diff --git a/.gitattributes b/.gitattributes index 59cdc5cae5fda7852e9952dae783e70ebada5851..4b44bbcbe0d53e35b94fd4932c44700535c7df43 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2938,3 +2938,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/GraphEcho_[[:space:]]Graph-Driven[[:space:]]Unsupervised[[:space:]]Domain[[:space:]]Adaptation[[:space:]]for[[:space:]]Echocardiogram[[:space:]]Video[[:space:]]Segmentation/594040d6-53b4-4c11-aa75-a08204fcd0de_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Graphics2RAW_[[:space:]]Mapping[[:space:]]Computer[[:space:]]Graphics[[:space:]]Images[[:space:]]to[[:space:]]Sensor[[:space:]]RAW[[:space:]]Images/aa2b7962-2dff-4878-8a58-f91afd985cfb_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/GridMM_[[:space:]]Grid[[:space:]]Memory[[:space:]]Map[[:space:]]for[[:space:]]Vision-and-Language[[:space:]]Navigation/3a3a473d-aae4-427d-8997-1caf744d0ceb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/GridPull_[[:space:]]Towards[[:space:]]Scalability[[:space:]]in[[:space:]]Learning[[:space:]]Implicit[[:space:]]Representations[[:space:]]from[[:space:]]3D[[:space:]]Point[[:space:]]Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Grounded[[:space:]]Entity-Landmark[[:space:]]Adaptive[[:space:]]Pre-Training[[:space:]]for[[:space:]]Vision-and-Language[[:space:]]Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Grounded[[:space:]]Image[[:space:]]Text[[:space:]]Matching[[:space:]]with[[:space:]]Mismatched[[:space:]]Relation[[:space:]]Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Grounding[[:space:]]3D[[:space:]]Object[[:space:]]Affordance[[:space:]]from[[:space:]]2D[[:space:]]Interactions[[:space:]]in[[:space:]]Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Group[[:space:]]DETR_[[:space:]]Fast[[:space:]]DETR[[:space:]]Training[[:space:]]with[[:space:]]Group-Wise[[:space:]]One-to-Many[[:space:]]Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Group[[:space:]]Pose_[[:space:]]A[[:space:]]Simple[[:space:]]Baseline[[:space:]]for[[:space:]]End-to-End[[:space:]]Multi-Person[[:space:]]Pose[[:space:]]Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/GrowCLIP_[[:space:]]Data-Aware[[:space:]]Automatic[[:space:]]Model[[:space:]]Growing[[:space:]]for[[:space:]]Large-scale[[:space:]]Contrastive[[:space:]]Language-Image[[:space:]]Pre-Training/84827675-3aac-4157-b576-07a125b493f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Growing[[:space:]]a[[:space:]]Brain[[:space:]]with[[:space:]]Sparsity-Inducing[[:space:]]Generation[[:space:]]for[[:space:]]Continual[[:space:]]Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Guided[[:space:]]Motion[[:space:]]Diffusion[[:space:]]for[[:space:]]Controllable[[:space:]]Human[[:space:]]Motion[[:space:]]Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Guiding[[:space:]]Image[[:space:]]Captioning[[:space:]]Models[[:space:]]Toward[[:space:]]More[[:space:]]Specific[[:space:]]Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Guiding[[:space:]]Local[[:space:]]Feature[[:space:]]Matching[[:space:]]with[[:space:]]Surface[[:space:]]Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/H3WB_[[:space:]]Human3.6M[[:space:]]3D[[:space:]]WholeBody[[:space:]]Dataset[[:space:]]and[[:space:]]Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HAL3D_[[:space:]]Hierarchical[[:space:]]Active[[:space:]]Learning[[:space:]]for[[:space:]]Fine-Grained[[:space:]]3D[[:space:]]Part[[:space:]]Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HDG-ODE_[[:space:]]A[[:space:]]Hierarchical[[:space:]]Continuous-Time[[:space:]]Model[[:space:]]for[[:space:]]Human[[:space:]]Pose[[:space:]]Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HM-ViT_[[:space:]]Hetero-Modal[[:space:]]Vehicle-to-Vehicle[[:space:]]Cooperative[[:space:]]Perception[[:space:]]with[[:space:]]Vision[[:space:]]Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HMD-NeMo_[[:space:]]Online[[:space:]]3D[[:space:]]Avatar[[:space:]]Motion[[:space:]]Generation[[:space:]]From[[:space:]]Sparse[[:space:]]Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HOSNeRF_[[:space:]]Dynamic[[:space:]]Human-Object-Scene[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HRS-Bench_[[:space:]]Holistic,[[:space:]]Reliable[[:space:]]and[[:space:]]Scalable[[:space:]]Benchmark[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HSE_[[:space:]]Hybrid[[:space:]]Species[[:space:]]Embedding[[:space:]]for[[:space:]]Deep[[:space:]]Metric[[:space:]]Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HSR-Diff_[[:space:]]Hyperspectral[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]via[[:space:]]Conditional[[:space:]]Diffusion[[:space:]]Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HTML_[[:space:]]Hybrid[[:space:]]Temporal-scale[[:space:]]Multimodal[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Referring[[:space:]]Video[[:space:]]Object[[:space:]]Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HaMuCo_[[:space:]]Hand[[:space:]]Pose[[:space:]]Estimation[[:space:]]via[[:space:]]Multiview[[:space:]]Collaborative[[:space:]]Self-Supervised[[:space:]]Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HairCLIPv2_[[:space:]]Unifying[[:space:]]Hair[[:space:]]Editing[[:space:]]via[[:space:]]Proxy[[:space:]]Feature[[:space:]]Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HairNeRF_[[:space:]]Geometry-Aware[[:space:]]Image[[:space:]]Synthesis[[:space:]]for[[:space:]]Hairstyle[[:space:]]Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hallucination[[:space:]]Improves[[:space:]]the[[:space:]]Performance[[:space:]]of[[:space:]]Unsupervised[[:space:]]Visual[[:space:]]Representation[[:space:]]Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HandR2N2_[[:space:]]Iterative[[:space:]]3D[[:space:]]Hand[[:space:]]Pose[[:space:]]Estimation[[:space:]]Using[[:space:]]a[[:space:]]Residual[[:space:]]Recurrent[[:space:]]Neural[[:space:]]Network/431e0555-f342-4159-8204-85714e05b7c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Handwritten[[:space:]]and[[:space:]]Printed[[:space:]]Text[[:space:]]Segmentation_[[:space:]]A[[:space:]]Signature[[:space:]]Case[[:space:]]Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hard[[:space:]]No-Box[[:space:]]Adversarial[[:space:]]Attack[[:space:]]on[[:space:]]Skeleton-Based[[:space:]]Human[[:space:]]Action[[:space:]]Recognition[[:space:]]with[[:space:]]Skeleton-Motion-Informed[[:space:]]Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Harnessing[[:space:]]the[[:space:]]Spatial-Temporal[[:space:]]Attention[[:space:]]of[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]High-Fidelity[[:space:]]Text-to-Image[[:space:]]Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Harvard[[:space:]]Glaucoma[[:space:]]Detection[[:space:]]and[[:space:]]Progression_[[:space:]]A[[:space:]]Multimodal[[:space:]]Multitask[[:space:]]Dataset[[:space:]]and[[:space:]]Generalization-Reinforced[[:space:]]Semi-Supervised[[:space:]]Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hashing[[:space:]]Neural[[:space:]]Video[[:space:]]Decomposition[[:space:]]with[[:space:]]Multiplicative[[:space:]]Residuals[[:space:]]in[[:space:]]Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Helping[[:space:]]Hands_[[:space:]]An[[:space:]]Object-Aware[[:space:]]Ego-Centric[[:space:]]Video[[:space:]]Recognition[[:space:]]Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Heterogeneous[[:space:]]Diversity[[:space:]]Driven[[:space:]]Active[[:space:]]Learning[[:space:]]for[[:space:]]Multi-Object[[:space:]]Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Heterogeneous[[:space:]]Forgetting[[:space:]]Compensation[[:space:]]for[[:space:]]Class-Incremental[[:space:]]Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HiFace_[[:space:]]High-Fidelity[[:space:]]3D[[:space:]]Face[[:space:]]Reconstruction[[:space:]]by[[:space:]]Learning[[:space:]]Static[[:space:]]and[[:space:]]Dynamic[[:space:]]Details/a0404738-6577-4978-90b9-452e8141d3d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HiLo_[[:space:]]Exploiting[[:space:]]High[[:space:]]Low[[:space:]]Frequency[[:space:]]Relations[[:space:]]for[[:space:]]Unbiased[[:space:]]Panoptic[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HiTeA_[[:space:]]Hierarchical[[:space:]]Temporal-Aware[[:space:]]Video-Language[[:space:]]Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HiVLP_[[:space:]]Hierarchical[[:space:]]Interactive[[:space:]]Video-Language[[:space:]]Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hidden[[:space:]]Biases[[:space:]]of[[:space:]]End-to-End[[:space:]]Driving[[:space:]]Models/8f4d684a-1555-4674-8f25-206ee2d4af90_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hiding[[:space:]]Visual[[:space:]]Information[[:space:]]via[[:space:]]Obfuscating[[:space:]]Adversarial[[:space:]]Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchical[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Pattern-Generalizable[[:space:]]Image[[:space:]]Corruption[[:space:]]Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchical[[:space:]]Generation[[:space:]]of[[:space:]]Human-Object[[:space:]]Interactions[[:space:]]with[[:space:]]Diffusion[[:space:]]Probabilistic[[:space:]]Models/c8490beb-3559-4e1e-833f-b399932e108c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchical[[:space:]]Point-based[[:space:]]Active[[:space:]]Learning[[:space:]]for[[:space:]]Semi-supervised[[:space:]]Point[[:space:]]Cloud[[:space:]]Semantic[[:space:]]Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchical[[:space:]]Prior[[:space:]]Mining[[:space:]]for[[:space:]]Non-local[[:space:]]Multi-View[[:space:]]Stereo/32540140-b677-4721-9ff6-75d372d0916d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchical[[:space:]]Spatio-Temporal[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Gait[[:space:]]Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchical[[:space:]]Visual[[:space:]]Categories[[:space:]]Modeling_[[:space:]]A[[:space:]]Joint[[:space:]]Representation[[:space:]]Learning[[:space:]]and[[:space:]]Density[[:space:]]Estimation[[:space:]]Framework[[:space:]]for[[:space:]]Out-of-Distribution[[:space:]]Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchical[[:space:]]Visual[[:space:]]Primitive[[:space:]]Experts[[:space:]]for[[:space:]]Compositional[[:space:]]Zero-Shot[[:space:]]Learning/d892901c-ab65-4154-a39e-e041473b0ce9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Hierarchically[[:space:]]Decomposed[[:space:]]Graph[[:space:]]Convolutional[[:space:]]Networks[[:space:]]for[[:space:]]Skeleton-Based[[:space:]]Action[[:space:]]Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/High[[:space:]]Quality[[:space:]]Entity[[:space:]]Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/High-Resolution[[:space:]]Document[[:space:]]Shadow[[:space:]]Removal[[:space:]]via[[:space:]]A[[:space:]]Large-Scale[[:space:]]Real-World[[:space:]]Dataset[[:space:]]and[[:space:]]A[[:space:]]Frequency-Aware[[:space:]]Shadow[[:space:]]Erasing[[:space:]]Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Holistic[[:space:]]Geometric[[:space:]]Feature[[:space:]]Learning[[:space:]]for[[:space:]]Structured[[:space:]]Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Holistic[[:space:]]Label[[:space:]]Correction[[:space:]]for[[:space:]]Noisy[[:space:]]Multi-Label[[:space:]]Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HollowNeRF_[[:space:]]Pruning[[:space:]]Hashgrid-Based[[:space:]]NeRFs[[:space:]]with[[:space:]]Trainable[[:space:]]Collision[[:space:]]Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HoloAssist_[[:space:]]an[[:space:]]Egocentric[[:space:]]Human[[:space:]]Interaction[[:space:]]Dataset[[:space:]]for[[:space:]]Interactive[[:space:]]AI[[:space:]]Assistants[[:space:]]in[[:space:]]the[[:space:]]Real[[:space:]]World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HoloFusion_[[:space:]]Towards[[:space:]]Photo-realistic[[:space:]]3D[[:space:]]Generative[[:space:]]Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Homeomorphism[[:space:]]Alignment[[:space:]]for[[:space:]]Unsupervised[[:space:]]Domain[[:space:]]Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Homography[[:space:]]Guided[[:space:]]Temporal[[:space:]]Fusion[[:space:]]for[[:space:]]Road[[:space:]]Line[[:space:]]and[[:space:]]Marking[[:space:]]Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/HopFIR_[[:space:]]Hop-wise[[:space:]]GraphFormer[[:space:]]with[[:space:]]Intragroup[[:space:]]Joint[[:space:]]Refinement[[:space:]]for[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Householder[[:space:]]Projector[[:space:]]for[[:space:]]Unsupervised[[:space:]]Latent[[:space:]]Semantics[[:space:]]Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/How[[:space:]]Far[[:space:]]Pre-trained[[:space:]]Models[[:space:]]Are[[:space:]]from[[:space:]]Neural[[:space:]]Collapse[[:space:]]on[[:space:]]the[[:space:]]Target[[:space:]]Dataset[[:space:]]Informs[[:space:]]their[[:space:]]Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/How[[:space:]]Much[[:space:]]Temporal[[:space:]]Long-Term[[:space:]]Context[[:space:]]is[[:space:]]Needed[[:space:]]for[[:space:]]Action[[:space:]]Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/How[[:space:]]to[[:space:]]Boost[[:space:]]Face[[:space:]]Recognition[[:space:]]with[[:space:]]StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/How[[:space:]]to[[:space:]]Choose[[:space:]]your[[:space:]]Best[[:space:]]Allies[[:space:]]for[[:space:]]a[[:space:]]Transferable[[:space:]]Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Human[[:space:]]Part-wise[[:space:]]3D[[:space:]]Motion[[:space:]]Context[[:space:]]Learning[[:space:]]for[[:space:]]Sign[[:space:]]Language[[:space:]]Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_content_list.json b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d46d5282380ee35ff227fd5ff02243a99f18a251 --- /dev/null +++ b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_content_list.json @@ -0,0 +1,2299 @@ +[ + { + "type": "text", + "text": "GridPull: Towards Scalability in Learning Implicit Representations from 3D Point Clouds", + "text_level": 1, + "bbox": [ + 99, + 128, + 870, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chao Chen $^{1}$ Yu-Shen Liu $^{1}$ * Zhizhong Han $^{2}$", + "bbox": [ + 292, + 202, + 674, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Software, Tsinghua University, Beijing, China", + "bbox": [ + 276, + 222, + 692, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, Wayne State University, Detroit, USA", + "bbox": [ + 217, + 239, + 751, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenchao19@mails.tsinghua.edu.cn liuyushen@tsinghua.edu.cn h312h@wayne.edu", + "bbox": [ + 140, + 258, + 828, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning implicit representations has been a widely used solution for surface reconstruction from 3D point clouds. The latest methods infer a distance or occupancy field by overfitting a neural network on a single point cloud. However, these methods suffer from a slow inference due to the slow convergence of neural networks and the extensive calculation of distances to surface points, which limits them to small scale points. To resolve the scalability issue in surface reconstruction, we propose GridPull to improve the efficiency of learning implicit representations from large scale point clouds. Our novelty lies in the fast inference of a discrete distance field defined on grids without using any neural components. To remedy the lack of continuousness brought by neural networks, we introduce a loss function to encourage continuous distances and consistent gradients in the field during pulling queries onto the surface in grids near to the surface. We use uniform grids for a fast grid search to localize sampled queries, and organize surface points in a tree structure to speed up the calculation of distances to the surface. We do not rely on learning priors or normal supervision during optimization, and achieve superiority over the latest methods in terms of complexity and accuracy. We evaluate our method on shape and scene benchmarks, and report numerical and visual comparisons with the latest methods to justify our effectiveness and superiority. The code is available at https://github.com/chenchao15/GridPull.", + "bbox": [ + 75, + 340, + 472, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 781, + 209, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is vital to reconstruct surfaces from 3D point clouds for downstream applications. A widely used strategy is to learn", + "bbox": [ + 76, + 806, + 468, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "implicit representations [59, 62, 26, 10, 91, 77, 54, 68, 88] from 3D point clouds in a data-driven manner. With the learned implicit representations, we can reconstruct surfaces in meshes by running the marching cubes algorithm [48]. By learning priors from large scale datasets during training, previous methods [57, 21, 31, 46, 78, 73, 17] generalize the learned priors to either global implicit functions [89, 80, 8, 32, 6, 50] or local ones for unseen point clouds. However, the generalization ability limits their performances on large structure or geometry variations of training samples.", + "bbox": [ + 500, + 309, + 890, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "More recent methods [23, 12, 2, 100, 3, 95, 4, 49, 14, 65, 35] achieved better generalization by directly overfitting neural networks on single unseen point clouds. They rely on extensive calculations of distances between queries and surface points to probe the space, which supervises neural networks to converge to a distance or occupancy field [49, 23, 2, 100, 3, 11, 52, 83, 40]. However, this inference procedure is time consuming due to the extensive distance calculations and the slow convergence of neural networks. This demerit makes these methods hard to scale up to large scale point clouds for surface reconstruction.", + "bbox": [ + 496, + 477, + 892, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the scalability challenge, we propose GridPull to speed up the learning of implicit function from large scale point clouds. GridPull does not require learned priors or point normal, and directly infers a distance field from a point cloud without using any neural components. We infer the distance field on grids near the surface, which reduces the number of grids we need to infer. Moreover, we organize surface points in a tree structure to speed up the nearest neighbor search for the calculation of distances to the surface. Specifically, we infer the discrete distance field by pulling queries onto the surface in grids of interests. Our loss function encourages continuous distances and consistent gradients in the field, which makes up the lack of continuousness brought by neural networks. We justify our effectiveness and highlight our superiority by numerical and visual comparisons with the latest methods on the widely used benchmarks. Our contributions are listed below.", + "bbox": [ + 496, + 643, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The corresponding author is Yu-Shen Liu. This work was supported by National Key R&D Program of China (2022YFC3800600), the National Natural Science Foundation of China (62272263, 62072268), and in part by Tsinghua-Kuaishou Institute of Future Media Data.", + "bbox": [ + 75, + 851, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18322", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "i) We propose GridPull to reconstruct surfaces from large scale point clouds without using neural networks. GridPull speeds up the learning of implicit function, which addresses the scalability challenge in surface reconstruction.", + "ii) We introduce a loss function to directly infer a discrete distance field defined on grids but achieving continuous distances and consistent gradients in the field.", + "iii) Our method outperforms state-of-the-art methods in surface reconstruction in terms of speed and accuracy on the widely used benchmarks." + ], + "bbox": [ + 80, + 90, + 467, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 297, + 217, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural implicit representations have made a huge progress in various tasks [59, 62, 26, 10, 101, 91, 77, 54, 68, 53, 39, 36, 60, 28, 7, 41, 76]. We can use different supervision including 3D supervision [58, 64, 63, 56, 13, 92, 87], multi-view [72, 45, 34, 99, 44, 90, 61, 42, 94, 93, 19, 84, 98, 86, 81, 82, 25, 22, 29, 69, 74, 33, 55], and point clouds [89, 43, 57, 21] to learn neural implicit representations. In the following, we focus on reviewing works on surface reconstruction by learning implicit representations point clouds below.", + "bbox": [ + 75, + 324, + 467, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning with Priors. With neural networks, one intuitive strategy is to use neural networks to learn priors from a training set and then generalize the learned priors to unseen samples. We can learn either global priors to map a shape level point cloud into a global implicit function [57, 21, 31, 46, 78, 73, 17] or local priors for local implicit functions to represent parts or patches [89, 80, 8, 32, 6, 50] which are further used to approximate a global implicit function.", + "bbox": [ + 75, + 476, + 467, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These methods require large scale datasets to learn priors. However, the learned priors may not generalize well to unseen point clouds that have large geometric variations compared to training samples. Our method does not require priors, and falls into the following category.", + "bbox": [ + 75, + 613, + 467, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning with Overfitting. To improve the generalization, we can learn implicit functions by overfitting neural networks on single point clouds. Methods using this strategy introduce novel constraints [23, 2, 100, 3, 95, 4], ways of leveraging gradients [49, 14], differentiable poisson solver [65] or specially designed priors [50, 51] to learn signed [49, 23, 2, 100, 3, 11] or unsigned distance functions [14, 102, 47].", + "bbox": [ + 75, + 688, + 467, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These methods rely on neural networks to infer implicit functions without learning priors. However, the slow convergence of neural networks is a big limit for them to scale up to large scale point clouds. Our method addresses the scalability challenge by directly inferring a discrete distance field without using neural networks. Our novel loss leads", + "bbox": [ + 75, + 810, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to less complexity and higher accuracy than the overfitting based methods.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning with Grids. Learning implicit functions with grids has been used in prior-based methods [77, 54, 66, 79, 46, 34, 85]. Using neural networks, these methods learn features at vertices of grids and further map the interpolated features into a distance or occupancy field. Some methods [70, 75, 96, 38] inferred a discrete radiance field defined on grids to speed up the learning of a radiance field, while they cared more about the quality of synthesized views than the underlying geometry. Without using neural components, some methods [67, 65] directly infer discrete implicit functions on grids. To make up the continuousness of neural networks, they employed different constraints, such as Possion equations or Viscosity priors, to pursue a continuous field from a single point cloud. However, these methods are slowly converged, which makes them hard to scale up to large scale point clouds. Although we also use grids to speed up the inference, we introduce a more efficient training strategy to decrease optimization space and more effective losses to increase inference efficiency.", + "bbox": [ + 496, + 119, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 420, + 589, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overview. GridPull aims to achieve a fast inference of a distance field, such as a signed distance field, from a 3D point cloud $S = \\{s_j | j \\in [1, J]\\}$ without using any neural components. As shown in Fig. 1 (b), we represent the distance field by a signed distance function (SDF) $f$ defined on a set of discrete grids $G$ with a resolution of $R^3$ , where the vertices $V = \\{v_i | i \\in [1, I], I = (R + 1)^3\\}$ shared among these grids hold learnable signed distances $D = \\{d_i\\}$ . For an arbitrary location $q$ in Fig. 1 (e), such as a randomly sampled query or a surface point, we use trilinear interpolation to approximate the signed distance $f(q)$ from the eight nearest grid vertices below,", + "bbox": [ + 496, + 446, + 890, + 626 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf (q) = \\operatorname {t r i I n t e r} (q, D). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 642, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to conduct direct optimization in Fig. 1 (c) to infer the signed distances $D$ on grid vertices by minimizing our loss function $L$ on queries $q$ below,", + "bbox": [ + 496, + 662, + 890, + 709 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {D} L (f (q)). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 723, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the learned function $f$ , we run the marching cubes algorithm [48] to reconstruct a mesh as the surface as shown in Fig. 1 (d).", + "bbox": [ + 496, + 750, + 890, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Discrete Grids. We split the space occupied by the point cloud $S$ into discrete grids $G$ with a resolution of $R$ in Fig. 1 (b). GridPull can work with high resolutions, such as $R = 512$ , because of our efficient and effective loss function and training strategy. Within each grid, we interpolate the signed distance $f(q)$ of query $q$ using the trilinear interpolation in Eq. 1. We organize surface points in a tree", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18323", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08f150614fe630839ea0ac6e26ac5b2b1d0c579a5990a364a1b683d9b7cebbe8.jpg", + "image_caption": [ + "(a)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 145, + 106, + 285, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/76e610d2943f6f27312a06785d9ac9c9d7367472872d7f7efba2d0a24f353c92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 89, + 442, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8690051359f5626af01eb76e2a360584f30fdcb4748706ab0a49e2b01e1baab1.jpg", + "image_caption": [ + "(c)", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 442, + 89, + 707, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd13e9eb07f3043383b1a37bd78aa1ddd5a6eccbe4649f185cff8405e64bd0be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 106, + 795, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b53338c2fa128870cd35d21e28f143ad68a8ea9e806f5dc4f35e28675ad9801.jpg", + "image_caption": [ + "Figure 1. We proposed GridPull for fast inference of SDF defined on discrete grids using effective losses." + ], + "image_footnote": [], + "bbox": [ + 76, + 228, + 245, + 351 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b5487feb83f0d32d23ec8b6d9557ce9d8c53c817fd30259cf13ae3a9593b9ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 228, + 375, + 347 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2da1b71496b018182b8f8dcde2174440a5a555c983587028f59e98cd9d0a6c23.jpg", + "image_caption": [ + "(g) Total Variation Loss" + ], + "image_footnote": [], + "bbox": [ + 397, + 220, + 555, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33498d87d523024d17748da6e4788e52089a07a76e909f8d7114a435685bc580.jpg", + "image_caption": [ + "(h) Surface Loss" + ], + "image_footnote": [], + "bbox": [ + 558, + 220, + 722, + 325 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23a7055f4c28076833010b9ef76ab69350afabfe54a82aaed36c36f15355ccb6.jpg", + "image_caption": [ + "(i) Gradient Loss" + ], + "image_footnote": [], + "bbox": [ + 720, + 228, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "structure to speed up the search of the nearest surface point for queries $q$ . Since we need to extensively conduct trilinear interpolation on signed distances for queries $q$ , we use uniform grids for fast nearest grid search for queries. We did not use an Octree to organize grids, due to its inefficient and inaccurate grid search for queries.", + "bbox": [ + 75, + 363, + 468, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometric Initialization. We initialize the signed distances $d_{i}$ on grid vertices which represent a sphere in the field. The initialization can speed up the convergence during optimization. Our preliminary results show that a small sphere with a radius of one or two grids works well. Since enlarging a small sphere to the target shape can calculate the continuous distance loss over smaller regions than shrinking a large sphere.", + "bbox": [ + 75, + 452, + 468, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Signed Distance Inference. For a query $q$ , we infer its signed distance $f(q)$ through pulling it onto its nearest surface point in the point cloud $S$ . As shown in Fig. 1 (e), we first obtain $f(q)$ as the interpolation of the signed distances $d_{i}$ on the nearest eight vertices of $q$ . We implement the trilinear interpolation by solving $\\mathbf{a} = [a_0,..,a_7]$ in a linear regression system on the eight vertices $v_{i}$ below,", + "bbox": [ + 75, + 574, + 468, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (q) = a _ {0} + a _ {1} x + a _ {2} y + a _ {3} z + a _ {4} x y + a _ {5} x z + a _ {6} y z + a _ {7} x y z, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 694, + 468, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[x,y,z]$ is the coordinate of $q$ and the eight constraints are $\\{f(v_i) = d_i\\}$ .", + "bbox": [ + 75, + 724, + 467, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the predicted signed distance $f(q)$ and the gradient $\\nabla f(q) = \\partial f / \\partial q$ , we pull a query $q$ onto the surface and obtain a pulled query $p$ by $p = q - f(q) \\times \\nabla f(q) / ||\\nabla f(q)||_2$ in Fig. 1 (f).", + "bbox": [ + 75, + 753, + 468, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to minimize the distance between the pulled query $p$ and the nearest surface point $NN(q) \\in S$ of query $q$ below,", + "bbox": [ + 75, + 815, + 467, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {P u l l} = \\left\\| p - N N (q) \\right\\| _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 862, + 468, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt this pulling loss from [49]. Instead, we do not", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rely on neural network in the pulling procedure. We use the trilinear interpolation on signed distances $d_{i}$ on grid vertices to obtain signed distances and gradients according to Eq. 3.", + "bbox": [ + 496, + 363, + 890, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We sample queries $q$ around surface points in each epoch. For each randomly sampled surface point, we sample queries using a gaussian distribution with the surface point as the center and a two-grid length variance. We build a KD-tree on surface points so that we can speed up the searching of the nearest neighbor from a large number of surface points for each query.", + "bbox": [ + 496, + 410, + 890, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To speed up the optimization, we only calculate the pulling loss on grids near the surface. We regard the grids containing surface points as the base and find the union of their $M_{1}$ nearest neighboring grids. These grids form a $M_{1}$ -bandwidth near the surface in Fig. 1 (c).", + "bbox": [ + 496, + 517, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continuous Distance Fields. Although the pulling loss can infer correct signed distances at grid vertices, it fails to learn continuous signed distances across the field, especially near the surface. Fig. 2 (a) shows severe artifacts on the surface and gradients near the surface are messy and inconsistent, which are caused by the discontinuous signed distances across neighboring regions.", + "bbox": [ + 496, + 595, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural network based methods [49, 51] do not have this issue, since the continuous character of neural networks makes the sudden change of signs on the same shape side hard. While our method uses signed distances $d_{i}$ on grid vertices to represent a discrete field, this makes that one $d_{i}$ can change separately without considering the change of neighboring $d_{i}$ , which leads to a poor continuousness in the field.", + "bbox": [ + 496, + 702, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To resolve this issue, we impose a total variation (TV) loss as a continuous constraint on our signed distances $d_{i}$ on grid vertices. Our key idea is to make the change of one $d_{i}$ affect the change of its neighboring $NN(d_{i})$ , which smoothes the local distance field and propagates local up", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18324", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14712135528e45347a8f1b81ce55d4d449a3f94d17b61fc1ce25ec57ec91b9eb.jpg", + "image_caption": [ + "Figure 2. The effect of losses on reconstruction.", + "dates to the rest grids. Hence, as shown in a 2D case in Fig. 1 (g), we approximate one $d_{i}$ to each one of its six neighbors by minimizing the difference between each two below," + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 472, + 234 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T V} = \\sqrt {\\sum_ {a = \\{x , y , z \\}} e _ {a + 1} \\left(d _ {i}\\right) ^ {2} + e _ {a - 1} \\left(d _ {i}\\right) ^ {2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 318, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e_{a + 1}(d_i)$ is the difference between $d_i$ and the signed distance along the $a$ axis, i.e., $x, y, z$ axis. Fig. 2 (b) shows that the continuous constraints can significantly improve the signed distance field in terms of continuousness and gradients.", + "bbox": [ + 75, + 362, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One remaining issue is its computational complexity. Since we are targeting grids at a high resolution, such as $R = 512$ , it is very time consuming to impose this constraint on all $d_{i}$ across the field.", + "bbox": [ + 76, + 436, + 468, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To resolve this issue, we focus on grids near the surface. Similar to imposing the pulling loss on grids in the $M_{1}$ -bandwidth near the surface, we impose our continuous constraints on the grids in the $M_{2}$ -bandwidth near the surface, which significantly reduces the number of learnable parameters to optimize.", + "bbox": [ + 75, + 498, + 468, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Signed Distance Supervision. For signed distances, one supervision we can directly use is the input point cloud $S$ . It is a discrete surface indicating the zero level-set of the signed distance function. Hence, as introduced in Fig. 1 (h), we can constrain the signed distances at the points $s$ on the surface below,", + "bbox": [ + 75, + 588, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {S u r f a c e}} = \\left\\| f (s) \\right\\| _ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 680, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although Fig. 2 (c) indicates that $L_{Surface}$ slightly improves the signed distance field and leads to a little bit more continuous surface, it is helpful to determine the zero level set with other losses. For noises on $S$ , our continuous constraint can relieve the impact of noises on this supervision. We will show this in experiments.", + "bbox": [ + 75, + 704, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consistent Gradients. Fig. 2 (a) indicates that the pulling loss produces gradients that are messy and non-orthogonal to the surface. This leads to inaccurate signed distance inference during the pulling procedure.", + "bbox": [ + 75, + 795, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We additionally introduce a constraint on gradients to improve the field near the surface. As shown in Fig. 1 (i), we aim to encourage the gradient at a query $q$ to point to the", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "same direction of the gradient at its nearest surface point $NN(q)$ . We use a cosine distance below to achieve more consistent gradients,", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\nabla} = 1 - \\cos (\\nabla f (q), \\nabla f (N N (q))). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 138, + 890, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{\\nabla}$ constrains the gradients in the field and corrects inaccurate gradients. Fig. 2 (d) shows that more consistent gradients can improve the surface to be more continuous using small planes on the surface which however the surface not smooth at all.", + "bbox": [ + 498, + 169, + 890, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Function. We learn signed distances on grid vertices $D$ by minimizing the loss function below,", + "bbox": [ + 500, + 244, + 890, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {P u l l}} + \\alpha L _ {T V} + \\beta L _ {\\text {S u r f a c e}} + \\gamma L _ {\\nabla}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 277, + 890, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ , $\\beta$ , and $\\gamma$ are balance weights to make each term contribute equally. Fig. 3 shows the signed distance field on a slide of $D$ during optimization. Our optimization starts from an initial sphere and only optimizes variables in the bandwidth. We skip the grids that are not optimized after the initialization during the marching cubes for surface reconstruction. Please watch our video for more details.", + "bbox": [ + 496, + 300, + 890, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments and Analysis", + "text_level": 1, + "bbox": [ + 498, + 421, + 743, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate GridPull by numerical and visual comparisons with the latest methods on synthetic and real datasets in surface reconstruction.", + "bbox": [ + 496, + 446, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets and Metrics. We use benchmarks for shapes and scenes in evaluations. For shapes, we conduct evaluations on five datasets including a subset of ShapeNet [1], FAMOUS [17], Thingi10k [103], Surface Reconstruction Benchmark (SRB) [89] and D-FAUST [5]. For scenes, we also report comparisons on five datasets including 3DScene [104], SceneNet [27], 3DFRONT [18], Matterport [9] and KITTI [20].", + "bbox": [ + 496, + 492, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use L2 Chamfer distance $(CD_{L2})$ , $CD_{L1}$ and Hausdorff distance(HD) to measure the error between the reconstructed surface and the ground truth. Moreover, we use normal consistency (NC) and F-score to evaluate the accuracy of normal on the reconstructed surface. We also report our time and storage complexity to highlight our advantage towards scalability. We also use the intersection over union (IoU) to measure the reconstruction for fair comparison with the latest methods.", + "bbox": [ + 496, + 613, + 890, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Details. We use $R = 256$ to infer a distance field. We calculate the pulling loss on grids in a bandwidth with $M_{1} = 3$ . We impose the continuous constraint on grids in a bandwidth with $M_{2} = 14$ for shapes and $M_{2} = 25$ for scenes, since scenes contain open surfaces with larger empty spaces. We run the marching cubes for surface reconstruction at a resolution of 256.", + "bbox": [ + 496, + 750, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the Adam optimizer with an initial learning rate of 1.0. We decrease the learning rate with a decay rate of 0.3 every 400 iterations. We use 50,000 queries in each", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18325", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb201be88e6afbc6a4dc5e742d25ee257e3bb221f91021bf5fbfb201996b9b34.jpg", + "image_caption": [ + "0th Iteration" + ], + "image_footnote": [], + "bbox": [ + 81, + 75, + 215, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27abef32dbbe34dec1596d0ba6aab2d3cf3a43d5ee0f85709d5f62edc7bd2fc4.jpg", + "image_caption": [ + "50th Iteration" + ], + "image_footnote": [], + "bbox": [ + 217, + 77, + 349, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d676db783771bad12e0d9d115cbaa9ba3d1d6b26c5b8c84b62d9d2d23d0a3276.jpg", + "image_caption": [ + "100th Iteration" + ], + "image_footnote": [], + "bbox": [ + 351, + 77, + 482, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4bad84de10b76dcd23cd53ac6bfb4617f79a26cd074e09af20a9a05e34c9fc7.jpg", + "image_caption": [ + "150th Iteration" + ], + "image_footnote": [], + "bbox": [ + 483, + 77, + 616, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30b683a5ee4e3421fd1acaeaf3ac760458a43704d20ce825d1e0b9bd6d4c4e62.jpg", + "image_caption": [ + "200th Iteration" + ], + "image_footnote": [], + "bbox": [ + 617, + 77, + 750, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/10b3e5fbc48694f67a753a9571d9727bc46adc17040db454ae645aec55ca26cd.jpg", + "image_caption": [ + "1000th Iteration" + ], + "image_footnote": [], + "bbox": [ + 751, + 77, + 885, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31cff11eaad808fa8900f90bf875e00b8b4febf4cd424e84e47c335acb46bf1c.jpg", + "image_caption": [ + "Figure 3. Visualization of signed distances in optimization.", + "Figure 4. Visual comparison under ShapetNet dataset." + ], + "image_footnote": [], + "bbox": [ + 80, + 179, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iteration, and run 1600 iterations for each point cloud during optimization. We set the balance weights $\\alpha = 1$ , $\\beta = 1$ , and $\\gamma = 0.005$ for equal contribution from the three terms.", + "bbox": [ + 76, + 417, + 468, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Surface Reconstruction for Shapes", + "text_level": 1, + "bbox": [ + 76, + 470, + 379, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ShapeNet. Following predictive context prior (PCP) [51] and NeuralPull (NP) [49], we report our results on a subset in ShapeNet in terms of $CD_{L2}$ in Tab. 1, Normal Consistency (NC) in Tab. 2, and F-Score with thresholds of 0.002 and 0.004 in Tab. 3 and Tab. 4, which show our superiority over PCP and NP in numerical comparisons over all classes, even though PCP employs a local shape prior learned from a large scale dataset. The visual comparison in Fig. 4 shows that we can reveal more accurate and detailed geometry.", + "bbox": [ + 75, + 494, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5f62ea01f2a104405f9ccdf3b8e8689c1cc757dbce941f38598de672fcbe0d73.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ClassNP [49]PCP [51]Ours
Display0.0390.00870.0082
Lamp0.0800.03800.0347
Airplane0.0080.00650.0007
Cabinet0.0260.01530.0112
Vessel0.0220.00790.0033
Table0.0600.01310.0052
Chair0.0540.01100.0043
Sofa0.0120.00860.0015
Mean0.0380.01360.0086
", + "bbox": [ + 138, + 633, + 410, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FAMOUS. We follow the experimental setting in PCP [51] and NP [49] to evaluate GridPull on FAMOUS. We compare GridPull with the latest methods with priors including PCP, GenSDF, FGC and no priors including NP and IGR. GridPull outperforms these methods in terms of $CD_{L2}$ and running time. Although FGC is fast, it does not reconstruct", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ccc7c3f3bc66a7920fcad613e6fedf83a1aa949502913cf7cd2d761df47d9fc5.jpg", + "table_caption": [ + "Table 1. Reconstruction accuracy under ShapeNet in terms of $CD_{L2} \\times 100$ ." + ], + "table_footnote": [], + "table_body": "
ClassNP [49]PCP [51]Ours
Display0.9640.97750.9847
Lamp0.9300.94500.9693
Airplane0.9470.94900.9614
Cabinet0.9300.96000.9689
Vessel0.9410.95460.9667
Table0.9080.95950.9755
Chair0.9370.95800.9733
Sofa0.9510.96800.9792
Mean0.9390.95900.9723
", + "bbox": [ + 560, + 185, + 831, + 327 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/408516d0d80f45dd19d03e0f22eea0ea81d36d76006438cc3050e9189b6b62f1.jpg", + "table_caption": [ + "Table 2. Reconstruction accuracy under ShapeNet in terms of NC." + ], + "table_footnote": [], + "table_body": "
ClassNP [49]PCP [51]Ours
Display0.9890.99390.9963
Lamp0.8910.93820.9455
Airplane0.9960.99420.9976
Cabinet0.9800.98880.9901
Vessel0.9850.99350.9956
Table0.9220.99690.9977
Chair0.9540.99700.9979
Sofa0.9680.99430.9974
Mean0.9610.98710.9896
", + "bbox": [ + 560, + 347, + 831, + 489 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1116c82b1fc2f147b6d41dadbef3b1213645e344012cbe6a8787622dc5f5d651.jpg", + "table_caption": [ + "Table 3. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.002." + ], + "table_footnote": [], + "table_body": "
ClassNP [49]PCP [51]Ours
Display0.9910.99580.9963
Lamp0.9240.94020.9538
Airplane0.9970.99720.9989
Cabinet0.9890.99390.9946
Vessel0.9900.99580.9972
Table0.9730.99850.9990
Chair0.9690.99910.9990
Sofa0.9740.99870.9992
Mean0.9760.98990.9923
", + "bbox": [ + 560, + 521, + 831, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a watertight mesh but just a polygon soup with no normal, which contains almost no geometry details. Worse than that, it fails to reconstruct some point clouds in more than eight hours, hence we remove these shapes from its results. We highlight our superiority in visual comparison in Fig. 5. We can reveal details like hair and expressions more clearly.", + "bbox": [ + 500, + 699, + 890, + 789 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/17e782dffee1c3a9698ea9a9ad7ab145e3d4597f7a5201f7482d2cf69b93de8c.jpg", + "table_caption": [ + "Table 4. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.004." + ], + "table_footnote": [], + "table_body": "
MethodCDL2×100Min timeMean time
IGR [24]1.650402s407s
GenSDF [15]0.668230s232s
NP [49]0.220593s605s
FGC [97]0.05510sTimeout
PCP [51]0.0443416s3534s
Ours0.040198s201s
", + "bbox": [ + 540, + 801, + 849, + 890 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 5. Reconstruction accuracy under FAMOUS.", + "bbox": [ + 542, + 893, + 846, + 906 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18326", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f86dd0e21b1991b0f68701504a745ceb960bcd52bb17aa918ce78f4c79169300.jpg", + "image_caption": [ + "Figure 5. Visual comparison under FAMOUS dataset." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 467, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d1af343e0df8fb9ceaec63400ed409efda84caa525fc04d7c9419a5cd3d9be3.jpg", + "image_caption": [ + "Figure 6. Visual comparison under SRB dataset." + ], + "image_footnote": [], + "bbox": [ + 78, + 321, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SRB. We follow the experimental setting in VisCo [67] to evaluate GridPull on real scans in SRB. We report the comparison with the latest overfitting based methods in terms of $CD_{L1}$ in Tab. 6. Meanwhile, we also report their one-sided distances between the reconstructed mesh and the input noisy point cloud. The numerical comparison shows that our method significantly outperforms the latest. Moreover, we also report our results of learning unsigned distances in Tab. 7. We replace our pulling loss with the pulling loss introduced for unsigned distances in CAP-UDF [102], and remove the gradient loss due to the nondifferentiable character of UDF on surface. Visual comparison in Fig. 6 shows that our method can reconstruct surfaces that are more compact than the methods learning SDF like SIREN or more smooth than the methods learning UDF like CAP-UDF.", + "bbox": [ + 76, + 532, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thingi10K. We follow the experimental setting in SAP [65] to evaluate GridPull in Thingi10K. Tab. 8 indicates our superiority over the overfitting based methods in the numerical comparison. Visual comparisons in Fig. 7 show our more accurate surfaces with complex details.", + "bbox": [ + 76, + 762, + 468, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "D-FAUST. We also follow the experimental setting in S-AP [65] to evaluate GridPull in D-FAUST. Tab. 8 shows that we outperform the overfitting based methods with more surface details. As shown in Fig. 7, we recover more accurate", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c9e5fe75be731eacddc7b593e7f699d2664e4b0df1449b02f5e1f2f0d8971220.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Poisson [37]IGR [24]SIREN [71]VisCo [67]SAP [65]Ours
Anchor\\(CD_{L1}\\)0.600.220.320.210.120.093
\\(HD\\)14.894.718.193.002.381.804
\\(d_{\\vec{C}}\\)0.600.120.100.150.080.066
\\(d_{\\vec{H}}\\)14.891.322.431.070.830.460
Daratech\\(CD_{L1}\\)0.440.250.210.260.070.062
\\(HD\\)7.244.014.304.060.870.648
\\(d_{\\vec{C}}\\)0.440.080.090.140.040.039
\\(d_{\\vec{H}}\\)7.241.591.771.760.410.293
DC\\(CD_{L1}\\)0.270.170.150.150.070.066
\\(HD\\)3.102.222.182.221.171.103
\\(d_{\\vec{C}}\\)0.270.090.060.090.040.036
\\(d_{\\vec{H}}\\)3.102.612.762.760.530.539
Gargoyle\\(CD_{L1}\\)0.260.160.170.170.070.063
\\(HD\\)6.803.524.644.401.491.129
\\(d_{\\vec{C}}\\)0.260.060.080.110.050.045
\\(d_{\\vec{H}}\\)6.800.810.910.960.780.700
Lord Quas\\(CD_{L1}\\)0.200.120.170.120.050.047
\\(HD\\)4.611.170.821.060.980.569
\\(d_{\\vec{C}}\\)0.200.070.120.070.040.031
\\(d_{\\vec{H}}\\)4.610.980.760.640.510.370
", + "bbox": [ + 504, + 106, + 890, + 297 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/527a717b9a7266c9b127f58b23afd7748d9bcdc8898f8e3569493b8f91096133.jpg", + "table_caption": [ + "Table 6. Reconstruction accuracy under SRB." + ], + "table_footnote": [], + "table_body": "
NDF [14]CAP-UDF [102]Ours(UDF)
CDL10.2380.0730.070
F-Score68.684.585.1
", + "bbox": [ + 560, + 316, + 831, + 352 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6998a1e2a06a36e1005cdeb696ed9fdb3464e3220e086208a519e37dda5cdf85.jpg", + "image_caption": [ + "Table 7. Accuracy of reconstruction with UDF under SRB.", + "Figure 7. Visual comparison under Thingi10k and D-FAUST." + ], + "image_footnote": [], + "bbox": [ + 501, + 371, + 887, + 630 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f2427a3e6633596af31a752aaee7cea901b85d38db8749942df94a42506bb9fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IGR [24]SAP [65]Ours
Thing10kCDL10.4400.0540.051
F-Score0.5050.9400.948
NC0.6920.9470.965
DFAUSTCDL10.2350.0430.015
F-Score0.8050.9660.975
NC0.9110.9590.978
", + "bbox": [ + 560, + 652, + 831, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 8. Reconstruction accuracy under Thingi10k and D-FAUST in terms of $CD_{L1}$ and F-Score with a threshold of 0.01.", + "bbox": [ + 500, + 734, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "human bodies and more detailed expressions on faces.", + "bbox": [ + 500, + 777, + 857, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Surface Reconstruction for Scenes", + "text_level": 1, + "bbox": [ + 500, + 801, + 797, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3DScene. We follow PCP [51] to report $CD_{L1}$ , $CD_{L2}$ and Normal Consistency(NC) for evaluation. We report the comparisons with the latest methods in Tab. 9. We outperform both kinds of methods with learned priors such as ConvOcc [73] and PCP [51] and overfitting based NP [49]", + "bbox": [ + 500, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18327", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/99d914b16119c2ec4cd76a1aa3e4c74e4f2db9ca5bbcc69f30384563f8658bb7.jpg", + "image_caption": [ + "ConvOcc" + ], + "image_footnote": [], + "bbox": [ + 81, + 79, + 243, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0b03ddc233ebc125ac799649dd24c3a48b0328a941ac603036c200377755e8b.jpg", + "image_caption": [ + "NP" + ], + "image_footnote": [], + "bbox": [ + 246, + 80, + 406, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/726a1d1207ba43fc36d35ceb2ba62a5618f64596babfaea182e970ba9ea5599b.jpg", + "image_caption": [ + "PCP" + ], + "image_footnote": [], + "bbox": [ + 408, + 80, + 566, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e15184067e4c73650994685d61f50203310af44e75d4e2cb6b0f48bd2193374a.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 566, + 80, + 722, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7576ebcaebb72a25bfc8743731bb47ab16c91e14a0c4f2a0a07a8eb8e2103048.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 723, + 80, + 888, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b5004b4b02b3108d2f182b8f5c2c0bed111468b24e76b526b3cd624b244a534.jpg", + "image_caption": [ + "NP PCP Ours Figure 9. Visual comparison under SceneNet dataset." + ], + "image_footnote": [], + "bbox": [ + 80, + 205, + 207, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/692d5de9e25cfe3f42a3d65d054294fe01c8cd70f846d58003ccdceed3682c3e.jpg", + "image_caption": [ + "Figure 8. Visual comparison under 3DScene dataset." + ], + "image_footnote": [], + "bbox": [ + 209, + 205, + 338, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9fd1564499f529be4c258b65137ff4e5878e41de19f8401ed56c4ad6ec1b688.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 205, + 468, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in all scenes. The visual comparisons 8 shows that our method can work well with real scans on scenes and reveal more geometry details on surfaces.", + "bbox": [ + 76, + 363, + 468, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e2a939aabec28703e4f3de1ce1269425db87d5b0e10f8bf471262a20e0083130.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConvOcc [73]NP [49]PCP[51]Ours
BurghersCDL2× 10026.691.760.2670.246
CDL10.0770.0100.0080.008
NC0.8650.8830.9140.926
LoungeCDL2× 1008.6839.710.0610.055
CDL10.0420.0590.0060.005
NC0.8570.8570.9280.922
CopyroomCDL2× 10010.990.510.0760.069
CDL10.0450.0110.0070.006
NC0.8480.8840.9180.929
StonewallCDL2× 10019.120.0630.0610.058
CDL10.0660.0070.00650.006
NC0.8660.8680.8880.893
TotepoleCDL2× 1001.160.190.100.093
CDL10.0160.0100.0080.007
NC0.9250.7650.7840.847
", + "bbox": [ + 81, + 412, + 468, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SceneNet. Following the experimental setting in PCP [51], we report our $CD_{L1}$ , NC, and F-score in Tab. 10. The comparison indicates that we outperform these methods by producing more accurate and smooth surfaces. This is also verified by our visual comparison in Fig. 9, where the latest methods reveal either no or inaccurate indoor geometry.", + "bbox": [ + 75, + 627, + 468, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D-FRONT. We follow the experimental setting in NeuralPoisson [16] to report $CD_{L1}$ , IoU and L2 distance over the voxelization of meshes. IoU and L2 distances are calculated as the error between the reconstructed surfaces and the ground truth over the voxelization at a resolution of 128. The comparison in Tab. 11 shows that our results are more accurate than the latest method. Our visual comparison with SIREN in Fig. 10 shows that GridPull can infer an accurate zero level set which leads to very sharp corners.", + "bbox": [ + 75, + 718, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Matterport. We follow the experimental setting in NGS [30], and report the comparison in terms of $CD_{L1}$ , NC, and F-score in Tab. 12. We achieve the best perfor", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/32d9bf6b2ae00dc36dae035779ce9886e9cbb805a338c265cfe2a1fac059fd7e.jpg", + "table_caption": [ + "Table 9. Reconstruction accuracy under 3DScene." + ], + "table_footnote": [], + "table_body": "
NP [49]PCP [51]Ours
LivingroomCDL10.0880.0170.015
NC0.8810.9330.922
FScore0.8010.9660.953
BathroomCDL10.0360.0160.013
NC0.9120.9450.952
FScore0.8600.9770.974
BedroomCDL10.0340.0140.014
NC0.9050.9480.951
FScore0.8760.9800.986
KitchenCDL10.0490.0150.014
NC0.9000.9450.952
FScore0.8250.9760.984
OfficeCDL10.0620.0240.022
NC0.8790.9190.931
FScore0.7299250.940
MeanCDL10.0540.0170.0156
NC0.8950.9380.9416
FScore0.8189650.9674
", + "bbox": [ + 540, + 205, + 852, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b0fda8a33aa1e8759e1975383b9e53745aafc09b759d40b8e63eddc7d4f7bff.jpg", + "image_caption": [ + "Figure 10. Visual comparison under 3Dfront dataset." + ], + "image_footnote": [], + "bbox": [ + 500, + 474, + 890, + 655 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7edd1a49f48a1b7e3c7b1168f31f95374247e966a974c32831028176a1e02928.jpg", + "table_caption": [ + "Table 10. Reconstruction accuracy under SceneNet. FScore is with a threshold of 0.025." + ], + "table_footnote": [], + "table_body": "
MethodCDIOUL2
SIREN [71]0.01830.5242.373
NeuralPoisson [16]0.01610.5452.073
Ours0.01410.5621.984
", + "bbox": [ + 578, + 683, + 812, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 11. Reconstruction accuracy under 3DFRONT.", + "bbox": [ + 537, + 734, + 852, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance in terms of accuracy. NGS [30] has a pretty fast inference but it costs about four to five days to learn priors from the training set in Matterport. Visual comparison in Fig. 11 shows that we can reconstruct more accurate surfaces while NGS is poorly generalized to unseen scenes.", + "bbox": [ + 496, + 763, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KITTI. We evaluate GridPull on a large scale real scan from KITTI odometry (Sequence00, frame 3000 to 4000) which contains about 13.8 million points. The comparison in Fig. 12 shows that we can reconstruct a more accurate", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18328", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47e2860efde063c793afb59182941a65fc86b588c467ed0d4cd65644d4d94b02.jpg", + "image_caption": [ + "Figure 11. Visual comparison under Matterport3D dataset." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 893, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4c7c19071dc506bff46b81312b521853ec3323365f932e776ef44a583d0104c.jpg", + "image_caption": [ + "Figure 12. Visual comparison under Kitti dataset." + ], + "image_footnote": [], + "bbox": [ + 78, + 353, + 472, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "surface in much less time (3810s) than the overfitting based method SAP (13005s) [65].", + "bbox": [ + 76, + 640, + 468, + 671 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9cdfcaabdb9eeff4f8c266b35ae2266682e3f887a47b80e4c3ab3ca4f27c3411.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCDL1× 1000F-Score(0.002)NCTime
ConvOcc [73]6.2188.992.034s
SAP [65]4.1791.291.4643s
NGS [30]3.0497.395.72.1s
Ours2.9397.796.3223s
", + "bbox": [ + 98, + 681, + 447, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 770, + 240, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies on FAMOUS [17] to justify the effectiveness of modules in GridPull.", + "bbox": [ + 76, + 794, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss. We justify the effectiveness of our loss in terms of $CD_{L2}$ in Tab. 13. For the pulling loss, it is able to estimate a coarse shape but with discontinuous meshes and artifacts. The gradient loss can improve the continuousness on the surface. Using the TV loss, we obtained a more continu-", + "bbox": [ + 76, + 824, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c04bd206759f675ef696837d34dcb1a2739ae0ae0dbe511b19630e0eb00f766a.jpg", + "image_caption": [ + "Figure 13. Effect of Resolution $R$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 348, + 890, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ous field, which removes most of artifacts and significantly improve the accuracy. The surface loss encourages a more accurate zero level-set, which also improves the accuracy.", + "bbox": [ + 498, + 448, + 890, + 493 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7e22fc53a056911fbe481cf4e1f2b9b16c3035961a8fc68c8ddcd085e9aefee2.jpg", + "table_caption": [ + "Table 12. Reconstruction accuracy under Matterport3D." + ], + "table_footnote": [], + "table_body": "
LossCDL2×100
LPull0.184
LPull + L∇0.126
LPull + L∇ + LTV0.056
LPull + L∇ + LTV + LSurface0.040
", + "bbox": [ + 560, + 493, + 831, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Resolution $R$ . We compare the effect of resolution $R$ on the reconstructed surfaces. We try several candidates $\\{32, 64, 128, 256, 512\\}$ to learn the discrete distance field. We use the same bandwidth for the calculation of the pulling loss and the TV loss, and reconstruct meshes by running the marching cubes at the same resolution. The numerical and visual comparisons show that higher resolutions can infer more geometry details but also cost more inference time.", + "bbox": [ + 498, + 571, + 890, + 691 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f01a7b88562beb560b675b7b2fb150615dfe0d261fcc56b1735a6017ceb2ae5.jpg", + "table_caption": [ + "Table 13. Effect of losses." + ], + "table_footnote": [], + "table_body": "
3264128256512
CDL2×1000.1860.0970.0480.0400.035
Time50s67s84s201s645s
", + "bbox": [ + 540, + 700, + 849, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 14. Effect of Resolution $R$ .", + "bbox": [ + 594, + 742, + 795, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Point Number. We compare our performance on d-different numbers of points. We use the same set of shapes but with different numbers of points including $\\{10K, 100K, 1000K\\}$ . The numerical comparison in Tab. 15 shows that our method can estimate a discrete distance field from different numbers of points well, which leads to more accurate reconstructed surfaces than the latest overfitting based methods. In addition, our time complexity is not affected by the number of points a lot, while the latest", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18329", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods require more time to converge on more points.", + "bbox": [ + 76, + 90, + 442, + 106 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1fc64d14df98c96b9c085a891cb12cca77f9b1112857e5e77c9a6a071a19970c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 15. Effect of point numbers." + ], + "table_body": "
NumberMethodCDL2×100TimeMemory
10KNP0.351548s4.5G
SAP0.275147s3.5G
Ours0.12282s3.3G
100KNP0.224615s4.6G
SAP0.076354s3.6G
Ours0.04084s3.3G
1000KNP0.1751304s4.8G
SAP0.063841s3.7G
Ours0.03990s3.4G
", + "bbox": [ + 117, + 119, + 426, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Initialization. We highlight our geometric initialization of distances on grid vertices by comparing it with random initialization. The comparison in Tab. 16 shows that random initialization makes the optimization converge hard, and cannot work well with the TV loss, especially with a bandwidth, which produces lots of artifacts in empty spaces. Moreover, we try an Eikonal constraint on gradients based on the geometric initialization, but did not get improvement.", + "bbox": [ + 75, + 267, + 468, + 387 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2da4d5018056408303595ee688c28f893483270407d0395bbd786b2319a8ae22.jpg", + "table_caption": [], + "table_footnote": [ + "Table 16. Effect of Initialization." + ], + "table_body": "
RandomGeometric||∇|| = 1
CDL2 × 1000.1550.0400.054
", + "bbox": [ + 117, + 396, + 426, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bandwidth. We report the effect of bandwidth $M_{1}$ and $M_{2}$ in the pulling loss and the TV loss in Tab. 17. For both losses, the bandwidth controls the scope of query sampling and continuous constraints. A larger bandwidth would cover more parameters which require more time and make the optimization converge slowly. For instance, if we sample queries and impose the pulling in all $R^3$ grids, the speed is even slower than NeuralPull which uses neural network for the pulling, since neural networks can generalize signed distances nearby and do not need to infer in all grids. Although a small bandwidth would save time, a too small bandwidth would cover too few grids around the surface, which leads to discontinuousness and artifacts. While if the bandwidth is large enough, we may not see surface improvement but just cost more time. Moreover, even with the same loss, NeuralPull does not produce better results than ours. Comparing with methods using neural networks, our discrete S-DF can have faster convergence, as indicated by the angular error of gradients in Fig. 14.", + "bbox": [ + 75, + 450, + 468, + 738 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/03a6f9da4462539730a351275d21e2199af3192c3911f2f85acea774872c26a3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 17. Effect of Bandwidth." + ], + "table_body": "
BandwidthCDL2×100Time
LPull (M1=3)0.04084s
LPull (All Grids)0.043628s
LTV (M2=2)0.07983s
LTV (M2=14)0.04084s
LTV (M2=30)0.04189s
LTV (M2=50)0.04096s
LTV (All Grids)0.040101s
NeuralPull0.220552s
NeuralPull+L∇+LSurface0.203554s
", + "bbox": [ + 119, + 742, + 426, + 877 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6298761a19a7f8987f7988b7acd038e56b09aaf3d4436cf5813001e50a8dda81.jpg", + "image_caption": [ + "Figure 14. Gradients comparison with neural networks." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 890, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Noises. We report our performance on point clouds with noise. We follow the experimental setting in PCP, and report our results with middle and large level noise in Tab. 18. To handle noise, we weight more on the continuous constraint $L_{TV}$ using a larger $\\alpha = 2$ to decrease the impact of noise. The comparison shows that our method can handle noise better than the neural network based methods. Since the continuousness of neural networks make gradients near the surface over smooth to overfit all noise, this makes it hard to infer accurate zero level-set.", + "bbox": [ + 496, + 228, + 890, + 378 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f9cbe42091cccb5a1bda68d2c2f47e7ca581fee89f91e914bd242cfedbe801b3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 18. Effect of noises." + ], + "table_body": "
Noise levelNP [49]PCP [51]Ours
F-med-noise0.2800.0710.044
F-max-noise0.3100.2980.060
", + "bbox": [ + 560, + 391, + 831, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Normals. We report our performance on point clouds with normals. If GT normal is available, we could add a supervision on normals at points on the surface. What we do is to supervise the gradients $\\lambda f$ at points on the surface using their ground truth normals. The comparisons in Tab. 19 show that the normal supervision improves the performance a little. This indicates that our method can infer pretty accurate gradients on the surface or near the surface, where gradients on the surface can be highly accurate to the normal ground truth on surfaces.", + "bbox": [ + 496, + 459, + 890, + 609 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/65b5e19079f8cea75c35e8b398e385312107c35cda92a6eeb321ac8650ba9ad7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 19. Effect of normals." + ], + "table_body": "
ClassOursOurs(Normals)
CDL20.04010.0403
", + "bbox": [ + 578, + 616, + 810, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 617, + 693 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduce GridPull for the fast inference of distance fields from large scale point clouds without using neural components. We infer a distance field using learnable parameters defined on discrete grids, which we can directly optimize in a more efficient way than neural networks. Besides the complexity advantage, our loss function manages to achieve more continuous distance fields with more consistent gradients, which leads to higher accuracy in surface reconstruction. We evaluate GridPull and justify its effectiveness on benchmarks. Numerical and visual comparisons show that GridPull outperforms the latest methods in terms of accuracy and complexity, even without using priors, normal, and neural networks.", + "bbox": [ + 496, + 703, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18330", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] ShapeNet. https://www.shapenet.org/.4", + "[2] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2", + "[3] Matan Atzmon and yaron Lipman. SALD: sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021. 1, 2", + "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. CoRR, abs/2106.10811, 2021. 1, 2", + "[5] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Computer Vision and Pattern Recognition, 2017. 4", + "[6] Alexandre Boulch and Renaud Marlet. POCO: Point convolution for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2", + "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2", + "[8] Rohan Chabra, Jan Eric Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard A. Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conference on Computer Vision, volume 12374, pages 608-625, 2020. 1, 2", + "[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 4", + "[10] Chao Chen, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Unsupervised learning of fine structure generation for 3D point clouds by 2D projections matching. In IEEE International Conference on Computer Vision, 2021. 1, 2", + "[11] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Latent partition implicit with surface codes for 3d representation. In European Conference on Computer Vision, 2022. 1, 2", + "[12] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Unsupervised inference of signed distance functions from single-sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17712-17723, 2023. 1", + "[13] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[14] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. arXiv, 2010.13938, 2020. 1, 2, 6", + "[15] Gene Chou, Ilya Chugunov, and Felix Heide. GenSDF: Two-stage learning of generalizable signed distance functions. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 5" + ], + "bbox": [ + 86, + 114, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Angela Dai and Matthias Nießner. Neural Poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 7", + "[17] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J. Mitra, and Michael Wimmer. Points2Surf: Learning implicit surfaces from point clouds. In European Conference on Computer Vision, 2020. 1, 2, 4, 8", + "[18] Huan Fu, Bowen Cai, Lin Gao, Lingxiao Zhang, Cao Li, Zengqi Xun, Chengyue Sun, Yiyun Fei, Yu Zheng, Ying Li, Yi Liu, Peng Liu, Lin Ma, Le Weng, Xiaohang Hu, Xin Ma, Qian Qian, Rongfei Jia, Binqiang Zhao, and Hao Zhang. 3d-front: 3d furnished rooms with layouts and semantics. CoRR, abs/2011.09127, 2020. 4", + "[19] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. 2022. 2", + "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition, 2012. 4", + "[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In International Conference on Computer Vision, 2019. 1, 2", + "[22] Lily Goli, Daniel Rebain, Sara Sabour, Animesh Garg, and Andrea Tagliasacchi. nerf2nerf: Pairwise registration of neural radiance fields. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9354-9361. IEEE, 2023. 2", + "[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799, 2020. 1, 2", + "[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv, 2002.10099, 2020. 5, 6", + "[25] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. ShapeCaptioner: Generative caption network for 3D shapes by learning a mapping from parts detected in multiple views to sentences. ArXiv, abs/1908.00120, 2019. 2", + "[26] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. DRWR: A differentiable renderer without rendering for unsupervised 3D structure learning from silhouette images. In International Conference on Machine Learning, 2020. 1, 2", + "[27] A. Handa, V. Patraucean, V. Badrinarayanan, S. Stent, and R. Cipolla. Understanding realworld indoor scenes with synthetic data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4077-4085, 2016. 4", + "[28] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18331", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[30] Jiahui Huang, Hao-Xiang Chen, and Shi-Min Hu. A neural galerkin solver for accurate surface reconstruction. ACM Trans. Graph., 41(6), 2022. 7, 8", + "[31] Meng Jia and Matthew Kyan. Learning occupancy function from point clouds for surface reconstruction. arXiv, 2010.11378, 2020.1, 2", + "[32] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3D scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2", + "[33] Sijia Jiang, Jing Hua, and Zhizhong Han. Coordinate quantized neural implicit representations for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2", + "[34] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[35] Chuan Jin, Tieru Wu, and Junsheng Zhou. Multi-grid representation with field regularization for self-supervised surface reconstruction from point clouds. Computers & Graphics, 2023. 1", + "[36] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2", + "[37] Michael M. Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions Graphics, 32(3):29:1-29:13, 2013. 6", + "[38] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based implicit surface representation. CoRR, abs/2208.10925, 2022. 2", + "[39] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2", + "[40] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3D shapes with dynamic code clouds. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12830-12840, 2022. 1", + "[41] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[42] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. SDF-SRN: Learning signed distance 3D object reconstruction from static images. In Advances in Neural Information Processing Systems, 2020. 2", + "[43] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In European Conference on Computer vision, 2020. 2" + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Shichen Liu, Shunsuke Saito, Weikai Chen, and Hao Li. Learning to infer implicit surfaces without 3D supervision. In Advances in Neural Information Processing Systems, 2019. 2", + "[45] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Pengshuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3D reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2", + "[47] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. NeUDF: Leaning neural unsigned distance fields with volume rendering. In Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[48] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. Computer Graphics, 21(4):163-169, 1987. 1, 2", + "[49] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-Pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In International Conference on Machine Learning, 2021. 1, 2, 3, 5, 6, 7, 9", + "[50] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6305-6315, 2022. 1, 2", + "[51] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6316-6327, 2022. 2, 3, 5, 6, 7, 9", + "[52] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3d point clouds via noise to noise mapping. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1", + "[53] Baorui Ma, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Towards better gradient consistency for neural signed distance functions via level set alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17724-17734, 2023. 2", + "[54] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. A-CORN: adaptive coordinate networks for neural scene representation. CoRR, abs/2105.02788, 2021. 1, 2", + "[55] Xiaoxu Meng, Weikai Chen, and Bo Yang. NeAT: Learning neural implicit surfaces with arbitrary topologies from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 2", + "[56] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18332", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Zhenxing Mi, Yiming Luo, and Wenbing Tao. SSRNet: Scalable 3D surface reconstruction network. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2", + "[58] Mateusz Michalkiewicz, Jhony K. Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3D shape inference. CoRR, abs/1901.06802, 2019. 2", + "[59] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 1, 2", + "[60] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2", + "[61] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[62] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision, 2021. 1, 2", + "[63] Amine Ouasfi and Adnane Boukhayma. Few 'zero level set'-shot learning of shape signed distance functions in feature space. In European Conference on Computer Vision, 2022. 2", + "[64] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[65] Songyou Peng, Chiyu “Max” Jiang, Yiji Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 8", + "[66] Songyou Peng, Michael Niemeyer, Lars M. Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, volume 12348, pages 523-540, 2020. 2", + "[67] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 2, 6", + "[68] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. In International Conference on Machine Learning, 2021. 1, 2", + "[69] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[70] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa." + ], + "bbox": [ + 86, + 90, + 470, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Plenoxels: Radiance fields without neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[71] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 6, 7", + "[72] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2", + "[73] Lars Mescheder Marc Pollefeys Andreas Geiger Songyou Peng, Michael Niemeyer. Convolutional occupancy networks. In European Conference on Computer Vision, 2020. 1, 2, 6, 7, 8", + "[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CoRR, abs/2111.11215, 2021. 2", + "[75] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[76] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2", + "[77] Towaki Takikawa, Joel Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2", + "[78] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. SA-ConvONet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2", + "[79] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Annual Conference on Neural Information Processing Systems, 2021. 2", + "[80] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. PatchNets: Patch-Based Generalizable Deep Implicit 3D Shape Representations. European Conference on Computer Vision, 2020. 1, 2", + "[81] Delio Vicini, Sbastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Transactions on Graphics, 41(4):125:1-125:18, 2022. 2", + "[82] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. NeuRIS: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, 2022. 2", + "[83] Meng Wang, Yu-Shen Liu, Yue Gao, Kanle Shi, Yi Fang, and Zhizhong Han. Lp-dif: Learning local pattern-specific deep implicit function for 3d objects and scenes. In Pro" + ], + "bbox": [ + 509, + 90, + 890, + 901 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18333", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21856-21865, June 2023. 1", + "[84] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems, pages 27171-27183, 2021. 2", + "[85] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics, 41(4):1-15, 2022. 2", + "[86] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. HF-NeuS: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems, 2022. 2", + "[87] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net++: Point cloud completion by transformer-enhanced multi-step point moving paths. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):852-867, 2022. 2", + "[88] Xin Wen, Junsheng Zhou, Yu-Shen Liu, Hua Su, Zhen Dong, and Zhizhong Han. 3d shape reconstruction from 2d images with disentangled attribute flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3803-3813, 2022. 1", + "[89] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 4", + "[90] Yunjie Wu and Zhengxing Sun. DFR: differentiable function rendering for learning 3D generation from images. Computer Graphics Forum, 39(5):241-252, 2020. 2", + "[91] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In IEEE International Conference on Computer Vision, 2021. 1, 2", + "[92] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflake point deconvolution for point cloud completion and generation with skip-transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6320-6338, 2022. 2", + "[93] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems, 2021. 2", + "[94] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2", + "[95] Wang Yifan, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Iso-Points: Optimizing neural implicit surfaces with hybrid representations. CoRR, abs/2012.06434, 2020. 1, 2", + "[96] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of" + ], + "bbox": [ + 86, + 90, + 468, + 901 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural radiance fields. In IEEE International Conference on Computer Vision, 2021. 2", + "[97] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, June 2022. 5", + "[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. ArXiv, abs/2022.00665, 2022. 2", + "[99] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3D objects with differentiable rendering of sdf shape priors. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[100] Wenbin Zhao, Jiabao Lei, Yuxin Wen, Jianguo Zhang, and Kui Jia. Sign-agnostic implicit learning of surface self-similarities for shape modeling and reconstruction from raw point clouds. CoRR, abs/2012.07498, 2020. 1, 2", + "[101] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, and Zhizhong Han. Learning a more continuous zero level set in unsigned distance fields through level set projection. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2", + "[102] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. Advances in Neural Information Processing Systems, 35:16481-16494, 2022. 2, 6", + "[103] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. CoRR, abs/1605.04797, 2016. 4", + "[104] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics, 32(4):112:1-112:8, 2013. 4" + ], + "bbox": [ + 503, + 92, + 890, + 585 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "18334", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_model.json b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..84f27a71c9ec52333679c8bd37929bb236c24631 --- /dev/null +++ b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_model.json @@ -0,0 +1,3702 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.872, + 0.175 + ], + "angle": 0, + "content": "GridPull: Towards Scalability in Learning Implicit Representations from 3D Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.203, + 0.676, + 0.222 + ], + "angle": 0, + "content": "Chao Chen\\(^{1}\\) Yu-Shen Liu\\(^{1}\\)* Zhizhong Han\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.223, + 0.694, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Software, Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.24, + 0.753, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science, Wayne State University, Detroit, USA" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.26, + 0.829, + 0.275 + ], + "angle": 0, + "content": "chenchao19@mails.tsinghua.edu.cn liuyushen@tsinghua.edu.cn h312h@wayne.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.751 + ], + "angle": 0, + "content": "Learning implicit representations has been a widely used solution for surface reconstruction from 3D point clouds. The latest methods infer a distance or occupancy field by overfitting a neural network on a single point cloud. However, these methods suffer from a slow inference due to the slow convergence of neural networks and the extensive calculation of distances to surface points, which limits them to small scale points. To resolve the scalability issue in surface reconstruction, we propose GridPull to improve the efficiency of learning implicit representations from large scale point clouds. Our novelty lies in the fast inference of a discrete distance field defined on grids without using any neural components. To remedy the lack of continuousness brought by neural networks, we introduce a loss function to encourage continuous distances and consistent gradients in the field during pulling queries onto the surface in grids near to the surface. We use uniform grids for a fast grid search to localize sampled queries, and organize surface points in a tree structure to speed up the calculation of distances to the surface. We do not rely on learning priors or normal supervision during optimization, and achieve superiority over the latest methods in terms of complexity and accuracy. We evaluate our method on shape and scene benchmarks, and report numerical and visual comparisons with the latest methods to justify our effectiveness and superiority. The code is available at https://github.com/chenchao15/GridPull." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.782, + 0.21, + 0.798 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.808, + 0.47, + 0.839 + ], + "angle": 0, + "content": "It is vital to reconstruct surfaces from 3D point clouds for downstream applications. A widely used strategy is to learn" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.31, + 0.892, + 0.476 + ], + "angle": 0, + "content": "implicit representations [59, 62, 26, 10, 91, 77, 54, 68, 88] from 3D point clouds in a data-driven manner. With the learned implicit representations, we can reconstruct surfaces in meshes by running the marching cubes algorithm [48]. By learning priors from large scale datasets during training, previous methods [57, 21, 31, 46, 78, 73, 17] generalize the learned priors to either global implicit functions [89, 80, 8, 32, 6, 50] or local ones for unseen point clouds. However, the generalization ability limits their performances on large structure or geometry variations of training samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.644 + ], + "angle": 0, + "content": "More recent methods [23, 12, 2, 100, 3, 95, 4, 49, 14, 65, 35] achieved better generalization by directly overfitting neural networks on single unseen point clouds. They rely on extensive calculations of distances between queries and surface points to probe the space, which supervises neural networks to converge to a distance or occupancy field [49, 23, 2, 100, 3, 11, 52, 83, 40]. However, this inference procedure is time consuming due to the extensive distance calculations and the slow convergence of neural networks. This demerit makes these methods hard to scale up to large scale point clouds for surface reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To address the scalability challenge, we propose GridPull to speed up the learning of implicit function from large scale point clouds. GridPull does not require learned priors or point normal, and directly infers a distance field from a point cloud without using any neural components. We infer the distance field on grids near the surface, which reduces the number of grids we need to infer. Moreover, we organize surface points in a tree structure to speed up the nearest neighbor search for the calculation of distances to the surface. Specifically, we infer the discrete distance field by pulling queries onto the surface in grids of interests. Our loss function encourages continuous distances and consistent gradients in the field, which makes up the lack of continuousness brought by neural networks. We justify our effectiveness and highlight our superiority by numerical and visual comparisons with the latest methods on the widely used benchmarks. Our contributions are listed below." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.852, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*The corresponding author is Yu-Shen Liu. This work was supported by National Key R&D Program of China (2022YFC3800600), the National Natural Science Foundation of China (62272263, 62072268), and in part by Tsinghua-Kuaishou Institute of Future Media Data." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18322" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "i) We propose GridPull to reconstruct surfaces from large scale point clouds without using neural networks. GridPull speeds up the learning of implicit function, which addresses the scalability challenge in surface reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.179, + 0.468, + 0.225 + ], + "angle": 0, + "content": "ii) We introduce a loss function to directly infer a discrete distance field defined on grids but achieving continuous distances and consistent gradients in the field." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.238, + 0.468, + 0.282 + ], + "angle": 0, + "content": "iii) Our method outperforms state-of-the-art methods in surface reconstruction in terms of speed and accuracy on the widely used benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.299, + 0.218, + 0.315 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.468, + 0.475 + ], + "angle": 0, + "content": "Neural implicit representations have made a huge progress in various tasks [59, 62, 26, 10, 101, 91, 77, 54, 68, 53, 39, 36, 60, 28, 7, 41, 76]. We can use different supervision including 3D supervision [58, 64, 63, 56, 13, 92, 87], multi-view [72, 45, 34, 99, 44, 90, 61, 42, 94, 93, 19, 84, 98, 86, 81, 82, 25, 22, 29, 69, 74, 33, 55], and point clouds [89, 43, 57, 21] to learn neural implicit representations. In the following, we focus on reviewing works on surface reconstruction by learning implicit representations point clouds below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.468, + 0.612 + ], + "angle": 0, + "content": "Learning with Priors. With neural networks, one intuitive strategy is to use neural networks to learn priors from a training set and then generalize the learned priors to unseen samples. We can learn either global priors to map a shape level point cloud into a global implicit function [57, 21, 31, 46, 78, 73, 17] or local priors for local implicit functions to represent parts or patches [89, 80, 8, 32, 6, 50] which are further used to approximate a global implicit function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.688 + ], + "angle": 0, + "content": "These methods require large scale datasets to learn priors. However, the learned priors may not generalize well to unseen point clouds that have large geometric variations compared to training samples. Our method does not require priors, and falls into the following category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.809 + ], + "angle": 0, + "content": "Learning with Overfitting. To improve the generalization, we can learn implicit functions by overfitting neural networks on single point clouds. Methods using this strategy introduce novel constraints [23, 2, 100, 3, 95, 4], ways of leveraging gradients [49, 14], differentiable poisson solver [65] or specially designed priors [50, 51] to learn signed [49, 23, 2, 100, 3, 11] or unsigned distance functions [14, 102, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.899 + ], + "angle": 0, + "content": "These methods rely on neural networks to infer implicit functions without learning priors. However, the slow convergence of neural networks is a big limit for them to scale up to large scale point clouds. Our method addresses the scalability challenge by directly inferring a discrete distance field without using neural networks. Our novel loss leads" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.12 + ], + "angle": 0, + "content": "to less complexity and higher accuracy than the overfitting based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.409 + ], + "angle": 0, + "content": "Learning with Grids. Learning implicit functions with grids has been used in prior-based methods [77, 54, 66, 79, 46, 34, 85]. Using neural networks, these methods learn features at vertices of grids and further map the interpolated features into a distance or occupancy field. Some methods [70, 75, 96, 38] inferred a discrete radiance field defined on grids to speed up the learning of a radiance field, while they cared more about the quality of synthesized views than the underlying geometry. Without using neural components, some methods [67, 65] directly infer discrete implicit functions on grids. To make up the continuousness of neural networks, they employed different constraints, such as Possion equations or Viscosity priors, to pursue a continuous field from a single point cloud. However, these methods are slowly converged, which makes them hard to scale up to large scale point clouds. Although we also use grids to speed up the inference, we introduce a more efficient training strategy to decrease optimization space and more effective losses to increase inference efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.59, + 0.436 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Overview. GridPull aims to achieve a fast inference of a distance field, such as a signed distance field, from a 3D point cloud \\( S = \\{s_j | j \\in [1, J]\\} \\) without using any neural components. As shown in Fig. 1 (b), we represent the distance field by a signed distance function (SDF) \\( f \\) defined on a set of discrete grids \\( G \\) with a resolution of \\( R^3 \\), where the vertices \\( V = \\{v_i | i \\in [1, I], I = (R + 1)^3\\} \\) shared among these grids hold learnable signed distances \\( D = \\{d_i\\} \\). For an arbitrary location \\( q \\) in Fig. 1 (e), such as a randomly sampled query or a surface point, we use trilinear interpolation to approximate the signed distance \\( f(q) \\) from the eight nearest grid vertices below," + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.643, + 0.891, + 0.659 + ], + "angle": 0, + "content": "\\[\nf (q) = \\operatorname {t r i I n t e r} (q, D). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.891, + 0.71 + ], + "angle": 0, + "content": "Our goal is to conduct direct optimization in Fig. 1 (c) to infer the signed distances \\( D \\) on grid vertices by minimizing our loss function \\( L \\) on queries \\( q \\) below," + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.724, + 0.891, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\min _ {D} L (f (q)). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.891, + 0.795 + ], + "angle": 0, + "content": "With the learned function \\( f \\), we run the marching cubes algorithm [48] to reconstruct a mesh as the surface as shown in Fig. 1 (d)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Discrete Grids. We split the space occupied by the point cloud \\(S\\) into discrete grids \\(G\\) with a resolution of \\(R\\) in Fig. 1 (b). GridPull can work with high resolutions, such as \\(R = 512\\), because of our efficient and effective loss function and training strategy. Within each grid, we interpolate the signed distance \\(f(q)\\) of query \\(q\\) using the trilinear interpolation in Eq. 1. We organize surface points in a tree" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18323" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.091, + 0.173, + 0.105 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.107, + 0.286, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.09, + 0.279, + 0.105 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.09, + 0.444, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.091, + 0.493, + 0.105 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.09, + 0.709, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.091, + 0.723, + 0.105 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.107, + 0.796, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.229, + 0.246, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.229, + 0.377, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.221, + 0.557, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.335, + 0.558, + 0.347 + ], + "angle": 0, + "content": "(g) Total Variation Loss" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.221, + 0.723, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.334, + 0.689, + 0.347 + ], + "angle": 0, + "content": "(h) Surface Loss" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.229, + 0.892, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.334, + 0.857, + 0.347 + ], + "angle": 0, + "content": "(i) Gradient Loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.35, + 0.794, + 0.363 + ], + "angle": 0, + "content": "Figure 1. We proposed GridPull for fast inference of SDF defined on discrete grids using effective losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.469, + 0.453 + ], + "angle": 0, + "content": "structure to speed up the search of the nearest surface point for queries \\( q \\). Since we need to extensively conduct trilinear interpolation on signed distances for queries \\( q \\), we use uniform grids for fast nearest grid search for queries. We did not use an Octree to organize grids, due to its inefficient and inaccurate grid search for queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.469, + 0.573 + ], + "angle": 0, + "content": "Geometric Initialization. We initialize the signed distances \\( d_{i} \\) on grid vertices which represent a sphere in the field. The initialization can speed up the convergence during optimization. Our preliminary results show that a small sphere with a radius of one or two grids works well. Since enlarging a small sphere to the target shape can calculate the continuous distance loss over smaller regions than shrinking a large sphere." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.68 + ], + "angle": 0, + "content": "Signed Distance Inference. For a query \\( q \\), we infer its signed distance \\( f(q) \\) through pulling it onto its nearest surface point in the point cloud \\( S \\). As shown in Fig. 1 (e), we first obtain \\( f(q) \\) as the interpolation of the signed distances \\( d_{i} \\) on the nearest eight vertices of \\( q \\). We implement the trilinear interpolation by solving \\( \\mathbf{a} = [a_0,..,a_7] \\) in a linear regression system on the eight vertices \\( v_{i} \\) below," + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.695, + 0.469, + 0.724 + ], + "angle": 0, + "content": "\\[\nf (q) = a _ {0} + a _ {1} x + a _ {2} y + a _ {3} z + a _ {4} x y + a _ {5} x z + a _ {6} y z + a _ {7} x y z, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.468, + 0.754 + ], + "angle": 0, + "content": "where \\([x,y,z]\\) is the coordinate of \\(q\\) and the eight constraints are \\(\\{f(v_i) = d_i\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.469, + 0.815 + ], + "angle": 0, + "content": "With the predicted signed distance \\( f(q) \\) and the gradient \\( \\nabla f(q) = \\partial f / \\partial q \\), we pull a query \\( q \\) onto the surface and obtain a pulled query \\( p \\) by \\( p = q - f(q) \\times \\nabla f(q) / ||\\nabla f(q)||_2 \\) in Fig. 1 (f)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.816, + 0.468, + 0.86 + ], + "angle": 0, + "content": "Our goal is to minimize the distance between the pulled query \\( p \\) and the nearest surface point \\( NN(q) \\in S \\) of query \\( q \\) below," + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.863, + 0.469, + 0.879 + ], + "angle": 0, + "content": "\\[\nL _ {P u l l} = \\left\\| p - N N (q) \\right\\| _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We adopt this pulling loss from [49]. Instead, we do not" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.409 + ], + "angle": 0, + "content": "rely on neural network in the pulling procedure. We use the trilinear interpolation on signed distances \\( d_{i} \\) on grid vertices to obtain signed distances and gradients according to Eq. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.516 + ], + "angle": 0, + "content": "We sample queries \\( q \\) around surface points in each epoch. For each randomly sampled surface point, we sample queries using a gaussian distribution with the surface point as the center and a two-grid length variance. We build a KD-tree on surface points so that we can speed up the searching of the nearest neighbor from a large number of surface points for each query." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.593 + ], + "angle": 0, + "content": "To speed up the optimization, we only calculate the pulling loss on grids near the surface. We regard the grids containing surface points as the base and find the union of their \\(M_{1}\\) nearest neighboring grids. These grids form a \\(M_{1}\\)-bandwidth near the surface in Fig. 1 (c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Continuous Distance Fields. Although the pulling loss can infer correct signed distances at grid vertices, it fails to learn continuous signed distances across the field, especially near the surface. Fig. 2 (a) shows severe artifacts on the surface and gradients near the surface are messy and inconsistent, which are caused by the discontinuous signed distances across neighboring regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Neural network based methods [49, 51] do not have this issue, since the continuous character of neural networks makes the sudden change of signs on the same shape side hard. While our method uses signed distances \\( d_{i} \\) on grid vertices to represent a discrete field, this makes that one \\( d_{i} \\) can change separately without considering the change of neighboring \\( d_{i} \\), which leads to a poor continuousness in the field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To resolve this issue, we impose a total variation (TV) loss as a continuous constraint on our signed distances \\( d_{i} \\) on grid vertices. Our key idea is to make the change of one \\( d_{i} \\) affect the change of its neighboring \\( NN(d_{i}) \\), which smoothes the local distance field and propagates local up" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18324" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.473, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.236, + 0.414, + 0.247 + ], + "angle": 0, + "content": "Figure 2. The effect of losses on reconstruction." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.248, + 0.47, + 0.307 + ], + "angle": 0, + "content": "dates to the rest grids. Hence, as shown in a 2D case in Fig. 1 (g), we approximate one \\(d_{i}\\) to each one of its six neighbors by minimizing the difference between each two below," + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.319, + 0.47, + 0.352 + ], + "angle": 0, + "content": "\\[\nL _ {T V} = \\sqrt {\\sum_ {a = \\{x , y , z \\}} e _ {a + 1} \\left(d _ {i}\\right) ^ {2} + e _ {a - 1} \\left(d _ {i}\\right) ^ {2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.437 + ], + "angle": 0, + "content": "where \\( e_{a + 1}(d_i) \\) is the difference between \\( d_i \\) and the signed distance along the \\( a \\) axis, i.e., \\( x, y, z \\) axis. Fig. 2 (b) shows that the continuous constraints can significantly improve the signed distance field in terms of continuousness and gradients." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.438, + 0.47, + 0.497 + ], + "angle": 0, + "content": "One remaining issue is its computational complexity. Since we are targeting grids at a high resolution, such as \\( R = 512 \\), it is very time consuming to impose this constraint on all \\( d_{i} \\) across the field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.469, + 0.588 + ], + "angle": 0, + "content": "To resolve this issue, we focus on grids near the surface. Similar to imposing the pulling loss on grids in the \\(M_{1}\\)-bandwidth near the surface, we impose our continuous constraints on the grids in the \\(M_{2}\\)-bandwidth near the surface, which significantly reduces the number of learnable parameters to optimize." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.469, + 0.678 + ], + "angle": 0, + "content": "Signed Distance Supervision. For signed distances, one supervision we can directly use is the input point cloud \\( S \\). It is a discrete surface indicating the zero level-set of the signed distance function. Hence, as introduced in Fig. 1 (h), we can constrain the signed distances at the points \\( s \\) on the surface below," + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.681, + 0.469, + 0.699 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {S u r f a c e}} = \\left\\| f (s) \\right\\| _ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Although Fig. 2 (c) indicates that \\( L_{Surface} \\) slightly improves the signed distance field and leads to a little bit more continuous surface, it is helpful to determine the zero level set with other losses. For noises on \\( S \\), our continuous constraint can relieve the impact of noises on this supervision. We will show this in experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Consistent Gradients. Fig. 2 (a) indicates that the pulling loss produces gradients that are messy and non-orthogonal to the surface. This leads to inaccurate signed distance inference during the pulling procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We additionally introduce a constraint on gradients to improve the field near the surface. As shown in Fig. 1 (i), we aim to encourage the gradient at a query \\( q \\) to point to the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "same direction of the gradient at its nearest surface point \\( NN(q) \\). We use a cosine distance below to achieve more consistent gradients," + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.14, + 0.892, + 0.158 + ], + "angle": 0, + "content": "\\[\nL _ {\\nabla} = 1 - \\cos (\\nabla f (q), \\nabla f (N N (q))). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.17, + 0.892, + 0.243 + ], + "angle": 0, + "content": "\\(L_{\\nabla}\\) constrains the gradients in the field and corrects inaccurate gradients. Fig. 2 (d) shows that more consistent gradients can improve the surface to be more continuous using small planes on the surface which however the surface not smooth at all." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.245, + 0.892, + 0.275 + ], + "angle": 0, + "content": "Loss Function. We learn signed distances on grid vertices \\(D\\) by minimizing the loss function below," + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.278, + 0.892, + 0.295 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {P u l l}} + \\alpha L _ {T V} + \\beta L _ {\\text {S u r f a c e}} + \\gamma L _ {\\nabla}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.406 + ], + "angle": 0, + "content": "where \\(\\alpha\\), \\(\\beta\\), and \\(\\gamma\\) are balance weights to make each term contribute equally. Fig. 3 shows the signed distance field on a slide of \\(D\\) during optimization. Our optimization starts from an initial sphere and only optimizes variables in the bandwidth. We skip the grids that are not optimized after the initialization during the marching cubes for surface reconstruction. Please watch our video for more details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.744, + 0.439 + ], + "angle": 0, + "content": "4. Experiments and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.491 + ], + "angle": 0, + "content": "We evaluate GridPull by numerical and visual comparisons with the latest methods on synthetic and real datasets in surface reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Datasets and Metrics. We use benchmarks for shapes and scenes in evaluations. For shapes, we conduct evaluations on five datasets including a subset of ShapeNet [1], FAMOUS [17], Thingi10k [103], Surface Reconstruction Benchmark (SRB) [89] and D-FAUST [5]. For scenes, we also report comparisons on five datasets including 3DScene [104], SceneNet [27], 3DFRONT [18], Matterport [9] and KITTI [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.748 + ], + "angle": 0, + "content": "We use L2 Chamfer distance \\((CD_{L2})\\), \\(CD_{L1}\\) and Hausdorff distance(HD) to measure the error between the reconstructed surface and the ground truth. Moreover, we use normal consistency (NC) and F-score to evaluate the accuracy of normal on the reconstructed surface. We also report our time and storage complexity to highlight our advantage towards scalability. We also use the intersection over union (IoU) to measure the reconstruction for fair comparison with the latest methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Details. We use \\( R = 256 \\) to infer a distance field. We calculate the pulling loss on grids in a bandwidth with \\( M_{1} = 3 \\). We impose the continuous constraint on grids in a bandwidth with \\( M_{2} = 14 \\) for shapes and \\( M_{2} = 25 \\) for scenes, since scenes contain open surfaces with larger empty spaces. We run the marching cubes for surface reconstruction at a resolution of 256." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use the Adam optimizer with an initial learning rate of 1.0. We decrease the learning rate with a decay rate of 0.3 every 400 iterations. We use 50,000 queries in each" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18325" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.077, + 0.216, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.156, + 0.182, + 0.166 + ], + "angle": 0, + "content": "0th Iteration" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.078, + 0.351, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.156, + 0.323, + 0.166 + ], + "angle": 0, + "content": "50th Iteration" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.078, + 0.483, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.156, + 0.459, + 0.166 + ], + "angle": 0, + "content": "100th Iteration" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.078, + 0.617, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.156, + 0.592, + 0.166 + ], + "angle": 0, + "content": "150th Iteration" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.078, + 0.751, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.156, + 0.728, + 0.166 + ], + "angle": 0, + "content": "200th Iteration" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.078, + 0.887, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.156, + 0.863, + 0.166 + ], + "angle": 0, + "content": "1000th Iteration" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.166, + 0.658, + 0.178 + ], + "angle": 0, + "content": "Figure 3. Visualization of signed distances in optimization." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.18, + 0.47, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.393, + 0.433, + 0.407 + ], + "angle": 0, + "content": "Figure 4. Visual comparison under ShapetNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.418, + 0.47, + 0.464 + ], + "angle": 0, + "content": "iteration, and run 1600 iterations for each point cloud during optimization. We set the balance weights \\(\\alpha = 1\\), \\(\\beta = 1\\), and \\(\\gamma = 0.005\\) for equal contribution from the three terms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.38, + 0.489 + ], + "angle": 0, + "content": "4.1. Surface Reconstruction for Shapes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.631 + ], + "angle": 0, + "content": "ShapeNet. Following predictive context prior (PCP) [51] and NeuralPull (NP) [49], we report our results on a subset in ShapeNet in terms of \\( CD_{L2} \\) in Tab. 1, Normal Consistency (NC) in Tab. 2, and F-Score with thresholds of 0.002 and 0.004 in Tab. 3 and Tab. 4, which show our superiority over PCP and NP in numerical comparisons over all classes, even though PCP employs a local shape prior learned from a large scale dataset. The visual comparison in Fig. 4 shows that we can reveal more accurate and detailed geometry." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.634, + 0.411, + 0.776 + ], + "angle": 0, + "content": "
ClassNP [49]PCP [51]Ours
Display0.0390.00870.0082
Lamp0.0800.03800.0347
Airplane0.0080.00650.0007
Cabinet0.0260.01530.0112
Vessel0.0220.00790.0033
Table0.0600.01310.0052
Chair0.0540.01100.0043
Sofa0.0120.00860.0015
Mean0.0380.01360.0086
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.778, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Table 1. Reconstruction accuracy under ShapeNet in terms of \\(CD_{L2} \\times 100\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "FAMOUS. We follow the experimental setting in PCP [51] and NP [49] to evaluate GridPull on FAMOUS. We compare GridPull with the latest methods with priors including PCP, GenSDF, FGC and no priors including NP and IGR. GridPull outperforms these methods in terms of \\(CD_{L2}\\) and running time. Although FGC is fast, it does not reconstruct" + }, + { + "type": "table", + "bbox": [ + 0.562, + 0.186, + 0.832, + 0.328 + ], + "angle": 0, + "content": "
ClassNP [49]PCP [51]Ours
Display0.9640.97750.9847
Lamp0.9300.94500.9693
Airplane0.9470.94900.9614
Cabinet0.9300.96000.9689
Vessel0.9410.95460.9667
Table0.9080.95950.9755
Chair0.9370.95800.9733
Sofa0.9510.96800.9792
Mean0.9390.95900.9723
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.331, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Table 2. Reconstruction accuracy under ShapeNet in terms of NC." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.348, + 0.832, + 0.49 + ], + "angle": 0, + "content": "
ClassNP [49]PCP [51]Ours
Display0.9890.99390.9963
Lamp0.8910.93820.9455
Airplane0.9960.99420.9976
Cabinet0.9800.98880.9901
Vessel0.9850.99350.9956
Table0.9220.99690.9977
Chair0.9540.99700.9979
Sofa0.9680.99430.9974
Mean0.9610.98710.9896
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.493, + 0.891, + 0.52 + ], + "angle": 0, + "content": "Table 3. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.002." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.522, + 0.832, + 0.664 + ], + "angle": 0, + "content": "
ClassNP [49]PCP [51]Ours
Display0.9910.99580.9963
Lamp0.9240.94020.9538
Airplane0.9970.99720.9989
Cabinet0.9890.99390.9946
Vessel0.9900.99580.9972
Table0.9730.99850.9990
Chair0.9690.99910.9990
Sofa0.9740.99870.9992
Mean0.9760.98990.9923
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.667, + 0.891, + 0.693 + ], + "angle": 0, + "content": "Table 4. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.004." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.7, + 0.892, + 0.79 + ], + "angle": 0, + "content": "a watertight mesh but just a polygon soup with no normal, which contains almost no geometry details. Worse than that, it fails to reconstruct some point clouds in more than eight hours, hence we remove these shapes from its results. We highlight our superiority in visual comparison in Fig. 5. We can reveal details like hair and expressions more clearly." + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.803, + 0.851, + 0.891 + ], + "angle": 0, + "content": "
MethodCDL2×100Min timeMean time
IGR [24]1.650402s407s
GenSDF [15]0.668230s232s
NP [49]0.220593s605s
FGC [97]0.05510sTimeout
PCP [51]0.0443416s3534s
Ours0.040198s201s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.544, + 0.894, + 0.848, + 0.907 + ], + "angle": 0, + "content": "Table 5. Reconstruction accuracy under FAMOUS." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18326" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.468, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.307, + 0.432, + 0.32 + ], + "angle": 0, + "content": "Figure 5. Visual comparison under FAMOUS dataset." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.323, + 0.47, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.517, + 0.416, + 0.53 + ], + "angle": 0, + "content": "Figure 6. Visual comparison under SRB dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.534, + 0.47, + 0.761 + ], + "angle": 0, + "content": "SRB. We follow the experimental setting in VisCo [67] to evaluate GridPull on real scans in SRB. We report the comparison with the latest overfitting based methods in terms of \\( CD_{L1} \\) in Tab. 6. Meanwhile, we also report their one-sided distances between the reconstructed mesh and the input noisy point cloud. The numerical comparison shows that our method significantly outperforms the latest. Moreover, we also report our results of learning unsigned distances in Tab. 7. We replace our pulling loss with the pulling loss introduced for unsigned distances in CAP-UDF [102], and remove the gradient loss due to the nondifferentiable character of UDF on surface. Visual comparison in Fig. 6 shows that our method can reconstruct surfaces that are more compact than the methods learning SDF like SIREN or more smooth than the methods learning UDF like CAP-UDF." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.763, + 0.469, + 0.838 + ], + "angle": 0, + "content": "Thingi10K. We follow the experimental setting in SAP [65] to evaluate GridPull in Thingi10K. Tab. 8 indicates our superiority over the overfitting based methods in the numerical comparison. Visual comparisons in Fig. 7 show our more accurate surfaces with complex details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "D-FAUST. We also follow the experimental setting in S-AP [65] to evaluate GridPull in D-FAUST. Tab. 8 shows that we outperform the overfitting based methods with more surface details. As shown in Fig. 7, we recover more accurate" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.107, + 0.892, + 0.299 + ], + "angle": 0, + "content": "
Poisson [37]IGR [24]SIREN [71]VisCo [67]SAP [65]Ours
Anchor\\(CD_{L1}\\)0.600.220.320.210.120.093
\\(HD\\)14.894.718.193.002.381.804
\\(d_{\\vec{C}}\\)0.600.120.100.150.080.066
\\(d_{\\vec{H}}\\)14.891.322.431.070.830.460
Daratech\\(CD_{L1}\\)0.440.250.210.260.070.062
\\(HD\\)7.244.014.304.060.870.648
\\(d_{\\vec{C}}\\)0.440.080.090.140.040.039
\\(d_{\\vec{H}}\\)7.241.591.771.760.410.293
DC\\(CD_{L1}\\)0.270.170.150.150.070.066
\\(HD\\)3.102.222.182.221.171.103
\\(d_{\\vec{C}}\\)0.270.090.060.090.040.036
\\(d_{\\vec{H}}\\)3.102.612.762.760.530.539
Gargoyle\\(CD_{L1}\\)0.260.160.170.170.070.063
\\(HD\\)6.803.524.644.401.491.129
\\(d_{\\vec{C}}\\)0.260.060.080.110.050.045
\\(d_{\\vec{H}}\\)6.800.810.910.960.780.700
Lord Quas\\(CD_{L1}\\)0.200.120.170.120.050.047
\\(HD\\)4.611.170.821.060.980.569
\\(d_{\\vec{C}}\\)0.200.070.120.070.040.031
\\(d_{\\vec{H}}\\)4.610.980.760.640.510.370
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.301, + 0.831, + 0.314 + ], + "angle": 0, + "content": "Table 6. Reconstruction accuracy under SRB." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.317, + 0.832, + 0.353 + ], + "angle": 0, + "content": "
NDF [14]CAP-UDF [102]Ours(UDF)
CDL10.2380.0730.070
F-Score68.684.585.1
" + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.355, + 0.869, + 0.368 + ], + "angle": 0, + "content": "Table 7. Accuracy of reconstruction with UDF under SRB." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.372, + 0.888, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.632, + 0.864, + 0.646 + ], + "angle": 0, + "content": "Figure 7. Visual comparison under Thingi10k and D-FAUST." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.653, + 0.832, + 0.734 + ], + "angle": 0, + "content": "
IGR [24]SAP [65]Ours
Thing10kCDL10.4400.0540.051
F-Score0.5050.9400.948
NC0.6920.9470.965
DFAUSTCDL10.2350.0430.015
F-Score0.8050.9660.975
NC0.9110.9590.978
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.736, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Table 8. Reconstruction accuracy under Thingi10k and D-FAUST in terms of \\(CD_{L1}\\) and F-Score with a threshold of 0.01." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.778, + 0.858, + 0.793 + ], + "angle": 0, + "content": "human bodies and more detailed expressions on faces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.799, + 0.817 + ], + "angle": 0, + "content": "4.2. Surface Reconstruction for Scenes" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3DScene. We follow PCP [51] to report \\( CD_{L1} \\), \\( CD_{L2} \\) and Normal Consistency(NC) for evaluation. We report the comparisons with the latest methods in Tab. 9. We outperform both kinds of methods with learned priors such as ConvOcc [73] and PCP [51] and overfitting based NP [49]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18327" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.08, + 0.245, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.174, + 0.182, + 0.184 + ], + "angle": 0, + "content": "ConvOcc" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.081, + 0.408, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.174, + 0.331, + 0.184 + ], + "angle": 0, + "content": "NP" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.081, + 0.567, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.173, + 0.496, + 0.183 + ], + "angle": 0, + "content": "PCP" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.081, + 0.723, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.173, + 0.663, + 0.183 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.082, + 0.889, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.173, + 0.824, + 0.183 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.183, + 0.641, + 0.197 + ], + "angle": 0, + "content": "Figure 8. Visual comparison under 3DScene dataset." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.206, + 0.208, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.327, + 0.431, + 0.346 + ], + "angle": 0, + "content": "NP PCP Ours Figure 9. Visual comparison under SceneNet dataset." + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.206, + 0.339, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.206, + 0.469, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.364, + 0.47, + 0.41 + ], + "angle": 0, + "content": "in all scenes. The visual comparisons 8 shows that our method can work well with real scans on scenes and reveal more geometry details on surfaces." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.414, + 0.47, + 0.601 + ], + "angle": 0, + "content": "
ConvOcc [73]NP [49]PCP[51]Ours
BurghersCDL2× 10026.691.760.2670.246
CDL10.0770.0100.0080.008
NC0.8650.8830.9140.926
LoungeCDL2× 1008.6839.710.0610.055
CDL10.0420.0590.0060.005
NC0.8570.8570.9280.922
CopyroomCDL2× 10010.990.510.0760.069
CDL10.0450.0110.0070.006
NC0.8480.8840.9180.929
StonewallCDL2× 10019.120.0630.0610.058
CDL10.0660.0070.00650.006
NC0.8660.8680.8880.893
TotepoleCDL2× 1001.160.190.100.093
CDL10.0160.0100.0080.007
NC0.9250.7650.7840.847
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.606, + 0.422, + 0.62 + ], + "angle": 0, + "content": "Table 9. Reconstruction accuracy under 3DScene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.718 + ], + "angle": 0, + "content": "SceneNet. Following the experimental setting in PCP [51], we report our \\(CD_{L1}\\), NC, and F-score in Tab. 10. The comparison indicates that we outperform these methods by producing more accurate and smooth surfaces. This is also verified by our visual comparison in Fig. 9, where the latest methods reveal either no or inaccurate indoor geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.856 + ], + "angle": 0, + "content": "3D-FRONT. We follow the experimental setting in NeuralPoisson [16] to report \\(CD_{L1}\\), IoU and L2 distance over the voxelization of meshes. IoU and L2 distances are calculated as the error between the reconstructed surfaces and the ground truth over the voxelization at a resolution of 128. The comparison in Tab. 11 shows that our results are more accurate than the latest method. Our visual comparison with SIREN in Fig. 10 shows that GridPull can infer an accurate zero level set which leads to very sharp corners." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Matterport. We follow the experimental setting in NGS [30], and report the comparison in terms of \\(CD_{L1}\\), NC, and F-score in Tab. 12. We achieve the best perfor" + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.206, + 0.853, + 0.442 + ], + "angle": 0, + "content": "
NP [49]PCP [51]Ours
LivingroomCDL10.0880.0170.015
NC0.8810.9330.922
FScore0.8010.9660.953
BathroomCDL10.0360.0160.013
NC0.9120.9450.952
FScore0.8600.9770.974
BedroomCDL10.0340.0140.014
NC0.9050.9480.951
FScore0.8760.9800.986
KitchenCDL10.0490.0150.014
NC0.9000.9450.952
FScore0.8250.9760.984
OfficeCDL10.0620.0240.022
NC0.8790.9190.931
FScore0.7299250.940
MeanCDL10.0540.0170.0156
NC0.8950.9380.9416
FScore0.8189650.9674
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.446, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Table 10. Reconstruction accuracy under SceneNet. FScore is with a threshold of 0.025." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.475, + 0.892, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.658, + 0.852, + 0.672 + ], + "angle": 0, + "content": "Figure 10. Visual comparison under 3Dfront dataset." + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.684, + 0.813, + 0.732 + ], + "angle": 0, + "content": "
MethodCDIOUL2
SIREN [71]0.01830.5242.373
NeuralPoisson [16]0.01610.5452.073
Ours0.01410.5621.984
" + }, + { + "type": "table_caption", + "bbox": [ + 0.539, + 0.735, + 0.854, + 0.748 + ], + "angle": 0, + "content": "Table 11. Reconstruction accuracy under 3DFRONT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.84 + ], + "angle": 0, + "content": "mance in terms of accuracy. NGS [30] has a pretty fast inference but it costs about four to five days to learn priors from the training set in Matterport. Visual comparison in Fig. 11 shows that we can reconstruct more accurate surfaces while NGS is poorly generalized to unseen scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "KITTI. We evaluate GridPull on a large scale real scan from KITTI odometry (Sequence00, frame 3000 to 4000) which contains about 13.8 million points. The comparison in Fig. 12 shows that we can reconstruct a more accurate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18328" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.895, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.328, + 0.658, + 0.342 + ], + "angle": 0, + "content": "Figure 11. Visual comparison under Matterport3D dataset." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.354, + 0.473, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.619, + 0.42, + 0.633 + ], + "angle": 0, + "content": "Figure 12. Visual comparison under Kitti dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.641, + 0.47, + 0.672 + ], + "angle": 0, + "content": "surface in much less time (3810s) than the overfitting based method SAP (13005s) [65]." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.683, + 0.449, + 0.747 + ], + "angle": 0, + "content": "
MethodCDL1× 1000F-Score(0.002)NCTime
ConvOcc [73]6.2188.992.034s
SAP [65]4.1791.291.4643s
NGS [30]3.0497.395.72.1s
Ours2.9397.796.3223s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.749, + 0.438, + 0.763 + ], + "angle": 0, + "content": "Table 12. Reconstruction accuracy under Matterport3D." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.241, + 0.786 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.824 + ], + "angle": 0, + "content": "We conduct ablation studies on FAMOUS [17] to justify the effectiveness of modules in GridPull." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Loss. We justify the effectiveness of our loss in terms of \\( CD_{L2} \\) in Tab. 13. For the pulling loss, it is able to estimate a coarse shape but with discontinuous meshes and artifacts. The gradient loss can improve the continuousness on the surface. Using the TV loss, we obtained a more continu-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.349, + 0.891, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.428, + 0.799, + 0.441 + ], + "angle": 0, + "content": "Figure 13. Effect of Resolution \\( R \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.449, + 0.892, + 0.494 + ], + "angle": 0, + "content": "ous field, which removes most of artifacts and significantly improve the accuracy. The surface loss encourages a more accurate zero level-set, which also improves the accuracy." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.494, + 0.832, + 0.557 + ], + "angle": 0, + "content": "
LossCDL2×100
LPull0.184
LPull + L∇0.126
LPull + L∇ + LTV0.056
LPull + L∇ + LTV + LSurface0.040
" + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.559, + 0.774, + 0.571 + ], + "angle": 0, + "content": "Table 13. Effect of losses." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Resolution \\( R \\). We compare the effect of resolution \\( R \\) on the reconstructed surfaces. We try several candidates \\( \\{32, 64, 128, 256, 512\\} \\) to learn the discrete distance field. We use the same bandwidth for the calculation of the pulling loss and the TV loss, and reconstruct meshes by running the marching cubes at the same resolution. The numerical and visual comparisons show that higher resolutions can infer more geometry details but also cost more inference time." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.701, + 0.851, + 0.741 + ], + "angle": 0, + "content": "
3264128256512
CDL2×1000.1860.0970.0480.0400.035
Time50s67s84s201s645s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.596, + 0.743, + 0.796, + 0.755 + ], + "angle": 0, + "content": "Table 14. Effect of Resolution \\( R \\) ." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Point Number. We compare our performance on d-different numbers of points. We use the same set of shapes but with different numbers of points including \\(\\{10K, 100K, 1000K\\}\\). The numerical comparison in Tab. 15 shows that our method can estimate a discrete distance field from different numbers of points well, which leads to more accurate reconstructed surfaces than the latest overfitting based methods. In addition, our time complexity is not affected by the number of points a lot, while the latest" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18329" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.444, + 0.107 + ], + "angle": 0, + "content": "methods require more time to converge on more points." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.12, + 0.428, + 0.244 + ], + "angle": 0, + "content": "
NumberMethodCDL2×100TimeMemory
10KNP0.351548s4.5G
SAP0.275147s3.5G
Ours0.12282s3.3G
100KNP0.224615s4.6G
SAP0.076354s3.6G
Ours0.04084s3.3G
1000KNP0.1751304s4.8G
SAP0.063841s3.7G
Ours0.03990s3.4G
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.17, + 0.247, + 0.375, + 0.261 + ], + "angle": 0, + "content": "Table 15. Effect of point numbers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Initialization. We highlight our geometric initialization of distances on grid vertices by comparing it with random initialization. The comparison in Tab. 16 shows that random initialization makes the optimization converge hard, and cannot work well with the TV loss, especially with a bandwidth, which produces lots of artifacts in empty spaces. Moreover, we try an Eikonal constraint on gradients based on the geometric initialization, but did not get improvement." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.397, + 0.428, + 0.428 + ], + "angle": 0, + "content": "
RandomGeometric||∇|| = 1
CDL2 × 1000.1550.0400.054
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.175, + 0.43, + 0.371, + 0.442 + ], + "angle": 0, + "content": "Table 16. Effect of Initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.47, + 0.739 + ], + "angle": 0, + "content": "Bandwidth. We report the effect of bandwidth \\( M_{1} \\) and \\( M_{2} \\) in the pulling loss and the TV loss in Tab. 17. For both losses, the bandwidth controls the scope of query sampling and continuous constraints. A larger bandwidth would cover more parameters which require more time and make the optimization converge slowly. For instance, if we sample queries and impose the pulling in all \\( R^3 \\) grids, the speed is even slower than NeuralPull which uses neural network for the pulling, since neural networks can generalize signed distances nearby and do not need to infer in all grids. Although a small bandwidth would save time, a too small bandwidth would cover too few grids around the surface, which leads to discontinuousness and artifacts. While if the bandwidth is large enough, we may not see surface improvement but just cost more time. Moreover, even with the same loss, NeuralPull does not produce better results than ours. Comparing with methods using neural networks, our discrete S-DF can have faster convergence, as indicated by the angular error of gradients in Fig. 14." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.743, + 0.428, + 0.878 + ], + "angle": 0, + "content": "
BandwidthCDL2×100Time
LPull (M1=3)0.04084s
LPull (All Grids)0.043628s
LTV (M2=2)0.07983s
LTV (M2=14)0.04084s
LTV (M2=30)0.04189s
LTV (M2=50)0.04096s
LTV (All Grids)0.040101s
NeuralPull0.220552s
NeuralPull+L∇+LSurface0.203554s
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.18, + 0.88, + 0.365, + 0.893 + ], + "angle": 0, + "content": "Table 17. Effect of Bandwidth." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.891, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.206, + 0.86, + 0.22 + ], + "angle": 0, + "content": "Figure 14. Gradients comparison with neural networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Noises. We report our performance on point clouds with noise. We follow the experimental setting in PCP, and report our results with middle and large level noise in Tab. 18. To handle noise, we weight more on the continuous constraint \\( L_{TV} \\) using a larger \\( \\alpha = 2 \\) to decrease the impact of noise. The comparison shows that our method can handle noise better than the neural network based methods. Since the continuousness of neural networks make gradients near the surface over smooth to overfit all noise, this makes it hard to infer accurate zero level-set." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.392, + 0.832, + 0.435 + ], + "angle": 0, + "content": "
Noise levelNP [49]PCP [51]Ours
F-med-noise0.2800.0710.044
F-max-noise0.3100.2980.060
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.617, + 0.438, + 0.774, + 0.451 + ], + "angle": 0, + "content": "Table 18. Effect of noises." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Normals. We report our performance on point clouds with normals. If GT normal is available, we could add a supervision on normals at points on the surface. What we do is to supervise the gradients \\(\\lambda f\\) at points on the surface using their ground truth normals. The comparisons in Tab. 19 show that the normal supervision improves the performance a little. This indicates that our method can infer pretty accurate gradients on the surface or near the surface, where gradients on the surface can be highly accurate to the normal ground truth on surfaces." + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.617, + 0.812, + 0.649 + ], + "angle": 0, + "content": "
ClassOursOurs(Normals)
CDL20.04010.0403
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.612, + 0.651, + 0.779, + 0.664 + ], + "angle": 0, + "content": "Table 19. Effect of normals." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.619, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We introduce GridPull for the fast inference of distance fields from large scale point clouds without using neural components. We infer a distance field using learnable parameters defined on discrete grids, which we can directly optimize in a more efficient way than neural networks. Besides the complexity advantage, our loss function manages to achieve more continuous distance fields with more consistent gradients, which leads to higher accuracy in surface reconstruction. We evaluate GridPull and justify its effectiveness on benchmarks. Numerical and visual comparisons show that GridPull outperforms the latest methods in terms of accuracy and complexity, even without using priors, normal, and neural networks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18330" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.424, + 0.129 + ], + "angle": 0, + "content": "[1] ShapeNet. https://www.shapenet.org/.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.131, + 0.469, + 0.171 + ], + "angle": 0, + "content": "[2] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.173, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[3] Matan Atzmon and yaron Lipman. SALD: sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.215, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. CoRR, abs/2106.10811, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.271, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[5] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Computer Vision and Pattern Recognition, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.327, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[6] Alexandre Boulch and Renaud Marlet. POCO: Point convolution for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.383, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.425, + 0.469, + 0.506 + ], + "angle": 0, + "content": "[8] Rohan Chabra, Jan Eric Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard A. Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conference on Computer Vision, volume 12374, pages 608-625, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.509, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.58, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[10] Chao Chen, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Unsupervised learning of fine structure generation for 3D point clouds by 2D projections matching. In IEEE International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.649, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[11] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Latent partition implicit with surface codes for 3d representation. In European Conference on Computer Vision, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.692, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[12] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Unsupervised inference of signed distance functions from single-sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17712-17723, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.762, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[13] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[14] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. arXiv, 2010.13938, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[15] Gene Chou, Ilya Chugunov, and Felix Heide. GenSDF: Two-stage learning of generalizable signed distance functions. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[16] Angela Dai and Matthias Nießner. Neural Poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[17] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J. Mitra, and Michael Wimmer. Points2Surf: Learning implicit surfaces from point clouds. In European Conference on Computer Vision, 2020. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.196, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[18] Huan Fu, Bowen Cai, Lin Gao, Lingxiao Zhang, Cao Li, Zengqi Xun, Chengyue Sun, Yiyun Fei, Yu Zheng, Ying Li, Yi Liu, Peng Liu, Lin Ma, Le Weng, Xiaohang Hu, Xin Ma, Qian Qian, Rongfei Jia, Binqiang Zhao, and Hao Zhang. 3d-front: 3d furnished rooms with layouts and semantics. CoRR, abs/2011.09127, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[19] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.326, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.383, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In International Conference on Computer Vision, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[22] Lily Goli, Daniel Rebain, Sara Sabour, Animesh Garg, and Andrea Tagliasacchi. nerf2nerf: Pairwise registration of neural radiance fields. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9354-9361. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.514, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.585, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv, 2002.10099, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.63, + 0.892, + 0.697 + ], + "angle": 0, + "content": "[25] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. ShapeCaptioner: Generative caption network for 3D shapes by learning a mapping from parts detected in multiple views to sentences. ArXiv, abs/1908.00120, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.702, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[26] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. DRWR: A differentiable renderer without rendering for unsupervised 3D structure learning from silhouette images. In International Conference on Machine Learning, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[27] A. Handa, V. Patraucean, V. Badrinarayanan, S. Stent, and R. Cipolla. Understanding realworld indoor scenes with synthetic data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4077-4085, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[28] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18331" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[30] Jiahui Huang, Hao-Xiang Chen, and Shi-Min Hu. A neural galerkin solver for accurate surface reconstruction. ACM Trans. Graph., 41(6), 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.191, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[31] Meng Jia and Matthew Kyan. Learning occupancy function from point clouds for surface reconstruction. arXiv, 2010.11378, 2020.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[32] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3D scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[33] Sijia Jiang, Jing Hua, and Zhizhong Han. Coordinate quantized neural implicit representations for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[34] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.414, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[35] Chuan Jin, Tieru Wu, and Junsheng Zhou. Multi-grid representation with field regularization for self-supervised surface reconstruction from point clouds. Computers & Graphics, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.47, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[36] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[37] Michael M. Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions Graphics, 32(3):29:1-29:13, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.553, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[38] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based implicit surface representation. CoRR, abs/2208.10925, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[39] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.651, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[40] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3D shapes with dynamic code clouds. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12830-12840, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[41] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[42] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. SDF-SRN: Learning signed distance 3D object reconstruction from static images. In Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[43] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In European Conference on Computer vision, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[44] Shichen Liu, Shunsuke Saito, Weikai Chen, and Hao Li. Learning to infer implicit surfaces without 3D supervision. In Advances in Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[45] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.219, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Pengshuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3D reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[47] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. NeUDF: Leaning neural unsigned distance fields with volume rendering. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[48] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. Computer Graphics, 21(4):163-169, 1987. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.386, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[49] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-Pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In International Conference on Machine Learning, 2021. 1, 2, 3, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[50] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6305-6315, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[51] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6316-6327, 2022. 2, 3, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.58, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[52] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3d point clouds via noise to noise mapping. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.636, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[53] Baorui Ma, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Towards better gradient consistency for neural signed distance functions via level set alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17724-17734, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.706, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[54] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. A-CORN: adaptive coordinate networks for neural scene representation. CoRR, abs/2105.02788, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[55] Xiaoxu Meng, Weikai Chen, and Bo Yang. NeAT: Learning neural implicit surfaces with arbitrary topologies from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[56] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18332" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[57] Zhenxing Mi, Yiming Luo, and Wenbing Tao. SSRNet: Scalable 3D surface reconstruction network. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[58] Mateusz Michalkiewicz, Jhony K. Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3D shape inference. CoRR, abs/1901.06802, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[59] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.277, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[60] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[61] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.471, + 0.461 + ], + "angle": 0, + "content": "[62] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.462, + 0.471, + 0.516 + ], + "angle": 0, + "content": "[63] Amine Ouasfi and Adnane Boukhayma. Few 'zero level set'-shot learning of shape signed distance functions in feature space. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.471, + 0.587 + ], + "angle": 0, + "content": "[64] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.471, + 0.644 + ], + "angle": 0, + "content": "[65] Songyou Peng, Chiyu “Max” Jiang, Yiji Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.646, + 0.471, + 0.701 + ], + "angle": 0, + "content": "[66] Songyou Peng, Michael Niemeyer, Lars M. Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, volume 12348, pages 523-540, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.703, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[67] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.76, + 0.471, + 0.815 + ], + "angle": 0, + "content": "[68] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. In International Conference on Machine Learning, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[69] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[70] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Plenoxels: Radiance fields without neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[71] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.206, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[72] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.264, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[73] Lars Mescheder Marc Pollefeys Andreas Geiger Songyou Peng, Michael Niemeyer. Convolutional occupancy networks. In European Conference on Computer Vision, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.321, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CoRR, abs/2111.11215, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.363, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[75] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.42, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[76] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.463, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[77] Towaki Takikawa, Joel Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.548, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[78] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. SA-ConvONet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[79] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Annual Conference on Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.676, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[80] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. PatchNets: Patch-Based Generalizable Deep Implicit 3D Shape Representations. European Conference on Computer Vision, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.746, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[81] Delio Vicini, Sbastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Transactions on Graphics, 41(4):125:1-125:18, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[82] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. NeuRIS: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[83] Meng Wang, Yu-Shen Liu, Yue Gao, Kanle Shi, Yi Fang, and Zhizhong Han. Lp-dif: Learning local pattern-specific deep implicit function for 3d objects and scenes. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18333" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21856-21865, June 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[84] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems, pages 27171-27183, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.207, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[85] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics, 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.264, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[86] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. HF-NeuS: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.321, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[87] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net++: Point cloud completion by transformer-enhanced multi-step point moving paths. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):852-867, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.391, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[88] Xin Wen, Junsheng Zhou, Yu-Shen Liu, Hua Su, Zhen Dong, and Zhizhong Han. 3d shape reconstruction from 2d images with disentangled attribute flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3803-3813, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.462, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[89] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.519, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[90] Yunjie Wu and Zhengxing Sun. DFR: differentiable function rendering for learning 3D generation from images. Computer Graphics Forum, 39(5):241-252, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.562, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[91] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In IEEE International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.633, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[92] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflake point deconvolution for point cloud completion and generation with skip-transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6320-6338, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.703, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[93] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.746, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[94] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[95] Wang Yifan, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Iso-Points: Optimizing neural implicit surfaces with hybrid representations. CoRR, abs/2012.06434, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[96] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "neural radiance fields. In IEEE International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[97] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, June 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. ArXiv, abs/2022.00665, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[99] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3D objects with differentiable rendering of sdf shape priors. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[100] Wenbin Zhao, Jiabao Lei, Yuxin Wen, Jianguo Zhang, and Kui Jia. Sign-agnostic implicit learning of surface self-similarities for shape modeling and reconstruction from raw point clouds. CoRR, abs/2012.07498, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[101] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, and Zhizhong Han. Learning a more continuous zero level set in unsigned distance fields through level set projection. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[102] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. Advances in Neural Information Processing Systems, 35:16481-16494, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.503, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[103] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. CoRR, abs/1605.04797, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[104] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics, 32(4):112:1-112:8, 2013. 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18334" + } + ] +] \ No newline at end of file diff --git a/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_origin.pdf b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6550d3e8820a473e5dc64101748a588ff999686e --- /dev/null +++ b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/694f653d-a2d1-4ef0-9ac5-0fdfb6b8b37d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f66025f269ceedea301c2b2de8f7b8b9ce94a808d998afea69785e2a8b2ddef +size 9742001 diff --git a/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/full.md b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..419bef92341bf5549aac5afba6b09162095fe5d6 --- /dev/null +++ b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/full.md @@ -0,0 +1,498 @@ +# GridPull: Towards Scalability in Learning Implicit Representations from 3D Point Clouds + +Chao Chen $^{1}$ Yu-Shen Liu $^{1}$ * Zhizhong Han $^{2}$ + +$^{1}$ School of Software, Tsinghua University, Beijing, China + +$^{2}$ Department of Computer Science, Wayne State University, Detroit, USA + +chenchao19@mails.tsinghua.edu.cn liuyushen@tsinghua.edu.cn h312h@wayne.edu + +# Abstract + +Learning implicit representations has been a widely used solution for surface reconstruction from 3D point clouds. The latest methods infer a distance or occupancy field by overfitting a neural network on a single point cloud. However, these methods suffer from a slow inference due to the slow convergence of neural networks and the extensive calculation of distances to surface points, which limits them to small scale points. To resolve the scalability issue in surface reconstruction, we propose GridPull to improve the efficiency of learning implicit representations from large scale point clouds. Our novelty lies in the fast inference of a discrete distance field defined on grids without using any neural components. To remedy the lack of continuousness brought by neural networks, we introduce a loss function to encourage continuous distances and consistent gradients in the field during pulling queries onto the surface in grids near to the surface. We use uniform grids for a fast grid search to localize sampled queries, and organize surface points in a tree structure to speed up the calculation of distances to the surface. We do not rely on learning priors or normal supervision during optimization, and achieve superiority over the latest methods in terms of complexity and accuracy. We evaluate our method on shape and scene benchmarks, and report numerical and visual comparisons with the latest methods to justify our effectiveness and superiority. The code is available at https://github.com/chenchao15/GridPull. + +# 1. Introduction + +It is vital to reconstruct surfaces from 3D point clouds for downstream applications. A widely used strategy is to learn + +implicit representations [59, 62, 26, 10, 91, 77, 54, 68, 88] from 3D point clouds in a data-driven manner. With the learned implicit representations, we can reconstruct surfaces in meshes by running the marching cubes algorithm [48]. By learning priors from large scale datasets during training, previous methods [57, 21, 31, 46, 78, 73, 17] generalize the learned priors to either global implicit functions [89, 80, 8, 32, 6, 50] or local ones for unseen point clouds. However, the generalization ability limits their performances on large structure or geometry variations of training samples. + +More recent methods [23, 12, 2, 100, 3, 95, 4, 49, 14, 65, 35] achieved better generalization by directly overfitting neural networks on single unseen point clouds. They rely on extensive calculations of distances between queries and surface points to probe the space, which supervises neural networks to converge to a distance or occupancy field [49, 23, 2, 100, 3, 11, 52, 83, 40]. However, this inference procedure is time consuming due to the extensive distance calculations and the slow convergence of neural networks. This demerit makes these methods hard to scale up to large scale point clouds for surface reconstruction. + +To address the scalability challenge, we propose GridPull to speed up the learning of implicit function from large scale point clouds. GridPull does not require learned priors or point normal, and directly infers a distance field from a point cloud without using any neural components. We infer the distance field on grids near the surface, which reduces the number of grids we need to infer. Moreover, we organize surface points in a tree structure to speed up the nearest neighbor search for the calculation of distances to the surface. Specifically, we infer the discrete distance field by pulling queries onto the surface in grids of interests. Our loss function encourages continuous distances and consistent gradients in the field, which makes up the lack of continuousness brought by neural networks. We justify our effectiveness and highlight our superiority by numerical and visual comparisons with the latest methods on the widely used benchmarks. Our contributions are listed below. + +i) We propose GridPull to reconstruct surfaces from large scale point clouds without using neural networks. GridPull speeds up the learning of implicit function, which addresses the scalability challenge in surface reconstruction. +ii) We introduce a loss function to directly infer a discrete distance field defined on grids but achieving continuous distances and consistent gradients in the field. +iii) Our method outperforms state-of-the-art methods in surface reconstruction in terms of speed and accuracy on the widely used benchmarks. + +# 2. Related Work + +Neural implicit representations have made a huge progress in various tasks [59, 62, 26, 10, 101, 91, 77, 54, 68, 53, 39, 36, 60, 28, 7, 41, 76]. We can use different supervision including 3D supervision [58, 64, 63, 56, 13, 92, 87], multi-view [72, 45, 34, 99, 44, 90, 61, 42, 94, 93, 19, 84, 98, 86, 81, 82, 25, 22, 29, 69, 74, 33, 55], and point clouds [89, 43, 57, 21] to learn neural implicit representations. In the following, we focus on reviewing works on surface reconstruction by learning implicit representations point clouds below. + +Learning with Priors. With neural networks, one intuitive strategy is to use neural networks to learn priors from a training set and then generalize the learned priors to unseen samples. We can learn either global priors to map a shape level point cloud into a global implicit function [57, 21, 31, 46, 78, 73, 17] or local priors for local implicit functions to represent parts or patches [89, 80, 8, 32, 6, 50] which are further used to approximate a global implicit function. + +These methods require large scale datasets to learn priors. However, the learned priors may not generalize well to unseen point clouds that have large geometric variations compared to training samples. Our method does not require priors, and falls into the following category. + +Learning with Overfitting. To improve the generalization, we can learn implicit functions by overfitting neural networks on single point clouds. Methods using this strategy introduce novel constraints [23, 2, 100, 3, 95, 4], ways of leveraging gradients [49, 14], differentiable poisson solver [65] or specially designed priors [50, 51] to learn signed [49, 23, 2, 100, 3, 11] or unsigned distance functions [14, 102, 47]. + +These methods rely on neural networks to infer implicit functions without learning priors. However, the slow convergence of neural networks is a big limit for them to scale up to large scale point clouds. Our method addresses the scalability challenge by directly inferring a discrete distance field without using neural networks. Our novel loss leads + +to less complexity and higher accuracy than the overfitting based methods. + +Learning with Grids. Learning implicit functions with grids has been used in prior-based methods [77, 54, 66, 79, 46, 34, 85]. Using neural networks, these methods learn features at vertices of grids and further map the interpolated features into a distance or occupancy field. Some methods [70, 75, 96, 38] inferred a discrete radiance field defined on grids to speed up the learning of a radiance field, while they cared more about the quality of synthesized views than the underlying geometry. Without using neural components, some methods [67, 65] directly infer discrete implicit functions on grids. To make up the continuousness of neural networks, they employed different constraints, such as Possion equations or Viscosity priors, to pursue a continuous field from a single point cloud. However, these methods are slowly converged, which makes them hard to scale up to large scale point clouds. Although we also use grids to speed up the inference, we introduce a more efficient training strategy to decrease optimization space and more effective losses to increase inference efficiency. + +# 3. Method + +Overview. GridPull aims to achieve a fast inference of a distance field, such as a signed distance field, from a 3D point cloud $S = \{s_j | j \in [1, J]\}$ without using any neural components. As shown in Fig. 1 (b), we represent the distance field by a signed distance function (SDF) $f$ defined on a set of discrete grids $G$ with a resolution of $R^3$ , where the vertices $V = \{v_i | i \in [1, I], I = (R + 1)^3\}$ shared among these grids hold learnable signed distances $D = \{d_i\}$ . For an arbitrary location $q$ in Fig. 1 (e), such as a randomly sampled query or a surface point, we use trilinear interpolation to approximate the signed distance $f(q)$ from the eight nearest grid vertices below, + +$$ +f (q) = \operatorname {t r i I n t e r} (q, D). \tag {1} +$$ + +Our goal is to conduct direct optimization in Fig. 1 (c) to infer the signed distances $D$ on grid vertices by minimizing our loss function $L$ on queries $q$ below, + +$$ +\min _ {D} L (f (q)). \tag {2} +$$ + +With the learned function $f$ , we run the marching cubes algorithm [48] to reconstruct a mesh as the surface as shown in Fig. 1 (d). + +Discrete Grids. We split the space occupied by the point cloud $S$ into discrete grids $G$ with a resolution of $R$ in Fig. 1 (b). GridPull can work with high resolutions, such as $R = 512$ , because of our efficient and effective loss function and training strategy. Within each grid, we interpolate the signed distance $f(q)$ of query $q$ using the trilinear interpolation in Eq. 1. We organize surface points in a tree + +![](images/08f150614fe630839ea0ac6e26ac5b2b1d0c579a5990a364a1b683d9b7cebbe8.jpg) +(a) +(b) + +![](images/76e610d2943f6f27312a06785d9ac9c9d7367472872d7f7efba2d0a24f353c92.jpg) + +![](images/8690051359f5626af01eb76e2a360584f30fdcb4748706ab0a49e2b01e1baab1.jpg) +(c) +(d) + +![](images/fd13e9eb07f3043383b1a37bd78aa1ddd5a6eccbe4649f185cff8405e64bd0be.jpg) + +![](images/1b53338c2fa128870cd35d21e28f143ad68a8ea9e806f5dc4f35e28675ad9801.jpg) +Figure 1. We proposed GridPull for fast inference of SDF defined on discrete grids using effective losses. + +![](images/4b5487feb83f0d32d23ec8b6d9557ce9d8c53c817fd30259cf13ae3a9593b9ae.jpg) + +![](images/2da1b71496b018182b8f8dcde2174440a5a555c983587028f59e98cd9d0a6c23.jpg) +(g) Total Variation Loss + +![](images/33498d87d523024d17748da6e4788e52089a07a76e909f8d7114a435685bc580.jpg) +(h) Surface Loss + +![](images/23a7055f4c28076833010b9ef76ab69350afabfe54a82aaed36c36f15355ccb6.jpg) +(i) Gradient Loss + +structure to speed up the search of the nearest surface point for queries $q$ . Since we need to extensively conduct trilinear interpolation on signed distances for queries $q$ , we use uniform grids for fast nearest grid search for queries. We did not use an Octree to organize grids, due to its inefficient and inaccurate grid search for queries. + +Geometric Initialization. We initialize the signed distances $d_{i}$ on grid vertices which represent a sphere in the field. The initialization can speed up the convergence during optimization. Our preliminary results show that a small sphere with a radius of one or two grids works well. Since enlarging a small sphere to the target shape can calculate the continuous distance loss over smaller regions than shrinking a large sphere. + +Signed Distance Inference. For a query $q$ , we infer its signed distance $f(q)$ through pulling it onto its nearest surface point in the point cloud $S$ . As shown in Fig. 1 (e), we first obtain $f(q)$ as the interpolation of the signed distances $d_{i}$ on the nearest eight vertices of $q$ . We implement the trilinear interpolation by solving $\mathbf{a} = [a_0,..,a_7]$ in a linear regression system on the eight vertices $v_{i}$ below, + +$$ +f (q) = a _ {0} + a _ {1} x + a _ {2} y + a _ {3} z + a _ {4} x y + a _ {5} x z + a _ {6} y z + a _ {7} x y z, \tag {3} +$$ + +where $[x,y,z]$ is the coordinate of $q$ and the eight constraints are $\{f(v_i) = d_i\}$ . + +With the predicted signed distance $f(q)$ and the gradient $\nabla f(q) = \partial f / \partial q$ , we pull a query $q$ onto the surface and obtain a pulled query $p$ by $p = q - f(q) \times \nabla f(q) / ||\nabla f(q)||_2$ in Fig. 1 (f). + +Our goal is to minimize the distance between the pulled query $p$ and the nearest surface point $NN(q) \in S$ of query $q$ below, + +$$ +L _ {P u l l} = \left\| p - N N (q) \right\| _ {2}. \tag {4} +$$ + +We adopt this pulling loss from [49]. Instead, we do not + +rely on neural network in the pulling procedure. We use the trilinear interpolation on signed distances $d_{i}$ on grid vertices to obtain signed distances and gradients according to Eq. 3. + +We sample queries $q$ around surface points in each epoch. For each randomly sampled surface point, we sample queries using a gaussian distribution with the surface point as the center and a two-grid length variance. We build a KD-tree on surface points so that we can speed up the searching of the nearest neighbor from a large number of surface points for each query. + +To speed up the optimization, we only calculate the pulling loss on grids near the surface. We regard the grids containing surface points as the base and find the union of their $M_{1}$ nearest neighboring grids. These grids form a $M_{1}$ -bandwidth near the surface in Fig. 1 (c). + +Continuous Distance Fields. Although the pulling loss can infer correct signed distances at grid vertices, it fails to learn continuous signed distances across the field, especially near the surface. Fig. 2 (a) shows severe artifacts on the surface and gradients near the surface are messy and inconsistent, which are caused by the discontinuous signed distances across neighboring regions. + +Neural network based methods [49, 51] do not have this issue, since the continuous character of neural networks makes the sudden change of signs on the same shape side hard. While our method uses signed distances $d_{i}$ on grid vertices to represent a discrete field, this makes that one $d_{i}$ can change separately without considering the change of neighboring $d_{i}$ , which leads to a poor continuousness in the field. + +To resolve this issue, we impose a total variation (TV) loss as a continuous constraint on our signed distances $d_{i}$ on grid vertices. Our key idea is to make the change of one $d_{i}$ affect the change of its neighboring $NN(d_{i})$ , which smoothes the local distance field and propagates local up + +![](images/14712135528e45347a8f1b81ce55d4d449a3f94d17b61fc1ce25ec57ec91b9eb.jpg) +Figure 2. The effect of losses on reconstruction. +dates to the rest grids. Hence, as shown in a 2D case in Fig. 1 (g), we approximate one $d_{i}$ to each one of its six neighbors by minimizing the difference between each two below, + +$$ +L _ {T V} = \sqrt {\sum_ {a = \{x , y , z \}} e _ {a + 1} \left(d _ {i}\right) ^ {2} + e _ {a - 1} \left(d _ {i}\right) ^ {2}}, \tag {5} +$$ + +where $e_{a + 1}(d_i)$ is the difference between $d_i$ and the signed distance along the $a$ axis, i.e., $x, y, z$ axis. Fig. 2 (b) shows that the continuous constraints can significantly improve the signed distance field in terms of continuousness and gradients. + +One remaining issue is its computational complexity. Since we are targeting grids at a high resolution, such as $R = 512$ , it is very time consuming to impose this constraint on all $d_{i}$ across the field. + +To resolve this issue, we focus on grids near the surface. Similar to imposing the pulling loss on grids in the $M_{1}$ -bandwidth near the surface, we impose our continuous constraints on the grids in the $M_{2}$ -bandwidth near the surface, which significantly reduces the number of learnable parameters to optimize. + +Signed Distance Supervision. For signed distances, one supervision we can directly use is the input point cloud $S$ . It is a discrete surface indicating the zero level-set of the signed distance function. Hence, as introduced in Fig. 1 (h), we can constrain the signed distances at the points $s$ on the surface below, + +$$ +L _ {\text {S u r f a c e}} = \left\| f (s) \right\| _ {2}. \tag {6} +$$ + +Although Fig. 2 (c) indicates that $L_{Surface}$ slightly improves the signed distance field and leads to a little bit more continuous surface, it is helpful to determine the zero level set with other losses. For noises on $S$ , our continuous constraint can relieve the impact of noises on this supervision. We will show this in experiments. + +Consistent Gradients. Fig. 2 (a) indicates that the pulling loss produces gradients that are messy and non-orthogonal to the surface. This leads to inaccurate signed distance inference during the pulling procedure. + +We additionally introduce a constraint on gradients to improve the field near the surface. As shown in Fig. 1 (i), we aim to encourage the gradient at a query $q$ to point to the + +same direction of the gradient at its nearest surface point $NN(q)$ . We use a cosine distance below to achieve more consistent gradients, + +$$ +L _ {\nabla} = 1 - \cos (\nabla f (q), \nabla f (N N (q))). \tag {7} +$$ + +$L_{\nabla}$ constrains the gradients in the field and corrects inaccurate gradients. Fig. 2 (d) shows that more consistent gradients can improve the surface to be more continuous using small planes on the surface which however the surface not smooth at all. + +Loss Function. We learn signed distances on grid vertices $D$ by minimizing the loss function below, + +$$ +L = L _ {\text {P u l l}} + \alpha L _ {T V} + \beta L _ {\text {S u r f a c e}} + \gamma L _ {\nabla}. \tag {8} +$$ + +where $\alpha$ , $\beta$ , and $\gamma$ are balance weights to make each term contribute equally. Fig. 3 shows the signed distance field on a slide of $D$ during optimization. Our optimization starts from an initial sphere and only optimizes variables in the bandwidth. We skip the grids that are not optimized after the initialization during the marching cubes for surface reconstruction. Please watch our video for more details. + +# 4. Experiments and Analysis + +We evaluate GridPull by numerical and visual comparisons with the latest methods on synthetic and real datasets in surface reconstruction. + +Datasets and Metrics. We use benchmarks for shapes and scenes in evaluations. For shapes, we conduct evaluations on five datasets including a subset of ShapeNet [1], FAMOUS [17], Thingi10k [103], Surface Reconstruction Benchmark (SRB) [89] and D-FAUST [5]. For scenes, we also report comparisons on five datasets including 3DScene [104], SceneNet [27], 3DFRONT [18], Matterport [9] and KITTI [20]. + +We use L2 Chamfer distance $(CD_{L2})$ , $CD_{L1}$ and Hausdorff distance(HD) to measure the error between the reconstructed surface and the ground truth. Moreover, we use normal consistency (NC) and F-score to evaluate the accuracy of normal on the reconstructed surface. We also report our time and storage complexity to highlight our advantage towards scalability. We also use the intersection over union (IoU) to measure the reconstruction for fair comparison with the latest methods. + +Details. We use $R = 256$ to infer a distance field. We calculate the pulling loss on grids in a bandwidth with $M_{1} = 3$ . We impose the continuous constraint on grids in a bandwidth with $M_{2} = 14$ for shapes and $M_{2} = 25$ for scenes, since scenes contain open surfaces with larger empty spaces. We run the marching cubes for surface reconstruction at a resolution of 256. + +We use the Adam optimizer with an initial learning rate of 1.0. We decrease the learning rate with a decay rate of 0.3 every 400 iterations. We use 50,000 queries in each + +![](images/cb201be88e6afbc6a4dc5e742d25ee257e3bb221f91021bf5fbfb201996b9b34.jpg) +0th Iteration + +![](images/27abef32dbbe34dec1596d0ba6aab2d3cf3a43d5ee0f85709d5f62edc7bd2fc4.jpg) +50th Iteration + +![](images/d676db783771bad12e0d9d115cbaa9ba3d1d6b26c5b8c84b62d9d2d23d0a3276.jpg) +100th Iteration + +![](images/c4bad84de10b76dcd23cd53ac6bfb4617f79a26cd074e09af20a9a05e34c9fc7.jpg) +150th Iteration + +![](images/30b683a5ee4e3421fd1acaeaf3ac760458a43704d20ce825d1e0b9bd6d4c4e62.jpg) +200th Iteration + +![](images/10b3e5fbc48694f67a753a9571d9727bc46adc17040db454ae645aec55ca26cd.jpg) +1000th Iteration + +![](images/31cff11eaad808fa8900f90bf875e00b8b4febf4cd424e84e47c335acb46bf1c.jpg) +Figure 3. Visualization of signed distances in optimization. +Figure 4. Visual comparison under ShapetNet dataset. + +iteration, and run 1600 iterations for each point cloud during optimization. We set the balance weights $\alpha = 1$ , $\beta = 1$ , and $\gamma = 0.005$ for equal contribution from the three terms. + +# 4.1. Surface Reconstruction for Shapes + +ShapeNet. Following predictive context prior (PCP) [51] and NeuralPull (NP) [49], we report our results on a subset in ShapeNet in terms of $CD_{L2}$ in Tab. 1, Normal Consistency (NC) in Tab. 2, and F-Score with thresholds of 0.002 and 0.004 in Tab. 3 and Tab. 4, which show our superiority over PCP and NP in numerical comparisons over all classes, even though PCP employs a local shape prior learned from a large scale dataset. The visual comparison in Fig. 4 shows that we can reveal more accurate and detailed geometry. + +
ClassNP [49]PCP [51]Ours
Display0.0390.00870.0082
Lamp0.0800.03800.0347
Airplane0.0080.00650.0007
Cabinet0.0260.01530.0112
Vessel0.0220.00790.0033
Table0.0600.01310.0052
Chair0.0540.01100.0043
Sofa0.0120.00860.0015
Mean0.0380.01360.0086
+ +FAMOUS. We follow the experimental setting in PCP [51] and NP [49] to evaluate GridPull on FAMOUS. We compare GridPull with the latest methods with priors including PCP, GenSDF, FGC and no priors including NP and IGR. GridPull outperforms these methods in terms of $CD_{L2}$ and running time. Although FGC is fast, it does not reconstruct + +Table 1. Reconstruction accuracy under ShapeNet in terms of $CD_{L2} \times 100$ . + +
ClassNP [49]PCP [51]Ours
Display0.9640.97750.9847
Lamp0.9300.94500.9693
Airplane0.9470.94900.9614
Cabinet0.9300.96000.9689
Vessel0.9410.95460.9667
Table0.9080.95950.9755
Chair0.9370.95800.9733
Sofa0.9510.96800.9792
Mean0.9390.95900.9723
+ +Table 2. Reconstruction accuracy under ShapeNet in terms of NC. + +
ClassNP [49]PCP [51]Ours
Display0.9890.99390.9963
Lamp0.8910.93820.9455
Airplane0.9960.99420.9976
Cabinet0.9800.98880.9901
Vessel0.9850.99350.9956
Table0.9220.99690.9977
Chair0.9540.99700.9979
Sofa0.9680.99430.9974
Mean0.9610.98710.9896
+ +Table 3. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.002. + +
ClassNP [49]PCP [51]Ours
Display0.9910.99580.9963
Lamp0.9240.94020.9538
Airplane0.9970.99720.9989
Cabinet0.9890.99390.9946
Vessel0.9900.99580.9972
Table0.9730.99850.9990
Chair0.9690.99910.9990
Sofa0.9740.99870.9992
Mean0.9760.98990.9923
+ +a watertight mesh but just a polygon soup with no normal, which contains almost no geometry details. Worse than that, it fails to reconstruct some point clouds in more than eight hours, hence we remove these shapes from its results. We highlight our superiority in visual comparison in Fig. 5. We can reveal details like hair and expressions more clearly. + +Table 4. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.004. + +
MethodCDL2×100Min timeMean time
IGR [24]1.650402s407s
GenSDF [15]0.668230s232s
NP [49]0.220593s605s
FGC [97]0.05510sTimeout
PCP [51]0.0443416s3534s
Ours0.040198s201s
+ +Table 5. Reconstruction accuracy under FAMOUS. + +![](images/f86dd0e21b1991b0f68701504a745ceb960bcd52bb17aa918ce78f4c79169300.jpg) +Figure 5. Visual comparison under FAMOUS dataset. + +![](images/8d1af343e0df8fb9ceaec63400ed409efda84caa525fc04d7c9419a5cd3d9be3.jpg) +Figure 6. Visual comparison under SRB dataset. + +SRB. We follow the experimental setting in VisCo [67] to evaluate GridPull on real scans in SRB. We report the comparison with the latest overfitting based methods in terms of $CD_{L1}$ in Tab. 6. Meanwhile, we also report their one-sided distances between the reconstructed mesh and the input noisy point cloud. The numerical comparison shows that our method significantly outperforms the latest. Moreover, we also report our results of learning unsigned distances in Tab. 7. We replace our pulling loss with the pulling loss introduced for unsigned distances in CAP-UDF [102], and remove the gradient loss due to the nondifferentiable character of UDF on surface. Visual comparison in Fig. 6 shows that our method can reconstruct surfaces that are more compact than the methods learning SDF like SIREN or more smooth than the methods learning UDF like CAP-UDF. + +Thingi10K. We follow the experimental setting in SAP [65] to evaluate GridPull in Thingi10K. Tab. 8 indicates our superiority over the overfitting based methods in the numerical comparison. Visual comparisons in Fig. 7 show our more accurate surfaces with complex details. + +D-FAUST. We also follow the experimental setting in S-AP [65] to evaluate GridPull in D-FAUST. Tab. 8 shows that we outperform the overfitting based methods with more surface details. As shown in Fig. 7, we recover more accurate + +
Poisson [37]IGR [24]SIREN [71]VisCo [67]SAP [65]Ours
Anchor\(CD_{L1}\)0.600.220.320.210.120.093
\(HD\)14.894.718.193.002.381.804
\(d_{\vec{C}}\)0.600.120.100.150.080.066
\(d_{\vec{H}}\)14.891.322.431.070.830.460
Daratech\(CD_{L1}\)0.440.250.210.260.070.062
\(HD\)7.244.014.304.060.870.648
\(d_{\vec{C}}\)0.440.080.090.140.040.039
\(d_{\vec{H}}\)7.241.591.771.760.410.293
DC\(CD_{L1}\)0.270.170.150.150.070.066
\(HD\)3.102.222.182.221.171.103
\(d_{\vec{C}}\)0.270.090.060.090.040.036
\(d_{\vec{H}}\)3.102.612.762.760.530.539
Gargoyle\(CD_{L1}\)0.260.160.170.170.070.063
\(HD\)6.803.524.644.401.491.129
\(d_{\vec{C}}\)0.260.060.080.110.050.045
\(d_{\vec{H}}\)6.800.810.910.960.780.700
Lord Quas\(CD_{L1}\)0.200.120.170.120.050.047
\(HD\)4.611.170.821.060.980.569
\(d_{\vec{C}}\)0.200.070.120.070.040.031
\(d_{\vec{H}}\)4.610.980.760.640.510.370
+ +Table 6. Reconstruction accuracy under SRB. + +
NDF [14]CAP-UDF [102]Ours(UDF)
CDL10.2380.0730.070
F-Score68.684.585.1
+ +![](images/6998a1e2a06a36e1005cdeb696ed9fdb3464e3220e086208a519e37dda5cdf85.jpg) +Table 7. Accuracy of reconstruction with UDF under SRB. +Figure 7. Visual comparison under Thingi10k and D-FAUST. + +
IGR [24]SAP [65]Ours
Thing10kCDL10.4400.0540.051
F-Score0.5050.9400.948
NC0.6920.9470.965
DFAUSTCDL10.2350.0430.015
F-Score0.8050.9660.975
NC0.9110.9590.978
+ +Table 8. Reconstruction accuracy under Thingi10k and D-FAUST in terms of $CD_{L1}$ and F-Score with a threshold of 0.01. + +human bodies and more detailed expressions on faces. + +# 4.2. Surface Reconstruction for Scenes + +3DScene. We follow PCP [51] to report $CD_{L1}$ , $CD_{L2}$ and Normal Consistency(NC) for evaluation. We report the comparisons with the latest methods in Tab. 9. We outperform both kinds of methods with learned priors such as ConvOcc [73] and PCP [51] and overfitting based NP [49] + +![](images/99d914b16119c2ec4cd76a1aa3e4c74e4f2db9ca5bbcc69f30384563f8658bb7.jpg) +ConvOcc + +![](images/b0b03ddc233ebc125ac799649dd24c3a48b0328a941ac603036c200377755e8b.jpg) +NP + +![](images/726a1d1207ba43fc36d35ceb2ba62a5618f64596babfaea182e970ba9ea5599b.jpg) +PCP + +![](images/e15184067e4c73650994685d61f50203310af44e75d4e2cb6b0f48bd2193374a.jpg) +Ours + +![](images/7576ebcaebb72a25bfc8743731bb47ab16c91e14a0c4f2a0a07a8eb8e2103048.jpg) +GT + +![](images/2b5004b4b02b3108d2f182b8f5c2c0bed111468b24e76b526b3cd624b244a534.jpg) +NP PCP Ours Figure 9. Visual comparison under SceneNet dataset. + +![](images/692d5de9e25cfe3f42a3d65d054294fe01c8cd70f846d58003ccdceed3682c3e.jpg) +Figure 8. Visual comparison under 3DScene dataset. + +![](images/d9fd1564499f529be4c258b65137ff4e5878e41de19f8401ed56c4ad6ec1b688.jpg) + +in all scenes. The visual comparisons 8 shows that our method can work well with real scans on scenes and reveal more geometry details on surfaces. + +
ConvOcc [73]NP [49]PCP[51]Ours
BurghersCDL2× 10026.691.760.2670.246
CDL10.0770.0100.0080.008
NC0.8650.8830.9140.926
LoungeCDL2× 1008.6839.710.0610.055
CDL10.0420.0590.0060.005
NC0.8570.8570.9280.922
CopyroomCDL2× 10010.990.510.0760.069
CDL10.0450.0110.0070.006
NC0.8480.8840.9180.929
StonewallCDL2× 10019.120.0630.0610.058
CDL10.0660.0070.00650.006
NC0.8660.8680.8880.893
TotepoleCDL2× 1001.160.190.100.093
CDL10.0160.0100.0080.007
NC0.9250.7650.7840.847
+ +SceneNet. Following the experimental setting in PCP [51], we report our $CD_{L1}$ , NC, and F-score in Tab. 10. The comparison indicates that we outperform these methods by producing more accurate and smooth surfaces. This is also verified by our visual comparison in Fig. 9, where the latest methods reveal either no or inaccurate indoor geometry. + +3D-FRONT. We follow the experimental setting in NeuralPoisson [16] to report $CD_{L1}$ , IoU and L2 distance over the voxelization of meshes. IoU and L2 distances are calculated as the error between the reconstructed surfaces and the ground truth over the voxelization at a resolution of 128. The comparison in Tab. 11 shows that our results are more accurate than the latest method. Our visual comparison with SIREN in Fig. 10 shows that GridPull can infer an accurate zero level set which leads to very sharp corners. + +Matterport. We follow the experimental setting in NGS [30], and report the comparison in terms of $CD_{L1}$ , NC, and F-score in Tab. 12. We achieve the best perfor + +Table 9. Reconstruction accuracy under 3DScene. + +
NP [49]PCP [51]Ours
LivingroomCDL10.0880.0170.015
NC0.8810.9330.922
FScore0.8010.9660.953
BathroomCDL10.0360.0160.013
NC0.9120.9450.952
FScore0.8600.9770.974
BedroomCDL10.0340.0140.014
NC0.9050.9480.951
FScore0.8760.9800.986
KitchenCDL10.0490.0150.014
NC0.9000.9450.952
FScore0.8250.9760.984
OfficeCDL10.0620.0240.022
NC0.8790.9190.931
FScore0.7299250.940
MeanCDL10.0540.0170.0156
NC0.8950.9380.9416
FScore0.8189650.9674
+ +![](images/8b0fda8a33aa1e8759e1975383b9e53745aafc09b759d40b8e63eddc7d4f7bff.jpg) +Figure 10. Visual comparison under 3Dfront dataset. + +Table 10. Reconstruction accuracy under SceneNet. FScore is with a threshold of 0.025. + +
MethodCDIOUL2
SIREN [71]0.01830.5242.373
NeuralPoisson [16]0.01610.5452.073
Ours0.01410.5621.984
+ +Table 11. Reconstruction accuracy under 3DFRONT. + +mance in terms of accuracy. NGS [30] has a pretty fast inference but it costs about four to five days to learn priors from the training set in Matterport. Visual comparison in Fig. 11 shows that we can reconstruct more accurate surfaces while NGS is poorly generalized to unseen scenes. + +KITTI. We evaluate GridPull on a large scale real scan from KITTI odometry (Sequence00, frame 3000 to 4000) which contains about 13.8 million points. The comparison in Fig. 12 shows that we can reconstruct a more accurate + +![](images/47e2860efde063c793afb59182941a65fc86b588c467ed0d4cd65644d4d94b02.jpg) +Figure 11. Visual comparison under Matterport3D dataset. + +![](images/f4c7c19071dc506bff46b81312b521853ec3323365f932e776ef44a583d0104c.jpg) +Figure 12. Visual comparison under Kitti dataset. + +surface in much less time (3810s) than the overfitting based method SAP (13005s) [65]. + +
MethodCDL1× 1000F-Score(0.002)NCTime
ConvOcc [73]6.2188.992.034s
SAP [65]4.1791.291.4643s
NGS [30]3.0497.395.72.1s
Ours2.9397.796.3223s
+ +# 4.3. Ablation Studies + +We conduct ablation studies on FAMOUS [17] to justify the effectiveness of modules in GridPull. + +Loss. We justify the effectiveness of our loss in terms of $CD_{L2}$ in Tab. 13. For the pulling loss, it is able to estimate a coarse shape but with discontinuous meshes and artifacts. The gradient loss can improve the continuousness on the surface. Using the TV loss, we obtained a more continu- + +![](images/c04bd206759f675ef696837d34dcb1a2739ae0ae0dbe511b19630e0eb00f766a.jpg) +Figure 13. Effect of Resolution $R$ . + +ous field, which removes most of artifacts and significantly improve the accuracy. The surface loss encourages a more accurate zero level-set, which also improves the accuracy. + +Table 12. Reconstruction accuracy under Matterport3D. + +
LossCDL2×100
LPull0.184
LPull + L∇0.126
LPull + L∇ + LTV0.056
LPull + L∇ + LTV + LSurface0.040
+ +Resolution $R$ . We compare the effect of resolution $R$ on the reconstructed surfaces. We try several candidates $\{32, 64, 128, 256, 512\}$ to learn the discrete distance field. We use the same bandwidth for the calculation of the pulling loss and the TV loss, and reconstruct meshes by running the marching cubes at the same resolution. The numerical and visual comparisons show that higher resolutions can infer more geometry details but also cost more inference time. + +Table 13. Effect of losses. + +
3264128256512
CDL2×1000.1860.0970.0480.0400.035
Time50s67s84s201s645s
+ +Table 14. Effect of Resolution $R$ . + +Point Number. We compare our performance on d-different numbers of points. We use the same set of shapes but with different numbers of points including $\{10K, 100K, 1000K\}$ . The numerical comparison in Tab. 15 shows that our method can estimate a discrete distance field from different numbers of points well, which leads to more accurate reconstructed surfaces than the latest overfitting based methods. In addition, our time complexity is not affected by the number of points a lot, while the latest + +methods require more time to converge on more points. + +
NumberMethodCDL2×100TimeMemory
10KNP0.351548s4.5G
SAP0.275147s3.5G
Ours0.12282s3.3G
100KNP0.224615s4.6G
SAP0.076354s3.6G
Ours0.04084s3.3G
1000KNP0.1751304s4.8G
SAP0.063841s3.7G
Ours0.03990s3.4G
+ +Table 15. Effect of point numbers. + +Initialization. We highlight our geometric initialization of distances on grid vertices by comparing it with random initialization. The comparison in Tab. 16 shows that random initialization makes the optimization converge hard, and cannot work well with the TV loss, especially with a bandwidth, which produces lots of artifacts in empty spaces. Moreover, we try an Eikonal constraint on gradients based on the geometric initialization, but did not get improvement. + +
RandomGeometric||∇|| = 1
CDL2 × 1000.1550.0400.054
+ +Table 16. Effect of Initialization. + +Bandwidth. We report the effect of bandwidth $M_{1}$ and $M_{2}$ in the pulling loss and the TV loss in Tab. 17. For both losses, the bandwidth controls the scope of query sampling and continuous constraints. A larger bandwidth would cover more parameters which require more time and make the optimization converge slowly. For instance, if we sample queries and impose the pulling in all $R^3$ grids, the speed is even slower than NeuralPull which uses neural network for the pulling, since neural networks can generalize signed distances nearby and do not need to infer in all grids. Although a small bandwidth would save time, a too small bandwidth would cover too few grids around the surface, which leads to discontinuousness and artifacts. While if the bandwidth is large enough, we may not see surface improvement but just cost more time. Moreover, even with the same loss, NeuralPull does not produce better results than ours. Comparing with methods using neural networks, our discrete S-DF can have faster convergence, as indicated by the angular error of gradients in Fig. 14. + +
BandwidthCDL2×100Time
LPull (M1=3)0.04084s
LPull (All Grids)0.043628s
LTV (M2=2)0.07983s
LTV (M2=14)0.04084s
LTV (M2=30)0.04189s
LTV (M2=50)0.04096s
LTV (All Grids)0.040101s
NeuralPull0.220552s
NeuralPull+L∇+LSurface0.203554s
+ +Table 17. Effect of Bandwidth. + +![](images/6298761a19a7f8987f7988b7acd038e56b09aaf3d4436cf5813001e50a8dda81.jpg) +Figure 14. Gradients comparison with neural networks. + +Noises. We report our performance on point clouds with noise. We follow the experimental setting in PCP, and report our results with middle and large level noise in Tab. 18. To handle noise, we weight more on the continuous constraint $L_{TV}$ using a larger $\alpha = 2$ to decrease the impact of noise. The comparison shows that our method can handle noise better than the neural network based methods. Since the continuousness of neural networks make gradients near the surface over smooth to overfit all noise, this makes it hard to infer accurate zero level-set. + +
Noise levelNP [49]PCP [51]Ours
F-med-noise0.2800.0710.044
F-max-noise0.3100.2980.060
+ +Table 18. Effect of noises. + +Normals. We report our performance on point clouds with normals. If GT normal is available, we could add a supervision on normals at points on the surface. What we do is to supervise the gradients $\lambda f$ at points on the surface using their ground truth normals. The comparisons in Tab. 19 show that the normal supervision improves the performance a little. This indicates that our method can infer pretty accurate gradients on the surface or near the surface, where gradients on the surface can be highly accurate to the normal ground truth on surfaces. + +
ClassOursOurs(Normals)
CDL20.04010.0403
+ +Table 19. Effect of normals. + +# 5. Conclusion + +We introduce GridPull for the fast inference of distance fields from large scale point clouds without using neural components. We infer a distance field using learnable parameters defined on discrete grids, which we can directly optimize in a more efficient way than neural networks. Besides the complexity advantage, our loss function manages to achieve more continuous distance fields with more consistent gradients, which leads to higher accuracy in surface reconstruction. We evaluate GridPull and justify its effectiveness on benchmarks. Numerical and visual comparisons show that GridPull outperforms the latest methods in terms of accuracy and complexity, even without using priors, normal, and neural networks. + +# References + +[1] ShapeNet. https://www.shapenet.org/.4 +[2] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2 +[3] Matan Atzmon and yaron Lipman. SALD: sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021. 1, 2 +[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. CoRR, abs/2106.10811, 2021. 1, 2 +[5] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Computer Vision and Pattern Recognition, 2017. 4 +[6] Alexandre Boulch and Renaud Marlet. POCO: Point convolution for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2 +[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2 +[8] Rohan Chabra, Jan Eric Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard A. Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conference on Computer Vision, volume 12374, pages 608-625, 2020. 1, 2 +[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 4 +[10] Chao Chen, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Unsupervised learning of fine structure generation for 3D point clouds by 2D projections matching. In IEEE International Conference on Computer Vision, 2021. 1, 2 +[11] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Latent partition implicit with surface codes for 3d representation. In European Conference on Computer Vision, 2022. 1, 2 +[12] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Unsupervised inference of signed distance functions from single-sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17712-17723, 2023. 1 +[13] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[14] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. arXiv, 2010.13938, 2020. 1, 2, 6 +[15] Gene Chou, Ilya Chugunov, and Felix Heide. GenSDF: Two-stage learning of generalizable signed distance functions. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 5 + +[16] Angela Dai and Matthias Nießner. Neural Poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 7 +[17] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J. Mitra, and Michael Wimmer. Points2Surf: Learning implicit surfaces from point clouds. In European Conference on Computer Vision, 2020. 1, 2, 4, 8 +[18] Huan Fu, Bowen Cai, Lin Gao, Lingxiao Zhang, Cao Li, Zengqi Xun, Chengyue Sun, Yiyun Fei, Yu Zheng, Ying Li, Yi Liu, Peng Liu, Lin Ma, Le Weng, Xiaohang Hu, Xin Ma, Qian Qian, Rongfei Jia, Binqiang Zhao, and Hao Zhang. 3d-front: 3d furnished rooms with layouts and semantics. CoRR, abs/2011.09127, 2020. 4 +[19] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. 2022. 2 +[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition, 2012. 4 +[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In International Conference on Computer Vision, 2019. 1, 2 +[22] Lily Goli, Daniel Rebain, Sara Sabour, Animesh Garg, and Andrea Tagliasacchi. nerf2nerf: Pairwise registration of neural radiance fields. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9354-9361. IEEE, 2023. 2 +[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799, 2020. 1, 2 +[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv, 2002.10099, 2020. 5, 6 +[25] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. ShapeCaptioner: Generative caption network for 3D shapes by learning a mapping from parts detected in multiple views to sentences. ArXiv, abs/1908.00120, 2019. 2 +[26] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. DRWR: A differentiable renderer without rendering for unsupervised 3D structure learning from silhouette images. In International Conference on Machine Learning, 2020. 1, 2 +[27] A. Handa, V. Patraucean, V. Badrinarayanan, S. Stent, and R. Cipolla. Understanding realworld indoor scenes with synthetic data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4077-4085, 2016. 4 +[28] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2 + +[29] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[30] Jiahui Huang, Hao-Xiang Chen, and Shi-Min Hu. A neural galerkin solver for accurate surface reconstruction. ACM Trans. Graph., 41(6), 2022. 7, 8 +[31] Meng Jia and Matthew Kyan. Learning occupancy function from point clouds for surface reconstruction. arXiv, 2010.11378, 2020.1, 2 +[32] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3D scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2 +[33] Sijia Jiang, Jing Hua, and Zhizhong Han. Coordinate quantized neural implicit representations for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2 +[34] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[35] Chuan Jin, Tieru Wu, and Junsheng Zhou. Multi-grid representation with field regularization for self-supervised surface reconstruction from point clouds. Computers & Graphics, 2023. 1 +[36] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2 +[37] Michael M. Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions Graphics, 32(3):29:1-29:13, 2013. 6 +[38] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based implicit surface representation. CoRR, abs/2208.10925, 2022. 2 +[39] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2 +[40] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3D shapes with dynamic code clouds. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12830-12840, 2022. 1 +[41] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2 +[42] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. SDF-SRN: Learning signed distance 3D object reconstruction from static images. In Advances in Neural Information Processing Systems, 2020. 2 +[43] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In European Conference on Computer vision, 2020. 2 + +[44] Shichen Liu, Shunsuke Saito, Weikai Chen, and Hao Li. Learning to infer implicit surfaces without 3D supervision. In Advances in Neural Information Processing Systems, 2019. 2 +[45] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Pengshuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3D reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2 +[47] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. NeUDF: Leaning neural unsigned distance fields with volume rendering. In Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[48] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. Computer Graphics, 21(4):163-169, 1987. 1, 2 +[49] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-Pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In International Conference on Machine Learning, 2021. 1, 2, 3, 5, 6, 7, 9 +[50] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6305-6315, 2022. 1, 2 +[51] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6316-6327, 2022. 2, 3, 5, 6, 7, 9 +[52] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3d point clouds via noise to noise mapping. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1 +[53] Baorui Ma, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Towards better gradient consistency for neural signed distance functions via level set alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17724-17734, 2023. 2 +[54] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. A-CORN: adaptive coordinate networks for neural scene representation. CoRR, abs/2105.02788, 2021. 1, 2 +[55] Xiaoxu Meng, Weikai Chen, and Bo Yang. NeAT: Learning neural implicit surfaces with arbitrary topologies from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 2 +[56] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 + +[57] Zhenxing Mi, Yiming Luo, and Wenbing Tao. SSRNet: Scalable 3D surface reconstruction network. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2 +[58] Mateusz Michalkiewicz, Jhony K. Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3D shape inference. CoRR, abs/1901.06802, 2019. 2 +[59] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 1, 2 +[60] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2 +[61] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[62] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision, 2021. 1, 2 +[63] Amine Ouasfi and Adnane Boukhayma. Few 'zero level set'-shot learning of shape signed distance functions in feature space. In European Conference on Computer Vision, 2022. 2 +[64] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[65] Songyou Peng, Chiyu “Max” Jiang, Yiji Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 8 +[66] Songyou Peng, Michael Niemeyer, Lars M. Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, volume 12348, pages 523-540, 2020. 2 +[67] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 2, 6 +[68] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. In International Conference on Machine Learning, 2021. 1, 2 +[69] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[70] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa. + +Plenoxels: Radiance fields without neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[71] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 6, 7 +[72] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2 +[73] Lars Mescheder Marc Pollefeys Andreas Geiger Songyou Peng, Michael Niemeyer. Convolutional occupancy networks. In European Conference on Computer Vision, 2020. 1, 2, 6, 7, 8 +[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CoRR, abs/2111.11215, 2021. 2 +[75] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[76] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2 +[77] Towaki Takikawa, Joel Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2 +[78] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. SA-ConvONet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2 +[79] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Annual Conference on Neural Information Processing Systems, 2021. 2 +[80] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. PatchNets: Patch-Based Generalizable Deep Implicit 3D Shape Representations. European Conference on Computer Vision, 2020. 1, 2 +[81] Delio Vicini, Sbastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Transactions on Graphics, 41(4):125:1-125:18, 2022. 2 +[82] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. NeuRIS: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, 2022. 2 +[83] Meng Wang, Yu-Shen Liu, Yue Gao, Kanle Shi, Yi Fang, and Zhizhong Han. Lp-dif: Learning local pattern-specific deep implicit function for 3d objects and scenes. In Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21856-21865, June 2023. 1 +[84] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems, pages 27171-27183, 2021. 2 +[85] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics, 41(4):1-15, 2022. 2 +[86] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. HF-NeuS: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems, 2022. 2 +[87] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net++: Point cloud completion by transformer-enhanced multi-step point moving paths. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):852-867, 2022. 2 +[88] Xin Wen, Junsheng Zhou, Yu-Shen Liu, Hua Su, Zhen Dong, and Zhizhong Han. 3d shape reconstruction from 2d images with disentangled attribute flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3803-3813, 2022. 1 +[89] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 4 +[90] Yunjie Wu and Zhengxing Sun. DFR: differentiable function rendering for learning 3D generation from images. Computer Graphics Forum, 39(5):241-252, 2020. 2 +[91] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In IEEE International Conference on Computer Vision, 2021. 1, 2 +[92] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflake point deconvolution for point cloud completion and generation with skip-transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6320-6338, 2022. 2 +[93] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems, 2021. 2 +[94] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2 +[95] Wang Yifan, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Iso-Points: Optimizing neural implicit surfaces with hybrid representations. CoRR, abs/2012.06434, 2020. 1, 2 +[96] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of + +neural radiance fields. In IEEE International Conference on Computer Vision, 2021. 2 +[97] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, June 2022. 5 +[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. ArXiv, abs/2022.00665, 2022. 2 +[99] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3D objects with differentiable rendering of sdf shape priors. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[100] Wenbin Zhao, Jiabao Lei, Yuxin Wen, Jianguo Zhang, and Kui Jia. Sign-agnostic implicit learning of surface self-similarities for shape modeling and reconstruction from raw point clouds. CoRR, abs/2012.07498, 2020. 1, 2 +[101] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, and Zhizhong Han. Learning a more continuous zero level set in unsigned distance fields through level set projection. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2 +[102] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. Advances in Neural Information Processing Systems, 35:16481-16494, 2022. 2, 6 +[103] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. CoRR, abs/1605.04797, 2016. 4 +[104] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics, 32(4):112:1-112:8, 2013. 4 \ No newline at end of file diff --git a/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/images.zip b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..045085a3855892d9dd47daa328c8e5e3991304f0 --- /dev/null +++ b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17790db5784ae3f5fe8ed4ebef7626c1e69d6356ed74cd2995107226e7673195 +size 1091771 diff --git a/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/layout.json b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2fbf15d66ef3ee4e85de17bc5333533133abb0d8 --- /dev/null +++ b/2023/GridPull_ Towards Scalability in Learning Implicit Representations from 3D Point Clouds/layout.json @@ -0,0 +1,13771 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 533, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 533, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 533, + 138 + ], + "type": "text", + "content": "GridPull: Towards Scalability in Learning Implicit Representations from 3D Point Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "spans": [ + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "type": "text", + "content": "Chao Chen" + }, + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "type": "text", + "content": " Yu-Shen Liu" + }, + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "type": "text", + "content": "* Zhizhong Han" + }, + { + "bbox": [ + 179, + 160, + 413, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 176, + 424, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 176, + 424, + 189 + ], + "spans": [ + { + "bbox": [ + 169, + 176, + 424, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 176, + 424, + 189 + ], + "type": "text", + "content": "School of Software, Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 190, + 460, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 460, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 460, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 190, + 460, + 203 + ], + "type": "text", + "content": "Department of Computer Science, Wayne State University, Detroit, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 205, + 507, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 205, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 86, + 205, + 507, + 217 + ], + "type": "text", + "content": "chenchao19@mails.tsinghua.edu.cn liuyushen@tsinghua.edu.cn h312h@wayne.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 594 + ], + "type": "text", + "content": "Learning implicit representations has been a widely used solution for surface reconstruction from 3D point clouds. The latest methods infer a distance or occupancy field by overfitting a neural network on a single point cloud. However, these methods suffer from a slow inference due to the slow convergence of neural networks and the extensive calculation of distances to surface points, which limits them to small scale points. To resolve the scalability issue in surface reconstruction, we propose GridPull to improve the efficiency of learning implicit representations from large scale point clouds. Our novelty lies in the fast inference of a discrete distance field defined on grids without using any neural components. To remedy the lack of continuousness brought by neural networks, we introduce a loss function to encourage continuous distances and consistent gradients in the field during pulling queries onto the surface in grids near to the surface. We use uniform grids for a fast grid search to localize sampled queries, and organize surface points in a tree structure to speed up the calculation of distances to the surface. We do not rely on learning priors or normal supervision during optimization, and achieve superiority over the latest methods in terms of complexity and accuracy. We evaluate our method on shape and scene benchmarks, and report numerical and visual comparisons with the latest methods to justify our effectiveness and superiority. The code is available at https://github.com/chenchao15/GridPull." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 619, + 128, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 128, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 128, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 639, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 287, + 664 + ], + "type": "text", + "content": "It is vital to reconstruct surfaces from 3D point clouds for downstream applications. A widely used strategy is to learn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 245, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 245, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 245, + 545, + 376 + ], + "type": "text", + "content": "implicit representations [59, 62, 26, 10, 91, 77, 54, 68, 88] from 3D point clouds in a data-driven manner. With the learned implicit representations, we can reconstruct surfaces in meshes by running the marching cubes algorithm [48]. By learning priors from large scale datasets during training, previous methods [57, 21, 31, 46, 78, 73, 17] generalize the learned priors to either global implicit functions [89, 80, 8, 32, 6, 50] or local ones for unseen point clouds. However, the generalization ability limits their performances on large structure or geometry variations of training samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 510 + ], + "type": "text", + "content": "More recent methods [23, 12, 2, 100, 3, 95, 4, 49, 14, 65, 35] achieved better generalization by directly overfitting neural networks on single unseen point clouds. They rely on extensive calculations of distances between queries and surface points to probe the space, which supervises neural networks to converge to a distance or occupancy field [49, 23, 2, 100, 3, 11, 52, 83, 40]. However, this inference procedure is time consuming due to the extensive distance calculations and the slow convergence of neural networks. This demerit makes these methods hard to scale up to large scale point clouds for surface reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": "To address the scalability challenge, we propose GridPull to speed up the learning of implicit function from large scale point clouds. GridPull does not require learned priors or point normal, and directly infers a distance field from a point cloud without using any neural components. We infer the distance field on grids near the surface, which reduces the number of grids we need to infer. Moreover, we organize surface points in a tree structure to speed up the nearest neighbor search for the calculation of distances to the surface. Specifically, we infer the discrete distance field by pulling queries onto the surface in grids of interests. Our loss function encourages continuous distances and consistent gradients in the field, which makes up the lack of continuousness brought by neural networks. We justify our effectiveness and highlight our superiority by numerical and visual comparisons with the latest methods on the widely used benchmarks. Our contributions are listed below." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 674, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 674, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 287, + 713 + ], + "type": "text", + "content": "*The corresponding author is Yu-Shen Liu. This work was supported by National Key R&D Program of China (2022YFC3800600), the National Natural Science Foundation of China (62272263, 62072268), and in part by Tsinghua-Kuaishou Institute of Future Media Data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18322" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 223 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 54, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 286, + 131 + ], + "type": "text", + "content": "i) We propose GridPull to reconstruct surfaces from large scale point clouds without using neural networks. GridPull speeds up the learning of implicit function, which addresses the scalability challenge in surface reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 141, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 141, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 52, + 141, + 286, + 178 + ], + "type": "text", + "content": "ii) We introduce a loss function to directly infer a discrete distance field defined on grids but achieving continuous distances and consistent gradients in the field." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 188, + 286, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 188, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 188, + 286, + 223 + ], + "type": "text", + "content": "iii) Our method outperforms state-of-the-art methods in surface reconstruction in terms of speed and accuracy on the widely used benchmarks." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 236, + 133, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 133, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 133, + 249 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 257, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 286, + 376 + ], + "type": "text", + "content": "Neural implicit representations have made a huge progress in various tasks [59, 62, 26, 10, 101, 91, 77, 54, 68, 53, 39, 36, 60, 28, 7, 41, 76]. We can use different supervision including 3D supervision [58, 64, 63, 56, 13, 92, 87], multi-view [72, 45, 34, 99, 44, 90, 61, 42, 94, 93, 19, 84, 98, 86, 81, 82, 25, 22, 29, 69, 74, 33, 55], and point clouds [89, 43, 57, 21] to learn neural implicit representations. In the following, we focus on reviewing works on surface reconstruction by learning implicit representations point clouds below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 377, + 286, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 286, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 286, + 484 + ], + "type": "text", + "content": "Learning with Priors. With neural networks, one intuitive strategy is to use neural networks to learn priors from a training set and then generalize the learned priors to unseen samples. We can learn either global priors to map a shape level point cloud into a global implicit function [57, 21, 31, 46, 78, 73, 17] or local priors for local implicit functions to represent parts or patches [89, 80, 8, 32, 6, 50] which are further used to approximate a global implicit function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 544 + ], + "type": "text", + "content": "These methods require large scale datasets to learn priors. However, the learned priors may not generalize well to unseen point clouds that have large geometric variations compared to training samples. Our method does not require priors, and falls into the following category." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 640 + ], + "type": "text", + "content": "Learning with Overfitting. To improve the generalization, we can learn implicit functions by overfitting neural networks on single point clouds. Methods using this strategy introduce novel constraints [23, 2, 100, 3, 95, 4], ways of leveraging gradients [49, 14], differentiable poisson solver [65] or specially designed priors [50, 51] to learn signed [49, 23, 2, 100, 3, 11] or unsigned distance functions [14, 102, 47]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 712 + ], + "type": "text", + "content": "These methods rely on neural networks to infer implicit functions without learning priors. However, the slow convergence of neural networks is a big limit for them to scale up to large scale point clouds. Our method addresses the scalability challenge by directly inferring a discrete distance field without using neural networks. Our novel loss leads" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "to less complexity and higher accuracy than the overfitting based methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 95, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 323 + ], + "type": "text", + "content": "Learning with Grids. Learning implicit functions with grids has been used in prior-based methods [77, 54, 66, 79, 46, 34, 85]. Using neural networks, these methods learn features at vertices of grids and further map the interpolated features into a distance or occupancy field. Some methods [70, 75, 96, 38] inferred a discrete radiance field defined on grids to speed up the learning of a radiance field, while they cared more about the quality of synthesized views than the underlying geometry. Without using neural components, some methods [67, 65] directly infer discrete implicit functions on grids. To make up the continuousness of neural networks, they employed different constraints, such as Possion equations or Viscosity priors, to pursue a continuous field from a single point cloud. However, these methods are slowly converged, which makes them hard to scale up to large scale point clouds. Although we also use grids to speed up the inference, we introduce a more efficient training strategy to decrease optimization space and more effective losses to increase inference efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 333, + 361, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 361, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 361, + 345 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": "Overview. GridPull aims to achieve a fast inference of a distance field, such as a signed distance field, from a 3D point cloud " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "S = \\{s_j | j \\in [1, J]\\}" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": " without using any neural components. As shown in Fig. 1 (b), we represent the distance field by a signed distance function (SDF) " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": " defined on a set of discrete grids " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": " with a resolution of " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "R^3" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": ", where the vertices " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "V = \\{v_i | i \\in [1, I], I = (R + 1)^3\\}" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": " shared among these grids hold learnable signed distances " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "D = \\{d_i\\}" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": ". For an arbitrary location " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": " in Fig. 1 (e), such as a randomly sampled query or a surface point, we use trilinear interpolation to approximate the signed distance " + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "inline_equation", + "content": "f(q)" + }, + { + "bbox": [ + 304, + 354, + 545, + 496 + ], + "type": "text", + "content": " from the eight nearest grid vertices below," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 509, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 509, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 376, + 509, + 545, + 521 + ], + "type": "interline_equation", + "content": "f (q) = \\operatorname {t r i I n t e r} (q, D). \\tag {1}", + "image_path": "aaa802f73859e9557ad220c81e06dcb54523f0084b146cbbeedfa3664fe5c76a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "text", + "content": "Our goal is to conduct direct optimization in Fig. 1 (c) to infer the signed distances " + }, + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "text", + "content": " on grid vertices by minimizing our loss function " + }, + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "text", + "content": " on queries " + }, + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 525, + 545, + 562 + ], + "type": "text", + "content": " below," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 396, + 573, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 573, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 396, + 573, + 545, + 590 + ], + "type": "interline_equation", + "content": "\\min _ {D} L (f (q)). \\tag {2}", + "image_path": "fb281379db61c69466a782f018dca9c808cde5fe47396ded2415fcc9ca0482b3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "content": "With the learned function " + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "content": ", we run the marching cubes algorithm [48] to reconstruct a mesh as the surface as shown in Fig. 1 (d)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Discrete Grids. We split the space occupied by the point cloud " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " into discrete grids " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " with a resolution of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " in Fig. 1 (b). GridPull can work with high resolutions, such as " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "R = 512" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", because of our efficient and effective loss function and training strategy. Within each grid, we interpolate the signed distance " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(q)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " of query " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " using the trilinear interpolation in Eq. 1. We organize surface points in a tree" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18323" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 84, + 175, + 169 + ], + "blocks": [ + { + "bbox": [ + 93, + 72, + 105, + 83 + ], + "lines": [ + { + "bbox": [ + 93, + 72, + 105, + 83 + ], + "spans": [ + { + "bbox": [ + 93, + 72, + 105, + 83 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 84, + 175, + 169 + ], + "lines": [ + { + "bbox": [ + 89, + 84, + 175, + 169 + ], + "spans": [ + { + "bbox": [ + 89, + 84, + 175, + 169 + ], + "type": "image", + "image_path": "08f150614fe630839ea0ac6e26ac5b2b1d0c579a5990a364a1b683d9b7cebbe8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 71, + 170, + 83 + ], + "lines": [ + { + "bbox": [ + 158, + 71, + 170, + 83 + ], + "spans": [ + { + "bbox": [ + 158, + 71, + 170, + 83 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 175, + 71, + 271, + 171 + ], + "blocks": [ + { + "bbox": [ + 175, + 71, + 271, + 171 + ], + "lines": [ + { + "bbox": [ + 175, + 71, + 271, + 171 + ], + "spans": [ + { + "bbox": [ + 175, + 71, + 271, + 171 + ], + "type": "image", + "image_path": "76e610d2943f6f27312a06785d9ac9c9d7367472872d7f7efba2d0a24f353c92.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 271, + 71, + 433, + 171 + ], + "blocks": [ + { + "bbox": [ + 289, + 72, + 301, + 83 + ], + "lines": [ + { + "bbox": [ + 289, + 72, + 301, + 83 + ], + "spans": [ + { + "bbox": [ + 289, + 72, + 301, + 83 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 271, + 71, + 433, + 171 + ], + "lines": [ + { + "bbox": [ + 271, + 71, + 433, + 171 + ], + "spans": [ + { + "bbox": [ + 271, + 71, + 433, + 171 + ], + "type": "image", + "image_path": "8690051359f5626af01eb76e2a360584f30fdcb4748706ab0a49e2b01e1baab1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 72, + 442, + 83 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 442, + 83 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 442, + 83 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 427, + 84, + 487, + 168 + ], + "blocks": [ + { + "bbox": [ + 427, + 84, + 487, + 168 + ], + "lines": [ + { + "bbox": [ + 427, + 84, + 487, + 168 + ], + "spans": [ + { + "bbox": [ + 427, + 84, + 487, + 168 + ], + "type": "image", + "image_path": "fd13e9eb07f3043383b1a37bd78aa1ddd5a6eccbe4649f185cff8405e64bd0be.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 47, + 181, + 150, + 278 + ], + "blocks": [ + { + "bbox": [ + 47, + 181, + 150, + 278 + ], + "lines": [ + { + "bbox": [ + 47, + 181, + 150, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 150, + 278 + ], + "type": "image", + "image_path": "1b53338c2fa128870cd35d21e28f143ad68a8ea9e806f5dc4f35e28675ad9801.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 277, + 485, + 287 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 485, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 485, + 287 + ], + "type": "text", + "content": "Figure 1. We proposed GridPull for fast inference of SDF defined on discrete grids using effective losses." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 155, + 181, + 230, + 275 + ], + "blocks": [ + { + "bbox": [ + 155, + 181, + 230, + 275 + ], + "lines": [ + { + "bbox": [ + 155, + 181, + 230, + 275 + ], + "spans": [ + { + "bbox": [ + 155, + 181, + 230, + 275 + ], + "type": "image", + "image_path": "4b5487feb83f0d32d23ec8b6d9557ce9d8c53c817fd30259cf13ae3a9593b9ae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 243, + 175, + 340, + 264 + ], + "blocks": [ + { + "bbox": [ + 243, + 175, + 340, + 264 + ], + "lines": [ + { + "bbox": [ + 243, + 175, + 340, + 264 + ], + "spans": [ + { + "bbox": [ + 243, + 175, + 340, + 264 + ], + "type": "image", + "image_path": "2da1b71496b018182b8f8dcde2174440a5a555c983587028f59e98cd9d0a6c23.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 265, + 341, + 274 + ], + "lines": [ + { + "bbox": [ + 255, + 265, + 341, + 274 + ], + "spans": [ + { + "bbox": [ + 255, + 265, + 341, + 274 + ], + "type": "text", + "content": "(g) Total Variation Loss" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 342, + 175, + 442, + 258 + ], + "blocks": [ + { + "bbox": [ + 342, + 175, + 442, + 258 + ], + "lines": [ + { + "bbox": [ + 342, + 175, + 442, + 258 + ], + "spans": [ + { + "bbox": [ + 342, + 175, + 442, + 258 + ], + "type": "image", + "image_path": "33498d87d523024d17748da6e4788e52089a07a76e909f8d7114a435685bc580.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 264, + 421, + 274 + ], + "lines": [ + { + "bbox": [ + 361, + 264, + 421, + 274 + ], + "spans": [ + { + "bbox": [ + 361, + 264, + 421, + 274 + ], + "type": "text", + "content": "(h) Surface Loss" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 441, + 181, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 441, + 181, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 441, + 181, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 441, + 181, + 545, + 257 + ], + "type": "image", + "image_path": "23a7055f4c28076833010b9ef76ab69350afabfe54a82aaed36c36f15355ccb6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 264, + 524, + 274 + ], + "lines": [ + { + "bbox": [ + 459, + 264, + 524, + 274 + ], + "spans": [ + { + "bbox": [ + 459, + 264, + 524, + 274 + ], + "type": "text", + "content": "(i) Gradient Loss" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 288, + 287, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 358 + ], + "type": "text", + "content": "structure to speed up the search of the nearest surface point for queries " + }, + { + "bbox": [ + 46, + 288, + 287, + 358 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 288, + 287, + 358 + ], + "type": "text", + "content": ". Since we need to extensively conduct trilinear interpolation on signed distances for queries " + }, + { + "bbox": [ + 46, + 288, + 287, + 358 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 288, + 287, + 358 + ], + "type": "text", + "content": ", we use uniform grids for fast nearest grid search for queries. We did not use an Octree to organize grids, due to its inefficient and inaccurate grid search for queries." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 358, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 453 + ], + "type": "text", + "content": "Geometric Initialization. We initialize the signed distances " + }, + { + "bbox": [ + 46, + 358, + 287, + 453 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 358, + 287, + 453 + ], + "type": "text", + "content": " on grid vertices which represent a sphere in the field. The initialization can speed up the convergence during optimization. Our preliminary results show that a small sphere with a radius of one or two grids works well. Since enlarging a small sphere to the target shape can calculate the continuous distance loss over smaller regions than shrinking a large sphere." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": "Signed Distance Inference. For a query " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": ", we infer its signed distance " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "f(q)" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": " through pulling it onto its nearest surface point in the point cloud " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": ". As shown in Fig. 1 (e), we first obtain " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "f(q)" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": " as the interpolation of the signed distances " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": " on the nearest eight vertices of " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": ". We implement the trilinear interpolation by solving " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = [a_0,..,a_7]" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": " in a linear regression system on the eight vertices " + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 455, + 287, + 538 + ], + "type": "text", + "content": " below," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 550, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 287, + 573 + ], + "type": "interline_equation", + "content": "f (q) = a _ {0} + a _ {1} x + a _ {2} y + a _ {3} z + a _ {4} x y + a _ {5} x z + a _ {6} y z + a _ {7} x y z, \\tag {3}", + "image_path": "9cf401e837f70aead326ea0a5f7488dbf6fbfc409e22b79a9fccd92e4fbc25cc.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "inline_equation", + "content": "[x,y,z]" + }, + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "text", + "content": " is the coordinate of " + }, + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "text", + "content": " and the eight constraints are " + }, + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\{f(v_i) = d_i\\}" + }, + { + "bbox": [ + 46, + 574, + 286, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "text", + "content": "With the predicted signed distance " + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "inline_equation", + "content": "f(q)" + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "text", + "content": " and the gradient " + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "inline_equation", + "content": "\\nabla f(q) = \\partial f / \\partial q" + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "text", + "content": ", we pull a query " + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "text", + "content": " onto the surface and obtain a pulled query " + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "inline_equation", + "content": "p = q - f(q) \\times \\nabla f(q) / ||\\nabla f(q)||_2" + }, + { + "bbox": [ + 46, + 597, + 287, + 645 + ], + "type": "text", + "content": " in Fig. 1 (f)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "text", + "content": "Our goal is to minimize the distance between the pulled query " + }, + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "text", + "content": " and the nearest surface point " + }, + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "inline_equation", + "content": "NN(q) \\in S" + }, + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "text", + "content": " of query " + }, + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 646, + 286, + 681 + ], + "type": "text", + "content": " below," + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 683, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 683, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 115, + 683, + 287, + 696 + ], + "type": "interline_equation", + "content": "L _ {P u l l} = \\left\\| p - N N (q) \\right\\| _ {2}. \\tag {4}", + "image_path": "b77b6a52a51c531365f5288e1304d0c1dd22b9135c64d127b7fbe5f499e85685.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "We adopt this pulling loss from [49]. Instead, we do not" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 288, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 323 + ], + "type": "text", + "content": "rely on neural network in the pulling procedure. We use the trilinear interpolation on signed distances " + }, + { + "bbox": [ + 304, + 288, + 545, + 323 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 288, + 545, + 323 + ], + "type": "text", + "content": " on grid vertices to obtain signed distances and gradients according to Eq. 3." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 325, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 408 + ], + "type": "text", + "content": "We sample queries " + }, + { + "bbox": [ + 304, + 325, + 545, + 408 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 325, + 545, + 408 + ], + "type": "text", + "content": " around surface points in each epoch. For each randomly sampled surface point, we sample queries using a gaussian distribution with the surface point as the center and a two-grid length variance. We build a KD-tree on surface points so that we can speed up the searching of the nearest neighbor from a large number of surface points for each query." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 410, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 469 + ], + "type": "text", + "content": "To speed up the optimization, we only calculate the pulling loss on grids near the surface. We regard the grids containing surface points as the base and find the union of their " + }, + { + "bbox": [ + 304, + 410, + 545, + 469 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 304, + 410, + 545, + 469 + ], + "type": "text", + "content": " nearest neighboring grids. These grids form a " + }, + { + "bbox": [ + 304, + 410, + 545, + 469 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 304, + 410, + 545, + 469 + ], + "type": "text", + "content": "-bandwidth near the surface in Fig. 1 (c)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": "Continuous Distance Fields. Although the pulling loss can infer correct signed distances at grid vertices, it fails to learn continuous signed distances across the field, especially near the surface. Fig. 2 (a) shows severe artifacts on the surface and gradients near the surface are messy and inconsistent, which are caused by the discontinuous signed distances across neighboring regions." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "text", + "content": "Neural network based methods [49, 51] do not have this issue, since the continuous character of neural networks makes the sudden change of signs on the same shape side hard. While our method uses signed distances " + }, + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "text", + "content": " on grid vertices to represent a discrete field, this makes that one " + }, + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "text", + "content": " can change separately without considering the change of neighboring " + }, + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 556, + 545, + 651 + ], + "type": "text", + "content": ", which leads to a poor continuousness in the field." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "To resolve this issue, we impose a total variation (TV) loss as a continuous constraint on our signed distances " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " on grid vertices. Our key idea is to make the change of one " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " affect the change of its neighboring " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "NN(d_{i})" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", which smoothes the local distance field and propagates local up" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18324" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 289, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 289, + 186 + ], + "type": "image", + "image_path": "14712135528e45347a8f1b81ce55d4d449a3f94d17b61fc1ce25ec57ec91b9eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 186, + 253, + 195 + ], + "lines": [ + { + "bbox": [ + 82, + 186, + 253, + 195 + ], + "spans": [ + { + "bbox": [ + 82, + 186, + 253, + 195 + ], + "type": "text", + "content": "Figure 2. The effect of losses on reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 196, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 243 + ], + "type": "text", + "content": "dates to the rest grids. Hence, as shown in a 2D case in Fig. 1 (g), we approximate one " + }, + { + "bbox": [ + 46, + 196, + 287, + 243 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 196, + 287, + 243 + ], + "type": "text", + "content": " to each one of its six neighbors by minimizing the difference between each two below," + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 252, + 287, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 252, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 64, + 252, + 287, + 278 + ], + "type": "interline_equation", + "content": "L _ {T V} = \\sqrt {\\sum_ {a = \\{x , y , z \\}} e _ {a + 1} \\left(d _ {i}\\right) ^ {2} + e _ {a - 1} \\left(d _ {i}\\right) ^ {2}}, \\tag {5}", + "image_path": "f6204836fcddd408f8944aba8dc319629fb10916856c2e31aae0996f4bbdc416.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "inline_equation", + "content": "e_{a + 1}(d_i)" + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "text", + "content": " is the difference between " + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "text", + "content": " and the signed distance along the " + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "text", + "content": " axis, i.e., " + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 46, + 287, + 287, + 346 + ], + "type": "text", + "content": " axis. Fig. 2 (b) shows that the continuous constraints can significantly improve the signed distance field in terms of continuousness and gradients." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 346, + 287, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 393 + ], + "type": "text", + "content": "One remaining issue is its computational complexity. Since we are targeting grids at a high resolution, such as " + }, + { + "bbox": [ + 47, + 346, + 287, + 393 + ], + "type": "inline_equation", + "content": "R = 512" + }, + { + "bbox": [ + 47, + 346, + 287, + 393 + ], + "type": "text", + "content": ", it is very time consuming to impose this constraint on all " + }, + { + "bbox": [ + 47, + 346, + 287, + 393 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 47, + 346, + 287, + 393 + ], + "type": "text", + "content": " across the field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 395, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 465 + ], + "type": "text", + "content": "To resolve this issue, we focus on grids near the surface. Similar to imposing the pulling loss on grids in the " + }, + { + "bbox": [ + 46, + 395, + 287, + 465 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 46, + 395, + 287, + 465 + ], + "type": "text", + "content": "-bandwidth near the surface, we impose our continuous constraints on the grids in the " + }, + { + "bbox": [ + 46, + 395, + 287, + 465 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 46, + 395, + 287, + 465 + ], + "type": "text", + "content": "-bandwidth near the surface, which significantly reduces the number of learnable parameters to optimize." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 466, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 536 + ], + "type": "text", + "content": "Signed Distance Supervision. For signed distances, one supervision we can directly use is the input point cloud " + }, + { + "bbox": [ + 46, + 466, + 287, + 536 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 466, + 287, + 536 + ], + "type": "text", + "content": ". It is a discrete surface indicating the zero level-set of the signed distance function. Hence, as introduced in Fig. 1 (h), we can constrain the signed distances at the points " + }, + { + "bbox": [ + 46, + 466, + 287, + 536 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 466, + 287, + 536 + ], + "type": "text", + "content": " on the surface below," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 539, + 287, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 539, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 122, + 539, + 287, + 553 + ], + "type": "interline_equation", + "content": "L _ {\\text {S u r f a c e}} = \\left\\| f (s) \\right\\| _ {2}. \\tag {6}", + "image_path": "310837e7083ffd4227fe26ecb38f96db00a6f24102c0dbe912f1fb6adaee2b84.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": "Although Fig. 2 (c) indicates that " + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "inline_equation", + "content": "L_{Surface}" + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": " slightly improves the signed distance field and leads to a little bit more continuous surface, it is helpful to determine the zero level set with other losses. For noises on " + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": ", our continuous constraint can relieve the impact of noises on this supervision. We will show this in experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "content": "Consistent Gradients. Fig. 2 (a) indicates that the pulling loss produces gradients that are messy and non-orthogonal to the surface. This leads to inaccurate signed distance inference during the pulling procedure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "We additionally introduce a constraint on gradients to improve the field near the surface. As shown in Fig. 1 (i), we aim to encourage the gradient at a query " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " to point to the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "same direction of the gradient at its nearest surface point " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "NN(q)" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": ". We use a cosine distance below to achieve more consistent gradients," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 110, + 545, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 110, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 348, + 110, + 545, + 125 + ], + "type": "interline_equation", + "content": "L _ {\\nabla} = 1 - \\cos (\\nabla f (q), \\nabla f (N N (q))). \\tag {7}", + "image_path": "9aa5781313307aa09a20c25270f1357c20536da2cdc50f4b6bfa2682dd46169f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 134, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 134, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 134, + 545, + 192 + ], + "type": "inline_equation", + "content": "L_{\\nabla}" + }, + { + "bbox": [ + 305, + 134, + 545, + 192 + ], + "type": "text", + "content": " constrains the gradients in the field and corrects inaccurate gradients. Fig. 2 (d) shows that more consistent gradients can improve the surface to be more continuous using small planes on the surface which however the surface not smooth at all." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 194, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 194, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 545, + 217 + ], + "type": "text", + "content": "Loss Function. We learn signed distances on grid vertices " + }, + { + "bbox": [ + 306, + 194, + 545, + 217 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 194, + 545, + 217 + ], + "type": "text", + "content": " by minimizing the loss function below," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 338, + 220, + 545, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 220, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 338, + 220, + 545, + 233 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {P u l l}} + \\alpha L _ {T V} + \\beta L _ {\\text {S u r f a c e}} + \\gamma L _ {\\nabla}. \\tag {8}", + "image_path": "9af21fa259fda62e1553dc9944d5a5ec12a36491f0b13555de2a273354ec8ccd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "text", + "content": " are balance weights to make each term contribute equally. Fig. 3 shows the signed distance field on a slide of " + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 238, + 545, + 321 + ], + "type": "text", + "content": " during optimization. Our optimization starts from an initial sphere and only optimizes variables in the bandwidth. We skip the grids that are not optimized after the initialization during the marching cubes for surface reconstruction. Please watch our video for more details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 334, + 455, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 455, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 455, + 347 + ], + "type": "text", + "content": "4. Experiments and Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 354, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 388 + ], + "type": "text", + "content": "We evaluate GridPull by numerical and visual comparisons with the latest methods on synthetic and real datasets in surface reconstruction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 485 + ], + "type": "text", + "content": "Datasets and Metrics. We use benchmarks for shapes and scenes in evaluations. For shapes, we conduct evaluations on five datasets including a subset of ShapeNet [1], FAMOUS [17], Thingi10k [103], Surface Reconstruction Benchmark (SRB) [89] and D-FAUST [5]. For scenes, we also report comparisons on five datasets including 3DScene [104], SceneNet [27], 3DFRONT [18], Matterport [9] and KITTI [20]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 486, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 592 + ], + "type": "text", + "content": "We use L2 Chamfer distance " + }, + { + "bbox": [ + 304, + 486, + 545, + 592 + ], + "type": "inline_equation", + "content": "(CD_{L2})" + }, + { + "bbox": [ + 304, + 486, + 545, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 486, + 545, + 592 + ], + "type": "inline_equation", + "content": "CD_{L1}" + }, + { + "bbox": [ + 304, + 486, + 545, + 592 + ], + "type": "text", + "content": " and Hausdorff distance(HD) to measure the error between the reconstructed surface and the ground truth. Moreover, we use normal consistency (NC) and F-score to evaluate the accuracy of normal on the reconstructed surface. We also report our time and storage complexity to highlight our advantage towards scalability. We also use the intersection over union (IoU) to measure the reconstruction for fair comparison with the latest methods." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "content": "Details. We use " + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "inline_equation", + "content": "R = 256" + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "content": " to infer a distance field. We calculate the pulling loss on grids in a bandwidth with " + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "inline_equation", + "content": "M_{1} = 3" + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "content": ". We impose the continuous constraint on grids in a bandwidth with " + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "inline_equation", + "content": "M_{2} = 14" + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "content": " for shapes and " + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "inline_equation", + "content": "M_{2} = 25" + }, + { + "bbox": [ + 304, + 594, + 545, + 676 + ], + "type": "text", + "content": " for scenes, since scenes contain open surfaces with larger empty spaces. We run the marching cubes for surface reconstruction at a resolution of 256." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We use the Adam optimizer with an initial learning rate of 1.0. We decrease the learning rate with a decay rate of 0.3 every 400 iterations. We use 50,000 queries in each" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18325" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 60, + 132, + 123 + ], + "blocks": [ + { + "bbox": [ + 50, + 60, + 132, + 123 + ], + "lines": [ + { + "bbox": [ + 50, + 60, + 132, + 123 + ], + "spans": [ + { + "bbox": [ + 50, + 60, + 132, + 123 + ], + "type": "image", + "image_path": "cb201be88e6afbc6a4dc5e742d25ee257e3bb221f91021bf5fbfb201996b9b34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 123, + 111, + 131 + ], + "lines": [ + { + "bbox": [ + 69, + 123, + 111, + 131 + ], + "spans": [ + { + "bbox": [ + 69, + 123, + 111, + 131 + ], + "type": "text", + "content": "0th Iteration" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 133, + 61, + 214, + 123 + ], + "blocks": [ + { + "bbox": [ + 133, + 61, + 214, + 123 + ], + "lines": [ + { + "bbox": [ + 133, + 61, + 214, + 123 + ], + "spans": [ + { + "bbox": [ + 133, + 61, + 214, + 123 + ], + "type": "image", + "image_path": "27abef32dbbe34dec1596d0ba6aab2d3cf3a43d5ee0f85709d5f62edc7bd2fc4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 123, + 197, + 131 + ], + "lines": [ + { + "bbox": [ + 152, + 123, + 197, + 131 + ], + "spans": [ + { + "bbox": [ + 152, + 123, + 197, + 131 + ], + "type": "text", + "content": "50th Iteration" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 61, + 295, + 123 + ], + "blocks": [ + { + "bbox": [ + 215, + 61, + 295, + 123 + ], + "lines": [ + { + "bbox": [ + 215, + 61, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 215, + 61, + 295, + 123 + ], + "type": "image", + "image_path": "d676db783771bad12e0d9d115cbaa9ba3d1d6b26c5b8c84b62d9d2d23d0a3276.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 123, + 280, + 131 + ], + "lines": [ + { + "bbox": [ + 231, + 123, + 280, + 131 + ], + "spans": [ + { + "bbox": [ + 231, + 123, + 280, + 131 + ], + "type": "text", + "content": "100th Iteration" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 61, + 377, + 123 + ], + "blocks": [ + { + "bbox": [ + 296, + 61, + 377, + 123 + ], + "lines": [ + { + "bbox": [ + 296, + 61, + 377, + 123 + ], + "spans": [ + { + "bbox": [ + 296, + 61, + 377, + 123 + ], + "type": "image", + "image_path": "c4bad84de10b76dcd23cd53ac6bfb4617f79a26cd074e09af20a9a05e34c9fc7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 123, + 362, + 131 + ], + "lines": [ + { + "bbox": [ + 313, + 123, + 362, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 362, + 131 + ], + "type": "text", + "content": "150th Iteration" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 61, + 459, + 123 + ], + "blocks": [ + { + "bbox": [ + 378, + 61, + 459, + 123 + ], + "lines": [ + { + "bbox": [ + 378, + 61, + 459, + 123 + ], + "spans": [ + { + "bbox": [ + 378, + 61, + 459, + 123 + ], + "type": "image", + "image_path": "30b683a5ee4e3421fd1acaeaf3ac760458a43704d20ce825d1e0b9bd6d4c4e62.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 123, + 445, + 131 + ], + "lines": [ + { + "bbox": [ + 395, + 123, + 445, + 131 + ], + "spans": [ + { + "bbox": [ + 395, + 123, + 445, + 131 + ], + "type": "text", + "content": "200th Iteration" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 460, + 61, + 542, + 123 + ], + "blocks": [ + { + "bbox": [ + 460, + 61, + 542, + 123 + ], + "lines": [ + { + "bbox": [ + 460, + 61, + 542, + 123 + ], + "spans": [ + { + "bbox": [ + 460, + 61, + 542, + 123 + ], + "type": "image", + "image_path": "10b3e5fbc48694f67a753a9571d9727bc46adc17040db454ae645aec55ca26cd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 123, + 528, + 131 + ], + "lines": [ + { + "bbox": [ + 475, + 123, + 528, + 131 + ], + "spans": [ + { + "bbox": [ + 475, + 123, + 528, + 131 + ], + "type": "text", + "content": "1000th Iteration" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 49, + 142, + 287, + 308 + ], + "blocks": [ + { + "bbox": [ + 189, + 131, + 402, + 140 + ], + "lines": [ + { + "bbox": [ + 189, + 131, + 402, + 140 + ], + "spans": [ + { + "bbox": [ + 189, + 131, + 402, + 140 + ], + "type": "text", + "content": "Figure 3. Visualization of signed distances in optimization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 142, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 49, + 142, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 142, + 287, + 308 + ], + "type": "image", + "image_path": "31cff11eaad808fa8900f90bf875e00b8b4febf4cd424e84e47c335acb46bf1c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 311, + 264, + 322 + ], + "lines": [ + { + "bbox": [ + 69, + 311, + 264, + 322 + ], + "spans": [ + { + "bbox": [ + 69, + 311, + 264, + 322 + ], + "type": "text", + "content": "Figure 4. Visual comparison under ShapetNet dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "text", + "content": "iteration, and run 1600 iterations for each point cloud during optimization. We set the balance weights " + }, + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\gamma = 0.005" + }, + { + "bbox": [ + 47, + 331, + 287, + 367 + ], + "type": "text", + "content": " for equal contribution from the three terms." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 373, + 232, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 232, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 232, + 387 + ], + "type": "text", + "content": "4.1. Surface Reconstruction for Shapes" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": "ShapeNet. Following predictive context prior (PCP) [51] and NeuralPull (NP) [49], we report our results on a subset in ShapeNet in terms of " + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "inline_equation", + "content": "CD_{L2}" + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": " in Tab. 1, Normal Consistency (NC) in Tab. 2, and F-Score with thresholds of 0.002 and 0.004 in Tab. 3 and Tab. 4, which show our superiority over PCP and NP in numerical comparisons over all classes, even though PCP employs a local shape prior learned from a large scale dataset. The visual comparison in Fig. 4 shows that we can reveal more accurate and detailed geometry." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 85, + 502, + 251, + 614 + ], + "blocks": [ + { + "bbox": [ + 85, + 502, + 251, + 614 + ], + "lines": [ + { + "bbox": [ + 85, + 502, + 251, + 614 + ], + "spans": [ + { + "bbox": [ + 85, + 502, + 251, + 614 + ], + "type": "table", + "html": "
ClassNP [49]PCP [51]Ours
Display0.0390.00870.0082
Lamp0.0800.03800.0347
Airplane0.0080.00650.0007
Cabinet0.0260.01530.0112
Vessel0.0220.00790.0033
Table0.0600.01310.0052
Chair0.0540.01100.0043
Sofa0.0120.00860.0015
Mean0.0380.01360.0086
", + "image_path": "5f62ea01f2a104405f9ccdf3b8e8689c1cc757dbce941f38598de672fcbe0d73.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "FAMOUS. We follow the experimental setting in PCP [51] and NP [49] to evaluate GridPull on FAMOUS. We compare GridPull with the latest methods with priors including PCP, GenSDF, FGC and no priors including NP and IGR. GridPull outperforms these methods in terms of " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "CD_{L2}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " and running time. Although FGC is fast, it does not reconstruct" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 343, + 147, + 509, + 259 + ], + "blocks": [ + { + "bbox": [ + 47, + 616, + 287, + 638 + ], + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 638 + ], + "type": "text", + "content": "Table 1. Reconstruction accuracy under ShapeNet in terms of " + }, + { + "bbox": [ + 47, + 616, + 287, + 638 + ], + "type": "inline_equation", + "content": "CD_{L2} \\times 100" + }, + { + "bbox": [ + 47, + 616, + 287, + 638 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 147, + 509, + 259 + ], + "lines": [ + { + "bbox": [ + 343, + 147, + 509, + 259 + ], + "spans": [ + { + "bbox": [ + 343, + 147, + 509, + 259 + ], + "type": "table", + "html": "
ClassNP [49]PCP [51]Ours
Display0.9640.97750.9847
Lamp0.9300.94500.9693
Airplane0.9470.94900.9614
Cabinet0.9300.96000.9689
Vessel0.9410.95460.9667
Table0.9080.95950.9755
Chair0.9370.95800.9733
Sofa0.9510.96800.9792
Mean0.9390.95900.9723
", + "image_path": "ccc7c3f3bc66a7920fcad613e6fedf83a1aa949502913cf7cd2d761df47d9fc5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 343, + 275, + 509, + 388 + ], + "blocks": [ + { + "bbox": [ + 306, + 262, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 306, + 262, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 545, + 273 + ], + "type": "text", + "content": "Table 2. Reconstruction accuracy under ShapeNet in terms of NC." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 275, + 509, + 388 + ], + "lines": [ + { + "bbox": [ + 343, + 275, + 509, + 388 + ], + "spans": [ + { + "bbox": [ + 343, + 275, + 509, + 388 + ], + "type": "table", + "html": "
ClassNP [49]PCP [51]Ours
Display0.9890.99390.9963
Lamp0.8910.93820.9455
Airplane0.9960.99420.9976
Cabinet0.9800.98880.9901
Vessel0.9850.99350.9956
Table0.9220.99690.9977
Chair0.9540.99700.9979
Sofa0.9680.99430.9974
Mean0.9610.98710.9896
", + "image_path": "408516d0d80f45dd19d03e0f22eea0ea81d36d76006438cc3050e9189b6b62f1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 343, + 413, + 509, + 525 + ], + "blocks": [ + { + "bbox": [ + 306, + 390, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 306, + 390, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 545, + 411 + ], + "type": "text", + "content": "Table 3. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.002." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 413, + 509, + 525 + ], + "lines": [ + { + "bbox": [ + 343, + 413, + 509, + 525 + ], + "spans": [ + { + "bbox": [ + 343, + 413, + 509, + 525 + ], + "type": "table", + "html": "
ClassNP [49]PCP [51]Ours
Display0.9910.99580.9963
Lamp0.9240.94020.9538
Airplane0.9970.99720.9989
Cabinet0.9890.99390.9946
Vessel0.9900.99580.9972
Table0.9730.99850.9990
Chair0.9690.99910.9990
Sofa0.9740.99870.9992
Mean0.9760.98990.9923
", + "image_path": "1116c82b1fc2f147b6d41dadbef3b1213645e344012cbe6a8787622dc5f5d651.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 554, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 545, + 625 + ], + "type": "text", + "content": "a watertight mesh but just a polygon soup with no normal, which contains almost no geometry details. Worse than that, it fails to reconstruct some point clouds in more than eight hours, hence we remove these shapes from its results. We highlight our superiority in visual comparison in Fig. 5. We can reveal details like hair and expressions more clearly." + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 331, + 635, + 520, + 705 + ], + "blocks": [ + { + "bbox": [ + 306, + 528, + 545, + 548 + ], + "lines": [ + { + "bbox": [ + 306, + 528, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 545, + 548 + ], + "type": "text", + "content": "Table 4. Reconstruction accuracy under ShapeNet in terms of F-Score with a threshold of 0.004." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 635, + 520, + 705 + ], + "lines": [ + { + "bbox": [ + 331, + 635, + 520, + 705 + ], + "spans": [ + { + "bbox": [ + 331, + 635, + 520, + 705 + ], + "type": "table", + "html": "
MethodCDL2×100Min timeMean time
IGR [24]1.650402s407s
GenSDF [15]0.668230s232s
NP [49]0.220593s605s
FGC [97]0.05510sTimeout
PCP [51]0.0443416s3534s
Ours0.040198s201s
", + "image_path": "17e782dffee1c3a9698ea9a9ad7ab145e3d4597f7a5201f7482d2cf69b93de8c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 332, + 708, + 518, + 718 + ], + "lines": [ + { + "bbox": [ + 332, + 708, + 518, + 718 + ], + "spans": [ + { + "bbox": [ + 332, + 708, + 518, + 718 + ], + "type": "text", + "content": "Table 5. Reconstruction accuracy under FAMOUS." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18326" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 286, + 242 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 286, + 242 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 286, + 242 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 286, + 242 + ], + "type": "image", + "image_path": "f86dd0e21b1991b0f68701504a745ceb960bcd52bb17aa918ce78f4c79169300.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 243, + 264, + 253 + ], + "lines": [ + { + "bbox": [ + 69, + 243, + 264, + 253 + ], + "spans": [ + { + "bbox": [ + 69, + 243, + 264, + 253 + ], + "type": "text", + "content": "Figure 5. Visual comparison under FAMOUS dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 255, + 287, + 407 + ], + "blocks": [ + { + "bbox": [ + 48, + 255, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 407 + ], + "type": "image", + "image_path": "8d1af343e0df8fb9ceaec63400ed409efda84caa525fc04d7c9419a5cd3d9be3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 409, + 254, + 419 + ], + "lines": [ + { + "bbox": [ + 79, + 409, + 254, + 419 + ], + "spans": [ + { + "bbox": [ + 79, + 409, + 254, + 419 + ], + "type": "text", + "content": "Figure 6. Visual comparison under SRB dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 422, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 602 + ], + "type": "text", + "content": "SRB. We follow the experimental setting in VisCo [67] to evaluate GridPull on real scans in SRB. We report the comparison with the latest overfitting based methods in terms of " + }, + { + "bbox": [ + 47, + 422, + 287, + 602 + ], + "type": "inline_equation", + "content": "CD_{L1}" + }, + { + "bbox": [ + 47, + 422, + 287, + 602 + ], + "type": "text", + "content": " in Tab. 6. Meanwhile, we also report their one-sided distances between the reconstructed mesh and the input noisy point cloud. The numerical comparison shows that our method significantly outperforms the latest. Moreover, we also report our results of learning unsigned distances in Tab. 7. We replace our pulling loss with the pulling loss introduced for unsigned distances in CAP-UDF [102], and remove the gradient loss due to the nondifferentiable character of UDF on surface. Visual comparison in Fig. 6 shows that our method can reconstruct surfaces that are more compact than the methods learning SDF like SIREN or more smooth than the methods learning UDF like CAP-UDF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 604, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 287, + 663 + ], + "type": "text", + "content": "Thingi10K. We follow the experimental setting in SAP [65] to evaluate GridPull in Thingi10K. Tab. 8 indicates our superiority over the overfitting based methods in the numerical comparison. Visual comparisons in Fig. 7 show our more accurate surfaces with complex details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "D-FAUST. We also follow the experimental setting in S-AP [65] to evaluate GridPull in D-FAUST. Tab. 8 shows that we outperform the overfitting based methods with more surface details. As shown in Fig. 7, we recover more accurate" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 84, + 545, + 236 + ], + "blocks": [ + { + "bbox": [ + 309, + 84, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 309, + 84, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 309, + 84, + 545, + 236 + ], + "type": "table", + "html": "
Poisson [37]IGR [24]SIREN [71]VisCo [67]SAP [65]Ours
Anchor\\(CD_{L1}\\)0.600.220.320.210.120.093
\\(HD\\)14.894.718.193.002.381.804
\\(d_{\\vec{C}}\\)0.600.120.100.150.080.066
\\(d_{\\vec{H}}\\)14.891.322.431.070.830.460
Daratech\\(CD_{L1}\\)0.440.250.210.260.070.062
\\(HD\\)7.244.014.304.060.870.648
\\(d_{\\vec{C}}\\)0.440.080.090.140.040.039
\\(d_{\\vec{H}}\\)7.241.591.771.760.410.293
DC\\(CD_{L1}\\)0.270.170.150.150.070.066
\\(HD\\)3.102.222.182.221.171.103
\\(d_{\\vec{C}}\\)0.270.090.060.090.040.036
\\(d_{\\vec{H}}\\)3.102.612.762.760.530.539
Gargoyle\\(CD_{L1}\\)0.260.160.170.170.070.063
\\(HD\\)6.803.524.644.401.491.129
\\(d_{\\vec{C}}\\)0.260.060.080.110.050.045
\\(d_{\\vec{H}}\\)6.800.810.910.960.780.700
Lord Quas\\(CD_{L1}\\)0.200.120.170.120.050.047
\\(HD\\)4.611.170.821.060.980.569
\\(d_{\\vec{C}}\\)0.200.070.120.070.040.031
\\(d_{\\vec{H}}\\)4.610.980.760.640.510.370
", + "image_path": "c9e5fe75be731eacddc7b593e7f699d2664e4b0df1449b02f5e1f2f0d8971220.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 343, + 251, + 509, + 279 + ], + "blocks": [ + { + "bbox": [ + 343, + 238, + 508, + 248 + ], + "lines": [ + { + "bbox": [ + 343, + 238, + 508, + 248 + ], + "spans": [ + { + "bbox": [ + 343, + 238, + 508, + 248 + ], + "type": "text", + "content": "Table 6. Reconstruction accuracy under SRB." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 251, + 509, + 279 + ], + "lines": [ + { + "bbox": [ + 343, + 251, + 509, + 279 + ], + "spans": [ + { + "bbox": [ + 343, + 251, + 509, + 279 + ], + "type": "table", + "html": "
NDF [14]CAP-UDF [102]Ours(UDF)
CDL10.2380.0730.070
F-Score68.684.585.1
", + "image_path": "527a717b9a7266c9b127f58b23afd7748d9bcdc8898f8e3569493b8f91096133.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 294, + 543, + 499 + ], + "blocks": [ + { + "bbox": [ + 319, + 281, + 531, + 291 + ], + "lines": [ + { + "bbox": [ + 319, + 281, + 531, + 291 + ], + "spans": [ + { + "bbox": [ + 319, + 281, + 531, + 291 + ], + "type": "text", + "content": "Table 7. Accuracy of reconstruction with UDF under SRB." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 294, + 543, + 499 + ], + "lines": [ + { + "bbox": [ + 307, + 294, + 543, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 543, + 499 + ], + "type": "image", + "image_path": "6998a1e2a06a36e1005cdeb696ed9fdb3464e3220e086208a519e37dda5cdf85.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 500, + 528, + 511 + ], + "lines": [ + { + "bbox": [ + 306, + 500, + 528, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 528, + 511 + ], + "type": "text", + "content": "Figure 7. Visual comparison under Thingi10k and D-FAUST." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 343, + 517, + 509, + 581 + ], + "blocks": [ + { + "bbox": [ + 343, + 517, + 509, + 581 + ], + "lines": [ + { + "bbox": [ + 343, + 517, + 509, + 581 + ], + "spans": [ + { + "bbox": [ + 343, + 517, + 509, + 581 + ], + "type": "table", + "html": "
IGR [24]SAP [65]Ours
Thing10kCDL10.4400.0540.051
F-Score0.5050.9400.948
NC0.6920.9470.965
DFAUSTCDL10.2350.0430.015
F-Score0.8050.9660.975
NC0.9110.9590.978
", + "image_path": "f2427a3e6633596af31a752aaee7cea901b85d38db8749942df94a42506bb9fa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "lines": [ + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "type": "text", + "content": "Table 8. Reconstruction accuracy under Thingi10k and D-FAUST in terms of " + }, + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "type": "inline_equation", + "content": "CD_{L1}" + }, + { + "bbox": [ + 306, + 582, + 545, + 604 + ], + "type": "text", + "content": " and F-Score with a threshold of 0.01." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 616, + 525, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 525, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 525, + 628 + ], + "type": "text", + "content": "human bodies and more detailed expressions on faces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 635, + 488, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 488, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 488, + 647 + ], + "type": "text", + "content": "4.2. Surface Reconstruction for Scenes" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": "3DScene. We follow PCP [51] to report " + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "CD_{L1}" + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "CD_{L2}" + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": " and Normal Consistency(NC) for evaluation. We report the comparisons with the latest methods in Tab. 9. We outperform both kinds of methods with learned priors such as ConvOcc [73] and PCP [51] and overfitting based NP [49]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18327" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 63, + 149, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 63, + 149, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 63, + 149, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 63, + 149, + 137 + ], + "type": "image", + "image_path": "99d914b16119c2ec4cd76a1aa3e4c74e4f2db9ca5bbcc69f30384563f8658bb7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 137, + 111, + 145 + ], + "lines": [ + { + "bbox": [ + 78, + 137, + 111, + 145 + ], + "spans": [ + { + "bbox": [ + 78, + 137, + 111, + 145 + ], + "type": "text", + "content": "ConvOcc" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 151, + 64, + 249, + 135 + ], + "blocks": [ + { + "bbox": [ + 151, + 64, + 249, + 135 + ], + "lines": [ + { + "bbox": [ + 151, + 64, + 249, + 135 + ], + "spans": [ + { + "bbox": [ + 151, + 64, + 249, + 135 + ], + "type": "image", + "image_path": "b0b03ddc233ebc125ac799649dd24c3a48b0328a941ac603036c200377755e8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 137, + 202, + 145 + ], + "lines": [ + { + "bbox": [ + 189, + 137, + 202, + 145 + ], + "spans": [ + { + "bbox": [ + 189, + 137, + 202, + 145 + ], + "type": "text", + "content": "NP" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 64, + 347, + 133 + ], + "blocks": [ + { + "bbox": [ + 250, + 64, + 347, + 133 + ], + "lines": [ + { + "bbox": [ + 250, + 64, + 347, + 133 + ], + "spans": [ + { + "bbox": [ + 250, + 64, + 347, + 133 + ], + "type": "image", + "image_path": "726a1d1207ba43fc36d35ceb2ba62a5618f64596babfaea182e970ba9ea5599b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 137, + 303, + 144 + ], + "lines": [ + { + "bbox": [ + 287, + 137, + 303, + 144 + ], + "spans": [ + { + "bbox": [ + 287, + 137, + 303, + 144 + ], + "type": "text", + "content": "PCP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 64, + 442, + 133 + ], + "blocks": [ + { + "bbox": [ + 347, + 64, + 442, + 133 + ], + "lines": [ + { + "bbox": [ + 347, + 64, + 442, + 133 + ], + "spans": [ + { + "bbox": [ + 347, + 64, + 442, + 133 + ], + "type": "image", + "image_path": "e15184067e4c73650994685d61f50203310af44e75d4e2cb6b0f48bd2193374a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 137, + 405, + 144 + ], + "lines": [ + { + "bbox": [ + 388, + 137, + 405, + 144 + ], + "spans": [ + { + "bbox": [ + 388, + 137, + 405, + 144 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 443, + 64, + 544, + 133 + ], + "blocks": [ + { + "bbox": [ + 443, + 64, + 544, + 133 + ], + "lines": [ + { + "bbox": [ + 443, + 64, + 544, + 133 + ], + "spans": [ + { + "bbox": [ + 443, + 64, + 544, + 133 + ], + "type": "image", + "image_path": "7576ebcaebb72a25bfc8743731bb47ab16c91e14a0c4f2a0a07a8eb8e2103048.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 137, + 504, + 144 + ], + "lines": [ + { + "bbox": [ + 492, + 137, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 492, + 137, + 504, + 144 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 163, + 127, + 258 + ], + "blocks": [ + { + "bbox": [ + 49, + 163, + 127, + 258 + ], + "lines": [ + { + "bbox": [ + 49, + 163, + 127, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 127, + 258 + ], + "type": "image", + "image_path": "2b5004b4b02b3108d2f182b8f5c2c0bed111468b24e76b526b3cd624b244a534.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 258, + 263, + 274 + ], + "lines": [ + { + "bbox": [ + 69, + 258, + 263, + 274 + ], + "spans": [ + { + "bbox": [ + 69, + 258, + 263, + 274 + ], + "type": "text", + "content": "NP PCP Ours Figure 9. Visual comparison under SceneNet dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 128, + 163, + 207, + 258 + ], + "blocks": [ + { + "bbox": [ + 200, + 144, + 392, + 156 + ], + "lines": [ + { + "bbox": [ + 200, + 144, + 392, + 156 + ], + "spans": [ + { + "bbox": [ + 200, + 144, + 392, + 156 + ], + "type": "text", + "content": "Figure 8. Visual comparison under 3DScene dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 128, + 163, + 207, + 258 + ], + "lines": [ + { + "bbox": [ + 128, + 163, + 207, + 258 + ], + "spans": [ + { + "bbox": [ + 128, + 163, + 207, + 258 + ], + "type": "image", + "image_path": "692d5de9e25cfe3f42a3d65d054294fe01c8cd70f846d58003ccdceed3682c3e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 207, + 163, + 287, + 258 + ], + "blocks": [ + { + "bbox": [ + 207, + 163, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 207, + 163, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 207, + 163, + 287, + 258 + ], + "type": "image", + "image_path": "d9fd1564499f529be4c258b65137ff4e5878e41de19f8401ed56c4ad6ec1b688.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 288, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 287, + 324 + ], + "type": "text", + "content": "in all scenes. The visual comparisons 8 shows that our method can work well with real scans on scenes and reveal more geometry details on surfaces." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 50, + 327, + 287, + 475 + ], + "blocks": [ + { + "bbox": [ + 50, + 327, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 50, + 327, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 50, + 327, + 287, + 475 + ], + "type": "table", + "html": "
ConvOcc [73]NP [49]PCP[51]Ours
BurghersCDL2× 10026.691.760.2670.246
CDL10.0770.0100.0080.008
NC0.8650.8830.9140.926
LoungeCDL2× 1008.6839.710.0610.055
CDL10.0420.0590.0060.005
NC0.8570.8570.9280.922
CopyroomCDL2× 10010.990.510.0760.069
CDL10.0450.0110.0070.006
NC0.8480.8840.9180.929
StonewallCDL2× 10019.120.0630.0610.058
CDL10.0660.0070.00650.006
NC0.8660.8680.8880.893
TotepoleCDL2× 1001.160.190.100.093
CDL10.0160.0100.0080.007
NC0.9250.7650.7840.847
", + "image_path": "e2a939aabec28703e4f3de1ce1269425db87d5b0e10f8bf471262a20e0083130.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "content": "SceneNet. Following the experimental setting in PCP [51], we report our " + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "inline_equation", + "content": "CD_{L1}" + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "content": ", NC, and F-score in Tab. 10. The comparison indicates that we outperform these methods by producing more accurate and smooth surfaces. This is also verified by our visual comparison in Fig. 9, where the latest methods reveal either no or inaccurate indoor geometry." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "text", + "content": "3D-FRONT. We follow the experimental setting in NeuralPoisson [16] to report " + }, + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "inline_equation", + "content": "CD_{L1}" + }, + { + "bbox": [ + 46, + 569, + 287, + 677 + ], + "type": "text", + "content": ", IoU and L2 distance over the voxelization of meshes. IoU and L2 distances are calculated as the error between the reconstructed surfaces and the ground truth over the voxelization at a resolution of 128. The comparison in Tab. 11 shows that our results are more accurate than the latest method. Our visual comparison with SIREN in Fig. 10 shows that GridPull can infer an accurate zero level set which leads to very sharp corners." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Matterport. We follow the experimental setting in NGS [30], and report the comparison in terms of " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "CD_{L1}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ", NC, and F-score in Tab. 12. We achieve the best perfor" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 331, + 163, + 522, + 350 + ], + "blocks": [ + { + "bbox": [ + 75, + 479, + 258, + 491 + ], + "lines": [ + { + "bbox": [ + 75, + 479, + 258, + 491 + ], + "spans": [ + { + "bbox": [ + 75, + 479, + 258, + 491 + ], + "type": "text", + "content": "Table 9. Reconstruction accuracy under 3DScene." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 163, + 522, + 350 + ], + "lines": [ + { + "bbox": [ + 331, + 163, + 522, + 350 + ], + "spans": [ + { + "bbox": [ + 331, + 163, + 522, + 350 + ], + "type": "table", + "html": "
NP [49]PCP [51]Ours
LivingroomCDL10.0880.0170.015
NC0.8810.9330.922
FScore0.8010.9660.953
BathroomCDL10.0360.0160.013
NC0.9120.9450.952
FScore0.8600.9770.974
BedroomCDL10.0340.0140.014
NC0.9050.9480.951
FScore0.8760.9800.986
KitchenCDL10.0490.0150.014
NC0.9000.9450.952
FScore0.8250.9760.984
OfficeCDL10.0620.0240.022
NC0.8790.9190.931
FScore0.7299250.940
MeanCDL10.0540.0170.0156
NC0.8950.9380.9416
FScore0.8189650.9674
", + "image_path": "32d9bf6b2ae00dc36dae035779ce9886e9cbb805a338c265cfe2a1fac059fd7e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 306, + 376, + 545, + 519 + ], + "blocks": [ + { + "bbox": [ + 306, + 376, + 545, + 519 + ], + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 519 + ], + "type": "image", + "image_path": "8b0fda8a33aa1e8759e1975383b9e53745aafc09b759d40b8e63eddc7d4f7bff.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 521, + 521, + 532 + ], + "lines": [ + { + "bbox": [ + 329, + 521, + 521, + 532 + ], + "spans": [ + { + "bbox": [ + 329, + 521, + 521, + 532 + ], + "type": "text", + "content": "Figure 10. Visual comparison under 3Dfront dataset." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 354, + 541, + 497, + 579 + ], + "blocks": [ + { + "bbox": [ + 306, + 353, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 306, + 353, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 545, + 373 + ], + "type": "text", + "content": "Table 10. Reconstruction accuracy under SceneNet. FScore is with a threshold of 0.025." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 541, + 497, + 579 + ], + "lines": [ + { + "bbox": [ + 354, + 541, + 497, + 579 + ], + "spans": [ + { + "bbox": [ + 354, + 541, + 497, + 579 + ], + "type": "table", + "html": "
MethodCDIOUL2
SIREN [71]0.01830.5242.373
NeuralPoisson [16]0.01610.5452.073
Ours0.01410.5621.984
", + "image_path": "7edd1a49f48a1b7e3c7b1168f31f95374247e966a974c32831028176a1e02928.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 329, + 582, + 522, + 592 + ], + "lines": [ + { + "bbox": [ + 329, + 582, + 522, + 592 + ], + "spans": [ + { + "bbox": [ + 329, + 582, + 522, + 592 + ], + "type": "text", + "content": "Table 11. Reconstruction accuracy under 3DFRONT." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": "mance in terms of accuracy. NGS [30] has a pretty fast inference but it costs about four to five days to learn priors from the training set in Matterport. Visual comparison in Fig. 11 shows that we can reconstruct more accurate surfaces while NGS is poorly generalized to unseen scenes." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "KITTI. We evaluate GridPull on a large scale real scan from KITTI odometry (Sequence00, frame 3000 to 4000) which contains about 13.8 million points. The comparison in Fig. 12 shows that we can reconstruct a more accurate" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18328" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 547, + 258 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 258 + ], + "type": "image", + "image_path": "47e2860efde063c793afb59182941a65fc86b588c467ed0d4cd65644d4d94b02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 259, + 402, + 270 + ], + "lines": [ + { + "bbox": [ + 189, + 259, + 402, + 270 + ], + "spans": [ + { + "bbox": [ + 189, + 259, + 402, + 270 + ], + "type": "text", + "content": "Figure 11. Visual comparison under Matterport3D dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 280, + 289, + 485 + ], + "blocks": [ + { + "bbox": [ + 48, + 280, + 289, + 485 + ], + "lines": [ + { + "bbox": [ + 48, + 280, + 289, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 289, + 485 + ], + "type": "image", + "image_path": "f4c7c19071dc506bff46b81312b521853ec3323365f932e776ef44a583d0104c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 490, + 257, + 501 + ], + "lines": [ + { + "bbox": [ + 77, + 490, + 257, + 501 + ], + "spans": [ + { + "bbox": [ + 77, + 490, + 257, + 501 + ], + "type": "text", + "content": "Figure 12. Visual comparison under Kitti dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 507, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 287, + 532 + ], + "type": "text", + "content": "surface in much less time (3810s) than the overfitting based method SAP (13005s) [65]." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 60, + 540, + 274, + 591 + ], + "blocks": [ + { + "bbox": [ + 60, + 540, + 274, + 591 + ], + "lines": [ + { + "bbox": [ + 60, + 540, + 274, + 591 + ], + "spans": [ + { + "bbox": [ + 60, + 540, + 274, + 591 + ], + "type": "table", + "html": "
MethodCDL1× 1000F-Score(0.002)NCTime
ConvOcc [73]6.2188.992.034s
SAP [65]4.1791.291.4643s
NGS [30]3.0497.395.72.1s
Ours2.9397.796.3223s
", + "image_path": "9cdfcaabdb9eeff4f8c266b35ae2266682e3f887a47b80e4c3ab3ca4f27c3411.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 610, + 147, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 147, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 147, + 622 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 629, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 652 + ], + "type": "text", + "content": "We conduct ablation studies on FAMOUS [17] to justify the effectiveness of modules in GridPull." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": "Loss. We justify the effectiveness of our loss in terms of " + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "inline_equation", + "content": "CD_{L2}" + }, + { + "bbox": [ + 47, + 653, + 287, + 712 + ], + "type": "text", + "content": " in Tab. 13. For the pulling loss, it is able to estimate a coarse shape but with discontinuous meshes and artifacts. The gradient loss can improve the continuousness on the surface. Using the TV loss, we obtained a more continu-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 276, + 545, + 338 + ], + "blocks": [ + { + "bbox": [ + 307, + 276, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 338 + ], + "type": "image", + "image_path": "c04bd206759f675ef696837d34dcb1a2739ae0ae0dbe511b19630e0eb00f766a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 338, + 488, + 349 + ], + "lines": [ + { + "bbox": [ + 362, + 338, + 488, + 349 + ], + "spans": [ + { + "bbox": [ + 362, + 338, + 488, + 349 + ], + "type": "text", + "content": "Figure 13. Effect of Resolution " + }, + { + "bbox": [ + 362, + 338, + 488, + 349 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 362, + 338, + 488, + 349 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 355, + 545, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 355, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 545, + 391 + ], + "type": "text", + "content": "ous field, which removes most of artifacts and significantly improve the accuracy. The surface loss encourages a more accurate zero level-set, which also improves the accuracy." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 343, + 391, + 509, + 441 + ], + "blocks": [ + { + "bbox": [ + 64, + 593, + 268, + 604 + ], + "lines": [ + { + "bbox": [ + 64, + 593, + 268, + 604 + ], + "spans": [ + { + "bbox": [ + 64, + 593, + 268, + 604 + ], + "type": "text", + "content": "Table 12. Reconstruction accuracy under Matterport3D." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 391, + 509, + 441 + ], + "lines": [ + { + "bbox": [ + 343, + 391, + 509, + 441 + ], + "spans": [ + { + "bbox": [ + 343, + 391, + 509, + 441 + ], + "type": "table", + "html": "
LossCDL2×100
LPull0.184
LPull + L∇0.126
LPull + L∇ + LTV0.056
LPull + L∇ + LTV + LSurface0.040
", + "image_path": "7e22fc53a056911fbe481cf4e1f2b9b16c3035961a8fc68c8ddcd085e9aefee2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "text", + "content": "Resolution " + }, + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "text", + "content": ". We compare the effect of resolution " + }, + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "text", + "content": " on the reconstructed surfaces. We try several candidates " + }, + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\{32, 64, 128, 256, 512\\}" + }, + { + "bbox": [ + 305, + 453, + 545, + 548 + ], + "type": "text", + "content": " to learn the discrete distance field. We use the same bandwidth for the calculation of the pulling loss and the TV loss, and reconstruct meshes by running the marching cubes at the same resolution. The numerical and visual comparisons show that higher resolutions can infer more geometry details but also cost more inference time." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 331, + 555, + 520, + 586 + ], + "blocks": [ + { + "bbox": [ + 378, + 442, + 473, + 452 + ], + "lines": [ + { + "bbox": [ + 378, + 442, + 473, + 452 + ], + "spans": [ + { + "bbox": [ + 378, + 442, + 473, + 452 + ], + "type": "text", + "content": "Table 13. Effect of losses." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 555, + 520, + 586 + ], + "lines": [ + { + "bbox": [ + 331, + 555, + 520, + 586 + ], + "spans": [ + { + "bbox": [ + 331, + 555, + 520, + 586 + ], + "type": "table", + "html": "
3264128256512
CDL2×1000.1860.0970.0480.0400.035
Time50s67s84s201s645s
", + "image_path": "9f01a7b88562beb560b675b7b2fb150615dfe0d261fcc56b1735a6017ceb2ae5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 364, + 588, + 487, + 597 + ], + "lines": [ + { + "bbox": [ + 364, + 588, + 487, + 597 + ], + "spans": [ + { + "bbox": [ + 364, + 588, + 487, + 597 + ], + "type": "text", + "content": "Table 14. Effect of Resolution " + }, + { + "bbox": [ + 364, + 588, + 487, + 597 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 364, + 588, + 487, + 597 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Point Number. We compare our performance on d-different numbers of points. We use the same set of shapes but with different numbers of points including " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{10K, 100K, 1000K\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". The numerical comparison in Tab. 15 shows that our method can estimate a discrete distance field from different numbers of points well, which leads to more accurate reconstructed surfaces than the latest overfitting based methods. In addition, our time complexity is not affected by the number of points a lot, while the latest" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18329" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 271, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 271, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 271, + 84 + ], + "type": "text", + "content": "methods require more time to converge on more points." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 95, + 261, + 193 + ], + "blocks": [ + { + "bbox": [ + 72, + 95, + 261, + 193 + ], + "lines": [ + { + "bbox": [ + 72, + 95, + 261, + 193 + ], + "spans": [ + { + "bbox": [ + 72, + 95, + 261, + 193 + ], + "type": "table", + "html": "
NumberMethodCDL2×100TimeMemory
10KNP0.351548s4.5G
SAP0.275147s3.5G
Ours0.12282s3.3G
100KNP0.224615s4.6G
SAP0.076354s3.6G
Ours0.04084s3.3G
1000KNP0.1751304s4.8G
SAP0.063841s3.7G
Ours0.03990s3.4G
", + "image_path": "1fc64d14df98c96b9c085a891cb12cca77f9b1112857e5e77c9a6a071a19970c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 104, + 195, + 229, + 206 + ], + "lines": [ + { + "bbox": [ + 104, + 195, + 229, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 195, + 229, + 206 + ], + "type": "text", + "content": "Table 15. Effect of point numbers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 212, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 307 + ], + "type": "text", + "content": "Initialization. We highlight our geometric initialization of distances on grid vertices by comparing it with random initialization. The comparison in Tab. 16 shows that random initialization makes the optimization converge hard, and cannot work well with the TV loss, especially with a bandwidth, which produces lots of artifacts in empty spaces. Moreover, we try an Eikonal constraint on gradients based on the geometric initialization, but did not get improvement." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 72, + 314, + 261, + 338 + ], + "blocks": [ + { + "bbox": [ + 72, + 314, + 261, + 338 + ], + "lines": [ + { + "bbox": [ + 72, + 314, + 261, + 338 + ], + "spans": [ + { + "bbox": [ + 72, + 314, + 261, + 338 + ], + "type": "table", + "html": "
RandomGeometric||∇|| = 1
CDL2 × 1000.1550.0400.054
", + "image_path": "2da4d5018056408303595ee688c28f893483270407d0395bbd786b2319a8ae22.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 107, + 340, + 227, + 350 + ], + "lines": [ + { + "bbox": [ + 107, + 340, + 227, + 350 + ], + "spans": [ + { + "bbox": [ + 107, + 340, + 227, + 350 + ], + "type": "text", + "content": "Table 16. Effect of Initialization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "text", + "content": "Bandwidth. We report the effect of bandwidth " + }, + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "text", + "content": " in the pulling loss and the TV loss in Tab. 17. For both losses, the bandwidth controls the scope of query sampling and continuous constraints. A larger bandwidth would cover more parameters which require more time and make the optimization converge slowly. For instance, if we sample queries and impose the pulling in all " + }, + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "inline_equation", + "content": "R^3" + }, + { + "bbox": [ + 46, + 357, + 287, + 585 + ], + "type": "text", + "content": " grids, the speed is even slower than NeuralPull which uses neural network for the pulling, since neural networks can generalize signed distances nearby and do not need to infer in all grids. Although a small bandwidth would save time, a too small bandwidth would cover too few grids around the surface, which leads to discontinuousness and artifacts. While if the bandwidth is large enough, we may not see surface improvement but just cost more time. Moreover, even with the same loss, NeuralPull does not produce better results than ours. Comparing with methods using neural networks, our discrete S-DF can have faster convergence, as indicated by the angular error of gradients in Fig. 14." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 73, + 588, + 261, + 695 + ], + "blocks": [ + { + "bbox": [ + 73, + 588, + 261, + 695 + ], + "lines": [ + { + "bbox": [ + 73, + 588, + 261, + 695 + ], + "spans": [ + { + "bbox": [ + 73, + 588, + 261, + 695 + ], + "type": "table", + "html": "
BandwidthCDL2×100Time
LPull (M1=3)0.04084s
LPull (All Grids)0.043628s
LTV (M2=2)0.07983s
LTV (M2=14)0.04084s
LTV (M2=30)0.04189s
LTV (M2=50)0.04096s
LTV (All Grids)0.040101s
NeuralPull0.220552s
NeuralPull+L∇+LSurface0.203554s
", + "image_path": "03a6f9da4462539730a351275d21e2199af3192c3911f2f85acea774872c26a3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 110, + 696, + 223, + 707 + ], + "lines": [ + { + "bbox": [ + 110, + 696, + 223, + 707 + ], + "spans": [ + { + "bbox": [ + 110, + 696, + 223, + 707 + ], + "type": "text", + "content": "Table 17. Effect of Bandwidth." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 545, + 162 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 162 + ], + "type": "image", + "image_path": "6298761a19a7f8987f7988b7acd038e56b09aaf3d4436cf5813001e50a8dda81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 163, + 526, + 174 + ], + "lines": [ + { + "bbox": [ + 324, + 163, + 526, + 174 + ], + "spans": [ + { + "bbox": [ + 324, + 163, + 526, + 174 + ], + "type": "text", + "content": "Figure 14. Gradients comparison with neural networks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "content": "Noises. We report our performance on point clouds with noise. We follow the experimental setting in PCP, and report our results with middle and large level noise in Tab. 18. To handle noise, we weight more on the continuous constraint " + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "inline_equation", + "content": "L_{TV}" + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "content": " using a larger " + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\alpha = 2" + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "content": " to decrease the impact of noise. The comparison shows that our method can handle noise better than the neural network based methods. Since the continuousness of neural networks make gradients near the surface over smooth to overfit all noise, this makes it hard to infer accurate zero level-set." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 343, + 310, + 509, + 344 + ], + "blocks": [ + { + "bbox": [ + 343, + 310, + 509, + 344 + ], + "lines": [ + { + "bbox": [ + 343, + 310, + 509, + 344 + ], + "spans": [ + { + "bbox": [ + 343, + 310, + 509, + 344 + ], + "type": "table", + "html": "
Noise levelNP [49]PCP [51]Ours
F-med-noise0.2800.0710.044
F-max-noise0.3100.2980.060
", + "image_path": "f9cbe42091cccb5a1bda68d2c2f47e7ca581fee89f91e914bd242cfedbe801b3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 377, + 346, + 473, + 357 + ], + "lines": [ + { + "bbox": [ + 377, + 346, + 473, + 357 + ], + "spans": [ + { + "bbox": [ + 377, + 346, + 473, + 357 + ], + "type": "text", + "content": "Table 18. Effect of noises." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 364, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 483 + ], + "type": "text", + "content": "Normals. We report our performance on point clouds with normals. If GT normal is available, we could add a supervision on normals at points on the surface. What we do is to supervise the gradients " + }, + { + "bbox": [ + 304, + 364, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\lambda f" + }, + { + "bbox": [ + 304, + 364, + 545, + 483 + ], + "type": "text", + "content": " at points on the surface using their ground truth normals. The comparisons in Tab. 19 show that the normal supervision improves the performance a little. This indicates that our method can infer pretty accurate gradients on the surface or near the surface, where gradients on the surface can be highly accurate to the normal ground truth on surfaces." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 354, + 488, + 496, + 514 + ], + "blocks": [ + { + "bbox": [ + 354, + 488, + 496, + 514 + ], + "lines": [ + { + "bbox": [ + 354, + 488, + 496, + 514 + ], + "spans": [ + { + "bbox": [ + 354, + 488, + 496, + 514 + ], + "type": "table", + "html": "
ClassOursOurs(Normals)
CDL20.04010.0403
", + "image_path": "65b5e19079f8cea75c35e8b398e385312107c35cda92a6eeb321ac8650ba9ad7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 374, + 515, + 476, + 525 + ], + "lines": [ + { + "bbox": [ + 374, + 515, + 476, + 525 + ], + "spans": [ + { + "bbox": [ + 374, + 515, + 476, + 525 + ], + "type": "text", + "content": "Table 19. Effect of normals." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 712 + ], + "type": "text", + "content": "We introduce GridPull for the fast inference of distance fields from large scale point clouds without using neural components. We infer a distance field using learnable parameters defined on discrete grids, which we can directly optimize in a more efficient way than neural networks. Besides the complexity advantage, our loss function manages to achieve more continuous distance fields with more consistent gradients, which leads to higher accuracy in surface reconstruction. We evaluate GridPull and justify its effectiveness on benchmarks. Numerical and visual comparisons show that GridPull outperforms the latest methods in terms of accuracy and complexity, even without using priors, normal, and neural networks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18330" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 58, + 91, + 259, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 259, + 102 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 259, + 102 + ], + "type": "text", + "content": "[1] ShapeNet. https://www.shapenet.org/.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 103, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 287, + 135 + ], + "type": "text", + "content": "[2] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 137, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 137, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 287, + 168 + ], + "type": "text", + "content": "[3] Matan Atzmon and yaron Lipman. SALD: sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 170, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 170, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 170, + 287, + 213 + ], + "type": "text", + "content": "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. DiGS: Divergence guided shape implicit neural representation for unoriented point clouds. CoRR, abs/2106.10811, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 214, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 214, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 58, + 214, + 287, + 257 + ], + "type": "text", + "content": "[5] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Computer Vision and Pattern Recognition, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 258, + 287, + 300 + ], + "type": "text", + "content": "[6] Alexandre Boulch and Renaud Marlet. POCO: Point convolution for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 303, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 303, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 58, + 303, + 287, + 334 + ], + "type": "text", + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 336, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 336, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 58, + 336, + 287, + 400 + ], + "type": "text", + "content": "[8] Rohan Chabra, Jan Eric Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard A. Newcombe. Deep local shapes: Learning local SDF priors for detailed 3D reconstruction. In European Conference on Computer Vision, volume 12374, pages 608-625, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 403, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 403, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 287, + 456 + ], + "type": "text", + "content": "[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 459, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 511 + ], + "type": "text", + "content": "[10] Chao Chen, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Unsupervised learning of fine structure generation for 3D point clouds by 2D projections matching. In IEEE International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 514, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 546 + ], + "type": "text", + "content": "[11] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Latent partition implicit with surface codes for 3d representation. In European Conference on Computer Vision, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 287, + 601 + ], + "type": "text", + "content": "[12] Chao Chen, Yu-Shen Liu, and Zhizhong Han. Unsupervised inference of signed distance functions from single-sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17712-17723, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "text", + "content": "[13] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "type": "text", + "content": "[14] Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. arXiv, 2010.13938, 2020. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[15] Gene Chou, Ilya Chugunov, and Felix Heide. GenSDF: Two-stage learning of generalizable signed distance functions. In Proc. of Neural Information Processing Systems (NeurIPS), 2022. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 105 + ], + "type": "text", + "content": "[16] Angela Dai and Matthias Nießner. Neural Poisson: Indicator functions for neural fields. arXiv preprint arXiv:2211.14249, 2022. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 152 + ], + "type": "text", + "content": "[17] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J. Mitra, and Michael Wimmer. Points2Surf: Learning implicit surfaces from point clouds. In European Conference on Computer Vision, 2020. 1, 2, 4, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 155, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 155, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 155, + 545, + 219 + ], + "type": "text", + "content": "[18] Huan Fu, Bowen Cai, Lin Gao, Lingxiao Zhang, Cao Li, Zengqi Xun, Chengyue Sun, Yiyun Fei, Yu Zheng, Ying Li, Yi Liu, Peng Liu, Lin Ma, Le Weng, Xiaohang Hu, Xin Ma, Qian Qian, Rongfei Jia, Binqiang Zhao, and Hao Zhang. 3d-front: 3d furnished rooms with layouts and semantics. CoRR, abs/2011.09127, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 222, + 545, + 255 + ], + "type": "text", + "content": "[19] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 258, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 258, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 258, + 545, + 299 + ], + "type": "text", + "content": "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition, 2012. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 303, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 303, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 312, + 303, + 545, + 346 + ], + "type": "text", + "content": "[21] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. In International Conference on Computer Vision, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 545, + 403 + ], + "type": "text", + "content": "[22] Lily Goli, Daniel Rebain, Sara Sabour, Animesh Garg, and Andrea Tagliasacchi. nerf2nerf: Pairwise registration of neural radiance fields. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9354-9361. IEEE, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 407, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 407, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 312, + 407, + 545, + 460 + ], + "type": "text", + "content": "[23] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 463, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 463, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 312, + 463, + 545, + 495 + ], + "type": "text", + "content": "[24] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv, 2002.10099, 2020. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 498, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 545, + 552 + ], + "type": "text", + "content": "[25] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. ShapeCaptioner: Generative caption network for 3D shapes by learning a mapping from parts detected in multiple views to sentences. ArXiv, abs/1908.00120, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 545, + 609 + ], + "type": "text", + "content": "[26] Zhizhong Han, Chao Chen, Yu-Shen Liu, and Matthias Zwicker. DRWR: A differentiable renderer without rendering for unsupervised 3D structure learning from silhouette images. In International Conference on Machine Learning, 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 545, + 656 + ], + "type": "text", + "content": "[27] A. Handa, V. Patraucean, V. Badrinarayanan, S. Stent, and R. Cipolla. Understanding realworld indoor scenes with synthetic data. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4077-4085, 2016. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "text", + "content": "[28] Ayaan Haque, Matthew Tancik, Alexei Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18331" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[29] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 150 + ], + "type": "text", + "content": "[30] Jiahui Huang, Hao-Xiang Chen, and Shi-Min Hu. A neural galerkin solver for accurate surface reconstruction. ACM Trans. Graph., 41(6), 2022. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "type": "text", + "content": "[31] Meng Jia and Matthew Kyan. Learning occupancy function from point clouds for surface reconstruction. arXiv, 2010.11378, 2020.1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 237 + ], + "type": "text", + "content": "[32] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3D scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[33] Sijia Jiang, Jing Hua, and Zhizhong Han. Coordinate quantized neural implicit representations for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 327 + ], + "type": "text", + "content": "[34] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. SDFDiff: Differentiable rendering of signed distance fields for 3D shape optimization. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "text", + "content": "[35] Chuan Jin, Tieru Wu, and Junsheng Zhou. Multi-grid representation with field regularization for self-supervised surface reconstruction from point clouds. Computers & Graphics, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 372, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 288, + 403 + ], + "type": "text", + "content": "[36] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 437 + ], + "type": "text", + "content": "[37] Michael M. Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions Graphics, 32(3):29:1-29:13, 2013. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 437, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 470 + ], + "type": "text", + "content": "[38] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-Surf: Voxel-based implicit surface representation. CoRR, abs/2208.10925, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "type": "text", + "content": "[39] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 515, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 515, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 515, + 288, + 568 + ], + "type": "text", + "content": "[40] Tianyang Li, Xin Wen, Yu-Shen Liu, Hua Su, and Zhizhong Han. Learning deep implicit functions for 3D shapes with dynamic code clouds. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12830-12840, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 288, + 624 + ], + "type": "text", + "content": "[41] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 288, + 669 + ], + "type": "text", + "content": "[42] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. SDF-SRN: Learning signed distance 3D object reconstruction from static images. In Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "type": "text", + "content": "[43] Minghua Liu, Xiaoshuai Zhang, and Hao Su. Meshing point clouds with predicted intrinsic-extrinsic ratio guidance. In European Conference on Computer vision, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[44] Shichen Liu, Shunsuke Saito, Weikai Chen, and Hao Li. Learning to infer implicit surfaces without 3D supervision. In Advances in Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 172 + ], + "type": "text", + "content": "[45] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 173, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 545, + 225 + ], + "type": "text", + "content": "[46] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Pengshuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3D reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 271 + ], + "type": "text", + "content": "[47] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. NeUDF: Leaning neural unsigned distance fields with volume rendering. In Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 272, + 545, + 304 + ], + "type": "text", + "content": "[48] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. Computer Graphics, 21(4):163-169, 1987. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 305, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 305, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 545, + 360 + ], + "type": "text", + "content": "[49] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-Pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In International Conference on Machine Learning, 2021. 1, 2, 3, 5, 6, 7, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 404 + ], + "type": "text", + "content": "[50] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Reconstructing surfaces for sparse point clouds with on-surface priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6305-6315, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 459 + ], + "type": "text", + "content": "[51] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6316-6327, 2022. 2, 3, 5, 6, 7, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 459, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 459, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 312, + 459, + 545, + 503 + ], + "type": "text", + "content": "[52] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3d point clouds via noise to noise mapping. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 503, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 503, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 312, + 503, + 545, + 559 + ], + "type": "text", + "content": "[53] Baorui Ma, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Towards better gradient consistency for neural signed distance functions via level set alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17724-17734, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 545, + 603 + ], + "type": "text", + "content": "[54] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. A-CORN: adaptive coordinate networks for neural scene representation. CoRR, abs/2105.02788, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 312, + 603, + 545, + 657 + ], + "type": "text", + "content": "[55] Xiaoxu Meng, Weikai Chen, and Bo Yang. NeAT: Learning neural implicit surfaces with arbitrary topologies from multi-view images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "text", + "content": "[56] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18332" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "text", + "content": "[57] Zhenxing Mi, Yiming Luo, and Wenbing Tao. SSRNet: Scalable 3D surface reconstruction network. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 162 + ], + "type": "text", + "content": "[58] Mateusz Michalkiewicz, Jhony K. Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders P. Eriksson. Deep level sets: Implicit surface representations for 3D shape inference. CoRR, abs/1901.06802, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 217 + ], + "type": "text", + "content": "[59] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. N-eRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 288, + 262 + ], + "type": "text", + "content": "[60] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "text", + "content": "[61] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 320, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 288, + 365 + ], + "type": "text", + "content": "[62] Michael Oechsle, Songyou Peng, and Andreas Geiger. U-NISURF: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 365, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 288, + 408 + ], + "type": "text", + "content": "[63] Amine Ouasfi and Adnane Boukhayma. Few 'zero level set'-shot learning of shape signed distance functions in feature space. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "type": "text", + "content": "[64] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 466, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 510 + ], + "type": "text", + "content": "[65] Songyou Peng, Chiyu “Max” Jiang, Yiji Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. In Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 511, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 511, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 511, + 288, + 555 + ], + "type": "text", + "content": "[66] Songyou Peng, Michael Niemeyer, Lars M. Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, volume 12348, pages 523-540, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 556, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 288, + 600 + ], + "type": "text", + "content": "[67] Albert Pumarola, Artsiom Sanakoyeu, Lior Yariv, Ali Thabet, and Yaron Lipman. Visco grids: Surface reconstruction with viscosity and coarea grids. In Advances in Neural Information Processing Systems, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "text", + "content": "[68] Konstantinos Rematas, Ricardo Martin-Brualla, and Vittorio Ferrari. Sharf: Shape-conditioned radiance fields from a single view. In International Conference on Machine Learning, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 689 + ], + "type": "text", + "content": "[69] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 715 + ], + "type": "text", + "content": "[70] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qin-hong Chen, Benjamin Recht, and Angjoo Kanazawa." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 545, + 106 + ], + "type": "text", + "content": "Plenoxels: Radiance fields without neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "text", + "content": "[71] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, 2020. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 207 + ], + "type": "text", + "content": "[72] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 251 + ], + "type": "text", + "content": "[73] Lars Mescheder Marc Pollefeys Andreas Geiger Songyou Peng, Michael Niemeyer. Convolutional occupancy networks. In European Conference on Computer Vision, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 545, + 285 + ], + "type": "text", + "content": "[74] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CoRR, abs/2111.11215, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 287, + 545, + 331 + ], + "type": "text", + "content": "[75] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 545, + 365 + ], + "type": "text", + "content": "[76] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arXiv preprint arXiv:2206.05085, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 366, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 366, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 545, + 431 + ], + "type": "text", + "content": "[77] Towaki Takikawa, Joel Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 545, + 487 + ], + "type": "text", + "content": "[78] Jiapeng Tang, Jiabao Lei, Dan Xu, Feiying Ma, Kui Jia, and Lei Zhang. SA-ConvONet: Sign-agnostic optimization of convolutional occupancy networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 489, + 545, + 533 + ], + "type": "text", + "content": "[79] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Annual Conference on Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 535, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 535, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 535, + 545, + 588 + ], + "type": "text", + "content": "[80] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, and Christian Theobalt. PatchNets: Patch-Based Generalizable Deep Implicit 3D Shape Representations. European Conference on Computer Vision, 2020. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 590, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 312, + 590, + 545, + 624 + ], + "type": "text", + "content": "[81] Delio Vicini, Sbastien Speierer, and Wenzel Jakob. Differentiable signed distance function rendering. ACM Transactions on Graphics, 41(4):125:1-125:18, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 677 + ], + "type": "text", + "content": "[82] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. NeuRIS: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 680, + 545, + 714 + ], + "type": "text", + "content": "[83] Meng Wang, Yu-Shen Liu, Yue Gao, Kanle Shi, Yi Fang, and Zhizhong Han. Lp-dif: Learning local pattern-specific deep implicit function for 3d objects and scenes. In Pro" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18333" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 105 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21856-21865, June 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "type": "text", + "content": "[84] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems, pages 27171-27183, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 163, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 163, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 287, + 205 + ], + "type": "text", + "content": "[85] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics, 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 209, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 287, + 251 + ], + "type": "text", + "content": "[86] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. HF-NeuS: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 287, + 308 + ], + "type": "text", + "content": "[87] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net++: Point cloud completion by transformer-enhanced multi-step point moving paths. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):852-867, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 309, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 287, + 364 + ], + "type": "text", + "content": "[88] Xin Wen, Junsheng Zhou, Yu-Shen Liu, Hua Su, Zhen Dong, and Zhizhong Han. 3d shape reconstruction from 2d images with disentangled attribute flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3803-3813, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 287, + 409 + ], + "type": "text", + "content": "[89] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep geometric prior for surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 287, + 443 + ], + "type": "text", + "content": "[90] Yunjie Wu and Zhengxing Sun. DFR: differentiable function rendering for learning 3D generation from images. Computer Graphics Forum, 39(5):241-252, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 287, + 498 + ], + "type": "text", + "content": "[91] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In IEEE International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 501, + 287, + 555 + ], + "type": "text", + "content": "[92] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflake point deconvolution for point cloud completion and generation with skip-transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):6320-6338, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 556, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 556, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 556, + 287, + 589 + ], + "type": "text", + "content": "[93] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 54, + 590, + 287, + 644 + ], + "type": "text", + "content": "[94] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 647, + 287, + 689 + ], + "type": "text", + "content": "[95] Wang Yifan, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Iso-Points: Optimizing neural implicit surfaces with hybrid representations. CoRR, abs/2012.06434, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 691, + 287, + 714 + ], + "type": "text", + "content": "[96] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 464 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "neural radiance fields. In IEEE International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "type": "text", + "content": "[97] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, June 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 195 + ], + "type": "text", + "content": "[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. ArXiv, abs/2022.00665, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 312, + 197, + 545, + 240 + ], + "type": "text", + "content": "[99] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3D objects with differentiable rendering of sdf shape priors. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 285 + ], + "type": "text", + "content": "[100] Wenbin Zhao, Jiabao Lei, Yuxin Wen, Jianguo Zhang, and Kui Jia. Sign-agnostic implicit learning of surface self-similarities for shape modeling and reconstruction from raw point clouds. CoRR, abs/2012.07498, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[101] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, and Zhizhong Han. Learning a more continuous zero level set in unsigned distance fields through level set projection. In Proceedings of the IEEE/CVF international conference on computer vision, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "text", + "content": "[102] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. Advances in Neural Information Processing Systems, 35:16481-16494, 2022. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 398, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 398, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 545, + 430 + ], + "type": "text", + "content": "[103] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. CoRR, abs/1605.04797, 2016. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 432, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 545, + 464 + ], + "type": "text", + "content": "[104] Qian-Yi Zhou and Vladlen Koltun. Dense scene reconstruction with points of interest. ACM Transactions on Graphics, 32(4):112:1-112:8, 2013. 4" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18334" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_content_list.json b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6c4f22ff620a2647ceeb41351fdbb336cce3e618 --- /dev/null +++ b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_content_list.json @@ -0,0 +1,1558 @@ +[ + { + "type": "text", + "text": "Grounded Entity-Landmark Adaptive Pre-training for Vision-and-Language Navigation", + "text_level": 1, + "bbox": [ + 225, + 130, + 743, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yibo Cui $^{1,2}$ , Liang Xie $^{1,2}$ , Yakun Zhang $^{1,2}$ , Meishan Zhang $^{3}$ , Ye Yan $^{1,2}$ , Erwei Yin $^{1,2}$ $^{1}$ Defense Innovation Institute, Chinese Academy of Military Science \n $^{2}$ Tianjin Artificial Intelligence Innovation Center $^{3}$ Harbin Institute of Technology (Shenzhen)", + "bbox": [ + 104, + 202, + 861, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/CSir1996/VLN-GELA", + "bbox": [ + 323, + 258, + 645, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-modal alignment is one key challenge for Vision- and Language Navigation (VLN). Most existing studies concentrate on mapping the global instruction or single sub-instruction to the corresponding trajectory. However, another critical problem of achieving fine-grained alignment at the entity level is seldom considered. To address this problem, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN tasks. To achieve the adaptive pre-training paradigm, we first introduce grounded entity-landmark human annotations into the Room-to-Room (R2R) dataset, named GEL-R2R. Additionally, we adopt three grounded entity-landmark adaptive pretraining objectives: 1) entity phrase prediction, 2) landmark bounding box prediction, and 3) entity-landmark semantic alignment, which explicitly supervise the learning of fine-grained cross-modal alignment between entity phrases and environment landmarks. Finally, we validate our model on two downstream benchmarks: VLN with descriptive instructions (R2R) and dialogue instructions (CVDN). The comprehensive experiments show that our GELA model achieves state-of-the-art results on both tasks, demonstrating its effectiveness and generalizability.", + "bbox": [ + 75, + 340, + 473, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-and-Language Navigation (VLN) [3] is an important task in the Embodied Vision community, which has gained great attention [66, 40, 50]. It aims to ask an agent to reach the target location inside photo-realistic environments by following natural language instructions. In VLN, cross-modal alignment is one critical step to accurately make the action decision for the agent [60], since matching the mentioned landmarks (objects or scenes) with visual observations can help the comprehensive understanding of the en", + "bbox": [ + 75, + 741, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0a06bdb3d822c15e1cfd882814e2deccc812240946ea9928f2f8d73bf9aca8e.jpg", + "image_caption": [ + "Figure 1. An illustration of an embodied agent navigating in a 3D photo-realistic environment. The agent is expected to navigate based on environment landmarks that correspond to the entity phrase in the instruction." + ], + "image_footnote": [], + "bbox": [ + 514, + 309, + 879, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "virements and instructions [60, 15, 68]. However, most of the available datasets could only provide the coarse-grained text-image alignment signals [15], i.e., instruction-level correspondences to the complete trajectories, where finer ones are required to learn the cross-modality alignment for well-performed navigation [18, 67, 47, 43, 57].", + "bbox": [ + 496, + 625, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A large part of previous works [66, 11, 40, 50] has concentrated on the global alignment of instructions and trajectories, which matches instructions to the overall temporal visual trajectory via reinforcement learning [60], auxiliary reasoning [66], or transformer-based matching pre-training [7, 50, 40]. Others [18, 67, 15] have attempted to align sub-instructions and sub-trajectories locally. At this granularity, the agents are designed to segment long instruction [67, 18] and determine which sub-instruction to focus on. To guide local cross-modal alignment, He et al. [15] introduced the Landmark-RxR dataset by human-annotating sub-instructions and sub-trajectories alignment. They ad", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 101, + 887, + 228, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12043", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vanced the granularity of visual-textual alignment, but the agent learning is still supervised at the sentence level. The alignment within (sub-)instructions should be considered.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Identifying environment landmarks that correspond to entities in the instruction, as shown in Figure 1, is the next step to refine alignment and improve navigation. Recent studies [47, 43, 57] investigated the same problem, suggesting the alignment of entities and object regions [47] or the scene- and object-aware transformer model [43]. But the object regions were not under the direct supervision of the corresponding entity phrases in the instruction. To generate navigation instructions enriched with landmark phrases, Wang et al. [57] introduced grounded landmark annotations using a dependency parser and weak supervision from the pose traces, leading to the entity-landmark alignments of insufficient precision. Therefore, a dataset with high-quality entity-landmark level grounding annotations and powerful supervision for entity-landmark level cross-modal alignment is highly desired for VLN.", + "bbox": [ + 75, + 138, + 470, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above limitations, we first enhance the Room-to-Room (R2R) dataset [3] by introducing additional high-quality grounded entity-landmark human annotations, known as the Grounded Entity-Landmark R2R dataset (GEL-R2R). The GEL-R2R dataset is annotated with abundant precisely matched entity-landmark pairs (as magenta text and bounding box illustrated in Figure 1), which could provide the VLN models with direct supervision of fine-grained cross-modal alignment.", + "bbox": [ + 75, + 382, + 468, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To verify the value of our GEL-R2R dataset, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm to improve the learning of entity-landmark level alignment for the VLN pre-trained model. Specifically, we suggest three grounded entity-landmark adaptive pre-training objectives: 1) Entity Phrase Prediction (EPP), locating entity phrases that refer to environment landmarks from the instruction; 2) Landmark Bounding box Prediction (LBP), predicting the bounding box of environment landmarks that match with entity phrases; and 3) Entity-Landmark Semantic Alignment (ELSA), aligning the matched pairs of landmark patches and entity tokens in the feature space by contrastive loss. These three tasks explicitly equip the model with the ability to comprehend the entity-level grounding between human instructions and visual environment observations.", + "bbox": [ + 75, + 520, + 470, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we conduct extensive experiments on two downstream tasks to evaluate our proposed dataset GEL-R2R and adaptive pre-training methods GELA: Room-to-Room (R2R) and Vision-and-Dialog Navigation (CVDN). The instructions in R2R are fine-grained descriptions for the navigation trajectory, whereas the instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. We use HAMT [7] as the backbone VLN model for the GELA pre-training. The results demonstrate our", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "suggested GELA achieves state-of-the-art (SoTA) performance in both seen and unseen environments of the above two benchmarks: $62\\%$ SPL on R2R and 5.87 GP on CVDN.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are three-fold:", + "bbox": [ + 517, + 137, + 834, + 148 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct a new dataset GEL-R2R, which is the first dataset with high-quality grounded entity-landmark human annotations in the VLN domain.", + "- We propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN, explicitly supervising the models to learn fine-grained cross-modal semantic alignment between entity phrases and environment landmarks.", + "- Our suggested GELA achieves state-of-the-art results on two challenging VLN downstream benchmarks, demonstrating its effectiveness and generalizability." + ], + "bbox": [ + 517, + 151, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 330, + 640, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-and-Language Navigation. Since the R2R benchmark was proposed by [3], studies of the VLN have made considerable progress in various aspects. VLN datasets have become increasingly diverse, existing in indoor [4, 30, 49, 55, 65] or outdoor [5, 16, 42] environments, and in discrete [3, 49] or continuous [29, 28, 34, 19] environments. The architecture of VLN models has become more complex, progressing from LSTM-based [11, 39] to transformer-based [14, 43, 45, 7]. To develop the competence of getting closer to the destination and to the ground truth path [23], VLN agents have been enhanced mainly through action strategy learning [63, 8, 48, 17, 21] and multi-modal representation learning [22, 59, 27, 26, 1]. Many excellent VLN models use imitation and reinforcement learning-based training paradigms for action strategy learning [60, 44, 54]. Recently, multi-modal transformer-based models that use effective joint representations for instructions and visual observations have achieved promising performance in VLN [12, 7, 50]. In this passage, we focus on incorporating more accurate and fine-grained cross-modal grounding information to improve the performance of VLN models by supervising the alignment between entity phrases and environment landmarks.", + "bbox": [ + 496, + 357, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Phrase-to-Region Grounding. Phrase-to-region grounding is an important task in the Vision-Language (VL) domain that involves localizing textual entities in an image, commonly abbreviated as phrase grounding [25, 33, 38]. This task has seen significant progress since the introduction of the Flickr30k Entities dataset [46] and has played a crucial role in learning fine-grained semantic visual representation. Various models such as VisualBERT [32], MDETR [25], and GLIP [33] have explored this task under different architectures, and grounded pre-training has shown to facilitate fine-grained semantic understanding. However, the task of fine-grained cross-modal semantic alignment in embodied VLN tasks is more challenging due", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12044", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/716acb5eec926d3d0ee0b955169d5a56901936bb313df764bc11bdf028d96b6e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AllObjectScene
TrajectoryInstructionPhraseP/IBoxB/IPhraseBoxP/BPhraseBoxP/B
Train467514039577884.121211468.6330756632482.0627032579022.14
Val Seen340102141964.1187418.56193940652.10225746762.07
Val Unseen775232594834.08202968.73432489652.075159113312.20
Total579017385714674.111501838.6437019762782.0634448739092.15
", + "bbox": [ + 81, + 89, + 890, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Statistics on our GEL-R2R dataset. P/I (resp. B/I) denotes the average number of entity phrases contained (resp. landmark bounding boxes matched with entity phrases) in each instruction. And P/B represents the average number of landmark bounding boxes matched with each entity phrase.", + "bbox": [ + 75, + 213, + 892, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to more complex scenes and various objects. In this paper, we introduce high-quality entity-landmark grounding human annotations into the R2R dataset and propose a grounded entity-landmark adaptive pre-training scheme for VLN pre-trained models to address this problem.", + "bbox": [ + 75, + 280, + 467, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptive Pre-training. In the field of natural language processing, a pre-trained language model in the general domain is continuously pre-trained to learn the knowledge appropriate for a particular task or domain. This process is known as adaptive pre-training [13], which is conducted between pre-training and fine-tuning. Previous work [41, 64] gains consistent improvements by continuously pre-training an adaptive language model with the Masked Language Model (MLM) [10] objective. Furthermore, through three pre-training objectives — the MLM, Span Boundary Objective (SBO) [24], and Perturbation Masking Objective (PMO) — Wu et al. [62] improved the overall performance of a dialogue understanding model. In this work, we adopt adaptive pre-training into the area of VLN. Specifically, based on our human-annotated GEL-R2R dataset, we continuously train the SoTA pre-trained model [7] with three different supervised entity-landmark grounding objectives.", + "bbox": [ + 75, + 356, + 467, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. GEL-R2R Dataset", + "text_level": 1, + "bbox": [ + 76, + 626, + 256, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Construction", + "text_level": 1, + "bbox": [ + 76, + 650, + 276, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To establish alignment between entity phrases in instructions and their corresponding landmarks in the surrounding environment, we introduce GEL-R2R, a dataset that includes human annotations of grounded entity-landmark pairs. The construction process consists of five stages:", + "bbox": [ + 75, + 674, + 467, + 748 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Raw data preparation. We collect panoramicas of the R2R dataset from the Matterport3D (MP3D) simulator [4] for each viewpoint. To improve the accuracy and efficiency of annotating landmarks, we mark the next action direction in panoramas and match each panorama with the corresponding sub-instruction based on [18].", + "(2) Annotation tool development. We develop a web-based annotation platform using the label-studio [56] to facilitate annotations. The annotation interface is shown in the supplementary material. Annotators first mark the entity" + ], + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "phrases in the instructions and then identify the corresponding landmarks in the images using the same labels.", + "bbox": [ + 498, + 280, + 888, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(3) Annotation guideline standardization. Through pre-annotation, we standardized the annotation guideline by establishing four rules to ensure the consistency of the annotations. The four rules are as follows:", + "bbox": [ + 498, + 311, + 888, + 369 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Alignment Rule: The entity phrase in instructions should match landmarks in panoramas precisely.", + "- Free Text Rule: Free text instead of the class should be annotated, for instance, \"the white dining table\" instead of \"table\".", + "- Text Coreference Rule: Entity phrases referring to the same landmark are marked with the same label.", + "- Unique Landmark Rule: For an entity phrase, only one corresponding landmark bounding box should be annotated in a panorama." + ], + "bbox": [ + 517, + 371, + 888, + 521 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(4) Data annotation and revision. We first select annotators by testing the participants using 50 test instruction-path pairs. The qualification process leaves us 43 qualified annotators to complete the annotation task. Following the annotation, our experts double-check the annotations and correct any errors to ensure compliance with the above four rules.", + "(5) Data processing. In order to ensure the quality of the annotations, we first eliminate the incorrect annotations violating the above rules and then correct a few incorrect words. Finally, we established the GEL-R2R dataset by introducing additional grounded entity-landmark annotations to the R2R dataset. More details about data collection are provided in the supplementary material." + ], + "bbox": [ + 498, + 522, + 888, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Statistics and Analysis", + "text_level": 1, + "bbox": [ + 500, + 726, + 707, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1 presents statistics for the train, validation seen, and validation unseen splits of GEL-R2R. The dataset contains a total of 71,467 entity phrases, with 57,788 in the train split, 4,196 in the validation seen split, and 9,483 in the validation unseen split. It also includes 150,183 landmark boxes, of which 121,146 are in the train split, 8,741 are in the validation seen split, and 20,296 are in the validation unseen split. On average, each instruction contains 4.1 entity phrases, which is roughly equal to the number of actions in a path. Furthermore, each instruction uses about", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12045", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b96f6303d0c785c7f6689f039b172cf1fdb49b227d1b2130b6db16f230a1172b.jpg", + "image_caption": [ + "Figure 2. Overview of GELA. The three adaptive pre-training objectives served by our scheme encourage: 1) the annotated entity to predict the bounding box of its corresponding landmark in panoramas, 2) the annotated landmark to predict the positions of its corresponding entity phrases, and 3) the cross-modal encoder to establish the semantic alignment between landmarks and entities representations. After grounded entity-landmark adaptive pre-training, the GELA model is fine-tuned for the two VLN downstream tasks, R2R and CVDN." + ], + "image_footnote": [], + "bbox": [ + 133, + 92, + 836, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8.5 landmark boxes on average. As a result, the agent has access to a diverse set of landmarks that can aid in decision-making during each navigation episode.", + "bbox": [ + 75, + 453, + 470, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compute statistics on the object and scene, taking into account their varying identification difficulties. In total, we gather 37019 object entity phrases and 34448 scene entity phrases, referring to 76278 object landmark boxes and 73909 scene landmark boxes, respectively. On average, there are about 2 landmark boxes corresponding to each exact entity phrase, whether for objects or scenes. This indicates that the same landmark appears in approximately two adjacent viewpoint panoramicas from different sizes and angles, which enables the agent to perceive the landmark from multiple views. In summary, the GEL-R2R dataset is the first to provide high-quality entity-landmark alignment human annotations, which are essential for the area of VLN, such as cross-modal representation learning, data augmentation, and interpretable navigation.", + "bbox": [ + 75, + 500, + 473, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 743, + 169, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Problem Setup and Overview", + "text_level": 1, + "bbox": [ + 76, + 768, + 339, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Setup. In VLN tasks, an embodied agent should move to a target location from a starting pose inside a 3D photo-realistic environment by following natural language navigable instructions. The instruction is a sequence of $L$ words, denoted as $I = \\{w_0, \\dots, w_{L-1}\\}$ , and it guides the agent to traverse a connectivity graph $G$ in order to reach the intended destination. At each step $t$ , the agent ac", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "quires a new panoramic visual observation $O_{t} = \\{o_{t,i}\\}_{i=1}^{36}$ from neighboring environment, where $o_{t,i} = [v_{t,i};a_{t,i}]$ consists RGB image $v_{t,i}$ and orientation angle $a_{t,i}$ of $i$ -th view. The relative angles of $n$ navigable viewpoints to the current viewpoint and the stop action constitute the action space at step $t$ , denoted as $A_{t} = \\{act_{t,1},\\dots ,act_{t,n},[STOP]\\}$ . The agent selects an action to move to a navigable viewpoint in $G$ or stop at the current location. Once the agent stops, the navigation episode is completed.", + "bbox": [ + 496, + 450, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. In this study, we propose a novel grounded entity-landmark adaptive pre-training scheme for a VLN pre-trained model, whose architecture is shown in Figure 2. The training scheme introduces explicit supervision on entity-landmark grounding to enhance fine-grained cross-modal representation learning. The scheme comprises two distinct stages: GELA pre-training on several proxy objectives and fine-tuning on the VLN downstream tasks.", + "bbox": [ + 496, + 590, + 892, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Pre-trained Model", + "text_level": 1, + "bbox": [ + 500, + 724, + 679, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in the bottom-left of Figure 2, our pretrained model is a fully transformer-based architecture for multi-modal decision-making, which is modified from the classical cross-modal model LXMERT [53]. The pretrained model takes three inputs: a global instruction $I$ , history information $H_{t}$ , and current panoramic visual observation $O_{t}$ , which are fed in a language encoder, a history encoder, and a vision encoder, respectively. And then the textual and visual modalities exchange the signals through cross-attention layers in the cross-modal en", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12046", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coder. Specifically, the visual modality is the concatenation of history and visual observation. Finally, the representation of tokens in instruction, history, and visual state is $Z = \\{z_{\\mathrm{cls}}, z_1, \\dots, z_L\\}$ , $Y_t = \\{y_{\\mathrm{cls}}, y_1, \\dots, y_{t-1}\\}$ , and $S_t = \\{s_1, \\dots, s_{36}, s_{\\mathrm{stop}}\\}$ , respectively.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To learn effective uni-modal and multi-modal representation, transformer-based models for VLN commonly undergo pre-training on in-domain datasets using several proxy tasks [14, 50, 7]. These tasks include common vision-language pre-training tasks as well as VLN-specific auxiliary tasks. In this study, we utilize five proxy tasks: Masked Language Modeling (MLM) [10], Masked Region Classification (MRC) [37], Instruction Trajectory Matching (ITM) [40], Single-step Action Prediction (SAP) [7], and Spatial Relationship Prediction (SPREL) [7]. Details are demonstrated in the supplementary material.", + "bbox": [ + 75, + 167, + 470, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. GELA Pre-training", + "text_level": 1, + "bbox": [ + 76, + 347, + 264, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the top-left of Figure 2, we adopt three adaptive pre-training objectives for learning entity-landmark level alignment, which are Entity Phrase Prediction (EPP), Landmark Bounding box Prediction (LBP), and Entity-Landmark Semantic Alignment (ELSA). The three proxy objectives are designed by mimicking the objectives used in pre-training vision-language Transformers, particularly with the goal of learning phase-level grounding information. LBP is adapted from the training objective of TransVG [9]. EPP and ELSA are extended from the soft token prediction and contrastive alignment of MDETR [25], respectively.", + "bbox": [ + 75, + 371, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Entity Phrase Prediction (EPP). In this objective, we predict the positions of entity phrases that correspond to annotated environment landmarks. We first transfer the human-annotated entity location to a mask vector $M_z$ of $L + 1$ dimensions, which is the same as $Z$ . Similarly, we transfer the human-annotated landmark bounding box to a 37-dim (the same as $S_t$ ) mask vector $M_s$ . Then we train the model to infer a uniform distribution over all token positions that refer to the corresponding landmark patches and supervise this process by the mask vector $M_z$ . Specifically, we average the patches representation of the landmark: $S_t \\times M_s^\\top$ , where $\\times$ denotes the matrix product, obtaining a 768-dim vector. Then we put it into a two-layer feedforward network (FFN) to predict a distribution over the token positions in the instruction sequence:", + "bbox": [ + 75, + 553, + 470, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl o g i t s = \\operatorname {S o f t m a x} \\left(\\operatorname {F F N} \\left(S _ {t} \\times M _ {s} ^ {\\top}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 792, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the FFN maps a vector of 768-dim to a vector of $L + 1$ dimensions. Finally, we minimize the cross-entropy loss between logits and $M_z$ :", + "bbox": [ + 76, + 824, + 468, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E P P}} = \\text {C r o s s E n t r o p y} (\\operatorname {l o g i t s}, M _ {z}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Landmark Bounding Box Prediction (LBP). In this objective, we predict the bounding box of the landmark that matches with annotated entity phrases. We train the model to directly predict a 4-dim vector $box' = (x', y', w', h')$ as the coordinates of the bounding box for each entity phrases and supervise this process by the human-annotated bounding box $box = (x, y, w, h)$ . Specifically, we first average the token embeddings of the entity phrase: $Z \\times M_z^\\top$ and then we predict the coordinates of the box through a two-layer FFN and sigmoid function:", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nb o x ^ {\\prime} = \\operatorname {S i g m o i d} \\left(\\operatorname {F F N} \\left(Z \\times M _ {z} ^ {\\top}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 251, + 890, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the FFN maps a vector of 768-dim to a vector of 4-dim. Finally, we apply the smooth L1 loss and generalized IoU loss (GIoU loss) to optimize the box coordinates $M_z$ :", + "bbox": [ + 496, + 279, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L B P}} = \\mathcal {L} _ {\\text {s m o o t h} - l _ {1}} \\left(\\text {b o x}, \\text {b o x} ^ {\\prime}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {G I o U}} \\left(\\text {b o x}, \\text {b o x} ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 335, + 890, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weight coefficient of GIoU loss to balance these two losses.", + "bbox": [ + 496, + 361, + 890, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Entity-Landmark Semantic Alignment (ELSA). While the above two unidirectional prediction tasks use positional information to match the entity and landmark, the entity-landmark semantic alignment loss enforces alignment between the hidden embeddings of the landmark and entity at the output of the cross-modal encoder. This additional contrastive alignment loss ensures that the representations of the landmark patches and the corresponding entity tokens are closer in the feature space compared to representations of unrelated tokens. This constraint is stronger than the above two unidirectional prediction losses as it directly operates on the representations and is not single-handedly based on the positional information. Specifically, inspired by InfoNCE loss [36], the objective is the mean of two contrastive losses as follows:", + "bbox": [ + 496, + 392, + 890, + 618 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {i = 0} ^ {3 6} \\frac {1}{\\left| Z _ {i} ^ {+} \\right|} \\sum_ {j \\in Z _ {i} ^ {+}} - \\log \\left(\\frac {\\exp \\left(s _ {i} ^ {\\top} z _ {j} / \\tau\\right)}{\\sum_ {k = 0} ^ {L} \\exp \\left(s _ {i} ^ {\\top} z _ {k} / \\tau\\right)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 640, + 890, + 689 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {z} = \\sum_ {i = 0} ^ {L} \\frac {1}{\\left| S _ {i} ^ {+} \\right|} \\sum_ {j \\in s _ {i} ^ {+}} - \\log \\left(\\frac {\\exp \\left(z _ {i} ^ {\\top} s _ {j} / \\tau\\right)}{\\sum_ {k = 0} ^ {3 6} \\exp \\left(z _ {i} ^ {\\top} s _ {k} / \\tau\\right)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 712, + 890, + 761 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {E L S A}} = \\left(\\mathcal {L} _ {s} + \\mathcal {L} _ {z}\\right) / 2, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 776, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Z_{i}^{+}$ is the token set to be matched with an annotated patch $s_i$ , and $S_{i}^{+}$ is the landmark patch set that should be matched with an annotated token $z_{i}$ , and $\\tau$ is a temperature parameter regulating attention to negative samples.", + "bbox": [ + 496, + 797, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the full adaptive pre-training objective is:", + "bbox": [ + 517, + 859, + 864, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {G E L A}} = \\alpha \\mathcal {L} _ {\\text {E P P}} + \\beta \\mathcal {L} _ {\\text {L B P}} + \\gamma \\mathcal {L} _ {\\text {E L S A}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12047", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Fine-tuning for VLN Tasks", + "text_level": 1, + "bbox": [ + 76, + 90, + 321, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The fine-tuning stage is illustrated in the right of Figure 2, where we generalize the GELA model for two VLN downstream tasks under the scheme of imitation learning (IL) and reinforcement learning (RL) by following previous work [20, 7]. IL supervises the agent to clone the behavior of the expert while RL encourages the agent to explore the trajectory according to the learning policy. Firstly, the GELA model navigates in the environment following the ground-truth action and generates gradients by IL. Secondly, using the same instruction, the model samples the action space to make decisions and uses RL to generate gradients. Finally, we combine the gradients and optimize the pre-trained model. For the CVDN task, our setting is the same to [7]. For the R2R task, we apply three settings with different data augmentation following [31]:", + "bbox": [ + 75, + 113, + 468, + 340 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- none: without data augmentation on environments.", + "- st: with data augmentation by style transformation on original environments. Specifically, each discrete view of panoramas is transferred with a random style.", + "- smo: with data augmentation by semantic class masking on synthesis environments, which are generated by image synthesis with the same style as the original environments. One semantic class is randomly masked out during image generation." + ], + "bbox": [ + 94, + 342, + 468, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment and Results", + "text_level": 1, + "bbox": [ + 76, + 492, + 303, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 518, + 266, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our proposed method on the two VLN datasets: Room-to-Room (R2R) [3] and Vision-and-Dialog Navigation (CVDN) [55], which are all based upon the MP3D indoor environments. The instructions in R2R are detailed descriptions of the navigation trajectory. There are four splits in 21,558 trajectory-instruction pairs of the R2R dataset: training (14,039), validation seen (1,021), validation unseen (2,325), and test unseen (4,173). Our grounded entity-landmark human annotations are based on the training, validation seen, and validation unseen splits. The instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. These kinds of instructions are frequently vague and unspecific. Therefore, the CVDN task is more challenging, serving to evaluate the generalization ability for new downstream tasks.", + "bbox": [ + 75, + 541, + 468, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For R2R, we report four evaluation metrics: Trajectory Length (TL); Navigation Error (NE ↓) - the average distance in meters between the agent's final position and the goal viewpoint; Success Rate (SR ↑) - the proportion of paths where the agent stopped within 3 meters of the goal viewpoint; Success rate weighted by Path Length (SPL ↑) [2]². SR and SPL are the recommended", + "bbox": [ + 75, + 768, + 468, + 875 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/15ce38c40fee0f11472c347373cc2443c108bc90643612d25fd8cb5ce60337fd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFeatureValidation Unseen
TLNE↓SR↑SPL↑
HAMT[7]none11.462.2965.760.9
st11.783.4267.362.6
smo12.133.2267.962.9
GELA (ours)none11.753.3369.263.4
st11.563.2669.364.2
smo11.733.1171.165.0
", + "bbox": [ + 524, + 89, + 870, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with the HAMT baseline on the R2R validation unseen split. Black denotes the best results.", + "bbox": [ + 498, + 232, + 890, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "key metrics. For CVDN, we use the average progress in meters of the agent towards the goal viewpoint as the primary evaluation metric, denoted as Goal Progress (GP $\\uparrow$ ).", + "bbox": [ + 496, + 285, + 890, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We typically conform to the architecture of [7] and its associated hyper-parameters. We conduct adaptive pre-training on our GEL-R2R training dataset and the augmented dataset from [14]. In the adaptive pre-training stage, we inherited the pre-training tasks to prevent knowledge forgetting. The GEL-R2R training set serves for both pre-training tasks and adaptive pre-training tasks. Due to the lack of entity-landmark grounding annotations, the augmented dataset serves only for pre-training tasks. We set $\\lambda = 1.0$ in Eq. (4), and $\\alpha = 1.0$ , $\\beta = 1.0$ , $\\gamma = 1.0$ in Eq. (8). And we set $\\tau$ to 0.07 in Eq. (5) and Eq. (6) following [25]. And We adopt adaptive pre-training for 200k iterations with a learning rate of 5e-5 by using 2 NVIDIA RTX 3090 GPUs, and the batch size for each GPU is set to 64. We adopt fine-tuning on the R2R and CVDN tasks respectively. For both R2R and CVDN tasks, we train the GELA model for 100k iterations with a learning rate of 1e-5 on a single NVIDIA RTX 3090 GPU. We set the batch size to 16 for the R2R task and 8 for the CVDN task.", + "bbox": [ + 496, + 330, + 892, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. The Effect of GELA Pre-training", + "text_level": 1, + "bbox": [ + 500, + 627, + 790, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we demonstrate the effects of our GELA pre-training paradigm for VLN models in terms of navigation performance and cross-modal representation.", + "bbox": [ + 496, + 651, + 890, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effect of GELA pre-training on navigation performance. Table 2 presents a comparison with the HAMT baseline model on the R2R validation unseen split under three fine-tuning settings. The results of the first three lines are obtained from [31]. The results show that our GELA model outperforms HAMT by a large margin under all three settings. Specifically, GELA achieves absolute improvements of $3.5\\%$ , $2\\%$ , and $3.2\\%$ in SR under none, st, and smo respectively. Furthermore, we show the single-step action prediction (SAP) accuracy of GELA and HAMT on the R2R validation unseen split during adaptive pre-training in the left panel of Figure 3. As the training progresses, we", + "bbox": [ + 496, + 696, + 892, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2We denote a $\\downarrow$ to demonstrate lower is better and an $\\uparrow$ to demonstrate", + "bbox": [ + 94, + 886, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "higher is better.", + "bbox": [ + 500, + 887, + 584, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12048", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/72075c5622ceb43b369be2911c3995e95f53461ab04d7faa47fad2f115acd7ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq[3]11.336.0139-8.397.8122-8.137.852018
SSM [58]14.703.10716220.704.32624520.404.576146
EnvDrop[54]11.003.99625910.705.22524811.665.235147
AuxRN[66]-3.337067-5.285550-5.155551
SEvol[6]11.973.56676312.263.99625713.404.136257
PREVALENT[14]10.323.67696510.194.71585310.515.305451
AirBERT[12]11.092.68757011.784.01625612.414.136257
RecBERT[20]11.132.90726812.013.93635712.354.096357
HOP[50]11.262.72757012.273.80645712.683.836459
REM[35]10.882.48757212.443.89645813.113.876559
HAMT[7]11.152.51767211.462.29666112.273.936560
GELA (ours)11.192.39767311.733.11716512.993.596762
", + "bbox": [ + 122, + 88, + 848, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with SoTA methods on the R2R dataset. The methods in the top group are trained from scratch. The methods in the second group are based on pre-training.", + "bbox": [ + 75, + 330, + 892, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db0760fb6bf398b0ed5754dfe4d4e4884b8fc8e62703cd674b48a7ec9ad8877a.jpg", + "image_caption": [ + "Figure 3. The effect of GELA pre-training. Single-step Action Prediction (SAP) accuracy (left) and Effective Attention (EA) score (right) of GELA and HAMT on the R2R validation splits." + ], + "image_footnote": [], + "bbox": [ + 78, + 383, + 269, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e6ff82e40c075fd263c0b4d958059aad6f0c70a808d2b5b05ebeba20522e16d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 383, + 464, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "observe a gradual increase in the performance gap between GELA and HAMT, with GELA and HAMT achieving prediction accuracies of approximately $78\\%$ and $75\\%$ , respectively. These significant improvements demonstrate that fine-grained cross-modal alignment between entity phrases and environmental landmarks is highly effective in enhancing agents' navigation performance.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of GELA pre-training on cross-modal representation. To directly demonstrate the effect of GELA pretraining on cross-modal semantic alignment, we analyze cross-modal attention weights at the last Transformer layer during inference. First, we visualize the attention weights normalized to a scale of 0 to 1 in a heatmap. We then utilize the annotations in GEL-R2R to mark tokens, which are involved in the grounded entity-landmark pairs, and calculate the attention weights received by each token from the corresponding tokens in the other modality, defined as the Effective Attention (EA) score. In the right panel of Figure 3, we compare the average EA score of each marked token in", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GELA and HAMT on the R2R validation seen and unseen splits, where $e2l$ (resp. $l2e$ ) represents the attention of the entity token to the landmark token (resp. the attention of the landmark token to the entity token). Among all comparisons, GELA achieves significantly higher EA scores than HAMT (with a p-value in the two-sample t-test approaching 0), indicating that the grounded entity-landmark pairs have a stronger correspondence in GELA. While HAMT has a strong ability for uni-modal representation based on the large-scale transformer and end-to-end pre-training, it can be confused about matching the mentioned landmarks with visual observations. The explicit supervision for entity-landmark level alignment in GELA can improve the cross-modal representation and substantially enhance the agent's decision-making ability during navigation. Qualitative examples are provided in the supplementary material.", + "bbox": [ + 496, + 383, + 893, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison to SoTA", + "text_level": 1, + "bbox": [ + 498, + 635, + 697, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Room-to-Room: R2R. Table 3 reports the performance comparison of various models on the R2R benchmark. The ensemble models, SE-Mixed [51] and EnvEdit [31], are excluded for a fair comparison. Our proposed GELA model outperforms all the other models in primary metrics (SR and SPL) across all the dataset splits. This indicates that GELA agents can navigate more accurately and efficiently than the other models. Specifically, on the validation seen split, GELA performs comparably to the SoTA HAMT. However, on the validation unseen and test unseen splits, GELA outperforms HAMT by absolute $5\\%$ and $2\\%$ (SR) and $4\\%$ and $2\\%$ (SPL) improvements, respectively. This demonstrates that GELA exhibits better generalization ability to unseen environments, which can be attributed to its stronger capability in capturing the semantic features of landmarks after entity-landmark grounding learning.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12049", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea5978fffb2f5241a5c64ad90a2942a2ea404d78dce48d482f87bc822db29ced.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsVal SeenVal UnseenTest Unseen
Seq2Seq[55]5.922.102.35
CMN[68]7.052.972.95
PREVALENT[14]-3.152.44
VISITRON[52]5.113.253.11
ORIST[61]-3.553.15
HOP[50]-4.373.31
MT-RCM+EnvAg[61]5.074.653.91
HAMT[7]6.915.135.58
GELA (ours)8.575.865.87
", + "bbox": [ + 81, + 89, + 467, + 242 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1f2c581fecbc1ff9dbcc7ae332017f6d10c12cca3d730ae62bc3758cd66f48e3.jpg", + "table_caption": [ + "Table 4. Comparison with SoTA methods on the CVDN dataset." + ], + "table_footnote": [], + "table_body": "
MethodsTaskTLNE↓SR↑SPL↑
0None11.462.2965.760.9
1EPP11.833.6365.860.7
2LBP12.713.7165.660.1
3ELSA11.733.5866.361.7
4EPP+LBP12.103.5967.061.4
5EPP+LBP+ELSA11.753.3369.263.4
", + "bbox": [ + 81, + 279, + 467, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vision-and-Dialog Navigation: CVDN. Table 4 presents the results on the CVDN benchmark, which utilizes Goal Progress (GP) in meters as the key performance metric. During training, we use a mixture of two types of demonstrations (navigator and oracle) as supervision for the trajectory. The results indicate that GELA consistently outperforms the other models on both the validation and test unseen environments. Particularly, our model achieves up to 1.66-meter improvement over the SoTA model HAMT on the validation seen split. Additionally, GELA outperforms the SoTA results by 0.73 meters and 0.29 meters on validation unseen and test splits, respectively. These results clearly demonstrate the GELA model is effective and generalizable to different types of instructions in more challenging VLN downstream tasks.", + "bbox": [ + 75, + 460, + 467, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 710, + 236, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Pre-training Objective. To evaluate the effectiveness of different adaptive pre-training objectives, we conduct an ablation study on R2R validation unseen split under none setting. The results are presented in Table 5. Model 0 represents the HAMT baseline, while Models 1-3 show the results of combining the original proxy tasks in [7] with EPP, LBP, and ELSA, respectively. The results demonstrate that ELSA can further improve navigation performance, while the other two unidirectional prediction objectives solely based on positional information do not enhance the pre-trained model. This indicates that ELSA pro", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29a558390d9bc5059cc922d2543521b7acf137adbcf6446511cc744227093539.jpg", + "table_caption": [ + "Table 5. Ablation study of adaptive pre-training objectives on the R2R validation unseen split." + ], + "table_footnote": [], + "table_body": "
CategoryTLNE↓SR↑SPL↑
None11.462.2965.760.9
Scene11.843.4466.761.2
Object11.233.5466.861.9
All11.753.3369.263.4
", + "bbox": [ + 563, + 89, + 826, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study of landmark categories on the R2R validation unseen split.", + "bbox": [ + 498, + 186, + 890, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vides more effective supervision of cross-modal alignment by directly operating on the feature representations. Subsequently, we evaluate the effect of combining EPP and LBP. The results show that the combined performance (Model 4) is much better than the separate performance of Model 1 and Model 2, demonstrating that these two objectives are complementary. Finally, we combined all three objectives, and the results show that this performance (Model 5) improves even further. These findings highlight the importance of effective cross-modal alignment and the complementary nature of different adaptive pre-training objectives in improving navigation performance.", + "bbox": [ + 496, + 255, + 890, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Landmark Category. Table 6 presents our investigation into the impact of two category entities (object and scene) from GEL-R2R under none setting. Our results demonstrate that integrating both object and scene entities can enhance navigation performance. Of the two entity types, object entities have a greater influence on performance, as agents find it easier to comprehend individual objects compared to complex scenes. Furthermore, the best performance is achieved by combining both entity types, enabling agents to navigate more precisely by utilizing both object and scene landmarks. This improvement can be attributed to the fact that the amount of available data for each entity type is only half that of the complete dataset. Therefore, it is plausible that the performance of GELA has not reached saturation due to the limited volume of available data.", + "bbox": [ + 496, + 445, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 500, + 703, + 764, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GELA is a preliminary investigation into cross-modal alignment at the entity-landmark level in VLN based on our human-annotated GEL-R2R dataset. Due to the high cost of human annotation, the volume of GEL-R2R is relatively limited. In the future, we plan to expand the dataset using two cost-effective approaches: 1) developing a data augmentation model by using GEL-R2R to produce grounded entity-landmark annotations, and 2) incorporating large-scale phrase grounding datasets in the VL domain. Additionally, we intend to explore interpretable navigation based on our introduced dataset and method.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12050", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 90, + 194, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A well-performed VLN agent should have powerful competency in fine-grained cross-modal semantic alignment between entities and landmarks. In this paper, we introduce the manually annotated grounded entity-landmark dataset GEL-R2R, which provides powerful cross-modal alignment for VLN at the entity-landmark level. We then adopt three grounded entity-landmark adaptive pre-training objectives based on GEL-R2R to facilitate cross-modal semantic alignment learning under explicit supervision. Comprehensive experimental results on two downstream VLN tasks, R2R and CVDN, demonstrate the effectiveness and generalizability of our proposed model, GELA.", + "bbox": [ + 75, + 114, + 472, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 313, + 246, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank ICCV reviewers for their constructive suggestions. This work is supported in part by the National Natural Science Foundation of China under Grant 62076250, Grant 61703407, and Grant 61901505.", + "bbox": [ + 76, + 339, + 470, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 416, + 174, + 431 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In MM '21: ACM Multimedia Conference, pages 5101-5109, 2021. 2", + "[2] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir R. Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6", + "[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 6, 7", + "[4] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 2, 3", + "[5] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 2", + "[6] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In CVPR, pages 15429-15438, 2022. 7", + "[7] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 1, 2, 3, 5, 6, 7, 8", + "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local:" + ], + "bbox": [ + 86, + 441, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16516-16526, 2022. 2", + "[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In ICCV, pages 1749-1759, 2021. 5", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In *NAACL-HLT*, pages 4171–4186, 2019. 3, 5", + "[11] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, pages 3318–3329, 2018. 1, 2", + "[12] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1614-1623, 2021. 2, 7", + "[13] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 3", + "[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, pages 13134-13143, 2020. 2, 5, 6, 7, 8", + "[15] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In NeurIPS, pages 652-663, 2021. 1", + "[16] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, pages 11773-11781, 2020. 2", + "[17] Yicong Hong, Cristian Rodriguez Opazo, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 2", + "[18] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In EMNLP, pages 3360-3376, 2020. 1, 3", + "[19] Yicong Hong, Zun Wang, Qi Wu, and Stephen Gould. Bridging the gap between learning in discrete and continuous environments for vision-and-language navigation. In CVPR, pages 15418-15428, 2022. 2", + "[20] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 6, 7", + "[21] Haoshuo Huang, Vihan Jain, Harsh Mehta, Alexander Ku, Gabriel Magalhães, Jason Baldridge, and Eugene Ie. Transferable representation learning in vision-and-language navigation. In ICCV, pages 7403-7412, 2019. 2", + "[22] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. In NeurIPS Workshop, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, pages 1862-1872, 2019. 2", + "[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Trans. Assoc. Comput. Linguistics, 8:64-77, 2020. 3", + "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1760-1770, 2021. 2, 5, 6", + "[26] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha S. Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In CVPR, pages 6741-6749, 2019. 2", + "[27] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, pages 14718-14728, 2021. 2", + "[28] Jacob Krantz and Stefan Lee. Sim-2-sim transfer for vision-and-language navigation in continuous environments. In ECCV, volume 13699 of Lecture Notes in Computer Science, pages 588–603, 2022. 2", + "[29] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, volume 12373 of Lecture Notes in Computer Science, pages 104-120, 2020. 2", + "[30] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 2", + "[31] Jialu Li, Hao Tan, and Mohit Bansal. Envedit: Environment editing for vision-and-language navigation. In CVPR, pages 15386-15396, 2022. 6, 7", + "[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 2", + "[33] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955–10965, 2022. 2", + "[34] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. REVE-CE: remote embodied visual referring expression in continuous environment. IEEE Robotics Autom. Lett., 7(2):1494–1501, 2022. 2", + "[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, pages 1624-1634, 2021. 7", + "[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 5", + "[38] Panzhong Lu, Xin Zhang, Meishan Zhang, and Min Zhang. Extending phrase grounding with pronouns in visual dialogues. In Proc. of the EMNLP, pages 7614-7625, Abu Dhabi, United Arab Emirates, Dec. 2022. Association for Computational Linguistics. 2", + "[39] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 2", + "[40] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, volume 12351 of Lecture Notes in Computer Science, pages 259-274, 2020. 1, 5", + "[41] Shikib Mehri, Mihail Eric, and Dilek Hakkani-Tür. Dialogue: A natural language understanding benchmark for task-oriented dialogue. CoRR, abs/2009.13570, 2020. 3", + "[42] Dipendra Kumar Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3d environments with visual goal prediction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 2667-2678, 2018. 2", + "[43] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. SOAT: A scene- and object-aware transformer for vision-and-language navigation. In NeurIPS, pages 7357-7367, 2021. 1, 2", + "[44] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In CVPR, pages 12527-12537, 2019. 2", + "[45] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, pages 15922-15932, 2021. 2", + "[46] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. Int. J. Comput. Vis., 123(1):74-93, 2017. 2", + "[47] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential BERT for indoor vision-language navigation. In ICCV, pages 1635-1644, 2021. 1, 2", + "[48] Yuankai Qi, Zizheng Pan, Shengping Zhang, Anton van den Hengel, and Qi Wu. Object-and-action aware model for visual language navigation. In ECCV, volume 12355 of Lecture Notes in Computer Science, pages 303-317, 2020. 2", + "[49] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 2" + ], + "bbox": [ + 501, + 90, + 892, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12052", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pre-training for vision-and-language navigation. CoRR, abs/2203.11591, 2022. 1, 2, 5, 7, 8", + "[51] Wenda Qin, Teruhisa Mitsu, and Derry Wijaya. Explore the potential performance of vision-and-language navigation model: a snapshot ensemble method. CoRR, abs/2111.14267, 2021. 7", + "[52] Ayush Shrivastava, Karthik Gopalakrishnan, Yang Liu, Robinson Piramuthu, Gokhan Tur, Devi Parikh, and Dilek Hakkani-Tur. VISITRON: visual semantics-aligned interactively trained object-navigator. In Findings of ACL, pages 1984-1994, 2022. 8", + "[53] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In EMNLP-IJCNLP, pages 5099-5110, 2019. 4", + "[54] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In *NAACL-HLT*, pages 2610–2621, 2019. 2, 7", + "[55] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, volume 100 of Proceedings of Machine Learning Research, pages 394-406, 2019. 2, 6, 8", + "[56] Maxim Tkachenko, Mikhail Malyuk, Andrey Holmanyuk, and Nikolai Liubimov. Label Studio: Data labeling software, 2020-2022. Open source software available from https://github.com/heartxlabs label-studio.3", + "[57] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, pages 15450-15460, 2022. 1, 2", + "[58] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, pages 8455-8464, 2021. 7", + "[59] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, volume 12367 of Lecture Notes in Computer Science, pages 307-322, 2020. 2", + "[60] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 1, 2", + "[61] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In ECCV, volume 12369 of Lecture Notes in Computer Science, pages 413–430, 2020. 8", + "[62] Han Wu, Kun Xu, Linfeng Song, Lifeng Jin, Haisong Zhang, and Linqi Song. Domain-adaptive pretraining methods for dialogue understanding. In ACL/IJCNLP, pages 665-669, 2021. 3", + "[63] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In IJCAI, pages 890-897, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Jeffrey Zhao, Mahdis Mahdieh, Ye Zhang, Yuan Cao, and Yonghui Wu. Effective sequence-to-sequence dialogue state tracking. In EMNLP, pages 7486-7493, 2021. 3", + "[65] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 2", + "[66] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, pages 10009-10019, 2020. 1, 7", + "[67] Wang Zhu, Hexiang Hu, Jiacheng Chen, Zhiwei Deng, Vihan Jain, Eugene Ie, and Fei Sha. Babywalk: Going farther in vision-and-language navigation by taking baby steps. In ACL, pages 2539-2556, 2020. 1", + "[68] Yi Zhu, Fengda Zhu, Zhaohuan Zhan, Bingqian Lin, Jianbin Jiao, Xiaojun Chang, and Xiaodan Liang. Vision-dialog navigation by exploring cross-modal memory. In CVPR, pages 10727-10736, 2020. 1, 8" + ], + "bbox": [ + 501, + 92, + 890, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12053", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_model.json b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..873658ad58e918d3e86cd0c2b5f99bd2052f4711 --- /dev/null +++ b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.227, + 0.131, + 0.745, + 0.176 + ], + "angle": 0, + "content": "Grounded Entity-Landmark Adaptive Pre-training for Vision-and-Language Navigation" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.203, + 0.862, + 0.257 + ], + "angle": 0, + "content": "Yibo Cui\\(^{1,2}\\), Liang Xie\\(^{1,2}\\), Yakun Zhang\\(^{1,2}\\), Meishan Zhang\\(^{3}\\), Ye Yan\\(^{1,2}\\), Erwei Yin\\(^{1,2}\\) \n\\(^{1}\\) Defense Innovation Institute, Chinese Academy of Military Science \n\\(^{2}\\) Tianjin Artificial Intelligence Innovation Center \\(^{3}\\) Harbin Institute of Technology (Shenzhen)" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.26, + 0.647, + 0.274 + ], + "angle": 0, + "content": "https://github.com/CSir1996/VLN-GELA" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.474, + 0.675 + ], + "angle": 0, + "content": "Cross-modal alignment is one key challenge for Vision- and Language Navigation (VLN). Most existing studies concentrate on mapping the global instruction or single sub-instruction to the corresponding trajectory. However, another critical problem of achieving fine-grained alignment at the entity level is seldom considered. To address this problem, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN tasks. To achieve the adaptive pre-training paradigm, we first introduce grounded entity-landmark human annotations into the Room-to-Room (R2R) dataset, named GEL-R2R. Additionally, we adopt three grounded entity-landmark adaptive pretraining objectives: 1) entity phrase prediction, 2) landmark bounding box prediction, and 3) entity-landmark semantic alignment, which explicitly supervise the learning of fine-grained cross-modal alignment between entity phrases and environment landmarks. Finally, we validate our model on two downstream benchmarks: VLN with descriptive instructions (R2R) and dialogue instructions (CVDN). The comprehensive experiments show that our GELA model achieves state-of-the-art results on both tasks, demonstrating its effectiveness and generalizability." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Vision-and-Language Navigation (VLN) [3] is an important task in the Embodied Vision community, which has gained great attention [66, 40, 50]. It aims to ask an agent to reach the target location inside photo-realistic environments by following natural language instructions. In VLN, cross-modal alignment is one critical step to accurately make the action decision for the agent [60], since matching the mentioned landmarks (objects or scenes) with visual observations can help the comprehensive understanding of the en" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.31, + 0.88, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.538, + 0.895, + 0.595 + ], + "angle": 0, + "content": "Figure 1. An illustration of an embodied agent navigating in a 3D photo-realistic environment. The agent is expected to navigate based on environment landmarks that correspond to the entity phrase in the instruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.894, + 0.717 + ], + "angle": 0, + "content": "virements and instructions [60, 15, 68]. However, most of the available datasets could only provide the coarse-grained text-image alignment signals [15], i.e., instruction-level correspondences to the complete trajectories, where finer ones are required to learn the cross-modality alignment for well-performed navigation [18, 67, 47, 43, 57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "A large part of previous works [66, 11, 40, 50] has concentrated on the global alignment of instructions and trajectories, which matches instructions to the overall temporal visual trajectory via reinforcement learning [60], auxiliary reasoning [66], or transformer-based matching pre-training [7, 50, 40]. Others [18, 67, 15] have attempted to align sub-instructions and sub-trajectories locally. At this granularity, the agents are designed to segment long instruction [67, 18] and determine which sub-instruction to focus on. To guide local cross-modal alignment, He et al. [15] introduced the Landmark-RxR dataset by human-annotating sub-instructions and sub-trajectories alignment. They ad" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.23, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12043" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "vanced the granularity of visual-textual alignment, but the agent learning is still supervised at the sentence level. The alignment within (sub-)instructions should be considered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.471, + 0.381 + ], + "angle": 0, + "content": "Identifying environment landmarks that correspond to entities in the instruction, as shown in Figure 1, is the next step to refine alignment and improve navigation. Recent studies [47, 43, 57] investigated the same problem, suggesting the alignment of entities and object regions [47] or the scene- and object-aware transformer model [43]. But the object regions were not under the direct supervision of the corresponding entity phrases in the instruction. To generate navigation instructions enriched with landmark phrases, Wang et al. [57] introduced grounded landmark annotations using a dependency parser and weak supervision from the pose traces, leading to the entity-landmark alignments of insufficient precision. Therefore, a dataset with high-quality entity-landmark level grounding annotations and powerful supervision for entity-landmark level cross-modal alignment is highly desired for VLN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.519 + ], + "angle": 0, + "content": "To address the above limitations, we first enhance the Room-to-Room (R2R) dataset [3] by introducing additional high-quality grounded entity-landmark human annotations, known as the Grounded Entity-Landmark R2R dataset (GEL-R2R). The GEL-R2R dataset is annotated with abundant precisely matched entity-landmark pairs (as magenta text and bounding box illustrated in Figure 1), which could provide the VLN models with direct supervision of fine-grained cross-modal alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.471, + 0.763 + ], + "angle": 0, + "content": "To verify the value of our GEL-R2R dataset, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm to improve the learning of entity-landmark level alignment for the VLN pre-trained model. Specifically, we suggest three grounded entity-landmark adaptive pre-training objectives: 1) Entity Phrase Prediction (EPP), locating entity phrases that refer to environment landmarks from the instruction; 2) Landmark Bounding box Prediction (LBP), predicting the bounding box of environment landmarks that match with entity phrases; and 3) Entity-Landmark Semantic Alignment (ELSA), aligning the matched pairs of landmark patches and entity tokens in the feature space by contrastive loss. These three tasks explicitly equip the model with the ability to comprehend the entity-level grounding between human instructions and visual environment observations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Finally, we conduct extensive experiments on two downstream tasks to evaluate our proposed dataset GEL-R2R and adaptive pre-training methods GELA: Room-to-Room (R2R) and Vision-and-Dialog Navigation (CVDN). The instructions in R2R are fine-grained descriptions for the navigation trajectory, whereas the instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. We use HAMT [7] as the backbone VLN model for the GELA pre-training. The results demonstrate our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "suggested GELA achieves state-of-the-art (SoTA) performance in both seen and unseen environments of the above two benchmarks: \\(62\\%\\) SPL on R2R and 5.87 GP on CVDN." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.138, + 0.835, + 0.15 + ], + "angle": 0, + "content": "To summarize, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.152, + 0.892, + 0.196 + ], + "angle": 0, + "content": "- We construct a new dataset GEL-R2R, which is the first dataset with high-quality grounded entity-landmark human annotations in the VLN domain." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.198, + 0.892, + 0.271 + ], + "angle": 0, + "content": "- We propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN, explicitly supervising the models to learn fine-grained cross-modal semantic alignment between entity phrases and environment landmarks." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.273, + 0.892, + 0.318 + ], + "angle": 0, + "content": "- Our suggested GELA achieves state-of-the-art results on two challenging VLN downstream benchmarks, demonstrating its effectiveness and generalizability." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.152, + 0.892, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.642, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Vision-and-Language Navigation. Since the R2R benchmark was proposed by [3], studies of the VLN have made considerable progress in various aspects. VLN datasets have become increasingly diverse, existing in indoor [4, 30, 49, 55, 65] or outdoor [5, 16, 42] environments, and in discrete [3, 49] or continuous [29, 28, 34, 19] environments. The architecture of VLN models has become more complex, progressing from LSTM-based [11, 39] to transformer-based [14, 43, 45, 7]. To develop the competence of getting closer to the destination and to the ground truth path [23], VLN agents have been enhanced mainly through action strategy learning [63, 8, 48, 17, 21] and multi-modal representation learning [22, 59, 27, 26, 1]. Many excellent VLN models use imitation and reinforcement learning-based training paradigms for action strategy learning [60, 44, 54]. Recently, multi-modal transformer-based models that use effective joint representations for instructions and visual observations have achieved promising performance in VLN [12, 7, 50]. In this passage, we focus on incorporating more accurate and fine-grained cross-modal grounding information to improve the performance of VLN models by supervising the alignment between entity phrases and environment landmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Phrase-to-Region Grounding. Phrase-to-region grounding is an important task in the Vision-Language (VL) domain that involves localizing textual entities in an image, commonly abbreviated as phrase grounding [25, 33, 38]. This task has seen significant progress since the introduction of the Flickr30k Entities dataset [46] and has played a crucial role in learning fine-grained semantic visual representation. Various models such as VisualBERT [32], MDETR [25], and GLIP [33] have explored this task under different architectures, and grounded pre-training has shown to facilitate fine-grained semantic understanding. However, the task of fine-grained cross-modal semantic alignment in embodied VLN tasks is more challenging due" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12044" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.205 + ], + "angle": 0, + "content": "
AllObjectScene
TrajectoryInstructionPhraseP/IBoxB/IPhraseBoxP/BPhraseBoxP/B
Train467514039577884.121211468.6330756632482.0627032579022.14
Val Seen340102141964.1187418.56193940652.10225746762.07
Val Unseen775232594834.08202968.73432489652.075159113312.20
Total579017385714674.111501838.6437019762782.0634448739092.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.214, + 0.893, + 0.257 + ], + "angle": 0, + "content": "Table 1. Statistics on our GEL-R2R dataset. P/I (resp. B/I) denotes the average number of entity phrases contained (resp. landmark bounding boxes matched with entity phrases) in each instruction. And P/B represents the average number of landmark bounding boxes matched with each entity phrase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.468, + 0.356 + ], + "angle": 0, + "content": "to more complex scenes and various objects. In this paper, we introduce high-quality entity-landmark grounding human annotations into the R2R dataset and propose a grounded entity-landmark adaptive pre-training scheme for VLN pre-trained models to address this problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.468, + 0.613 + ], + "angle": 0, + "content": "Adaptive Pre-training. In the field of natural language processing, a pre-trained language model in the general domain is continuously pre-trained to learn the knowledge appropriate for a particular task or domain. This process is known as adaptive pre-training [13], which is conducted between pre-training and fine-tuning. Previous work [41, 64] gains consistent improvements by continuously pre-training an adaptive language model with the Masked Language Model (MLM) [10] objective. Furthermore, through three pre-training objectives — the MLM, Span Boundary Objective (SBO) [24], and Perturbation Masking Objective (PMO) — Wu et al. [62] improved the overall performance of a dialogue understanding model. In this work, we adopt adaptive pre-training into the area of VLN. Specifically, based on our human-annotated GEL-R2R dataset, we continuously train the SoTA pre-trained model [7] with three different supervised entity-landmark grounding objectives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.257, + 0.643 + ], + "angle": 0, + "content": "3. GEL-R2R Dataset" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.277, + 0.667 + ], + "angle": 0, + "content": "3.1. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.749 + ], + "angle": 0, + "content": "To establish alignment between entity phrases in instructions and their corresponding landmarks in the surrounding environment, we introduce GEL-R2R, a dataset that includes human annotations of grounded entity-landmark pairs. The construction process consists of five stages:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.84 + ], + "angle": 0, + "content": "(1) Raw data preparation. We collect panoramicas of the R2R dataset from the Matterport3D (MP3D) simulator [4] for each viewpoint. To improve the accuracy and efficiency of annotating landmarks, we mark the next action direction in panoramas and match each panorama with the corresponding sub-instruction based on [18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "(2) Annotation tool development. We develop a web-based annotation platform using the label-studio [56] to facilitate annotations. The annotation interface is shown in the supplementary material. Annotators first mark the entity" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.281, + 0.89, + 0.311 + ], + "angle": 0, + "content": "phrases in the instructions and then identify the corresponding landmarks in the images using the same labels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.312, + 0.89, + 0.37 + ], + "angle": 0, + "content": "(3) Annotation guideline standardization. Through pre-annotation, we standardized the annotation guideline by establishing four rules to ensure the consistency of the annotations. The four rules are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.372, + 0.89, + 0.402 + ], + "angle": 0, + "content": "- Alignment Rule: The entity phrase in instructions should match landmarks in panoramas precisely." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.402, + 0.89, + 0.445 + ], + "angle": 0, + "content": "- Free Text Rule: Free text instead of the class should be annotated, for instance, \"the white dining table\" instead of \"table\"." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.448, + 0.89, + 0.476 + ], + "angle": 0, + "content": "- Text Coreference Rule: Entity phrases referring to the same landmark are marked with the same label." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.478, + 0.89, + 0.522 + ], + "angle": 0, + "content": "- Unique Landmark Rule: For an entity phrase, only one corresponding landmark bounding box should be annotated in a panorama." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.372, + 0.89, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.89, + 0.613 + ], + "angle": 0, + "content": "(4) Data annotation and revision. We first select annotators by testing the participants using 50 test instruction-path pairs. The qualification process leaves us 43 qualified annotators to complete the annotation task. Following the annotation, our experts double-check the annotations and correct any errors to ensure compliance with the above four rules." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.89, + 0.718 + ], + "angle": 0, + "content": "(5) Data processing. In order to ensure the quality of the annotations, we first eliminate the incorrect annotations violating the above rules and then correct a few incorrect words. Finally, we established the GEL-R2R dataset by introducing additional grounded entity-landmark annotations to the R2R dataset. More details about data collection are provided in the supplementary material." + }, + { + "type": "list", + "bbox": [ + 0.499, + 0.523, + 0.89, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.708, + 0.743 + ], + "angle": 0, + "content": "3.2. Statistics and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents statistics for the train, validation seen, and validation unseen splits of GEL-R2R. The dataset contains a total of 71,467 entity phrases, with 57,788 in the train split, 4,196 in the validation seen split, and 9,483 in the validation unseen split. It also includes 150,183 landmark boxes, of which 121,146 are in the train split, 8,741 are in the validation seen split, and 20,296 are in the validation unseen split. On average, each instruction contains 4.1 entity phrases, which is roughly equal to the number of actions in a path. Furthermore, each instruction uses about" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12045" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.093, + 0.837, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Overview of GELA. The three adaptive pre-training objectives served by our scheme encourage: 1) the annotated entity to predict the bounding box of its corresponding landmark in panoramas, 2) the annotated landmark to predict the positions of its corresponding entity phrases, and 3) the cross-modal encoder to establish the semantic alignment between landmarks and entities representations. After grounded entity-landmark adaptive pre-training, the GELA model is fine-tuned for the two VLN downstream tasks, R2R and CVDN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.471, + 0.499 + ], + "angle": 0, + "content": "8.5 landmark boxes on average. As a result, the agent has access to a diverse set of landmarks that can aid in decision-making during each navigation episode." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.474, + 0.728 + ], + "angle": 0, + "content": "We compute statistics on the object and scene, taking into account their varying identification difficulties. In total, we gather 37019 object entity phrases and 34448 scene entity phrases, referring to 76278 object landmark boxes and 73909 scene landmark boxes, respectively. On average, there are about 2 landmark boxes corresponding to each exact entity phrase, whether for objects or scenes. This indicates that the same landmark appears in approximately two adjacent viewpoint panoramicas from different sizes and angles, which enables the agent to perceive the landmark from multiple views. In summary, the GEL-R2R dataset is the first to provide high-quality entity-landmark alignment human annotations, which are essential for the area of VLN, such as cross-modal representation learning, data augmentation, and interpretable navigation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.744, + 0.17, + 0.76 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.34, + 0.788 + ], + "angle": 0, + "content": "4.1. Problem Setup and Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Problem Setup. In VLN tasks, an embodied agent should move to a target location from a starting pose inside a 3D photo-realistic environment by following natural language navigable instructions. The instruction is a sequence of \\( L \\) words, denoted as \\( I = \\{w_0, \\dots, w_{L-1}\\} \\), and it guides the agent to traverse a connectivity graph \\( G \\) in order to reach the intended destination. At each step \\( t \\), the agent ac" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.59 + ], + "angle": 0, + "content": "quires a new panoramic visual observation \\( O_{t} = \\{o_{t,i}\\}_{i=1}^{36} \\) from neighboring environment, where \\( o_{t,i} = [v_{t,i};a_{t,i}] \\) consists RGB image \\( v_{t,i} \\) and orientation angle \\( a_{t,i} \\) of \\( i \\)-th view. The relative angles of \\( n \\) navigable viewpoints to the current viewpoint and the stop action constitute the action space at step \\( t \\), denoted as \\( A_{t} = \\{act_{t,1},\\dots ,act_{t,n},[STOP]\\} \\). The agent selects an action to move to a navigable viewpoint in \\( G \\) or stop at the current location. Once the agent stops, the navigation episode is completed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Overview. In this study, we propose a novel grounded entity-landmark adaptive pre-training scheme for a VLN pre-trained model, whose architecture is shown in Figure 2. The training scheme introduces explicit supervision on entity-landmark grounding to enhance fine-grained cross-modal representation learning. The scheme comprises two distinct stages: GELA pre-training on several proxy objectives and fine-tuning on the VLN downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.681, + 0.74 + ], + "angle": 0, + "content": "4.2. Pre-trained Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As illustrated in the bottom-left of Figure 2, our pretrained model is a fully transformer-based architecture for multi-modal decision-making, which is modified from the classical cross-modal model LXMERT [53]. The pretrained model takes three inputs: a global instruction \\( I \\), history information \\( H_{t} \\), and current panoramic visual observation \\( O_{t} \\), which are fed in a language encoder, a history encoder, and a vision encoder, respectively. And then the textual and visual modalities exchange the signals through cross-attention layers in the cross-modal en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12046" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "coder. Specifically, the visual modality is the concatenation of history and visual observation. Finally, the representation of tokens in instruction, history, and visual state is \\( Z = \\{z_{\\mathrm{cls}}, z_1, \\dots, z_L\\} \\), \\( Y_t = \\{y_{\\mathrm{cls}}, y_1, \\dots, y_{t-1}\\} \\), and \\( S_t = \\{s_1, \\dots, s_{36}, s_{\\mathrm{stop}}\\} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.471, + 0.334 + ], + "angle": 0, + "content": "To learn effective uni-modal and multi-modal representation, transformer-based models for VLN commonly undergo pre-training on in-domain datasets using several proxy tasks [14, 50, 7]. These tasks include common vision-language pre-training tasks as well as VLN-specific auxiliary tasks. In this study, we utilize five proxy tasks: Masked Language Modeling (MLM) [10], Masked Region Classification (MRC) [37], Instruction Trajectory Matching (ITM) [40], Single-step Action Prediction (SAP) [7], and Spatial Relationship Prediction (SPREL) [7]. Details are demonstrated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.266, + 0.365 + ], + "angle": 0, + "content": "4.3. GELA Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.553 + ], + "angle": 0, + "content": "As shown in the top-left of Figure 2, we adopt three adaptive pre-training objectives for learning entity-landmark level alignment, which are Entity Phrase Prediction (EPP), Landmark Bounding box Prediction (LBP), and Entity-Landmark Semantic Alignment (ELSA). The three proxy objectives are designed by mimicking the objectives used in pre-training vision-language Transformers, particularly with the goal of learning phase-level grounding information. LBP is adapted from the training objective of TransVG [9]. EPP and ELSA are extended from the soft token prediction and contrastive alignment of MDETR [25], respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.782 + ], + "angle": 0, + "content": "Entity Phrase Prediction (EPP). In this objective, we predict the positions of entity phrases that correspond to annotated environment landmarks. We first transfer the human-annotated entity location to a mask vector \\( M_z \\) of \\( L + 1 \\) dimensions, which is the same as \\( Z \\). Similarly, we transfer the human-annotated landmark bounding box to a 37-dim (the same as \\( S_t \\)) mask vector \\( M_s \\). Then we train the model to infer a uniform distribution over all token positions that refer to the corresponding landmark patches and supervise this process by the mask vector \\( M_z \\). Specifically, we average the patches representation of the landmark: \\( S_t \\times M_s^\\top \\), where \\( \\times \\) denotes the matrix product, obtaining a 768-dim vector. Then we put it into a two-layer feedforward network (FFN) to predict a distribution over the token positions in the instruction sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.794, + 0.47, + 0.813 + ], + "angle": 0, + "content": "\\[\nl o g i t s = \\operatorname {S o f t m a x} \\left(\\operatorname {F F N} \\left(S _ {t} \\times M _ {s} ^ {\\top}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.872 + ], + "angle": 0, + "content": "where the FFN maps a vector of 768-dim to a vector of \\( L + 1 \\) dimensions. Finally, we minimize the cross-entropy loss between logits and \\( M_z \\):" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E P P}} = \\text {C r o s s E n t r o p y} (\\operatorname {l o g i t s}, M _ {z}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Landmark Bounding Box Prediction (LBP). In this objective, we predict the bounding box of the landmark that matches with annotated entity phrases. We train the model to directly predict a 4-dim vector \\(box' = (x', y', w', h')\\) as the coordinates of the bounding box for each entity phrases and supervise this process by the human-annotated bounding box \\(box = (x, y, w, h)\\). Specifically, we first average the token embeddings of the entity phrase: \\(Z \\times M_z^\\top\\) and then we predict the coordinates of the box through a two-layer FFN and sigmoid function:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.252, + 0.892, + 0.271 + ], + "angle": 0, + "content": "\\[\nb o x ^ {\\prime} = \\operatorname {S i g m o i d} \\left(\\operatorname {F F N} \\left(Z \\times M _ {z} ^ {\\top}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.326 + ], + "angle": 0, + "content": "where the FFN maps a vector of 768-dim to a vector of 4-dim. Finally, we apply the smooth L1 loss and generalized IoU loss (GIoU loss) to optimize the box coordinates \\( M_z \\):" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.336, + 0.892, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L B P}} = \\mathcal {L} _ {\\text {s m o o t h} - l _ {1}} \\left(\\text {b o x}, \\text {b o x} ^ {\\prime}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {G I o U}} \\left(\\text {b o x}, \\text {b o x} ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.392 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weight coefficient of GIoU loss to balance these two losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Entity-Landmark Semantic Alignment (ELSA). While the above two unidirectional prediction tasks use positional information to match the entity and landmark, the entity-landmark semantic alignment loss enforces alignment between the hidden embeddings of the landmark and entity at the output of the cross-modal encoder. This additional contrastive alignment loss ensures that the representations of the landmark patches and the corresponding entity tokens are closer in the feature space compared to representations of unrelated tokens. This constraint is stronger than the above two unidirectional prediction losses as it directly operates on the representations and is not single-handedly based on the positional information. Specifically, inspired by InfoNCE loss [36], the objective is the mean of two contrastive losses as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.641, + 0.892, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {i = 0} ^ {3 6} \\frac {1}{\\left| Z _ {i} ^ {+} \\right|} \\sum_ {j \\in Z _ {i} ^ {+}} - \\log \\left(\\frac {\\exp \\left(s _ {i} ^ {\\top} z _ {j} / \\tau\\right)}{\\sum_ {k = 0} ^ {L} \\exp \\left(s _ {i} ^ {\\top} z _ {k} / \\tau\\right)}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.713, + 0.892, + 0.762 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {z} = \\sum_ {i = 0} ^ {L} \\frac {1}{\\left| S _ {i} ^ {+} \\right|} \\sum_ {j \\in s _ {i} ^ {+}} - \\log \\left(\\frac {\\exp \\left(z _ {i} ^ {\\top} s _ {j} / \\tau\\right)}{\\sum_ {k = 0} ^ {3 6} \\exp \\left(z _ {i} ^ {\\top} s _ {k} / \\tau\\right)}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.777, + 0.891, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {E L S A}} = \\left(\\mathcal {L} _ {s} + \\mathcal {L} _ {z}\\right) / 2, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.799, + 0.892, + 0.86 + ], + "angle": 0, + "content": "where \\( Z_{i}^{+} \\) is the token set to be matched with an annotated patch \\( s_i \\), and \\( S_{i}^{+} \\) is the landmark patch set that should be matched with an annotated token \\( z_{i} \\), and \\( \\tau \\) is a temperature parameter regulating attention to negative samples." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.86, + 0.866, + 0.875 + ], + "angle": 0, + "content": "Therefore, the full adaptive pre-training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {G E L A}} = \\alpha \\mathcal {L} _ {\\text {E P P}} + \\beta \\mathcal {L} _ {\\text {L B P}} + \\gamma \\mathcal {L} _ {\\text {E L S A}}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12047" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.322, + 0.108 + ], + "angle": 0, + "content": "4.4. Fine-tuning for VLN Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.341 + ], + "angle": 0, + "content": "The fine-tuning stage is illustrated in the right of Figure 2, where we generalize the GELA model for two VLN downstream tasks under the scheme of imitation learning (IL) and reinforcement learning (RL) by following previous work [20, 7]. IL supervises the agent to clone the behavior of the expert while RL encourages the agent to explore the trajectory according to the learning policy. Firstly, the GELA model navigates in the environment following the ground-truth action and generates gradients by IL. Secondly, using the same instruction, the model samples the action space to make decisions and uses RL to generate gradients. Finally, we combine the gradients and optimize the pre-trained model. For the CVDN task, our setting is the same to [7]. For the R2R task, we apply three settings with different data augmentation following [31]:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.343, + 0.453, + 0.357 + ], + "angle": 0, + "content": "- none: without data augmentation on environments." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.358, + 0.469, + 0.401 + ], + "angle": 0, + "content": "- st: with data augmentation by style transformation on original environments. Specifically, each discrete view of panoramas is transferred with a random style." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.403, + 0.469, + 0.478 + ], + "angle": 0, + "content": "- smo: with data augmentation by semantic class masking on synthesis environments, which are generated by image synthesis with the same style as the original environments. One semantic class is randomly masked out during image generation." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.343, + 0.469, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.493, + 0.304, + 0.51 + ], + "angle": 0, + "content": "5. Experiment and Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.519, + 0.267, + 0.535 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Datasets. We evaluate our proposed method on the two VLN datasets: Room-to-Room (R2R) [3] and Vision-and-Dialog Navigation (CVDN) [55], which are all based upon the MP3D indoor environments. The instructions in R2R are detailed descriptions of the navigation trajectory. There are four splits in 21,558 trajectory-instruction pairs of the R2R dataset: training (14,039), validation seen (1,021), validation unseen (2,325), and test unseen (4,173). Our grounded entity-landmark human annotations are based on the training, validation seen, and validation unseen splits. The instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. These kinds of instructions are frequently vague and unspecific. Therefore, the CVDN task is more challenging, serving to evaluate the generalization ability for new downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.469, + 0.875 + ], + "angle": 0, + "content": "Evaluation Metrics. For R2R, we report four evaluation metrics: Trajectory Length (TL); Navigation Error (NE ↓) - the average distance in meters between the agent's final position and the goal viewpoint; Success Rate (SR ↑) - the proportion of paths where the agent stopped within 3 meters of the goal viewpoint; Success rate weighted by Path Length (SPL ↑) [2]². SR and SPL are the recommended" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.871, + 0.224 + ], + "angle": 0, + "content": "
ModelFeatureValidation Unseen
TLNE↓SR↑SPL↑
HAMT[7]none11.462.2965.760.9
st11.783.4267.362.6
smo12.133.2267.962.9
GELA (ours)none11.753.3369.263.4
st11.563.2669.364.2
smo11.733.1171.165.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.233, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Table 2. Comparison with the HAMT baseline on the R2R validation unseen split. Black denotes the best results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.892, + 0.331 + ], + "angle": 0, + "content": "key metrics. For CVDN, we use the average progress in meters of the agent towards the goal viewpoint as the primary evaluation metric, denoted as Goal Progress (GP \\(\\uparrow\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.893, + 0.618 + ], + "angle": 0, + "content": "Implementation Details. We typically conform to the architecture of [7] and its associated hyper-parameters. We conduct adaptive pre-training on our GEL-R2R training dataset and the augmented dataset from [14]. In the adaptive pre-training stage, we inherited the pre-training tasks to prevent knowledge forgetting. The GEL-R2R training set serves for both pre-training tasks and adaptive pre-training tasks. Due to the lack of entity-landmark grounding annotations, the augmented dataset serves only for pre-training tasks. We set \\(\\lambda = 1.0\\) in Eq. (4), and \\(\\alpha = 1.0\\), \\(\\beta = 1.0\\), \\(\\gamma = 1.0\\) in Eq. (8). And we set \\(\\tau\\) to 0.07 in Eq. (5) and Eq. (6) following [25]. And We adopt adaptive pre-training for 200k iterations with a learning rate of 5e-5 by using 2 NVIDIA RTX 3090 GPUs, and the batch size for each GPU is set to 64. We adopt fine-tuning on the R2R and CVDN tasks respectively. For both R2R and CVDN tasks, we train the GELA model for 100k iterations with a learning rate of 1e-5 on a single NVIDIA RTX 3090 GPU. We set the batch size to 16 for the R2R task and 8 for the CVDN task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.792, + 0.644 + ], + "angle": 0, + "content": "5.2. The Effect of GELA Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.697 + ], + "angle": 0, + "content": "In this section, we demonstrate the effects of our GELA pre-training paradigm for VLN models in terms of navigation performance and cross-modal representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.893, + 0.879 + ], + "angle": 0, + "content": "The effect of GELA pre-training on navigation performance. Table 2 presents a comparison with the HAMT baseline model on the R2R validation unseen split under three fine-tuning settings. The results of the first three lines are obtained from [31]. The results show that our GELA model outperforms HAMT by a large margin under all three settings. Specifically, GELA achieves absolute improvements of \\(3.5\\%\\), \\(2\\%\\), and \\(3.2\\%\\) in SR under none, st, and smo respectively. Furthermore, we show the single-step action prediction (SAP) accuracy of GELA and HAMT on the R2R validation unseen split during adaptive pre-training in the left panel of Figure 3. As the training progresses, we" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2We denote a \\(\\downarrow\\) to demonstrate lower is better and an \\(\\uparrow\\) to demonstrate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.888, + 0.586, + 0.9 + ], + "angle": 0, + "content": "higher is better." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12048" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.849, + 0.323 + ], + "angle": 0, + "content": "
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq[3]11.336.0139-8.397.8122-8.137.852018
SSM [58]14.703.10716220.704.32624520.404.576146
EnvDrop[54]11.003.99625910.705.22524811.665.235147
AuxRN[66]-3.337067-5.285550-5.155551
SEvol[6]11.973.56676312.263.99625713.404.136257
PREVALENT[14]10.323.67696510.194.71585310.515.305451
AirBERT[12]11.092.68757011.784.01625612.414.136257
RecBERT[20]11.132.90726812.013.93635712.354.096357
HOP[50]11.262.72757012.273.80645712.683.836459
REM[35]10.882.48757212.443.89645813.113.876559
HAMT[7]11.152.51767211.462.29666112.273.936560
GELA (ours)11.192.39767311.733.11716512.993.596762
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.332, + 0.893, + 0.361 + ], + "angle": 0, + "content": "Table 3. Comparison with SoTA methods on the R2R dataset. The methods in the top group are trained from scratch. The methods in the second group are based on pre-training." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.385, + 0.271, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.385, + 0.465, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.577 + ], + "angle": 0, + "content": "Figure 3. The effect of GELA pre-training. Single-step Action Prediction (SAP) accuracy (left) and Effective Attention (EA) score (right) of GELA and HAMT on the R2R validation splits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.716 + ], + "angle": 0, + "content": "observe a gradual increase in the performance gap between GELA and HAMT, with GELA and HAMT achieving prediction accuracies of approximately \\(78\\%\\) and \\(75\\%\\), respectively. These significant improvements demonstrate that fine-grained cross-modal alignment between entity phrases and environmental landmarks is highly effective in enhancing agents' navigation performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The effect of GELA pre-training on cross-modal representation. To directly demonstrate the effect of GELA pretraining on cross-modal semantic alignment, we analyze cross-modal attention weights at the last Transformer layer during inference. First, we visualize the attention weights normalized to a scale of 0 to 1 in a heatmap. We then utilize the annotations in GEL-R2R to mark tokens, which are involved in the grounded entity-landmark pairs, and calculate the attention weights received by each token from the corresponding tokens in the other modality, defined as the Effective Attention (EA) score. In the right panel of Figure 3, we compare the average EA score of each marked token in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.895, + 0.628 + ], + "angle": 0, + "content": "GELA and HAMT on the R2R validation seen and unseen splits, where \\( e2l \\) (resp. \\( l2e \\)) represents the attention of the entity token to the landmark token (resp. the attention of the landmark token to the entity token). Among all comparisons, GELA achieves significantly higher EA scores than HAMT (with a p-value in the two-sample t-test approaching 0), indicating that the grounded entity-landmark pairs have a stronger correspondence in GELA. While HAMT has a strong ability for uni-modal representation based on the large-scale transformer and end-to-end pre-training, it can be confused about matching the mentioned landmarks with visual observations. The explicit supervision for entity-landmark level alignment in GELA can improve the cross-modal representation and substantially enhance the agent's decision-making ability during navigation. Qualitative examples are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.698, + 0.652 + ], + "angle": 0, + "content": "5.3. Comparison to SoTA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Room-to-Room: R2R. Table 3 reports the performance comparison of various models on the R2R benchmark. The ensemble models, SE-Mixed [51] and EnvEdit [31], are excluded for a fair comparison. Our proposed GELA model outperforms all the other models in primary metrics (SR and SPL) across all the dataset splits. This indicates that GELA agents can navigate more accurately and efficiently than the other models. Specifically, on the validation seen split, GELA performs comparably to the SoTA HAMT. However, on the validation unseen and test unseen splits, GELA outperforms HAMT by absolute \\(5\\%\\) and \\(2\\%\\) (SR) and \\(4\\%\\) and \\(2\\%\\) (SPL) improvements, respectively. This demonstrates that GELA exhibits better generalization ability to unseen environments, which can be attributed to its stronger capability in capturing the semantic features of landmarks after entity-landmark grounding learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12049" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.243 + ], + "angle": 0, + "content": "
MethodsVal SeenVal UnseenTest Unseen
Seq2Seq[55]5.922.102.35
CMN[68]7.052.972.95
PREVALENT[14]-3.152.44
VISITRON[52]5.113.253.11
ORIST[61]-3.553.15
HOP[50]-4.373.31
MT-RCM+EnvAg[61]5.074.653.91
HAMT[7]6.915.135.58
GELA (ours)8.575.865.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.253, + 0.462, + 0.267 + ], + "angle": 0, + "content": "Table 4. Comparison with SoTA methods on the CVDN dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.28, + 0.468, + 0.398 + ], + "angle": 0, + "content": "
MethodsTaskTLNE↓SR↑SPL↑
0None11.462.2965.760.9
1EPP11.833.6365.860.7
2LBP12.713.7165.660.1
3ELSA11.733.5866.361.7
4EPP+LBP12.103.5967.061.4
5EPP+LBP+ELSA11.753.3369.263.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.408, + 0.468, + 0.436 + ], + "angle": 0, + "content": "Table 5. Ablation study of adaptive pre-training objectives on the R2R validation unseen split." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.468, + 0.687 + ], + "angle": 0, + "content": "Vision-and-Dialog Navigation: CVDN. Table 4 presents the results on the CVDN benchmark, which utilizes Goal Progress (GP) in meters as the key performance metric. During training, we use a mixture of two types of demonstrations (navigator and oracle) as supervision for the trajectory. The results indicate that GELA consistently outperforms the other models on both the validation and test unseen environments. Particularly, our model achieves up to 1.66-meter improvement over the SoTA model HAMT on the validation seen split. Additionally, GELA outperforms the SoTA results by 0.73 meters and 0.29 meters on validation unseen and test splits, respectively. These results clearly demonstrate the GELA model is effective and generalizable to different types of instructions in more challenging VLN downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.238, + 0.725 + ], + "angle": 0, + "content": "5.4. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Adaptive Pre-training Objective. To evaluate the effectiveness of different adaptive pre-training objectives, we conduct an ablation study on R2R validation unseen split under none setting. The results are presented in Table 5. Model 0 represents the HAMT baseline, while Models 1-3 show the results of combining the original proxy tasks in [7] with EPP, LBP, and ELSA, respectively. The results demonstrate that ELSA can further improve navigation performance, while the other two unidirectional prediction objectives solely based on positional information do not enhance the pre-trained model. This indicates that ELSA pro" + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.09, + 0.828, + 0.179 + ], + "angle": 0, + "content": "
CategoryTLNE↓SR↑SPL↑
None11.462.2965.760.9
Scene11.843.4466.761.2
Object11.233.5466.861.9
All11.753.3369.263.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.188, + 0.892, + 0.217 + ], + "angle": 0, + "content": "Table 6. Ablation study of landmark categories on the R2R validation unseen split." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.439 + ], + "angle": 0, + "content": "vides more effective supervision of cross-modal alignment by directly operating on the feature representations. Subsequently, we evaluate the effect of combining EPP and LBP. The results show that the combined performance (Model 4) is much better than the separate performance of Model 1 and Model 2, demonstrating that these two objectives are complementary. Finally, we combined all three objectives, and the results show that this performance (Model 5) improves even further. These findings highlight the importance of effective cross-modal alignment and the complementary nature of different adaptive pre-training objectives in improving navigation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Landmark Category. Table 6 presents our investigation into the impact of two category entities (object and scene) from GEL-R2R under none setting. Our results demonstrate that integrating both object and scene entities can enhance navigation performance. Of the two entity types, object entities have a greater influence on performance, as agents find it easier to comprehend individual objects compared to complex scenes. Furthermore, the best performance is achieved by combining both entity types, enabling agents to navigate more precisely by utilizing both object and scene landmarks. This improvement can be attributed to the fact that the amount of available data for each entity type is only half that of the complete dataset. Therefore, it is plausible that the performance of GELA has not reached saturation due to the limited volume of available data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.704, + 0.765, + 0.719 + ], + "angle": 0, + "content": "5.5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "GELA is a preliminary investigation into cross-modal alignment at the entity-landmark level in VLN based on our human-annotated GEL-R2R dataset. Due to the high cost of human annotation, the volume of GEL-R2R is relatively limited. In the future, we plan to expand the dataset using two cost-effective approaches: 1) developing a data augmentation model by using GEL-R2R to produce grounded entity-landmark annotations, and 2) incorporating large-scale phrase grounding datasets in the VL domain. Additionally, we intend to explore interpretable navigation based on our introduced dataset and method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12050" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.196, + 0.106 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.299 + ], + "angle": 0, + "content": "A well-performed VLN agent should have powerful competency in fine-grained cross-modal semantic alignment between entities and landmarks. In this paper, we introduce the manually annotated grounded entity-landmark dataset GEL-R2R, which provides powerful cross-modal alignment for VLN at the entity-landmark level. We then adopt three grounded entity-landmark adaptive pre-training objectives based on GEL-R2R to facilitate cross-modal semantic alignment learning under explicit supervision. Comprehensive experimental results on two downstream VLN tasks, R2R and CVDN, demonstrate the effectiveness and generalizability of our proposed model, GELA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.314, + 0.248, + 0.332 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.471, + 0.4 + ], + "angle": 0, + "content": "We thank ICCV reviewers for their constructive suggestions. This work is supported in part by the National Natural Science Foundation of China under Grant 62076250, Grant 61703407, and Grant 61901505." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.417, + 0.175, + 0.433 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.443, + 0.47, + 0.498 + ], + "angle": 0, + "content": "[1] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In MM '21: ACM Multimedia Conference, pages 5101-5109, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.5, + 0.472, + 0.569 + ], + "angle": 0, + "content": "[2] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir R. Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.571, + 0.47, + 0.641 + ], + "angle": 0, + "content": "[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.643, + 0.47, + 0.711 + ], + "angle": 0, + "content": "[4] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.713, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[5] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.771, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[6] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In CVPR, pages 15429-15438, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.816, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local:" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.443, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16516-16526, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.123, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In ICCV, pages 1749-1759, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.222 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In *NAACL-HLT*, pages 4171–4186, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[11] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, pages 3318–3329, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[12] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1614-1623, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[13] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.412, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, pages 13134-13143, 2020. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[15] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In NeurIPS, pages 652-663, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[16] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, pages 11773-11781, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[17] Yicong Hong, Cristian Rodriguez Opazo, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[18] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In EMNLP, pages 3360-3376, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[19] Yicong Hong, Zun Wang, Qi Wu, and Stephen Gould. Bridging the gap between learning in discrete and continuous environments for vision-and-language navigation. In CVPR, pages 15418-15428, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[20] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[21] Haoshuo Huang, Vihan Jain, Harsh Mehta, Alexander Ku, Gabriel Magalhães, Jason Baldridge, and Eugene Ie. Transferable representation learning in vision-and-language navigation. In ICCV, pages 7403-7412, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[22] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. In NeurIPS Workshop, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12051" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[23] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, pages 1862-1872, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Trans. Assoc. Comput. Linguistics, 8:64-77, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1760-1770, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[26] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha S. Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In CVPR, pages 6741-6749, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[27] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, pages 14718-14728, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[28] Jacob Krantz and Stefan Lee. Sim-2-sim transfer for vision-and-language navigation in continuous environments. In ECCV, volume 13699 of Lecture Notes in Computer Science, pages 588–603, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.501 + ], + "angle": 0, + "content": "[29] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, volume 12373 of Lecture Notes in Computer Science, pages 104-120, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[30] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[31] Jialu Li, Hao Tan, and Mohit Bansal. Envedit: Environment editing for vision-and-language navigation. In CVPR, pages 15386-15396, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[33] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955–10965, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[34] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. REVE-CE: remote embodied visual referring expression in continuous environment. IEEE Robotics Autom. Lett., 7(2):1494–1501, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, pages 1624-1634, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.146 + ], + "angle": 0, + "content": "[37] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.217 + ], + "angle": 0, + "content": "[38] Panzhong Lu, Xin Zhang, Meishan Zhang, and Min Zhang. Extending phrase grounding with pronouns in visual dialogues. In Proc. of the EMNLP, pages 7614-7625, Abu Dhabi, United Arab Emirates, Dec. 2022. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.272 + ], + "angle": 0, + "content": "[39] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.342 + ], + "angle": 0, + "content": "[40] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, volume 12351 of Lecture Notes in Computer Science, pages 259-274, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[41] Shikib Mehri, Mihail Eric, and Dilek Hakkani-Tür. Dialogue: A natural language understanding benchmark for task-oriented dialogue. CoRR, abs/2009.13570, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.893, + 0.481 + ], + "angle": 0, + "content": "[42] Dipendra Kumar Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3d environments with visual goal prediction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 2667-2678, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[43] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. SOAT: A scene- and object-aware transformer for vision-and-language navigation. In NeurIPS, pages 7357-7367, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.54, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[44] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In CVPR, pages 12527-12537, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.893, + 0.637 + ], + "angle": 0, + "content": "[45] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, pages 15922-15932, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.893, + 0.705 + ], + "angle": 0, + "content": "[46] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. Int. J. Comput. Vis., 123(1):74-93, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.893, + 0.774 + ], + "angle": 0, + "content": "[47] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential BERT for indoor vision-language navigation. In ICCV, pages 1635-1644, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.831 + ], + "angle": 0, + "content": "[48] Yuankai Qi, Zizheng Pan, Shengping Zhang, Anton van den Hengel, and Qi Wu. Object-and-action aware model for visual language navigation. In ECCV, volume 12355 of Lecture Notes in Computer Science, pages 303-317, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[49] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12052" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[50] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pre-training for vision-and-language navigation. CoRR, abs/2203.11591, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[51] Wenda Qin, Teruhisa Mitsu, and Derry Wijaya. Explore the potential performance of vision-and-language navigation model: a snapshot ensemble method. CoRR, abs/2111.14267, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[52] Ayush Shrivastava, Karthik Gopalakrishnan, Yang Liu, Robinson Piramuthu, Gokhan Tur, Devi Parikh, and Dilek Hakkani-Tur. VISITRON: visual semantics-aligned interactively trained object-navigator. In Findings of ACL, pages 1984-1994, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[53] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In EMNLP-IJCNLP, pages 5099-5110, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.469, + 0.373 + ], + "angle": 0, + "content": "[54] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In *NAACL-HLT*, pages 2610–2621, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[55] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, volume 100 of Proceedings of Machine Learning Research, pages 394-406, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.435, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[56] Maxim Tkachenko, Mikhail Malyuk, Andrey Holmanyuk, and Nikolai Liubimov. Label Studio: Data labeling software, 2020-2022. Open source software available from https://github.com/heartxlabs label-studio.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[57] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, pages 15450-15460, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[58] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, pages 8455-8464, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.605, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[59] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, volume 12367 of Lecture Notes in Computer Science, pages 307-322, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[60] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[61] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In ECCV, volume 12369 of Lecture Notes in Computer Science, pages 413–430, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[62] Han Wu, Kun Xu, Linfeng Song, Lifeng Jin, Haisong Zhang, and Linqi Song. Domain-adaptive pretraining methods for dialogue understanding. In ACL/IJCNLP, pages 665-669, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[63] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In IJCAI, pages 890-897, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[64] Jeffrey Zhao, Mahdis Mahdieh, Ye Zhang, Yuan Cao, and Yonghui Wu. Effective sequence-to-sequence dialogue state tracking. In EMNLP, pages 7486-7493, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[65] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[66] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, pages 10009-10019, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[67] Wang Zhu, Hexiang Hu, Jiacheng Chen, Zhiwei Deng, Vihan Jain, Eugene Ie, and Fei Sha. Babywalk: Going farther in vision-and-language navigation by taking baby steps. In ACL, pages 2539-2556, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[68] Yi Zhu, Fengda Zhu, Zhaohuan Zhan, Bingqian Lin, Jianbin Jiao, Xiaojun Chang, and Xiaodan Liang. Vision-dialog navigation by exploring cross-modal memory. In CVPR, pages 10727-10736, 2020. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12053" + } + ] +] \ No newline at end of file diff --git a/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_origin.pdf b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30356dbfa223b49dad7c28a801a6891922b4e019 --- /dev/null +++ b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/66797f14-cf57-48b6-98a1-7ce74385a6fb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:511b33c769fb77a0a0efad6bb50db951c5e5b24a4ce54a785b4c314be03d4966 +size 4537842 diff --git a/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/full.md b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0c3d0274c0427938aec684af56cdaa133ad2110 --- /dev/null +++ b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/full.md @@ -0,0 +1,315 @@ +# Grounded Entity-Landmark Adaptive Pre-training for Vision-and-Language Navigation + +Yibo Cui $^{1,2}$ , Liang Xie $^{1,2}$ , Yakun Zhang $^{1,2}$ , Meishan Zhang $^{3}$ , Ye Yan $^{1,2}$ , Erwei Yin $^{1,2}$ $^{1}$ Defense Innovation Institute, Chinese Academy of Military Science + $^{2}$ Tianjin Artificial Intelligence Innovation Center $^{3}$ Harbin Institute of Technology (Shenzhen) + +https://github.com/CSir1996/VLN-GELA + +# Abstract + +Cross-modal alignment is one key challenge for Vision- and Language Navigation (VLN). Most existing studies concentrate on mapping the global instruction or single sub-instruction to the corresponding trajectory. However, another critical problem of achieving fine-grained alignment at the entity level is seldom considered. To address this problem, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN tasks. To achieve the adaptive pre-training paradigm, we first introduce grounded entity-landmark human annotations into the Room-to-Room (R2R) dataset, named GEL-R2R. Additionally, we adopt three grounded entity-landmark adaptive pretraining objectives: 1) entity phrase prediction, 2) landmark bounding box prediction, and 3) entity-landmark semantic alignment, which explicitly supervise the learning of fine-grained cross-modal alignment between entity phrases and environment landmarks. Finally, we validate our model on two downstream benchmarks: VLN with descriptive instructions (R2R) and dialogue instructions (CVDN). The comprehensive experiments show that our GELA model achieves state-of-the-art results on both tasks, demonstrating its effectiveness and generalizability. + +# 1. Introduction + +Vision-and-Language Navigation (VLN) [3] is an important task in the Embodied Vision community, which has gained great attention [66, 40, 50]. It aims to ask an agent to reach the target location inside photo-realistic environments by following natural language instructions. In VLN, cross-modal alignment is one critical step to accurately make the action decision for the agent [60], since matching the mentioned landmarks (objects or scenes) with visual observations can help the comprehensive understanding of the en + +![](images/e0a06bdb3d822c15e1cfd882814e2deccc812240946ea9928f2f8d73bf9aca8e.jpg) +Figure 1. An illustration of an embodied agent navigating in a 3D photo-realistic environment. The agent is expected to navigate based on environment landmarks that correspond to the entity phrase in the instruction. + +virements and instructions [60, 15, 68]. However, most of the available datasets could only provide the coarse-grained text-image alignment signals [15], i.e., instruction-level correspondences to the complete trajectories, where finer ones are required to learn the cross-modality alignment for well-performed navigation [18, 67, 47, 43, 57]. + +A large part of previous works [66, 11, 40, 50] has concentrated on the global alignment of instructions and trajectories, which matches instructions to the overall temporal visual trajectory via reinforcement learning [60], auxiliary reasoning [66], or transformer-based matching pre-training [7, 50, 40]. Others [18, 67, 15] have attempted to align sub-instructions and sub-trajectories locally. At this granularity, the agents are designed to segment long instruction [67, 18] and determine which sub-instruction to focus on. To guide local cross-modal alignment, He et al. [15] introduced the Landmark-RxR dataset by human-annotating sub-instructions and sub-trajectories alignment. They ad + +vanced the granularity of visual-textual alignment, but the agent learning is still supervised at the sentence level. The alignment within (sub-)instructions should be considered. + +Identifying environment landmarks that correspond to entities in the instruction, as shown in Figure 1, is the next step to refine alignment and improve navigation. Recent studies [47, 43, 57] investigated the same problem, suggesting the alignment of entities and object regions [47] or the scene- and object-aware transformer model [43]. But the object regions were not under the direct supervision of the corresponding entity phrases in the instruction. To generate navigation instructions enriched with landmark phrases, Wang et al. [57] introduced grounded landmark annotations using a dependency parser and weak supervision from the pose traces, leading to the entity-landmark alignments of insufficient precision. Therefore, a dataset with high-quality entity-landmark level grounding annotations and powerful supervision for entity-landmark level cross-modal alignment is highly desired for VLN. + +To address the above limitations, we first enhance the Room-to-Room (R2R) dataset [3] by introducing additional high-quality grounded entity-landmark human annotations, known as the Grounded Entity-Landmark R2R dataset (GEL-R2R). The GEL-R2R dataset is annotated with abundant precisely matched entity-landmark pairs (as magenta text and bounding box illustrated in Figure 1), which could provide the VLN models with direct supervision of fine-grained cross-modal alignment. + +To verify the value of our GEL-R2R dataset, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm to improve the learning of entity-landmark level alignment for the VLN pre-trained model. Specifically, we suggest three grounded entity-landmark adaptive pre-training objectives: 1) Entity Phrase Prediction (EPP), locating entity phrases that refer to environment landmarks from the instruction; 2) Landmark Bounding box Prediction (LBP), predicting the bounding box of environment landmarks that match with entity phrases; and 3) Entity-Landmark Semantic Alignment (ELSA), aligning the matched pairs of landmark patches and entity tokens in the feature space by contrastive loss. These three tasks explicitly equip the model with the ability to comprehend the entity-level grounding between human instructions and visual environment observations. + +Finally, we conduct extensive experiments on two downstream tasks to evaluate our proposed dataset GEL-R2R and adaptive pre-training methods GELA: Room-to-Room (R2R) and Vision-and-Dialog Navigation (CVDN). The instructions in R2R are fine-grained descriptions for the navigation trajectory, whereas the instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. We use HAMT [7] as the backbone VLN model for the GELA pre-training. The results demonstrate our + +suggested GELA achieves state-of-the-art (SoTA) performance in both seen and unseen environments of the above two benchmarks: $62\%$ SPL on R2R and 5.87 GP on CVDN. + +To summarize, our contributions are three-fold: + +- We construct a new dataset GEL-R2R, which is the first dataset with high-quality grounded entity-landmark human annotations in the VLN domain. +- We propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN, explicitly supervising the models to learn fine-grained cross-modal semantic alignment between entity phrases and environment landmarks. +- Our suggested GELA achieves state-of-the-art results on two challenging VLN downstream benchmarks, demonstrating its effectiveness and generalizability. + +# 2. Related Work + +Vision-and-Language Navigation. Since the R2R benchmark was proposed by [3], studies of the VLN have made considerable progress in various aspects. VLN datasets have become increasingly diverse, existing in indoor [4, 30, 49, 55, 65] or outdoor [5, 16, 42] environments, and in discrete [3, 49] or continuous [29, 28, 34, 19] environments. The architecture of VLN models has become more complex, progressing from LSTM-based [11, 39] to transformer-based [14, 43, 45, 7]. To develop the competence of getting closer to the destination and to the ground truth path [23], VLN agents have been enhanced mainly through action strategy learning [63, 8, 48, 17, 21] and multi-modal representation learning [22, 59, 27, 26, 1]. Many excellent VLN models use imitation and reinforcement learning-based training paradigms for action strategy learning [60, 44, 54]. Recently, multi-modal transformer-based models that use effective joint representations for instructions and visual observations have achieved promising performance in VLN [12, 7, 50]. In this passage, we focus on incorporating more accurate and fine-grained cross-modal grounding information to improve the performance of VLN models by supervising the alignment between entity phrases and environment landmarks. + +Phrase-to-Region Grounding. Phrase-to-region grounding is an important task in the Vision-Language (VL) domain that involves localizing textual entities in an image, commonly abbreviated as phrase grounding [25, 33, 38]. This task has seen significant progress since the introduction of the Flickr30k Entities dataset [46] and has played a crucial role in learning fine-grained semantic visual representation. Various models such as VisualBERT [32], MDETR [25], and GLIP [33] have explored this task under different architectures, and grounded pre-training has shown to facilitate fine-grained semantic understanding. However, the task of fine-grained cross-modal semantic alignment in embodied VLN tasks is more challenging due + +
AllObjectScene
TrajectoryInstructionPhraseP/IBoxB/IPhraseBoxP/BPhraseBoxP/B
Train467514039577884.121211468.6330756632482.0627032579022.14
Val Seen340102141964.1187418.56193940652.10225746762.07
Val Unseen775232594834.08202968.73432489652.075159113312.20
Total579017385714674.111501838.6437019762782.0634448739092.15
+ +Table 1. Statistics on our GEL-R2R dataset. P/I (resp. B/I) denotes the average number of entity phrases contained (resp. landmark bounding boxes matched with entity phrases) in each instruction. And P/B represents the average number of landmark bounding boxes matched with each entity phrase. + +to more complex scenes and various objects. In this paper, we introduce high-quality entity-landmark grounding human annotations into the R2R dataset and propose a grounded entity-landmark adaptive pre-training scheme for VLN pre-trained models to address this problem. + +Adaptive Pre-training. In the field of natural language processing, a pre-trained language model in the general domain is continuously pre-trained to learn the knowledge appropriate for a particular task or domain. This process is known as adaptive pre-training [13], which is conducted between pre-training and fine-tuning. Previous work [41, 64] gains consistent improvements by continuously pre-training an adaptive language model with the Masked Language Model (MLM) [10] objective. Furthermore, through three pre-training objectives — the MLM, Span Boundary Objective (SBO) [24], and Perturbation Masking Objective (PMO) — Wu et al. [62] improved the overall performance of a dialogue understanding model. In this work, we adopt adaptive pre-training into the area of VLN. Specifically, based on our human-annotated GEL-R2R dataset, we continuously train the SoTA pre-trained model [7] with three different supervised entity-landmark grounding objectives. + +# 3. GEL-R2R Dataset + +# 3.1. Dataset Construction + +To establish alignment between entity phrases in instructions and their corresponding landmarks in the surrounding environment, we introduce GEL-R2R, a dataset that includes human annotations of grounded entity-landmark pairs. The construction process consists of five stages: + +(1) Raw data preparation. We collect panoramicas of the R2R dataset from the Matterport3D (MP3D) simulator [4] for each viewpoint. To improve the accuracy and efficiency of annotating landmarks, we mark the next action direction in panoramas and match each panorama with the corresponding sub-instruction based on [18]. +(2) Annotation tool development. We develop a web-based annotation platform using the label-studio [56] to facilitate annotations. The annotation interface is shown in the supplementary material. Annotators first mark the entity + +phrases in the instructions and then identify the corresponding landmarks in the images using the same labels. + +(3) Annotation guideline standardization. Through pre-annotation, we standardized the annotation guideline by establishing four rules to ensure the consistency of the annotations. The four rules are as follows: + +- Alignment Rule: The entity phrase in instructions should match landmarks in panoramas precisely. +- Free Text Rule: Free text instead of the class should be annotated, for instance, "the white dining table" instead of "table". +- Text Coreference Rule: Entity phrases referring to the same landmark are marked with the same label. +- Unique Landmark Rule: For an entity phrase, only one corresponding landmark bounding box should be annotated in a panorama. + +(4) Data annotation and revision. We first select annotators by testing the participants using 50 test instruction-path pairs. The qualification process leaves us 43 qualified annotators to complete the annotation task. Following the annotation, our experts double-check the annotations and correct any errors to ensure compliance with the above four rules. +(5) Data processing. In order to ensure the quality of the annotations, we first eliminate the incorrect annotations violating the above rules and then correct a few incorrect words. Finally, we established the GEL-R2R dataset by introducing additional grounded entity-landmark annotations to the R2R dataset. More details about data collection are provided in the supplementary material. + +# 3.2. Statistics and Analysis + +Table 1 presents statistics for the train, validation seen, and validation unseen splits of GEL-R2R. The dataset contains a total of 71,467 entity phrases, with 57,788 in the train split, 4,196 in the validation seen split, and 9,483 in the validation unseen split. It also includes 150,183 landmark boxes, of which 121,146 are in the train split, 8,741 are in the validation seen split, and 20,296 are in the validation unseen split. On average, each instruction contains 4.1 entity phrases, which is roughly equal to the number of actions in a path. Furthermore, each instruction uses about + +![](images/b96f6303d0c785c7f6689f039b172cf1fdb49b227d1b2130b6db16f230a1172b.jpg) +Figure 2. Overview of GELA. The three adaptive pre-training objectives served by our scheme encourage: 1) the annotated entity to predict the bounding box of its corresponding landmark in panoramas, 2) the annotated landmark to predict the positions of its corresponding entity phrases, and 3) the cross-modal encoder to establish the semantic alignment between landmarks and entities representations. After grounded entity-landmark adaptive pre-training, the GELA model is fine-tuned for the two VLN downstream tasks, R2R and CVDN. + +8.5 landmark boxes on average. As a result, the agent has access to a diverse set of landmarks that can aid in decision-making during each navigation episode. + +We compute statistics on the object and scene, taking into account their varying identification difficulties. In total, we gather 37019 object entity phrases and 34448 scene entity phrases, referring to 76278 object landmark boxes and 73909 scene landmark boxes, respectively. On average, there are about 2 landmark boxes corresponding to each exact entity phrase, whether for objects or scenes. This indicates that the same landmark appears in approximately two adjacent viewpoint panoramicas from different sizes and angles, which enables the agent to perceive the landmark from multiple views. In summary, the GEL-R2R dataset is the first to provide high-quality entity-landmark alignment human annotations, which are essential for the area of VLN, such as cross-modal representation learning, data augmentation, and interpretable navigation. + +# 4. Method + +# 4.1. Problem Setup and Overview + +Problem Setup. In VLN tasks, an embodied agent should move to a target location from a starting pose inside a 3D photo-realistic environment by following natural language navigable instructions. The instruction is a sequence of $L$ words, denoted as $I = \{w_0, \dots, w_{L-1}\}$ , and it guides the agent to traverse a connectivity graph $G$ in order to reach the intended destination. At each step $t$ , the agent ac + +quires a new panoramic visual observation $O_{t} = \{o_{t,i}\}_{i=1}^{36}$ from neighboring environment, where $o_{t,i} = [v_{t,i};a_{t,i}]$ consists RGB image $v_{t,i}$ and orientation angle $a_{t,i}$ of $i$ -th view. The relative angles of $n$ navigable viewpoints to the current viewpoint and the stop action constitute the action space at step $t$ , denoted as $A_{t} = \{act_{t,1},\dots ,act_{t,n},[STOP]\}$ . The agent selects an action to move to a navigable viewpoint in $G$ or stop at the current location. Once the agent stops, the navigation episode is completed. + +Overview. In this study, we propose a novel grounded entity-landmark adaptive pre-training scheme for a VLN pre-trained model, whose architecture is shown in Figure 2. The training scheme introduces explicit supervision on entity-landmark grounding to enhance fine-grained cross-modal representation learning. The scheme comprises two distinct stages: GELA pre-training on several proxy objectives and fine-tuning on the VLN downstream tasks. + +# 4.2. Pre-trained Model + +As illustrated in the bottom-left of Figure 2, our pretrained model is a fully transformer-based architecture for multi-modal decision-making, which is modified from the classical cross-modal model LXMERT [53]. The pretrained model takes three inputs: a global instruction $I$ , history information $H_{t}$ , and current panoramic visual observation $O_{t}$ , which are fed in a language encoder, a history encoder, and a vision encoder, respectively. And then the textual and visual modalities exchange the signals through cross-attention layers in the cross-modal en + +coder. Specifically, the visual modality is the concatenation of history and visual observation. Finally, the representation of tokens in instruction, history, and visual state is $Z = \{z_{\mathrm{cls}}, z_1, \dots, z_L\}$ , $Y_t = \{y_{\mathrm{cls}}, y_1, \dots, y_{t-1}\}$ , and $S_t = \{s_1, \dots, s_{36}, s_{\mathrm{stop}}\}$ , respectively. + +To learn effective uni-modal and multi-modal representation, transformer-based models for VLN commonly undergo pre-training on in-domain datasets using several proxy tasks [14, 50, 7]. These tasks include common vision-language pre-training tasks as well as VLN-specific auxiliary tasks. In this study, we utilize five proxy tasks: Masked Language Modeling (MLM) [10], Masked Region Classification (MRC) [37], Instruction Trajectory Matching (ITM) [40], Single-step Action Prediction (SAP) [7], and Spatial Relationship Prediction (SPREL) [7]. Details are demonstrated in the supplementary material. + +# 4.3. GELA Pre-training + +As shown in the top-left of Figure 2, we adopt three adaptive pre-training objectives for learning entity-landmark level alignment, which are Entity Phrase Prediction (EPP), Landmark Bounding box Prediction (LBP), and Entity-Landmark Semantic Alignment (ELSA). The three proxy objectives are designed by mimicking the objectives used in pre-training vision-language Transformers, particularly with the goal of learning phase-level grounding information. LBP is adapted from the training objective of TransVG [9]. EPP and ELSA are extended from the soft token prediction and contrastive alignment of MDETR [25], respectively. + +Entity Phrase Prediction (EPP). In this objective, we predict the positions of entity phrases that correspond to annotated environment landmarks. We first transfer the human-annotated entity location to a mask vector $M_z$ of $L + 1$ dimensions, which is the same as $Z$ . Similarly, we transfer the human-annotated landmark bounding box to a 37-dim (the same as $S_t$ ) mask vector $M_s$ . Then we train the model to infer a uniform distribution over all token positions that refer to the corresponding landmark patches and supervise this process by the mask vector $M_z$ . Specifically, we average the patches representation of the landmark: $S_t \times M_s^\top$ , where $\times$ denotes the matrix product, obtaining a 768-dim vector. Then we put it into a two-layer feedforward network (FFN) to predict a distribution over the token positions in the instruction sequence: + +$$ +l o g i t s = \operatorname {S o f t m a x} \left(\operatorname {F F N} \left(S _ {t} \times M _ {s} ^ {\top}\right)\right), \tag {1} +$$ + +where the FFN maps a vector of 768-dim to a vector of $L + 1$ dimensions. Finally, we minimize the cross-entropy loss between logits and $M_z$ : + +$$ +\mathcal {L} _ {\mathrm {E P P}} = \text {C r o s s E n t r o p y} (\operatorname {l o g i t s}, M _ {z}). \tag {2} +$$ + +Landmark Bounding Box Prediction (LBP). In this objective, we predict the bounding box of the landmark that matches with annotated entity phrases. We train the model to directly predict a 4-dim vector $box' = (x', y', w', h')$ as the coordinates of the bounding box for each entity phrases and supervise this process by the human-annotated bounding box $box = (x, y, w, h)$ . Specifically, we first average the token embeddings of the entity phrase: $Z \times M_z^\top$ and then we predict the coordinates of the box through a two-layer FFN and sigmoid function: + +$$ +b o x ^ {\prime} = \operatorname {S i g m o i d} \left(\operatorname {F F N} \left(Z \times M _ {z} ^ {\top}\right)\right), \tag {3} +$$ + +where the FFN maps a vector of 768-dim to a vector of 4-dim. Finally, we apply the smooth L1 loss and generalized IoU loss (GIoU loss) to optimize the box coordinates $M_z$ : + +$$ +\mathcal {L} _ {\mathrm {L B P}} = \mathcal {L} _ {\text {s m o o t h} - l _ {1}} \left(\text {b o x}, \text {b o x} ^ {\prime}\right) + \lambda \mathcal {L} _ {\mathrm {G I o U}} \left(\text {b o x}, \text {b o x} ^ {\prime}\right), \tag {4} +$$ + +where $\lambda$ is the weight coefficient of GIoU loss to balance these two losses. + +Entity-Landmark Semantic Alignment (ELSA). While the above two unidirectional prediction tasks use positional information to match the entity and landmark, the entity-landmark semantic alignment loss enforces alignment between the hidden embeddings of the landmark and entity at the output of the cross-modal encoder. This additional contrastive alignment loss ensures that the representations of the landmark patches and the corresponding entity tokens are closer in the feature space compared to representations of unrelated tokens. This constraint is stronger than the above two unidirectional prediction losses as it directly operates on the representations and is not single-handedly based on the positional information. Specifically, inspired by InfoNCE loss [36], the objective is the mean of two contrastive losses as follows: + +$$ +\mathcal {L} _ {s} = \sum_ {i = 0} ^ {3 6} \frac {1}{\left| Z _ {i} ^ {+} \right|} \sum_ {j \in Z _ {i} ^ {+}} - \log \left(\frac {\exp \left(s _ {i} ^ {\top} z _ {j} / \tau\right)}{\sum_ {k = 0} ^ {L} \exp \left(s _ {i} ^ {\top} z _ {k} / \tau\right)}\right), \tag {5} +$$ + +$$ +\mathcal {L} _ {z} = \sum_ {i = 0} ^ {L} \frac {1}{\left| S _ {i} ^ {+} \right|} \sum_ {j \in s _ {i} ^ {+}} - \log \left(\frac {\exp \left(z _ {i} ^ {\top} s _ {j} / \tau\right)}{\sum_ {k = 0} ^ {3 6} \exp \left(z _ {i} ^ {\top} s _ {k} / \tau\right)}\right), \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {E L S A}} = \left(\mathcal {L} _ {s} + \mathcal {L} _ {z}\right) / 2, \tag {7} +$$ + +where $Z_{i}^{+}$ is the token set to be matched with an annotated patch $s_i$ , and $S_{i}^{+}$ is the landmark patch set that should be matched with an annotated token $z_{i}$ , and $\tau$ is a temperature parameter regulating attention to negative samples. + +Therefore, the full adaptive pre-training objective is: + +$$ +\mathcal {L} _ {\text {G E L A}} = \alpha \mathcal {L} _ {\text {E P P}} + \beta \mathcal {L} _ {\text {L B P}} + \gamma \mathcal {L} _ {\text {E L S A}}. \tag {8} +$$ + +# 4.4. Fine-tuning for VLN Tasks + +The fine-tuning stage is illustrated in the right of Figure 2, where we generalize the GELA model for two VLN downstream tasks under the scheme of imitation learning (IL) and reinforcement learning (RL) by following previous work [20, 7]. IL supervises the agent to clone the behavior of the expert while RL encourages the agent to explore the trajectory according to the learning policy. Firstly, the GELA model navigates in the environment following the ground-truth action and generates gradients by IL. Secondly, using the same instruction, the model samples the action space to make decisions and uses RL to generate gradients. Finally, we combine the gradients and optimize the pre-trained model. For the CVDN task, our setting is the same to [7]. For the R2R task, we apply three settings with different data augmentation following [31]: + +- none: without data augmentation on environments. +- st: with data augmentation by style transformation on original environments. Specifically, each discrete view of panoramas is transferred with a random style. +- smo: with data augmentation by semantic class masking on synthesis environments, which are generated by image synthesis with the same style as the original environments. One semantic class is randomly masked out during image generation. + +# 5. Experiment and Results + +# 5.1. Experimental Setup + +Datasets. We evaluate our proposed method on the two VLN datasets: Room-to-Room (R2R) [3] and Vision-and-Dialog Navigation (CVDN) [55], which are all based upon the MP3D indoor environments. The instructions in R2R are detailed descriptions of the navigation trajectory. There are four splits in 21,558 trajectory-instruction pairs of the R2R dataset: training (14,039), validation seen (1,021), validation unseen (2,325), and test unseen (4,173). Our grounded entity-landmark human annotations are based on the training, validation seen, and validation unseen splits. The instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. These kinds of instructions are frequently vague and unspecific. Therefore, the CVDN task is more challenging, serving to evaluate the generalization ability for new downstream tasks. + +Evaluation Metrics. For R2R, we report four evaluation metrics: Trajectory Length (TL); Navigation Error (NE ↓) - the average distance in meters between the agent's final position and the goal viewpoint; Success Rate (SR ↑) - the proportion of paths where the agent stopped within 3 meters of the goal viewpoint; Success rate weighted by Path Length (SPL ↑) [2]². SR and SPL are the recommended + +
ModelFeatureValidation Unseen
TLNE↓SR↑SPL↑
HAMT[7]none11.462.2965.760.9
st11.783.4267.362.6
smo12.133.2267.962.9
GELA (ours)none11.753.3369.263.4
st11.563.2669.364.2
smo11.733.1171.165.0
+ +Table 2. Comparison with the HAMT baseline on the R2R validation unseen split. Black denotes the best results. + +key metrics. For CVDN, we use the average progress in meters of the agent towards the goal viewpoint as the primary evaluation metric, denoted as Goal Progress (GP $\uparrow$ ). + +Implementation Details. We typically conform to the architecture of [7] and its associated hyper-parameters. We conduct adaptive pre-training on our GEL-R2R training dataset and the augmented dataset from [14]. In the adaptive pre-training stage, we inherited the pre-training tasks to prevent knowledge forgetting. The GEL-R2R training set serves for both pre-training tasks and adaptive pre-training tasks. Due to the lack of entity-landmark grounding annotations, the augmented dataset serves only for pre-training tasks. We set $\lambda = 1.0$ in Eq. (4), and $\alpha = 1.0$ , $\beta = 1.0$ , $\gamma = 1.0$ in Eq. (8). And we set $\tau$ to 0.07 in Eq. (5) and Eq. (6) following [25]. And We adopt adaptive pre-training for 200k iterations with a learning rate of 5e-5 by using 2 NVIDIA RTX 3090 GPUs, and the batch size for each GPU is set to 64. We adopt fine-tuning on the R2R and CVDN tasks respectively. For both R2R and CVDN tasks, we train the GELA model for 100k iterations with a learning rate of 1e-5 on a single NVIDIA RTX 3090 GPU. We set the batch size to 16 for the R2R task and 8 for the CVDN task. + +# 5.2. The Effect of GELA Pre-training + +In this section, we demonstrate the effects of our GELA pre-training paradigm for VLN models in terms of navigation performance and cross-modal representation. + +The effect of GELA pre-training on navigation performance. Table 2 presents a comparison with the HAMT baseline model on the R2R validation unseen split under three fine-tuning settings. The results of the first three lines are obtained from [31]. The results show that our GELA model outperforms HAMT by a large margin under all three settings. Specifically, GELA achieves absolute improvements of $3.5\%$ , $2\%$ , and $3.2\%$ in SR under none, st, and smo respectively. Furthermore, we show the single-step action prediction (SAP) accuracy of GELA and HAMT on the R2R validation unseen split during adaptive pre-training in the left panel of Figure 3. As the training progresses, we + +
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq[3]11.336.0139-8.397.8122-8.137.852018
SSM [58]14.703.10716220.704.32624520.404.576146
EnvDrop[54]11.003.99625910.705.22524811.665.235147
AuxRN[66]-3.337067-5.285550-5.155551
SEvol[6]11.973.56676312.263.99625713.404.136257
PREVALENT[14]10.323.67696510.194.71585310.515.305451
AirBERT[12]11.092.68757011.784.01625612.414.136257
RecBERT[20]11.132.90726812.013.93635712.354.096357
HOP[50]11.262.72757012.273.80645712.683.836459
REM[35]10.882.48757212.443.89645813.113.876559
HAMT[7]11.152.51767211.462.29666112.273.936560
GELA (ours)11.192.39767311.733.11716512.993.596762
+ +Table 3. Comparison with SoTA methods on the R2R dataset. The methods in the top group are trained from scratch. The methods in the second group are based on pre-training. + +![](images/db0760fb6bf398b0ed5754dfe4d4e4884b8fc8e62703cd674b48a7ec9ad8877a.jpg) +Figure 3. The effect of GELA pre-training. Single-step Action Prediction (SAP) accuracy (left) and Effective Attention (EA) score (right) of GELA and HAMT on the R2R validation splits. + +![](images/6e6ff82e40c075fd263c0b4d958059aad6f0c70a808d2b5b05ebeba20522e16d.jpg) + +observe a gradual increase in the performance gap between GELA and HAMT, with GELA and HAMT achieving prediction accuracies of approximately $78\%$ and $75\%$ , respectively. These significant improvements demonstrate that fine-grained cross-modal alignment between entity phrases and environmental landmarks is highly effective in enhancing agents' navigation performance. + +The effect of GELA pre-training on cross-modal representation. To directly demonstrate the effect of GELA pretraining on cross-modal semantic alignment, we analyze cross-modal attention weights at the last Transformer layer during inference. First, we visualize the attention weights normalized to a scale of 0 to 1 in a heatmap. We then utilize the annotations in GEL-R2R to mark tokens, which are involved in the grounded entity-landmark pairs, and calculate the attention weights received by each token from the corresponding tokens in the other modality, defined as the Effective Attention (EA) score. In the right panel of Figure 3, we compare the average EA score of each marked token in + +GELA and HAMT on the R2R validation seen and unseen splits, where $e2l$ (resp. $l2e$ ) represents the attention of the entity token to the landmark token (resp. the attention of the landmark token to the entity token). Among all comparisons, GELA achieves significantly higher EA scores than HAMT (with a p-value in the two-sample t-test approaching 0), indicating that the grounded entity-landmark pairs have a stronger correspondence in GELA. While HAMT has a strong ability for uni-modal representation based on the large-scale transformer and end-to-end pre-training, it can be confused about matching the mentioned landmarks with visual observations. The explicit supervision for entity-landmark level alignment in GELA can improve the cross-modal representation and substantially enhance the agent's decision-making ability during navigation. Qualitative examples are provided in the supplementary material. + +# 5.3. Comparison to SoTA + +Room-to-Room: R2R. Table 3 reports the performance comparison of various models on the R2R benchmark. The ensemble models, SE-Mixed [51] and EnvEdit [31], are excluded for a fair comparison. Our proposed GELA model outperforms all the other models in primary metrics (SR and SPL) across all the dataset splits. This indicates that GELA agents can navigate more accurately and efficiently than the other models. Specifically, on the validation seen split, GELA performs comparably to the SoTA HAMT. However, on the validation unseen and test unseen splits, GELA outperforms HAMT by absolute $5\%$ and $2\%$ (SR) and $4\%$ and $2\%$ (SPL) improvements, respectively. This demonstrates that GELA exhibits better generalization ability to unseen environments, which can be attributed to its stronger capability in capturing the semantic features of landmarks after entity-landmark grounding learning. + +
MethodsVal SeenVal UnseenTest Unseen
Seq2Seq[55]5.922.102.35
CMN[68]7.052.972.95
PREVALENT[14]-3.152.44
VISITRON[52]5.113.253.11
ORIST[61]-3.553.15
HOP[50]-4.373.31
MT-RCM+EnvAg[61]5.074.653.91
HAMT[7]6.915.135.58
GELA (ours)8.575.865.87
+ +Table 4. Comparison with SoTA methods on the CVDN dataset. + +
MethodsTaskTLNE↓SR↑SPL↑
0None11.462.2965.760.9
1EPP11.833.6365.860.7
2LBP12.713.7165.660.1
3ELSA11.733.5866.361.7
4EPP+LBP12.103.5967.061.4
5EPP+LBP+ELSA11.753.3369.263.4
+ +Vision-and-Dialog Navigation: CVDN. Table 4 presents the results on the CVDN benchmark, which utilizes Goal Progress (GP) in meters as the key performance metric. During training, we use a mixture of two types of demonstrations (navigator and oracle) as supervision for the trajectory. The results indicate that GELA consistently outperforms the other models on both the validation and test unseen environments. Particularly, our model achieves up to 1.66-meter improvement over the SoTA model HAMT on the validation seen split. Additionally, GELA outperforms the SoTA results by 0.73 meters and 0.29 meters on validation unseen and test splits, respectively. These results clearly demonstrate the GELA model is effective and generalizable to different types of instructions in more challenging VLN downstream tasks. + +# 5.4. Ablation studies + +Adaptive Pre-training Objective. To evaluate the effectiveness of different adaptive pre-training objectives, we conduct an ablation study on R2R validation unseen split under none setting. The results are presented in Table 5. Model 0 represents the HAMT baseline, while Models 1-3 show the results of combining the original proxy tasks in [7] with EPP, LBP, and ELSA, respectively. The results demonstrate that ELSA can further improve navigation performance, while the other two unidirectional prediction objectives solely based on positional information do not enhance the pre-trained model. This indicates that ELSA pro + +Table 5. Ablation study of adaptive pre-training objectives on the R2R validation unseen split. + +
CategoryTLNE↓SR↑SPL↑
None11.462.2965.760.9
Scene11.843.4466.761.2
Object11.233.5466.861.9
All11.753.3369.263.4
+ +Table 6. Ablation study of landmark categories on the R2R validation unseen split. + +vides more effective supervision of cross-modal alignment by directly operating on the feature representations. Subsequently, we evaluate the effect of combining EPP and LBP. The results show that the combined performance (Model 4) is much better than the separate performance of Model 1 and Model 2, demonstrating that these two objectives are complementary. Finally, we combined all three objectives, and the results show that this performance (Model 5) improves even further. These findings highlight the importance of effective cross-modal alignment and the complementary nature of different adaptive pre-training objectives in improving navigation performance. + +Landmark Category. Table 6 presents our investigation into the impact of two category entities (object and scene) from GEL-R2R under none setting. Our results demonstrate that integrating both object and scene entities can enhance navigation performance. Of the two entity types, object entities have a greater influence on performance, as agents find it easier to comprehend individual objects compared to complex scenes. Furthermore, the best performance is achieved by combining both entity types, enabling agents to navigate more precisely by utilizing both object and scene landmarks. This improvement can be attributed to the fact that the amount of available data for each entity type is only half that of the complete dataset. Therefore, it is plausible that the performance of GELA has not reached saturation due to the limited volume of available data. + +# 5.5. Limitations and Future Work + +GELA is a preliminary investigation into cross-modal alignment at the entity-landmark level in VLN based on our human-annotated GEL-R2R dataset. Due to the high cost of human annotation, the volume of GEL-R2R is relatively limited. In the future, we plan to expand the dataset using two cost-effective approaches: 1) developing a data augmentation model by using GEL-R2R to produce grounded entity-landmark annotations, and 2) incorporating large-scale phrase grounding datasets in the VL domain. Additionally, we intend to explore interpretable navigation based on our introduced dataset and method. + +# 6. Conclusion + +A well-performed VLN agent should have powerful competency in fine-grained cross-modal semantic alignment between entities and landmarks. In this paper, we introduce the manually annotated grounded entity-landmark dataset GEL-R2R, which provides powerful cross-modal alignment for VLN at the entity-landmark level. We then adopt three grounded entity-landmark adaptive pre-training objectives based on GEL-R2R to facilitate cross-modal semantic alignment learning under explicit supervision. Comprehensive experimental results on two downstream VLN tasks, R2R and CVDN, demonstrate the effectiveness and generalizability of our proposed model, GELA. + +# 7. Acknowledgement + +We thank ICCV reviewers for their constructive suggestions. This work is supported in part by the National Natural Science Foundation of China under Grant 62076250, Grant 61703407, and Grant 61901505. + +# References + +[1] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In MM '21: ACM Multimedia Conference, pages 5101-5109, 2021. 2 +[2] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir R. Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6 +[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 6, 7 +[4] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 2, 3 +[5] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 2 +[6] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In CVPR, pages 15429-15438, 2022. 7 +[7] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 1, 2, 3, 5, 6, 7, 8 +[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: + +Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16516-16526, 2022. 2 +[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In ICCV, pages 1749-1759, 2021. 5 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In *NAACL-HLT*, pages 4171–4186, 2019. 3, 5 +[11] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, pages 3318–3329, 2018. 1, 2 +[12] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1614-1623, 2021. 2, 7 +[13] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 3 +[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, pages 13134-13143, 2020. 2, 5, 6, 7, 8 +[15] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In NeurIPS, pages 652-663, 2021. 1 +[16] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, pages 11773-11781, 2020. 2 +[17] Yicong Hong, Cristian Rodriguez Opazo, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 2 +[18] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In EMNLP, pages 3360-3376, 2020. 1, 3 +[19] Yicong Hong, Zun Wang, Qi Wu, and Stephen Gould. Bridging the gap between learning in discrete and continuous environments for vision-and-language navigation. In CVPR, pages 15418-15428, 2022. 2 +[20] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 6, 7 +[21] Haoshuo Huang, Vihan Jain, Harsh Mehta, Alexander Ku, Gabriel Magalhães, Jason Baldridge, and Eugene Ie. Transferable representation learning in vision-and-language navigation. In ICCV, pages 7403-7412, 2019. 2 +[22] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. In NeurIPS Workshop, 2019. 2 + +[23] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, pages 1862-1872, 2019. 2 +[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Trans. Assoc. Comput. Linguistics, 8:64-77, 2020. 3 +[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1760-1770, 2021. 2, 5, 6 +[26] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha S. Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In CVPR, pages 6741-6749, 2019. 2 +[27] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, pages 14718-14728, 2021. 2 +[28] Jacob Krantz and Stefan Lee. Sim-2-sim transfer for vision-and-language navigation in continuous environments. In ECCV, volume 13699 of Lecture Notes in Computer Science, pages 588–603, 2022. 2 +[29] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, volume 12373 of Lecture Notes in Computer Science, pages 104-120, 2020. 2 +[30] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 2 +[31] Jialu Li, Hao Tan, and Mohit Bansal. Envedit: Environment editing for vision-and-language navigation. In CVPR, pages 15386-15396, 2022. 6, 7 +[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 2 +[33] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955–10965, 2022. 2 +[34] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. REVE-CE: remote embodied visual referring expression in continuous environment. IEEE Robotics Autom. Lett., 7(2):1494–1501, 2022. 2 +[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, pages 1624-1634, 2021. 7 +[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 5 + +[37] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 5 +[38] Panzhong Lu, Xin Zhang, Meishan Zhang, and Min Zhang. Extending phrase grounding with pronouns in visual dialogues. In Proc. of the EMNLP, pages 7614-7625, Abu Dhabi, United Arab Emirates, Dec. 2022. Association for Computational Linguistics. 2 +[39] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 2 +[40] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, volume 12351 of Lecture Notes in Computer Science, pages 259-274, 2020. 1, 5 +[41] Shikib Mehri, Mihail Eric, and Dilek Hakkani-Tür. Dialogue: A natural language understanding benchmark for task-oriented dialogue. CoRR, abs/2009.13570, 2020. 3 +[42] Dipendra Kumar Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3d environments with visual goal prediction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 2667-2678, 2018. 2 +[43] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. SOAT: A scene- and object-aware transformer for vision-and-language navigation. In NeurIPS, pages 7357-7367, 2021. 1, 2 +[44] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In CVPR, pages 12527-12537, 2019. 2 +[45] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, pages 15922-15932, 2021. 2 +[46] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. Int. J. Comput. Vis., 123(1):74-93, 2017. 2 +[47] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential BERT for indoor vision-language navigation. In ICCV, pages 1635-1644, 2021. 1, 2 +[48] Yuankai Qi, Zizheng Pan, Shengping Zhang, Anton van den Hengel, and Qi Wu. Object-and-action aware model for visual language navigation. In ECCV, volume 12355 of Lecture Notes in Computer Science, pages 303-317, 2020. 2 +[49] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 2 + +[50] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pre-training for vision-and-language navigation. CoRR, abs/2203.11591, 2022. 1, 2, 5, 7, 8 +[51] Wenda Qin, Teruhisa Mitsu, and Derry Wijaya. Explore the potential performance of vision-and-language navigation model: a snapshot ensemble method. CoRR, abs/2111.14267, 2021. 7 +[52] Ayush Shrivastava, Karthik Gopalakrishnan, Yang Liu, Robinson Piramuthu, Gokhan Tur, Devi Parikh, and Dilek Hakkani-Tur. VISITRON: visual semantics-aligned interactively trained object-navigator. In Findings of ACL, pages 1984-1994, 2022. 8 +[53] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In EMNLP-IJCNLP, pages 5099-5110, 2019. 4 +[54] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In *NAACL-HLT*, pages 2610–2621, 2019. 2, 7 +[55] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, volume 100 of Proceedings of Machine Learning Research, pages 394-406, 2019. 2, 6, 8 +[56] Maxim Tkachenko, Mikhail Malyuk, Andrey Holmanyuk, and Nikolai Liubimov. Label Studio: Data labeling software, 2020-2022. Open source software available from https://github.com/heartxlabs label-studio.3 +[57] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, pages 15450-15460, 2022. 1, 2 +[58] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, pages 8455-8464, 2021. 7 +[59] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, volume 12367 of Lecture Notes in Computer Science, pages 307-322, 2020. 2 +[60] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 1, 2 +[61] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In ECCV, volume 12369 of Lecture Notes in Computer Science, pages 413–430, 2020. 8 +[62] Han Wu, Kun Xu, Linfeng Song, Lifeng Jin, Haisong Zhang, and Linqi Song. Domain-adaptive pretraining methods for dialogue understanding. In ACL/IJCNLP, pages 665-669, 2021. 3 +[63] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In IJCAI, pages 890-897, 2020. 2 + +[64] Jeffrey Zhao, Mahdis Mahdieh, Ye Zhang, Yuan Cao, and Yonghui Wu. Effective sequence-to-sequence dialogue state tracking. In EMNLP, pages 7486-7493, 2021. 3 +[65] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 2 +[66] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, pages 10009-10019, 2020. 1, 7 +[67] Wang Zhu, Hexiang Hu, Jiacheng Chen, Zhiwei Deng, Vihan Jain, Eugene Ie, and Fei Sha. Babywalk: Going farther in vision-and-language navigation by taking baby steps. In ACL, pages 2539-2556, 2020. 1 +[68] Yi Zhu, Fengda Zhu, Zhaohuan Zhan, Bingqian Lin, Jianbin Jiao, Xiaojun Chang, and Xiaodan Liang. Vision-dialog navigation by exploring cross-modal memory. In CVPR, pages 10727-10736, 2020. 1, 8 \ No newline at end of file diff --git a/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/images.zip b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ddc74129bf97eb6cfccc6ca965185dd0a61b2b5a --- /dev/null +++ b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dacbc7498ac4d8213562cfc7cbf3bdf4ba3b44ffe1dbad7d37deacd928cad05 +size 453934 diff --git a/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/layout.json b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..479e1ece5fad2d6778cb8a3fbfa27bf5c046493f --- /dev/null +++ b/2023/Grounded Entity-Landmark Adaptive Pre-Training for Vision-and-Language Navigation/layout.json @@ -0,0 +1,8464 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 103, + 455, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 455, + 139 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 455, + 139 + ], + "type": "text", + "content": "Grounded Entity-Landmark Adaptive Pre-training for Vision-and-Language Navigation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": "Yibo Cui" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Liang Xie" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Yakun Zhang" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Meishan Zhang" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Ye Yan" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Erwei Yin" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": " Defense Innovation Institute, Chinese Academy of Military Science \n" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": " Tianjin Artificial Intelligence Innovation Center " + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 527, + 203 + ], + "type": "text", + "content": " Harbin Institute of Technology (Shenzhen)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 205, + 395, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 205, + 395, + 217 + ], + "spans": [ + { + "bbox": [ + 198, + 205, + 395, + 217 + ], + "type": "text", + "content": "https://github.com/CSir1996/VLN-GELA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 270, + 290, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 534 + ], + "type": "text", + "content": "Cross-modal alignment is one key challenge for Vision- and Language Navigation (VLN). Most existing studies concentrate on mapping the global instruction or single sub-instruction to the corresponding trajectory. However, another critical problem of achieving fine-grained alignment at the entity level is seldom considered. To address this problem, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN tasks. To achieve the adaptive pre-training paradigm, we first introduce grounded entity-landmark human annotations into the Room-to-Room (R2R) dataset, named GEL-R2R. Additionally, we adopt three grounded entity-landmark adaptive pretraining objectives: 1) entity phrase prediction, 2) landmark bounding box prediction, and 3) entity-landmark semantic alignment, which explicitly supervise the learning of fine-grained cross-modal alignment between entity phrases and environment landmarks. Finally, we validate our model on two downstream benchmarks: VLN with descriptive instructions (R2R) and dialogue instructions (CVDN). The comprehensive experiments show that our GELA model achieves state-of-the-art results on both tasks, demonstrating its effectiveness and generalizability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "content": "Vision-and-Language Navigation (VLN) [3] is an important task in the Embodied Vision community, which has gained great attention [66, 40, 50]. It aims to ask an agent to reach the target location inside photo-realistic environments by following natural language instructions. In VLN, cross-modal alignment is one critical step to accurately make the action decision for the agent [60], since matching the mentioned landmarks (objects or scenes) with visual observations can help the comprehensive understanding of the en" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 245, + 538, + 423 + ], + "blocks": [ + { + "bbox": [ + 315, + 245, + 538, + 423 + ], + "lines": [ + { + "bbox": [ + 315, + 245, + 538, + 423 + ], + "spans": [ + { + "bbox": [ + 315, + 245, + 538, + 423 + ], + "type": "image", + "image_path": "e0a06bdb3d822c15e1cfd882814e2deccc812240946ea9928f2f8d73bf9aca8e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 426, + 547, + 471 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 471 + ], + "type": "text", + "content": "Figure 1. An illustration of an embodied agent navigating in a 3D photo-realistic environment. The agent is expected to navigate based on environment landmarks that correspond to the entity phrase in the instruction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 495, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 567 + ], + "type": "text", + "content": "virements and instructions [60, 15, 68]. However, most of the available datasets could only provide the coarse-grained text-image alignment signals [15], i.e., instruction-level correspondences to the complete trajectories, where finer ones are required to learn the cross-modality alignment for well-performed navigation [18, 67, 47, 43, 57]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "A large part of previous works [66, 11, 40, 50] has concentrated on the global alignment of instructions and trajectories, which matches instructions to the overall temporal visual trajectory via reinforcement learning [60], auxiliary reasoning [66], or transformer-based matching pre-training [7, 50, 40]. Others [18, 67, 15] have attempted to align sub-instructions and sub-trajectories locally. At this granularity, the agents are designed to segment long instruction [67, 18] and determine which sub-instruction to focus on. To guide local cross-modal alignment, He et al. [15] introduced the Landmark-RxR dataset by human-annotating sub-instructions and sub-trajectories alignment. They ad" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 140, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 140, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 140, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12043" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "vanced the granularity of visual-textual alignment, but the agent learning is still supervised at the sentence level. The alignment within (sub-)instructions should be considered." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 301 + ], + "type": "text", + "content": "Identifying environment landmarks that correspond to entities in the instruction, as shown in Figure 1, is the next step to refine alignment and improve navigation. Recent studies [47, 43, 57] investigated the same problem, suggesting the alignment of entities and object regions [47] or the scene- and object-aware transformer model [43]. But the object regions were not under the direct supervision of the corresponding entity phrases in the instruction. To generate navigation instructions enriched with landmark phrases, Wang et al. [57] introduced grounded landmark annotations using a dependency parser and weak supervision from the pose traces, leading to the entity-landmark alignments of insufficient precision. Therefore, a dataset with high-quality entity-landmark level grounding annotations and powerful supervision for entity-landmark level cross-modal alignment is highly desired for VLN." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 411 + ], + "type": "text", + "content": "To address the above limitations, we first enhance the Room-to-Room (R2R) dataset [3] by introducing additional high-quality grounded entity-landmark human annotations, known as the Grounded Entity-Landmark R2R dataset (GEL-R2R). The GEL-R2R dataset is annotated with abundant precisely matched entity-landmark pairs (as magenta text and bounding box illustrated in Figure 1), which could provide the VLN models with direct supervision of fine-grained cross-modal alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 288, + 604 + ], + "type": "text", + "content": "To verify the value of our GEL-R2R dataset, we propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm to improve the learning of entity-landmark level alignment for the VLN pre-trained model. Specifically, we suggest three grounded entity-landmark adaptive pre-training objectives: 1) Entity Phrase Prediction (EPP), locating entity phrases that refer to environment landmarks from the instruction; 2) Landmark Bounding box Prediction (LBP), predicting the bounding box of environment landmarks that match with entity phrases; and 3) Entity-Landmark Semantic Alignment (ELSA), aligning the matched pairs of landmark patches and entity tokens in the feature space by contrastive loss. These three tasks explicitly equip the model with the ability to comprehend the entity-level grounding between human instructions and visual environment observations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Finally, we conduct extensive experiments on two downstream tasks to evaluate our proposed dataset GEL-R2R and adaptive pre-training methods GELA: Room-to-Room (R2R) and Vision-and-Dialog Navigation (CVDN). The instructions in R2R are fine-grained descriptions for the navigation trajectory, whereas the instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. We use HAMT [7] as the backbone VLN model for the GELA pre-training. The results demonstrate our" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "suggested GELA achieves state-of-the-art (SoTA) performance in both seen and unseen environments of the above two benchmarks: " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " SPL on R2R and 5.87 GP on CVDN." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 109, + 511, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 109, + 511, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 109, + 511, + 118 + ], + "type": "text", + "content": "To summarize, our contributions are three-fold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 120, + 545, + 251 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 120, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 545, + 155 + ], + "type": "text", + "content": "- We construct a new dataset GEL-R2R, which is the first dataset with high-quality grounded entity-landmark human annotations in the VLN domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 156, + 545, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 156, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 545, + 214 + ], + "type": "text", + "content": "- We propose a novel Grounded Entity-Landmark Adaptive (GELA) pre-training paradigm for VLN, explicitly supervising the models to learn fine-grained cross-modal semantic alignment between entity phrases and environment landmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 216, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 216, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 216, + 545, + 251 + ], + "type": "text", + "content": "- Our suggested GELA achieves state-of-the-art results on two challenging VLN downstream benchmarks, demonstrating its effectiveness and generalizability." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 283, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 557 + ], + "type": "text", + "content": "Vision-and-Language Navigation. Since the R2R benchmark was proposed by [3], studies of the VLN have made considerable progress in various aspects. VLN datasets have become increasingly diverse, existing in indoor [4, 30, 49, 55, 65] or outdoor [5, 16, 42] environments, and in discrete [3, 49] or continuous [29, 28, 34, 19] environments. The architecture of VLN models has become more complex, progressing from LSTM-based [11, 39] to transformer-based [14, 43, 45, 7]. To develop the competence of getting closer to the destination and to the ground truth path [23], VLN agents have been enhanced mainly through action strategy learning [63, 8, 48, 17, 21] and multi-modal representation learning [22, 59, 27, 26, 1]. Many excellent VLN models use imitation and reinforcement learning-based training paradigms for action strategy learning [60, 44, 54]. Recently, multi-modal transformer-based models that use effective joint representations for instructions and visual observations have achieved promising performance in VLN [12, 7, 50]. In this passage, we focus on incorporating more accurate and fine-grained cross-modal grounding information to improve the performance of VLN models by supervising the alignment between entity phrases and environment landmarks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Phrase-to-Region Grounding. Phrase-to-region grounding is an important task in the Vision-Language (VL) domain that involves localizing textual entities in an image, commonly abbreviated as phrase grounding [25, 33, 38]. This task has seen significant progress since the introduction of the Flickr30k Entities dataset [46] and has played a crucial role in learning fine-grained semantic visual representation. Various models such as VisualBERT [32], MDETR [25], and GLIP [33] have explored this task under different architectures, and grounded pre-training has shown to facilitate fine-grained semantic understanding. However, the task of fine-grained cross-modal semantic alignment in embodied VLN tasks is more challenging due" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12044" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 162 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 162 + ], + "type": "table", + "html": "
AllObjectScene
TrajectoryInstructionPhraseP/IBoxB/IPhraseBoxP/BPhraseBoxP/B
Train467514039577884.121211468.6330756632482.0627032579022.14
Val Seen340102141964.1187418.56193940652.10225746762.07
Val Unseen775232594834.08202968.73432489652.075159113312.20
Total579017385714674.111501838.6437019762782.0634448739092.15
", + "image_path": "716acb5eec926d3d0ee0b955169d5a56901936bb313df764bc11bdf028d96b6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 546, + 203 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 546, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 546, + 203 + ], + "type": "text", + "content": "Table 1. Statistics on our GEL-R2R dataset. P/I (resp. B/I) denotes the average number of entity phrases contained (resp. landmark bounding boxes matched with entity phrases) in each instruction. And P/B represents the average number of landmark bounding boxes matched with each entity phrase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 222, + 286, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 286, + 281 + ], + "type": "text", + "content": "to more complex scenes and various objects. In this paper, we introduce high-quality entity-landmark grounding human annotations into the R2R dataset and propose a grounded entity-landmark adaptive pre-training scheme for VLN pre-trained models to address this problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 286, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 286, + 485 + ], + "type": "text", + "content": "Adaptive Pre-training. In the field of natural language processing, a pre-trained language model in the general domain is continuously pre-trained to learn the knowledge appropriate for a particular task or domain. This process is known as adaptive pre-training [13], which is conducted between pre-training and fine-tuning. Previous work [41, 64] gains consistent improvements by continuously pre-training an adaptive language model with the Masked Language Model (MLM) [10] objective. Furthermore, through three pre-training objectives — the MLM, Span Boundary Objective (SBO) [24], and Perturbation Masking Objective (PMO) — Wu et al. [62] improved the overall performance of a dialogue understanding model. In this work, we adopt adaptive pre-training into the area of VLN. Specifically, based on our human-annotated GEL-R2R dataset, we continuously train the SoTA pre-trained model [7] with three different supervised entity-landmark grounding objectives." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 496, + 157, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 157, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 157, + 509 + ], + "type": "text", + "content": "3. GEL-R2R Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 515, + 169, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 169, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 169, + 528 + ], + "type": "text", + "content": "3.1. Dataset Construction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 593 + ], + "type": "text", + "content": "To establish alignment between entity phrases in instructions and their corresponding landmarks in the surrounding environment, we introduce GEL-R2R, a dataset that includes human annotations of grounded entity-landmark pairs. The construction process consists of five stages:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "text", + "content": "(1) Raw data preparation. We collect panoramicas of the R2R dataset from the Matterport3D (MP3D) simulator [4] for each viewpoint. To improve the accuracy and efficiency of annotating landmarks, we mark the next action direction in panoramas and match each panorama with the corresponding sub-instruction based on [18]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "(2) Annotation tool development. We develop a web-based annotation platform using the label-studio [56] to facilitate annotations. The annotation interface is shown in the supplementary material. Annotators first mark the entity" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 222, + 544, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 222, + 544, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 544, + 246 + ], + "type": "text", + "content": "phrases in the instructions and then identify the corresponding landmarks in the images using the same labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 247, + 544, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 544, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 544, + 293 + ], + "type": "text", + "content": "(3) Annotation guideline standardization. Through pre-annotation, we standardized the annotation guideline by establishing four rules to ensure the consistency of the annotations. The four rules are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 294, + 544, + 413 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 317, + 294, + 544, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 544, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 544, + 318 + ], + "type": "text", + "content": "- Alignment Rule: The entity phrase in instructions should match landmarks in panoramas precisely." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 318, + 544, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 318, + 544, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 318, + 544, + 352 + ], + "type": "text", + "content": "- Free Text Rule: Free text instead of the class should be annotated, for instance, \"the white dining table\" instead of \"table\"." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 354, + 544, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 544, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 544, + 376 + ], + "type": "text", + "content": "- Text Coreference Rule: Entity phrases referring to the same landmark are marked with the same label." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 378, + 544, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 544, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 544, + 413 + ], + "type": "text", + "content": "- Unique Landmark Rule: For an entity phrase, only one corresponding landmark bounding box should be annotated in a panorama." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 414, + 544, + 568 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 305, + 414, + 544, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 544, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 544, + 485 + ], + "type": "text", + "content": "(4) Data annotation and revision. We first select annotators by testing the participants using 50 test instruction-path pairs. The qualification process leaves us 43 qualified annotators to complete the annotation task. Following the annotation, our experts double-check the annotations and correct any errors to ensure compliance with the above four rules." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 485, + 544, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 544, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 544, + 568 + ], + "type": "text", + "content": "(5) Data processing. In order to ensure the quality of the annotations, we first eliminate the incorrect annotations violating the above rules and then correct a few incorrect words. Finally, we established the GEL-R2R dataset by introducing additional grounded entity-landmark annotations to the R2R dataset. More details about data collection are provided in the supplementary material." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 575, + 433, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 433, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 433, + 588 + ], + "type": "text", + "content": "3.2. Statistics and Analysis" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Table 1 presents statistics for the train, validation seen, and validation unseen splits of GEL-R2R. The dataset contains a total of 71,467 entity phrases, with 57,788 in the train split, 4,196 in the validation seen split, and 9,483 in the validation unseen split. It also includes 150,183 landmark boxes, of which 121,146 are in the train split, 8,741 are in the validation seen split, and 20,296 are in the validation unseen split. On average, each instruction contains 4.1 entity phrases, which is roughly equal to the number of actions in a path. Furthermore, each instruction uses about" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12045" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 73, + 512, + 289 + ], + "blocks": [ + { + "bbox": [ + 82, + 73, + 512, + 289 + ], + "lines": [ + { + "bbox": [ + 82, + 73, + 512, + 289 + ], + "spans": [ + { + "bbox": [ + 82, + 73, + 512, + 289 + ], + "type": "image", + "image_path": "b96f6303d0c785c7f6689f039b172cf1fdb49b227d1b2130b6db16f230a1172b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 295, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 547, + 340 + ], + "type": "text", + "content": "Figure 2. Overview of GELA. The three adaptive pre-training objectives served by our scheme encourage: 1) the annotated entity to predict the bounding box of its corresponding landmark in panoramas, 2) the annotated landmark to predict the positions of its corresponding entity phrases, and 3) the cross-modal encoder to establish the semantic alignment between landmarks and entities representations. After grounded entity-landmark adaptive pre-training, the GELA model is fine-tuned for the two VLN downstream tasks, R2R and CVDN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 359, + 288, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 288, + 395 + ], + "type": "text", + "content": "8.5 landmark boxes on average. As a result, the agent has access to a diverse set of landmarks that can aid in decision-making during each navigation episode." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 290, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 290, + 576 + ], + "type": "text", + "content": "We compute statistics on the object and scene, taking into account their varying identification difficulties. In total, we gather 37019 object entity phrases and 34448 scene entity phrases, referring to 76278 object landmark boxes and 73909 scene landmark boxes, respectively. On average, there are about 2 landmark boxes corresponding to each exact entity phrase, whether for objects or scenes. This indicates that the same landmark appears in approximately two adjacent viewpoint panoramicas from different sizes and angles, which enables the agent to perceive the landmark from multiple views. In summary, the GEL-R2R dataset is the first to provide high-quality entity-landmark alignment human annotations, which are essential for the area of VLN, such as cross-modal representation learning, data augmentation, and interpretable navigation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 589, + 104, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 104, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 104, + 601 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 609, + 208, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 208, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 208, + 624 + ], + "type": "text", + "content": "4.1. Problem Setup and Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Problem Setup. In VLN tasks, an embodied agent should move to a target location from a starting pose inside a 3D photo-realistic environment by following natural language navigable instructions. The instruction is a sequence of " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " words, denoted as " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "I = \\{w_0, \\dots, w_{L-1}\\}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ", and it guides the agent to traverse a connectivity graph " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " in order to reach the intended destination. At each step " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ", the agent ac" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": "quires a new panoramic visual observation " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "O_{t} = \\{o_{t,i}\\}_{i=1}^{36}" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": " from neighboring environment, where " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "o_{t,i} = [v_{t,i};a_{t,i}]" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": " consists RGB image " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "v_{t,i}" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": " and orientation angle " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "a_{t,i}" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": "-th view. The relative angles of " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": " navigable viewpoints to the current viewpoint and the stop action constitute the action space at step " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "A_{t} = \\{act_{t,1},\\dots ,act_{t,n},[STOP]\\}" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": ". The agent selects an action to move to a navigable viewpoint in " + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 357, + 545, + 467 + ], + "type": "text", + "content": " or stop at the current location. Once the agent stops, the navigation episode is completed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 468, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 546, + 564 + ], + "type": "text", + "content": "Overview. In this study, we propose a novel grounded entity-landmark adaptive pre-training scheme for a VLN pre-trained model, whose architecture is shown in Figure 2. The training scheme introduces explicit supervision on entity-landmark grounding to enhance fine-grained cross-modal representation learning. The scheme comprises two distinct stages: GELA pre-training on several proxy objectives and fine-tuning on the VLN downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 574, + 416, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 416, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 416, + 586 + ], + "type": "text", + "content": "4.2. Pre-trained Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "As illustrated in the bottom-left of Figure 2, our pretrained model is a fully transformer-based architecture for multi-modal decision-making, which is modified from the classical cross-modal model LXMERT [53]. The pretrained model takes three inputs: a global instruction " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", history information " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", and current panoramic visual observation " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_{t}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", which are fed in a language encoder, a history encoder, and a vision encoder, respectively. And then the textual and visual modalities exchange the signals through cross-attention layers in the cross-modal en" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12046" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "coder. Specifically, the visual modality is the concatenation of history and visual observation. Finally, the representation of tokens in instruction, history, and visual state is " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "Z = \\{z_{\\mathrm{cls}}, z_1, \\dots, z_L\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "Y_t = \\{y_{\\mathrm{cls}}, y_1, \\dots, y_{t-1}\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "S_t = \\{s_1, \\dots, s_{36}, s_{\\mathrm{stop}}\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 264 + ], + "type": "text", + "content": "To learn effective uni-modal and multi-modal representation, transformer-based models for VLN commonly undergo pre-training on in-domain datasets using several proxy tasks [14, 50, 7]. These tasks include common vision-language pre-training tasks as well as VLN-specific auxiliary tasks. In this study, we utilize five proxy tasks: Masked Language Modeling (MLM) [10], Masked Region Classification (MRC) [37], Instruction Trajectory Matching (ITM) [40], Single-step Action Prediction (SAP) [7], and Spatial Relationship Prediction (SPREL) [7]. Details are demonstrated in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 275, + 162, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 162, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 162, + 289 + ], + "type": "text", + "content": "4.3. GELA Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "text", + "content": "As shown in the top-left of Figure 2, we adopt three adaptive pre-training objectives for learning entity-landmark level alignment, which are Entity Phrase Prediction (EPP), Landmark Bounding box Prediction (LBP), and Entity-Landmark Semantic Alignment (ELSA). The three proxy objectives are designed by mimicking the objectives used in pre-training vision-language Transformers, particularly with the goal of learning phase-level grounding information. LBP is adapted from the training objective of TransVG [9]. EPP and ELSA are extended from the soft token prediction and contrastive alignment of MDETR [25], respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": "Entity Phrase Prediction (EPP). In this objective, we predict the positions of entity phrases that correspond to annotated environment landmarks. We first transfer the human-annotated entity location to a mask vector " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "M_z" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "L + 1" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": " dimensions, which is the same as " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": ". Similarly, we transfer the human-annotated landmark bounding box to a 37-dim (the same as " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": ") mask vector " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "M_s" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": ". Then we train the model to infer a uniform distribution over all token positions that refer to the corresponding landmark patches and supervise this process by the mask vector " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "M_z" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": ". Specifically, we average the patches representation of the landmark: " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "S_t \\times M_s^\\top" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 438, + 288, + 619 + ], + "type": "text", + "content": " denotes the matrix product, obtaining a 768-dim vector. Then we put it into a two-layer feedforward network (FFN) to predict a distribution over the token positions in the instruction sequence:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 628, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 628, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 91, + 628, + 287, + 643 + ], + "type": "interline_equation", + "content": "l o g i t s = \\operatorname {S o f t m a x} \\left(\\operatorname {F F N} \\left(S _ {t} \\times M _ {s} ^ {\\top}\\right)\\right), \\tag {1}", + "image_path": "c084aead1be385e4cdaa0f89852d0ede81720f5a45ecda9337be98ce0221c675.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 653, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 690 + ], + "type": "text", + "content": "where the FFN maps a vector of 768-dim to a vector of " + }, + { + "bbox": [ + 47, + 653, + 287, + 690 + ], + "type": "inline_equation", + "content": "L + 1" + }, + { + "bbox": [ + 47, + 653, + 287, + 690 + ], + "type": "text", + "content": " dimensions. Finally, we minimize the cross-entropy loss between logits and " + }, + { + "bbox": [ + 47, + 653, + 287, + 690 + ], + "type": "inline_equation", + "content": "M_z" + }, + { + "bbox": [ + 47, + 653, + 287, + 690 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 92, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E P P}} = \\text {C r o s s E n t r o p y} (\\operatorname {l o g i t s}, M _ {z}). \\tag {2}", + "image_path": "38337952287b070f615fb4977b0e302178620258b26b76073aeca55aac23a61b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "Landmark Bounding Box Prediction (LBP). In this objective, we predict the bounding box of the landmark that matches with annotated entity phrases. We train the model to directly predict a 4-dim vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "box' = (x', y', w', h')" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " as the coordinates of the bounding box for each entity phrases and supervise this process by the human-annotated bounding box " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "box = (x, y, w, h)" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ". Specifically, we first average the token embeddings of the entity phrase: " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "Z \\times M_z^\\top" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " and then we predict the coordinates of the box through a two-layer FFN and sigmoid function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 354, + 199, + 545, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 199, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 354, + 199, + 545, + 214 + ], + "type": "interline_equation", + "content": "b o x ^ {\\prime} = \\operatorname {S i g m o i d} \\left(\\operatorname {F F N} \\left(Z \\times M _ {z} ^ {\\top}\\right)\\right), \\tag {3}", + "image_path": "d7caa17632c1f6bea114aebf5a34ddf6f3b26065e200fc5fc0139786f97ff39c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 221, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 258 + ], + "type": "text", + "content": "where the FFN maps a vector of 768-dim to a vector of 4-dim. Finally, we apply the smooth L1 loss and generalized IoU loss (GIoU loss) to optimize the box coordinates " + }, + { + "bbox": [ + 304, + 221, + 545, + 258 + ], + "type": "inline_equation", + "content": "M_z" + }, + { + "bbox": [ + 304, + 221, + 545, + 258 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 266, + 545, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 545, + 279 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L B P}} = \\mathcal {L} _ {\\text {s m o o t h} - l _ {1}} \\left(\\text {b o x}, \\text {b o x} ^ {\\prime}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {G I o U}} \\left(\\text {b o x}, \\text {b o x} ^ {\\prime}\\right), \\tag {4}", + "image_path": "6c707280c07deaa2d01af67b1c4c6ce60d210108a236a98b97a96fe38cfc5d2e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 286, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 286, + 545, + 310 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 286, + 545, + 310 + ], + "type": "text", + "content": " is the weight coefficient of GIoU loss to balance these two losses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 311, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 490 + ], + "type": "text", + "content": "Entity-Landmark Semantic Alignment (ELSA). While the above two unidirectional prediction tasks use positional information to match the entity and landmark, the entity-landmark semantic alignment loss enforces alignment between the hidden embeddings of the landmark and entity at the output of the cross-modal encoder. This additional contrastive alignment loss ensures that the representations of the landmark patches and the corresponding entity tokens are closer in the feature space compared to representations of unrelated tokens. This constraint is stronger than the above two unidirectional prediction losses as it directly operates on the representations and is not single-handedly based on the positional information. Specifically, inspired by InfoNCE loss [36], the objective is the mean of two contrastive losses as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 507, + 545, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 545, + 546 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {i = 0} ^ {3 6} \\frac {1}{\\left| Z _ {i} ^ {+} \\right|} \\sum_ {j \\in Z _ {i} ^ {+}} - \\log \\left(\\frac {\\exp \\left(s _ {i} ^ {\\top} z _ {j} / \\tau\\right)}{\\sum_ {k = 0} ^ {L} \\exp \\left(s _ {i} ^ {\\top} z _ {k} / \\tau\\right)}\\right), \\tag {5}", + "image_path": "969176142cb370d77eab4b1dee8608618c1c27458dabc0c2501142c40a4c0090.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 564, + 545, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 564, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 314, + 564, + 545, + 603 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {z} = \\sum_ {i = 0} ^ {L} \\frac {1}{\\left| S _ {i} ^ {+} \\right|} \\sum_ {j \\in s _ {i} ^ {+}} - \\log \\left(\\frac {\\exp \\left(z _ {i} ^ {\\top} s _ {j} / \\tau\\right)}{\\sum_ {k = 0} ^ {3 6} \\exp \\left(z _ {i} ^ {\\top} s _ {k} / \\tau\\right)}\\right), \\tag {6}", + "image_path": "9d19ce31c708ca6801828cb0a16a151b48f509769eddebfc98219da0e51a0338.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 377, + 615, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 615, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 377, + 615, + 545, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {E L S A}} = \\left(\\mathcal {L} _ {s} + \\mathcal {L} _ {z}\\right) / 2, \\tag {7}", + "image_path": "79f97e27200d23a837218ad89af07f05994e9873721d7d42fdc5aa78712a21b1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "Z_{i}^{+}" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": " is the token set to be matched with an annotated patch " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "S_{i}^{+}" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": " is the landmark patch set that should be matched with an annotated token " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 632, + 545, + 681 + ], + "type": "text", + "content": " is a temperature parameter regulating attention to negative samples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 681, + 529, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 529, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 529, + 693 + ], + "type": "text", + "content": "Therefore, the full adaptive pre-training objective is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 350, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {G E L A}} = \\alpha \\mathcal {L} _ {\\text {E P P}} + \\beta \\mathcal {L} _ {\\text {L B P}} + \\gamma \\mathcal {L} _ {\\text {E L S A}}. \\tag {8}", + "image_path": "82dfd0becda18916bc02815dc8ba98a4dc0062e9bc99aec08c3fdf5191ee4cbd.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "12047" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 197, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 197, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 197, + 85 + ], + "type": "text", + "content": "4.4. Fine-tuning for VLN Tasks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 270 + ], + "type": "text", + "content": "The fine-tuning stage is illustrated in the right of Figure 2, where we generalize the GELA model for two VLN downstream tasks under the scheme of imitation learning (IL) and reinforcement learning (RL) by following previous work [20, 7]. IL supervises the agent to clone the behavior of the expert while RL encourages the agent to explore the trajectory according to the learning policy. Firstly, the GELA model navigates in the environment following the ground-truth action and generates gradients by IL. Secondly, using the same instruction, the model samples the action space to make decisions and uses RL to generate gradients. Finally, we combine the gradients and optimize the pre-trained model. For the CVDN task, our setting is the same to [7]. For the R2R task, we apply three settings with different data augmentation following [31]:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 271, + 287, + 378 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 271, + 277, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 271, + 277, + 282 + ], + "spans": [ + { + "bbox": [ + 58, + 271, + 277, + 282 + ], + "type": "text", + "content": "- none: without data augmentation on environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 283, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 283, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 58, + 283, + 287, + 317 + ], + "type": "text", + "content": "- st: with data augmentation by style transformation on original environments. Specifically, each discrete view of panoramas is transferred with a random style." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 319, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 319, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 319, + 287, + 378 + ], + "type": "text", + "content": "- smo: with data augmentation by semantic class masking on synthesis environments, which are generated by image synthesis with the same style as the original environments. One semantic class is randomly masked out during image generation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 390, + 186, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 186, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 186, + 403 + ], + "type": "text", + "content": "5. Experiment and Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 411, + 163, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 163, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 163, + 423 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 429, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 608 + ], + "type": "text", + "content": "Datasets. We evaluate our proposed method on the two VLN datasets: Room-to-Room (R2R) [3] and Vision-and-Dialog Navigation (CVDN) [55], which are all based upon the MP3D indoor environments. The instructions in R2R are detailed descriptions of the navigation trajectory. There are four splits in 21,558 trajectory-instruction pairs of the R2R dataset: training (14,039), validation seen (1,021), validation unseen (2,325), and test unseen (4,173). Our grounded entity-landmark human annotations are based on the training, validation seen, and validation unseen splits. The instructions in CVDN are multi-turn dialogs between the agent and the oracle during navigation. These kinds of instructions are frequently vague and unspecific. Therefore, the CVDN task is more challenging, serving to evaluate the generalization ability for new downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 609, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 693 + ], + "type": "text", + "content": "Evaluation Metrics. For R2R, we report four evaluation metrics: Trajectory Length (TL); Navigation Error (NE ↓) - the average distance in meters between the agent's final position and the goal viewpoint; Success Rate (SR ↑) - the proportion of paths where the agent stopped within 3 meters of the goal viewpoint; Success rate weighted by Path Length (SPL ↑) [2]². SR and SPL are the recommended" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 533, + 177 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 533, + 177 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 533, + 177 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 533, + 177 + ], + "type": "table", + "html": "
ModelFeatureValidation Unseen
TLNE↓SR↑SPL↑
HAMT[7]none11.462.2965.760.9
st11.783.4267.362.6
smo12.133.2267.962.9
GELA (ours)none11.753.3369.263.4
st11.563.2669.364.2
smo11.733.1171.165.0
", + "image_path": "15ce38c40fee0f11472c347373cc2443c108bc90643612d25fd8cb5ce60337fd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 184, + 545, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 184, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 184, + 545, + 206 + ], + "type": "text", + "content": "Table 2. Comparison with the HAMT baseline on the R2R validation unseen split. Black denotes the best results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 226, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 545, + 262 + ], + "type": "text", + "content": "key metrics. For CVDN, we use the average progress in meters of the agent towards the goal viewpoint as the primary evaluation metric, denoted as Goal Progress (GP " + }, + { + "bbox": [ + 304, + 226, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 226, + 545, + 262 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "text", + "content": "Implementation Details. We typically conform to the architecture of [7] and its associated hyper-parameters. We conduct adaptive pre-training on our GEL-R2R training dataset and the augmented dataset from [14]. In the adaptive pre-training stage, we inherited the pre-training tasks to prevent knowledge forgetting. The GEL-R2R training set serves for both pre-training tasks and adaptive pre-training tasks. Due to the lack of entity-landmark grounding annotations, the augmented dataset serves only for pre-training tasks. We set " + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "text", + "content": " in Eq. (4), and " + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "inline_equation", + "content": "\\alpha = 1.0" + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "inline_equation", + "content": "\\beta = 1.0" + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "text", + "content": " in Eq. (8). And we set " + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 262, + 546, + 489 + ], + "type": "text", + "content": " to 0.07 in Eq. (5) and Eq. (6) following [25]. And We adopt adaptive pre-training for 200k iterations with a learning rate of 5e-5 by using 2 NVIDIA RTX 3090 GPUs, and the batch size for each GPU is set to 64. We adopt fine-tuning on the R2R and CVDN tasks respectively. For both R2R and CVDN tasks, we train the GELA model for 100k iterations with a learning rate of 1e-5 on a single NVIDIA RTX 3090 GPU. We set the batch size to 16 for the R2R task and 8 for the CVDN task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 497, + 484, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 484, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 484, + 510 + ], + "type": "text", + "content": "5.2. The Effect of GELA Pre-training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "text", + "content": "In this section, we demonstrate the effects of our GELA pre-training paradigm for VLN models in terms of navigation performance and cross-modal representation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "text", + "content": "The effect of GELA pre-training on navigation performance. Table 2 presents a comparison with the HAMT baseline model on the R2R validation unseen split under three fine-tuning settings. The results of the first three lines are obtained from [31]. The results show that our GELA model outperforms HAMT by a large margin under all three settings. Specifically, GELA achieves absolute improvements of " + }, + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 304, + 552, + 546, + 696 + ], + "type": "text", + "content": " in SR under none, st, and smo respectively. Furthermore, we show the single-step action prediction (SAP) accuracy of GELA and HAMT on the R2R validation unseen split during adaptive pre-training in the left panel of Figure 3. As the training progresses, we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "text", + "content": "2We denote a " + }, + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "text", + "content": " to demonstrate lower is better and an " + }, + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "text", + "content": " to demonstrate" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 703, + 358, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 703, + 358, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 703, + 358, + 712 + ], + "type": "text", + "content": "higher is better." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12048" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 519, + 255 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 255 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 255 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 255 + ], + "type": "table", + "html": "
MethodsValidation SeenValidation UnseenTest Unseen
TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑TLNE↓SR↑SPL↑
Seq2Seq[3]11.336.0139-8.397.8122-8.137.852018
SSM [58]14.703.10716220.704.32624520.404.576146
EnvDrop[54]11.003.99625910.705.22524811.665.235147
AuxRN[66]-3.337067-5.285550-5.155551
SEvol[6]11.973.56676312.263.99625713.404.136257
PREVALENT[14]10.323.67696510.194.71585310.515.305451
AirBERT[12]11.092.68757011.784.01625612.414.136257
RecBERT[20]11.132.90726812.013.93635712.354.096357
HOP[50]11.262.72757012.273.80645712.683.836459
REM[35]10.882.48757212.443.89645813.113.876559
HAMT[7]11.152.51767211.462.29666112.273.936560
GELA (ours)11.192.39767311.733.11716512.993.596762
", + "image_path": "72075c5622ceb43b369be2911c3995e95f53461ab04d7faa47fad2f115acd7ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 546, + 285 + ], + "type": "text", + "content": "Table 3. Comparison with SoTA methods on the R2R dataset. The methods in the top group are trained from scratch. The methods in the second group are based on pre-training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 304, + 165, + 411 + ], + "blocks": [ + { + "bbox": [ + 48, + 304, + 165, + 411 + ], + "lines": [ + { + "bbox": [ + 48, + 304, + 165, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 165, + 411 + ], + "type": "image", + "image_path": "db0760fb6bf398b0ed5754dfe4d4e4884b8fc8e62703cd674b48a7ec9ad8877a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 422, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 456 + ], + "type": "text", + "content": "Figure 3. The effect of GELA pre-training. Single-step Action Prediction (SAP) accuracy (left) and Effective Attention (EA) score (right) of GELA and HAMT on the R2R validation splits." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 304, + 284, + 410 + ], + "blocks": [ + { + "bbox": [ + 173, + 304, + 284, + 410 + ], + "lines": [ + { + "bbox": [ + 173, + 304, + 284, + 410 + ], + "spans": [ + { + "bbox": [ + 173, + 304, + 284, + 410 + ], + "type": "image", + "image_path": "6e6ff82e40c075fd263c0b4d958059aad6f0c70a808d2b5b05ebeba20522e16d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "observe a gradual increase in the performance gap between GELA and HAMT, with GELA and HAMT achieving prediction accuracies of approximately " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": ", respectively. These significant improvements demonstrate that fine-grained cross-modal alignment between entity phrases and environmental landmarks is highly effective in enhancing agents' navigation performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "The effect of GELA pre-training on cross-modal representation. To directly demonstrate the effect of GELA pretraining on cross-modal semantic alignment, we analyze cross-modal attention weights at the last Transformer layer during inference. First, we visualize the attention weights normalized to a scale of 0 to 1 in a heatmap. We then utilize the annotations in GEL-R2R to mark tokens, which are involved in the grounded entity-landmark pairs, and calculate the attention weights received by each token from the corresponding tokens in the other modality, defined as the Effective Attention (EA) score. In the right panel of Figure 3, we compare the average EA score of each marked token in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 304, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 497 + ], + "type": "text", + "content": "GELA and HAMT on the R2R validation seen and unseen splits, where " + }, + { + "bbox": [ + 304, + 304, + 547, + 497 + ], + "type": "inline_equation", + "content": "e2l" + }, + { + "bbox": [ + 304, + 304, + 547, + 497 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 304, + 304, + 547, + 497 + ], + "type": "inline_equation", + "content": "l2e" + }, + { + "bbox": [ + 304, + 304, + 547, + 497 + ], + "type": "text", + "content": ") represents the attention of the entity token to the landmark token (resp. the attention of the landmark token to the entity token). Among all comparisons, GELA achieves significantly higher EA scores than HAMT (with a p-value in the two-sample t-test approaching 0), indicating that the grounded entity-landmark pairs have a stronger correspondence in GELA. While HAMT has a strong ability for uni-modal representation based on the large-scale transformer and end-to-end pre-training, it can be confused about matching the mentioned landmarks with visual observations. The explicit supervision for entity-landmark level alignment in GELA can improve the cross-modal representation and substantially enhance the agent's decision-making ability during navigation. Qualitative examples are provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 503, + 427, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 427, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 427, + 516 + ], + "type": "text", + "content": "5.3. Comparison to SoTA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Room-to-Room: R2R. Table 3 reports the performance comparison of various models on the R2R benchmark. The ensemble models, SE-Mixed [51] and EnvEdit [31], are excluded for a fair comparison. Our proposed GELA model outperforms all the other models in primary metrics (SR and SPL) across all the dataset splits. This indicates that GELA agents can navigate more accurately and efficiently than the other models. Specifically, on the validation seen split, GELA performs comparably to the SoTA HAMT. However, on the validation unseen and test unseen splits, GELA outperforms HAMT by absolute " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " (SR) and " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " (SPL) improvements, respectively. This demonstrates that GELA exhibits better generalization ability to unseen environments, which can be attributed to its stronger capability in capturing the semantic features of landmarks after entity-landmark grounding learning." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12049" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 192 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 192 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 192 + ], + "type": "table", + "html": "
MethodsVal SeenVal UnseenTest Unseen
Seq2Seq[55]5.922.102.35
CMN[68]7.052.972.95
PREVALENT[14]-3.152.44
VISITRON[52]5.113.253.11
ORIST[61]-3.553.15
HOP[50]-4.373.31
MT-RCM+EnvAg[61]5.074.653.91
HAMT[7]6.915.135.58
GELA (ours)8.575.865.87
", + "image_path": "ea5978fffb2f5241a5c64ad90a2942a2ea404d78dce48d482f87bc822db29ced.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 221, + 286, + 315 + ], + "blocks": [ + { + "bbox": [ + 47, + 200, + 282, + 211 + ], + "lines": [ + { + "bbox": [ + 47, + 200, + 282, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 282, + 211 + ], + "type": "text", + "content": "Table 4. Comparison with SoTA methods on the CVDN dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 221, + 286, + 315 + ], + "lines": [ + { + "bbox": [ + 50, + 221, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 50, + 221, + 286, + 315 + ], + "type": "table", + "html": "
MethodsTaskTLNE↓SR↑SPL↑
0None11.462.2965.760.9
1EPP11.833.6365.860.7
2LBP12.713.7165.660.1
3ELSA11.733.5866.361.7
4EPP+LBP12.103.5967.061.4
5EPP+LBP+ELSA11.753.3369.263.4
", + "image_path": "1f2c581fecbc1ff9dbcc7ae332017f6d10c12cca3d730ae62bc3758cd66f48e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 544 + ], + "type": "text", + "content": "Vision-and-Dialog Navigation: CVDN. Table 4 presents the results on the CVDN benchmark, which utilizes Goal Progress (GP) in meters as the key performance metric. During training, we use a mixture of two types of demonstrations (navigator and oracle) as supervision for the trajectory. The results indicate that GELA consistently outperforms the other models on both the validation and test unseen environments. Particularly, our model achieves up to 1.66-meter improvement over the SoTA model HAMT on the validation seen split. Additionally, GELA outperforms the SoTA results by 0.73 meters and 0.29 meters on validation unseen and test splits, respectively. These results clearly demonstrate the GELA model is effective and generalizable to different types of instructions in more challenging VLN downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 145, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 145, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 145, + 574 + ], + "type": "text", + "content": "5.4. Ablation studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Adaptive Pre-training Objective. To evaluate the effectiveness of different adaptive pre-training objectives, we conduct an ablation study on R2R validation unseen split under none setting. The results are presented in Table 5. Model 0 represents the HAMT baseline, while Models 1-3 show the results of combining the original proxy tasks in [7] with EPP, LBP, and ELSA, respectively. The results demonstrate that ELSA can further improve navigation performance, while the other two unidirectional prediction objectives solely based on positional information do not enhance the pre-trained model. This indicates that ELSA pro" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 345, + 71, + 506, + 141 + ], + "blocks": [ + { + "bbox": [ + 47, + 323, + 286, + 345 + ], + "lines": [ + { + "bbox": [ + 47, + 323, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 286, + 345 + ], + "type": "text", + "content": "Table 5. Ablation study of adaptive pre-training objectives on the R2R validation unseen split." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 71, + 506, + 141 + ], + "lines": [ + { + "bbox": [ + 345, + 71, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 345, + 71, + 506, + 141 + ], + "type": "table", + "html": "
CategoryTLNE↓SR↑SPL↑
None11.462.2965.760.9
Scene11.843.4466.761.2
Object11.233.5466.861.9
All11.753.3369.263.4
", + "image_path": "29a558390d9bc5059cc922d2543521b7acf137adbcf6446511cc744227093539.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 148, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 171 + ], + "type": "text", + "content": "Table 6. Ablation study of landmark categories on the R2R validation unseen split." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 202, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 347 + ], + "type": "text", + "content": "vides more effective supervision of cross-modal alignment by directly operating on the feature representations. Subsequently, we evaluate the effect of combining EPP and LBP. The results show that the combined performance (Model 4) is much better than the separate performance of Model 1 and Model 2, demonstrating that these two objectives are complementary. Finally, we combined all three objectives, and the results show that this performance (Model 5) improves even further. These findings highlight the importance of effective cross-modal alignment and the complementary nature of different adaptive pre-training objectives in improving navigation performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 353, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 533 + ], + "type": "text", + "content": "Landmark Category. Table 6 presents our investigation into the impact of two category entities (object and scene) from GEL-R2R under none setting. Our results demonstrate that integrating both object and scene entities can enhance navigation performance. Of the two entity types, object entities have a greater influence on performance, as agents find it easier to comprehend individual objects compared to complex scenes. Furthermore, the best performance is achieved by combining both entity types, enabling agents to navigate more precisely by utilizing both object and scene landmarks. This improvement can be attributed to the fact that the amount of available data for each entity type is only half that of the complete dataset. Therefore, it is plausible that the performance of GELA has not reached saturation due to the limited volume of available data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 557, + 468, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 468, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 468, + 569 + ], + "type": "text", + "content": "5.5. Limitations and Future Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "GELA is a preliminary investigation into cross-modal alignment at the entity-landmark level in VLN based on our human-annotated GEL-R2R dataset. Due to the high cost of human annotation, the volume of GEL-R2R is relatively limited. In the future, we plan to expand the dataset using two cost-effective approaches: 1) developing a data augmentation model by using GEL-R2R to produce grounded entity-landmark annotations, and 2) incorporating large-scale phrase grounding datasets in the VL domain. Additionally, we intend to explore interpretable navigation based on our introduced dataset and method." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12050" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 119, + 83 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": "A well-performed VLN agent should have powerful competency in fine-grained cross-modal semantic alignment between entities and landmarks. In this paper, we introduce the manually annotated grounded entity-landmark dataset GEL-R2R, which provides powerful cross-modal alignment for VLN at the entity-landmark level. We then adopt three grounded entity-landmark adaptive pre-training objectives based on GEL-R2R to facilitate cross-modal semantic alignment learning under explicit supervision. Comprehensive experimental results on two downstream VLN tasks, R2R and CVDN, demonstrate the effectiveness and generalizability of our proposed model, GELA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 248, + 151, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 151, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 151, + 262 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 269, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 288, + 316 + ], + "type": "text", + "content": "We thank ICCV reviewers for their constructive suggestions. This work is supported in part by the National Natural Science Foundation of China under Grant 62076250, Grant 61703407, and Grant 61901505." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 330, + 107, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 107, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 107, + 342 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 350, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 350, + 287, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 287, + 394 + ], + "type": "text", + "content": "[1] Dong An, Yuankai Qi, Yan Huang, Qi Wu, Liang Wang, and Tieniu Tan. Neighbor-view enhanced model for vision and language navigation. In MM '21: ACM Multimedia Conference, pages 5101-5109, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 396, + 288, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 450 + ], + "type": "text", + "content": "[2] Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir R. Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 452, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 452, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 452, + 287, + 507 + ], + "type": "text", + "content": "[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian D. Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674-3683, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 509, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 287, + 563 + ], + "type": "text", + "content": "[4] Angel X. Chang, Angela Dai, Thomas A. Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. In 3DV, pages 667-676, 2017. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 564, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 564, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 564, + 287, + 609 + ], + "type": "text", + "content": "[5] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. TOUCHDOWN: natural language navigation and spatial reasoning in visual street environments. In CVPR, pages 12538-12547, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 610, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 610, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 610, + 287, + 644 + ], + "type": "text", + "content": "[6] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In CVPR, pages 15429-15438, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 646, + 287, + 689 + ], + "type": "text", + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, pages 5834-5847, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 690, + 287, + 714 + ], + "type": "text", + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16516-16526, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 97, + 545, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 97, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 311, + 97, + 545, + 130 + ], + "type": "text", + "content": "[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In ICCV, pages 1749-1759, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In *NAACL-HLT*, pages 4171–4186, 2019. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 177, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 232 + ], + "type": "text", + "content": "[11] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, pages 3318–3329, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 234, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 277 + ], + "type": "text", + "content": "[12] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation. In ICCV, pages 1614-1623, 2021. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 323 + ], + "type": "text", + "content": "[13] Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. Don't stop pretraining: Adapt language models to domains and tasks. In ACL, pages 8342-8360, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 326, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 326, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 326, + 545, + 369 + ], + "type": "text", + "content": "[14] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In CVPR, pages 13134-13143, 2020. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[15] Keji He, Yan Huang, Qi Wu, Jianhua Yang, Dong An, Shuanglin Sima, and Liang Wang. Landmark-rxr: Solving vision-and-language navigation with fine-grained alignment supervision. In NeurIPS, pages 652-663, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 460 + ], + "type": "text", + "content": "[16] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, pages 11773-11781, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 462, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 496 + ], + "type": "text", + "content": "[17] Yicong Hong, Cristian Rodriguez Opazo, Yuankai Qi, Qi Wu, and Stephen Gould. Language and visual entity relationship graph for agent navigation. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "type": "text", + "content": "[18] Yicong Hong, Cristian Rodriguez Opazo, Qi Wu, and Stephen Gould. Sub-instruction aware vision-and-language navigation. In EMNLP, pages 3360-3376, 2020. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 576 + ], + "type": "text", + "content": "[19] Yicong Hong, Zun Wang, Qi Wu, and Stephen Gould. Bridging the gap between learning in discrete and continuous environments for vision-and-language navigation. In CVPR, pages 15418-15428, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 621 + ], + "type": "text", + "content": "[20] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez Opazo, and Stephen Gould. VLN BERT: A recurrent vision-and-language BERT for navigation. In CVPR, pages 1643-1653, 2021. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[21] Haoshuo Huang, Vihan Jain, Harsh Mehta, Alexander Ku, Gabriel Magalhães, Jason Baldridge, and Eugene Ie. Transferable representation learning in vision-and-language navigation. In ICCV, pages 7403-7412, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[22] Gabriel Ilharco, Vihan Jain, Alexander Ku, Eugene Ie, and Jason Baldridge. General evaluation for instruction conditioned navigation using dynamic time warping. In NeurIPS Workshop, 2019. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12051" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[23] Vihan Jain, Gabriel Magalhães, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. In ACL, pages 1862-1872, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[24] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Trans. Assoc. Comput. Linguistics, 8:64-77, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "text", + "content": "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1760-1770, 2021. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[26] Liyiming Ke, Xiujun Li, Yonatan Bisk, Ari Holtzman, Zhe Gan, Jingjing Liu, Jianfeng Gao, Yejin Choi, and Siddhartha S. Srinivasa. Tactical rewind: Self-correction via backtracking in vision-and-language navigation. In CVPR, pages 6741-6749, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[27] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. In ICCV, pages 14718-14728, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[28] Jacob Krantz and Stefan Lee. Sim-2-sim transfer for vision-and-language navigation in continuous environments. In ECCV, volume 13699 of Lecture Notes in Computer Science, pages 588–603, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "text", + "content": "[29] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, volume 12373 of Lecture Notes in Computer Science, pages 104-120, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[30] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In EMNLP, pages 4392-4412, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "type": "text", + "content": "[31] Jialu Li, Hao Tan, and Mohit Bansal. Envedit: Environment editing for vision-and-language navigation. In CVPR, pages 15386-15396, 2022. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "text", + "content": "[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. CoRR, abs/1908.03557, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[33] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, pages 10955–10965, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "text", + "content": "[34] Xinghang Li, Di Guo, Huaping Liu, and Fuchun Sun. REVE-CE: remote embodied visual referring expression in continuous environment. IEEE Robotics Autom. Lett., 7(2):1494–1501, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup. In ICCV, pages 1624-1634, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 115 + ], + "type": "text", + "content": "[37] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, pages 13–23, 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 171 + ], + "type": "text", + "content": "[38] Panzhong Lu, Xin Zhang, Meishan Zhang, and Min Zhang. Extending phrase grounding with pronouns in visual dialogues. In Proc. of the EMNLP, pages 7614-7625, Abu Dhabi, United Arab Emirates, Dec. 2022. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 546, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 215 + ], + "type": "text", + "content": "[39] Chih-Yao Ma, Jiasen Lu, Zuxuan Wu, Ghassan AlRegib, Zsolt Kira, Richard Socher, and Caiming Xiong. Self-monitoring navigation agent via auxiliary progress estimation. In ICLR, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 546, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 270 + ], + "type": "text", + "content": "[40] Arjun Majumdar, Ayush Shrivastava, Stefan Lee, Peter Anderson, Devi Parikh, and Dhruv Batra. Improving vision-and-language navigation with image-text pairs from the web. In ECCV, volume 12351 of Lecture Notes in Computer Science, pages 259-274, 2020. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 546, + 304 + ], + "type": "text", + "content": "[41] Shikib Mehri, Mihail Eric, and Dilek Hakkani-Tür. Dialogue: A natural language understanding benchmark for task-oriented dialogue. CoRR, abs/2009.13570, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 305, + 546, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 546, + 380 + ], + "type": "text", + "content": "[42] Dipendra Kumar Misra, Andrew Bennett, Valts Blukis, Eyvind Niklasson, Max Shatkhin, and Yoav Artzi. Mapping instructions to actions in 3d environments with visual goal prediction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 2667-2678, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 546, + 426 + ], + "type": "text", + "content": "[43] Abhinav Moudgil, Arjun Majumdar, Harsh Agrawal, Stefan Lee, and Dhruv Batra. SOAT: A scene- and object-aware transformer for vision-and-language navigation. In NeurIPS, pages 7357-7367, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 427, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 427, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 546, + 470 + ], + "type": "text", + "content": "[44] Khanh Nguyen, Debadeepta Dey, Chris Brockett, and Bill Dolan. Vision-based navigation with language-based assistance via imitation learning with indirect intervention. In CVPR, pages 12527-12537, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 472, + 546, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 546, + 504 + ], + "type": "text", + "content": "[45] Alexander Pashevich, Cordelia Schmid, and Chen Sun. Episodic transformer for vision-and-language navigation. In ICCV, pages 15922-15932, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 505, + 546, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 546, + 558 + ], + "type": "text", + "content": "[46] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. Int. J. Comput. Vis., 123(1):74-93, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 546, + 613 + ], + "type": "text", + "content": "[47] Yuankai Qi, Zizheng Pan, Yicong Hong, Ming-Hsuan Yang, Anton van den Hengel, and Qi Wu. The road to knowwhere: An object-and-room informed sequential BERT for indoor vision-language navigation. In ICCV, pages 1635-1644, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 546, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 658 + ], + "type": "text", + "content": "[48] Yuankai Qi, Zizheng Pan, Shengping Zhang, Anton van den Hengel, and Qi Wu. Object-and-action aware model for visual language navigation. In ECCV, volume 12355 of Lecture Notes in Computer Science, pages 303-317, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "type": "text", + "content": "[49] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: remote embodied visual referring expression in real indoor environments. In CVPR, pages 9979-9988, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12052" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[50] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: history-and-order aware pre-training for vision-and-language navigation. CoRR, abs/2203.11591, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[51] Wenda Qin, Teruhisa Mitsu, and Derry Wijaya. Explore the potential performance of vision-and-language navigation model: a snapshot ensemble method. CoRR, abs/2111.14267, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "text", + "content": "[52] Ayush Shrivastava, Karthik Gopalakrishnan, Yang Liu, Robinson Piramuthu, Gokhan Tur, Devi Parikh, and Dilek Hakkani-Tur. VISITRON: visual semantics-aligned interactively trained object-navigator. In Findings of ACL, pages 1984-1994, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 252 + ], + "type": "text", + "content": "[53] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In EMNLP-IJCNLP, pages 5099-5110, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 287, + 295 + ], + "type": "text", + "content": "[54] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. In *NAACL-HLT*, pages 2610–2621, 2019. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 287, + 342 + ], + "type": "text", + "content": "[55] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In CoRL, volume 100 of Proceedings of Machine Learning Research, pages 394-406, 2019. 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 344, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 344, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 344, + 287, + 387 + ], + "type": "text", + "content": "[56] Maxim Tkachenko, Mikhail Malyuk, Andrey Holmanyuk, and Nikolai Liubimov. Label Studio: Data labeling software, 2020-2022. Open source software available from https://github.com/heartxlabs label-studio.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 442 + ], + "type": "text", + "content": "[57] Hanqing Wang, Wei Liang, Jianbing Shen, Luc Van Gool, and Wenguan Wang. Counterfactual cycle-consistent learning for instruction following and generation in vision-language navigation. In CVPR, pages 15450-15460, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 445, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 477 + ], + "type": "text", + "content": "[58] Hanqing Wang, Wenguan Wang, Wei Liang, Caiming Xiong, and Jianbing Shen. Structured scene memory for vision-language navigation. In CVPR, pages 8455-8464, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 479, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 479, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 479, + 287, + 521 + ], + "type": "text", + "content": "[59] Hanqing Wang, Wenguan Wang, Tianmin Shu, Wei Liang, and Jianbing Shen. Active visual information gathering for vision-language navigation. In ECCV, volume 12367 of Lecture Notes in Computer Science, pages 307-322, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 578 + ], + "type": "text", + "content": "[60] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, pages 6629-6638, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 580, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 287, + 634 + ], + "type": "text", + "content": "[61] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In ECCV, volume 12369 of Lecture Notes in Computer Science, pages 413–430, 2020. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "type": "text", + "content": "[62] Han Wu, Kun Xu, Linfeng Song, Lifeng Jin, Haisong Zhang, and Linqi Song. Domain-adaptive pretraining methods for dialogue understanding. In ACL/IJCNLP, pages 665-669, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[63] Yubo Zhang, Hao Tan, and Mohit Bansal. Diagnosing the environment bias in vision-and-language navigation. In IJCAI, pages 890-897, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 285 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[64] Jeffrey Zhao, Mahdis Mahdieh, Ye Zhang, Yuan Cao, and Yonghui Wu. Effective sequence-to-sequence dialogue state tracking. In EMNLP, pages 7486-7493, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "type": "text", + "content": "[65] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. SOON: scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689-12699, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 194 + ], + "type": "text", + "content": "[66] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In CVPR, pages 10009-10019, 2020. 1, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "text", + "content": "[67] Wang Zhu, Hexiang Hu, Jiacheng Chen, Zhiwei Deng, Vihan Jain, Eugene Ie, and Fei Sha. Babywalk: Going farther in vision-and-language navigation by taking baby steps. In ACL, pages 2539-2556, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 285 + ], + "type": "text", + "content": "[68] Yi Zhu, Fengda Zhu, Zhaohuan Zhan, Bingqian Lin, Jianbin Jiao, Xiaojun Chang, and Xiaodan Liang. Vision-dialog navigation by exploring cross-modal memory. In CVPR, pages 10727-10736, 2020. 1, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12053" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_content_list.json b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd6a5327a240e0a02efe41a07e94106d10ce5e31 --- /dev/null +++ b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_content_list.json @@ -0,0 +1,1799 @@ +[ + { + "type": "text", + "text": "Grounded Image Text Matching with Mismatched Relation Reasoning", + "text_level": 1, + "bbox": [ + 161, + 162, + 810, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Wu $^{1,*}$ Yana Wei $^{1,*}$ Haozhe Wang $^{1}$ Yongfei Liu $^{1}$ Sibei Yang $^{1,2}$ Xuming He $^{1,2}$", + "bbox": [ + 210, + 207, + 815, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ShanghaiTech University, $^{2}$ Shanghai Engineering Research Center of Intelligent Vision and Imaging", + "bbox": [ + 148, + 224, + 877, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wuyul, weiynl, yangsb, hexm}@shanghaiitech.edu.cn jasper.whz@outlook.com liuyongfei314@gmail.com", + "bbox": [ + 122, + 242, + 903, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 287, + 328, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces Grounded Image Text Matching with Mismatched Relation (GITM-MR), a novel visual-linguistic joint task that evaluates the relation understanding capabilities of transformer-based pre-trained models. GITM-MR requires a model to first determine if an expression describes an image, then localize referred objects or ground the mismatched parts of the text. We provide a benchmark for evaluating vision-language (VL) models on this task, with a focus on the challenging settings of limited training data and out-of-distribution sentence lengths. Our evaluation demonstrates that pre-trained VL models often lack data efficiency and length generalization ability. To address this, we propose the Relation-sensitive Correspondence Reasoning Network (RCRN), which incorporates relation-aware reasoning via bi-directional message propagation guided by language structure. Our RCRN can be interpreted as a modular program and delivers strong performance in terms of both length generalization and data efficiency. The code and data are available on https://github.com/SHTUPLUS/GITM-MR.", + "bbox": [ + 111, + 316, + 473, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 619, + 233, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, transformer-based vision-language (VL) pretrained models have made significant progress on various VL tasks by fine-tuning on downstream tasks [8, 64, 58, 29, 16]. Despite their success, the representation capacity of such VL pre-trained models remains poorly understood. As a result, an increasing number of studies start to probe the limitations of those learned models from several aspects. The first aspect focuses on the lack of fine-grained understanding of multi-modal data, which is indicated by the weak VL correspondence learned on relations among the entities [36]. In particular, the researchers constructed", + "bbox": [ + 111, + 642, + 473, + 794 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08e7c43eec4ca58a4f5c3986ca78d6eacee85e98ec4273a39768efa687a4fac4.jpg", + "image_caption": [ + "Figure 1. Illustration of the GITM-MR task and the proposed model RCRN. The task is to first identify whether the expression describes the image, and then localize the referred object if the pair are matched, or otherwise ground the mismatched language part. Both data efficiency and length generalization are required." + ], + "image_footnote": [], + "bbox": [ + 532, + 290, + 692, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc45aeb4d866410428d817cf035afcbf356cc3425003ff5f707755d3e2cc8e84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 290, + 826, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "subtle relation phrase differences in text to reveal the poor matching capability between visual relations and linguistic relations [46, 21, 38], or required the model to output fine-grained predictions to thoroughly examine the model's understanding of VL correspondence on different uni-modal component representations [52]. Additionally, while large models generally deliver superior performance, they tend to overfit on small datasets, emphasizing the need to enhance the data-efficiency on fine-tuning [15]. Furthermore, when the text is composed of multiple relations, inferring VL correspondence typically involves combining those semantic relations. This combination requires the model to reason the global semantic correspondence, where the length generalization is shown to be a critical weakness in reasoning capability of pre-trained transformers [3, 54]. However, most of those studies only focus on a single aspect of the VL models and hence produce a partial view on their limitations.", + "bbox": [ + 496, + 516, + 859, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better understand the VL pre-trained models, we introduce a novel VL joint task named Grounded Image Text Matching with Mismatched Relation (GITM-MR) which focuses on distinguishing nuance in relations within text and image, and requires a model to show explicit comprehension on relations in both modalities. As shown in Fig. 1, the task first requires the model to identify whether an expression describes a given image, and then grounds the referred", + "bbox": [ + 496, + 755, + 859, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 138, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 259, + 42, + 781, + 56 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 336, + 56, + 702, + 70 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 308, + 70, + 732, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Both authors contributed equally to this work, which was supported by Shanghai Science and Technology Program 21010502700, Shanghai Frontiers Science Center of Human-centered Artificial Intelligence and MoE Key Lab of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University).", + "bbox": [ + 111, + 809, + 472, + 867 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2976", + "bbox": [ + 483, + 891, + 514, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "object if the image-text pair are matched or localizes the mismatched relation in the expression otherwise. The two fine-grained sub-tasks verify the specific understanding on relations both in match and mismatch scenarios. Moreover, the proposed task also allows us to evaluate the data efficiency of model's fine-tuning and its generalization ability. In particular, we design a learning setting with limited annotations, where only the data in a limited size and domain have annotations (i.e., a mix of in- and out-of-distribution (OOD) test data). Given the task design, we then build a new benchmark based on the Ref-Reasoning [61] dataset, which is more general and challenging than previous probing benchmarks on relations and with low linguistic bias.", + "bbox": [ + 109, + 126, + 472, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging our proposed benchmark, we evaluate several representative state-of-the-art models on the task, assessing their performance under limited data sizes and varying domains of input length. Our results reveal their limitations in the both settings. To alleviate these drawbacks, we propose a unified modular framework that jointly solves subtasks in the GITM-MR problem. By explicitly capturing relation correspondence and utilizing it to propagate messages, our approach is capable of obtaining a comprehensive understanding of global correspondence. Additionally, its lightweight and compositional module design allows for excellent data efficiency and length generalization, accompanied by a more transparent reasoning process.", + "bbox": [ + 109, + 308, + 472, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we develop a graph neural network on a parsed language scene graph (LSG), named Relation-sensitive Correspondence Reasoning Network (RCRN), which exploits the language structure to perform relation-aware reasoning in a belief space of visual-linguistic alignment. Our RCRN uses a set of primitive reasoning modules to operate on the pre-trained vision-language features under the guidance of LSG, and predict the fine-level grounding in both match and mismatch scenarios. As demonstrated by the experiments, our method outperforms in both data efficiency and length generalization settings, demonstrating its effectiveness in learning relation correspondences with data efficiency and OOD generalization capability.", + "bbox": [ + 109, + 489, + 472, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contribution of our work is summarized as follows:", + "bbox": [ + 130, + 671, + 467, + 684 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new VL task GITM-MR that challenges relation correspondence learning in VL pre-trained model, and focusing on both data efficiency and out-of-distribution text length setting, with a new benchmark for GITM-MR.", + "- We identify the shortcomings of current VL models through the benchmark, and develop a modular graph network RCRN to address these issues.", + "- Our method achieves superior results on both settings, validating its relation learning efficiency and generalization ability, which hopefully inspires the following exploration of these challenges." + ], + "bbox": [ + 130, + 700, + 470, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 125, + 635, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Probing Tasks for VL Models Despite the rapid development on VL pre-trained models [8, 41, 19, 64, 58, 29], emergent works reveal the limitations and gain some deep understanding on pre-trained representations, including implausible scenes understanding [10], language and vision priors exploration [20], and modality ablation [18]. A considerable number of them have revealed the weakness of VL models in relation understanding. Early work proposed a benchmark FOIL [46, 45] by substituting a word in caption to test non-pretrained VL models on identifying minor differences. Subsequent works [21, 38, 36, 53, 63] further evaluate pre-trained models on detecting subtle differences in relations or predicate-noun dependencies in zero-shot scenarios, and reach a consensus on the poor quality of pre-trained VL representation for relations. FGVE [52] extends this evaluation to the task of fine-grained entailment with model fine-tuning, but lacks evaluation in vision modality. Recently, a new benchmark [42] is proposed to simultaneously assess the concept understanding of representations learned from both modalities. In contrast, our task focuses on generating fine-grained multi-modal output under different fine-tuning settings in order to systematically investigate the relation understanding problem in VL pre-training.", + "bbox": [ + 496, + 147, + 859, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OOD Generalization The general problem of OOD generalization has attracted much attention recently [33, 6, 47, 30, 48]. Among them, several studies have considered the tasks with generalization to processing or generating longer sequences. Some works [3, 54] have studied length generalization in different uni-modal seq-to-seq tasks, and find that fine-tuning pre-trained transformer models can lead to poor out-of-distribution (OOD) performance. To address this issue, various solutions have been proposed, including modifications to attention mechanism [39, 11, 17] and the design of recurrent networks [44, 4], but their application in VL tasks is nontrivial. Other recent works have also revealed the challenge of OOD generalization for the VL tasks, but most of them focuses on the answer bias of Visual Question Answering (VQA) [50, 1, 2]. The attempted solutions include the causal learning [37] and stable learning [51] methods. Perhaps most related to us is the modular network approaches [28, 49, 65], which use program-based design to achieve good generalization for certain OOD scenario such as question domains in VQA. We follow the modular design but focus on the OOD problem induced by the expression complexity in the grounded image text matching task.", + "bbox": [ + 496, + 486, + 859, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language Matching Image-Text Matching (ITM) and Referring Expression Grounding (REG) are common cross-modal matching tasks in vision-language domain. The problem of ITM aims to predict a global similarity between a text and an image. Classical methods for ITM fol", + "bbox": [ + 496, + 797, + 859, + 867 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2977", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/84876a2a1d96c35f5193bef9971e7406e9b2a81684c02549173b9a97e2ad43ad.jpg", + "image_caption": [ + "Figure 2. The length distribution of sentences containing different numbers of entities in the validation set. Different colors represent sentences with different number of entities." + ], + "image_footnote": [], + "bbox": [ + 143, + 125, + 439, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "low the core idea of learning a common space where linguistic and visual features are compatible [27, 31, 7, 14, 9]. Recent approach [41] concentrates on learning the relation-aware global similarity by a regularization strategy on training. Our model derives a global similarity measure by propagating messages on local correspondences, capable of capturing subtle changes in relations.", + "bbox": [ + 111, + 266, + 470, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of REG is to locate an object in an image based on a language expression. Most methods focus on learning a holistic visual and linguistic representation in a joint space for aligning the object and text, based on CNN+RNN architectures [35, 66, 25] or the Transformer [12, 13, 32, 40, 59]. Another line of research aims to exploit the linguistic structure in the text for better context modeling [62, 34, 60, 56]. In particular, SGMN [61] parses the text into language components and performs reasoning under the guidance of the linguistic structure. Our method introduces a flexible contextual representation and a bi-directional propagation to extend correspondence reasoning beyond the task of REG.", + "bbox": [ + 111, + 363, + 473, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Setup and Benchmark", + "text_level": 1, + "bbox": [ + 112, + 542, + 375, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to study the representation capacity of the VL pretrained models for fine-grained downstream tasks. To this end, we introduce the Grounded Image Text Matching with Mismatched Relation (GITM-MR) task, which is a joint matching and grounding task focusing on relations in vision and language domain. In this section, we present the problem setting of GITM-MR, followed by building a new benchmark for the task.", + "bbox": [ + 111, + 566, + 472, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GITM-MR Task Given an image and a referring expression, our goal is to determine if the expression describes an object in the image or not (termed as Image-Text Matching (ITM)), and then to localize the referred object if it is matched (termed as Referring Expression Grounding (REG)), or otherwise ground the mismatched relation phrase (termed as Mismatched Relation Reasoning (MRR)). Notably, the only difference between match and mismatch is a relation phrase in the expression, making our task particularly challenging.", + "bbox": [ + 111, + 685, + 470, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, given an image $I$ and a text description $L$ , we denote their matching state as $y$ , indicating whether the image matches the description $(y = 1)$ or not $(y = 0)$ . For the", + "bbox": [ + 111, + 825, + 472, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "matched case, the referred object has a corresponding location in the image, represented by its bounding box $z \\in \\mathbb{R}^4$ . For the mismatch case, we denote the set of all possible relation phrases (i.e. predicates) for mismatch reasoning as $\\mathcal{R} = \\{r_i\\}_{i=1}^{N_e}$ , and there is a target mismatched relation $r_i^*$ in the relation phrase set $\\mathcal{R}$ . Given the input $I$ and $L$ , our goal is to first predict $y$ for the input pair, and then predict $z$ for the referred object in $L$ when $y = 1$ , or otherwise predict $r_i^*$ from $\\mathcal{R}$ to explain why $y = 0$ .", + "bbox": [ + 496, + 126, + 857, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For model training, we follow the convention of ITM and REG tasks, and assume the annotations for those two subtasks are provided. However, it is usually costly to collect annotations for the subtle MRR task. Consequently, we assume a weakly-supervised learning setting where only the matching state label $y$ is given for mismatched data.", + "bbox": [ + 496, + 251, + 857, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Benchmark To conduct systematical study, we construct a benchmark for the GITM-MR task, including a new dataset and two challenging evaluation settings. We first build a dataset of image-text pairs from the public REC dataset, Ref-Reasoning [61], by substituting the relation phrases to generate mismatched referring expressions for images. From the provided relation phrases in the LSG offered by the dataset, we manually select a subset of 27 commonly-occurred relations, and assign some contextual close but semantically different ones for each relation in the subset as their mismatched candidates. We control the linguistic bias by several measures, including keeping the relation distribution and checking by language-only model, the details of which are stated in the Suppl.", + "bbox": [ + 496, + 357, + 857, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, replacing the relation phrases may introduce falsely mismatched cases. To ensure the quality of our data, we filter out those false mismatch cases in the test set with the help of annotators. After filtering, the dataset contains more than 1M referring expressions in 60K images. It has 1M, 50K and 10K expression-referent pairs for training, validation, testing, respectively. We only select the original image text pairs which have generated mismatched pairs to keep the matching label balanced. Fig. 2 shows the length distribution of sentences in the validation set of the benchmark. We find it is more general and challenging than previous benchmarks on relation understanding, and refer the reader to see more details of the dataset in Suppl.", + "bbox": [ + 496, + 551, + 857, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation Settings Given the dataset, we develop two evaluation settings to test pretrained VL models on their relation understanding. The first setting is designed to investigate data efficiency during the fine-tuning process. We train the models on small-scale datasets and tested on indistribution test sets. The second setting is designed to evaluate the generalization performance of models with regards to different text input lengths. Here the test set contains", + "bbox": [ + 496, + 756, + 857, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2978", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cbab5a3af3c37e1c2667eb5303a249a44386c3fc3579c02d583c00d5f32fb365.jpg", + "image_caption": [ + "Figure 3. Model overview with a mismatched case. Given an image and an expression, we first generate visual and linguistic candidates by a detector and a language parser, and compute their representations. Then we use the Context-sensitive Propagation Network to inter alignments between visual-linguistic candidates, which conducts bi-directional message propagation based on the language graph. The propagation initializes the messages by computing local beliefs, selectively aggregates the context information and updates the belief with a context-sensitive gating function. Predictions for this case are obtained by exploiting the beliefs from the propagation. In the language graph, the ellipses represent entity phrases and the rectangles stand for relation phrases. Further elaboration can be found in the main text." + ], + "image_footnote": [], + "bbox": [ + 178, + 126, + 802, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "longer and more complex samples on relation combinations compared to the training set.", + "bbox": [ + 111, + 422, + 470, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, according to the sentence length distribution in Fig. 2, we construct two training sets, named TrainLen16 and Train-Len11, which includes sentences containing less than 16 and 11 words respectively. In the test phase, the model needs to perform the reasoning task on sentences with lengths less than 11 and 16 (in-distribution), as well as sentences longer than these lengths (out-of-distribution). As two training sets accounts for $5\\%$ and $10\\%$ of the overall training data, those two settings allow us to evaluate the model's data efficiency in fine-tuning and OOD generalization ability simultaneously. Detailed data partition rules are explained in Suppl.", + "bbox": [ + 111, + 450, + 473, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Our Approach", + "text_level": 1, + "bbox": [ + 112, + 627, + 246, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our model design for the GITM-MR task and the learning objectives for training with weakly-annotated data. To tackle the problem, we develop a unified and generalizable framework as outlined in Fig. 1, which consists of two main components: a representation network that computes an initial representation of the linguistic and visual components, and a context-aware correspondence reasoning network for final match predictions. Below we first introduce the two main model components in Sec. 4.1 and 4.2, and then the model learning in Sec. 4.3.", + "bbox": [ + 111, + 650, + 470, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Candidate Generation and Representation", + "text_level": 1, + "bbox": [ + 112, + 796, + 442, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1 Candidate Generation", + "text_level": 1, + "bbox": [ + 112, + 816, + 305, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method first generates a set of visual objects from the image and phrase candidates from the corresponding text.", + "bbox": [ + 111, + 838, + 470, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for the visual object candidates, we adopt a pre-trained object detector [64] to generate $N_{o}$ box proposals $\\mathcal{O} = \\{o_i\\}_{i=1}^{N_o}$ from the image $I$ . The detector also provides an initial visual feature (i.e., pooled ROI feature) for each object proposal. For the language phrase candidates, we utilize an off-the-shelf parser [43] to produce a language scene graph $\\mathcal{G} = (\\mathcal{E}, \\mathcal{R})$ of the text expression $L$ , which consists of a set of entity phrases $\\mathcal{E} = \\{e_j\\}_{i=1}^{N_e}$ as its nodes and relation phrases as its edges $\\mathcal{R} = \\{r_{ij}\\}$ , where $r_{ij}$ connects to subject $e_i$ and object $e_j$ . We denote the number of relation phrases as $|\\mathcal{R}| = N_r$ .", + "bbox": [ + 496, + 422, + 859, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2 Candidate Representation", + "text_level": 1, + "bbox": [ + 500, + 590, + 718, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the visual and linguistic candidates, we compute their representations in two steps, as detailed below.", + "bbox": [ + 498, + 612, + 857, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token-level Representation The first step generates an initial representation for the word tokens and visual objects based on a pre-trained vision-language model. Here we adopt a shallow version of the UNITER model [8], aiming to compute a feature representation for the input tokens in two modalities with a generic vision-language alignment from the pre-trained model.", + "bbox": [ + 496, + 643, + 859, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Candidate Features In the second step, we then compile the token-level representations into a set of features for", + "bbox": [ + 498, + 758, + 857, + 786 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "*We note that such a generic representation module has a limited capacity and is typically insufficient for our downstream tasks via a simple fine-tuning procedure as in [8, 58]. Nevertheless, it provides rich semantic features for the subsequent alignment reasoning network (c.f. Sec. 4.2) after a minimal fine-tuning (described in Sec. 5.1), which complements our explicit reasoning process.", + "bbox": [ + 498, + 799, + 859, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2979", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "our visual and linguistic components used in the subsequent reasoning network. Here we adopt a similar strategy as in SGMN [61] for computing the features.", + "bbox": [ + 111, + 126, + 470, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concretely, for visual objects $o_i$ , we augment its token-level feature $\\mathbf{o}_i$ by its spatial location $\\mathbf{l}_i^o$ (i.e., bounding box parameters). In addition, for every object pair $(o_i,o_j)$ in the image, we represent its relational feature as $\\mathbf{r}_{ij}^{o}$ , which is computed from the features of its two objects. For the linguistic entities in $\\mathcal{E}$ and relation phrases in $\\mathcal{R}$ , we feed the corresponding word tokens into a Bi-LSTM [22] with self-attention [24] to generate two sets of features. In the first set, each entities $e_i$ is encoded by a tuple of three features $(\\mathbf{e}_i,\\mathbf{l}_i^e,\\mathbf{h}_i^e)$ , where $\\mathbf{e}_i$ is an embedding into the space of object feature $\\mathbf{o}_i$ , $\\mathbf{l}_i^e$ is an embedding into the space of object location $\\mathbf{l}_i^o$ , and $\\mathbf{h}_i^e$ is a linguistic feature from the LSTM states. For the second feature set, we generate a representation $\\mathbf{r}_{ij}^{e}$ for each relation phrase $r_{ij}\\in \\mathcal{R}$ . We refer the reader to Suppl. for the details of our feature design.", + "bbox": [ + 111, + 167, + 473, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Context-sensitive Propagation Network", + "text_level": 1, + "bbox": [ + 112, + 385, + 421, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now introduce our second model component, the Context-sensitive Propagation Network (CPN), which aims to infer the alignment between the visual-linguistic candidates and predict the match label of input image-text pair. To this end, we build a modular graph neural network on the language graph $\\mathcal{G}$ , which takes the candidate features as input and performs context-aware reasoning in a belief space of visual-linguistic alignment. Such a structured reasoning process enables us to exploit the regularity in language, i.e. the connections of entities by relations, leading to a relation-aware and compositional alignment model with potentially better generalization.", + "bbox": [ + 111, + 407, + 473, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, as illustrated in the language graph in Fig. 3, the $i$ -th node of the graph network corresponds to an entity phrase $e_i$ and the edge between the node $i$ and $j$ corresponds to a relation phrase $r_{ij}$ . We follow the common assumption that the corresponding $\\mathcal{G}$ has a tree structure [61, 34] and the referred entity is regarded as the root node. To perform reasoning, the graph network uses a bi-directional message propagation procedure to infer a relation-aware correspondence belief (w.r.t. the visual candidates) for all the entity nodes (described in Sec. 4.2.1), followed by a readout head module to predict the fine-grained and global match scores (described in Sec. 4.2.2). Finally, we formulate each key operation in the graph reasoning as an easy-to-interpret primitive module, and convert our reasoning process into an explainable program in Sec. 4.2.3.", + "bbox": [ + 111, + 574, + 470, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Bi-directional Message Propagation", + "text_level": 1, + "bbox": [ + 112, + 801, + 390, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the graph network, we aim to compute a message for each node, representing a belief on the correspondence between the associated entity phrase and all the visual objects", + "bbox": [ + 111, + 825, + 472, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in the image. In order to resolve the ambiguity in the local matching of entities, we introduce a bi-directional message propagation procedure to aggregate all the local beliefs on phrase-visual region correspondence throughout the graph.", + "bbox": [ + 496, + 126, + 857, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concretely, inspired by belief propagation [5], we perform two parallel passes of message propagation on the tree graph $\\mathcal{G}$ along two directions: 1) bottom-up direction (denoted as bp) first computes messages from the leaf nodes and then updates their parents recursively until the root is reached; 2) top-down direction (denoted as td) starts from the root node and updates the children nodes recursively until reaching all the leaves. For each direction, the message propagation aggregates the context information on cross-modal matching and update the belief on phrase alignment at each node in the following three steps:", + "bbox": [ + 496, + 181, + 857, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(1) Message initialization: We initialize a local belief $\\mathbf{b}_i^{\\mathrm{loc}}$ at node $i$ based on a similarity measure between the entity $e_i$ and all the visual objects. Here we define the similarity by a weighted combination of feature similarities in appearance and spatial location space. Specifically, for each pair of entity $e_i$ and object $o_j$ , we compute two similarities as follows,", + "bbox": [ + 496, + 333, + 857, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nb _ {i j} ^ {\\mathrm {a p p}} = \\mathrm {F} _ {\\mathrm {s i m}} (\\mathbf {e} _ {i}, \\mathbf {o} _ {j}), \\quad b _ {i j} ^ {\\mathrm {p o s}} = \\mathrm {F} _ {\\mathrm {s i m}} \\left(\\mathrm {I} _ {i} ^ {e}, \\mathrm {I} _ {j} ^ {o}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 436, + 857, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{F}_{\\mathrm{sim}}$ is an MLP network computing the similarity between features from two modalities. The local belief $\\mathbf{b}_i^{\\mathrm{loc}}$ then integrates two similarities as follows:", + "bbox": [ + 496, + 460, + 857, + 502 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {i} ^ {\\text {l o c}} = \\mathrm {F} _ {\\text {n o r m}} \\left(\\beta_ {i} ^ {\\text {a p p}} \\mathbf {b} _ {i} ^ {\\text {a p p}} + \\beta_ {i} ^ {\\text {p o s}} \\mathbf {b} _ {i} ^ {\\text {p o s}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 510, + 857, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{b}_i^t = [b_{ij}^t]_{j=1}^{N_o}$ , $t \\in \\{\\mathrm{app, pos}\\}$ . Here $\\beta_i^t$ is the weight for combining similarities and is computed as $\\beta_i^t = \\mathrm{sigmoid}(\\mathbf{W}_t^\\top[\\mathbf{h}_i^e, \\mathbf{b}_i^{\\mathrm{app}}, \\mathbf{b}_i^{\\mathrm{pos}}])$ , where $\\mathrm{F}_{\\mathrm{norm}}$ is a normalization function to rescale its input elements to $[0,1]$ .", + "bbox": [ + 496, + 536, + 857, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Message aggregation: According to the direction $\\mathbf{td}$ or $\\mathbf{bp}$ , message aggregation collects beliefs from the upstream neighbors in a context-sensitive manner. Formally, we first prune the local beliefs by selecting top $K$ maximum values, which significantly reduces the computation complexity for large $N_{o}$ and is written as,", + "bbox": [ + 496, + 594, + 857, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {j} ^ {\\text {s e l}} = \\mathrm {F} _ {\\text {t o p k}} \\left(\\mathbf {b} _ {j} ^ {\\text {l o c}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 683, + 857, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to [61], our selection is belief-based and more informative for propagation. Given the sparsified beliefs, we then aggregate the messages from the neighboring nodes by taking into account the alignment of the relation phrases on the edges as follows,", + "bbox": [ + 496, + 708, + 857, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {i} ^ {\\mathrm {a g g}} = \\prod_ {j \\in \\operatorname {C t x} (i)} \\mathbf {A} _ {i j} \\mathbf {b} _ {j} ^ {\\mathrm {s e l}}, [ \\mathbf {A} _ {i j} ] _ {k l} = \\mathrm {F} _ {\\text {s i m}} \\left(\\mathbf {r} _ {i j} ^ {e}, \\mathbf {r} _ {k l} ^ {o}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 785, + 857, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Ctx}(i)$ denotes the upstream neighbor set of node $v_{i}$ according to the propagation direction. We model the messages provided by relations explicitly through alignment on", + "bbox": [ + 496, + 824, + 857, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2980", + "bbox": [ + 483, + 890, + 514, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "edges, which forces the model to capture the nuances of different relations. For the aggregation operator, we take the element-wise product of beliefs, which can be viewed as a soft version of logical AND. Unlike the min pooling used in [49, 23], our design leads to a more robust and smooth learning procedure.", + "bbox": [ + 109, + 126, + 472, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(3) Message update: Finally, we update the belief $\\mathbf{b}_i$ at node $i$ by a context-sensitive integration of the aggregated neighboring belief and the local belief. To achieve this, we design a gated product as follows:", + "bbox": [ + 111, + 208, + 472, + 263 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {i} ^ {\\mathrm {r e l}} = \\operatorname {s i g m o i d} \\left(\\mathbf {W} _ {\\mathrm {r e l}} ^ {\\top} \\left[ \\mathbf {h} _ {i} ^ {e}, \\mathbf {b} _ {i} ^ {\\text {a p p}}, \\mathbf {b} _ {i} ^ {\\text {p o s}} \\right]\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 270, + 472, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {i} = \\mathrm {F} _ {\\text {n o r m}} \\left(\\mathbf {b} _ {i} ^ {\\text {l o c}} \\circ \\left(\\mathbf {b} _ {i} ^ {\\text {a g g}}\\right) ^ {\\beta_ {i} ^ {\\text {r e l}}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 289, + 472, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\circ$ denotes the element-wise product. Instead of using the language features as in [61, 55], our gating function also takes the local beliefs as input, making the network more sensitive to the image context.", + "bbox": [ + 111, + 319, + 472, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the updated beliefs, we compute a confidence score for each node matching to any object in the image by applying an MLP network, denoted as $p_i = \\mathrm{sigmoid}(\\mathrm{MLP}_{\\mathrm{match}}(\\mathbf{b}_i))$ . In the end, the message propagation along two directions generates two belief vectors $(\\mathbf{b}_i^{\\mathrm{bp}}$ and $\\mathbf{b}_i^{\\mathrm{td}})$ and two confidence scores $(p_i^{\\mathrm{bp}}$ and $p_i^{\\mathrm{td}})$ for each node $i$ .", + "bbox": [ + 111, + 375, + 472, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Readout Head", + "text_level": 1, + "bbox": [ + 112, + 479, + 254, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the graph-based reasoning, we now introduce a readout head network to generate the predictions for three subtasks, i.e., ITM, REG and MRR, by exploiting the beliefs and confidence scores from the bi-directional propagation.", + "bbox": [ + 111, + 502, + 472, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-Text Matching To predict the global matching label $y$ , we take a minimum pooling of the matching confidence scores across the entire graph as below,", + "bbox": [ + 111, + 561, + 472, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (y = 1 | I, L) = \\min _ {i} \\left(\\min \\left(p _ {i} ^ {\\mathrm {b p}}, p _ {i} ^ {\\mathrm {t d}}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 609, + 472, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "REG Task For the referent grounding, we take the belief vector of the root node $\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}}$ from the bottom-up direction and locate the matching visual object $o_{i^{*}}$ by choosing the largest belief value: $i^{*} = \\arg \\max_{i}\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}}$ . A standard regression head is also employed for more precise localization and its details are left to the Suppl.", + "bbox": [ + 111, + 638, + 472, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MRR Task For mismatched relation prediction, we exploits the inconsistency of the match confidence scores along the graph edges. We found the mismatched relation typically leads to a confidence gap between its two incident nodes. As such, we choose the edge with the largest match confidence difference between its two nodes:", + "bbox": [ + 111, + 724, + 472, + 808 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {r} = \\arg \\min _ {r _ {i j} \\in \\mathcal {R}} \\left(\\left\\{\\left(p _ {i} ^ {\\mathrm {b p}} - p _ {j} ^ {\\mathrm {b p}}\\right) + \\left(p _ {j} ^ {\\mathrm {t d}} - p _ {i} ^ {\\mathrm {t d}}\\right) \\right\\}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 814, + 472, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we sum the differences from two propagation directions to make the decision more robust.", + "bbox": [ + 111, + 845, + 472, + 872 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4bcebd092f06acf17769977c3da07fa4e8a4eca409dbe55e2686d4fe5ee9a8f9.jpg", + "table_caption": [ + "Table 1. The full list of the reasoning modules and their implementations. We use bold lowercase variables to represent vectors and bold uppercase variables to represent matrices. Regular variables are scalars. W is trainable parameter matrix." + ], + "table_footnote": [], + "table_body": "
ModulesInOutImplementation
Local Correspondences:
Simv'txt, v'visbFsim(v'txt, v'vis)
GateSum{vi}, {bi}bFnorm(∑i sigmoid(W^T[vi, bi])bi)
Correspondence Reasoning:
Selectbb'Ftopk(b)
Aggregate{Ai}, {bi}b'Πi(AiBi)
GateProdb, b', βib''Fnorm(b o (b')βi)
ClassifybpMLPmatch(b)
Readout Head:
Locatepiarg min(p)
Comparepi, pjdpi - pj
And{pi}pminpi
", + "bbox": [ + 504, + 174, + 861, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 Modular Network Interpretation", + "text_level": 1, + "bbox": [ + 500, + 375, + 759, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now introduce an interpretation of our using the modular network framework. As shown in Tab. 1, we summarize the key operations in the RCRN into a set of primitive modules with semantic meanings. This allows us to build a explainable reasoning program: (1) In the message initialization step, local beliefs are computed using Sim, and are combined by GateSum to generate the initial belief on each node. (2) For message aggregation, we first use Select to prune each node and then apply Aggregate to collect the messages from all neighbors, considering the alignment of relations particularly. (3) For message update, GateProd is used to combine the aggregated message with the local belief to generate the updated belief on each node. The Classify module provides a confidence score indicating the likelihood that the phrase entity matches an visual object. (4) In the readout head, we use two Compare modules in two directions respectively for MRR and one And operation for the ITM task, and finally generate predictions for all three subtasks via Locate module. For instance, the grounding prediction is achieved by $\\text{Locate}(-\\mathbf{b}_{\\text{ref}})$ . Several example are provided in Sec. 5.4 and Suppl. to illustrate the interpretable reasoning process, along with further detailed implementations of the modules.", + "bbox": [ + 496, + 400, + 859, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that such a program formulation indicates that the model can enjoy a good out-of-domain generalization due to the compositionality of language graphs. Even facing with more complex text structures, we can perform this reasoning process by reusing those modules on new graphs.", + "bbox": [ + 496, + 719, + 859, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Model Learning", + "text_level": 1, + "bbox": [ + 500, + 801, + 648, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train the network on a multitask loss with end-to-end back propagation. Cross-Entropy (CE) loss and Binary Cross-Entropy (BCE) loss are used for the REG and ITM", + "bbox": [ + 496, + 825, + 859, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2981", + "bbox": [ + 483, + 891, + 511, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tasks, respectively:", + "bbox": [ + 112, + 126, + 230, + 140 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g r d}} = \\mathrm {C E} \\left(\\mathbf {p} _ {\\mathrm {g r d}}, \\mathbf {p} _ {\\mathrm {g r d}} ^ {*}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 148, + 470, + 167 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h}} = \\mathrm {B C E} \\left(p _ {\\text {m a t c h}}, y ^ {*}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 186, + 470, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{\\mathrm{grad}} = \\mathrm{softmax}(\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}})$ stands for the predicted grounding distribution, $\\mathbf{p}_{\\mathrm{grad}}^{*}$ denotes the ground-truth one-hot grounding label, and $p_{\\mathrm{match}} = P(y = 1|I,L)$ denotes the predicted match probability. When the object boxes are from a detector, the box with the maximum IoU with the ground-truth box of the referent is labeled as the localization output. The overall loss is defined as:", + "bbox": [ + 111, + 209, + 472, + 306 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {g r d}} + \\mu \\mathcal {L} _ {\\mathrm {m a t c h}} + \\omega \\mathcal {L} _ {\\mathrm {r e g}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 316, + 470, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\omega$ are hyper-parameters to balance the three terms. $\\omega$ would be zero if the IoU of the ground-truth box and the selected proposal is less than 0.5. The regression loss is standard and can be found in Suppl.", + "bbox": [ + 111, + 342, + 472, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 112, + 409, + 233, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conducted experiments on the proposed benchmark to test several baselines of VL pre-trained models on inferring relation correspondences, with a specific focus on their data efficiency and length generalization capabilities. We then compare our model with those strong baselines to show its effectiveness in these settings. Finally we illustrate the visualizations and interpretations of the reasoning process of our model.", + "bbox": [ + 111, + 431, + 473, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 112, + 550, + 274, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics The evaluation metrics include classification accuracy for three subtasks. The grounding result for a matched case is considered as correct when it is identified as matching and the predicted box has at least 0.5 IoU with its ground-truth location. The grounding accuracy (i.e. Recall@1) is the ratio of correctly grounded cases. For mismatch reasoning, a mismatched case needs to be correctly classified and the mismatched relation should be accurately selected from the candidate set. The MRR accuracy is the top-1 accuracy among the candidates.", + "bbox": [ + 111, + 571, + 470, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VL Pre-trained Models We benchmark five VL models on the proposed dataset. We first choose TCL [58], UNITER [8] and FIBER [16], where TCL is the state-of-the-art pre-trained model for image-text retrieval, UNITER is a vanilla transformer-based VL pre-trained model for matching and grounding tasks, and FIBER uses a specific fine-grained pre-training design. Additionally, we adapt the latest fine-grained visual entailment (FGVE) model [52] to our setting. We compare its potential fine-grained multitask capability with ours, which is designed based on the VL pre-trained model Oscar+[64]. Because none of the", + "bbox": [ + 111, + 714, + 472, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pretrained VL models can fulfill three subtasks jointly, we also apply the multiple instance learning strategy [57] on UNITER to cope with the weakly-supervised MRR task. As shown in Suppl, we modify those models by adding subtask heads to perform three subtasks. We note that TCL and UNITER use the same pre-training datasets and the pretraining corpus of FIBER and FGVE is larger while all the models are fine-tuned on the GITM-MR, which leads to a relatively fair comparison.", + "bbox": [ + 496, + 126, + 859, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details For generating object candidates, we take the off-the-shelf detector, VinVL [64], to extract at most 100 proposals with object score great than 0.1 for each image. For model training, the reasoning component of RCRN is trained by the Adam optimizer with the learning rate set to $5\\mathrm{e} - 4$ for the proposed dataset. The learning rate of candidate representation module stated in Sec. 4.1 is set to $5\\mathrm{e} - 7$ . We train 80K iterations with batchsize 64 totally. The hyper-parameter $K$ , $\\mu$ and $\\omega$ are set as 5, 3 and 1 based on validation, respectively. The representation module uses the first 6 transformer encoder layers in the pre-trained UNITER. The ablation study on layers and more implementation details are included in the Suppl.", + "bbox": [ + 496, + 255, + 859, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results and Comparisons", + "text_level": 1, + "bbox": [ + 500, + 443, + 712, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relation Understanding with Limited Data Tab. 2 (In-Distribution part) shows the performances of pre-trained models on the relation understanding under the condition of limited data. The results indicate that TCL struggles to achieve a match accuracy of more than chance. This could be attributed to its emphasis on global image-text matching during pre-training, which may hinder its ability to perform fine-grained understanding tasks. Though FGVE is designed for fine-grained entailment prediction, it still lacks capability on mismatch relation reasoning in our scenario. This is likely due to its reliance on the pre-trained VL model for the alignment between vision and language relations. UNITER and FIBER perform relatively better on the match task than TCL because they both use the World-Region Alignment (WRA) as a pre-training task, and FIBER is also pre-trained with phrase grounding, which helps to establish fine-grained correspondence between image and text. The overall conclusion drawn from these findings is that pre-trained models exhibit unsatisfying performance on the challenging benchmark under the limited data condition, indicating the need for further development in establishing effective visual-linguistic correspondence for relation understanding.", + "bbox": [ + 496, + 465, + 859, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In comparison, our RCRN outperforms the SOTA pretrained models on match and grounding tasks in the indistribution scenarios, even with fewer parameters. The model's performance with limited training data indicates the effectiveness of its lightweight and structured design. The design incorporates appropriate inductive biases based", + "bbox": [ + 496, + 784, + 859, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2982", + "bbox": [ + 483, + 890, + 514, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ebff7d1d41b2283aee690ca70896344327407618422caeec8e333435c3f6c20e.jpg", + "table_caption": [ + "Table 2. Experiment results of testing models' relation understanding with limited data and ability of length generalization. UNITER+MIL represents the UNITER with multiple instance learning strategy applied. FGVE+MAX represents the fine-grained visual entailment model with max pooling on knowledge elements to predict mismatched relation. - means the model can't handle the corresponding task." + ], + "table_footnote": [], + "table_body": "
Training SetMethod#Pretrain Images#ParamFull TestIn-DistributionOut-of-Distribution
Match%Grounding%MRR%Match%Grounding%MRR%Match%Grounding%MRR%
Train-Len16TCL [58]4M210M52.77--53.11--52.70--
UNITER [8]4M111M61.9324.16-71.2940.63-59.9420.67-
UNITER+MIL [57]4M111M59.1318.4647.568.8341.5361.3857.0713.5644.53
FIBER [16]4.8M254M52.319.01-59.0437.15-50.883.03-
FGVE+MAX [52]5.7M113M54.30-33.2358.37-51.7953.43-29.26
RCRN (Ours)4M90M66.5929.1055.7271.8546.2459.0465.4725.4655.01
Train-Len11TCL [58]4M210M51.71--52.08--51.70--
UNITER [8]4M111M55.5110.24-68.2230.05-54.979.42-
UNITER+MIL [57]4M111M55.4311.4442.7365.0429.5661.6555.0210.6941.93
FIBER [16]4.8M254M53.5211.74-57.2129.56-53.3611.00-
FGVE+MAX [52]5.7M113M53.40-38.0555.26-40.7853.31-37.93
RCRN (Ours)4M90M63.4121.4552.6768.7037.9357.7763.1920.7752.46
", + "bbox": [ + 116, + 160, + 857, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on sentence structure to establish both relation and entity correspondence in a visual-linguistic context, and learns the reasoning strategy effectively.", + "bbox": [ + 111, + 359, + 472, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Length Generalization Results about length generalization ability are shown in the Out-of-Distribution part of Tab. 2. Overall, the OOD performance of pre-trained models on both test settings was generally poor compared to their in-distribution performance, especially on the grounding and MRR subtasks, which explicitly reflect the models' understanding of relations. In particular, in the OOD setting, FIBER's match accuracy is similar to TCL, with both models only slightly exceeding $50\\%$ . The OOD results suggest that these models tend to overfit to the training data biases, without truly establishing the universal VL correspondence of the relations, thus preventing them from generalizing to longer sentences.", + "bbox": [ + 111, + 407, + 472, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, RCRN achieved notably better performance than the other models on all three subtasks in both OOD scenarios. Specifically, RCRN trained with Train-Len11 outperformed the best performance of other models by $8.17\\%$ , $10.08\\%$ , and $10.53\\%$ on all three tasks. This success can be attributed to the modular design, which helps the model learn to establish local visual-linguistic correspondence and the universal reasoning strategy. During test, the model can exploit the regularity in sentence structures to compose the modules and deal with longer sentences.", + "bbox": [ + 111, + 589, + 472, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion These experiments not only help us investigate the two challenges that the pre-trained models faced, but also lead to other interesting findings. First, our observation that fine-grained pre-training tasks, such as WRA, can help models learn local correspondences on downstream tasks, is consistent with the findings of [36]. Second, our findings in the OOD scenario are consistent with previous works [3, 54], which suggest that pre-trained models are prone to overfitting to the training data length biases.", + "bbox": [ + 111, + 742, + 472, + 867 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e158282e63d81c6d6c5cef53500e5f6fee2e30f84231fde1c4696d72d3f80b67.jpg", + "table_caption": [ + "Table 3. Ablation study on the benchmark's OOD validation set." + ], + "table_footnote": [], + "table_body": "
LangβCtxβBi-MPVLPMatGrdMRR
---√(w/o MP)54.7913.9941.47
--✓✓59.6523.8045.58
✓-✓✓60.8824.1849.57
✓✓-✓60.0723.8848.62
✓✓✓-60.7029.2536.65
✓✓✓✓61.4723.3753.23
", + "bbox": [ + 513, + 369, + 846, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, to independently investigate the capability on two subtasks, namely grounding and MRR, we also evaluate on an oracle setting, where we suppose that the models always predict correctly on the matching task. The experiment demonstrates our RCRN has strong performance on both subtasks, and the details of this experiment are provided in the Suppl.", + "bbox": [ + 496, + 483, + 857, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, our experiments provide insights into the challenges and potential solutions for the relation understanding in VL models. The fine-grained pre-training tasks can help models learn local correspondences, while the modular design can facilitate the composition of modules to deal with longer sentences. We believe that our findings offer a promising direction to effectively address these challenges.", + "bbox": [ + 496, + 580, + 857, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 690, + 635, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present the effectiveness of each component in the proposed RCRN through ablative experiments trained on the Train-Len16 set, with the results in Tab. 3.", + "bbox": [ + 496, + 713, + 857, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation study by removing some specific designs. $\\mathbf{Lang}\\beta$ stands for using the language features as one input for gate functions in Eq. 2 and Eq. 6. Removing it indicates constant $\\beta$ 's in the propagation. $\\mathbf{Ctx}\\beta$ means feeding initial similarities into the gate functions. Bi-MP represents propagating twice individually using different $\\beta^{\\mathrm{rel}}$ 's for matching task reasoning. Without this part, only the propagation from the root to leaves is conducted. VLP", + "bbox": [ + 496, + 755, + 859, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2983", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8cb73cd665b761423235616b8428a77eaf47f7e582b22b34e696a407186fb651.jpg", + "image_caption": [ + "(a) There is a guy wearing the black shirt and to the right of the boy that holding a cell phone." + ], + "image_footnote": [], + "bbox": [ + 125, + 127, + 478, + 289 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8e689158bbf2041f9a120a5acaea4aa9336b1d2692096169a6ea00abcab481a1.jpg", + "image_caption": [ + "(b) There is a guy wearing the black shirt and to the left of the boy that holding a cell phone.", + "Figure 4. The visualization of the reasoning process of RCRN. The entity and relation correspondence maps are shown by color maps on boxes and box pairs respectively, where only those with top 5 correspondence scores for each phrase are included. Small color maps in (a) are the local correspondence maps and the large ones are the propagated contextualized maps. The boxes of parents in box pairs are dotted on their centers. Dashed boxes indicate the pooling results." + ], + "image_footnote": [], + "bbox": [ + 127, + 304, + 475, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "stands for introducing vision-language pre-trained transformer layers to encode features. Particularly, the first line with VLP only stands for the model which consists of 6 transformer layers from UNITER and the UNITER task heads stated in Sec. 5.1 and Sec. 5.2.", + "bbox": [ + 111, + 553, + 470, + 621 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results show that all of these designs improve the matching performance. VLP model (first line in Tab. 3) with the same setting as the VLP layers in RCRN performs poorly on OOD scenario, which demonstrates our footnote statement in Sec. 4.1. Nevertheless, VLP component of RCRN is significant on MRR performance, probably because it provides accurate preliminary vision-language alignments on relations.", + "bbox": [ + 111, + 623, + 470, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.4. Visualization and Interpretability", + "text_level": 1, + "bbox": [ + 112, + 746, + 380, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 4, we visualize a pair of matched and mismatched samples of our RCRN. Fig. 4(a) shows the bottom-up propagation process followed by grounding prediction on a matched image-text pair, and Fig. 4(b) presents the bi-directional propagation results with readout process for MRR on a mismatched pair. In Fig. 4(a), the propagation eliminates the ambiguities on entities under the guid", + "bbox": [ + 111, + 768, + 470, + 866 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f77a01f54304cddd60e4a844308c0635f72532728ce759fb7c89dd1308097d88.jpg", + "table_caption": [ + "Table 4. Intermediate diagnosis results of RCRN. E.G.R. means Entity Grounding Recall. R.G.R. means Referent Grounding Recall. R.C.S. means Relation Correspondence Scores." + ], + "table_footnote": [], + "table_body": "
Inference MethodE.G.R.R.G.R.R.C.S.
R@1R@3R@5R@1R@3R@5GTMeanMax
By chance6.3718.2528.785.7416.9427.22---
w/o MP50.5080.5989.7938.8875.8488.110.610.170.69
w/ MP60.8285.5892.1669.4690.9295.590.610.170.69
", + "bbox": [ + 503, + 160, + 857, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ance of explicit relation correspondences. Fig. 4(b) illustrates that the intermediate reasons of the predictions can be traced. The sum of match confidence differences along the mismatched relation, denoted as $d$ in the dotted box, are notably lower than other $d$ 's.", + "bbox": [ + 496, + 250, + 857, + 318 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 4 demonstrates some interpretable intermediate diagnosis results of RCRN on a subset of the validation set, with ground-truth box proposals. We analyze from the aspect of phrase-level visual grounding recall and statistics on relation correspondence scores. To obtain the true correspondence between the language scene graph and the bounding boxes, we search the parsed language scene graph in the original image scene graph in GQA dataset [26] with a noisy graph matching algorithm. See Suppl. for the detail of the algorithm.", + "bbox": [ + 496, + 319, + 857, + 457 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results verify that reasonable local correspondences are built, since the mean of initial recall (i.e. inference without MP) on entity phrase grounding is much higher than the random guess. As for relation correspondences, the average values on the ground-truth vision-language relation correspondences are far above the average mean value on all the pairs and very close to the average maximum. Moreover, message propagation notably helps refine the correspondence maps, especially on the referent entities.", + "bbox": [ + 496, + 457, + 857, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 607, + 607, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have introduced a novel VL joint task, Grounded Image Text Matching with Mismatched Relation (GITM-MR), which provides a challenging benchmark for evaluating pre-trained models on the relation reasoning. Our experiments have shown that some state-of-the-art pretrained models struggle with the task under the setting of limited data and out-of-distribution sentence lengths. To address this problem, we develop a Relation-sensitive Correspondence Reasoning Network (RCRN) to compute contextualized cross-modal alignment of both entities and relations, and ground the fine-grained result in the image or expression. The proposed method outperforms prior SOTA pre-trained models, which demonstrates its strong generalization and data efficiency. Our work sheds light on the limitations of current pre-trained models and the importance of establishing fine-grained vision-language correspondence for relation understanding.", + "bbox": [ + 496, + 630, + 857, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2984", + "bbox": [ + 483, + 890, + 513, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 125, + 200, + 138 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2", + "[2] Arjun R Akula. Question generation for evaluating cross-dataset shifts in multi-modal grounding. arXiv preprint arXiv:2201.09639, 2022. 2", + "[3] Cem Anil, Yuhuai Wu, Anders Andreassen, Aitor Lewkowycz, Vedant Misra, Vinay Ramasesh, Ambrose Slone, Guy Gur-Ari, Ethan Dyer, and Behnam Neyshabur. Exploring length generalization in large language models. arXiv preprint arXiv:2207.04901, 2022. 1, 2, 8", + "[4] Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. End-to-end algorithm synthesis with recurrent networks: Logical extrapolation without overthinking. arXiv preprint arXiv:2202.05826, 2022. 2", + "[5] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006. 5", + "[6] Peter Buhlmann. Invariance, causality and robustness. Statistical Science, 35(3):404-426, 2020. 2", + "[7] Tianlang Chen and Jiebo Luo. Expressing objects just like words: Recurrent visual embedding for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10583-10590, 2020. 3", + "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2, 4, 7, 8", + "[9] Mengjun Cheng, Yipeng Sun, Longchao Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, et al. Vista: vision and scene text aggregation for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5184-5193, 2022. 3", + "[10] Myung Jin Choi, Antonio Torralba, and Alan S Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012. 2", + "[11] JU Da, Stephen Roller, Sainbayar Sukhbaatar, and Jason E Weston. Staircase attention for recurrent processing of sequences. In Advances in Neural Information Processing Systems. 2", + "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 3", + "[13] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. arXiv preprint arXiv:2206.06619, 2022. 3" + ], + "bbox": [ + 112, + 147, + 473, + 864 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Haiwen Diao, Ying Zhang, Lin Ma, and Huchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 3", + "[15] Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305, 2020. 1", + "[16] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 1, 7, 8", + "[17] Yann Dubois, Gautier Dagan, Dieuwke Hupkes, and Elia Bruni. Location attention for extrapolation to longer sequences. arXiv preprint arXiv:1911.03872, 2019. 2", + "[18] Stella Frank, Emanuele Bugliarello, and Desmond Elliott. Vision-and-language or vision-for-language. On CrossModal Influence in Multimodal Transformers.(2021). DOI: https://doi.org/10.18653/v1/2021.emnlp-main, 775, 2021. 2", + "[19] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 2", + "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2", + "[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 3635-3644, 2021. 1, 2", + "[22] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5", + "[23] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 804-813, 2017. 6", + "[24] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1115-1124, 2017. 5", + "[25] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16888-16897, 2021. 3", + "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 9" + ], + "bbox": [ + 501, + 127, + 857, + 864 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2985", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Zhong Ji, Haoran Wang, Jungong Han, and Yanwei Pang. Saliency-guided attention network for image-sentence matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5754-5763, 2019. 3", + "[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. In Proceedings of the IEEE international conference on computer vision, pages 2989-2998, 2017. 2", + "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 2", + "[30] Kun Kuang, Ruoxuan Xiong, Peng Cui, Susan Athey, and Bo Li. Stable prediction with model misspecification and ag- nostic distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4485-4492, 2020. 2", + "[31] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF International conference on computer vision, pages 4654-4662, 2019. 3", + "[32] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 3", + "[33] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 2", + "[34] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11645-11652, 2020. 3, 5", + "[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 3", + "[36] Mitja Nikolaus, Emmanuelle Salin, Stephane Ayache, Abdellah Fourtassi, and Benoit Favre. Do vision-and-language transformers learn grounded predicate-noun dependencies? arXiv preprint arXiv:2210.12079, 2022. 1, 2, 8", + "[37] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2", + "[38] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, 2022. 1, 2" + ], + "bbox": [ + 114, + 126, + 472, + 864 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2", + "[40] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In European Conference on Computer Vision, pages 546-562. Springer, 2022. 3", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3", + "[42] Emmanuelle Salin, Badreddine Farah, Stéphane Ayache, and Benoit Favre. Are vision-language transformers learning multimodal representations? a probing perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11248-11257, 2022. 2", + "[43] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4", + "[44] Avi Schwarzschild, Eitan Borgnia, Arjun Gupta, Furong Huang, Uzi Vishkin, Micah Goldblum, and Tom Goldstein. Can you learn an algorithm? generalizing from easy to hard problems with recurrent networks. Advances in Neural Information Processing Systems, 34:6695-6706, 2021. 2", + "[45] Ravi Shekhar, Sandro Pezzelle, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Vision and language integration: Moving beyond objects. In IWCS 2017—12th International Conference on Computational Semantics—Short papers, 2017. 2", + "[46] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, 2017. 1, 2", + "[47] Xinwei Shen, Furui Liu, Hanze Dong, Qing Lian, Zhitang Chen, and Tong Zhang. Disentangled generative causal representation learning. arXiv preprint arXiv:2010.02637, 2020.2", + "[48] Zheyan Shen, Peng Cui, Tong Zhang, and Kun Kunag. Stable learning via sample reweighting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5692-5699, 2020. 2", + "[49] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8384, 2019. 2, 6", + "[50] Damien Teney, Ehsan Abbasnejad, Kushal Kafle, Robik Shrestha, Christopher Kanan, and Anton Van Den Hengel. On the value of out-of-distribution testing: An example of goodhart's law. Advances in Neural Information Processing Systems, 33:407-417, 2020. 2" + ], + "bbox": [ + 501, + 126, + 856, + 866 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2986", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1417-1427, 2021. 2", + "[52] Christopher Thomas, Yipeng Zhang, and Shih-Fu Chang. Fine-grained visual entailment. arXiv preprint arXiv:2203.15704, 2022. 1, 2, 7, 8", + "[53] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2", + "[54] Dušan Variš and Ondřej Bojar. Sequence length is a domain: Length-based overfitting in transformer models. arXiv preprint arXiv:2109.07276, 2021. 1, 2, 8", + "[55] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 6", + "[56] Jia Wang, Jingcheng Ke, Hong-Han Shuai, Yung-Hui Li, and Wen-Huang Cheng. Referring expression comprehension via enhanced cross-modal graph attention networks. ACM Journal of the ACM (JACM), 2022. 3", + "[57] Yun Wang, Juncheng Li, and Florian Metze. A comparison of five multiple instance learning pooling functions for sound event detection with weak labeling. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 31-35. IEEE, 2019. 7, 8", + "[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquen Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 1, 2, 4, 7, 8", + "[59] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 3", + "[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3", + "[61] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2, 3, 5, 6", + "[62] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L. Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3", + "[63] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 2" + ], + "bbox": [ + 114, + 126, + 472, + 866 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 1, 2, 4, 7", + "[65] Zelin Zhao, Karan Samel, Binghong Chen, et al. Proto: Program-guided transformer for program-guided tasks. Advances in Neural Information Processing Systems, 34:17021-17036, 2021. 2", + "[66] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4252-4261, 2018. 3" + ], + "bbox": [ + 501, + 126, + 857, + 319 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2987", + "bbox": [ + 483, + 891, + 513, + 901 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_model.json b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ed2f94e322d5c34e21a9c6939e9f05f0a1f03ce --- /dev/null +++ b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_model.json @@ -0,0 +1,2534 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.261, + 0.044, + 0.782, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.338, + 0.058, + 0.704, + 0.071 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.31, + 0.071, + 0.733, + 0.085 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.163, + 0.812, + 0.183 + ], + "angle": 0, + "content": "Grounded Image Text Matching with Mismatched Relation Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.208, + 0.816, + 0.224 + ], + "angle": 0, + "content": "Yu Wu\\(^{1,*}\\) Yana Wei\\(^{1,*}\\) Haozhe Wang\\(^{1}\\) Yongfei Liu\\(^{1}\\) Sibei Yang\\(^{1,2}\\) Xuming He\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.225, + 0.879, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)ShanghaiTech University, \\(^{2}\\)Shanghai Engineering Research Center of Intelligent Vision and Imaging" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.243, + 0.905, + 0.258 + ], + "angle": 0, + "content": "{wuyul, weiynl, yangsb, hexm}@shanghaiitech.edu.cn jasper.whz@outlook.com liuyongfei314@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.289, + 0.33, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.318, + 0.474, + 0.594 + ], + "angle": 0, + "content": "This paper introduces Grounded Image Text Matching with Mismatched Relation (GITM-MR), a novel visual-linguistic joint task that evaluates the relation understanding capabilities of transformer-based pre-trained models. GITM-MR requires a model to first determine if an expression describes an image, then localize referred objects or ground the mismatched parts of the text. We provide a benchmark for evaluating vision-language (VL) models on this task, with a focus on the challenging settings of limited training data and out-of-distribution sentence lengths. Our evaluation demonstrates that pre-trained VL models often lack data efficiency and length generalization ability. To address this, we propose the Relation-sensitive Correspondence Reasoning Network (RCRN), which incorporates relation-aware reasoning via bi-directional message propagation guided by language structure. Our RCRN can be interpreted as a modular program and delivers strong performance in terms of both length generalization and data efficiency. The code and data are available on https://github.com/SHTUPLUS/GITM-MR." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.62, + 0.234, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.643, + 0.474, + 0.795 + ], + "angle": 0, + "content": "Recently, transformer-based vision-language (VL) pretrained models have made significant progress on various VL tasks by fine-tuning on downstream tasks [8, 64, 58, 29, 16]. Despite their success, the representation capacity of such VL pre-trained models remains poorly understood. As a result, an increasing number of studies start to probe the limitations of those learned models from several aspects. The first aspect focuses on the lack of fine-grained understanding of multi-modal data, which is indicated by the weak VL correspondence learned on relations among the entities [36]. In particular, the researchers constructed" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.291, + 0.693, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.291, + 0.828, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.428, + 0.86, + 0.493 + ], + "angle": 0, + "content": "Figure 1. Illustration of the GITM-MR task and the proposed model RCRN. The task is to first identify whether the expression describes the image, and then localize the referred object if the pair are matched, or otherwise ground the mismatched language part. Both data efficiency and length generalization are required." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.861, + 0.752 + ], + "angle": 0, + "content": "subtle relation phrase differences in text to reveal the poor matching capability between visual relations and linguistic relations [46, 21, 38], or required the model to output fine-grained predictions to thoroughly examine the model's understanding of VL correspondence on different uni-modal component representations [52]. Additionally, while large models generally deliver superior performance, they tend to overfit on small datasets, emphasizing the need to enhance the data-efficiency on fine-tuning [15]. Furthermore, when the text is composed of multiple relations, inferring VL correspondence typically involves combining those semantic relations. This combination requires the model to reason the global semantic correspondence, where the length generalization is shown to be a critical weakness in reasoning capability of pre-trained transformers [3, 54]. However, most of those studies only focus on a single aspect of the VL models and hence produce a partial view on their limitations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.861, + 0.868 + ], + "angle": 0, + "content": "To better understand the VL pre-trained models, we introduce a novel VL joint task named Grounded Image Text Matching with Mismatched Relation (GITM-MR) which focuses on distinguishing nuance in relations within text and image, and requires a model to show explicit comprehension on relations in both modalities. As shown in Fig. 1, the task first requires the model to identify whether an expression describes a given image, and then grounds the referred" + }, + { + "type": "page_footnote", + "bbox": [ + 0.112, + 0.81, + 0.473, + 0.868 + ], + "angle": 0, + "content": "*Both authors contributed equally to this work, which was supported by Shanghai Science and Technology Program 21010502700, Shanghai Frontiers Science Center of Human-centered Artificial Intelligence and MoE Key Lab of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.515, + 0.903 + ], + "angle": 0, + "content": "2976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.127, + 0.473, + 0.306 + ], + "angle": 0, + "content": "object if the image-text pair are matched or localizes the mismatched relation in the expression otherwise. The two fine-grained sub-tasks verify the specific understanding on relations both in match and mismatch scenarios. Moreover, the proposed task also allows us to evaluate the data efficiency of model's fine-tuning and its generalization ability. In particular, we design a learning setting with limited annotations, where only the data in a limited size and domain have annotations (i.e., a mix of in- and out-of-distribution (OOD) test data). Given the task design, we then build a new benchmark based on the Ref-Reasoning [61] dataset, which is more general and challenging than previous probing benchmarks on relations and with low linguistic bias." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.309, + 0.473, + 0.488 + ], + "angle": 0, + "content": "Leveraging our proposed benchmark, we evaluate several representative state-of-the-art models on the task, assessing their performance under limited data sizes and varying domains of input length. Our results reveal their limitations in the both settings. To alleviate these drawbacks, we propose a unified modular framework that jointly solves subtasks in the GITM-MR problem. By explicitly capturing relation correspondence and utilizing it to propagate messages, our approach is capable of obtaining a comprehensive understanding of global correspondence. Additionally, its lightweight and compositional module design allows for excellent data efficiency and length generalization, accompanied by a more transparent reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.49, + 0.473, + 0.669 + ], + "angle": 0, + "content": "Specifically, we develop a graph neural network on a parsed language scene graph (LSG), named Relation-sensitive Correspondence Reasoning Network (RCRN), which exploits the language structure to perform relation-aware reasoning in a belief space of visual-linguistic alignment. Our RCRN uses a set of primitive reasoning modules to operate on the pre-trained vision-language features under the guidance of LSG, and predict the fine-level grounding in both match and mismatch scenarios. As demonstrated by the experiments, our method outperforms in both data efficiency and length generalization settings, demonstrating its effectiveness in learning relation correspondences with data efficiency and OOD generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.672, + 0.468, + 0.685 + ], + "angle": 0, + "content": "The contribution of our work is summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.701, + 0.472, + 0.769 + ], + "angle": 0, + "content": "- We propose a new VL task GITM-MR that challenges relation correspondence learning in VL pre-trained model, and focusing on both data efficiency and out-of-distribution text length setting, with a new benchmark for GITM-MR." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.77, + 0.472, + 0.811 + ], + "angle": 0, + "content": "- We identify the shortcomings of current VL models through the benchmark, and develop a modular graph network RCRN to address these issues." + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.812, + 0.472, + 0.868 + ], + "angle": 0, + "content": "- Our method achieves superior results on both settings, validating its relation learning efficiency and generalization ability, which hopefully inspires the following exploration of these challenges." + }, + { + "type": "list", + "bbox": [ + 0.131, + 0.701, + 0.472, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.125, + 0.637, + 0.14 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.148, + 0.86, + 0.481 + ], + "angle": 0, + "content": "Probing Tasks for VL Models Despite the rapid development on VL pre-trained models [8, 41, 19, 64, 58, 29], emergent works reveal the limitations and gain some deep understanding on pre-trained representations, including implausible scenes understanding [10], language and vision priors exploration [20], and modality ablation [18]. A considerable number of them have revealed the weakness of VL models in relation understanding. Early work proposed a benchmark FOIL [46, 45] by substituting a word in caption to test non-pretrained VL models on identifying minor differences. Subsequent works [21, 38, 36, 53, 63] further evaluate pre-trained models on detecting subtle differences in relations or predicate-noun dependencies in zero-shot scenarios, and reach a consensus on the poor quality of pre-trained VL representation for relations. FGVE [52] extends this evaluation to the task of fine-grained entailment with model fine-tuning, but lacks evaluation in vision modality. Recently, a new benchmark [42] is proposed to simultaneously assess the concept understanding of representations learned from both modalities. In contrast, our task focuses on generating fine-grained multi-modal output under different fine-tuning settings in order to systematically investigate the relation understanding problem in VL pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.86, + 0.791 + ], + "angle": 0, + "content": "OOD Generalization The general problem of OOD generalization has attracted much attention recently [33, 6, 47, 30, 48]. Among them, several studies have considered the tasks with generalization to processing or generating longer sequences. Some works [3, 54] have studied length generalization in different uni-modal seq-to-seq tasks, and find that fine-tuning pre-trained transformer models can lead to poor out-of-distribution (OOD) performance. To address this issue, various solutions have been proposed, including modifications to attention mechanism [39, 11, 17] and the design of recurrent networks [44, 4], but their application in VL tasks is nontrivial. Other recent works have also revealed the challenge of OOD generalization for the VL tasks, but most of them focuses on the answer bias of Visual Question Answering (VQA) [50, 1, 2]. The attempted solutions include the causal learning [37] and stable learning [51] methods. Perhaps most related to us is the modular network approaches [28, 49, 65], which use program-based design to achieve good generalization for certain OOD scenario such as question domains in VQA. We follow the modular design but focus on the OOD problem induced by the expression complexity in the grounded image text matching task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.86, + 0.868 + ], + "angle": 0, + "content": "Vision-language Matching Image-Text Matching (ITM) and Referring Expression Grounding (REG) are common cross-modal matching tasks in vision-language domain. The problem of ITM aims to predict a global similarity between a text and an image. Classical methods for ITM fol" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2977" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.125, + 0.44, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.219, + 0.473, + 0.257 + ], + "angle": 0, + "content": "Figure 2. The length distribution of sentences containing different numbers of entities in the validation set. Different colors represent sentences with different number of entities." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.267, + 0.472, + 0.363 + ], + "angle": 0, + "content": "low the core idea of learning a common space where linguistic and visual features are compatible [27, 31, 7, 14, 9]. Recent approach [41] concentrates on learning the relation-aware global similarity by a regularization strategy on training. Our model derives a global similarity measure by propagating messages on local correspondences, capable of capturing subtle changes in relations." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.364, + 0.474, + 0.53 + ], + "angle": 0, + "content": "The goal of REG is to locate an object in an image based on a language expression. Most methods focus on learning a holistic visual and linguistic representation in a joint space for aligning the object and text, based on CNN+RNN architectures [35, 66, 25] or the Transformer [12, 13, 32, 40, 59]. Another line of research aims to exploit the linguistic structure in the text for better context modeling [62, 34, 60, 56]. In particular, SGMN [61] parses the text into language components and performs reasoning under the guidance of the linguistic structure. Our method introduces a flexible contextual representation and a bi-directional propagation to extend correspondence reasoning beyond the task of REG." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.544, + 0.377, + 0.56 + ], + "angle": 0, + "content": "3. Problem Setup and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.568, + 0.473, + 0.679 + ], + "angle": 0, + "content": "We aim to study the representation capacity of the VL pretrained models for fine-grained downstream tasks. To this end, we introduce the Grounded Image Text Matching with Mismatched Relation (GITM-MR) task, which is a joint matching and grounding task focusing on relations in vision and language domain. In this section, we present the problem setting of GITM-MR, followed by building a new benchmark for the task." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.686, + 0.472, + 0.825 + ], + "angle": 0, + "content": "GITM-MR Task Given an image and a referring expression, our goal is to determine if the expression describes an object in the image or not (termed as Image-Text Matching (ITM)), and then to localize the referred object if it is matched (termed as Referring Expression Grounding (REG)), or otherwise ground the mismatched relation phrase (termed as Mismatched Relation Reasoning (MRR)). Notably, the only difference between match and mismatch is a relation phrase in the expression, making our task particularly challenging." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.826, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Formally, given an image \\(I\\) and a text description \\(L\\), we denote their matching state as \\(y\\), indicating whether the image matches the description \\((y = 1)\\) or not \\((y = 0)\\). For the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.858, + 0.251 + ], + "angle": 0, + "content": "matched case, the referred object has a corresponding location in the image, represented by its bounding box \\( z \\in \\mathbb{R}^4 \\). For the mismatch case, we denote the set of all possible relation phrases (i.e. predicates) for mismatch reasoning as \\( \\mathcal{R} = \\{r_i\\}_{i=1}^{N_e} \\), and there is a target mismatched relation \\( r_i^* \\) in the relation phrase set \\( \\mathcal{R} \\). Given the input \\( I \\) and \\( L \\), our goal is to first predict \\( y \\) for the input pair, and then predict \\( z \\) for the referred object in \\( L \\) when \\( y = 1 \\), or otherwise predict \\( r_i^* \\) from \\( \\mathcal{R} \\) to explain why \\( y = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.252, + 0.859, + 0.335 + ], + "angle": 0, + "content": "For model training, we follow the convention of ITM and REG tasks, and assume the annotations for those two subtasks are provided. However, it is usually costly to collect annotations for the subtle MRR task. Consequently, we assume a weakly-supervised learning setting where only the matching state label \\( y \\) is given for mismatched data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.859, + 0.552 + ], + "angle": 0, + "content": "Our Benchmark To conduct systematical study, we construct a benchmark for the GITM-MR task, including a new dataset and two challenging evaluation settings. We first build a dataset of image-text pairs from the public REC dataset, Ref-Reasoning [61], by substituting the relation phrases to generate mismatched referring expressions for images. From the provided relation phrases in the LSG offered by the dataset, we manually select a subset of 27 commonly-occurred relations, and assign some contextual close but semantically different ones for each relation in the subset as their mismatched candidates. We control the linguistic bias by several measures, including keeping the relation distribution and checking by language-only model, the details of which are stated in the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.859, + 0.733 + ], + "angle": 0, + "content": "However, replacing the relation phrases may introduce falsely mismatched cases. To ensure the quality of our data, we filter out those false mismatch cases in the test set with the help of annotators. After filtering, the dataset contains more than 1M referring expressions in 60K images. It has 1M, 50K and 10K expression-referent pairs for training, validation, testing, respectively. We only select the original image text pairs which have generated mismatched pairs to keep the matching label balanced. Fig. 2 shows the length distribution of sentences in the validation set of the benchmark. We find it is more general and challenging than previous benchmarks on relation understanding, and refer the reader to see more details of the dataset in Suppl." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.859, + 0.868 + ], + "angle": 0, + "content": "Evaluation Settings Given the dataset, we develop two evaluation settings to test pretrained VL models on their relation understanding. The first setting is designed to investigate data efficiency during the fine-tuning process. We train the models on small-scale datasets and tested on indistribution test sets. The second setting is designed to evaluate the generalization performance of models with regards to different text input lengths. Here the test set contains" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2978" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.127, + 0.803, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.337, + 0.862, + 0.416 + ], + "angle": 0, + "content": "Figure 3. Model overview with a mismatched case. Given an image and an expression, we first generate visual and linguistic candidates by a detector and a language parser, and compute their representations. Then we use the Context-sensitive Propagation Network to inter alignments between visual-linguistic candidates, which conducts bi-directional message propagation based on the language graph. The propagation initializes the messages by computing local beliefs, selectively aggregates the context information and updates the belief with a context-sensitive gating function. Predictions for this case are obtained by exploiting the beliefs from the propagation. In the language graph, the ellipses represent entity phrases and the rectangles stand for relation phrases. Further elaboration can be found in the main text." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.424, + 0.471, + 0.451 + ], + "angle": 0, + "content": "longer and more complex samples on relation combinations compared to the training set." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.451, + 0.474, + 0.618 + ], + "angle": 0, + "content": "Specifically, according to the sentence length distribution in Fig. 2, we construct two training sets, named TrainLen16 and Train-Len11, which includes sentences containing less than 16 and 11 words respectively. In the test phase, the model needs to perform the reasoning task on sentences with lengths less than 11 and 16 (in-distribution), as well as sentences longer than these lengths (out-of-distribution). As two training sets accounts for \\(5\\%\\) and \\(10\\%\\) of the overall training data, those two settings allow us to evaluate the model's data efficiency in fine-tuning and OOD generalization ability simultaneously. Detailed data partition rules are explained in Suppl." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.628, + 0.248, + 0.644 + ], + "angle": 0, + "content": "4. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.651, + 0.472, + 0.79 + ], + "angle": 0, + "content": "In this section, we present our model design for the GITM-MR task and the learning objectives for training with weakly-annotated data. To tackle the problem, we develop a unified and generalizable framework as outlined in Fig. 1, which consists of two main components: a representation network that computes an initial representation of the linguistic and visual components, and a context-aware correspondence reasoning network for final match predictions. Below we first introduce the two main model components in Sec. 4.1 and 4.2, and then the model learning in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.797, + 0.444, + 0.811 + ], + "angle": 0, + "content": "4.1. Candidate Generation and Representation" + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.818, + 0.306, + 0.831 + ], + "angle": 0, + "content": "4.1.1 Candidate Generation" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.839, + 0.472, + 0.868 + ], + "angle": 0, + "content": "Our method first generates a set of visual objects from the image and phrase candidates from the corresponding text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.86, + 0.577 + ], + "angle": 0, + "content": "Specifically, for the visual object candidates, we adopt a pre-trained object detector [64] to generate \\(N_{o}\\) box proposals \\(\\mathcal{O} = \\{o_i\\}_{i=1}^{N_o}\\) from the image \\(I\\). The detector also provides an initial visual feature (i.e., pooled ROI feature) for each object proposal. For the language phrase candidates, we utilize an off-the-shelf parser [43] to produce a language scene graph \\(\\mathcal{G} = (\\mathcal{E}, \\mathcal{R})\\) of the text expression \\(L\\), which consists of a set of entity phrases \\(\\mathcal{E} = \\{e_j\\}_{i=1}^{N_e}\\) as its nodes and relation phrases as its edges \\(\\mathcal{R} = \\{r_{ij}\\}\\), where \\(r_{ij}\\) connects to subject \\(e_i\\) and object \\(e_j\\). We denote the number of relation phrases as \\(|\\mathcal{R}| = N_r\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.592, + 0.719, + 0.606 + ], + "angle": 0, + "content": "4.1.2 Candidate Representation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.859, + 0.642 + ], + "angle": 0, + "content": "Given the visual and linguistic candidates, we compute their representations in two steps, as detailed below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.86, + 0.743 + ], + "angle": 0, + "content": "Token-level Representation The first step generates an initial representation for the word tokens and visual objects based on a pre-trained vision-language model. Here we adopt a shallow version of the UNITER model [8], aiming to compute a feature representation for the input tokens in two modalities with a generic vision-language alignment from the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.759, + 0.859, + 0.787 + ], + "angle": 0, + "content": "Candidate Features In the second step, we then compile the token-level representations into a set of features for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.8, + 0.86, + 0.868 + ], + "angle": 0, + "content": "*We note that such a generic representation module has a limited capacity and is typically insufficient for our downstream tasks via a simple fine-tuning procedure as in [8, 58]. Nevertheless, it provides rich semantic features for the subsequent alignment reasoning network (c.f. Sec. 4.2) after a minimal fine-tuning (described in Sec. 5.1), which complements our explicit reasoning process." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2979" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.127, + 0.471, + 0.168 + ], + "angle": 0, + "content": "our visual and linguistic components used in the subsequent reasoning network. Here we adopt a similar strategy as in SGMN [61] for computing the features." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.169, + 0.474, + 0.378 + ], + "angle": 0, + "content": "Concretely, for visual objects \\(o_i\\), we augment its token-level feature \\(\\mathbf{o}_i\\) by its spatial location \\(\\mathbf{l}_i^o\\) (i.e., bounding box parameters). In addition, for every object pair \\((o_i,o_j)\\) in the image, we represent its relational feature as \\(\\mathbf{r}_{ij}^{o}\\), which is computed from the features of its two objects. For the linguistic entities in \\(\\mathcal{E}\\) and relation phrases in \\(\\mathcal{R}\\), we feed the corresponding word tokens into a Bi-LSTM [22] with self-attention [24] to generate two sets of features. In the first set, each entities \\(e_i\\) is encoded by a tuple of three features \\((\\mathbf{e}_i,\\mathbf{l}_i^e,\\mathbf{h}_i^e)\\), where \\(\\mathbf{e}_i\\) is an embedding into the space of object feature \\(\\mathbf{o}_i\\), \\(\\mathbf{l}_i^e\\) is an embedding into the space of object location \\(\\mathbf{l}_i^o\\), and \\(\\mathbf{h}_i^e\\) is a linguistic feature from the LSTM states. For the second feature set, we generate a representation \\(\\mathbf{r}_{ij}^{e}\\) for each relation phrase \\(r_{ij}\\in \\mathcal{R}\\). We refer the reader to Suppl. for the details of our feature design." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.386, + 0.422, + 0.401 + ], + "angle": 0, + "content": "4.2. Context-sensitive Propagation Network" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.408, + 0.474, + 0.573 + ], + "angle": 0, + "content": "We now introduce our second model component, the Context-sensitive Propagation Network (CPN), which aims to infer the alignment between the visual-linguistic candidates and predict the match label of input image-text pair. To this end, we build a modular graph neural network on the language graph \\(\\mathcal{G}\\), which takes the candidate features as input and performs context-aware reasoning in a belief space of visual-linguistic alignment. Such a structured reasoning process enables us to exploit the regularity in language, i.e. the connections of entities by relations, leading to a relation-aware and compositional alignment model with potentially better generalization." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.575, + 0.472, + 0.783 + ], + "angle": 0, + "content": "Specifically, as illustrated in the language graph in Fig. 3, the \\(i\\)-th node of the graph network corresponds to an entity phrase \\(e_i\\) and the edge between the node \\(i\\) and \\(j\\) corresponds to a relation phrase \\(r_{ij}\\). We follow the common assumption that the corresponding \\(\\mathcal{G}\\) has a tree structure [61, 34] and the referred entity is regarded as the root node. To perform reasoning, the graph network uses a bi-directional message propagation procedure to infer a relation-aware correspondence belief (w.r.t. the visual candidates) for all the entity nodes (described in Sec. 4.2.1), followed by a readout head module to predict the fine-grained and global match scores (described in Sec. 4.2.2). Finally, we formulate each key operation in the graph reasoning as an easy-to-interpret primitive module, and convert our reasoning process into an explainable program in Sec. 4.2.3." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.802, + 0.391, + 0.817 + ], + "angle": 0, + "content": "4.2.1 Bi-directional Message Propagation" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.826, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Given the graph network, we aim to compute a message for each node, representing a belief on the correspondence between the associated entity phrase and all the visual objects" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.858, + 0.181 + ], + "angle": 0, + "content": "in the image. In order to resolve the ambiguity in the local matching of entities, we introduce a bi-directional message propagation procedure to aggregate all the local beliefs on phrase-visual region correspondence throughout the graph." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.859, + 0.333 + ], + "angle": 0, + "content": "Concretely, inspired by belief propagation [5], we perform two parallel passes of message propagation on the tree graph \\(\\mathcal{G}\\) along two directions: 1) bottom-up direction (denoted as bp) first computes messages from the leaf nodes and then updates their parents recursively until the root is reached; 2) top-down direction (denoted as td) starts from the root node and updates the children nodes recursively until reaching all the leaves. For each direction, the message propagation aggregates the context information on cross-modal matching and update the belief on phrase alignment at each node in the following three steps:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.859, + 0.43 + ], + "angle": 0, + "content": "(1) Message initialization: We initialize a local belief \\(\\mathbf{b}_i^{\\mathrm{loc}}\\) at node \\(i\\) based on a similarity measure between the entity \\(e_i\\) and all the visual objects. Here we define the similarity by a weighted combination of feature similarities in appearance and spatial location space. Specifically, for each pair of entity \\(e_i\\) and object \\(o_j\\), we compute two similarities as follows," + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.438, + 0.858, + 0.456 + ], + "angle": 0, + "content": "\\[\nb _ {i j} ^ {\\mathrm {a p p}} = \\mathrm {F} _ {\\mathrm {s i m}} (\\mathbf {e} _ {i}, \\mathbf {o} _ {j}), \\quad b _ {i j} ^ {\\mathrm {p o s}} = \\mathrm {F} _ {\\mathrm {s i m}} \\left(\\mathrm {I} _ {i} ^ {e}, \\mathrm {I} _ {j} ^ {o}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.859, + 0.503 + ], + "angle": 0, + "content": "where \\( \\mathrm{F}_{\\mathrm{sim}} \\) is an MLP network computing the similarity between features from two modalities. The local belief \\( \\mathbf{b}_i^{\\mathrm{loc}} \\) then integrates two similarities as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.511, + 0.858, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {i} ^ {\\text {l o c}} = \\mathrm {F} _ {\\text {n o r m}} \\left(\\beta_ {i} ^ {\\text {a p p}} \\mathbf {b} _ {i} ^ {\\text {a p p}} + \\beta_ {i} ^ {\\text {p o s}} \\mathbf {b} _ {i} ^ {\\text {p o s}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.859, + 0.594 + ], + "angle": 0, + "content": "where \\(\\mathbf{b}_i^t = [b_{ij}^t]_{j=1}^{N_o}\\), \\(t \\in \\{\\mathrm{app, pos}\\}\\). Here \\(\\beta_i^t\\) is the weight for combining similarities and is computed as \\(\\beta_i^t = \\mathrm{sigmoid}(\\mathbf{W}_t^\\top[\\mathbf{h}_i^e, \\mathbf{b}_i^{\\mathrm{app}}, \\mathbf{b}_i^{\\mathrm{pos}}])\\), where \\(\\mathrm{F}_{\\mathrm{norm}}\\) is a normalization function to rescale its input elements to \\([0,1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.859, + 0.678 + ], + "angle": 0, + "content": "(2) Message aggregation: According to the direction \\( \\mathbf{td} \\) or \\( \\mathbf{bp} \\), message aggregation collects beliefs from the upstream neighbors in a context-sensitive manner. Formally, we first prune the local beliefs by selecting top \\( K \\) maximum values, which significantly reduces the computation complexity for large \\( N_{o} \\) and is written as," + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.684, + 0.858, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {j} ^ {\\text {s e l}} = \\mathrm {F} _ {\\text {t o p k}} \\left(\\mathbf {b} _ {j} ^ {\\text {l o c}}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.859, + 0.779 + ], + "angle": 0, + "content": "In contrast to [61], our selection is belief-based and more informative for propagation. Given the sparsified beliefs, we then aggregate the messages from the neighboring nodes by taking into account the alignment of the relation phrases on the edges as follows," + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.786, + 0.858, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {i} ^ {\\mathrm {a g g}} = \\prod_ {j \\in \\operatorname {C t x} (i)} \\mathbf {A} _ {i j} \\mathbf {b} _ {j} ^ {\\mathrm {s e l}}, [ \\mathbf {A} _ {i j} ] _ {k l} = \\mathrm {F} _ {\\text {s i m}} \\left(\\mathbf {r} _ {i j} ^ {e}, \\mathbf {r} _ {k l} ^ {o}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.859, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathrm{Ctx}(i)\\) denotes the upstream neighbor set of node \\(v_{i}\\) according to the propagation direction. We model the messages provided by relations explicitly through alignment on" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.515, + 0.902 + ], + "angle": 0, + "content": "2980" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.127, + 0.473, + 0.209 + ], + "angle": 0, + "content": "edges, which forces the model to capture the nuances of different relations. For the aggregation operator, we take the element-wise product of beliefs, which can be viewed as a soft version of logical AND. Unlike the min pooling used in [49, 23], our design leads to a more robust and smooth learning procedure." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.209, + 0.473, + 0.265 + ], + "angle": 0, + "content": "(3) Message update: Finally, we update the belief \\(\\mathbf{b}_i\\) at node \\(i\\) by a context-sensitive integration of the aggregated neighboring belief and the local belief. To achieve this, we design a gated product as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.271, + 0.473, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\beta_ {i} ^ {\\mathrm {r e l}} = \\operatorname {s i g m o i d} \\left(\\mathbf {W} _ {\\mathrm {r e l}} ^ {\\top} \\left[ \\mathbf {h} _ {i} ^ {e}, \\mathbf {b} _ {i} ^ {\\text {a p p}}, \\mathbf {b} _ {i} ^ {\\text {p o s}} \\right]\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.29, + 0.473, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {i} = \\mathrm {F} _ {\\text {n o r m}} \\left(\\mathbf {b} _ {i} ^ {\\text {l o c}} \\circ \\left(\\mathbf {b} _ {i} ^ {\\text {a g g}}\\right) ^ {\\beta_ {i} ^ {\\text {r e l}}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.32, + 0.473, + 0.375 + ], + "angle": 0, + "content": "where \\(\\circ\\) denotes the element-wise product. Instead of using the language features as in [61, 55], our gating function also takes the local beliefs as input, making the network more sensitive to the image context." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.376, + 0.473, + 0.471 + ], + "angle": 0, + "content": "Given the updated beliefs, we compute a confidence score for each node matching to any object in the image by applying an MLP network, denoted as \\( p_i = \\mathrm{sigmoid}(\\mathrm{MLP}_{\\mathrm{match}}(\\mathbf{b}_i)) \\). In the end, the message propagation along two directions generates two belief vectors \\( (\\mathbf{b}_i^{\\mathrm{bp}} \\) and \\( \\mathbf{b}_i^{\\mathrm{td}}) \\) and two confidence scores \\( (p_i^{\\mathrm{bp}} \\) and \\( p_i^{\\mathrm{td}}) \\) for each node \\( i \\)." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.481, + 0.255, + 0.494 + ], + "angle": 0, + "content": "4.2.2 Readout Head" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.503, + 0.473, + 0.56 + ], + "angle": 0, + "content": "After the graph-based reasoning, we now introduce a readout head network to generate the predictions for three subtasks, i.e., ITM, REG and MRR, by exploiting the beliefs and confidence scores from the bi-directional propagation." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.562, + 0.473, + 0.603 + ], + "angle": 0, + "content": "Image-Text Matching To predict the global matching label \\(y\\), we take a minimum pooling of the matching confidence scores across the entire graph as below," + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.61, + 0.473, + 0.634 + ], + "angle": 0, + "content": "\\[\nP (y = 1 | I, L) = \\min _ {i} \\left(\\min \\left(p _ {i} ^ {\\mathrm {b p}}, p _ {i} ^ {\\mathrm {t d}}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.64, + 0.473, + 0.724 + ], + "angle": 0, + "content": "REG Task For the referent grounding, we take the belief vector of the root node \\(\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}}\\) from the bottom-up direction and locate the matching visual object \\(o_{i^{*}}\\) by choosing the largest belief value: \\(i^{*} = \\arg \\max_{i}\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}}\\). A standard regression head is also employed for more precise localization and its details are left to the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.726, + 0.473, + 0.809 + ], + "angle": 0, + "content": "MRR Task For mismatched relation prediction, we exploits the inconsistency of the match confidence scores along the graph edges. We found the mismatched relation typically leads to a confidence gap between its two incident nodes. As such, we choose the edge with the largest match confidence difference between its two nodes:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.815, + 0.473, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\hat {r} = \\arg \\min _ {r _ {i j} \\in \\mathcal {R}} \\left(\\left\\{\\left(p _ {i} ^ {\\mathrm {b p}} - p _ {j} ^ {\\mathrm {b p}}\\right) + \\left(p _ {j} ^ {\\mathrm {t d}} - p _ {i} ^ {\\mathrm {t d}}\\right) \\right\\}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.846, + 0.473, + 0.873 + ], + "angle": 0, + "content": "where we sum the differences from two propagation directions to make the decision more robust." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.123, + 0.86, + 0.174 + ], + "angle": 0, + "content": "Table 1. The full list of the reasoning modules and their implementations. We use bold lowercase variables to represent vectors and bold uppercase variables to represent matrices. Regular variables are scalars. W is trainable parameter matrix." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.175, + 0.862, + 0.364 + ], + "angle": 0, + "content": "
ModulesInOutImplementation
Local Correspondences:
Simv'txt, v'visbFsim(v'txt, v'vis)
GateSum{vi}, {bi}bFnorm(∑i sigmoid(W^T[vi, bi])bi)
Correspondence Reasoning:
Selectbb'Ftopk(b)
Aggregate{Ai}, {bi}b'Πi(AiBi)
GateProdb, b', βib''Fnorm(b o (b')βi)
ClassifybpMLPmatch(b)
Readout Head:
Locatepiarg min(p)
Comparepi, pjdpi - pj
And{pi}pminpi
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.761, + 0.391 + ], + "angle": 0, + "content": "4.2.3 Modular Network Interpretation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.86, + 0.719 + ], + "angle": 0, + "content": "We now introduce an interpretation of our using the modular network framework. As shown in Tab. 1, we summarize the key operations in the RCRN into a set of primitive modules with semantic meanings. This allows us to build a explainable reasoning program: (1) In the message initialization step, local beliefs are computed using Sim, and are combined by GateSum to generate the initial belief on each node. (2) For message aggregation, we first use Select to prune each node and then apply Aggregate to collect the messages from all neighbors, considering the alignment of relations particularly. (3) For message update, GateProd is used to combine the aggregated message with the local belief to generate the updated belief on each node. The Classify module provides a confidence score indicating the likelihood that the phrase entity matches an visual object. (4) In the readout head, we use two Compare modules in two directions respectively for MRR and one And operation for the ITM task, and finally generate predictions for all three subtasks via Locate module. For instance, the grounding prediction is achieved by \\( \\text{Locate}(-\\mathbf{b}_{\\text{ref}}) \\). Several example are provided in Sec. 5.4 and Suppl. to illustrate the interpretable reasoning process, along with further detailed implementations of the modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.86, + 0.79 + ], + "angle": 0, + "content": "We note that such a program formulation indicates that the model can enjoy a good out-of-domain generalization due to the compositionality of language graphs. Even facing with more complex text structures, we can perform this reasoning process by reusing those modules on new graphs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.803, + 0.649, + 0.818 + ], + "angle": 0, + "content": "4.3. Model Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.86, + 0.868 + ], + "angle": 0, + "content": "We train the network on a multitask loss with end-to-end back propagation. Cross-Entropy (CE) loss and Binary Cross-Entropy (BCE) loss are used for the REG and ITM" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.513, + 0.902 + ], + "angle": 0, + "content": "2981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.127, + 0.232, + 0.141 + ], + "angle": 0, + "content": "tasks, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.15, + 0.472, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g r d}} = \\mathrm {C E} \\left(\\mathbf {p} _ {\\mathrm {g r d}}, \\mathbf {p} _ {\\mathrm {g r d}} ^ {*}\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.188, + 0.472, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h}} = \\mathrm {B C E} \\left(p _ {\\text {m a t c h}}, y ^ {*}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.21, + 0.473, + 0.307 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{\\mathrm{grad}} = \\mathrm{softmax}(\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}})\\) stands for the predicted grounding distribution, \\(\\mathbf{p}_{\\mathrm{grad}}^{*}\\) denotes the ground-truth one-hot grounding label, and \\(p_{\\mathrm{match}} = P(y = 1|I,L)\\) denotes the predicted match probability. When the object boxes are from a detector, the box with the maximum IoU with the ground-truth box of the referent is labeled as the localization output. The overall loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.318, + 0.472, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {g r d}} + \\mu \\mathcal {L} _ {\\mathrm {m a t c h}} + \\omega \\mathcal {L} _ {\\mathrm {r e g}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.343, + 0.473, + 0.399 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\omega\\) are hyper-parameters to balance the three terms. \\(\\omega\\) would be zero if the IoU of the ground-truth box and the selected proposal is less than 0.5. The regression loss is standard and can be found in Suppl." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.41, + 0.235, + 0.426 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.433, + 0.474, + 0.543 + ], + "angle": 0, + "content": "In this section, we conducted experiments on the proposed benchmark to test several baselines of VL pre-trained models on inferring relation correspondences, with a specific focus on their data efficiency and length generalization capabilities. We then compare our model with those strong baselines to show its effectiveness in these settings. Finally we illustrate the visualizations and interpretations of the reasoning process of our model." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.551, + 0.275, + 0.567 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.573, + 0.472, + 0.711 + ], + "angle": 0, + "content": "Metrics The evaluation metrics include classification accuracy for three subtasks. The grounding result for a matched case is considered as correct when it is identified as matching and the predicted box has at least 0.5 IoU with its ground-truth location. The grounding accuracy (i.e. Recall@1) is the ratio of correctly grounded cases. For mismatch reasoning, a mismatched case needs to be correctly classified and the mismatched relation should be accurately selected from the candidate set. The MRR accuracy is the top-1 accuracy among the candidates." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.715, + 0.473, + 0.867 + ], + "angle": 0, + "content": "VL Pre-trained Models We benchmark five VL models on the proposed dataset. We first choose TCL [58], UNITER [8] and FIBER [16], where TCL is the state-of-the-art pre-trained model for image-text retrieval, UNITER is a vanilla transformer-based VL pre-trained model for matching and grounding tasks, and FIBER uses a specific fine-grained pre-training design. Additionally, we adapt the latest fine-grained visual entailment (FGVE) model [52] to our setting. We compare its potential fine-grained multitask capability with ours, which is designed based on the VL pre-trained model Oscar+[64]. Because none of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.86, + 0.251 + ], + "angle": 0, + "content": "pretrained VL models can fulfill three subtasks jointly, we also apply the multiple instance learning strategy [57] on UNITER to cope with the weakly-supervised MRR task. As shown in Suppl, we modify those models by adding subtask heads to perform three subtasks. We note that TCL and UNITER use the same pre-training datasets and the pretraining corpus of FIBER and FGVE is larger while all the models are fine-tuned on the GITM-MR, which leads to a relatively fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.86, + 0.436 + ], + "angle": 0, + "content": "Implementation Details For generating object candidates, we take the off-the-shelf detector, VinVL [64], to extract at most 100 proposals with object score great than 0.1 for each image. For model training, the reasoning component of RCRN is trained by the Adam optimizer with the learning rate set to \\(5\\mathrm{e} - 4\\) for the proposed dataset. The learning rate of candidate representation module stated in Sec. 4.1 is set to \\(5\\mathrm{e} - 7\\). We train 80K iterations with batchsize 64 totally. The hyper-parameter \\(K\\), \\(\\mu\\) and \\(\\omega\\) are set as 5, 3 and 1 based on validation, respectively. The representation module uses the first 6 transformer encoder layers in the pre-trained UNITER. The ablation study on layers and more implementation details are included in the Suppl." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.714, + 0.459 + ], + "angle": 0, + "content": "5.2. Results and Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.86, + 0.784 + ], + "angle": 0, + "content": "Relation Understanding with Limited Data Tab. 2 (In-Distribution part) shows the performances of pre-trained models on the relation understanding under the condition of limited data. The results indicate that TCL struggles to achieve a match accuracy of more than chance. This could be attributed to its emphasis on global image-text matching during pre-training, which may hinder its ability to perform fine-grained understanding tasks. Though FGVE is designed for fine-grained entailment prediction, it still lacks capability on mismatch relation reasoning in our scenario. This is likely due to its reliance on the pre-trained VL model for the alignment between vision and language relations. UNITER and FIBER perform relatively better on the match task than TCL because they both use the World-Region Alignment (WRA) as a pre-training task, and FIBER is also pre-trained with phrase grounding, which helps to establish fine-grained correspondence between image and text. The overall conclusion drawn from these findings is that pre-trained models exhibit unsatisfying performance on the challenging benchmark under the limited data condition, indicating the need for further development in establishing effective visual-linguistic correspondence for relation understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.861, + 0.868 + ], + "angle": 0, + "content": "In comparison, our RCRN outperforms the SOTA pretrained models on match and grounding tasks in the indistribution scenarios, even with fewer parameters. The model's performance with limited training data indicates the effectiveness of its lightweight and structured design. The design incorporates appropriate inductive biases based" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.516, + 0.903 + ], + "angle": 0, + "content": "2982" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.122, + 0.859, + 0.161 + ], + "angle": 0, + "content": "Table 2. Experiment results of testing models' relation understanding with limited data and ability of length generalization. UNITER+MIL represents the UNITER with multiple instance learning strategy applied. FGVE+MAX represents the fine-grained visual entailment model with max pooling on knowledge elements to predict mismatched relation. - means the model can't handle the corresponding task." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.161, + 0.859, + 0.337 + ], + "angle": 0, + "content": "
Training SetMethod#Pretrain Images#ParamFull TestIn-DistributionOut-of-Distribution
Match%Grounding%MRR%Match%Grounding%MRR%Match%Grounding%MRR%
Train-Len16TCL [58]4M210M52.77--53.11--52.70--
UNITER [8]4M111M61.9324.16-71.2940.63-59.9420.67-
UNITER+MIL [57]4M111M59.1318.4647.568.8341.5361.3857.0713.5644.53
FIBER [16]4.8M254M52.319.01-59.0437.15-50.883.03-
FGVE+MAX [52]5.7M113M54.30-33.2358.37-51.7953.43-29.26
RCRN (Ours)4M90M66.5929.1055.7271.8546.2459.0465.4725.4655.01
Train-Len11TCL [58]4M210M51.71--52.08--51.70--
UNITER [8]4M111M55.5110.24-68.2230.05-54.979.42-
UNITER+MIL [57]4M111M55.4311.4442.7365.0429.5661.6555.0210.6941.93
FIBER [16]4.8M254M53.5211.74-57.2129.56-53.3611.00-
FGVE+MAX [52]5.7M113M53.40-38.0555.26-40.7853.31-37.93
RCRN (Ours)4M90M63.4121.4552.6768.7037.9357.7763.1920.7752.46
" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.36, + 0.473, + 0.401 + ], + "angle": 0, + "content": "on sentence structure to establish both relation and entity correspondence in a visual-linguistic context, and learns the reasoning strategy effectively." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.409, + 0.473, + 0.589 + ], + "angle": 0, + "content": "Length Generalization Results about length generalization ability are shown in the Out-of-Distribution part of Tab. 2. Overall, the OOD performance of pre-trained models on both test settings was generally poor compared to their in-distribution performance, especially on the grounding and MRR subtasks, which explicitly reflect the models' understanding of relations. In particular, in the OOD setting, FIBER's match accuracy is similar to TCL, with both models only slightly exceeding \\(50\\%\\). The OOD results suggest that these models tend to overfit to the training data biases, without truly establishing the universal VL correspondence of the relations, thus preventing them from generalizing to longer sentences." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.59, + 0.473, + 0.728 + ], + "angle": 0, + "content": "In contrast, RCRN achieved notably better performance than the other models on all three subtasks in both OOD scenarios. Specifically, RCRN trained with Train-Len11 outperformed the best performance of other models by \\(8.17\\%\\), \\(10.08\\%\\), and \\(10.53\\%\\) on all three tasks. This success can be attributed to the modular design, which helps the model learn to establish local visual-linguistic correspondence and the universal reasoning strategy. During test, the model can exploit the regularity in sentence structures to compose the modules and deal with longer sentences." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.743, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Discussion These experiments not only help us investigate the two challenges that the pre-trained models faced, but also lead to other interesting findings. First, our observation that fine-grained pre-training tasks, such as WRA, can help models learn local correspondences on downstream tasks, is consistent with the findings of [36]. Second, our findings in the OOD scenario are consistent with previous works [3, 54], which suggest that pre-trained models are prone to overfitting to the training data length biases." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.356, + 0.848, + 0.369 + ], + "angle": 0, + "content": "Table 3. Ablation study on the benchmark's OOD validation set." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.371, + 0.848, + 0.458 + ], + "angle": 0, + "content": "
LangβCtxβBi-MPVLPMatGrdMRR
---√(w/o MP)54.7913.9941.47
--✓✓59.6523.8045.58
✓-✓✓60.8824.1849.57
✓✓-✓60.0723.8848.62
✓✓✓-60.7029.2536.65
✓✓✓✓61.4723.3753.23
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.859, + 0.581 + ], + "angle": 0, + "content": "Moreover, to independently investigate the capability on two subtasks, namely grounding and MRR, we also evaluate on an oracle setting, where we suppose that the models always predict correctly on the matching task. The experiment demonstrates our RCRN has strong performance on both subtasks, and the details of this experiment are provided in the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.859, + 0.68 + ], + "angle": 0, + "content": "Overall, our experiments provide insights into the challenges and potential solutions for the relation understanding in VL models. The fine-grained pre-training tasks can help models learn local correspondences, while the modular design can facilitate the composition of modules to deal with longer sentences. We believe that our findings offer a promising direction to effectively address these challenges." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.691, + 0.637, + 0.706 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.859, + 0.755 + ], + "angle": 0, + "content": "In this section, we present the effectiveness of each component in the proposed RCRN through ablative experiments trained on the Train-Len16 set, with the results in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.86, + 0.868 + ], + "angle": 0, + "content": "We conduct ablation study by removing some specific designs. \\(\\mathbf{Lang}\\beta\\) stands for using the language features as one input for gate functions in Eq. 2 and Eq. 6. Removing it indicates constant \\(\\beta\\)'s in the propagation. \\(\\mathbf{Ctx}\\beta\\) means feeding initial similarities into the gate functions. Bi-MP represents propagating twice individually using different \\(\\beta^{\\mathrm{rel}}\\)'s for matching task reasoning. Without this part, only the propagation from the root to leaves is conducted. VLP" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.128, + 0.48, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.292, + 0.472, + 0.301 + ], + "angle": 0, + "content": "(a) There is a guy wearing the black shirt and to the right of the boy that holding a cell phone." + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.305, + 0.477, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.411, + 0.466, + 0.42 + ], + "angle": 0, + "content": "(b) There is a guy wearing the black shirt and to the left of the boy that holding a cell phone." + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.426, + 0.473, + 0.528 + ], + "angle": 0, + "content": "Figure 4. The visualization of the reasoning process of RCRN. The entity and relation correspondence maps are shown by color maps on boxes and box pairs respectively, where only those with top 5 correspondence scores for each phrase are included. Small color maps in (a) are the local correspondence maps and the large ones are the propagated contextualized maps. The boxes of parents in box pairs are dotted on their centers. Dashed boxes indicate the pooling results." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.554, + 0.472, + 0.622 + ], + "angle": 0, + "content": "stands for introducing vision-language pre-trained transformer layers to encode features. Particularly, the first line with VLP only stands for the model which consists of 6 transformer layers from UNITER and the UNITER task heads stated in Sec. 5.1 and Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.624, + 0.472, + 0.734 + ], + "angle": 0, + "content": "The results show that all of these designs improve the matching performance. VLP model (first line in Tab. 3) with the same setting as the VLP layers in RCRN performs poorly on OOD scenario, which demonstrates our footnote statement in Sec. 4.1. Nevertheless, VLP component of RCRN is significant on MRR performance, probably because it provides accurate preliminary vision-language alignments on relations." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.747, + 0.382, + 0.763 + ], + "angle": 0, + "content": "5.4. Visualization and Interpretability" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.77, + 0.472, + 0.867 + ], + "angle": 0, + "content": "As shown in Fig. 4, we visualize a pair of matched and mismatched samples of our RCRN. Fig. 4(a) shows the bottom-up propagation process followed by grounding prediction on a matched image-text pair, and Fig. 4(b) presents the bi-directional propagation results with readout process for MRR on a mismatched pair. In Fig. 4(a), the propagation eliminates the ambiguities on entities under the guid" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.123, + 0.858, + 0.16 + ], + "angle": 0, + "content": "Table 4. Intermediate diagnosis results of RCRN. E.G.R. means Entity Grounding Recall. R.G.R. means Referent Grounding Recall. R.C.S. means Relation Correspondence Scores." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.161, + 0.858, + 0.228 + ], + "angle": 0, + "content": "
Inference MethodE.G.R.R.G.R.R.C.S.
R@1R@3R@5R@1R@3R@5GTMeanMax
By chance6.3718.2528.785.7416.9427.22---
w/o MP50.5080.5989.7938.8875.8488.110.610.170.69
w/ MP60.8285.5892.1669.4690.9295.590.610.170.69
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.858, + 0.319 + ], + "angle": 0, + "content": "ance of explicit relation correspondences. Fig. 4(b) illustrates that the intermediate reasons of the predictions can be traced. The sum of match confidence differences along the mismatched relation, denoted as \\(d\\) in the dotted box, are notably lower than other \\(d\\)'s." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.858, + 0.458 + ], + "angle": 0, + "content": "Tab. 4 demonstrates some interpretable intermediate diagnosis results of RCRN on a subset of the validation set, with ground-truth box proposals. We analyze from the aspect of phrase-level visual grounding recall and statistics on relation correspondence scores. To obtain the true correspondence between the language scene graph and the bounding boxes, we search the parsed language scene graph in the original image scene graph in GQA dataset [26] with a noisy graph matching algorithm. See Suppl. for the detail of the algorithm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.858, + 0.582 + ], + "angle": 0, + "content": "The results verify that reasonable local correspondences are built, since the mean of initial recall (i.e. inference without MP) on entity phrase grounding is much higher than the random guess. As for relation correspondences, the average values on the ground-truth vision-language relation correspondences are far above the average mean value on all the pairs and very close to the average maximum. Moreover, message propagation notably helps refine the correspondence maps, especially on the referent entities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.608, + 0.623 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.859, + 0.867 + ], + "angle": 0, + "content": "In this paper, we have introduced a novel VL joint task, Grounded Image Text Matching with Mismatched Relation (GITM-MR), which provides a challenging benchmark for evaluating pre-trained models on the relation reasoning. Our experiments have shown that some state-of-the-art pretrained models struggle with the task under the setting of limited data and out-of-distribution sentence lengths. To address this problem, we develop a Relation-sensitive Correspondence Reasoning Network (RCRN) to compute contextualized cross-modal alignment of both entities and relations, and ground the fine-grained result in the image or expression. The proposed method outperforms prior SOTA pre-trained models, which demonstrates its strong generalization and data efficiency. Our work sheds light on the limitations of current pre-trained models and the importance of establishing fine-grained vision-language correspondence for relation understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.891, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2984" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.125, + 0.201, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.148, + 0.474, + 0.212 + ], + "angle": 0, + "content": "[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.12, + 0.214, + 0.473, + 0.251 + ], + "angle": 0, + "content": "[2] Arjun R Akula. Question generation for evaluating cross-dataset shifts in multi-modal grounding. arXiv preprint arXiv:2201.09639, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.254, + 0.474, + 0.317 + ], + "angle": 0, + "content": "[3] Cem Anil, Yuhuai Wu, Anders Andreassen, Aitor Lewkowycz, Vedant Misra, Vinay Ramasesh, Ambrose Slone, Guy Gur-Ari, Ethan Dyer, and Behnam Neyshabur. Exploring length generalization in large language models. arXiv preprint arXiv:2207.04901, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.319, + 0.473, + 0.382 + ], + "angle": 0, + "content": "[4] Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. End-to-end algorithm synthesis with recurrent networks: Logical extrapolation without overthinking. arXiv preprint arXiv:2202.05826, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.384, + 0.473, + 0.421 + ], + "angle": 0, + "content": "[5] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.424, + 0.473, + 0.448 + ], + "angle": 0, + "content": "[6] Peter Buhlmann. Invariance, causality and robustness. Statistical Science, 35(3):404-426, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.451, + 0.473, + 0.5 + ], + "angle": 0, + "content": "[7] Tianlang Chen and Jiebo Luo. Expressing objects just like words: Recurrent visual embedding for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10583-10590, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.503, + 0.473, + 0.566 + ], + "angle": 0, + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.568, + 0.473, + 0.644 + ], + "angle": 0, + "content": "[9] Mengjun Cheng, Yipeng Sun, Longchao Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, et al. Vista: vision and scene text aggregation for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5184-5193, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.647, + 0.472, + 0.684 + ], + "angle": 0, + "content": "[10] Myung Jin Choi, Antonio Torralba, and Alan S Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.686, + 0.473, + 0.735 + ], + "angle": 0, + "content": "[11] JU Da, Stephen Roller, Sainbayar Sukhbaatar, and Jason E Weston. Staircase attention for recurrent processing of sequences. In Advances in Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.738, + 0.473, + 0.8 + ], + "angle": 0, + "content": "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.803, + 0.473, + 0.866 + ], + "angle": 0, + "content": "[13] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. arXiv preprint arXiv:2206.06619, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.148, + 0.474, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.858, + 0.178 + ], + "angle": 0, + "content": "[14] Haiwen Diao, Ying Zhang, Lin Ma, and Huchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.858, + 0.242 + ], + "angle": 0, + "content": "[15] Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.245, + 0.858, + 0.307 + ], + "angle": 0, + "content": "[16] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.858, + 0.347 + ], + "angle": 0, + "content": "[17] Yann Dubois, Gautier Dagan, Dieuwke Hupkes, and Elia Bruni. Location attention for extrapolation to longer sequences. arXiv preprint arXiv:1911.03872, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.858, + 0.41 + ], + "angle": 0, + "content": "[18] Stella Frank, Emanuele Bugliarello, and Desmond Elliott. Vision-and-language or vision-for-language. On CrossModal Influence in Multimodal Transformers.(2021). DOI: https://doi.org/10.18653/v1/2021.emnlp-main, 775, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.858, + 0.463 + ], + "angle": 0, + "content": "[19] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.465, + 0.858, + 0.529 + ], + "angle": 0, + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.858, + 0.58 + ], + "angle": 0, + "content": "[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 3635-3644, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.582, + 0.858, + 0.607 + ], + "angle": 0, + "content": "[22] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.858, + 0.672 + ], + "angle": 0, + "content": "[23] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 804-813, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.858, + 0.737 + ], + "angle": 0, + "content": "[24] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1115-1124, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.739, + 0.858, + 0.801 + ], + "angle": 0, + "content": "[25] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16888-16897, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.858, + 0.866 + ], + "angle": 0, + "content": "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.858, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2985" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.178 + ], + "angle": 0, + "content": "[27] Zhong Ji, Haoran Wang, Jungong Han, and Yanwei Pang. Saliency-guided attention network for image-sentence matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5754-5763, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.181, + 0.473, + 0.245 + ], + "angle": 0, + "content": "[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. In Proceedings of the IEEE international conference on computer vision, pages 2989-2998, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.247, + 0.473, + 0.31 + ], + "angle": 0, + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.313, + 0.473, + 0.375 + ], + "angle": 0, + "content": "[30] Kun Kuang, Ruoxuan Xiong, Peng Cui, Susan Athey, and Bo Li. Stable prediction with model misspecification and ag- nostic distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4485-4492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.379, + 0.473, + 0.43 + ], + "angle": 0, + "content": "[31] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF International conference on computer vision, pages 4654-4662, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.433, + 0.473, + 0.482 + ], + "angle": 0, + "content": "[32] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.485, + 0.473, + 0.549 + ], + "angle": 0, + "content": "[33] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.552, + 0.473, + 0.602 + ], + "angle": 0, + "content": "[34] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11645-11652, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.605, + 0.473, + 0.657 + ], + "angle": 0, + "content": "[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.659, + 0.473, + 0.709 + ], + "angle": 0, + "content": "[36] Mitja Nikolaus, Emmanuelle Salin, Stephane Ayache, Abdellah Fourtassi, and Benoit Favre. Do vision-and-language transformers learn grounded predicate-noun dependencies? arXiv preprint arXiv:2210.12079, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.712, + 0.473, + 0.776 + ], + "angle": 0, + "content": "[37] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.778, + 0.473, + 0.866 + ], + "angle": 0, + "content": "[38] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.857, + 0.165 + ], + "angle": 0, + "content": "[39] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.857, + 0.23 + ], + "angle": 0, + "content": "[40] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In European Conference on Computer Vision, pages 546-562. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.857, + 0.307 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.857, + 0.373 + ], + "angle": 0, + "content": "[42] Emmanuelle Salin, Badreddine Farah, Stéphane Ayache, and Benoit Favre. Are vision-language transformers learning multimodal representations? a probing perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11248-11257, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.857, + 0.438 + ], + "angle": 0, + "content": "[43] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.857, + 0.503 + ], + "angle": 0, + "content": "[44] Avi Schwarzschild, Eitan Borgnia, Arjun Gupta, Furong Huang, Uzi Vishkin, Micah Goldblum, and Tom Goldstein. Can you learn an algorithm? generalizing from easy to hard problems with recurrent networks. Advances in Neural Information Processing Systems, 34:6695-6706, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.857, + 0.568 + ], + "angle": 0, + "content": "[45] Ravi Shekhar, Sandro Pezzelle, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Vision and language integration: Moving beyond objects. In IWCS 2017—12th International Conference on Computational Semantics—Short papers, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.857, + 0.645 + ], + "angle": 0, + "content": "[46] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.857, + 0.696 + ], + "angle": 0, + "content": "[47] Xinwei Shen, Furui Liu, Hanze Dong, Qing Lian, Zhitang Chen, and Tong Zhang. Disentangled generative causal representation learning. arXiv preprint arXiv:2010.02637, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.857, + 0.749 + ], + "angle": 0, + "content": "[48] Zheyan Shen, Peng Cui, Tong Zhang, and Kun Kunag. Stable learning via sample reweighting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5692-5699, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.751, + 0.857, + 0.802 + ], + "angle": 0, + "content": "[49] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8384, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.857, + 0.867 + ], + "angle": 0, + "content": "[50] Damien Teney, Ehsan Abbasnejad, Kushal Kafle, Robik Shrestha, Christopher Kanan, and Anton Van Den Hengel. On the value of out-of-distribution testing: An example of goodhart's law. Advances in Neural Information Processing Systems, 33:407-417, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.857, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2986" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.189 + ], + "angle": 0, + "content": "[51] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1417-1427, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.191, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[52] Christopher Thomas, Yipeng Zhang, and Shih-Fu Chang. Fine-grained visual entailment. arXiv preprint arXiv:2203.15704, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.23, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[53] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.307, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[54] Dušan Variš and Ondřej Bojar. Sequence length is a domain: Length-based overfitting in transformer models. arXiv preprint arXiv:2109.07276, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.345, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[55] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.396, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[56] Jia Wang, Jingcheng Ke, Hong-Han Shuai, Yung-Hui Li, and Wen-Huang Cheng. Referring expression comprehension via enhanced cross-modal graph attention networks. ACM Journal of the ACM (JACM), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.447, + 0.472, + 0.521 + ], + "angle": 0, + "content": "[57] Yun Wang, Juncheng Li, and Florian Metze. A comparison of five multiple instance learning pooling functions for sound event detection with weak labeling. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 31-35. IEEE, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.523, + 0.472, + 0.573 + ], + "angle": 0, + "content": "[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquen Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 1, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.574, + 0.472, + 0.637 + ], + "angle": 0, + "content": "[59] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.638, + 0.472, + 0.688 + ], + "angle": 0, + "content": "[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.689, + 0.472, + 0.739 + ], + "angle": 0, + "content": "[61] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.74, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[62] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L. Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.804, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[63] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.127, + 0.473, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.858, + 0.202 + ], + "angle": 0, + "content": "[64] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.204, + 0.858, + 0.254 + ], + "angle": 0, + "content": "[65] Zelin Zhao, Karan Samel, Binghong Chen, et al. Proto: Program-guided transformer for program-guided tasks. Advances in Neural Information Processing Systems, 34:17021-17036, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.256, + 0.858, + 0.32 + ], + "angle": 0, + "content": "[66] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4252-4261, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.858, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.892, + 0.514, + 0.902 + ], + "angle": 0, + "content": "2987" + } + ] +] \ No newline at end of file diff --git a/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_origin.pdf b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..86341785437c2baea1233a423cd442c37b229222 --- /dev/null +++ b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/37f1017e-5814-46cc-ba5e-1fd9ffaa6a28_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cead15192ecd7fcc1e46a2f34d48d820a4ec3d5c4c612d61ddbf3fc652dc2f01 +size 1675427 diff --git a/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/full.md b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..58f68c61b8dd5d74f61ce534a60610a5d2f98240 --- /dev/null +++ b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/full.md @@ -0,0 +1,359 @@ +# Grounded Image Text Matching with Mismatched Relation Reasoning + +Yu Wu $^{1,*}$ Yana Wei $^{1,*}$ Haozhe Wang $^{1}$ Yongfei Liu $^{1}$ Sibei Yang $^{1,2}$ Xuming He $^{1,2}$ + +$^{1}$ ShanghaiTech University, $^{2}$ Shanghai Engineering Research Center of Intelligent Vision and Imaging + +{wuyul, weiynl, yangsb, hexm}@shanghaiitech.edu.cn jasper.whz@outlook.com liuyongfei314@gmail.com + +# Abstract + +This paper introduces Grounded Image Text Matching with Mismatched Relation (GITM-MR), a novel visual-linguistic joint task that evaluates the relation understanding capabilities of transformer-based pre-trained models. GITM-MR requires a model to first determine if an expression describes an image, then localize referred objects or ground the mismatched parts of the text. We provide a benchmark for evaluating vision-language (VL) models on this task, with a focus on the challenging settings of limited training data and out-of-distribution sentence lengths. Our evaluation demonstrates that pre-trained VL models often lack data efficiency and length generalization ability. To address this, we propose the Relation-sensitive Correspondence Reasoning Network (RCRN), which incorporates relation-aware reasoning via bi-directional message propagation guided by language structure. Our RCRN can be interpreted as a modular program and delivers strong performance in terms of both length generalization and data efficiency. The code and data are available on https://github.com/SHTUPLUS/GITM-MR. + +# 1. Introduction + +Recently, transformer-based vision-language (VL) pretrained models have made significant progress on various VL tasks by fine-tuning on downstream tasks [8, 64, 58, 29, 16]. Despite their success, the representation capacity of such VL pre-trained models remains poorly understood. As a result, an increasing number of studies start to probe the limitations of those learned models from several aspects. The first aspect focuses on the lack of fine-grained understanding of multi-modal data, which is indicated by the weak VL correspondence learned on relations among the entities [36]. In particular, the researchers constructed + +![](images/08e7c43eec4ca58a4f5c3986ca78d6eacee85e98ec4273a39768efa687a4fac4.jpg) +Figure 1. Illustration of the GITM-MR task and the proposed model RCRN. The task is to first identify whether the expression describes the image, and then localize the referred object if the pair are matched, or otherwise ground the mismatched language part. Both data efficiency and length generalization are required. + +![](images/bc45aeb4d866410428d817cf035afcbf356cc3425003ff5f707755d3e2cc8e84.jpg) + +subtle relation phrase differences in text to reveal the poor matching capability between visual relations and linguistic relations [46, 21, 38], or required the model to output fine-grained predictions to thoroughly examine the model's understanding of VL correspondence on different uni-modal component representations [52]. Additionally, while large models generally deliver superior performance, they tend to overfit on small datasets, emphasizing the need to enhance the data-efficiency on fine-tuning [15]. Furthermore, when the text is composed of multiple relations, inferring VL correspondence typically involves combining those semantic relations. This combination requires the model to reason the global semantic correspondence, where the length generalization is shown to be a critical weakness in reasoning capability of pre-trained transformers [3, 54]. However, most of those studies only focus on a single aspect of the VL models and hence produce a partial view on their limitations. + +To better understand the VL pre-trained models, we introduce a novel VL joint task named Grounded Image Text Matching with Mismatched Relation (GITM-MR) which focuses on distinguishing nuance in relations within text and image, and requires a model to show explicit comprehension on relations in both modalities. As shown in Fig. 1, the task first requires the model to identify whether an expression describes a given image, and then grounds the referred + +object if the image-text pair are matched or localizes the mismatched relation in the expression otherwise. The two fine-grained sub-tasks verify the specific understanding on relations both in match and mismatch scenarios. Moreover, the proposed task also allows us to evaluate the data efficiency of model's fine-tuning and its generalization ability. In particular, we design a learning setting with limited annotations, where only the data in a limited size and domain have annotations (i.e., a mix of in- and out-of-distribution (OOD) test data). Given the task design, we then build a new benchmark based on the Ref-Reasoning [61] dataset, which is more general and challenging than previous probing benchmarks on relations and with low linguistic bias. + +Leveraging our proposed benchmark, we evaluate several representative state-of-the-art models on the task, assessing their performance under limited data sizes and varying domains of input length. Our results reveal their limitations in the both settings. To alleviate these drawbacks, we propose a unified modular framework that jointly solves subtasks in the GITM-MR problem. By explicitly capturing relation correspondence and utilizing it to propagate messages, our approach is capable of obtaining a comprehensive understanding of global correspondence. Additionally, its lightweight and compositional module design allows for excellent data efficiency and length generalization, accompanied by a more transparent reasoning process. + +Specifically, we develop a graph neural network on a parsed language scene graph (LSG), named Relation-sensitive Correspondence Reasoning Network (RCRN), which exploits the language structure to perform relation-aware reasoning in a belief space of visual-linguistic alignment. Our RCRN uses a set of primitive reasoning modules to operate on the pre-trained vision-language features under the guidance of LSG, and predict the fine-level grounding in both match and mismatch scenarios. As demonstrated by the experiments, our method outperforms in both data efficiency and length generalization settings, demonstrating its effectiveness in learning relation correspondences with data efficiency and OOD generalization capability. + +The contribution of our work is summarized as follows: + +- We propose a new VL task GITM-MR that challenges relation correspondence learning in VL pre-trained model, and focusing on both data efficiency and out-of-distribution text length setting, with a new benchmark for GITM-MR. +- We identify the shortcomings of current VL models through the benchmark, and develop a modular graph network RCRN to address these issues. +- Our method achieves superior results on both settings, validating its relation learning efficiency and generalization ability, which hopefully inspires the following exploration of these challenges. + +# 2. Related Works + +Probing Tasks for VL Models Despite the rapid development on VL pre-trained models [8, 41, 19, 64, 58, 29], emergent works reveal the limitations and gain some deep understanding on pre-trained representations, including implausible scenes understanding [10], language and vision priors exploration [20], and modality ablation [18]. A considerable number of them have revealed the weakness of VL models in relation understanding. Early work proposed a benchmark FOIL [46, 45] by substituting a word in caption to test non-pretrained VL models on identifying minor differences. Subsequent works [21, 38, 36, 53, 63] further evaluate pre-trained models on detecting subtle differences in relations or predicate-noun dependencies in zero-shot scenarios, and reach a consensus on the poor quality of pre-trained VL representation for relations. FGVE [52] extends this evaluation to the task of fine-grained entailment with model fine-tuning, but lacks evaluation in vision modality. Recently, a new benchmark [42] is proposed to simultaneously assess the concept understanding of representations learned from both modalities. In contrast, our task focuses on generating fine-grained multi-modal output under different fine-tuning settings in order to systematically investigate the relation understanding problem in VL pre-training. + +OOD Generalization The general problem of OOD generalization has attracted much attention recently [33, 6, 47, 30, 48]. Among them, several studies have considered the tasks with generalization to processing or generating longer sequences. Some works [3, 54] have studied length generalization in different uni-modal seq-to-seq tasks, and find that fine-tuning pre-trained transformer models can lead to poor out-of-distribution (OOD) performance. To address this issue, various solutions have been proposed, including modifications to attention mechanism [39, 11, 17] and the design of recurrent networks [44, 4], but their application in VL tasks is nontrivial. Other recent works have also revealed the challenge of OOD generalization for the VL tasks, but most of them focuses on the answer bias of Visual Question Answering (VQA) [50, 1, 2]. The attempted solutions include the causal learning [37] and stable learning [51] methods. Perhaps most related to us is the modular network approaches [28, 49, 65], which use program-based design to achieve good generalization for certain OOD scenario such as question domains in VQA. We follow the modular design but focus on the OOD problem induced by the expression complexity in the grounded image text matching task. + +Vision-language Matching Image-Text Matching (ITM) and Referring Expression Grounding (REG) are common cross-modal matching tasks in vision-language domain. The problem of ITM aims to predict a global similarity between a text and an image. Classical methods for ITM fol + +![](images/84876a2a1d96c35f5193bef9971e7406e9b2a81684c02549173b9a97e2ad43ad.jpg) +Figure 2. The length distribution of sentences containing different numbers of entities in the validation set. Different colors represent sentences with different number of entities. + +low the core idea of learning a common space where linguistic and visual features are compatible [27, 31, 7, 14, 9]. Recent approach [41] concentrates on learning the relation-aware global similarity by a regularization strategy on training. Our model derives a global similarity measure by propagating messages on local correspondences, capable of capturing subtle changes in relations. + +The goal of REG is to locate an object in an image based on a language expression. Most methods focus on learning a holistic visual and linguistic representation in a joint space for aligning the object and text, based on CNN+RNN architectures [35, 66, 25] or the Transformer [12, 13, 32, 40, 59]. Another line of research aims to exploit the linguistic structure in the text for better context modeling [62, 34, 60, 56]. In particular, SGMN [61] parses the text into language components and performs reasoning under the guidance of the linguistic structure. Our method introduces a flexible contextual representation and a bi-directional propagation to extend correspondence reasoning beyond the task of REG. + +# 3. Problem Setup and Benchmark + +We aim to study the representation capacity of the VL pretrained models for fine-grained downstream tasks. To this end, we introduce the Grounded Image Text Matching with Mismatched Relation (GITM-MR) task, which is a joint matching and grounding task focusing on relations in vision and language domain. In this section, we present the problem setting of GITM-MR, followed by building a new benchmark for the task. + +GITM-MR Task Given an image and a referring expression, our goal is to determine if the expression describes an object in the image or not (termed as Image-Text Matching (ITM)), and then to localize the referred object if it is matched (termed as Referring Expression Grounding (REG)), or otherwise ground the mismatched relation phrase (termed as Mismatched Relation Reasoning (MRR)). Notably, the only difference between match and mismatch is a relation phrase in the expression, making our task particularly challenging. + +Formally, given an image $I$ and a text description $L$ , we denote their matching state as $y$ , indicating whether the image matches the description $(y = 1)$ or not $(y = 0)$ . For the + +matched case, the referred object has a corresponding location in the image, represented by its bounding box $z \in \mathbb{R}^4$ . For the mismatch case, we denote the set of all possible relation phrases (i.e. predicates) for mismatch reasoning as $\mathcal{R} = \{r_i\}_{i=1}^{N_e}$ , and there is a target mismatched relation $r_i^*$ in the relation phrase set $\mathcal{R}$ . Given the input $I$ and $L$ , our goal is to first predict $y$ for the input pair, and then predict $z$ for the referred object in $L$ when $y = 1$ , or otherwise predict $r_i^*$ from $\mathcal{R}$ to explain why $y = 0$ . + +For model training, we follow the convention of ITM and REG tasks, and assume the annotations for those two subtasks are provided. However, it is usually costly to collect annotations for the subtle MRR task. Consequently, we assume a weakly-supervised learning setting where only the matching state label $y$ is given for mismatched data. + +Our Benchmark To conduct systematical study, we construct a benchmark for the GITM-MR task, including a new dataset and two challenging evaluation settings. We first build a dataset of image-text pairs from the public REC dataset, Ref-Reasoning [61], by substituting the relation phrases to generate mismatched referring expressions for images. From the provided relation phrases in the LSG offered by the dataset, we manually select a subset of 27 commonly-occurred relations, and assign some contextual close but semantically different ones for each relation in the subset as their mismatched candidates. We control the linguistic bias by several measures, including keeping the relation distribution and checking by language-only model, the details of which are stated in the Suppl. + +However, replacing the relation phrases may introduce falsely mismatched cases. To ensure the quality of our data, we filter out those false mismatch cases in the test set with the help of annotators. After filtering, the dataset contains more than 1M referring expressions in 60K images. It has 1M, 50K and 10K expression-referent pairs for training, validation, testing, respectively. We only select the original image text pairs which have generated mismatched pairs to keep the matching label balanced. Fig. 2 shows the length distribution of sentences in the validation set of the benchmark. We find it is more general and challenging than previous benchmarks on relation understanding, and refer the reader to see more details of the dataset in Suppl. + +Evaluation Settings Given the dataset, we develop two evaluation settings to test pretrained VL models on their relation understanding. The first setting is designed to investigate data efficiency during the fine-tuning process. We train the models on small-scale datasets and tested on indistribution test sets. The second setting is designed to evaluate the generalization performance of models with regards to different text input lengths. Here the test set contains + +![](images/cbab5a3af3c37e1c2667eb5303a249a44386c3fc3579c02d583c00d5f32fb365.jpg) +Figure 3. Model overview with a mismatched case. Given an image and an expression, we first generate visual and linguistic candidates by a detector and a language parser, and compute their representations. Then we use the Context-sensitive Propagation Network to inter alignments between visual-linguistic candidates, which conducts bi-directional message propagation based on the language graph. The propagation initializes the messages by computing local beliefs, selectively aggregates the context information and updates the belief with a context-sensitive gating function. Predictions for this case are obtained by exploiting the beliefs from the propagation. In the language graph, the ellipses represent entity phrases and the rectangles stand for relation phrases. Further elaboration can be found in the main text. + +longer and more complex samples on relation combinations compared to the training set. + +Specifically, according to the sentence length distribution in Fig. 2, we construct two training sets, named TrainLen16 and Train-Len11, which includes sentences containing less than 16 and 11 words respectively. In the test phase, the model needs to perform the reasoning task on sentences with lengths less than 11 and 16 (in-distribution), as well as sentences longer than these lengths (out-of-distribution). As two training sets accounts for $5\%$ and $10\%$ of the overall training data, those two settings allow us to evaluate the model's data efficiency in fine-tuning and OOD generalization ability simultaneously. Detailed data partition rules are explained in Suppl. + +# 4. Our Approach + +In this section, we present our model design for the GITM-MR task and the learning objectives for training with weakly-annotated data. To tackle the problem, we develop a unified and generalizable framework as outlined in Fig. 1, which consists of two main components: a representation network that computes an initial representation of the linguistic and visual components, and a context-aware correspondence reasoning network for final match predictions. Below we first introduce the two main model components in Sec. 4.1 and 4.2, and then the model learning in Sec. 4.3. + +# 4.1. Candidate Generation and Representation + +# 4.1.1 Candidate Generation + +Our method first generates a set of visual objects from the image and phrase candidates from the corresponding text. + +Specifically, for the visual object candidates, we adopt a pre-trained object detector [64] to generate $N_{o}$ box proposals $\mathcal{O} = \{o_i\}_{i=1}^{N_o}$ from the image $I$ . The detector also provides an initial visual feature (i.e., pooled ROI feature) for each object proposal. For the language phrase candidates, we utilize an off-the-shelf parser [43] to produce a language scene graph $\mathcal{G} = (\mathcal{E}, \mathcal{R})$ of the text expression $L$ , which consists of a set of entity phrases $\mathcal{E} = \{e_j\}_{i=1}^{N_e}$ as its nodes and relation phrases as its edges $\mathcal{R} = \{r_{ij}\}$ , where $r_{ij}$ connects to subject $e_i$ and object $e_j$ . We denote the number of relation phrases as $|\mathcal{R}| = N_r$ . + +# 4.1.2 Candidate Representation + +Given the visual and linguistic candidates, we compute their representations in two steps, as detailed below. + +Token-level Representation The first step generates an initial representation for the word tokens and visual objects based on a pre-trained vision-language model. Here we adopt a shallow version of the UNITER model [8], aiming to compute a feature representation for the input tokens in two modalities with a generic vision-language alignment from the pre-trained model. + +Candidate Features In the second step, we then compile the token-level representations into a set of features for + +our visual and linguistic components used in the subsequent reasoning network. Here we adopt a similar strategy as in SGMN [61] for computing the features. + +Concretely, for visual objects $o_i$ , we augment its token-level feature $\mathbf{o}_i$ by its spatial location $\mathbf{l}_i^o$ (i.e., bounding box parameters). In addition, for every object pair $(o_i,o_j)$ in the image, we represent its relational feature as $\mathbf{r}_{ij}^{o}$ , which is computed from the features of its two objects. For the linguistic entities in $\mathcal{E}$ and relation phrases in $\mathcal{R}$ , we feed the corresponding word tokens into a Bi-LSTM [22] with self-attention [24] to generate two sets of features. In the first set, each entities $e_i$ is encoded by a tuple of three features $(\mathbf{e}_i,\mathbf{l}_i^e,\mathbf{h}_i^e)$ , where $\mathbf{e}_i$ is an embedding into the space of object feature $\mathbf{o}_i$ , $\mathbf{l}_i^e$ is an embedding into the space of object location $\mathbf{l}_i^o$ , and $\mathbf{h}_i^e$ is a linguistic feature from the LSTM states. For the second feature set, we generate a representation $\mathbf{r}_{ij}^{e}$ for each relation phrase $r_{ij}\in \mathcal{R}$ . We refer the reader to Suppl. for the details of our feature design. + +# 4.2. Context-sensitive Propagation Network + +We now introduce our second model component, the Context-sensitive Propagation Network (CPN), which aims to infer the alignment between the visual-linguistic candidates and predict the match label of input image-text pair. To this end, we build a modular graph neural network on the language graph $\mathcal{G}$ , which takes the candidate features as input and performs context-aware reasoning in a belief space of visual-linguistic alignment. Such a structured reasoning process enables us to exploit the regularity in language, i.e. the connections of entities by relations, leading to a relation-aware and compositional alignment model with potentially better generalization. + +Specifically, as illustrated in the language graph in Fig. 3, the $i$ -th node of the graph network corresponds to an entity phrase $e_i$ and the edge between the node $i$ and $j$ corresponds to a relation phrase $r_{ij}$ . We follow the common assumption that the corresponding $\mathcal{G}$ has a tree structure [61, 34] and the referred entity is regarded as the root node. To perform reasoning, the graph network uses a bi-directional message propagation procedure to infer a relation-aware correspondence belief (w.r.t. the visual candidates) for all the entity nodes (described in Sec. 4.2.1), followed by a readout head module to predict the fine-grained and global match scores (described in Sec. 4.2.2). Finally, we formulate each key operation in the graph reasoning as an easy-to-interpret primitive module, and convert our reasoning process into an explainable program in Sec. 4.2.3. + +# 4.2.1 Bi-directional Message Propagation + +Given the graph network, we aim to compute a message for each node, representing a belief on the correspondence between the associated entity phrase and all the visual objects + +in the image. In order to resolve the ambiguity in the local matching of entities, we introduce a bi-directional message propagation procedure to aggregate all the local beliefs on phrase-visual region correspondence throughout the graph. + +Concretely, inspired by belief propagation [5], we perform two parallel passes of message propagation on the tree graph $\mathcal{G}$ along two directions: 1) bottom-up direction (denoted as bp) first computes messages from the leaf nodes and then updates their parents recursively until the root is reached; 2) top-down direction (denoted as td) starts from the root node and updates the children nodes recursively until reaching all the leaves. For each direction, the message propagation aggregates the context information on cross-modal matching and update the belief on phrase alignment at each node in the following three steps: + +(1) Message initialization: We initialize a local belief $\mathbf{b}_i^{\mathrm{loc}}$ at node $i$ based on a similarity measure between the entity $e_i$ and all the visual objects. Here we define the similarity by a weighted combination of feature similarities in appearance and spatial location space. Specifically, for each pair of entity $e_i$ and object $o_j$ , we compute two similarities as follows, + +$$ +b _ {i j} ^ {\mathrm {a p p}} = \mathrm {F} _ {\mathrm {s i m}} (\mathbf {e} _ {i}, \mathbf {o} _ {j}), \quad b _ {i j} ^ {\mathrm {p o s}} = \mathrm {F} _ {\mathrm {s i m}} \left(\mathrm {I} _ {i} ^ {e}, \mathrm {I} _ {j} ^ {o}\right) \tag {1} +$$ + +where $\mathrm{F}_{\mathrm{sim}}$ is an MLP network computing the similarity between features from two modalities. The local belief $\mathbf{b}_i^{\mathrm{loc}}$ then integrates two similarities as follows: + +$$ +\mathbf {b} _ {i} ^ {\text {l o c}} = \mathrm {F} _ {\text {n o r m}} \left(\beta_ {i} ^ {\text {a p p}} \mathbf {b} _ {i} ^ {\text {a p p}} + \beta_ {i} ^ {\text {p o s}} \mathbf {b} _ {i} ^ {\text {p o s}}\right) \tag {2} +$$ + +where $\mathbf{b}_i^t = [b_{ij}^t]_{j=1}^{N_o}$ , $t \in \{\mathrm{app, pos}\}$ . Here $\beta_i^t$ is the weight for combining similarities and is computed as $\beta_i^t = \mathrm{sigmoid}(\mathbf{W}_t^\top[\mathbf{h}_i^e, \mathbf{b}_i^{\mathrm{app}}, \mathbf{b}_i^{\mathrm{pos}}])$ , where $\mathrm{F}_{\mathrm{norm}}$ is a normalization function to rescale its input elements to $[0,1]$ . + +(2) Message aggregation: According to the direction $\mathbf{td}$ or $\mathbf{bp}$ , message aggregation collects beliefs from the upstream neighbors in a context-sensitive manner. Formally, we first prune the local beliefs by selecting top $K$ maximum values, which significantly reduces the computation complexity for large $N_{o}$ and is written as, + +$$ +\mathbf {b} _ {j} ^ {\text {s e l}} = \mathrm {F} _ {\text {t o p k}} \left(\mathbf {b} _ {j} ^ {\text {l o c}}\right) \tag {3} +$$ + +In contrast to [61], our selection is belief-based and more informative for propagation. Given the sparsified beliefs, we then aggregate the messages from the neighboring nodes by taking into account the alignment of the relation phrases on the edges as follows, + +$$ +\mathbf {b} _ {i} ^ {\mathrm {a g g}} = \prod_ {j \in \operatorname {C t x} (i)} \mathbf {A} _ {i j} \mathbf {b} _ {j} ^ {\mathrm {s e l}}, [ \mathbf {A} _ {i j} ] _ {k l} = \mathrm {F} _ {\text {s i m}} \left(\mathbf {r} _ {i j} ^ {e}, \mathbf {r} _ {k l} ^ {o}\right) \tag {4} +$$ + +where $\mathrm{Ctx}(i)$ denotes the upstream neighbor set of node $v_{i}$ according to the propagation direction. We model the messages provided by relations explicitly through alignment on + +edges, which forces the model to capture the nuances of different relations. For the aggregation operator, we take the element-wise product of beliefs, which can be viewed as a soft version of logical AND. Unlike the min pooling used in [49, 23], our design leads to a more robust and smooth learning procedure. + +(3) Message update: Finally, we update the belief $\mathbf{b}_i$ at node $i$ by a context-sensitive integration of the aggregated neighboring belief and the local belief. To achieve this, we design a gated product as follows: + +$$ +\beta_ {i} ^ {\mathrm {r e l}} = \operatorname {s i g m o i d} \left(\mathbf {W} _ {\mathrm {r e l}} ^ {\top} \left[ \mathbf {h} _ {i} ^ {e}, \mathbf {b} _ {i} ^ {\text {a p p}}, \mathbf {b} _ {i} ^ {\text {p o s}} \right]\right) \tag {5} +$$ + +$$ +\mathbf {b} _ {i} = \mathrm {F} _ {\text {n o r m}} \left(\mathbf {b} _ {i} ^ {\text {l o c}} \circ \left(\mathbf {b} _ {i} ^ {\text {a g g}}\right) ^ {\beta_ {i} ^ {\text {r e l}}}\right) \tag {6} +$$ + +where $\circ$ denotes the element-wise product. Instead of using the language features as in [61, 55], our gating function also takes the local beliefs as input, making the network more sensitive to the image context. + +Given the updated beliefs, we compute a confidence score for each node matching to any object in the image by applying an MLP network, denoted as $p_i = \mathrm{sigmoid}(\mathrm{MLP}_{\mathrm{match}}(\mathbf{b}_i))$ . In the end, the message propagation along two directions generates two belief vectors $(\mathbf{b}_i^{\mathrm{bp}}$ and $\mathbf{b}_i^{\mathrm{td}})$ and two confidence scores $(p_i^{\mathrm{bp}}$ and $p_i^{\mathrm{td}})$ for each node $i$ . + +# 4.2.2 Readout Head + +After the graph-based reasoning, we now introduce a readout head network to generate the predictions for three subtasks, i.e., ITM, REG and MRR, by exploiting the beliefs and confidence scores from the bi-directional propagation. + +Image-Text Matching To predict the global matching label $y$ , we take a minimum pooling of the matching confidence scores across the entire graph as below, + +$$ +P (y = 1 | I, L) = \min _ {i} \left(\min \left(p _ {i} ^ {\mathrm {b p}}, p _ {i} ^ {\mathrm {t d}}\right)\right) \tag {7} +$$ + +REG Task For the referent grounding, we take the belief vector of the root node $\mathbf{b}_{\mathrm{root}}^{\mathrm{bp}}$ from the bottom-up direction and locate the matching visual object $o_{i^{*}}$ by choosing the largest belief value: $i^{*} = \arg \max_{i}\mathbf{b}_{\mathrm{root}}^{\mathrm{bp}}$ . A standard regression head is also employed for more precise localization and its details are left to the Suppl. + +MRR Task For mismatched relation prediction, we exploits the inconsistency of the match confidence scores along the graph edges. We found the mismatched relation typically leads to a confidence gap between its two incident nodes. As such, we choose the edge with the largest match confidence difference between its two nodes: + +$$ +\hat {r} = \arg \min _ {r _ {i j} \in \mathcal {R}} \left(\left\{\left(p _ {i} ^ {\mathrm {b p}} - p _ {j} ^ {\mathrm {b p}}\right) + \left(p _ {j} ^ {\mathrm {t d}} - p _ {i} ^ {\mathrm {t d}}\right) \right\}\right), \tag {8} +$$ + +where we sum the differences from two propagation directions to make the decision more robust. + +Table 1. The full list of the reasoning modules and their implementations. We use bold lowercase variables to represent vectors and bold uppercase variables to represent matrices. Regular variables are scalars. W is trainable parameter matrix. + +
ModulesInOutImplementation
Local Correspondences:
Simv'txt, v'visbFsim(v'txt, v'vis)
GateSum{vi}, {bi}bFnorm(∑i sigmoid(W^T[vi, bi])bi)
Correspondence Reasoning:
Selectbb'Ftopk(b)
Aggregate{Ai}, {bi}b'Πi(AiBi)
GateProdb, b', βib''Fnorm(b o (b')βi)
ClassifybpMLPmatch(b)
Readout Head:
Locatepiarg min(p)
Comparepi, pjdpi - pj
And{pi}pminpi
+ +# 4.2.3 Modular Network Interpretation + +We now introduce an interpretation of our using the modular network framework. As shown in Tab. 1, we summarize the key operations in the RCRN into a set of primitive modules with semantic meanings. This allows us to build a explainable reasoning program: (1) In the message initialization step, local beliefs are computed using Sim, and are combined by GateSum to generate the initial belief on each node. (2) For message aggregation, we first use Select to prune each node and then apply Aggregate to collect the messages from all neighbors, considering the alignment of relations particularly. (3) For message update, GateProd is used to combine the aggregated message with the local belief to generate the updated belief on each node. The Classify module provides a confidence score indicating the likelihood that the phrase entity matches an visual object. (4) In the readout head, we use two Compare modules in two directions respectively for MRR and one And operation for the ITM task, and finally generate predictions for all three subtasks via Locate module. For instance, the grounding prediction is achieved by $\text{Locate}(-\mathbf{b}_{\text{ref}})$ . Several example are provided in Sec. 5.4 and Suppl. to illustrate the interpretable reasoning process, along with further detailed implementations of the modules. + +We note that such a program formulation indicates that the model can enjoy a good out-of-domain generalization due to the compositionality of language graphs. Even facing with more complex text structures, we can perform this reasoning process by reusing those modules on new graphs. + +# 4.3. Model Learning + +We train the network on a multitask loss with end-to-end back propagation. Cross-Entropy (CE) loss and Binary Cross-Entropy (BCE) loss are used for the REG and ITM + +tasks, respectively: + +$$ +\mathcal {L} _ {\mathrm {g r d}} = \mathrm {C E} \left(\mathbf {p} _ {\mathrm {g r d}}, \mathbf {p} _ {\mathrm {g r d}} ^ {*}\right) \tag {9} +$$ + +$$ +\mathcal {L} _ {\text {m a t c h}} = \mathrm {B C E} \left(p _ {\text {m a t c h}}, y ^ {*}\right) \tag {10} +$$ + +where $\mathbf{p}_{\mathrm{grad}} = \mathrm{softmax}(\mathbf{b}_{\mathrm{root}}^{\mathrm{bp}})$ stands for the predicted grounding distribution, $\mathbf{p}_{\mathrm{grad}}^{*}$ denotes the ground-truth one-hot grounding label, and $p_{\mathrm{match}} = P(y = 1|I,L)$ denotes the predicted match probability. When the object boxes are from a detector, the box with the maximum IoU with the ground-truth box of the referent is labeled as the localization output. The overall loss is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {g r d}} + \mu \mathcal {L} _ {\mathrm {m a t c h}} + \omega \mathcal {L} _ {\mathrm {r e g}} \tag {11} +$$ + +where $\mu$ and $\omega$ are hyper-parameters to balance the three terms. $\omega$ would be zero if the IoU of the ground-truth box and the selected proposal is less than 0.5. The regression loss is standard and can be found in Suppl. + +# 5. Experiments + +In this section, we conducted experiments on the proposed benchmark to test several baselines of VL pre-trained models on inferring relation correspondences, with a specific focus on their data efficiency and length generalization capabilities. We then compare our model with those strong baselines to show its effectiveness in these settings. Finally we illustrate the visualizations and interpretations of the reasoning process of our model. + +# 5.1. Experiment Setup + +Metrics The evaluation metrics include classification accuracy for three subtasks. The grounding result for a matched case is considered as correct when it is identified as matching and the predicted box has at least 0.5 IoU with its ground-truth location. The grounding accuracy (i.e. Recall@1) is the ratio of correctly grounded cases. For mismatch reasoning, a mismatched case needs to be correctly classified and the mismatched relation should be accurately selected from the candidate set. The MRR accuracy is the top-1 accuracy among the candidates. + +VL Pre-trained Models We benchmark five VL models on the proposed dataset. We first choose TCL [58], UNITER [8] and FIBER [16], where TCL is the state-of-the-art pre-trained model for image-text retrieval, UNITER is a vanilla transformer-based VL pre-trained model for matching and grounding tasks, and FIBER uses a specific fine-grained pre-training design. Additionally, we adapt the latest fine-grained visual entailment (FGVE) model [52] to our setting. We compare its potential fine-grained multitask capability with ours, which is designed based on the VL pre-trained model Oscar+[64]. Because none of the + +pretrained VL models can fulfill three subtasks jointly, we also apply the multiple instance learning strategy [57] on UNITER to cope with the weakly-supervised MRR task. As shown in Suppl, we modify those models by adding subtask heads to perform three subtasks. We note that TCL and UNITER use the same pre-training datasets and the pretraining corpus of FIBER and FGVE is larger while all the models are fine-tuned on the GITM-MR, which leads to a relatively fair comparison. + +Implementation Details For generating object candidates, we take the off-the-shelf detector, VinVL [64], to extract at most 100 proposals with object score great than 0.1 for each image. For model training, the reasoning component of RCRN is trained by the Adam optimizer with the learning rate set to $5\mathrm{e} - 4$ for the proposed dataset. The learning rate of candidate representation module stated in Sec. 4.1 is set to $5\mathrm{e} - 7$ . We train 80K iterations with batchsize 64 totally. The hyper-parameter $K$ , $\mu$ and $\omega$ are set as 5, 3 and 1 based on validation, respectively. The representation module uses the first 6 transformer encoder layers in the pre-trained UNITER. The ablation study on layers and more implementation details are included in the Suppl. + +# 5.2. Results and Comparisons + +Relation Understanding with Limited Data Tab. 2 (In-Distribution part) shows the performances of pre-trained models on the relation understanding under the condition of limited data. The results indicate that TCL struggles to achieve a match accuracy of more than chance. This could be attributed to its emphasis on global image-text matching during pre-training, which may hinder its ability to perform fine-grained understanding tasks. Though FGVE is designed for fine-grained entailment prediction, it still lacks capability on mismatch relation reasoning in our scenario. This is likely due to its reliance on the pre-trained VL model for the alignment between vision and language relations. UNITER and FIBER perform relatively better on the match task than TCL because they both use the World-Region Alignment (WRA) as a pre-training task, and FIBER is also pre-trained with phrase grounding, which helps to establish fine-grained correspondence between image and text. The overall conclusion drawn from these findings is that pre-trained models exhibit unsatisfying performance on the challenging benchmark under the limited data condition, indicating the need for further development in establishing effective visual-linguistic correspondence for relation understanding. + +In comparison, our RCRN outperforms the SOTA pretrained models on match and grounding tasks in the indistribution scenarios, even with fewer parameters. The model's performance with limited training data indicates the effectiveness of its lightweight and structured design. The design incorporates appropriate inductive biases based + +Table 2. Experiment results of testing models' relation understanding with limited data and ability of length generalization. UNITER+MIL represents the UNITER with multiple instance learning strategy applied. FGVE+MAX represents the fine-grained visual entailment model with max pooling on knowledge elements to predict mismatched relation. - means the model can't handle the corresponding task. + +
Training SetMethod#Pretrain Images#ParamFull TestIn-DistributionOut-of-Distribution
Match%Grounding%MRR%Match%Grounding%MRR%Match%Grounding%MRR%
Train-Len16TCL [58]4M210M52.77--53.11--52.70--
UNITER [8]4M111M61.9324.16-71.2940.63-59.9420.67-
UNITER+MIL [57]4M111M59.1318.4647.568.8341.5361.3857.0713.5644.53
FIBER [16]4.8M254M52.319.01-59.0437.15-50.883.03-
FGVE+MAX [52]5.7M113M54.30-33.2358.37-51.7953.43-29.26
RCRN (Ours)4M90M66.5929.1055.7271.8546.2459.0465.4725.4655.01
Train-Len11TCL [58]4M210M51.71--52.08--51.70--
UNITER [8]4M111M55.5110.24-68.2230.05-54.979.42-
UNITER+MIL [57]4M111M55.4311.4442.7365.0429.5661.6555.0210.6941.93
FIBER [16]4.8M254M53.5211.74-57.2129.56-53.3611.00-
FGVE+MAX [52]5.7M113M53.40-38.0555.26-40.7853.31-37.93
RCRN (Ours)4M90M63.4121.4552.6768.7037.9357.7763.1920.7752.46
+ +on sentence structure to establish both relation and entity correspondence in a visual-linguistic context, and learns the reasoning strategy effectively. + +Length Generalization Results about length generalization ability are shown in the Out-of-Distribution part of Tab. 2. Overall, the OOD performance of pre-trained models on both test settings was generally poor compared to their in-distribution performance, especially on the grounding and MRR subtasks, which explicitly reflect the models' understanding of relations. In particular, in the OOD setting, FIBER's match accuracy is similar to TCL, with both models only slightly exceeding $50\%$ . The OOD results suggest that these models tend to overfit to the training data biases, without truly establishing the universal VL correspondence of the relations, thus preventing them from generalizing to longer sentences. + +In contrast, RCRN achieved notably better performance than the other models on all three subtasks in both OOD scenarios. Specifically, RCRN trained with Train-Len11 outperformed the best performance of other models by $8.17\%$ , $10.08\%$ , and $10.53\%$ on all three tasks. This success can be attributed to the modular design, which helps the model learn to establish local visual-linguistic correspondence and the universal reasoning strategy. During test, the model can exploit the regularity in sentence structures to compose the modules and deal with longer sentences. + +Discussion These experiments not only help us investigate the two challenges that the pre-trained models faced, but also lead to other interesting findings. First, our observation that fine-grained pre-training tasks, such as WRA, can help models learn local correspondences on downstream tasks, is consistent with the findings of [36]. Second, our findings in the OOD scenario are consistent with previous works [3, 54], which suggest that pre-trained models are prone to overfitting to the training data length biases. + +Table 3. Ablation study on the benchmark's OOD validation set. + +
LangβCtxβBi-MPVLPMatGrdMRR
---√(w/o MP)54.7913.9941.47
--✓✓59.6523.8045.58
✓-✓✓60.8824.1849.57
✓✓-✓60.0723.8848.62
✓✓✓-60.7029.2536.65
✓✓✓✓61.4723.3753.23
+ +Moreover, to independently investigate the capability on two subtasks, namely grounding and MRR, we also evaluate on an oracle setting, where we suppose that the models always predict correctly on the matching task. The experiment demonstrates our RCRN has strong performance on both subtasks, and the details of this experiment are provided in the Suppl. + +Overall, our experiments provide insights into the challenges and potential solutions for the relation understanding in VL models. The fine-grained pre-training tasks can help models learn local correspondences, while the modular design can facilitate the composition of modules to deal with longer sentences. We believe that our findings offer a promising direction to effectively address these challenges. + +# 5.3. Ablation study + +In this section, we present the effectiveness of each component in the proposed RCRN through ablative experiments trained on the Train-Len16 set, with the results in Tab. 3. + +We conduct ablation study by removing some specific designs. $\mathbf{Lang}\beta$ stands for using the language features as one input for gate functions in Eq. 2 and Eq. 6. Removing it indicates constant $\beta$ 's in the propagation. $\mathbf{Ctx}\beta$ means feeding initial similarities into the gate functions. Bi-MP represents propagating twice individually using different $\beta^{\mathrm{rel}}$ 's for matching task reasoning. Without this part, only the propagation from the root to leaves is conducted. VLP + +![](images/8cb73cd665b761423235616b8428a77eaf47f7e582b22b34e696a407186fb651.jpg) +(a) There is a guy wearing the black shirt and to the right of the boy that holding a cell phone. + +![](images/8e689158bbf2041f9a120a5acaea4aa9336b1d2692096169a6ea00abcab481a1.jpg) +(b) There is a guy wearing the black shirt and to the left of the boy that holding a cell phone. +Figure 4. The visualization of the reasoning process of RCRN. The entity and relation correspondence maps are shown by color maps on boxes and box pairs respectively, where only those with top 5 correspondence scores for each phrase are included. Small color maps in (a) are the local correspondence maps and the large ones are the propagated contextualized maps. The boxes of parents in box pairs are dotted on their centers. Dashed boxes indicate the pooling results. + +stands for introducing vision-language pre-trained transformer layers to encode features. Particularly, the first line with VLP only stands for the model which consists of 6 transformer layers from UNITER and the UNITER task heads stated in Sec. 5.1 and Sec. 5.2. + +The results show that all of these designs improve the matching performance. VLP model (first line in Tab. 3) with the same setting as the VLP layers in RCRN performs poorly on OOD scenario, which demonstrates our footnote statement in Sec. 4.1. Nevertheless, VLP component of RCRN is significant on MRR performance, probably because it provides accurate preliminary vision-language alignments on relations. + +# 5.4. Visualization and Interpretability + +As shown in Fig. 4, we visualize a pair of matched and mismatched samples of our RCRN. Fig. 4(a) shows the bottom-up propagation process followed by grounding prediction on a matched image-text pair, and Fig. 4(b) presents the bi-directional propagation results with readout process for MRR on a mismatched pair. In Fig. 4(a), the propagation eliminates the ambiguities on entities under the guid + +Table 4. Intermediate diagnosis results of RCRN. E.G.R. means Entity Grounding Recall. R.G.R. means Referent Grounding Recall. R.C.S. means Relation Correspondence Scores. + +
Inference MethodE.G.R.R.G.R.R.C.S.
R@1R@3R@5R@1R@3R@5GTMeanMax
By chance6.3718.2528.785.7416.9427.22---
w/o MP50.5080.5989.7938.8875.8488.110.610.170.69
w/ MP60.8285.5892.1669.4690.9295.590.610.170.69
+ +ance of explicit relation correspondences. Fig. 4(b) illustrates that the intermediate reasons of the predictions can be traced. The sum of match confidence differences along the mismatched relation, denoted as $d$ in the dotted box, are notably lower than other $d$ 's. + +Tab. 4 demonstrates some interpretable intermediate diagnosis results of RCRN on a subset of the validation set, with ground-truth box proposals. We analyze from the aspect of phrase-level visual grounding recall and statistics on relation correspondence scores. To obtain the true correspondence between the language scene graph and the bounding boxes, we search the parsed language scene graph in the original image scene graph in GQA dataset [26] with a noisy graph matching algorithm. See Suppl. for the detail of the algorithm. + +The results verify that reasonable local correspondences are built, since the mean of initial recall (i.e. inference without MP) on entity phrase grounding is much higher than the random guess. As for relation correspondences, the average values on the ground-truth vision-language relation correspondences are far above the average mean value on all the pairs and very close to the average maximum. Moreover, message propagation notably helps refine the correspondence maps, especially on the referent entities. + +# 6. Conclusion + +In this paper, we have introduced a novel VL joint task, Grounded Image Text Matching with Mismatched Relation (GITM-MR), which provides a challenging benchmark for evaluating pre-trained models on the relation reasoning. Our experiments have shown that some state-of-the-art pretrained models struggle with the task under the setting of limited data and out-of-distribution sentence lengths. To address this problem, we develop a Relation-sensitive Correspondence Reasoning Network (RCRN) to compute contextualized cross-modal alignment of both entities and relations, and ground the fine-grained result in the image or expression. The proposed method outperforms prior SOTA pre-trained models, which demonstrates its strong generalization and data efficiency. Our work sheds light on the limitations of current pre-trained models and the importance of establishing fine-grained vision-language correspondence for relation understanding. + +# References + +[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2 +[2] Arjun R Akula. Question generation for evaluating cross-dataset shifts in multi-modal grounding. arXiv preprint arXiv:2201.09639, 2022. 2 +[3] Cem Anil, Yuhuai Wu, Anders Andreassen, Aitor Lewkowycz, Vedant Misra, Vinay Ramasesh, Ambrose Slone, Guy Gur-Ari, Ethan Dyer, and Behnam Neyshabur. Exploring length generalization in large language models. arXiv preprint arXiv:2207.04901, 2022. 1, 2, 8 +[4] Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. End-to-end algorithm synthesis with recurrent networks: Logical extrapolation without overthinking. arXiv preprint arXiv:2202.05826, 2022. 2 +[5] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006. 5 +[6] Peter Buhlmann. Invariance, causality and robustness. Statistical Science, 35(3):404-426, 2020. 2 +[7] Tianlang Chen and Jiebo Luo. Expressing objects just like words: Recurrent visual embedding for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10583-10590, 2020. 3 +[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2, 4, 7, 8 +[9] Mengjun Cheng, Yipeng Sun, Longchao Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, et al. Vista: vision and scene text aggregation for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5184-5193, 2022. 3 +[10] Myung Jin Choi, Antonio Torralba, and Alan S Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012. 2 +[11] JU Da, Stephen Roller, Sainbayar Sukhbaatar, and Jason E Weston. Staircase attention for recurrent processing of sequences. In Advances in Neural Information Processing Systems. 2 +[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 3 +[13] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. arXiv preprint arXiv:2206.06619, 2022. 3 + +[14] Haiwen Diao, Ying Zhang, Lin Ma, and Huchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 3 +[15] Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305, 2020. 1 +[16] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 1, 7, 8 +[17] Yann Dubois, Gautier Dagan, Dieuwke Hupkes, and Elia Bruni. Location attention for extrapolation to longer sequences. arXiv preprint arXiv:1911.03872, 2019. 2 +[18] Stella Frank, Emanuele Bugliarello, and Desmond Elliott. Vision-and-language or vision-for-language. On CrossModal Influence in Multimodal Transformers.(2021). DOI: https://doi.org/10.18653/v1/2021.emnlp-main, 775, 2021. 2 +[19] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 2 +[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2 +[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 3635-3644, 2021. 1, 2 +[22] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5 +[23] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 804-813, 2017. 6 +[24] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1115-1124, 2017. 5 +[25] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16888-16897, 2021. 3 +[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 9 + +[27] Zhong Ji, Haoran Wang, Jungong Han, and Yanwei Pang. Saliency-guided attention network for image-sentence matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5754-5763, 2019. 3 +[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. In Proceedings of the IEEE international conference on computer vision, pages 2989-2998, 2017. 2 +[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 2 +[30] Kun Kuang, Ruoxuan Xiong, Peng Cui, Susan Athey, and Bo Li. Stable prediction with model misspecification and ag- nostic distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4485-4492, 2020. 2 +[31] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF International conference on computer vision, pages 4654-4662, 2019. 3 +[32] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 3 +[33] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 2 +[34] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11645-11652, 2020. 3, 5 +[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 3 +[36] Mitja Nikolaus, Emmanuelle Salin, Stephane Ayache, Abdellah Fourtassi, and Benoit Favre. Do vision-and-language transformers learn grounded predicate-noun dependencies? arXiv preprint arXiv:2210.12079, 2022. 1, 2, 8 +[37] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2 +[38] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, 2022. 1, 2 + +[39] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2 +[40] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In European Conference on Computer Vision, pages 546-562. Springer, 2022. 3 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3 +[42] Emmanuelle Salin, Badreddine Farah, Stéphane Ayache, and Benoit Favre. Are vision-language transformers learning multimodal representations? a probing perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11248-11257, 2022. 2 +[43] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4 +[44] Avi Schwarzschild, Eitan Borgnia, Arjun Gupta, Furong Huang, Uzi Vishkin, Micah Goldblum, and Tom Goldstein. Can you learn an algorithm? generalizing from easy to hard problems with recurrent networks. Advances in Neural Information Processing Systems, 34:6695-6706, 2021. 2 +[45] Ravi Shekhar, Sandro Pezzelle, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Vision and language integration: Moving beyond objects. In IWCS 2017—12th International Conference on Computational Semantics—Short papers, 2017. 2 +[46] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, 2017. 1, 2 +[47] Xinwei Shen, Furui Liu, Hanze Dong, Qing Lian, Zhitang Chen, and Tong Zhang. Disentangled generative causal representation learning. arXiv preprint arXiv:2010.02637, 2020.2 +[48] Zheyan Shen, Peng Cui, Tong Zhang, and Kun Kunag. Stable learning via sample reweighting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5692-5699, 2020. 2 +[49] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8384, 2019. 2, 6 +[50] Damien Teney, Ehsan Abbasnejad, Kushal Kafle, Robik Shrestha, Christopher Kanan, and Anton Van Den Hengel. On the value of out-of-distribution testing: An example of goodhart's law. Advances in Neural Information Processing Systems, 33:407-417, 2020. 2 + +[51] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1417-1427, 2021. 2 +[52] Christopher Thomas, Yipeng Zhang, and Shih-Fu Chang. Fine-grained visual entailment. arXiv preprint arXiv:2203.15704, 2022. 1, 2, 7, 8 +[53] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2 +[54] Dušan Variš and Ondřej Bojar. Sequence length is a domain: Length-based overfitting in transformer models. arXiv preprint arXiv:2109.07276, 2021. 1, 2, 8 +[55] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 6 +[56] Jia Wang, Jingcheng Ke, Hong-Han Shuai, Yung-Hui Li, and Wen-Huang Cheng. Referring expression comprehension via enhanced cross-modal graph attention networks. ACM Journal of the ACM (JACM), 2022. 3 +[57] Yun Wang, Juncheng Li, and Florian Metze. A comparison of five multiple instance learning pooling functions for sound event detection with weak labeling. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 31-35. IEEE, 2019. 7, 8 +[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquen Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 1, 2, 4, 7, 8 +[59] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 3 +[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3 +[61] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2, 3, 5, 6 +[62] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L. Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3 +[63] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 2 + +[64] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 1, 2, 4, 7 +[65] Zelin Zhao, Karan Samel, Binghong Chen, et al. Proto: Program-guided transformer for program-guided tasks. Advances in Neural Information Processing Systems, 34:17021-17036, 2021. 2 +[66] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4252-4261, 2018. 3 \ No newline at end of file diff --git a/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/images.zip b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..35f7ead0a6cd9381753e7b6b88356b6a95646302 --- /dev/null +++ b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0a1a325e577decacd34fc49fce895e914738616542dc1baee60f36329f5617b +size 370523 diff --git a/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/layout.json b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..534e9a7c648d1910c1b5c132ef35d4cc28615559 --- /dev/null +++ b/2023/Grounded Image Text Matching with Mismatched Relation Reasoning/layout.json @@ -0,0 +1,10004 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 129, + 496, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 129, + 496, + 144 + ], + "spans": [ + { + "bbox": [ + 99, + 129, + 496, + 144 + ], + "type": "text", + "content": "Grounded Image Text Matching with Mismatched Relation Reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "spans": [ + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "text", + "content": "Yu Wu" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "text", + "content": " Yana Wei" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "text", + "content": " Haozhe Wang" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "text", + "content": " Yongfei Liu" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "text", + "content": " Sibei Yang" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "text", + "content": " Xuming He" + }, + { + "bbox": [ + 129, + 164, + 499, + 177 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 178, + 537, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 178, + 537, + 191 + ], + "spans": [ + { + "bbox": [ + 91, + 178, + 537, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 178, + 537, + 191 + ], + "type": "text", + "content": "ShanghaiTech University, " + }, + { + "bbox": [ + 91, + 178, + 537, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 178, + 537, + 191 + ], + "type": "text", + "content": "Shanghai Engineering Research Center of Intelligent Vision and Imaging" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 192, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 192, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 75, + 192, + 553, + 204 + ], + "type": "text", + "content": "{wuyul, weiynl, yangsb, hexm}@shanghaiitech.edu.cn jasper.whz@outlook.com liuyongfei314@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 157, + 228, + 201, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 251, + 290, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 251, + 290, + 470 + ], + "spans": [ + { + "bbox": [ + 68, + 251, + 290, + 470 + ], + "type": "text", + "content": "This paper introduces Grounded Image Text Matching with Mismatched Relation (GITM-MR), a novel visual-linguistic joint task that evaluates the relation understanding capabilities of transformer-based pre-trained models. GITM-MR requires a model to first determine if an expression describes an image, then localize referred objects or ground the mismatched parts of the text. We provide a benchmark for evaluating vision-language (VL) models on this task, with a focus on the challenging settings of limited training data and out-of-distribution sentence lengths. Our evaluation demonstrates that pre-trained VL models often lack data efficiency and length generalization ability. To address this, we propose the Relation-sensitive Correspondence Reasoning Network (RCRN), which incorporates relation-aware reasoning via bi-directional message propagation guided by language structure. Our RCRN can be interpreted as a modular program and delivers strong performance in terms of both length generalization and data efficiency. The code and data are available on https://github.com/SHTUPLUS/GITM-MR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 491, + 143, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 143, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 143, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 509, + 290, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 509, + 290, + 629 + ], + "spans": [ + { + "bbox": [ + 68, + 509, + 290, + 629 + ], + "type": "text", + "content": "Recently, transformer-based vision-language (VL) pretrained models have made significant progress on various VL tasks by fine-tuning on downstream tasks [8, 64, 58, 29, 16]. Despite their success, the representation capacity of such VL pre-trained models remains poorly understood. As a result, an increasing number of studies start to probe the limitations of those learned models from several aspects. The first aspect focuses on the lack of fine-grained understanding of multi-modal data, which is indicated by the weak VL correspondence learned on relations among the entities [36]. In particular, the researchers constructed" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 230, + 424, + 335 + ], + "blocks": [ + { + "bbox": [ + 326, + 230, + 424, + 335 + ], + "lines": [ + { + "bbox": [ + 326, + 230, + 424, + 335 + ], + "spans": [ + { + "bbox": [ + 326, + 230, + 424, + 335 + ], + "type": "image", + "image_path": "08e7c43eec4ca58a4f5c3986ca78d6eacee85e98ec4273a39768efa687a4fac4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 526, + 390 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 526, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 526, + 390 + ], + "type": "text", + "content": "Figure 1. Illustration of the GITM-MR task and the proposed model RCRN. The task is to first identify whether the expression describes the image, and then localize the referred object if the pair are matched, or otherwise ground the mismatched language part. Both data efficiency and length generalization are required." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 424, + 230, + 506, + 335 + ], + "blocks": [ + { + "bbox": [ + 424, + 230, + 506, + 335 + ], + "lines": [ + { + "bbox": [ + 424, + 230, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 424, + 230, + 506, + 335 + ], + "type": "image", + "image_path": "bc45aeb4d866410428d817cf035afcbf356cc3425003ff5f707755d3e2cc8e84.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 409, + 526, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 526, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 526, + 595 + ], + "type": "text", + "content": "subtle relation phrase differences in text to reveal the poor matching capability between visual relations and linguistic relations [46, 21, 38], or required the model to output fine-grained predictions to thoroughly examine the model's understanding of VL correspondence on different uni-modal component representations [52]. Additionally, while large models generally deliver superior performance, they tend to overfit on small datasets, emphasizing the need to enhance the data-efficiency on fine-tuning [15]. Furthermore, when the text is composed of multiple relations, inferring VL correspondence typically involves combining those semantic relations. This combination requires the model to reason the global semantic correspondence, where the length generalization is shown to be a critical weakness in reasoning capability of pre-trained transformers [3, 54]. However, most of those studies only focus on a single aspect of the VL models and hence produce a partial view on their limitations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "content": "To better understand the VL pre-trained models, we introduce a novel VL joint task named Grounded Image Text Matching with Mismatched Relation (GITM-MR) which focuses on distinguishing nuance in relations within text and image, and requires a model to show explicit comprehension on relations in both modalities. As shown in Fig. 1, the task first requires the model to identify whether an expression describes a given image, and then grounds the referred" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "spans": [ + { + "bbox": [ + 159, + 34, + 478, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 45, + 430, + 56 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 45, + 430, + 56 + ], + "spans": [ + { + "bbox": [ + 206, + 45, + 430, + 56 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 56, + 448, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 56, + 448, + 67 + ], + "spans": [ + { + "bbox": [ + 189, + 56, + 448, + 67 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 641, + 289, + 687 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 641, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 641, + 289, + 687 + ], + "type": "text", + "content": "*Both authors contributed equally to this work, which was supported by Shanghai Science and Technology Program 21010502700, Shanghai Frontiers Science Center of Human-centered Artificial Intelligence and MoE Key Lab of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 706, + 315, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 315, + 715 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 315, + 715 + ], + "type": "text", + "content": "2976" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 100, + 289, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 100, + 289, + 242 + ], + "spans": [ + { + "bbox": [ + 67, + 100, + 289, + 242 + ], + "type": "text", + "content": "object if the image-text pair are matched or localizes the mismatched relation in the expression otherwise. The two fine-grained sub-tasks verify the specific understanding on relations both in match and mismatch scenarios. Moreover, the proposed task also allows us to evaluate the data efficiency of model's fine-tuning and its generalization ability. In particular, we design a learning setting with limited annotations, where only the data in a limited size and domain have annotations (i.e., a mix of in- and out-of-distribution (OOD) test data). Given the task design, we then build a new benchmark based on the Ref-Reasoning [61] dataset, which is more general and challenging than previous probing benchmarks on relations and with low linguistic bias." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 244, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 244, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 67, + 244, + 289, + 386 + ], + "type": "text", + "content": "Leveraging our proposed benchmark, we evaluate several representative state-of-the-art models on the task, assessing their performance under limited data sizes and varying domains of input length. Our results reveal their limitations in the both settings. To alleviate these drawbacks, we propose a unified modular framework that jointly solves subtasks in the GITM-MR problem. By explicitly capturing relation correspondence and utilizing it to propagate messages, our approach is capable of obtaining a comprehensive understanding of global correspondence. Additionally, its lightweight and compositional module design allows for excellent data efficiency and length generalization, accompanied by a more transparent reasoning process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 388, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 388, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 67, + 388, + 289, + 529 + ], + "type": "text", + "content": "Specifically, we develop a graph neural network on a parsed language scene graph (LSG), named Relation-sensitive Correspondence Reasoning Network (RCRN), which exploits the language structure to perform relation-aware reasoning in a belief space of visual-linguistic alignment. Our RCRN uses a set of primitive reasoning modules to operate on the pre-trained vision-language features under the guidance of LSG, and predict the fine-level grounding in both match and mismatch scenarios. As demonstrated by the experiments, our method outperforms in both data efficiency and length generalization settings, demonstrating its effectiveness in learning relation correspondences with data efficiency and OOD generalization capability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 532, + 286, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 532, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 80, + 532, + 286, + 542 + ], + "type": "text", + "content": "The contribution of our work is summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 555, + 288, + 687 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 80, + 555, + 288, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 555, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 80, + 555, + 288, + 609 + ], + "type": "text", + "content": "- We propose a new VL task GITM-MR that challenges relation correspondence learning in VL pre-trained model, and focusing on both data efficiency and out-of-distribution text length setting, with a new benchmark for GITM-MR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 609, + 288, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 609, + 288, + 642 + ], + "spans": [ + { + "bbox": [ + 80, + 609, + 288, + 642 + ], + "type": "text", + "content": "- We identify the shortcomings of current VL models through the benchmark, and develop a modular graph network RCRN to address these issues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 643, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 643, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 80, + 643, + 288, + 687 + ], + "type": "text", + "content": "- Our method achieves superior results on both settings, validating its relation learning efficiency and generalization ability, which hopefully inspires the following exploration of these challenges." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 99, + 389, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 99, + 389, + 110 + ], + "spans": [ + { + "bbox": [ + 306, + 99, + 389, + 110 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 117, + 526, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 117, + 526, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 117, + 526, + 380 + ], + "type": "text", + "content": "Probing Tasks for VL Models Despite the rapid development on VL pre-trained models [8, 41, 19, 64, 58, 29], emergent works reveal the limitations and gain some deep understanding on pre-trained representations, including implausible scenes understanding [10], language and vision priors exploration [20], and modality ablation [18]. A considerable number of them have revealed the weakness of VL models in relation understanding. Early work proposed a benchmark FOIL [46, 45] by substituting a word in caption to test non-pretrained VL models on identifying minor differences. Subsequent works [21, 38, 36, 53, 63] further evaluate pre-trained models on detecting subtle differences in relations or predicate-noun dependencies in zero-shot scenarios, and reach a consensus on the poor quality of pre-trained VL representation for relations. FGVE [52] extends this evaluation to the task of fine-grained entailment with model fine-tuning, but lacks evaluation in vision modality. Recently, a new benchmark [42] is proposed to simultaneously assess the concept understanding of representations learned from both modalities. In contrast, our task focuses on generating fine-grained multi-modal output under different fine-tuning settings in order to systematically investigate the relation understanding problem in VL pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 385, + 526, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 526, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 526, + 626 + ], + "type": "text", + "content": "OOD Generalization The general problem of OOD generalization has attracted much attention recently [33, 6, 47, 30, 48]. Among them, several studies have considered the tasks with generalization to processing or generating longer sequences. Some works [3, 54] have studied length generalization in different uni-modal seq-to-seq tasks, and find that fine-tuning pre-trained transformer models can lead to poor out-of-distribution (OOD) performance. To address this issue, various solutions have been proposed, including modifications to attention mechanism [39, 11, 17] and the design of recurrent networks [44, 4], but their application in VL tasks is nontrivial. Other recent works have also revealed the challenge of OOD generalization for the VL tasks, but most of them focuses on the answer bias of Visual Question Answering (VQA) [50, 1, 2]. The attempted solutions include the causal learning [37] and stable learning [51] methods. Perhaps most related to us is the modular network approaches [28, 49, 65], which use program-based design to achieve good generalization for certain OOD scenario such as question domains in VQA. We follow the modular design but focus on the OOD problem induced by the expression complexity in the grounded image text matching task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 632, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 526, + 687 + ], + "type": "text", + "content": "Vision-language Matching Image-Text Matching (ITM) and Referring Expression Grounding (REG) are common cross-modal matching tasks in vision-language domain. The problem of ITM aims to predict a global similarity between a text and an image. Classical methods for ITM fol" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2977" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 99, + 269, + 171 + ], + "blocks": [ + { + "bbox": [ + 88, + 99, + 269, + 171 + ], + "lines": [ + { + "bbox": [ + 88, + 99, + 269, + 171 + ], + "spans": [ + { + "bbox": [ + 88, + 99, + 269, + 171 + ], + "type": "image", + "image_path": "84876a2a1d96c35f5193bef9971e7406e9b2a81684c02549173b9a97e2ad43ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 173, + 289, + 203 + ], + "lines": [ + { + "bbox": [ + 68, + 173, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 68, + 173, + 289, + 203 + ], + "type": "text", + "content": "Figure 2. The length distribution of sentences containing different numbers of entities in the validation set. Different colors represent sentences with different number of entities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 211, + 288, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 211, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 68, + 211, + 288, + 287 + ], + "type": "text", + "content": "low the core idea of learning a common space where linguistic and visual features are compatible [27, 31, 7, 14, 9]. Recent approach [41] concentrates on learning the relation-aware global similarity by a regularization strategy on training. Our model derives a global similarity measure by propagating messages on local correspondences, capable of capturing subtle changes in relations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 288, + 290, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 288, + 290, + 419 + ], + "spans": [ + { + "bbox": [ + 68, + 288, + 290, + 419 + ], + "type": "text", + "content": "The goal of REG is to locate an object in an image based on a language expression. Most methods focus on learning a holistic visual and linguistic representation in a joint space for aligning the object and text, based on CNN+RNN architectures [35, 66, 25] or the Transformer [12, 13, 32, 40, 59]. Another line of research aims to exploit the linguistic structure in the text for better context modeling [62, 34, 60, 56]. In particular, SGMN [61] parses the text into language components and performs reasoning under the guidance of the linguistic structure. Our method introduces a flexible contextual representation and a bi-directional propagation to extend correspondence reasoning beyond the task of REG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 430, + 230, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 430, + 230, + 443 + ], + "spans": [ + { + "bbox": [ + 69, + 430, + 230, + 443 + ], + "type": "text", + "content": "3. Problem Setup and Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 449, + 289, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 449, + 289, + 537 + ], + "spans": [ + { + "bbox": [ + 68, + 449, + 289, + 537 + ], + "type": "text", + "content": "We aim to study the representation capacity of the VL pretrained models for fine-grained downstream tasks. To this end, we introduce the Grounded Image Text Matching with Mismatched Relation (GITM-MR) task, which is a joint matching and grounding task focusing on relations in vision and language domain. In this section, we present the problem setting of GITM-MR, followed by building a new benchmark for the task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 543, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 543, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 68, + 543, + 288, + 653 + ], + "type": "text", + "content": "GITM-MR Task Given an image and a referring expression, our goal is to determine if the expression describes an object in the image or not (termed as Image-Text Matching (ITM)), and then to localize the referred object if it is matched (termed as Referring Expression Grounding (REG)), or otherwise ground the mismatched relation phrase (termed as Mismatched Relation Reasoning (MRR)). Notably, the only difference between match and mismatch is a relation phrase in the expression, making our task particularly challenging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "content": "Formally, given an image " + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "content": " and a text description " + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "content": ", we denote their matching state as " + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "content": ", indicating whether the image matches the description " + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "inline_equation", + "content": "(y = 1)" + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "content": " or not " + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "inline_equation", + "content": "(y = 0)" + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "content": ". For the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": "matched case, the referred object has a corresponding location in the image, represented by its bounding box " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": ". For the mismatch case, we denote the set of all possible relation phrases (i.e. predicates) for mismatch reasoning as " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{r_i\\}_{i=1}^{N_e}" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": ", and there is a target mismatched relation " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "r_i^*" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": " in the relation phrase set " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": ". Given the input " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": ", our goal is to first predict " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": " for the input pair, and then predict " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": " for the referred object in " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "y = 1" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": ", or otherwise predict " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "r_i^*" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": " to explain why " + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 304, + 100, + 525, + 198 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 199, + 525, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 199, + 525, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 525, + 265 + ], + "type": "text", + "content": "For model training, we follow the convention of ITM and REG tasks, and assume the annotations for those two subtasks are provided. However, it is usually costly to collect annotations for the subtle MRR task. Consequently, we assume a weakly-supervised learning setting where only the matching state label " + }, + { + "bbox": [ + 304, + 199, + 525, + 265 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 199, + 525, + 265 + ], + "type": "text", + "content": " is given for mismatched data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 283, + 525, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 525, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 525, + 437 + ], + "type": "text", + "content": "Our Benchmark To conduct systematical study, we construct a benchmark for the GITM-MR task, including a new dataset and two challenging evaluation settings. We first build a dataset of image-text pairs from the public REC dataset, Ref-Reasoning [61], by substituting the relation phrases to generate mismatched referring expressions for images. From the provided relation phrases in the LSG offered by the dataset, we manually select a subset of 27 commonly-occurred relations, and assign some contextual close but semantically different ones for each relation in the subset as their mismatched candidates. We control the linguistic bias by several measures, including keeping the relation distribution and checking by language-only model, the details of which are stated in the Suppl." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 525, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 525, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 525, + 580 + ], + "type": "text", + "content": "However, replacing the relation phrases may introduce falsely mismatched cases. To ensure the quality of our data, we filter out those false mismatch cases in the test set with the help of annotators. After filtering, the dataset contains more than 1M referring expressions in 60K images. It has 1M, 50K and 10K expression-referent pairs for training, validation, testing, respectively. We only select the original image text pairs which have generated mismatched pairs to keep the matching label balanced. Fig. 2 shows the length distribution of sentences in the validation set of the benchmark. We find it is more general and challenging than previous benchmarks on relation understanding, and refer the reader to see more details of the dataset in Suppl." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 599, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 525, + 687 + ], + "type": "text", + "content": "Evaluation Settings Given the dataset, we develop two evaluation settings to test pretrained VL models on their relation understanding. The first setting is designed to investigate data efficiency during the fine-tuning process. We train the models on small-scale datasets and tested on indistribution test sets. The second setting is designed to evaluate the generalization performance of models with regards to different text input lengths. Here the test set contains" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2978" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 100, + 491, + 262 + ], + "blocks": [ + { + "bbox": [ + 109, + 100, + 491, + 262 + ], + "lines": [ + { + "bbox": [ + 109, + 100, + 491, + 262 + ], + "spans": [ + { + "bbox": [ + 109, + 100, + 491, + 262 + ], + "type": "image", + "image_path": "cbab5a3af3c37e1c2667eb5303a249a44386c3fc3579c02d583c00d5f32fb365.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 266, + 527, + 329 + ], + "lines": [ + { + "bbox": [ + 67, + 266, + 527, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 266, + 527, + 329 + ], + "type": "text", + "content": "Figure 3. Model overview with a mismatched case. Given an image and an expression, we first generate visual and linguistic candidates by a detector and a language parser, and compute their representations. Then we use the Context-sensitive Propagation Network to inter alignments between visual-linguistic candidates, which conducts bi-directional message propagation based on the language graph. The propagation initializes the messages by computing local beliefs, selectively aggregates the context information and updates the belief with a context-sensitive gating function. Predictions for this case are obtained by exploiting the beliefs from the propagation. In the language graph, the ellipses represent entity phrases and the rectangles stand for relation phrases. Further elaboration can be found in the main text." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 335, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 335, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 68, + 335, + 288, + 357 + ], + "type": "text", + "content": "longer and more complex samples on relation combinations compared to the training set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 357, + 290, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 357, + 290, + 489 + ], + "spans": [ + { + "bbox": [ + 68, + 357, + 290, + 489 + ], + "type": "text", + "content": "Specifically, according to the sentence length distribution in Fig. 2, we construct two training sets, named TrainLen16 and Train-Len11, which includes sentences containing less than 16 and 11 words respectively. In the test phase, the model needs to perform the reasoning task on sentences with lengths less than 11 and 16 (in-distribution), as well as sentences longer than these lengths (out-of-distribution). As two training sets accounts for " + }, + { + "bbox": [ + 68, + 357, + 290, + 489 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 68, + 357, + 290, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 357, + 290, + 489 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 68, + 357, + 290, + 489 + ], + "type": "text", + "content": " of the overall training data, those two settings allow us to evaluate the model's data efficiency in fine-tuning and OOD generalization ability simultaneously. Detailed data partition rules are explained in Suppl." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 497, + 151, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 497, + 151, + 510 + ], + "spans": [ + { + "bbox": [ + 69, + 497, + 151, + 510 + ], + "type": "text", + "content": "4. Our Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 515, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 515, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 68, + 515, + 288, + 625 + ], + "type": "text", + "content": "In this section, we present our model design for the GITM-MR task and the learning objectives for training with weakly-annotated data. To tackle the problem, we develop a unified and generalizable framework as outlined in Fig. 1, which consists of two main components: a representation network that computes an initial representation of the linguistic and visual components, and a context-aware correspondence reasoning network for final match predictions. Below we first introduce the two main model components in Sec. 4.1 and 4.2, and then the model learning in Sec. 4.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 631, + 271, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 631, + 271, + 642 + ], + "spans": [ + { + "bbox": [ + 69, + 631, + 271, + 642 + ], + "type": "text", + "content": "4.1. Candidate Generation and Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 647, + 187, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 647, + 187, + 658 + ], + "spans": [ + { + "bbox": [ + 69, + 647, + 187, + 658 + ], + "type": "text", + "content": "4.1.1 Candidate Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 664, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 664, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 664, + 288, + 687 + ], + "type": "text", + "content": "Our method first generates a set of visual objects from the image and phrase candidates from the corresponding text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": "Specifically, for the visual object candidates, we adopt a pre-trained object detector [64] to generate " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "N_{o}" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": " box proposals " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{o_i\\}_{i=1}^{N_o}" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": " from the image " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": ". The detector also provides an initial visual feature (i.e., pooled ROI feature) for each object proposal. For the language phrase candidates, we utilize an off-the-shelf parser [43] to produce a language scene graph " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{E}, \\mathcal{R})" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": " of the text expression " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": ", which consists of a set of entity phrases " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_j\\}_{i=1}^{N_e}" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": " as its nodes and relation phrases as its edges " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{r_{ij}\\}" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": " connects to subject " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": ". We denote the number of relation phrases as " + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "inline_equation", + "content": "|\\mathcal{R}| = N_r" + }, + { + "bbox": [ + 304, + 335, + 526, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 468, + 440, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 440, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 440, + 479 + ], + "type": "text", + "content": "4.1.2 Candidate Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 485, + 525, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 525, + 508 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 525, + 508 + ], + "type": "text", + "content": "Given the visual and linguistic candidates, we compute their representations in two steps, as detailed below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 526, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 526, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 526, + 588 + ], + "type": "text", + "content": "Token-level Representation The first step generates an initial representation for the word tokens and visual objects based on a pre-trained vision-language model. Here we adopt a shallow version of the UNITER model [8], aiming to compute a feature representation for the input tokens in two modalities with a generic vision-language alignment from the pre-trained model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 601, + 525, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 525, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 525, + 623 + ], + "type": "text", + "content": "Candidate Features In the second step, we then compile the token-level representations into a set of features for" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 633, + 526, + 687 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 633, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 633, + 526, + 687 + ], + "type": "text", + "content": "*We note that such a generic representation module has a limited capacity and is typically insufficient for our downstream tasks via a simple fine-tuning procedure as in [8, 58]. Nevertheless, it provides rich semantic features for the subsequent alignment reasoning network (c.f. Sec. 4.2) after a minimal fine-tuning (described in Sec. 5.1), which complements our explicit reasoning process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2979" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 100, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 100, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 68, + 100, + 288, + 133 + ], + "type": "text", + "content": "our visual and linguistic components used in the subsequent reasoning network. Here we adopt a similar strategy as in SGMN [61] for computing the features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "spans": [ + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": "Concretely, for visual objects " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": ", we augment its token-level feature " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " by its spatial location " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i^o" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " (i.e., bounding box parameters). In addition, for every object pair " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "(o_i,o_j)" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " in the image, we represent its relational feature as " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ij}^{o}" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": ", which is computed from the features of its two objects. For the linguistic entities in " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " and relation phrases in " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": ", we feed the corresponding word tokens into a Bi-LSTM [22] with self-attention [24] to generate two sets of features. In the first set, each entities " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " is encoded by a tuple of three features " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "(\\mathbf{e}_i,\\mathbf{l}_i^e,\\mathbf{h}_i^e)" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " is an embedding into the space of object feature " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i^e" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " is an embedding into the space of object location " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i^o" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^e" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " is a linguistic feature from the LSTM states. For the second feature set, we generate a representation " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{ij}^{e}" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": " for each relation phrase " + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "inline_equation", + "content": "r_{ij}\\in \\mathcal{R}" + }, + { + "bbox": [ + 68, + 133, + 290, + 299 + ], + "type": "text", + "content": ". We refer the reader to Suppl. for the details of our feature design." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 305, + 258, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 305, + 258, + 317 + ], + "spans": [ + { + "bbox": [ + 69, + 305, + 258, + 317 + ], + "type": "text", + "content": "4.2. Context-sensitive Propagation Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 323, + 290, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 323, + 290, + 453 + ], + "spans": [ + { + "bbox": [ + 68, + 323, + 290, + 453 + ], + "type": "text", + "content": "We now introduce our second model component, the Context-sensitive Propagation Network (CPN), which aims to infer the alignment between the visual-linguistic candidates and predict the match label of input image-text pair. To this end, we build a modular graph neural network on the language graph " + }, + { + "bbox": [ + 68, + 323, + 290, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 68, + 323, + 290, + 453 + ], + "type": "text", + "content": ", which takes the candidate features as input and performs context-aware reasoning in a belief space of visual-linguistic alignment. Such a structured reasoning process enables us to exploit the regularity in language, i.e. the connections of entities by relations, leading to a relation-aware and compositional alignment model with potentially better generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "content": "Specifically, as illustrated in the language graph in Fig. 3, the " + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "content": "-th node of the graph network corresponds to an entity phrase " + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "content": " and the edge between the node " + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "content": " corresponds to a relation phrase " + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "content": ". We follow the common assumption that the corresponding " + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 68, + 455, + 288, + 620 + ], + "type": "text", + "content": " has a tree structure [61, 34] and the referred entity is regarded as the root node. To perform reasoning, the graph network uses a bi-directional message propagation procedure to infer a relation-aware correspondence belief (w.r.t. the visual candidates) for all the entity nodes (described in Sec. 4.2.1), followed by a readout head module to predict the fine-grained and global match scores (described in Sec. 4.2.2). Finally, we formulate each key operation in the graph reasoning as an easy-to-interpret primitive module, and convert our reasoning process into an explainable program in Sec. 4.2.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 635, + 239, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 635, + 239, + 647 + ], + "spans": [ + { + "bbox": [ + 69, + 635, + 239, + 647 + ], + "type": "text", + "content": "4.2.1 Bi-directional Message Propagation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 654, + 289, + 687 + ], + "type": "text", + "content": "Given the graph network, we aim to compute a message for each node, representing a belief on the correspondence between the associated entity phrase and all the visual objects" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 100, + 525, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 525, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 525, + 143 + ], + "type": "text", + "content": "in the image. In order to resolve the ambiguity in the local matching of entities, we introduce a bi-directional message propagation procedure to aggregate all the local beliefs on phrase-visual region correspondence throughout the graph." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 525, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 525, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 525, + 263 + ], + "type": "text", + "content": "Concretely, inspired by belief propagation [5], we perform two parallel passes of message propagation on the tree graph " + }, + { + "bbox": [ + 304, + 144, + 525, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 144, + 525, + 263 + ], + "type": "text", + "content": " along two directions: 1) bottom-up direction (denoted as bp) first computes messages from the leaf nodes and then updates their parents recursively until the root is reached; 2) top-down direction (denoted as td) starts from the root node and updates the children nodes recursively until reaching all the leaves. For each direction, the message propagation aggregates the context information on cross-modal matching and update the belief on phrase alignment at each node in the following three steps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "text", + "content": "(1) Message initialization: We initialize a local belief " + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i^{\\mathrm{loc}}" + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "text", + "content": " at node " + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "text", + "content": " based on a similarity measure between the entity " + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "text", + "content": " and all the visual objects. Here we define the similarity by a weighted combination of feature similarities in appearance and spatial location space. Specifically, for each pair of entity " + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 304, + 264, + 525, + 340 + ], + "type": "text", + "content": ", we compute two similarities as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 335, + 346, + 525, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 346, + 525, + 361 + ], + "spans": [ + { + "bbox": [ + 335, + 346, + 525, + 361 + ], + "type": "interline_equation", + "content": "b _ {i j} ^ {\\mathrm {a p p}} = \\mathrm {F} _ {\\mathrm {s i m}} (\\mathbf {e} _ {i}, \\mathbf {o} _ {j}), \\quad b _ {i j} ^ {\\mathrm {p o s}} = \\mathrm {F} _ {\\mathrm {s i m}} \\left(\\mathrm {I} _ {i} ^ {e}, \\mathrm {I} _ {j} ^ {o}\\right) \\tag {1}", + "image_path": "5afb97e85b4811464715aac49dd5913615e44bd120a0f8999f18f90ed79aa451.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 365, + 525, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 525, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 525, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 365, + 525, + 398 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\mathrm{sim}}" + }, + { + "bbox": [ + 304, + 365, + 525, + 398 + ], + "type": "text", + "content": " is an MLP network computing the similarity between features from two modalities. The local belief " + }, + { + "bbox": [ + 304, + 365, + 525, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i^{\\mathrm{loc}}" + }, + { + "bbox": [ + 304, + 365, + 525, + 398 + ], + "type": "text", + "content": " then integrates two similarities as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 343, + 404, + 525, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 404, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 343, + 404, + 525, + 418 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {i} ^ {\\text {l o c}} = \\mathrm {F} _ {\\text {n o r m}} \\left(\\beta_ {i} ^ {\\text {a p p}} \\mathbf {b} _ {i} ^ {\\text {a p p}} + \\beta_ {i} ^ {\\text {p o s}} \\mathbf {b} _ {i} ^ {\\text {p o s}}\\right) \\tag {2}", + "image_path": "12b25f33160d4180094007af489bcafb4bcb684093c362f610698282b88a2a23.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i^t = [b_{ij}^t]_{j=1}^{N_o}" + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "inline_equation", + "content": "t \\in \\{\\mathrm{app, pos}\\}" + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "inline_equation", + "content": "\\beta_i^t" + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "content": " is the weight for combining similarities and is computed as " + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "inline_equation", + "content": "\\beta_i^t = \\mathrm{sigmoid}(\\mathbf{W}_t^\\top[\\mathbf{h}_i^e, \\mathbf{b}_i^{\\mathrm{app}}, \\mathbf{b}_i^{\\mathrm{pos}}])" + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\mathrm{norm}}" + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "content": " is a normalization function to rescale its input elements to " + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 304, + 425, + 525, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "text", + "content": "(2) Message aggregation: According to the direction " + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{td}" + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{bp}" + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "text", + "content": ", message aggregation collects beliefs from the upstream neighbors in a context-sensitive manner. Formally, we first prune the local beliefs by selecting top " + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "text", + "content": " maximum values, which significantly reduces the computation complexity for large " + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "inline_equation", + "content": "N_{o}" + }, + { + "bbox": [ + 304, + 471, + 525, + 536 + ], + "type": "text", + "content": " and is written as," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 377, + 541, + 525, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 541, + 525, + 556 + ], + "spans": [ + { + "bbox": [ + 377, + 541, + 525, + 556 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {j} ^ {\\text {s e l}} = \\mathrm {F} _ {\\text {t o p k}} \\left(\\mathbf {b} _ {j} ^ {\\text {l o c}}\\right) \\tag {3}", + "image_path": "94add33a9d282bba5c617fdf701ea2754a6ce7d3358928f353e05a00371b0ed2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 561, + 525, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 525, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 525, + 616 + ], + "type": "text", + "content": "In contrast to [61], our selection is belief-based and more informative for propagation. Given the sparsified beliefs, we then aggregate the messages from the neighboring nodes by taking into account the alignment of the relation phrases on the edges as follows," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 622, + 525, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 525, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 525, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {i} ^ {\\mathrm {a g g}} = \\prod_ {j \\in \\operatorname {C t x} (i)} \\mathbf {A} _ {i j} \\mathbf {b} _ {j} ^ {\\mathrm {s e l}}, [ \\mathbf {A} _ {i j} ] _ {k l} = \\mathrm {F} _ {\\text {s i m}} \\left(\\mathbf {r} _ {i j} ^ {e}, \\mathbf {r} _ {k l} ^ {o}\\right) \\tag {4}", + "image_path": "47789a7877b3f97bf92ca854717107198906ac8f9ad19c7badd5b7b261de836d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{Ctx}(i)" + }, + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "text", + "content": " denotes the upstream neighbor set of node " + }, + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 653, + 525, + 687 + ], + "type": "text", + "content": " according to the propagation direction. We model the messages provided by relations explicitly through alignment on" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 315, + 714 + ], + "type": "text", + "content": "2980" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 100, + 289, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 100, + 289, + 165 + ], + "spans": [ + { + "bbox": [ + 67, + 100, + 289, + 165 + ], + "type": "text", + "content": "edges, which forces the model to capture the nuances of different relations. For the aggregation operator, we take the element-wise product of beliefs, which can be viewed as a soft version of logical AND. Unlike the min pooling used in [49, 23], our design leads to a more robust and smooth learning procedure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 165, + 289, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 165, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 68, + 165, + 289, + 209 + ], + "type": "text", + "content": "(3) Message update: Finally, we update the belief " + }, + { + "bbox": [ + 68, + 165, + 289, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i" + }, + { + "bbox": [ + 68, + 165, + 289, + 209 + ], + "type": "text", + "content": " at node " + }, + { + "bbox": [ + 68, + 165, + 289, + 209 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 165, + 289, + 209 + ], + "type": "text", + "content": " by a context-sensitive integration of the aggregated neighboring belief and the local belief. To achieve this, we design a gated product as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 214, + 289, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 214, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 289, + 228 + ], + "type": "interline_equation", + "content": "\\beta_ {i} ^ {\\mathrm {r e l}} = \\operatorname {s i g m o i d} \\left(\\mathbf {W} _ {\\mathrm {r e l}} ^ {\\top} \\left[ \\mathbf {h} _ {i} ^ {e}, \\mathbf {b} _ {i} ^ {\\text {a p p}}, \\mathbf {b} _ {i} ^ {\\text {p o s}} \\right]\\right) \\tag {5}", + "image_path": "ea72e4b97d1205c06f468c412cc795cb8e97554e9d45fc7a8c060465a31164a1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 229, + 289, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 229, + 289, + 248 + ], + "spans": [ + { + "bbox": [ + 110, + 229, + 289, + 248 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {i} = \\mathrm {F} _ {\\text {n o r m}} \\left(\\mathbf {b} _ {i} ^ {\\text {l o c}} \\circ \\left(\\mathbf {b} _ {i} ^ {\\text {a g g}}\\right) ^ {\\beta_ {i} ^ {\\text {r e l}}}\\right) \\tag {6}", + "image_path": "84b0ed4bf47dab55186b983411a81924b7cfe19156f55ea5fe2bd7d76827712f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 253, + 289, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 253, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 68, + 253, + 289, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 253, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 68, + 253, + 289, + 297 + ], + "type": "text", + "content": " denotes the element-wise product. Instead of using the language features as in [61, 55], our gating function also takes the local beliefs as input, making the network more sensitive to the image context." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "content": "Given the updated beliefs, we compute a confidence score for each node matching to any object in the image by applying an MLP network, denoted as " + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "inline_equation", + "content": "p_i = \\mathrm{sigmoid}(\\mathrm{MLP}_{\\mathrm{match}}(\\mathbf{b}_i))" + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "content": ". In the end, the message propagation along two directions generates two belief vectors " + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "inline_equation", + "content": "(\\mathbf{b}_i^{\\mathrm{bp}}" + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i^{\\mathrm{td}})" + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "content": " and two confidence scores " + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "inline_equation", + "content": "(p_i^{\\mathrm{bp}}" + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "inline_equation", + "content": "p_i^{\\mathrm{td}})" + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "content": " for each node " + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 68, + 297, + 289, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 380, + 156, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 380, + 156, + 391 + ], + "spans": [ + { + "bbox": [ + 69, + 380, + 156, + 391 + ], + "type": "text", + "content": "4.2.2 Readout Head" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 398, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 398, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 68, + 398, + 289, + 443 + ], + "type": "text", + "content": "After the graph-based reasoning, we now introduce a readout head network to generate the predictions for three subtasks, i.e., ITM, REG and MRR, by exploiting the beliefs and confidence scores from the bi-directional propagation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 445, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 445, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 68, + 445, + 289, + 477 + ], + "type": "text", + "content": "Image-Text Matching To predict the global matching label " + }, + { + "bbox": [ + 68, + 445, + 289, + 477 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 68, + 445, + 289, + 477 + ], + "type": "text", + "content": ", we take a minimum pooling of the matching confidence scores across the entire graph as below," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 483, + 289, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 483, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 107, + 483, + 289, + 502 + ], + "type": "interline_equation", + "content": "P (y = 1 | I, L) = \\min _ {i} \\left(\\min \\left(p _ {i} ^ {\\mathrm {b p}}, p _ {i} ^ {\\mathrm {t d}}\\right)\\right) \\tag {7}", + "image_path": "cbc5f37a4888b1c56df61add402fa4fc30c03de11f247a3a98cce36bfc732568.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "spans": [ + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "text", + "content": "REG Task For the referent grounding, we take the belief vector of the root node " + }, + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}}" + }, + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "text", + "content": " from the bottom-up direction and locate the matching visual object " + }, + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "inline_equation", + "content": "o_{i^{*}}" + }, + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "text", + "content": " by choosing the largest belief value: " + }, + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "inline_equation", + "content": "i^{*} = \\arg \\max_{i}\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}}" + }, + { + "bbox": [ + 68, + 506, + 289, + 573 + ], + "type": "text", + "content": ". A standard regression head is also employed for more precise localization and its details are left to the Suppl." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 574, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 574, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 68, + 574, + 289, + 640 + ], + "type": "text", + "content": "MRR Task For mismatched relation prediction, we exploits the inconsistency of the match confidence scores along the graph edges. We found the mismatched relation typically leads to a confidence gap between its two incident nodes. As such, we choose the edge with the largest match confidence difference between its two nodes:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 645, + 289, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 645, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 93, + 645, + 289, + 664 + ], + "type": "interline_equation", + "content": "\\hat {r} = \\arg \\min _ {r _ {i j} \\in \\mathcal {R}} \\left(\\left\\{\\left(p _ {i} ^ {\\mathrm {b p}} - p _ {j} ^ {\\mathrm {b p}}\\right) + \\left(p _ {j} ^ {\\mathrm {t d}} - p _ {i} ^ {\\mathrm {t d}}\\right) \\right\\}\\right), \\tag {8}", + "image_path": "1c7ab859b00d9374022c1a01417300d4c7ca86ba8c857cba61de6022adbca287.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 68, + 670, + 289, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 670, + 289, + 691 + ], + "spans": [ + { + "bbox": [ + 68, + 670, + 289, + 691 + ], + "type": "text", + "content": "where we sum the differences from two propagation directions to make the decision more robust." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 138, + 527, + 288 + ], + "blocks": [ + { + "bbox": [ + 304, + 97, + 526, + 137 + ], + "lines": [ + { + "bbox": [ + 304, + 97, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 526, + 137 + ], + "type": "text", + "content": "Table 1. The full list of the reasoning modules and their implementations. We use bold lowercase variables to represent vectors and bold uppercase variables to represent matrices. Regular variables are scalars. W is trainable parameter matrix." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 138, + 527, + 288 + ], + "lines": [ + { + "bbox": [ + 309, + 138, + 527, + 288 + ], + "spans": [ + { + "bbox": [ + 309, + 138, + 527, + 288 + ], + "type": "table", + "html": "
ModulesInOutImplementation
Local Correspondences:
Simv'txt, v'visbFsim(v'txt, v'vis)
GateSum{vi}, {bi}bFnorm(∑i sigmoid(W^T[vi, bi])bi)
Correspondence Reasoning:
Selectbb'Ftopk(b)
Aggregate{Ai}, {bi}b'Πi(AiBi)
GateProdb, b', βib''Fnorm(b o (b')βi)
ClassifybpMLPmatch(b)
Readout Head:
Locatepiarg min(p)
Comparepi, pjdpi - pj
And{pi}pminpi
", + "image_path": "4bcebd092f06acf17769977c3da07fa4e8a4eca409dbe55e2686d4fe5ee9a8f9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 297, + 465, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 465, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 465, + 309 + ], + "type": "text", + "content": "4.2.3 Modular Network Interpretation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 317, + 526, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 526, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 526, + 569 + ], + "type": "text", + "content": "We now introduce an interpretation of our using the modular network framework. As shown in Tab. 1, we summarize the key operations in the RCRN into a set of primitive modules with semantic meanings. This allows us to build a explainable reasoning program: (1) In the message initialization step, local beliefs are computed using Sim, and are combined by GateSum to generate the initial belief on each node. (2) For message aggregation, we first use Select to prune each node and then apply Aggregate to collect the messages from all neighbors, considering the alignment of relations particularly. (3) For message update, GateProd is used to combine the aggregated message with the local belief to generate the updated belief on each node. The Classify module provides a confidence score indicating the likelihood that the phrase entity matches an visual object. (4) In the readout head, we use two Compare modules in two directions respectively for MRR and one And operation for the ITM task, and finally generate predictions for all three subtasks via Locate module. For instance, the grounding prediction is achieved by " + }, + { + "bbox": [ + 304, + 317, + 526, + 569 + ], + "type": "inline_equation", + "content": "\\text{Locate}(-\\mathbf{b}_{\\text{ref}})" + }, + { + "bbox": [ + 304, + 317, + 526, + 569 + ], + "type": "text", + "content": ". Several example are provided in Sec. 5.4 and Suppl. to illustrate the interpretable reasoning process, along with further detailed implementations of the modules." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 526, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 526, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 526, + 625 + ], + "type": "text", + "content": "We note that such a program formulation indicates that the model can enjoy a good out-of-domain generalization due to the compositionality of language graphs. Even facing with more complex text structures, we can perform this reasoning process by reusing those modules on new graphs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 397, + 647 + ], + "type": "text", + "content": "4.3. Model Learning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 526, + 687 + ], + "type": "text", + "content": "We train the network on a multitask loss with end-to-end back propagation. Cross-Entropy (CE) loss and Binary Cross-Entropy (BCE) loss are used for the REG and ITM" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 313, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 313, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 313, + 714 + ], + "type": "text", + "content": "2981" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 100, + 141, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 100, + 141, + 111 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 141, + 111 + ], + "type": "text", + "content": "tasks, respectively:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 118, + 288, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 118, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 288, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g r d}} = \\mathrm {C E} \\left(\\mathbf {p} _ {\\mathrm {g r d}}, \\mathbf {p} _ {\\mathrm {g r d}} ^ {*}\\right) \\tag {9}", + "image_path": "ba716ef9e6ad2aa8973a066c2672f7986a0e3e6454b32c6c966e3f4459888a87.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 148, + 288, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 148, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 125, + 148, + 288, + 162 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h}} = \\mathrm {B C E} \\left(p _ {\\text {m a t c h}}, y ^ {*}\\right) \\tag {10}", + "image_path": "69fccddff3d1274a80c57da248d7a1cb2064ccabce402c43180fcc0308d3ab79.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "spans": [ + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{grad}} = \\mathrm{softmax}(\\mathbf{b}_{\\mathrm{root}}^{\\mathrm{bp}})" + }, + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "text", + "content": " stands for the predicted grounding distribution, " + }, + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{grad}}^{*}" + }, + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "text", + "content": " denotes the ground-truth one-hot grounding label, and " + }, + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{match}} = P(y = 1|I,L)" + }, + { + "bbox": [ + 68, + 166, + 289, + 243 + ], + "type": "text", + "content": " denotes the predicted match probability. When the object boxes are from a detector, the box with the maximum IoU with the ground-truth box of the referent is labeled as the localization output. The overall loss is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 251, + 288, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 251, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 121, + 251, + 288, + 264 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {g r d}} + \\mu \\mathcal {L} _ {\\mathrm {m a t c h}} + \\omega \\mathcal {L} _ {\\mathrm {r e g}} \\tag {11}", + "image_path": "d817c24d50082a4eae520ac15f86fec438c7752c7b5b503901efad37c4b41be5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "text", + "content": " are hyper-parameters to balance the three terms. " + }, + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 68, + 271, + 289, + 316 + ], + "type": "text", + "content": " would be zero if the IoU of the ground-truth box and the selected proposal is less than 0.5. The regression loss is standard and can be found in Suppl." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 324, + 143, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 324, + 143, + 337 + ], + "spans": [ + { + "bbox": [ + 69, + 324, + 143, + 337 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 342, + 290, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 342, + 290, + 430 + ], + "spans": [ + { + "bbox": [ + 68, + 342, + 290, + 430 + ], + "type": "text", + "content": "In this section, we conducted experiments on the proposed benchmark to test several baselines of VL pre-trained models on inferring relation correspondences, with a specific focus on their data efficiency and length generalization capabilities. We then compare our model with those strong baselines to show its effectiveness in these settings. Finally we illustrate the visualizations and interpretations of the reasoning process of our model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 436, + 168, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 436, + 168, + 449 + ], + "spans": [ + { + "bbox": [ + 69, + 436, + 168, + 449 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 453, + 288, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 453, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 68, + 453, + 288, + 563 + ], + "type": "text", + "content": "Metrics The evaluation metrics include classification accuracy for three subtasks. The grounding result for a matched case is considered as correct when it is identified as matching and the predicted box has at least 0.5 IoU with its ground-truth location. The grounding accuracy (i.e. Recall@1) is the ratio of correctly grounded cases. For mismatch reasoning, a mismatched case needs to be correctly classified and the mismatched relation should be accurately selected from the candidate set. The MRR accuracy is the top-1 accuracy among the candidates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 566, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 566, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 566, + 289, + 686 + ], + "type": "text", + "content": "VL Pre-trained Models We benchmark five VL models on the proposed dataset. We first choose TCL [58], UNITER [8] and FIBER [16], where TCL is the state-of-the-art pre-trained model for image-text retrieval, UNITER is a vanilla transformer-based VL pre-trained model for matching and grounding tasks, and FIBER uses a specific fine-grained pre-training design. Additionally, we adapt the latest fine-grained visual entailment (FGVE) model [52] to our setting. We compare its potential fine-grained multitask capability with ours, which is designed based on the VL pre-trained model Oscar+[64]. Because none of the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 100, + 526, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 526, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 526, + 198 + ], + "type": "text", + "content": "pretrained VL models can fulfill three subtasks jointly, we also apply the multiple instance learning strategy [57] on UNITER to cope with the weakly-supervised MRR task. As shown in Suppl, we modify those models by adding subtask heads to perform three subtasks. We note that TCL and UNITER use the same pre-training datasets and the pretraining corpus of FIBER and FGVE is larger while all the models are fine-tuned on the GITM-MR, which leads to a relatively fair comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "text", + "content": "Implementation Details For generating object candidates, we take the off-the-shelf detector, VinVL [64], to extract at most 100 proposals with object score great than 0.1 for each image. For model training, the reasoning component of RCRN is trained by the Adam optimizer with the learning rate set to " + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "inline_equation", + "content": "5\\mathrm{e} - 4" + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "text", + "content": " for the proposed dataset. The learning rate of candidate representation module stated in Sec. 4.1 is set to " + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "inline_equation", + "content": "5\\mathrm{e} - 7" + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "text", + "content": ". We train 80K iterations with batchsize 64 totally. The hyper-parameter " + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 202, + 526, + 345 + ], + "type": "text", + "content": " are set as 5, 3 and 1 based on validation, respectively. The representation module uses the first 6 transformer encoder layers in the pre-trained UNITER. The ablation study on layers and more implementation details are included in the Suppl." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 351, + 436, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 436, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 436, + 363 + ], + "type": "text", + "content": "5.2. Results and Comparisons" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 369, + 526, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 526, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 526, + 620 + ], + "type": "text", + "content": "Relation Understanding with Limited Data Tab. 2 (In-Distribution part) shows the performances of pre-trained models on the relation understanding under the condition of limited data. The results indicate that TCL struggles to achieve a match accuracy of more than chance. This could be attributed to its emphasis on global image-text matching during pre-training, which may hinder its ability to perform fine-grained understanding tasks. Though FGVE is designed for fine-grained entailment prediction, it still lacks capability on mismatch relation reasoning in our scenario. This is likely due to its reliance on the pre-trained VL model for the alignment between vision and language relations. UNITER and FIBER perform relatively better on the match task than TCL because they both use the World-Region Alignment (WRA) as a pre-training task, and FIBER is also pre-trained with phrase grounding, which helps to establish fine-grained correspondence between image and text. The overall conclusion drawn from these findings is that pre-trained models exhibit unsatisfying performance on the challenging benchmark under the limited data condition, indicating the need for further development in establishing effective visual-linguistic correspondence for relation understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 621, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 526, + 687 + ], + "type": "text", + "content": "In comparison, our RCRN outperforms the SOTA pretrained models on match and grounding tasks in the indistribution scenarios, even with fewer parameters. The model's performance with limited training data indicates the effectiveness of its lightweight and structured design. The design incorporates appropriate inductive biases based" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 315, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 315, + 715 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 315, + 715 + ], + "type": "text", + "content": "2982" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 127, + 525, + 266 + ], + "blocks": [ + { + "bbox": [ + 68, + 96, + 525, + 127 + ], + "lines": [ + { + "bbox": [ + 68, + 96, + 525, + 127 + ], + "spans": [ + { + "bbox": [ + 68, + 96, + 525, + 127 + ], + "type": "text", + "content": "Table 2. Experiment results of testing models' relation understanding with limited data and ability of length generalization. UNITER+MIL represents the UNITER with multiple instance learning strategy applied. FGVE+MAX represents the fine-grained visual entailment model with max pooling on knowledge elements to predict mismatched relation. - means the model can't handle the corresponding task." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 127, + 525, + 266 + ], + "lines": [ + { + "bbox": [ + 71, + 127, + 525, + 266 + ], + "spans": [ + { + "bbox": [ + 71, + 127, + 525, + 266 + ], + "type": "table", + "html": "
Training SetMethod#Pretrain Images#ParamFull TestIn-DistributionOut-of-Distribution
Match%Grounding%MRR%Match%Grounding%MRR%Match%Grounding%MRR%
Train-Len16TCL [58]4M210M52.77--53.11--52.70--
UNITER [8]4M111M61.9324.16-71.2940.63-59.9420.67-
UNITER+MIL [57]4M111M59.1318.4647.568.8341.5361.3857.0713.5644.53
FIBER [16]4.8M254M52.319.01-59.0437.15-50.883.03-
FGVE+MAX [52]5.7M113M54.30-33.2358.37-51.7953.43-29.26
RCRN (Ours)4M90M66.5929.1055.7271.8546.2459.0465.4725.4655.01
Train-Len11TCL [58]4M210M51.71--52.08--51.70--
UNITER [8]4M111M55.5110.24-68.2230.05-54.979.42-
UNITER+MIL [57]4M111M55.4311.4442.7365.0429.5661.6555.0210.6941.93
FIBER [16]4.8M254M53.5211.74-57.2129.56-53.3611.00-
FGVE+MAX [52]5.7M113M53.40-38.0555.26-40.7853.31-37.93
RCRN (Ours)4M90M63.4121.4552.6768.7037.9357.7763.1920.7752.46
", + "image_path": "ebff7d1d41b2283aee690ca70896344327407618422caeec8e333435c3f6c20e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 285, + 289, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 285, + 289, + 317 + ], + "spans": [ + { + "bbox": [ + 68, + 285, + 289, + 317 + ], + "type": "text", + "content": "on sentence structure to establish both relation and entity correspondence in a visual-linguistic context, and learns the reasoning strategy effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 323, + 289, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 323, + 289, + 466 + ], + "spans": [ + { + "bbox": [ + 68, + 323, + 289, + 466 + ], + "type": "text", + "content": "Length Generalization Results about length generalization ability are shown in the Out-of-Distribution part of Tab. 2. Overall, the OOD performance of pre-trained models on both test settings was generally poor compared to their in-distribution performance, especially on the grounding and MRR subtasks, which explicitly reflect the models' understanding of relations. In particular, in the OOD setting, FIBER's match accuracy is similar to TCL, with both models only slightly exceeding " + }, + { + "bbox": [ + 68, + 323, + 289, + 466 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 68, + 323, + 289, + 466 + ], + "type": "text", + "content": ". The OOD results suggest that these models tend to overfit to the training data biases, without truly establishing the universal VL correspondence of the relations, thus preventing them from generalizing to longer sentences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "spans": [ + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "text", + "content": "In contrast, RCRN achieved notably better performance than the other models on all three subtasks in both OOD scenarios. Specifically, RCRN trained with Train-Len11 outperformed the best performance of other models by " + }, + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "inline_equation", + "content": "8.17\\%" + }, + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "inline_equation", + "content": "10.08\\%" + }, + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "inline_equation", + "content": "10.53\\%" + }, + { + "bbox": [ + 68, + 467, + 289, + 576 + ], + "type": "text", + "content": " on all three tasks. This success can be attributed to the modular design, which helps the model learn to establish local visual-linguistic correspondence and the universal reasoning strategy. During test, the model can exploit the regularity in sentence structures to compose the modules and deal with longer sentences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 588, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 588, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 68, + 588, + 289, + 687 + ], + "type": "text", + "content": "Discussion These experiments not only help us investigate the two challenges that the pre-trained models faced, but also lead to other interesting findings. First, our observation that fine-grained pre-training tasks, such as WRA, can help models learn local correspondences on downstream tasks, is consistent with the findings of [36]. Second, our findings in the OOD scenario are consistent with previous works [3, 54], which suggest that pre-trained models are prone to overfitting to the training data length biases." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 314, + 293, + 518, + 362 + ], + "blocks": [ + { + "bbox": [ + 306, + 281, + 518, + 292 + ], + "lines": [ + { + "bbox": [ + 306, + 281, + 518, + 292 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 518, + 292 + ], + "type": "text", + "content": "Table 3. Ablation study on the benchmark's OOD validation set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 293, + 518, + 362 + ], + "lines": [ + { + "bbox": [ + 314, + 293, + 518, + 362 + ], + "spans": [ + { + "bbox": [ + 314, + 293, + 518, + 362 + ], + "type": "table", + "html": "
LangβCtxβBi-MPVLPMatGrdMRR
---√(w/o MP)54.7913.9941.47
--✓✓59.6523.8045.58
✓-✓✓60.8824.1849.57
✓✓-✓60.0723.8848.62
✓✓✓-60.7029.2536.65
✓✓✓✓61.4723.3753.23
", + "image_path": "e158282e63d81c6d6c5cef53500e5f6fee2e30f84231fde1c4696d72d3f80b67.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 383, + 525, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 525, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 525, + 460 + ], + "type": "text", + "content": "Moreover, to independently investigate the capability on two subtasks, namely grounding and MRR, we also evaluate on an oracle setting, where we suppose that the models always predict correctly on the matching task. The experiment demonstrates our RCRN has strong performance on both subtasks, and the details of this experiment are provided in the Suppl." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 460, + 525, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 525, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 525, + 538 + ], + "type": "text", + "content": "Overall, our experiments provide insights into the challenges and potential solutions for the relation understanding in VL models. The fine-grained pre-training tasks can help models learn local correspondences, while the modular design can facilitate the composition of modules to deal with longer sentences. We believe that our findings offer a promising direction to effectively address these challenges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 547, + 389, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 547, + 389, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 547, + 389, + 559 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 565, + 525, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 525, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 525, + 597 + ], + "type": "text", + "content": "In this section, we present the effectiveness of each component in the proposed RCRN through ablative experiments trained on the Train-Len16 set, with the results in Tab. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "content": "We conduct ablation study by removing some specific designs. " + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{Lang}\\beta" + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "content": " stands for using the language features as one input for gate functions in Eq. 2 and Eq. 6. Removing it indicates constant " + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "content": "'s in the propagation. " + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{Ctx}\\beta" + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "content": " means feeding initial similarities into the gate functions. Bi-MP represents propagating twice individually using different " + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "inline_equation", + "content": "\\beta^{\\mathrm{rel}}" + }, + { + "bbox": [ + 304, + 598, + 526, + 687 + ], + "type": "text", + "content": "'s for matching task reasoning. Without this part, only the propagation from the root to leaves is conducted. VLP" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2983" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 101, + 293, + 229 + ], + "blocks": [ + { + "bbox": [ + 77, + 101, + 293, + 229 + ], + "lines": [ + { + "bbox": [ + 77, + 101, + 293, + 229 + ], + "spans": [ + { + "bbox": [ + 77, + 101, + 293, + 229 + ], + "type": "image", + "image_path": "8cb73cd665b761423235616b8428a77eaf47f7e582b22b34e696a407186fb651.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 231, + 288, + 238 + ], + "lines": [ + { + "bbox": [ + 79, + 231, + 288, + 238 + ], + "spans": [ + { + "bbox": [ + 79, + 231, + 288, + 238 + ], + "type": "text", + "content": "(a) There is a guy wearing the black shirt and to the right of the boy that holding a cell phone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 241, + 291, + 324 + ], + "blocks": [ + { + "bbox": [ + 78, + 241, + 291, + 324 + ], + "lines": [ + { + "bbox": [ + 78, + 241, + 291, + 324 + ], + "spans": [ + { + "bbox": [ + 78, + 241, + 291, + 324 + ], + "type": "image", + "image_path": "8e689158bbf2041f9a120a5acaea4aa9336b1d2692096169a6ea00abcab481a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 325, + 285, + 332 + ], + "lines": [ + { + "bbox": [ + 79, + 325, + 285, + 332 + ], + "spans": [ + { + "bbox": [ + 79, + 325, + 285, + 332 + ], + "type": "text", + "content": "(b) There is a guy wearing the black shirt and to the left of the boy that holding a cell phone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 68, + 337, + 289, + 418 + ], + "lines": [ + { + "bbox": [ + 68, + 337, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 68, + 337, + 289, + 418 + ], + "type": "text", + "content": "Figure 4. The visualization of the reasoning process of RCRN. The entity and relation correspondence maps are shown by color maps on boxes and box pairs respectively, where only those with top 5 correspondence scores for each phrase are included. Small color maps in (a) are the local correspondence maps and the large ones are the propagated contextualized maps. The boxes of parents in box pairs are dotted on their centers. Dashed boxes indicate the pooling results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 438, + 288, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 438, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 68, + 438, + 288, + 492 + ], + "type": "text", + "content": "stands for introducing vision-language pre-trained transformer layers to encode features. Particularly, the first line with VLP only stands for the model which consists of 6 transformer layers from UNITER and the UNITER task heads stated in Sec. 5.1 and Sec. 5.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 494, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 494, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 68, + 494, + 288, + 581 + ], + "type": "text", + "content": "The results show that all of these designs improve the matching performance. VLP model (first line in Tab. 3) with the same setting as the VLP layers in RCRN performs poorly on OOD scenario, which demonstrates our footnote statement in Sec. 4.1. Nevertheless, VLP component of RCRN is significant on MRR performance, probably because it provides accurate preliminary vision-language alignments on relations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 591, + 233, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 591, + 233, + 604 + ], + "spans": [ + { + "bbox": [ + 69, + 591, + 233, + 604 + ], + "type": "text", + "content": "5.4. Visualization and Interpretability" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 609, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 609, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 68, + 609, + 288, + 686 + ], + "type": "text", + "content": "As shown in Fig. 4, we visualize a pair of matched and mismatched samples of our RCRN. Fig. 4(a) shows the bottom-up propagation process followed by grounding prediction on a matched image-text pair, and Fig. 4(b) presents the bi-directional propagation results with readout process for MRR on a mismatched pair. In Fig. 4(a), the propagation eliminates the ambiguities on entities under the guid" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 127, + 525, + 180 + ], + "blocks": [ + { + "bbox": [ + 305, + 97, + 525, + 126 + ], + "lines": [ + { + "bbox": [ + 305, + 97, + 525, + 126 + ], + "spans": [ + { + "bbox": [ + 305, + 97, + 525, + 126 + ], + "type": "text", + "content": "Table 4. Intermediate diagnosis results of RCRN. E.G.R. means Entity Grounding Recall. R.G.R. means Referent Grounding Recall. R.C.S. means Relation Correspondence Scores." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 127, + 525, + 180 + ], + "lines": [ + { + "bbox": [ + 308, + 127, + 525, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 127, + 525, + 180 + ], + "type": "table", + "html": "
Inference MethodE.G.R.R.G.R.R.C.S.
R@1R@3R@5R@1R@3R@5GTMeanMax
By chance6.3718.2528.785.7416.9427.22---
w/o MP50.5080.5989.7938.8875.8488.110.610.170.69
w/ MP60.8285.5892.1669.4690.9295.590.610.170.69
", + "image_path": "f77a01f54304cddd60e4a844308c0635f72532728ce759fb7c89dd1308097d88.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 198, + 525, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 525, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 525, + 252 + ], + "type": "text", + "content": "ance of explicit relation correspondences. Fig. 4(b) illustrates that the intermediate reasons of the predictions can be traced. The sum of match confidence differences along the mismatched relation, denoted as " + }, + { + "bbox": [ + 304, + 198, + 525, + 252 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 198, + 525, + 252 + ], + "type": "text", + "content": " in the dotted box, are notably lower than other " + }, + { + "bbox": [ + 304, + 198, + 525, + 252 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 198, + 525, + 252 + ], + "type": "text", + "content": "'s." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 253, + 525, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 525, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 525, + 362 + ], + "type": "text", + "content": "Tab. 4 demonstrates some interpretable intermediate diagnosis results of RCRN on a subset of the validation set, with ground-truth box proposals. We analyze from the aspect of phrase-level visual grounding recall and statistics on relation correspondence scores. To obtain the true correspondence between the language scene graph and the bounding boxes, we search the parsed language scene graph in the original image scene graph in GQA dataset [26] with a noisy graph matching algorithm. See Suppl. for the detail of the algorithm." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 362, + 525, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 525, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 525, + 460 + ], + "type": "text", + "content": "The results verify that reasonable local correspondences are built, since the mean of initial recall (i.e. inference without MP) on entity phrase grounding is much higher than the random guess. As for relation correspondences, the average values on the ground-truth vision-language relation correspondences are far above the average mean value on all the pairs and very close to the average maximum. Moreover, message propagation notably helps refine the correspondence maps, especially on the referent entities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 481, + 372, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 372, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 372, + 493 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 499, + 525, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 525, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 525, + 686 + ], + "type": "text", + "content": "In this paper, we have introduced a novel VL joint task, Grounded Image Text Matching with Mismatched Relation (GITM-MR), which provides a challenging benchmark for evaluating pre-trained models on the relation reasoning. Our experiments have shown that some state-of-the-art pretrained models struggle with the task under the setting of limited data and out-of-distribution sentence lengths. To address this problem, we develop a Relation-sensitive Correspondence Reasoning Network (RCRN) to compute contextualized cross-modal alignment of both entities and relations, and ground the fine-grained result in the image or expression. The proposed method outperforms prior SOTA pre-trained models, which demonstrates its strong generalization and data efficiency. Our work sheds light on the limitations of current pre-trained models and the importance of establishing fine-grained vision-language correspondence for relation understanding." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 705, + 314, + 714 + ], + "type": "text", + "content": "2984" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 123, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 117, + 290, + 685 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 73, + 117, + 290, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 117, + 290, + 167 + ], + "spans": [ + { + "bbox": [ + 73, + 117, + 290, + 167 + ], + "type": "text", + "content": "[1] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 169, + 289, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 169, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 73, + 169, + 289, + 198 + ], + "type": "text", + "content": "[2] Arjun R Akula. Question generation for evaluating cross-dataset shifts in multi-modal grounding. arXiv preprint arXiv:2201.09639, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 201, + 290, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 201, + 290, + 251 + ], + "spans": [ + { + "bbox": [ + 74, + 201, + 290, + 251 + ], + "type": "text", + "content": "[3] Cem Anil, Yuhuai Wu, Anders Andreassen, Aitor Lewkowycz, Vedant Misra, Vinay Ramasesh, Ambrose Slone, Guy Gur-Ari, Ethan Dyer, and Behnam Neyshabur. Exploring length generalization in large language models. arXiv preprint arXiv:2207.04901, 2022. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 252, + 289, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 252, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 74, + 252, + 289, + 302 + ], + "type": "text", + "content": "[4] Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. End-to-end algorithm synthesis with recurrent networks: Logical extrapolation without overthinking. arXiv preprint arXiv:2202.05826, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 304, + 289, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 304, + 289, + 333 + ], + "spans": [ + { + "bbox": [ + 74, + 304, + 289, + 333 + ], + "type": "text", + "content": "[5] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 335, + 289, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 335, + 289, + 354 + ], + "spans": [ + { + "bbox": [ + 74, + 335, + 289, + 354 + ], + "type": "text", + "content": "[6] Peter Buhlmann. Invariance, causality and robustness. Statistical Science, 35(3):404-426, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 357, + 289, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 357, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 74, + 357, + 289, + 396 + ], + "type": "text", + "content": "[7] Tianlang Chen and Jiebo Luo. Expressing objects just like words: Recurrent visual embedding for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10583-10590, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 398, + 289, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 398, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 74, + 398, + 289, + 448 + ], + "type": "text", + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 449, + 289, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 449, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 74, + 449, + 289, + 510 + ], + "type": "text", + "content": "[9] Mengjun Cheng, Yipeng Sun, Longchao Wang, Xiongwei Zhu, Kun Yao, Jie Chen, Guoli Song, Junyu Han, Jingtuo Liu, Errui Ding, et al. Vista: vision and scene text aggregation for cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5184-5193, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 512, + 288, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 512, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 69, + 512, + 288, + 541 + ], + "type": "text", + "content": "[10] Myung Jin Choi, Antonio Torralba, and Alan S Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 543, + 289, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 543, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 543, + 289, + 582 + ], + "type": "text", + "content": "[11] JU Da, Stephen Roller, Sainbayar Sukhbaatar, and Jason E Weston. Staircase attention for recurrent processing of sequences. In Advances in Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 584, + 289, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 584, + 289, + 633 + ], + "spans": [ + { + "bbox": [ + 69, + 584, + 289, + 633 + ], + "type": "text", + "content": "[12] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 635, + 289, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 635, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 635, + 289, + 685 + ], + "type": "text", + "content": "[13] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. arXiv preprint arXiv:2206.06619, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 525, + 685 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 101, + 525, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 101, + 525, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 101, + 525, + 140 + ], + "type": "text", + "content": "[14] Haiwen Diao, Ying Zhang, Lin Ma, and Huchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 142, + 525, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 525, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 525, + 191 + ], + "type": "text", + "content": "[15] Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 194, + 525, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 525, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 525, + 243 + ], + "type": "text", + "content": "[16] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 1, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 244, + 525, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 525, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 525, + 274 + ], + "type": "text", + "content": "[17] Yann Dubois, Gautier Dagan, Dieuwke Hupkes, and Elia Bruni. Location attention for extrapolation to longer sequences. arXiv preprint arXiv:1911.03872, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 525, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 525, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 525, + 324 + ], + "type": "text", + "content": "[18] Stella Frank, Emanuele Bugliarello, and Desmond Elliott. Vision-and-language or vision-for-language. On CrossModal Influence in Multimodal Transformers.(2021). DOI: https://doi.org/10.18653/v1/2021.emnlp-main, 775, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 525, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 525, + 366 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 525, + 366 + ], + "type": "text", + "content": "[19] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 368, + 525, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 368, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 525, + 418 + ], + "type": "text", + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 420, + 525, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 525, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 525, + 459 + ], + "type": "text", + "content": "[21] Lisa Anne Hendricks and Aida Nematzadeh. Probing image-language transformers for verb understanding. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 3635-3644, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 460, + 525, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 525, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 525, + 480 + ], + "type": "text", + "content": "[22] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 482, + 525, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 525, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 525, + 532 + ], + "type": "text", + "content": "[23] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 804-813, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 525, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 525, + 583 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 525, + 583 + ], + "type": "text", + "content": "[24] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1115-1124, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 585, + 525, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 585, + 525, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 585, + 525, + 634 + ], + "type": "text", + "content": "[25] Binbin Huang, Dongze Lian, Weixin Luo, and Shenghua Gao. Look before you leap: Learning landmark features for one-stage visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16888-16897, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 525, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 525, + 685 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 525, + 685 + ], + "type": "text", + "content": "[26] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 9" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2985" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 685 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 100, + 289, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 100, + 289, + 140 + ], + "type": "text", + "content": "[27] Zhong Ji, Haoran Wang, Jungong Han, and Yanwei Pang. Saliency-guided attention network for image-sentence matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5754-5763, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 143, + 289, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 143, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 70, + 143, + 289, + 194 + ], + "type": "text", + "content": "[28] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Inferring and executing programs for visual reasoning. In Proceedings of the IEEE international conference on computer vision, pages 2989-2998, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 195, + 289, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 195, + 289, + 245 + ], + "spans": [ + { + "bbox": [ + 70, + 195, + 289, + 245 + ], + "type": "text", + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 247, + 289, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 247, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 70, + 247, + 289, + 297 + ], + "type": "text", + "content": "[30] Kun Kuang, Ruoxuan Xiong, Peng Cui, Susan Athey, and Bo Li. Stable prediction with model misspecification and ag- nostic distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4485-4492, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 300, + 289, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 300, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 70, + 300, + 289, + 340 + ], + "type": "text", + "content": "[31] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF International conference on computer vision, pages 4654-4662, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 342, + 289, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 342, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 70, + 342, + 289, + 381 + ], + "type": "text", + "content": "[32] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 384, + 289, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 384, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 70, + 384, + 289, + 434 + ], + "type": "text", + "content": "[33] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 437, + 289, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 437, + 289, + 476 + ], + "spans": [ + { + "bbox": [ + 70, + 437, + 289, + 476 + ], + "type": "text", + "content": "[34] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11645-11652, 2020. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 479, + 289, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 479, + 289, + 520 + ], + "spans": [ + { + "bbox": [ + 70, + 479, + 289, + 520 + ], + "type": "text", + "content": "[35] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In European Conference on Computer Vision, pages 792-807. Springer, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 521, + 289, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 521, + 289, + 561 + ], + "spans": [ + { + "bbox": [ + 70, + 521, + 289, + 561 + ], + "type": "text", + "content": "[36] Mitja Nikolaus, Emmanuelle Salin, Stephane Ayache, Abdellah Fourtassi, and Benoit Favre. Do vision-and-language transformers learn grounded predicate-noun dependencies? arXiv preprint arXiv:2210.12079, 2022. 1, 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 563, + 289, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 563, + 289, + 614 + ], + "spans": [ + { + "bbox": [ + 70, + 563, + 289, + 614 + ], + "type": "text", + "content": "[37] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 616, + 289, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 616, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 70, + 616, + 289, + 685 + ], + "type": "text", + "content": "[38] Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. Valse: A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8253-8280, 2022. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 100, + 524, + 686 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 100, + 524, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 100, + 524, + 130 + ], + "spans": [ + { + "bbox": [ + 307, + 100, + 524, + 130 + ], + "type": "text", + "content": "[39] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 132, + 524, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 524, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 524, + 182 + ], + "type": "text", + "content": "[40] Mengxue Qu, Yu Wu, Wu Liu, Qiqi Gong, Xiaodan Liang, Olga Russakovsky, Yao Zhao, and Yunchao Wei. Siri: A simple selective retraining mechanism for transformer-based visual grounding. In European Conference on Computer Vision, pages 546-562. Springer, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 183, + 524, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 524, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 524, + 243 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 244, + 524, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 524, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 524, + 295 + ], + "type": "text", + "content": "[42] Emmanuelle Salin, Badreddine Farah, Stéphane Ayache, and Benoit Favre. Are vision-language transformers learning multimodal representations? a probing perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11248-11257, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 296, + 524, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 524, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 524, + 346 + ], + "type": "text", + "content": "[43] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 347, + 524, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 524, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 524, + 398 + ], + "type": "text", + "content": "[44] Avi Schwarzschild, Eitan Borgnia, Arjun Gupta, Furong Huang, Uzi Vishkin, Micah Goldblum, and Tom Goldstein. Can you learn an algorithm? generalizing from easy to hard problems with recurrent networks. Advances in Neural Information Processing Systems, 34:6695-6706, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 399, + 524, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 524, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 524, + 449 + ], + "type": "text", + "content": "[45] Ravi Shekhar, Sandro Pezzelle, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Vision and language integration: Moving beyond objects. In IWCS 2017—12th International Conference on Computational Semantics—Short papers, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 450, + 524, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 450, + 524, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 524, + 510 + ], + "type": "text", + "content": "[46] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 255-265, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 512, + 524, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 524, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 524, + 551 + ], + "type": "text", + "content": "[47] Xinwei Shen, Furui Liu, Hanze Dong, Qing Lian, Zhitang Chen, and Tong Zhang. Disentangled generative causal representation learning. arXiv preprint arXiv:2010.02637, 2020.2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 553, + 524, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 553, + 524, + 593 + ], + "spans": [ + { + "bbox": [ + 307, + 553, + 524, + 593 + ], + "type": "text", + "content": "[48] Zheyan Shen, Peng Cui, Tong Zhang, and Kun Kunag. Stable learning via sample reweighting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5692-5699, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 594, + 524, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 594, + 524, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 594, + 524, + 635 + ], + "type": "text", + "content": "[49] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8376-8384, 2019. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 635, + 524, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 524, + 686 + ], + "type": "text", + "content": "[50] Damien Teney, Ehsan Abbasnejad, Kushal Kafle, Robik Shrestha, Christopher Kanan, and Anton Van Den Hengel. On the value of out-of-distribution testing: An example of goodhart's law. Advances in Neural Information Processing Systems, 33:407-417, 2020. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2986" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 686 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 70, + 100, + 289, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 100, + 289, + 149 + ], + "spans": [ + { + "bbox": [ + 70, + 100, + 289, + 149 + ], + "type": "text", + "content": "[51] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1417-1427, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 151, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 151, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 70, + 151, + 288, + 180 + ], + "type": "text", + "content": "[52] Christopher Thomas, Yipeng Zhang, and Shih-Fu Chang. Fine-grained visual entailment. arXiv preprint arXiv:2203.15704, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 182, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 182, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 70, + 182, + 288, + 241 + ], + "type": "text", + "content": "[53] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 243, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 243, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 70, + 243, + 288, + 271 + ], + "type": "text", + "content": "[54] Dušan Variš and Ondřej Bojar. Sequence length is a domain: Length-based overfitting in transformer models. arXiv preprint arXiv:2109.07276, 2021. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 273, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 273, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 70, + 273, + 288, + 312 + ], + "type": "text", + "content": "[55] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 313, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 313, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 70, + 313, + 288, + 352 + ], + "type": "text", + "content": "[56] Jia Wang, Jingcheng Ke, Hong-Han Shuai, Yung-Hui Li, and Wen-Huang Cheng. Referring expression comprehension via enhanced cross-modal graph attention networks. ACM Journal of the ACM (JACM), 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 354, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 354, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 70, + 354, + 288, + 412 + ], + "type": "text", + "content": "[57] Yun Wang, Juncheng Li, and Florian Metze. A comparison of five multiple instance learning pooling functions for sound event detection with weak labeling. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 31-35. IEEE, 2019. 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 414, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 414, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 70, + 414, + 288, + 453 + ], + "type": "text", + "content": "[58] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquen Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 454, + 288, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 454, + 288, + 504 + ], + "spans": [ + { + "bbox": [ + 70, + 454, + 288, + 504 + ], + "type": "text", + "content": "[59] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 505, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 505, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 70, + 505, + 288, + 544 + ], + "type": "text", + "content": "[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 545, + 288, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 545, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 70, + 545, + 288, + 585 + ], + "type": "text", + "content": "[61] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 586, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 586, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 70, + 586, + 288, + 635 + ], + "type": "text", + "content": "[62] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L. Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 636, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 636, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 70, + 636, + 288, + 686 + ], + "type": "text", + "content": "[63] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 100, + 525, + 253 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 100, + 525, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 100, + 525, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 100, + 525, + 159 + ], + "type": "text", + "content": "[64] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 1, 2, 4, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 161, + 525, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 525, + 201 + ], + "type": "text", + "content": "[65] Zelin Zhao, Karan Samel, Binghong Chen, et al. Proto: Program-guided transformer for program-guided tasks. Advances in Neural Information Processing Systems, 34:17021-17036, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 202, + 525, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 202, + 525, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 525, + 253 + ], + "type": "text", + "content": "[66] Bohan Zhuang, Qi Wu, Chunhua Shen, Ian Reid, and Anton Van Den Hengel. Parallel attention: A unified framework for visual object discovery through dialogs and queries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4252-4261, 2018. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 314, + 714 + ], + "type": "text", + "content": "2987" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_content_list.json b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..de2c8af9fdb69d37d2c1265192a74094ef7e66be --- /dev/null +++ b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_content_list.json @@ -0,0 +1,1413 @@ +[ + { + "type": "text", + "text": "Grounding 3D Object Affordance from 2D Interactions in Images", + "text_level": 1, + "bbox": [ + 153, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhang Yang $^{1}$ , Wei Zhai $^{1,*}$ , Hongchen Luo $^{1}$ , Yang Cao $^{1,2}$ , Jiebo Luo $^{3}$ , Zheng-Jun Zha $^{1}$ $^{1}$ University of Science and Technology of China \n $^{3}$ University of Rochester \n $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 140, + 179, + 826, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yyuhang@mail., wzhai056@mail., lhcl2@mail., forrest@ustc.edu.cn", + "bbox": [ + 197, + 236, + 772, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jluo@cs.rochester.edu zhazj@ustc.edu.cn", + "bbox": [ + 300, + 255, + 671, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 318, + 313, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Grounding 3D object affordance seeks to locate objects' \"action possibilities\" regions in the 3D space, which serves as a link between perception and operation for embodied agents. Existing studies primarily focus on connecting visual affordances with geometry structures, e.g., relying on annotations to declare interactive regions of interest on the object and establishing a mapping between the regions and affordances. However, the essence of learning object affordance is to understand how to use it, and the manner that detaches interactions is limited in generalization. Normally, humans possess the ability to perceive object affordances in the physical world through demonstration images or videos. Motivated by this, we introduce a novel task setting: grounding 3D object affordance from 2D interactions in images, which faces the challenge of anticipating affordance through interactions of different sources. To address this problem, we devise a novel Interaction-driven 3D Affordance Grounding Network (IAG), which aligns the region feature of objects from different sources and models the interactive contexts for 3D object affordance grounding. Besides, we collect a Point-Image Affordance Dataset (PIAD) to support the proposed task. Comprehensive experiments on PIAD demonstrate the reliability of the proposed task and the superiority of our method. The project is available at https://github.com/yyvhang/IAGNet.", + "bbox": [ + 75, + 351, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The term \"affordance\" is described as \"opportunities of interaction\" by J. Gibson [14]. Grounding 3D object affordance aims to comprehend the interactive regions of objects in 3D space, which is not only simply to predict which interaction an object affords, but also to identify specific points on the object that could support the interaction. It consti", + "bbox": [ + 75, + 784, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8247bf0739f3d10b9f0296881df42b876a940e11b48883728ea15b3f7f6866b7.jpg", + "image_caption": [ + "Figure 1. Grounding Affordance from Interactions. We propose to ground 3D object affordance through 2D interactions. Given an object point cloud with an interactive image, grounding the corresponding affordance on the 3D object." + ], + "image_footnote": [], + "bbox": [ + 517, + 301, + 875, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tutes a link between perception and operation for embodied agents, which has the potential to serve numerous practical applications, e.g. action prediction [21, 63], robot manipulation [36, 45, 51], imitation learning [17, 50], and augmented/virtual reality [6, 9].", + "bbox": [ + 496, + 551, + 892, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "So far, the paradigm of perceiving 3D object affordance has several branches. One of them involves establishing an explicit mapping between affordance categories and geometry structures [10, 20, 46, 66], based on visual appearance. However, affordance is dynamic and multiple, these geometric-specific manners have limited generalization for unseen structures. Besides, locking the geometry with a specific affordance category may lead to the anticipated region being inconsistent with its affordance when objects possess multiple similar geometrics, resulting in affordance regional confusion. Another paradigm is based on reinforcement learning, which puts the agent in 3D synthetic scenarios to interact with several objects actively, taking the reward mechanism to optimize the whole process [48]. While this type of approach transitions agents from passive recognition to active reasoning, it needs repeated attempts in a huge search space when meeting a novel structure, and is therefore time-consuming.", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Corresponding Author.", + "bbox": [ + 101, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4cc3bd9669f23af051f7af326d58d307576221a3458726e494f5e37aa5a0c742.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 132, + 88, + 450, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8539b518ee166ef2e0c94c58af4233460fc128c48b8e84c8173cb0eeb30dd9de.jpg", + "image_caption": [ + "Figure 2. Motivation. (a) Clues to correlate regions of objects from different sources. Objects with the same category possess a similar combination scheme to meet certain affordance properties, it is a similar inherent relation among different instances. And some structures hint at these properties exhibit closer representations with higher similarity in the latent space. (b) Object affordance may be affected by dynamic factors, such as the position of the object itself in the scene, other objects in the scene, and the body part of the interacting subject. These factors can be decomposed into object-subject and object-scene interaction contexts." + ], + "image_footnote": [], + "bbox": [ + 460, + 89, + 836, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These limitations motivate us to explore an affordance-compatible learning paradigm. Typically, humans can infer object affordance in the 3D physical world by watching images or videos that demonstrate the interactions. Some studies in cognitive science [37, 56] point out the existence of \"body image\" in human cognition, which claims that humans have a perceptual experience of the objects they see, thus facilitating their ability to organize the perception [73] and operate novel objects. Hence, human-object interaction conveys the perception that object structure could perform certain affordance, which is a crucial clue to reason object affordance. In light of this, we present a novel task setting: grounding 3D object affordance from 2D interactions in images, which is shown in Fig. 1.", + "bbox": [ + 75, + 371, + 468, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This challenging task includes several essential issues that should be properly addressed. 1) Alignment ambiguity. To ground 3D object affordance from 2D source interactions, the premise is to correspond the regions of the object in different sources. The object in 2D demonstrations and the 3D object we face are usually derived from different physical instances in different locations and scales. This discrepancy may lead to confusion when matching the affordance regions, causing alignment ambiguity. While objects are commonly designed to satisfy certain needs of human beings, so the same category generally follows a similar combination scheme of object components to meet certain affordances, and these affordances are hinted at by some structures (Fig. 2 (a)). These invariant properties are across instances and could be utilized to correlate object regions from different sources. 2) Affordance ambiguity. Affordance has properties of dynamic and multiplicity, which means the object affordance may change according to the situation, the same part of an object could afford multiple interactions, as shown in Fig. 2 (b), \"Chair\" affords \"Sit\" or \"Move\" depends on the human actions, \"Mug\" af", + "bbox": [ + 75, + 584, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fords \"Wrapgrasp\" or \"Contain\" according to the scene context, these properties may make ambiguity when extracting affordance. However, these dynamic factors for affordance extraction can be decomposed into the interaction between subject-object and object-scene. Modeling these interactions is possible to extract explicit affordance.", + "bbox": [ + 496, + 371, + 890, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose the Interaction-driven 3D Affordance Grounding Network (IAG) to align object region features from different sources and model interaction contexts to reveal affordance. In detail, it contains two sub-modules, one is Joint Region Alignment Module (JRA), which is devised to eliminate alignment ambiguity. It takes the relative difference in dense cross-similarity to refine analogous shape regions and employs learnable layers to map the invariant combination scheme, taking them to match local regions of objects. For affordance ambiguity, the other one Affordance Revealed Module (ARM) takes the object representation as a shared factor and jointly models its interaction contexts with the affordance-related factors to reveal explicit affordance. Moreover, we collect Point-Image Affordance Dataset (PIAD) that contains plenty of paired image-point cloud affordance data and make a benchmark to support the model training and evaluation on the proposed setting.", + "bbox": [ + 496, + 462, + 890, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions are summarized as follows: 1) We introduce grounding 3D object affordance through the 2D interactions, which facilitates the generalization to 3D object affordance perception. 2) We propose the IAG framework, which aligns the region feature of objects from different sources and jointly models affordance-related interactions to locate the 3D object affordance. 3) We collect a dataset named PIAD to support the proposed task setting, which contains paired image-point cloud affordance data. Besides, we establish a benchmark on PIAD, and the experiments on PIAD exhibit the reliability of the method and the setting.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10906", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 89, + 227, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Affordance Learning", + "text_level": 1, + "bbox": [ + 76, + 114, + 276, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At present, the affordance area has made achievements in multiple tasks (Tab. 1). Some works are devoted to detecting the affordance region from 2D sources i.e. images and videos [7, 11, 25, 33, 54, 60, 78], and there are also some studies accomplish this task with the assistance of natural language [32, 38, 39]. They seek to detect or segment objects that afford the action in 2D data, but cannot perceive the specific parts corresponding to the affordance. Thus, another type of method brings a leap to locate the affordance region of objects from 2D sources [12, 35, 47, 34]. However, the affordance knowledge derived from 2D sources is hard to extrapolate specific interactive locations of objects in the 3D environment. With several 3D object datasets proposed [13, 29, 44], some researchers explore grounding object affordance from 3D data [10, 43, 66, 48]. Such methods directly map the connection between semantic affordances and 3D structures, detach the real interaction, and may deter the generalization ability: structures that do not map to a specific affordance are usually hard to generalize through this type of method. T. Nagarajan et al. [48] give a fresh mind, taking the reinforcement learning to make agents actively interact with the 3D synthetic scenarios, while it requires repeated attempts by the agents in a given search space, and is time-consuming. In robotics, affordance is utilized to provide priors for object manipulation and achieve considerable results, especially the articulated 3D objects [42, 64, 79]. Several methods utilize the 2.5D data e.g. RGB-D image to understand the object affordance and serve the manipulation task [21, 22, 23, 49], for these methods, the image and depth information are corresponding in spatial and need to be collected in the same scene. In contrast, our task focuses on grounding 3D object affordance from 2D interactions, in which the 2D interactions provide direct clues to excavate object affordance efficiently and make the affordance could generalize to some unseen structures, and the 2D-3D data is collected from different sources, freeing the constraints on spatial correspondence.", + "bbox": [ + 76, + 140, + 472, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Image-Point Cloud Cross-Modal Learning", + "text_level": 1, + "bbox": [ + 76, + 709, + 439, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The combination of point cloud and image data can capture both semantic and geometric information, enabling cross-modal learning among them has great potential application value in scenarios like autopilot [8]. For this type of work, aligning features is the premise for completing downstream tasks e.g. detection, segmentation, and registration. It is aimed at establishing correspondences between instances from different modalities [2, 26, 69], either spatial or semantic. To achieve this, many methods use camera intrinsic to correspond spatial position of pixels and points, then align per pixel-point feature or fuse the raw", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c7e8c35d4d199dc50b7c69ccb0ff6efe9dfd3816589b8610805918f726a2d5f1.jpg", + "table_caption": [ + "Table 1. Affordance Learning. Various works for several tasks in the affordance community." + ], + "table_footnote": [], + "table_body": "
WorkInputOutputTask
[11, 60, 78]Image/Video2D MaskDetection
[32, 38, 39]Image,Text2D MaskDetection
[21, 23, 49]RGBDHeatmap & ActionManipulation
[42, 64, 79]Point CloudHeatmap & ActionManipulation
[12, 35, 47]Image/Video2D HeatmapGrounding
[10, 43, 66]Point Cloud3D HeatmapGrounding
", + "bbox": [ + 501, + 119, + 893, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data [24, 59, 62, 68, 77, 80]. Some works utilize depth information to project image features into 3D space and then fuse them with point-wise features [18, 41, 72, 75, 76]. The above methods rely on the spatial prior information to align features, while in our task, images and point clouds are collected from distinct physical instances, there are no priors on camera pose or intrinsics, also no corresponding depth information of the image. Therefore, we align the object region features that are derived from different sources in the feature space with the assistance of the correlation invariance between affordances and appearance structures.", + "bbox": [ + 496, + 258, + 893, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 436, + 591, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 462, + 609, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to anticipate the affordance regions on the point cloud that correspond to the interaction in the image. Given a sample $\\{P, I, \\mathcal{B}, y\\}$ , where $P$ is a point cloud with the coordinates $P_{c} \\in \\mathbb{R}^{N \\times 3}$ and the affordance annotation $P_{label} \\in \\mathbb{R}^{N \\times 1}$ , $I \\in \\mathbb{R}^{3 \\times H \\times W}$ is an RGB image, $\\mathcal{B} = \\{B_{sub}, B_{obj}\\}$ denotes the bounding box of the subject and object in $I$ , and $y$ is the affordance category label. The IAG (Fig. 3) capture localized features $\\mathbf{F}_{\\mathbf{I}} \\in \\mathbb{R}^{C \\times H' \\times W'}$ and $\\mathbf{F}_{p} \\in \\mathbb{R}^{C \\times N_{p}}$ of the image and point cloud by two feature extractors ResNet [16] and PointNet++ [52]. Then, utilizing $B_{obj} / B_{sub}$ to locate the object/subject region in $\\mathbf{F}_{\\mathbf{I}}$ , outside the $B_{obj}$ and $B_{sub}$ is the scene mask $M_{sce}$ , use ROI-Align [15] to get the object, subject, and scene features $\\mathbf{F}_{i}, \\mathbf{F}_{s}, \\mathbf{F}_{e} \\in \\mathbb{R}^{C \\times H_{1} \\times W_{1}}$ , and reshape them to $\\mathbb{R}^{C \\times N_{i}}$ ( $N_{i} = H_{1} \\times W_{1}$ ). Next, the JRA module takes $\\mathbf{F}_{i}, \\mathbf{F}_{p}$ as input and jointly mines the correspondence to align them, obtaining the joint feature $\\mathbf{F}_{j}$ (Sec. 3.2). Following, the ARM module takes $\\mathbf{F}_{j}, \\mathbf{F}_{s}, \\mathbf{F}_{e}$ to reveal the affordance $\\mathbf{F}_{\\alpha}$ through cross-attention mechanism (Sec. 3.3). Ultimately, $\\mathbf{F}_{\\alpha}$ and $\\mathbf{F}_{j}$ are sent to the decoder to compute the affordance logits $\\hat{y}$ and the final 3D object affordance $\\hat{\\phi}$ (Sec. 3.4). The process is expressed as $\\hat{\\phi}, \\hat{y} = f_{\\theta}(P_{c}, I, \\mathcal{B}; \\theta)$ , where $f_{\\theta}$ denotes the network and $\\theta$ is the parameter.", + "bbox": [ + 496, + 486, + 893, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint Region Alignment Module", + "text_level": 1, + "bbox": [ + 500, + 847, + 782, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The JRA module calculates the high-dimensional dense similarity in feature space to approximate analogous shape", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10907", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4930cba0022500dd36f1fbd31252ca1d519670a9d66912c97e0111f5293fdea2.jpg", + "image_caption": [ + "Figure 3. Method. Overview of Interaction-driven 3D Affordance Grounding Network (IAG), it firstly extracts localized features $\\mathbf{F}_i, \\mathbf{F}_p$ respectively, then takes JRA (Sec. 3.2) to align them and get the joint feature $\\mathbf{F}_j$ . Next, ARM (Sec. 3.3) utilizes $\\mathbf{F}_j$ to reveal affordance $\\mathbf{F}_{\\alpha}$ with $\\mathbf{F}_s, \\mathbf{F}_e$ by cross-attention. Eventually, $\\mathbf{F}_j$ and $\\mathbf{F}_{\\alpha}$ are sent to the decoder (Sec. 3.4) to obtain the final results $\\hat{\\phi}$ and $\\hat{y}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 885, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "regions that contain relatively higher similarity [2]. Meanwhile, to map the inherent combination scheme of object components, the JRA models the inherent relations between modality-specific regions and takes it to drive the alignment of other regions by a transformer-based joint-attention $f_{\\delta}$ . With the network trained to capture the correspondences among regions from different sources, the alignment is performed implicitly during the optimization.", + "bbox": [ + 75, + 335, + 467, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, $\\mathbf{F}_p$ and $\\mathbf{F}_i$ are projected into a feature space by shared convolution layers $f_{\\sigma}$ , obtaining the features $\\mathbf{P} \\in \\mathbb{R}^{C \\times N_p}, \\mathbf{I} \\in \\mathbb{R}^{C \\times N_i}$ . Then, to correlate the analogous shape regions, the dense cross-similarity is calculated between each region of $\\mathbf{P}$ and $\\mathbf{I}$ , formulated as:", + "bbox": [ + 75, + 455, + 468, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {i, j} = \\frac {e ^ {\\left(\\mathbf {P} _ {i} , \\mathbf {I} _ {j}\\right)}}{\\sum_ {i = 1} ^ {N _ {p}} \\sum_ {j = 1} ^ {N _ {i}} e ^ {\\left(\\mathbf {P} _ {i} , \\mathbf {I} _ {j}\\right)}}, \\varphi \\in \\mathbb {R} ^ {N _ {p} \\times N _ {i}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 541, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varphi_{i,j}$ denotes the cross-similarity between the i-th region of $\\mathbf{P}$ and the j-th region of $\\mathbf{I}$ . Taking the relative difference in $\\varphi$ to refine and correlate analogous regions in $\\mathbf{P}$ and $\\mathbf{I}$ , next, applying self-attention layers $f_{i}, f_{p}$ to model the intra-structural inherent relation of objects in respective modalities, the process is expressed as:", + "bbox": [ + 75, + 590, + 468, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {P}} = f _ {p} \\left(\\mathbf {I} \\cdot \\varphi^ {T}\\right), \\bar {\\mathbf {I}} = f _ {i} \\left(\\mathbf {P} \\cdot \\varphi\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 690, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{P}}\\in \\mathbb{R}^{C\\times N_p},\\bar{\\mathbf{I}}\\in \\mathbb{R}^{C\\times N_i}$ . Following, we perform a joint-attention on features with structural relevance to map the similar inherent combination scheme and drive the alignment of the remaining regions. The process is formulated as: $\\mathbf{F}_j = f_{\\delta}[\\bar{\\mathbf{P}},\\bar{\\mathbf{I}}]$ , where $\\mathbf{F}_j\\in \\mathbb{R}^{C\\times (N_p + N_i)}$ and $[\\cdot ]$ denotes joining the image feature sequence and point cloud feature sequence into a whole one, and $\\mathbf{F}_j$ denotes the joint representation. The $\\mathbf{F}_j$ is sent to the affordance revealed module to excavate interaction contexts with $\\mathbf{F}_s,\\mathbf{F}_e$ . And the affordance knowledge revealed by the interaction contexts could mutually optimize the alignment process during training (see Sec. 3.4).", + "bbox": [ + 75, + 718, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Affordance Revealed Module", + "text_level": 1, + "bbox": [ + 500, + 334, + 759, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, the dynamic factors related to affordance are reflected in the interaction, affecting the extraction of affordance representation, and the interaction is mostly considered to exist between humans and objects. However, the interaction between objects and scenes (including other objects) may also affect affordance [43]. To consider both of these factors, the ARM module utilizes the cross-attention technique to extract these interaction contexts respectively and jointly embed them to reveal explicit object affordance.", + "bbox": [ + 496, + 358, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, $\\mathbf{F}_j$ is projected to form the shared query $\\mathbf{Q} = \\mathbf{F}_j\\mathbf{W}_1,\\mathbf{F}_s$ and $\\mathbf{F}_e$ are projected to form different keys and values $\\mathbf{K}_{1 / 2} = \\mathbf{F}_s\\mathbf{W}_{2 / 3},\\mathbf{V}_{1 / 2} = \\mathbf{F}_e\\mathbf{W}_{4 / 5}$ where $\\mathbf{W}_{1\\sim 5}$ are projection weights. Then cross-attention aggregates them to excavate interaction contexts, expressed as:", + "bbox": [ + 498, + 494, + 890, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Theta} _ {1 / 2} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} ^ {T} \\cdot \\mathbf {K} _ {1 / 2} / \\sqrt {d}\\right) \\cdot \\mathbf {V} _ {1 / 2} ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 582, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q} \\in \\mathbb{R}^{d \\times (N_p + N_i)}$ , $\\mathbf{K}_{1/2}$ , $\\mathbf{V}_{1/2} \\in \\mathbb{R}^{d \\times N_i}$ , $d$ is the dimension of projection, $\\Theta_1, \\Theta_2$ indicate the excavated interaction contexts. Then, jointly embedding $\\Theta_1$ and $\\Theta_2$ to reveal affordance, expressed as:", + "bbox": [ + 496, + 607, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\alpha} = f _ {\\xi} (\\boldsymbol {\\Theta} _ {1}, \\boldsymbol {\\Theta} _ {2}), \\mathbf {F} _ {\\alpha} \\in \\mathbb {R} ^ {C \\times (N _ {p} + N _ {i})}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 676, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{\\alpha}$ is the joint affordance representation, it is split in the decoder to compute final results, $f_{\\xi}$ denotes the concatenation, followed by a convolution layer to fuse the feature.", + "bbox": [ + 496, + 705, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Decoder and Loss Function", + "text_level": 1, + "bbox": [ + 500, + 758, + 745, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\mathbf{F}_j$ is split to $\\hat{\\mathbf{F}}_p\\in \\mathbb{R}^{C\\times N_p}$ and $\\hat{\\mathbf{F}}_i\\in \\mathbb{R}^{C\\times N_i}$ , $\\mathbf{F}_{\\alpha}$ is split to $\\mathbf{F}_{p\\alpha}\\in \\mathbb{R}^{C\\times N_p}$ and $\\mathbf{F}_{i\\alpha}\\in \\mathbb{R}^{C\\times N_i}$ in the decoder. Pooling $\\mathbf{F}_{p\\alpha}$ and $\\mathbf{F}_{i\\alpha}$ respectively and concatenate them to compute the affordance logits $\\hat{y}$ . $\\hat{\\mathbf{F}}_p$ is upsampled to $\\mathbb{R}^{C\\times N}$ by Feature Propagation Layers (FP) [52], and the 3D object affordance $\\hat{\\phi}$ is computed as:", + "bbox": [ + 496, + 780, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\phi} = f _ {\\phi} \\left(\\mathbf {F P} \\left(\\hat {\\mathbf {F}} _ {p}\\right) \\odot \\Gamma \\left(\\mathbf {F} _ {p \\alpha}\\right)\\right), \\hat {\\phi} \\in \\mathbb {R} ^ {N \\times 1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 882, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10908", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ae5d64006c44d4a3924455e449838133da73909bd8fbf5077308aa609f61a97.jpg", + "image_caption": [ + "Figure 4. Properties of the PIAD dataset. (a) Data pairs in the PIAD, the red region in point clouds is the affordance annotation. (b) Distribution of the image data. The horizontal axis represents the category of affordance, the vertical axis represents quantity, and different colors represent different objects. (c) Distribution of the point cloud data. (d) The ratio of images and point clouds in each affordance class. It shows that images and point clouds are not fixed one-to-one pairing, they can form multiple pairs. Please refer to supp. for more details." + ], + "image_footnote": [], + "bbox": [ + 106, + 74, + 867, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\phi}$ is an output head, $\\mathbf{FP}$ is the upsample layer, $\\Gamma$ denotes pooling $\\mathbf{F}_{p\\alpha}$ and expand it to the shape of $\\mathbb{R}^{C\\times N}$ .", + "bbox": [ + 75, + 292, + 467, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss comprises three parts: $\\mathcal{L}_{CE}$ , $\\mathcal{L}_{KL}$ , $\\mathcal{L}_{HM}$ . Where $\\mathcal{L}_{CE}$ computes the cross-entropy loss between $y$ and $\\hat{y}$ , it supervises the extraction of $\\mathbf{F}_{\\alpha}$ and implicitly optimizes the alignment process. To enable the network focus on the alignment of affordance regions, we apply the KL Divergence (KLD) [3] to constrain the distribution between $\\mathbf{F}_{i\\alpha}$ and $\\hat{\\mathbf{F}}_i$ , formulated as $\\mathcal{L}_{KL} = (\\mathbf{F}_{i\\alpha}||\\hat{\\mathbf{F}}_i)$ . The reason is that $\\mathbf{F}_{i\\alpha}$ exhibits the affordance distribution of each object region in the image, and the affordance-related regions keep more significant features. Constraining the feature distribution of $\\hat{\\mathbf{F}}_i$ to enhance the affordance region features in $\\hat{\\mathbf{F}}_i$ , and with the region correlations established by the alignment process, $\\hat{\\mathbf{F}}_p$ also tends to exhibit this property, similar to distillation [19, 71]. Which makes the alignment and affordance extraction optimize mutually. $\\mathcal{L}_{HM}$ is a focal loss [28] combined with a dice loss [40], it is calculated by $\\hat{\\phi}$ and $P_{label}$ , which supervise the point-wise heatmap on point clouds. Eventually, the total loss is formulated as:", + "bbox": [ + 75, + 323, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {C E} + \\lambda_ {2} \\mathcal {L} _ {K L} + \\lambda_ {3} \\mathcal {L} _ {H M}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 611, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1},\\lambda_{2}$ and $\\lambda_3$ are hyper-parameters to balance the total loss. See more details in supplementary materials.", + "bbox": [ + 75, + 633, + 468, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dataset", + "text_level": 1, + "bbox": [ + 76, + 678, + 165, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collection Details. We collect Point-Image Affordance Dataset (PIAD), which contains paired image-point cloud affordance data. The point clouds are mainly collected from 3D-AffordanceNet [10], including the point cloud coordinates and affordance annotation. Images are mainly collected from HICO [4], AGD20K [35], and websites with free licenses. The collection criteria is that the image should demonstrate interactions that the object in the point cloud could afford. For example, if the object point cloud is a \"Chair\", it affords \"Sit\", then the image should depict a subject (usually humans) sitting on a chair. The final dataset comprises 7012 point clouds and 5162 images, spanning 23 object classes and 17 affordance categories. Notably, ob", + "bbox": [ + 75, + 703, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "jects in the images and point clouds do not sample from the same physical instance, but they belong to the same object category. Paired examples are shown in Fig. 4 (a).", + "bbox": [ + 496, + 292, + 890, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotation Details. For point clouds, each point may afford one or multiple interactions, e.g. one annotation is a matrix of (2048, 17), 2048 is the point number, 17 represents the number of affordance types. Each element in the matrix indicates the probability of a point affording a specific affordance. Meanwhile, images are annotated with bounding boxes of the interactive subject and object, as well as an affordance category label. In our task, we only use the heatmap in the point cloud annotation that corresponds to the affordance of the image for training, resulting in a matrix of (2048, 1). By doing so, the affordance category is detached from the point cloud during inference, and the anticipation of the 3D object affordance category only relies on 2D interactions. As a result, different affordances can be anticipated on a point cloud through distinct interactions.", + "bbox": [ + 496, + 339, + 892, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Statistic Analysis. Since images and point clouds are sampled from different instances, they do not need a fixed one-to-one pairing, one image could be paired with multiple point clouds and the count of them is not strictly consistent. Fig. 4 (b) and (c) show the count and distribution of affordances in images and point clouds. Fig. 4 (d) illustrates the ratio of images and point clouds in each affordance category. PIAD has two partitions: Seen and Unseen. In Seen, both objects and affordances in the training and testing sets are consistent, while in Unseen, some objects or affordances in the testing set do not exist in the training set.", + "bbox": [ + 496, + 565, + 893, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 744, + 633, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Benchmark Setting", + "text_level": 1, + "bbox": [ + 500, + 770, + 684, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To provide a comprehensive and effective evaluation, we compare serval advanced works in the affordance area [10, 47, 74] and finally chose four evaluation metrics: AUC [31], aIOU [53], SIMilarity [57] and Mean Absolute Error [65] to benchmark the PIAD.", + "bbox": [ + 496, + 794, + 890, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modular Baselines. Since there are no prior works using paired image-point cloud data to ground 3D object", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10909", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c464df2efdb1d25281ad68f76789123161f837fc7201f233bea0d54bbac8e98b.jpg", + "table_caption": [ + "Table 2. Comparison Results on PIAD. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. \"Green\" and \"Orange\" indicate two types of the comparative modular baselines. \"Base.\" represents the baseline. $\\diamond$ denotes the relative improvement of our method over other methods. AUC and aIOU are shown in percentage." + ], + "table_footnote": [], + "table_body": "
MetricsBase.MBDF [59]PMF [80]FRCNN [68]ILN [5]PFusion [67]XMF [1]Ours
SeenAUC ↑68.4974.88◇13.3%75.05◇13.0%76.05◇11.5%75.84◇11.9%77.50◇9.5%78.24◇8.4%84.85 ± 0.3
aIOU ↑6.859.34◇119.5%10.13◇102.4%11.97◇71.3%11.52◇78.0%12.31◇66.6%12.94◇58.5%20.51 ± 0.7
SIM ↑0.3670.415◇31.3%0.425◇28.2%0.429◇27.0%0.427◇27.6%0.432◇26.1%0.441◇23.6%0.545 ± 0.02
MAE ↓0.1520.143◇31.4%0.141◇30.5%0.136◇27.9%0.137◇28.4%0.135◇27.4%0.127◇22.8%0.098 ± 0.01
UnseenAUC ↑57.3458.23◇23.3%60.25◇19.2%61.92◇16.0%59.69◇20.3%61.87◇16.1%62.58◇14.8%71.84 ± 1.8
aIOU ↑3.954.22◇88.4%4.67◇70.2%5.12◇55.2%4.71◇68.8%5.33◇49.1%5.68◇39.9%7.95 ± 0.8
SIM ↑0.3180.325◇8.3%0.330◇6.6%0.332◇6.0%0.325◇8.3%0.330◇6.6%0.342◇2.9%0.352 ± 0.03
MAE ↓0.2350.213◇40.3%0.211◇39.8%0.195◇34.8%0.207◇38.6%0.193◇34.2%0.188◇32.4%0.127 ± 0.01
", + "bbox": [ + 78, + 133, + 897, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9e82174288403b9f88e23bcf64ecb47e40ce83150a93739f94932c250539725.jpg", + "image_caption": [ + "Figure 5. Paradigm Comparison. (a) Ground truth. (b) The results of 3D-AffordanceNet [10]. (c) Our results. The top row shows the regional confusion case in Seen, and the bottom row displays the generalization in Unseen." + ], + "image_footnote": [], + "bbox": [ + 81, + 297, + 454, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "affordance. For a thorough comparison of our method, we select several advanced image-point cloud cross-modal works as modular baselines. Methods of comparison are broadly divided into two types, one is that utilizes camera intrinsics to facilitate cross-modal feature alignment or fusion, i.e. MBDF-Net (MBDF) [59], PMF [80] and FusionRCNN (FRCNN) [68], for this type of method, we remove the step of using intrinsic parameters to align raw data or features to explore their effectiveness on PIAD. Another type performs feature alignment directly in the feature space, without relying on camera intrinsic, i.e. ImLoveNet (ILN) [5], PointFusion (PFusion) [67] and XMFnet (XMF) [1]. To ensure a fair comparison, all methods use the same feature extractors, with the only variation coming from cross-modal alignment or fusion block. For the Baseline, we directly concatenate the features that are output from extractors, regarding the concatenation as the cross-fusion block. More details about baselines are provided in the supplementary materials.", + "bbox": [ + 75, + 534, + 472, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison Results", + "text_level": 1, + "bbox": [ + 76, + 830, + 269, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The comparison results of evaluation metrics are shown in Tab. 2. As can be seen, our method outperforms the compared baselines, across all metrics in both partitions.", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1d527988452bc1ae190a3567a219a8670d7005293fce4ca17e52d0a99dd984d7.jpg", + "table_caption": [ + "Table 3. Ablation Study. We investigate the improvement of JRA and ARM on the model performance based on the baseline." + ], + "table_footnote": [], + "table_body": "
JRAARMAUCaIOUSIMMAE
Seen69.928.850.4270.132
✓80.2914.310.4950.121
✓78.6713.950.4750.126
✓✓85.1621.200.5640.088
Unseen59.144.050.3380.202
✓66.256.270.3630.159
✓65.795.990.3580.162
✓✓73.698.700.3830.117
", + "bbox": [ + 503, + 335, + 888, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, to display the limitation of methods that lock geometrics with specific semantic affordance categories, we conduct an experiment to compare one of these methods [10] with ours. As shown in Fig. 5, the top row indicates that the result of this method exhibits regional confusion, e.g. the region where the chair could be moved or sat is geometrically rectangular, and it directly anticipates results on these similar geometries, inconsistent with affordance property. Plus, the bottom row shows that directly establishing a link between object structures and affordances may fail to anticipate correct 3D affordance in Unseen. In contrast, our method anticipates precise results by mining affordance clues provided by 2D interactions. Additionally, visual comparative results of our method and other baselines are shown in Fig. 6. As can be seen, the comparative baselines could anticipate some 3D object affordance under our setting, but in comparison, our method obviously achieves better results, which validates the rationality of our setting, and also demonstrates the superiority of our method.", + "bbox": [ + 496, + 513, + 890, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 815, + 653, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of JRA. Tab. 3 reports the impact on evaluation metrics of JRA. Fig. 7 provides visual comparison results. It shows that without the JRA, the result is anticipated over the entire region that contains the interactive compo", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10910", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a151ae1999e8e56175ae92d6788ee6eff41373c50d0619773e3c11b280511ed6.jpg", + "image_caption": [ + "Figure 6. Visualization Results. The first row is the interactive image, which demonstrates the interaction that the object can afford. The last row is the ground truth of 3D object affordance in the point cloud. The four columns on the left are the visual comparison results for distinct 3D object affordances in the Seen partition. The four columns on the right are the results in the Unseen partition." + ], + "image_footnote": [], + "bbox": [ + 106, + 74, + 478, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2af1ca3d6a1399f4edc9c3898c54d77582ba08f36e6f0de79cd9fdda39af7dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 74, + 848, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/475f80aafcadb50818b19301c1c3f06561aa57ef916c17361447c7da55397623.jpg", + "image_caption": [ + "Figure 7. Ablation of JRA. The top row is the result of the anticipated affordance with/without JRA. The bottom row is a part of the cross-similarity matrix of a sample during the training process." + ], + "image_footnote": [], + "bbox": [ + 94, + 517, + 450, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nents, which means the 2D-3D affordance regions of objects do not match well. Besides, we visualize a part of the cross-similarity matrix between $\\hat{\\mathbf{F}}_p$ and $\\hat{\\mathbf{F}}_i$ in Fig. 7. In the initial stage, only a few analogous shape regions keep explicit correspondence, as training proceeds, the JRA maps the correlation among regions with lower similarity. Meanwhile, the model extracts explicit affordance progressively, and the affordance introduced into the optimization process reveals the corresponding affordance regions.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/112e6f64757112bf60e8460adea407ead05a29f6fe2b90b12f89e5c05d05fc26.jpg", + "image_caption": [ + "Figure 8. Ablation of ARM. The activation map and t-SNE [61] results with $(\\boldsymbol{w})$ , without $(\\boldsymbol{w} / \\boldsymbol{o})$ ARM in both partitions." + ], + "image_footnote": [], + "bbox": [ + 521, + 515, + 867, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of ARM. The influence of ARM on evaluation metrics is also shown in Tab. 3. To visually evaluate the effectiveness of ARM, we employ GradCAM [55] to generate the activation map, and utilize t-SNE [61] to demonstrate the clustering of affordances in both partitions. The results are shown in Fig. 8. It shows that ARM makes the model focus on the interactive regions to excavate the interaction contexts, and enables the model to differentiate various affordances from interactions in both partitions.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10911", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71aaa6b56d0d763a90f12a62142c759cc862dc7934c3ed2e7060ac825eef7fdd.jpg", + "image_caption": [ + "Figure 9. Same Image w.r.t. Multiple Point Clouds. (a) Same object category. (b) Different object categories, similar geometrics. (c) Different object categories and geometrics." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 433, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95750804648c10db1e24627c1aed71739502066ba80d481cb494b2d8b2703b8b.jpg", + "image_caption": [ + "Figure 10. Same Point Cloud w.r.t. Multiple Images. Grounding affordance on the same point cloud with images that contain similar or disparate interactions." + ], + "image_footnote": [], + "bbox": [ + 94, + 268, + 452, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Performance Analysis", + "text_level": 1, + "bbox": [ + 76, + 469, + 282, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Instances. We conduct experiments to verify whether the model could ground 3D affordance on instances from different sources: (i) using an image and different point clouds to infer respectively (Fig. 9 (a)). (ii) using multiple images and a single point cloud (Fig. 10). The results indicate that the model can anticipate 3D affordances on different instances with the same 2D interaction, and can also anticipate distinct 3D affordances on the same point cloud through different 2D interactions. Showing that the model maintains the mapping between object structures and affordances by learning from different 2D interactions.", + "bbox": [ + 75, + 492, + 468, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Object Categories. What will happen if the object category is different in the image and point cloud? To explore this issue, we perform experiments with mismatched object categories, shown in Fig. 9 (b) and (c). When objects are of different categories but with similar geometric primitives, e.g. the handle of \"Mug\" and the shoulder strap of \"Bag\", our model still properly infer the 3D affordance. However, when geometric structures are also dissimilar (Fig. 9 (c)), the model does not make random predictions. This indicates that the model maps cross-category invariance between affordance and geometries, which can be generalized to new instances of geometry.", + "bbox": [ + 75, + 657, + 468, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiplicity. One defining property of affordance is multiplicity: some points may correspond to multiple affordances. Fig. 11 shows that the model makes objects' multiple affordances compatible, and verifies the model does not", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40d0e1faa8821e7386c55c5b4ade70c48db66c317502d8b8d726aca61c1320c7.jpg", + "image_caption": [ + "Figure 11. Multiple Affordances. Some objects like \"Bag\" contains the region that corresponds to multiple affordances." + ], + "image_footnote": [], + "bbox": [ + 506, + 85, + 883, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47fb743a30689270c03b964f9a1639fbd1ec23e5296161b32bab46c2e57f20cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 186, + 879, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74028dde10d445c71a2f715e6dc0635dc3b1e4f77e03f0c50c68de6d82c75422.jpg", + "image_caption": [ + "Figure 12. Real-Wrold. The first row is scanned by iPhone, the second row comes from [29], point clouds are sampled from them.", + "Figure 13. Failure Cases. Over prediction in the small regions." + ], + "image_footnote": [], + "bbox": [ + 514, + 344, + 874, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "anticipate affordance by linking a region with specific affordances. Instead, it predicts by considering the presence of affordance-related interactions in the same region.", + "bbox": [ + 498, + 465, + 890, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real World. To validate the model in real-world scenarios, we use an iPhone to scan objects and real-scan dataset [29] to test our model, shown in Fig. 12. It shows that our model exhibits a certain degree of generalization to the real world. Limitations. Our model exhibits over predictions for some small affordance regions, shown in Fig. 13. This may be due to the limited understanding of fine-grained geometric parts, and aligning the object features in a single scale may lead to overly large receptive fields. Our feature extension will refer [13, 27] to tackle this problem.", + "bbox": [ + 498, + 511, + 892, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 617, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel setting for affordance grounding, which utilizes the 2D interactive semantics to guide the grounding of 3D object affordance, it has the potential to serve embodied systems when collaborating with multimodal grounding systems [30, 58, 70]. Besides, We collect the PIAD dataset as the first test bed for the proposed setting, it contains paired image-point cloud affordance data. Plus, we propose a novel framework to correlate affordance regions of objects that are from different sources and model interactive contexts to ground 3D object affordance. Comprehensive experiments on PIAD display the reliability of the setting, and we believe it could offer fresh insights and facilitate research in the affordance area.", + "bbox": [ + 496, + 704, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10912", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002.", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 162, + 171, + 178 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. arXiv preprint arXiv:2209.09552, 2022. 6, 7", + "[2] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3, 4", + "[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE transactions on pattern analysis and machine intelligence, 41(3):740-757, 2018. 5", + "[4] Yu-Wei Chao, Zhan Wang, Yugeng He, Jiaxuan Wang, and Jia Deng. Hico: A benchmark for recognizing human-object interactions in images. In Proceedings of the IEEE international conference on computer vision, pages 1017–1025, 2015. 5", + "[5] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. Imlovenet: Misaligned image-supported registration network for low-overlap point cloud pairs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 6", + "[6] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22(4):449-462, 2013. 1", + "[7] Ching-Yao Chuang, Jiaman Li, Antonio Torralba, and Sanja Fidler. Learning to act properly: Predicting and explaining affordances from images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 975–983, 2018. 3", + "[8] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2):722-739, 2021. 3", + "[9] Barney Dalgarno and Mark JW Lee. What are the learning affordances of 3-d virtual environments? British Journal of Educational Technology, 41(1):10-32, 2010. 1", + "[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 1, 3, 5, 6", + "[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3", + "[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3" + ], + "bbox": [ + 78, + 188, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, June 2023. 3, 8", + "[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[17] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1", + "[18] Maximilian Jaritz, Jiayuan Gu, and Hao Su. Multi-view pointnet for 3d scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3", + "[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 5", + "[20] David Inkyu Kim and Gaurav S Sukhatme. Semantic labeling of 3d point clouds with object affordance for robot manipulation. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 5578-5584. IEEE, 2014. 1", + "[21] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 1, 3", + "[22] Hema S Koppula and Ashutosh Saxena. Physically grounded spatio-temporal object affordances. In European Conference on Computer Vision, pages 831-847. Springer, 2014. 3", + "[23] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 3", + "[24] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 3", + "[25] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhi-hong Zhu, and Yuexian Zou. G21: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3", + "[26] Yaowei Li, Bang Yang, Xuxin Cheng, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. Unify, align and refine: Multi-level semantic alignment for radiology report generation. In" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10913", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3", + "[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 8", + "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[29] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, June 2022. 3, 8", + "[30] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 8", + "[31] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5", + "[32] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Transactions on Artificial Intelligence, 2022. 3", + "[33] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot affordance detection. arXiv preprint arXiv:2106.14747, 2021. 3", + "[34] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Grounded affordance from exocentric view. arXiv preprint arXiv:2208.13196, 2022. 3", + "[35] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2252-2261, 2022. 3, 5", + "[36] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 1", + "[37] Maurice Merleau-Ponty. Phenomenologie de la perception (1945). Libraqire Gallimard, Paris, 1976. 2", + "[38] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers in Neurorobotics, 14:26, 2020. 3", + "[39] Jinpeng Mi, Song Tang, Zhen Deng, Michael Goerner, and Jianwei Zhang. Object affordance based multimodal fusion for natural human-robot interaction. Cognitive Systems Research, 54:128–137, 2019. 3", + "[40] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5", + "[41] Niluthpol Chowdhury Mithun, Karan Sikka, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Rgb2lidar: Towards solving large-scale cross-modal visual localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 934-954, 2020. 3", + "[42] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 3", + "[43] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3, 4", + "[44] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3", + "[45] Bogdan Moldovan, Plinio Moreno, Martijn Van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE international conference on robotics and automation, pages 4373-4378. IEEE, 2012. 1", + "[46] Austin Myers, Ching L Teo, Cornelia Fermüller, and Yiannis Aloimonos. Affordance detection of tool parts from geometric features. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 1374-1381. IEEE, 2015. 1", + "[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 5", + "[48] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3", + "[49] Anh Nguyen, Dimitrios Kanoulas, Darwin G Caldwell, and Nikos G Tsagarakis. Detecting object affordances with convolutional neural networks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2765-2770. IEEE, 2016. 3", + "[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 7(1-2):1-179, 2018. 1", + "[51] François Osiurak and Arnaud Badets. Tool use and affordance: Manipulation-based versus reasoning-based approaches. Psychological review, 123(5):534, 2016. 1", + "[52] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3, 4" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10914", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Md Atiqur Rahman and Yang Wang. Optimizing intersection-over-union in deep neural networks for image segmentation. In International symposium on visual computing, pages 234–244. Springer, 2016. 5", + "[54] Anirban Roy and Sinisa Todorovic. A multi-scale cnn for affordance segmentation in rgb images. In European conference on computer vision, pages 186-201. Springer, 2016. 3", + "[55] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7", + "[56] Chris Shilling. The body in culture, technology and society. Sage, 2004. 2", + "[57] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991. 5", + "[58] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.8", + "[59] Xun Tan, Xingyu Chen, Guowei Zhang, Jishiyu Ding, and Xuguang Lan. Mbdf-net: Multi-branch deep fusion network for 3d object detection. In Proceedings of the 1st International Workshop on Multimedia Computing for Urban Data, pages 9–17, 2021. 3, 6", + "[60] Spyridon Thermos, Petros Daras, and Gerasimos Potamianos. A deep learning approach to object affordance segmentation. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2358-2362. IEEE, 2020. 3", + "[61] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7", + "[62] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 3", + "[63] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In European Conference on Computer Vision, pages 421-436. Springer, 2014. 1", + "[64] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 90-107. Springer, 2022. 3", + "[65] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5", + "[66] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022. 1, 3", + "[67] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 6", + "[68] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 3, 6, 7", + "[69] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement, 2023. 3", + "[70] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 8", + "[71] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73-91. Springer, 2022. 5", + "[72] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 3", + "[73] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 2", + "[74] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 5", + "[75] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4534-4543, 2020. 3", + "[76] Linqing Zhao, Jiwen Lu, and Jie Zhou. Similarity-aware fusion network for 3d semantic segmentation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1585-1592. IEEE, 2021. 3", + "[77] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 3", + "[78] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Computing and Applications, 32(18):14321-14333, 2020. 3", + "[79] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022. 3", + "[80] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10915", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_model.json b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1832e2c0870b60a4dafd03614d57b2bccc3566d0 --- /dev/null +++ b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "Grounding 3D Object Affordance from 2D Interactions in Images" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.18, + 0.827, + 0.234 + ], + "angle": 0, + "content": "Yuhang Yang\\(^{1}\\), Wei Zhai\\(^{1,*}\\), Hongchen Luo\\(^{1}\\), Yang Cao\\(^{1,2}\\), Jiebo Luo\\(^{3}\\), Zheng-Jun Zha\\(^{1}\\) \n\\(^{1}\\) University of Science and Technology of China \n\\(^{3}\\) University of Rochester \n\\(^{2}\\) Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.237, + 0.773, + 0.253 + ], + "angle": 0, + "content": "yyuhang@mail., wzhai056@mail., lhcl2@mail., forrest@ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.256, + 0.672, + 0.269 + ], + "angle": 0, + "content": "jluo@cs.rochester.edu zhazj@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.319, + 0.314, + 0.335 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Grounding 3D object affordance seeks to locate objects' \"action possibilities\" regions in the 3D space, which serves as a link between perception and operation for embodied agents. Existing studies primarily focus on connecting visual affordances with geometry structures, e.g., relying on annotations to declare interactive regions of interest on the object and establishing a mapping between the regions and affordances. However, the essence of learning object affordance is to understand how to use it, and the manner that detaches interactions is limited in generalization. Normally, humans possess the ability to perceive object affordances in the physical world through demonstration images or videos. Motivated by this, we introduce a novel task setting: grounding 3D object affordance from 2D interactions in images, which faces the challenge of anticipating affordance through interactions of different sources. To address this problem, we devise a novel Interaction-driven 3D Affordance Grounding Network (IAG), which aligns the region feature of objects from different sources and models the interactive contexts for 3D object affordance grounding. Besides, we collect a Point-Image Affordance Dataset (PIAD) to support the proposed task. Comprehensive experiments on PIAD demonstrate the reliability of the proposed task and the superiority of our method. The project is available at https://github.com/yyvhang/IAGNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.471, + 0.877 + ], + "angle": 0, + "content": "The term \"affordance\" is described as \"opportunities of interaction\" by J. Gibson [14]. Grounding 3D object affordance aims to comprehend the interactive regions of objects in 3D space, which is not only simply to predict which interaction an object affords, but also to identify specific points on the object that could support the interaction. It consti" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.303, + 0.876, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Figure 1. Grounding Affordance from Interactions. We propose to ground 3D object affordance through 2D interactions. Given an object point cloud with an interactive image, grounding the corresponding affordance on the 3D object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.629 + ], + "angle": 0, + "content": "tutes a link between perception and operation for embodied agents, which has the potential to serve numerous practical applications, e.g. action prediction [21, 63], robot manipulation [36, 45, 51], imitation learning [17, 50], and augmented/virtual reality [6, 9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.902 + ], + "angle": 0, + "content": "So far, the paradigm of perceiving 3D object affordance has several branches. One of them involves establishing an explicit mapping between affordance categories and geometry structures [10, 20, 46, 66], based on visual appearance. However, affordance is dynamic and multiple, these geometric-specific manners have limited generalization for unseen structures. Besides, locking the geometry with a specific affordance category may lead to the anticipated region being inconsistent with its affordance when objects possess multiple similar geometrics, resulting in affordance regional confusion. Another paradigm is based on reinforcement learning, which puts the agent in 3D synthetic scenarios to interact with several objects actively, taking the reward mechanism to optimize the whole process [48]. While this type of approach transitions agents from passive recognition to active reasoning, it needs repeated attempts in a huge search space when meeting a novel structure, and is therefore time-consuming." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.234, + 0.901 + ], + "angle": 0, + "content": "\\*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10905" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.089, + 0.452, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.09, + 0.838, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Motivation. (a) Clues to correlate regions of objects from different sources. Objects with the same category possess a similar combination scheme to meet certain affordance properties, it is a similar inherent relation among different instances. And some structures hint at these properties exhibit closer representations with higher similarity in the latent space. (b) Object affordance may be affected by dynamic factors, such as the position of the object itself in the scene, other objects in the scene, and the body part of the interacting subject. These factors can be decomposed into object-subject and object-scene interaction contexts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.584 + ], + "angle": 0, + "content": "These limitations motivate us to explore an affordance-compatible learning paradigm. Typically, humans can infer object affordance in the 3D physical world by watching images or videos that demonstrate the interactions. Some studies in cognitive science [37, 56] point out the existence of \"body image\" in human cognition, which claims that humans have a perceptual experience of the objects they see, thus facilitating their ability to organize the perception [73] and operate novel objects. Hence, human-object interaction conveys the perception that object structure could perform certain affordance, which is a crucial clue to reason object affordance. In light of this, we present a novel task setting: grounding 3D object affordance from 2D interactions in images, which is shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.471, + 0.902 + ], + "angle": 0, + "content": "This challenging task includes several essential issues that should be properly addressed. 1) Alignment ambiguity. To ground 3D object affordance from 2D source interactions, the premise is to correspond the regions of the object in different sources. The object in 2D demonstrations and the 3D object we face are usually derived from different physical instances in different locations and scales. This discrepancy may lead to confusion when matching the affordance regions, causing alignment ambiguity. While objects are commonly designed to satisfy certain needs of human beings, so the same category generally follows a similar combination scheme of object components to meet certain affordances, and these affordances are hinted at by some structures (Fig. 2 (a)). These invariant properties are across instances and could be utilized to correlate object regions from different sources. 2) Affordance ambiguity. Affordance has properties of dynamic and multiplicity, which means the object affordance may change according to the situation, the same part of an object could afford multiple interactions, as shown in Fig. 2 (b), \"Chair\" affords \"Sit\" or \"Move\" depends on the human actions, \"Mug\" af" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.463 + ], + "angle": 0, + "content": "fords \"Wrapgrasp\" or \"Contain\" according to the scene context, these properties may make ambiguity when extracting affordance. However, these dynamic factors for affordance extraction can be decomposed into the interaction between subject-object and object-scene. Modeling these interactions is possible to extract explicit affordance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.735 + ], + "angle": 0, + "content": "To address these issues, we propose the Interaction-driven 3D Affordance Grounding Network (IAG) to align object region features from different sources and model interaction contexts to reveal affordance. In detail, it contains two sub-modules, one is Joint Region Alignment Module (JRA), which is devised to eliminate alignment ambiguity. It takes the relative difference in dense cross-similarity to refine analogous shape regions and employs learnable layers to map the invariant combination scheme, taking them to match local regions of objects. For affordance ambiguity, the other one Affordance Revealed Module (ARM) takes the object representation as a shared factor and jointly models its interaction contexts with the affordance-related factors to reveal explicit affordance. Moreover, we collect Point-Image Affordance Dataset (PIAD) that contains plenty of paired image-point cloud affordance data and make a benchmark to support the model training and evaluation on the proposed setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The contributions are summarized as follows: 1) We introduce grounding 3D object affordance through the 2D interactions, which facilitates the generalization to 3D object affordance perception. 2) We propose the IAG framework, which aligns the region feature of objects from different sources and jointly models affordance-related interactions to locate the 3D object affordance. 3) We collect a dataset named PIAD to support the proposed task setting, which contains paired image-point cloud affordance data. Besides, we establish a benchmark on PIAD, and the experiments on PIAD exhibit the reliability of the method and the setting." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10906" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.228, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.277, + 0.133 + ], + "angle": 0, + "content": "2.1. Affordance Learning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.141, + 0.473, + 0.701 + ], + "angle": 0, + "content": "At present, the affordance area has made achievements in multiple tasks (Tab. 1). Some works are devoted to detecting the affordance region from 2D sources i.e. images and videos [7, 11, 25, 33, 54, 60, 78], and there are also some studies accomplish this task with the assistance of natural language [32, 38, 39]. They seek to detect or segment objects that afford the action in 2D data, but cannot perceive the specific parts corresponding to the affordance. Thus, another type of method brings a leap to locate the affordance region of objects from 2D sources [12, 35, 47, 34]. However, the affordance knowledge derived from 2D sources is hard to extrapolate specific interactive locations of objects in the 3D environment. With several 3D object datasets proposed [13, 29, 44], some researchers explore grounding object affordance from 3D data [10, 43, 66, 48]. Such methods directly map the connection between semantic affordances and 3D structures, detach the real interaction, and may deter the generalization ability: structures that do not map to a specific affordance are usually hard to generalize through this type of method. T. Nagarajan et al. [48] give a fresh mind, taking the reinforcement learning to make agents actively interact with the 3D synthetic scenarios, while it requires repeated attempts by the agents in a given search space, and is time-consuming. In robotics, affordance is utilized to provide priors for object manipulation and achieve considerable results, especially the articulated 3D objects [42, 64, 79]. Several methods utilize the 2.5D data e.g. RGB-D image to understand the object affordance and serve the manipulation task [21, 22, 23, 49], for these methods, the image and depth information are corresponding in spatial and need to be collected in the same scene. In contrast, our task focuses on grounding 3D object affordance from 2D interactions, in which the 2D interactions provide direct clues to excavate object affordance efficiently and make the affordance could generalize to some unseen structures, and the 2D-3D data is collected from different sources, freeing the constraints on spatial correspondence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.441, + 0.728 + ], + "angle": 0, + "content": "2.2. Image-Point Cloud Cross-Modal Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The combination of point cloud and image data can capture both semantic and geometric information, enabling cross-modal learning among them has great potential application value in scenarios like autopilot [8]. For this type of work, aligning features is the premise for completing downstream tasks e.g. detection, segmentation, and registration. It is aimed at establishing correspondences between instances from different modalities [2, 26, 69], either spatial or semantic. To achieve this, many methods use camera intrinsic to correspond spatial position of pixels and points, then align per pixel-point feature or fuse the raw" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.894, + 0.116 + ], + "angle": 0, + "content": "Table 1. Affordance Learning. Various works for several tasks in the affordance community." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.121, + 0.895, + 0.235 + ], + "angle": 0, + "content": "
WorkInputOutputTask
[11, 60, 78]Image/Video2D MaskDetection
[32, 38, 39]Image,Text2D MaskDetection
[21, 23, 49]RGBDHeatmap & ActionManipulation
[42, 64, 79]Point CloudHeatmap & ActionManipulation
[12, 35, 47]Image/Video2D HeatmapGrounding
[10, 43, 66]Point Cloud3D HeatmapGrounding
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.895, + 0.427 + ], + "angle": 0, + "content": "data [24, 59, 62, 68, 77, 80]. Some works utilize depth information to project image features into 3D space and then fuse them with point-wise features [18, 41, 72, 75, 76]. The above methods rely on the spatial prior information to align features, while in our task, images and point clouds are collected from distinct physical instances, there are no priors on camera pose or intrinsics, also no corresponding depth information of the image. Therefore, we align the object region features that are derived from different sources in the feature space with the assistance of the correlation invariance between affordances and appearance structures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.593, + 0.453 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.611, + 0.478 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Our goal is to anticipate the affordance regions on the point cloud that correspond to the interaction in the image. Given a sample \\(\\{P, I, \\mathcal{B}, y\\}\\), where \\(P\\) is a point cloud with the coordinates \\(P_{c} \\in \\mathbb{R}^{N \\times 3}\\) and the affordance annotation \\(P_{label} \\in \\mathbb{R}^{N \\times 1}\\), \\(I \\in \\mathbb{R}^{3 \\times H \\times W}\\) is an RGB image, \\(\\mathcal{B} = \\{B_{sub}, B_{obj}\\}\\) denotes the bounding box of the subject and object in \\(I\\), and \\(y\\) is the affordance category label. The IAG (Fig. 3) capture localized features \\(\\mathbf{F}_{\\mathbf{I}} \\in \\mathbb{R}^{C \\times H' \\times W'}\\) and \\(\\mathbf{F}_{p} \\in \\mathbb{R}^{C \\times N_{p}}\\) of the image and point cloud by two feature extractors ResNet [16] and PointNet++ [52]. Then, utilizing \\(B_{obj} / B_{sub}\\) to locate the object/subject region in \\(\\mathbf{F}_{\\mathbf{I}}\\), outside the \\(B_{obj}\\) and \\(B_{sub}\\) is the scene mask \\(M_{sce}\\), use ROI-Align [15] to get the object, subject, and scene features \\(\\mathbf{F}_{i}, \\mathbf{F}_{s}, \\mathbf{F}_{e} \\in \\mathbb{R}^{C \\times H_{1} \\times W_{1}}\\), and reshape them to \\(\\mathbb{R}^{C \\times N_{i}}\\) (\\(N_{i} = H_{1} \\times W_{1}\\)). Next, the JRA module takes \\(\\mathbf{F}_{i}, \\mathbf{F}_{p}\\) as input and jointly mines the correspondence to align them, obtaining the joint feature \\(\\mathbf{F}_{j}\\) (Sec. 3.2). Following, the ARM module takes \\(\\mathbf{F}_{j}, \\mathbf{F}_{s}, \\mathbf{F}_{e}\\) to reveal the affordance \\(\\mathbf{F}_{\\alpha}\\) through cross-attention mechanism (Sec. 3.3). Ultimately, \\(\\mathbf{F}_{\\alpha}\\) and \\(\\mathbf{F}_{j}\\) are sent to the decoder to compute the affordance logits \\(\\hat{y}\\) and the final 3D object affordance \\(\\hat{\\phi}\\) (Sec. 3.4). The process is expressed as \\(\\hat{\\phi}, \\hat{y} = f_{\\theta}(P_{c}, I, \\mathcal{B}; \\theta)\\), where \\(f_{\\theta}\\) denotes the network and \\(\\theta\\) is the parameter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.783, + 0.864 + ], + "angle": 0, + "content": "3.2. Joint Region Alignment Module" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The JRA module calculates the high-dimensional dense similarity in feature space to approximate analogous shape" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10907" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.887, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Figure 3. Method. Overview of Interaction-driven 3D Affordance Grounding Network (IAG), it firstly extracts localized features \\(\\mathbf{F}_i, \\mathbf{F}_p\\) respectively, then takes JRA (Sec. 3.2) to align them and get the joint feature \\(\\mathbf{F}_j\\). Next, ARM (Sec. 3.3) utilizes \\(\\mathbf{F}_j\\) to reveal affordance \\(\\mathbf{F}_{\\alpha}\\) with \\(\\mathbf{F}_s, \\mathbf{F}_e\\) by cross-attention. Eventually, \\(\\mathbf{F}_j\\) and \\(\\mathbf{F}_{\\alpha}\\) are sent to the decoder (Sec. 3.4) to obtain the final results \\(\\hat{\\phi}\\) and \\(\\hat{y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.468, + 0.456 + ], + "angle": 0, + "content": "regions that contain relatively higher similarity [2]. Meanwhile, to map the inherent combination scheme of object components, the JRA models the inherent relations between modality-specific regions and takes it to drive the alignment of other regions by a transformer-based joint-attention \\( f_{\\delta} \\). With the network trained to capture the correspondences among regions from different sources, the alignment is performed implicitly during the optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.532 + ], + "angle": 0, + "content": "Initially, \\(\\mathbf{F}_p\\) and \\(\\mathbf{F}_i\\) are projected into a feature space by shared convolution layers \\(f_{\\sigma}\\), obtaining the features \\(\\mathbf{P} \\in \\mathbb{R}^{C \\times N_p}, \\mathbf{I} \\in \\mathbb{R}^{C \\times N_i}\\). Then, to correlate the analogous shape regions, the dense cross-similarity is calculated between each region of \\(\\mathbf{P}\\) and \\(\\mathbf{I}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.542, + 0.469, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {i, j} = \\frac {e ^ {\\left(\\mathbf {P} _ {i} , \\mathbf {I} _ {j}\\right)}}{\\sum_ {i = 1} ^ {N _ {p}} \\sum_ {j = 1} ^ {N _ {i}} e ^ {\\left(\\mathbf {P} _ {i} , \\mathbf {I} _ {j}\\right)}}, \\varphi \\in \\mathbb {R} ^ {N _ {p} \\times N _ {i}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.469, + 0.681 + ], + "angle": 0, + "content": "where \\(\\varphi_{i,j}\\) denotes the cross-similarity between the i-th region of \\(\\mathbf{P}\\) and the j-th region of \\(\\mathbf{I}\\). Taking the relative difference in \\(\\varphi\\) to refine and correlate analogous regions in \\(\\mathbf{P}\\) and \\(\\mathbf{I}\\), next, applying self-attention layers \\(f_{i}, f_{p}\\) to model the intra-structural inherent relation of objects in respective modalities, the process is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.691, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {P}} = f _ {p} \\left(\\mathbf {I} \\cdot \\varphi^ {T}\\right), \\bar {\\mathbf {I}} = f _ {i} \\left(\\mathbf {P} \\cdot \\varphi\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{P}}\\in \\mathbb{R}^{C\\times N_p},\\bar{\\mathbf{I}}\\in \\mathbb{R}^{C\\times N_i}\\). Following, we perform a joint-attention on features with structural relevance to map the similar inherent combination scheme and drive the alignment of the remaining regions. The process is formulated as: \\(\\mathbf{F}_j = f_{\\delta}[\\bar{\\mathbf{P}},\\bar{\\mathbf{I}}]\\), where \\(\\mathbf{F}_j\\in \\mathbb{R}^{C\\times (N_p + N_i)}\\) and \\([\\cdot ]\\) denotes joining the image feature sequence and point cloud feature sequence into a whole one, and \\(\\mathbf{F}_j\\) denotes the joint representation. The \\(\\mathbf{F}_j\\) is sent to the affordance revealed module to excavate interaction contexts with \\(\\mathbf{F}_s,\\mathbf{F}_e\\). And the affordance knowledge revealed by the interaction contexts could mutually optimize the alignment process during training (see Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.335, + 0.76, + 0.35 + ], + "angle": 0, + "content": "3.3. Affordance Revealed Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.495 + ], + "angle": 0, + "content": "In general, the dynamic factors related to affordance are reflected in the interaction, affecting the extraction of affordance representation, and the interaction is mostly considered to exist between humans and objects. However, the interaction between objects and scenes (including other objects) may also affect affordance [43]. To consider both of these factors, the ARM module utilizes the cross-attention technique to extract these interaction contexts respectively and jointly embed them to reveal explicit object affordance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.495, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Specifically, \\(\\mathbf{F}_j\\) is projected to form the shared query \\(\\mathbf{Q} = \\mathbf{F}_j\\mathbf{W}_1,\\mathbf{F}_s\\) and \\(\\mathbf{F}_e\\) are projected to form different keys and values \\(\\mathbf{K}_{1 / 2} = \\mathbf{F}_s\\mathbf{W}_{2 / 3},\\mathbf{V}_{1 / 2} = \\mathbf{F}_e\\mathbf{W}_{4 / 5}\\) where \\(\\mathbf{W}_{1\\sim 5}\\) are projection weights. Then cross-attention aggregates them to excavate interaction contexts, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.583, + 0.891, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Theta} _ {1 / 2} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} ^ {T} \\cdot \\mathbf {K} _ {1 / 2} / \\sqrt {d}\\right) \\cdot \\mathbf {V} _ {1 / 2} ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.67 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q} \\in \\mathbb{R}^{d \\times (N_p + N_i)}\\), \\(\\mathbf{K}_{1/2}\\), \\(\\mathbf{V}_{1/2} \\in \\mathbb{R}^{d \\times N_i}\\), \\(d\\) is the dimension of projection, \\(\\Theta_1, \\Theta_2\\) indicate the excavated interaction contexts. Then, jointly embedding \\(\\Theta_1\\) and \\(\\Theta_2\\) to reveal affordance, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.678, + 0.891, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\alpha} = f _ {\\xi} (\\boldsymbol {\\Theta} _ {1}, \\boldsymbol {\\Theta} _ {2}), \\mathbf {F} _ {\\alpha} \\in \\mathbb {R} ^ {C \\times (N _ {p} + N _ {i})}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.892, + 0.751 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{\\alpha}\\) is the joint affordance representation, it is split in the decoder to compute final results, \\(f_{\\xi}\\) denotes the concatenation, followed by a convolution layer to fuse the feature." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.746, + 0.774 + ], + "angle": 0, + "content": "3.4. Decoder and Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.876 + ], + "angle": 0, + "content": "The \\(\\mathbf{F}_j\\) is split to \\(\\hat{\\mathbf{F}}_p\\in \\mathbb{R}^{C\\times N_p}\\) and \\(\\hat{\\mathbf{F}}_i\\in \\mathbb{R}^{C\\times N_i}\\), \\(\\mathbf{F}_{\\alpha}\\) is split to \\(\\mathbf{F}_{p\\alpha}\\in \\mathbb{R}^{C\\times N_p}\\) and \\(\\mathbf{F}_{i\\alpha}\\in \\mathbb{R}^{C\\times N_i}\\) in the decoder. Pooling \\(\\mathbf{F}_{p\\alpha}\\) and \\(\\mathbf{F}_{i\\alpha}\\) respectively and concatenate them to compute the affordance logits \\(\\hat{y}\\). \\(\\hat{\\mathbf{F}}_p\\) is upsampled to \\(\\mathbb{R}^{C\\times N}\\) by Feature Propagation Layers (FP) [52], and the 3D object affordance \\(\\hat{\\phi}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.883, + 0.891, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {\\phi} = f _ {\\phi} \\left(\\mathbf {F P} \\left(\\hat {\\mathbf {F}} _ {p}\\right) \\odot \\Gamma \\left(\\mathbf {F} _ {p \\alpha}\\right)\\right), \\hat {\\phi} \\in \\mathbb {R} ^ {N \\times 1}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10908" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.075, + 0.868, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.894, + 0.27 + ], + "angle": 0, + "content": "Figure 4. Properties of the PIAD dataset. (a) Data pairs in the PIAD, the red region in point clouds is the affordance annotation. (b) Distribution of the image data. The horizontal axis represents the category of affordance, the vertical axis represents quantity, and different colors represent different objects. (c) Distribution of the point cloud data. (d) The ratio of images and point clouds in each affordance class. It shows that images and point clouds are not fixed one-to-one pairing, they can form multiple pairs. Please refer to supp. for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.468, + 0.323 + ], + "angle": 0, + "content": "where \\(f_{\\phi}\\) is an output head, \\(\\mathbf{FP}\\) is the upsample layer, \\(\\Gamma\\) denotes pooling \\(\\mathbf{F}_{p\\alpha}\\) and expand it to the shape of \\(\\mathbb{R}^{C\\times N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.596 + ], + "angle": 0, + "content": "The total loss comprises three parts: \\(\\mathcal{L}_{CE}\\), \\(\\mathcal{L}_{KL}\\), \\(\\mathcal{L}_{HM}\\). Where \\(\\mathcal{L}_{CE}\\) computes the cross-entropy loss between \\(y\\) and \\(\\hat{y}\\), it supervises the extraction of \\(\\mathbf{F}_{\\alpha}\\) and implicitly optimizes the alignment process. To enable the network focus on the alignment of affordance regions, we apply the KL Divergence (KLD) [3] to constrain the distribution between \\(\\mathbf{F}_{i\\alpha}\\) and \\(\\hat{\\mathbf{F}}_i\\), formulated as \\(\\mathcal{L}_{KL} = (\\mathbf{F}_{i\\alpha}||\\hat{\\mathbf{F}}_i)\\). The reason is that \\(\\mathbf{F}_{i\\alpha}\\) exhibits the affordance distribution of each object region in the image, and the affordance-related regions keep more significant features. Constraining the feature distribution of \\(\\hat{\\mathbf{F}}_i\\) to enhance the affordance region features in \\(\\hat{\\mathbf{F}}_i\\), and with the region correlations established by the alignment process, \\(\\hat{\\mathbf{F}}_p\\) also tends to exhibit this property, similar to distillation [19, 71]. Which makes the alignment and affordance extraction optimize mutually. \\(\\mathcal{L}_{HM}\\) is a focal loss [28] combined with a dice loss [40], it is calculated by \\(\\hat{\\phi}\\) and \\(P_{label}\\), which supervise the point-wise heatmap on point clouds. Eventually, the total loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.612, + 0.47, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {C E} + \\lambda_ {2} \\mathcal {L} _ {K L} + \\lambda_ {3} \\mathcal {L} _ {H M}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.666 + ], + "angle": 0, + "content": "where \\(\\lambda_{1},\\lambda_{2}\\) and \\(\\lambda_3\\) are hyper-parameters to balance the total loss. See more details in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.166, + 0.695 + ], + "angle": 0, + "content": "4. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Collection Details. We collect Point-Image Affordance Dataset (PIAD), which contains paired image-point cloud affordance data. The point clouds are mainly collected from 3D-AffordanceNet [10], including the point cloud coordinates and affordance annotation. Images are mainly collected from HICO [4], AGD20K [35], and websites with free licenses. The collection criteria is that the image should demonstrate interactions that the object in the point cloud could afford. For example, if the object point cloud is a \"Chair\", it affords \"Sit\", then the image should depict a subject (usually humans) sitting on a chair. The final dataset comprises 7012 point clouds and 5162 images, spanning 23 object classes and 17 affordance categories. Notably, ob" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.339 + ], + "angle": 0, + "content": "jects in the images and point clouds do not sample from the same physical instance, but they belong to the same object category. Paired examples are shown in Fig. 4 (a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.565 + ], + "angle": 0, + "content": "Annotation Details. For point clouds, each point may afford one or multiple interactions, e.g. one annotation is a matrix of (2048, 17), 2048 is the point number, 17 represents the number of affordance types. Each element in the matrix indicates the probability of a point affording a specific affordance. Meanwhile, images are annotated with bounding boxes of the interactive subject and object, as well as an affordance category label. In our task, we only use the heatmap in the point cloud annotation that corresponds to the affordance of the image for training, resulting in a matrix of (2048, 1). By doing so, the affordance category is detached from the point cloud during inference, and the anticipation of the 3D object affordance category only relies on 2D interactions. As a result, different affordances can be anticipated on a point cloud through distinct interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.894, + 0.732 + ], + "angle": 0, + "content": "Statistic Analysis. Since images and point clouds are sampled from different instances, they do not need a fixed one-to-one pairing, one image could be paired with multiple point clouds and the count of them is not strictly consistent. Fig. 4 (b) and (c) show the count and distribution of affordances in images and point clouds. Fig. 4 (d) illustrates the ratio of images and point clouds in each affordance category. PIAD has two partitions: Seen and Unseen. In Seen, both objects and affordances in the training and testing sets are consistent, while in Unseen, some objects or affordances in the testing set do not exist in the training set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.746, + 0.634, + 0.763 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.686, + 0.788 + ], + "angle": 0, + "content": "5.1. Benchmark Setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Evaluation Metrics. To provide a comprehensive and effective evaluation, we compare serval advanced works in the affordance area [10, 47, 74] and finally chose four evaluation metrics: AUC [31], aIOU [53], SIMilarity [57] and Mean Absolute Error [65] to benchmark the PIAD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Modular Baselines. Since there are no prior works using paired image-point cloud data to ground 3D object" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10909" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.088, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 2. Comparison Results on PIAD. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. \"Green\" and \"Orange\" indicate two types of the comparative modular baselines. \"Base.\" represents the baseline. \\(\\diamond\\) denotes the relative improvement of our method over other methods. AUC and aIOU are shown in percentage." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.135, + 0.898, + 0.284 + ], + "angle": 0, + "content": "
MetricsBase.MBDF [59]PMF [80]FRCNN [68]ILN [5]PFusion [67]XMF [1]Ours
SeenAUC ↑68.4974.88◇13.3%75.05◇13.0%76.05◇11.5%75.84◇11.9%77.50◇9.5%78.24◇8.4%84.85 ± 0.3
aIOU ↑6.859.34◇119.5%10.13◇102.4%11.97◇71.3%11.52◇78.0%12.31◇66.6%12.94◇58.5%20.51 ± 0.7
SIM ↑0.3670.415◇31.3%0.425◇28.2%0.429◇27.0%0.427◇27.6%0.432◇26.1%0.441◇23.6%0.545 ± 0.02
MAE ↓0.1520.143◇31.4%0.141◇30.5%0.136◇27.9%0.137◇28.4%0.135◇27.4%0.127◇22.8%0.098 ± 0.01
UnseenAUC ↑57.3458.23◇23.3%60.25◇19.2%61.92◇16.0%59.69◇20.3%61.87◇16.1%62.58◇14.8%71.84 ± 1.8
aIOU ↑3.954.22◇88.4%4.67◇70.2%5.12◇55.2%4.71◇68.8%5.33◇49.1%5.68◇39.9%7.95 ± 0.8
SIM ↑0.3180.325◇8.3%0.330◇6.6%0.332◇6.0%0.325◇8.3%0.330◇6.6%0.342◇2.9%0.352 ± 0.03
MAE ↓0.2350.213◇40.3%0.211◇39.8%0.195◇34.8%0.207◇38.6%0.193◇34.2%0.188◇32.4%0.127 ± 0.01
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.298, + 0.455, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.453, + 0.472, + 0.51 + ], + "angle": 0, + "content": "Figure 5. Paradigm Comparison. (a) Ground truth. (b) The results of 3D-AffordanceNet [10]. (c) Our results. The top row shows the regional confusion case in Seen, and the bottom row displays the generalization in Unseen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.822 + ], + "angle": 0, + "content": "affordance. For a thorough comparison of our method, we select several advanced image-point cloud cross-modal works as modular baselines. Methods of comparison are broadly divided into two types, one is that utilizes camera intrinsics to facilitate cross-modal feature alignment or fusion, i.e. MBDF-Net (MBDF) [59], PMF [80] and FusionRCNN (FRCNN) [68], for this type of method, we remove the step of using intrinsic parameters to align raw data or features to explore their effectiveness on PIAD. Another type performs feature alignment directly in the feature space, without relying on camera intrinsic, i.e. ImLoveNet (ILN) [5], PointFusion (PFusion) [67] and XMFnet (XMF) [1]. To ensure a fair comparison, all methods use the same feature extractors, with the only variation coming from cross-modal alignment or fusion block. For the Baseline, we directly concatenate the features that are output from extractors, regarding the concatenation as the cross-fusion block. More details about baselines are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.27, + 0.849 + ], + "angle": 0, + "content": "5.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The comparison results of evaluation metrics are shown in Tab. 2. As can be seen, our method outperforms the compared baselines, across all metrics in both partitions." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Table 3. Ablation Study. We investigate the improvement of JRA and ARM on the model performance based on the baseline." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.337, + 0.89, + 0.485 + ], + "angle": 0, + "content": "
JRAARMAUCaIOUSIMMAE
Seen69.928.850.4270.132
✓80.2914.310.4950.121
✓78.6713.950.4750.126
✓✓85.1621.200.5640.088
Unseen59.144.050.3380.202
✓66.256.270.3630.159
✓65.795.990.3580.162
✓✓73.698.700.3830.117
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.801 + ], + "angle": 0, + "content": "Besides, to display the limitation of methods that lock geometrics with specific semantic affordance categories, we conduct an experiment to compare one of these methods [10] with ours. As shown in Fig. 5, the top row indicates that the result of this method exhibits regional confusion, e.g. the region where the chair could be moved or sat is geometrically rectangular, and it directly anticipates results on these similar geometries, inconsistent with affordance property. Plus, the bottom row shows that directly establishing a link between object structures and affordances may fail to anticipate correct 3D affordance in Unseen. In contrast, our method anticipates precise results by mining affordance clues provided by 2D interactions. Additionally, visual comparative results of our method and other baselines are shown in Fig. 6. As can be seen, the comparative baselines could anticipate some 3D object affordance under our setting, but in comparison, our method obviously achieves better results, which validates the rationality of our setting, and also demonstrates the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.655, + 0.832 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of JRA. Tab. 3 reports the impact on evaluation metrics of JRA. Fig. 7 provides visual comparison results. It shows that without the JRA, the result is anticipated over the entire region that contains the interactive compo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.075, + 0.48, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.075, + 0.849, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.46, + 0.894, + 0.503 + ], + "angle": 0, + "content": "Figure 6. Visualization Results. The first row is the interactive image, which demonstrates the interaction that the object can afford. The last row is the ground truth of 3D object affordance in the point cloud. The four columns on the left are the visual comparison results for distinct 3D object affordances in the Seen partition. The four columns on the right are the results in the Unseen partition." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.518, + 0.451, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.697, + 0.471, + 0.739 + ], + "angle": 0, + "content": "Figure 7. Ablation of JRA. The top row is the result of the anticipated affordance with/without JRA. The bottom row is a part of the cross-similarity matrix of a sample during the training process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "nents, which means the 2D-3D affordance regions of objects do not match well. Besides, we visualize a part of the cross-similarity matrix between \\(\\hat{\\mathbf{F}}_p\\) and \\(\\hat{\\mathbf{F}}_i\\) in Fig. 7. In the initial stage, only a few analogous shape regions keep explicit correspondence, as training proceeds, the JRA maps the correlation among regions with lower similarity. Meanwhile, the model extracts explicit affordance progressively, and the affordance introduced into the optimization process reveals the corresponding affordance regions." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.516, + 0.868, + 0.705 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.711, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Figure 8. Ablation of ARM. The activation map and t-SNE [61] results with \\((\\boldsymbol{w})\\), without \\((\\boldsymbol{w} / \\boldsymbol{o})\\) ARM in both partitions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of ARM. The influence of ARM on evaluation metrics is also shown in Tab. 3. To visually evaluate the effectiveness of ARM, we employ GradCAM [55] to generate the activation map, and utilize t-SNE [61] to demonstrate the clustering of affordances in both partitions. The results are shown in Fig. 8. It shows that ARM makes the model focus on the interactive regions to excavate the interaction contexts, and enables the model to differentiate various affordances from interactions in both partitions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10911" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.434, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.217, + 0.47, + 0.258 + ], + "angle": 0, + "content": "Figure 9. Same Image w.r.t. Multiple Point Clouds. (a) Same object category. (b) Different object categories, similar geometrics. (c) Different object categories and geometrics." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.27, + 0.454, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.404, + 0.47, + 0.445 + ], + "angle": 0, + "content": "Figure 10. Same Point Cloud w.r.t. Multiple Images. Grounding affordance on the same point cloud with images that contain similar or disparate interactions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.47, + 0.283, + 0.485 + ], + "angle": 0, + "content": "5.4. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Different Instances. We conduct experiments to verify whether the model could ground 3D affordance on instances from different sources: (i) using an image and different point clouds to infer respectively (Fig. 9 (a)). (ii) using multiple images and a single point cloud (Fig. 10). The results indicate that the model can anticipate 3D affordances on different instances with the same 2D interaction, and can also anticipate distinct 3D affordances on the same point cloud through different 2D interactions. Showing that the model maintains the mapping between object structures and affordances by learning from different 2D interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Different Object Categories. What will happen if the object category is different in the image and point cloud? To explore this issue, we perform experiments with mismatched object categories, shown in Fig. 9 (b) and (c). When objects are of different categories but with similar geometric primitives, e.g. the handle of \"Mug\" and the shoulder strap of \"Bag\", our model still properly infer the 3D affordance. However, when geometric structures are also dissimilar (Fig. 9 (c)), the model does not make random predictions. This indicates that the model maps cross-category invariance between affordance and geometries, which can be generalized to new instances of geometry." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Multiplicity. One defining property of affordance is multiplicity: some points may correspond to multiple affordances. Fig. 11 shows that the model makes objects' multiple affordances compatible, and verifies the model does not" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.086, + 0.885, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.155, + 0.89, + 0.184 + ], + "angle": 0, + "content": "Figure 11. Multiple Affordances. Some objects like \"Bag\" contains the region that corresponds to multiple affordances." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.188, + 0.88, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.313, + 0.892, + 0.341 + ], + "angle": 0, + "content": "Figure 12. Real-Wrold. The first row is scanned by iPhone, the second row comes from [29], point clouds are sampled from them." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.345, + 0.875, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.425, + 0.875, + 0.44 + ], + "angle": 0, + "content": "Figure 13. Failure Cases. Over prediction in the small regions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.892, + 0.511 + ], + "angle": 0, + "content": "anticipate affordance by linking a region with specific affordances. Instead, it predicts by considering the presence of affordance-related interactions in the same region." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.512, + 0.893, + 0.664 + ], + "angle": 0, + "content": "Real World. To validate the model in real-world scenarios, we use an iPhone to scan objects and real-scan dataset [29] to test our model, shown in Fig. 12. It shows that our model exhibits a certain degree of generalization to the real world. Limitations. Our model exhibits over predictions for some small affordance regions, shown in Fig. 13. This may be due to the limited understanding of fine-grained geometric parts, and aligning the object features in a single scale may lead to overly large receptive fields. Our feature extension will refer [13, 27] to tackle this problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.619, + 0.694 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.9 + ], + "angle": 0, + "content": "We present a novel setting for affordance grounding, which utilizes the 2D interactive semantics to guide the grounding of 3D object affordance, it has the potential to serve embodied systems when collaborating with multimodal grounding systems [30, 58, 70]. Besides, We collect the PIAD dataset as the first test bed for the proposed setting, it contains paired image-point cloud affordance data. Plus, we propose a novel framework to correlate affordance regions of objects that are from different sources and model interactive contexts to ground 3D object affordance. Comprehensive experiments on PIAD display the reliability of the setting, and we believe it could offer fresh insights and facilitate research in the affordance area." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10912" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "Acknowledgments This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.164, + 0.173, + 0.179 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.47, + 0.23 + ], + "angle": 0, + "content": "[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. arXiv preprint arXiv:2209.09552, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[2] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE transactions on pattern analysis and machine intelligence, 41(3):740-757, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.47, + 0.411 + ], + "angle": 0, + "content": "[4] Yu-Wei Chao, Zhan Wang, Yugeng He, Jiaxuan Wang, and Jia Deng. Hico: A benchmark for recognizing human-object interactions in images. In Proceedings of the IEEE international conference on computer vision, pages 1017–1025, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[5] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. Imlovenet: Misaligned image-supported registration network for low-overlap point cloud pairs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[6] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22(4):449-462, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.47, + 0.593 + ], + "angle": 0, + "content": "[7] Ching-Yao Chuang, Jiaman Li, Antonio Torralba, and Sanja Fidler. Learning to act properly: Predicting and explaining affordances from images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 975–983, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[8] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2):722-739, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[9] Barney Dalgarno and Mark JW Lee. What are the learning affordances of 3-d virtual environments? British Journal of Educational Technology, 41(1):10-32, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.189, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, June 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[17] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[18] Maximilian Jaritz, Jiayuan Gu, and Hao Su. Multi-view pointnet for 3d scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.894, + 0.49 + ], + "angle": 0, + "content": "[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[20] David Inkyu Kim and Gaurav S Sukhatme. Semantic labeling of 3d point clouds with object affordance for robot manipulation. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 5578-5584. IEEE, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[21] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[22] Hema S Koppula and Ashutosh Saxena. Physically grounded spatio-temporal object affordances. In European Conference on Computer Vision, pages 831-847. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[23] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[24] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhi-hong Zhu, and Yuexian Zou. G21: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Yaowei Li, Bang Yang, Xuxin Cheng, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. Unify, align and refine: Multi-level semantic alignment for radiology report generation. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10913" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[29] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, June 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[30] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[31] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[32] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Transactions on Artificial Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[33] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot affordance detection. arXiv preprint arXiv:2106.14747, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[34] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Grounded affordance from exocentric view. arXiv preprint arXiv:2208.13196, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.658 + ], + "angle": 0, + "content": "[35] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2252-2261, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.716 + ], + "angle": 0, + "content": "[36] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.744 + ], + "angle": 0, + "content": "[37] Maurice Merleau-Ponty. Phenomenologie de la perception (1945). Libraqire Gallimard, Paris, 1976. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[38] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers in Neurorobotics, 14:26, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[39] Jinpeng Mi, Song Tang, Zhen Deng, Michael Goerner, and Jianwei Zhang. Object affordance based multimodal fusion for natural human-robot interaction. Cognitive Systems Research, 54:128–137, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[40] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.133 + ], + "angle": 0, + "content": "medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[41] Niluthpol Chowdhury Mithun, Karan Sikka, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Rgb2lidar: Towards solving large-scale cross-modal visual localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 934-954, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.893, + 0.276 + ], + "angle": 0, + "content": "[42] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[43] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[44] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.893, + 0.488 + ], + "angle": 0, + "content": "[45] Bogdan Moldovan, Plinio Moreno, Martijn Van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE international conference on robotics and automation, pages 4373-4378. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.893, + 0.558 + ], + "angle": 0, + "content": "[46] Austin Myers, Ching L Teo, Cornelia Fermüller, and Yiannis Aloimonos. Affordance detection of tool parts from geometric features. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 1374-1381. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.893, + 0.616 + ], + "angle": 0, + "content": "[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[48] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.893, + 0.743 + ], + "angle": 0, + "content": "[49] Anh Nguyen, Dimitrios Kanoulas, Darwin G Caldwell, and Nikos G Tsagarakis. Detecting object affordances with convolutional neural networks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2765-2770. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.893, + 0.8 + ], + "angle": 0, + "content": "[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 7(1-2):1-179, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.893, + 0.844 + ], + "angle": 0, + "content": "[51] François Osiurak and Arnaud Badets. Tool use and affordance: Manipulation-based versus reasoning-based approaches. Psychological review, 123(5):534, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[52] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10914" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[53] Md Atiqur Rahman and Yang Wang. Optimizing intersection-over-union in deep neural networks for image segmentation. In International symposium on visual computing, pages 234–244. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[54] Anirban Roy and Sinisa Todorovic. A multi-scale cnn for affordance segmentation in rgb images. In European conference on computer vision, pages 186-201. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[55] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[56] Chris Shilling. The body in culture, technology and society. Sage, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[57] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[58] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[59] Xun Tan, Xingyu Chen, Guowei Zhang, Jishiyu Ding, and Xuguang Lan. Mbdf-net: Multi-branch deep fusion network for 3d object detection. In Proceedings of the 1st International Workshop on Multimedia Computing for Urban Data, pages 9–17, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[60] Spyridon Thermos, Petros Daras, and Gerasimos Potamianos. A deep learning approach to object affordance segmentation. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2358-2362. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[61] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[62] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[63] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In European Conference on Computer Vision, pages 421-436. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[64] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 90-107. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[65] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[66] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[67] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[68] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[69] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[70] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[71] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73-91. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[72] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[73] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[74] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[75] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4534-4543, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[76] Linqing Zhao, Jiwen Lu, and Jie Zhou. Similarity-aware fusion network for 3d semantic segmentation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1585-1592. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[77] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.713, + 0.892, + 0.754 + ], + "angle": 0, + "content": "[78] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Computing and Applications, 32(18):14321-14333, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.756, + 0.892, + 0.823 + ], + "angle": 0, + "content": "[79] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.826, + 0.892, + 0.895 + ], + "angle": 0, + "content": "[80] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10915" + } + ] +] \ No newline at end of file diff --git a/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_origin.pdf b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8944ced6ad785cd851c5665fa500e06162ec48be --- /dev/null +++ b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/7f72f78a-6126-465f-bd5b-53bbbe4b1d49_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3929a721cfc8c2a21d5fba4e9574bde94c6156b3e835835ebdfea1b7ea335ba +size 6260017 diff --git a/2023/Grounding 3D Object Affordance from 2D Interactions in Images/full.md b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2e20ea2871198b26504385b6323ec54d78457460 --- /dev/null +++ b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/full.md @@ -0,0 +1,303 @@ +# Grounding 3D Object Affordance from 2D Interactions in Images + +Yuhang Yang $^{1}$ , Wei Zhai $^{1,*}$ , Hongchen Luo $^{1}$ , Yang Cao $^{1,2}$ , Jiebo Luo $^{3}$ , Zheng-Jun Zha $^{1}$ $^{1}$ University of Science and Technology of China + $^{3}$ University of Rochester + $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +yyuhang@mail., wzhai056@mail., lhcl2@mail., forrest@ustc.edu.cn + +jluo@cs.rochester.edu zhazj@ustc.edu.cn + +# Abstract + +Grounding 3D object affordance seeks to locate objects' "action possibilities" regions in the 3D space, which serves as a link between perception and operation for embodied agents. Existing studies primarily focus on connecting visual affordances with geometry structures, e.g., relying on annotations to declare interactive regions of interest on the object and establishing a mapping between the regions and affordances. However, the essence of learning object affordance is to understand how to use it, and the manner that detaches interactions is limited in generalization. Normally, humans possess the ability to perceive object affordances in the physical world through demonstration images or videos. Motivated by this, we introduce a novel task setting: grounding 3D object affordance from 2D interactions in images, which faces the challenge of anticipating affordance through interactions of different sources. To address this problem, we devise a novel Interaction-driven 3D Affordance Grounding Network (IAG), which aligns the region feature of objects from different sources and models the interactive contexts for 3D object affordance grounding. Besides, we collect a Point-Image Affordance Dataset (PIAD) to support the proposed task. Comprehensive experiments on PIAD demonstrate the reliability of the proposed task and the superiority of our method. The project is available at https://github.com/yyvhang/IAGNet. + +# 1. Introduction + +The term "affordance" is described as "opportunities of interaction" by J. Gibson [14]. Grounding 3D object affordance aims to comprehend the interactive regions of objects in 3D space, which is not only simply to predict which interaction an object affords, but also to identify specific points on the object that could support the interaction. It consti + +![](images/8247bf0739f3d10b9f0296881df42b876a940e11b48883728ea15b3f7f6866b7.jpg) +Figure 1. Grounding Affordance from Interactions. We propose to ground 3D object affordance through 2D interactions. Given an object point cloud with an interactive image, grounding the corresponding affordance on the 3D object. + +tutes a link between perception and operation for embodied agents, which has the potential to serve numerous practical applications, e.g. action prediction [21, 63], robot manipulation [36, 45, 51], imitation learning [17, 50], and augmented/virtual reality [6, 9]. + +So far, the paradigm of perceiving 3D object affordance has several branches. One of them involves establishing an explicit mapping between affordance categories and geometry structures [10, 20, 46, 66], based on visual appearance. However, affordance is dynamic and multiple, these geometric-specific manners have limited generalization for unseen structures. Besides, locking the geometry with a specific affordance category may lead to the anticipated region being inconsistent with its affordance when objects possess multiple similar geometrics, resulting in affordance regional confusion. Another paradigm is based on reinforcement learning, which puts the agent in 3D synthetic scenarios to interact with several objects actively, taking the reward mechanism to optimize the whole process [48]. While this type of approach transitions agents from passive recognition to active reasoning, it needs repeated attempts in a huge search space when meeting a novel structure, and is therefore time-consuming. + +![](images/4cc3bd9669f23af051f7af326d58d307576221a3458726e494f5e37aa5a0c742.jpg) + +![](images/8539b518ee166ef2e0c94c58af4233460fc128c48b8e84c8173cb0eeb30dd9de.jpg) +Figure 2. Motivation. (a) Clues to correlate regions of objects from different sources. Objects with the same category possess a similar combination scheme to meet certain affordance properties, it is a similar inherent relation among different instances. And some structures hint at these properties exhibit closer representations with higher similarity in the latent space. (b) Object affordance may be affected by dynamic factors, such as the position of the object itself in the scene, other objects in the scene, and the body part of the interacting subject. These factors can be decomposed into object-subject and object-scene interaction contexts. + +These limitations motivate us to explore an affordance-compatible learning paradigm. Typically, humans can infer object affordance in the 3D physical world by watching images or videos that demonstrate the interactions. Some studies in cognitive science [37, 56] point out the existence of "body image" in human cognition, which claims that humans have a perceptual experience of the objects they see, thus facilitating their ability to organize the perception [73] and operate novel objects. Hence, human-object interaction conveys the perception that object structure could perform certain affordance, which is a crucial clue to reason object affordance. In light of this, we present a novel task setting: grounding 3D object affordance from 2D interactions in images, which is shown in Fig. 1. + +This challenging task includes several essential issues that should be properly addressed. 1) Alignment ambiguity. To ground 3D object affordance from 2D source interactions, the premise is to correspond the regions of the object in different sources. The object in 2D demonstrations and the 3D object we face are usually derived from different physical instances in different locations and scales. This discrepancy may lead to confusion when matching the affordance regions, causing alignment ambiguity. While objects are commonly designed to satisfy certain needs of human beings, so the same category generally follows a similar combination scheme of object components to meet certain affordances, and these affordances are hinted at by some structures (Fig. 2 (a)). These invariant properties are across instances and could be utilized to correlate object regions from different sources. 2) Affordance ambiguity. Affordance has properties of dynamic and multiplicity, which means the object affordance may change according to the situation, the same part of an object could afford multiple interactions, as shown in Fig. 2 (b), "Chair" affords "Sit" or "Move" depends on the human actions, "Mug" af + +fords "Wrapgrasp" or "Contain" according to the scene context, these properties may make ambiguity when extracting affordance. However, these dynamic factors for affordance extraction can be decomposed into the interaction between subject-object and object-scene. Modeling these interactions is possible to extract explicit affordance. + +To address these issues, we propose the Interaction-driven 3D Affordance Grounding Network (IAG) to align object region features from different sources and model interaction contexts to reveal affordance. In detail, it contains two sub-modules, one is Joint Region Alignment Module (JRA), which is devised to eliminate alignment ambiguity. It takes the relative difference in dense cross-similarity to refine analogous shape regions and employs learnable layers to map the invariant combination scheme, taking them to match local regions of objects. For affordance ambiguity, the other one Affordance Revealed Module (ARM) takes the object representation as a shared factor and jointly models its interaction contexts with the affordance-related factors to reveal explicit affordance. Moreover, we collect Point-Image Affordance Dataset (PIAD) that contains plenty of paired image-point cloud affordance data and make a benchmark to support the model training and evaluation on the proposed setting. + +The contributions are summarized as follows: 1) We introduce grounding 3D object affordance through the 2D interactions, which facilitates the generalization to 3D object affordance perception. 2) We propose the IAG framework, which aligns the region feature of objects from different sources and jointly models affordance-related interactions to locate the 3D object affordance. 3) We collect a dataset named PIAD to support the proposed task setting, which contains paired image-point cloud affordance data. Besides, we establish a benchmark on PIAD, and the experiments on PIAD exhibit the reliability of the method and the setting. + +# 2. Related Works + +# 2.1. Affordance Learning + +At present, the affordance area has made achievements in multiple tasks (Tab. 1). Some works are devoted to detecting the affordance region from 2D sources i.e. images and videos [7, 11, 25, 33, 54, 60, 78], and there are also some studies accomplish this task with the assistance of natural language [32, 38, 39]. They seek to detect or segment objects that afford the action in 2D data, but cannot perceive the specific parts corresponding to the affordance. Thus, another type of method brings a leap to locate the affordance region of objects from 2D sources [12, 35, 47, 34]. However, the affordance knowledge derived from 2D sources is hard to extrapolate specific interactive locations of objects in the 3D environment. With several 3D object datasets proposed [13, 29, 44], some researchers explore grounding object affordance from 3D data [10, 43, 66, 48]. Such methods directly map the connection between semantic affordances and 3D structures, detach the real interaction, and may deter the generalization ability: structures that do not map to a specific affordance are usually hard to generalize through this type of method. T. Nagarajan et al. [48] give a fresh mind, taking the reinforcement learning to make agents actively interact with the 3D synthetic scenarios, while it requires repeated attempts by the agents in a given search space, and is time-consuming. In robotics, affordance is utilized to provide priors for object manipulation and achieve considerable results, especially the articulated 3D objects [42, 64, 79]. Several methods utilize the 2.5D data e.g. RGB-D image to understand the object affordance and serve the manipulation task [21, 22, 23, 49], for these methods, the image and depth information are corresponding in spatial and need to be collected in the same scene. In contrast, our task focuses on grounding 3D object affordance from 2D interactions, in which the 2D interactions provide direct clues to excavate object affordance efficiently and make the affordance could generalize to some unseen structures, and the 2D-3D data is collected from different sources, freeing the constraints on spatial correspondence. + +# 2.2. Image-Point Cloud Cross-Modal Learning + +The combination of point cloud and image data can capture both semantic and geometric information, enabling cross-modal learning among them has great potential application value in scenarios like autopilot [8]. For this type of work, aligning features is the premise for completing downstream tasks e.g. detection, segmentation, and registration. It is aimed at establishing correspondences between instances from different modalities [2, 26, 69], either spatial or semantic. To achieve this, many methods use camera intrinsic to correspond spatial position of pixels and points, then align per pixel-point feature or fuse the raw + +Table 1. Affordance Learning. Various works for several tasks in the affordance community. + +
WorkInputOutputTask
[11, 60, 78]Image/Video2D MaskDetection
[32, 38, 39]Image,Text2D MaskDetection
[21, 23, 49]RGBDHeatmap & ActionManipulation
[42, 64, 79]Point CloudHeatmap & ActionManipulation
[12, 35, 47]Image/Video2D HeatmapGrounding
[10, 43, 66]Point Cloud3D HeatmapGrounding
+ +data [24, 59, 62, 68, 77, 80]. Some works utilize depth information to project image features into 3D space and then fuse them with point-wise features [18, 41, 72, 75, 76]. The above methods rely on the spatial prior information to align features, while in our task, images and point clouds are collected from distinct physical instances, there are no priors on camera pose or intrinsics, also no corresponding depth information of the image. Therefore, we align the object region features that are derived from different sources in the feature space with the assistance of the correlation invariance between affordances and appearance structures. + +# 3. Method + +# 3.1. Overview + +Our goal is to anticipate the affordance regions on the point cloud that correspond to the interaction in the image. Given a sample $\{P, I, \mathcal{B}, y\}$ , where $P$ is a point cloud with the coordinates $P_{c} \in \mathbb{R}^{N \times 3}$ and the affordance annotation $P_{label} \in \mathbb{R}^{N \times 1}$ , $I \in \mathbb{R}^{3 \times H \times W}$ is an RGB image, $\mathcal{B} = \{B_{sub}, B_{obj}\}$ denotes the bounding box of the subject and object in $I$ , and $y$ is the affordance category label. The IAG (Fig. 3) capture localized features $\mathbf{F}_{\mathbf{I}} \in \mathbb{R}^{C \times H' \times W'}$ and $\mathbf{F}_{p} \in \mathbb{R}^{C \times N_{p}}$ of the image and point cloud by two feature extractors ResNet [16] and PointNet++ [52]. Then, utilizing $B_{obj} / B_{sub}$ to locate the object/subject region in $\mathbf{F}_{\mathbf{I}}$ , outside the $B_{obj}$ and $B_{sub}$ is the scene mask $M_{sce}$ , use ROI-Align [15] to get the object, subject, and scene features $\mathbf{F}_{i}, \mathbf{F}_{s}, \mathbf{F}_{e} \in \mathbb{R}^{C \times H_{1} \times W_{1}}$ , and reshape them to $\mathbb{R}^{C \times N_{i}}$ ( $N_{i} = H_{1} \times W_{1}$ ). Next, the JRA module takes $\mathbf{F}_{i}, \mathbf{F}_{p}$ as input and jointly mines the correspondence to align them, obtaining the joint feature $\mathbf{F}_{j}$ (Sec. 3.2). Following, the ARM module takes $\mathbf{F}_{j}, \mathbf{F}_{s}, \mathbf{F}_{e}$ to reveal the affordance $\mathbf{F}_{\alpha}$ through cross-attention mechanism (Sec. 3.3). Ultimately, $\mathbf{F}_{\alpha}$ and $\mathbf{F}_{j}$ are sent to the decoder to compute the affordance logits $\hat{y}$ and the final 3D object affordance $\hat{\phi}$ (Sec. 3.4). The process is expressed as $\hat{\phi}, \hat{y} = f_{\theta}(P_{c}, I, \mathcal{B}; \theta)$ , where $f_{\theta}$ denotes the network and $\theta$ is the parameter. + +# 3.2. Joint Region Alignment Module + +The JRA module calculates the high-dimensional dense similarity in feature space to approximate analogous shape + +![](images/4930cba0022500dd36f1fbd31252ca1d519670a9d66912c97e0111f5293fdea2.jpg) +Figure 3. Method. Overview of Interaction-driven 3D Affordance Grounding Network (IAG), it firstly extracts localized features $\mathbf{F}_i, \mathbf{F}_p$ respectively, then takes JRA (Sec. 3.2) to align them and get the joint feature $\mathbf{F}_j$ . Next, ARM (Sec. 3.3) utilizes $\mathbf{F}_j$ to reveal affordance $\mathbf{F}_{\alpha}$ with $\mathbf{F}_s, \mathbf{F}_e$ by cross-attention. Eventually, $\mathbf{F}_j$ and $\mathbf{F}_{\alpha}$ are sent to the decoder (Sec. 3.4) to obtain the final results $\hat{\phi}$ and $\hat{y}$ . + +regions that contain relatively higher similarity [2]. Meanwhile, to map the inherent combination scheme of object components, the JRA models the inherent relations between modality-specific regions and takes it to drive the alignment of other regions by a transformer-based joint-attention $f_{\delta}$ . With the network trained to capture the correspondences among regions from different sources, the alignment is performed implicitly during the optimization. + +Initially, $\mathbf{F}_p$ and $\mathbf{F}_i$ are projected into a feature space by shared convolution layers $f_{\sigma}$ , obtaining the features $\mathbf{P} \in \mathbb{R}^{C \times N_p}, \mathbf{I} \in \mathbb{R}^{C \times N_i}$ . Then, to correlate the analogous shape regions, the dense cross-similarity is calculated between each region of $\mathbf{P}$ and $\mathbf{I}$ , formulated as: + +$$ +\varphi_ {i, j} = \frac {e ^ {\left(\mathbf {P} _ {i} , \mathbf {I} _ {j}\right)}}{\sum_ {i = 1} ^ {N _ {p}} \sum_ {j = 1} ^ {N _ {i}} e ^ {\left(\mathbf {P} _ {i} , \mathbf {I} _ {j}\right)}}, \varphi \in \mathbb {R} ^ {N _ {p} \times N _ {i}}, \tag {1} +$$ + +where $\varphi_{i,j}$ denotes the cross-similarity between the i-th region of $\mathbf{P}$ and the j-th region of $\mathbf{I}$ . Taking the relative difference in $\varphi$ to refine and correlate analogous regions in $\mathbf{P}$ and $\mathbf{I}$ , next, applying self-attention layers $f_{i}, f_{p}$ to model the intra-structural inherent relation of objects in respective modalities, the process is expressed as: + +$$ +\bar {\mathbf {P}} = f _ {p} \left(\mathbf {I} \cdot \varphi^ {T}\right), \bar {\mathbf {I}} = f _ {i} \left(\mathbf {P} \cdot \varphi\right), \tag {2} +$$ + +where $\bar{\mathbf{P}}\in \mathbb{R}^{C\times N_p},\bar{\mathbf{I}}\in \mathbb{R}^{C\times N_i}$ . Following, we perform a joint-attention on features with structural relevance to map the similar inherent combination scheme and drive the alignment of the remaining regions. The process is formulated as: $\mathbf{F}_j = f_{\delta}[\bar{\mathbf{P}},\bar{\mathbf{I}}]$ , where $\mathbf{F}_j\in \mathbb{R}^{C\times (N_p + N_i)}$ and $[\cdot ]$ denotes joining the image feature sequence and point cloud feature sequence into a whole one, and $\mathbf{F}_j$ denotes the joint representation. The $\mathbf{F}_j$ is sent to the affordance revealed module to excavate interaction contexts with $\mathbf{F}_s,\mathbf{F}_e$ . And the affordance knowledge revealed by the interaction contexts could mutually optimize the alignment process during training (see Sec. 3.4). + +# 3.3. Affordance Revealed Module + +In general, the dynamic factors related to affordance are reflected in the interaction, affecting the extraction of affordance representation, and the interaction is mostly considered to exist between humans and objects. However, the interaction between objects and scenes (including other objects) may also affect affordance [43]. To consider both of these factors, the ARM module utilizes the cross-attention technique to extract these interaction contexts respectively and jointly embed them to reveal explicit object affordance. + +Specifically, $\mathbf{F}_j$ is projected to form the shared query $\mathbf{Q} = \mathbf{F}_j\mathbf{W}_1,\mathbf{F}_s$ and $\mathbf{F}_e$ are projected to form different keys and values $\mathbf{K}_{1 / 2} = \mathbf{F}_s\mathbf{W}_{2 / 3},\mathbf{V}_{1 / 2} = \mathbf{F}_e\mathbf{W}_{4 / 5}$ where $\mathbf{W}_{1\sim 5}$ are projection weights. Then cross-attention aggregates them to excavate interaction contexts, expressed as: + +$$ +\boldsymbol {\Theta} _ {1 / 2} = \operatorname {s o f t m a x} \left(\mathbf {Q} ^ {T} \cdot \mathbf {K} _ {1 / 2} / \sqrt {d}\right) \cdot \mathbf {V} _ {1 / 2} ^ {T}, \tag {3} +$$ + +where $\mathbf{Q} \in \mathbb{R}^{d \times (N_p + N_i)}$ , $\mathbf{K}_{1/2}$ , $\mathbf{V}_{1/2} \in \mathbb{R}^{d \times N_i}$ , $d$ is the dimension of projection, $\Theta_1, \Theta_2$ indicate the excavated interaction contexts. Then, jointly embedding $\Theta_1$ and $\Theta_2$ to reveal affordance, expressed as: + +$$ +\mathbf {F} _ {\alpha} = f _ {\xi} (\boldsymbol {\Theta} _ {1}, \boldsymbol {\Theta} _ {2}), \mathbf {F} _ {\alpha} \in \mathbb {R} ^ {C \times (N _ {p} + N _ {i})}, \tag {4} +$$ + +where $\mathbf{F}_{\alpha}$ is the joint affordance representation, it is split in the decoder to compute final results, $f_{\xi}$ denotes the concatenation, followed by a convolution layer to fuse the feature. + +# 3.4. Decoder and Loss Function + +The $\mathbf{F}_j$ is split to $\hat{\mathbf{F}}_p\in \mathbb{R}^{C\times N_p}$ and $\hat{\mathbf{F}}_i\in \mathbb{R}^{C\times N_i}$ , $\mathbf{F}_{\alpha}$ is split to $\mathbf{F}_{p\alpha}\in \mathbb{R}^{C\times N_p}$ and $\mathbf{F}_{i\alpha}\in \mathbb{R}^{C\times N_i}$ in the decoder. Pooling $\mathbf{F}_{p\alpha}$ and $\mathbf{F}_{i\alpha}$ respectively and concatenate them to compute the affordance logits $\hat{y}$ . $\hat{\mathbf{F}}_p$ is upsampled to $\mathbb{R}^{C\times N}$ by Feature Propagation Layers (FP) [52], and the 3D object affordance $\hat{\phi}$ is computed as: + +$$ +\hat {\phi} = f _ {\phi} \left(\mathbf {F P} \left(\hat {\mathbf {F}} _ {p}\right) \odot \Gamma \left(\mathbf {F} _ {p \alpha}\right)\right), \hat {\phi} \in \mathbb {R} ^ {N \times 1}, \tag {5} +$$ + +![](images/2ae5d64006c44d4a3924455e449838133da73909bd8fbf5077308aa609f61a97.jpg) +Figure 4. Properties of the PIAD dataset. (a) Data pairs in the PIAD, the red region in point clouds is the affordance annotation. (b) Distribution of the image data. The horizontal axis represents the category of affordance, the vertical axis represents quantity, and different colors represent different objects. (c) Distribution of the point cloud data. (d) The ratio of images and point clouds in each affordance class. It shows that images and point clouds are not fixed one-to-one pairing, they can form multiple pairs. Please refer to supp. for more details. + +where $f_{\phi}$ is an output head, $\mathbf{FP}$ is the upsample layer, $\Gamma$ denotes pooling $\mathbf{F}_{p\alpha}$ and expand it to the shape of $\mathbb{R}^{C\times N}$ . + +The total loss comprises three parts: $\mathcal{L}_{CE}$ , $\mathcal{L}_{KL}$ , $\mathcal{L}_{HM}$ . Where $\mathcal{L}_{CE}$ computes the cross-entropy loss between $y$ and $\hat{y}$ , it supervises the extraction of $\mathbf{F}_{\alpha}$ and implicitly optimizes the alignment process. To enable the network focus on the alignment of affordance regions, we apply the KL Divergence (KLD) [3] to constrain the distribution between $\mathbf{F}_{i\alpha}$ and $\hat{\mathbf{F}}_i$ , formulated as $\mathcal{L}_{KL} = (\mathbf{F}_{i\alpha}||\hat{\mathbf{F}}_i)$ . The reason is that $\mathbf{F}_{i\alpha}$ exhibits the affordance distribution of each object region in the image, and the affordance-related regions keep more significant features. Constraining the feature distribution of $\hat{\mathbf{F}}_i$ to enhance the affordance region features in $\hat{\mathbf{F}}_i$ , and with the region correlations established by the alignment process, $\hat{\mathbf{F}}_p$ also tends to exhibit this property, similar to distillation [19, 71]. Which makes the alignment and affordance extraction optimize mutually. $\mathcal{L}_{HM}$ is a focal loss [28] combined with a dice loss [40], it is calculated by $\hat{\phi}$ and $P_{label}$ , which supervise the point-wise heatmap on point clouds. Eventually, the total loss is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {C E} + \lambda_ {2} \mathcal {L} _ {K L} + \lambda_ {3} \mathcal {L} _ {H M}, \tag {6} +$$ + +where $\lambda_{1},\lambda_{2}$ and $\lambda_3$ are hyper-parameters to balance the total loss. See more details in supplementary materials. + +# 4. Dataset + +Collection Details. We collect Point-Image Affordance Dataset (PIAD), which contains paired image-point cloud affordance data. The point clouds are mainly collected from 3D-AffordanceNet [10], including the point cloud coordinates and affordance annotation. Images are mainly collected from HICO [4], AGD20K [35], and websites with free licenses. The collection criteria is that the image should demonstrate interactions that the object in the point cloud could afford. For example, if the object point cloud is a "Chair", it affords "Sit", then the image should depict a subject (usually humans) sitting on a chair. The final dataset comprises 7012 point clouds and 5162 images, spanning 23 object classes and 17 affordance categories. Notably, ob + +jects in the images and point clouds do not sample from the same physical instance, but they belong to the same object category. Paired examples are shown in Fig. 4 (a). + +Annotation Details. For point clouds, each point may afford one or multiple interactions, e.g. one annotation is a matrix of (2048, 17), 2048 is the point number, 17 represents the number of affordance types. Each element in the matrix indicates the probability of a point affording a specific affordance. Meanwhile, images are annotated with bounding boxes of the interactive subject and object, as well as an affordance category label. In our task, we only use the heatmap in the point cloud annotation that corresponds to the affordance of the image for training, resulting in a matrix of (2048, 1). By doing so, the affordance category is detached from the point cloud during inference, and the anticipation of the 3D object affordance category only relies on 2D interactions. As a result, different affordances can be anticipated on a point cloud through distinct interactions. + +Statistic Analysis. Since images and point clouds are sampled from different instances, they do not need a fixed one-to-one pairing, one image could be paired with multiple point clouds and the count of them is not strictly consistent. Fig. 4 (b) and (c) show the count and distribution of affordances in images and point clouds. Fig. 4 (d) illustrates the ratio of images and point clouds in each affordance category. PIAD has two partitions: Seen and Unseen. In Seen, both objects and affordances in the training and testing sets are consistent, while in Unseen, some objects or affordances in the testing set do not exist in the training set. + +# 5. Experiments + +# 5.1. Benchmark Setting + +Evaluation Metrics. To provide a comprehensive and effective evaluation, we compare serval advanced works in the affordance area [10, 47, 74] and finally chose four evaluation metrics: AUC [31], aIOU [53], SIMilarity [57] and Mean Absolute Error [65] to benchmark the PIAD. + +Modular Baselines. Since there are no prior works using paired image-point cloud data to ground 3D object + +Table 2. Comparison Results on PIAD. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. "Green" and "Orange" indicate two types of the comparative modular baselines. "Base." represents the baseline. $\diamond$ denotes the relative improvement of our method over other methods. AUC and aIOU are shown in percentage. + +
MetricsBase.MBDF [59]PMF [80]FRCNN [68]ILN [5]PFusion [67]XMF [1]Ours
SeenAUC ↑68.4974.88◇13.3%75.05◇13.0%76.05◇11.5%75.84◇11.9%77.50◇9.5%78.24◇8.4%84.85 ± 0.3
aIOU ↑6.859.34◇119.5%10.13◇102.4%11.97◇71.3%11.52◇78.0%12.31◇66.6%12.94◇58.5%20.51 ± 0.7
SIM ↑0.3670.415◇31.3%0.425◇28.2%0.429◇27.0%0.427◇27.6%0.432◇26.1%0.441◇23.6%0.545 ± 0.02
MAE ↓0.1520.143◇31.4%0.141◇30.5%0.136◇27.9%0.137◇28.4%0.135◇27.4%0.127◇22.8%0.098 ± 0.01
UnseenAUC ↑57.3458.23◇23.3%60.25◇19.2%61.92◇16.0%59.69◇20.3%61.87◇16.1%62.58◇14.8%71.84 ± 1.8
aIOU ↑3.954.22◇88.4%4.67◇70.2%5.12◇55.2%4.71◇68.8%5.33◇49.1%5.68◇39.9%7.95 ± 0.8
SIM ↑0.3180.325◇8.3%0.330◇6.6%0.332◇6.0%0.325◇8.3%0.330◇6.6%0.342◇2.9%0.352 ± 0.03
MAE ↓0.2350.213◇40.3%0.211◇39.8%0.195◇34.8%0.207◇38.6%0.193◇34.2%0.188◇32.4%0.127 ± 0.01
+ +![](images/f9e82174288403b9f88e23bcf64ecb47e40ce83150a93739f94932c250539725.jpg) +Figure 5. Paradigm Comparison. (a) Ground truth. (b) The results of 3D-AffordanceNet [10]. (c) Our results. The top row shows the regional confusion case in Seen, and the bottom row displays the generalization in Unseen. + +affordance. For a thorough comparison of our method, we select several advanced image-point cloud cross-modal works as modular baselines. Methods of comparison are broadly divided into two types, one is that utilizes camera intrinsics to facilitate cross-modal feature alignment or fusion, i.e. MBDF-Net (MBDF) [59], PMF [80] and FusionRCNN (FRCNN) [68], for this type of method, we remove the step of using intrinsic parameters to align raw data or features to explore their effectiveness on PIAD. Another type performs feature alignment directly in the feature space, without relying on camera intrinsic, i.e. ImLoveNet (ILN) [5], PointFusion (PFusion) [67] and XMFnet (XMF) [1]. To ensure a fair comparison, all methods use the same feature extractors, with the only variation coming from cross-modal alignment or fusion block. For the Baseline, we directly concatenate the features that are output from extractors, regarding the concatenation as the cross-fusion block. More details about baselines are provided in the supplementary materials. + +# 5.2. Comparison Results + +The comparison results of evaluation metrics are shown in Tab. 2. As can be seen, our method outperforms the compared baselines, across all metrics in both partitions. + +Table 3. Ablation Study. We investigate the improvement of JRA and ARM on the model performance based on the baseline. + +
JRAARMAUCaIOUSIMMAE
Seen69.928.850.4270.132
✓80.2914.310.4950.121
✓78.6713.950.4750.126
✓✓85.1621.200.5640.088
Unseen59.144.050.3380.202
✓66.256.270.3630.159
✓65.795.990.3580.162
✓✓73.698.700.3830.117
+ +Besides, to display the limitation of methods that lock geometrics with specific semantic affordance categories, we conduct an experiment to compare one of these methods [10] with ours. As shown in Fig. 5, the top row indicates that the result of this method exhibits regional confusion, e.g. the region where the chair could be moved or sat is geometrically rectangular, and it directly anticipates results on these similar geometries, inconsistent with affordance property. Plus, the bottom row shows that directly establishing a link between object structures and affordances may fail to anticipate correct 3D affordance in Unseen. In contrast, our method anticipates precise results by mining affordance clues provided by 2D interactions. Additionally, visual comparative results of our method and other baselines are shown in Fig. 6. As can be seen, the comparative baselines could anticipate some 3D object affordance under our setting, but in comparison, our method obviously achieves better results, which validates the rationality of our setting, and also demonstrates the superiority of our method. + +# 5.3. Ablation Study + +Effectiveness of JRA. Tab. 3 reports the impact on evaluation metrics of JRA. Fig. 7 provides visual comparison results. It shows that without the JRA, the result is anticipated over the entire region that contains the interactive compo + +![](images/a151ae1999e8e56175ae92d6788ee6eff41373c50d0619773e3c11b280511ed6.jpg) +Figure 6. Visualization Results. The first row is the interactive image, which demonstrates the interaction that the object can afford. The last row is the ground truth of 3D object affordance in the point cloud. The four columns on the left are the visual comparison results for distinct 3D object affordances in the Seen partition. The four columns on the right are the results in the Unseen partition. + +![](images/d2af1ca3d6a1399f4edc9c3898c54d77582ba08f36e6f0de79cd9fdda39af7dc.jpg) + +![](images/475f80aafcadb50818b19301c1c3f06561aa57ef916c17361447c7da55397623.jpg) +Figure 7. Ablation of JRA. The top row is the result of the anticipated affordance with/without JRA. The bottom row is a part of the cross-similarity matrix of a sample during the training process. + +nents, which means the 2D-3D affordance regions of objects do not match well. Besides, we visualize a part of the cross-similarity matrix between $\hat{\mathbf{F}}_p$ and $\hat{\mathbf{F}}_i$ in Fig. 7. In the initial stage, only a few analogous shape regions keep explicit correspondence, as training proceeds, the JRA maps the correlation among regions with lower similarity. Meanwhile, the model extracts explicit affordance progressively, and the affordance introduced into the optimization process reveals the corresponding affordance regions. + +![](images/112e6f64757112bf60e8460adea407ead05a29f6fe2b90b12f89e5c05d05fc26.jpg) +Figure 8. Ablation of ARM. The activation map and t-SNE [61] results with $(\boldsymbol{w})$ , without $(\boldsymbol{w} / \boldsymbol{o})$ ARM in both partitions. + +Effectiveness of ARM. The influence of ARM on evaluation metrics is also shown in Tab. 3. To visually evaluate the effectiveness of ARM, we employ GradCAM [55] to generate the activation map, and utilize t-SNE [61] to demonstrate the clustering of affordances in both partitions. The results are shown in Fig. 8. It shows that ARM makes the model focus on the interactive regions to excavate the interaction contexts, and enables the model to differentiate various affordances from interactions in both partitions. + +![](images/71aaa6b56d0d763a90f12a62142c759cc862dc7934c3ed2e7060ac825eef7fdd.jpg) +Figure 9. Same Image w.r.t. Multiple Point Clouds. (a) Same object category. (b) Different object categories, similar geometrics. (c) Different object categories and geometrics. + +![](images/95750804648c10db1e24627c1aed71739502066ba80d481cb494b2d8b2703b8b.jpg) +Figure 10. Same Point Cloud w.r.t. Multiple Images. Grounding affordance on the same point cloud with images that contain similar or disparate interactions. + +# 5.4. Performance Analysis + +Different Instances. We conduct experiments to verify whether the model could ground 3D affordance on instances from different sources: (i) using an image and different point clouds to infer respectively (Fig. 9 (a)). (ii) using multiple images and a single point cloud (Fig. 10). The results indicate that the model can anticipate 3D affordances on different instances with the same 2D interaction, and can also anticipate distinct 3D affordances on the same point cloud through different 2D interactions. Showing that the model maintains the mapping between object structures and affordances by learning from different 2D interactions. + +Different Object Categories. What will happen if the object category is different in the image and point cloud? To explore this issue, we perform experiments with mismatched object categories, shown in Fig. 9 (b) and (c). When objects are of different categories but with similar geometric primitives, e.g. the handle of "Mug" and the shoulder strap of "Bag", our model still properly infer the 3D affordance. However, when geometric structures are also dissimilar (Fig. 9 (c)), the model does not make random predictions. This indicates that the model maps cross-category invariance between affordance and geometries, which can be generalized to new instances of geometry. + +Multiplicity. One defining property of affordance is multiplicity: some points may correspond to multiple affordances. Fig. 11 shows that the model makes objects' multiple affordances compatible, and verifies the model does not + +![](images/40d0e1faa8821e7386c55c5b4ade70c48db66c317502d8b8d726aca61c1320c7.jpg) +Figure 11. Multiple Affordances. Some objects like "Bag" contains the region that corresponds to multiple affordances. + +![](images/47fb743a30689270c03b964f9a1639fbd1ec23e5296161b32bab46c2e57f20cd.jpg) + +![](images/74028dde10d445c71a2f715e6dc0635dc3b1e4f77e03f0c50c68de6d82c75422.jpg) +Figure 12. Real-Wrold. The first row is scanned by iPhone, the second row comes from [29], point clouds are sampled from them. +Figure 13. Failure Cases. Over prediction in the small regions. + +anticipate affordance by linking a region with specific affordances. Instead, it predicts by considering the presence of affordance-related interactions in the same region. + +Real World. To validate the model in real-world scenarios, we use an iPhone to scan objects and real-scan dataset [29] to test our model, shown in Fig. 12. It shows that our model exhibits a certain degree of generalization to the real world. Limitations. Our model exhibits over predictions for some small affordance regions, shown in Fig. 13. This may be due to the limited understanding of fine-grained geometric parts, and aligning the object features in a single scale may lead to overly large receptive fields. Our feature extension will refer [13, 27] to tackle this problem. + +# 6. Conclusion + +We present a novel setting for affordance grounding, which utilizes the 2D interactive semantics to guide the grounding of 3D object affordance, it has the potential to serve embodied systems when collaborating with multimodal grounding systems [30, 58, 70]. Besides, We collect the PIAD dataset as the first test bed for the proposed setting, it contains paired image-point cloud affordance data. Plus, we propose a novel framework to correlate affordance regions of objects that are from different sources and model interactive contexts to ground 3D object affordance. Comprehensive experiments on PIAD display the reliability of the setting, and we believe it could offer fresh insights and facilitate research in the affordance area. + +Acknowledgments This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002. + +# References + +[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. arXiv preprint arXiv:2209.09552, 2022. 6, 7 +[2] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3, 4 +[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE transactions on pattern analysis and machine intelligence, 41(3):740-757, 2018. 5 +[4] Yu-Wei Chao, Zhan Wang, Yugeng He, Jiaxuan Wang, and Jia Deng. Hico: A benchmark for recognizing human-object interactions in images. In Proceedings of the IEEE international conference on computer vision, pages 1017–1025, 2015. 5 +[5] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. Imlovenet: Misaligned image-supported registration network for low-overlap point cloud pairs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 6 +[6] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22(4):449-462, 2013. 1 +[7] Ching-Yao Chuang, Jiaman Li, Antonio Torralba, and Sanja Fidler. Learning to act properly: Predicting and explaining affordances from images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 975–983, 2018. 3 +[8] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2):722-739, 2021. 3 +[9] Barney Dalgarno and Mark JW Lee. What are the learning affordances of 3-d virtual environments? British Journal of Educational Technology, 41(1):10-32, 2010. 1 +[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 1, 3, 5, 6 +[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3 +[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3 + +[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, June 2023. 3, 8 +[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[17] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1 +[18] Maximilian Jaritz, Jiayuan Gu, and Hao Su. Multi-view pointnet for 3d scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3 +[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 5 +[20] David Inkyu Kim and Gaurav S Sukhatme. Semantic labeling of 3d point clouds with object affordance for robot manipulation. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 5578-5584. IEEE, 2014. 1 +[21] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 1, 3 +[22] Hema S Koppula and Ashutosh Saxena. Physically grounded spatio-temporal object affordances. In European Conference on Computer Vision, pages 831-847. Springer, 2014. 3 +[23] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 3 +[24] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 3 +[25] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhi-hong Zhu, and Yuexian Zou. G21: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3 +[26] Yaowei Li, Bang Yang, Xuxin Cheng, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. Unify, align and refine: Multi-level semantic alignment for radiology report generation. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3 +[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 8 +[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[29] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, June 2022. 3, 8 +[30] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 8 +[31] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5 +[32] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Transactions on Artificial Intelligence, 2022. 3 +[33] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot affordance detection. arXiv preprint arXiv:2106.14747, 2021. 3 +[34] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Grounded affordance from exocentric view. arXiv preprint arXiv:2208.13196, 2022. 3 +[35] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2252-2261, 2022. 3, 5 +[36] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 1 +[37] Maurice Merleau-Ponty. Phenomenologie de la perception (1945). Libraqire Gallimard, Paris, 1976. 2 +[38] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers in Neurorobotics, 14:26, 2020. 3 +[39] Jinpeng Mi, Song Tang, Zhen Deng, Michael Goerner, and Jianwei Zhang. Object affordance based multimodal fusion for natural human-robot interaction. Cognitive Systems Research, 54:128–137, 2019. 3 +[40] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric + +medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5 +[41] Niluthpol Chowdhury Mithun, Karan Sikka, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Rgb2lidar: Towards solving large-scale cross-modal visual localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 934-954, 2020. 3 +[42] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 3 +[43] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3, 4 +[44] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3 +[45] Bogdan Moldovan, Plinio Moreno, Martijn Van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE international conference on robotics and automation, pages 4373-4378. IEEE, 2012. 1 +[46] Austin Myers, Ching L Teo, Cornelia Fermüller, and Yiannis Aloimonos. Affordance detection of tool parts from geometric features. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 1374-1381. IEEE, 2015. 1 +[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 5 +[48] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3 +[49] Anh Nguyen, Dimitrios Kanoulas, Darwin G Caldwell, and Nikos G Tsagarakis. Detecting object affordances with convolutional neural networks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2765-2770. IEEE, 2016. 3 +[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 7(1-2):1-179, 2018. 1 +[51] François Osiurak and Arnaud Badets. Tool use and affordance: Manipulation-based versus reasoning-based approaches. Psychological review, 123(5):534, 2016. 1 +[52] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3, 4 + +[53] Md Atiqur Rahman and Yang Wang. Optimizing intersection-over-union in deep neural networks for image segmentation. In International symposium on visual computing, pages 234–244. Springer, 2016. 5 +[54] Anirban Roy and Sinisa Todorovic. A multi-scale cnn for affordance segmentation in rgb images. In European conference on computer vision, pages 186-201. Springer, 2016. 3 +[55] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7 +[56] Chris Shilling. The body in culture, technology and society. Sage, 2004. 2 +[57] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991. 5 +[58] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.8 +[59] Xun Tan, Xingyu Chen, Guowei Zhang, Jishiyu Ding, and Xuguang Lan. Mbdf-net: Multi-branch deep fusion network for 3d object detection. In Proceedings of the 1st International Workshop on Multimedia Computing for Urban Data, pages 9–17, 2021. 3, 6 +[60] Spyridon Thermos, Petros Daras, and Gerasimos Potamianos. A deep learning approach to object affordance segmentation. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2358-2362. IEEE, 2020. 3 +[61] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7 +[62] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 3 +[63] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In European Conference on Computer Vision, pages 421-436. Springer, 2014. 1 +[64] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 90-107. Springer, 2022. 3 +[65] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5 +[66] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance + +discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022. 1, 3 +[67] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 6 +[68] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 3, 6, 7 +[69] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement, 2023. 3 +[70] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 8 +[71] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73-91. Springer, 2022. 5 +[72] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 3 +[73] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 2 +[74] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 5 +[75] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4534-4543, 2020. 3 +[76] Linqing Zhao, Jiwen Lu, and Jie Zhou. Similarity-aware fusion network for 3d semantic segmentation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1585-1592. IEEE, 2021. 3 +[77] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 3 +[78] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Computing and Applications, 32(18):14321-14333, 2020. 3 +[79] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022. 3 +[80] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 3, 6 \ No newline at end of file diff --git a/2023/Grounding 3D Object Affordance from 2D Interactions in Images/images.zip b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14a5e479bab425b5bb743fe28142d03546d712b3 --- /dev/null +++ b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6715f1ba12b5914a86e7a698bcaca0b362ed46b4b61fafe6dd67d0d9e59ccc6 +size 822926 diff --git a/2023/Grounding 3D Object Affordance from 2D Interactions in Images/layout.json b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d26cedc6cf3d4ac7199c2045608b6f9c4bfd9f97 --- /dev/null +++ b/2023/Grounding 3D Object Affordance from 2D Interactions in Images/layout.json @@ -0,0 +1,9618 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 501, + 121 + ], + "type": "text", + "content": "Grounding 3D Object Affordance from 2D Interactions in Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": "Yuhang Yang" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": ", Wei Zhai" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": ", Hongchen Luo" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": ", Yang Cao" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": ", Jiebo Luo" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": ", Zheng-Jun Zha" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": " University of Science and Technology of China \n" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": " University of Rochester \n" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 506, + 185 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 187, + 473, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 187, + 473, + 200 + ], + "spans": [ + { + "bbox": [ + 121, + 187, + 473, + 200 + ], + "type": "text", + "content": "yyuhang@mail., wzhai056@mail., lhcl2@mail., forrest@ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 202, + 411, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 202, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 184, + 202, + 411, + 213 + ], + "type": "text", + "content": "jluo@cs.rochester.edu zhazj@ustc.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "spans": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 278, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 290, + 579 + ], + "type": "text", + "content": "Grounding 3D object affordance seeks to locate objects' \"action possibilities\" regions in the 3D space, which serves as a link between perception and operation for embodied agents. Existing studies primarily focus on connecting visual affordances with geometry structures, e.g., relying on annotations to declare interactive regions of interest on the object and establishing a mapping between the regions and affordances. However, the essence of learning object affordance is to understand how to use it, and the manner that detaches interactions is limited in generalization. Normally, humans possess the ability to perceive object affordances in the physical world through demonstration images or videos. Motivated by this, we introduce a novel task setting: grounding 3D object affordance from 2D interactions in images, which faces the challenge of anticipating affordance through interactions of different sources. To address this problem, we devise a novel Interaction-driven 3D Affordance Grounding Network (IAG), which aligns the region feature of objects from different sources and models the interactive contexts for 3D object affordance grounding. Besides, we collect a Point-Image Affordance Dataset (PIAD) to support the proposed task. Comprehensive experiments on PIAD demonstrate the reliability of the proposed task and the superiority of our method. The project is available at https://github.com/yyvhang/IAGNet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 621, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 288, + 694 + ], + "type": "text", + "content": "The term \"affordance\" is described as \"opportunities of interaction\" by J. Gibson [14]. Grounding 3D object affordance aims to comprehend the interactive regions of objects in 3D space, which is not only simply to predict which interaction an object affords, but also to identify specific points on the object that could support the interaction. It consti" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 239, + 536, + 369 + ], + "blocks": [ + { + "bbox": [ + 317, + 239, + 536, + 369 + ], + "lines": [ + { + "bbox": [ + 317, + 239, + 536, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 536, + 369 + ], + "type": "image", + "image_path": "8247bf0739f3d10b9f0296881df42b876a940e11b48883728ea15b3f7f6866b7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 372, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 416 + ], + "type": "text", + "content": "Figure 1. Grounding Affordance from Interactions. We propose to ground 3D object affordance through 2D interactions. Given an object point cloud with an interactive image, grounding the corresponding affordance on the 3D object." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 498 + ], + "type": "text", + "content": "tutes a link between perception and operation for embodied agents, which has the potential to serve numerous practical applications, e.g. action prediction [21, 63], robot manipulation [36, 45, 51], imitation learning [17, 50], and augmented/virtual reality [6, 9]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "So far, the paradigm of perceiving 3D object affordance has several branches. One of them involves establishing an explicit mapping between affordance categories and geometry structures [10, 20, 46, 66], based on visual appearance. However, affordance is dynamic and multiple, these geometric-specific manners have limited generalization for unseen structures. Besides, locking the geometry with a specific affordance category may lead to the anticipated region being inconsistent with its affordance when objects possess multiple similar geometrics, resulting in affordance regional confusion. Another paradigm is based on reinforcement learning, which puts the agent in 3D synthetic scenarios to interact with several objects actively, taking the reward mechanism to optimize the whole process [48]. While this type of approach transitions agents from passive recognition to active reasoning, it needs repeated attempts in a huge search space when meeting a novel structure, and is therefore time-consuming." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 143, + 713 + ], + "type": "text", + "content": "\\*Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10905" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 70, + 276, + 219 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 276, + 219 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 276, + 219 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 276, + 219 + ], + "type": "image", + "image_path": "4cc3bd9669f23af051f7af326d58d307576221a3458726e494f5e37aa5a0c742.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 282, + 71, + 512, + 220 + ], + "blocks": [ + { + "bbox": [ + 282, + 71, + 512, + 220 + ], + "lines": [ + { + "bbox": [ + 282, + 71, + 512, + 220 + ], + "spans": [ + { + "bbox": [ + 282, + 71, + 512, + 220 + ], + "type": "image", + "image_path": "8539b518ee166ef2e0c94c58af4233460fc128c48b8e84c8173cb0eeb30dd9de.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 545, + 275 + ], + "type": "text", + "content": "Figure 2. Motivation. (a) Clues to correlate regions of objects from different sources. Objects with the same category possess a similar combination scheme to meet certain affordance properties, it is a similar inherent relation among different instances. And some structures hint at these properties exhibit closer representations with higher similarity in the latent space. (b) Object affordance may be affected by dynamic factors, such as the position of the object itself in the scene, other objects in the scene, and the body part of the interacting subject. These factors can be decomposed into object-subject and object-scene interaction contexts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 294, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 462 + ], + "type": "text", + "content": "These limitations motivate us to explore an affordance-compatible learning paradigm. Typically, humans can infer object affordance in the 3D physical world by watching images or videos that demonstrate the interactions. Some studies in cognitive science [37, 56] point out the existence of \"body image\" in human cognition, which claims that humans have a perceptual experience of the objects they see, thus facilitating their ability to organize the perception [73] and operate novel objects. Hence, human-object interaction conveys the perception that object structure could perform certain affordance, which is a crucial clue to reason object affordance. In light of this, we present a novel task setting: grounding 3D object affordance from 2D interactions in images, which is shown in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 288, + 714 + ], + "type": "text", + "content": "This challenging task includes several essential issues that should be properly addressed. 1) Alignment ambiguity. To ground 3D object affordance from 2D source interactions, the premise is to correspond the regions of the object in different sources. The object in 2D demonstrations and the 3D object we face are usually derived from different physical instances in different locations and scales. This discrepancy may lead to confusion when matching the affordance regions, causing alignment ambiguity. While objects are commonly designed to satisfy certain needs of human beings, so the same category generally follows a similar combination scheme of object components to meet certain affordances, and these affordances are hinted at by some structures (Fig. 2 (a)). These invariant properties are across instances and could be utilized to correlate object regions from different sources. 2) Affordance ambiguity. Affordance has properties of dynamic and multiplicity, which means the object affordance may change according to the situation, the same part of an object could afford multiple interactions, as shown in Fig. 2 (b), \"Chair\" affords \"Sit\" or \"Move\" depends on the human actions, \"Mug\" af" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 294, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 366 + ], + "type": "text", + "content": "fords \"Wrapgrasp\" or \"Contain\" according to the scene context, these properties may make ambiguity when extracting affordance. However, these dynamic factors for affordance extraction can be decomposed into the interaction between subject-object and object-scene. Modeling these interactions is possible to extract explicit affordance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 366, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 582 + ], + "type": "text", + "content": "To address these issues, we propose the Interaction-driven 3D Affordance Grounding Network (IAG) to align object region features from different sources and model interaction contexts to reveal affordance. In detail, it contains two sub-modules, one is Joint Region Alignment Module (JRA), which is devised to eliminate alignment ambiguity. It takes the relative difference in dense cross-similarity to refine analogous shape regions and employs learnable layers to map the invariant combination scheme, taking them to match local regions of objects. For affordance ambiguity, the other one Affordance Revealed Module (ARM) takes the object representation as a shared factor and jointly models its interaction contexts with the affordance-related factors to reveal explicit affordance. Moreover, we collect Point-Image Affordance Dataset (PIAD) that contains plenty of paired image-point cloud affordance data and make a benchmark to support the model training and evaluation on the proposed setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "The contributions are summarized as follows: 1) We introduce grounding 3D object affordance through the 2D interactions, which facilitates the generalization to 3D object affordance perception. 2) We propose the IAG framework, which aligns the region feature of objects from different sources and jointly models affordance-related interactions to locate the 3D object affordance. 3) We collect a dataset named PIAD to support the proposed task setting, which contains paired image-point cloud affordance data. Besides, we establish a benchmark on PIAD, and the experiments on PIAD exhibit the reliability of the method and the setting." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10906" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 139, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 139, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 139, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 169, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 169, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 169, + 105 + ], + "type": "text", + "content": "2.1. Affordance Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 111, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 111, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 111, + 289, + 555 + ], + "type": "text", + "content": "At present, the affordance area has made achievements in multiple tasks (Tab. 1). Some works are devoted to detecting the affordance region from 2D sources i.e. images and videos [7, 11, 25, 33, 54, 60, 78], and there are also some studies accomplish this task with the assistance of natural language [32, 38, 39]. They seek to detect or segment objects that afford the action in 2D data, but cannot perceive the specific parts corresponding to the affordance. Thus, another type of method brings a leap to locate the affordance region of objects from 2D sources [12, 35, 47, 34]. However, the affordance knowledge derived from 2D sources is hard to extrapolate specific interactive locations of objects in the 3D environment. With several 3D object datasets proposed [13, 29, 44], some researchers explore grounding object affordance from 3D data [10, 43, 66, 48]. Such methods directly map the connection between semantic affordances and 3D structures, detach the real interaction, and may deter the generalization ability: structures that do not map to a specific affordance are usually hard to generalize through this type of method. T. Nagarajan et al. [48] give a fresh mind, taking the reinforcement learning to make agents actively interact with the 3D synthetic scenarios, while it requires repeated attempts by the agents in a given search space, and is time-consuming. In robotics, affordance is utilized to provide priors for object manipulation and achieve considerable results, especially the articulated 3D objects [42, 64, 79]. Several methods utilize the 2.5D data e.g. RGB-D image to understand the object affordance and serve the manipulation task [21, 22, 23, 49], for these methods, the image and depth information are corresponding in spatial and need to be collected in the same scene. In contrast, our task focuses on grounding 3D object affordance from 2D interactions, in which the 2D interactions provide direct clues to excavate object affordance efficiently and make the affordance could generalize to some unseen structures, and the 2D-3D data is collected from different sources, freeing the constraints on spatial correspondence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 562, + 269, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 269, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 269, + 576 + ], + "type": "text", + "content": "2.2. Image-Point Cloud Cross-Modal Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "The combination of point cloud and image data can capture both semantic and geometric information, enabling cross-modal learning among them has great potential application value in scenarios like autopilot [8]. For this type of work, aligning features is the premise for completing downstream tasks e.g. detection, segmentation, and registration. It is aimed at establishing correspondences between instances from different modalities [2, 26, 69], either spatial or semantic. To achieve this, many methods use camera intrinsic to correspond spatial position of pixels and points, then align per pixel-point feature or fuse the raw" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 95, + 547, + 186 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 547, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 547, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 547, + 91 + ], + "type": "text", + "content": "Table 1. Affordance Learning. Various works for several tasks in the affordance community." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 95, + 547, + 186 + ], + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 186 + ], + "type": "table", + "html": "
WorkInputOutputTask
[11, 60, 78]Image/Video2D MaskDetection
[32, 38, 39]Image,Text2D MaskDetection
[21, 23, 49]RGBDHeatmap & ActionManipulation
[42, 64, 79]Point CloudHeatmap & ActionManipulation
[12, 35, 47]Image/Video2D HeatmapGrounding
[10, 43, 66]Point Cloud3D HeatmapGrounding
", + "image_path": "c7e8c35d4d199dc50b7c69ccb0ff6efe9dfd3816589b8610805918f726a2d5f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 205, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 338 + ], + "type": "text", + "content": "data [24, 59, 62, 68, 77, 80]. Some works utilize depth information to project image features into 3D space and then fuse them with point-wise features [18, 41, 72, 75, 76]. The above methods rely on the spatial prior information to align features, while in our task, images and point clouds are collected from distinct physical instances, there are no priors on camera pose or intrinsics, also no corresponding depth information of the image. Therefore, we align the object region features that are derived from different sources in the feature space with the assistance of the correlation invariance between affordances and appearance structures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 346, + 362, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 362, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 362, + 358 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 366, + 373, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 373, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 373, + 378 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": "Our goal is to anticipate the affordance regions on the point cloud that correspond to the interaction in the image. Given a sample " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\{P, I, \\mathcal{B}, y\\}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " is a point cloud with the coordinates " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "P_{c} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " and the affordance annotation " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "P_{label} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " is an RGB image, " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{B_{sub}, B_{obj}\\}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " denotes the bounding box of the subject and object in " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " is the affordance category label. The IAG (Fig. 3) capture localized features " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}} \\in \\mathbb{R}^{C \\times H' \\times W'}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{p} \\in \\mathbb{R}^{C \\times N_{p}}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " of the image and point cloud by two feature extractors ResNet [16] and PointNet++ [52]. Then, utilizing " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "B_{obj} / B_{sub}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " to locate the object/subject region in " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": ", outside the " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "B_{obj}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "B_{sub}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " is the scene mask " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "M_{sce}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": ", use ROI-Align [15] to get the object, subject, and scene features " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i}, \\mathbf{F}_{s}, \\mathbf{F}_{e} \\in \\mathbb{R}^{C \\times H_{1} \\times W_{1}}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": ", and reshape them to " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times N_{i}}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "N_{i} = H_{1} \\times W_{1}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": "). Next, the JRA module takes " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i}, \\mathbf{F}_{p}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " as input and jointly mines the correspondence to align them, obtaining the joint feature " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{j}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " (Sec. 3.2). Following, the ARM module takes " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{j}, \\mathbf{F}_{s}, \\mathbf{F}_{e}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " to reveal the affordance " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\alpha}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " through cross-attention mechanism (Sec. 3.3). Ultimately, " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\alpha}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{j}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " are sent to the decoder to compute the affordance logits " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " and the final 3D object affordance " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " (Sec. 3.4). The process is expressed as " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}, \\hat{y} = f_{\\theta}(P_{c}, I, \\mathcal{B}; \\theta)" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " denotes the network and " + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 385, + 547, + 665 + ], + "type": "text", + "content": " is the parameter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 479, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 479, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 479, + 684 + ], + "type": "text", + "content": "3.2. Joint Region Alignment Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "The JRA module calculates the high-dimensional dense similarity in feature space to approximate analogous shape" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10907" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 542, + 210 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 542, + 210 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 542, + 210 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 542, + 210 + ], + "type": "image", + "image_path": "4930cba0022500dd36f1fbd31252ca1d519670a9d66912c97e0111f5293fdea2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": "Figure 3. Method. Overview of Interaction-driven 3D Affordance Grounding Network (IAG), it firstly extracts localized features " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i, \\mathbf{F}_p" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": " respectively, then takes JRA (Sec. 3.2) to align them and get the joint feature " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": ". Next, ARM (Sec. 3.3) utilizes " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": " to reveal affordance " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\alpha}" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s, \\mathbf{F}_e" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": " by cross-attention. Eventually, " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\alpha}" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": " are sent to the decoder (Sec. 3.4) to obtain the final results " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 213, + 547, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 286, + 361 + ], + "type": "text", + "content": "regions that contain relatively higher similarity [2]. Meanwhile, to map the inherent combination scheme of object components, the JRA models the inherent relations between modality-specific regions and takes it to drive the alignment of other regions by a transformer-based joint-attention " + }, + { + "bbox": [ + 46, + 266, + 286, + 361 + ], + "type": "inline_equation", + "content": "f_{\\delta}" + }, + { + "bbox": [ + 46, + 266, + 286, + 361 + ], + "type": "text", + "content": ". With the network trained to capture the correspondences among regions from different sources, the alignment is performed implicitly during the optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "content": "Initially, " + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "content": " are projected into a feature space by shared convolution layers " + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "inline_equation", + "content": "f_{\\sigma}" + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "content": ", obtaining the features " + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{C \\times N_p}, \\mathbf{I} \\in \\mathbb{R}^{C \\times N_i}" + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "content": ". Then, to correlate the analogous shape regions, the dense cross-similarity is calculated between each region of " + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 361, + 287, + 421 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 429, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 429, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 78, + 429, + 287, + 460 + ], + "type": "interline_equation", + "content": "\\varphi_ {i, j} = \\frac {e ^ {\\left(\\mathbf {P} _ {i} , \\mathbf {I} _ {j}\\right)}}{\\sum_ {i = 1} ^ {N _ {p}} \\sum_ {j = 1} ^ {N _ {i}} e ^ {\\left(\\mathbf {P} _ {i} , \\mathbf {I} _ {j}\\right)}}, \\varphi \\in \\mathbb {R} ^ {N _ {p} \\times N _ {i}}, \\tag {1}", + "image_path": "12a8e7a612d2008c88745391f86d321ba4d7ef9d61857925960d065ef06cca86.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\varphi_{i,j}" + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": " denotes the cross-similarity between the i-th region of " + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": " and the j-th region of " + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": ". Taking the relative difference in " + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": " to refine and correlate analogous regions in " + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": ", next, applying self-attention layers " + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "inline_equation", + "content": "f_{i}, f_{p}" + }, + { + "bbox": [ + 46, + 468, + 287, + 539 + ], + "type": "text", + "content": " to model the intra-structural inherent relation of objects in respective modalities, the process is expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 547, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 547, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 102, + 547, + 287, + 562 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {P}} = f _ {p} \\left(\\mathbf {I} \\cdot \\varphi^ {T}\\right), \\bar {\\mathbf {I}} = f _ {i} \\left(\\mathbf {P} \\cdot \\varphi\\right), \\tag {2}", + "image_path": "005f5916812ff29799e13bdae73dab25dcb3a32812a47a76f6cb1776669d16b6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{P}}\\in \\mathbb{R}^{C\\times N_p},\\bar{\\mathbf{I}}\\in \\mathbb{R}^{C\\times N_i}" + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": ". Following, we perform a joint-attention on features with structural relevance to map the similar inherent combination scheme and drive the alignment of the remaining regions. The process is formulated as: " + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j = f_{\\delta}[\\bar{\\mathbf{P}},\\bar{\\mathbf{I}}]" + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j\\in \\mathbb{R}^{C\\times (N_p + N_i)}" + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": " denotes joining the image feature sequence and point cloud feature sequence into a whole one, and " + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": " denotes the joint representation. The " + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": " is sent to the affordance revealed module to excavate interaction contexts with " + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s,\\mathbf{F}_e" + }, + { + "bbox": [ + 46, + 569, + 287, + 713 + ], + "type": "text", + "content": ". And the affordance knowledge revealed by the interaction contexts could mutually optimize the alignment process during training (see Sec. 3.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 265, + 465, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 265, + 465, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 465, + 277 + ], + "type": "text", + "content": "3.3. Affordance Revealed Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 392 + ], + "type": "text", + "content": "In general, the dynamic factors related to affordance are reflected in the interaction, affecting the extraction of affordance representation, and the interaction is mostly considered to exist between humans and objects. However, the interaction between objects and scenes (including other objects) may also affect affordance [43]. To consider both of these factors, the ARM module utilizes the cross-attention technique to extract these interaction contexts respectively and jointly embed them to reveal explicit object affordance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "text", + "content": "Specifically, " + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "text", + "content": " is projected to form the shared query " + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{F}_j\\mathbf{W}_1,\\mathbf{F}_s" + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_e" + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "text", + "content": " are projected to form different keys and values " + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{1 / 2} = \\mathbf{F}_s\\mathbf{W}_{2 / 3},\\mathbf{V}_{1 / 2} = \\mathbf{F}_e\\mathbf{W}_{4 / 5}" + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{1\\sim 5}" + }, + { + "bbox": [ + 305, + 392, + 545, + 452 + ], + "type": "text", + "content": " are projection weights. Then cross-attention aggregates them to excavate interaction contexts, expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 336, + 461, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 461, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 336, + 461, + 545, + 478 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Theta} _ {1 / 2} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} ^ {T} \\cdot \\mathbf {K} _ {1 / 2} / \\sqrt {d}\\right) \\cdot \\mathbf {V} _ {1 / 2} ^ {T}, \\tag {3}", + "image_path": "6edd5612779dd7cc7995f41ab9664c15d873360559ea359d59d172b2a6c6c31a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{d \\times (N_p + N_i)}" + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{1/2}" + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{1/2} \\in \\mathbb{R}^{d \\times N_i}" + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": " is the dimension of projection, " + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\Theta_1, \\Theta_2" + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": " indicate the excavated interaction contexts. Then, jointly embedding " + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\Theta_1" + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\Theta_2" + }, + { + "bbox": [ + 304, + 481, + 545, + 530 + ], + "type": "text", + "content": " to reveal affordance, expressed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 536, + 545, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 536, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 344, + 536, + 545, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\alpha} = f _ {\\xi} (\\boldsymbol {\\Theta} _ {1}, \\boldsymbol {\\Theta} _ {2}), \\mathbf {F} _ {\\alpha} \\in \\mathbb {R} ^ {C \\times (N _ {p} + N _ {i})}, \\tag {4}", + "image_path": "e4a6c0da6b8de6036ddaa16308d837066968d3458e6c34ec11f2e10fa0861ea2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 559, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 559, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\alpha}" + }, + { + "bbox": [ + 304, + 559, + 545, + 594 + ], + "type": "text", + "content": " is the joint affordance representation, it is split in the decoder to compute final results, " + }, + { + "bbox": [ + 304, + 559, + 545, + 594 + ], + "type": "inline_equation", + "content": "f_{\\xi}" + }, + { + "bbox": [ + 304, + 559, + 545, + 594 + ], + "type": "text", + "content": " denotes the concatenation, followed by a convolution layer to fuse the feature." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 601, + 456, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 456, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 456, + 613 + ], + "type": "text", + "content": "3.4. Decoder and Loss Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " is split to " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p\\in \\mathbb{R}^{C\\times N_p}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i\\in \\mathbb{R}^{C\\times N_i}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\alpha}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " is split to " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{p\\alpha}\\in \\mathbb{R}^{C\\times N_p}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i\\alpha}\\in \\mathbb{R}^{C\\times N_i}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " in the decoder. Pooling " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{p\\alpha}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i\\alpha}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " respectively and concatenate them to compute the affordance logits " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " is upsampled to " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times N}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " by Feature Propagation Layers (FP) [52], and the 3D object affordance " + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}" + }, + { + "bbox": [ + 304, + 618, + 545, + 693 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 341, + 699, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 699, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 341, + 699, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\hat {\\phi} = f _ {\\phi} \\left(\\mathbf {F P} \\left(\\hat {\\mathbf {F}} _ {p}\\right) \\odot \\Gamma \\left(\\mathbf {F} _ {p \\alpha}\\right)\\right), \\hat {\\phi} \\in \\mathbb {R} ^ {N \\times 1}, \\tag {5}", + "image_path": "c352b0b92704818d61acbebab750b559d688c028dd8358b26f7bfc95555245e6.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10908" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 59, + 531, + 167 + ], + "blocks": [ + { + "bbox": [ + 65, + 59, + 531, + 167 + ], + "lines": [ + { + "bbox": [ + 65, + 59, + 531, + 167 + ], + "spans": [ + { + "bbox": [ + 65, + 59, + 531, + 167 + ], + "type": "image", + "image_path": "2ae5d64006c44d4a3924455e449838133da73909bd8fbf5077308aa609f61a97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 547, + 213 + ], + "type": "text", + "content": "Figure 4. Properties of the PIAD dataset. (a) Data pairs in the PIAD, the red region in point clouds is the affordance annotation. (b) Distribution of the image data. The horizontal axis represents the category of affordance, the vertical axis represents quantity, and different colors represent different objects. (c) Distribution of the point cloud data. (d) The ratio of images and point clouds in each affordance class. It shows that images and point clouds are not fixed one-to-one pairing, they can form multiple pairs. Please refer to supp. for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "text", + "content": " is an output head, " + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{FP}" + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "text", + "content": " is the upsample layer, " + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "text", + "content": " denotes pooling " + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{p\\alpha}" + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "text", + "content": " and expand it to the shape of " + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times N}" + }, + { + "bbox": [ + 46, + 232, + 286, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": "The total loss comprises three parts: " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{HM}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " computes the cross-entropy loss between " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ", it supervises the extraction of " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\alpha}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " and implicitly optimizes the alignment process. To enable the network focus on the alignment of affordance regions, we apply the KL Divergence (KLD) [3] to constrain the distribution between " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i\\alpha}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ", formulated as " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL} = (\\mathbf{F}_{i\\alpha}||\\hat{\\mathbf{F}}_i)" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ". The reason is that " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i\\alpha}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " exhibits the affordance distribution of each object region in the image, and the affordance-related regions keep more significant features. Constraining the feature distribution of " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " to enhance the affordance region features in " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ", and with the region correlations established by the alignment process, " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " also tends to exhibit this property, similar to distillation [19, 71]. Which makes the alignment and affordance extraction optimize mutually. " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{HM}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " is a focal loss [28] combined with a dice loss [40], it is calculated by " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "inline_equation", + "content": "P_{label}" + }, + { + "bbox": [ + 46, + 256, + 287, + 472 + ], + "type": "text", + "content": ", which supervise the point-wise heatmap on point clouds. Eventually, the total loss is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 484, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 484, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 86, + 484, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {C E} + \\lambda_ {2} \\mathcal {L} _ {K L} + \\lambda_ {3} \\mathcal {L} _ {H M}, \\tag {6}", + "image_path": "d952789058befd6c701fac257ba3f3a2d0a702b4937835c93b3ed95a65797cf7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "text", + "content": " are hyper-parameters to balance the total loss. See more details in supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 537, + 101, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 101, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 101, + 550 + ], + "type": "text", + "content": "4. Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 714 + ], + "type": "text", + "content": "Collection Details. We collect Point-Image Affordance Dataset (PIAD), which contains paired image-point cloud affordance data. The point clouds are mainly collected from 3D-AffordanceNet [10], including the point cloud coordinates and affordance annotation. Images are mainly collected from HICO [4], AGD20K [35], and websites with free licenses. The collection criteria is that the image should demonstrate interactions that the object in the point cloud could afford. For example, if the object point cloud is a \"Chair\", it affords \"Sit\", then the image should depict a subject (usually humans) sitting on a chair. The final dataset comprises 7012 point clouds and 5162 images, spanning 23 object classes and 17 affordance categories. Notably, ob" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 232, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 268 + ], + "type": "text", + "content": "jects in the images and point clouds do not sample from the same physical instance, but they belong to the same object category. Paired examples are shown in Fig. 4 (a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 269, + 546, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 447 + ], + "type": "text", + "content": "Annotation Details. For point clouds, each point may afford one or multiple interactions, e.g. one annotation is a matrix of (2048, 17), 2048 is the point number, 17 represents the number of affordance types. Each element in the matrix indicates the probability of a point affording a specific affordance. Meanwhile, images are annotated with bounding boxes of the interactive subject and object, as well as an affordance category label. In our task, we only use the heatmap in the point cloud annotation that corresponds to the affordance of the image for training, resulting in a matrix of (2048, 1). By doing so, the affordance category is detached from the point cloud during inference, and the anticipation of the 3D object affordance category only relies on 2D interactions. As a result, different affordances can be anticipated on a point cloud through distinct interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 448, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 579 + ], + "type": "text", + "content": "Statistic Analysis. Since images and point clouds are sampled from different instances, they do not need a fixed one-to-one pairing, one image could be paired with multiple point clouds and the count of them is not strictly consistent. Fig. 4 (b) and (c) show the count and distribution of affordances in images and point clouds. Fig. 4 (d) illustrates the ratio of images and point clouds in each affordance category. PIAD has two partitions: Seen and Unseen. In Seen, both objects and affordances in the training and testing sets are consistent, while in Unseen, some objects or affordances in the testing set do not exist in the training set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 590, + 388, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 388, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 388, + 604 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 610, + 419, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 419, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 419, + 624 + ], + "type": "text", + "content": "5.1. Benchmark Setting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 688 + ], + "type": "text", + "content": "Evaluation Metrics. To provide a comprehensive and effective evaluation, we compare serval advanced works in the affordance area [10, 47, 74] and finally chose four evaluation metrics: AUC [31], aIOU [53], SIMilarity [57] and Mean Absolute Error [65] to benchmark the PIAD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Modular Baselines. Since there are no prior works using paired image-point cloud data to ground 3D object" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10909" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 106, + 549, + 224 + ], + "blocks": [ + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "type": "text", + "content": "Table 2. Comparison Results on PIAD. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. \"Green\" and \"Orange\" indicate two types of the comparative modular baselines. \"Base.\" represents the baseline. " + }, + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "type": "text", + "content": " denotes the relative improvement of our method over other methods. AUC and aIOU are shown in percentage." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 106, + 549, + 224 + ], + "lines": [ + { + "bbox": [ + 48, + 106, + 549, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 549, + 224 + ], + "type": "table", + "html": "
MetricsBase.MBDF [59]PMF [80]FRCNN [68]ILN [5]PFusion [67]XMF [1]Ours
SeenAUC ↑68.4974.88◇13.3%75.05◇13.0%76.05◇11.5%75.84◇11.9%77.50◇9.5%78.24◇8.4%84.85 ± 0.3
aIOU ↑6.859.34◇119.5%10.13◇102.4%11.97◇71.3%11.52◇78.0%12.31◇66.6%12.94◇58.5%20.51 ± 0.7
SIM ↑0.3670.415◇31.3%0.425◇28.2%0.429◇27.0%0.427◇27.6%0.432◇26.1%0.441◇23.6%0.545 ± 0.02
MAE ↓0.1520.143◇31.4%0.141◇30.5%0.136◇27.9%0.137◇28.4%0.135◇27.4%0.127◇22.8%0.098 ± 0.01
UnseenAUC ↑57.3458.23◇23.3%60.25◇19.2%61.92◇16.0%59.69◇20.3%61.87◇16.1%62.58◇14.8%71.84 ± 1.8
aIOU ↑3.954.22◇88.4%4.67◇70.2%5.12◇55.2%4.71◇68.8%5.33◇49.1%5.68◇39.9%7.95 ± 0.8
SIM ↑0.3180.325◇8.3%0.330◇6.6%0.332◇6.0%0.325◇8.3%0.330◇6.6%0.342◇2.9%0.352 ± 0.03
MAE ↓0.2350.213◇40.3%0.211◇39.8%0.195◇34.8%0.207◇38.6%0.193◇34.2%0.188◇32.4%0.127 ± 0.01
", + "image_path": "c464df2efdb1d25281ad68f76789123161f837fc7201f233bea0d54bbac8e98b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 236, + 278, + 357 + ], + "blocks": [ + { + "bbox": [ + 50, + 236, + 278, + 357 + ], + "lines": [ + { + "bbox": [ + 50, + 236, + 278, + 357 + ], + "spans": [ + { + "bbox": [ + 50, + 236, + 278, + 357 + ], + "type": "image", + "image_path": "f9e82174288403b9f88e23bcf64ecb47e40ce83150a93739f94932c250539725.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 358, + 288, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 288, + 403 + ], + "type": "text", + "content": "Figure 5. Paradigm Comparison. (a) Ground truth. (b) The results of 3D-AffordanceNet [10]. (c) Our results. The top row shows the regional confusion case in Seen, and the bottom row displays the generalization in Unseen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 651 + ], + "type": "text", + "content": "affordance. For a thorough comparison of our method, we select several advanced image-point cloud cross-modal works as modular baselines. Methods of comparison are broadly divided into two types, one is that utilizes camera intrinsics to facilitate cross-modal feature alignment or fusion, i.e. MBDF-Net (MBDF) [59], PMF [80] and FusionRCNN (FRCNN) [68], for this type of method, we remove the step of using intrinsic parameters to align raw data or features to explore their effectiveness on PIAD. Another type performs feature alignment directly in the feature space, without relying on camera intrinsic, i.e. ImLoveNet (ILN) [5], PointFusion (PFusion) [67] and XMFnet (XMF) [1]. To ensure a fair comparison, all methods use the same feature extractors, with the only variation coming from cross-modal alignment or fusion block. For the Baseline, we directly concatenate the features that are output from extractors, regarding the concatenation as the cross-fusion block. More details about baselines are provided in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 165, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 165, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 165, + 672 + ], + "type": "text", + "content": "5.2. Comparison Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "The comparison results of evaluation metrics are shown in Tab. 2. As can be seen, our method outperforms the compared baselines, across all metrics in both partitions." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 266, + 544, + 384 + ], + "blocks": [ + { + "bbox": [ + 305, + 240, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 262 + ], + "type": "text", + "content": "Table 3. Ablation Study. We investigate the improvement of JRA and ARM on the model performance based on the baseline." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 266, + 544, + 384 + ], + "lines": [ + { + "bbox": [ + 308, + 266, + 544, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 544, + 384 + ], + "type": "table", + "html": "
JRAARMAUCaIOUSIMMAE
Seen69.928.850.4270.132
✓80.2914.310.4950.121
✓78.6713.950.4750.126
✓✓85.1621.200.5640.088
Unseen59.144.050.3380.202
✓66.256.270.3630.159
✓65.795.990.3580.162
✓✓73.698.700.3830.117
", + "image_path": "1d527988452bc1ae190a3567a219a8670d7005293fce4ca17e52d0a99dd984d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 407, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 634 + ], + "type": "text", + "content": "Besides, to display the limitation of methods that lock geometrics with specific semantic affordance categories, we conduct an experiment to compare one of these methods [10] with ours. As shown in Fig. 5, the top row indicates that the result of this method exhibits regional confusion, e.g. the region where the chair could be moved or sat is geometrically rectangular, and it directly anticipates results on these similar geometries, inconsistent with affordance property. Plus, the bottom row shows that directly establishing a link between object structures and affordances may fail to anticipate correct 3D affordance in Unseen. In contrast, our method anticipates precise results by mining affordance clues provided by 2D interactions. Additionally, visual comparative results of our method and other baselines are shown in Fig. 6. As can be seen, the comparative baselines could anticipate some 3D object affordance under our setting, but in comparison, our method obviously achieves better results, which validates the rationality of our setting, and also demonstrates the superiority of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 646, + 400, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 400, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 400, + 658 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Effectiveness of JRA. Tab. 3 reports the impact on evaluation metrics of JRA. Fig. 7 provides visual comparison results. It shows that without the JRA, the result is anticipated over the entire region that contains the interactive compo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10910" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 59, + 293, + 364 + ], + "blocks": [ + { + "bbox": [ + 65, + 59, + 293, + 364 + ], + "lines": [ + { + "bbox": [ + 65, + 59, + 293, + 364 + ], + "spans": [ + { + "bbox": [ + 65, + 59, + 293, + 364 + ], + "type": "image", + "image_path": "a151ae1999e8e56175ae92d6788ee6eff41373c50d0619773e3c11b280511ed6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 364, + 547, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 364, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 547, + 398 + ], + "type": "text", + "content": "Figure 6. Visualization Results. The first row is the interactive image, which demonstrates the interaction that the object can afford. The last row is the ground truth of 3D object affordance in the point cloud. The four columns on the left are the visual comparison results for distinct 3D object affordances in the Seen partition. The four columns on the right are the results in the Unseen partition." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 59, + 519, + 364 + ], + "blocks": [ + { + "bbox": [ + 294, + 59, + 519, + 364 + ], + "lines": [ + { + "bbox": [ + 294, + 59, + 519, + 364 + ], + "spans": [ + { + "bbox": [ + 294, + 59, + 519, + 364 + ], + "type": "image", + "image_path": "d2af1ca3d6a1399f4edc9c3898c54d77582ba08f36e6f0de79cd9fdda39af7dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 410, + 276, + 550 + ], + "blocks": [ + { + "bbox": [ + 58, + 410, + 276, + 550 + ], + "lines": [ + { + "bbox": [ + 58, + 410, + 276, + 550 + ], + "spans": [ + { + "bbox": [ + 58, + 410, + 276, + 550 + ], + "type": "image", + "image_path": "475f80aafcadb50818b19301c1c3f06561aa57ef916c17361447c7da55397623.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 552, + 288, + 585 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 585 + ], + "type": "text", + "content": "Figure 7. Ablation of JRA. The top row is the result of the anticipated affordance with/without JRA. The bottom row is a part of the cross-similarity matrix of a sample during the training process." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "nents, which means the 2D-3D affordance regions of objects do not match well. Besides, we visualize a part of the cross-similarity matrix between " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " in Fig. 7. In the initial stage, only a few analogous shape regions keep explicit correspondence, as training proceeds, the JRA maps the correlation among regions with lower similarity. Meanwhile, the model extracts explicit affordance progressively, and the affordance introduced into the optimization process reveals the corresponding affordance regions." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 408, + 531, + 558 + ], + "blocks": [ + { + "bbox": [ + 319, + 408, + 531, + 558 + ], + "lines": [ + { + "bbox": [ + 319, + 408, + 531, + 558 + ], + "spans": [ + { + "bbox": [ + 319, + 408, + 531, + 558 + ], + "type": "image", + "image_path": "112e6f64757112bf60e8460adea407ead05a29f6fe2b90b12f89e5c05d05fc26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "type": "text", + "content": "Figure 8. Ablation of ARM. The activation map and t-SNE [61] results with " + }, + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{w})" + }, + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "type": "text", + "content": ", without " + }, + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{w} / \\boldsymbol{o})" + }, + { + "bbox": [ + 305, + 563, + 545, + 586 + ], + "type": "text", + "content": " ARM in both partitions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of ARM. The influence of ARM on evaluation metrics is also shown in Tab. 3. To visually evaluate the effectiveness of ARM, we employ GradCAM [55] to generate the activation map, and utilize t-SNE [61] to demonstrate the clustering of affordances in both partitions. The results are shown in Fig. 8. It shows that ARM makes the model focus on the interactive regions to excavate the interaction contexts, and enables the model to differentiate various affordances from interactions in both partitions." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10911" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 265, + 171 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 265, + 171 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 265, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 265, + 171 + ], + "type": "image", + "image_path": "71aaa6b56d0d763a90f12a62142c759cc862dc7934c3ed2e7060ac825eef7fdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 171, + 287, + 204 + ], + "lines": [ + { + "bbox": [ + 47, + 171, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 287, + 204 + ], + "type": "text", + "content": "Figure 9. Same Image w.r.t. Multiple Point Clouds. (a) Same object category. (b) Different object categories, similar geometrics. (c) Different object categories and geometrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 213, + 277, + 320 + ], + "blocks": [ + { + "bbox": [ + 58, + 213, + 277, + 320 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 277, + 320 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 277, + 320 + ], + "type": "image", + "image_path": "95750804648c10db1e24627c1aed71739502066ba80d481cb494b2d8b2703b8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 319, + 287, + 352 + ], + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 352 + ], + "type": "text", + "content": "Figure 10. Same Point Cloud w.r.t. Multiple Images. Grounding affordance on the same point cloud with images that contain similar or disparate interactions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 372, + 173, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 173, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 173, + 384 + ], + "type": "text", + "content": "5.4. Performance Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 390, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 521 + ], + "type": "text", + "content": "Different Instances. We conduct experiments to verify whether the model could ground 3D affordance on instances from different sources: (i) using an image and different point clouds to infer respectively (Fig. 9 (a)). (ii) using multiple images and a single point cloud (Fig. 10). The results indicate that the model can anticipate 3D affordances on different instances with the same 2D interaction, and can also anticipate distinct 3D affordances on the same point cloud through different 2D interactions. Showing that the model maintains the mapping between object structures and affordances by learning from different 2D interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": "Different Object Categories. What will happen if the object category is different in the image and point cloud? To explore this issue, we perform experiments with mismatched object categories, shown in Fig. 9 (b) and (c). When objects are of different categories but with similar geometric primitives, e.g. the handle of \"Mug\" and the shoulder strap of \"Bag\", our model still properly infer the 3D affordance. However, when geometric structures are also dissimilar (Fig. 9 (c)), the model does not make random predictions. This indicates that the model maps cross-category invariance between affordance and geometries, which can be generalized to new instances of geometry." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "Multiplicity. One defining property of affordance is multiplicity: some points may correspond to multiple affordances. Fig. 11 shows that the model makes objects' multiple affordances compatible, and verifies the model does not" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 68, + 541, + 120 + ], + "blocks": [ + { + "bbox": [ + 310, + 68, + 541, + 120 + ], + "lines": [ + { + "bbox": [ + 310, + 68, + 541, + 120 + ], + "spans": [ + { + "bbox": [ + 310, + 68, + 541, + 120 + ], + "type": "image", + "image_path": "40d0e1faa8821e7386c55c5b4ade70c48db66c317502d8b8d726aca61c1320c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 122, + 544, + 145 + ], + "lines": [ + { + "bbox": [ + 306, + 122, + 544, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 122, + 544, + 145 + ], + "type": "text", + "content": "Figure 11. Multiple Affordances. Some objects like \"Bag\" contains the region that corresponds to multiple affordances." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 148, + 538, + 243 + ], + "blocks": [ + { + "bbox": [ + 307, + 148, + 538, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 148, + 538, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 538, + 243 + ], + "type": "image", + "image_path": "47fb743a30689270c03b964f9a1639fbd1ec23e5296161b32bab46c2e57f20cd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 273, + 535, + 334 + ], + "blocks": [ + { + "bbox": [ + 306, + 247, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 306, + 247, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 545, + 270 + ], + "type": "text", + "content": "Figure 12. Real-Wrold. The first row is scanned by iPhone, the second row comes from [29], point clouds are sampled from them." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 273, + 535, + 334 + ], + "lines": [ + { + "bbox": [ + 315, + 273, + 535, + 334 + ], + "spans": [ + { + "bbox": [ + 315, + 273, + 535, + 334 + ], + "type": "image", + "image_path": "74028dde10d445c71a2f715e6dc0635dc3b1e4f77e03f0c50c68de6d82c75422.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 336, + 535, + 348 + ], + "lines": [ + { + "bbox": [ + 306, + 336, + 535, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 336, + 535, + 348 + ], + "type": "text", + "content": "Figure 13. Failure Cases. Over prediction in the small regions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 369, + 545, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 404 + ], + "type": "text", + "content": "anticipate affordance by linking a region with specific affordances. Instead, it predicts by considering the presence of affordance-related interactions in the same region." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 405, + 546, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 546, + 525 + ], + "type": "text", + "content": "Real World. To validate the model in real-world scenarios, we use an iPhone to scan objects and real-scan dataset [29] to test our model, shown in Fig. 12. It shows that our model exhibits a certain degree of generalization to the real world. Limitations. Our model exhibits over predictions for some small affordance regions, shown in Fig. 13. This may be due to the limited understanding of fine-grained geometric parts, and aligning the object features in a single scale may lead to overly large receptive fields. Our feature extension will refer [13, 27] to tackle this problem." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "type": "text", + "content": "We present a novel setting for affordance grounding, which utilizes the 2D interactive semantics to guide the grounding of 3D object affordance, it has the potential to serve embodied systems when collaborating with multimodal grounding systems [30, 58, 70]. Besides, We collect the PIAD dataset as the first test bed for the proposed setting, it contains paired image-point cloud affordance data. Plus, we propose a novel framework to correlate affordance regions of objects that are from different sources and model interactive contexts to ground 3D object affordance. Comprehensive experiments on PIAD display the reliability of the setting, and we believe it could offer fresh insights and facilitate research in the affordance area." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10912" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "Acknowledgments This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 105, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 105, + 141 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 105, + 141 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 149, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 149, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 287, + 182 + ], + "type": "text", + "content": "[1] Emanuele Aiello, Diego Valsesia, and Enrico Magli. Cross-modal learning for image-guided point cloud shape completion. arXiv preprint arXiv:2209.09552, 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 183, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 288, + 227 + ], + "type": "text", + "content": "[2] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "text", + "content": "[3] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE transactions on pattern analysis and machine intelligence, 41(3):740-757, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 271, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 287, + 325 + ], + "type": "text", + "content": "[4] Yu-Wei Chao, Zhan Wang, Yugeng He, Jiaxuan Wang, and Jia Deng. Hico: A benchmark for recognizing human-object interactions in images. In Proceedings of the IEEE international conference on computer vision, pages 1017–1025, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 287, + 370 + ], + "type": "text", + "content": "[5] Honghua Chen, Zeyong Wei, Yabin Xu, Mingqiang Wei, and Jun Wang. Imlovenet: Misaligned image-supported registration network for low-overlap point cloud pairs. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 287, + 414 + ], + "type": "text", + "content": "[6] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22(4):449-462, 2013. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 287, + 469 + ], + "type": "text", + "content": "[7] Ching-Yao Chuang, Jiaman Li, Antonio Torralba, and Sanja Fidler. Learning to act properly: Predicting and explaining affordances from images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 975–983, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 287, + 525 + ], + "type": "text", + "content": "[8] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2):722-739, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 525, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 558 + ], + "type": "text", + "content": "[9] Barney Dalgarno and Mark JW Lee. What are the learning affordances of 3-d virtual environments? British Journal of Educational Technology, 41(1):10-32, 2010. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7081-7091, June 2023. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 162 + ], + "type": "text", + "content": "[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[17] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 331 + ], + "type": "text", + "content": "[18] Maximilian Jaritz, Jiayuan Gu, and Hao Su. Multi-view pointnet for 3d scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 332, + 547, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 388 + ], + "type": "text", + "content": "[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "text", + "content": "[20] David Inkyu Kim and Gaurav S Sukhatme. Semantic labeling of 3d point clouds with object affordance for robot manipulation. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 5578-5584. IEEE, 2014. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 487 + ], + "type": "text", + "content": "[21] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 489, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 522 + ], + "type": "text", + "content": "[22] Hema S Koppula and Ashutosh Saxena. Physically grounded spatio-temporal object affordances. In European Conference on Computer Vision, pages 831-847. Springer, 2014. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[23] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "text", + "content": "[24] Georg Krispel, Michael Opitz, Georg Waltner, Horst Possegger, and Horst Bischof. Fuseseg: Lidar point cloud segmentation fusing multi-modal data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1874-1883, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[25] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhi-hong Zhu, and Yuexian Zou. G21: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Yaowei Li, Bang Yang, Xuxin Cheng, Zhihong Zhu, Hongxiang Li, and Yuexian Zou. Unify, align and refine: Multi-level semantic alignment for radiology report generation. In" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10913" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "text", + "content": "[29] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, June 2022. 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "type": "text", + "content": "[30] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 353 + ], + "type": "text", + "content": "[31] Jorge M Lobo, Alberto Jiménez-Valverde, and Raimundo Real. Auc: a misleading measure of the performance of predictive distribution models. Global ecology and Biogeography, 17(2):145–151, 2008. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "text", + "content": "[32] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Transactions on Artificial Intelligence, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 431 + ], + "type": "text", + "content": "[33] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot affordance detection. arXiv preprint arXiv:2106.14747, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 433, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 465 + ], + "type": "text", + "content": "[34] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Grounded affordance from exocentric view. arXiv preprint arXiv:2208.13196, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "text", + "content": "[35] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2252-2261, 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "text", + "content": "[36] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 589 + ], + "type": "text", + "content": "[37] Maurice Merleau-Ponty. Phenomenologie de la perception (1945). Libraqire Gallimard, Paris, 1976. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "text", + "content": "[38] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers in Neurorobotics, 14:26, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[39] Jinpeng Mi, Song Tang, Zhen Deng, Michael Goerner, and Jianwei Zhang. Object affordance based multimodal fusion for natural human-robot interaction. Cognitive Systems Research, 54:128–137, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[40] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 105 + ], + "type": "text", + "content": "medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 162 + ], + "type": "text", + "content": "[41] Niluthpol Chowdhury Mithun, Karan Sikka, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Rgb2lidar: Towards solving large-scale cross-modal visual localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 934-954, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 546, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 218 + ], + "type": "text", + "content": "[42] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 546, + 262 + ], + "type": "text", + "content": "[43] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 546, + 330 + ], + "type": "text", + "content": "[44] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 331, + 546, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 386 + ], + "type": "text", + "content": "[45] Bogdan Moldovan, Plinio Moreno, Martijn Van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE international conference on robotics and automation, pages 4373-4378. IEEE, 2012. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 546, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 441 + ], + "type": "text", + "content": "[46] Austin Myers, Ching L Teo, Cornelia Fermüller, and Yiannis Aloimonos. Affordance detection of tool parts from geometric features. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 1374-1381. IEEE, 2015. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 546, + 487 + ], + "type": "text", + "content": "[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 546, + 532 + ], + "type": "text", + "content": "[48] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 546, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 546, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 546, + 588 + ], + "type": "text", + "content": "[49] Anh Nguyen, Dimitrios Kanoulas, Darwin G Caldwell, and Nikos G Tsagarakis. Detecting object affordances with convolutional neural networks. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2765-2770. IEEE, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 546, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 546, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 546, + 633 + ], + "type": "text", + "content": "[50] Takayuki Osa, Joni Pajarinen, Gerhard Neumann, J Andrew Bagnell, Pieter Abbeel, Jan Peters, et al. An algorithmic perspective on imitation learning. Foundations and Trends® in Robotics, 7(1-2):1-179, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 546, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 668 + ], + "type": "text", + "content": "[51] François Osiurak and Arnaud Badets. Tool use and affordance: Manipulation-based versus reasoning-based approaches. Psychological review, 123(5):534, 2016. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "type": "text", + "content": "[52] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3, 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10914" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[53] Md Atiqur Rahman and Yang Wang. Optimizing intersection-over-union in deep neural networks for image segmentation. In International symposium on visual computing, pages 234–244. Springer, 2016. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[54] Anirban Roy and Sinisa Todorovic. A multi-scale cnn for affordance segmentation in rgb images. In European conference on computer vision, pages 186-201. Springer, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 228 + ], + "type": "text", + "content": "[55] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 252 + ], + "type": "text", + "content": "[56] Chris Shilling. The body in culture, technology and society. Sage, 2004. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 275 + ], + "type": "text", + "content": "[57] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "text", + "content": "[58] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "text", + "content": "[59] Xun Tan, Xingyu Chen, Guowei Zhang, Jishiyu Ding, and Xuguang Lan. Mbdf-net: Multi-branch deep fusion network for 3d object detection. In Proceedings of the 1st International Workshop on Multimedia Computing for Urban Data, pages 9–17, 2021. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 431 + ], + "type": "text", + "content": "[60] Spyridon Thermos, Petros Daras, and Gerasimos Potamianos. A deep learning approach to object affordance segmentation. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2358-2362. IEEE, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "text", + "content": "[61] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "text", + "content": "[62] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[63] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In European Conference on Computer Vision, pages 421-436. Springer, 2014. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 644 + ], + "type": "text", + "content": "[64] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 90-107. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[65] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[66] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[67] Danfei Xu, Dragomir Anguelov, and Ashesh Jain. Pointfusion: Deep sensor fusion for 3d bounding box estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 244-253, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[68] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[69] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[70] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 296 + ], + "type": "text", + "content": "[71] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73-91. Springer, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "type": "text", + "content": "[72] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 373 + ], + "type": "text", + "content": "[73] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "text", + "content": "[74] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 420, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 473 + ], + "type": "text", + "content": "[75] Jiazhao Zhang, Chenyang Zhu, Lintao Zheng, and Kai Xu. Fusion-aware point convolution for online semantic 3d scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4534-4543, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 475, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 519 + ], + "type": "text", + "content": "[76] Linqing Zhao, Jiwen Lu, and Jie Zhou. Similarity-aware fusion network for 3d semantic segmentation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1585-1592. IEEE, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "text", + "content": "[77] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 564, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 545, + 597 + ], + "type": "text", + "content": "[78] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Computing and Applications, 32(18):14321-14333, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 598, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 598, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 598, + 545, + 651 + ], + "type": "text", + "content": "[79] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 654, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 654, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 307, + 654, + 545, + 708 + ], + "type": "text", + "content": "[80] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 3, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10915" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_content_list.json b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..faae96504be4d875f8bc705c62331e3b662afaf0 --- /dev/null +++ b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_content_list.json @@ -0,0 +1,1787 @@ +[ + { + "type": "text", + "text": "Group DETR: Fast DETR Training with Group-Wise One-to-Many Assignment", + "text_level": 1, + "bbox": [ + 81, + 128, + 888, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiang Chen $^{1*}$ , Xiaokang Chen $^{2*}$ , Jian Wang $^{1}$ , Shan Zhang $^{3}$ \nKun Yao $^{1}$ , Haocheng Feng $^{1}$ , Junyu Han $^{1}$ , Errui Ding $^{1}$ , Gang Zeng $^{2}$ , Jingdong Wang $^{1\\dagger}$ $^{1}$ Baidu VIS", + "bbox": [ + 148, + 179, + 826, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Lab. of Machine Perception (MoE), School of IST, Peking University", + "bbox": [ + 186, + 233, + 781, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Australian National University", + "bbox": [ + 357, + 252, + 611, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenqiang13, wangjian33} @baidu.com", + "bbox": [ + 295, + 268, + 676, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fenghaocheng,hanjunyu,dingerrui,wangjingdong}@baidu.com", + "bbox": [ + 166, + 287, + 805, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pkucxk, gang.zeng}@oku.edu.cn, shan.zhang@anu.edu.au", + "bbox": [ + 214, + 305, + 756, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 313, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detection transformer (DETR) relies on one-to-one assignment, assigning one ground-truth object to one prediction, for end-to-end detection without NMS post-processing. It is known that one-to-many assignment, assigning one ground-truth object to multiple predictions, succeeds in detection methods such as Faster R-CNN and FCOS. While the naive one-to-many assignment does not work for DETR, and it remains challenging to apply one-to-many assignment for DETR training. In this paper, we introduce Group DETR, a simple yet efficient DETR training approach that introduces a group-wise way for one-to-many assignment. This approach involves using multiple groups of object queries, conducting one-to-one assignment within each group, and performing decoder self-attention separately. It resembles data augmentation with automatically learned object query augmentation. It is also equivalent to simultaneously training parameter-sharing networks of the same architecture, introducing more supervision and thus improving DETR training. The inference process is the same as DETR trained normally and only needs one group of queries without any architecture modification. Group DETR is versatile and is applicable to various DETR variants. The experiments show that Group DETR significantly speeds up the training convergence and improves the performance of various DETR-based models. Code will be available at https://github.com/Atten4Vis/GroupDETR.", + "bbox": [ + 76, + 388, + 473, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 811, + 209, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detection Transformer (DETR) [2] conducts end-to-end object detection without the need of many hand", + "bbox": [ + 76, + 837, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/15cd1f6221ce434f13f6a8b2fbbaf3fb527b6259acd9a221e47859e1731f413b.jpg", + "image_caption": [ + "Figure 1. Group DETR accelerates the training process for DETR variants. The training convergence curves are obtained on COCO val2017 [24] with ResNet-50 [13]. Dashed and bold curves correspond to the baseline models and the Group DETR counterparts. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 501, + 354, + 901, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "crafted components, such as non-maximum suppression (NMS) [14] and anchor generation [33, 23, 32]. The architecture consists of a CNN [13] and transformer encoder [37], and a transformer decoder that consists of self-attention, cross-attention and FFNs, followed by class and box prediction FFNs. During training, one-to-one assignment, where one ground-truth object is assigned to one single prediction, is applied for learning to only promote the predictions assigned to ground-truth objects, and demote the duplicate predictions.", + "bbox": [ + 496, + 598, + 893, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work explores the solutions to accelerate the DETR training process. Previous solutions contain two main lines. The one line is to modify cross-attention so that informative image regions are selected for effectively and efficiently collecting the information from image features. Example methods include sparse sampling, through deformable attention [47], and spatial modulations with modifying object queries [8, 30, 4, 40, 43, 25, 9]. The other line is to stabilize one-to-one assignment during training, e.g., feeding ground-truth bounding boxes with noises into transformer", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6633", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "decoder [20, 44].", + "bbox": [ + 76, + 90, + 192, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We are interested in the second line. Instead of focusing on stabilizing the assignment like DN-DETR [20], we study the assignment scheme for efficient DETR training from a new perspective: introducing more supervision. It has been proven that assigning one ground-truth object to multiple predictions, i.e., one-to-many assignment, is successful in traditional object detection methods, e.g., Faster R-CNN [33] and FCOS [36] with more anchors and pixels assigned to one ground-truth object. Unfortunately, naive one-to-many assignment does not work for DETR training. It remains a challenge to apply one-to-many assignment to DETR training.", + "bbox": [ + 75, + 108, + 468, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present a simple yet efficient DETR training approach that uses a group-wise way for one-to-many assignment, called Group DETR. Our approach is based on that end-to-end detection with successful removal of NMS post-processing for DETR comes from the joint effect of two components [2, 30]: decoder self-attention, which collects the information of other predictions, and one-to-one assignment, which expects to learn to score one prediction higher and other duplicate predictions lower for one ground-truth object.", + "bbox": [ + 75, + 291, + 468, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach adopts $K$ groups of object queries, and introduces group-wise one-to-many assignment. This assignment scheme conducts one-to-one assignment within each group of object queries, resulting in that one ground-truth object is assigned to multiple predictions. It is encouraged that the prediction assigned to the ground-truth object gets a high score, and other duplicate predictions from the same group of queries get low scores. In other words, the predictions make competition within each group. Thus, our approach uses separate self-attention, i.e., self-attention is done for each group separately, eliminating the influence of predictions from other groups and easing DETR training. Regarding inference, it is the same as DETR trained normally, and only needs a single group of object queries.", + "bbox": [ + 75, + 444, + 468, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The resulting architecture is equivalent to DETR with a group of parallel decoders, illustrated in Figure 2 (a). During training, the parallel decoders boost each other through sharing decoder parameters and using different object queries. On the other hand, using more groups of object queries resembles data augmentation, and behaves as query augmentation. It introduces more supervision and improves the decoder training. In addition, it is empirically observed that the encoder training is also improved, presumably with the help of the improved decoder.", + "bbox": [ + 75, + 657, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Group DETR is versatile and is applicable to various DETR variants. Extensive experiments demonstrate that our approach is effective in achieving fast training convergence, shown in Figure 1. Group DETR obtains consistent improvements on various DETR-based methods [30, 25, 20, 44]. For instance, Group DETR significantly improves", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional DETR-C5 by $5.0\\mathrm{mAP}$ with 12-epoch training on COCO [24]. The non-trivial improvements hold when we adopt longer training schedules (e.g., 36 epochs and 50 epochs). Furthermore, Group DETR outperforms baseline methods for multi-view 3D object detection [26, 27] and instance segmentation [5].", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 500, + 195, + 627, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DETR Architecture. DETR [2] is composed of an encoder, a transformer decoder, and object class and box position predictors. The encoder takes an image $\\mathbf{I}$ as input, and outputs the image feature $\\mathbf{X}$ ,", + "bbox": [ + 496, + 220, + 890, + 281 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E n c o d e r} (\\mathbf {I}) \\rightarrow \\mathbf {X}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 292, + 890, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The decoder receives the image feature $\\mathbf{X}$ and the object queries, denoted by a matrix $\\mathbf{Q} (= [\\mathbf{q}_1\\mathbf{q}_2\\dots \\mathbf{q}_N])$ as input, and outputs the embeddings $\\tilde{\\mathbf{Q}}$ , followed by the predictors with the output denoted by $\\mathbf{Y} (= [\\mathbf{y}_1\\mathbf{y}_2\\dots \\mathbf{y}_N])$ ,", + "bbox": [ + 496, + 320, + 890, + 381 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {D e c o d e r} (\\mathbf {X}, \\mathbf {Q}) \\rightarrow \\tilde {\\mathbf {Q}}, \\quad \\operatorname {P r e d i c t o r} (\\tilde {\\mathbf {Q}}) \\rightarrow \\mathbf {Y}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 391, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The decoder is a sequence of multiple layers. Each layer includes: (i) self-attention over object queries, which performs interactions among queries for collecting the information about duplicate detection; (ii) cross-attention between queries and image features, which collects the information from image features that is useful for object detection; (iii) feed-forward network that processes the queries separately to benefit object detection.", + "bbox": [ + 496, + 419, + 890, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DETR Training. The predictions during DETR training are in the set form, and have no correspondence to the ground-truth objects. DETR uses one-to-one assignment, i.e., one ground-truth object is assigned to one predictions and vice versa, through building a bipartite matching between the predictions and the ground-truth objects:", + "bbox": [ + 496, + 544, + 890, + 635 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {y} _ {\\sigma (1)}, \\bar {\\mathbf {y}} _ {1}\\right), \\left(\\mathbf {y} _ {\\sigma (2)}, \\bar {\\mathbf {y}} _ {2}\\right),, \\dots , \\left(\\mathbf {y} _ {\\sigma (N)}, \\bar {\\mathbf {y}} _ {N}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 647, + 890, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, $\\sigma(\\cdot)$ is the optimal permutation of $N$ indices, and $[\\bar{\\mathbf{y}}_1\\bar{\\mathbf{y}}_2,\\dots \\bar{\\mathbf{y}}_N] = \\bar{\\mathbf{Y}}$ corresponds to ground truth. The loss is then formulated as below:", + "bbox": [ + 496, + 674, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {n = 1} ^ {N} \\ell \\left(\\mathbf {y} _ {\\sigma (n)}, \\bar {\\mathbf {y}} _ {n}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 727, + 890, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\ell(\\cdot)$ is a combination of the classification loss and the box regression loss between the ground-truth object $\\bar{\\mathbf{y}}$ and the prediction $\\mathbf{y}$ [2, 47, 30].", + "bbox": [ + 496, + 763, + 890, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimization with one-to-one assignment aims to score the predictions for promoting one prediction for one ground-truth object, and demoting duplicate predictions. Such scoring needs the comparison of one prediction with other predictions, and the information of other predictions is provided from decoder self-attention over queries. The", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6634", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c6b5d212cc9768883036fc403828e99b452d717bae2f20fec2a51049edce473.jpg", + "image_caption": [ + "Figure 2. Architecture illustration. (a) Our Group DETR: group-wise one-to-many assignment and separate self-attention, architecturally equivalent to parallel decoder. (b) Group-wise one-to-many assignment only. (c) Naive one-to-many assignment. We use two groups of 4 object queries as an example. X: image features; Y: predictions; $\\bar{\\mathbf{Y}}$ : ground-truth objects, where two color boxes mean two objects and two gray boxes mean dummy objects (no objects). The color lines between $\\mathbf{Y}$ and $\\bar{\\mathbf{Y}}$ correspond to the assignment for ground-truth objects, and the gray lines for dummy objects. For clarity, the predictors are not explicitly included." + ], + "image_footnote": [], + "bbox": [ + 83, + 94, + 344, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0c936e43e665cc9a91553d92361731d2ed508aa8eb4560df8d7deb7c46e1eaf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 94, + 614, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9c2225c2f3c5b82bb1720bcb3809bdf46b326f5bab2075d8676acc38ca25e521.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 95, + 890, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "two designs, one-to-one assignment and self-attention over object queries, are critical for end-to-end detection without the need of the post-processing NMS.", + "bbox": [ + 75, + 313, + 468, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One-to-many assignment for non-end-to-end detection. One-to-many assignment is successfully adopted for introducing more supervision to non-end-to-end detection training, such as Faster R-CNN [33], FCOS [36], and so on [12, 23, 32, 45, 10, 3, 11]. One ground-truth object is assigned to multiple anchors or multiple pixels. During inference, a post-processing NMS is conducted for duplicate detection removal.", + "bbox": [ + 75, + 362, + 467, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Group DETR", + "text_level": 1, + "bbox": [ + 76, + 496, + 217, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Algorithm", + "text_level": 1, + "bbox": [ + 76, + 521, + 194, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Naive one-to-many assignment. We start from a naive way for one-to-many assignment depicted in Figure 2 (c). We replace one-to-one assignment with one-to-many assignment: assign one ground-truth object to multiple predictions. It does not work and the performance is much low. The reason is that the model is trained to output multiple predictions for one ground-truth object, and lacks the scoring mechanism to promote one single prediction and demote duplicate predictions for one ground-truth object.", + "bbox": [ + 75, + 546, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Group-wise one-to-many assignment. We adopt the multi-group object query mechanism: form the initial $N$ queries as the primary group and introduce more $(K - 1)$ groups of $N$ queries, totally $K$ groups, $\\{\\mathbf{Q}_1,\\mathbf{Q}_2,\\dots ,\\mathbf{Q}_K\\}$ . Accordingly, we have $K$ groups of predictions, $\\{\\mathbf{Y}_1,\\mathbf{Y}_2,\\ldots ,\\mathbf{Y}_K\\}$ . We perform one-to-one assignment for each group, and find a bipartite matching $\\sigma_{k}(\\cdot)$ , between each group of predictions and the ground-truth objects $(\\mathbf{Y}_k,\\bar{\\mathbf{Y}})$ . This results in that only one prediction for one ground-truth object is expected to score higher, and duplicate predictions is expected to score lower within one group other than within all the groups.", + "bbox": [ + 75, + 685, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Separate self-attention. One-to-one assignment in one group means that the prediction assigned to one ground", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudocode of one Group Decoder Layer" + ], + "code_body": "SA: Self-Attention in the decoder layer \nCA: Cross-Attention in the decoder layer \nFFN: FFN in the decoder layer \nX: output image features of the encoder \nQ: object queries, with size (KxN, B, C) \nN, K, B, C: object query number, group number, batch size, feature dimension \n# group decoder \nif training: \n # split object queries to K groups \n Q_list = Q.split(N, dim=0) # a list of K tensors \n parallel_Q = cat(Q_list, dim=1) # (N, KxB, C) \n # parallel self-attention \n out = SA(parallel_Q) # (N, KxB, C) \n # concat all groups: (KxN, B, C) \n out = cat(out.split(B, dim=1), dim=0) \n # cross-attention and ffn \n out = FFN(CA(out, X)) \nelse: \n # in inference, only one group is kept \n Q = Q[:N] # (N, B, C) \n # self-attention, cross-attention, and ffn \n out = SA(Q) \n out = FFN(CA(out, X))", + "guess_lang": "python", + "bbox": [ + 500, + 335, + 879, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "truth object is superior to other predictions within the same group. This implies that we only need to collect the information of the predictions only from the same group, rather than from all the groups. Thus we perform self-attention (abbreviated as SA) over queries for each group separately:", + "bbox": [ + 496, + 627, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S A} \\left(\\mathbf {Q} _ {1}\\right), \\mathrm {S A} \\left(\\mathbf {Q} _ {2}\\right), \\dots , \\mathrm {S A} \\left(\\mathbf {Q} _ {\\mathrm {K}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 726, + 890, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training architecture. The resulting architecture for training is very simple: the encoder keeps the same, and the decoder contains $K$ separate parallel decoders as shown in Figure 2 (a):", + "bbox": [ + 496, + 753, + 890, + 815 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_1)\\to \\mathbf{Q}_1$ , Predictor $(\\mathbf{Q}_1)\\rightarrow \\mathbf{Y}_1$ $\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_2)\\to \\mathbf{Q}_2$ , Predictor $(\\mathbf{Q}_2)\\rightarrow \\mathbf{Y}_2$ 1 \n $\\dots \\dots$ $\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_K)\\to \\mathbf{Q}_K$ , Predictor $(\\mathbf{Q}_K)\\to \\mathbf{Y}_K$ . (6)", + "guess_lang": "latex", + "bbox": [ + 517, + 823, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6635", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9522b55fa2a5018eea96cfef98640f945b8f375448fa95c87ebec7111cdfc7f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 289, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1081bee59684bcc400d95ca07793734df052ae08daf2e1846c13781e9c9f0c43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 88, + 500, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec32b411ec159224dbbb9dad53b723b2a6da59224e62342efb92d5184f3d16f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 607, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ffbd37c77265d99cf668502cff7206be81f70d0865c06e4352ee936debff5b27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 89, + 702, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a04d147afc4193b479049f5ff7baf1009cabffc5f0e4aa8385383f64689b55a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 89, + 890, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9419d36f2984515e26de5ab598ba248c424dd4b7471d64d948754b3bada74e9.jpg", + "image_caption": [ + "Figure 3. Illustrating object queries. The predicted boxes and reference points corresponding to object queries in different groups for the same ground-truth object are plotted in different colors with one color for one group. It can be seen that these queries are spatially close and can be viewed as an augmentation of other queries. The results are from Group DETR over Conditional DETR-R50 [30]. The predicted boxes and reference points may overlap. Best view in color and zoom in." + ], + "image_footnote": [], + "bbox": [ + 78, + 199, + 289, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9784ef6fd49f08ccb23918db15c8cdf5e64b54086b83bf129f75b43c007aaa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 199, + 500, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a26acde4218ddb55477e95d0e32474574290cd4f2270864bed14e148083040ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 198, + 607, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce1cca17dbc00d3a9371199a56bfdb612ad38bb2f994b3edf859b3054d033904.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 198, + 700, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3145b97eb497989bfa3cd2a3a9e8420a41dc3a0de3683f38ce3f5ec8651c728b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 198, + 890, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/70d1ddf3064f2c3698fdd64b81105afc41dfaa3b642e2b963097a35535d94018.jpg", + "image_caption": [ + "Figure 4. The performance across groups of queries are similar. Only a $\\pm 0.1$ mAP is observed over the median (37.5 mAP). The mAP scores over the COCO val2017 are reported by a 12-epoch trained Conditional DETR-R50 with Group DETR." + ], + "image_footnote": [], + "bbox": [ + 106, + 386, + 455, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the parameters of the decoder and the predictor for the $K$ groups are shared. Decoder separation and parallelism are feasible in that there is no interaction among queries for the other two operations, cross-attention and FFN. Our approach is called Group Decoder. In model inference, the process is the same as DETR trained normally and only needs one group of queries without any architecture modification. The pseudo-code is shown in Algorithm 1.", + "bbox": [ + 75, + 537, + 468, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss function. The loss is an aggregation of $K$ losses, each for one decoder. It is written as follows,", + "bbox": [ + 75, + 661, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {k} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sum_ {n = 1} ^ {N} \\ell \\left(\\mathbf {y} _ {\\sigma_ {k} (n)}, \\bar {\\mathbf {y}} _ {k n}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 699, + 468, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{k}(\\cdot)$ is the optimal permutation of $N$ indices for the $k$ th decoder.", + "bbox": [ + 75, + 748, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Analysis", + "text_level": 1, + "bbox": [ + 76, + 786, + 179, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explanation with parameter-shared models. We discuss Group DETR from the perspective of training multiple models with parameter sharing. Training with Group DETR can be regarded as simultaneously training $K$ DETR models, which share the parameters of the encoder, the decoder, and the predictor, and only differ in the initialization", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e49ac70fdb560e444b88fa461442230a463c441fda2202918a54fd0d721f8642.jpg", + "image_caption": [ + "Figure 5. More stable assignment. The $x$ -axis corresponds to #epoch, and the $y$ -axis corresponds to instability score (the score is introduced by DN-DETR [20], the lower the instability score, the more stable the label assignment) over COCO val2017. One can see that the assignment in Group DETR is more stable than DN-DETR and its baseline DAB-DETR." + ], + "image_footnote": [], + "bbox": [ + 508, + 391, + 882, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of object queries. This leads to the shared parameters receive more back-propagated gradients. Thus, these parameters are better trained and accordingly the training process converges faster.", + "bbox": [ + 496, + 652, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a side benefit, we observe that Group DETR makes the assignment more stable, as shown in Figure 5. We speculate that the stability is because the improved network leads to more reliable predictions, and thus the assignment quality is better.", + "bbox": [ + 496, + 714, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explanation with object query augmentation. The multi-group object query mechanism introduces additional $(K - 1)$ group of queries, which can be regarded as an augmentation of the primary group of queries. This is empirically illustrated in Figure 3. The reference points predicting the same objects are spatially close, and thus the corresponding object queries are similar. This may suggest that the multi", + "bbox": [ + 496, + 795, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6636", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32d577bbcc44946ebca4f0e7dbb036985926944b4507c1694d099df05d404f7a.jpg", + "image_caption": [ + "Figure 6. The parallel decoders in Group DETR are efficiently implemented as parallel self-attention, cross attention and FFN." + ], + "image_footnote": [], + "bbox": [ + 135, + 85, + 411, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "group object query mechanism resembles data augmentation, and at each iteration, more automatically-learned augmented queries are included, which equivalently introduces more supervision for decoder training. The results in Figure 4 empirically suggest that different groups of augmented queries lead to similar results.", + "bbox": [ + 75, + 281, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The point about more supervision is also observed from the comparison between Equation 6 (for training with Group DETR) and Equation 2 (for normal DETR training). Group DETR training includes $K$ pairs of image feature and object query group $\\{(\\mathbf{X},\\mathbf{Q}_1),(\\mathbf{X},\\mathbf{Q}_2),\\ldots ,(\\mathbf{X},\\mathbf{Q}_K)\\}$ , and thus the loss contains more components as shown in Equation 7.", + "bbox": [ + 75, + 371, + 468, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Illustrating that training with Group DETR improves both encoder and decoder. The encoder, including CNN and transformer encoders, is initialized from a trained Conditional DETR-R50 [30] with 50 epochs and the decoder is random initialized. (a) (Fixed, Single) = the encoder is not retrained, and the decoder is trained normally without using Group DETR. (b) (Fixed, Group) = the encoder is not retrained, and the decoder is with Group DETR. (c) (Group, Group) = the encoder and the decoder are trained with Group DETR. All the results are got through training with 50 epochs. (c) > (b) implies that Group DETR also improves the encoder training.", + "bbox": [ + 75, + 488, + 468, + 642 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/13217cd77d7c4ca36a885af00bc0119b911cdd9e59a2e9eb1ea48e4216ddd162.jpg", + "table_caption": [], + "table_footnote": [ + "*: Training Conditional DETR with a trained encoder gives slightly lower performances than the one trained regularly, even though we train all components. New hyper-parameters may need to get better results." + ], + "table_body": "
EncoderDecodermAPAPsAPmAPl
(a)FixedSingle40.6*20.244.059.3
(b)FixedGroup41.521.245.060.2
(c)GroupGroup42.9*22.246.661.6
", + "bbox": [ + 84, + 646, + 464, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Encoder training improvement. The additional supervision introduces more box regression and classification supervision from more queries assigned to each ground-truth object. The gradients with more supervision are also backpropagated from the decoder to the encoder. It is presumable that the encoder also gets benefit, verified by the empirical results in Table 1.", + "bbox": [ + 75, + 760, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computation and memory complexity. Group DETR uses more decoders during training. It is expected that", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30792bb91a4c8244db2788dc760d3e1f07d74facc80e163426dd6e72e8fe5e24.jpg", + "image_caption": [ + "Figure 7. Baseline models vs their Group DETR counterparts w.r.t training memory. The gray baseline represents using the naive implementation of attention modules. With a memory-efficient implementation [6], Group DETR does not bring much memory burden during training, only requires $1.2\\mathrm{G}$ and $1.7\\mathrm{G}$ more GPU memory with Conditional DETR [30] ('C-DETR' for short) and DAB-DETR [25]." + ], + "image_footnote": [], + "bbox": [ + 506, + 94, + 893, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Group DETR outperforms baseline models with a similar training time. Conditional DETR [30] and DAB-DETR [25] serve as baseline models to compare the performances on COCO val2017 [24]. 'C-DETR' and 'w/ Group' are the abbreviations of 'Conditional DETR' and 'with Group DETR'. The entries noted by grey are the results of baseline models with the same training epochs (12 or 50 epochs) as Group DETR. To match the training times of Group DETR, we adopt longer training sheds for baselines (15 or 60 epochs). The training times are measured on 8 A100 GPUs in hours.", + "bbox": [ + 496, + 337, + 892, + 474 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a01e6c14ede9b0898daa423aa12258ec34300ae477091f22b86956123d3babc1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelw/ GroupHoursmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓4.632.614.735.048.3
5.834.415.137.351.3
5.637.618.240.755.9
C-DETR✓19.240.920.544.259.6
23.041.621.445.160.0
23.343.423.047.362.3
DAB-DETR✓5.635.216.738.651.6
7.036.317.139.452.5
6.639.119.742.556.8
DAB-DETR✓23.342.221.545.760.3
28.042.922.846.461.9
27.544.524.248.563.2
", + "bbox": [ + 501, + 479, + 893, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Group DETR will bring additional training computation costs (FLOPs) as well as training memory costs. But the parallel decoders can be implemented as a single decoder by replacing normal self-attention with parallel self-attention (depicted in Figure 6) and we can use an efficient attention implementation, FlashAttention [6, 19]. As a result, Group DETR only takes a small increase in training GPU memory and training time. For example, with Conditional DETR [30] and DAB-DETR [25], the memory increases are just $1.2\\mathrm{G}$ and $1.7\\mathrm{G}$ (Figure 7). The training time is increased by 5 minutes per epoch (from 23 minutes to 28 minutes and from 28 minutes to 33 minutes, respectively).", + "bbox": [ + 496, + 702, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide the results by increasing the training time", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6637", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1308bcc871b6bf390a46f7638013b7f1e0dfbeb8dc058fcccc1d78decfe3c630.jpg", + "image_caption": [ + "Figure 8. Comparisons with DN-DETR. Group DETR outperforms DN-DETR on DAB-DETR [25] ( $y$ -axis). Combining those two methods give better results, indicating they are complementary to each other. The $x$ -axis is the mAP scores with a 12-epoch schedule on COCO val2017. * represents that we report the best results of DN-DETR among different numbers of denoising queries (detailed results are provided in Appendix)." + ], + "image_footnote": [], + "bbox": [ + 81, + 98, + 433, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for normal DETR training to see if Group DETR benefits simply from more training time. The results given in Table 2 show that normal training with more training time brings a little benefit and the performance is still much lower than Group DETR, implying that the performance gain from our approach is not from training time increase.", + "bbox": [ + 75, + 305, + 468, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Connection to DN-DETR. DN-DETR [20] aims to stabilize one-to-one assignment during DETR training. DN-DETR forms the additional queries by adding the noises to ground-truth objects, which can be regarded as a variant of our multi-group mechanism with clear differences. In DN-DETR [20], on the one hand, the number of queries within each additional group is the same as the number of ground-truth objects. Each one corresponds to one ground-truth object, and there is no query corresponding to no-object. In contrast, our approach automatically learns a number of $N$ (e.g., 300) object queries that correspond to both ground-truth objects and no-object.", + "bbox": [ + 75, + 400, + 470, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the other hand, DN-DETR performs self-attention over noised queries, mainly for collecting the information from predictions for other objects other than from duplicate predictions. Self-attention in Group DETR instead collects both duplicate predictions and predictions for other objects.", + "bbox": [ + 75, + 582, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above two comparisons imply that DN-DETR brings the major help for the box and classification prediction, through the introduction of more positive queries corresponding to ground-truth objects (like FCOS), and no direct help for duplicate prediction removal. Our approach introduces both positive queries and negative queries (no-object), also brings the help for duplicate prediction removal.", + "bbox": [ + 75, + 657, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 8 shows that the performance of Group DETR is better than DN-DETR. We further investigate if Group DETR still benefits from introducing more positive queries with noised queries. As shown in Figure 8, the performance gain over Group DETR is non-trivial, a $1.5\\mathrm{mAP}$ . This implies that Group DETR and DN-DETR are complementary and their major roles are different, though they have some similarities.", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 633, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of Group DETR in various DETR variants, and its extension to 3D detection and instance segmentation [30, 25, 20, 47, 44, 26, 27, 5]. The training setting is almost the same as baseline models, for illustrating the effectiveness of our Group DETR. We adopt the same training settings and hyper-parameters as the baseline models, such as learning rate, optimizer, pre-trained model, initialization methods, and data augmentations1.", + "bbox": [ + 496, + 114, + 893, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Object Detection", + "text_level": 1, + "bbox": [ + 498, + 247, + 666, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setting. We study various representative DETR-based detectors, such as basic baselines (Conditional DETR [30], DAB-DETR [25], DN-DETR [20]) with dense attentions, and strong baselines (DAB-Deformable-DETR [25, 47] and DINO [44, 47]) with deformable attentions. We report the results on two training schedules, training for 12 epochs and training for more epochs (36 or 50). Unless specified, the models are trained with ResNet-50 [13] as the backbone on the COCO train2017 and evaluated on the COCO val2017. More implementation details are provided in Appendix.", + "bbox": [ + 496, + 270, + 893, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We first report the results of training with 12 epochs in Table 3. Group DETR brings consistent improvements over the baselines with dense attentions that already are superior to the original DETR [2]. It boosts Conditional DETR (-DC5) [30] by 5.0 (4.8) mAP, improves DAB-DETR (-DC5) [25] by 3.9 (4.4) mAP, and brings a 2.0 (2.6) mAP gain to DN-DETR (-DC5) [20].", + "bbox": [ + 496, + 425, + 893, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Group DETR also works well on those strong baselines with deformable attentions that are equipped with two or more accelerating techniques. It gives a 1.5 mAP improvements over DAB-Deformable-DETR [25, 47]. When applying to DINO [44, 20, 47], Group DETR also exceeds it by 0.7 mAP. The gain is non-trivial over such a stronger baseline, considering that DINO is a well-tuned model based on DAB-Deformable-DETR that combines improved hyperparameters, improved two-stage design, improved query denoising task, and other tricks.", + "bbox": [ + 496, + 532, + 893, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we report the results with 50 training epochs that is commonly adopted in many acceleration methods [47, 30, 4, 25]. Table 4 presents that Group DETR outperforms baseline models by large margins. For the stronger backbone, Swin-Large [29], our approach achieves $58.4\\mathrm{mAP}$ (still a $0.4\\mathrm{mAP}$ higher than its baseline DINO [44] ( $58.0\\mathrm{mAP}$ with Swin-Large)). This verifies the generalization ability of our Group DETR.", + "bbox": [ + 496, + 683, + 893, + 805 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We may adjust the batch size due to the limitation of the GPU memory size for both the baseline model and our approach so that the batch size is the same.", + "bbox": [ + 500, + 814, + 890, + 851 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2In fact, our approach is compatible with query denoising and two-stage. In Table 3, for example, DN-DETR [20] utilizes query denoising, and our method improves it by 2.0. Similarly, DAB-D-DETR [25] adopts a two-stage structure, and our method achieves a 1.5 improvement.", + "bbox": [ + 500, + 851, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6638", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Effectiveness of Group DETR with 12 epochs. Group DETR gives consistent gains over various DETR-based baselines on COCO val2017 [24], highlighted with brackets. All experiments adopt ResNet-50 [13] and do not use multiple patterns [40]. For DN-DETR, an improved version of DN, dynamic DN groups [44] with 100 DN queries, is used, making the results slightly different from the ones (with 3 patterns) reported in the original paper [20] (more results about the number of DN queries can be found in Appendix. 'C-DETR', 'DAB-D-DETR', and 'w/ Group' are 'Conditional DETR' [30], 'DAB-Deformable DETR' [25, 47], and 'with Group DETR', respectively, for neat representation.", + "bbox": [ + 75, + 85, + 472, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/889ff06c856a073276fd2b5959bc062dd8e48733a8848235b328c5fbdcf0bd02.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelw/ GroupmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓32.614.735.048.3
37.6 (+5.0)18.240.755.9
C-DETR-DC5✓36.418.039.652.5
41.2 (+4.8)21.445.058.7
DAB-DETR✓35.216.738.651.6
39.1 (+3.9)19.742.556.8
DAB-DETR-DC5✓37.519.440.653.2
41.9 (+4.4)23.345.658.4
DN-DETR✓38.617.941.657.7
40.6 (+2.0)19.843.959.4
DN-DETR-DC5✓41.922.245.159.8
44.5 (+2.6)25.948.262.2
DAB-D-DETR✓44.227.547.158.6
45.7 (+1.5)28.149.060.6
DINO-4scale✓49.432.352.563.2
50.1 (+0.7)32.453.264.7
", + "bbox": [ + 78, + 258, + 472, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Last, we compare the training convergence curves of the baseline models and their Group DETR counterparts. The results, as shown in Figure 1, provide more evidence that Group DETR speeds DETR training convergence on various DETR variants.", + "bbox": [ + 75, + 592, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "System-level Results on COCO test-dev with ViT-Huge. We also have the system-level performance on COCO test-dev [24] with ViT-Huge [7]. We apply Group DETR to DINO [44] and follow its training pipeline and settings: pretrain the encoder with a self-supervised method, then pretrain the whole model on Object365 [34], and last fine-tune the whole model on COCO [24]. Our model is the first to achieve 64.5 mAP on COCO test-dev, which is still superior to other methods with larger encoder and more pretraining data [28, 39, 41, 42]. The details and comparisons with other methods are provided in Appendix.", + "bbox": [ + 75, + 671, + 470, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. More Applications", + "text_level": 1, + "bbox": [ + 76, + 847, + 256, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Group DETR is applicable to DETR-style techniques to other vision problems. We report the results for two ad", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Effectiveness of Group DETR with more epochs. Group DETR still outperforms baselines by non-trivial margins with more training epochs (36 or 50 epochs). Settings and notations are consistent with Table 3, except for the training epochs (36 epochs for DINO-4scale by following the original paper [44] and 50 epochs for other models). 'DINO-4scale-Swin-L' means it adopts Swin-Large [29] as the backbone.", + "bbox": [ + 496, + 85, + 893, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df0724423e89d3944475f399c0b89d7e8e9319b1a646126ce01c0a46fe05a94b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelw/ GroupmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓40.920.544.259.6
43.4 (+2.5)23.047.362.3
C-DETR-DC5✓43.723.947.660.1
45.8 (+2.1)26.849.763.1
DAB-DETR✓42.221.545.760.3
44.5 (+2.3)24.248.563.2
DAB-DETR-DC5✓44.525.348.262.3
46.7 (+2.2)27.650.964.0
DN-DETR✓44.023.947.762.9
45.4 (+1.4)25.149.363.8
DN-DETR-DC5✓47.527.950.765.9
48.0 (+0.5)29.352.165.4
DAB-D-DETR✓48.131.451.463.4
49.7 (+1.6)31.452.565.6
DINO-4scale✓50.934.654.164.6
51.3 (+0.4)34.754.565.3
DINO-4scale-Swin-L✓58.041.361.974.0
58.4 (+0.4)41.062.573.9
", + "bbox": [ + 501, + 189, + 893, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ditional problems: multi-view 3D object detection [15, 22, 26, 27] and instance segmentation [5, 21], to further demonstrate the effectiveness.", + "bbox": [ + 496, + 503, + 890, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-view 3D object detection. We report the results over PETR [26] and PETR v2 [27] on the nuScenes val dataset [1]. Table 5 shows that Group DETR brings significant gains to PETR and PETR v2 with 24 training epochs in terms of both the nuScenes Detection Score (NDS) and mAP scores.", + "bbox": [ + 496, + 556, + 893, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results on multi-view 3D object detection. All experiments are evaluated on the nuScenes val set [1]. We train these experiments for 24 epochs with VoVNetV2 [18] as the backbone and with the image size of $800 \\times 320$ . We follow all the settings and hyper-parameters of PETR [26] and PETR v2 [27].", + "bbox": [ + 496, + 660, + 893, + 732 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eeb91a45ad2ba4467d184ea2e4e6eb2e71e467ca07c37d38581188fd86334b88.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelw/ GroupNDSmAP
PETR✓42.037.4
45.0 (+3.0)38.8 (+1.4)
PETR v2✓50.340.7
51.3 (+1.0)41.9 (+1.2)
", + "bbox": [ + 532, + 736, + 857, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instance segmentation. We demonstrate the effectiveness of the representative method, Mask2Former [5]. The results are given in Table 6. Group DETR achieves a 1.2 (0.3) $\\mathrm{MAP}^m$ gain with 12 (50) epochs.", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6639", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b92bd8ed40e83ee34a4c636907af621de756bc96f716f430ddce8dfbf6f2e9b9.jpg", + "table_caption": [ + "Table 6. Results on instance segmentation. The mask mAP $(\\mathrm{mAP}^m)$ is used for instance segmentation on COCO val2017. We adopt Mask2Former [5] as the baseline. The experiments are conducted with ResNet-50 [13] as the backbone, following all the settings of Mask2Former." + ], + "table_footnote": [], + "table_body": "
Epochsw/ GroupmAPmAPSmAPmAPl
1238.517.641.460.4
12✓39.7 (+1.2)18.742.860.8
5043.723.447.264.8
50✓44.0 (+0.3)23.847.165.1
", + "bbox": [ + 91, + 166, + 452, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9664a764953eb2e4791994bdf9b81fbe6a5196f20a84fe48afe50ad99d5d52a2.jpg", + "table_caption": [ + "Table 7. Effects of group-wise one-to-many assignment and separate self-attention. (a) baseline: one-to-one assignment with 300 object queries. (b) naive one-to-many assignment with 3300 object queries for training and inference. (c) group-wise one-to-many assignment and no separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. (d) group-wise one-to-many assignment and separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. o2m = one-to-many, Sep. SA = separate self-attention." + ], + "table_footnote": [], + "table_body": "
o2mSep. SAmAPAPsAPmAPl
(a)××32.614.434.948.6
(b)Naive×8.48.013.213.3
(c)Group×34.816.437.751.4
(d)Group✓37.618.240.755.9
", + "bbox": [ + 78, + 387, + 468, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 497, + 230, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct the ablation study by using Conditional DETR [30] as the baseline. The CNN backbone is ResNet-50 [13], and the training epoch nubmer is 12. The performances are evaluated on COCO val2017 [24]. We mainly study the effects of the key design: group-wise one-to-many assignment, separate self-attention, and group number.", + "bbox": [ + 75, + 523, + 468, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Group-wise one-to-many assignment and separate self-attention. Table 7 shows how group-wise one-to-many (o2m) assignment and separate self-attention make contributions. In comparison to the baseline (a), group-wise o2m assignment improves the mAP score from $32.6\\mathrm{mAP}$ to $34.8\\mathrm{mAP}$ : with the gain 2.2. The separate self attention (Sep. SA) further gets a $2.8\\mathrm{mAP}$ gain. In addition, we report naive one-to-many assignment. The results are very poor, which is reasonable in that there are duplicate predictions and there is a lack of scoring mechanisms for demoting them. The results suggest that both group-wise o2m assignment and separate self-attention are effective.", + "bbox": [ + 75, + 622, + 468, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Group number. Figure 9 shows the influence of the number of groups $K$ in Group DETR. The detection performance improves when increasing the number of groups, and becomes stable when the group number reaches 11. Thus, we adopt $K = 11$ by default in Group DETR in our experiments.", + "bbox": [ + 75, + 809, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0cac6c4c1638a11f233bde704afcfa502bcb1d6b62f3a5b04992320bb67813a.jpg", + "image_caption": [ + "Figure 9. Influence of group number. The $x$ -axis is the number of groups. It can be seen that the performance becomes stable when the number of groups reaches 11." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 885, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Works", + "text_level": 1, + "bbox": [ + 500, + 253, + 650, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are two main lines for accelerating DETR training: modify cross-attention and stabilize one-to-one assignment. The two are complementary and can be combined to further boost the performance.", + "bbox": [ + 498, + 279, + 892, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modifying cross-attention. Cross-attention module aims to collect the information from the image features useful to classification and localization. Various methods are proposed to select the informative image regions more efficiently and effectively [8, 4, 40, 43, 25, 9]. For example, Deformable attention [47] selects the highly informative positions dynamically according to the previous decoder embedding. Conditional DETR [4] instead continues to use the normal global attention, and dynamically computes the spatial attention to softly select the informative regions. SMCA [8] uses the Gaussian-like weight for spatial modulation.", + "bbox": [ + 496, + 344, + 892, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stabilizing one-to-one assignment. DETR [2] relies on one-to-one assignment, where each ground-truth object is assigned to a single prediction through building a bipartite matching between the predictions and the ground-truth objects. DN-DETR [20] finds the assignment process is unstable and attributes the slow convergence issue to the instabilities. Thus, DN-DETR [20] introduces groups of noisy queries by adding noises to ground-truth objects, to stabilize the assignment, leading to faster convergence. DINO [44] makes further improvement through contrastive denoising training to generate both positive and negative noise queries with different noise levels. Our approach studies the assignment mechanism instead for introducing more supervision.", + "bbox": [ + 496, + 530, + 893, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One-to-many assignment. One-to-many assignment is widely adopted in deep detectors [33, 12, 23, 36], and has attracted a lot of interest [45, 17, 46, 10, 3, 38, 35]. For example, Faster R-CNN [33] and FCOS [36] produce multiple positive anchors and pixels for each ground-truth object. In this paper, we investigate one-to-many assignment in a feasible manner for the end-to-end detector DETR.", + "bbox": [ + 496, + 733, + 893, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concurrent with our work, $\\mathcal{H}$ -DETR [16] also uses one-to-many assignment to speed up DETR training convergence. Our Group DETR and $\\mathcal{H}$ -DETR are related, but different: (1) Group DETR introduces group-wise one-to", + "bbox": [ + 498, + 839, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6640", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "many assignment with separate self-attention with the same number of object queries in each group. $\\mathcal{H}$ -DETR adopts hybrid assignments in two different groups: One group uses one-to-one assignment and another uses one-to-many assignment with more object queries. (2) All the decoders in Group DETR can be used for inference. But the additional decoder in $\\mathcal{H}$ -DETR is not directly used and requires NMS for inference. (3) During training, our architecture introduces one parameter: the number of groups. In contrast, $\\mathcal{H}$ -DETR introduces the number of additional queries and the number of additional positive queries.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "DETA [31] is another concurrent work with our Group DETR. DETA directly uses one-to-many assignment and brings NMS back to DETR frameworks. While our method provides group-wise one-to-many assignment and maintains end-to-end detection.", + "bbox": [ + 75, + 258, + 472, + 334 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 352, + 194, + 368 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The key points in Group DETR include group-wise one-to-many assignment and parallel self-attention. The success stems from involving more groups of object queries as an addition to the primary group of object queries, and thus introducing more supervision. Group-wise assignment mechanism makes sure that the competition among predictions happens within each group separately, and separate self-attention eases the training. Thus, the NMS pose-processing is not necessary, and the inference process is kept the same as normally trained DETR and not dependent on the group design. Our approach is simple, easily implemented, and general.", + "bbox": [ + 75, + 378, + 472, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision.", + "bbox": [ + 75, + 561, + 472, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 700, + 174, + 717 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 7", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 6, 8", + "[3] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In CVPR, pages 13039-13048, 2021. 3, 8" + ], + "bbox": [ + 84, + 728, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[4] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1, 6, 8", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. arXiv preprint arXiv:2112.01527, 2021. 2, 6, 7, 8", + "[6] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022. 5", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7", + "[8] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In ICCV, pages 3621–3630, 2021. 1, 8", + "[9] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5364-5373, 2022. 1, 8", + "[10] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In CVPR, pages 303-312, 2021. 3, 8", + "[11] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3", + "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 3, 8", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 6, 7, 8", + "[14] J Hosang, R Benenson, and B Schiele. Learning nonmaximum suppression. PAMI, 2017. 1", + "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7", + "[16] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 8", + "[17] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371. Springer, 2020. 8", + "[18] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In CVPR, pages 13906-13915, 2020. 7", + "[19] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5", + "[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 2, 4, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6641", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 7", + "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 7", + "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 1, 3, 8", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5, 7, 8", + "[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 5, 6, 7, 8", + "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 2, 6, 7", + "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 2, 6, 7", + "[28] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 7", + "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021. 6, 7", + "[30] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, pages 3651-3660, 2021. 1, 2, 4, 5, 6, 7, 8", + "[31] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 9", + "[32] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1, 3", + "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 2, 3, 8", + "[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7", + "[35] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In ICML, pages 9934-9944. PMLR, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 887 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 2, 3, 8", + "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1", + "[38] Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end object detection with fully convolutional network. In CVPR, pages 15849-15858, 2021. 8", + "[39] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 7", + "[40] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 7, 8", + "[41] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. arXiv preprint arXiv:2205.14141, 2022.7", + "[42] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 7", + "[43] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 1, 8", + "[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 6, 7, 8", + "[45] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 3, 8", + "[46] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.8", + "[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. CoRR, abs/2010.04159, 2020. 1, 2, 6, 7, 8" + ], + "bbox": [ + 501, + 90, + 890, + 714 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 112, + 888, + 120, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6642", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_model.json b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d829a0e03e6177dbebd44371bad71bf1cdedcddf --- /dev/null +++ b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.13, + 0.89, + 0.155 + ], + "angle": 0, + "content": "Group DETR: Fast DETR Training with Group-Wise One-to-Many Assignment" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.18, + 0.827, + 0.233 + ], + "angle": 0, + "content": "Qiang Chen\\(^{1*}\\), Xiaokang Chen\\(^{2*}\\), Jian Wang\\(^{1}\\), Shan Zhang\\(^{3}\\) \nKun Yao\\(^{1}\\), Haocheng Feng\\(^{1}\\), Junyu Han\\(^{1}\\), Errui Ding\\(^{1}\\), Gang Zeng\\(^{2}\\), Jingdong Wang\\(^{1\\dagger}\\) \n\\(^{1}\\)Baidu VIS" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.234, + 0.782, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Key Lab. of Machine Perception (MoE), School of IST, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.253, + 0.612, + 0.27 + ], + "angle": 0, + "content": "3Australian National University" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.27, + 0.677, + 0.288 + ], + "angle": 0, + "content": "{chenqiang13, wangjian33} @baidu.com" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.288, + 0.806, + 0.305 + ], + "angle": 0, + "content": "{fenghaocheng,hanjunyu,dingerrui,wangjingdong}@baidu.com" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.306, + 0.758, + 0.323 + ], + "angle": 0, + "content": "{pkucxk, gang.zeng}@oku.edu.cn, shan.zhang@anu.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.474, + 0.797 + ], + "angle": 0, + "content": "Detection transformer (DETR) relies on one-to-one assignment, assigning one ground-truth object to one prediction, for end-to-end detection without NMS post-processing. It is known that one-to-many assignment, assigning one ground-truth object to multiple predictions, succeeds in detection methods such as Faster R-CNN and FCOS. While the naive one-to-many assignment does not work for DETR, and it remains challenging to apply one-to-many assignment for DETR training. In this paper, we introduce Group DETR, a simple yet efficient DETR training approach that introduces a group-wise way for one-to-many assignment. This approach involves using multiple groups of object queries, conducting one-to-one assignment within each group, and performing decoder self-attention separately. It resembles data augmentation with automatically learned object query augmentation. It is also equivalent to simultaneously training parameter-sharing networks of the same architecture, introducing more supervision and thus improving DETR training. The inference process is the same as DETR trained normally and only needs one group of queries without any architecture modification. Group DETR is versatile and is applicable to various DETR variants. The experiments show that Group DETR significantly speeds up the training convergence and improves the performance of various DETR-based models. Code will be available at https://github.com/Atten4Vis/GroupDETR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.813, + 0.21, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.838, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Detection Transformer (DETR) [2] conducts end-to-end object detection without the need of many hand" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.356, + 0.903, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Group DETR accelerates the training process for DETR variants. The training convergence curves are obtained on COCO val2017 [24] with ResNet-50 [13]. Dashed and bold curves correspond to the baseline models and the Group DETR counterparts. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.749 + ], + "angle": 0, + "content": "crafted components, such as non-maximum suppression (NMS) [14] and anchor generation [33, 23, 32]. The architecture consists of a CNN [13] and transformer encoder [37], and a transformer decoder that consists of self-attention, cross-attention and FFNs, followed by class and box prediction FFNs. During training, one-to-one assignment, where one ground-truth object is assigned to one single prediction, is applied for learning to only promote the predictions assigned to ground-truth objects, and demote the duplicate predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.896, + 0.903 + ], + "angle": 0, + "content": "This work explores the solutions to accelerate the DETR training process. Previous solutions contain two main lines. The one line is to modify cross-attention so that informative image regions are selected for effectively and efficiently collecting the information from image features. Example methods include sparse sampling, through deformable attention [47], and spatial modulations with modifying object queries [8, 30, 4, 40, 43, 25, 9]. The other line is to stabilize one-to-one assignment during training, e.g., feeding ground-truth bounding boxes with noises into transformer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6633" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.194, + 0.106 + ], + "angle": 0, + "content": "decoder [20, 44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.47, + 0.29 + ], + "angle": 0, + "content": "We are interested in the second line. Instead of focusing on stabilizing the assignment like DN-DETR [20], we study the assignment scheme for efficient DETR training from a new perspective: introducing more supervision. It has been proven that assigning one ground-truth object to multiple predictions, i.e., one-to-many assignment, is successful in traditional object detection methods, e.g., Faster R-CNN [33] and FCOS [36] with more anchors and pixels assigned to one ground-truth object. Unfortunately, naive one-to-many assignment does not work for DETR training. It remains a challenge to apply one-to-many assignment to DETR training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.47, + 0.442 + ], + "angle": 0, + "content": "We present a simple yet efficient DETR training approach that uses a group-wise way for one-to-many assignment, called Group DETR. Our approach is based on that end-to-end detection with successful removal of NMS post-processing for DETR comes from the joint effect of two components [2, 30]: decoder self-attention, which collects the information of other predictions, and one-to-one assignment, which expects to learn to score one prediction higher and other duplicate predictions lower for one ground-truth object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Our approach adopts \\( K \\) groups of object queries, and introduces group-wise one-to-many assignment. This assignment scheme conducts one-to-one assignment within each group of object queries, resulting in that one ground-truth object is assigned to multiple predictions. It is encouraged that the prediction assigned to the ground-truth object gets a high score, and other duplicate predictions from the same group of queries get low scores. In other words, the predictions make competition within each group. Thus, our approach uses separate self-attention, i.e., self-attention is done for each group separately, eliminating the influence of predictions from other groups and easing DETR training. Regarding inference, it is the same as DETR trained normally, and only needs a single group of object queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.809 + ], + "angle": 0, + "content": "The resulting architecture is equivalent to DETR with a group of parallel decoders, illustrated in Figure 2 (a). During training, the parallel decoders boost each other through sharing decoder parameters and using different object queries. On the other hand, using more groups of object queries resembles data augmentation, and behaves as query augmentation. It introduces more supervision and improves the decoder training. In addition, it is empirically observed that the encoder training is also improved, presumably with the help of the improved decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Group DETR is versatile and is applicable to various DETR variants. Extensive experiments demonstrate that our approach is effective in achieving fast training convergence, shown in Figure 1. Group DETR obtains consistent improvements on various DETR-based methods [30, 25, 20, 44]. For instance, Group DETR significantly improves" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "Conditional DETR-C5 by \\(5.0\\mathrm{mAP}\\) with 12-epoch training on COCO [24]. The non-trivial improvements hold when we adopt longer training schedules (e.g., 36 epochs and 50 epochs). Furthermore, Group DETR outperforms baseline methods for multi-view 3D object detection [26, 27] and instance segmentation [5]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.196, + 0.628, + 0.213 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.221, + 0.892, + 0.282 + ], + "angle": 0, + "content": "DETR Architecture. DETR [2] is composed of an encoder, a transformer decoder, and object class and box position predictors. The encoder takes an image \\(\\mathbf{I}\\) as input, and outputs the image feature \\(\\mathbf{X}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.293, + 0.892, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E n c o d e r} (\\mathbf {I}) \\rightarrow \\mathbf {X}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.892, + 0.382 + ], + "angle": 0, + "content": "The decoder receives the image feature \\(\\mathbf{X}\\) and the object queries, denoted by a matrix \\(\\mathbf{Q} (= [\\mathbf{q}_1\\mathbf{q}_2\\dots \\mathbf{q}_N])\\) as input, and outputs the embeddings \\(\\tilde{\\mathbf{Q}}\\), followed by the predictors with the output denoted by \\(\\mathbf{Y} (= [\\mathbf{y}_1\\mathbf{y}_2\\dots \\mathbf{y}_N])\\)," + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.392, + 0.892, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\operatorname {D e c o d e r} (\\mathbf {X}, \\mathbf {Q}) \\rightarrow \\tilde {\\mathbf {Q}}, \\quad \\operatorname {P r e d i c t o r} (\\tilde {\\mathbf {Q}}) \\rightarrow \\mathbf {Y}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.542 + ], + "angle": 0, + "content": "The decoder is a sequence of multiple layers. Each layer includes: (i) self-attention over object queries, which performs interactions among queries for collecting the information about duplicate detection; (ii) cross-attention between queries and image features, which collects the information from image features that is useful for object detection; (iii) feed-forward network that processes the queries separately to benefit object detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.636 + ], + "angle": 0, + "content": "DETR Training. The predictions during DETR training are in the set form, and have no correspondence to the ground-truth objects. DETR uses one-to-one assignment, i.e., one ground-truth object is assigned to one predictions and vice versa, through building a bipartite matching between the predictions and the ground-truth objects:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.648, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {y} _ {\\sigma (1)}, \\bar {\\mathbf {y}} _ {1}\\right), \\left(\\mathbf {y} _ {\\sigma (2)}, \\bar {\\mathbf {y}} _ {2}\\right),, \\dots , \\left(\\mathbf {y} _ {\\sigma (N)}, \\bar {\\mathbf {y}} _ {N}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Here, \\(\\sigma(\\cdot)\\) is the optimal permutation of \\(N\\) indices, and \\([\\bar{\\mathbf{y}}_1\\bar{\\mathbf{y}}_2,\\dots \\bar{\\mathbf{y}}_N] = \\bar{\\mathbf{Y}}\\) corresponds to ground truth. The loss is then formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.728, + 0.892, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {n = 1} ^ {N} \\ell \\left(\\mathbf {y} _ {\\sigma (n)}, \\bar {\\mathbf {y}} _ {n}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.811 + ], + "angle": 0, + "content": "where \\(\\ell(\\cdot)\\) is a combination of the classification loss and the box regression loss between the ground-truth object \\(\\bar{\\mathbf{y}}\\) and the prediction \\(\\mathbf{y}\\) [2, 47, 30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Optimization with one-to-one assignment aims to score the predictions for promoting one prediction for one ground-truth object, and demoting duplicate predictions. Such scoring needs the comparison of one prediction with other predictions, and the information of other predictions is provided from decoder self-attention over queries. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6634" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.095, + 0.345, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.095, + 0.616, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.096, + 0.891, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.219, + 0.894, + 0.289 + ], + "angle": 0, + "content": "Figure 2. Architecture illustration. (a) Our Group DETR: group-wise one-to-many assignment and separate self-attention, architecturally equivalent to parallel decoder. (b) Group-wise one-to-many assignment only. (c) Naive one-to-many assignment. We use two groups of 4 object queries as an example. X: image features; Y: predictions; \\(\\bar{\\mathbf{Y}}\\): ground-truth objects, where two color boxes mean two objects and two gray boxes mean dummy objects (no objects). The color lines between \\(\\mathbf{Y}\\) and \\(\\bar{\\mathbf{Y}}\\) correspond to the assignment for ground-truth objects, and the gray lines for dummy objects. For clarity, the predictors are not explicitly included." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.47, + 0.359 + ], + "angle": 0, + "content": "two designs, one-to-one assignment and self-attention over object queries, are critical for end-to-end detection without the need of the post-processing NMS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.468, + 0.482 + ], + "angle": 0, + "content": "One-to-many assignment for non-end-to-end detection. One-to-many assignment is successfully adopted for introducing more supervision to non-end-to-end detection training, such as Faster R-CNN [33], FCOS [36], and so on [12, 23, 32, 45, 10, 3, 11]. One ground-truth object is assigned to multiple anchors or multiple pixels. During inference, a post-processing NMS is conducted for duplicate detection removal." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.218, + 0.515 + ], + "angle": 0, + "content": "3. Group DETR" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.522, + 0.196, + 0.539 + ], + "angle": 0, + "content": "3.1. Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Naive one-to-many assignment. We start from a naive way for one-to-many assignment depicted in Figure 2 (c). We replace one-to-one assignment with one-to-many assignment: assign one ground-truth object to multiple predictions. It does not work and the performance is much low. The reason is that the model is trained to output multiple predictions for one ground-truth object, and lacks the scoring mechanism to promote one single prediction and demote duplicate predictions for one ground-truth object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Group-wise one-to-many assignment. We adopt the multi-group object query mechanism: form the initial \\(N\\) queries as the primary group and introduce more \\((K - 1)\\) groups of \\(N\\) queries, totally \\(K\\) groups, \\(\\{\\mathbf{Q}_1,\\mathbf{Q}_2,\\dots ,\\mathbf{Q}_K\\}\\). Accordingly, we have \\(K\\) groups of predictions, \\(\\{\\mathbf{Y}_1,\\mathbf{Y}_2,\\ldots ,\\mathbf{Y}_K\\}\\). We perform one-to-one assignment for each group, and find a bipartite matching \\(\\sigma_{k}(\\cdot)\\), between each group of predictions and the ground-truth objects \\((\\mathbf{Y}_k,\\bar{\\mathbf{Y}})\\). This results in that only one prediction for one ground-truth object is expected to score higher, and duplicate predictions is expected to score lower within one group other than within all the groups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Separate self-attention. One-to-one assignment in one group means that the prediction assigned to one ground" + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.313, + 0.866, + 0.329 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of one Group Decoder Layer" + }, + { + "type": "code", + "bbox": [ + 0.501, + 0.336, + 0.88, + 0.605 + ], + "angle": 0, + "content": "SA: Self-Attention in the decoder layer \nCA: Cross-Attention in the decoder layer \nFFN: FFN in the decoder layer \nX: output image features of the encoder \nQ: object queries, with size (KxN, B, C) \nN, K, B, C: object query number, group number, batch size, feature dimension \n# group decoder \nif training: \n # split object queries to K groups \n Q_list = Q.split(N, dim=0) # a list of K tensors \n parallel_Q = cat(Q_list, dim=1) # (N, KxB, C) \n # parallel self-attention \n out = SA(parallel_Q) # (N, KxB, C) \n # concat all groups: (KxN, B, C) \n out = cat(out.split(B, dim=1), dim=0) \n # cross-attention and ffn \n out = FFN(CA(out, X)) \nelse: \n # in inference, only one group is kept \n Q = Q[:N] # (N, B, C) \n # self-attention, cross-attention, and ffn \n out = SA(Q) \n out = FFN(CA(out, X))" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.704 + ], + "angle": 0, + "content": "truth object is superior to other predictions within the same group. This implies that we only need to collect the information of the predictions only from the same group, rather than from all the groups. Thus we perform self-attention (abbreviated as SA) over queries for each group separately:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.727, + 0.892, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S A} \\left(\\mathbf {Q} _ {1}\\right), \\mathrm {S A} \\left(\\mathbf {Q} _ {2}\\right), \\dots , \\mathrm {S A} \\left(\\mathbf {Q} _ {\\mathrm {K}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Training architecture. The resulting architecture for training is very simple: the encoder keeps the same, and the decoder contains \\( K \\) separate parallel decoders as shown in Figure 2 (a):" + }, + { + "type": "code", + "bbox": [ + 0.519, + 0.824, + 0.892, + 0.899 + ], + "angle": 0, + "content": "\\(\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_1)\\to \\mathbf{Q}_1\\) , Predictor \\((\\mathbf{Q}_1)\\rightarrow \\mathbf{Y}_1\\) \n\\(\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_2)\\to \\mathbf{Q}_2\\) , Predictor \\((\\mathbf{Q}_2)\\rightarrow \\mathbf{Y}_2\\) 1 \n\\(\\dots \\dots\\) \n\\(\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_K)\\to \\mathbf{Q}_K\\) , Predictor \\((\\mathbf{Q}_K)\\to \\mathbf{Y}_K\\) . (6)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6635" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.29, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.501, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.609, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.09, + 0.703, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.09, + 0.892, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.2, + 0.29, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.2, + 0.501, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.199, + 0.609, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.199, + 0.702, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.199, + 0.892, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.31, + 0.894, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Illustrating object queries. The predicted boxes and reference points corresponding to object queries in different groups for the same ground-truth object are plotted in different colors with one color for one group. It can be seen that these queries are spatially close and can be viewed as an augmentation of other queries. The results are from Group DETR over Conditional DETR-R50 [30]. The predicted boxes and reference points may overlap. Best view in color and zoom in." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.387, + 0.456, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.524 + ], + "angle": 0, + "content": "Figure 4. The performance across groups of queries are similar. Only a \\(\\pm 0.1\\) mAP is observed over the median (37.5 mAP). The mAP scores over the COCO val2017 are reported by a 12-epoch trained Conditional DETR-R50 with Group DETR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Here, the parameters of the decoder and the predictor for the \\(K\\) groups are shared. Decoder separation and parallelism are feasible in that there is no interaction among queries for the other two operations, cross-attention and FFN. Our approach is called Group Decoder. In model inference, the process is the same as DETR trained normally and only needs one group of queries without any architecture modification. The pseudo-code is shown in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Loss function. The loss is an aggregation of \\(K\\) losses, each for one decoder. It is written as follows," + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.7, + 0.47, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {k} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sum_ {n = 1} ^ {N} \\ell \\left(\\mathbf {y} _ {\\sigma_ {k} (n)}, \\bar {\\mathbf {y}} _ {k n}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.779 + ], + "angle": 0, + "content": "where \\(\\sigma_{k}(\\cdot)\\) is the optimal permutation of \\(N\\) indices for the \\(k\\)th decoder." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.181, + 0.803 + ], + "angle": 0, + "content": "3.2. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Explanation with parameter-shared models. We discuss Group DETR from the perspective of training multiple models with parameter sharing. Training with Group DETR can be regarded as simultaneously training \\( K \\) DETR models, which share the parameters of the encoder, the decoder, and the predictor, and only differ in the initialization" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.392, + 0.883, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.545, + 0.894, + 0.629 + ], + "angle": 0, + "content": "Figure 5. More stable assignment. The \\( x \\)-axis corresponds to #epoch, and the \\( y \\)-axis corresponds to instability score (the score is introduced by DN-DETR [20], the lower the instability score, the more stable the label assignment) over COCO val2017. One can see that the assignment in Group DETR is more stable than DN-DETR and its baseline DAB-DETR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.715 + ], + "angle": 0, + "content": "of object queries. This leads to the shared parameters receive more back-propagated gradients. Thus, these parameters are better trained and accordingly the training process converges faster." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.79 + ], + "angle": 0, + "content": "As a side benefit, we observe that Group DETR makes the assignment more stable, as shown in Figure 5. We speculate that the stability is because the improved network leads to more reliable predictions, and thus the assignment quality is better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Explanation with object query augmentation. The multi-group object query mechanism introduces additional \\((K - 1)\\) group of queries, which can be regarded as an augmentation of the primary group of queries. This is empirically illustrated in Figure 3. The reference points predicting the same objects are spatially close, and thus the corresponding object queries are similar. This may suggest that the multi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6636" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.087, + 0.413, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.241, + 0.47, + 0.269 + ], + "angle": 0, + "content": "Figure 6. The parallel decoders in Group DETR are efficiently implemented as parallel self-attention, cross attention and FFN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.469, + 0.371 + ], + "angle": 0, + "content": "group object query mechanism resembles data augmentation, and at each iteration, more automatically-learned augmented queries are included, which equivalently introduces more supervision for decoder training. The results in Figure 4 empirically suggest that different groups of augmented queries lead to similar results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.478 + ], + "angle": 0, + "content": "The point about more supervision is also observed from the comparison between Equation 6 (for training with Group DETR) and Equation 2 (for normal DETR training). Group DETR training includes \\(K\\) pairs of image feature and object query group \\(\\{(\\mathbf{X},\\mathbf{Q}_1),(\\mathbf{X},\\mathbf{Q}_2),\\ldots ,(\\mathbf{X},\\mathbf{Q}_K)\\}\\), and thus the loss contains more components as shown in Equation 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Table 1. Illustrating that training with Group DETR improves both encoder and decoder. The encoder, including CNN and transformer encoders, is initialized from a trained Conditional DETR-R50 [30] with 50 epochs and the decoder is random initialized. (a) (Fixed, Single) = the encoder is not retrained, and the decoder is trained normally without using Group DETR. (b) (Fixed, Group) = the encoder is not retrained, and the decoder is with Group DETR. (c) (Group, Group) = the encoder and the decoder are trained with Group DETR. All the results are got through training with 50 epochs. (c) > (b) implies that Group DETR also improves the encoder training." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.647, + 0.465, + 0.711 + ], + "angle": 0, + "content": "
EncoderDecodermAPAPsAPmAPl
(a)FixedSingle40.6*20.244.059.3
(b)FixedGroup41.521.245.060.2
(c)GroupGroup42.9*22.246.661.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.712, + 0.469, + 0.747 + ], + "angle": 0, + "content": "*: Training Conditional DETR with a trained encoder gives slightly lower performances than the one trained regularly, even though we train all components. New hyper-parameters may need to get better results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Encoder training improvement. The additional supervision introduces more box regression and classification supervision from more queries assigned to each ground-truth object. The gradients with more supervision are also backpropagated from the decoder to the encoder. It is presumable that the encoder also gets benefit, verified by the empirical results in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Computation and memory complexity. Group DETR uses more decoders during training. It is expected that" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.095, + 0.895, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.228, + 0.894, + 0.326 + ], + "angle": 0, + "content": "Figure 7. Baseline models vs their Group DETR counterparts w.r.t training memory. The gray baseline represents using the naive implementation of attention modules. With a memory-efficient implementation [6], Group DETR does not bring much memory burden during training, only requires \\(1.2\\mathrm{G}\\) and \\(1.7\\mathrm{G}\\) more GPU memory with Conditional DETR [30] ('C-DETR' for short) and DAB-DETR [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Table 2. Group DETR outperforms baseline models with a similar training time. Conditional DETR [30] and DAB-DETR [25] serve as baseline models to compare the performances on COCO val2017 [24]. 'C-DETR' and 'w/ Group' are the abbreviations of 'Conditional DETR' and 'with Group DETR'. The entries noted by grey are the results of baseline models with the same training epochs (12 or 50 epochs) as Group DETR. To match the training times of Group DETR, we adopt longer training sheds for baselines (15 or 60 epochs). The training times are measured on 8 A100 GPUs in hours." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.481, + 0.895, + 0.681 + ], + "angle": 0, + "content": "
Modelw/ GroupHoursmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓4.632.614.735.048.3
5.834.415.137.351.3
5.637.618.240.755.9
C-DETR✓19.240.920.544.259.6
23.041.621.445.160.0
23.343.423.047.362.3
DAB-DETR✓5.635.216.738.651.6
7.036.317.139.452.5
6.639.119.742.556.8
DAB-DETR✓23.342.221.545.760.3
28.042.922.846.461.9
27.544.524.248.563.2
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Group DETR will bring additional training computation costs (FLOPs) as well as training memory costs. But the parallel decoders can be implemented as a single decoder by replacing normal self-attention with parallel self-attention (depicted in Figure 6) and we can use an efficient attention implementation, FlashAttention [6, 19]. As a result, Group DETR only takes a small increase in training GPU memory and training time. For example, with Conditional DETR [30] and DAB-DETR [25], the memory increases are just \\(1.2\\mathrm{G}\\) and \\(1.7\\mathrm{G}\\) (Figure 7). The training time is increased by 5 minutes per epoch (from 23 minutes to 28 minutes and from 28 minutes to 33 minutes, respectively)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We provide the results by increasing the training time" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6637" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.099, + 0.434, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.189, + 0.471, + 0.287 + ], + "angle": 0, + "content": "Figure 8. Comparisons with DN-DETR. Group DETR outperforms DN-DETR on DAB-DETR [25] (\\(y\\)-axis). Combining those two methods give better results, indicating they are complementary to each other. The \\(x\\)-axis is the mAP scores with a 12-epoch schedule on COCO val2017. * represents that we report the best results of DN-DETR among different numbers of denoising queries (detailed results are provided in Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.398 + ], + "angle": 0, + "content": "for normal DETR training to see if Group DETR benefits simply from more training time. The results given in Table 2 show that normal training with more training time brings a little benefit and the performance is still much lower than Group DETR, implying that the performance gain from our approach is not from training time increase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.471, + 0.581 + ], + "angle": 0, + "content": "Connection to DN-DETR. DN-DETR [20] aims to stabilize one-to-one assignment during DETR training. DN-DETR forms the additional queries by adding the noises to ground-truth objects, which can be regarded as a variant of our multi-group mechanism with clear differences. In DN-DETR [20], on the one hand, the number of queries within each additional group is the same as the number of ground-truth objects. Each one corresponds to one ground-truth object, and there is no query corresponding to no-object. In contrast, our approach automatically learns a number of \\( N \\) (e.g., 300) object queries that correspond to both ground-truth objects and no-object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.658 + ], + "angle": 0, + "content": "On the other hand, DN-DETR performs self-attention over noised queries, mainly for collecting the information from predictions for other objects other than from duplicate predictions. Self-attention in Group DETR instead collects both duplicate predictions and predictions for other objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.779 + ], + "angle": 0, + "content": "The above two comparisons imply that DN-DETR brings the major help for the box and classification prediction, through the introduction of more positive queries corresponding to ground-truth objects (like FCOS), and no direct help for duplicate prediction removal. Our approach introduces both positive queries and negative queries (no-object), also brings the help for duplicate prediction removal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Figure 8 shows that the performance of Group DETR is better than DN-DETR. We further investigate if Group DETR still benefits from introducing more positive queries with noised queries. As shown in Figure 8, the performance gain over Group DETR is non-trivial, a \\(1.5\\mathrm{mAP}\\). This implies that Group DETR and DN-DETR are complementary and their major roles are different, though they have some similarities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.238 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of Group DETR in various DETR variants, and its extension to 3D detection and instance segmentation [30, 25, 20, 47, 44, 26, 27, 5]. The training setting is almost the same as baseline models, for illustrating the effectiveness of our Group DETR. We adopt the same training settings and hyper-parameters as the baseline models, such as learning rate, optimizer, pre-trained model, initialization methods, and data augmentations1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.248, + 0.667, + 0.264 + ], + "angle": 0, + "content": "4.1. Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.895, + 0.424 + ], + "angle": 0, + "content": "Setting. We study various representative DETR-based detectors, such as basic baselines (Conditional DETR [30], DAB-DETR [25], DN-DETR [20]) with dense attentions, and strong baselines (DAB-Deformable-DETR [25, 47] and DINO [44, 47]) with deformable attentions. We report the results on two training schedules, training for 12 epochs and training for more epochs (36 or 50). Unless specified, the models are trained with ResNet-50 [13] as the backbone on the COCO train2017 and evaluated on the COCO val2017. More implementation details are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.532 + ], + "angle": 0, + "content": "Results. We first report the results of training with 12 epochs in Table 3. Group DETR brings consistent improvements over the baselines with dense attentions that already are superior to the original DETR [2]. It boosts Conditional DETR (-DC5) [30] by 5.0 (4.8) mAP, improves DAB-DETR (-DC5) [25] by 3.9 (4.4) mAP, and brings a 2.0 (2.6) mAP gain to DN-DETR (-DC5) [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.683 + ], + "angle": 0, + "content": "Group DETR also works well on those strong baselines with deformable attentions that are equipped with two or more accelerating techniques. It gives a 1.5 mAP improvements over DAB-Deformable-DETR [25, 47]. When applying to DINO [44, 20, 47], Group DETR also exceeds it by 0.7 mAP. The gain is non-trivial over such a stronger baseline, considering that DINO is a well-tuned model based on DAB-Deformable-DETR that combines improved hyperparameters, improved two-stage design, improved query denoising task, and other tricks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.895, + 0.806 + ], + "angle": 0, + "content": "Furthermore, we report the results with 50 training epochs that is commonly adopted in many acceleration methods [47, 30, 4, 25]. Table 4 presents that Group DETR outperforms baseline models by large margins. For the stronger backbone, Swin-Large [29], our approach achieves \\(58.4\\mathrm{mAP}\\) (still a \\(0.4\\mathrm{mAP}\\) higher than its baseline DINO [44] (\\(58.0\\mathrm{mAP}\\) with Swin-Large)). This verifies the generalization ability of our Group DETR." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.815, + 0.892, + 0.852 + ], + "angle": 0, + "content": "1We may adjust the batch size due to the limitation of the GPU memory size for both the baseline model and our approach so that the batch size is the same." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.852, + 0.894, + 0.901 + ], + "angle": 0, + "content": "2In fact, our approach is compatible with query denoising and two-stage. In Table 3, for example, DN-DETR [20] utilizes query denoising, and our method improves it by 2.0. Similarly, DAB-D-DETR [25] adopts a two-stage structure, and our method achieves a 1.5 improvement." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.815, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6638" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.087, + 0.473, + 0.255 + ], + "angle": 0, + "content": "Table 3. Effectiveness of Group DETR with 12 epochs. Group DETR gives consistent gains over various DETR-based baselines on COCO val2017 [24], highlighted with brackets. All experiments adopt ResNet-50 [13] and do not use multiple patterns [40]. For DN-DETR, an improved version of DN, dynamic DN groups [44] with 100 DN queries, is used, making the results slightly different from the ones (with 3 patterns) reported in the original paper [20] (more results about the number of DN queries can be found in Appendix. 'C-DETR', 'DAB-D-DETR', and 'w/ Group' are 'Conditional DETR' [30], 'DAB-Deformable DETR' [25, 47], and 'with Group DETR', respectively, for neat representation." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.26, + 0.473, + 0.575 + ], + "angle": 0, + "content": "
Modelw/ GroupmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓32.614.735.048.3
37.6 (+5.0)18.240.755.9
C-DETR-DC5✓36.418.039.652.5
41.2 (+4.8)21.445.058.7
DAB-DETR✓35.216.738.651.6
39.1 (+3.9)19.742.556.8
DAB-DETR-DC5✓37.519.440.653.2
41.9 (+4.4)23.345.658.4
DN-DETR✓38.617.941.657.7
40.6 (+2.0)19.843.959.4
DN-DETR-DC5✓41.922.245.159.8
44.5 (+2.6)25.948.262.2
DAB-D-DETR✓44.227.547.158.6
45.7 (+1.5)28.149.060.6
DINO-4scale✓49.432.352.563.2
50.1 (+0.7)32.453.264.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Last, we compare the training convergence curves of the baseline models and their Group DETR counterparts. The results, as shown in Figure 1, provide more evidence that Group DETR speeds DETR training convergence on various DETR variants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.472, + 0.839 + ], + "angle": 0, + "content": "System-level Results on COCO test-dev with ViT-Huge. We also have the system-level performance on COCO test-dev [24] with ViT-Huge [7]. We apply Group DETR to DINO [44] and follow its training pipeline and settings: pretrain the encoder with a self-supervised method, then pretrain the whole model on Object365 [34], and last fine-tune the whole model on COCO [24]. Our model is the first to achieve 64.5 mAP on COCO test-dev, which is still superior to other methods with larger encoder and more pretraining data [28, 39, 41, 42]. The details and comparisons with other methods are provided in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.258, + 0.865 + ], + "angle": 0, + "content": "4.2. More Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Group DETR is applicable to DETR-style techniques to other vision problems. We report the results for two ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.087, + 0.895, + 0.187 + ], + "angle": 0, + "content": "Table 4. Effectiveness of Group DETR with more epochs. Group DETR still outperforms baselines by non-trivial margins with more training epochs (36 or 50 epochs). Settings and notations are consistent with Table 3, except for the training epochs (36 epochs for DINO-4scale by following the original paper [44] and 50 epochs for other models). 'DINO-4scale-Swin-L' means it adopts Swin-Large [29] as the backbone." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.19, + 0.895, + 0.486 + ], + "angle": 0, + "content": "
Modelw/ GroupmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓40.920.544.259.6
43.4 (+2.5)23.047.362.3
C-DETR-DC5✓43.723.947.660.1
45.8 (+2.1)26.849.763.1
DAB-DETR✓42.221.545.760.3
44.5 (+2.3)24.248.563.2
DAB-DETR-DC5✓44.525.348.262.3
46.7 (+2.2)27.650.964.0
DN-DETR✓44.023.947.762.9
45.4 (+1.4)25.149.363.8
DN-DETR-DC5✓47.527.950.765.9
48.0 (+0.5)29.352.165.4
DAB-D-DETR✓48.131.451.463.4
49.7 (+1.6)31.452.565.6
DINO-4scale✓50.934.654.164.6
51.3 (+0.4)34.754.565.3
DINO-4scale-Swin-L✓58.041.361.974.0
58.4 (+0.4)41.062.573.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.551 + ], + "angle": 0, + "content": "ditional problems: multi-view 3D object detection [15, 22, 26, 27] and instance segmentation [5, 21], to further demonstrate the effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.895, + 0.648 + ], + "angle": 0, + "content": "Multi-view 3D object detection. We report the results over PETR [26] and PETR v2 [27] on the nuScenes val dataset [1]. Table 5 shows that Group DETR brings significant gains to PETR and PETR v2 with 24 training epochs in terms of both the nuScenes Detection Score (NDS) and mAP scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.895, + 0.733 + ], + "angle": 0, + "content": "Table 5. Results on multi-view 3D object detection. All experiments are evaluated on the nuScenes val set [1]. We train these experiments for 24 epochs with VoVNetV2 [18] as the backbone and with the image size of \\(800 \\times 320\\). We follow all the settings and hyper-parameters of PETR [26] and PETR v2 [27]." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.737, + 0.858, + 0.819 + ], + "angle": 0, + "content": "
Modelw/ GroupNDSmAP
PETR✓42.037.4
45.0 (+3.0)38.8 (+1.4)
PETR v2✓50.340.7
51.3 (+1.0)41.9 (+1.2)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Instance segmentation. We demonstrate the effectiveness of the representative method, Mask2Former [5]. The results are given in Table 6. Group DETR achieves a 1.2 (0.3) \\(\\mathrm{MAP}^m\\) gain with 12 (50) epochs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6639" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.472, + 0.158 + ], + "angle": 0, + "content": "Table 6. Results on instance segmentation. The mask mAP \\((\\mathrm{mAP}^m)\\) is used for instance segmentation on COCO val2017. We adopt Mask2Former [5] as the baseline. The experiments are conducted with ResNet-50 [13] as the backbone, following all the settings of Mask2Former." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.167, + 0.454, + 0.248 + ], + "angle": 0, + "content": "
Epochsw/ GroupmAPmAPSmAPmAPl
1238.517.641.460.4
12✓39.7 (+1.2)18.742.860.8
5043.723.447.264.8
50✓44.0 (+0.3)23.847.165.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.257, + 0.472, + 0.384 + ], + "angle": 0, + "content": "Table 7. Effects of group-wise one-to-many assignment and separate self-attention. (a) baseline: one-to-one assignment with 300 object queries. (b) naive one-to-many assignment with 3300 object queries for training and inference. (c) group-wise one-to-many assignment and no separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. (d) group-wise one-to-many assignment and separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. o2m = one-to-many, Sep. SA = separate self-attention." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.388, + 0.47, + 0.47 + ], + "angle": 0, + "content": "
o2mSep. SAmAPAPsAPmAPl
(a)××32.614.434.948.6
(b)Naive×8.48.013.213.3
(c)Group×34.816.437.751.4
(d)Group✓37.618.240.755.9
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.231, + 0.515 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.616 + ], + "angle": 0, + "content": "We conduct the ablation study by using Conditional DETR [30] as the baseline. The CNN backbone is ResNet-50 [13], and the training epoch nubmer is 12. The performances are evaluated on COCO val2017 [24]. We mainly study the effects of the key design: group-wise one-to-many assignment, separate self-attention, and group number." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.804 + ], + "angle": 0, + "content": "Group-wise one-to-many assignment and separate self-attention. Table 7 shows how group-wise one-to-many (o2m) assignment and separate self-attention make contributions. In comparison to the baseline (a), group-wise o2m assignment improves the mAP score from \\(32.6\\mathrm{mAP}\\) to \\(34.8\\mathrm{mAP}\\): with the gain 2.2. The separate self attention (Sep. SA) further gets a \\(2.8\\mathrm{mAP}\\) gain. In addition, we report naive one-to-many assignment. The results are very poor, which is reasonable in that there are duplicate predictions and there is a lack of scoring mechanisms for demoting them. The results suggest that both group-wise o2m assignment and separate self-attention are effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Group number. Figure 9 shows the influence of the number of groups \\( K \\) in Group DETR. The detection performance improves when increasing the number of groups, and becomes stable when the group number reaches 11. Thus, we adopt \\( K = 11 \\) by default in Group DETR in our experiments." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.092, + 0.887, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.185, + 0.894, + 0.228 + ], + "angle": 0, + "content": "Figure 9. Influence of group number. The \\( x \\)-axis is the number of groups. It can be seen that the performance becomes stable when the number of groups reaches 11." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.254, + 0.651, + 0.27 + ], + "angle": 0, + "content": "5. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.28, + 0.893, + 0.342 + ], + "angle": 0, + "content": "There are two main lines for accelerating DETR training: modify cross-attention and stabilize one-to-one assignment. The two are complementary and can be combined to further boost the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.893, + 0.526 + ], + "angle": 0, + "content": "Modifying cross-attention. Cross-attention module aims to collect the information from the image features useful to classification and localization. Various methods are proposed to select the informative image regions more efficiently and effectively [8, 4, 40, 43, 25, 9]. For example, Deformable attention [47] selects the highly informative positions dynamically according to the previous decoder embedding. Conditional DETR [4] instead continues to use the normal global attention, and dynamically computes the spatial attention to softly select the informative regions. SMCA [8] uses the Gaussian-like weight for spatial modulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.729 + ], + "angle": 0, + "content": "Stabilizing one-to-one assignment. DETR [2] relies on one-to-one assignment, where each ground-truth object is assigned to a single prediction through building a bipartite matching between the predictions and the ground-truth objects. DN-DETR [20] finds the assignment process is unstable and attributes the slow convergence issue to the instabilities. Thus, DN-DETR [20] introduces groups of noisy queries by adding noises to ground-truth objects, to stabilize the assignment, leading to faster convergence. DINO [44] makes further improvement through contrastive denoising training to generate both positive and negative noise queries with different noise levels. Our approach studies the assignment mechanism instead for introducing more supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.894, + 0.839 + ], + "angle": 0, + "content": "One-to-many assignment. One-to-many assignment is widely adopted in deep detectors [33, 12, 23, 36], and has attracted a lot of interest [45, 17, 46, 10, 3, 38, 35]. For example, Faster R-CNN [33] and FCOS [36] produce multiple positive anchors and pixels for each ground-truth object. In this paper, we investigate one-to-many assignment in a feasible manner for the end-to-end detector DETR." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Concurrent with our work, \\(\\mathcal{H}\\)-DETR [16] also uses one-to-many assignment to speed up DETR training convergence. Our Group DETR and \\(\\mathcal{H}\\)-DETR are related, but different: (1) Group DETR introduces group-wise one-to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6640" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "many assignment with separate self-attention with the same number of object queries in each group. \\(\\mathcal{H}\\)-DETR adopts hybrid assignments in two different groups: One group uses one-to-one assignment and another uses one-to-many assignment with more object queries. (2) All the decoders in Group DETR can be used for inference. But the additional decoder in \\(\\mathcal{H}\\)-DETR is not directly used and requires NMS for inference. (3) During training, our architecture introduces one parameter: the number of groups. In contrast, \\(\\mathcal{H}\\)-DETR introduces the number of additional queries and the number of additional positive queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.335 + ], + "angle": 0, + "content": "DETA [31] is another concurrent work with our Group DETR. DETA directly uses one-to-many assignment and brings NMS back to DETR frameworks. While our method provides group-wise one-to-many assignment and maintains end-to-end detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.353, + 0.196, + 0.369 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.562 + ], + "angle": 0, + "content": "The key points in Group DETR include group-wise one-to-many assignment and parallel self-attention. The success stems from involving more groups of object queries as an addition to the primary group of object queries, and thus introducing more supervision. Group-wise assignment mechanism makes sure that the competition among predictions happens within each group separately, and separate self-attention eases the training. Thus, the NMS pose-processing is not necessary, and the inference process is kept the same as normally trained DETR and not dependent on the group design. Our approach is simple, easily implemented, and general." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.473, + 0.685 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.702, + 0.175, + 0.718 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.729, + 0.47, + 0.798 + ], + "angle": 0, + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.801, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[3] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In CVPR, pages 13039-13048, 2021. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.729, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[4] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.149, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. arXiv preprint arXiv:2112.01527, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.205, + 0.895, + 0.258 + ], + "angle": 0, + "content": "[6] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.26, + 0.895, + 0.328 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.33, + 0.895, + 0.383 + ], + "angle": 0, + "content": "[8] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In ICCV, pages 3621–3630, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.385, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[9] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5364-5373, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[10] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In CVPR, pages 303-312, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[11] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[14] J Hosang, R Benenson, and B Schiele. Learning nonmaximum suppression. PAMI, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[16] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.693, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[17] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[18] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In CVPR, pages 13906-13915, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.777, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[19] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 2, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6641" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[21] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.217 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.371 + ], + "angle": 0, + "content": "[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[28] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[30] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, pages 3651-3660, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[31] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[32] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.888 + ], + "angle": 0, + "content": "[35] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In ICML, pages 9934-9944. PMLR, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[36] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[38] Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end object detection with fully convolutional network. In CVPR, pages 15849-15858, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[39] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[40] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[41] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. arXiv preprint arXiv:2205.14141, 2022.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[42] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[43] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[45] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[46] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. CoRR, abs/2010.04159, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.113, + 0.89, + 0.122, + 0.898 + ], + "angle": 0, + "content": "8" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6642" + } + ] +] \ No newline at end of file diff --git a/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_origin.pdf b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8ab0b6ab3a285ed71be7e9ef0519d6f5c09da541 --- /dev/null +++ b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/6e9ced6c-ddf3-4570-9bfd-71f343a00212_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9aad5c1f068b349c67805b86247732750f0526586106151a4915fbac2908905 +size 4818399 diff --git a/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/full.md b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..976de4409719be3254f4bc58f45ab59729470475 --- /dev/null +++ b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/full.md @@ -0,0 +1,358 @@ +# Group DETR: Fast DETR Training with Group-Wise One-to-Many Assignment + +Qiang Chen $^{1*}$ , Xiaokang Chen $^{2*}$ , Jian Wang $^{1}$ , Shan Zhang $^{3}$ +Kun Yao $^{1}$ , Haocheng Feng $^{1}$ , Junyu Han $^{1}$ , Errui Ding $^{1}$ , Gang Zeng $^{2}$ , Jingdong Wang $^{1\dagger}$ $^{1}$ Baidu VIS + +$^{2}$ Key Lab. of Machine Perception (MoE), School of IST, Peking University + +3Australian National University + +{chenqiang13, wangjian33} @baidu.com + +{fenghaocheng,hanjunyu,dingerrui,wangjingdong}@baidu.com + +{pkucxk, gang.zeng}@oku.edu.cn, shan.zhang@anu.edu.au + +# Abstract + +Detection transformer (DETR) relies on one-to-one assignment, assigning one ground-truth object to one prediction, for end-to-end detection without NMS post-processing. It is known that one-to-many assignment, assigning one ground-truth object to multiple predictions, succeeds in detection methods such as Faster R-CNN and FCOS. While the naive one-to-many assignment does not work for DETR, and it remains challenging to apply one-to-many assignment for DETR training. In this paper, we introduce Group DETR, a simple yet efficient DETR training approach that introduces a group-wise way for one-to-many assignment. This approach involves using multiple groups of object queries, conducting one-to-one assignment within each group, and performing decoder self-attention separately. It resembles data augmentation with automatically learned object query augmentation. It is also equivalent to simultaneously training parameter-sharing networks of the same architecture, introducing more supervision and thus improving DETR training. The inference process is the same as DETR trained normally and only needs one group of queries without any architecture modification. Group DETR is versatile and is applicable to various DETR variants. The experiments show that Group DETR significantly speeds up the training convergence and improves the performance of various DETR-based models. Code will be available at https://github.com/Atten4Vis/GroupDETR. + +# 1. Introduction + +Detection Transformer (DETR) [2] conducts end-to-end object detection without the need of many hand + +![](images/15cd1f6221ce434f13f6a8b2fbbaf3fb527b6259acd9a221e47859e1731f413b.jpg) +Figure 1. Group DETR accelerates the training process for DETR variants. The training convergence curves are obtained on COCO val2017 [24] with ResNet-50 [13]. Dashed and bold curves correspond to the baseline models and the Group DETR counterparts. Best viewed in color. + +crafted components, such as non-maximum suppression (NMS) [14] and anchor generation [33, 23, 32]. The architecture consists of a CNN [13] and transformer encoder [37], and a transformer decoder that consists of self-attention, cross-attention and FFNs, followed by class and box prediction FFNs. During training, one-to-one assignment, where one ground-truth object is assigned to one single prediction, is applied for learning to only promote the predictions assigned to ground-truth objects, and demote the duplicate predictions. + +This work explores the solutions to accelerate the DETR training process. Previous solutions contain two main lines. The one line is to modify cross-attention so that informative image regions are selected for effectively and efficiently collecting the information from image features. Example methods include sparse sampling, through deformable attention [47], and spatial modulations with modifying object queries [8, 30, 4, 40, 43, 25, 9]. The other line is to stabilize one-to-one assignment during training, e.g., feeding ground-truth bounding boxes with noises into transformer + +decoder [20, 44]. + +We are interested in the second line. Instead of focusing on stabilizing the assignment like DN-DETR [20], we study the assignment scheme for efficient DETR training from a new perspective: introducing more supervision. It has been proven that assigning one ground-truth object to multiple predictions, i.e., one-to-many assignment, is successful in traditional object detection methods, e.g., Faster R-CNN [33] and FCOS [36] with more anchors and pixels assigned to one ground-truth object. Unfortunately, naive one-to-many assignment does not work for DETR training. It remains a challenge to apply one-to-many assignment to DETR training. + +We present a simple yet efficient DETR training approach that uses a group-wise way for one-to-many assignment, called Group DETR. Our approach is based on that end-to-end detection with successful removal of NMS post-processing for DETR comes from the joint effect of two components [2, 30]: decoder self-attention, which collects the information of other predictions, and one-to-one assignment, which expects to learn to score one prediction higher and other duplicate predictions lower for one ground-truth object. + +Our approach adopts $K$ groups of object queries, and introduces group-wise one-to-many assignment. This assignment scheme conducts one-to-one assignment within each group of object queries, resulting in that one ground-truth object is assigned to multiple predictions. It is encouraged that the prediction assigned to the ground-truth object gets a high score, and other duplicate predictions from the same group of queries get low scores. In other words, the predictions make competition within each group. Thus, our approach uses separate self-attention, i.e., self-attention is done for each group separately, eliminating the influence of predictions from other groups and easing DETR training. Regarding inference, it is the same as DETR trained normally, and only needs a single group of object queries. + +The resulting architecture is equivalent to DETR with a group of parallel decoders, illustrated in Figure 2 (a). During training, the parallel decoders boost each other through sharing decoder parameters and using different object queries. On the other hand, using more groups of object queries resembles data augmentation, and behaves as query augmentation. It introduces more supervision and improves the decoder training. In addition, it is empirically observed that the encoder training is also improved, presumably with the help of the improved decoder. + +Group DETR is versatile and is applicable to various DETR variants. Extensive experiments demonstrate that our approach is effective in achieving fast training convergence, shown in Figure 1. Group DETR obtains consistent improvements on various DETR-based methods [30, 25, 20, 44]. For instance, Group DETR significantly improves + +Conditional DETR-C5 by $5.0\mathrm{mAP}$ with 12-epoch training on COCO [24]. The non-trivial improvements hold when we adopt longer training schedules (e.g., 36 epochs and 50 epochs). Furthermore, Group DETR outperforms baseline methods for multi-view 3D object detection [26, 27] and instance segmentation [5]. + +# 2. Background + +DETR Architecture. DETR [2] is composed of an encoder, a transformer decoder, and object class and box position predictors. The encoder takes an image $\mathbf{I}$ as input, and outputs the image feature $\mathbf{X}$ , + +$$ +\operatorname {E n c o d e r} (\mathbf {I}) \rightarrow \mathbf {X}. \tag {1} +$$ + +The decoder receives the image feature $\mathbf{X}$ and the object queries, denoted by a matrix $\mathbf{Q} (= [\mathbf{q}_1\mathbf{q}_2\dots \mathbf{q}_N])$ as input, and outputs the embeddings $\tilde{\mathbf{Q}}$ , followed by the predictors with the output denoted by $\mathbf{Y} (= [\mathbf{y}_1\mathbf{y}_2\dots \mathbf{y}_N])$ , + +$$ +\operatorname {D e c o d e r} (\mathbf {X}, \mathbf {Q}) \rightarrow \tilde {\mathbf {Q}}, \quad \operatorname {P r e d i c t o r} (\tilde {\mathbf {Q}}) \rightarrow \mathbf {Y}. \tag {2} +$$ + +The decoder is a sequence of multiple layers. Each layer includes: (i) self-attention over object queries, which performs interactions among queries for collecting the information about duplicate detection; (ii) cross-attention between queries and image features, which collects the information from image features that is useful for object detection; (iii) feed-forward network that processes the queries separately to benefit object detection. + +DETR Training. The predictions during DETR training are in the set form, and have no correspondence to the ground-truth objects. DETR uses one-to-one assignment, i.e., one ground-truth object is assigned to one predictions and vice versa, through building a bipartite matching between the predictions and the ground-truth objects: + +$$ +\left(\mathbf {y} _ {\sigma (1)}, \bar {\mathbf {y}} _ {1}\right), \left(\mathbf {y} _ {\sigma (2)}, \bar {\mathbf {y}} _ {2}\right),, \dots , \left(\mathbf {y} _ {\sigma (N)}, \bar {\mathbf {y}} _ {N}\right). \tag {3} +$$ + +Here, $\sigma(\cdot)$ is the optimal permutation of $N$ indices, and $[\bar{\mathbf{y}}_1\bar{\mathbf{y}}_2,\dots \bar{\mathbf{y}}_N] = \bar{\mathbf{Y}}$ corresponds to ground truth. The loss is then formulated as below: + +$$ +\mathcal {L} = \sum_ {n = 1} ^ {N} \ell \left(\mathbf {y} _ {\sigma (n)}, \bar {\mathbf {y}} _ {n}\right), \tag {4} +$$ + +where $\ell(\cdot)$ is a combination of the classification loss and the box regression loss between the ground-truth object $\bar{\mathbf{y}}$ and the prediction $\mathbf{y}$ [2, 47, 30]. + +Optimization with one-to-one assignment aims to score the predictions for promoting one prediction for one ground-truth object, and demoting duplicate predictions. Such scoring needs the comparison of one prediction with other predictions, and the information of other predictions is provided from decoder self-attention over queries. The + +![](images/8c6b5d212cc9768883036fc403828e99b452d717bae2f20fec2a51049edce473.jpg) +Figure 2. Architecture illustration. (a) Our Group DETR: group-wise one-to-many assignment and separate self-attention, architecturally equivalent to parallel decoder. (b) Group-wise one-to-many assignment only. (c) Naive one-to-many assignment. We use two groups of 4 object queries as an example. X: image features; Y: predictions; $\bar{\mathbf{Y}}$ : ground-truth objects, where two color boxes mean two objects and two gray boxes mean dummy objects (no objects). The color lines between $\mathbf{Y}$ and $\bar{\mathbf{Y}}$ correspond to the assignment for ground-truth objects, and the gray lines for dummy objects. For clarity, the predictors are not explicitly included. + +![](images/0c936e43e665cc9a91553d92361731d2ed508aa8eb4560df8d7deb7c46e1eaf9.jpg) + +![](images/9c2225c2f3c5b82bb1720bcb3809bdf46b326f5bab2075d8676acc38ca25e521.jpg) + +two designs, one-to-one assignment and self-attention over object queries, are critical for end-to-end detection without the need of the post-processing NMS. + +One-to-many assignment for non-end-to-end detection. One-to-many assignment is successfully adopted for introducing more supervision to non-end-to-end detection training, such as Faster R-CNN [33], FCOS [36], and so on [12, 23, 32, 45, 10, 3, 11]. One ground-truth object is assigned to multiple anchors or multiple pixels. During inference, a post-processing NMS is conducted for duplicate detection removal. + +# 3. Group DETR + +# 3.1. Algorithm + +Naive one-to-many assignment. We start from a naive way for one-to-many assignment depicted in Figure 2 (c). We replace one-to-one assignment with one-to-many assignment: assign one ground-truth object to multiple predictions. It does not work and the performance is much low. The reason is that the model is trained to output multiple predictions for one ground-truth object, and lacks the scoring mechanism to promote one single prediction and demote duplicate predictions for one ground-truth object. + +Group-wise one-to-many assignment. We adopt the multi-group object query mechanism: form the initial $N$ queries as the primary group and introduce more $(K - 1)$ groups of $N$ queries, totally $K$ groups, $\{\mathbf{Q}_1,\mathbf{Q}_2,\dots ,\mathbf{Q}_K\}$ . Accordingly, we have $K$ groups of predictions, $\{\mathbf{Y}_1,\mathbf{Y}_2,\ldots ,\mathbf{Y}_K\}$ . We perform one-to-one assignment for each group, and find a bipartite matching $\sigma_{k}(\cdot)$ , between each group of predictions and the ground-truth objects $(\mathbf{Y}_k,\bar{\mathbf{Y}})$ . This results in that only one prediction for one ground-truth object is expected to score higher, and duplicate predictions is expected to score lower within one group other than within all the groups. + +Separate self-attention. One-to-one assignment in one group means that the prediction assigned to one ground + +Algorithm 1 Pseudocode of one Group Decoder Layer +```python +SA: Self-Attention in the decoder layer +CA: Cross-Attention in the decoder layer +FFN: FFN in the decoder layer +X: output image features of the encoder +Q: object queries, with size (KxN, B, C) +N, K, B, C: object query number, group number, batch size, feature dimension +# group decoder +if training: + # split object queries to K groups + Q_list = Q.split(N, dim=0) # a list of K tensors + parallel_Q = cat(Q_list, dim=1) # (N, KxB, C) + # parallel self-attention + out = SA(parallel_Q) # (N, KxB, C) + # concat all groups: (KxN, B, C) + out = cat(out.split(B, dim=1), dim=0) + # cross-attention and ffn + out = FFN(CA(out, X)) +else: + # in inference, only one group is kept + Q = Q[:N] # (N, B, C) + # self-attention, cross-attention, and ffn + out = SA(Q) + out = FFN(CA(out, X)) +``` + +truth object is superior to other predictions within the same group. This implies that we only need to collect the information of the predictions only from the same group, rather than from all the groups. Thus we perform self-attention (abbreviated as SA) over queries for each group separately: + +$$ +\mathrm {S A} \left(\mathbf {Q} _ {1}\right), \mathrm {S A} \left(\mathbf {Q} _ {2}\right), \dots , \mathrm {S A} \left(\mathbf {Q} _ {\mathrm {K}}\right). \tag {5} +$$ + +Training architecture. The resulting architecture for training is very simple: the encoder keeps the same, and the decoder contains $K$ separate parallel decoders as shown in Figure 2 (a): + +```latex +$\mathrm{Decoder}(\mathbf{X},\mathbf{Q}_1)\to \mathbf{Q}_1$ , Predictor $(\mathbf{Q}_1)\rightarrow \mathbf{Y}_1$ $\mathrm{Decoder}(\mathbf{X},\mathbf{Q}_2)\to \mathbf{Q}_2$ , Predictor $(\mathbf{Q}_2)\rightarrow \mathbf{Y}_2$ 1 + $\dots \dots$ $\mathrm{Decoder}(\mathbf{X},\mathbf{Q}_K)\to \mathbf{Q}_K$ , Predictor $(\mathbf{Q}_K)\to \mathbf{Y}_K$ . (6) +``` + +![](images/9522b55fa2a5018eea96cfef98640f945b8f375448fa95c87ebec7111cdfc7f8.jpg) + +![](images/1081bee59684bcc400d95ca07793734df052ae08daf2e1846c13781e9c9f0c43.jpg) + +![](images/ec32b411ec159224dbbb9dad53b723b2a6da59224e62342efb92d5184f3d16f3.jpg) + +![](images/ffbd37c77265d99cf668502cff7206be81f70d0865c06e4352ee936debff5b27.jpg) + +![](images/a04d147afc4193b479049f5ff7baf1009cabffc5f0e4aa8385383f64689b55a9.jpg) + +![](images/a9419d36f2984515e26de5ab598ba248c424dd4b7471d64d948754b3bada74e9.jpg) +Figure 3. Illustrating object queries. The predicted boxes and reference points corresponding to object queries in different groups for the same ground-truth object are plotted in different colors with one color for one group. It can be seen that these queries are spatially close and can be viewed as an augmentation of other queries. The results are from Group DETR over Conditional DETR-R50 [30]. The predicted boxes and reference points may overlap. Best view in color and zoom in. + +![](images/a9784ef6fd49f08ccb23918db15c8cdf5e64b54086b83bf129f75b43c007aaa7.jpg) + +![](images/a26acde4218ddb55477e95d0e32474574290cd4f2270864bed14e148083040ac.jpg) + +![](images/ce1cca17dbc00d3a9371199a56bfdb612ad38bb2f994b3edf859b3054d033904.jpg) + +![](images/3145b97eb497989bfa3cd2a3a9e8420a41dc3a0de3683f38ce3f5ec8651c728b.jpg) + +![](images/70d1ddf3064f2c3698fdd64b81105afc41dfaa3b642e2b963097a35535d94018.jpg) +Figure 4. The performance across groups of queries are similar. Only a $\pm 0.1$ mAP is observed over the median (37.5 mAP). The mAP scores over the COCO val2017 are reported by a 12-epoch trained Conditional DETR-R50 with Group DETR. + +Here, the parameters of the decoder and the predictor for the $K$ groups are shared. Decoder separation and parallelism are feasible in that there is no interaction among queries for the other two operations, cross-attention and FFN. Our approach is called Group Decoder. In model inference, the process is the same as DETR trained normally and only needs one group of queries without any architecture modification. The pseudo-code is shown in Algorithm 1. + +Loss function. The loss is an aggregation of $K$ losses, each for one decoder. It is written as follows, + +$$ +\mathcal {L} = \frac {1}{K} \sum_ {k = 1} ^ {K} \mathcal {L} _ {k} = \frac {1}{K} \sum_ {k = 1} ^ {K} \sum_ {n = 1} ^ {N} \ell \left(\mathbf {y} _ {\sigma_ {k} (n)}, \bar {\mathbf {y}} _ {k n}\right), \tag {7} +$$ + +where $\sigma_{k}(\cdot)$ is the optimal permutation of $N$ indices for the $k$ th decoder. + +# 3.2. Analysis + +Explanation with parameter-shared models. We discuss Group DETR from the perspective of training multiple models with parameter sharing. Training with Group DETR can be regarded as simultaneously training $K$ DETR models, which share the parameters of the encoder, the decoder, and the predictor, and only differ in the initialization + +![](images/e49ac70fdb560e444b88fa461442230a463c441fda2202918a54fd0d721f8642.jpg) +Figure 5. More stable assignment. The $x$ -axis corresponds to #epoch, and the $y$ -axis corresponds to instability score (the score is introduced by DN-DETR [20], the lower the instability score, the more stable the label assignment) over COCO val2017. One can see that the assignment in Group DETR is more stable than DN-DETR and its baseline DAB-DETR. + +of object queries. This leads to the shared parameters receive more back-propagated gradients. Thus, these parameters are better trained and accordingly the training process converges faster. + +As a side benefit, we observe that Group DETR makes the assignment more stable, as shown in Figure 5. We speculate that the stability is because the improved network leads to more reliable predictions, and thus the assignment quality is better. + +Explanation with object query augmentation. The multi-group object query mechanism introduces additional $(K - 1)$ group of queries, which can be regarded as an augmentation of the primary group of queries. This is empirically illustrated in Figure 3. The reference points predicting the same objects are spatially close, and thus the corresponding object queries are similar. This may suggest that the multi + +![](images/32d577bbcc44946ebca4f0e7dbb036985926944b4507c1694d099df05d404f7a.jpg) +Figure 6. The parallel decoders in Group DETR are efficiently implemented as parallel self-attention, cross attention and FFN. + +group object query mechanism resembles data augmentation, and at each iteration, more automatically-learned augmented queries are included, which equivalently introduces more supervision for decoder training. The results in Figure 4 empirically suggest that different groups of augmented queries lead to similar results. + +The point about more supervision is also observed from the comparison between Equation 6 (for training with Group DETR) and Equation 2 (for normal DETR training). Group DETR training includes $K$ pairs of image feature and object query group $\{(\mathbf{X},\mathbf{Q}_1),(\mathbf{X},\mathbf{Q}_2),\ldots ,(\mathbf{X},\mathbf{Q}_K)\}$ , and thus the loss contains more components as shown in Equation 7. + +Table 1. Illustrating that training with Group DETR improves both encoder and decoder. The encoder, including CNN and transformer encoders, is initialized from a trained Conditional DETR-R50 [30] with 50 epochs and the decoder is random initialized. (a) (Fixed, Single) = the encoder is not retrained, and the decoder is trained normally without using Group DETR. (b) (Fixed, Group) = the encoder is not retrained, and the decoder is with Group DETR. (c) (Group, Group) = the encoder and the decoder are trained with Group DETR. All the results are got through training with 50 epochs. (c) > (b) implies that Group DETR also improves the encoder training. + +
EncoderDecodermAPAPsAPmAPl
(a)FixedSingle40.6*20.244.059.3
(b)FixedGroup41.521.245.060.2
(c)GroupGroup42.9*22.246.661.6
+ +*: Training Conditional DETR with a trained encoder gives slightly lower performances than the one trained regularly, even though we train all components. New hyper-parameters may need to get better results. + +Encoder training improvement. The additional supervision introduces more box regression and classification supervision from more queries assigned to each ground-truth object. The gradients with more supervision are also backpropagated from the decoder to the encoder. It is presumable that the encoder also gets benefit, verified by the empirical results in Table 1. + +Computation and memory complexity. Group DETR uses more decoders during training. It is expected that + +![](images/30792bb91a4c8244db2788dc760d3e1f07d74facc80e163426dd6e72e8fe5e24.jpg) +Figure 7. Baseline models vs their Group DETR counterparts w.r.t training memory. The gray baseline represents using the naive implementation of attention modules. With a memory-efficient implementation [6], Group DETR does not bring much memory burden during training, only requires $1.2\mathrm{G}$ and $1.7\mathrm{G}$ more GPU memory with Conditional DETR [30] ('C-DETR' for short) and DAB-DETR [25]. + +Table 2. Group DETR outperforms baseline models with a similar training time. Conditional DETR [30] and DAB-DETR [25] serve as baseline models to compare the performances on COCO val2017 [24]. 'C-DETR' and 'w/ Group' are the abbreviations of 'Conditional DETR' and 'with Group DETR'. The entries noted by grey are the results of baseline models with the same training epochs (12 or 50 epochs) as Group DETR. To match the training times of Group DETR, we adopt longer training sheds for baselines (15 or 60 epochs). The training times are measured on 8 A100 GPUs in hours. + +
Modelw/ GroupHoursmAP\( AP_s \)\( AP_m \)\( AP_l \)
C-DETR✓4.632.614.735.048.3
5.834.415.137.351.3
5.637.618.240.755.9
C-DETR✓19.240.920.544.259.6
23.041.621.445.160.0
23.343.423.047.362.3
DAB-DETR✓5.635.216.738.651.6
7.036.317.139.452.5
6.639.119.742.556.8
DAB-DETR✓23.342.221.545.760.3
28.042.922.846.461.9
27.544.524.248.563.2
+ +Group DETR will bring additional training computation costs (FLOPs) as well as training memory costs. But the parallel decoders can be implemented as a single decoder by replacing normal self-attention with parallel self-attention (depicted in Figure 6) and we can use an efficient attention implementation, FlashAttention [6, 19]. As a result, Group DETR only takes a small increase in training GPU memory and training time. For example, with Conditional DETR [30] and DAB-DETR [25], the memory increases are just $1.2\mathrm{G}$ and $1.7\mathrm{G}$ (Figure 7). The training time is increased by 5 minutes per epoch (from 23 minutes to 28 minutes and from 28 minutes to 33 minutes, respectively). + +We provide the results by increasing the training time + +![](images/1308bcc871b6bf390a46f7638013b7f1e0dfbeb8dc058fcccc1d78decfe3c630.jpg) +Figure 8. Comparisons with DN-DETR. Group DETR outperforms DN-DETR on DAB-DETR [25] ( $y$ -axis). Combining those two methods give better results, indicating they are complementary to each other. The $x$ -axis is the mAP scores with a 12-epoch schedule on COCO val2017. * represents that we report the best results of DN-DETR among different numbers of denoising queries (detailed results are provided in Appendix). + +for normal DETR training to see if Group DETR benefits simply from more training time. The results given in Table 2 show that normal training with more training time brings a little benefit and the performance is still much lower than Group DETR, implying that the performance gain from our approach is not from training time increase. + +Connection to DN-DETR. DN-DETR [20] aims to stabilize one-to-one assignment during DETR training. DN-DETR forms the additional queries by adding the noises to ground-truth objects, which can be regarded as a variant of our multi-group mechanism with clear differences. In DN-DETR [20], on the one hand, the number of queries within each additional group is the same as the number of ground-truth objects. Each one corresponds to one ground-truth object, and there is no query corresponding to no-object. In contrast, our approach automatically learns a number of $N$ (e.g., 300) object queries that correspond to both ground-truth objects and no-object. + +On the other hand, DN-DETR performs self-attention over noised queries, mainly for collecting the information from predictions for other objects other than from duplicate predictions. Self-attention in Group DETR instead collects both duplicate predictions and predictions for other objects. + +The above two comparisons imply that DN-DETR brings the major help for the box and classification prediction, through the introduction of more positive queries corresponding to ground-truth objects (like FCOS), and no direct help for duplicate prediction removal. Our approach introduces both positive queries and negative queries (no-object), also brings the help for duplicate prediction removal. + +Figure 8 shows that the performance of Group DETR is better than DN-DETR. We further investigate if Group DETR still benefits from introducing more positive queries with noised queries. As shown in Figure 8, the performance gain over Group DETR is non-trivial, a $1.5\mathrm{mAP}$ . This implies that Group DETR and DN-DETR are complementary and their major roles are different, though they have some similarities. + +# 4. Experiments + +We demonstrate the effectiveness of Group DETR in various DETR variants, and its extension to 3D detection and instance segmentation [30, 25, 20, 47, 44, 26, 27, 5]. The training setting is almost the same as baseline models, for illustrating the effectiveness of our Group DETR. We adopt the same training settings and hyper-parameters as the baseline models, such as learning rate, optimizer, pre-trained model, initialization methods, and data augmentations1. + +# 4.1. Object Detection + +Setting. We study various representative DETR-based detectors, such as basic baselines (Conditional DETR [30], DAB-DETR [25], DN-DETR [20]) with dense attentions, and strong baselines (DAB-Deformable-DETR [25, 47] and DINO [44, 47]) with deformable attentions. We report the results on two training schedules, training for 12 epochs and training for more epochs (36 or 50). Unless specified, the models are trained with ResNet-50 [13] as the backbone on the COCO train2017 and evaluated on the COCO val2017. More implementation details are provided in Appendix. + +Results. We first report the results of training with 12 epochs in Table 3. Group DETR brings consistent improvements over the baselines with dense attentions that already are superior to the original DETR [2]. It boosts Conditional DETR (-DC5) [30] by 5.0 (4.8) mAP, improves DAB-DETR (-DC5) [25] by 3.9 (4.4) mAP, and brings a 2.0 (2.6) mAP gain to DN-DETR (-DC5) [20]. + +Group DETR also works well on those strong baselines with deformable attentions that are equipped with two or more accelerating techniques. It gives a 1.5 mAP improvements over DAB-Deformable-DETR [25, 47]. When applying to DINO [44, 20, 47], Group DETR also exceeds it by 0.7 mAP. The gain is non-trivial over such a stronger baseline, considering that DINO is a well-tuned model based on DAB-Deformable-DETR that combines improved hyperparameters, improved two-stage design, improved query denoising task, and other tricks. + +Furthermore, we report the results with 50 training epochs that is commonly adopted in many acceleration methods [47, 30, 4, 25]. Table 4 presents that Group DETR outperforms baseline models by large margins. For the stronger backbone, Swin-Large [29], our approach achieves $58.4\mathrm{mAP}$ (still a $0.4\mathrm{mAP}$ higher than its baseline DINO [44] ( $58.0\mathrm{mAP}$ with Swin-Large)). This verifies the generalization ability of our Group DETR. + +Table 3. Effectiveness of Group DETR with 12 epochs. Group DETR gives consistent gains over various DETR-based baselines on COCO val2017 [24], highlighted with brackets. All experiments adopt ResNet-50 [13] and do not use multiple patterns [40]. For DN-DETR, an improved version of DN, dynamic DN groups [44] with 100 DN queries, is used, making the results slightly different from the ones (with 3 patterns) reported in the original paper [20] (more results about the number of DN queries can be found in Appendix. 'C-DETR', 'DAB-D-DETR', and 'w/ Group' are 'Conditional DETR' [30], 'DAB-Deformable DETR' [25, 47], and 'with Group DETR', respectively, for neat representation. + +
Modelw/ GroupmAP\( AP_s \)\( AP_m \)\( AP_l \)
C-DETR✓32.614.735.048.3
37.6 (+5.0)18.240.755.9
C-DETR-DC5✓36.418.039.652.5
41.2 (+4.8)21.445.058.7
DAB-DETR✓35.216.738.651.6
39.1 (+3.9)19.742.556.8
DAB-DETR-DC5✓37.519.440.653.2
41.9 (+4.4)23.345.658.4
DN-DETR✓38.617.941.657.7
40.6 (+2.0)19.843.959.4
DN-DETR-DC5✓41.922.245.159.8
44.5 (+2.6)25.948.262.2
DAB-D-DETR✓44.227.547.158.6
45.7 (+1.5)28.149.060.6
DINO-4scale✓49.432.352.563.2
50.1 (+0.7)32.453.264.7
+ +Last, we compare the training convergence curves of the baseline models and their Group DETR counterparts. The results, as shown in Figure 1, provide more evidence that Group DETR speeds DETR training convergence on various DETR variants. + +System-level Results on COCO test-dev with ViT-Huge. We also have the system-level performance on COCO test-dev [24] with ViT-Huge [7]. We apply Group DETR to DINO [44] and follow its training pipeline and settings: pretrain the encoder with a self-supervised method, then pretrain the whole model on Object365 [34], and last fine-tune the whole model on COCO [24]. Our model is the first to achieve 64.5 mAP on COCO test-dev, which is still superior to other methods with larger encoder and more pretraining data [28, 39, 41, 42]. The details and comparisons with other methods are provided in Appendix. + +# 4.2. More Applications + +Group DETR is applicable to DETR-style techniques to other vision problems. We report the results for two ad + +Table 4. Effectiveness of Group DETR with more epochs. Group DETR still outperforms baselines by non-trivial margins with more training epochs (36 or 50 epochs). Settings and notations are consistent with Table 3, except for the training epochs (36 epochs for DINO-4scale by following the original paper [44] and 50 epochs for other models). 'DINO-4scale-Swin-L' means it adopts Swin-Large [29] as the backbone. + +
Modelw/ GroupmAP\( AP_s \)\( AP_m \)\( AP_l \)
C-DETR✓40.920.544.259.6
43.4 (+2.5)23.047.362.3
C-DETR-DC5✓43.723.947.660.1
45.8 (+2.1)26.849.763.1
DAB-DETR✓42.221.545.760.3
44.5 (+2.3)24.248.563.2
DAB-DETR-DC5✓44.525.348.262.3
46.7 (+2.2)27.650.964.0
DN-DETR✓44.023.947.762.9
45.4 (+1.4)25.149.363.8
DN-DETR-DC5✓47.527.950.765.9
48.0 (+0.5)29.352.165.4
DAB-D-DETR✓48.131.451.463.4
49.7 (+1.6)31.452.565.6
DINO-4scale✓50.934.654.164.6
51.3 (+0.4)34.754.565.3
DINO-4scale-Swin-L✓58.041.361.974.0
58.4 (+0.4)41.062.573.9
+ +ditional problems: multi-view 3D object detection [15, 22, 26, 27] and instance segmentation [5, 21], to further demonstrate the effectiveness. + +Multi-view 3D object detection. We report the results over PETR [26] and PETR v2 [27] on the nuScenes val dataset [1]. Table 5 shows that Group DETR brings significant gains to PETR and PETR v2 with 24 training epochs in terms of both the nuScenes Detection Score (NDS) and mAP scores. + +Table 5. Results on multi-view 3D object detection. All experiments are evaluated on the nuScenes val set [1]. We train these experiments for 24 epochs with VoVNetV2 [18] as the backbone and with the image size of $800 \times 320$ . We follow all the settings and hyper-parameters of PETR [26] and PETR v2 [27]. + +
Modelw/ GroupNDSmAP
PETR✓42.037.4
45.0 (+3.0)38.8 (+1.4)
PETR v2✓50.340.7
51.3 (+1.0)41.9 (+1.2)
+ +Instance segmentation. We demonstrate the effectiveness of the representative method, Mask2Former [5]. The results are given in Table 6. Group DETR achieves a 1.2 (0.3) $\mathrm{MAP}^m$ gain with 12 (50) epochs. + +Table 6. Results on instance segmentation. The mask mAP $(\mathrm{mAP}^m)$ is used for instance segmentation on COCO val2017. We adopt Mask2Former [5] as the baseline. The experiments are conducted with ResNet-50 [13] as the backbone, following all the settings of Mask2Former. + +
Epochsw/ GroupmAPmAPSmAPmAPl
1238.517.641.460.4
12✓39.7 (+1.2)18.742.860.8
5043.723.447.264.8
50✓44.0 (+0.3)23.847.165.1
+ +Table 7. Effects of group-wise one-to-many assignment and separate self-attention. (a) baseline: one-to-one assignment with 300 object queries. (b) naive one-to-many assignment with 3300 object queries for training and inference. (c) group-wise one-to-many assignment and no separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. (d) group-wise one-to-many assignment and separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. o2m = one-to-many, Sep. SA = separate self-attention. + +
o2mSep. SAmAPAPsAPmAPl
(a)××32.614.434.948.6
(b)Naive×8.48.013.213.3
(c)Group×34.816.437.751.4
(d)Group✓37.618.240.755.9
+ +# 4.3. Ablation Study + +We conduct the ablation study by using Conditional DETR [30] as the baseline. The CNN backbone is ResNet-50 [13], and the training epoch nubmer is 12. The performances are evaluated on COCO val2017 [24]. We mainly study the effects of the key design: group-wise one-to-many assignment, separate self-attention, and group number. + +Group-wise one-to-many assignment and separate self-attention. Table 7 shows how group-wise one-to-many (o2m) assignment and separate self-attention make contributions. In comparison to the baseline (a), group-wise o2m assignment improves the mAP score from $32.6\mathrm{mAP}$ to $34.8\mathrm{mAP}$ : with the gain 2.2. The separate self attention (Sep. SA) further gets a $2.8\mathrm{mAP}$ gain. In addition, we report naive one-to-many assignment. The results are very poor, which is reasonable in that there are duplicate predictions and there is a lack of scoring mechanisms for demoting them. The results suggest that both group-wise o2m assignment and separate self-attention are effective. + +Group number. Figure 9 shows the influence of the number of groups $K$ in Group DETR. The detection performance improves when increasing the number of groups, and becomes stable when the group number reaches 11. Thus, we adopt $K = 11$ by default in Group DETR in our experiments. + +![](images/e0cac6c4c1638a11f233bde704afcfa502bcb1d6b62f3a5b04992320bb67813a.jpg) +Figure 9. Influence of group number. The $x$ -axis is the number of groups. It can be seen that the performance becomes stable when the number of groups reaches 11. + +# 5. Related Works + +There are two main lines for accelerating DETR training: modify cross-attention and stabilize one-to-one assignment. The two are complementary and can be combined to further boost the performance. + +Modifying cross-attention. Cross-attention module aims to collect the information from the image features useful to classification and localization. Various methods are proposed to select the informative image regions more efficiently and effectively [8, 4, 40, 43, 25, 9]. For example, Deformable attention [47] selects the highly informative positions dynamically according to the previous decoder embedding. Conditional DETR [4] instead continues to use the normal global attention, and dynamically computes the spatial attention to softly select the informative regions. SMCA [8] uses the Gaussian-like weight for spatial modulation. + +Stabilizing one-to-one assignment. DETR [2] relies on one-to-one assignment, where each ground-truth object is assigned to a single prediction through building a bipartite matching between the predictions and the ground-truth objects. DN-DETR [20] finds the assignment process is unstable and attributes the slow convergence issue to the instabilities. Thus, DN-DETR [20] introduces groups of noisy queries by adding noises to ground-truth objects, to stabilize the assignment, leading to faster convergence. DINO [44] makes further improvement through contrastive denoising training to generate both positive and negative noise queries with different noise levels. Our approach studies the assignment mechanism instead for introducing more supervision. + +One-to-many assignment. One-to-many assignment is widely adopted in deep detectors [33, 12, 23, 36], and has attracted a lot of interest [45, 17, 46, 10, 3, 38, 35]. For example, Faster R-CNN [33] and FCOS [36] produce multiple positive anchors and pixels for each ground-truth object. In this paper, we investigate one-to-many assignment in a feasible manner for the end-to-end detector DETR. + +Concurrent with our work, $\mathcal{H}$ -DETR [16] also uses one-to-many assignment to speed up DETR training convergence. Our Group DETR and $\mathcal{H}$ -DETR are related, but different: (1) Group DETR introduces group-wise one-to + +many assignment with separate self-attention with the same number of object queries in each group. $\mathcal{H}$ -DETR adopts hybrid assignments in two different groups: One group uses one-to-one assignment and another uses one-to-many assignment with more object queries. (2) All the decoders in Group DETR can be used for inference. But the additional decoder in $\mathcal{H}$ -DETR is not directly used and requires NMS for inference. (3) During training, our architecture introduces one parameter: the number of groups. In contrast, $\mathcal{H}$ -DETR introduces the number of additional queries and the number of additional positive queries. + +DETA [31] is another concurrent work with our Group DETR. DETA directly uses one-to-many assignment and brings NMS back to DETR frameworks. While our method provides group-wise one-to-many assignment and maintains end-to-end detection. + +# 6. Conclusion + +The key points in Group DETR include group-wise one-to-many assignment and parallel self-attention. The success stems from involving more groups of object queries as an addition to the primary group of object queries, and thus introducing more supervision. Group-wise assignment mechanism makes sure that the competition among predictions happens within each group separately, and separate self-attention eases the training. Thus, the NMS pose-processing is not necessary, and the inference process is kept the same as normally trained DETR and not dependent on the group design. Our approach is simple, easily implemented, and general. + +Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision. + +# References + +[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 7 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 6, 8 +[3] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In CVPR, pages 13039-13048, 2021. 3, 8 + +[4] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1, 6, 8 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. arXiv preprint arXiv:2112.01527, 2021. 2, 6, 7, 8 +[6] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022. 5 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7 +[8] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In ICCV, pages 3621–3630, 2021. 1, 8 +[9] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5364-5373, 2022. 1, 8 +[10] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In CVPR, pages 303-312, 2021. 3, 8 +[11] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3 +[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 3, 8 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 6, 7, 8 +[14] J Hosang, R Benenson, and B Schiele. Learning nonmaximum suppression. PAMI, 2017. 1 +[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7 +[16] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 8 +[17] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371. Springer, 2020. 8 +[18] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In CVPR, pages 13906-13915, 2020. 7 +[19] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5 +[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 2, 4, 6, 7, 8 + +[21] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 7 +[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 7 +[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 1, 3, 8 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5, 7, 8 +[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 5, 6, 7, 8 +[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 2, 6, 7 +[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 2, 6, 7 +[28] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 7 +[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021. 6, 7 +[30] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, pages 3651-3660, 2021. 1, 2, 4, 5, 6, 7, 8 +[31] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 9 +[32] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1, 3 +[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 2, 3, 8 +[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7 +[35] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In ICML, pages 9934-9944. PMLR, 2021. + +[36] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 2, 3, 8 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1 +[38] Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end object detection with fully convolutional network. In CVPR, pages 15849-15858, 2021. 8 +[39] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 7 +[40] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 7, 8 +[41] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. arXiv preprint arXiv:2205.14141, 2022.7 +[42] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 7 +[43] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 1, 8 +[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 6, 7, 8 +[45] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 3, 8 +[46] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.8 +[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. CoRR, abs/2010.04159, 2020. 1, 2, 6, 7, 8 \ No newline at end of file diff --git a/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/images.zip b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ac4eaeff09796a92808ae9e818c8cf159ca14d3 --- /dev/null +++ b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ff0fb7d663435545a5e42c87f77a92e94932617ea06cc6d99a77834a0baf7ec +size 501080 diff --git a/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/layout.json b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1f45c4d9976b7b9608930be5ef3f9bbfcd9934be --- /dev/null +++ b/2023/Group DETR_ Fast DETR Training with Group-Wise One-to-Many Assignment/layout.json @@ -0,0 +1,8846 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 102, + 544, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 102, + 544, + 122 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 544, + 122 + ], + "type": "text", + "content": "Group DETR: Fast DETR Training with Group-Wise One-to-Many Assignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": "Qiang Chen" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Xiaokang Chen" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Jian Wang" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Shan Zhang" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": " \nKun Yao" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Haocheng Feng" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Junyu Han" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Errui Ding" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Gang Zeng" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": ", Jingdong Wang" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 506, + 184 + ], + "type": "text", + "content": "Baidu VIS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 185, + 478, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 185, + 478, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 478, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 185, + 478, + 199 + ], + "type": "text", + "content": "Key Lab. of Machine Perception (MoE), School of IST, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 200, + 374, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 200, + 374, + 213 + ], + "spans": [ + { + "bbox": [ + 219, + 200, + 374, + 213 + ], + "type": "text", + "content": "3Australian National University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 181, + 213, + 414, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 213, + 414, + 228 + ], + "spans": [ + { + "bbox": [ + 181, + 213, + 414, + 228 + ], + "type": "text", + "content": "{chenqiang13, wangjian33} @baidu.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 228, + 493, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 228, + 493, + 241 + ], + "spans": [ + { + "bbox": [ + 102, + 228, + 493, + 241 + ], + "type": "text", + "content": "{fenghaocheng,hanjunyu,dingerrui,wangjingdong}@baidu.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 242, + 463, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 463, + 255 + ], + "type": "text", + "content": "{pkucxk, gang.zeng}@oku.edu.cn, shan.zhang@anu.edu.au" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 308, + 290, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 290, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 290, + 631 + ], + "type": "text", + "content": "Detection transformer (DETR) relies on one-to-one assignment, assigning one ground-truth object to one prediction, for end-to-end detection without NMS post-processing. It is known that one-to-many assignment, assigning one ground-truth object to multiple predictions, succeeds in detection methods such as Faster R-CNN and FCOS. While the naive one-to-many assignment does not work for DETR, and it remains challenging to apply one-to-many assignment for DETR training. In this paper, we introduce Group DETR, a simple yet efficient DETR training approach that introduces a group-wise way for one-to-many assignment. This approach involves using multiple groups of object queries, conducting one-to-one assignment within each group, and performing decoder self-attention separately. It resembles data augmentation with automatically learned object query augmentation. It is also equivalent to simultaneously training parameter-sharing networks of the same architecture, introducing more supervision and thus improving DETR training. The inference process is the same as DETR trained normally and only needs one group of queries without any architecture modification. Group DETR is versatile and is applicable to various DETR variants. The experiments show that Group DETR significantly speeds up the training convergence and improves the performance of various DETR-based models. Code will be available at https://github.com/Atten4Vis/GroupDETR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 128, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 663, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 287, + 688 + ], + "type": "text", + "content": "Detection Transformer (DETR) [2] conducts end-to-end object detection without the need of many hand" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 281, + 552, + 405 + ], + "blocks": [ + { + "bbox": [ + 307, + 281, + 552, + 405 + ], + "lines": [ + { + "bbox": [ + 307, + 281, + 552, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 552, + 405 + ], + "type": "image", + "image_path": "15cd1f6221ce434f13f6a8b2fbbaf3fb527b6259acd9a221e47859e1731f413b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 407, + 547, + 462 + ], + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 462 + ], + "type": "text", + "content": "Figure 1. Group DETR accelerates the training process for DETR variants. The training convergence curves are obtained on COCO val2017 [24] with ResNet-50 [13]. Dashed and bold curves correspond to the baseline models and the Group DETR counterparts. Best viewed in color." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 593 + ], + "type": "text", + "content": "crafted components, such as non-maximum suppression (NMS) [14] and anchor generation [33, 23, 32]. The architecture consists of a CNN [13] and transformer encoder [37], and a transformer decoder that consists of self-attention, cross-attention and FFNs, followed by class and box prediction FFNs. During training, one-to-one assignment, where one ground-truth object is assigned to one single prediction, is applied for learning to only promote the predictions assigned to ground-truth objects, and demote the duplicate predictions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "This work explores the solutions to accelerate the DETR training process. Previous solutions contain two main lines. The one line is to modify cross-attention so that informative image regions are selected for effectively and efficiently collecting the information from image features. Example methods include sparse sampling, through deformable attention [47], and spatial modulations with modifying object queries [8, 30, 4, 40, 43, 25, 9]. The other line is to stabilize one-to-one assignment during training, e.g., feeding ground-truth bounding boxes with noises into transformer" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6633" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 118, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 118, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 118, + 83 + ], + "type": "text", + "content": "decoder [20, 44]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 287, + 229 + ], + "type": "text", + "content": "We are interested in the second line. Instead of focusing on stabilizing the assignment like DN-DETR [20], we study the assignment scheme for efficient DETR training from a new perspective: introducing more supervision. It has been proven that assigning one ground-truth object to multiple predictions, i.e., one-to-many assignment, is successful in traditional object detection methods, e.g., Faster R-CNN [33] and FCOS [36] with more anchors and pixels assigned to one ground-truth object. Unfortunately, naive one-to-many assignment does not work for DETR training. It remains a challenge to apply one-to-many assignment to DETR training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 350 + ], + "type": "text", + "content": "We present a simple yet efficient DETR training approach that uses a group-wise way for one-to-many assignment, called Group DETR. Our approach is based on that end-to-end detection with successful removal of NMS post-processing for DETR comes from the joint effect of two components [2, 30]: decoder self-attention, which collects the information of other predictions, and one-to-one assignment, which expects to learn to score one prediction higher and other duplicate predictions lower for one ground-truth object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "type": "text", + "content": "Our approach adopts " + }, + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 352, + 287, + 519 + ], + "type": "text", + "content": " groups of object queries, and introduces group-wise one-to-many assignment. This assignment scheme conducts one-to-one assignment within each group of object queries, resulting in that one ground-truth object is assigned to multiple predictions. It is encouraged that the prediction assigned to the ground-truth object gets a high score, and other duplicate predictions from the same group of queries get low scores. In other words, the predictions make competition within each group. Thus, our approach uses separate self-attention, i.e., self-attention is done for each group separately, eliminating the influence of predictions from other groups and easing DETR training. Regarding inference, it is the same as DETR trained normally, and only needs a single group of object queries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 640 + ], + "type": "text", + "content": "The resulting architecture is equivalent to DETR with a group of parallel decoders, illustrated in Figure 2 (a). During training, the parallel decoders boost each other through sharing decoder parameters and using different object queries. On the other hand, using more groups of object queries resembles data augmentation, and behaves as query augmentation. It introduces more supervision and improves the decoder training. In addition, it is empirically observed that the encoder training is also improved, presumably with the help of the improved decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Group DETR is versatile and is applicable to various DETR variants. Extensive experiments demonstrate that our approach is effective in achieving fast training convergence, shown in Figure 1. Group DETR obtains consistent improvements on various DETR-based methods [30, 25, 20, 44]. For instance, Group DETR significantly improves" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "Conditional DETR-C5 by " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "5.0\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " with 12-epoch training on COCO [24]. The non-trivial improvements hold when we adopt longer training schedules (e.g., 36 epochs and 50 epochs). Furthermore, Group DETR outperforms baseline methods for multi-view 3D object detection [26, 27] and instance segmentation [5]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 155, + 384, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 155, + 384, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 155, + 384, + 168 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 175, + 545, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 223 + ], + "type": "text", + "content": "DETR Architecture. DETR [2] is composed of an encoder, a transformer decoder, and object class and box position predictors. The encoder takes an image " + }, + { + "bbox": [ + 304, + 175, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 175, + 545, + 223 + ], + "type": "text", + "content": " as input, and outputs the image feature " + }, + { + "bbox": [ + 304, + 175, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 175, + 545, + 223 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 386, + 232, + 545, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 232, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 386, + 232, + 545, + 245 + ], + "type": "interline_equation", + "content": "\\operatorname {E n c o d e r} (\\mathbf {I}) \\rightarrow \\mathbf {X}. \\tag {1}", + "image_path": "25f56b63e57ca529a215b6c1f8b6076952322c537f7f3e30070a415f79ceaf3a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "text", + "content": "The decoder receives the image feature " + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "text", + "content": " and the object queries, denoted by a matrix " + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} (= [\\mathbf{q}_1\\mathbf{q}_2\\dots \\mathbf{q}_N])" + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "text", + "content": " as input, and outputs the embeddings " + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}" + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "text", + "content": ", followed by the predictors with the output denoted by " + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} (= [\\mathbf{y}_1\\mathbf{y}_2\\dots \\mathbf{y}_N])" + }, + { + "bbox": [ + 304, + 254, + 545, + 302 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 310, + 545, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 310, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 331, + 310, + 545, + 324 + ], + "type": "interline_equation", + "content": "\\operatorname {D e c o d e r} (\\mathbf {X}, \\mathbf {Q}) \\rightarrow \\tilde {\\mathbf {Q}}, \\quad \\operatorname {P r e d i c t o r} (\\tilde {\\mathbf {Q}}) \\rightarrow \\mathbf {Y}. \\tag {2}", + "image_path": "dfdaed4d7bd014707be0e642db0c1a27ed64357f07dd8d28c251e4b634442728.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 332, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 429 + ], + "type": "text", + "content": "The decoder is a sequence of multiple layers. Each layer includes: (i) self-attention over object queries, which performs interactions among queries for collecting the information about duplicate detection; (ii) cross-attention between queries and image features, which collects the information from image features that is useful for object detection; (iii) feed-forward network that processes the queries separately to benefit object detection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "type": "text", + "content": "DETR Training. The predictions during DETR training are in the set form, and have no correspondence to the ground-truth objects. DETR uses one-to-one assignment, i.e., one ground-truth object is assigned to one predictions and vice versa, through building a bipartite matching between the predictions and the ground-truth objects:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 513, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 513, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 337, + 513, + 545, + 526 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {y} _ {\\sigma (1)}, \\bar {\\mathbf {y}} _ {1}\\right), \\left(\\mathbf {y} _ {\\sigma (2)}, \\bar {\\mathbf {y}} _ {2}\\right),, \\dots , \\left(\\mathbf {y} _ {\\sigma (N)}, \\bar {\\mathbf {y}} _ {N}\\right). \\tag {3}", + "image_path": "e1b912387a52f95078e3e34c0f2fb486720a75c76bacca0a303e64fba7fce54c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "text", + "content": " is the optimal permutation of " + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "text", + "content": " indices, and " + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "inline_equation", + "content": "[\\bar{\\mathbf{y}}_1\\bar{\\mathbf{y}}_2,\\dots \\bar{\\mathbf{y}}_N] = \\bar{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "text", + "content": " corresponds to ground truth. The loss is then formulated as below:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 371, + 576, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 576, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 371, + 576, + 545, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {n = 1} ^ {N} \\ell \\left(\\mathbf {y} _ {\\sigma (n)}, \\bar {\\mathbf {y}} _ {n}\\right), \\tag {4}", + "image_path": "20771f8c14306e46864df1da3ab080b7101617f244a928726c173c4549afffa4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "text", + "content": " is a combination of the classification loss and the box regression loss between the ground-truth object " + }, + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{y}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "text", + "content": " and the prediction " + }, + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 605, + 545, + 642 + ], + "type": "text", + "content": " [2, 47, 30]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Optimization with one-to-one assignment aims to score the predictions for promoting one prediction for one ground-truth object, and demoting duplicate predictions. Such scoring needs the comparison of one prediction with other predictions, and the information of other predictions is provided from decoder self-attention over queries. The" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6634" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 75, + 211, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 75, + 211, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 75, + 211, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 75, + 211, + 171 + ], + "type": "image", + "image_path": "8c6b5d212cc9768883036fc403828e99b452d717bae2f20fec2a51049edce473.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "type": "text", + "content": "Figure 2. Architecture illustration. (a) Our Group DETR: group-wise one-to-many assignment and separate self-attention, architecturally equivalent to parallel decoder. (b) Group-wise one-to-many assignment only. (c) Naive one-to-many assignment. We use two groups of 4 object queries as an example. X: image features; Y: predictions; " + }, + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}" + }, + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "type": "text", + "content": ": ground-truth objects, where two color boxes mean two objects and two gray boxes mean dummy objects (no objects). The color lines between " + }, + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}" + }, + { + "bbox": [ + 46, + 173, + 547, + 228 + ], + "type": "text", + "content": " correspond to the assignment for ground-truth objects, and the gray lines for dummy objects. For clarity, the predictors are not explicitly included." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 75, + 376, + 171 + ], + "blocks": [ + { + "bbox": [ + 217, + 75, + 376, + 171 + ], + "lines": [ + { + "bbox": [ + 217, + 75, + 376, + 171 + ], + "spans": [ + { + "bbox": [ + 217, + 75, + 376, + 171 + ], + "type": "image", + "image_path": "0c936e43e665cc9a91553d92361731d2ed508aa8eb4560df8d7deb7c46e1eaf9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 386, + 76, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 386, + 76, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 386, + 76, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 386, + 76, + 545, + 171 + ], + "type": "image", + "image_path": "9c2225c2f3c5b82bb1720bcb3809bdf46b326f5bab2075d8676acc38ca25e521.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 248, + 287, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 284 + ], + "type": "text", + "content": "two designs, one-to-one assignment and self-attention over object queries, are critical for end-to-end detection without the need of the post-processing NMS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 287, + 286, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 286, + 381 + ], + "type": "text", + "content": "One-to-many assignment for non-end-to-end detection. One-to-many assignment is successfully adopted for introducing more supervision to non-end-to-end detection training, such as Faster R-CNN [33], FCOS [36], and so on [12, 23, 32, 45, 10, 3, 11]. One ground-truth object is assigned to multiple anchors or multiple pixels. During inference, a post-processing NMS is conducted for duplicate detection removal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 393, + 133, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 133, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 133, + 407 + ], + "type": "text", + "content": "3. Group DETR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 413, + 119, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 119, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 119, + 426 + ], + "type": "text", + "content": "3.1. Algorithm" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 540 + ], + "type": "text", + "content": "Naive one-to-many assignment. We start from a naive way for one-to-many assignment depicted in Figure 2 (c). We replace one-to-one assignment with one-to-many assignment: assign one ground-truth object to multiple predictions. It does not work and the performance is much low. The reason is that the model is trained to output multiple predictions for one ground-truth object, and lacks the scoring mechanism to promote one single prediction and demote duplicate predictions for one ground-truth object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": "Group-wise one-to-many assignment. We adopt the multi-group object query mechanism: form the initial " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " queries as the primary group and introduce more " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "(K - 1)" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " groups of " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " queries, totally " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " groups, " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}_1,\\mathbf{Q}_2,\\dots ,\\mathbf{Q}_K\\}" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": ". Accordingly, we have " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": " groups of predictions, " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Y}_1,\\mathbf{Y}_2,\\ldots ,\\mathbf{Y}_K\\}" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": ". We perform one-to-one assignment for each group, and find a bipartite matching " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\sigma_{k}(\\cdot)" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": ", between each group of predictions and the ground-truth objects " + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "inline_equation", + "content": "(\\mathbf{Y}_k,\\bar{\\mathbf{Y}})" + }, + { + "bbox": [ + 46, + 543, + 287, + 686 + ], + "type": "text", + "content": ". This results in that only one prediction for one ground-truth object is expected to score higher, and duplicate predictions is expected to score lower within one group other than within all the groups." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Separate self-attention. One-to-one assignment in one group means that the prediction assigned to one ground" + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 306, + 266, + 538, + 479 + ], + "blocks": [ + { + "bbox": [ + 307, + 247, + 529, + 260 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 529, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 529, + 260 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of one Group Decoder Layer" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 266, + 538, + 479 + ], + "lines": [ + { + "bbox": [ + 306, + 266, + 538, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 538, + 479 + ], + "type": "text", + "content": "SA: Self-Attention in the decoder layer \nCA: Cross-Attention in the decoder layer \nFFN: FFN in the decoder layer \nX: output image features of the encoder \nQ: object queries, with size (KxN, B, C) \nN, K, B, C: object query number, group number, batch size, feature dimension \n# group decoder \nif training: \n # split object queries to K groups \n Q_list = Q.split(N, dim=0) # a list of K tensors \n parallel_Q = cat(Q_list, dim=1) # (N, KxB, C) \n # parallel self-attention \n out = SA(parallel_Q) # (N, KxB, C) \n # concat all groups: (KxN, B, C) \n out = cat(out.split(B, dim=1), dim=0) \n # cross-attention and ffn \n out = FFN(CA(out, X)) \nelse: \n # in inference, only one group is kept \n Q = Q[:N] # (N, B, C) \n # self-attention, cross-attention, and ffn \n out = SA(Q) \n out = FFN(CA(out, X))" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 304, + 497, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 557 + ], + "type": "text", + "content": "truth object is superior to other predictions within the same group. This implies that we only need to collect the information of the predictions only from the same group, rather than from all the groups. Thus we perform self-attention (abbreviated as SA) over queries for each group separately:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 575, + 545, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 575, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 356, + 575, + 545, + 589 + ], + "type": "interline_equation", + "content": "\\mathrm {S A} \\left(\\mathbf {Q} _ {1}\\right), \\mathrm {S A} \\left(\\mathbf {Q} _ {2}\\right), \\dots , \\mathrm {S A} \\left(\\mathbf {Q} _ {\\mathrm {K}}\\right). \\tag {5}", + "image_path": "e2c4b86ce0a540eba7d0f25f7a37a7c644971cac0fcdffe021b9a979de72a182.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 597, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 646 + ], + "type": "text", + "content": "Training architecture. The resulting architecture for training is very simple: the encoder keeps the same, and the decoder contains " + }, + { + "bbox": [ + 304, + 597, + 545, + 646 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 597, + 545, + 646 + ], + "type": "text", + "content": " separate parallel decoders as shown in Figure 2 (a):" + } + ] + } + ], + "index": 15 + }, + { + "type": "code", + "bbox": [ + 317, + 652, + 545, + 712 + ], + "blocks": [ + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "lines": [ + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_1)\\to \\mathbf{Q}_1" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "text", + "content": " , Predictor " + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_1)\\rightarrow \\mathbf{Y}_1" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_2)\\to \\mathbf{Q}_2" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "text", + "content": " , Predictor " + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_2)\\rightarrow \\mathbf{Y}_2" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "text", + "content": " 1 \n" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\dots \\dots" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}(\\mathbf{X},\\mathbf{Q}_K)\\to \\mathbf{Q}_K" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "text", + "content": " , Predictor " + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_K)\\to \\mathbf{Y}_K" + }, + { + "bbox": [ + 317, + 652, + 545, + 712 + ], + "type": "text", + "content": " . (6)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "code", + "guess_lang": "latex" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6635" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 177, + 158 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 177, + 158 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 177, + 158 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 177, + 158 + ], + "type": "image", + "image_path": "9522b55fa2a5018eea96cfef98640f945b8f375448fa95c87ebec7111cdfc7f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 306, + 158 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 306, + 158 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 306, + 158 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 306, + 158 + ], + "type": "image", + "image_path": "1081bee59684bcc400d95ca07793734df052ae08daf2e1846c13781e9c9f0c43.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 372, + 156 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 372, + 156 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 372, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 372, + 156 + ], + "type": "image", + "image_path": "ec32b411ec159224dbbb9dad53b723b2a6da59224e62342efb92d5184f3d16f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 373, + 71, + 430, + 157 + ], + "blocks": [ + { + "bbox": [ + 373, + 71, + 430, + 157 + ], + "lines": [ + { + "bbox": [ + 373, + 71, + 430, + 157 + ], + "spans": [ + { + "bbox": [ + 373, + 71, + 430, + 157 + ], + "type": "image", + "image_path": "ffbd37c77265d99cf668502cff7206be81f70d0865c06e4352ee936debff5b27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 430, + 71, + 545, + 157 + ], + "blocks": [ + { + "bbox": [ + 430, + 71, + 545, + 157 + ], + "lines": [ + { + "bbox": [ + 430, + 71, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 545, + 157 + ], + "type": "image", + "image_path": "a04d147afc4193b479049f5ff7baf1009cabffc5f0e4aa8385383f64689b55a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 158, + 177, + 243 + ], + "blocks": [ + { + "bbox": [ + 48, + 158, + 177, + 243 + ], + "lines": [ + { + "bbox": [ + 48, + 158, + 177, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 177, + 243 + ], + "type": "image", + "image_path": "a9419d36f2984515e26de5ab598ba248c424dd4b7471d64d948754b3bada74e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 245, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 245, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 547, + 291 + ], + "type": "text", + "content": "Figure 3. Illustrating object queries. The predicted boxes and reference points corresponding to object queries in different groups for the same ground-truth object are plotted in different colors with one color for one group. It can be seen that these queries are spatially close and can be viewed as an augmentation of other queries. The results are from Group DETR over Conditional DETR-R50 [30]. The predicted boxes and reference points may overlap. Best view in color and zoom in." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 178, + 158, + 306, + 243 + ], + "blocks": [ + { + "bbox": [ + 178, + 158, + 306, + 243 + ], + "lines": [ + { + "bbox": [ + 178, + 158, + 306, + 243 + ], + "spans": [ + { + "bbox": [ + 178, + 158, + 306, + 243 + ], + "type": "image", + "image_path": "a9784ef6fd49f08ccb23918db15c8cdf5e64b54086b83bf129f75b43c007aaa7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 157, + 372, + 243 + ], + "blocks": [ + { + "bbox": [ + 306, + 157, + 372, + 243 + ], + "lines": [ + { + "bbox": [ + 306, + 157, + 372, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 372, + 243 + ], + "type": "image", + "image_path": "a26acde4218ddb55477e95d0e32474574290cd4f2270864bed14e148083040ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 373, + 157, + 429, + 243 + ], + "blocks": [ + { + "bbox": [ + 373, + 157, + 429, + 243 + ], + "lines": [ + { + "bbox": [ + 373, + 157, + 429, + 243 + ], + "spans": [ + { + "bbox": [ + 373, + 157, + 429, + 243 + ], + "type": "image", + "image_path": "ce1cca17dbc00d3a9371199a56bfdb612ad38bb2f994b3edf859b3054d033904.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 430, + 157, + 545, + 243 + ], + "blocks": [ + { + "bbox": [ + 430, + 157, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 430, + 157, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 430, + 157, + 545, + 243 + ], + "type": "image", + "image_path": "3145b97eb497989bfa3cd2a3a9e8420a41dc3a0de3683f38ce3f5ec8651c728b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 65, + 306, + 279, + 360 + ], + "blocks": [ + { + "bbox": [ + 65, + 306, + 279, + 360 + ], + "lines": [ + { + "bbox": [ + 65, + 306, + 279, + 360 + ], + "spans": [ + { + "bbox": [ + 65, + 306, + 279, + 360 + ], + "type": "image", + "image_path": "70d1ddf3064f2c3698fdd64b81105afc41dfaa3b642e2b963097a35535d94018.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 415 + ], + "type": "text", + "content": "Figure 4. The performance across groups of queries are similar. Only a " + }, + { + "bbox": [ + 46, + 369, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\pm 0.1" + }, + { + "bbox": [ + 46, + 369, + 287, + 415 + ], + "type": "text", + "content": " mAP is observed over the median (37.5 mAP). The mAP scores over the COCO val2017 are reported by a 12-epoch trained Conditional DETR-R50 with Group DETR." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 426, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 522 + ], + "type": "text", + "content": "Here, the parameters of the decoder and the predictor for the " + }, + { + "bbox": [ + 46, + 426, + 287, + 522 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 426, + 287, + 522 + ], + "type": "text", + "content": " groups are shared. Decoder separation and parallelism are feasible in that there is no interaction among queries for the other two operations, cross-attention and FFN. Our approach is called Group Decoder. In model inference, the process is the same as DETR trained normally and only needs one group of queries without any architecture modification. The pseudo-code is shown in Algorithm 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 524, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 287, + 548 + ], + "type": "text", + "content": "Loss function. The loss is an aggregation of " + }, + { + "bbox": [ + 46, + 524, + 287, + 548 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 524, + 287, + 548 + ], + "type": "text", + "content": " losses, each for one decoder. It is written as follows," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 73, + 554, + 287, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 554, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 73, + 554, + 287, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {k} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\sum_ {n = 1} ^ {N} \\ell \\left(\\mathbf {y} _ {\\sigma_ {k} (n)}, \\bar {\\mathbf {y}} _ {k n}\\right), \\tag {7}", + "image_path": "efd9b26b7edf42d50a0763e56d0e6b869e6b70c21977a77769a3897b2bbfd070.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\sigma_{k}(\\cdot)" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": " is the optimal permutation of " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": " indices for the " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": "th decoder." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 623, + 110, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 110, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 110, + 635 + ], + "type": "text", + "content": "3.2. Analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Explanation with parameter-shared models. We discuss Group DETR from the perspective of training multiple models with parameter sharing. Training with Group DETR can be regarded as simultaneously training " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " DETR models, which share the parameters of the encoder, the decoder, and the predictor, and only differ in the initialization" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 311, + 310, + 540, + 429 + ], + "blocks": [ + { + "bbox": [ + 311, + 310, + 540, + 429 + ], + "lines": [ + { + "bbox": [ + 311, + 310, + 540, + 429 + ], + "spans": [ + { + "bbox": [ + 311, + 310, + 540, + 429 + ], + "type": "image", + "image_path": "e49ac70fdb560e444b88fa461442230a463c441fda2202918a54fd0d721f8642.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 431, + 547, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 498 + ], + "type": "text", + "content": "Figure 5. More stable assignment. The " + }, + { + "bbox": [ + 304, + 431, + 547, + 498 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 431, + 547, + 498 + ], + "type": "text", + "content": "-axis corresponds to #epoch, and the " + }, + { + "bbox": [ + 304, + 431, + 547, + 498 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 431, + 547, + 498 + ], + "type": "text", + "content": "-axis corresponds to instability score (the score is introduced by DN-DETR [20], the lower the instability score, the more stable the label assignment) over COCO val2017. One can see that the assignment in Group DETR is more stable than DN-DETR and its baseline DAB-DETR." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "text", + "content": "of object queries. This leads to the shared parameters receive more back-propagated gradients. Thus, these parameters are better trained and accordingly the training process converges faster." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 566, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 625 + ], + "type": "text", + "content": "As a side benefit, we observe that Group DETR makes the assignment more stable, as shown in Figure 5. We speculate that the stability is because the improved network leads to more reliable predictions, and thus the assignment quality is better." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": "Explanation with object query augmentation. The multi-group object query mechanism introduces additional " + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "inline_equation", + "content": "(K - 1)" + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": " group of queries, which can be regarded as an augmentation of the primary group of queries. This is empirically illustrated in Figure 3. The reference points predicting the same objects are spatially close, and thus the corresponding object queries are similar. This may suggest that the multi" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6636" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 68, + 252, + 187 + ], + "blocks": [ + { + "bbox": [ + 83, + 68, + 252, + 187 + ], + "lines": [ + { + "bbox": [ + 83, + 68, + 252, + 187 + ], + "spans": [ + { + "bbox": [ + 83, + 68, + 252, + 187 + ], + "type": "image", + "image_path": "32d577bbcc44946ebca4f0e7dbb036985926944b4507c1694d099df05d404f7a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 287, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 213 + ], + "type": "text", + "content": "Figure 6. The parallel decoders in Group DETR are efficiently implemented as parallel self-attention, cross attention and FFN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 223, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 293 + ], + "type": "text", + "content": "group object query mechanism resembles data augmentation, and at each iteration, more automatically-learned augmented queries are included, which equivalently introduces more supervision for decoder training. The results in Figure 4 empirically suggest that different groups of augmented queries lead to similar results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 378 + ], + "type": "text", + "content": "The point about more supervision is also observed from the comparison between Equation 6 (for training with Group DETR) and Equation 2 (for normal DETR training). Group DETR training includes " + }, + { + "bbox": [ + 46, + 294, + 287, + 378 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 294, + 287, + 378 + ], + "type": "text", + "content": " pairs of image feature and object query group " + }, + { + "bbox": [ + 46, + 294, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{X},\\mathbf{Q}_1),(\\mathbf{X},\\mathbf{Q}_2),\\ldots ,(\\mathbf{X},\\mathbf{Q}_K)\\}" + }, + { + "bbox": [ + 46, + 294, + 287, + 378 + ], + "type": "text", + "content": ", and thus the loss contains more components as shown in Equation 7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 387, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 509 + ], + "type": "text", + "content": "Table 1. Illustrating that training with Group DETR improves both encoder and decoder. The encoder, including CNN and transformer encoders, is initialized from a trained Conditional DETR-R50 [30] with 50 epochs and the decoder is random initialized. (a) (Fixed, Single) = the encoder is not retrained, and the decoder is trained normally without using Group DETR. (b) (Fixed, Group) = the encoder is not retrained, and the decoder is with Group DETR. (c) (Group, Group) = the encoder and the decoder are trained with Group DETR. All the results are got through training with 50 epochs. (c) > (b) implies that Group DETR also improves the encoder training." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 52, + 512, + 284, + 563 + ], + "blocks": [ + { + "bbox": [ + 52, + 512, + 284, + 563 + ], + "lines": [ + { + "bbox": [ + 52, + 512, + 284, + 563 + ], + "spans": [ + { + "bbox": [ + 52, + 512, + 284, + 563 + ], + "type": "table", + "html": "
EncoderDecodermAPAPsAPmAPl
(a)FixedSingle40.6*20.244.059.3
(b)FixedGroup41.521.245.060.2
(c)GroupGroup42.9*22.246.661.6
", + "image_path": "13217cd77d7c4ca36a885af00bc0119b911cdd9e59a2e9eb1ea48e4216ddd162.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 563, + 287, + 591 + ], + "lines": [ + { + "bbox": [ + 48, + 563, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 287, + 591 + ], + "type": "text", + "content": "*: Training Conditional DETR with a trained encoder gives slightly lower performances than the one trained regularly, even though we train all components. New hyper-parameters may need to get better results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 686 + ], + "type": "text", + "content": "Encoder training improvement. The additional supervision introduces more box regression and classification supervision from more queries assigned to each ground-truth object. The gradients with more supervision are also backpropagated from the decoder to the encoder. It is presumable that the encoder also gets benefit, verified by the empirical results in Table 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Computation and memory complexity. Group DETR uses more decoders during training. It is expected that" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 75, + 547, + 167 + ], + "blocks": [ + { + "bbox": [ + 310, + 75, + 547, + 167 + ], + "lines": [ + { + "bbox": [ + 310, + 75, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 310, + 75, + 547, + 167 + ], + "type": "image", + "image_path": "30792bb91a4c8244db2788dc760d3e1f07d74facc80e163426dd6e72e8fe5e24.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 180, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 258 + ], + "type": "text", + "content": "Figure 7. Baseline models vs their Group DETR counterparts w.r.t training memory. The gray baseline represents using the naive implementation of attention modules. With a memory-efficient implementation [6], Group DETR does not bring much memory burden during training, only requires " + }, + { + "bbox": [ + 304, + 180, + 547, + 258 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{G}" + }, + { + "bbox": [ + 304, + 180, + 547, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 180, + 547, + 258 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{G}" + }, + { + "bbox": [ + 304, + 180, + 547, + 258 + ], + "type": "text", + "content": " more GPU memory with Conditional DETR [30] ('C-DETR' for short) and DAB-DETR [25]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 267, + 546, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 546, + 376 + ], + "type": "text", + "content": "Table 2. Group DETR outperforms baseline models with a similar training time. Conditional DETR [30] and DAB-DETR [25] serve as baseline models to compare the performances on COCO val2017 [24]. 'C-DETR' and 'w/ Group' are the abbreviations of 'Conditional DETR' and 'with Group DETR'. The entries noted by grey are the results of baseline models with the same training epochs (12 or 50 epochs) as Group DETR. To match the training times of Group DETR, we adopt longer training sheds for baselines (15 or 60 epochs). The training times are measured on 8 A100 GPUs in hours." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 380, + 547, + 539 + ], + "blocks": [ + { + "bbox": [ + 307, + 380, + 547, + 539 + ], + "lines": [ + { + "bbox": [ + 307, + 380, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 547, + 539 + ], + "type": "table", + "html": "
Modelw/ GroupHoursmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓4.632.614.735.048.3
5.834.415.137.351.3
5.637.618.240.755.9
C-DETR✓19.240.920.544.259.6
23.041.621.445.160.0
23.343.423.047.362.3
DAB-DETR✓5.635.216.738.651.6
7.036.317.139.452.5
6.639.119.742.556.8
DAB-DETR✓23.342.221.545.760.3
28.042.922.846.461.9
27.544.524.248.563.2
", + "image_path": "a01e6c14ede9b0898daa423aa12258ec34300ae477091f22b86956123d3babc1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": "Group DETR will bring additional training computation costs (FLOPs) as well as training memory costs. But the parallel decoders can be implemented as a single decoder by replacing normal self-attention with parallel self-attention (depicted in Figure 6) and we can use an efficient attention implementation, FlashAttention [6, 19]. As a result, Group DETR only takes a small increase in training GPU memory and training time. For example, with Conditional DETR [30] and DAB-DETR [25], the memory increases are just " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{G}" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{G}" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": " (Figure 7). The training time is increased by 5 minutes per epoch (from 23 minutes to 28 minutes and from 28 minutes to 33 minutes, respectively)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "We provide the results by increasing the training time" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6637" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 78, + 265, + 142 + ], + "blocks": [ + { + "bbox": [ + 50, + 78, + 265, + 142 + ], + "lines": [ + { + "bbox": [ + 50, + 78, + 265, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 78, + 265, + 142 + ], + "type": "image", + "image_path": "1308bcc871b6bf390a46f7638013b7f1e0dfbeb8dc058fcccc1d78decfe3c630.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 149, + 288, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 288, + 227 + ], + "type": "text", + "content": "Figure 8. Comparisons with DN-DETR. Group DETR outperforms DN-DETR on DAB-DETR [25] (" + }, + { + "bbox": [ + 46, + 149, + 288, + 227 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 149, + 288, + 227 + ], + "type": "text", + "content": "-axis). Combining those two methods give better results, indicating they are complementary to each other. The " + }, + { + "bbox": [ + 46, + 149, + 288, + 227 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 149, + 288, + 227 + ], + "type": "text", + "content": "-axis is the mAP scores with a 12-epoch schedule on COCO val2017. * represents that we report the best results of DN-DETR among different numbers of denoising queries (detailed results are provided in Appendix)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 315 + ], + "type": "text", + "content": "for normal DETR training to see if Group DETR benefits simply from more training time. The results given in Table 2 show that normal training with more training time brings a little benefit and the performance is still much lower than Group DETR, implying that the performance gain from our approach is not from training time increase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 317, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 288, + 460 + ], + "type": "text", + "content": "Connection to DN-DETR. DN-DETR [20] aims to stabilize one-to-one assignment during DETR training. DN-DETR forms the additional queries by adding the noises to ground-truth objects, which can be regarded as a variant of our multi-group mechanism with clear differences. In DN-DETR [20], on the one hand, the number of queries within each additional group is the same as the number of ground-truth objects. Each one corresponds to one ground-truth object, and there is no query corresponding to no-object. In contrast, our approach automatically learns a number of " + }, + { + "bbox": [ + 46, + 317, + 288, + 460 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 317, + 288, + 460 + ], + "type": "text", + "content": " (e.g., 300) object queries that correspond to both ground-truth objects and no-object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": "On the other hand, DN-DETR performs self-attention over noised queries, mainly for collecting the information from predictions for other objects other than from duplicate predictions. Self-attention in Group DETR instead collects both duplicate predictions and predictions for other objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "content": "The above two comparisons imply that DN-DETR brings the major help for the box and classification prediction, through the introduction of more positive queries corresponding to ground-truth objects (like FCOS), and no direct help for duplicate prediction removal. Our approach introduces both positive queries and negative queries (no-object), also brings the help for duplicate prediction removal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Figure 8 shows that the performance of Group DETR is better than DN-DETR. We further investigate if Group DETR still benefits from introducing more positive queries with noised queries. As shown in Figure 8, the performance gain over Group DETR is non-trivial, a " + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": ". This implies that Group DETR and DN-DETR are complementary and their major roles are different, though they have some similarities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 188 + ], + "type": "text", + "content": "We demonstrate the effectiveness of Group DETR in various DETR variants, and its extension to 3D detection and instance segmentation [30, 25, 20, 47, 44, 26, 27, 5]. The training setting is almost the same as baseline models, for illustrating the effectiveness of our Group DETR. We adopt the same training settings and hyper-parameters as the baseline models, such as learning rate, optimizer, pre-trained model, initialization methods, and data augmentations1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 196, + 408, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 196, + 408, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 408, + 209 + ], + "type": "text", + "content": "4.1. Object Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 214, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 547, + 335 + ], + "type": "text", + "content": "Setting. We study various representative DETR-based detectors, such as basic baselines (Conditional DETR [30], DAB-DETR [25], DN-DETR [20]) with dense attentions, and strong baselines (DAB-Deformable-DETR [25, 47] and DINO [44, 47]) with deformable attentions. We report the results on two training schedules, training for 12 epochs and training for more epochs (36 or 50). Unless specified, the models are trained with ResNet-50 [13] as the backbone on the COCO train2017 and evaluated on the COCO val2017. More implementation details are provided in Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 337, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 421 + ], + "type": "text", + "content": "Results. We first report the results of training with 12 epochs in Table 3. Group DETR brings consistent improvements over the baselines with dense attentions that already are superior to the original DETR [2]. It boosts Conditional DETR (-DC5) [30] by 5.0 (4.8) mAP, improves DAB-DETR (-DC5) [25] by 3.9 (4.4) mAP, and brings a 2.0 (2.6) mAP gain to DN-DETR (-DC5) [20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 540 + ], + "type": "text", + "content": "Group DETR also works well on those strong baselines with deformable attentions that are equipped with two or more accelerating techniques. It gives a 1.5 mAP improvements over DAB-Deformable-DETR [25, 47]. When applying to DINO [44, 20, 47], Group DETR also exceeds it by 0.7 mAP. The gain is non-trivial over such a stronger baseline, considering that DINO is a well-tuned model based on DAB-Deformable-DETR that combines improved hyperparameters, improved two-stage design, improved query denoising task, and other tricks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "text", + "content": "Furthermore, we report the results with 50 training epochs that is commonly adopted in many acceleration methods [47, 30, 4, 25]. Table 4 presents that Group DETR outperforms baseline models by large margins. For the stronger backbone, Swin-Large [29], our approach achieves " + }, + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "inline_equation", + "content": "58.4\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "text", + "content": " (still a " + }, + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "text", + "content": " higher than its baseline DINO [44] (" + }, + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "inline_equation", + "content": "58.0\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 541, + 547, + 638 + ], + "type": "text", + "content": " with Swin-Large)). This verifies the generalization ability of our Group DETR." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 645, + 545, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 545, + 674 + ], + "type": "text", + "content": "1We may adjust the batch size due to the limitation of the GPU memory size for both the baseline model and our approach so that the batch size is the same." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 674, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 674, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 674, + 547, + 713 + ], + "type": "text", + "content": "2In fact, our approach is compatible with query denoising and two-stage. In Table 3, for example, DN-DETR [20] utilizes query denoising, and our method improves it by 2.0. Similarly, DAB-D-DETR [25] adopts a two-stage structure, and our method achieves a 1.5 improvement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6638" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 68, + 289, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 68, + 289, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 289, + 201 + ], + "type": "text", + "content": "Table 3. Effectiveness of Group DETR with 12 epochs. Group DETR gives consistent gains over various DETR-based baselines on COCO val2017 [24], highlighted with brackets. All experiments adopt ResNet-50 [13] and do not use multiple patterns [40]. For DN-DETR, an improved version of DN, dynamic DN groups [44] with 100 DN queries, is used, making the results slightly different from the ones (with 3 patterns) reported in the original paper [20] (more results about the number of DN queries can be found in Appendix. 'C-DETR', 'DAB-D-DETR', and 'w/ Group' are 'Conditional DETR' [30], 'DAB-Deformable DETR' [25, 47], and 'with Group DETR', respectively, for neat representation." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 205, + 289, + 455 + ], + "blocks": [ + { + "bbox": [ + 48, + 205, + 289, + 455 + ], + "lines": [ + { + "bbox": [ + 48, + 205, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 289, + 455 + ], + "type": "table", + "html": "
Modelw/ GroupmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓32.614.735.048.3
37.6 (+5.0)18.240.755.9
C-DETR-DC5✓36.418.039.652.5
41.2 (+4.8)21.445.058.7
DAB-DETR✓35.216.738.651.6
39.1 (+3.9)19.742.556.8
DAB-DETR-DC5✓37.519.440.653.2
41.9 (+4.4)23.345.658.4
DN-DETR✓38.617.941.657.7
40.6 (+2.0)19.843.959.4
DN-DETR-DC5✓41.922.245.159.8
44.5 (+2.6)25.948.262.2
DAB-D-DETR✓44.227.547.158.6
45.7 (+1.5)28.149.060.6
DINO-4scale✓49.432.352.563.2
50.1 (+0.7)32.453.264.7
", + "image_path": "889ff06c856a073276fd2b5959bc062dd8e48733a8848235b328c5fbdcf0bd02.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 469, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 529 + ], + "type": "text", + "content": "Last, we compare the training convergence curves of the baseline models and their Group DETR counterparts. The results, as shown in Figure 1, provide more evidence that Group DETR speeds DETR training convergence on various DETR variants." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 664 + ], + "type": "text", + "content": "System-level Results on COCO test-dev with ViT-Huge. We also have the system-level performance on COCO test-dev [24] with ViT-Huge [7]. We apply Group DETR to DINO [44] and follow its training pipeline and settings: pretrain the encoder with a self-supervised method, then pretrain the whole model on Object365 [34], and last fine-tune the whole model on COCO [24]. Our model is the first to achieve 64.5 mAP on COCO test-dev, which is still superior to other methods with larger encoder and more pretraining data [28, 39, 41, 42]. The details and comparisons with other methods are provided in Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 671, + 157, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 157, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 157, + 685 + ], + "type": "text", + "content": "4.2. More Applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Group DETR is applicable to DETR-style techniques to other vision problems. We report the results for two ad" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 68, + 547, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 68, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 547, + 148 + ], + "type": "text", + "content": "Table 4. Effectiveness of Group DETR with more epochs. Group DETR still outperforms baselines by non-trivial margins with more training epochs (36 or 50 epochs). Settings and notations are consistent with Table 3, except for the training epochs (36 epochs for DINO-4scale by following the original paper [44] and 50 epochs for other models). 'DINO-4scale-Swin-L' means it adopts Swin-Large [29] as the backbone." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 150, + 547, + 384 + ], + "blocks": [ + { + "bbox": [ + 307, + 150, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 384 + ], + "type": "table", + "html": "
Modelw/ GroupmAP\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
C-DETR✓40.920.544.259.6
43.4 (+2.5)23.047.362.3
C-DETR-DC5✓43.723.947.660.1
45.8 (+2.1)26.849.763.1
DAB-DETR✓42.221.545.760.3
44.5 (+2.3)24.248.563.2
DAB-DETR-DC5✓44.525.348.262.3
46.7 (+2.2)27.650.964.0
DN-DETR✓44.023.947.762.9
45.4 (+1.4)25.149.363.8
DN-DETR-DC5✓47.527.950.765.9
48.0 (+0.5)29.352.165.4
DAB-D-DETR✓48.131.451.463.4
49.7 (+1.6)31.452.565.6
DINO-4scale✓50.934.654.164.6
51.3 (+0.4)34.754.565.3
DINO-4scale-Swin-L✓58.041.361.974.0
58.4 (+0.4)41.062.573.9
", + "image_path": "df0724423e89d3944475f399c0b89d7e8e9319b1a646126ce01c0a46fe05a94b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 399, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 436 + ], + "type": "text", + "content": "ditional problems: multi-view 3D object detection [15, 22, 26, 27] and instance segmentation [5, 21], to further demonstrate the effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 441, + 547, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 513 + ], + "type": "text", + "content": "Multi-view 3D object detection. We report the results over PETR [26] and PETR v2 [27] on the nuScenes val dataset [1]. Table 5 shows that Group DETR brings significant gains to PETR and PETR v2 with 24 training epochs in terms of both the nuScenes Detection Score (NDS) and mAP scores." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 523, + 547, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 547, + 580 + ], + "type": "text", + "content": "Table 5. Results on multi-view 3D object detection. All experiments are evaluated on the nuScenes val set [1]. We train these experiments for 24 epochs with VoVNetV2 [18] as the backbone and with the image size of " + }, + { + "bbox": [ + 304, + 523, + 547, + 580 + ], + "type": "inline_equation", + "content": "800 \\times 320" + }, + { + "bbox": [ + 304, + 523, + 547, + 580 + ], + "type": "text", + "content": ". We follow all the settings and hyper-parameters of PETR [26] and PETR v2 [27]." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 326, + 583, + 525, + 648 + ], + "blocks": [ + { + "bbox": [ + 326, + 583, + 525, + 648 + ], + "lines": [ + { + "bbox": [ + 326, + 583, + 525, + 648 + ], + "spans": [ + { + "bbox": [ + 326, + 583, + 525, + 648 + ], + "type": "table", + "html": "
Modelw/ GroupNDSmAP
PETR✓42.037.4
45.0 (+3.0)38.8 (+1.4)
PETR v2✓50.340.7
51.3 (+1.0)41.9 (+1.2)
", + "image_path": "eeb91a45ad2ba4467d184ea2e4e6eb2e71e467ca07c37d38581188fd86334b88.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Instance segmentation. We demonstrate the effectiveness of the representative method, Mask2Former [5]. The results are given in Table 6. Group DETR achieves a 1.2 (0.3) " + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP}^m" + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": " gain with 12 (50) epochs." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6639" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 132, + 277, + 196 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "type": "text", + "content": "Table 6. Results on instance segmentation. The mask mAP " + }, + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "type": "inline_equation", + "content": "(\\mathrm{mAP}^m)" + }, + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "type": "text", + "content": " is used for instance segmentation on COCO val2017. We adopt Mask2Former [5] as the baseline. The experiments are conducted with ResNet-50 [13] as the backbone, following all the settings of Mask2Former." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 132, + 277, + 196 + ], + "lines": [ + { + "bbox": [ + 56, + 132, + 277, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 277, + 196 + ], + "type": "table", + "html": "
Epochsw/ GroupmAPmAPSmAPmAPl
1238.517.641.460.4
12✓39.7 (+1.2)18.742.860.8
5043.723.447.264.8
50✓44.0 (+0.3)23.847.165.1
", + "image_path": "b92bd8ed40e83ee34a4c636907af621de756bc96f716f430ddce8dfbf6f2e9b9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 307, + 287, + 372 + ], + "blocks": [ + { + "bbox": [ + 47, + 203, + 288, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 203, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 288, + 304 + ], + "type": "text", + "content": "Table 7. Effects of group-wise one-to-many assignment and separate self-attention. (a) baseline: one-to-one assignment with 300 object queries. (b) naive one-to-many assignment with 3300 object queries for training and inference. (c) group-wise one-to-many assignment and no separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. (d) group-wise one-to-many assignment and separate self-attention with 11 groups of 300 queries, inference with a group of 300 queries. o2m = one-to-many, Sep. SA = separate self-attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 307, + 287, + 372 + ], + "lines": [ + { + "bbox": [ + 48, + 307, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 287, + 372 + ], + "type": "table", + "html": "
o2mSep. SAmAPAPsAPmAPl
(a)××32.614.434.948.6
(b)Naive×8.48.013.213.3
(c)Group×34.816.437.751.4
(d)Group✓37.618.240.755.9
", + "image_path": "9664a764953eb2e4791994bdf9b81fbe6a5196f20a84fe48afe50ad99d5d52a2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 394, + 141, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 141, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 141, + 407 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 415, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 487 + ], + "type": "text", + "content": "We conduct the ablation study by using Conditional DETR [30] as the baseline. The CNN backbone is ResNet-50 [13], and the training epoch nubmer is 12. The performances are evaluated on COCO val2017 [24]. We mainly study the effects of the key design: group-wise one-to-many assignment, separate self-attention, and group number." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "text", + "content": "Group-wise one-to-many assignment and separate self-attention. Table 7 shows how group-wise one-to-many (o2m) assignment and separate self-attention make contributions. In comparison to the baseline (a), group-wise o2m assignment improves the mAP score from " + }, + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "inline_equation", + "content": "32.6\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "inline_equation", + "content": "34.8\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "text", + "content": ": with the gain 2.2. The separate self attention (Sep. SA) further gets a " + }, + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "inline_equation", + "content": "2.8\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 493, + 287, + 636 + ], + "type": "text", + "content": " gain. In addition, we report naive one-to-many assignment. The results are very poor, which is reasonable in that there are duplicate predictions and there is a lack of scoring mechanisms for demoting them. The results suggest that both group-wise o2m assignment and separate self-attention are effective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "content": "Group number. Figure 9 shows the influence of the number of groups " + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "content": " in Group DETR. The detection performance improves when increasing the number of groups, and becomes stable when the group number reaches 11. Thus, we adopt " + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "inline_equation", + "content": "K = 11" + }, + { + "bbox": [ + 46, + 641, + 287, + 712 + ], + "type": "text", + "content": " by default in Group DETR in our experiments." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 542, + 143 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 542, + 143 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 542, + 143 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 542, + 143 + ], + "type": "image", + "image_path": "e0cac6c4c1638a11f233bde704afcfa502bcb1d6b62f3a5b04992320bb67813a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 146, + 547, + 180 + ], + "lines": [ + { + "bbox": [ + 305, + 146, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 547, + 180 + ], + "type": "text", + "content": "Figure 9. Influence of group number. The " + }, + { + "bbox": [ + 305, + 146, + 547, + 180 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 146, + 547, + 180 + ], + "type": "text", + "content": "-axis is the number of groups. It can be seen that the performance becomes stable when the number of groups reaches 11." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 201, + 398, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 398, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 398, + 213 + ], + "type": "text", + "content": "5. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 221, + 546, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 546, + 270 + ], + "type": "text", + "content": "There are two main lines for accelerating DETR training: modify cross-attention and stabilize one-to-one assignment. The two are complementary and can be combined to further boost the performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 273, + 546, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 546, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 546, + 416 + ], + "type": "text", + "content": "Modifying cross-attention. Cross-attention module aims to collect the information from the image features useful to classification and localization. Various methods are proposed to select the informative image regions more efficiently and effectively [8, 4, 40, 43, 25, 9]. For example, Deformable attention [47] selects the highly informative positions dynamically according to the previous decoder embedding. Conditional DETR [4] instead continues to use the normal global attention, and dynamically computes the spatial attention to softly select the informative regions. SMCA [8] uses the Gaussian-like weight for spatial modulation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 420, + 547, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 577 + ], + "type": "text", + "content": "Stabilizing one-to-one assignment. DETR [2] relies on one-to-one assignment, where each ground-truth object is assigned to a single prediction through building a bipartite matching between the predictions and the ground-truth objects. DN-DETR [20] finds the assignment process is unstable and attributes the slow convergence issue to the instabilities. Thus, DN-DETR [20] introduces groups of noisy queries by adding noises to ground-truth objects, to stabilize the assignment, leading to faster convergence. DINO [44] makes further improvement through contrastive denoising training to generate both positive and negative noise queries with different noise levels. Our approach studies the assignment mechanism instead for introducing more supervision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 581, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 664 + ], + "type": "text", + "content": "One-to-many assignment. One-to-many assignment is widely adopted in deep detectors [33, 12, 23, 36], and has attracted a lot of interest [45, 17, 46, 10, 3, 38, 35]. For example, Faster R-CNN [33] and FCOS [36] produce multiple positive anchors and pixels for each ground-truth object. In this paper, we investigate one-to-many assignment in a feasible manner for the end-to-end detector DETR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "content": "Concurrent with our work, " + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "content": "-DETR [16] also uses one-to-many assignment to speed up DETR training convergence. Our Group DETR and " + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "content": "-DETR are related, but different: (1) Group DETR introduces group-wise one-to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6640" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "many assignment with separate self-attention with the same number of object queries in each group. " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "-DETR adopts hybrid assignments in two different groups: One group uses one-to-one assignment and another uses one-to-many assignment with more object queries. (2) All the decoders in Group DETR can be used for inference. But the additional decoder in " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "-DETR is not directly used and requires NMS for inference. (3) During training, our architecture introduces one parameter: the number of groups. In contrast, " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "-DETR introduces the number of additional queries and the number of additional positive queries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 265 + ], + "type": "text", + "content": "DETA [31] is another concurrent work with our Group DETR. DETA directly uses one-to-many assignment and brings NMS back to DETR frameworks. While our method provides group-wise one-to-many assignment and maintains end-to-end detection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 279, + 119, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 119, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 119, + 292 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 445 + ], + "type": "text", + "content": "The key points in Group DETR include group-wise one-to-many assignment and parallel self-attention. The success stems from involving more groups of object queries as an addition to the primary group of object queries, and thus introducing more supervision. Group-wise assignment mechanism makes sure that the competition among predictions happens within each group separately, and separate self-attention eases the training. Thus, the NMS pose-processing is not necessary, and the inference process is kept the same as normally trained DETR and not dependent on the group design. Our approach is simple, easily implemented, and general." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 445, + 289, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 289, + 542 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the Sichuan Science and Technology Program (2023YFSY0008), National Natural Science Foundation of China (61632003, 61375022, 61403005), Grant SCITLAB-20017 of Intelligent Terminal Key Laboratory of SiChuan Province, Beijing Advanced Innovation Center for Intelligent Robots and Systems (2018IRS11), and PEK-SenseTime Joint Laboratory of Machine Vision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 555, + 107, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 107, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 107, + 568 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 577, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 52, + 577, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 577, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 52, + 577, + 287, + 632 + ], + "type": "text", + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 634, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 634, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 634, + 287, + 677 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 287, + 714 + ], + "type": "text", + "content": "[3] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In CVPR, pages 13039-13048, 2021. 3, 8" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "type": "text", + "content": "[4] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 118, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 118, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 311, + 118, + 547, + 160 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. arXiv preprint arXiv:2112.01527, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 162, + 547, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 162, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 311, + 162, + 547, + 204 + ], + "type": "text", + "content": "[6] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 205, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 205, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 205, + 547, + 259 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 261, + 547, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 547, + 303 + ], + "type": "text", + "content": "[8] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In ICCV, pages 3621–3630, 2021. 1, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 304, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 304, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 545, + 337 + ], + "type": "text", + "content": "[9] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5364-5373, 2022. 1, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "text", + "content": "[10] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In CVPR, pages 303-312, 2021. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "text", + "content": "[11] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 404, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 425 + ], + "type": "text", + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 460, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 480 + ], + "type": "text", + "content": "[14] J Hosang, R Benenson, and B Schiele. Learning nonmaximum suppression. PAMI, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "text", + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 548 + ], + "type": "text", + "content": "[16] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "text", + "content": "[17] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371. Springer, 2020. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "text", + "content": "[18] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In CVPR, pages 13906-13915, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 615, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 615, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 615, + 545, + 679 + ], + "type": "text", + "content": "[19] Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[20] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In CVPR, 2022. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6641" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[21] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "text", + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 1, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 249 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "type": "text", + "content": "[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 337 + ], + "type": "text", + "content": "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "text", + "content": "[28] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "text", + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "text", + "content": "[30] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, pages 3651-3660, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 557 + ], + "type": "text", + "content": "[31] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 581 + ], + "type": "text", + "content": "[32] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 613 + ], + "type": "text", + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 2, 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 669 + ], + "type": "text", + "content": "[34] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 703 + ], + "type": "text", + "content": "[35] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In ICML, pages 9934-9944. PMLR, 2021." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 566 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[36] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 2, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[38] Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end object detection with fully convolutional network. In CVPR, pages 15849-15858, 2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 185, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 240 + ], + "type": "text", + "content": "[39] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[40] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 275, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 328 + ], + "type": "text", + "content": "[41] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. arXiv preprint arXiv:2205.14141, 2022.7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 353 + ], + "type": "text", + "content": "[42] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[43] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 1, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 476 + ], + "type": "text", + "content": "[45] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "text", + "content": "[46] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[47] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. CoRR, abs/2010.04159, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 69, + 704, + 74, + 711 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 704, + 74, + 711 + ], + "spans": [ + { + "bbox": [ + 69, + 704, + 74, + 711 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6642" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_content_list.json b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..badd56dab6f15231646310db8778d7c80b9ac474 --- /dev/null +++ b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_content_list.json @@ -0,0 +1,1609 @@ +[ + { + "type": "text", + "text": "Group Pose: A Simple Baseline for End-to-End Multi-person Pose Estimation", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huan Liu $^{1,3*}$ Qiang Chen $^{2*}$ Zichang Tan $^{2}$ Jiang-Jiang Liu $^{2}$ Jian Wang $^{2}$ Xiangbo Su $^{2}$ Xiaolong Li $^{1,3}$ Kun Yao $^{2}$ Junyu Han $^{2}$ Errui Ding $^{2}$ Yao Zhao $^{1,3\\dagger}$ Jingdong Wang $^{2}$", + "bbox": [ + 122, + 179, + 846, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Science, Beijing Jiaotong University $^{2}$ Baidu VIS $^{3}$ Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing, China", + "bbox": [ + 86, + 219, + 880, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 290, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we study the problem of end-to-end multiperson pose estimation. State-of-the-art solutions adopt the DETR-like framework, and mainly develop the complex decoder, e.g., regarding pose estimation as keypoint box detection and combining with human detection in ED-Pose [38], hierarchically predicting with pose decoder and joint (keypoint) decoder in PETR [27].", + "bbox": [ + 75, + 321, + 468, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a simple yet effective transformer approach, named Group Pose. We simply regard $K$ -keypoint pose estimation as predicting a set of $N \\times K$ keypoint positions, each from a keypoint query, as well as representing each pose with an instance query for scoring $N$ pose predictions.", + "bbox": [ + 75, + 428, + 468, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by the intuition that the interaction, among across-instance queries of different types, is not directly helpful, we make a simple modification to decoder self-attention. We replace single self-attention over all the $N \\times (K + 1)$ queries with two subsequent group self-attention: (i) $N$ within-instance self-attention, with each over $K$ keypoint queries and one instance query, and (ii) $(K + 1)$ same-type across-instance self-attention, each over $N$ queries of the same type. The resulting decoder removes the interaction among across-instance type-different queries, easing the optimization and thus improving the performance. Experimental results on MS COCO and CrowdPose show that our approach without human box supervision is superior to previous methods with complex decoders, and even is slightly better than ED-Pose that uses human box supervision. Paddle1 and PyTorch2 codes are available.", + "bbox": [ + 75, + 503, + 470, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 787, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-person pose estimation aims to detect the corresponding human keypoints for all human instances in an", + "bbox": [ + 76, + 811, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/437a2e132d7408f4d6ab28929d8ba413a7667cc3aa9779c1387b4dd615c4c1bb.jpg", + "image_caption": [ + "(a) PETR" + ], + "image_footnote": [], + "bbox": [ + 509, + 287, + 875, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ddb363828897d62896c274eff862023799eacec0f75c98e3634b7eb5e5fd16d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 333, + 874, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73df08414980fcafaa87f02a383e1d47af7fa6b19897415042a65ab8f4ec3e3f.jpg", + "image_caption": [ + "(b) ED-Pose", + "(c) Group Pose", + "Figure 1: Comparison of transformer decoders. Here we mainly illustrate the overview of the decoder part of PETR [27], ED-Pose [38], and our Group Pose. The three end-to-end frameworks differ in query design and decoder architecture. Group Pose only uses a simple transformer decoder rather than developing complex decoders. 'inst' and 'kpt' represent for instance and keypoint." + ], + "image_footnote": [], + "bbox": [ + 509, + 383, + 872, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image. Previous frameworks include top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5] that divide the task into two sequential sub-tasks: human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping [24]. Another line in previous frameworks is one-stage methods [22, 26, 31, 34], which directly predict instance-aware keypoints. These frameworks rely on non-differentiable hand-crafted post-processes [10, 24], which complicate the pipelines and challenge the optimizations. Inspired by the success of DETR [2] in object detection, building an end-to-end framework for multi-person pose estimation has seen significant interest.", + "bbox": [ + 496, + 549, + 890, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches follow the DETR framework [2, 41, 39], with the transformer encoder-decoder architecture for multi-person pose estimation, as shown in Figure 1. PETR [27] hierarchically predicts the keypoint positions and uses two subsequent decoders, pose decoder and joint decoder, with two different queries, pose query (a person has one pose query) for pose decoder, and keypoint queries for joint decoder. ED-Pose [38] transfers pose estimation to a keypoint box detection problem, and learns a content query and a box query for each keypoint position prediction", + "bbox": [ + 496, + 750, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. Work done when H. Liu is an intern at Baidu.", + "bbox": [ + 93, + 849, + 450, + 862 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. Email: yzhao@bjtu.edu.cn", + "bbox": [ + 98, + 862, + 364, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code of Paddle: https://github.com/Michel-liu/GroupPose-Paddle", + "bbox": [ + 98, + 875, + 447, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Code of PyTorch: https://github.com/Michel-liu/GroupPose", + "bbox": [ + 96, + 887, + 418, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15029", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with using the box size to process the query.", + "bbox": [ + 76, + 90, + 372, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a simple yet effective transformer approach, named Group Pose, for end-to-end multiperson pose estimation. Instead of using a single query to predict and score one pose, similar to ED-Pose, we use the $N \\times K$ keypoint queries to regress the $N \\times K$ positions, by regarding each keypoint as an object, as well as $N$ instance queries, each representing a $K$ -keypoint pose for scoring the $K$ -keypoint pose prediction.", + "bbox": [ + 76, + 117, + 468, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We make a simple modification for the decoder architecture. We replace standard self-attention in the decoder with two subsequent group self-attention: $N$ parallel self-attention with each over $K$ keypoint queries and the corresponding instance query for exploiting kinematic relation and gathering information for scoring pose predictions, and $(K + 1)$ parallel self-attention with each over $N$ queries of the same type for collecting duplicate prediction information like self-attention of the original DETR.", + "bbox": [ + 76, + 250, + 470, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The two self-attention captures two kinds of interactions: (i) $N$ within-instance interactions over $K$ keypoint queries and one instance query, (ii) $(K + 1)$ across-instance interactions over $N$ queries of the same type (e.g., nose keypoint query or instance query). The extra interactions in standard self-attention are (iii) across-instance interactions for queries with different types, which is not directly useful. Empirical results show that the removal of the third kind of interactions eases the optimization and thus improves the performance.", + "bbox": [ + 76, + 395, + 470, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The design about self-attention is different from the closely-related approach ED-Pose [38]. On the one hand, in addition to removing the third interactions, ED-Pose only models across-instance interactions for the instance queries. In contrast, our approach also models across-instance interactions for the $K$ keypoint types. On the other hand, our approach separates the two interactions using two subsequent group self-attention, explicitly exploring the information about queries belonging to the same human instance, and with the same type. ED-Pose couples the two interactions using a single masked self-attention with the third interactions masked. It is empirically demonstrated that the two differences benefit the pose estimation performance.", + "bbox": [ + 76, + 556, + 468, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results show that our simple approach Group Pose without human box supervision surprisingly outperforms the recent end-to-end methods with human box supervisions on MS COCO [16] and CrowdPose [14]. Notably, Group Pose achieves 72.0 AP with ResNet-50 [9] and 74.8 AP with Swin-Large [18] on MS COCO val2017. We hope our simple transformer decoder in Group Pose will motivate people to simplify the design in end-to-end multiperson pose estimation.", + "bbox": [ + 76, + 763, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-person pose estimation is a challenging task that aims to detect the corresponding human keypoints for all human instances in an image. Previous methods usually adopt complex frameworks to address it, which are divided into non-end-to-end and end-to-end methods.", + "bbox": [ + 498, + 119, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-end-to-end methods. There are typically two types: two-stage and one-stage. Two-stage frameworks, including top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5], split the multi-person pose estimation task into two sequential sub-tasks, human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping. In top-down methods, an object detector is first employed to detect the boxes of human instances, which is then cropped for single-person pose estimation in each box. Bottom-up methods first predict all human keypoints in a human-agnostic way then group them into instances. While one-stage frameworks [40, 22, 26, 31, 34] directly predict instance-aware keypoints. These methods require hand-crafted pose-processes, such as NMS [10] or grouping [24], which complicate the pipelines and challenge the optimizations. In this paper, we focus on concise ways, end-to-end frameworks.", + "bbox": [ + 498, + 205, + 892, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end methods. Current end-to-end multi-person pose estimation frameworks [27, 36, 38] are built by following the designs of DETR [2] and its variants [41, 23, 3, 17, 13, 39]. They adopt a paradigm of splitting the multiperson pose estimation task into two sub-processes. For example, PETR [27] views the task as a hierarchical set prediction problem. It first determines human instances by predicting a set of poses with a pose decoder and then refines the keypoints in each pose with a joint (keypoint) decoder. There are also two different types of queries, pose query (a person has one pose query) for pose decoder and keypoint queries for joint decoder. QueryPose [36] and ED-Pose [38] follow this end-to-end paradigm but further incorporate an extra human detection task. QueryPose [36] follows Sparse R-CNN [30] to build two parallel RoIAlign-based [8] decoders to perform human detection and pose estimation, respectively. ED-Pose [38] transfers pose estimation task to a keypoint box detection problem. It first employs a human detection decoder [39] to determine the human instances with box queries. Then it builds a human-to-keypoint detection decoder with a content query and a box query for each keypoint position, collecting contextual information near keypoint positions.", + "bbox": [ + 498, + 470, + 892, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these end-to-end methods show promising results in multi-person pose estimation, they rely on complex decoders. Our Group Pose, on the other hand, adopts a simple transformer decoder, improving the performance and simplifying the process.", + "bbox": [ + 498, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15030", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/27218fbf40913cf0091d027d05ad545804fd92c209cec145c29f801bee2eff7a.jpg", + "image_caption": [ + "Figure 2: Our Group Pose architecture. The backbone takes an image as input and outputs image features, which are refined by the transformer encoder. To directly predict $N$ human poses with $K$ keypoint positions in each pose, we adopt $N \\times (K + 1)$ queries, containing $N$ instance queries for scoring poses and $N \\times K$ keypoint queries for regressing positions." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 851, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Group Pose", + "text_level": 1, + "bbox": [ + 76, + 273, + 199, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Group Pose is a simple yet effective end-to-end multi-person pose estimation framework. We follow previous end-to-end frameworks [27, 38] to view the multi-person pose estimation task as a set prediction problem, but directly adopt a simple transformer decoder [41] instead of complex decoders, simplifying the process. Next, we introduce the key elements of Group Pose.", + "bbox": [ + 75, + 299, + 467, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 415, + 186, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall structure of Group Pose is depicted in Figure 2. Group Pose consists of a backbone [9, 18], a transformer encoder [32], a transformer decoder, and task-specific prediction heads. This framework enables Group Pose to simultaneously regresses $K$ keypoints (e.g., $K = 17$ on MS COCO) for $N$ human instances given an image.", + "bbox": [ + 75, + 439, + 467, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backbone and transformer encoder. We directly follow DETR frameworks [41] to build the backbone and the transformer encoder (with 6 deformable transformer layers [41]) for Group Pose. It takes an image as input and outputs the extracted multi-level features, which serve as inputs for the following transformer decoder. We use 4 feature levels in Group Pose, with downsampling rates of $\\{8, 16, 32, 64\\}$ .", + "bbox": [ + 75, + 534, + 467, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer decoder. For transformer decoder, we adopt a combination of $N \\times K$ keypoint queries and $N$ instance queries as input instead of using a single query to predict and score one pose. The keypoint queries regard each keypoint as an object and are used to regress the $N \\times K$ keypoint positions, while each instance query is for scoring the corresponding $K$ -keypoint pose prediction. Besides, the architecture of transformer decoder is simple, which stacks 6 same decoder layers [41]. In each decoder layer, we follow the macro design of previous DETR frameworks by building self-attention, cross-attention implemented with deformable attention, and FFN. We only make simple modifications to self-attention in our Group Pose. Specifically, we replace the standard self-attention with two subsequent group self-attention, enabling decoder layers to perform interactions over queries belonging to the same human instance and with the same type.", + "bbox": [ + 75, + 643, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prediction heads. There are two prediction heads implemented with FFNs in Group Pose for human classification and human keypoints regression. Group Pose predicts for $N$ human poses, each contains a classification score and $K$ keypoints positions for the corresponding $K$ -keypoint pose.", + "bbox": [ + 498, + 273, + 888, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss function. The Hungarian matching algorithm [2] is employed for one-to-one assignment between predicted poses and ground-truth poses. Our loss function comprises solely of classification loss $(\\mathcal{L}_{cls})$ and keypoint regression loss $(\\mathcal{L}_{kpt})$ , without any extra supervisions such as human detection loss in QueryPose [36] and ED-Pose [38] or heatmap loss in PETR [27]. The keypoint regression loss $(\\mathcal{L}_{kpt})$ is a combination of a normal $\\ell_1$ loss and a constrained $\\ell_1$ loss named Object Keypoint Similarity (OKS) [27]. We directly use the cost coefficients and loss weights of ED-Pose [38] in the hungarian matching and the calculation of losses.", + "bbox": [ + 498, + 356, + 888, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. $N \\times K$ keypoint queries and $N$ instance queries", + "text_level": 1, + "bbox": [ + 500, + 551, + 888, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In multi-person pose estimation, frameworks are required to predict $N$ human poses with $K$ keypoint positions in each pose given an image. We directly use $N \\times K$ keypoints queries to predict the poses, with each $K$ predicted keypoint positions to represent a corresponding $K$ -keypoint pose for a human instance. To classify if it is a human instance by scoring the predicted $N$ human poses, we also introduce $N$ instance queries. Thus, in Group Pose, a human instance can be represented with a combination of $K$ keypoint queries and one instance query. As these two types of queries are responsible for different tasks, we construct and initialize them differently.", + "bbox": [ + 496, + 578, + 888, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query construction. We follow the previous end-to-end frameworks [27, 38] to first identify human instances and predict human poses in each position of the output memory from the transformer encoder. Then we select $N$ human instances ( $N = 100$ ) based on the classification scores, resulting $N$ human poses and the corresponding output memory features of the selected $N$ positions (with the shape of $N \\times D$ , where $D$ is the channel dimension). We then construct and initialize the keypoint queries and the instance", + "bbox": [ + 496, + 763, + 888, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15031", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92ed9b5f62fd89906b8ec4e8043e59b1d924df628c3bb2ae823ad810a7170d57.jpg", + "image_caption": [ + "$(N\\times (K + 1))\\times (N\\times (K + 1))$", + "(a) standard self-attention" + ], + "image_footnote": [], + "bbox": [ + 158, + 85, + 336, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7adcb61177411121267fa4a2e6d0d63cca5e8d2983a91ee894b32ed1c50c216f.jpg", + "image_caption": [ + "$(N\\times (K + 1))\\times (N\\times (K + 1))$", + "(b) standard self-attention with mask", + "Figure 3: Conceptual comparison of self-attention implementations. Given $N \\times (K + 1)$ queries, three different implementations of self-attention are considered in the decoder layers. (a) standard self-attention over all queries. (b) standard self-attention with an attention mask, removing the across-instance interactions for queries with different types. (c) our proposed group self-attention. Dashed boxes indicate the masked-out interactions." + ], + "image_footnote": [], + "bbox": [ + 362, + 85, + 540, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/24322e3e8f6c28a6ebdf809954965d38494d3cd194fa52f1b16cc4ba454b5b7e.jpg", + "image_caption": [ + "$(N\\times ((K + 1)\\times (K + 1)) + ((K + 1)\\times (N\\times N))$", + "(c) group self-attention" + ], + "image_footnote": [], + "bbox": [ + 571, + 85, + 813, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "queries based on the above results.", + "bbox": [ + 75, + 310, + 307, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For $N \\times K$ keypoint queries, we construct the content part of $K$ keypoint queries $(K \\times D)$ in each human instance by combining $K$ randomly initialized learnable keypoint embeddings $(K \\times D)$ and the corresponding output memory feature $(1 \\times D)$ . And the position part of $K$ keypoint queries $(K \\times 2)$ in each human instance is initialized with the corresponding predicted $K$ -keypoint pose $(K \\times 2)$ . For $N$ instance queries, we only consider the content part and use a randomly initialized learnable instance embedding $(1 \\times D)$ for each human instance. This is because instance query is for the classification task, which does not require explicit position information. When performing cross-attention in the decoder layers, we simply use the mean of $K$ keypoint positions as the reference point for the instance query in each human instance.", + "bbox": [ + 75, + 327, + 473, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Group self-attention", + "text_level": 1, + "bbox": [ + 76, + 566, + 276, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In DETR frameworks [2, 41, 23], self-attention in the transformer decoder is usually applied to model interactions among queries, collecting information from other queries and facilitating the duplicate removal for instances. While the situation is slightly different in Group Pose, as there two different types of queries, $K$ keypoint queries and one instance query for each human instance. Motivated by the intuition that the interactions among across-instance queries of different types, e.g., the interactions between the instance query and the keypoint queries across human instances, may not be directly helpful for the above purposes. We thus replace the standard self-attention over all the $N \\times (K + 1)$ queries with two subsequent group self-attention: $N$ within-instance self-attention and $K + 1$ same-type across-instance self-attention.", + "bbox": [ + 75, + 593, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$N$ within-instance self-attention. We build interactions among the queries within a human instance, exploiting kinematic relations and gathering information for scoring pose predictions. It calculates $N$ self-attention maps for $N$ human instances in parallel with shapes of $(K + 1)\\times (K + 1)$", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "omitting the dimensions of batch size and attention heads.", + "bbox": [ + 496, + 309, + 880, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(K + 1)$ same-type across-instance self-attention. Similar to the within-instance group self-attention, We add another group self-attention, which collect information from the same-type queries in other instances and help remove duplicate predictions. We build interactions across human instances with the same-type queries, the instance query and each keypoint queries, resulting in $K + 1$ same-type across-instance self-attention. It also can be implemented in parallel with one self-attention module, calculating $K + 1$ self-attention maps with shapes of $N \\times N$ .", + "bbox": [ + 496, + 329, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with the standard self-attention, our two subsequent group self-attention explicitly explores the information about queries belong to the same human instance and with the same type, while remove the across-instance interactions for queries with different types. Empirical results in Section 4.3 show that the removal of this kind of interactions eases the optimization and thus improves the performance for our Group Pose.", + "bbox": [ + 496, + 481, + 892, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 619, + 632, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 500, + 645, + 596, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. Our experiments are conducted on two representative human pose estimation datasets, MS COCO [16] and CrowdPose [14]. MS COCO contains 200K images and 250K person instances with 17 keypoint annotations per instance (we set the number of keypoint queries as $K = 17$ on MS COCO). CrowdPose has 20K images and 80K person instances with 14 keypoint annotations per instance ( $K = 14$ on CrowdPose). CrowdPose is more challenging as it includes many crowd and occlusion scenes. We train Group Pose on COCO train2017 set and evaluate it on COCO val2017 set and test-dev set. On CrowdPose, we train our model on the train set and evaluate it on the test set.", + "bbox": [ + 496, + 669, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation metric. The OKS-based average precision (AP) scores are reported as the main metric for both", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15032", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/6b392ef3f83010271628be205771e16ac8641fdf13053b9657e1cee680dd06e0.jpg", + "table_caption": [ + "Table 1: Comparisons with state-of-the-art methods on MS COCO va12017. We also provide the reference (Ref) for previous frameworks. The 'HM', 'BR', and 'KR' denote heatmap-based losses, human box regression losses, and keypoint regression losses. 'RLE' represents the residual log-likelihood estimation in Poseur [21]. † denotes the flipping test. ‡ removes the prediction uncertainty estimation in Poseur as a fair regression comparison. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodRefBackboneLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_M \\)\\( AP_L \\)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM65.587.271.161.373.4
Mask R-CNN [8]CVPR 17ResNet-101HM66.187.472.061.574.4
\\( PRTR^† \\)[15]CVPR 21ResNet-50KR68.288.275.263.276.2
\\( Poseur^‡ \\)[21]ECCV 22ResNet-50RLE70.0----
Poseur [21]ECCV 22ResNet-50RLE74.289.881.371.180.1
Bottom-Up\\( HrHRNet^† \\)[5]CVPR 20HRNet-w32HM67.186.273.061.576.1
\\( DEKR^† \\)[7]CVPR 21HRNet-w32HM68.086.774.562.177.7
\\( SWAHR^† \\)[20]CVPR 21HRNet-w32HM68.987.874.963.077.4
\\( LOGO-CAP^† \\)[37]CVPR 22HRNet-w32HM69.687.575.964.178.0
One-StageDirectPose [31]-ResNet-50KR63.185.668.857.771.3
\\( CenterNet^† \\)[40]-Hourglass-104KR+HM64.0----
FCPose [22]CVPR 21ResNet-50KR+HM63.085.968.959.170.3
InsPose [28]ACM MM 21ResNet-50KR+HM63.186.268.558.570.1
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR68.887.576.362.777.7
PETR [27]CVPR 22Swin-LHM+KR73.190.780.967.281.7
QueryPose [36]NeurIPS 22ResNet-50BR+RLE68.788.674.463.876.5
QueryPose [36]NeurIPS 22Swin-LBR+RLE73.391.379.568.581.2
ED-Pose [38]ICLR 23ResNet-50BR+KR71.689.678.165.979.8
ED-Pose [38]ICLR 23Swin-LBR+KR74.391.581.668.682.6
OursGroupPose-ResNet-50KR72.089.479.166.879.7
GroupPose-Swin-TKR73.690.480.568.781.2
GroupPose-Swin-LKR74.891.682.169.483.0
", + "bbox": [ + 76, + 155, + 890, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets. For MS COCO, we adopt AP with different thresholds and different object sizes (medium and large), denoted as AP, $\\mathrm{AP}_{50}$ , $\\mathrm{AP}_{75}$ , $\\mathrm{AP}_M$ , and $\\mathrm{AP}_L$ , following the standard evaluation process3. On CrowdPose, to better evaluate the model performance in different crowded scenarios, we adopt AP with different thresholds and different crowding levels, denoted as AP, $\\mathrm{AP}_{50}$ , and $\\mathrm{AP}_{75}$ , as well as $\\mathrm{AP}_E$ , $\\mathrm{AP}_M$ and $\\mathrm{AP}_H$ for images with easy, medium and hard crowding levels.", + "bbox": [ + 75, + 518, + 470, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our training and testing settings follow ED-Pose [38]. During training, we adopt the widely-used data augmentations in DETR frameworks [2, 41, 39, 38], including random flip, random crop, and random resize with the short sides in [480, 800] and the long side less or equal to 1333. We use the AdamW optimizer[11, 19] with the weight decay $1 \\times 10^{-4}$ and train 60 epochs and 80 epochs on MS COCO [16] and CrowdPose [14], respectively. We adopt a total batch size of 16, and set the base learning rate as $1 \\times 10^{-4}$ . The base learning rate for the backbone is $1 \\times 10^{-5}$ following the DETR frameworks. The learning rates are decayed at the 50-th epoch and 70-th by a factor of 0.1 for MS COCO and CrowdPose, respectively. During testing, we resize the input images with their short", + "bbox": [ + 75, + 660, + 470, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sides being 800 and long sides less or equal to 1333.", + "bbox": [ + 500, + 518, + 846, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 550, + 638, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our purpose is to build a simple baseline for end-to-end multi-person pose estimation. Thus, we mainly compare our Group Pose with previous end-to-end frameworks, including PETR [27], QueryPose [36], and ED-Pose [38]. Besides, to show the effectiveness of our method, we also add comparisons with non-end-to-end frameworks, such as top-down [8, 15], bottom-up [5, 7, 20, 37], and one-stage methods [31, 40, 22, 28].", + "bbox": [ + 496, + 575, + 893, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparisons with end-to-end frameworks on COCO. Table 1 and Table 2 present the comparisons on COCO val2017 set and test-dev set. Results show that Group Pose outperforms PETR [27], QueryPose [36], and ED-Pose [38] consistently.", + "bbox": [ + 496, + 702, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On COCO val2017, Group Pose surpasses PETR and QueryPose by over a significant 3.0 AP with ResNet-50 [9] and the gaps remain $1.5 + \\mathrm{AP}$ with a strong backbone, SwinLarge [18]. When comparing with the recently proposed ED-Pose [38], which transfers pose estimation to a keypoint box detection problem and combines human box detection task, Group Pose can also exceed it with non-negligible margins. Moreover, unlike these methods that use complex", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3https://cocodataset.org/#keypoints-eval", + "bbox": [ + 94, + 886, + 308, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15033", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fa909cecbdf25594fa60d16f4da10cbb595496d5cccb5456c535807c67757aaf.jpg", + "table_caption": [ + "Table 2: Comparisons with state-of-the-art methods on MS COCO test-dev2017 dataset. Notations are consistent with Table 1." + ], + "table_footnote": [], + "table_body": "
MethodRefBackboneLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_M \\)\\( AP_L \\)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM63.987.769.959.771.5
Mask R-CNN [8]CVPR 17ResNet-101HM64.388.270.660.171.9
\\( PRTR^† \\)[15]CVPR 21ResNet-101KR68.889.976.964.775.8
\\( PRTR^† \\)[15]CVPR 21HRNet-w32KR71.790.679.667.678.4
Bottom-Up\\( HrHRNet^† \\)[5]CVPR 20HRNet-w32HM66.487.572.861.274.2
\\( DEKR^† \\)[7]CVPR 21HRNet-w32HM67.387.974.161.576.1
\\( SWAHR^† \\)[20]CVPR 21HRNet-w32HM67.988.974.562.475.5
\\( LOGO-CAP^† \\)[37]CVPR 22HRNet-w32HM68.288.774.962.876.0
One-StageDirectPose [31]-ResNet-50KR62.286.468.256.769.8
\\( CenterNet^† \\)[40]-Hourglass-104KR+HM63.086.869.658.970.4
FCPose [22]CVPR 21ResNet-50KR+HM64.387.371.061.670.5
InsPose [28]ACM MM 21ResNet-50KR+HM65.488.971.760.272.7
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR67.689.875.361.676.0
PETR [27]CVPR 22Swin-LHM+KR70.591.578.765.278.0
QueryPose [36]NeurIPS 22Swin-LBR+RLE72.292.078.867.379.4
ED-Pose [38]ICLR 23ResNet-50BR+KR69.890.277.264.377.4
ED-Pose [38]ICLR 23Swin-LBR+KR72.792.380.967.680.0
OursGroupPose-ResNet-50KR70.290.577.864.778.0
GroupPose-Swin-TKR72.191.479.966.779.5
GroupPose-Swin-LKR72.892.581.067.780.3
", + "bbox": [ + 83, + 114, + 883, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "decoders and add extra supervisions, e.g., extra heatmap or box supervisions, our Group Pose only use a simple decoder and are trained only with keypoint regression targets. The above evidences indicate that complex design for end-to-end multi-person pose estimation may not be necessary.", + "bbox": [ + 75, + 448, + 468, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On COCO test-dev, our Group Pose achieves 70.2 AP, 72.1 AP, and 72.8 AP with ResNet-50 [9], Swin-Tiny [18], and Swin-Large [18] as the backbone. Compared with other end-to-end frameworks, similar trends are observed with the ones on COCO val2017.", + "bbox": [ + 75, + 526, + 468, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with end-to-end frameworks on CrowdPose. To further demonstrate the effectiveness of Group Pose, we provide comparisons with previous end-to-end frameworks on the challenging CrowdPose dataset [14]. Table 3 reports the results of PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose with the same backbone Swin-Large [18]. Overall, Group Pose achieve 74.1 AP, performing the best over all methods.", + "bbox": [ + 75, + 609, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, it is interesting when we compare the AP scores with easy, medium, and hard crowding levels of different models. PETR [27] performs worse with the easy crowding level while giving the second best results with the hard level. ED-Pose [38] performs worst with the hard crowding level. We conjecture that this phenomenon is caused by the differences in their decoders, e.g., human instances crowded with each other have similar boxes, which challenges the human keypoint decoder. Our Group Pose, which uses $K$ keypoint queries and one instance query for each human instance and considers different interactions", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/93da4c097e7113641fa6b5bb641c4e74138829a42ba2390c7979fafdc73f00b8.jpg", + "table_caption": [ + "Table 3: Comparisons with state-of-the-art methods on Crowd-Posetest dataset. Swin-L is adopted as the backbone. Other notations are consistent with Tabel 1." + ], + "table_footnote": [], + "table_body": "
MethodLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_E \\)\\( AP_M \\)\\( AP_H \\)
PETR [27]HM+KR71.690.478.377.372.065.8
QueryPose [36]BR+RLE72.791.778.179.573.465.4
ED-Pose [38]BR+KR73.190.579.880.573.863.8
GroupPoseKR74.191.380.480.874.766.4
", + "bbox": [ + 503, + 500, + 890, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "among queries, can achieve reasonable good results with easy, medium, and hard crowding levels.", + "bbox": [ + 498, + 614, + 890, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with various non-end-to-end frameworks on COCO. We also compare Group Pose with representative non-end-to-end frameworks in Table 1 and Table 2. Group Pose can easily outperform previous bottom-up methods [5, 7, 20, 37] and one-stage methods [31, 40, 22, 28]. For example, Group Pose surpasses the recent proposed LOGO-CAP [37] with flipping test by over 2.0 AP on both COCO val2017 and test-dev, even with a smaller backbone (ResNet-50 [9]) than it (HRNetw32 [29]). Group Pose is more concise and precise than those bottom-up and one-stage methods. Surprisingly, Group Pose can also beat previous top-down methods like PRTR [15] and Poseur [21]. Given the simple design and the end-to-end property, our Group Pose can serve as a simple baseline for pursuing higher performance on multi-person pose estimation.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15034", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(a) Query designs for human instances. Both instance (inst) and keypoint (kpt) queries are essential in Group Pose, especially the keypoint ones.", + "bbox": [ + 76, + 109, + 326, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) Benefits of the instance query. On the model only with kpt queries, the inst query is first added but not for classification (cls). Then we use the inst query for the cls task.", + "bbox": [ + 339, + 109, + 629, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(c) Number of instances. When the number is smaller than 100, increasing instance numbers provide gains. 100 is set by default for the number of instances.", + "bbox": [ + 642, + 109, + 890, + 157 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/22ad8e861f7f4f22f994cb05842db6c107175e4a2def8237d538a3ea6ca65047.jpg", + "table_caption": [ + "Table 4: Ablation experiments for Group Pose. Evaluated on MS COCO val2017. Default settings are marked in gray ." + ], + "table_footnote": [], + "table_body": "
query typesAPAPMAPL
inst & kpt72.066.879.7
only kpt71.266.079.1
only inst64.561.169.9
", + "bbox": [ + 98, + 164, + 305, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b854030b12b889a951c8b2e72f46f70abecd5431b9475f5075d69324875031f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
query typescls taskAPAPMAPL
only kptavg kpt71.266.079.1
inst & kptavg kpt71.766.879.3
inst & kptinst72.066.879.7
", + "bbox": [ + 348, + 165, + 616, + 217 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/383d8be24d22b0b01c88de55bac12ec0136139ff5826f27c74f97971ecbe0a6e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#instanceAPAPMAPL
5071.466.479.1
10072.066.879.7
20072.066.979.7
", + "bbox": [ + 669, + 164, + 864, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(d) Self-attention implementations. In the self-attention module, removing the across-instance interactions for queries with different types with attention (attn) mask or group self-attention eases the optimization and improves the performance.", + "bbox": [ + 76, + 223, + 470, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bcebdabe9bc0011801eeec04fa440856c316cbb248a54ce3cd43acab0c7abec7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
self-attention implementationsw/ attn maskAPAPMAPL
standard self-attention×69.464.077.5
standard self-attention✓70.765.578.4
group self-attention×72.066.879.7
", + "bbox": [ + 88, + 277, + 460, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 342, + 228, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We run a number of ablation experiments to verify the effectiveness of key elements in our Group Pose. We adopt ResNet-50 [9] as the backbone. Unless specified, we report the results on COCO val2017 with 60 epochs training.", + "bbox": [ + 76, + 369, + 468, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation: query designs for human instance. Table 4a gives the results of representing a human instance with different query designs. Both instance query and keypoint queries are essential in Group Pose. As the multi-person pose estimation is to predict human poses given an image, it is reasonable that keypoint queries themselves can achieve good results, while removing keypoint queries gives a significant performance drop (from 72.0 AP to 64.5 AP).", + "bbox": [ + 75, + 436, + 468, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the instance query, its benefits come from two aspects: (i) gather information within human instance and help model training and (ii) decouple the classification task and keypoint positions regression task. Table 4b shows the improvements brought by these two benefits.", + "bbox": [ + 75, + 561, + 468, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation: self-attention implementations. Table 4d provides the comparisons between different self-attention in the decoder layers, including (i) standard self-attention over all the $N \\times (K + 1)$ queries (Figure 3 (a)), (ii) standard self-attention with an attention mask (Figure 3 (b)), which masks out the across-instance interactions for queries with different types, and (iii) our group self-attention (Figure 3 (c)). Results validate that the across-instance interactions for queries with different types are not directly helpful for the multi-person pose estimation task. Removing this type of interactions eases the model optimization, helps model converge faster, thus improves the performances. The comparisons on convergence curves are given in Figure 4. Besides, we find that it is better to separately perform different types of interactions, the within-instance and same-type across-instance interactions, with different parameters. Table 4d shows that our sequential implementation group self", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(e) Group self-attentions. Both types of self-attentions are important in group self-attentions. We also conduct an experiment by removing both within-instance and across-instance self-attentions. The performance further drops to 65.3 AP.", + "bbox": [ + 491, + 223, + 887, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45b5b720399ff2b8715f693c0da6754afd0ec069209448245559f0e63ba14ec7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
group self-attentionssAPAPMAPL
within-instance & across-instance self-attentionss72.066.879.7
within-instance self-attentionss66.363.870.8
across-instance self-attentionss67.463.074.4
", + "bbox": [ + 504, + 277, + 879, + 329 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c13e0a49672a4448492403bfa91f66ad7518684121ae72b6a932ac09f40b3829.jpg", + "image_caption": [ + "Figure 4: Faster and better convergence. The $x$ -axis corresponds to #epoch, and the $y$ -axis corresponds to AP score. One can see that group self-attention converges faster and better than other self-attention implementations." + ], + "image_footnote": [], + "bbox": [ + 506, + 342, + 883, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": " attentions bring a 1.3 AP gain over the parallel implementation of standard self-attention with an attention mask.", + "bbox": [ + 498, + 544, + 890, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation: group self-attentions. We ablate the effects of $N$ within-instance self-attentions over $K$ keypoint queries and one instance query and $(K + 1)$ across-instance self-attentions over $N$ queries of the same type in Table 4e. Large performance gaps (over $-5.0\\mathrm{AP}$ ) are observed when we remove either of these two group self-attentions. Both the interactions are important for multi-person pose estimation. The within-instance self-attentions explore kinematic relations and gather information for scoring pose predictions, which help the model predict precise keypoint positions for human poses. The same-type across-instance self-attentions collect information from the same-type of keypoint queries or instance queries, removing duplicate predictions for poses among human instances. As shown in Figure 5, model without the same-type across-instance self-attentions produces more duplicated pose predictions for the same human instance than Group Pose.", + "bbox": [ + 496, + 578, + 892, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation: number of instances. Table 4c gives how predefined number of human instances affect the results. We find 100 human instances are enough and can achieve comparable results with 200 human instances. We set the num", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15035", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6cfa798a371f8ac372d2e5ac660f503b0db2e62b16a338e37339baa83c611c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 89, + 289, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae478084818c4857514ecf386cff4996c3c093570d8af5df33aebf1935235c22.jpg", + "image_caption": [ + "Figure 5: Comparison of removing duplicated predictions. We visualize the predicted poses of Group Pose (second row), and Group Pose without same-type across-instance self-attention (first row) according to the number of objects. Group Pose produces less duplicated pose predictions. The yellow dashed ellipse indicates the duplicated predicted human instance, leading to a mismatch of other instances. Best view in zoom in." + ], + "image_footnote": [], + "bbox": [ + 89, + 178, + 287, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f5613600a75b8f370e97a191151358c33d1b25694901b57ed62d14014eacfe2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 90, + 457, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/971f93e3d4e76b0bc9976722630cc3991ce7c80fc05d8fb004ff6f3c5b37ef49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 178, + 457, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a767049aebf04137063dcf2ee5cc92a663c52149d4380d08c2fdc6ea14286d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 90, + 627, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afb2ebd454ecf6bc8840ab8f47594b35f4d5d7338ecf7b89b12a07459fc6608d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 178, + 627, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8071d082320ae6a4252f76095b86c5a9f5f434f51e567b5714845b300c923a33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 90, + 790, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4c51b53b30675e8f416da4963151e6ee570c60cd3cb2f3c218b6c36c1e5e0e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 178, + 790, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/706640c2966ce82b31341ba3e38be81d65d73ec2712e388a736441a29713af11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 90, + 879, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f19f5546c2f9314533c5db1e68c48d65b6f2a8b6dbe4361df0827695486b2341.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 178, + 879, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3255f3aec7190a2d9783622e7f3fd5179cce8095bf92dfed6ba8307c5fade54c.jpg", + "table_caption": [ + "Table 5: Analysis on model convergence. Group Pose without human detection already can outperform ED-Pose [38]. Besides, Group Pose can also benefit from better human instance initialization with human detection. We report AP on COCO val2017 dataset. 'Det Dec' = human detection decoder. The results of ED-Pose are from the original paper [38]." + ], + "table_footnote": [], + "table_body": "
Methodw/ Det Dec12e24e36e48e60e
ED-Pose [38]✓60.567.569.770.871.6
GroupPose×61.067.670.171.472.0
GroupPose✓61.468.170.371.672.2
", + "bbox": [ + 78, + 443, + 468, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ber of human instances as 100 by default. Thus, Group Pose has $100 \\times (17 + 1)$ queries on COCO [16] and contains $100 \\times (14 + 1)$ queries on CrowdPose [14].", + "bbox": [ + 75, + 531, + 468, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. More Analysis", + "text_level": 1, + "bbox": [ + 76, + 585, + 225, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide more analysis about our Group Pose in this section, including the analysis on model convergence and the analysis on model inference speed, detailed next.", + "bbox": [ + 75, + 609, + 468, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on model convergence. A common wisdom about training models is to provide a good initialization and then refine the model based on it. ED-Pose [38] splits the multi-person pose estimation task into two sub-processes, which first detect human instances with a human detection decoder and then use a human-to-keypoint detection decoder for predicting human poses. The human detection decoder gives good initialization for human instances, which can help ED-Pose learn and converge faster. Although our Group Pose adopts a different design for transformer decoder, it can also be built upon a human detection decoder and enjoys the benefits brought by better initialization of human instances. Based on this observation, we build a variant of our Group Pose, whose transformer decoder consists of 2 human detection decoder layers and 4 simple transformer decoder layers for multi-person pose estimation, following", + "bbox": [ + 75, + 659, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ED-Pose [38].", + "bbox": [ + 500, + 349, + 598, + 364 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/faa4b26ac65609b287f3b0f94e194ccabc82c9bae52421c1d7c1be7755751f07.jpg", + "table_caption": [ + "Table 6: Analysis on model inference speed. Frames per second (FPS) and inference time (Time) are measured with ResNet-50 and different image resolutions on one NVIDIA A100 GPU." + ], + "table_footnote": [], + "table_body": "
MethodInput ResolutionFPS ↑Time [ms] ↓
PETR [27]480 × 80020.050
800 × 133312.183
QueryPose [36]480 × 80019.056
800 × 133313.475
ED-Pose [38]480 × 80042.424
800 × 133324.740
GroupPose480 × 80068.615
800 × 133331.332
", + "bbox": [ + 500, + 445, + 903, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 shows the AP scores of three models, ED-Pose, Group Pose, and Group Pose with human detection decoder, under different training schedules. Group Pose already shows superior results than ED-Pose with 12 epochs, 24 epochs, 36 epochs, 48 epochs, and 60 epochs training, even without the help of human detection decoder. Moreover, the comparison between Group Pose and Group Pose with human detection decoder verifies that good initialization for human instances leads to faster convergence and better results.", + "bbox": [ + 496, + 602, + 893, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on model inference speed. Table 6 provides the comparisons of inference time and FPS among end-to-end frameworks, including PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose. Simple designs usually show efficiency. Results show that our simple transformer decoder can be faster than complex decoders in previous end-to-end frameworks. Even with an image in a large size $800 \\times 1333$ , our Group Pose can also achieve real-time speed (above 30 FPS) on a single A100 GPU.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15036", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 89, + 194, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we present a simple baseline, Group Pose, for end-to-end multi-person pose estimation. With simple designs for queries and decoder self-attention, the approach outperforms previous end-to-end frameworks while being faster. Besides, the transformer decoder in Group Pose is also flexible, which can be built solely or upon human detection decoders. We hope Group Pose can provide insights for exploring concise and effective end-to-end multi-person pose estimation frameworks.", + "bbox": [ + 75, + 116, + 468, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.U1936212, No.62120106009, No.62261160653).", + "bbox": [ + 76, + 253, + 470, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 330, + 174, + 345 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 1, 2", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 2, 3, 4, 5", + "[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2", + "[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018. 1, 2", + "[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5386-5395, 2020. 1, 2, 5, 6", + "[6] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2334-2343, 2017. 1, 2", + "[7] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 5, 6", + "[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2961-2969, 2017. 2, 5, 6" + ], + "bbox": [ + 86, + 357, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 3, 5, 6, 7", + "[10] Jan Hosang, Rodrigo Benenson, and Bernt Schiele. Learning non-maximum suppression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4507-4515, 2017. 1, 2", + "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[12] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11977-11986, 2019. 1, 2", + "[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2", + "[14] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10863-10872, 2019. 2, 4, 5, 6, 8", + "[15] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 5, 6", + "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 2, 4, 5, 8", + "[17] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2", + "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5, 6", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[20] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 5, 6", + "[21] Weian Mao, Yongtao Ge, Chunhua Shen, Zhi Tian, Xinlong Wang, Zhibin Wang, and Anton van den Hengel. Poseur: Direct human pose regression with transformers. In Proceedings of the European Conference on Computer Vision, pages 72-88, 2022. 5, 6", + "[22] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15037", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 1, 2, 5, 6", + "[23] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 4", + "[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in Neural Information Processing Systems, 30, 2017. 1, 2", + "[25] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision, pages 483-499. Springer, 2016. 1, 2", + "[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6951-6960, 2019. 1, 2", + "[27] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022. 1, 2, 3, 5, 6, 8", + "[28] Dahu Shi, Xing Wei, Xiaodong Yu, Wenming Tan, Ye Ren, and Shiliang Pu. Inspose: instance-aware networks for single-stage multi-person pose estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3079-3087, 2021. 5, 6", + "[29] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019. 1, 2, 6", + "[30] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2", + "[31] Zhi Tian, Hao Chen, and Chunhua Shen. Directpose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 1, 2, 5, 6", + "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3", + "[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 2", + "[34] Fangyun Wei, Xiao Sun, Hongyang Li, Jingdong Wang, and Stephen Lin. Point-set anchors for object detection, instance segmentation and pose estimation. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision, pages 527-544. Springer, 2020. 1, 2", + "[35] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision, pages 466-481, 2018. 1, 2", + "[36] Yabo Xiao, Kai Su, Xiaojuan Wang, Dongdong Yu, Lei Jin, Mingshu He, and Zehuan Yuan. Querypose: Sparse multiperson pose regression via spatial-aware part-level query. In Advances in Neural Information Processing Systems, pages 1-14, 2022. 2, 3, 5, 6, 8", + "[37] Nan Xue, Tianfu Wu, Gui-Song Xia, and Liangpei Zhang. Learning local-global contextual adaptation for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13065-13074, 2022. 5, 6", + "[38] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In International Conference on Learning Representations, pages 1-17, 2023. 1, 2, 3, 5, 6, 8", + "[39] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, pages 1-18, 2022. 1, 2, 5", + "[40] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 5, 6", + "[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, pages 1-16, 2021. 1, 2, 3, 4, 5" + ], + "bbox": [ + 501, + 92, + 890, + 556 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15038", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_model.json b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce3265b984624c822f2140231e58154b8e5a5382 --- /dev/null +++ b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "Group Pose: A Simple Baseline for End-to-End Multi-person Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.18, + 0.847, + 0.218 + ], + "angle": 0, + "content": "Huan Liu\\(^{1,3*}\\) Qiang Chen\\(^{2*}\\) Zichang Tan\\(^{2}\\) Jiang-Jiang Liu\\(^{2}\\) Jian Wang\\(^{2}\\) Xiangbo Su\\(^{2}\\) Xiaolong Li\\(^{1,3}\\) Kun Yao\\(^{2}\\) Junyu Han\\(^{2}\\) Errui Ding\\(^{2}\\) Yao Zhao\\(^{1,3\\dagger}\\) Jingdong Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.22, + 0.882, + 0.259 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Information Science, Beijing Jiaotong University \\(^{2}\\)Baidu VIS \\(^{3}\\)Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.291, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.428 + ], + "angle": 0, + "content": "In this paper, we study the problem of end-to-end multiperson pose estimation. State-of-the-art solutions adopt the DETR-like framework, and mainly develop the complex decoder, e.g., regarding pose estimation as keypoint box detection and combining with human detection in ED-Pose [38], hierarchically predicting with pose decoder and joint (keypoint) decoder in PETR [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.504 + ], + "angle": 0, + "content": "We present a simple yet effective transformer approach, named Group Pose. We simply regard \\( K \\)-keypoint pose estimation as predicting a set of \\( N \\times K \\) keypoint positions, each from a keypoint query, as well as representing each pose with an instance query for scoring \\( N \\) pose predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.76 + ], + "angle": 0, + "content": "Motivated by the intuition that the interaction, among across-instance queries of different types, is not directly helpful, we make a simple modification to decoder self-attention. We replace single self-attention over all the \\( N \\times (K + 1) \\) queries with two subsequent group self-attention: (i) \\( N \\) within-instance self-attention, with each over \\( K \\) keypoint queries and one instance query, and (ii) \\( (K + 1) \\) same-type across-instance self-attention, each over \\( N \\) queries of the same type. The resulting decoder removes the interaction among across-instance type-different queries, easing the optimization and thus improving the performance. Experimental results on MS COCO and CrowdPose show that our approach without human box supervision is superior to previous methods with complex decoders, and even is slightly better than ED-Pose that uses human box supervision. Paddle1 and PyTorch2 codes are available." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.788, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.813, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Multi-person pose estimation aims to detect the corresponding human keypoints for all human instances in an" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.288, + 0.877, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.32, + 0.725, + 0.331 + ], + "angle": 0, + "content": "(a) PETR" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.334, + 0.875, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.37, + 0.732, + 0.381 + ], + "angle": 0, + "content": "(b) ED-Pose" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.385, + 0.874, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.422, + 0.737, + 0.433 + ], + "angle": 0, + "content": "(c) Group Pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.445, + 0.894, + 0.53 + ], + "angle": 0, + "content": "Figure 1: Comparison of transformer decoders. Here we mainly illustrate the overview of the decoder part of PETR [27], ED-Pose [38], and our Group Pose. The three end-to-end frameworks differ in query design and decoder architecture. Group Pose only uses a simple transformer decoder rather than developing complex decoders. 'inst' and 'kpt' represent for instance and keypoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.747 + ], + "angle": 0, + "content": "image. Previous frameworks include top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5] that divide the task into two sequential sub-tasks: human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping [24]. Another line in previous frameworks is one-stage methods [22, 26, 31, 34], which directly predict instance-aware keypoints. These frameworks rely on non-differentiable hand-crafted post-processes [10, 24], which complicate the pipelines and challenge the optimizations. Inspired by the success of DETR [2] in object detection, building an end-to-end framework for multi-person pose estimation has seen significant interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Recent approaches follow the DETR framework [2, 41, 39], with the transformer encoder-decoder architecture for multi-person pose estimation, as shown in Figure 1. PETR [27] hierarchically predicts the keypoint positions and uses two subsequent decoders, pose decoder and joint decoder, with two different queries, pose query (a person has one pose query) for pose decoder, and keypoint queries for joint decoder. ED-Pose [38] transfers pose estimation to a keypoint box detection problem, and learns a content query and a box query for each keypoint position prediction" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.85, + 0.451, + 0.863 + ], + "angle": 0, + "content": "*Equal Contribution. Work done when H. Liu is an intern at Baidu." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.863, + 0.366, + 0.876 + ], + "angle": 0, + "content": "†Corresponding author. Email: yzhao@bjtu.edu.cn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.876, + 0.449, + 0.888 + ], + "angle": 0, + "content": "\\(^{1}\\)Code of Paddle: https://github.com/Michel-liu/GroupPose-Paddle" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.419, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)Code of PyTorch: https://github.com/Michel-liu/GroupPose" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.85, + 0.451, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15029" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.373, + 0.108 + ], + "angle": 0, + "content": "with using the box size to process the query." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.118, + 0.47, + 0.239 + ], + "angle": 0, + "content": "In this paper, we present a simple yet effective transformer approach, named Group Pose, for end-to-end multiperson pose estimation. Instead of using a single query to predict and score one pose, similar to ED-Pose, we use the \\( N \\times K \\) keypoint queries to regress the \\( N \\times K \\) positions, by regarding each keypoint as an object, as well as \\( N \\) instance queries, each representing a \\( K \\)-keypoint pose for scoring the \\( K \\)-keypoint pose prediction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.25, + 0.471, + 0.386 + ], + "angle": 0, + "content": "We make a simple modification for the decoder architecture. We replace standard self-attention in the decoder with two subsequent group self-attention: \\(N\\) parallel self-attention with each over \\(K\\) keypoint queries and the corresponding instance query for exploiting kinematic relation and gathering information for scoring pose predictions, and \\((K + 1)\\) parallel self-attention with each over \\(N\\) queries of the same type for collecting duplicate prediction information like self-attention of the original DETR." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.471, + 0.547 + ], + "angle": 0, + "content": "The two self-attention captures two kinds of interactions: (i) \\(N\\) within-instance interactions over \\(K\\) keypoint queries and one instance query, (ii) \\((K + 1)\\) across-instance interactions over \\(N\\) queries of the same type (e.g., nose keypoint query or instance query). The extra interactions in standard self-attention are (iii) across-instance interactions for queries with different types, which is not directly useful. Empirical results show that the removal of the third kind of interactions eases the optimization and thus improves the performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.47, + 0.755 + ], + "angle": 0, + "content": "The design about self-attention is different from the closely-related approach ED-Pose [38]. On the one hand, in addition to removing the third interactions, ED-Pose only models across-instance interactions for the instance queries. In contrast, our approach also models across-instance interactions for the \\(K\\) keypoint types. On the other hand, our approach separates the two interactions using two subsequent group self-attention, explicitly exploring the information about queries belonging to the same human instance, and with the same type. ED-Pose couples the two interactions using a single masked self-attention with the third interactions masked. It is empirically demonstrated that the two differences benefit the pose estimation performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Experimental results show that our simple approach Group Pose without human box supervision surprisingly outperforms the recent end-to-end methods with human box supervisions on MS COCO [16] and CrowdPose [14]. Notably, Group Pose achieves 72.0 AP with ResNet-50 [9] and 74.8 AP with Swin-Large [18] on MS COCO val2017. We hope our simple transformer decoder in Group Pose will motivate people to simplify the design in end-to-end multiperson pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.121, + 0.894, + 0.198 + ], + "angle": 0, + "content": "Multi-person pose estimation is a challenging task that aims to detect the corresponding human keypoints for all human instances in an image. Previous methods usually adopt complex frameworks to address it, which are divided into non-end-to-end and end-to-end methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.206, + 0.893, + 0.465 + ], + "angle": 0, + "content": "Non-end-to-end methods. There are typically two types: two-stage and one-stage. Two-stage frameworks, including top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5], split the multi-person pose estimation task into two sequential sub-tasks, human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping. In top-down methods, an object detector is first employed to detect the boxes of human instances, which is then cropped for single-person pose estimation in each box. Bottom-up methods first predict all human keypoints in a human-agnostic way then group them into instances. While one-stage frameworks [40, 22, 26, 31, 34] directly predict instance-aware keypoints. These methods require hand-crafted pose-processes, such as NMS [10] or grouping [24], which complicate the pipelines and challenge the optimizations. In this paper, we focus on concise ways, end-to-end frameworks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.893, + 0.821 + ], + "angle": 0, + "content": "End-to-end methods. Current end-to-end multi-person pose estimation frameworks [27, 36, 38] are built by following the designs of DETR [2] and its variants [41, 23, 3, 17, 13, 39]. They adopt a paradigm of splitting the multiperson pose estimation task into two sub-processes. For example, PETR [27] views the task as a hierarchical set prediction problem. It first determines human instances by predicting a set of poses with a pose decoder and then refines the keypoints in each pose with a joint (keypoint) decoder. There are also two different types of queries, pose query (a person has one pose query) for pose decoder and keypoint queries for joint decoder. QueryPose [36] and ED-Pose [38] follow this end-to-end paradigm but further incorporate an extra human detection task. QueryPose [36] follows Sparse R-CNN [30] to build two parallel RoIAlign-based [8] decoders to perform human detection and pose estimation, respectively. ED-Pose [38] transfers pose estimation task to a keypoint box detection problem. It first employs a human detection decoder [39] to determine the human instances with box queries. Then it builds a human-to-keypoint detection decoder with a content query and a box query for each keypoint position, collecting contextual information near keypoint positions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Although these end-to-end methods show promising results in multi-person pose estimation, they rely on complex decoders. Our Group Pose, on the other hand, adopts a simple transformer decoder, improving the performance and simplifying the process." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15030" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.852, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.89, + 0.268 + ], + "angle": 0, + "content": "Figure 2: Our Group Pose architecture. The backbone takes an image as input and outputs image features, which are refined by the transformer encoder. To directly predict \\(N\\) human poses with \\(K\\) keypoint positions in each pose, we adopt \\(N \\times (K + 1)\\) queries, containing \\(N\\) instance queries for scoring poses and \\(N \\times K\\) keypoint queries for regressing positions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.274, + 0.2, + 0.291 + ], + "angle": 0, + "content": "3. Group Pose" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.468, + 0.405 + ], + "angle": 0, + "content": "Group Pose is a simple yet effective end-to-end multi-person pose estimation framework. We follow previous end-to-end frameworks [27, 38] to view the multi-person pose estimation task as a set prediction problem, but directly adopt a simple transformer decoder [41] instead of complex decoders, simplifying the process. Next, we introduce the key elements of Group Pose." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.416, + 0.187, + 0.431 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.468, + 0.531 + ], + "angle": 0, + "content": "The overall structure of Group Pose is depicted in Figure 2. Group Pose consists of a backbone [9, 18], a transformer encoder [32], a transformer decoder, and task-specific prediction heads. This framework enables Group Pose to simultaneously regresses \\( K \\) keypoints (e.g., \\( K = 17 \\) on MS COCO) for \\( N \\) human instances given an image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.468, + 0.642 + ], + "angle": 0, + "content": "Backbone and transformer encoder. We directly follow DETR frameworks [41] to build the backbone and the transformer encoder (with 6 deformable transformer layers [41]) for Group Pose. It takes an image as input and outputs the extracted multi-level features, which serve as inputs for the following transformer decoder. We use 4 feature levels in Group Pose, with downsampling rates of \\(\\{8, 16, 32, 64\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Transformer decoder. For transformer decoder, we adopt a combination of \\( N \\times K \\) keypoint queries and \\( N \\) instance queries as input instead of using a single query to predict and score one pose. The keypoint queries regard each keypoint as an object and are used to regress the \\( N \\times K \\) keypoint positions, while each instance query is for scoring the corresponding \\( K \\)-keypoint pose prediction. Besides, the architecture of transformer decoder is simple, which stacks 6 same decoder layers [41]. In each decoder layer, we follow the macro design of previous DETR frameworks by building self-attention, cross-attention implemented with deformable attention, and FFN. We only make simple modifications to self-attention in our Group Pose. Specifically, we replace the standard self-attention with two subsequent group self-attention, enabling decoder layers to perform interactions over queries belonging to the same human instance and with the same type." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.275, + 0.89, + 0.352 + ], + "angle": 0, + "content": "Prediction heads. There are two prediction heads implemented with FFNs in Group Pose for human classification and human keypoints regression. Group Pose predicts for \\(N\\) human poses, each contains a classification score and \\(K\\) keypoints positions for the corresponding \\(K\\)-keypoint pose." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.357, + 0.89, + 0.537 + ], + "angle": 0, + "content": "Loss function. The Hungarian matching algorithm [2] is employed for one-to-one assignment between predicted poses and ground-truth poses. Our loss function comprises solely of classification loss \\((\\mathcal{L}_{cls})\\) and keypoint regression loss \\((\\mathcal{L}_{kpt})\\), without any extra supervisions such as human detection loss in QueryPose [36] and ED-Pose [38] or heatmap loss in PETR [27]. The keypoint regression loss \\((\\mathcal{L}_{kpt})\\) is a combination of a normal \\(\\ell_1\\) loss and a constrained \\(\\ell_1\\) loss named Object Keypoint Similarity (OKS) [27]. We directly use the cost coefficients and loss weights of ED-Pose [38] in the hungarian matching and the calculation of losses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.89, + 0.57 + ], + "angle": 0, + "content": "3.2. \\(N \\times K\\) keypoint queries and \\(N\\) instance queries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.89, + 0.76 + ], + "angle": 0, + "content": "In multi-person pose estimation, frameworks are required to predict \\(N\\) human poses with \\(K\\) keypoint positions in each pose given an image. We directly use \\(N \\times K\\) keypoints queries to predict the poses, with each \\(K\\) predicted keypoint positions to represent a corresponding \\(K\\)-keypoint pose for a human instance. To classify if it is a human instance by scoring the predicted \\(N\\) human poses, we also introduce \\(N\\) instance queries. Thus, in Group Pose, a human instance can be represented with a combination of \\(K\\) keypoint queries and one instance query. As these two types of queries are responsible for different tasks, we construct and initialize them differently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.89, + 0.901 + ], + "angle": 0, + "content": "Query construction. We follow the previous end-to-end frameworks [27, 38] to first identify human instances and predict human poses in each position of the output memory from the transformer encoder. Then we select \\( N \\) human instances (\\( N = 100 \\)) based on the classification scores, resulting \\( N \\) human poses and the corresponding output memory features of the selected \\( N \\) positions (with the shape of \\( N \\times D \\), where \\( D \\) is the channel dimension). We then construct and initialize the keypoint queries and the instance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15031" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.087, + 0.337, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.211, + 0.331, + 0.223 + ], + "angle": 0, + "content": "\\((N\\times (K + 1))\\times (N\\times (K + 1))\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.224, + 0.319, + 0.234 + ], + "angle": 0, + "content": "(a) standard self-attention" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.087, + 0.541, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.212, + 0.536, + 0.223 + ], + "angle": 0, + "content": "\\((N\\times (K + 1))\\times (N\\times (K + 1))\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.224, + 0.551, + 0.234 + ], + "angle": 0, + "content": "(b) standard self-attention with mask" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.087, + 0.815, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.212, + 0.809, + 0.223 + ], + "angle": 0, + "content": "\\((N\\times ((K + 1)\\times (K + 1)) + ((K + 1)\\times (N\\times N))\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.224, + 0.747, + 0.234 + ], + "angle": 0, + "content": "(c) group self-attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 3: Conceptual comparison of self-attention implementations. Given \\( N \\times (K + 1) \\) queries, three different implementations of self-attention are considered in the decoder layers. (a) standard self-attention over all queries. (b) standard self-attention with an attention mask, removing the across-instance interactions for queries with different types. (c) our proposed group self-attention. Dashed boxes indicate the masked-out interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.308, + 0.325 + ], + "angle": 0, + "content": "queries based on the above results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.474, + 0.552 + ], + "angle": 0, + "content": "For \\(N \\times K\\) keypoint queries, we construct the content part of \\(K\\) keypoint queries \\((K \\times D)\\) in each human instance by combining \\(K\\) randomly initialized learnable keypoint embeddings \\((K \\times D)\\) and the corresponding output memory feature \\((1 \\times D)\\). And the position part of \\(K\\) keypoint queries \\((K \\times 2)\\) in each human instance is initialized with the corresponding predicted \\(K\\)-keypoint pose \\((K \\times 2)\\). For \\(N\\) instance queries, we only consider the content part and use a randomly initialized learnable instance embedding \\((1 \\times D)\\) for each human instance. This is because instance query is for the classification task, which does not require explicit position information. When performing cross-attention in the decoder layers, we simply use the mean of \\(K\\) keypoint positions as the reference point for the instance query in each human instance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.277, + 0.584 + ], + "angle": 0, + "content": "3.3. Group self-attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.82 + ], + "angle": 0, + "content": "In DETR frameworks [2, 41, 23], self-attention in the transformer decoder is usually applied to model interactions among queries, collecting information from other queries and facilitating the duplicate removal for instances. While the situation is slightly different in Group Pose, as there two different types of queries, \\( K \\) keypoint queries and one instance query for each human instance. Motivated by the intuition that the interactions among across-instance queries of different types, e.g., the interactions between the instance query and the keypoint queries across human instances, may not be directly helpful for the above purposes. We thus replace the standard self-attention over all the \\( N \\times (K + 1) \\) queries with two subsequent group self-attention: \\( N \\) within-instance self-attention and \\( K + 1 \\) same-type across-instance self-attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\(N\\) within-instance self-attention. We build interactions among the queries within a human instance, exploiting kinematic relations and gathering information for scoring pose predictions. It calculates \\(N\\) self-attention maps for \\(N\\) human instances in parallel with shapes of \\((K + 1)\\times (K + 1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.882, + 0.325 + ], + "angle": 0, + "content": "omitting the dimensions of batch size and attention heads." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.481 + ], + "angle": 0, + "content": "\\((K + 1)\\) same-type across-instance self-attention. Similar to the within-instance group self-attention, We add another group self-attention, which collect information from the same-type queries in other instances and help remove duplicate predictions. We build interactions across human instances with the same-type queries, the instance query and each keypoint queries, resulting in \\(K + 1\\) same-type across-instance self-attention. It also can be implemented in parallel with one self-attention module, calculating \\(K + 1\\) self-attention maps with shapes of \\(N \\times N\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.893, + 0.603 + ], + "angle": 0, + "content": "Compared with the standard self-attention, our two subsequent group self-attention explicitly explores the information about queries belong to the same human instance and with the same type, while remove the across-instance interactions for queries with different types. Empirical results in Section 4.3 show that the removal of this kind of interactions eases the optimization and thus improves the performance for our Group Pose." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.633, + 0.637 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.598, + 0.661 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.865 + ], + "angle": 0, + "content": "Datasets. Our experiments are conducted on two representative human pose estimation datasets, MS COCO [16] and CrowdPose [14]. MS COCO contains 200K images and 250K person instances with 17 keypoint annotations per instance (we set the number of keypoint queries as \\( K = 17 \\) on MS COCO). CrowdPose has 20K images and 80K person instances with 14 keypoint annotations per instance (\\( K = 14 \\) on CrowdPose). CrowdPose is more challenging as it includes many crowd and occlusion scenes. We train Group Pose on COCO train2017 set and evaluate it on COCO val2017 set and test-dev set. On CrowdPose, we train our model on the train set and evaluate it on the test set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation metric. The OKS-based average precision (AP) scores are reported as the main metric for both" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15032" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.148 + ], + "angle": 0, + "content": "Table 1: Comparisons with state-of-the-art methods on MS COCO va12017. We also provide the reference (Ref) for previous frameworks. The 'HM', 'BR', and 'KR' denote heatmap-based losses, human box regression losses, and keypoint regression losses. 'RLE' represents the residual log-likelihood estimation in Poseur [21]. † denotes the flipping test. ‡ removes the prediction uncertainty estimation in Poseur as a fair regression comparison. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.156, + 0.891, + 0.507 + ], + "angle": 0, + "content": "
MethodRefBackboneLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_M \\)\\( AP_L \\)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM65.587.271.161.373.4
Mask R-CNN [8]CVPR 17ResNet-101HM66.187.472.061.574.4
\\( PRTR^† \\)[15]CVPR 21ResNet-50KR68.288.275.263.276.2
\\( Poseur^‡ \\)[21]ECCV 22ResNet-50RLE70.0----
Poseur [21]ECCV 22ResNet-50RLE74.289.881.371.180.1
Bottom-Up\\( HrHRNet^† \\)[5]CVPR 20HRNet-w32HM67.186.273.061.576.1
\\( DEKR^† \\)[7]CVPR 21HRNet-w32HM68.086.774.562.177.7
\\( SWAHR^† \\)[20]CVPR 21HRNet-w32HM68.987.874.963.077.4
\\( LOGO-CAP^† \\)[37]CVPR 22HRNet-w32HM69.687.575.964.178.0
One-StageDirectPose [31]-ResNet-50KR63.185.668.857.771.3
\\( CenterNet^† \\)[40]-Hourglass-104KR+HM64.0----
FCPose [22]CVPR 21ResNet-50KR+HM63.085.968.959.170.3
InsPose [28]ACM MM 21ResNet-50KR+HM63.186.268.558.570.1
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR68.887.576.362.777.7
PETR [27]CVPR 22Swin-LHM+KR73.190.780.967.281.7
QueryPose [36]NeurIPS 22ResNet-50BR+RLE68.788.674.463.876.5
QueryPose [36]NeurIPS 22Swin-LBR+RLE73.391.379.568.581.2
ED-Pose [38]ICLR 23ResNet-50BR+KR71.689.678.165.979.8
ED-Pose [38]ICLR 23Swin-LBR+KR74.391.581.668.682.6
OursGroupPose-ResNet-50KR72.089.479.166.879.7
GroupPose-Swin-TKR73.690.480.568.781.2
GroupPose-Swin-LKR74.891.682.169.483.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.471, + 0.656 + ], + "angle": 0, + "content": "datasets. For MS COCO, we adopt AP with different thresholds and different object sizes (medium and large), denoted as AP, \\(\\mathrm{AP}_{50}\\), \\(\\mathrm{AP}_{75}\\), \\(\\mathrm{AP}_M\\), and \\(\\mathrm{AP}_L\\), following the standard evaluation process3. On CrowdPose, to better evaluate the model performance in different crowded scenarios, we adopt AP with different thresholds and different crowding levels, denoted as AP, \\(\\mathrm{AP}_{50}\\), and \\(\\mathrm{AP}_{75}\\), as well as \\(\\mathrm{AP}_E\\), \\(\\mathrm{AP}_M\\) and \\(\\mathrm{AP}_H\\) for images with easy, medium and hard crowding levels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.472, + 0.875 + ], + "angle": 0, + "content": "Implementation details. Our training and testing settings follow ED-Pose [38]. During training, we adopt the widely-used data augmentations in DETR frameworks [2, 41, 39, 38], including random flip, random crop, and random resize with the short sides in [480, 800] and the long side less or equal to 1333. We use the AdamW optimizer[11, 19] with the weight decay \\(1 \\times 10^{-4}\\) and train 60 epochs and 80 epochs on MS COCO [16] and CrowdPose [14], respectively. We adopt a total batch size of 16, and set the base learning rate as \\(1 \\times 10^{-4}\\). The base learning rate for the backbone is \\(1 \\times 10^{-5}\\) following the DETR frameworks. The learning rates are decayed at the 50-th epoch and 70-th by a factor of 0.1 for MS COCO and CrowdPose, respectively. During testing, we resize the input images with their short" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.52, + 0.848, + 0.536 + ], + "angle": 0, + "content": "sides being 800 and long sides less or equal to 1333." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.64, + 0.566 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.698 + ], + "angle": 0, + "content": "Our purpose is to build a simple baseline for end-to-end multi-person pose estimation. Thus, we mainly compare our Group Pose with previous end-to-end frameworks, including PETR [27], QueryPose [36], and ED-Pose [38]. Besides, to show the effectiveness of our method, we also add comparisons with non-end-to-end frameworks, such as top-down [8, 15], bottom-up [5, 7, 20, 37], and one-stage methods [31, 40, 22, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Comparisons with end-to-end frameworks on COCO. Table 1 and Table 2 present the comparisons on COCO val2017 set and test-dev set. Results show that Group Pose outperforms PETR [27], QueryPose [36], and ED-Pose [38] consistently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "On COCO val2017, Group Pose surpasses PETR and QueryPose by over a significant 3.0 AP with ResNet-50 [9] and the gaps remain \\(1.5 + \\mathrm{AP}\\) with a strong backbone, SwinLarge [18]. When comparing with the recently proposed ED-Pose [38], which transfers pose estimation to a keypoint box detection problem and combines human box detection task, Group Pose can also exceed it with non-negligible margins. Moreover, unlike these methods that use complex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.31, + 0.902 + ], + "angle": 0, + "content": "3https://cocodataset.org/#keypoints-eval" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15033" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.089, + 0.887, + 0.103 + ], + "angle": 0, + "content": "Table 2: Comparisons with state-of-the-art methods on MS COCO test-dev2017 dataset. Notations are consistent with Table 1." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.116, + 0.885, + 0.435 + ], + "angle": 0, + "content": "
MethodRefBackboneLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_M \\)\\( AP_L \\)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM63.987.769.959.771.5
Mask R-CNN [8]CVPR 17ResNet-101HM64.388.270.660.171.9
\\( PRTR^† \\)[15]CVPR 21ResNet-101KR68.889.976.964.775.8
\\( PRTR^† \\)[15]CVPR 21HRNet-w32KR71.790.679.667.678.4
Bottom-Up\\( HrHRNet^† \\)[5]CVPR 20HRNet-w32HM66.487.572.861.274.2
\\( DEKR^† \\)[7]CVPR 21HRNet-w32HM67.387.974.161.576.1
\\( SWAHR^† \\)[20]CVPR 21HRNet-w32HM67.988.974.562.475.5
\\( LOGO-CAP^† \\)[37]CVPR 22HRNet-w32HM68.288.774.962.876.0
One-StageDirectPose [31]-ResNet-50KR62.286.468.256.769.8
\\( CenterNet^† \\)[40]-Hourglass-104KR+HM63.086.869.658.970.4
FCPose [22]CVPR 21ResNet-50KR+HM64.387.371.061.670.5
InsPose [28]ACM MM 21ResNet-50KR+HM65.488.971.760.272.7
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR67.689.875.361.676.0
PETR [27]CVPR 22Swin-LHM+KR70.591.578.765.278.0
QueryPose [36]NeurIPS 22Swin-LBR+RLE72.292.078.867.379.4
ED-Pose [38]ICLR 23ResNet-50BR+KR69.890.277.264.377.4
ED-Pose [38]ICLR 23Swin-LBR+KR72.792.380.967.680.0
OursGroupPose-ResNet-50KR70.290.577.864.778.0
GroupPose-Swin-TKR72.191.479.966.779.5
GroupPose-Swin-LKR72.892.581.067.780.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.525 + ], + "angle": 0, + "content": "decoders and add extra supervisions, e.g., extra heatmap or box supervisions, our Group Pose only use a simple decoder and are trained only with keypoint regression targets. The above evidences indicate that complex design for end-to-end multi-person pose estimation may not be necessary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.47, + 0.603 + ], + "angle": 0, + "content": "On COCO test-dev, our Group Pose achieves 70.2 AP, 72.1 AP, and 72.8 AP with ResNet-50 [9], Swin-Tiny [18], and Swin-Large [18] as the backbone. Compared with other end-to-end frameworks, similar trends are observed with the ones on COCO val2017." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.469, + 0.732 + ], + "angle": 0, + "content": "Comparisons with end-to-end frameworks on CrowdPose. To further demonstrate the effectiveness of Group Pose, we provide comparisons with previous end-to-end frameworks on the challenging CrowdPose dataset [14]. Table 3 reports the results of PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose with the same backbone Swin-Large [18]. Overall, Group Pose achieve 74.1 AP, performing the best over all methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Moreover, it is interesting when we compare the AP scores with easy, medium, and hard crowding levels of different models. PETR [27] performs worse with the easy crowding level while giving the second best results with the hard level. ED-Pose [38] performs worst with the hard crowding level. We conjecture that this phenomenon is caused by the differences in their decoders, e.g., human instances crowded with each other have similar boxes, which challenges the human keypoint decoder. Our Group Pose, which uses \\( K \\) keypoint queries and one instance query for each human instance and considers different interactions" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Table 3: Comparisons with state-of-the-art methods on Crowd-Posetest dataset. Swin-L is adopted as the backbone. Other notations are consistent with Tabel 1." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.5, + 0.892, + 0.578 + ], + "angle": 0, + "content": "
MethodLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_E \\)\\( AP_M \\)\\( AP_H \\)
PETR [27]HM+KR71.690.478.377.372.065.8
QueryPose [36]BR+RLE72.791.778.179.573.465.4
ED-Pose [38]BR+KR73.190.579.880.573.863.8
GroupPoseKR74.191.380.480.874.766.4
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.615, + 0.892, + 0.646 + ], + "angle": 0, + "content": "among queries, can achieve reasonable good results with easy, medium, and hard crowding levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparisons with various non-end-to-end frameworks on COCO. We also compare Group Pose with representative non-end-to-end frameworks in Table 1 and Table 2. Group Pose can easily outperform previous bottom-up methods [5, 7, 20, 37] and one-stage methods [31, 40, 22, 28]. For example, Group Pose surpasses the recent proposed LOGO-CAP [37] with flipping test by over 2.0 AP on both COCO val2017 and test-dev, even with a smaller backbone (ResNet-50 [9]) than it (HRNetw32 [29]). Group Pose is more concise and precise than those bottom-up and one-stage methods. Surprisingly, Group Pose can also beat previous top-down methods like PRTR [15] and Poseur [21]. Given the simple design and the end-to-end property, our Group Pose can serve as a simple baseline for pursuing higher performance on multi-person pose estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15034" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.089, + 0.854, + 0.104 + ], + "angle": 0, + "content": "Table 4: Ablation experiments for Group Pose. Evaluated on MS COCO val2017. Default settings are marked in gray ." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.11, + 0.327, + 0.159 + ], + "angle": 0, + "content": "(a) Query designs for human instances. Both instance (inst) and keypoint (kpt) queries are essential in Group Pose, especially the keypoint ones." + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.11, + 0.63, + 0.158 + ], + "angle": 0, + "content": "(b) Benefits of the instance query. On the model only with kpt queries, the inst query is first added but not for classification (cls). Then we use the inst query for the cls task." + }, + { + "type": "text", + "bbox": [ + 0.643, + 0.11, + 0.892, + 0.158 + ], + "angle": 0, + "content": "(c) Number of instances. When the number is smaller than 100, increasing instance numbers provide gains. 100 is set by default for the number of instances." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.165, + 0.307, + 0.216 + ], + "angle": 0, + "content": "
query typesAPAPMAPL
inst & kpt72.066.879.7
only kpt71.266.079.1
only inst64.561.169.9
" + }, + { + "type": "table", + "bbox": [ + 0.349, + 0.166, + 0.617, + 0.218 + ], + "angle": 0, + "content": "
query typescls taskAPAPMAPL
only kptavg kpt71.266.079.1
inst & kptavg kpt71.766.879.3
inst & kptinst72.066.879.7
" + }, + { + "type": "table", + "bbox": [ + 0.671, + 0.165, + 0.866, + 0.216 + ], + "angle": 0, + "content": "
#instanceAPAPMAPL
5071.466.479.1
10072.066.879.7
20072.066.979.7
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.224, + 0.471, + 0.273 + ], + "angle": 0, + "content": "(d) Self-attention implementations. In the self-attention module, removing the across-instance interactions for queries with different types with attention (attn) mask or group self-attention eases the optimization and improves the performance." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.279, + 0.462, + 0.33 + ], + "angle": 0, + "content": "
self-attention implementationsw/ attn maskAPAPMAPL
standard self-attention×69.464.077.5
standard self-attention✓70.765.578.4
group self-attention×72.066.879.7
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.343, + 0.23, + 0.359 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.37, + 0.47, + 0.431 + ], + "angle": 0, + "content": "We run a number of ablation experiments to verify the effectiveness of key elements in our Group Pose. We adopt ResNet-50 [9] as the backbone. Unless specified, we report the results on COCO val2017 with 60 epochs training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Ablation: query designs for human instance. Table 4a gives the results of representing a human instance with different query designs. Both instance query and keypoint queries are essential in Group Pose. As the multi-person pose estimation is to predict human poses given an image, it is reasonable that keypoint queries themselves can achieve good results, while removing keypoint queries gives a significant performance drop (from 72.0 AP to 64.5 AP)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.638 + ], + "angle": 0, + "content": "For the instance query, its benefits come from two aspects: (i) gather information within human instance and help model training and (ii) decouple the classification task and keypoint positions regression task. Table 4b shows the improvements brought by these two benefits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Ablation: self-attention implementations. Table 4d provides the comparisons between different self-attention in the decoder layers, including (i) standard self-attention over all the \\( N \\times (K + 1) \\) queries (Figure 3 (a)), (ii) standard self-attention with an attention mask (Figure 3 (b)), which masks out the across-instance interactions for queries with different types, and (iii) our group self-attention (Figure 3 (c)). Results validate that the across-instance interactions for queries with different types are not directly helpful for the multi-person pose estimation task. Removing this type of interactions eases the model optimization, helps model converge faster, thus improves the performances. The comparisons on convergence curves are given in Figure 4. Besides, we find that it is better to separately perform different types of interactions, the within-instance and same-type across-instance interactions, with different parameters. Table 4d shows that our sequential implementation group self" + }, + { + "type": "text", + "bbox": [ + 0.493, + 0.224, + 0.888, + 0.273 + ], + "angle": 0, + "content": "(e) Group self-attentions. Both types of self-attentions are important in group self-attentions. We also conduct an experiment by removing both within-instance and across-instance self-attentions. The performance further drops to 65.3 AP." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.279, + 0.88, + 0.33 + ], + "angle": 0, + "content": "
group self-attentionssAPAPMAPL
within-instance & across-instance self-attentionss72.066.879.7
within-instance self-attentionss66.363.870.8
across-instance self-attentionss67.463.074.4
" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.343, + 0.885, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.471, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Figure 4: Faster and better convergence. The \\( x \\)-axis corresponds to #epoch, and the \\( y \\)-axis corresponds to AP score. One can see that group self-attention converges faster and better than other self-attention implementations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.545, + 0.892, + 0.574 + ], + "angle": 0, + "content": " attentions bring a 1.3 AP gain over the parallel implementation of standard self-attention with an attention mask." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.893, + 0.837 + ], + "angle": 0, + "content": "Ablation: group self-attentions. We ablate the effects of \\( N \\) within-instance self-attentions over \\( K \\) keypoint queries and one instance query and \\( (K + 1) \\) across-instance self-attentions over \\( N \\) queries of the same type in Table 4e. Large performance gaps (over \\( -5.0\\mathrm{AP} \\)) are observed when we remove either of these two group self-attentions. Both the interactions are important for multi-person pose estimation. The within-instance self-attentions explore kinematic relations and gather information for scoring pose predictions, which help the model predict precise keypoint positions for human poses. The same-type across-instance self-attentions collect information from the same-type of keypoint queries or instance queries, removing duplicate predictions for poses among human instances. As shown in Figure 5, model without the same-type across-instance self-attentions produces more duplicated pose predictions for the same human instance than Group Pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Ablation: number of instances. Table 4c gives how predefined number of human instances affect the results. We find 100 human instances are enough and can achieve comparable results with 200 human instances. We set the num" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15035" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.29, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.179, + 0.289, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.091, + 0.458, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.179, + 0.458, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.091, + 0.629, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.179, + 0.629, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.091, + 0.791, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.179, + 0.791, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.091, + 0.88, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.179, + 0.88, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.895, + 0.335 + ], + "angle": 0, + "content": "Figure 5: Comparison of removing duplicated predictions. We visualize the predicted poses of Group Pose (second row), and Group Pose without same-type across-instance self-attention (first row) according to the number of objects. Group Pose produces less duplicated pose predictions. The yellow dashed ellipse indicates the duplicated predicted human instance, leading to a mismatch of other instances. Best view in zoom in." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.349, + 0.473, + 0.433 + ], + "angle": 0, + "content": "Table 5: Analysis on model convergence. Group Pose without human detection already can outperform ED-Pose [38]. Besides, Group Pose can also benefit from better human instance initialization with human detection. We report AP on COCO val2017 dataset. 'Det Dec' = human detection decoder. The results of ED-Pose are from the original paper [38]." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.444, + 0.469, + 0.52 + ], + "angle": 0, + "content": "
Methodw/ Det Dec12e24e36e48e60e
ED-Pose [38]✓60.567.569.770.871.6
GroupPose×61.067.670.171.472.0
GroupPose✓61.468.170.371.672.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.579 + ], + "angle": 0, + "content": "ber of human instances as 100 by default. Thus, Group Pose has \\(100 \\times (17 + 1)\\) queries on COCO [16] and contains \\(100 \\times (14 + 1)\\) queries on CrowdPose [14]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.226, + 0.603 + ], + "angle": 0, + "content": "4.4. More Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.656 + ], + "angle": 0, + "content": "We provide more analysis about our Group Pose in this section, including the analysis on model convergence and the analysis on model inference speed, detailed next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Analysis on model convergence. A common wisdom about training models is to provide a good initialization and then refine the model based on it. ED-Pose [38] splits the multi-person pose estimation task into two sub-processes, which first detect human instances with a human detection decoder and then use a human-to-keypoint detection decoder for predicting human poses. The human detection decoder gives good initialization for human instances, which can help ED-Pose learn and converge faster. Although our Group Pose adopts a different design for transformer decoder, it can also be built upon a human detection decoder and enjoys the benefits brought by better initialization of human instances. Based on this observation, we build a variant of our Group Pose, whose transformer decoder consists of 2 human detection decoder layers and 4 simple transformer decoder layers for multi-person pose estimation, following" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.351, + 0.599, + 0.366 + ], + "angle": 0, + "content": "ED-Pose [38]." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.394, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Table 6: Analysis on model inference speed. Frames per second (FPS) and inference time (Time) are measured with ResNet-50 and different image resolutions on one NVIDIA A100 GPU." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.446, + 0.905, + 0.578 + ], + "angle": 0, + "content": "
MethodInput ResolutionFPS ↑Time [ms] ↓
PETR [27]480 × 80020.050
800 × 133312.183
QueryPose [36]480 × 80019.056
800 × 133313.475
ED-Pose [38]480 × 80042.424
800 × 133324.740
GroupPose480 × 80068.615
800 × 133331.332
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.894, + 0.753 + ], + "angle": 0, + "content": "Table 5 shows the AP scores of three models, ED-Pose, Group Pose, and Group Pose with human detection decoder, under different training schedules. Group Pose already shows superior results than ED-Pose with 12 epochs, 24 epochs, 36 epochs, 48 epochs, and 60 epochs training, even without the help of human detection decoder. Moreover, the comparison between Group Pose and Group Pose with human detection decoder verifies that good initialization for human instances leads to faster convergence and better results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Analysis on model inference speed. Table 6 provides the comparisons of inference time and FPS among end-to-end frameworks, including PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose. Simple designs usually show efficiency. Results show that our simple transformer decoder can be faster than complex decoders in previous end-to-end frameworks. Even with an image in a large size \\(800 \\times 1333\\), our Group Pose can also achieve real-time speed (above 30 FPS) on a single A100 GPU." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15036" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.196, + 0.106 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.117, + 0.47, + 0.252 + ], + "angle": 0, + "content": "In this paper, we present a simple baseline, Group Pose, for end-to-end multi-person pose estimation. With simple designs for queries and decoder self-attention, the approach outperforms previous end-to-end frameworks while being faster. Besides, the transformer decoder in Group Pose is also flexible, which can be built solely or upon human detection decoders. We hope Group Pose can provide insights for exploring concise and effective end-to-end multi-person pose estimation frameworks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.254, + 0.471, + 0.314 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.U1936212, No.62120106009, No.62261160653)." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.331, + 0.175, + 0.347 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.472, + 0.497 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.473, + 0.556 + ], + "angle": 0, + "content": "[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.473, + 0.628 + ], + "angle": 0, + "content": "[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5386-5395, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.716, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[6] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2334-2343, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[7] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2961-2969, 2017. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.358, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[10] Jan Hosang, Rodrigo Benenson, and Bernt Schiele. Learning non-maximum suppression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4507-4515, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[12] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11977-11986, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[14] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10863-10872, 2019. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[15] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[17] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[20] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[21] Weian Mao, Yongtao Ge, Chunhua Shen, Zhi Tian, Xinlong Wang, Zhibin Wang, and Anton van den Hengel. Poseur: Direct human pose regression with transformers. In Proceedings of the European Conference on Computer Vision, pages 72-88, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[22] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15037" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[23] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in Neural Information Processing Systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[25] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision, pages 483-499. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.376 + ], + "angle": 0, + "content": "[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6951-6960, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[27] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.471, + 0.518 + ], + "angle": 0, + "content": "[28] Dahu Shi, Xing Wei, Xiaodong Yu, Wenming Tan, Ye Ren, and Shiliang Pu. Inspose: instance-aware networks for single-stage multi-person pose estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3079-3087, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[29] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[30] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.716 + ], + "angle": 0, + "content": "[31] Zhi Tian, Hao Chen, and Chunhua Shen. Directpose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[34] Fangyun Wei, Xiao Sun, Hongyang Li, Jingdong Wang, and Stephen Lin. Point-set anchors for object detection, instance segmentation and pose estimation. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision, pages 527-544. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[35] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision, pages 466-481, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[36] Yabo Xiao, Kai Su, Xiaojuan Wang, Dongdong Yu, Lei Jin, Mingshu He, and Zehuan Yuan. Querypose: Sparse multiperson pose regression via spatial-aware part-level query. In Advances in Neural Information Processing Systems, pages 1-14, 2022. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[37] Nan Xue, Tianfu Wu, Gui-Song Xia, and Liangpei Zhang. Learning local-global contextual adaptation for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13065-13074, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[38] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In International Conference on Learning Representations, pages 1-17, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[39] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, pages 1-18, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[40] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, pages 1-16, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15038" + } + ] +] \ No newline at end of file diff --git a/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_origin.pdf b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30d05d7a75be498d7891418b54d3581fc9757d28 --- /dev/null +++ b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/ebcfd119-a15f-45f3-824b-f01682d5c7b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9c5ee79b831463a5a055692b9ea19a4f0b64f31505140cc0d747605fde4e30d +size 1973118 diff --git a/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/full.md b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..feee6c4fea36796a40690497198ba6ef5514fd86 --- /dev/null +++ b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/full.md @@ -0,0 +1,292 @@ +# Group Pose: A Simple Baseline for End-to-End Multi-person Pose Estimation + +Huan Liu $^{1,3*}$ Qiang Chen $^{2*}$ Zichang Tan $^{2}$ Jiang-Jiang Liu $^{2}$ Jian Wang $^{2}$ Xiangbo Su $^{2}$ Xiaolong Li $^{1,3}$ Kun Yao $^{2}$ Junyu Han $^{2}$ Errui Ding $^{2}$ Yao Zhao $^{1,3\dagger}$ Jingdong Wang $^{2}$ + +$^{1}$ Institute of Information Science, Beijing Jiaotong University $^{2}$ Baidu VIS $^{3}$ Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing, China + +# Abstract + +In this paper, we study the problem of end-to-end multiperson pose estimation. State-of-the-art solutions adopt the DETR-like framework, and mainly develop the complex decoder, e.g., regarding pose estimation as keypoint box detection and combining with human detection in ED-Pose [38], hierarchically predicting with pose decoder and joint (keypoint) decoder in PETR [27]. + +We present a simple yet effective transformer approach, named Group Pose. We simply regard $K$ -keypoint pose estimation as predicting a set of $N \times K$ keypoint positions, each from a keypoint query, as well as representing each pose with an instance query for scoring $N$ pose predictions. + +Motivated by the intuition that the interaction, among across-instance queries of different types, is not directly helpful, we make a simple modification to decoder self-attention. We replace single self-attention over all the $N \times (K + 1)$ queries with two subsequent group self-attention: (i) $N$ within-instance self-attention, with each over $K$ keypoint queries and one instance query, and (ii) $(K + 1)$ same-type across-instance self-attention, each over $N$ queries of the same type. The resulting decoder removes the interaction among across-instance type-different queries, easing the optimization and thus improving the performance. Experimental results on MS COCO and CrowdPose show that our approach without human box supervision is superior to previous methods with complex decoders, and even is slightly better than ED-Pose that uses human box supervision. Paddle1 and PyTorch2 codes are available. + +# 1. Introduction + +Multi-person pose estimation aims to detect the corresponding human keypoints for all human instances in an + +![](images/437a2e132d7408f4d6ab28929d8ba413a7667cc3aa9779c1387b4dd615c4c1bb.jpg) +(a) PETR + +![](images/4ddb363828897d62896c274eff862023799eacec0f75c98e3634b7eb5e5fd16d.jpg) + +![](images/73df08414980fcafaa87f02a383e1d47af7fa6b19897415042a65ab8f4ec3e3f.jpg) +(b) ED-Pose +(c) Group Pose +Figure 1: Comparison of transformer decoders. Here we mainly illustrate the overview of the decoder part of PETR [27], ED-Pose [38], and our Group Pose. The three end-to-end frameworks differ in query design and decoder architecture. Group Pose only uses a simple transformer decoder rather than developing complex decoders. 'inst' and 'kpt' represent for instance and keypoint. + +image. Previous frameworks include top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5] that divide the task into two sequential sub-tasks: human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping [24]. Another line in previous frameworks is one-stage methods [22, 26, 31, 34], which directly predict instance-aware keypoints. These frameworks rely on non-differentiable hand-crafted post-processes [10, 24], which complicate the pipelines and challenge the optimizations. Inspired by the success of DETR [2] in object detection, building an end-to-end framework for multi-person pose estimation has seen significant interest. + +Recent approaches follow the DETR framework [2, 41, 39], with the transformer encoder-decoder architecture for multi-person pose estimation, as shown in Figure 1. PETR [27] hierarchically predicts the keypoint positions and uses two subsequent decoders, pose decoder and joint decoder, with two different queries, pose query (a person has one pose query) for pose decoder, and keypoint queries for joint decoder. ED-Pose [38] transfers pose estimation to a keypoint box detection problem, and learns a content query and a box query for each keypoint position prediction + +with using the box size to process the query. + +In this paper, we present a simple yet effective transformer approach, named Group Pose, for end-to-end multiperson pose estimation. Instead of using a single query to predict and score one pose, similar to ED-Pose, we use the $N \times K$ keypoint queries to regress the $N \times K$ positions, by regarding each keypoint as an object, as well as $N$ instance queries, each representing a $K$ -keypoint pose for scoring the $K$ -keypoint pose prediction. + +We make a simple modification for the decoder architecture. We replace standard self-attention in the decoder with two subsequent group self-attention: $N$ parallel self-attention with each over $K$ keypoint queries and the corresponding instance query for exploiting kinematic relation and gathering information for scoring pose predictions, and $(K + 1)$ parallel self-attention with each over $N$ queries of the same type for collecting duplicate prediction information like self-attention of the original DETR. + +The two self-attention captures two kinds of interactions: (i) $N$ within-instance interactions over $K$ keypoint queries and one instance query, (ii) $(K + 1)$ across-instance interactions over $N$ queries of the same type (e.g., nose keypoint query or instance query). The extra interactions in standard self-attention are (iii) across-instance interactions for queries with different types, which is not directly useful. Empirical results show that the removal of the third kind of interactions eases the optimization and thus improves the performance. + +The design about self-attention is different from the closely-related approach ED-Pose [38]. On the one hand, in addition to removing the third interactions, ED-Pose only models across-instance interactions for the instance queries. In contrast, our approach also models across-instance interactions for the $K$ keypoint types. On the other hand, our approach separates the two interactions using two subsequent group self-attention, explicitly exploring the information about queries belonging to the same human instance, and with the same type. ED-Pose couples the two interactions using a single masked self-attention with the third interactions masked. It is empirically demonstrated that the two differences benefit the pose estimation performance. + +Experimental results show that our simple approach Group Pose without human box supervision surprisingly outperforms the recent end-to-end methods with human box supervisions on MS COCO [16] and CrowdPose [14]. Notably, Group Pose achieves 72.0 AP with ResNet-50 [9] and 74.8 AP with Swin-Large [18] on MS COCO val2017. We hope our simple transformer decoder in Group Pose will motivate people to simplify the design in end-to-end multiperson pose estimation. + +# 2. Related Work + +Multi-person pose estimation is a challenging task that aims to detect the corresponding human keypoints for all human instances in an image. Previous methods usually adopt complex frameworks to address it, which are divided into non-end-to-end and end-to-end methods. + +Non-end-to-end methods. There are typically two types: two-stage and one-stage. Two-stage frameworks, including top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5], split the multi-person pose estimation task into two sequential sub-tasks, human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping. In top-down methods, an object detector is first employed to detect the boxes of human instances, which is then cropped for single-person pose estimation in each box. Bottom-up methods first predict all human keypoints in a human-agnostic way then group them into instances. While one-stage frameworks [40, 22, 26, 31, 34] directly predict instance-aware keypoints. These methods require hand-crafted pose-processes, such as NMS [10] or grouping [24], which complicate the pipelines and challenge the optimizations. In this paper, we focus on concise ways, end-to-end frameworks. + +End-to-end methods. Current end-to-end multi-person pose estimation frameworks [27, 36, 38] are built by following the designs of DETR [2] and its variants [41, 23, 3, 17, 13, 39]. They adopt a paradigm of splitting the multiperson pose estimation task into two sub-processes. For example, PETR [27] views the task as a hierarchical set prediction problem. It first determines human instances by predicting a set of poses with a pose decoder and then refines the keypoints in each pose with a joint (keypoint) decoder. There are also two different types of queries, pose query (a person has one pose query) for pose decoder and keypoint queries for joint decoder. QueryPose [36] and ED-Pose [38] follow this end-to-end paradigm but further incorporate an extra human detection task. QueryPose [36] follows Sparse R-CNN [30] to build two parallel RoIAlign-based [8] decoders to perform human detection and pose estimation, respectively. ED-Pose [38] transfers pose estimation task to a keypoint box detection problem. It first employs a human detection decoder [39] to determine the human instances with box queries. Then it builds a human-to-keypoint detection decoder with a content query and a box query for each keypoint position, collecting contextual information near keypoint positions. + +Although these end-to-end methods show promising results in multi-person pose estimation, they rely on complex decoders. Our Group Pose, on the other hand, adopts a simple transformer decoder, improving the performance and simplifying the process. + +![](images/27218fbf40913cf0091d027d05ad545804fd92c209cec145c29f801bee2eff7a.jpg) +Figure 2: Our Group Pose architecture. The backbone takes an image as input and outputs image features, which are refined by the transformer encoder. To directly predict $N$ human poses with $K$ keypoint positions in each pose, we adopt $N \times (K + 1)$ queries, containing $N$ instance queries for scoring poses and $N \times K$ keypoint queries for regressing positions. + +# 3. Group Pose + +Group Pose is a simple yet effective end-to-end multi-person pose estimation framework. We follow previous end-to-end frameworks [27, 38] to view the multi-person pose estimation task as a set prediction problem, but directly adopt a simple transformer decoder [41] instead of complex decoders, simplifying the process. Next, we introduce the key elements of Group Pose. + +# 3.1. Overview + +The overall structure of Group Pose is depicted in Figure 2. Group Pose consists of a backbone [9, 18], a transformer encoder [32], a transformer decoder, and task-specific prediction heads. This framework enables Group Pose to simultaneously regresses $K$ keypoints (e.g., $K = 17$ on MS COCO) for $N$ human instances given an image. + +Backbone and transformer encoder. We directly follow DETR frameworks [41] to build the backbone and the transformer encoder (with 6 deformable transformer layers [41]) for Group Pose. It takes an image as input and outputs the extracted multi-level features, which serve as inputs for the following transformer decoder. We use 4 feature levels in Group Pose, with downsampling rates of $\{8, 16, 32, 64\}$ . + +Transformer decoder. For transformer decoder, we adopt a combination of $N \times K$ keypoint queries and $N$ instance queries as input instead of using a single query to predict and score one pose. The keypoint queries regard each keypoint as an object and are used to regress the $N \times K$ keypoint positions, while each instance query is for scoring the corresponding $K$ -keypoint pose prediction. Besides, the architecture of transformer decoder is simple, which stacks 6 same decoder layers [41]. In each decoder layer, we follow the macro design of previous DETR frameworks by building self-attention, cross-attention implemented with deformable attention, and FFN. We only make simple modifications to self-attention in our Group Pose. Specifically, we replace the standard self-attention with two subsequent group self-attention, enabling decoder layers to perform interactions over queries belonging to the same human instance and with the same type. + +Prediction heads. There are two prediction heads implemented with FFNs in Group Pose for human classification and human keypoints regression. Group Pose predicts for $N$ human poses, each contains a classification score and $K$ keypoints positions for the corresponding $K$ -keypoint pose. + +Loss function. The Hungarian matching algorithm [2] is employed for one-to-one assignment between predicted poses and ground-truth poses. Our loss function comprises solely of classification loss $(\mathcal{L}_{cls})$ and keypoint regression loss $(\mathcal{L}_{kpt})$ , without any extra supervisions such as human detection loss in QueryPose [36] and ED-Pose [38] or heatmap loss in PETR [27]. The keypoint regression loss $(\mathcal{L}_{kpt})$ is a combination of a normal $\ell_1$ loss and a constrained $\ell_1$ loss named Object Keypoint Similarity (OKS) [27]. We directly use the cost coefficients and loss weights of ED-Pose [38] in the hungarian matching and the calculation of losses. + +# 3.2. $N \times K$ keypoint queries and $N$ instance queries + +In multi-person pose estimation, frameworks are required to predict $N$ human poses with $K$ keypoint positions in each pose given an image. We directly use $N \times K$ keypoints queries to predict the poses, with each $K$ predicted keypoint positions to represent a corresponding $K$ -keypoint pose for a human instance. To classify if it is a human instance by scoring the predicted $N$ human poses, we also introduce $N$ instance queries. Thus, in Group Pose, a human instance can be represented with a combination of $K$ keypoint queries and one instance query. As these two types of queries are responsible for different tasks, we construct and initialize them differently. + +Query construction. We follow the previous end-to-end frameworks [27, 38] to first identify human instances and predict human poses in each position of the output memory from the transformer encoder. Then we select $N$ human instances ( $N = 100$ ) based on the classification scores, resulting $N$ human poses and the corresponding output memory features of the selected $N$ positions (with the shape of $N \times D$ , where $D$ is the channel dimension). We then construct and initialize the keypoint queries and the instance + +![](images/92ed9b5f62fd89906b8ec4e8043e59b1d924df628c3bb2ae823ad810a7170d57.jpg) +$(N\times (K + 1))\times (N\times (K + 1))$ +(a) standard self-attention + +![](images/7adcb61177411121267fa4a2e6d0d63cca5e8d2983a91ee894b32ed1c50c216f.jpg) +$(N\times (K + 1))\times (N\times (K + 1))$ +(b) standard self-attention with mask +Figure 3: Conceptual comparison of self-attention implementations. Given $N \times (K + 1)$ queries, three different implementations of self-attention are considered in the decoder layers. (a) standard self-attention over all queries. (b) standard self-attention with an attention mask, removing the across-instance interactions for queries with different types. (c) our proposed group self-attention. Dashed boxes indicate the masked-out interactions. + +![](images/24322e3e8f6c28a6ebdf809954965d38494d3cd194fa52f1b16cc4ba454b5b7e.jpg) +$(N\times ((K + 1)\times (K + 1)) + ((K + 1)\times (N\times N))$ +(c) group self-attention + +queries based on the above results. + +For $N \times K$ keypoint queries, we construct the content part of $K$ keypoint queries $(K \times D)$ in each human instance by combining $K$ randomly initialized learnable keypoint embeddings $(K \times D)$ and the corresponding output memory feature $(1 \times D)$ . And the position part of $K$ keypoint queries $(K \times 2)$ in each human instance is initialized with the corresponding predicted $K$ -keypoint pose $(K \times 2)$ . For $N$ instance queries, we only consider the content part and use a randomly initialized learnable instance embedding $(1 \times D)$ for each human instance. This is because instance query is for the classification task, which does not require explicit position information. When performing cross-attention in the decoder layers, we simply use the mean of $K$ keypoint positions as the reference point for the instance query in each human instance. + +# 3.3. Group self-attention + +In DETR frameworks [2, 41, 23], self-attention in the transformer decoder is usually applied to model interactions among queries, collecting information from other queries and facilitating the duplicate removal for instances. While the situation is slightly different in Group Pose, as there two different types of queries, $K$ keypoint queries and one instance query for each human instance. Motivated by the intuition that the interactions among across-instance queries of different types, e.g., the interactions between the instance query and the keypoint queries across human instances, may not be directly helpful for the above purposes. We thus replace the standard self-attention over all the $N \times (K + 1)$ queries with two subsequent group self-attention: $N$ within-instance self-attention and $K + 1$ same-type across-instance self-attention. + +$N$ within-instance self-attention. We build interactions among the queries within a human instance, exploiting kinematic relations and gathering information for scoring pose predictions. It calculates $N$ self-attention maps for $N$ human instances in parallel with shapes of $(K + 1)\times (K + 1)$ + +omitting the dimensions of batch size and attention heads. + +$(K + 1)$ same-type across-instance self-attention. Similar to the within-instance group self-attention, We add another group self-attention, which collect information from the same-type queries in other instances and help remove duplicate predictions. We build interactions across human instances with the same-type queries, the instance query and each keypoint queries, resulting in $K + 1$ same-type across-instance self-attention. It also can be implemented in parallel with one self-attention module, calculating $K + 1$ self-attention maps with shapes of $N \times N$ . + +Compared with the standard self-attention, our two subsequent group self-attention explicitly explores the information about queries belong to the same human instance and with the same type, while remove the across-instance interactions for queries with different types. Empirical results in Section 4.3 show that the removal of this kind of interactions eases the optimization and thus improves the performance for our Group Pose. + +# 4. Experiments + +# 4.1. Settings + +Datasets. Our experiments are conducted on two representative human pose estimation datasets, MS COCO [16] and CrowdPose [14]. MS COCO contains 200K images and 250K person instances with 17 keypoint annotations per instance (we set the number of keypoint queries as $K = 17$ on MS COCO). CrowdPose has 20K images and 80K person instances with 14 keypoint annotations per instance ( $K = 14$ on CrowdPose). CrowdPose is more challenging as it includes many crowd and occlusion scenes. We train Group Pose on COCO train2017 set and evaluate it on COCO val2017 set and test-dev set. On CrowdPose, we train our model on the train set and evaluate it on the test set. + +Evaluation metric. The OKS-based average precision (AP) scores are reported as the main metric for both + +Table 1: Comparisons with state-of-the-art methods on MS COCO va12017. We also provide the reference (Ref) for previous frameworks. The 'HM', 'BR', and 'KR' denote heatmap-based losses, human box regression losses, and keypoint regression losses. 'RLE' represents the residual log-likelihood estimation in Poseur [21]. † denotes the flipping test. ‡ removes the prediction uncertainty estimation in Poseur as a fair regression comparison. The best results are highlighted in bold. + +
MethodRefBackboneLossAP\( AP_{50} \)\( AP_{75} \)\( AP_M \)\( AP_L \)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM65.587.271.161.373.4
Mask R-CNN [8]CVPR 17ResNet-101HM66.187.472.061.574.4
\( PRTR^† \)[15]CVPR 21ResNet-50KR68.288.275.263.276.2
\( Poseur^‡ \)[21]ECCV 22ResNet-50RLE70.0----
Poseur [21]ECCV 22ResNet-50RLE74.289.881.371.180.1
Bottom-Up\( HrHRNet^† \)[5]CVPR 20HRNet-w32HM67.186.273.061.576.1
\( DEKR^† \)[7]CVPR 21HRNet-w32HM68.086.774.562.177.7
\( SWAHR^† \)[20]CVPR 21HRNet-w32HM68.987.874.963.077.4
\( LOGO-CAP^† \)[37]CVPR 22HRNet-w32HM69.687.575.964.178.0
One-StageDirectPose [31]-ResNet-50KR63.185.668.857.771.3
\( CenterNet^† \)[40]-Hourglass-104KR+HM64.0----
FCPose [22]CVPR 21ResNet-50KR+HM63.085.968.959.170.3
InsPose [28]ACM MM 21ResNet-50KR+HM63.186.268.558.570.1
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR68.887.576.362.777.7
PETR [27]CVPR 22Swin-LHM+KR73.190.780.967.281.7
QueryPose [36]NeurIPS 22ResNet-50BR+RLE68.788.674.463.876.5
QueryPose [36]NeurIPS 22Swin-LBR+RLE73.391.379.568.581.2
ED-Pose [38]ICLR 23ResNet-50BR+KR71.689.678.165.979.8
ED-Pose [38]ICLR 23Swin-LBR+KR74.391.581.668.682.6
OursGroupPose-ResNet-50KR72.089.479.166.879.7
GroupPose-Swin-TKR73.690.480.568.781.2
GroupPose-Swin-LKR74.891.682.169.483.0
+ +datasets. For MS COCO, we adopt AP with different thresholds and different object sizes (medium and large), denoted as AP, $\mathrm{AP}_{50}$ , $\mathrm{AP}_{75}$ , $\mathrm{AP}_M$ , and $\mathrm{AP}_L$ , following the standard evaluation process3. On CrowdPose, to better evaluate the model performance in different crowded scenarios, we adopt AP with different thresholds and different crowding levels, denoted as AP, $\mathrm{AP}_{50}$ , and $\mathrm{AP}_{75}$ , as well as $\mathrm{AP}_E$ , $\mathrm{AP}_M$ and $\mathrm{AP}_H$ for images with easy, medium and hard crowding levels. + +Implementation details. Our training and testing settings follow ED-Pose [38]. During training, we adopt the widely-used data augmentations in DETR frameworks [2, 41, 39, 38], including random flip, random crop, and random resize with the short sides in [480, 800] and the long side less or equal to 1333. We use the AdamW optimizer[11, 19] with the weight decay $1 \times 10^{-4}$ and train 60 epochs and 80 epochs on MS COCO [16] and CrowdPose [14], respectively. We adopt a total batch size of 16, and set the base learning rate as $1 \times 10^{-4}$ . The base learning rate for the backbone is $1 \times 10^{-5}$ following the DETR frameworks. The learning rates are decayed at the 50-th epoch and 70-th by a factor of 0.1 for MS COCO and CrowdPose, respectively. During testing, we resize the input images with their short + +sides being 800 and long sides less or equal to 1333. + +# 4.2. Main Results + +Our purpose is to build a simple baseline for end-to-end multi-person pose estimation. Thus, we mainly compare our Group Pose with previous end-to-end frameworks, including PETR [27], QueryPose [36], and ED-Pose [38]. Besides, to show the effectiveness of our method, we also add comparisons with non-end-to-end frameworks, such as top-down [8, 15], bottom-up [5, 7, 20, 37], and one-stage methods [31, 40, 22, 28]. + +Comparisons with end-to-end frameworks on COCO. Table 1 and Table 2 present the comparisons on COCO val2017 set and test-dev set. Results show that Group Pose outperforms PETR [27], QueryPose [36], and ED-Pose [38] consistently. + +On COCO val2017, Group Pose surpasses PETR and QueryPose by over a significant 3.0 AP with ResNet-50 [9] and the gaps remain $1.5 + \mathrm{AP}$ with a strong backbone, SwinLarge [18]. When comparing with the recently proposed ED-Pose [38], which transfers pose estimation to a keypoint box detection problem and combines human box detection task, Group Pose can also exceed it with non-negligible margins. Moreover, unlike these methods that use complex + +Table 2: Comparisons with state-of-the-art methods on MS COCO test-dev2017 dataset. Notations are consistent with Table 1. + +
MethodRefBackboneLossAP\( AP_{50} \)\( AP_{75} \)\( AP_M \)\( AP_L \)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM63.987.769.959.771.5
Mask R-CNN [8]CVPR 17ResNet-101HM64.388.270.660.171.9
\( PRTR^† \)[15]CVPR 21ResNet-101KR68.889.976.964.775.8
\( PRTR^† \)[15]CVPR 21HRNet-w32KR71.790.679.667.678.4
Bottom-Up\( HrHRNet^† \)[5]CVPR 20HRNet-w32HM66.487.572.861.274.2
\( DEKR^† \)[7]CVPR 21HRNet-w32HM67.387.974.161.576.1
\( SWAHR^† \)[20]CVPR 21HRNet-w32HM67.988.974.562.475.5
\( LOGO-CAP^† \)[37]CVPR 22HRNet-w32HM68.288.774.962.876.0
One-StageDirectPose [31]-ResNet-50KR62.286.468.256.769.8
\( CenterNet^† \)[40]-Hourglass-104KR+HM63.086.869.658.970.4
FCPose [22]CVPR 21ResNet-50KR+HM64.387.371.061.670.5
InsPose [28]ACM MM 21ResNet-50KR+HM65.488.971.760.272.7
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR67.689.875.361.676.0
PETR [27]CVPR 22Swin-LHM+KR70.591.578.765.278.0
QueryPose [36]NeurIPS 22Swin-LBR+RLE72.292.078.867.379.4
ED-Pose [38]ICLR 23ResNet-50BR+KR69.890.277.264.377.4
ED-Pose [38]ICLR 23Swin-LBR+KR72.792.380.967.680.0
OursGroupPose-ResNet-50KR70.290.577.864.778.0
GroupPose-Swin-TKR72.191.479.966.779.5
GroupPose-Swin-LKR72.892.581.067.780.3
+ +decoders and add extra supervisions, e.g., extra heatmap or box supervisions, our Group Pose only use a simple decoder and are trained only with keypoint regression targets. The above evidences indicate that complex design for end-to-end multi-person pose estimation may not be necessary. + +On COCO test-dev, our Group Pose achieves 70.2 AP, 72.1 AP, and 72.8 AP with ResNet-50 [9], Swin-Tiny [18], and Swin-Large [18] as the backbone. Compared with other end-to-end frameworks, similar trends are observed with the ones on COCO val2017. + +Comparisons with end-to-end frameworks on CrowdPose. To further demonstrate the effectiveness of Group Pose, we provide comparisons with previous end-to-end frameworks on the challenging CrowdPose dataset [14]. Table 3 reports the results of PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose with the same backbone Swin-Large [18]. Overall, Group Pose achieve 74.1 AP, performing the best over all methods. + +Moreover, it is interesting when we compare the AP scores with easy, medium, and hard crowding levels of different models. PETR [27] performs worse with the easy crowding level while giving the second best results with the hard level. ED-Pose [38] performs worst with the hard crowding level. We conjecture that this phenomenon is caused by the differences in their decoders, e.g., human instances crowded with each other have similar boxes, which challenges the human keypoint decoder. Our Group Pose, which uses $K$ keypoint queries and one instance query for each human instance and considers different interactions + +Table 3: Comparisons with state-of-the-art methods on Crowd-Posetest dataset. Swin-L is adopted as the backbone. Other notations are consistent with Tabel 1. + +
MethodLossAP\( AP_{50} \)\( AP_{75} \)\( AP_E \)\( AP_M \)\( AP_H \)
PETR [27]HM+KR71.690.478.377.372.065.8
QueryPose [36]BR+RLE72.791.778.179.573.465.4
ED-Pose [38]BR+KR73.190.579.880.573.863.8
GroupPoseKR74.191.380.480.874.766.4
+ +among queries, can achieve reasonable good results with easy, medium, and hard crowding levels. + +Comparisons with various non-end-to-end frameworks on COCO. We also compare Group Pose with representative non-end-to-end frameworks in Table 1 and Table 2. Group Pose can easily outperform previous bottom-up methods [5, 7, 20, 37] and one-stage methods [31, 40, 22, 28]. For example, Group Pose surpasses the recent proposed LOGO-CAP [37] with flipping test by over 2.0 AP on both COCO val2017 and test-dev, even with a smaller backbone (ResNet-50 [9]) than it (HRNetw32 [29]). Group Pose is more concise and precise than those bottom-up and one-stage methods. Surprisingly, Group Pose can also beat previous top-down methods like PRTR [15] and Poseur [21]. Given the simple design and the end-to-end property, our Group Pose can serve as a simple baseline for pursuing higher performance on multi-person pose estimation. + +(a) Query designs for human instances. Both instance (inst) and keypoint (kpt) queries are essential in Group Pose, especially the keypoint ones. + +(b) Benefits of the instance query. On the model only with kpt queries, the inst query is first added but not for classification (cls). Then we use the inst query for the cls task. + +(c) Number of instances. When the number is smaller than 100, increasing instance numbers provide gains. 100 is set by default for the number of instances. + +Table 4: Ablation experiments for Group Pose. Evaluated on MS COCO val2017. Default settings are marked in gray . + +
query typesAPAPMAPL
inst & kpt72.066.879.7
only kpt71.266.079.1
only inst64.561.169.9
+ +
query typescls taskAPAPMAPL
only kptavg kpt71.266.079.1
inst & kptavg kpt71.766.879.3
inst & kptinst72.066.879.7
+ +
#instanceAPAPMAPL
5071.466.479.1
10072.066.879.7
20072.066.979.7
+ +(d) Self-attention implementations. In the self-attention module, removing the across-instance interactions for queries with different types with attention (attn) mask or group self-attention eases the optimization and improves the performance. + +
self-attention implementationsw/ attn maskAPAPMAPL
standard self-attention×69.464.077.5
standard self-attention✓70.765.578.4
group self-attention×72.066.879.7
+ +# 4.3. Ablation Study + +We run a number of ablation experiments to verify the effectiveness of key elements in our Group Pose. We adopt ResNet-50 [9] as the backbone. Unless specified, we report the results on COCO val2017 with 60 epochs training. + +Ablation: query designs for human instance. Table 4a gives the results of representing a human instance with different query designs. Both instance query and keypoint queries are essential in Group Pose. As the multi-person pose estimation is to predict human poses given an image, it is reasonable that keypoint queries themselves can achieve good results, while removing keypoint queries gives a significant performance drop (from 72.0 AP to 64.5 AP). + +For the instance query, its benefits come from two aspects: (i) gather information within human instance and help model training and (ii) decouple the classification task and keypoint positions regression task. Table 4b shows the improvements brought by these two benefits. + +Ablation: self-attention implementations. Table 4d provides the comparisons between different self-attention in the decoder layers, including (i) standard self-attention over all the $N \times (K + 1)$ queries (Figure 3 (a)), (ii) standard self-attention with an attention mask (Figure 3 (b)), which masks out the across-instance interactions for queries with different types, and (iii) our group self-attention (Figure 3 (c)). Results validate that the across-instance interactions for queries with different types are not directly helpful for the multi-person pose estimation task. Removing this type of interactions eases the model optimization, helps model converge faster, thus improves the performances. The comparisons on convergence curves are given in Figure 4. Besides, we find that it is better to separately perform different types of interactions, the within-instance and same-type across-instance interactions, with different parameters. Table 4d shows that our sequential implementation group self + +(e) Group self-attentions. Both types of self-attentions are important in group self-attentions. We also conduct an experiment by removing both within-instance and across-instance self-attentions. The performance further drops to 65.3 AP. + +
group self-attentionssAPAPMAPL
within-instance & across-instance self-attentionss72.066.879.7
within-instance self-attentionss66.363.870.8
across-instance self-attentionss67.463.074.4
+ +![](images/c13e0a49672a4448492403bfa91f66ad7518684121ae72b6a932ac09f40b3829.jpg) +Figure 4: Faster and better convergence. The $x$ -axis corresponds to #epoch, and the $y$ -axis corresponds to AP score. One can see that group self-attention converges faster and better than other self-attention implementations. + +attentions bring a 1.3 AP gain over the parallel implementation of standard self-attention with an attention mask. + +Ablation: group self-attentions. We ablate the effects of $N$ within-instance self-attentions over $K$ keypoint queries and one instance query and $(K + 1)$ across-instance self-attentions over $N$ queries of the same type in Table 4e. Large performance gaps (over $-5.0\mathrm{AP}$ ) are observed when we remove either of these two group self-attentions. Both the interactions are important for multi-person pose estimation. The within-instance self-attentions explore kinematic relations and gather information for scoring pose predictions, which help the model predict precise keypoint positions for human poses. The same-type across-instance self-attentions collect information from the same-type of keypoint queries or instance queries, removing duplicate predictions for poses among human instances. As shown in Figure 5, model without the same-type across-instance self-attentions produces more duplicated pose predictions for the same human instance than Group Pose. + +Ablation: number of instances. Table 4c gives how predefined number of human instances affect the results. We find 100 human instances are enough and can achieve comparable results with 200 human instances. We set the num + +![](images/e6cfa798a371f8ac372d2e5ac660f503b0db2e62b16a338e37339baa83c611c7.jpg) + +![](images/ae478084818c4857514ecf386cff4996c3c093570d8af5df33aebf1935235c22.jpg) +Figure 5: Comparison of removing duplicated predictions. We visualize the predicted poses of Group Pose (second row), and Group Pose without same-type across-instance self-attention (first row) according to the number of objects. Group Pose produces less duplicated pose predictions. The yellow dashed ellipse indicates the duplicated predicted human instance, leading to a mismatch of other instances. Best view in zoom in. + +![](images/2f5613600a75b8f370e97a191151358c33d1b25694901b57ed62d14014eacfe2.jpg) + +![](images/971f93e3d4e76b0bc9976722630cc3991ce7c80fc05d8fb004ff6f3c5b37ef49.jpg) + +![](images/0a767049aebf04137063dcf2ee5cc92a663c52149d4380d08c2fdc6ea14286d7.jpg) + +![](images/afb2ebd454ecf6bc8840ab8f47594b35f4d5d7338ecf7b89b12a07459fc6608d.jpg) + +![](images/8071d082320ae6a4252f76095b86c5a9f5f434f51e567b5714845b300c923a33.jpg) + +![](images/a4c51b53b30675e8f416da4963151e6ee570c60cd3cb2f3c218b6c36c1e5e0e2.jpg) + +![](images/706640c2966ce82b31341ba3e38be81d65d73ec2712e388a736441a29713af11.jpg) + +![](images/f19f5546c2f9314533c5db1e68c48d65b6f2a8b6dbe4361df0827695486b2341.jpg) + +Table 5: Analysis on model convergence. Group Pose without human detection already can outperform ED-Pose [38]. Besides, Group Pose can also benefit from better human instance initialization with human detection. We report AP on COCO val2017 dataset. 'Det Dec' = human detection decoder. The results of ED-Pose are from the original paper [38]. + +
Methodw/ Det Dec12e24e36e48e60e
ED-Pose [38]✓60.567.569.770.871.6
GroupPose×61.067.670.171.472.0
GroupPose✓61.468.170.371.672.2
+ +ber of human instances as 100 by default. Thus, Group Pose has $100 \times (17 + 1)$ queries on COCO [16] and contains $100 \times (14 + 1)$ queries on CrowdPose [14]. + +# 4.4. More Analysis + +We provide more analysis about our Group Pose in this section, including the analysis on model convergence and the analysis on model inference speed, detailed next. + +Analysis on model convergence. A common wisdom about training models is to provide a good initialization and then refine the model based on it. ED-Pose [38] splits the multi-person pose estimation task into two sub-processes, which first detect human instances with a human detection decoder and then use a human-to-keypoint detection decoder for predicting human poses. The human detection decoder gives good initialization for human instances, which can help ED-Pose learn and converge faster. Although our Group Pose adopts a different design for transformer decoder, it can also be built upon a human detection decoder and enjoys the benefits brought by better initialization of human instances. Based on this observation, we build a variant of our Group Pose, whose transformer decoder consists of 2 human detection decoder layers and 4 simple transformer decoder layers for multi-person pose estimation, following + +ED-Pose [38]. + +Table 6: Analysis on model inference speed. Frames per second (FPS) and inference time (Time) are measured with ResNet-50 and different image resolutions on one NVIDIA A100 GPU. + +
MethodInput ResolutionFPS ↑Time [ms] ↓
PETR [27]480 × 80020.050
800 × 133312.183
QueryPose [36]480 × 80019.056
800 × 133313.475
ED-Pose [38]480 × 80042.424
800 × 133324.740
GroupPose480 × 80068.615
800 × 133331.332
+ +Table 5 shows the AP scores of three models, ED-Pose, Group Pose, and Group Pose with human detection decoder, under different training schedules. Group Pose already shows superior results than ED-Pose with 12 epochs, 24 epochs, 36 epochs, 48 epochs, and 60 epochs training, even without the help of human detection decoder. Moreover, the comparison between Group Pose and Group Pose with human detection decoder verifies that good initialization for human instances leads to faster convergence and better results. + +Analysis on model inference speed. Table 6 provides the comparisons of inference time and FPS among end-to-end frameworks, including PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose. Simple designs usually show efficiency. Results show that our simple transformer decoder can be faster than complex decoders in previous end-to-end frameworks. Even with an image in a large size $800 \times 1333$ , our Group Pose can also achieve real-time speed (above 30 FPS) on a single A100 GPU. + +# 5. Conclusion + +In this paper, we present a simple baseline, Group Pose, for end-to-end multi-person pose estimation. With simple designs for queries and decoder self-attention, the approach outperforms previous end-to-end frameworks while being faster. Besides, the transformer decoder in Group Pose is also flexible, which can be built solely or upon human detection decoders. We hope Group Pose can provide insights for exploring concise and effective end-to-end multi-person pose estimation frameworks. + +Acknowledgements. This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.U1936212, No.62120106009, No.62261160653). + +# References + +[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 1, 2 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 2, 3, 4, 5 +[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2 +[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018. 1, 2 +[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5386-5395, 2020. 1, 2, 5, 6 +[6] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2334-2343, 2017. 1, 2 +[7] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 5, 6 +[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2961-2969, 2017. 2, 5, 6 + +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 3, 5, 6, 7 +[10] Jan Hosang, Rodrigo Benenson, and Bernt Schiele. Learning non-maximum suppression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4507-4515, 2017. 1, 2 +[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[12] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11977-11986, 2019. 1, 2 +[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2 +[14] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10863-10872, 2019. 2, 4, 5, 6, 8 +[15] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 5, 6 +[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 2, 4, 5, 8 +[17] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2 +[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5, 6 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[20] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 5, 6 +[21] Weian Mao, Yongtao Ge, Chunhua Shen, Zhi Tian, Xinlong Wang, Zhibin Wang, and Anton van den Hengel. Poseur: Direct human pose regression with transformers. In Proceedings of the European Conference on Computer Vision, pages 72-88, 2022. 5, 6 +[22] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation + +with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 1, 2, 5, 6 +[23] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 4 +[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in Neural Information Processing Systems, 30, 2017. 1, 2 +[25] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision, pages 483-499. Springer, 2016. 1, 2 +[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6951-6960, 2019. 1, 2 +[27] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022. 1, 2, 3, 5, 6, 8 +[28] Dahu Shi, Xing Wei, Xiaodong Yu, Wenming Tan, Ye Ren, and Shiliang Pu. Inspose: instance-aware networks for single-stage multi-person pose estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3079-3087, 2021. 5, 6 +[29] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019. 1, 2, 6 +[30] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2 +[31] Zhi Tian, Hao Chen, and Chunhua Shen. Directpose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 1, 2, 5, 6 +[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3 +[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 2 +[34] Fangyun Wei, Xiao Sun, Hongyang Li, Jingdong Wang, and Stephen Lin. Point-set anchors for object detection, instance segmentation and pose estimation. In Proceedings of the + +European Conference on Computer Vision, pages 527-544. Springer, 2020. 1, 2 +[35] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision, pages 466-481, 2018. 1, 2 +[36] Yabo Xiao, Kai Su, Xiaojuan Wang, Dongdong Yu, Lei Jin, Mingshu He, and Zehuan Yuan. Querypose: Sparse multiperson pose regression via spatial-aware part-level query. In Advances in Neural Information Processing Systems, pages 1-14, 2022. 2, 3, 5, 6, 8 +[37] Nan Xue, Tianfu Wu, Gui-Song Xia, and Liangpei Zhang. Learning local-global contextual adaptation for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13065-13074, 2022. 5, 6 +[38] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In International Conference on Learning Representations, pages 1-17, 2023. 1, 2, 3, 5, 6, 8 +[39] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, pages 1-18, 2022. 1, 2, 5 +[40] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 5, 6 +[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, pages 1-16, 2021. 1, 2, 3, 4, 5 \ No newline at end of file diff --git a/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/images.zip b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa786102943c3c54299097d4f48183c930fdf73e --- /dev/null +++ b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35b3f17ea3f2acb603da35fa35f053245e0d23d98cb9d81488a74e6a63916aa4 +size 651726 diff --git a/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/layout.json b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..66e3cb3d9a2d03ddc707a9cfc38d349b644da850 --- /dev/null +++ b/2023/Group Pose_ A Simple Baseline for End-to-End Multi-Person Pose Estimation/layout.json @@ -0,0 +1,9298 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "text", + "content": "Group Pose: A Simple Baseline for End-to-End Multi-person Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": "Huan Liu" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Qiang Chen" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Zichang Tan" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Jiang-Jiang Liu" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Jian Wang" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Xiangbo Su" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Xiaolong Li" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Kun Yao" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Junyu Han" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Errui Ding" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Yao Zhao" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "text", + "content": " Jingdong Wang" + }, + { + "bbox": [ + 75, + 142, + 518, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "type": "text", + "content": "Institute of Information Science, Beijing Jiaotong University " + }, + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "type": "text", + "content": "Baidu VIS " + }, + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 174, + 539, + 205 + ], + "type": "text", + "content": "Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 230, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 255, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 338 + ], + "type": "text", + "content": "In this paper, we study the problem of end-to-end multiperson pose estimation. State-of-the-art solutions adopt the DETR-like framework, and mainly develop the complex decoder, e.g., regarding pose estimation as keypoint box detection and combining with human detection in ED-Pose [38], hierarchically predicting with pose decoder and joint (keypoint) decoder in PETR [27]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "text", + "content": "We present a simple yet effective transformer approach, named Group Pose. We simply regard " + }, + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "text", + "content": "-keypoint pose estimation as predicting a set of " + }, + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "text", + "content": " keypoint positions, each from a keypoint query, as well as representing each pose with an instance query for scoring " + }, + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 339, + 287, + 399 + ], + "type": "text", + "content": " pose predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "text", + "content": "Motivated by the intuition that the interaction, among across-instance queries of different types, is not directly helpful, we make a simple modification to decoder self-attention. We replace single self-attention over all the " + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "inline_equation", + "content": "N \\times (K + 1)" + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "text", + "content": " queries with two subsequent group self-attention: (i) " + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "text", + "content": " within-instance self-attention, with each over " + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "text", + "content": " keypoint queries and one instance query, and (ii) " + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "text", + "content": " same-type across-instance self-attention, each over " + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 399, + 288, + 601 + ], + "type": "text", + "content": " queries of the same type. The resulting decoder removes the interaction among across-instance type-different queries, easing the optimization and thus improving the performance. Experimental results on MS COCO and CrowdPose show that our approach without human box supervision is superior to previous methods with complex decoders, and even is slightly better than ED-Pose that uses human box supervision. Paddle1 and PyTorch2 codes are available." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 624, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 643, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 287, + 668 + ], + "type": "text", + "content": "Multi-person pose estimation aims to detect the corresponding human keypoints for all human instances in an" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 228, + 536, + 253 + ], + "blocks": [ + { + "bbox": [ + 312, + 228, + 536, + 253 + ], + "lines": [ + { + "bbox": [ + 312, + 228, + 536, + 253 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 536, + 253 + ], + "type": "image", + "image_path": "437a2e132d7408f4d6ab28929d8ba413a7667cc3aa9779c1387b4dd615c4c1bb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 253, + 443, + 262 + ], + "lines": [ + { + "bbox": [ + 413, + 253, + 443, + 262 + ], + "spans": [ + { + "bbox": [ + 413, + 253, + 443, + 262 + ], + "type": "text", + "content": "(a) PETR" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 264, + 535, + 291 + ], + "blocks": [ + { + "bbox": [ + 312, + 264, + 535, + 291 + ], + "lines": [ + { + "bbox": [ + 312, + 264, + 535, + 291 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 535, + 291 + ], + "type": "image", + "image_path": "4ddb363828897d62896c274eff862023799eacec0f75c98e3634b7eb5e5fd16d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 312, + 304, + 534, + 331 + ], + "blocks": [ + { + "bbox": [ + 410, + 293, + 447, + 301 + ], + "lines": [ + { + "bbox": [ + 410, + 293, + 447, + 301 + ], + "spans": [ + { + "bbox": [ + 410, + 293, + 447, + 301 + ], + "type": "text", + "content": "(b) ED-Pose" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 304, + 534, + 331 + ], + "lines": [ + { + "bbox": [ + 312, + 304, + 534, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 534, + 331 + ], + "type": "image", + "image_path": "73df08414980fcafaa87f02a383e1d47af7fa6b19897415042a65ab8f4ec3e3f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 334, + 451, + 342 + ], + "lines": [ + { + "bbox": [ + 406, + 334, + 451, + 342 + ], + "spans": [ + { + "bbox": [ + 406, + 334, + 451, + 342 + ], + "type": "text", + "content": "(c) Group Pose" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 352, + 547, + 419 + ], + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 419 + ], + "type": "text", + "content": "Figure 1: Comparison of transformer decoders. Here we mainly illustrate the overview of the decoder part of PETR [27], ED-Pose [38], and our Group Pose. The three end-to-end frameworks differ in query design and decoder architecture. Group Pose only uses a simple transformer decoder rather than developing complex decoders. 'inst' and 'kpt' represent for instance and keypoint." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 435, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 591 + ], + "type": "text", + "content": "image. Previous frameworks include top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5] that divide the task into two sequential sub-tasks: human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping [24]. Another line in previous frameworks is one-stage methods [22, 26, 31, 34], which directly predict instance-aware keypoints. These frameworks rely on non-differentiable hand-crafted post-processes [10, 24], which complicate the pipelines and challenge the optimizations. Inspired by the success of DETR [2] in object detection, building an end-to-end framework for multi-person pose estimation has seen significant interest." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 715 + ], + "type": "text", + "content": "Recent approaches follow the DETR framework [2, 41, 39], with the transformer encoder-decoder architecture for multi-person pose estimation, as shown in Figure 1. PETR [27] hierarchically predicts the keypoint positions and uses two subsequent decoders, pose decoder and joint decoder, with two different queries, pose query (a person has one pose query) for pose decoder, and keypoint queries for joint decoder. ED-Pose [38] transfers pose estimation to a keypoint box detection problem, and learns a content query and a box query for each keypoint position prediction" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 673, + 276, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 673, + 276, + 683 + ], + "spans": [ + { + "bbox": [ + 57, + 673, + 276, + 683 + ], + "type": "text", + "content": "*Equal Contribution. Work done when H. Liu is an intern at Baidu." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 60, + 683, + 223, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 683, + 223, + 693 + ], + "spans": [ + { + "bbox": [ + 60, + 683, + 223, + 693 + ], + "type": "text", + "content": "†Corresponding author. Email: yzhao@bjtu.edu.cn" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 60, + 693, + 274, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 693, + 274, + 703 + ], + "spans": [ + { + "bbox": [ + 60, + 693, + 274, + 703 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 693, + 274, + 703 + ], + "type": "text", + "content": "Code of Paddle: https://github.com/Michel-liu/GroupPose-Paddle" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 703, + 256, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 256, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 256, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 59, + 703, + 256, + 713 + ], + "type": "text", + "content": "Code of PyTorch: https://github.com/Michel-liu/GroupPose" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15029" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 228, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 228, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 228, + 85 + ], + "type": "text", + "content": "with using the box size to process the query." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "text", + "content": "In this paper, we present a simple yet effective transformer approach, named Group Pose, for end-to-end multiperson pose estimation. Instead of using a single query to predict and score one pose, similar to ED-Pose, we use the " + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "text", + "content": " keypoint queries to regress the " + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "text", + "content": " positions, by regarding each keypoint as an object, as well as " + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "text", + "content": " instance queries, each representing a " + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "text", + "content": "-keypoint pose for scoring the " + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 93, + 287, + 189 + ], + "type": "text", + "content": "-keypoint pose prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "text", + "content": "We make a simple modification for the decoder architecture. We replace standard self-attention in the decoder with two subsequent group self-attention: " + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "text", + "content": " parallel self-attention with each over " + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "text", + "content": " keypoint queries and the corresponding instance query for exploiting kinematic relation and gathering information for scoring pose predictions, and " + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "text", + "content": " parallel self-attention with each over " + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 198, + 288, + 305 + ], + "type": "text", + "content": " queries of the same type for collecting duplicate prediction information like self-attention of the original DETR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "text", + "content": "The two self-attention captures two kinds of interactions: (i) " + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "text", + "content": " within-instance interactions over " + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "text", + "content": " keypoint queries and one instance query, (ii) " + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "text", + "content": " across-instance interactions over " + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 313, + 288, + 433 + ], + "type": "text", + "content": " queries of the same type (e.g., nose keypoint query or instance query). The extra interactions in standard self-attention are (iii) across-instance interactions for queries with different types, which is not directly useful. Empirical results show that the removal of the third kind of interactions eases the optimization and thus improves the performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 441, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 597 + ], + "type": "text", + "content": "The design about self-attention is different from the closely-related approach ED-Pose [38]. On the one hand, in addition to removing the third interactions, ED-Pose only models across-instance interactions for the instance queries. In contrast, our approach also models across-instance interactions for the " + }, + { + "bbox": [ + 47, + 441, + 287, + 597 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 441, + 287, + 597 + ], + "type": "text", + "content": " keypoint types. On the other hand, our approach separates the two interactions using two subsequent group self-attention, explicitly exploring the information about queries belonging to the same human instance, and with the same type. ED-Pose couples the two interactions using a single masked self-attention with the third interactions masked. It is empirically demonstrated that the two differences benefit the pose estimation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 715 + ], + "type": "text", + "content": "Experimental results show that our simple approach Group Pose without human box supervision surprisingly outperforms the recent end-to-end methods with human box supervisions on MS COCO [16] and CrowdPose [14]. Notably, Group Pose achieves 72.0 AP with ResNet-50 [9] and 74.8 AP with Swin-Large [18] on MS COCO val2017. We hope our simple transformer decoder in Group Pose will motivate people to simplify the design in end-to-end multiperson pose estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 95, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 95, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 95, + 547, + 156 + ], + "type": "text", + "content": "Multi-person pose estimation is a challenging task that aims to detect the corresponding human keypoints for all human instances in an image. Previous methods usually adopt complex frameworks to address it, which are divided into non-end-to-end and end-to-end methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 163, + 546, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 163, + 546, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 546, + 368 + ], + "type": "text", + "content": "Non-end-to-end methods. There are typically two types: two-stage and one-stage. Two-stage frameworks, including top-down [25, 6, 35, 4, 29, 33] and bottom-up methods [1, 24, 12, 5], split the multi-person pose estimation task into two sequential sub-tasks, human detection with single-person pose estimation or human-agnostic keypoint detection with human instance grouping. In top-down methods, an object detector is first employed to detect the boxes of human instances, which is then cropped for single-person pose estimation in each box. Bottom-up methods first predict all human keypoints in a human-agnostic way then group them into instances. While one-stage frameworks [40, 22, 26, 31, 34] directly predict instance-aware keypoints. These methods require hand-crafted pose-processes, such as NMS [10] or grouping [24], which complicate the pipelines and challenge the optimizations. In this paper, we focus on concise ways, end-to-end frameworks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 373, + 546, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 546, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 546, + 650 + ], + "type": "text", + "content": "End-to-end methods. Current end-to-end multi-person pose estimation frameworks [27, 36, 38] are built by following the designs of DETR [2] and its variants [41, 23, 3, 17, 13, 39]. They adopt a paradigm of splitting the multiperson pose estimation task into two sub-processes. For example, PETR [27] views the task as a hierarchical set prediction problem. It first determines human instances by predicting a set of poses with a pose decoder and then refines the keypoints in each pose with a joint (keypoint) decoder. There are also two different types of queries, pose query (a person has one pose query) for pose decoder and keypoint queries for joint decoder. QueryPose [36] and ED-Pose [38] follow this end-to-end paradigm but further incorporate an extra human detection task. QueryPose [36] follows Sparse R-CNN [30] to build two parallel RoIAlign-based [8] decoders to perform human detection and pose estimation, respectively. ED-Pose [38] transfers pose estimation task to a keypoint box detection problem. It first employs a human detection decoder [39] to determine the human instances with box queries. Then it builds a human-to-keypoint detection decoder with a content query and a box query for each keypoint position, collecting contextual information near keypoint positions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "type": "text", + "content": "Although these end-to-end methods show promising results in multi-person pose estimation, they rely on complex decoders. Our Group Pose, on the other hand, adopts a simple transformer decoder, improving the performance and simplifying the process." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15030" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 169 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 169 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 169 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 169 + ], + "type": "image", + "image_path": "27218fbf40913cf0091d027d05ad545804fd92c209cec145c29f801bee2eff7a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "text", + "content": "Figure 2: Our Group Pose architecture. The backbone takes an image as input and outputs image features, which are refined by the transformer encoder. To directly predict " + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "text", + "content": " human poses with " + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "text", + "content": " keypoint positions in each pose, we adopt " + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "inline_equation", + "content": "N \\times (K + 1)" + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "text", + "content": " queries, containing " + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "text", + "content": " instance queries for scoring poses and " + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 46, + 178, + 544, + 212 + ], + "type": "text", + "content": " keypoint queries for regressing positions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 122, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 122, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 122, + 230 + ], + "type": "text", + "content": "3. Group Pose" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 286, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 286, + 320 + ], + "type": "text", + "content": "Group Pose is a simple yet effective end-to-end multi-person pose estimation framework. We follow previous end-to-end frameworks [27, 38] to view the multi-person pose estimation task as a set prediction problem, but directly adopt a simple transformer decoder [41] instead of complex decoders, simplifying the process. Next, we introduce the key elements of Group Pose." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 329, + 114, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 114, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 114, + 341 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "text", + "content": "The overall structure of Group Pose is depicted in Figure 2. Group Pose consists of a backbone [9, 18], a transformer encoder [32], a transformer decoder, and task-specific prediction heads. This framework enables Group Pose to simultaneously regresses " + }, + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "text", + "content": " keypoints (e.g., " + }, + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "text", + "content": " on MS COCO) for " + }, + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 348, + 286, + 420 + ], + "type": "text", + "content": " human instances given an image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 423, + 286, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 286, + 508 + ], + "type": "text", + "content": "Backbone and transformer encoder. We directly follow DETR frameworks [41] to build the backbone and the transformer encoder (with 6 deformable transformer layers [41]) for Group Pose. It takes an image as input and outputs the extracted multi-level features, which serve as inputs for the following transformer decoder. We use 4 feature levels in Group Pose, with downsampling rates of " + }, + { + "bbox": [ + 46, + 423, + 286, + 508 + ], + "type": "inline_equation", + "content": "\\{8, 16, 32, 64\\}" + }, + { + "bbox": [ + 46, + 423, + 286, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": "Transformer decoder. For transformer decoder, we adopt a combination of " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " keypoint queries and " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " instance queries as input instead of using a single query to predict and score one pose. The keypoint queries regard each keypoint as an object and are used to regress the " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " keypoint positions, while each instance query is for scoring the corresponding " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": "-keypoint pose prediction. Besides, the architecture of transformer decoder is simple, which stacks 6 same decoder layers [41]. In each decoder layer, we follow the macro design of previous DETR frameworks by building self-attention, cross-attention implemented with deformable attention, and FFN. We only make simple modifications to self-attention in our Group Pose. Specifically, we replace the standard self-attention with two subsequent group self-attention, enabling decoder layers to perform interactions over queries belonging to the same human instance and with the same type." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "text", + "content": "Prediction heads. There are two prediction heads implemented with FFNs in Group Pose for human classification and human keypoints regression. Group Pose predicts for " + }, + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "text", + "content": " human poses, each contains a classification score and " + }, + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "text", + "content": " keypoints positions for the corresponding " + }, + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 217, + 544, + 278 + ], + "type": "text", + "content": "-keypoint pose." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "text", + "content": "Loss function. The Hungarian matching algorithm [2] is employed for one-to-one assignment between predicted poses and ground-truth poses. Our loss function comprises solely of classification loss " + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{cls})" + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "text", + "content": " and keypoint regression loss " + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{kpt})" + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "text", + "content": ", without any extra supervisions such as human detection loss in QueryPose [36] and ED-Pose [38] or heatmap loss in PETR [27]. The keypoint regression loss " + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{kpt})" + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "text", + "content": " is a combination of a normal " + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "text", + "content": " loss and a constrained " + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 305, + 282, + 544, + 425 + ], + "type": "text", + "content": " loss named Object Keypoint Similarity (OKS) [27]. We directly use the cost coefficients and loss weights of ED-Pose [38] in the hungarian matching and the calculation of losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 437, + 544, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 544, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 544, + 451 + ], + "type": "text", + "content": "3.2. " + }, + { + "bbox": [ + 306, + 437, + 544, + 451 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 306, + 437, + 544, + 451 + ], + "type": "text", + "content": " keypoint queries and " + }, + { + "bbox": [ + 306, + 437, + 544, + 451 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 437, + 544, + 451 + ], + "type": "text", + "content": " instance queries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": "In multi-person pose estimation, frameworks are required to predict " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": " human poses with " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": " keypoint positions in each pose given an image. We directly use " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": " keypoints queries to predict the poses, with each " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": " predicted keypoint positions to represent a corresponding " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": "-keypoint pose for a human instance. To classify if it is a human instance by scoring the predicted " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": " human poses, we also introduce " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": " instance queries. Thus, in Group Pose, a human instance can be represented with a combination of " + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 458, + 544, + 601 + ], + "type": "text", + "content": " keypoint queries and one instance query. As these two types of queries are responsible for different tasks, we construct and initialize them differently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": "Query construction. We follow the previous end-to-end frameworks [27, 38] to first identify human instances and predict human poses in each position of the output memory from the transformer encoder. Then we select " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": " human instances (" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": ") based on the classification scores, resulting " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": " human poses and the corresponding output memory features of the selected " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": " positions (with the shape of " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "N \\times D" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": " is the channel dimension). We then construct and initialize the keypoint queries and the instance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15031" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 68, + 206, + 166 + ], + "blocks": [ + { + "bbox": [ + 97, + 68, + 206, + 166 + ], + "lines": [ + { + "bbox": [ + 97, + 68, + 206, + 166 + ], + "spans": [ + { + "bbox": [ + 97, + 68, + 206, + 166 + ], + "type": "image", + "image_path": "92ed9b5f62fd89906b8ec4e8043e59b1d924df628c3bb2ae823ad810a7170d57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 167, + 202, + 176 + ], + "lines": [ + { + "bbox": [ + 108, + 167, + 202, + 176 + ], + "spans": [ + { + "bbox": [ + 108, + 167, + 202, + 176 + ], + "type": "inline_equation", + "content": "(N\\times (K + 1))\\times (N\\times (K + 1))" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 177, + 195, + 185 + ], + "lines": [ + { + "bbox": [ + 116, + 177, + 195, + 185 + ], + "spans": [ + { + "bbox": [ + 116, + 177, + 195, + 185 + ], + "type": "text", + "content": "(a) standard self-attention" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 68, + 331, + 166 + ], + "blocks": [ + { + "bbox": [ + 222, + 68, + 331, + 166 + ], + "lines": [ + { + "bbox": [ + 222, + 68, + 331, + 166 + ], + "spans": [ + { + "bbox": [ + 222, + 68, + 331, + 166 + ], + "type": "image", + "image_path": "7adcb61177411121267fa4a2e6d0d63cca5e8d2983a91ee894b32ed1c50c216f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 167, + 328, + 176 + ], + "lines": [ + { + "bbox": [ + 234, + 167, + 328, + 176 + ], + "spans": [ + { + "bbox": [ + 234, + 167, + 328, + 176 + ], + "type": "inline_equation", + "content": "(N\\times (K + 1))\\times (N\\times (K + 1))" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 177, + 337, + 185 + ], + "lines": [ + { + "bbox": [ + 225, + 177, + 337, + 185 + ], + "spans": [ + { + "bbox": [ + 225, + 177, + 337, + 185 + ], + "type": "text", + "content": "(b) standard self-attention with mask" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 194, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 237 + ], + "type": "text", + "content": "Figure 3: Conceptual comparison of self-attention implementations. Given " + }, + { + "bbox": [ + 46, + 194, + 547, + 237 + ], + "type": "inline_equation", + "content": "N \\times (K + 1)" + }, + { + "bbox": [ + 46, + 194, + 547, + 237 + ], + "type": "text", + "content": " queries, three different implementations of self-attention are considered in the decoder layers. (a) standard self-attention over all queries. (b) standard self-attention with an attention mask, removing the across-instance interactions for queries with different types. (c) our proposed group self-attention. Dashed boxes indicate the masked-out interactions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 68, + 498, + 166 + ], + "blocks": [ + { + "bbox": [ + 350, + 68, + 498, + 166 + ], + "lines": [ + { + "bbox": [ + 350, + 68, + 498, + 166 + ], + "spans": [ + { + "bbox": [ + 350, + 68, + 498, + 166 + ], + "type": "image", + "image_path": "24322e3e8f6c28a6ebdf809954965d38494d3cd194fa52f1b16cc4ba454b5b7e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 167, + 495, + 176 + ], + "lines": [ + { + "bbox": [ + 347, + 167, + 495, + 176 + ], + "spans": [ + { + "bbox": [ + 347, + 167, + 495, + 176 + ], + "type": "inline_equation", + "content": "(N\\times ((K + 1)\\times (K + 1)) + ((K + 1)\\times (N\\times N))" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 177, + 457, + 185 + ], + "lines": [ + { + "bbox": [ + 385, + 177, + 457, + 185 + ], + "spans": [ + { + "bbox": [ + 385, + 177, + 457, + 185 + ], + "type": "text", + "content": "(c) group self-attention" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 246, + 188, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 188, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 188, + 257 + ], + "type": "text", + "content": "queries based on the above results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " keypoint queries, we construct the content part of " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " keypoint queries " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "(K \\times D)" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " in each human instance by combining " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " randomly initialized learnable keypoint embeddings " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "(K \\times D)" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " and the corresponding output memory feature " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "(1 \\times D)" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": ". And the position part of " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " keypoint queries " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "(K \\times 2)" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " in each human instance is initialized with the corresponding predicted " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": "-keypoint pose " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "(K \\times 2)" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " instance queries, we only consider the content part and use a randomly initialized learnable instance embedding " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "(1 \\times D)" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " for each human instance. This is because instance query is for the classification task, which does not require explicit position information. When performing cross-attention in the decoder layers, we simply use the mean of " + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 259, + 290, + 437 + ], + "type": "text", + "content": " keypoint positions as the reference point for the instance query in each human instance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 449, + 169, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 169, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 169, + 462 + ], + "type": "text", + "content": "3.3. Group self-attention" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": "In DETR frameworks [2, 41, 23], self-attention in the transformer decoder is usually applied to model interactions among queries, collecting information from other queries and facilitating the duplicate removal for instances. While the situation is slightly different in Group Pose, as there two different types of queries, " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": " keypoint queries and one instance query for each human instance. Motivated by the intuition that the interactions among across-instance queries of different types, e.g., the interactions between the instance query and the keypoint queries across human instances, may not be directly helpful for the above purposes. We thus replace the standard self-attention over all the " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "N \\times (K + 1)" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": " queries with two subsequent group self-attention: " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": " within-instance self-attention and " + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 46, + 470, + 287, + 649 + ], + "type": "text", + "content": " same-type across-instance self-attention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " within-instance self-attention. We build interactions among the queries within a human instance, exploiting kinematic relations and gathering information for scoring pose predictions. It calculates " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " self-attention maps for " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " human instances in parallel with shapes of " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "(K + 1)\\times (K + 1)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 245, + 539, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 539, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 539, + 257 + ], + "type": "text", + "content": "omitting the dimensions of batch size and attention heads." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "text", + "content": " same-type across-instance self-attention. Similar to the within-instance group self-attention, We add another group self-attention, which collect information from the same-type queries in other instances and help remove duplicate predictions. We build interactions across human instances with the same-type queries, the instance query and each keypoint queries, resulting in " + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "text", + "content": " same-type across-instance self-attention. It also can be implemented in parallel with one self-attention module, calculating " + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "text", + "content": " self-attention maps with shapes of " + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 304, + 261, + 545, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 381, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 546, + 477 + ], + "type": "text", + "content": "Compared with the standard self-attention, our two subsequent group self-attention explicitly explores the information about queries belong to the same human instance and with the same type, while remove the across-instance interactions for queries with different types. Empirical results in Section 4.3 show that the removal of this kind of interactions eases the optimization and thus improves the performance for our Group Pose." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 491, + 387, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 387, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 387, + 504 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 511, + 365, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 365, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 365, + 523 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 530, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 685 + ], + "type": "text", + "content": "Datasets. Our experiments are conducted on two representative human pose estimation datasets, MS COCO [16] and CrowdPose [14]. MS COCO contains 200K images and 250K person instances with 17 keypoint annotations per instance (we set the number of keypoint queries as " + }, + { + "bbox": [ + 304, + 530, + 545, + 685 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 304, + 530, + 545, + 685 + ], + "type": "text", + "content": " on MS COCO). CrowdPose has 20K images and 80K person instances with 14 keypoint annotations per instance (" + }, + { + "bbox": [ + 304, + 530, + 545, + 685 + ], + "type": "inline_equation", + "content": "K = 14" + }, + { + "bbox": [ + 304, + 530, + 545, + 685 + ], + "type": "text", + "content": " on CrowdPose). CrowdPose is more challenging as it includes many crowd and occlusion scenes. We train Group Pose on COCO train2017 set and evaluate it on COCO val2017 set and test-dev set. On CrowdPose, we train our model on the train set and evaluate it on the test set." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Evaluation metric. The OKS-based average precision (AP) scores are reported as the main metric for both" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15032" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 123, + 545, + 401 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 117 + ], + "type": "text", + "content": "Table 1: Comparisons with state-of-the-art methods on MS COCO va12017. We also provide the reference (Ref) for previous frameworks. The 'HM', 'BR', and 'KR' denote heatmap-based losses, human box regression losses, and keypoint regression losses. 'RLE' represents the residual log-likelihood estimation in Poseur [21]. † denotes the flipping test. ‡ removes the prediction uncertainty estimation in Poseur as a fair regression comparison. The best results are highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 123, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 47, + 123, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 545, + 401 + ], + "type": "table", + "html": "
MethodRefBackboneLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_M \\)\\( AP_L \\)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM65.587.271.161.373.4
Mask R-CNN [8]CVPR 17ResNet-101HM66.187.472.061.574.4
\\( PRTR^† \\)[15]CVPR 21ResNet-50KR68.288.275.263.276.2
\\( Poseur^‡ \\)[21]ECCV 22ResNet-50RLE70.0----
Poseur [21]ECCV 22ResNet-50RLE74.289.881.371.180.1
Bottom-Up\\( HrHRNet^† \\)[5]CVPR 20HRNet-w32HM67.186.273.061.576.1
\\( DEKR^† \\)[7]CVPR 21HRNet-w32HM68.086.774.562.177.7
\\( SWAHR^† \\)[20]CVPR 21HRNet-w32HM68.987.874.963.077.4
\\( LOGO-CAP^† \\)[37]CVPR 22HRNet-w32HM69.687.575.964.178.0
One-StageDirectPose [31]-ResNet-50KR63.185.668.857.771.3
\\( CenterNet^† \\)[40]-Hourglass-104KR+HM64.0----
FCPose [22]CVPR 21ResNet-50KR+HM63.085.968.959.170.3
InsPose [28]ACM MM 21ResNet-50KR+HM63.186.268.558.570.1
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR68.887.576.362.777.7
PETR [27]CVPR 22Swin-LHM+KR73.190.780.967.281.7
QueryPose [36]NeurIPS 22ResNet-50BR+RLE68.788.674.463.876.5
QueryPose [36]NeurIPS 22Swin-LBR+RLE73.391.379.568.581.2
ED-Pose [38]ICLR 23ResNet-50BR+KR71.689.678.165.979.8
ED-Pose [38]ICLR 23Swin-LBR+KR74.391.581.668.682.6
OursGroupPose-ResNet-50KR72.089.479.166.879.7
GroupPose-Swin-TKR73.690.480.568.781.2
GroupPose-Swin-LKR74.891.682.169.483.0
", + "image_path": "6b392ef3f83010271628be205771e16ac8641fdf13053b9657e1cee680dd06e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": "datasets. For MS COCO, we adopt AP with different thresholds and different object sizes (medium and large), denoted as AP, " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_M" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_L" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": ", following the standard evaluation process3. On CrowdPose, to better evaluate the model performance in different crowded scenarios, we adopt AP with different thresholds and different crowding levels, denoted as AP, " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_E" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_M" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_H" + }, + { + "bbox": [ + 46, + 411, + 288, + 519 + ], + "type": "text", + "content": " for images with easy, medium and hard crowding levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "text", + "content": "Implementation details. Our training and testing settings follow ED-Pose [38]. During training, we adopt the widely-used data augmentations in DETR frameworks [2, 41, 39, 38], including random flip, random crop, and random resize with the short sides in [480, 800] and the long side less or equal to 1333. We use the AdamW optimizer[11, 19] with the weight decay " + }, + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "text", + "content": " and train 60 epochs and 80 epochs on MS COCO [16] and CrowdPose [14], respectively. We adopt a total batch size of 16, and set the base learning rate as " + }, + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "text", + "content": ". The base learning rate for the backbone is " + }, + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 523, + 288, + 693 + ], + "type": "text", + "content": " following the DETR frameworks. The learning rates are decayed at the 50-th epoch and 70-th by a factor of 0.1 for MS COCO and CrowdPose, respectively. During testing, we resize the input images with their short" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 411, + 518, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 518, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 518, + 424 + ], + "type": "text", + "content": "sides being 800 and long sides less or equal to 1333." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 436, + 391, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 391, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 391, + 448 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 456, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 552 + ], + "type": "text", + "content": "Our purpose is to build a simple baseline for end-to-end multi-person pose estimation. Thus, we mainly compare our Group Pose with previous end-to-end frameworks, including PETR [27], QueryPose [36], and ED-Pose [38]. Besides, to show the effectiveness of our method, we also add comparisons with non-end-to-end frameworks, such as top-down [8, 15], bottom-up [5, 7, 20, 37], and one-stage methods [31, 40, 22, 28]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 556, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 617 + ], + "type": "text", + "content": "Comparisons with end-to-end frameworks on COCO. Table 1 and Table 2 present the comparisons on COCO val2017 set and test-dev set. Results show that Group Pose outperforms PETR [27], QueryPose [36], and ED-Pose [38] consistently." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "On COCO val2017, Group Pose surpasses PETR and QueryPose by over a significant 3.0 AP with ResNet-50 [9] and the gaps remain " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.5 + \\mathrm{AP}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " with a strong backbone, SwinLarge [18]. When comparing with the recently proposed ED-Pose [38], which transfers pose estimation to a keypoint box detection problem and combines human box detection task, Group Pose can also exceed it with non-negligible margins. Moreover, unlike these methods that use complex" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 189, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 189, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 189, + 714 + ], + "type": "text", + "content": "3https://cocodataset.org/#keypoints-eval" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15033" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 91, + 541, + 344 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 542, + 81 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 542, + 81 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 542, + 81 + ], + "type": "text", + "content": "Table 2: Comparisons with state-of-the-art methods on MS COCO test-dev2017 dataset. Notations are consistent with Table 1." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 91, + 541, + 344 + ], + "lines": [ + { + "bbox": [ + 51, + 91, + 541, + 344 + ], + "spans": [ + { + "bbox": [ + 51, + 91, + 541, + 344 + ], + "type": "table", + "html": "
MethodRefBackboneLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_M \\)\\( AP_L \\)
Non-End-to-EndTop-DownMask R-CNN [8]CVPR 17ResNet-50HM63.987.769.959.771.5
Mask R-CNN [8]CVPR 17ResNet-101HM64.388.270.660.171.9
\\( PRTR^† \\)[15]CVPR 21ResNet-101KR68.889.976.964.775.8
\\( PRTR^† \\)[15]CVPR 21HRNet-w32KR71.790.679.667.678.4
Bottom-Up\\( HrHRNet^† \\)[5]CVPR 20HRNet-w32HM66.487.572.861.274.2
\\( DEKR^† \\)[7]CVPR 21HRNet-w32HM67.387.974.161.576.1
\\( SWAHR^† \\)[20]CVPR 21HRNet-w32HM67.988.974.562.475.5
\\( LOGO-CAP^† \\)[37]CVPR 22HRNet-w32HM68.288.774.962.876.0
One-StageDirectPose [31]-ResNet-50KR62.286.468.256.769.8
\\( CenterNet^† \\)[40]-Hourglass-104KR+HM63.086.869.658.970.4
FCPose [22]CVPR 21ResNet-50KR+HM64.387.371.061.670.5
InsPose [28]ACM MM 21ResNet-50KR+HM65.488.971.760.272.7
End-to-EndPrevious WorksPETR [27]CVPR 22ResNet-50HM+KR67.689.875.361.676.0
PETR [27]CVPR 22Swin-LHM+KR70.591.578.765.278.0
QueryPose [36]NeurIPS 22Swin-LBR+RLE72.292.078.867.379.4
ED-Pose [38]ICLR 23ResNet-50BR+KR69.890.277.264.377.4
ED-Pose [38]ICLR 23Swin-LBR+KR72.792.380.967.680.0
OursGroupPose-ResNet-50KR70.290.577.864.778.0
GroupPose-Swin-TKR72.191.479.966.779.5
GroupPose-Swin-LKR72.892.581.067.780.3
", + "image_path": "fa909cecbdf25594fa60d16f4da10cbb595496d5cccb5456c535807c67757aaf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 355, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 415 + ], + "type": "text", + "content": "decoders and add extra supervisions, e.g., extra heatmap or box supervisions, our Group Pose only use a simple decoder and are trained only with keypoint regression targets. The above evidences indicate that complex design for end-to-end multi-person pose estimation may not be necessary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 477 + ], + "type": "text", + "content": "On COCO test-dev, our Group Pose achieves 70.2 AP, 72.1 AP, and 72.8 AP with ResNet-50 [9], Swin-Tiny [18], and Swin-Large [18] as the backbone. Compared with other end-to-end frameworks, similar trends are observed with the ones on COCO val2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 483, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 579 + ], + "type": "text", + "content": "Comparisons with end-to-end frameworks on CrowdPose. To further demonstrate the effectiveness of Group Pose, we provide comparisons with previous end-to-end frameworks on the challenging CrowdPose dataset [14]. Table 3 reports the results of PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose with the same backbone Swin-Large [18]. Overall, Group Pose achieve 74.1 AP, performing the best over all methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Moreover, it is interesting when we compare the AP scores with easy, medium, and hard crowding levels of different models. PETR [27] performs worse with the easy crowding level while giving the second best results with the hard level. ED-Pose [38] performs worst with the hard crowding level. We conjecture that this phenomenon is caused by the differences in their decoders, e.g., human instances crowded with each other have similar boxes, which challenges the human keypoint decoder. Our Group Pose, which uses " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " keypoint queries and one instance query for each human instance and considers different interactions" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 396, + 545, + 457 + ], + "blocks": [ + { + "bbox": [ + 305, + 354, + 545, + 386 + ], + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 386 + ], + "type": "text", + "content": "Table 3: Comparisons with state-of-the-art methods on Crowd-Posetest dataset. Swin-L is adopted as the backbone. Other notations are consistent with Tabel 1." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 396, + 545, + 457 + ], + "lines": [ + { + "bbox": [ + 308, + 396, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 545, + 457 + ], + "type": "table", + "html": "
MethodLossAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_E \\)\\( AP_M \\)\\( AP_H \\)
PETR [27]HM+KR71.690.478.377.372.065.8
QueryPose [36]BR+RLE72.791.778.179.573.465.4
ED-Pose [38]BR+KR73.190.579.880.573.863.8
GroupPoseKR74.191.380.480.874.766.4
", + "image_path": "93da4c097e7113641fa6b5bb641c4e74138829a42ba2390c7979fafdc73f00b8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 487, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 545, + 511 + ], + "type": "text", + "content": "among queries, can achieve reasonable good results with easy, medium, and hard crowding levels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Comparisons with various non-end-to-end frameworks on COCO. We also compare Group Pose with representative non-end-to-end frameworks in Table 1 and Table 2. Group Pose can easily outperform previous bottom-up methods [5, 7, 20, 37] and one-stage methods [31, 40, 22, 28]. For example, Group Pose surpasses the recent proposed LOGO-CAP [37] with flipping test by over 2.0 AP on both COCO val2017 and test-dev, even with a smaller backbone (ResNet-50 [9]) than it (HRNetw32 [29]). Group Pose is more concise and precise than those bottom-up and one-stage methods. Surprisingly, Group Pose can also beat previous top-down methods like PRTR [15] and Poseur [21]. Given the simple design and the end-to-end property, our Group Pose can serve as a simple baseline for pursuing higher performance on multi-person pose estimation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15034" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 87, + 200, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 87, + 200, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 87, + 200, + 125 + ], + "type": "text", + "content": "(a) Query designs for human instances. Both instance (inst) and keypoint (kpt) queries are essential in Group Pose, especially the keypoint ones." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 208, + 87, + 385, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 87, + 385, + 125 + ], + "spans": [ + { + "bbox": [ + 208, + 87, + 385, + 125 + ], + "type": "text", + "content": "(b) Benefits of the instance query. On the model only with kpt queries, the inst query is first added but not for classification (cls). Then we use the inst query for the cls task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 393, + 87, + 545, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 87, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 393, + 87, + 545, + 125 + ], + "type": "text", + "content": "(c) Number of instances. When the number is smaller than 100, increasing instance numbers provide gains. 100 is set by default for the number of instances." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 60, + 130, + 187, + 171 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 522, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 522, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 522, + 82 + ], + "type": "text", + "content": "Table 4: Ablation experiments for Group Pose. Evaluated on MS COCO val2017. Default settings are marked in gray ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 130, + 187, + 171 + ], + "lines": [ + { + "bbox": [ + 60, + 130, + 187, + 171 + ], + "spans": [ + { + "bbox": [ + 60, + 130, + 187, + 171 + ], + "type": "table", + "html": "
query typesAPAPMAPL
inst & kpt72.066.879.7
only kpt71.266.079.1
only inst64.561.169.9
", + "image_path": "22ad8e861f7f4f22f994cb05842db6c107175e4a2def8237d538a3ea6ca65047.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 213, + 131, + 377, + 172 + ], + "blocks": [ + { + "bbox": [ + 213, + 131, + 377, + 172 + ], + "lines": [ + { + "bbox": [ + 213, + 131, + 377, + 172 + ], + "spans": [ + { + "bbox": [ + 213, + 131, + 377, + 172 + ], + "type": "table", + "html": "
query typescls taskAPAPMAPL
only kptavg kpt71.266.079.1
inst & kptavg kpt71.766.879.3
inst & kptinst72.066.879.7
", + "image_path": "6b854030b12b889a951c8b2e72f46f70abecd5431b9475f5075d69324875031f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 410, + 130, + 529, + 171 + ], + "blocks": [ + { + "bbox": [ + 410, + 130, + 529, + 171 + ], + "lines": [ + { + "bbox": [ + 410, + 130, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 410, + 130, + 529, + 171 + ], + "type": "table", + "html": "
#instanceAPAPMAPL
5071.466.479.1
10072.066.879.7
20072.066.979.7
", + "image_path": "383d8be24d22b0b01c88de55bac12ec0136139ff5826f27c74f97971ecbe0a6e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 177, + 288, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 288, + 216 + ], + "type": "text", + "content": "(d) Self-attention implementations. In the self-attention module, removing the across-instance interactions for queries with different types with attention (attn) mask or group self-attention eases the optimization and improves the performance." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 54, + 220, + 282, + 261 + ], + "blocks": [ + { + "bbox": [ + 54, + 220, + 282, + 261 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 282, + 261 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 282, + 261 + ], + "type": "table", + "html": "
self-attention implementationsw/ attn maskAPAPMAPL
standard self-attention×69.464.077.5
standard self-attention✓70.765.578.4
group self-attention×72.066.879.7
", + "image_path": "bcebdabe9bc0011801eeec04fa440856c316cbb248a54ce3cd43acab0c7abec7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 271, + 140, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 140, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 140, + 284 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 293, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 341 + ], + "type": "text", + "content": "We run a number of ablation experiments to verify the effectiveness of key elements in our Group Pose. We adopt ResNet-50 [9] as the backbone. Unless specified, we report the results on COCO val2017 with 60 epochs training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "content": "Ablation: query designs for human instance. Table 4a gives the results of representing a human instance with different query designs. Both instance query and keypoint queries are essential in Group Pose. As the multi-person pose estimation is to predict human poses given an image, it is reasonable that keypoint queries themselves can achieve good results, while removing keypoint queries gives a significant performance drop (from 72.0 AP to 64.5 AP)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 505 + ], + "type": "text", + "content": "For the instance query, its benefits come from two aspects: (i) gather information within human instance and help model training and (ii) decouple the classification task and keypoint positions regression task. Table 4b shows the improvements brought by these two benefits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Ablation: self-attention implementations. Table 4d provides the comparisons between different self-attention in the decoder layers, including (i) standard self-attention over all the " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "N \\times (K + 1)" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " queries (Figure 3 (a)), (ii) standard self-attention with an attention mask (Figure 3 (b)), which masks out the across-instance interactions for queries with different types, and (iii) our group self-attention (Figure 3 (c)). Results validate that the across-instance interactions for queries with different types are not directly helpful for the multi-person pose estimation task. Removing this type of interactions eases the model optimization, helps model converge faster, thus improves the performances. The comparisons on convergence curves are given in Figure 4. Besides, we find that it is better to separately perform different types of interactions, the within-instance and same-type across-instance interactions, with different parameters. Table 4d shows that our sequential implementation group self" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 301, + 177, + 543, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 177, + 543, + 216 + ], + "spans": [ + { + "bbox": [ + 301, + 177, + 543, + 216 + ], + "type": "text", + "content": "(e) Group self-attentions. Both types of self-attentions are important in group self-attentions. We also conduct an experiment by removing both within-instance and across-instance self-attentions. The performance further drops to 65.3 AP." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 309, + 220, + 538, + 261 + ], + "blocks": [ + { + "bbox": [ + 309, + 220, + 538, + 261 + ], + "lines": [ + { + "bbox": [ + 309, + 220, + 538, + 261 + ], + "spans": [ + { + "bbox": [ + 309, + 220, + 538, + 261 + ], + "type": "table", + "html": "
group self-attentionssAPAPMAPL
within-instance & across-instance self-attentionss72.066.879.7
within-instance self-attentionss66.363.870.8
across-instance self-attentionss67.463.074.4
", + "image_path": "45b5b720399ff2b8715f693c0da6754afd0ec069209448245559f0e63ba14ec7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 310, + 271, + 541, + 363 + ], + "blocks": [ + { + "bbox": [ + 310, + 271, + 541, + 363 + ], + "lines": [ + { + "bbox": [ + 310, + 271, + 541, + 363 + ], + "spans": [ + { + "bbox": [ + 310, + 271, + 541, + 363 + ], + "type": "image", + "image_path": "c13e0a49672a4448492403bfa91f66ad7518684121ae72b6a932ac09f40b3829.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 373, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 417 + ], + "type": "text", + "content": "Figure 4: Faster and better convergence. The " + }, + { + "bbox": [ + 305, + 373, + 545, + 417 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 373, + 545, + 417 + ], + "type": "text", + "content": "-axis corresponds to #epoch, and the " + }, + { + "bbox": [ + 305, + 373, + 545, + 417 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 373, + 545, + 417 + ], + "type": "text", + "content": "-axis corresponds to AP score. One can see that group self-attention converges faster and better than other self-attention implementations." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 431, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 545, + 454 + ], + "type": "text", + "content": " attentions bring a 1.3 AP gain over the parallel implementation of standard self-attention with an attention mask." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "text", + "content": "Ablation: group self-attentions. We ablate the effects of " + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "text", + "content": " within-instance self-attentions over " + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "text", + "content": " keypoint queries and one instance query and " + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "text", + "content": " across-instance self-attentions over " + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "text", + "content": " queries of the same type in Table 4e. Large performance gaps (over " + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "inline_equation", + "content": "-5.0\\mathrm{AP}" + }, + { + "bbox": [ + 304, + 458, + 546, + 662 + ], + "type": "text", + "content": ") are observed when we remove either of these two group self-attentions. Both the interactions are important for multi-person pose estimation. The within-instance self-attentions explore kinematic relations and gather information for scoring pose predictions, which help the model predict precise keypoint positions for human poses. The same-type across-instance self-attentions collect information from the same-type of keypoint queries or instance queries, removing duplicate predictions for poses among human instances. As shown in Figure 5, model without the same-type across-instance self-attentions produces more duplicated pose predictions for the same human instance than Group Pose." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Ablation: number of instances. Table 4c gives how predefined number of human instances affect the results. We find 100 human instances are enough and can achieve comparable results with 200 human instances. We set the num" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15035" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 177, + 141 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 177, + 141 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 177, + 141 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 177, + 141 + ], + "type": "image", + "image_path": "e6cfa798a371f8ac372d2e5ac660f503b0db2e62b16a338e37339baa83c611c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 141, + 176, + 210 + ], + "blocks": [ + { + "bbox": [ + 55, + 141, + 176, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 141, + 176, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 176, + 210 + ], + "type": "image", + "image_path": "ae478084818c4857514ecf386cff4996c3c093570d8af5df33aebf1935235c22.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 265 + ], + "type": "text", + "content": "Figure 5: Comparison of removing duplicated predictions. We visualize the predicted poses of Group Pose (second row), and Group Pose without same-type across-instance self-attention (first row) according to the number of objects. Group Pose produces less duplicated pose predictions. The yellow dashed ellipse indicates the duplicated predicted human instance, leading to a mismatch of other instances. Best view in zoom in." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 178, + 72, + 280, + 140 + ], + "blocks": [ + { + "bbox": [ + 178, + 72, + 280, + 140 + ], + "lines": [ + { + "bbox": [ + 178, + 72, + 280, + 140 + ], + "spans": [ + { + "bbox": [ + 178, + 72, + 280, + 140 + ], + "type": "image", + "image_path": "2f5613600a75b8f370e97a191151358c33d1b25694901b57ed62d14014eacfe2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 141, + 280, + 210 + ], + "blocks": [ + { + "bbox": [ + 178, + 141, + 280, + 210 + ], + "lines": [ + { + "bbox": [ + 178, + 141, + 280, + 210 + ], + "spans": [ + { + "bbox": [ + 178, + 141, + 280, + 210 + ], + "type": "image", + "image_path": "971f93e3d4e76b0bc9976722630cc3991ce7c80fc05d8fb004ff6f3c5b37ef49.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 281, + 72, + 384, + 140 + ], + "blocks": [ + { + "bbox": [ + 281, + 72, + 384, + 140 + ], + "lines": [ + { + "bbox": [ + 281, + 72, + 384, + 140 + ], + "spans": [ + { + "bbox": [ + 281, + 72, + 384, + 140 + ], + "type": "image", + "image_path": "0a767049aebf04137063dcf2ee5cc92a663c52149d4380d08c2fdc6ea14286d7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 281, + 141, + 384, + 210 + ], + "blocks": [ + { + "bbox": [ + 281, + 141, + 384, + 210 + ], + "lines": [ + { + "bbox": [ + 281, + 141, + 384, + 210 + ], + "spans": [ + { + "bbox": [ + 281, + 141, + 384, + 210 + ], + "type": "image", + "image_path": "afb2ebd454ecf6bc8840ab8f47594b35f4d5d7338ecf7b89b12a07459fc6608d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 385, + 72, + 484, + 140 + ], + "blocks": [ + { + "bbox": [ + 385, + 72, + 484, + 140 + ], + "lines": [ + { + "bbox": [ + 385, + 72, + 484, + 140 + ], + "spans": [ + { + "bbox": [ + 385, + 72, + 484, + 140 + ], + "type": "image", + "image_path": "8071d082320ae6a4252f76095b86c5a9f5f434f51e567b5714845b300c923a33.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 385, + 141, + 484, + 210 + ], + "blocks": [ + { + "bbox": [ + 385, + 141, + 484, + 210 + ], + "lines": [ + { + "bbox": [ + 385, + 141, + 484, + 210 + ], + "spans": [ + { + "bbox": [ + 385, + 141, + 484, + 210 + ], + "type": "image", + "image_path": "a4c51b53b30675e8f416da4963151e6ee570c60cd3cb2f3c218b6c36c1e5e0e2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 484, + 72, + 538, + 140 + ], + "blocks": [ + { + "bbox": [ + 484, + 72, + 538, + 140 + ], + "lines": [ + { + "bbox": [ + 484, + 72, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 484, + 72, + 538, + 140 + ], + "type": "image", + "image_path": "706640c2966ce82b31341ba3e38be81d65d73ec2712e388a736441a29713af11.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 484, + 141, + 538, + 210 + ], + "blocks": [ + { + "bbox": [ + 484, + 141, + 538, + 210 + ], + "lines": [ + { + "bbox": [ + 484, + 141, + 538, + 210 + ], + "spans": [ + { + "bbox": [ + 484, + 141, + 538, + 210 + ], + "type": "image", + "image_path": "f19f5546c2f9314533c5db1e68c48d65b6f2a8b6dbe4361df0827695486b2341.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 48, + 351, + 287, + 411 + ], + "blocks": [ + { + "bbox": [ + 46, + 276, + 289, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 289, + 342 + ], + "type": "text", + "content": "Table 5: Analysis on model convergence. Group Pose without human detection already can outperform ED-Pose [38]. Besides, Group Pose can also benefit from better human instance initialization with human detection. We report AP on COCO val2017 dataset. 'Det Dec' = human detection decoder. The results of ED-Pose are from the original paper [38]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 351, + 287, + 411 + ], + "lines": [ + { + "bbox": [ + 48, + 351, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 351, + 287, + 411 + ], + "type": "table", + "html": "
Methodw/ Det Dec12e24e36e48e60e
ED-Pose [38]✓60.567.569.770.871.6
GroupPose×61.067.670.171.472.0
GroupPose✓61.468.170.371.672.2
", + "image_path": "3255f3aec7190a2d9783622e7f3fd5179cce8095bf92dfed6ba8307c5fade54c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 421, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 458 + ], + "type": "text", + "content": "ber of human instances as 100 by default. Thus, Group Pose has " + }, + { + "bbox": [ + 46, + 421, + 287, + 458 + ], + "type": "inline_equation", + "content": "100 \\times (17 + 1)" + }, + { + "bbox": [ + 46, + 421, + 287, + 458 + ], + "type": "text", + "content": " queries on COCO [16] and contains " + }, + { + "bbox": [ + 46, + 421, + 287, + 458 + ], + "type": "inline_equation", + "content": "100 \\times (14 + 1)" + }, + { + "bbox": [ + 46, + 421, + 287, + 458 + ], + "type": "text", + "content": " queries on CrowdPose [14]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 464, + 138, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 138, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 138, + 477 + ], + "type": "text", + "content": "4.4. More Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 483, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 519 + ], + "type": "text", + "content": "We provide more analysis about our Group Pose in this section, including the analysis on model convergence and the analysis on model inference speed, detailed next." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": "Analysis on model convergence. A common wisdom about training models is to provide a good initialization and then refine the model based on it. ED-Pose [38] splits the multi-person pose estimation task into two sub-processes, which first detect human instances with a human detection decoder and then use a human-to-keypoint detection decoder for predicting human poses. The human detection decoder gives good initialization for human instances, which can help ED-Pose learn and converge faster. Although our Group Pose adopts a different design for transformer decoder, it can also be built upon a human detection decoder and enjoys the benefits brought by better initialization of human instances. Based on this observation, we build a variant of our Group Pose, whose transformer decoder consists of 2 human detection decoder layers and 4 simple transformer decoder layers for multi-person pose estimation, following" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 277, + 366, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 366, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 366, + 289 + ], + "type": "text", + "content": "ED-Pose [38]." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 306, + 353, + 553, + 457 + ], + "blocks": [ + { + "bbox": [ + 306, + 312, + 546, + 344 + ], + "lines": [ + { + "bbox": [ + 306, + 312, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 546, + 344 + ], + "type": "text", + "content": "Table 6: Analysis on model inference speed. Frames per second (FPS) and inference time (Time) are measured with ResNet-50 and different image resolutions on one NVIDIA A100 GPU." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 353, + 553, + 457 + ], + "lines": [ + { + "bbox": [ + 306, + 353, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 553, + 457 + ], + "type": "table", + "html": "
MethodInput ResolutionFPS ↑Time [ms] ↓
PETR [27]480 × 80020.050
800 × 133312.183
QueryPose [36]480 × 80019.056
800 × 133313.475
ED-Pose [38]480 × 80042.424
800 × 133324.740
GroupPose480 × 80068.615
800 × 133331.332
", + "image_path": "faa4b26ac65609b287f3b0f94e194ccabc82c9bae52421c1d7c1be7755751f07.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 477, + 547, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 547, + 596 + ], + "type": "text", + "content": "Table 5 shows the AP scores of three models, ED-Pose, Group Pose, and Group Pose with human detection decoder, under different training schedules. Group Pose already shows superior results than ED-Pose with 12 epochs, 24 epochs, 36 epochs, 48 epochs, and 60 epochs training, even without the help of human detection decoder. Moreover, the comparison between Group Pose and Group Pose with human detection decoder verifies that good initialization for human instances leads to faster convergence and better results." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Analysis on model inference speed. Table 6 provides the comparisons of inference time and FPS among end-to-end frameworks, including PETR [27], QueryPose [36], ED-Pose [38], and our Group Pose. Simple designs usually show efficiency. Results show that our simple transformer decoder can be faster than complex decoders in previous end-to-end frameworks. Even with an image in a large size " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "800 \\times 1333" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", our Group Pose can also achieve real-time speed (above 30 FPS) on a single A100 GPU." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15036" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 92, + 287, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 92, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 92, + 287, + 199 + ], + "type": "text", + "content": "In this paper, we present a simple baseline, Group Pose, for end-to-end multi-person pose estimation. With simple designs for queries and decoder self-attention, the approach outperforms previous end-to-end frameworks while being faster. Besides, the transformer decoder in Group Pose is also flexible, which can be built solely or upon human detection decoders. We hope Group Pose can provide insights for exploring concise and effective end-to-end multi-person pose estimation frameworks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 201, + 288, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 288, + 248 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.U1936212, No.62120106009, No.62261160653)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 262, + 107, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 107, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 107, + 274 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 283, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 283, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 287, + 338 + ], + "type": "text", + "content": "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7291-7299, 2017. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 339, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 288, + 393 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision, pages 213-229, 2020. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 396, + 289, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 289, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 289, + 440 + ], + "type": "text", + "content": "[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 442, + 289, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 289, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 289, + 497 + ], + "type": "text", + "content": "[4] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7103-7112, 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 499, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 564 + ], + "type": "text", + "content": "[5] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5386-5395, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 287, + 610 + ], + "type": "text", + "content": "[6] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rumpe: Regional multi-person pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2334-2343, 2017. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 287, + 667 + ], + "type": "text", + "content": "[7] Zigang Geng, Ke Sun, Bin Xiao, Zhaoxiang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14676-14686, 2021. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2961-2969, 2017. 2, 5, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 117 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[10] Jan Hosang, Rodrigo Benenson, and Bernt Schiele. Learning non-maximum suppression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4507-4515, 2017. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 195, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 547, + 239 + ], + "type": "text", + "content": "[12] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11977-11986, 2019. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[13] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "type": "text", + "content": "[14] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10863-10872, 2019. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[15] Ke Li, Shijie Wang, Xiang Zhang, Yifan Xu, Weijian Xu, and Zhuowen Tu. Pose recognition with cascade transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1944-1953, 2021. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 2, 4, 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[17] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "text", + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 581 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[20] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13264-13273, 2021. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[21] Weian Mao, Yongtao Ge, Chunhua Shen, Zhi Tian, Xinlong Wang, Zhibin Wang, and Anton van den Hengel. Poseur: Direct human pose regression with transformers. In Proceedings of the European Conference on Computer Vision, pages 72-88, 2022. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[22] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15037" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "with dynamic instance-aware convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9034-9043, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[23] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "text", + "content": "[24] Alejandro Newell, Zhiao Huang, and Jia Deng. Associative embedding: End-to-end learning for joint detection and grouping. Advances in Neural Information Processing Systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "text", + "content": "[25] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Proceedings of the European Conference on Computer Vision, pages 483-499. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "text", + "content": "[26] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6951-6960, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 288, + 353 + ], + "type": "text", + "content": "[27] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11069-11078, 2022. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 410 + ], + "type": "text", + "content": "[28] Dahu Shi, Xing Wei, Xiaodong Yu, Wenming Tan, Ye Ren, and Shiliang Pu. Inspose: instance-aware networks for single-stage multi-person pose estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3079-3087, 2021. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "text", + "content": "[29] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5693-5703, 2019. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 533 + ], + "type": "text", + "content": "[30] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14454-14463, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 567 + ], + "type": "text", + "content": "[31] Zhi Tian, Hao Chen, and Chunhua Shen. Directpose: Direct end-to-end multi-person pose estimation. arXiv preprint arXiv:1911.07451, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "text", + "content": "[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[34] Fangyun Wei, Xiao Sun, Hongyang Li, Jingdong Wang, and Stephen Lin. Point-set anchors for object detection, instance segmentation and pose estimation. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 441 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision, pages 527-544. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[35] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European Conference on Computer Vision, pages 466-481, 2018. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[36] Yabo Xiao, Kai Su, Xiaojuan Wang, Dongdong Yu, Lei Jin, Mingshu He, and Zehuan Yuan. Querypose: Sparse multiperson pose regression via spatial-aware part-level query. In Advances in Neural Information Processing Systems, pages 1-14, 2022. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[37] Nan Xue, Tianfu Wu, Gui-Song Xia, and Liangpei Zhang. Learning local-global contextual adaptation for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13065-13074, 2022. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "text", + "content": "[38] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In International Conference on Learning Representations, pages 1-17, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[39] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, pages 1-18, 2022. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 396 + ], + "type": "text", + "content": "[40] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 398, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 441 + ], + "type": "text", + "content": "[41] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, pages 1-16, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15038" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_content_list.json b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..24a53212d223aae3f02c02e3bf0727270e109c95 --- /dev/null +++ b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_content_list.json @@ -0,0 +1,1614 @@ +[ + { + "type": "text", + "text": "GrowCLIP: Data-aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-training", + "text_level": 1, + "bbox": [ + 207, + 130, + 764, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinchi Deng $^{1}$ Han Shi $^{2}$ Runhui Huang $^{1}$ Changlin Li $^{3}$ Hang Xu $^{2}$ \nJianhua Han $^{2}$ James Kwok $^{4}$ Shen Zhao $^{1*}$ Wei Zhang $^{2}$ Xiaodan Liang $^{1}$ $^{1}$ Sun Yat-sen University $^{2}$ Huawei Noah's Ark Lab \n $^{3}$ University of Technology Sydney $^{4}$ The Hong Kong University of Science and Technology \ndengxch5@mail2.sysu.edu.cn", + "bbox": [ + 181, + 202, + 805, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-modal pre-training has shown impressive performance on a wide range of downstream tasks, benefiting from massive image-text pairs collected from the Internet. In practice, online data are growing constantly, highlighting the importance of the ability of pre-trained model to learn from data that is continuously growing. Existing works on cross-modal pre-training mainly focus on training a network with fixed architecture. However, it is impractical to limit the model capacity when considering the continuously growing nature of pre-training data in real-world applications. On the other hand, it is important to utilize the knowledge in the current model to obtain efficient training and better performance. To address the above issues, in this paper, we propose GrowCLIP, a data-driven automatic model growing algorithm for contrastive language-image pre-training with continuous image-text pairs as input. Specifically, we adopt a dynamic growth space and seek out the optimal architecture at each growth step to adapt to online learning scenarios. And the shared encoder is proposed in our growth space to enhance the degree of cross-modal fusion. Besides, we explore the effect of growth in different dimensions, which could provide future references for the design of cross-modal model architecture. Finally, we employ parameter inheriting with momentum (PIM) to maintain the previous knowledge and address the issue of the local minimum dilemma. Compared with the existing methods, GrowCLIP improves $2.3\\%$ average top-1 accuracy on zero-shot image classification of 9 downstream tasks. As for zero-shot image retrieval, GrowCLIP can improve $1.2\\%$ for top-1 image-to-text recall on Flickr30K dataset.", + "bbox": [ + 75, + 368, + 472, + 821 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75efdd0b644164ce01f11930b0362fa1ddf453c61cbbaaa819538b4866d032d6.jpg", + "image_caption": [ + "Figure 1. Top-1 accuracy $(\\%)$ of zero-shot image classification on ImageNet of GrowCLIP and baselines during training at step 4, where the horizontal dotted lines mean the process of supernet training. Our GrowCLIP has the best performance and is more efficient compared with other baselines." + ], + "image_footnote": [], + "bbox": [ + 513, + 325, + 857, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 614, + 632, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large-scale pre-trained models illustrate the potential performance among different fields including computer vision (CV) [24, 25, 11] and natural language processing (NLP) [15, 44, 33]. Cross-modal models like CLIP [52], ALIGN [28], FILIP [65], BLIP [36] also demonstrate remarkable success across various vision-language downstream tasks. Note that these models are usually built in dual-stream architectures, which consist of the image and text encoders to extract the feature of image and text inputs, respectively. The alignment between these two features is then performed under the contrastive objective to learn the alignment between different modals.", + "bbox": [ + 496, + 642, + 890, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, training these cross-modal models requires a large amount of image-text pairs collected from the Internet, e.g., 400M image-text pairs for CLIP [52] and 300M for FILIP [65]. Most existing cross-modal methods [52, 28, 65, 39, 36, 68, 63, 1] directly train the final large-", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22178", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/a0ea8e71f475b5a9f206990de9099cf88db1c7c657cc546f24cf8d1835736645.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Result of different model sizes with different data sizes: Top-1 accuracy (\\%) of zero-shot image classification on ImageNet. The depth and the width of 0.5-ViT-B/16 are both one half of ViT-B/16. The relative performance ranking is dependent on the size of the training data." + ], + "table_body": "
CC12M10%20%50%100%
CLIP-0.5-ViT-B/166.611.119.025.7
CLIP-ViT-B/166.511.620.928.3
", + "bbox": [ + 94, + 88, + 450, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scale model using the completed image-text pair datasets collected at some point in time. However, it is impractical to regard cross-modal pre-training as \"disposable\" training without considering the continuously growing nature of pre-training data or domain in real-world applications. For instance, mountains of new knowledge is generated on the Internet every day, which can be used to further improve the performance of existing pre-trained cross-modal models. Therefore, in this paper, we consider the training of large-scale cross-modal pre-trained models as an online learning problem [17, 56, 57] with continuous image-text pairs as input. Different from the standard continual learning setting [59, 55, 22, 31, 2, 7], the previous image-text pairs can also be achieved in our setting since we assume the size of training data is gradually increased over time.", + "bbox": [ + 75, + 231, + 468, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One tough problem in the online learning scenario is that the model capacity should be related to the size of training data [16, 69, 29, 26]. For example, it's observed that large ViT models perform worse than ResNets when pre-trained on small datasets, while the result is opposite when they are both pre-trained on larger datasets [16]. To verify the relationship between model capability and data size, we split Conceptual 12M (CC12M) dataset [9] and test the performance of ViT-B/16 with different scales [52]. As shown in Table 1, the relative performance ranking is dependent on the size of the training data. Given a smaller dataset (e.g., $10\\%$ CC12M), the performance of 0.5-ViT-B/16 is comparable with ViT-B/16. In contrast, when data is sufficient (e.g., $100\\%$ CC12M), the performance of ViT-B/16 is better than the smaller model. Thus, the unchanged model is not practical for this real-time learning setting and it's still an open issue on how to modify and train our model with the growing data. Besides, how to efficiently make use of knowledge of previous model when new data is coming remains an open problem. One direct solution is to fine-tune the model with the updated training dataset. However, training with previous pre-trained parameters of the same model will deteriorate the performance [4] due to the parameter inheriting issue. As shown in Figure 1, the CLIP training with pre-training (TWP) have worse performance than the one training from scratch (TFS).", + "bbox": [ + 75, + 460, + 468, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose a data-aware automatic model growing method (denoted as GrowCLIP) for large-scale contrastive language-image pre-training, which", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performs a model growth process considering the gradually increased pre-training data. Specially, when training data grows dynamically, we adopt different cross-modal network architectures via the customized neural architecture search (NAS) to make the network pre-training more efficient. Different from traditional NAS approaches [70, 53, 64], we introduce a cross-modal customized NAS by defining a dynamic search space named growth space, which is enlarged when more data comes, and proposing a shared encoder search space to enhance the degree of cross-modal fusion. To utilize the architecture at the previous step more efficiently, the parameters of the new architecture are also inherited from the old one with momentum to maintain the previous knowledge and address the issue of local minimum dilemma. Finally, growth architecture selection procedure is performed to select the optimal model architecture at each step, considering the performance and model size.", + "bbox": [ + 496, + 90, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments are conducted by averagely dividing the Conceptual 12M (CC12M) into 4 growth steps under the online learning setting. As depicted in Table 1, compared with the existing methods, our GrowCLIP has the best performance and is more efficient. Specially, experimental results show that our GrowCLIP can improve up to $2.3\\%$ average top-1 accuracy on zero-shot image classification of 9 downstream tasks compared with the existing methods. As for zero-shot image-text retrieval, GrowCLIP has a $1.2\\%$ improvement for top-1 image-to-text recall on Flickr30K [50] dataset and $0.8\\%$ on MSCOCO [40].", + "bbox": [ + 496, + 347, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the contributions of this paper are listed as follows: (i) To adapt to the growing data scenario, we propose a data-aware automatic model growing method, named GrowCLIP. (ii) We provide some insights for the design of cross-modal model architecture. (iii) The extensive experiment results illustrate the effectiveness of our approach on zero-shot classification and retrieval tasks.", + "bbox": [ + 496, + 513, + 890, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 633, + 640, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Pre-training Models Inspired by the success of pre-training in computer vision (CV) and natural language processing (NLP), a boosting number of research works in the domain of vision-language pretraining (VLP) has recently surged to pursue a unified multi-modality representation. Vision-language pretraining trains the model on large-scale image-text pairs to improve performance of downstream vision and language tasks including image classification [14, 35, 32], retrieval [50, 40], Image Captioning (IC) [43, 13, 27, 26], Visual Question and Answer (VQA) [3, 46, 67] and so on. The VLP model can be categorized into two categories: (i) The dual-stream model, e.g., CLIP [52], ALIGN [28], FILIP [65], processes visual and textual tokens separately with two parallel streams to acquire the representation and then fuse them through interactive module. (ii) The single-", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22179", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stream model, e.g., Oscar [38], ViLT [30], VisualBERT [37], directly combines the extracted visual and textual tokens and feeds them into the transformer-based model. The public pre-training datasets of VLP model include available datasets like YFCC100M [61] and CC12M [9]. CLIP [52], ALIGN [28], FILIP [65] and LEMON [26] adopt large-scale image-text pairs crawled from the web to show more powerful models. Different from these works that treat this process as offline learning, we investigate the online learning case in vision-language pre-training.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online Learning Online learning is a problem setting of machine learning in which data becomes available in a sequential order [17, 56, 57]. Different from continual learning [22, 31, 2, 7], online learning can reach previous data all the time. As opposed to batch learning techniques which generate the best model by learning on the entire training dataset at once, online learning is used to update our predictor for future data at each growth step. Follow the leader [58] is the simplest learning rule which assumes the hypothesis that the leader has the least loss overall past rounds. Follow the regularised leader [51] is a modification variant and obtain better regret bounds by learning a regularisation function. Online convex optimization [23] is a general algorithm framework for decision making which targets convex optimization. To adapt online learning scenario, we propose a data-aware automatic model growing algorithm that grows the model with increasing training data.", + "bbox": [ + 75, + 242, + 472, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Architecture Search Neural Architecture Search (NAS) aims to automatically design the state-of-the-art neural networks. Early works [70, 54, 60] are based on reinforcement learning and evolutionary algorithms. In the later work, various techniques are used to reduce the computational complexity. Gradient-based methods [42, 41] regard the network architecture as a set of parameters and adopt back-propagation algorithm to optimize them. One-shot NAS methods [49, 10, 20] train a supernet to avoid training each subnet from scratch. Unfortunately, one-shot NAS methods have to spend lots of resources to train the supernet. In contrast, the method we propose inherits the parameters of the old model with momentum, which only takes few epochs to train the supernet. Furthermore, GrowCLIP can modify the neural architecture to adapt to the size of the training data.", + "bbox": [ + 75, + 513, + 472, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 768, + 212, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the online learning problem setting and its challenges for cross-modal pre-training (Section 3.1). To adapt the growth data, we propose a dynamic growth space, including a proposed shared encoder (Section 3.2). To rise to the challenge of the local minimum dilemma, we propose parameter inheriting with momentum (Section 3.3). The whole pipeline of our method is illus", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trated in Section 3.4.", + "bbox": [ + 498, + 90, + 640, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting and Challenges", + "text_level": 1, + "bbox": [ + 498, + 114, + 782, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setting The target of online learning is obtaining the optimal model at each growth step of the data. Let $\\mathcal{D}^{t - 1} = \\{\\pmb{x}_i^I,\\pmb{x}_i^T\\}_{i = 1}^k$ be the training data at growth step $t - 1$ , where $\\pmb{x}^I$ and $\\pmb{x}^T$ are image and text samples and $k$ is the training data size. At the next growth step $t$ , the enlarged training dataset $\\mathcal{D}^t$ satisfies $\\mathcal{D}^{t - 1}\\subseteq \\mathcal{D}^t$ due to the incoming data. Given an optimal model from growth step $t - 1$ , our aim is to achieve the optimal model (with optimal architecture $\\psi^{*}$ and parameters $\\omega^{*}$ ) based on the enlarged dataset $\\mathcal{D}^t$ :", + "bbox": [ + 496, + 138, + 893, + 290 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\psi^ {*}, \\omega^ {*} \\right\\} = \\underset {\\psi , \\omega} {\\arg \\min } \\mathcal {L} (\\psi , \\omega ; \\mathcal {D} ^ {t}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 305, + 890, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge of Data Growth As discussed in Table 1, the size of the training dataset is related to the optimal architecture selection. Given a smaller dataset, the small-scale architecture has comparable performance to large-scale architecture but fewer parameters are deployed. When data grows, the small-scale architecture may limit the final performance due to the limited parameters. However, few works notice the effect of data size in online learning settings.", + "bbox": [ + 496, + 337, + 893, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge of Parameter Inheriting As investigated, the performance of model trained from scratch is much better than trained with pre-training. This is probably caused by the influence of the inheriting parameters trained with the previous smaller dataset, which makes it difficult for the model to escape from the local minima in the grown dataset. However, training from scratch is costly due to multiple growth steps in online learning. The balance between training from scratch and training with pre-training is worthy of consideration.", + "bbox": [ + 496, + 474, + 893, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To solve the above problems, we propose an automatic model growing method - GrowCLIP (shown in Figure 2), which can modify the neural architecture at different growth steps in online learning case. To adapt to the growing data scenario, we propose a growth space for each growth step, where the scale of architecture candidates among this space is larger when data grows. To alleviate the parameter inhering issue with efficiency, we perform parameter inhering with momentum such that the trade off between training from scratch and training with pre-training is balanced.", + "bbox": [ + 496, + 625, + 893, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Growth Space", + "text_level": 1, + "bbox": [ + 500, + 786, + 648, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Architecture Following the basic neural architecture in CLIP [52], we adopt a dual-stream model with image encoder $f(\\cdot)$ and text encoder $g(\\cdot)$ . To adapt the cross-modality scenarios, we propose a shared encoder $h(\\cdot)$ , in which the image and text share the transformers except layernorm layers. The shared encoder not only can reduce", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22180", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6edcf2db6e9d07d8e12003be64e20d8d8870f81efafe8db811bdde3147900670.jpg", + "image_caption": [ + "Figure 2. The overview of GrowCLIP. As shown in Figure (a), when the data grows from growth step $t - 1$ to growth step $t$ , our model grows adaptively as well. Figure (b) illustrates the detail of model growing and we simplify the representation of model architecture." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 524, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9d953690f661e72d8c03171fbc2af26326642285bc46b92416cd579b8764279.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 88, + 893, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model parameters, but also enhance the degree of cross-modal fusion. Given an image $\\boldsymbol{x}_i^I$ and a text $\\boldsymbol{x}_j^T$ , the model outputs the image representations $I_i = h(f(\\boldsymbol{x}_i^I))$ and text representations $T_j = h(g(\\boldsymbol{x}_j^T))$ . The similarity score computation of between image $\\boldsymbol{x}_i^I$ and text $\\boldsymbol{x}_j^T$ is:", + "bbox": [ + 75, + 433, + 468, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = h \\left(f \\left(\\boldsymbol {x} _ {i} ^ {I}\\right)\\right) ^ {\\top} h \\left(g \\left(\\boldsymbol {x} _ {j} ^ {T}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 523, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_{i,j}$ denotes the similarity of the $i$ -th image to the $j$ -th text.", + "bbox": [ + 75, + 551, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each training batch, we sample $n$ image-text pairs $\\{\\pmb{x}_i^I,\\pmb{x}_i^T\\}_{i = 1}^n$ from the dataset. For image $\\pmb{x}_i^I$ in the image-text pair $< x_{i}^{I},x_{i}^{T}>$ $x_{i}^{T}$ is its positive, while the other texts will be used as in-batch negatives. The image-to-text contrastive loss $\\mathcal{L}_i^I$ for $\\pmb{x}_i^I$ can be formulated as:", + "bbox": [ + 76, + 582, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} ^ {I} \\left(\\boldsymbol {x} _ {i} ^ {I}, \\left\\{\\boldsymbol {x} _ {j} ^ {T} \\right\\} _ {j = 1} ^ {n}\\right) = - \\frac {1}{n} \\log \\frac {\\exp \\left(s _ {i , i}\\right)}{\\sum_ {j} \\exp \\left(s _ {i , j}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 667, + 468, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, the text-to-image contrastive loss for $\\pmb{x}_i^T$ is:", + "bbox": [ + 76, + 710, + 433, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} ^ {T} \\left(\\boldsymbol {x} _ {i} ^ {T}, \\left\\{\\boldsymbol {x} _ {j} ^ {I} \\right\\} _ {j = 1} ^ {n}\\right) = - \\frac {1}{n} \\log \\frac {\\exp \\left(s _ {i , i}\\right)}{\\sum_ {j} \\exp \\left(s _ {i , j}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 736, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Totally, the contrastive loss of this mini-batch can be represented by $\\mathcal{L} = \\frac{1}{2}\\sum_{i=1}^{a}\\left(\\mathcal{L}_i^I + \\mathcal{L}_i^T\\right)$ .", + "bbox": [ + 75, + 780, + 468, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Growth Space of image encoder For the image encoder, we adopt a mixed architecture of convolutional layer and transformer block [62]. There are three variable factors in our growth space design: (i) the number of convolutional layers $l^I$ ; (ii) the number of transformer blocks $b^I$ ; (iii) the number of transformer heads $h^I$ .", + "bbox": [ + 75, + 811, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Growth Space of text encoder The text encoder part is a transformer-based architecture, whose growth space includes two variable factors: (i) the number of transformer blocks $b^{T}$ and (ii) the number of transformer heads $h^{T}$ .", + "bbox": [ + 496, + 433, + 892, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Growth Space of shared encoder The shared encoder is a transformer-based architecture, whose growth space includes the number of transformer blocks $b^{S}$ .", + "bbox": [ + 496, + 494, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ff35109c874407bdfefd856922e95d6938a4de29830c8b0f4838c6635dfa59ad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Image encoderText encoderShared encoder
Transformer blocks{bIt-1, bIt-1+4}{bTt-1, bTt-1+4}{bSt-1, bSt-1+4}
Transformer heads{hIt-1, hIt-1+4}{hTt-1, hTt-1+4}-
Convolutional layers{lIt-1, lIt-1+2}--
", + "bbox": [ + 501, + 547, + 893, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. The growth space $\\Phi$ of GrowCLIP at the growth step t. ${b}_{t - 1}^{I},{b}_{t - 1}^{T},{b}_{t - 1}^{S},{h}_{t - 1}^{I},{h}_{t - 1}^{T},{l}_{t - 1}^{I}$ are the hyperparameters of model architecture at growth step $t - 1$ .", + "bbox": [ + 498, + 616, + 893, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detailed descriptions of growth space for three encoders are shown in Table 2. At each new growth step, the variable factors are enlarged to adapt the growth data. Specifically, the numbers of transformer blocks and transformer heads are both increased by 4, and the number of convolutional layers is increased by 2.", + "bbox": [ + 496, + 672, + 890, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Parameter inheriting with momentum", + "text_level": 1, + "bbox": [ + 500, + 771, + 828, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some previous papers utilize the training results of the old model and transfer the knowledge to the new model [18, 19, 25, 4, 34]. Inspired by the above methods, we propose parameter inheriting with momentum (PIM) to avoid local minimum dilemma. The illustration of our parameter inheriting with momentum is shown in Figure 3. To preserve the knowledge from the old model, we extend the size", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8bab3a1ee52d8ce8e2c6e01e9e660787ba328cfc961bb08276127fe556278dde.jpg", + "image_caption": [ + "Figure 3. Take image encoder for example of our proposed parameter inheriting with momentum. As for the initialization of the enlarged model, we preserve the existing parameters of the old model for exploitation and combine it with random initialization for exploration." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the old model to the enlarged model and maintain the parameters in existing parts. As for the extended layers, we copy the parameters from the last layers directly. $\\omega_{old}$ denotes the parameters of extension of the old model. Besides the above parameter inheriting term, we also add a re-initialization term with randomly initialized parameters $\\omega_{rand}$ to avoid the parameter inheriting issue. The final initialized supernet is a combination of the above two terms as follows:", + "bbox": [ + 75, + 289, + 468, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {\\text {i n h e r i t}} = \\beta \\cdot \\omega_ {\\text {o l d}} + \\gamma \\cdot \\omega_ {\\text {r a n d}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 426, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\gamma$ are two hyperparameters related to exploitation and exploration. Additionally, the parameters of new transformer heads are randomly initialized, which makes that different heads in transformer are orthogonal and the new heads learn new features from different perspectives.", + "bbox": [ + 75, + 446, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Growth Pipeline", + "text_level": 1, + "bbox": [ + 76, + 529, + 238, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The growth pipeline at growth step $t$ (except the first one) includes three sequential phases as shown in Figure 2 (b).", + "bbox": [ + 75, + 551, + 468, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase 1: Supernet Fine-tuning As for a new growth step, a supernet $\\Psi$ covering the whole growth space is constructed by enlarging the current model. First, we use the PIM method described in Section 3.3 to initialize the supernet. However, there are some newly added layers in the supernet compared with the old model from the last growth step, which brings a performance gap between the subnets with new layers and others. To make the supernet more robust, we slightly fine-tune the parameters of the full supernet for a few epochs to mitigate the performance gap among subnetworks after PIM initialization.", + "bbox": [ + 75, + 583, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase 2: Growth Architecture Selection Growth Architecture Selection (GAS) is the main phase in our algorithm where we select the optimal architecture in the growth space. The whole GAS process can be divided into two steps: training and searching.", + "bbox": [ + 75, + 750, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, to find the optimal architecture $\\psi^{*}$ , we need to fully optimize all the subnet in the supernet, which covers all the candidate growth architecture:", + "bbox": [ + 75, + 824, + 468, + 868 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {s u p e r} ^ {*} = \\underset {\\omega_ {s u p e r}} {\\arg \\min } \\mathbb {E} _ {\\psi \\in \\Psi} \\left\\{\\mathcal {L} \\left(\\psi , \\omega_ {s u p e r}; \\mathcal {D} ^ {t}\\right) \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 876, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by one-shot neural architecture search [8], we optimize $\\omega_{super}$ in Equation 6 by uniformly sampling subnets $\\psi$ from the growth space $\\Phi$ and update the parameters of the sampled subnets in each iteration. More specifically, at each step, we first randomly sample a subnet $\\psi$ from the growth space $\\Psi$ , then optimize the parameter of this subnet by minimizing the contrastive loss between image and text. This can be considered as solving Equation 6 via Monte Carlo sampling. By iterative training, these subnets can be fully optimized in a few epochs.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, after training the supernet, we search for the optimal subnet in the growth space $\\Phi$ . Here, we consider both the performance and the number of parameters of the architecture, and select the optimal architecture $\\psi^{*}$ using the composite metric:", + "bbox": [ + 496, + 242, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi^ {*} = \\underset {\\psi \\in \\Psi} {\\arg \\max } \\left\\{\\mathcal {A} (\\psi) + \\alpha \\frac {\\mathcal {N} ^ {t - 1}}{\\mathcal {N} ^ {t}} \\frac {\\mathcal {P} (\\Psi)}{\\mathcal {P} (\\psi)} \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 328, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}(\\cdot)$ is the top-1 accuracy of test dataset, $\\mathcal{N}^t$ denotes the number of data on growth step $t$ , $\\mathcal{P}(\\cdot)$ represents the total amount of model parameters and $\\alpha$ is the trade-off hyperparameter. When we get a large number of new coming data, $\\frac{\\mathcal{N}^{t - 1}}{\\mathcal{N}^t}$ will get small, which means that we focus more on the performance and encourage the growth of models. As our growth space is relatively small, traversal evaluation of all the candidate architectures is affordable. Therefore, we simply test all candidate subnets in the growth space $\\Phi$ using the parameters inherited from the supernet $\\Psi$ and select the optimal one for the next stage via the composite metric of Equation 7.", + "bbox": [ + 496, + 375, + 890, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase 3: Selected Model Training To alleviate the influence of supernet training on the grown model, we also use PIM to initialize the parameter $\\omega$ of selected model $\\psi^{*}$ :", + "bbox": [ + 496, + 559, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\omega} = \\beta \\cdot \\boldsymbol {\\omega} _ {s u p e r} ^ {*} + \\gamma \\cdot \\boldsymbol {\\omega} _ {r a n d}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 619, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the parameter inheriting, we keep training the selected model, to obtain its final optimal parameters:", + "bbox": [ + 496, + 643, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega^ {*} = \\underset {\\omega} {\\arg \\min } \\mathcal {L} \\left(\\psi^ {*}, \\omega ; \\mathcal {D} ^ {t}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 685, + 890, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 720, + 632, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we firstly describe the experiment setting and implementation details (Section 4.1). Then we show the results on zero-shot image classification (Section 4.2) and image-text retrieval (Section 4.3). Finally, we conduct ablation study on our GrowCLIP to validate the effectiveness of components (Section 4.4) and analyse (Section 4.5).", + "bbox": [ + 496, + 746, + 890, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setting", + "text_level": 1, + "bbox": [ + 500, + 847, + 686, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Protocol Conceptual 12M (CC12M) [9] collects about 12 million image-text pairs from the Internet.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22182", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4ce9ff3adeab9b9a16db0b6532f742e1b70f58989e0a2cbd39bab2f6c08c8f2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Growth stepMethodTraining dataCaltech101CIFAR10CIFAR100DTDFlowers102Food101OxfordPedsSUN397ImageNetAverage
ParameterParameterParameterParameterParameterParameterParameterParameterParameterParameter
Step 1CLIP (TWP)3M149.6M42.837.614.5109.316.121.622.014.020.9
CLIP (TFS)[52]3M149.6M42.837.614.5109.316.121.622.014.020.9
MS-CLIP (TFS)[21]3M129M47.135.514.410.012.917.330.124.718.123.3
CLIP (SAP)[4]3M149.6M42.837.614.5109.316.121.622.014.020.9
NAS-S13M30.0M40.531.213.39.08.915.820.223.213.719.5
GrowCLIP-S1 (ours)3M30.0M40.531.213.39.08.915.820.223.213.719.5
Step 2CLIP (TWP)6M149.6M50.336.116.011.312.421.227.125.817.924.2
CLIP (TFS)[52]6M149.6M58.552.022.312.113.125.033.630.120.929.7
MS-CLIP (TFS)[21]6M129M60.650.018.313.417.42742.432.625.231.8
CLIP (SAP)[4]6M149.6M61.949.021.412.415.524.635.431.622.630.5
NAS-S26M96.9M59.441.818.313.315.826.737.333.824.430.1
GrowCLIP-S2 (ours)6M116.6M57.748.021.213.718.029.539.534.525.732.0
Step 3CLIP (TWP)9M149.6M55.240.119.713.115.723.830.928.920.427.5
CLIP (TFS)[52]9M149.6M63.354.624.514.517.529.536.933.825.333.3
MS-CLIP (TFS)[21]9M129M65.048.222.513.920.631.042.635.729.233.0
CLIP (SAP)[4]9M149.6M67.853.926.715.121.232.941.436.028.235.9
NAS-S39M287.4M66.746.822.214.919.735.446.139.330.135.8
GrowCLIP-S3 (ours)9M168.6M69.653.925.818.418.736.346.241.632.638.1
Step 4CLIP (TWP)12M149.6M57.544.420.512.114.224.134.730.322.228.9
CLIP (TFS)[52]12M149.6M65.957.329.716.318.532.945.536.728.336.8
MS-CLIP (TFS)[21]12M129M65.350.926.316.221.935.950.138.432.037.4
CLIP (SAP)[4]12M149.6M69.860.631.912.421.036.747.040.432.239.1
NAS-S412M325.9M70.853.928.316.721.739.551.241.334.239.7
GrowCLIP-S4 (ours)12M188.4M71.960.728.317.323.342.552.445.536.142.0
", + "bbox": [ + 86, + 88, + 883, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Top-1 accuracy(%) of zero-shot image classification on 9 datasets. Our GrowCLIP outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at each growth step except step 1.", + "bbox": [ + 76, + 503, + 890, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The CC12M is randomly divided into 4 subsets. We select one subset as the training set for the first growth step and adds one subset to the training set at each subsequent growth step such that there are total 4 growth steps in our experiments. At each growth step, we evaluate the corresponding model on the downstream tasks including zero-shot image classification and zero-shot image-text retrieval. The grown architecture of each growth step is shown in Appendix.", + "bbox": [ + 75, + 556, + 468, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Methods To demonstrate the effectiveness of our method, we compare it with the following baselines: (i) Training with pre-training (TWP) is a usual method in transfer learning. (ii) Training from scratch (TFS) [52] is used as a basic method to train language-image pre-trained model. (iii) MS-CLIP [21] is a handcrafted framework with a shareable module and modality-specific auxiliary module (the early specialization). (iv) The shrink and perturb (SAP) [4] is a method of parameter inheriting in warm-starting learning. (v) Neural architecture search (NAS) [49, 10] is an important research field of auto machine learning, in which we use the same search space as the growth space of GrowCLIP.", + "bbox": [ + 75, + 683, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details A dual-stream model follow", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing CLIP [52] without shared encoder is used as basic architecture at the first growth step. The image encoder is stacked with 6 transformer blocks with the following hyperparameters: number of self-attention heads is 6 and the hidden layer size is 384. As for the text encoder, 6 transformer blocks are stacked and the number of the attention heads and the width are 4 and 256 respectively. The resolution of input in the images encoder is resized to 224 and the maximum sequence in the text encoder is limited to 77. The center crop is the only data augmentation method for training. We use the prompts like $\\mathrm{CLIP}^1$ on the zero-shot image classification. We adopt LAMB optimizer [66] with learning rate 0.01 and weight decay 0.0001, and cosine learning rate schedule [45] with a 4000-iteration linear warmup. For the hyperparameters, the proportion of the old model and the new one in the parameter inheriting $\\beta = 0.3$ , $\\gamma = 0.001$ . The model parameter coefficient in growth architecture selection $\\alpha = 0.5$ . We train the model with a total batch size of 1536 for 30 epochs at each growth step, which includes 2 epochs for supernet fine-tuning, 2 epochs for su", + "bbox": [ + 496, + 556, + 890, + 859 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/openai/CLIP/blob/main/data/prompts.md", + "bbox": [ + 500, + 875, + 887, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1a8dbc25c08c355e4e9d0d0493d7763600afc31f7b1ea7b53f4f993bf4d3e08d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Growth stepMethodFlickr30KMSCOCO
image-to-texttext-to-imageimage-to-texttext-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
Step 1CLIP (TWP)19.645.356.414.333.544.110.726.336.87.019.127.5
CLIP (TFS) [52]19.645.356.414.333.544.110.726.336.87.019.127.5
MS-CLIP (TFS)[21]27.451.863.318.139.250.712.931.943.28.823.533.1
CLIP (SAP) [4]19.645.356.414.333.544.110.726.336.87.019.127.5
NAS-S122.845.257.813.534.246.010.627.437.57.019.928.8
GrowCLIP-S1 (ours)22.845.257.813.534.246.010.627.437.57.019.928.8
Step 2CLIP (TWP)26.651.062.017.439.850.712.830.540.88.322.531.9
CLIP (TFS) [52]31.457.768.121.246.757.117.137.448.710.927.337.4
MS-CLIP (TFS)[21]35.865.175.025.853.164.619.343.055.213.131.542.6
CLIP (SAP) [4]33.660.972.024.148.259.317.138.050.210.928.028.1
NAS-S236.963.074.025.251.762.318.441.653.712.930.941.7
GrowCLIP-S2 (ours)37.563.774.225.951.162.819.142.854.712.930.841.9
Step 3CLIP (TWP)28.354.466.620.642.254.714.434.145.09.625.234.9
CLIP (TFS) [52]37.563.574.226.052.164.019.743.255.113.331.842.6
MS-CLIP (TFS)[21]40.067.778.229.857.068.822.947.959.115.335.246.6
CLIP (SAP) [4]40.768.078.328.155.367.122.845.758.814.233.444.6
NAS-S343.272.880.732.159.970.223.448.361.015.736.848.3
GrowCLIP-S3 (ours)44.673.180.733.260.770.923.548.861.015.836.948.4
Step 4CLIP (TWP)30.256.366.321.746.657.615.636.247.810.426.536.9
CLIP (TFS) [52]41.568.680.030.056.968.021.845.557.314.333.845.1
MS-CLIP (TFS)[21]44.372.881.132.661.372.125.250.562.016.637.749.0
CLIP (SAP) [4]46.974.783.733.761.372.125.250.862.816.637.449.1
NAS-S449.675.283.337.065.275.426.853.064.617.739.651.5
GrowCLIP-S4 (ours)50.875.884.037.265.376.227.653.264.918.140.352.0
", + "bbox": [ + 76, + 90, + 903, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results of zero-shot image-text retrieval on Flickr30K and MSCOCO datasets. R@K means top-K recall.", + "bbox": [ + 145, + 463, + 821, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pernet training and 26 epochs for grown model training. For the baselines, the CLIP adopts the same usual architecture (ViT-B/16), which has 12-layer 768-wide transformer blocks with 12 attention heads for image encoder and 12-layer 512-wide transformer blocks with 8 attention heads for text encoder and are trained for 30 epochs. For NAS, we first train the supernet for 30 epochs to select the best architecture, and then train the selected model from scratch for another 30 epochs. All experiments are performed on 16 NVIDIA Tesla V100 GPUs.", + "bbox": [ + 75, + 503, + 470, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Zero-shot Image Classification", + "text_level": 1, + "bbox": [ + 76, + 676, + 346, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At each growth step, we evaluate the model on zero-shot image classification task. The downstream datasets include 9 typical settings, Caltech101 [35], CIFAR10 [32], CIFAR100 [32], Describable Textures Dataset (DTD) [12], Flowers102 [47], Food101 [6], OxfordPets [48], SUN39 [5] and ImageNet [14]. We apply prompt ensemble evaluation like CLIP for each dataset. Results are given in Table 3. Our model performs growth process at each growth step except step 1, where the parameter grows from $30.0M$ to $188.4M$ . It outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at the last three growth steps. For example, at the third growth step, our model improves the average top-1 accuracy from $35.9\\%$ to $38.1\\%$ $(+2.2\\%)$ .", + "bbox": [ + 75, + 704, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Zero-shot Image-Text Retrieval", + "text_level": 1, + "bbox": [ + 500, + 502, + 777, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The zero-shot image-text retrieval consists of two subtasks: image-to-text retrieval and text-to-image retrieval. We evaluate the model of each growth step on two retrieval benchmark datasets: Flickr30K [50] and MSCOCO [40]. As shown in Table 4, our model gets better as model grows and data grows. At the last two growth steps, it can be observed that our method has great performance on all metrics. Particularly, our method improves from $49.6\\%$ to $50.8\\% (+1.2\\%)$ for top-1 image-to-text recall on Flickr30K dataset and from $26.8\\%$ to $27.6\\% (+0.8\\%)$ on MSCOCO at the last growth step.", + "bbox": [ + 496, + 527, + 890, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 709, + 653, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we conduct the ablation study on our GrowCLIP. All experiments are built at the growth step 2 and we evaluate the models on zero-shot image classification of ImageNet. To verify the effectiveness of each module in our proposed GrowCLIP, we conduct ablation experiments at some core steps mentioned in Section 3.4 as follows. (i) Parameter inheriting with momentum in phase 1 (PIM@1). (ii) Supernet fine-tuning in phase 1 (ST). (iii) Growth architecture selection in phase 2 (GAS). (iv) Parameter inheriting with momentum in phase 3 (PIM@3). The accuracy on zero-shot image classification of ImageNet is re", + "bbox": [ + 496, + 734, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22184", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d18fa5e133fd627a64fd719114babb9cb41e01ef04456c730d574cda77f6705d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PIM@1STGASPIM@3Para.Acc.
✓✓✓✓116.6M25.7
✓✓✓X116.6M24.5
✓✓XX109.5M22.6
✓XXX109.5M21.1
XXXX109.5M20.8
", + "bbox": [ + 104, + 88, + 442, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89b7ccf15eb81ae9dc0b6ffcaec912bdbe857cbe1e3016868caf6a7292d02c55.jpg", + "table_caption": [ + "Table 5. Ablation analysis of each module in GrowCLIP at Growth Step 2. (Para.: parameter of the model. Acc.: top-1 accuracy (\\%) of zero-shot image classification on ImageNet.)" + ], + "table_footnote": [], + "table_body": "
Para.Acc.
CLIP-ViT-B/16 (TFS)149.6M28.3
CLIP-ViT-L/16 (TFS)431.1M29.9
GrowCLIP-S4 (TFS)188.4M33.3
GrowCLIP-S4 (ours)188.4M36.1
", + "bbox": [ + 84, + 258, + 460, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ported for comparison and the results of the ablation study are shown in Table 5. As can be seen, all above four components have positive effects on the final results. Specifically, the supernet training will influence the grown model without PIM@3, which results in $1.2\\%$ drop. Compared with GAS method, we randomly choose an architecture in the growth space, which has a worse performance from $24.5\\%$ to $22.6\\%$ . With ST, we observe a remarkable improvement of $1.5\\%$ . GrowCLIP with PIM@1 solves the problem of local minimum dilemma and brings a $0.3\\%$ improvement.", + "bbox": [ + 75, + 398, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis", + "text_level": 1, + "bbox": [ + 76, + 560, + 179, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Why can GrowCLIP improve the performance?", + "text_level": 1, + "bbox": [ + 96, + 583, + 452, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared with baseline models, the gains of our GrowCLIP come from two folds: (i) GrowCLIP can search the architecture that matches the current data with Growth architecture selection (GAS). (ii) Parameter inheriting with momentum (PIM) benefits GrowCLIP from yesterday's model. To demonstrate their effectiveness, we train the architecture selected by GrowCLIP at step 4 from scratch, and the result is shown in Table 6. Compared with ViT-B/16, the growth architecture selection can bring $5\\%$ improvement on zero-shot image classification of ImageNet. Furthermore, it is even better than ViT-L/16, which is two and a half times as big as our model. More experiments have proved the effectiveness of GAS in the appendix. Comparing our GrowCLIP-S4 with GrowCLIP-S4 (TFS), we observe a remarkable improvement of $2.8\\%$ with PIM.", + "bbox": [ + 75, + 598, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Why can GrowCLIP improve the training efficiency?", + "text_level": 1, + "bbox": [ + 96, + 825, + 467, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our GrowCLIP can improve training efficiency, due to GAS and PIM, and the result is shown in Table 7. (i) Compared with the fixed architecture method, GrowCLIP has a smaller model with the GAS, so we reduce training costs", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/028dfca84c8aa208e45199e92aca4aef48988266632af37820f259d0e523dddc.jpg", + "table_caption": [ + "Table 6. Ablation analysis of GrowCLIP-S4." + ], + "table_footnote": [], + "table_body": "
Para.Acc.EpochCost
CLIP-S2(TFS)149.6M20.930336
GrowCLIP-S2 (ours)116.6M25.730288
GrowCLIP-S4 (TFS)188.4M33.3301230
GrowCLIP-S4 (ours)188.4M33.39722
NAS-S4325.9M34.2605280
GrowCLIP-S4 (ours)188.4M36.1301583
", + "bbox": [ + 506, + 88, + 887, + 223 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37dbeed2f385563ca656283f6f47b267d017df95916fede62539f1afb81f8130.jpg", + "table_caption": [ + "Table 7. Efficiency analysis. (Cost: GPU hours)" + ], + "table_footnote": [], + "table_body": "
Para.Acc.
GrowCLIP-S1431.1M14.6
GrowCLIP-S2431.2M26.2
", + "bbox": [ + 506, + 260, + 885, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. The architecture of GrowCLIP-S1 with bigger initialized model (ViT-L/16).", + "bbox": [ + 500, + 332, + 890, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "at the early steps (48 GPU hours saved at step 2). At the later steps, GrowCLIP uses fewer epochs to achieve good performance because of PIM. As shown in Figure 1, GrowCLIP has far outperformed CLIP (TFS) only after supernet training. To demonstrate the effectiveness of PIM, we compare GrowCLIP-S4 with GrowCLIP-S4 (TFS). GrowCLIP only uses 9 epochs to achieve comparable performance with PIM. (ii) Compared with the NAS method, GrowCLIP inherits the parameters of the old model with momentum, which only takes 2 epochs to fine-tune the supernet and 2 epochs to train the supernet. In contrast, one-shot NAS methods consume about 3.3X resources.", + "bbox": [ + 496, + 386, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Can GrowCLIP find the relationship between model and data? When the model scale is sufficient large, there is no need to enlarge the model. We conduct the experiment started from ViT-L/16. And the result in Table 8 shows that the larger model startpoint hardly grows at step 2, which means that our GrowCLIP can find the relationship between the model and data. In our online learning scenario, the initial architecture of the model is not important due to our growth pipeline. If the model scale is smaller at the beginning, GrowCLIP will select a larger enough model for growth. In contrast, if the model scale is too large initially, GrowCLIP will stop growing.", + "bbox": [ + 496, + 568, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What is the influence when growth happens in different dimensions? As shown in Table 9, different growth dimensions have different influences on the performance. By analyzing the experimental results, we conclude 4 laws, which can provide a reference for the design of cross-modal model architecture: (i) Enlarging the smaller model bring a greater improvement than the bigger one. For example, when the growth appears in transformer blocks in the image encoder, the accuracy can gain $0.93\\%$ improvement at step 2 but only $0.12\\%$ at step 4. (ii) We need to leave the image", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22185", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5691aa04d4e705a3b23dd00643a0e579d846fe8a0063400ed3ac04c164c876a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Step 2Step 3Step 4
DimensionPara.Acc.Para.Acc.Para.Acc.
+4 bI+7.1M+0.93+19.7M+0.46+19.7M+0.12
+4 hI+20.1M+0.34+67.3M+0.12+105.1M+0.03
+2 lI+0.1M+0.84+0.1M+0.07+0.1M+0.10
+4 bT+3.2M+0.01+12.6M+0.08+12.6M+0.01
+4 hT+27.0M+0.22+45.0M+0.01+69.5M-0.04
+4 bS+7.1M+1.03+19.7M+0.19+19.7M-0.07
", + "bbox": [ + 78, + 89, + 470, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 9. The ablation of growth space corresponding to different dimensions. $(+4b^{I}$ : the number of transformer blocks in image encoder is increased by 4)", + "bbox": [ + 75, + 243, + 468, + 285 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "encoder with more capacity. Compared with text encoder, the improvement of growing the image encoder is more significant at each step. (iii) The convolutional layers is a very important component in the image encoder, which can help the model to better extract visual feature. (iv) The shared encoder can enhance the degree of cross-modal fusion and bring a remarkable improvement. However, do not use a large number of transformers in the shared encoder, or the model become redundant with no noticeable performance improvement.", + "bbox": [ + 75, + 311, + 468, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 478, + 194, + 493 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we study the online learning case in cross-modal pre-trained models. The small-scale architecture may limit the final performance due to the limited parameters when data grows. Based on this real-world scenario, we propose a data-driven automatic model growing algorithm for language-image pre-training - GrowCLIP, which upgrades the model size when abundant new data arrives. We adopt parameter inheriting with momentum (PIM) to utilize the knowledge in yesterday's model and seek out the optimal growth architecture at each growth step with growth architecture selection (GAS). GrowCLIP improves the performance on downstream tasks, including zero-shot image classification and retrieval. The effectiveness of each component of GrowCLIP is proved on ablation studies. And we provide a reference for the design of cross-modal model architecture.", + "bbox": [ + 75, + 503, + 468, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations In this work, we only prove the effectiveness of GrowCLIP with public pre-training dataset - CC12M. In future works, we will extend our method to a real-world scenario, where the VLP model keeps training with constantly updated data crawled from the web.", + "bbox": [ + 75, + 746, + 468, + 821 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 76, + 837, + 228, + 853 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Key R&D Program of China under Grant No.2020AAA0109700, Guangdong Outstanding Youth Fund (Grant No.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "2021B1515020061), Shenzhen Science and Technology Program (Grant No.RCYX20200714114642083), Shenzhen Fundamental Research Program(Grant No. JCYJ20190807154211365), Nansha Key RD Program under Grant No.2022ZD014 and Sun Yat-sen University under Grant No.22lgqb38 and 76160-12220011. We thank MindSpore for the partial support of this work, which is a new deep learning computing framework $^{2}$ .", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www.mindspore.cn/", + "bbox": [ + 517, + 886, + 669, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22186", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems.", + "[2] R. Aljundi, F. Babiloni, M. Elhoseiny, M. Rohrbach, and T. Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European conference on computer vision, 2018.", + "[3] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Zitnick, and D. Parikh. Vqa: Visual question answering. In International Conference on Computer Vision, 2015.", + "[4] J. Ash and R. Adams. On warm-starting neural network training. 2020.", + "[5] A. Barriuso and A. Torralba. Notes on image annotation. Preprint arXiv:1210.3448, 2012.", + "[6] L. Bossard, M. Guillaumin, and L. Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, 2014.", + "[7] P. Buzzega, M. Boschini, A. Porrello, D. Abati, and S. Calderara. Dark experience for general continual learning: a strong, simple baseline. In Advances in Neural Information Processing Systems, 2020.", + "[8] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019.", + "[9] S. Changpinyo, P. Sharma, N. Ding, and R. Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Computer Vision and Pattern Recognition, 2021.", + "[10] M. Chen, H. Peng, J. Fu, and H. Ling. Autoformer: Searching transformers for visual recognition. In International Conference on Computer Vision, 2021.", + "[11] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020.", + "[12] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Computer Vision and Pattern Recognition, 2014.", + "[13] M. Cornia, L. Baraldi, and R. Cucchiara. Show, control and tell: A framework for generating controllable and grounded captions. In Computer Vision and Pattern Recognition, 2019.", + "[14] J. Deng. A large-scale hierarchical image database. 2009.", + "[15] J. Devlin, M. Chang, K. Lee, and K. Toutanova. BERT: Pretraining of Deep Bidirectional Transformers for Language Understanding. In North American Chapter of the Association for Computational Linguistics, 2019.", + "[16] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.", + "[17] H. Grabner and H. Bischof. On-line boosting and vision. In Computer Vision and Pattern Recognition, 2006.", + "[18] J. Grill, F. Strub, F. Altché, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Ghesh" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. 2020.", + "[19] Z. Guo, B. Avila Pires, B. Piot, J. Grill, F. Altché, R. Munos, and M. Azar. Bootstrap latent-predictive representations for multitask reinforcement learning. In International Conference on Machine Learning, 2020.", + "[20] Z. Guo, X. Zhang, H. Mu, W. Heng, Z. Liu, Y. Wei, and J. Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, 2020.", + "[21] L Zhou B Xiao C. Noel Y Cheng R Xu S Chang H, You and Yuan L. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision, pages 69-87. Springer, 2022.", + "[22] R. Hadsell, D. Rao, A. Rusu, and R. Pascanu. Embracing change: Continual learning in deep neural networks. Trends in Cognitive Sciences, 2020.", + "[23] E. Hazan et al. Introduction to online convex optimization. Foundations and Trends® in Optimization, 2016.", + "[24] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In Conference on Computer Vision and Pattern Recognition, 2022.", + "[25] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In Computer Vision and Pattern Recognition, 2020.", + "[26] X. Hu, Z. Gan, J. Wang, Z. Yang, Z. Liu, Y. Lu, and L. Wang. Scaling up vision-language pre-training for image captioning. Preprint arXiv:2111.12233, 2021.", + "[27] X. Hu, X. Yin, K. Lin, L. Wang, L. Zhang, J. Gao, and Z. Liu. Vivo: Surpassing human performance in novel object captioning with visual vocabulary pre-training. 2020.", + "[28] C. Jia, Y. Yang, Y. Xia, Y. Chen, Z. Parekh, H. Pham, Q. Le, Y. Sung, Z. Li, and T. Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021.", + "[29] J. Kaplan, S. McCandlish, T. Henighan, T. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei. Scaling laws for neural language models. Preprint arXiv:2001.08361, 2020.", + "[30] W. Kim, B. Son, and I. Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, 2021.", + "[31] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. National Academy of Sciences, 2017.", + "[32] A. Krizhevsky, G. Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "[33] Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, and R. Soricut. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. In International Conference on Learning Representations, 2020.", + "[34] C Li, B. Zhuang, G. Wang, X. Liang, X. Chang, and Y. Yang. Automated progressive learning for efficient training of vision transformers. In Computer Vision and Pattern Recognition, 2022." + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22187", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] F. Li, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 2006.", + "[36] J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022.", + "[37] L. Li, M. Yatskar, D. Yin, C. Hsieh, and K. Chang. Visualbert: A simple and performant baseline for vision and language. Preprint arXiv:1908.03557, 2019.", + "[38] X. Li, X. Yin, C. Li, P. Zhang, X. Hu, L. Zhang, L. Wang, H. Hu, L. Dong, F. Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, 2020.", + "[39] Y. Li, F. Liang, L. Zhao, Y. Cui, W. Ouyang, J. Shao, F. Yu, and J. Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. Preprint arXiv:2110.05208, 2021.", + "[40] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014.", + "[41] C. Liu, B. Zoph, M. Neumann, J. Shlens, W. Hua, L. Li, F. Li, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. In European conference on computer vision, 2018.", + "[42] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. Preprint arXiv:1806.09055, 2018.", + "[43] W. Liu, S. Chen, L. Guo, X. Zhu, and J. Liu. Cptr: Full transformer network for image captioning. Preprint arXiv:2101.10804, 2021.", + "[44] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Preprint arXiv:1907.11692, 2019.", + "[45] I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016.", + "[46] J. Lu, V. Goswami, M. Rohrbach, D. Parikh, and S. Lee. 12-in-1: Multi-task vision and language representation learning. In Computer Vision and Pattern Recognition, 2020.", + "[47] M. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Computer Vision, Graphics & Image Processing, 2008.", + "[48] O. Parkhi, A. Vedaldi, A. Zisserman, and C. Jawahar. Cats and dogs. In Computer Vision and Pattern Recognition, 2012.", + "[49] H. Peng, H. Du, H. Yu, Q. Li, J. Liao, and J. Fu. Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. In Advances in Neural Information Processing Systems, 2020.", + "[50] B. Plummer, L. Wang, C. Cervantes, J. Caicedo, J. Hockenmaier, and S. Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-tosentence models. In International Conference on Computer Vision, 2015." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] R. Pogodin and T. Lattimore. On first-order bounds, variance and gap-dependent bounds for adversarial bandits. In Uncertainty in Artificial Intelligence, 2020.", + "[52] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021.", + "[53] E. Real, A. Aggarwal, Y. Huang, and Q. Le. Regularized Evolution for Image Classifier Architecture Search. In AAAI Conference on Artificial Intelligence, 2019.", + "[54] E. Real, S. Moore, A. Selle, S. Saxena, Y. Suematsu, J. Tan, Q. Le, and A. Kurakin. Large-scale evolution of image classifiers. In International Conference on Machine Learning, 2017.", + "[55] H Xu J. Han T. Tuytelaars Z. Li S. Yan, L. Hong and X. He. Generative negative text replay for continual vision-language pretraining. In European Conference on Computer Vision, pages 22-38. Springer, 2022.", + "[56] A. Saffari, C. Leistner, J. Santner, M. Godec, and H. Bischof. On-line random forests. In International Conference on Computer Vision workshops, 2009.", + "[57] S. Shalev-Shwartz et al. Online learning and online convex optimization. Foundations and Trends® in Machine Learning, 2012.", + "[58] S. Song and B. Choi. A study on continuous follow-the-leader (ftl) gaits: an effective walking algorithm over rough terrain. Mathematical biosciences, 1989.", + "[59] T Srinivasan, T Chang, L. Alva, G Chochlakis, M Rostami, and J Thomason. Climb: A continual learning benchmark for vision-and-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track.", + "[60] M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard, and Q. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Computer Vision and Pattern Recognition, 2019.", + "[61] B. Thomee, D. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland, D. Borth, and L. Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016.", + "[62] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, and I. Polosukhin. Attention Is All You Need. In Neural Information Processing Systems, 2017.", + "[63] Haonan Wang, Minbin Huang, Runhui Huang, Lanqing Hong, Hang Xu, Tianyang Hu, Xiaodan Liang, and Zhenguo Li. Boosting visual-language models by exploiting hard samples. arXiv preprint arXiv:2305.05208, 2023.", + "[64] S. Xie, H. Zheng, C. Liu, and L. Lin. SNAS: Stochastic Neural Architecture Search. In International Conference on Learning Representations, 2019.", + "[65] L. Yao, R. Huang, L. Hou, G. Lu, M. Niu, H. Xu, X. Liang, Z. Li, X. Jiang, and C. Xu. Filip: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022.", + "[66] Y. You, J. Li, S. Reddi, J. Hseu, S. Kumar, S. Bhojanapalli, X. Song, J. Demmel, K. Keutzer, and C. Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22188", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In International Conference on Learning Representations, 2020.", + "[67] F. Yu, J. Tang, W. Yin, Y. Sun, H. Tian, H. Wu, and H. Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3208-3216, 2021.", + "[68] J. Yu, Z. Wang, V. Vasudevan, L. Yeung, M. Seyedhosseini, and Y. Wu. Coca: Contrastive captioners are image-text foundation models. Preprint arXiv:2205.01917, 2022.", + "[69] X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022.", + "[70] B. Zoph, V. Vasudevan, J. Shlens, and Q. Le. Learning Transferable Architectures for Scalable Image Recognition. In Computer Vision and Pattern Recognition, 2018." + ], + "bbox": [ + 78, + 90, + 468, + 332 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22189", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_model.json b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3502091d8395ed5a5491f6c8e883542259467a9 --- /dev/null +++ b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_model.json @@ -0,0 +1,2347 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.765, + 0.177 + ], + "angle": 0, + "content": "GrowCLIP: Data-aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.203, + 0.806, + 0.293 + ], + "angle": 0, + "content": "Xinchi Deng\\(^{1}\\) Han Shi\\(^{2}\\) Runhui Huang\\(^{1}\\) Changlin Li\\(^{3}\\) Hang Xu\\(^{2}\\) \nJianhua Han\\(^{2}\\) James Kwok\\(^{4}\\) Shen Zhao\\(^{1*}\\) Wei Zhang\\(^{2}\\) Xiaodan Liang\\(^{1}\\) \n\\(^{1}\\)Sun Yat-sen University \\(^{2}\\)Huawei Noah's Ark Lab \n\\(^{3}\\)University of Technology Sydney \\(^{4}\\)The Hong Kong University of Science and Technology \ndengxch5@mail2.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.473, + 0.822 + ], + "angle": 0, + "content": "Cross-modal pre-training has shown impressive performance on a wide range of downstream tasks, benefiting from massive image-text pairs collected from the Internet. In practice, online data are growing constantly, highlighting the importance of the ability of pre-trained model to learn from data that is continuously growing. Existing works on cross-modal pre-training mainly focus on training a network with fixed architecture. However, it is impractical to limit the model capacity when considering the continuously growing nature of pre-training data in real-world applications. On the other hand, it is important to utilize the knowledge in the current model to obtain efficient training and better performance. To address the above issues, in this paper, we propose GrowCLIP, a data-driven automatic model growing algorithm for contrastive language-image pre-training with continuous image-text pairs as input. Specifically, we adopt a dynamic growth space and seek out the optimal architecture at each growth step to adapt to online learning scenarios. And the shared encoder is proposed in our growth space to enhance the degree of cross-modal fusion. Besides, we explore the effect of growth in different dimensions, which could provide future references for the design of cross-modal model architecture. Finally, we employ parameter inheriting with momentum (PIM) to maintain the previous knowledge and address the issue of the local minimum dilemma. Compared with the existing methods, GrowCLIP improves \\(2.3\\%\\) average top-1 accuracy on zero-shot image classification of 9 downstream tasks. As for zero-shot image retrieval, GrowCLIP can improve \\(1.2\\%\\) for top-1 image-to-text recall on Flickr30K dataset." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.326, + 0.859, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.52, + 0.895, + 0.591 + ], + "angle": 0, + "content": "Figure 1. Top-1 accuracy \\((\\%)\\) of zero-shot image classification on ImageNet of GrowCLIP and baselines during training at step 4, where the horizontal dotted lines mean the process of supernet training. Our GrowCLIP has the best performance and is more efficient compared with other baselines." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.616, + 0.633, + 0.632 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Recently, large-scale pre-trained models illustrate the potential performance among different fields including computer vision (CV) [24, 25, 11] and natural language processing (NLP) [15, 44, 33]. Cross-modal models like CLIP [52], ALIGN [28], FILIP [65], BLIP [36] also demonstrate remarkable success across various vision-language downstream tasks. Note that these models are usually built in dual-stream architectures, which consist of the image and text encoders to extract the feature of image and text inputs, respectively. The alignment between these two features is then performed under the contrastive objective to learn the alignment between different modals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "However, training these cross-modal models requires a large amount of image-text pairs collected from the Internet, e.g., 400M image-text pairs for CLIP [52] and 300M for FILIP [65]. Most existing cross-modal methods [52, 28, 65, 39, 36, 68, 63, 1] directly train the final large-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.224, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22178" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.452, + 0.15 + ], + "angle": 0, + "content": "
CC12M10%20%50%100%
CLIP-0.5-ViT-B/166.611.119.025.7
CLIP-ViT-B/166.511.620.928.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.153, + 0.47, + 0.224 + ], + "angle": 0, + "content": "Table 1. Result of different model sizes with different data sizes: Top-1 accuracy (\\%) of zero-shot image classification on ImageNet. The depth and the width of 0.5-ViT-B/16 are both one half of ViT-B/16. The relative performance ranking is dependent on the size of the training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.457 + ], + "angle": 0, + "content": "scale model using the completed image-text pair datasets collected at some point in time. However, it is impractical to regard cross-modal pre-training as \"disposable\" training without considering the continuously growing nature of pre-training data or domain in real-world applications. For instance, mountains of new knowledge is generated on the Internet every day, which can be used to further improve the performance of existing pre-trained cross-modal models. Therefore, in this paper, we consider the training of large-scale cross-modal pre-trained models as an online learning problem [17, 56, 57] with continuous image-text pairs as input. Different from the standard continual learning setting [59, 55, 22, 31, 2, 7], the previous image-text pairs can also be achieved in our setting since we assume the size of training data is gradually increased over time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.852 + ], + "angle": 0, + "content": "One tough problem in the online learning scenario is that the model capacity should be related to the size of training data [16, 69, 29, 26]. For example, it's observed that large ViT models perform worse than ResNets when pre-trained on small datasets, while the result is opposite when they are both pre-trained on larger datasets [16]. To verify the relationship between model capability and data size, we split Conceptual 12M (CC12M) dataset [9] and test the performance of ViT-B/16 with different scales [52]. As shown in Table 1, the relative performance ranking is dependent on the size of the training data. Given a smaller dataset (e.g., \\(10\\%\\) CC12M), the performance of 0.5-ViT-B/16 is comparable with ViT-B/16. In contrast, when data is sufficient (e.g., \\(100\\%\\) CC12M), the performance of ViT-B/16 is better than the smaller model. Thus, the unchanged model is not practical for this real-time learning setting and it's still an open issue on how to modify and train our model with the growing data. Besides, how to efficiently make use of knowledge of previous model when new data is coming remains an open problem. One direct solution is to fine-tune the model with the updated training dataset. However, training with previous pre-trained parameters of the same model will deteriorate the performance [4] due to the parameter inheriting issue. As shown in Figure 1, the CLIP training with pre-training (TWP) have worse performance than the one training from scratch (TFS)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To address the above issues, we propose a data-aware automatic model growing method (denoted as GrowCLIP) for large-scale contrastive language-image pre-training, which" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.348 + ], + "angle": 0, + "content": "performs a model growth process considering the gradually increased pre-training data. Specially, when training data grows dynamically, we adopt different cross-modal network architectures via the customized neural architecture search (NAS) to make the network pre-training more efficient. Different from traditional NAS approaches [70, 53, 64], we introduce a cross-modal customized NAS by defining a dynamic search space named growth space, which is enlarged when more data comes, and proposing a shared encoder search space to enhance the degree of cross-modal fusion. To utilize the architecture at the previous step more efficiently, the parameters of the new architecture are also inherited from the old one with momentum to maintain the previous knowledge and address the issue of local minimum dilemma. Finally, growth architecture selection procedure is performed to select the optimal model architecture at each step, considering the performance and model size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Experiments are conducted by averagely dividing the Conceptual 12M (CC12M) into 4 growth steps under the online learning setting. As depicted in Table 1, compared with the existing methods, our GrowCLIP has the best performance and is more efficient. Specially, experimental results show that our GrowCLIP can improve up to \\(2.3\\%\\) average top-1 accuracy on zero-shot image classification of 9 downstream tasks compared with the existing methods. As for zero-shot image-text retrieval, GrowCLIP has a \\(1.2\\%\\) improvement for top-1 image-to-text recall on Flickr30K [50] dataset and \\(0.8\\%\\) on MSCOCO [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.62 + ], + "angle": 0, + "content": "To summarize, the contributions of this paper are listed as follows: (i) To adapt to the growing data scenario, we propose a data-aware automatic model growing method, named GrowCLIP. (ii) We provide some insights for the design of cross-modal model architecture. (iii) The extensive experiment results illustrate the effectiveness of our approach on zero-shot classification and retrieval tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.642, + 0.65 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Vision-Language Pre-training Models Inspired by the success of pre-training in computer vision (CV) and natural language processing (NLP), a boosting number of research works in the domain of vision-language pretraining (VLP) has recently surged to pursue a unified multi-modality representation. Vision-language pretraining trains the model on large-scale image-text pairs to improve performance of downstream vision and language tasks including image classification [14, 35, 32], retrieval [50, 40], Image Captioning (IC) [43, 13, 27, 26], Visual Question and Answer (VQA) [3, 46, 67] and so on. The VLP model can be categorized into two categories: (i) The dual-stream model, e.g., CLIP [52], ALIGN [28], FILIP [65], processes visual and textual tokens separately with two parallel streams to acquire the representation and then fuse them through interactive module. (ii) The single-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22179" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "stream model, e.g., Oscar [38], ViLT [30], VisualBERT [37], directly combines the extracted visual and textual tokens and feeds them into the transformer-based model. The public pre-training datasets of VLP model include available datasets like YFCC100M [61] and CC12M [9]. CLIP [52], ALIGN [28], FILIP [65] and LEMON [26] adopt large-scale image-text pairs crawled from the web to show more powerful models. Different from these works that treat this process as offline learning, we investigate the online learning case in vision-language pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.514 + ], + "angle": 0, + "content": "Online Learning Online learning is a problem setting of machine learning in which data becomes available in a sequential order [17, 56, 57]. Different from continual learning [22, 31, 2, 7], online learning can reach previous data all the time. As opposed to batch learning techniques which generate the best model by learning on the entire training dataset at once, online learning is used to update our predictor for future data at each growth step. Follow the leader [58] is the simplest learning rule which assumes the hypothesis that the leader has the least loss overall past rounds. Follow the regularised leader [51] is a modification variant and obtain better regret bounds by learning a regularisation function. Online convex optimization [23] is a general algorithm framework for decision making which targets convex optimization. To adapt online learning scenario, we propose a data-aware automatic model growing algorithm that grows the model with increasing training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.473, + 0.758 + ], + "angle": 0, + "content": "Neural Architecture Search Neural Architecture Search (NAS) aims to automatically design the state-of-the-art neural networks. Early works [70, 54, 60] are based on reinforcement learning and evolutionary algorithms. In the later work, various techniques are used to reduce the computational complexity. Gradient-based methods [42, 41] regard the network architecture as a set of parameters and adopt back-propagation algorithm to optimize them. One-shot NAS methods [49, 10, 20] train a supernet to avoid training each subnet from scratch. Unfortunately, one-shot NAS methods have to spend lots of resources to train the supernet. In contrast, the method we propose inherits the parameters of the old model with momentum, which only takes few epochs to train the supernet. Furthermore, GrowCLIP can modify the neural architecture to adapt to the size of the training data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.214, + 0.788 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this section, we describe the online learning problem setting and its challenges for cross-modal pre-training (Section 3.1). To adapt the growth data, we propose a dynamic growth space, including a proposed shared encoder (Section 3.2). To rise to the challenge of the local minimum dilemma, we propose parameter inheriting with momentum (Section 3.3). The whole pipeline of our method is illus" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.642, + 0.106 + ], + "angle": 0, + "content": "trated in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.116, + 0.783, + 0.133 + ], + "angle": 0, + "content": "3.1. Problem Setting and Challenges" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.895, + 0.291 + ], + "angle": 0, + "content": "Problem Setting The target of online learning is obtaining the optimal model at each growth step of the data. Let \\(\\mathcal{D}^{t - 1} = \\{\\pmb{x}_i^I,\\pmb{x}_i^T\\}_{i = 1}^k\\) be the training data at growth step \\(t - 1\\), where \\(\\pmb{x}^I\\) and \\(\\pmb{x}^T\\) are image and text samples and \\(k\\) is the training data size. At the next growth step \\(t\\), the enlarged training dataset \\(\\mathcal{D}^t\\) satisfies \\(\\mathcal{D}^{t - 1}\\subseteq \\mathcal{D}^t\\) due to the incoming data. Given an optimal model from growth step \\(t - 1\\), our aim is to achieve the optimal model (with optimal architecture \\(\\psi^{*}\\) and parameters \\(\\omega^{*}\\)) based on the enlarged dataset \\(\\mathcal{D}^t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.306, + 0.892, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\psi^ {*}, \\omega^ {*} \\right\\} = \\underset {\\psi , \\omega} {\\arg \\min } \\mathcal {L} (\\psi , \\omega ; \\mathcal {D} ^ {t}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.895, + 0.474 + ], + "angle": 0, + "content": "Challenge of Data Growth As discussed in Table 1, the size of the training dataset is related to the optimal architecture selection. Given a smaller dataset, the small-scale architecture has comparable performance to large-scale architecture but fewer parameters are deployed. When data grows, the small-scale architecture may limit the final performance due to the limited parameters. However, few works notice the effect of data size in online learning settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.895, + 0.626 + ], + "angle": 0, + "content": "Challenge of Parameter Inheriting As investigated, the performance of model trained from scratch is much better than trained with pre-training. This is probably caused by the influence of the inheriting parameters trained with the previous smaller dataset, which makes it difficult for the model to escape from the local minima in the grown dataset. However, training from scratch is costly due to multiple growth steps in online learning. The balance between training from scratch and training with pre-training is worthy of consideration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.779 + ], + "angle": 0, + "content": "To solve the above problems, we propose an automatic model growing method - GrowCLIP (shown in Figure 2), which can modify the neural architecture at different growth steps in online learning case. To adapt to the growing data scenario, we propose a growth space for each growth step, where the scale of architecture candidates among this space is larger when data grows. To alleviate the parameter inhering issue with efficiency, we perform parameter inhering with momentum such that the trade off between training from scratch and training with pre-training is balanced." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.649, + 0.804 + ], + "angle": 0, + "content": "3.2. Growth Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Model Architecture Following the basic neural architecture in CLIP [52], we adopt a dual-stream model with image encoder \\( f(\\cdot) \\) and text encoder \\( g(\\cdot) \\). To adapt the cross-modality scenarios, we propose a shared encoder \\( h(\\cdot) \\), in which the image and text share the transformers except layernorm layers. The shared encoder not only can reduce" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22180" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.526, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.895, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.381, + 0.893, + 0.41 + ], + "angle": 0, + "content": "Figure 2. The overview of GrowCLIP. As shown in Figure (a), when the data grows from growth step \\( t - 1 \\) to growth step \\( t \\), our model grows adaptively as well. Figure (b) illustrates the detail of model growing and we simplify the representation of model architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.516 + ], + "angle": 0, + "content": "model parameters, but also enhance the degree of cross-modal fusion. Given an image \\( \\boldsymbol{x}_i^I \\) and a text \\( \\boldsymbol{x}_j^T \\), the model outputs the image representations \\( I_i = h(f(\\boldsymbol{x}_i^I)) \\) and text representations \\( T_j = h(g(\\boldsymbol{x}_j^T)) \\). The similarity score computation of between image \\( \\boldsymbol{x}_i^I \\) and text \\( \\boldsymbol{x}_j^T \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.524, + 0.469, + 0.544 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = h \\left(f \\left(\\boldsymbol {x} _ {i} ^ {I}\\right)\\right) ^ {\\top} h \\left(g \\left(\\boldsymbol {x} _ {j} ^ {T}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.58 + ], + "angle": 0, + "content": "where \\( s_{i,j} \\) denotes the similarity of the \\( i \\)-th image to the \\( j \\)-th text." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.583, + 0.47, + 0.659 + ], + "angle": 0, + "content": "In each training batch, we sample \\(n\\) image-text pairs \\(\\{\\pmb{x}_i^I,\\pmb{x}_i^T\\}_{i = 1}^n\\) from the dataset. For image \\(\\pmb{x}_i^I\\) in the image-text pair \\(< x_{i}^{I},x_{i}^{T}>\\) \\(x_{i}^{T}\\) is its positive, while the other texts will be used as in-batch negatives. The image-to-text contrastive loss \\(\\mathcal{L}_i^I\\) for \\(\\pmb{x}_i^I\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.668, + 0.469, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} ^ {I} \\left(\\boldsymbol {x} _ {i} ^ {I}, \\left\\{\\boldsymbol {x} _ {j} ^ {T} \\right\\} _ {j = 1} ^ {n}\\right) = - \\frac {1}{n} \\log \\frac {\\exp \\left(s _ {i , i}\\right)}{\\sum_ {j} \\exp \\left(s _ {i , j}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.711, + 0.434, + 0.728 + ], + "angle": 0, + "content": "Similarly, the text-to-image contrastive loss for \\( \\pmb{x}_i^T \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.737, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} ^ {T} \\left(\\boldsymbol {x} _ {i} ^ {T}, \\left\\{\\boldsymbol {x} _ {j} ^ {I} \\right\\} _ {j = 1} ^ {n}\\right) = - \\frac {1}{n} \\log \\frac {\\exp \\left(s _ {i , i}\\right)}{\\sum_ {j} \\exp \\left(s _ {i , j}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.811 + ], + "angle": 0, + "content": "Totally, the contrastive loss of this mini-batch can be represented by \\(\\mathcal{L} = \\frac{1}{2}\\sum_{i=1}^{a}\\left(\\mathcal{L}_i^I + \\mathcal{L}_i^T\\right)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Growth Space of image encoder For the image encoder, we adopt a mixed architecture of convolutional layer and transformer block [62]. There are three variable factors in our growth space design: (i) the number of convolutional layers \\( l^I \\); (ii) the number of transformer blocks \\( b^I \\); (iii) the number of transformer heads \\( h^I \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.434, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Growth Space of text encoder The text encoder part is a transformer-based architecture, whose growth space includes two variable factors: (i) the number of transformer blocks \\( b^{T} \\) and (ii) the number of transformer heads \\( h^{T} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.539 + ], + "angle": 0, + "content": "Growth Space of shared encoder The shared encoder is a transformer-based architecture, whose growth space includes the number of transformer blocks \\( b^{S} \\)." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.548, + 0.895, + 0.607 + ], + "angle": 0, + "content": "
Image encoderText encoderShared encoder
Transformer blocks{bIt-1, bIt-1+4}{bTt-1, bTt-1+4}{bSt-1, bSt-1+4}
Transformer heads{hIt-1, hIt-1+4}{hTt-1, hTt-1+4}-
Convolutional layers{lIt-1, lIt-1+2}--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.617, + 0.894, + 0.659 + ], + "angle": 0, + "content": "Table 2. The growth space \\( \\Phi \\) of GrowCLIP at the growth step t. \\( {b}_{t - 1}^{I},{b}_{t - 1}^{T},{b}_{t - 1}^{S},{h}_{t - 1}^{I},{h}_{t - 1}^{T},{l}_{t - 1}^{I} \\) are the hyperparameters of model architecture at growth step \\( t - 1 \\) ." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.766 + ], + "angle": 0, + "content": "The detailed descriptions of growth space for three encoders are shown in Table 2. At each new growth step, the variable factors are enlarged to adapt the growth data. Specifically, the numbers of transformer blocks and transformer heads are both increased by 4, and the number of convolutional layers is increased by 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.829, + 0.789 + ], + "angle": 0, + "content": "3.3. Parameter inheriting with momentum" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Some previous papers utilize the training results of the old model and transfer the knowledge to the new model [18, 19, 25, 4, 34]. Inspired by the above methods, we propose parameter inheriting with momentum (PIM) to avoid local minimum dilemma. The illustration of our parameter inheriting with momentum is shown in Figure 3. To preserve the knowledge from the old model, we extend the size" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.47, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.471, + 0.286 + ], + "angle": 0, + "content": "Figure 3. Take image encoder for example of our proposed parameter inheriting with momentum. As for the initialization of the enlarged model, we preserve the existing parameters of the old model for exploitation and combine it with random initialization for exploration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.425 + ], + "angle": 0, + "content": "of the old model to the enlarged model and maintain the parameters in existing parts. As for the extended layers, we copy the parameters from the last layers directly. \\(\\omega_{old}\\) denotes the parameters of extension of the old model. Besides the above parameter inheriting term, we also add a re-initialization term with randomly initialized parameters \\(\\omega_{rand}\\) to avoid the parameter inheriting issue. The final initialized supernet is a combination of the above two terms as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.427, + 0.469, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\omega_ {\\text {i n h e r i t}} = \\beta \\cdot \\omega_ {\\text {o l d}} + \\gamma \\cdot \\omega_ {\\text {r a n d}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.522 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\gamma\\) are two hyperparameters related to exploitation and exploration. Additionally, the parameters of new transformer heads are randomly initialized, which makes that different heads in transformer are orthogonal and the new heads learn new features from different perspectives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.24, + 0.545 + ], + "angle": 0, + "content": "3.4. Growth Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.583 + ], + "angle": 0, + "content": "The growth pipeline at growth step \\( t \\) (except the first one) includes three sequential phases as shown in Figure 2 (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.748 + ], + "angle": 0, + "content": "Phase 1: Supernet Fine-tuning As for a new growth step, a supernet \\(\\Psi\\) covering the whole growth space is constructed by enlarging the current model. First, we use the PIM method described in Section 3.3 to initialize the supernet. However, there are some newly added layers in the supernet compared with the old model from the last growth step, which brings a performance gap between the subnets with new layers and others. To make the supernet more robust, we slightly fine-tune the parameters of the full supernet for a few epochs to mitigate the performance gap among subnetworks after PIM initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Phase 2: Growth Architecture Selection Growth Architecture Selection (GAS) is the main phase in our algorithm where we select the optimal architecture in the growth space. The whole GAS process can be divided into two steps: training and searching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.869 + ], + "angle": 0, + "content": "First, to find the optimal architecture \\(\\psi^{*}\\), we need to fully optimize all the subnet in the supernet, which covers all the candidate growth architecture:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.877, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\omega_ {s u p e r} ^ {*} = \\underset {\\omega_ {s u p e r}} {\\arg \\min } \\mathbb {E} _ {\\psi \\in \\Psi} \\left\\{\\mathcal {L} \\left(\\psi , \\omega_ {s u p e r}; \\mathcal {D} ^ {t}\\right) \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Inspired by one-shot neural architecture search [8], we optimize \\(\\omega_{super}\\) in Equation 6 by uniformly sampling subnets \\(\\psi\\) from the growth space \\(\\Phi\\) and update the parameters of the sampled subnets in each iteration. More specifically, at each step, we first randomly sample a subnet \\(\\psi\\) from the growth space \\(\\Psi\\), then optimize the parameter of this subnet by minimizing the contrastive loss between image and text. This can be considered as solving Equation 6 via Monte Carlo sampling. By iterative training, these subnets can be fully optimized in a few epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Second, after training the supernet, we search for the optimal subnet in the growth space \\(\\Phi\\). Here, we consider both the performance and the number of parameters of the architecture, and select the optimal architecture \\(\\psi^{*}\\) using the composite metric:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.329, + 0.892, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\psi^ {*} = \\underset {\\psi \\in \\Psi} {\\arg \\max } \\left\\{\\mathcal {A} (\\psi) + \\alpha \\frac {\\mathcal {N} ^ {t - 1}}{\\mathcal {N} ^ {t}} \\frac {\\mathcal {P} (\\Psi)}{\\mathcal {P} (\\psi)} \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.559 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}(\\cdot)\\) is the top-1 accuracy of test dataset, \\(\\mathcal{N}^t\\) denotes the number of data on growth step \\(t\\), \\(\\mathcal{P}(\\cdot)\\) represents the total amount of model parameters and \\(\\alpha\\) is the trade-off hyperparameter. When we get a large number of new coming data, \\(\\frac{\\mathcal{N}^{t - 1}}{\\mathcal{N}^t}\\) will get small, which means that we focus more on the performance and encourage the growth of models. As our growth space is relatively small, traversal evaluation of all the candidate architectures is affordable. Therefore, we simply test all candidate subnets in the growth space \\(\\Phi\\) using the parameters inherited from the supernet \\(\\Psi\\) and select the optimal one for the next stage via the composite metric of Equation 7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.605 + ], + "angle": 0, + "content": "Phase 3: Selected Model Training To alleviate the influence of supernet training on the grown model, we also use PIM to initialize the parameter \\(\\omega\\) of selected model \\(\\psi^{*}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.621, + 0.891, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\omega} = \\beta \\cdot \\boldsymbol {\\omega} _ {s u p e r} ^ {*} + \\gamma \\cdot \\boldsymbol {\\omega} _ {r a n d}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.891, + 0.674 + ], + "angle": 0, + "content": "After the parameter inheriting, we keep training the selected model, to obtain its final optimal parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.686, + 0.891, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\omega^ {*} = \\underset {\\omega} {\\arg \\min } \\mathcal {L} \\left(\\psi^ {*}, \\omega ; \\mathcal {D} ^ {t}\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.633, + 0.738 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.837 + ], + "angle": 0, + "content": "In this section, we firstly describe the experiment setting and implementation details (Section 4.1). Then we show the results on zero-shot image classification (Section 4.2) and image-text retrieval (Section 4.3). Finally, we conduct ablation study on our GrowCLIP to validate the effectiveness of components (Section 4.4) and analyse (Section 4.5)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.687, + 0.864 + ], + "angle": 0, + "content": "4.1. Experiment Setting" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Benchmark Protocol Conceptual 12M (CC12M) [9] collects about 12 million image-text pairs from the Internet." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22182" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.089, + 0.884, + 0.495 + ], + "angle": 0, + "content": "
Growth stepMethodTraining dataCaltech101CIFAR10CIFAR100DTDFlowers102Food101OxfordPedsSUN397ImageNetAverage
ParameterParameterParameterParameterParameterParameterParameterParameterParameterParameter
Step 1CLIP (TWP)3M149.6M42.837.614.5109.316.121.622.014.020.9
CLIP (TFS)[52]3M149.6M42.837.614.5109.316.121.622.014.020.9
MS-CLIP (TFS)[21]3M129M47.135.514.410.012.917.330.124.718.123.3
CLIP (SAP)[4]3M149.6M42.837.614.5109.316.121.622.014.020.9
NAS-S13M30.0M40.531.213.39.08.915.820.223.213.719.5
GrowCLIP-S1 (ours)3M30.0M40.531.213.39.08.915.820.223.213.719.5
Step 2CLIP (TWP)6M149.6M50.336.116.011.312.421.227.125.817.924.2
CLIP (TFS)[52]6M149.6M58.552.022.312.113.125.033.630.120.929.7
MS-CLIP (TFS)[21]6M129M60.650.018.313.417.42742.432.625.231.8
CLIP (SAP)[4]6M149.6M61.949.021.412.415.524.635.431.622.630.5
NAS-S26M96.9M59.441.818.313.315.826.737.333.824.430.1
GrowCLIP-S2 (ours)6M116.6M57.748.021.213.718.029.539.534.525.732.0
Step 3CLIP (TWP)9M149.6M55.240.119.713.115.723.830.928.920.427.5
CLIP (TFS)[52]9M149.6M63.354.624.514.517.529.536.933.825.333.3
MS-CLIP (TFS)[21]9M129M65.048.222.513.920.631.042.635.729.233.0
CLIP (SAP)[4]9M149.6M67.853.926.715.121.232.941.436.028.235.9
NAS-S39M287.4M66.746.822.214.919.735.446.139.330.135.8
GrowCLIP-S3 (ours)9M168.6M69.653.925.818.418.736.346.241.632.638.1
Step 4CLIP (TWP)12M149.6M57.544.420.512.114.224.134.730.322.228.9
CLIP (TFS)[52]12M149.6M65.957.329.716.318.532.945.536.728.336.8
MS-CLIP (TFS)[21]12M129M65.350.926.316.221.935.950.138.432.037.4
CLIP (SAP)[4]12M149.6M69.860.631.912.421.036.747.040.432.239.1
NAS-S412M325.9M70.853.928.316.721.739.551.241.334.239.7
GrowCLIP-S4 (ours)12M188.4M71.960.728.317.323.342.552.445.536.142.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.504, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Table 3. Top-1 accuracy(%) of zero-shot image classification on 9 datasets. Our GrowCLIP outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at each growth step except step 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.679 + ], + "angle": 0, + "content": "The CC12M is randomly divided into 4 subsets. We select one subset as the training set for the first growth step and adds one subset to the training set at each subsequent growth step such that there are total 4 growth steps in our experiments. At each growth step, we evaluate the corresponding model on the downstream tasks including zero-shot image classification and zero-shot image-text retrieval. The grown architecture of each growth step is shown in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Comparison Methods To demonstrate the effectiveness of our method, we compare it with the following baselines: (i) Training with pre-training (TWP) is a usual method in transfer learning. (ii) Training from scratch (TFS) [52] is used as a basic method to train language-image pre-trained model. (iii) MS-CLIP [21] is a handcrafted framework with a shareable module and modality-specific auxiliary module (the early specialization). (iv) The shrink and perturb (SAP) [4] is a method of parameter inheriting in warm-starting learning. (v) Neural architecture search (NAS) [49, 10] is an important research field of auto machine learning, in which we use the same search space as the growth space of GrowCLIP." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implementation Details A dual-stream model follow" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.86 + ], + "angle": 0, + "content": "ing CLIP [52] without shared encoder is used as basic architecture at the first growth step. The image encoder is stacked with 6 transformer blocks with the following hyperparameters: number of self-attention heads is 6 and the hidden layer size is 384. As for the text encoder, 6 transformer blocks are stacked and the number of the attention heads and the width are 4 and 256 respectively. The resolution of input in the images encoder is resized to 224 and the maximum sequence in the text encoder is limited to 77. The center crop is the only data augmentation method for training. We use the prompts like \\(\\mathrm{CLIP}^1\\) on the zero-shot image classification. We adopt LAMB optimizer [66] with learning rate 0.01 and weight decay 0.0001, and cosine learning rate schedule [45] with a 4000-iteration linear warmup. For the hyperparameters, the proportion of the old model and the new one in the parameter inheriting \\(\\beta = 0.3\\), \\(\\gamma = 0.001\\). The model parameter coefficient in growth architecture selection \\(\\alpha = 0.5\\). We train the model with a total batch size of 1536 for 30 epochs at each growth step, which includes 2 epochs for supernet fine-tuning, 2 epochs for su" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.888, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/openai/CLIP/blob/main/data/prompts.md" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22183" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.091, + 0.904, + 0.455 + ], + "angle": 0, + "content": "
Growth stepMethodFlickr30KMSCOCO
image-to-texttext-to-imageimage-to-texttext-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
Step 1CLIP (TWP)19.645.356.414.333.544.110.726.336.87.019.127.5
CLIP (TFS) [52]19.645.356.414.333.544.110.726.336.87.019.127.5
MS-CLIP (TFS)[21]27.451.863.318.139.250.712.931.943.28.823.533.1
CLIP (SAP) [4]19.645.356.414.333.544.110.726.336.87.019.127.5
NAS-S122.845.257.813.534.246.010.627.437.57.019.928.8
GrowCLIP-S1 (ours)22.845.257.813.534.246.010.627.437.57.019.928.8
Step 2CLIP (TWP)26.651.062.017.439.850.712.830.540.88.322.531.9
CLIP (TFS) [52]31.457.768.121.246.757.117.137.448.710.927.337.4
MS-CLIP (TFS)[21]35.865.175.025.853.164.619.343.055.213.131.542.6
CLIP (SAP) [4]33.660.972.024.148.259.317.138.050.210.928.028.1
NAS-S236.963.074.025.251.762.318.441.653.712.930.941.7
GrowCLIP-S2 (ours)37.563.774.225.951.162.819.142.854.712.930.841.9
Step 3CLIP (TWP)28.354.466.620.642.254.714.434.145.09.625.234.9
CLIP (TFS) [52]37.563.574.226.052.164.019.743.255.113.331.842.6
MS-CLIP (TFS)[21]40.067.778.229.857.068.822.947.959.115.335.246.6
CLIP (SAP) [4]40.768.078.328.155.367.122.845.758.814.233.444.6
NAS-S343.272.880.732.159.970.223.448.361.015.736.848.3
GrowCLIP-S3 (ours)44.673.180.733.260.770.923.548.861.015.836.948.4
Step 4CLIP (TWP)30.256.366.321.746.657.615.636.247.810.426.536.9
CLIP (TFS) [52]41.568.680.030.056.968.021.845.557.314.333.845.1
MS-CLIP (TFS)[21]44.372.881.132.661.372.125.250.562.016.637.749.0
CLIP (SAP) [4]46.974.783.733.761.372.125.250.862.816.637.449.1
NAS-S449.675.283.337.065.275.426.853.064.617.739.651.5
GrowCLIP-S4 (ours)50.875.884.037.265.376.227.653.264.918.140.352.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.464, + 0.822, + 0.479 + ], + "angle": 0, + "content": "Table 4. Results of zero-shot image-text retrieval on Flickr30K and MSCOCO datasets. R@K means top-K recall." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.471, + 0.653 + ], + "angle": 0, + "content": "pernet training and 26 epochs for grown model training. For the baselines, the CLIP adopts the same usual architecture (ViT-B/16), which has 12-layer 768-wide transformer blocks with 12 attention heads for image encoder and 12-layer 512-wide transformer blocks with 8 attention heads for text encoder and are trained for 30 epochs. For NAS, we first train the supernet for 30 epochs to select the best architecture, and then train the selected model from scratch for another 30 epochs. All experiments are performed on 16 NVIDIA Tesla V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.348, + 0.693 + ], + "angle": 0, + "content": "4.2. Zero-shot Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.903 + ], + "angle": 0, + "content": "At each growth step, we evaluate the model on zero-shot image classification task. The downstream datasets include 9 typical settings, Caltech101 [35], CIFAR10 [32], CIFAR100 [32], Describable Textures Dataset (DTD) [12], Flowers102 [47], Food101 [6], OxfordPets [48], SUN39 [5] and ImageNet [14]. We apply prompt ensemble evaluation like CLIP for each dataset. Results are given in Table 3. Our model performs growth process at each growth step except step 1, where the parameter grows from \\(30.0M\\) to \\(188.4M\\). It outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at the last three growth steps. For example, at the third growth step, our model improves the average top-1 accuracy from \\(35.9\\%\\) to \\(38.1\\%\\) \\((+2.2\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.779, + 0.519 + ], + "angle": 0, + "content": "4.3. Zero-shot Image-Text Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.695 + ], + "angle": 0, + "content": "The zero-shot image-text retrieval consists of two subtasks: image-to-text retrieval and text-to-image retrieval. We evaluate the model of each growth step on two retrieval benchmark datasets: Flickr30K [50] and MSCOCO [40]. As shown in Table 4, our model gets better as model grows and data grows. At the last two growth steps, it can be observed that our method has great performance on all metrics. Particularly, our method improves from \\(49.6\\%\\) to \\(50.8\\% (+1.2\\%)\\) for top-1 image-to-text recall on Flickr30K dataset and from \\(26.8\\%\\) to \\(27.6\\% (+0.8\\%)\\) on MSCOCO at the last growth step." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.655, + 0.726 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Here, we conduct the ablation study on our GrowCLIP. All experiments are built at the growth step 2 and we evaluate the models on zero-shot image classification of ImageNet. To verify the effectiveness of each module in our proposed GrowCLIP, we conduct ablation experiments at some core steps mentioned in Section 3.4 as follows. (i) Parameter inheriting with momentum in phase 1 (PIM@1). (ii) Supernet fine-tuning in phase 1 (ST). (iii) Growth architecture selection in phase 2 (GAS). (iv) Parameter inheriting with momentum in phase 3 (PIM@3). The accuracy on zero-shot image classification of ImageNet is re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22184" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.105, + 0.089, + 0.444, + 0.196 + ], + "angle": 0, + "content": "
PIM@1STGASPIM@3Para.Acc.
✓✓✓✓116.6M25.7
✓✓✓X116.6M24.5
✓✓XX109.5M22.6
✓XXX109.5M21.1
XXXX109.5M20.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.206, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Table 5. Ablation analysis of each module in GrowCLIP at Growth Step 2. (Para.: parameter of the model. Acc.: top-1 accuracy (\\%) of zero-shot image classification on ImageNet.)" + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.26, + 0.462, + 0.352 + ], + "angle": 0, + "content": "
Para.Acc.
CLIP-ViT-B/16 (TFS)149.6M28.3
CLIP-ViT-L/16 (TFS)431.1M29.9
GrowCLIP-S4 (TFS)188.4M33.3
GrowCLIP-S4 (ours)188.4M36.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.362, + 0.405, + 0.375 + ], + "angle": 0, + "content": "Table 6. Ablation analysis of GrowCLIP-S4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.552 + ], + "angle": 0, + "content": "ported for comparison and the results of the ablation study are shown in Table 5. As can be seen, all above four components have positive effects on the final results. Specifically, the supernet training will influence the grown model without PIM@3, which results in \\(1.2\\%\\) drop. Compared with GAS method, we randomly choose an architecture in the growth space, which has a worse performance from \\(24.5\\%\\) to \\(22.6\\%\\). With ST, we observe a remarkable improvement of \\(1.5\\%\\). GrowCLIP with PIM@1 solves the problem of local minimum dilemma and brings a \\(0.3\\%\\) improvement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.18, + 0.577 + ], + "angle": 0, + "content": "4.5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.584, + 0.454, + 0.599 + ], + "angle": 0, + "content": "Why can GrowCLIP improve the performance?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Compared with baseline models, the gains of our GrowCLIP come from two folds: (i) GrowCLIP can search the architecture that matches the current data with Growth architecture selection (GAS). (ii) Parameter inheriting with momentum (PIM) benefits GrowCLIP from yesterday's model. To demonstrate their effectiveness, we train the architecture selected by GrowCLIP at step 4 from scratch, and the result is shown in Table 6. Compared with ViT-B/16, the growth architecture selection can bring \\(5\\%\\) improvement on zero-shot image classification of ImageNet. Furthermore, it is even better than ViT-L/16, which is two and a half times as big as our model. More experiments have proved the effectiveness of GAS in the appendix. Comparing our GrowCLIP-S4 with GrowCLIP-S4 (TFS), we observe a remarkable improvement of \\(2.8\\%\\) with PIM." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.826, + 0.468, + 0.841 + ], + "angle": 0, + "content": "Why can GrowCLIP improve the training efficiency?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our GrowCLIP can improve training efficiency, due to GAS and PIM, and the result is shown in Table 7. (i) Compared with the fixed architecture method, GrowCLIP has a smaller model with the GAS, so we reduce training costs" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.888, + 0.224 + ], + "angle": 0, + "content": "
Para.Acc.EpochCost
CLIP-S2(TFS)149.6M20.930336
GrowCLIP-S2 (ours)116.6M25.730288
GrowCLIP-S4 (TFS)188.4M33.3301230
GrowCLIP-S4 (ours)188.4M33.39722
NAS-S4325.9M34.2605280
GrowCLIP-S4 (ours)188.4M36.1301583
" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.234, + 0.839, + 0.248 + ], + "angle": 0, + "content": "Table 7. Efficiency analysis. (Cost: GPU hours)" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.261, + 0.886, + 0.323 + ], + "angle": 0, + "content": "
Para.Acc.
GrowCLIP-S1431.1M14.6
GrowCLIP-S2431.2M26.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.333, + 0.892, + 0.361 + ], + "angle": 0, + "content": "Table 8. The architecture of GrowCLIP-S1 with bigger initialized model (ViT-L/16)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.567 + ], + "angle": 0, + "content": "at the early steps (48 GPU hours saved at step 2). At the later steps, GrowCLIP uses fewer epochs to achieve good performance because of PIM. As shown in Figure 1, GrowCLIP has far outperformed CLIP (TFS) only after supernet training. To demonstrate the effectiveness of PIM, we compare GrowCLIP-S4 with GrowCLIP-S4 (TFS). GrowCLIP only uses 9 epochs to achieve comparable performance with PIM. (ii) Compared with the NAS method, GrowCLIP inherits the parameters of the old model with momentum, which only takes 2 epochs to fine-tune the supernet and 2 epochs to train the supernet. In contrast, one-shot NAS methods consume about 3.3X resources." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Can GrowCLIP find the relationship between model and data? When the model scale is sufficient large, there is no need to enlarge the model. We conduct the experiment started from ViT-L/16. And the result in Table 8 shows that the larger model startpoint hardly grows at step 2, which means that our GrowCLIP can find the relationship between the model and data. In our online learning scenario, the initial architecture of the model is not important due to our growth pipeline. If the model scale is smaller at the beginning, GrowCLIP will select a larger enough model for growth. In contrast, if the model scale is too large initially, GrowCLIP will stop growing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "What is the influence when growth happens in different dimensions? As shown in Table 9, different growth dimensions have different influences on the performance. By analyzing the experimental results, we conclude 4 laws, which can provide a reference for the design of cross-modal model architecture: (i) Enlarging the smaller model bring a greater improvement than the bigger one. For example, when the growth appears in transformer blocks in the image encoder, the accuracy can gain \\(0.93\\%\\) improvement at step 2 but only \\(0.12\\%\\) at step 4. (ii) We need to leave the image" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22185" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.472, + 0.235 + ], + "angle": 0, + "content": "
Step 2Step 3Step 4
DimensionPara.Acc.Para.Acc.Para.Acc.
+4 bI+7.1M+0.93+19.7M+0.46+19.7M+0.12
+4 hI+20.1M+0.34+67.3M+0.12+105.1M+0.03
+2 lI+0.1M+0.84+0.1M+0.07+0.1M+0.10
+4 bT+3.2M+0.01+12.6M+0.08+12.6M+0.01
+4 hT+27.0M+0.22+45.0M+0.01+69.5M-0.04
+4 bS+7.1M+1.03+19.7M+0.19+19.7M-0.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Table 9. The ablation of growth space corresponding to different dimensions. \\((+4b^{I}\\) : the number of transformer blocks in image encoder is increased by 4)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.47, + 0.463 + ], + "angle": 0, + "content": "encoder with more capacity. Compared with text encoder, the improvement of growing the image encoder is more significant at each step. (iii) The convolutional layers is a very important component in the image encoder, which can help the model to better extract visual feature. (iv) The shared encoder can enhance the degree of cross-modal fusion and bring a remarkable improvement. However, do not use a large number of transformers in the shared encoder, or the model become redundant with no noticeable performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.479, + 0.195, + 0.494 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.744 + ], + "angle": 0, + "content": "In this paper, we study the online learning case in cross-modal pre-trained models. The small-scale architecture may limit the final performance due to the limited parameters when data grows. Based on this real-world scenario, we propose a data-driven automatic model growing algorithm for language-image pre-training - GrowCLIP, which upgrades the model size when abundant new data arrives. We adopt parameter inheriting with momentum (PIM) to utilize the knowledge in yesterday's model and seek out the optimal growth architecture at each growth step with growth architecture selection (GAS). GrowCLIP improves the performance on downstream tasks, including zero-shot image classification and retrieval. The effectiveness of each component of GrowCLIP is proved on ablation studies. And we provide a reference for the design of cross-modal model architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Limitations In this work, we only prove the effectiveness of GrowCLIP with public pre-training dataset - CC12M. In future works, we will extend our method to a real-world scenario, where the VLP model keeps training with constantly updated data crawled from the web." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.838, + 0.229, + 0.854 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by National Key R&D Program of China under Grant No.2020AAA0109700, Guangdong Outstanding Youth Fund (Grant No." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "2021B1515020061), Shenzhen Science and Technology Program (Grant No.RCYX20200714114642083), Shenzhen Fundamental Research Program(Grant No. JCYJ20190807154211365), Nansha Key RD Program under Grant No.2022ZD014 and Sun Yat-sen University under Grant No.22lgqb38 and 76160-12220011. We thank MindSpore for the partial support of this work, which is a new deep learning computing framework\\(^{2}\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.671, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://www.mindspore.cn/" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22186" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] R. Aljundi, F. Babiloni, M. Elhoseiny, M. Rohrbach, and T. Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European conference on computer vision, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.469, + 0.257 + ], + "angle": 0, + "content": "[3] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Zitnick, and D. Parikh. Vqa: Visual question answering. In International Conference on Computer Vision, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.469, + 0.285 + ], + "angle": 0, + "content": "[4] J. Ash and R. Adams. On warm-starting neural network training. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[5] A. Barriuso and A. Torralba. Notes on image annotation. Preprint arXiv:1210.3448, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[6] L. Bossard, M. Guillaumin, and L. Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.469, + 0.414 + ], + "angle": 0, + "content": "[7] P. Buzzega, M. Boschini, A. Porrello, D. Abati, and S. Calderara. Dark experience for general continual learning: a strong, simple baseline. In Advances in Neural Information Processing Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.469, + 0.47 + ], + "angle": 0, + "content": "[8] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[9] S. Changpinyo, P. Sharma, N. Ding, and R. Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[10] M. Chen, H. Peng, J. Fu, and H. Ling. Autoformer: Searching transformers for visual recognition. In International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[11] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[12] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Computer Vision and Pattern Recognition, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[13] M. Cornia, L. Baraldi, and R. Cucchiara. Show, control and tell: A framework for generating controllable and grounded captions. In Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.454, + 0.716 + ], + "angle": 0, + "content": "[14] J. Deng. A large-scale hierarchical image database. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[15] J. Devlin, M. Chang, K. Lee, and K. Toutanova. BERT: Pretraining of Deep Bidirectional Transformers for Language Understanding. In North American Chapter of the Association for Computational Linguistics, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[16] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[17] H. Grabner and H. Bischof. On-line boosting and vision. In Computer Vision and Pattern Recognition, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[18] J. Grill, F. Strub, F. Altché, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Ghesh" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[19] Z. Guo, B. Avila Pires, B. Piot, J. Grill, F. Altché, R. Munos, and M. Azar. Bootstrap latent-predictive representations for multitask reinforcement learning. In International Conference on Machine Learning, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[20] Z. Guo, X. Zhang, H. Mu, W. Heng, Z. Liu, Y. Wei, and J. Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.299 + ], + "angle": 0, + "content": "[21] L Zhou B Xiao C. Noel Y Cheng R Xu S Chang H, You and Yuan L. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision, pages 69-87. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[22] R. Hadsell, D. Rao, A. Rusu, and R. Pascanu. Embracing change: Continual learning in deep neural networks. Trends in Cognitive Sciences, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[23] E. Hazan et al. Introduction to online convex optimization. Foundations and Trends® in Optimization, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[24] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.415, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[25] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In Computer Vision and Pattern Recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[26] X. Hu, Z. Gan, J. Wang, Z. Yang, Z. Liu, Y. Lu, and L. Wang. Scaling up vision-language pre-training for image captioning. Preprint arXiv:2111.12233, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[27] X. Hu, X. Yin, K. Lin, L. Wang, L. Zhang, J. Gao, and Z. Liu. Vivo: Surpassing human performance in novel object captioning with visual vocabulary pre-training. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[28] C. Jia, Y. Yang, Y. Xia, Y. Chen, Z. Parekh, H. Pham, Q. Le, Y. Sung, Z. Li, and T. Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[29] J. Kaplan, S. McCandlish, T. Henighan, T. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei. Scaling laws for neural language models. Preprint arXiv:2001.08361, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[30] W. Kim, B. Son, and I. Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.694, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[31] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. National Academy of Sciences, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[32] A. Krizhevsky, G. Hinton, et al. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.791, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[33] Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, and R. Soricut. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] C Li, B. Zhuang, G. Wang, X. Liang, X. Chang, and Y. Yang. Automated progressive learning for efficient training of vision transformers. In Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22187" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[35] F. Li, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[36] J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.47, + 0.235 + ], + "angle": 0, + "content": "[37] L. Li, M. Yatskar, D. Yin, C. Hsieh, and K. Chang. Visualbert: A simple and performant baseline for vision and language. Preprint arXiv:1908.03557, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.47, + 0.293 + ], + "angle": 0, + "content": "[38] X. Li, X. Yin, C. Li, P. Zhang, X. Hu, L. Zhang, L. Wang, H. Hu, L. Dong, F. Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.47, + 0.35 + ], + "angle": 0, + "content": "[39] Y. Li, F. Liang, L. Zhao, Y. Cui, W. Ouyang, J. Shao, F. Yu, and J. Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. Preprint arXiv:2110.05208, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.47, + 0.407 + ], + "angle": 0, + "content": "[40] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[41] C. Liu, B. Zoph, M. Neumann, J. Shlens, W. Hua, L. Li, F. Li, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. In European conference on computer vision, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[42] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. Preprint arXiv:1806.09055, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[43] W. Liu, S. Chen, L. Guo, X. Zhu, and J. Liu. Cptr: Full transformer network for image captioning. Preprint arXiv:2101.10804, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[44] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Preprint arXiv:1907.11692, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[45] I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[46] J. Lu, V. Goswami, M. Rohrbach, D. Parikh, and S. Lee. 12-in-1: Multi-task vision and language representation learning. In Computer Vision and Pattern Recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[47] M. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Computer Vision, Graphics & Image Processing, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[48] O. Parkhi, A. Vedaldi, A. Zisserman, and C. Jawahar. Cats and dogs. In Computer Vision and Pattern Recognition, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[49] H. Peng, H. Du, H. Yu, Q. Li, J. Liao, and J. Fu. Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. In Advances in Neural Information Processing Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[50] B. Plummer, L. Wang, C. Cervantes, J. Caicedo, J. Hockenmaier, and S. Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-tosentence models. In International Conference on Computer Vision, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[51] R. Pogodin and T. Lattimore. On first-order bounds, variance and gap-dependent bounds for adversarial bandits. In Uncertainty in Artificial Intelligence, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[52] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[53] E. Real, A. Aggarwal, Y. Huang, and Q. Le. Regularized Evolution for Image Classifier Architecture Search. In AAAI Conference on Artificial Intelligence, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[54] E. Real, S. Moore, A. Selle, S. Saxena, Y. Suematsu, J. Tan, Q. Le, and A. Kurakin. Large-scale evolution of image classifiers. In International Conference on Machine Learning, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[55] H Xu J. Han T. Tuytelaars Z. Li S. Yan, L. Hong and X. He. Generative negative text replay for continual vision-language pretraining. In European Conference on Computer Vision, pages 22-38. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[56] A. Saffari, C. Leistner, J. Santner, M. Godec, and H. Bischof. On-line random forests. In International Conference on Computer Vision workshops, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[57] S. Shalev-Shwartz et al. Online learning and online convex optimization. Foundations and Trends® in Machine Learning, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[58] S. Song and B. Choi. A study on continuous follow-the-leader (ftl) gaits: an effective walking algorithm over rough terrain. Mathematical biosciences, 1989." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[59] T Srinivasan, T Chang, L. Alva, G Chochlakis, M Rostami, and J Thomason. Climb: A continual learning benchmark for vision-and-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[60] M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard, and Q. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[61] B. Thomee, D. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland, D. Borth, and L. Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[62] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, and I. Polosukhin. Attention Is All You Need. In Neural Information Processing Systems, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[63] Haonan Wang, Minbin Huang, Runhui Huang, Lanqing Hong, Hang Xu, Tianyang Hu, Xiaodan Liang, and Zhenguo Li. Boosting visual-language models by exploiting hard samples. arXiv preprint arXiv:2305.05208, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[64] S. Xie, H. Zheng, C. Liu, and L. Lin. SNAS: Stochastic Neural Architecture Search. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[65] L. Yao, R. Huang, L. Hou, G. Lu, M. Niu, H. Xu, X. Liang, Z. Li, X. Jiang, and C. Xu. Filip: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[66] Y. You, J. Li, S. Reddi, J. Hseu, S. Kumar, S. Bhojanapalli, X. Song, J. Demmel, K. Keutzer, and C. Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22188" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[67] F. Yu, J. Tang, W. Yin, Y. Sun, H. Tian, H. Wu, and H. Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3208-3216, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[68] J. Yu, Z. Wang, V. Vasudevan, L. Yeung, M. Seyedhosseini, and Y. Wu. Coca: Contrastive captioners are image-text foundation models. Preprint arXiv:2205.01917, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[69] X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.468, + 0.333 + ], + "angle": 0, + "content": "[70] B. Zoph, V. Vasudevan, J. Shlens, and Q. Le. Learning Transferable Architectures for Scalable Image Recognition. In Computer Vision and Pattern Recognition, 2018." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22189" + } + ] +] \ No newline at end of file diff --git a/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_origin.pdf b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8a272102933afd2301ef6bd2d5d9af53c0e4448 --- /dev/null +++ b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/84827675-3aac-4157-b576-07a125b493f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3937b2275ed067f68e0de2e2637fa6fb66ee42cae9817b1cd1490fb59b7c6f68 +size 631753 diff --git a/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/full.md b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd4a90088f421d5247eb4db3d48b9728878ea767 --- /dev/null +++ b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/full.md @@ -0,0 +1,328 @@ +# GrowCLIP: Data-aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-training + +Xinchi Deng $^{1}$ Han Shi $^{2}$ Runhui Huang $^{1}$ Changlin Li $^{3}$ Hang Xu $^{2}$ +Jianhua Han $^{2}$ James Kwok $^{4}$ Shen Zhao $^{1*}$ Wei Zhang $^{2}$ Xiaodan Liang $^{1}$ $^{1}$ Sun Yat-sen University $^{2}$ Huawei Noah's Ark Lab + $^{3}$ University of Technology Sydney $^{4}$ The Hong Kong University of Science and Technology +dengxch5@mail2.sysu.edu.cn + +# Abstract + +Cross-modal pre-training has shown impressive performance on a wide range of downstream tasks, benefiting from massive image-text pairs collected from the Internet. In practice, online data are growing constantly, highlighting the importance of the ability of pre-trained model to learn from data that is continuously growing. Existing works on cross-modal pre-training mainly focus on training a network with fixed architecture. However, it is impractical to limit the model capacity when considering the continuously growing nature of pre-training data in real-world applications. On the other hand, it is important to utilize the knowledge in the current model to obtain efficient training and better performance. To address the above issues, in this paper, we propose GrowCLIP, a data-driven automatic model growing algorithm for contrastive language-image pre-training with continuous image-text pairs as input. Specifically, we adopt a dynamic growth space and seek out the optimal architecture at each growth step to adapt to online learning scenarios. And the shared encoder is proposed in our growth space to enhance the degree of cross-modal fusion. Besides, we explore the effect of growth in different dimensions, which could provide future references for the design of cross-modal model architecture. Finally, we employ parameter inheriting with momentum (PIM) to maintain the previous knowledge and address the issue of the local minimum dilemma. Compared with the existing methods, GrowCLIP improves $2.3\%$ average top-1 accuracy on zero-shot image classification of 9 downstream tasks. As for zero-shot image retrieval, GrowCLIP can improve $1.2\%$ for top-1 image-to-text recall on Flickr30K dataset. + +![](images/75efdd0b644164ce01f11930b0362fa1ddf453c61cbbaaa819538b4866d032d6.jpg) +Figure 1. Top-1 accuracy $(\%)$ of zero-shot image classification on ImageNet of GrowCLIP and baselines during training at step 4, where the horizontal dotted lines mean the process of supernet training. Our GrowCLIP has the best performance and is more efficient compared with other baselines. + +# 1. Introduction + +Recently, large-scale pre-trained models illustrate the potential performance among different fields including computer vision (CV) [24, 25, 11] and natural language processing (NLP) [15, 44, 33]. Cross-modal models like CLIP [52], ALIGN [28], FILIP [65], BLIP [36] also demonstrate remarkable success across various vision-language downstream tasks. Note that these models are usually built in dual-stream architectures, which consist of the image and text encoders to extract the feature of image and text inputs, respectively. The alignment between these two features is then performed under the contrastive objective to learn the alignment between different modals. + +However, training these cross-modal models requires a large amount of image-text pairs collected from the Internet, e.g., 400M image-text pairs for CLIP [52] and 300M for FILIP [65]. Most existing cross-modal methods [52, 28, 65, 39, 36, 68, 63, 1] directly train the final large- + +
CC12M10%20%50%100%
CLIP-0.5-ViT-B/166.611.119.025.7
CLIP-ViT-B/166.511.620.928.3
+ +Table 1. Result of different model sizes with different data sizes: Top-1 accuracy (\%) of zero-shot image classification on ImageNet. The depth and the width of 0.5-ViT-B/16 are both one half of ViT-B/16. The relative performance ranking is dependent on the size of the training data. + +scale model using the completed image-text pair datasets collected at some point in time. However, it is impractical to regard cross-modal pre-training as "disposable" training without considering the continuously growing nature of pre-training data or domain in real-world applications. For instance, mountains of new knowledge is generated on the Internet every day, which can be used to further improve the performance of existing pre-trained cross-modal models. Therefore, in this paper, we consider the training of large-scale cross-modal pre-trained models as an online learning problem [17, 56, 57] with continuous image-text pairs as input. Different from the standard continual learning setting [59, 55, 22, 31, 2, 7], the previous image-text pairs can also be achieved in our setting since we assume the size of training data is gradually increased over time. + +One tough problem in the online learning scenario is that the model capacity should be related to the size of training data [16, 69, 29, 26]. For example, it's observed that large ViT models perform worse than ResNets when pre-trained on small datasets, while the result is opposite when they are both pre-trained on larger datasets [16]. To verify the relationship between model capability and data size, we split Conceptual 12M (CC12M) dataset [9] and test the performance of ViT-B/16 with different scales [52]. As shown in Table 1, the relative performance ranking is dependent on the size of the training data. Given a smaller dataset (e.g., $10\%$ CC12M), the performance of 0.5-ViT-B/16 is comparable with ViT-B/16. In contrast, when data is sufficient (e.g., $100\%$ CC12M), the performance of ViT-B/16 is better than the smaller model. Thus, the unchanged model is not practical for this real-time learning setting and it's still an open issue on how to modify and train our model with the growing data. Besides, how to efficiently make use of knowledge of previous model when new data is coming remains an open problem. One direct solution is to fine-tune the model with the updated training dataset. However, training with previous pre-trained parameters of the same model will deteriorate the performance [4] due to the parameter inheriting issue. As shown in Figure 1, the CLIP training with pre-training (TWP) have worse performance than the one training from scratch (TFS). + +To address the above issues, we propose a data-aware automatic model growing method (denoted as GrowCLIP) for large-scale contrastive language-image pre-training, which + +performs a model growth process considering the gradually increased pre-training data. Specially, when training data grows dynamically, we adopt different cross-modal network architectures via the customized neural architecture search (NAS) to make the network pre-training more efficient. Different from traditional NAS approaches [70, 53, 64], we introduce a cross-modal customized NAS by defining a dynamic search space named growth space, which is enlarged when more data comes, and proposing a shared encoder search space to enhance the degree of cross-modal fusion. To utilize the architecture at the previous step more efficiently, the parameters of the new architecture are also inherited from the old one with momentum to maintain the previous knowledge and address the issue of local minimum dilemma. Finally, growth architecture selection procedure is performed to select the optimal model architecture at each step, considering the performance and model size. + +Experiments are conducted by averagely dividing the Conceptual 12M (CC12M) into 4 growth steps under the online learning setting. As depicted in Table 1, compared with the existing methods, our GrowCLIP has the best performance and is more efficient. Specially, experimental results show that our GrowCLIP can improve up to $2.3\%$ average top-1 accuracy on zero-shot image classification of 9 downstream tasks compared with the existing methods. As for zero-shot image-text retrieval, GrowCLIP has a $1.2\%$ improvement for top-1 image-to-text recall on Flickr30K [50] dataset and $0.8\%$ on MSCOCO [40]. + +To summarize, the contributions of this paper are listed as follows: (i) To adapt to the growing data scenario, we propose a data-aware automatic model growing method, named GrowCLIP. (ii) We provide some insights for the design of cross-modal model architecture. (iii) The extensive experiment results illustrate the effectiveness of our approach on zero-shot classification and retrieval tasks. + +# 2. Related Work + +Vision-Language Pre-training Models Inspired by the success of pre-training in computer vision (CV) and natural language processing (NLP), a boosting number of research works in the domain of vision-language pretraining (VLP) has recently surged to pursue a unified multi-modality representation. Vision-language pretraining trains the model on large-scale image-text pairs to improve performance of downstream vision and language tasks including image classification [14, 35, 32], retrieval [50, 40], Image Captioning (IC) [43, 13, 27, 26], Visual Question and Answer (VQA) [3, 46, 67] and so on. The VLP model can be categorized into two categories: (i) The dual-stream model, e.g., CLIP [52], ALIGN [28], FILIP [65], processes visual and textual tokens separately with two parallel streams to acquire the representation and then fuse them through interactive module. (ii) The single- + +stream model, e.g., Oscar [38], ViLT [30], VisualBERT [37], directly combines the extracted visual and textual tokens and feeds them into the transformer-based model. The public pre-training datasets of VLP model include available datasets like YFCC100M [61] and CC12M [9]. CLIP [52], ALIGN [28], FILIP [65] and LEMON [26] adopt large-scale image-text pairs crawled from the web to show more powerful models. Different from these works that treat this process as offline learning, we investigate the online learning case in vision-language pre-training. + +Online Learning Online learning is a problem setting of machine learning in which data becomes available in a sequential order [17, 56, 57]. Different from continual learning [22, 31, 2, 7], online learning can reach previous data all the time. As opposed to batch learning techniques which generate the best model by learning on the entire training dataset at once, online learning is used to update our predictor for future data at each growth step. Follow the leader [58] is the simplest learning rule which assumes the hypothesis that the leader has the least loss overall past rounds. Follow the regularised leader [51] is a modification variant and obtain better regret bounds by learning a regularisation function. Online convex optimization [23] is a general algorithm framework for decision making which targets convex optimization. To adapt online learning scenario, we propose a data-aware automatic model growing algorithm that grows the model with increasing training data. + +Neural Architecture Search Neural Architecture Search (NAS) aims to automatically design the state-of-the-art neural networks. Early works [70, 54, 60] are based on reinforcement learning and evolutionary algorithms. In the later work, various techniques are used to reduce the computational complexity. Gradient-based methods [42, 41] regard the network architecture as a set of parameters and adopt back-propagation algorithm to optimize them. One-shot NAS methods [49, 10, 20] train a supernet to avoid training each subnet from scratch. Unfortunately, one-shot NAS methods have to spend lots of resources to train the supernet. In contrast, the method we propose inherits the parameters of the old model with momentum, which only takes few epochs to train the supernet. Furthermore, GrowCLIP can modify the neural architecture to adapt to the size of the training data. + +# 3. Methodology + +In this section, we describe the online learning problem setting and its challenges for cross-modal pre-training (Section 3.1). To adapt the growth data, we propose a dynamic growth space, including a proposed shared encoder (Section 3.2). To rise to the challenge of the local minimum dilemma, we propose parameter inheriting with momentum (Section 3.3). The whole pipeline of our method is illus + +trated in Section 3.4. + +# 3.1. Problem Setting and Challenges + +Problem Setting The target of online learning is obtaining the optimal model at each growth step of the data. Let $\mathcal{D}^{t - 1} = \{\pmb{x}_i^I,\pmb{x}_i^T\}_{i = 1}^k$ be the training data at growth step $t - 1$ , where $\pmb{x}^I$ and $\pmb{x}^T$ are image and text samples and $k$ is the training data size. At the next growth step $t$ , the enlarged training dataset $\mathcal{D}^t$ satisfies $\mathcal{D}^{t - 1}\subseteq \mathcal{D}^t$ due to the incoming data. Given an optimal model from growth step $t - 1$ , our aim is to achieve the optimal model (with optimal architecture $\psi^{*}$ and parameters $\omega^{*}$ ) based on the enlarged dataset $\mathcal{D}^t$ : + +$$ +\left\{\psi^ {*}, \omega^ {*} \right\} = \underset {\psi , \omega} {\arg \min } \mathcal {L} (\psi , \omega ; \mathcal {D} ^ {t}). \tag {1} +$$ + +Challenge of Data Growth As discussed in Table 1, the size of the training dataset is related to the optimal architecture selection. Given a smaller dataset, the small-scale architecture has comparable performance to large-scale architecture but fewer parameters are deployed. When data grows, the small-scale architecture may limit the final performance due to the limited parameters. However, few works notice the effect of data size in online learning settings. + +Challenge of Parameter Inheriting As investigated, the performance of model trained from scratch is much better than trained with pre-training. This is probably caused by the influence of the inheriting parameters trained with the previous smaller dataset, which makes it difficult for the model to escape from the local minima in the grown dataset. However, training from scratch is costly due to multiple growth steps in online learning. The balance between training from scratch and training with pre-training is worthy of consideration. + +To solve the above problems, we propose an automatic model growing method - GrowCLIP (shown in Figure 2), which can modify the neural architecture at different growth steps in online learning case. To adapt to the growing data scenario, we propose a growth space for each growth step, where the scale of architecture candidates among this space is larger when data grows. To alleviate the parameter inhering issue with efficiency, we perform parameter inhering with momentum such that the trade off between training from scratch and training with pre-training is balanced. + +# 3.2. Growth Space + +Model Architecture Following the basic neural architecture in CLIP [52], we adopt a dual-stream model with image encoder $f(\cdot)$ and text encoder $g(\cdot)$ . To adapt the cross-modality scenarios, we propose a shared encoder $h(\cdot)$ , in which the image and text share the transformers except layernorm layers. The shared encoder not only can reduce + +![](images/6edcf2db6e9d07d8e12003be64e20d8d8870f81efafe8db811bdde3147900670.jpg) +Figure 2. The overview of GrowCLIP. As shown in Figure (a), when the data grows from growth step $t - 1$ to growth step $t$ , our model grows adaptively as well. Figure (b) illustrates the detail of model growing and we simplify the representation of model architecture. + +![](images/c9d953690f661e72d8c03171fbc2af26326642285bc46b92416cd579b8764279.jpg) + +model parameters, but also enhance the degree of cross-modal fusion. Given an image $\boldsymbol{x}_i^I$ and a text $\boldsymbol{x}_j^T$ , the model outputs the image representations $I_i = h(f(\boldsymbol{x}_i^I))$ and text representations $T_j = h(g(\boldsymbol{x}_j^T))$ . The similarity score computation of between image $\boldsymbol{x}_i^I$ and text $\boldsymbol{x}_j^T$ is: + +$$ +s _ {i, j} = h \left(f \left(\boldsymbol {x} _ {i} ^ {I}\right)\right) ^ {\top} h \left(g \left(\boldsymbol {x} _ {j} ^ {T}\right)\right), \tag {2} +$$ + +where $s_{i,j}$ denotes the similarity of the $i$ -th image to the $j$ -th text. + +In each training batch, we sample $n$ image-text pairs $\{\pmb{x}_i^I,\pmb{x}_i^T\}_{i = 1}^n$ from the dataset. For image $\pmb{x}_i^I$ in the image-text pair $< x_{i}^{I},x_{i}^{T}>$ $x_{i}^{T}$ is its positive, while the other texts will be used as in-batch negatives. The image-to-text contrastive loss $\mathcal{L}_i^I$ for $\pmb{x}_i^I$ can be formulated as: + +$$ +\mathcal {L} _ {i} ^ {I} \left(\boldsymbol {x} _ {i} ^ {I}, \left\{\boldsymbol {x} _ {j} ^ {T} \right\} _ {j = 1} ^ {n}\right) = - \frac {1}{n} \log \frac {\exp \left(s _ {i , i}\right)}{\sum_ {j} \exp \left(s _ {i , j}\right)}, \tag {3} +$$ + +Similarly, the text-to-image contrastive loss for $\pmb{x}_i^T$ is: + +$$ +\mathcal {L} _ {i} ^ {T} \left(\boldsymbol {x} _ {i} ^ {T}, \left\{\boldsymbol {x} _ {j} ^ {I} \right\} _ {j = 1} ^ {n}\right) = - \frac {1}{n} \log \frac {\exp \left(s _ {i , i}\right)}{\sum_ {j} \exp \left(s _ {i , j}\right)}, \tag {4} +$$ + +Totally, the contrastive loss of this mini-batch can be represented by $\mathcal{L} = \frac{1}{2}\sum_{i=1}^{a}\left(\mathcal{L}_i^I + \mathcal{L}_i^T\right)$ . + +Growth Space of image encoder For the image encoder, we adopt a mixed architecture of convolutional layer and transformer block [62]. There are three variable factors in our growth space design: (i) the number of convolutional layers $l^I$ ; (ii) the number of transformer blocks $b^I$ ; (iii) the number of transformer heads $h^I$ . + +Growth Space of text encoder The text encoder part is a transformer-based architecture, whose growth space includes two variable factors: (i) the number of transformer blocks $b^{T}$ and (ii) the number of transformer heads $h^{T}$ . + +Growth Space of shared encoder The shared encoder is a transformer-based architecture, whose growth space includes the number of transformer blocks $b^{S}$ . + +
Image encoderText encoderShared encoder
Transformer blocks{bIt-1, bIt-1+4}{bTt-1, bTt-1+4}{bSt-1, bSt-1+4}
Transformer heads{hIt-1, hIt-1+4}{hTt-1, hTt-1+4}-
Convolutional layers{lIt-1, lIt-1+2}--
+ +Table 2. The growth space $\Phi$ of GrowCLIP at the growth step t. ${b}_{t - 1}^{I},{b}_{t - 1}^{T},{b}_{t - 1}^{S},{h}_{t - 1}^{I},{h}_{t - 1}^{T},{l}_{t - 1}^{I}$ are the hyperparameters of model architecture at growth step $t - 1$ . + +The detailed descriptions of growth space for three encoders are shown in Table 2. At each new growth step, the variable factors are enlarged to adapt the growth data. Specifically, the numbers of transformer blocks and transformer heads are both increased by 4, and the number of convolutional layers is increased by 2. + +# 3.3. Parameter inheriting with momentum + +Some previous papers utilize the training results of the old model and transfer the knowledge to the new model [18, 19, 25, 4, 34]. Inspired by the above methods, we propose parameter inheriting with momentum (PIM) to avoid local minimum dilemma. The illustration of our parameter inheriting with momentum is shown in Figure 3. To preserve the knowledge from the old model, we extend the size + +![](images/8bab3a1ee52d8ce8e2c6e01e9e660787ba328cfc961bb08276127fe556278dde.jpg) +Figure 3. Take image encoder for example of our proposed parameter inheriting with momentum. As for the initialization of the enlarged model, we preserve the existing parameters of the old model for exploitation and combine it with random initialization for exploration. + +of the old model to the enlarged model and maintain the parameters in existing parts. As for the extended layers, we copy the parameters from the last layers directly. $\omega_{old}$ denotes the parameters of extension of the old model. Besides the above parameter inheriting term, we also add a re-initialization term with randomly initialized parameters $\omega_{rand}$ to avoid the parameter inheriting issue. The final initialized supernet is a combination of the above two terms as follows: + +$$ +\omega_ {\text {i n h e r i t}} = \beta \cdot \omega_ {\text {o l d}} + \gamma \cdot \omega_ {\text {r a n d}}, \tag {5} +$$ + +where $\beta$ and $\gamma$ are two hyperparameters related to exploitation and exploration. Additionally, the parameters of new transformer heads are randomly initialized, which makes that different heads in transformer are orthogonal and the new heads learn new features from different perspectives. + +# 3.4. Growth Pipeline + +The growth pipeline at growth step $t$ (except the first one) includes three sequential phases as shown in Figure 2 (b). + +Phase 1: Supernet Fine-tuning As for a new growth step, a supernet $\Psi$ covering the whole growth space is constructed by enlarging the current model. First, we use the PIM method described in Section 3.3 to initialize the supernet. However, there are some newly added layers in the supernet compared with the old model from the last growth step, which brings a performance gap between the subnets with new layers and others. To make the supernet more robust, we slightly fine-tune the parameters of the full supernet for a few epochs to mitigate the performance gap among subnetworks after PIM initialization. + +Phase 2: Growth Architecture Selection Growth Architecture Selection (GAS) is the main phase in our algorithm where we select the optimal architecture in the growth space. The whole GAS process can be divided into two steps: training and searching. + +First, to find the optimal architecture $\psi^{*}$ , we need to fully optimize all the subnet in the supernet, which covers all the candidate growth architecture: + +$$ +\omega_ {s u p e r} ^ {*} = \underset {\omega_ {s u p e r}} {\arg \min } \mathbb {E} _ {\psi \in \Psi} \left\{\mathcal {L} \left(\psi , \omega_ {s u p e r}; \mathcal {D} ^ {t}\right) \right\}. \tag {6} +$$ + +Inspired by one-shot neural architecture search [8], we optimize $\omega_{super}$ in Equation 6 by uniformly sampling subnets $\psi$ from the growth space $\Phi$ and update the parameters of the sampled subnets in each iteration. More specifically, at each step, we first randomly sample a subnet $\psi$ from the growth space $\Psi$ , then optimize the parameter of this subnet by minimizing the contrastive loss between image and text. This can be considered as solving Equation 6 via Monte Carlo sampling. By iterative training, these subnets can be fully optimized in a few epochs. + +Second, after training the supernet, we search for the optimal subnet in the growth space $\Phi$ . Here, we consider both the performance and the number of parameters of the architecture, and select the optimal architecture $\psi^{*}$ using the composite metric: + +$$ +\psi^ {*} = \underset {\psi \in \Psi} {\arg \max } \left\{\mathcal {A} (\psi) + \alpha \frac {\mathcal {N} ^ {t - 1}}{\mathcal {N} ^ {t}} \frac {\mathcal {P} (\Psi)}{\mathcal {P} (\psi)} \right\}, \tag {7} +$$ + +where $\mathcal{A}(\cdot)$ is the top-1 accuracy of test dataset, $\mathcal{N}^t$ denotes the number of data on growth step $t$ , $\mathcal{P}(\cdot)$ represents the total amount of model parameters and $\alpha$ is the trade-off hyperparameter. When we get a large number of new coming data, $\frac{\mathcal{N}^{t - 1}}{\mathcal{N}^t}$ will get small, which means that we focus more on the performance and encourage the growth of models. As our growth space is relatively small, traversal evaluation of all the candidate architectures is affordable. Therefore, we simply test all candidate subnets in the growth space $\Phi$ using the parameters inherited from the supernet $\Psi$ and select the optimal one for the next stage via the composite metric of Equation 7. + +Phase 3: Selected Model Training To alleviate the influence of supernet training on the grown model, we also use PIM to initialize the parameter $\omega$ of selected model $\psi^{*}$ : + +$$ +\boldsymbol {\omega} = \beta \cdot \boldsymbol {\omega} _ {s u p e r} ^ {*} + \gamma \cdot \boldsymbol {\omega} _ {r a n d}. \tag {8} +$$ + +After the parameter inheriting, we keep training the selected model, to obtain its final optimal parameters: + +$$ +\omega^ {*} = \underset {\omega} {\arg \min } \mathcal {L} \left(\psi^ {*}, \omega ; \mathcal {D} ^ {t}\right). \tag {9} +$$ + +# 4. Experiments + +In this section, we firstly describe the experiment setting and implementation details (Section 4.1). Then we show the results on zero-shot image classification (Section 4.2) and image-text retrieval (Section 4.3). Finally, we conduct ablation study on our GrowCLIP to validate the effectiveness of components (Section 4.4) and analyse (Section 4.5). + +# 4.1. Experiment Setting + +Benchmark Protocol Conceptual 12M (CC12M) [9] collects about 12 million image-text pairs from the Internet. + +
Growth stepMethodTraining dataCaltech101CIFAR10CIFAR100DTDFlowers102Food101OxfordPedsSUN397ImageNetAverage
ParameterParameterParameterParameterParameterParameterParameterParameterParameterParameter
Step 1CLIP (TWP)3M149.6M42.837.614.5109.316.121.622.014.020.9
CLIP (TFS)[52]3M149.6M42.837.614.5109.316.121.622.014.020.9
MS-CLIP (TFS)[21]3M129M47.135.514.410.012.917.330.124.718.123.3
CLIP (SAP)[4]3M149.6M42.837.614.5109.316.121.622.014.020.9
NAS-S13M30.0M40.531.213.39.08.915.820.223.213.719.5
GrowCLIP-S1 (ours)3M30.0M40.531.213.39.08.915.820.223.213.719.5
Step 2CLIP (TWP)6M149.6M50.336.116.011.312.421.227.125.817.924.2
CLIP (TFS)[52]6M149.6M58.552.022.312.113.125.033.630.120.929.7
MS-CLIP (TFS)[21]6M129M60.650.018.313.417.42742.432.625.231.8
CLIP (SAP)[4]6M149.6M61.949.021.412.415.524.635.431.622.630.5
NAS-S26M96.9M59.441.818.313.315.826.737.333.824.430.1
GrowCLIP-S2 (ours)6M116.6M57.748.021.213.718.029.539.534.525.732.0
Step 3CLIP (TWP)9M149.6M55.240.119.713.115.723.830.928.920.427.5
CLIP (TFS)[52]9M149.6M63.354.624.514.517.529.536.933.825.333.3
MS-CLIP (TFS)[21]9M129M65.048.222.513.920.631.042.635.729.233.0
CLIP (SAP)[4]9M149.6M67.853.926.715.121.232.941.436.028.235.9
NAS-S39M287.4M66.746.822.214.919.735.446.139.330.135.8
GrowCLIP-S3 (ours)9M168.6M69.653.925.818.418.736.346.241.632.638.1
Step 4CLIP (TWP)12M149.6M57.544.420.512.114.224.134.730.322.228.9
CLIP (TFS)[52]12M149.6M65.957.329.716.318.532.945.536.728.336.8
MS-CLIP (TFS)[21]12M129M65.350.926.316.221.935.950.138.432.037.4
CLIP (SAP)[4]12M149.6M69.860.631.912.421.036.747.040.432.239.1
NAS-S412M325.9M70.853.928.316.721.739.551.241.334.239.7
GrowCLIP-S4 (ours)12M188.4M71.960.728.317.323.342.552.445.536.142.0
+ +Table 3. Top-1 accuracy(%) of zero-shot image classification on 9 datasets. Our GrowCLIP outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at each growth step except step 1. + +The CC12M is randomly divided into 4 subsets. We select one subset as the training set for the first growth step and adds one subset to the training set at each subsequent growth step such that there are total 4 growth steps in our experiments. At each growth step, we evaluate the corresponding model on the downstream tasks including zero-shot image classification and zero-shot image-text retrieval. The grown architecture of each growth step is shown in Appendix. + +Comparison Methods To demonstrate the effectiveness of our method, we compare it with the following baselines: (i) Training with pre-training (TWP) is a usual method in transfer learning. (ii) Training from scratch (TFS) [52] is used as a basic method to train language-image pre-trained model. (iii) MS-CLIP [21] is a handcrafted framework with a shareable module and modality-specific auxiliary module (the early specialization). (iv) The shrink and perturb (SAP) [4] is a method of parameter inheriting in warm-starting learning. (v) Neural architecture search (NAS) [49, 10] is an important research field of auto machine learning, in which we use the same search space as the growth space of GrowCLIP. + +Implementation Details A dual-stream model follow + +ing CLIP [52] without shared encoder is used as basic architecture at the first growth step. The image encoder is stacked with 6 transformer blocks with the following hyperparameters: number of self-attention heads is 6 and the hidden layer size is 384. As for the text encoder, 6 transformer blocks are stacked and the number of the attention heads and the width are 4 and 256 respectively. The resolution of input in the images encoder is resized to 224 and the maximum sequence in the text encoder is limited to 77. The center crop is the only data augmentation method for training. We use the prompts like $\mathrm{CLIP}^1$ on the zero-shot image classification. We adopt LAMB optimizer [66] with learning rate 0.01 and weight decay 0.0001, and cosine learning rate schedule [45] with a 4000-iteration linear warmup. For the hyperparameters, the proportion of the old model and the new one in the parameter inheriting $\beta = 0.3$ , $\gamma = 0.001$ . The model parameter coefficient in growth architecture selection $\alpha = 0.5$ . We train the model with a total batch size of 1536 for 30 epochs at each growth step, which includes 2 epochs for supernet fine-tuning, 2 epochs for su + +
Growth stepMethodFlickr30KMSCOCO
image-to-texttext-to-imageimage-to-texttext-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
Step 1CLIP (TWP)19.645.356.414.333.544.110.726.336.87.019.127.5
CLIP (TFS) [52]19.645.356.414.333.544.110.726.336.87.019.127.5
MS-CLIP (TFS)[21]27.451.863.318.139.250.712.931.943.28.823.533.1
CLIP (SAP) [4]19.645.356.414.333.544.110.726.336.87.019.127.5
NAS-S122.845.257.813.534.246.010.627.437.57.019.928.8
GrowCLIP-S1 (ours)22.845.257.813.534.246.010.627.437.57.019.928.8
Step 2CLIP (TWP)26.651.062.017.439.850.712.830.540.88.322.531.9
CLIP (TFS) [52]31.457.768.121.246.757.117.137.448.710.927.337.4
MS-CLIP (TFS)[21]35.865.175.025.853.164.619.343.055.213.131.542.6
CLIP (SAP) [4]33.660.972.024.148.259.317.138.050.210.928.028.1
NAS-S236.963.074.025.251.762.318.441.653.712.930.941.7
GrowCLIP-S2 (ours)37.563.774.225.951.162.819.142.854.712.930.841.9
Step 3CLIP (TWP)28.354.466.620.642.254.714.434.145.09.625.234.9
CLIP (TFS) [52]37.563.574.226.052.164.019.743.255.113.331.842.6
MS-CLIP (TFS)[21]40.067.778.229.857.068.822.947.959.115.335.246.6
CLIP (SAP) [4]40.768.078.328.155.367.122.845.758.814.233.444.6
NAS-S343.272.880.732.159.970.223.448.361.015.736.848.3
GrowCLIP-S3 (ours)44.673.180.733.260.770.923.548.861.015.836.948.4
Step 4CLIP (TWP)30.256.366.321.746.657.615.636.247.810.426.536.9
CLIP (TFS) [52]41.568.680.030.056.968.021.845.557.314.333.845.1
MS-CLIP (TFS)[21]44.372.881.132.661.372.125.250.562.016.637.749.0
CLIP (SAP) [4]46.974.783.733.761.372.125.250.862.816.637.449.1
NAS-S449.675.283.337.065.275.426.853.064.617.739.651.5
GrowCLIP-S4 (ours)50.875.884.037.265.376.227.653.264.918.140.352.0
+ +Table 4. Results of zero-shot image-text retrieval on Flickr30K and MSCOCO datasets. R@K means top-K recall. + +pernet training and 26 epochs for grown model training. For the baselines, the CLIP adopts the same usual architecture (ViT-B/16), which has 12-layer 768-wide transformer blocks with 12 attention heads for image encoder and 12-layer 512-wide transformer blocks with 8 attention heads for text encoder and are trained for 30 epochs. For NAS, we first train the supernet for 30 epochs to select the best architecture, and then train the selected model from scratch for another 30 epochs. All experiments are performed on 16 NVIDIA Tesla V100 GPUs. + +# 4.2. Zero-shot Image Classification + +At each growth step, we evaluate the model on zero-shot image classification task. The downstream datasets include 9 typical settings, Caltech101 [35], CIFAR10 [32], CIFAR100 [32], Describable Textures Dataset (DTD) [12], Flowers102 [47], Food101 [6], OxfordPets [48], SUN39 [5] and ImageNet [14]. We apply prompt ensemble evaluation like CLIP for each dataset. Results are given in Table 3. Our model performs growth process at each growth step except step 1, where the parameter grows from $30.0M$ to $188.4M$ . It outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at the last three growth steps. For example, at the third growth step, our model improves the average top-1 accuracy from $35.9\%$ to $38.1\%$ $(+2.2\%)$ . + +# 4.3. Zero-shot Image-Text Retrieval + +The zero-shot image-text retrieval consists of two subtasks: image-to-text retrieval and text-to-image retrieval. We evaluate the model of each growth step on two retrieval benchmark datasets: Flickr30K [50] and MSCOCO [40]. As shown in Table 4, our model gets better as model grows and data grows. At the last two growth steps, it can be observed that our method has great performance on all metrics. Particularly, our method improves from $49.6\%$ to $50.8\% (+1.2\%)$ for top-1 image-to-text recall on Flickr30K dataset and from $26.8\%$ to $27.6\% (+0.8\%)$ on MSCOCO at the last growth step. + +# 4.4. Ablation Study + +Here, we conduct the ablation study on our GrowCLIP. All experiments are built at the growth step 2 and we evaluate the models on zero-shot image classification of ImageNet. To verify the effectiveness of each module in our proposed GrowCLIP, we conduct ablation experiments at some core steps mentioned in Section 3.4 as follows. (i) Parameter inheriting with momentum in phase 1 (PIM@1). (ii) Supernet fine-tuning in phase 1 (ST). (iii) Growth architecture selection in phase 2 (GAS). (iv) Parameter inheriting with momentum in phase 3 (PIM@3). The accuracy on zero-shot image classification of ImageNet is re + +
PIM@1STGASPIM@3Para.Acc.
✓✓✓✓116.6M25.7
✓✓✓X116.6M24.5
✓✓XX109.5M22.6
✓XXX109.5M21.1
XXXX109.5M20.8
+ +Table 5. Ablation analysis of each module in GrowCLIP at Growth Step 2. (Para.: parameter of the model. Acc.: top-1 accuracy (\%) of zero-shot image classification on ImageNet.) + +
Para.Acc.
CLIP-ViT-B/16 (TFS)149.6M28.3
CLIP-ViT-L/16 (TFS)431.1M29.9
GrowCLIP-S4 (TFS)188.4M33.3
GrowCLIP-S4 (ours)188.4M36.1
+ +ported for comparison and the results of the ablation study are shown in Table 5. As can be seen, all above four components have positive effects on the final results. Specifically, the supernet training will influence the grown model without PIM@3, which results in $1.2\%$ drop. Compared with GAS method, we randomly choose an architecture in the growth space, which has a worse performance from $24.5\%$ to $22.6\%$ . With ST, we observe a remarkable improvement of $1.5\%$ . GrowCLIP with PIM@1 solves the problem of local minimum dilemma and brings a $0.3\%$ improvement. + +# 4.5. Analysis + +# Why can GrowCLIP improve the performance? + +Compared with baseline models, the gains of our GrowCLIP come from two folds: (i) GrowCLIP can search the architecture that matches the current data with Growth architecture selection (GAS). (ii) Parameter inheriting with momentum (PIM) benefits GrowCLIP from yesterday's model. To demonstrate their effectiveness, we train the architecture selected by GrowCLIP at step 4 from scratch, and the result is shown in Table 6. Compared with ViT-B/16, the growth architecture selection can bring $5\%$ improvement on zero-shot image classification of ImageNet. Furthermore, it is even better than ViT-L/16, which is two and a half times as big as our model. More experiments have proved the effectiveness of GAS in the appendix. Comparing our GrowCLIP-S4 with GrowCLIP-S4 (TFS), we observe a remarkable improvement of $2.8\%$ with PIM. + +# Why can GrowCLIP improve the training efficiency? + +Our GrowCLIP can improve training efficiency, due to GAS and PIM, and the result is shown in Table 7. (i) Compared with the fixed architecture method, GrowCLIP has a smaller model with the GAS, so we reduce training costs + +Table 6. Ablation analysis of GrowCLIP-S4. + +
Para.Acc.EpochCost
CLIP-S2(TFS)149.6M20.930336
GrowCLIP-S2 (ours)116.6M25.730288
GrowCLIP-S4 (TFS)188.4M33.3301230
GrowCLIP-S4 (ours)188.4M33.39722
NAS-S4325.9M34.2605280
GrowCLIP-S4 (ours)188.4M36.1301583
+ +Table 7. Efficiency analysis. (Cost: GPU hours) + +
Para.Acc.
GrowCLIP-S1431.1M14.6
GrowCLIP-S2431.2M26.2
+ +Table 8. The architecture of GrowCLIP-S1 with bigger initialized model (ViT-L/16). + +at the early steps (48 GPU hours saved at step 2). At the later steps, GrowCLIP uses fewer epochs to achieve good performance because of PIM. As shown in Figure 1, GrowCLIP has far outperformed CLIP (TFS) only after supernet training. To demonstrate the effectiveness of PIM, we compare GrowCLIP-S4 with GrowCLIP-S4 (TFS). GrowCLIP only uses 9 epochs to achieve comparable performance with PIM. (ii) Compared with the NAS method, GrowCLIP inherits the parameters of the old model with momentum, which only takes 2 epochs to fine-tune the supernet and 2 epochs to train the supernet. In contrast, one-shot NAS methods consume about 3.3X resources. + +Can GrowCLIP find the relationship between model and data? When the model scale is sufficient large, there is no need to enlarge the model. We conduct the experiment started from ViT-L/16. And the result in Table 8 shows that the larger model startpoint hardly grows at step 2, which means that our GrowCLIP can find the relationship between the model and data. In our online learning scenario, the initial architecture of the model is not important due to our growth pipeline. If the model scale is smaller at the beginning, GrowCLIP will select a larger enough model for growth. In contrast, if the model scale is too large initially, GrowCLIP will stop growing. + +What is the influence when growth happens in different dimensions? As shown in Table 9, different growth dimensions have different influences on the performance. By analyzing the experimental results, we conclude 4 laws, which can provide a reference for the design of cross-modal model architecture: (i) Enlarging the smaller model bring a greater improvement than the bigger one. For example, when the growth appears in transformer blocks in the image encoder, the accuracy can gain $0.93\%$ improvement at step 2 but only $0.12\%$ at step 4. (ii) We need to leave the image + +
Step 2Step 3Step 4
DimensionPara.Acc.Para.Acc.Para.Acc.
+4 bI+7.1M+0.93+19.7M+0.46+19.7M+0.12
+4 hI+20.1M+0.34+67.3M+0.12+105.1M+0.03
+2 lI+0.1M+0.84+0.1M+0.07+0.1M+0.10
+4 bT+3.2M+0.01+12.6M+0.08+12.6M+0.01
+4 hT+27.0M+0.22+45.0M+0.01+69.5M-0.04
+4 bS+7.1M+1.03+19.7M+0.19+19.7M-0.07
+ +Table 9. The ablation of growth space corresponding to different dimensions. $(+4b^{I}$ : the number of transformer blocks in image encoder is increased by 4) + +encoder with more capacity. Compared with text encoder, the improvement of growing the image encoder is more significant at each step. (iii) The convolutional layers is a very important component in the image encoder, which can help the model to better extract visual feature. (iv) The shared encoder can enhance the degree of cross-modal fusion and bring a remarkable improvement. However, do not use a large number of transformers in the shared encoder, or the model become redundant with no noticeable performance improvement. + +# 5. Conclusion + +In this paper, we study the online learning case in cross-modal pre-trained models. The small-scale architecture may limit the final performance due to the limited parameters when data grows. Based on this real-world scenario, we propose a data-driven automatic model growing algorithm for language-image pre-training - GrowCLIP, which upgrades the model size when abundant new data arrives. We adopt parameter inheriting with momentum (PIM) to utilize the knowledge in yesterday's model and seek out the optimal growth architecture at each growth step with growth architecture selection (GAS). GrowCLIP improves the performance on downstream tasks, including zero-shot image classification and retrieval. The effectiveness of each component of GrowCLIP is proved on ablation studies. And we provide a reference for the design of cross-modal model architecture. + +Limitations In this work, we only prove the effectiveness of GrowCLIP with public pre-training dataset - CC12M. In future works, we will extend our method to a real-world scenario, where the VLP model keeps training with constantly updated data crawled from the web. + +# Acknowledgment + +This work was supported in part by National Key R&D Program of China under Grant No.2020AAA0109700, Guangdong Outstanding Youth Fund (Grant No. + +2021B1515020061), Shenzhen Science and Technology Program (Grant No.RCYX20200714114642083), Shenzhen Fundamental Research Program(Grant No. JCYJ20190807154211365), Nansha Key RD Program under Grant No.2022ZD014 and Sun Yat-sen University under Grant No.22lgqb38 and 76160-12220011. We thank MindSpore for the partial support of this work, which is a new deep learning computing framework $^{2}$ . + +# References + +[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems. +[2] R. Aljundi, F. Babiloni, M. Elhoseiny, M. Rohrbach, and T. Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European conference on computer vision, 2018. +[3] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Zitnick, and D. Parikh. Vqa: Visual question answering. In International Conference on Computer Vision, 2015. +[4] J. Ash and R. Adams. On warm-starting neural network training. 2020. +[5] A. Barriuso and A. Torralba. Notes on image annotation. Preprint arXiv:1210.3448, 2012. +[6] L. Bossard, M. Guillaumin, and L. Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, 2014. +[7] P. Buzzega, M. Boschini, A. Porrello, D. Abati, and S. Calderara. Dark experience for general continual learning: a strong, simple baseline. In Advances in Neural Information Processing Systems, 2020. +[8] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. +[9] S. Changpinyo, P. Sharma, N. Ding, and R. Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Computer Vision and Pattern Recognition, 2021. +[10] M. Chen, H. Peng, J. Fu, and H. Ling. Autoformer: Searching transformers for visual recognition. In International Conference on Computer Vision, 2021. +[11] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020. +[12] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Computer Vision and Pattern Recognition, 2014. +[13] M. Cornia, L. Baraldi, and R. Cucchiara. Show, control and tell: A framework for generating controllable and grounded captions. In Computer Vision and Pattern Recognition, 2019. +[14] J. Deng. A large-scale hierarchical image database. 2009. +[15] J. Devlin, M. Chang, K. Lee, and K. Toutanova. BERT: Pretraining of Deep Bidirectional Transformers for Language Understanding. In North American Chapter of the Association for Computational Linguistics, 2019. +[16] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. +[17] H. Grabner and H. Bischof. On-line boosting and vision. In Computer Vision and Pattern Recognition, 2006. +[18] J. Grill, F. Strub, F. Altché, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Ghesh + +laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. 2020. +[19] Z. Guo, B. Avila Pires, B. Piot, J. Grill, F. Altché, R. Munos, and M. Azar. Bootstrap latent-predictive representations for multitask reinforcement learning. In International Conference on Machine Learning, 2020. +[20] Z. Guo, X. Zhang, H. Mu, W. Heng, Z. Liu, Y. Wei, and J. Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, 2020. +[21] L Zhou B Xiao C. Noel Y Cheng R Xu S Chang H, You and Yuan L. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision, pages 69-87. Springer, 2022. +[22] R. Hadsell, D. Rao, A. Rusu, and R. Pascanu. Embracing change: Continual learning in deep neural networks. Trends in Cognitive Sciences, 2020. +[23] E. Hazan et al. Introduction to online convex optimization. Foundations and Trends® in Optimization, 2016. +[24] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In Conference on Computer Vision and Pattern Recognition, 2022. +[25] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In Computer Vision and Pattern Recognition, 2020. +[26] X. Hu, Z. Gan, J. Wang, Z. Yang, Z. Liu, Y. Lu, and L. Wang. Scaling up vision-language pre-training for image captioning. Preprint arXiv:2111.12233, 2021. +[27] X. Hu, X. Yin, K. Lin, L. Wang, L. Zhang, J. Gao, and Z. Liu. Vivo: Surpassing human performance in novel object captioning with visual vocabulary pre-training. 2020. +[28] C. Jia, Y. Yang, Y. Xia, Y. Chen, Z. Parekh, H. Pham, Q. Le, Y. Sung, Z. Li, and T. Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. +[29] J. Kaplan, S. McCandlish, T. Henighan, T. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei. Scaling laws for neural language models. Preprint arXiv:2001.08361, 2020. +[30] W. Kim, B. Son, and I. Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, 2021. +[31] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. National Academy of Sciences, 2017. +[32] A. Krizhevsky, G. Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[33] Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, and R. Soricut. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. In International Conference on Learning Representations, 2020. +[34] C Li, B. Zhuang, G. Wang, X. Liang, X. Chang, and Y. Yang. Automated progressive learning for efficient training of vision transformers. In Computer Vision and Pattern Recognition, 2022. + +[35] F. Li, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 2006. +[36] J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. +[37] L. Li, M. Yatskar, D. Yin, C. Hsieh, and K. Chang. Visualbert: A simple and performant baseline for vision and language. Preprint arXiv:1908.03557, 2019. +[38] X. Li, X. Yin, C. Li, P. Zhang, X. Hu, L. Zhang, L. Wang, H. Hu, L. Dong, F. Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, 2020. +[39] Y. Li, F. Liang, L. Zhao, Y. Cui, W. Ouyang, J. Shao, F. Yu, and J. Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. Preprint arXiv:2110.05208, 2021. +[40] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. +[41] C. Liu, B. Zoph, M. Neumann, J. Shlens, W. Hua, L. Li, F. Li, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. In European conference on computer vision, 2018. +[42] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. Preprint arXiv:1806.09055, 2018. +[43] W. Liu, S. Chen, L. Guo, X. Zhu, and J. Liu. Cptr: Full transformer network for image captioning. Preprint arXiv:2101.10804, 2021. +[44] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Preprint arXiv:1907.11692, 2019. +[45] I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016. +[46] J. Lu, V. Goswami, M. Rohrbach, D. Parikh, and S. Lee. 12-in-1: Multi-task vision and language representation learning. In Computer Vision and Pattern Recognition, 2020. +[47] M. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Computer Vision, Graphics & Image Processing, 2008. +[48] O. Parkhi, A. Vedaldi, A. Zisserman, and C. Jawahar. Cats and dogs. In Computer Vision and Pattern Recognition, 2012. +[49] H. Peng, H. Du, H. Yu, Q. Li, J. Liao, and J. Fu. Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. In Advances in Neural Information Processing Systems, 2020. +[50] B. Plummer, L. Wang, C. Cervantes, J. Caicedo, J. Hockenmaier, and S. Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-tosentence models. In International Conference on Computer Vision, 2015. + +[51] R. Pogodin and T. Lattimore. On first-order bounds, variance and gap-dependent bounds for adversarial bandits. In Uncertainty in Artificial Intelligence, 2020. +[52] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. +[53] E. Real, A. Aggarwal, Y. Huang, and Q. Le. Regularized Evolution for Image Classifier Architecture Search. In AAAI Conference on Artificial Intelligence, 2019. +[54] E. Real, S. Moore, A. Selle, S. Saxena, Y. Suematsu, J. Tan, Q. Le, and A. Kurakin. Large-scale evolution of image classifiers. In International Conference on Machine Learning, 2017. +[55] H Xu J. Han T. Tuytelaars Z. Li S. Yan, L. Hong and X. He. Generative negative text replay for continual vision-language pretraining. In European Conference on Computer Vision, pages 22-38. Springer, 2022. +[56] A. Saffari, C. Leistner, J. Santner, M. Godec, and H. Bischof. On-line random forests. In International Conference on Computer Vision workshops, 2009. +[57] S. Shalev-Shwartz et al. Online learning and online convex optimization. Foundations and Trends® in Machine Learning, 2012. +[58] S. Song and B. Choi. A study on continuous follow-the-leader (ftl) gaits: an effective walking algorithm over rough terrain. Mathematical biosciences, 1989. +[59] T Srinivasan, T Chang, L. Alva, G Chochlakis, M Rostami, and J Thomason. Climb: A continual learning benchmark for vision-and-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track. +[60] M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard, and Q. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Computer Vision and Pattern Recognition, 2019. +[61] B. Thomee, D. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland, D. Borth, and L. Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. +[62] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, and I. Polosukhin. Attention Is All You Need. In Neural Information Processing Systems, 2017. +[63] Haonan Wang, Minbin Huang, Runhui Huang, Lanqing Hong, Hang Xu, Tianyang Hu, Xiaodan Liang, and Zhenguo Li. Boosting visual-language models by exploiting hard samples. arXiv preprint arXiv:2305.05208, 2023. +[64] S. Xie, H. Zheng, C. Liu, and L. Lin. SNAS: Stochastic Neural Architecture Search. In International Conference on Learning Representations, 2019. +[65] L. Yao, R. Huang, L. Hou, G. Lu, M. Niu, H. Xu, X. Liang, Z. Li, X. Jiang, and C. Xu. Filip: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. +[66] Y. You, J. Li, S. Reddi, J. Hseu, S. Kumar, S. Bhojanapalli, X. Song, J. Demmel, K. Keutzer, and C. Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. + +In International Conference on Learning Representations, 2020. +[67] F. Yu, J. Tang, W. Yin, Y. Sun, H. Tian, H. Wu, and H. Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3208-3216, 2021. +[68] J. Yu, Z. Wang, V. Vasudevan, L. Yeung, M. Seyedhosseini, and Y. Wu. Coca: Contrastive captioners are image-text foundation models. Preprint arXiv:2205.01917, 2022. +[69] X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022. +[70] B. Zoph, V. Vasudevan, J. Shlens, and Q. Le. Learning Transferable Architectures for Scalable Image Recognition. In Computer Vision and Pattern Recognition, 2018. \ No newline at end of file diff --git a/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/images.zip b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7b103e8597110b32fe394dcf8c28230cb8728bb --- /dev/null +++ b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4612d151ad8f7317cbce1ebea49737ab538122b09360c8ab3dbc94cf46e565ad +size 758050 diff --git a/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/layout.json b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8850e8b5001a9fb5e23ca5c481e7ce3b76b050dc --- /dev/null +++ b/2023/GrowCLIP_ Data-Aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-Training/layout.json @@ -0,0 +1,9440 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 468, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 468, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 468, + 140 + ], + "type": "text", + "content": "GrowCLIP: Data-aware Automatic Model Growing for Large-scale Contrastive Language-Image Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": "Xinchi Deng" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " Han Shi" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " Runhui Huang" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " Changlin Li" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " Hang Xu" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " \nJianhua Han" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " James Kwok" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " Shen Zhao" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " Wei Zhang" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": " Xiaodan Liang" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": "Sun Yat-sen University " + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab \n" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": "University of Technology Sydney " + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 160, + 493, + 232 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology \ndengxch5@mail2.sysu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 292, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 289, + 651 + ], + "type": "text", + "content": "Cross-modal pre-training has shown impressive performance on a wide range of downstream tasks, benefiting from massive image-text pairs collected from the Internet. In practice, online data are growing constantly, highlighting the importance of the ability of pre-trained model to learn from data that is continuously growing. Existing works on cross-modal pre-training mainly focus on training a network with fixed architecture. However, it is impractical to limit the model capacity when considering the continuously growing nature of pre-training data in real-world applications. On the other hand, it is important to utilize the knowledge in the current model to obtain efficient training and better performance. To address the above issues, in this paper, we propose GrowCLIP, a data-driven automatic model growing algorithm for contrastive language-image pre-training with continuous image-text pairs as input. Specifically, we adopt a dynamic growth space and seek out the optimal architecture at each growth step to adapt to online learning scenarios. And the shared encoder is proposed in our growth space to enhance the degree of cross-modal fusion. Besides, we explore the effect of growth in different dimensions, which could provide future references for the design of cross-modal model architecture. Finally, we employ parameter inheriting with momentum (PIM) to maintain the previous knowledge and address the issue of the local minimum dilemma. Compared with the existing methods, GrowCLIP improves " + }, + { + "bbox": [ + 46, + 292, + 289, + 651 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 292, + 289, + 651 + ], + "type": "text", + "content": " average top-1 accuracy on zero-shot image classification of 9 downstream tasks. As for zero-shot image retrieval, GrowCLIP can improve " + }, + { + "bbox": [ + 46, + 292, + 289, + 651 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 292, + 289, + 651 + ], + "type": "text", + "content": " for top-1 image-to-text recall on Flickr30K dataset." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 258, + 525, + 410 + ], + "blocks": [ + { + "bbox": [ + 314, + 258, + 525, + 410 + ], + "lines": [ + { + "bbox": [ + 314, + 258, + 525, + 410 + ], + "spans": [ + { + "bbox": [ + 314, + 258, + 525, + 410 + ], + "type": "image", + "image_path": "75efdd0b644164ce01f11930b0362fa1ddf453c61cbbaaa819538b4866d032d6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 411, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 411, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 547, + 468 + ], + "type": "text", + "content": "Figure 1. Top-1 accuracy " + }, + { + "bbox": [ + 305, + 411, + 547, + 468 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 411, + 547, + 468 + ], + "type": "text", + "content": " of zero-shot image classification on ImageNet of GrowCLIP and baselines during training at step 4, where the horizontal dotted lines mean the process of supernet training. Our GrowCLIP has the best performance and is more efficient compared with other baselines." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 487, + 387, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 387, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 387, + 500 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "text", + "content": "Recently, large-scale pre-trained models illustrate the potential performance among different fields including computer vision (CV) [24, 25, 11] and natural language processing (NLP) [15, 44, 33]. Cross-modal models like CLIP [52], ALIGN [28], FILIP [65], BLIP [36] also demonstrate remarkable success across various vision-language downstream tasks. Note that these models are usually built in dual-stream architectures, which consist of the image and text encoders to extract the feature of image and text inputs, respectively. The alignment between these two features is then performed under the contrastive objective to learn the alignment between different modals." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "However, training these cross-modal models requires a large amount of image-text pairs collected from the Internet, e.g., 400M image-text pairs for CLIP [52] and 300M for FILIP [65]. Most existing cross-modal methods [52, 28, 65, 39, 36, 68, 63, 1] directly train the final large-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22178" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 276, + 118 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 276, + 118 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 276, + 118 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 276, + 118 + ], + "type": "table", + "html": "
CC12M10%20%50%100%
CLIP-0.5-ViT-B/166.611.119.025.7
CLIP-ViT-B/166.511.620.928.3
", + "image_path": "a0ea8e71f475b5a9f206990de9099cf88db1c7c657cc546f24cf8d1835736645.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 121, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 177 + ], + "type": "text", + "content": "Table 1. Result of different model sizes with different data sizes: Top-1 accuracy (\\%) of zero-shot image classification on ImageNet. The depth and the width of 0.5-ViT-B/16 are both one half of ViT-B/16. The relative performance ranking is dependent on the size of the training data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 361 + ], + "type": "text", + "content": "scale model using the completed image-text pair datasets collected at some point in time. However, it is impractical to regard cross-modal pre-training as \"disposable\" training without considering the continuously growing nature of pre-training data or domain in real-world applications. For instance, mountains of new knowledge is generated on the Internet every day, which can be used to further improve the performance of existing pre-trained cross-modal models. Therefore, in this paper, we consider the training of large-scale cross-modal pre-trained models as an online learning problem [17, 56, 57] with continuous image-text pairs as input. Different from the standard continual learning setting [59, 55, 22, 31, 2, 7], the previous image-text pairs can also be achieved in our setting since we assume the size of training data is gradually increased over time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 674 + ], + "type": "text", + "content": "One tough problem in the online learning scenario is that the model capacity should be related to the size of training data [16, 69, 29, 26]. For example, it's observed that large ViT models perform worse than ResNets when pre-trained on small datasets, while the result is opposite when they are both pre-trained on larger datasets [16]. To verify the relationship between model capability and data size, we split Conceptual 12M (CC12M) dataset [9] and test the performance of ViT-B/16 with different scales [52]. As shown in Table 1, the relative performance ranking is dependent on the size of the training data. Given a smaller dataset (e.g., " + }, + { + "bbox": [ + 46, + 365, + 287, + 674 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 674 + ], + "type": "text", + "content": " CC12M), the performance of 0.5-ViT-B/16 is comparable with ViT-B/16. In contrast, when data is sufficient (e.g., " + }, + { + "bbox": [ + 46, + 365, + 287, + 674 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 674 + ], + "type": "text", + "content": " CC12M), the performance of ViT-B/16 is better than the smaller model. Thus, the unchanged model is not practical for this real-time learning setting and it's still an open issue on how to modify and train our model with the growing data. Besides, how to efficiently make use of knowledge of previous model when new data is coming remains an open problem. One direct solution is to fine-tune the model with the updated training dataset. However, training with previous pre-trained parameters of the same model will deteriorate the performance [4] due to the parameter inheriting issue. As shown in Figure 1, the CLIP training with pre-training (TWP) have worse performance than the one training from scratch (TFS)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "To address the above issues, we propose a data-aware automatic model growing method (denoted as GrowCLIP) for large-scale contrastive language-image pre-training, which" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 275 + ], + "type": "text", + "content": "performs a model growth process considering the gradually increased pre-training data. Specially, when training data grows dynamically, we adopt different cross-modal network architectures via the customized neural architecture search (NAS) to make the network pre-training more efficient. Different from traditional NAS approaches [70, 53, 64], we introduce a cross-modal customized NAS by defining a dynamic search space named growth space, which is enlarged when more data comes, and proposing a shared encoder search space to enhance the degree of cross-modal fusion. To utilize the architecture at the previous step more efficiently, the parameters of the new architecture are also inherited from the old one with momentum to maintain the previous knowledge and address the issue of local minimum dilemma. Finally, growth architecture selection procedure is performed to select the optimal model architecture at each step, considering the performance and model size." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "text", + "content": "Experiments are conducted by averagely dividing the Conceptual 12M (CC12M) into 4 growth steps under the online learning setting. As depicted in Table 1, compared with the existing methods, our GrowCLIP has the best performance and is more efficient. Specially, experimental results show that our GrowCLIP can improve up to " + }, + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "text", + "content": " average top-1 accuracy on zero-shot image classification of 9 downstream tasks compared with the existing methods. As for zero-shot image-text retrieval, GrowCLIP has a " + }, + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "text", + "content": " improvement for top-1 image-to-text recall on Flickr30K [50] dataset and " + }, + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 275, + 545, + 407 + ], + "type": "text", + "content": " on MSCOCO [40]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 491 + ], + "type": "text", + "content": "To summarize, the contributions of this paper are listed as follows: (i) To adapt to the growing data scenario, we propose a data-aware automatic model growing method, named GrowCLIP. (ii) We provide some insights for the design of cross-modal model architecture. (iii) The extensive experiment results illustrate the effectiveness of our approach on zero-shot classification and retrieval tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Vision-Language Pre-training Models Inspired by the success of pre-training in computer vision (CV) and natural language processing (NLP), a boosting number of research works in the domain of vision-language pretraining (VLP) has recently surged to pursue a unified multi-modality representation. Vision-language pretraining trains the model on large-scale image-text pairs to improve performance of downstream vision and language tasks including image classification [14, 35, 32], retrieval [50, 40], Image Captioning (IC) [43, 13, 27, 26], Visual Question and Answer (VQA) [3, 46, 67] and so on. The VLP model can be categorized into two categories: (i) The dual-stream model, e.g., CLIP [52], ALIGN [28], FILIP [65], processes visual and textual tokens separately with two parallel streams to acquire the representation and then fuse them through interactive module. (ii) The single-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22179" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "stream model, e.g., Oscar [38], ViLT [30], VisualBERT [37], directly combines the extracted visual and textual tokens and feeds them into the transformer-based model. The public pre-training datasets of VLP model include available datasets like YFCC100M [61] and CC12M [9]. CLIP [52], ALIGN [28], FILIP [65] and LEMON [26] adopt large-scale image-text pairs crawled from the web to show more powerful models. Different from these works that treat this process as offline learning, we investigate the online learning case in vision-language pre-training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "type": "text", + "content": "Online Learning Online learning is a problem setting of machine learning in which data becomes available in a sequential order [17, 56, 57]. Different from continual learning [22, 31, 2, 7], online learning can reach previous data all the time. As opposed to batch learning techniques which generate the best model by learning on the entire training dataset at once, online learning is used to update our predictor for future data at each growth step. Follow the leader [58] is the simplest learning rule which assumes the hypothesis that the leader has the least loss overall past rounds. Follow the regularised leader [51] is a modification variant and obtain better regret bounds by learning a regularisation function. Online convex optimization [23] is a general algorithm framework for decision making which targets convex optimization. To adapt online learning scenario, we propose a data-aware automatic model growing algorithm that grows the model with increasing training data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 407, + 289, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 289, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 289, + 600 + ], + "type": "text", + "content": "Neural Architecture Search Neural Architecture Search (NAS) aims to automatically design the state-of-the-art neural networks. Early works [70, 54, 60] are based on reinforcement learning and evolutionary algorithms. In the later work, various techniques are used to reduce the computational complexity. Gradient-based methods [42, 41] regard the network architecture as a set of parameters and adopt back-propagation algorithm to optimize them. One-shot NAS methods [49, 10, 20] train a supernet to avoid training each subnet from scratch. Unfortunately, one-shot NAS methods have to spend lots of resources to train the supernet. In contrast, the method we propose inherits the parameters of the old model with momentum, which only takes few epochs to train the supernet. Furthermore, GrowCLIP can modify the neural architecture to adapt to the size of the training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 130, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 130, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 130, + 624 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "In this section, we describe the online learning problem setting and its challenges for cross-modal pre-training (Section 3.1). To adapt the growth data, we propose a dynamic growth space, including a proposed shared encoder (Section 3.2). To rise to the challenge of the local minimum dilemma, we propose parameter inheriting with momentum (Section 3.3). The whole pipeline of our method is illus" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 392, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 392, + 83 + ], + "type": "text", + "content": "trated in Section 3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 91, + 479, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 479, + 105 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 479, + 105 + ], + "type": "text", + "content": "3.1. Problem Setting and Challenges" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": "Problem Setting The target of online learning is obtaining the optimal model at each growth step of the data. Let " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{t - 1} = \\{\\pmb{x}_i^I,\\pmb{x}_i^T\\}_{i = 1}^k" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": " be the training data at growth step " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\pmb{x}^I" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\pmb{x}^T" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": " are image and text samples and " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": " is the training data size. At the next growth step " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": ", the enlarged training dataset " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": " satisfies " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{t - 1}\\subseteq \\mathcal{D}^t" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": " due to the incoming data. Given an optimal model from growth step " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": ", our aim is to achieve the optimal model (with optimal architecture " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\psi^{*}" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": " and parameters " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\omega^{*}" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": ") based on the enlarged dataset " + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 304, + 110, + 547, + 230 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 242, + 545, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 242, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 354, + 242, + 545, + 262 + ], + "type": "interline_equation", + "content": "\\left\\{\\psi^ {*}, \\omega^ {*} \\right\\} = \\underset {\\psi , \\omega} {\\arg \\min } \\mathcal {L} (\\psi , \\omega ; \\mathcal {D} ^ {t}). \\tag {1}", + "image_path": "1c004b6bca51255d5c1357e8ea56bce2cc488ec28c8dfb7a467ca6fcd13d54ec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 267, + 547, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 547, + 375 + ], + "type": "text", + "content": "Challenge of Data Growth As discussed in Table 1, the size of the training dataset is related to the optimal architecture selection. Given a smaller dataset, the small-scale architecture has comparable performance to large-scale architecture but fewer parameters are deployed. When data grows, the small-scale architecture may limit the final performance due to the limited parameters. However, few works notice the effect of data size in online learning settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 376, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 495 + ], + "type": "text", + "content": "Challenge of Parameter Inheriting As investigated, the performance of model trained from scratch is much better than trained with pre-training. This is probably caused by the influence of the inheriting parameters trained with the previous smaller dataset, which makes it difficult for the model to escape from the local minima in the grown dataset. However, training from scratch is costly due to multiple growth steps in online learning. The balance between training from scratch and training with pre-training is worthy of consideration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "text", + "content": "To solve the above problems, we propose an automatic model growing method - GrowCLIP (shown in Figure 2), which can modify the neural architecture at different growth steps in online learning case. To adapt to the growing data scenario, we propose a growth space for each growth step, where the scale of architecture candidates among this space is larger when data grows. To alleviate the parameter inhering issue with efficiency, we perform parameter inhering with momentum such that the trade off between training from scratch and training with pre-training is balanced." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 397, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 397, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 397, + 636 + ], + "type": "text", + "content": "3.2. Growth Space" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Model Architecture Following the basic neural architecture in CLIP [52], we adopt a dual-stream model with image encoder " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": ". To adapt the cross-modality scenarios, we propose a shared encoder " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": ", in which the image and text share the transformers except layernorm layers. The shared encoder not only can reduce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22180" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 321, + 290 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 321, + 290 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 321, + 290 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 321, + 290 + ], + "type": "image", + "image_path": "6edcf2db6e9d07d8e12003be64e20d8d8870f81efafe8db811bdde3147900670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 301, + 546, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 546, + 324 + ], + "type": "text", + "content": "Figure 2. The overview of GrowCLIP. As shown in Figure (a), when the data grows from growth step " + }, + { + "bbox": [ + 46, + 301, + 546, + 324 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 301, + 546, + 324 + ], + "type": "text", + "content": " to growth step " + }, + { + "bbox": [ + 46, + 301, + 546, + 324 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 301, + 546, + 324 + ], + "type": "text", + "content": ", our model grows adaptively as well. Figure (b) illustrates the detail of model growing and we simplify the representation of model architecture." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 547, + 290 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 547, + 290 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 547, + 290 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 547, + 290 + ], + "type": "image", + "image_path": "c9d953690f661e72d8c03171fbc2af26326642285bc46b92416cd579b8764279.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "content": "model parameters, but also enhance the degree of cross-modal fusion. Given an image " + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^I" + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "content": " and a text " + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_j^T" + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "content": ", the model outputs the image representations " + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "inline_equation", + "content": "I_i = h(f(\\boldsymbol{x}_i^I))" + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "content": " and text representations " + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "inline_equation", + "content": "T_j = h(g(\\boldsymbol{x}_j^T))" + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "content": ". The similarity score computation of between image " + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^I" + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_j^T" + }, + { + "bbox": [ + 46, + 343, + 287, + 408 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 415, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 415, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 108, + 415, + 287, + 430 + ], + "type": "interline_equation", + "content": "s _ {i, j} = h \\left(f \\left(\\boldsymbol {x} _ {i} ^ {I}\\right)\\right) ^ {\\top} h \\left(g \\left(\\boldsymbol {x} _ {j} ^ {T}\\right)\\right), \\tag {2}", + "image_path": "f8fc630b9e1525e4d05f2df5346915253a4a72bca1d93eb01dd05a3a4adcd76b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "inline_equation", + "content": "s_{i,j}" + }, + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "text", + "content": " denotes the similarity of the " + }, + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "text", + "content": "-th image to the " + }, + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 437, + 287, + 459 + ], + "type": "text", + "content": "-th text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "content": "In each training batch, we sample " + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "content": " image-text pairs " + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_i^I,\\pmb{x}_i^T\\}_{i = 1}^n" + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "content": " from the dataset. For image " + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^I" + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "content": " in the image-text pair " + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "< x_{i}^{I},x_{i}^{T}>" + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "x_{i}^{T}" + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "content": " is its positive, while the other texts will be used as in-batch negatives. The image-to-text contrastive loss " + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i^I" + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^I" + }, + { + "bbox": [ + 47, + 461, + 287, + 521 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 529, + 287, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 529, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 76, + 529, + 287, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} ^ {I} \\left(\\boldsymbol {x} _ {i} ^ {I}, \\left\\{\\boldsymbol {x} _ {j} ^ {T} \\right\\} _ {j = 1} ^ {n}\\right) = - \\frac {1}{n} \\log \\frac {\\exp \\left(s _ {i , i}\\right)}{\\sum_ {j} \\exp \\left(s _ {i , j}\\right)}, \\tag {3}", + "image_path": "dfcc4152da6cafeee1b10a1b39895a9d4329ee5d1daa07c93a6ad89baf016f1e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 265, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 265, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 265, + 576 + ], + "type": "text", + "content": "Similarly, the text-to-image contrastive loss for " + }, + { + "bbox": [ + 47, + 563, + 265, + 576 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^T" + }, + { + "bbox": [ + 47, + 563, + 265, + 576 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 583, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 583, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 75, + 583, + 287, + 611 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} ^ {T} \\left(\\boldsymbol {x} _ {i} ^ {T}, \\left\\{\\boldsymbol {x} _ {j} ^ {I} \\right\\} _ {j = 1} ^ {n}\\right) = - \\frac {1}{n} \\log \\frac {\\exp \\left(s _ {i , i}\\right)}{\\sum_ {j} \\exp \\left(s _ {i , j}\\right)}, \\tag {4}", + "image_path": "9748462b2e06daa18738ff29ae7fa04ec929007f1784d15a690500625a2863dd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "type": "text", + "content": "Totally, the contrastive loss of this mini-batch can be represented by " + }, + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\frac{1}{2}\\sum_{i=1}^{a}\\left(\\mathcal{L}_i^I + \\mathcal{L}_i^T\\right)" + }, + { + "bbox": [ + 46, + 618, + 287, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "text", + "content": "Growth Space of image encoder For the image encoder, we adopt a mixed architecture of convolutional layer and transformer block [62]. There are three variable factors in our growth space design: (i) the number of convolutional layers " + }, + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "inline_equation", + "content": "l^I" + }, + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "text", + "content": "; (ii) the number of transformer blocks " + }, + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "inline_equation", + "content": "b^I" + }, + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "text", + "content": "; (iii) the number of transformer heads " + }, + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "inline_equation", + "content": "h^I" + }, + { + "bbox": [ + 46, + 643, + 287, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 343, + 546, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 343, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 546, + 390 + ], + "type": "text", + "content": "Growth Space of text encoder The text encoder part is a transformer-based architecture, whose growth space includes two variable factors: (i) the number of transformer blocks " + }, + { + "bbox": [ + 304, + 343, + 546, + 390 + ], + "type": "inline_equation", + "content": "b^{T}" + }, + { + "bbox": [ + 304, + 343, + 546, + 390 + ], + "type": "text", + "content": " and (ii) the number of transformer heads " + }, + { + "bbox": [ + 304, + 343, + 546, + 390 + ], + "type": "inline_equation", + "content": "h^{T}" + }, + { + "bbox": [ + 304, + 343, + 546, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 392, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 426 + ], + "type": "text", + "content": "Growth Space of shared encoder The shared encoder is a transformer-based architecture, whose growth space includes the number of transformer blocks " + }, + { + "bbox": [ + 304, + 392, + 545, + 426 + ], + "type": "inline_equation", + "content": "b^{S}" + }, + { + "bbox": [ + 304, + 392, + 545, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 307, + 434, + 547, + 480 + ], + "blocks": [ + { + "bbox": [ + 307, + 434, + 547, + 480 + ], + "lines": [ + { + "bbox": [ + 307, + 434, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 547, + 480 + ], + "type": "table", + "html": "
Image encoderText encoderShared encoder
Transformer blocks{bIt-1, bIt-1+4}{bTt-1, bTt-1+4}{bSt-1, bSt-1+4}
Transformer heads{hIt-1, hIt-1+4}{hTt-1, hTt-1+4}-
Convolutional layers{lIt-1, lIt-1+2}--
", + "image_path": "ff35109c874407bdfefd856922e95d6938a4de29830c8b0f4838c6635dfa59ad.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "type": "text", + "content": "Table 2. The growth space " + }, + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "type": "text", + "content": " of GrowCLIP at the growth step t. " + }, + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "type": "inline_equation", + "content": "{b}_{t - 1}^{I},{b}_{t - 1}^{T},{b}_{t - 1}^{S},{h}_{t - 1}^{I},{h}_{t - 1}^{T},{l}_{t - 1}^{I}" + }, + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "type": "text", + "content": " are the hyperparameters of model architecture at growth step " + }, + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 305, + 488, + 547, + 521 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "content": "The detailed descriptions of growth space for three encoders are shown in Table 2. At each new growth step, the variable factors are enlarged to adapt the growth data. Specifically, the numbers of transformer blocks and transformer heads are both increased by 4, and the number of convolutional layers is increased by 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 611, + 507, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 507, + 624 + ], + "type": "text", + "content": "3.3. Parameter inheriting with momentum" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Some previous papers utilize the training results of the old model and transfer the knowledge to the new model [18, 19, 25, 4, 34]. Inspired by the above methods, we propose parameter inheriting with momentum (PIM) to avoid local minimum dilemma. The illustration of our parameter inheriting with momentum is shown in Figure 3. To preserve the knowledge from the old model, we extend the size" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22181" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 287, + 168 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 287, + 168 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 287, + 168 + ], + "type": "image", + "image_path": "8bab3a1ee52d8ce8e2c6e01e9e660787ba328cfc961bb08276127fe556278dde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 288, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 288, + 226 + ], + "type": "text", + "content": "Figure 3. Take image encoder for example of our proposed parameter inheriting with momentum. As for the initialization of the enlarged model, we preserve the existing parameters of the old model for exploitation and combine it with random initialization for exploration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 287, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 336 + ], + "type": "text", + "content": "of the old model to the enlarged model and maintain the parameters in existing parts. As for the extended layers, we copy the parameters from the last layers directly. " + }, + { + "bbox": [ + 46, + 229, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\omega_{old}" + }, + { + "bbox": [ + 46, + 229, + 287, + 336 + ], + "type": "text", + "content": " denotes the parameters of extension of the old model. Besides the above parameter inheriting term, we also add a re-initialization term with randomly initialized parameters " + }, + { + "bbox": [ + 46, + 229, + 287, + 336 + ], + "type": "inline_equation", + "content": "\\omega_{rand}" + }, + { + "bbox": [ + 46, + 229, + 287, + 336 + ], + "type": "text", + "content": " to avoid the parameter inheriting issue. The final initialized supernet is a combination of the above two terms as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 338, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 338, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\omega_ {\\text {i n h e r i t}} = \\beta \\cdot \\omega_ {\\text {o l d}} + \\gamma \\cdot \\omega_ {\\text {r a n d}}, \\tag {5}", + "image_path": "192555c3e38c33ae5e5dcfc3e8d6a25038a3da5c97dddf91bc2b78c5f3143213.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 354, + 287, + 413 + ], + "type": "text", + "content": " are two hyperparameters related to exploitation and exploration. Additionally, the parameters of new transformer heads are randomly initialized, which makes that different heads in transformer are orthogonal and the new heads learn new features from different perspectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 146, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 146, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 146, + 431 + ], + "type": "text", + "content": "3.4. Growth Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "text", + "content": "The growth pipeline at growth step " + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 437, + 287, + 461 + ], + "type": "text", + "content": " (except the first one) includes three sequential phases as shown in Figure 2 (b)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 462, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 592 + ], + "type": "text", + "content": "Phase 1: Supernet Fine-tuning As for a new growth step, a supernet " + }, + { + "bbox": [ + 46, + 462, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 462, + 287, + 592 + ], + "type": "text", + "content": " covering the whole growth space is constructed by enlarging the current model. First, we use the PIM method described in Section 3.3 to initialize the supernet. However, there are some newly added layers in the supernet compared with the old model from the last growth step, which brings a performance gap between the subnets with new layers and others. To make the supernet more robust, we slightly fine-tune the parameters of the full supernet for a few epochs to mitigate the performance gap among subnetworks after PIM initialization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "content": "Phase 2: Growth Architecture Selection Growth Architecture Selection (GAS) is the main phase in our algorithm where we select the optimal architecture in the growth space. The whole GAS process can be divided into two steps: training and searching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 688 + ], + "type": "text", + "content": "First, to find the optimal architecture " + }, + { + "bbox": [ + 46, + 653, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\psi^{*}" + }, + { + "bbox": [ + 46, + 653, + 287, + 688 + ], + "type": "text", + "content": ", we need to fully optimize all the subnet in the supernet, which covers all the candidate growth architecture:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 694, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 694, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 77, + 694, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\omega_ {s u p e r} ^ {*} = \\underset {\\omega_ {s u p e r}} {\\arg \\min } \\mathbb {E} _ {\\psi \\in \\Psi} \\left\\{\\mathcal {L} \\left(\\psi , \\omega_ {s u p e r}; \\mathcal {D} ^ {t}\\right) \\right\\}. \\tag {6}", + "image_path": "cbc6b0630bcf69c9d3d21f187a392b4018bf9dfebe325eebc67093da796502b1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "Inspired by one-shot neural architecture search [8], we optimize " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\omega_{super}" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " in Equation 6 by uniformly sampling subnets " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " from the growth space " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " and update the parameters of the sampled subnets in each iteration. More specifically, at each step, we first randomly sample a subnet " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " from the growth space " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": ", then optimize the parameter of this subnet by minimizing the contrastive loss between image and text. This can be considered as solving Equation 6 via Monte Carlo sampling. By iterative training, these subnets can be fully optimized in a few epochs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "text", + "content": "Second, after training the supernet, we search for the optimal subnet in the growth space " + }, + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "text", + "content": ". Here, we consider both the performance and the number of parameters of the architecture, and select the optimal architecture " + }, + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\psi^{*}" + }, + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "text", + "content": " using the composite metric:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 336, + 260, + 545, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 260, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 336, + 260, + 545, + 287 + ], + "type": "interline_equation", + "content": "\\psi^ {*} = \\underset {\\psi \\in \\Psi} {\\arg \\max } \\left\\{\\mathcal {A} (\\psi) + \\alpha \\frac {\\mathcal {N} ^ {t - 1}}{\\mathcal {N} ^ {t}} \\frac {\\mathcal {P} (\\Psi)}{\\mathcal {P} (\\psi)} \\right\\}, \\tag {7}", + "image_path": "fa706add93f1e694f12fc19d07dfa8cf33c3bc15ee32066a825447ef79a6cc5c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": " is the top-1 accuracy of test dataset, " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{N}^t" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": " denotes the number of data on growth step " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot)" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": " represents the total amount of model parameters and " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": " is the trade-off hyperparameter. When we get a large number of new coming data, " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\frac{\\mathcal{N}^{t - 1}}{\\mathcal{N}^t}" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": " will get small, which means that we focus more on the performance and encourage the growth of models. As our growth space is relatively small, traversal evaluation of all the candidate architectures is affordable. Therefore, we simply test all candidate subnets in the growth space " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": " using the parameters inherited from the supernet " + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 297, + 545, + 442 + ], + "type": "text", + "content": " and select the optimal one for the next stage via the composite metric of Equation 7." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "text", + "content": "Phase 3: Selected Model Training To alleviate the influence of supernet training on the grown model, we also use PIM to initialize the parameter " + }, + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "text", + "content": " of selected model " + }, + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\psi^{*}" + }, + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 491, + 545, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 491, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 370, + 491, + 545, + 504 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\omega} = \\beta \\cdot \\boldsymbol {\\omega} _ {s u p e r} ^ {*} + \\gamma \\cdot \\boldsymbol {\\omega} _ {r a n d}. \\tag {8}", + "image_path": "92442c7ea2aeecae44a320b543b3c89fb3b1f196229204dab2dbda07aa7cfeaf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "content": "After the parameter inheriting, we keep training the selected model, to obtain its final optimal parameters:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 543, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 543, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 365, + 543, + 545, + 561 + ], + "type": "interline_equation", + "content": "\\omega^ {*} = \\underset {\\omega} {\\arg \\min } \\mathcal {L} \\left(\\psi^ {*}, \\omega ; \\mathcal {D} ^ {t}\\right). \\tag {9}", + "image_path": "452a35c225f7f515037fd10e0bdcb2b7b1fa8269adcc1ff30f417505a40b981f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 571, + 387, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 387, + 584 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 387, + 584 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "type": "text", + "content": "In this section, we firstly describe the experiment setting and implementation details (Section 4.1). Then we show the results on zero-shot image classification (Section 4.2) and image-text retrieval (Section 4.3). Finally, we conduct ablation study on our GrowCLIP to validate the effectiveness of components (Section 4.4) and analyse (Section 4.5)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 671, + 420, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 420, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 420, + 684 + ], + "type": "text", + "content": "4.1. Experiment Setting" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Benchmark Protocol Conceptual 12M (CC12M) [9] collects about 12 million image-text pairs from the Internet." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22182" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 541, + 392 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 392 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 392 + ], + "type": "table", + "html": "
Growth stepMethodTraining dataCaltech101CIFAR10CIFAR100DTDFlowers102Food101OxfordPedsSUN397ImageNetAverage
ParameterParameterParameterParameterParameterParameterParameterParameterParameterParameter
Step 1CLIP (TWP)3M149.6M42.837.614.5109.316.121.622.014.020.9
CLIP (TFS)[52]3M149.6M42.837.614.5109.316.121.622.014.020.9
MS-CLIP (TFS)[21]3M129M47.135.514.410.012.917.330.124.718.123.3
CLIP (SAP)[4]3M149.6M42.837.614.5109.316.121.622.014.020.9
NAS-S13M30.0M40.531.213.39.08.915.820.223.213.719.5
GrowCLIP-S1 (ours)3M30.0M40.531.213.39.08.915.820.223.213.719.5
Step 2CLIP (TWP)6M149.6M50.336.116.011.312.421.227.125.817.924.2
CLIP (TFS)[52]6M149.6M58.552.022.312.113.125.033.630.120.929.7
MS-CLIP (TFS)[21]6M129M60.650.018.313.417.42742.432.625.231.8
CLIP (SAP)[4]6M149.6M61.949.021.412.415.524.635.431.622.630.5
NAS-S26M96.9M59.441.818.313.315.826.737.333.824.430.1
GrowCLIP-S2 (ours)6M116.6M57.748.021.213.718.029.539.534.525.732.0
Step 3CLIP (TWP)9M149.6M55.240.119.713.115.723.830.928.920.427.5
CLIP (TFS)[52]9M149.6M63.354.624.514.517.529.536.933.825.333.3
MS-CLIP (TFS)[21]9M129M65.048.222.513.920.631.042.635.729.233.0
CLIP (SAP)[4]9M149.6M67.853.926.715.121.232.941.436.028.235.9
NAS-S39M287.4M66.746.822.214.919.735.446.139.330.135.8
GrowCLIP-S3 (ours)9M168.6M69.653.925.818.418.736.346.241.632.638.1
Step 4CLIP (TWP)12M149.6M57.544.420.512.114.224.134.730.322.228.9
CLIP (TFS)[52]12M149.6M65.957.329.716.318.532.945.536.728.336.8
MS-CLIP (TFS)[21]12M129M65.350.926.316.221.935.950.138.432.037.4
CLIP (SAP)[4]12M149.6M69.860.631.912.421.036.747.040.432.239.1
NAS-S412M325.9M70.853.928.316.721.739.551.241.334.239.7
GrowCLIP-S4 (ours)12M188.4M71.960.728.317.323.342.552.445.536.142.0
", + "image_path": "4ce9ff3adeab9b9a16db0b6532f742e1b70f58989e0a2cbd39bab2f6c08c8f2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 399, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 47, + 399, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 545, + 422 + ], + "type": "text", + "content": "Table 3. Top-1 accuracy(%) of zero-shot image classification on 9 datasets. Our GrowCLIP outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at each growth step except step 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 441, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 537 + ], + "type": "text", + "content": "The CC12M is randomly divided into 4 subsets. We select one subset as the training set for the first growth step and adds one subset to the training set at each subsequent growth step such that there are total 4 growth steps in our experiments. At each growth step, we evaluate the corresponding model on the downstream tasks including zero-shot image classification and zero-shot image-text retrieval. The grown architecture of each growth step is shown in Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 541, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 696 + ], + "type": "text", + "content": "Comparison Methods To demonstrate the effectiveness of our method, we compare it with the following baselines: (i) Training with pre-training (TWP) is a usual method in transfer learning. (ii) Training from scratch (TFS) [52] is used as a basic method to train language-image pre-trained model. (iii) MS-CLIP [21] is a handcrafted framework with a shareable module and modality-specific auxiliary module (the early specialization). (iv) The shrink and perturb (SAP) [4] is a method of parameter inheriting in warm-starting learning. (v) Neural architecture search (NAS) [49, 10] is an important research field of auto machine learning, in which we use the same search space as the growth space of GrowCLIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Implementation Details A dual-stream model follow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "text", + "content": "ing CLIP [52] without shared encoder is used as basic architecture at the first growth step. The image encoder is stacked with 6 transformer blocks with the following hyperparameters: number of self-attention heads is 6 and the hidden layer size is 384. As for the text encoder, 6 transformer blocks are stacked and the number of the attention heads and the width are 4 and 256 respectively. The resolution of input in the images encoder is resized to 224 and the maximum sequence in the text encoder is limited to 77. The center crop is the only data augmentation method for training. We use the prompts like " + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^1" + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "text", + "content": " on the zero-shot image classification. We adopt LAMB optimizer [66] with learning rate 0.01 and weight decay 0.0001, and cosine learning rate schedule [45] with a 4000-iteration linear warmup. For the hyperparameters, the proportion of the old model and the new one in the parameter inheriting " + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\beta = 0.3" + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\gamma = 0.001" + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "text", + "content": ". The model parameter coefficient in growth architecture selection " + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 304, + 441, + 545, + 681 + ], + "type": "text", + "content": ". We train the model with a total batch size of 1536 for 30 epochs at each growth step, which includes 2 epochs for supernet fine-tuning, 2 epochs for su" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 543, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 543, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 543, + 713 + ], + "type": "text", + "content": "1https://github.com/openai/CLIP/blob/main/data/prompts.md" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22183" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 72, + 553, + 360 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 553, + 360 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 553, + 360 + ], + "type": "table", + "html": "
Growth stepMethodFlickr30KMSCOCO
image-to-texttext-to-imageimage-to-texttext-to-image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
Step 1CLIP (TWP)19.645.356.414.333.544.110.726.336.87.019.127.5
CLIP (TFS) [52]19.645.356.414.333.544.110.726.336.87.019.127.5
MS-CLIP (TFS)[21]27.451.863.318.139.250.712.931.943.28.823.533.1
CLIP (SAP) [4]19.645.356.414.333.544.110.726.336.87.019.127.5
NAS-S122.845.257.813.534.246.010.627.437.57.019.928.8
GrowCLIP-S1 (ours)22.845.257.813.534.246.010.627.437.57.019.928.8
Step 2CLIP (TWP)26.651.062.017.439.850.712.830.540.88.322.531.9
CLIP (TFS) [52]31.457.768.121.246.757.117.137.448.710.927.337.4
MS-CLIP (TFS)[21]35.865.175.025.853.164.619.343.055.213.131.542.6
CLIP (SAP) [4]33.660.972.024.148.259.317.138.050.210.928.028.1
NAS-S236.963.074.025.251.762.318.441.653.712.930.941.7
GrowCLIP-S2 (ours)37.563.774.225.951.162.819.142.854.712.930.841.9
Step 3CLIP (TWP)28.354.466.620.642.254.714.434.145.09.625.234.9
CLIP (TFS) [52]37.563.574.226.052.164.019.743.255.113.331.842.6
MS-CLIP (TFS)[21]40.067.778.229.857.068.822.947.959.115.335.246.6
CLIP (SAP) [4]40.768.078.328.155.367.122.845.758.814.233.444.6
NAS-S343.272.880.732.159.970.223.448.361.015.736.848.3
GrowCLIP-S3 (ours)44.673.180.733.260.770.923.548.861.015.836.948.4
Step 4CLIP (TWP)30.256.366.321.746.657.615.636.247.810.426.536.9
CLIP (TFS) [52]41.568.680.030.056.968.021.845.557.314.333.845.1
MS-CLIP (TFS)[21]44.372.881.132.661.372.125.250.562.016.637.749.0
CLIP (SAP) [4]46.974.783.733.761.372.125.250.862.816.637.449.1
NAS-S449.675.283.337.065.275.426.853.064.617.739.651.5
GrowCLIP-S4 (ours)50.875.884.037.265.376.227.653.264.918.140.352.0
", + "image_path": "1a8dbc25c08c355e4e9d0d0493d7763600afc31f7b1ea7b53f4f993bf4d3e08d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 367, + 503, + 379 + ], + "lines": [ + { + "bbox": [ + 89, + 367, + 503, + 379 + ], + "spans": [ + { + "bbox": [ + 89, + 367, + 503, + 379 + ], + "type": "text", + "content": "Table 4. Results of zero-shot image-text retrieval on Flickr30K and MSCOCO datasets. R@K means top-K recall." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 517 + ], + "type": "text", + "content": "pernet training and 26 epochs for grown model training. For the baselines, the CLIP adopts the same usual architecture (ViT-B/16), which has 12-layer 768-wide transformer blocks with 12 attention heads for image encoder and 12-layer 512-wide transformer blocks with 8 attention heads for text encoder and are trained for 30 epochs. For NAS, we first train the supernet for 30 epochs to select the best architecture, and then train the selected model from scratch for another 30 epochs. All experiments are performed on 16 NVIDIA Tesla V100 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 212, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 212, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 212, + 548 + ], + "type": "text", + "content": "4.2. Zero-shot Image Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": "At each growth step, we evaluate the model on zero-shot image classification task. The downstream datasets include 9 typical settings, Caltech101 [35], CIFAR10 [32], CIFAR100 [32], Describable Textures Dataset (DTD) [12], Flowers102 [47], Food101 [6], OxfordPets [48], SUN39 [5] and ImageNet [14]. We apply prompt ensemble evaluation like CLIP for each dataset. Results are given in Table 3. Our model performs growth process at each growth step except step 1, where the parameter grows from " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "30.0M" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "188.4M" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": ". It outperforms the baseline methods in terms of average top-1 accuracy over 9 datasets at the last three growth steps. For example, at the third growth step, our model improves the average top-1 accuracy from " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "35.9\\%" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "38.1\\%" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "inline_equation", + "content": "(+2.2\\%)" + }, + { + "bbox": [ + 46, + 558, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 398, + 476, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 476, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 476, + 411 + ], + "type": "text", + "content": "4.3. Zero-shot Image-Text Retrieval" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "text", + "content": "The zero-shot image-text retrieval consists of two subtasks: image-to-text retrieval and text-to-image retrieval. We evaluate the model of each growth step on two retrieval benchmark datasets: Flickr30K [50] and MSCOCO [40]. As shown in Table 4, our model gets better as model grows and data grows. At the last two growth steps, it can be observed that our method has great performance on all metrics. Particularly, our method improves from " + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "inline_equation", + "content": "49.6\\%" + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "inline_equation", + "content": "50.8\\% (+1.2\\%)" + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "text", + "content": " for top-1 image-to-text recall on Flickr30K dataset and from " + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "inline_equation", + "content": "26.8\\%" + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "inline_equation", + "content": "27.6\\% (+0.8\\%)" + }, + { + "bbox": [ + 304, + 418, + 545, + 550 + ], + "type": "text", + "content": " on MSCOCO at the last growth step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 562, + 400, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 400, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 400, + 574 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "type": "text", + "content": "Here, we conduct the ablation study on our GrowCLIP. All experiments are built at the growth step 2 and we evaluate the models on zero-shot image classification of ImageNet. To verify the effectiveness of each module in our proposed GrowCLIP, we conduct ablation experiments at some core steps mentioned in Section 3.4 as follows. (i) Parameter inheriting with momentum in phase 1 (PIM@1). (ii) Supernet fine-tuning in phase 1 (ST). (iii) Growth architecture selection in phase 2 (GAS). (iv) Parameter inheriting with momentum in phase 3 (PIM@3). The accuracy on zero-shot image classification of ImageNet is re" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22184" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 70, + 271, + 155 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 271, + 155 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 271, + 155 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 271, + 155 + ], + "type": "table", + "html": "
PIM@1STGASPIM@3Para.Acc.
✓✓✓✓116.6M25.7
✓✓✓X116.6M24.5
✓✓XX109.5M22.6
✓XXX109.5M21.1
XXXX109.5M20.8
", + "image_path": "d18fa5e133fd627a64fd719114babb9cb41e01ef04456c730d574cda77f6705d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 205, + 282, + 278 + ], + "blocks": [ + { + "bbox": [ + 47, + 163, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 196 + ], + "type": "text", + "content": "Table 5. Ablation analysis of each module in GrowCLIP at Growth Step 2. (Para.: parameter of the model. Acc.: top-1 accuracy (\\%) of zero-shot image classification on ImageNet.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 205, + 282, + 278 + ], + "lines": [ + { + "bbox": [ + 52, + 205, + 282, + 278 + ], + "spans": [ + { + "bbox": [ + 52, + 205, + 282, + 278 + ], + "type": "table", + "html": "
Para.Acc.
CLIP-ViT-B/16 (TFS)149.6M28.3
CLIP-ViT-L/16 (TFS)431.1M29.9
GrowCLIP-S4 (TFS)188.4M33.3
GrowCLIP-S4 (ours)188.4M36.1
", + "image_path": "89b7ccf15eb81ae9dc0b6ffcaec912bdbe857cbe1e3016868caf6a7292d02c55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "text", + "content": "ported for comparison and the results of the ablation study are shown in Table 5. As can be seen, all above four components have positive effects on the final results. Specifically, the supernet training will influence the grown model without PIM@3, which results in " + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "text", + "content": " drop. Compared with GAS method, we randomly choose an architecture in the growth space, which has a worse performance from " + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "inline_equation", + "content": "24.5\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "inline_equation", + "content": "22.6\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "text", + "content": ". With ST, we observe a remarkable improvement of " + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "text", + "content": ". GrowCLIP with PIM@1 solves the problem of local minimum dilemma and brings a " + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 316, + 287, + 437 + ], + "type": "text", + "content": " improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 444, + 110, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 110, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 110, + 456 + ], + "type": "text", + "content": "4.5. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 462, + 277, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 462, + 277, + 474 + ], + "spans": [ + { + "bbox": [ + 59, + 462, + 277, + 474 + ], + "type": "text", + "content": "Why can GrowCLIP improve the performance?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 653 + ], + "type": "text", + "content": "Compared with baseline models, the gains of our GrowCLIP come from two folds: (i) GrowCLIP can search the architecture that matches the current data with Growth architecture selection (GAS). (ii) Parameter inheriting with momentum (PIM) benefits GrowCLIP from yesterday's model. To demonstrate their effectiveness, we train the architecture selected by GrowCLIP at step 4 from scratch, and the result is shown in Table 6. Compared with ViT-B/16, the growth architecture selection can bring " + }, + { + "bbox": [ + 46, + 474, + 287, + 653 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 474, + 287, + 653 + ], + "type": "text", + "content": " improvement on zero-shot image classification of ImageNet. Furthermore, it is even better than ViT-L/16, which is two and a half times as big as our model. More experiments have proved the effectiveness of GAS in the appendix. Comparing our GrowCLIP-S4 with GrowCLIP-S4 (TFS), we observe a remarkable improvement of " + }, + { + "bbox": [ + 46, + 474, + 287, + 653 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 46, + 474, + 287, + 653 + ], + "type": "text", + "content": " with PIM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 654, + 286, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 654, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 59, + 654, + 286, + 666 + ], + "type": "text", + "content": "Why can GrowCLIP improve the training efficiency?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Our GrowCLIP can improve training efficiency, due to GAS and PIM, and the result is shown in Table 7. (i) Compared with the fixed architecture method, GrowCLIP has a smaller model with the GAS, so we reduce training costs" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 543, + 177 + ], + "blocks": [ + { + "bbox": [ + 84, + 286, + 247, + 297 + ], + "lines": [ + { + "bbox": [ + 84, + 286, + 247, + 297 + ], + "spans": [ + { + "bbox": [ + 84, + 286, + 247, + 297 + ], + "type": "text", + "content": "Table 6. Ablation analysis of GrowCLIP-S4." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 543, + 177 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 543, + 177 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 543, + 177 + ], + "type": "table", + "html": "
Para.Acc.EpochCost
CLIP-S2(TFS)149.6M20.930336
GrowCLIP-S2 (ours)116.6M25.730288
GrowCLIP-S4 (TFS)188.4M33.3301230
GrowCLIP-S4 (ours)188.4M33.39722
NAS-S4325.9M34.2605280
GrowCLIP-S4 (ours)188.4M36.1301583
", + "image_path": "028dfca84c8aa208e45199e92aca4aef48988266632af37820f259d0e523dddc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 206, + 542, + 255 + ], + "blocks": [ + { + "bbox": [ + 338, + 185, + 513, + 196 + ], + "lines": [ + { + "bbox": [ + 338, + 185, + 513, + 196 + ], + "spans": [ + { + "bbox": [ + 338, + 185, + 513, + 196 + ], + "type": "text", + "content": "Table 7. Efficiency analysis. (Cost: GPU hours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 206, + 542, + 255 + ], + "lines": [ + { + "bbox": [ + 310, + 206, + 542, + 255 + ], + "spans": [ + { + "bbox": [ + 310, + 206, + 542, + 255 + ], + "type": "table", + "html": "
Para.Acc.
GrowCLIP-S1431.1M14.6
GrowCLIP-S2431.2M26.2
", + "image_path": "37dbeed2f385563ca656283f6f47b267d017df95916fede62539f1afb81f8130.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 263, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 306, + 263, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 545, + 285 + ], + "type": "text", + "content": "Table 8. The architecture of GrowCLIP-S1 with bigger initialized model (ViT-L/16)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 306, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 449 + ], + "type": "text", + "content": "at the early steps (48 GPU hours saved at step 2). At the later steps, GrowCLIP uses fewer epochs to achieve good performance because of PIM. As shown in Figure 1, GrowCLIP has far outperformed CLIP (TFS) only after supernet training. To demonstrate the effectiveness of PIM, we compare GrowCLIP-S4 with GrowCLIP-S4 (TFS). GrowCLIP only uses 9 epochs to achieve comparable performance with PIM. (ii) Compared with the NAS method, GrowCLIP inherits the parameters of the old model with momentum, which only takes 2 epochs to fine-tune the supernet and 2 epochs to train the supernet. In contrast, one-shot NAS methods consume about 3.3X resources." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 450, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 594 + ], + "type": "text", + "content": "Can GrowCLIP find the relationship between model and data? When the model scale is sufficient large, there is no need to enlarge the model. We conduct the experiment started from ViT-L/16. And the result in Table 8 shows that the larger model startpoint hardly grows at step 2, which means that our GrowCLIP can find the relationship between the model and data. In our online learning scenario, the initial architecture of the model is not important due to our growth pipeline. If the model scale is smaller at the beginning, GrowCLIP will select a larger enough model for growth. In contrast, if the model scale is too large initially, GrowCLIP will stop growing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "What is the influence when growth happens in different dimensions? As shown in Table 9, different growth dimensions have different influences on the performance. By analyzing the experimental results, we conclude 4 laws, which can provide a reference for the design of cross-modal model architecture: (i) Enlarging the smaller model bring a greater improvement than the bigger one. For example, when the growth appears in transformer blocks in the image encoder, the accuracy can gain " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "0.93\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " improvement at step 2 but only " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "0.12\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " at step 4. (ii) We need to leave the image" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22185" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 288, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 288, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 288, + 186 + ], + "type": "table", + "html": "
Step 2Step 3Step 4
DimensionPara.Acc.Para.Acc.Para.Acc.
+4 bI+7.1M+0.93+19.7M+0.46+19.7M+0.12
+4 hI+20.1M+0.34+67.3M+0.12+105.1M+0.03
+2 lI+0.1M+0.84+0.1M+0.07+0.1M+0.10
+4 bT+3.2M+0.01+12.6M+0.08+12.6M+0.01
+4 hT+27.0M+0.22+45.0M+0.01+69.5M-0.04
+4 bS+7.1M+1.03+19.7M+0.19+19.7M-0.07
", + "image_path": "5691aa04d4e705a3b23dd00643a0e579d846fe8a0063400ed3ac04c164c876a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 226 + ], + "type": "text", + "content": "Table 9. The ablation of growth space corresponding to different dimensions. " + }, + { + "bbox": [ + 46, + 193, + 287, + 226 + ], + "type": "inline_equation", + "content": "(+4b^{I}" + }, + { + "bbox": [ + 46, + 193, + 287, + 226 + ], + "type": "text", + "content": " : the number of transformer blocks in image encoder is increased by 4)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 247, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 366 + ], + "type": "text", + "content": "encoder with more capacity. Compared with text encoder, the improvement of growing the image encoder is more significant at each step. (iii) The convolutional layers is a very important component in the image encoder, which can help the model to better extract visual feature. (iv) The shared encoder can enhance the degree of cross-modal fusion and bring a remarkable improvement. However, do not use a large number of transformers in the shared encoder, or the model become redundant with no noticeable performance improvement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 379, + 119, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 119, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 119, + 391 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 399, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 589 + ], + "type": "text", + "content": "In this paper, we study the online learning case in cross-modal pre-trained models. The small-scale architecture may limit the final performance due to the limited parameters when data grows. Based on this real-world scenario, we propose a data-driven automatic model growing algorithm for language-image pre-training - GrowCLIP, which upgrades the model size when abundant new data arrives. We adopt parameter inheriting with momentum (PIM) to utilize the knowledge in yesterday's model and seek out the optimal growth architecture at each growth step with growth architecture selection (GAS). GrowCLIP improves the performance on downstream tasks, including zero-shot image classification and retrieval. The effectiveness of each component of GrowCLIP is proved on ablation studies. And we provide a reference for the design of cross-modal model architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "content": "Limitations In this work, we only prove the effectiveness of GrowCLIP with public pre-training dataset - CC12M. In future works, we will extend our method to a real-world scenario, where the VLP model keeps training with constantly updated data crawled from the web." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 663, + 140, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 140, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 140, + 676 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "This work was supported in part by National Key R&D Program of China under Grant No.2020AAA0109700, Guangdong Outstanding Youth Fund (Grant No." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "2021B1515020061), Shenzhen Science and Technology Program (Grant No.RCYX20200714114642083), Shenzhen Fundamental Research Program(Grant No. JCYJ20190807154211365), Nansha Key RD Program under Grant No.2022ZD014 and Sun Yat-sen University under Grant No.22lgqb38 and 76160-12220011. We thank MindSpore for the partial support of this work, which is a new deep learning computing framework" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 410, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 410, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 410, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 317, + 702, + 410, + 713 + ], + "type": "text", + "content": "https://www.mindspore.cn/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22186" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "text", + "content": "[2] R. Aljundi, F. Babiloni, M. Elhoseiny, M. Rohrbach, and T. Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European conference on computer vision, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "type": "text", + "content": "[3] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Zitnick, and D. Parikh. Vqa: Visual question answering. In International Conference on Computer Vision, 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 225 + ], + "type": "text", + "content": "[4] J. Ash and R. Adams. On warm-starting neural network training. 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 248 + ], + "type": "text", + "content": "[5] A. Barriuso and A. Torralba. Notes on image annotation. Preprint arXiv:1210.3448, 2012." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "type": "text", + "content": "[6] L. Bossard, M. Guillaumin, and L. Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, 2014." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "text", + "content": "[7] P. Buzzega, M. Boschini, A. Porrello, D. Abati, and S. Calderara. Dark experience for general continual learning: a strong, simple baseline. In Advances in Neural Information Processing Systems, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 330, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 372 + ], + "type": "text", + "content": "[8] H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "text", + "content": "[9] S. Changpinyo, P. Sharma, N. Ding, and R. Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 420, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 452 + ], + "type": "text", + "content": "[10] M. Chen, H. Peng, J. Fu, and H. Ling. Autoformer: Searching transformers for visual recognition. In International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 453, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 486 + ], + "type": "text", + "content": "[11] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "text", + "content": "[12] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Computer Vision and Pattern Recognition, 2014." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "text", + "content": "[13] M. Cornia, L. Baraldi, and R. Cucchiara. Show, control and tell: A framework for generating controllable and grounded captions. In Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 555, + 277, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 277, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 277, + 567 + ], + "type": "text", + "content": "[14] J. Deng. A large-scale hierarchical image database. 2009." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[15] J. Devlin, M. Chang, K. Lee, and K. Toutanova. BERT: Pretraining of Deep Bidirectional Transformers for Language Understanding. In North American Chapter of the Association for Computational Linguistics, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[16] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 669, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 689 + ], + "type": "text", + "content": "[17] H. Grabner and H. Bischof. On-line boosting and vision. In Computer Vision and Pattern Recognition, 2006." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[18] J. Grill, F. Strub, F. Altché, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Ghesh" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[19] Z. Guo, B. Avila Pires, B. Piot, J. Grill, F. Altché, R. Munos, and M. Azar. Bootstrap latent-predictive representations for multitask reinforcement learning. In International Conference on Machine Learning, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[20] Z. Guo, X. Zhang, H. Mu, W. Heng, Z. Liu, Y. Wei, and J. Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 184, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 236 + ], + "type": "text", + "content": "[21] L Zhou B Xiao C. Noel Y Cheng R Xu S Chang H, You and Yuan L. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision, pages 69-87. Springer, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[22] R. Hadsell, D. Rao, A. Rusu, and R. Pascanu. Embracing change: Continual learning in deep neural networks. Trends in Cognitive Sciences, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 273, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 293 + ], + "type": "text", + "content": "[23] E. Hazan et al. Introduction to online convex optimization. Foundations and Trends® in Optimization, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 295, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 327 + ], + "type": "text", + "content": "[24] K. He, X. Chen, S. Xie, Y. Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 328, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 328, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 328, + 545, + 360 + ], + "type": "text", + "content": "[25] K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In Computer Vision and Pattern Recognition, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "text", + "content": "[26] X. Hu, Z. Gan, J. Wang, Z. Yang, Z. Liu, Y. Lu, and L. Wang. Scaling up vision-language pre-training for image captioning. Preprint arXiv:2111.12233, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "text", + "content": "[27] X. Hu, X. Yin, K. Lin, L. Wang, L. Zhang, J. Gao, and Z. Liu. Vivo: Surpassing human performance in novel object captioning with visual vocabulary pre-training. 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "text", + "content": "[28] C. Jia, Y. Yang, Y. Xia, Y. Chen, Z. Parekh, H. Pham, Q. Le, Y. Sung, Z. Li, and T. Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "text", + "content": "[29] J. Kaplan, S. McCandlish, T. Henighan, T. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei. Scaling laws for neural language models. Preprint arXiv:2001.08361, 2020." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "text", + "content": "[30] W. Kim, B. Son, and I. Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, 2021." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 549, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 549, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 549, + 545, + 601 + ], + "type": "text", + "content": "[31] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. National Academy of Sciences, 2017." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "type": "text", + "content": "[32] A. Krizhevsky, G. Hinton, et al. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 626, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 626, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 626, + 545, + 669 + ], + "type": "text", + "content": "[33] Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, and R. Soricut. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[34] C Li, B. Zhuang, G. Wang, X. Liang, X. Chang, and Y. Yang. Automated progressive learning for efficient training of vision transformers. In Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22187" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[35] F. Li, R. Fergus, and P. Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 2006." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[36] J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "text", + "content": "[37] L. Li, M. Yatskar, D. Yin, C. Hsieh, and K. Chang. Visualbert: A simple and performant baseline for vision and language. Preprint arXiv:1908.03557, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "text", + "content": "[38] X. Li, X. Yin, C. Li, P. Zhang, X. Hu, L. Zhang, L. Wang, H. Hu, L. Dong, F. Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 277 + ], + "type": "text", + "content": "[39] Y. Li, F. Liang, L. Zhao, Y. Cui, W. Ouyang, J. Shao, F. Yu, and J. Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. Preprint arXiv:2110.05208, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 322 + ], + "type": "text", + "content": "[40] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 324, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 287, + 368 + ], + "type": "text", + "content": "[41] C. Liu, B. Zoph, M. Neumann, J. Shlens, W. Hua, L. Li, F. Li, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. In European conference on computer vision, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 403 + ], + "type": "text", + "content": "[42] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. Preprint arXiv:1806.09055, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "text", + "content": "[43] W. Liu, S. Chen, L. Guo, X. Zhu, and J. Liu. Cptr: Full transformer network for image captioning. Preprint arXiv:2101.10804, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "text", + "content": "[44] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. Preprint arXiv:1907.11692, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 485, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 506 + ], + "type": "text", + "content": "[45] I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "type": "text", + "content": "[46] J. Lu, V. Goswami, M. Rohrbach, D. Parikh, and S. Lee. 12-in-1: Multi-task vision and language representation learning. In Computer Vision and Pattern Recognition, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "text", + "content": "[47] M. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Computer Vision, Graphics & Image Processing, 2008." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[48] O. Parkhi, A. Vedaldi, A. Zisserman, and C. Jawahar. Cats and dogs. In Computer Vision and Pattern Recognition, 2012." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[49] H. Peng, H. Du, H. Yu, Q. Li, J. Liao, and J. Fu. Cream of the crop: Distilling prioritized paths for one-shot neural architecture search. In Advances in Neural Information Processing Systems, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[50] B. Plummer, L. Wang, C. Cervantes, J. Caicedo, J. Hockenmaier, and S. Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-tosentence models. In International Conference on Computer Vision, 2015." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[51] R. Pogodin and T. Lattimore. On first-order bounds, variance and gap-dependent bounds for adversarial bandits. In Uncertainty in Artificial Intelligence, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[52] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[53] E. Real, A. Aggarwal, Y. Huang, and Q. Le. Regularized Evolution for Image Classifier Architecture Search. In AAAI Conference on Artificial Intelligence, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[54] E. Real, S. Moore, A. Selle, S. Saxena, Y. Suematsu, J. Tan, Q. Le, and A. Kurakin. Large-scale evolution of image classifiers. In International Conference on Machine Learning, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[55] H Xu J. Han T. Tuytelaars Z. Li S. Yan, L. Hong and X. He. Generative negative text replay for continual vision-language pretraining. In European Conference on Computer Vision, pages 22-38. Springer, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[56] A. Saffari, C. Leistner, J. Santner, M. Godec, and H. Bischof. On-line random forests. In International Conference on Computer Vision workshops, 2009." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "text", + "content": "[57] S. Shalev-Shwartz et al. Online learning and online convex optimization. Foundations and Trends® in Machine Learning, 2012." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[58] S. Song and B. Choi. A study on continuous follow-the-leader (ftl) gaits: an effective walking algorithm over rough terrain. Mathematical biosciences, 1989." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "text", + "content": "[59] T Srinivasan, T Chang, L. Alva, G Chochlakis, M Rostami, and J Thomason. Climb: A continual learning benchmark for vision-and-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[60] M. Tan, B. Chen, R. Pang, V. Vasudevan, M. Sandler, A. Howard, and Q. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "text", + "content": "[61] B. Thomee, D. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland, D. Borth, and L. Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[62] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. Gomez, L. Kaiser, and I. Polosukhin. Attention Is All You Need. In Neural Information Processing Systems, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "text", + "content": "[63] Haonan Wang, Minbin Huang, Runhui Huang, Lanqing Hong, Hang Xu, Tianyang Hu, Xiaodan Liang, and Zhenguo Li. Boosting visual-language models by exploiting hard samples. arXiv preprint arXiv:2305.05208, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[64] S. Xie, H. Zheng, C. Liu, and L. Lin. SNAS: Stochastic Neural Architecture Search. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[65] L. Yao, R. Huang, L. Hou, G. Lu, M. Niu, H. Xu, X. Liang, Z. Li, X. Jiang, and C. Xu. Filip: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[66] Y. You, J. Li, S. Reddi, J. Hseu, S. Kumar, S. Bhojanapalli, X. Song, J. Demmel, K. Keutzer, and C. Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22188" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 263 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[67] F. Yu, J. Tang, W. Yin, Y. Sun, H. Tian, H. Wu, and H. Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3208-3216, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "text", + "content": "[68] J. Yu, Z. Wang, V. Vasudevan, L. Yeung, M. Seyedhosseini, and Y. Wu. Coca: Contrastive captioners are image-text foundation models. Preprint arXiv:2205.01917, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "type": "text", + "content": "[69] X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104-12113, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 286, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 286, + 263 + ], + "type": "text", + "content": "[70] B. Zoph, V. Vasudevan, J. Shlens, and Q. Le. Learning Transferable Architectures for Scalable Image Recognition. In Computer Vision and Pattern Recognition, 2018." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22189" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_content_list.json b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d78673dff208dabe485d07242ba8f01b8f06b5 --- /dev/null +++ b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_content_list.json @@ -0,0 +1,1475 @@ +[ + { + "type": "text", + "text": "Growing a Brain with Sparsity-Inducing Generation for Continual Learning", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyundong Jin1 Gyeong-hyeon Kim1 Chanho Ahn2* Eunwoo Kim1 \n1School of Computer Science and Engineering, Chung-Ang University \n2Samsung Advanced Institute of Technology (SAIT)", + "bbox": [ + 173, + 179, + 795, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jude0316,leonardkkh,eunwoo}@cau.ac.kr chanho.ahn@samsung.com", + "bbox": [ + 192, + 236, + 777, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks suffer from catastrophic forgetting in continual learning, where they tend to lose information about previously learned tasks when optimizing a new incoming task. Recent strategies isolate the important parameters for previous tasks to retain old knowledge while learning the new task. However, using the fixed old knowledge might act as an obstacle to capturing novel representations. To overcome this limitation, we propose a framework that evolves the previously allocated parameters by absorbing the knowledge of the new task. The approach performs under two different networks. The base network learns knowledge of sequential tasks, and the sparsity-inducing hypernetwork generates parameters for each time step for evolving old knowledge. The generated parameters transform old parameters of the base network to reflect the new knowledge. We design the hypernetwork to generate sparse parameters conditional to the task-specific information and the structural information of the base network. We evaluate the proposed approach on class-incremental and task-incremental learning scenarios for image classification and video action recognition tasks. Experimental results show that the proposed method consistently outperforms a large variety of continual learning approaches for those scenarios by evolving old knowledge.", + "bbox": [ + 76, + 319, + 472, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 712, + 209, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability of deep learning to incrementally accumulate knowledge for sequentially incoming tasks is an important capability required for modern AI algorithms. However, deep neural networks tend to lose previously acquired knowledge while learning new data. This limitation of deep neural networks, called catastrophic forgetting, has been investigated by previous studies [29, 17]. Catastrophic forgetting significantly degrades performance when data from previous tasks (task-incremental) or already-learned class", + "bbox": [ + 76, + 738, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data (class-incremental) cannot be accessed. Under those incremental learning scenarios, continual learning methods have toiled to preserve the previous knowledge.", + "bbox": [ + 500, + 287, + 890, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various approaches have been proposed for continual learning and are categorized into three types: regularization [2, 4, 20, 17, 48], data rehearsal [5, 26, 32, 35, 38, 47], and parameter isolation [1, 30, 28]. The conventional continual learning methods introduced regularization approaches to suppress changes in important parameters [48, 2] or input-output mappings from a network trained up to previous tasks [22, 7]. The network is trained in a way that contributable parameters to the performance of the previously trained model remain unchanged. On the other hand, data rehearsal methods [25, 35] address catastrophic forgetting by accessing old data. The old data can be obtained by storing some data from old tasks [35, 47] or leveraging a generative network [38]. However, the privacy of data and the difficulty of learning generative models limit their versatility. Recently, parameter isolation approaches [28, 30, 15] construct a disjoint set of parameters for each task, preventing it from mixing old and new knowledge. These methods learn a new set of parameters by associating all [28] or some helpful old knowledge [15].", + "bbox": [ + 496, + 335, + 892, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike other approaches, the parameter isolation approach is designed to prevent forgetting without requiring access to old data by disallowing updates to previous sets of parameters. Despite this advantage, these methods can limit the ability to learn new tasks by relying on parameters that were learned for older tasks and remain unchanged. We observe that this limitation arises when there is a low correlation between the old and new data. This raises the question: In parameter isolation-based continual learning, is it possible to transform the knowledge from old tasks to fit newer tasks better?", + "bbox": [ + 496, + 641, + 893, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel method for transforming prior knowledge into a form that can aid in learning new tasks. The approach involves reframing the parameter selection problem, which previous works [15] have addressed, as a generation problem. The proposed framework emphasizes a subset of pre-trained knowledge, selecting and highlight-", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done independently, without any support from SAIT.", + "bbox": [ + 94, + 886, + 460, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18961", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing it for use in learning new tasks. Specifically, the framework consists of two networks: base network and hypernetwork. The base network constructs disjoint parameter sets for sequential tasks. The proposed transformer-based [42] hypernetwork generates a small set of influential parameters that significantly impact the loss. The impact of each element of generated parameters is evaluated by the difference in loss computed with and without its inclusion in the learning process. Also, generated parameters are conditioned on the information of the current task (task token) and the layer of the base network (layer embedding). The generated parameters transform the previous set of parameters in the base network to further accommodate new knowledge. In addition, compared to existing hypernetworks that generate a small set of parameters and apply it to all the layers [13, 44], the proposed hypernetwork adaptively generates parameters according to the layer configurations defined in the base network. Through this strategy, the quantity of layer embeddings for ResNet-50 is reduced by a substantial factor of 24. Furthermore, by leveraging a matrix decomposition technique to represent generated parameters compactly, we achieve a noteworthy reduction in the parameters necessary for the hypernetwork.", + "bbox": [ + 76, + 90, + 472, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of the proposed method in both class-incremental and task-incremental learning scenarios using classification benchmark datasets. Furthermore, we demonstrate the efficacy of the proposed method [28, 15] in these scenarios by applying it to video action recognition datasets [19]. Experimental results show that the proposed method performs better than its competitors for most learning scenarios while introducing minor memory overhead from the hypernetwork. The contributions of our work are mainly three-fold: (1) We present a novel approach to evolving previous knowledge in continual learning, aimed at reflecting the knowledge of new tasks. (2) We propose a transformer-based hypernetwork that generates customized parameters for new tasks while suppressing the generation of redundant parameters that have little impact on the loss. (3) Experimental results on diverse benchmarks for video action recognition and image classification demonstrate that ours outperforms its competitors with lower network capacity requirements to perform tasks.", + "bbox": [ + 76, + 439, + 472, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 739, + 218, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Learning. Generally, continual learning methods can be classified into three types. Regularization methods [23, 7, 17, 48, 4, 2, 20] add a penalty term that restrains the previously learned knowledge from being updated when learning a new task. In order not to hurt previous knowledge, a new network is introduced to mimic the features obtained from the old network [22, 7]. While other regularization techniques, such as those based on the Fisher information matrix [17], change of loss [48], and variation of", + "bbox": [ + 76, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2d8c0b91120742ed46a2160189708b9340e73b523ac00cd4bc347e6ca22661e5.jpg", + "image_caption": [ + "Figure 1. An illustration of the proposed method (right) and baseline methods (middle, left) for the third (new) task. The figure represents task-specific allocated parameters, represented by task numbers. Gray numbers represent the sets of discarded parameters from a selection method. The baseline methods do not update old knowledge (red boxes) while updating a new task (blue boxes). The proposed method transforms the old sets of parameters, represented by $+$ , to fit the new task. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 880, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prediction [2], suppress the update of crucial parameters for each task, they experience significant performance decline as the number of tasks grows [12]. Data rehearsal methods [5, 26, 35, 33, 25, 38] rely on the availability of previous task data, which is achieved through selective storage of data [35, 47, 25]. These methods employ a few previous samples jointly with a new dataset [25, 38] or constraining the gradient based on the angle between that obtained from stored samples and the gradient obtained from the new samples [5, 26]. These methods incur a data imbalance problem due to a small number of stored previous tasks compared to a large amount of new task data. Parameter isolation methods [28, 1, 30, 15, 16] involve constructing disjoint sets of parameters for old tasks, respectively, and updating another set of parameters for a newer task by associating it with old parameters during training. This is usually done by using all previous sets of parameters [28], using a learnable mask to select some previous set of parameters [27], jointly selecting a location to allocate a new set of parameters [16], or searching a subnetwork with coarse-to-fine task association [15].", + "bbox": [ + 496, + 353, + 893, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, this approach can impede efficient learning because the parameters optimized for old tasks remain the same when learning newer tasks, meaning they need to absorb the information from the newer tasks. In contrast, the proposed method allows the old knowledge to grow by incorporating the knowledge from new tasks. The proposed method and its baselines are illustrated in Figure 1.", + "bbox": [ + 496, + 672, + 893, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hypernetworks. Hypernetworks generate parameters for a target network [13, 44, 39, 21, 34]. A recent work [34] proposes a method for generating parameters for layers with few parameters, such as activation of ResNet blocks or batch normalization layers. However, generating parameters for convolutional and linear layers that take most of the parameters in the base network can be challenging because it is difficult to represent the entire set of parameters from a", + "bbox": [ + 496, + 779, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18962", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c5d4fe3df336ce739587dcec0b7ab5dd0688a6e13c1fa1176f88f1d65470b0fd.jpg", + "image_caption": [ + "Figure 2. The proposed method for transforming the parameters of the $l$ -th layer of the base network $F(\\cdot)$ when learning the $t$ -th task. The numbers in the grid indicate the indices of tasks. The hypernetwork $H(\\cdot)$ generates parameters using the task-layer attention block (TLA), which discovers the relationship between the task token and layer embedding. The sparse parameters are generated while suppressing the generation of the redundant ones by evaluating the impact of parameters on the loss. The previous sets of parameters of the $l$ -th layer are transformed by element-wise multiplication with $G_{l}^{t}$ . The transformed parameters are combined with the current parameter set through the union operation $(\\cup)$ and released as a complete set of parameters to learn new knowledge. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 870, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hypernetwork. To avoid this, existing works generate small sets of parameters and concatenates them [13, 21]. Consequently, to generate a base network of large scale, allowing a large number of input embeddings corresponding to the network is inevitable due to the repetitive generation mechanism. In contrast to the approach, the parameters are generated according to the dimensions defined in the layers of the base network in our method. Note that as these dimensions expand, the corresponding demands on hypernetwork parameters also escalate. To address this, we employ a matrix decomposition technique for a compact representation of the required parameters in the hypernetwork.", + "bbox": [ + 75, + 446, + 472, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, in the continual learning literature, [44] and [10] learn a hypernetwork to generate parameters with task-specific embeddings for new tasks while preventing changes to previously generated parameters for previous tasks. The approaches produce dense parameters for the base network through a task-conditional hypernetwork. In contrast, the proposed transformer-based hypernetwork generates sparse parameters by measuring the effect of individual parameters on the loss. Furthermore, our method generates parameters to evolve previous knowledge to reflect new knowledge, whereas the aforementioned approaches generate parameters directly for the base network.", + "bbox": [ + 75, + 628, + 472, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. GrowBrain", + "text_level": 1, + "bbox": [ + 76, + 821, + 199, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 76, + 847, + 199, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The base network $F(\\cdot)$ constructs a disjoint set of parameters for each task under an iterative train-prune", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "retrain framework [28]. At a time step $t$ , the base network is learned for the $t$ -th task with the dataset $T^t = \\{(x_n^t, y_n^t)\\}_{n=1}^{N^t}$ , where $x_i^t$ is the $i$ -th sample and $y_i^t$ is corresponding target, respectively. Before training the task, we allocate parameters for the new task to the pruned positions in the base network trained for $t-1$ tasks. As a typical strategy, the allocated parameters $W^t$ are learned from the help of all previous parameter sets [28]. After training, redundant parameters in $W^t$ are pruned, and the unpruned parameters, $\\bar{W}^t$ , are retrained. Finally, after learning for $t$ tasks, $F(\\cdot; \\mathcal{W}^t)$ is parameterized with $\\mathcal{W}^t = \\{\\bar{W}^1, \\dots, \\bar{W}^t\\}$ .", + "bbox": [ + 496, + 446, + 893, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although using previous knowledge can be beneficial in learning new tasks, there is also the possibility of negative knowledge transfer. If earlier tasks are irrelevant to a new task, leveraging their knowledge leads to suboptimal performance of the new task. A recent approach involving the selective use of helpful parameters [15] has been proposed. However, it is based on fixed old knowledge, which limits the model's ability to evolve in response to new information. Consequently, the expressiveness of the model remains constrained by its dependence on immutable prior knowledge. To resolve the issue, we propose a novel framework that transforms old parameters into those absorbing the knowledge of newer tasks. To selectively transform the parameter sets of old tasks $\\{\\bar{W}^1,\\dots ,\\bar{W}^{t - 1}\\}$ for the time $t$ , we generate parameters by presenting a sparsity-inducing hypernetwork. The generated ones enhance the previous parameters to fit the new task better. The entire procedure of the proposed method is illustrated in Figure 2.", + "bbox": [ + 496, + 628, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18963", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Evolving Old Knowledge", + "text_level": 1, + "bbox": [ + 76, + 90, + 305, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach involves transforming previously learned parameters into more adaptable ones to incorporate knowledge from a new task while retaining the old knowledge. This allows for the evolution of existing knowledge to accommodate new information better. Let us denote the hypernetwork as $H(\\cdot; \\mathcal{H}^t)$ parameterized by $\\mathcal{H}^t$ based on the transformer architecture [42]. Note that $\\mathcal{H}^t$ is initialized with the parameters of the hypernetwork after training the previous task, $t - 1$ . Specifically, the hypernetwork receives the task token $e^t \\in \\mathbb{R}^D$ and the set of layer embeddings $\\mathcal{C}_l = \\{c_n \\in \\mathbb{R}^D | 1 \\leq n \\leq N_l\\}$ and generates the parameter set $G_l^t$ , where $N_l$ and $l \\in \\{1 \\dots L\\}$ denote the number of embeddings for layer $l$ and the layer index of the base network, respectively.", + "bbox": [ + 75, + 113, + 468, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate $G_{l}^{t}$ the task token $e^{t}$ and layer embedding $c_{n}$ from the set $C_l$ are concatenated to construct the input of the hypernetwork, i.e., $z_{n,0}^{t} = \\mathrm{Concat}(e^{t},c_{n})\\in \\mathbb{R}^{2\\times D}$ . The input $z_{n,m - 1}^{t}$ is fed into the $m$ -th task-layer attention block (TLA) as", + "bbox": [ + 75, + 324, + 468, + 400 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {z} _ {n, m - 1} ^ {t} = z _ {n, m - 1} ^ {t} + \\operatorname {M S A} (\\mathrm {L N} (z _ {n, m - 1} ^ {t})), \\tag {1} \\\\ z _ {n, m} ^ {t} = \\bar {z} _ {n, m - 1} ^ {t} + \\operatorname {M L P} (\\operatorname {L N} (\\bar {z} _ {n, m - 1} ^ {t})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 409, + 468, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m \\in \\{1, \\dots, N_b\\}$ . $N_b$ , LN, MSA, and MLP denote the number of TLA blocks, layer normalization, multi-head self-attention, and multi-layer perceptron, respectively. The structure of TLA is illustrated in Figure 3. MSA is defined for the input $\\hat{z}_{n,m-1}^t \\triangleq \\mathrm{LN}(z_{n,m-1}^t)$ as", + "bbox": [ + 75, + 457, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M S A} \\left(\\hat {z} _ {n, m - 1} ^ {t}\\right) = W ^ {O} \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {N _ {h}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 545, + 468, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$W^{O} \\in \\mathbb{R}^{2 \\times D}$ is a projection matrix and $N_{h}$ denotes the number of heads in the MSA layer, where $\\mathrm{head}_{i} \\triangleq \\operatorname{softmax}(Q_{i}K_{i}^{T} / \\sqrt{D / N_{h}})V_{i}$ and", + "bbox": [ + 76, + 574, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} = W _ {Q _ {i}} \\hat {z} _ {n, m - 1} ^ {t}, K _ {i} = W _ {K _ {i}} \\hat {z} _ {n, m - 1} ^ {t}, V _ {i} = W _ {V _ {i}} \\hat {z} _ {n, m - 1} ^ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 632, + 472, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{Q_i}, W_{K_i}$ , and $W_{V_i}$ are projection matrices in the $i$ -th head. The parameters for the $l$ -th layer are generated as", + "bbox": [ + 76, + 662, + 468, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} G _ {l} ^ {t} = \\operatorname {C o n c a t} \\left(g \\left(\\theta_ {1} ^ {t}\\right), \\dots , g \\left(\\theta_ {N _ {l}} ^ {t}\\right)\\right) \\tag {3} \\\\ = \\operatorname {C o n c a t} \\left(H \\left(z _ {1, 0} ^ {t}; \\mathcal {H} ^ {t}\\right), \\dots , H \\left(z _ {N _ {l}, 0} ^ {t}; \\mathcal {H} ^ {t}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 713, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g(\\cdot)$ denotes the fully connected layer followed by a sigmoid function. $\\theta_{n}^{t}\\in \\mathbb{R}^{D}$ is the first row of $z_{n,N_b}^t$ , which acts as the final embedding in a transformer architecture [8].", + "bbox": [ + 76, + 762, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To refine old knowledge to align with the knowledge of the $t$ -th task, let us denote the set of old parameters in the $l$ -th layer as $\\bar{W}_l^{1:t-1}$ . We obtain the evolved parameters $\\hat{W}_l^{1:t-1}$ by fusing $\\bar{W}_l^{1:t-1}$ and $G_l^t$ as", + "bbox": [ + 76, + 808, + 468, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {W} _ {l} ^ {1: t - 1} = \\bar {W} _ {l} ^ {1: t - 1} \\cdot G _ {l} ^ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 881, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/580be8affda3d993b9c77be535a414575ea27b86e97208bdaa156a6be0e9ba73.jpg", + "image_caption": [ + "Figure 3. The task-layer attention block (TLA) consists of a layer normalization, multi-head self-attention (MSA), and MLP with a single hidden layer. $z_{n,N_b}^t$ is generated by $N_{b}$ stacked TLAs." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 883, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we define $\\cdot$ as the element-wise multiplication. The base network accommodates the evolved sets of parameters for $t - 1$ tasks together with the $t$ -th parameter set as $\\hat{\\mathcal{W}}^t = \\{\\hat{W}^1, \\dots, \\hat{W}^{t-1}, W^t\\}$ . Finally, the loss function to learn the $t$ -th task using the evolved parameter sets becomes", + "bbox": [ + 498, + 250, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t a s k}} = - \\sum_ {n = 1} ^ {N ^ {t}} y _ {n} ^ {t} \\log \\left(F \\left(x _ {n} ^ {t}; \\hat {\\mathcal {W}} ^ {t}\\right)\\right)). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 335, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another loss function to generate parameters from $H(\\cdot ;\\mathcal{H}^t)$ similar to those learned in the previous task is defined as", + "bbox": [ + 498, + 388, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\sum_ {k = 1} ^ {t - 1} \\sum_ {l = 1} ^ {L} \\| \\bar {G} _ {l} ^ {k} - G _ {l} ^ {k} \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 431, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{G}_l^k$ represents the set parameters generated from the hypernetwork after training the task $t - 1$ .", + "bbox": [ + 498, + 484, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sparsity-Inducing Generation", + "text_level": 1, + "bbox": [ + 500, + 523, + 766, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To renew the network without redundant parameters, we propose to generate sparse parameters in the hypernetwork. As $G_{l}^{t}$ is generated through the sigmoid function, they can be interpreted as the likelihood that each element of the old set of parameters is selected. Following the practice in [41], we minimize $\\sum_{l=1}^{L} -\\log(1 - G_{l}^{t})$ to achieve sparsity in $G_{l}^{t}$ . However, to prevent $G_{l}^{t}$ from becoming overly sparse, we ensure that parameters significantly impacting the loss remain survived. Inspired by the work in [24], we evaluate the impact of individual elements in $G_{l}^{t1}$ to the loss. We compute the set of all entries of the score $S_{l}^{t}$ by comparing the network losses computed with and without the inclusion of parameters in the learning process. To estimate the effect of removing a parameter on the loss function, we apply a Taylor expansion approximation of the network loss $\\mathcal{L}$ for $N$ samples. We obtain the $k$ -th element of the score, $s_{k}$ , as", + "bbox": [ + 496, + 547, + 890, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s _ {k} = \\mathcal {L} \\left(\\vec {G} _ {l} ^ {t} - M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t}\\right) - \\mathcal {L} \\left(\\vec {G} _ {l} ^ {t}\\right) \\tag {7} \\\\ \\simeq - M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t} \\cdot \\nabla_ {\\vec {G} _ {l} ^ {t}} \\mathcal {L} + \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\frac {\\partial \\mathcal {L} _ {n}}{\\partial M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t}}\\right) ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 797, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We denote the vectorized form of $G_{l}^{t}$ as $\\vec{G}_{l}^{t}$ to represent individual entries.", + "bbox": [ + 500, + 873, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18964", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{\\backslash (k)}$ is an all-zero vector except that the $k$ -th element is 1, and $\\mathcal{L}_n$ is the loss of the $n$ -th example. We introduce an additional term minimizing the discrepancy between the scores of the complete set of parameters (i.e., the original scores) and the scores of the sparse set of parameters produced as follows:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s p} = \\sum_ {l = 1} ^ {L} - \\log (1 - G _ {l} ^ {t}) + \\left\\| S _ {l} ^ {t} \\cdot \\left(1 - \\mathbb {1} (\\vec {G} _ {l} ^ {t} \\neq 0)\\right) \\right\\| _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 193, + 468, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_1$ and $\\mathbb{1}(\\cdot)$ denote the $l_1$ -norm and the indicator function, respectively.", + "bbox": [ + 75, + 244, + 468, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total Loss. The proposed framework learns the $t$ -th task with the three loss functions. To learn the $t$ -th task, we minimize the cross-entropy loss $\\mathcal{L}_{task}$ by updating $W^t$ in the base network. Also, $\\mathcal{H}^t$ , $e^t$ , and $\\{C_l\\}_{l=1}^L$ are updated from two losses, $\\mathcal{L}_{task}$ and $\\mathcal{L}_{sp}$ . Last, we update $\\mathcal{H}^t$ and $\\{C_l\\}_{l=1}^L$ by minimizing $\\mathcal{L}_{reg}$ to preserve the previously generated parameters. The total loss is", + "bbox": [ + 75, + 276, + 468, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {W ^ {t}, \\mathcal {H} ^ {t}, e ^ {t}, \\{C _ {l} \\} _ {l = 1} ^ {L}} \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {s p} \\mathcal {L} _ {s p} + \\lambda_ {r e g} \\mathcal {L} _ {r e g}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 393, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{sp}$ and $\\lambda_{reg}$ are parameters to balance the losses.", + "bbox": [ + 75, + 430, + 450, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scalability. A hypernetwork is designed to generate parameters for every layer of the base network, and the parameters were composed of compact segments of a consistent size [13, 21]. While it can generate parameters for all layers, this approach requires a substantial quantity of layer embeddings, particularly for deeper layers. To mitigate this, we adjust the scale of the segments for each layer depending on the number of channels. However, it incurs a memory overhead stemming from the substantial dimensions of the heads. To address this, we apply a matrix decomposition technique [39] to generate two reduced sets of low-rank $r$ . By doing so, we achieve a reduction in memory requirements for layer embeddings, accompanied by a marginal increase in hypernetwork parameters. For detailed results, please refer to Section 4.5.1.", + "bbox": [ + 75, + 446, + 472, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Scenarios", + "text_level": 1, + "bbox": [ + 76, + 710, + 187, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We applied the proposed method, GrowBrain, to image classification and video action recognition problems under class incremental and task-incremental learning scenarios. Image classification. We conducted experiments using ImageNet [6] and five fine-grained datasets including CUBS [45], Stanford Cars [18], Flowers [31], Wikiart [36], and Sketch [11], where each dataset is considered a task. We call the datasets ImageNet+FG for short. In addition, we demonstrated the proposed method with the CIFAR-100. We divided the dataset into subsets of classes based on the random order provided by [35, 9] and considered", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/633e537d4b70efcde203f3f7ccb546bf9bba75488e54935a641ce5713b8b6bd1.jpg", + "table_caption": [ + "Table 1. Datasets used in this work." + ], + "table_footnote": [], + "table_body": "
Image Dataset# Train Images# Test Images# Classes
ImageNet [6]1,287,16750,0001,000
CUBS [45]5,9945,794200
Stanford Cars [18]8,1448,041196
Flowers [31]2,0406,149102
Wikiart [36]42,12910,628195
Sketch [11]16,0004,000250
CIFAR-100 [19]50,00010,000100
Video Dataset# Train Videos# Test Videos# Classes
ActivityNet [3]10,0244,926200
UCF-101 [40]9,2802,720101
", + "bbox": [ + 511, + 99, + 880, + 241 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/21b7498a5d3c3a748e3447ad45f4a1bd2d45c1a7a3c8dbf9981ce15644901e3e.jpg", + "table_caption": [ + "Table 2. The specification of the hypernetwork." + ], + "table_footnote": [], + "table_body": "
Base networkHypernetwork
Embed dim (D)Hidden dim in MLP# Heads (Nh)# Blocks (Nb)
ResNet-506412887
ResNet-34
ResNet-18326485
", + "bbox": [ + 511, + 265, + 880, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each subset a task. Video action recognition. We used a trimmed version of ActivityNet [3] and UCF-101 [40] video datasets. To conduct experiments, we divided each dataset into subsets containing randomly selected classes provided by [43] and treated each subset as a task. We summarize the datasets as mentioned above in Table 1.", + "bbox": [ + 496, + 367, + 890, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 468, + 714, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image classification. ImageNet+FG and CIFAR-100 were resized to $224 \\times 224$ and $32 \\times 32$ pixels, respectively. We applied random cropping and horizontal flip as augmentation. For those datasets, we used the ResNet-50 [14] and ResNet-18 as the base networks, respectively.", + "bbox": [ + 496, + 491, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video action recognition. We used TSN [46] with the ResNet-34 backbone network for video action recognition. Each video was split into three segments of equal duration, and a frame was randomly selected from each segment. The action class prediction was performed by aggregating the predictions of the selected frames with a consensus function. We used an average of predictions as a consensus function [46]. The video datasets were augmented with random cropping and horizontal flipping following [46]. We resized each frame of ActivityNet and UCF-101 to $224 \\times 224$ pixels.", + "bbox": [ + 496, + 566, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 reports the base networks and corresponding hypernetworks, which include the embedding dimension, the hidden dimension of the MLP, the number of heads, and the number of TLA blocks. We set the rank $r$ for matrix decomposition as 7. We trained the base and hypernetwork using stochastic gradient descent. In all our experiments, we set $\\lambda_{reg}$ to 1 and $\\lambda_{sp}$ to 0.1, except ImageNet+FG where we set $\\lambda_{sp}$ to 1. We report the average results over five independent runs for all the experiments. To enable class-incremental learning in scenarios where the task identity is unknown, the proposed method first predicts the task index and then conducts classification. Following [44], we predict the task", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18965", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/35848c83cbeeb73fb2f371f4ec1d28c5bdcc29126789e29cc57567bd7dfc3156.jpg", + "table_caption": [ + "Table 3. Performance of the class-incremental learning scenario using ImageNet and five fine-grained datasets (ImageNet+FG), where the average accuracy was measured after training each task. # P represents the average number of parameters $(\\times 10^{6})$ to perform the tasks." + ], + "table_footnote": [], + "table_body": "
MethodImageNetCUBSStanford CarsFlowersWikiartSketch
Acc#PAcc#PAcc#PAcc#PAcc#PAcc#P
EWC-On76.123.431.623.428.623.429.223.423.123.411.323.4
LwF76.123.468.423.458.223.455.823.449.223.428.423.4
iCaRL76.123.470.223.461.523.452.723.448.323.442.323.4
PackNet75.712.562.813.365.214.066.514.654.315.245.315.7
Piggyback76.123.460.122.754.822.256.422.049.721.747.421.4
WSN74.712.559.612.542.212.529.212.531.012.525.012.5
H275.712.569.513.067.312.968.413.061.812.948.712.6
GrowBrain75.712.569.211.069.310.371.39.864.39.660.19.4
", + "bbox": [ + 94, + 114, + 875, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a44403d2458280411c417bba241f1ad75828a91eca21a4190b80f5006df31cb2.jpg", + "image_caption": [ + "Figure 4. Accuracy of class-incremental learning for image classification on ImageNet+FG (left), video action recognition on UCF-101 (middle) and ActivityNet (right)." + ], + "image_footnote": [], + "bbox": [ + 84, + 258, + 339, + 462 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e98dedb86e4c1568827fdaec08a34720fcb9f3adf90a47b284ad3de833a6892.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 258, + 612, + 464 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee8d531d8ec70c4bb451d10a2eda8c0c6e3279c02b4bb822f3edb3ddf8a53f0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 258, + 883, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "index using the network with the least entropy that has been learned thus far.", + "bbox": [ + 75, + 525, + 468, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Class-incremental learning", + "text_level": 1, + "bbox": [ + 76, + 563, + 320, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Image classification", + "text_level": 1, + "bbox": [ + 76, + 585, + 267, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated the proposed method of evolving the previous knowledge to the class-incremental learning scenario using ImageNet+FG. We compared with regularization methods, EWC-On [37] and LwF [22], and data rehearsal methods, iCaRL [35], and parameter isolation methods, PackNet [28], Piggyback [27], WSN [16], and $\\mathrm{H}^2$ [15]. The data rehearsal methods stored 1,000 raw instances (taking 150 $\\mathrm{MB}^2$ ). We applied the same task prediction mechanism as ours to the comparing methods, excluding iCaRL.", + "bbox": [ + 75, + 609, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 shows the class-incremental learning results for ImageNet and five fine-grained datasets. Overall, the proposed method outperforms competitors (more detailed results are shown on the left of Figure 4). The regularization approaches, EwC-On and LwF, give unsatisfactory results with performance gaps of $48.8\\%$ and $31.7\\%$ compared to ours, respectively. LwF shows a $20.8\\%$ accuracy drop before and after learning Sketch, which is the most significant", + "bbox": [ + 75, + 746, + 470, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "among the comparing methods. The data rehearsal method, iCaRL, shows a $17.1\\%$ performance drop from GrowBrain despite storing 1,000 old samples. PackNet shows large performance drops of $12.2\\%$ and $9.0\\%$ after training on WikiArt and Sketch, respectively. This is because negative interference occurs when using all previous sets of parameters optimized for different task domains. The methods of associating prior knowledge, Piggyback, WSN, and $\\mathrm{H}^2$ , show notable performance drop ( $12.7\\%$ , $35.1\\%$ , and $11.4\\%$ , respectively) compared to the proposed method that improves the previous knowledge from the new task. The results demonstrate that the proposed approach surpasses a wide range of continual learning methods, confirming its superiority.", + "bbox": [ + 504, + 525, + 890, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Video action recognition", + "text_level": 1, + "bbox": [ + 500, + 755, + 723, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We applied our method and the competitors to the video action recognition tasks using the benchmark sets proposed for continual learning [43]. For this experiment, iCaRL was excluded from the comparison since it demands approximately 2TB and 52GB of memory for replay [43] on ActivityNet and UCF-101, respectively. Since Piggyback requires a pre-trained backbone, we pre-trained the backbone model [14] using the validation set of UCF-101 for video action", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2 We calculated the rehearsal memory assuming that each image (or frame of video) has $224 \\times 224 \\times 3$ pixels.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18966", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dfd6c5ed318d4c6a5095845e10c0d7701737161be5874e0c4ff195be3741aec4.jpg", + "table_caption": [ + "Table 4. Results of the class-incremental learning scenario using the trimmed version of UCF-101 and ActivityNet, where the Avg and Last accuracies for the time steps 5 and 10." + ], + "table_footnote": [], + "table_body": "
MethodUCF-101ActivityNet
T5T10T5T10
AvgLast# PAvgLast# PAvgLast# PAvgLast# P
EWC-On48.725.821.232.713.921.238.420.721.226.612.421.2
LwF47.027.721.233.415.421.244.129.121.232.718.421.2
PackNet59.354.211.552.540.415.242.430.211.532.419.015.2
Piggyback65.451.720.453.536.520.422.212.720.515.67.520.4
WSN69.158.810.161.149.710.5------
H263.654.88.356.947.011.437.422.53.426.813.32.1
GrowBrain79.372.27.273.564.78.942.829.911.533.220.615.2
", + "bbox": [ + 163, + 114, + 805, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "recognition experiments. We report the $Avg$ and Last accuracies [9] for the time steps 5 and 10. Figure 4 shows the Last accuracy measured on the UCF-101 and ActivityNet datasets.", + "bbox": [ + 76, + 281, + 468, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UCF-101. The results of the class-incremental learning scenario using the UCF-101 dataset are reported in Table 4 (left). The aspects of the experimental results on UCF-101 of the compared approaches are similar to those of image classification. Despite utilizing a large number of parameters, regularization approaches, EWC-On and LwF, give unsatisfactory performance compared to parameter isolation approaches. PackNet, which uses all parameters from previous tasks, performs less than other parameter isolation approaches, WSN and $\\mathrm{H}^2$ , using selectively chosen parameters associated with previous knowledge. The proposed method shows the highest performance compared to other comparison methods. Ours outperforms WSN, $\\mathrm{H}^2$ , PackNet, and Piggyback with the significant accuracy gap on the Last measure from $15.0\\%$ to $28.2\\%$ .", + "bbox": [ + 76, + 357, + 468, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ActivityNet. Table 4 (right) reports the results on ActivityNet. WSN was excluded from the experiment because it did not give promising results due to the lack of data. Overall, the results of ActivityNet show a different pattern compared to those of UCF-101. The regularization method, LwF, outperforms the parameter isolation method, PackNet, with $1.7\\%$ and $0.3\\%$ Avg accuracy gap measured after training the 5-th and 10-th tasks. The proposed approach performs better than Piggyback, PackNet, and $\\mathrm{H}^2$ by large margins of $13.1\\%$ , $1.6\\%$ , and $7.3\\%$ , respectively, in terms of the Last accuracy measured after training 10-th task. The proposed method outperforms other competitors for image classification and video action recognition problems. In contrast to the UCF-101 dataset, where the proposal can acquire new knowledge by using a small number of parameters from previous tasks and achieve excellent results, the ActivityNet dataset requires previous knowledge to effectively train the current task due to its inadequate data (the number of videos per class is only half compared to UCF-101).", + "bbox": [ + 76, + 598, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a26e33be965b83186eb7607b0d109f30cfdabf23d398dc9da31f6e0fbbf36836.jpg", + "table_caption": [ + "Table 5. Task-incremental learning scenarios using video action recognition and image classification datasets, where average accuracy and parameter consumption are reported." + ], + "table_footnote": [], + "table_body": "
MethodImage classificationVideo action recognition
ImageNet+FGCIFAR-100UCF-101ActivityNet
Acc#PAcc#PAcc#PAcc#P
EWC-On55.323.444.511.165.821.261.521.2
LwF53.623.439.711.171.221.265.821.2
PackNet80.215.782.110.492.215.265.315.2
Piggyback82.221.475.610.586.620.442.120.4
WSN61.712.584.65.692.710.5--
H282.812.685.91.194.511.444.02.1
GrowBrain83.09.486.82.596.78.966.915.2
", + "bbox": [ + 501, + 319, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Task-incremental learning", + "text_level": 1, + "bbox": [ + 500, + 483, + 738, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we conducted experiments on the task-incremental learning scenario where task-id is given during the inference step. In this scenario, we used ImageNet+FG and CIFAR-100 as the image classification datasets. Also, we used ActivityNet and UCF-101 for video action recognition. We divided CIFAR-100 into 10 tasks based on [35]. We followed the same experimental setting as in Tables 3 and 4. For the experiments of Piggyback on CIFAR-100, the backbone network was pre-trained on a randomly selected subset (40%) of the CIFAR-100 training set.", + "bbox": [ + 496, + 507, + 890, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image classification. Table 5 reports the results of task-incremental learning conducted on ImageNet+FG and CIFAR-100. Even if the regularization methods, EwC-On and LwF, try to maintain previous knowledge, they perform significantly less than the parameter isolation methods. GrowBrain consistently achieves high performance compared to other parameter isolation methods, PackNet, Piggyback, WSN, and $\\mathrm{H}^2$ , in both image classification experiments (e.g., $0.2\\%$ and $0.9\\%$ performance gap on ImageNet+FG and CIFAR-100 with the most competitive method, $\\mathrm{H}^2$ ).", + "bbox": [ + 496, + 657, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video action recognition. Table 5 reports the results of task-incremental learning conducted on UCF-101 and ActivityNet. Overall, the results are similar to those of class-incremental learning scenarios. The regularization method, LwF, achieves higher performance than PackNet on Activ", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18967", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/30dd18089bb50a9adbb811434576c333f65769981cb84407cd1b3a41bb00e802.jpg", + "table_caption": [ + "Table 6. Ablation study of the proposed method for image classification and video action recognition tasks, where average accuracy, the number of parameters for the base network and hypernetwork (in millions), and the number of layer embeddings are reported." + ], + "table_footnote": [], + "table_body": "
MethodImageNet+FGUCF-101
Acc# (F + H)# cnAcc# (F + H)# cn
Ours60.19.4 + 2.813364.78.9 + 2.858
w/o transform [28]45.315.7 + 0040.415.2 + 00
w/o sparsity58.115.7 + 2.813364.014.1 + 2.858
w/o scalability59.815.7 + 2.9326264.711.0 + 2.9614
", + "bbox": [ + 78, + 142, + 468, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ityNet. The proposed method outperforms both parameter isolation and regularization methods. Notably, our method outperforms WSN, PackNet, and $\\mathrm{H}^2$ with a performance gap of $4\\%$ , $4.5\\%$ , and $2.2\\%$ , respectively, on UCF-101.", + "bbox": [ + 75, + 237, + 468, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis", + "text_level": 1, + "bbox": [ + 76, + 308, + 179, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.1 Ablation study", + "text_level": 1, + "bbox": [ + 76, + 332, + 233, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our method, we conducted an ablation study. We compared ours by removing transformation of previous knowledge [28], sparsity-inducing generation $(\\lambda_{sp} = 0)$ , and scalability of hypernetwork, i.e., channel adaptive segments and matrix decomposition. We applied these methods to class-incremental learning scenarios using the ImageNet+FG and the UCF-101 datasets. We report the Last accuracy measured after training all tasks, the average parameter consumption of the base network $(F)$ and hypernetwork $(H)$ to perform each task, and the number of layer embeddings $c_{n}$ .", + "bbox": [ + 75, + 356, + 468, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 reports the results of the ablation study. Our method outperforms that without transformation by performance gaps of $14.8\\%$ and $24.3\\%$ for those datasets, emphasizing that the evolution of prior knowledge is critical in enhancing performance. For video action recognition using UCF-101, the method without sparsity requires more parameters and performs less than the proposed method. The proposed method for ImageNet+FG achieves higher performance while taking fewer parameters than the nonsparse alternative. The reason is that sparsity-inducing hypernetwork prevents the generation of redundant parameters and regularizes the generated parameters to achieve generalization performance. The method without scalability requires 24 and 10.5 times larger layer embeddings on ImageNet+FG and UCF-101 than the proposed method, respectively, without accompanying performance improvement.", + "bbox": [ + 75, + 523, + 468, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.2 Impact of $\\lambda_{sp}$", + "text_level": 1, + "bbox": [ + 76, + 785, + 225, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further conducted a study investigating the impact of $\\lambda_{sp}$ , which regulates the sparsity of generated parameters. We experimented with varying values of $\\lambda_{sp}$ in the ImageNet+FG experiment under the class-incremental scenario using the same setup presented in Table 3. Figure 5 shows the parameter consumption (left) and average ac", + "bbox": [ + 76, + 809, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89486965460d5420072724b4798da3eb4be9cabae17d24a4a876116dc731659d.jpg", + "image_caption": [ + "Figure 5. Parameter consumption and average accuracy of the proposed method based on varying $\\lambda_{sp}$ in class-incremental learning for ImageNet+FG." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 694, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1fe5106a3da215418b6ab96e9515f51cfa3834f5cf3824f27c28c231544552a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 88, + 887, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "curacy (right) with respect to different $\\lambda_{sp}$ after training all tasks. Ours with $\\lambda_{sp} = 0.1$ uses a similar amount of parameters as the baseline [28] but outperforms the competing approach by leveraging enhanced prior knowledge. Notably, after training all tasks, $\\lambda_{sp} = 1.0$ takes $9.4\\mathrm{M}$ parameters while showing $0.7\\%$ performance improvement compared to $\\lambda_{sp} = 0.1$ consuming $15.7\\mathrm{M}$ parameters.", + "bbox": [ + 496, + 281, + 890, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 422, + 617, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a novel approach, GrowBrain, to transform learned old knowledge to absorb the information of newer tasks in parameter isolation-based continual learning. GrowBrain reinforces the previous knowledge with a transformer-based hypernetwork that generates optimized parameters for new tasks while eliminating redundant parameters that show negligible impact on the loss. The proposed method overcomes the limitation of using fixed old knowledge that can constrain the ability to capture novel representations while taking fewer parameters. In the experiments, we have evaluated the proposal compared to a variety of continual learning methods in class- and task-incremental learning scenarios for image classification and video action recognition. Experimental results show that the proposed approach of leveraging transformed prior knowledge generates parameters adapted to the new task without losing the prior knowledge, resulting in significant performance improvements over its strong competitors.", + "bbox": [ + 496, + 455, + 890, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT2002-05, Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (2022-0-01341, Artificial Intelligence Graduate School Program(Chung-Ang University)), and BK21 FOUR (Fostering Outstanding Universities for Research) Program funded by Ministry of Education of Korea (No.I22SS7609062).", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18968", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3931-3940, 2020. 1, 2", + "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European Conference on Computer Vision, pages 144-161. Springer, 2018. 1, 2", + "[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5", + "[4] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 556-572. Springer, 2018. 1, 2", + "[5] Arslan Chaudhry, Ranzato Marc'Aurelio, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In 7th International Conference on Learning Representations, ICLR 2019. 1, 2", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5", + "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1, 2", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 4", + "[9] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 5, 7", + "[10] Benjamin Ehret, Christian Henning, Maria Cervera, Alexander Meulemans, Johannes Von Oswald, and Benjamin F Grewe. Continual learning in recurrent neural networks. In International Conference on Learning Representations, 2020. 3", + "[11] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Transactions on Graphics, 31(4):1-10, 2012. 5", + "[12] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3, 5", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6", + "[15] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, pages 519-535. Springer, 2022. 1, 2, 3, 6", + "[16] Haeyong Kang, Rusty John Lloyd Mina, Sultan Rizky Hikmawan Madjid, Jaehong Yoon, Mark Hasegawa-Johnson, Sung Ju Hwang, and Chang D Yoo. Forget-free continual learning with winning subnetworks. In International Conference on Machine Learning, pages 10734-10750. PMLR, 2022. 2, 6", + "[17] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America, 114(13):3521-3526, 2017. 1, 2", + "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561. IEEE, 2013. 5", + "[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 2, 5", + "[20] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. Advances in Neural Information Processing Systems, 30:4652-4662, 2017. 1, 2", + "[21] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. DHP: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2, 3, 5", + "[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. In European Conference on Computer Vision, pages 614-629. Springer, 2016. 1, 2, 6", + "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2", + "[24] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group fisher pruning for practical network compression. In International Conference on Machine Learning, pages 7021-7032. PMLR, 2021. 4", + "[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2544-2553, 2021. 1, 2", + "[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6470-6479, 2017. 1, 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18969", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In Proceedings of the European Conference on Computer Vision, pages 67-82, 2018. 2, 6", + "[28] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 1, 2, 3, 6, 8", + "[29] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, volume 24, pages 109–165. Elsevier, 1989. 1", + "[30] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1, 2", + "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5", + "[32] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 1", + "[33] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13698-13707, 2021. 2", + "[34] Dripta S Raychaudhuri, Yumin Suh, Samuel Schulter, Xiang Yu, Masoud Faraki, Amit K Roy-Chowdhury, and Manmohan Chandraker. Controllable dynamic multi-task architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2022. 2", + "[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 5, 6, 7", + "[36] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. International Journal for Digital Art History, (2), 2016. 5", + "[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International Conference on Machine Learning, pages 4528-4537. PMLR, 2018. 6", + "[38] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 2994-3003, 2017. 1, 2", + "[39] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5", + "[41] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. AdaShare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 4", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4", + "[43] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vCLIMB: A novel video class incremental learning benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19035-19044, 2022. 5, 6", + "[44] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. In 8th International Conference on Learning Representations, 2020. 2, 3, 5", + "[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 5", + "[47] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2", + "[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3987-3995, 2017. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 628 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18970", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_model.json b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9da3a3fa4e09b3e28af9ff934548f226a601b980 --- /dev/null +++ b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "Growing a Brain with Sparsity-Inducing Generation for Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.18, + 0.796, + 0.235 + ], + "angle": 0, + "content": "Hyundong Jin1 Gyeong-hyeon Kim1 Chanho Ahn2* Eunwoo Kim1 \n1School of Computer Science and Engineering, Chung-Ang University \n2Samsung Advanced Institute of Technology (SAIT)" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.237, + 0.778, + 0.253 + ], + "angle": 0, + "content": "{jude0316,leonardkkh,eunwoo}@cau.ac.kr chanho.ahn@samsung.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.473, + 0.683 + ], + "angle": 0, + "content": "Deep neural networks suffer from catastrophic forgetting in continual learning, where they tend to lose information about previously learned tasks when optimizing a new incoming task. Recent strategies isolate the important parameters for previous tasks to retain old knowledge while learning the new task. However, using the fixed old knowledge might act as an obstacle to capturing novel representations. To overcome this limitation, we propose a framework that evolves the previously allocated parameters by absorbing the knowledge of the new task. The approach performs under two different networks. The base network learns knowledge of sequential tasks, and the sparsity-inducing hypernetwork generates parameters for each time step for evolving old knowledge. The generated parameters transform old parameters of the base network to reflect the new knowledge. We design the hypernetwork to generate sparse parameters conditional to the task-specific information and the structural information of the base network. We evaluate the proposed approach on class-incremental and task-incremental learning scenarios for image classification and video action recognition tasks. Experimental results show that the proposed method consistently outperforms a large variety of continual learning approaches for those scenarios by evolving old knowledge." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.21, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.47, + 0.876 + ], + "angle": 0, + "content": "The ability of deep learning to incrementally accumulate knowledge for sequentially incoming tasks is an important capability required for modern AI algorithms. However, deep neural networks tend to lose previously acquired knowledge while learning new data. This limitation of deep neural networks, called catastrophic forgetting, has been investigated by previous studies [29, 17]. Catastrophic forgetting significantly degrades performance when data from previous tasks (task-incremental) or already-learned class" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.334 + ], + "angle": 0, + "content": "data (class-incremental) cannot be accessed. Under those incremental learning scenarios, continual learning methods have toiled to preserve the previous knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.893, + 0.639 + ], + "angle": 0, + "content": "Various approaches have been proposed for continual learning and are categorized into three types: regularization [2, 4, 20, 17, 48], data rehearsal [5, 26, 32, 35, 38, 47], and parameter isolation [1, 30, 28]. The conventional continual learning methods introduced regularization approaches to suppress changes in important parameters [48, 2] or input-output mappings from a network trained up to previous tasks [22, 7]. The network is trained in a way that contributable parameters to the performance of the previously trained model remain unchanged. On the other hand, data rehearsal methods [25, 35] address catastrophic forgetting by accessing old data. The old data can be obtained by storing some data from old tasks [35, 47] or leveraging a generative network [38]. However, the privacy of data and the difficulty of learning generative models limit their versatility. Recently, parameter isolation approaches [28, 30, 15] construct a disjoint set of parameters for each task, preventing it from mixing old and new knowledge. These methods learn a new set of parameters by associating all [28] or some helpful old knowledge [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.806 + ], + "angle": 0, + "content": "Unlike other approaches, the parameter isolation approach is designed to prevent forgetting without requiring access to old data by disallowing updates to previous sets of parameters. Despite this advantage, these methods can limit the ability to learn new tasks by relying on parameters that were learned for older tasks and remain unchanged. We observe that this limitation arises when there is a low correlation between the old and new data. This raises the question: In parameter isolation-based continual learning, is it possible to transform the knowledge from old tasks to fit newer tasks better?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce a novel method for transforming prior knowledge into a form that can aid in learning new tasks. The approach involves reframing the parameter selection problem, which previous works [15] have addressed, as a generation problem. The proposed framework emphasizes a subset of pre-trained knowledge, selecting and highlight-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.462, + 0.902 + ], + "angle": 0, + "content": "*This work was done independently, without any support from SAIT." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18961" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.438 + ], + "angle": 0, + "content": "ing it for use in learning new tasks. Specifically, the framework consists of two networks: base network and hypernetwork. The base network constructs disjoint parameter sets for sequential tasks. The proposed transformer-based [42] hypernetwork generates a small set of influential parameters that significantly impact the loss. The impact of each element of generated parameters is evaluated by the difference in loss computed with and without its inclusion in the learning process. Also, generated parameters are conditioned on the information of the current task (task token) and the layer of the base network (layer embedding). The generated parameters transform the previous set of parameters in the base network to further accommodate new knowledge. In addition, compared to existing hypernetworks that generate a small set of parameters and apply it to all the layers [13, 44], the proposed hypernetwork adaptively generates parameters according to the layer configurations defined in the base network. Through this strategy, the quantity of layer embeddings for ResNet-50 is reduced by a substantial factor of 24. Furthermore, by leveraging a matrix decomposition technique to represent generated parameters compactly, we achieve a noteworthy reduction in the parameters necessary for the hypernetwork." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.473, + 0.727 + ], + "angle": 0, + "content": "We evaluate the effectiveness of the proposed method in both class-incremental and task-incremental learning scenarios using classification benchmark datasets. Furthermore, we demonstrate the efficacy of the proposed method [28, 15] in these scenarios by applying it to video action recognition datasets [19]. Experimental results show that the proposed method performs better than its competitors for most learning scenarios while introducing minor memory overhead from the hypernetwork. The contributions of our work are mainly three-fold: (1) We present a novel approach to evolving previous knowledge in continual learning, aimed at reflecting the knowledge of new tasks. (2) We propose a transformer-based hypernetwork that generates customized parameters for new tasks while suppressing the generation of redundant parameters that have little impact on the loss. (3) Experimental results on diverse benchmarks for video action recognition and image classification demonstrate that ours outperforms its competitors with lower network capacity requirements to perform tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.22, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Continual Learning. Generally, continual learning methods can be classified into three types. Regularization methods [23, 7, 17, 48, 4, 2, 20] add a penalty term that restrains the previously learned knowledge from being updated when learning a new task. In order not to hurt previous knowledge, a new network is introduced to mimic the features obtained from the old network [22, 7]. While other regularization techniques, such as those based on the Fisher information matrix [17], change of loss [48], and variation of" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.882, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.214, + 0.895, + 0.326 + ], + "angle": 0, + "content": "Figure 1. An illustration of the proposed method (right) and baseline methods (middle, left) for the third (new) task. The figure represents task-specific allocated parameters, represented by task numbers. Gray numbers represent the sets of discarded parameters from a selection method. The baseline methods do not update old knowledge (red boxes) while updating a new task (blue boxes). The proposed method transforms the old sets of parameters, represented by \\(+\\), to fit the new task. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.895, + 0.672 + ], + "angle": 0, + "content": "prediction [2], suppress the update of crucial parameters for each task, they experience significant performance decline as the number of tasks grows [12]. Data rehearsal methods [5, 26, 35, 33, 25, 38] rely on the availability of previous task data, which is achieved through selective storage of data [35, 47, 25]. These methods employ a few previous samples jointly with a new dataset [25, 38] or constraining the gradient based on the angle between that obtained from stored samples and the gradient obtained from the new samples [5, 26]. These methods incur a data imbalance problem due to a small number of stored previous tasks compared to a large amount of new task data. Parameter isolation methods [28, 1, 30, 15, 16] involve constructing disjoint sets of parameters for old tasks, respectively, and updating another set of parameters for a newer task by associating it with old parameters during training. This is usually done by using all previous sets of parameters [28], using a learnable mask to select some previous set of parameters [27], jointly selecting a location to allocate a new set of parameters [16], or searching a subnetwork with coarse-to-fine task association [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.779 + ], + "angle": 0, + "content": "However, this approach can impede efficient learning because the parameters optimized for old tasks remain the same when learning newer tasks, meaning they need to absorb the information from the newer tasks. In contrast, the proposed method allows the old knowledge to grow by incorporating the knowledge from new tasks. The proposed method and its baselines are illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Hypernetworks. Hypernetworks generate parameters for a target network [13, 44, 39, 21, 34]. A recent work [34] proposes a method for generating parameters for layers with few parameters, such as activation of ResNet blocks or batch normalization layers. However, generating parameters for convolutional and linear layers that take most of the parameters in the base network can be challenging because it is difficult to represent the entire set of parameters from a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18962" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.872, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.34, + 0.896, + 0.427 + ], + "angle": 0, + "content": "Figure 2. The proposed method for transforming the parameters of the \\(l\\)-th layer of the base network \\(F(\\cdot)\\) when learning the \\(t\\)-th task. The numbers in the grid indicate the indices of tasks. The hypernetwork \\(H(\\cdot)\\) generates parameters using the task-layer attention block (TLA), which discovers the relationship between the task token and layer embedding. The sparse parameters are generated while suppressing the generation of the redundant ones by evaluating the impact of parameters on the loss. The previous sets of parameters of the \\(l\\)-th layer are transformed by element-wise multiplication with \\(G_{l}^{t}\\). The transformed parameters are combined with the current parameter set through the union operation \\((\\cup)\\) and released as a complete set of parameters to learn new knowledge. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.629 + ], + "angle": 0, + "content": "hypernetwork. To avoid this, existing works generate small sets of parameters and concatenates them [13, 21]. Consequently, to generate a base network of large scale, allowing a large number of input embeddings corresponding to the network is inevitable due to the repetitive generation mechanism. In contrast to the approach, the parameters are generated according to the dimensions defined in the layers of the base network in our method. Note that as these dimensions expand, the corresponding demands on hypernetwork parameters also escalate. To address this, we employ a matrix decomposition technique for a compact representation of the required parameters in the hypernetwork." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.473, + 0.812 + ], + "angle": 0, + "content": "Recently, in the continual learning literature, [44] and [10] learn a hypernetwork to generate parameters with task-specific embeddings for new tasks while preventing changes to previously generated parameters for previous tasks. The approaches produce dense parameters for the base network through a task-conditional hypernetwork. In contrast, the proposed transformer-based hypernetwork generates sparse parameters by measuring the effect of individual parameters on the loss. Furthermore, our method generates parameters to evolve previous knowledge to reflect new knowledge, whereas the aforementioned approaches generate parameters directly for the base network." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.2, + 0.837 + ], + "angle": 0, + "content": "3. GrowBrain" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.2, + 0.862 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The base network \\( F(\\cdot) \\) constructs a disjoint set of parameters for each task under an iterative train-prune" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.618 + ], + "angle": 0, + "content": "retrain framework [28]. At a time step \\(t\\), the base network is learned for the \\(t\\)-th task with the dataset \\(T^t = \\{(x_n^t, y_n^t)\\}_{n=1}^{N^t}\\), where \\(x_i^t\\) is the \\(i\\)-th sample and \\(y_i^t\\) is corresponding target, respectively. Before training the task, we allocate parameters for the new task to the pruned positions in the base network trained for \\(t-1\\) tasks. As a typical strategy, the allocated parameters \\(W^t\\) are learned from the help of all previous parameter sets [28]. After training, redundant parameters in \\(W^t\\) are pruned, and the unpruned parameters, \\(\\bar{W}^t\\), are retrained. Finally, after learning for \\(t\\) tasks, \\(F(\\cdot; \\mathcal{W}^t)\\) is parameterized with \\(\\mathcal{W}^t = \\{\\bar{W}^1, \\dots, \\bar{W}^t\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Although using previous knowledge can be beneficial in learning new tasks, there is also the possibility of negative knowledge transfer. If earlier tasks are irrelevant to a new task, leveraging their knowledge leads to suboptimal performance of the new task. A recent approach involving the selective use of helpful parameters [15] has been proposed. However, it is based on fixed old knowledge, which limits the model's ability to evolve in response to new information. Consequently, the expressiveness of the model remains constrained by its dependence on immutable prior knowledge. To resolve the issue, we propose a novel framework that transforms old parameters into those absorbing the knowledge of newer tasks. To selectively transform the parameter sets of old tasks \\(\\{\\bar{W}^1,\\dots ,\\bar{W}^{t - 1}\\}\\) for the time \\(t\\), we generate parameters by presenting a sparsity-inducing hypernetwork. The generated ones enhance the previous parameters to fit the new task better. The entire procedure of the proposed method is illustrated in Figure 2." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18963" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.307, + 0.108 + ], + "angle": 0, + "content": "3.2. Evolving Old Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Our approach involves transforming previously learned parameters into more adaptable ones to incorporate knowledge from a new task while retaining the old knowledge. This allows for the evolution of existing knowledge to accommodate new information better. Let us denote the hypernetwork as \\( H(\\cdot; \\mathcal{H}^t) \\) parameterized by \\( \\mathcal{H}^t \\) based on the transformer architecture [42]. Note that \\( \\mathcal{H}^t \\) is initialized with the parameters of the hypernetwork after training the previous task, \\( t - 1 \\). Specifically, the hypernetwork receives the task token \\( e^t \\in \\mathbb{R}^D \\) and the set of layer embeddings \\( \\mathcal{C}_l = \\{c_n \\in \\mathbb{R}^D | 1 \\leq n \\leq N_l\\} \\) and generates the parameter set \\( G_l^t \\), where \\( N_l \\) and \\( l \\in \\{1 \\dots L\\} \\) denote the number of embeddings for layer \\( l \\) and the layer index of the base network, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.401 + ], + "angle": 0, + "content": "To generate \\( G_{l}^{t} \\) the task token \\( e^{t} \\) and layer embedding \\( c_{n} \\) from the set \\( C_l \\) are concatenated to construct the input of the hypernetwork, i.e., \\( z_{n,0}^{t} = \\mathrm{Concat}(e^{t},c_{n})\\in \\mathbb{R}^{2\\times D} \\). The input \\( z_{n,m - 1}^{t} \\) is fed into the \\( m \\)-th task-layer attention block (TLA) as" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.41, + 0.47, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {z} _ {n, m - 1} ^ {t} = z _ {n, m - 1} ^ {t} + \\operatorname {M S A} (\\mathrm {L N} (z _ {n, m - 1} ^ {t})), \\tag {1} \\\\ z _ {n, m} ^ {t} = \\bar {z} _ {n, m - 1} ^ {t} + \\operatorname {M L P} (\\operatorname {L N} (\\bar {z} _ {n, m - 1} ^ {t})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.537 + ], + "angle": 0, + "content": "where \\(m \\in \\{1, \\dots, N_b\\}\\). \\(N_b\\), LN, MSA, and MLP denote the number of TLA blocks, layer normalization, multi-head self-attention, and multi-layer perceptron, respectively. The structure of TLA is illustrated in Figure 3. MSA is defined for the input \\(\\hat{z}_{n,m-1}^t \\triangleq \\mathrm{LN}(z_{n,m-1}^t)\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.546, + 0.47, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M S A} \\left(\\hat {z} _ {n, m - 1} ^ {t}\\right) = W ^ {O} \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {N _ {h}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.575, + 0.47, + 0.623 + ], + "angle": 0, + "content": "\\(W^{O} \\in \\mathbb{R}^{2 \\times D}\\) is a projection matrix and \\(N_{h}\\) denotes the number of heads in the MSA layer, where \\(\\mathrm{head}_{i} \\triangleq \\operatorname{softmax}(Q_{i}K_{i}^{T} / \\sqrt{D / N_{h}})V_{i}\\) and" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.633, + 0.473, + 0.653 + ], + "angle": 0, + "content": "\\[\nQ _ {i} = W _ {Q _ {i}} \\hat {z} _ {n, m - 1} ^ {t}, K _ {i} = W _ {K _ {i}} \\hat {z} _ {n, m - 1} ^ {t}, V _ {i} = W _ {V _ {i}} \\hat {z} _ {n, m - 1} ^ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.663, + 0.47, + 0.694 + ], + "angle": 0, + "content": "where \\( W_{Q_i}, W_{K_i} \\), and \\( W_{V_i} \\) are projection matrices in the \\( i \\)-th head. The parameters for the \\( l \\)-th layer are generated as" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.714, + 0.47, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} G _ {l} ^ {t} = \\operatorname {C o n c a t} \\left(g \\left(\\theta_ {1} ^ {t}\\right), \\dots , g \\left(\\theta_ {N _ {l}} ^ {t}\\right)\\right) \\tag {3} \\\\ = \\operatorname {C o n c a t} \\left(H \\left(z _ {1, 0} ^ {t}; \\mathcal {H} ^ {t}\\right), \\dots , H \\left(z _ {N _ {l}, 0} ^ {t}; \\mathcal {H} ^ {t}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.763, + 0.469, + 0.809 + ], + "angle": 0, + "content": "where \\(g(\\cdot)\\) denotes the fully connected layer followed by a sigmoid function. \\(\\theta_{n}^{t}\\in \\mathbb{R}^{D}\\) is the first row of \\(z_{n,N_b}^t\\) , which acts as the final embedding in a transformer architecture [8]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.809, + 0.47, + 0.873 + ], + "angle": 0, + "content": "To refine old knowledge to align with the knowledge of the \\(t\\)-th task, let us denote the set of old parameters in the \\(l\\)-th layer as \\(\\bar{W}_l^{1:t-1}\\). We obtain the evolved parameters \\(\\hat{W}_l^{1:t-1}\\) by fusing \\(\\bar{W}_l^{1:t-1}\\) and \\(G_l^t\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.882, + 0.47, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\hat {W} _ {l} ^ {1: t - 1} = \\bar {W} _ {l} ^ {1: t - 1} \\cdot G _ {l} ^ {t}, \\tag {4}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.885, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.182, + 0.892, + 0.225 + ], + "angle": 0, + "content": "Figure 3. The task-layer attention block (TLA) consists of a layer normalization, multi-head self-attention (MSA), and MLP with a single hidden layer. \\( z_{n,N_b}^t \\) is generated by \\( N_{b} \\) stacked TLAs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.326 + ], + "angle": 0, + "content": "where we define \\(\\cdot\\) as the element-wise multiplication. The base network accommodates the evolved sets of parameters for \\(t - 1\\) tasks together with the \\(t\\)-th parameter set as \\(\\hat{\\mathcal{W}}^t = \\{\\hat{W}^1, \\dots, \\hat{W}^{t-1}, W^t\\}\\). Finally, the loss function to learn the \\(t\\)-th task using the evolved parameter sets becomes" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.337, + 0.892, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t a s k}} = - \\sum_ {n = 1} ^ {N ^ {t}} y _ {n} ^ {t} \\log \\left(F \\left(x _ {n} ^ {t}; \\hat {\\mathcal {W}} ^ {t}\\right)\\right)). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.39, + 0.891, + 0.422 + ], + "angle": 0, + "content": "Another loss function to generate parameters from \\(H(\\cdot ;\\mathcal{H}^t)\\) similar to those learned in the previous task is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.432, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\sum_ {k = 1} ^ {t - 1} \\sum_ {l = 1} ^ {L} \\| \\bar {G} _ {l} ^ {k} - G _ {l} ^ {k} \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.485, + 0.891, + 0.516 + ], + "angle": 0, + "content": "where \\(\\bar{G}_l^k\\) represents the set parameters generated from the hypernetwork after training the task \\(t - 1\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.767, + 0.541 + ], + "angle": 0, + "content": "3.3. Sparsity-Inducing Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.791 + ], + "angle": 0, + "content": "To renew the network without redundant parameters, we propose to generate sparse parameters in the hypernetwork. As \\( G_{l}^{t} \\) is generated through the sigmoid function, they can be interpreted as the likelihood that each element of the old set of parameters is selected. Following the practice in [41], we minimize \\( \\sum_{l=1}^{L} -\\log(1 - G_{l}^{t}) \\) to achieve sparsity in \\( G_{l}^{t} \\). However, to prevent \\( G_{l}^{t} \\) from becoming overly sparse, we ensure that parameters significantly impacting the loss remain survived. Inspired by the work in [24], we evaluate the impact of individual elements in \\( G_{l}^{t1} \\) to the loss. We compute the set of all entries of the score \\( S_{l}^{t} \\) by comparing the network losses computed with and without the inclusion of parameters in the learning process. To estimate the effect of removing a parameter on the loss function, we apply a Taylor expansion approximation of the network loss \\( \\mathcal{L} \\) for \\( N \\) samples. We obtain the \\( k \\)-th element of the score, \\( s_{k} \\), as" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.799, + 0.891, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s _ {k} = \\mathcal {L} \\left(\\vec {G} _ {l} ^ {t} - M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t}\\right) - \\mathcal {L} \\left(\\vec {G} _ {l} ^ {t}\\right) \\tag {7} \\\\ \\simeq - M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t} \\cdot \\nabla_ {\\vec {G} _ {l} ^ {t}} \\mathcal {L} + \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\frac {\\partial \\mathcal {L} _ {n}}{\\partial M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t}}\\right) ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "1We denote the vectorized form of \\( G_{l}^{t} \\) as \\( \\vec{G}_{l}^{t} \\) to represent individual entries." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18964" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "where \\( M_{\\backslash (k)} \\) is an all-zero vector except that the \\( k \\)-th element is 1, and \\( \\mathcal{L}_n \\) is the loss of the \\( n \\)-th example. We introduce an additional term minimizing the discrepancy between the scores of the complete set of parameters (i.e., the original scores) and the scores of the sparse set of parameters produced as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.194, + 0.47, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s p} = \\sum_ {l = 1} ^ {L} - \\log (1 - G _ {l} ^ {t}) + \\left\\| S _ {l} ^ {t} \\cdot \\left(1 - \\mathbb {1} (\\vec {G} _ {l} ^ {t} \\neq 0)\\right) \\right\\| _ {1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.47, + 0.277 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_1 \\) and \\( \\mathbb{1}(\\cdot) \\) denote the \\( l_1 \\)-norm and the indicator function, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.385 + ], + "angle": 0, + "content": "Total Loss. The proposed framework learns the \\(t\\)-th task with the three loss functions. To learn the \\(t\\)-th task, we minimize the cross-entropy loss \\(\\mathcal{L}_{task}\\) by updating \\(W^t\\) in the base network. Also, \\(\\mathcal{H}^t\\), \\(e^t\\), and \\(\\{C_l\\}_{l=1}^L\\) are updated from two losses, \\(\\mathcal{L}_{task}\\) and \\(\\mathcal{L}_{sp}\\). Last, we update \\(\\mathcal{H}^t\\) and \\(\\{C_l\\}_{l=1}^L\\) by minimizing \\(\\mathcal{L}_{reg}\\) to preserve the previously generated parameters. The total loss is" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.395, + 0.47, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\min _ {W ^ {t}, \\mathcal {H} ^ {t}, e ^ {t}, \\{C _ {l} \\} _ {l = 1} ^ {L}} \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {s p} \\mathcal {L} _ {s p} + \\lambda_ {r e g} \\mathcal {L} _ {r e g}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.452, + 0.447 + ], + "angle": 0, + "content": "where \\(\\lambda_{sp}\\) and \\(\\lambda_{reg}\\) are parameters to balance the losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.473, + 0.675 + ], + "angle": 0, + "content": "Scalability. A hypernetwork is designed to generate parameters for every layer of the base network, and the parameters were composed of compact segments of a consistent size [13, 21]. While it can generate parameters for all layers, this approach requires a substantial quantity of layer embeddings, particularly for deeper layers. To mitigate this, we adjust the scale of the segments for each layer depending on the number of channels. However, it incurs a memory overhead stemming from the substantial dimensions of the heads. To address this, we apply a matrix decomposition technique [39] to generate two reduced sets of low-rank \\( r \\). By doing so, we achieve a reduction in memory requirements for layer embeddings, accompanied by a marginal increase in hypernetwork parameters. For detailed results, please refer to Section 4.5.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.687, + 0.21, + 0.704 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.189, + 0.726 + ], + "angle": 0, + "content": "4.1. Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We applied the proposed method, GrowBrain, to image classification and video action recognition problems under class incremental and task-incremental learning scenarios. Image classification. We conducted experiments using ImageNet [6] and five fine-grained datasets including CUBS [45], Stanford Cars [18], Flowers [31], Wikiart [36], and Sketch [11], where each dataset is considered a task. We call the datasets ImageNet+FG for short. In addition, we demonstrated the proposed method with the CIFAR-100. We divided the dataset into subsets of classes based on the random order provided by [35, 9] and considered" + }, + { + "type": "table_caption", + "bbox": [ + 0.59, + 0.088, + 0.8, + 0.1 + ], + "angle": 0, + "content": "Table 1. Datasets used in this work." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.1, + 0.882, + 0.242 + ], + "angle": 0, + "content": "
Image Dataset# Train Images# Test Images# Classes
ImageNet [6]1,287,16750,0001,000
CUBS [45]5,9945,794200
Stanford Cars [18]8,1448,041196
Flowers [31]2,0406,149102
Wikiart [36]42,12910,628195
Sketch [11]16,0004,000250
CIFAR-100 [19]50,00010,000100
Video Dataset# Train Videos# Test Videos# Classes
ActivityNet [3]10,0244,926200
UCF-101 [40]9,2802,720101
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.253, + 0.836, + 0.266 + ], + "angle": 0, + "content": "Table 2. The specification of the hypernetwork." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.266, + 0.882, + 0.343 + ], + "angle": 0, + "content": "
Base networkHypernetwork
Embed dim (D)Hidden dim in MLP# Heads (Nh)# Blocks (Nb)
ResNet-506412887
ResNet-34
ResNet-18326485
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.459 + ], + "angle": 0, + "content": "each subset a task. Video action recognition. We used a trimmed version of ActivityNet [3] and UCF-101 [40] video datasets. To conduct experiments, we divided each dataset into subsets containing randomly selected classes provided by [43] and treated each subset as a task. We summarize the datasets as mentioned above in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.715, + 0.485 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Image classification. ImageNet+FG and CIFAR-100 were resized to \\(224 \\times 224\\) and \\(32 \\times 32\\) pixels, respectively. We applied random cropping and horizontal flip as augmentation. For those datasets, we used the ResNet-50 [14] and ResNet-18 as the base networks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Video action recognition. We used TSN [46] with the ResNet-34 backbone network for video action recognition. Each video was split into three segments of equal duration, and a frame was randomly selected from each segment. The action class prediction was performed by aggregating the predictions of the selected frames with a consensus function. We used an average of predictions as a consensus function [46]. The video datasets were augmented with random cropping and horizontal flipping following [46]. We resized each frame of ActivityNet and UCF-101 to \\(224 \\times 224\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Table 2 reports the base networks and corresponding hypernetworks, which include the embedding dimension, the hidden dimension of the MLP, the number of heads, and the number of TLA blocks. We set the rank \\( r \\) for matrix decomposition as 7. We trained the base and hypernetwork using stochastic gradient descent. In all our experiments, we set \\( \\lambda_{reg} \\) to 1 and \\( \\lambda_{sp} \\) to 0.1, except ImageNet+FG where we set \\( \\lambda_{sp} \\) to 1. We report the average results over five independent runs for all the experiments. To enable class-incremental learning in scenarios where the task identity is unknown, the proposed method first predicts the task index and then conducts classification. Following [44], we predict the task" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18965" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.114 + ], + "angle": 0, + "content": "Table 3. Performance of the class-incremental learning scenario using ImageNet and five fine-grained datasets (ImageNet+FG), where the average accuracy was measured after training each task. # P represents the average number of parameters \\((\\times 10^{6})\\) to perform the tasks." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.877, + 0.248 + ], + "angle": 0, + "content": "
MethodImageNetCUBSStanford CarsFlowersWikiartSketch
Acc#PAcc#PAcc#PAcc#PAcc#PAcc#P
EWC-On76.123.431.623.428.623.429.223.423.123.411.323.4
LwF76.123.468.423.458.223.455.823.449.223.428.423.4
iCaRL76.123.470.223.461.523.452.723.448.323.442.323.4
PackNet75.712.562.813.365.214.066.514.654.315.245.315.7
Piggyback76.123.460.122.754.822.256.422.049.721.747.421.4
WSN74.712.559.612.542.212.529.212.531.012.525.012.5
H275.712.569.513.067.312.968.413.061.812.948.712.6
GrowBrain75.712.569.211.069.310.371.39.864.39.660.19.4
" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.259, + 0.341, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.259, + 0.614, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.26, + 0.884, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.473, + 0.89, + 0.501 + ], + "angle": 0, + "content": "Figure 4. Accuracy of class-incremental learning for image classification on ImageNet+FG (left), video action recognition on UCF-101 (middle) and ActivityNet (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.556 + ], + "angle": 0, + "content": "index using the network with the least entropy that has been learned thus far." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.564, + 0.321, + 0.581 + ], + "angle": 0, + "content": "4.3. Class-incremental learning" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.268, + 0.603 + ], + "angle": 0, + "content": "4.3.1 Image classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.746 + ], + "angle": 0, + "content": "We evaluated the proposed method of evolving the previous knowledge to the class-incremental learning scenario using ImageNet+FG. We compared with regularization methods, EWC-On [37] and LwF [22], and data rehearsal methods, iCaRL [35], and parameter isolation methods, PackNet [28], Piggyback [27], WSN [16], and \\(\\mathrm{H}^2\\) [15]. The data rehearsal methods stored 1,000 raw instances (taking 150 \\(\\mathrm{MB}^2\\)). We applied the same task prediction mechanism as ours to the comparing methods, excluding iCaRL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Table 3 shows the class-incremental learning results for ImageNet and five fine-grained datasets. Overall, the proposed method outperforms competitors (more detailed results are shown on the left of Figure 4). The regularization approaches, EwC-On and LwF, give unsatisfactory results with performance gaps of \\(48.8\\%\\) and \\(31.7\\%\\) compared to ours, respectively. LwF shows a \\(20.8\\%\\) accuracy drop before and after learning Sketch, which is the most significant" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.526, + 0.892, + 0.738 + ], + "angle": 0, + "content": "among the comparing methods. The data rehearsal method, iCaRL, shows a \\(17.1\\%\\) performance drop from GrowBrain despite storing 1,000 old samples. PackNet shows large performance drops of \\(12.2\\%\\) and \\(9.0\\%\\) after training on WikiArt and Sketch, respectively. This is because negative interference occurs when using all previous sets of parameters optimized for different task domains. The methods of associating prior knowledge, Piggyback, WSN, and \\(\\mathrm{H}^2\\), show notable performance drop (\\(12.7\\%\\), \\(35.1\\%\\), and \\(11.4\\%\\), respectively) compared to the proposed method that improves the previous knowledge from the new task. The results demonstrate that the proposed approach surpasses a wide range of continual learning methods, confirming its superiority." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.724, + 0.772 + ], + "angle": 0, + "content": "4.3.2 Video action recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We applied our method and the competitors to the video action recognition tasks using the benchmark sets proposed for continual learning [43]. For this experiment, iCaRL was excluded from the comparison since it demands approximately 2TB and 52GB of memory for replay [43] on ActivityNet and UCF-101, respectively. Since Piggyback requires a pre-trained backbone, we pre-trained the backbone model [14] using the validation set of UCF-101 for video action" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "2 We calculated the rehearsal memory assuming that each image (or frame of video) has \\( 224 \\times 224 \\times 3 \\) pixels." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18966" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.114 + ], + "angle": 0, + "content": "Table 4. Results of the class-incremental learning scenario using the trimmed version of UCF-101 and ActivityNet, where the Avg and Last accuracies for the time steps 5 and 10." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.115, + 0.807, + 0.257 + ], + "angle": 0, + "content": "
MethodUCF-101ActivityNet
T5T10T5T10
AvgLast# PAvgLast# PAvgLast# PAvgLast# P
EWC-On48.725.821.232.713.921.238.420.721.226.612.421.2
LwF47.027.721.233.415.421.244.129.121.232.718.421.2
PackNet59.354.211.552.540.415.242.430.211.532.419.015.2
Piggyback65.451.720.453.536.520.422.212.720.515.67.520.4
WSN69.158.810.161.149.710.5------
H263.654.88.356.947.011.437.422.53.426.813.32.1
GrowBrain79.372.27.273.564.78.942.829.911.533.220.615.2
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.47, + 0.342 + ], + "angle": 0, + "content": "recognition experiments. We report the \\(Avg\\) and Last accuracies [9] for the time steps 5 and 10. Figure 4 shows the Last accuracy measured on the UCF-101 and ActivityNet datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.358, + 0.47, + 0.584 + ], + "angle": 0, + "content": "UCF-101. The results of the class-incremental learning scenario using the UCF-101 dataset are reported in Table 4 (left). The aspects of the experimental results on UCF-101 of the compared approaches are similar to those of image classification. Despite utilizing a large number of parameters, regularization approaches, EWC-On and LwF, give unsatisfactory performance compared to parameter isolation approaches. PackNet, which uses all parameters from previous tasks, performs less than other parameter isolation approaches, WSN and \\(\\mathrm{H}^2\\), using selectively chosen parameters associated with previous knowledge. The proposed method shows the highest performance compared to other comparison methods. Ours outperforms WSN, \\(\\mathrm{H}^2\\), PackNet, and Piggyback with the significant accuracy gap on the Last measure from \\(15.0\\%\\) to \\(28.2\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.47, + 0.901 + ], + "angle": 0, + "content": "ActivityNet. Table 4 (right) reports the results on ActivityNet. WSN was excluded from the experiment because it did not give promising results due to the lack of data. Overall, the results of ActivityNet show a different pattern compared to those of UCF-101. The regularization method, LwF, outperforms the parameter isolation method, PackNet, with \\(1.7\\%\\) and \\(0.3\\%\\) Avg accuracy gap measured after training the 5-th and 10-th tasks. The proposed approach performs better than Piggyback, PackNet, and \\(\\mathrm{H}^2\\) by large margins of \\(13.1\\%\\), \\(1.6\\%\\), and \\(7.3\\%\\), respectively, in terms of the Last accuracy measured after training 10-th task. The proposed method outperforms other competitors for image classification and video action recognition problems. In contrast to the UCF-101 dataset, where the proposal can acquire new knowledge by using a small number of parameters from previous tasks and achieve excellent results, the ActivityNet dataset requires previous knowledge to effectively train the current task due to its inadequate data (the number of videos per class is only half compared to UCF-101)." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.279, + 0.892, + 0.319 + ], + "angle": 0, + "content": "Table 5. Task-incremental learning scenarios using video action recognition and image classification datasets, where average accuracy and parameter consumption are reported." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.457 + ], + "angle": 0, + "content": "
MethodImage classificationVideo action recognition
ImageNet+FGCIFAR-100UCF-101ActivityNet
Acc#PAcc#PAcc#PAcc#P
EWC-On55.323.444.511.165.821.261.521.2
LwF53.623.439.711.171.221.265.821.2
PackNet80.215.782.110.492.215.265.315.2
Piggyback82.221.475.610.586.620.442.120.4
WSN61.712.584.65.692.710.5--
H282.812.685.91.194.511.444.02.1
GrowBrain83.09.486.82.596.78.966.915.2
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.74, + 0.5 + ], + "angle": 0, + "content": "4.4. Task-incremental learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Additionally, we conducted experiments on the task-incremental learning scenario where task-id is given during the inference step. In this scenario, we used ImageNet+FG and CIFAR-100 as the image classification datasets. Also, we used ActivityNet and UCF-101 for video action recognition. We divided CIFAR-100 into 10 tasks based on [35]. We followed the same experimental setting as in Tables 3 and 4. For the experiments of Piggyback on CIFAR-100, the backbone network was pre-trained on a randomly selected subset (40%) of the CIFAR-100 training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Image classification. Table 5 reports the results of task-incremental learning conducted on ImageNet+FG and CIFAR-100. Even if the regularization methods, EwC-On and LwF, try to maintain previous knowledge, they perform significantly less than the parameter isolation methods. GrowBrain consistently achieves high performance compared to other parameter isolation methods, PackNet, Piggyback, WSN, and \\(\\mathrm{H}^2\\), in both image classification experiments (e.g., \\(0.2\\%\\) and \\(0.9\\%\\) performance gap on ImageNet+FG and CIFAR-100 with the most competitive method, \\(\\mathrm{H}^2\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video action recognition. Table 5 reports the results of task-incremental learning conducted on UCF-101 and ActivityNet. Overall, the results are similar to those of class-incremental learning scenarios. The regularization method, LwF, achieves higher performance than PackNet on Activ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18967" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.142 + ], + "angle": 0, + "content": "Table 6. Ablation study of the proposed method for image classification and video action recognition tasks, where average accuracy, the number of parameters for the base network and hypernetwork (in millions), and the number of layer embeddings are reported." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.143, + 0.469, + 0.212 + ], + "angle": 0, + "content": "
MethodImageNet+FGUCF-101
Acc# (F + H)# cnAcc# (F + H)# cn
Ours60.19.4 + 2.813364.78.9 + 2.858
w/o transform [28]45.315.7 + 0040.415.2 + 00
w/o sparsity58.115.7 + 2.813364.014.1 + 2.858
w/o scalability59.815.7 + 2.9326264.711.0 + 2.9614
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.469, + 0.299 + ], + "angle": 0, + "content": "ityNet. The proposed method outperforms both parameter isolation and regularization methods. Notably, our method outperforms WSN, PackNet, and \\(\\mathrm{H}^2\\) with a performance gap of \\(4\\%\\), \\(4.5\\%\\), and \\(2.2\\%\\), respectively, on UCF-101." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.309, + 0.18, + 0.325 + ], + "angle": 0, + "content": "4.5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.333, + 0.235, + 0.348 + ], + "angle": 0, + "content": "4.5.1 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.469, + 0.522 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our method, we conducted an ablation study. We compared ours by removing transformation of previous knowledge [28], sparsity-inducing generation \\((\\lambda_{sp} = 0)\\), and scalability of hypernetwork, i.e., channel adaptive segments and matrix decomposition. We applied these methods to class-incremental learning scenarios using the ImageNet+FG and the UCF-101 datasets. We report the Last accuracy measured after training all tasks, the average parameter consumption of the base network \\((F)\\) and hypernetwork \\((H)\\) to perform each task, and the number of layer embeddings \\(c_{n}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.766 + ], + "angle": 0, + "content": "Table 6 reports the results of the ablation study. Our method outperforms that without transformation by performance gaps of \\(14.8\\%\\) and \\(24.3\\%\\) for those datasets, emphasizing that the evolution of prior knowledge is critical in enhancing performance. For video action recognition using UCF-101, the method without sparsity requires more parameters and performs less than the proposed method. The proposed method for ImageNet+FG achieves higher performance while taking fewer parameters than the nonsparse alternative. The reason is that sparsity-inducing hypernetwork prevents the generation of redundant parameters and regularizes the generated parameters to achieve generalization performance. The method without scalability requires 24 and 10.5 times larger layer embeddings on ImageNet+FG and UCF-101 than the proposed method, respectively, without accompanying performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.226, + 0.802 + ], + "angle": 0, + "content": "4.5.2 Impact of \\(\\lambda_{sp}\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We further conducted a study investigating the impact of \\(\\lambda_{sp}\\), which regulates the sparsity of generated parameters. We experimented with varying values of \\(\\lambda_{sp}\\) in the ImageNet+FG experiment under the class-incremental scenario using the same setup presented in Table 3. Figure 5 shows the parameter consumption (left) and average ac" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.695, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.089, + 0.888, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.2, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Figure 5. Parameter consumption and average accuracy of the proposed method based on varying \\(\\lambda_{sp}\\) in class-incremental learning for ImageNet+FG." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.388 + ], + "angle": 0, + "content": "curacy (right) with respect to different \\(\\lambda_{sp}\\) after training all tasks. Ours with \\(\\lambda_{sp} = 0.1\\) uses a similar amount of parameters as the baseline [28] but outperforms the competing approach by leveraging enhanced prior knowledge. Notably, after training all tasks, \\(\\lambda_{sp} = 1.0\\) takes \\(9.4\\mathrm{M}\\) parameters while showing \\(0.7\\%\\) performance improvement compared to \\(\\lambda_{sp} = 0.1\\) consuming \\(15.7\\mathrm{M}\\) parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.619, + 0.438 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.728 + ], + "angle": 0, + "content": "We have proposed a novel approach, GrowBrain, to transform learned old knowledge to absorb the information of newer tasks in parameter isolation-based continual learning. GrowBrain reinforces the previous knowledge with a transformer-based hypernetwork that generates optimized parameters for new tasks while eliminating redundant parameters that show negligible impact on the loss. The proposed method overcomes the limitation of using fixed old knowledge that can constrain the ability to capture novel representations while taking fewer parameters. In the experiments, we have evaluated the proposal compared to a variety of continual learning methods in class- and task-incremental learning scenarios for image classification and video action recognition. Experimental results show that the proposed approach of leveraging transformed prior knowledge generates parameters adapted to the new task without losing the prior knowledge, resulting in significant performance improvements over its strong competitors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT2002-05, Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (2022-0-01341, Artificial Intelligence Graduate School Program(Chung-Ang University)), and BK21 FOUR (Fostering Outstanding Universities for Research) Program funded by Ministry of Education of Korea (No.I22SS7609062)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18968" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3931-3940, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European Conference on Computer Vision, pages 144-161. Springer, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.339 + ], + "angle": 0, + "content": "[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[4] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 556-572. Springer, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[5] Arslan Chaudhry, Ranzato Marc'Aurelio, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In 7th International Conference on Learning Representations, ICLR 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.591 + ], + "angle": 0, + "content": "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.677, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[9] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[10] Benjamin Ehret, Christian Henning, Maria Cervera, Alexander Meulemans, Johannes Von Oswald, and Benjamin F Grewe. Continual learning in recurrent neural networks. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[11] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Transactions on Graphics, 31(4):1-10, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[15] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, pages 519-535. Springer, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[16] Haeyong Kang, Rusty John Lloyd Mina, Sultan Rizky Hikmawan Madjid, Jaehong Yoon, Mark Hasegawa-Johnson, Sung Ju Hwang, and Chang D Yoo. Forget-free continual learning with winning subnetworks. In International Conference on Machine Learning, pages 10734-10750. PMLR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[17] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America, 114(13):3521-3526, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561. IEEE, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. Advances in Neural Information Processing Systems, 30:4652-4662, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[21] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. DHP: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. In European Conference on Computer Vision, pages 614-629. Springer, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[24] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group fisher pruning for practical network compression. In International Conference on Machine Learning, pages 7021-7032. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2544-2553, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6470-6479, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18969" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[27] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In Proceedings of the European Conference on Computer Vision, pages 67-82, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[28] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[29] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, volume 24, pages 109–165. Elsevier, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[30] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.471, + 0.362 + ], + "angle": 0, + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.364, + 0.471, + 0.433 + ], + "angle": 0, + "content": "[32] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.435, + 0.471, + 0.49 + ], + "angle": 0, + "content": "[33] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13698-13707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.492, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[34] Dripta S Raychaudhuri, Yumin Suh, Samuel Schulter, Xiang Yu, Masoud Faraki, Amit K Roy-Chowdhury, and Manmohan Chandraker. Controllable dynamic multi-task architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.471, + 0.646 + ], + "angle": 0, + "content": "[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[36] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. International Journal for Digital Art History, (2), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International Conference on Machine Learning, pages 4528-4537. PMLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[38] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 2994-3003, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[39] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[41] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. AdaShare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[43] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vCLIMB: A novel video class incremental learning benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19035-19044, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[44] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. In 8th International Conference on Learning Representations, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.503, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[47] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3987-3995, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18970" + } + ] +] \ No newline at end of file diff --git a/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_origin.pdf b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a650560d409a0dccf1c496bbbc19314eddbc49de --- /dev/null +++ b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/d35af5fa-bfe4-478d-92d6-454b21b1d4e0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d08db2c33105fb5a85b11849d395056b7c2030e7bd29210929fac593dbe7b801 +size 2465340 diff --git a/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/full.md b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..72bf6f1b6c003254fe2d09a348d52b499d547f2e --- /dev/null +++ b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/full.md @@ -0,0 +1,294 @@ +# Growing a Brain with Sparsity-Inducing Generation for Continual Learning + +Hyundong Jin1 Gyeong-hyeon Kim1 Chanho Ahn2* Eunwoo Kim1 +1School of Computer Science and Engineering, Chung-Ang University +2Samsung Advanced Institute of Technology (SAIT) + +{jude0316,leonardkkh,eunwoo}@cau.ac.kr chanho.ahn@samsung.com + +# Abstract + +Deep neural networks suffer from catastrophic forgetting in continual learning, where they tend to lose information about previously learned tasks when optimizing a new incoming task. Recent strategies isolate the important parameters for previous tasks to retain old knowledge while learning the new task. However, using the fixed old knowledge might act as an obstacle to capturing novel representations. To overcome this limitation, we propose a framework that evolves the previously allocated parameters by absorbing the knowledge of the new task. The approach performs under two different networks. The base network learns knowledge of sequential tasks, and the sparsity-inducing hypernetwork generates parameters for each time step for evolving old knowledge. The generated parameters transform old parameters of the base network to reflect the new knowledge. We design the hypernetwork to generate sparse parameters conditional to the task-specific information and the structural information of the base network. We evaluate the proposed approach on class-incremental and task-incremental learning scenarios for image classification and video action recognition tasks. Experimental results show that the proposed method consistently outperforms a large variety of continual learning approaches for those scenarios by evolving old knowledge. + +# 1. Introduction + +The ability of deep learning to incrementally accumulate knowledge for sequentially incoming tasks is an important capability required for modern AI algorithms. However, deep neural networks tend to lose previously acquired knowledge while learning new data. This limitation of deep neural networks, called catastrophic forgetting, has been investigated by previous studies [29, 17]. Catastrophic forgetting significantly degrades performance when data from previous tasks (task-incremental) or already-learned class + +data (class-incremental) cannot be accessed. Under those incremental learning scenarios, continual learning methods have toiled to preserve the previous knowledge. + +Various approaches have been proposed for continual learning and are categorized into three types: regularization [2, 4, 20, 17, 48], data rehearsal [5, 26, 32, 35, 38, 47], and parameter isolation [1, 30, 28]. The conventional continual learning methods introduced regularization approaches to suppress changes in important parameters [48, 2] or input-output mappings from a network trained up to previous tasks [22, 7]. The network is trained in a way that contributable parameters to the performance of the previously trained model remain unchanged. On the other hand, data rehearsal methods [25, 35] address catastrophic forgetting by accessing old data. The old data can be obtained by storing some data from old tasks [35, 47] or leveraging a generative network [38]. However, the privacy of data and the difficulty of learning generative models limit their versatility. Recently, parameter isolation approaches [28, 30, 15] construct a disjoint set of parameters for each task, preventing it from mixing old and new knowledge. These methods learn a new set of parameters by associating all [28] or some helpful old knowledge [15]. + +Unlike other approaches, the parameter isolation approach is designed to prevent forgetting without requiring access to old data by disallowing updates to previous sets of parameters. Despite this advantage, these methods can limit the ability to learn new tasks by relying on parameters that were learned for older tasks and remain unchanged. We observe that this limitation arises when there is a low correlation between the old and new data. This raises the question: In parameter isolation-based continual learning, is it possible to transform the knowledge from old tasks to fit newer tasks better? + +In this paper, we introduce a novel method for transforming prior knowledge into a form that can aid in learning new tasks. The approach involves reframing the parameter selection problem, which previous works [15] have addressed, as a generation problem. The proposed framework emphasizes a subset of pre-trained knowledge, selecting and highlight- + +ing it for use in learning new tasks. Specifically, the framework consists of two networks: base network and hypernetwork. The base network constructs disjoint parameter sets for sequential tasks. The proposed transformer-based [42] hypernetwork generates a small set of influential parameters that significantly impact the loss. The impact of each element of generated parameters is evaluated by the difference in loss computed with and without its inclusion in the learning process. Also, generated parameters are conditioned on the information of the current task (task token) and the layer of the base network (layer embedding). The generated parameters transform the previous set of parameters in the base network to further accommodate new knowledge. In addition, compared to existing hypernetworks that generate a small set of parameters and apply it to all the layers [13, 44], the proposed hypernetwork adaptively generates parameters according to the layer configurations defined in the base network. Through this strategy, the quantity of layer embeddings for ResNet-50 is reduced by a substantial factor of 24. Furthermore, by leveraging a matrix decomposition technique to represent generated parameters compactly, we achieve a noteworthy reduction in the parameters necessary for the hypernetwork. + +We evaluate the effectiveness of the proposed method in both class-incremental and task-incremental learning scenarios using classification benchmark datasets. Furthermore, we demonstrate the efficacy of the proposed method [28, 15] in these scenarios by applying it to video action recognition datasets [19]. Experimental results show that the proposed method performs better than its competitors for most learning scenarios while introducing minor memory overhead from the hypernetwork. The contributions of our work are mainly three-fold: (1) We present a novel approach to evolving previous knowledge in continual learning, aimed at reflecting the knowledge of new tasks. (2) We propose a transformer-based hypernetwork that generates customized parameters for new tasks while suppressing the generation of redundant parameters that have little impact on the loss. (3) Experimental results on diverse benchmarks for video action recognition and image classification demonstrate that ours outperforms its competitors with lower network capacity requirements to perform tasks. + +# 2. Related Work + +Continual Learning. Generally, continual learning methods can be classified into three types. Regularization methods [23, 7, 17, 48, 4, 2, 20] add a penalty term that restrains the previously learned knowledge from being updated when learning a new task. In order not to hurt previous knowledge, a new network is introduced to mimic the features obtained from the old network [22, 7]. While other regularization techniques, such as those based on the Fisher information matrix [17], change of loss [48], and variation of + +![](images/2d8c0b91120742ed46a2160189708b9340e73b523ac00cd4bc347e6ca22661e5.jpg) +Figure 1. An illustration of the proposed method (right) and baseline methods (middle, left) for the third (new) task. The figure represents task-specific allocated parameters, represented by task numbers. Gray numbers represent the sets of discarded parameters from a selection method. The baseline methods do not update old knowledge (red boxes) while updating a new task (blue boxes). The proposed method transforms the old sets of parameters, represented by $+$ , to fit the new task. Best viewed in color. + +prediction [2], suppress the update of crucial parameters for each task, they experience significant performance decline as the number of tasks grows [12]. Data rehearsal methods [5, 26, 35, 33, 25, 38] rely on the availability of previous task data, which is achieved through selective storage of data [35, 47, 25]. These methods employ a few previous samples jointly with a new dataset [25, 38] or constraining the gradient based on the angle between that obtained from stored samples and the gradient obtained from the new samples [5, 26]. These methods incur a data imbalance problem due to a small number of stored previous tasks compared to a large amount of new task data. Parameter isolation methods [28, 1, 30, 15, 16] involve constructing disjoint sets of parameters for old tasks, respectively, and updating another set of parameters for a newer task by associating it with old parameters during training. This is usually done by using all previous sets of parameters [28], using a learnable mask to select some previous set of parameters [27], jointly selecting a location to allocate a new set of parameters [16], or searching a subnetwork with coarse-to-fine task association [15]. + +However, this approach can impede efficient learning because the parameters optimized for old tasks remain the same when learning newer tasks, meaning they need to absorb the information from the newer tasks. In contrast, the proposed method allows the old knowledge to grow by incorporating the knowledge from new tasks. The proposed method and its baselines are illustrated in Figure 1. + +Hypernetworks. Hypernetworks generate parameters for a target network [13, 44, 39, 21, 34]. A recent work [34] proposes a method for generating parameters for layers with few parameters, such as activation of ResNet blocks or batch normalization layers. However, generating parameters for convolutional and linear layers that take most of the parameters in the base network can be challenging because it is difficult to represent the entire set of parameters from a + +![](images/c5d4fe3df336ce739587dcec0b7ab5dd0688a6e13c1fa1176f88f1d65470b0fd.jpg) +Figure 2. The proposed method for transforming the parameters of the $l$ -th layer of the base network $F(\cdot)$ when learning the $t$ -th task. The numbers in the grid indicate the indices of tasks. The hypernetwork $H(\cdot)$ generates parameters using the task-layer attention block (TLA), which discovers the relationship between the task token and layer embedding. The sparse parameters are generated while suppressing the generation of the redundant ones by evaluating the impact of parameters on the loss. The previous sets of parameters of the $l$ -th layer are transformed by element-wise multiplication with $G_{l}^{t}$ . The transformed parameters are combined with the current parameter set through the union operation $(\cup)$ and released as a complete set of parameters to learn new knowledge. Best viewed in color. + +hypernetwork. To avoid this, existing works generate small sets of parameters and concatenates them [13, 21]. Consequently, to generate a base network of large scale, allowing a large number of input embeddings corresponding to the network is inevitable due to the repetitive generation mechanism. In contrast to the approach, the parameters are generated according to the dimensions defined in the layers of the base network in our method. Note that as these dimensions expand, the corresponding demands on hypernetwork parameters also escalate. To address this, we employ a matrix decomposition technique for a compact representation of the required parameters in the hypernetwork. + +Recently, in the continual learning literature, [44] and [10] learn a hypernetwork to generate parameters with task-specific embeddings for new tasks while preventing changes to previously generated parameters for previous tasks. The approaches produce dense parameters for the base network through a task-conditional hypernetwork. In contrast, the proposed transformer-based hypernetwork generates sparse parameters by measuring the effect of individual parameters on the loss. Furthermore, our method generates parameters to evolve previous knowledge to reflect new knowledge, whereas the aforementioned approaches generate parameters directly for the base network. + +# 3. GrowBrain + +# 3.1. Motivation + +The base network $F(\cdot)$ constructs a disjoint set of parameters for each task under an iterative train-prune + +retrain framework [28]. At a time step $t$ , the base network is learned for the $t$ -th task with the dataset $T^t = \{(x_n^t, y_n^t)\}_{n=1}^{N^t}$ , where $x_i^t$ is the $i$ -th sample and $y_i^t$ is corresponding target, respectively. Before training the task, we allocate parameters for the new task to the pruned positions in the base network trained for $t-1$ tasks. As a typical strategy, the allocated parameters $W^t$ are learned from the help of all previous parameter sets [28]. After training, redundant parameters in $W^t$ are pruned, and the unpruned parameters, $\bar{W}^t$ , are retrained. Finally, after learning for $t$ tasks, $F(\cdot; \mathcal{W}^t)$ is parameterized with $\mathcal{W}^t = \{\bar{W}^1, \dots, \bar{W}^t\}$ . + +Although using previous knowledge can be beneficial in learning new tasks, there is also the possibility of negative knowledge transfer. If earlier tasks are irrelevant to a new task, leveraging their knowledge leads to suboptimal performance of the new task. A recent approach involving the selective use of helpful parameters [15] has been proposed. However, it is based on fixed old knowledge, which limits the model's ability to evolve in response to new information. Consequently, the expressiveness of the model remains constrained by its dependence on immutable prior knowledge. To resolve the issue, we propose a novel framework that transforms old parameters into those absorbing the knowledge of newer tasks. To selectively transform the parameter sets of old tasks $\{\bar{W}^1,\dots ,\bar{W}^{t - 1}\}$ for the time $t$ , we generate parameters by presenting a sparsity-inducing hypernetwork. The generated ones enhance the previous parameters to fit the new task better. The entire procedure of the proposed method is illustrated in Figure 2. + +# 3.2. Evolving Old Knowledge + +Our approach involves transforming previously learned parameters into more adaptable ones to incorporate knowledge from a new task while retaining the old knowledge. This allows for the evolution of existing knowledge to accommodate new information better. Let us denote the hypernetwork as $H(\cdot; \mathcal{H}^t)$ parameterized by $\mathcal{H}^t$ based on the transformer architecture [42]. Note that $\mathcal{H}^t$ is initialized with the parameters of the hypernetwork after training the previous task, $t - 1$ . Specifically, the hypernetwork receives the task token $e^t \in \mathbb{R}^D$ and the set of layer embeddings $\mathcal{C}_l = \{c_n \in \mathbb{R}^D | 1 \leq n \leq N_l\}$ and generates the parameter set $G_l^t$ , where $N_l$ and $l \in \{1 \dots L\}$ denote the number of embeddings for layer $l$ and the layer index of the base network, respectively. + +To generate $G_{l}^{t}$ the task token $e^{t}$ and layer embedding $c_{n}$ from the set $C_l$ are concatenated to construct the input of the hypernetwork, i.e., $z_{n,0}^{t} = \mathrm{Concat}(e^{t},c_{n})\in \mathbb{R}^{2\times D}$ . The input $z_{n,m - 1}^{t}$ is fed into the $m$ -th task-layer attention block (TLA) as + +$$ +\begin{array}{l} \bar {z} _ {n, m - 1} ^ {t} = z _ {n, m - 1} ^ {t} + \operatorname {M S A} (\mathrm {L N} (z _ {n, m - 1} ^ {t})), \tag {1} \\ z _ {n, m} ^ {t} = \bar {z} _ {n, m - 1} ^ {t} + \operatorname {M L P} (\operatorname {L N} (\bar {z} _ {n, m - 1} ^ {t})), \\ \end{array} +$$ + +where $m \in \{1, \dots, N_b\}$ . $N_b$ , LN, MSA, and MLP denote the number of TLA blocks, layer normalization, multi-head self-attention, and multi-layer perceptron, respectively. The structure of TLA is illustrated in Figure 3. MSA is defined for the input $\hat{z}_{n,m-1}^t \triangleq \mathrm{LN}(z_{n,m-1}^t)$ as + +$$ +\operatorname {M S A} \left(\hat {z} _ {n, m - 1} ^ {t}\right) = W ^ {O} \operatorname {C o n c a t} \left(\operatorname {h e a d} _ {1}, \dots , \operatorname {h e a d} _ {N _ {h}}\right). \tag {2} +$$ + +$W^{O} \in \mathbb{R}^{2 \times D}$ is a projection matrix and $N_{h}$ denotes the number of heads in the MSA layer, where $\mathrm{head}_{i} \triangleq \operatorname{softmax}(Q_{i}K_{i}^{T} / \sqrt{D / N_{h}})V_{i}$ and + +$$ +Q _ {i} = W _ {Q _ {i}} \hat {z} _ {n, m - 1} ^ {t}, K _ {i} = W _ {K _ {i}} \hat {z} _ {n, m - 1} ^ {t}, V _ {i} = W _ {V _ {i}} \hat {z} _ {n, m - 1} ^ {t}, +$$ + +where $W_{Q_i}, W_{K_i}$ , and $W_{V_i}$ are projection matrices in the $i$ -th head. The parameters for the $l$ -th layer are generated as + +$$ +\begin{array}{l} G _ {l} ^ {t} = \operatorname {C o n c a t} \left(g \left(\theta_ {1} ^ {t}\right), \dots , g \left(\theta_ {N _ {l}} ^ {t}\right)\right) \tag {3} \\ = \operatorname {C o n c a t} \left(H \left(z _ {1, 0} ^ {t}; \mathcal {H} ^ {t}\right), \dots , H \left(z _ {N _ {l}, 0} ^ {t}; \mathcal {H} ^ {t}\right)\right), \\ \end{array} +$$ + +where $g(\cdot)$ denotes the fully connected layer followed by a sigmoid function. $\theta_{n}^{t}\in \mathbb{R}^{D}$ is the first row of $z_{n,N_b}^t$ , which acts as the final embedding in a transformer architecture [8]. + +To refine old knowledge to align with the knowledge of the $t$ -th task, let us denote the set of old parameters in the $l$ -th layer as $\bar{W}_l^{1:t-1}$ . We obtain the evolved parameters $\hat{W}_l^{1:t-1}$ by fusing $\bar{W}_l^{1:t-1}$ and $G_l^t$ as + +$$ +\hat {W} _ {l} ^ {1: t - 1} = \bar {W} _ {l} ^ {1: t - 1} \cdot G _ {l} ^ {t}, \tag {4} +$$ + +![](images/580be8affda3d993b9c77be535a414575ea27b86e97208bdaa156a6be0e9ba73.jpg) +Figure 3. The task-layer attention block (TLA) consists of a layer normalization, multi-head self-attention (MSA), and MLP with a single hidden layer. $z_{n,N_b}^t$ is generated by $N_{b}$ stacked TLAs. + +where we define $\cdot$ as the element-wise multiplication. The base network accommodates the evolved sets of parameters for $t - 1$ tasks together with the $t$ -th parameter set as $\hat{\mathcal{W}}^t = \{\hat{W}^1, \dots, \hat{W}^{t-1}, W^t\}$ . Finally, the loss function to learn the $t$ -th task using the evolved parameter sets becomes + +$$ +\mathcal {L} _ {\text {t a s k}} = - \sum_ {n = 1} ^ {N ^ {t}} y _ {n} ^ {t} \log \left(F \left(x _ {n} ^ {t}; \hat {\mathcal {W}} ^ {t}\right)\right)). \tag {5} +$$ + +Another loss function to generate parameters from $H(\cdot ;\mathcal{H}^t)$ similar to those learned in the previous task is defined as + +$$ +\mathcal {L} _ {r e g} = \sum_ {k = 1} ^ {t - 1} \sum_ {l = 1} ^ {L} \| \bar {G} _ {l} ^ {k} - G _ {l} ^ {k} \| _ {2} ^ {2}, \tag {6} +$$ + +where $\bar{G}_l^k$ represents the set parameters generated from the hypernetwork after training the task $t - 1$ . + +# 3.3. Sparsity-Inducing Generation + +To renew the network without redundant parameters, we propose to generate sparse parameters in the hypernetwork. As $G_{l}^{t}$ is generated through the sigmoid function, they can be interpreted as the likelihood that each element of the old set of parameters is selected. Following the practice in [41], we minimize $\sum_{l=1}^{L} -\log(1 - G_{l}^{t})$ to achieve sparsity in $G_{l}^{t}$ . However, to prevent $G_{l}^{t}$ from becoming overly sparse, we ensure that parameters significantly impacting the loss remain survived. Inspired by the work in [24], we evaluate the impact of individual elements in $G_{l}^{t1}$ to the loss. We compute the set of all entries of the score $S_{l}^{t}$ by comparing the network losses computed with and without the inclusion of parameters in the learning process. To estimate the effect of removing a parameter on the loss function, we apply a Taylor expansion approximation of the network loss $\mathcal{L}$ for $N$ samples. We obtain the $k$ -th element of the score, $s_{k}$ , as + +$$ +\begin{array}{l} s _ {k} = \mathcal {L} \left(\vec {G} _ {l} ^ {t} - M _ {\backslash (k)} \cdot \vec {G} _ {l} ^ {t}\right) - \mathcal {L} \left(\vec {G} _ {l} ^ {t}\right) \tag {7} \\ \simeq - M _ {\backslash (k)} \cdot \vec {G} _ {l} ^ {t} \cdot \nabla_ {\vec {G} _ {l} ^ {t}} \mathcal {L} + \frac {1}{N} \sum_ {n = 1} ^ {N} \left(\frac {\partial \mathcal {L} _ {n}}{\partial M _ {\backslash (k)} \cdot \vec {G} _ {l} ^ {t}}\right) ^ {2}, \\ \end{array} +$$ + +where $M_{\backslash (k)}$ is an all-zero vector except that the $k$ -th element is 1, and $\mathcal{L}_n$ is the loss of the $n$ -th example. We introduce an additional term minimizing the discrepancy between the scores of the complete set of parameters (i.e., the original scores) and the scores of the sparse set of parameters produced as follows: + +$$ +\mathcal {L} _ {s p} = \sum_ {l = 1} ^ {L} - \log (1 - G _ {l} ^ {t}) + \left\| S _ {l} ^ {t} \cdot \left(1 - \mathbb {1} (\vec {G} _ {l} ^ {t} \neq 0)\right) \right\| _ {1}, \tag {8} +$$ + +where $\| \cdot \|_1$ and $\mathbb{1}(\cdot)$ denote the $l_1$ -norm and the indicator function, respectively. + +Total Loss. The proposed framework learns the $t$ -th task with the three loss functions. To learn the $t$ -th task, we minimize the cross-entropy loss $\mathcal{L}_{task}$ by updating $W^t$ in the base network. Also, $\mathcal{H}^t$ , $e^t$ , and $\{C_l\}_{l=1}^L$ are updated from two losses, $\mathcal{L}_{task}$ and $\mathcal{L}_{sp}$ . Last, we update $\mathcal{H}^t$ and $\{C_l\}_{l=1}^L$ by minimizing $\mathcal{L}_{reg}$ to preserve the previously generated parameters. The total loss is + +$$ +\min _ {W ^ {t}, \mathcal {H} ^ {t}, e ^ {t}, \{C _ {l} \} _ {l = 1} ^ {L}} \mathcal {L} _ {\text {t a s k}} + \lambda_ {s p} \mathcal {L} _ {s p} + \lambda_ {r e g} \mathcal {L} _ {r e g}, \tag {9} +$$ + +where $\lambda_{sp}$ and $\lambda_{reg}$ are parameters to balance the losses. + +Scalability. A hypernetwork is designed to generate parameters for every layer of the base network, and the parameters were composed of compact segments of a consistent size [13, 21]. While it can generate parameters for all layers, this approach requires a substantial quantity of layer embeddings, particularly for deeper layers. To mitigate this, we adjust the scale of the segments for each layer depending on the number of channels. However, it incurs a memory overhead stemming from the substantial dimensions of the heads. To address this, we apply a matrix decomposition technique [39] to generate two reduced sets of low-rank $r$ . By doing so, we achieve a reduction in memory requirements for layer embeddings, accompanied by a marginal increase in hypernetwork parameters. For detailed results, please refer to Section 4.5.1. + +# 4. Experiments + +# 4.1. Scenarios + +We applied the proposed method, GrowBrain, to image classification and video action recognition problems under class incremental and task-incremental learning scenarios. Image classification. We conducted experiments using ImageNet [6] and five fine-grained datasets including CUBS [45], Stanford Cars [18], Flowers [31], Wikiart [36], and Sketch [11], where each dataset is considered a task. We call the datasets ImageNet+FG for short. In addition, we demonstrated the proposed method with the CIFAR-100. We divided the dataset into subsets of classes based on the random order provided by [35, 9] and considered + +Table 1. Datasets used in this work. + +
Image Dataset# Train Images# Test Images# Classes
ImageNet [6]1,287,16750,0001,000
CUBS [45]5,9945,794200
Stanford Cars [18]8,1448,041196
Flowers [31]2,0406,149102
Wikiart [36]42,12910,628195
Sketch [11]16,0004,000250
CIFAR-100 [19]50,00010,000100
Video Dataset# Train Videos# Test Videos# Classes
ActivityNet [3]10,0244,926200
UCF-101 [40]9,2802,720101
+ +Table 2. The specification of the hypernetwork. + +
Base networkHypernetwork
Embed dim (D)Hidden dim in MLP# Heads (Nh)# Blocks (Nb)
ResNet-506412887
ResNet-34
ResNet-18326485
+ +each subset a task. Video action recognition. We used a trimmed version of ActivityNet [3] and UCF-101 [40] video datasets. To conduct experiments, we divided each dataset into subsets containing randomly selected classes provided by [43] and treated each subset as a task. We summarize the datasets as mentioned above in Table 1. + +# 4.2. Implementation details + +Image classification. ImageNet+FG and CIFAR-100 were resized to $224 \times 224$ and $32 \times 32$ pixels, respectively. We applied random cropping and horizontal flip as augmentation. For those datasets, we used the ResNet-50 [14] and ResNet-18 as the base networks, respectively. + +Video action recognition. We used TSN [46] with the ResNet-34 backbone network for video action recognition. Each video was split into three segments of equal duration, and a frame was randomly selected from each segment. The action class prediction was performed by aggregating the predictions of the selected frames with a consensus function. We used an average of predictions as a consensus function [46]. The video datasets were augmented with random cropping and horizontal flipping following [46]. We resized each frame of ActivityNet and UCF-101 to $224 \times 224$ pixels. + +Table 2 reports the base networks and corresponding hypernetworks, which include the embedding dimension, the hidden dimension of the MLP, the number of heads, and the number of TLA blocks. We set the rank $r$ for matrix decomposition as 7. We trained the base and hypernetwork using stochastic gradient descent. In all our experiments, we set $\lambda_{reg}$ to 1 and $\lambda_{sp}$ to 0.1, except ImageNet+FG where we set $\lambda_{sp}$ to 1. We report the average results over five independent runs for all the experiments. To enable class-incremental learning in scenarios where the task identity is unknown, the proposed method first predicts the task index and then conducts classification. Following [44], we predict the task + +Table 3. Performance of the class-incremental learning scenario using ImageNet and five fine-grained datasets (ImageNet+FG), where the average accuracy was measured after training each task. # P represents the average number of parameters $(\times 10^{6})$ to perform the tasks. + +
MethodImageNetCUBSStanford CarsFlowersWikiartSketch
Acc#PAcc#PAcc#PAcc#PAcc#PAcc#P
EWC-On76.123.431.623.428.623.429.223.423.123.411.323.4
LwF76.123.468.423.458.223.455.823.449.223.428.423.4
iCaRL76.123.470.223.461.523.452.723.448.323.442.323.4
PackNet75.712.562.813.365.214.066.514.654.315.245.315.7
Piggyback76.123.460.122.754.822.256.422.049.721.747.421.4
WSN74.712.559.612.542.212.529.212.531.012.525.012.5
H275.712.569.513.067.312.968.413.061.812.948.712.6
GrowBrain75.712.569.211.069.310.371.39.864.39.660.19.4
+ +![](images/a44403d2458280411c417bba241f1ad75828a91eca21a4190b80f5006df31cb2.jpg) +Figure 4. Accuracy of class-incremental learning for image classification on ImageNet+FG (left), video action recognition on UCF-101 (middle) and ActivityNet (right). + +![](images/9e98dedb86e4c1568827fdaec08a34720fcb9f3adf90a47b284ad3de833a6892.jpg) + +![](images/ee8d531d8ec70c4bb451d10a2eda8c0c6e3279c02b4bb822f3edb3ddf8a53f0b.jpg) + +index using the network with the least entropy that has been learned thus far. + +# 4.3. Class-incremental learning + +# 4.3.1 Image classification + +We evaluated the proposed method of evolving the previous knowledge to the class-incremental learning scenario using ImageNet+FG. We compared with regularization methods, EWC-On [37] and LwF [22], and data rehearsal methods, iCaRL [35], and parameter isolation methods, PackNet [28], Piggyback [27], WSN [16], and $\mathrm{H}^2$ [15]. The data rehearsal methods stored 1,000 raw instances (taking 150 $\mathrm{MB}^2$ ). We applied the same task prediction mechanism as ours to the comparing methods, excluding iCaRL. + +Table 3 shows the class-incremental learning results for ImageNet and five fine-grained datasets. Overall, the proposed method outperforms competitors (more detailed results are shown on the left of Figure 4). The regularization approaches, EwC-On and LwF, give unsatisfactory results with performance gaps of $48.8\%$ and $31.7\%$ compared to ours, respectively. LwF shows a $20.8\%$ accuracy drop before and after learning Sketch, which is the most significant + +among the comparing methods. The data rehearsal method, iCaRL, shows a $17.1\%$ performance drop from GrowBrain despite storing 1,000 old samples. PackNet shows large performance drops of $12.2\%$ and $9.0\%$ after training on WikiArt and Sketch, respectively. This is because negative interference occurs when using all previous sets of parameters optimized for different task domains. The methods of associating prior knowledge, Piggyback, WSN, and $\mathrm{H}^2$ , show notable performance drop ( $12.7\%$ , $35.1\%$ , and $11.4\%$ , respectively) compared to the proposed method that improves the previous knowledge from the new task. The results demonstrate that the proposed approach surpasses a wide range of continual learning methods, confirming its superiority. + +# 4.3.2 Video action recognition + +We applied our method and the competitors to the video action recognition tasks using the benchmark sets proposed for continual learning [43]. For this experiment, iCaRL was excluded from the comparison since it demands approximately 2TB and 52GB of memory for replay [43] on ActivityNet and UCF-101, respectively. Since Piggyback requires a pre-trained backbone, we pre-trained the backbone model [14] using the validation set of UCF-101 for video action + +Table 4. Results of the class-incremental learning scenario using the trimmed version of UCF-101 and ActivityNet, where the Avg and Last accuracies for the time steps 5 and 10. + +
MethodUCF-101ActivityNet
T5T10T5T10
AvgLast# PAvgLast# PAvgLast# PAvgLast# P
EWC-On48.725.821.232.713.921.238.420.721.226.612.421.2
LwF47.027.721.233.415.421.244.129.121.232.718.421.2
PackNet59.354.211.552.540.415.242.430.211.532.419.015.2
Piggyback65.451.720.453.536.520.422.212.720.515.67.520.4
WSN69.158.810.161.149.710.5------
H263.654.88.356.947.011.437.422.53.426.813.32.1
GrowBrain79.372.27.273.564.78.942.829.911.533.220.615.2
+ +recognition experiments. We report the $Avg$ and Last accuracies [9] for the time steps 5 and 10. Figure 4 shows the Last accuracy measured on the UCF-101 and ActivityNet datasets. + +UCF-101. The results of the class-incremental learning scenario using the UCF-101 dataset are reported in Table 4 (left). The aspects of the experimental results on UCF-101 of the compared approaches are similar to those of image classification. Despite utilizing a large number of parameters, regularization approaches, EWC-On and LwF, give unsatisfactory performance compared to parameter isolation approaches. PackNet, which uses all parameters from previous tasks, performs less than other parameter isolation approaches, WSN and $\mathrm{H}^2$ , using selectively chosen parameters associated with previous knowledge. The proposed method shows the highest performance compared to other comparison methods. Ours outperforms WSN, $\mathrm{H}^2$ , PackNet, and Piggyback with the significant accuracy gap on the Last measure from $15.0\%$ to $28.2\%$ . + +ActivityNet. Table 4 (right) reports the results on ActivityNet. WSN was excluded from the experiment because it did not give promising results due to the lack of data. Overall, the results of ActivityNet show a different pattern compared to those of UCF-101. The regularization method, LwF, outperforms the parameter isolation method, PackNet, with $1.7\%$ and $0.3\%$ Avg accuracy gap measured after training the 5-th and 10-th tasks. The proposed approach performs better than Piggyback, PackNet, and $\mathrm{H}^2$ by large margins of $13.1\%$ , $1.6\%$ , and $7.3\%$ , respectively, in terms of the Last accuracy measured after training 10-th task. The proposed method outperforms other competitors for image classification and video action recognition problems. In contrast to the UCF-101 dataset, where the proposal can acquire new knowledge by using a small number of parameters from previous tasks and achieve excellent results, the ActivityNet dataset requires previous knowledge to effectively train the current task due to its inadequate data (the number of videos per class is only half compared to UCF-101). + +Table 5. Task-incremental learning scenarios using video action recognition and image classification datasets, where average accuracy and parameter consumption are reported. + +
MethodImage classificationVideo action recognition
ImageNet+FGCIFAR-100UCF-101ActivityNet
Acc#PAcc#PAcc#PAcc#P
EWC-On55.323.444.511.165.821.261.521.2
LwF53.623.439.711.171.221.265.821.2
PackNet80.215.782.110.492.215.265.315.2
Piggyback82.221.475.610.586.620.442.120.4
WSN61.712.584.65.692.710.5--
H282.812.685.91.194.511.444.02.1
GrowBrain83.09.486.82.596.78.966.915.2
+ +# 4.4. Task-incremental learning + +Additionally, we conducted experiments on the task-incremental learning scenario where task-id is given during the inference step. In this scenario, we used ImageNet+FG and CIFAR-100 as the image classification datasets. Also, we used ActivityNet and UCF-101 for video action recognition. We divided CIFAR-100 into 10 tasks based on [35]. We followed the same experimental setting as in Tables 3 and 4. For the experiments of Piggyback on CIFAR-100, the backbone network was pre-trained on a randomly selected subset (40%) of the CIFAR-100 training set. + +Image classification. Table 5 reports the results of task-incremental learning conducted on ImageNet+FG and CIFAR-100. Even if the regularization methods, EwC-On and LwF, try to maintain previous knowledge, they perform significantly less than the parameter isolation methods. GrowBrain consistently achieves high performance compared to other parameter isolation methods, PackNet, Piggyback, WSN, and $\mathrm{H}^2$ , in both image classification experiments (e.g., $0.2\%$ and $0.9\%$ performance gap on ImageNet+FG and CIFAR-100 with the most competitive method, $\mathrm{H}^2$ ). + +Video action recognition. Table 5 reports the results of task-incremental learning conducted on UCF-101 and ActivityNet. Overall, the results are similar to those of class-incremental learning scenarios. The regularization method, LwF, achieves higher performance than PackNet on Activ + +Table 6. Ablation study of the proposed method for image classification and video action recognition tasks, where average accuracy, the number of parameters for the base network and hypernetwork (in millions), and the number of layer embeddings are reported. + +
MethodImageNet+FGUCF-101
Acc# (F + H)# cnAcc# (F + H)# cn
Ours60.19.4 + 2.813364.78.9 + 2.858
w/o transform [28]45.315.7 + 0040.415.2 + 00
w/o sparsity58.115.7 + 2.813364.014.1 + 2.858
w/o scalability59.815.7 + 2.9326264.711.0 + 2.9614
+ +ityNet. The proposed method outperforms both parameter isolation and regularization methods. Notably, our method outperforms WSN, PackNet, and $\mathrm{H}^2$ with a performance gap of $4\%$ , $4.5\%$ , and $2.2\%$ , respectively, on UCF-101. + +# 4.5. Analysis + +# 4.5.1 Ablation study + +To demonstrate the effectiveness of our method, we conducted an ablation study. We compared ours by removing transformation of previous knowledge [28], sparsity-inducing generation $(\lambda_{sp} = 0)$ , and scalability of hypernetwork, i.e., channel adaptive segments and matrix decomposition. We applied these methods to class-incremental learning scenarios using the ImageNet+FG and the UCF-101 datasets. We report the Last accuracy measured after training all tasks, the average parameter consumption of the base network $(F)$ and hypernetwork $(H)$ to perform each task, and the number of layer embeddings $c_{n}$ . + +Table 6 reports the results of the ablation study. Our method outperforms that without transformation by performance gaps of $14.8\%$ and $24.3\%$ for those datasets, emphasizing that the evolution of prior knowledge is critical in enhancing performance. For video action recognition using UCF-101, the method without sparsity requires more parameters and performs less than the proposed method. The proposed method for ImageNet+FG achieves higher performance while taking fewer parameters than the nonsparse alternative. The reason is that sparsity-inducing hypernetwork prevents the generation of redundant parameters and regularizes the generated parameters to achieve generalization performance. The method without scalability requires 24 and 10.5 times larger layer embeddings on ImageNet+FG and UCF-101 than the proposed method, respectively, without accompanying performance improvement. + +# 4.5.2 Impact of $\lambda_{sp}$ + +We further conducted a study investigating the impact of $\lambda_{sp}$ , which regulates the sparsity of generated parameters. We experimented with varying values of $\lambda_{sp}$ in the ImageNet+FG experiment under the class-incremental scenario using the same setup presented in Table 3. Figure 5 shows the parameter consumption (left) and average ac + +![](images/89486965460d5420072724b4798da3eb4be9cabae17d24a4a876116dc731659d.jpg) +Figure 5. Parameter consumption and average accuracy of the proposed method based on varying $\lambda_{sp}$ in class-incremental learning for ImageNet+FG. + +![](images/e1fe5106a3da215418b6ab96e9515f51cfa3834f5cf3824f27c28c231544552a.jpg) + +curacy (right) with respect to different $\lambda_{sp}$ after training all tasks. Ours with $\lambda_{sp} = 0.1$ uses a similar amount of parameters as the baseline [28] but outperforms the competing approach by leveraging enhanced prior knowledge. Notably, after training all tasks, $\lambda_{sp} = 1.0$ takes $9.4\mathrm{M}$ parameters while showing $0.7\%$ performance improvement compared to $\lambda_{sp} = 0.1$ consuming $15.7\mathrm{M}$ parameters. + +# 5. Conclusion + +We have proposed a novel approach, GrowBrain, to transform learned old knowledge to absorb the information of newer tasks in parameter isolation-based continual learning. GrowBrain reinforces the previous knowledge with a transformer-based hypernetwork that generates optimized parameters for new tasks while eliminating redundant parameters that show negligible impact on the loss. The proposed method overcomes the limitation of using fixed old knowledge that can constrain the ability to capture novel representations while taking fewer parameters. In the experiments, we have evaluated the proposal compared to a variety of continual learning methods in class- and task-incremental learning scenarios for image classification and video action recognition. Experimental results show that the proposed approach of leveraging transformed prior knowledge generates parameters adapted to the new task without losing the prior knowledge, resulting in significant performance improvements over its strong competitors. + +Acknowledgements. This work was supported in part by Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT2002-05, Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (2022-0-01341, Artificial Intelligence Graduate School Program(Chung-Ang University)), and BK21 FOUR (Fostering Outstanding Universities for Research) Program funded by Ministry of Education of Korea (No.I22SS7609062). + +# References + +[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3931-3940, 2020. 1, 2 +[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European Conference on Computer Vision, pages 144-161. Springer, 2018. 1, 2 +[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5 +[4] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 556-572. Springer, 2018. 1, 2 +[5] Arslan Chaudhry, Ranzato Marc'Aurelio, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In 7th International Conference on Learning Representations, ICLR 2019. 1, 2 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5 +[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1, 2 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 4 +[9] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 5, 7 +[10] Benjamin Ehret, Christian Henning, Maria Cervera, Alexander Meulemans, Johannes Von Oswald, and Benjamin F Grewe. Continual learning in recurrent neural networks. In International Conference on Learning Representations, 2020. 3 +[11] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Transactions on Graphics, 31(4):1-10, 2012. 5 +[12] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2 + +[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3, 5 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6 +[15] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, pages 519-535. Springer, 2022. 1, 2, 3, 6 +[16] Haeyong Kang, Rusty John Lloyd Mina, Sultan Rizky Hikmawan Madjid, Jaehong Yoon, Mark Hasegawa-Johnson, Sung Ju Hwang, and Chang D Yoo. Forget-free continual learning with winning subnetworks. In International Conference on Machine Learning, pages 10734-10750. PMLR, 2022. 2, 6 +[17] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America, 114(13):3521-3526, 2017. 1, 2 +[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561. IEEE, 2013. 5 +[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 2, 5 +[20] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. Advances in Neural Information Processing Systems, 30:4652-4662, 2017. 1, 2 +[21] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. DHP: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2, 3, 5 +[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. In European Conference on Computer Vision, pages 614-629. Springer, 2016. 1, 2, 6 +[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2 +[24] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group fisher pruning for practical network compression. In International Conference on Machine Learning, pages 7021-7032. PMLR, 2021. 4 +[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2544-2553, 2021. 1, 2 +[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6470-6479, 2017. 1, 2 + +[27] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In Proceedings of the European Conference on Computer Vision, pages 67-82, 2018. 2, 6 +[28] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 1, 2, 3, 6, 8 +[29] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, volume 24, pages 109–165. Elsevier, 1989. 1 +[30] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1, 2 +[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5 +[32] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 1 +[33] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13698-13707, 2021. 2 +[34] Dripta S Raychaudhuri, Yumin Suh, Samuel Schulter, Xiang Yu, Masoud Faraki, Amit K Roy-Chowdhury, and Manmohan Chandraker. Controllable dynamic multi-task architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2022. 2 +[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 5, 6, 7 +[36] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. International Journal for Digital Art History, (2), 2016. 5 +[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International Conference on Machine Learning, pages 4528-4537. PMLR, 2018. 6 +[38] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 2994-3003, 2017. 1, 2 +[39] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 5 + +[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5 +[41] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. AdaShare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 4 +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4 +[43] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vCLIMB: A novel video class incremental learning benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19035-19044, 2022. 5, 6 +[44] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. In 8th International Conference on Learning Representations, 2020. 2, 3, 5 +[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 5 +[47] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2 +[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3987-3995, 2017. 1, 2 \ No newline at end of file diff --git a/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/images.zip b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..779d85ef30917e2584329d8136f2ea7cb8ee75dd --- /dev/null +++ b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:090d1c8c060fc9cdf8d6a0a94cddcf7d281715af150c7a51f6303bb5d08c89e3 +size 635528 diff --git a/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/layout.json b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9aa9e02555e27434bdb3e4949425908597decf9e --- /dev/null +++ b/2023/Growing a Brain with Sparsity-Inducing Generation for Continual Learning/layout.json @@ -0,0 +1,9364 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "text", + "content": "Growing a Brain with Sparsity-Inducing Generation for Continual Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 487, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 142, + 487, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 487, + 186 + ], + "type": "text", + "content": "Hyundong Jin1 Gyeong-hyeon Kim1 Chanho Ahn2* Eunwoo Kim1 \n1School of Computer Science and Engineering, Chung-Ang University \n2Samsung Advanced Institute of Technology (SAIT)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 187, + 476, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 187, + 476, + 200 + ], + "spans": [ + { + "bbox": [ + 118, + 187, + 476, + 200 + ], + "type": "text", + "content": "{jude0316,leonardkkh,eunwoo}@cau.ac.kr chanho.ahn@samsung.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 253, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 289, + 540 + ], + "type": "text", + "content": "Deep neural networks suffer from catastrophic forgetting in continual learning, where they tend to lose information about previously learned tasks when optimizing a new incoming task. Recent strategies isolate the important parameters for previous tasks to retain old knowledge while learning the new task. However, using the fixed old knowledge might act as an obstacle to capturing novel representations. To overcome this limitation, we propose a framework that evolves the previously allocated parameters by absorbing the knowledge of the new task. The approach performs under two different networks. The base network learns knowledge of sequential tasks, and the sparsity-inducing hypernetwork generates parameters for each time step for evolving old knowledge. The generated parameters transform old parameters of the base network to reflect the new knowledge. We design the hypernetwork to generate sparse parameters conditional to the task-specific information and the structural information of the base network. We evaluate the proposed approach on class-incremental and task-incremental learning scenarios for image classification and video action recognition tasks. Experimental results show that the proposed method consistently outperforms a large variety of continual learning approaches for those scenarios by evolving old knowledge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 287, + 693 + ], + "type": "text", + "content": "The ability of deep learning to incrementally accumulate knowledge for sequentially incoming tasks is an important capability required for modern AI algorithms. However, deep neural networks tend to lose previously acquired knowledge while learning new data. This limitation of deep neural networks, called catastrophic forgetting, has been investigated by previous studies [29, 17]. Catastrophic forgetting significantly degrades performance when data from previous tasks (task-incremental) or already-learned class" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "content": "data (class-incremental) cannot be accessed. Under those incremental learning scenarios, continual learning methods have toiled to preserve the previous knowledge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 266, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 546, + 506 + ], + "type": "text", + "content": "Various approaches have been proposed for continual learning and are categorized into three types: regularization [2, 4, 20, 17, 48], data rehearsal [5, 26, 32, 35, 38, 47], and parameter isolation [1, 30, 28]. The conventional continual learning methods introduced regularization approaches to suppress changes in important parameters [48, 2] or input-output mappings from a network trained up to previous tasks [22, 7]. The network is trained in a way that contributable parameters to the performance of the previously trained model remain unchanged. On the other hand, data rehearsal methods [25, 35] address catastrophic forgetting by accessing old data. The old data can be obtained by storing some data from old tasks [35, 47] or leveraging a generative network [38]. However, the privacy of data and the difficulty of learning generative models limit their versatility. Recently, parameter isolation approaches [28, 30, 15] construct a disjoint set of parameters for each task, preventing it from mixing old and new knowledge. These methods learn a new set of parameters by associating all [28] or some helpful old knowledge [15]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 508, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 638 + ], + "type": "text", + "content": "Unlike other approaches, the parameter isolation approach is designed to prevent forgetting without requiring access to old data by disallowing updates to previous sets of parameters. Despite this advantage, these methods can limit the ability to learn new tasks by relying on parameters that were learned for older tasks and remain unchanged. We observe that this limitation arises when there is a low correlation between the old and new data. This raises the question: In parameter isolation-based continual learning, is it possible to transform the knowledge from old tasks to fit newer tasks better?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we introduce a novel method for transforming prior knowledge into a form that can aid in learning new tasks. The approach involves reframing the parameter selection problem, which previous works [15] have addressed, as a generation problem. The proposed framework emphasizes a subset of pre-trained knowledge, selecting and highlight-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 282, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 282, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 282, + 714 + ], + "type": "text", + "content": "*This work was done independently, without any support from SAIT." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18961" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 346 + ], + "type": "text", + "content": "ing it for use in learning new tasks. Specifically, the framework consists of two networks: base network and hypernetwork. The base network constructs disjoint parameter sets for sequential tasks. The proposed transformer-based [42] hypernetwork generates a small set of influential parameters that significantly impact the loss. The impact of each element of generated parameters is evaluated by the difference in loss computed with and without its inclusion in the learning process. Also, generated parameters are conditioned on the information of the current task (task token) and the layer of the base network (layer embedding). The generated parameters transform the previous set of parameters in the base network to further accommodate new knowledge. In addition, compared to existing hypernetworks that generate a small set of parameters and apply it to all the layers [13, 44], the proposed hypernetwork adaptively generates parameters according to the layer configurations defined in the base network. Through this strategy, the quantity of layer embeddings for ResNet-50 is reduced by a substantial factor of 24. Furthermore, by leveraging a matrix decomposition technique to represent generated parameters compactly, we achieve a noteworthy reduction in the parameters necessary for the hypernetwork." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 348, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 289, + 575 + ], + "type": "text", + "content": "We evaluate the effectiveness of the proposed method in both class-incremental and task-incremental learning scenarios using classification benchmark datasets. Furthermore, we demonstrate the efficacy of the proposed method [28, 15] in these scenarios by applying it to video action recognition datasets [19]. Experimental results show that the proposed method performs better than its competitors for most learning scenarios while introducing minor memory overhead from the hypernetwork. The contributions of our work are mainly three-fold: (1) We present a novel approach to evolving previous knowledge in continual learning, aimed at reflecting the knowledge of new tasks. (2) We propose a transformer-based hypernetwork that generates customized parameters for new tasks while suppressing the generation of redundant parameters that have little impact on the loss. (3) Experimental results on diverse benchmarks for video action recognition and image classification demonstrate that ours outperforms its competitors with lower network capacity requirements to perform tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 288, + 715 + ], + "type": "text", + "content": "Continual Learning. Generally, continual learning methods can be classified into three types. Regularization methods [23, 7, 17, 48, 4, 2, 20] add a penalty term that restrains the previously learned knowledge from being updated when learning a new task. In order not to hurt previous knowledge, a new network is introduced to mimic the features obtained from the old network [22, 7]. While other regularization techniques, such as those based on the Fisher information matrix [17], change of loss [48], and variation of" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 539, + 167 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 539, + 167 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 539, + 167 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 539, + 167 + ], + "type": "image", + "image_path": "2d8c0b91120742ed46a2160189708b9340e73b523ac00cd4bc347e6ca22661e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 169, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 169, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 547, + 258 + ], + "type": "text", + "content": "Figure 1. An illustration of the proposed method (right) and baseline methods (middle, left) for the third (new) task. The figure represents task-specific allocated parameters, represented by task numbers. Gray numbers represent the sets of discarded parameters from a selection method. The baseline methods do not update old knowledge (red boxes) while updating a new task (blue boxes). The proposed method transforms the old sets of parameters, represented by " + }, + { + "bbox": [ + 304, + 169, + 547, + 258 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 304, + 169, + 547, + 258 + ], + "type": "text", + "content": ", to fit the new task. Best viewed in color." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 280, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 547, + 532 + ], + "type": "text", + "content": "prediction [2], suppress the update of crucial parameters for each task, they experience significant performance decline as the number of tasks grows [12]. Data rehearsal methods [5, 26, 35, 33, 25, 38] rely on the availability of previous task data, which is achieved through selective storage of data [35, 47, 25]. These methods employ a few previous samples jointly with a new dataset [25, 38] or constraining the gradient based on the angle between that obtained from stored samples and the gradient obtained from the new samples [5, 26]. These methods incur a data imbalance problem due to a small number of stored previous tasks compared to a large amount of new task data. Parameter isolation methods [28, 1, 30, 15, 16] involve constructing disjoint sets of parameters for old tasks, respectively, and updating another set of parameters for a newer task by associating it with old parameters during training. This is usually done by using all previous sets of parameters [28], using a learnable mask to select some previous set of parameters [27], jointly selecting a location to allocate a new set of parameters [16], or searching a subnetwork with coarse-to-fine task association [15]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 533, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 616 + ], + "type": "text", + "content": "However, this approach can impede efficient learning because the parameters optimized for old tasks remain the same when learning newer tasks, meaning they need to absorb the information from the newer tasks. In contrast, the proposed method allows the old knowledge to grow by incorporating the knowledge from new tasks. The proposed method and its baselines are illustrated in Figure 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": "Hypernetworks. Hypernetworks generate parameters for a target network [13, 44, 39, 21, 34]. A recent work [34] proposes a method for generating parameters for layers with few parameters, such as activation of ResNet blocks or batch normalization layers. However, generating parameters for convolutional and linear layers that take most of the parameters in the base network can be challenging because it is difficult to represent the entire set of parameters from a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18962" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 533, + 265 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 533, + 265 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 533, + 265 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 533, + 265 + ], + "type": "image", + "image_path": "c5d4fe3df336ce739587dcec0b7ab5dd0688a6e13c1fa1176f88f1d65470b0fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": "Figure 2. The proposed method for transforming the parameters of the " + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": "-th layer of the base network " + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": " when learning the " + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": "-th task. The numbers in the grid indicate the indices of tasks. The hypernetwork " + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": " generates parameters using the task-layer attention block (TLA), which discovers the relationship between the task token and layer embedding. The sparse parameters are generated while suppressing the generation of the redundant ones by evaluating the impact of parameters on the loss. The previous sets of parameters of the " + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": "-th layer are transformed by element-wise multiplication with " + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "inline_equation", + "content": "G_{l}^{t}" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": ". The transformed parameters are combined with the current parameter set through the union operation " + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "inline_equation", + "content": "(\\cup)" + }, + { + "bbox": [ + 46, + 269, + 548, + 338 + ], + "type": "text", + "content": " and released as a complete set of parameters to learn new knowledge. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 498 + ], + "type": "text", + "content": "hypernetwork. To avoid this, existing works generate small sets of parameters and concatenates them [13, 21]. Consequently, to generate a base network of large scale, allowing a large number of input embeddings corresponding to the network is inevitable due to the repetitive generation mechanism. In contrast to the approach, the parameters are generated according to the dimensions defined in the layers of the base network in our method. Note that as these dimensions expand, the corresponding demands on hypernetwork parameters also escalate. To address this, we employ a matrix decomposition technique for a compact representation of the required parameters in the hypernetwork." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 643 + ], + "type": "text", + "content": "Recently, in the continual learning literature, [44] and [10] learn a hypernetwork to generate parameters with task-specific embeddings for new tasks while preventing changes to previously generated parameters for previous tasks. The approaches produce dense parameters for the base network through a task-conditional hypernetwork. In contrast, the proposed transformer-based hypernetwork generates sparse parameters by measuring the effect of individual parameters on the loss. Furthermore, our method generates parameters to evolve previous knowledge to reflect new knowledge, whereas the aforementioned approaches generate parameters directly for the base network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 651, + 122, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 122, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 122, + 662 + ], + "type": "text", + "content": "3. GrowBrain" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 122, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 122, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 122, + 682 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "The base network " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " constructs a disjoint set of parameters for each task under an iterative train-prune" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": "retrain framework [28]. At a time step " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": ", the base network is learned for the " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": "-th task with the dataset " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "T^t = \\{(x_n^t, y_n^t)\\}_{n=1}^{N^t}" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "x_i^t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": "-th sample and " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "y_i^t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": " is corresponding target, respectively. Before training the task, we allocate parameters for the new task to the pruned positions in the base network trained for " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": " tasks. As a typical strategy, the allocated parameters " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "W^t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": " are learned from the help of all previous parameter sets [28]. After training, redundant parameters in " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "W^t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": " are pruned, and the unpruned parameters, " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "\\bar{W}^t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": ", are retrained. Finally, after learning for " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": " tasks, " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "F(\\cdot; \\mathcal{W}^t)" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": " is parameterized with " + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^t = \\{\\bar{W}^1, \\dots, \\bar{W}^t\\}" + }, + { + "bbox": [ + 304, + 354, + 547, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": "Although using previous knowledge can be beneficial in learning new tasks, there is also the possibility of negative knowledge transfer. If earlier tasks are irrelevant to a new task, leveraging their knowledge leads to suboptimal performance of the new task. A recent approach involving the selective use of helpful parameters [15] has been proposed. However, it is based on fixed old knowledge, which limits the model's ability to evolve in response to new information. Consequently, the expressiveness of the model remains constrained by its dependence on immutable prior knowledge. To resolve the issue, we propose a novel framework that transforms old parameters into those absorbing the knowledge of newer tasks. To selectively transform the parameter sets of old tasks " + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\bar{W}^1,\\dots ,\\bar{W}^{t - 1}\\}" + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": " for the time " + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 498, + 548, + 715 + ], + "type": "text", + "content": ", we generate parameters by presenting a sparsity-inducing hypernetwork. The generated ones enhance the previous parameters to fit the new task better. The entire procedure of the proposed method is illustrated in Figure 2." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18963" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 187, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 187, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 187, + 85 + ], + "type": "text", + "content": "3.2. Evolving Old Knowledge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": "Our approach involves transforming previously learned parameters into more adaptable ones to incorporate knowledge from a new task while retaining the old knowledge. This allows for the evolution of existing knowledge to accommodate new information better. Let us denote the hypernetwork as " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "H(\\cdot; \\mathcal{H}^t)" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " based on the transformer architecture [42]. Note that " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " is initialized with the parameters of the hypernetwork after training the previous task, " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": ". Specifically, the hypernetwork receives the task token " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "e^t \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " and the set of layer embeddings " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l = \\{c_n \\in \\mathbb{R}^D | 1 \\leq n \\leq N_l\\}" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " and generates the parameter set " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "G_l^t" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "N_l" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "l \\in \\{1 \\dots L\\}" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " denote the number of embeddings for layer " + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 90, + 287, + 257 + ], + "type": "text", + "content": " and the layer index of the base network, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": "To generate " + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "inline_equation", + "content": "G_{l}^{t}" + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": " the task token " + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "inline_equation", + "content": "e^{t}" + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": " and layer embedding " + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "inline_equation", + "content": "c_{n}" + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": " are concatenated to construct the input of the hypernetwork, i.e., " + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "inline_equation", + "content": "z_{n,0}^{t} = \\mathrm{Concat}(e^{t},c_{n})\\in \\mathbb{R}^{2\\times D}" + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": ". The input " + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "inline_equation", + "content": "z_{n,m - 1}^{t}" + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": " is fed into the " + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 257, + 287, + 317 + ], + "type": "text", + "content": "-th task-layer attention block (TLA) as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 324, + 287, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 324, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 81, + 324, + 287, + 355 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {z} _ {n, m - 1} ^ {t} = z _ {n, m - 1} ^ {t} + \\operatorname {M S A} (\\mathrm {L N} (z _ {n, m - 1} ^ {t})), \\tag {1} \\\\ z _ {n, m} ^ {t} = \\bar {z} _ {n, m - 1} ^ {t} + \\operatorname {M L P} (\\operatorname {L N} (\\bar {z} _ {n, m - 1} ^ {t})), \\\\ \\end{array}", + "image_path": "7af7124c64136e93e23cf25ac5244180bbe88d3f0a07e4ef0d171304ccc99b74.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "inline_equation", + "content": "m \\in \\{1, \\dots, N_b\\}" + }, + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "inline_equation", + "content": "N_b" + }, + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "text", + "content": ", LN, MSA, and MLP denote the number of TLA blocks, layer normalization, multi-head self-attention, and multi-layer perceptron, respectively. The structure of TLA is illustrated in Figure 3. MSA is defined for the input " + }, + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\hat{z}_{n,m-1}^t \\triangleq \\mathrm{LN}(z_{n,m-1}^t)" + }, + { + "bbox": [ + 46, + 362, + 287, + 425 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 432, + 287, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 287, + 447 + ], + "type": "interline_equation", + "content": "\\operatorname {M S A} \\left(\\hat {z} _ {n, m - 1} ^ {t}\\right) = W ^ {O} \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {N _ {h}}\\right). \\tag {2}", + "image_path": "dd3d560080911653756da038bacb693dafa28b0ce1b384c6b02aa1156d82f80b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "type": "inline_equation", + "content": "W^{O} \\in \\mathbb{R}^{2 \\times D}" + }, + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "type": "text", + "content": " is a projection matrix and " + }, + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "type": "inline_equation", + "content": "N_{h}" + }, + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "type": "text", + "content": " denotes the number of heads in the MSA layer, where " + }, + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{head}_{i} \\triangleq \\operatorname{softmax}(Q_{i}K_{i}^{T} / \\sqrt{D / N_{h}})V_{i}" + }, + { + "bbox": [ + 47, + 455, + 287, + 493 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 501, + 289, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 289, + 517 + ], + "type": "interline_equation", + "content": "Q _ {i} = W _ {Q _ {i}} \\hat {z} _ {n, m - 1} ^ {t}, K _ {i} = W _ {K _ {i}} \\hat {z} _ {n, m - 1} ^ {t}, V _ {i} = W _ {V _ {i}} \\hat {z} _ {n, m - 1} ^ {t},", + "image_path": "6fe09c5bad42f151020b77ea2e8f13d8e3fff8b3b21c1a68de7cc172e9060e9b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "inline_equation", + "content": "W_{Q_i}, W_{K_i}" + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "inline_equation", + "content": "W_{V_i}" + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "text", + "content": " are projection matrices in the " + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "text", + "content": "-th head. The parameters for the " + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 525, + 287, + 549 + ], + "type": "text", + "content": "-th layer are generated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 565, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 565, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 70, + 565, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} G _ {l} ^ {t} = \\operatorname {C o n c a t} \\left(g \\left(\\theta_ {1} ^ {t}\\right), \\dots , g \\left(\\theta_ {N _ {l}} ^ {t}\\right)\\right) \\tag {3} \\\\ = \\operatorname {C o n c a t} \\left(H \\left(z _ {1, 0} ^ {t}; \\mathcal {H} ^ {t}\\right), \\dots , H \\left(z _ {N _ {l}, 0} ^ {t}; \\mathcal {H} ^ {t}\\right)\\right), \\\\ \\end{array}", + "image_path": "13ae5e68f1dbd324babcef50863597a47eed527b16937b4efb2f1f9db30f84f0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "text", + "content": " denotes the fully connected layer followed by a sigmoid function. " + }, + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\theta_{n}^{t}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "text", + "content": " is the first row of " + }, + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "inline_equation", + "content": "z_{n,N_b}^t" + }, + { + "bbox": [ + 47, + 604, + 287, + 640 + ], + "type": "text", + "content": " , which acts as the final embedding in a transformer architecture [8]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "content": "To refine old knowledge to align with the knowledge of the " + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "content": "-th task, let us denote the set of old parameters in the " + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "content": "-th layer as " + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\bar{W}_l^{1:t-1}" + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "content": ". We obtain the evolved parameters " + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\hat{W}_l^{1:t-1}" + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "content": " by fusing " + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\bar{W}_l^{1:t-1}" + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "inline_equation", + "content": "G_l^t" + }, + { + "bbox": [ + 47, + 640, + 287, + 691 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 698, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 698, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 117, + 698, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\hat {W} _ {l} ^ {1: t - 1} = \\bar {W} _ {l} ^ {1: t - 1} \\cdot G _ {l} ^ {t}, \\tag {4}", + "image_path": "14132732d0698aa17377a4bf36c3739b1dfd4076f4b8b7e8f591088aed1a0e61.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 541, + 142 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 541, + 142 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 541, + 142 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 541, + 142 + ], + "type": "image", + "image_path": "580be8affda3d993b9c77be535a414575ea27b86e97208bdaa156a6be0e9ba73.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 144, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 178 + ], + "type": "text", + "content": "Figure 3. The task-layer attention block (TLA) consists of a layer normalization, multi-head self-attention (MSA), and MLP with a single hidden layer. " + }, + { + "bbox": [ + 305, + 144, + 545, + 178 + ], + "type": "inline_equation", + "content": "z_{n,N_b}^t" + }, + { + "bbox": [ + 305, + 144, + 545, + 178 + ], + "type": "text", + "content": " is generated by " + }, + { + "bbox": [ + 305, + 144, + 545, + 178 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 305, + 144, + 545, + 178 + ], + "type": "text", + "content": " stacked TLAs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "text", + "content": "where we define " + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "text", + "content": " as the element-wise multiplication. The base network accommodates the evolved sets of parameters for " + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "text", + "content": " tasks together with the " + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "text", + "content": "-th parameter set as " + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{W}}^t = \\{\\hat{W}^1, \\dots, \\hat{W}^{t-1}, W^t\\}" + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "text", + "content": ". Finally, the loss function to learn the " + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 198, + 545, + 258 + ], + "type": "text", + "content": "-th task using the evolved parameter sets becomes" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 266, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 266, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 349, + 266, + 545, + 300 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t a s k}} = - \\sum_ {n = 1} ^ {N ^ {t}} y _ {n} ^ {t} \\log \\left(F \\left(x _ {n} ^ {t}; \\hat {\\mathcal {W}} ^ {t}\\right)\\right)). \\tag {5}", + "image_path": "4fdf8cebac5ea38f16bbb5dc3b4d59b36fc8f8845e1d9200c7ede01e8c660b1d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 308, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 334 + ], + "type": "text", + "content": "Another loss function to generate parameters from " + }, + { + "bbox": [ + 305, + 308, + 545, + 334 + ], + "type": "inline_equation", + "content": "H(\\cdot ;\\mathcal{H}^t)" + }, + { + "bbox": [ + 305, + 308, + 545, + 334 + ], + "type": "text", + "content": " similar to those learned in the previous task is defined as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 342, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 361, + 342, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\sum_ {k = 1} ^ {t - 1} \\sum_ {l = 1} ^ {L} \\| \\bar {G} _ {l} ^ {k} - G _ {l} ^ {k} \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "9dbf0b05faf935afe5a45c952ba6bb2b332f9ae6d97e50267a2eb3816bb44d5d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 384, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 384, + 545, + 408 + ], + "type": "inline_equation", + "content": "\\bar{G}_l^k" + }, + { + "bbox": [ + 305, + 384, + 545, + 408 + ], + "type": "text", + "content": " represents the set parameters generated from the hypernetwork after training the task " + }, + { + "bbox": [ + 305, + 384, + 545, + 408 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 305, + 384, + 545, + 408 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 415, + 469, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 469, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 469, + 428 + ], + "type": "text", + "content": "3.3. Sparsity-Inducing Generation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": "To renew the network without redundant parameters, we propose to generate sparse parameters in the hypernetwork. As " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "G_{l}^{t}" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": " is generated through the sigmoid function, they can be interpreted as the likelihood that each element of the old set of parameters is selected. Following the practice in [41], we minimize " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\sum_{l=1}^{L} -\\log(1 - G_{l}^{t})" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": " to achieve sparsity in " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "G_{l}^{t}" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": ". However, to prevent " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "G_{l}^{t}" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": " from becoming overly sparse, we ensure that parameters significantly impacting the loss remain survived. Inspired by the work in [24], we evaluate the impact of individual elements in " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "G_{l}^{t1}" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": " to the loss. We compute the set of all entries of the score " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "S_{l}^{t}" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": " by comparing the network losses computed with and without the inclusion of parameters in the learning process. To estimate the effect of removing a parameter on the loss function, we apply a Taylor expansion approximation of the network loss " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": " samples. We obtain the " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": "-th element of the score, " + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "inline_equation", + "content": "s_{k}" + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 632, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 632, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 311, + 632, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s _ {k} = \\mathcal {L} \\left(\\vec {G} _ {l} ^ {t} - M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t}\\right) - \\mathcal {L} \\left(\\vec {G} _ {l} ^ {t}\\right) \\tag {7} \\\\ \\simeq - M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t} \\cdot \\nabla_ {\\vec {G} _ {l} ^ {t}} \\mathcal {L} + \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left(\\frac {\\partial \\mathcal {L} _ {n}}{\\partial M _ {\\backslash (k)} \\cdot \\vec {G} _ {l} ^ {t}}\\right) ^ {2}, \\\\ \\end{array}", + "image_path": "3f1159ee9a51e136ee0789e7524bec9fc00ab96a7fb3dda64786ac22d0708a65.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 692, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 692, + 545, + 712 + ], + "type": "text", + "content": "1We denote the vectorized form of " + }, + { + "bbox": [ + 306, + 692, + 545, + 712 + ], + "type": "inline_equation", + "content": "G_{l}^{t}" + }, + { + "bbox": [ + 306, + 692, + 545, + 712 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 306, + 692, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\vec{G}_{l}^{t}" + }, + { + "bbox": [ + 306, + 692, + 545, + 712 + ], + "type": "text", + "content": " to represent individual entries." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18964" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "M_{\\backslash (k)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is an all-zero vector except that the " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "-th element is 1, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is the loss of the " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "-th example. We introduce an additional term minimizing the discrepancy between the scores of the complete set of parameters (i.e., the original scores) and the scores of the sparse set of parameters produced as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 153, + 287, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 52, + 153, + 287, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s p} = \\sum_ {l = 1} ^ {L} - \\log (1 - G _ {l} ^ {t}) + \\left\\| S _ {l} ^ {t} \\cdot \\left(1 - \\mathbb {1} (\\vec {G} _ {l} ^ {t} \\neq 0)\\right) \\right\\| _ {1}, \\tag {8}", + "image_path": "7e91eae82ef16d402da64249bfbf579b4869a3ebf7cc63b50fa601ecaae1792c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_1" + }, + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 46, + 194, + 287, + 219 + ], + "type": "text", + "content": "-norm and the indicator function, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": "Total Loss. The proposed framework learns the " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": "-th task with the three loss functions. To learn the " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": "-th task, we minimize the cross-entropy loss " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": " by updating " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "W^t" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": " in the base network. Also, " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "e^t" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\{C_l\\}_{l=1}^L" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": " are updated from two losses, " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sp}" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": ". Last, we update " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^t" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\{C_l\\}_{l=1}^L" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 46, + 219, + 287, + 304 + ], + "type": "text", + "content": " to preserve the previously generated parameters. The total loss is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 312, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 312, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 74, + 312, + 287, + 334 + ], + "type": "interline_equation", + "content": "\\min _ {W ^ {t}, \\mathcal {H} ^ {t}, e ^ {t}, \\{C _ {l} \\} _ {l = 1} ^ {L}} \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {s p} \\mathcal {L} _ {s p} + \\lambda_ {r e g} \\mathcal {L} _ {r e g}, \\tag {9}", + "image_path": "5ceb958c539bdbba260b23f7f9d131c2f7660f5bb2a254f669d322d0d4c81558.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 341, + 276, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 276, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 276, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 341, + 276, + 354 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + }, + { + "bbox": [ + 46, + 341, + 276, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 341, + 276, + 354 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 46, + 341, + 276, + 354 + ], + "type": "text", + "content": " are parameters to balance the losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": "Scalability. A hypernetwork is designed to generate parameters for every layer of the base network, and the parameters were composed of compact segments of a consistent size [13, 21]. While it can generate parameters for all layers, this approach requires a substantial quantity of layer embeddings, particularly for deeper layers. To mitigate this, we adjust the scale of the segments for each layer depending on the number of channels. However, it incurs a memory overhead stemming from the substantial dimensions of the heads. To address this, we apply a matrix decomposition technique [39] to generate two reduced sets of low-rank " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": ". By doing so, we achieve a reduction in memory requirements for layer embeddings, accompanied by a marginal increase in hypernetwork parameters. For detailed results, please refer to Section 4.5.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "type": "text", + "content": "4.1. Scenarios" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "We applied the proposed method, GrowBrain, to image classification and video action recognition problems under class incremental and task-incremental learning scenarios. Image classification. We conducted experiments using ImageNet [6] and five fine-grained datasets including CUBS [45], Stanford Cars [18], Flowers [31], Wikiart [36], and Sketch [11], where each dataset is considered a task. We call the datasets ImageNet+FG for short. In addition, we demonstrated the proposed method with the CIFAR-100. We divided the dataset into subsets of classes based on the random order provided by [35, 9] and considered" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 79, + 539, + 191 + ], + "blocks": [ + { + "bbox": [ + 361, + 69, + 489, + 79 + ], + "lines": [ + { + "bbox": [ + 361, + 69, + 489, + 79 + ], + "spans": [ + { + "bbox": [ + 361, + 69, + 489, + 79 + ], + "type": "text", + "content": "Table 1. Datasets used in this work." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 79, + 539, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 79, + 539, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 79, + 539, + 191 + ], + "type": "table", + "html": "
Image Dataset# Train Images# Test Images# Classes
ImageNet [6]1,287,16750,0001,000
CUBS [45]5,9945,794200
Stanford Cars [18]8,1448,041196
Flowers [31]2,0406,149102
Wikiart [36]42,12910,628195
Sketch [11]16,0004,000250
CIFAR-100 [19]50,00010,000100
Video Dataset# Train Videos# Test Videos# Classes
ActivityNet [3]10,0244,926200
UCF-101 [40]9,2802,720101
", + "image_path": "633e537d4b70efcde203f3f7ccb546bf9bba75488e54935a641ce5713b8b6bd1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 313, + 210, + 539, + 271 + ], + "blocks": [ + { + "bbox": [ + 336, + 200, + 511, + 210 + ], + "lines": [ + { + "bbox": [ + 336, + 200, + 511, + 210 + ], + "spans": [ + { + "bbox": [ + 336, + 200, + 511, + 210 + ], + "type": "text", + "content": "Table 2. The specification of the hypernetwork." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 210, + 539, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 210, + 539, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 539, + 271 + ], + "type": "table", + "html": "
Base networkHypernetwork
Embed dim (D)Hidden dim in MLP# Heads (Nh)# Blocks (Nb)
ResNet-506412887
ResNet-34
ResNet-18326485
", + "image_path": "21b7498a5d3c3a748e3447ad45f4a1bd2d45c1a7a3c8dbf9981ce15644901e3e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 291, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 363 + ], + "type": "text", + "content": "each subset a task. Video action recognition. We used a trimmed version of ActivityNet [3] and UCF-101 [40] video datasets. To conduct experiments, we divided each dataset into subsets containing randomly selected classes provided by [43] and treated each subset as a task. We summarize the datasets as mentioned above in Table 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 371, + 437, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 437, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 437, + 384 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 389, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 449 + ], + "type": "text", + "content": "Image classification. ImageNet+FG and CIFAR-100 were resized to " + }, + { + "bbox": [ + 304, + 389, + 545, + 449 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 389, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 389, + 545, + 449 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 389, + 545, + 449 + ], + "type": "text", + "content": " pixels, respectively. We applied random cropping and horizontal flip as augmentation. For those datasets, we used the ResNet-50 [14] and ResNet-18 as the base networks, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "text", + "content": "Video action recognition. We used TSN [46] with the ResNet-34 backbone network for video action recognition. Each video was split into three segments of equal duration, and a frame was randomly selected from each segment. The action class prediction was performed by aggregating the predictions of the selected frames with a consensus function. We used an average of predictions as a consensus function [46]. The video datasets were augmented with random cropping and horizontal flipping following [46]. We resized each frame of ActivityNet and UCF-101 to " + }, + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 449, + 545, + 569 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Table 2 reports the base networks and corresponding hypernetworks, which include the embedding dimension, the hidden dimension of the MLP, the number of heads, and the number of TLA blocks. We set the rank " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " for matrix decomposition as 7. We trained the base and hypernetwork using stochastic gradient descent. In all our experiments, we set " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " to 1 and " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " to 0.1, except ImageNet+FG where we set " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " to 1. We report the average results over five independent runs for all the experiments. To enable class-incremental learning in scenarios where the task identity is unknown, the proposed method first predicts the task index and then conducts classification. Following [44], we predict the task" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18965" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 536, + 196 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "type": "text", + "content": "Table 3. Performance of the class-incremental learning scenario using ImageNet and five fine-grained datasets (ImageNet+FG), where the average accuracy was measured after training each task. # P represents the average number of parameters " + }, + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "type": "inline_equation", + "content": "(\\times 10^{6})" + }, + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "type": "text", + "content": " to perform the tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 536, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 536, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 536, + 196 + ], + "type": "table", + "html": "
MethodImageNetCUBSStanford CarsFlowersWikiartSketch
Acc#PAcc#PAcc#PAcc#PAcc#PAcc#P
EWC-On76.123.431.623.428.623.429.223.423.123.411.323.4
LwF76.123.468.423.458.223.455.823.449.223.428.423.4
iCaRL76.123.470.223.461.523.452.723.448.323.442.323.4
PackNet75.712.562.813.365.214.066.514.654.315.245.315.7
Piggyback76.123.460.122.754.822.256.422.049.721.747.421.4
WSN74.712.559.612.542.212.529.212.531.012.525.012.5
H275.712.569.513.067.312.968.413.061.812.948.712.6
GrowBrain75.712.569.211.069.310.371.39.864.39.660.19.4
", + "image_path": "35848c83cbeeb73fb2f371f4ec1d28c5bdcc29126789e29cc57567bd7dfc3156.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 205, + 208, + 366 + ], + "blocks": [ + { + "bbox": [ + 52, + 205, + 208, + 366 + ], + "lines": [ + { + "bbox": [ + 52, + 205, + 208, + 366 + ], + "spans": [ + { + "bbox": [ + 52, + 205, + 208, + 366 + ], + "type": "image", + "image_path": "a44403d2458280411c417bba241f1ad75828a91eca21a4190b80f5006df31cb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 374, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 374, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 544, + 396 + ], + "type": "text", + "content": "Figure 4. Accuracy of class-incremental learning for image classification on ImageNet+FG (left), video action recognition on UCF-101 (middle) and ActivityNet (right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 205, + 375, + 368 + ], + "blocks": [ + { + "bbox": [ + 216, + 205, + 375, + 368 + ], + "lines": [ + { + "bbox": [ + 216, + 205, + 375, + 368 + ], + "spans": [ + { + "bbox": [ + 216, + 205, + 375, + 368 + ], + "type": "image", + "image_path": "9e98dedb86e4c1568827fdaec08a34720fcb9f3adf90a47b284ad3de833a6892.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 384, + 205, + 541, + 368 + ], + "blocks": [ + { + "bbox": [ + 384, + 205, + 541, + 368 + ], + "lines": [ + { + "bbox": [ + 384, + 205, + 541, + 368 + ], + "spans": [ + { + "bbox": [ + 384, + 205, + 541, + 368 + ], + "type": "image", + "image_path": "ee8d531d8ec70c4bb451d10a2eda8c0c6e3279c02b4bb822f3edb3ddf8a53f0b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "type": "text", + "content": "index using the network with the least entropy that has been learned thus far." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 446, + 196, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 196, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 196, + 460 + ], + "type": "text", + "content": "4.3. Class-incremental learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 464, + 164, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 164, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 164, + 477 + ], + "type": "text", + "content": "4.3.1 Image classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": "We evaluated the proposed method of evolving the previous knowledge to the class-incremental learning scenario using ImageNet+FG. We compared with regularization methods, EWC-On [37] and LwF [22], and data rehearsal methods, iCaRL [35], and parameter isolation methods, PackNet [28], Piggyback [27], WSN [16], and " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": " [15]. The data rehearsal methods stored 1,000 raw instances (taking 150 " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{MB}^2" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": "). We applied the same task prediction mechanism as ours to the comparing methods, excluding iCaRL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "text", + "content": "Table 3 shows the class-incremental learning results for ImageNet and five fine-grained datasets. Overall, the proposed method outperforms competitors (more detailed results are shown on the left of Figure 4). The regularization approaches, EwC-On and LwF, give unsatisfactory results with performance gaps of " + }, + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "inline_equation", + "content": "48.8\\%" + }, + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "inline_equation", + "content": "31.7\\%" + }, + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "text", + "content": " compared to ours, respectively. LwF shows a " + }, + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 46, + 591, + 288, + 687 + ], + "type": "text", + "content": " accuracy drop before and after learning Sketch, which is the most significant" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": "among the comparing methods. The data rehearsal method, iCaRL, shows a " + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "inline_equation", + "content": "17.1\\%" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": " performance drop from GrowBrain despite storing 1,000 old samples. PackNet shows large performance drops of " + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "inline_equation", + "content": "12.2\\%" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "inline_equation", + "content": "9.0\\%" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": " after training on WikiArt and Sketch, respectively. This is because negative interference occurs when using all previous sets of parameters optimized for different task domains. The methods of associating prior knowledge, Piggyback, WSN, and " + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": ", show notable performance drop (" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "inline_equation", + "content": "35.1\\%" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 309, + 416, + 545, + 584 + ], + "type": "text", + "content": ", respectively) compared to the proposed method that improves the previous knowledge from the new task. The results demonstrate that the proposed approach surpasses a wide range of continual learning methods, confirming its superiority." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 598, + 443, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 443, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 443, + 611 + ], + "type": "text", + "content": "4.3.2 Video action recognition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We applied our method and the competitors to the video action recognition tasks using the benchmark sets proposed for continual learning [43]. For this experiment, iCaRL was excluded from the comparison since it demands approximately 2TB and 52GB of memory for replay [43] on ActivityNet and UCF-101, respectively. Since Piggyback requires a pre-trained backbone, we pre-trained the backbone model [14] using the validation set of UCF-101 for video action" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "2 We calculated the rehearsal memory assuming that each image (or frame of video) has " + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18966" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 91, + 493, + 203 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 90 + ], + "type": "text", + "content": "Table 4. Results of the class-incremental learning scenario using the trimmed version of UCF-101 and ActivityNet, where the Avg and Last accuracies for the time steps 5 and 10." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 91, + 493, + 203 + ], + "lines": [ + { + "bbox": [ + 100, + 91, + 493, + 203 + ], + "spans": [ + { + "bbox": [ + 100, + 91, + 493, + 203 + ], + "type": "table", + "html": "
MethodUCF-101ActivityNet
T5T10T5T10
AvgLast# PAvgLast# PAvgLast# PAvgLast# P
EWC-On48.725.821.232.713.921.238.420.721.226.612.421.2
LwF47.027.721.233.415.421.244.129.121.232.718.421.2
PackNet59.354.211.552.540.415.242.430.211.532.419.015.2
Piggyback65.451.720.453.536.520.422.212.720.515.67.520.4
WSN69.158.810.161.149.710.5------
H263.654.88.356.947.011.437.422.53.426.813.32.1
GrowBrain79.372.27.273.564.78.942.829.911.533.220.615.2
", + "image_path": "dfd6c5ed318d4c6a5095845e10c0d7701737161be5874e0c4ff195be3741aec4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 223, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 287, + 270 + ], + "type": "text", + "content": "recognition experiments. We report the " + }, + { + "bbox": [ + 47, + 223, + 287, + 270 + ], + "type": "inline_equation", + "content": "Avg" + }, + { + "bbox": [ + 47, + 223, + 287, + 270 + ], + "type": "text", + "content": " and Last accuracies [9] for the time steps 5 and 10. Figure 4 shows the Last accuracy measured on the UCF-101 and ActivityNet datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "text", + "content": "UCF-101. The results of the class-incremental learning scenario using the UCF-101 dataset are reported in Table 4 (left). The aspects of the experimental results on UCF-101 of the compared approaches are similar to those of image classification. Despite utilizing a large number of parameters, regularization approaches, EWC-On and LwF, give unsatisfactory performance compared to parameter isolation approaches. PackNet, which uses all parameters from previous tasks, performs less than other parameter isolation approaches, WSN and " + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "text", + "content": ", using selectively chosen parameters associated with previous knowledge. The proposed method shows the highest performance compared to other comparison methods. Ours outperforms WSN, " + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "text", + "content": ", PackNet, and Piggyback with the significant accuracy gap on the Last measure from " + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "inline_equation", + "content": "15.0\\%" + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "inline_equation", + "content": "28.2\\%" + }, + { + "bbox": [ + 47, + 283, + 287, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "content": "ActivityNet. Table 4 (right) reports the results on ActivityNet. WSN was excluded from the experiment because it did not give promising results due to the lack of data. Overall, the results of ActivityNet show a different pattern compared to those of UCF-101. The regularization method, LwF, outperforms the parameter isolation method, PackNet, with " + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "content": " Avg accuracy gap measured after training the 5-th and 10-th tasks. The proposed approach performs better than Piggyback, PackNet, and " + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "content": " by large margins of " + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 47, + 474, + 287, + 713 + ], + "type": "text", + "content": ", respectively, in terms of the Last accuracy measured after training 10-th task. The proposed method outperforms other competitors for image classification and video action recognition problems. In contrast to the UCF-101 dataset, where the proposal can acquire new knowledge by using a small number of parameters from previous tasks and achieve excellent results, the ActivityNet dataset requires previous knowledge to effectively train the current task due to its inadequate data (the number of videos per class is only half compared to UCF-101)." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 253, + 545, + 361 + ], + "blocks": [ + { + "bbox": [ + 306, + 220, + 545, + 252 + ], + "lines": [ + { + "bbox": [ + 306, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 545, + 252 + ], + "type": "text", + "content": "Table 5. Task-incremental learning scenarios using video action recognition and image classification datasets, where average accuracy and parameter consumption are reported." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 253, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 361 + ], + "type": "table", + "html": "
MethodImage classificationVideo action recognition
ImageNet+FGCIFAR-100UCF-101ActivityNet
Acc#PAcc#PAcc#PAcc#P
EWC-On55.323.444.511.165.821.261.521.2
LwF53.623.439.711.171.221.265.821.2
PackNet80.215.782.110.492.215.265.315.2
Piggyback82.221.475.610.586.620.442.120.4
WSN61.712.584.65.692.710.5--
H282.812.685.91.194.511.444.02.1
GrowBrain83.09.486.82.596.78.966.915.2
", + "image_path": "a26e33be965b83186eb7607b0d109f30cfdabf23d398dc9da31f6e0fbbf36836.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 383, + 452, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 452, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 452, + 396 + ], + "type": "text", + "content": "4.4. Task-incremental learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "text", + "content": "Additionally, we conducted experiments on the task-incremental learning scenario where task-id is given during the inference step. In this scenario, we used ImageNet+FG and CIFAR-100 as the image classification datasets. Also, we used ActivityNet and UCF-101 for video action recognition. We divided CIFAR-100 into 10 tasks based on [35]. We followed the same experimental setting as in Tables 3 and 4. For the experiments of Piggyback on CIFAR-100, the backbone network was pre-trained on a randomly selected subset (40%) of the CIFAR-100 training set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "content": "Image classification. Table 5 reports the results of task-incremental learning conducted on ImageNet+FG and CIFAR-100. Even if the regularization methods, EwC-On and LwF, try to maintain previous knowledge, they perform significantly less than the parameter isolation methods. GrowBrain consistently achieves high performance compared to other parameter isolation methods, PackNet, Piggyback, WSN, and " + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "content": ", in both image classification experiments (e.g., " + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "content": " performance gap on ImageNet+FG and CIFAR-100 with the most competitive method, " + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 304, + 521, + 545, + 652 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Video action recognition. Table 5 reports the results of task-incremental learning conducted on UCF-101 and ActivityNet. Overall, the results are similar to those of class-incremental learning scenarios. The regularization method, LwF, achieves higher performance than PackNet on Activ" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18967" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 113, + 287, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 112 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 112 + ], + "type": "text", + "content": "Table 6. Ablation study of the proposed method for image classification and video action recognition tasks, where average accuracy, the number of parameters for the base network and hypernetwork (in millions), and the number of layer embeddings are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 113, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 48, + 113, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 113, + 287, + 167 + ], + "type": "table", + "html": "
MethodImageNet+FGUCF-101
Acc# (F + H)# cnAcc# (F + H)# cn
Ours60.19.4 + 2.813364.78.9 + 2.858
w/o transform [28]45.315.7 + 0040.415.2 + 00
w/o sparsity58.115.7 + 2.813364.014.1 + 2.858
w/o scalability59.815.7 + 2.9326264.711.0 + 2.9614
", + "image_path": "30dd18089bb50a9adbb811434576c333f65769981cb84407cd1b3a41bb00e802.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "text", + "content": "ityNet. The proposed method outperforms both parameter isolation and regularization methods. Notably, our method outperforms WSN, PackNet, and " + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2" + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "text", + "content": " with a performance gap of " + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 188, + 287, + 236 + ], + "type": "text", + "content": ", respectively, on UCF-101." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 244, + 110, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 244, + 110, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 244, + 110, + 257 + ], + "type": "text", + "content": "4.5. Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 263, + 143, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 143, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 143, + 275 + ], + "type": "text", + "content": "4.5.1 Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our method, we conducted an ablation study. We compared ours by removing transformation of previous knowledge [28], sparsity-inducing generation " + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "inline_equation", + "content": "(\\lambda_{sp} = 0)" + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "text", + "content": ", and scalability of hypernetwork, i.e., channel adaptive segments and matrix decomposition. We applied these methods to class-incremental learning scenarios using the ImageNet+FG and the UCF-101 datasets. We report the Last accuracy measured after training all tasks, the average parameter consumption of the base network " + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "text", + "content": " and hypernetwork " + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "inline_equation", + "content": "(H)" + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "text", + "content": " to perform each task, and the number of layer embeddings " + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "inline_equation", + "content": "c_{n}" + }, + { + "bbox": [ + 46, + 282, + 287, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "text", + "content": "Table 6 reports the results of the ablation study. Our method outperforms that without transformation by performance gaps of " + }, + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "inline_equation", + "content": "14.8\\%" + }, + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "inline_equation", + "content": "24.3\\%" + }, + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "text", + "content": " for those datasets, emphasizing that the evolution of prior knowledge is critical in enhancing performance. For video action recognition using UCF-101, the method without sparsity requires more parameters and performs less than the proposed method. The proposed method for ImageNet+FG achieves higher performance while taking fewer parameters than the nonsparse alternative. The reason is that sparsity-inducing hypernetwork prevents the generation of redundant parameters and regularizes the generated parameters to achieve generalization performance. The method without scalability requires 24 and 10.5 times larger layer embeddings on ImageNet+FG and UCF-101 than the proposed method, respectively, without accompanying performance improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 622, + 138, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 138, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 138, + 635 + ], + "type": "text", + "content": "4.5.2 Impact of " + }, + { + "bbox": [ + 47, + 622, + 138, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": "We further conducted a study investigating the impact of " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": ", which regulates the sparsity of generated parameters. We experimented with varying values of " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " in the ImageNet+FG experiment under the class-incremental scenario using the same setup presented in Table 3. Figure 5 shows the parameter consumption (left) and average ac" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 425, + 153 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 425, + 153 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 425, + 153 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 425, + 153 + ], + "type": "image", + "image_path": "89486965460d5420072724b4798da3eb4be9cabae17d24a4a876116dc731659d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "text", + "content": "Figure 5. Parameter consumption and average accuracy of the proposed method based on varying " + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "text", + "content": " in class-incremental learning for ImageNet+FG." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 70, + 543, + 153 + ], + "blocks": [ + { + "bbox": [ + 426, + 70, + 543, + 153 + ], + "lines": [ + { + "bbox": [ + 426, + 70, + 543, + 153 + ], + "spans": [ + { + "bbox": [ + 426, + 70, + 543, + 153 + ], + "type": "image", + "image_path": "e1fe5106a3da215418b6ab96e9515f51cfa3834f5cf3824f27c28c231544552a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": "curacy (right) with respect to different " + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_{sp}" + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": " after training all tasks. Ours with " + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_{sp} = 0.1" + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": " uses a similar amount of parameters as the baseline [28] but outperforms the competing approach by leveraging enhanced prior knowledge. Notably, after training all tasks, " + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_{sp} = 1.0" + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "inline_equation", + "content": "9.4\\mathrm{M}" + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": " parameters while showing " + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": " performance improvement compared to " + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_{sp} = 0.1" + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": " consuming " + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "inline_equation", + "content": "15.7\\mathrm{M}" + }, + { + "bbox": [ + 304, + 223, + 545, + 307 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 335, + 378, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 378, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 378, + 346 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "text", + "content": "We have proposed a novel approach, GrowBrain, to transform learned old knowledge to absorb the information of newer tasks in parameter isolation-based continual learning. GrowBrain reinforces the previous knowledge with a transformer-based hypernetwork that generates optimized parameters for new tasks while eliminating redundant parameters that show negligible impact on the loss. The proposed method overcomes the limitation of using fixed old knowledge that can constrain the ability to capture novel representations while taking fewer parameters. In the experiments, we have evaluated the proposal compared to a variety of continual learning methods in class- and task-incremental learning scenarios for image classification and video action recognition. Experimental results show that the proposed approach of leveraging transformed prior knowledge generates parameters adapted to the new task without losing the prior knowledge, resulting in significant performance improvements over its strong competitors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT2002-05, Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (2022-0-01341, Artificial Intelligence Graduate School Program(Chung-Ang University)), and BK21 FOUR (Fostering Outstanding Universities for Research) Program funded by Ministry of Education of Korea (No.I22SS7609062)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18968" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3931-3940, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In European Conference on Computer Vision, pages 144-161. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 268 + ], + "type": "text", + "content": "[3] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[4] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 556-572. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "text", + "content": "[5] Arslan Chaudhry, Ranzato Marc'Aurelio, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In 7th International Conference on Learning Representations, ICLR 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 468 + ], + "type": "text", + "content": "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 534 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "type": "text", + "content": "[9] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[10] Benjamin Ehret, Christian Henning, Maria Cervera, Alexander Meulemans, Johannes Von Oswald, and Benjamin F Grewe. Continual learning in recurrent neural networks. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[11] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Transactions on Graphics, 31(4):1-10, 2012. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[12] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[13] David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv preprint arXiv:1609.09106, 2016. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 184 + ], + "type": "text", + "content": "[15] Hyundong Jin and Eunwoo Kim. Helpful or harmful: Intertask association in continual learning. In European Conference on Computer Vision, pages 519-535. Springer, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 251 + ], + "type": "text", + "content": "[16] Haeyong Kang, Rusty John Lloyd Mina, Sultan Rizky Hikmawan Madjid, Jaehong Yoon, Mark Hasegawa-Johnson, Sung Ju Hwang, and Chang D Yoo. Forget-free continual learning with winning subnetworks. In International Conference on Machine Learning, pages 10734-10750. PMLR, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 547, + 329 + ], + "type": "text", + "content": "[17] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America, 114(13):3521-3526, 2017. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 375 + ], + "type": "text", + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561. IEEE, 2013. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 399 + ], + "type": "text", + "content": "[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[20] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forgetting by incremental moment matching. Advances in Neural Information Processing Systems, 30:4652-4662, 2017. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 444, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 499 + ], + "type": "text", + "content": "[21] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. DHP: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "text", + "content": "[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. In European Conference on Computer Vision, pages 614-629. Springer, 2016. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "text", + "content": "[23] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "text", + "content": "[24] Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group fisher pruning for practical network compression. In International Conference on Machine Learning, pages 7021-7032. PMLR, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[25] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Adaptive aggregation networks for class-incremental learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2544-2553, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6470-6479, 2017. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18969" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[27] Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In Proceedings of the European Conference on Computer Vision, pages 67-82, 2018. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[28] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 207 + ], + "type": "text", + "content": "[29] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In *Psychology of learning and motivation*, volume 24, pages 109–165. Elsevier, 1989. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 288, + 241 + ], + "type": "text", + "content": "[30] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 243, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 288, + 286 + ], + "type": "text", + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 288, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 288, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 288, + 288, + 342 + ], + "type": "text", + "content": "[32] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11321-11329, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 344, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 344, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 344, + 288, + 388 + ], + "type": "text", + "content": "[33] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13698-13707, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 389, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 389, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 389, + 288, + 453 + ], + "type": "text", + "content": "[34] Dripta S Raychaudhuri, Yumin Suh, Samuel Schulter, Xiang Yu, Masoud Faraki, Amit K Roy-Chowdhury, and Manmohan Chandraker. Controllable dynamic multi-task architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10955-10964, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 456, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 288, + 511 + ], + "type": "text", + "content": "[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. iCaRL: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 512, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 288, + 555 + ], + "type": "text", + "content": "[36] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. International Journal for Digital Art History, (2), 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 557, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 288, + 622 + ], + "type": "text", + "content": "[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International Conference on Machine Learning, pages 4528-4537. PMLR, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 288, + 668 + ], + "type": "text", + "content": "[38] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 2994-3003, 2017. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 288, + 713 + ], + "type": "text", + "content": "[39] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 498 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[41] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. AdaShare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems, 33:8728-8740, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[43] Andrés Villa, Kumail Alhamoud, Victor Escorcia, Fabian Caba, Juan León Alcazar, and Bernard Ghanem. vCLIMB: A novel video class incremental learning benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19035-19044, 2022. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 263, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 306 + ], + "type": "text", + "content": "[44] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F Grewe. Continual learning with hypernetworks. In 8th International Conference on Learning Representations, 2020. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 340 + ], + "type": "text", + "content": "[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "text", + "content": "[46] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 398, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 398, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 398, + 545, + 452 + ], + "type": "text", + "content": "[47] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 454, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 498 + ], + "type": "text", + "content": "[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3987-3995, 2017. 1, 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18970" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_content_list.json b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1969b1ba935040ae9b619bf100e778e77dc22e56 --- /dev/null +++ b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_content_list.json @@ -0,0 +1,2024 @@ +[ + { + "type": "text", + "text": "Guided Motion Diffusion for Controllable Human Motion Synthesis", + "text_level": 1, + "bbox": [ + 142, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korrawe Karunratanakul1", + "bbox": [ + 122, + 179, + 326, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Konpat Preechakul2", + "bbox": [ + 348, + 180, + 509, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Supasorn Suwajanakorn2", + "bbox": [ + 527, + 180, + 728, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siyu Tang", + "bbox": [ + 750, + 180, + 839, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zürich, Switzerland $^{2}$ VISTEC, Thailand", + "bbox": [ + 290, + 208, + 679, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://korrawe.github.io/gmd-project/", + "bbox": [ + 310, + 229, + 651, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/478f1f5716b592c29ae9039e4fdef8b13da30f3560adf01a83d493c6779e945e.jpg", + "image_caption": [ + "(a) Text only" + ], + "image_footnote": [], + "bbox": [ + 86, + 286, + 285, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a55b1394bc0dd7e07665aac903cea39797258e11936deb30b1fea9e7da953fc.jpg", + "image_caption": [ + "(b) Text + Trajectory" + ], + "image_footnote": [], + "bbox": [ + 287, + 286, + 486, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af433ba802f473570529c40f23fb907d0a7594ffe095492dead1705ae8b78ef9.jpg", + "image_caption": [ + "(c) Text + Keyframe locations" + ], + "image_footnote": [], + "bbox": [ + 488, + 286, + 683, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b848983346230cd514c9ca12b5b1a33e4396fd4c66322b8f89cbee9bbfbbc421.jpg", + "image_caption": [ + "(d) Text + Keyframe locations +Obstacle avoidance", + "Figure 1. Our proposed Guided Motion Diffusion (GMD) can generate high-quality and diverse motions given a text prompt and a goal function. We demonstrate the controllability of GMD on four different tasks, guided by the following conditions: (a) text only, (b) text and trajectory, (c) text and keyframe locations (double circles), and (d) with obstacle avoidance (red-cross areas represent obstacles). The darker the colors, the later in time." + ], + "image_footnote": [], + "bbox": [ + 684, + 286, + 883, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 530, + 313, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Denoising diffusion models have shown great promise in human motion synthesis conditioned on natural language descriptions. However, integrating spatial constraints, such as pre-defined motion trajectories and obstacles, remains a challenge despite being essential for bridging the gap between isolated human motion and its surrounding environment. To address this issue, we propose Guided Motion Diffusion (GMD), a method that incorporates spatial constraints into the motion generation process. Specifically, we propose an effective feature projection scheme that manipulates motion representation to enhance the coherency between spatial information and local poses. Together with a new imputation formulation, the generated motion can reliably conform to spatial constraints such as global motion trajectories. Furthermore, given sparse spatial constraints (e.g. sparse keyframes), we introduce a new dense guidance approach to turn a sparse signal, which is susceptible to being ignored during the reverse steps, into denser signals to guide the generated motion to the given constraints. Our extensive experiments justify the development of GMD, which achieves a significant improvement over state-of-the-art methods in text-based motion generation while allowing control of the synthesized motions with spatial constraints.", + "bbox": [ + 75, + 561, + 472, + 910 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 530, + 630, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, denoising diffusion models have emerged as a promising approach for human motion generation [11, 62, 64] outperforming other alternatives such as GAN or VAE in terms of both quality and diversity [7, 51, 58]. Several studies have focused on generating motion based on expressive text prompts [7, 51], or music [52, 64]. The state-of-the-art motion generation methods, such as MDM [51], utilize classifier-free guidance to generate motion conditioned on text prompts. However, incorporating spatial constraints into diffusion models remains underexplored. Human motions consist of both semantic and spatial information, where the semantic aspect can be described using natural languages or action labels and the spatial aspect governs physical interaction with surroundings. To generate realistic human motion in a 3D environment, both aspects must be incorporated. Our experiments show that simply adding spatial constraint guidance, such as global trajectories, into the state-of-the-art models or using imputation and in-painting approaches do not yield satisfactory results.", + "bbox": [ + 496, + 561, + 890, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We identify two main issues that make the motion diffusion models likely to ignore the guidance when conditioned on spatial objectives: the sparseness of global orientation", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2151", + "bbox": [ + 482, + 949, + 513, + 960 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7447fc7616774b2d60110837a33d8a1f87f13a593266904511ff55fc8a4004c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 88, + 465, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/40e7fbf622dd39aa2616e07e88d250edcf1a17718619671b817c7348d117b601.jpg", + "image_caption": [ + "Figure 2. We tackle the problem of spatially conditioned motion generation with GMD, depicted in a). Our main contributions are b) Emphasis projection, for better trajectory-motion coherence, and c) Dense signal propagation, for a more controllable generation even under sparse guidance signal." + ], + "image_footnote": [], + "bbox": [ + 78, + 263, + 464, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the motion representation and sparse frame-wise guidance signals. By design, the diffusion models are a denoising model that consecutively denoises the target output over multiple steps. With sparse guidance, a small portion of the output that receives guidance will be inconsistent with all other parts that do not, therefore, are more likely to be treated as noise and discarded in subsequent steps.", + "bbox": [ + 75, + 474, + 468, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, the sparseness within a frame is a result of common motion representations that separate local pose information, like joint rotations, from global orientations, such as pelvis translations and rotations [43], usually with more focus on local poses. For instance, the common motion representation [14] uses 4 values to represent global orientation and 259 values for local pose in each frame. Such imbalance can cause the model to focus excessively on local pose information, and consequently, perceive guided global orientation as noise, resulting in a discrepancy such as foot skating.", + "bbox": [ + 75, + 582, + 468, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, in many applications such as character animation, gaming, and virtual reality, the spatial control signals are defined on only a few keyframes such as target locations on the ground. We show that the current diffusion-based motion generation models struggle to follow such sparse guidance as doing so is equivalent to guiding an image diffusion model with only a few pixels. As a result, either the guidance at the provided keyframes will be ignored during the denoising process or the output motion will contain an artifact where the character warps to satisfy the guidance only in those specific keyframes.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively incorporate sparse spatial constraints into the motion generation process, we propose GMD, a novel and principled Guided Motion Diffusion model. To alleviate the discrepancy between local pose and global orientation in the guided denoising steps, we introduce emphasis projection, a general representation manipulation method that we use to increase the importance of spatial information during training. Additionally, we derive a new imputation and inpainting formulation that enables the existing inpainting techniques to operate in the projected space, which we leverage to generate significantly more coherent motion under guidance by spatial conditions. Then, to address the highly sparse guidance, we draw inspiration from the credit assignment problem in Reinforcement Learning [50, 54], where sparse rewards can be distributed along a trajectory to allow for efficient learning [3]. Our key insight is that motion denoisers, including the diffusion model itself, can be used to expand the spatial guidance signal at a specific location to its neighboring locations without any additional model. By turning a sparse signal into a dense one by backpropagating through a denoiser, it enables us to achieve high-quality controllable motion synthesis, even with extremely sparse guidance signals.", + "bbox": [ + 496, + 90, + 890, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (1) Emphasis projection, a method to adjust relative importance between different parts of the representation vector, which we use to encourage coherency between spatial information and local poses to allow spatial guidance. (2) Dense signal propagation, a conditioning method to tackle the sparse guidance problem. (3) GMD, an effective spatially controllable motion generation method that enables the unexplored synthesizing of motions based on free-text and spatial conditioning by integrating the above contributions into our proposed Unet-based architecture. We provide extensive analysis to support our design decisions and show the versatility of GMD on three tasks: trajectory conditioning, keyframe conditioning, and obstacle avoidance. Additionally, GMD's model also significantly outperforms the state-of-the-art in traditional text-to-motion tasks.", + "bbox": [ + 496, + 439, + 890, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 699, + 640, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based probabilistic generative models (DPM). DPMs [20,47,48,49] have gained significant attention in recent years due to their impressive performance across multiple fields of research. They have been used for tasks such as image generation [12], image super-resolution [30, 46], speech synthesis [26, 26, 39], video generation [19, 22], 3D shape generation [38, 55], and reinforcement learning [23].", + "bbox": [ + 496, + 732, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The surge in interest in DPMs may be attributed to their impressive controllable generation capabilities, including text-conditioned generation [42, 44, 45] and image editing [4, 6, 9, 18, 34]. Latent diffusion models (LDM) are an-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2152", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "other area of interest, which includes representation learning [27,40] and more efficient modeling techniques [7,44].", + "bbox": [ + 75, + 90, + 467, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, DPMs exhibit a high degree of versatility in terms of conditioning. There are various methods for conditional generation, such as imputation and inpainting [9, 10, 34], classifier guidance [10, 12], and classifier-free guidance [21, 42, 44, 45]. Inpainting and classifier guidance can be applied to any pretrained DPM, which extends the model's capabilities further without the need for retraining.", + "bbox": [ + 75, + 122, + 467, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human motion generation. The goal of the human motion generation task is to generate motions based on the conditioning signals. Various conditions have been explored such as partial poses [13, 17, 51], trajectories [24, 53, 60], images [8,43], music [28,29,31], text [1, 14, 15, 25, 37], objects [56], action labels [16,36], or unconditioned [35,57,61,63]. Recently, many diffusion-based motion generation models have been proposed [11, 25, 32, 62, 64] and demonstrate better quality compared to alternative models such as GAN or VAE. Employing the CLIP model [41], these models showed great improvements in the challenging text-to-motion generation task [7,58,59] as well as allowing conditioning on partial motions [51] or music [2, 52]. However, they do not support conditioning signals that are not specifically trained, for example, following keyframe locations or avoiding obstacles. Maintaining the capabilities of the diffusion models, we propose methods to enable spatial guidance without retraining the model for each new objective.", + "bbox": [ + 75, + 234, + 467, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 76, + 525, + 205, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion-based generative models", + "text_level": 1, + "bbox": [ + 76, + 550, + 375, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based probabilistic generative models (DPMs) are a family of generative models that learn a sequential denoising process of an input $\\mathbf{x}_t$ with varying noise levels $t$ . The noising process of DPM is defined cumulatively as $q(\\mathbf{x}_t|\\mathbf{x}_0) = \\mathcal{N}(\\sqrt{\\alpha_t}\\mathbf{x}_0,(1 - \\alpha_t)\\mathbf{I})$ , where $\\mathbf{x}_0$ is the clean input, $\\alpha_{t} = \\prod_{s = 1}^{t}(1 - \\beta_{s})$ , and $\\beta_{t}$ is a noise scheduler. The denoising model $p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)$ with parameters $\\theta$ learns to reverse the noising process by modeling the Gaussian posterior distribution $q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0)$ . DPMs can map a prior distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ to any distribution $p(\\mathbf{x})$ after $T$ successive denoising steps.", + "bbox": [ + 75, + 575, + 467, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To draw samples from a DPM, we start from a sample $\\mathbf{x}_T$ from the prior distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . Then, for each $t$ , we sample $\\mathbf{x}_{t - 1}\\sim \\mathcal{N}(\\mu_t,\\Sigma_t)$ until $t = 0$ , where", + "bbox": [ + 75, + 744, + 467, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t} = \\frac {\\sqrt {\\alpha_ {t - 1}} \\beta_ {t}}{1 - \\alpha_ {t}} \\mathbf {x} _ {0} + \\frac {\\sqrt {1 - \\beta_ {t}} (1 - \\alpha_ {t - 1})}{1 - \\alpha_ {t}} \\mathbf {x} _ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 804, + 467, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $\\Sigma_{t}$ is a variance scheduler of choice, usually $\\Sigma_{t} = \\frac{1 - \\alpha_{t-1}}{1 - \\alpha_{t}}\\beta_{t}$ [20]. $\\mathbf{x}_0$ in Eq. 1 is the prediction from a denoising model. For an $\\epsilon_{\\theta}$ model, $\\mathbf{x}_0 = \\frac{1}{\\sqrt{\\alpha_t}}\\mathbf{x}_t + \\frac{\\sqrt{1 - \\alpha_t}}{\\sqrt{\\alpha_t}}\\epsilon_\\theta (\\mathbf{x}_t)$ .", + "bbox": [ + 76, + 849, + 467, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are multiple choices for the denoising model to predict including the clean input $\\mathbf{x}_0$ , the noise $\\epsilon$ , and the one-step denoised target $\\mu_t$ . An $\\mathbf{x}_{0,\\theta}$ model is trained using the squared loss to the clean input $\\| \\mathbf{x}_{0,\\theta}(\\mathbf{x}_t) - \\mathbf{x}_0\\|^2$ , an $\\epsilon_{\\theta}$ model is trained using the squared loss $\\| \\epsilon_{\\theta}(\\mathbf{x}_t) - \\epsilon \\| ^2$ , and $\\mu_{t,\\theta}$ model is trained using the squared loss $\\| \\mu_{t,\\theta}(\\mathbf{x}_t) - \\mu_t\\|^2$ .", + "bbox": [ + 498, + 90, + 890, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Controllable generation with diffusion models.", + "text_level": 1, + "bbox": [ + 500, + 210, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free guidance. The conditioning signals are treated as additional inputs to the denoiser $p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,d)$ where $d$ is the conditioning signals which can be omitted $d = \\varnothing$ to generate unconditionally. Classifier-free guidance has been shown to generate very high-quality results [44, 45, 51]. To draw samples, the effective denoiser becomes $\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_t,d) = w\\epsilon_{\\theta}(\\mathbf{x}_t,d) + (1 - w)\\epsilon_{\\theta}(\\mathbf{x}_t,\\varnothing)$ , where $w$ controls the conditional strength. The new $\\hat{\\epsilon}_{\\theta}$ model can be used in Eq. 1. Two downsides of this method are that the nature of conditioning signals need to be known before hand and the denoiser needs to be adjusted and retrained for each specific case restricting its flexibility.", + "bbox": [ + 498, + 241, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier guidance. We can also obtain $p(\\mathbf{x}_{t-1}|\\mathbf{x}_t,d)$ from $p_{\\theta}(\\mathbf{x}_{t-1}|\\mathbf{x}_t)p(d|\\mathbf{x}_t)$ [12], where $p(d|\\mathbf{x}_t)$ is any probability function that we can approximate its score function $\\nabla_{\\mathbf{x}_t}\\log p(d|\\mathbf{x}_t)$ effectively. The new sampling process is similar to the original (Eq. 1) but with the mean shifted by the scaled score function as", + "bbox": [ + 498, + 429, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t} = \\mu_ {t} ^ {\\prime} + s \\Sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (d | \\mathbf {x} _ {t}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 532, + 890, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_t^\\prime$ is the original mean, $s$ controls the conditioning strength, and $\\Sigma_{t}$ is a variance scheduler which can be the same as in Eq. 1. Since $\\Sigma_{t}$ is a decreasing sequence, the guidance signal diminishes as $t\\to 0$ which corresponds to the characteristic of DPMs that tend to modify $\\mathbf{x}_t$ less and less as time goes. Classifier guidance is a post-hoc method, i.e., there is no change to the DPM model, one only needs to come up with $p(d|\\mathbf{x}_t)$ which is extremely flexible.", + "bbox": [ + 498, + 561, + 890, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Imputation and inpainting. To generate human motion sequences from partial observations, such as global motion trajectories or keyframe locations, inpainting is used. These partial observations, called imputing signals, are used to adjust the generative process towards the observations. Imputation and inpainting are two sides of the same coin.", + "bbox": [ + 498, + 690, + 890, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{y}$ be a partial target value in an input $\\mathbf{x}$ that we want to impute. The imputation region of $\\mathbf{y}$ on $\\mathbf{x}$ is denoted by $M_y^x$ , and a projection $P_y^x$ that resizes $\\mathbf{y}$ to that of $\\mathbf{x}$ by filling in zeros. In DPMs, imputation can be done on the sample $\\mathbf{x}_{t - 1}$ after every denoising step [10]. We have the new imputed sample $\\tilde{\\mathbf{x}}_{t - 1}$ as", + "bbox": [ + 498, + 781, + 890, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {t - 1} = \\left(1 - M _ {y} ^ {x}\\right) \\odot \\mathbf {x} _ {t - 1} + M _ {y} ^ {x} \\odot P _ {y} ^ {x} \\mathbf {y} _ {t - 1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2153", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is a Hadamard product and $\\mathbf{y}_{t - 1}$ is a noised target value. $\\mathbf{y}_{t - 1}\\sim \\mathcal{N}(\\sqrt{\\alpha_{t - 1}}\\mathbf{y},(1 - \\alpha_{t - 1})\\mathbf{I})$ following Ho et al. [20] is one of the simplest choices of $\\mathbf{y}_{t - 1}$ .", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that all three modes of conditioning presented here are not mutually exclusive. One could apply one or more tricks in a single pipeline.", + "bbox": [ + 75, + 137, + 468, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Guided Motion Diffusion", + "text_level": 1, + "bbox": [ + 76, + 198, + 310, + 214 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 GMD's two-stage guided motion diffusion" + ], + "code_body": "Require: A trajectory DPM $\\mathbf{z}_{0,\\phi}$ , a motion DPM $\\mathbf{x}_{0,\\theta}$ , a goal function $G_{\\mathbf{z}}(\\cdot)$ , and keyframe locations $\\mathbf{y}$ (if any). \n1: # Stage 1: Trajectory generation \n2: $\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ \n3: for all $t$ from $T$ to 1 do \n4: $\\mathbf{z}_0 \\gets \\mathbf{z}_{0,\\phi}(\\mathbf{z}_t)$ \n5: $\\mu, \\Sigma \\gets \\mu(\\mathbf{z}_0, \\mathbf{z}_t), \\Sigma_t$ \n6: # Classifier guidance (Eq. 2) \n7: # Dense signal propagation \n8: $\\mathbf{z}_{t-1} \\sim \\mathcal{N}(\\mu - s\\Sigma\\nabla_{\\mathbf{z}_t}G_z(\\mathbf{z}_0), \\Sigma)$ \n9: # Impute y on z (Eq. 3) (if any) \n10: $\\mathbf{z}_{t-1} \\gets (1 - M_y^z) \\odot \\mathbf{z}_{t-1} + M_y^z\\mathbf{y}_{t-1}$ \n11: end for \n12: # Stage 2: Trajectory-conditioned motion generation \n13: $\\mathbf{x}_T^{\\mathrm{proj}} \\gets$ sample from $\\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ \n14: for all $t$ from $T$ to 1 do \n15: $M \\gets P_z^x M_y^z$ # Imputation region of y on x \n16: $\\mathbf{x}_0^{\\mathrm{proj}} \\gets \\mathbf{x}_{0,\\theta}^{\\mathrm{proj}}(\\mathbf{x}_t^{\\mathrm{proj}})$ # Emphasis projection \n17: # Impute y on x $^{\\mathrm{proj}}$ (Eq. 6) \n18: $\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}} \\gets A((1 - M) \\odot A^{-1}\\mathbf{x}_0^{\\mathrm{proj}} + MP_z^x\\mathbf{y})$ \n19: $\\mu, \\Sigma \\gets \\mu(\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}},\\mathbf{x}_t^{\\mathrm{proj}}), \\Sigma_t$ \n20: # Masked classifier guidance (Eq. 9) \n21: # Dense signal propagation \n22: $\\Delta \\gets -s\\Sigma A^{-1}\\nabla_{\\mathbf{x}_t^{\\mathrm{proj}}}G_z(P_x^zA^{-1}\\mathbf{x}_0^{\\mathrm{proj}})$ \n23: $\\mu \\gets \\mu + A(1 - M) \\odot \\Delta$ \n24: $\\mathbf{x}_{t-1} \\sim \\mathcal{N}(\\mu, \\Sigma)$ \n25: end for \n26: return $\\mathbf{z}_0$", + "bbox": [ + 80, + 255, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to generate realistic human motions that can be guided by spatial constraints, enabling the generated human motion to achieve specific goals, such as following a global trajectory, reaching certain locations, or avoiding obstacles. Although diffusion-based models have significantly improved text-to-motion modeling [7, 51], generating motions that achieve specific goals is still beyond the reach of the current models. Our work addresses this limitation and advances the state-of-the-art in human motion modeling.", + "bbox": [ + 75, + 718, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are interested in modeling a full-body human motion that satisfies a certain scalar goal function $G_{x}(\\cdot)$ that takes a motion representation $\\mathbf{x}$ and measures how far the", + "bbox": [ + 75, + 854, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "motion $\\mathbf{x}$ is from the goal (lower is better). More specifically, $\\mathbf{x} \\in \\mathbb{R}^{N \\times M}$ represents a sequence of human poses for $M$ motion steps, where $N$ is the dimension of human pose representations, e.g., $N = 263$ in the HumanML3D [14] dataset. Let $X$ be the random variable associated with $\\mathbf{x}$ . Our goal is to model the following conditional probability using a motion DPM", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} \\mid G _ {x} (X) = 0\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 207, + 890, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This can be extended to $p\\big(\\mathbf{x}|G_x(X) = 0,d\\big)$ , where $d$ is any additional signal, such as text prompts. From now on, we omit $d$ to reduce clutter.", + "bbox": [ + 496, + 236, + 890, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Many challenging tasks in motion modeling can be encapsulated within a goal function $G_{z}$ that only depends on the trajectory $\\mathbf{z}$ of the human motion, not the whole motion $\\mathbf{x}$ . Let us define $\\mathbf{z} \\in \\mathbb{R}^{L \\times M}$ to be the trajectory part of $\\mathbf{x}$ with length $M$ and $L = 2$ describing the ground location of the pelvis of a human body. A particular location $\\mathbf{z}^{(i)}$ at motion step $i$ describes the pelvis location of the human body on the ground plane. We define a projection $P_x^z$ that resizes $\\mathbf{x}$ to match $\\mathbf{z}$ by taking only the $\\mathbf{z}$ part, and its reverse $P_z^x$ that resizes $\\mathbf{z}$ to match $\\mathbf{x}$ by filling in zeros. With this, our conditional probability becomes $p\\big(\\mathbf{x}|G_z(P_x^z X) = 0\\big)$ .", + "bbox": [ + 496, + 281, + 890, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we will show how text-to-motion DPMs can be extended to solve several challenging tasks, including trajectory-conditioned motion generation, location-conditioned trajectory planning, and obstacle avoidance trajectory planning. Using our proposed Emphasis projection and dense signal propagation, we alleviate the sparse guidance problem and enable motion generation based on spatial conditions. The overview of our methods is shown in Fig. 3.", + "bbox": [ + 496, + 448, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Emphasis projection", + "text_level": 1, + "bbox": [ + 500, + 578, + 692, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One of the most straightforward approaches for minimizing the goal function $G_{\\mathbf{z}}(\\cdot)$ is by analyzing what trajectories that minimize $\\mathbf{z}^{*} = \\arg \\min_{\\mathbf{z}}G_{z}(\\mathbf{z})$ look like. For a trajectory conditioning task, a whole trajectory $\\mathbf{z}^*$ is directly given. Our task is to generate the rest of the motion $\\mathbf{x}$ . With such knowledge, we can employ imputation & inpainting technique by supplying the motion DPM with the $\\mathbf{x}$ -shaped $P_z^x\\mathbf{z}^*$ to guide the generation process.", + "bbox": [ + 496, + 601, + 890, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem 1: Motion incoherence", + "text_level": 1, + "bbox": [ + 500, + 728, + 723, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the imputing trajectory $\\mathbf{z}^*$ is only a small part of the whole motion $\\mathbf{x}$ ( $L \\ll N$ ), we often observe that the DPM ignores the change from imputation and fails to make appropriate changes on the rest of $\\mathbf{x}$ . This results in an incoherent local motion that is not aligned or well coordinated with the imputing trajectory.", + "bbox": [ + 496, + 742, + 890, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Solution 1: Emphasis projection", + "text_level": 1, + "bbox": [ + 500, + 840, + 728, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We tackle this problem by giving more emphasis on the trajectory part of motion $\\mathbf{x}$ . More specifically, we propose an Emphasis projection method that increases the trajectory's", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2154", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2aada7e8aaab10721a928ae76978d7292c606cbb8db4108324dd3cf30e9c082c.jpg", + "image_caption": [ + "Figure 3. (a) Under standard motion representation and guiding method, only a few values in the motion representation are updated according to the guidance. (b) With Emphasis projection, all values in each frame describing the motion receive gradients w.r.t. the guidance, leading to better coherence between global orientation and local pose in each frame. (c) With dense gradient propagation, all frames are updated according to the guidance at the keyframes, making the guidance less likely to be ignored." + ], + "image_footnote": [], + "bbox": [ + 140, + 92, + 826, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "relative importance within motion $\\mathbf{x}$ . We achieve this by utilizing a random matrix $A = A'B$ , where $A' \\in \\mathbb{R}^{N \\times N}$ is a matrix with elements randomly sampled from $\\mathcal{N}(0,1)$ and $B \\in \\mathbb{R}^{N \\times N}$ is a diagonal matrix whose trajectory-related diagonal indexes are $c$ and the rest are 1 for emphasizing those trajectory elements. In our case, we emphasize the rotation and ground location of the pelvis, $(\\mathrm{rot},x,z)$ , in $\\mathbf{x}$ by $c$ times. We now have a projected motion $\\mathbf{x}^{\\mathrm{proj}} = \\frac{1}{N - 3 + 3c^2} A\\mathbf{x}$ . Note that the fractional term is to maintain the unit variance on $\\mathbf{x}_{\\mathrm{proj}}$ . The noisng process of the projected motion becomes $q(\\mathbf{x}_t^{\\mathrm{proj}}|\\mathbf{x}_0^{\\mathrm{proj}}) = \\mathcal{N}(\\sqrt{\\alpha_t}\\mathbf{x}_0^{\\mathrm{proj}},(1 - \\alpha_t)\\mathbf{I})$ . There is no change on how a DPM that works on the projected motion $p_{\\theta}(\\mathbf{x}_{t - 1}^{\\mathrm{proj}}|\\mathbf{x}_t^{\\mathrm{proj}})$ operates and treats $\\mathbf{x}_t^{\\mathrm{proj}}$ .", + "bbox": [ + 75, + 364, + 472, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Section 6.3, we show that emphasis projection is an effective way of solving the motion incoherence problem, and is shown to be substantially better than a straightforward approach of retraining a DPM with an increased loss weight on the trajectory.", + "bbox": [ + 75, + 565, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Imputation on the projected motion $\\mathbf{x}^{\\mathrm{proj}}$ . We have discussed imputing on the sample $\\mathbf{x}_{t-1}$ in Eq. 3. Here, we introduce an imputation on $\\mathbf{x}_0$ which modifies the DPM's belief on the final outcome $\\mathbf{x}_{0,\\theta}$ by imputing it with $\\mathbf{z}$ . We have found this technique useful in many tasks we are interested in.", + "bbox": [ + 75, + 647, + 468, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let us define the imputation region of $\\mathbf{z}$ on $\\mathbf{x}$ as $M_z^x$ . We obtain the imputed $\\tilde{\\mathbf{x}}_0$ from", + "bbox": [ + 76, + 739, + 468, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {0} = \\left(1 - M _ {z} ^ {x}\\right) \\odot \\mathbf {x} _ {0, \\theta} + M _ {z} ^ {x} \\odot \\underbrace {P _ {z} ^ {x} \\mathbf {z} ^ {*}} _ {\\mathbf {x} \\text {s h a p e d}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 785, + 468, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now operating on the projected motion $\\mathbf{x}^{\\mathrm{proj}}$ , before we can do imputation, we need to unproject it back to the original motion using $\\mathbf{x}_0 = A^{-1}\\mathbf{x}_0^{\\mathrm{proj}}$ , and then project the imputed $\\tilde{\\mathbf{x}}_0$ back using $\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}} = A\\tilde{\\mathbf{x}}_0$ . We obtain the imputed", + "bbox": [ + 76, + 838, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "motion under emphasis projection $\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}}$ from", + "bbox": [ + 500, + 363, + 797, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {0} ^ {\\text {p r o j}} = A \\left(\\left(1 - M _ {z} ^ {x}\\right) \\odot \\left(A ^ {- 1} \\mathbf {x} _ {0, \\theta} ^ {\\text {p r o j}}\\right) + M _ {z} ^ {x} \\odot P _ {z} ^ {x} \\mathbf {z} ^ {*}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 387, + 890, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Substituting $\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}}$ into Eq. 1, we obtain the new mean $\\tilde{\\mu}_t^{\\mathrm{proj}}$ for sampling $\\mathbf{x}_{t - 1}^{\\mathrm{proj}}\\sim \\mathcal{N}(\\tilde{\\mu}_t^{\\mathrm{proj}},\\Sigma_t)$", + "bbox": [ + 498, + 421, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Dense guidance signal with a learned denoiser", + "text_level": 1, + "bbox": [ + 498, + 463, + 888, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another way to minimize the goal function $G_{z}(\\cdot)$ is by adjusting the sample of each diffusion step $\\mathbf{x}_{t-1}$ toward a region with lower $G_{z}$ . This trick is called classifier guidance [12]. The direction of change corresponds to a score function $\\nabla_{\\mathbf{x}_t}\\log p\\big(G_x(X_t) = 0|\\mathbf{x}_t\\big)$ which can be approximated as a direction $\\Delta_{\\mathbf{x}_0} = -\\nabla_{\\mathbf{x}_0}G_z(P_x^z\\mathbf{x}_{0,\\theta})$ that reduces the goal function. We can guide the generative process by nudging the DPM's prediction as $\\mathbf{x}_0 = \\mathbf{x}_{0,\\theta} + \\Delta_{\\mathbf{x}_0}$ . While imputation requires the minimizer $\\mathbf{z}^*$ of $G_{z}$ , which might not be easy to obtain or may not be unique, this trick only requires the easier-to-obtain direction of change.", + "bbox": [ + 496, + 486, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem 2: Sparse guidance signal", + "text_level": 1, + "bbox": [ + 498, + 659, + 746, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the motion domain, conditioning signals can often be sparse. There are two types of sparsity that can occur: sparsity in feature and sparsity in time. Sparsity in feature is when the conditioning signal is a small part of the feature dimension of $\\mathbf{x}$ . For example, in trajectory-conditioned generation, $\\mathbf{z}$ may only consist of a sequence of ground locations over time. This type of sparsity can be addressed by emphasis projection, as explained in Section 4.1. Sparsity in time refers to cases where the conditioning signal consists of small segments of a trajectory spread out over time. For instance, in keyframe location conditioning task, only a sparse set of keyframe locations are given. When the conditioning signal-to-noise ratio becomes too small, the conditioning signal may be mistaken as noise and ignored during the denoising process.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2155", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solution 2: Dense signal propagation", + "text_level": 1, + "bbox": [ + 76, + 90, + 338, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To turn a sparse signal into a dense signal, we need domain knowledge. One way to achieve this is by using a denoising function $f(\\mathbf{x}_t) = \\mathbf{x}_0$ , which is trained on a motion dataset to denoise by gathering information from the nearby motion frames. With the ability to relate a single frame to many other frames, the denoising function is capable of expanding a sparse signal into a denser one.", + "bbox": [ + 75, + 106, + 467, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can use backward propagation through the denoising function $f$ to take advantage of this. Therefore, a dense classifier guidance can be obtained as follows:", + "bbox": [ + 75, + 210, + 467, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {x} _ {t}} \\log p \\left(G _ {x} \\left(X _ {t}\\right) = 0 \\mid \\mathbf {x} _ {t}\\right) \\approx - \\nabla_ {\\mathbf {x} _ {t}} G _ {z} \\left(\\underbrace {P _ {x} ^ {z} f (\\mathbf {x} _ {t})} _ {\\mathbf {z} \\text {s h a p e d}}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 267, + 467, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While an external function can be used as $f$ , we observe that the existing DPM model $\\mathbf{x}_{0,\\theta}(\\mathbf{x}_t)$ itself is a motion de-noiser, and thus can be used to turn a sparse signal into a dense signal without the need for an additional model. In practice, this process amounts to computing the gradient of $G$ with respect to $\\mathbf{x}_t$ through $\\mathbf{x}_{0,\\theta}(\\mathbf{x}_t)$ using autodiff.", + "bbox": [ + 75, + 311, + 467, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Applying classifier guidance together with imputation. Whenever available, we want to utilize signals from both imputation and classifier guidance techniques to help guide the generative process. Imputation is explicit but may encounter sparsity in time, while classifier guidance is indirect but dense. We want to use the direct signal from imputation wherever available (with mask $M_z^x$ ), and the rest from classifier guidance (with mask $1 - M_z^x$ ). Based on Eq. 2, imputation-aware classifier guidance can be written as", + "bbox": [ + 75, + 409, + 467, + 545 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t} = \\tilde {\\mu} _ {t} - \\left(1 - M _ {z} ^ {x}\\right) \\odot s \\Sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} G _ {z} \\left(P _ {x} ^ {z} f (\\mathbf {x} _ {t})\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 556, + 467, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{\\mu}$ is an imputed sampling mean. By replacing $\\tilde{\\mu}$ with $\\tilde{\\mu}^{\\mathrm{proj}}$ , we get classifier guidance together with imputation that works with emphasis projection as", + "bbox": [ + 75, + 584, + 467, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\mu} = - s \\Sigma_ {t} A ^ {- 1} \\nabla_ {\\mathbf {x} _ {t} ^ {\\mathrm {p r o j}}} G _ {z} \\left(P _ {x} ^ {z} A ^ {- 1} f \\left(\\mathbf {x} _ {t} ^ {\\mathrm {p r o j}}\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 638, + 467, + 659 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t} ^ {\\text {p r o j}} = \\tilde {\\mu} _ {t} ^ {\\text {p r o j}} + A \\left(1 - M _ {z} ^ {x}\\right) \\odot \\Delta_ {\\mu} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 662, + 467, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Problem 3: DPM's bias hinders the guidance signal", + "text_level": 1, + "bbox": [ + 76, + 698, + 437, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A DPM removes noise from an input based on the distribution of the training data it has seen. This could be problematic when it comes to conditional generation because the conditioning signal may be outside of the training distribution. As a result, any changes made to the classifier guidance may be reversed by the DPM in the next time step, due to its inherent bias towards the data, shown in Figure 4.", + "bbox": [ + 75, + 713, + 467, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Solution 3: Epsilon modeling", + "text_level": 1, + "bbox": [ + 76, + 825, + 284, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While it is unlikely to train an unbiased DPM model, there are ways to minimize the influence of model's bias under the guidance signal. Conceptually, the DPM model usually makes less and less change near the final outcome. This is", + "bbox": [ + 75, + 839, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in tandem with the guidance signal that gradually decreases over time due to $\\Sigma_t$ (Eq. 2).", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We investigate the coefficient $\\frac{\\sqrt{\\alpha_{t-1}\\beta_t}}{1-\\alpha_t}$ of $\\mathbf{x}_0$ in the sampling mean $\\mu_t$ (Eq. 1). This coefficient reaches its maximum value at $t = 0$ , meaning that an $\\mathbf{x}_{0,\\theta}$ model could have a significant impact on the sampling mean even at $t = 0$ , which contradicts the weak guidance signal at that time.", + "bbox": [ + 496, + 121, + 888, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the other hand, an $\\epsilon_{\\theta}$ model will have the most influence on the sampling mean at $t = T$ , which aligns with our intuition. In Section 6.4 and Figure 4, we demonstrate that modeling $\\epsilon_{\\theta}$ instead of $\\mathbf{x}_{0,\\theta}$ is a successful approach for managing the bias effect of the DPM model in classifier guidance. We further discuss this point in Supplementary.", + "bbox": [ + 496, + 200, + 890, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Applications", + "text_level": 1, + "bbox": [ + 500, + 306, + 629, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Trajectory-conditioned generation", + "text_level": 1, + "bbox": [ + 500, + 333, + 799, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This task aims at generating a realistic motion $\\mathbf{x}$ that matches a given trajectory $\\mathbf{z}$ . Our objective is to minimize the distance between the generated motion and the given trajectory, which we define as", + "bbox": [ + 496, + 358, + 890, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG _ {x} (\\mathbf {x}) := \\left\\| \\mathbf {z} - \\underbrace {P _ {x} ^ {z} \\mathbf {x}} _ {\\mathbf {z} \\text {p a r t o f} \\mathbf {x}} \\right\\| _ {p} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 430, + 888, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite the apparent simplicity of this task, a traditional DPM faces the challenge of ensuring coherence in the generated motion. However, our emphasis projection method can effectively address this problem.", + "bbox": [ + 496, + 482, + 888, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Keyframe-conditioned generation", + "text_level": 1, + "bbox": [ + 500, + 556, + 794, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The locations of ground positions at specific times can be used to define locations that we wish the generated motion to reach. This task is a generalized version of the trajectory-conditioned generation where only a partial and potentially sparse trajectory is given. Let $\\mathbf{y} \\in \\mathbb{R}^{2 \\times M}$ be a trajectory describing keyframe locations and a mask $M_y^z$ describe the key motion steps. Our goal function of a motion $\\mathbf{x}$ is", + "bbox": [ + 496, + 580, + 888, + 685 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG _ {x} (\\mathbf {x}) := \\sum_ {i} \\left\\| M _ {y} ^ {z} \\left(P _ {x} ^ {z} \\mathbf {x} - \\mathbf {y}\\right) \\right\\| _ {p} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 698, + 888, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consequently, $G_{z}(\\mathbf{z}) = \\sum_{i}\\left\\| M_{y}^{z}(\\mathbf{z} - \\mathbf{y})\\right\\|_{p}$ . Due to the partial trajectory $\\mathbf{y}$ , the imputation region of $\\mathbf{y}$ on $\\mathbf{x}$ becomes $M_y^x = P_z^x M_y^z$ .", + "bbox": [ + 496, + 739, + 888, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Two-stage guided motion generation. Generating both the trajectory and motion simultaneously under a conditioning signal can be challenging and may result in lower quality motion. To address this issue, we propose a two-step approach. First, we generate a trajectory $\\mathbf{z}$ that satisfies the keyframe locations and then generate the motion $\\mathbf{x}$ given the trajectory (following Section 5.1). Our overall pipeline", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2156", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is depicted in Figure 2 (a). We offer two options for generating the trajectory from keyframe locations $\\mathbf{y}$ : a point-to-point trajectory and a trajectory DPM.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The point-to-point trajectory connects consecutive keyframe locations with a straight line. These unrealistic trajectories can be used as imputation signals for the motion DPM during the early phase $(t \\geq \\tau)$ . If $\\tau$ is large enough, the DPM will adjust the given trajectory to a higher quality one. However, if $\\tau$ is too large, the DPM may generate a motion that does not perform well on $G_{z}$ .", + "bbox": [ + 75, + 137, + 467, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The trajectory DPM $p_{\\phi}(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)$ , which is trained using the same dataset but with a smaller network, can be used to generate the trajectory under the guidance signal from $G_{z}$ . We summarize our two-stage approach in Algorithm 1.", + "bbox": [ + 76, + 242, + 467, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is also possible to combine the two methods, as the point-to-point trajectory can serve as a useful guidance signal for the trajectory DPM during $t \\geq \\tau$ . After that, the trajectory DPM is subject to the usual imputation and classifier guidance from $G_{z}$ . By tuning $\\tau$ , we can balance between trajectory diversity and lower scores on $G_{z}$ .", + "bbox": [ + 75, + 303, + 467, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Obstacle avoidance motion generation", + "text_level": 1, + "bbox": [ + 76, + 400, + 405, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Humans have the ability to navigate around obstacles while traveling from point A to B. Under our framework, this problem can be defined using two goal functions: one that navigates from A to B, called $G_{x}^{\\mathrm{loc}}$ (defined as in Eq. 12), and another that pushes back when the human model crosses the obstacle's boundary, called $G_{x}^{\\mathrm{obs}}$ , as follows", + "bbox": [ + 75, + 422, + 467, + 513 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nG _ {x} ^ {\\mathrm {o b s}} (\\mathbf {x}) := \\sum_ {i} - \\operatorname {c l i p m a x} \\left(\\operatorname {S D F} \\left(\\left(P _ {x} ^ {z} \\mathbf {x}\\right) ^ {(i)}\\right), c\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 518, + 467, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c$ is the safe distance from the obstacle. These two goal functions are combined additively to obtain the final goal function, $G_{x}(\\mathbf{x}) = G_{x}^{\\mathrm{loc}}(\\mathbf{x}) + G_{x}^{\\mathrm{obs}}(\\mathbf{x})$ , for this task.", + "bbox": [ + 75, + 555, + 467, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We utilize the same pipeline as in Section 5.2, with the exception that imputation is not possible for obstacle avoidance. Therefore, minimizing the obstacle avoidance goal relies solely on classifier guidance.", + "bbox": [ + 75, + 601, + 467, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 672, + 207, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate our methods, we perform experiments on the standard human motion generation task conditioned on text descriptors and spatial objectives. In particular, we evaluate (1) the performance of our model in standard text-condition motion generation tasks, (2) the effect of emphasis projection to alleviate incoherence between spatial locations and local poses, (3) the ability to conditionally generate motion based on spatial information by conditioning with given trajectories, keyframe locations, and obstacles.", + "bbox": [ + 75, + 696, + 467, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 840, + 174, + 856 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate generative text-to-motion models using standard metrics introduced by Guo et al.", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/077c8037bb07fad7206e414702650fa0d547c21ea319a5990dba16dea1675079.jpg", + "table_caption": [ + "Table 1. Text-to-motion evaluation on the HumanML3D [14] dataset. The right arrow $\\rightarrow$ means closer to real data is better." + ], + "table_footnote": [], + "table_body": "
FID ↓R-precision ↑ (Top-3)Diversity →
Real0.0020.7979.503
JL2P [1]11.020.4867.676
Text2Gesture [5]7.6640.3456.409
T2M [14]1.0670.7409.188
MotionDiffuse [59]0.6300.7829.410
MDM [51]0.5560.6089.446
MLD [7]0.4730.7729.724
PhysDiff [58]0.4330.631-
Ours0.2120.6709.440
Ours xproj0.2350.6529.726
", + "bbox": [ + 503, + 119, + 885, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[14]. These include Fréchet Inception Distance (FID), R- Precision, and Diversity. FID measures the distance between the distributions of ground truth and generated motion using a pretrained motion encoder. R- Precision evaluates the relevance of the generated motion and its text prompt, while Diversity measures the variability within the generated motion. We also report Foot skating ratio, which measures the proportion of frames in which either foot skids more than a certain distance $(2.5\\mathrm{cm})$ while maintaining contact with the ground (foot height $< 5\\mathrm{cm}$ ), as a proxy for the incoherence between trajectory and human motion.", + "bbox": [ + 496, + 310, + 890, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, for conditional generation with keyframe locations, we use Trajectory diversity, Trajectory error, Location error, and Average error of keyframe locations. Trajectory diversity measures the root mean square distance of each location of each motion step from the average location of that motion step across multiple samples with the same settings. Trajectory error is the ratio of unsuccessful trajectories, defined as those with any keyframe location error exceeding a threshold. Location error is the ratio of keyframe locations that are not reached within a threshold distance. Average error measures the mean distance between the generated motion locations and the keyframe locations measured at the keyframe motion steps.", + "bbox": [ + 496, + 478, + 890, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We evaluate the text-to-motion generation using the HumanML3D [14] dataset, which is a collection of text-annotate motion sequences from AMASS [33] and HumanAct12 [16] datasets. It contains 14,646 motions and 44,970 motion annotations.", + "bbox": [ + 496, + 681, + 890, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. Both our motion DPM and trajectory DPM are based on UNET with AdaGN [12] depicted in details in the Supplementary. The motion DPM is an $\\mathbf{x}_0$ model, while the trajectory DPM is an $\\epsilon$ model, as explained in Section 4.2, to enhance controllability. We utilized DDPM [20] with $T = 1,000$ denoising steps for training and inference of both models. Additionally, we condition the generation process on text prompts in a classifier-free [21] manner, similar to MDM [51], and use the CLIP [41]", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2157", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5307deeb383aba53e95407c09ff31ce3784e060c878b56630fc93128636c648.jpg", + "table_caption": [ + "Table 2. Trajectory-conditioned motions evaluation. The ground truth trajectory is used for imputing after each diffusion step. Comparing the effect of an original $\\mathbf{x}$ with emphasis loss functions to the emphasis projection $\\mathbf{x}^{\\mathrm{proj}}$ after imputing whole trajectories after each diffusion step." + ], + "table_footnote": [], + "table_body": "
ModelSpaceEmphasisFID ↓Foot skating ↓ ratio
MDM [51]xloss 1×0.9040.284
xprojc = 10.6320.304
c = 20.4640.309
c = 50.4660.256
c = 101.0290.161
Oursxloss 1×0.2780.262
loss 22×0.2560.250
loss 52×0.2400.249
loss 102×0.3200.265
xprojc = 10.3070.268
c = 20.2900.257
c = 50.2740.199
c = 100.1980.128
", + "bbox": [ + 91, + 161, + 452, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model as the text encoder across all tasks.", + "bbox": [ + 76, + 402, + 354, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational resources Our GMD architecture is capable of running both the motion and trajectory models on a single commercial GPU, such as the Nvidia RTX 2080 Ti, 3080, or 3090. The trajectory model achieved a throughput of 2,048 samples per second when run on an RTX 3090, with a training time of approximately 4.34 GPU hours. Meanwhile, the motion model achieved a throughput of 256 samples per second on an RTX 3090, with a training time of around 34.7 GPU hours. The total inference time for one sample is approximately 110 seconds.", + "bbox": [ + 75, + 424, + 470, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Text-to-motion generation", + "text_level": 1, + "bbox": [ + 76, + 585, + 313, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section evaluates our model's performance in the standard text-to-motion generation task and compares it with other motion DPM baselines: MotionDiffuse [59], MDM [51], MLD [7], and PhysDiff [58]. Tab. 1 shows the results where our model architecture outperforms the baselines significantly in terms of motion quality measured by FID, while maintaining similar R-Precision and Diversity.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Trajectory-conditioned generation", + "text_level": 1, + "bbox": [ + 76, + 724, + 377, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section demonstrates how our emphasis projection method can address the issue of incoherent motion caused by spatial conditioning, specifically in the trajectory conditioning task, where the model is provided with ground-truth trajectories for imputation at each denoising step and is required to generate corresponding local poses. Both quantitative and qualitative results support that our emphasis projection leads to a reduction in Foot skating ratio, as evidenced in Tab. 2 and a more coherent motion in Fig. 5 compared to the MDM [51] model.", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f06bfcd27c56b984d76b2308cb8bbe7690b497403a79a5a9c44138f544871d35.jpg", + "image_caption": [ + "Figure 4. Comparing the evolution of the clean trajectory subject to classifier guidance from $\\mathbf{x}_0$ and $\\epsilon$ DPMs. The $\\mathbf{x}_0$ DPM shows significant resistance on the guidance signal as exhibited by the trajectory \"contraction\" behavior at $t\\to 0$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 92, + 890, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/159c388c323ff7de7a83b347165a0566f84fcc767a9165a6eb8a7bb7207009b3.jpg", + "image_caption": [ + "Figure 5. Generated motion, conditioned a given trajectory and text \"walking forward\". MDM [51] exhibits motion incoherence where the model disregards the trajectory and generates an inconsistent motion. Our method, improved by emphasis projection, deals effectively with the conditioning." + ], + "image_footnote": [], + "bbox": [ + 504, + 301, + 694, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f64d7f192fdad6bad7a73d0ea65c795dd1bb50bdcd0c8c19f07b28e63e655d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 301, + 887, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca8ba36a0fc61da6ac016941d9b88c5a4eaa55b07059cfc44b199f1a4fcf9924.jpg", + "image_caption": [ + "Figure 6. Generated motion trajectories, conditioned on target locations at given keyframes. Without dense signal propagation, the model ignores the target conditions." + ], + "image_footnote": [], + "bbox": [ + 504, + 494, + 885, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also compare our emphasis projection method with an alternative approach of increasing the trajectory loss strength during training. We include loss $k^2 \\times$ baselines, where $k \\in \\{1, 2, 5, 10\\}$ , for comparison. The results in Tab. 2 indicate that, while increasing the loss strength marginally improves both FID and Foot skating ratio, increasing it beyond a certain point leads to a decline in both FID and Foot skating ratio. By contrast, our emphasis projection method consistently leads to improvements in both metrics. We discuss this topic further in the Supplementary.", + "bbox": [ + 496, + 681, + 890, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Keyframe-conditioned generation", + "text_level": 1, + "bbox": [ + 500, + 845, + 794, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section evaluates the quality and adherence of the generated motion to the desired goal. A viable solution", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2158", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8b4b0f004d2567b4b9108b37df890a804cb90ff64c269ea75cb9b416a433e125.jpg", + "table_caption": [ + "Table 3. The effect of different conditioning strategies tested on keyframe-conditioning task. The keyframes $(N = 5)$ are sampled from the ground truth motion trajectories with the same text prompts in the HumanML3D [14] test set." + ], + "table_footnote": [], + "table_body": "
ModelConditioningFID ↓Foot ↓ skating ratioTraj. ↑ diversity (m.)Traj. err. ↓ (50 cm)Loc. err. ↓ (50 cm)Avg. err. ↓R-precision ↑ (Top-3)
MDM [51]Single stagex + τ=01.2560.2020.1340.0000.0000.0000.631
xproj + τ=02.9940.1510.1340.0000.0000.0000.554
xproj + τ=1002.2130.0950.2140.3260.1270.2360.555
xproj + no p2p1.6790.0920.3940.5190.3260.5430.548
Ours (xproj)Single stageτ=00.9020.1270.1170.0000.0000.0000.594
τ=1000.5230.0860.1570.1760.0490.1390.599
Two stageτ=1000.9370.0980.1200.0760.0200.1090.574
τ=3000.9380.0980.1270.1180.0310.1280.573
τ=5000.9080.0980.1400.1570.0430.1400.577
τ=7000.8980.0980.1620.1960.0580.1530.580
τ=9000.8740.0980.1920.2380.0800.1800.581
no p2p0.8620.1040.2220.2870.1180.2820.577
", + "bbox": [ + 78, + 119, + 890, + 318 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "must meet both criteria to an acceptable degree.", + "bbox": [ + 76, + 333, + 392, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To achieve high-quality motion, both FID and Foot skating ratio are essential since FID alone cannot adequately measure the trajectory-motion coherence. Our Emphasis projection technique significantly improves motion coherence, reducing foot skating as shown in Tab. 3 while MDM [51] is unsuitable for this task due to the high motion incoherence. Furthermore, our improved architecture significantly improves motion quality in all cases. Note that without dense signal propagation, the model ignores the keyframe conditioning as shown in Fig 6.", + "bbox": [ + 75, + 349, + 468, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While a single-stage model performs reasonably well due to emphasis projection, it is too restrictive at $\\tau = 0$ (forced trajectory), resulting in relatively high Foot skating. This issue can be addressed by allowing more modification (increasing to $\\tau$ to 100) but at the cost of higher Loc. error.", + "bbox": [ + 75, + 500, + 468, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Lastly, the trajectory model's better controllability reduces Location error by more than half compared to the single-stage model at $\\tau = 100$ . As expected, increasing $\\tau$ leads to more freedom in the model, resulting in increased Trajectory diversity, lower FID, and higher Location error.", + "bbox": [ + 75, + 575, + 468, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.5. Obstacle avoidance motion generation", + "text_level": 1, + "bbox": [ + 76, + 659, + 405, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, we demonstrate our model's ability to generate motion on additional guidance on the obstacle avoidance task. In this task, we randomly sample the target point that the human needs to reach at a specific motion step along with a set of obstacles it needs to avoid, represented as a 2D SDF (Sec. 5.3). We show the qualitative results in Fig 7.", + "bbox": [ + 75, + 681, + 468, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 76, + 784, + 328, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose GMD, a controllable human motion generation method based on goal functions. GMD produces high-quality and diverse motions and supports diverse possibilities for goal functions. Since obtaining necessary data and designing a classifier-free learning method for non-explicit goals, such as obstacle avoidance,", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0f7203132172e6f7bf31a65a030ba7fc2f52329d35f5fbdaa3f5e35c1bb705b2.jpg", + "image_caption": [ + "Figure 7. Qualitative results from the obstacle avoidance task given keyframe locations and obstacles. The red crossed areas represent obstacles to avoid. More results are in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 504, + 332, + 890, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "can be challenging, our GMD utilizes a classifier-based method which allows for more conditioning flexibility without retraining the model. Thus, our studies on effective classifier guidance will be useful for further including more guiding signals.", + "bbox": [ + 496, + 517, + 890, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the SNSF project grant 200021 204840.", + "bbox": [ + 498, + 599, + 890, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 646, + 596, + 661 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 7", + "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. arXiv preprint arXiv:2211.09707, 2022. 3", + "[3] Jose A Arjona-Medina, Michael Gillhofer, Michael Widrich, Thomas Unterthiner, Johannes Brandstetter, and Sepp Hochreiter. RUDDER: Return decomposition for delayed rewards. In Advances in Neural Information Processing Systems, 2019. 2", + "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu" + ], + "bbox": [ + 506, + 672, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2159", + "bbox": [ + 480, + 949, + 514, + 960 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu. eDiff-I: Text-to-Image diffusion models with an ensemble of expert denoisers. Nov. 2022. 2", + "[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 7", + "[6] T Brooks, A Holynski, and A A Efros. InstructPix2Pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2", + "[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. arXiv preprint arXiv:2212.04048, 2022. 1, 3, 4, 7, 8", + "[8] Xin Chen, Zhuo Su, Lingbo Yang, Pei Cheng, Lan Xu, Bin Fu, and Gang Yu. Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134, 2022. 3", + "[9] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. Aug. 2021. 2, 3", + "[10] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. June 2022. 3", + "[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. arXiv preprint arXiv:2212.04495, 2022. 1, 3", + "[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. May 2021. 2, 3, 5, 7", + "[13] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 3", + "[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 4, 7, 9", + "[15] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 580-597. Springer, 2022. 3", + "[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3, 7", + "[17] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-between. ACM Transactions on Graphics (TOG), 39(4):60-1, 2020. 3", + "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. Aug. 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. Oct. 2022. 2", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. June 2020. 2, 3, 4, 7", + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications. 3, 7", + "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Apr. 2022. 2", + "[23] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. May 2022. 2", + "[24] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927. IEEE, 2020. 3", + "[25] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. 3", + "[26] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. Sept. 2020. 2", + "[27] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. Feb. 2023. 3", + "[28] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 3", + "[29] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 3", + "[30] Haoying Li, Yifan Yang, Meng Chang, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. SRDiff: Single image SuperResolution with diffusion probabilistic models. Apr. 2021. 2", + "[31] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3", + "[32] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 3", + "[33] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 7", + "[34] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Image synthesis and" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2160", + "bbox": [ + 482, + 949, + 516, + 960 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "editing with stochastic differential equations. Aug. 2021. 2, 3", + "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3", + "[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3", + "[37] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 480-497. Springer, 2022. 3", + "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. Sept. 2022. 2", + "[39] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-TTS: A diffusion probabilistic model for Text-to-Speech. May 2021. 2", + "[40] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7", + "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. ArXiv, 2022. 2, 3", + "[43] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 3", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution image synthesis with latent diffusion models. Dec. 2021. 2, 3", + "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image diffusion models with deep language understanding. May 2022. 2, 3", + "[46] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image SuperResolution via iterative refinement. Apr. 2021. 2", + "[47] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning using nonequilibrium thermodynamics. Mar. 2015. 2", + "[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. July 2019. 2", + "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based generative modeling through stochastic differential equations. Nov. 2020. 2", + "[50] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, Aug. 1988. 2", + "[51] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. Sept. 2022. 1, 3, 4, 7, 8, 9", + "[52] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. Edge: Editable dance generation from music. arXiv preprint arXiv:2211.10658, 2022. 1, 3", + "[53] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 3", + "[54] Christopher John Cornish Hellaby Watkins. Learning from delayed rewards. PhD thesis, King's College, Cambridge, 1989. 2", + "[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. Oct. 2022. 2", + "[56] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 257-274. Springer, 2022. 3", + "[57] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019. 3", + "[58] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500, 2022. 1, 3, 7, 8", + "[59] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 7, 8", + "[60] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3", + "[61] Yan Zhang, Michael J Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020. 3", + "[62] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion gener" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2161", + "bbox": [ + 482, + 949, + 513, + 960 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 1, 3", + "[63] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6225-6234, 2020. 3", + "[64] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. arXiv preprint arXiv:2211.16016, 2022. 1, 3" + ], + "bbox": [ + 78, + 90, + 468, + 219 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2162", + "bbox": [ + 482, + 950, + 516, + 960 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_model.json b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b4dd14b3081410e4d9f8e1e633c16c7bc04c05c --- /dev/null +++ b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_model.json @@ -0,0 +1,2776 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.829, + 0.154 + ], + "angle": 0, + "content": "Guided Motion Diffusion for Controllable Human Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.18, + 0.328, + 0.199 + ], + "angle": 0, + "content": "Korrawe Karunratanakul1" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.181, + 0.51, + 0.2 + ], + "angle": 0, + "content": "Konpat Preechakul2" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.181, + 0.73, + 0.2 + ], + "angle": 0, + "content": "Supasorn Suwajanakorn2" + }, + { + "type": "text", + "bbox": [ + 0.75, + 0.181, + 0.84, + 0.2 + ], + "angle": 0, + "content": "Siyu Tang" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.209, + 0.681, + 0.227 + ], + "angle": 0, + "content": "\\(^{1}\\)ETH Zürich, Switzerland \\(^{2}\\)VISTEC, Thailand" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.23, + 0.652, + 0.246 + ], + "angle": 0, + "content": "https://korrawe.github.io/gmd-project/" + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.273, + 0.222, + 0.285 + ], + "angle": 0, + "content": "(a) Text only" + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.273, + 0.44, + 0.285 + ], + "angle": 0, + "content": "(b) Text + Trajectory" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.273, + 0.663, + 0.285 + ], + "angle": 0, + "content": "(c) Text + Keyframe locations" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.264, + 0.862, + 0.284 + ], + "angle": 0, + "content": "(d) Text + Keyframe locations +Obstacle avoidance" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.287, + 0.287, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.287, + 0.488, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.287, + 0.684, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.287, + 0.885, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.451, + 0.894, + 0.507 + ], + "angle": 0, + "content": "Figure 1. Our proposed Guided Motion Diffusion (GMD) can generate high-quality and diverse motions given a text prompt and a goal function. We demonstrate the controllability of GMD on four different tasks, guided by the following conditions: (a) text only, (b) text and trajectory, (c) text and keyframe locations (double circles), and (d) with obstacle avoidance (red-cross areas represent obstacles). The darker the colors, the later in time." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.531, + 0.314, + 0.546 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.473, + 0.911 + ], + "angle": 0, + "content": "Denoising diffusion models have shown great promise in human motion synthesis conditioned on natural language descriptions. However, integrating spatial constraints, such as pre-defined motion trajectories and obstacles, remains a challenge despite being essential for bridging the gap between isolated human motion and its surrounding environment. To address this issue, we propose Guided Motion Diffusion (GMD), a method that incorporates spatial constraints into the motion generation process. Specifically, we propose an effective feature projection scheme that manipulates motion representation to enhance the coherency between spatial information and local poses. Together with a new imputation formulation, the generated motion can reliably conform to spatial constraints such as global motion trajectories. Furthermore, given sparse spatial constraints (e.g. sparse keyframes), we introduce a new dense guidance approach to turn a sparse signal, which is susceptible to being ignored during the reverse steps, into denser signals to guide the generated motion to the given constraints. Our extensive experiments justify the development of GMD, which achieves a significant improvement over state-of-the-art methods in text-based motion generation while allowing control of the synthesized motions with spatial constraints." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.531, + 0.631, + 0.546 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.851 + ], + "angle": 0, + "content": "Recently, denoising diffusion models have emerged as a promising approach for human motion generation [11, 62, 64] outperforming other alternatives such as GAN or VAE in terms of both quality and diversity [7, 51, 58]. Several studies have focused on generating motion based on expressive text prompts [7, 51], or music [52, 64]. The state-of-the-art motion generation methods, such as MDM [51], utilize classifier-free guidance to generate motion conditioned on text prompts. However, incorporating spatial constraints into diffusion models remains underexplored. Human motions consist of both semantic and spatial information, where the semantic aspect can be described using natural languages or action labels and the spatial aspect governs physical interaction with surroundings. To generate realistic human motion in a 3D environment, both aspects must be incorporated. Our experiments show that simply adding spatial constraint guidance, such as global trajectories, into the state-of-the-art models or using imputation and in-painting approaches do not yield satisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We identify two main issues that make the motion diffusion models likely to ignore the guidance when conditioned on spatial objectives: the sparseness of global orientation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.514, + 0.962 + ], + "angle": 0, + "content": "2151" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.466, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.264, + 0.465, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.449 + ], + "angle": 0, + "content": "Figure 2. We tackle the problem of spatially conditioned motion generation with GMD, depicted in a). Our main contributions are b) Emphasis projection, for better trajectory-motion coherence, and c) Dense signal propagation, for a more controllable generation even under sparse guidance signal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.581 + ], + "angle": 0, + "content": "in the motion representation and sparse frame-wise guidance signals. By design, the diffusion models are a denoising model that consecutively denoises the target output over multiple steps. With sparse guidance, a small portion of the output that receives guidance will be inconsistent with all other parts that do not, therefore, are more likely to be treated as noise and discarded in subsequent steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.734 + ], + "angle": 0, + "content": "First, the sparseness within a frame is a result of common motion representations that separate local pose information, like joint rotations, from global orientations, such as pelvis translations and rotations [43], usually with more focus on local poses. For instance, the common motion representation [14] uses 4 values to represent global orientation and 259 values for local pose in each frame. Such imbalance can cause the model to focus excessively on local pose information, and consequently, perceive guided global orientation as noise, resulting in a discrepancy such as foot skating." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Second, in many applications such as character animation, gaming, and virtual reality, the spatial control signals are defined on only a few keyframes such as target locations on the ground. We show that the current diffusion-based motion generation models struggle to follow such sparse guidance as doing so is equivalent to guiding an image diffusion model with only a few pixels. As a result, either the guidance at the provided keyframes will be ignored during the denoising process or the output motion will contain an artifact where the character warps to satisfy the guidance only in those specific keyframes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.439 + ], + "angle": 0, + "content": "To effectively incorporate sparse spatial constraints into the motion generation process, we propose GMD, a novel and principled Guided Motion Diffusion model. To alleviate the discrepancy between local pose and global orientation in the guided denoising steps, we introduce emphasis projection, a general representation manipulation method that we use to increase the importance of spatial information during training. Additionally, we derive a new imputation and inpainting formulation that enables the existing inpainting techniques to operate in the projected space, which we leverage to generate significantly more coherent motion under guidance by spatial conditions. Then, to address the highly sparse guidance, we draw inspiration from the credit assignment problem in Reinforcement Learning [50, 54], where sparse rewards can be distributed along a trajectory to allow for efficient learning [3]. Our key insight is that motion denoisers, including the diffusion model itself, can be used to expand the spatial guidance signal at a specific location to its neighboring locations without any additional model. By turning a sparse signal into a dense one by backpropagating through a denoiser, it enables us to achieve high-quality controllable motion synthesis, even with extremely sparse guidance signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.681 + ], + "angle": 0, + "content": "In summary, our contributions are: (1) Emphasis projection, a method to adjust relative importance between different parts of the representation vector, which we use to encourage coherency between spatial information and local poses to allow spatial guidance. (2) Dense signal propagation, a conditioning method to tackle the sparse guidance problem. (3) GMD, an effective spatially controllable motion generation method that enables the unexplored synthesizing of motions based on free-text and spatial conditioning by integrating the above contributions into our proposed Unet-based architecture. We provide extensive analysis to support our design decisions and show the versatility of GMD on three tasks: trajectory conditioning, keyframe conditioning, and obstacle avoidance. Additionally, GMD's model also significantly outperforms the state-of-the-art in traditional text-to-motion tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.7, + 0.642, + 0.715 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Diffusion-based probabilistic generative models (DPM). DPMs [20,47,48,49] have gained significant attention in recent years due to their impressive performance across multiple fields of research. They have been used for tasks such as image generation [12], image super-resolution [30, 46], speech synthesis [26, 26, 39], video generation [19, 22], 3D shape generation [38, 55], and reinforcement learning [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The surge in interest in DPMs may be attributed to their impressive controllable generation capabilities, including text-conditioned generation [42, 44, 45] and image editing [4, 6, 9, 18, 34]. Latent diffusion models (LDM) are an-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2152" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.122 + ], + "angle": 0, + "content": "other area of interest, which includes representation learning [27,40] and more efficient modeling techniques [7,44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.468, + 0.229 + ], + "angle": 0, + "content": "Moreover, DPMs exhibit a high degree of versatility in terms of conditioning. There are various methods for conditional generation, such as imputation and inpainting [9, 10, 34], classifier guidance [10, 12], and classifier-free guidance [21, 42, 44, 45]. Inpainting and classifier guidance can be applied to any pretrained DPM, which extends the model's capabilities further without the need for retraining." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.468, + 0.508 + ], + "angle": 0, + "content": "Human motion generation. The goal of the human motion generation task is to generate motions based on the conditioning signals. Various conditions have been explored such as partial poses [13, 17, 51], trajectories [24, 53, 60], images [8,43], music [28,29,31], text [1, 14, 15, 25, 37], objects [56], action labels [16,36], or unconditioned [35,57,61,63]. Recently, many diffusion-based motion generation models have been proposed [11, 25, 32, 62, 64] and demonstrate better quality compared to alternative models such as GAN or VAE. Employing the CLIP model [41], these models showed great improvements in the challenging text-to-motion generation task [7,58,59] as well as allowing conditioning on partial motions [51] or music [2, 52]. However, they do not support conditioning signals that are not specifically trained, for example, following keyframe locations or avoiding obstacles. Maintaining the capabilities of the diffusion models, we propose methods to enable spatial guidance without retraining the model for each new objective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.206, + 0.543 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.377, + 0.567 + ], + "angle": 0, + "content": "3.1. Diffusion-based generative models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.468, + 0.744 + ], + "angle": 0, + "content": "Diffusion-based probabilistic generative models (DPMs) are a family of generative models that learn a sequential denoising process of an input \\(\\mathbf{x}_t\\) with varying noise levels \\(t\\). The noising process of DPM is defined cumulatively as \\(q(\\mathbf{x}_t|\\mathbf{x}_0) = \\mathcal{N}(\\sqrt{\\alpha_t}\\mathbf{x}_0,(1 - \\alpha_t)\\mathbf{I})\\), where \\(\\mathbf{x}_0\\) is the clean input, \\(\\alpha_{t} = \\prod_{s = 1}^{t}(1 - \\beta_{s})\\), and \\(\\beta_{t}\\) is a noise scheduler. The denoising model \\(p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)\\) with parameters \\(\\theta\\) learns to reverse the noising process by modeling the Gaussian posterior distribution \\(q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0)\\). DPMs can map a prior distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) to any distribution \\(p(\\mathbf{x})\\) after \\(T\\) successive denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.468, + 0.791 + ], + "angle": 0, + "content": "To draw samples from a DPM, we start from a sample \\(\\mathbf{x}_T\\) from the prior distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). Then, for each \\(t\\), we sample \\(\\mathbf{x}_{t - 1}\\sim \\mathcal{N}(\\mu_t,\\Sigma_t)\\) until \\(t = 0\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.805, + 0.468, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t} = \\frac {\\sqrt {\\alpha_ {t - 1}} \\beta_ {t}}{1 - \\alpha_ {t}} \\mathbf {x} _ {0} + \\frac {\\sqrt {1 - \\beta_ {t}} (1 - \\alpha_ {t - 1})}{1 - \\alpha_ {t}} \\mathbf {x} _ {t} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.85, + 0.468, + 0.905 + ], + "angle": 0, + "content": "and \\(\\Sigma_{t}\\) is a variance scheduler of choice, usually \\(\\Sigma_{t} = \\frac{1 - \\alpha_{t-1}}{1 - \\alpha_{t}}\\beta_{t}\\) [20]. \\(\\mathbf{x}_0\\) in Eq. 1 is the prediction from a denoising model. For an \\(\\epsilon_{\\theta}\\) model, \\(\\mathbf{x}_0 = \\frac{1}{\\sqrt{\\alpha_t}}\\mathbf{x}_t + \\frac{\\sqrt{1 - \\alpha_t}}{\\sqrt{\\alpha_t}}\\epsilon_\\theta (\\mathbf{x}_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.202 + ], + "angle": 0, + "content": "There are multiple choices for the denoising model to predict including the clean input \\(\\mathbf{x}_0\\), the noise \\(\\epsilon\\), and the one-step denoised target \\(\\mu_t\\). An \\(\\mathbf{x}_{0,\\theta}\\) model is trained using the squared loss to the clean input \\(\\| \\mathbf{x}_{0,\\theta}(\\mathbf{x}_t) - \\mathbf{x}_0\\|^2\\), an \\(\\epsilon_{\\theta}\\) model is trained using the squared loss \\(\\| \\epsilon_{\\theta}(\\mathbf{x}_t) - \\epsilon \\| ^2\\), and \\(\\mu_{t,\\theta}\\) model is trained using the squared loss \\(\\| \\mu_{t,\\theta}(\\mathbf{x}_t) - \\mu_t\\|^2\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.212, + 0.891, + 0.228 + ], + "angle": 0, + "content": "3.2. Controllable generation with diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Classifier-free guidance. The conditioning signals are treated as additional inputs to the denoiser \\( p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,d) \\) where \\( d \\) is the conditioning signals which can be omitted \\( d = \\varnothing \\) to generate unconditionally. Classifier-free guidance has been shown to generate very high-quality results [44, 45, 51]. To draw samples, the effective denoiser becomes \\( \\hat{\\epsilon}_{\\theta}(\\mathbf{x}_t,d) = w\\epsilon_{\\theta}(\\mathbf{x}_t,d) + (1 - w)\\epsilon_{\\theta}(\\mathbf{x}_t,\\varnothing) \\), where \\( w \\) controls the conditional strength. The new \\( \\hat{\\epsilon}_{\\theta} \\) model can be used in Eq. 1. Two downsides of this method are that the nature of conditioning signals need to be known before hand and the denoiser needs to be adjusted and retrained for each specific case restricting its flexibility." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.43, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Classifier guidance. We can also obtain \\( p(\\mathbf{x}_{t-1}|\\mathbf{x}_t,d) \\) from \\( p_{\\theta}(\\mathbf{x}_{t-1}|\\mathbf{x}_t)p(d|\\mathbf{x}_t) \\) [12], where \\( p(d|\\mathbf{x}_t) \\) is any probability function that we can approximate its score function \\( \\nabla_{\\mathbf{x}_t}\\log p(d|\\mathbf{x}_t) \\) effectively. The new sampling process is similar to the original (Eq. 1) but with the mean shifted by the scaled score function as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.534, + 0.891, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t} = \\mu_ {t} ^ {\\prime} + s \\Sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (d | \\mathbf {x} _ {t}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.563, + 0.892, + 0.685 + ], + "angle": 0, + "content": "where \\(\\mu_t^\\prime\\) is the original mean, \\(s\\) controls the conditioning strength, and \\(\\Sigma_{t}\\) is a variance scheduler which can be the same as in Eq. 1. Since \\(\\Sigma_{t}\\) is a decreasing sequence, the guidance signal diminishes as \\(t\\to 0\\) which corresponds to the characteristic of DPMs that tend to modify \\(\\mathbf{x}_t\\) less and less as time goes. Classifier guidance is a post-hoc method, i.e., there is no change to the DPM model, one only needs to come up with \\(p(d|\\mathbf{x}_t)\\) which is extremely flexible." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.691, + 0.891, + 0.781 + ], + "angle": 0, + "content": "Imputation and inpainting. To generate human motion sequences from partial observations, such as global motion trajectories or keyframe locations, inpainting is used. These partial observations, called imputing signals, are used to adjust the generative process towards the observations. Imputation and inpainting are two sides of the same coin." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.782, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Let \\(\\mathbf{y}\\) be a partial target value in an input \\(\\mathbf{x}\\) that we want to impute. The imputation region of \\(\\mathbf{y}\\) on \\(\\mathbf{x}\\) is denoted by \\(M_y^x\\), and a projection \\(P_y^x\\) that resizes \\(\\mathbf{y}\\) to that of \\(\\mathbf{x}\\) by filling in zeros. In DPMs, imputation can be done on the sample \\(\\mathbf{x}_{t - 1}\\) after every denoising step [10]. We have the new imputed sample \\(\\tilde{\\mathbf{x}}_{t - 1}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.885, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {t - 1} = \\left(1 - M _ {y} ^ {x}\\right) \\odot \\mathbf {x} _ {t - 1} + M _ {y} ^ {x} \\odot P _ {y} ^ {x} \\mathbf {y} _ {t - 1} \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2153" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "where \\(\\odot\\) is a Hadamard product and \\(\\mathbf{y}_{t - 1}\\) is a noised target value. \\(\\mathbf{y}_{t - 1}\\sim \\mathcal{N}(\\sqrt{\\alpha_{t - 1}}\\mathbf{y},(1 - \\alpha_{t - 1})\\mathbf{I})\\) following Ho et al. [20] is one of the simplest choices of \\(\\mathbf{y}_{t - 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.469, + 0.184 + ], + "angle": 0, + "content": "Note that all three modes of conditioning presented here are not mutually exclusive. One could apply one or more tricks in a single pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.199, + 0.312, + 0.215 + ], + "angle": 0, + "content": "4. Guided Motion Diffusion" + }, + { + "type": "code_caption", + "bbox": [ + 0.081, + 0.239, + 0.45, + 0.254 + ], + "angle": 0, + "content": "Algorithm 1 GMD's two-stage guided motion diffusion" + }, + { + "type": "algorithm", + "bbox": [ + 0.081, + 0.256, + 0.469, + 0.695 + ], + "angle": 0, + "content": "Require: A trajectory DPM \\(\\mathbf{z}_{0,\\phi}\\), a motion DPM \\(\\mathbf{x}_{0,\\theta}\\), a goal function \\(G_{\\mathbf{z}}(\\cdot)\\), and keyframe locations \\(\\mathbf{y}\\) (if any). \n1: # Stage 1: Trajectory generation \n2: \\(\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) \n3: for all \\(t\\) from \\(T\\) to 1 do \n4: \\(\\mathbf{z}_0 \\gets \\mathbf{z}_{0,\\phi}(\\mathbf{z}_t)\\) \n5: \\(\\mu, \\Sigma \\gets \\mu(\\mathbf{z}_0, \\mathbf{z}_t), \\Sigma_t\\) \n6: # Classifier guidance (Eq. 2) \n7: # Dense signal propagation \n8: \\(\\mathbf{z}_{t-1} \\sim \\mathcal{N}(\\mu - s\\Sigma\\nabla_{\\mathbf{z}_t}G_z(\\mathbf{z}_0), \\Sigma)\\) \n9: # Impute y on z (Eq. 3) (if any) \n10: \\(\\mathbf{z}_{t-1} \\gets (1 - M_y^z) \\odot \\mathbf{z}_{t-1} + M_y^z\\mathbf{y}_{t-1}\\) \n11: end for \n12: # Stage 2: Trajectory-conditioned motion generation \n13: \\(\\mathbf{x}_T^{\\mathrm{proj}} \\gets\\) sample from \\(\\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) \n14: for all \\(t\\) from \\(T\\) to 1 do \n15: \\(M \\gets P_z^x M_y^z\\) # Imputation region of y on x \n16: \\(\\mathbf{x}_0^{\\mathrm{proj}} \\gets \\mathbf{x}_{0,\\theta}^{\\mathrm{proj}}(\\mathbf{x}_t^{\\mathrm{proj}})\\) # Emphasis projection \n17: # Impute y on x\\(^{\\mathrm{proj}}\\) (Eq. 6) \n18: \\(\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}} \\gets A((1 - M) \\odot A^{-1}\\mathbf{x}_0^{\\mathrm{proj}} + MP_z^x\\mathbf{y})\\) \n19: \\(\\mu, \\Sigma \\gets \\mu(\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}},\\mathbf{x}_t^{\\mathrm{proj}}), \\Sigma_t\\) \n20: # Masked classifier guidance (Eq. 9) \n21: # Dense signal propagation \n22: \\(\\Delta \\gets -s\\Sigma A^{-1}\\nabla_{\\mathbf{x}_t^{\\mathrm{proj}}}G_z(P_x^zA^{-1}\\mathbf{x}_0^{\\mathrm{proj}})\\) \n23: \\(\\mu \\gets \\mu + A(1 - M) \\odot \\Delta\\) \n24: \\(\\mathbf{x}_{t-1} \\sim \\mathcal{N}(\\mu, \\Sigma)\\) \n25: end for \n26: return \\(\\mathbf{z}_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.854 + ], + "angle": 0, + "content": "We aim to generate realistic human motions that can be guided by spatial constraints, enabling the generated human motion to achieve specific goals, such as following a global trajectory, reaching certain locations, or avoiding obstacles. Although diffusion-based models have significantly improved text-to-motion modeling [7, 51], generating motions that achieve specific goals is still beyond the reach of the current models. Our work addresses this limitation and advances the state-of-the-art in human motion modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We are interested in modeling a full-body human motion that satisfies a certain scalar goal function \\( G_{x}(\\cdot) \\) that takes a motion representation \\( \\mathbf{x} \\) and measures how far the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "motion \\(\\mathbf{x}\\) is from the goal (lower is better). More specifically, \\(\\mathbf{x} \\in \\mathbb{R}^{N \\times M}\\) represents a sequence of human poses for \\(M\\) motion steps, where \\(N\\) is the dimension of human pose representations, e.g., \\(N = 263\\) in the HumanML3D [14] dataset. Let \\(X\\) be the random variable associated with \\(\\mathbf{x}\\). Our goal is to model the following conditional probability using a motion DPM" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.208, + 0.892, + 0.226 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} \\mid G _ {x} (X) = 0\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.892, + 0.281 + ], + "angle": 0, + "content": "This can be extended to \\( p\\big(\\mathbf{x}|G_x(X) = 0,d\\big) \\), where \\( d \\) is any additional signal, such as text prompts. From now on, we omit \\( d \\) to reduce clutter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.449 + ], + "angle": 0, + "content": "Many challenging tasks in motion modeling can be encapsulated within a goal function \\( G_{z} \\) that only depends on the trajectory \\( \\mathbf{z} \\) of the human motion, not the whole motion \\( \\mathbf{x} \\). Let us define \\( \\mathbf{z} \\in \\mathbb{R}^{L \\times M} \\) to be the trajectory part of \\( \\mathbf{x} \\) with length \\( M \\) and \\( L = 2 \\) describing the ground location of the pelvis of a human body. A particular location \\( \\mathbf{z}^{(i)} \\) at motion step \\( i \\) describes the pelvis location of the human body on the ground plane. We define a projection \\( P_x^z \\) that resizes \\( \\mathbf{x} \\) to match \\( \\mathbf{z} \\) by taking only the \\( \\mathbf{z} \\) part, and its reverse \\( P_z^x \\) that resizes \\( \\mathbf{z} \\) to match \\( \\mathbf{x} \\) by filling in zeros. With this, our conditional probability becomes \\( p\\big(\\mathbf{x}|G_z(P_x^z X) = 0\\big) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.569 + ], + "angle": 0, + "content": "In this work, we will show how text-to-motion DPMs can be extended to solve several challenging tasks, including trajectory-conditioned motion generation, location-conditioned trajectory planning, and obstacle avoidance trajectory planning. Using our proposed Emphasis projection and dense signal propagation, we alleviate the sparse guidance problem and enable motion generation based on spatial conditions. The overview of our methods is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.579, + 0.694, + 0.595 + ], + "angle": 0, + "content": "4.1. Emphasis projection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.723 + ], + "angle": 0, + "content": "One of the most straightforward approaches for minimizing the goal function \\( G_{\\mathbf{z}}(\\cdot) \\) is by analyzing what trajectories that minimize \\( \\mathbf{z}^{*} = \\arg \\min_{\\mathbf{z}}G_{z}(\\mathbf{z}) \\) look like. For a trajectory conditioning task, a whole trajectory \\( \\mathbf{z}^* \\) is directly given. Our task is to generate the rest of the motion \\( \\mathbf{x} \\). With such knowledge, we can employ imputation & inpainting technique by supplying the motion DPM with the \\( \\mathbf{x} \\)-shaped \\( P_z^x\\mathbf{z}^* \\) to guide the generation process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.729, + 0.725, + 0.742 + ], + "angle": 0, + "content": "Problem 1: Motion incoherence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Since the imputing trajectory \\(\\mathbf{z}^*\\) is only a small part of the whole motion \\(\\mathbf{x}\\) (\\(L \\ll N\\)), we often observe that the DPM ignores the change from imputation and fails to make appropriate changes on the rest of \\(\\mathbf{x}\\). This results in an incoherent local motion that is not aligned or well coordinated with the imputing trajectory." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.841, + 0.729, + 0.856 + ], + "angle": 0, + "content": "Solution 1: Emphasis projection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We tackle this problem by giving more emphasis on the trajectory part of motion \\(\\mathbf{x}\\). More specifically, we propose an Emphasis projection method that increases the trajectory's" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.093, + 0.827, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.893, + 0.34 + ], + "angle": 0, + "content": "Figure 3. (a) Under standard motion representation and guiding method, only a few values in the motion representation are updated according to the guidance. (b) With Emphasis projection, all values in each frame describing the motion receive gradients w.r.t. the guidance, leading to better coherence between global orientation and local pose in each frame. (c) With dense gradient propagation, all frames are updated according to the guidance at the keyframes, making the guidance less likely to be ignored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.565 + ], + "angle": 0, + "content": "relative importance within motion \\(\\mathbf{x}\\). We achieve this by utilizing a random matrix \\(A = A'B\\), where \\(A' \\in \\mathbb{R}^{N \\times N}\\) is a matrix with elements randomly sampled from \\(\\mathcal{N}(0,1)\\) and \\(B \\in \\mathbb{R}^{N \\times N}\\) is a diagonal matrix whose trajectory-related diagonal indexes are \\(c\\) and the rest are 1 for emphasizing those trajectory elements. In our case, we emphasize the rotation and ground location of the pelvis, \\((\\mathrm{rot},x,z)\\), in \\(\\mathbf{x}\\) by \\(c\\) times. We now have a projected motion \\(\\mathbf{x}^{\\mathrm{proj}} = \\frac{1}{N - 3 + 3c^2} A\\mathbf{x}\\). Note that the fractional term is to maintain the unit variance on \\(\\mathbf{x}_{\\mathrm{proj}}\\). The noisng process of the projected motion becomes \\(q(\\mathbf{x}_t^{\\mathrm{proj}}|\\mathbf{x}_0^{\\mathrm{proj}}) = \\mathcal{N}(\\sqrt{\\alpha_t}\\mathbf{x}_0^{\\mathrm{proj}},(1 - \\alpha_t)\\mathbf{I})\\). There is no change on how a DPM that works on the projected motion \\(p_{\\theta}(\\mathbf{x}_{t - 1}^{\\mathrm{proj}}|\\mathbf{x}_t^{\\mathrm{proj}})\\) operates and treats \\(\\mathbf{x}_t^{\\mathrm{proj}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.641 + ], + "angle": 0, + "content": "In Section 6.3, we show that emphasis projection is an effective way of solving the motion incoherence problem, and is shown to be substantially better than a straightforward approach of retraining a DPM with an increased loss weight on the trajectory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Imputation on the projected motion \\(\\mathbf{x}^{\\mathrm{proj}}\\). We have discussed imputing on the sample \\(\\mathbf{x}_{t-1}\\) in Eq. 3. Here, we introduce an imputation on \\(\\mathbf{x}_0\\) which modifies the DPM's belief on the final outcome \\(\\mathbf{x}_{0,\\theta}\\) by imputing it with \\(\\mathbf{z}\\). We have found this technique useful in many tasks we are interested in." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.741, + 0.47, + 0.772 + ], + "angle": 0, + "content": "Let us define the imputation region of \\(\\mathbf{z}\\) on \\(\\mathbf{x}\\) as \\(M_z^x\\). We obtain the imputed \\(\\tilde{\\mathbf{x}}_0\\) from" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.786, + 0.469, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {0} = \\left(1 - M _ {z} ^ {x}\\right) \\odot \\mathbf {x} _ {0, \\theta} + M _ {z} ^ {x} \\odot \\underbrace {P _ {z} ^ {x} \\mathbf {z} ^ {*}} _ {\\mathbf {x} \\text {s h a p e d}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Now operating on the projected motion \\(\\mathbf{x}^{\\mathrm{proj}}\\), before we can do imputation, we need to unproject it back to the original motion using \\(\\mathbf{x}_0 = A^{-1}\\mathbf{x}_0^{\\mathrm{proj}}\\), and then project the imputed \\(\\tilde{\\mathbf{x}}_0\\) back using \\(\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}} = A\\tilde{\\mathbf{x}}_0\\). We obtain the imputed" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.799, + 0.381 + ], + "angle": 0, + "content": "motion under emphasis projection \\(\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}}\\) from" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.388, + 0.892, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {0} ^ {\\text {p r o j}} = A \\left(\\left(1 - M _ {z} ^ {x}\\right) \\odot \\left(A ^ {- 1} \\mathbf {x} _ {0, \\theta} ^ {\\text {p r o j}}\\right) + M _ {z} ^ {x} \\odot P _ {z} ^ {x} \\mathbf {z} ^ {*}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Substituting \\(\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}}\\) into Eq. 1, we obtain the new mean \\(\\tilde{\\mu}_t^{\\mathrm{proj}}\\) for sampling \\(\\mathbf{x}_{t - 1}^{\\mathrm{proj}}\\sim \\mathcal{N}(\\tilde{\\mu}_t^{\\mathrm{proj}},\\Sigma_t)\\)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.464, + 0.889, + 0.48 + ], + "angle": 0, + "content": "4.2. Dense guidance signal with a learned denoiser" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Another way to minimize the goal function \\( G_{z}(\\cdot) \\) is by adjusting the sample of each diffusion step \\( \\mathbf{x}_{t-1} \\) toward a region with lower \\( G_{z} \\). This trick is called classifier guidance [12]. The direction of change corresponds to a score function \\( \\nabla_{\\mathbf{x}_t}\\log p\\big(G_x(X_t) = 0|\\mathbf{x}_t\\big) \\) which can be approximated as a direction \\( \\Delta_{\\mathbf{x}_0} = -\\nabla_{\\mathbf{x}_0}G_z(P_x^z\\mathbf{x}_{0,\\theta}) \\) that reduces the goal function. We can guide the generative process by nudging the DPM's prediction as \\( \\mathbf{x}_0 = \\mathbf{x}_{0,\\theta} + \\Delta_{\\mathbf{x}_0} \\). While imputation requires the minimizer \\( \\mathbf{z}^* \\) of \\( G_{z} \\), which might not be easy to obtain or may not be unique, this trick only requires the easier-to-obtain direction of change." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.66, + 0.747, + 0.674 + ], + "angle": 0, + "content": "Problem 2: Sparse guidance signal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the motion domain, conditioning signals can often be sparse. There are two types of sparsity that can occur: sparsity in feature and sparsity in time. Sparsity in feature is when the conditioning signal is a small part of the feature dimension of \\(\\mathbf{x}\\). For example, in trajectory-conditioned generation, \\(\\mathbf{z}\\) may only consist of a sequence of ground locations over time. This type of sparsity can be addressed by emphasis projection, as explained in Section 4.1. Sparsity in time refers to cases where the conditioning signal consists of small segments of a trajectory spread out over time. For instance, in keyframe location conditioning task, only a sparse set of keyframe locations are given. When the conditioning signal-to-noise ratio becomes too small, the conditioning signal may be mistaken as noise and ignored during the denoising process." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2155" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.339, + 0.106 + ], + "angle": 0, + "content": "Solution 2: Dense signal propagation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.211 + ], + "angle": 0, + "content": "To turn a sparse signal into a dense signal, we need domain knowledge. One way to achieve this is by using a denoising function \\( f(\\mathbf{x}_t) = \\mathbf{x}_0 \\), which is trained on a motion dataset to denoise by gathering information from the nearby motion frames. With the ability to relate a single frame to many other frames, the denoising function is capable of expanding a sparse signal into a denser one." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.468, + 0.258 + ], + "angle": 0, + "content": "We can use backward propagation through the denoising function \\( f \\) to take advantage of this. Therefore, a dense classifier guidance can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.268, + 0.468, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {x} _ {t}} \\log p \\left(G _ {x} \\left(X _ {t}\\right) = 0 \\mid \\mathbf {x} _ {t}\\right) \\approx - \\nabla_ {\\mathbf {x} _ {t}} G _ {z} \\left(\\underbrace {P _ {x} ^ {z} f (\\mathbf {x} _ {t})} _ {\\mathbf {z} \\text {s h a p e d}}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.468, + 0.404 + ], + "angle": 0, + "content": "While an external function can be used as \\( f \\), we observe that the existing DPM model \\( \\mathbf{x}_{0,\\theta}(\\mathbf{x}_t) \\) itself is a motion de-noiser, and thus can be used to turn a sparse signal into a dense signal without the need for an additional model. In practice, this process amounts to computing the gradient of \\( G \\) with respect to \\( \\mathbf{x}_t \\) through \\( \\mathbf{x}_{0,\\theta}(\\mathbf{x}_t) \\) using autodiff." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.468, + 0.546 + ], + "angle": 0, + "content": "Applying classifier guidance together with imputation. Whenever available, we want to utilize signals from both imputation and classifier guidance techniques to help guide the generative process. Imputation is explicit but may encounter sparsity in time, while classifier guidance is indirect but dense. We want to use the direct signal from imputation wherever available (with mask \\( M_z^x \\)), and the rest from classifier guidance (with mask \\( 1 - M_z^x \\)). Based on Eq. 2, imputation-aware classifier guidance can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.557, + 0.468, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t} = \\tilde {\\mu} _ {t} - \\left(1 - M _ {z} ^ {x}\\right) \\odot s \\Sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} G _ {z} \\left(P _ {x} ^ {z} f (\\mathbf {x} _ {t})\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.468, + 0.629 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mu}\\) is an imputed sampling mean. By replacing \\(\\tilde{\\mu}\\) with \\(\\tilde{\\mu}^{\\mathrm{proj}}\\), we get classifier guidance together with imputation that works with emphasis projection as" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.64, + 0.468, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\mu} = - s \\Sigma_ {t} A ^ {- 1} \\nabla_ {\\mathbf {x} _ {t} ^ {\\mathrm {p r o j}}} G _ {z} \\left(P _ {x} ^ {z} A ^ {- 1} f \\left(\\mathbf {x} _ {t} ^ {\\mathrm {p r o j}}\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.664, + 0.468, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t} ^ {\\text {p r o j}} = \\tilde {\\mu} _ {t} ^ {\\text {p r o j}} + A \\left(1 - M _ {z} ^ {x}\\right) \\odot \\Delta_ {\\mu} \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.699, + 0.439, + 0.714 + ], + "angle": 0, + "content": "Problem 3: DPM's bias hinders the guidance signal" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.468, + 0.819 + ], + "angle": 0, + "content": "A DPM removes noise from an input based on the distribution of the training data it has seen. This could be problematic when it comes to conditional generation because the conditioning signal may be outside of the training distribution. As a result, any changes made to the classifier guidance may be reversed by the DPM in the next time step, due to its inherent bias towards the data, shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.826, + 0.285, + 0.84 + ], + "angle": 0, + "content": "Solution 3: Epsilon modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "While it is unlikely to train an unbiased DPM model, there are ways to minimize the influence of model's bias under the guidance signal. Conceptually, the DPM model usually makes less and less change near the final outcome. This is" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "in tandem with the guidance signal that gradually decreases over time due to \\(\\Sigma_t\\) (Eq. 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.89, + 0.2 + ], + "angle": 0, + "content": "We investigate the coefficient \\(\\frac{\\sqrt{\\alpha_{t-1}\\beta_t}}{1-\\alpha_t}\\) of \\(\\mathbf{x}_0\\) in the sampling mean \\(\\mu_t\\) (Eq. 1). This coefficient reaches its maximum value at \\(t = 0\\), meaning that an \\(\\mathbf{x}_{0,\\theta}\\) model could have a significant impact on the sampling mean even at \\(t = 0\\), which contradicts the weak guidance signal at that time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.891, + 0.292 + ], + "angle": 0, + "content": "On the other hand, an \\(\\epsilon_{\\theta}\\) model will have the most influence on the sampling mean at \\(t = T\\), which aligns with our intuition. In Section 6.4 and Figure 4, we demonstrate that modeling \\(\\epsilon_{\\theta}\\) instead of \\(\\mathbf{x}_{0,\\theta}\\) is a successful approach for managing the bias effect of the DPM model in classifier guidance. We further discuss this point in Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.308, + 0.63, + 0.326 + ], + "angle": 0, + "content": "5. Applications" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.8, + 0.351 + ], + "angle": 0, + "content": "5.1. Trajectory-conditioned generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.891, + 0.418 + ], + "angle": 0, + "content": "This task aims at generating a realistic motion \\(\\mathbf{x}\\) that matches a given trajectory \\(\\mathbf{z}\\). Our objective is to minimize the distance between the generated motion and the given trajectory, which we define as" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.431, + 0.89, + 0.47 + ], + "angle": 0, + "content": "\\[\nG _ {x} (\\mathbf {x}) := \\left\\| \\mathbf {z} - \\underbrace {P _ {x} ^ {z} \\mathbf {x}} _ {\\mathbf {z} \\text {p a r t o f} \\mathbf {x}} \\right\\| _ {p} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.89, + 0.544 + ], + "angle": 0, + "content": "Despite the apparent simplicity of this task, a traditional DPM faces the challenge of ensuring coherence in the generated motion. However, our emphasis projection method can effectively address this problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.795, + 0.572 + ], + "angle": 0, + "content": "5.2. Keyframe-conditioned generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.89, + 0.686 + ], + "angle": 0, + "content": "The locations of ground positions at specific times can be used to define locations that we wish the generated motion to reach. This task is a generalized version of the trajectory-conditioned generation where only a partial and potentially sparse trajectory is given. Let \\(\\mathbf{y} \\in \\mathbb{R}^{2 \\times M}\\) be a trajectory describing keyframe locations and a mask \\(M_y^z\\) describe the key motion steps. Our goal function of a motion \\(\\mathbf{x}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.699, + 0.89, + 0.727 + ], + "angle": 0, + "content": "\\[\nG _ {x} (\\mathbf {x}) := \\sum_ {i} \\left\\| M _ {y} ^ {z} \\left(P _ {x} ^ {z} \\mathbf {x} - \\mathbf {y}\\right) \\right\\| _ {p} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.89, + 0.79 + ], + "angle": 0, + "content": "Consequently, \\( G_{z}(\\mathbf{z}) = \\sum_{i}\\left\\| M_{y}^{z}(\\mathbf{z} - \\mathbf{y})\\right\\|_{p} \\). Due to the partial trajectory \\( \\mathbf{y} \\), the imputation region of \\( \\mathbf{y} \\) on \\( \\mathbf{x} \\) becomes \\( M_y^x = P_z^x M_y^z \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Two-stage guided motion generation. Generating both the trajectory and motion simultaneously under a conditioning signal can be challenging and may result in lower quality motion. To address this issue, we propose a two-step approach. First, we generate a trajectory \\(\\mathbf{z}\\) that satisfies the keyframe locations and then generate the motion \\(\\mathbf{x}\\) given the trajectory (following Section 5.1). Our overall pipeline" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "is depicted in Figure 2 (a). We offer two options for generating the trajectory from keyframe locations \\(\\mathbf{y}\\): a point-to-point trajectory and a trajectory DPM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.242 + ], + "angle": 0, + "content": "The point-to-point trajectory connects consecutive keyframe locations with a straight line. These unrealistic trajectories can be used as imputation signals for the motion DPM during the early phase \\((t \\geq \\tau)\\). If \\(\\tau\\) is large enough, the DPM will adjust the given trajectory to a higher quality one. However, if \\(\\tau\\) is too large, the DPM may generate a motion that does not perform well on \\(G_{z}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.243, + 0.468, + 0.303 + ], + "angle": 0, + "content": "The trajectory DPM \\( p_{\\phi}(\\mathbf{z}_{t - 1}|\\mathbf{z}_t) \\), which is trained using the same dataset but with a smaller network, can be used to generate the trajectory under the guidance signal from \\( G_{z} \\). We summarize our two-stage approach in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.468, + 0.394 + ], + "angle": 0, + "content": "It is also possible to combine the two methods, as the point-to-point trajectory can serve as a useful guidance signal for the trajectory DPM during \\( t \\geq \\tau \\). After that, the trajectory DPM is subject to the usual imputation and classifier guidance from \\( G_{z} \\). By tuning \\( \\tau \\), we can balance between trajectory diversity and lower scores on \\( G_{z} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.401, + 0.406, + 0.417 + ], + "angle": 0, + "content": "5.3. Obstacle avoidance motion generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.468, + 0.514 + ], + "angle": 0, + "content": "Humans have the ability to navigate around obstacles while traveling from point A to B. Under our framework, this problem can be defined using two goal functions: one that navigates from A to B, called \\( G_{x}^{\\mathrm{loc}} \\) (defined as in Eq. 12), and another that pushes back when the human model crosses the obstacle's boundary, called \\( G_{x}^{\\mathrm{obs}} \\), as follows" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.52, + 0.468, + 0.55 + ], + "angle": 0, + "content": "\\[\nG _ {x} ^ {\\mathrm {o b s}} (\\mathbf {x}) := \\sum_ {i} - \\operatorname {c l i p m a x} \\left(\\operatorname {S D F} \\left(\\left(P _ {x} ^ {z} \\mathbf {x}\\right) ^ {(i)}\\right), c\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.468, + 0.602 + ], + "angle": 0, + "content": "where \\( c \\) is the safe distance from the obstacle. These two goal functions are combined additively to obtain the final goal function, \\( G_{x}(\\mathbf{x}) = G_{x}^{\\mathrm{loc}}(\\mathbf{x}) + G_{x}^{\\mathrm{obs}}(\\mathbf{x}) \\), for this task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.468, + 0.661 + ], + "angle": 0, + "content": "We utilize the same pipeline as in Section 5.2, with the exception that imputation is not possible for obstacle avoidance. Therefore, minimizing the obstacle avoidance goal relies solely on classifier guidance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.673, + 0.209, + 0.69 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.468, + 0.834 + ], + "angle": 0, + "content": "To evaluate our methods, we perform experiments on the standard human motion generation task conditioned on text descriptors and spatial objectives. In particular, we evaluate (1) the performance of our model in standard text-condition motion generation tasks, (2) the effect of emphasis projection to alleviate incoherence between spatial locations and local poses, (3) the ability to conditionally generate motion based on spatial information by conditioning with given trajectories, keyframe locations, and obstacles." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.841, + 0.175, + 0.857 + ], + "angle": 0, + "content": "6.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate generative text-to-motion models using standard metrics introduced by Guo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.115 + ], + "angle": 0, + "content": "Table 1. Text-to-motion evaluation on the HumanML3D [14] dataset. The right arrow \\(\\rightarrow\\) means closer to real data is better." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.12, + 0.887, + 0.294 + ], + "angle": 0, + "content": "
FID ↓R-precision ↑ (Top-3)Diversity →
Real0.0020.7979.503
JL2P [1]11.020.4867.676
Text2Gesture [5]7.6640.3456.409
T2M [14]1.0670.7409.188
MotionDiffuse [59]0.6300.7829.410
MDM [51]0.5560.6089.446
MLD [7]0.4730.7729.724
PhysDiff [58]0.4330.631-
Ours0.2120.6709.440
Ours xproj0.2350.6529.726
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.477 + ], + "angle": 0, + "content": "[14]. These include Fréchet Inception Distance (FID), R- Precision, and Diversity. FID measures the distance between the distributions of ground truth and generated motion using a pretrained motion encoder. R- Precision evaluates the relevance of the generated motion and its text prompt, while Diversity measures the variability within the generated motion. We also report Foot skating ratio, which measures the proportion of frames in which either foot skids more than a certain distance \\((2.5\\mathrm{cm})\\) while maintaining contact with the ground (foot height \\(< 5\\mathrm{cm}\\)), as a proxy for the incoherence between trajectory and human motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.675 + ], + "angle": 0, + "content": "In addition, for conditional generation with keyframe locations, we use Trajectory diversity, Trajectory error, Location error, and Average error of keyframe locations. Trajectory diversity measures the root mean square distance of each location of each motion step from the average location of that motion step across multiple samples with the same settings. Trajectory error is the ratio of unsuccessful trajectories, defined as those with any keyframe location error exceeding a threshold. Location error is the ratio of keyframe locations that are not reached within a threshold distance. Average error measures the mean distance between the generated motion locations and the keyframe locations measured at the keyframe motion steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.757 + ], + "angle": 0, + "content": "Datasets. We evaluate the text-to-motion generation using the HumanML3D [14] dataset, which is a collection of text-annotate motion sequences from AMASS [33] and HumanAct12 [16] datasets. It contains 14,646 motions and 44,970 motion annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. Both our motion DPM and trajectory DPM are based on UNET with AdaGN [12] depicted in details in the Supplementary. The motion DPM is an \\(\\mathbf{x}_0\\) model, while the trajectory DPM is an \\(\\epsilon\\) model, as explained in Section 4.2, to enhance controllability. We utilized DDPM [20] with \\(T = 1,000\\) denoising steps for training and inference of both models. Additionally, we condition the generation process on text prompts in a classifier-free [21] manner, similar to MDM [51], and use the CLIP [41]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2157" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.158 + ], + "angle": 0, + "content": "Table 2. Trajectory-conditioned motions evaluation. The ground truth trajectory is used for imputing after each diffusion step. Comparing the effect of an original \\(\\mathbf{x}\\) with emphasis loss functions to the emphasis projection \\(\\mathbf{x}^{\\mathrm{proj}}\\) after imputing whole trajectories after each diffusion step." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.162, + 0.454, + 0.377 + ], + "angle": 0, + "content": "
ModelSpaceEmphasisFID ↓Foot skating ↓ ratio
MDM [51]xloss 1×0.9040.284
xprojc = 10.6320.304
c = 20.4640.309
c = 50.4660.256
c = 101.0290.161
Oursxloss 1×0.2780.262
loss 22×0.2560.250
loss 52×0.2400.249
loss 102×0.3200.265
xprojc = 10.3070.268
c = 20.2900.257
c = 50.2740.199
c = 100.1980.128
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.403, + 0.355, + 0.417 + ], + "angle": 0, + "content": "model as the text encoder across all tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.576 + ], + "angle": 0, + "content": "Computational resources Our GMD architecture is capable of running both the motion and trajectory models on a single commercial GPU, such as the Nvidia RTX 2080 Ti, 3080, or 3090. The trajectory model achieved a throughput of 2,048 samples per second when run on an RTX 3090, with a training time of approximately 4.34 GPU hours. Meanwhile, the motion model achieved a throughput of 256 samples per second on an RTX 3090, with a training time of around 34.7 GPU hours. The total inference time for one sample is approximately 110 seconds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.315, + 0.603 + ], + "angle": 0, + "content": "6.2. Text-to-motion generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.716 + ], + "angle": 0, + "content": "This section evaluates our model's performance in the standard text-to-motion generation task and compares it with other motion DPM baselines: MotionDiffuse [59], MDM [51], MLD [7], and PhysDiff [58]. Tab. 1 shows the results where our model architecture outperforms the baselines significantly in terms of motion quality measured by FID, while maintaining similar R-Precision and Diversity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.378, + 0.743 + ], + "angle": 0, + "content": "6.3. Trajectory-conditioned generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This section demonstrates how our emphasis projection method can address the issue of incoherent motion caused by spatial conditioning, specifically in the trajectory conditioning task, where the model is provided with ground-truth trajectories for imputation at each denoising step and is required to generate corresponding local poses. Both quantitative and qualitative results support that our emphasis projection leads to a reduction in Foot skating ratio, as evidenced in Tab. 2 and a more coherent motion in Fig. 5 compared to the MDM [51] model." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.093, + 0.891, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.242, + 0.894, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Comparing the evolution of the clean trajectory subject to classifier guidance from \\(\\mathbf{x}_0\\) and \\(\\epsilon\\) DPMs. The \\(\\mathbf{x}_0\\) DPM shows significant resistance on the guidance signal as exhibited by the trajectory \"contraction\" behavior at \\(t\\to 0\\)." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.303, + 0.696, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.302, + 0.888, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.407, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Figure 5. Generated motion, conditioned a given trajectory and text \"walking forward\". MDM [51] exhibits motion incoherence where the model disregards the trajectory and generates an inconsistent motion. Our method, improved by emphasis projection, deals effectively with the conditioning." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.495, + 0.887, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.613, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Figure 6. Generated motion trajectories, conditioned on target locations at given keyframes. Without dense signal propagation, the model ignores the target conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.834 + ], + "angle": 0, + "content": "We also compare our emphasis projection method with an alternative approach of increasing the trajectory loss strength during training. We include loss \\( k^2 \\times \\) baselines, where \\( k \\in \\{1, 2, 5, 10\\} \\), for comparison. The results in Tab. 2 indicate that, while increasing the loss strength marginally improves both FID and Foot skating ratio, increasing it beyond a certain point leads to a decline in both FID and Foot skating ratio. By contrast, our emphasis projection method consistently leads to improvements in both metrics. We discuss this topic further in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.795, + 0.862 + ], + "angle": 0, + "content": "6.4. Keyframe-conditioned generation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This section evaluates the quality and adherence of the generated motion to the desired goal. A viable solution" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2158" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 3. The effect of different conditioning strategies tested on keyframe-conditioning task. The keyframes \\((N = 5)\\) are sampled from the ground truth motion trajectories with the same text prompts in the HumanML3D [14] test set." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.121, + 0.892, + 0.319 + ], + "angle": 0, + "content": "
ModelConditioningFID ↓Foot ↓ skating ratioTraj. ↑ diversity (m.)Traj. err. ↓ (50 cm)Loc. err. ↓ (50 cm)Avg. err. ↓R-precision ↑ (Top-3)
MDM [51]Single stagex + τ=01.2560.2020.1340.0000.0000.0000.631
xproj + τ=02.9940.1510.1340.0000.0000.0000.554
xproj + τ=1002.2130.0950.2140.3260.1270.2360.555
xproj + no p2p1.6790.0920.3940.5190.3260.5430.548
Ours (xproj)Single stageτ=00.9020.1270.1170.0000.0000.0000.594
τ=1000.5230.0860.1570.1760.0490.1390.599
Two stageτ=1000.9370.0980.1200.0760.0200.1090.574
τ=3000.9380.0980.1270.1180.0310.1280.573
τ=5000.9080.0980.1400.1570.0430.1400.577
τ=7000.8980.0980.1620.1960.0580.1530.580
τ=9000.8740.0980.1920.2380.0800.1800.581
no p2p0.8620.1040.2220.2870.1180.2820.577
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.334, + 0.393, + 0.349 + ], + "angle": 0, + "content": "must meet both criteria to an acceptable degree." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.5 + ], + "angle": 0, + "content": "To achieve high-quality motion, both FID and Foot skating ratio are essential since FID alone cannot adequately measure the trajectory-motion coherence. Our Emphasis projection technique significantly improves motion coherence, reducing foot skating as shown in Tab. 3 while MDM [51] is unsuitable for this task due to the high motion incoherence. Furthermore, our improved architecture significantly improves motion quality in all cases. Note that without dense signal propagation, the model ignores the keyframe conditioning as shown in Fig 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.469, + 0.576 + ], + "angle": 0, + "content": "While a single-stage model performs reasonably well due to emphasis projection, it is too restrictive at \\(\\tau = 0\\) (forced trajectory), resulting in relatively high Foot skating. This issue can be addressed by allowing more modification (increasing to \\(\\tau\\) to 100) but at the cost of higher Loc. error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.469, + 0.651 + ], + "angle": 0, + "content": "Lastly, the trajectory model's better controllability reduces Location error by more than half compared to the single-stage model at \\(\\tau = 100\\). As expected, increasing \\(\\tau\\) leads to more freedom in the model, resulting in increased Trajectory diversity, lower FID, and higher Location error." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.66, + 0.406, + 0.675 + ], + "angle": 0, + "content": "6.5. Obstacle avoidance motion generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Finally, we demonstrate our model's ability to generate motion on additional guidance on the obstacle avoidance task. In this task, we randomly sample the target point that the human needs to reach at a specific motion step along with a set of obstacles it needs to avoid, represented as a 2D SDF (Sec. 5.3). We show the qualitative results in Fig 7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.329, + 0.801 + ], + "angle": 0, + "content": "7. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose GMD, a controllable human motion generation method based on goal functions. GMD produces high-quality and diverse motions and supports diverse possibilities for goal functions. Since obtaining necessary data and designing a classifier-free learning method for non-explicit goals, such as obstacle avoidance," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.333, + 0.891, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.45, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Figure 7. Qualitative results from the obstacle avoidance task given keyframe locations and obstacles. The red crossed areas represent obstacles to avoid. More results are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.595 + ], + "angle": 0, + "content": "can be challenging, our GMD utilizes a classifier-based method which allows for more conditioning flexibility without retraining the model. Thus, our studies on effective classifier guidance will be useful for further including more guiding signals." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.601, + 0.892, + 0.632 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the SNSF project grant 200021 204840." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.647, + 0.597, + 0.662 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.673, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.731, + 0.894, + 0.785 + ], + "angle": 0, + "content": "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. arXiv preprint arXiv:2211.09707, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.788, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[3] Jose A Arjona-Medina, Michael Gillhofer, Michael Widrich, Thomas Unterthiner, Johannes Brandstetter, and Sepp Hochreiter. RUDDER: Return decomposition for delayed rewards. In Advances in Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.673, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.95, + 0.516, + 0.962 + ], + "angle": 0, + "content": "2159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Liu. eDiff-I: Text-to-Image diffusion models with an ensemble of expert denoisers. Nov. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.207, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[6] T Brooks, A Holynski, and A A Efros. InstructPix2Pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.25, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. arXiv preprint arXiv:2212.04048, 2022. 1, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.305, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[8] Xin Chen, Zhuo Su, Lingbo Yang, Pei Cheng, Lan Xu, Bin Fu, and Gang Yu. Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.362, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[9] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. Aug. 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[10] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. arXiv preprint arXiv:2212.04495, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. May 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[13] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.66 + ], + "angle": 0, + "content": "[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 4, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[15] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 580-597. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[17] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-between. ACM Transactions on Graphics (TOG), 39(4):60-1, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. Aug. 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. Oct. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. June 2020. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Apr. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[23] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. May 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[24] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[25] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[26] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. Sept. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[27] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. Feb. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[28] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[29] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[30] Haoying Li, Yifan Yang, Meng Chang, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. SRDiff: Single image SuperResolution with diffusion probabilistic models. Apr. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[31] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[32] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[33] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Image synthesis and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.517, + 0.962 + ], + "angle": 0, + "content": "2160" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "editing with stochastic differential equations. Aug. 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[37] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 480-497. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. Sept. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[39] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-TTS: A diffusion probabilistic model for Text-to-Speech. May 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[40] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. ArXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[43] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution image synthesis with latent diffusion models. Dec. 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image diffusion models with deep language understanding. May 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[46] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image SuperResolution via iterative refinement. Apr. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.119 + ], + "angle": 0, + "content": "learning using nonequilibrium thermodynamics. Mar. 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.891, + 0.15 + ], + "angle": 0, + "content": "[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. July 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based generative modeling through stochastic differential equations. Nov. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[50] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, Aug. 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[51] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. Sept. 2022. 1, 3, 4, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[52] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. Edge: Editable dance generation from music. arXiv preprint arXiv:2211.10658, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[53] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[54] Christopher John Cornish Hellaby Watkins. Learning from delayed rewards. PhD thesis, King's College, Cambridge, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. Oct. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[56] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 257-274. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[57] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[58] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500, 2022. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[59] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[60] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[61] Yan Zhang, Michael J Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[62] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion gener" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.95, + 0.514, + 0.962 + ], + "angle": 0, + "content": "2161" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[63] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6225-6234, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[64] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. arXiv preprint arXiv:2211.16016, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.951, + 0.517, + 0.962 + ], + "angle": 0, + "content": "2162" + } + ] +] \ No newline at end of file diff --git a/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_origin.pdf b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1de8af3b47c6ed61d1a6ccb3c97fe6055b2776bb --- /dev/null +++ b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/e3abf2fb-5ad1-4ea8-9128-1e0bdc10e0b0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b55ec514e06f8c31113bebfcbde189985992fd3204aec8d85e2e1cd8f4cc0ea +size 4917236 diff --git a/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/full.md b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..66b4aaf8801a0d1be9aa5deaa208db6b94c9bfce --- /dev/null +++ b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/full.md @@ -0,0 +1,437 @@ +# Guided Motion Diffusion for Controllable Human Motion Synthesis + +Korrawe Karunratanakul1 + +Konpat Preechakul2 + +Supasorn Suwajanakorn2 + +Siyu Tang + +$^{1}$ ETH Zürich, Switzerland $^{2}$ VISTEC, Thailand + +https://korrawe.github.io/gmd-project/ + +![](images/478f1f5716b592c29ae9039e4fdef8b13da30f3560adf01a83d493c6779e945e.jpg) +(a) Text only + +![](images/7a55b1394bc0dd7e07665aac903cea39797258e11936deb30b1fea9e7da953fc.jpg) +(b) Text + Trajectory + +![](images/af433ba802f473570529c40f23fb907d0a7594ffe095492dead1705ae8b78ef9.jpg) +(c) Text + Keyframe locations + +![](images/b848983346230cd514c9ca12b5b1a33e4396fd4c66322b8f89cbee9bbfbbc421.jpg) +(d) Text + Keyframe locations +Obstacle avoidance +Figure 1. Our proposed Guided Motion Diffusion (GMD) can generate high-quality and diverse motions given a text prompt and a goal function. We demonstrate the controllability of GMD on four different tasks, guided by the following conditions: (a) text only, (b) text and trajectory, (c) text and keyframe locations (double circles), and (d) with obstacle avoidance (red-cross areas represent obstacles). The darker the colors, the later in time. + +# Abstract + +Denoising diffusion models have shown great promise in human motion synthesis conditioned on natural language descriptions. However, integrating spatial constraints, such as pre-defined motion trajectories and obstacles, remains a challenge despite being essential for bridging the gap between isolated human motion and its surrounding environment. To address this issue, we propose Guided Motion Diffusion (GMD), a method that incorporates spatial constraints into the motion generation process. Specifically, we propose an effective feature projection scheme that manipulates motion representation to enhance the coherency between spatial information and local poses. Together with a new imputation formulation, the generated motion can reliably conform to spatial constraints such as global motion trajectories. Furthermore, given sparse spatial constraints (e.g. sparse keyframes), we introduce a new dense guidance approach to turn a sparse signal, which is susceptible to being ignored during the reverse steps, into denser signals to guide the generated motion to the given constraints. Our extensive experiments justify the development of GMD, which achieves a significant improvement over state-of-the-art methods in text-based motion generation while allowing control of the synthesized motions with spatial constraints. + +# 1. Introduction + +Recently, denoising diffusion models have emerged as a promising approach for human motion generation [11, 62, 64] outperforming other alternatives such as GAN or VAE in terms of both quality and diversity [7, 51, 58]. Several studies have focused on generating motion based on expressive text prompts [7, 51], or music [52, 64]. The state-of-the-art motion generation methods, such as MDM [51], utilize classifier-free guidance to generate motion conditioned on text prompts. However, incorporating spatial constraints into diffusion models remains underexplored. Human motions consist of both semantic and spatial information, where the semantic aspect can be described using natural languages or action labels and the spatial aspect governs physical interaction with surroundings. To generate realistic human motion in a 3D environment, both aspects must be incorporated. Our experiments show that simply adding spatial constraint guidance, such as global trajectories, into the state-of-the-art models or using imputation and in-painting approaches do not yield satisfactory results. + +We identify two main issues that make the motion diffusion models likely to ignore the guidance when conditioned on spatial objectives: the sparseness of global orientation + +![](images/7447fc7616774b2d60110837a33d8a1f87f13a593266904511ff55fc8a4004c6.jpg) + +![](images/40e7fbf622dd39aa2616e07e88d250edcf1a17718619671b817c7348d117b601.jpg) +Figure 2. We tackle the problem of spatially conditioned motion generation with GMD, depicted in a). Our main contributions are b) Emphasis projection, for better trajectory-motion coherence, and c) Dense signal propagation, for a more controllable generation even under sparse guidance signal. + +in the motion representation and sparse frame-wise guidance signals. By design, the diffusion models are a denoising model that consecutively denoises the target output over multiple steps. With sparse guidance, a small portion of the output that receives guidance will be inconsistent with all other parts that do not, therefore, are more likely to be treated as noise and discarded in subsequent steps. + +First, the sparseness within a frame is a result of common motion representations that separate local pose information, like joint rotations, from global orientations, such as pelvis translations and rotations [43], usually with more focus on local poses. For instance, the common motion representation [14] uses 4 values to represent global orientation and 259 values for local pose in each frame. Such imbalance can cause the model to focus excessively on local pose information, and consequently, perceive guided global orientation as noise, resulting in a discrepancy such as foot skating. + +Second, in many applications such as character animation, gaming, and virtual reality, the spatial control signals are defined on only a few keyframes such as target locations on the ground. We show that the current diffusion-based motion generation models struggle to follow such sparse guidance as doing so is equivalent to guiding an image diffusion model with only a few pixels. As a result, either the guidance at the provided keyframes will be ignored during the denoising process or the output motion will contain an artifact where the character warps to satisfy the guidance only in those specific keyframes. + +To effectively incorporate sparse spatial constraints into the motion generation process, we propose GMD, a novel and principled Guided Motion Diffusion model. To alleviate the discrepancy between local pose and global orientation in the guided denoising steps, we introduce emphasis projection, a general representation manipulation method that we use to increase the importance of spatial information during training. Additionally, we derive a new imputation and inpainting formulation that enables the existing inpainting techniques to operate in the projected space, which we leverage to generate significantly more coherent motion under guidance by spatial conditions. Then, to address the highly sparse guidance, we draw inspiration from the credit assignment problem in Reinforcement Learning [50, 54], where sparse rewards can be distributed along a trajectory to allow for efficient learning [3]. Our key insight is that motion denoisers, including the diffusion model itself, can be used to expand the spatial guidance signal at a specific location to its neighboring locations without any additional model. By turning a sparse signal into a dense one by backpropagating through a denoiser, it enables us to achieve high-quality controllable motion synthesis, even with extremely sparse guidance signals. + +In summary, our contributions are: (1) Emphasis projection, a method to adjust relative importance between different parts of the representation vector, which we use to encourage coherency between spatial information and local poses to allow spatial guidance. (2) Dense signal propagation, a conditioning method to tackle the sparse guidance problem. (3) GMD, an effective spatially controllable motion generation method that enables the unexplored synthesizing of motions based on free-text and spatial conditioning by integrating the above contributions into our proposed Unet-based architecture. We provide extensive analysis to support our design decisions and show the versatility of GMD on three tasks: trajectory conditioning, keyframe conditioning, and obstacle avoidance. Additionally, GMD's model also significantly outperforms the state-of-the-art in traditional text-to-motion tasks. + +# 2. Related Work + +Diffusion-based probabilistic generative models (DPM). DPMs [20,47,48,49] have gained significant attention in recent years due to their impressive performance across multiple fields of research. They have been used for tasks such as image generation [12], image super-resolution [30, 46], speech synthesis [26, 26, 39], video generation [19, 22], 3D shape generation [38, 55], and reinforcement learning [23]. + +The surge in interest in DPMs may be attributed to their impressive controllable generation capabilities, including text-conditioned generation [42, 44, 45] and image editing [4, 6, 9, 18, 34]. Latent diffusion models (LDM) are an- + +other area of interest, which includes representation learning [27,40] and more efficient modeling techniques [7,44]. + +Moreover, DPMs exhibit a high degree of versatility in terms of conditioning. There are various methods for conditional generation, such as imputation and inpainting [9, 10, 34], classifier guidance [10, 12], and classifier-free guidance [21, 42, 44, 45]. Inpainting and classifier guidance can be applied to any pretrained DPM, which extends the model's capabilities further without the need for retraining. + +Human motion generation. The goal of the human motion generation task is to generate motions based on the conditioning signals. Various conditions have been explored such as partial poses [13, 17, 51], trajectories [24, 53, 60], images [8,43], music [28,29,31], text [1, 14, 15, 25, 37], objects [56], action labels [16,36], or unconditioned [35,57,61,63]. Recently, many diffusion-based motion generation models have been proposed [11, 25, 32, 62, 64] and demonstrate better quality compared to alternative models such as GAN or VAE. Employing the CLIP model [41], these models showed great improvements in the challenging text-to-motion generation task [7,58,59] as well as allowing conditioning on partial motions [51] or music [2, 52]. However, they do not support conditioning signals that are not specifically trained, for example, following keyframe locations or avoiding obstacles. Maintaining the capabilities of the diffusion models, we propose methods to enable spatial guidance without retraining the model for each new objective. + +# 3. Background + +# 3.1. Diffusion-based generative models + +Diffusion-based probabilistic generative models (DPMs) are a family of generative models that learn a sequential denoising process of an input $\mathbf{x}_t$ with varying noise levels $t$ . The noising process of DPM is defined cumulatively as $q(\mathbf{x}_t|\mathbf{x}_0) = \mathcal{N}(\sqrt{\alpha_t}\mathbf{x}_0,(1 - \alpha_t)\mathbf{I})$ , where $\mathbf{x}_0$ is the clean input, $\alpha_{t} = \prod_{s = 1}^{t}(1 - \beta_{s})$ , and $\beta_{t}$ is a noise scheduler. The denoising model $p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t)$ with parameters $\theta$ learns to reverse the noising process by modeling the Gaussian posterior distribution $q(\mathbf{x}_{t - 1}|\mathbf{x}_t,\mathbf{x}_0)$ . DPMs can map a prior distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ to any distribution $p(\mathbf{x})$ after $T$ successive denoising steps. + +To draw samples from a DPM, we start from a sample $\mathbf{x}_T$ from the prior distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ . Then, for each $t$ , we sample $\mathbf{x}_{t - 1}\sim \mathcal{N}(\mu_t,\Sigma_t)$ until $t = 0$ , where + +$$ +\mu_ {t} = \frac {\sqrt {\alpha_ {t - 1}} \beta_ {t}}{1 - \alpha_ {t}} \mathbf {x} _ {0} + \frac {\sqrt {1 - \beta_ {t}} (1 - \alpha_ {t - 1})}{1 - \alpha_ {t}} \mathbf {x} _ {t} \tag {1} +$$ + +and $\Sigma_{t}$ is a variance scheduler of choice, usually $\Sigma_{t} = \frac{1 - \alpha_{t-1}}{1 - \alpha_{t}}\beta_{t}$ [20]. $\mathbf{x}_0$ in Eq. 1 is the prediction from a denoising model. For an $\epsilon_{\theta}$ model, $\mathbf{x}_0 = \frac{1}{\sqrt{\alpha_t}}\mathbf{x}_t + \frac{\sqrt{1 - \alpha_t}}{\sqrt{\alpha_t}}\epsilon_\theta (\mathbf{x}_t)$ . + +There are multiple choices for the denoising model to predict including the clean input $\mathbf{x}_0$ , the noise $\epsilon$ , and the one-step denoised target $\mu_t$ . An $\mathbf{x}_{0,\theta}$ model is trained using the squared loss to the clean input $\| \mathbf{x}_{0,\theta}(\mathbf{x}_t) - \mathbf{x}_0\|^2$ , an $\epsilon_{\theta}$ model is trained using the squared loss $\| \epsilon_{\theta}(\mathbf{x}_t) - \epsilon \| ^2$ , and $\mu_{t,\theta}$ model is trained using the squared loss $\| \mu_{t,\theta}(\mathbf{x}_t) - \mu_t\|^2$ . + +# 3.2. Controllable generation with diffusion models. + +Classifier-free guidance. The conditioning signals are treated as additional inputs to the denoiser $p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t,d)$ where $d$ is the conditioning signals which can be omitted $d = \varnothing$ to generate unconditionally. Classifier-free guidance has been shown to generate very high-quality results [44, 45, 51]. To draw samples, the effective denoiser becomes $\hat{\epsilon}_{\theta}(\mathbf{x}_t,d) = w\epsilon_{\theta}(\mathbf{x}_t,d) + (1 - w)\epsilon_{\theta}(\mathbf{x}_t,\varnothing)$ , where $w$ controls the conditional strength. The new $\hat{\epsilon}_{\theta}$ model can be used in Eq. 1. Two downsides of this method are that the nature of conditioning signals need to be known before hand and the denoiser needs to be adjusted and retrained for each specific case restricting its flexibility. + +Classifier guidance. We can also obtain $p(\mathbf{x}_{t-1}|\mathbf{x}_t,d)$ from $p_{\theta}(\mathbf{x}_{t-1}|\mathbf{x}_t)p(d|\mathbf{x}_t)$ [12], where $p(d|\mathbf{x}_t)$ is any probability function that we can approximate its score function $\nabla_{\mathbf{x}_t}\log p(d|\mathbf{x}_t)$ effectively. The new sampling process is similar to the original (Eq. 1) but with the mean shifted by the scaled score function as + +$$ +\mu_ {t} = \mu_ {t} ^ {\prime} + s \Sigma_ {t} \nabla_ {\mathbf {x} _ {t}} \log p (d | \mathbf {x} _ {t}) \tag {2} +$$ + +where $\mu_t^\prime$ is the original mean, $s$ controls the conditioning strength, and $\Sigma_{t}$ is a variance scheduler which can be the same as in Eq. 1. Since $\Sigma_{t}$ is a decreasing sequence, the guidance signal diminishes as $t\to 0$ which corresponds to the characteristic of DPMs that tend to modify $\mathbf{x}_t$ less and less as time goes. Classifier guidance is a post-hoc method, i.e., there is no change to the DPM model, one only needs to come up with $p(d|\mathbf{x}_t)$ which is extremely flexible. + +Imputation and inpainting. To generate human motion sequences from partial observations, such as global motion trajectories or keyframe locations, inpainting is used. These partial observations, called imputing signals, are used to adjust the generative process towards the observations. Imputation and inpainting are two sides of the same coin. + +Let $\mathbf{y}$ be a partial target value in an input $\mathbf{x}$ that we want to impute. The imputation region of $\mathbf{y}$ on $\mathbf{x}$ is denoted by $M_y^x$ , and a projection $P_y^x$ that resizes $\mathbf{y}$ to that of $\mathbf{x}$ by filling in zeros. In DPMs, imputation can be done on the sample $\mathbf{x}_{t - 1}$ after every denoising step [10]. We have the new imputed sample $\tilde{\mathbf{x}}_{t - 1}$ as + +$$ +\tilde {\mathbf {x}} _ {t - 1} = \left(1 - M _ {y} ^ {x}\right) \odot \mathbf {x} _ {t - 1} + M _ {y} ^ {x} \odot P _ {y} ^ {x} \mathbf {y} _ {t - 1} \tag {3} +$$ + +where $\odot$ is a Hadamard product and $\mathbf{y}_{t - 1}$ is a noised target value. $\mathbf{y}_{t - 1}\sim \mathcal{N}(\sqrt{\alpha_{t - 1}}\mathbf{y},(1 - \alpha_{t - 1})\mathbf{I})$ following Ho et al. [20] is one of the simplest choices of $\mathbf{y}_{t - 1}$ . + +Note that all three modes of conditioning presented here are not mutually exclusive. One could apply one or more tricks in a single pipeline. + +# 4. Guided Motion Diffusion + +Algorithm 1 GMD's two-stage guided motion diffusion +Require: A trajectory DPM $\mathbf{z}_{0,\phi}$ , a motion DPM $\mathbf{x}_{0,\theta}$ , a goal function $G_{\mathbf{z}}(\cdot)$ , and keyframe locations $\mathbf{y}$ (if any). +1: # Stage 1: Trajectory generation +2: $\mathbf{z}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ +3: for all $t$ from $T$ to 1 do +4: $\mathbf{z}_0 \gets \mathbf{z}_{0,\phi}(\mathbf{z}_t)$ +5: $\mu, \Sigma \gets \mu(\mathbf{z}_0, \mathbf{z}_t), \Sigma_t$ +6: # Classifier guidance (Eq. 2) +7: # Dense signal propagation +8: $\mathbf{z}_{t-1} \sim \mathcal{N}(\mu - s\Sigma\nabla_{\mathbf{z}_t}G_z(\mathbf{z}_0), \Sigma)$ +9: # Impute y on z (Eq. 3) (if any) +10: $\mathbf{z}_{t-1} \gets (1 - M_y^z) \odot \mathbf{z}_{t-1} + M_y^z\mathbf{y}_{t-1}$ +11: end for +12: # Stage 2: Trajectory-conditioned motion generation +13: $\mathbf{x}_T^{\mathrm{proj}} \gets$ sample from $\mathcal{N}(\mathbf{0}, \mathbf{I})$ +14: for all $t$ from $T$ to 1 do +15: $M \gets P_z^x M_y^z$ # Imputation region of y on x +16: $\mathbf{x}_0^{\mathrm{proj}} \gets \mathbf{x}_{0,\theta}^{\mathrm{proj}}(\mathbf{x}_t^{\mathrm{proj}})$ # Emphasis projection +17: # Impute y on x $^{\mathrm{proj}}$ (Eq. 6) +18: $\tilde{\mathbf{x}}_0^{\mathrm{proj}} \gets A((1 - M) \odot A^{-1}\mathbf{x}_0^{\mathrm{proj}} + MP_z^x\mathbf{y})$ +19: $\mu, \Sigma \gets \mu(\tilde{\mathbf{x}}_0^{\mathrm{proj}},\mathbf{x}_t^{\mathrm{proj}}), \Sigma_t$ +20: # Masked classifier guidance (Eq. 9) +21: # Dense signal propagation +22: $\Delta \gets -s\Sigma A^{-1}\nabla_{\mathbf{x}_t^{\mathrm{proj}}}G_z(P_x^zA^{-1}\mathbf{x}_0^{\mathrm{proj}})$ +23: $\mu \gets \mu + A(1 - M) \odot \Delta$ +24: $\mathbf{x}_{t-1} \sim \mathcal{N}(\mu, \Sigma)$ +25: end for +26: return $\mathbf{z}_0$ + +We aim to generate realistic human motions that can be guided by spatial constraints, enabling the generated human motion to achieve specific goals, such as following a global trajectory, reaching certain locations, or avoiding obstacles. Although diffusion-based models have significantly improved text-to-motion modeling [7, 51], generating motions that achieve specific goals is still beyond the reach of the current models. Our work addresses this limitation and advances the state-of-the-art in human motion modeling. + +We are interested in modeling a full-body human motion that satisfies a certain scalar goal function $G_{x}(\cdot)$ that takes a motion representation $\mathbf{x}$ and measures how far the + +motion $\mathbf{x}$ is from the goal (lower is better). More specifically, $\mathbf{x} \in \mathbb{R}^{N \times M}$ represents a sequence of human poses for $M$ motion steps, where $N$ is the dimension of human pose representations, e.g., $N = 263$ in the HumanML3D [14] dataset. Let $X$ be the random variable associated with $\mathbf{x}$ . Our goal is to model the following conditional probability using a motion DPM + +$$ +p \left(\mathbf {x} \mid G _ {x} (X) = 0\right) \tag {4} +$$ + +This can be extended to $p\big(\mathbf{x}|G_x(X) = 0,d\big)$ , where $d$ is any additional signal, such as text prompts. From now on, we omit $d$ to reduce clutter. + +Many challenging tasks in motion modeling can be encapsulated within a goal function $G_{z}$ that only depends on the trajectory $\mathbf{z}$ of the human motion, not the whole motion $\mathbf{x}$ . Let us define $\mathbf{z} \in \mathbb{R}^{L \times M}$ to be the trajectory part of $\mathbf{x}$ with length $M$ and $L = 2$ describing the ground location of the pelvis of a human body. A particular location $\mathbf{z}^{(i)}$ at motion step $i$ describes the pelvis location of the human body on the ground plane. We define a projection $P_x^z$ that resizes $\mathbf{x}$ to match $\mathbf{z}$ by taking only the $\mathbf{z}$ part, and its reverse $P_z^x$ that resizes $\mathbf{z}$ to match $\mathbf{x}$ by filling in zeros. With this, our conditional probability becomes $p\big(\mathbf{x}|G_z(P_x^z X) = 0\big)$ . + +In this work, we will show how text-to-motion DPMs can be extended to solve several challenging tasks, including trajectory-conditioned motion generation, location-conditioned trajectory planning, and obstacle avoidance trajectory planning. Using our proposed Emphasis projection and dense signal propagation, we alleviate the sparse guidance problem and enable motion generation based on spatial conditions. The overview of our methods is shown in Fig. 3. + +# 4.1. Emphasis projection + +One of the most straightforward approaches for minimizing the goal function $G_{\mathbf{z}}(\cdot)$ is by analyzing what trajectories that minimize $\mathbf{z}^{*} = \arg \min_{\mathbf{z}}G_{z}(\mathbf{z})$ look like. For a trajectory conditioning task, a whole trajectory $\mathbf{z}^*$ is directly given. Our task is to generate the rest of the motion $\mathbf{x}$ . With such knowledge, we can employ imputation & inpainting technique by supplying the motion DPM with the $\mathbf{x}$ -shaped $P_z^x\mathbf{z}^*$ to guide the generation process. + +# Problem 1: Motion incoherence + +Since the imputing trajectory $\mathbf{z}^*$ is only a small part of the whole motion $\mathbf{x}$ ( $L \ll N$ ), we often observe that the DPM ignores the change from imputation and fails to make appropriate changes on the rest of $\mathbf{x}$ . This results in an incoherent local motion that is not aligned or well coordinated with the imputing trajectory. + +# Solution 1: Emphasis projection + +We tackle this problem by giving more emphasis on the trajectory part of motion $\mathbf{x}$ . More specifically, we propose an Emphasis projection method that increases the trajectory's + +![](images/2aada7e8aaab10721a928ae76978d7292c606cbb8db4108324dd3cf30e9c082c.jpg) +Figure 3. (a) Under standard motion representation and guiding method, only a few values in the motion representation are updated according to the guidance. (b) With Emphasis projection, all values in each frame describing the motion receive gradients w.r.t. the guidance, leading to better coherence between global orientation and local pose in each frame. (c) With dense gradient propagation, all frames are updated according to the guidance at the keyframes, making the guidance less likely to be ignored. + +relative importance within motion $\mathbf{x}$ . We achieve this by utilizing a random matrix $A = A'B$ , where $A' \in \mathbb{R}^{N \times N}$ is a matrix with elements randomly sampled from $\mathcal{N}(0,1)$ and $B \in \mathbb{R}^{N \times N}$ is a diagonal matrix whose trajectory-related diagonal indexes are $c$ and the rest are 1 for emphasizing those trajectory elements. In our case, we emphasize the rotation and ground location of the pelvis, $(\mathrm{rot},x,z)$ , in $\mathbf{x}$ by $c$ times. We now have a projected motion $\mathbf{x}^{\mathrm{proj}} = \frac{1}{N - 3 + 3c^2} A\mathbf{x}$ . Note that the fractional term is to maintain the unit variance on $\mathbf{x}_{\mathrm{proj}}$ . The noisng process of the projected motion becomes $q(\mathbf{x}_t^{\mathrm{proj}}|\mathbf{x}_0^{\mathrm{proj}}) = \mathcal{N}(\sqrt{\alpha_t}\mathbf{x}_0^{\mathrm{proj}},(1 - \alpha_t)\mathbf{I})$ . There is no change on how a DPM that works on the projected motion $p_{\theta}(\mathbf{x}_{t - 1}^{\mathrm{proj}}|\mathbf{x}_t^{\mathrm{proj}})$ operates and treats $\mathbf{x}_t^{\mathrm{proj}}$ . + +In Section 6.3, we show that emphasis projection is an effective way of solving the motion incoherence problem, and is shown to be substantially better than a straightforward approach of retraining a DPM with an increased loss weight on the trajectory. + +Imputation on the projected motion $\mathbf{x}^{\mathrm{proj}}$ . We have discussed imputing on the sample $\mathbf{x}_{t-1}$ in Eq. 3. Here, we introduce an imputation on $\mathbf{x}_0$ which modifies the DPM's belief on the final outcome $\mathbf{x}_{0,\theta}$ by imputing it with $\mathbf{z}$ . We have found this technique useful in many tasks we are interested in. + +Let us define the imputation region of $\mathbf{z}$ on $\mathbf{x}$ as $M_z^x$ . We obtain the imputed $\tilde{\mathbf{x}}_0$ from + +$$ +\tilde {\mathbf {x}} _ {0} = \left(1 - M _ {z} ^ {x}\right) \odot \mathbf {x} _ {0, \theta} + M _ {z} ^ {x} \odot \underbrace {P _ {z} ^ {x} \mathbf {z} ^ {*}} _ {\mathbf {x} \text {s h a p e d}} \tag {5} +$$ + +Now operating on the projected motion $\mathbf{x}^{\mathrm{proj}}$ , before we can do imputation, we need to unproject it back to the original motion using $\mathbf{x}_0 = A^{-1}\mathbf{x}_0^{\mathrm{proj}}$ , and then project the imputed $\tilde{\mathbf{x}}_0$ back using $\tilde{\mathbf{x}}_0^{\mathrm{proj}} = A\tilde{\mathbf{x}}_0$ . We obtain the imputed + +motion under emphasis projection $\tilde{\mathbf{x}}_0^{\mathrm{proj}}$ from + +$$ +\tilde {\mathbf {x}} _ {0} ^ {\text {p r o j}} = A \left(\left(1 - M _ {z} ^ {x}\right) \odot \left(A ^ {- 1} \mathbf {x} _ {0, \theta} ^ {\text {p r o j}}\right) + M _ {z} ^ {x} \odot P _ {z} ^ {x} \mathbf {z} ^ {*}\right) \tag {6} +$$ + +Substituting $\tilde{\mathbf{x}}_0^{\mathrm{proj}}$ into Eq. 1, we obtain the new mean $\tilde{\mu}_t^{\mathrm{proj}}$ for sampling $\mathbf{x}_{t - 1}^{\mathrm{proj}}\sim \mathcal{N}(\tilde{\mu}_t^{\mathrm{proj}},\Sigma_t)$ + +# 4.2. Dense guidance signal with a learned denoiser + +Another way to minimize the goal function $G_{z}(\cdot)$ is by adjusting the sample of each diffusion step $\mathbf{x}_{t-1}$ toward a region with lower $G_{z}$ . This trick is called classifier guidance [12]. The direction of change corresponds to a score function $\nabla_{\mathbf{x}_t}\log p\big(G_x(X_t) = 0|\mathbf{x}_t\big)$ which can be approximated as a direction $\Delta_{\mathbf{x}_0} = -\nabla_{\mathbf{x}_0}G_z(P_x^z\mathbf{x}_{0,\theta})$ that reduces the goal function. We can guide the generative process by nudging the DPM's prediction as $\mathbf{x}_0 = \mathbf{x}_{0,\theta} + \Delta_{\mathbf{x}_0}$ . While imputation requires the minimizer $\mathbf{z}^*$ of $G_{z}$ , which might not be easy to obtain or may not be unique, this trick only requires the easier-to-obtain direction of change. + +# Problem 2: Sparse guidance signal + +In the motion domain, conditioning signals can often be sparse. There are two types of sparsity that can occur: sparsity in feature and sparsity in time. Sparsity in feature is when the conditioning signal is a small part of the feature dimension of $\mathbf{x}$ . For example, in trajectory-conditioned generation, $\mathbf{z}$ may only consist of a sequence of ground locations over time. This type of sparsity can be addressed by emphasis projection, as explained in Section 4.1. Sparsity in time refers to cases where the conditioning signal consists of small segments of a trajectory spread out over time. For instance, in keyframe location conditioning task, only a sparse set of keyframe locations are given. When the conditioning signal-to-noise ratio becomes too small, the conditioning signal may be mistaken as noise and ignored during the denoising process. + +# Solution 2: Dense signal propagation + +To turn a sparse signal into a dense signal, we need domain knowledge. One way to achieve this is by using a denoising function $f(\mathbf{x}_t) = \mathbf{x}_0$ , which is trained on a motion dataset to denoise by gathering information from the nearby motion frames. With the ability to relate a single frame to many other frames, the denoising function is capable of expanding a sparse signal into a denser one. + +We can use backward propagation through the denoising function $f$ to take advantage of this. Therefore, a dense classifier guidance can be obtained as follows: + +$$ +\nabla_ {\mathbf {x} _ {t}} \log p \left(G _ {x} \left(X _ {t}\right) = 0 \mid \mathbf {x} _ {t}\right) \approx - \nabla_ {\mathbf {x} _ {t}} G _ {z} \left(\underbrace {P _ {x} ^ {z} f (\mathbf {x} _ {t})} _ {\mathbf {z} \text {s h a p e d}}\right) \tag {7} +$$ + +While an external function can be used as $f$ , we observe that the existing DPM model $\mathbf{x}_{0,\theta}(\mathbf{x}_t)$ itself is a motion de-noiser, and thus can be used to turn a sparse signal into a dense signal without the need for an additional model. In practice, this process amounts to computing the gradient of $G$ with respect to $\mathbf{x}_t$ through $\mathbf{x}_{0,\theta}(\mathbf{x}_t)$ using autodiff. + +Applying classifier guidance together with imputation. Whenever available, we want to utilize signals from both imputation and classifier guidance techniques to help guide the generative process. Imputation is explicit but may encounter sparsity in time, while classifier guidance is indirect but dense. We want to use the direct signal from imputation wherever available (with mask $M_z^x$ ), and the rest from classifier guidance (with mask $1 - M_z^x$ ). Based on Eq. 2, imputation-aware classifier guidance can be written as + +$$ +\mu_ {t} = \tilde {\mu} _ {t} - \left(1 - M _ {z} ^ {x}\right) \odot s \Sigma_ {t} \nabla_ {\mathbf {x} _ {t}} G _ {z} \left(P _ {x} ^ {z} f (\mathbf {x} _ {t})\right) \tag {8} +$$ + +where $\tilde{\mu}$ is an imputed sampling mean. By replacing $\tilde{\mu}$ with $\tilde{\mu}^{\mathrm{proj}}$ , we get classifier guidance together with imputation that works with emphasis projection as + +$$ +\Delta_ {\mu} = - s \Sigma_ {t} A ^ {- 1} \nabla_ {\mathbf {x} _ {t} ^ {\mathrm {p r o j}}} G _ {z} \left(P _ {x} ^ {z} A ^ {- 1} f \left(\mathbf {x} _ {t} ^ {\mathrm {p r o j}}\right)\right) \tag {9} +$$ + +$$ +\mu_ {t} ^ {\text {p r o j}} = \tilde {\mu} _ {t} ^ {\text {p r o j}} + A \left(1 - M _ {z} ^ {x}\right) \odot \Delta_ {\mu} \tag {10} +$$ + +# Problem 3: DPM's bias hinders the guidance signal + +A DPM removes noise from an input based on the distribution of the training data it has seen. This could be problematic when it comes to conditional generation because the conditioning signal may be outside of the training distribution. As a result, any changes made to the classifier guidance may be reversed by the DPM in the next time step, due to its inherent bias towards the data, shown in Figure 4. + +# Solution 3: Epsilon modeling + +While it is unlikely to train an unbiased DPM model, there are ways to minimize the influence of model's bias under the guidance signal. Conceptually, the DPM model usually makes less and less change near the final outcome. This is + +in tandem with the guidance signal that gradually decreases over time due to $\Sigma_t$ (Eq. 2). + +We investigate the coefficient $\frac{\sqrt{\alpha_{t-1}\beta_t}}{1-\alpha_t}$ of $\mathbf{x}_0$ in the sampling mean $\mu_t$ (Eq. 1). This coefficient reaches its maximum value at $t = 0$ , meaning that an $\mathbf{x}_{0,\theta}$ model could have a significant impact on the sampling mean even at $t = 0$ , which contradicts the weak guidance signal at that time. + +On the other hand, an $\epsilon_{\theta}$ model will have the most influence on the sampling mean at $t = T$ , which aligns with our intuition. In Section 6.4 and Figure 4, we demonstrate that modeling $\epsilon_{\theta}$ instead of $\mathbf{x}_{0,\theta}$ is a successful approach for managing the bias effect of the DPM model in classifier guidance. We further discuss this point in Supplementary. + +# 5. Applications + +# 5.1. Trajectory-conditioned generation + +This task aims at generating a realistic motion $\mathbf{x}$ that matches a given trajectory $\mathbf{z}$ . Our objective is to minimize the distance between the generated motion and the given trajectory, which we define as + +$$ +G _ {x} (\mathbf {x}) := \left\| \mathbf {z} - \underbrace {P _ {x} ^ {z} \mathbf {x}} _ {\mathbf {z} \text {p a r t o f} \mathbf {x}} \right\| _ {p} \tag {11} +$$ + +Despite the apparent simplicity of this task, a traditional DPM faces the challenge of ensuring coherence in the generated motion. However, our emphasis projection method can effectively address this problem. + +# 5.2. Keyframe-conditioned generation + +The locations of ground positions at specific times can be used to define locations that we wish the generated motion to reach. This task is a generalized version of the trajectory-conditioned generation where only a partial and potentially sparse trajectory is given. Let $\mathbf{y} \in \mathbb{R}^{2 \times M}$ be a trajectory describing keyframe locations and a mask $M_y^z$ describe the key motion steps. Our goal function of a motion $\mathbf{x}$ is + +$$ +G _ {x} (\mathbf {x}) := \sum_ {i} \left\| M _ {y} ^ {z} \left(P _ {x} ^ {z} \mathbf {x} - \mathbf {y}\right) \right\| _ {p} \tag {12} +$$ + +Consequently, $G_{z}(\mathbf{z}) = \sum_{i}\left\| M_{y}^{z}(\mathbf{z} - \mathbf{y})\right\|_{p}$ . Due to the partial trajectory $\mathbf{y}$ , the imputation region of $\mathbf{y}$ on $\mathbf{x}$ becomes $M_y^x = P_z^x M_y^z$ . + +Two-stage guided motion generation. Generating both the trajectory and motion simultaneously under a conditioning signal can be challenging and may result in lower quality motion. To address this issue, we propose a two-step approach. First, we generate a trajectory $\mathbf{z}$ that satisfies the keyframe locations and then generate the motion $\mathbf{x}$ given the trajectory (following Section 5.1). Our overall pipeline + +is depicted in Figure 2 (a). We offer two options for generating the trajectory from keyframe locations $\mathbf{y}$ : a point-to-point trajectory and a trajectory DPM. + +The point-to-point trajectory connects consecutive keyframe locations with a straight line. These unrealistic trajectories can be used as imputation signals for the motion DPM during the early phase $(t \geq \tau)$ . If $\tau$ is large enough, the DPM will adjust the given trajectory to a higher quality one. However, if $\tau$ is too large, the DPM may generate a motion that does not perform well on $G_{z}$ . + +The trajectory DPM $p_{\phi}(\mathbf{z}_{t - 1}|\mathbf{z}_t)$ , which is trained using the same dataset but with a smaller network, can be used to generate the trajectory under the guidance signal from $G_{z}$ . We summarize our two-stage approach in Algorithm 1. + +It is also possible to combine the two methods, as the point-to-point trajectory can serve as a useful guidance signal for the trajectory DPM during $t \geq \tau$ . After that, the trajectory DPM is subject to the usual imputation and classifier guidance from $G_{z}$ . By tuning $\tau$ , we can balance between trajectory diversity and lower scores on $G_{z}$ . + +# 5.3. Obstacle avoidance motion generation + +Humans have the ability to navigate around obstacles while traveling from point A to B. Under our framework, this problem can be defined using two goal functions: one that navigates from A to B, called $G_{x}^{\mathrm{loc}}$ (defined as in Eq. 12), and another that pushes back when the human model crosses the obstacle's boundary, called $G_{x}^{\mathrm{obs}}$ , as follows + +$$ +G _ {x} ^ {\mathrm {o b s}} (\mathbf {x}) := \sum_ {i} - \operatorname {c l i p m a x} \left(\operatorname {S D F} \left(\left(P _ {x} ^ {z} \mathbf {x}\right) ^ {(i)}\right), c\right) \tag {13} +$$ + +where $c$ is the safe distance from the obstacle. These two goal functions are combined additively to obtain the final goal function, $G_{x}(\mathbf{x}) = G_{x}^{\mathrm{loc}}(\mathbf{x}) + G_{x}^{\mathrm{obs}}(\mathbf{x})$ , for this task. + +We utilize the same pipeline as in Section 5.2, with the exception that imputation is not possible for obstacle avoidance. Therefore, minimizing the obstacle avoidance goal relies solely on classifier guidance. + +# 6. Experiments + +To evaluate our methods, we perform experiments on the standard human motion generation task conditioned on text descriptors and spatial objectives. In particular, we evaluate (1) the performance of our model in standard text-condition motion generation tasks, (2) the effect of emphasis projection to alleviate incoherence between spatial locations and local poses, (3) the ability to conditionally generate motion based on spatial information by conditioning with given trajectories, keyframe locations, and obstacles. + +# 6.1. Settings + +Evaluation metrics. We evaluate generative text-to-motion models using standard metrics introduced by Guo et al. + +Table 1. Text-to-motion evaluation on the HumanML3D [14] dataset. The right arrow $\rightarrow$ means closer to real data is better. + +
FID ↓R-precision ↑ (Top-3)Diversity →
Real0.0020.7979.503
JL2P [1]11.020.4867.676
Text2Gesture [5]7.6640.3456.409
T2M [14]1.0670.7409.188
MotionDiffuse [59]0.6300.7829.410
MDM [51]0.5560.6089.446
MLD [7]0.4730.7729.724
PhysDiff [58]0.4330.631-
Ours0.2120.6709.440
Ours xproj0.2350.6529.726
+ +[14]. These include Fréchet Inception Distance (FID), R- Precision, and Diversity. FID measures the distance between the distributions of ground truth and generated motion using a pretrained motion encoder. R- Precision evaluates the relevance of the generated motion and its text prompt, while Diversity measures the variability within the generated motion. We also report Foot skating ratio, which measures the proportion of frames in which either foot skids more than a certain distance $(2.5\mathrm{cm})$ while maintaining contact with the ground (foot height $< 5\mathrm{cm}$ ), as a proxy for the incoherence between trajectory and human motion. + +In addition, for conditional generation with keyframe locations, we use Trajectory diversity, Trajectory error, Location error, and Average error of keyframe locations. Trajectory diversity measures the root mean square distance of each location of each motion step from the average location of that motion step across multiple samples with the same settings. Trajectory error is the ratio of unsuccessful trajectories, defined as those with any keyframe location error exceeding a threshold. Location error is the ratio of keyframe locations that are not reached within a threshold distance. Average error measures the mean distance between the generated motion locations and the keyframe locations measured at the keyframe motion steps. + +Datasets. We evaluate the text-to-motion generation using the HumanML3D [14] dataset, which is a collection of text-annotate motion sequences from AMASS [33] and HumanAct12 [16] datasets. It contains 14,646 motions and 44,970 motion annotations. + +Implementation details. Both our motion DPM and trajectory DPM are based on UNET with AdaGN [12] depicted in details in the Supplementary. The motion DPM is an $\mathbf{x}_0$ model, while the trajectory DPM is an $\epsilon$ model, as explained in Section 4.2, to enhance controllability. We utilized DDPM [20] with $T = 1,000$ denoising steps for training and inference of both models. Additionally, we condition the generation process on text prompts in a classifier-free [21] manner, similar to MDM [51], and use the CLIP [41] + +Table 2. Trajectory-conditioned motions evaluation. The ground truth trajectory is used for imputing after each diffusion step. Comparing the effect of an original $\mathbf{x}$ with emphasis loss functions to the emphasis projection $\mathbf{x}^{\mathrm{proj}}$ after imputing whole trajectories after each diffusion step. + +
ModelSpaceEmphasisFID ↓Foot skating ↓ ratio
MDM [51]xloss 1×0.9040.284
xprojc = 10.6320.304
c = 20.4640.309
c = 50.4660.256
c = 101.0290.161
Oursxloss 1×0.2780.262
loss 22×0.2560.250
loss 52×0.2400.249
loss 102×0.3200.265
xprojc = 10.3070.268
c = 20.2900.257
c = 50.2740.199
c = 100.1980.128
+ +model as the text encoder across all tasks. + +Computational resources Our GMD architecture is capable of running both the motion and trajectory models on a single commercial GPU, such as the Nvidia RTX 2080 Ti, 3080, or 3090. The trajectory model achieved a throughput of 2,048 samples per second when run on an RTX 3090, with a training time of approximately 4.34 GPU hours. Meanwhile, the motion model achieved a throughput of 256 samples per second on an RTX 3090, with a training time of around 34.7 GPU hours. The total inference time for one sample is approximately 110 seconds. + +# 6.2. Text-to-motion generation + +This section evaluates our model's performance in the standard text-to-motion generation task and compares it with other motion DPM baselines: MotionDiffuse [59], MDM [51], MLD [7], and PhysDiff [58]. Tab. 1 shows the results where our model architecture outperforms the baselines significantly in terms of motion quality measured by FID, while maintaining similar R-Precision and Diversity. + +# 6.3. Trajectory-conditioned generation + +This section demonstrates how our emphasis projection method can address the issue of incoherent motion caused by spatial conditioning, specifically in the trajectory conditioning task, where the model is provided with ground-truth trajectories for imputation at each denoising step and is required to generate corresponding local poses. Both quantitative and qualitative results support that our emphasis projection leads to a reduction in Foot skating ratio, as evidenced in Tab. 2 and a more coherent motion in Fig. 5 compared to the MDM [51] model. + +![](images/f06bfcd27c56b984d76b2308cb8bbe7690b497403a79a5a9c44138f544871d35.jpg) +Figure 4. Comparing the evolution of the clean trajectory subject to classifier guidance from $\mathbf{x}_0$ and $\epsilon$ DPMs. The $\mathbf{x}_0$ DPM shows significant resistance on the guidance signal as exhibited by the trajectory "contraction" behavior at $t\to 0$ . + +![](images/159c388c323ff7de7a83b347165a0566f84fcc767a9165a6eb8a7bb7207009b3.jpg) +Figure 5. Generated motion, conditioned a given trajectory and text "walking forward". MDM [51] exhibits motion incoherence where the model disregards the trajectory and generates an inconsistent motion. Our method, improved by emphasis projection, deals effectively with the conditioning. + +![](images/5f64d7f192fdad6bad7a73d0ea65c795dd1bb50bdcd0c8c19f07b28e63e655d2.jpg) + +![](images/ca8ba36a0fc61da6ac016941d9b88c5a4eaa55b07059cfc44b199f1a4fcf9924.jpg) +Figure 6. Generated motion trajectories, conditioned on target locations at given keyframes. Without dense signal propagation, the model ignores the target conditions. + +We also compare our emphasis projection method with an alternative approach of increasing the trajectory loss strength during training. We include loss $k^2 \times$ baselines, where $k \in \{1, 2, 5, 10\}$ , for comparison. The results in Tab. 2 indicate that, while increasing the loss strength marginally improves both FID and Foot skating ratio, increasing it beyond a certain point leads to a decline in both FID and Foot skating ratio. By contrast, our emphasis projection method consistently leads to improvements in both metrics. We discuss this topic further in the Supplementary. + +# 6.4. Keyframe-conditioned generation + +This section evaluates the quality and adherence of the generated motion to the desired goal. A viable solution + +Table 3. The effect of different conditioning strategies tested on keyframe-conditioning task. The keyframes $(N = 5)$ are sampled from the ground truth motion trajectories with the same text prompts in the HumanML3D [14] test set. + +
ModelConditioningFID ↓Foot ↓ skating ratioTraj. ↑ diversity (m.)Traj. err. ↓ (50 cm)Loc. err. ↓ (50 cm)Avg. err. ↓R-precision ↑ (Top-3)
MDM [51]Single stagex + τ=01.2560.2020.1340.0000.0000.0000.631
xproj + τ=02.9940.1510.1340.0000.0000.0000.554
xproj + τ=1002.2130.0950.2140.3260.1270.2360.555
xproj + no p2p1.6790.0920.3940.5190.3260.5430.548
Ours (xproj)Single stageτ=00.9020.1270.1170.0000.0000.0000.594
τ=1000.5230.0860.1570.1760.0490.1390.599
Two stageτ=1000.9370.0980.1200.0760.0200.1090.574
τ=3000.9380.0980.1270.1180.0310.1280.573
τ=5000.9080.0980.1400.1570.0430.1400.577
τ=7000.8980.0980.1620.1960.0580.1530.580
τ=9000.8740.0980.1920.2380.0800.1800.581
no p2p0.8620.1040.2220.2870.1180.2820.577
+ +must meet both criteria to an acceptable degree. + +To achieve high-quality motion, both FID and Foot skating ratio are essential since FID alone cannot adequately measure the trajectory-motion coherence. Our Emphasis projection technique significantly improves motion coherence, reducing foot skating as shown in Tab. 3 while MDM [51] is unsuitable for this task due to the high motion incoherence. Furthermore, our improved architecture significantly improves motion quality in all cases. Note that without dense signal propagation, the model ignores the keyframe conditioning as shown in Fig 6. + +While a single-stage model performs reasonably well due to emphasis projection, it is too restrictive at $\tau = 0$ (forced trajectory), resulting in relatively high Foot skating. This issue can be addressed by allowing more modification (increasing to $\tau$ to 100) but at the cost of higher Loc. error. + +Lastly, the trajectory model's better controllability reduces Location error by more than half compared to the single-stage model at $\tau = 100$ . As expected, increasing $\tau$ leads to more freedom in the model, resulting in increased Trajectory diversity, lower FID, and higher Location error. + +# 6.5. Obstacle avoidance motion generation + +Finally, we demonstrate our model's ability to generate motion on additional guidance on the obstacle avoidance task. In this task, we randomly sample the target point that the human needs to reach at a specific motion step along with a set of obstacles it needs to avoid, represented as a 2D SDF (Sec. 5.3). We show the qualitative results in Fig 7. + +# 7. Discussion and Limitations + +In this work, we propose GMD, a controllable human motion generation method based on goal functions. GMD produces high-quality and diverse motions and supports diverse possibilities for goal functions. Since obtaining necessary data and designing a classifier-free learning method for non-explicit goals, such as obstacle avoidance, + +![](images/0f7203132172e6f7bf31a65a030ba7fc2f52329d35f5fbdaa3f5e35c1bb705b2.jpg) +Figure 7. Qualitative results from the obstacle avoidance task given keyframe locations and obstacles. The red crossed areas represent obstacles to avoid. More results are in the supplementary. + +can be challenging, our GMD utilizes a classifier-based method which allows for more conditioning flexibility without retraining the model. Thus, our studies on effective classifier guidance will be useful for further including more guiding signals. + +Acknowledgement. This work was supported by the SNSF project grant 200021 204840. + +# References + +[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 7 +[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. arXiv preprint arXiv:2211.09707, 2022. 3 +[3] Jose A Arjona-Medina, Michael Gillhofer, Michael Widrich, Thomas Unterthiner, Johannes Brandstetter, and Sepp Hochreiter. RUDDER: Return decomposition for delayed rewards. In Advances in Neural Information Processing Systems, 2019. 2 +[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu + +Liu. eDiff-I: Text-to-Image diffusion models with an ensemble of expert denoisers. Nov. 2022. 2 +[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 7 +[6] T Brooks, A Holynski, and A A Efros. InstructPix2Pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2 +[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. arXiv preprint arXiv:2212.04048, 2022. 1, 3, 4, 7, 8 +[8] Xin Chen, Zhuo Su, Lingbo Yang, Pei Cheng, Lan Xu, Bin Fu, and Gang Yu. Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134, 2022. 3 +[9] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. Aug. 2021. 2, 3 +[10] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. June 2022. 3 +[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. arXiv preprint arXiv:2212.04495, 2022. 1, 3 +[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. May 2021. 2, 3, 5, 7 +[13] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 3 +[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 4, 7, 9 +[15] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 580-597. Springer, 2022. 3 +[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3, 7 +[17] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-between. ACM Transactions on Graphics (TOG), 39(4):60-1, 2020. 3 +[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. Aug. 2022. 2 + +[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. Oct. 2022. 2 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. June 2020. 2, 3, 4, 7 +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications. 3, 7 +[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Apr. 2022. 2 +[23] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. May 2022. 2 +[24] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927. IEEE, 2020. 3 +[25] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. 3 +[26] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. Sept. 2020. 2 +[27] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. Feb. 2023. 3 +[28] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 3 +[29] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 3 +[30] Haoying Li, Yifan Yang, Meng Chang, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. SRDiff: Single image SuperResolution with diffusion probabilistic models. Apr. 2021. 2 +[31] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3 +[32] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 3 +[33] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 7 +[34] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Image synthesis and + +editing with stochastic differential equations. Aug. 2021. 2, 3 +[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3 +[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3 +[37] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 480-497. Springer, 2022. 3 +[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. Sept. 2022. 2 +[39] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-TTS: A diffusion probabilistic model for Text-to-Speech. May 2021. 2 +[40] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7 +[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. ArXiv, 2022. 2, 3 +[43] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 3 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution image synthesis with latent diffusion models. Dec. 2021. 2, 3 +[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image diffusion models with deep language understanding. May 2022. 2, 3 +[46] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image SuperResolution via iterative refinement. Apr. 2021. 2 +[47] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised + +learning using nonequilibrium thermodynamics. Mar. 2015. 2 +[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. July 2019. 2 +[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based generative modeling through stochastic differential equations. Nov. 2020. 2 +[50] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, Aug. 1988. 2 +[51] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. Sept. 2022. 1, 3, 4, 7, 8, 9 +[52] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. Edge: Editable dance generation from music. arXiv preprint arXiv:2211.10658, 2022. 1, 3 +[53] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 3 +[54] Christopher John Cornish Hellaby Watkins. Learning from delayed rewards. PhD thesis, King's College, Cambridge, 1989. 2 +[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. Oct. 2022. 2 +[56] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 257-274. Springer, 2022. 3 +[57] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019. 3 +[58] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500, 2022. 1, 3, 7, 8 +[59] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 7, 8 +[60] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3 +[61] Yan Zhang, Michael J Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020. 3 +[62] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion gener + +ation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 1, 3 +[63] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6225-6234, 2020. 3 +[64] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. arXiv preprint arXiv:2211.16016, 2022. 1, 3 \ No newline at end of file diff --git a/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/images.zip b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3ec460aaa5de4d34228218afa632bc80e442f05 --- /dev/null +++ b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:434ce3d2ff5c308a4ce5dcf5680a8b4b1d05f4d86d9872ab9bdd4f798a795b75 +size 412100 diff --git a/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/layout.json b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e0ea12ce2eee3ec21fe318fd2fa6a3aba081b5ef --- /dev/null +++ b/2023/Guided Motion Diffusion for Controllable Human Motion Synthesis/layout.json @@ -0,0 +1,13213 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 507, + 121 + ], + "type": "text", + "content": "Guided Motion Diffusion for Controllable Human Motion Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 200, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 200, + 157 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 200, + 157 + ], + "type": "text", + "content": "Korrawe Karunratanakul1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 213, + 143, + 312, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 143, + 312, + 158 + ], + "spans": [ + { + "bbox": [ + 213, + 143, + 312, + 158 + ], + "type": "text", + "content": "Konpat Preechakul2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 143, + 446, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 143, + 446, + 158 + ], + "spans": [ + { + "bbox": [ + 323, + 143, + 446, + 158 + ], + "type": "text", + "content": "Supasorn Suwajanakorn2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 459, + 143, + 514, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 459, + 143, + 514, + 158 + ], + "spans": [ + { + "bbox": [ + 459, + 143, + 514, + 158 + ], + "type": "text", + "content": "Siyu Tang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 178, + 165, + 416, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 165, + 416, + 179 + ], + "spans": [ + { + "bbox": [ + 178, + 165, + 416, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 165, + 416, + 179 + ], + "type": "text", + "content": "ETH Zürich, Switzerland " + }, + { + "bbox": [ + 178, + 165, + 416, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 165, + 416, + 179 + ], + "type": "text", + "content": "VISTEC, Thailand" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 190, + 182, + 399, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 182, + 399, + 194 + ], + "spans": [ + { + "bbox": [ + 190, + 182, + 399, + 194 + ], + "type": "text", + "content": "https://korrawe.github.io/gmd-project/" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 227, + 175, + 350 + ], + "blocks": [ + { + "bbox": [ + 94, + 216, + 135, + 225 + ], + "lines": [ + { + "bbox": [ + 94, + 216, + 135, + 225 + ], + "spans": [ + { + "bbox": [ + 94, + 216, + 135, + 225 + ], + "type": "text", + "content": "(a) Text only" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 227, + 175, + 350 + ], + "lines": [ + { + "bbox": [ + 53, + 227, + 175, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 175, + 350 + ], + "type": "image", + "image_path": "478f1f5716b592c29ae9039e4fdef8b13da30f3560adf01a83d493c6779e945e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 176, + 227, + 298, + 349 + ], + "blocks": [ + { + "bbox": [ + 204, + 216, + 269, + 225 + ], + "lines": [ + { + "bbox": [ + 204, + 216, + 269, + 225 + ], + "spans": [ + { + "bbox": [ + 204, + 216, + 269, + 225 + ], + "type": "text", + "content": "(b) Text + Trajectory" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 227, + 298, + 349 + ], + "lines": [ + { + "bbox": [ + 176, + 227, + 298, + 349 + ], + "spans": [ + { + "bbox": [ + 176, + 227, + 298, + 349 + ], + "type": "image", + "image_path": "7a55b1394bc0dd7e07665aac903cea39797258e11936deb30b1fea9e7da953fc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 299, + 227, + 418, + 349 + ], + "blocks": [ + { + "bbox": [ + 310, + 216, + 405, + 225 + ], + "lines": [ + { + "bbox": [ + 310, + 216, + 405, + 225 + ], + "spans": [ + { + "bbox": [ + 310, + 216, + 405, + 225 + ], + "type": "text", + "content": "(c) Text + Keyframe locations" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 227, + 418, + 349 + ], + "lines": [ + { + "bbox": [ + 299, + 227, + 418, + 349 + ], + "spans": [ + { + "bbox": [ + 299, + 227, + 418, + 349 + ], + "type": "image", + "image_path": "af433ba802f473570529c40f23fb907d0a7594ffe095492dead1705ae8b78ef9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 419, + 227, + 541, + 350 + ], + "blocks": [ + { + "bbox": [ + 432, + 209, + 527, + 224 + ], + "lines": [ + { + "bbox": [ + 432, + 209, + 527, + 224 + ], + "spans": [ + { + "bbox": [ + 432, + 209, + 527, + 224 + ], + "type": "text", + "content": "(d) Text + Keyframe locations +Obstacle avoidance" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 227, + 541, + 350 + ], + "lines": [ + { + "bbox": [ + 419, + 227, + 541, + 350 + ], + "spans": [ + { + "bbox": [ + 419, + 227, + 541, + 350 + ], + "type": "image", + "image_path": "b848983346230cd514c9ca12b5b1a33e4396fd4c66322b8f89cbee9bbfbbc421.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 547, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 547, + 401 + ], + "type": "text", + "content": "Figure 1. Our proposed Guided Motion Diffusion (GMD) can generate high-quality and diverse motions given a text prompt and a goal function. We demonstrate the controllability of GMD on four different tasks, guided by the following conditions: (a) text only, (b) text and trajectory, (c) text and keyframe locations (double circles), and (d) with obstacle avoidance (red-cross areas represent obstacles). The darker the colors, the later in time." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 420, + 192, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 420, + 192, + 432 + ], + "spans": [ + { + "bbox": [ + 143, + 420, + 192, + 432 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 445, + 289, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 289, + 721 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 289, + 721 + ], + "type": "text", + "content": "Denoising diffusion models have shown great promise in human motion synthesis conditioned on natural language descriptions. However, integrating spatial constraints, such as pre-defined motion trajectories and obstacles, remains a challenge despite being essential for bridging the gap between isolated human motion and its surrounding environment. To address this issue, we propose Guided Motion Diffusion (GMD), a method that incorporates spatial constraints into the motion generation process. Specifically, we propose an effective feature projection scheme that manipulates motion representation to enhance the coherency between spatial information and local poses. Together with a new imputation formulation, the generated motion can reliably conform to spatial constraints such as global motion trajectories. Furthermore, given sparse spatial constraints (e.g. sparse keyframes), we introduce a new dense guidance approach to turn a sparse signal, which is susceptible to being ignored during the reverse steps, into denser signals to guide the generated motion to the given constraints. Our extensive experiments justify the development of GMD, which achieves a significant improvement over state-of-the-art methods in text-based motion generation while allowing control of the synthesized motions with spatial constraints." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 420, + 386, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 386, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 386, + 432 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 445, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 673 + ], + "type": "text", + "content": "Recently, denoising diffusion models have emerged as a promising approach for human motion generation [11, 62, 64] outperforming other alternatives such as GAN or VAE in terms of both quality and diversity [7, 51, 58]. Several studies have focused on generating motion based on expressive text prompts [7, 51], or music [52, 64]. The state-of-the-art motion generation methods, such as MDM [51], utilize classifier-free guidance to generate motion conditioned on text prompts. However, incorporating spatial constraints into diffusion models remains underexplored. Human motions consist of both semantic and spatial information, where the semantic aspect can be described using natural languages or action labels and the spatial aspect governs physical interaction with surroundings. To generate realistic human motion in a 3D environment, both aspects must be incorporated. Our experiments show that simply adding spatial constraint guidance, such as global trajectories, into the state-of-the-art models or using imputation and in-painting approaches do not yield satisfactory results." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "We identify two main issues that make the motion diffusion models likely to ignore the guidance when conditioned on spatial objectives: the sparseness of global orientation" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "type": "text", + "content": "2151" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 285, + 204 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 285, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 285, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 285, + 204 + ], + "type": "image", + "image_path": "7447fc7616774b2d60110837a33d8a1f87f13a593266904511ff55fc8a4004c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 209, + 284, + 294 + ], + "blocks": [ + { + "bbox": [ + 48, + 209, + 284, + 294 + ], + "lines": [ + { + "bbox": [ + 48, + 209, + 284, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 284, + 294 + ], + "type": "image", + "image_path": "40e7fbf622dd39aa2616e07e88d250edcf1a17718619671b817c7348d117b601.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 299, + 287, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 355 + ], + "type": "text", + "content": "Figure 2. We tackle the problem of spatially conditioned motion generation with GMD, depicted in a). Our main contributions are b) Emphasis projection, for better trajectory-motion coherence, and c) Dense signal propagation, for a more controllable generation even under sparse guidance signal." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": "in the motion representation and sparse frame-wise guidance signals. By design, the diffusion models are a denoising model that consecutively denoises the target output over multiple steps. With sparse guidance, a small portion of the output that receives guidance will be inconsistent with all other parts that do not, therefore, are more likely to be treated as noise and discarded in subsequent steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": "First, the sparseness within a frame is a result of common motion representations that separate local pose information, like joint rotations, from global orientations, such as pelvis translations and rotations [43], usually with more focus on local poses. For instance, the common motion representation [14] uses 4 values to represent global orientation and 259 values for local pose in each frame. Such imbalance can cause the model to focus excessively on local pose information, and consequently, perceive guided global orientation as noise, resulting in a discrepancy such as foot skating." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Second, in many applications such as character animation, gaming, and virtual reality, the spatial control signals are defined on only a few keyframes such as target locations on the ground. We show that the current diffusion-based motion generation models struggle to follow such sparse guidance as doing so is equivalent to guiding an image diffusion model with only a few pixels. As a result, either the guidance at the provided keyframes will be ignored during the denoising process or the output motion will contain an artifact where the character warps to satisfy the guidance only in those specific keyframes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 347 + ], + "type": "text", + "content": "To effectively incorporate sparse spatial constraints into the motion generation process, we propose GMD, a novel and principled Guided Motion Diffusion model. To alleviate the discrepancy between local pose and global orientation in the guided denoising steps, we introduce emphasis projection, a general representation manipulation method that we use to increase the importance of spatial information during training. Additionally, we derive a new imputation and inpainting formulation that enables the existing inpainting techniques to operate in the projected space, which we leverage to generate significantly more coherent motion under guidance by spatial conditions. Then, to address the highly sparse guidance, we draw inspiration from the credit assignment problem in Reinforcement Learning [50, 54], where sparse rewards can be distributed along a trajectory to allow for efficient learning [3]. Our key insight is that motion denoisers, including the diffusion model itself, can be used to expand the spatial guidance signal at a specific location to its neighboring locations without any additional model. By turning a sparse signal into a dense one by backpropagating through a denoiser, it enables us to achieve high-quality controllable motion synthesis, even with extremely sparse guidance signals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 348, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 539 + ], + "type": "text", + "content": "In summary, our contributions are: (1) Emphasis projection, a method to adjust relative importance between different parts of the representation vector, which we use to encourage coherency between spatial information and local poses to allow spatial guidance. (2) Dense signal propagation, a conditioning method to tackle the sparse guidance problem. (3) GMD, an effective spatially controllable motion generation method that enables the unexplored synthesizing of motions based on free-text and spatial conditioning by integrating the above contributions into our proposed Unet-based architecture. We provide extensive analysis to support our design decisions and show the versatility of GMD on three tasks: trajectory conditioning, keyframe conditioning, and obstacle avoidance. Additionally, GMD's model also significantly outperforms the state-of-the-art in traditional text-to-motion tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 554, + 392, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 392, + 566 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 392, + 566 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "content": "Diffusion-based probabilistic generative models (DPM). DPMs [20,47,48,49] have gained significant attention in recent years due to their impressive performance across multiple fields of research. They have been used for tasks such as image generation [12], image super-resolution [30, 46], speech synthesis [26, 26, 39], video generation [19, 22], 3D shape generation [38, 55], and reinforcement learning [23]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The surge in interest in DPMs may be attributed to their impressive controllable generation capabilities, including text-conditioned generation [42, 44, 45] and image editing [4, 6, 9, 18, 34]. Latent diffusion models (LDM) are an-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2152" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 96 + ], + "type": "text", + "content": "other area of interest, which includes representation learning [27,40] and more efficient modeling techniques [7,44]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 286, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 286, + 181 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 286, + 181 + ], + "type": "text", + "content": "Moreover, DPMs exhibit a high degree of versatility in terms of conditioning. There are various methods for conditional generation, such as imputation and inpainting [9, 10, 34], classifier guidance [10, 12], and classifier-free guidance [21, 42, 44, 45]. Inpainting and classifier guidance can be applied to any pretrained DPM, which extends the model's capabilities further without the need for retraining." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 186, + 286, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 286, + 402 + ], + "type": "text", + "content": "Human motion generation. The goal of the human motion generation task is to generate motions based on the conditioning signals. Various conditions have been explored such as partial poses [13, 17, 51], trajectories [24, 53, 60], images [8,43], music [28,29,31], text [1, 14, 15, 25, 37], objects [56], action labels [16,36], or unconditioned [35,57,61,63]. Recently, many diffusion-based motion generation models have been proposed [11, 25, 32, 62, 64] and demonstrate better quality compared to alternative models such as GAN or VAE. Employing the CLIP model [41], these models showed great improvements in the challenging text-to-motion generation task [7,58,59] as well as allowing conditioning on partial motions [51] or music [2, 52]. However, they do not support conditioning signals that are not specifically trained, for example, following keyframe locations or avoiding obstacles. Maintaining the capabilities of the diffusion models, we propose methods to enable spatial guidance without retraining the model for each new objective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 416, + 126, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 126, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 126, + 430 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 436, + 230, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 230, + 449 + ], + "type": "text", + "content": "3.1. Diffusion-based generative models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": "Diffusion-based probabilistic generative models (DPMs) are a family of generative models that learn a sequential denoising process of an input " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " with varying noise levels " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": ". The noising process of DPM is defined cumulatively as " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t|\\mathbf{x}_0) = \\mathcal{N}(\\sqrt{\\alpha_t}\\mathbf{x}_0,(1 - \\alpha_t)\\mathbf{I})" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " is the clean input, " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\prod_{s = 1}^{t}(1 - \\beta_{s})" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " is a noise scheduler. The denoising model " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " learns to reverse the noising process by modeling the Gaussian posterior distribution " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": ". DPMs can map a prior distribution " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " to any distribution " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 456, + 286, + 589 + ], + "type": "text", + "content": " successive denoising steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "text", + "content": "To draw samples from a DPM, we start from a sample " + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "text", + "content": " from the prior distribution " + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "text", + "content": ". Then, for each " + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}\\sim \\mathcal{N}(\\mu_t,\\Sigma_t)" + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 590, + 286, + 626 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 637, + 286, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 637, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 76, + 637, + 286, + 663 + ], + "type": "interline_equation", + "content": "\\mu_ {t} = \\frac {\\sqrt {\\alpha_ {t - 1}} \\beta_ {t}}{1 - \\alpha_ {t}} \\mathbf {x} _ {0} + \\frac {\\sqrt {1 - \\beta_ {t}} (1 - \\alpha_ {t - 1})}{1 - \\alpha_ {t}} \\mathbf {x} _ {t} \\tag {1}", + "image_path": "20fdf7ddda46f13e92eec7c8be00c04ce2e46b5f1a938a26a7f9c8bd498ef46b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "inline_equation", + "content": "\\Sigma_{t}" + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "text", + "content": " is a variance scheduler of choice, usually " + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "inline_equation", + "content": "\\Sigma_{t} = \\frac{1 - \\alpha_{t-1}}{1 - \\alpha_{t}}\\beta_{t}" + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "text", + "content": " [20]. " + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "text", + "content": " in Eq. 1 is the prediction from a denoising model. For an " + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "text", + "content": " model, " + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 = \\frac{1}{\\sqrt{\\alpha_t}}\\mathbf{x}_t + \\frac{\\sqrt{1 - \\alpha_t}}{\\sqrt{\\alpha_t}}\\epsilon_\\theta (\\mathbf{x}_t)" + }, + { + "bbox": [ + 47, + 673, + 286, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": "There are multiple choices for the denoising model to predict including the clean input " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": ", the noise " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": ", and the one-step denoised target " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": ". An " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,\\theta}" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": " model is trained using the squared loss to the clean input " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{x}_{0,\\theta}(\\mathbf{x}_t) - \\mathbf{x}_0\\|^2" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": ", an " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": " model is trained using the squared loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\| \\epsilon_{\\theta}(\\mathbf{x}_t) - \\epsilon \\| ^2" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mu_{t,\\theta}" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": " model is trained using the squared loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\| \\mu_{t,\\theta}(\\mathbf{x}_t) - \\mu_t\\|^2" + }, + { + "bbox": [ + 305, + 72, + 545, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 167, + 545, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 545, + 180 + ], + "type": "text", + "content": "3.2. Controllable generation with diffusion models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "content": "Classifier-free guidance. The conditioning signals are treated as additional inputs to the denoiser " + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,d)" + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "content": " is the conditioning signals which can be omitted " + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "inline_equation", + "content": "d = \\varnothing" + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "content": " to generate unconditionally. Classifier-free guidance has been shown to generate very high-quality results [44, 45, 51]. To draw samples, the effective denoiser becomes " + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_t,d) = w\\epsilon_{\\theta}(\\mathbf{x}_t,d) + (1 - w)\\epsilon_{\\theta}(\\mathbf{x}_t,\\varnothing)" + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "content": " controls the conditional strength. The new " + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 305, + 191, + 545, + 335 + ], + "type": "text", + "content": " model can be used in Eq. 1. Two downsides of this method are that the nature of conditioning signals need to be known before hand and the denoiser needs to be adjusted and retrained for each specific case restricting its flexibility." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "text", + "content": "Classifier guidance. We can also obtain " + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{t-1}|\\mathbf{x}_t,d)" + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t-1}|\\mathbf{x}_t)p(d|\\mathbf{x}_t)" + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "text", + "content": " [12], where " + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "p(d|\\mathbf{x}_t)" + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "text", + "content": " is any probability function that we can approximate its score function " + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(d|\\mathbf{x}_t)" + }, + { + "bbox": [ + 305, + 340, + 545, + 411 + ], + "type": "text", + "content": " effectively. The new sampling process is similar to the original (Eq. 1) but with the mean shifted by the scaled score function as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 422, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 422, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 361, + 422, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\mu_ {t} = \\mu_ {t} ^ {\\prime} + s \\Sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (d | \\mathbf {x} _ {t}) \\tag {2}", + "image_path": "a299c404cf117b6a78bedec19eece99688ceb0f1c639d60a67191dfd04b07c32.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\mu_t^\\prime" + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": " is the original mean, " + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": " controls the conditioning strength, and " + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\Sigma_{t}" + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": " is a variance scheduler which can be the same as in Eq. 1. Since " + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\Sigma_{t}" + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": " is a decreasing sequence, the guidance signal diminishes as " + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "t\\to 0" + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": " which corresponds to the characteristic of DPMs that tend to modify " + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": " less and less as time goes. Classifier guidance is a post-hoc method, i.e., there is no change to the DPM model, one only needs to come up with " + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "inline_equation", + "content": "p(d|\\mathbf{x}_t)" + }, + { + "bbox": [ + 305, + 445, + 545, + 542 + ], + "type": "text", + "content": " which is extremely flexible." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 547, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 547, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 305, + 547, + 545, + 618 + ], + "type": "text", + "content": "Imputation and inpainting. To generate human motion sequences from partial observations, such as global motion trajectories or keyframe locations, inpainting is used. These partial observations, called imputing signals, are used to adjust the generative process towards the observations. Imputation and inpainting are two sides of the same coin." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " be a partial target value in an input " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " that we want to impute. The imputation region of " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "M_y^x" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": ", and a projection " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "P_y^x" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " that resizes " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " to that of " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " by filling in zeros. In DPMs, imputation can be done on the sample " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " after every denoising step [10]. We have the new imputed sample " + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{t - 1}" + }, + { + "bbox": [ + 305, + 619, + 545, + 691 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 334, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {t - 1} = \\left(1 - M _ {y} ^ {x}\\right) \\odot \\mathbf {x} _ {t - 1} + M _ {y} ^ {x} \\odot P _ {y} ^ {x} \\mathbf {y} _ {t - 1} \\tag {3}", + "image_path": "c2bdd63f18b44afe0ba614066883d2cf4e5fdf2baf0fd40b71f28aa84ca75a71.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2153" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " is a Hadamard product and " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{t - 1}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " is a noised target value. " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{t - 1}\\sim \\mathcal{N}(\\sqrt{\\alpha_{t - 1}}\\mathbf{y},(1 - \\alpha_{t - 1})\\mathbf{I})" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " following Ho et al. [20] is one of the simplest choices of " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{t - 1}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 145 + ], + "type": "text", + "content": "Note that all three modes of conditioning presented here are not mutually exclusive. One could apply one or more tricks in a single pipeline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 157, + 190, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 157, + 190, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 190, + 170 + ], + "type": "text", + "content": "4. Guided Motion Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 49, + 202, + 287, + 550 + ], + "blocks": [ + { + "bbox": [ + 49, + 189, + 275, + 201 + ], + "lines": [ + { + "bbox": [ + 49, + 189, + 275, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 275, + 201 + ], + "type": "text", + "content": "Algorithm 1 GMD's two-stage guided motion diffusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "lines": [ + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": "Require: A trajectory DPM " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0,\\phi}" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": ", a motion DPM " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,\\theta}" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": ", a goal function " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "G_{\\mathbf{z}}(\\cdot)" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": ", and keyframe locations " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " (if any). \n1: # Stage 1: Trajectory generation \n2: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n3: for all " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " to 1 do \n4: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 \\gets \\mathbf{z}_{0,\\phi}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mu, \\Sigma \\gets \\mu(\\mathbf{z}_0, \\mathbf{z}_t), \\Sigma_t" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n6: # Classifier guidance (Eq. 2) \n7: # Dense signal propagation \n8: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t-1} \\sim \\mathcal{N}(\\mu - s\\Sigma\\nabla_{\\mathbf{z}_t}G_z(\\mathbf{z}_0), \\Sigma)" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n9: # Impute y on z (Eq. 3) (if any) \n10: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t-1} \\gets (1 - M_y^z) \\odot \\mathbf{z}_{t-1} + M_y^z\\mathbf{y}_{t-1}" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n11: end for \n12: # Stage 2: Trajectory-conditioned motion generation \n13: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{proj}} \\gets" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " sample from " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n14: for all " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " to 1 do \n15: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "M \\gets P_z^x M_y^z" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " # Imputation region of y on x \n16: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{\\mathrm{proj}} \\gets \\mathbf{x}_{0,\\theta}^{\\mathrm{proj}}(\\mathbf{x}_t^{\\mathrm{proj}})" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " # Emphasis projection \n17: # Impute y on x" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "^{\\mathrm{proj}}" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " (Eq. 6) \n18: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}} \\gets A((1 - M) \\odot A^{-1}\\mathbf{x}_0^{\\mathrm{proj}} + MP_z^x\\mathbf{y})" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n19: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mu, \\Sigma \\gets \\mu(\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}},\\mathbf{x}_t^{\\mathrm{proj}}), \\Sigma_t" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n20: # Masked classifier guidance (Eq. 9) \n21: # Dense signal propagation \n22: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\Delta \\gets -s\\Sigma A^{-1}\\nabla_{\\mathbf{x}_t^{\\mathrm{proj}}}G_z(P_x^zA^{-1}\\mathbf{x}_0^{\\mathrm{proj}})" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n23: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mu \\gets \\mu + A(1 - M) \\odot \\Delta" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n24: " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1} \\sim \\mathcal{N}(\\mu, \\Sigma)" + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "text", + "content": " \n25: end for \n26: return " + }, + { + "bbox": [ + 49, + 202, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "content": "We aim to generate realistic human motions that can be guided by spatial constraints, enabling the generated human motion to achieve specific goals, such as following a global trajectory, reaching certain locations, or avoiding obstacles. Although diffusion-based models have significantly improved text-to-motion modeling [7, 51], generating motions that achieve specific goals is still beyond the reach of the current models. Our work addresses this limitation and advances the state-of-the-art in human motion modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": "We are interested in modeling a full-body human motion that satisfies a certain scalar goal function " + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "G_{x}(\\cdot)" + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": " that takes a motion representation " + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": " and measures how far the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "motion " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is from the goal (lower is better). More specifically, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{N \\times M}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " represents a sequence of human poses for " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " motion steps, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the dimension of human pose representations, e.g., " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "N = 263" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " in the HumanML3D [14] dataset. Let " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " be the random variable associated with " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Our goal is to model the following conditional probability using a motion DPM" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 388, + 164, + 545, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 164, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 388, + 164, + 545, + 178 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} \\mid G _ {x} (X) = 0\\right) \\tag {4}", + "image_path": "0e43a2d6f3fe0c8b36fd918a1983bcc1c31d8bc31926bd8dce15855270737521.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "text", + "content": "This can be extended to " + }, + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "inline_equation", + "content": "p\\big(\\mathbf{x}|G_x(X) = 0,d\\big)" + }, + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "text", + "content": " is any additional signal, such as text prompts. From now on, we omit " + }, + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 187, + 545, + 222 + ], + "type": "text", + "content": " to reduce clutter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": "Many challenging tasks in motion modeling can be encapsulated within a goal function " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "G_{z}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " that only depends on the trajectory " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " of the human motion, not the whole motion " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": ". Let us define " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{L \\times M}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " to be the trajectory part of " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "L = 2" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " describing the ground location of the pelvis of a human body. A particular location " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{(i)}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " at motion step " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " describes the pelvis location of the human body on the ground plane. We define a projection " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "P_x^z" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " that resizes " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " to match " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " by taking only the " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " part, and its reverse " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "P_z^x" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " that resizes " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " to match " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": " by filling in zeros. With this, our conditional probability becomes " + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "inline_equation", + "content": "p\\big(\\mathbf{x}|G_z(P_x^z X) = 0\\big)" + }, + { + "bbox": [ + 304, + 223, + 545, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 355, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 450 + ], + "type": "text", + "content": "In this work, we will show how text-to-motion DPMs can be extended to solve several challenging tasks, including trajectory-conditioned motion generation, location-conditioned trajectory planning, and obstacle avoidance trajectory planning. Using our proposed Emphasis projection and dense signal propagation, we alleviate the sparse guidance problem and enable motion generation based on spatial conditions. The overview of our methods is shown in Fig. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 458, + 424, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 458, + 424, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 458, + 424, + 471 + ], + "type": "text", + "content": "4.1. Emphasis projection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "content": "One of the most straightforward approaches for minimizing the goal function " + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "inline_equation", + "content": "G_{\\mathbf{z}}(\\cdot)" + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "content": " is by analyzing what trajectories that minimize " + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{*} = \\arg \\min_{\\mathbf{z}}G_{z}(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "content": " look like. For a trajectory conditioning task, a whole trajectory " + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^*" + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "content": " is directly given. Our task is to generate the rest of the motion " + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "content": ". With such knowledge, we can employ imputation & inpainting technique by supplying the motion DPM with the " + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "content": "-shaped " + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "inline_equation", + "content": "P_z^x\\mathbf{z}^*" + }, + { + "bbox": [ + 304, + 476, + 545, + 572 + ], + "type": "text", + "content": " to guide the generation process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 577, + 443, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 443, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 443, + 587 + ], + "type": "text", + "content": "Problem 1: Motion incoherence" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "text", + "content": "Since the imputing trajectory " + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^*" + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "text", + "content": " is only a small part of the whole motion " + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "inline_equation", + "content": "L \\ll N" + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "text", + "content": "), we often observe that the DPM ignores the change from imputation and fails to make appropriate changes on the rest of " + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 588, + 545, + 661 + ], + "type": "text", + "content": ". This results in an incoherent local motion that is not aligned or well coordinated with the imputing trajectory." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 666, + 446, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 446, + 677 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 446, + 677 + ], + "type": "text", + "content": "Solution 1: Emphasis projection" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We tackle this problem by giving more emphasis on the trajectory part of motion " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". More specifically, we propose an Emphasis projection method that increases the trajectory's" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2154" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 73, + 506, + 221 + ], + "blocks": [ + { + "bbox": [ + 86, + 73, + 506, + 221 + ], + "lines": [ + { + "bbox": [ + 86, + 73, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 86, + 73, + 506, + 221 + ], + "type": "image", + "image_path": "2aada7e8aaab10721a928ae76978d7292c606cbb8db4108324dd3cf30e9c082c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 546, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 546, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 546, + 269 + ], + "type": "text", + "content": "Figure 3. (a) Under standard motion representation and guiding method, only a few values in the motion representation are updated according to the guidance. (b) With Emphasis projection, all values in each frame describing the motion receive gradients w.r.t. the guidance, leading to better coherence between global orientation and local pose in each frame. (c) With dense gradient propagation, all frames are updated according to the guidance at the keyframes, making the guidance less likely to be ignored." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": "relative importance within motion " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": ". We achieve this by utilizing a random matrix " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "A = A'B" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "A' \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": " is a matrix with elements randomly sampled from " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": " is a diagonal matrix whose trajectory-related diagonal indexes are " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": " and the rest are 1 for emphasizing those trajectory elements. In our case, we emphasize the rotation and ground location of the pelvis, " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "(\\mathrm{rot},x,z)" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": ", in " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": " times. We now have a projected motion " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{proj}} = \\frac{1}{N - 3 + 3c^2} A\\mathbf{x}" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": ". Note that the fractional term is to maintain the unit variance on " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": ". The noisng process of the projected motion becomes " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t^{\\mathrm{proj}}|\\mathbf{x}_0^{\\mathrm{proj}}) = \\mathcal{N}(\\sqrt{\\alpha_t}\\mathbf{x}_0^{\\mathrm{proj}},(1 - \\alpha_t)\\mathbf{I})" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": ". There is no change on how a DPM that works on the projected motion " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}^{\\mathrm{proj}}|\\mathbf{x}_t^{\\mathrm{proj}})" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": " operates and treats " + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 289, + 289, + 447 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 448, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 507 + ], + "type": "text", + "content": "In Section 6.3, we show that emphasis projection is an effective way of solving the motion incoherence problem, and is shown to be substantially better than a straightforward approach of retraining a DPM with an increased loss weight on the trajectory." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "text", + "content": "Imputation on the projected motion " + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "text", + "content": ". We have discussed imputing on the sample " + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1}" + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "text", + "content": " in Eq. 3. Here, we introduce an imputation on " + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "text", + "content": " which modifies the DPM's belief on the final outcome " + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,\\theta}" + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "text", + "content": " by imputing it with " + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 513, + 287, + 583 + ], + "type": "text", + "content": ". We have found this technique useful in many tasks we are interested in." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": "Let us define the imputation region of " + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "inline_equation", + "content": "M_z^x" + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": ". We obtain the imputed " + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 47, + 586, + 287, + 611 + ], + "type": "text", + "content": " from" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 622, + 287, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 622, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 85, + 622, + 287, + 650 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {0} = \\left(1 - M _ {z} ^ {x}\\right) \\odot \\mathbf {x} _ {0, \\theta} + M _ {z} ^ {x} \\odot \\underbrace {P _ {z} ^ {x} \\mathbf {z} ^ {*}} _ {\\mathbf {x} \\text {s h a p e d}} \\tag {5}", + "image_path": "1ba34ed77f5ffa47b3e560eb51219d1306dfaf686aff70900b0ba6885eef9abb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": "Now operating on the projected motion " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{proj}}" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": ", before we can do imputation, we need to unproject it back to the original motion using " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 = A^{-1}\\mathbf{x}_0^{\\mathrm{proj}}" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": ", and then project the imputed " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": " back using " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}} = A\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": ". We obtain the imputed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 288, + 488, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 488, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 488, + 301 + ], + "type": "text", + "content": "motion under emphasis projection " + }, + { + "bbox": [ + 306, + 288, + 488, + 301 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}}" + }, + { + "bbox": [ + 306, + 288, + 488, + 301 + ], + "type": "text", + "content": " from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 307, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 307, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 312, + 307, + 545, + 328 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {0} ^ {\\text {p r o j}} = A \\left(\\left(1 - M _ {z} ^ {x}\\right) \\odot \\left(A ^ {- 1} \\mathbf {x} _ {0, \\theta} ^ {\\text {p r o j}}\\right) + M _ {z} ^ {x} \\odot P _ {z} ^ {x} \\mathbf {z} ^ {*}\\right) \\tag {6}", + "image_path": "b4e46487fbcebd00c00316ea1aa63d5db2ca861367ee6cc942f0b74eda46cc5e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "type": "text", + "content": "Substituting " + }, + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0^{\\mathrm{proj}}" + }, + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "type": "text", + "content": " into Eq. 1, we obtain the new mean " + }, + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t^{\\mathrm{proj}}" + }, + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "type": "text", + "content": " for sampling " + }, + { + "bbox": [ + 305, + 334, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}^{\\mathrm{proj}}\\sim \\mathcal{N}(\\tilde{\\mu}_t^{\\mathrm{proj}},\\Sigma_t)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 367, + 544, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 367, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 544, + 380 + ], + "type": "text", + "content": "4.2. Dense guidance signal with a learned denoiser" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": "Another way to minimize the goal function " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "G_{z}(\\cdot)" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": " is by adjusting the sample of each diffusion step " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1}" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": " toward a region with lower " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "G_{z}" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": ". This trick is called classifier guidance [12]. The direction of change corresponds to a score function " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p\\big(G_x(X_t) = 0|\\mathbf{x}_t\\big)" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": " which can be approximated as a direction " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathbf{x}_0} = -\\nabla_{\\mathbf{x}_0}G_z(P_x^z\\mathbf{x}_{0,\\theta})" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": " that reduces the goal function. We can guide the generative process by nudging the DPM's prediction as " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 = \\mathbf{x}_{0,\\theta} + \\Delta_{\\mathbf{x}_0}" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": ". While imputation requires the minimizer " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^*" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "inline_equation", + "content": "G_{z}" + }, + { + "bbox": [ + 304, + 385, + 545, + 517 + ], + "type": "text", + "content": ", which might not be easy to obtain or may not be unique, this trick only requires the easier-to-obtain direction of change." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 522, + 457, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 522, + 457, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 457, + 533 + ], + "type": "text", + "content": "Problem 2: Sparse guidance signal" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "In the motion domain, conditioning signals can often be sparse. There are two types of sparsity that can occur: sparsity in feature and sparsity in time. Sparsity in feature is when the conditioning signal is a small part of the feature dimension of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ". For example, in trajectory-conditioned generation, " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " may only consist of a sequence of ground locations over time. This type of sparsity can be addressed by emphasis projection, as explained in Section 4.1. Sparsity in time refers to cases where the conditioning signal consists of small segments of a trajectory spread out over time. For instance, in keyframe location conditioning task, only a sparse set of keyframe locations are given. When the conditioning signal-to-noise ratio becomes too small, the conditioning signal may be mistaken as noise and ignored during the denoising process." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2155" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 207, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 207, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 207, + 83 + ], + "type": "text", + "content": "Solution 2: Dense signal propagation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": "To turn a sparse signal into a dense signal, we need domain knowledge. One way to achieve this is by using a denoising function " + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_t) = \\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": ", which is trained on a motion dataset to denoise by gathering information from the nearby motion frames. With the ability to relate a single frame to many other frames, the denoising function is capable of expanding a sparse signal into a denser one." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "content": "We can use backward propagation through the denoising function " + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "content": " to take advantage of this. Therefore, a dense classifier guidance can be obtained as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 286, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 286, + 240 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {x} _ {t}} \\log p \\left(G _ {x} \\left(X _ {t}\\right) = 0 \\mid \\mathbf {x} _ {t}\\right) \\approx - \\nabla_ {\\mathbf {x} _ {t}} G _ {z} \\left(\\underbrace {P _ {x} ^ {z} f (\\mathbf {x} _ {t})} _ {\\mathbf {z} \\text {s h a p e d}}\\right) \\tag {7}", + "image_path": "fffc989c71f2721497eee3f40e1c05c0a7672c550447d38bdced5126d541328f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "text", + "content": "While an external function can be used as " + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "text", + "content": ", we observe that the existing DPM model " + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,\\theta}(\\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "text", + "content": " itself is a motion de-noiser, and thus can be used to turn a sparse signal into a dense signal without the need for an additional model. In practice, this process amounts to computing the gradient of " + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,\\theta}(\\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 247, + 286, + 319 + ], + "type": "text", + "content": " using autodiff." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 324, + 286, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 286, + 432 + ], + "type": "text", + "content": "Applying classifier guidance together with imputation. Whenever available, we want to utilize signals from both imputation and classifier guidance techniques to help guide the generative process. Imputation is explicit but may encounter sparsity in time, while classifier guidance is indirect but dense. We want to use the direct signal from imputation wherever available (with mask " + }, + { + "bbox": [ + 46, + 324, + 286, + 432 + ], + "type": "inline_equation", + "content": "M_z^x" + }, + { + "bbox": [ + 46, + 324, + 286, + 432 + ], + "type": "text", + "content": "), and the rest from classifier guidance (with mask " + }, + { + "bbox": [ + 46, + 324, + 286, + 432 + ], + "type": "inline_equation", + "content": "1 - M_z^x" + }, + { + "bbox": [ + 46, + 324, + 286, + 432 + ], + "type": "text", + "content": "). Based on Eq. 2, imputation-aware classifier guidance can be written as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 441, + 286, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 441, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 71, + 441, + 286, + 455 + ], + "type": "interline_equation", + "content": "\\mu_ {t} = \\tilde {\\mu} _ {t} - \\left(1 - M _ {z} ^ {x}\\right) \\odot s \\Sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} G _ {z} \\left(P _ {x} ^ {z} f (\\mathbf {x} _ {t})\\right) \\tag {8}", + "image_path": "ca53167ddbab9ec75ddac3d700905ab3597d2531e9325fc66e02cccab7c89c36.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}" + }, + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "text", + "content": " is an imputed sampling mean. By replacing " + }, + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}" + }, + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}^{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 463, + 286, + 498 + ], + "type": "text", + "content": ", we get classifier guidance together with imputation that works with emphasis projection as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 506, + 286, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 506, + 286, + 522 + ], + "spans": [ + { + "bbox": [ + 79, + 506, + 286, + 522 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\mu} = - s \\Sigma_ {t} A ^ {- 1} \\nabla_ {\\mathbf {x} _ {t} ^ {\\mathrm {p r o j}}} G _ {z} \\left(P _ {x} ^ {z} A ^ {- 1} f \\left(\\mathbf {x} _ {t} ^ {\\mathrm {p r o j}}\\right)\\right) \\tag {9}", + "image_path": "3113a61b2a777757a5e93afe907fcd07b3ce2b81cd827befa7a7f26b4c61cca8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 525, + 286, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 525, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 75, + 525, + 286, + 540 + ], + "type": "interline_equation", + "content": "\\mu_ {t} ^ {\\text {p r o j}} = \\tilde {\\mu} _ {t} ^ {\\text {p r o j}} + A \\left(1 - M _ {z} ^ {x}\\right) \\odot \\Delta_ {\\mu} \\tag {10}", + "image_path": "b1ccf8379a5459ed64f6b597989bc853ccb2401f375db67402c93b4d47c245b7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 553, + 268, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 268, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 268, + 565 + ], + "type": "text", + "content": "Problem 3: DPM's bias hinders the guidance signal" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 565, + 286, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 286, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 286, + 648 + ], + "type": "text", + "content": "A DPM removes noise from an input based on the distribution of the training data it has seen. This could be problematic when it comes to conditional generation because the conditioning signal may be outside of the training distribution. As a result, any changes made to the classifier guidance may be reversed by the DPM in the next time step, due to its inherent bias towards the data, shown in Figure 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 654, + 174, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 174, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 174, + 665 + ], + "type": "text", + "content": "Solution 3: Epsilon modeling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": "While it is unlikely to train an unbiased DPM model, there are ways to minimize the influence of model's bias under the guidance signal. Conceptually, the DPM model usually makes less and less change near the final outcome. This is" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": "in tandem with the guidance signal that gradually decreases over time due to " + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "inline_equation", + "content": "\\Sigma_t" + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": " (Eq. 2)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "content": "We investigate the coefficient " + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "inline_equation", + "content": "\\frac{\\sqrt{\\alpha_{t-1}\\beta_t}}{1-\\alpha_t}" + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "content": " in the sampling mean " + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "content": " (Eq. 1). This coefficient reaches its maximum value at " + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "content": ", meaning that an " + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,\\theta}" + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "content": " model could have a significant impact on the sampling mean even at " + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 96, + 544, + 158 + ], + "type": "text", + "content": ", which contradicts the weak guidance signal at that time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "text", + "content": "On the other hand, an " + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "text", + "content": " model will have the most influence on the sampling mean at " + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "text", + "content": ", which aligns with our intuition. In Section 6.4 and Figure 4, we demonstrate that modeling " + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0,\\theta}" + }, + { + "bbox": [ + 304, + 159, + 545, + 231 + ], + "type": "text", + "content": " is a successful approach for managing the bias effect of the DPM model in classifier guidance. We further discuss this point in Supplementary." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 243, + 385, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 385, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 385, + 258 + ], + "type": "text", + "content": "5. Applications" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 264, + 489, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 489, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 489, + 277 + ], + "type": "text", + "content": "5.1. Trajectory-conditioned generation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 284, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 331 + ], + "type": "text", + "content": "This task aims at generating a realistic motion " + }, + { + "bbox": [ + 304, + 284, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 284, + 545, + 331 + ], + "type": "text", + "content": " that matches a given trajectory " + }, + { + "bbox": [ + 304, + 284, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 284, + 545, + 331 + ], + "type": "text", + "content": ". Our objective is to minimize the distance between the generated motion and the given trajectory, which we define as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 341, + 544, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 341, + 544, + 372 + ], + "spans": [ + { + "bbox": [ + 369, + 341, + 544, + 372 + ], + "type": "interline_equation", + "content": "G _ {x} (\\mathbf {x}) := \\left\\| \\mathbf {z} - \\underbrace {P _ {x} ^ {z} \\mathbf {x}} _ {\\mathbf {z} \\text {p a r t o f} \\mathbf {x}} \\right\\| _ {p} \\tag {11}", + "image_path": "4cf392113c5836cadf768077efaddd2945631a6513def45837a8d2a755f79221.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 382, + 544, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 544, + 430 + ], + "type": "text", + "content": "Despite the apparent simplicity of this task, a traditional DPM faces the challenge of ensuring coherence in the generated motion. However, our emphasis projection method can effectively address this problem." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 441, + 486, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 486, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 486, + 453 + ], + "type": "text", + "content": "5.2. Keyframe-conditioned generation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "text", + "content": "The locations of ground positions at specific times can be used to define locations that we wish the generated motion to reach. This task is a generalized version of the trajectory-conditioned generation where only a partial and potentially sparse trajectory is given. Let " + }, + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{2 \\times M}" + }, + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "text", + "content": " be a trajectory describing keyframe locations and a mask " + }, + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "inline_equation", + "content": "M_y^z" + }, + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "text", + "content": " describe the key motion steps. Our goal function of a motion " + }, + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 460, + 544, + 543 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 354, + 553, + 544, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 553, + 544, + 575 + ], + "spans": [ + { + "bbox": [ + 354, + 553, + 544, + 575 + ], + "type": "interline_equation", + "content": "G _ {x} (\\mathbf {x}) := \\sum_ {i} \\left\\| M _ {y} ^ {z} \\left(P _ {x} ^ {z} \\mathbf {x} - \\mathbf {y}\\right) \\right\\| _ {p} \\tag {12}", + "image_path": "4aaf52327a334c5749b4ba2377db1aa0c1a93e0ac74ef9dcc29ba1b3642709e2.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "text", + "content": "Consequently, " + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "inline_equation", + "content": "G_{z}(\\mathbf{z}) = \\sum_{i}\\left\\| M_{y}^{z}(\\mathbf{z} - \\mathbf{y})\\right\\|_{p}" + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "text", + "content": ". Due to the partial trajectory " + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "text", + "content": ", the imputation region of " + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "text", + "content": " becomes " + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "inline_equation", + "content": "M_y^x = P_z^x M_y^z" + }, + { + "bbox": [ + 304, + 586, + 544, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Two-stage guided motion generation. Generating both the trajectory and motion simultaneously under a conditioning signal can be challenging and may result in lower quality motion. To address this issue, we propose a two-step approach. First, we generate a trajectory " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " that satisfies the keyframe locations and then generate the motion " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " given the trajectory (following Section 5.1). Our overall pipeline" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2156" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "is depicted in Figure 2 (a). We offer two options for generating the trajectory from keyframe locations " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ": a point-to-point trajectory and a trajectory DPM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "content": "The point-to-point trajectory connects consecutive keyframe locations with a straight line. These unrealistic trajectories can be used as imputation signals for the motion DPM during the early phase " + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "inline_equation", + "content": "(t \\geq \\tau)" + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "content": " is large enough, the DPM will adjust the given trajectory to a higher quality one. However, if " + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "content": " is too large, the DPM may generate a motion that does not perform well on " + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "inline_equation", + "content": "G_{z}" + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 192, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 286, + 239 + ], + "type": "text", + "content": "The trajectory DPM " + }, + { + "bbox": [ + 47, + 192, + 286, + 239 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{z}_{t - 1}|\\mathbf{z}_t)" + }, + { + "bbox": [ + 47, + 192, + 286, + 239 + ], + "type": "text", + "content": ", which is trained using the same dataset but with a smaller network, can be used to generate the trajectory under the guidance signal from " + }, + { + "bbox": [ + 47, + 192, + 286, + 239 + ], + "type": "inline_equation", + "content": "G_{z}" + }, + { + "bbox": [ + 47, + 192, + 286, + 239 + ], + "type": "text", + "content": ". We summarize our two-stage approach in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "text", + "content": "It is also possible to combine the two methods, as the point-to-point trajectory can serve as a useful guidance signal for the trajectory DPM during " + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "inline_equation", + "content": "t \\geq \\tau" + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "text", + "content": ". After that, the trajectory DPM is subject to the usual imputation and classifier guidance from " + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "inline_equation", + "content": "G_{z}" + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "text", + "content": ". By tuning " + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "text", + "content": ", we can balance between trajectory diversity and lower scores on " + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "inline_equation", + "content": "G_{z}" + }, + { + "bbox": [ + 46, + 240, + 286, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 317, + 248, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 248, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 248, + 330 + ], + "type": "text", + "content": "5.3. Obstacle avoidance motion generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 335, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 286, + 407 + ], + "type": "text", + "content": "Humans have the ability to navigate around obstacles while traveling from point A to B. Under our framework, this problem can be defined using two goal functions: one that navigates from A to B, called " + }, + { + "bbox": [ + 46, + 335, + 286, + 407 + ], + "type": "inline_equation", + "content": "G_{x}^{\\mathrm{loc}}" + }, + { + "bbox": [ + 46, + 335, + 286, + 407 + ], + "type": "text", + "content": " (defined as in Eq. 12), and another that pushes back when the human model crosses the obstacle's boundary, called " + }, + { + "bbox": [ + 46, + 335, + 286, + 407 + ], + "type": "inline_equation", + "content": "G_{x}^{\\mathrm{obs}}" + }, + { + "bbox": [ + 46, + 335, + 286, + 407 + ], + "type": "text", + "content": ", as follows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 411, + 286, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 411, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 64, + 411, + 286, + 435 + ], + "type": "interline_equation", + "content": "G _ {x} ^ {\\mathrm {o b s}} (\\mathbf {x}) := \\sum_ {i} - \\operatorname {c l i p m a x} \\left(\\operatorname {S D F} \\left(\\left(P _ {x} ^ {z} \\mathbf {x}\\right) ^ {(i)}\\right), c\\right) \\tag {13}", + "image_path": "3d8cf58e74e96dd5828caaabb0530c7914f5824b85a802d3812aab87ae069658.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 440, + 286, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 286, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 440, + 286, + 476 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 440, + 286, + 476 + ], + "type": "text", + "content": " is the safe distance from the obstacle. These two goal functions are combined additively to obtain the final goal function, " + }, + { + "bbox": [ + 46, + 440, + 286, + 476 + ], + "type": "inline_equation", + "content": "G_{x}(\\mathbf{x}) = G_{x}^{\\mathrm{loc}}(\\mathbf{x}) + G_{x}^{\\mathrm{obs}}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 440, + 286, + 476 + ], + "type": "text", + "content": ", for this task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 476, + 286, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 286, + 523 + ], + "type": "text", + "content": "We utilize the same pipeline as in Section 5.2, with the exception that imputation is not possible for obstacle avoidance. Therefore, minimizing the obstacle avoidance goal relies solely on classifier guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 533, + 127, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 127, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 127, + 546 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 552, + 286, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 286, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 286, + 660 + ], + "type": "text", + "content": "To evaluate our methods, we perform experiments on the standard human motion generation task conditioned on text descriptors and spatial objectives. In particular, we evaluate (1) the performance of our model in standard text-condition motion generation tasks, (2) the effect of emphasis projection to alleviate incoherence between spatial locations and local poses, (3) the ability to conditionally generate motion based on spatial information by conditioning with given trajectories, keyframe locations, and obstacles." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 666, + 107, + 678 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 107, + 678 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 107, + 678 + ], + "type": "text", + "content": "6.1. Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate generative text-to-motion models using standard metrics introduced by Guo et al." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 95, + 542, + 232 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "type": "text", + "content": "Table 1. Text-to-motion evaluation on the HumanML3D [14] dataset. The right arrow " + }, + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "type": "text", + "content": " means closer to real data is better." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 95, + 542, + 232 + ], + "lines": [ + { + "bbox": [ + 308, + 95, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 542, + 232 + ], + "type": "table", + "html": "
FID ↓R-precision ↑ (Top-3)Diversity →
Real0.0020.7979.503
JL2P [1]11.020.4867.676
Text2Gesture [5]7.6640.3456.409
T2M [14]1.0670.7409.188
MotionDiffuse [59]0.6300.7829.410
MDM [51]0.5560.6089.446
MLD [7]0.4730.7729.724
PhysDiff [58]0.4330.631-
Ours0.2120.6709.440
Ours xproj0.2350.6529.726
", + "image_path": "077c8037bb07fad7206e414702650fa0d547c21ea319a5990dba16dea1675079.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 246, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 377 + ], + "type": "text", + "content": "[14]. These include Fréchet Inception Distance (FID), R- Precision, and Diversity. FID measures the distance between the distributions of ground truth and generated motion using a pretrained motion encoder. R- Precision evaluates the relevance of the generated motion and its text prompt, while Diversity measures the variability within the generated motion. We also report Foot skating ratio, which measures the proportion of frames in which either foot skids more than a certain distance " + }, + { + "bbox": [ + 304, + 246, + 545, + 377 + ], + "type": "inline_equation", + "content": "(2.5\\mathrm{cm})" + }, + { + "bbox": [ + 304, + 246, + 545, + 377 + ], + "type": "text", + "content": " while maintaining contact with the ground (foot height " + }, + { + "bbox": [ + 304, + 246, + 545, + 377 + ], + "type": "inline_equation", + "content": "< 5\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 246, + 545, + 377 + ], + "type": "text", + "content": "), as a proxy for the incoherence between trajectory and human motion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 379, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 534 + ], + "type": "text", + "content": "In addition, for conditional generation with keyframe locations, we use Trajectory diversity, Trajectory error, Location error, and Average error of keyframe locations. Trajectory diversity measures the root mean square distance of each location of each motion step from the average location of that motion step across multiple samples with the same settings. Trajectory error is the ratio of unsuccessful trajectories, defined as those with any keyframe location error exceeding a threshold. Location error is the ratio of keyframe locations that are not reached within a threshold distance. Average error measures the mean distance between the generated motion locations and the keyframe locations measured at the keyframe motion steps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 540, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 599 + ], + "type": "text", + "content": "Datasets. We evaluate the text-to-motion generation using the HumanML3D [14] dataset, which is a collection of text-annotate motion sequences from AMASS [33] and HumanAct12 [16] datasets. It contains 14,646 motions and 44,970 motion annotations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. Both our motion DPM and trajectory DPM are based on UNET with AdaGN [12] depicted in details in the Supplementary. The motion DPM is an " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " model, while the trajectory DPM is an " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " model, as explained in Section 4.2, to enhance controllability. We utilized DDPM [20] with " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "T = 1,000" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " denoising steps for training and inference of both models. Additionally, we condition the generation process on text prompts in a classifier-free [21] manner, similar to MDM [51], and use the CLIP [41]" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2157" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 128, + 277, + 298 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "text", + "content": "Table 2. Trajectory-conditioned motions evaluation. The ground truth trajectory is used for imputing after each diffusion step. Comparing the effect of an original " + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "text", + "content": " with emphasis loss functions to the emphasis projection " + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{proj}}" + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "text", + "content": " after imputing whole trajectories after each diffusion step." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 128, + 277, + 298 + ], + "lines": [ + { + "bbox": [ + 56, + 128, + 277, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 277, + 298 + ], + "type": "table", + "html": "
ModelSpaceEmphasisFID ↓Foot skating ↓ ratio
MDM [51]xloss 1×0.9040.284
xprojc = 10.6320.304
c = 20.4640.309
c = 50.4660.256
c = 101.0290.161
Oursxloss 1×0.2780.262
loss 22×0.2560.250
loss 52×0.2400.249
loss 102×0.3200.265
xprojc = 10.3070.268
c = 20.2900.257
c = 50.2740.199
c = 100.1980.128
", + "image_path": "f5307deeb383aba53e95407c09ff31ce3784e060c878b56630fc93128636c648.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 319, + 217, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 217, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 217, + 330 + ], + "type": "text", + "content": "model as the text encoder across all tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 336, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 456 + ], + "type": "text", + "content": "Computational resources Our GMD architecture is capable of running both the motion and trajectory models on a single commercial GPU, such as the Nvidia RTX 2080 Ti, 3080, or 3090. The trajectory model achieved a throughput of 2,048 samples per second when run on an RTX 3090, with a training time of approximately 4.34 GPU hours. Meanwhile, the motion model achieved a throughput of 256 samples per second on an RTX 3090, with a training time of around 34.7 GPU hours. The total inference time for one sample is approximately 110 seconds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 464, + 192, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 192, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 192, + 477 + ], + "type": "text", + "content": "6.2. Text-to-motion generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "This section evaluates our model's performance in the standard text-to-motion generation task and compares it with other motion DPM baselines: MotionDiffuse [59], MDM [51], MLD [7], and PhysDiff [58]. Tab. 1 shows the results where our model architecture outperforms the baselines significantly in terms of motion quality measured by FID, while maintaining similar R-Precision and Diversity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 231, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 231, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 231, + 588 + ], + "type": "text", + "content": "6.3. Trajectory-conditioned generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "This section demonstrates how our emphasis projection method can address the issue of incoherent motion caused by spatial conditioning, specifically in the trajectory conditioning task, where the model is provided with ground-truth trajectories for imputation at each denoising step and is required to generate corresponding local poses. Both quantitative and qualitative results support that our emphasis projection leads to a reduction in Foot skating ratio, as evidenced in Tab. 2 and a more coherent motion in Fig. 5 compared to the MDM [51] model." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 545, + 188 + ], + "blocks": [ + { + "bbox": [ + 309, + 73, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 545, + 188 + ], + "type": "image", + "image_path": "f06bfcd27c56b984d76b2308cb8bbe7690b497403a79a5a9c44138f544871d35.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "text", + "content": "Figure 4. Comparing the evolution of the clean trajectory subject to classifier guidance from " + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "text", + "content": " DPMs. The " + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "text", + "content": " DPM shows significant resistance on the guidance signal as exhibited by the trajectory \"contraction\" behavior at " + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "inline_equation", + "content": "t\\to 0" + }, + { + "bbox": [ + 305, + 191, + 547, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 239, + 425, + 318 + ], + "blocks": [ + { + "bbox": [ + 309, + 239, + 425, + 318 + ], + "lines": [ + { + "bbox": [ + 309, + 239, + 425, + 318 + ], + "spans": [ + { + "bbox": [ + 309, + 239, + 425, + 318 + ], + "type": "image", + "image_path": "159c388c323ff7de7a83b347165a0566f84fcc767a9165a6eb8a7bb7207009b3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 322, + 546, + 377 + ], + "lines": [ + { + "bbox": [ + 305, + 322, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 546, + 377 + ], + "type": "text", + "content": "Figure 5. Generated motion, conditioned a given trajectory and text \"walking forward\". MDM [51] exhibits motion incoherence where the model disregards the trajectory and generates an inconsistent motion. Our method, improved by emphasis projection, deals effectively with the conditioning." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 239, + 543, + 318 + ], + "blocks": [ + { + "bbox": [ + 425, + 239, + 543, + 318 + ], + "lines": [ + { + "bbox": [ + 425, + 239, + 543, + 318 + ], + "spans": [ + { + "bbox": [ + 425, + 239, + 543, + 318 + ], + "type": "image", + "image_path": "5f64d7f192fdad6bad7a73d0ea65c795dd1bb50bdcd0c8c19f07b28e63e655d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 392, + 542, + 480 + ], + "blocks": [ + { + "bbox": [ + 309, + 392, + 542, + 480 + ], + "lines": [ + { + "bbox": [ + 309, + 392, + 542, + 480 + ], + "spans": [ + { + "bbox": [ + 309, + 392, + 542, + 480 + ], + "type": "image", + "image_path": "ca8ba36a0fc61da6ac016941d9b88c5a4eaa55b07059cfc44b199f1a4fcf9924.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 485, + 546, + 518 + ], + "lines": [ + { + "bbox": [ + 305, + 485, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 546, + 518 + ], + "type": "text", + "content": "Figure 6. Generated motion trajectories, conditioned on target locations at given keyframes. Without dense signal propagation, the model ignores the target conditions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 540, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 660 + ], + "type": "text", + "content": "We also compare our emphasis projection method with an alternative approach of increasing the trajectory loss strength during training. We include loss " + }, + { + "bbox": [ + 304, + 540, + 545, + 660 + ], + "type": "inline_equation", + "content": "k^2 \\times" + }, + { + "bbox": [ + 304, + 540, + 545, + 660 + ], + "type": "text", + "content": " baselines, where " + }, + { + "bbox": [ + 304, + 540, + 545, + 660 + ], + "type": "inline_equation", + "content": "k \\in \\{1, 2, 5, 10\\}" + }, + { + "bbox": [ + 304, + 540, + 545, + 660 + ], + "type": "text", + "content": ", for comparison. The results in Tab. 2 indicate that, while increasing the loss strength marginally improves both FID and Foot skating ratio, increasing it beyond a certain point leads to a decline in both FID and Foot skating ratio. By contrast, our emphasis projection method consistently leads to improvements in both metrics. We discuss this topic further in the Supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 670, + 486, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 486, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 486, + 682 + ], + "type": "text", + "content": "6.4. Keyframe-conditioned generation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "This section evaluates the quality and adherence of the generated motion to the desired goal. A viable solution" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2158" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 95, + 545, + 252 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "text", + "content": "Table 3. The effect of different conditioning strategies tested on keyframe-conditioning task. The keyframes " + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "inline_equation", + "content": "(N = 5)" + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "text", + "content": " are sampled from the ground truth motion trajectories with the same text prompts in the HumanML3D [14] test set." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 95, + 545, + 252 + ], + "lines": [ + { + "bbox": [ + 48, + 95, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 545, + 252 + ], + "type": "table", + "html": "
ModelConditioningFID ↓Foot ↓ skating ratioTraj. ↑ diversity (m.)Traj. err. ↓ (50 cm)Loc. err. ↓ (50 cm)Avg. err. ↓R-precision ↑ (Top-3)
MDM [51]Single stagex + τ=01.2560.2020.1340.0000.0000.0000.631
xproj + τ=02.9940.1510.1340.0000.0000.0000.554
xproj + τ=1002.2130.0950.2140.3260.1270.2360.555
xproj + no p2p1.6790.0920.3940.5190.3260.5430.548
Ours (xproj)Single stageτ=00.9020.1270.1170.0000.0000.0000.594
τ=1000.5230.0860.1570.1760.0490.1390.599
Two stageτ=1000.9370.0980.1200.0760.0200.1090.574
τ=3000.9380.0980.1270.1180.0310.1280.573
τ=5000.9080.0980.1400.1570.0430.1400.577
τ=7000.8980.0980.1620.1960.0580.1530.580
τ=9000.8740.0980.1920.2380.0800.1800.581
no p2p0.8620.1040.2220.2870.1180.2820.577
", + "image_path": "8b4b0f004d2567b4b9108b37df890a804cb90ff64c269ea75cb9b416a433e125.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 264, + 240, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 240, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 240, + 276 + ], + "type": "text", + "content": "must meet both criteria to an acceptable degree." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 396 + ], + "type": "text", + "content": "To achieve high-quality motion, both FID and Foot skating ratio are essential since FID alone cannot adequately measure the trajectory-motion coherence. Our Emphasis projection technique significantly improves motion coherence, reducing foot skating as shown in Tab. 3 while MDM [51] is unsuitable for this task due to the high motion incoherence. Furthermore, our improved architecture significantly improves motion quality in all cases. Note that without dense signal propagation, the model ignores the keyframe conditioning as shown in Fig 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": "While a single-stage model performs reasonably well due to emphasis projection, it is too restrictive at " + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": " (forced trajectory), resulting in relatively high Foot skating. This issue can be addressed by allowing more modification (increasing to " + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": " to 100) but at the cost of higher Loc. error." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": "Lastly, the trajectory model's better controllability reduces Location error by more than half compared to the single-stage model at " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\tau = 100" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": ". As expected, increasing " + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 456, + 287, + 515 + ], + "type": "text", + "content": " leads to more freedom in the model, resulting in increased Trajectory diversity, lower FID, and higher Location error." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 522, + 248, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 248, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 248, + 534 + ], + "type": "text", + "content": "6.5. Obstacle avoidance motion generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "type": "text", + "content": "Finally, we demonstrate our model's ability to generate motion on additional guidance on the obstacle avoidance task. In this task, we randomly sample the target point that the human needs to reach at a specific motion step along with a set of obstacles it needs to avoid, represented as a 2D SDF (Sec. 5.3). We show the qualitative results in Fig 7." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 621, + 201, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 201, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 201, + 634 + ], + "type": "text", + "content": "7. Discussion and Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "In this work, we propose GMD, a controllable human motion generation method based on goal functions. GMD produces high-quality and diverse motions and supports diverse possibilities for goal functions. Since obtaining necessary data and designing a classifier-free learning method for non-explicit goals, such as obstacle avoidance," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 263, + 545, + 353 + ], + "blocks": [ + { + "bbox": [ + 309, + 263, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 309, + 263, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 309, + 263, + 545, + 353 + ], + "type": "image", + "image_path": "0f7203132172e6f7bf31a65a030ba7fc2f52329d35f5fbdaa3f5e35c1bb705b2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 356, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 545, + 389 + ], + "type": "text", + "content": "Figure 7. Qualitative results from the obstacle avoidance task given keyframe locations and obstacles. The red crossed areas represent obstacles to avoid. More results are in the supplementary." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 410, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 471 + ], + "type": "text", + "content": "can be challenging, our GMD utilizes a classifier-based method which allows for more conditioning flexibility without retraining the model. Thus, our studies on effective classifier guidance will be useful for further including more guiding signals." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 475, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 500 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the SNSF project grant 200021 204840." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 512, + 365, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 512, + 365, + 524 + ], + "spans": [ + { + "bbox": [ + 306, + 512, + 365, + 524 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 533, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 310, + 533, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 533, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 310, + 533, + 545, + 576 + ], + "type": "text", + "content": "[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 578, + 547, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 578, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 311, + 578, + 547, + 621 + ], + "type": "text", + "content": "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. arXiv preprint arXiv:2211.09707, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 545, + 677 + ], + "type": "text", + "content": "[3] Jose A Arjona-Medina, Michael Gillhofer, Michael Widrich, Thomas Unterthiner, Johannes Brandstetter, and Sepp Hochreiter. RUDDER: Return decomposition for delayed rewards. In Advances in Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 680, + 545, + 714 + ], + "type": "text", + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 752, + 315, + 761 + ], + "type": "text", + "content": "2159" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Liu. eDiff-I: Text-to-Image diffusion models with an ensemble of expert denoisers. Nov. 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 162 + ], + "type": "text", + "content": "[5] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE virtual reality and 3D user interfaces (VR), pages 1-10. IEEE, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 196 + ], + "type": "text", + "content": "[6] T Brooks, A Holynski, and A A Efros. InstructPix2Pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 287, + 239 + ], + "type": "text", + "content": "[7] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. arXiv preprint arXiv:2212.04048, 2022. 1, 3, 4, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "type": "text", + "content": "[8] Xin Chen, Zhuo Su, Lingbo Yang, Pei Cheng, Lan Xu, Bin Fu, and Gang Yu. Learning variational motion prior for video-based motion capture. arXiv preprint arXiv:2210.15134, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 286, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 288, + 319 + ], + "type": "text", + "content": "[9] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. Aug. 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "text", + "content": "[10] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. June 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. Mofusion: A framework for denoising-diffusion-based motion synthesis. arXiv preprint arXiv:2212.04495, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 422 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. May 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 423, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 465 + ], + "type": "text", + "content": "[13] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 522 + ], + "type": "text", + "content": "[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 4, 7, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "text", + "content": "[15] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 580-597. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[17] Félix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in-between. ACM Transactions on Graphics (TOG), 39(4):60-1, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt image editing with cross attention control. Aug. 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[19] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. Oct. 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. June 2020. 2, 3, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Apr. 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "text", + "content": "[23] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. May 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "type": "text", + "content": "[24] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927. IEEE, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[25] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "text", + "content": "[26] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A versatile diffusion model for audio synthesis. Sept. 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 377, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 399 + ], + "type": "text", + "content": "[27] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. Feb. 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "text", + "content": "[28] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[29] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[30] Haoying Li, Yifan Yang, Meng Chang, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. SRDiff: Single image SuperResolution with diffusion probabilistic models. Apr. 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "text", + "content": "[31] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "text", + "content": "[32] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained diffusion models for unified human motion synthesis. arXiv preprint arXiv:2212.02837, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[33] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[34] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Image synthesis and" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 316, + 761 + ], + "type": "text", + "content": "2160" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "editing with stochastic differential equations. Aug. 2021. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 161 + ], + "type": "text", + "content": "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "text", + "content": "[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "text", + "content": "[37] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 480-497. Springer, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "text", + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. Sept. 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "type": "text", + "content": "[39] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-TTS: A diffusion probabilistic model for Text-to-Speech. May 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 392 + ], + "type": "text", + "content": "[40] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. ArXiv, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "text", + "content": "[43] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution image synthesis with latent diffusion models. Dec. 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 658 + ], + "type": "text", + "content": "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image diffusion models with deep language understanding. May 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[46] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image SuperResolution via iterative refinement. Apr. 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[47] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "learning using nonequilibrium thermodynamics. Mar. 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "text", + "content": "[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. July 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based generative modeling through stochastic differential equations. Nov. 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[50] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, Aug. 1988. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 230 + ], + "type": "text", + "content": "[51] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. Sept. 2022. 1, 3, 4, 7, 8, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "text", + "content": "[52] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. Edge: Editable dance generation from music. arXiv preprint arXiv:2211.10658, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 265, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 319 + ], + "type": "text", + "content": "[53] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "text", + "content": "[54] Christopher John Cornish Hellaby Watkins. Learning from delayed rewards. PhD thesis, King's College, Cambridge, 1989. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "text", + "content": "[55] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. Oct. 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 466 + ], + "type": "text", + "content": "[56] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI, pages 257-274. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "text", + "content": "[57] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4394-4402, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[58] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2212.02500, 2022. 1, 3, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "text", + "content": "[59] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 655 + ], + "type": "text", + "content": "[60] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "text", + "content": "[61] Yan Zhang, Michael J Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv preprint arXiv:2007.13886, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[62] Mengyi Zhao, Mengyuan Liu, Bin Ren, Shuling Dai, and Nicu Sebe. Modiff: Action-conditioned 3d motion gener" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 752, + 314, + 761 + ], + "type": "text", + "content": "2161" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 174 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "ation with denoising diffusion probabilistic models. arXiv preprint arXiv:2301.03949, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[63] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6225-6234, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[64] Zixiang Zhou and Baoyuan Wang. Ude: A unified driving engine for human motion generation. arXiv preprint arXiv:2211.16016, 2022. 1, 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 753, + 316, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 753, + 316, + 761 + ], + "spans": [ + { + "bbox": [ + 295, + 753, + 316, + 761 + ], + "type": "text", + "content": "2162" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_content_list.json b/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f30eeaa4e81f57f221cd24d985e058a829108046 --- /dev/null +++ b/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_content_list.json @@ -0,0 +1,1514 @@ +[ + { + "type": "text", + "text": "Guiding image captioning models toward more specific captions", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simon Kornblith", + "bbox": [ + 168, + 179, + 308, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google DeepMind", + "bbox": [ + 210, + 198, + 367, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lala Li", + "bbox": [ + 383, + 180, + 452, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Apple AI/ML", + "bbox": [ + 406, + 198, + 524, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zirui Wang $^{2*}$", + "bbox": [ + 527, + 180, + 633, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ University of Washington", + "bbox": [ + 566, + 198, + 777, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thao Nguyen $^{3\\dagger}$", + "bbox": [ + 700, + 179, + 823, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 250, + 312, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image captioning is conventionally formulated as the task of generating captions for images that match the distribution of reference image-caption pairs. However, reference captions in standard captioning datasets are short and may not uniquely identify the images they describe. These problems are further exacerbated when models are trained directly on image-alt text pairs collected from the internet. In this work, we show that it is possible to generate more specific captions with minimal changes to the training process. We implement classifier-free guidance [14] for an autoregressive captioning model by fine-tuning it to estimate both conditional and unconditional distributions over captions. The guidance scale applied at decoding controls a trade-off between maximizing $p(\\text{caption} | \\text{image})$ and $p(\\text{image} | \\text{caption})$ . Compared to standard greedy decoding, decoding with a guidance scale of 2 substantially improves reference-free metrics such as CLIPScore (0.808 vs. 0.775) and caption→image retrieval performance in the CLIP embedding space (recall@1 44.6% vs. 26.5%), but worsens standard reference-based captioning metrics (e.g., CIDEr 78.6 vs 126.1). We further explore the use of language models to guide the decoding process, obtaining small improvements over the Pareto frontier of reference-free vs. reference-based captioning metrics that arises from classifier-free guidance, and substantially improving the quality of captions generated from a model trained only on minimally curated web data.", + "bbox": [ + 75, + 282, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 719, + 207, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image captioning is both a difficult task for computer vision systems to perform and a difficult task to evaluate. Although automated captioning metrics rank the best captioning systems higher than humans, human raters still show a strong preference for human-generated captions [20], suggesting shortcomings in both captioning models and metrics. One shortcoming relates to the lack of specificity in generated captions. Conventional maximum likelihood-", + "bbox": [ + 75, + 744, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac98db833f7ae2e815e3f274219bb5ef911b530cefa419804e8ec00b668373c4.jpg", + "image_caption": [ + "$\\gamma = 1.0$ a man riding a blue motorcycle on a dirt road" + ], + "image_footnote": [], + "bbox": [ + 506, + 250, + 696, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc5d4f0ca1f455ed06eee5c5ae79fdd45a350774a01056ea77c2299f646d214b.jpg", + "image_caption": [ + "$\\gamma = 1.5$ a man riding a blue motorcycle on a straw bale", + "$\\gamma = 2.0$ rider on blue suzuki motorcycle near straw bales", + "$\\gamma = 3.0$ racer on blue suzuki motorcycle leaning up against straw bales", + "GT A person riding a baby blue motorcycle near haystacks", + "$y = 1.0$ $y = 2.0$", + "$y = 1,2$ $y = 3,0$", + "$\\gamma = 1.5$ $\\gamma = 4.0$", + "Figure 1. Using classifier-free guidance $(\\gamma > 1)$ results in more specific captions that are farther from the reference distribution. Left: Example of captions generated at different guidance scales for a single image. Right: Caption $\\rightarrow$ image recall@1 with CLIP ViT-B/32 vs. CIDEr score, for captions generated with different guidance scales $\\gamma$ on MS-COCO. Higher scales improve retrieval accuracy at the expense of CIDEr." + ], + "image_footnote": [], + "bbox": [ + 697, + 263, + 885, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based image captioning models attempt to generate captions such that the $p(\\text{caption} | \\text{image})$ is high. However, captions from the ground truth distribution are often nonspecific, e.g., human annotators will usually describe a German Shepard only as a dog. Moreover, previous work has emphasized \"reference-based\" captioning metrics that measure the match between generated captions and human-provided ground truth captions [28, 23, 42]. These metrics intrinsically penalize captions that are more specific than ground truth.", + "bbox": [ + 496, + 580, + 890, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we explore strategies to guide image captioning models to produce more specific captions by modifying the decoding distribution, and explore the tradeoffs in captioning metrics that result. We first investigate the application of classifier-free guidance (CFG) [14] to image captioning with autoregressive models. Classifier-free guidance increases $p(\\text{image|caption})$ at the expense of $p(\\text{caption|image})$ . Although CFG hurts reference-based image captioning metrics such as BLEU [28], ROUGE [23], and CIDEr [42], it improves \"reference-free\" metrics that measure captions' specificity via the similarity between the", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work performed while at Google.", + "bbox": [ + 93, + 875, + 281, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work performed as a student researcher at Google.", + "bbox": [ + 96, + 887, + 369, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15259", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image and the generated caption in the embedding space of image-text models [13] or caption $\\rightarrow$ image retrieval performance. Qualitatively, we find that captions generated with CFG are more specific than both the ground truth captions and captions generated without CFG, but they are less grammatical, particularly at high CFG scales.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond classifier-free guidance, we experiment with guiding image captioning models using the probability distribution obtained from a few shot-prompted language model (LM). We find that using a language model to guide a captioning model trained on MS-COCO [24] with descriptive manually written captions can allow it to achieve slightly better trade-offs between reference-free vs. reference-based captioning metrics than those observed with CFG. LM guidance also substantially improves the captions produced by a model trained exclusively on minimally curated web data. Although this model achieves a CIDEr score of only 21.8 without guidance, this CIDEr score improves to 57.4 when guided by a language model prompted with 20 captions from the MS-COCO training set. In summary, our contributions are as follows:", + "bbox": [ + 75, + 183, + 470, + 409 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose two strategies to guide image captioning models to produce more specific captions: classifier-free guidance and language model guidance.", + "- We demonstrate that classifier-free guidance yields captions that are closer to the corresponding image in the embedding space of image-text models, but are farther from human-provided reference captions.", + "- We show that language model guidance can alter caption styles, substantially improving captions produced by a model trained only on minimal curated web data and marginally improving the trade-off between captioning metrics observed with classifier-free guidance." + ], + "bbox": [ + 76, + 415, + 467, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 617, + 215, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Measuring specificity of captions. Early work using neural networks for image captioning found that models have a propensity to regurgitate captions from their training data, and as a result, the generated captions are not descriptive enough to uniquely identify images [43, 11]. To address this shortcoming, Lindh et al. [25] proposed to use caption $\\rightarrow$ image recall with an image retrieval model to examine whether images can be retrieved from generated captions, and further attempt to differentiate through this retrieval process to train a captioning model. Their approach marginally improves retrieval accuracy, but worsens reference-based captioning metrics. More recent work has adopted approaches to evaluate the specificity of captions based on the CLIP image-text model [30]. Hessel et al. [13] propose CLIPScore, an image captioning metric based on the cosine similarity between CLIP embeddings of the image and the generated caption. Kasai et al. [20] re", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "port that CLIPScore-based metrics align better with human judgments compared to reference-based captioning metrics.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Improving specificity of captions. Recent work has attempted to train caption models to directly maximize the similarity of captions with corresponding images in the CLIP embedding space using reinforcement learning. This work has found that adding such losses worsens standard reference-based captioning metrics but improve similarity and retrieval in the CLIP embedding space [16, 6, 51], similar to our observations regarding CFG. Wen et al. [44] attempt to generate prompts for text-to-image generative models that correspond to specific images without a captioning model, by directly optimizing the similarity between the text and image in the CLIP embedding space using a gradient-based discrete optimization procedure, but the resulting text is not grammatical.", + "bbox": [ + 496, + 123, + 890, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other work has attempted to generate more descriptive captions through different means. Dense captioning [46] aims to detect and caption all objects in an image, but concatenating all of these captions leads to long and unnatural captions, whereas CFG produces single-sentence captions. The Localized Narratives dataset [29] contains visually grounded captions for MS-COCO images collected through voice annotation. These captions are substantially more descriptive than the captions in the MS-COCO dataset and can be used for model training. Two recent papers propose to combine multiple captions from a single model [5] or outputs of different vision models [34], and then combine them using a language model. The resulting captions are longer, but achieve greater caption→image recall.", + "bbox": [ + 496, + 337, + 890, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Captioning from uncurated data. In Section 4.2, we explore the use of LM guidance for captioning with access to uncurated image-text data from the web and a small number of captions but not images from the target distribution. This setting, which does not rely on aligned images and captions from the target distribution, is often referred to as \"zero-shot\" captioning, and previous work has pursued a number of alternative approaches. Flamingo [3] and CM3 [1] perform zero-shot captioning by pretraining on interleaved image/text data. MAGIC [39] and ZeroCap [41] generate captions using a combination of guidance from CLIP and a large language model. Other recent work adapts CLIP to perform captioning by training a text decoder using only captions, with no corresponding images [27, 22].", + "bbox": [ + 496, + 551, + 890, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classifier-free guidance. CFG is widely used in diffusion-based and autoregressive text-to-image models [26, 32, 35, 33, 12, 48]. Because of the popularity of the combination of CFG and diffusion, previous work that has performed image captioning with diffusion models has also examined the use of CFG. This work finds either no benefit to using CFG [45] or a small and inconsistent benefit that appears to vary with minor changes in training settings [52]. However, these studies do not seek to generate", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15260", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "more specific captions, and thus measure only reference-based captioning metrics, which we likewise find do not benefit from CFG. Concurrently with out work, [36] propose to use classifier-free guidance to improve prompt following in large language models.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 180, + 174, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Classifier-free guidance for image captioning", + "text_level": 1, + "bbox": [ + 76, + 205, + 455, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}$ be an image caption and $\\mathbf{y}$ be the corresponding image. A standard captioning model aims to model the likelihood $p(\\mathrm{x}|\\mathrm{y})$ , factorized autoregressively in terms of the probability of each token given previous tokens", + "bbox": [ + 75, + 228, + 467, + 289 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathrm {x} | \\mathrm {y}) = p \\left(\\mathrm {x} _ {n} \\mid \\mathrm {x} _ {n - 1}, \\dots \\mathrm {x} _ {1}, \\mathrm {y}\\right) \\dots p \\left(\\mathrm {x} _ {1} \\mid \\mathrm {y}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 292, + 468, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The network is trained so that its output distribution $q_{\\theta}(\\mathbf{x}_n|\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y}) \\stackrel{\\mathrm{def}}{=} \\text{softmax}(f_{\\theta}(\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y}))$ approximates $p(\\mathbf{x}_n|\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y})$ . At inference time, one typically uses beam search or greedy decoding to produce a caption that has a particularly high probability. In this work, we use greedy decoding because it is the more common choice and it is also simpler to implement.", + "bbox": [ + 75, + 314, + 467, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free guidance (CFG) [14] modifies the inference procedure so it instead aims to achieve high values of", + "bbox": [ + 75, + 421, + 467, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\boldsymbol {\\theta}, \\gamma} (\\mathrm {x}, \\mathrm {y}) \\stackrel {\\text {d e f}} {=} p (\\mathrm {x}) \\left(\\frac {p (\\mathrm {x} | \\mathrm {y})}{p (\\mathrm {x})}\\right) ^ {\\gamma} \\propto p (\\mathrm {x}) p (\\mathrm {y} | \\mathrm {x}) ^ {\\gamma}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 455, + 468, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where proportionality holds because $p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x}) = p(\\mathrm{y}|\\mathrm{x}) / p(\\mathrm{y})$ and $p(\\mathrm{y})$ is fixed. The parameter $\\gamma$ is called the guidance scale and controls the trade-off between maximization of $p(\\mathrm{x}|\\mathrm{y})$ and $p(\\mathrm{y}|\\mathrm{x})$ . When $\\gamma = 1$ , $l_{\\theta, \\gamma}(\\mathrm{x}, \\mathrm{y}) = p(\\mathrm{x}|\\mathrm{y})$ and guidance has no effect. Setting $\\gamma > 1$ inflates the probability of the image given the caption $p(\\mathrm{y}|\\mathrm{x})$ relative to the unconditional probability of the caption $p(\\mathrm{x})$ .", + "bbox": [ + 75, + 492, + 467, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ho and Salimans [14] originally proposed CFG in the context of diffusion models, which estimate the score functions $\\nabla \\log p(\\mathbf{x}|\\mathbf{y})$ and $\\nabla \\log p(\\mathbf{x})$ . Although $l_{\\theta,\\gamma}(\\mathbf{x},\\mathbf{y})$ factorizes autoregressively, it is not a normalized probability distribution, so it is not entirely clear how one should sample tokens when performing autoregressive generation. Crowson [8] suggested to sample from", + "bbox": [ + 75, + 598, + 467, + 704 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {q} _ {\\boldsymbol {\\theta}, \\gamma} (\\mathrm {x} _ {n} | \\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathrm {y}) \\stackrel {\\mathrm {d e f}} {=} \\operatorname {s o f t m a x} (f _ {\\boldsymbol {\\theta}} (\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathbf {0}) \\\\ + \\gamma \\left(f _ {\\boldsymbol {\\theta}} \\left(\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathrm {y}\\right) - f _ {\\boldsymbol {\\theta}} \\left(\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathbf {0}\\right)\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 708, + 467, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{\\theta}(\\mathbf{x}_{n - 1},\\ldots ,\\mathbf{x}_1,\\mathbf{0})$ are logits generated by the model without conditioning, usually by passing zeros in place of the conditioning information. This formulation has been successfully applied in autoregressive image models [12, 48]. In our experiments, since we decode greedily, i.e., at each step we take the token that maximizes $\\tilde{q}_{\\theta ,\\gamma}(\\mathrm{x}_n|\\mathrm{x}_{n - 1},\\dots ,\\mathrm{x}_1,\\mathrm{y})$ and thus $l_{\\theta ,\\gamma}(\\mathrm{x},\\mathrm{y})$ , any form of normalization of the per-step sampling distribution would produce the same captions. We provide pseudocode in Appendix A.1. Crucially, the only change required to train", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing is to learn an unconditional model of captions alongside the image-conditional model, which we accomplish by randomly masking the image embeddings.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Language model guidance", + "text_level": 1, + "bbox": [ + 500, + 148, + 736, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by classifier-free guidance, we consider language model (LM) guidance, which attempts to maximize", + "bbox": [ + 498, + 172, + 890, + 203 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\boldsymbol {\\theta}, \\gamma} ^ {\\prime} (\\mathrm {x}, \\mathrm {y}) \\stackrel {\\text {d e f}} {=} q (\\mathrm {x}) \\left(\\frac {p (\\mathrm {x} | \\mathrm {y}) ^ {\\alpha}}{p (\\mathrm {x}) ^ {\\beta}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 207, + 890, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(x)$ and $p(x|y)$ are obtained from a captioning model as in CFG but $q(\\mathbf{x})$ is obtained from a language model that was trained independently (but with the same vocabulary) on a large text corpus. The quantity $p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x}) = p(\\mathrm{x},\\mathrm{y}) / (p(\\mathrm{x})p(\\mathrm{y}))$ measures the strength of the association between a caption and an image; its logarithm is the pointwise mutual information (PMI). LM guidance relies on the assumption that, even for large shifts in the prior distribution of captions $p(\\mathbf{x})$ , the shift in PMI will be small. Empirically, we obtain better results by allowing different exponents for the numerator and denominator, with $\\alpha > \\beta$ . This decoupling resembles $\\mathrm{PMI}^k$ [9], which reduces the bias of PMI toward rare associations. We provide a more detailed derivation in Appendix A.2.", + "bbox": [ + 496, + 243, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We investigate two applications of LM guidance. First, we combine a captioning model fine-tuned on MS-COCO with a LM prompted with manually written descriptive captions to alter the style of the captions the model produces. The manually written prompts are shown in Appendix A.4. Second, we combine a captioning model trained only on low-quality web data with a LM prompted with varying numbers of examples from the MS-COCO training set to evaluate the ability of LM guidance to elicit higher-quality captions without high-quality paired data. We randomly select a different set of captions for each minibatch of four test examples. In both cases, we separate the captions with two newlines. Because this format leads the LM to place probability mass on the newline token to end the caption, we transfer the probability mass from the newline token to the EOS token. See Appendix A.3 for pseudocode.", + "bbox": [ + 496, + 455, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Models and training", + "text_level": 1, + "bbox": [ + 500, + 710, + 692, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our captioning model is a \"bottleneck\" variant of CoCa-Base [47], which combines a contrastive loss with a captioning loss to simultaneously learn aligned image and text embeddings as well as a captioner. The architecture consists of an image encoder, a unimodal text decoder, and a multimodal text decoder, each of which are Transformers with 12 layers, 768 hidden dimensions, an MLP of size 3072, and 12 self-attention heads, matching $\\mathrm{BERT}_{\\mathrm{BASE}}$ [10] and GPT-1 [31]. The image encoder is a ViT-B/18 that processes $288 \\times 288$ input and produces an embedding such that images are embedded close to their corresponding text.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15261", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CoCa's multimodal text decoder processes the representations of the image encoder to produce a caption. Whereas [47] conditions the multimodal text decoder using crossattention to pooled representations, our bottleneck variant uses only the contrastive image embedding. Appendix A.5 shows a diagram of the resulting architecture. We adopt this bottleneck variant because of its simplicity and the conceptual appeal: When CFG is used, the captioner's role is to invert the image embedding, providing a caption that, when embedded by the text encoder, lies close to it. However, as we show in Appendix B.1, this choice of the bottleneck model is not critical, and CFG is equally effective with the standard CoCa architecture with attention pooling.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For CFG experiments, we pretrain our model on an image-text dataset comprising images from the JFT-5B dataset [40, 49] paired with their corresponding label names substituted into a randomly selected prompt from the list provided by Radford et al. [30], web images paired with noisy alt text from the ALIGNN dataset [17], and a small amount of data from other sources. We do not mask conditioning information during pretraining. We then fine-tune on the combined MS-COCO train and Karpathy validation splits [18] using Adam with batch size 128. We linearly warm up to a learning rate of $1 \\times 10^{-5}$ over the first 1,000 steps and linearly decay to zero over the rest of training. We vary $\\gamma \\in \\{1.0, 1.2, 1.5, 2.0, 3.0, 4.0\\}$ , conditioning masking proportion in $\\{0.0, 0.25, 0.5, 0.75\\}$ , and numbers of steps in $\\{5,000, 10,000, 20,000, 50,000\\}$ . We report results from the model trained for 20,000 steps with masking proportion 0.5, which achieves near-optimal results, in Tables 1 and B.4, and sample example captions from it. To ensure that results generalize across datasets, we also experiment with a model fine-tuned on Conceptual Captions [37] for 100,000 steps with masking proportion 0.5.", + "bbox": [ + 75, + 287, + 472, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For LM guidance experiments, we pretrain on the JFT-5B and ALIGN datasets, again following the recipe of [47]. For zero-shot captioning experiments, we fine-tune this model on the same datasets for an additional 50,000 steps with conditioning masking proportion of 0.5 to improve our ability to sample unconditionally. For LM guidance on MS-COCO, we first fine-tune on ALIGN, JFT-5B images backcaptioned by an MS-COCO fine-tuned CoCa2B model, and a small amount of internal data before finetuning on MS-COCO. Our language model is a variant of Primer [38] with 2 billion parameters, trained on a similar dataset to that used to train PaLM [7].", + "bbox": [ + 75, + 606, + 472, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 797, + 199, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the standard reference-based captioning metrics BLEU-4, METEOR, ROUGE, and CIDEr, as well", + "bbox": [ + 76, + 821, + 468, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as reference-free captioning metrics based on CLIP ViT-B/32 [30]. The first reference-free captioning metric is CLIPScore [13], which is defined as $\\mathrm{CLIP - S}(c,v) = 2.5\\cdot \\max (\\cos (c,v),0)$ where $c$ and $v$ are the CLIP embeddings of the caption and image respectively. The second reference-free metric measures the accuracy with which we can retrieve an image from the generated caption within a given test split by taking the $k$ nearest neighbors of the caption in the CLIP embedding space. Because recall@ $k$ for $k > 1$ is highly correlated with recall@1 (R@5: $r = 0.99$ , R@10: $r = 0.98$ ), we plot only recall@1. We additionally report RefOnlyCLIP-S, a reference-based metric that uses the CLIP text encoder to compute the similarity of CLIP embeddings of the generated captions with embeddings of ground truth captions, and RefCLIP-S, which takes the average of the per-image harmonic means of CLIP-S and RefOnlyCLIP-S [13]. Unless otherwise stated, all evaluation is performed on the MS-COCO Karpathy test split [18].", + "bbox": [ + 500, + 90, + 893, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 378, + 586, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Classifier-free guidance", + "text_level": 1, + "bbox": [ + 500, + 405, + 718, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first investigate the trade-off between reference-based and reference-free image captioning metrics as a function of guidance scale. Because different guidance scales and metrics could conceivably benefit from different fine-tuning hyperparameter combinations, we plot all results from our hyperparameter grid in Figure 2. Although standard greedy decoding $(\\gamma = 1.0)$ produces the highest CIDEr, METEOR, ROUGE, and BLEU-4 scores, higher guidance weights consistently yield higher values of reference-free captioning metrics. In particular, $\\gamma = 3.0$ offers both the best caption→image recall and the best CLIP-Score.", + "bbox": [ + 496, + 430, + 893, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 compares our results, obtained from a single model evaluated at different guidance scales, with previous work that reports either CLIPScore or CLIP ViT-B/32 caption $\\rightarrow$ image retrieval performance. Although our model is trained with standard cross-entropy loss rather than a CLIP-based loss and our pretraining dataset is distinct from CLIP's, sampling from our model with CFG yields higher CLIPScores than all other models trained without CLIP-based losses, and better CLIP caption $\\rightarrow$ image retrieval even when compared with models that use CLIP-based losses.", + "bbox": [ + 496, + 612, + 895, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present examples of captions generated at different CFG scales in Figure 3. Higher CFG strengths lead to more descriptive captions. At $\\gamma = 1.0$ , the central object in the top left image is described as a \"car\" as in the ground truth caption, whereas at $\\gamma > 1.0$ it is a \"station wagon.\" Similarly, at low CFG strengths, the birds in the center image are described simply as \"birds,\" whereas at $\\gamma = 2.0$ they become \"crested cranes.\" However, at $\\gamma = 3.0$ , captions clearly become less grammatical, containing repeated", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We find that passing an all-zero image embedding to the pretrained model yields samples that resemble the unconditional distribution, suggesting that it implicitly learns to model the unconditional distribution.", + "bbox": [ + 76, + 862, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15262", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66ceb179a19a7d97a20c9c8a9512f0e8d35e690abc03d9acd40efaac31ae6f19.jpg", + "image_caption": [ + "Figure 2. Classifier-free guidance controls a trade off between reference-free and reference-based captioning metrics. Each point reflects a model trained with a different hyperparameter combination; each color represents a $\\gamma$ value used to decode. Models are evaluated with different guidance scales $\\gamma$ , using reference-free captioning metrics based on CLIP ViT-B/32 (y-axes; top: CLIPScore, bottom: recall@1) and reference-based captioning metrics (x-axes). The dashed line reflects the value of the reference-free captioning metric for the ground-truth captions obtained from MS-COCO.", + "Figure 3. Caption descriptiveness increases with CFG strength, but high CFG strengths produce agrammatical captions. Here we show examples of captions generated with different classifier-free guidance scales, for randomly selected images without human faces from the MS-COCO Karpathy test split. Captions labeled $\\gamma = 1.0$ are obtained without CFG; $\\gamma > 1$ uses CFG; GT = ground truth." + ], + "image_footnote": [], + "bbox": [ + 148, + 85, + 816, + 296 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c61f430cc7fd0f6ee4d9217300ecca01f8c4ce44f97cfe57dc0488ae4c760fce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelReference-Based MetricsReference-Free MetricsRefCLIP-S
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10
Models trained with CLIP features or losses:
CLIP-Captioner [4]38.729.358.6126.00.8110.7540.814
UMT-BITG [16]37.328.257.9122.60.772
X-LAN+SCST+GEG [51]36.528.757.5121.728.150.367.2
CIDEr + CLIP-S Reward [6]37.728.858.3124.60.77224.450.263.1
CLIP-S Reward [6]6.218.731.611.20.86042.571.682.2
ZeroCap [41]2.611.514.60.870.79
Models trained without access to CLIP:
UMT-BITG w/o CLIP loss [16]37.628.358.1122.50.725
VinVL-large [50]41.030.959.4*140.90.91*0.78*0.84*
Ours (γ = 1.0)36.130.558.2126.10.9000.77526.551.964.10.830
Ours (γ = 1.2)35.130.057.5124.10.8990.78531.357.469.30.835
Ours (γ = 1.5)31.528.454.4113.20.8910.79636.664.075.00.838
Ours (γ = 2.0)20.923.343.078.60.8620.80844.671.781.70.831
Ours (γ = 3.0)11.517.129.441.70.8200.80849.475.784.70.811
Ours (γ = 4.0)6.512.318.417.30.7660.78244.771.380.90.771
", + "bbox": [ + 84, + 383, + 885, + 609 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/278d2d12a045cd2f87d31f31a841670fac7bdd71fc843f943724616bb5c8f58e.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of our approach with results from previous work that reports CLIP-based metrics. For VinVL-large, * indicates metrics from [19]." + ], + "table_footnote": [], + "table_body": "
γ=1.0: a car with a surfboard on top of it parked next to a car\nγ=1.5: a vintage station wagon with a surfboard on top\nγ=2.0: a antique station wagons and a car buick station wagon\nγ=3.0: buick woody wyddo staitonwagon and surf green station wagon parked in front of car show establishment\nGT: An old green and brownγ=1.0: a knife is sitting on a cutting board next to a sliced carrot\nγ=1.5: a knife is sitting on a cutting board next to an orange\nγ=2.0: knife sitting on cutting board next to whole one\nγ=3.0: knife sitting on cutting board next to misschappi carrot on cutting board\nGT: A knife sticking out of the side of a block of cheese.γ=1.0: a herd of sheep grazing on a grass covered road\nγ=1.5: sheep grazing on a high-way with a truck on the road\nγ=2.0: sheep graze on freeway medians where grass is grown\nγ=3.0: grazing trucks blocking sheep on roadway grazing grass\nGT: A large herd of sheep are grazing by the busy road.γ=1.0: two birds standing in a cage with their heads in the air\nγ=1.5: two birds standing inside of a cage in a zoo\nγ=2.0: two crested cranes inside a wire cage\nγ=3.0: crested tantalus cranes caged together in bird-cage enclosure\nGT: Two birds who are looking out of the cage they are in.γ=1.0: a view of a city with a clock tower in the background\nγ=1.5: a city with steeples and trees and buildings\nγ=2.0: spires of churches line a city skyline\nγ=3.0: spires steeples buildings trees church spires and trees\nGT: A clock that is on the side of a tower.γ=1.0: a kitchen with a microwave and refrigerator\nγ=1.5: a kitchen with a mi-crowage and a refrigerator\nγ=2.0: a kitchen with red ap-pliances and white cup-boards\nγ=3.0: appliances sit in a small empty dingroom red and white kitchen\nGT: A kitchen that has a tile floor, a refrigerator, a microwave, and a toaster.γ=1.0: a bathroom with a large mirror and a bathtub\nγ=1.5: a bathroom with a large mirror and a bathtub\nγ=2.0: a spacious bathroom with a large mirror and a large tub\nγ=3.0: spacious bathroom with chandelier over tub mir-rors and tv\nGT: A bathroom with a tub, sinks, lights and a televi-sion.
", + "bbox": [ + 84, + 654, + 879, + 829 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15263", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "words (\"woody woody\") and nonsense words (\"misshappi\", \"dingroomy\"). Figure 4 shows captions obtained with and without CFG next to the top 5 closest images in the embedding space of CoCa 2B [47],² where it is clear that CFG adds details to captions that help to distinguish them from other captions in the test split. We provide additional examples in Appendix C.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To provide additional quantitative assessments of the specificity of elicited captions, we perform two additional evaluations, described further in Appendix B.2. First, we generate captions for the Stanford Dogs [21] test set, which consists of 8,580 images in total of 120 breeds of dogs, and examine their properties. Without guidance, only $1.9\\%$ of captions contain one of the 120 breed names, whereas at $\\gamma = 2.0$ , $42.4\\%$ do. The percentage of these breed names that are correct changes little, from $61.7\\%$ without guidance to $58.5\\%$ at $\\gamma = 2.0$ . Second, we performed a human evaluation comparing captions of MS-COCO test set images obtained without guidance and at $\\gamma = 2.0$ . We asked subjects to select the caption that is \"better\" and \"more descriptive\" or to indicate that they are both equal. When we asked these questions separately, we found that the two sets of captions are statistically indistinguishable. However, when asking both questions on the same survey, we found that captions generated without guidance are slightly \"better\" (50.5% vs. $46.6\\%$ , $p = 0.006$ , binomial test) but captions generated at $\\gamma = 2.0$ are \"more descriptive\" (52.7% vs. $45.8\\%$ , $p = 1 \\times 10^{-6}$ ).", + "bbox": [ + 75, + 198, + 473, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the reliability of our results, we further measure the impact of CFG on three additional datasets, nocaps [2], Flickr-8k [15], and Conceptual Captions (CC3M) [37], as well as with alternative retrieval models. nocaps is a test set for captioning models with objects not present in MS-COCO; Flickr-8k is a small captioning dataset collected using a different procedure than MS-COCO; and Conceptual Captions is a set of 3.3M captions collected from filtered alt-text. We fine-tune the bottleneck CoCa-Base model directly on CC3M, and use our model fine-tuned on MS-COCO to caption images on nocaps and Flickr-8K. As shown in Figure 5, we find trade-offs between reference-based and reference-free captioning metrics similar to those above. In Appendix B.3, we report reference-free captioning metrics on MS-COCO computed with two additional retrieval models: the pretrained CoCa 2B model from [47] and the fine-tuned CoCa Base model that we use to generate captions. With both models, CFG substantially increases recall, in line with results obtained with CLIP ViT-B/32.", + "bbox": [ + 75, + 515, + 473, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although CFG produces captions that are more successful at uniquely identifying images than decoding from the conditional distribution, caption lengths are similar for", + "bbox": [ + 75, + 816, + 472, + 864 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fd0d825af643485344d7b9cc5c1af8faf84bc57f9bea56ace03c9b449b3f99c.jpg", + "image_caption": [ + "Figure 4. Captions generated with CFG contain specific details that improve retrieval. For each reference image (far left), we show captions at $\\gamma = 1.0$ (no guidance) and $\\gamma = 2.0$ . To the right, we show the closest images to each caption in the CoCa embedding space. Reference images are selected at random subject to the constraints that the closest image differs between $\\gamma$ values and there are no identifiable human faces." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 893, + 534 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2d56f207cc27833fea0be98669f6dc37a54d07155aa418bb1d081a3eb1a6c4a.jpg", + "image_caption": [ + "Figure 5. CFG also yields trade-offs between captioning metrics on nocaps, Flickr-8K, and CC3M." + ], + "image_footnote": [], + "bbox": [ + 500, + 640, + 893, + 868 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2We use CoCa 2B rather than CLIP because, quantitatively and qualitatively, it provides better retrieval results both with and without guidance.", + "bbox": [ + 75, + 875, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15264", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ec1eee5c9892209172572e6c8a0943f202c9b00974e2da1d2759c711ae06251b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
γWordsCharacters
1.09.6 ± 1.444.2 ± 7.2
1.29.6 ± 1.444.7 ± 7.4
1.59.4 ± 1.445.7 ± 7.8
2.09.3 ± 2.450.3 ± 18.6
3.010.7 ± 7.669.0 ± 56.1
4.019.9 ± 16.9161.2 ± 140.0
", + "bbox": [ + 160, + 88, + 387, + 178 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Moderate CFG scales do not substantially change caption lengths, although higher CFG scales result in longer captions. Numbers are mean $\\pm$ standard deviation.", + "bbox": [ + 75, + 188, + 468, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eff26af2f82da7153684253b61b1e109761bee454bd2e4357f22683717325faa.jpg", + "image_caption": [ + "Figure 6. Language model guidance produces captions that slightly exceed the Pareto frontier of CIDEr vs. caption $\\rightarrow$ image retrieval accuracy on MS-COCO." + ], + "image_footnote": [], + "bbox": [ + 124, + 241, + 416, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\gamma \\in [1,2]$ , as shown in Table 2. Thus, at low guidance strengths, CFG improves recall by making more efficient use of words, rather than by producing more verbose captions. Higher CFG strengths lead to longer captions but, as described above, these captions are agrammatical and contain nonsense words.", + "bbox": [ + 75, + 430, + 468, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Language model guidance", + "text_level": 1, + "bbox": [ + 76, + 530, + 313, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first experiment with guiding a captioning model fine-tuned on MS-COCO to produce more descriptive captions using a language model prompted with manually written prompts. We first manually wrote a prompt containing 10 descriptive captions of COCO test set images (Appendix A.4). We then sweep over $\\alpha \\in \\{1,2,3,4,5,6,7,8,9,10,12,15\\}$ and $\\beta \\in \\{0,\\alpha /4,\\alpha /2,3 / 4\\alpha ,\\alpha \\}$ , and compare the resulting retrieval/CIDEr trade-off to that produced by the same model with CFG. We observe that it is possible to obtain small improvements upon the Pareto frontier provided by CFG, as shown in Figure 6. With $\\alpha = 5$ , $\\beta = -5 / 2$ , LM guidance achieves CLIP ViT-B/32 R@1 of $39.6\\%$ and CIDEr of 114.4, whereas CFG with $\\gamma = 1.6$ is worse on both metrics, achieving R@1 of $39.0\\%$ and CIDEr of 109.3.", + "bbox": [ + 75, + 553, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further experiment with using prompting to control the captioner using a manually written prompt of 25 captions in the form of \"a photo of NUMBER OBJECTS\" (e.g., \"a photo of eight apples\"; see Appendix A.4). With $\\alpha = \\beta = 1$ , the guided model is able to match this format and counts the number of objects in images (Figure 7).", + "bbox": [ + 75, + 780, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We next investigate whether language model guidance can elicit better captions from a model trained only on", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86ea8725a151896a78c9bc2d7ca0dd77994637f50668653f9582caf42b0fdcc3.jpg", + "image_caption": [ + "a photo of two dogs" + ], + "image_footnote": [], + "bbox": [ + 504, + 87, + 599, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04ff0335e5d94d7e00d59a44f0a1d6d6a969e82fbd910d6d5fd04b378bfbef3b.jpg", + "image_caption": [ + "a photo of one bird" + ], + "image_footnote": [], + "bbox": [ + 602, + 88, + 700, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4dd6affcc4456e67eb59b154b72d1e6d0c652ae69d794473dc5449318f79ef18.jpg", + "image_caption": [ + "a photo of four suitcases", + "Figure 7. Captions generated with LM guidance with a prompt of 25 captions in the form of \"a photo of NUMBER OBJECTS\". Examples are selected to show different numbers of objects." + ], + "image_footnote": [], + "bbox": [ + 702, + 88, + 800, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a2e42dd24594938eb170c50661f9dc971aa37ccbc9747a69b8fa313483c379a.jpg", + "image_caption": [ + "a photo of five sheep" + ], + "image_footnote": [], + "bbox": [ + 803, + 88, + 898, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d47a7b746a450aefdfc7122b35fd5c0bd04085c33a4baf0caaa5cfd4906cfc94.jpg", + "image_caption": [ + "Figure 8. LM guidance substantially improves CIDEr and retrieval scores of a model trained solely on minimally curated web data and evaluated on MS-COCO. The x-axis shows the number of captions used to prompt the LM; we do not prompt with images." + ], + "image_footnote": [], + "bbox": [ + 511, + 241, + 687, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf52968589ca7511b10d22284360cd466927fd7cfc51cc8e7966d06cf21efd8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 241, + 880, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "low-quality data. Here, we use a CoCa model that is pretrained on image-alt text pairs from the web (the ALIGN dataset [17]) and classification labels converted to text (the JFT-5B dataset [49]), without any additional fine-tuning. Because the data distribution places higher probability mass on short, non-descriptive captions than on longer captions, the resulting model is of limited utility for captioning, and would generally need to be fine-tuned on another dataset such as MS-COCO before being applied to a captioning task. Rather than fine-tune, we use LM guidance to prompt the model with captions from the MS-COCO training set.", + "bbox": [ + 496, + 443, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LM guidance substantially improves the quality of the captions produced by the original pretrained CoCa model without any clean parallel data. With LM guidance, we achieve CIDEr scores of 48.6 with 5 shots and 59.7 with 50 shots, far exceeding the CIDEr score of 21.8 obtained with no guidance. Figure 8 shows CIDEr and CLIP recall@1 scores for LM guidance of this pretrained CoCa model as a function of the number of shots, with $\\alpha = \\beta = 1$ . Table 3 compares classifier-free guidance and LM guidance. CFG yields higher CLIP-Scores and retrieval accuracy than LM guidance with $\\alpha = \\beta = 1$ , but LM guidance provides much higher CIDEr scores.", + "bbox": [ + 496, + 611, + 892, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare captions generated with CFG to those generated with LM guidance for four images in Figure 9. In general, CFG produces agrammatical captions, whereas LM guidance produces grammatical captions but hallucinates details. For example, the image in the upper left shows two elephants and no zebras, but LM guidance leads to the caption \"an elephant and a zebra in a field.\"", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15265", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c05c2217211674894cd2b0c0f83992bacc004820b3e9d17b42f8a27da6594aed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelReference-Based MetricsReference-Free Metrics
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10RefCLIP-S
Classifier-free guidance:
γ = 1.08.28.321.821.80.7660.6949.019.526.10.725
γ = 1.28.69.524.525.00.7810.71812.727.235.10.745
γ = 1.58.910.025.625.20.7800.72816.733.843.00.750
γ = 2.08.19.723.922.90.7770.74121.240.851.10.755
γ = 3.07.18.720.018.50.7670.75325.847.858.30.756
γ = 4.06.47.516.313.90.7490.74327.348.558.10.742
Language model guidance with α = β = 1:
2 captions12.714.634.739.30.8060.68810.023.732.40.740
5 captions15.016.639.148.60.8270.71212.427.537.50.763
10 captions16.217.740.553.10.8350.72313.030.541.00.773
20 captions17.418.441.657.40.8390.72814.432.242.70.777
50 captions18.119.142.559.70.8400.72913.432.543.90.778
Other models trained without aligned MS-COCO images and captions:
ZeroCap [41]2.611.514.60.870.79
MAGIC [39]12.917.439.949.3
Flamingo [3]84.3
DeCap (560 captions) [22]51.4
DeCap (full train set) [22]24.725.091.2
CapDec (full train set) [27]26.425.151.891.8
", + "bbox": [ + 76, + 87, + 890, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison of decoding strategies for a captioning model trained only on minimally curated web data (JFT-5B and ALIGN) and evaluated on MS-COCO. At the bottom, we report metrics for other models trained without aligned MS-COCO images and captions. These models may not be directly comparable since they use different pretraining data. DeCap and CapDec use all 560K captions in the MS-COCO training set to train their decoders; we include CIDEr for DeCap with 560 captions (0.1% of the training data) for comparison.", + "bbox": [ + 75, + 364, + 892, + 421 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f48e3bb4252d02890f0913187077bc55a3e0f64d7f5d1598c9f29bba6a389f9c.jpg", + "image_caption": [ + "$\\gamma = 1.0$ : a photo of the small elephant", + "$\\gamma = 2.0$ : elephants in the", + "ruaha national park", + "$\\gamma = 3.0$ : elephants chobe", + "np", + "LM: a elephant and a", + "zebra in a field", + "OT: Two elephants.", + "G1: Two elephants standing on a grassy field next to a tree.", + "Figure 9. Examples of captions generated from a model pretrained only on minimally curated data, for randomly selected images without human faces. Captions labeled $\\gamma = 1.0$ are obtained without CFG; $\\gamma > 1$ uses CFG; LM indicates LM guidance with $\\alpha = \\beta = 1$ and 20 shots; GT indicates ground truth." + ], + "image_footnote": [], + "bbox": [ + 80, + 435, + 176, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19d46518ea0a222e081981c6bbd809b9d405903d5b122d2521587282c3a8265c.jpg", + "image_caption": [ + "$\\gamma = 1.0$ a photo of the", + "small coffee.", + "$\\alpha = {2.0}$ ; a coffee in a video", + "，", + "game:", + "1", + "video game.", + "LM: a banana with a", + "cup of coffee", + "G1: A close up of a banana next to a cup with liquid." + ], + "image_footnote": [], + "bbox": [ + 179, + 435, + 272, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c40f15b728b5e5805118f1fd22d4a71ff411ef01445d1c39d38b5c9471bdd159.jpg", + "image_caption": [ + "$\\gamma = 1.0$ a photo of the large", + "windsurfing.", + "$\\gamma = 3.0$ : windsurfing in tarifa", + "$x = 3.0$ : windsurfing walls", + "pors 1300x1034", + "1M", + "EM. a Win", + "water", + "GT: A man riding a", + "wind sail in the", + "ocean filled with", + "waves" + ], + "image_footnote": [], + "bbox": [ + 274, + 435, + 370, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4eff774343483a48497c359ed640f66e928da88211ab34e660cc43d94972ed83.jpg", + "image_caption": [ + "$\\gamma = 1.0$ a photo of the large", + "giraffe.", + "$\\alpha = 2.0$ a piraffe in a video", + "game", + "gamer:", + "1", + "photo 1", + "LM: a giraffe standing in a tall tower.", + "in a tall tree", + "GI: A giraffe in a dry", + "savannah with dry", + "shrubs" + ], + "image_footnote": [], + "bbox": [ + 372, + 435, + 467, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 693, + 194, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our study indicates that it is possible to substantially improve the extent to which generated captions uniquely describe the images they correspond to, raising questions regarding the goal of image captioning and how it should be evaluated. As it is conventionally formulated, image captioning aims not to provide text that can substitute for an image, but to write the text that a human annotator would have written. This formulation penalizes captions that are more descriptive than ground truth, even when a human might prefer them. On the other hand, treating image captioning as a problem of generating a caption that lies close to the image in the embedding space of an image-text model is", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also inadequate, because captions that lie close to the image need not be grammatical and may contain gibberish. Our proposed methods leveraging classifier-free guidance and language model guidance modulate the trade-offs between these two goals, as captured by various reference-based and reference-free metrics.", + "bbox": [ + 496, + 436, + 892, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are several possible extensions to our work. First, our present experiments use only greedy decoding. Although greedy decoding appears to perform reasonably well in our setup, it may be suboptimal for LM guidance with prompts that impose greater structure upon the captions. If the LM is prompted to output either \"there is a person in this image\" or \"there is no person this image\", greedy decoding is likely to fail even if the captioner properly scores the two possible captions, because when choosing between the tokens \"a\" and \"no\", the captioner has no knowledge of the structure that the LM will impose on future tokens. Since beam search could explore both tokens, it may offer better results in this scenario. Second, our method could be combined with RL-based methods to increase similarity in a contrastive embedding space, which may further improve retrieval performance and CLIPScore. Finally, with a perfect captioning model, $p(\\mathrm{image}|\\mathrm{caption})$ should increase with $\\gamma$ . However, in practice we find that $\\gamma > 3$ leads to a decrease in retrieval performance. This discrepancy suggests that the difference between the conditional and unconditional model distributions may be a noisy estimator of the pointwise mutual information. Although selecting $\\gamma$ is one way to regularize this estimator, there may also be strategies to regularize $p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x})$ at training time.", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15266", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Kevin Clark, David Fleet, Geoffrey Hinton, and the rest of the Google DeepMind Toronto team for inspiration, comments, and discussion.", + "bbox": [ + 76, + 114, + 468, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 172, + 174, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2", + "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2, 8", + "[4] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022. 5", + "[5] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. $ic^3$ : Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 2", + "[6] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with CLIP reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, Seattle, United States, July 2022. Association for Computational Linguistics. 2, 5", + "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. PaLM: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 4", + "[8] Katherine Crowson. You can apply a similar trick to classifier-free guidance to autoregressive transformers to sample from a synthetic \"super-conditioned\" distribution. https://twitter.com/RiversHaveWings/status/1478093658716966912, 2022. 3", + "[9] Béatrice Daille. Approche mixte pour l'extraction automatique de terminologie: statistiques lexicales et filtrres linguistiques. PhD thesis, Ph.D. thesis, Université Paris 7, 1994. 3", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language" + ], + "bbox": [ + 78, + 198, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. 3", + "[11] Jacob Devlin, Hao Cheng, Hao Fang, Saurabh Gupta, Li Deng, Xiaodong He, Geoffrey Zweig, and Margaret Mitchell. Language models for image captioning: The quirks and what works. arXiv preprint arXiv:1505.01809, 2015. 2", + "[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2, 3", + "[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics. 2, 4", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3", + "[15] Micah Hodosh, Peter Young, and Julia Hockenmaier. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853-899, 2013. 6", + "[16] Yupan Huang, Hongwei Xue, Bei Liu, and Yutong Lu. Unifying multimodal transformer for bi-directional image and text generation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1138-1147, 2021. 2, 5", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 4, 7", + "[18] Andrej Karpathy and Fei-Fei Li. Deep visual-semantic alignments for generating image descriptions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3128-3137. IEEE Computer Society, 2015. 4", + "[19] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander R Fabbri, Yejin Choi, and Noah A Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. arXiv preprint arXiv:2112.04139, 2021. 5", + "[20] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Morrison, Ronan Le Bras, Yejin Choi, and Noah A Smith. Transparent human evaluation for image captioning. arXiv preprint arXiv:2111.08940, 2021. 1, 2", + "[21] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011. 6", + "[22] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning. In International Conference on Learning Representations, 2022. 2, 8" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15267", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 1", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2", + "[25] Annika Lindh, Robert J Ross, Abhijit Mahalunkar, Giancarlo Salton, and John D Kelleher. Generating diverse and meaningful captions. In International Conference on Artificial Neural Networks, pages 176-187. Springer, 2018. 2", + "[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2", + "[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 2, 8", + "[28] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 1", + "[29] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 2", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4", + "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3", + "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2", + "[34] Noam Rotstein, David Bensaid, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models to fuse visual data into enriched image captions. arXiv preprint arXiv:2305.17718, 2023. 2", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems. 2", + "[36] Guillaume Sanchez, Honglu Fan, Alexander Spangher, Elad Levi, Pawan Sasanka Ammanamanchi, and Stella Biderman. Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806, 2023. 3", + "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 4, 6", + "[38] David R So, Wojciech Manke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V Le. Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668, 2021. 4", + "[39] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 2, 8", + "[40] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 4", + "[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 2, 5, 8", + "[42] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 1", + "[43] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015. 2", + "[44] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 2", + "[45] Shitong Xu. Clip-diffusion-lm: Apply diffusion model on image captioning. arXiv preprint arXiv:2210.04559, 2022. 2", + "[46] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 2", + "[47] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3, 4, 6, 14", + "[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15268", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3", + "[49] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104–12113, 2022. 4, 7", + "[50] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 5", + "[51] Youuyuan Zhang, Jiuniu Wang, Hao Wu, and Wenjia Xu. Distinctive image captioning via clip guided group optimization. arXiv preprint arXiv:2208.04254, 2022. 2, 5", + "[52] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang, Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint arXiv:2211.11694, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 375 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15269", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_model.json b/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b208fe7e4798e178cefe748617cc1c2236d7dff8 --- /dev/null +++ b/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.164, + 0.131, + 0.806, + 0.154 + ], + "angle": 0, + "content": "Guiding image captioning models toward more specific captions" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.18, + 0.309, + 0.198 + ], + "angle": 0, + "content": "Simon Kornblith" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.199, + 0.368, + 0.218 + ], + "angle": 0, + "content": "Google DeepMind" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.181, + 0.453, + 0.197 + ], + "angle": 0, + "content": "Lala Li" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.199, + 0.526, + 0.217 + ], + "angle": 0, + "content": "2Apple AI/ML" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.181, + 0.634, + 0.199 + ], + "angle": 0, + "content": "Zirui Wang\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.568, + 0.199, + 0.779, + 0.218 + ], + "angle": 0, + "content": "\\(^{3}\\)University of Washington" + }, + { + "type": "text", + "bbox": [ + 0.701, + 0.18, + 0.825, + 0.199 + ], + "angle": 0, + "content": "Thao Nguyen\\(^{3\\dagger}\\)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.251, + 0.313, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Image captioning is conventionally formulated as the task of generating captions for images that match the distribution of reference image-caption pairs. However, reference captions in standard captioning datasets are short and may not uniquely identify the images they describe. These problems are further exacerbated when models are trained directly on image-alt text pairs collected from the internet. In this work, we show that it is possible to generate more specific captions with minimal changes to the training process. We implement classifier-free guidance [14] for an autoregressive captioning model by fine-tuning it to estimate both conditional and unconditional distributions over captions. The guidance scale applied at decoding controls a trade-off between maximizing \\( p(\\text{caption} | \\text{image}) \\) and \\( p(\\text{image} | \\text{caption}) \\). Compared to standard greedy decoding, decoding with a guidance scale of 2 substantially improves reference-free metrics such as CLIPScore (0.808 vs. 0.775) and caption→image retrieval performance in the CLIP embedding space (recall@1 44.6% vs. 26.5%), but worsens standard reference-based captioning metrics (e.g., CIDEr 78.6 vs 126.1). We further explore the use of language models to guide the decoding process, obtaining small improvements over the Pareto frontier of reference-free vs. reference-based captioning metrics that arises from classifier-free guidance, and substantially improving the quality of captions generated from a model trained only on minimally curated web data." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.72, + 0.208, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Image captioning is both a difficult task for computer vision systems to perform and a difficult task to evaluate. Although automated captioning metrics rank the best captioning systems higher than humans, human raters still show a strong preference for human-generated captions [20], suggesting shortcomings in both captioning models and metrics. One shortcoming relates to the lack of specificity in generated captions. Conventional maximum likelihood-" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.251, + 0.697, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.397, + 0.687, + 0.404 + ], + "angle": 0, + "content": "\\(\\gamma = 1.0\\) a man riding a blue motorcycle on a dirt road" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.405, + 0.692, + 0.412 + ], + "angle": 0, + "content": "\\(\\gamma = 1.5\\) a man riding a blue motorcycle on a straw bale" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.412, + 0.692, + 0.425 + ], + "angle": 0, + "content": "\\(\\gamma = 2.0\\) rider on blue suzuki motorcycle near straw bales" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.426, + 0.692, + 0.439 + ], + "angle": 0, + "content": "\\(\\gamma = 3.0\\) racer on blue suzuki motorcycle leaning up against straw bales" + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.44, + 0.692, + 0.455 + ], + "angle": 0, + "content": "GT A person riding a baby blue motorcycle near haystacks" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.264, + 0.887, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.406, + 0.862, + 0.415 + ], + "angle": 0, + "content": "\\(y = 1.0\\) \\(y = 2.0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.416, + 0.861, + 0.425 + ], + "angle": 0, + "content": "\\(y = 1,2\\) \\(y = 3,0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.426, + 0.861, + 0.436 + ], + "angle": 0, + "content": "\\(\\gamma = 1.5\\) \\(\\gamma = 4.0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Figure 1. Using classifier-free guidance \\((\\gamma > 1)\\) results in more specific captions that are farther from the reference distribution. Left: Example of captions generated at different guidance scales for a single image. Right: Caption \\(\\rightarrow\\) image recall@1 with CLIP ViT-B/32 vs. CIDEr score, for captions generated with different guidance scales \\(\\gamma\\) on MS-COCO. Higher scales improve retrieval accuracy at the expense of CIDEr." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.732 + ], + "angle": 0, + "content": "based image captioning models attempt to generate captions such that the \\( p(\\text{caption} | \\text{image}) \\) is high. However, captions from the ground truth distribution are often nonspecific, e.g., human annotators will usually describe a German Shepard only as a dog. Moreover, previous work has emphasized \"reference-based\" captioning metrics that measure the match between generated captions and human-provided ground truth captions [28, 23, 42]. These metrics intrinsically penalize captions that are more specific than ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this work, we explore strategies to guide image captioning models to produce more specific captions by modifying the decoding distribution, and explore the tradeoffs in captioning metrics that result. We first investigate the application of classifier-free guidance (CFG) [14] to image captioning with autoregressive models. Classifier-free guidance increases \\( p(\\text{image|caption}) \\) at the expense of \\( p(\\text{caption|image}) \\). Although CFG hurts reference-based image captioning metrics such as BLEU [28], ROUGE [23], and CIDEr [42], it improves \"reference-free\" metrics that measure captions' specificity via the similarity between the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.282, + 0.888 + ], + "angle": 0, + "content": "*Work performed while at Google." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.37, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work performed as a student researcher at Google." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.37, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15259" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "image and the generated caption in the embedding space of image-text models [13] or caption \\(\\rightarrow\\) image retrieval performance. Qualitatively, we find that captions generated with CFG are more specific than both the ground truth captions and captions generated without CFG, but they are less grammatical, particularly at high CFG scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.471, + 0.41 + ], + "angle": 0, + "content": "Beyond classifier-free guidance, we experiment with guiding image captioning models using the probability distribution obtained from a few shot-prompted language model (LM). We find that using a language model to guide a captioning model trained on MS-COCO [24] with descriptive manually written captions can allow it to achieve slightly better trade-offs between reference-free vs. reference-based captioning metrics than those observed with CFG. LM guidance also substantially improves the captions produced by a model trained exclusively on minimally curated web data. Although this model achieves a CIDEr score of only 21.8 without guidance, this CIDEr score improves to 57.4 when guided by a language model prompted with 20 captions from the MS-COCO training set. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.416, + 0.468, + 0.46 + ], + "angle": 0, + "content": "- We propose two strategies to guide image captioning models to produce more specific captions: classifier-free guidance and language model guidance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.465, + 0.468, + 0.525 + ], + "angle": 0, + "content": "- We demonstrate that classifier-free guidance yields captions that are closer to the corresponding image in the embedding space of image-text models, but are farther from human-provided reference captions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.529, + 0.468, + 0.604 + ], + "angle": 0, + "content": "- We show that language model guidance can alter caption styles, substantially improving captions produced by a model trained only on minimal curated web data and marginally improving the trade-off between captioning metrics observed with classifier-free guidance." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.416, + 0.468, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.618, + 0.216, + 0.634 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Measuring specificity of captions. Early work using neural networks for image captioning found that models have a propensity to regurgitate captions from their training data, and as a result, the generated captions are not descriptive enough to uniquely identify images [43, 11]. To address this shortcoming, Lindh et al. [25] proposed to use caption \\(\\rightarrow\\) image recall with an image retrieval model to examine whether images can be retrieved from generated captions, and further attempt to differentiate through this retrieval process to train a captioning model. Their approach marginally improves retrieval accuracy, but worsens reference-based captioning metrics. More recent work has adopted approaches to evaluate the specificity of captions based on the CLIP image-text model [30]. Hessel et al. [13] propose CLIPScore, an image captioning metric based on the cosine similarity between CLIP embeddings of the image and the generated caption. Kasai et al. [20] re" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "port that CLIPScore-based metrics align better with human judgments compared to reference-based captioning metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Improving specificity of captions. Recent work has attempted to train caption models to directly maximize the similarity of captions with corresponding images in the CLIP embedding space using reinforcement learning. This work has found that adding such losses worsens standard reference-based captioning metrics but improve similarity and retrieval in the CLIP embedding space [16, 6, 51], similar to our observations regarding CFG. Wen et al. [44] attempt to generate prompts for text-to-image generative models that correspond to specific images without a captioning model, by directly optimizing the similarity between the text and image in the CLIP embedding space using a gradient-based discrete optimization procedure, but the resulting text is not grammatical." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.549 + ], + "angle": 0, + "content": "Other work has attempted to generate more descriptive captions through different means. Dense captioning [46] aims to detect and caption all objects in an image, but concatenating all of these captions leads to long and unnatural captions, whereas CFG produces single-sentence captions. The Localized Narratives dataset [29] contains visually grounded captions for MS-COCO images collected through voice annotation. These captions are substantially more descriptive than the captions in the MS-COCO dataset and can be used for model training. Two recent papers propose to combine multiple captions from a single model [5] or outputs of different vision models [34], and then combine them using a language model. The resulting captions are longer, but achieve greater caption→image recall." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Captioning from uncurated data. In Section 4.2, we explore the use of LM guidance for captioning with access to uncurated image-text data from the web and a small number of captions but not images from the target distribution. This setting, which does not rely on aligned images and captions from the target distribution, is often referred to as \"zero-shot\" captioning, and previous work has pursued a number of alternative approaches. Flamingo [3] and CM3 [1] perform zero-shot captioning by pretraining on interleaved image/text data. MAGIC [39] and ZeroCap [41] generate captions using a combination of guidance from CLIP and a large language model. Other recent work adapts CLIP to perform captioning by training a text decoder using only captions, with no corresponding images [27, 22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Classifier-free guidance. CFG is widely used in diffusion-based and autoregressive text-to-image models [26, 32, 35, 33, 12, 48]. Because of the popularity of the combination of CFG and diffusion, previous work that has performed image captioning with diffusion models has also examined the use of CFG. This work finds either no benefit to using CFG [45] or a small and inconsistent benefit that appears to vary with minor changes in training settings [52]. However, these studies do not seek to generate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15260" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "more specific captions, and thus measure only reference-based captioning metrics, which we likewise find do not benefit from CFG. Concurrently with out work, [36] propose to use classifier-free guidance to improve prompt following in large language models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.181, + 0.176, + 0.197 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.206, + 0.457, + 0.223 + ], + "angle": 0, + "content": "3.1. Classifier-free guidance for image captioning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.468, + 0.29 + ], + "angle": 0, + "content": "Let \\( \\mathbf{x} \\) be an image caption and \\( \\mathbf{y} \\) be the corresponding image. A standard captioning model aims to model the likelihood \\( p(\\mathrm{x}|\\mathrm{y}) \\), factorized autoregressively in terms of the probability of each token given previous tokens" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.294, + 0.469, + 0.311 + ], + "angle": 0, + "content": "\\[\np (\\mathrm {x} | \\mathrm {y}) = p \\left(\\mathrm {x} _ {n} \\mid \\mathrm {x} _ {n - 1}, \\dots \\mathrm {x} _ {1}, \\mathrm {y}\\right) \\dots p \\left(\\mathrm {x} _ {1} \\mid \\mathrm {y}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.468, + 0.421 + ], + "angle": 0, + "content": "The network is trained so that its output distribution \\( q_{\\theta}(\\mathbf{x}_n|\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y}) \\stackrel{\\mathrm{def}}{=} \\text{softmax}(f_{\\theta}(\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y})) \\) approximates \\( p(\\mathbf{x}_n|\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y}) \\). At inference time, one typically uses beam search or greedy decoding to produce a caption that has a particularly high probability. In this work, we use greedy decoding because it is the more common choice and it is also simpler to implement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.468, + 0.452 + ], + "angle": 0, + "content": "Classifier-free guidance (CFG) [14] modifies the inference procedure so it instead aims to achieve high values of" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.456, + 0.469, + 0.49 + ], + "angle": 0, + "content": "\\[\nl _ {\\boldsymbol {\\theta}, \\gamma} (\\mathrm {x}, \\mathrm {y}) \\stackrel {\\text {d e f}} {=} p (\\mathrm {x}) \\left(\\frac {p (\\mathrm {x} | \\mathrm {y})}{p (\\mathrm {x})}\\right) ^ {\\gamma} \\propto p (\\mathrm {x}) p (\\mathrm {y} | \\mathrm {x}) ^ {\\gamma}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.598 + ], + "angle": 0, + "content": "where proportionality holds because \\( p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x}) = p(\\mathrm{y}|\\mathrm{x}) / p(\\mathrm{y}) \\) and \\( p(\\mathrm{y}) \\) is fixed. The parameter \\( \\gamma \\) is called the guidance scale and controls the trade-off between maximization of \\( p(\\mathrm{x}|\\mathrm{y}) \\) and \\( p(\\mathrm{y}|\\mathrm{x}) \\). When \\( \\gamma = 1 \\), \\( l_{\\theta, \\gamma}(\\mathrm{x}, \\mathrm{y}) = p(\\mathrm{x}|\\mathrm{y}) \\) and guidance has no effect. Setting \\( \\gamma > 1 \\) inflates the probability of the image given the caption \\( p(\\mathrm{y}|\\mathrm{x}) \\) relative to the unconditional probability of the caption \\( p(\\mathrm{x}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.705 + ], + "angle": 0, + "content": "Ho and Salimans [14] originally proposed CFG in the context of diffusion models, which estimate the score functions \\(\\nabla \\log p(\\mathbf{x}|\\mathbf{y})\\) and \\(\\nabla \\log p(\\mathbf{x})\\). Although \\(l_{\\theta,\\gamma}(\\mathbf{x},\\mathbf{y})\\) factorizes autoregressively, it is not a normalized probability distribution, so it is not entirely clear how one should sample tokens when performing autoregressive generation. Crowson [8] suggested to sample from" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.709, + 0.468, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {q} _ {\\boldsymbol {\\theta}, \\gamma} (\\mathrm {x} _ {n} | \\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathrm {y}) \\stackrel {\\mathrm {d e f}} {=} \\operatorname {s o f t m a x} (f _ {\\boldsymbol {\\theta}} (\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathbf {0}) \\\\ + \\gamma \\left(f _ {\\boldsymbol {\\theta}} \\left(\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathrm {y}\\right) - f _ {\\boldsymbol {\\theta}} \\left(\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathbf {0}\\right)\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where \\(f_{\\theta}(\\mathbf{x}_{n - 1},\\ldots ,\\mathbf{x}_1,\\mathbf{0})\\) are logits generated by the model without conditioning, usually by passing zeros in place of the conditioning information. This formulation has been successfully applied in autoregressive image models [12, 48]. In our experiments, since we decode greedily, i.e., at each step we take the token that maximizes \\(\\tilde{q}_{\\theta ,\\gamma}(\\mathrm{x}_n|\\mathrm{x}_{n - 1},\\dots ,\\mathrm{x}_1,\\mathrm{y})\\) and thus \\(l_{\\theta ,\\gamma}(\\mathrm{x},\\mathrm{y})\\), any form of normalization of the per-step sampling distribution would produce the same captions. We provide pseudocode in Appendix A.1. Crucially, the only change required to train" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "ing is to learn an unconditional model of captions alongside the image-conditional model, which we accomplish by randomly masking the image embeddings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.15, + 0.738, + 0.166 + ], + "angle": 0, + "content": "3.2. Language model guidance" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.174, + 0.892, + 0.204 + ], + "angle": 0, + "content": "Inspired by classifier-free guidance, we consider language model (LM) guidance, which attempts to maximize" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.208, + 0.892, + 0.241 + ], + "angle": 0, + "content": "\\[\nl _ {\\boldsymbol {\\theta}, \\gamma} ^ {\\prime} (\\mathrm {x}, \\mathrm {y}) \\stackrel {\\text {d e f}} {=} q (\\mathrm {x}) \\left(\\frac {p (\\mathrm {x} | \\mathrm {y}) ^ {\\alpha}}{p (\\mathrm {x}) ^ {\\beta}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.456 + ], + "angle": 0, + "content": "where \\( p(x) \\) and \\( p(x|y) \\) are obtained from a captioning model as in CFG but \\( q(\\mathbf{x}) \\) is obtained from a language model that was trained independently (but with the same vocabulary) on a large text corpus. The quantity \\( p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x}) = p(\\mathrm{x},\\mathrm{y}) / (p(\\mathrm{x})p(\\mathrm{y})) \\) measures the strength of the association between a caption and an image; its logarithm is the pointwise mutual information (PMI). LM guidance relies on the assumption that, even for large shifts in the prior distribution of captions \\( p(\\mathbf{x}) \\), the shift in PMI will be small. Empirically, we obtain better results by allowing different exponents for the numerator and denominator, with \\( \\alpha > \\beta \\). This decoupling resembles \\( \\mathrm{PMI}^k \\) [9], which reduces the bias of PMI toward rare associations. We provide a more detailed derivation in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.699 + ], + "angle": 0, + "content": "We investigate two applications of LM guidance. First, we combine a captioning model fine-tuned on MS-COCO with a LM prompted with manually written descriptive captions to alter the style of the captions the model produces. The manually written prompts are shown in Appendix A.4. Second, we combine a captioning model trained only on low-quality web data with a LM prompted with varying numbers of examples from the MS-COCO training set to evaluate the ability of LM guidance to elicit higher-quality captions without high-quality paired data. We randomly select a different set of captions for each minibatch of four test examples. In both cases, we separate the captions with two newlines. Because this format leads the LM to place probability mass on the newline token to end the caption, we transfer the probability mass from the newline token to the EOS token. See Appendix A.3 for pseudocode." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.694, + 0.727 + ], + "angle": 0, + "content": "3.3. Models and training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our captioning model is a \"bottleneck\" variant of CoCa-Base [47], which combines a contrastive loss with a captioning loss to simultaneously learn aligned image and text embeddings as well as a captioner. The architecture consists of an image encoder, a unimodal text decoder, and a multimodal text decoder, each of which are Transformers with 12 layers, 768 hidden dimensions, an MLP of size 3072, and 12 self-attention heads, matching \\(\\mathrm{BERT}_{\\mathrm{BASE}}\\) [10] and GPT-1 [31]. The image encoder is a ViT-B/18 that processes \\(288 \\times 288\\) input and produces an embedding such that images are embedded close to their corresponding text." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15261" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "CoCa's multimodal text decoder processes the representations of the image encoder to produce a caption. Whereas [47] conditions the multimodal text decoder using crossattention to pooled representations, our bottleneck variant uses only the contrastive image embedding. Appendix A.5 shows a diagram of the resulting architecture. We adopt this bottleneck variant because of its simplicity and the conceptual appeal: When CFG is used, the captioner's role is to invert the image embedding, providing a caption that, when embedded by the text encoder, lies close to it. However, as we show in Appendix B.1, this choice of the bottleneck model is not critical, and CFG is equally effective with the standard CoCa architecture with attention pooling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.606 + ], + "angle": 0, + "content": "For CFG experiments, we pretrain our model on an image-text dataset comprising images from the JFT-5B dataset [40, 49] paired with their corresponding label names substituted into a randomly selected prompt from the list provided by Radford et al. [30], web images paired with noisy alt text from the ALIGNN dataset [17], and a small amount of data from other sources. We do not mask conditioning information during pretraining. We then fine-tune on the combined MS-COCO train and Karpathy validation splits [18] using Adam with batch size 128. We linearly warm up to a learning rate of \\(1 \\times 10^{-5}\\) over the first 1,000 steps and linearly decay to zero over the rest of training. We vary \\(\\gamma \\in \\{1.0, 1.2, 1.5, 2.0, 3.0, 4.0\\}\\), conditioning masking proportion in \\(\\{0.0, 0.25, 0.5, 0.75\\}\\), and numbers of steps in \\(\\{5,000, 10,000, 20,000, 50,000\\}\\). We report results from the model trained for 20,000 steps with masking proportion 0.5, which achieves near-optimal results, in Tables 1 and B.4, and sample example captions from it. To ensure that results generalize across datasets, we also experiment with a model fine-tuned on Conceptual Captions [37] for 100,000 steps with masking proportion 0.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.473, + 0.789 + ], + "angle": 0, + "content": "For LM guidance experiments, we pretrain on the JFT-5B and ALIGN datasets, again following the recipe of [47]. For zero-shot captioning experiments, we fine-tune this model on the same datasets for an additional 50,000 steps with conditioning masking proportion of 0.5 to improve our ability to sample unconditionally. For LM guidance on MS-COCO, we first fine-tune on ALIGN, JFT-5B images backcaptioned by an MS-COCO fine-tuned CoCa2B model, and a small amount of internal data before finetuning on MS-COCO. Our language model is a variant of Primer [38] with 2 billion parameters, trained on a similar dataset to that used to train PaLM [7]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.799, + 0.2, + 0.813 + ], + "angle": 0, + "content": "3.4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.47, + 0.853 + ], + "angle": 0, + "content": "We adopt the standard reference-based captioning metrics BLEU-4, METEOR, ROUGE, and CIDEr, as well" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.365 + ], + "angle": 0, + "content": "as reference-free captioning metrics based on CLIP ViT-B/32 [30]. The first reference-free captioning metric is CLIPScore [13], which is defined as \\(\\mathrm{CLIP - S}(c,v) = 2.5\\cdot \\max (\\cos (c,v),0)\\) where \\(c\\) and \\(v\\) are the CLIP embeddings of the caption and image respectively. The second reference-free metric measures the accuracy with which we can retrieve an image from the generated caption within a given test split by taking the \\(k\\) nearest neighbors of the caption in the CLIP embedding space. Because recall@ \\(k\\) for \\(k > 1\\) is highly correlated with recall@1 (R@5: \\(r = 0.99\\), R@10: \\(r = 0.98\\)), we plot only recall@1. We additionally report RefOnlyCLIP-S, a reference-based metric that uses the CLIP text encoder to compute the similarity of CLIP embeddings of the generated captions with embeddings of ground truth captions, and RefCLIP-S, which takes the average of the per-image harmonic means of CLIP-S and RefOnlyCLIP-S [13]. Unless otherwise stated, all evaluation is performed on the MS-COCO Karpathy test split [18]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.38, + 0.587, + 0.396 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.406, + 0.719, + 0.422 + ], + "angle": 0, + "content": "4.1. Classifier-free guidance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.612 + ], + "angle": 0, + "content": "We first investigate the trade-off between reference-based and reference-free image captioning metrics as a function of guidance scale. Because different guidance scales and metrics could conceivably benefit from different fine-tuning hyperparameter combinations, we plot all results from our hyperparameter grid in Figure 2. Although standard greedy decoding \\((\\gamma = 1.0)\\) produces the highest CIDEr, METEOR, ROUGE, and BLEU-4 scores, higher guidance weights consistently yield higher values of reference-free captioning metrics. In particular, \\(\\gamma = 3.0\\) offers both the best caption→image recall and the best CLIP-Score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.896, + 0.764 + ], + "angle": 0, + "content": "Table 1 compares our results, obtained from a single model evaluated at different guidance scales, with previous work that reports either CLIPScore or CLIP ViT-B/32 caption \\(\\rightarrow\\) image retrieval performance. Although our model is trained with standard cross-entropy loss rather than a CLIP-based loss and our pretraining dataset is distinct from CLIP's, sampling from our model with CFG yields higher CLIPScores than all other models trained without CLIP-based losses, and better CLIP caption \\(\\rightarrow\\) image retrieval even when compared with models that use CLIP-based losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "We present examples of captions generated at different CFG scales in Figure 3. Higher CFG strengths lead to more descriptive captions. At \\(\\gamma = 1.0\\), the central object in the top left image is described as a \"car\" as in the ground truth caption, whereas at \\(\\gamma > 1.0\\) it is a \"station wagon.\" Similarly, at low CFG strengths, the birds in the center image are described simply as \"birds,\" whereas at \\(\\gamma = 2.0\\) they become \"crested cranes.\" However, at \\(\\gamma = 3.0\\), captions clearly become less grammatical, containing repeated" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1We find that passing an all-zero image embedding to the pretrained model yields samples that resemble the unconditional distribution, suggesting that it implicitly learns to model the unconditional distribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15262" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.15, + 0.087, + 0.818, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.306, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Figure 2. Classifier-free guidance controls a trade off between reference-free and reference-based captioning metrics. Each point reflects a model trained with a different hyperparameter combination; each color represents a \\(\\gamma\\) value used to decode. Models are evaluated with different guidance scales \\(\\gamma\\), using reference-free captioning metrics based on CLIP ViT-B/32 (y-axes; top: CLIPScore, bottom: recall@1) and reference-based captioning metrics (x-axes). The dashed line reflects the value of the reference-free captioning metric for the ground-truth captions obtained from MS-COCO." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.385, + 0.887, + 0.61 + ], + "angle": 0, + "content": "
ModelReference-Based MetricsReference-Free MetricsRefCLIP-S
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10
Models trained with CLIP features or losses:
CLIP-Captioner [4]38.729.358.6126.00.8110.7540.814
UMT-BITG [16]37.328.257.9122.60.772
X-LAN+SCST+GEG [51]36.528.757.5121.728.150.367.2
CIDEr + CLIP-S Reward [6]37.728.858.3124.60.77224.450.263.1
CLIP-S Reward [6]6.218.731.611.20.86042.571.682.2
ZeroCap [41]2.611.514.60.870.79
Models trained without access to CLIP:
UMT-BITG w/o CLIP loss [16]37.628.358.1122.50.725
VinVL-large [50]41.030.959.4*140.90.91*0.78*0.84*
Ours (γ = 1.0)36.130.558.2126.10.9000.77526.551.964.10.830
Ours (γ = 1.2)35.130.057.5124.10.8990.78531.357.469.30.835
Ours (γ = 1.5)31.528.454.4113.20.8910.79636.664.075.00.838
Ours (γ = 2.0)20.923.343.078.60.8620.80844.671.781.70.831
Ours (γ = 3.0)11.517.129.441.70.8200.80849.475.784.70.811
Ours (γ = 4.0)6.512.318.417.30.7660.78244.771.380.90.771
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.619, + 0.89, + 0.647 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of our approach with results from previous work that reports CLIP-based metrics. For VinVL-large, * indicates metrics from [19]." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.655, + 0.88, + 0.83 + ], + "angle": 0, + "content": "
γ=1.0: a car with a surfboard on top of it parked next to a car\nγ=1.5: a vintage station wagon with a surfboard on top\nγ=2.0: a antique station wagons and a car buick station wagon\nγ=3.0: buick woody wyddo staitonwagon and surf green station wagon parked in front of car show establishment\nGT: An old green and brownγ=1.0: a knife is sitting on a cutting board next to a sliced carrot\nγ=1.5: a knife is sitting on a cutting board next to an orange\nγ=2.0: knife sitting on cutting board next to whole one\nγ=3.0: knife sitting on cutting board next to misschappi carrot on cutting board\nGT: A knife sticking out of the side of a block of cheese.γ=1.0: a herd of sheep grazing on a grass covered road\nγ=1.5: sheep grazing on a high-way with a truck on the road\nγ=2.0: sheep graze on freeway medians where grass is grown\nγ=3.0: grazing trucks blocking sheep on roadway grazing grass\nGT: A large herd of sheep are grazing by the busy road.γ=1.0: two birds standing in a cage with their heads in the air\nγ=1.5: two birds standing inside of a cage in a zoo\nγ=2.0: two crested cranes inside a wire cage\nγ=3.0: crested tantalus cranes caged together in bird-cage enclosure\nGT: Two birds who are looking out of the cage they are in.γ=1.0: a view of a city with a clock tower in the background\nγ=1.5: a city with steeples and trees and buildings\nγ=2.0: spires of churches line a city skyline\nγ=3.0: spires steeples buildings trees church spires and trees\nGT: A clock that is on the side of a tower.γ=1.0: a kitchen with a microwave and refrigerator\nγ=1.5: a kitchen with a mi-crowage and a refrigerator\nγ=2.0: a kitchen with red ap-pliances and white cup-boards\nγ=3.0: appliances sit in a small empty dingroom red and white kitchen\nGT: A kitchen that has a tile floor, a refrigerator, a microwave, and a toaster.γ=1.0: a bathroom with a large mirror and a bathtub\nγ=1.5: a bathroom with a large mirror and a bathtub\nγ=2.0: a spacious bathroom with a large mirror and a large tub\nγ=3.0: spacious bathroom with chandelier over tub mir-rors and tv\nGT: A bathroom with a tub, sinks, lights and a televi-sion.
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.848, + 0.895, + 0.892 + ], + "angle": 0, + "content": "Figure 3. Caption descriptiveness increases with CFG strength, but high CFG strengths produce agrammatical captions. Here we show examples of captions generated with different classifier-free guidance scales, for randomly selected images without human faces from the MS-COCO Karpathy test split. Captions labeled \\(\\gamma = 1.0\\) are obtained without CFG; \\(\\gamma > 1\\) uses CFG; GT = ground truth." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15263" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.198 + ], + "angle": 0, + "content": "words (\"woody woody\") and nonsense words (\"misshappi\", \"dingroomy\"). Figure 4 shows captions obtained with and without CFG next to the top 5 closest images in the embedding space of CoCa 2B [47],² where it is clear that CFG adds details to captions that help to distinguish them from other captions in the test split. We provide additional examples in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.474, + 0.514 + ], + "angle": 0, + "content": "To provide additional quantitative assessments of the specificity of elicited captions, we perform two additional evaluations, described further in Appendix B.2. First, we generate captions for the Stanford Dogs [21] test set, which consists of 8,580 images in total of 120 breeds of dogs, and examine their properties. Without guidance, only \\(1.9\\%\\) of captions contain one of the 120 breed names, whereas at \\(\\gamma = 2.0\\), \\(42.4\\%\\) do. The percentage of these breed names that are correct changes little, from \\(61.7\\%\\) without guidance to \\(58.5\\%\\) at \\(\\gamma = 2.0\\). Second, we performed a human evaluation comparing captions of MS-COCO test set images obtained without guidance and at \\(\\gamma = 2.0\\). We asked subjects to select the caption that is \"better\" and \"more descriptive\" or to indicate that they are both equal. When we asked these questions separately, we found that the two sets of captions are statistically indistinguishable. However, when asking both questions on the same survey, we found that captions generated without guidance are slightly \"better\" (50.5% vs. \\(46.6\\%\\), \\(p = 0.006\\), binomial test) but captions generated at \\(\\gamma = 2.0\\) are \"more descriptive\" (52.7% vs. \\(45.8\\%\\), \\(p = 1 \\times 10^{-6}\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.474, + 0.818 + ], + "angle": 0, + "content": "To validate the reliability of our results, we further measure the impact of CFG on three additional datasets, nocaps [2], Flickr-8k [15], and Conceptual Captions (CC3M) [37], as well as with alternative retrieval models. nocaps is a test set for captioning models with objects not present in MS-COCO; Flickr-8k is a small captioning dataset collected using a different procedure than MS-COCO; and Conceptual Captions is a set of 3.3M captions collected from filtered alt-text. We fine-tune the bottleneck CoCa-Base model directly on CC3M, and use our model fine-tuned on MS-COCO to caption images on nocaps and Flickr-8K. As shown in Figure 5, we find trade-offs between reference-based and reference-free captioning metrics similar to those above. In Appendix B.3, we report reference-free captioning metrics on MS-COCO computed with two additional retrieval models: the pretrained CoCa 2B model from [47] and the fine-tuned CoCa Base model that we use to generate captions. With both models, CFG substantially increases recall, in line with results obtained with CLIP ViT-B/32." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.473, + 0.866 + ], + "angle": 0, + "content": "Although CFG produces captions that are more successful at uniquely identifying images than decoding from the conditional distribution, caption lengths are similar for" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.895, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.633 + ], + "angle": 0, + "content": "Figure 4. Captions generated with CFG contain specific details that improve retrieval. For each reference image (far left), we show captions at \\(\\gamma = 1.0\\) (no guidance) and \\(\\gamma = 2.0\\). To the right, we show the closest images to each caption in the CoCa embedding space. Reference images are selected at random subject to the constraints that the closest image differs between \\(\\gamma\\) values and there are no identifiable human faces." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.641, + 0.895, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.872, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Figure 5. CFG also yields trade-offs between captioning metrics on nocaps, Flickr-8K, and CC3M." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.473, + 0.903 + ], + "angle": 0, + "content": "2We use CoCa 2B rather than CLIP because, quantitatively and qualitatively, it provides better retrieval results both with and without guidance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15264" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.161, + 0.089, + 0.388, + 0.179 + ], + "angle": 0, + "content": "
γWordsCharacters
1.09.6 ± 1.444.2 ± 7.2
1.29.6 ± 1.444.7 ± 7.4
1.59.4 ± 1.445.7 ± 7.8
2.09.3 ± 2.450.3 ± 18.6
3.010.7 ± 7.669.0 ± 56.1
4.019.9 ± 16.9161.2 ± 140.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.189, + 0.47, + 0.23 + ], + "angle": 0, + "content": "Table 2. Moderate CFG scales do not substantially change caption lengths, although higher CFG scales result in longer captions. Numbers are mean \\(\\pm\\) standard deviation." + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.242, + 0.418, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.414 + ], + "angle": 0, + "content": "Figure 6. Language model guidance produces captions that slightly exceed the Pareto frontier of CIDEr vs. caption \\(\\rightarrow\\) image retrieval accuracy on MS-COCO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.522 + ], + "angle": 0, + "content": "\\(\\gamma \\in [1,2]\\), as shown in Table 2. Thus, at low guidance strengths, CFG improves recall by making more efficient use of words, rather than by producing more verbose captions. Higher CFG strengths lead to longer captions but, as described above, these captions are agrammatical and contain nonsense words." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.315, + 0.547 + ], + "angle": 0, + "content": "4.2. Language model guidance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.78 + ], + "angle": 0, + "content": "We first experiment with guiding a captioning model fine-tuned on MS-COCO to produce more descriptive captions using a language model prompted with manually written prompts. We first manually wrote a prompt containing 10 descriptive captions of COCO test set images (Appendix A.4). We then sweep over \\(\\alpha \\in \\{1,2,3,4,5,6,7,8,9,10,12,15\\}\\) and \\(\\beta \\in \\{0,\\alpha /4,\\alpha /2,3 / 4\\alpha ,\\alpha \\}\\), and compare the resulting retrieval/CIDEr trade-off to that produced by the same model with CFG. We observe that it is possible to obtain small improvements upon the Pareto frontier provided by CFG, as shown in Figure 6. With \\(\\alpha = 5\\), \\(\\beta = -5 / 2\\), LM guidance achieves CLIP ViT-B/32 R@1 of \\(39.6\\%\\) and CIDEr of 114.4, whereas CFG with \\(\\gamma = 1.6\\) is worse on both metrics, achieving R@1 of \\(39.0\\%\\) and CIDEr of 109.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.871 + ], + "angle": 0, + "content": "We further experiment with using prompting to control the captioner using a manually written prompt of 25 captions in the form of \"a photo of NUMBER OBJECTS\" (e.g., \"a photo of eight apples\"; see Appendix A.4). With \\(\\alpha = \\beta = 1\\), the guided model is able to match this format and counts the number of objects in images (Figure 7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We next investigate whether language model guidance can elicit better captions from a model trained only on" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.601, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.163, + 0.587, + 0.171 + ], + "angle": 0, + "content": "a photo of two dogs" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.089, + 0.701, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.163, + 0.685, + 0.171 + ], + "angle": 0, + "content": "a photo of one bird" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.089, + 0.801, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.163, + 0.793, + 0.171 + ], + "angle": 0, + "content": "a photo of four suitcases" + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.089, + 0.899, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.163, + 0.886, + 0.171 + ], + "angle": 0, + "content": "a photo of five sheep" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.177, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Figure 7. Captions generated with LM guidance with a prompt of 25 captions in the form of \"a photo of NUMBER OBJECTS\". Examples are selected to show different numbers of objects." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.242, + 0.689, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.242, + 0.882, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.372, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Figure 8. LM guidance substantially improves CIDEr and retrieval scores of a model trained solely on minimally curated web data and evaluated on MS-COCO. The x-axis shows the number of captions used to prompt the LM; we do not prompt with images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.61 + ], + "angle": 0, + "content": "low-quality data. Here, we use a CoCa model that is pretrained on image-alt text pairs from the web (the ALIGN dataset [17]) and classification labels converted to text (the JFT-5B dataset [49]), without any additional fine-tuning. Because the data distribution places higher probability mass on short, non-descriptive captions than on longer captions, the resulting model is of limited utility for captioning, and would generally need to be fine-tuned on another dataset such as MS-COCO before being applied to a captioning task. Rather than fine-tune, we use LM guidance to prompt the model with captions from the MS-COCO training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.893, + 0.793 + ], + "angle": 0, + "content": "LM guidance substantially improves the quality of the captions produced by the original pretrained CoCa model without any clean parallel data. With LM guidance, we achieve CIDEr scores of 48.6 with 5 shots and 59.7 with 50 shots, far exceeding the CIDEr score of 21.8 obtained with no guidance. Figure 8 shows CIDEr and CLIP recall@1 scores for LM guidance of this pretrained CoCa model as a function of the number of shots, with \\(\\alpha = \\beta = 1\\). Table 3 compares classifier-free guidance and LM guidance. CFG yields higher CLIP-Scores and retrieval accuracy than LM guidance with \\(\\alpha = \\beta = 1\\), but LM guidance provides much higher CIDEr scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare captions generated with CFG to those generated with LM guidance for four images in Figure 9. In general, CFG produces agrammatical captions, whereas LM guidance produces grammatical captions but hallucinates details. For example, the image in the upper left shows two elephants and no zebras, but LM guidance leads to the caption \"an elephant and a zebra in a field.\"" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15265" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.088, + 0.891, + 0.362 + ], + "angle": 0, + "content": "
ModelReference-Based MetricsReference-Free Metrics
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10RefCLIP-S
Classifier-free guidance:
γ = 1.08.28.321.821.80.7660.6949.019.526.10.725
γ = 1.28.69.524.525.00.7810.71812.727.235.10.745
γ = 1.58.910.025.625.20.7800.72816.733.843.00.750
γ = 2.08.19.723.922.90.7770.74121.240.851.10.755
γ = 3.07.18.720.018.50.7670.75325.847.858.30.756
γ = 4.06.47.516.313.90.7490.74327.348.558.10.742
Language model guidance with α = β = 1:
2 captions12.714.634.739.30.8060.68810.023.732.40.740
5 captions15.016.639.148.60.8270.71212.427.537.50.763
10 captions16.217.740.553.10.8350.72313.030.541.00.773
20 captions17.418.441.657.40.8390.72814.432.242.70.777
50 captions18.119.142.559.70.8400.72913.432.543.90.778
Other models trained without aligned MS-COCO images and captions:
ZeroCap [41]2.611.514.60.870.79
MAGIC [39]12.917.439.949.3
Flamingo [3]84.3
DeCap (560 captions) [22]51.4
DeCap (full train set) [22]24.725.091.2
CapDec (full train set) [27]26.425.151.891.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.365, + 0.893, + 0.422 + ], + "angle": 0, + "content": "Table 3. Comparison of decoding strategies for a captioning model trained only on minimally curated web data (JFT-5B and ALIGN) and evaluated on MS-COCO. At the bottom, we report metrics for other models trained without aligned MS-COCO images and captions. These models may not be directly comparable since they use different pretraining data. DeCap and CapDec use all 560K captions in the MS-COCO training set to train their decoders; we include CIDEr for DeCap with 560 captions (0.1% of the training data) for comparison." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.436, + 0.177, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.511, + 0.157, + 0.526 + ], + "angle": 0, + "content": "\\(\\gamma = 1.0\\) : a photo of the small elephant" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.526, + 0.174, + 0.532 + ], + "angle": 0, + "content": "\\(\\gamma = 2.0\\) : elephants in the" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.532, + 0.175, + 0.539 + ], + "angle": 0, + "content": "ruaha national park" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.54, + 0.166, + 0.547 + ], + "angle": 0, + "content": "\\(\\gamma = 3.0\\) : elephants chobe" + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.547, + 0.122, + 0.553 + ], + "angle": 0, + "content": "np" + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.554, + 0.167, + 0.56 + ], + "angle": 0, + "content": "LM: a elephant and a" + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.56, + 0.166, + 0.566 + ], + "angle": 0, + "content": "zebra in a field" + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.567, + 0.166, + 0.572 + ], + "angle": 0, + "content": "OT: Two elephants." + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.573, + 0.172, + 0.597 + ], + "angle": 0, + "content": "G1: Two elephants standing on a grassy field next to a tree." + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.436, + 0.273, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.511, + 0.253, + 0.517 + ], + "angle": 0, + "content": "\\(\\gamma = 1.0\\) a photo of the" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.518, + 0.249, + 0.524 + ], + "angle": 0, + "content": "small coffee." + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.524, + 0.268, + 0.53 + ], + "angle": 0, + "content": "\\( \\alpha = {2.0} \\) ; a coffee in a video" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.531, + 0.204, + 0.537 + ], + "angle": 0, + "content": "，" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.538, + 0.254, + 0.544 + ], + "angle": 0, + "content": "game:" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.545, + 0.253, + 0.55 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.551, + 0.261, + 0.557 + ], + "angle": 0, + "content": "video game." + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.558, + 0.261, + 0.564 + ], + "angle": 0, + "content": "LM: a banana with a" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.565, + 0.262, + 0.57 + ], + "angle": 0, + "content": "cup of coffee" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.571, + 0.263, + 0.591 + ], + "angle": 0, + "content": "G1: A close up of a banana next to a cup with liquid." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.436, + 0.371, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.51, + 0.37, + 0.517 + ], + "angle": 0, + "content": "\\(\\gamma = 1.0\\) a photo of the large" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.518, + 0.345, + 0.524 + ], + "angle": 0, + "content": "windsurfing." + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.524, + 0.369, + 0.53 + ], + "angle": 0, + "content": "\\(\\gamma = 3.0\\) : windsurfing in tarifa" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.531, + 0.368, + 0.537 + ], + "angle": 0, + "content": "\\(x = 3.0\\) : windsurfing walls" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.538, + 0.368, + 0.544 + ], + "angle": 0, + "content": "pors 1300x1034" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.545, + 0.368, + 0.55 + ], + "angle": 0, + "content": "1M" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.551, + 0.324, + 0.557 + ], + "angle": 0, + "content": "EM. a Win" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.558, + 0.354, + 0.563 + ], + "angle": 0, + "content": "water" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.564, + 0.356, + 0.57 + ], + "angle": 0, + "content": "GT: A man riding a" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.571, + 0.357, + 0.576 + ], + "angle": 0, + "content": "wind sail in the" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.577, + 0.357, + 0.582 + ], + "angle": 0, + "content": "ocean filled with" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.583, + 0.324, + 0.589 + ], + "angle": 0, + "content": "waves" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.436, + 0.468, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.511, + 0.465, + 0.517 + ], + "angle": 0, + "content": "\\(\\gamma = 1.0\\) a photo of the large" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.518, + 0.424, + 0.524 + ], + "angle": 0, + "content": "giraffe." + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.524, + 0.461, + 0.53 + ], + "angle": 0, + "content": "\\(\\alpha = 2.0\\) a piraffe in a video" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.531, + 0.461, + 0.537 + ], + "angle": 0, + "content": "game" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.538, + 0.451, + 0.544 + ], + "angle": 0, + "content": "gamer:" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.545, + 0.424, + 0.55 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.551, + 0.458, + 0.557 + ], + "angle": 0, + "content": "photo 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.558, + 0.459, + 0.564 + ], + "angle": 0, + "content": "LM: a giraffe standing in a tall tower." + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.565, + 0.459, + 0.57 + ], + "angle": 0, + "content": "in a tall tree" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.571, + 0.461, + 0.577 + ], + "angle": 0, + "content": "GI: A giraffe in a dry" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.577, + 0.461, + 0.582 + ], + "angle": 0, + "content": "savannah with dry" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.583, + 0.452, + 0.59 + ], + "angle": 0, + "content": "shrubs" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Figure 9. Examples of captions generated from a model pretrained only on minimally curated data, for randomly selected images without human faces. Captions labeled \\(\\gamma = 1.0\\) are obtained without CFG; \\(\\gamma > 1\\) uses CFG; LM indicates LM guidance with \\(\\alpha = \\beta = 1\\) and 20 shots; GT indicates ground truth." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.196, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our study indicates that it is possible to substantially improve the extent to which generated captions uniquely describe the images they correspond to, raising questions regarding the goal of image captioning and how it should be evaluated. As it is conventionally formulated, image captioning aims not to provide text that can substitute for an image, but to write the text that a human annotator would have written. This formulation penalizes captions that are more descriptive than ground truth, even when a human might prefer them. On the other hand, treating image captioning as a problem of generating a caption that lies close to the image in the embedding space of an image-text model is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.893, + 0.528 + ], + "angle": 0, + "content": "also inadequate, because captions that lie close to the image need not be grammatical and may contain gibberish. Our proposed methods leveraging classifier-free guidance and language model guidance modulate the trade-offs between these two goals, as captured by various reference-based and reference-free metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "There are several possible extensions to our work. First, our present experiments use only greedy decoding. Although greedy decoding appears to perform reasonably well in our setup, it may be suboptimal for LM guidance with prompts that impose greater structure upon the captions. If the LM is prompted to output either \"there is a person in this image\" or \"there is no person this image\", greedy decoding is likely to fail even if the captioner properly scores the two possible captions, because when choosing between the tokens \"a\" and \"no\", the captioner has no knowledge of the structure that the LM will impose on future tokens. Since beam search could explore both tokens, it may offer better results in this scenario. Second, our method could be combined with RL-based methods to increase similarity in a contrastive embedding space, which may further improve retrieval performance and CLIPScore. Finally, with a perfect captioning model, \\( p(\\mathrm{image}|\\mathrm{caption}) \\) should increase with \\( \\gamma \\). However, in practice we find that \\( \\gamma > 3 \\) leads to a decrease in retrieval performance. This discrepancy suggests that the difference between the conditional and unconditional model distributions may be a noisy estimator of the pointwise mutual information. Although selecting \\( \\gamma \\) is one way to regularize this estimator, there may also be strategies to regularize \\( p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x}) \\) at training time." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15266" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.161 + ], + "angle": 0, + "content": "We thank Kevin Clark, David Fleet, Geoffrey Hinton, and the rest of the Google DeepMind Toronto team for inspiration, comments, and discussion." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.173, + 0.175, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.199, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.472, + 0.339 + ], + "angle": 0, + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[4] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[5] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. \\(ic^3\\): Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[6] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with CLIP reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, Seattle, United States, July 2022. Association for Computational Linguistics. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. PaLM: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[8] Katherine Crowson. You can apply a similar trick to classifier-free guidance to autoregressive transformers to sample from a synthetic \"super-conditioned\" distribution. https://twitter.com/RiversHaveWings/status/1478093658716966912, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.775, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[9] Béatrice Daille. Approche mixte pour l'extraction automatique de terminologie: statistiques lexicales et filtrres linguistiques. PhD thesis, Ph.D. thesis, Université Paris 7, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.199, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Hao Cheng, Hao Fang, Saurabh Gupta, Li Deng, Xiaodong He, Geoffrey Zweig, and Margaret Mitchell. Language models for image captioning: The quirks and what works. arXiv preprint arXiv:1505.01809, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[15] Micah Hodosh, Peter Young, and Julia Hockenmaier. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853-899, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[16] Yupan Huang, Hongwei Xue, Bei Liu, and Yutong Lu. Unifying multimodal transformer for bi-directional image and text generation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1138-1147, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[18] Andrej Karpathy and Fei-Fei Li. Deep visual-semantic alignments for generating image descriptions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3128-3137. IEEE Computer Society, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[19] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander R Fabbri, Yejin Choi, and Noah A Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. arXiv preprint arXiv:2112.04139, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[20] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Morrison, Ronan Le Bras, Yejin Choi, and Noah A Smith. Transparent human evaluation for image captioning. arXiv preprint arXiv:2111.08940, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[21] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[22] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning. In International Conference on Learning Representations, 2022. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15267" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[23] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[25] Annika Lindh, Robert J Ross, Abhijit Mahalunkar, Giancarlo Salton, and John D Kelleher. Generating diverse and meaningful captions. In International Conference on Artificial Neural Networks, pages 176-187. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[28] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[29] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.561, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.688, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.745, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[34] Noam Rotstein, David Bensaid, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models to fuse visual data into enriched image captions. arXiv preprint arXiv:2305.17718, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[36] Guillaume Sanchez, Honglu Fan, Alexander Spangher, Elad Levi, Pawan Sasanka Ammanamanchi, and Stella Biderman. Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[38] David R So, Wojciech Manke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V Le. Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[39] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[40] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[42] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[43] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[44] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[45] Shitong Xu. Clip-diffusion-lm: Apply diffusion model on image captioning. arXiv preprint arXiv:2210.04559, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[46] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[47] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3, 4, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15268" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "fei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[49] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104–12113, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[50] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.471, + 0.318 + ], + "angle": 0, + "content": "[51] Youuyuan Zhang, Jiuniu Wang, Hao Wu, and Wenjia Xu. Distinctive image captioning via clip guided group optimization. arXiv preprint arXiv:2208.04254, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[52] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang, Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint arXiv:2211.11694, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15269" + } + ] +] \ No newline at end of file diff --git a/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_origin.pdf b/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9186f7f95329bbb0295186cebdf1c0c0e7993a34 --- /dev/null +++ b/2023/Guiding Image Captioning Models Toward More Specific Captions/17ad3fd0-4183-40a9-ac84-18391fe9effe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e8fa9dfd7ebd868d5b6d8adbf502481941af921141bd877c93f433f1209051e +size 3471208 diff --git a/2023/Guiding Image Captioning Models Toward More Specific Captions/full.md b/2023/Guiding Image Captioning Models Toward More Specific Captions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ca7b06f5d1b53cbb98286e51cace82f61f3cf4d9 --- /dev/null +++ b/2023/Guiding Image Captioning Models Toward More Specific Captions/full.md @@ -0,0 +1,355 @@ +# Guiding image captioning models toward more specific captions + +Simon Kornblith + +Google DeepMind + +Lala Li + +2Apple AI/ML + +Zirui Wang $^{2*}$ + +$^{3}$ University of Washington + +Thao Nguyen $^{3\dagger}$ + +# Abstract + +Image captioning is conventionally formulated as the task of generating captions for images that match the distribution of reference image-caption pairs. However, reference captions in standard captioning datasets are short and may not uniquely identify the images they describe. These problems are further exacerbated when models are trained directly on image-alt text pairs collected from the internet. In this work, we show that it is possible to generate more specific captions with minimal changes to the training process. We implement classifier-free guidance [14] for an autoregressive captioning model by fine-tuning it to estimate both conditional and unconditional distributions over captions. The guidance scale applied at decoding controls a trade-off between maximizing $p(\text{caption} | \text{image})$ and $p(\text{image} | \text{caption})$ . Compared to standard greedy decoding, decoding with a guidance scale of 2 substantially improves reference-free metrics such as CLIPScore (0.808 vs. 0.775) and caption→image retrieval performance in the CLIP embedding space (recall@1 44.6% vs. 26.5%), but worsens standard reference-based captioning metrics (e.g., CIDEr 78.6 vs 126.1). We further explore the use of language models to guide the decoding process, obtaining small improvements over the Pareto frontier of reference-free vs. reference-based captioning metrics that arises from classifier-free guidance, and substantially improving the quality of captions generated from a model trained only on minimally curated web data. + +# 1. Introduction + +Image captioning is both a difficult task for computer vision systems to perform and a difficult task to evaluate. Although automated captioning metrics rank the best captioning systems higher than humans, human raters still show a strong preference for human-generated captions [20], suggesting shortcomings in both captioning models and metrics. One shortcoming relates to the lack of specificity in generated captions. Conventional maximum likelihood- + +![](images/ac98db833f7ae2e815e3f274219bb5ef911b530cefa419804e8ec00b668373c4.jpg) +$\gamma = 1.0$ a man riding a blue motorcycle on a dirt road + +![](images/fc5d4f0ca1f455ed06eee5c5ae79fdd45a350774a01056ea77c2299f646d214b.jpg) +$\gamma = 1.5$ a man riding a blue motorcycle on a straw bale +$\gamma = 2.0$ rider on blue suzuki motorcycle near straw bales +$\gamma = 3.0$ racer on blue suzuki motorcycle leaning up against straw bales +GT A person riding a baby blue motorcycle near haystacks +$y = 1.0$ $y = 2.0$ +$y = 1,2$ $y = 3,0$ +$\gamma = 1.5$ $\gamma = 4.0$ +Figure 1. Using classifier-free guidance $(\gamma > 1)$ results in more specific captions that are farther from the reference distribution. Left: Example of captions generated at different guidance scales for a single image. Right: Caption $\rightarrow$ image recall@1 with CLIP ViT-B/32 vs. CIDEr score, for captions generated with different guidance scales $\gamma$ on MS-COCO. Higher scales improve retrieval accuracy at the expense of CIDEr. + +based image captioning models attempt to generate captions such that the $p(\text{caption} | \text{image})$ is high. However, captions from the ground truth distribution are often nonspecific, e.g., human annotators will usually describe a German Shepard only as a dog. Moreover, previous work has emphasized "reference-based" captioning metrics that measure the match between generated captions and human-provided ground truth captions [28, 23, 42]. These metrics intrinsically penalize captions that are more specific than ground truth. + +In this work, we explore strategies to guide image captioning models to produce more specific captions by modifying the decoding distribution, and explore the tradeoffs in captioning metrics that result. We first investigate the application of classifier-free guidance (CFG) [14] to image captioning with autoregressive models. Classifier-free guidance increases $p(\text{image|caption})$ at the expense of $p(\text{caption|image})$ . Although CFG hurts reference-based image captioning metrics such as BLEU [28], ROUGE [23], and CIDEr [42], it improves "reference-free" metrics that measure captions' specificity via the similarity between the + +image and the generated caption in the embedding space of image-text models [13] or caption $\rightarrow$ image retrieval performance. Qualitatively, we find that captions generated with CFG are more specific than both the ground truth captions and captions generated without CFG, but they are less grammatical, particularly at high CFG scales. + +Beyond classifier-free guidance, we experiment with guiding image captioning models using the probability distribution obtained from a few shot-prompted language model (LM). We find that using a language model to guide a captioning model trained on MS-COCO [24] with descriptive manually written captions can allow it to achieve slightly better trade-offs between reference-free vs. reference-based captioning metrics than those observed with CFG. LM guidance also substantially improves the captions produced by a model trained exclusively on minimally curated web data. Although this model achieves a CIDEr score of only 21.8 without guidance, this CIDEr score improves to 57.4 when guided by a language model prompted with 20 captions from the MS-COCO training set. In summary, our contributions are as follows: + +- We propose two strategies to guide image captioning models to produce more specific captions: classifier-free guidance and language model guidance. +- We demonstrate that classifier-free guidance yields captions that are closer to the corresponding image in the embedding space of image-text models, but are farther from human-provided reference captions. +- We show that language model guidance can alter caption styles, substantially improving captions produced by a model trained only on minimal curated web data and marginally improving the trade-off between captioning metrics observed with classifier-free guidance. + +# 2. Related work + +Measuring specificity of captions. Early work using neural networks for image captioning found that models have a propensity to regurgitate captions from their training data, and as a result, the generated captions are not descriptive enough to uniquely identify images [43, 11]. To address this shortcoming, Lindh et al. [25] proposed to use caption $\rightarrow$ image recall with an image retrieval model to examine whether images can be retrieved from generated captions, and further attempt to differentiate through this retrieval process to train a captioning model. Their approach marginally improves retrieval accuracy, but worsens reference-based captioning metrics. More recent work has adopted approaches to evaluate the specificity of captions based on the CLIP image-text model [30]. Hessel et al. [13] propose CLIPScore, an image captioning metric based on the cosine similarity between CLIP embeddings of the image and the generated caption. Kasai et al. [20] re + +port that CLIPScore-based metrics align better with human judgments compared to reference-based captioning metrics. + +Improving specificity of captions. Recent work has attempted to train caption models to directly maximize the similarity of captions with corresponding images in the CLIP embedding space using reinforcement learning. This work has found that adding such losses worsens standard reference-based captioning metrics but improve similarity and retrieval in the CLIP embedding space [16, 6, 51], similar to our observations regarding CFG. Wen et al. [44] attempt to generate prompts for text-to-image generative models that correspond to specific images without a captioning model, by directly optimizing the similarity between the text and image in the CLIP embedding space using a gradient-based discrete optimization procedure, but the resulting text is not grammatical. + +Other work has attempted to generate more descriptive captions through different means. Dense captioning [46] aims to detect and caption all objects in an image, but concatenating all of these captions leads to long and unnatural captions, whereas CFG produces single-sentence captions. The Localized Narratives dataset [29] contains visually grounded captions for MS-COCO images collected through voice annotation. These captions are substantially more descriptive than the captions in the MS-COCO dataset and can be used for model training. Two recent papers propose to combine multiple captions from a single model [5] or outputs of different vision models [34], and then combine them using a language model. The resulting captions are longer, but achieve greater caption→image recall. + +Captioning from uncurated data. In Section 4.2, we explore the use of LM guidance for captioning with access to uncurated image-text data from the web and a small number of captions but not images from the target distribution. This setting, which does not rely on aligned images and captions from the target distribution, is often referred to as "zero-shot" captioning, and previous work has pursued a number of alternative approaches. Flamingo [3] and CM3 [1] perform zero-shot captioning by pretraining on interleaved image/text data. MAGIC [39] and ZeroCap [41] generate captions using a combination of guidance from CLIP and a large language model. Other recent work adapts CLIP to perform captioning by training a text decoder using only captions, with no corresponding images [27, 22]. + +Classifier-free guidance. CFG is widely used in diffusion-based and autoregressive text-to-image models [26, 32, 35, 33, 12, 48]. Because of the popularity of the combination of CFG and diffusion, previous work that has performed image captioning with diffusion models has also examined the use of CFG. This work finds either no benefit to using CFG [45] or a small and inconsistent benefit that appears to vary with minor changes in training settings [52]. However, these studies do not seek to generate + +more specific captions, and thus measure only reference-based captioning metrics, which we likewise find do not benefit from CFG. Concurrently with out work, [36] propose to use classifier-free guidance to improve prompt following in large language models. + +# 3. Methods + +# 3.1. Classifier-free guidance for image captioning + +Let $\mathbf{x}$ be an image caption and $\mathbf{y}$ be the corresponding image. A standard captioning model aims to model the likelihood $p(\mathrm{x}|\mathrm{y})$ , factorized autoregressively in terms of the probability of each token given previous tokens + +$$ +p (\mathrm {x} | \mathrm {y}) = p \left(\mathrm {x} _ {n} \mid \mathrm {x} _ {n - 1}, \dots \mathrm {x} _ {1}, \mathrm {y}\right) \dots p \left(\mathrm {x} _ {1} \mid \mathrm {y}\right). \tag {1} +$$ + +The network is trained so that its output distribution $q_{\theta}(\mathbf{x}_n|\mathbf{x}_{n-1}, \ldots, \mathbf{x}_1, \mathbf{y}) \stackrel{\mathrm{def}}{=} \text{softmax}(f_{\theta}(\mathbf{x}_{n-1}, \ldots, \mathbf{x}_1, \mathbf{y}))$ approximates $p(\mathbf{x}_n|\mathbf{x}_{n-1}, \ldots, \mathbf{x}_1, \mathbf{y})$ . At inference time, one typically uses beam search or greedy decoding to produce a caption that has a particularly high probability. In this work, we use greedy decoding because it is the more common choice and it is also simpler to implement. + +Classifier-free guidance (CFG) [14] modifies the inference procedure so it instead aims to achieve high values of + +$$ +l _ {\boldsymbol {\theta}, \gamma} (\mathrm {x}, \mathrm {y}) \stackrel {\text {d e f}} {=} p (\mathrm {x}) \left(\frac {p (\mathrm {x} | \mathrm {y})}{p (\mathrm {x})}\right) ^ {\gamma} \propto p (\mathrm {x}) p (\mathrm {y} | \mathrm {x}) ^ {\gamma}, \tag {2} +$$ + +where proportionality holds because $p(\mathrm{x}|\mathrm{y}) / p(\mathrm{x}) = p(\mathrm{y}|\mathrm{x}) / p(\mathrm{y})$ and $p(\mathrm{y})$ is fixed. The parameter $\gamma$ is called the guidance scale and controls the trade-off between maximization of $p(\mathrm{x}|\mathrm{y})$ and $p(\mathrm{y}|\mathrm{x})$ . When $\gamma = 1$ , $l_{\theta, \gamma}(\mathrm{x}, \mathrm{y}) = p(\mathrm{x}|\mathrm{y})$ and guidance has no effect. Setting $\gamma > 1$ inflates the probability of the image given the caption $p(\mathrm{y}|\mathrm{x})$ relative to the unconditional probability of the caption $p(\mathrm{x})$ . + +Ho and Salimans [14] originally proposed CFG in the context of diffusion models, which estimate the score functions $\nabla \log p(\mathbf{x}|\mathbf{y})$ and $\nabla \log p(\mathbf{x})$ . Although $l_{\theta,\gamma}(\mathbf{x},\mathbf{y})$ factorizes autoregressively, it is not a normalized probability distribution, so it is not entirely clear how one should sample tokens when performing autoregressive generation. Crowson [8] suggested to sample from + +$$ +\begin{array}{l} \tilde {q} _ {\boldsymbol {\theta}, \gamma} (\mathrm {x} _ {n} | \mathrm {x} _ {n - 1}, \dots , \mathrm {x} _ {1}, \mathrm {y}) \stackrel {\mathrm {d e f}} {=} \operatorname {s o f t m a x} (f _ {\boldsymbol {\theta}} (\mathrm {x} _ {n - 1}, \dots , \mathrm {x} _ {1}, \mathbf {0}) \\ + \gamma \left(f _ {\boldsymbol {\theta}} \left(\mathrm {x} _ {n - 1}, \dots , \mathrm {x} _ {1}, \mathrm {y}\right) - f _ {\boldsymbol {\theta}} \left(\mathrm {x} _ {n - 1}, \dots , \mathrm {x} _ {1}, \mathbf {0}\right)\right), \tag {3} \\ \end{array} +$$ + +where $f_{\theta}(\mathbf{x}_{n - 1},\ldots ,\mathbf{x}_1,\mathbf{0})$ are logits generated by the model without conditioning, usually by passing zeros in place of the conditioning information. This formulation has been successfully applied in autoregressive image models [12, 48]. In our experiments, since we decode greedily, i.e., at each step we take the token that maximizes $\tilde{q}_{\theta ,\gamma}(\mathrm{x}_n|\mathrm{x}_{n - 1},\dots ,\mathrm{x}_1,\mathrm{y})$ and thus $l_{\theta ,\gamma}(\mathrm{x},\mathrm{y})$ , any form of normalization of the per-step sampling distribution would produce the same captions. We provide pseudocode in Appendix A.1. Crucially, the only change required to train + +ing is to learn an unconditional model of captions alongside the image-conditional model, which we accomplish by randomly masking the image embeddings. + +# 3.2. Language model guidance + +Inspired by classifier-free guidance, we consider language model (LM) guidance, which attempts to maximize + +$$ +l _ {\boldsymbol {\theta}, \gamma} ^ {\prime} (\mathrm {x}, \mathrm {y}) \stackrel {\text {d e f}} {=} q (\mathrm {x}) \left(\frac {p (\mathrm {x} | \mathrm {y}) ^ {\alpha}}{p (\mathrm {x}) ^ {\beta}}\right), \tag {4} +$$ + +where $p(x)$ and $p(x|y)$ are obtained from a captioning model as in CFG but $q(\mathbf{x})$ is obtained from a language model that was trained independently (but with the same vocabulary) on a large text corpus. The quantity $p(\mathrm{x}|\mathrm{y}) / p(\mathrm{x}) = p(\mathrm{x},\mathrm{y}) / (p(\mathrm{x})p(\mathrm{y}))$ measures the strength of the association between a caption and an image; its logarithm is the pointwise mutual information (PMI). LM guidance relies on the assumption that, even for large shifts in the prior distribution of captions $p(\mathbf{x})$ , the shift in PMI will be small. Empirically, we obtain better results by allowing different exponents for the numerator and denominator, with $\alpha > \beta$ . This decoupling resembles $\mathrm{PMI}^k$ [9], which reduces the bias of PMI toward rare associations. We provide a more detailed derivation in Appendix A.2. + +We investigate two applications of LM guidance. First, we combine a captioning model fine-tuned on MS-COCO with a LM prompted with manually written descriptive captions to alter the style of the captions the model produces. The manually written prompts are shown in Appendix A.4. Second, we combine a captioning model trained only on low-quality web data with a LM prompted with varying numbers of examples from the MS-COCO training set to evaluate the ability of LM guidance to elicit higher-quality captions without high-quality paired data. We randomly select a different set of captions for each minibatch of four test examples. In both cases, we separate the captions with two newlines. Because this format leads the LM to place probability mass on the newline token to end the caption, we transfer the probability mass from the newline token to the EOS token. See Appendix A.3 for pseudocode. + +# 3.3. Models and training + +Our captioning model is a "bottleneck" variant of CoCa-Base [47], which combines a contrastive loss with a captioning loss to simultaneously learn aligned image and text embeddings as well as a captioner. The architecture consists of an image encoder, a unimodal text decoder, and a multimodal text decoder, each of which are Transformers with 12 layers, 768 hidden dimensions, an MLP of size 3072, and 12 self-attention heads, matching $\mathrm{BERT}_{\mathrm{BASE}}$ [10] and GPT-1 [31]. The image encoder is a ViT-B/18 that processes $288 \times 288$ input and produces an embedding such that images are embedded close to their corresponding text. + +CoCa's multimodal text decoder processes the representations of the image encoder to produce a caption. Whereas [47] conditions the multimodal text decoder using crossattention to pooled representations, our bottleneck variant uses only the contrastive image embedding. Appendix A.5 shows a diagram of the resulting architecture. We adopt this bottleneck variant because of its simplicity and the conceptual appeal: When CFG is used, the captioner's role is to invert the image embedding, providing a caption that, when embedded by the text encoder, lies close to it. However, as we show in Appendix B.1, this choice of the bottleneck model is not critical, and CFG is equally effective with the standard CoCa architecture with attention pooling. + +For CFG experiments, we pretrain our model on an image-text dataset comprising images from the JFT-5B dataset [40, 49] paired with their corresponding label names substituted into a randomly selected prompt from the list provided by Radford et al. [30], web images paired with noisy alt text from the ALIGNN dataset [17], and a small amount of data from other sources. We do not mask conditioning information during pretraining. We then fine-tune on the combined MS-COCO train and Karpathy validation splits [18] using Adam with batch size 128. We linearly warm up to a learning rate of $1 \times 10^{-5}$ over the first 1,000 steps and linearly decay to zero over the rest of training. We vary $\gamma \in \{1.0, 1.2, 1.5, 2.0, 3.0, 4.0\}$ , conditioning masking proportion in $\{0.0, 0.25, 0.5, 0.75\}$ , and numbers of steps in $\{5,000, 10,000, 20,000, 50,000\}$ . We report results from the model trained for 20,000 steps with masking proportion 0.5, which achieves near-optimal results, in Tables 1 and B.4, and sample example captions from it. To ensure that results generalize across datasets, we also experiment with a model fine-tuned on Conceptual Captions [37] for 100,000 steps with masking proportion 0.5. + +For LM guidance experiments, we pretrain on the JFT-5B and ALIGN datasets, again following the recipe of [47]. For zero-shot captioning experiments, we fine-tune this model on the same datasets for an additional 50,000 steps with conditioning masking proportion of 0.5 to improve our ability to sample unconditionally. For LM guidance on MS-COCO, we first fine-tune on ALIGN, JFT-5B images backcaptioned by an MS-COCO fine-tuned CoCa2B model, and a small amount of internal data before finetuning on MS-COCO. Our language model is a variant of Primer [38] with 2 billion parameters, trained on a similar dataset to that used to train PaLM [7]. + +# 3.4. Evaluation + +We adopt the standard reference-based captioning metrics BLEU-4, METEOR, ROUGE, and CIDEr, as well + +as reference-free captioning metrics based on CLIP ViT-B/32 [30]. The first reference-free captioning metric is CLIPScore [13], which is defined as $\mathrm{CLIP - S}(c,v) = 2.5\cdot \max (\cos (c,v),0)$ where $c$ and $v$ are the CLIP embeddings of the caption and image respectively. The second reference-free metric measures the accuracy with which we can retrieve an image from the generated caption within a given test split by taking the $k$ nearest neighbors of the caption in the CLIP embedding space. Because recall@ $k$ for $k > 1$ is highly correlated with recall@1 (R@5: $r = 0.99$ , R@10: $r = 0.98$ ), we plot only recall@1. We additionally report RefOnlyCLIP-S, a reference-based metric that uses the CLIP text encoder to compute the similarity of CLIP embeddings of the generated captions with embeddings of ground truth captions, and RefCLIP-S, which takes the average of the per-image harmonic means of CLIP-S and RefOnlyCLIP-S [13]. Unless otherwise stated, all evaluation is performed on the MS-COCO Karpathy test split [18]. + +# 4. Results + +# 4.1. Classifier-free guidance + +We first investigate the trade-off between reference-based and reference-free image captioning metrics as a function of guidance scale. Because different guidance scales and metrics could conceivably benefit from different fine-tuning hyperparameter combinations, we plot all results from our hyperparameter grid in Figure 2. Although standard greedy decoding $(\gamma = 1.0)$ produces the highest CIDEr, METEOR, ROUGE, and BLEU-4 scores, higher guidance weights consistently yield higher values of reference-free captioning metrics. In particular, $\gamma = 3.0$ offers both the best caption→image recall and the best CLIP-Score. + +Table 1 compares our results, obtained from a single model evaluated at different guidance scales, with previous work that reports either CLIPScore or CLIP ViT-B/32 caption $\rightarrow$ image retrieval performance. Although our model is trained with standard cross-entropy loss rather than a CLIP-based loss and our pretraining dataset is distinct from CLIP's, sampling from our model with CFG yields higher CLIPScores than all other models trained without CLIP-based losses, and better CLIP caption $\rightarrow$ image retrieval even when compared with models that use CLIP-based losses. + +We present examples of captions generated at different CFG scales in Figure 3. Higher CFG strengths lead to more descriptive captions. At $\gamma = 1.0$ , the central object in the top left image is described as a "car" as in the ground truth caption, whereas at $\gamma > 1.0$ it is a "station wagon." Similarly, at low CFG strengths, the birds in the center image are described simply as "birds," whereas at $\gamma = 2.0$ they become "crested cranes." However, at $\gamma = 3.0$ , captions clearly become less grammatical, containing repeated + +![](images/66ceb179a19a7d97a20c9c8a9512f0e8d35e690abc03d9acd40efaac31ae6f19.jpg) +Figure 2. Classifier-free guidance controls a trade off between reference-free and reference-based captioning metrics. Each point reflects a model trained with a different hyperparameter combination; each color represents a $\gamma$ value used to decode. Models are evaluated with different guidance scales $\gamma$ , using reference-free captioning metrics based on CLIP ViT-B/32 (y-axes; top: CLIPScore, bottom: recall@1) and reference-based captioning metrics (x-axes). The dashed line reflects the value of the reference-free captioning metric for the ground-truth captions obtained from MS-COCO. +Figure 3. Caption descriptiveness increases with CFG strength, but high CFG strengths produce agrammatical captions. Here we show examples of captions generated with different classifier-free guidance scales, for randomly selected images without human faces from the MS-COCO Karpathy test split. Captions labeled $\gamma = 1.0$ are obtained without CFG; $\gamma > 1$ uses CFG; GT = ground truth. + +
ModelReference-Based MetricsReference-Free MetricsRefCLIP-S
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10
Models trained with CLIP features or losses:
CLIP-Captioner [4]38.729.358.6126.00.8110.7540.814
UMT-BITG [16]37.328.257.9122.60.772
X-LAN+SCST+GEG [51]36.528.757.5121.728.150.367.2
CIDEr + CLIP-S Reward [6]37.728.858.3124.60.77224.450.263.1
CLIP-S Reward [6]6.218.731.611.20.86042.571.682.2
ZeroCap [41]2.611.514.60.870.79
Models trained without access to CLIP:
UMT-BITG w/o CLIP loss [16]37.628.358.1122.50.725
VinVL-large [50]41.030.959.4*140.90.91*0.78*0.84*
Ours (γ = 1.0)36.130.558.2126.10.9000.77526.551.964.10.830
Ours (γ = 1.2)35.130.057.5124.10.8990.78531.357.469.30.835
Ours (γ = 1.5)31.528.454.4113.20.8910.79636.664.075.00.838
Ours (γ = 2.0)20.923.343.078.60.8620.80844.671.781.70.831
Ours (γ = 3.0)11.517.129.441.70.8200.80849.475.784.70.811
Ours (γ = 4.0)6.512.318.417.30.7660.78244.771.380.90.771
+ +Table 1. Quantitative comparison of our approach with results from previous work that reports CLIP-based metrics. For VinVL-large, * indicates metrics from [19]. + +
γ=1.0: a car with a surfboard on top of it parked next to a car +γ=1.5: a vintage station wagon with a surfboard on top +γ=2.0: a antique station wagons and a car buick station wagon +γ=3.0: buick woody wyddo staitonwagon and surf green station wagon parked in front of car show establishment +GT: An old green and brownγ=1.0: a knife is sitting on a cutting board next to a sliced carrot +γ=1.5: a knife is sitting on a cutting board next to an orange +γ=2.0: knife sitting on cutting board next to whole one +γ=3.0: knife sitting on cutting board next to misschappi carrot on cutting board +GT: A knife sticking out of the side of a block of cheese.γ=1.0: a herd of sheep grazing on a grass covered road +γ=1.5: sheep grazing on a high-way with a truck on the road +γ=2.0: sheep graze on freeway medians where grass is grown +γ=3.0: grazing trucks blocking sheep on roadway grazing grass +GT: A large herd of sheep are grazing by the busy road.γ=1.0: two birds standing in a cage with their heads in the air +γ=1.5: two birds standing inside of a cage in a zoo +γ=2.0: two crested cranes inside a wire cage +γ=3.0: crested tantalus cranes caged together in bird-cage enclosure +GT: Two birds who are looking out of the cage they are in.γ=1.0: a view of a city with a clock tower in the background +γ=1.5: a city with steeples and trees and buildings +γ=2.0: spires of churches line a city skyline +γ=3.0: spires steeples buildings trees church spires and trees +GT: A clock that is on the side of a tower.γ=1.0: a kitchen with a microwave and refrigerator +γ=1.5: a kitchen with a mi-crowage and a refrigerator +γ=2.0: a kitchen with red ap-pliances and white cup-boards +γ=3.0: appliances sit in a small empty dingroom red and white kitchen +GT: A kitchen that has a tile floor, a refrigerator, a microwave, and a toaster.γ=1.0: a bathroom with a large mirror and a bathtub +γ=1.5: a bathroom with a large mirror and a bathtub +γ=2.0: a spacious bathroom with a large mirror and a large tub +γ=3.0: spacious bathroom with chandelier over tub mir-rors and tv +GT: A bathroom with a tub, sinks, lights and a televi-sion.
+ +words ("woody woody") and nonsense words ("misshappi", "dingroomy"). Figure 4 shows captions obtained with and without CFG next to the top 5 closest images in the embedding space of CoCa 2B [47],² where it is clear that CFG adds details to captions that help to distinguish them from other captions in the test split. We provide additional examples in Appendix C. + +To provide additional quantitative assessments of the specificity of elicited captions, we perform two additional evaluations, described further in Appendix B.2. First, we generate captions for the Stanford Dogs [21] test set, which consists of 8,580 images in total of 120 breeds of dogs, and examine their properties. Without guidance, only $1.9\%$ of captions contain one of the 120 breed names, whereas at $\gamma = 2.0$ , $42.4\%$ do. The percentage of these breed names that are correct changes little, from $61.7\%$ without guidance to $58.5\%$ at $\gamma = 2.0$ . Second, we performed a human evaluation comparing captions of MS-COCO test set images obtained without guidance and at $\gamma = 2.0$ . We asked subjects to select the caption that is "better" and "more descriptive" or to indicate that they are both equal. When we asked these questions separately, we found that the two sets of captions are statistically indistinguishable. However, when asking both questions on the same survey, we found that captions generated without guidance are slightly "better" (50.5% vs. $46.6\%$ , $p = 0.006$ , binomial test) but captions generated at $\gamma = 2.0$ are "more descriptive" (52.7% vs. $45.8\%$ , $p = 1 \times 10^{-6}$ ). + +To validate the reliability of our results, we further measure the impact of CFG on three additional datasets, nocaps [2], Flickr-8k [15], and Conceptual Captions (CC3M) [37], as well as with alternative retrieval models. nocaps is a test set for captioning models with objects not present in MS-COCO; Flickr-8k is a small captioning dataset collected using a different procedure than MS-COCO; and Conceptual Captions is a set of 3.3M captions collected from filtered alt-text. We fine-tune the bottleneck CoCa-Base model directly on CC3M, and use our model fine-tuned on MS-COCO to caption images on nocaps and Flickr-8K. As shown in Figure 5, we find trade-offs between reference-based and reference-free captioning metrics similar to those above. In Appendix B.3, we report reference-free captioning metrics on MS-COCO computed with two additional retrieval models: the pretrained CoCa 2B model from [47] and the fine-tuned CoCa Base model that we use to generate captions. With both models, CFG substantially increases recall, in line with results obtained with CLIP ViT-B/32. + +Although CFG produces captions that are more successful at uniquely identifying images than decoding from the conditional distribution, caption lengths are similar for + +![](images/2fd0d825af643485344d7b9cc5c1af8faf84bc57f9bea56ace03c9b449b3f99c.jpg) +Figure 4. Captions generated with CFG contain specific details that improve retrieval. For each reference image (far left), we show captions at $\gamma = 1.0$ (no guidance) and $\gamma = 2.0$ . To the right, we show the closest images to each caption in the CoCa embedding space. Reference images are selected at random subject to the constraints that the closest image differs between $\gamma$ values and there are no identifiable human faces. + +![](images/c2d56f207cc27833fea0be98669f6dc37a54d07155aa418bb1d081a3eb1a6c4a.jpg) +Figure 5. CFG also yields trade-offs between captioning metrics on nocaps, Flickr-8K, and CC3M. + +
γWordsCharacters
1.09.6 ± 1.444.2 ± 7.2
1.29.6 ± 1.444.7 ± 7.4
1.59.4 ± 1.445.7 ± 7.8
2.09.3 ± 2.450.3 ± 18.6
3.010.7 ± 7.669.0 ± 56.1
4.019.9 ± 16.9161.2 ± 140.0
+ +Table 2. Moderate CFG scales do not substantially change caption lengths, although higher CFG scales result in longer captions. Numbers are mean $\pm$ standard deviation. + +![](images/eff26af2f82da7153684253b61b1e109761bee454bd2e4357f22683717325faa.jpg) +Figure 6. Language model guidance produces captions that slightly exceed the Pareto frontier of CIDEr vs. caption $\rightarrow$ image retrieval accuracy on MS-COCO. + +$\gamma \in [1,2]$ , as shown in Table 2. Thus, at low guidance strengths, CFG improves recall by making more efficient use of words, rather than by producing more verbose captions. Higher CFG strengths lead to longer captions but, as described above, these captions are agrammatical and contain nonsense words. + +# 4.2. Language model guidance + +We first experiment with guiding a captioning model fine-tuned on MS-COCO to produce more descriptive captions using a language model prompted with manually written prompts. We first manually wrote a prompt containing 10 descriptive captions of COCO test set images (Appendix A.4). We then sweep over $\alpha \in \{1,2,3,4,5,6,7,8,9,10,12,15\}$ and $\beta \in \{0,\alpha /4,\alpha /2,3 / 4\alpha ,\alpha \}$ , and compare the resulting retrieval/CIDEr trade-off to that produced by the same model with CFG. We observe that it is possible to obtain small improvements upon the Pareto frontier provided by CFG, as shown in Figure 6. With $\alpha = 5$ , $\beta = -5 / 2$ , LM guidance achieves CLIP ViT-B/32 R@1 of $39.6\%$ and CIDEr of 114.4, whereas CFG with $\gamma = 1.6$ is worse on both metrics, achieving R@1 of $39.0\%$ and CIDEr of 109.3. + +We further experiment with using prompting to control the captioner using a manually written prompt of 25 captions in the form of "a photo of NUMBER OBJECTS" (e.g., "a photo of eight apples"; see Appendix A.4). With $\alpha = \beta = 1$ , the guided model is able to match this format and counts the number of objects in images (Figure 7). + +We next investigate whether language model guidance can elicit better captions from a model trained only on + +![](images/86ea8725a151896a78c9bc2d7ca0dd77994637f50668653f9582caf42b0fdcc3.jpg) +a photo of two dogs + +![](images/04ff0335e5d94d7e00d59a44f0a1d6d6a969e82fbd910d6d5fd04b378bfbef3b.jpg) +a photo of one bird + +![](images/4dd6affcc4456e67eb59b154b72d1e6d0c652ae69d794473dc5449318f79ef18.jpg) +a photo of four suitcases +Figure 7. Captions generated with LM guidance with a prompt of 25 captions in the form of "a photo of NUMBER OBJECTS". Examples are selected to show different numbers of objects. + +![](images/8a2e42dd24594938eb170c50661f9dc971aa37ccbc9747a69b8fa313483c379a.jpg) +a photo of five sheep + +![](images/d47a7b746a450aefdfc7122b35fd5c0bd04085c33a4baf0caaa5cfd4906cfc94.jpg) +Figure 8. LM guidance substantially improves CIDEr and retrieval scores of a model trained solely on minimally curated web data and evaluated on MS-COCO. The x-axis shows the number of captions used to prompt the LM; we do not prompt with images. + +![](images/cf52968589ca7511b10d22284360cd466927fd7cfc51cc8e7966d06cf21efd8d.jpg) + +low-quality data. Here, we use a CoCa model that is pretrained on image-alt text pairs from the web (the ALIGN dataset [17]) and classification labels converted to text (the JFT-5B dataset [49]), without any additional fine-tuning. Because the data distribution places higher probability mass on short, non-descriptive captions than on longer captions, the resulting model is of limited utility for captioning, and would generally need to be fine-tuned on another dataset such as MS-COCO before being applied to a captioning task. Rather than fine-tune, we use LM guidance to prompt the model with captions from the MS-COCO training set. + +LM guidance substantially improves the quality of the captions produced by the original pretrained CoCa model without any clean parallel data. With LM guidance, we achieve CIDEr scores of 48.6 with 5 shots and 59.7 with 50 shots, far exceeding the CIDEr score of 21.8 obtained with no guidance. Figure 8 shows CIDEr and CLIP recall@1 scores for LM guidance of this pretrained CoCa model as a function of the number of shots, with $\alpha = \beta = 1$ . Table 3 compares classifier-free guidance and LM guidance. CFG yields higher CLIP-Scores and retrieval accuracy than LM guidance with $\alpha = \beta = 1$ , but LM guidance provides much higher CIDEr scores. + +We compare captions generated with CFG to those generated with LM guidance for four images in Figure 9. In general, CFG produces agrammatical captions, whereas LM guidance produces grammatical captions but hallucinates details. For example, the image in the upper left shows two elephants and no zebras, but LM guidance leads to the caption "an elephant and a zebra in a field." + +
ModelReference-Based MetricsReference-Free Metrics
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10RefCLIP-S
Classifier-free guidance:
γ = 1.08.28.321.821.80.7660.6949.019.526.10.725
γ = 1.28.69.524.525.00.7810.71812.727.235.10.745
γ = 1.58.910.025.625.20.7800.72816.733.843.00.750
γ = 2.08.19.723.922.90.7770.74121.240.851.10.755
γ = 3.07.18.720.018.50.7670.75325.847.858.30.756
γ = 4.06.47.516.313.90.7490.74327.348.558.10.742
Language model guidance with α = β = 1:
2 captions12.714.634.739.30.8060.68810.023.732.40.740
5 captions15.016.639.148.60.8270.71212.427.537.50.763
10 captions16.217.740.553.10.8350.72313.030.541.00.773
20 captions17.418.441.657.40.8390.72814.432.242.70.777
50 captions18.119.142.559.70.8400.72913.432.543.90.778
Other models trained without aligned MS-COCO images and captions:
ZeroCap [41]2.611.514.60.870.79
MAGIC [39]12.917.439.949.3
Flamingo [3]84.3
DeCap (560 captions) [22]51.4
DeCap (full train set) [22]24.725.091.2
CapDec (full train set) [27]26.425.151.891.8
+ +Table 3. Comparison of decoding strategies for a captioning model trained only on minimally curated web data (JFT-5B and ALIGN) and evaluated on MS-COCO. At the bottom, we report metrics for other models trained without aligned MS-COCO images and captions. These models may not be directly comparable since they use different pretraining data. DeCap and CapDec use all 560K captions in the MS-COCO training set to train their decoders; we include CIDEr for DeCap with 560 captions (0.1% of the training data) for comparison. + +![](images/f48e3bb4252d02890f0913187077bc55a3e0f64d7f5d1598c9f29bba6a389f9c.jpg) +$\gamma = 1.0$ : a photo of the small elephant +$\gamma = 2.0$ : elephants in the +ruaha national park +$\gamma = 3.0$ : elephants chobe +np +LM: a elephant and a +zebra in a field +OT: Two elephants. +G1: Two elephants standing on a grassy field next to a tree. +Figure 9. Examples of captions generated from a model pretrained only on minimally curated data, for randomly selected images without human faces. Captions labeled $\gamma = 1.0$ are obtained without CFG; $\gamma > 1$ uses CFG; LM indicates LM guidance with $\alpha = \beta = 1$ and 20 shots; GT indicates ground truth. + +![](images/19d46518ea0a222e081981c6bbd809b9d405903d5b122d2521587282c3a8265c.jpg) +$\gamma = 1.0$ a photo of the +small coffee. +$\alpha = {2.0}$ ; a coffee in a video +， +game: +1 +video game. +LM: a banana with a +cup of coffee +G1: A close up of a banana next to a cup with liquid. + +![](images/c40f15b728b5e5805118f1fd22d4a71ff411ef01445d1c39d38b5c9471bdd159.jpg) +$\gamma = 1.0$ a photo of the large +windsurfing. +$\gamma = 3.0$ : windsurfing in tarifa +$x = 3.0$ : windsurfing walls +pors 1300x1034 +1M +EM. a Win +water +GT: A man riding a +wind sail in the +ocean filled with +waves + +![](images/4eff774343483a48497c359ed640f66e928da88211ab34e660cc43d94972ed83.jpg) +$\gamma = 1.0$ a photo of the large +giraffe. +$\alpha = 2.0$ a piraffe in a video +game +gamer: +1 +photo 1 +LM: a giraffe standing in a tall tower. +in a tall tree +GI: A giraffe in a dry +savannah with dry +shrubs + +# 5. Conclusion + +Our study indicates that it is possible to substantially improve the extent to which generated captions uniquely describe the images they correspond to, raising questions regarding the goal of image captioning and how it should be evaluated. As it is conventionally formulated, image captioning aims not to provide text that can substitute for an image, but to write the text that a human annotator would have written. This formulation penalizes captions that are more descriptive than ground truth, even when a human might prefer them. On the other hand, treating image captioning as a problem of generating a caption that lies close to the image in the embedding space of an image-text model is + +also inadequate, because captions that lie close to the image need not be grammatical and may contain gibberish. Our proposed methods leveraging classifier-free guidance and language model guidance modulate the trade-offs between these two goals, as captured by various reference-based and reference-free metrics. + +There are several possible extensions to our work. First, our present experiments use only greedy decoding. Although greedy decoding appears to perform reasonably well in our setup, it may be suboptimal for LM guidance with prompts that impose greater structure upon the captions. If the LM is prompted to output either "there is a person in this image" or "there is no person this image", greedy decoding is likely to fail even if the captioner properly scores the two possible captions, because when choosing between the tokens "a" and "no", the captioner has no knowledge of the structure that the LM will impose on future tokens. Since beam search could explore both tokens, it may offer better results in this scenario. Second, our method could be combined with RL-based methods to increase similarity in a contrastive embedding space, which may further improve retrieval performance and CLIPScore. Finally, with a perfect captioning model, $p(\mathrm{image}|\mathrm{caption})$ should increase with $\gamma$ . However, in practice we find that $\gamma > 3$ leads to a decrease in retrieval performance. This discrepancy suggests that the difference between the conditional and unconditional model distributions may be a noisy estimator of the pointwise mutual information. Although selecting $\gamma$ is one way to regularize this estimator, there may also be strategies to regularize $p(\mathrm{x}|\mathrm{y}) / p(\mathrm{x})$ at training time. + +# Acknowledgements + +We thank Kevin Clark, David Fleet, Geoffrey Hinton, and the rest of the Google DeepMind Toronto team for inspiration, comments, and discussion. + +# References + +[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2 +[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2, 8 +[4] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022. 5 +[5] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. $ic^3$ : Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 2 +[6] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with CLIP reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, Seattle, United States, July 2022. Association for Computational Linguistics. 2, 5 +[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. PaLM: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 4 +[8] Katherine Crowson. You can apply a similar trick to classifier-free guidance to autoregressive transformers to sample from a synthetic "super-conditioned" distribution. https://twitter.com/RiversHaveWings/status/1478093658716966912, 2022. 3 +[9] Béatrice Daille. Approche mixte pour l'extraction automatique de terminologie: statistiques lexicales et filtrres linguistiques. PhD thesis, Ph.D. thesis, Université Paris 7, 1994. 3 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language + +Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. 3 +[11] Jacob Devlin, Hao Cheng, Hao Fang, Saurabh Gupta, Li Deng, Xiaodong He, Geoffrey Zweig, and Margaret Mitchell. Language models for image captioning: The quirks and what works. arXiv preprint arXiv:1505.01809, 2015. 2 +[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2, 3 +[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics. 2, 4 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3 +[15] Micah Hodosh, Peter Young, and Julia Hockenmaier. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853-899, 2013. 6 +[16] Yupan Huang, Hongwei Xue, Bei Liu, and Yutong Lu. Unifying multimodal transformer for bi-directional image and text generation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1138-1147, 2021. 2, 5 +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 4, 7 +[18] Andrej Karpathy and Fei-Fei Li. Deep visual-semantic alignments for generating image descriptions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3128-3137. IEEE Computer Society, 2015. 4 +[19] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander R Fabbri, Yejin Choi, and Noah A Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. arXiv preprint arXiv:2112.04139, 2021. 5 +[20] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Morrison, Ronan Le Bras, Yejin Choi, and Noah A Smith. Transparent human evaluation for image captioning. arXiv preprint arXiv:2111.08940, 2021. 1, 2 +[21] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011. 6 +[22] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning. In International Conference on Learning Representations, 2022. 2, 8 + +[23] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 1 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2 +[25] Annika Lindh, Robert J Ross, Abhijit Mahalunkar, Giancarlo Salton, and John D Kelleher. Generating diverse and meaningful captions. In International Conference on Artificial Neural Networks, pages 176-187. Springer, 2018. 2 +[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2 +[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 2, 8 +[28] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 1 +[29] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 2 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4 +[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2 +[34] Noam Rotstein, David Bensaid, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models to fuse visual data into enriched image captions. arXiv preprint arXiv:2305.17718, 2023. 2 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed + +Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems. 2 +[36] Guillaume Sanchez, Honglu Fan, Alexander Spangher, Elad Levi, Pawan Sasanka Ammanamanchi, and Stella Biderman. Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806, 2023. 3 +[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 4, 6 +[38] David R So, Wojciech Manke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V Le. Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668, 2021. 4 +[39] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 2, 8 +[40] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 4 +[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 2, 5, 8 +[42] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 1 +[43] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015. 2 +[44] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 2 +[45] Shitong Xu. Clip-diffusion-lm: Apply diffusion model on image captioning. arXiv preprint arXiv:2210.04559, 2022. 2 +[46] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 2 +[47] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3, 4, 6, 14 +[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin + +fei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3 +[49] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104–12113, 2022. 4, 7 +[50] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 5 +[51] Youuyuan Zhang, Jiuniu Wang, Hao Wu, and Wenjia Xu. Distinctive image captioning via clip guided group optimization. arXiv preprint arXiv:2208.04254, 2022. 2, 5 +[52] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang, Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint arXiv:2211.11694, 2022. 2 \ No newline at end of file diff --git a/2023/Guiding Image Captioning Models Toward More Specific Captions/images.zip b/2023/Guiding Image Captioning Models Toward More Specific Captions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da42e00d84d198a550513d10465c5df5a263d224 --- /dev/null +++ b/2023/Guiding Image Captioning Models Toward More Specific Captions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25551074162ac128854028957a0efa0a46ad65c2d4843227f3554c2782a71dd0 +size 717509 diff --git a/2023/Guiding Image Captioning Models Toward More Specific Captions/layout.json b/2023/Guiding Image Captioning Models Toward More Specific Captions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..045a48ebabfa292a6c138040c7a0261932b3cac4 --- /dev/null +++ b/2023/Guiding Image Captioning Models Toward More Specific Captions/layout.json @@ -0,0 +1,10434 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "text", + "content": "Guiding image captioning models toward more specific captions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 142, + 189, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 142, + 189, + 156 + ], + "spans": [ + { + "bbox": [ + 103, + 142, + 189, + 156 + ], + "type": "text", + "content": "Simon Kornblith" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 157, + 225, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 157, + 225, + 172 + ], + "spans": [ + { + "bbox": [ + 129, + 157, + 225, + 172 + ], + "type": "text", + "content": "Google DeepMind" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 143, + 277, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 143, + 277, + 156 + ], + "spans": [ + { + "bbox": [ + 235, + 143, + 277, + 156 + ], + "type": "text", + "content": "Lala Li" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 157, + 321, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 157, + 321, + 171 + ], + "spans": [ + { + "bbox": [ + 249, + 157, + 321, + 171 + ], + "type": "text", + "content": "2Apple AI/ML" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 143, + 388, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 143, + 388, + 157 + ], + "spans": [ + { + "bbox": [ + 323, + 143, + 388, + 157 + ], + "type": "text", + "content": "Zirui Wang" + }, + { + "bbox": [ + 323, + 143, + 388, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 157, + 476, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 157, + 476, + 172 + ], + "spans": [ + { + "bbox": [ + 347, + 157, + 476, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 347, + 157, + 476, + 172 + ], + "type": "text", + "content": "University of Washington" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 429, + 142, + 504, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 142, + 504, + 157 + ], + "spans": [ + { + "bbox": [ + 429, + 142, + 504, + 157 + ], + "type": "text", + "content": "Thao Nguyen" + }, + { + "bbox": [ + 429, + 142, + 504, + 157 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 198, + 191, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 198, + 191, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 198, + 191, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 224, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 547 + ], + "type": "text", + "content": "Image captioning is conventionally formulated as the task of generating captions for images that match the distribution of reference image-caption pairs. However, reference captions in standard captioning datasets are short and may not uniquely identify the images they describe. These problems are further exacerbated when models are trained directly on image-alt text pairs collected from the internet. In this work, we show that it is possible to generate more specific captions with minimal changes to the training process. We implement classifier-free guidance [14] for an autoregressive captioning model by fine-tuning it to estimate both conditional and unconditional distributions over captions. The guidance scale applied at decoding controls a trade-off between maximizing " + }, + { + "bbox": [ + 46, + 224, + 290, + 547 + ], + "type": "inline_equation", + "content": "p(\\text{caption} | \\text{image})" + }, + { + "bbox": [ + 46, + 224, + 290, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 224, + 290, + 547 + ], + "type": "inline_equation", + "content": "p(\\text{image} | \\text{caption})" + }, + { + "bbox": [ + 46, + 224, + 290, + 547 + ], + "type": "text", + "content": ". Compared to standard greedy decoding, decoding with a guidance scale of 2 substantially improves reference-free metrics such as CLIPScore (0.808 vs. 0.775) and caption→image retrieval performance in the CLIP embedding space (recall@1 44.6% vs. 26.5%), but worsens standard reference-based captioning metrics (e.g., CIDEr 78.6 vs 126.1). We further explore the use of language models to guide the decoding process, obtaining small improvements over the Pareto frontier of reference-free vs. reference-based captioning metrics that arises from classifier-free guidance, and substantially improving the quality of captions generated from a model trained only on minimally curated web data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 570, + 127, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 127, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 127, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 590, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 685 + ], + "type": "text", + "content": "Image captioning is both a difficult task for computer vision systems to perform and a difficult task to evaluate. Although automated captioning metrics rank the best captioning systems higher than humans, human raters still show a strong preference for human-generated captions [20], suggesting shortcomings in both captioning models and metrics. One shortcoming relates to the lack of specificity in generated captions. Conventional maximum likelihood-" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 198, + 426, + 313 + ], + "blocks": [ + { + "bbox": [ + 310, + 198, + 426, + 313 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 426, + 313 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 426, + 313 + ], + "type": "image", + "image_path": "ac98db833f7ae2e815e3f274219bb5ef911b530cefa419804e8ec00b668373c4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 314, + 420, + 319 + ], + "lines": [ + { + "bbox": [ + 312, + 314, + 420, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 314, + 420, + 319 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 312, + 314, + 420, + 319 + ], + "type": "text", + "content": " a man riding a blue motorcycle on a dirt road" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 209, + 542, + 317 + ], + "blocks": [ + { + "bbox": [ + 313, + 320, + 423, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 320, + 423, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 423, + 326 + ], + "type": "inline_equation", + "content": "\\gamma = 1.5" + }, + { + "bbox": [ + 313, + 320, + 423, + 326 + ], + "type": "text", + "content": " a man riding a blue motorcycle on a straw bale" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 326, + 423, + 336 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 423, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 423, + 336 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 313, + 326, + 423, + 336 + ], + "type": "text", + "content": " rider on blue suzuki motorcycle near straw bales" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 337, + 423, + 347 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 423, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 423, + 347 + ], + "type": "inline_equation", + "content": "\\gamma = 3.0" + }, + { + "bbox": [ + 313, + 337, + 423, + 347 + ], + "type": "text", + "content": " racer on blue suzuki motorcycle leaning up against straw bales" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 348, + 423, + 360 + ], + "lines": [ + { + "bbox": [ + 318, + 348, + 423, + 360 + ], + "spans": [ + { + "bbox": [ + 318, + 348, + 423, + 360 + ], + "type": "text", + "content": "GT A person riding a baby blue motorcycle near haystacks" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 209, + 542, + 317 + ], + "lines": [ + { + "bbox": [ + 427, + 209, + 542, + 317 + ], + "spans": [ + { + "bbox": [ + 427, + 209, + 542, + 317 + ], + "type": "image", + "image_path": "fc5d4f0ca1f455ed06eee5c5ae79fdd45a350774a01056ea77c2299f646d214b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 321, + 527, + 328 + ], + "lines": [ + { + "bbox": [ + 461, + 321, + 527, + 328 + ], + "spans": [ + { + "bbox": [ + 461, + 321, + 527, + 328 + ], + "type": "inline_equation", + "content": "y = 1.0" + }, + { + "bbox": [ + 461, + 321, + 527, + 328 + ], + "type": "inline_equation", + "content": "y = 2.0" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 461, + 329, + 526, + 336 + ], + "lines": [ + { + "bbox": [ + 461, + 329, + 526, + 336 + ], + "spans": [ + { + "bbox": [ + 461, + 329, + 526, + 336 + ], + "type": "inline_equation", + "content": "y = 1,2" + }, + { + "bbox": [ + 461, + 329, + 526, + 336 + ], + "type": "inline_equation", + "content": "y = 3,0" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 461, + 337, + 526, + 345 + ], + "lines": [ + { + "bbox": [ + 461, + 337, + 526, + 345 + ], + "spans": [ + { + "bbox": [ + 461, + 337, + 526, + 345 + ], + "type": "inline_equation", + "content": "\\gamma = 1.5" + }, + { + "bbox": [ + 461, + 337, + 526, + 345 + ], + "type": "inline_equation", + "content": "\\gamma = 4.0" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "text", + "content": "Figure 1. Using classifier-free guidance " + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "inline_equation", + "content": "(\\gamma > 1)" + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "text", + "content": " results in more specific captions that are farther from the reference distribution. Left: Example of captions generated at different guidance scales for a single image. Right: Caption " + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "text", + "content": " image recall@1 with CLIP ViT-B/32 vs. CIDEr score, for captions generated with different guidance scales " + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "text", + "content": " on MS-COCO. Higher scales improve retrieval accuracy at the expense of CIDEr." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "content": "based image captioning models attempt to generate captions such that the " + }, + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "inline_equation", + "content": "p(\\text{caption} | \\text{image})" + }, + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "content": " is high. However, captions from the ground truth distribution are often nonspecific, e.g., human annotators will usually describe a German Shepard only as a dog. Moreover, previous work has emphasized \"reference-based\" captioning metrics that measure the match between generated captions and human-provided ground truth captions [28, 23, 42]. These metrics intrinsically penalize captions that are more specific than ground truth." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "In this work, we explore strategies to guide image captioning models to produce more specific captions by modifying the decoding distribution, and explore the tradeoffs in captioning metrics that result. We first investigate the application of classifier-free guidance (CFG) [14] to image captioning with autoregressive models. Classifier-free guidance increases " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "p(\\text{image|caption})" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " at the expense of " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "p(\\text{caption|image})" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": ". Although CFG hurts reference-based image captioning metrics such as BLEU [28], ROUGE [23], and CIDEr [42], it improves \"reference-free\" metrics that measure captions' specificity via the similarity between the" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 172, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 172, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 172, + 703 + ], + "type": "text", + "content": "*Work performed while at Google." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 59, + 703, + 226, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 226, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 226, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 703, + 226, + 712 + ], + "type": "text", + "content": " Work performed as a student researcher at Google." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15259" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "image and the generated caption in the embedding space of image-text models [13] or caption " + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": " image retrieval performance. Qualitatively, we find that captions generated with CFG are more specific than both the ground truth captions and captions generated without CFG, but they are less grammatical, particularly at high CFG scales." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "content": "Beyond classifier-free guidance, we experiment with guiding image captioning models using the probability distribution obtained from a few shot-prompted language model (LM). We find that using a language model to guide a captioning model trained on MS-COCO [24] with descriptive manually written captions can allow it to achieve slightly better trade-offs between reference-free vs. reference-based captioning metrics than those observed with CFG. LM guidance also substantially improves the captions produced by a model trained exclusively on minimally curated web data. Although this model achieves a CIDEr score of only 21.8 without guidance, this CIDEr score improves to 57.4 when guided by a language model prompted with 20 captions from the MS-COCO training set. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 329, + 286, + 478 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 329, + 286, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 286, + 364 + ], + "type": "text", + "content": "- We propose two strategies to guide image captioning models to produce more specific captions: classifier-free guidance and language model guidance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 368, + 286, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 286, + 415 + ], + "type": "text", + "content": "- We demonstrate that classifier-free guidance yields captions that are closer to the corresponding image in the embedding space of image-text models, but are farther from human-provided reference captions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 286, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 286, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 286, + 478 + ], + "type": "text", + "content": "- We show that language model guidance can alter caption styles, substantially improving captions produced by a model trained only on minimal curated web data and marginally improving the trade-off between captioning metrics observed with classifier-free guidance." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 489, + 132, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 132, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 132, + 502 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "Measuring specificity of captions. Early work using neural networks for image captioning found that models have a propensity to regurgitate captions from their training data, and as a result, the generated captions are not descriptive enough to uniquely identify images [43, 11]. To address this shortcoming, Lindh et al. [25] proposed to use caption " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " image recall with an image retrieval model to examine whether images can be retrieved from generated captions, and further attempt to differentiate through this retrieval process to train a captioning model. Their approach marginally improves retrieval accuracy, but worsens reference-based captioning metrics. More recent work has adopted approaches to evaluate the specificity of captions based on the CLIP image-text model [30]. Hessel et al. [13] propose CLIPScore, an image captioning metric based on the cosine similarity between CLIP embeddings of the image and the generated caption. Kasai et al. [20] re" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "port that CLIPScore-based metrics align better with human judgments compared to reference-based captioning metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 98, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 545, + 266 + ], + "type": "text", + "content": "Improving specificity of captions. Recent work has attempted to train caption models to directly maximize the similarity of captions with corresponding images in the CLIP embedding space using reinforcement learning. This work has found that adding such losses worsens standard reference-based captioning metrics but improve similarity and retrieval in the CLIP embedding space [16, 6, 51], similar to our observations regarding CFG. Wen et al. [44] attempt to generate prompts for text-to-image generative models that correspond to specific images without a captioning model, by directly optimizing the similarity between the text and image in the CLIP embedding space using a gradient-based discrete optimization procedure, but the resulting text is not grammatical." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 267, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 434 + ], + "type": "text", + "content": "Other work has attempted to generate more descriptive captions through different means. Dense captioning [46] aims to detect and caption all objects in an image, but concatenating all of these captions leads to long and unnatural captions, whereas CFG produces single-sentence captions. The Localized Narratives dataset [29] contains visually grounded captions for MS-COCO images collected through voice annotation. These captions are substantially more descriptive than the captions in the MS-COCO dataset and can be used for model training. Two recent papers propose to combine multiple captions from a single model [5] or outputs of different vision models [34], and then combine them using a language model. The resulting captions are longer, but achieve greater caption→image recall." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 604 + ], + "type": "text", + "content": "Captioning from uncurated data. In Section 4.2, we explore the use of LM guidance for captioning with access to uncurated image-text data from the web and a small number of captions but not images from the target distribution. This setting, which does not rely on aligned images and captions from the target distribution, is often referred to as \"zero-shot\" captioning, and previous work has pursued a number of alternative approaches. Flamingo [3] and CM3 [1] perform zero-shot captioning by pretraining on interleaved image/text data. MAGIC [39] and ZeroCap [41] generate captions using a combination of guidance from CLIP and a large language model. Other recent work adapts CLIP to perform captioning by training a text decoder using only captions, with no corresponding images [27, 22]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Classifier-free guidance. CFG is widely used in diffusion-based and autoregressive text-to-image models [26, 32, 35, 33, 12, 48]. Because of the popularity of the combination of CFG and diffusion, previous work that has performed image captioning with diffusion models has also examined the use of CFG. This work finds either no benefit to using CFG [45] or a small and inconsistent benefit that appears to vary with minor changes in training settings [52]. However, these studies do not seek to generate" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15260" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "more specific captions, and thus measure only reference-based captioning metrics, which we likewise find do not benefit from CFG. Concurrently with out work, [36] propose to use classifier-free guidance to improve prompt following in large language models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 107, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 107, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 107, + 156 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 279, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 279, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 279, + 176 + ], + "type": "text", + "content": "3.1. Classifier-free guidance for image captioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "text", + "content": " be an image caption and " + }, + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "text", + "content": " be the corresponding image. A standard captioning model aims to model the likelihood " + }, + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "inline_equation", + "content": "p(\\mathrm{x}|\\mathrm{y})" + }, + { + "bbox": [ + 46, + 181, + 286, + 229 + ], + "type": "text", + "content": ", factorized autoregressively in terms of the probability of each token given previous tokens" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 232, + 287, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 232, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 80, + 232, + 287, + 246 + ], + "type": "interline_equation", + "content": "p (\\mathrm {x} | \\mathrm {y}) = p \\left(\\mathrm {x} _ {n} \\mid \\mathrm {x} _ {n - 1}, \\dots \\mathrm {x} _ {1}, \\mathrm {y}\\right) \\dots p \\left(\\mathrm {x} _ {1} \\mid \\mathrm {y}\\right). \\tag {1}", + "image_path": "9536fcfe73c7d80ab52dc8bd43f71245dc1bddfe4ad9a8ebd5e40fcbcc8b58ae.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 249, + 286, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 286, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 286, + 333 + ], + "type": "text", + "content": "The network is trained so that its output distribution " + }, + { + "bbox": [ + 46, + 249, + 286, + 333 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\mathbf{x}_n|\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y}) \\stackrel{\\mathrm{def}}{=} \\text{softmax}(f_{\\theta}(\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y}))" + }, + { + "bbox": [ + 46, + 249, + 286, + 333 + ], + "type": "text", + "content": " approximates " + }, + { + "bbox": [ + 46, + 249, + 286, + 333 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_n|\\mathbf{x}_{n-1}, \\ldots, \\mathbf{x}_1, \\mathbf{y})" + }, + { + "bbox": [ + 46, + 249, + 286, + 333 + ], + "type": "text", + "content": ". At inference time, one typically uses beam search or greedy decoding to produce a caption that has a particularly high probability. In this work, we use greedy decoding because it is the more common choice and it is also simpler to implement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 334, + 286, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 286, + 357 + ], + "type": "text", + "content": "Classifier-free guidance (CFG) [14] modifies the inference procedure so it instead aims to achieve high values of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 361, + 287, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 361, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 73, + 361, + 287, + 388 + ], + "type": "interline_equation", + "content": "l _ {\\boldsymbol {\\theta}, \\gamma} (\\mathrm {x}, \\mathrm {y}) \\stackrel {\\text {d e f}} {=} p (\\mathrm {x}) \\left(\\frac {p (\\mathrm {x} | \\mathrm {y})}{p (\\mathrm {x})}\\right) ^ {\\gamma} \\propto p (\\mathrm {x}) p (\\mathrm {y} | \\mathrm {x}) ^ {\\gamma}, \\tag {2}", + "image_path": "473d037268261d8dbea18d103ae6d09b4cc0cf5837732f5572114fc67477056a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": "where proportionality holds because " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x}) = p(\\mathrm{y}|\\mathrm{x}) / p(\\mathrm{y})" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "p(\\mathrm{y})" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " is fixed. The parameter " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " is called the guidance scale and controls the trade-off between maximization of " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "p(\\mathrm{x}|\\mathrm{y})" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "p(\\mathrm{y}|\\mathrm{x})" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "l_{\\theta, \\gamma}(\\mathrm{x}, \\mathrm{y}) = p(\\mathrm{x}|\\mathrm{y})" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " and guidance has no effect. Setting " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " inflates the probability of the image given the caption " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "p(\\mathrm{y}|\\mathrm{x})" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " relative to the unconditional probability of the caption " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "p(\\mathrm{x})" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "text", + "content": "Ho and Salimans [14] originally proposed CFG in the context of diffusion models, which estimate the score functions " + }, + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "inline_equation", + "content": "\\nabla \\log p(\\mathbf{x}|\\mathbf{y})" + }, + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "inline_equation", + "content": "\\nabla \\log p(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "text", + "content": ". Although " + }, + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "inline_equation", + "content": "l_{\\theta,\\gamma}(\\mathbf{x},\\mathbf{y})" + }, + { + "bbox": [ + 46, + 474, + 286, + 558 + ], + "type": "text", + "content": " factorizes autoregressively, it is not a normalized probability distribution, so it is not entirely clear how one should sample tokens when performing autoregressive generation. Crowson [8] suggested to sample from" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 561, + 286, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 561, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 561, + 286, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {q} _ {\\boldsymbol {\\theta}, \\gamma} (\\mathrm {x} _ {n} | \\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathrm {y}) \\stackrel {\\mathrm {d e f}} {=} \\operatorname {s o f t m a x} (f _ {\\boldsymbol {\\theta}} (\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathbf {0}) \\\\ + \\gamma \\left(f _ {\\boldsymbol {\\theta}} \\left(\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathrm {y}\\right) - f _ {\\boldsymbol {\\theta}} \\left(\\mathrm {x} _ {n - 1}, \\dots , \\mathrm {x} _ {1}, \\mathbf {0}\\right)\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "8ba3e3b0cbffe8998a34d65dff3c7b33d42aba2e2a73ec29f81f3189da4d8f6e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x}_{n - 1},\\ldots ,\\mathbf{x}_1,\\mathbf{0})" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " are logits generated by the model without conditioning, usually by passing zeros in place of the conditioning information. This formulation has been successfully applied in autoregressive image models [12, 48]. In our experiments, since we decode greedily, i.e., at each step we take the token that maximizes " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{q}_{\\theta ,\\gamma}(\\mathrm{x}_n|\\mathrm{x}_{n - 1},\\dots ,\\mathrm{x}_1,\\mathrm{y})" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " and thus " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "l_{\\theta ,\\gamma}(\\mathrm{x},\\mathrm{y})" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": ", any form of normalization of the per-step sampling distribution would produce the same captions. We provide pseudocode in Appendix A.1. Crucially, the only change required to train" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "ing is to learn an unconditional model of captions alongside the image-conditional model, which we accomplish by randomly masking the image embeddings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 118, + 451, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 118, + 451, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 451, + 131 + ], + "type": "text", + "content": "3.2. Language model guidance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 137, + 545, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 161 + ], + "type": "text", + "content": "Inspired by classifier-free guidance, we consider language model (LM) guidance, which attempts to maximize" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 164, + 545, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 164, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 361, + 164, + 545, + 190 + ], + "type": "interline_equation", + "content": "l _ {\\boldsymbol {\\theta}, \\gamma} ^ {\\prime} (\\mathrm {x}, \\mathrm {y}) \\stackrel {\\text {d e f}} {=} q (\\mathrm {x}) \\left(\\frac {p (\\mathrm {x} | \\mathrm {y}) ^ {\\alpha}}{p (\\mathrm {x}) ^ {\\beta}}\\right), \\tag {4}", + "image_path": "bef0b8c77285c7e73826395bf2e65854ae04df5d6b47ce59428c8fabb5fb01d9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "p(x|y)" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " are obtained from a captioning model as in CFG but " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " is obtained from a language model that was trained independently (but with the same vocabulary) on a large text corpus. The quantity " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x}) = p(\\mathrm{x},\\mathrm{y}) / (p(\\mathrm{x})p(\\mathrm{y}))" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " measures the strength of the association between a caption and an image; its logarithm is the pointwise mutual information (PMI). LM guidance relies on the assumption that, even for large shifts in the prior distribution of captions " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": ", the shift in PMI will be small. Empirically, we obtain better results by allowing different exponents for the numerator and denominator, with " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\alpha > \\beta" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": ". This decoupling resembles " + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{PMI}^k" + }, + { + "bbox": [ + 304, + 193, + 545, + 361 + ], + "type": "text", + "content": " [9], which reduces the bias of PMI toward rare associations. We provide a more detailed derivation in Appendix A.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "text", + "content": "We investigate two applications of LM guidance. First, we combine a captioning model fine-tuned on MS-COCO with a LM prompted with manually written descriptive captions to alter the style of the captions the model produces. The manually written prompts are shown in Appendix A.4. Second, we combine a captioning model trained only on low-quality web data with a LM prompted with varying numbers of examples from the MS-COCO training set to evaluate the ability of LM guidance to elicit higher-quality captions without high-quality paired data. We randomly select a different set of captions for each minibatch of four test examples. In both cases, we separate the captions with two newlines. Because this format leads the LM to place probability mass on the newline token to end the caption, we transfer the probability mass from the newline token to the EOS token. See Appendix A.3 for pseudocode." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 563, + 424, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 424, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 424, + 575 + ], + "type": "text", + "content": "3.3. Models and training" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Our captioning model is a \"bottleneck\" variant of CoCa-Base [47], which combines a contrastive loss with a captioning loss to simultaneously learn aligned image and text embeddings as well as a captioner. The architecture consists of an image encoder, a unimodal text decoder, and a multimodal text decoder, each of which are Transformers with 12 layers, 768 hidden dimensions, an MLP of size 3072, and 12 self-attention heads, matching " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{BERT}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " [10] and GPT-1 [31]. The image encoder is a ViT-B/18 that processes " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "288 \\times 288" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " input and produces an embedding such that images are embedded close to their corresponding text." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15261" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "CoCa's multimodal text decoder processes the representations of the image encoder to produce a caption. Whereas [47] conditions the multimodal text decoder using crossattention to pooled representations, our bottleneck variant uses only the contrastive image embedding. Appendix A.5 shows a diagram of the resulting architecture. We adopt this bottleneck variant because of its simplicity and the conceptual appeal: When CFG is used, the captioner's role is to invert the image embedding, providing a caption that, when embedded by the text encoder, lies close to it. However, as we show in Appendix B.1, this choice of the bottleneck model is not critical, and CFG is equally effective with the standard CoCa architecture with attention pooling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "text", + "content": "For CFG experiments, we pretrain our model on an image-text dataset comprising images from the JFT-5B dataset [40, 49] paired with their corresponding label names substituted into a randomly selected prompt from the list provided by Radford et al. [30], web images paired with noisy alt text from the ALIGNN dataset [17], and a small amount of data from other sources. We do not mask conditioning information during pretraining. We then fine-tune on the combined MS-COCO train and Karpathy validation splits [18] using Adam with batch size 128. We linearly warm up to a learning rate of " + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "text", + "content": " over the first 1,000 steps and linearly decay to zero over the rest of training. We vary " + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\{1.0, 1.2, 1.5, 2.0, 3.0, 4.0\\}" + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "text", + "content": ", conditioning masking proportion in " + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\{0.0, 0.25, 0.5, 0.75\\}" + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "text", + "content": ", and numbers of steps in " + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\{5,000, 10,000, 20,000, 50,000\\}" + }, + { + "bbox": [ + 46, + 228, + 289, + 479 + ], + "type": "text", + "content": ". We report results from the model trained for 20,000 steps with masking proportion 0.5, which achieves near-optimal results, in Tables 1 and B.4, and sample example captions from it. To ensure that results generalize across datasets, we also experiment with a model fine-tuned on Conceptual Captions [37] for 100,000 steps with masking proportion 0.5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 480, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 289, + 624 + ], + "type": "text", + "content": "For LM guidance experiments, we pretrain on the JFT-5B and ALIGN datasets, again following the recipe of [47]. For zero-shot captioning experiments, we fine-tune this model on the same datasets for an additional 50,000 steps with conditioning masking proportion of 0.5 to improve our ability to sample unconditionally. For LM guidance on MS-COCO, we first fine-tune on ALIGN, JFT-5B images backcaptioned by an MS-COCO fine-tuned CoCa2B model, and a small amount of internal data before finetuning on MS-COCO. Our language model is a variant of Primer [38] with 2 billion parameters, trained on a similar dataset to that used to train PaLM [7]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 632, + 122, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 122, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 122, + 643 + ], + "type": "text", + "content": "3.4. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 651, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 675 + ], + "type": "text", + "content": "We adopt the standard reference-based captioning metrics BLEU-4, METEOR, ROUGE, and CIDEr, as well" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": "as reference-free captioning metrics based on CLIP ViT-B/32 [30]. The first reference-free captioning metric is CLIPScore [13], which is defined as " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - S}(c,v) = 2.5\\cdot \\max (\\cos (c,v),0)" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": " are the CLIP embeddings of the caption and image respectively. The second reference-free metric measures the accuracy with which we can retrieve an image from the generated caption within a given test split by taking the " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": " nearest neighbors of the caption in the CLIP embedding space. Because recall@ " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "k > 1" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": " is highly correlated with recall@1 (R@5: " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "r = 0.99" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": ", R@10: " + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "inline_equation", + "content": "r = 0.98" + }, + { + "bbox": [ + 306, + 72, + 547, + 289 + ], + "type": "text", + "content": "), we plot only recall@1. We additionally report RefOnlyCLIP-S, a reference-based metric that uses the CLIP text encoder to compute the similarity of CLIP embeddings of the generated captions with embeddings of ground truth captions, and RefCLIP-S, which takes the average of the per-image harmonic means of CLIP-S and RefOnlyCLIP-S [13]. Unless otherwise stated, all evaluation is performed on the MS-COCO Karpathy test split [18]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 300, + 359, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 359, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 359, + 313 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 321, + 440, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 321, + 440, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 440, + 334 + ], + "type": "text", + "content": "4.1. Classifier-free guidance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 341, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 484 + ], + "type": "text", + "content": "We first investigate the trade-off between reference-based and reference-free image captioning metrics as a function of guidance scale. Because different guidance scales and metrics could conceivably benefit from different fine-tuning hyperparameter combinations, we plot all results from our hyperparameter grid in Figure 2. Although standard greedy decoding " + }, + { + "bbox": [ + 304, + 341, + 547, + 484 + ], + "type": "inline_equation", + "content": "(\\gamma = 1.0)" + }, + { + "bbox": [ + 304, + 341, + 547, + 484 + ], + "type": "text", + "content": " produces the highest CIDEr, METEOR, ROUGE, and BLEU-4 scores, higher guidance weights consistently yield higher values of reference-free captioning metrics. In particular, " + }, + { + "bbox": [ + 304, + 341, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\gamma = 3.0" + }, + { + "bbox": [ + 304, + 341, + 547, + 484 + ], + "type": "text", + "content": " offers both the best caption→image recall and the best CLIP-Score." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 548, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 548, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 548, + 605 + ], + "type": "text", + "content": "Table 1 compares our results, obtained from a single model evaluated at different guidance scales, with previous work that reports either CLIPScore or CLIP ViT-B/32 caption " + }, + { + "bbox": [ + 304, + 485, + 548, + 605 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 485, + 548, + 605 + ], + "type": "text", + "content": " image retrieval performance. Although our model is trained with standard cross-entropy loss rather than a CLIP-based loss and our pretraining dataset is distinct from CLIP's, sampling from our model with CFG yields higher CLIPScores than all other models trained without CLIP-based losses, and better CLIP caption " + }, + { + "bbox": [ + 304, + 485, + 548, + 605 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 485, + 548, + 605 + ], + "type": "text", + "content": " image retrieval even when compared with models that use CLIP-based losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "We present examples of captions generated at different CFG scales in Figure 3. Higher CFG strengths lead to more descriptive captions. At " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ", the central object in the top left image is described as a \"car\" as in the ground truth caption, whereas at " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\gamma > 1.0" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " it is a \"station wagon.\" Similarly, at low CFG strengths, the birds in the center image are described simply as \"birds,\" whereas at " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " they become \"crested cranes.\" However, at " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\gamma = 3.0" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ", captions clearly become less grammatical, containing repeated" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "text", + "content": "1We find that passing an all-zero image embedding to the pretrained model yields samples that resemble the unconditional distribution, suggesting that it implicitly learns to model the unconditional distribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15262" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 68, + 500, + 235 + ], + "blocks": [ + { + "bbox": [ + 91, + 68, + 500, + 235 + ], + "lines": [ + { + "bbox": [ + 91, + 68, + 500, + 235 + ], + "spans": [ + { + "bbox": [ + 91, + 68, + 500, + 235 + ], + "type": "image", + "image_path": "66ceb179a19a7d97a20c9c8a9512f0e8d35e690abc03d9acd40efaac31ae6f19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 242, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 547, + 298 + ], + "type": "text", + "content": "Figure 2. Classifier-free guidance controls a trade off between reference-free and reference-based captioning metrics. Each point reflects a model trained with a different hyperparameter combination; each color represents a " + }, + { + "bbox": [ + 46, + 242, + 547, + 298 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 242, + 547, + 298 + ], + "type": "text", + "content": " value used to decode. Models are evaluated with different guidance scales " + }, + { + "bbox": [ + 46, + 242, + 547, + 298 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 242, + 547, + 298 + ], + "type": "text", + "content": ", using reference-free captioning metrics based on CLIP ViT-B/32 (y-axes; top: CLIPScore, bottom: recall@1) and reference-based captioning metrics (x-axes). The dashed line reflects the value of the reference-free captioning metric for the ground-truth captions obtained from MS-COCO." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 671, + 547, + 706 + ], + "lines": [ + { + "bbox": [ + 46, + 671, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 671, + 547, + 706 + ], + "type": "text", + "content": "Figure 3. Caption descriptiveness increases with CFG strength, but high CFG strengths produce agrammatical captions. Here we show examples of captions generated with different classifier-free guidance scales, for randomly selected images without human faces from the MS-COCO Karpathy test split. Captions labeled " + }, + { + "bbox": [ + 46, + 671, + 547, + 706 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 46, + 671, + 547, + 706 + ], + "type": "text", + "content": " are obtained without CFG; " + }, + { + "bbox": [ + 46, + 671, + 547, + 706 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 46, + 671, + 547, + 706 + ], + "type": "text", + "content": " uses CFG; GT = ground truth." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 304, + 542, + 483 + ], + "blocks": [ + { + "bbox": [ + 52, + 304, + 542, + 483 + ], + "lines": [ + { + "bbox": [ + 52, + 304, + 542, + 483 + ], + "spans": [ + { + "bbox": [ + 52, + 304, + 542, + 483 + ], + "type": "table", + "html": "
ModelReference-Based MetricsReference-Free MetricsRefCLIP-S
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10
Models trained with CLIP features or losses:
CLIP-Captioner [4]38.729.358.6126.00.8110.7540.814
UMT-BITG [16]37.328.257.9122.60.772
X-LAN+SCST+GEG [51]36.528.757.5121.728.150.367.2
CIDEr + CLIP-S Reward [6]37.728.858.3124.60.77224.450.263.1
CLIP-S Reward [6]6.218.731.611.20.86042.571.682.2
ZeroCap [41]2.611.514.60.870.79
Models trained without access to CLIP:
UMT-BITG w/o CLIP loss [16]37.628.358.1122.50.725
VinVL-large [50]41.030.959.4*140.90.91*0.78*0.84*
Ours (γ = 1.0)36.130.558.2126.10.9000.77526.551.964.10.830
Ours (γ = 1.2)35.130.057.5124.10.8990.78531.357.469.30.835
Ours (γ = 1.5)31.528.454.4113.20.8910.79636.664.075.00.838
Ours (γ = 2.0)20.923.343.078.60.8620.80844.671.781.70.831
Ours (γ = 3.0)11.517.129.441.70.8200.80849.475.784.70.811
Ours (γ = 4.0)6.512.318.417.30.7660.78244.771.380.90.771
", + "image_path": "c61f430cc7fd0f6ee4d9217300ecca01f8c4ce44f97cfe57dc0488ae4c760fce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 518, + 538, + 657 + ], + "blocks": [ + { + "bbox": [ + 46, + 490, + 544, + 512 + ], + "lines": [ + { + "bbox": [ + 46, + 490, + 544, + 512 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 544, + 512 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of our approach with results from previous work that reports CLIP-based metrics. For VinVL-large, * indicates metrics from [19]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 518, + 538, + 657 + ], + "lines": [ + { + "bbox": [ + 52, + 518, + 538, + 657 + ], + "spans": [ + { + "bbox": [ + 52, + 518, + 538, + 657 + ], + "type": "table", + "html": "
γ=1.0: a car with a surfboard on top of it parked next to a car\nγ=1.5: a vintage station wagon with a surfboard on top\nγ=2.0: a antique station wagons and a car buick station wagon\nγ=3.0: buick woody wyddo staitonwagon and surf green station wagon parked in front of car show establishment\nGT: An old green and brownγ=1.0: a knife is sitting on a cutting board next to a sliced carrot\nγ=1.5: a knife is sitting on a cutting board next to an orange\nγ=2.0: knife sitting on cutting board next to whole one\nγ=3.0: knife sitting on cutting board next to misschappi carrot on cutting board\nGT: A knife sticking out of the side of a block of cheese.γ=1.0: a herd of sheep grazing on a grass covered road\nγ=1.5: sheep grazing on a high-way with a truck on the road\nγ=2.0: sheep graze on freeway medians where grass is grown\nγ=3.0: grazing trucks blocking sheep on roadway grazing grass\nGT: A large herd of sheep are grazing by the busy road.γ=1.0: two birds standing in a cage with their heads in the air\nγ=1.5: two birds standing inside of a cage in a zoo\nγ=2.0: two crested cranes inside a wire cage\nγ=3.0: crested tantalus cranes caged together in bird-cage enclosure\nGT: Two birds who are looking out of the cage they are in.γ=1.0: a view of a city with a clock tower in the background\nγ=1.5: a city with steeples and trees and buildings\nγ=2.0: spires of churches line a city skyline\nγ=3.0: spires steeples buildings trees church spires and trees\nGT: A clock that is on the side of a tower.γ=1.0: a kitchen with a microwave and refrigerator\nγ=1.5: a kitchen with a mi-crowage and a refrigerator\nγ=2.0: a kitchen with red ap-pliances and white cup-boards\nγ=3.0: appliances sit in a small empty dingroom red and white kitchen\nGT: A kitchen that has a tile floor, a refrigerator, a microwave, and a toaster.γ=1.0: a bathroom with a large mirror and a bathtub\nγ=1.5: a bathroom with a large mirror and a bathtub\nγ=2.0: a spacious bathroom with a large mirror and a large tub\nγ=3.0: spacious bathroom with chandelier over tub mir-rors and tv\nGT: A bathroom with a tub, sinks, lights and a televi-sion.
", + "image_path": "278d2d12a045cd2f87d31f31a841670fac7bdd71fc843f943724616bb5c8f58e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15263" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "words (\"woody woody\") and nonsense words (\"misshappi\", \"dingroomy\"). Figure 4 shows captions obtained with and without CFG next to the top 5 closest images in the embedding space of CoCa 2B [47],² where it is clear that CFG adds details to captions that help to distinguish them from other captions in the test split. We provide additional examples in Appendix C." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": "To provide additional quantitative assessments of the specificity of elicited captions, we perform two additional evaluations, described further in Appendix B.2. First, we generate captions for the Stanford Dogs [21] test set, which consists of 8,580 images in total of 120 breeds of dogs, and examine their properties. Without guidance, only " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": " of captions contain one of the 120 breed names, whereas at " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "42.4\\%" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": " do. The percentage of these breed names that are correct changes little, from " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "61.7\\%" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": " without guidance to " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "58.5\\%" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": ". Second, we performed a human evaluation comparing captions of MS-COCO test set images obtained without guidance and at " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": ". We asked subjects to select the caption that is \"better\" and \"more descriptive\" or to indicate that they are both equal. When we asked these questions separately, we found that the two sets of captions are statistically indistinguishable. However, when asking both questions on the same survey, we found that captions generated without guidance are slightly \"better\" (50.5% vs. " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "46.6\\%" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "p = 0.006" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": ", binomial test) but captions generated at " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": " are \"more descriptive\" (52.7% vs. " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "45.8\\%" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "inline_equation", + "content": "p = 1 \\times 10^{-6}" + }, + { + "bbox": [ + 46, + 157, + 290, + 407 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 408, + 290, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 290, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 290, + 647 + ], + "type": "text", + "content": "To validate the reliability of our results, we further measure the impact of CFG on three additional datasets, nocaps [2], Flickr-8k [15], and Conceptual Captions (CC3M) [37], as well as with alternative retrieval models. nocaps is a test set for captioning models with objects not present in MS-COCO; Flickr-8k is a small captioning dataset collected using a different procedure than MS-COCO; and Conceptual Captions is a set of 3.3M captions collected from filtered alt-text. We fine-tune the bottleneck CoCa-Base model directly on CC3M, and use our model fine-tuned on MS-COCO to caption images on nocaps and Flickr-8K. As shown in Figure 5, we find trade-offs between reference-based and reference-free captioning metrics similar to those above. In Appendix B.3, we report reference-free captioning metrics on MS-COCO computed with two additional retrieval models: the pretrained CoCa 2B model from [47] and the fine-tuned CoCa Base model that we use to generate captions. With both models, CFG substantially increases recall, in line with results obtained with CLIP ViT-B/32." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 647, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 289, + 685 + ], + "type": "text", + "content": "Although CFG produces captions that are more successful at uniquely identifying images than decoding from the conditional distribution, caption lengths are similar for" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 547, + 423 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 423 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 423 + ], + "type": "image", + "image_path": "2fd0d825af643485344d7b9cc5c1af8faf84bc57f9bea56ace03c9b449b3f99c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "type": "text", + "content": "Figure 4. Captions generated with CFG contain specific details that improve retrieval. For each reference image (far left), we show captions at " + }, + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "type": "text", + "content": " (no guidance) and " + }, + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "type": "text", + "content": ". To the right, we show the closest images to each caption in the CoCa embedding space. Reference images are selected at random subject to the constraints that the closest image differs between " + }, + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 425, + 547, + 501 + ], + "type": "text", + "content": " values and there are no identifiable human faces." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 507, + 547, + 688 + ], + "blocks": [ + { + "bbox": [ + 306, + 507, + 547, + 688 + ], + "lines": [ + { + "bbox": [ + 306, + 507, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 547, + 688 + ], + "type": "image", + "image_path": "c2d56f207cc27833fea0be98669f6dc37a54d07155aa418bb1d081a3eb1a6c4a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 690, + 547, + 712 + ], + "lines": [ + { + "bbox": [ + 304, + 690, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 690, + 547, + 712 + ], + "type": "text", + "content": "Figure 5. CFG also yields trade-offs between captioning metrics on nocaps, Flickr-8K, and CC3M." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 289, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 289, + 715 + ], + "type": "text", + "content": "2We use CoCa 2B rather than CLIP because, quantitatively and qualitatively, it provides better retrieval results both with and without guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15264" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 98, + 70, + 237, + 141 + ], + "blocks": [ + { + "bbox": [ + 98, + 70, + 237, + 141 + ], + "lines": [ + { + "bbox": [ + 98, + 70, + 237, + 141 + ], + "spans": [ + { + "bbox": [ + 98, + 70, + 237, + 141 + ], + "type": "table", + "html": "
γWordsCharacters
1.09.6 ± 1.444.2 ± 7.2
1.29.6 ± 1.444.7 ± 7.4
1.59.4 ± 1.445.7 ± 7.8
2.09.3 ± 2.450.3 ± 18.6
3.010.7 ± 7.669.0 ± 56.1
4.019.9 ± 16.9161.2 ± 140.0
", + "image_path": "ec1eee5c9892209172572e6c8a0943f202c9b00974e2da1d2759c711ae06251b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 149, + 287, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 182 + ], + "type": "text", + "content": "Table 2. Moderate CFG scales do not substantially change caption lengths, although higher CFG scales result in longer captions. Numbers are mean " + }, + { + "bbox": [ + 46, + 149, + 287, + 182 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 46, + 149, + 287, + 182 + ], + "type": "text", + "content": " standard deviation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 76, + 191, + 255, + 289 + ], + "blocks": [ + { + "bbox": [ + 76, + 191, + 255, + 289 + ], + "lines": [ + { + "bbox": [ + 76, + 191, + 255, + 289 + ], + "spans": [ + { + "bbox": [ + 76, + 191, + 255, + 289 + ], + "type": "image", + "image_path": "eff26af2f82da7153684253b61b1e109761bee454bd2e4357f22683717325faa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 294, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 327 + ], + "type": "text", + "content": "Figure 6. Language model guidance produces captions that slightly exceed the Pareto frontier of CIDEr vs. caption " + }, + { + "bbox": [ + 46, + 294, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 294, + 287, + 327 + ], + "type": "text", + "content": " image retrieval accuracy on MS-COCO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 341, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\gamma \\in [1,2]" + }, + { + "bbox": [ + 46, + 341, + 287, + 413 + ], + "type": "text", + "content": ", as shown in Table 2. Thus, at low guidance strengths, CFG improves recall by making more efficient use of words, rather than by producing more verbose captions. Higher CFG strengths lead to longer captions but, as described above, these captions are agrammatical and contain nonsense words." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 420, + 192, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 192, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 192, + 433 + ], + "type": "text", + "content": "4.2. Language model guidance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": "We first experiment with guiding a captioning model fine-tuned on MS-COCO to produce more descriptive captions using a language model prompted with manually written prompts. We first manually wrote a prompt containing 10 descriptive captions of COCO test set images (Appendix A.4). We then sweep over " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{1,2,3,4,5,6,7,8,9,10,12,15\\}" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\beta \\in \\{0,\\alpha /4,\\alpha /2,3 / 4\\alpha ,\\alpha \\}" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": ", and compare the resulting retrieval/CIDEr trade-off to that produced by the same model with CFG. We observe that it is possible to obtain small improvements upon the Pareto frontier provided by CFG, as shown in Figure 6. With " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\alpha = 5" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\beta = -5 / 2" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": ", LM guidance achieves CLIP ViT-B/32 R@1 of " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "39.6\\%" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " and CIDEr of 114.4, whereas CFG with " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\gamma = 1.6" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " is worse on both metrics, achieving R@1 of " + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "inline_equation", + "content": "39.0\\%" + }, + { + "bbox": [ + 46, + 438, + 287, + 617 + ], + "type": "text", + "content": " and CIDEr of 109.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "type": "text", + "content": "We further experiment with using prompting to control the captioner using a manually written prompt of 25 captions in the form of \"a photo of NUMBER OBJECTS\" (e.g., \"a photo of eight apples\"; see Appendix A.4). With " + }, + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = 1" + }, + { + "bbox": [ + 46, + 618, + 287, + 689 + ], + "type": "text", + "content": ", the guided model is able to match this format and counts the number of objects in images (Figure 7)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We next investigate whether language model guidance can elicit better captions from a model trained only on" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 367, + 127 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 367, + 127 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 367, + 127 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 367, + 127 + ], + "type": "image", + "image_path": "86ea8725a151896a78c9bc2d7ca0dd77994637f50668653f9582caf42b0fdcc3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 129, + 359, + 135 + ], + "lines": [ + { + "bbox": [ + 318, + 129, + 359, + 135 + ], + "spans": [ + { + "bbox": [ + 318, + 129, + 359, + 135 + ], + "type": "text", + "content": "a photo of two dogs" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 70, + 429, + 128 + ], + "blocks": [ + { + "bbox": [ + 369, + 70, + 429, + 128 + ], + "lines": [ + { + "bbox": [ + 369, + 70, + 429, + 128 + ], + "spans": [ + { + "bbox": [ + 369, + 70, + 429, + 128 + ], + "type": "image", + "image_path": "04ff0335e5d94d7e00d59a44f0a1d6d6a969e82fbd910d6d5fd04b378bfbef3b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 129, + 419, + 135 + ], + "lines": [ + { + "bbox": [ + 380, + 129, + 419, + 135 + ], + "spans": [ + { + "bbox": [ + 380, + 129, + 419, + 135 + ], + "type": "text", + "content": "a photo of one bird" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 430, + 70, + 490, + 128 + ], + "blocks": [ + { + "bbox": [ + 430, + 70, + 490, + 128 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 490, + 128 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 490, + 128 + ], + "type": "image", + "image_path": "4dd6affcc4456e67eb59b154b72d1e6d0c652ae69d794473dc5449318f79ef18.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 129, + 485, + 135 + ], + "lines": [ + { + "bbox": [ + 435, + 129, + 485, + 135 + ], + "spans": [ + { + "bbox": [ + 435, + 129, + 485, + 135 + ], + "type": "text", + "content": "a photo of four suitcases" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 140, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 173 + ], + "type": "text", + "content": "Figure 7. Captions generated with LM guidance with a prompt of 25 captions in the form of \"a photo of NUMBER OBJECTS\". Examples are selected to show different numbers of objects." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 492, + 70, + 550, + 128 + ], + "blocks": [ + { + "bbox": [ + 492, + 70, + 550, + 128 + ], + "lines": [ + { + "bbox": [ + 492, + 70, + 550, + 128 + ], + "spans": [ + { + "bbox": [ + 492, + 70, + 550, + 128 + ], + "type": "image", + "image_path": "8a2e42dd24594938eb170c50661f9dc971aa37ccbc9747a69b8fa313483c379a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 129, + 542, + 135 + ], + "lines": [ + { + "bbox": [ + 499, + 129, + 542, + 135 + ], + "spans": [ + { + "bbox": [ + 499, + 129, + 542, + 135 + ], + "type": "text", + "content": "a photo of five sheep" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 313, + 191, + 421, + 289 + ], + "blocks": [ + { + "bbox": [ + 313, + 191, + 421, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 191, + 421, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 421, + 289 + ], + "type": "image", + "image_path": "d47a7b746a450aefdfc7122b35fd5c0bd04085c33a4baf0caaa5cfd4906cfc94.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 294, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 545, + 338 + ], + "type": "text", + "content": "Figure 8. LM guidance substantially improves CIDEr and retrieval scores of a model trained solely on minimally curated web data and evaluated on MS-COCO. The x-axis shows the number of captions used to prompt the LM; we do not prompt with images." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 425, + 191, + 539, + 289 + ], + "blocks": [ + { + "bbox": [ + 425, + 191, + 539, + 289 + ], + "lines": [ + { + "bbox": [ + 425, + 191, + 539, + 289 + ], + "spans": [ + { + "bbox": [ + 425, + 191, + 539, + 289 + ], + "type": "image", + "image_path": "cf52968589ca7511b10d22284360cd466927fd7cfc51cc8e7966d06cf21efd8d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 483 + ], + "type": "text", + "content": "low-quality data. Here, we use a CoCa model that is pretrained on image-alt text pairs from the web (the ALIGN dataset [17]) and classification labels converted to text (the JFT-5B dataset [49]), without any additional fine-tuning. Because the data distribution places higher probability mass on short, non-descriptive captions than on longer captions, the resulting model is of limited utility for captioning, and would generally need to be fine-tuned on another dataset such as MS-COCO before being applied to a captioning task. Rather than fine-tune, we use LM guidance to prompt the model with captions from the MS-COCO training set." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 484, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 546, + 628 + ], + "type": "text", + "content": "LM guidance substantially improves the quality of the captions produced by the original pretrained CoCa model without any clean parallel data. With LM guidance, we achieve CIDEr scores of 48.6 with 5 shots and 59.7 with 50 shots, far exceeding the CIDEr score of 21.8 obtained with no guidance. Figure 8 shows CIDEr and CLIP recall@1 scores for LM guidance of this pretrained CoCa model as a function of the number of shots, with " + }, + { + "bbox": [ + 304, + 484, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = 1" + }, + { + "bbox": [ + 304, + 484, + 546, + 628 + ], + "type": "text", + "content": ". Table 3 compares classifier-free guidance and LM guidance. CFG yields higher CLIP-Scores and retrieval accuracy than LM guidance with " + }, + { + "bbox": [ + 304, + 484, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = 1" + }, + { + "bbox": [ + 304, + 484, + 546, + 628 + ], + "type": "text", + "content": ", but LM guidance provides much higher CIDEr scores." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "We compare captions generated with CFG to those generated with LM guidance for four images in Figure 9. In general, CFG produces agrammatical captions, whereas LM guidance produces grammatical captions but hallucinates details. For example, the image in the upper left shows two elephants and no zebras, but LM guidance leads to the caption \"an elephant and a zebra in a field.\"" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15265" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 69, + 545, + 286 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 286 + ], + "type": "table", + "html": "
ModelReference-Based MetricsReference-Free Metrics
BLEU-4METEORROUGECIDErRefOnlyCLIP-SCLIP-SR@1R@5R@10RefCLIP-S
Classifier-free guidance:
γ = 1.08.28.321.821.80.7660.6949.019.526.10.725
γ = 1.28.69.524.525.00.7810.71812.727.235.10.745
γ = 1.58.910.025.625.20.7800.72816.733.843.00.750
γ = 2.08.19.723.922.90.7770.74121.240.851.10.755
γ = 3.07.18.720.018.50.7670.75325.847.858.30.756
γ = 4.06.47.516.313.90.7490.74327.348.558.10.742
Language model guidance with α = β = 1:
2 captions12.714.634.739.30.8060.68810.023.732.40.740
5 captions15.016.639.148.60.8270.71212.427.537.50.763
10 captions16.217.740.553.10.8350.72313.030.541.00.773
20 captions17.418.441.657.40.8390.72814.432.242.70.777
50 captions18.119.142.559.70.8400.72913.432.543.90.778
Other models trained without aligned MS-COCO images and captions:
ZeroCap [41]2.611.514.60.870.79
MAGIC [39]12.917.439.949.3
Flamingo [3]84.3
DeCap (560 captions) [22]51.4
DeCap (full train set) [22]24.725.091.2
CapDec (full train set) [27]26.425.151.891.8
", + "image_path": "c05c2217211674894cd2b0c0f83992bacc004820b3e9d17b42f8a27da6594aed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 546, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 546, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 546, + 334 + ], + "type": "text", + "content": "Table 3. Comparison of decoding strategies for a captioning model trained only on minimally curated web data (JFT-5B and ALIGN) and evaluated on MS-COCO. At the bottom, we report metrics for other models trained without aligned MS-COCO images and captions. These models may not be directly comparable since they use different pretraining data. DeCap and CapDec use all 560K captions in the MS-COCO training set to train their decoders; we include CIDEr for DeCap with 560 captions (0.1% of the training data) for comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 345, + 108, + 403 + ], + "blocks": [ + { + "bbox": [ + 49, + 345, + 108, + 403 + ], + "lines": [ + { + "bbox": [ + 49, + 345, + 108, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 345, + 108, + 403 + ], + "type": "image", + "image_path": "f48e3bb4252d02890f0913187077bc55a3e0f64d7f5d1598c9f29bba6a389f9c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 404, + 96, + 416 + ], + "lines": [ + { + "bbox": [ + 52, + 404, + 96, + 416 + ], + "spans": [ + { + "bbox": [ + 52, + 404, + 96, + 416 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 52, + 404, + 96, + 416 + ], + "type": "text", + "content": " : a photo of the small elephant" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 416, + 106, + 421 + ], + "lines": [ + { + "bbox": [ + 52, + 416, + 106, + 421 + ], + "spans": [ + { + "bbox": [ + 52, + 416, + 106, + 421 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 52, + 416, + 106, + 421 + ], + "type": "text", + "content": " : elephants in the" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 421, + 107, + 426 + ], + "lines": [ + { + "bbox": [ + 52, + 421, + 107, + 426 + ], + "spans": [ + { + "bbox": [ + 52, + 421, + 107, + 426 + ], + "type": "text", + "content": "ruaha national park" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 427, + 101, + 433 + ], + "lines": [ + { + "bbox": [ + 52, + 427, + 101, + 433 + ], + "spans": [ + { + "bbox": [ + 52, + 427, + 101, + 433 + ], + "type": "inline_equation", + "content": "\\gamma = 3.0" + }, + { + "bbox": [ + 52, + 427, + 101, + 433 + ], + "type": "text", + "content": " : elephants chobe" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 433, + 74, + 437 + ], + "lines": [ + { + "bbox": [ + 58, + 433, + 74, + 437 + ], + "spans": [ + { + "bbox": [ + 58, + 433, + 74, + 437 + ], + "type": "text", + "content": "np" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 438, + 102, + 443 + ], + "lines": [ + { + "bbox": [ + 58, + 438, + 102, + 443 + ], + "spans": [ + { + "bbox": [ + 58, + 438, + 102, + 443 + ], + "type": "text", + "content": "LM: a elephant and a" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 443, + 101, + 448 + ], + "lines": [ + { + "bbox": [ + 58, + 443, + 101, + 448 + ], + "spans": [ + { + "bbox": [ + 58, + 443, + 101, + 448 + ], + "type": "text", + "content": "zebra in a field" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 449, + 101, + 453 + ], + "lines": [ + { + "bbox": [ + 58, + 449, + 101, + 453 + ], + "spans": [ + { + "bbox": [ + 58, + 449, + 101, + 453 + ], + "type": "text", + "content": "OT: Two elephants." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 453, + 105, + 472 + ], + "lines": [ + { + "bbox": [ + 58, + 453, + 105, + 472 + ], + "spans": [ + { + "bbox": [ + 58, + 453, + 105, + 472 + ], + "type": "text", + "content": "G1: Two elephants standing on a grassy field next to a tree." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "type": "text", + "content": "Figure 9. Examples of captions generated from a model pretrained only on minimally curated data, for randomly selected images without human faces. Captions labeled " + }, + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "type": "text", + "content": " are obtained without CFG; " + }, + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "type": "text", + "content": " uses CFG; LM indicates LM guidance with " + }, + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = 1" + }, + { + "bbox": [ + 46, + 482, + 287, + 537 + ], + "type": "text", + "content": " and 20 shots; GT indicates ground truth." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 110, + 345, + 167, + 403 + ], + "blocks": [ + { + "bbox": [ + 110, + 345, + 167, + 403 + ], + "lines": [ + { + "bbox": [ + 110, + 345, + 167, + 403 + ], + "spans": [ + { + "bbox": [ + 110, + 345, + 167, + 403 + ], + "type": "image", + "image_path": "19d46518ea0a222e081981c6bbd809b9d405903d5b122d2521587282c3a8265c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 404, + 154, + 409 + ], + "lines": [ + { + "bbox": [ + 110, + 404, + 154, + 409 + ], + "spans": [ + { + "bbox": [ + 110, + 404, + 154, + 409 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 110, + 404, + 154, + 409 + ], + "type": "text", + "content": " a photo of the" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 410, + 152, + 415 + ], + "lines": [ + { + "bbox": [ + 110, + 410, + 152, + 415 + ], + "spans": [ + { + "bbox": [ + 110, + 410, + 152, + 415 + ], + "type": "text", + "content": "small coffee." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 415, + 164, + 419 + ], + "lines": [ + { + "bbox": [ + 110, + 415, + 164, + 419 + ], + "spans": [ + { + "bbox": [ + 110, + 415, + 164, + 419 + ], + "type": "inline_equation", + "content": "\\alpha = {2.0}" + }, + { + "bbox": [ + 110, + 415, + 164, + 419 + ], + "type": "text", + "content": " ; a coffee in a video" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 420, + 124, + 425 + ], + "lines": [ + { + "bbox": [ + 110, + 420, + 124, + 425 + ], + "spans": [ + { + "bbox": [ + 110, + 420, + 124, + 425 + ], + "type": "text", + "content": "，" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 426, + 155, + 430 + ], + "lines": [ + { + "bbox": [ + 110, + 426, + 155, + 430 + ], + "spans": [ + { + "bbox": [ + 110, + 426, + 155, + 430 + ], + "type": "text", + "content": "game:" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 431, + 154, + 435 + ], + "lines": [ + { + "bbox": [ + 110, + 431, + 154, + 435 + ], + "spans": [ + { + "bbox": [ + 110, + 431, + 154, + 435 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 436, + 159, + 441 + ], + "lines": [ + { + "bbox": [ + 110, + 436, + 159, + 441 + ], + "spans": [ + { + "bbox": [ + 110, + 436, + 159, + 441 + ], + "type": "text", + "content": "video game." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 441, + 159, + 446 + ], + "lines": [ + { + "bbox": [ + 110, + 441, + 159, + 446 + ], + "spans": [ + { + "bbox": [ + 110, + 441, + 159, + 446 + ], + "type": "text", + "content": "LM: a banana with a" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 447, + 160, + 451 + ], + "lines": [ + { + "bbox": [ + 110, + 447, + 160, + 451 + ], + "spans": [ + { + "bbox": [ + 110, + 447, + 160, + 451 + ], + "type": "text", + "content": "cup of coffee" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 452, + 160, + 468 + ], + "lines": [ + { + "bbox": [ + 110, + 452, + 160, + 468 + ], + "spans": [ + { + "bbox": [ + 110, + 452, + 160, + 468 + ], + "type": "text", + "content": "G1: A close up of a banana next to a cup with liquid." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 168, + 345, + 227, + 403 + ], + "blocks": [ + { + "bbox": [ + 168, + 345, + 227, + 403 + ], + "lines": [ + { + "bbox": [ + 168, + 345, + 227, + 403 + ], + "spans": [ + { + "bbox": [ + 168, + 345, + 227, + 403 + ], + "type": "image", + "image_path": "c40f15b728b5e5805118f1fd22d4a71ff411ef01445d1c39d38b5c9471bdd159.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 403, + 226, + 409 + ], + "lines": [ + { + "bbox": [ + 169, + 403, + 226, + 409 + ], + "spans": [ + { + "bbox": [ + 169, + 403, + 226, + 409 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 169, + 403, + 226, + 409 + ], + "type": "text", + "content": " a photo of the large" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 410, + 211, + 415 + ], + "lines": [ + { + "bbox": [ + 169, + 410, + 211, + 415 + ], + "spans": [ + { + "bbox": [ + 169, + 410, + 211, + 415 + ], + "type": "text", + "content": "windsurfing." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 415, + 225, + 419 + ], + "lines": [ + { + "bbox": [ + 169, + 415, + 225, + 419 + ], + "spans": [ + { + "bbox": [ + 169, + 415, + 225, + 419 + ], + "type": "inline_equation", + "content": "\\gamma = 3.0" + }, + { + "bbox": [ + 169, + 415, + 225, + 419 + ], + "type": "text", + "content": " : windsurfing in tarifa" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 420, + 225, + 425 + ], + "lines": [ + { + "bbox": [ + 169, + 420, + 225, + 425 + ], + "spans": [ + { + "bbox": [ + 169, + 420, + 225, + 425 + ], + "type": "inline_equation", + "content": "x = 3.0" + }, + { + "bbox": [ + 169, + 420, + 225, + 425 + ], + "type": "text", + "content": " : windsurfing walls" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 426, + 225, + 430 + ], + "lines": [ + { + "bbox": [ + 169, + 426, + 225, + 430 + ], + "spans": [ + { + "bbox": [ + 169, + 426, + 225, + 430 + ], + "type": "text", + "content": "pors 1300x1034" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 431, + 225, + 435 + ], + "lines": [ + { + "bbox": [ + 169, + 431, + 225, + 435 + ], + "spans": [ + { + "bbox": [ + 169, + 431, + 225, + 435 + ], + "type": "text", + "content": "1M" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 436, + 198, + 441 + ], + "lines": [ + { + "bbox": [ + 169, + 436, + 198, + 441 + ], + "spans": [ + { + "bbox": [ + 169, + 436, + 198, + 441 + ], + "type": "text", + "content": "EM. a Win" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 441, + 216, + 445 + ], + "lines": [ + { + "bbox": [ + 176, + 441, + 216, + 445 + ], + "spans": [ + { + "bbox": [ + 176, + 441, + 216, + 445 + ], + "type": "text", + "content": "water" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 446, + 217, + 451 + ], + "lines": [ + { + "bbox": [ + 176, + 446, + 217, + 451 + ], + "spans": [ + { + "bbox": [ + 176, + 446, + 217, + 451 + ], + "type": "text", + "content": "GT: A man riding a" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 452, + 218, + 456 + ], + "lines": [ + { + "bbox": [ + 176, + 452, + 218, + 456 + ], + "spans": [ + { + "bbox": [ + 176, + 452, + 218, + 456 + ], + "type": "text", + "content": "wind sail in the" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 456, + 218, + 460 + ], + "lines": [ + { + "bbox": [ + 176, + 456, + 218, + 460 + ], + "spans": [ + { + "bbox": [ + 176, + 456, + 218, + 460 + ], + "type": "text", + "content": "ocean filled with" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 461, + 198, + 466 + ], + "lines": [ + { + "bbox": [ + 176, + 461, + 198, + 466 + ], + "spans": [ + { + "bbox": [ + 176, + 461, + 198, + 466 + ], + "type": "text", + "content": "waves" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 228, + 345, + 286, + 403 + ], + "blocks": [ + { + "bbox": [ + 228, + 345, + 286, + 403 + ], + "lines": [ + { + "bbox": [ + 228, + 345, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 228, + 345, + 286, + 403 + ], + "type": "image", + "image_path": "4eff774343483a48497c359ed640f66e928da88211ab34e660cc43d94972ed83.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 404, + 284, + 409 + ], + "lines": [ + { + "bbox": [ + 228, + 404, + 284, + 409 + ], + "spans": [ + { + "bbox": [ + 228, + 404, + 284, + 409 + ], + "type": "inline_equation", + "content": "\\gamma = 1.0" + }, + { + "bbox": [ + 228, + 404, + 284, + 409 + ], + "type": "text", + "content": " a photo of the large" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 410, + 259, + 415 + ], + "lines": [ + { + "bbox": [ + 228, + 410, + 259, + 415 + ], + "spans": [ + { + "bbox": [ + 228, + 410, + 259, + 415 + ], + "type": "text", + "content": "giraffe." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 415, + 282, + 419 + ], + "lines": [ + { + "bbox": [ + 228, + 415, + 282, + 419 + ], + "spans": [ + { + "bbox": [ + 228, + 415, + 282, + 419 + ], + "type": "inline_equation", + "content": "\\alpha = 2.0" + }, + { + "bbox": [ + 228, + 415, + 282, + 419 + ], + "type": "text", + "content": " a piraffe in a video" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 420, + 282, + 425 + ], + "lines": [ + { + "bbox": [ + 228, + 420, + 282, + 425 + ], + "spans": [ + { + "bbox": [ + 228, + 420, + 282, + 425 + ], + "type": "text", + "content": "game" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 426, + 276, + 430 + ], + "lines": [ + { + "bbox": [ + 228, + 426, + 276, + 430 + ], + "spans": [ + { + "bbox": [ + 228, + 426, + 276, + 430 + ], + "type": "text", + "content": "gamer:" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 431, + 259, + 435 + ], + "lines": [ + { + "bbox": [ + 228, + 431, + 259, + 435 + ], + "spans": [ + { + "bbox": [ + 228, + 431, + 259, + 435 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 436, + 280, + 441 + ], + "lines": [ + { + "bbox": [ + 228, + 436, + 280, + 441 + ], + "spans": [ + { + "bbox": [ + 228, + 436, + 280, + 441 + ], + "type": "text", + "content": "photo 1" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 441, + 280, + 446 + ], + "lines": [ + { + "bbox": [ + 228, + 441, + 280, + 446 + ], + "spans": [ + { + "bbox": [ + 228, + 441, + 280, + 446 + ], + "type": "text", + "content": "LM: a giraffe standing in a tall tower." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 447, + 280, + 451 + ], + "lines": [ + { + "bbox": [ + 228, + 447, + 280, + 451 + ], + "spans": [ + { + "bbox": [ + 228, + 447, + 280, + 451 + ], + "type": "text", + "content": "in a tall tree" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 452, + 282, + 456 + ], + "lines": [ + { + "bbox": [ + 228, + 452, + 282, + 456 + ], + "spans": [ + { + "bbox": [ + 228, + 452, + 282, + 456 + ], + "type": "text", + "content": "GI: A giraffe in a dry" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 456, + 282, + 460 + ], + "lines": [ + { + "bbox": [ + 228, + 456, + 282, + 460 + ], + "spans": [ + { + "bbox": [ + 228, + 456, + 282, + 460 + ], + "type": "text", + "content": "savannah with dry" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 461, + 276, + 467 + ], + "lines": [ + { + "bbox": [ + 228, + 461, + 276, + 467 + ], + "spans": [ + { + "bbox": [ + 228, + 461, + 276, + 467 + ], + "type": "text", + "content": "shrubs" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 47, + 549, + 119, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 119, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 119, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "Our study indicates that it is possible to substantially improve the extent to which generated captions uniquely describe the images they correspond to, raising questions regarding the goal of image captioning and how it should be evaluated. As it is conventionally formulated, image captioning aims not to provide text that can substitute for an image, but to write the text that a human annotator would have written. This formulation penalizes captions that are more descriptive than ground truth, even when a human might prefer them. On the other hand, treating image captioning as a problem of generating a caption that lies close to the image in the embedding space of an image-text model is" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 346, + 546, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 546, + 418 + ], + "type": "text", + "content": "also inadequate, because captions that lie close to the image need not be grammatical and may contain gibberish. Our proposed methods leveraging classifier-free guidance and language model guidance modulate the trade-offs between these two goals, as captured by various reference-based and reference-free metrics." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "There are several possible extensions to our work. First, our present experiments use only greedy decoding. Although greedy decoding appears to perform reasonably well in our setup, it may be suboptimal for LM guidance with prompts that impose greater structure upon the captions. If the LM is prompted to output either \"there is a person in this image\" or \"there is no person this image\", greedy decoding is likely to fail even if the captioner properly scores the two possible captions, because when choosing between the tokens \"a\" and \"no\", the captioner has no knowledge of the structure that the LM will impose on future tokens. Since beam search could explore both tokens, it may offer better results in this scenario. Second, our method could be combined with RL-based methods to increase similarity in a contrastive embedding space, which may further improve retrieval performance and CLIPScore. Finally, with a perfect captioning model, " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathrm{image}|\\mathrm{caption})" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " should increase with " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ". However, in practice we find that " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma > 3" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " leads to a decrease in retrieval performance. This discrepancy suggests that the difference between the conditional and unconditional model distributions may be a noisy estimator of the pointwise mutual information. Although selecting " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " is one way to regularize this estimator, there may also be strategies to regularize " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathrm{x}|\\mathrm{y}) / p(\\mathrm{x})" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " at training time." + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15266" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "type": "text", + "content": "We thank Kevin Clark, David Fleet, Geoffrey Hinton, and the rest of the Google DeepMind Toronto team for inspiration, comments, and discussion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 157, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 157, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 287, + 212 + ], + "type": "text", + "content": "[1] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 268 + ], + "type": "text", + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In Proceedings of the IEEE International Conference on Computer Vision, pages 8948-8957, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 323 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 390 + ], + "type": "text", + "content": "[4] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[5] David M Chan, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. " + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "inline_equation", + "content": "ic^3" + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": ": Image captioning by committee consensus. arXiv preprint arXiv:2302.01328, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 435, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 502 + ], + "type": "text", + "content": "[6] Jaemin Cho, Seunghyun Yoon, Ajinkya Kale, Franck Dernoncourt, Trung Bui, and Mohit Bansal. Fine-grained image captioning with CLIP reward. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 517-527, Seattle, United States, July 2022. Association for Computational Linguistics. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "type": "text", + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. PaLM: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 288, + 612 + ], + "type": "text", + "content": "[8] Katherine Crowson. You can apply a similar trick to classifier-free guidance to autoregressive transformers to sample from a synthetic \"super-conditioned\" distribution. https://twitter.com/RiversHaveWings/status/1478093658716966912, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 287, + 656 + ], + "type": "text", + "content": "[9] Béatrice Daille. Approche mixte pour l'extraction automatique de terminologie: statistiques lexicales et filtrres linguistiques. PhD thesis, Ph.D. thesis, Université Paris 7, 1994. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[11] Jacob Devlin, Hao Cheng, Hao Fang, Saurabh Gupta, Li Deng, Xiaodong He, Geoffrey Zweig, and Margaret Mitchell. Language models for image captioning: The quirks and what works. arXiv preprint arXiv:1505.01809, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 195, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 271 + ], + "type": "text", + "content": "[13] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[15] Micah Hodosh, Peter Young, and Julia Hockenmaier. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853-899, 2013. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[16] Yupan Huang, Hongwei Xue, Bei Liu, and Yutong Lu. Unifying multimodal transformer for bi-directional image and text generation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1138-1147, 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[18] Andrej Karpathy and Fei-Fei Li. Deep visual-semantic alignments for generating image descriptions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 3128-3137. IEEE Computer Society, 2015. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "type": "text", + "content": "[19] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander R Fabbri, Yejin Choi, and Noah A Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. arXiv preprint arXiv:2112.04139, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[20] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Morrison, Ronan Le Bras, Yejin Choi, and Noah A Smith. Transparent human evaluation for image captioning. arXiv preprint arXiv:2111.08940, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "text", + "content": "[21] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Li Fei-Fei. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 714 + ], + "type": "text", + "content": "[22] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding clip latents for zero-shot captioning. In International Conference on Learning Representations, 2022. 2, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15267" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[23] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 173 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 288, + 218 + ], + "type": "text", + "content": "[25] Annika Lindh, Robert J Ross, Abhijit Mahalunkar, Giancarlo Salton, and John D Kelleher. Generating diverse and meaningful captions. In International Conference on Artificial Neural Networks, pages 176-187. Springer, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 284 + ], + "type": "text", + "content": "[26] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 286, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 287, + 319 + ], + "type": "text", + "content": "[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 320, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 373 + ], + "type": "text", + "content": "[28] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 376, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 441 + ], + "type": "text", + "content": "[29] Jordi Pont-Tuset, Jasper Uijlings, Soravit Changpinyo, Radu Soricut, and Vittorio Ferrari. Connecting vision and language with localized narratives. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 647-664. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 444, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 444, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 444, + 287, + 509 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "text", + "content": "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 544, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 287, + 586 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 644 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[34] Noam Rotstein, David Bensaid, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models to fuse visual data into enriched image captions. arXiv preprint arXiv:2305.17718, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "text", + "content": "Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 119, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 161 + ], + "type": "text", + "content": "[36] Guillaume Sanchez, Honglu Fan, Alexander Spangher, Elad Levi, Pawan Sasanka Ammanamanchi, and Stella Biderman. Stay on topic with classifier-free guidance. arXiv preprint arXiv:2306.17806, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "text", + "content": "[37] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "text", + "content": "[38] David R So, Wojciech Manke, Hanxiao Liu, Zihang Dai, Noam Shazeer, and Quoc V Le. Primer: Searching for efficient transformers for language modeling. arXiv preprint arXiv:2109.08668, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "text", + "content": "[39] Yixuan Su, Tian Lan, Yahui Liu, Fangyu Liu, Dani Yogatama, Yan Wang, Lingpeng Kong, and Nigel Collier. Language models can see: plugging visual controls in text generation. arXiv preprint arXiv:2205.02655, 2022. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "type": "text", + "content": "[40] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "text", + "content": "[41] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zerocap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918-17928, 2022. 2, 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "text", + "content": "[42] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "text", + "content": "[43] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 564 + ], + "type": "text", + "content": "[44] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. arXiv preprint arXiv:2302.03668, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 599 + ], + "type": "text", + "content": "[45] Shitong Xu. Clip-diffusion-lm: Apply diffusion model on image captioning. arXiv preprint arXiv:2210.04559, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "text", + "content": "[46] Linjie Yang, Kevin Tang, Jianchao Yang, and Li-Jia Li. Dense captioning with joint inference and visual context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2193-2202, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[47] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3, 4, 6, 14" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[48] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15268" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 297 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "fei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 150 + ], + "type": "text", + "content": "[49] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12104–12113, 2022. 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "text", + "content": "[50] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 251 + ], + "type": "text", + "content": "[51] Youuyuan Zhang, Jiuniu Wang, Hao Wu, and Wenjia Xu. Distinctive image captioning via clip guided group optimization. arXiv preprint arXiv:2208.04254, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "type": "text", + "content": "[52] Zixin Zhu, Yixuan Wei, Jianfeng Wang, Zhe Gan, Zheng Zhang, Le Wang, Gang Hua, Lijuan Wang, Zicheng Liu, and Han Hu. Exploring discrete diffusion models for image captioning. arXiv preprint arXiv:2211.11694, 2022. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15269" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_content_list.json b/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e3dcf7573c21d98249a8a27d1e45a8fac5a0a5ef --- /dev/null +++ b/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_content_list.json @@ -0,0 +1,1861 @@ +[ + { + "type": "text", + "text": "Guiding Local Feature Matching with Surface Curvature", + "text_level": 1, + "bbox": [ + 194, + 130, + 774, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuzhe Wang $^{1*}$", + "bbox": [ + 153, + 179, + 281, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juho kannala", + "bbox": [ + 320, + 180, + 433, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marc Pollefeys $^{2,3}$", + "bbox": [ + 480, + 179, + 622, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Barath2", + "bbox": [ + 669, + 180, + 789, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Aalto University", + "bbox": [ + 233, + 222, + 377, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ ETH zurich", + "bbox": [ + 436, + 222, + 540, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Microsoft", + "bbox": [ + 599, + 223, + 691, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 292, + 313, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a new method, called curvature similarity extractor (CSE), for improving local feature matching across images. CSE calculates the curvature of the local 3D surface patch for each detected feature point in a viewpoint-invariant manner via fitting quadrics to predicted monocular depth maps. This curvature is then leveraged as an additional signal in feature matching with off-the-shelf matchers like SuperGlue and LoFTR. Additionally, CSE enables end-to-end joint training by connecting the matcher and depth predictor networks. Our experiments demonstrate on large-scale real-world datasets that CSE consistently improves the accuracy of state-of-the-art methods. Fine-tuning the depth prediction network further enhances the accuracy. The proposed approach achieves state-of-the-art results on the ScanNet dataset, showcasing the effectiveness of incorporating 3D geometric information into feature matching.", + "bbox": [ + 75, + 325, + 473, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 613, + 209, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Local feature matching is a crucial component for many geometric computer vision tasks, including visual localization [50, 51, 52, 48, 14, 66], structure-from-motion (SfM) [67, 53, 29], and simultaneous localization and mapping (SLAM) [38, 39, 10]. Given a pair of images observing a 3D scene, the task is to find reliable tentative point-to-point correspondences in the two images. Forming such feature matches is often a challenging task as the images may undergo large viewpoint and illumination changes, have occlusions or repetitive patterns.", + "bbox": [ + 75, + 638, + 468, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The standard pipeline for local feature matching typically involves two steps: (1) keypoint detection and description and (2) point-wise feature matching. Traditional approaches mainly focus on improving the robustness of", + "bbox": [ + 75, + 791, + 470, + 852 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c181ac8fd349f5b1758fbde035aacee3f778dd4a3e8c2db5c48d649a3a09ef39.jpg", + "image_caption": [ + "Figure 1. Curvature-guided match selection: (a) Correct matches in two views must share the same underlying surface curvature, e.g., predicted from monocular depth. (b) Correspondences from different surfaces are guaranteed to be incorrect." + ], + "image_footnote": [], + "bbox": [ + 501, + 292, + 694, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/374c70d2a8eba1f187d384f647deb1c62e8cd64440dfb6c6f094578c8d30b59d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 295, + 890, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "keypoint detection and description through, for example, extending the Harris [21] detector to handle affine transformations and multiple scales [35, 36] or using more discriminative or efficient descriptors [32, 7, 34, 1, 3]. However, despite their unbroken popularity, these algorithms often fail to cope with the challenges that arise in real-world environments, leading to low accuracy.", + "bbox": [ + 496, + 522, + 892, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in deep learning-based feature matching have made significant progress in addressing the limitations of hand-crafted approaches, such as by jointly training detectors and descriptors [69, 12, 15, 44, 63, 40] with convolutional neural networks (CNNs), or combining hand-crafted and learning-based descriptors in a hybrid manner [41, 6]. SuperGlue [49] introduced the use of transformer networks to learn the matching process and formulate the problem as an optimal transport task [64]. LoFTR [55] and its recent variants [58, 8, 65] leverage both the global and local context from raw images by jointly learning the feature extraction and matching in a single network. While these approaches have led to state-of-the-art performance on several benchmarks, they work entirely in the 2D image domain and ignore the underlying 3D geometry of the scene. As feature matching is, essentially, finding the corresponding projections of an actual 3D point, this could be a critical limitation in real scenarios where the", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Part of the work was done during the author's visit to ETH zurich.", + "bbox": [ + 94, + 862, + 450, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code and trained models are available at https://github.com/AaltoVision/surface-curvature-estimator.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17981", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "matched pixels from different views must share the geometry of the underlying surface patch, as shown in Fig. 1.", + "bbox": [ + 76, + 90, + 467, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on investigating and exploiting 3D geometry cues, e.g. coming from monocular depth predictions, for feature matching. The only prior work is [61], which adopts depth priors to find planar image regions that are then rectified to eliminate viewpoint changes. Finally, a handcrafted detector and descriptor are applied to the rectified image regions. This, however, relies on the heavy assumption that the images consist mostly of dominant planes, which severely restricts its out-of-the-box applicability. In this work, we go further by exploring the local surface geometry coming from depth priors. This additional geometric clue is coupled with any off-the-shelf feature matcher increasing its accuracy. Benefiting from advanced monocular depth predictors [42, 43, 16], obtaining depth priors is easy and costs only a few milliseconds. We use these depth predictions to extract curvature at the observed feature points. Exploiting the fact that the curvature is invariant to viewpoint changes (i.e., rotation, translation, and scaling), we enforce that matched features must lie on similar surfaces. Our contributions are as follows:", + "bbox": [ + 76, + 122, + 467, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose an approach for feature matching, leveraging dense depth via utilizing local surface curvature similarity, which is invariant to viewpoint changes. This approach is a departure from traditional matching methods that rely solely on image information.", + "2. The proposed curvature similarity extractor can be seamlessly integrated with any recent feature matcher, making it a versatile tool that can be easily adopted. The experiments show that it improves several recent algorithms on various benchmarks by $1 - 3\\%$ .", + "3. The proposed algorithm can be used to train feature matchers and monocular depth prediction networks jointly in an end-to-end fashion. To demonstrate this, we fine-tune the state-of-the-art MiDaS [42] depth predictor to increase the feature matching accuracy." + ], + "bbox": [ + 88, + 435, + 467, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 700, + 212, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Local Feature Matching", + "text_level": 1, + "bbox": [ + 76, + 724, + 295, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detector-based Matchers build sparse correspondences on top of the detected keypoints. Therefore, robust keypoint detection and feature extraction are essential for these approaches. Traditional hand-crafted descriptors, such as SIFT [32] and SURF [7], follow a detect-then-describe pipeline and have shown great success since the 2000s. More recently, learned detectors and descriptors [6, 69, 12, 15, 44, 63, 60, 30] with convolutional neural networks show their superiority in matching images with large viewpoint and appearance changes. This traditional pipeline is further", + "bbox": [ + 76, + 750, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "developed to detect-and-describe [12, 15, 44] and describe-to-detect [60, 30] strategies leveraging high-level image information for accurate and reliable matching in challenging conditions. Besides substituting the traditional detectors and descriptors with their learned counterparts for image matching, the recent SuperGlue [49] focuses on the matching stage itself. It replaces the naive mutual nearest neighbour search with Graph Neural Networks (GNN). SuperGlue takes the sparse descriptors and their positional encoding as inputs, then leverage a transformer-based [64] network to create a more robust feature representation for the optimal partial assignment. The follow-up work [54] further improves the performance and efficiency by adaptively clustering sparse features into different subgraphs and using a coarse-to-fine paradigm. Although showing significant improvements with the recent components, the detector-based matchers are naturally limited by the accuracy of detected keypoints coming from independent detectors.", + "bbox": [ + 496, + 90, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detector-free Matchers directly estimate the dense correspondences from raw images without an independent keypoint detection stage. These approaches leverage low and high-level image information with neural networks and can distinguish indistinctive regions across images. The pioneer work of [46] constructs 4D cost volumes, significantly increasing the matching accuracy. However, it is computationally expensive as the complexity is $\\mathcal{O}(n^2)$ , where $n$ is the number of pixels or patches. [45, 26] alleviate the problem by applying either sparse convolutions or a coarse-to-fine paradigm. More recently, transformer-based detector-free matchers [55, 58, 65, 8, 24, 57] have received attention due to their strong performance on local feature matching. The most representative work, LoFTR [55], inherits the advantages of graph matching from SuperGlue and leverages a linear transformer [64] for efficient dense matching. [58, 8] utilise more efficient transformer structures to improve the performance. Detector-free matchers show a promising direction towards local feature matching by encoding rich image information. We claim that extracting 3D geometric information could further enhance accuracy.", + "bbox": [ + 496, + 367, + 890, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Correspondence Pruning approaches [17, 4, 5, 72, 70, 56] apply a consensus mechanism to filter out the outliers from putative correspondences coming from the feature matcher procedure. RANSAC [17] and its follow-up works [4, 5] are the most popular correspondence pruning algorithms. In the era of deep learning, OANet [72] infers the probabilities of correspondences being inliers with an order-aware network. Other works improve the accuracy of correspondence pruning by applying motion coherence constraints [31, 33], leveraging local-to-global consensus learning procedure [74, 11], and adopting the attention mechanism [56]. These approaches provide inlier probability predictions for each putative correspondence that can be used either to filter or order matches. Finally, least squares fitting", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17982", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or traditional robust estimation is applied to the matches.", + "bbox": [ + 76, + 90, + 452, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D Geometric Priors to Image Matching", + "text_level": 1, + "bbox": [ + 76, + 114, + 423, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Incorporating 3D geometric priors into various vision tasks has been widely explored in recent years. [23] leverages the RGB-D reconstructions to learn view-invariant, geometry-aware 2D representations for downstream tasks. [71] introduces the predicted monocular depth to guide the optimization of neural scene representation. The surface curvature, a popular 3D geometric cue, is widely applied to 3D vision tasks such as 3D point cloud registration [62], multi-view stereo [68, 19]. However, integrating 3D geometric cues into 2D image matching has received little attention. [37, 18] assume that the underlying surface patch is locally planar for feature detection and description, while [61] uses prior depth information to improve feature extraction by rectifying large planar regions. None of the above approaches considers the 3D geometric information in the local region, and we are the first to include 3D surface geometry in general 2D feature matching.", + "bbox": [ + 75, + 137, + 472, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Feature Matching with Surface Curvature", + "text_level": 1, + "bbox": [ + 76, + 406, + 455, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section proposes an approach to improve any local feature matcher by extracting information about the underlying local 3D surface predicted by a deep network. The overview of our method coupled with LoFTR [55] is shown in Fig. 2. Although the proposed algorithm is compatible with any matcher, we demonstrate its effectiveness with LoFTR. Combining it with other ones is straightforward.", + "bbox": [ + 75, + 431, + 468, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. LoFTR-style Matchers", + "text_level": 1, + "bbox": [ + 76, + 546, + 290, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides a brief overview of LoFTR. Given an image pair $I_A, I_B \\in \\mathbb{R}^{H*W*3}$ , the end-to-end trainable matcher, first, extracts coarse-level features $\\mathcal{F}_A^c, \\mathcal{F}_B^c$ and, then, fine-level ones $\\mathcal{F}_A^f, \\mathcal{F}_B^f$ with a local feature CNN [28]. At the coarse level transformer module, the features are flattened and processed by a transformer-based architecture with multiple self- and cross-attention layers [64]. A matching confidence map $\\mathcal{P}_c$ is predicted at the end of this module, and the matches $\\mathcal{M}_c$ with high confidence are selected from $\\mathcal{P}_c$ to fine-tune the prediction of fine-level features matching at the coarse-to-fine transformer module. The final outputs are the fine-level matches $\\mathcal{M}_f$ .", + "bbox": [ + 75, + 569, + 468, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Curvature Similarity Extractor", + "text_level": 1, + "bbox": [ + 76, + 760, + 354, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will discuss multiple potential ways of extracting local surface curvature at detected features. This curvature is then used for improving feature matching.", + "bbox": [ + 75, + 782, + 470, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Curvature Similarity", + "text_level": 1, + "bbox": [ + 76, + 845, + 277, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For an at least two times continuously differentiable surface $S \\in \\mathbb{C}^2$ in the 3D space, given a point $p \\in \\mathbb{R}^3$ and", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a direction $d \\in \\mathbb{R}^3$ on $S$ ; the normal curvature $k_n \\in \\mathbb{R}$ measures how curved the surface is at this point along the direction [13]. Since there are infinitely many directions that travel through a point on the surface, there are also infinitely many normal curvatures at a given point. In differential geometry, the minimum normal curvature $k_1$ and the maximum curvature $k_2$ at point $p$ are defined as the principal curvatures, where $k_1 \\leq k_2$ , and the directions of the principal curvatures are the smoothest and steepest directions on the surface. We also define the mean curvature $H$ and Gaussian curvature $K$ as", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH = \\frac {1}{2} \\left(k _ {1} + k _ {2}\\right), K = k _ {1} k _ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 265, + 890, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A highly beneficial property of such curvatures is their invariance to surface rotation and translation. Both are important when describing local surface patches centered on features detected in 2D images. However, these measures do not preserve scale invariance [13], which is essential in our case due to using predicted relative depth maps instead of a metric one. This means that, in each image, the depth is only defined up-to-scale.", + "bbox": [ + 496, + 304, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To seek the scale invariance of the curvatures for local patch matching, we follow [47] for the idea of scale-invariant curvature measure and define the curvature similarity function $S: \\mathbb{R} \\times \\mathbb{R} \\to \\mathbb{R}$ as follows:", + "bbox": [ + 496, + 426, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS \\left(k _ {1}, k _ {2}\\right) = \\frac {\\operatorname* {m i n} \\left(\\left| k _ {1} \\right| , \\left| k _ {2} \\right|\\right)}{\\operatorname* {m a x} \\left(\\left| k _ {1} \\right| , \\left| k _ {2} \\right|\\right)}, (0 \\leq S (k) \\leq 1), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 496, + 890, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $k_{1}$ and $k_{2}$ are the minimum and maximum curvatures, respectively. It is easy to see that $S(k_{1},k_{2})$ is rotation, translation and scale invariant. Case $S(k_{1},k_{2}) = 1$ is interpreted as the curvature of a point on a spherical surface. In the case when $k_{1} = k_{2} = 0$ , we define $S(k_{1},k_{2})$ to be 0. Detecting points on planes is also straightforward from $k_{1}$ and $k_{2}$ as they both will become $\\infty$ . Proof of the invariances is provided in the supplementary material. Next, we will discuss ways of obtaining curvature similarities.", + "bbox": [ + 496, + 540, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2 Surface Curvatures Extraction", + "text_level": 1, + "bbox": [ + 500, + 695, + 769, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces our approximate form of principal curvature estimation. Different from prior work [68, 19] that extract the curvatures from image intensity, we estimate the principal curvatures based on the depth map. This 3D geometric cue could be obtained either from the active sensors (LiDAR, Kinect) or monocular depth predictors [42, 43]. Since our target curvature similarity $S(k_{1}, k_{2})$ is scaling invariant, the scaling ambiguity in monocular depth estimation is neglected and the predicted relative depth is sufficient for the curvature extraction.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we consider only a pair of RGB images as input and leverage any off-the-shelf depth predictor, e.g.", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17983", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d0dd7ea3279a63ac0d14048d072e0d10beec8d485d747ea8a11b7992c9e121a5.jpg", + "image_caption": [ + "Figure 2. Overview of the CSE with LoFTR. For a pair of images, it follows the LoFTR [55] matching pipeline to extract the coarse and fine level image features and generate a coarse matching confidence map $\\mathcal{P}_c$ . In the meantime, the images are fed to a monocular depth network to predict the depth map and extract the curvature similarity upon the depth map. We formulate a new curvature score matrix $\\mathcal{P}_{curv}$ based on the pixel-wise $L2$ distance of the curvature similarities between two images and mix it with $\\mathcal{P}_c$ . The mixed score matrix is then processed by the coarse-to-fine transformer module together with the fine-level features to predict the final matches $\\mathcal{M}_f$ ." + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 848, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MiDaS [42], to extract the depth map $Z_{i}:\\mathbb{R}\\times \\mathbb{R}\\to \\mathbb{R}$ for each input image, $i\\in \\{1,2\\}$ . Let $\\mathbf{r}_i(u,v,Z_i(u,v))$ be the 3D surface patch corresponding to the image region at coordinates $(u,v)$ with window size $\\delta \\in \\mathbb{R}^{+}$ in the $i$ th image, where $Z_{i}(u,v)$ is the depth of point $(u,v)$ .", + "bbox": [ + 75, + 412, + 468, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us formulate a local point set $\\mathcal{P}_i = \\{(u,v,Z_i(u,v))\\mid (u,v)\\in \\mathcal{M}_f^i\\} \\subseteq \\mathbb{R}^3$ , where $\\mathcal{M}_f^i$ are the points detected in the $i$ th image.", + "bbox": [ + 76, + 489, + 470, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Principal Curvature via Surface Fitting. The most straightforward approach is first applying the quadratic surface approximation method [20] to estimate the surface $\\mathbf{r}_i(u,v,Z_i(u,v))$ at all points in $\\mathcal{P}$ , then employing the first and second fundamental forms of a surface. The normal curvature is calculated as follows:", + "bbox": [ + 75, + 535, + 472, + 626 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk _ {n} = \\frac {I I}{I} = \\frac {L d u ^ {2} + 2 M d u d v + N d v ^ {2}}{E d u ^ {2} + 2 F d u d v + G d v ^ {2}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 640, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L = \\mathbf{r}_{uu}\\mathbf{n}$ , $M = \\mathbf{r}_{uv}\\mathbf{n}$ , $N = \\mathbf{r}_{vv}\\mathbf{n}$ , $E = \\mathbf{r}_u\\mathbf{r}_u$ , $F = \\mathbf{r}_u\\mathbf{r}_v$ , $G = \\mathbf{r}_v\\mathbf{r}_v$ ; $\\mathbf{r}_u, \\mathbf{r}_v, \\mathbf{r}_{uu}, \\mathbf{r}_{uv}, \\mathbf{r}_{vv}$ are the first and second-order derivatives of the surface patch $\\mathbf{r}_i(u,v,Z_i(u,v))$ , and $\\mathbf{n} \\in \\mathbb{R}^3$ is the normal.", + "bbox": [ + 75, + 680, + 468, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk _ {n} = \\frac {L + 2 M \\lambda + N \\lambda^ {2}}{E + 2 F \\lambda + G \\lambda^ {2}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 755, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once we have the normal curvature $k_{n}$ , we then calculate the partial derivative of $k_{n}$ w.r.t. $\\lambda$ to obtain the principal curvatures $k_{1}, k_{2}$ . The details are presented in the supplementary. While this approach works accurately and efficiently in practice, we observed that its gradients become unstable during back-propagation. Therefore, we explore another way as well for calculating the surface curvature.", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Curvature Extraction via Quadrics.** Another lightweight procedure for estimating the curvature is via fitting quadric surfaces, e.g. ellipsoid, locally to the neighborhood of the observed point. Thus, instead of directly estimating a quadratic surface from the given depth map, we constrain the surface to be a quadric and, more specifically, an ellipsoid. An ellipsoid has the beneficial property that all its curvatures are semi-positive. Its principal axes define three directions which we can use the define the similarity. We use the radii $\\alpha, \\beta, \\gamma \\in \\mathbb{R}^{+}$ along the three axes and define the curvature similarity score as follows:", + "bbox": [ + 496, + 412, + 893, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (\\alpha , \\beta , \\gamma) = \\frac {\\operatorname* {m i n} \\left(\\alpha^ {2} , \\beta^ {2} , \\gamma^ {2}\\right)}{\\operatorname* {m a x} \\left(\\alpha^ {2} , \\beta^ {2} , \\gamma^ {2}\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 587, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarity $S$ is rotation, translation, scale and thus, viewpoint invariant. The advantage from Eq. (5) is that the computation of first and second-order gradients are no longer required for the curvature similarity calculation.", + "bbox": [ + 496, + 631, + 893, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The general constraint that a quadric imposes on a point $x \\in \\mathcal{P}_i$ lying on the surface is", + "bbox": [ + 496, + 691, + 893, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} \\mathbf {Q} _ {i} \\mathbf {x} ^ {\\mathrm {T}} + \\mathbf {P} _ {i} \\mathbf {x} ^ {\\mathrm {T}} + \\mathbf {R} _ {i} = 0,\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 731, + 782, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_i$ , $\\mathbf{P}_i$ , and $\\mathbf{R}_i$ are the parameters in a matrix form. This can be written as", + "bbox": [ + 496, + 758, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} a x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \\tag {6} \\\\ 2 f y z + 2 g x + 2 h y + 2 i z + j = 0. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 796, + 890, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As described in [25], a sufficient condition that guarantees the quadric surface to be an ellipsoid is", + "bbox": [ + 496, + 843, + 890, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n4 M - N ^ {2} = 1,\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 883, + 750, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17984", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 92, + 124, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nN = a + b + c,\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 118, + 374, + 132 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM = a b + b c + a c - d ^ {2} - e ^ {2} - f ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 135, + 375, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Also, since we aim at measuring the curvature of the surface patch located on the observed point, we have to constrain the ellipsoid so that the observed point $\\mathbf{p}$ is on its surface. To do so, we translate the point set $\\mathcal{P}$ as $\\mathcal{P}' = \\{\\mathbf{q} - \\mathbf{p} \\mid \\mathbf{q} \\in \\mathcal{P}\\}$ and fix $j = 0$ . The central point now is $\\mathbf{p}' = [0,0,0]^{\\mathrm{T}}$ .", + "bbox": [ + 76, + 164, + 468, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Substituting all these constraints into Eq. (6), we obtain", + "bbox": [ + 94, + 239, + 465, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 263, + 468, + 289 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n2 f y z + 2 g x + 2 h y + 2 i z = 0.\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 286, + 388, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From Eq. (7), only nine variables need to be estimated in our case. For each point $[u, v, z]^{\\mathrm{T}} \\in \\mathcal{P}^{\\prime}$ , we define $\\mathbf{X}_i$ as", + "bbox": [ + 76, + 311, + 468, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {i} = \\left[ u _ {i} ^ {2}, v _ {i} ^ {2}, z _ {i} ^ {2}, 2 v _ {i} z _ {i}, 2 u _ {i} z _ {i}, 2 u _ {i} v _ {i}, 2 u _ {i}, 2 v _ {i}, 2 z _ {i} \\right] ^ {\\mathrm {T}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 351, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and coefficient matrix $\\mathbf{C} = [\\mathbf{X}_1, \\mathbf{X}_2, \\mathbf{X}_3, \\dots, \\mathbf{X}_n]$ . Then, we follow the method in [25] for the ellipsoid fitting and estimate its radii $R = (\\alpha, \\beta, \\gamma)$ . The algorithms for fitting and calculating the radii are shown in the supplementary.", + "bbox": [ + 76, + 378, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Curvature-Guided Match Selection", + "text_level": 1, + "bbox": [ + 76, + 449, + 383, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed curvature similarity extractor (CSE) is a depth-based plug-in component that can be combined with both detector-based [32, 49] and detector-free [55, 58, 8] matchers. For an image pair $(I_A, I_B)$ , the monocular depth predictor first estimates the depth map for each image independently. CSE extracts the curvature map $(S_A, S_B)$ on the top of the depth maps. Note that LoFTR selects the candidate matches $M_c$ at the coarse level of 1/8-resolution grids where the ground truth matches are defined based on the reprojection distance of the centers of the grid cells. Thus, it is not necessary to extract the curvature similarity at each pixel. It is calculated only at the center point of each cell, i.e., $S_A, S_B \\in \\mathbb{R}^{(h * w)*1}$ , where $h = H / 8$ , $w = W / 8$ .", + "bbox": [ + 75, + 472, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the proposed curvature similarity is rotation, translation and scale invariant, $S_{A}$ at point $\\mathbf{p}_A$ in the first image should be equal to the similarity $S_B$ of its corresponding pair $\\mathbf{p}_B$ in the second image. The correct matches $M_c$ , ideally, have the same curvature values. To measure this, we define the normalised curvature score matrix as", + "bbox": [ + 75, + 669, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {\\text {c u r v}} = 1 - L _ {2} (S (k) _ {A}, S (k) _ {B}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 771, + 468, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{2}(\\cdot ,\\cdot)$ is the L2-norm of $S(k)_A$ and $S(k)_B$ . Since both $S(k)_A$ and $S(k)_B$ are in-between zero and one, the curvature score is also normalized into range [0, 1].", + "bbox": [ + 76, + 797, + 468, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The new matching score matrix is the linear combination of $\\mathcal{P}_{\\mathrm{curv}}$ and LoFTR matching probability $\\mathcal{P}_c$ as follows:", + "bbox": [ + 76, + 844, + 468, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} = \\lambda \\mathcal {P} _ {c} + (1 - \\lambda) \\mathcal {P} _ {\\text {c u r v}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/47e53b10f054a133c4167ac138e30950c20b6933afa6ee0f90c7b7be82554861.jpg", + "image_caption": [ + "Figure 3. Multi-scale curvature extraction. We extract the curvature similarity from ellipsoids centered on the observed point, fitted with different window (i.e., 3D neighborhood) sizes." + ], + "image_footnote": [], + "bbox": [ + 513, + 89, + 880, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the mixing parameter for balancing the two matching scores. Since our $\\mathcal{P}_{\\mathrm{curv}}$ is an additional supporting score for coarse-level matching selection, we set the parameter $\\lambda$ to 0.9 in all our experiments. We adopt the scheme of LoFTR for the coarse-level match selection as", + "bbox": [ + 496, + 258, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {c} = \\{(i, j) \\mid \\forall (i, j) \\in \\operatorname {M N N} (\\mathcal {P}), \\mathcal {P} (i, j) \\geq \\theta \\}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 340, + 890, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where MNN is the mutual nearest neighbor operation, $\\theta$ is a threshold so that we only select the matches with confidence higher than $\\theta$ . Note that $\\mathcal{P}_{\\mathrm{curv}}$ can be similarly combined with the matching score matrix of SuperGlue [49].", + "bbox": [ + 496, + 364, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-scale Curvature Similarity. Although the proposed curvature similarity is scaling invariant in theory, it still depends on the neighboring pixels of the points. Thus, the multi-scale extraction of this geometric cue could also be adopted for robustness to scale change in the images. For detector-based approaches, such as SuperGlue [49], we use a set of window size $\\{\\delta_1,\\delta_2,\\dots,\\delta_n\\}$ with the window centered on detected keypoints. Hence, the curvature similarity maps are $S(k)_A\\in \\mathbb{R}^{\\mathcal{A}*n}$ and $S(k)_B\\in \\mathbb{R}^{\\mathcal{B}*n}$ , $\\mathcal{A}$ and $\\mathcal{B}$ are the detected keypoints at $I_A$ and $I_B$ , respectively. For detector-free matchers, the multi-scale extraction is conducted within the patch grid. Fig. 3 describes the curvature similarity extraction at multiple scales.", + "bbox": [ + 496, + 424, + 892, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Depth Prediction Fine-tuning", + "text_level": 1, + "bbox": [ + 500, + 628, + 763, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the proposed curvature similarity extraction is fully differentiable, it is possible to use it to train feature matching and depth prediction networks jointly end-to-end. To demonstrate this, we fine-tune the depth predictor with the curvature similarity extractor so that it takes the feature matching into account when being trained. The new confidence matrix $\\mathcal{P}$ is minimized with the negative log-likelihood, similarly as in [49, 55], where the loss is", + "bbox": [ + 496, + 651, + 890, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = - \\frac {1}{\\left| M _ {c} ^ {g t} \\right|} \\sum_ {(i, j) \\in M _ {c} ^ {g t}} \\log \\mathcal {P} (i, j), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 777, + 890, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $M_c^{gt}$ is the ground truth matches defined at the coarse-level in [55]. Since the loss term focuses only on the matched patches, long training time would crash the depth prediction model. Thus, we fine-tune the depth predictor with only a few epochs and small learning rates.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17985", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/afd6f1ec42bf39954ee3175271851e6a826101a24861aa23cd78a894b478218c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsYFCC100M [59]MegaDepth [27]ScanNet [9]
AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)
SuperGlue [49]38.959.475.798.723.643.762.277.099.632.516.133.851.884.431.5
SuperGlue + CSE40.060.176.395.626.546.264.878.798.335.716.234.052.280.731.7
LoFTR [55]43.463.178.095.28.253.770.182.496.89.022.140.857.687.47.9
LoFTR + CSE43.563.378.193.48.554.470.482.195.59.222.942.460.086.49.3
LoFTR + CSE (w/ FT)43.963.578.193.38.554.370.482.295.59.323.943.560.586.19.5
QuadTree [58]44.163.578.396.58.953.570.282.298.59.624.944.761.689.49.8
QuadTree + CSE45.064.378.895.39.154.070.382.197.99.825.245.562.589.411.2
QuadTree + CSE (w/ FT)45.064.478.995.59.154.571.182.998.09.725.846.163.489.211.5
", + "bbox": [ + 80, + 88, + 893, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results of relative pose estimation on YFCC100M, MegaDepth, and ScanNet datasets. We report AUC score of the translation and rotation errors with different thresholds $(5^{\\circ}, 10^{\\circ}, 20^{\\circ})$ , precision $(\\%)$ , and matching score $(\\%)$ . Our methods and the best performance are marked in bold. w/FT indicates the results with fine-tuning. Note that comparison among baselines is not straightforward as some sensitive default parameters, e.g. input image size, are different.", + "bbox": [ + 75, + 258, + 893, + 316 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/551650444ceed7e5e1679b6f9abcde28daab7aad78256588a1848430af233c86.jpg", + "image_caption": [ + "Figure 4. Qualitative matching results. We add the curvature similarity extractor module to LoFTR [55] and QuadTree [58] and compare the two approaches in both indoor and outdoor scenes. By adding the CSE module, we are able to obtain more correct matches (green lines) and better pose accuracy while keeping similar matching precision. The matches are coloured by the epipolar error threshold in [55]." + ], + "image_footnote": [], + "bbox": [ + 76, + 327, + 890, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 595, + 209, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first demonstrate the effectiveness of the proposed similarity curvature extractor on relative pose estimation and the downstream visual localization tasks. Next, we analyze the choice of architecture and hyper-parameters in the ablation study. Finally, we provide qualitative results to help understand the similarity curvature extractor.", + "bbox": [ + 75, + 623, + 470, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our proposed CSE adopts Mi-DAS + ViT [42] as the default depth predictor. For detector-based approaches [55, 58], we extract the curvature similarity at keypoint locations and use multi-scales $\\{7 * 7, 9 * 9, 11 * 11\\}$ . For detector-free matchers, we use the default grid patch size $8 * 8$ for curvature estimation.", + "bbox": [ + 75, + 715, + 470, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the depth predictor training, following the default setting in LoFTR [55], we fine-tune the predictor with the outdoor matching model on MegaDepth [27]. The model is optimized using Adam with a fixed learning rate of $r = 1e - 5$ and batch size of 2. The indoor model is fine-tuned on ScanNet [9] with $r = 5e - 6$ . In the training process, the", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "weights from the matcher are frozen, and only the weights from refineNets in MiDAS + ViT are updated. Instead of sampling 200 pairs from each scene, we reduce it to 10 pairs. The full training and evaluation process of the outdoor model is conducted on 2 Telsa-V100 GPUs with 32GB of memory. We train and evaluate the indoor model on 2 RTX 2080 Ti GPUs with a memory of 12GB.", + "bbox": [ + 496, + 595, + 893, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Relative Pose Estimation", + "text_level": 1, + "bbox": [ + 498, + 710, + 725, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our CSE on three popular datasets, ScanNet [9], MegaDepth [27], and YFCC100M [59], for the two-view pose estimation task. ScanNet is an indoor RGB-D dataset consisting of 1,613 sequences with 2.5 million views, and each view is associated with a ground truth camera pose and depth map. Similar to the setting in LoFTR [49], the images and depth maps are resized to $640*480$ for training and evaluation. MegaDepth and YFCC100M are two outdoor datasets which consist of multiple scenarios. We fine-tune the outdoor model on MegaDepth with the longest dimension of the images re", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17986", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sized to 840 and evaluate our method on MegaDepth-1500 from [55] and YFCC100M-4000 from [49].", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. The proposed CSE is a plug-in module that can be applied with both detector-based and detector-free matchers. We add this module to detector-based SuperGlue with SuperPoint [12] features, detector-free LoFTR, and its recent variant QuadTree [58]. The experiments follow the default setting from the original papers for all the baselines. For all the experiments on ScanNet, the image resolution is consistently $640 \\times 480$ . For evaluation on MegaDepth and YFCC100M, the default input sizes are different among the approaches. Thus, it is not reasonable to compare among the baselines as the input image size significantly impacts the final accuracy. We detail the default input sizes of different approaches in the supplementary material.", + "bbox": [ + 75, + 121, + 468, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. Similar to the previous methods [49, 55], we report the Area Under the recall Curve (AUC) of the translation and rotation errors with multiple thresholds $(5^{\\circ}, 10^{\\circ}, 20^{\\circ})$ . For each setting, we consider a pose to be correct if the maximum error from translation and rotation error is below the threshold. The estimated camera poses are recovered from the essential matrix estimated by RANSAC [17] with the inlier-outlier threshold set at 0.5 pixels. Besides, we report the matching precision and score [69, 12]. To calculate matching scores on detector-free matchers, we consider the number of coarse-level features as the number of detected keypoints, i.e., $h * w$ .", + "bbox": [ + 75, + 319, + 470, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We report the results of the CSE add-on on SuperGlue, LoFTR, and QuadTree in Tab. 1. For LoFTR and QuadTree, we report the results of directly adding the CSE module and adding CSE with fine-tuning. As the official training code for SuperGlue is not publicly available, we evaluated the SuperGlue + CSE without fine-tuning. However, we believe that fine-tuning would further improve the results of SuperGlue as well. We also notice that the MegaDepth-1500 is a part of the training set for SuperGlue. However, this does not invalidate the fact that the proposed method helps in finding better features. The proposed CSE improves the pose AUC scores of all methods on almost all datasets. Fine-tuning further increases the accuracy, and QuadTree combined with the proposed CSE method and fine-tuned on depth predictions achieves state-of-the-art AUC scores on all tested datasets.", + "bbox": [ + 75, + 501, + 468, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for matching score and precision, we also visualize the matching results of LoFTR and QuadTree in Fig. 4. By adding CSE, we obtain more correct matches and better pose accuracy on surface regions with distinctive geometry, while keeping similar matching precision.", + "bbox": [ + 75, + 743, + 468, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Visual Localization", + "text_level": 1, + "bbox": [ + 76, + 830, + 261, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Metrics. We select the Aachen DAY-NIGHT v1.1 dataset [52] to demonstrate the effectiveness of our proposed approach on the visual localization task.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/12dc6cdca5d4178357a03f0e052be14856340065336cdc816531be4c3693ddee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDayNight
(0.25m, 2°) / (0.5m, 5°) / (5m,10°)
SP [12] + SG [49]89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
SP + SG + CSE90.7 / 96.5 / 99.475.9 / 91.1 / 100.0
", + "bbox": [ + 504, + 89, + 890, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aachen DAY-NIGHT is a challenging urban-scale outdoor dataset which is collected by handheld devices with 6,697 reference images and 1,015 queries, including 191 nighttime images. The ground truth 6DoF camera pose is obtained by COLMAP [53] and refined by [73]. We follow the evaluation protocols of the visual localization benchmark [52] reporting the translation and rotation error recalls at $0.25\\mathrm{m} / 2^{\\circ}$ , $0.5\\mathrm{m} / 5^{\\circ}$ , and $5\\mathrm{m} / 10^{\\circ}$ .", + "bbox": [ + 496, + 255, + 890, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We add our CSE to the SuperPoint [12] + SuperGlue [49] pipeline in HLoc [48, 49] without fine-tuning any models. The matching pairs are provided by HLoc with top-50 candidate images from NetVLAD [2]. Tab. 2 reports the pose error recalls. We observe that the plug-in CSE improves the SP + SG pipeline on all thresholds on the Day sequences. On the Night sequence, it improves at threshold $(0.5\\mathrm{m} / 5^{\\circ})$ and $(5\\mathrm{m} / 10^{\\circ})$ . We believe the slight decrease at $(0.25\\mathrm{m} / 1^{\\circ})$ stems from the inaccuracies of depth prediction on nighttime images. Overall, by adding the CSE module, the localization accuracy is improved.", + "bbox": [ + 496, + 376, + 892, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 554, + 653, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyse the impact of different design components on our curvature similarity extractor in this section. We evaluate two main design choices: (1) multi-scale curvature extraction, (2) choice of depth predictors, and (3) mixing parameter value $\\lambda$ selection. All experiments are conducted with the Quadtree [58] matcher.", + "bbox": [ + 496, + 579, + 890, + 670 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d9251fcd1393f767a6f045cad79ec023821990f38e9bda8bda2a61f849ea2d56.jpg", + "table_caption": [ + "Table 2. Visual localization. We report the pose recall at $(0.25\\mathrm{m} / 2^{\\circ}$ $0.5\\mathrm{m} / 5^{\\circ}$ $5\\mathrm{m} / 10^{\\circ})$ on the Aachen Day-Night [52] dataset with the HLoc [48] algorithm. The best results are bold." + ], + "table_footnote": [], + "table_body": "
Multi-ScaleScanNet [9]MegaDepth [27]Time (ms)
AUC@5°/ AUC@10°/ AUC@20°
4*424.444.661.952.769.682.0113
6*625.345.662.853.369.981.8115
8*8 (default)25.245.562.554.070.382.1112
4*4+6*6+8*825.645.262.454.070.082.1338
", + "bbox": [ + 503, + 683, + 890, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Multi-scale and different grid sizes. The AUC scores and average CSE run-times (ms) of Quadtree [58] + CSE with different grid sizes and on multi-scale estimation.", + "bbox": [ + 498, + 796, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-scale. As described in Sec. 3.3, we conduct multiscale curvature similarity extraction on a $8*8$ patch grid with QuadTree by default. In Tab. 3, we show the results", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17987", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c98367b5e4af7a7e71cf8278c4b400303765747e06d48d9c431449ebc67e6504.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Depth PredictorsCSEFTAUC@5@10@20Model
Ground TruthQuadricX25.145.062.0153 MB
MiDAS (Res101)Quadric✓25.545.562.5575 MB
MiDAS (Real-time)Quadric✓25.545.462.9239 MB
MiDAS (ViT)Quadric✓25.846.163.41.6 GB
MiDAS (ViT)Surface FitX24.844.761.61.6 GB
", + "bbox": [ + 81, + 88, + 467, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "when using $4 * 4$ , $6 * 6$ , and multi-scale grids without fin-tuning. Even though the proposed CSE is robust to the grid size and the accuracy is similar, it slightly increases together with the grid size. Size $4 * 4$ leads to the worst solutions on average, while $8 * 8$ leads to the best ones. As expected, curvature similarity extraction with multi-scale shows the best performance. However, it increases the average run-time by three times.", + "bbox": [ + 75, + 263, + 467, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth Predictors and Curvature Extractors. We then explore the impact of different depth predictors on the final pose accuracy and model size. We select the ground truth depth, MiDAS with ResNet101 [22] architecture and MiDAS small (real-time) to compare with the default predictor MiDAS with ViT [64]. Tab. 4 reports the pose accuracy of these approaches on the ScanNet dataset. MiDAS with ViT, which shows the best performance, has the largest model size due to the heavy depth prediction network. When we employ lighter models for depth prediction, we observe that the performance has only a small drop, indicating that our method is still effective when working with lightweight networks. We also notice that the pose accuracy with ground truth depth is worse than the fine-tuned models. We believe the reason for this is the missing depth regions in the ground truth. We visualize the depth maps in different conditions at supplementary material.", + "bbox": [ + 75, + 385, + 467, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The last two rows of the table show the results with the proposed ellipse and with quadratic surface fitting. The ellipse fitting performs more accurately in the experiments.", + "bbox": [ + 75, + 642, + 467, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mixing Parameter Value. We finetune our model with different values for $\\lambda$ in Eq. (10). The results are reported in Tab. 5. The proposed model performs best with $\\lambda = 0.9$ for outdoor and $\\lambda = 0.8$ for indoor scenes. This is expected since the proposed CSE is additional guidance for coarse-level matching that helps reduce the score of incorrect matches stemming from different surfaces. Assigning a large mixing weight could degrade the performance.", + "bbox": [ + 75, + 686, + 467, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Curvature Similarity Understanding", + "text_level": 1, + "bbox": [ + 76, + 816, + 393, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we visualize how the curvature similarity score impacts the key points selection and matching results. We first show in the curvature images that similar surfaces share similar curvature values (more visualizations are presented", + "bbox": [ + 76, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a7b1191f08f16c471a9b092dd9432e7f357cdd849b6351d5b9f248cef639e60d.jpg", + "table_caption": [ + "Table 4. Depth study. The AUC scores of QuadTree [58] with the proposed CSE using the GT depth or MiDaS [42] with different models. Quadric and surface fitting results are also shown." + ], + "table_footnote": [], + "table_body": "
λ (AUC@5°)1.00.90.80.70.50.3
ScanNet [9]24.925.826.123.821.017.8
MegaDepth [27]53.554.552.847.045.539.8
", + "bbox": [ + 509, + 89, + 880, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Mixing Parameter Value Selection. We finetune the Quadtree + SCE with different $\\lambda$ values and report the AUC@ $5^{\\circ}$ on ScanNet and MegaDepth datasets.", + "bbox": [ + 498, + 161, + 890, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fa22869d52a9886e5c4e8f3d9af57eb42671b6d379237f65f8f5101ca1afa67.jpg", + "image_caption": [ + "(a) Points found by QuadTree alone (left) and when combined with CSE (right)" + ], + "image_footnote": [], + "bbox": [ + 501, + 217, + 888, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2aeab890e7e71470619bb07ef73c9eb3e83731bbeb1377e30a7685f4aee942d3.jpg", + "image_caption": [ + "(b)Matches found by QuadTree alone (left) and when combined with CSE (right)", + "Figure 5. Keypoints (top row) and matches (bottom) found by the QuadTree matcher [58] with and without the proposed CSE, overlayed on the RGB (left) and extracted curvature images (right)." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 888, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the supplementary). Then, we observe that, after adding CSE, more matches are detected from the same surface, as we increase the matching confidence of those matches by assigning a larger curvature similarity score. Besides, the curvature score helps in lowering the matching confidence of matches from different surfaces, e.g., the incorrect (red) matches in Fig. 5 (b) left are eliminated in the right.", + "bbox": [ + 496, + 434, + 890, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 556, + 625, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a curvature similarity extractor (CSE), a new algorithm to leverage 3D geometric cues in local feature matching. By fitting quadrics to depth maps obtained from off-the-shelf monocular depth predictors, we extract curvature similarity, which is invariant to translation, rotation, and scaling. CSE can be seamlessly integrated with matchers to guide the match selection. Also, it allows training feature matches and monocular depth networks jointly minimizing the matching loss in an end-to-end manner to further improve the accuracy. Our experiments in Sec. 4 demonstrate the effectiveness of the proposed CSE module, which consistently improves the pose AUC scores of all methods on almost all datasets, achieving state-of-the-art performance on all benchmarks when combined with QuadTree.", + "bbox": [ + 496, + 582, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. First, the integration of depth prediction and curvature similarity extraction into feature matching leads to increased memory consumption and inference time. An alternative solution for time-sensitive applications is to use a lightweight depth predictor. As shown in Tab. 4, such an approach still increases the accuracy. Second, the discussed invariance properties do", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17988", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "not hold at depth discontinuities, and limitations may arise in scenes lacking accurate depth or clear surfaces. Even though this did not pose an issue in our experiments, this is a potential future direction that might lead to further improvements.", + "bbox": [ + 76, + 90, + 468, + 165 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the Academy of Finland (grant No. 327911, No. 353138). Daniel Barath is supported by ETH Postdoc fellowship. We acknowledge the computational resources provided by the Aalto Science-IT project and CSC-IT Center for Science, Finland.", + "bbox": [ + 76, + 167, + 472, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 273, + 173, + 287 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pablo Fernández Alcantarilla, Adrien Bartoli, and Andrew J Davison. Kaze features. In Proceedings of European Conference on Computer Vision, pages 214-227, 2012. 1", + "[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 7", + "[3] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2911-2918, 2012. 1", + "[4] Daniel Barath and Jií Matas. Graph-cut ransac. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733-6741, 2018. 2", + "[5] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1304-1312, 2020. 2", + "[6] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 1, 2", + "[7] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 1, 2", + "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In Proceedings of the European Conference on Computer Vision, 2022. 1, 2, 5", + "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 6, 7, 8", + "[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly sur" + ], + "bbox": [ + 78, + 299, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "face reintegration. ACM Transactions on Graphics, 36(4):1, 2017. 1", + "[11] Luanyuan Dai, Yizhang Liu, Jiayi Ma, Lifang Wei, Taotao Lai, Changcai Yang, and Riqing Chen. Ms2dg-net: Progressive correspondence learning via multiple sparse semantics dynamic graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8973-8982, 2022. 2", + "[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 224-236, 2018. 1, 2, 7", + "[13] M.P. do Carmo. Differential Geometry of Curves and Surfaces: Revised and Updated Second Edition. Dover Books on Mathematics. Dover Publications, 2016. 3", + "[14] Siyan Dong, Shuzhe Wang, Yixin Zhuang, Juho Kannala, Marc Pollefeys, and Baoquan Chen. Visual localization via few-shot scene region classification. International Conference on 3D Vision, 2022. 1", + "[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 1, 2", + "[16] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2", + "[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 7", + "[18] Per-Erik Forssén and David G Lowe. Shape descriptors for maximally stable extremal regions. In International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3", + "[19] Khang Truong Giang, Soohwan Song, and Sungho Jo. Curvature-guided dynamic scale networks for multi-view stereo. In International Conference on Learning Representations, 2022. 3", + "[20] Jack Goldfeather and Victoria Interrante. A novel cubic-order algorithm for approximating principal direction vectors. ACM Trans. Graph., 23(1):45-63, 2004. 4", + "[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. CiteSeer, 1988. 1", + "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 8", + "[23] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17989", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 2", + "[25] Qingde Li and J.G. Griffiths. Least squares ellipsoid specific fitting. In Geometric Modeling and Processing, 2004. Proceedings, pages 335-340, 2004. 4, 5", + "[26] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. Advances in Neural Information Processing Systems, 33:17346-17357, 2020. 2", + "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 6, 7, 8", + "[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3", + "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1", + "[30] Xiaotao Liu, Chen Meng, Fei-Peng Tian, and Wei Feng. Dgd-net: Local descriptor guided keypoint detection network. In 2021 IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 2", + "[31] Yuan Liu, Lingjie Liu, Cheng Lin, Zhen Dong, and Wenping Wang. Learnable motion coherence for correspondence pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2021. 2", + "[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5", + "[33] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019. 2", + "[34] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004. 1", + "[35] Krystian Mokolajczyk and Cordelia Schmid. Scale & affine invariant interest point detectors. International Journal of Computer Vision, 60(1):63-86, 2004. 1", + "[36] Krystian Mikolajczyk, Tinne Tuytelaars, Cordelia Schmid, Andrew Zisserman, Jiri Matas, Frederik Schaffalitzky, Timor Kadir, and L Van Gool. A comparison of affine region detectors. International Journal of Computer Vision, 65(1):43-72, 2005. 1", + "[37] Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Repeatability is not enough: Learning affine regions via discriminability. In Proceedings of the European Conference on Computer Vision, pages 284-300, 2018. 3", + "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1", + "[39] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 1", + "[40] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3456-3465, 2017. 1", + "[41] Rémi Pautrat, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Online invariance selection for local feature descriptors. In European Conference on Computer Vision, pages 707-724. Springer, 2020. 1", + "[42] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 2, 3, 4, 6, 8", + "[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2, 3", + "[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 1, 2", + "[45] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2", + "[46] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[47] John Rugis and Reinhard Klette. A scale invariant surface curvature estimator. In Pacific-Rim Symposium on Image and Video Technology, pages 138-147. Springer, 2006. 3", + "[48] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 7", + "[49] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 5, 6, 7", + "[50] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In European conference on computer vision, pages 752-765. Springer, 2012. 1", + "[51] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17990", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 1, 7", + "[53] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 1, 7", + "[54] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 2", + "[55] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7", + "[56] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11286-11295, 2020. 2", + "[57] Dongli Tan, Jiang-Jiang Liu, Xingyu Chen, Chao Chen, Ruixin Zhang, Yunhang Shen, Shouhong Ding, and Rongrong Ji. Eco-tr: Efficient correspondences finding via coarse-to-fine refinement. Proceedings of the European Conference on Computer Vision, 2022. 2", + "[58] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1, 2, 5, 6, 7, 8", + "[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 6", + "[60] Yurun Tian, Vassileios Balntas, Tony Ng, Axel BarrosoLaguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In Proceedings of the Asian Conference on Computer Vision, 2020. 2", + "[61] Carl Toft, Daniyar Turmukhambetov, Torsten Sattler, Fredrik Kahl, and Gabriel J Brostow. Single-image depth prediction makes feature matching easier. In European Conference on Computer Vision, pages 473-492. Springer, 2020. 2, 3", + "[62] Lijing Tong and Xiang Ying. 3d point cloud initial registration using surface curvature and surf matching. 3D Research, 9:1-16, 2018. 3", + "[63] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 1, 2", + "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 8" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, 2022. 1, 2", + "[66] Shuzhe Wang, Zakaria Laskar, Iaroslav Melekhov, Xiaotian Li, and Juho Kannala. Continual learning for image-based camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3252-3262, 2021. 1", + "[67] Changchang Wu. Towards linear-time incremental structure from motion. In International Conference on 3D Vision, pages 127-134. IEEE, 2013. 1", + "[68] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 3", + "[69] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 1, 2, 7", + "[70] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2018. 2", + "[71] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 3", + "[72] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Hongen Liao. Learning two-view correspondences and geometry using order-aware network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5845-5854, 2019. 2", + "[73] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7", + "[74] Chen Zhao, Yixiao Ge, Feng Zhu, Rui Zhao, Hongsheng Li, and Mathieu Salzmann. Progressive correspondence pruning by consensus learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6464-6473, 2021. 2" + ], + "bbox": [ + 501, + 90, + 890, + 738 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17991", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_model.json b/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7dfc085390455d7511a44eb1faa77cc404a08003 --- /dev/null +++ b/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.776, + 0.154 + ], + "angle": 0, + "content": "Guiding Local Feature Matching with Surface Curvature" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.18, + 0.282, + 0.2 + ], + "angle": 0, + "content": "Shuzhe Wang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.181, + 0.434, + 0.198 + ], + "angle": 0, + "content": "Juho kannala" + }, + { + "type": "text", + "bbox": [ + 0.482, + 0.18, + 0.623, + 0.199 + ], + "angle": 0, + "content": "Marc Pollefeys\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.67, + 0.181, + 0.79, + 0.198 + ], + "angle": 0, + "content": "Daniel Barath2" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.223, + 0.378, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)Aalto University" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.223, + 0.542, + 0.241 + ], + "angle": 0, + "content": "\\(^{2}\\)ETH zurich" + }, + { + "type": "text", + "bbox": [ + 0.601, + 0.224, + 0.692, + 0.241 + ], + "angle": 0, + "content": "3Microsoft" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.294, + 0.314, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.474, + 0.57 + ], + "angle": 0, + "content": "We propose a new method, called curvature similarity extractor (CSE), for improving local feature matching across images. CSE calculates the curvature of the local 3D surface patch for each detected feature point in a viewpoint-invariant manner via fitting quadrics to predicted monocular depth maps. This curvature is then leveraged as an additional signal in feature matching with off-the-shelf matchers like SuperGlue and LoFTR. Additionally, CSE enables end-to-end joint training by connecting the matcher and depth predictor networks. Our experiments demonstrate on large-scale real-world datasets that CSE consistently improves the accuracy of state-of-the-art methods. Fine-tuning the depth prediction network further enhances the accuracy. The proposed approach achieves state-of-the-art results on the ScanNet dataset, showcasing the effectiveness of incorporating 3D geometric information into feature matching." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.614, + 0.21, + 0.63 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Local feature matching is a crucial component for many geometric computer vision tasks, including visual localization [50, 51, 52, 48, 14, 66], structure-from-motion (SfM) [67, 53, 29], and simultaneous localization and mapping (SLAM) [38, 39, 10]. Given a pair of images observing a 3D scene, the task is to find reliable tentative point-to-point correspondences in the two images. Forming such feature matches is often a challenging task as the images may undergo large viewpoint and illumination changes, have occlusions or repetitive patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.471, + 0.853 + ], + "angle": 0, + "content": "The standard pipeline for local feature matching typically involves two steps: (1) keypoint detection and description and (2) point-wise feature matching. Traditional approaches mainly focus on improving the robustness of" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.293, + 0.695, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.296, + 0.891, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.441, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Figure 1. Curvature-guided match selection: (a) Correct matches in two views must share the same underlying surface curvature, e.g., predicted from monocular depth. (b) Correspondences from different surfaces are guaranteed to be incorrect." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.893, + 0.629 + ], + "angle": 0, + "content": "keypoint detection and description through, for example, extending the Harris [21] detector to handle affine transformations and multiple scales [35, 36] or using more discriminative or efficient descriptors [32, 7, 34, 1, 3]. However, despite their unbroken popularity, these algorithms often fail to cope with the challenges that arise in real-world environments, leading to low accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Recent advances in deep learning-based feature matching have made significant progress in addressing the limitations of hand-crafted approaches, such as by jointly training detectors and descriptors [69, 12, 15, 44, 63, 40] with convolutional neural networks (CNNs), or combining hand-crafted and learning-based descriptors in a hybrid manner [41, 6]. SuperGlue [49] introduced the use of transformer networks to learn the matching process and formulate the problem as an optimal transport task [64]. LoFTR [55] and its recent variants [58, 8, 65] leverage both the global and local context from raw images by jointly learning the feature extraction and matching in a single network. While these approaches have led to state-of-the-art performance on several benchmarks, they work entirely in the 2D image domain and ignore the underlying 3D geometry of the scene. As feature matching is, essentially, finding the corresponding projections of an actual 3D point, this could be a critical limitation in real scenarios where the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.451, + 0.876 + ], + "angle": 0, + "content": "*Part of the work was done during the author's visit to ETH zurich." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.468, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code and trained models are available at https://github.com/AaltoVision/surface-curvature-estimator." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "matched pixels from different views must share the geometry of the underlying surface patch, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.123, + 0.468, + 0.424 + ], + "angle": 0, + "content": "In this paper, we focus on investigating and exploiting 3D geometry cues, e.g. coming from monocular depth predictions, for feature matching. The only prior work is [61], which adopts depth priors to find planar image regions that are then rectified to eliminate viewpoint changes. Finally, a handcrafted detector and descriptor are applied to the rectified image regions. This, however, relies on the heavy assumption that the images consist mostly of dominant planes, which severely restricts its out-of-the-box applicability. In this work, we go further by exploring the local surface geometry coming from depth priors. This additional geometric clue is coupled with any off-the-shelf feature matcher increasing its accuracy. Benefiting from advanced monocular depth predictors [42, 43, 16], obtaining depth priors is easy and costs only a few milliseconds. We use these depth predictions to extract curvature at the observed feature points. Exploiting the fact that the curvature is invariant to viewpoint changes (i.e., rotation, translation, and scaling), we enforce that matched features must lie on similar surfaces. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.436, + 0.468, + 0.511 + ], + "angle": 0, + "content": "1. We propose an approach for feature matching, leveraging dense depth via utilizing local surface curvature similarity, which is invariant to viewpoint changes. This approach is a departure from traditional matching methods that rely solely on image information." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.468, + 0.598 + ], + "angle": 0, + "content": "2. The proposed curvature similarity extractor can be seamlessly integrated with any recent feature matcher, making it a versatile tool that can be easily adopted. The experiments show that it improves several recent algorithms on various benchmarks by \\(1 - 3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.611, + 0.468, + 0.686 + ], + "angle": 0, + "content": "3. The proposed algorithm can be used to train feature matchers and monocular depth prediction networks jointly in an end-to-end fashion. To demonstrate this, we fine-tune the state-of-the-art MiDaS [42] depth predictor to increase the feature matching accuracy." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.436, + 0.468, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.702, + 0.214, + 0.716 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.726, + 0.297, + 0.743 + ], + "angle": 0, + "content": "2.1. Local Feature Matching" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Detector-based Matchers build sparse correspondences on top of the detected keypoints. Therefore, robust keypoint detection and feature extraction are essential for these approaches. Traditional hand-crafted descriptors, such as SIFT [32] and SURF [7], follow a detect-then-describe pipeline and have shown great success since the 2000s. More recently, learned detectors and descriptors [6, 69, 12, 15, 44, 63, 60, 30] with convolutional neural networks show their superiority in matching images with large viewpoint and appearance changes. This traditional pipeline is further" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.363 + ], + "angle": 0, + "content": "developed to detect-and-describe [12, 15, 44] and describe-to-detect [60, 30] strategies leveraging high-level image information for accurate and reliable matching in challenging conditions. Besides substituting the traditional detectors and descriptors with their learned counterparts for image matching, the recent SuperGlue [49] focuses on the matching stage itself. It replaces the naive mutual nearest neighbour search with Graph Neural Networks (GNN). SuperGlue takes the sparse descriptors and their positional encoding as inputs, then leverage a transformer-based [64] network to create a more robust feature representation for the optimal partial assignment. The follow-up work [54] further improves the performance and efficiency by adaptively clustering sparse features into different subgraphs and using a coarse-to-fine paradigm. Although showing significant improvements with the recent components, the detector-based matchers are naturally limited by the accuracy of detected keypoints coming from independent detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Detector-free Matchers directly estimate the dense correspondences from raw images without an independent keypoint detection stage. These approaches leverage low and high-level image information with neural networks and can distinguish indistinctive regions across images. The pioneer work of [46] constructs 4D cost volumes, significantly increasing the matching accuracy. However, it is computationally expensive as the complexity is \\(\\mathcal{O}(n^2)\\), where \\(n\\) is the number of pixels or patches. [45, 26] alleviate the problem by applying either sparse convolutions or a coarse-to-fine paradigm. More recently, transformer-based detector-free matchers [55, 58, 65, 8, 24, 57] have received attention due to their strong performance on local feature matching. The most representative work, LoFTR [55], inherits the advantages of graph matching from SuperGlue and leverages a linear transformer [64] for efficient dense matching. [58, 8] utilise more efficient transformer structures to improve the performance. Detector-free matchers show a promising direction towards local feature matching by encoding rich image information. We claim that extracting 3D geometric information could further enhance accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Correspondence Pruning approaches [17, 4, 5, 72, 70, 56] apply a consensus mechanism to filter out the outliers from putative correspondences coming from the feature matcher procedure. RANSAC [17] and its follow-up works [4, 5] are the most popular correspondence pruning algorithms. In the era of deep learning, OANet [72] infers the probabilities of correspondences being inliers with an order-aware network. Other works improve the accuracy of correspondence pruning by applying motion coherence constraints [31, 33], leveraging local-to-global consensus learning procedure [74, 11], and adopting the attention mechanism [56]. These approaches provide inlier probability predictions for each putative correspondence that can be used either to filter or order matches. Finally, least squares fitting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "17982" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.454, + 0.108 + ], + "angle": 0, + "content": "or traditional robust estimation is applied to the matches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.115, + 0.424, + 0.132 + ], + "angle": 0, + "content": "2.2. 3D Geometric Priors to Image Matching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.473, + 0.396 + ], + "angle": 0, + "content": "Incorporating 3D geometric priors into various vision tasks has been widely explored in recent years. [23] leverages the RGB-D reconstructions to learn view-invariant, geometry-aware 2D representations for downstream tasks. [71] introduces the predicted monocular depth to guide the optimization of neural scene representation. The surface curvature, a popular 3D geometric cue, is widely applied to 3D vision tasks such as 3D point cloud registration [62], multi-view stereo [68, 19]. However, integrating 3D geometric cues into 2D image matching has received little attention. [37, 18] assume that the underlying surface patch is locally planar for feature detection and description, while [61] uses prior depth information to improve feature extraction by rectifying large planar regions. None of the above approaches considers the 3D geometric information in the local region, and we are the first to include 3D surface geometry in general 2D feature matching." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.407, + 0.456, + 0.425 + ], + "angle": 0, + "content": "3. Feature Matching with Surface Curvature" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.54 + ], + "angle": 0, + "content": "This section proposes an approach to improve any local feature matcher by extracting information about the underlying local 3D surface predicted by a deep network. The overview of our method coupled with LoFTR [55] is shown in Fig. 2. Although the proposed algorithm is compatible with any matcher, we demonstrate its effectiveness with LoFTR. Combining it with other ones is straightforward." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.547, + 0.291, + 0.563 + ], + "angle": 0, + "content": "3.1. LoFTR-style Matchers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.754 + ], + "angle": 0, + "content": "This section provides a brief overview of LoFTR. Given an image pair \\( I_A, I_B \\in \\mathbb{R}^{H*W*3} \\), the end-to-end trainable matcher, first, extracts coarse-level features \\( \\mathcal{F}_A^c, \\mathcal{F}_B^c \\) and, then, fine-level ones \\( \\mathcal{F}_A^f, \\mathcal{F}_B^f \\) with a local feature CNN [28]. At the coarse level transformer module, the features are flattened and processed by a transformer-based architecture with multiple self- and cross-attention layers [64]. A matching confidence map \\( \\mathcal{P}_c \\) is predicted at the end of this module, and the matches \\( \\mathcal{M}_c \\) with high confidence are selected from \\( \\mathcal{P}_c \\) to fine-tune the prediction of fine-level features matching at the coarse-to-fine transformer module. The final outputs are the fine-level matches \\( \\mathcal{M}_f \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.761, + 0.355, + 0.777 + ], + "angle": 0, + "content": "3.2. Curvature Similarity Extractor" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.471, + 0.83 + ], + "angle": 0, + "content": "In this section, we will discuss multiple potential ways of extracting local surface curvature at detected features. This curvature is then used for improving feature matching." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.279, + 0.862 + ], + "angle": 0, + "content": "3.2.1 Curvature Similarity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For an at least two times continuously differentiable surface \\(S \\in \\mathbb{C}^2\\) in the 3D space, given a point \\(p \\in \\mathbb{R}^3\\) and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "a direction \\(d \\in \\mathbb{R}^3\\) on \\(S\\); the normal curvature \\(k_n \\in \\mathbb{R}\\) measures how curved the surface is at this point along the direction [13]. Since there are infinitely many directions that travel through a point on the surface, there are also infinitely many normal curvatures at a given point. In differential geometry, the minimum normal curvature \\(k_1\\) and the maximum curvature \\(k_2\\) at point \\(p\\) are defined as the principal curvatures, where \\(k_1 \\leq k_2\\), and the directions of the principal curvatures are the smoothest and steepest directions on the surface. We also define the mean curvature \\(H\\) and Gaussian curvature \\(K\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.266, + 0.892, + 0.295 + ], + "angle": 0, + "content": "\\[\nH = \\frac {1}{2} \\left(k _ {1} + k _ {2}\\right), K = k _ {1} k _ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.426 + ], + "angle": 0, + "content": "A highly beneficial property of such curvatures is their invariance to surface rotation and translation. Both are important when describing local surface patches centered on features detected in 2D images. However, these measures do not preserve scale invariance [13], which is essential in our case due to using predicted relative depth maps instead of a metric one. This means that, in each image, the depth is only defined up-to-scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.486 + ], + "angle": 0, + "content": "To seek the scale invariance of the curvatures for local patch matching, we follow [47] for the idea of scale-invariant curvature measure and define the curvature similarity function \\(S: \\mathbb{R} \\times \\mathbb{R} \\to \\mathbb{R}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.497, + 0.892, + 0.531 + ], + "angle": 0, + "content": "\\[\nS \\left(k _ {1}, k _ {2}\\right) = \\frac {\\operatorname* {m i n} \\left(\\left| k _ {1} \\right| , \\left| k _ {2} \\right|\\right)}{\\operatorname* {m a x} \\left(\\left| k _ {1} \\right| , \\left| k _ {2} \\right|\\right)}, (0 \\leq S (k) \\leq 1), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.678 + ], + "angle": 0, + "content": "where \\( k_{1} \\) and \\( k_{2} \\) are the minimum and maximum curvatures, respectively. It is easy to see that \\( S(k_{1},k_{2}) \\) is rotation, translation and scale invariant. Case \\( S(k_{1},k_{2}) = 1 \\) is interpreted as the curvature of a point on a spherical surface. In the case when \\( k_{1} = k_{2} = 0 \\), we define \\( S(k_{1},k_{2}) \\) to be 0. Detecting points on planes is also straightforward from \\( k_{1} \\) and \\( k_{2} \\) as they both will become \\( \\infty \\). Proof of the invariances is provided in the supplementary material. Next, we will discuss ways of obtaining curvature similarities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.771, + 0.71 + ], + "angle": 0, + "content": "3.2.2 Surface Curvatures Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "This section introduces our approximate form of principal curvature estimation. Different from prior work [68, 19] that extract the curvatures from image intensity, we estimate the principal curvatures based on the depth map. This 3D geometric cue could be obtained either from the active sensors (LiDAR, Kinect) or monocular depth predictors [42, 43]. Since our target curvature similarity \\( S(k_{1}, k_{2}) \\) is scaling invariant, the scaling ambiguity in monocular depth estimation is neglected and the predicted relative depth is sufficient for the curvature extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this work, we consider only a pair of RGB images as input and leverage any off-the-shelf depth predictor, e.g." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.092, + 0.849, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.896, + 0.39 + ], + "angle": 0, + "content": "Figure 2. Overview of the CSE with LoFTR. For a pair of images, it follows the LoFTR [55] matching pipeline to extract the coarse and fine level image features and generate a coarse matching confidence map \\(\\mathcal{P}_c\\). In the meantime, the images are fed to a monocular depth network to predict the depth map and extract the curvature similarity upon the depth map. We formulate a new curvature score matrix \\(\\mathcal{P}_{curv}\\) based on the pixel-wise \\(L2\\) distance of the curvature similarities between two images and mix it with \\(\\mathcal{P}_c\\). The mixed score matrix is then processed by the coarse-to-fine transformer module together with the fine-level features to predict the final matches \\(\\mathcal{M}_f\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.489 + ], + "angle": 0, + "content": "MiDaS [42], to extract the depth map \\( Z_{i}:\\mathbb{R}\\times \\mathbb{R}\\to \\mathbb{R} \\) for each input image, \\( i\\in \\{1,2\\} \\). Let \\( \\mathbf{r}_i(u,v,Z_i(u,v)) \\) be the 3D surface patch corresponding to the image region at coordinates \\( (u,v) \\) with window size \\( \\delta \\in \\mathbb{R}^{+} \\) in the \\( i \\)th image, where \\( Z_{i}(u,v) \\) is the depth of point \\( (u,v) \\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.49, + 0.472, + 0.535 + ], + "angle": 0, + "content": "Let us formulate a local point set \\(\\mathcal{P}_i = \\{(u,v,Z_i(u,v))\\mid (u,v)\\in \\mathcal{M}_f^i\\} \\subseteq \\mathbb{R}^3\\) , where \\(\\mathcal{M}_f^i\\) are the points detected in the \\(i\\) th image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.473, + 0.627 + ], + "angle": 0, + "content": "Principal Curvature via Surface Fitting. The most straightforward approach is first applying the quadratic surface approximation method [20] to estimate the surface \\(\\mathbf{r}_i(u,v,Z_i(u,v))\\) at all points in \\(\\mathcal{P}\\), then employing the first and second fundamental forms of a surface. The normal curvature is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.641, + 0.47, + 0.673 + ], + "angle": 0, + "content": "\\[\nk _ {n} = \\frac {I I}{I} = \\frac {L d u ^ {2} + 2 M d u d v + N d v ^ {2}}{E d u ^ {2} + 2 F d u d v + G d v ^ {2}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.743 + ], + "angle": 0, + "content": "where \\(L = \\mathbf{r}_{uu}\\mathbf{n}\\), \\(M = \\mathbf{r}_{uv}\\mathbf{n}\\), \\(N = \\mathbf{r}_{vv}\\mathbf{n}\\), \\(E = \\mathbf{r}_u\\mathbf{r}_u\\), \\(F = \\mathbf{r}_u\\mathbf{r}_v\\), \\(G = \\mathbf{r}_v\\mathbf{r}_v\\); \\(\\mathbf{r}_u, \\mathbf{r}_v, \\mathbf{r}_{uu}, \\mathbf{r}_{uv}, \\mathbf{r}_{vv}\\) are the first and second-order derivatives of the surface patch \\(\\mathbf{r}_i(u,v,Z_i(u,v))\\), and \\(\\mathbf{n} \\in \\mathbb{R}^3\\) is the normal." + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.756, + 0.47, + 0.788 + ], + "angle": 0, + "content": "\\[\nk _ {n} = \\frac {L + 2 M \\lambda + N \\lambda^ {2}}{E + 2 F \\lambda + G \\lambda^ {2}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Once we have the normal curvature \\( k_{n} \\), we then calculate the partial derivative of \\( k_{n} \\) w.r.t. \\( \\lambda \\) to obtain the principal curvatures \\( k_{1}, k_{2} \\). The details are presented in the supplementary. While this approach works accurately and efficiently in practice, we observed that its gradients become unstable during back-propagation. Therefore, we explore another way as well for calculating the surface curvature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.895, + 0.58 + ], + "angle": 0, + "content": "**Curvature Extraction via Quadrics.** Another lightweight procedure for estimating the curvature is via fitting quadric surfaces, e.g. ellipsoid, locally to the neighborhood of the observed point. Thus, instead of directly estimating a quadratic surface from the given depth map, we constrain the surface to be a quadric and, more specifically, an ellipsoid. An ellipsoid has the beneficial property that all its curvatures are semi-positive. Its principal axes define three directions which we can use the define the similarity. We use the radii \\(\\alpha, \\beta, \\gamma \\in \\mathbb{R}^{+}\\) along the three axes and define the curvature similarity score as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.588, + 0.892, + 0.625 + ], + "angle": 0, + "content": "\\[\nS (\\alpha , \\beta , \\gamma) = \\frac {\\operatorname* {m i n} \\left(\\alpha^ {2} , \\beta^ {2} , \\gamma^ {2}\\right)}{\\operatorname* {m a x} \\left(\\alpha^ {2} , \\beta^ {2} , \\gamma^ {2}\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.894, + 0.692 + ], + "angle": 0, + "content": "Similarity \\( S \\) is rotation, translation, scale and thus, viewpoint invariant. The advantage from Eq. (5) is that the computation of first and second-order gradients are no longer required for the curvature similarity calculation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.894, + 0.723 + ], + "angle": 0, + "content": "The general constraint that a quadric imposes on a point \\( x \\in \\mathcal{P}_i \\) lying on the surface is" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.732, + 0.783, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} \\mathbf {Q} _ {i} \\mathbf {x} ^ {\\mathrm {T}} + \\mathbf {P} _ {i} \\mathbf {x} ^ {\\mathrm {T}} + \\mathbf {R} _ {i} = 0,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.79 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_i\\), \\(\\mathbf{P}_i\\), and \\(\\mathbf{R}_i\\) are the parameters in a matrix form. This can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.797, + 0.892, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} a x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \\tag {6} \\\\ 2 f y z + 2 g x + 2 h y + 2 i z + j = 0. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.844, + 0.892, + 0.874 + ], + "angle": 0, + "content": "As described in [25], a sufficient condition that guarantees the quadric surface to be an ellipsoid is" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.884, + 0.751, + 0.901 + ], + "angle": 0, + "content": "\\[\n4 M - N ^ {2} = 1,\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "17984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.125, + 0.105 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.119, + 0.375, + 0.133 + ], + "angle": 0, + "content": "\\[\nN = a + b + c,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.136, + 0.377, + 0.153 + ], + "angle": 0, + "content": "\\[\nM = a b + b c + a c - d ^ {2} - e ^ {2} - f ^ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.165, + 0.469, + 0.24 + ], + "angle": 0, + "content": "Also, since we aim at measuring the curvature of the surface patch located on the observed point, we have to constrain the ellipsoid so that the observed point \\(\\mathbf{p}\\) is on its surface. To do so, we translate the point set \\(\\mathcal{P}\\) as \\(\\mathcal{P}' = \\{\\mathbf{q} - \\mathbf{p} \\mid \\mathbf{q} \\in \\mathcal{P}\\}\\) and fix \\(j = 0\\). The central point now is \\(\\mathbf{p}' = [0,0,0]^{\\mathrm{T}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.24, + 0.466, + 0.255 + ], + "angle": 0, + "content": "Substituting all these constraints into Eq. (6), we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.265, + 0.469, + 0.29 + ], + "angle": 0, + "content": "\\[\na x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.287, + 0.39, + 0.301 + ], + "angle": 0, + "content": "\\[\n2 f y z + 2 g x + 2 h y + 2 i z = 0.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.312, + 0.469, + 0.342 + ], + "angle": 0, + "content": "From Eq. (7), only nine variables need to be estimated in our case. For each point \\([u, v, z]^{\\mathrm{T}} \\in \\mathcal{P}^{\\prime}\\), we define \\(\\mathbf{X}_i\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.352, + 0.469, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {i} = \\left[ u _ {i} ^ {2}, v _ {i} ^ {2}, z _ {i} ^ {2}, 2 v _ {i} z _ {i}, 2 u _ {i} z _ {i}, 2 u _ {i} v _ {i}, 2 u _ {i}, 2 v _ {i}, 2 z _ {i} \\right] ^ {\\mathrm {T}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.38, + 0.469, + 0.442 + ], + "angle": 0, + "content": "and coefficient matrix \\(\\mathbf{C} = [\\mathbf{X}_1, \\mathbf{X}_2, \\mathbf{X}_3, \\dots, \\mathbf{X}_n]\\). Then, we follow the method in [25] for the ellipsoid fitting and estimate its radii \\(R = (\\alpha, \\beta, \\gamma)\\). The algorithms for fitting and calculating the radii are shown in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.45, + 0.385, + 0.464 + ], + "angle": 0, + "content": "3.3. Curvature-Guided Match Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.669 + ], + "angle": 0, + "content": "The proposed curvature similarity extractor (CSE) is a depth-based plug-in component that can be combined with both detector-based [32, 49] and detector-free [55, 58, 8] matchers. For an image pair \\((I_A, I_B)\\), the monocular depth predictor first estimates the depth map for each image independently. CSE extracts the curvature map \\((S_A, S_B)\\) on the top of the depth maps. Note that LoFTR selects the candidate matches \\(M_c\\) at the coarse level of 1/8-resolution grids where the ground truth matches are defined based on the reprojection distance of the centers of the grid cells. Thus, it is not necessary to extract the curvature similarity at each pixel. It is calculated only at the center point of each cell, i.e., \\(S_A, S_B \\in \\mathbb{R}^{(h * w)*1}\\), where \\(h = H / 8\\), \\(w = W / 8\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.759 + ], + "angle": 0, + "content": "Since the proposed curvature similarity is rotation, translation and scale invariant, \\(S_{A}\\) at point \\(\\mathbf{p}_A\\) in the first image should be equal to the similarity \\(S_B\\) of its corresponding pair \\(\\mathbf{p}_B\\) in the second image. The correct matches \\(M_c\\), ideally, have the same curvature values. To measure this, we define the normalised curvature score matrix as" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.772, + 0.469, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {\\text {c u r v}} = 1 - L _ {2} (S (k) _ {A}, S (k) _ {B}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.798, + 0.469, + 0.844 + ], + "angle": 0, + "content": "where \\(L_{2}(\\cdot ,\\cdot)\\) is the L2-norm of \\(S(k)_A\\) and \\(S(k)_B\\). Since both \\(S(k)_A\\) and \\(S(k)_B\\) are in-between zero and one, the curvature score is also normalized into range [0, 1]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.845, + 0.469, + 0.874 + ], + "angle": 0, + "content": "The new matching score matrix is the linear combination of \\(\\mathcal{P}_{\\mathrm{curv}}\\) and LoFTR matching probability \\(\\mathcal{P}_c\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} = \\lambda \\mathcal {P} _ {c} + (1 - \\lambda) \\mathcal {P} _ {\\text {c u r v}}, \\tag {10}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.09, + 0.882, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.238 + ], + "angle": 0, + "content": "Figure 3. Multi-scale curvature extraction. We extract the curvature similarity from ellipsoids centered on the observed point, fitted with different window (i.e., 3D neighborhood) sizes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.333 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the mixing parameter for balancing the two matching scores. Since our \\(\\mathcal{P}_{\\mathrm{curv}}\\) is an additional supporting score for coarse-level matching selection, we set the parameter \\(\\lambda\\) to 0.9 in all our experiments. We adopt the scheme of LoFTR for the coarse-level match selection as" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.341, + 0.892, + 0.358 + ], + "angle": 0, + "content": "\\[\nM _ {c} = \\{(i, j) \\mid \\forall (i, j) \\in \\operatorname {M N N} (\\mathcal {P}), \\mathcal {P} (i, j) \\geq \\theta \\}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.424 + ], + "angle": 0, + "content": "where MNN is the mutual nearest neighbor operation, \\(\\theta\\) is a threshold so that we only select the matches with confidence higher than \\(\\theta\\). Note that \\(\\mathcal{P}_{\\mathrm{curv}}\\) can be similarly combined with the matching score matrix of SuperGlue [49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.893, + 0.622 + ], + "angle": 0, + "content": "Multi-scale Curvature Similarity. Although the proposed curvature similarity is scaling invariant in theory, it still depends on the neighboring pixels of the points. Thus, the multi-scale extraction of this geometric cue could also be adopted for robustness to scale change in the images. For detector-based approaches, such as SuperGlue [49], we use a set of window size \\(\\{\\delta_1,\\delta_2,\\dots,\\delta_n\\}\\) with the window centered on detected keypoints. Hence, the curvature similarity maps are \\(S(k)_A\\in \\mathbb{R}^{\\mathcal{A}*n}\\) and \\(S(k)_B\\in \\mathbb{R}^{\\mathcal{B}*n}\\), \\(\\mathcal{A}\\) and \\(\\mathcal{B}\\) are the detected keypoints at \\(I_A\\) and \\(I_B\\), respectively. For detector-free matchers, the multi-scale extraction is conducted within the patch grid. Fig. 3 describes the curvature similarity extraction at multiple scales." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.764, + 0.645 + ], + "angle": 0, + "content": "3.4. Depth Prediction Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Since the proposed curvature similarity extraction is fully differentiable, it is possible to use it to train feature matching and depth prediction networks jointly end-to-end. To demonstrate this, we fine-tune the depth predictor with the curvature similarity extractor so that it takes the feature matching into account when being trained. The new confidence matrix \\(\\mathcal{P}\\) is minimized with the negative log-likelihood, similarly as in [49, 55], where the loss is" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.779, + 0.891, + 0.818 + ], + "angle": 0, + "content": "\\[\nL = - \\frac {1}{\\left| M _ {c} ^ {g t} \\right|} \\sum_ {(i, j) \\in M _ {c} ^ {g t}} \\log \\mathcal {P} (i, j), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "and \\( M_c^{gt} \\) is the ground truth matches defined at the coarse-level in [55]. Since the loss term focuses only on the matched patches, long training time would crash the depth prediction model. Thus, we fine-tune the depth predictor with only a few epochs and small learning rates." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.955 + ], + "angle": 0, + "content": "17985" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.251 + ], + "angle": 0, + "content": "
MethodsYFCC100M [59]MegaDepth [27]ScanNet [9]
AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)
SuperGlue [49]38.959.475.798.723.643.762.277.099.632.516.133.851.884.431.5
SuperGlue + CSE40.060.176.395.626.546.264.878.798.335.716.234.052.280.731.7
LoFTR [55]43.463.178.095.28.253.770.182.496.89.022.140.857.687.47.9
LoFTR + CSE43.563.378.193.48.554.470.482.195.59.222.942.460.086.49.3
LoFTR + CSE (w/ FT)43.963.578.193.38.554.370.482.295.59.323.943.560.586.19.5
QuadTree [58]44.163.578.396.58.953.570.282.298.59.624.944.761.689.49.8
QuadTree + CSE45.064.378.895.39.154.070.382.197.99.825.245.562.589.411.2
QuadTree + CSE (w/ FT)45.064.478.995.59.154.571.182.998.09.725.846.163.489.211.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.259, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Table 1. Results of relative pose estimation on YFCC100M, MegaDepth, and ScanNet datasets. We report AUC score of the translation and rotation errors with different thresholds \\((5^{\\circ}, 10^{\\circ}, 20^{\\circ})\\), precision \\((\\%)\\), and matching score \\((\\%)\\). Our methods and the best performance are marked in bold. w/FT indicates the results with fine-tuning. Note that comparison among baselines is not straightforward as some sensitive default parameters, e.g. input image size, are different." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.328, + 0.892, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.53, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Figure 4. Qualitative matching results. We add the curvature similarity extractor module to LoFTR [55] and QuadTree [58] and compare the two approaches in both indoor and outdoor scenes. By adding the CSE module, we are able to obtain more correct matches (green lines) and better pose accuracy while keeping similar matching precision. The matches are coloured by the epipolar error threshold in [55]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.596, + 0.21, + 0.614 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.471, + 0.716 + ], + "angle": 0, + "content": "We first demonstrate the effectiveness of the proposed similarity curvature extractor on relative pose estimation and the downstream visual localization tasks. Next, we analyze the choice of architecture and hyper-parameters in the ablation study. Finally, we provide qualitative results to help understand the similarity curvature extractor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.471, + 0.809 + ], + "angle": 0, + "content": "Implementation Details. Our proposed CSE adopts Mi-DAS + ViT [42] as the default depth predictor. For detector-based approaches [55, 58], we extract the curvature similarity at keypoint locations and use multi-scales \\(\\{7 * 7, 9 * 9, 11 * 11\\}\\). For detector-free matchers, we use the default grid patch size \\(8 * 8\\) for curvature estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For the depth predictor training, following the default setting in LoFTR [55], we fine-tune the predictor with the outdoor matching model on MegaDepth [27]. The model is optimized using Adam with a fixed learning rate of \\( r = 1e - 5 \\) and batch size of 2. The indoor model is fine-tuned on ScanNet [9] with \\( r = 5e - 6 \\). In the training process, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.706 + ], + "angle": 0, + "content": "weights from the matcher are frozen, and only the weights from refineNets in MiDAS + ViT are updated. Instead of sampling 200 pairs from each scene, we reduce it to 10 pairs. The full training and evaluation process of the outdoor model is conducted on 2 Telsa-V100 GPUs with 32GB of memory. We train and evaluate the indoor model on 2 RTX 2080 Ti GPUs with a memory of 12GB." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.712, + 0.727, + 0.727 + ], + "angle": 0, + "content": "4.1. Relative Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Datasets. We evaluate our CSE on three popular datasets, ScanNet [9], MegaDepth [27], and YFCC100M [59], for the two-view pose estimation task. ScanNet is an indoor RGB-D dataset consisting of 1,613 sequences with 2.5 million views, and each view is associated with a ground truth camera pose and depth map. Similar to the setting in LoFTR [49], the images and depth maps are resized to \\(640*480\\) for training and evaluation. MegaDepth and YFCC100M are two outdoor datasets which consist of multiple scenarios. We fine-tune the outdoor model on MegaDepth with the longest dimension of the images re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "17986" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "sized to 840 and evaluate our method on MegaDepth-1500 from [55] and YFCC100M-4000 from [49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.319 + ], + "angle": 0, + "content": "Baselines. The proposed CSE is a plug-in module that can be applied with both detector-based and detector-free matchers. We add this module to detector-based SuperGlue with SuperPoint [12] features, detector-free LoFTR, and its recent variant QuadTree [58]. The experiments follow the default setting from the original papers for all the baselines. For all the experiments on ScanNet, the image resolution is consistently \\(640 \\times 480\\). For evaluation on MegaDepth and YFCC100M, the default input sizes are different among the approaches. Thus, it is not reasonable to compare among the baselines as the input image size significantly impacts the final accuracy. We detail the default input sizes of different approaches in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.471, + 0.501 + ], + "angle": 0, + "content": "Metrics. Similar to the previous methods [49, 55], we report the Area Under the recall Curve (AUC) of the translation and rotation errors with multiple thresholds \\((5^{\\circ}, 10^{\\circ}, 20^{\\circ})\\). For each setting, we consider a pose to be correct if the maximum error from translation and rotation error is below the threshold. The estimated camera poses are recovered from the essential matrix estimated by RANSAC [17] with the inlier-outlier threshold set at 0.5 pixels. Besides, we report the matching precision and score [69, 12]. To calculate matching scores on detector-free matchers, we consider the number of coarse-level features as the number of detected keypoints, i.e., \\(h * w\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Results. We report the results of the CSE add-on on SuperGlue, LoFTR, and QuadTree in Tab. 1. For LoFTR and QuadTree, we report the results of directly adding the CSE module and adding CSE with fine-tuning. As the official training code for SuperGlue is not publicly available, we evaluated the SuperGlue + CSE without fine-tuning. However, we believe that fine-tuning would further improve the results of SuperGlue as well. We also notice that the MegaDepth-1500 is a part of the training set for SuperGlue. However, this does not invalidate the fact that the proposed method helps in finding better features. The proposed CSE improves the pose AUC scores of all methods on almost all datasets. Fine-tuning further increases the accuracy, and QuadTree combined with the proposed CSE method and fine-tuned on depth predictions achieves state-of-the-art AUC scores on all tested datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.47, + 0.82 + ], + "angle": 0, + "content": "As for matching score and precision, we also visualize the matching results of LoFTR and QuadTree in Fig. 4. By adding CSE, we obtain more correct matches and better pose accuracy on surface regions with distinctive geometry, while keeping similar matching precision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.262, + 0.846 + ], + "angle": 0, + "content": "4.2. Visual Localization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets and Metrics. We select the Aachen DAY-NIGHT v1.1 dataset [52] to demonstrate the effectiveness of our proposed approach on the visual localization task." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.178 + ], + "angle": 0, + "content": "
MethodsDayNight
(0.25m, 2°) / (0.5m, 5°) / (5m,10°)
SP [12] + SG [49]89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
SP + SG + CSE90.7 / 96.5 / 99.475.9 / 91.1 / 100.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.187, + 0.894, + 0.23 + ], + "angle": 0, + "content": "Table 2. Visual localization. We report the pose recall at \\((0.25\\mathrm{m} / 2^{\\circ}\\) \\(0.5\\mathrm{m} / 5^{\\circ}\\) \\(5\\mathrm{m} / 10^{\\circ})\\) on the Aachen Day-Night [52] dataset with the HLoc [48] algorithm. The best results are bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Aachen DAY-NIGHT is a challenging urban-scale outdoor dataset which is collected by handheld devices with 6,697 reference images and 1,015 queries, including 191 nighttime images. The ground truth 6DoF camera pose is obtained by COLMAP [53] and refined by [73]. We follow the evaluation protocols of the visual localization benchmark [52] reporting the translation and rotation error recalls at \\(0.25\\mathrm{m} / 2^{\\circ}\\), \\(0.5\\mathrm{m} / 5^{\\circ}\\), and \\(5\\mathrm{m} / 10^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.893, + 0.545 + ], + "angle": 0, + "content": "Results. We add our CSE to the SuperPoint [12] + SuperGlue [49] pipeline in HLoc [48, 49] without fine-tuning any models. The matching pairs are provided by HLoc with top-50 candidate images from NetVLAD [2]. Tab. 2 reports the pose error recalls. We observe that the plug-in CSE improves the SP + SG pipeline on all thresholds on the Day sequences. On the Night sequence, it improves at threshold \\((0.5\\mathrm{m} / 5^{\\circ})\\) and \\((5\\mathrm{m} / 10^{\\circ})\\). We believe the slight decrease at \\((0.25\\mathrm{m} / 1^{\\circ})\\) stems from the inaccuracies of depth prediction on nighttime images. Overall, by adding the CSE module, the localization accuracy is improved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.655, + 0.571 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.671 + ], + "angle": 0, + "content": "We analyse the impact of different design components on our curvature similarity extractor in this section. We evaluate two main design choices: (1) multi-scale curvature extraction, (2) choice of depth predictors, and (3) mixing parameter value \\(\\lambda\\) selection. All experiments are conducted with the Quadtree [58] matcher." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.684, + 0.892, + 0.786 + ], + "angle": 0, + "content": "
Multi-ScaleScanNet [9]MegaDepth [27]Time (ms)
AUC@5°/ AUC@10°/ AUC@20°
4*424.444.661.952.769.682.0113
6*625.345.662.853.369.981.8115
8*8 (default)25.245.562.554.070.382.1112
4*4+6*6+8*825.645.262.454.070.082.1338
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.797, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Table 3. Multi-scale and different grid sizes. The AUC scores and average CSE run-times (ms) of Quadtree [58] + CSE with different grid sizes and on multi-scale estimation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-scale. As described in Sec. 3.3, we conduct multiscale curvature similarity extraction on a \\(8*8\\) patch grid with QuadTree by default. In Tab. 3, we show the results" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17987" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.187 + ], + "angle": 0, + "content": "
Depth PredictorsCSEFTAUC@5@10@20Model
Ground TruthQuadricX25.145.062.0153 MB
MiDAS (Res101)Quadric✓25.545.562.5575 MB
MiDAS (Real-time)Quadric✓25.545.462.9239 MB
MiDAS (ViT)Quadric✓25.846.163.41.6 GB
MiDAS (ViT)Surface FitX24.844.761.61.6 GB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.198, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Table 4. Depth study. The AUC scores of QuadTree [58] with the proposed CSE using the GT depth or MiDaS [42] with different models. Quadric and surface fitting results are also shown." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.468, + 0.384 + ], + "angle": 0, + "content": "when using \\(4 * 4\\), \\(6 * 6\\), and multi-scale grids without fin-tuning. Even though the proposed CSE is robust to the grid size and the accuracy is similar, it slightly increases together with the grid size. Size \\(4 * 4\\) leads to the worst solutions on average, while \\(8 * 8\\) leads to the best ones. As expected, curvature similarity extraction with multi-scale shows the best performance. However, it increases the average run-time by three times." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.468, + 0.642 + ], + "angle": 0, + "content": "Depth Predictors and Curvature Extractors. We then explore the impact of different depth predictors on the final pose accuracy and model size. We select the ground truth depth, MiDAS with ResNet101 [22] architecture and MiDAS small (real-time) to compare with the default predictor MiDAS with ViT [64]. Tab. 4 reports the pose accuracy of these approaches on the ScanNet dataset. MiDAS with ViT, which shows the best performance, has the largest model size due to the heavy depth prediction network. When we employ lighter models for depth prediction, we observe that the performance has only a small drop, indicating that our method is still effective when working with lightweight networks. We also notice that the pose accuracy with ground truth depth is worse than the fine-tuned models. We believe the reason for this is the missing depth regions in the ground truth. We visualize the depth maps in different conditions at supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.687 + ], + "angle": 0, + "content": "The last two rows of the table show the results with the proposed ellipse and with quadratic surface fitting. The ellipse fitting performs more accurately in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.468, + 0.809 + ], + "angle": 0, + "content": "Mixing Parameter Value. We finetune our model with different values for \\(\\lambda\\) in Eq. (10). The results are reported in Tab. 5. The proposed model performs best with \\(\\lambda = 0.9\\) for outdoor and \\(\\lambda = 0.8\\) for indoor scenes. This is expected since the proposed CSE is additional guidance for coarse-level matching that helps reduce the score of incorrect matches stemming from different surfaces. Assigning a large mixing weight could degrade the performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.395, + 0.834 + ], + "angle": 0, + "content": "4.4. Curvature Similarity Understanding" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In Fig. 5, we visualize how the curvature similarity score impacts the key points selection and matching results. We first show in the curvature images that similar surfaces share similar curvature values (more visualizations are presented" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.09, + 0.882, + 0.153 + ], + "angle": 0, + "content": "
λ (AUC@5°)1.00.90.80.70.50.3
ScanNet [9]24.925.826.123.821.017.8
MegaDepth [27]53.554.552.847.045.539.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.162, + 0.892, + 0.204 + ], + "angle": 0, + "content": "Table 5. Mixing Parameter Value Selection. We finetune the Quadtree + SCE with different \\(\\lambda\\) values and report the AUC@ \\(5^{\\circ}\\) on ScanNet and MegaDepth datasets." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.218, + 0.89, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.276, + 0.882, + 0.287 + ], + "angle": 0, + "content": "(a) Points found by QuadTree alone (left) and when combined with CSE (right)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.289, + 0.89, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.349, + 0.885, + 0.361 + ], + "angle": 0, + "content": "(b)Matches found by QuadTree alone (left) and when combined with CSE (right)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.366, + 0.892, + 0.409 + ], + "angle": 0, + "content": "Figure 5. Keypoints (top row) and matches (bottom) found by the QuadTree matcher [58] with and without the proposed CSE, overlayed on the RGB (left) and extracted curvature images (right)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.542 + ], + "angle": 0, + "content": "in the supplementary). Then, we observe that, after adding CSE, more matches are detected from the same surface, as we increase the matching confidence of those matches by assigning a larger curvature similarity score. Besides, the curvature score helps in lowering the matching confidence of matches from different surfaces, e.g., the incorrect (red) matches in Fig. 5 (b) left are eliminated in the right." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.627, + 0.573 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.794 + ], + "angle": 0, + "content": "We present a curvature similarity extractor (CSE), a new algorithm to leverage 3D geometric cues in local feature matching. By fitting quadrics to depth maps obtained from off-the-shelf monocular depth predictors, we extract curvature similarity, which is invariant to translation, rotation, and scaling. CSE can be seamlessly integrated with matchers to guide the match selection. Also, it allows training feature matches and monocular depth networks jointly minimizing the matching loss in an end-to-end manner to further improve the accuracy. Our experiments in Sec. 4 demonstrate the effectiveness of the proposed CSE module, which consistently improves the pose AUC scores of all methods on almost all datasets, achieving state-of-the-art performance on all benchmarks when combined with QuadTree." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations and Future Work. First, the integration of depth prediction and curvature similarity extraction into feature matching leads to increased memory consumption and inference time. An alternative solution for time-sensitive applications is to use a lightweight depth predictor. As shown in Tab. 4, such an approach still increases the accuracy. Second, the discussed invariance properties do" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "17988" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "not hold at depth discontinuities, and limitations may arise in scenes lacking accurate depth or clear surfaces. Even though this did not pose an issue in our experiments, this is a potential future direction that might lead to further improvements." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.168, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the Academy of Finland (grant No. 327911, No. 353138). Daniel Barath is supported by ETH Postdoc fellowship. We acknowledge the computational resources provided by the Aalto Science-IT project and CSC-IT Center for Science, Finland." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.274, + 0.174, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.3, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[1] Pablo Fernández Alcantarilla, Adrien Bartoli, and Andrew J Davison. Kaze features. In Proceedings of European Conference on Computer Vision, pages 214-227, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.415, + 0.471, + 0.47 + ], + "angle": 0, + "content": "[3] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2911-2918, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.473, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[4] Daniel Barath and Jií Matas. Graph-cut ransac. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733-6741, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.517, + 0.47, + 0.584 + ], + "angle": 0, + "content": "[5] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1304-1312, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.588, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[6] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.66, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[7] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.704, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In Proceedings of the European Conference on Computer Vision, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.775, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly sur" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.3, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "face reintegration. ACM Transactions on Graphics, 36(4):1, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[11] Luanyuan Dai, Yizhang Liu, Jiayi Ma, Lifang Wei, Taotao Lai, Changcai Yang, and Riqing Chen. Ms2dg-net: Progressive correspondence learning via multiple sparse semantics dynamic graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8973-8982, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 224-236, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[13] M.P. do Carmo. Differential Geometry of Curves and Surfaces: Revised and Updated Second Edition. Dover Books on Mathematics. Dover Publications, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[14] Siyan Dong, Shuzhe Wang, Yixin Zhuang, Juho Kannala, Marc Pollefeys, and Baoquan Chen. Visual localization via few-shot scene region classification. International Conference on 3D Vision, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[16] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[18] Per-Erik Forssén and David G Lowe. Shape descriptors for maximally stable extremal regions. In International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[19] Khang Truong Giang, Soohwan Song, and Sungho Jo. Curvature-guided dynamic scale networks for multi-view stereo. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[20] Jack Goldfeather and Victoria Interrante. A novel cubic-order algorithm for approximating principal direction vectors. ACM Trans. Graph., 23(1):45-63, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. CiteSeer, 1988. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[23] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "17989" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[24] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[25] Qingde Li and J.G. Griffiths. Least squares ellipsoid specific fitting. In Geometric Modeling and Processing, 2004. Proceedings, pages 335-340, 2004. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[26] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. Advances in Neural Information Processing Systems, 33:17346-17357, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[30] Xiaotao Liu, Chen Meng, Fei-Peng Tian, and Wei Feng. Dgd-net: Local descriptor guided keypoint detection network. In 2021 IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[31] Yuan Liu, Lingjie Liu, Cheng Lin, Zhen Dong, and Wenping Wang. Learnable motion coherence for correspondence pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.472, + 0.65 + ], + "angle": 0, + "content": "[33] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[34] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.472, + 0.748 + ], + "angle": 0, + "content": "[35] Krystian Mokolajczyk and Cordelia Schmid. Scale & affine invariant interest point detectors. International Journal of Computer Vision, 60(1):63-86, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[36] Krystian Mikolajczyk, Tinne Tuytelaars, Cordelia Schmid, Andrew Zisserman, Jiri Matas, Frederik Schaffalitzky, Timor Kadir, and L Van Gool. A comparison of affine region detectors. International Journal of Computer Vision, 65(1):43-72, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[37] Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Repeatability is not enough: Learning affine regions via discriminability. In Proceedings of the European Conference on Computer Vision, pages 284-300, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[39] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[40] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3456-3465, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[41] Rémi Pautrat, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Online invariance selection for local feature descriptors. In European Conference on Computer Vision, pages 707-724. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[42] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[45] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[46] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.647 + ], + "angle": 0, + "content": "[47] John Rugis and Reinhard Klette. A scale invariant surface curvature estimator. In Pacific-Rim Symposium on Image and Video Technology, pages 138-147. Springer, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[48] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[49] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[50] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In European conference on computer vision, pages 752-765. Springer, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "17990" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.187 + ], + "angle": 0, + "content": "[52] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[53] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[54] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[55] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[56] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11286-11295, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[57] Dongli Tan, Jiang-Jiang Liu, Xingyu Chen, Chao Chen, Ruixin Zhang, Yunhang Shen, Shouhong Ding, and Rongrong Ji. Eco-tr: Efficient correspondences finding via coarse-to-fine refinement. Proceedings of the European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[58] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[60] Yurun Tian, Vassileios Balntas, Tony Ng, Axel BarrosoLaguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In Proceedings of the Asian Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[61] Carl Toft, Daniyar Turmukhambetov, Torsten Sattler, Fredrik Kahl, and Gabriel J Brostow. Single-image depth prediction makes feature matching easier. In European Conference on Computer Vision, pages 473-492. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[62] Lijing Tong and Xiang Ying. 3d point cloud initial registration using surface curvature and surf matching. 3D Research, 9:1-16, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[63] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[65] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[66] Shuzhe Wang, Zakaria Laskar, Iaroslav Melekhov, Xiaotian Li, and Juho Kannala. Continual learning for image-based camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3252-3262, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[67] Changchang Wu. Towards linear-time incremental structure from motion. In International Conference on 3D Vision, pages 127-134. IEEE, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[68] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[69] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[70] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[71] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[72] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Hongen Liao. Learning two-view correspondences and geometry using order-aware network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5845-5854, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[73] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[74] Chen Zhao, Yixiao Ge, Feng Zhu, Rui Zhao, Hongsheng Li, and Mathieu Salzmann. Progressive correspondence pruning by consensus learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6464-6473, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.955 + ], + "angle": 0, + "content": "17991" + } + ] +] \ No newline at end of file diff --git a/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_origin.pdf b/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4538f94b1905fc53fbb2b256c207ad88aee5e57f --- /dev/null +++ b/2023/Guiding Local Feature Matching with Surface Curvature/e3a43d92-ec36-432b-94e9-c4f32da40a2e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96e1d152f0c23bd9f91f989968b0a5a98ae7f14b896300f6f94d00f00e9bec7a +size 2084819 diff --git a/2023/Guiding Local Feature Matching with Surface Curvature/full.md b/2023/Guiding Local Feature Matching with Surface Curvature/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d9996824587e43acccc6a7787d4d584f1e275f99 --- /dev/null +++ b/2023/Guiding Local Feature Matching with Surface Curvature/full.md @@ -0,0 +1,392 @@ +# Guiding Local Feature Matching with Surface Curvature + +Shuzhe Wang $^{1*}$ + +Juho kannala + +Marc Pollefeys $^{2,3}$ + +Daniel Barath2 + +$^{1}$ Aalto University + +$^{2}$ ETH zurich + +3Microsoft + +# Abstract + +We propose a new method, called curvature similarity extractor (CSE), for improving local feature matching across images. CSE calculates the curvature of the local 3D surface patch for each detected feature point in a viewpoint-invariant manner via fitting quadrics to predicted monocular depth maps. This curvature is then leveraged as an additional signal in feature matching with off-the-shelf matchers like SuperGlue and LoFTR. Additionally, CSE enables end-to-end joint training by connecting the matcher and depth predictor networks. Our experiments demonstrate on large-scale real-world datasets that CSE consistently improves the accuracy of state-of-the-art methods. Fine-tuning the depth prediction network further enhances the accuracy. The proposed approach achieves state-of-the-art results on the ScanNet dataset, showcasing the effectiveness of incorporating 3D geometric information into feature matching. + +# 1. Introduction + +Local feature matching is a crucial component for many geometric computer vision tasks, including visual localization [50, 51, 52, 48, 14, 66], structure-from-motion (SfM) [67, 53, 29], and simultaneous localization and mapping (SLAM) [38, 39, 10]. Given a pair of images observing a 3D scene, the task is to find reliable tentative point-to-point correspondences in the two images. Forming such feature matches is often a challenging task as the images may undergo large viewpoint and illumination changes, have occlusions or repetitive patterns. + +The standard pipeline for local feature matching typically involves two steps: (1) keypoint detection and description and (2) point-wise feature matching. Traditional approaches mainly focus on improving the robustness of + +![](images/c181ac8fd349f5b1758fbde035aacee3f778dd4a3e8c2db5c48d649a3a09ef39.jpg) +Figure 1. Curvature-guided match selection: (a) Correct matches in two views must share the same underlying surface curvature, e.g., predicted from monocular depth. (b) Correspondences from different surfaces are guaranteed to be incorrect. + +![](images/374c70d2a8eba1f187d384f647deb1c62e8cd64440dfb6c6f094578c8d30b59d.jpg) + +keypoint detection and description through, for example, extending the Harris [21] detector to handle affine transformations and multiple scales [35, 36] or using more discriminative or efficient descriptors [32, 7, 34, 1, 3]. However, despite their unbroken popularity, these algorithms often fail to cope with the challenges that arise in real-world environments, leading to low accuracy. + +Recent advances in deep learning-based feature matching have made significant progress in addressing the limitations of hand-crafted approaches, such as by jointly training detectors and descriptors [69, 12, 15, 44, 63, 40] with convolutional neural networks (CNNs), or combining hand-crafted and learning-based descriptors in a hybrid manner [41, 6]. SuperGlue [49] introduced the use of transformer networks to learn the matching process and formulate the problem as an optimal transport task [64]. LoFTR [55] and its recent variants [58, 8, 65] leverage both the global and local context from raw images by jointly learning the feature extraction and matching in a single network. While these approaches have led to state-of-the-art performance on several benchmarks, they work entirely in the 2D image domain and ignore the underlying 3D geometry of the scene. As feature matching is, essentially, finding the corresponding projections of an actual 3D point, this could be a critical limitation in real scenarios where the + +matched pixels from different views must share the geometry of the underlying surface patch, as shown in Fig. 1. + +In this paper, we focus on investigating and exploiting 3D geometry cues, e.g. coming from monocular depth predictions, for feature matching. The only prior work is [61], which adopts depth priors to find planar image regions that are then rectified to eliminate viewpoint changes. Finally, a handcrafted detector and descriptor are applied to the rectified image regions. This, however, relies on the heavy assumption that the images consist mostly of dominant planes, which severely restricts its out-of-the-box applicability. In this work, we go further by exploring the local surface geometry coming from depth priors. This additional geometric clue is coupled with any off-the-shelf feature matcher increasing its accuracy. Benefiting from advanced monocular depth predictors [42, 43, 16], obtaining depth priors is easy and costs only a few milliseconds. We use these depth predictions to extract curvature at the observed feature points. Exploiting the fact that the curvature is invariant to viewpoint changes (i.e., rotation, translation, and scaling), we enforce that matched features must lie on similar surfaces. Our contributions are as follows: + +1. We propose an approach for feature matching, leveraging dense depth via utilizing local surface curvature similarity, which is invariant to viewpoint changes. This approach is a departure from traditional matching methods that rely solely on image information. +2. The proposed curvature similarity extractor can be seamlessly integrated with any recent feature matcher, making it a versatile tool that can be easily adopted. The experiments show that it improves several recent algorithms on various benchmarks by $1 - 3\%$ . +3. The proposed algorithm can be used to train feature matchers and monocular depth prediction networks jointly in an end-to-end fashion. To demonstrate this, we fine-tune the state-of-the-art MiDaS [42] depth predictor to increase the feature matching accuracy. + +# 2. Related work + +# 2.1. Local Feature Matching + +Detector-based Matchers build sparse correspondences on top of the detected keypoints. Therefore, robust keypoint detection and feature extraction are essential for these approaches. Traditional hand-crafted descriptors, such as SIFT [32] and SURF [7], follow a detect-then-describe pipeline and have shown great success since the 2000s. More recently, learned detectors and descriptors [6, 69, 12, 15, 44, 63, 60, 30] with convolutional neural networks show their superiority in matching images with large viewpoint and appearance changes. This traditional pipeline is further + +developed to detect-and-describe [12, 15, 44] and describe-to-detect [60, 30] strategies leveraging high-level image information for accurate and reliable matching in challenging conditions. Besides substituting the traditional detectors and descriptors with their learned counterparts for image matching, the recent SuperGlue [49] focuses on the matching stage itself. It replaces the naive mutual nearest neighbour search with Graph Neural Networks (GNN). SuperGlue takes the sparse descriptors and their positional encoding as inputs, then leverage a transformer-based [64] network to create a more robust feature representation for the optimal partial assignment. The follow-up work [54] further improves the performance and efficiency by adaptively clustering sparse features into different subgraphs and using a coarse-to-fine paradigm. Although showing significant improvements with the recent components, the detector-based matchers are naturally limited by the accuracy of detected keypoints coming from independent detectors. + +Detector-free Matchers directly estimate the dense correspondences from raw images without an independent keypoint detection stage. These approaches leverage low and high-level image information with neural networks and can distinguish indistinctive regions across images. The pioneer work of [46] constructs 4D cost volumes, significantly increasing the matching accuracy. However, it is computationally expensive as the complexity is $\mathcal{O}(n^2)$ , where $n$ is the number of pixels or patches. [45, 26] alleviate the problem by applying either sparse convolutions or a coarse-to-fine paradigm. More recently, transformer-based detector-free matchers [55, 58, 65, 8, 24, 57] have received attention due to their strong performance on local feature matching. The most representative work, LoFTR [55], inherits the advantages of graph matching from SuperGlue and leverages a linear transformer [64] for efficient dense matching. [58, 8] utilise more efficient transformer structures to improve the performance. Detector-free matchers show a promising direction towards local feature matching by encoding rich image information. We claim that extracting 3D geometric information could further enhance accuracy. + +Correspondence Pruning approaches [17, 4, 5, 72, 70, 56] apply a consensus mechanism to filter out the outliers from putative correspondences coming from the feature matcher procedure. RANSAC [17] and its follow-up works [4, 5] are the most popular correspondence pruning algorithms. In the era of deep learning, OANet [72] infers the probabilities of correspondences being inliers with an order-aware network. Other works improve the accuracy of correspondence pruning by applying motion coherence constraints [31, 33], leveraging local-to-global consensus learning procedure [74, 11], and adopting the attention mechanism [56]. These approaches provide inlier probability predictions for each putative correspondence that can be used either to filter or order matches. Finally, least squares fitting + +or traditional robust estimation is applied to the matches. + +# 2.2. 3D Geometric Priors to Image Matching + +Incorporating 3D geometric priors into various vision tasks has been widely explored in recent years. [23] leverages the RGB-D reconstructions to learn view-invariant, geometry-aware 2D representations for downstream tasks. [71] introduces the predicted monocular depth to guide the optimization of neural scene representation. The surface curvature, a popular 3D geometric cue, is widely applied to 3D vision tasks such as 3D point cloud registration [62], multi-view stereo [68, 19]. However, integrating 3D geometric cues into 2D image matching has received little attention. [37, 18] assume that the underlying surface patch is locally planar for feature detection and description, while [61] uses prior depth information to improve feature extraction by rectifying large planar regions. None of the above approaches considers the 3D geometric information in the local region, and we are the first to include 3D surface geometry in general 2D feature matching. + +# 3. Feature Matching with Surface Curvature + +This section proposes an approach to improve any local feature matcher by extracting information about the underlying local 3D surface predicted by a deep network. The overview of our method coupled with LoFTR [55] is shown in Fig. 2. Although the proposed algorithm is compatible with any matcher, we demonstrate its effectiveness with LoFTR. Combining it with other ones is straightforward. + +# 3.1. LoFTR-style Matchers + +This section provides a brief overview of LoFTR. Given an image pair $I_A, I_B \in \mathbb{R}^{H*W*3}$ , the end-to-end trainable matcher, first, extracts coarse-level features $\mathcal{F}_A^c, \mathcal{F}_B^c$ and, then, fine-level ones $\mathcal{F}_A^f, \mathcal{F}_B^f$ with a local feature CNN [28]. At the coarse level transformer module, the features are flattened and processed by a transformer-based architecture with multiple self- and cross-attention layers [64]. A matching confidence map $\mathcal{P}_c$ is predicted at the end of this module, and the matches $\mathcal{M}_c$ with high confidence are selected from $\mathcal{P}_c$ to fine-tune the prediction of fine-level features matching at the coarse-to-fine transformer module. The final outputs are the fine-level matches $\mathcal{M}_f$ . + +# 3.2. Curvature Similarity Extractor + +In this section, we will discuss multiple potential ways of extracting local surface curvature at detected features. This curvature is then used for improving feature matching. + +# 3.2.1 Curvature Similarity + +For an at least two times continuously differentiable surface $S \in \mathbb{C}^2$ in the 3D space, given a point $p \in \mathbb{R}^3$ and + +a direction $d \in \mathbb{R}^3$ on $S$ ; the normal curvature $k_n \in \mathbb{R}$ measures how curved the surface is at this point along the direction [13]. Since there are infinitely many directions that travel through a point on the surface, there are also infinitely many normal curvatures at a given point. In differential geometry, the minimum normal curvature $k_1$ and the maximum curvature $k_2$ at point $p$ are defined as the principal curvatures, where $k_1 \leq k_2$ , and the directions of the principal curvatures are the smoothest and steepest directions on the surface. We also define the mean curvature $H$ and Gaussian curvature $K$ as + +$$ +H = \frac {1}{2} \left(k _ {1} + k _ {2}\right), K = k _ {1} k _ {2} \tag {1} +$$ + +A highly beneficial property of such curvatures is their invariance to surface rotation and translation. Both are important when describing local surface patches centered on features detected in 2D images. However, these measures do not preserve scale invariance [13], which is essential in our case due to using predicted relative depth maps instead of a metric one. This means that, in each image, the depth is only defined up-to-scale. + +To seek the scale invariance of the curvatures for local patch matching, we follow [47] for the idea of scale-invariant curvature measure and define the curvature similarity function $S: \mathbb{R} \times \mathbb{R} \to \mathbb{R}$ as follows: + +$$ +S \left(k _ {1}, k _ {2}\right) = \frac {\operatorname* {m i n} \left(\left| k _ {1} \right| , \left| k _ {2} \right|\right)}{\operatorname* {m a x} \left(\left| k _ {1} \right| , \left| k _ {2} \right|\right)}, (0 \leq S (k) \leq 1), \tag {2} +$$ + +where $k_{1}$ and $k_{2}$ are the minimum and maximum curvatures, respectively. It is easy to see that $S(k_{1},k_{2})$ is rotation, translation and scale invariant. Case $S(k_{1},k_{2}) = 1$ is interpreted as the curvature of a point on a spherical surface. In the case when $k_{1} = k_{2} = 0$ , we define $S(k_{1},k_{2})$ to be 0. Detecting points on planes is also straightforward from $k_{1}$ and $k_{2}$ as they both will become $\infty$ . Proof of the invariances is provided in the supplementary material. Next, we will discuss ways of obtaining curvature similarities. + +# 3.2.2 Surface Curvatures Extraction + +This section introduces our approximate form of principal curvature estimation. Different from prior work [68, 19] that extract the curvatures from image intensity, we estimate the principal curvatures based on the depth map. This 3D geometric cue could be obtained either from the active sensors (LiDAR, Kinect) or monocular depth predictors [42, 43]. Since our target curvature similarity $S(k_{1}, k_{2})$ is scaling invariant, the scaling ambiguity in monocular depth estimation is neglected and the predicted relative depth is sufficient for the curvature extraction. + +In this work, we consider only a pair of RGB images as input and leverage any off-the-shelf depth predictor, e.g. + +![](images/d0dd7ea3279a63ac0d14048d072e0d10beec8d485d747ea8a11b7992c9e121a5.jpg) +Figure 2. Overview of the CSE with LoFTR. For a pair of images, it follows the LoFTR [55] matching pipeline to extract the coarse and fine level image features and generate a coarse matching confidence map $\mathcal{P}_c$ . In the meantime, the images are fed to a monocular depth network to predict the depth map and extract the curvature similarity upon the depth map. We formulate a new curvature score matrix $\mathcal{P}_{curv}$ based on the pixel-wise $L2$ distance of the curvature similarities between two images and mix it with $\mathcal{P}_c$ . The mixed score matrix is then processed by the coarse-to-fine transformer module together with the fine-level features to predict the final matches $\mathcal{M}_f$ . + +MiDaS [42], to extract the depth map $Z_{i}:\mathbb{R}\times \mathbb{R}\to \mathbb{R}$ for each input image, $i\in \{1,2\}$ . Let $\mathbf{r}_i(u,v,Z_i(u,v))$ be the 3D surface patch corresponding to the image region at coordinates $(u,v)$ with window size $\delta \in \mathbb{R}^{+}$ in the $i$ th image, where $Z_{i}(u,v)$ is the depth of point $(u,v)$ . + +Let us formulate a local point set $\mathcal{P}_i = \{(u,v,Z_i(u,v))\mid (u,v)\in \mathcal{M}_f^i\} \subseteq \mathbb{R}^3$ , where $\mathcal{M}_f^i$ are the points detected in the $i$ th image. + +Principal Curvature via Surface Fitting. The most straightforward approach is first applying the quadratic surface approximation method [20] to estimate the surface $\mathbf{r}_i(u,v,Z_i(u,v))$ at all points in $\mathcal{P}$ , then employing the first and second fundamental forms of a surface. The normal curvature is calculated as follows: + +$$ +k _ {n} = \frac {I I}{I} = \frac {L d u ^ {2} + 2 M d u d v + N d v ^ {2}}{E d u ^ {2} + 2 F d u d v + G d v ^ {2}}, \tag {3} +$$ + +where $L = \mathbf{r}_{uu}\mathbf{n}$ , $M = \mathbf{r}_{uv}\mathbf{n}$ , $N = \mathbf{r}_{vv}\mathbf{n}$ , $E = \mathbf{r}_u\mathbf{r}_u$ , $F = \mathbf{r}_u\mathbf{r}_v$ , $G = \mathbf{r}_v\mathbf{r}_v$ ; $\mathbf{r}_u, \mathbf{r}_v, \mathbf{r}_{uu}, \mathbf{r}_{uv}, \mathbf{r}_{vv}$ are the first and second-order derivatives of the surface patch $\mathbf{r}_i(u,v,Z_i(u,v))$ , and $\mathbf{n} \in \mathbb{R}^3$ is the normal. + +$$ +k _ {n} = \frac {L + 2 M \lambda + N \lambda^ {2}}{E + 2 F \lambda + G \lambda^ {2}}. \tag {4} +$$ + +Once we have the normal curvature $k_{n}$ , we then calculate the partial derivative of $k_{n}$ w.r.t. $\lambda$ to obtain the principal curvatures $k_{1}, k_{2}$ . The details are presented in the supplementary. While this approach works accurately and efficiently in practice, we observed that its gradients become unstable during back-propagation. Therefore, we explore another way as well for calculating the surface curvature. + +**Curvature Extraction via Quadrics.** Another lightweight procedure for estimating the curvature is via fitting quadric surfaces, e.g. ellipsoid, locally to the neighborhood of the observed point. Thus, instead of directly estimating a quadratic surface from the given depth map, we constrain the surface to be a quadric and, more specifically, an ellipsoid. An ellipsoid has the beneficial property that all its curvatures are semi-positive. Its principal axes define three directions which we can use the define the similarity. We use the radii $\alpha, \beta, \gamma \in \mathbb{R}^{+}$ along the three axes and define the curvature similarity score as follows: + +$$ +S (\alpha , \beta , \gamma) = \frac {\operatorname* {m i n} \left(\alpha^ {2} , \beta^ {2} , \gamma^ {2}\right)}{\operatorname* {m a x} \left(\alpha^ {2} , \beta^ {2} , \gamma^ {2}\right)}. \tag {5} +$$ + +Similarity $S$ is rotation, translation, scale and thus, viewpoint invariant. The advantage from Eq. (5) is that the computation of first and second-order gradients are no longer required for the curvature similarity calculation. + +The general constraint that a quadric imposes on a point $x \in \mathcal{P}_i$ lying on the surface is + +$$ +\mathbf {x} \mathbf {Q} _ {i} \mathbf {x} ^ {\mathrm {T}} + \mathbf {P} _ {i} \mathbf {x} ^ {\mathrm {T}} + \mathbf {R} _ {i} = 0, +$$ + +where $\mathbf{Q}_i$ , $\mathbf{P}_i$ , and $\mathbf{R}_i$ are the parameters in a matrix form. This can be written as + +$$ +\begin{array}{l} a x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \tag {6} \\ 2 f y z + 2 g x + 2 h y + 2 i z + j = 0. \\ \end{array} +$$ + +As described in [25], a sufficient condition that guarantees the quadric surface to be an ellipsoid is + +$$ +4 M - N ^ {2} = 1, +$$ + +where + +$$ +N = a + b + c, +$$ + +$$ +M = a b + b c + a c - d ^ {2} - e ^ {2} - f ^ {2}. +$$ + +Also, since we aim at measuring the curvature of the surface patch located on the observed point, we have to constrain the ellipsoid so that the observed point $\mathbf{p}$ is on its surface. To do so, we translate the point set $\mathcal{P}$ as $\mathcal{P}' = \{\mathbf{q} - \mathbf{p} \mid \mathbf{q} \in \mathcal{P}\}$ and fix $j = 0$ . The central point now is $\mathbf{p}' = [0,0,0]^{\mathrm{T}}$ . + +Substituting all these constraints into Eq. (6), we obtain + +$$ +a x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \tag {7} +$$ + +$$ +2 f y z + 2 g x + 2 h y + 2 i z = 0. +$$ + +From Eq. (7), only nine variables need to be estimated in our case. For each point $[u, v, z]^{\mathrm{T}} \in \mathcal{P}^{\prime}$ , we define $\mathbf{X}_i$ as + +$$ +\mathbf {X} _ {i} = \left[ u _ {i} ^ {2}, v _ {i} ^ {2}, z _ {i} ^ {2}, 2 v _ {i} z _ {i}, 2 u _ {i} z _ {i}, 2 u _ {i} v _ {i}, 2 u _ {i}, 2 v _ {i}, 2 z _ {i} \right] ^ {\mathrm {T}}, \tag {8} +$$ + +and coefficient matrix $\mathbf{C} = [\mathbf{X}_1, \mathbf{X}_2, \mathbf{X}_3, \dots, \mathbf{X}_n]$ . Then, we follow the method in [25] for the ellipsoid fitting and estimate its radii $R = (\alpha, \beta, \gamma)$ . The algorithms for fitting and calculating the radii are shown in the supplementary. + +# 3.3. Curvature-Guided Match Selection + +The proposed curvature similarity extractor (CSE) is a depth-based plug-in component that can be combined with both detector-based [32, 49] and detector-free [55, 58, 8] matchers. For an image pair $(I_A, I_B)$ , the monocular depth predictor first estimates the depth map for each image independently. CSE extracts the curvature map $(S_A, S_B)$ on the top of the depth maps. Note that LoFTR selects the candidate matches $M_c$ at the coarse level of 1/8-resolution grids where the ground truth matches are defined based on the reprojection distance of the centers of the grid cells. Thus, it is not necessary to extract the curvature similarity at each pixel. It is calculated only at the center point of each cell, i.e., $S_A, S_B \in \mathbb{R}^{(h * w)*1}$ , where $h = H / 8$ , $w = W / 8$ . + +Since the proposed curvature similarity is rotation, translation and scale invariant, $S_{A}$ at point $\mathbf{p}_A$ in the first image should be equal to the similarity $S_B$ of its corresponding pair $\mathbf{p}_B$ in the second image. The correct matches $M_c$ , ideally, have the same curvature values. To measure this, we define the normalised curvature score matrix as + +$$ +\mathcal {P} _ {\text {c u r v}} = 1 - L _ {2} (S (k) _ {A}, S (k) _ {B}), \tag {9} +$$ + +where $L_{2}(\cdot ,\cdot)$ is the L2-norm of $S(k)_A$ and $S(k)_B$ . Since both $S(k)_A$ and $S(k)_B$ are in-between zero and one, the curvature score is also normalized into range [0, 1]. + +The new matching score matrix is the linear combination of $\mathcal{P}_{\mathrm{curv}}$ and LoFTR matching probability $\mathcal{P}_c$ as follows: + +$$ +\mathcal {P} = \lambda \mathcal {P} _ {c} + (1 - \lambda) \mathcal {P} _ {\text {c u r v}}, \tag {10} +$$ + +![](images/47e53b10f054a133c4167ac138e30950c20b6933afa6ee0f90c7b7be82554861.jpg) +Figure 3. Multi-scale curvature extraction. We extract the curvature similarity from ellipsoids centered on the observed point, fitted with different window (i.e., 3D neighborhood) sizes. + +where $\lambda$ is the mixing parameter for balancing the two matching scores. Since our $\mathcal{P}_{\mathrm{curv}}$ is an additional supporting score for coarse-level matching selection, we set the parameter $\lambda$ to 0.9 in all our experiments. We adopt the scheme of LoFTR for the coarse-level match selection as + +$$ +M _ {c} = \{(i, j) \mid \forall (i, j) \in \operatorname {M N N} (\mathcal {P}), \mathcal {P} (i, j) \geq \theta \}, \tag {11} +$$ + +where MNN is the mutual nearest neighbor operation, $\theta$ is a threshold so that we only select the matches with confidence higher than $\theta$ . Note that $\mathcal{P}_{\mathrm{curv}}$ can be similarly combined with the matching score matrix of SuperGlue [49]. + +Multi-scale Curvature Similarity. Although the proposed curvature similarity is scaling invariant in theory, it still depends on the neighboring pixels of the points. Thus, the multi-scale extraction of this geometric cue could also be adopted for robustness to scale change in the images. For detector-based approaches, such as SuperGlue [49], we use a set of window size $\{\delta_1,\delta_2,\dots,\delta_n\}$ with the window centered on detected keypoints. Hence, the curvature similarity maps are $S(k)_A\in \mathbb{R}^{\mathcal{A}*n}$ and $S(k)_B\in \mathbb{R}^{\mathcal{B}*n}$ , $\mathcal{A}$ and $\mathcal{B}$ are the detected keypoints at $I_A$ and $I_B$ , respectively. For detector-free matchers, the multi-scale extraction is conducted within the patch grid. Fig. 3 describes the curvature similarity extraction at multiple scales. + +# 3.4. Depth Prediction Fine-tuning + +Since the proposed curvature similarity extraction is fully differentiable, it is possible to use it to train feature matching and depth prediction networks jointly end-to-end. To demonstrate this, we fine-tune the depth predictor with the curvature similarity extractor so that it takes the feature matching into account when being trained. The new confidence matrix $\mathcal{P}$ is minimized with the negative log-likelihood, similarly as in [49, 55], where the loss is + +$$ +L = - \frac {1}{\left| M _ {c} ^ {g t} \right|} \sum_ {(i, j) \in M _ {c} ^ {g t}} \log \mathcal {P} (i, j), \tag {12} +$$ + +and $M_c^{gt}$ is the ground truth matches defined at the coarse-level in [55]. Since the loss term focuses only on the matched patches, long training time would crash the depth prediction model. Thus, we fine-tune the depth predictor with only a few epochs and small learning rates. + +
MethodsYFCC100M [59]MegaDepth [27]ScanNet [9]
AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)
SuperGlue [49]38.959.475.798.723.643.762.277.099.632.516.133.851.884.431.5
SuperGlue + CSE40.060.176.395.626.546.264.878.798.335.716.234.052.280.731.7
LoFTR [55]43.463.178.095.28.253.770.182.496.89.022.140.857.687.47.9
LoFTR + CSE43.563.378.193.48.554.470.482.195.59.222.942.460.086.49.3
LoFTR + CSE (w/ FT)43.963.578.193.38.554.370.482.295.59.323.943.560.586.19.5
QuadTree [58]44.163.578.396.58.953.570.282.298.59.624.944.761.689.49.8
QuadTree + CSE45.064.378.895.39.154.070.382.197.99.825.245.562.589.411.2
QuadTree + CSE (w/ FT)45.064.478.995.59.154.571.182.998.09.725.846.163.489.211.5
+ +Table 1. Results of relative pose estimation on YFCC100M, MegaDepth, and ScanNet datasets. We report AUC score of the translation and rotation errors with different thresholds $(5^{\circ}, 10^{\circ}, 20^{\circ})$ , precision $(\%)$ , and matching score $(\%)$ . Our methods and the best performance are marked in bold. w/FT indicates the results with fine-tuning. Note that comparison among baselines is not straightforward as some sensitive default parameters, e.g. input image size, are different. + +![](images/551650444ceed7e5e1679b6f9abcde28daab7aad78256588a1848430af233c86.jpg) +Figure 4. Qualitative matching results. We add the curvature similarity extractor module to LoFTR [55] and QuadTree [58] and compare the two approaches in both indoor and outdoor scenes. By adding the CSE module, we are able to obtain more correct matches (green lines) and better pose accuracy while keeping similar matching precision. The matches are coloured by the epipolar error threshold in [55]. + +# 4. Experiments + +We first demonstrate the effectiveness of the proposed similarity curvature extractor on relative pose estimation and the downstream visual localization tasks. Next, we analyze the choice of architecture and hyper-parameters in the ablation study. Finally, we provide qualitative results to help understand the similarity curvature extractor. + +Implementation Details. Our proposed CSE adopts Mi-DAS + ViT [42] as the default depth predictor. For detector-based approaches [55, 58], we extract the curvature similarity at keypoint locations and use multi-scales $\{7 * 7, 9 * 9, 11 * 11\}$ . For detector-free matchers, we use the default grid patch size $8 * 8$ for curvature estimation. + +For the depth predictor training, following the default setting in LoFTR [55], we fine-tune the predictor with the outdoor matching model on MegaDepth [27]. The model is optimized using Adam with a fixed learning rate of $r = 1e - 5$ and batch size of 2. The indoor model is fine-tuned on ScanNet [9] with $r = 5e - 6$ . In the training process, the + +weights from the matcher are frozen, and only the weights from refineNets in MiDAS + ViT are updated. Instead of sampling 200 pairs from each scene, we reduce it to 10 pairs. The full training and evaluation process of the outdoor model is conducted on 2 Telsa-V100 GPUs with 32GB of memory. We train and evaluate the indoor model on 2 RTX 2080 Ti GPUs with a memory of 12GB. + +# 4.1. Relative Pose Estimation + +Datasets. We evaluate our CSE on three popular datasets, ScanNet [9], MegaDepth [27], and YFCC100M [59], for the two-view pose estimation task. ScanNet is an indoor RGB-D dataset consisting of 1,613 sequences with 2.5 million views, and each view is associated with a ground truth camera pose and depth map. Similar to the setting in LoFTR [49], the images and depth maps are resized to $640*480$ for training and evaluation. MegaDepth and YFCC100M are two outdoor datasets which consist of multiple scenarios. We fine-tune the outdoor model on MegaDepth with the longest dimension of the images re + +sized to 840 and evaluate our method on MegaDepth-1500 from [55] and YFCC100M-4000 from [49]. + +Baselines. The proposed CSE is a plug-in module that can be applied with both detector-based and detector-free matchers. We add this module to detector-based SuperGlue with SuperPoint [12] features, detector-free LoFTR, and its recent variant QuadTree [58]. The experiments follow the default setting from the original papers for all the baselines. For all the experiments on ScanNet, the image resolution is consistently $640 \times 480$ . For evaluation on MegaDepth and YFCC100M, the default input sizes are different among the approaches. Thus, it is not reasonable to compare among the baselines as the input image size significantly impacts the final accuracy. We detail the default input sizes of different approaches in the supplementary material. + +Metrics. Similar to the previous methods [49, 55], we report the Area Under the recall Curve (AUC) of the translation and rotation errors with multiple thresholds $(5^{\circ}, 10^{\circ}, 20^{\circ})$ . For each setting, we consider a pose to be correct if the maximum error from translation and rotation error is below the threshold. The estimated camera poses are recovered from the essential matrix estimated by RANSAC [17] with the inlier-outlier threshold set at 0.5 pixels. Besides, we report the matching precision and score [69, 12]. To calculate matching scores on detector-free matchers, we consider the number of coarse-level features as the number of detected keypoints, i.e., $h * w$ . + +Results. We report the results of the CSE add-on on SuperGlue, LoFTR, and QuadTree in Tab. 1. For LoFTR and QuadTree, we report the results of directly adding the CSE module and adding CSE with fine-tuning. As the official training code for SuperGlue is not publicly available, we evaluated the SuperGlue + CSE without fine-tuning. However, we believe that fine-tuning would further improve the results of SuperGlue as well. We also notice that the MegaDepth-1500 is a part of the training set for SuperGlue. However, this does not invalidate the fact that the proposed method helps in finding better features. The proposed CSE improves the pose AUC scores of all methods on almost all datasets. Fine-tuning further increases the accuracy, and QuadTree combined with the proposed CSE method and fine-tuned on depth predictions achieves state-of-the-art AUC scores on all tested datasets. + +As for matching score and precision, we also visualize the matching results of LoFTR and QuadTree in Fig. 4. By adding CSE, we obtain more correct matches and better pose accuracy on surface regions with distinctive geometry, while keeping similar matching precision. + +# 4.2. Visual Localization + +Datasets and Metrics. We select the Aachen DAY-NIGHT v1.1 dataset [52] to demonstrate the effectiveness of our proposed approach on the visual localization task. + +
MethodsDayNight
(0.25m, 2°) / (0.5m, 5°) / (5m,10°)
SP [12] + SG [49]89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
SP + SG + CSE90.7 / 96.5 / 99.475.9 / 91.1 / 100.0
+ +Aachen DAY-NIGHT is a challenging urban-scale outdoor dataset which is collected by handheld devices with 6,697 reference images and 1,015 queries, including 191 nighttime images. The ground truth 6DoF camera pose is obtained by COLMAP [53] and refined by [73]. We follow the evaluation protocols of the visual localization benchmark [52] reporting the translation and rotation error recalls at $0.25\mathrm{m} / 2^{\circ}$ , $0.5\mathrm{m} / 5^{\circ}$ , and $5\mathrm{m} / 10^{\circ}$ . + +Results. We add our CSE to the SuperPoint [12] + SuperGlue [49] pipeline in HLoc [48, 49] without fine-tuning any models. The matching pairs are provided by HLoc with top-50 candidate images from NetVLAD [2]. Tab. 2 reports the pose error recalls. We observe that the plug-in CSE improves the SP + SG pipeline on all thresholds on the Day sequences. On the Night sequence, it improves at threshold $(0.5\mathrm{m} / 5^{\circ})$ and $(5\mathrm{m} / 10^{\circ})$ . We believe the slight decrease at $(0.25\mathrm{m} / 1^{\circ})$ stems from the inaccuracies of depth prediction on nighttime images. Overall, by adding the CSE module, the localization accuracy is improved. + +# 4.3. Ablation Study + +We analyse the impact of different design components on our curvature similarity extractor in this section. We evaluate two main design choices: (1) multi-scale curvature extraction, (2) choice of depth predictors, and (3) mixing parameter value $\lambda$ selection. All experiments are conducted with the Quadtree [58] matcher. + +Table 2. Visual localization. We report the pose recall at $(0.25\mathrm{m} / 2^{\circ}$ $0.5\mathrm{m} / 5^{\circ}$ $5\mathrm{m} / 10^{\circ})$ on the Aachen Day-Night [52] dataset with the HLoc [48] algorithm. The best results are bold. + +
Multi-ScaleScanNet [9]MegaDepth [27]Time (ms)
AUC@5°/ AUC@10°/ AUC@20°
4*424.444.661.952.769.682.0113
6*625.345.662.853.369.981.8115
8*8 (default)25.245.562.554.070.382.1112
4*4+6*6+8*825.645.262.454.070.082.1338
+ +Table 3. Multi-scale and different grid sizes. The AUC scores and average CSE run-times (ms) of Quadtree [58] + CSE with different grid sizes and on multi-scale estimation. + +Multi-scale. As described in Sec. 3.3, we conduct multiscale curvature similarity extraction on a $8*8$ patch grid with QuadTree by default. In Tab. 3, we show the results + +
Depth PredictorsCSEFTAUC@5@10@20Model
Ground TruthQuadricX25.145.062.0153 MB
MiDAS (Res101)Quadric✓25.545.562.5575 MB
MiDAS (Real-time)Quadric✓25.545.462.9239 MB
MiDAS (ViT)Quadric✓25.846.163.41.6 GB
MiDAS (ViT)Surface FitX24.844.761.61.6 GB
+ +when using $4 * 4$ , $6 * 6$ , and multi-scale grids without fin-tuning. Even though the proposed CSE is robust to the grid size and the accuracy is similar, it slightly increases together with the grid size. Size $4 * 4$ leads to the worst solutions on average, while $8 * 8$ leads to the best ones. As expected, curvature similarity extraction with multi-scale shows the best performance. However, it increases the average run-time by three times. + +Depth Predictors and Curvature Extractors. We then explore the impact of different depth predictors on the final pose accuracy and model size. We select the ground truth depth, MiDAS with ResNet101 [22] architecture and MiDAS small (real-time) to compare with the default predictor MiDAS with ViT [64]. Tab. 4 reports the pose accuracy of these approaches on the ScanNet dataset. MiDAS with ViT, which shows the best performance, has the largest model size due to the heavy depth prediction network. When we employ lighter models for depth prediction, we observe that the performance has only a small drop, indicating that our method is still effective when working with lightweight networks. We also notice that the pose accuracy with ground truth depth is worse than the fine-tuned models. We believe the reason for this is the missing depth regions in the ground truth. We visualize the depth maps in different conditions at supplementary material. + +The last two rows of the table show the results with the proposed ellipse and with quadratic surface fitting. The ellipse fitting performs more accurately in the experiments. + +Mixing Parameter Value. We finetune our model with different values for $\lambda$ in Eq. (10). The results are reported in Tab. 5. The proposed model performs best with $\lambda = 0.9$ for outdoor and $\lambda = 0.8$ for indoor scenes. This is expected since the proposed CSE is additional guidance for coarse-level matching that helps reduce the score of incorrect matches stemming from different surfaces. Assigning a large mixing weight could degrade the performance. + +# 4.4. Curvature Similarity Understanding + +In Fig. 5, we visualize how the curvature similarity score impacts the key points selection and matching results. We first show in the curvature images that similar surfaces share similar curvature values (more visualizations are presented + +Table 4. Depth study. The AUC scores of QuadTree [58] with the proposed CSE using the GT depth or MiDaS [42] with different models. Quadric and surface fitting results are also shown. + +
λ (AUC@5°)1.00.90.80.70.50.3
ScanNet [9]24.925.826.123.821.017.8
MegaDepth [27]53.554.552.847.045.539.8
+ +Table 5. Mixing Parameter Value Selection. We finetune the Quadtree + SCE with different $\lambda$ values and report the AUC@ $5^{\circ}$ on ScanNet and MegaDepth datasets. + +![](images/1fa22869d52a9886e5c4e8f3d9af57eb42671b6d379237f65f8f5101ca1afa67.jpg) +(a) Points found by QuadTree alone (left) and when combined with CSE (right) + +![](images/2aeab890e7e71470619bb07ef73c9eb3e83731bbeb1377e30a7685f4aee942d3.jpg) +(b)Matches found by QuadTree alone (left) and when combined with CSE (right) +Figure 5. Keypoints (top row) and matches (bottom) found by the QuadTree matcher [58] with and without the proposed CSE, overlayed on the RGB (left) and extracted curvature images (right). + +in the supplementary). Then, we observe that, after adding CSE, more matches are detected from the same surface, as we increase the matching confidence of those matches by assigning a larger curvature similarity score. Besides, the curvature score helps in lowering the matching confidence of matches from different surfaces, e.g., the incorrect (red) matches in Fig. 5 (b) left are eliminated in the right. + +# 5. Conclusions + +We present a curvature similarity extractor (CSE), a new algorithm to leverage 3D geometric cues in local feature matching. By fitting quadrics to depth maps obtained from off-the-shelf monocular depth predictors, we extract curvature similarity, which is invariant to translation, rotation, and scaling. CSE can be seamlessly integrated with matchers to guide the match selection. Also, it allows training feature matches and monocular depth networks jointly minimizing the matching loss in an end-to-end manner to further improve the accuracy. Our experiments in Sec. 4 demonstrate the effectiveness of the proposed CSE module, which consistently improves the pose AUC scores of all methods on almost all datasets, achieving state-of-the-art performance on all benchmarks when combined with QuadTree. + +Limitations and Future Work. First, the integration of depth prediction and curvature similarity extraction into feature matching leads to increased memory consumption and inference time. An alternative solution for time-sensitive applications is to use a lightweight depth predictor. As shown in Tab. 4, such an approach still increases the accuracy. Second, the discussed invariance properties do + +not hold at depth discontinuities, and limitations may arise in scenes lacking accurate depth or clear surfaces. Even though this did not pose an issue in our experiments, this is a potential future direction that might lead to further improvements. + +Acknowledgements. This work is supported by the Academy of Finland (grant No. 327911, No. 353138). Daniel Barath is supported by ETH Postdoc fellowship. We acknowledge the computational resources provided by the Aalto Science-IT project and CSC-IT Center for Science, Finland. + +# References + +[1] Pablo Fernández Alcantarilla, Adrien Bartoli, and Andrew J Davison. Kaze features. In Proceedings of European Conference on Computer Vision, pages 214-227, 2012. 1 +[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 7 +[3] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2911-2918, 2012. 1 +[4] Daniel Barath and Jií Matas. Graph-cut ransac. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733-6741, 2018. 2 +[5] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1304-1312, 2020. 2 +[6] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 1, 2 +[7] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 1, 2 +[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In Proceedings of the European Conference on Computer Vision, 2022. 1, 2, 5 +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 6, 7, 8 +[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly sur + +face reintegration. ACM Transactions on Graphics, 36(4):1, 2017. 1 +[11] Luanyuan Dai, Yizhang Liu, Jiayi Ma, Lifang Wei, Taotao Lai, Changcai Yang, and Riqing Chen. Ms2dg-net: Progressive correspondence learning via multiple sparse semantics dynamic graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8973-8982, 2022. 2 +[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 224-236, 2018. 1, 2, 7 +[13] M.P. do Carmo. Differential Geometry of Curves and Surfaces: Revised and Updated Second Edition. Dover Books on Mathematics. Dover Publications, 2016. 3 +[14] Siyan Dong, Shuzhe Wang, Yixin Zhuang, Juho Kannala, Marc Pollefeys, and Baoquan Chen. Visual localization via few-shot scene region classification. International Conference on 3D Vision, 2022. 1 +[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 1, 2 +[16] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2 +[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 7 +[18] Per-Erik Forssén and David G Lowe. Shape descriptors for maximally stable extremal regions. In International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3 +[19] Khang Truong Giang, Soohwan Song, and Sungho Jo. Curvature-guided dynamic scale networks for multi-view stereo. In International Conference on Learning Representations, 2022. 3 +[20] Jack Goldfeather and Victoria Interrante. A novel cubic-order algorithm for approximating principal direction vectors. ACM Trans. Graph., 23(1):45-63, 2004. 4 +[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. CiteSeer, 1988. 1 +[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 8 +[23] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 3 + +[24] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 2 +[25] Qingde Li and J.G. Griffiths. Least squares ellipsoid specific fitting. In Geometric Modeling and Processing, 2004. Proceedings, pages 335-340, 2004. 4, 5 +[26] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. Advances in Neural Information Processing Systems, 33:17346-17357, 2020. 2 +[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 6, 7, 8 +[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3 +[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1 +[30] Xiaotao Liu, Chen Meng, Fei-Peng Tian, and Wei Feng. Dgd-net: Local descriptor guided keypoint detection network. In 2021 IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 2 +[31] Yuan Liu, Lingjie Liu, Cheng Lin, Zhen Dong, and Wenping Wang. Learnable motion coherence for correspondence pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2021. 2 +[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5 +[33] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019. 2 +[34] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004. 1 +[35] Krystian Mokolajczyk and Cordelia Schmid. Scale & affine invariant interest point detectors. International Journal of Computer Vision, 60(1):63-86, 2004. 1 +[36] Krystian Mikolajczyk, Tinne Tuytelaars, Cordelia Schmid, Andrew Zisserman, Jiri Matas, Frederik Schaffalitzky, Timor Kadir, and L Van Gool. A comparison of affine region detectors. International Journal of Computer Vision, 65(1):43-72, 2005. 1 +[37] Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Repeatability is not enough: Learning affine regions via discriminability. In Proceedings of the European Conference on Computer Vision, pages 284-300, 2018. 3 +[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam + +system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1 +[39] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 1 +[40] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3456-3465, 2017. 1 +[41] Rémi Pautrat, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Online invariance selection for local feature descriptors. In European Conference on Computer Vision, pages 707-724. Springer, 2020. 1 +[42] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 2, 3, 4, 6, 8 +[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2, 3 +[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 1, 2 +[45] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2 +[46] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in Neural Information Processing Systems, 31, 2018. 2 +[47] John Rugis and Reinhard Klette. A scale invariant surface curvature estimator. In Pacific-Rim Symposium on Image and Video Technology, pages 138-147. Springer, 2006. 3 +[48] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 7 +[49] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 5, 6, 7 +[50] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In European conference on computer vision, pages 752-765. Springer, 2012. 1 +[51] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1 + +[52] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 1, 7 +[53] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 1, 7 +[54] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 2 +[55] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7 +[56] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11286-11295, 2020. 2 +[57] Dongli Tan, Jiang-Jiang Liu, Xingyu Chen, Chao Chen, Ruixin Zhang, Yunhang Shen, Shouhong Ding, and Rongrong Ji. Eco-tr: Efficient correspondences finding via coarse-to-fine refinement. Proceedings of the European Conference on Computer Vision, 2022. 2 +[58] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1, 2, 5, 6, 7, 8 +[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 6 +[60] Yurun Tian, Vassileios Balntas, Tony Ng, Axel BarrosoLaguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In Proceedings of the Asian Conference on Computer Vision, 2020. 2 +[61] Carl Toft, Daniyar Turmukhambetov, Torsten Sattler, Fredrik Kahl, and Gabriel J Brostow. Single-image depth prediction makes feature matching easier. In European Conference on Computer Vision, pages 473-492. Springer, 2020. 2, 3 +[62] Lijing Tong and Xiang Ying. 3d point cloud initial registration using surface curvature and surf matching. 3D Research, 9:1-16, 2018. 3 +[63] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 1, 2 +[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 8 + +[65] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, 2022. 1, 2 +[66] Shuzhe Wang, Zakaria Laskar, Iaroslav Melekhov, Xiaotian Li, and Juho Kannala. Continual learning for image-based camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3252-3262, 2021. 1 +[67] Changchang Wu. Towards linear-time incremental structure from motion. In International Conference on 3D Vision, pages 127-134. IEEE, 2013. 1 +[68] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 3 +[69] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 1, 2, 7 +[70] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2018. 2 +[71] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 3 +[72] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Hongen Liao. Learning two-view correspondences and geometry using order-aware network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5845-5854, 2019. 2 +[73] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7 +[74] Chen Zhao, Yixiao Ge, Feng Zhu, Rui Zhao, Hongsheng Li, and Mathieu Salzmann. Progressive correspondence pruning by consensus learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6464-6473, 2021. 2 \ No newline at end of file diff --git a/2023/Guiding Local Feature Matching with Surface Curvature/images.zip b/2023/Guiding Local Feature Matching with Surface Curvature/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c838f31608a745a518f6b867405859330043c7c3 --- /dev/null +++ b/2023/Guiding Local Feature Matching with Surface Curvature/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a13ddfc5365db7e481069d78d1e7718b1c44b68dc3f6cb9b17726127498676dd +size 510641 diff --git a/2023/Guiding Local Feature Matching with Surface Curvature/layout.json b/2023/Guiding Local Feature Matching with Surface Curvature/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a29f774fba16b801ae11a22f9897e5079880a888 --- /dev/null +++ b/2023/Guiding Local Feature Matching with Surface Curvature/layout.json @@ -0,0 +1,10983 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 474, + 121 + ], + "type": "text", + "content": "Guiding Local Feature Matching with Surface Curvature" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 142, + 172, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 142, + 172, + 158 + ], + "spans": [ + { + "bbox": [ + 94, + 142, + 172, + 158 + ], + "type": "text", + "content": "Shuzhe Wang" + }, + { + "bbox": [ + 94, + 142, + 172, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 143, + 265, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 143, + 265, + 156 + ], + "spans": [ + { + "bbox": [ + 196, + 143, + 265, + 156 + ], + "type": "text", + "content": "Juho kannala" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 294, + 142, + 381, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 142, + 381, + 157 + ], + "spans": [ + { + "bbox": [ + 294, + 142, + 381, + 157 + ], + "type": "text", + "content": "Marc Pollefeys" + }, + { + "bbox": [ + 294, + 142, + 381, + 157 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 410, + 143, + 483, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 143, + 483, + 156 + ], + "spans": [ + { + "bbox": [ + 410, + 143, + 483, + 156 + ], + "type": "text", + "content": "Daniel Barath2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 176, + 231, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 176, + 231, + 191 + ], + "spans": [ + { + "bbox": [ + 143, + 176, + 231, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 176, + 231, + 191 + ], + "type": "text", + "content": "Aalto University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 267, + 176, + 331, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 176, + 331, + 190 + ], + "spans": [ + { + "bbox": [ + 267, + 176, + 331, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 267, + 176, + 331, + 190 + ], + "type": "text", + "content": "ETH zurich" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 177, + 423, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 177, + 423, + 190 + ], + "spans": [ + { + "bbox": [ + 367, + 177, + 423, + 190 + ], + "type": "text", + "content": "3Microsoft" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "spans": [ + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 258, + 290, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 451 + ], + "type": "text", + "content": "We propose a new method, called curvature similarity extractor (CSE), for improving local feature matching across images. CSE calculates the curvature of the local 3D surface patch for each detected feature point in a viewpoint-invariant manner via fitting quadrics to predicted monocular depth maps. This curvature is then leveraged as an additional signal in feature matching with off-the-shelf matchers like SuperGlue and LoFTR. Additionally, CSE enables end-to-end joint training by connecting the matcher and depth predictor networks. Our experiments demonstrate on large-scale real-world datasets that CSE consistently improves the accuracy of state-of-the-art methods. Fine-tuning the depth prediction network further enhances the accuracy. The proposed approach achieves state-of-the-art results on the ScanNet dataset, showcasing the effectiveness of incorporating 3D geometric information into feature matching." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 486, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 128, + 498 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 506, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 625 + ], + "type": "text", + "content": "Local feature matching is a crucial component for many geometric computer vision tasks, including visual localization [50, 51, 52, 48, 14, 66], structure-from-motion (SfM) [67, 53, 29], and simultaneous localization and mapping (SLAM) [38, 39, 10]. Given a pair of images observing a 3D scene, the task is to find reliable tentative point-to-point correspondences in the two images. Forming such feature matches is often a challenging task as the images may undergo large viewpoint and illumination changes, have occlusions or repetitive patterns." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 627, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 288, + 675 + ], + "type": "text", + "content": "The standard pipeline for local feature matching typically involves two steps: (1) keypoint detection and description and (2) point-wise feature matching. Traditional approaches mainly focus on improving the robustness of" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 425, + 346 + ], + "blocks": [ + { + "bbox": [ + 307, + 232, + 425, + 346 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 425, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 425, + 346 + ], + "type": "image", + "image_path": "c181ac8fd349f5b1758fbde035aacee3f778dd4a3e8c2db5c48d649a3a09ef39.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 349, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 349, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 546, + 392 + ], + "type": "text", + "content": "Figure 1. Curvature-guided match selection: (a) Correct matches in two views must share the same underlying surface curvature, e.g., predicted from monocular depth. (b) Correspondences from different surfaces are guaranteed to be incorrect." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 426, + 234, + 545, + 346 + ], + "blocks": [ + { + "bbox": [ + 426, + 234, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 426, + 234, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 426, + 234, + 545, + 346 + ], + "type": "image", + "image_path": "374c70d2a8eba1f187d384f647deb1c62e8cd64440dfb6c6f094578c8d30b59d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 414, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 546, + 498 + ], + "type": "text", + "content": "keypoint detection and description through, for example, extending the Harris [21] detector to handle affine transformations and multiple scales [35, 36] or using more discriminative or efficient descriptors [32, 7, 34, 1, 3]. However, despite their unbroken popularity, these algorithms often fail to cope with the challenges that arise in real-world environments, leading to low accuracy." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "Recent advances in deep learning-based feature matching have made significant progress in addressing the limitations of hand-crafted approaches, such as by jointly training detectors and descriptors [69, 12, 15, 44, 63, 40] with convolutional neural networks (CNNs), or combining hand-crafted and learning-based descriptors in a hybrid manner [41, 6]. SuperGlue [49] introduced the use of transformer networks to learn the matching process and formulate the problem as an optimal transport task [64]. LoFTR [55] and its recent variants [58, 8, 65] leverage both the global and local context from raw images by jointly learning the feature extraction and matching in a single network. While these approaches have led to state-of-the-art performance on several benchmarks, they work entirely in the 2D image domain and ignore the underlying 3D geometry of the scene. As feature matching is, essentially, finding the corresponding projections of an actual 3D point, this could be a critical limitation in real scenarios where the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 276, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 276, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 276, + 693 + ], + "type": "text", + "content": "*Part of the work was done during the author's visit to ETH zurich." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "Code and trained models are available at https://github.com/AaltoVision/surface-curvature-estimator." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "17981" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "type": "text", + "content": "matched pixels from different views must share the geometry of the underlying surface patch, as shown in Fig. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 97, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 97, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 97, + 286, + 335 + ], + "type": "text", + "content": "In this paper, we focus on investigating and exploiting 3D geometry cues, e.g. coming from monocular depth predictions, for feature matching. The only prior work is [61], which adopts depth priors to find planar image regions that are then rectified to eliminate viewpoint changes. Finally, a handcrafted detector and descriptor are applied to the rectified image regions. This, however, relies on the heavy assumption that the images consist mostly of dominant planes, which severely restricts its out-of-the-box applicability. In this work, we go further by exploring the local surface geometry coming from depth priors. This additional geometric clue is coupled with any off-the-shelf feature matcher increasing its accuracy. Benefiting from advanced monocular depth predictors [42, 43, 16], obtaining depth priors is easy and costs only a few milliseconds. We use these depth predictions to extract curvature at the observed feature points. Exploiting the fact that the curvature is invariant to viewpoint changes (i.e., rotation, translation, and scaling), we enforce that matched features must lie on similar surfaces. Our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 345, + 286, + 543 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 345, + 286, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 286, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 286, + 404 + ], + "type": "text", + "content": "1. We propose an approach for feature matching, leveraging dense depth via utilizing local surface curvature similarity, which is invariant to viewpoint changes. This approach is a departure from traditional matching methods that rely solely on image information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 414, + 286, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 286, + 473 + ], + "type": "text", + "content": "2. The proposed curvature similarity extractor can be seamlessly integrated with any recent feature matcher, making it a versatile tool that can be easily adopted. The experiments show that it improves several recent algorithms on various benchmarks by " + }, + { + "bbox": [ + 54, + 414, + 286, + 473 + ], + "type": "inline_equation", + "content": "1 - 3\\%" + }, + { + "bbox": [ + 54, + 414, + 286, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 483, + 286, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 286, + 543 + ], + "type": "text", + "content": "3. The proposed algorithm can be used to train feature matchers and monocular depth prediction networks jointly in an end-to-end fashion. To demonstrate this, we fine-tune the state-of-the-art MiDaS [42] depth predictor to increase the feature matching accuracy." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 555, + 130, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 130, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 130, + 567 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 574, + 181, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 181, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 181, + 588 + ], + "type": "text", + "content": "2.1. Local Feature Matching" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 713 + ], + "type": "text", + "content": "Detector-based Matchers build sparse correspondences on top of the detected keypoints. Therefore, robust keypoint detection and feature extraction are essential for these approaches. Traditional hand-crafted descriptors, such as SIFT [32] and SURF [7], follow a detect-then-describe pipeline and have shown great success since the 2000s. More recently, learned detectors and descriptors [6, 69, 12, 15, 44, 63, 60, 30] with convolutional neural networks show their superiority in matching images with large viewpoint and appearance changes. This traditional pipeline is further" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "content": "developed to detect-and-describe [12, 15, 44] and describe-to-detect [60, 30] strategies leveraging high-level image information for accurate and reliable matching in challenging conditions. Besides substituting the traditional detectors and descriptors with their learned counterparts for image matching, the recent SuperGlue [49] focuses on the matching stage itself. It replaces the naive mutual nearest neighbour search with Graph Neural Networks (GNN). SuperGlue takes the sparse descriptors and their positional encoding as inputs, then leverage a transformer-based [64] network to create a more robust feature representation for the optimal partial assignment. The follow-up work [54] further improves the performance and efficiency by adaptively clustering sparse features into different subgraphs and using a coarse-to-fine paradigm. Although showing significant improvements with the recent components, the detector-based matchers are naturally limited by the accuracy of detected keypoints coming from independent detectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 291, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 542 + ], + "type": "text", + "content": "Detector-free Matchers directly estimate the dense correspondences from raw images without an independent keypoint detection stage. These approaches leverage low and high-level image information with neural networks and can distinguish indistinctive regions across images. The pioneer work of [46] constructs 4D cost volumes, significantly increasing the matching accuracy. However, it is computationally expensive as the complexity is " + }, + { + "bbox": [ + 304, + 291, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^2)" + }, + { + "bbox": [ + 304, + 291, + 545, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 291, + 545, + 542 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 291, + 545, + 542 + ], + "type": "text", + "content": " is the number of pixels or patches. [45, 26] alleviate the problem by applying either sparse convolutions or a coarse-to-fine paradigm. More recently, transformer-based detector-free matchers [55, 58, 65, 8, 24, 57] have received attention due to their strong performance on local feature matching. The most representative work, LoFTR [55], inherits the advantages of graph matching from SuperGlue and leverages a linear transformer [64] for efficient dense matching. [58, 8] utilise more efficient transformer structures to improve the performance. Detector-free matchers show a promising direction towards local feature matching by encoding rich image information. We claim that extracting 3D geometric information could further enhance accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Correspondence Pruning approaches [17, 4, 5, 72, 70, 56] apply a consensus mechanism to filter out the outliers from putative correspondences coming from the feature matcher procedure. RANSAC [17] and its follow-up works [4, 5] are the most popular correspondence pruning algorithms. In the era of deep learning, OANet [72] infers the probabilities of correspondences being inliers with an order-aware network. Other works improve the accuracy of correspondence pruning by applying motion coherence constraints [31, 33], leveraging local-to-global consensus learning procedure [74, 11], and adopting the attention mechanism [56]. These approaches provide inlier probability predictions for each putative correspondence that can be used either to filter or order matches. Finally, least squares fitting" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "17982" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 277, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 277, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 277, + 85 + ], + "type": "text", + "content": "or traditional robust estimation is applied to the matches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 259, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 259, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 259, + 104 + ], + "type": "text", + "content": "2.2. 3D Geometric Priors to Image Matching" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 109, + 289, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 289, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 289, + 313 + ], + "type": "text", + "content": "Incorporating 3D geometric priors into various vision tasks has been widely explored in recent years. [23] leverages the RGB-D reconstructions to learn view-invariant, geometry-aware 2D representations for downstream tasks. [71] introduces the predicted monocular depth to guide the optimization of neural scene representation. The surface curvature, a popular 3D geometric cue, is widely applied to 3D vision tasks such as 3D point cloud registration [62], multi-view stereo [68, 19]. However, integrating 3D geometric cues into 2D image matching has received little attention. [37, 18] assume that the underlying surface patch is locally planar for feature detection and description, while [61] uses prior depth information to improve feature extraction by rectifying large planar regions. None of the above approaches considers the 3D geometric information in the local region, and we are the first to include 3D surface geometry in general 2D feature matching." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 322, + 279, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 279, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 279, + 336 + ], + "type": "text", + "content": "3. Feature Matching with Surface Curvature" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 342, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 427 + ], + "type": "text", + "content": "This section proposes an approach to improve any local feature matcher by extracting information about the underlying local 3D surface predicted by a deep network. The overview of our method coupled with LoFTR [55] is shown in Fig. 2. Although the proposed algorithm is compatible with any matcher, we demonstrate its effectiveness with LoFTR. Combining it with other ones is straightforward." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 433, + 178, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 178, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 178, + 445 + ], + "type": "text", + "content": "3.1. LoFTR-style Matchers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": "This section provides a brief overview of LoFTR. Given an image pair " + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "inline_equation", + "content": "I_A, I_B \\in \\mathbb{R}^{H*W*3}" + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": ", the end-to-end trainable matcher, first, extracts coarse-level features " + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_A^c, \\mathcal{F}_B^c" + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": " and, then, fine-level ones " + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_A^f, \\mathcal{F}_B^f" + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": " with a local feature CNN [28]. At the coarse level transformer module, the features are flattened and processed by a transformer-based architecture with multiple self- and cross-attention layers [64]. A matching confidence map " + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c" + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": " is predicted at the end of this module, and the matches " + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": " with high confidence are selected from " + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c" + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": " to fine-tune the prediction of fine-level features matching at the coarse-to-fine transformer module. The final outputs are the fine-level matches " + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_f" + }, + { + "bbox": [ + 46, + 451, + 287, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 602, + 217, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 217, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 217, + 615 + ], + "type": "text", + "content": "3.2. Curvature Similarity Extractor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 620, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 288, + 657 + ], + "type": "text", + "content": "In this section, we will discuss multiple potential ways of extracting local surface curvature at detected features. This curvature is then used for improving feature matching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 170, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 170, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 170, + 682 + ], + "type": "text", + "content": "3.2.1 Curvature Similarity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "For an at least two times continuously differentiable surface " + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{C}^2" + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": " in the 3D space, given a point " + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "a direction " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "; the normal curvature " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "k_n \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " measures how curved the surface is at this point along the direction [13]. Since there are infinitely many directions that travel through a point on the surface, there are also infinitely many normal curvatures at a given point. In differential geometry, the minimum normal curvature " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "k_1" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " and the maximum curvature " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "k_2" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " at point " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " are defined as the principal curvatures, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "k_1 \\leq k_2" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ", and the directions of the principal curvatures are the smoothest and steepest directions on the surface. We also define the mean curvature " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " and Gaussian curvature " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 210, + 545, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 210, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 364, + 210, + 545, + 233 + ], + "type": "interline_equation", + "content": "H = \\frac {1}{2} \\left(k _ {1} + k _ {2}\\right), K = k _ {1} k _ {2} \\tag {1}", + "image_path": "dccd9ff27652e7c0c9396bec59a450fcfc8c7eeeae0ed1993bd042e154a6a7d3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 241, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 337 + ], + "type": "text", + "content": "A highly beneficial property of such curvatures is their invariance to surface rotation and translation. Both are important when describing local surface patches centered on features detected in 2D images. However, these measures do not preserve scale invariance [13], which is essential in our case due to using predicted relative depth maps instead of a metric one. This means that, in each image, the depth is only defined up-to-scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 338, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 384 + ], + "type": "text", + "content": "To seek the scale invariance of the curvatures for local patch matching, we follow [47] for the idea of scale-invariant curvature measure and define the curvature similarity function " + }, + { + "bbox": [ + 304, + 338, + 545, + 384 + ], + "type": "inline_equation", + "content": "S: \\mathbb{R} \\times \\mathbb{R} \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 338, + 545, + 384 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 393, + 545, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 393, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 393, + 545, + 420 + ], + "type": "interline_equation", + "content": "S \\left(k _ {1}, k _ {2}\\right) = \\frac {\\operatorname* {m i n} \\left(\\left| k _ {1} \\right| , \\left| k _ {2} \\right|\\right)}{\\operatorname* {m a x} \\left(\\left| k _ {1} \\right| , \\left| k _ {2} \\right|\\right)}, (0 \\leq S (k) \\leq 1), \\tag {2}", + "image_path": "8dcf757e51c1502aeec076dba312052d5425dd0bc28b0fd2fffa28dbfef56892.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "k_{1}" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "k_{2}" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": " are the minimum and maximum curvatures, respectively. It is easy to see that " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "S(k_{1},k_{2})" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": " is rotation, translation and scale invariant. Case " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "S(k_{1},k_{2}) = 1" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": " is interpreted as the curvature of a point on a spherical surface. In the case when " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "k_{1} = k_{2} = 0" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "S(k_{1},k_{2})" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": " to be 0. Detecting points on planes is also straightforward from " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "k_{1}" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "k_{2}" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": " as they both will become " + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 304, + 428, + 545, + 536 + ], + "type": "text", + "content": ". Proof of the invariances is provided in the supplementary material. Next, we will discuss ways of obtaining curvature similarities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 551, + 471, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 471, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 471, + 562 + ], + "type": "text", + "content": "3.2.2 Surface Curvatures Extraction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "This section introduces our approximate form of principal curvature estimation. Different from prior work [68, 19] that extract the curvatures from image intensity, we estimate the principal curvatures based on the depth map. This 3D geometric cue could be obtained either from the active sensors (LiDAR, Kinect) or monocular depth predictors [42, 43]. Since our target curvature similarity " + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "inline_equation", + "content": "S(k_{1}, k_{2})" + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": " is scaling invariant, the scaling ambiguity in monocular depth estimation is neglected and the predicted relative depth is sufficient for the curvature extraction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "In this work, we consider only a pair of RGB images as input and leverage any off-the-shelf depth predictor, e.g." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "17983" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 519, + 247 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 519, + 247 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 519, + 247 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 519, + 247 + ], + "type": "image", + "image_path": "d0dd7ea3279a63ac0d14048d072e0d10beec8d485d747ea8a11b7992c9e121a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "text", + "content": "Figure 2. Overview of the CSE with LoFTR. For a pair of images, it follows the LoFTR [55] matching pipeline to extract the coarse and fine level image features and generate a coarse matching confidence map " + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c" + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "text", + "content": ". In the meantime, the images are fed to a monocular depth network to predict the depth map and extract the curvature similarity upon the depth map. We formulate a new curvature score matrix " + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{curv}" + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "text", + "content": " based on the pixel-wise " + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "text", + "content": " distance of the curvature similarities between two images and mix it with " + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c" + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "text", + "content": ". The mixed score matrix is then processed by the coarse-to-fine transformer module together with the fine-level features to predict the final matches " + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_f" + }, + { + "bbox": [ + 46, + 251, + 548, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": "MiDaS [42], to extract the depth map " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "Z_{i}:\\mathbb{R}\\times \\mathbb{R}\\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": " for each input image, " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2\\}" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i(u,v,Z_i(u,v))" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": " be the 3D surface patch corresponding to the image region at coordinates " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": " with window size " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": "th image, where " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "Z_{i}(u,v)" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": " is the depth of point " + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 327, + 287, + 387 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "text", + "content": "Let us formulate a local point set " + }, + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\{(u,v,Z_i(u,v))\\mid (u,v)\\in \\mathcal{M}_f^i\\} \\subseteq \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_f^i" + }, + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "text", + "content": " are the points detected in the " + }, + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 388, + 288, + 423 + ], + "type": "text", + "content": " th image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 289, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 289, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 289, + 496 + ], + "type": "text", + "content": "Principal Curvature via Surface Fitting. The most straightforward approach is first applying the quadratic surface approximation method [20] to estimate the surface " + }, + { + "bbox": [ + 46, + 424, + 289, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i(u,v,Z_i(u,v))" + }, + { + "bbox": [ + 46, + 424, + 289, + 496 + ], + "type": "text", + "content": " at all points in " + }, + { + "bbox": [ + 46, + 424, + 289, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 424, + 289, + 496 + ], + "type": "text", + "content": ", then employing the first and second fundamental forms of a surface. The normal curvature is calculated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 507, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 507, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 83, + 507, + 287, + 533 + ], + "type": "interline_equation", + "content": "k _ {n} = \\frac {I I}{I} = \\frac {L d u ^ {2} + 2 M d u d v + N d v ^ {2}}{E d u ^ {2} + 2 F d u d v + G d v ^ {2}}, \\tag {3}", + "image_path": "3684cf273b31b092868b6b4d5d2bfb90eaa3caf74b9c12043be6f2023e301eb1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "L = \\mathbf{r}_{uu}\\mathbf{n}" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "M = \\mathbf{r}_{uv}\\mathbf{n}" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "N = \\mathbf{r}_{vv}\\mathbf{n}" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "E = \\mathbf{r}_u\\mathbf{r}_u" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "F = \\mathbf{r}_u\\mathbf{r}_v" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "G = \\mathbf{r}_v\\mathbf{r}_v" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_u, \\mathbf{r}_v, \\mathbf{r}_{uu}, \\mathbf{r}_{uv}, \\mathbf{r}_{vv}" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": " are the first and second-order derivatives of the surface patch " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i(u,v,Z_i(u,v))" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 539, + 287, + 588 + ], + "type": "text", + "content": " is the normal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 598, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 598, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 114, + 598, + 287, + 624 + ], + "type": "interline_equation", + "content": "k _ {n} = \\frac {L + 2 M \\lambda + N \\lambda^ {2}}{E + 2 F \\lambda + G \\lambda^ {2}}. \\tag {4}", + "image_path": "cfa98c73512817581a2b81d952bfdd0893995d7e5db021f0c6f08fd0962f410c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Once we have the normal curvature " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "k_{n}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", we then calculate the partial derivative of " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "k_{n}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " to obtain the principal curvatures " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "k_{1}, k_{2}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ". The details are presented in the supplementary. While this approach works accurately and efficiently in practice, we observed that its gradients become unstable during back-propagation. Therefore, we explore another way as well for calculating the surface curvature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 327, + 547, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 459 + ], + "type": "text", + "content": "**Curvature Extraction via Quadrics.** Another lightweight procedure for estimating the curvature is via fitting quadric surfaces, e.g. ellipsoid, locally to the neighborhood of the observed point. Thus, instead of directly estimating a quadratic surface from the given depth map, we constrain the surface to be a quadric and, more specifically, an ellipsoid. An ellipsoid has the beneficial property that all its curvatures are semi-positive. Its principal axes define three directions which we can use the define the similarity. We use the radii " + }, + { + "bbox": [ + 304, + 327, + 547, + 459 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 304, + 327, + 547, + 459 + ], + "type": "text", + "content": " along the three axes and define the curvature similarity score as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 360, + 465, + 545, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 465, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 360, + 465, + 545, + 495 + ], + "type": "interline_equation", + "content": "S (\\alpha , \\beta , \\gamma) = \\frac {\\operatorname* {m i n} \\left(\\alpha^ {2} , \\beta^ {2} , \\gamma^ {2}\\right)}{\\operatorname* {m a x} \\left(\\alpha^ {2} , \\beta^ {2} , \\gamma^ {2}\\right)}. \\tag {5}", + "image_path": "89ac9ea325643563283f6350ffded134eafa4de11c8d23e01edf1b00da82bd00.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "text", + "content": "Similarity " + }, + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 500, + 547, + 548 + ], + "type": "text", + "content": " is rotation, translation, scale and thus, viewpoint invariant. The advantage from Eq. (5) is that the computation of first and second-order gradients are no longer required for the curvature similarity calculation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 548, + 547, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 572 + ], + "type": "text", + "content": "The general constraint that a quadric imposes on a point " + }, + { + "bbox": [ + 304, + 548, + 547, + 572 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{P}_i" + }, + { + "bbox": [ + 304, + 548, + 547, + 572 + ], + "type": "text", + "content": " lying on the surface is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 579, + 479, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 579, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 373, + 579, + 479, + 594 + ], + "type": "interline_equation", + "content": "\\mathbf {x} \\mathbf {Q} _ {i} \\mathbf {x} ^ {\\mathrm {T}} + \\mathbf {P} _ {i} \\mathbf {x} ^ {\\mathrm {T}} + \\mathbf {R} _ {i} = 0,", + "image_path": "e7dc65dfad38a2e55f8ae0b9bda4337c4e84150638ec80b095ad6a5039dcd2b4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 304, + 601, + 545, + 625 + ], + "type": "text", + "content": " are the parameters in a matrix form. This can be written as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 631, + 545, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 631, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 353, + 631, + 545, + 660 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} a x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \\tag {6} \\\\ 2 f y z + 2 g x + 2 h y + 2 i z + j = 0. \\\\ \\end{array}", + "image_path": "c7caa50add9654b2e6cd6ae712f25eda04448ebf44e77c7def49ab574c6f535e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 668, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 668, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 668, + 545, + 692 + ], + "type": "text", + "content": "As described in [25], a sufficient condition that guarantees the quadric surface to be an ellipsoid is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 392, + 700, + 459, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 700, + 459, + 713 + ], + "spans": [ + { + "bbox": [ + 392, + 700, + 459, + 713 + ], + "type": "interline_equation", + "content": "4 M - N ^ {2} = 1,", + "image_path": "c60f2f3b4059d24eb57a9c764c06b8882634545534dc302b3b93700c27d17885.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "17984" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 76, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 76, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 76, + 83 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 164, + 94, + 229, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 94, + 229, + 105 + ], + "spans": [ + { + "bbox": [ + 164, + 94, + 229, + 105 + ], + "type": "interline_equation", + "content": "N = a + b + c,", + "image_path": "21847d37b80fab74e03a1e1c84e9531fe033626996efa07e9dd57fd4dee81480.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 84, + 107, + 230, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 107, + 230, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 107, + 230, + 121 + ], + "type": "interline_equation", + "content": "M = a b + b c + a c - d ^ {2} - e ^ {2} - f ^ {2}.", + "image_path": "591f8f9f52f1c83b0e60cd8dbd4e55c9a1360b7fcfaf7353ad7c4d35f76babd3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "text", + "content": "Also, since we aim at measuring the curvature of the surface patch located on the observed point, we have to constrain the ellipsoid so that the observed point " + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "text", + "content": " is on its surface. To do so, we translate the point set " + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{P}' = \\{\\mathbf{q} - \\mathbf{p} \\mid \\mathbf{q} \\in \\mathcal{P}\\}" + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "text", + "content": " and fix " + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "inline_equation", + "content": "j = 0" + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "text", + "content": ". The central point now is " + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' = [0,0,0]^{\\mathrm{T}}" + }, + { + "bbox": [ + 47, + 130, + 287, + 190 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 190, + 285, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 190, + 285, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 190, + 285, + 201 + ], + "type": "text", + "content": "Substituting all these constraints into Eq. (6), we obtain" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 209, + 287, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 209, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 94, + 209, + 287, + 229 + ], + "type": "interline_equation", + "content": "a x ^ {2} + b y ^ {2} + c z ^ {2} + 2 d y z + 2 e x z + \\tag {7}", + "image_path": "252e23dfdd59c975eeb33fbec26c4280594924859d3cc964c6c88c748c6ada5f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 227, + 238, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 227, + 238, + 238 + ], + "spans": [ + { + "bbox": [ + 113, + 227, + 238, + 238 + ], + "type": "interline_equation", + "content": "2 f y z + 2 g x + 2 h y + 2 i z = 0.", + "image_path": "73f29eea25070f5d2d13bb0e48571d7be69f41316179da2c5a8de137631794c5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 247, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 287, + 270 + ], + "type": "text", + "content": "From Eq. (7), only nine variables need to be estimated in our case. For each point " + }, + { + "bbox": [ + 47, + 247, + 287, + 270 + ], + "type": "inline_equation", + "content": "[u, v, z]^{\\mathrm{T}} \\in \\mathcal{P}^{\\prime}" + }, + { + "bbox": [ + 47, + 247, + 287, + 270 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 47, + 247, + 287, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 47, + 247, + 287, + 270 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 278, + 287, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 287, + 293 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {i} = \\left[ u _ {i} ^ {2}, v _ {i} ^ {2}, z _ {i} ^ {2}, 2 v _ {i} z _ {i}, 2 u _ {i} z _ {i}, 2 u _ {i} v _ {i}, 2 u _ {i}, 2 v _ {i}, 2 z _ {i} \\right] ^ {\\mathrm {T}}, \\tag {8}", + "image_path": "10a8c8bfcb24aeffe97c3c14ce5fc3648c6721c27dd458abc33efa53ac45514d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 300, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 350 + ], + "type": "text", + "content": "and coefficient matrix " + }, + { + "bbox": [ + 47, + 300, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = [\\mathbf{X}_1, \\mathbf{X}_2, \\mathbf{X}_3, \\dots, \\mathbf{X}_n]" + }, + { + "bbox": [ + 47, + 300, + 287, + 350 + ], + "type": "text", + "content": ". Then, we follow the method in [25] for the ellipsoid fitting and estimate its radii " + }, + { + "bbox": [ + 47, + 300, + 287, + 350 + ], + "type": "inline_equation", + "content": "R = (\\alpha, \\beta, \\gamma)" + }, + { + "bbox": [ + 47, + 300, + 287, + 350 + ], + "type": "text", + "content": ". The algorithms for fitting and calculating the radii are shown in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 356, + 235, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 235, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 235, + 367 + ], + "type": "text", + "content": "3.3. Curvature-Guided Match Selection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": "The proposed curvature similarity extractor (CSE) is a depth-based plug-in component that can be combined with both detector-based [32, 49] and detector-free [55, 58, 8] matchers. For an image pair " + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "inline_equation", + "content": "(I_A, I_B)" + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": ", the monocular depth predictor first estimates the depth map for each image independently. CSE extracts the curvature map " + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "inline_equation", + "content": "(S_A, S_B)" + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": " on the top of the depth maps. Note that LoFTR selects the candidate matches " + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "inline_equation", + "content": "M_c" + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": " at the coarse level of 1/8-resolution grids where the ground truth matches are defined based on the reprojection distance of the centers of the grid cells. Thus, it is not necessary to extract the curvature similarity at each pixel. It is calculated only at the center point of each cell, i.e., " + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "inline_equation", + "content": "S_A, S_B \\in \\mathbb{R}^{(h * w)*1}" + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "inline_equation", + "content": "h = H / 8" + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "inline_equation", + "content": "w = W / 8" + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "text", + "content": "Since the proposed curvature similarity is rotation, translation and scale invariant, " + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "inline_equation", + "content": "S_{A}" + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "text", + "content": " at point " + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_A" + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "text", + "content": " in the first image should be equal to the similarity " + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "inline_equation", + "content": "S_B" + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "text", + "content": " of its corresponding pair " + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_B" + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "text", + "content": " in the second image. The correct matches " + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "inline_equation", + "content": "M_c" + }, + { + "bbox": [ + 46, + 530, + 287, + 601 + ], + "type": "text", + "content": ", ideally, have the same curvature values. To measure this, we define the normalised curvature score matrix as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 100, + 611, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 611, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 100, + 611, + 287, + 623 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {\\text {c u r v}} = 1 - L _ {2} (S (k) _ {A}, S (k) _ {B}), \\tag {9}", + "image_path": "c144a428c1c1fdb0a1bfef58ee1cf182596ed77a6bf99fafefea390698c2d1a9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "inline_equation", + "content": "L_{2}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "text", + "content": " is the L2-norm of " + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "inline_equation", + "content": "S(k)_A" + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "inline_equation", + "content": "S(k)_B" + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "text", + "content": ". Since both " + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "inline_equation", + "content": "S(k)_A" + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "inline_equation", + "content": "S(k)_B" + }, + { + "bbox": [ + 47, + 632, + 287, + 668 + ], + "type": "text", + "content": " are in-between zero and one, the curvature score is also normalized into range [0, 1]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 669, + 287, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 287, + 692 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 287, + 692 + ], + "type": "text", + "content": "The new matching score matrix is the linear combination of " + }, + { + "bbox": [ + 47, + 669, + 287, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 47, + 669, + 287, + 692 + ], + "type": "text", + "content": " and LoFTR matching probability " + }, + { + "bbox": [ + 47, + 669, + 287, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c" + }, + { + "bbox": [ + 47, + 669, + 287, + 692 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 701, + 287, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 113, + 701, + 287, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {P} = \\lambda \\mathcal {P} _ {c} + (1 - \\lambda) \\mathcal {P} _ {\\text {c u r v}}, \\tag {10}", + "image_path": "f9dee5fdd8fd1e5a527ced1fc066e639c8eaeed673a12787d295d3b5444b0245.jpg" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 314, + 71, + 539, + 152 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 539, + 152 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 539, + 152 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 539, + 152 + ], + "type": "image", + "image_path": "47e53b10f054a133c4167ac138e30950c20b6933afa6ee0f90c7b7be82554861.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 155, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 188 + ], + "type": "text", + "content": "Figure 3. Multi-scale curvature extraction. We extract the curvature similarity from ellipsoids centered on the observed point, fitted with different window (i.e., 3D neighborhood) sizes." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "text", + "content": " is the mixing parameter for balancing the two matching scores. Since our " + }, + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "text", + "content": " is an additional supporting score for coarse-level matching selection, we set the parameter " + }, + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 205, + 545, + 263 + ], + "type": "text", + "content": " to 0.9 in all our experiments. We adopt the scheme of LoFTR for the coarse-level match selection as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 270, + 545, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 270, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 318, + 270, + 545, + 283 + ], + "type": "interline_equation", + "content": "M _ {c} = \\{(i, j) \\mid \\forall (i, j) \\in \\operatorname {M N N} (\\mathcal {P}), \\mathcal {P} (i, j) \\geq \\theta \\}, \\tag {11}", + "image_path": "af6ee0e6acae3aa5dbf7a929fdadb9c7f80652b7315a5b821d5400c34fcc3e21.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "text", + "content": "where MNN is the mutual nearest neighbor operation, " + }, + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "text", + "content": " is a threshold so that we only select the matches with confidence higher than " + }, + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 304, + 289, + 545, + 335 + ], + "type": "text", + "content": " can be similarly combined with the matching score matrix of SuperGlue [49]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": "Multi-scale Curvature Similarity. Although the proposed curvature similarity is scaling invariant in theory, it still depends on the neighboring pixels of the points. Thus, the multi-scale extraction of this geometric cue could also be adopted for robustness to scale change in the images. For detector-based approaches, such as SuperGlue [49], we use a set of window size " + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\{\\delta_1,\\delta_2,\\dots,\\delta_n\\}" + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": " with the window centered on detected keypoints. Hence, the curvature similarity maps are " + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "inline_equation", + "content": "S(k)_A\\in \\mathbb{R}^{\\mathcal{A}*n}" + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "inline_equation", + "content": "S(k)_B\\in \\mathbb{R}^{\\mathcal{B}*n}" + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": " are the detected keypoints at " + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "inline_equation", + "content": "I_A" + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "inline_equation", + "content": "I_B" + }, + { + "bbox": [ + 304, + 336, + 546, + 492 + ], + "type": "text", + "content": ", respectively. For detector-free matchers, the multi-scale extraction is conducted within the patch grid. Fig. 3 describes the curvature similarity extraction at multiple scales." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 498, + 467, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 467, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 467, + 510 + ], + "type": "text", + "content": "3.4. Depth Prediction Fine-tuning" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "type": "text", + "content": "Since the proposed curvature similarity extraction is fully differentiable, it is possible to use it to train feature matching and depth prediction networks jointly end-to-end. To demonstrate this, we fine-tune the depth predictor with the curvature similarity extractor so that it takes the feature matching into account when being trained. The new confidence matrix " + }, + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 516, + 545, + 612 + ], + "type": "text", + "content": " is minimized with the negative log-likelihood, similarly as in [49, 55], where the loss is" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 354, + 616, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 616, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 354, + 616, + 545, + 647 + ], + "type": "interline_equation", + "content": "L = - \\frac {1}{\\left| M _ {c} ^ {g t} \\right|} \\sum_ {(i, j) \\in M _ {c} ^ {g t}} \\log \\mathcal {P} (i, j), \\tag {12}", + "image_path": "20ba153c605974ba080cd6acd60a160ad66088dc7eac2e7f0614547a9007c325.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_c^{gt}" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " is the ground truth matches defined at the coarse-level in [55]. Since the loss term focuses only on the matched patches, long training time would crash the depth prediction model. Thus, we fine-tune the depth predictor with only a few epochs and small learning rates." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "17985" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 198 + ], + "type": "table", + "html": "
MethodsYFCC100M [59]MegaDepth [27]ScanNet [9]
AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)AUC@5°@10°@20°P(%)MS(%)
SuperGlue [49]38.959.475.798.723.643.762.277.099.632.516.133.851.884.431.5
SuperGlue + CSE40.060.176.395.626.546.264.878.798.335.716.234.052.280.731.7
LoFTR [55]43.463.178.095.28.253.770.182.496.89.022.140.857.687.47.9
LoFTR + CSE43.563.378.193.48.554.470.482.195.59.222.942.460.086.49.3
LoFTR + CSE (w/ FT)43.963.578.193.38.554.370.482.295.59.323.943.560.586.19.5
QuadTree [58]44.163.578.396.58.953.570.282.298.59.624.944.761.689.49.8
QuadTree + CSE45.064.378.895.39.154.070.382.197.99.825.245.562.589.411.2
QuadTree + CSE (w/ FT)45.064.478.995.59.154.571.182.998.09.725.846.163.489.211.5
", + "image_path": "afd6f1ec42bf39954ee3175271851e6a826101a24861aa23cd78a894b478218c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "text", + "content": "Table 1. Results of relative pose estimation on YFCC100M, MegaDepth, and ScanNet datasets. We report AUC score of the translation and rotation errors with different thresholds " + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "inline_equation", + "content": "(5^{\\circ}, 10^{\\circ}, 20^{\\circ})" + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "text", + "content": ", precision " + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "text", + "content": ", and matching score " + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 205, + 547, + 251 + ], + "type": "text", + "content": ". Our methods and the best performance are marked in bold. w/FT indicates the results with fine-tuning. Note that comparison among baselines is not straightforward as some sensitive default parameters, e.g. input image size, are different." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 47, + 259, + 545, + 417 + ], + "blocks": [ + { + "bbox": [ + 47, + 259, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 47, + 259, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 545, + 417 + ], + "type": "image", + "image_path": "551650444ceed7e5e1679b6f9abcde28daab7aad78256588a1848430af233c86.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 419, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 419, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 547, + 456 + ], + "type": "text", + "content": "Figure 4. Qualitative matching results. We add the curvature similarity extractor module to LoFTR [55] and QuadTree [58] and compare the two approaches in both indoor and outdoor scenes. By adding the CSE module, we are able to obtain more correct matches (green lines) and better pose accuracy while keeping similar matching precision. The matches are coloured by the epipolar error threshold in [55]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 472, + 128, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 128, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 128, + 486 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 494, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 288, + 567 + ], + "type": "text", + "content": "We first demonstrate the effectiveness of the proposed similarity curvature extractor on relative pose estimation and the downstream visual localization tasks. Next, we analyze the choice of architecture and hyper-parameters in the ablation study. Finally, we provide qualitative results to help understand the similarity curvature extractor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 640 + ], + "type": "text", + "content": "Implementation Details. Our proposed CSE adopts Mi-DAS + ViT [42] as the default depth predictor. For detector-based approaches [55, 58], we extract the curvature similarity at keypoint locations and use multi-scales " + }, + { + "bbox": [ + 46, + 567, + 288, + 640 + ], + "type": "inline_equation", + "content": "\\{7 * 7, 9 * 9, 11 * 11\\}" + }, + { + "bbox": [ + 46, + 567, + 288, + 640 + ], + "type": "text", + "content": ". For detector-free matchers, we use the default grid patch size " + }, + { + "bbox": [ + 46, + 567, + 288, + 640 + ], + "type": "inline_equation", + "content": "8 * 8" + }, + { + "bbox": [ + 46, + 567, + 288, + 640 + ], + "type": "text", + "content": " for curvature estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "For the depth predictor training, following the default setting in LoFTR [55], we fine-tune the predictor with the outdoor matching model on MegaDepth [27]. The model is optimized using Adam with a fixed learning rate of " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "r = 1e - 5" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " and batch size of 2. The indoor model is fine-tuned on ScanNet [9] with " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "r = 5e - 6" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ". In the training process, the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 559 + ], + "type": "text", + "content": "weights from the matcher are frozen, and only the weights from refineNets in MiDAS + ViT are updated. Instead of sampling 200 pairs from each scene, we reduce it to 10 pairs. The full training and evaluation process of the outdoor model is conducted on 2 Telsa-V100 GPUs with 32GB of memory. We train and evaluate the indoor model on 2 RTX 2080 Ti GPUs with a memory of 12GB." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 563, + 444, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 444, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 444, + 575 + ], + "type": "text", + "content": "4.1. Relative Pose Estimation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Datasets. We evaluate our CSE on three popular datasets, ScanNet [9], MegaDepth [27], and YFCC100M [59], for the two-view pose estimation task. ScanNet is an indoor RGB-D dataset consisting of 1,613 sequences with 2.5 million views, and each view is associated with a ground truth camera pose and depth map. Similar to the setting in LoFTR [49], the images and depth maps are resized to " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "640*480" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " for training and evaluation. MegaDepth and YFCC100M are two outdoor datasets which consist of multiple scenarios. We fine-tune the outdoor model on MegaDepth with the longest dimension of the images re" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "17986" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "sized to 840 and evaluate our method on MegaDepth-1500 from [55] and YFCC100M-4000 from [49]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 252 + ], + "type": "text", + "content": "Baselines. The proposed CSE is a plug-in module that can be applied with both detector-based and detector-free matchers. We add this module to detector-based SuperGlue with SuperPoint [12] features, detector-free LoFTR, and its recent variant QuadTree [58]. The experiments follow the default setting from the original papers for all the baselines. For all the experiments on ScanNet, the image resolution is consistently " + }, + { + "bbox": [ + 46, + 96, + 287, + 252 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 96, + 287, + 252 + ], + "type": "text", + "content": ". For evaluation on MegaDepth and YFCC100M, the default input sizes are different among the approaches. Thus, it is not reasonable to compare among the baselines as the input image size significantly impacts the final accuracy. We detail the default input sizes of different approaches in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 288, + 396 + ], + "type": "text", + "content": "Metrics. Similar to the previous methods [49, 55], we report the Area Under the recall Curve (AUC) of the translation and rotation errors with multiple thresholds " + }, + { + "bbox": [ + 46, + 253, + 288, + 396 + ], + "type": "inline_equation", + "content": "(5^{\\circ}, 10^{\\circ}, 20^{\\circ})" + }, + { + "bbox": [ + 46, + 253, + 288, + 396 + ], + "type": "text", + "content": ". For each setting, we consider a pose to be correct if the maximum error from translation and rotation error is below the threshold. The estimated camera poses are recovered from the essential matrix estimated by RANSAC [17] with the inlier-outlier threshold set at 0.5 pixels. Besides, we report the matching precision and score [69, 12]. To calculate matching scores on detector-free matchers, we consider the number of coarse-level features as the number of detected keypoints, i.e., " + }, + { + "bbox": [ + 46, + 253, + 288, + 396 + ], + "type": "inline_equation", + "content": "h * w" + }, + { + "bbox": [ + 46, + 253, + 288, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 397, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 588 + ], + "type": "text", + "content": "Results. We report the results of the CSE add-on on SuperGlue, LoFTR, and QuadTree in Tab. 1. For LoFTR and QuadTree, we report the results of directly adding the CSE module and adding CSE with fine-tuning. As the official training code for SuperGlue is not publicly available, we evaluated the SuperGlue + CSE without fine-tuning. However, we believe that fine-tuning would further improve the results of SuperGlue as well. We also notice that the MegaDepth-1500 is a part of the training set for SuperGlue. However, this does not invalidate the fact that the proposed method helps in finding better features. The proposed CSE improves the pose AUC scores of all methods on almost all datasets. Fine-tuning further increases the accuracy, and QuadTree combined with the proposed CSE method and fine-tuned on depth predictions achieves state-of-the-art AUC scores on all tested datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 649 + ], + "type": "text", + "content": "As for matching score and precision, we also visualize the matching results of LoFTR and QuadTree in Fig. 4. By adding CSE, we obtain more correct matches and better pose accuracy on surface regions with distinctive geometry, while keeping similar matching precision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 160, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 160, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 160, + 670 + ], + "type": "text", + "content": "4.2. Visual Localization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Datasets and Metrics. We select the Aachen DAY-NIGHT v1.1 dataset [52] to demonstrate the effectiveness of our proposed approach on the visual localization task." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 140 + ], + "type": "table", + "html": "
MethodsDayNight
(0.25m, 2°) / (0.5m, 5°) / (5m,10°)
SP [12] + SG [49]89.8 / 96.1 / 99.477.0 / 90.6 / 100.0
SP + SG + CSE90.7 / 96.5 / 99.475.9 / 91.1 / 100.0
", + "image_path": "12dc6cdca5d4178357a03f0e052be14856340065336cdc816531be4c3693ddee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "text", + "content": "Aachen DAY-NIGHT is a challenging urban-scale outdoor dataset which is collected by handheld devices with 6,697 reference images and 1,015 queries, including 191 nighttime images. The ground truth 6DoF camera pose is obtained by COLMAP [53] and refined by [73]. We follow the evaluation protocols of the visual localization benchmark [52] reporting the translation and rotation error recalls at " + }, + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{m} / 2^{\\circ}" + }, + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m} / 5^{\\circ}" + }, + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "inline_equation", + "content": "5\\mathrm{m} / 10^{\\circ}" + }, + { + "bbox": [ + 304, + 202, + 545, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "text", + "content": "Results. We add our CSE to the SuperPoint [12] + SuperGlue [49] pipeline in HLoc [48, 49] without fine-tuning any models. The matching pairs are provided by HLoc with top-50 candidate images from NetVLAD [2]. Tab. 2 reports the pose error recalls. We observe that the plug-in CSE improves the SP + SG pipeline on all thresholds on the Day sequences. On the Night sequence, it improves at threshold " + }, + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "inline_equation", + "content": "(0.5\\mathrm{m} / 5^{\\circ})" + }, + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "inline_equation", + "content": "(5\\mathrm{m} / 10^{\\circ})" + }, + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "text", + "content": ". We believe the slight decrease at " + }, + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "inline_equation", + "content": "(0.25\\mathrm{m} / 1^{\\circ})" + }, + { + "bbox": [ + 304, + 298, + 546, + 431 + ], + "type": "text", + "content": " stems from the inaccuracies of depth prediction on nighttime images. Overall, by adding the CSE module, the localization accuracy is improved." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 439, + 400, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 400, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 400, + 452 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "text", + "content": "We analyse the impact of different design components on our curvature similarity extractor in this section. We evaluate two main design choices: (1) multi-scale curvature extraction, (2) choice of depth predictors, and (3) mixing parameter value " + }, + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "text", + "content": " selection. All experiments are conducted with the Quadtree [58] matcher." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 541, + 545, + 622 + ], + "blocks": [ + { + "bbox": [ + 305, + 148, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 547, + 182 + ], + "type": "text", + "content": "Table 2. Visual localization. We report the pose recall at " + }, + { + "bbox": [ + 305, + 148, + 547, + 182 + ], + "type": "inline_equation", + "content": "(0.25\\mathrm{m} / 2^{\\circ}" + }, + { + "bbox": [ + 305, + 148, + 547, + 182 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m} / 5^{\\circ}" + }, + { + "bbox": [ + 305, + 148, + 547, + 182 + ], + "type": "inline_equation", + "content": "5\\mathrm{m} / 10^{\\circ})" + }, + { + "bbox": [ + 305, + 148, + 547, + 182 + ], + "type": "text", + "content": " on the Aachen Day-Night [52] dataset with the HLoc [48] algorithm. The best results are bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 541, + 545, + 622 + ], + "lines": [ + { + "bbox": [ + 308, + 541, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 541, + 545, + 622 + ], + "type": "table", + "html": "
Multi-ScaleScanNet [9]MegaDepth [27]Time (ms)
AUC@5°/ AUC@10°/ AUC@20°
4*424.444.661.952.769.682.0113
6*625.345.662.853.369.981.8115
8*8 (default)25.245.562.554.070.382.1112
4*4+6*6+8*825.645.262.454.070.082.1338
", + "image_path": "d9251fcd1393f767a6f045cad79ec023821990f38e9bda8bda2a61f849ea2d56.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 631, + 545, + 664 + ], + "lines": [ + { + "bbox": [ + 305, + 631, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 305, + 631, + 545, + 664 + ], + "type": "text", + "content": "Table 3. Multi-scale and different grid sizes. The AUC scores and average CSE run-times (ms) of Quadtree [58] + CSE with different grid sizes and on multi-scale estimation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Multi-scale. As described in Sec. 3.3, we conduct multiscale curvature similarity extraction on a " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "8*8" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " patch grid with QuadTree by default. In Tab. 3, we show the results" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "17987" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 148 + ], + "type": "table", + "html": "
Depth PredictorsCSEFTAUC@5@10@20Model
Ground TruthQuadricX25.145.062.0153 MB
MiDAS (Res101)Quadric✓25.545.562.5575 MB
MiDAS (Real-time)Quadric✓25.545.462.9239 MB
MiDAS (ViT)Quadric✓25.846.163.41.6 GB
MiDAS (ViT)Surface FitX24.844.761.61.6 GB
", + "image_path": "c98367b5e4af7a7e71cf8278c4b400303765747e06d48d9c431449ebc67e6504.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": "when using " + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "inline_equation", + "content": "4 * 4" + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "inline_equation", + "content": "6 * 6" + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": ", and multi-scale grids without fin-tuning. Even though the proposed CSE is robust to the grid size and the accuracy is similar, it slightly increases together with the grid size. Size " + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "inline_equation", + "content": "4 * 4" + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": " leads to the worst solutions on average, while " + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "inline_equation", + "content": "8 * 8" + }, + { + "bbox": [ + 46, + 209, + 286, + 304 + ], + "type": "text", + "content": " leads to the best ones. As expected, curvature similarity extraction with multi-scale shows the best performance. However, it increases the average run-time by three times." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 286, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 286, + 508 + ], + "type": "text", + "content": "Depth Predictors and Curvature Extractors. We then explore the impact of different depth predictors on the final pose accuracy and model size. We select the ground truth depth, MiDAS with ResNet101 [22] architecture and MiDAS small (real-time) to compare with the default predictor MiDAS with ViT [64]. Tab. 4 reports the pose accuracy of these approaches on the ScanNet dataset. MiDAS with ViT, which shows the best performance, has the largest model size due to the heavy depth prediction network. When we employ lighter models for depth prediction, we observe that the performance has only a small drop, indicating that our method is still effective when working with lightweight networks. We also notice that the pose accuracy with ground truth depth is worse than the fine-tuned models. We believe the reason for this is the missing depth regions in the ground truth. We visualize the depth maps in different conditions at supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 544 + ], + "type": "text", + "content": "The last two rows of the table show the results with the proposed ellipse and with quadratic surface fitting. The ellipse fitting performs more accurately in the experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "text", + "content": "Mixing Parameter Value. We finetune our model with different values for " + }, + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "text", + "content": " in Eq. (10). The results are reported in Tab. 5. The proposed model performs best with " + }, + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\lambda = 0.9" + }, + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "text", + "content": " for outdoor and " + }, + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "inline_equation", + "content": "\\lambda = 0.8" + }, + { + "bbox": [ + 46, + 544, + 286, + 640 + ], + "type": "text", + "content": " for indoor scenes. This is expected since the proposed CSE is additional guidance for coarse-level matching that helps reduce the score of incorrect matches stemming from different surfaces. Assigning a large mixing weight could degrade the performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 241, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 241, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 241, + 660 + ], + "type": "text", + "content": "4.4. Curvature Similarity Understanding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "In Fig. 5, we visualize how the curvature similarity score impacts the key points selection and matching results. We first show in the curvature images that similar surfaces share similar curvature values (more visualizations are presented" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 312, + 71, + 539, + 121 + ], + "blocks": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "content": "Table 4. Depth study. The AUC scores of QuadTree [58] with the proposed CSE using the GT depth or MiDaS [42] with different models. Quadric and surface fitting results are also shown." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 71, + 539, + 121 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 539, + 121 + ], + "type": "table", + "html": "
λ (AUC@5°)1.00.90.80.70.50.3
ScanNet [9]24.925.826.123.821.017.8
MegaDepth [27]53.554.552.847.045.539.8
", + "image_path": "a7b1191f08f16c471a9b092dd9432e7f357cdd849b6351d5b9f248cef639e60d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 128, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 305, + 128, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 305, + 128, + 545, + 161 + ], + "type": "text", + "content": "Table 5. Mixing Parameter Value Selection. We finetune the Quadtree + SCE with different " + }, + { + "bbox": [ + 305, + 128, + 545, + 161 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 128, + 545, + 161 + ], + "type": "text", + "content": " values and report the AUC@ " + }, + { + "bbox": [ + 305, + 128, + 545, + 161 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 305, + 128, + 545, + 161 + ], + "type": "text", + "content": " on ScanNet and MegaDepth datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 172, + 544, + 217 + ], + "blocks": [ + { + "bbox": [ + 307, + 172, + 544, + 217 + ], + "lines": [ + { + "bbox": [ + 307, + 172, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 544, + 217 + ], + "type": "image", + "image_path": "1fa22869d52a9886e5c4e8f3d9af57eb42671b6d379237f65f8f5101ca1afa67.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 218, + 539, + 227 + ], + "lines": [ + { + "bbox": [ + 315, + 218, + 539, + 227 + ], + "spans": [ + { + "bbox": [ + 315, + 218, + 539, + 227 + ], + "type": "text", + "content": "(a) Points found by QuadTree alone (left) and when combined with CSE (right)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 544, + 274 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 544, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 544, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 544, + 274 + ], + "type": "image", + "image_path": "2aeab890e7e71470619bb07ef73c9eb3e83731bbeb1377e30a7685f4aee942d3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 276, + 541, + 285 + ], + "lines": [ + { + "bbox": [ + 310, + 276, + 541, + 285 + ], + "spans": [ + { + "bbox": [ + 310, + 276, + 541, + 285 + ], + "type": "text", + "content": "(b)Matches found by QuadTree alone (left) and when combined with CSE (right)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 289, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 323 + ], + "type": "text", + "content": "Figure 5. Keypoints (top row) and matches (bottom) found by the QuadTree matcher [58] with and without the proposed CSE, overlayed on the RGB (left) and extracted curvature images (right)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 344, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 429 + ], + "type": "text", + "content": "in the supplementary). Then, we observe that, after adding CSE, more matches are detected from the same surface, as we increase the matching confidence of those matches by assigning a larger curvature similarity score. Besides, the curvature score helps in lowering the matching confidence of matches from different surfaces, e.g., the incorrect (red) matches in Fig. 5 (b) left are eliminated in the right." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 441, + 383, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 383, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 383, + 453 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "text", + "content": "We present a curvature similarity extractor (CSE), a new algorithm to leverage 3D geometric cues in local feature matching. By fitting quadrics to depth maps obtained from off-the-shelf monocular depth predictors, we extract curvature similarity, which is invariant to translation, rotation, and scaling. CSE can be seamlessly integrated with matchers to guide the match selection. Also, it allows training feature matches and monocular depth networks jointly minimizing the matching loss in an end-to-end manner to further improve the accuracy. Our experiments in Sec. 4 demonstrate the effectiveness of the proposed CSE module, which consistently improves the pose AUC scores of all methods on almost all datasets, achieving state-of-the-art performance on all benchmarks when combined with QuadTree." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Limitations and Future Work. First, the integration of depth prediction and curvature similarity extraction into feature matching leads to increased memory consumption and inference time. An alternative solution for time-sensitive applications is to use a lightweight depth predictor. As shown in Tab. 4, such an approach still increases the accuracy. Second, the discussed invariance properties do" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "17988" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": "not hold at depth discontinuities, and limitations may arise in scenes lacking accurate depth or clear surfaces. Even though this did not pose an issue in our experiments, this is a potential future direction that might lead to further improvements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 289, + 204 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the Academy of Finland (grant No. 327911, No. 353138). Daniel Barath is supported by ETH Postdoc fellowship. We acknowledge the computational resources provided by the Aalto Science-IT project and CSC-IT Center for Science, Finland." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 217, + 106, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 106, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 106, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 237, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 237, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 237, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 237, + 287, + 270 + ], + "type": "text", + "content": "[1] Pablo Fernández Alcantarilla, Adrien Bartoli, and Andrew J Davison. Kaze features. In Proceedings of European Conference on Computer Vision, pages 214-227, 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 327 + ], + "type": "text", + "content": "[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 372 + ], + "type": "text", + "content": "[3] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2911-2918, 2012. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 374, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 288, + 407 + ], + "type": "text", + "content": "[4] Daniel Barath and Jií Matas. Graph-cut ransac. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6733-6741, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 462 + ], + "type": "text", + "content": "[5] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1304-1312, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 520 + ], + "type": "text", + "content": "[6] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5836-5844, 2019. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 522, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 287, + 555 + ], + "type": "text", + "content": "[7] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 610 + ], + "type": "text", + "content": "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In Proceedings of the European Conference on Computer Vision, 2022. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 613, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 287, + 677 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly sur" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "face reintegration. ACM Transactions on Graphics, 36(4):1, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 161 + ], + "type": "text", + "content": "[11] Luanyuan Dai, Yizhang Liu, Jiayi Ma, Lifang Wei, Taotao Lai, Changcai Yang, and Riqing Chen. Ms2dg-net: Progressive correspondence learning via multiple sparse semantics dynamic graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8973-8982, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 224-236, 2018. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "type": "text", + "content": "[13] M.P. do Carmo. Differential Geometry of Curves and Surfaces: Revised and Updated Second Edition. Dover Books on Mathematics. Dover Publications, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "text", + "content": "[14] Siyan Dong, Shuzhe Wang, Yixin Zhuang, Juho Kannala, Marc Pollefeys, and Baoquan Chen. Visual localization via few-shot scene region classification. International Conference on 3D Vision, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "type": "text", + "content": "[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 420 + ], + "type": "text", + "content": "[16] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "type": "text", + "content": "[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 467, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 499 + ], + "type": "text", + "content": "[18] Per-Erik Forssén and David G Lowe. Shape descriptors for maximally stable extremal regions. In International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[19] Khang Truong Giang, Soohwan Song, and Sungho Jo. Curvature-guided dynamic scale networks for multi-view stereo. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[20] Jack Goldfeather and Victoria Interrante. A novel cubic-order algorithm for approximating principal direction vectors. ACM Trans. Graph., 23(1):45-63, 2004. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 580, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 611 + ], + "type": "text", + "content": "[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10-5244. CiteSeer, 1988. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[23] Ji Hou, Saining Xie, Benjamin Graham, Angela Dai, and Matthias Nießner. Pri3d: Can 3d priors help 2d representation learning? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5693-5702, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "17989" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[24] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "text", + "content": "[25] Qingde Li and J.G. Griffiths. Least squares ellipsoid specific fitting. In Geometric Modeling and Processing, 2004. Proceedings, pages 335-340, 2004. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 195 + ], + "type": "text", + "content": "[26] Xinghui Li, Kai Han, Shuda Li, and Victor Prisacariu. Dualresolution correspondence networks. Advances in Neural Information Processing Systems, 33:17346-17357, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 239 + ], + "type": "text", + "content": "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 350 + ], + "type": "text", + "content": "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 393 + ], + "type": "text", + "content": "[30] Xiaotao Liu, Chen Meng, Fei-Peng Tian, and Wei Feng. Dgd-net: Local descriptor guided keypoint detection network. In 2021 IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "text", + "content": "[31] Yuan Liu, Lingjie Liu, Cheng Lin, Zhen Dong, and Wenping Wang. Learnable motion coherence for correspondence pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "text", + "content": "[32] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 514 + ], + "type": "text", + "content": "[33] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "text", + "content": "[34] Jiri Matas, Ondrej Chum, Martin Urban, and Tomás Pajdla. Robust wide-baseline stereo from maximally stable extremal regions. Image and vision computing, 22(10):761-767, 2004. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 592 + ], + "type": "text", + "content": "[35] Krystian Mokolajczyk and Cordelia Schmid. Scale & affine invariant interest point detectors. International Journal of Computer Vision, 60(1):63-86, 2004. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "type": "text", + "content": "[36] Krystian Mikolajczyk, Tinne Tuytelaars, Cordelia Schmid, Andrew Zisserman, Jiri Matas, Frederik Schaffalitzky, Timor Kadir, and L Van Gool. A comparison of affine region detectors. International Journal of Computer Vision, 65(1):43-72, 2005. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[37] Dmytro Mishkin, Filip Radenovic, and Jiri Matas. Repeatability is not enough: Learning affine regions via discriminability. In Proceedings of the European Conference on Computer Vision, pages 284-300, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[39] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[40] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3456-3465, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 196, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 242 + ], + "type": "text", + "content": "[41] Rémi Pautrat, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Online invariance selection for local feature descriptors. In European Conference on Computer Vision, pages 707-724. Springer, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 242, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 286 + ], + "type": "text", + "content": "[42] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "text", + "content": "[43] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 388 + ], + "type": "text", + "content": "[44] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 388, + 545, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 433 + ], + "type": "text", + "content": "[45] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. Efficient neighbourhood consensus networks via submanifold sparse convolutions. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 478 + ], + "type": "text", + "content": "[46] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 478, + 545, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 512 + ], + "type": "text", + "content": "[47] John Rugis and Reinhard Klette. A scale invariant surface curvature estimator. In Pacific-Rim Symposium on Image and Video Technology, pages 138-147. Springer, 2006. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 567 + ], + "type": "text", + "content": "[48] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 624 + ], + "type": "text", + "content": "[49] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "type": "text", + "content": "[50] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In European conference on computer vision, pages 752-765. Springer, 2012. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[51] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2017. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "17990" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "text", + "content": "[52] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6dof outdoor visual localization in changing conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 1, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 150, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 184 + ], + "type": "text", + "content": "[53] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 1, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "text", + "content": "[54] Yan Shi, Jun-Xiong Cai, Yoli Shavit, Tai-Jiang Mu, Wensen Feng, and Kai Zhang. Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12517-12526, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "text", + "content": "[55] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 352 + ], + "type": "text", + "content": "[56] Weiwei Sun, Wei Jiang, Eduard Trulls, Andrea Tagliasacchi, and Kwang Moo Yi. Acne: Attentive context normalization for robust permutation-equivariant learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11286-11295, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 353, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 408 + ], + "type": "text", + "content": "[57] Dongli Tan, Jiang-Jiang Liu, Xingyu Chen, Chao Chen, Ruixin Zhang, Yunhang Shen, Shouhong Ding, and Rongrong Ji. Eco-tr: Efficient correspondences finding via coarse-to-fine refinement. Proceedings of the European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 409, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 442 + ], + "type": "text", + "content": "[58] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In International Conference on Learning Representations, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 443, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 287, + 487 + ], + "type": "text", + "content": "[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 542 + ], + "type": "text", + "content": "[60] Yurun Tian, Vassileios Balntas, Tony Ng, Axel BarrosoLaguna, Yiannis Demiris, and Krystian Mikolajczyk. D2d: Keypoint extraction with describe to detect approach. In Proceedings of the Asian Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "type": "text", + "content": "[61] Carl Toft, Daniyar Turmukhambetov, Torsten Sattler, Fredrik Kahl, and Gabriel J Brostow. Single-image depth prediction makes feature matching easier. In European Conference on Computer Vision, pages 473-492. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[62] Lijing Tong and Xiang Ying. 3d point cloud initial registration using surface curvature and surf matching. 3D Research, 9:1-16, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[63] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 585 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 117 + ], + "type": "text", + "content": "[65] Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching. In Proceedings of the Asian Conference on Computer Vision, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[66] Shuzhe Wang, Zakaria Laskar, Iaroslav Melekhov, Xiaotian Li, and Juho Kannala. Continual learning for image-based camera localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3252-3262, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "type": "text", + "content": "[67] Changchang Wu. Towards linear-time incremental structure from motion. In International Conference on 3D Vision, pages 127-134. IEEE, 2013. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "text", + "content": "[68] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 262, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 306 + ], + "type": "text", + "content": "[69] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "text", + "content": "[70] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666-2674, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 363, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 545, + 418 + ], + "type": "text", + "content": "[71] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 419, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 484 + ], + "type": "text", + "content": "[72] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Hongen Liao. Learning two-view correspondences and geometry using order-aware network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5845-5854, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "text", + "content": "[73] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "type": "text", + "content": "[74] Chen Zhao, Yixiao Ge, Feng Zhu, Rui Zhao, Hongsheng Li, and Mathieu Salzmann. Progressive correspondence pruning by consensus learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6464-6473, 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "17991" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_content_list.json b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2c36ae399474036ef0ab3aa6c478f31e891dc2bd --- /dev/null +++ b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_content_list.json @@ -0,0 +1,1768 @@ +[ + { + "type": "text", + "text": "H3WB: Human3.6M 3D WholeBody Dataset and Benchmark", + "text_level": 1, + "bbox": [ + 174, + 130, + 795, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yue Zhu Nermin Samet David Picard LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France {yue.zhu, nermin.samet, david.picard}@enpc.fr Code and dataset: https://github.com/wholebody3d/wholebody3d", + "bbox": [ + 176, + 181, + 792, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a benchmark for 3D human whole-body pose estimation, which involves identifying accurate 3D keypoints on the entire human body, including face, hands, body, and feet. Currently, the lack of a fully annotated and accurate 3D whole-body dataset results in deep networks being trained separately on specific body parts, which are combined during inference. Or they rely on pseudogroundtruth provided by parametric body models which are not as accurate as detection based methods. To overcome these issues, we introduce the Human3.6M 3D WholeBody (H3WB) dataset, which provides whole-body annotations for the Human3.6M dataset using the COCO Wholebody layout. H3WB comprises 133 whole-body keypoint annotations on 100K images, made possible by our new multiview pipeline. We also propose three tasks: i) 3D whole-body pose lifting from 2D complete whole-body pose, ii) 3D whole-body pose lifting from 2D incomplete whole-body pose, and iii) 3D whole-body pose estimation from a single RGB image. Additionally, we report several baselines from popular methods for these tasks. Furthermore, we also provide automated 3D whole-body annotations of TotalCapture and experimentally show that when used with H3WB it helps to improve the performance.", + "bbox": [ + 75, + 316, + 472, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 693, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Human pose estimation is the task of localizing human body keypoints in images which is critical to analyze human behavior, expressions, emotions, intentions, and how people communicate and interact with the physical world. As a result, 3D human pose estimation has an important role in several vision tasks and applications such as robotics [28, 25, 29] or augmented/virtual reality [55, 3, 73, 81]. However, to make more accurate predictions about human behaviors, we need more than a few body keypoints. To that end, 3D whole-body pose estimation aims to detect face, hand and foot keypoints in addition to the standard human body keypoints of classical 3D hu", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "man pose estimation.", + "bbox": [ + 500, + 287, + 642, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The lack of accurate 3D datasets has made 3D whole-body pose estimation a challenging task, leading previous works to focus on separate body parts and train separate models on different datasets for 3D body pose [1, 30, 55, 56, 63, 3, 44, 45, 46, 65, 74, 77, 80, 26], 3D hand pose [8, 58, 84, 7, 27, 37, 82, 31], or 3D face landmarks [68, 9, 13]. However, directly ensembling separate body part models during inference suffers from issues arising from datasets' biases, pose and scales, and complex inference pipelines. Distillation from pretrained models has been used to overcome these issues, with FrankMocap [66] using three specialized pretrained models to estimate 65 whole-body keypoints (22 on the body, 40 on the hands, and 3 on the face), and DOPE [75] also using three specialized models to output 139 whole-body keypoints (13 on the body, 42 on the hands, and 84 on the face).", + "bbox": [ + 496, + 305, + 893, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alternatively, parametric body models can be fitted to obtain whole-body pose, as has been proposed in Ex-Pose [18], SMPLify-X [60] or Monocular Total Capture (MTC) [77]. While parametric models enable sampling an almost infinite number of keypoints from the mesh [24, 22, 23], their accuracy is usually less than that of detection based methods on fine body parts like hands and feet (see supplementary material for examples from the literature). Indeed, parametric models are tailored for visual applications such as realistic motion generation [62] or avatar capture [67], where a realistic capture of the body shape and pose is more important than fine grain accuracy. Relying on a small set of data-driven pose parameters ensures realism but also limits their flexibility in representing complex unusual poses. For applications where the body shape is not needed but the accuracy of the keypoints is essential, like high performance sport analysis or ergonomics, detection based methods are thus the preferred solution.", + "bbox": [ + 496, + 550, + 895, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, 3D whole-body pose estimation has not been fully explored in the literature due to the absence of a representative and accurate benchmark. As previously mentioned, existing 3D whole-body methods either rely on specific datasets and models for different body parts, lead-", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20166", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25f628b325ee5dc9efdbfe6a0f0299a05223c6d2c26ff5de4853d5ee48af82d5.jpg", + "image_caption": [ + "Figure 1. The H3WB dataset has 133 whole-body keypoint annotations in 3D as well as their respective projections in 2D." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 854, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing to complex training pipelines and heterogeneous evaluations, or utilize parametric models that prioritize shape capture over highly precise keypoints. In addition, unified methods vary significantly in terms of keypoint layout definition, number of keypoints and distribution of keypoints across body parts (see Table 1). These significant dataset disparities and the absence of a standard benchmark make it challenging to compare methods fairly.", + "bbox": [ + 75, + 238, + 472, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose a new large-scale dataset for accurate 3D whole-body pose estimation called Human3.6M 3D WholeBody, or H3WB for short (see Figure 1). Our dataset extends Human3.6M [36, 12] with 3D whole-body keypoint annotations. It consists of 133 paired 2D and 3D whole-body keypoint annotations for a set of 100k images from Human3.6M, following the same layout used in COCO WholeBody [40]. More specifically, in addition to the standard 17 body keypoints, the dataset has 42 hand keypoints, 6 foot keypoints and 68 facial landmarks. H3WB was automatically created in a 3 step process: We obtain an initial set of 3D annotations using multiview geometry. Then, we trained a masked auto-encoder to complete the initial annotations. Finally, we refine the whole-body keypoints via a diffusion model. A manual annotation of 80K keypoints from 600 images shows our labels have an average error of $17\\mathrm{mm}$ which suggest the H3WB keypoints are very accurate for such a complex task. We propose 3 tasks and benchmarks on the H3WB dataset for which we provide baselines: i) 3D whole-body pose lifting from a complete 2D whole-body keypoints, ii) 3D whole-body pose lifting from incomplete 2D whole-body keypoints (i.e. 2D whole-body with missing keypoints, which is more realistic), and iii) 3D whole-body pose estimation from a single RGB image.", + "bbox": [ + 75, + 364, + 472, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows. 1) We propose a method to create detailed 3D human pose keypoints from multi-view images. 2) We propose H3WB, the first accurate public benchmark dataset for 3D whole-body pose estimation, using the aforementioned method. Our benchmark can easily leverage existing results in 2D and enables the community to build upon existing high-quality 2D detectors on COCO. Unifying the 3D whole-body pose estimation with the COCO 2D benchmark will greatly benefit the research community. 3) We provide baselines for", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the 3 tasks of H3WB, which we believe will encourage the community to explore 3D whole-body pose estimation more and accelerate progress in the field. 4) Additionally, we provide 3D whole-body annotations for the TotalCapture [43] dataset, and show that when combined with the H3WB dataset it improves the performance of pose lifting tasks.", + "bbox": [ + 496, + 238, + 893, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 498, + 359, + 638, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Body, hand and face pose estimation. There are two main groups of prominent approaches in 3D human pose estimation. The first group directly estimates 3D body pose from a single RGB image [61, 54, 55, 56, 61, 63]. The second group follows two stage approach where they first localize 2D keypoints and then lift 2D human pose to 3D space [53, 72, 38, 6, 49, 53]. Several optimization based methods [6, 49], utilize 2D keypoints to initialize a parametric model of the human body such as SMPL [51]. Several works attempt to eliminate the requirement of 3D annotations using 2D multi-view supervision to estimate 3D human pose [72, 78] or temporal supervision with video [17, 50]. 3D hand pose estimation methods share similar approaches as the body counterparts. First group of works, estimates hand pose from a single RGB image by directly regressing 3D hand keypoints [79], mesh vertices [27, 48], and parameters of parametric 3D hand models such as MANO [7, 64, 2, 14, 15, 82]. Second groups of works rely on intermediate 2D representations such as 2D keypoints and feature maps [8, 58, 82, 84, 34]. Simi", + "bbox": [ + 496, + 383, + 893, + 688 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/3a00357557fb72d26244e78c35fcd107f0ca93504783f86994eef5016b577540.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSizeKeypointsBodyHandFace
Human3.6M[36]3.6M1717
3DPW[71]51k2424
LSP[41]10k1414
3DHP[71]>1.3M1717
Panoptic[42]1.5M1515
MTC[77]834K2020
InterHand2.6M[57]2.6M2121
FreiHAND[85]37k2121
RHD[84]44K2121
MTC[77]111K2121
TotalCapture[43]1.9M1272116+1674
ExPose[18]33K1442515+1589
H3WB100k1332321+2168
", + "bbox": [ + 511, + 709, + 880, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Overview of datasets for 3D human pose estimation.", + "bbox": [ + 500, + 885, + 864, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20167", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "larly, predominant 3D face pose estimation methods regress the dense 3D face landmarks [19, 21, 39] and face model parameters [13, 68, 69, 20, 76] based on 3DMM [5].", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Whole-body pose estimation. There are a several methods [60, 77, 83, 18, 75, 66] jointly estimating 3D whole-body pose. The first group of works is based on parametric human body models such as Adam [43] and SMPL-X [60]. MTC [77] is based on the Adam model [43], and first gets 2.5D predictions, then optimizes the parameters of Adam. SMPLify-X optimizes the parameter of the SMPL-X model [60] to fit it to 2D keypoints. As a major drawback, optimization-based methods are relatively slow and highly sensitive to parameter initializations. Nonparametric methods [75, 18, 66] follow different approaches to avoid heavy optimization procedure. DOPE [75] and FrankMocap [66] first train separate body, hand, and face models. Next, they combine those models within a learning framework. DOPE [75] curates pseudo-ground truths from separate body models and uses those ground-truths to supervise the distillation model. Similar to DOPE, ExPose [18] first obtains a pseudo-ground truth dataset by fitting SMPL-X model on in-the-wild images, and trains a joint model to output whole-body poses. All these methods utilize many part-based datasets. Moreover, all output different whole-body layouts with a different numbers of whole-body keypoints. FrankMocap, DOPE, and SMPLify-X estimate whole-body pose with 65, 139 and 144 keypoints, respectively.", + "bbox": [ + 75, + 137, + 472, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose completion completes a partially estimated pose by localizing missing keypoints. Carissimi et al. [10] propose a denoising variational autoencoder network to fill the missing keypoints in 2D pose completion. Bautembach et al. [4] selects a small subset of poses from a database based on their distance to an incomplete 3D pose, and replaces the missing keypoints with the corresponding averaged keypoints in the subset. Despite being critical for real-world scenarios, pose completion has not been sufficiently explored due to the lack of annotated keypoint datasets. Our 3D whole-body dataset can facilitate more exploration of 3D pose estimation from 2D incomplete human poses.", + "bbox": [ + 75, + 515, + 472, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The H3WB dataset", + "text_level": 1, + "bbox": [ + 76, + 712, + 264, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the making of the H3WB dataset1. Our objective is to build a keypoint based 3D whole-body dataset including keypoints on the body, the face and the hands, and propose a benchmark. We use the same keypoint layout as COCO WholeBody dataset [40] with 133 keypoints. To that end, we build on the widely used Human3.6M dataset [36] for which we provide 3D whole-body keypoints. The H3WB building process is as follows: First, we use an off-the-shelf 2D whole-body de", + "bbox": [ + 75, + 738, + 468, + 875 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b562750440474c9e446610dc3cd872e9bfdbfa07547960d199c400c847c61e4.jpg", + "image_caption": [ + "Figure 2. OpenPifPaf detects most of the non-occluded keypoints inside the image (orange keypoints). The occluded or undetected keypoints (cyan keypoints) are reprojections after 3D multi-view reconstruction. Notice that these reprojections do not always align with the images, like the right hand in the last view, which is probably due to OpenPifPaf not being perfectly accurate." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 893, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tector combined with multi-view reconstruction to obtain an initial set of incomplete 3D whole-body keypoints. Next, we implement a completion network to fill in the keypoints missed by the multi-view geometric approach. Then, we develop a refinement method for the hands and the face to obtain more accurate keypoints. Finally, we perform quality assessment to select 25k 3D whole-body poses with high confidence and the 100k associated images from 4-view.", + "bbox": [ + 498, + 342, + 893, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Initial 3D whole-body dataset with OpenPifPaf", + "text_level": 1, + "bbox": [ + 500, + 484, + 893, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We run the 2D whole-body detector from OpenPif-Paf [47] on all the 4 views from the training set of Human3.6M (S1, S5, S6, S7 and S8, 1 image per 5 frames). Since the cameras of Human3.6M are well calibrated, we can reconstruct keypoints in 3D using standard multi-view geometry.", + "bbox": [ + 498, + 513, + 893, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The OpenPifPaf 2D whole-body detector can miss keypoints due to self-occlusions (hands, feet) or unfavorable camera viewpoints (facial landmarks). However, the four-view setup allows us to recover missing keypoints and obtain a complete 3D whole-body pose, provided each keypoint appears in at least two non-opposing views. An example of this process is shown in Figure 2. Using this method, we obtained 11,426 fully complete 3D whole-body poses with all 133 keypoints and 26,333 incomplete 3D whole-body poses where all keypoints appear in at least one view, resulting in a total of 37,759 3D whole-body poses with each keypoint appearing in at least one view.", + "bbox": [ + 496, + 608, + 893, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We did not rely on the video information because the reconstruction problem becomes significantly more difficult when there is a motion between two frames. In the absence of motion, an additional frame does not help solve the occlusion problem. The results of our study demonstrate that multi-view labeling is sufficiently effective for our task (see Table 2, \"Geometry\" line).", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We consider the feet keypoints as a part of the body.", + "bbox": [ + 94, + 886, + 377, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20168", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c90ee42565437b0f8f4c0ae560ef5e062d578337d9277b28f83963ef941ea4d7.jpg", + "image_caption": [ + "Figure 3. The completion network consists of one linear input layer, 4 transformer encoder blocks (each of them containing 2 transformer encoder layer with $d\\_model = 64$ and $n\\_head = 1$ ), and a linear output layer. At the end of each encoder block, the features are decoded by the output layer into a predicted position in a curriculum way where later blocks decode more keypoints." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 472, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59eb85bf0820086b5487f607946a9d30b6fba0d4af33000117121f833f2ea449.jpg", + "image_caption": [ + "Figure 4. Example outputs of the completion network. The orange color denotes the keypoints that were detected by OpenPifPaf. The cyan color shows the missed keypoints by OpenPifPaf but completed by our completion network. The left hand is detected in only 1 view by OpenPifPaf and thus fully predicted by the completion network." + ], + "image_footnote": [], + "bbox": [ + 107, + 309, + 439, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Completion network", + "text_level": 1, + "bbox": [ + 76, + 607, + 272, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to complete the 26,333 incomplete 3D whole-body poses, we develop a completion network as shown in Figure 3. We designed our completion network using Transformer architecture [70] as they can easily handle the conditional dependencies introduced by the skeleton's topology through masking. Since each skeleton always has exactly 133 keypoints, which can be considered as 133 tokens of 3 coordinate values. Token values are expanded from 3 coordinates to $3 \\times 16 = 48$ features using Fourier encoding. We use learnable positional encoding since each keypoint is uniquely identified.", + "bbox": [ + 75, + 631, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the completion network on the 11,426 complete skeletons using a masked auto-encoder strategy [32] where the missing keypoints are masked at the input and will be predicted using the unmasked keypoints. The masking strategy is as follows:", + "bbox": [ + 75, + 797, + 470, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- With a $50\\%$ chance, we perform a keypoint wise mask", + "bbox": [ + 94, + 885, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each keypoint has $15\\%$ chance of being masked,", + "bbox": [ + 531, + 90, + 890, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- with the remaining $50\\%$ chance, we perform a block wise mask in which either the body, the left hand, the right hand, the left or the right part of the face are masked (uniform probability).", + "bbox": [ + 517, + 109, + 893, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ease the learning process and take into account the causal link between some keypoints (e.g., the tip of a finger depends on the position of its parent phalanges), we introduce a curriculum approach. We compute the loss at different levels following a hierarchy where early levels consider only keypoints closer to the root, while later levels consider more deformable keypoints which highly depend on their parents. We illustrate the completion network and learning process in Figure 3. The loss function is", + "bbox": [ + 496, + 178, + 893, + 315 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (X, X _ {g t 3 D}, X _ {g t 2 D}) = \\mathcal {L} _ {3 D} (X, X _ {g t 3 D}) \\\\ + \\alpha \\mathcal {L} _ {2 D} (X, X _ {g t 2 D}) \\\\ + \\beta \\mathcal {L} _ {\\text {s y m}} (X), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 321, + 890, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_3D$ is an $\\ell_1$ loss of 3D coordinates, $\\mathcal{L}_2D$ is an $\\ell_1$ loss of 2D projection of the 3D coordinates if we have the 2D annotation from OpenPifPaf, and $\\mathcal{L}_{sym}$ is a symmetric loss which is applied to make sure the left part and right part of the human have the same length on corresponding body parts.", + "bbox": [ + 496, + 383, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show an example output from our completion network in Figure 4. The completion network results on missing body parts are visually realistic and appealing. However, since the completion network does not rely on the image content, its output does not always align with the image", + "bbox": [ + 496, + 474, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8186813d0a00cae85898e215ffbacc269bbbcdea2a45ec07e5c6e5f4ec3d60f0.jpg", + "image_caption": [ + "Figure 5. Refinement network architecture and training process. Gaussian noise is added to the groundtruth coordinates with increasing variance, and the network is iteratively trained to recover the less noisy coordinates." + ], + "image_footnote": [], + "bbox": [ + 508, + 569, + 888, + 665 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/812730bc15fa31542f2333210c7659bcb5a9382badfea8201d0269a29593f2ba.jpg", + "image_caption": [ + "Figure 6. Example outputs from the face (top row) and hand (bottom row) refinement networks during inference time. We observe that the predictions almost converge to the correct locations in 5-iteration." + ], + "image_footnote": [], + "bbox": [ + 509, + 742, + 883, + 837 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20169", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1aa1b698d6d8c39350d1eb21896b1f59fc1cbc5393b505b32270bf5748f6187c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 93, + 272, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef90dd24ea822845f670e725b88db5893cb7c52ac2f8cb6ac7e467297780f1dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 93, + 478, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66b3ef0869e0545bddbc02ef2b55c8a4e7f086ca07a5de673ee934dfc5edb0d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 93, + 684, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b701b4be651867ef5f4300293d84dfef82f6f2d943b1a092ea460c297677502f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 94, + 888, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b03a7dcff89ef6a2abdac401749abfb4723c609361f6f61937a983d057305c18.jpg", + "image_caption": [ + "Figure 7. Examples of the 3D whole body skeleton. They are visually realistic humans. The strange looking faces (fatter or thinner) in different views are due to viewing artifacts of the default perspective projection." + ], + "image_footnote": [], + "bbox": [ + 83, + 213, + 272, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65c139ca9019b262e7a8a5c628490692d408503841d2ac2d9f886ab79f85dc6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 213, + 478, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9689e8d23e4ef3256860f35c41edd36fb4bc051cb754c112c981b0401121b9e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 213, + 679, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5d16a1484e229f1c5fb37d83880c8129e01f6117cabaa6a05025b7e1bc0fcf6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 213, + 887, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and may only reflect the most common poses of the training set. This can quantitatively be seen in the line “+ Completion” of Table 2.", + "bbox": [ + 75, + 398, + 470, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Hands and face 2D refinements", + "text_level": 1, + "bbox": [ + 76, + 453, + 352, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to correct the alignment problem, we propose another neural network that refines the 2D position of keypoints on the face and the hands. Previous studies have explored and demonstrated the effectiveness of 2D human pose refinement using an iterative error feedback framework [11]. Motivated by this, we build upon recent conditional diffusion models [35] and we consider the prediction from the completion network as noisy such that the refinement network denotes it to conditionally fit the image.", + "bbox": [ + 75, + 478, + 468, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train separate refinement models for the face and the hands, while keeping the same network architecture and the same training strategy. We used a simple MLP and found it to be effective, preventing the need to explore more complex architectures. We illustrate the refinement process in Figure 5. During training, we add Gaussian noise to the groundtruth poses with an increasing variance from 5 to 25 pixels, and annotate them as step $t = 1 \\dots 5$ (step $t = 0$ is the groundtruth). The network learns to predict the pose at step $t$ given the image and the noisier step $t + 1$ with a 2D supervision loss.", + "bbox": [ + 75, + 614, + 470, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We build two small datasets, each consisting of 22,000 non-occluded faces and hands respectively, with their corresponding OpenPifPaf predictions. Each image is resized to $384 \\times 384$ pixels. We use a random crop of size $224 \\times 224$ pixels to have the face and hands located in diverse regions of the images. We split the datasets into training and validation sets with 20,000 images and 2,000 images, respectively.", + "bbox": [ + 75, + 779, + 470, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitatively, the face predictions achieve an average", + "bbox": [ + 96, + 886, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "error less than 3 pixels and the hand predictions achieve an average error less than 7 pixels on the validation sets. We show example qualitative results in Figure 6.", + "bbox": [ + 496, + 398, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we run the refinement networks on the 2D-projections of the 3D poses predicted by our completion network. For each 3D skeleton, we project it into the 4 different 2D views. We then crop the regions around the hands and face and denoise the corresponding predictions using the refinement network with 10 iterations to obtain refined 2D poses in each of the 4 views.", + "bbox": [ + 496, + 445, + 892, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the refinement network is not always correct due to its training on non-occluded faces or hands, we only need 2 non-opposing views to perform geometric reconstruction. Since bad refinements tend to collapse all keypoints into the same location, we select the two non-opposing views with the highest variance in keypoint positions to avoid disruptions caused by occlusions. Using this method, we obtain 151,036 triplets of 3D whole-body keypoints, corresponding image, and 2D projected keypoints from the original set. Examples of resulting 3D whole-body skeletons and their image-aligned 2D counterparts are shown in Figure 7 and Figure 8, respectively.", + "bbox": [ + 496, + 550, + 893, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Quality assessment", + "text_level": 1, + "bbox": [ + 500, + 741, + 684, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To select the most accurate triplets from our dataset, we reuse the refinement networks and employ a multi-crop strategy that accounts for the variance of the prediction. We project each 3D whole-body skeleton onto all 4 views, and produce four cropped images for each region of interest around the face and hands. The refinement network is run on these 4 crops, and the resulting predictions are aligned with the original prediction to compute the 2D error compared to the original 2D projection. We score the", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88e3a0895ec6b6bcc4d8a3794619fdf511795d5a54b52be6203abdf53c0ac0cc.jpg", + "image_caption": [ + "Figure 8. Examples of the 3D whole-body skeleton projected in 2D onto their corresponding images. They are visually accurate, though still there are small errors in detail which we do not expect to overcome due to the initial resolution and ambiguity of the images." + ], + "image_footnote": [], + "bbox": [ + 155, + 87, + 816, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D skeletons by averaging the errors of all 4 projected views, and select the 5k lowest error skeletons from each subject of Human3.6M (S1, S5, S6, S7, S8) to form the $5\\mathrm{k}\\times 4(\\mathrm{view})\\times 5(\\mathrm{subject}) = 100\\mathrm{k}$ triplets of {image, 2D coordinates, 3D coordinates in camera space} of our 3D whole-body dataset.", + "bbox": [ + 75, + 382, + 473, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e396a735b6833949744e58a7cc5d5daebac14a18b032ff4d143d0b9db6928242.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 488, + 272, + 594 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba06ddb24d0ebd65ec0c05339fc52ccf3143079f80f78442cc13c12df00bcdfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 488, + 460, + 594 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f77eac3dfb9a2fbe6bb732f41a339861a14e9e70e6f875f299ac31f2022e2339.jpg", + "image_caption": [ + "Figure 9. 3D error distributions calculated from 80k manually corrected keypoint annotations. 3D error distributions are presented for whole-body, body, hand and face in mm. We observe that 3D errors are mostly concentrated between 10mm and 20mm." + ], + "image_footnote": [], + "bbox": [ + 83, + 595, + 271, + 703 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4d4cf203561e59006798740cb78aa13631d658369392a1831ad2bf9ec668bdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 595, + 460, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the quality of the H3WB dataset, we conducted a cross-check study on 600 randomly selected images from the dataset. In this study, annotators were presented an image with the 2D projection of the 3D skeleton on top and were asked to manually correct mis-aligned keypoints by drag and drop. Using multi-view geometry, we reconstructed these corrected skeletons in 3D and compared them to our original skeletons. To validate our process, we show", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a728589516765baee46eb5894551e626ebdb3108306cee363a860b425c6c66e9.jpg", + "image_caption": [ + "Figure 10. Per keypoints error statistics. Please zoom in." + ], + "image_footnote": [], + "bbox": [ + 506, + 380, + 888, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the influence of each step in Table 2. The geometric approach produced good results but unfortunately cannot provide a large enough dataset. The completion step allows to obtain all labels but at the cost of degraded accuracy due to lack of alignment as explained in section 3.2. The diffusion recovers the original accuracy of the geometric approach. $2\\mathrm{mm}$ difference is irrelevant given the initial resolution of the images. We obtain a final average error of $17\\mathrm{mm}$ which is very accurate for such a difficult task, and leads to a benchmark which we believe will not be saturated until methods reach around $35\\mathrm{mm}$ .", + "bbox": [ + 496, + 604, + 893, + 771 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24f158fe78521b3f71b489bddee65ef86048a2caf4d2be9e721bdb8ef3d847f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Steps# keypoints available3D error (mm)
AllBodyFaceHand
Geometry4812714.8717.7213.2915.87
+ Completion7980029.3125.5726.0236.67
+ Diffusion7980016.9818.6315.0819.16
", + "bbox": [ + 501, + 780, + 890, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative analysis of each intermediate step in our pipeline.", + "bbox": [ + 500, + 840, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D error distributions: In Figure 9, we illustrate 3D", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20171", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "error distributions in mm for all whole-body keypoints and each whole-body part keypoints (i.e. body, hand and face) separately. The distributions of errors are well concentrated around low values.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Per keypoint errors: All per keypoint error statistics are shown in Figure 10. $98.3\\%$ of the images are below 5cm error (97.4% for body joints, 99.7% hands, and 96.4% face). Similarly, $82.6\\%$ of the images are below 2cm error (75.5% for body joints, 89.2% hands, and 75.8% face).", + "bbox": [ + 75, + 151, + 467, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5. TotalCapture 3D WholeBody Dataset", + "text_level": 1, + "bbox": [ + 76, + 237, + 398, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to H3WB dataset, we also prepare whole-body annotations for training sequence of poses from TotalCapture [43] dataset using our proposed multi-view pipeline. We call this dataset TotalCapture 3D WholeBody, or T3WB for short. To create TotalCapture 3D WholeBody, we first obtain 2D whole-body keypoints from OpenPif-Paf [47]. At this stage we discard the frames without any human. Next, we finetune the completion network initialized by H3WB weights using 7000 samples of TotalCapture with complete 8-views. At the end we obtain 125,960 triplets of 3D whole-body keypoint, corresponding image and 2D projected keypoints. This shows our pipeline can be used with any multiview dataset.", + "bbox": [ + 75, + 261, + 467, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We do not conduct a quality assessment study for T3WB and therefore we cannot guarantee the precision of the annotations. Instead, we use T3WB together with H3WB for training models. To alleviate the effect of the noisy annotations in T3WB, we sample more from H3WB than T3WB in each batch. More specifically, we follow 4:1 ratio for each batch during $H3WB + T3WB$ trainings.", + "bbox": [ + 75, + 457, + 467, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. The H3WB benchmark", + "text_level": 1, + "bbox": [ + 76, + 577, + 295, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the H3WB dataset to propose a benchmark and the associated leaderboard. We split the dataset into training and test sets. The training set contains all samples from S1, S5, S6 and S7, including 80k {image,2D,3D} triplets. The test set contains all samples from S8, including 20k triplets. The test set labels are retained to prevent involuntary overfitting on the test set. Evaluation is accessible only by submitting results to the maintainers. We do not provide a validation set. We encourage researchers to report 5-fold cross-validation average and standard deviation (see supplementary).", + "bbox": [ + 75, + 603, + 467, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The corresponding benchmark has 3 different tasks:", + "bbox": [ + 96, + 770, + 437, + 784 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. 3D whole-body lifting from complete 2D whole-body skeletons, or $2D \\rightarrow 3D$ for short.", + "2. 3D whole-body lifting from incomplete 2D whole-body skeletons, or $I2D \\rightarrow 3D$ for short.", + "3. 3D whole-body skeleton prediction from image, or $RGB \\rightarrow 3D$ for short." + ], + "bbox": [ + 88, + 797, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each task, we report the following MPJPE (Mean Per Joint Position Error) metrics:", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MPJPE for the whole-body, the body (keypoint 1-23), the face (keypoint 24-91) and the hands (keypoint 92-133) when whole-body is centered on the root joint, i.e. aligned with the pelvis, which in our case is the middle of two hip joints2,", + "- MPJPE for the face when it is centered on the nose, i.e. aligned with keypoint 1,", + "- MPJPE for the hands when hands are centered on the wrist, i.e left hand aligned with keypoint 92 and right hand aligned with keypoint 113." + ], + "bbox": [ + 517, + 133, + 890, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To create baselines on each task, we adapt popular methods from the literature by changing the number of keypoints to that of our whole-body dataset. Notice that we keep the training recipes of the original papers to avoid over-fitting to this new benchmark. In practice, we recommend to perform model selection and hyper-parameters tuning using 5-fold cross-validation.", + "bbox": [ + 496, + 311, + 890, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. 3D whole-body lifting from complete 2D whole-body keypoints $(2D\\rightarrow 3D)$", + "text_level": 1, + "bbox": [ + 498, + 430, + 888, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This task is similar to the standard 3D human pose estimation from 2D keypoints but using whole-body keypoints. The training set contains 80k 2D-3D pairs. The test set contains only a half of all the test samples, i.e. 10k 2D poses3.", + "bbox": [ + 496, + 469, + 888, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate 6 methods on this task. SimpleBaseline [53] is a well-established model, consisting of a 6-layer MLP. We propose a modification, replacing the network architecture with an 8-layer MLP, which we call Large Simple-Baseline inspired by CanonPose [72]. Jointformer [52] is a recent transformer-based method. CanonPose is trained only with 2D supervision [72]. We also adapt CanonPose to work with additional 3D supervision by manually creating 3 fixed camera views and rotating the 3D skeletons into the corresponding view before projecting them into 2D, training it with multi-view weak-supervision. Finally, we report results for the parametric model SMPLify-X [60] by running optimizations on each input sample.", + "bbox": [ + 496, + 531, + 890, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train SimpleBaseline models using their official training setting as described in [53]. The inputs and targets are normalized by subtracting the mean and dividing by the standard deviation. Similarly, we train CanonPose [72] models following their official training setup where the inputs and targets are centered on the pelvis and scaled by the Forbenius norm. We train the Jointformer model in the two stages as described in [52].", + "bbox": [ + 496, + 728, + 890, + 849 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2We provide the whole-body keypoint ids in supplementary material.", + "bbox": [ + 514, + 862, + 880, + 875 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3The other half is reserved for the task I2D $\\rightarrow$ 3D to prevent access to the missing keypoints.", + "bbox": [ + 501, + 875, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20172", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9450472f4e46be61b2bd59dcabd51e778b13310f52527cd201bbbac171885d69.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Comparing different methods for $2\\mathrm{D}\\rightarrow 3\\mathrm{D}$ on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with \\* output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except $\\dagger$ and $\\ddagger$ show nose and wrist aligned results for face and hands, respectively. Sv. is supervision." + ], + "table_body": "
MethodAllBodyFace / aligned†Hand / aligned‡
H3WB
SMPL-X[60]188.9166.0208.3 / 23.7170.2 / 44.4
CanonPose[72]*186.7193.7188.4 / 24.6180.2 / 48.9
SimpleBaseline [53]*125.4125.7115.9 / 24.6140.7 / 42.5
CanonPose[72] w 3D sv.*117.7117.5112.0 / 17.9126.9 / 38.3
Large SimpleBaseline[53]*112.3112.6110.6 / 14.6114.8 / 31.7
Jointformer[52]88.384.966.5 / 17.8125.3 / 43.7
H3WB + T3WB
CanonPose[72]*164.7161.1174.5 / 21.5150.8 / 43.6
SimpleBaseline [53]*115.3114.8109.4 / 15.8125.1 / 33.5
Jointformer[52]81.578.060.4 / 16.2117.6 / 38.8
", + "bbox": [ + 81, + 88, + 467, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SimpleBaseline and CanonPose models output normalized whole-body keypoints which requires re-scaling at inference. We use statistics from the training set to adjust the test predictions. We calculate a scaling factor using the ratio of 3D to 2D bounding boxes. The formula is: $X_{\\mathrm{final}} = X_{\\mathrm{unit}} \\times \\overline{\\sigma_{3d}} \\times \\frac{\\sigma_{2d}}{\\sigma_{2d}}$ , where $X_{\\mathrm{unit}}$ is the normalized prediction, $\\overline{\\sigma_{3d}}$ is the average size of the 3D training boxes, $\\sigma_{2d}$ is the size of the current 2D box, and $\\overline{\\sigma_{2d}}$ is the average size of the 2D training boxes.", + "bbox": [ + 76, + 349, + 467, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since SMPLify-X has 144 keypoints with a different layout, we use interpolation to transform between the WholeBody skeleton and SMPL-X and run SMPL-X's optimization for 2,000 iterations (4 minutes/sample).", + "bbox": [ + 76, + 493, + 467, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the results in Table 3. SMPLify-X performs the worst, showing that parametric models struggle more than discriminative approaches. SimpleBaseline[53] is a solid method, and Large SimpleBaseline improves its performance further. CanonPose[72] can be improved with additional 3D supervision, but still performs worse than Large SimpleBaseline. CanonPose also predicts the camera view, and the uncertainty in this prediction can lead to more error. Jointformer[52] achieves the best results among all methods, but still has room for improvement. All methods perform worse on our benchmark than on Human3.6M because of pelvis centering, which creates higher numerical error on extremities like hands and face, the parts that contain most of the whole-body keypoints.", + "bbox": [ + 75, + 561, + 467, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we conducted experiments with SimpleBaseline, CanonPose, and Jointformer, leveraging the merged dataset $H3WB + T3WB$ , which combines both H3WB and T3WB. The results in Table 3 show that when T3WB is integrated with H3WB, the performances are improved significantly. For instance, on all whole-body keypoints, it yields 22 pt, 10.1 pt and 6.8 pt improvement for SimpleBaseline, CanonPose and Jointformer, respectively.", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. 3D whole-body lifting from incomplete 2D whole-body keypoints $(I2D\\rightarrow 3D)$", + "text_level": 1, + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a second task where we want to obtain 3D complete whole-body poses from 2D incomplete pose. This task aims to simulate the more realistic case when there are occlusions and the 2D whole-body detector outputs an incomplete skeleton. We do not provide masks for the training skeletons to allow for online data-augmentation. Instead, we propose a masking strategy as follows:", + "bbox": [ + 496, + 128, + 890, + 234 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- With $40\\%$ probability, each keypoint has a $25\\%$ chance of being masked,", + "with $20\\%$ probability, the face is entirely masked,", + "with $20\\%$ probability, the left hand is entirely masked,", + "- with $20\\%$ probability, the right hand is entirely masked." + ], + "bbox": [ + 517, + 244, + 890, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The second half of the test set (10k 2D) is devoted to this task. The masking strategy is applied only once on the 2D poses of the test set, which are directly provided as incomplete 2D skeletons for fair comparison between methods.", + "bbox": [ + 496, + 359, + 890, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results for the I2D $\\rightarrow$ 3D task are shown in Table 4. All methods perform worse than in the $2\\mathrm{D}\\rightarrow 3\\mathrm{D}$ task. SimpleBaseline[53] has low capacity and uses batch normalization that struggles with missing data, resulting in poor performance. The Large SimpleBaseline model, without batch normalization layers, achieves good results for the task's complexity. CanonPose[72] performs poorly due to errors in camera rotation prediction, which are magnified since most of the 133 keypoints are on the face and hands. The addition of 3D supervision partly solves this problem. The transformer-based Jointformer[52] method outperforms others. Sample outputs obtained by Large SimpleBaseline are shown in Figure 11, where predicted skeletons, although not accurate, are realistic.", + "bbox": [ + 496, + 420, + 890, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar to 2D $\\rightarrow$ 3D task, we experimented with Simple-Baseline, CanonPose and Jointformer on $H3WB + T3WB$ dataset and obtained significant improvements. For all", + "bbox": [ + 496, + 631, + 890, + 676 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/430741569708e1616ae10db9d8ef2ee9eccc5e1610ab1aefe21207cc92747819.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Comparing different methods for I2D→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively. Sv. is supervision." + ], + "table_body": "
MethodAllBodyFace / aligned†Hand / aligned‡
CanonPose[72]*285.0264.4319.7 / 31.9240.0 / 56.2
SimpleBaseline[53]*268.8252.0227.9 / 34.0344.3 / 83.4
CanonPose[72] + 3D sv.*163.6155.9161.3 / 22.2171.4 / 47.4
Large SimpleBaseline[53]*131.4131.6120.6 / 19.8148.8 / 44.8
Jointformer[52]109.2103.082.4 / 19.8155.9 / 53.5
H3WB + T3WB
CanonPose[72]*261.5243.3291.3 / 31.3223.1 / 53.7
SimpleBaseline[53]*260.5238.0221.1 / 32.2336.5 / 80.4
Jointformer[52]84.280.159.4 / 16.3126.5 / 44.5
", + "bbox": [ + 504, + 698, + 890, + 813 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20173", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8295350e297817f1107eb4c9651d09d26675c62a4f7f45b586afc8cb7d378038.jpg", + "image_caption": [ + "Figure 11. Example predictions from Large SimpleBaseline model for task I2D $\\rightarrow$ 3D. Colored skeletons correspond to predictions and gray skeletons correspond to groundtruths." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "whole-body keypoints, the combined dataset yields 23.5 pt, 8.3 pt and 25 pt improvement for SimpleBaseline, Canon-Pose and Jointformer, respectively. Those results further validate the effectiveness of our multi-view pipeline for creating whole-body keypoint annotations for any multi-view dataset.", + "bbox": [ + 75, + 369, + 468, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3. 3D whole-body pose estimation from a single image $(RGB\\rightarrow 3D)$", + "text_level": 1, + "bbox": [ + 76, + 469, + 468, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This task is the standard monocular 3D human pose estimation task extended to whole-body pose estimation. We provide a script to split the original Human3.6M videos into images with our indexing in order to establish image-3D correspondences. The training set contains $80\\mathrm{k}$ {image paths,3D} pairs, as well as the 2D bounding box of the human in the image. The test set contains all the test samples, including $20\\mathrm{k}$ image paths and their 2D bounding boxes. 2D coordinates are not given in order to avoid collisions with $2\\mathrm{D}\\rightarrow 3\\mathrm{D}$ and I2D $\\rightarrow 3\\mathrm{D}$ .", + "bbox": [ + 75, + 508, + 468, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For this task, we train 2 two-stage models and 1 single-stage model. For our first two-stage model, we first train a Stacked Hourglass Network (SHN)[59] to predict 2D whole-body keypoints. Then, SimpleBaseline[53] takes 2D keypoint predictions as input and lifts them to 3D space. Similarly, the second two-stage model utilizes CPN[16] to output 2D keypoints and then Jointformer[52] lifts the 2D predictions to obtain 3D whole-body poses. For our single-stage model, we modify the last layer of Resnet50[33] to directly output the 3D whole-body keypoints. We regress the 3D whole-body keypoint coordinates using L1 loss.", + "bbox": [ + 75, + 659, + 468, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results in Table 5 show the two-stage $CPN +$ Jointformer model obtains the best results. Our simple single-stage method performs better than the two-stage $SHN +$ SimpleBaseline model. Learning 2D whole-body keypoints is challenging for SHN as very close keypoints on face", + "bbox": [ + 73, + 825, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/79d0b3c1fc73808ad2441a6681e8d6fad21ac4c4b54b0b6397e3d16a91aab27d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAllBodyFace / aligned†Hand / aligned‡
RGB→2D+2D→3D:
SHN[59]+SimpleBaseline*182.5189.6138.7 / 32.5249.4 / 64.3
CPN[16]+Jointformer[52]132.6142.891.9 / 20.7192.7 / 56.9
RGB→3D:
Resnet50[33]166.7151.6123.6 / 26.3244.9 / 63.1
DOPE[75]191.3199.7187.3 / 66.0193.3 / 78.2
", + "bbox": [ + 501, + 88, + 890, + 169 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5. Comparing different methods for RGB→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively.", + "bbox": [ + 498, + 170, + 890, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and hands may introduce noise to the predicted keypoint heatmaps. The error in the 2D keypoints then makes the lifting task much more challenging. Surprisingly, RGB $\\rightarrow$ 3D seems to be harder that the I2D $\\rightarrow$ 3D task. Although there are also missing body parts due to self occlusion, RGB $\\rightarrow$ 3D contains more contextual information that should allow to better disambiguate the pose. Compared to 2D $\\rightarrow$ 3D and I2D $\\rightarrow$ 3D, direct prediction of 3D whole-body pose from images remains thus as a challenging task which we hope this benchmark can help improve over time.", + "bbox": [ + 496, + 277, + 890, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In order to show the importance of training body parts jointly, we evaluate DOPE [75] on our benchmark. Unfortunately, it fails to address occluded body parts only predicts the whole-body keypoints for $35\\%$ of the test set. For each missing keypoint, we use the (topological) nearest predicted joint as a proxy. Even so, a disjointed model like DOPE fails to achieve significant accuracy.", + "bbox": [ + 496, + 429, + 890, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 549, + 617, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduce the H3WB dataset, which extends the Human3.6M dataset with 2D and 3D keypoint annotations for body, face, and hands, containing 100k images with 133 keypoints with an average accuracy of $17\\mathrm{mm}$ . We propose three tasks based on this dataset: 3D whole-body lifting from complete 2D keypoints, 3D whole-body lifting from incomplete 2D keypoints, and 3D whole-body prediction from monocular images. We evaluate several baselines on these tasks and demonstrate promising accuracy, but with room for improvement. Lifting from incomplete 2D skeletons and direct estimation from monocular images remain challenging, and we hope that our dataset and benchmark will spur future research in these areas.", + "bbox": [ + 496, + 573, + 890, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 784, + 658, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by ANR project TOSAI ANR-20-IADJ-0009 and Ergonova Conseil, and was granted access to the HPC resources of IDRIS under the allocation 2023-AD011012640R2 and 2023-AD011013267R1 made by GENCI.", + "bbox": [ + 496, + 809, + 890, + 885 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20174", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019. 1", + "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2", + "[3] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. TOG, 2021. 1", + "[4] Dennis Bautembach, Iason Oikonomidis, and Antonis Argyros. Filling the joints: Completion and recovery of incomplete 3d human poses. Technologies, 2018. 3", + "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 3", + "[6] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: automatic estimation of 3d human pose and shape from a single image. ECCV, 2016. 2", + "[7] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2", + "[8] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In ECCV, 2018. 1, 2", + "[9] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In CVPR, 2018. 1", + "[10] Nicolò Carissimi, Paolo Rota, Cigdem Beyer, and Vittorio Murino. Filling the gaps: Predicting missing joints of human poses using denoising autoencoders. In ECCV Workshops, 2018. 3", + "[11] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 5", + "[12] Cristian Sminchisescu Catalin Ionescu, Fuxin Li. Latent structured models for human pose estimation. In ICCV, 2011. 2", + "[13] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Expnet: Landmark-free, deep, 3d facial expressions. In FG 2018, 2018. 1, 3", + "[14] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2", + "[15] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 116, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. CVPR, 2017. 9", + "[17] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, 2021. 2", + "[18] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 1, 2, 3", + "[19] Daniel Crispell and Maxim Bazik. Pix2face: Direct 3d face model estimation. In ICCV Workshops, pages 2512-2518, 2017. 3", + "[20] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. CVPR Workshop, 2019. 3", + "[21] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In ECCV, pages 534-551, 2018. 3", + "[22] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 1", + "[23] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Learning complex 3d human self-contact. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1343-1351, 2021. 1", + "[24] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9919–9928, 2021. 1", + "[25] Mercedes Garcia-Salguero, Javier Gonzalez-Jimenez, and Francisco-Angel Moreno. Human 3d pose estimation with a tilting camera for social mobile robot interaction. Sensors, 2019. 1", + "[26] Erik Gartner, Aleksis Pirinen, and Cristian Sminchisescu. Deep reinforcement learning for active human pose estimation. AAAI, 2020. 1", + "[27] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, pages 10833-10842, 2019. 1, 2", + "[28] Yiwen Gu, Shreya Pandit, Elham Saraee, Timothy Nordahl, Terry Ellis, and Margrit Betke. Home-based physical therapy with an interactive computer vision system. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019. 1", + "[29] Liang-Yan Gui, Kevin Zhang, Yu-Xiong Wang, Xiaodan Liang, José MF Moura, and Manuela Veloso. Teaching robots to predict human motion. In IROS, 2018. 1", + "[30] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In the wild human pose" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimation using explicit 2d features and intermediate 3d representations. CVPR, 2019. 1", + "[31] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Handsformer: Keypoint transformer for monocular 3d pose estimation of hands and object in interaction. CVPR, 2022. 1", + "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 4", + "[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 9", + "[34] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. CVPR, 2020. 2", + "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 5", + "[36] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPMAI, 2014. 2, 3", + "[37] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In ECCV, 2018. 1", + "[38] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images in the wild. In CVPR, 2020. 2", + "[39] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In ICCV, 2017. 3", + "[40] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In ECCV, 2020. 2, 3", + "[41] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, 2010. 2", + "[42] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015. 2", + "[43] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. CVPR, 2018. 2, 3, 7", + "[44] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1", + "[45] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. ICCV, 2019. 1", + "[46] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 1", + "[47] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. OpenPifPaf: Composite Fields for Semantic Keypoint Detection and Spatio-Temporal Association. IEEE Transactions on Intelligent Transportation Systems, 2021. 3, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2", + "[49] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In CVPR, 2017. 2", + "[50] Ruixu Liu, Ju Shen, He Wang, Chen Chen, Sen-ching Cheung, and Vijayan Asari. Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In CVPR, 2020. 2", + "[51] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 2015. 2", + "[52] Sebastian Lutz, Richard Blythman, Koustav Ghosal, Matthew Moynihan, Ciaran Simms, and Aljosa Smolic. Jointformer: Single-frame lifting transformer with error prediction and refinement for 3d human pose estimation. ArXiv, 2022. 7, 8, 9", + "[53] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 7, 8, 9", + "[54] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 2", + "[55] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, 2018. 1, 2", + "[56] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single RGB image. ICCV, 2019. 1, 2", + "[57] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. ECCV, 2020. 2", + "[58] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In CVPR, 2018. 1, 2", + "[59] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. ECCV, 2016. 9", + "[60] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed Osman, Dimitrios Tzionas, and Michael Black. Expressive body capture: 3D hands, face, and body from a single image. CVPR, 2019. 1, 3, 7, 8", + "[61] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. CVPR, 2017. 2", + "[62] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20176", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Grégory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net++: Multi-person 2d and 3d pose detection in natural images. TPAMI, 2019. 1, 2", + "[64] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2", + "[65] Yu Rong, Ziwei Liu, and Chen Change Loy. Chasing the tail in monocular 3d human reconstruction with prototype memory. TIP, 2022. 1", + "[66] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. ICCV, 2021. 1, 3", + "[67] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 1", + "[68] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 1, 3", + "[69] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In CVPR, 2017. 3", + "[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 4", + "[71] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 2", + "[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In CVPR, 2021. 2, 7, 8", + "[73] Jingbo Wang, Sijie Yan, Bo Dai, and Dahua Lin. Scene-aware generative network for human motion synthesis. In CVPR, 2021. 1", + "[74] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 1", + "[75] Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, Vincent Leroy, and Grégory Rogez. DOPE: distillation of part experts for whole-body 3d pose estimation in the wild. ECCV, 2020. 1, 3, 9", + "[76] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. ICCV, 2021. 3", + "[77] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In CVPR, 2019. 1, 2, 3", + "[78] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In CVPR, 2020. 2", + "[79] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In ECCV, 2020. 1", + "[81] Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Andrei Zanfir, and Cristian Sminchisescu. Human synthesis and scene compositing. AAAI, 2019. 1", + "[82] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, pages 2354-2364, 2019. 1, 2", + "[83] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 3", + "[84] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single RGB images. ICCV, 2017. 1, 2", + "[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan C. Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single RGB images. ICCV, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 388 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_model.json b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4497f9ba89265e6e8925f1ab6bffb5a4c429e80f --- /dev/null +++ b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_model.json @@ -0,0 +1,2787 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.797, + 0.152 + ], + "angle": 0, + "content": "H3WB: Human3.6M 3D WholeBody Dataset and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.182, + 0.793, + 0.252 + ], + "angle": 0, + "content": "Yue Zhu Nermin Samet David Picard LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France {yue.zhu, nermin.samet, david.picard}@enpc.fr Code and dataset: https://github.com/wholebody3d/wholebody3d" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.667 + ], + "angle": 0, + "content": "We present a benchmark for 3D human whole-body pose estimation, which involves identifying accurate 3D keypoints on the entire human body, including face, hands, body, and feet. Currently, the lack of a fully annotated and accurate 3D whole-body dataset results in deep networks being trained separately on specific body parts, which are combined during inference. Or they rely on pseudogroundtruth provided by parametric body models which are not as accurate as detection based methods. To overcome these issues, we introduce the Human3.6M 3D WholeBody (H3WB) dataset, which provides whole-body annotations for the Human3.6M dataset using the COCO Wholebody layout. H3WB comprises 133 whole-body keypoint annotations on 100K images, made possible by our new multiview pipeline. We also propose three tasks: i) 3D whole-body pose lifting from 2D complete whole-body pose, ii) 3D whole-body pose lifting from 2D incomplete whole-body pose, and iii) 3D whole-body pose estimation from a single RGB image. Additionally, we report several baselines from popular methods for these tasks. Furthermore, we also provide automated 3D whole-body annotations of TotalCapture and experimentally show that when used with H3WB it helps to improve the performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "3D Human pose estimation is the task of localizing human body keypoints in images which is critical to analyze human behavior, expressions, emotions, intentions, and how people communicate and interact with the physical world. As a result, 3D human pose estimation has an important role in several vision tasks and applications such as robotics [28, 25, 29] or augmented/virtual reality [55, 3, 73, 81]. However, to make more accurate predictions about human behaviors, we need more than a few body keypoints. To that end, 3D whole-body pose estimation aims to detect face, hand and foot keypoints in addition to the standard human body keypoints of classical 3D hu" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.643, + 0.303 + ], + "angle": 0, + "content": "man pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.548 + ], + "angle": 0, + "content": "The lack of accurate 3D datasets has made 3D whole-body pose estimation a challenging task, leading previous works to focus on separate body parts and train separate models on different datasets for 3D body pose [1, 30, 55, 56, 63, 3, 44, 45, 46, 65, 74, 77, 80, 26], 3D hand pose [8, 58, 84, 7, 27, 37, 82, 31], or 3D face landmarks [68, 9, 13]. However, directly ensembling separate body part models during inference suffers from issues arising from datasets' biases, pose and scales, and complex inference pipelines. Distillation from pretrained models has been used to overcome these issues, with FrankMocap [66] using three specialized pretrained models to estimate 65 whole-body keypoints (22 on the body, 40 on the hands, and 3 on the face), and DOPE [75] also using three specialized models to output 139 whole-body keypoints (13 on the body, 42 on the hands, and 84 on the face)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.896, + 0.824 + ], + "angle": 0, + "content": "Alternatively, parametric body models can be fitted to obtain whole-body pose, as has been proposed in Ex-Pose [18], SMPLify-X [60] or Monocular Total Capture (MTC) [77]. While parametric models enable sampling an almost infinite number of keypoints from the mesh [24, 22, 23], their accuracy is usually less than that of detection based methods on fine body parts like hands and feet (see supplementary material for examples from the literature). Indeed, parametric models are tailored for visual applications such as realistic motion generation [62] or avatar capture [67], where a realistic capture of the body shape and pose is more important than fine grain accuracy. Relying on a small set of data-driven pose parameters ensures realism but also limits their flexibility in representing complex unusual poses. For applications where the body shape is not needed but the accuracy of the keypoints is essential, like high performance sport analysis or ergonomics, detection based methods are thus the preferred solution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Furthermore, 3D whole-body pose estimation has not been fully explored in the literature due to the absence of a representative and accurate benchmark. As previously mentioned, existing 3D whole-body methods either rely on specific datasets and models for different body parts, lead-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20166" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.855, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.199, + 0.846, + 0.215 + ], + "angle": 0, + "content": "Figure 1. The H3WB dataset has 133 whole-body keypoint annotations in 3D as well as their respective projections in 2D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.473, + 0.36 + ], + "angle": 0, + "content": "ing to complex training pipelines and heterogeneous evaluations, or utilize parametric models that prioritize shape capture over highly precise keypoints. In addition, unified methods vary significantly in terms of keypoint layout definition, number of keypoints and distribution of keypoints across body parts (see Table 1). These significant dataset disparities and the absence of a standard benchmark make it challenging to compare methods fairly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.744 + ], + "angle": 0, + "content": "To address the above issues, we propose a new large-scale dataset for accurate 3D whole-body pose estimation called Human3.6M 3D WholeBody, or H3WB for short (see Figure 1). Our dataset extends Human3.6M [36, 12] with 3D whole-body keypoint annotations. It consists of 133 paired 2D and 3D whole-body keypoint annotations for a set of 100k images from Human3.6M, following the same layout used in COCO WholeBody [40]. More specifically, in addition to the standard 17 body keypoints, the dataset has 42 hand keypoints, 6 foot keypoints and 68 facial landmarks. H3WB was automatically created in a 3 step process: We obtain an initial set of 3D annotations using multiview geometry. Then, we trained a masked auto-encoder to complete the initial annotations. Finally, we refine the whole-body keypoints via a diffusion model. A manual annotation of 80K keypoints from 600 images shows our labels have an average error of \\(17\\mathrm{mm}\\) which suggest the H3WB keypoints are very accurate for such a complex task. We propose 3 tasks and benchmarks on the H3WB dataset for which we provide baselines: i) 3D whole-body pose lifting from a complete 2D whole-body keypoints, ii) 3D whole-body pose lifting from incomplete 2D whole-body keypoints (i.e. 2D whole-body with missing keypoints, which is more realistic), and iii) 3D whole-body pose estimation from a single RGB image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows. 1) We propose a method to create detailed 3D human pose keypoints from multi-view images. 2) We propose H3WB, the first accurate public benchmark dataset for 3D whole-body pose estimation, using the aforementioned method. Our benchmark can easily leverage existing results in 2D and enables the community to build upon existing high-quality 2D detectors on COCO. Unifying the 3D whole-body pose estimation with the COCO 2D benchmark will greatly benefit the research community. 3) We provide baselines for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.239, + 0.895, + 0.345 + ], + "angle": 0, + "content": "the 3 tasks of H3WB, which we believe will encourage the community to explore 3D whole-body pose estimation more and accelerate progress in the field. 4) Additionally, we provide 3D whole-body annotations for the TotalCapture [43] dataset, and show that when combined with the H3WB dataset it improves the performance of pose lifting tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.36, + 0.64, + 0.377 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.895, + 0.689 + ], + "angle": 0, + "content": "3D Body, hand and face pose estimation. There are two main groups of prominent approaches in 3D human pose estimation. The first group directly estimates 3D body pose from a single RGB image [61, 54, 55, 56, 61, 63]. The second group follows two stage approach where they first localize 2D keypoints and then lift 2D human pose to 3D space [53, 72, 38, 6, 49, 53]. Several optimization based methods [6, 49], utilize 2D keypoints to initialize a parametric model of the human body such as SMPL [51]. Several works attempt to eliminate the requirement of 3D annotations using 2D multi-view supervision to estimate 3D human pose [72, 78] or temporal supervision with video [17, 50]. 3D hand pose estimation methods share similar approaches as the body counterparts. First group of works, estimates hand pose from a single RGB image by directly regressing 3D hand keypoints [79], mesh vertices [27, 48], and parameters of parametric 3D hand models such as MANO [7, 64, 2, 14, 15, 82]. Second groups of works rely on intermediate 2D representations such as 2D keypoints and feature maps [8, 58, 82, 84, 34]. Simi" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.71, + 0.882, + 0.886 + ], + "angle": 0, + "content": "
DatasetSizeKeypointsBodyHandFace
Human3.6M[36]3.6M1717
3DPW[71]51k2424
LSP[41]10k1414
3DHP[71]>1.3M1717
Panoptic[42]1.5M1515
MTC[77]834K2020
InterHand2.6M[57]2.6M2121
FreiHAND[85]37k2121
RHD[84]44K2121
MTC[77]111K2121
TotalCapture[43]1.9M1272116+1674
ExPose[18]33K1442515+1589
H3WB100k1332321+2168
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.886, + 0.866, + 0.899 + ], + "angle": 0, + "content": "Table 1. Overview of datasets for 3D human pose estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "larly, predominant 3D face pose estimation methods regress the dense 3D face landmarks [19, 21, 39] and face model parameters [13, 68, 69, 20, 76] based on 3DMM [5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.473, + 0.515 + ], + "angle": 0, + "content": "3D Whole-body pose estimation. There are a several methods [60, 77, 83, 18, 75, 66] jointly estimating 3D whole-body pose. The first group of works is based on parametric human body models such as Adam [43] and SMPL-X [60]. MTC [77] is based on the Adam model [43], and first gets 2.5D predictions, then optimizes the parameters of Adam. SMPLify-X optimizes the parameter of the SMPL-X model [60] to fit it to 2D keypoints. As a major drawback, optimization-based methods are relatively slow and highly sensitive to parameter initializations. Nonparametric methods [75, 18, 66] follow different approaches to avoid heavy optimization procedure. DOPE [75] and FrankMocap [66] first train separate body, hand, and face models. Next, they combine those models within a learning framework. DOPE [75] curates pseudo-ground truths from separate body models and uses those ground-truths to supervise the distillation model. Similar to DOPE, ExPose [18] first obtains a pseudo-ground truth dataset by fitting SMPL-X model on in-the-wild images, and trains a joint model to output whole-body poses. All these methods utilize many part-based datasets. Moreover, all output different whole-body layouts with a different numbers of whole-body keypoints. FrankMocap, DOPE, and SMPLify-X estimate whole-body pose with 65, 139 and 144 keypoints, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.473, + 0.698 + ], + "angle": 0, + "content": "Pose completion completes a partially estimated pose by localizing missing keypoints. Carissimi et al. [10] propose a denoising variational autoencoder network to fill the missing keypoints in 2D pose completion. Bautembach et al. [4] selects a small subset of poses from a database based on their distance to an incomplete 3D pose, and replaces the missing keypoints with the corresponding averaged keypoints in the subset. Despite being critical for real-world scenarios, pose completion has not been sufficiently explored due to the lack of annotated keypoint datasets. Our 3D whole-body dataset can facilitate more exploration of 3D pose estimation from 2D incomplete human poses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.265, + 0.729 + ], + "angle": 0, + "content": "3. The H3WB dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.875 + ], + "angle": 0, + "content": "In this section, we describe the making of the H3WB dataset1. Our objective is to build a keypoint based 3D whole-body dataset including keypoints on the body, the face and the hands, and propose a benchmark. We use the same keypoint layout as COCO WholeBody dataset [40] with 133 keypoints. To that end, we build on the widely used Human3.6M dataset [36] for which we provide 3D whole-body keypoints. The H3WB building process is as follows: First, we use an off-the-shelf 2D whole-body de" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.895, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.225, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Figure 2. OpenPifPaf detects most of the non-occluded keypoints inside the image (orange keypoints). The occluded or undetected keypoints (cyan keypoints) are reprojections after 3D multi-view reconstruction. Notice that these reprojections do not always align with the images, like the right hand in the last view, which is probably due to OpenPifPaf not being perfectly accurate." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.343, + 0.895, + 0.465 + ], + "angle": 0, + "content": "tector combined with multi-view reconstruction to obtain an initial set of incomplete 3D whole-body keypoints. Next, we implement a completion network to fill in the keypoints missed by the multi-view geometric approach. Then, we develop a refinement method for the hands and the face to obtain more accurate keypoints. Finally, we perform quality assessment to select 25k 3D whole-body poses with high confidence and the 100k associated images from 4-view." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.486, + 0.894, + 0.503 + ], + "angle": 0, + "content": "3.1. Initial 3D whole-body dataset with OpenPifPaf" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.894, + 0.606 + ], + "angle": 0, + "content": "We run the 2D whole-body detector from OpenPif-Paf [47] on all the 4 views from the training set of Human3.6M (S1, S5, S6, S7 and S8, 1 image per 5 frames). Since the cameras of Human3.6M are well calibrated, we can reconstruct keypoints in 3D using standard multi-view geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.895, + 0.791 + ], + "angle": 0, + "content": "The OpenPifPaf 2D whole-body detector can miss keypoints due to self-occlusions (hands, feet) or unfavorable camera viewpoints (facial landmarks). However, the four-view setup allows us to recover missing keypoints and obtain a complete 3D whole-body pose, provided each keypoint appears in at least two non-opposing views. An example of this process is shown in Figure 2. Using this method, we obtained 11,426 fully complete 3D whole-body poses with all 133 keypoints and 26,333 incomplete 3D whole-body poses where all keypoints appear in at least one view, resulting in a total of 37,759 3D whole-body poses with each keypoint appearing in at least one view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We did not rely on the video information because the reconstruction problem becomes significantly more difficult when there is a motion between two frames. In the absence of motion, an additional frame does not help solve the occlusion problem. The results of our study demonstrate that multi-view labeling is sufficiently effective for our task (see Table 2, \"Geometry\" line)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.378, + 0.901 + ], + "angle": 0, + "content": "1We consider the feet keypoints as a part of the body." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.298 + ], + "angle": 0, + "content": "Figure 3. The completion network consists of one linear input layer, 4 transformer encoder blocks (each of them containing 2 transformer encoder layer with \\( d\\_model = 64 \\) and \\( n\\_head = 1 \\)), and a linear output layer. At the end of each encoder block, the features are decoded by the output layer into a predicted position in a curriculum way where later blocks decode more keypoints." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.31, + 0.441, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.499, + 0.471, + 0.584 + ], + "angle": 0, + "content": "Figure 4. Example outputs of the completion network. The orange color denotes the keypoints that were detected by OpenPifPaf. The cyan color shows the missed keypoints by OpenPifPaf but completed by our completion network. The left hand is detected in only 1 view by OpenPifPaf and thus fully predicted by the completion network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.608, + 0.273, + 0.625 + ], + "angle": 0, + "content": "3.2. Completion network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.798 + ], + "angle": 0, + "content": "In order to complete the 26,333 incomplete 3D whole-body poses, we develop a completion network as shown in Figure 3. We designed our completion network using Transformer architecture [70] as they can easily handle the conditional dependencies introduced by the skeleton's topology through masking. Since each skeleton always has exactly 133 keypoints, which can be considered as 133 tokens of 3 coordinate values. Token values are expanded from 3 coordinates to \\(3 \\times 16 = 48\\) features using Fourier encoding. We use learnable positional encoding since each keypoint is uniquely identified." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.799, + 0.471, + 0.875 + ], + "angle": 0, + "content": "We train the completion network on the 11,426 complete skeletons using a masked auto-encoder strategy [32] where the missing keypoints are masked at the input and will be predicted using the unmasked keypoints. The masking strategy is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- With a \\(50\\%\\) chance, we perform a keypoint wise mask" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.892, + 0.108 + ], + "angle": 0, + "content": "where each keypoint has \\(15\\%\\) chance of being masked," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.111, + 0.894, + 0.173 + ], + "angle": 0, + "content": "- with the remaining \\(50\\%\\) chance, we perform a block wise mask in which either the body, the left hand, the right hand, the left or the right part of the face are masked (uniform probability)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.179, + 0.895, + 0.316 + ], + "angle": 0, + "content": "To ease the learning process and take into account the causal link between some keypoints (e.g., the tip of a finger depends on the position of its parent phalanges), we introduce a curriculum approach. We compute the loss at different levels following a hierarchy where early levels consider only keypoints closer to the root, while later levels consider more deformable keypoints which highly depend on their parents. We illustrate the completion network and learning process in Figure 3. The loss function is" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.322, + 0.891, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (X, X _ {g t 3 D}, X _ {g t 2 D}) = \\mathcal {L} _ {3 D} (X, X _ {g t 3 D}) \\\\ + \\alpha \\mathcal {L} _ {2 D} (X, X _ {g t 2 D}) \\\\ + \\beta \\mathcal {L} _ {\\text {s y m}} (X), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.476 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_3D\\) is an \\(\\ell_1\\) loss of 3D coordinates, \\(\\mathcal{L}_2D\\) is an \\(\\ell_1\\) loss of 2D projection of the 3D coordinates if we have the 2D annotation from OpenPifPaf, and \\(\\mathcal{L}_{sym}\\) is a symmetric loss which is applied to make sure the left part and right part of the human have the same length on corresponding body parts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.552 + ], + "angle": 0, + "content": "We show an example output from our completion network in Figure 4. The completion network results on missing body parts are visually realistic and appealing. However, since the completion network does not rely on the image content, its output does not always align with the image" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.57, + 0.89, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Figure 5. Refinement network architecture and training process. Gaussian noise is added to the groundtruth coordinates with increasing variance, and the network is iteratively trained to recover the less noisy coordinates." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.743, + 0.885, + 0.838 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Figure 6. Example outputs from the face (top row) and hand (bottom row) refinement networks during inference time. We observe that the predictions almost converge to the correct locations in 5-iteration." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.273, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.094, + 0.48, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.094, + 0.685, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.095, + 0.889, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.214, + 0.273, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.214, + 0.48, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.214, + 0.681, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.214, + 0.888, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Figure 7. Examples of the 3D whole body skeleton. They are visually realistic humans. The strange looking faces (fatter or thinner) in different views are due to viewing artifacts of the default perspective projection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.445 + ], + "angle": 0, + "content": "and may only reflect the most common poses of the training set. This can quantitatively be seen in the line “+ Completion” of Table 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.354, + 0.47 + ], + "angle": 0, + "content": "3.3. Hands and face 2D refinements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.614 + ], + "angle": 0, + "content": "In order to correct the alignment problem, we propose another neural network that refines the 2D position of keypoints on the face and the hands. Previous studies have explored and demonstrated the effectiveness of 2D human pose refinement using an iterative error feedback framework [11]. Motivated by this, we build upon recent conditional diffusion models [35] and we consider the prediction from the completion network as noisy such that the refinement network denotes it to conditionally fit the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.471, + 0.78 + ], + "angle": 0, + "content": "We train separate refinement models for the face and the hands, while keeping the same network architecture and the same training strategy. We used a simple MLP and found it to be effective, preventing the need to explore more complex architectures. We illustrate the refinement process in Figure 5. During training, we add Gaussian noise to the groundtruth poses with an increasing variance from 5 to 25 pixels, and annotate them as step \\( t = 1 \\dots 5 \\) (step \\( t = 0 \\) is the groundtruth). The network learns to predict the pose at step \\( t \\) given the image and the noisier step \\( t + 1 \\) with a 2D supervision loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.886 + ], + "angle": 0, + "content": "We build two small datasets, each consisting of 22,000 non-occluded faces and hands respectively, with their corresponding OpenPifPaf predictions. Each image is resized to \\(384 \\times 384\\) pixels. We use a random crop of size \\(224 \\times 224\\) pixels to have the face and hands located in diverse regions of the images. We split the datasets into training and validation sets with 20,000 images and 2,000 images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.887, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Quantitatively, the face predictions achieve an average" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.445 + ], + "angle": 0, + "content": "error less than 3 pixels and the hand predictions achieve an average error less than 7 pixels on the validation sets. We show example qualitative results in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.893, + 0.55 + ], + "angle": 0, + "content": "Finally, we run the refinement networks on the 2D-projections of the 3D poses predicted by our completion network. For each 3D skeleton, we project it into the 4 different 2D views. We then crop the regions around the hands and face and denoise the corresponding predictions using the refinement network with 10 iterations to obtain refined 2D poses in each of the 4 views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.894, + 0.734 + ], + "angle": 0, + "content": "Although the refinement network is not always correct due to its training on non-occluded faces or hands, we only need 2 non-opposing views to perform geometric reconstruction. Since bad refinements tend to collapse all keypoints into the same location, we select the two non-opposing views with the highest variance in keypoint positions to avoid disruptions caused by occlusions. Using this method, we obtain 151,036 triplets of 3D whole-body keypoints, corresponding image, and 2D projected keypoints from the original set. Examples of resulting 3D whole-body skeletons and their image-aligned 2D counterparts are shown in Figure 7 and Figure 8, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.685, + 0.758 + ], + "angle": 0, + "content": "3.4. Quality assessment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "To select the most accurate triplets from our dataset, we reuse the refinement networks and employ a multi-crop strategy that accounts for the variance of the prediction. We project each 3D whole-body skeleton onto all 4 views, and produce four cropped images for each region of interest around the face and hands. The refinement network is run on these 4 crops, and the resulting predictions are aligned with the original prediction to compute the 2D error compared to the original 2D projection. We score the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.088, + 0.817, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 8. Examples of the 3D whole-body skeleton projected in 2D onto their corresponding images. They are visually accurate, though still there are small errors in detail which we do not expect to overcome due to the initial resolution and ambiguity of the images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.475, + 0.475 + ], + "angle": 0, + "content": "3D skeletons by averaging the errors of all 4 projected views, and select the 5k lowest error skeletons from each subject of Human3.6M (S1, S5, S6, S7, S8) to form the \\(5\\mathrm{k}\\times 4(\\mathrm{view})\\times 5(\\mathrm{subject}) = 100\\mathrm{k}\\) triplets of {image, 2D coordinates, 3D coordinates in camera space} of our 3D whole-body dataset." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.489, + 0.273, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.489, + 0.462, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.597, + 0.272, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.597, + 0.462, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.707, + 0.471, + 0.765 + ], + "angle": 0, + "content": "Figure 9. 3D error distributions calculated from 80k manually corrected keypoint annotations. 3D error distributions are presented for whole-body, body, hand and face in mm. We observe that 3D errors are mostly concentrated between 10mm and 20mm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To assess the quality of the H3WB dataset, we conducted a cross-check study on 600 randomly selected images from the dataset. In this study, annotators were presented an image with the 2D projection of the 3D skeleton on top and were asked to manually correct mis-aligned keypoints by drag and drop. Using multi-view geometry, we reconstructed these corrected skeletons in 3D and compared them to our original skeletons. To validate our process, we show" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.381, + 0.889, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.566, + 0.866, + 0.58 + ], + "angle": 0, + "content": "Figure 10. Per keypoints error statistics. Please zoom in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.895, + 0.772 + ], + "angle": 0, + "content": "the influence of each step in Table 2. The geometric approach produced good results but unfortunately cannot provide a large enough dataset. The completion step allows to obtain all labels but at the cost of degraded accuracy due to lack of alignment as explained in section 3.2. The diffusion recovers the original accuracy of the geometric approach. \\(2\\mathrm{mm}\\) difference is irrelevant given the initial resolution of the images. We obtain a final average error of \\(17\\mathrm{mm}\\) which is very accurate for such a difficult task, and leads to a benchmark which we believe will not be saturated until methods reach around \\(35\\mathrm{mm}\\)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.781, + 0.892, + 0.842 + ], + "angle": 0, + "content": "
Steps# keypoints available3D error (mm)
AllBodyFaceHand
Geometry4812714.8717.7213.2915.87
+ Completion7980029.3125.5726.0236.67
+ Diffusion7980016.9818.6315.0819.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.842, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Table 2. Quantitative analysis of each intermediate step in our pipeline." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "3D error distributions: In Figure 9, we illustrate 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "error distributions in mm for all whole-body keypoints and each whole-body part keypoints (i.e. body, hand and face) separately. The distributions of errors are well concentrated around low values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Per keypoint errors: All per keypoint error statistics are shown in Figure 10. \\(98.3\\%\\) of the images are below 5cm error (97.4% for body joints, 99.7% hands, and 96.4% face). Similarly, \\(82.6\\%\\) of the images are below 2cm error (75.5% for body joints, 89.2% hands, and 75.8% face)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.238, + 0.4, + 0.254 + ], + "angle": 0, + "content": "3.5. TotalCapture 3D WholeBody Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.468, + 0.456 + ], + "angle": 0, + "content": "In addition to H3WB dataset, we also prepare whole-body annotations for training sequence of poses from TotalCapture [43] dataset using our proposed multi-view pipeline. We call this dataset TotalCapture 3D WholeBody, or T3WB for short. To create TotalCapture 3D WholeBody, we first obtain 2D whole-body keypoints from OpenPif-Paf [47]. At this stage we discard the frames without any human. Next, we finetune the completion network initialized by H3WB weights using 7000 samples of TotalCapture with complete 8-views. At the end we obtain 125,960 triplets of 3D whole-body keypoint, corresponding image and 2D projected keypoints. This shows our pipeline can be used with any multiview dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.468, + 0.563 + ], + "angle": 0, + "content": "We do not conduct a quality assessment study for T3WB and therefore we cannot guarantee the precision of the annotations. Instead, we use T3WB together with H3WB for training models. To alleviate the effect of the noisy annotations in T3WB, we sample more from H3WB than T3WB in each batch. More specifically, we follow 4:1 ratio for each batch during \\(H3WB + T3WB\\) trainings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.578, + 0.296, + 0.594 + ], + "angle": 0, + "content": "4. The H3WB benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.468, + 0.77 + ], + "angle": 0, + "content": "We use the H3WB dataset to propose a benchmark and the associated leaderboard. We split the dataset into training and test sets. The training set contains all samples from S1, S5, S6 and S7, including 80k {image,2D,3D} triplets. The test set contains all samples from S8, including 20k triplets. The test set labels are retained to prevent involuntary overfitting on the test set. Evaluation is accessible only by submitting results to the maintainers. We do not provide a validation set. We encourage researchers to report 5-fold cross-validation average and standard deviation (see supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.771, + 0.439, + 0.785 + ], + "angle": 0, + "content": "The corresponding benchmark has 3 different tasks:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.468, + 0.828 + ], + "angle": 0, + "content": "1. 3D whole-body lifting from complete 2D whole-body skeletons, or \\(2D \\rightarrow 3D\\) for short." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.835, + 0.468, + 0.864 + ], + "angle": 0, + "content": "2. 3D whole-body lifting from incomplete 2D whole-body skeletons, or \\( I2D \\rightarrow 3D \\) for short." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.871, + 0.468, + 0.899 + ], + "angle": 0, + "content": "3. 3D whole-body skeleton prediction from image, or \\( RGB \\rightarrow 3D \\) for short." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.798, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "For each task, we report the following MPJPE (Mean Per Joint Position Error) metrics:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.135, + 0.89, + 0.21 + ], + "angle": 0, + "content": "- MPJPE for the whole-body, the body (keypoint 1-23), the face (keypoint 24-91) and the hands (keypoint 92-133) when whole-body is centered on the root joint, i.e. aligned with the pelvis, which in our case is the middle of two hip joints2," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.217, + 0.89, + 0.248 + ], + "angle": 0, + "content": "- MPJPE for the face when it is centered on the nose, i.e. aligned with keypoint 1," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.255, + 0.891, + 0.3 + ], + "angle": 0, + "content": "- MPJPE for the hands when hands are centered on the wrist, i.e left hand aligned with keypoint 92 and right hand aligned with keypoint 113." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.135, + 0.891, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.418 + ], + "angle": 0, + "content": "To create baselines on each task, we adapt popular methods from the literature by changing the number of keypoints to that of our whole-body dataset. Notice that we keep the training recipes of the original papers to avoid over-fitting to this new benchmark. In practice, we recommend to perform model selection and hyper-parameters tuning using 5-fold cross-validation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.431, + 0.89, + 0.462 + ], + "angle": 0, + "content": "4.1. 3D whole-body lifting from complete 2D whole-body keypoints \\((2D\\rightarrow 3D)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.89, + 0.531 + ], + "angle": 0, + "content": "This task is similar to the standard 3D human pose estimation from 2D keypoints but using whole-body keypoints. The training set contains 80k 2D-3D pairs. The test set contains only a half of all the test samples, i.e. 10k 2D poses3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.729 + ], + "angle": 0, + "content": "We evaluate 6 methods on this task. SimpleBaseline [53] is a well-established model, consisting of a 6-layer MLP. We propose a modification, replacing the network architecture with an 8-layer MLP, which we call Large Simple-Baseline inspired by CanonPose [72]. Jointformer [52] is a recent transformer-based method. CanonPose is trained only with 2D supervision [72]. We also adapt CanonPose to work with additional 3D supervision by manually creating 3 fixed camera views and rotating the 3D skeletons into the corresponding view before projecting them into 2D, training it with multi-view weak-supervision. Finally, we report results for the parametric model SMPLify-X [60] by running optimizations on each input sample." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.85 + ], + "angle": 0, + "content": "We train SimpleBaseline models using their official training setting as described in [53]. The inputs and targets are normalized by subtracting the mean and dividing by the standard deviation. Similarly, we train CanonPose [72] models following their official training setup where the inputs and targets are centered on the pelvis and scaled by the Forbenius norm. We train the Jointformer model in the two stages as described in [52]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.863, + 0.882, + 0.876 + ], + "angle": 0, + "content": "2We provide the whole-body keypoint ids in supplementary material." + }, + { + "type": "page_footnote", + "bbox": [ + 0.502, + 0.876, + 0.891, + 0.9 + ], + "angle": 0, + "content": "3The other half is reserved for the task I2D \\(\\rightarrow\\) 3D to prevent access to the missing keypoints." + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.863, + 0.891, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20172" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.229 + ], + "angle": 0, + "content": "
MethodAllBodyFace / aligned†Hand / aligned‡
H3WB
SMPL-X[60]188.9166.0208.3 / 23.7170.2 / 44.4
CanonPose[72]*186.7193.7188.4 / 24.6180.2 / 48.9
SimpleBaseline [53]*125.4125.7115.9 / 24.6140.7 / 42.5
CanonPose[72] w 3D sv.*117.7117.5112.0 / 17.9126.9 / 38.3
Large SimpleBaseline[53]*112.3112.6110.6 / 14.6114.8 / 31.7
Jointformer[52]88.384.966.5 / 17.8125.3 / 43.7
H3WB + T3WB
CanonPose[72]*164.7161.1174.5 / 21.5150.8 / 43.6
SimpleBaseline [53]*115.3114.8109.4 / 15.8125.1 / 33.5
Jointformer[52]81.578.060.4 / 16.2117.6 / 38.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.23, + 0.468, + 0.311 + ], + "angle": 0, + "content": "Table 3. Comparing different methods for \\(2\\mathrm{D}\\rightarrow 3\\mathrm{D}\\) on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with \\* output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except \\(\\dagger\\) and \\(\\ddagger\\) show nose and wrist aligned results for face and hands, respectively. Sv. is supervision." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.351, + 0.468, + 0.487 + ], + "angle": 0, + "content": "SimpleBaseline and CanonPose models output normalized whole-body keypoints which requires re-scaling at inference. We use statistics from the training set to adjust the test predictions. We calculate a scaling factor using the ratio of 3D to 2D bounding boxes. The formula is: \\( X_{\\mathrm{final}} = X_{\\mathrm{unit}} \\times \\overline{\\sigma_{3d}} \\times \\frac{\\sigma_{2d}}{\\sigma_{2d}} \\), where \\( X_{\\mathrm{unit}} \\) is the normalized prediction, \\( \\overline{\\sigma_{3d}} \\) is the average size of the 3D training boxes, \\( \\sigma_{2d} \\) is the size of the current 2D box, and \\( \\overline{\\sigma_{2d}} \\) is the average size of the 2D training boxes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.494, + 0.468, + 0.554 + ], + "angle": 0, + "content": "Since SMPLify-X has 144 keypoints with a different layout, we use interpolation to transform between the WholeBody skeleton and SMPL-X and run SMPL-X's optimization for 2,000 iterations (4 minutes/sample)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.468, + 0.773 + ], + "angle": 0, + "content": "We present the results in Table 3. SMPLify-X performs the worst, showing that parametric models struggle more than discriminative approaches. SimpleBaseline[53] is a solid method, and Large SimpleBaseline improves its performance further. CanonPose[72] can be improved with additional 3D supervision, but still performs worse than Large SimpleBaseline. CanonPose also predicts the camera view, and the uncertainty in this prediction can lead to more error. Jointformer[52] achieves the best results among all methods, but still has room for improvement. All methods perform worse on our benchmark than on Human3.6M because of pelvis centering, which creates higher numerical error on extremities like hands and face, the parts that contain most of the whole-body keypoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Additionally, we conducted experiments with SimpleBaseline, CanonPose, and Jointformer, leveraging the merged dataset \\( H3WB + T3WB \\), which combines both H3WB and T3WB. The results in Table 3 show that when T3WB is integrated with H3WB, the performances are improved significantly. For instance, on all whole-body keypoints, it yields 22 pt, 10.1 pt and 6.8 pt improvement for SimpleBaseline, CanonPose and Jointformer, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.891, + 0.122 + ], + "angle": 0, + "content": "4.2. 3D whole-body lifting from incomplete 2D whole-body keypoints \\((I2D\\rightarrow 3D)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.129, + 0.892, + 0.235 + ], + "angle": 0, + "content": "We propose a second task where we want to obtain 3D complete whole-body poses from 2D incomplete pose. This task aims to simulate the more realistic case when there are occlusions and the 2D whole-body detector outputs an incomplete skeleton. We do not provide masks for the training skeletons to allow for online data-augmentation. Instead, we propose a masking strategy as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.245, + 0.891, + 0.275 + ], + "angle": 0, + "content": "- With \\(40\\%\\) probability, each keypoint has a \\(25\\%\\) chance of being masked," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.28, + 0.86, + 0.295 + ], + "angle": 0, + "content": "with \\(20\\%\\) probability, the face is entirely masked," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.3, + 0.89, + 0.315 + ], + "angle": 0, + "content": "with \\(20\\%\\) probability, the left hand is entirely masked," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.32, + 0.89, + 0.349 + ], + "angle": 0, + "content": "- with \\(20\\%\\) probability, the right hand is entirely masked." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.245, + 0.891, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.891, + 0.42 + ], + "angle": 0, + "content": "The second half of the test set (10k 2D) is devoted to this task. The masking strategy is applied only once on the 2D poses of the test set, which are directly provided as incomplete 2D skeletons for fair comparison between methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.631 + ], + "angle": 0, + "content": "The results for the I2D \\(\\rightarrow\\) 3D task are shown in Table 4. All methods perform worse than in the \\(2\\mathrm{D}\\rightarrow 3\\mathrm{D}\\) task. SimpleBaseline[53] has low capacity and uses batch normalization that struggles with missing data, resulting in poor performance. The Large SimpleBaseline model, without batch normalization layers, achieves good results for the task's complexity. CanonPose[72] performs poorly due to errors in camera rotation prediction, which are magnified since most of the 133 keypoints are on the face and hands. The addition of 3D supervision partly solves this problem. The transformer-based Jointformer[52] method outperforms others. Sample outputs obtained by Large SimpleBaseline are shown in Figure 11, where predicted skeletons, although not accurate, are realistic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.677 + ], + "angle": 0, + "content": "Similar to 2D \\(\\rightarrow\\) 3D task, we experimented with Simple-Baseline, CanonPose and Jointformer on \\(H3WB + T3WB\\) dataset and obtained significant improvements. For all" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.699, + 0.892, + 0.814 + ], + "angle": 0, + "content": "
MethodAllBodyFace / aligned†Hand / aligned‡
CanonPose[72]*285.0264.4319.7 / 31.9240.0 / 56.2
SimpleBaseline[53]*268.8252.0227.9 / 34.0344.3 / 83.4
CanonPose[72] + 3D sv.*163.6155.9161.3 / 22.2171.4 / 47.4
Large SimpleBaseline[53]*131.4131.6120.6 / 19.8148.8 / 44.8
Jointformer[52]109.2103.082.4 / 19.8155.9 / 53.5
H3WB + T3WB
CanonPose[72]*261.5243.3291.3 / 31.3223.1 / 53.7
SimpleBaseline[53]*260.5238.0221.1 / 32.2336.5 / 80.4
Jointformer[52]84.280.159.4 / 16.3126.5 / 44.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.815, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Table 4. Comparing different methods for I2D→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively. Sv. is supervision." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.303, + 0.47, + 0.346 + ], + "angle": 0, + "content": "Figure 11. Example predictions from Large SimpleBaseline model for task I2D \\(\\rightarrow\\) 3D. Colored skeletons correspond to predictions and gray skeletons correspond to groundtruths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.462 + ], + "angle": 0, + "content": "whole-body keypoints, the combined dataset yields 23.5 pt, 8.3 pt and 25 pt improvement for SimpleBaseline, Canon-Pose and Jointformer, respectively. Those results further validate the effectiveness of our multi-view pipeline for creating whole-body keypoint annotations for any multi-view dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.469, + 0.501 + ], + "angle": 0, + "content": "4.3. 3D whole-body pose estimation from a single image \\((RGB\\rightarrow 3D)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.469, + 0.658 + ], + "angle": 0, + "content": "This task is the standard monocular 3D human pose estimation task extended to whole-body pose estimation. We provide a script to split the original Human3.6M videos into images with our indexing in order to establish image-3D correspondences. The training set contains \\(80\\mathrm{k}\\) {image paths,3D} pairs, as well as the 2D bounding box of the human in the image. The test set contains all the test samples, including \\(20\\mathrm{k}\\) image paths and their 2D bounding boxes. 2D coordinates are not given in order to avoid collisions with \\(2\\mathrm{D}\\rightarrow 3\\mathrm{D}\\) and I2D \\(\\rightarrow 3\\mathrm{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.825 + ], + "angle": 0, + "content": "For this task, we train 2 two-stage models and 1 single-stage model. For our first two-stage model, we first train a Stacked Hourglass Network (SHN)[59] to predict 2D whole-body keypoints. Then, SimpleBaseline[53] takes 2D keypoint predictions as input and lifts them to 3D space. Similarly, the second two-stage model utilizes CPN[16] to output 2D keypoints and then Jointformer[52] lifts the 2D predictions to obtain 3D whole-body poses. For our single-stage model, we modify the last layer of Resnet50[33] to directly output the 3D whole-body keypoints. We regress the 3D whole-body keypoint coordinates using L1 loss." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.826, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Results in Table 5 show the two-stage \\( CPN + \\) Jointformer model obtains the best results. Our simple single-stage method performs better than the two-stage \\( SHN + \\) SimpleBaseline model. Learning 2D whole-body keypoints is challenging for SHN as very close keypoints on face" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.17 + ], + "angle": 0, + "content": "
MethodAllBodyFace / aligned†Hand / aligned‡
RGB→2D+2D→3D:
SHN[59]+SimpleBaseline*182.5189.6138.7 / 32.5249.4 / 64.3
CPN[16]+Jointformer[52]132.6142.891.9 / 20.7192.7 / 56.9
RGB→3D:
Resnet50[33]166.7151.6123.6 / 26.3244.9 / 63.1
DOPE[75]191.3199.7187.3 / 66.0193.3 / 78.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.171, + 0.892, + 0.254 + ], + "angle": 0, + "content": "Table 5. Comparing different methods for RGB→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.43 + ], + "angle": 0, + "content": "and hands may introduce noise to the predicted keypoint heatmaps. The error in the 2D keypoints then makes the lifting task much more challenging. Surprisingly, RGB \\(\\rightarrow\\) 3D seems to be harder that the I2D \\(\\rightarrow\\) 3D task. Although there are also missing body parts due to self occlusion, RGB \\(\\rightarrow\\) 3D contains more contextual information that should allow to better disambiguate the pose. Compared to 2D \\(\\rightarrow\\) 3D and I2D \\(\\rightarrow\\) 3D, direct prediction of 3D whole-body pose from images remains thus as a challenging task which we hope this benchmark can help improve over time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.536 + ], + "angle": 0, + "content": "In order to show the importance of training body parts jointly, we evaluate DOPE [75] on our benchmark. Unfortunately, it fails to address occluded body parts only predicts the whole-body keypoints for \\(35\\%\\) of the test set. For each missing keypoint, we use the (topological) nearest predicted joint as a proxy. Even so, a disjointed model like DOPE fails to achieve significant accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.619, + 0.565 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.772 + ], + "angle": 0, + "content": "In this paper, we introduce the H3WB dataset, which extends the Human3.6M dataset with 2D and 3D keypoint annotations for body, face, and hands, containing 100k images with 133 keypoints with an average accuracy of \\(17\\mathrm{mm}\\). We propose three tasks based on this dataset: 3D whole-body lifting from complete 2D keypoints, 3D whole-body lifting from incomplete 2D keypoints, and 3D whole-body prediction from monocular images. We evaluate several baselines on these tasks and demonstrate promising accuracy, but with room for improvement. Lifting from incomplete 2D skeletons and direct estimation from monocular images remain challenging, and we hope that our dataset and benchmark will spur future research in these areas." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.659, + 0.802 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.886 + ], + "angle": 0, + "content": "This work was supported by ANR project TOSAI ANR-20-IADJ-0009 and Ergonova Conseil, and was granted access to the HPC resources of IDRIS under the allocation 2023-AD011012640R2 and 2023-AD011013267R1 made by GENCI." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20174" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[3] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. TOG, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[4] Dennis Bautembach, Iason Oikonomidis, and Antonis Argyros. Filling the joints: Completion and recovery of incomplete 3d human poses. Technologies, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.365, + 0.469, + 0.42 + ], + "angle": 0, + "content": "[6] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: automatic estimation of 3d human pose and shape from a single image. ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.423, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[7] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[8] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[9] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[10] Nicolò Carissimi, Paolo Rota, Cigdem Beyer, and Vittorio Murino. Filling the gaps: Predicting missing joints of human poses using denoising autoencoders. In ECCV Workshops, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.681 + ], + "angle": 0, + "content": "[11] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[12] Cristian Sminchisescu Catalin Ionescu, Fuxin Li. Latent structured models for human pose estimation. In ICCV, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[13] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Expnet: Landmark-free, deep, 3d facial expressions. In FG 2018, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[14] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[15] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. CVPR, 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[17] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[18] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[19] Daniel Crispell and Maxim Bazik. Pix2face: Direct 3d face model estimation. In ICCV Workshops, pages 2512-2518, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[20] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. CVPR Workshop, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[21] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In ECCV, pages 534-551, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[22] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[23] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Learning complex 3d human self-contact. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1343-1351, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[24] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9919–9928, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[25] Mercedes Garcia-Salguero, Javier Gonzalez-Jimenez, and Francisco-Angel Moreno. Human 3d pose estimation with a tilting camera for social mobile robot interaction. Sensors, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[26] Erik Gartner, Aleksis Pirinen, and Cristian Sminchisescu. Deep reinforcement learning for active human pose estimation. AAAI, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[27] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, pages 10833-10842, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[28] Yiwen Gu, Shreya Pandit, Elham Saraee, Timothy Nordahl, Terry Ellis, and Margrit Betke. Home-based physical therapy with an interactive computer vision system. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[29] Liang-Yan Gui, Kevin Zhang, Yu-Xiong Wang, Xiaodan Liang, José MF Moura, and Manuela Veloso. Teaching robots to predict human motion. In IROS, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In the wild human pose" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20175" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "estimation using explicit 2d features and intermediate 3d representations. CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[31] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Handsformer: Keypoint transformer for monocular 3d pose estimation of hands and object in interaction. CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.468, + 0.295 + ], + "angle": 0, + "content": "[34] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.468, + 0.324 + ], + "angle": 0, + "content": "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.327, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[36] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPMAI, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[37] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[38] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images in the wild. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[39] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.528, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[40] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[41] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[42] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[43] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. CVPR, 2018. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[44] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[45] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[46] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[47] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. OpenPifPaf: Composite Fields for Semantic Keypoint Detection and Spatio-Temporal Association. IEEE Transactions on Intelligent Transportation Systems, 2021. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[48] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[49] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[50] Ruixu Liu, Ju Shen, He Wang, Chen Chen, Sen-ching Cheung, and Vijayan Asari. Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[51] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[52] Sebastian Lutz, Richard Blythman, Koustav Ghosal, Matthew Moynihan, Ciaran Simms, and Aljosa Smolic. Jointformer: Single-frame lifting transformer with error prediction and refinement for 3d human pose estimation. ArXiv, 2022. 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[53] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[54] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[55] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[56] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single RGB image. ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[57] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[58] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[59] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. ECCV, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[60] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed Osman, Dimitrios Tzionas, and Michael Black. Expressive body capture: 3D hands, face, and body from a single image. CVPR, 2019. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[61] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[62] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20176" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[63] Grégory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net++: Multi-person 2d and 3d pose detection in natural images. TPAMI, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[64] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[65] Yu Rong, Ziwei Liu, and Chen Change Loy. Chasing the tail in monocular 3d human reconstruction with prototype memory. TIP, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[66] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[67] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[68] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[69] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[71] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.533, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In CVPR, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[73] Jingbo Wang, Sijie Yan, Bo Dai, and Dahua Lin. Scene-aware generative network for human motion synthesis. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[74] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.676, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[75] Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, Vincent Leroy, and Grégory Rogez. DOPE: distillation of part experts for whole-body 3d pose estimation in the wild. ECCV, 2020. 1, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.732, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[76] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.775, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[77] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In CVPR, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[78] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[79] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[80] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[81] Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Andrei Zanfir, and Cristian Sminchisescu. Human synthesis and scene compositing. AAAI, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[82] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, pages 2354-2364, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[83] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[84] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single RGB images. ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan C. Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single RGB images. ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20177" + } + ] +] \ No newline at end of file diff --git a/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_origin.pdf b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1b8b1fb7a0a5dca2062601046465e1a8ba469268 --- /dev/null +++ b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/a8f8b3db-4bb9-47fd-9ae5-943dc0af642d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8ce5a5de565446aab8c8d7c9df4f9e5c754e7f0dc5a3572394246a07808fa95 +size 3819122 diff --git a/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/full.md b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..207638f4ab106c768e52d56668efc30f01bb216b --- /dev/null +++ b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/full.md @@ -0,0 +1,351 @@ +# H3WB: Human3.6M 3D WholeBody Dataset and Benchmark + +Yue Zhu Nermin Samet David Picard LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France {yue.zhu, nermin.samet, david.picard}@enpc.fr Code and dataset: https://github.com/wholebody3d/wholebody3d + +# Abstract + +We present a benchmark for 3D human whole-body pose estimation, which involves identifying accurate 3D keypoints on the entire human body, including face, hands, body, and feet. Currently, the lack of a fully annotated and accurate 3D whole-body dataset results in deep networks being trained separately on specific body parts, which are combined during inference. Or they rely on pseudogroundtruth provided by parametric body models which are not as accurate as detection based methods. To overcome these issues, we introduce the Human3.6M 3D WholeBody (H3WB) dataset, which provides whole-body annotations for the Human3.6M dataset using the COCO Wholebody layout. H3WB comprises 133 whole-body keypoint annotations on 100K images, made possible by our new multiview pipeline. We also propose three tasks: i) 3D whole-body pose lifting from 2D complete whole-body pose, ii) 3D whole-body pose lifting from 2D incomplete whole-body pose, and iii) 3D whole-body pose estimation from a single RGB image. Additionally, we report several baselines from popular methods for these tasks. Furthermore, we also provide automated 3D whole-body annotations of TotalCapture and experimentally show that when used with H3WB it helps to improve the performance. + +# 1. Introduction + +3D Human pose estimation is the task of localizing human body keypoints in images which is critical to analyze human behavior, expressions, emotions, intentions, and how people communicate and interact with the physical world. As a result, 3D human pose estimation has an important role in several vision tasks and applications such as robotics [28, 25, 29] or augmented/virtual reality [55, 3, 73, 81]. However, to make more accurate predictions about human behaviors, we need more than a few body keypoints. To that end, 3D whole-body pose estimation aims to detect face, hand and foot keypoints in addition to the standard human body keypoints of classical 3D hu + +man pose estimation. + +The lack of accurate 3D datasets has made 3D whole-body pose estimation a challenging task, leading previous works to focus on separate body parts and train separate models on different datasets for 3D body pose [1, 30, 55, 56, 63, 3, 44, 45, 46, 65, 74, 77, 80, 26], 3D hand pose [8, 58, 84, 7, 27, 37, 82, 31], or 3D face landmarks [68, 9, 13]. However, directly ensembling separate body part models during inference suffers from issues arising from datasets' biases, pose and scales, and complex inference pipelines. Distillation from pretrained models has been used to overcome these issues, with FrankMocap [66] using three specialized pretrained models to estimate 65 whole-body keypoints (22 on the body, 40 on the hands, and 3 on the face), and DOPE [75] also using three specialized models to output 139 whole-body keypoints (13 on the body, 42 on the hands, and 84 on the face). + +Alternatively, parametric body models can be fitted to obtain whole-body pose, as has been proposed in Ex-Pose [18], SMPLify-X [60] or Monocular Total Capture (MTC) [77]. While parametric models enable sampling an almost infinite number of keypoints from the mesh [24, 22, 23], their accuracy is usually less than that of detection based methods on fine body parts like hands and feet (see supplementary material for examples from the literature). Indeed, parametric models are tailored for visual applications such as realistic motion generation [62] or avatar capture [67], where a realistic capture of the body shape and pose is more important than fine grain accuracy. Relying on a small set of data-driven pose parameters ensures realism but also limits their flexibility in representing complex unusual poses. For applications where the body shape is not needed but the accuracy of the keypoints is essential, like high performance sport analysis or ergonomics, detection based methods are thus the preferred solution. + +Furthermore, 3D whole-body pose estimation has not been fully explored in the literature due to the absence of a representative and accurate benchmark. As previously mentioned, existing 3D whole-body methods either rely on specific datasets and models for different body parts, lead- + +![](images/25f628b325ee5dc9efdbfe6a0f0299a05223c6d2c26ff5de4853d5ee48af82d5.jpg) +Figure 1. The H3WB dataset has 133 whole-body keypoint annotations in 3D as well as their respective projections in 2D. + +ing to complex training pipelines and heterogeneous evaluations, or utilize parametric models that prioritize shape capture over highly precise keypoints. In addition, unified methods vary significantly in terms of keypoint layout definition, number of keypoints and distribution of keypoints across body parts (see Table 1). These significant dataset disparities and the absence of a standard benchmark make it challenging to compare methods fairly. + +To address the above issues, we propose a new large-scale dataset for accurate 3D whole-body pose estimation called Human3.6M 3D WholeBody, or H3WB for short (see Figure 1). Our dataset extends Human3.6M [36, 12] with 3D whole-body keypoint annotations. It consists of 133 paired 2D and 3D whole-body keypoint annotations for a set of 100k images from Human3.6M, following the same layout used in COCO WholeBody [40]. More specifically, in addition to the standard 17 body keypoints, the dataset has 42 hand keypoints, 6 foot keypoints and 68 facial landmarks. H3WB was automatically created in a 3 step process: We obtain an initial set of 3D annotations using multiview geometry. Then, we trained a masked auto-encoder to complete the initial annotations. Finally, we refine the whole-body keypoints via a diffusion model. A manual annotation of 80K keypoints from 600 images shows our labels have an average error of $17\mathrm{mm}$ which suggest the H3WB keypoints are very accurate for such a complex task. We propose 3 tasks and benchmarks on the H3WB dataset for which we provide baselines: i) 3D whole-body pose lifting from a complete 2D whole-body keypoints, ii) 3D whole-body pose lifting from incomplete 2D whole-body keypoints (i.e. 2D whole-body with missing keypoints, which is more realistic), and iii) 3D whole-body pose estimation from a single RGB image. + +Our contributions can be summarized as follows. 1) We propose a method to create detailed 3D human pose keypoints from multi-view images. 2) We propose H3WB, the first accurate public benchmark dataset for 3D whole-body pose estimation, using the aforementioned method. Our benchmark can easily leverage existing results in 2D and enables the community to build upon existing high-quality 2D detectors on COCO. Unifying the 3D whole-body pose estimation with the COCO 2D benchmark will greatly benefit the research community. 3) We provide baselines for + +the 3 tasks of H3WB, which we believe will encourage the community to explore 3D whole-body pose estimation more and accelerate progress in the field. 4) Additionally, we provide 3D whole-body annotations for the TotalCapture [43] dataset, and show that when combined with the H3WB dataset it improves the performance of pose lifting tasks. + +# 2. Related work + +3D Body, hand and face pose estimation. There are two main groups of prominent approaches in 3D human pose estimation. The first group directly estimates 3D body pose from a single RGB image [61, 54, 55, 56, 61, 63]. The second group follows two stage approach where they first localize 2D keypoints and then lift 2D human pose to 3D space [53, 72, 38, 6, 49, 53]. Several optimization based methods [6, 49], utilize 2D keypoints to initialize a parametric model of the human body such as SMPL [51]. Several works attempt to eliminate the requirement of 3D annotations using 2D multi-view supervision to estimate 3D human pose [72, 78] or temporal supervision with video [17, 50]. 3D hand pose estimation methods share similar approaches as the body counterparts. First group of works, estimates hand pose from a single RGB image by directly regressing 3D hand keypoints [79], mesh vertices [27, 48], and parameters of parametric 3D hand models such as MANO [7, 64, 2, 14, 15, 82]. Second groups of works rely on intermediate 2D representations such as 2D keypoints and feature maps [8, 58, 82, 84, 34]. Simi + +
DatasetSizeKeypointsBodyHandFace
Human3.6M[36]3.6M1717
3DPW[71]51k2424
LSP[41]10k1414
3DHP[71]>1.3M1717
Panoptic[42]1.5M1515
MTC[77]834K2020
InterHand2.6M[57]2.6M2121
FreiHAND[85]37k2121
RHD[84]44K2121
MTC[77]111K2121
TotalCapture[43]1.9M1272116+1674
ExPose[18]33K1442515+1589
H3WB100k1332321+2168
+ +Table 1. Overview of datasets for 3D human pose estimation. + +larly, predominant 3D face pose estimation methods regress the dense 3D face landmarks [19, 21, 39] and face model parameters [13, 68, 69, 20, 76] based on 3DMM [5]. + +3D Whole-body pose estimation. There are a several methods [60, 77, 83, 18, 75, 66] jointly estimating 3D whole-body pose. The first group of works is based on parametric human body models such as Adam [43] and SMPL-X [60]. MTC [77] is based on the Adam model [43], and first gets 2.5D predictions, then optimizes the parameters of Adam. SMPLify-X optimizes the parameter of the SMPL-X model [60] to fit it to 2D keypoints. As a major drawback, optimization-based methods are relatively slow and highly sensitive to parameter initializations. Nonparametric methods [75, 18, 66] follow different approaches to avoid heavy optimization procedure. DOPE [75] and FrankMocap [66] first train separate body, hand, and face models. Next, they combine those models within a learning framework. DOPE [75] curates pseudo-ground truths from separate body models and uses those ground-truths to supervise the distillation model. Similar to DOPE, ExPose [18] first obtains a pseudo-ground truth dataset by fitting SMPL-X model on in-the-wild images, and trains a joint model to output whole-body poses. All these methods utilize many part-based datasets. Moreover, all output different whole-body layouts with a different numbers of whole-body keypoints. FrankMocap, DOPE, and SMPLify-X estimate whole-body pose with 65, 139 and 144 keypoints, respectively. + +Pose completion completes a partially estimated pose by localizing missing keypoints. Carissimi et al. [10] propose a denoising variational autoencoder network to fill the missing keypoints in 2D pose completion. Bautembach et al. [4] selects a small subset of poses from a database based on their distance to an incomplete 3D pose, and replaces the missing keypoints with the corresponding averaged keypoints in the subset. Despite being critical for real-world scenarios, pose completion has not been sufficiently explored due to the lack of annotated keypoint datasets. Our 3D whole-body dataset can facilitate more exploration of 3D pose estimation from 2D incomplete human poses. + +# 3. The H3WB dataset + +In this section, we describe the making of the H3WB dataset1. Our objective is to build a keypoint based 3D whole-body dataset including keypoints on the body, the face and the hands, and propose a benchmark. We use the same keypoint layout as COCO WholeBody dataset [40] with 133 keypoints. To that end, we build on the widely used Human3.6M dataset [36] for which we provide 3D whole-body keypoints. The H3WB building process is as follows: First, we use an off-the-shelf 2D whole-body de + +![](images/1b562750440474c9e446610dc3cd872e9bfdbfa07547960d199c400c847c61e4.jpg) +Figure 2. OpenPifPaf detects most of the non-occluded keypoints inside the image (orange keypoints). The occluded or undetected keypoints (cyan keypoints) are reprojections after 3D multi-view reconstruction. Notice that these reprojections do not always align with the images, like the right hand in the last view, which is probably due to OpenPifPaf not being perfectly accurate. + +tector combined with multi-view reconstruction to obtain an initial set of incomplete 3D whole-body keypoints. Next, we implement a completion network to fill in the keypoints missed by the multi-view geometric approach. Then, we develop a refinement method for the hands and the face to obtain more accurate keypoints. Finally, we perform quality assessment to select 25k 3D whole-body poses with high confidence and the 100k associated images from 4-view. + +# 3.1. Initial 3D whole-body dataset with OpenPifPaf + +We run the 2D whole-body detector from OpenPif-Paf [47] on all the 4 views from the training set of Human3.6M (S1, S5, S6, S7 and S8, 1 image per 5 frames). Since the cameras of Human3.6M are well calibrated, we can reconstruct keypoints in 3D using standard multi-view geometry. + +The OpenPifPaf 2D whole-body detector can miss keypoints due to self-occlusions (hands, feet) or unfavorable camera viewpoints (facial landmarks). However, the four-view setup allows us to recover missing keypoints and obtain a complete 3D whole-body pose, provided each keypoint appears in at least two non-opposing views. An example of this process is shown in Figure 2. Using this method, we obtained 11,426 fully complete 3D whole-body poses with all 133 keypoints and 26,333 incomplete 3D whole-body poses where all keypoints appear in at least one view, resulting in a total of 37,759 3D whole-body poses with each keypoint appearing in at least one view. + +We did not rely on the video information because the reconstruction problem becomes significantly more difficult when there is a motion between two frames. In the absence of motion, an additional frame does not help solve the occlusion problem. The results of our study demonstrate that multi-view labeling is sufficiently effective for our task (see Table 2, "Geometry" line). + +![](images/c90ee42565437b0f8f4c0ae560ef5e062d578337d9277b28f83963ef941ea4d7.jpg) +Figure 3. The completion network consists of one linear input layer, 4 transformer encoder blocks (each of them containing 2 transformer encoder layer with $d\_model = 64$ and $n\_head = 1$ ), and a linear output layer. At the end of each encoder block, the features are decoded by the output layer into a predicted position in a curriculum way where later blocks decode more keypoints. + +![](images/59eb85bf0820086b5487f607946a9d30b6fba0d4af33000117121f833f2ea449.jpg) +Figure 4. Example outputs of the completion network. The orange color denotes the keypoints that were detected by OpenPifPaf. The cyan color shows the missed keypoints by OpenPifPaf but completed by our completion network. The left hand is detected in only 1 view by OpenPifPaf and thus fully predicted by the completion network. + +# 3.2. Completion network + +In order to complete the 26,333 incomplete 3D whole-body poses, we develop a completion network as shown in Figure 3. We designed our completion network using Transformer architecture [70] as they can easily handle the conditional dependencies introduced by the skeleton's topology through masking. Since each skeleton always has exactly 133 keypoints, which can be considered as 133 tokens of 3 coordinate values. Token values are expanded from 3 coordinates to $3 \times 16 = 48$ features using Fourier encoding. We use learnable positional encoding since each keypoint is uniquely identified. + +We train the completion network on the 11,426 complete skeletons using a masked auto-encoder strategy [32] where the missing keypoints are masked at the input and will be predicted using the unmasked keypoints. The masking strategy is as follows: + +- With a $50\%$ chance, we perform a keypoint wise mask + +where each keypoint has $15\%$ chance of being masked, + +- with the remaining $50\%$ chance, we perform a block wise mask in which either the body, the left hand, the right hand, the left or the right part of the face are masked (uniform probability). + +To ease the learning process and take into account the causal link between some keypoints (e.g., the tip of a finger depends on the position of its parent phalanges), we introduce a curriculum approach. We compute the loss at different levels following a hierarchy where early levels consider only keypoints closer to the root, while later levels consider more deformable keypoints which highly depend on their parents. We illustrate the completion network and learning process in Figure 3. The loss function is + +$$ +\begin{array}{l} \mathcal {L} (X, X _ {g t 3 D}, X _ {g t 2 D}) = \mathcal {L} _ {3 D} (X, X _ {g t 3 D}) \\ + \alpha \mathcal {L} _ {2 D} (X, X _ {g t 2 D}) \\ + \beta \mathcal {L} _ {\text {s y m}} (X), \tag {1} \\ \end{array} +$$ + +where $\mathcal{L}_3D$ is an $\ell_1$ loss of 3D coordinates, $\mathcal{L}_2D$ is an $\ell_1$ loss of 2D projection of the 3D coordinates if we have the 2D annotation from OpenPifPaf, and $\mathcal{L}_{sym}$ is a symmetric loss which is applied to make sure the left part and right part of the human have the same length on corresponding body parts. + +We show an example output from our completion network in Figure 4. The completion network results on missing body parts are visually realistic and appealing. However, since the completion network does not rely on the image content, its output does not always align with the image + +![](images/8186813d0a00cae85898e215ffbacc269bbbcdea2a45ec07e5c6e5f4ec3d60f0.jpg) +Figure 5. Refinement network architecture and training process. Gaussian noise is added to the groundtruth coordinates with increasing variance, and the network is iteratively trained to recover the less noisy coordinates. + +![](images/812730bc15fa31542f2333210c7659bcb5a9382badfea8201d0269a29593f2ba.jpg) +Figure 6. Example outputs from the face (top row) and hand (bottom row) refinement networks during inference time. We observe that the predictions almost converge to the correct locations in 5-iteration. + +![](images/1aa1b698d6d8c39350d1eb21896b1f59fc1cbc5393b505b32270bf5748f6187c.jpg) + +![](images/ef90dd24ea822845f670e725b88db5893cb7c52ac2f8cb6ac7e467297780f1dc.jpg) + +![](images/66b3ef0869e0545bddbc02ef2b55c8a4e7f086ca07a5de673ee934dfc5edb0d5.jpg) + +![](images/b701b4be651867ef5f4300293d84dfef82f6f2d943b1a092ea460c297677502f.jpg) + +![](images/b03a7dcff89ef6a2abdac401749abfb4723c609361f6f61937a983d057305c18.jpg) +Figure 7. Examples of the 3D whole body skeleton. They are visually realistic humans. The strange looking faces (fatter or thinner) in different views are due to viewing artifacts of the default perspective projection. + +![](images/65c139ca9019b262e7a8a5c628490692d408503841d2ac2d9f886ab79f85dc6c.jpg) + +![](images/9689e8d23e4ef3256860f35c41edd36fb4bc051cb754c112c981b0401121b9e6.jpg) + +![](images/5d16a1484e229f1c5fb37d83880c8129e01f6117cabaa6a05025b7e1bc0fcf6a.jpg) + +and may only reflect the most common poses of the training set. This can quantitatively be seen in the line “+ Completion” of Table 2. + +# 3.3. Hands and face 2D refinements + +In order to correct the alignment problem, we propose another neural network that refines the 2D position of keypoints on the face and the hands. Previous studies have explored and demonstrated the effectiveness of 2D human pose refinement using an iterative error feedback framework [11]. Motivated by this, we build upon recent conditional diffusion models [35] and we consider the prediction from the completion network as noisy such that the refinement network denotes it to conditionally fit the image. + +We train separate refinement models for the face and the hands, while keeping the same network architecture and the same training strategy. We used a simple MLP and found it to be effective, preventing the need to explore more complex architectures. We illustrate the refinement process in Figure 5. During training, we add Gaussian noise to the groundtruth poses with an increasing variance from 5 to 25 pixels, and annotate them as step $t = 1 \dots 5$ (step $t = 0$ is the groundtruth). The network learns to predict the pose at step $t$ given the image and the noisier step $t + 1$ with a 2D supervision loss. + +We build two small datasets, each consisting of 22,000 non-occluded faces and hands respectively, with their corresponding OpenPifPaf predictions. Each image is resized to $384 \times 384$ pixels. We use a random crop of size $224 \times 224$ pixels to have the face and hands located in diverse regions of the images. We split the datasets into training and validation sets with 20,000 images and 2,000 images, respectively. + +Quantitatively, the face predictions achieve an average + +error less than 3 pixels and the hand predictions achieve an average error less than 7 pixels on the validation sets. We show example qualitative results in Figure 6. + +Finally, we run the refinement networks on the 2D-projections of the 3D poses predicted by our completion network. For each 3D skeleton, we project it into the 4 different 2D views. We then crop the regions around the hands and face and denoise the corresponding predictions using the refinement network with 10 iterations to obtain refined 2D poses in each of the 4 views. + +Although the refinement network is not always correct due to its training on non-occluded faces or hands, we only need 2 non-opposing views to perform geometric reconstruction. Since bad refinements tend to collapse all keypoints into the same location, we select the two non-opposing views with the highest variance in keypoint positions to avoid disruptions caused by occlusions. Using this method, we obtain 151,036 triplets of 3D whole-body keypoints, corresponding image, and 2D projected keypoints from the original set. Examples of resulting 3D whole-body skeletons and their image-aligned 2D counterparts are shown in Figure 7 and Figure 8, respectively. + +# 3.4. Quality assessment + +To select the most accurate triplets from our dataset, we reuse the refinement networks and employ a multi-crop strategy that accounts for the variance of the prediction. We project each 3D whole-body skeleton onto all 4 views, and produce four cropped images for each region of interest around the face and hands. The refinement network is run on these 4 crops, and the resulting predictions are aligned with the original prediction to compute the 2D error compared to the original 2D projection. We score the + +![](images/88e3a0895ec6b6bcc4d8a3794619fdf511795d5a54b52be6203abdf53c0ac0cc.jpg) +Figure 8. Examples of the 3D whole-body skeleton projected in 2D onto their corresponding images. They are visually accurate, though still there are small errors in detail which we do not expect to overcome due to the initial resolution and ambiguity of the images. + +3D skeletons by averaging the errors of all 4 projected views, and select the 5k lowest error skeletons from each subject of Human3.6M (S1, S5, S6, S7, S8) to form the $5\mathrm{k}\times 4(\mathrm{view})\times 5(\mathrm{subject}) = 100\mathrm{k}$ triplets of {image, 2D coordinates, 3D coordinates in camera space} of our 3D whole-body dataset. + +![](images/e396a735b6833949744e58a7cc5d5daebac14a18b032ff4d143d0b9db6928242.jpg) + +![](images/ba06ddb24d0ebd65ec0c05339fc52ccf3143079f80f78442cc13c12df00bcdfa.jpg) + +![](images/f77eac3dfb9a2fbe6bb732f41a339861a14e9e70e6f875f299ac31f2022e2339.jpg) +Figure 9. 3D error distributions calculated from 80k manually corrected keypoint annotations. 3D error distributions are presented for whole-body, body, hand and face in mm. We observe that 3D errors are mostly concentrated between 10mm and 20mm. + +![](images/f4d4cf203561e59006798740cb78aa13631d658369392a1831ad2bf9ec668bdf.jpg) + +To assess the quality of the H3WB dataset, we conducted a cross-check study on 600 randomly selected images from the dataset. In this study, annotators were presented an image with the 2D projection of the 3D skeleton on top and were asked to manually correct mis-aligned keypoints by drag and drop. Using multi-view geometry, we reconstructed these corrected skeletons in 3D and compared them to our original skeletons. To validate our process, we show + +![](images/a728589516765baee46eb5894551e626ebdb3108306cee363a860b425c6c66e9.jpg) +Figure 10. Per keypoints error statistics. Please zoom in. + +the influence of each step in Table 2. The geometric approach produced good results but unfortunately cannot provide a large enough dataset. The completion step allows to obtain all labels but at the cost of degraded accuracy due to lack of alignment as explained in section 3.2. The diffusion recovers the original accuracy of the geometric approach. $2\mathrm{mm}$ difference is irrelevant given the initial resolution of the images. We obtain a final average error of $17\mathrm{mm}$ which is very accurate for such a difficult task, and leads to a benchmark which we believe will not be saturated until methods reach around $35\mathrm{mm}$ . + +
Steps# keypoints available3D error (mm)
AllBodyFaceHand
Geometry4812714.8717.7213.2915.87
+ Completion7980029.3125.5726.0236.67
+ Diffusion7980016.9818.6315.0819.16
+ +Table 2. Quantitative analysis of each intermediate step in our pipeline. + +3D error distributions: In Figure 9, we illustrate 3D + +error distributions in mm for all whole-body keypoints and each whole-body part keypoints (i.e. body, hand and face) separately. The distributions of errors are well concentrated around low values. + +Per keypoint errors: All per keypoint error statistics are shown in Figure 10. $98.3\%$ of the images are below 5cm error (97.4% for body joints, 99.7% hands, and 96.4% face). Similarly, $82.6\%$ of the images are below 2cm error (75.5% for body joints, 89.2% hands, and 75.8% face). + +# 3.5. TotalCapture 3D WholeBody Dataset + +In addition to H3WB dataset, we also prepare whole-body annotations for training sequence of poses from TotalCapture [43] dataset using our proposed multi-view pipeline. We call this dataset TotalCapture 3D WholeBody, or T3WB for short. To create TotalCapture 3D WholeBody, we first obtain 2D whole-body keypoints from OpenPif-Paf [47]. At this stage we discard the frames without any human. Next, we finetune the completion network initialized by H3WB weights using 7000 samples of TotalCapture with complete 8-views. At the end we obtain 125,960 triplets of 3D whole-body keypoint, corresponding image and 2D projected keypoints. This shows our pipeline can be used with any multiview dataset. + +We do not conduct a quality assessment study for T3WB and therefore we cannot guarantee the precision of the annotations. Instead, we use T3WB together with H3WB for training models. To alleviate the effect of the noisy annotations in T3WB, we sample more from H3WB than T3WB in each batch. More specifically, we follow 4:1 ratio for each batch during $H3WB + T3WB$ trainings. + +# 4. The H3WB benchmark + +We use the H3WB dataset to propose a benchmark and the associated leaderboard. We split the dataset into training and test sets. The training set contains all samples from S1, S5, S6 and S7, including 80k {image,2D,3D} triplets. The test set contains all samples from S8, including 20k triplets. The test set labels are retained to prevent involuntary overfitting on the test set. Evaluation is accessible only by submitting results to the maintainers. We do not provide a validation set. We encourage researchers to report 5-fold cross-validation average and standard deviation (see supplementary). + +The corresponding benchmark has 3 different tasks: + +1. 3D whole-body lifting from complete 2D whole-body skeletons, or $2D \rightarrow 3D$ for short. +2. 3D whole-body lifting from incomplete 2D whole-body skeletons, or $I2D \rightarrow 3D$ for short. +3. 3D whole-body skeleton prediction from image, or $RGB \rightarrow 3D$ for short. + +For each task, we report the following MPJPE (Mean Per Joint Position Error) metrics: + +- MPJPE for the whole-body, the body (keypoint 1-23), the face (keypoint 24-91) and the hands (keypoint 92-133) when whole-body is centered on the root joint, i.e. aligned with the pelvis, which in our case is the middle of two hip joints2, +- MPJPE for the face when it is centered on the nose, i.e. aligned with keypoint 1, +- MPJPE for the hands when hands are centered on the wrist, i.e left hand aligned with keypoint 92 and right hand aligned with keypoint 113. + +To create baselines on each task, we adapt popular methods from the literature by changing the number of keypoints to that of our whole-body dataset. Notice that we keep the training recipes of the original papers to avoid over-fitting to this new benchmark. In practice, we recommend to perform model selection and hyper-parameters tuning using 5-fold cross-validation. + +# 4.1. 3D whole-body lifting from complete 2D whole-body keypoints $(2D\rightarrow 3D)$ + +This task is similar to the standard 3D human pose estimation from 2D keypoints but using whole-body keypoints. The training set contains 80k 2D-3D pairs. The test set contains only a half of all the test samples, i.e. 10k 2D poses3. + +We evaluate 6 methods on this task. SimpleBaseline [53] is a well-established model, consisting of a 6-layer MLP. We propose a modification, replacing the network architecture with an 8-layer MLP, which we call Large Simple-Baseline inspired by CanonPose [72]. Jointformer [52] is a recent transformer-based method. CanonPose is trained only with 2D supervision [72]. We also adapt CanonPose to work with additional 3D supervision by manually creating 3 fixed camera views and rotating the 3D skeletons into the corresponding view before projecting them into 2D, training it with multi-view weak-supervision. Finally, we report results for the parametric model SMPLify-X [60] by running optimizations on each input sample. + +We train SimpleBaseline models using their official training setting as described in [53]. The inputs and targets are normalized by subtracting the mean and dividing by the standard deviation. Similarly, we train CanonPose [72] models following their official training setup where the inputs and targets are centered on the pelvis and scaled by the Forbenius norm. We train the Jointformer model in the two stages as described in [52]. + +
MethodAllBodyFace / aligned†Hand / aligned‡
H3WB
SMPL-X[60]188.9166.0208.3 / 23.7170.2 / 44.4
CanonPose[72]*186.7193.7188.4 / 24.6180.2 / 48.9
SimpleBaseline [53]*125.4125.7115.9 / 24.6140.7 / 42.5
CanonPose[72] w 3D sv.*117.7117.5112.0 / 17.9126.9 / 38.3
Large SimpleBaseline[53]*112.3112.6110.6 / 14.6114.8 / 31.7
Jointformer[52]88.384.966.5 / 17.8125.3 / 43.7
H3WB + T3WB
CanonPose[72]*164.7161.1174.5 / 21.5150.8 / 43.6
SimpleBaseline [53]*115.3114.8109.4 / 15.8125.1 / 33.5
Jointformer[52]81.578.060.4 / 16.2117.6 / 38.8
+ +Table 3. Comparing different methods for $2\mathrm{D}\rightarrow 3\mathrm{D}$ on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with \* output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except $\dagger$ and $\ddagger$ show nose and wrist aligned results for face and hands, respectively. Sv. is supervision. + +SimpleBaseline and CanonPose models output normalized whole-body keypoints which requires re-scaling at inference. We use statistics from the training set to adjust the test predictions. We calculate a scaling factor using the ratio of 3D to 2D bounding boxes. The formula is: $X_{\mathrm{final}} = X_{\mathrm{unit}} \times \overline{\sigma_{3d}} \times \frac{\sigma_{2d}}{\sigma_{2d}}$ , where $X_{\mathrm{unit}}$ is the normalized prediction, $\overline{\sigma_{3d}}$ is the average size of the 3D training boxes, $\sigma_{2d}$ is the size of the current 2D box, and $\overline{\sigma_{2d}}$ is the average size of the 2D training boxes. + +Since SMPLify-X has 144 keypoints with a different layout, we use interpolation to transform between the WholeBody skeleton and SMPL-X and run SMPL-X's optimization for 2,000 iterations (4 minutes/sample). + +We present the results in Table 3. SMPLify-X performs the worst, showing that parametric models struggle more than discriminative approaches. SimpleBaseline[53] is a solid method, and Large SimpleBaseline improves its performance further. CanonPose[72] can be improved with additional 3D supervision, but still performs worse than Large SimpleBaseline. CanonPose also predicts the camera view, and the uncertainty in this prediction can lead to more error. Jointformer[52] achieves the best results among all methods, but still has room for improvement. All methods perform worse on our benchmark than on Human3.6M because of pelvis centering, which creates higher numerical error on extremities like hands and face, the parts that contain most of the whole-body keypoints. + +Additionally, we conducted experiments with SimpleBaseline, CanonPose, and Jointformer, leveraging the merged dataset $H3WB + T3WB$ , which combines both H3WB and T3WB. The results in Table 3 show that when T3WB is integrated with H3WB, the performances are improved significantly. For instance, on all whole-body keypoints, it yields 22 pt, 10.1 pt and 6.8 pt improvement for SimpleBaseline, CanonPose and Jointformer, respectively. + +# 4.2. 3D whole-body lifting from incomplete 2D whole-body keypoints $(I2D\rightarrow 3D)$ + +We propose a second task where we want to obtain 3D complete whole-body poses from 2D incomplete pose. This task aims to simulate the more realistic case when there are occlusions and the 2D whole-body detector outputs an incomplete skeleton. We do not provide masks for the training skeletons to allow for online data-augmentation. Instead, we propose a masking strategy as follows: + +- With $40\%$ probability, each keypoint has a $25\%$ chance of being masked, +with $20\%$ probability, the face is entirely masked, +with $20\%$ probability, the left hand is entirely masked, +- with $20\%$ probability, the right hand is entirely masked. + +The second half of the test set (10k 2D) is devoted to this task. The masking strategy is applied only once on the 2D poses of the test set, which are directly provided as incomplete 2D skeletons for fair comparison between methods. + +The results for the I2D $\rightarrow$ 3D task are shown in Table 4. All methods perform worse than in the $2\mathrm{D}\rightarrow 3\mathrm{D}$ task. SimpleBaseline[53] has low capacity and uses batch normalization that struggles with missing data, resulting in poor performance. The Large SimpleBaseline model, without batch normalization layers, achieves good results for the task's complexity. CanonPose[72] performs poorly due to errors in camera rotation prediction, which are magnified since most of the 133 keypoints are on the face and hands. The addition of 3D supervision partly solves this problem. The transformer-based Jointformer[52] method outperforms others. Sample outputs obtained by Large SimpleBaseline are shown in Figure 11, where predicted skeletons, although not accurate, are realistic. + +Similar to 2D $\rightarrow$ 3D task, we experimented with Simple-Baseline, CanonPose and Jointformer on $H3WB + T3WB$ dataset and obtained significant improvements. For all + +
MethodAllBodyFace / aligned†Hand / aligned‡
CanonPose[72]*285.0264.4319.7 / 31.9240.0 / 56.2
SimpleBaseline[53]*268.8252.0227.9 / 34.0344.3 / 83.4
CanonPose[72] + 3D sv.*163.6155.9161.3 / 22.2171.4 / 47.4
Large SimpleBaseline[53]*131.4131.6120.6 / 19.8148.8 / 44.8
Jointformer[52]109.2103.082.4 / 19.8155.9 / 53.5
H3WB + T3WB
CanonPose[72]*261.5243.3291.3 / 31.3223.1 / 53.7
SimpleBaseline[53]*260.5238.0221.1 / 32.2336.5 / 80.4
Jointformer[52]84.280.159.4 / 16.3126.5 / 44.5
+ +Table 4. Comparing different methods for I2D→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively. Sv. is supervision. + +![](images/8295350e297817f1107eb4c9651d09d26675c62a4f7f45b586afc8cb7d378038.jpg) +Figure 11. Example predictions from Large SimpleBaseline model for task I2D $\rightarrow$ 3D. Colored skeletons correspond to predictions and gray skeletons correspond to groundtruths. + +whole-body keypoints, the combined dataset yields 23.5 pt, 8.3 pt and 25 pt improvement for SimpleBaseline, Canon-Pose and Jointformer, respectively. Those results further validate the effectiveness of our multi-view pipeline for creating whole-body keypoint annotations for any multi-view dataset. + +# 4.3. 3D whole-body pose estimation from a single image $(RGB\rightarrow 3D)$ + +This task is the standard monocular 3D human pose estimation task extended to whole-body pose estimation. We provide a script to split the original Human3.6M videos into images with our indexing in order to establish image-3D correspondences. The training set contains $80\mathrm{k}$ {image paths,3D} pairs, as well as the 2D bounding box of the human in the image. The test set contains all the test samples, including $20\mathrm{k}$ image paths and their 2D bounding boxes. 2D coordinates are not given in order to avoid collisions with $2\mathrm{D}\rightarrow 3\mathrm{D}$ and I2D $\rightarrow 3\mathrm{D}$ . + +For this task, we train 2 two-stage models and 1 single-stage model. For our first two-stage model, we first train a Stacked Hourglass Network (SHN)[59] to predict 2D whole-body keypoints. Then, SimpleBaseline[53] takes 2D keypoint predictions as input and lifts them to 3D space. Similarly, the second two-stage model utilizes CPN[16] to output 2D keypoints and then Jointformer[52] lifts the 2D predictions to obtain 3D whole-body poses. For our single-stage model, we modify the last layer of Resnet50[33] to directly output the 3D whole-body keypoints. We regress the 3D whole-body keypoint coordinates using L1 loss. + +Results in Table 5 show the two-stage $CPN +$ Jointformer model obtains the best results. Our simple single-stage method performs better than the two-stage $SHN +$ SimpleBaseline model. Learning 2D whole-body keypoints is challenging for SHN as very close keypoints on face + +
MethodAllBodyFace / aligned†Hand / aligned‡
RGB→2D+2D→3D:
SHN[59]+SimpleBaseline*182.5189.6138.7 / 32.5249.4 / 64.3
CPN[16]+Jointformer[52]132.6142.891.9 / 20.7192.7 / 56.9
RGB→3D:
Resnet50[33]166.7151.6123.6 / 26.3244.9 / 63.1
DOPE[75]191.3199.7187.3 / 66.0193.3 / 78.2
+ +Table 5. Comparing different methods for RGB→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively. + +and hands may introduce noise to the predicted keypoint heatmaps. The error in the 2D keypoints then makes the lifting task much more challenging. Surprisingly, RGB $\rightarrow$ 3D seems to be harder that the I2D $\rightarrow$ 3D task. Although there are also missing body parts due to self occlusion, RGB $\rightarrow$ 3D contains more contextual information that should allow to better disambiguate the pose. Compared to 2D $\rightarrow$ 3D and I2D $\rightarrow$ 3D, direct prediction of 3D whole-body pose from images remains thus as a challenging task which we hope this benchmark can help improve over time. + +In order to show the importance of training body parts jointly, we evaluate DOPE [75] on our benchmark. Unfortunately, it fails to address occluded body parts only predicts the whole-body keypoints for $35\%$ of the test set. For each missing keypoint, we use the (topological) nearest predicted joint as a proxy. Even so, a disjointed model like DOPE fails to achieve significant accuracy. + +# 5. Conclusion + +In this paper, we introduce the H3WB dataset, which extends the Human3.6M dataset with 2D and 3D keypoint annotations for body, face, and hands, containing 100k images with 133 keypoints with an average accuracy of $17\mathrm{mm}$ . We propose three tasks based on this dataset: 3D whole-body lifting from complete 2D keypoints, 3D whole-body lifting from incomplete 2D keypoints, and 3D whole-body prediction from monocular images. We evaluate several baselines on these tasks and demonstrate promising accuracy, but with room for improvement. Lifting from incomplete 2D skeletons and direct estimation from monocular images remain challenging, and we hope that our dataset and benchmark will spur future research in these areas. + +# Acknowledgments + +This work was supported by ANR project TOSAI ANR-20-IADJ-0009 and Ergonova Conseil, and was granted access to the HPC resources of IDRIS under the allocation 2023-AD011012640R2 and 2023-AD011013267R1 made by GENCI. + +# References + +[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019. 1 +[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2 +[3] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. TOG, 2021. 1 +[4] Dennis Bautembach, Iason Oikonomidis, and Antonis Argyros. Filling the joints: Completion and recovery of incomplete 3d human poses. Technologies, 2018. 3 +[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 3 +[6] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: automatic estimation of 3d human pose and shape from a single image. ECCV, 2016. 2 +[7] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2 +[8] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In ECCV, 2018. 1, 2 +[9] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In CVPR, 2018. 1 +[10] Nicolò Carissimi, Paolo Rota, Cigdem Beyer, and Vittorio Murino. Filling the gaps: Predicting missing joints of human poses using denoising autoencoders. In ECCV Workshops, 2018. 3 +[11] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 5 +[12] Cristian Sminchisescu Catalin Ionescu, Fuxin Li. Latent structured models for human pose estimation. In ICCV, 2011. 2 +[13] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Expnet: Landmark-free, deep, 3d facial expressions. In FG 2018, 2018. 1, 3 +[14] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2 +[15] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2 + +[16] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. CVPR, 2017. 9 +[17] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, 2021. 2 +[18] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 1, 2, 3 +[19] Daniel Crispell and Maxim Bazik. Pix2face: Direct 3d face model estimation. In ICCV Workshops, pages 2512-2518, 2017. 3 +[20] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. CVPR Workshop, 2019. 3 +[21] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In ECCV, pages 534-551, 2018. 3 +[22] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 1 +[23] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Learning complex 3d human self-contact. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1343-1351, 2021. 1 +[24] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9919–9928, 2021. 1 +[25] Mercedes Garcia-Salguero, Javier Gonzalez-Jimenez, and Francisco-Angel Moreno. Human 3d pose estimation with a tilting camera for social mobile robot interaction. Sensors, 2019. 1 +[26] Erik Gartner, Aleksis Pirinen, and Cristian Sminchisescu. Deep reinforcement learning for active human pose estimation. AAAI, 2020. 1 +[27] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, pages 10833-10842, 2019. 1, 2 +[28] Yiwen Gu, Shreya Pandit, Elham Saraee, Timothy Nordahl, Terry Ellis, and Margrit Betke. Home-based physical therapy with an interactive computer vision system. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019. 1 +[29] Liang-Yan Gui, Kevin Zhang, Yu-Xiong Wang, Xiaodan Liang, José MF Moura, and Manuela Veloso. Teaching robots to predict human motion. In IROS, 2018. 1 +[30] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In the wild human pose + +estimation using explicit 2d features and intermediate 3d representations. CVPR, 2019. 1 +[31] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Handsformer: Keypoint transformer for monocular 3d pose estimation of hands and object in interaction. CVPR, 2022. 1 +[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 4 +[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 9 +[34] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. CVPR, 2020. 2 +[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 5 +[36] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPMAI, 2014. 2, 3 +[37] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In ECCV, 2018. 1 +[38] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images in the wild. In CVPR, 2020. 2 +[39] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In ICCV, 2017. 3 +[40] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In ECCV, 2020. 2, 3 +[41] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, 2010. 2 +[42] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015. 2 +[43] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. CVPR, 2018. 2, 3, 7 +[44] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1 +[45] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. ICCV, 2019. 1 +[46] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 1 +[47] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. OpenPifPaf: Composite Fields for Semantic Keypoint Detection and Spatio-Temporal Association. IEEE Transactions on Intelligent Transportation Systems, 2021. 3, 7 + +[48] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2 +[49] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In CVPR, 2017. 2 +[50] Ruixu Liu, Ju Shen, He Wang, Chen Chen, Sen-ching Cheung, and Vijayan Asari. Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In CVPR, 2020. 2 +[51] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 2015. 2 +[52] Sebastian Lutz, Richard Blythman, Koustav Ghosal, Matthew Moynihan, Ciaran Simms, and Aljosa Smolic. Jointformer: Single-frame lifting transformer with error prediction and refinement for 3d human pose estimation. ArXiv, 2022. 7, 8, 9 +[53] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 7, 8, 9 +[54] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 2 +[55] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, 2018. 1, 2 +[56] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single RGB image. ICCV, 2019. 1, 2 +[57] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. ECCV, 2020. 2 +[58] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In CVPR, 2018. 1, 2 +[59] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. ECCV, 2016. 9 +[60] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed Osman, Dimitrios Tzionas, and Michael Black. Expressive body capture: 3D hands, face, and body from a single image. CVPR, 2019. 1, 3, 7, 8 +[61] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. CVPR, 2017. 2 +[62] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1 + +[63] Grégory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net++: Multi-person 2d and 3d pose detection in natural images. TPAMI, 2019. 1, 2 +[64] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2 +[65] Yu Rong, Ziwei Liu, and Chen Change Loy. Chasing the tail in monocular 3d human reconstruction with prototype memory. TIP, 2022. 1 +[66] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. ICCV, 2021. 1, 3 +[67] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 1 +[68] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 1, 3 +[69] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In CVPR, 2017. 3 +[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 4 +[71] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 2 +[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In CVPR, 2021. 2, 7, 8 +[73] Jingbo Wang, Sijie Yan, Bo Dai, and Dahua Lin. Scene-aware generative network for human motion synthesis. In CVPR, 2021. 1 +[74] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 1 +[75] Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, Vincent Leroy, and Grégory Rogez. DOPE: distillation of part experts for whole-body 3d pose estimation in the wild. ECCV, 2020. 1, 3, 9 +[76] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. ICCV, 2021. 3 +[77] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In CVPR, 2019. 1, 2, 3 +[78] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In CVPR, 2020. 2 +[79] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 2 + +[80] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In ECCV, 2020. 1 +[81] Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Andrei Zanfir, and Cristian Sminchisescu. Human synthesis and scene compositing. AAAI, 2019. 1 +[82] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, pages 2354-2364, 2019. 1, 2 +[83] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 3 +[84] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single RGB images. ICCV, 2017. 1, 2 +[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan C. Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single RGB images. ICCV, 2019. 2 \ No newline at end of file diff --git a/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/images.zip b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e8220de791f2be02d33c84c3d048304e57e2173 --- /dev/null +++ b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce56774354d073962a6acc4e0cf0d902db77dd9b18f71003bd9c1acd04206d89 +size 585030 diff --git a/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/layout.json b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0da98dfec8cdb396e9543a3c04d792863e735fa0 --- /dev/null +++ b/2023/H3WB_ Human3.6M 3D WholeBody Dataset and Benchmark/layout.json @@ -0,0 +1,9785 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 487, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 487, + 120 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 487, + 120 + ], + "type": "text", + "content": "H3WB: Human3.6M 3D WholeBody Dataset and Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 144, + 485, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 144, + 485, + 199 + ], + "spans": [ + { + "bbox": [ + 108, + 144, + 485, + 199 + ], + "type": "text", + "content": "Yue Zhu Nermin Samet David Picard LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France {yue.zhu, nermin.samet, david.picard}@enpc.fr Code and dataset: https://github.com/wholebody3d/wholebody3d" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 251, + 289, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 528 + ], + "type": "text", + "content": "We present a benchmark for 3D human whole-body pose estimation, which involves identifying accurate 3D keypoints on the entire human body, including face, hands, body, and feet. Currently, the lack of a fully annotated and accurate 3D whole-body dataset results in deep networks being trained separately on specific body parts, which are combined during inference. Or they rely on pseudogroundtruth provided by parametric body models which are not as accurate as detection based methods. To overcome these issues, we introduce the Human3.6M 3D WholeBody (H3WB) dataset, which provides whole-body annotations for the Human3.6M dataset using the COCO Wholebody layout. H3WB comprises 133 whole-body keypoint annotations on 100K images, made possible by our new multiview pipeline. We also propose three tasks: i) 3D whole-body pose lifting from 2D complete whole-body pose, ii) 3D whole-body pose lifting from 2D incomplete whole-body pose, and iii) 3D whole-body pose estimation from a single RGB image. Additionally, we report several baselines from popular methods for these tasks. Furthermore, we also provide automated 3D whole-body annotations of TotalCapture and experimentally show that when used with H3WB it helps to improve the performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "3D Human pose estimation is the task of localizing human body keypoints in images which is critical to analyze human behavior, expressions, emotions, intentions, and how people communicate and interact with the physical world. As a result, 3D human pose estimation has an important role in several vision tasks and applications such as robotics [28, 25, 29] or augmented/virtual reality [55, 3, 73, 81]. However, to make more accurate predictions about human behaviors, we need more than a few body keypoints. To that end, 3D whole-body pose estimation aims to detect face, hand and foot keypoints in addition to the standard human body keypoints of classical 3D hu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 228, + 393, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 393, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 393, + 239 + ], + "type": "text", + "content": "man pose estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 242, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 434 + ], + "type": "text", + "content": "The lack of accurate 3D datasets has made 3D whole-body pose estimation a challenging task, leading previous works to focus on separate body parts and train separate models on different datasets for 3D body pose [1, 30, 55, 56, 63, 3, 44, 45, 46, 65, 74, 77, 80, 26], 3D hand pose [8, 58, 84, 7, 27, 37, 82, 31], or 3D face landmarks [68, 9, 13]. However, directly ensembling separate body part models during inference suffers from issues arising from datasets' biases, pose and scales, and complex inference pipelines. Distillation from pretrained models has been used to overcome these issues, with FrankMocap [66] using three specialized pretrained models to estimate 65 whole-body keypoints (22 on the body, 40 on the hands, and 3 on the face), and DOPE [75] also using three specialized models to output 139 whole-body keypoints (13 on the body, 42 on the hands, and 84 on the face)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 436, + 548, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 548, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 548, + 652 + ], + "type": "text", + "content": "Alternatively, parametric body models can be fitted to obtain whole-body pose, as has been proposed in Ex-Pose [18], SMPLify-X [60] or Monocular Total Capture (MTC) [77]. While parametric models enable sampling an almost infinite number of keypoints from the mesh [24, 22, 23], their accuracy is usually less than that of detection based methods on fine body parts like hands and feet (see supplementary material for examples from the literature). Indeed, parametric models are tailored for visual applications such as realistic motion generation [62] or avatar capture [67], where a realistic capture of the body shape and pose is more important than fine grain accuracy. Relying on a small set of data-driven pose parameters ensures realism but also limits their flexibility in representing complex unusual poses. For applications where the body shape is not needed but the accuracy of the keypoints is essential, like high performance sport analysis or ergonomics, detection based methods are thus the preferred solution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Furthermore, 3D whole-body pose estimation has not been fully explored in the literature due to the absence of a representative and accurate benchmark. As previously mentioned, existing 3D whole-body methods either rely on specific datasets and models for different body parts, lead-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20166" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 523, + 155 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 523, + 155 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 523, + 155 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 523, + 155 + ], + "type": "image", + "image_path": "25f628b325ee5dc9efdbfe6a0f0299a05223c6d2c26ff5de4853d5ee48af82d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 157, + 517, + 170 + ], + "lines": [ + { + "bbox": [ + 75, + 157, + 517, + 170 + ], + "spans": [ + { + "bbox": [ + 75, + 157, + 517, + 170 + ], + "type": "text", + "content": "Figure 1. The H3WB dataset has 133 whole-body keypoint annotations in 3D as well as their respective projections in 2D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 189, + 289, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 289, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 289, + 285 + ], + "type": "text", + "content": "ing to complex training pipelines and heterogeneous evaluations, or utilize parametric models that prioritize shape capture over highly precise keypoints. In addition, unified methods vary significantly in terms of keypoint layout definition, number of keypoints and distribution of keypoints across body parts (see Table 1). These significant dataset disparities and the absence of a standard benchmark make it challenging to compare methods fairly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 289, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 589 + ], + "type": "text", + "content": "To address the above issues, we propose a new large-scale dataset for accurate 3D whole-body pose estimation called Human3.6M 3D WholeBody, or H3WB for short (see Figure 1). Our dataset extends Human3.6M [36, 12] with 3D whole-body keypoint annotations. It consists of 133 paired 2D and 3D whole-body keypoint annotations for a set of 100k images from Human3.6M, following the same layout used in COCO WholeBody [40]. More specifically, in addition to the standard 17 body keypoints, the dataset has 42 hand keypoints, 6 foot keypoints and 68 facial landmarks. H3WB was automatically created in a 3 step process: We obtain an initial set of 3D annotations using multiview geometry. Then, we trained a masked auto-encoder to complete the initial annotations. Finally, we refine the whole-body keypoints via a diffusion model. A manual annotation of 80K keypoints from 600 images shows our labels have an average error of " + }, + { + "bbox": [ + 46, + 289, + 289, + 589 + ], + "type": "inline_equation", + "content": "17\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 289, + 289, + 589 + ], + "type": "text", + "content": " which suggest the H3WB keypoints are very accurate for such a complex task. We propose 3 tasks and benchmarks on the H3WB dataset for which we provide baselines: i) 3D whole-body pose lifting from a complete 2D whole-body keypoints, ii) 3D whole-body pose lifting from incomplete 2D whole-body keypoints (i.e. 2D whole-body with missing keypoints, which is more realistic), and iii) 3D whole-body pose estimation from a single RGB image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Our contributions can be summarized as follows. 1) We propose a method to create detailed 3D human pose keypoints from multi-view images. 2) We propose H3WB, the first accurate public benchmark dataset for 3D whole-body pose estimation, using the aforementioned method. Our benchmark can easily leverage existing results in 2D and enables the community to build upon existing high-quality 2D detectors on COCO. Unifying the 3D whole-body pose estimation with the COCO 2D benchmark will greatly benefit the research community. 3) We provide baselines for" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 189, + 547, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 547, + 273 + ], + "type": "text", + "content": "the 3 tasks of H3WB, which we believe will encourage the community to explore 3D whole-body pose estimation more and accelerate progress in the field. 4) Additionally, we provide 3D whole-body annotations for the TotalCapture [43] dataset, and show that when combined with the H3WB dataset it improves the performance of pose lifting tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 285, + 391, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 391, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 391, + 298 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 304, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 545 + ], + "type": "text", + "content": "3D Body, hand and face pose estimation. There are two main groups of prominent approaches in 3D human pose estimation. The first group directly estimates 3D body pose from a single RGB image [61, 54, 55, 56, 61, 63]. The second group follows two stage approach where they first localize 2D keypoints and then lift 2D human pose to 3D space [53, 72, 38, 6, 49, 53]. Several optimization based methods [6, 49], utilize 2D keypoints to initialize a parametric model of the human body such as SMPL [51]. Several works attempt to eliminate the requirement of 3D annotations using 2D multi-view supervision to estimate 3D human pose [72, 78] or temporal supervision with video [17, 50]. 3D hand pose estimation methods share similar approaches as the body counterparts. First group of works, estimates hand pose from a single RGB image by directly regressing 3D hand keypoints [79], mesh vertices [27, 48], and parameters of parametric 3D hand models such as MANO [7, 64, 2, 14, 15, 82]. Second groups of works rely on intermediate 2D representations such as 2D keypoints and feature maps [8, 58, 82, 84, 34]. Simi" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 562, + 539, + 701 + ], + "blocks": [ + { + "bbox": [ + 313, + 562, + 539, + 701 + ], + "lines": [ + { + "bbox": [ + 313, + 562, + 539, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 539, + 701 + ], + "type": "table", + "html": "
DatasetSizeKeypointsBodyHandFace
Human3.6M[36]3.6M1717
3DPW[71]51k2424
LSP[41]10k1414
3DHP[71]>1.3M1717
Panoptic[42]1.5M1515
MTC[77]834K2020
InterHand2.6M[57]2.6M2121
FreiHAND[85]37k2121
RHD[84]44K2121
MTC[77]111K2121
TotalCapture[43]1.9M1272116+1674
ExPose[18]33K1442515+1589
H3WB100k1332321+2168
", + "image_path": "3a00357557fb72d26244e78c35fcd107f0ca93504783f86994eef5016b577540.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 701, + 529, + 712 + ], + "lines": [ + { + "bbox": [ + 306, + 701, + 529, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 529, + 712 + ], + "type": "text", + "content": "Table 1. Overview of datasets for 3D human pose estimation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20167" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "larly, predominant 3D face pose estimation methods regress the dense 3D face landmarks [19, 21, 39] and face model parameters [13, 68, 69, 20, 76] based on 3DMM [5]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 289, + 407 + ], + "type": "text", + "content": "3D Whole-body pose estimation. There are a several methods [60, 77, 83, 18, 75, 66] jointly estimating 3D whole-body pose. The first group of works is based on parametric human body models such as Adam [43] and SMPL-X [60]. MTC [77] is based on the Adam model [43], and first gets 2.5D predictions, then optimizes the parameters of Adam. SMPLify-X optimizes the parameter of the SMPL-X model [60] to fit it to 2D keypoints. As a major drawback, optimization-based methods are relatively slow and highly sensitive to parameter initializations. Nonparametric methods [75, 18, 66] follow different approaches to avoid heavy optimization procedure. DOPE [75] and FrankMocap [66] first train separate body, hand, and face models. Next, they combine those models within a learning framework. DOPE [75] curates pseudo-ground truths from separate body models and uses those ground-truths to supervise the distillation model. Similar to DOPE, ExPose [18] first obtains a pseudo-ground truth dataset by fitting SMPL-X model on in-the-wild images, and trains a joint model to output whole-body poses. All these methods utilize many part-based datasets. Moreover, all output different whole-body layouts with a different numbers of whole-body keypoints. FrankMocap, DOPE, and SMPLify-X estimate whole-body pose with 65, 139 and 144 keypoints, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 408, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 289, + 552 + ], + "type": "text", + "content": "Pose completion completes a partially estimated pose by localizing missing keypoints. Carissimi et al. [10] propose a denoising variational autoencoder network to fill the missing keypoints in 2D pose completion. Bautembach et al. [4] selects a small subset of poses from a database based on their distance to an incomplete 3D pose, and replaces the missing keypoints with the corresponding averaged keypoints in the subset. Despite being critical for real-world scenarios, pose completion has not been sufficiently explored due to the lack of annotated keypoint datasets. Our 3D whole-body dataset can facilitate more exploration of 3D pose estimation from 2D incomplete human poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 564, + 162, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 162, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 162, + 577 + ], + "type": "text", + "content": "3. The H3WB dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "content": "In this section, we describe the making of the H3WB dataset1. Our objective is to build a keypoint based 3D whole-body dataset including keypoints on the body, the face and the hands, and propose a benchmark. We use the same keypoint layout as COCO WholeBody dataset [40] with 133 keypoints. To that end, we build on the widely used Human3.6M dataset [36] for which we provide 3D whole-body keypoints. The H3WB building process is as follows: First, we use an off-the-shelf 2D whole-body de" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 547, + 176 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 547, + 176 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 547, + 176 + ], + "type": "image", + "image_path": "1b562750440474c9e446610dc3cd872e9bfdbfa07547960d199c400c847c61e4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 547, + 245 + ], + "type": "text", + "content": "Figure 2. OpenPifPaf detects most of the non-occluded keypoints inside the image (orange keypoints). The occluded or undetected keypoints (cyan keypoints) are reprojections after 3D multi-view reconstruction. Notice that these reprojections do not always align with the images, like the right hand in the last view, which is probably due to OpenPifPaf not being perfectly accurate." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 271, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 271, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 547, + 368 + ], + "type": "text", + "content": "tector combined with multi-view reconstruction to obtain an initial set of incomplete 3D whole-body keypoints. Next, we implement a completion network to fill in the keypoints missed by the multi-view geometric approach. Then, we develop a refinement method for the hands and the face to obtain more accurate keypoints. Finally, we perform quality assessment to select 25k 3D whole-body poses with high confidence and the 100k associated images from 4-view." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 384, + 547, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 547, + 398 + ], + "type": "text", + "content": "3.1. Initial 3D whole-body dataset with OpenPifPaf" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 407, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 547, + 479 + ], + "type": "text", + "content": "We run the 2D whole-body detector from OpenPif-Paf [47] on all the 4 views from the training set of Human3.6M (S1, S5, S6, S7 and S8, 1 image per 5 frames). Since the cameras of Human3.6M are well calibrated, we can reconstruct keypoints in 3D using standard multi-view geometry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 482, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 547, + 626 + ], + "type": "text", + "content": "The OpenPifPaf 2D whole-body detector can miss keypoints due to self-occlusions (hands, feet) or unfavorable camera viewpoints (facial landmarks). However, the four-view setup allows us to recover missing keypoints and obtain a complete 3D whole-body pose, provided each keypoint appears in at least two non-opposing views. An example of this process is shown in Figure 2. Using this method, we obtained 11,426 fully complete 3D whole-body poses with all 133 keypoints and 26,333 incomplete 3D whole-body poses where all keypoints appear in at least one view, resulting in a total of 37,759 3D whole-body poses with each keypoint appearing in at least one view." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "We did not rely on the video information because the reconstruction problem becomes significantly more difficult when there is a motion between two frames. In the absence of motion, an additional frame does not help solve the occlusion problem. The results of our study demonstrate that multi-view labeling is sufficiently effective for our task (see Table 2, \"Geometry\" line)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "text", + "content": "1We consider the feet keypoints as a part of the body." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20168" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 289, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 167 + ], + "type": "image", + "image_path": "c90ee42565437b0f8f4c0ae560ef5e062d578337d9277b28f83963ef941ea4d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "text", + "content": "Figure 3. The completion network consists of one linear input layer, 4 transformer encoder blocks (each of them containing 2 transformer encoder layer with " + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "inline_equation", + "content": "d\\_model = 64" + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "inline_equation", + "content": "n\\_head = 1" + }, + { + "bbox": [ + 46, + 168, + 288, + 236 + ], + "type": "text", + "content": "), and a linear output layer. At the end of each encoder block, the features are decoded by the output layer into a predicted position in a curriculum way where later blocks decode more keypoints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 245, + 269, + 393 + ], + "blocks": [ + { + "bbox": [ + 66, + 245, + 269, + 393 + ], + "lines": [ + { + "bbox": [ + 66, + 245, + 269, + 393 + ], + "spans": [ + { + "bbox": [ + 66, + 245, + 269, + 393 + ], + "type": "image", + "image_path": "59eb85bf0820086b5487f607946a9d30b6fba0d4af33000117121f833f2ea449.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 395, + 288, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 395, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 288, + 462 + ], + "type": "text", + "content": "Figure 4. Example outputs of the completion network. The orange color denotes the keypoints that were detected by OpenPifPaf. The cyan color shows the missed keypoints by OpenPifPaf but completed by our completion network. The left hand is detected in only 1 view by OpenPifPaf and thus fully predicted by the completion network." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 481, + 167, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 167, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 167, + 495 + ], + "type": "text", + "content": "3.2. Completion network" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 500, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 632 + ], + "type": "text", + "content": "In order to complete the 26,333 incomplete 3D whole-body poses, we develop a completion network as shown in Figure 3. We designed our completion network using Transformer architecture [70] as they can easily handle the conditional dependencies introduced by the skeleton's topology through masking. Since each skeleton always has exactly 133 keypoints, which can be considered as 133 tokens of 3 coordinate values. Token values are expanded from 3 coordinates to " + }, + { + "bbox": [ + 46, + 500, + 287, + 632 + ], + "type": "inline_equation", + "content": "3 \\times 16 = 48" + }, + { + "bbox": [ + 46, + 500, + 287, + 632 + ], + "type": "text", + "content": " features using Fourier encoding. We use learnable positional encoding since each keypoint is uniquely identified." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 632, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 288, + 693 + ], + "type": "text", + "content": "We train the completion network on the 11,426 complete skeletons using a masked auto-encoder strategy [32] where the missing keypoints are masked at the input and will be predicted using the unmasked keypoints. The masking strategy is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "content": "- With a " + }, + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 58, + 701, + 288, + 714 + ], + "type": "text", + "content": " chance, we perform a keypoint wise mask" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 72, + 545, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 85 + ], + "type": "text", + "content": "where each keypoint has " + }, + { + "bbox": [ + 325, + 72, + 545, + 85 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 325, + 72, + 545, + 85 + ], + "type": "text", + "content": " chance of being masked," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 87, + 547, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 87, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 87, + 547, + 137 + ], + "type": "text", + "content": "- with the remaining " + }, + { + "bbox": [ + 317, + 87, + 547, + 137 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 317, + 87, + 547, + 137 + ], + "type": "text", + "content": " chance, we perform a block wise mask in which either the body, the left hand, the right hand, the left or the right part of the face are masked (uniform probability)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 141, + 547, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 141, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 141, + 547, + 250 + ], + "type": "text", + "content": "To ease the learning process and take into account the causal link between some keypoints (e.g., the tip of a finger depends on the position of its parent phalanges), we introduce a curriculum approach. We compute the loss at different levels following a hierarchy where early levels consider only keypoints closer to the root, while later levels consider more deformable keypoints which highly depend on their parents. We illustrate the completion network and learning process in Figure 3. The loss function is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 255, + 545, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 255, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 340, + 255, + 545, + 298 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (X, X _ {g t 3 D}, X _ {g t 2 D}) = \\mathcal {L} _ {3 D} (X, X _ {g t 3 D}) \\\\ + \\alpha \\mathcal {L} _ {2 D} (X, X _ {g t 2 D}) \\\\ + \\beta \\mathcal {L} _ {\\text {s y m}} (X), \\tag {1} \\\\ \\end{array}", + "image_path": "e919949b524c4f237666b2589f47dab8ded3186e6b419f2007258afb6b7e0246.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3D" + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": " loss of 3D coordinates, " + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2D" + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": " loss of 2D projection of the 3D coordinates if we have the 2D annotation from OpenPifPaf, and " + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sym}" + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": " is a symmetric loss which is applied to make sure the left part and right part of the human have the same length on corresponding body parts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 376, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 437 + ], + "type": "text", + "content": "We show an example output from our completion network in Figure 4. The completion network results on missing body parts are visually realistic and appealing. However, since the completion network does not rely on the image content, its output does not always align with the image" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 311, + 451, + 544, + 527 + ], + "blocks": [ + { + "bbox": [ + 311, + 451, + 544, + 527 + ], + "lines": [ + { + "bbox": [ + 311, + 451, + 544, + 527 + ], + "spans": [ + { + "bbox": [ + 311, + 451, + 544, + 527 + ], + "type": "image", + "image_path": "8186813d0a00cae85898e215ffbacc269bbbcdea2a45ec07e5c6e5f4ec3d60f0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 531, + 545, + 575 + ], + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 575 + ], + "type": "text", + "content": "Figure 5. Refinement network architecture and training process. Gaussian noise is added to the groundtruth coordinates with increasing variance, and the network is iteratively trained to recover the less noisy coordinates." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 588, + 541, + 663 + ], + "blocks": [ + { + "bbox": [ + 312, + 588, + 541, + 663 + ], + "lines": [ + { + "bbox": [ + 312, + 588, + 541, + 663 + ], + "spans": [ + { + "bbox": [ + 312, + 588, + 541, + 663 + ], + "type": "image", + "image_path": "812730bc15fa31542f2333210c7659bcb5a9382badfea8201d0269a29593f2ba.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 669, + 545, + 712 + ], + "lines": [ + { + "bbox": [ + 304, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 545, + 712 + ], + "type": "text", + "content": "Figure 6. Example outputs from the face (top row) and hand (bottom row) refinement networks during inference time. We observe that the predictions almost converge to the correct locations in 5-iteration." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20169" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 167, + 167 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 167, + 167 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 167, + 167 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 167, + 167 + ], + "type": "image", + "image_path": "1aa1b698d6d8c39350d1eb21896b1f59fc1cbc5393b505b32270bf5748f6187c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 74, + 293, + 167 + ], + "blocks": [ + { + "bbox": [ + 176, + 74, + 293, + 167 + ], + "lines": [ + { + "bbox": [ + 176, + 74, + 293, + 167 + ], + "spans": [ + { + "bbox": [ + 176, + 74, + 293, + 167 + ], + "type": "image", + "image_path": "ef90dd24ea822845f670e725b88db5893cb7c52ac2f8cb6ac7e467297780f1dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 74, + 419, + 167 + ], + "blocks": [ + { + "bbox": [ + 301, + 74, + 419, + 167 + ], + "lines": [ + { + "bbox": [ + 301, + 74, + 419, + 167 + ], + "spans": [ + { + "bbox": [ + 301, + 74, + 419, + 167 + ], + "type": "image", + "image_path": "66b3ef0869e0545bddbc02ef2b55c8a4e7f086ca07a5de673ee934dfc5edb0d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 427, + 75, + 544, + 167 + ], + "blocks": [ + { + "bbox": [ + 427, + 75, + 544, + 167 + ], + "lines": [ + { + "bbox": [ + 427, + 75, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 427, + 75, + 544, + 167 + ], + "type": "image", + "image_path": "b701b4be651867ef5f4300293d84dfef82f6f2d943b1a092ea460c297677502f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 169, + 167, + 262 + ], + "blocks": [ + { + "bbox": [ + 51, + 169, + 167, + 262 + ], + "lines": [ + { + "bbox": [ + 51, + 169, + 167, + 262 + ], + "spans": [ + { + "bbox": [ + 51, + 169, + 167, + 262 + ], + "type": "image", + "image_path": "b03a7dcff89ef6a2abdac401749abfb4723c609361f6f61937a983d057305c18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "type": "text", + "content": "Figure 7. Examples of the 3D whole body skeleton. They are visually realistic humans. The strange looking faces (fatter or thinner) in different views are due to viewing artifacts of the default perspective projection." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 169, + 293, + 262 + ], + "blocks": [ + { + "bbox": [ + 176, + 169, + 293, + 262 + ], + "lines": [ + { + "bbox": [ + 176, + 169, + 293, + 262 + ], + "spans": [ + { + "bbox": [ + 176, + 169, + 293, + 262 + ], + "type": "image", + "image_path": "65c139ca9019b262e7a8a5c628490692d408503841d2ac2d9f886ab79f85dc6c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 301, + 169, + 416, + 266 + ], + "blocks": [ + { + "bbox": [ + 301, + 169, + 416, + 266 + ], + "lines": [ + { + "bbox": [ + 301, + 169, + 416, + 266 + ], + "spans": [ + { + "bbox": [ + 301, + 169, + 416, + 266 + ], + "type": "image", + "image_path": "9689e8d23e4ef3256860f35c41edd36fb4bc051cb754c112c981b0401121b9e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 169, + 543, + 266 + ], + "blocks": [ + { + "bbox": [ + 427, + 169, + 543, + 266 + ], + "lines": [ + { + "bbox": [ + 427, + 169, + 543, + 266 + ], + "spans": [ + { + "bbox": [ + 427, + 169, + 543, + 266 + ], + "type": "image", + "image_path": "5d16a1484e229f1c5fb37d83880c8129e01f6117cabaa6a05025b7e1bc0fcf6a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 316, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 352 + ], + "type": "text", + "content": "and may only reflect the most common poses of the training set. This can quantitatively be seen in the line “+ Completion” of Table 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 359, + 216, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 216, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 216, + 372 + ], + "type": "text", + "content": "3.3. Hands and face 2D refinements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "type": "text", + "content": "In order to correct the alignment problem, we propose another neural network that refines the 2D position of keypoints on the face and the hands. Previous studies have explored and demonstrated the effectiveness of 2D human pose refinement using an iterative error feedback framework [11]. Motivated by this, we build upon recent conditional diffusion models [35] and we consider the prediction from the completion network as noisy such that the refinement network denotes it to conditionally fit the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "text", + "content": "We train separate refinement models for the face and the hands, while keeping the same network architecture and the same training strategy. We used a simple MLP and found it to be effective, preventing the need to explore more complex architectures. We illustrate the refinement process in Figure 5. During training, we add Gaussian noise to the groundtruth poses with an increasing variance from 5 to 25 pixels, and annotate them as step " + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "inline_equation", + "content": "t = 1 \\dots 5" + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "text", + "content": " (step " + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "text", + "content": " is the groundtruth). The network learns to predict the pose at step " + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "text", + "content": " given the image and the noisier step " + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 487, + 288, + 617 + ], + "type": "text", + "content": " with a 2D supervision loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 617, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 701 + ], + "type": "text", + "content": "We build two small datasets, each consisting of 22,000 non-occluded faces and hands respectively, with their corresponding OpenPifPaf predictions. Each image is resized to " + }, + { + "bbox": [ + 46, + 617, + 288, + 701 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 46, + 617, + 288, + 701 + ], + "type": "text", + "content": " pixels. We use a random crop of size " + }, + { + "bbox": [ + 46, + 617, + 288, + 701 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 617, + 288, + 701 + ], + "type": "text", + "content": " pixels to have the face and hands located in diverse regions of the images. We split the datasets into training and validation sets with 20,000 images and 2,000 images, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 702, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 288, + 714 + ], + "type": "text", + "content": "Quantitatively, the face predictions achieve an average" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 316, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 352 + ], + "type": "text", + "content": "error less than 3 pixels and the hand predictions achieve an average error less than 7 pixels on the validation sets. We show example qualitative results in Figure 6." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 353, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 546, + 435 + ], + "type": "text", + "content": "Finally, we run the refinement networks on the 2D-projections of the 3D poses predicted by our completion network. For each 3D skeleton, we project it into the 4 different 2D views. We then crop the regions around the hands and face and denoise the corresponding predictions using the refinement network with 10 iterations to obtain refined 2D poses in each of the 4 views." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "text", + "content": "Although the refinement network is not always correct due to its training on non-occluded faces or hands, we only need 2 non-opposing views to perform geometric reconstruction. Since bad refinements tend to collapse all keypoints into the same location, we select the two non-opposing views with the highest variance in keypoint positions to avoid disruptions caused by occlusions. Using this method, we obtain 151,036 triplets of 3D whole-body keypoints, corresponding image, and 2D projected keypoints from the original set. Examples of resulting 3D whole-body skeletons and their image-aligned 2D counterparts are shown in Figure 7 and Figure 8, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 587, + 419, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 419, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 419, + 600 + ], + "type": "text", + "content": "3.4. Quality assessment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "To select the most accurate triplets from our dataset, we reuse the refinement networks and employ a multi-crop strategy that accounts for the variance of the prediction. We project each 3D whole-body skeleton onto all 4 views, and produce four cropped images for each region of interest around the face and hands. The refinement network is run on these 4 crops, and the resulting predictions are aligned with the original prediction to compute the 2D error compared to the original 2D projection. We score the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20170" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 69, + 500, + 259 + ], + "blocks": [ + { + "bbox": [ + 95, + 69, + 500, + 259 + ], + "lines": [ + { + "bbox": [ + 95, + 69, + 500, + 259 + ], + "spans": [ + { + "bbox": [ + 95, + 69, + 500, + 259 + ], + "type": "image", + "image_path": "88e3a0895ec6b6bcc4d8a3794619fdf511795d5a54b52be6203abdf53c0ac0cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 285 + ], + "type": "text", + "content": "Figure 8. Examples of the 3D whole-body skeleton projected in 2D onto their corresponding images. They are visually accurate, though still there are small errors in detail which we do not expect to overcome due to the initial resolution and ambiguity of the images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 290, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 290, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 290, + 376 + ], + "type": "text", + "content": "3D skeletons by averaging the errors of all 4 projected views, and select the 5k lowest error skeletons from each subject of Human3.6M (S1, S5, S6, S7, S8) to form the " + }, + { + "bbox": [ + 46, + 303, + 290, + 376 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}\\times 4(\\mathrm{view})\\times 5(\\mathrm{subject}) = 100\\mathrm{k}" + }, + { + "bbox": [ + 46, + 303, + 290, + 376 + ], + "type": "text", + "content": " triplets of {image, 2D coordinates, 3D coordinates in camera space} of our 3D whole-body dataset." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 387, + 167, + 471 + ], + "blocks": [ + { + "bbox": [ + 52, + 387, + 167, + 471 + ], + "lines": [ + { + "bbox": [ + 52, + 387, + 167, + 471 + ], + "spans": [ + { + "bbox": [ + 52, + 387, + 167, + 471 + ], + "type": "image", + "image_path": "e396a735b6833949744e58a7cc5d5daebac14a18b032ff4d143d0b9db6928242.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 168, + 387, + 282, + 471 + ], + "blocks": [ + { + "bbox": [ + 168, + 387, + 282, + 471 + ], + "lines": [ + { + "bbox": [ + 168, + 387, + 282, + 471 + ], + "spans": [ + { + "bbox": [ + 168, + 387, + 282, + 471 + ], + "type": "image", + "image_path": "ba06ddb24d0ebd65ec0c05339fc52ccf3143079f80f78442cc13c12df00bcdfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 472, + 166, + 557 + ], + "blocks": [ + { + "bbox": [ + 51, + 472, + 166, + 557 + ], + "lines": [ + { + "bbox": [ + 51, + 472, + 166, + 557 + ], + "spans": [ + { + "bbox": [ + 51, + 472, + 166, + 557 + ], + "type": "image", + "image_path": "f77eac3dfb9a2fbe6bb732f41a339861a14e9e70e6f875f299ac31f2022e2339.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 559, + 288, + 605 + ], + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 605 + ], + "type": "text", + "content": "Figure 9. 3D error distributions calculated from 80k manually corrected keypoint annotations. 3D error distributions are presented for whole-body, body, hand and face in mm. We observe that 3D errors are mostly concentrated between 10mm and 20mm." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 472, + 282, + 557 + ], + "blocks": [ + { + "bbox": [ + 168, + 472, + 282, + 557 + ], + "lines": [ + { + "bbox": [ + 168, + 472, + 282, + 557 + ], + "spans": [ + { + "bbox": [ + 168, + 472, + 282, + 557 + ], + "type": "image", + "image_path": "f4d4cf203561e59006798740cb78aa13631d658369392a1831ad2bf9ec668bdf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "To assess the quality of the H3WB dataset, we conducted a cross-check study on 600 randomly selected images from the dataset. In this study, annotators were presented an image with the 2D projection of the 3D skeleton on top and were asked to manually correct mis-aligned keypoints by drag and drop. Using multi-view geometry, we reconstructed these corrected skeletons in 3D and compared them to our original skeletons. To validate our process, we show" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 301, + 544, + 446 + ], + "blocks": [ + { + "bbox": [ + 310, + 301, + 544, + 446 + ], + "lines": [ + { + "bbox": [ + 310, + 301, + 544, + 446 + ], + "spans": [ + { + "bbox": [ + 310, + 301, + 544, + 446 + ], + "type": "image", + "image_path": "a728589516765baee46eb5894551e626ebdb3108306cee363a860b425c6c66e9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 448, + 529, + 459 + ], + "lines": [ + { + "bbox": [ + 323, + 448, + 529, + 459 + ], + "spans": [ + { + "bbox": [ + 323, + 448, + 529, + 459 + ], + "type": "text", + "content": "Figure 10. Per keypoints error statistics. Please zoom in." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "text", + "content": "the influence of each step in Table 2. The geometric approach produced good results but unfortunately cannot provide a large enough dataset. The completion step allows to obtain all labels but at the cost of degraded accuracy due to lack of alignment as explained in section 3.2. The diffusion recovers the original accuracy of the geometric approach. " + }, + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "inline_equation", + "content": "2\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "text", + "content": " difference is irrelevant given the initial resolution of the images. We obtain a final average error of " + }, + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "inline_equation", + "content": "17\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "text", + "content": " which is very accurate for such a difficult task, and leads to a benchmark which we believe will not be saturated until methods reach around " + }, + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "inline_equation", + "content": "35\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 479, + 547, + 611 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 618, + 545, + 666 + ], + "blocks": [ + { + "bbox": [ + 307, + 618, + 545, + 666 + ], + "lines": [ + { + "bbox": [ + 307, + 618, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 618, + 545, + 666 + ], + "type": "table", + "html": "
Steps# keypoints available3D error (mm)
AllBodyFaceHand
Geometry4812714.8717.7213.2915.87
+ Completion7980029.3125.5726.0236.67
+ Diffusion7980016.9818.6315.0819.16
", + "image_path": "24f158fe78521b3f71b489bddee65ef86048a2caf4d2be9e721bdb8ef3d847f0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 666, + 545, + 688 + ], + "lines": [ + { + "bbox": [ + 306, + 666, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 545, + 688 + ], + "type": "text", + "content": "Table 2. Quantitative analysis of each intermediate step in our pipeline." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "3D error distributions: In Figure 9, we illustrate 3D" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20171" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "error distributions in mm for all whole-body keypoints and each whole-body part keypoints (i.e. body, hand and face) separately. The distributions of errors are well concentrated around low values." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 286, + 180 + ], + "type": "text", + "content": "Per keypoint errors: All per keypoint error statistics are shown in Figure 10. " + }, + { + "bbox": [ + 46, + 120, + 286, + 180 + ], + "type": "inline_equation", + "content": "98.3\\%" + }, + { + "bbox": [ + 46, + 120, + 286, + 180 + ], + "type": "text", + "content": " of the images are below 5cm error (97.4% for body joints, 99.7% hands, and 96.4% face). Similarly, " + }, + { + "bbox": [ + 46, + 120, + 286, + 180 + ], + "type": "inline_equation", + "content": "82.6\\%" + }, + { + "bbox": [ + 46, + 120, + 286, + 180 + ], + "type": "text", + "content": " of the images are below 2cm error (75.5% for body joints, 89.2% hands, and 75.8% face)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 188, + 244, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 188, + 244, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 244, + 201 + ], + "type": "text", + "content": "3.5. TotalCapture 3D WholeBody Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 207, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 286, + 361 + ], + "type": "text", + "content": "In addition to H3WB dataset, we also prepare whole-body annotations for training sequence of poses from TotalCapture [43] dataset using our proposed multi-view pipeline. We call this dataset TotalCapture 3D WholeBody, or T3WB for short. To create TotalCapture 3D WholeBody, we first obtain 2D whole-body keypoints from OpenPif-Paf [47]. At this stage we discard the frames without any human. Next, we finetune the completion network initialized by H3WB weights using 7000 samples of TotalCapture with complete 8-views. At the end we obtain 125,960 triplets of 3D whole-body keypoint, corresponding image and 2D projected keypoints. This shows our pipeline can be used with any multiview dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 362, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 286, + 445 + ], + "type": "text", + "content": "We do not conduct a quality assessment study for T3WB and therefore we cannot guarantee the precision of the annotations. Instead, we use T3WB together with H3WB for training models. To alleviate the effect of the noisy annotations in T3WB, we sample more from H3WB than T3WB in each batch. More specifically, we follow 4:1 ratio for each batch during " + }, + { + "bbox": [ + 46, + 362, + 286, + 445 + ], + "type": "inline_equation", + "content": "H3WB + T3WB" + }, + { + "bbox": [ + 46, + 362, + 286, + 445 + ], + "type": "text", + "content": " trainings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 457, + 181, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 181, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 181, + 470 + ], + "type": "text", + "content": "4. The H3WB benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 478, + 286, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 286, + 609 + ], + "type": "text", + "content": "We use the H3WB dataset to propose a benchmark and the associated leaderboard. We split the dataset into training and test sets. The training set contains all samples from S1, S5, S6 and S7, including 80k {image,2D,3D} triplets. The test set contains all samples from S8, including 20k triplets. The test set labels are retained to prevent involuntary overfitting on the test set. Evaluation is accessible only by submitting results to the maintainers. We do not provide a validation set. We encourage researchers to report 5-fold cross-validation average and standard deviation (see supplementary)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 610, + 268, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 610, + 268, + 621 + ], + "spans": [ + { + "bbox": [ + 59, + 610, + 268, + 621 + ], + "type": "text", + "content": "The corresponding benchmark has 3 different tasks:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 632, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 632, + 286, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 286, + 655 + ], + "type": "text", + "content": "1. 3D whole-body lifting from complete 2D whole-body skeletons, or " + }, + { + "bbox": [ + 55, + 632, + 286, + 655 + ], + "type": "inline_equation", + "content": "2D \\rightarrow 3D" + }, + { + "bbox": [ + 55, + 632, + 286, + 655 + ], + "type": "text", + "content": " for short." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 661, + 286, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 661, + 286, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 661, + 286, + 684 + ], + "type": "text", + "content": "2. 3D whole-body lifting from incomplete 2D whole-body skeletons, or " + }, + { + "bbox": [ + 54, + 661, + 286, + 684 + ], + "type": "inline_equation", + "content": "I2D \\rightarrow 3D" + }, + { + "bbox": [ + 54, + 661, + 286, + 684 + ], + "type": "text", + "content": " for short." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 689, + 286, + 712 + ], + "type": "text", + "content": "3. 3D whole-body skeleton prediction from image, or " + }, + { + "bbox": [ + 54, + 689, + 286, + 712 + ], + "type": "inline_equation", + "content": "RGB \\rightarrow 3D" + }, + { + "bbox": [ + 54, + 689, + 286, + 712 + ], + "type": "text", + "content": " for short." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "For each task, we report the following MPJPE (Mean Per Joint Position Error) metrics:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 106, + 545, + 237 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 317, + 106, + 544, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 544, + 166 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 544, + 166 + ], + "type": "text", + "content": "- MPJPE for the whole-body, the body (keypoint 1-23), the face (keypoint 24-91) and the hands (keypoint 92-133) when whole-body is centered on the root joint, i.e. aligned with the pelvis, which in our case is the middle of two hip joints2," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 171, + 544, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 171, + 544, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 171, + 544, + 196 + ], + "type": "text", + "content": "- MPJPE for the face when it is centered on the nose, i.e. aligned with keypoint 1," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 201, + 545, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 201, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 545, + 237 + ], + "type": "text", + "content": "- MPJPE for the hands when hands are centered on the wrist, i.e left hand aligned with keypoint 92 and right hand aligned with keypoint 113." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 247, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 331 + ], + "type": "text", + "content": "To create baselines on each task, we adapt popular methods from the literature by changing the number of keypoints to that of our whole-body dataset. Notice that we keep the training recipes of the original papers to avoid over-fitting to this new benchmark. In practice, we recommend to perform model selection and hyper-parameters tuning using 5-fold cross-validation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 341, + 544, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 341, + 544, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 544, + 365 + ], + "type": "text", + "content": "4.1. 3D whole-body lifting from complete 2D whole-body keypoints " + }, + { + "bbox": [ + 305, + 341, + 544, + 365 + ], + "type": "inline_equation", + "content": "(2D\\rightarrow 3D)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 372, + 544, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 544, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 544, + 420 + ], + "type": "text", + "content": "This task is similar to the standard 3D human pose estimation from 2D keypoints but using whole-body keypoints. The training set contains 80k 2D-3D pairs. The test set contains only a half of all the test samples, i.e. 10k 2D poses3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 421, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 577 + ], + "type": "text", + "content": "We evaluate 6 methods on this task. SimpleBaseline [53] is a well-established model, consisting of a 6-layer MLP. We propose a modification, replacing the network architecture with an 8-layer MLP, which we call Large Simple-Baseline inspired by CanonPose [72]. Jointformer [52] is a recent transformer-based method. CanonPose is trained only with 2D supervision [72]. We also adapt CanonPose to work with additional 3D supervision by manually creating 3 fixed camera views and rotating the 3D skeletons into the corresponding view before projecting them into 2D, training it with multi-view weak-supervision. Finally, we report results for the parametric model SMPLify-X [60] by running optimizations on each input sample." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 577, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 673 + ], + "type": "text", + "content": "We train SimpleBaseline models using their official training setting as described in [53]. The inputs and targets are normalized by subtracting the mean and dividing by the standard deviation. Similarly, we train CanonPose [72] models following their official training setup where the inputs and targets are centered on the pelvis and scaled by the Forbenius norm. We train the Jointformer model in the two stages as described in [52]." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 683, + 539, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 683, + 539, + 693 + ], + "spans": [ + { + "bbox": [ + 315, + 683, + 539, + 693 + ], + "type": "text", + "content": "2We provide the whole-body keypoint ids in supplementary material." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 693, + 545, + 712 + ], + "type": "text", + "content": "3The other half is reserved for the task I2D " + }, + { + "bbox": [ + 307, + 693, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 307, + 693, + 545, + 712 + ], + "type": "text", + "content": " 3D to prevent access to the missing keypoints." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20172" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 181 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 181 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 181 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 181 + ], + "type": "table", + "html": "
MethodAllBodyFace / aligned†Hand / aligned‡
H3WB
SMPL-X[60]188.9166.0208.3 / 23.7170.2 / 44.4
CanonPose[72]*186.7193.7188.4 / 24.6180.2 / 48.9
SimpleBaseline [53]*125.4125.7115.9 / 24.6140.7 / 42.5
CanonPose[72] w 3D sv.*117.7117.5112.0 / 17.9126.9 / 38.3
Large SimpleBaseline[53]*112.3112.6110.6 / 14.6114.8 / 31.7
Jointformer[52]88.384.966.5 / 17.8125.3 / 43.7
H3WB + T3WB
CanonPose[72]*164.7161.1174.5 / 21.5150.8 / 43.6
SimpleBaseline [53]*115.3114.8109.4 / 15.8125.1 / 33.5
Jointformer[52]81.578.060.4 / 16.2117.6 / 38.8
", + "image_path": "9450472f4e46be61b2bd59dcabd51e778b13310f52527cd201bbbac171885d69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "type": "text", + "content": "Table 3. Comparing different methods for " + }, + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "type": "inline_equation", + "content": "2\\mathrm{D}\\rightarrow 3\\mathrm{D}" + }, + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "type": "text", + "content": " on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with \\* output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except " + }, + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 47, + 182, + 286, + 246 + ], + "type": "text", + "content": " show nose and wrist aligned results for face and hands, respectively. Sv. is supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "text", + "content": "SimpleBaseline and CanonPose models output normalized whole-body keypoints which requires re-scaling at inference. We use statistics from the training set to adjust the test predictions. We calculate a scaling factor using the ratio of 3D to 2D bounding boxes. The formula is: " + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{final}} = X_{\\mathrm{unit}} \\times \\overline{\\sigma_{3d}} \\times \\frac{\\sigma_{2d}}{\\sigma_{2d}}" + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{unit}}" + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "text", + "content": " is the normalized prediction, " + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "inline_equation", + "content": "\\overline{\\sigma_{3d}}" + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "text", + "content": " is the average size of the 3D training boxes, " + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "inline_equation", + "content": "\\sigma_{2d}" + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "text", + "content": " is the size of the current 2D box, and " + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "inline_equation", + "content": "\\overline{\\sigma_{2d}}" + }, + { + "bbox": [ + 47, + 277, + 286, + 385 + ], + "type": "text", + "content": " is the average size of the 2D training boxes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 391, + 286, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 286, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 286, + 438 + ], + "type": "text", + "content": "Since SMPLify-X has 144 keypoints with a different layout, we use interpolation to transform between the WholeBody skeleton and SMPL-X and run SMPL-X's optimization for 2,000 iterations (4 minutes/sample)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 445, + 286, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 286, + 612 + ], + "type": "text", + "content": "We present the results in Table 3. SMPLify-X performs the worst, showing that parametric models struggle more than discriminative approaches. SimpleBaseline[53] is a solid method, and Large SimpleBaseline improves its performance further. CanonPose[72] can be improved with additional 3D supervision, but still performs worse than Large SimpleBaseline. CanonPose also predicts the camera view, and the uncertainty in this prediction can lead to more error. Jointformer[52] achieves the best results among all methods, but still has room for improvement. All methods perform worse on our benchmark than on Human3.6M because of pelvis centering, which creates higher numerical error on extremities like hands and face, the parts that contain most of the whole-body keypoints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Additionally, we conducted experiments with SimpleBaseline, CanonPose, and Jointformer, leveraging the merged dataset " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "H3WB + T3WB" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": ", which combines both H3WB and T3WB. The results in Table 3 show that when T3WB is integrated with H3WB, the performances are improved significantly. For instance, on all whole-body keypoints, it yields 22 pt, 10.1 pt and 6.8 pt improvement for SimpleBaseline, CanonPose and Jointformer, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "4.2. 3D whole-body lifting from incomplete 2D whole-body keypoints " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "(I2D\\rightarrow 3D)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 102, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 545, + 186 + ], + "type": "text", + "content": "We propose a second task where we want to obtain 3D complete whole-body poses from 2D incomplete pose. This task aims to simulate the more realistic case when there are occlusions and the 2D whole-body detector outputs an incomplete skeleton. We do not provide masks for the training skeletons to allow for online data-augmentation. Instead, we propose a masking strategy as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 194, + 545, + 276 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 194, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 545, + 217 + ], + "type": "text", + "content": "- With " + }, + { + "bbox": [ + 317, + 194, + 545, + 217 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 317, + 194, + 545, + 217 + ], + "type": "text", + "content": " probability, each keypoint has a " + }, + { + "bbox": [ + 317, + 194, + 545, + 217 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 317, + 194, + 545, + 217 + ], + "type": "text", + "content": " chance of being masked," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 221, + 526, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 221, + 526, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 221, + 526, + 233 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 317, + 221, + 526, + 233 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 317, + 221, + 526, + 233 + ], + "type": "text", + "content": " probability, the face is entirely masked," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 237, + 544, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 237, + 544, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 237, + 544, + 249 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 317, + 237, + 544, + 249 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 317, + 237, + 544, + 249 + ], + "type": "text", + "content": " probability, the left hand is entirely masked," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 253, + 544, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 544, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 544, + 276 + ], + "type": "text", + "content": "- with " + }, + { + "bbox": [ + 317, + 253, + 544, + 276 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 317, + 253, + 544, + 276 + ], + "type": "text", + "content": " probability, the right hand is entirely masked." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 285, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 332 + ], + "type": "text", + "content": "The second half of the test set (10k 2D) is devoted to this task. The masking strategy is applied only once on the 2D poses of the test set, which are directly provided as incomplete 2D skeletons for fair comparison between methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 333, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 499 + ], + "type": "text", + "content": "The results for the I2D " + }, + { + "bbox": [ + 304, + 333, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 333, + 545, + 499 + ], + "type": "text", + "content": " 3D task are shown in Table 4. All methods perform worse than in the " + }, + { + "bbox": [ + 304, + 333, + 545, + 499 + ], + "type": "inline_equation", + "content": "2\\mathrm{D}\\rightarrow 3\\mathrm{D}" + }, + { + "bbox": [ + 304, + 333, + 545, + 499 + ], + "type": "text", + "content": " task. SimpleBaseline[53] has low capacity and uses batch normalization that struggles with missing data, resulting in poor performance. The Large SimpleBaseline model, without batch normalization layers, achieves good results for the task's complexity. CanonPose[72] performs poorly due to errors in camera rotation prediction, which are magnified since most of the 133 keypoints are on the face and hands. The addition of 3D supervision partly solves this problem. The transformer-based Jointformer[52] method outperforms others. Sample outputs obtained by Large SimpleBaseline are shown in Figure 11, where predicted skeletons, although not accurate, are realistic." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 500, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 536 + ], + "type": "text", + "content": "Similar to 2D " + }, + { + "bbox": [ + 304, + 500, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 500, + 545, + 536 + ], + "type": "text", + "content": " 3D task, we experimented with Simple-Baseline, CanonPose and Jointformer on " + }, + { + "bbox": [ + 304, + 500, + 545, + 536 + ], + "type": "inline_equation", + "content": "H3WB + T3WB" + }, + { + "bbox": [ + 304, + 500, + 545, + 536 + ], + "type": "text", + "content": " dataset and obtained significant improvements. For all" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 309, + 553, + 545, + 644 + ], + "blocks": [ + { + "bbox": [ + 309, + 553, + 545, + 644 + ], + "lines": [ + { + "bbox": [ + 309, + 553, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 309, + 553, + 545, + 644 + ], + "type": "table", + "html": "
MethodAllBodyFace / aligned†Hand / aligned‡
CanonPose[72]*285.0264.4319.7 / 31.9240.0 / 56.2
SimpleBaseline[53]*268.8252.0227.9 / 34.0344.3 / 83.4
CanonPose[72] + 3D sv.*163.6155.9161.3 / 22.2171.4 / 47.4
Large SimpleBaseline[53]*131.4131.6120.6 / 19.8148.8 / 44.8
Jointformer[52]109.2103.082.4 / 19.8155.9 / 53.5
H3WB + T3WB
CanonPose[72]*261.5243.3291.3 / 31.3223.1 / 53.7
SimpleBaseline[53]*260.5238.0221.1 / 32.2336.5 / 80.4
Jointformer[52]84.280.159.4 / 16.3126.5 / 44.5
", + "image_path": "430741569708e1616ae10db9d8ef2ee9eccc5e1610ab1aefe21207cc92747819.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 645, + 545, + 711 + ], + "lines": [ + { + "bbox": [ + 304, + 645, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 545, + 711 + ], + "type": "text", + "content": "Table 4. Comparing different methods for I2D→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively. Sv. is supervision." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20173" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 240 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 240 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 240 + ], + "type": "image", + "image_path": "8295350e297817f1107eb4c9651d09d26675c62a4f7f45b586afc8cb7d378038.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 287, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 274 + ], + "type": "text", + "content": "Figure 11. Example predictions from Large SimpleBaseline model for task I2D " + }, + { + "bbox": [ + 46, + 239, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 239, + 287, + 274 + ], + "type": "text", + "content": " 3D. Colored skeletons correspond to predictions and gray skeletons correspond to groundtruths." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": "whole-body keypoints, the combined dataset yields 23.5 pt, 8.3 pt and 25 pt improvement for SimpleBaseline, Canon-Pose and Jointformer, respectively. Those results further validate the effectiveness of our multi-view pipeline for creating whole-body keypoint annotations for any multi-view dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "text", + "content": "4.3. 3D whole-body pose estimation from a single image " + }, + { + "bbox": [ + 47, + 372, + 287, + 396 + ], + "type": "inline_equation", + "content": "(RGB\\rightarrow 3D)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "content": "This task is the standard monocular 3D human pose estimation task extended to whole-body pose estimation. We provide a script to split the original Human3.6M videos into images with our indexing in order to establish image-3D correspondences. The training set contains " + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "80\\mathrm{k}" + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "content": " {image paths,3D} pairs, as well as the 2D bounding box of the human in the image. The test set contains all the test samples, including " + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "content": " image paths and their 2D bounding boxes. 2D coordinates are not given in order to avoid collisions with " + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "2\\mathrm{D}\\rightarrow 3\\mathrm{D}" + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "content": " and I2D " + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\rightarrow 3\\mathrm{D}" + }, + { + "bbox": [ + 46, + 403, + 287, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 653 + ], + "type": "text", + "content": "For this task, we train 2 two-stage models and 1 single-stage model. For our first two-stage model, we first train a Stacked Hourglass Network (SHN)[59] to predict 2D whole-body keypoints. Then, SimpleBaseline[53] takes 2D keypoint predictions as input and lifts them to 3D space. Similarly, the second two-stage model utilizes CPN[16] to output 2D keypoints and then Jointformer[52] lifts the 2D predictions to obtain 3D whole-body poses. For our single-stage model, we modify the last layer of Resnet50[33] to directly output the 3D whole-body keypoints. We regress the 3D whole-body keypoint coordinates using L1 loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 654, + 287, + 714 + ], + "type": "text", + "content": "Results in Table 5 show the two-stage " + }, + { + "bbox": [ + 45, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "CPN +" + }, + { + "bbox": [ + 45, + 654, + 287, + 714 + ], + "type": "text", + "content": " Jointformer model obtains the best results. Our simple single-stage method performs better than the two-stage " + }, + { + "bbox": [ + 45, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "SHN +" + }, + { + "bbox": [ + 45, + 654, + 287, + 714 + ], + "type": "text", + "content": " SimpleBaseline model. Learning 2D whole-body keypoints is challenging for SHN as very close keypoints on face" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 134 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 134 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 134 + ], + "type": "table", + "html": "
MethodAllBodyFace / aligned†Hand / aligned‡
RGB→2D+2D→3D:
SHN[59]+SimpleBaseline*182.5189.6138.7 / 32.5249.4 / 64.3
CPN[16]+Jointformer[52]132.6142.891.9 / 20.7192.7 / 56.9
RGB→3D:
Resnet50[33]166.7151.6123.6 / 26.3244.9 / 63.1
DOPE[75]191.3199.7187.3 / 66.0193.3 / 78.2
", + "image_path": "79d0b3c1fc73808ad2441a6681e8d6fad21ac4c4b54b0b6397e3d16a91aab27d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 135, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 135, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 135, + 545, + 201 + ], + "type": "text", + "content": "Table 5. Comparing different methods for RGB→3D on H3WB test set. Results are shown for the MPJPE metric in mm. Methods with * output normalized predictions. Results of normalized methods are re-scaled using our scaling formula. All results are pelvis aligned, except † and ‡ show nose and wrist aligned results for face and hands, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": "and hands may introduce noise to the predicted keypoint heatmaps. The error in the 2D keypoints then makes the lifting task much more challenging. Surprisingly, RGB " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": " 3D seems to be harder that the I2D " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": " 3D task. Although there are also missing body parts due to self occlusion, RGB " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": " 3D contains more contextual information that should allow to better disambiguate the pose. Compared to 2D " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": " 3D and I2D " + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 220, + 545, + 340 + ], + "type": "text", + "content": " 3D, direct prediction of 3D whole-body pose from images remains thus as a challenging task which we hope this benchmark can help improve over time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "type": "text", + "content": "In order to show the importance of training body parts jointly, we evaluate DOPE [75] on our benchmark. Unfortunately, it fails to address occluded body parts only predicts the whole-body keypoints for " + }, + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 304, + 340, + 545, + 424 + ], + "type": "text", + "content": " of the test set. For each missing keypoint, we use the (topological) nearest predicted joint as a proxy. Even so, a disjointed model like DOPE fails to achieve significant accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 454, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 611 + ], + "type": "text", + "content": "In this paper, we introduce the H3WB dataset, which extends the Human3.6M dataset with 2D and 3D keypoint annotations for body, face, and hands, containing 100k images with 133 keypoints with an average accuracy of " + }, + { + "bbox": [ + 304, + 454, + 545, + 611 + ], + "type": "inline_equation", + "content": "17\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 454, + 545, + 611 + ], + "type": "text", + "content": ". We propose three tasks based on this dataset: 3D whole-body lifting from complete 2D keypoints, 3D whole-body lifting from incomplete 2D keypoints, and 3D whole-body prediction from monocular images. We evaluate several baselines on these tasks and demonstrate promising accuracy, but with room for improvement. Lifting from incomplete 2D skeletons and direct estimation from monocular images remain challenging, and we hope that our dataset and benchmark will spur future research in these areas." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 621, + 403, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 403, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 403, + 635 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 701 + ], + "type": "text", + "content": "This work was supported by ANR project TOSAI ANR-20-IADJ-0009 and Ergonova Conseil, and was granted access to the HPC resources of IDRIS under the allocation 2023-AD011012640R2 and 2023-AD011013267R1 made by GENCI." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20174" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "text", + "content": "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In CVPR, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 127, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 287, + 159 + ], + "type": "text", + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 205 + ], + "type": "text", + "content": "[3] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabian Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. TOG, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 208, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 287, + 241 + ], + "type": "text", + "content": "[4] Dennis Bautembach, Iason Oikonomidis, and Antonis Argyros. Filling the joints: Completion and recovery of incomplete 3d human poses. Technologies, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 243, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 287, + 286 + ], + "type": "text", + "content": "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 289, + 287, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 287, + 332 + ], + "type": "text", + "content": "[6] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: automatic estimation of 3d human pose and shape from a single image. ECCV, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 367 + ], + "type": "text", + "content": "[7] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 403 + ], + "type": "text", + "content": "[8] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "text", + "content": "[9] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In CVPR, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "text", + "content": "[10] Nicolò Carissimi, Paolo Rota, Cigdem Beyer, and Vittorio Murino. Filling the gaps: Predicting missing joints of human poses using denoising autoencoders. In ECCV Workshops, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 485, + 287, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 539 + ], + "type": "text", + "content": "[11] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4733-4742, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 542, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 287, + 574 + ], + "type": "text", + "content": "[12] Cristian Sminchisescu Catalin Ionescu, Fuxin Li. Latent structured models for human pose estimation. In ICCV, 2011. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[13] Feng-Ju Chang, Anh Tuan Tran, Tal Hassner, Iacopo Masi, Ram Nevatia, and Gerard Medioni. Expnet: Landmark-free, deep, 3d facial expressions. In FG 2018, 2018. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[14] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[15] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. CVPR, 2017. 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "text", + "content": "[17] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consistent 3d human pose and shape from a video. In CVPR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 151, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 193 + ], + "type": "text", + "content": "[18] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In ECCV, 2020. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "text", + "content": "[19] Daniel Crispell and Maxim Bazik. Pix2face: Direct 3d face model estimation. In ICCV Workshops, pages 2512-2518, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[20] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. CVPR Workshop, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 272, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 314 + ], + "type": "text", + "content": "[21] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In ECCV, pages 534-551, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "text", + "content": "[22] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[23] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Learning complex 3d human self-contact. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1343-1351, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[24] Mihai Fieraru, Mihai Zanfir, Silviu Cristian Pirlela, Vlad Olaru, and Cristian Sminchisescu. Aifit: Automatic 3d human-interpretable feedback models for fitness training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9919–9928, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 482, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 524 + ], + "type": "text", + "content": "[25] Mercedes Garcia-Salguero, Javier Gonzalez-Jimenez, and Francisco-Angel Moreno. Human 3d pose estimation with a tilting camera for social mobile robot interaction. Sensors, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 525, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 557 + ], + "type": "text", + "content": "[26] Erik Gartner, Aleksis Pirinen, and Cristian Sminchisescu. Deep reinforcement learning for active human pose estimation. AAAI, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[27] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, pages 10833-10842, 2019. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "text", + "content": "[28] Yiwen Gu, Shreya Pandit, Elham Saraee, Timothy Nordahl, Terry Ellis, and Margrit Betke. Home-based physical therapy with an interactive computer vision system. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[29] Liang-Yan Gui, Kevin Zhang, Yu-Xiong Wang, Xiaodan Liang, José MF Moura, and Manuela Veloso. Teaching robots to predict human motion. In IROS, 2018. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[30] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In the wild human pose" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20175" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "estimation using explicit 2d features and intermediate 3d representations. CVPR, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[31] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Handsformer: Keypoint transformer for monocular 3d pose estimation of hands and object in interaction. CVPR, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 142, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 287, + 174 + ], + "type": "text", + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 176, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 208 + ], + "type": "text", + "content": "[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2015. 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 210, + 286, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 286, + 233 + ], + "type": "text", + "content": "[34] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 235, + 286, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 286, + 256 + ], + "type": "text", + "content": "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 287, + 300 + ], + "type": "text", + "content": "[36] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. TPMAI, 2014. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 303, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 335 + ], + "type": "text", + "content": "[37] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In ECCV, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 338, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 369 + ], + "type": "text", + "content": "[38] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images in the wild. In CVPR, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "text", + "content": "[39] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In ICCV, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 418, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 450 + ], + "type": "text", + "content": "[40] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 453, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 483 + ], + "type": "text", + "content": "[41] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, 2010. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 487, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 529 + ], + "type": "text", + "content": "[42] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 532, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 563 + ], + "type": "text", + "content": "[43] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. CVPR, 2018. 2, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[44] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[45] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. ICCV, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[46] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In CVPR, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[47] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. OpenPifPaf: Composite Fields for Semantic Keypoint Detection and Spatio-Temporal Association. IEEE Transactions on Intelligent Transportation Systems, 2021. 3, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[48] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[49] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In CVPR, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[50] Ruixu Liu, Ju Shen, He Wang, Chen Chen, Sen-ching Cheung, and Vijayan Asari. Attention mechanism exploits temporal contexts: Real-time 3d human pose reconstruction. In CVPR, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[51] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "text", + "content": "[52] Sebastian Lutz, Richard Blythman, Koustav Ghosal, Matthew Moynihan, Ciaran Simms, and Aljosa Smolic. Jointformer: Single-frame lifting transformer with error prediction and refinement for 3d human pose estimation. ArXiv, 2022. 7, 8, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "text", + "content": "[53] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 2, 7, 8, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "type": "text", + "content": "[54] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[55] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, 2018. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 433, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 475 + ], + "type": "text", + "content": "[56] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single RGB image. ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[57] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2.6m: A dataset and baseline for 3d interacting hand pose estimation from a single RGB image. ECCV, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "text", + "content": "[58] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 568, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 590 + ], + "type": "text", + "content": "[59] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. ECCV, 2016. 9" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 591, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 635 + ], + "type": "text", + "content": "[60] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed Osman, Dimitrios Tzionas, and Michael Black. Expressive body capture: 3D hands, face, and body from a single image. CVPR, 2019. 1, 3, 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "text", + "content": "[61] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. CVPR, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[62] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20176" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[63] Grégory Rogez, Philippe Weinzaepfel, and Cordelia Schmid. Lcr-net++: Multi-person 2d and 3d pose detection in natural images. TPAMI, 2019. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[64] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 174 + ], + "type": "text", + "content": "[65] Yu Rong, Ziwei Liu, and Chen Change Loy. Chasing the tail in monocular 3d human reconstruction with prototype memory. TIP, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 207 + ], + "type": "text", + "content": "[66] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "text", + "content": "[67] Shunsuke Saito, Jinlong Yang, Qianli Ma, and Michael J. Black. SCANimate: Weakly supervised learning of skinned clothed avatar networks. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "text", + "content": "[68] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 298, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 287, + 341 + ], + "type": "text", + "content": "[69] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In CVPR, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 375 + ], + "type": "text", + "content": "[70] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 377, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 287, + 419 + ], + "type": "text", + "content": "[71] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "type": "text", + "content": "[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In CVPR, 2021. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 467, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 498 + ], + "type": "text", + "content": "[73] Jingbo Wang, Sijie Yan, Bo Dai, and Dahua Lin. Scene-aware generative network for human motion synthesis. In CVPR, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 501, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 532 + ], + "type": "text", + "content": "[74] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In ECCV, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 535, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 535, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 535, + 287, + 577 + ], + "type": "text", + "content": "[75] Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, Vincent Leroy, and Grégory Rogez. DOPE: distillation of part experts for whole-body 3d pose estimation in the wild. ECCV, 2020. 1, 3, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 579, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 579, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 579, + 287, + 610 + ], + "type": "text", + "content": "[76] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. ICCV, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 644 + ], + "type": "text", + "content": "[77] Donglai Xiang, Hanbyul Joo, and Yaser Sheikh. Monocular total capture: Posing face, body, and hands in the wild. In CVPR, 2019. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 678 + ], + "type": "text", + "content": "[78] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[79] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 308 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[80] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, Bill Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In ECCV, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[81] Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Andrei Zanfir, and Cristian Sminchisescu. Human synthesis and scene compositing. AAAI, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[82] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, pages 2354-2364, 2019. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[83] Yuxiao Zhou, Marc Habermann, Ikhsanul Habibie, Ayush Tewari, Christian Theobalt, and Feng Xu. Monocular real-time full body capture with inter-part correlations. In CVPR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 231, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 261 + ], + "type": "text", + "content": "[84] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single RGB images. ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan C. Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single RGB images. ICCV, 2019. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20177" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_content_list.json b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f37c0081cf21cfa950dc7772304239c21a5bdbb1 --- /dev/null +++ b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_content_list.json @@ -0,0 +1,1704 @@ +[ + { + "type": "text", + "text": "HAL3D: Hierarchical Active Learning for Fine-Grained 3D Part Labeling", + "text_level": 1, + "bbox": [ + 109, + 80, + 857, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fenggen $\\mathrm{Yu}^{1,2*}$ Yiming Qian $^{1}$ Francisca Gil-Ureta $^{1}$", + "bbox": [ + 243, + 133, + 722, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brian Jackson1 Eric Bennett1 Hao Zhang1,2", + "bbox": [ + 294, + 152, + 714, + 169 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Amazon $^{2}$ Simon Fraser University", + "bbox": [ + 323, + 170, + 648, + 188 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 222, + 313, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present the first active learning tool for fine-grained 3D part labeling, a problem which challenges even the most advanced deep learning (DL) methods due to the significant structural variations among the intricate parts. For the same reason, the necessary effort to annotate training data is tremendous, motivating approaches to minimize human involvement. Our labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. To effectively reduce human efforts, we develop two novel features in our tool, hierarchical and symmetry-aware active labeling. Our human-in-the-loop approach, coined HAL3D, achieves close to error-free fine-grained annotations on any test set with pre-defined hierarchical part labels, with $80\\%$ time-saving over manual effort. We will release the finely labeled models to serve the community.", + "bbox": [ + 75, + 255, + 473, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 527, + 209, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic shape segmentation and labeling [8, 27] is a classical problem with numerous applications [18], including shape/part recognition, retrieval, indexing, and attribute transfer. We consider the problem of fine-grained 3D part labeling, a task that has not received as much attention as the typical semantic segmentation into coarse/major parts (e.g., backs, seats, and legs for chairs). While these coarse semantic groups support high-level visual perception, they do not possess sufficient granularity to address shape properties related to motion, function, interaction, or construction. Indeed, fine-grained object parts, e.g., the individual wheels and the small mechanical parts in a swivel leg, are both function-revealing and build-aware (see Fig. 1), with strong ties to how the objects were physically assembled or virtually built by artists. Real-world applications such as product inspection, assembly, customization, and robot interaction all operate on fine-grained object parts.", + "bbox": [ + 75, + 555, + 468, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning to label fine-grained parts is a highly challenging problem due to the great structural variations among the small and intricate parts, even for shapes that belong to the same category, as shown in Fig. 1. On a challenging", + "bbox": [ + 75, + 811, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49ad14b908fdcd3535a0d9333c611b64806e40137f3da5dbb12c277b7a6036d4.jpg", + "image_caption": [ + "Figure 1. Fine-grained parts of the 3D models shown exhibit the kind of geometric and structural complexity and diversity that part labeling has to handle. No existing methods, whether learned or heuristic-based, could obtain close-to-fully-accurate labeling in these challenging cases, while our human-in-the-loop active learning tool can approach full accuracy, barring human errors." + ], + "image_footnote": [], + "bbox": [ + 504, + 220, + 890, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dataset with average part counts per category ranging from 15 to more than 500, state-of-the-art learned models such as [33] can only achieve mean IoU scores below $50\\%$ , out of a maximum of $100\\%$ for ground truth (GT) results. Also challenging is the tremendous data annotation effort needed to train methods to segment and label fine-grained parts. On the other hand, for applications which demand de facto full accuracy, e.g., for diagnoses in medicine and quality assurance in e-commerce, one can hardly rely on a fully automatic method to reach the mark. It is arguable that the only means to attain full accuracy is to have humans validate all the results, as in the case of data annotation.", + "bbox": [ + 496, + 492, + 892, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present the first active learning tool for fine-grained 3D part labeling. Given a set of 3D shapes segmented into fine parts, our labeling tool assigns one of the predefined labels to each part. These input parts are deemed atomic (i.e., indivisible); they can be as low-level as mesh triangles and as high-level as results obtained by an unlabelled semantic segmentation, or mid-level components from an over-segmentation. In general, active learning iterates between automated labeling and human input for label rectification [44]. Compared to conventional learned models with full autonomy, such a human-in-the-loop approach provides a viable option to achieve close to error-free part labeling on test sets, barring human errors1.", + "bbox": [ + 496, + 675, + 893, + 872 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work carried out during internship at Amazon.", + "bbox": [ + 94, + 886, + 349, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1During testing, accuracy is measured against GT labeling which has", + "bbox": [ + 517, + 886, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "865", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e13cce0fef38766eeff728302de76c2575fac608ccbf640bb699e4aba298d5c2.jpg", + "image_caption": [ + "Figure 2. Our human-in-the-loop, hierarchical active learning (HAL3D) tool for fine-grained 3D part labeling. The input consists of a set of test shapes each pre-segmented into parts. The labeling proceeds hierarchically, following a tree structure (left) that organizes the hierarchical part labels, from coarse labels (top) to fine-grained labels (bottom). For a node in the label tree (with label $l =$ \"Regular Leg Base\" in the illustration), the input $I_0$ is the subset of parts, from the entire set of shapes, that are labeled $l$ by its parent. When labeling parts with $l$ , the label proposal module first assigns refined labels for each part. Then, proposals are sorted by the mean label probability over the parts of each shape, with the high-confidence (HC) proposals passed to the verification step, which stops once the rate of failed shapes reaches a threshold. The low-confidence (LC) proposals are passed to the label modification module. Correctly labeled shapes fine-tune the network, while skipped and failed shapes, in set $I_1$ , go to the next iteration for labeling. The iterations terminate when all shapes have passed human validation." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 885, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our interactive labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. Specifically, our system consists of three modules for label proposal, verification, and modification. Label proposals are produced by dynamic graph CNN (DGCNN) [37], with the resulting label probabilities dictating whether to pass the predicted labels to the verification or modification module. Both human-verified and corrected labels are passed back to the label proposal network to fine-tune it. The iteration terminates once all part labels have passed human verification.", + "bbox": [ + 75, + 450, + 472, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the key criterion for success with active learning is the minimization of human effort, we design and incorporate two novel features into our labeling tool:", + "bbox": [ + 76, + 632, + 468, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Hierarchical labeling: We organize all the part labels in a hierarchical tree structure, which guides our prediction-verification-modification process so that labels further down in the hierarchy are dealt with only after their parent labels have been fully verified. This leads to labeling efficiency, since our labeling process is coarse-to-fine, with fewer labels to process in the coarser levels of the hierarchy. In addition, with both", + "bbox": [ + 94, + 693, + 468, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "human users and the prediction network only needing to deal with labels at the same level of the hierarchy rather than across all levels, there is less load on the users to reduce possible errors and the prediction accuracy improves, which reduces the amount of label modifications, which are the most expensive task.", + "bbox": [ + 529, + 450, + 890, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Part symmetry: Since symmetric parts receive the same label, we employ detected symmetries to constrain and facilitate both label verification and modification.", + "bbox": [ + 517, + 553, + 893, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our hierarchical active learning tool is coined HAL3D. We adopt the hierarchical labels from PartNet [20] and train and test our labeling tool on both PartNet, which comes with semantic part segmentations, and Amazon-Berkeley Objects (ABO). ABO is a recently published dataset of high-quality, artist-created 3D models of real products sold online. These models are composed of build-aware connected components, However, since some of these components are too coarse, we first obtain an over-segmentation via convex decomposition [39] and then apply HAL3D.", + "bbox": [ + 496, + 613, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the core components of HAL3D, i.e., the label proposal network being fine-tuned by human feedback, and both hierarchical and symmetry-aware active labeling, by comparing to baselines and performing an ablation. The results demonstrate clear efficiency of HAL3D over labeling tools without active learning and non-hierarchical designs. Overall, HAL3D achieves $80\\%$ speed-up over manual fine-grained labeling, on the PartNet test set. A supplementary video shows our interactive tool at work. We will", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "been produced by humans. To achieve full accuracy, two criteria must be met: (a) the testing human labeler's judgement of what is the correct labeling agrees with the GT; (b) the labeler does not make any error, e.g., due to carelessness or time pressure. We regard both types of violations as \"human errors.\" One may also refer to the former as an inconsistency between human labeling, which is likely to occur near fuzzy part boundaries.", + "bbox": [ + 75, + 827, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "866", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "release the finely labeled ABO models to serve the community pending permission from the dataset owner.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 135, + 218, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most approaches to semantic 3D shape segmentation and labeling have been designed to reason about coarse- or high-level structures and to target fully autonomy. Considerably less work has been devoted to fine-grained segmentation and labeling or human-in-the-loop approaches.", + "bbox": [ + 76, + 160, + 468, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse 3D shape segmentation and labeling: Since the seminal work on learning 3D mesh segmentation and labeling [13] in 2010, many learning methods have been proposed [8]. Prominent supervised methods such as PointNet [24], PointNet++ [25], and DGCNN [37] perform feature learning over point clouds, while MeshCNN [7] develops convolution and pooling layers that operate on mesh edges. These methods, among many more, produce both parts and part labels, but they have only been trained and tested on datasets with coarse segmentations, such as the ShapeNet part [44] and the COSEG datasets [35]; most object categories therein were labeled with only 2-5 parts.", + "bbox": [ + 75, + 236, + 468, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-level parts, e.g., the legs, seats, and backs of chairs, typically possess sufficient structural consistency (e.g., in terms of relative positioning), hence predictability, to facilitate supervised learning. Most unsupervised or weakly supervised approaches also leverage such consistencies. In particular, earlier methods for co-segmentation, i.e., segmenting a set of related shapes altogether, all rely on alignment or clustering, either in the spatial domain [6] or feature spaces [29, 9]. More recently, resorting to DL, AdaCoseg [50] optimizes a group consistency loss defined by matrix ranks. BAE-Net [3] and RIM-Net [21] obtain co-segmentations by learning branched neural fields.", + "bbox": [ + 75, + 417, + 468, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-grained segmentation: Fine-level parts are important in many real-world applications since, whether large or small, each of them has a function to serve and often a motion to carry out. However, the structure consistency possessed by major semantic groups is mostly lost among fine parts. Such parts could be obtained via shape decomposition based on geometric properties such as convexity [12, 39], cylindricity [49], or through primitive fitting using cuboids [42, 30] or quadrics [23, 45], deformed spheres [22, 15]. However, the resulting segmentations are neither semantic nor labeled.", + "bbox": [ + 75, + 598, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most learning-based solutions are unsupervised or self-supervised, mainly because annotating fine-grained parts is an extremely tedious task. Li et al. [43] learns part labeling from noisy online part graphs annotated by human artists. Yu et al. [46] trains a recursive part decomposition network for category-specific, fine-grained, hierarchical shape segmentation. Wang et al. [33] and SHRED [11] develop deep clustering to learn part priors from fine-grained segmentation without part labels, where no semantic labels are pro", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duced on test shapes. Sharma et al. [28] perform projective shape analysis [36], where a fine-grained 3D segmentation is obtained by aggregating segmented multi-view 2D images. Both of these methods operate fully automatically after training, but their performance falls far short of $100\\%$ accuracy: $32.6\\%$ mean IoU score achieved by [28] on PartNet [20] with fine-grained (Level-3) parts (part counts ranging from 4 to 51), and below $50\\%$ mean IoU by [33] on a more challenging dataset with part counts up to $500+$ .", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of operating on low-level shape elements such as points [33] and pixels [28], the neurally-guided shape parser [10] (NGSP) is trained to assign fine-grained semantic labels to regions of a 3D shape with improved labeling accuracy: mean IoU score up to $58\\%$ on PartNet. NGSP parses region label proposals generated by a guide neural network using likelihood modules to evaluate the global coherence of each proposal. It is still designed as a fully automatic method, but the authors did suggest the need for human-in-the-loop approaches to \"scale beyond carefully-curated research datasets to 'in-the-wild' scenarios.\"", + "bbox": [ + 496, + 229, + 892, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hierarchical analysis: A fine-grained shape understanding is also attainable via hierarchical structural analysis. Some early works perform rule-based reasoning to construct symmetry hierarchies [38] and shape grammars [17], while others work with geometric priors, e.g., convexity [2], for hierarchical shape approximation. Similar to fine-grained segmentation, most learned hierarchical models are also unsupervised or weakly supervised, including co-hierarchical analysis [32], hierarchical abstraction [30], recursive implicit fields [21], and inverse constructed solid geometry (CSG) [45, 14]. All of these methods aim for full automation and work well only with shallow hierarchies.", + "bbox": [ + 496, + 397, + 892, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Active learning for part labeling: The classical and core question for active learning [1, 47, 26] is: \"How does one select instances from the underlying data to label, so as to achieve the most effective training for a given level of effort?\" Past works have studied uncertainty and diversity for interactive segmentation and labeling of 3D shapes [35, 5], medical images [31], and large-scale point clouds [40], to name a few. We utilize active learning, aiming to approach full labeling accuracy, regardless whether the test shapes belong to a \"research dataset\" or are \"in-the-wild.\"", + "bbox": [ + 496, + 580, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most closely related to our work is the active framework for region annotations by Yi et al. [44]. Both works share the common goal of producing accurate, human-verified labels. On the other hand, their work imposes a fixed human effort budget, while ours does not. Technically, their work relies on a conditional random field to automatically propagate human labels, while our work integrates a DL module [37] into an active part labeling tool. The only prior work that combines deep and active learning for 3D segmentation [5] focuses on boundary refinement, just as the earlier work [35] whose learning module is constrained $k$ -", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "867", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "means clustering. However, none of these tools were designed to work with fine-grained labels, e.g., they still only operate on four part labels for chairs [5, 35, 44, 41].", + "bbox": [ + 75, + 90, + 467, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, while our work is the first active learning tool for fine-grained 3D part labeling, hierarchical active learning has been employed in other data domains such as documents. In hierarchical text classification[16, 4], text labels can also be organized hierarchically. In comparison, semantic labeling of 3D parts in a shape collection poses various challenges including richer structural variations, part relations, and more costly and delicate user interactions.", + "bbox": [ + 75, + 136, + 467, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 268, + 166, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a shape test set $S$ , where each shape is presegmented into parts, our methods assign a label $l \\in L$ to each part of each shape. The set of labels, $L$ , is predefined and organized in a tree structure such that coarse labels are at the top, and fine-grained labels at the bottom. The labeling progresses hierarchically top-to-bottom, assigning first coarse labels and refining them at each step, achieving high-accuracy fine-grained labeling results.", + "bbox": [ + 75, + 295, + 467, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each node, we use an active learning based tool to iteratively engage human interactions into the part labeling process. With this tool, we obtain a close to error-free part labeling on each node on any test set.", + "bbox": [ + 75, + 416, + 467, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Hierarchical labeling strategy", + "text_level": 1, + "bbox": [ + 76, + 484, + 341, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We organize the part labels into a tree structure based on their semantics, as in PartNet [20]. Following the same terminology, two types of internal tree nodes are defined: an \"AND\" node indicates \"what part(s) it has\" (e.g., a regular base has legs and a runner), while an \"OR\" node indicates \"what type it is\" (e.g., a star- or pedestal-type base). The labeling on \"AND\" nodes is based on part features, while the labeling on \"OR\" nodes depends on group features; see Fig. 3. Each leaf node presents a final part label.", + "bbox": [ + 75, + 508, + 467, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each node of the tree, we combine a deep neural network with an active learning tool to perform part labeling. Consider a single shape $s_i \\in S$ and its set of parts $P_i$ . At the root node, the input are all the parts of shape $s_i$ , that is $P_{root} = P_i$ . After finishing the labeling task at the root node, the set of parts is split into one or more subsets $P_l \\subseteq P_{root}$ corresponding to the children labels $l$ . Then, each subset is input to the respective node for further labeling. By splitting the input set at each node, the size of the input decreases as the depth of the node increases. Finally, the labeling stops when we reach a leaf node.", + "bbox": [ + 75, + 643, + 467, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The complexity of our labeling scheme depends on the structure of the tree. On one extreme, if all labels are direct children of the root node, training at the root would be more complex but there is only one node to train. The other extreme is a binary tree, where each node is trained to distinguish between only two labels, but we have to train more", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nodes, resulting in an increased training time. As we have found the original PartNet trees to have too many internal nodes, we modified this initial tree structure using a heuristic rule: we remove internal \"AND\" nodes which contain less than three children and keep all \"OR\" nodes. This way, our tree structure is better balanced for our task to lead to reduced training time. Examples of our modified tree structures are provided in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Symmetry-assisted labeling", + "text_level": 1, + "bbox": [ + 498, + 222, + 746, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another novel feature to accelerate HAL3D makes use of detectable part symmetries on the test set. Symmetry is ubiquitous in human-made objects and it is common practice for 3D artists to copy-and-paste symmetric or repeated parts when building 3D assets. Symmetry detection has been a well-studied subject in geometry processing [19] and there are a variety of applicable techniques. For simplicity, we simply use sizes of the part oriented bounding boxes as a crude and conservative feature to detect part symmetries.", + "bbox": [ + 496, + 246, + 890, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As symmetric parts are expected to share the same label, we accelerate our active learning in two ways: if a proposal assigns different labels to symmetric parts, we automatically remove it from the \"HC (High confidence)\" set. Then, in the modification module, the user only needs to label one part of the symmetry group and our system automatically propagate it to the other parts. These features improve labeling efficiency by $20\\%$ ; see Section 5.2.", + "bbox": [ + 496, + 383, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Active labeling of fine-grained 3D parts", + "text_level": 1, + "bbox": [ + 498, + 513, + 836, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-grained 3D shape segmentation and annotation is challenging because of the wide structural diversity found in small and intricate parts of each shape. This diversity makes it difficult for existing DL methods to achieve high prediction accuracy. On the other side, active learning techniques are commonly used by researchers to improve the performance of the learned classifier with the help of human efforts. We adopt these techniques and, at each node of our hierarchical labeling strategy, we use a deep active learning based method to achieve high part labeling accuracy. Our method comprises three main modules: the DL-based part label proposal network, the part label verification module, and the part label modification module.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider a node $l$ in our tree structure, given the subset of parts labeled $l$ of each shape, we iteratively update the part labels until they all pass human verification. At each iteration, we first use a DL-based neural network to make part label proposals based on part label probabilities. Then, the HC proposals are passed to the part label verification module for human verification. Next, we collect the LC, i.e., low-confidence, proposals, which are more likely to be inaccurate, and pass them to the label modification module. Finally, the verified proposals and modified proposals are combined and feed back to the proposal network to fine-", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "868", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/81dc34a885c63960c9cb62dd7f78eea27acc172fe3218bed86ce55062b35a313.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 89, + 467, + 284 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a138e88054d68308ee3e38c29f851fe793b2764fee969279cee1b6e4a0be5f2b.jpg", + "image_caption": [ + "Figure 3. Label proposal architectures. Part label prediction is based on global and local, part-level features at the \"AND\" nodes and global features only for the \"OR\" nodes. The \"AND\" nodes would serve to predict the label for each part and the \"OR\" nodes would predict labels for the entire input shape. Notion-wise, $N$ is the number of points, $P$ is the number of parts, $K$ is the feature dimension, and $L$ is the number of labels." + ], + "image_footnote": [], + "bbox": [ + 83, + 286, + 467, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tune the network weights. The active learning iteration will stop if the number of left shapes are less than 40. We explain next how we make HC vs. LC proposals.", + "bbox": [ + 75, + 542, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Label proposal network. Given a set of parts from each shape $s_i \\in S$ , we sample $N = 8,192$ points over the part surfaces and input them into our proposal network. As shown in Fig. 3, our network uses the edge convolution (EdgeConv) module from DGCNN [37] to extract point features from the point clouds. This module consists of five EdgeConv layers whose feature dimensions are 64, 64, 128, 256, and 256, respectively. The point features from all layers are concatenated and passed into a convolution layer. The size of the resulting feature is 512.", + "bbox": [ + 75, + 598, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our tree structure has different types of nodes, we apply two different network structures from DGCNN, where the architecture for \"OR\" nodes is the same as in the original DGCNN classification network. At an \"AND\" node, differently from DGCNN which is the SOTA method to predict point labels based on point features, we use extra max-pooling operations to aggregate the local part features and global shape features, as inspired by NGSP [10]. The concatenated features are fed into the part classifier, which contains 3 MLP layers to predict part labels. As for an", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"OR\" node, labels of the input shape are predicted based on the global features only; see Fig. 3. Since different shapes have different part counts, we set the maximum part count $P_{max} = 150$ during training for batch processing and add zero-padding to the part feature matrix so that all the part feature matrices have the same dimension. More details of our network can be found in the supplementary material.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Label verification. Our goal is to get the labels of each shape verified by a human. To reduce verification and training times, we need to decide which shapes are verified first and how many to verify on each iteration. Shapes in the test set that have similar structure to shapes in the train set are more likely to be labeled well. We can quickly select these well-labeled shapes from the test set and use them to enhance our limited train set, improving the generalization ability of our proposal network. In our method, we first sort the test set based on average predicted part label probabilities from the proposal network of each shape and arrange the test set into batches, each containing $B = 10$ shapes. Users can mark a shape as well-labeled if they see all parts are labeled correctly, or bad-labeled when one part is wrong. We design a user-friendly interface to achieve this, where users can view and rotate a batch of 3D shapes at the same time. The interface also shows the corresponding label colors for reference. Please see supplementary material for more details about our part label verification interface.", + "bbox": [ + 496, + 208, + 892, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the sorted shapes, we need to decide how many of them to show the user as HC proposals before fine-tuning the prediction, to achieve a balance between processing efficiency and utilization of user feedback. To this end, we adopt an adaptive strategy: given the sorted proposals, we stop the verification step when the number of human-verified shapes on a batch is below a threshold, which is set to 4 in our experiments. This simple strategy is effective since early batches tend to have higher labeling accuracy than later batches. This way, the number of high confidence proposals presented to the user would adapt to the classification purity of the current test set.", + "bbox": [ + 496, + 496, + 893, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Label modification. Shapes in the test set whose part structures and geometries differ significantly from those in the train set are less likely to be labeled accurately by the network. These shapes can be regarded as outliers in the label regression task, but they can be especially valuable during fine-tuning. Human intervention is inevitable in this situation to correct the labels. Specifically, the shapes provided to the label modification module are collected from two sets: the $Q_{1} = 20$ LC proposals with the lowest mean label probabilities, and those shapes that failed user verification more than $H = 2$ times. We ask users to modify the incorrect labels using our modification interface.", + "bbox": [ + 496, + 688, + 892, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our part annotation interface, labels are organized hierarchically so users can quickly look up the correct labels.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "869", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d90bc4b873ecf52dd1365638070b576a3653beac36e1c54d3d7a471901c3fdce.jpg", + "image_caption": [ + "Figure 4. Comparison of pre-segmented parts in PartNet and ABO. ABO shapes were pre-segmented into connected components where each component may correspond to multiple labels, e.g., the back and the arm of the chair are grouped into the same component, as indicated by the red color in the second column. Instead, our part labeling operates on the convex pieces obtained by a decomposition. In the first three figures, the colors are used to distinguish between different parts, and we can see that the parts in PartNet are manually segmented and exhibiting imperfect boundaries. In the last figure, the colors reflect semantic labels." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The label-to-color mapping table is also provided to help find the incorrect labels. More details about this interface are provided in the supplementary material.", + "bbox": [ + 75, + 372, + 468, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shapes that have been verified and modified by humans are added into the initial training set together to fine-tune our proposal network. On the other hand, we skip shapes excluded from the HC and LC proposals. These shapes are combined with shapes that failed to pass the verification less than $H$ times to go to the next iteration for labeling.", + "bbox": [ + 75, + 417, + 468, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 525, + 277, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The Stanford PartNet [20] dataset is a common choice for fine-grained shape analysis [34, 10, 28]. We follow NGSP [10] and use the four major categories from PartNet for evaluation: chair, table, lamp, and storage furniture. Fine-grained labels in PartNet are organized into three granularity levels, and we adopt the second level in our experiments as in NGSP [10]. We first filter out noisy shapes with erroneous labels. Then from each category, we take 500 shapes and split them into 50/50/400 as train/val/test sets.", + "bbox": [ + 75, + 551, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate HAL3D on the recent Amazon Berkeley Objects (ABO) dataset, which is composed of about 8,000 high-quality 3D shapes created by artists from real product images, where each 3D model is a combination of build-aware connected components. Unlike the 3D shapes from PartNet in which each part corresponds to one single semantic label, a build-aware component from ABO may be appropriately associated with multiple labels, as shown in Fig. 4. To alleviate this issue, we apply an approximate convex decomposition [39] to further decompose each connected component into a fine-grained, i.e., over-segmented, set of convexes, where each convex piece is deemed to receive one and only one part label. Our active labeling is then applied to the set of convexes. When running the con", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "vex decomposition algorithm, we use the default parameters from the provided code, except for the concavity threshold setting, which is set as 0.2 in our experiments.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As GT labels are unavailable in ABO, three artists were hired to create them for the decomposed convex components. We built a test set containing 1,800 shapes from five categories: chair, table, lamp, storage furniture, and bed, with 200 for storage and 400 each for the other categories.", + "bbox": [ + 496, + 136, + 890, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Active learning systems are typically evaluated from two perspectives: labeling accuracy and the amount of human efforts or involvements. We employ two standard metrics to assess the labeling quality: (1) part label accuracy measuring how many parts are correctly labeled and (2) point cloud segmentation mIoU depicting how much of the shape surface area is correctly labeled.", + "bbox": [ + 496, + 220, + 890, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The amount of human efforts is measured by the consumed time in the steps of label verification and modification. Inspired from [44], we firstly give artists the same training courses and tools, and conduct user studies on Part-Net. We report the total recorded labeling time in Tab. 1. We also collect the operations they used in the user study and model the total human interaction time as a function of the number of processed parts and shapes:", + "bbox": [ + 496, + 325, + 890, + 448 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nt = T _ {v} ^ {p} \\left(P _ {v}, L\\right) + T _ {v} ^ {f} \\left(S _ {v}, L\\right) + T _ {v} ^ {p} \\left(P _ {m} ^ {c}, L\\right) + T _ {m} ^ {p} \\left(P _ {m} ^ {w}, L\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 458, + 890, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{v}$ is the number of correctly labeled parts in verification, $S_{v}$ is the number of incorrectly shapes in verification, $P_{m}^{c}$ is the number of correctly labeled parts in modification, $P_{m}^{w}$ is the number of incorrectly labeled parts in modification, and finally $L$ is the total number of labels.", + "bbox": [ + 496, + 489, + 890, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Function (1) is devised based on several observations: a) users need to go through and verify all the part labels to ensure accuracy; b) they can easily pick up failure cases during verification if any part is incorrectly labeled, without checking other parts; and c) users need to check all the parts in each shape to know whether modification is needed.", + "bbox": [ + 496, + 566, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The formulas and their coefficients in (2) are obtained via linear regression using the total operation and labeling times collected from all annotators in our user study:", + "bbox": [ + 496, + 656, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT _ {v} ^ {p} \\left(P _ {v}, L\\right) = \\left(0. 3 1 + 0. 0 3 * L\\right) * P _ {v},\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 712, + 821, + 729 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} T _ {v} ^ {f} \\left(S _ {v}, L\\right) = (1. 4 7 + 0. 0 5 * L) * S _ {v}, \\\\ \\text {T r} \\left(R _ {v} ^ {f}, L\\right) = \\left(0. 0 4 - 0. 0 2, L\\right) * R _ {v} ^ {f} \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 733, + 890, + 757 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT _ {v} ^ {p} \\left(P _ {m} ^ {c}, L\\right) = \\left(0. 3 1 + 0. 0 3 * L\\right) * P _ {m} ^ {c},\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 753, + 825, + 770 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT _ {m} ^ {p} \\left(P _ {m} ^ {w}, L\\right) = (5. 2 8 + 0. 2 3 * L) * P _ {m} ^ {w},\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 772, + 825, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T_{v}^{p}$ is the checking time for correctly labeled parts, $T_{v}^{f}$ is the verification time for incorrectly labeled shapes, and $T_{m}^{p}$ is the time for part label modification. We use these functions to evaluate the human interaction timing, in seconds, for hyper-parameters analysis on PartNet. We also report additional machine time cost from training and fine-tuning in the supplementary material.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "870", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have implemented our approach in PyTorch. The label proposal network at each node is pre-trained with the Adam optimizer on the training set of PartNet for 250 epochs, where the learning rate is 0.001 and it is decreased by 0.8 every 25 epochs. When fine-tuning the network in HAL3D, the learning rate is fixed to 0.0001, the maximum epoch number is 125, and the validation set is used for early stopping. Our experiments were ran on an NVIDIA Tesla V100 GPU, and the training time varies from 20 minutes to 3 hours according to the number of parts at each node.", + "bbox": [ + 75, + 114, + 470, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Competing methods", + "text_level": 1, + "bbox": [ + 76, + 277, + 267, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare HAL3D, our hierarchical active labeling tool, with two automatic part labeling methods,", + "bbox": [ + 76, + 303, + 468, + 333 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PartNet [20], which employs the PointNet++ [25] backbone for point-wise feature extraction and labels parts by a voting scheme over points belonging to the same part.", + "- NGSP [10], which aggregates point-wise features via PointNet++ [25] to obtain part features and then leverages a likelihood-aware beam search for optimization; this represents the state of the art for fine-grained labeling." + ], + "bbox": [ + 76, + 337, + 468, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "And two variants of our labeling approach,", + "bbox": [ + 96, + 450, + 380, + 465 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our prop. is one such variant without using the manual label verification or modification module; it predicts part labels using the label proposal network only.", + "- DAL3D is a deep active learning variant of our approach without using the hierarchical labeling strategy." + ], + "bbox": [ + 76, + 470, + 468, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Evaluation on PartNet", + "text_level": 1, + "bbox": [ + 76, + 560, + 285, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All methods presented in this section are pre-trained and tested on the same training and test sets of PartNet.", + "bbox": [ + 76, + 584, + 468, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. Tab. 1 compares the three automatic methods, where PartNet is inferior to both NGSP and our label proposal network due to its hand-crafted voting mechanism. Our simple network achieves comparable accuracy compared to NGSP (difference by 0.11 and 3.49 with respect to mIoU and part accuracy, respectively). However, NGSP suffers from significantly longer running time due to the intensive beam search; see the supplementary for a running time comparison. In contrast, our proposal network is effective and efficient, i.e., capable of quickly producing accurate initial labels to reduce the manual workload in verification and modification for active learning.", + "bbox": [ + 75, + 628, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 also shows that even with the sophisticated design in NGSP, the highest accuracy attained is only $70\\%$ . Hence the performance of automatic fine-grained labeling still falls far below that of active learning, where the largest and smallest accuracy gap is $40.77\\%$ and $22.49\\%$ between NGSP and HAL3D, respectively.", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison against the competing labeling methods and variants. In the table, \"AL\" indicates whether a method uses active learning. Since PartNet, NGSP, and Our prop. are automatic methods, their entries are marked with a '-' instead. \"Accu.\" denotes the part label accuracy, \"mIoU\" denotes the mean Intersection-over-Union, and \"Lab. time\" denotes the labeling time (hours in total) for the active learning methods.", + "bbox": [ + 496, + 85, + 890, + 184 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a32a499f1bf9c6e3afc0d0603f59b3d1b65051028652e762d48908fc49cab5d6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodALAccu.↑mIoU↑Lab. time↓(h)
Chair
PartNet [20]-54.9031.04-
NGSP [10]-68.9842.90-
Our prop.-67.4541.34-
DAL3D✓89.8472.305.99
HAL3D✓94.1384.924.34
Table
PartNet [20]-33.2114.92-
NGSP [10]-47.1127.94-
Our prop.-42.7723.50-
DAL3D✓78.5369.829.14
HAL3D✓84.1874.654.41
Storage furniture
PartNet [20]-55.8230.86-
NGSP [10]-70.6045.79-
Our prop.-69.8242.43-
DAL3D✓84.5771.994.81
HAL3D✓93.0988.723.87
Lamp
PartNet [20]-38.2811.02-
NGSP [10]-49.7623.49-
Our prop.-44.7719.27-
DAL3D✓83.1469.354.52
HAL3D✓90.5379.512.16
", + "bbox": [ + 500, + 195, + 890, + 583 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d22fb27ed6f71c25aa8df62f3e5419ef17b7423b225cc67a68efe6901027271.jpg", + "table_caption": [ + "Table 2. Ablation on key modules of HAL3D for chair category." + ], + "table_footnote": [], + "table_body": "
Row IDProp.Hier.Sym.ALLab-T↓Accu↑
2nd----22.0589.16
3rd✓---8.6588.53
4th✓✓✓-6.3793.87
5th✓-✓✓5.9989.84
6th✓✓-✓5.2193.45
7th✓✓✓✓4.3494.13
", + "bbox": [ + 501, + 633, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HAL3D consistently outperforms our DAL3D variant in all metrics, particularly 2.42 hours faster on average on labeling time, demonstrating the importance of our hierarchical labeling order for reducing manual workload. We also provide qualitative comparison in Fig. 5.", + "bbox": [ + 496, + 751, + 890, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study. Tab. 2 verifies the contributions of the main components of HAL3D on minimizing human labeling efforts, a key criterion for success for a human-in-the-loop approach. The second to fifth columns indicate:", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "871", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c1e75107ec5e6391a49c469ec5e33050be31affb8228fbafb28a8d366c373d5.jpg", + "image_caption": [ + "Figure 5. Visual comparison between labeling methods. HAL3D can achieve results approaching GT, barring human errors. The only mis-labeled part by HAL3D here is the bottom panel of the table (red oval), which could be easily labeled as shelf by users." + ], + "image_footnote": [], + "bbox": [ + 89, + 95, + 460, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c544f81bcf3b438fbe3703bd64fb85931533191054a39af0e796aa615ea1157.jpg", + "image_caption": [ + "Figure 6. A plot of the number of human-verified shapes (max = 400) at different nodes (only those which existed in more than 100 shapes) of the chair category. The plot shows that with hierarchical active learning, many more shapes pass the verification, hence saving on the costly label modification, especially at higher hierarchy levels (i.e., earlier in the labeling iterations)." + ], + "image_footnote": [], + "bbox": [ + 81, + 414, + 467, + 513 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Prop. if the label proposal network is used;", + "- Hier. if the hierarchical labeling strategy is used;", + "- Sym. if the symmetry checking is used.", + "- AL. if the active learning is used." + ], + "bbox": [ + 76, + 627, + 413, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our full approach against five baselines. The second row is a baseline that directly predicts labels by annotating with our label modification interface. The third row is a baseline that firstly uses the label proposal network and then manually modifies incorrect part labels. The fourth, fifth, and sixth rows represent variants of HAL3D without hierarchical labeling, symmetry checking, and active learning, respectively. The results in Tab. 2 clearly justify the design choices in our method, where our full approach (last row) is the most efficient.", + "bbox": [ + 75, + 703, + 467, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further evaluate the effectiveness of our hierarchical labeling, we plot in Fig. 6 the number of shapes which passed human verification over the active learning qua", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/58effd788152898314806f0b9c13f3b1e751ba971bdadd244fd49089d906273c.jpg", + "table_caption": [ + "Table 3. Inference time comparison between NGSP and our label proposal network. We report average seconds per shape." + ], + "table_footnote": [], + "table_body": "
CategoryChairTableLampStorage
NGSP5.41018.0112.6912.87
Our prop.0.0120.0140.0120.011
", + "bbox": [ + 531, + 114, + 859, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions. As we can see, HAL3D consistently verified more shapes as correct at all tree nodes when comparing to the baseline that had no hierarchical design, e.g., more than 300 shapes were marked as correct in the first iteration in the chair_base node. Essentially, our hierarchical verification splits the labeling task into smaller and easier tasks compared to the case of non-hierarchical design, which effectively reduces the burden arising from the subsequent and more costly modification step in terms of human effort.", + "bbox": [ + 496, + 183, + 890, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the supplementary material, we perform ablations on the four hyper parameters and thresholds employed in our method. In all the experiments, we fix these parameters and set them based only on a small dataset. A more careful parameter tuning should lead to better results.", + "bbox": [ + 496, + 319, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference time comparison. Although NGSP achieved higher prediction accuracy than our proposal network, as shown in Tab. 1, our network is more efficient and better suited to be incorporated into an active learning framework. Tab. 3 compares the inference times between NGSP and our proposal network. NGSP is highly time-consuming due to the use of beam search, whereas our label proposal network takes only 12ms per shape on average.", + "bbox": [ + 496, + 414, + 890, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Evaluation on ABO", + "text_level": 1, + "bbox": [ + 500, + 544, + 687, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We apply HAL3D to label our constructed test set for ABO, where the label proposal network is pre-trained on the training set of PartNet. As mentioned in Sec. 4, shapes in ABO are decomposed into convex shapes and HAL3D assigns a semantic label to each decomposed piece. Three artists were hired to use our HAL3D tool and interface for evaluation. Since GT labels are not provided in ABO, we inspect the labeling results from the artists via human cross validation with the final corrected labeling forming the GT. As different artists may report different accuracy and recorded labeling times, we report averages in the paper. The artists worked together with an expert to determine labels for uncertain shapes. The visual results from HAL3D, as shown in Fig. 7 and the supplemental material, were obtained on the test set by taking majority voting from labels provided by different artists.", + "bbox": [ + 496, + 568, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown by quantitative results in Tab. 4, HAL3D can accurately label all categories with a $94\\%$ average accuracy using 3-7 hours of manual annotation, demonstrating the effectiveness and efficiency of our method. The reduction on human efforts against manual annotation becomes more significant when comparing against the last four rows in Tab. 4,", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "872", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b49e1a85c46503b978f5a0c239aa3ccae1f14c0aabbb80d78f39b58a0392ece.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 93, + 189, + 448 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a437c65889b9456358cb69672ebe761430f3c1570568fb031840fd4c643f9943.jpg", + "image_caption": [ + "Figure 7. Visual comparison between several part labeling methods on the ABO dataset." + ], + "image_footnote": [], + "bbox": [ + 233, + 95, + 325, + 446 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3ac5519016de7b6d167d5eed1e5cb3557ea73f3c151f69553067a808903bce2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 95, + 460, + 448 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5f331945387e14b6e310897453f265c0cf60d137af862fc3b00a085a9fa5caca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 95, + 599, + 448 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/41cb17aebc1aba0b33304d3aa25e85e3c488feb7bc23a7b208c823d027dd8b5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 94, + 728, + 448 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a509111814633060d3a97fd8c3017776107e5f0acf8a1d9feb6025e609fd8e60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 94, + 861, + 448 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a925b174c12dcbe1618e9e04ba30a8d06bb859a9f3361ea04aa572be2af4f3cd.jpg", + "table_caption": [ + "Table 4. Comparison for different methods to pass human verification on test set of ABO. The time is in hours." + ], + "table_footnote": [], + "table_body": "
ChairTableLampCabinetBed
# shapes400400400200400
HAL3D
Time ↓Accu. ↑6.4693.077.4192.554.3896.293.3494.484.5193.71
PartNet + modification
Time ↓Accu. ↑28.2588.9633.9486.5131.5390.9113.3590.2425.0289.37
NGSP + modification
Time ↓Accu. ↑28.3088.7734.3286.4031.5691.3713.2290.5624.9089.93
", + "bbox": [ + 84, + 517, + 460, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where we use the same manual modification interface to directly correct the PartNet and NGSP results. Please see supplementary material for results on ABO.", + "bbox": [ + 75, + 709, + 468, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 768, + 196, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present HAL3D, the first semantic labeling tool that is designed to operate on fine-grained 3D parts and achieve close to error-free labeling through full verification by humans, barring human misjudgement. Our human-in-the-loop approach is based on active learning, combining deep label prediction and human inputs to iteratively fine-tune and improve the network prediction. Extensive experiments", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "demonstrate that human efforts are effectively reduced via hierarchical and symmetry-aware active labeling.", + "bbox": [ + 498, + 494, + 890, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "HAL3D is currently implemented with DGCNN [37] as the label prediction backbone, but it can be readily replaced by PointNet [24], which was adopted by PartNet [20] and NGSP [10], or a more performant alternative such as Point Transformer [48]. Aside from symmetry, inter-shape part correspondences can also be utilized to reduce labeling costs, assuming that they can be reliably obtained. In addition, our tool does not yet allow splitting of under-segmented parts. Since performing such operations by humans is time-consuming, incorporating auto-splitting schemes into HAL3D deserves future consideration.", + "bbox": [ + 496, + 527, + 892, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We regard our work as only a preliminary step towards deep active learning for fine-grained structural analysis of 3D shapes. A natural focus for future work is to develop a more sophisticated ranking or selection mechanism for label verification and modification to achieve the most effective training for a given level of human effort.", + "bbox": [ + 496, + 696, + 893, + 789 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 811, + 660, + 829 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank all the anonymous reviewers and area chairs for their constructive comments and the Amazon artists for performing the human annotations. Thanks also go to Jeff-Aram Van Gorp for helpful discussions and other support.", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "873", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Charu C. Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and Philip S. Yu. Active learning: A survey. In Data Classification: Algorithms and Applications, pages 571-597. 3", + "[2] Marco Attene, Michela Mortara, Michela Spagnuolo, and Bianca Falcidieno. Hierarchical convex approximation of 3D shapes for fast region selection. Comput. Graph. Forum, 27(5):1323-1332, 2008. 3", + "[3] Zhiqin Chen, Kangxue Yin, Matt Fisher, Siddhartha Chaudhuri, and Hao Zhang. BAE-NET: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 3", + "[4] Yu Cheng, Kunpeng Zhang, Yusheng Xie, Ankit Agrawal, and Alok Choudhary. On active learning in hierarchical classification. In Proceedings of the 21st ACM International Conference on Information and knowledge Management, pages 2467-2470, 2012. 4", + "[5] David George, Xianghua Xie, Yukun Lai, and Gary KL Tam. A deep learning driven active framework for segmentation of large 3d shape collections. Computer-Aided Design, 144:103179, 2022. 3, 4", + "[6] Aleksey Golovinskiy and Thomas Funkhouser. Consistent segmentation of 3D models. In Proc. Shape Modeling International (SMI), 2009. 3", + "[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: A network with an edge. ACM TOG, 38(4), 2019. 3", + "[8] Yong He, Hongshan Yu, Xiaoyan Liu, Zhengeng Yang, Wei Sun, Yaonan Wang, Qiang Fu, Yanmei Zou, and Ajmal Mian. Deep learning based 3d segmentation: A survey. CoRR, abs/2103.05423, 2021. 1, 3", + "[9] Ruizhen Hu, Lubin Fan, and Ligang Liu. Co-segmentation of 3D shapes via subspace clustering. Comput. Graph. Forum, 31(5):1703--1713, 2012. 3", + "[10] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3D shape regions with approximate inference. In CVPR, 2022. 3, 5, 6, 7, 9", + "[11] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. ACM Transactions on Graphics (TOG), 41(6):1-11, 2022. 3", + "[12] Oliver Van Kaick, Noa Fish, Yanir Kleiman, Shmuel Asafi, and Daniel Cohen-OR. Shape segmentation by approximate convexity analysis. ACM TOG, 34(1). 3", + "[13] Evangelos Kalogerakis, Aaron Hertzmann, and Karan Singh. Learning 3D mesh segmentation and labeling. ACM TOG, 29(4), 2010. 3", + "[14] Kacper Kania, Maciej Zięba, and Tomasz Kajdanowicz. UCSG-NET - unsupervised discovering of constructive solid geometry tree. In NeurIPS, 2020. 3", + "[15] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Neural star domain as primitive representation. Advances in Neural Information Processing Systems, 33:7875-7886, 2020. 3", + "[16] Xiao Li, Da Kuang, and Charles X Ling. Active learning for hierarchical text classification. In Pacific-Asia conference on knowledge discovery and data mining, pages 14–25. Springer, 2012. 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Tianqiang Liu, Siddhartha Chaudhuri, Vladimir G. Kim, Qixing Huang, Niloy J. Mitra, and Thomas Funkhouser. Creating consistent scene graphs using a probabilistic grammar. ACM TOG, 33, 2014. 3", + "[18] Niloy Mitra, Michael Wand, Hao Zhang, Daniel Cohen-Or, and Martin Bokeloh. Structure-aware shape processing. In Eurographics State-of-the-Art Report (STAR), 2013. 1", + "[19] Niloy J. Mitra, Mark Pauly, Michael Wand, and Duygu Ceylan. Symmetry in 3D geometry: Extraction and applications. 32(6), 2013. 4", + "[20] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, June 2019. 2, 3, 4, 6, 7, 9", + "[21] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. RIM-Net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 3", + "[22] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 3", + "[23] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, pages 10344—10353, 2019. 3", + "[24] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 3, 9", + "[25] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 7", + "[26] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B. Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning, 2020. 3", + "[27] Ariel Shamir. A survey on mesh segmentation techniques. Comput. Graph. Forum, 27(6):1539-1556. 1", + "[28] Gopal Sharma, Kangxue Yin, Subhransu Maji, Evangelos Kalogerakis, Or Litany, and Sanja Fidler. MvDeCor: Multiview dense correspondence learning for fine-grained 3D segmentation. In ECCV, 2022. 3, 6", + "[29] Oana Sidi, Oliver van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. ACM TOG, 30(6):126:1-126:9, 2011. 3", + "[30] Chun-Yu Sun, Qian-Fang Zou, Xin Tong, and Yang Liu. Learning adaptive hierarchical cuboid abstractions of 3D shape collections. 38(6), 2019. 3", + "[31] Andrew Top, Ghassan Hamarneh, and Rafeef Abugharbieh. Active learning for interactive 3D image segmentation. In Int. Conf. on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2011. 3", + "[32] Oliver van Kaick, Kai Xu, Hao Zhang, Yanzhen Wang, Shuyang Sun, Ariel Shamir, and Daniel Cohen-Or. Cohierarchical analysis of shape structures. ACM TOG, 32(4), 2013. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "874", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Xiaogang Wang, Xun Sun, Xinyu Cao, Kai Xu, and Bin Zhou. Learning fine-grained segmentation of 3D shapes without part labels. In CVPR, 2021. 1, 3", + "[34] Xiaogang Wang, Bin Zhou, Haiyue Fang, Xiaowu Chen, Qinping Zhao, and Kai Xu. Learning to group and label fine-grained shape components. 37(6), 2018. 6", + "[35] Yunhai Wang, Shmulik Asafi, Oliver van Kaick, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Active co-analysis of a set of shapes. ACM TOG, 31(6), 2012. 3, 4", + "[36] Yunhai Wang, Minglun Gong, Tianhua Wang, Daniel Cohen-Or, Hao Zhang, and Baoquan Chen. Projective analysis for 3D shape segmentation. ACM TOG, 32(6), 2013. 3", + "[37] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM TOG, 2019. 2, 3, 5, 9", + "[38] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. Comput. Graph. Forum, 30(2):287-296, 2011. 3", + "[39] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search. ACM TOG, 41(4), 2022. 2, 3, 6", + "[40] Tsung-Han Wu, Yueh-Cheng Liu1, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H. Hsu. ReDAL: Region-based and diversity-aware active learning for point cloud semantic segmentation. In ICCV, 2021. 3", + "[41] Zizhao Wu, Ruyang Shou, Yunhai Wang, and Xinguo Liu. Interactive shape co-segmentation via label propagation. In Proc. of CAD/Graphics, 2014. 4", + "[42] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM TOG, 40(4), 2021. 3", + "[43] Li Yi, Leonidas Guibas, Aaron Hertzmann, Vladimir G. Kim, Hao Su, and Ersin Yumer. Learning hierarchical shape segmentation and labeling from online repositories. SIGGRAPH, 2017. 3", + "[44] Li Yi, Vladimir G. Kim, I.-Chao Shen Duygu Ceylan, Mengyuan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6), 2016. 1, 3, 4, 6", + "[45] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact CAD shapes with adaptive primitive assembly. In CVPR, 2022. 3", + "[46] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. PartNet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, 2019. 3", + "[47] Xueying Zhan, Qingzhong Wang, Kuan-hao Huang, Haoyi Xiong, Dejing Dou, and Antoni B. Chan. A comparative survey of deep active learning, 2022. 3", + "[48] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In CVPR, 2021. 9" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yang Zhou, Kangxue Yin, Hui Huang, Hao Zhang, Minglun Gong, and Daniel Cohen-Or. Generalized cylinder decomposition. ACM TOG, 34(6), 2015. 3", + "[50] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J. Guibas, and Hao Zhang. AdaCoSeg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 3" + ], + "bbox": [ + 501, + 90, + 890, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "875", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_model.json b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b15b0c659a4ecf04a1d72de944fec1b9d7e859a2 --- /dev/null +++ b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.082, + 0.859, + 0.105 + ], + "angle": 0, + "content": "HAL3D: Hierarchical Active Learning for Fine-Grained 3D Part Labeling" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.134, + 0.723, + 0.153 + ], + "angle": 0, + "content": "Fenggen \\(\\mathrm{Yu}^{1,2*}\\) Yiming Qian\\(^{1}\\) Francisca Gil-Ureta\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.153, + 0.715, + 0.17 + ], + "angle": 0, + "content": "Brian Jackson1 Eric Bennett1 Hao Zhang1,2" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.171, + 0.649, + 0.189 + ], + "angle": 0, + "content": "\\(^{1}\\)Amazon \\(^{2}\\) Simon Fraser University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.223, + 0.314, + 0.239 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.256, + 0.474, + 0.499 + ], + "angle": 0, + "content": "We present the first active learning tool for fine-grained 3D part labeling, a problem which challenges even the most advanced deep learning (DL) methods due to the significant structural variations among the intricate parts. For the same reason, the necessary effort to annotate training data is tremendous, motivating approaches to minimize human involvement. Our labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. To effectively reduce human efforts, we develop two novel features in our tool, hierarchical and symmetry-aware active labeling. Our human-in-the-loop approach, coined HAL3D, achieves close to error-free fine-grained annotations on any test set with pre-defined hierarchical part labels, with \\(80\\%\\) time-saving over manual effort. We will release the finely labeled models to serve the community." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.529, + 0.21, + 0.545 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Semantic shape segmentation and labeling [8, 27] is a classical problem with numerous applications [18], including shape/part recognition, retrieval, indexing, and attribute transfer. We consider the problem of fine-grained 3D part labeling, a task that has not received as much attention as the typical semantic segmentation into coarse/major parts (e.g., backs, seats, and legs for chairs). While these coarse semantic groups support high-level visual perception, they do not possess sufficient granularity to address shape properties related to motion, function, interaction, or construction. Indeed, fine-grained object parts, e.g., the individual wheels and the small mechanical parts in a swivel leg, are both function-revealing and build-aware (see Fig. 1), with strong ties to how the objects were physically assembled or virtually built by artists. Real-world applications such as product inspection, assembly, customization, and robot interaction all operate on fine-grained object parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Learning to label fine-grained parts is a highly challenging problem due to the great structural variations among the small and intricate parts, even for shapes that belong to the same category, as shown in Fig. 1. On a challenging" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.222, + 0.891, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.381, + 0.893, + 0.465 + ], + "angle": 0, + "content": "Figure 1. Fine-grained parts of the 3D models shown exhibit the kind of geometric and structural complexity and diversity that part labeling has to handle. No existing methods, whether learned or heuristic-based, could obtain close-to-fully-accurate labeling in these challenging cases, while our human-in-the-loop active learning tool can approach full accuracy, barring human errors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.674 + ], + "angle": 0, + "content": "dataset with average part counts per category ranging from 15 to more than 500, state-of-the-art learned models such as [33] can only achieve mean IoU scores below \\(50\\%\\), out of a maximum of \\(100\\%\\) for ground truth (GT) results. Also challenging is the tremendous data annotation effort needed to train methods to segment and label fine-grained parts. On the other hand, for applications which demand de facto full accuracy, e.g., for diagnoses in medicine and quality assurance in e-commerce, one can hardly rely on a fully automatic method to reach the mark. It is arguable that the only means to attain full accuracy is to have humans validate all the results, as in the case of data annotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.895, + 0.873 + ], + "angle": 0, + "content": "In this paper, we present the first active learning tool for fine-grained 3D part labeling. Given a set of 3D shapes segmented into fine parts, our labeling tool assigns one of the predefined labels to each part. These input parts are deemed atomic (i.e., indivisible); they can be as low-level as mesh triangles and as high-level as results obtained by an unlabelled semantic segmentation, or mid-level components from an over-segmentation. In general, active learning iterates between automated labeling and human input for label rectification [44]. Compared to conventional learned models with full autonomy, such a human-in-the-loop approach provides a viable option to achieve close to error-free part labeling on test sets, barring human errors1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.351, + 0.9 + ], + "angle": 0, + "content": "*Work carried out during internship at Amazon." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1During testing, accuracy is measured against GT labeling which has" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.896, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Our human-in-the-loop, hierarchical active learning (HAL3D) tool for fine-grained 3D part labeling. The input consists of a set of test shapes each pre-segmented into parts. The labeling proceeds hierarchically, following a tree structure (left) that organizes the hierarchical part labels, from coarse labels (top) to fine-grained labels (bottom). For a node in the label tree (with label \\( l = \\) \"Regular Leg Base\" in the illustration), the input \\( I_0 \\) is the subset of parts, from the entire set of shapes, that are labeled \\( l \\) by its parent. When labeling parts with \\( l \\), the label proposal module first assigns refined labels for each part. Then, proposals are sorted by the mean label probability over the parts of each shape, with the high-confidence (HC) proposals passed to the verification step, which stops once the rate of failed shapes reaches a threshold. The low-confidence (LC) proposals are passed to the label modification module. Correctly labeled shapes fine-tune the network, while skipped and failed shapes, in set \\( I_1 \\), go to the next iteration for labeling. The iterations terminate when all shapes have passed human validation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.473, + 0.632 + ], + "angle": 0, + "content": "As shown in Fig. 2, our interactive labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. Specifically, our system consists of three modules for label proposal, verification, and modification. Label proposals are produced by dynamic graph CNN (DGCNN) [37], with the resulting label probabilities dictating whether to pass the predicted labels to the verification or modification module. Both human-verified and corrected labels are passed back to the label proposal network to fine-tune it. The iteration terminates once all part labels have passed human verification." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.633, + 0.47, + 0.679 + ], + "angle": 0, + "content": "As the key criterion for success with active learning is the minimization of human effort, we design and incorporate two novel features into our labeling tool:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.694, + 0.47, + 0.816 + ], + "angle": 0, + "content": "- Hierarchical labeling: We organize all the part labels in a hierarchical tree structure, which guides our prediction-verification-modification process so that labels further down in the hierarchy are dealt with only after their parent labels have been fully verified. This leads to labeling efficiency, since our labeling process is coarse-to-fine, with fewer labels to process in the coarser levels of the hierarchy. In addition, with both" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.828, + 0.471, + 0.902 + ], + "angle": 0, + "content": "been produced by humans. To achieve full accuracy, two criteria must be met: (a) the testing human labeler's judgement of what is the correct labeling agrees with the GT; (b) the labeler does not make any error, e.g., due to carelessness or time pressure. We regard both types of violations as \"human errors.\" One may also refer to the former as an inconsistency between human labeling, which is likely to occur near fuzzy part boundaries." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.451, + 0.892, + 0.542 + ], + "angle": 0, + "content": "human users and the prediction network only needing to deal with labels at the same level of the hierarchy rather than across all levels, there is less load on the users to reduce possible errors and the prediction accuracy improves, which reduces the amount of label modifications, which are the most expensive task." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.554, + 0.894, + 0.598 + ], + "angle": 0, + "content": "- Part symmetry: Since symmetric parts receive the same label, we employ detected symmetries to constrain and facilitate both label verification and modification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Our hierarchical active learning tool is coined HAL3D. We adopt the hierarchical labels from PartNet [20] and train and test our labeling tool on both PartNet, which comes with semantic part segmentations, and Amazon-Berkeley Objects (ABO). ABO is a recently published dataset of high-quality, artist-created 3D models of real products sold online. These models are composed of build-aware connected components, However, since some of these components are too coarse, we first obtain an over-segmentation via convex decomposition [39] and then apply HAL3D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We evaluate the core components of HAL3D, i.e., the label proposal network being fine-tuned by human feedback, and both hierarchical and symmetry-aware active labeling, by comparing to baselines and performing an ablation. The results demonstrate clear efficiency of HAL3D over labeling tools without active learning and non-hierarchical designs. Overall, HAL3D achieves \\(80\\%\\) speed-up over manual fine-grained labeling, on the PartNet test set. A supplementary video shows our interactive tool at work. We will" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "866" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "release the finely labeled ABO models to serve the community pending permission from the dataset owner." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.136, + 0.22, + 0.151 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.161, + 0.47, + 0.236 + ], + "angle": 0, + "content": "Most approaches to semantic 3D shape segmentation and labeling have been designed to reason about coarse- or high-level structures and to target fully autonomy. Considerably less work has been devoted to fine-grained segmentation and labeling or human-in-the-loop approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.469, + 0.417 + ], + "angle": 0, + "content": "Coarse 3D shape segmentation and labeling: Since the seminal work on learning 3D mesh segmentation and labeling [13] in 2010, many learning methods have been proposed [8]. Prominent supervised methods such as PointNet [24], PointNet++ [25], and DGCNN [37] perform feature learning over point clouds, while MeshCNN [7] develops convolution and pooling layers that operate on mesh edges. These methods, among many more, produce both parts and part labels, but they have only been trained and tested on datasets with coarse segmentations, such as the ShapeNet part [44] and the COSEG datasets [35]; most object categories therein were labeled with only 2-5 parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.599 + ], + "angle": 0, + "content": "High-level parts, e.g., the legs, seats, and backs of chairs, typically possess sufficient structural consistency (e.g., in terms of relative positioning), hence predictability, to facilitate supervised learning. Most unsupervised or weakly supervised approaches also leverage such consistencies. In particular, earlier methods for co-segmentation, i.e., segmenting a set of related shapes altogether, all rely on alignment or clustering, either in the spatial domain [6] or feature spaces [29, 9]. More recently, resorting to DL, AdaCoseg [50] optimizes a group consistency loss defined by matrix ranks. BAE-Net [3] and RIM-Net [21] obtain co-segmentations by learning branched neural fields." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Fine-grained segmentation: Fine-level parts are important in many real-world applications since, whether large or small, each of them has a function to serve and often a motion to carry out. However, the structure consistency possessed by major semantic groups is mostly lost among fine parts. Such parts could be obtained via shape decomposition based on geometric properties such as convexity [12, 39], cylindricity [49], or through primitive fitting using cuboids [42, 30] or quadrics [23, 45], deformed spheres [22, 15]. However, the resulting segmentations are neither semantic nor labeled." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Most learning-based solutions are unsupervised or self-supervised, mainly because annotating fine-grained parts is an extremely tedious task. Li et al. [43] learns part labeling from noisy online part graphs annotated by human artists. Yu et al. [46] trains a recursive part decomposition network for category-specific, fine-grained, hierarchical shape segmentation. Wang et al. [33] and SHRED [11] develop deep clustering to learn part priors from fine-grained segmentation without part labels, where no semantic labels are pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "duced on test shapes. Sharma et al. [28] perform projective shape analysis [36], where a fine-grained 3D segmentation is obtained by aggregating segmented multi-view 2D images. Both of these methods operate fully automatically after training, but their performance falls far short of \\(100\\%\\) accuracy: \\(32.6\\%\\) mean IoU score achieved by [28] on PartNet [20] with fine-grained (Level-3) parts (part counts ranging from 4 to 51), and below \\(50\\%\\) mean IoU by [33] on a more challenging dataset with part counts up to \\(500+\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.893, + 0.395 + ], + "angle": 0, + "content": "Instead of operating on low-level shape elements such as points [33] and pixels [28], the neurally-guided shape parser [10] (NGSP) is trained to assign fine-grained semantic labels to regions of a 3D shape with improved labeling accuracy: mean IoU score up to \\(58\\%\\) on PartNet. NGSP parses region label proposals generated by a guide neural network using likelihood modules to evaluate the global coherence of each proposal. It is still designed as a fully automatic method, but the authors did suggest the need for human-in-the-loop approaches to \"scale beyond carefully-curated research datasets to 'in-the-wild' scenarios.\"" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.893, + 0.579 + ], + "angle": 0, + "content": "Hierarchical analysis: A fine-grained shape understanding is also attainable via hierarchical structural analysis. Some early works perform rule-based reasoning to construct symmetry hierarchies [38] and shape grammars [17], while others work with geometric priors, e.g., convexity [2], for hierarchical shape approximation. Similar to fine-grained segmentation, most learned hierarchical models are also unsupervised or weakly supervised, including co-hierarchical analysis [32], hierarchical abstraction [30], recursive implicit fields [21], and inverse constructed solid geometry (CSG) [45, 14]. All of these methods aim for full automation and work well only with shallow hierarchies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Active learning for part labeling: The classical and core question for active learning [1, 47, 26] is: \"How does one select instances from the underlying data to label, so as to achieve the most effective training for a given level of effort?\" Past works have studied uncertainty and diversity for interactive segmentation and labeling of 3D shapes [35, 5], medical images [31], and large-scale point clouds [40], to name a few. We utilize active learning, aiming to approach full labeling accuracy, regardless whether the test shapes belong to a \"research dataset\" or are \"in-the-wild.\"" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Most closely related to our work is the active framework for region annotations by Yi et al. [44]. Both works share the common goal of producing accurate, human-verified labels. On the other hand, their work imposes a fixed human effort budget, while ours does not. Technically, their work relies on a conditional random field to automatically propagate human labels, while our work integrates a DL module [37] into an active part labeling tool. The only prior work that combines deep and active learning for 3D segmentation [5] focuses on boundary refinement, just as the earlier work [35] whose learning module is constrained \\( k \\)-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "867" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "means clustering. However, none of these tools were designed to work with fine-grained labels, e.g., they still only operate on four part labels for chairs [5, 35, 44, 41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Finally, while our work is the first active learning tool for fine-grained 3D part labeling, hierarchical active learning has been employed in other data domains such as documents. In hierarchical text classification[16, 4], text labels can also be organized hierarchically. In comparison, semantic labeling of 3D parts in a shape collection poses various challenges including richer structural variations, part relations, and more costly and delicate user interactions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.27, + 0.168, + 0.286 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.468, + 0.416 + ], + "angle": 0, + "content": "Given a shape test set \\( S \\), where each shape is presegmented into parts, our methods assign a label \\( l \\in L \\) to each part of each shape. The set of labels, \\( L \\), is predefined and organized in a tree structure such that coarse labels are at the top, and fine-grained labels at the bottom. The labeling progresses hierarchically top-to-bottom, assigning first coarse labels and refining them at each step, achieving high-accuracy fine-grained labeling results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.468, + 0.477 + ], + "angle": 0, + "content": "At each node, we use an active learning based tool to iteratively engage human interactions into the part labeling process. With this tool, we obtain a close to error-free part labeling on each node on any test set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.342, + 0.502 + ], + "angle": 0, + "content": "3.1. Hierarchical labeling strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.468, + 0.644 + ], + "angle": 0, + "content": "We organize the part labels into a tree structure based on their semantics, as in PartNet [20]. Following the same terminology, two types of internal tree nodes are defined: an \"AND\" node indicates \"what part(s) it has\" (e.g., a regular base has legs and a runner), while an \"OR\" node indicates \"what type it is\" (e.g., a star- or pedestal-type base). The labeling on \"AND\" nodes is based on part features, while the labeling on \"OR\" nodes depends on group features; see Fig. 3. Each leaf node presents a final part label." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.468, + 0.81 + ], + "angle": 0, + "content": "At each node of the tree, we combine a deep neural network with an active learning tool to perform part labeling. Consider a single shape \\( s_i \\in S \\) and its set of parts \\( P_i \\). At the root node, the input are all the parts of shape \\( s_i \\), that is \\( P_{root} = P_i \\). After finishing the labeling task at the root node, the set of parts is split into one or more subsets \\( P_l \\subseteq P_{root} \\) corresponding to the children labels \\( l \\). Then, each subset is input to the respective node for further labeling. By splitting the input set at each node, the size of the input decreases as the depth of the node increases. Finally, the labeling stops when we reach a leaf node." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The complexity of our labeling scheme depends on the structure of the tree. On one extreme, if all labels are direct children of the root node, training at the root would be more complex but there is only one node to train. The other extreme is a binary tree, where each node is trained to distinguish between only two labels, but we have to train more" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "nodes, resulting in an increased training time. As we have found the original PartNet trees to have too many internal nodes, we modified this initial tree structure using a heuristic rule: we remove internal \"AND\" nodes which contain less than three children and keep all \"OR\" nodes. This way, our tree structure is better balanced for our task to lead to reduced training time. Examples of our modified tree structures are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.223, + 0.747, + 0.239 + ], + "angle": 0, + "content": "3.2. Symmetry-assisted labeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.383 + ], + "angle": 0, + "content": "Another novel feature to accelerate HAL3D makes use of detectable part symmetries on the test set. Symmetry is ubiquitous in human-made objects and it is common practice for 3D artists to copy-and-paste symmetric or repeated parts when building 3D assets. Symmetry detection has been a well-studied subject in geometry processing [19] and there are a variety of applicable techniques. For simplicity, we simply use sizes of the part oriented bounding boxes as a crude and conservative feature to detect part symmetries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.504 + ], + "angle": 0, + "content": "As symmetric parts are expected to share the same label, we accelerate our active learning in two ways: if a proposal assigns different labels to symmetric parts, we automatically remove it from the \"HC (High confidence)\" set. Then, in the modification module, the user only needs to label one part of the symmetry group and our system automatically propagate it to the other parts. These features improve labeling efficiency by \\(20\\%\\); see Section 5.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.838, + 0.531 + ], + "angle": 0, + "content": "3.3. Active labeling of fine-grained 3D parts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Fine-grained 3D shape segmentation and annotation is challenging because of the wide structural diversity found in small and intricate parts of each shape. This diversity makes it difficult for existing DL methods to achieve high prediction accuracy. On the other side, active learning techniques are commonly used by researchers to improve the performance of the learned classifier with the help of human efforts. We adopt these techniques and, at each node of our hierarchical labeling strategy, we use a deep active learning based method to achieve high part labeling accuracy. Our method comprises three main modules: the DL-based part label proposal network, the part label verification module, and the part label modification module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Consider a node \\( l \\) in our tree structure, given the subset of parts labeled \\( l \\) of each shape, we iteratively update the part labels until they all pass human verification. At each iteration, we first use a DL-based neural network to make part label proposals based on part label probabilities. Then, the HC proposals are passed to the part label verification module for human verification. Next, we collect the LC, i.e., low-confidence, proposals, which are more likely to be inaccurate, and pass them to the label modification module. Finally, the verified proposals and modified proposals are combined and feed back to the proposal network to fine-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "868" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.468, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.287, + 0.468, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.471, + 0.519 + ], + "angle": 0, + "content": "Figure 3. Label proposal architectures. Part label prediction is based on global and local, part-level features at the \"AND\" nodes and global features only for the \"OR\" nodes. The \"AND\" nodes would serve to predict the label for each part and the \"OR\" nodes would predict labels for the entire input shape. Notion-wise, \\( N \\) is the number of points, \\( P \\) is the number of parts, \\( K \\) is the feature dimension, and \\( L \\) is the number of labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.59 + ], + "angle": 0, + "content": "tune the network weights. The active learning iteration will stop if the number of left shapes are less than 40. We explain next how we make HC vs. LC proposals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Label proposal network. Given a set of parts from each shape \\( s_i \\in S \\), we sample \\( N = 8,192 \\) points over the part surfaces and input them into our proposal network. As shown in Fig. 3, our network uses the edge convolution (EdgeConv) module from DGCNN [37] to extract point features from the point clouds. This module consists of five EdgeConv layers whose feature dimensions are 64, 64, 128, 256, and 256, respectively. The point features from all layers are concatenated and passed into a convolution layer. The size of the resulting feature is 512." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Since our tree structure has different types of nodes, we apply two different network structures from DGCNN, where the architecture for \"OR\" nodes is the same as in the original DGCNN classification network. At an \"AND\" node, differently from DGCNN which is the SOTA method to predict point labels based on point features, we use extra max-pooling operations to aggregate the local part features and global shape features, as inspired by NGSP [10]. The concatenated features are fed into the part classifier, which contains 3 MLP layers to predict part labels. As for an" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "\"OR\" node, labels of the input shape are predicted based on the global features only; see Fig. 3. Since different shapes have different part counts, we set the maximum part count \\( P_{max} = 150 \\) during training for batch processing and add zero-padding to the part feature matrix so that all the part feature matrices have the same dimension. More details of our network can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.893, + 0.495 + ], + "angle": 0, + "content": "Label verification. Our goal is to get the labels of each shape verified by a human. To reduce verification and training times, we need to decide which shapes are verified first and how many to verify on each iteration. Shapes in the test set that have similar structure to shapes in the train set are more likely to be labeled well. We can quickly select these well-labeled shapes from the test set and use them to enhance our limited train set, improving the generalization ability of our proposal network. In our method, we first sort the test set based on average predicted part label probabilities from the proposal network of each shape and arrange the test set into batches, each containing \\( B = 10 \\) shapes. Users can mark a shape as well-labeled if they see all parts are labeled correctly, or bad-labeled when one part is wrong. We design a user-friendly interface to achieve this, where users can view and rotate a batch of 3D shapes at the same time. The interface also shows the corresponding label colors for reference. Please see supplementary material for more details about our part label verification interface." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.895, + 0.678 + ], + "angle": 0, + "content": "Given the sorted shapes, we need to decide how many of them to show the user as HC proposals before fine-tuning the prediction, to achieve a balance between processing efficiency and utilization of user feedback. To this end, we adopt an adaptive strategy: given the sorted proposals, we stop the verification step when the number of human-verified shapes on a batch is below a threshold, which is set to 4 in our experiments. This simple strategy is effective since early batches tend to have higher labeling accuracy than later batches. This way, the number of high confidence proposals presented to the user would adapt to the classification purity of the current test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.87 + ], + "angle": 0, + "content": "Label modification. Shapes in the test set whose part structures and geometries differ significantly from those in the train set are less likely to be labeled accurately by the network. These shapes can be regarded as outliers in the label regression task, but they can be especially valuable during fine-tuning. Human intervention is inevitable in this situation to correct the labels. Specifically, the shapes provided to the label modification module are collected from two sets: the \\( Q_{1} = 20 \\) LC proposals with the lowest mean label probabilities, and those shapes that failed user verification more than \\( H = 2 \\) times. We ask users to modify the incorrect labels using our modification interface." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our part annotation interface, labels are organized hierarchically so users can quickly look up the correct labels." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "869" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.207, + 0.473, + 0.348 + ], + "angle": 0, + "content": "Figure 4. Comparison of pre-segmented parts in PartNet and ABO. ABO shapes were pre-segmented into connected components where each component may correspond to multiple labels, e.g., the back and the arm of the chair are grouped into the same component, as indicated by the red color in the second column. Instead, our part labeling operates on the convex pieces obtained by a decomposition. In the first three figures, the colors are used to distinguish between different parts, and we can see that the parts in PartNet are manually segmented and exhibiting imperfect boundaries. In the last figure, the colors reflect semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.418 + ], + "angle": 0, + "content": "The label-to-color mapping table is also provided to help find the incorrect labels. More details about this interface are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Shapes that have been verified and modified by humans are added into the initial training set together to fine-tune our proposal network. On the other hand, we skip shapes excluded from the HC and LC proposals. These shapes are combined with shapes that failed to pass the verification less than \\( H \\) times to go to the next iteration for labeling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.279, + 0.543 + ], + "angle": 0, + "content": "4. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Datasets. The Stanford PartNet [20] dataset is a common choice for fine-grained shape analysis [34, 10, 28]. We follow NGSP [10] and use the four major categories from PartNet for evaluation: chair, table, lamp, and storage furniture. Fine-grained labels in PartNet are organized into three granularity levels, and we adopt the second level in our experiments as in NGSP [10]. We first filter out noisy shapes with erroneous labels. Then from each category, we take 500 shapes and split them into 50/50/400 as train/val/test sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We also evaluate HAL3D on the recent Amazon Berkeley Objects (ABO) dataset, which is composed of about 8,000 high-quality 3D shapes created by artists from real product images, where each 3D model is a combination of build-aware connected components. Unlike the 3D shapes from PartNet in which each part corresponds to one single semantic label, a build-aware component from ABO may be appropriately associated with multiple labels, as shown in Fig. 4. To alleviate this issue, we apply an approximate convex decomposition [39] to further decompose each connected component into a fine-grained, i.e., over-segmented, set of convexes, where each convex piece is deemed to receive one and only one part label. Our active labeling is then applied to the set of convexes. When running the con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "vex decomposition algorithm, we use the default parameters from the provided code, except for the concavity threshold setting, which is set as 0.2 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.213 + ], + "angle": 0, + "content": "As GT labels are unavailable in ABO, three artists were hired to create them for the decomposed convex components. We built a test set containing 1,800 shapes from five categories: chair, table, lamp, storage furniture, and bed, with 200 for storage and 400 each for the other categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.221, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Metrics. Active learning systems are typically evaluated from two perspectives: labeling accuracy and the amount of human efforts or involvements. We employ two standard metrics to assess the labeling quality: (1) part label accuracy measuring how many parts are correctly labeled and (2) point cloud segmentation mIoU depicting how much of the shape surface area is correctly labeled." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.449 + ], + "angle": 0, + "content": "The amount of human efforts is measured by the consumed time in the steps of label verification and modification. Inspired from [44], we firstly give artists the same training courses and tools, and conduct user studies on Part-Net. We report the total recorded labeling time in Tab. 1. We also collect the operations they used in the user study and model the total human interaction time as a function of the number of processed parts and shapes:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.459, + 0.891, + 0.49 + ], + "angle": 0, + "content": "\\[\nt = T _ {v} ^ {p} \\left(P _ {v}, L\\right) + T _ {v} ^ {f} \\left(S _ {v}, L\\right) + T _ {v} ^ {p} \\left(P _ {m} ^ {c}, L\\right) + T _ {m} ^ {p} \\left(P _ {m} ^ {w}, L\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.566 + ], + "angle": 0, + "content": "where \\(P_{v}\\) is the number of correctly labeled parts in verification, \\(S_{v}\\) is the number of incorrectly shapes in verification, \\(P_{m}^{c}\\) is the number of correctly labeled parts in modification, \\(P_{m}^{w}\\) is the number of incorrectly labeled parts in modification, and finally \\(L\\) is the total number of labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Function (1) is devised based on several observations: a) users need to go through and verify all the part labels to ensure accuracy; b) they can easily pick up failure cases during verification if any part is incorrectly labeled, without checking other parts; and c) users need to check all the parts in each shape to know whether modification is needed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.703 + ], + "angle": 0, + "content": "The formulas and their coefficients in (2) are obtained via linear regression using the total operation and labeling times collected from all annotators in our user study:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.713, + 0.823, + 0.731 + ], + "angle": 0, + "content": "\\[\nT _ {v} ^ {p} \\left(P _ {v}, L\\right) = \\left(0. 3 1 + 0. 0 3 * L\\right) * P _ {v},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.734, + 0.892, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} T _ {v} ^ {f} \\left(S _ {v}, L\\right) = (1. 4 7 + 0. 0 5 * L) * S _ {v}, \\\\ \\text {T r} \\left(R _ {v} ^ {f}, L\\right) = \\left(0. 0 4 - 0. 0 2, L\\right) * R _ {v} ^ {f} \\end{array} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.754, + 0.826, + 0.771 + ], + "angle": 0, + "content": "\\[\nT _ {v} ^ {p} \\left(P _ {m} ^ {c}, L\\right) = \\left(0. 3 1 + 0. 0 3 * L\\right) * P _ {m} ^ {c},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.773, + 0.826, + 0.79 + ], + "angle": 0, + "content": "\\[\nT _ {m} ^ {p} \\left(P _ {m} ^ {w}, L\\right) = (5. 2 8 + 0. 2 3 * L) * P _ {m} ^ {w},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( T_{v}^{p} \\) is the checking time for correctly labeled parts, \\( T_{v}^{f} \\) is the verification time for incorrectly labeled shapes, and \\( T_{m}^{p} \\) is the time for part label modification. We use these functions to evaluate the human interaction timing, in seconds, for hyper-parameters analysis on PartNet. We also report additional machine time cost from training and fine-tuning in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "870" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.471, + 0.268 + ], + "angle": 0, + "content": "We have implemented our approach in PyTorch. The label proposal network at each node is pre-trained with the Adam optimizer on the training set of PartNet for 250 epochs, where the learning rate is 0.001 and it is decreased by 0.8 every 25 epochs. When fine-tuning the network in HAL3D, the learning rate is fixed to 0.0001, the maximum epoch number is 125, and the validation set is used for early stopping. Our experiments were ran on an NVIDIA Tesla V100 GPU, and the training time varies from 20 minutes to 3 hours according to the number of parts at each node." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.279, + 0.269, + 0.296 + ], + "angle": 0, + "content": "5.1. Competing methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.304, + 0.47, + 0.334 + ], + "angle": 0, + "content": "We compare HAL3D, our hierarchical active labeling tool, with two automatic part labeling methods," + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.338, + 0.469, + 0.384 + ], + "angle": 0, + "content": "- PartNet [20], which employs the PointNet++ [25] backbone for point-wise feature extraction and labels parts by a voting scheme over points belonging to the same part." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.469, + 0.448 + ], + "angle": 0, + "content": "- NGSP [10], which aggregates point-wise features via PointNet++ [25] to obtain part features and then leverages a likelihood-aware beam search for optimization; this represents the state of the art for fine-grained labeling." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.338, + 0.469, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.452, + 0.382, + 0.467 + ], + "angle": 0, + "content": "And two variants of our labeling approach," + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.471, + 0.469, + 0.516 + ], + "angle": 0, + "content": "- Our prop. is one such variant without using the manual label verification or modification module; it predicts part labels using the label proposal network only." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.519, + 0.468, + 0.551 + ], + "angle": 0, + "content": "- DAL3D is a deep active learning variant of our approach without using the hierarchical labeling strategy." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.471, + 0.469, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.287, + 0.576 + ], + "angle": 0, + "content": "5.2. Evaluation on PartNet" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.585, + 0.469, + 0.615 + ], + "angle": 0, + "content": "All methods presented in this section are pre-trained and tested on the same training and test sets of PartNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Quantitative comparison. Tab. 1 compares the three automatic methods, where PartNet is inferior to both NGSP and our label proposal network due to its hand-crafted voting mechanism. Our simple network achieves comparable accuracy compared to NGSP (difference by 0.11 and 3.49 with respect to mIoU and part accuracy, respectively). However, NGSP suffers from significantly longer running time due to the intensive beam search; see the supplementary for a running time comparison. In contrast, our proposal network is effective and efficient, i.e., capable of quickly producing accurate initial labels to reduce the manual workload in verification and modification for active learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Tab. 1 also shows that even with the sophisticated design in NGSP, the highest accuracy attained is only \\(70\\%\\). Hence the performance of automatic fine-grained labeling still falls far below that of active learning, where the largest and smallest accuracy gap is \\(40.77\\%\\) and \\(22.49\\%\\) between NGSP and HAL3D, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.087, + 0.892, + 0.185 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison against the competing labeling methods and variants. In the table, \"AL\" indicates whether a method uses active learning. Since PartNet, NGSP, and Our prop. are automatic methods, their entries are marked with a '-' instead. \"Accu.\" denotes the part label accuracy, \"mIoU\" denotes the mean Intersection-over-Union, and \"Lab. time\" denotes the labeling time (hours in total) for the active learning methods." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.196, + 0.892, + 0.584 + ], + "angle": 0, + "content": "
MethodALAccu.↑mIoU↑Lab. time↓(h)
Chair
PartNet [20]-54.9031.04-
NGSP [10]-68.9842.90-
Our prop.-67.4541.34-
DAL3D✓89.8472.305.99
HAL3D✓94.1384.924.34
Table
PartNet [20]-33.2114.92-
NGSP [10]-47.1127.94-
Our prop.-42.7723.50-
DAL3D✓78.5369.829.14
HAL3D✓84.1874.654.41
Storage furniture
PartNet [20]-55.8230.86-
NGSP [10]-70.6045.79-
Our prop.-69.8242.43-
DAL3D✓84.5771.994.81
HAL3D✓93.0988.723.87
Lamp
PartNet [20]-38.2811.02-
NGSP [10]-49.7623.49-
Our prop.-44.7719.27-
DAL3D✓83.1469.354.52
HAL3D✓90.5379.512.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.609, + 0.88, + 0.624 + ], + "angle": 0, + "content": "Table 2. Ablation on key modules of HAL3D for chair category." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.634, + 0.908, + 0.744 + ], + "angle": 0, + "content": "
Row IDProp.Hier.Sym.ALLab-T↓Accu↑
2nd----22.0589.16
3rd✓---8.6588.53
4th✓✓✓-6.3793.87
5th✓-✓✓5.9989.84
6th✓✓-✓5.2193.45
7th✓✓✓✓4.3494.13
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.829 + ], + "angle": 0, + "content": "HAL3D consistently outperforms our DAL3D variant in all metrics, particularly 2.42 hours faster on average on labeling time, demonstrating the importance of our hierarchical labeling order for reducing manual workload. We also provide qualitative comparison in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation study. Tab. 2 verifies the contributions of the main components of HAL3D on minimizing human labeling efforts, a key criterion for success for a human-in-the-loop approach. The second to fifth columns indicate:" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.096, + 0.462, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.472, + 0.399 + ], + "angle": 0, + "content": "Figure 5. Visual comparison between labeling methods. HAL3D can achieve results approaching GT, barring human errors. The only mis-labeled part by HAL3D here is the bottom panel of the table (red oval), which could be easily labeled as shelf by users." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.415, + 0.468, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.517, + 0.471, + 0.601 + ], + "angle": 0, + "content": "Figure 6. A plot of the number of human-verified shapes (max = 400) at different nodes (only those which existed in more than 100 shapes) of the chair category. The plot shows that with hierarchical active learning, many more shapes pass the verification, hence saving on the costly label modification, especially at higher hierarchy levels (i.e., earlier in the labeling iterations)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.628, + 0.378, + 0.643 + ], + "angle": 0, + "content": "- Prop. if the label proposal network is used;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.647, + 0.414, + 0.662 + ], + "angle": 0, + "content": "- Hier. if the hierarchical labeling strategy is used;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.665, + 0.353, + 0.681 + ], + "angle": 0, + "content": "- Sym. if the symmetry checking is used." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.685, + 0.312, + 0.7 + ], + "angle": 0, + "content": "- AL. if the active learning is used." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.628, + 0.414, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.853 + ], + "angle": 0, + "content": "We compare our full approach against five baselines. The second row is a baseline that directly predicts labels by annotating with our label modification interface. The third row is a baseline that firstly uses the label proposal network and then manually modifies incorrect part labels. The fourth, fifth, and sixth rows represent variants of HAL3D without hierarchical labeling, symmetry checking, and active learning, respectively. The results in Tab. 2 clearly justify the design choices in our method, where our full approach (last row) is the most efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To further evaluate the effectiveness of our hierarchical labeling, we plot in Fig. 6 the number of shapes which passed human verification over the active learning qua" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.115 + ], + "angle": 0, + "content": "Table 3. Inference time comparison between NGSP and our label proposal network. We report average seconds per shape." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.115, + 0.86, + 0.164 + ], + "angle": 0, + "content": "
CategoryChairTableLampStorage
NGSP5.41018.0112.6912.87
Our prop.0.0120.0140.0120.011
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.319 + ], + "angle": 0, + "content": "tions. As we can see, HAL3D consistently verified more shapes as correct at all tree nodes when comparing to the baseline that had no hierarchical design, e.g., more than 300 shapes were marked as correct in the first iteration in the chair_base node. Essentially, our hierarchical verification splits the labeling task into smaller and easier tasks compared to the case of non-hierarchical design, which effectively reduces the burden arising from the subsequent and more costly modification step in terms of human effort." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.395 + ], + "angle": 0, + "content": "In the supplementary material, we perform ablations on the four hyper parameters and thresholds employed in our method. In all the experiments, we fix these parameters and set them based only on a small dataset. A more careful parameter tuning should lead to better results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Inference time comparison. Although NGSP achieved higher prediction accuracy than our proposal network, as shown in Tab. 1, our network is more efficient and better suited to be incorporated into an active learning framework. Tab. 3 compares the inference times between NGSP and our proposal network. NGSP is highly time-consuming due to the use of beam search, whereas our label proposal network takes only 12ms per shape on average." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.688, + 0.56 + ], + "angle": 0, + "content": "5.3. Evaluation on ABO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We apply HAL3D to label our constructed test set for ABO, where the label proposal network is pre-trained on the training set of PartNet. As mentioned in Sec. 4, shapes in ABO are decomposed into convex shapes and HAL3D assigns a semantic label to each decomposed piece. Three artists were hired to use our HAL3D tool and interface for evaluation. Since GT labels are not provided in ABO, we inspect the labeling results from the artists via human cross validation with the final corrected labeling forming the GT. As different artists may report different accuracy and recorded labeling times, we report averages in the paper. The artists worked together with an expert to determine labels for uncertain shapes. The visual results from HAL3D, as shown in Fig. 7 and the supplemental material, were obtained on the test set by taking majority voting from labels provided by different artists." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown by quantitative results in Tab. 4, HAL3D can accurately label all categories with a \\(94\\%\\) average accuracy using 3-7 hours of manual annotation, demonstrating the effectiveness and efficiency of our method. The reduction on human efforts against manual annotation becomes more significant when comparing against the last four rows in Tab. 4," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.094, + 0.191, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.096, + 0.326, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.097, + 0.462, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.096, + 0.6, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.095, + 0.729, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.095, + 0.862, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.456, + 0.746, + 0.471 + ], + "angle": 0, + "content": "Figure 7. Visual comparison between several part labeling methods on the ABO dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.491, + 0.468, + 0.518 + ], + "angle": 0, + "content": "Table 4. Comparison for different methods to pass human verification on test set of ABO. The time is in hours." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.518, + 0.462, + 0.692 + ], + "angle": 0, + "content": "
ChairTableLampCabinetBed
# shapes400400400200400
HAL3D
Time ↓Accu. ↑6.4693.077.4192.554.3896.293.3494.484.5193.71
PartNet + modification
Time ↓Accu. ↑28.2588.9633.9486.5131.5390.9113.3590.2425.0289.37
NGSP + modification
Time ↓Accu. ↑28.3088.7734.3286.4031.5691.3713.2290.5624.9089.93
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.756 + ], + "angle": 0, + "content": "where we use the same manual modification interface to directly correct the PartNet and NGSP results. Please see supplementary material for results on ABO." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.197, + 0.785 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We present HAL3D, the first semantic labeling tool that is designed to operate on fine-grained 3D parts and achieve close to error-free labeling through full verification by humans, barring human misjudgement. Our human-in-the-loop approach is based on active learning, combining deep label prediction and human inputs to iteratively fine-tune and improve the network prediction. Extensive experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.495, + 0.892, + 0.526 + ], + "angle": 0, + "content": "demonstrate that human efforts are effectively reduced via hierarchical and symmetry-aware active labeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.893, + 0.695 + ], + "angle": 0, + "content": "HAL3D is currently implemented with DGCNN [37] as the label prediction backbone, but it can be readily replaced by PointNet [24], which was adopted by PartNet [20] and NGSP [10], or a more performant alternative such as Point Transformer [48]. Aside from symmetry, inter-shape part correspondences can also be utilized to reduce labeling costs, assuming that they can be reliably obtained. In addition, our tool does not yet allow splitting of under-segmented parts. Since performing such operations by humans is time-consuming, incorporating auto-splitting schemes into HAL3D deserves future consideration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.895, + 0.79 + ], + "angle": 0, + "content": "We regard our work as only a preliminary step towards deep active learning for fine-grained structural analysis of 3D shapes. A natural focus for future work is to develop a more sophisticated ranking or selection mechanism for label verification and modification to achieve the most effective training for a given level of human effort." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.812, + 0.661, + 0.83 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We thank all the anonymous reviewers and area chairs for their constructive comments and the Amazon artists for performing the human annotations. Thanks also go to Jeff-Aram Van Gorp for helpful discussions and other support." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "873" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Charu C. Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and Philip S. Yu. Active learning: A survey. In Data Classification: Algorithms and Applications, pages 571-597. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.225 + ], + "angle": 0, + "content": "[2] Marco Attene, Michela Mortara, Michela Spagnuolo, and Bianca Falcidieno. Hierarchical convex approximation of 3D shapes for fast region selection. Comput. Graph. Forum, 27(5):1323-1332, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Zhiqin Chen, Kangxue Yin, Matt Fisher, Siddhartha Chaudhuri, and Hao Zhang. BAE-NET: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[4] Yu Cheng, Kunpeng Zhang, Yusheng Xie, Ankit Agrawal, and Alok Choudhary. On active learning in hierarchical classification. In Proceedings of the 21st ACM International Conference on Information and knowledge Management, pages 2467-2470, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[5] David George, Xianghua Xie, Yukun Lai, and Gary KL Tam. A deep learning driven active framework for segmentation of large 3d shape collections. Computer-Aided Design, 144:103179, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[6] Aleksey Golovinskiy and Thomas Funkhouser. Consistent segmentation of 3D models. In Proc. Shape Modeling International (SMI), 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: A network with an edge. ACM TOG, 38(4), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[8] Yong He, Hongshan Yu, Xiaoyan Liu, Zhengeng Yang, Wei Sun, Yaonan Wang, Qiang Fu, Yanmei Zou, and Ajmal Mian. Deep learning based 3d segmentation: A survey. CoRR, abs/2103.05423, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[9] Ruizhen Hu, Lubin Fan, and Ligang Liu. Co-segmentation of 3D shapes via subspace clustering. Comput. Graph. Forum, 31(5):1703--1713, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[10] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3D shape regions with approximate inference. In CVPR, 2022. 3, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[11] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. ACM Transactions on Graphics (TOG), 41(6):1-11, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[12] Oliver Van Kaick, Noa Fish, Yanir Kleiman, Shmuel Asafi, and Daniel Cohen-OR. Shape segmentation by approximate convexity analysis. ACM TOG, 34(1). 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[13] Evangelos Kalogerakis, Aaron Hertzmann, and Karan Singh. Learning 3D mesh segmentation and labeling. ACM TOG, 29(4), 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[14] Kacper Kania, Maciej Zięba, and Tomasz Kajdanowicz. UCSG-NET - unsupervised discovering of constructive solid geometry tree. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[15] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Neural star domain as primitive representation. Advances in Neural Information Processing Systems, 33:7875-7886, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.847, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Xiao Li, Da Kuang, and Charles X Ling. Active learning for hierarchical text classification. In Pacific-Asia conference on knowledge discovery and data mining, pages 14–25. Springer, 2012. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Tianqiang Liu, Siddhartha Chaudhuri, Vladimir G. Kim, Qixing Huang, Niloy J. Mitra, and Thomas Funkhouser. Creating consistent scene graphs using a probabilistic grammar. ACM TOG, 33, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[18] Niloy Mitra, Michael Wand, Hao Zhang, Daniel Cohen-Or, and Martin Bokeloh. Structure-aware shape processing. In Eurographics State-of-the-Art Report (STAR), 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[19] Niloy J. Mitra, Mark Pauly, Michael Wand, and Duygu Ceylan. Symmetry in 3D geometry: Extraction and applications. 32(6), 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[20] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, June 2019. 2, 3, 4, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[21] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. RIM-Net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[22] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[23] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, pages 10344—10353, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[24] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[25] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[26] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B. Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[27] Ariel Shamir. A survey on mesh segmentation techniques. Comput. Graph. Forum, 27(6):1539-1556. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[28] Gopal Sharma, Kangxue Yin, Subhransu Maji, Evangelos Kalogerakis, Or Litany, and Sanja Fidler. MvDeCor: Multiview dense correspondence learning for fine-grained 3D segmentation. In ECCV, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[29] Oana Sidi, Oliver van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. ACM TOG, 30(6):126:1-126:9, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[30] Chun-Yu Sun, Qian-Fang Zou, Xin Tong, and Yang Liu. Learning adaptive hierarchical cuboid abstractions of 3D shape collections. 38(6), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[31] Andrew Top, Ghassan Hamarneh, and Rafeef Abugharbieh. Active learning for interactive 3D image segmentation. In Int. Conf. on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[32] Oliver van Kaick, Kai Xu, Hao Zhang, Yanzhen Wang, Shuyang Sun, Ariel Shamir, and Daniel Cohen-Or. Cohierarchical analysis of shape structures. ACM TOG, 32(4), 2013. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "874" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[33] Xiaogang Wang, Xun Sun, Xinyu Cao, Kai Xu, and Bin Zhou. Learning fine-grained segmentation of 3D shapes without part labels. In CVPR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[34] Xiaogang Wang, Bin Zhou, Haiyue Fang, Xiaowu Chen, Qinping Zhao, and Kai Xu. Learning to group and label fine-grained shape components. 37(6), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[35] Yunhai Wang, Shmulik Asafi, Oliver van Kaick, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Active co-analysis of a set of shapes. ACM TOG, 31(6), 2012. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[36] Yunhai Wang, Minglun Gong, Tianhua Wang, Daniel Cohen-Or, Hao Zhang, and Baoquan Chen. Projective analysis for 3D shape segmentation. ACM TOG, 32(6), 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[37] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM TOG, 2019. 2, 3, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[38] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. Comput. Graph. Forum, 30(2):287-296, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[39] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search. ACM TOG, 41(4), 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.501 + ], + "angle": 0, + "content": "[40] Tsung-Han Wu, Yueh-Cheng Liu1, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H. Hsu. ReDAL: Region-based and diversity-aware active learning for point cloud semantic segmentation. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[41] Zizhao Wu, Ruyang Shou, Yunhai Wang, and Xinguo Liu. Interactive shape co-segmentation via label propagation. In Proc. of CAD/Graphics, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[42] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM TOG, 40(4), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[43] Li Yi, Leonidas Guibas, Aaron Hertzmann, Vladimir G. Kim, Hao Su, and Ersin Yumer. Learning hierarchical shape segmentation and labeling from online repositories. SIGGRAPH, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[44] Li Yi, Vladimir G. Kim, I.-Chao Shen Duygu Ceylan, Mengyuan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6), 2016. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[45] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact CAD shapes with adaptive primitive assembly. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[46] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. PartNet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[47] Xueying Zhan, Qingzhong Wang, Kuan-hao Huang, Haoyi Xiong, Dejing Dou, and Antoni B. Chan. A comparative survey of deep active learning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[48] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In CVPR, 2021. 9" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Yang Zhou, Kangxue Yin, Hui Huang, Hao Zhang, Minglun Gong, and Daniel Cohen-Or. Generalized cylinder decomposition. ACM TOG, 34(6), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[50] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J. Guibas, and Hao Zhang. AdaCoSeg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "875" + } + ] +] \ No newline at end of file diff --git a/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_origin.pdf b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8d7253c50ce3ef790b2d6a03883730f2a51a7396 --- /dev/null +++ b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/4d9d11c1-a387-4840-b228-bd93e87c1665_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b0c4fbe1fad4c66db7a603a3c70bd83788826c612dafce3e5d13dd6eb2c22db +size 9113112 diff --git a/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/full.md b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..88bb2f58c5b0d69017c8b79a7776785d9f70d30b --- /dev/null +++ b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/full.md @@ -0,0 +1,323 @@ +# HAL3D: Hierarchical Active Learning for Fine-Grained 3D Part Labeling + +Fenggen $\mathrm{Yu}^{1,2*}$ Yiming Qian $^{1}$ Francisca Gil-Ureta $^{1}$ + +Brian Jackson1 Eric Bennett1 Hao Zhang1,2 + +$^{1}$ Amazon $^{2}$ Simon Fraser University + +# Abstract + +We present the first active learning tool for fine-grained 3D part labeling, a problem which challenges even the most advanced deep learning (DL) methods due to the significant structural variations among the intricate parts. For the same reason, the necessary effort to annotate training data is tremendous, motivating approaches to minimize human involvement. Our labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. To effectively reduce human efforts, we develop two novel features in our tool, hierarchical and symmetry-aware active labeling. Our human-in-the-loop approach, coined HAL3D, achieves close to error-free fine-grained annotations on any test set with pre-defined hierarchical part labels, with $80\%$ time-saving over manual effort. We will release the finely labeled models to serve the community. + +# 1. Introduction + +Semantic shape segmentation and labeling [8, 27] is a classical problem with numerous applications [18], including shape/part recognition, retrieval, indexing, and attribute transfer. We consider the problem of fine-grained 3D part labeling, a task that has not received as much attention as the typical semantic segmentation into coarse/major parts (e.g., backs, seats, and legs for chairs). While these coarse semantic groups support high-level visual perception, they do not possess sufficient granularity to address shape properties related to motion, function, interaction, or construction. Indeed, fine-grained object parts, e.g., the individual wheels and the small mechanical parts in a swivel leg, are both function-revealing and build-aware (see Fig. 1), with strong ties to how the objects were physically assembled or virtually built by artists. Real-world applications such as product inspection, assembly, customization, and robot interaction all operate on fine-grained object parts. + +Learning to label fine-grained parts is a highly challenging problem due to the great structural variations among the small and intricate parts, even for shapes that belong to the same category, as shown in Fig. 1. On a challenging + +![](images/49ad14b908fdcd3535a0d9333c611b64806e40137f3da5dbb12c277b7a6036d4.jpg) +Figure 1. Fine-grained parts of the 3D models shown exhibit the kind of geometric and structural complexity and diversity that part labeling has to handle. No existing methods, whether learned or heuristic-based, could obtain close-to-fully-accurate labeling in these challenging cases, while our human-in-the-loop active learning tool can approach full accuracy, barring human errors. + +dataset with average part counts per category ranging from 15 to more than 500, state-of-the-art learned models such as [33] can only achieve mean IoU scores below $50\%$ , out of a maximum of $100\%$ for ground truth (GT) results. Also challenging is the tremendous data annotation effort needed to train methods to segment and label fine-grained parts. On the other hand, for applications which demand de facto full accuracy, e.g., for diagnoses in medicine and quality assurance in e-commerce, one can hardly rely on a fully automatic method to reach the mark. It is arguable that the only means to attain full accuracy is to have humans validate all the results, as in the case of data annotation. + +In this paper, we present the first active learning tool for fine-grained 3D part labeling. Given a set of 3D shapes segmented into fine parts, our labeling tool assigns one of the predefined labels to each part. These input parts are deemed atomic (i.e., indivisible); they can be as low-level as mesh triangles and as high-level as results obtained by an unlabelled semantic segmentation, or mid-level components from an over-segmentation. In general, active learning iterates between automated labeling and human input for label rectification [44]. Compared to conventional learned models with full autonomy, such a human-in-the-loop approach provides a viable option to achieve close to error-free part labeling on test sets, barring human errors1. + +![](images/e13cce0fef38766eeff728302de76c2575fac608ccbf640bb699e4aba298d5c2.jpg) +Figure 2. Our human-in-the-loop, hierarchical active learning (HAL3D) tool for fine-grained 3D part labeling. The input consists of a set of test shapes each pre-segmented into parts. The labeling proceeds hierarchically, following a tree structure (left) that organizes the hierarchical part labels, from coarse labels (top) to fine-grained labels (bottom). For a node in the label tree (with label $l =$ "Regular Leg Base" in the illustration), the input $I_0$ is the subset of parts, from the entire set of shapes, that are labeled $l$ by its parent. When labeling parts with $l$ , the label proposal module first assigns refined labels for each part. Then, proposals are sorted by the mean label probability over the parts of each shape, with the high-confidence (HC) proposals passed to the verification step, which stops once the rate of failed shapes reaches a threshold. The low-confidence (LC) proposals are passed to the label modification module. Correctly labeled shapes fine-tune the network, while skipped and failed shapes, in set $I_1$ , go to the next iteration for labeling. The iterations terminate when all shapes have passed human validation. + +As shown in Fig. 2, our interactive labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. Specifically, our system consists of three modules for label proposal, verification, and modification. Label proposals are produced by dynamic graph CNN (DGCNN) [37], with the resulting label probabilities dictating whether to pass the predicted labels to the verification or modification module. Both human-verified and corrected labels are passed back to the label proposal network to fine-tune it. The iteration terminates once all part labels have passed human verification. + +As the key criterion for success with active learning is the minimization of human effort, we design and incorporate two novel features into our labeling tool: + +- Hierarchical labeling: We organize all the part labels in a hierarchical tree structure, which guides our prediction-verification-modification process so that labels further down in the hierarchy are dealt with only after their parent labels have been fully verified. This leads to labeling efficiency, since our labeling process is coarse-to-fine, with fewer labels to process in the coarser levels of the hierarchy. In addition, with both + +human users and the prediction network only needing to deal with labels at the same level of the hierarchy rather than across all levels, there is less load on the users to reduce possible errors and the prediction accuracy improves, which reduces the amount of label modifications, which are the most expensive task. + +- Part symmetry: Since symmetric parts receive the same label, we employ detected symmetries to constrain and facilitate both label verification and modification. + +Our hierarchical active learning tool is coined HAL3D. We adopt the hierarchical labels from PartNet [20] and train and test our labeling tool on both PartNet, which comes with semantic part segmentations, and Amazon-Berkeley Objects (ABO). ABO is a recently published dataset of high-quality, artist-created 3D models of real products sold online. These models are composed of build-aware connected components, However, since some of these components are too coarse, we first obtain an over-segmentation via convex decomposition [39] and then apply HAL3D. + +We evaluate the core components of HAL3D, i.e., the label proposal network being fine-tuned by human feedback, and both hierarchical and symmetry-aware active labeling, by comparing to baselines and performing an ablation. The results demonstrate clear efficiency of HAL3D over labeling tools without active learning and non-hierarchical designs. Overall, HAL3D achieves $80\%$ speed-up over manual fine-grained labeling, on the PartNet test set. A supplementary video shows our interactive tool at work. We will + +release the finely labeled ABO models to serve the community pending permission from the dataset owner. + +# 2. Related Work + +Most approaches to semantic 3D shape segmentation and labeling have been designed to reason about coarse- or high-level structures and to target fully autonomy. Considerably less work has been devoted to fine-grained segmentation and labeling or human-in-the-loop approaches. + +Coarse 3D shape segmentation and labeling: Since the seminal work on learning 3D mesh segmentation and labeling [13] in 2010, many learning methods have been proposed [8]. Prominent supervised methods such as PointNet [24], PointNet++ [25], and DGCNN [37] perform feature learning over point clouds, while MeshCNN [7] develops convolution and pooling layers that operate on mesh edges. These methods, among many more, produce both parts and part labels, but they have only been trained and tested on datasets with coarse segmentations, such as the ShapeNet part [44] and the COSEG datasets [35]; most object categories therein were labeled with only 2-5 parts. + +High-level parts, e.g., the legs, seats, and backs of chairs, typically possess sufficient structural consistency (e.g., in terms of relative positioning), hence predictability, to facilitate supervised learning. Most unsupervised or weakly supervised approaches also leverage such consistencies. In particular, earlier methods for co-segmentation, i.e., segmenting a set of related shapes altogether, all rely on alignment or clustering, either in the spatial domain [6] or feature spaces [29, 9]. More recently, resorting to DL, AdaCoseg [50] optimizes a group consistency loss defined by matrix ranks. BAE-Net [3] and RIM-Net [21] obtain co-segmentations by learning branched neural fields. + +Fine-grained segmentation: Fine-level parts are important in many real-world applications since, whether large or small, each of them has a function to serve and often a motion to carry out. However, the structure consistency possessed by major semantic groups is mostly lost among fine parts. Such parts could be obtained via shape decomposition based on geometric properties such as convexity [12, 39], cylindricity [49], or through primitive fitting using cuboids [42, 30] or quadrics [23, 45], deformed spheres [22, 15]. However, the resulting segmentations are neither semantic nor labeled. + +Most learning-based solutions are unsupervised or self-supervised, mainly because annotating fine-grained parts is an extremely tedious task. Li et al. [43] learns part labeling from noisy online part graphs annotated by human artists. Yu et al. [46] trains a recursive part decomposition network for category-specific, fine-grained, hierarchical shape segmentation. Wang et al. [33] and SHRED [11] develop deep clustering to learn part priors from fine-grained segmentation without part labels, where no semantic labels are pro + +duced on test shapes. Sharma et al. [28] perform projective shape analysis [36], where a fine-grained 3D segmentation is obtained by aggregating segmented multi-view 2D images. Both of these methods operate fully automatically after training, but their performance falls far short of $100\%$ accuracy: $32.6\%$ mean IoU score achieved by [28] on PartNet [20] with fine-grained (Level-3) parts (part counts ranging from 4 to 51), and below $50\%$ mean IoU by [33] on a more challenging dataset with part counts up to $500+$ . + +Instead of operating on low-level shape elements such as points [33] and pixels [28], the neurally-guided shape parser [10] (NGSP) is trained to assign fine-grained semantic labels to regions of a 3D shape with improved labeling accuracy: mean IoU score up to $58\%$ on PartNet. NGSP parses region label proposals generated by a guide neural network using likelihood modules to evaluate the global coherence of each proposal. It is still designed as a fully automatic method, but the authors did suggest the need for human-in-the-loop approaches to "scale beyond carefully-curated research datasets to 'in-the-wild' scenarios." + +Hierarchical analysis: A fine-grained shape understanding is also attainable via hierarchical structural analysis. Some early works perform rule-based reasoning to construct symmetry hierarchies [38] and shape grammars [17], while others work with geometric priors, e.g., convexity [2], for hierarchical shape approximation. Similar to fine-grained segmentation, most learned hierarchical models are also unsupervised or weakly supervised, including co-hierarchical analysis [32], hierarchical abstraction [30], recursive implicit fields [21], and inverse constructed solid geometry (CSG) [45, 14]. All of these methods aim for full automation and work well only with shallow hierarchies. + +Active learning for part labeling: The classical and core question for active learning [1, 47, 26] is: "How does one select instances from the underlying data to label, so as to achieve the most effective training for a given level of effort?" Past works have studied uncertainty and diversity for interactive segmentation and labeling of 3D shapes [35, 5], medical images [31], and large-scale point clouds [40], to name a few. We utilize active learning, aiming to approach full labeling accuracy, regardless whether the test shapes belong to a "research dataset" or are "in-the-wild." + +Most closely related to our work is the active framework for region annotations by Yi et al. [44]. Both works share the common goal of producing accurate, human-verified labels. On the other hand, their work imposes a fixed human effort budget, while ours does not. Technically, their work relies on a conditional random field to automatically propagate human labels, while our work integrates a DL module [37] into an active part labeling tool. The only prior work that combines deep and active learning for 3D segmentation [5] focuses on boundary refinement, just as the earlier work [35] whose learning module is constrained $k$ - + +means clustering. However, none of these tools were designed to work with fine-grained labels, e.g., they still only operate on four part labels for chairs [5, 35, 44, 41]. + +Finally, while our work is the first active learning tool for fine-grained 3D part labeling, hierarchical active learning has been employed in other data domains such as documents. In hierarchical text classification[16, 4], text labels can also be organized hierarchically. In comparison, semantic labeling of 3D parts in a shape collection poses various challenges including richer structural variations, part relations, and more costly and delicate user interactions. + +# 3. Method + +Given a shape test set $S$ , where each shape is presegmented into parts, our methods assign a label $l \in L$ to each part of each shape. The set of labels, $L$ , is predefined and organized in a tree structure such that coarse labels are at the top, and fine-grained labels at the bottom. The labeling progresses hierarchically top-to-bottom, assigning first coarse labels and refining them at each step, achieving high-accuracy fine-grained labeling results. + +At each node, we use an active learning based tool to iteratively engage human interactions into the part labeling process. With this tool, we obtain a close to error-free part labeling on each node on any test set. + +# 3.1. Hierarchical labeling strategy + +We organize the part labels into a tree structure based on their semantics, as in PartNet [20]. Following the same terminology, two types of internal tree nodes are defined: an "AND" node indicates "what part(s) it has" (e.g., a regular base has legs and a runner), while an "OR" node indicates "what type it is" (e.g., a star- or pedestal-type base). The labeling on "AND" nodes is based on part features, while the labeling on "OR" nodes depends on group features; see Fig. 3. Each leaf node presents a final part label. + +At each node of the tree, we combine a deep neural network with an active learning tool to perform part labeling. Consider a single shape $s_i \in S$ and its set of parts $P_i$ . At the root node, the input are all the parts of shape $s_i$ , that is $P_{root} = P_i$ . After finishing the labeling task at the root node, the set of parts is split into one or more subsets $P_l \subseteq P_{root}$ corresponding to the children labels $l$ . Then, each subset is input to the respective node for further labeling. By splitting the input set at each node, the size of the input decreases as the depth of the node increases. Finally, the labeling stops when we reach a leaf node. + +The complexity of our labeling scheme depends on the structure of the tree. On one extreme, if all labels are direct children of the root node, training at the root would be more complex but there is only one node to train. The other extreme is a binary tree, where each node is trained to distinguish between only two labels, but we have to train more + +nodes, resulting in an increased training time. As we have found the original PartNet trees to have too many internal nodes, we modified this initial tree structure using a heuristic rule: we remove internal "AND" nodes which contain less than three children and keep all "OR" nodes. This way, our tree structure is better balanced for our task to lead to reduced training time. Examples of our modified tree structures are provided in the supplementary material. + +# 3.2. Symmetry-assisted labeling + +Another novel feature to accelerate HAL3D makes use of detectable part symmetries on the test set. Symmetry is ubiquitous in human-made objects and it is common practice for 3D artists to copy-and-paste symmetric or repeated parts when building 3D assets. Symmetry detection has been a well-studied subject in geometry processing [19] and there are a variety of applicable techniques. For simplicity, we simply use sizes of the part oriented bounding boxes as a crude and conservative feature to detect part symmetries. + +As symmetric parts are expected to share the same label, we accelerate our active learning in two ways: if a proposal assigns different labels to symmetric parts, we automatically remove it from the "HC (High confidence)" set. Then, in the modification module, the user only needs to label one part of the symmetry group and our system automatically propagate it to the other parts. These features improve labeling efficiency by $20\%$ ; see Section 5.2. + +# 3.3. Active labeling of fine-grained 3D parts + +Fine-grained 3D shape segmentation and annotation is challenging because of the wide structural diversity found in small and intricate parts of each shape. This diversity makes it difficult for existing DL methods to achieve high prediction accuracy. On the other side, active learning techniques are commonly used by researchers to improve the performance of the learned classifier with the help of human efforts. We adopt these techniques and, at each node of our hierarchical labeling strategy, we use a deep active learning based method to achieve high part labeling accuracy. Our method comprises three main modules: the DL-based part label proposal network, the part label verification module, and the part label modification module. + +Consider a node $l$ in our tree structure, given the subset of parts labeled $l$ of each shape, we iteratively update the part labels until they all pass human verification. At each iteration, we first use a DL-based neural network to make part label proposals based on part label probabilities. Then, the HC proposals are passed to the part label verification module for human verification. Next, we collect the LC, i.e., low-confidence, proposals, which are more likely to be inaccurate, and pass them to the label modification module. Finally, the verified proposals and modified proposals are combined and feed back to the proposal network to fine- + +![](images/81dc34a885c63960c9cb62dd7f78eea27acc172fe3218bed86ce55062b35a313.jpg) + +![](images/a138e88054d68308ee3e38c29f851fe793b2764fee969279cee1b6e4a0be5f2b.jpg) +Figure 3. Label proposal architectures. Part label prediction is based on global and local, part-level features at the "AND" nodes and global features only for the "OR" nodes. The "AND" nodes would serve to predict the label for each part and the "OR" nodes would predict labels for the entire input shape. Notion-wise, $N$ is the number of points, $P$ is the number of parts, $K$ is the feature dimension, and $L$ is the number of labels. + +tune the network weights. The active learning iteration will stop if the number of left shapes are less than 40. We explain next how we make HC vs. LC proposals. + +Label proposal network. Given a set of parts from each shape $s_i \in S$ , we sample $N = 8,192$ points over the part surfaces and input them into our proposal network. As shown in Fig. 3, our network uses the edge convolution (EdgeConv) module from DGCNN [37] to extract point features from the point clouds. This module consists of five EdgeConv layers whose feature dimensions are 64, 64, 128, 256, and 256, respectively. The point features from all layers are concatenated and passed into a convolution layer. The size of the resulting feature is 512. + +Since our tree structure has different types of nodes, we apply two different network structures from DGCNN, where the architecture for "OR" nodes is the same as in the original DGCNN classification network. At an "AND" node, differently from DGCNN which is the SOTA method to predict point labels based on point features, we use extra max-pooling operations to aggregate the local part features and global shape features, as inspired by NGSP [10]. The concatenated features are fed into the part classifier, which contains 3 MLP layers to predict part labels. As for an + +"OR" node, labels of the input shape are predicted based on the global features only; see Fig. 3. Since different shapes have different part counts, we set the maximum part count $P_{max} = 150$ during training for batch processing and add zero-padding to the part feature matrix so that all the part feature matrices have the same dimension. More details of our network can be found in the supplementary material. + +Label verification. Our goal is to get the labels of each shape verified by a human. To reduce verification and training times, we need to decide which shapes are verified first and how many to verify on each iteration. Shapes in the test set that have similar structure to shapes in the train set are more likely to be labeled well. We can quickly select these well-labeled shapes from the test set and use them to enhance our limited train set, improving the generalization ability of our proposal network. In our method, we first sort the test set based on average predicted part label probabilities from the proposal network of each shape and arrange the test set into batches, each containing $B = 10$ shapes. Users can mark a shape as well-labeled if they see all parts are labeled correctly, or bad-labeled when one part is wrong. We design a user-friendly interface to achieve this, where users can view and rotate a batch of 3D shapes at the same time. The interface also shows the corresponding label colors for reference. Please see supplementary material for more details about our part label verification interface. + +Given the sorted shapes, we need to decide how many of them to show the user as HC proposals before fine-tuning the prediction, to achieve a balance between processing efficiency and utilization of user feedback. To this end, we adopt an adaptive strategy: given the sorted proposals, we stop the verification step when the number of human-verified shapes on a batch is below a threshold, which is set to 4 in our experiments. This simple strategy is effective since early batches tend to have higher labeling accuracy than later batches. This way, the number of high confidence proposals presented to the user would adapt to the classification purity of the current test set. + +Label modification. Shapes in the test set whose part structures and geometries differ significantly from those in the train set are less likely to be labeled accurately by the network. These shapes can be regarded as outliers in the label regression task, but they can be especially valuable during fine-tuning. Human intervention is inevitable in this situation to correct the labels. Specifically, the shapes provided to the label modification module are collected from two sets: the $Q_{1} = 20$ LC proposals with the lowest mean label probabilities, and those shapes that failed user verification more than $H = 2$ times. We ask users to modify the incorrect labels using our modification interface. + +In our part annotation interface, labels are organized hierarchically so users can quickly look up the correct labels. + +![](images/d90bc4b873ecf52dd1365638070b576a3653beac36e1c54d3d7a471901c3fdce.jpg) +Figure 4. Comparison of pre-segmented parts in PartNet and ABO. ABO shapes were pre-segmented into connected components where each component may correspond to multiple labels, e.g., the back and the arm of the chair are grouped into the same component, as indicated by the red color in the second column. Instead, our part labeling operates on the convex pieces obtained by a decomposition. In the first three figures, the colors are used to distinguish between different parts, and we can see that the parts in PartNet are manually segmented and exhibiting imperfect boundaries. In the last figure, the colors reflect semantic labels. + +The label-to-color mapping table is also provided to help find the incorrect labels. More details about this interface are provided in the supplementary material. + +Shapes that have been verified and modified by humans are added into the initial training set together to fine-tune our proposal network. On the other hand, we skip shapes excluded from the HC and LC proposals. These shapes are combined with shapes that failed to pass the verification less than $H$ times to go to the next iteration for labeling. + +# 4. Datasets and Metrics + +Datasets. The Stanford PartNet [20] dataset is a common choice for fine-grained shape analysis [34, 10, 28]. We follow NGSP [10] and use the four major categories from PartNet for evaluation: chair, table, lamp, and storage furniture. Fine-grained labels in PartNet are organized into three granularity levels, and we adopt the second level in our experiments as in NGSP [10]. We first filter out noisy shapes with erroneous labels. Then from each category, we take 500 shapes and split them into 50/50/400 as train/val/test sets. + +We also evaluate HAL3D on the recent Amazon Berkeley Objects (ABO) dataset, which is composed of about 8,000 high-quality 3D shapes created by artists from real product images, where each 3D model is a combination of build-aware connected components. Unlike the 3D shapes from PartNet in which each part corresponds to one single semantic label, a build-aware component from ABO may be appropriately associated with multiple labels, as shown in Fig. 4. To alleviate this issue, we apply an approximate convex decomposition [39] to further decompose each connected component into a fine-grained, i.e., over-segmented, set of convexes, where each convex piece is deemed to receive one and only one part label. Our active labeling is then applied to the set of convexes. When running the con + +vex decomposition algorithm, we use the default parameters from the provided code, except for the concavity threshold setting, which is set as 0.2 in our experiments. + +As GT labels are unavailable in ABO, three artists were hired to create them for the decomposed convex components. We built a test set containing 1,800 shapes from five categories: chair, table, lamp, storage furniture, and bed, with 200 for storage and 400 each for the other categories. + +Metrics. Active learning systems are typically evaluated from two perspectives: labeling accuracy and the amount of human efforts or involvements. We employ two standard metrics to assess the labeling quality: (1) part label accuracy measuring how many parts are correctly labeled and (2) point cloud segmentation mIoU depicting how much of the shape surface area is correctly labeled. + +The amount of human efforts is measured by the consumed time in the steps of label verification and modification. Inspired from [44], we firstly give artists the same training courses and tools, and conduct user studies on Part-Net. We report the total recorded labeling time in Tab. 1. We also collect the operations they used in the user study and model the total human interaction time as a function of the number of processed parts and shapes: + +$$ +t = T _ {v} ^ {p} \left(P _ {v}, L\right) + T _ {v} ^ {f} \left(S _ {v}, L\right) + T _ {v} ^ {p} \left(P _ {m} ^ {c}, L\right) + T _ {m} ^ {p} \left(P _ {m} ^ {w}, L\right), \tag {1} +$$ + +where $P_{v}$ is the number of correctly labeled parts in verification, $S_{v}$ is the number of incorrectly shapes in verification, $P_{m}^{c}$ is the number of correctly labeled parts in modification, $P_{m}^{w}$ is the number of incorrectly labeled parts in modification, and finally $L$ is the total number of labels. + +Function (1) is devised based on several observations: a) users need to go through and verify all the part labels to ensure accuracy; b) they can easily pick up failure cases during verification if any part is incorrectly labeled, without checking other parts; and c) users need to check all the parts in each shape to know whether modification is needed. + +The formulas and their coefficients in (2) are obtained via linear regression using the total operation and labeling times collected from all annotators in our user study: + +$$ +T _ {v} ^ {p} \left(P _ {v}, L\right) = \left(0. 3 1 + 0. 0 3 * L\right) * P _ {v}, +$$ + +$$ +\begin{array}{l} T _ {v} ^ {f} \left(S _ {v}, L\right) = (1. 4 7 + 0. 0 5 * L) * S _ {v}, \\ \text {T r} \left(R _ {v} ^ {f}, L\right) = \left(0. 0 4 - 0. 0 2, L\right) * R _ {v} ^ {f} \end{array} \tag {2} +$$ + +$$ +T _ {v} ^ {p} \left(P _ {m} ^ {c}, L\right) = \left(0. 3 1 + 0. 0 3 * L\right) * P _ {m} ^ {c}, +$$ + +$$ +T _ {m} ^ {p} \left(P _ {m} ^ {w}, L\right) = (5. 2 8 + 0. 2 3 * L) * P _ {m} ^ {w}, +$$ + +where $T_{v}^{p}$ is the checking time for correctly labeled parts, $T_{v}^{f}$ is the verification time for incorrectly labeled shapes, and $T_{m}^{p}$ is the time for part label modification. We use these functions to evaluate the human interaction timing, in seconds, for hyper-parameters analysis on PartNet. We also report additional machine time cost from training and fine-tuning in the supplementary material. + +# 5. Experiments + +We have implemented our approach in PyTorch. The label proposal network at each node is pre-trained with the Adam optimizer on the training set of PartNet for 250 epochs, where the learning rate is 0.001 and it is decreased by 0.8 every 25 epochs. When fine-tuning the network in HAL3D, the learning rate is fixed to 0.0001, the maximum epoch number is 125, and the validation set is used for early stopping. Our experiments were ran on an NVIDIA Tesla V100 GPU, and the training time varies from 20 minutes to 3 hours according to the number of parts at each node. + +# 5.1. Competing methods + +We compare HAL3D, our hierarchical active labeling tool, with two automatic part labeling methods, + +- PartNet [20], which employs the PointNet++ [25] backbone for point-wise feature extraction and labels parts by a voting scheme over points belonging to the same part. +- NGSP [10], which aggregates point-wise features via PointNet++ [25] to obtain part features and then leverages a likelihood-aware beam search for optimization; this represents the state of the art for fine-grained labeling. + +And two variants of our labeling approach, + +- Our prop. is one such variant without using the manual label verification or modification module; it predicts part labels using the label proposal network only. +- DAL3D is a deep active learning variant of our approach without using the hierarchical labeling strategy. + +# 5.2. Evaluation on PartNet + +All methods presented in this section are pre-trained and tested on the same training and test sets of PartNet. + +Quantitative comparison. Tab. 1 compares the three automatic methods, where PartNet is inferior to both NGSP and our label proposal network due to its hand-crafted voting mechanism. Our simple network achieves comparable accuracy compared to NGSP (difference by 0.11 and 3.49 with respect to mIoU and part accuracy, respectively). However, NGSP suffers from significantly longer running time due to the intensive beam search; see the supplementary for a running time comparison. In contrast, our proposal network is effective and efficient, i.e., capable of quickly producing accurate initial labels to reduce the manual workload in verification and modification for active learning. + +Tab. 1 also shows that even with the sophisticated design in NGSP, the highest accuracy attained is only $70\%$ . Hence the performance of automatic fine-grained labeling still falls far below that of active learning, where the largest and smallest accuracy gap is $40.77\%$ and $22.49\%$ between NGSP and HAL3D, respectively. + +Table 1. Quantitative comparison against the competing labeling methods and variants. In the table, "AL" indicates whether a method uses active learning. Since PartNet, NGSP, and Our prop. are automatic methods, their entries are marked with a '-' instead. "Accu." denotes the part label accuracy, "mIoU" denotes the mean Intersection-over-Union, and "Lab. time" denotes the labeling time (hours in total) for the active learning methods. + +
MethodALAccu.↑mIoU↑Lab. time↓(h)
Chair
PartNet [20]-54.9031.04-
NGSP [10]-68.9842.90-
Our prop.-67.4541.34-
DAL3D✓89.8472.305.99
HAL3D✓94.1384.924.34
Table
PartNet [20]-33.2114.92-
NGSP [10]-47.1127.94-
Our prop.-42.7723.50-
DAL3D✓78.5369.829.14
HAL3D✓84.1874.654.41
Storage furniture
PartNet [20]-55.8230.86-
NGSP [10]-70.6045.79-
Our prop.-69.8242.43-
DAL3D✓84.5771.994.81
HAL3D✓93.0988.723.87
Lamp
PartNet [20]-38.2811.02-
NGSP [10]-49.7623.49-
Our prop.-44.7719.27-
DAL3D✓83.1469.354.52
HAL3D✓90.5379.512.16
+ +Table 2. Ablation on key modules of HAL3D for chair category. + +
Row IDProp.Hier.Sym.ALLab-T↓Accu↑
2nd----22.0589.16
3rd✓---8.6588.53
4th✓✓✓-6.3793.87
5th✓-✓✓5.9989.84
6th✓✓-✓5.2193.45
7th✓✓✓✓4.3494.13
+ +HAL3D consistently outperforms our DAL3D variant in all metrics, particularly 2.42 hours faster on average on labeling time, demonstrating the importance of our hierarchical labeling order for reducing manual workload. We also provide qualitative comparison in Fig. 5. + +Ablation study. Tab. 2 verifies the contributions of the main components of HAL3D on minimizing human labeling efforts, a key criterion for success for a human-in-the-loop approach. The second to fifth columns indicate: + +![](images/5c1e75107ec5e6391a49c469ec5e33050be31affb8228fbafb28a8d366c373d5.jpg) +Figure 5. Visual comparison between labeling methods. HAL3D can achieve results approaching GT, barring human errors. The only mis-labeled part by HAL3D here is the bottom panel of the table (red oval), which could be easily labeled as shelf by users. + +![](images/8c544f81bcf3b438fbe3703bd64fb85931533191054a39af0e796aa615ea1157.jpg) +Figure 6. A plot of the number of human-verified shapes (max = 400) at different nodes (only those which existed in more than 100 shapes) of the chair category. The plot shows that with hierarchical active learning, many more shapes pass the verification, hence saving on the costly label modification, especially at higher hierarchy levels (i.e., earlier in the labeling iterations). + +- Prop. if the label proposal network is used; +- Hier. if the hierarchical labeling strategy is used; +- Sym. if the symmetry checking is used. +- AL. if the active learning is used. + +We compare our full approach against five baselines. The second row is a baseline that directly predicts labels by annotating with our label modification interface. The third row is a baseline that firstly uses the label proposal network and then manually modifies incorrect part labels. The fourth, fifth, and sixth rows represent variants of HAL3D without hierarchical labeling, symmetry checking, and active learning, respectively. The results in Tab. 2 clearly justify the design choices in our method, where our full approach (last row) is the most efficient. + +To further evaluate the effectiveness of our hierarchical labeling, we plot in Fig. 6 the number of shapes which passed human verification over the active learning qua + +Table 3. Inference time comparison between NGSP and our label proposal network. We report average seconds per shape. + +
CategoryChairTableLampStorage
NGSP5.41018.0112.6912.87
Our prop.0.0120.0140.0120.011
+ +tions. As we can see, HAL3D consistently verified more shapes as correct at all tree nodes when comparing to the baseline that had no hierarchical design, e.g., more than 300 shapes were marked as correct in the first iteration in the chair_base node. Essentially, our hierarchical verification splits the labeling task into smaller and easier tasks compared to the case of non-hierarchical design, which effectively reduces the burden arising from the subsequent and more costly modification step in terms of human effort. + +In the supplementary material, we perform ablations on the four hyper parameters and thresholds employed in our method. In all the experiments, we fix these parameters and set them based only on a small dataset. A more careful parameter tuning should lead to better results. + +Inference time comparison. Although NGSP achieved higher prediction accuracy than our proposal network, as shown in Tab. 1, our network is more efficient and better suited to be incorporated into an active learning framework. Tab. 3 compares the inference times between NGSP and our proposal network. NGSP is highly time-consuming due to the use of beam search, whereas our label proposal network takes only 12ms per shape on average. + +# 5.3. Evaluation on ABO + +We apply HAL3D to label our constructed test set for ABO, where the label proposal network is pre-trained on the training set of PartNet. As mentioned in Sec. 4, shapes in ABO are decomposed into convex shapes and HAL3D assigns a semantic label to each decomposed piece. Three artists were hired to use our HAL3D tool and interface for evaluation. Since GT labels are not provided in ABO, we inspect the labeling results from the artists via human cross validation with the final corrected labeling forming the GT. As different artists may report different accuracy and recorded labeling times, we report averages in the paper. The artists worked together with an expert to determine labels for uncertain shapes. The visual results from HAL3D, as shown in Fig. 7 and the supplemental material, were obtained on the test set by taking majority voting from labels provided by different artists. + +As shown by quantitative results in Tab. 4, HAL3D can accurately label all categories with a $94\%$ average accuracy using 3-7 hours of manual annotation, demonstrating the effectiveness and efficiency of our method. The reduction on human efforts against manual annotation becomes more significant when comparing against the last four rows in Tab. 4, + +![](images/3b49e1a85c46503b978f5a0c239aa3ccae1f14c0aabbb80d78f39b58a0392ece.jpg) + +![](images/a437c65889b9456358cb69672ebe761430f3c1570568fb031840fd4c643f9943.jpg) +Figure 7. Visual comparison between several part labeling methods on the ABO dataset. + +![](images/3ac5519016de7b6d167d5eed1e5cb3557ea73f3c151f69553067a808903bce2a.jpg) + +![](images/5f331945387e14b6e310897453f265c0cf60d137af862fc3b00a085a9fa5caca.jpg) + +![](images/41cb17aebc1aba0b33304d3aa25e85e3c488feb7bc23a7b208c823d027dd8b5a.jpg) + +![](images/a509111814633060d3a97fd8c3017776107e5f0acf8a1d9feb6025e609fd8e60.jpg) + +Table 4. Comparison for different methods to pass human verification on test set of ABO. The time is in hours. + +
ChairTableLampCabinetBed
# shapes400400400200400
HAL3D
Time ↓Accu. ↑6.4693.077.4192.554.3896.293.3494.484.5193.71
PartNet + modification
Time ↓Accu. ↑28.2588.9633.9486.5131.5390.9113.3590.2425.0289.37
NGSP + modification
Time ↓Accu. ↑28.3088.7734.3286.4031.5691.3713.2290.5624.9089.93
+ +where we use the same manual modification interface to directly correct the PartNet and NGSP results. Please see supplementary material for results on ABO. + +# 6. Conclusion + +We present HAL3D, the first semantic labeling tool that is designed to operate on fine-grained 3D parts and achieve close to error-free labeling through full verification by humans, barring human misjudgement. Our human-in-the-loop approach is based on active learning, combining deep label prediction and human inputs to iteratively fine-tune and improve the network prediction. Extensive experiments + +demonstrate that human efforts are effectively reduced via hierarchical and symmetry-aware active labeling. + +HAL3D is currently implemented with DGCNN [37] as the label prediction backbone, but it can be readily replaced by PointNet [24], which was adopted by PartNet [20] and NGSP [10], or a more performant alternative such as Point Transformer [48]. Aside from symmetry, inter-shape part correspondences can also be utilized to reduce labeling costs, assuming that they can be reliably obtained. In addition, our tool does not yet allow splitting of under-segmented parts. Since performing such operations by humans is time-consuming, incorporating auto-splitting schemes into HAL3D deserves future consideration. + +We regard our work as only a preliminary step towards deep active learning for fine-grained structural analysis of 3D shapes. A natural focus for future work is to develop a more sophisticated ranking or selection mechanism for label verification and modification to achieve the most effective training for a given level of human effort. + +# Acknowledgement + +We thank all the anonymous reviewers and area chairs for their constructive comments and the Amazon artists for performing the human annotations. Thanks also go to Jeff-Aram Van Gorp for helpful discussions and other support. + +# References + +[1] Charu C. Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and Philip S. Yu. Active learning: A survey. In Data Classification: Algorithms and Applications, pages 571-597. 3 +[2] Marco Attene, Michela Mortara, Michela Spagnuolo, and Bianca Falcidieno. Hierarchical convex approximation of 3D shapes for fast region selection. Comput. Graph. Forum, 27(5):1323-1332, 2008. 3 +[3] Zhiqin Chen, Kangxue Yin, Matt Fisher, Siddhartha Chaudhuri, and Hao Zhang. BAE-NET: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 3 +[4] Yu Cheng, Kunpeng Zhang, Yusheng Xie, Ankit Agrawal, and Alok Choudhary. On active learning in hierarchical classification. In Proceedings of the 21st ACM International Conference on Information and knowledge Management, pages 2467-2470, 2012. 4 +[5] David George, Xianghua Xie, Yukun Lai, and Gary KL Tam. A deep learning driven active framework for segmentation of large 3d shape collections. Computer-Aided Design, 144:103179, 2022. 3, 4 +[6] Aleksey Golovinskiy and Thomas Funkhouser. Consistent segmentation of 3D models. In Proc. Shape Modeling International (SMI), 2009. 3 +[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: A network with an edge. ACM TOG, 38(4), 2019. 3 +[8] Yong He, Hongshan Yu, Xiaoyan Liu, Zhengeng Yang, Wei Sun, Yaonan Wang, Qiang Fu, Yanmei Zou, and Ajmal Mian. Deep learning based 3d segmentation: A survey. CoRR, abs/2103.05423, 2021. 1, 3 +[9] Ruizhen Hu, Lubin Fan, and Ligang Liu. Co-segmentation of 3D shapes via subspace clustering. Comput. Graph. Forum, 31(5):1703--1713, 2012. 3 +[10] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3D shape regions with approximate inference. In CVPR, 2022. 3, 5, 6, 7, 9 +[11] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. ACM Transactions on Graphics (TOG), 41(6):1-11, 2022. 3 +[12] Oliver Van Kaick, Noa Fish, Yanir Kleiman, Shmuel Asafi, and Daniel Cohen-OR. Shape segmentation by approximate convexity analysis. ACM TOG, 34(1). 3 +[13] Evangelos Kalogerakis, Aaron Hertzmann, and Karan Singh. Learning 3D mesh segmentation and labeling. ACM TOG, 29(4), 2010. 3 +[14] Kacper Kania, Maciej Zięba, and Tomasz Kajdanowicz. UCSG-NET - unsupervised discovering of constructive solid geometry tree. In NeurIPS, 2020. 3 +[15] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Neural star domain as primitive representation. Advances in Neural Information Processing Systems, 33:7875-7886, 2020. 3 +[16] Xiao Li, Da Kuang, and Charles X Ling. Active learning for hierarchical text classification. In Pacific-Asia conference on knowledge discovery and data mining, pages 14–25. Springer, 2012. 4 + +[17] Tianqiang Liu, Siddhartha Chaudhuri, Vladimir G. Kim, Qixing Huang, Niloy J. Mitra, and Thomas Funkhouser. Creating consistent scene graphs using a probabilistic grammar. ACM TOG, 33, 2014. 3 +[18] Niloy Mitra, Michael Wand, Hao Zhang, Daniel Cohen-Or, and Martin Bokeloh. Structure-aware shape processing. In Eurographics State-of-the-Art Report (STAR), 2013. 1 +[19] Niloy J. Mitra, Mark Pauly, Michael Wand, and Duygu Ceylan. Symmetry in 3D geometry: Extraction and applications. 32(6), 2013. 4 +[20] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, June 2019. 2, 3, 4, 6, 7, 9 +[21] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. RIM-Net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 3 +[22] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 3 +[23] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, pages 10344—10353, 2019. 3 +[24] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 3, 9 +[25] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 7 +[26] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B. Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning, 2020. 3 +[27] Ariel Shamir. A survey on mesh segmentation techniques. Comput. Graph. Forum, 27(6):1539-1556. 1 +[28] Gopal Sharma, Kangxue Yin, Subhransu Maji, Evangelos Kalogerakis, Or Litany, and Sanja Fidler. MvDeCor: Multiview dense correspondence learning for fine-grained 3D segmentation. In ECCV, 2022. 3, 6 +[29] Oana Sidi, Oliver van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. ACM TOG, 30(6):126:1-126:9, 2011. 3 +[30] Chun-Yu Sun, Qian-Fang Zou, Xin Tong, and Yang Liu. Learning adaptive hierarchical cuboid abstractions of 3D shape collections. 38(6), 2019. 3 +[31] Andrew Top, Ghassan Hamarneh, and Rafeef Abugharbieh. Active learning for interactive 3D image segmentation. In Int. Conf. on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2011. 3 +[32] Oliver van Kaick, Kai Xu, Hao Zhang, Yanzhen Wang, Shuyang Sun, Ariel Shamir, and Daniel Cohen-Or. Cohierarchical analysis of shape structures. ACM TOG, 32(4), 2013. 3 + +[33] Xiaogang Wang, Xun Sun, Xinyu Cao, Kai Xu, and Bin Zhou. Learning fine-grained segmentation of 3D shapes without part labels. In CVPR, 2021. 1, 3 +[34] Xiaogang Wang, Bin Zhou, Haiyue Fang, Xiaowu Chen, Qinping Zhao, and Kai Xu. Learning to group and label fine-grained shape components. 37(6), 2018. 6 +[35] Yunhai Wang, Shmulik Asafi, Oliver van Kaick, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Active co-analysis of a set of shapes. ACM TOG, 31(6), 2012. 3, 4 +[36] Yunhai Wang, Minglun Gong, Tianhua Wang, Daniel Cohen-Or, Hao Zhang, and Baoquan Chen. Projective analysis for 3D shape segmentation. ACM TOG, 32(6), 2013. 3 +[37] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM TOG, 2019. 2, 3, 5, 9 +[38] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. Comput. Graph. Forum, 30(2):287-296, 2011. 3 +[39] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search. ACM TOG, 41(4), 2022. 2, 3, 6 +[40] Tsung-Han Wu, Yueh-Cheng Liu1, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H. Hsu. ReDAL: Region-based and diversity-aware active learning for point cloud semantic segmentation. In ICCV, 2021. 3 +[41] Zizhao Wu, Ruyang Shou, Yunhai Wang, and Xinguo Liu. Interactive shape co-segmentation via label propagation. In Proc. of CAD/Graphics, 2014. 4 +[42] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM TOG, 40(4), 2021. 3 +[43] Li Yi, Leonidas Guibas, Aaron Hertzmann, Vladimir G. Kim, Hao Su, and Ersin Yumer. Learning hierarchical shape segmentation and labeling from online repositories. SIGGRAPH, 2017. 3 +[44] Li Yi, Vladimir G. Kim, I.-Chao Shen Duygu Ceylan, Mengyuan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6), 2016. 1, 3, 4, 6 +[45] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact CAD shapes with adaptive primitive assembly. In CVPR, 2022. 3 +[46] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. PartNet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, 2019. 3 +[47] Xueying Zhan, Qingzhong Wang, Kuan-hao Huang, Haoyi Xiong, Dejing Dou, and Antoni B. Chan. A comparative survey of deep active learning, 2022. 3 +[48] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In CVPR, 2021. 9 + +[49] Yang Zhou, Kangxue Yin, Hui Huang, Hao Zhang, Minglun Gong, and Daniel Cohen-Or. Generalized cylinder decomposition. ACM TOG, 34(6), 2015. 3 +[50] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J. Guibas, and Hao Zhang. AdaCoSeg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 3 \ No newline at end of file diff --git a/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/images.zip b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..66dfc7722cfbeedcf62986c686b2cbc7a818b219 --- /dev/null +++ b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9565c4f2f03dba8804238fb10fae8478a4ee8253239c7067ea04853e85cf2161 +size 501793 diff --git a/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/layout.json b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..949b92bd2134b25911c88fb2aad55734e608b31b --- /dev/null +++ b/2023/HAL3D_ Hierarchical Active Learning for Fine-Grained 3D Part Labeling/layout.json @@ -0,0 +1,8184 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 64, + 525, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 64, + 525, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 64, + 525, + 83 + ], + "type": "text", + "content": "HAL3D: Hierarchical Active Learning for Fine-Grained 3D Part Labeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "spans": [ + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "type": "text", + "content": "Fenggen " + }, + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1,2*}" + }, + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "type": "text", + "content": " Yiming Qian" + }, + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "type": "text", + "content": " Francisca Gil-Ureta" + }, + { + "bbox": [ + 149, + 106, + 442, + 121 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 121, + 437, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 121, + 437, + 134 + ], + "spans": [ + { + "bbox": [ + 180, + 121, + 437, + 134 + ], + "type": "text", + "content": "Brian Jackson1 Eric Bennett1 Hao Zhang1,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 135, + 397, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 135, + 397, + 149 + ], + "spans": [ + { + "bbox": [ + 198, + 135, + 397, + 149 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 135, + 397, + 149 + ], + "type": "text", + "content": "Amazon " + }, + { + "bbox": [ + 198, + 135, + 397, + 149 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 135, + 397, + 149 + ], + "type": "text", + "content": " Simon Fraser University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 176, + 192, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 176, + 192, + 189 + ], + "spans": [ + { + "bbox": [ + 143, + 176, + 192, + 189 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 202, + 290, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 202, + 290, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 290, + 395 + ], + "type": "text", + "content": "We present the first active learning tool for fine-grained 3D part labeling, a problem which challenges even the most advanced deep learning (DL) methods due to the significant structural variations among the intricate parts. For the same reason, the necessary effort to annotate training data is tremendous, motivating approaches to minimize human involvement. Our labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. To effectively reduce human efforts, we develop two novel features in our tool, hierarchical and symmetry-aware active labeling. Our human-in-the-loop approach, coined HAL3D, achieves close to error-free fine-grained annotations on any test set with pre-defined hierarchical part labels, with " + }, + { + "bbox": [ + 46, + 202, + 290, + 395 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 202, + 290, + 395 + ], + "type": "text", + "content": " time-saving over manual effort. We will release the finely labeled models to serve the community." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 418, + 128, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 128, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 128, + 431 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 440, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 643 + ], + "type": "text", + "content": "Semantic shape segmentation and labeling [8, 27] is a classical problem with numerous applications [18], including shape/part recognition, retrieval, indexing, and attribute transfer. We consider the problem of fine-grained 3D part labeling, a task that has not received as much attention as the typical semantic segmentation into coarse/major parts (e.g., backs, seats, and legs for chairs). While these coarse semantic groups support high-level visual perception, they do not possess sufficient granularity to address shape properties related to motion, function, interaction, or construction. Indeed, fine-grained object parts, e.g., the individual wheels and the small mechanical parts in a swivel leg, are both function-revealing and build-aware (see Fig. 1), with strong ties to how the objects were physically assembled or virtually built by artists. Real-world applications such as product inspection, assembly, customization, and robot interaction all operate on fine-grained object parts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "content": "Learning to label fine-grained parts is a highly challenging problem due to the great structural variations among the small and intricate parts, even for shapes that belong to the same category, as shown in Fig. 1. On a challenging" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 175, + 545, + 299 + ], + "blocks": [ + { + "bbox": [ + 309, + 175, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 309, + 175, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 309, + 175, + 545, + 299 + ], + "type": "image", + "image_path": "49ad14b908fdcd3535a0d9333c611b64806e40137f3da5dbb12c277b7a6036d4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 301, + 546, + 368 + ], + "lines": [ + { + "bbox": [ + 305, + 301, + 546, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 546, + 368 + ], + "type": "text", + "content": "Figure 1. Fine-grained parts of the 3D models shown exhibit the kind of geometric and structural complexity and diversity that part labeling has to handle. No existing methods, whether learned or heuristic-based, could obtain close-to-fully-accurate labeling in these challenging cases, while our human-in-the-loop active learning tool can approach full accuracy, barring human errors." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 533 + ], + "type": "text", + "content": "dataset with average part counts per category ranging from 15 to more than 500, state-of-the-art learned models such as [33] can only achieve mean IoU scores below " + }, + { + "bbox": [ + 304, + 390, + 546, + 533 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 390, + 546, + 533 + ], + "type": "text", + "content": ", out of a maximum of " + }, + { + "bbox": [ + 304, + 390, + 546, + 533 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 390, + 546, + 533 + ], + "type": "text", + "content": " for ground truth (GT) results. Also challenging is the tremendous data annotation effort needed to train methods to segment and label fine-grained parts. On the other hand, for applications which demand de facto full accuracy, e.g., for diagnoses in medicine and quality assurance in e-commerce, one can hardly rely on a fully automatic method to reach the mark. It is arguable that the only means to attain full accuracy is to have humans validate all the results, as in the case of data annotation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 535, + 547, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 547, + 691 + ], + "type": "text", + "content": "In this paper, we present the first active learning tool for fine-grained 3D part labeling. Given a set of 3D shapes segmented into fine parts, our labeling tool assigns one of the predefined labels to each part. These input parts are deemed atomic (i.e., indivisible); they can be as low-level as mesh triangles and as high-level as results obtained by an unlabelled semantic segmentation, or mid-level components from an over-segmentation. In general, active learning iterates between automated labeling and human input for label rectification [44]. Compared to conventional learned models with full autonomy, such a human-in-the-loop approach provides a viable option to achieve close to error-free part labeling on test sets, barring human errors1." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 214, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 214, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 214, + 712 + ], + "type": "text", + "content": "*Work carried out during internship at Amazon." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 702, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 545, + 713 + ], + "type": "text", + "content": "1During testing, accuracy is measured against GT labeling which has" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "865" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 542, + 232 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 232 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 232 + ], + "type": "image", + "image_path": "e13cce0fef38766eeff728302de76c2575fac608ccbf640bb699e4aba298d5c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "text", + "content": "Figure 2. Our human-in-the-loop, hierarchical active learning (HAL3D) tool for fine-grained 3D part labeling. The input consists of a set of test shapes each pre-segmented into parts. The labeling proceeds hierarchically, following a tree structure (left) that organizes the hierarchical part labels, from coarse labels (top) to fine-grained labels (bottom). For a node in the label tree (with label " + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "inline_equation", + "content": "l =" + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "text", + "content": " \"Regular Leg Base\" in the illustration), the input " + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "text", + "content": " is the subset of parts, from the entire set of shapes, that are labeled " + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "text", + "content": " by its parent. When labeling parts with " + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "text", + "content": ", the label proposal module first assigns refined labels for each part. Then, proposals are sorted by the mean label probability over the parts of each shape, with the high-confidence (HC) proposals passed to the verification step, which stops once the rate of failed shapes reaches a threshold. The low-confidence (LC) proposals are passed to the label modification module. Correctly labeled shapes fine-tune the network, while skipped and failed shapes, in set " + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 46, + 237, + 548, + 338 + ], + "type": "text", + "content": ", go to the next iteration for labeling. The iterations terminate when all shapes have passed human validation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 289, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 500 + ], + "type": "text", + "content": "As shown in Fig. 2, our interactive labeling tool iteratively verifies or modifies part labels predicted by a deep neural network, with human feedback continually improving the network prediction. Specifically, our system consists of three modules for label proposal, verification, and modification. Label proposals are produced by dynamic graph CNN (DGCNN) [37], with the resulting label probabilities dictating whether to pass the predicted labels to the verification or modification module. Both human-verified and corrected labels are passed back to the label proposal network to fine-tune it. The iteration terminates once all part labels have passed human verification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 501, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 287, + 537 + ], + "type": "text", + "content": "As the key criterion for success with active learning is the minimization of human effort, we design and incorporate two novel features into our labeling tool:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 549, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 549, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 58, + 549, + 287, + 646 + ], + "type": "text", + "content": "- Hierarchical labeling: We organize all the part labels in a hierarchical tree structure, which guides our prediction-verification-modification process so that labels further down in the hierarchy are dealt with only after their parent labels have been fully verified. This leads to labeling efficiency, since our labeling process is coarse-to-fine, with fewer labels to process in the coarser levels of the hierarchy. In addition, with both" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 324, + 357, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 357, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 324, + 357, + 545, + 429 + ], + "type": "text", + "content": "human users and the prediction network only needing to deal with labels at the same level of the hierarchy rather than across all levels, there is less load on the users to reduce possible errors and the prediction accuracy improves, which reduces the amount of label modifications, which are the most expensive task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 438, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 547, + 473 + ], + "type": "text", + "content": "- Part symmetry: Since symmetric parts receive the same label, we employ detected symmetries to constrain and facilitate both label verification and modification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 486, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 605 + ], + "type": "text", + "content": "Our hierarchical active learning tool is coined HAL3D. We adopt the hierarchical labels from PartNet [20] and train and test our labeling tool on both PartNet, which comes with semantic part segmentations, and Amazon-Berkeley Objects (ABO). ABO is a recently published dataset of high-quality, artist-created 3D models of real products sold online. These models are composed of build-aware connected components, However, since some of these components are too coarse, we first obtain an over-segmentation via convex decomposition [39] and then apply HAL3D." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "We evaluate the core components of HAL3D, i.e., the label proposal network being fine-tuned by human feedback, and both hierarchical and symmetry-aware active labeling, by comparing to baselines and performing an ablation. The results demonstrate clear efficiency of HAL3D over labeling tools without active learning and non-hierarchical designs. Overall, HAL3D achieves " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " speed-up over manual fine-grained labeling, on the PartNet test set. A supplementary video shows our interactive tool at work. We will" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": "been produced by humans. To achieve full accuracy, two criteria must be met: (a) the testing human labeler's judgement of what is the correct labeling agrees with the GT; (b) the labeler does not make any error, e.g., due to carelessness or time pressure. We regard both types of violations as \"human errors.\" One may also refer to the former as an inconsistency between human labeling, which is likely to occur near fuzzy part boundaries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "866" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "release the finely labeled ABO models to serve the community pending permission from the dataset owner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 107, + 134, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 107, + 134, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 107, + 134, + 119 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 127, + 287, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 287, + 186 + ], + "type": "text", + "content": "Most approaches to semantic 3D shape segmentation and labeling have been designed to reason about coarse- or high-level structures and to target fully autonomy. Considerably less work has been devoted to fine-grained segmentation and labeling or human-in-the-loop approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 187, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 287, + 330 + ], + "type": "text", + "content": "Coarse 3D shape segmentation and labeling: Since the seminal work on learning 3D mesh segmentation and labeling [13] in 2010, many learning methods have been proposed [8]. Prominent supervised methods such as PointNet [24], PointNet++ [25], and DGCNN [37] perform feature learning over point clouds, while MeshCNN [7] develops convolution and pooling layers that operate on mesh edges. These methods, among many more, produce both parts and part labels, but they have only been trained and tested on datasets with coarse segmentations, such as the ShapeNet part [44] and the COSEG datasets [35]; most object categories therein were labeled with only 2-5 parts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 474 + ], + "type": "text", + "content": "High-level parts, e.g., the legs, seats, and backs of chairs, typically possess sufficient structural consistency (e.g., in terms of relative positioning), hence predictability, to facilitate supervised learning. Most unsupervised or weakly supervised approaches also leverage such consistencies. In particular, earlier methods for co-segmentation, i.e., segmenting a set of related shapes altogether, all rely on alignment or clustering, either in the spatial domain [6] or feature spaces [29, 9]. More recently, resorting to DL, AdaCoseg [50] optimizes a group consistency loss defined by matrix ranks. BAE-Net [3] and RIM-Net [21] obtain co-segmentations by learning branched neural fields." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 605 + ], + "type": "text", + "content": "Fine-grained segmentation: Fine-level parts are important in many real-world applications since, whether large or small, each of them has a function to serve and often a motion to carry out. However, the structure consistency possessed by major semantic groups is mostly lost among fine parts. Such parts could be obtained via shape decomposition based on geometric properties such as convexity [12, 39], cylindricity [49], or through primitive fitting using cuboids [42, 30] or quadrics [23, 45], deformed spheres [22, 15]. However, the resulting segmentations are neither semantic nor labeled." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Most learning-based solutions are unsupervised or self-supervised, mainly because annotating fine-grained parts is an extremely tedious task. Li et al. [43] learns part labeling from noisy online part graphs annotated by human artists. Yu et al. [46] trains a recursive part decomposition network for category-specific, fine-grained, hierarchical shape segmentation. Wang et al. [33] and SHRED [11] develop deep clustering to learn part priors from fine-grained segmentation without part labels, where no semantic labels are pro" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "duced on test shapes. Sharma et al. [28] perform projective shape analysis [36], where a fine-grained 3D segmentation is obtained by aggregating segmented multi-view 2D images. Both of these methods operate fully automatically after training, but their performance falls far short of " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " accuracy: " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " mean IoU score achieved by [28] on PartNet [20] with fine-grained (Level-3) parts (part counts ranging from 4 to 51), and below " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " mean IoU by [33] on a more challenging dataset with part counts up to " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "500+" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 182, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 546, + 312 + ], + "type": "text", + "content": "Instead of operating on low-level shape elements such as points [33] and pixels [28], the neurally-guided shape parser [10] (NGSP) is trained to assign fine-grained semantic labels to regions of a 3D shape with improved labeling accuracy: mean IoU score up to " + }, + { + "bbox": [ + 304, + 182, + 546, + 312 + ], + "type": "inline_equation", + "content": "58\\%" + }, + { + "bbox": [ + 304, + 182, + 546, + 312 + ], + "type": "text", + "content": " on PartNet. NGSP parses region label proposals generated by a guide neural network using likelihood modules to evaluate the global coherence of each proposal. It is still designed as a fully automatic method, but the authors did suggest the need for human-in-the-loop approaches to \"scale beyond carefully-curated research datasets to 'in-the-wild' scenarios.\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 315, + 546, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 546, + 458 + ], + "type": "text", + "content": "Hierarchical analysis: A fine-grained shape understanding is also attainable via hierarchical structural analysis. Some early works perform rule-based reasoning to construct symmetry hierarchies [38] and shape grammars [17], while others work with geometric priors, e.g., convexity [2], for hierarchical shape approximation. Similar to fine-grained segmentation, most learned hierarchical models are also unsupervised or weakly supervised, including co-hierarchical analysis [32], hierarchical abstraction [30], recursive implicit fields [21], and inverse constructed solid geometry (CSG) [45, 14]. All of these methods aim for full automation and work well only with shallow hierarchies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 579 + ], + "type": "text", + "content": "Active learning for part labeling: The classical and core question for active learning [1, 47, 26] is: \"How does one select instances from the underlying data to label, so as to achieve the most effective training for a given level of effort?\" Past works have studied uncertainty and diversity for interactive segmentation and labeling of 3D shapes [35, 5], medical images [31], and large-scale point clouds [40], to name a few. We utilize active learning, aiming to approach full labeling accuracy, regardless whether the test shapes belong to a \"research dataset\" or are \"in-the-wild.\"" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "Most closely related to our work is the active framework for region annotations by Yi et al. [44]. Both works share the common goal of producing accurate, human-verified labels. On the other hand, their work imposes a fixed human effort budget, while ours does not. Technically, their work relies on a conditional random field to automatically propagate human labels, while our work integrates a DL module [37] into an active part labeling tool. The only prior work that combines deep and active learning for 3D segmentation [5] focuses on boundary refinement, just as the earlier work [35] whose learning module is constrained " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "867" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": "means clustering. However, none of these tools were designed to work with fine-grained labels, e.g., they still only operate on four part labels for chairs [5, 35, 44, 41]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 204 + ], + "type": "text", + "content": "Finally, while our work is the first active learning tool for fine-grained 3D part labeling, hierarchical active learning has been employed in other data domains such as documents. In hierarchical text classification[16, 4], text labels can also be organized hierarchically. In comparison, semantic labeling of 3D parts in a shape collection poses various challenges including richer structural variations, part relations, and more costly and delicate user interactions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 213, + 102, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 102, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 102, + 226 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "text", + "content": "Given a shape test set " + }, + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "text", + "content": ", where each shape is presegmented into parts, our methods assign a label " + }, + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "inline_equation", + "content": "l \\in L" + }, + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "text", + "content": " to each part of each shape. The set of labels, " + }, + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 234, + 286, + 329 + ], + "type": "text", + "content": ", is predefined and organized in a tree structure such that coarse labels are at the top, and fine-grained labels at the bottom. The labeling progresses hierarchically top-to-bottom, assigning first coarse labels and refining them at each step, achieving high-accuracy fine-grained labeling results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 330, + 286, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 286, + 377 + ], + "type": "text", + "content": "At each node, we use an active learning based tool to iteratively engage human interactions into the part labeling process. With this tool, we obtain a close to error-free part labeling on each node on any test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 384, + 209, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 209, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 209, + 397 + ], + "type": "text", + "content": "3.1. Hierarchical labeling strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 403, + 286, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 286, + 510 + ], + "type": "text", + "content": "We organize the part labels into a tree structure based on their semantics, as in PartNet [20]. Following the same terminology, two types of internal tree nodes are defined: an \"AND\" node indicates \"what part(s) it has\" (e.g., a regular base has legs and a runner), while an \"OR\" node indicates \"what type it is\" (e.g., a star- or pedestal-type base). The labeling on \"AND\" nodes is based on part features, while the labeling on \"OR\" nodes depends on group features; see Fig. 3. Each leaf node presents a final part label." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": "At each node of the tree, we combine a deep neural network with an active learning tool to perform part labeling. Consider a single shape " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "s_i \\in S" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": " and its set of parts " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": ". At the root node, the input are all the parts of shape " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": ", that is " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "P_{root} = P_i" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": ". After finishing the labeling task at the root node, the set of parts is split into one or more subsets " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "P_l \\subseteq P_{root}" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": " corresponding to the children labels " + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 510, + 286, + 641 + ], + "type": "text", + "content": ". Then, each subset is input to the respective node for further labeling. By splitting the input set at each node, the size of the input decreases as the depth of the node increases. Finally, the labeling stops when we reach a leaf node." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "The complexity of our labeling scheme depends on the structure of the tree. On one extreme, if all labels are direct children of the root node, training at the root would be more complex but there is only one node to train. The other extreme is a binary tree, where each node is trained to distinguish between only two labels, but we have to train more" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "nodes, resulting in an increased training time. As we have found the original PartNet trees to have too many internal nodes, we modified this initial tree structure using a heuristic rule: we remove internal \"AND\" nodes which contain less than three children and keep all \"OR\" nodes. This way, our tree structure is better balanced for our task to lead to reduced training time. Examples of our modified tree structures are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 176, + 457, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 176, + 457, + 189 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 457, + 189 + ], + "type": "text", + "content": "3.2. Symmetry-assisted labeling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 195, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 303 + ], + "type": "text", + "content": "Another novel feature to accelerate HAL3D makes use of detectable part symmetries on the test set. Symmetry is ubiquitous in human-made objects and it is common practice for 3D artists to copy-and-paste symmetric or repeated parts when building 3D assets. Symmetry detection has been a well-studied subject in geometry processing [19] and there are a variety of applicable techniques. For simplicity, we simply use sizes of the part oriented bounding boxes as a crude and conservative feature to detect part symmetries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "type": "text", + "content": "As symmetric parts are expected to share the same label, we accelerate our active learning in two ways: if a proposal assigns different labels to symmetric parts, we automatically remove it from the \"HC (High confidence)\" set. Then, in the modification module, the user only needs to label one part of the symmetry group and our system automatically propagate it to the other parts. These features improve labeling efficiency by " + }, + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "type": "text", + "content": "; see Section 5.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 407, + 512, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 512, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 512, + 420 + ], + "type": "text", + "content": "3.3. Active labeling of fine-grained 3D parts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "Fine-grained 3D shape segmentation and annotation is challenging because of the wide structural diversity found in small and intricate parts of each shape. This diversity makes it difficult for existing DL methods to achieve high prediction accuracy. On the other side, active learning techniques are commonly used by researchers to improve the performance of the learned classifier with the help of human efforts. We adopt these techniques and, at each node of our hierarchical labeling strategy, we use a deep active learning based method to achieve high part labeling accuracy. Our method comprises three main modules: the DL-based part label proposal network, the part label verification module, and the part label modification module." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Consider a node " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " in our tree structure, given the subset of parts labeled " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " of each shape, we iteratively update the part labels until they all pass human verification. At each iteration, we first use a DL-based neural network to make part label proposals based on part label probabilities. Then, the HC proposals are passed to the part label verification module for human verification. Next, we collect the LC, i.e., low-confidence, proposals, which are more likely to be inaccurate, and pass them to the label modification module. Finally, the verified proposals and modified proposals are combined and feed back to the proposal network to fine-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "868" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 286, + 225 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 286, + 225 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 286, + 225 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 286, + 225 + ], + "type": "image", + "image_path": "81dc34a885c63960c9cb62dd7f78eea27acc172fe3218bed86ce55062b35a313.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 227, + 286, + 331 + ], + "blocks": [ + { + "bbox": [ + 51, + 227, + 286, + 331 + ], + "lines": [ + { + "bbox": [ + 51, + 227, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 51, + 227, + 286, + 331 + ], + "type": "image", + "image_path": "a138e88054d68308ee3e38c29f851fe793b2764fee969279cee1b6e4a0be5f2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "text", + "content": "Figure 3. Label proposal architectures. Part label prediction is based on global and local, part-level features at the \"AND\" nodes and global features only for the \"OR\" nodes. The \"AND\" nodes would serve to predict the label for each part and the \"OR\" nodes would predict labels for the entire input shape. Notion-wise, " + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "text", + "content": " is the number of points, " + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "text", + "content": " is the number of parts, " + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "text", + "content": " is the feature dimension, and " + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 334, + 288, + 411 + ], + "type": "text", + "content": " is the number of labels." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 430, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 467 + ], + "type": "text", + "content": "tune the network weights. The active learning iteration will stop if the number of left shapes are less than 40. We explain next how we make HC vs. LC proposals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 593 + ], + "type": "text", + "content": "Label proposal network. Given a set of parts from each shape " + }, + { + "bbox": [ + 46, + 474, + 287, + 593 + ], + "type": "inline_equation", + "content": "s_i \\in S" + }, + { + "bbox": [ + 46, + 474, + 287, + 593 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 46, + 474, + 287, + 593 + ], + "type": "inline_equation", + "content": "N = 8,192" + }, + { + "bbox": [ + 46, + 474, + 287, + 593 + ], + "type": "text", + "content": " points over the part surfaces and input them into our proposal network. As shown in Fig. 3, our network uses the edge convolution (EdgeConv) module from DGCNN [37] to extract point features from the point clouds. This module consists of five EdgeConv layers whose feature dimensions are 64, 64, 128, 256, and 256, respectively. The point features from all layers are concatenated and passed into a convolution layer. The size of the resulting feature is 512." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Since our tree structure has different types of nodes, we apply two different network structures from DGCNN, where the architecture for \"OR\" nodes is the same as in the original DGCNN classification network. At an \"AND\" node, differently from DGCNN which is the SOTA method to predict point labels based on point features, we use extra max-pooling operations to aggregate the local part features and global shape features, as inspired by NGSP [10]. The concatenated features are fed into the part classifier, which contains 3 MLP layers to predict part labels. As for an" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "\"OR\" node, labels of the input shape are predicted based on the global features only; see Fig. 3. Since different shapes have different part counts, we set the maximum part count " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "P_{max} = 150" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " during training for batch processing and add zero-padding to the part feature matrix so that all the part feature matrices have the same dimension. More details of our network can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 165, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 546, + 392 + ], + "type": "text", + "content": "Label verification. Our goal is to get the labels of each shape verified by a human. To reduce verification and training times, we need to decide which shapes are verified first and how many to verify on each iteration. Shapes in the test set that have similar structure to shapes in the train set are more likely to be labeled well. We can quickly select these well-labeled shapes from the test set and use them to enhance our limited train set, improving the generalization ability of our proposal network. In our method, we first sort the test set based on average predicted part label probabilities from the proposal network of each shape and arrange the test set into batches, each containing " + }, + { + "bbox": [ + 304, + 165, + 546, + 392 + ], + "type": "inline_equation", + "content": "B = 10" + }, + { + "bbox": [ + 304, + 165, + 546, + 392 + ], + "type": "text", + "content": " shapes. Users can mark a shape as well-labeled if they see all parts are labeled correctly, or bad-labeled when one part is wrong. We design a user-friendly interface to achieve this, where users can view and rotate a batch of 3D shapes at the same time. The interface also shows the corresponding label colors for reference. Please see supplementary material for more details about our part label verification interface." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 393, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 536 + ], + "type": "text", + "content": "Given the sorted shapes, we need to decide how many of them to show the user as HC proposals before fine-tuning the prediction, to achieve a balance between processing efficiency and utilization of user feedback. To this end, we adopt an adaptive strategy: given the sorted proposals, we stop the verification step when the number of human-verified shapes on a batch is below a threshold, which is set to 4 in our experiments. This simple strategy is effective since early batches tend to have higher labeling accuracy than later batches. This way, the number of high confidence proposals presented to the user would adapt to the classification purity of the current test set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "text", + "content": "Label modification. Shapes in the test set whose part structures and geometries differ significantly from those in the train set are less likely to be labeled accurately by the network. These shapes can be regarded as outliers in the label regression task, but they can be especially valuable during fine-tuning. Human intervention is inevitable in this situation to correct the labels. Specifically, the shapes provided to the label modification module are collected from two sets: the " + }, + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "inline_equation", + "content": "Q_{1} = 20" + }, + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "text", + "content": " LC proposals with the lowest mean label probabilities, and those shapes that failed user verification more than " + }, + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "inline_equation", + "content": "H = 2" + }, + { + "bbox": [ + 304, + 545, + 546, + 689 + ], + "type": "text", + "content": " times. We ask users to modify the incorrect labels using our modification interface." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In our part annotation interface, labels are organized hierarchically so users can quickly look up the correct labels." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "869" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 160 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 160 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 160 + ], + "type": "image", + "image_path": "d90bc4b873ecf52dd1365638070b576a3653beac36e1c54d3d7a471901c3fdce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 163, + 289, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 289, + 275 + ], + "type": "text", + "content": "Figure 4. Comparison of pre-segmented parts in PartNet and ABO. ABO shapes were pre-segmented into connected components where each component may correspond to multiple labels, e.g., the back and the arm of the chair are grouped into the same component, as indicated by the red color in the second column. Instead, our part labeling operates on the convex pieces obtained by a decomposition. In the first three figures, the colors are used to distinguish between different parts, and we can see that the parts in PartNet are manually segmented and exhibiting imperfect boundaries. In the last figure, the colors reflect semantic labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 295, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 331 + ], + "type": "text", + "content": "The label-to-color mapping table is also provided to help find the incorrect labels. More details about this interface are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": "Shapes that have been verified and modified by humans are added into the initial training set together to fine-tune our proposal network. On the other hand, we skip shapes excluded from the HC and LC proposals. These shapes are combined with shapes that failed to pass the verification less than " + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": " times to go to the next iteration for labeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 416, + 170, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 170, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 170, + 430 + ], + "type": "text", + "content": "4. Datasets and Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": "Datasets. The Stanford PartNet [20] dataset is a common choice for fine-grained shape analysis [34, 10, 28]. We follow NGSP [10] and use the four major categories from PartNet for evaluation: chair, table, lamp, and storage furniture. Fine-grained labels in PartNet are organized into three granularity levels, and we adopt the second level in our experiments as in NGSP [10]. We first filter out noisy shapes with erroneous labels. Then from each category, we take 500 shapes and split them into 50/50/400 as train/val/test sets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "We also evaluate HAL3D on the recent Amazon Berkeley Objects (ABO) dataset, which is composed of about 8,000 high-quality 3D shapes created by artists from real product images, where each 3D model is a combination of build-aware connected components. Unlike the 3D shapes from PartNet in which each part corresponds to one single semantic label, a build-aware component from ABO may be appropriately associated with multiple labels, as shown in Fig. 4. To alleviate this issue, we apply an approximate convex decomposition [39] to further decompose each connected component into a fine-grained, i.e., over-segmented, set of convexes, where each convex piece is deemed to receive one and only one part label. Our active labeling is then applied to the set of convexes. When running the con" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "vex decomposition algorithm, we use the default parameters from the provided code, except for the concavity threshold setting, which is set as 0.2 in our experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": "As GT labels are unavailable in ABO, three artists were hired to create them for the decomposed convex components. We built a test set containing 1,800 shapes from five categories: chair, table, lamp, storage furniture, and bed, with 200 for storage and 400 each for the other categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 175, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 258 + ], + "type": "text", + "content": "Metrics. Active learning systems are typically evaluated from two perspectives: labeling accuracy and the amount of human efforts or involvements. We employ two standard metrics to assess the labeling quality: (1) part label accuracy measuring how many parts are correctly labeled and (2) point cloud segmentation mIoU depicting how much of the shape surface area is correctly labeled." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 355 + ], + "type": "text", + "content": "The amount of human efforts is measured by the consumed time in the steps of label verification and modification. Inspired from [44], we firstly give artists the same training courses and tools, and conduct user studies on Part-Net. We report the total recorded labeling time in Tab. 1. We also collect the operations they used in the user study and model the total human interaction time as a function of the number of processed parts and shapes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 363, + 545, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 545, + 388 + ], + "type": "interline_equation", + "content": "t = T _ {v} ^ {p} \\left(P _ {v}, L\\right) + T _ {v} ^ {f} \\left(S _ {v}, L\\right) + T _ {v} ^ {p} \\left(P _ {m} ^ {c}, L\\right) + T _ {m} ^ {p} \\left(P _ {m} ^ {w}, L\\right), \\tag {1}", + "image_path": "7f574c84d192129503f6847384a4f67cdd6ac278636775f931d2e9184f96a81e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "P_{v}" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": " is the number of correctly labeled parts in verification, " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "S_{v}" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": " is the number of incorrectly shapes in verification, " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "P_{m}^{c}" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": " is the number of correctly labeled parts in modification, " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "P_{m}^{w}" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": " is the number of incorrectly labeled parts in modification, and finally " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": " is the total number of labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 519 + ], + "type": "text", + "content": "Function (1) is devised based on several observations: a) users need to go through and verify all the part labels to ensure accuracy; b) they can easily pick up failure cases during verification if any part is incorrectly labeled, without checking other parts; and c) users need to check all the parts in each shape to know whether modification is needed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 520, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 556 + ], + "type": "text", + "content": "The formulas and their coefficients in (2) are obtained via linear regression using the total operation and labeling times collected from all annotators in our user study:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 564, + 503, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 564, + 503, + 578 + ], + "spans": [ + { + "bbox": [ + 348, + 564, + 503, + 578 + ], + "type": "interline_equation", + "content": "T _ {v} ^ {p} \\left(P _ {v}, L\\right) = \\left(0. 3 1 + 0. 0 3 * L\\right) * P _ {v},", + "image_path": "bf4faddda521fd4bb79b76b53d3ee592a20ece1ac956f06107aed69a06aec10f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 348, + 581, + 545, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 581, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 348, + 581, + 545, + 600 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} T _ {v} ^ {f} \\left(S _ {v}, L\\right) = (1. 4 7 + 0. 0 5 * L) * S _ {v}, \\\\ \\text {T r} \\left(R _ {v} ^ {f}, L\\right) = \\left(0. 0 4 - 0. 0 2, L\\right) * R _ {v} ^ {f} \\end{array} \\tag {2}", + "image_path": "9a778cea99bd78eb814e02f2db7ccc97adc700aedfddaba608b104074a40903e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 347, + 597, + 505, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 347, + 597, + 505, + 610 + ], + "type": "interline_equation", + "content": "T _ {v} ^ {p} \\left(P _ {m} ^ {c}, L\\right) = \\left(0. 3 1 + 0. 0 3 * L\\right) * P _ {m} ^ {c},", + "image_path": "b9b1c7589e5b8c7744801044c139387d04e5b2a9a1a65eb764ee7d1c639dc796.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 612, + 505, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 345, + 612, + 505, + 625 + ], + "type": "interline_equation", + "content": "T _ {m} ^ {p} \\left(P _ {m} ^ {w}, L\\right) = (5. 2 8 + 0. 2 3 * L) * P _ {m} ^ {w},", + "image_path": "edb2b2bbe232b8bd51770df60be35fa15df119906e4412e75b0d7455d1ca43c1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{v}^{p}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the checking time for correctly labeled parts, " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{v}^{f}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the verification time for incorrectly labeled shapes, and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{m}^{p}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the time for part label modification. We use these functions to evaluate the human interaction timing, in seconds, for hyper-parameters analysis on PartNet. We also report additional machine time cost from training and fine-tuning in the supplementary material." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "870" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 212 + ], + "type": "text", + "content": "We have implemented our approach in PyTorch. The label proposal network at each node is pre-trained with the Adam optimizer on the training set of PartNet for 250 epochs, where the learning rate is 0.001 and it is decreased by 0.8 every 25 epochs. When fine-tuning the network in HAL3D, the learning rate is fixed to 0.0001, the maximum epoch number is 125, and the validation set is used for early stopping. Our experiments were ran on an NVIDIA Tesla V100 GPU, and the training time varies from 20 minutes to 3 hours according to the number of parts at each node." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 220, + 164, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 164, + 234 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 164, + 234 + ], + "type": "text", + "content": "5.1. Competing methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 240, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 287, + 264 + ], + "type": "text", + "content": "We compare HAL3D, our hierarchical active labeling tool, with two automatic part labeling methods," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 267, + 287, + 354 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 267, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 267, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 287, + 304 + ], + "type": "text", + "content": "- PartNet [20], which employs the PointNet++ [25] backbone for point-wise feature extraction and labels parts by a voting scheme over points belonging to the same part." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "text", + "content": "- NGSP [10], which aggregates point-wise features via PointNet++ [25] to obtain part features and then leverages a likelihood-aware beam search for optimization; this represents the state of the art for fine-grained labeling." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 357, + 233, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 357, + 233, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 357, + 233, + 369 + ], + "type": "text", + "content": "And two variants of our labeling approach," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 373, + 287, + 436 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 47, + 373, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 408 + ], + "type": "text", + "content": "- Our prop. is one such variant without using the manual label verification or modification module; it predicts part labels using the label proposal network only." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 411, + 286, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 286, + 436 + ], + "type": "text", + "content": "- DAL3D is a deep active learning variant of our approach without using the hierarchical labeling strategy." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 444, + 175, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 175, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 175, + 456 + ], + "type": "text", + "content": "5.2. Evaluation on PartNet" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 463, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 287, + 487 + ], + "type": "text", + "content": "All methods presented in this section are pre-trained and tested on the same training and test sets of PartNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 498, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 640 + ], + "type": "text", + "content": "Quantitative comparison. Tab. 1 compares the three automatic methods, where PartNet is inferior to both NGSP and our label proposal network due to its hand-crafted voting mechanism. Our simple network achieves comparable accuracy compared to NGSP (difference by 0.11 and 3.49 with respect to mIoU and part accuracy, respectively). However, NGSP suffers from significantly longer running time due to the intensive beam search; see the supplementary for a running time comparison. In contrast, our proposal network is effective and efficient, i.e., capable of quickly producing accurate initial labels to reduce the manual workload in verification and modification for active learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Tab. 1 also shows that even with the sophisticated design in NGSP, the highest accuracy attained is only " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ". Hence the performance of automatic fine-grained labeling still falls far below that of active learning, where the largest and smallest accuracy gap is " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "40.77\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "22.49\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " between NGSP and HAL3D, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 68, + 545, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 68, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 545, + 146 + ], + "type": "text", + "content": "Table 1. Quantitative comparison against the competing labeling methods and variants. In the table, \"AL\" indicates whether a method uses active learning. Since PartNet, NGSP, and Our prop. are automatic methods, their entries are marked with a '-' instead. \"Accu.\" denotes the part label accuracy, \"mIoU\" denotes the mean Intersection-over-Union, and \"Lab. time\" denotes the labeling time (hours in total) for the active learning methods." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 306, + 155, + 545, + 462 + ], + "blocks": [ + { + "bbox": [ + 306, + 155, + 545, + 462 + ], + "lines": [ + { + "bbox": [ + 306, + 155, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 155, + 545, + 462 + ], + "type": "table", + "html": "
MethodALAccu.↑mIoU↑Lab. time↓(h)
Chair
PartNet [20]-54.9031.04-
NGSP [10]-68.9842.90-
Our prop.-67.4541.34-
DAL3D✓89.8472.305.99
HAL3D✓94.1384.924.34
Table
PartNet [20]-33.2114.92-
NGSP [10]-47.1127.94-
Our prop.-42.7723.50-
DAL3D✓78.5369.829.14
HAL3D✓84.1874.654.41
Storage furniture
PartNet [20]-55.8230.86-
NGSP [10]-70.6045.79-
Our prop.-69.8242.43-
DAL3D✓84.5771.994.81
HAL3D✓93.0988.723.87
Lamp
PartNet [20]-38.2811.02-
NGSP [10]-49.7623.49-
Our prop.-44.7719.27-
DAL3D✓83.1469.354.52
HAL3D✓90.5379.512.16
", + "image_path": "a32a499f1bf9c6e3afc0d0603f59b3d1b65051028652e762d48908fc49cab5d6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 307, + 502, + 555, + 589 + ], + "blocks": [ + { + "bbox": [ + 306, + 482, + 538, + 494 + ], + "lines": [ + { + "bbox": [ + 306, + 482, + 538, + 494 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 538, + 494 + ], + "type": "text", + "content": "Table 2. Ablation on key modules of HAL3D for chair category." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 502, + 555, + 589 + ], + "lines": [ + { + "bbox": [ + 307, + 502, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 555, + 589 + ], + "type": "table", + "html": "
Row IDProp.Hier.Sym.ALLab-T↓Accu↑
2nd----22.0589.16
3rd✓---8.6588.53
4th✓✓✓-6.3793.87
5th✓-✓✓5.9989.84
6th✓✓-✓5.2193.45
7th✓✓✓✓4.3494.13
", + "image_path": "3d22fb27ed6f71c25aa8df62f3e5419ef17b7423b225cc67a68efe6901027271.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 595, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 656 + ], + "type": "text", + "content": "HAL3D consistently outperforms our DAL3D variant in all metrics, particularly 2.42 hours faster on average on labeling time, demonstrating the importance of our hierarchical labeling order for reducing manual workload. We also provide qualitative comparison in Fig. 5." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Ablation study. Tab. 2 verifies the contributions of the main components of HAL3D on minimizing human labeling efforts, a key criterion for success for a human-in-the-loop approach. The second to fifth columns indicate:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "871" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 76, + 282, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 76, + 282, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 76, + 282, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 76, + 282, + 266 + ], + "type": "image", + "image_path": "5c1e75107ec5e6391a49c469ec5e33050be31affb8228fbafb28a8d366c373d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 288, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 288, + 316 + ], + "type": "text", + "content": "Figure 5. Visual comparison between labeling methods. HAL3D can achieve results approaching GT, barring human errors. The only mis-labeled part by HAL3D here is the bottom panel of the table (red oval), which could be easily labeled as shelf by users." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 328, + 286, + 407 + ], + "blocks": [ + { + "bbox": [ + 50, + 328, + 286, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 328, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 328, + 286, + 407 + ], + "type": "image", + "image_path": "8c544f81bcf3b438fbe3703bd64fb85931533191054a39af0e796aa615ea1157.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 409, + 288, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 409, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 288, + 475 + ], + "type": "text", + "content": "Figure 6. A plot of the number of human-verified shapes (max = 400) at different nodes (only those which existed in more than 100 shapes) of the chair category. The plot shows that with hierarchical active learning, many more shapes pass the verification, hence saving on the costly label modification, especially at higher hierarchy levels (i.e., earlier in the labeling iterations)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 497, + 253, + 554 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 497, + 231, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 231, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 231, + 509 + ], + "type": "text", + "content": "- Prop. if the label proposal network is used;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 512, + 253, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 253, + 524 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 253, + 524 + ], + "type": "text", + "content": "- Hier. if the hierarchical labeling strategy is used;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 526, + 216, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 216, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 216, + 539 + ], + "type": "text", + "content": "- Sym. if the symmetry checking is used." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 542, + 190, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 190, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 190, + 554 + ], + "type": "text", + "content": "- AL. if the active learning is used." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 557, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 675 + ], + "type": "text", + "content": "We compare our full approach against five baselines. The second row is a baseline that directly predicts labels by annotating with our label modification interface. The third row is a baseline that firstly uses the label proposal network and then manually modifies incorrect part labels. The fourth, fifth, and sixth rows represent variants of HAL3D without hierarchical labeling, symmetry checking, and active learning, respectively. The results in Tab. 2 clearly justify the design choices in our method, where our full approach (last row) is the most efficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "To further evaluate the effectiveness of our hierarchical labeling, we plot in Fig. 6 the number of shapes which passed human verification over the active learning qua" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 325, + 91, + 526, + 129 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "type": "text", + "content": "Table 3. Inference time comparison between NGSP and our label proposal network. We report average seconds per shape." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 91, + 526, + 129 + ], + "lines": [ + { + "bbox": [ + 325, + 91, + 526, + 129 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 526, + 129 + ], + "type": "table", + "html": "
CategoryChairTableLampStorage
NGSP5.41018.0112.6912.87
Our prop.0.0120.0140.0120.011
", + "image_path": "58effd788152898314806f0b9c13f3b1e751ba971bdadd244fd49089d906273c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 252 + ], + "type": "text", + "content": "tions. As we can see, HAL3D consistently verified more shapes as correct at all tree nodes when comparing to the baseline that had no hierarchical design, e.g., more than 300 shapes were marked as correct in the first iteration in the chair_base node. Essentially, our hierarchical verification splits the labeling task into smaller and easier tasks compared to the case of non-hierarchical design, which effectively reduces the burden arising from the subsequent and more costly modification step in terms of human effort." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 253, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 312 + ], + "type": "text", + "content": "In the supplementary material, we perform ablations on the four hyper parameters and thresholds employed in our method. In all the experiments, we fix these parameters and set them based only on a small dataset. A more careful parameter tuning should lead to better results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 328, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 424 + ], + "type": "text", + "content": "Inference time comparison. Although NGSP achieved higher prediction accuracy than our proposal network, as shown in Tab. 1, our network is more efficient and better suited to be incorporated into an active learning framework. Tab. 3 compares the inference times between NGSP and our proposal network. NGSP is highly time-consuming due to the use of beam search, whereas our label proposal network takes only 12ms per shape on average." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 431, + 421, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 421, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 421, + 443 + ], + "type": "text", + "content": "5.3. Evaluation on ABO" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": "We apply HAL3D to label our constructed test set for ABO, where the label proposal network is pre-trained on the training set of PartNet. As mentioned in Sec. 4, shapes in ABO are decomposed into convex shapes and HAL3D assigns a semantic label to each decomposed piece. Three artists were hired to use our HAL3D tool and interface for evaluation. Since GT labels are not provided in ABO, we inspect the labeling results from the artists via human cross validation with the final corrected labeling forming the GT. As different artists may report different accuracy and recorded labeling times, we report averages in the paper. The artists worked together with an expert to determine labels for uncertain shapes. The visual results from HAL3D, as shown in Fig. 7 and the supplemental material, were obtained on the test set by taking majority voting from labels provided by different artists." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "As shown by quantitative results in Tab. 4, HAL3D can accurately label all categories with a " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " average accuracy using 3-7 hours of manual annotation, demonstrating the effectiveness and efficiency of our method. The reduction on human efforts against manual annotation becomes more significant when comparing against the last four rows in Tab. 4," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "872" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 74, + 116, + 355 + ], + "blocks": [ + { + "bbox": [ + 55, + 74, + 116, + 355 + ], + "lines": [ + { + "bbox": [ + 55, + 74, + 116, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 74, + 116, + 355 + ], + "type": "image", + "image_path": "3b49e1a85c46503b978f5a0c239aa3ccae1f14c0aabbb80d78f39b58a0392ece.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 76, + 199, + 354 + ], + "blocks": [ + { + "bbox": [ + 143, + 76, + 199, + 354 + ], + "lines": [ + { + "bbox": [ + 143, + 76, + 199, + 354 + ], + "spans": [ + { + "bbox": [ + 143, + 76, + 199, + 354 + ], + "type": "image", + "image_path": "a437c65889b9456358cb69672ebe761430f3c1570568fb031840fd4c643f9943.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 361, + 456, + 373 + ], + "lines": [ + { + "bbox": [ + 136, + 361, + 456, + 373 + ], + "spans": [ + { + "bbox": [ + 136, + 361, + 456, + 373 + ], + "type": "text", + "content": "Figure 7. Visual comparison between several part labeling methods on the ABO dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 225, + 76, + 282, + 355 + ], + "blocks": [ + { + "bbox": [ + 225, + 76, + 282, + 355 + ], + "lines": [ + { + "bbox": [ + 225, + 76, + 282, + 355 + ], + "spans": [ + { + "bbox": [ + 225, + 76, + 282, + 355 + ], + "type": "image", + "image_path": "3ac5519016de7b6d167d5eed1e5cb3557ea73f3c151f69553067a808903bce2a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 367, + 355 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 367, + 355 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 367, + 355 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 367, + 355 + ], + "type": "image", + "image_path": "5f331945387e14b6e310897453f265c0cf60d137af862fc3b00a085a9fa5caca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 75, + 446, + 355 + ], + "blocks": [ + { + "bbox": [ + 389, + 75, + 446, + 355 + ], + "lines": [ + { + "bbox": [ + 389, + 75, + 446, + 355 + ], + "spans": [ + { + "bbox": [ + 389, + 75, + 446, + 355 + ], + "type": "image", + "image_path": "41cb17aebc1aba0b33304d3aa25e85e3c488feb7bc23a7b208c823d027dd8b5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 472, + 75, + 527, + 355 + ], + "blocks": [ + { + "bbox": [ + 472, + 75, + 527, + 355 + ], + "lines": [ + { + "bbox": [ + 472, + 75, + 527, + 355 + ], + "spans": [ + { + "bbox": [ + 472, + 75, + 527, + 355 + ], + "type": "image", + "image_path": "a509111814633060d3a97fd8c3017776107e5f0acf8a1d9feb6025e609fd8e60.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 52, + 410, + 282, + 548 + ], + "blocks": [ + { + "bbox": [ + 47, + 388, + 286, + 410 + ], + "lines": [ + { + "bbox": [ + 47, + 388, + 286, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 286, + 410 + ], + "type": "text", + "content": "Table 4. Comparison for different methods to pass human verification on test set of ABO. The time is in hours." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 410, + 282, + 548 + ], + "lines": [ + { + "bbox": [ + 52, + 410, + 282, + 548 + ], + "spans": [ + { + "bbox": [ + 52, + 410, + 282, + 548 + ], + "type": "table", + "html": "
ChairTableLampCabinetBed
# shapes400400400200400
HAL3D
Time ↓Accu. ↑6.4693.077.4192.554.3896.293.3494.484.5193.71
PartNet + modification
Time ↓Accu. ↑28.2588.9633.9486.5131.5390.9113.3590.2425.0289.37
NGSP + modification
Time ↓Accu. ↑28.3088.7734.3286.4031.5691.3713.2290.5624.9089.93
", + "image_path": "a925b174c12dcbe1618e9e04ba30a8d06bb859a9f3361ea04aa572be2af4f3cd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 562, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 598 + ], + "type": "text", + "content": "where we use the same manual modification interface to directly correct the PartNet and NGSP results. Please see supplementary material for results on ABO." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 609, + 120, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 120, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 120, + 621 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "We present HAL3D, the first semantic labeling tool that is designed to operate on fine-grained 3D parts and achieve close to error-free labeling through full verification by humans, barring human misjudgement. Our human-in-the-loop approach is based on active learning, combining deep label prediction and human inputs to iteratively fine-tune and improve the network prediction. Extensive experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 392, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 416 + ], + "type": "text", + "content": "demonstrate that human efforts are effectively reduced via hierarchical and symmetry-aware active labeling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 418, + 546, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 550 + ], + "type": "text", + "content": "HAL3D is currently implemented with DGCNN [37] as the label prediction backbone, but it can be readily replaced by PointNet [24], which was adopted by PartNet [20] and NGSP [10], or a more performant alternative such as Point Transformer [48]. Aside from symmetry, inter-shape part correspondences can also be utilized to reduce labeling costs, assuming that they can be reliably obtained. In addition, our tool does not yet allow splitting of under-segmented parts. Since performing such operations by humans is time-consuming, incorporating auto-splitting schemes into HAL3D deserves future consideration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 625 + ], + "type": "text", + "content": "We regard our work as only a preliminary step towards deep active learning for fine-grained structural analysis of 3D shapes. A natural focus for future work is to develop a more sophisticated ranking or selection mechanism for label verification and modification to achieve the most effective training for a given level of human effort." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 643, + 404, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 404, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 404, + 657 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "We thank all the anonymous reviewers and area chairs for their constructive comments and the Amazon artists for performing the human annotations. Thanks also go to Jeff-Aram Van Gorp for helpful discussions and other support." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "873" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Charu C. Aggarwal, Xiangnan Kong, Quanquan Gu, Jiawei Han, and Philip S. Yu. Active learning: A survey. In Data Classification: Algorithms and Applications, pages 571-597. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "text", + "content": "[2] Marco Attene, Michela Mortara, Michela Spagnuolo, and Bianca Falcidieno. Hierarchical convex approximation of 3D shapes for fast region selection. Comput. Graph. Forum, 27(5):1323-1332, 2008. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 213 + ], + "type": "text", + "content": "[3] Zhiqin Chen, Kangxue Yin, Matt Fisher, Siddhartha Chaudhuri, and Hao Zhang. BAE-NET: Branched autoencoder for shape co-segmentation. In ICCV, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "type": "text", + "content": "[4] Yu Cheng, Kunpeng Zhang, Yusheng Xie, Ankit Agrawal, and Alok Choudhary. On active learning in hierarchical classification. In Proceedings of the 21st ACM International Conference on Information and knowledge Management, pages 2467-2470, 2012. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "text", + "content": "[5] David George, Xianghua Xie, Yukun Lai, and Gary KL Tam. A deep learning driven active framework for segmentation of large 3d shape collections. Computer-Aided Design, 144:103179, 2022. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "text", + "content": "[6] Aleksey Golovinskiy and Thomas Funkhouser. Consistent segmentation of 3D models. In Proc. Shape Modeling International (SMI), 2009. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "type": "text", + "content": "[7] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: A network with an edge. ACM TOG, 38(4), 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "text", + "content": "[8] Yong He, Hongshan Yu, Xiaoyan Liu, Zhengeng Yang, Wei Sun, Yaonan Wang, Qiang Fu, Yanmei Zou, and Ajmal Mian. Deep learning based 3d segmentation: A survey. CoRR, abs/2103.05423, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "type": "text", + "content": "[9] Ruizhen Hu, Lubin Fan, and Ligang Liu. Co-segmentation of 3D shapes via subspace clustering. Comput. Graph. Forum, 31(5):1703--1713, 2012. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 287, + 501 + ], + "type": "text", + "content": "[10] R. Kenny Jones, Aalia Habib, Rana Hanocka, and Daniel Ritchie. The neurally-guided shape parser: Grammar-based labeling of 3D shape regions with approximate inference. In CVPR, 2022. 3, 5, 6, 7, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "text", + "content": "[11] R Kenny Jones, Aalia Habib, and Daniel Ritchie. Shred: 3d shape region decomposition with learned local operations. ACM Transactions on Graphics (TOG), 41(6):1-11, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[12] Oliver Van Kaick, Noa Fish, Yanir Kleiman, Shmuel Asafi, and Daniel Cohen-OR. Shape segmentation by approximate convexity analysis. ACM TOG, 34(1). 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "text", + "content": "[13] Evangelos Kalogerakis, Aaron Hertzmann, and Karan Singh. Learning 3D mesh segmentation and labeling. ACM TOG, 29(4), 2010. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[14] Kacper Kania, Maciej Zięba, and Tomasz Kajdanowicz. UCSG-NET - unsupervised discovering of constructive solid geometry tree. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[15] Yuki Kawana, Yusuke Mukuta, and Tatsuya Harada. Neural star domain as primitive representation. Advances in Neural Information Processing Systems, 33:7875-7886, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[16] Xiao Li, Da Kuang, and Charles X Ling. Active learning for hierarchical text classification. In Pacific-Asia conference on knowledge discovery and data mining, pages 14–25. Springer, 2012. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Tianqiang Liu, Siddhartha Chaudhuri, Vladimir G. Kim, Qixing Huang, Niloy J. Mitra, and Thomas Funkhouser. Creating consistent scene graphs using a probabilistic grammar. ACM TOG, 33, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[18] Niloy Mitra, Michael Wand, Hao Zhang, Daniel Cohen-Or, and Martin Bokeloh. Structure-aware shape processing. In Eurographics State-of-the-Art Report (STAR), 2013. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[19] Niloy J. Mitra, Mark Pauly, Michael Wand, and Duygu Ceylan. Symmetry in 3D geometry: Extraction and applications. 32(6), 2013. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[20] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In CVPR, June 2019. 2, 3, 4, 6, 7, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "text", + "content": "[21] Chengjie Niu, Manyi Li, Kai Xu, and Hao Zhang. RIM-Net: Recursive implicit fields for unsupervised learning of hierarchical shape structures. In CVPR, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[22] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 332, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 373 + ], + "type": "text", + "content": "[23] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, pages 10344—10353, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[24] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 3, 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[25] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "text", + "content": "[26] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B. Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 479, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 499 + ], + "type": "text", + "content": "[27] Ariel Shamir. A survey on mesh segmentation techniques. Comput. Graph. Forum, 27(6):1539-1556. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "type": "text", + "content": "[28] Gopal Sharma, Kangxue Yin, Subhransu Maji, Evangelos Kalogerakis, Or Litany, and Sanja Fidler. MvDeCor: Multiview dense correspondence learning for fine-grained 3D segmentation. In ECCV, 2022. 3, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "type": "text", + "content": "[29] Oana Sidi, Oliver van Kaick, Yanir Kleiman, Hao Zhang, and Daniel Cohen-Or. Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering. ACM TOG, 30(6):126:1-126:9, 2011. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 591, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 624 + ], + "type": "text", + "content": "[30] Chun-Yu Sun, Qian-Fang Zou, Xin Tong, and Yang Liu. Learning adaptive hierarchical cuboid abstractions of 3D shape collections. 38(6), 2019. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "text", + "content": "[31] Andrew Top, Ghassan Hamarneh, and Rafeef Abugharbieh. Active learning for interactive 3D image segmentation. In Int. Conf. on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2011. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[32] Oliver van Kaick, Kai Xu, Hao Zhang, Yanzhen Wang, Shuyang Sun, Ariel Shamir, and Daniel Cohen-Or. Cohierarchical analysis of shape structures. ACM TOG, 32(4), 2013. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "874" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[33] Xiaogang Wang, Xun Sun, Xinyu Cao, Kai Xu, and Bin Zhou. Learning fine-grained segmentation of 3D shapes without part labels. In CVPR, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[34] Xiaogang Wang, Bin Zhou, Haiyue Fang, Xiaowu Chen, Qinping Zhao, and Kai Xu. Learning to group and label fine-grained shape components. 37(6), 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[35] Yunhai Wang, Shmulik Asafi, Oliver van Kaick, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Active co-analysis of a set of shapes. ACM TOG, 31(6), 2012. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[36] Yunhai Wang, Minglun Gong, Tianhua Wang, Daniel Cohen-Or, Hao Zhang, and Baoquan Chen. Projective analysis for 3D shape segmentation. ACM TOG, 32(6), 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "text", + "content": "[37] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph CNN for learning on point clouds. ACM TOG, 2019. 2, 3, 5, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[38] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhiquan Cheng, and Yueshan Xiong. Symmetry hierarchy of man-made objects. Comput. Graph. Forum, 30(2):287-296, 2011. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 341 + ], + "type": "text", + "content": "[39] Xinyue Wei, Minghua Liu, Zhan Ling, and Hao Su. Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search. ACM TOG, 41(4), 2022. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 396 + ], + "type": "text", + "content": "[40] Tsung-Han Wu, Yueh-Cheng Liu1, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H. Hsu. ReDAL: Region-based and diversity-aware active learning for point cloud semantic segmentation. In ICCV, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "type": "text", + "content": "[41] Zizhao Wu, Ruyang Shou, Yunhai Wang, and Xinguo Liu. Interactive shape co-segmentation via label propagation. In Proc. of CAD/Graphics, 2014. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "type": "text", + "content": "[42] Kaizhi Yang and Xuejin Chen. Unsupervised learning for cuboid shape abstraction via joint segmentation from point clouds. ACM TOG, 40(4), 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[43] Li Yi, Leonidas Guibas, Aaron Hertzmann, Vladimir G. Kim, Hao Su, and Ersin Yumer. Learning hierarchical shape segmentation and labeling from online repositories. SIGGRAPH, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 512, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 565 + ], + "type": "text", + "content": "[44] Li Yi, Vladimir G. Kim, I.-Chao Shen Duygu Ceylan, Mengyuan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas J. Guibas. A scalable active framework for region annotation in 3D shape collections. ACM TOG, 35(6), 2016. 1, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[45] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. CAPRI-Net: Learning compact CAD shapes with adaptive primitive assembly. In CVPR, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "text", + "content": "[46] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. PartNet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[47] Xueying Zhan, Qingzhong Wang, Kuan-hao Huang, Haoyi Xiong, Dejing Dou, and Antoni B. Chan. A comparative survey of deep active learning, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "text", + "content": "[48] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In CVPR, 2021. 9" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 150 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[49] Yang Zhou, Kangxue Yin, Hui Huang, Hao Zhang, Minglun Gong, and Daniel Cohen-Or. Generalized cylinder decomposition. ACM TOG, 34(6), 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[50] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J. Guibas, and Hao Zhang. AdaCoSeg: Adaptive shape co-segmentation with group consistency loss. In CVPR, 2020. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "875" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_content_list.json b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..37bad2a2e553462dac5c824473191f89ea6b13fb --- /dev/null +++ b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_content_list.json @@ -0,0 +1,1819 @@ +[ + { + "type": "text", + "text": "HDG-ODE: A Hierarchical Continuous-Time Model for Human Pose Forecasting", + "text_level": 1, + "bbox": [ + 76, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yucheng Xing Xin Wang \nDepartment of Electrical and Computer Engineering, Stony Brook University \n{yucheng.xing, x.wang}@stonybrook.edu", + "bbox": [ + 178, + 181, + 790, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, human pose estimation has attracted more and more attention due to its importance in many real applications. Although many efforts have been put on extracting 2D poses from static images, there are still some severe problems to be solved. A critical one is occlusion, which is more obvious in multi-person scenarios and makes it even more difficult to recover the corresponding 3D poses. When we consider a sequence of images, the temporal correlation among the contexts can be utilized to help us ease the problem, but most of the current works only rely on discrete-time models and estimate the joint locations of all people within a whole sparse graph. In this paper, we propose a new framework, Hierarchical Dynamic Graph Ordinary Differential Equation (HDG-ODE), to tackle the 3D pose forecasting task from 2D skeleton representations in videos. Our framework adopts ODE, a continuous-time model, as the base to predict the 3D joint positions at any time. Considering the structural-property of the skeleton data in representing human poses and the possible irregularity caused by occlusion, we propose the use of dynamic graph convolution as the basic operator. To reduce the computational complexity introduced by the sparsity of the pose graph, our model takes a hierarchical structure where the encoding process at the observation timestamp is done in a cascade manner while the propagation between observations is conducted in parallel. The performance studies on several datasets demonstrate that our model is effective and can out-perform other methods with fewer parameters.", + "bbox": [ + 76, + 300, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the progress in the computer vision field, human-related contents have attracted more and more attention, as they are very important in many real-world applications. Some examples are autonomous driving, human-computer interaction and anomaly detection. Among these contents, human poses are the most basic and important ones since their accurate and timely estimation or even forecasting usually serve as the first step for further analyzing complex", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "activities.", + "bbox": [ + 500, + 270, + 566, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human pose estimation is not a novel topic but has been studied for many years. Its performance, however, is still not satisfactory enough to meet the demands of many practical applications. Since A. Toshev and C. Szegedy proposed DeepPose [69] to represent human poses by the locations of a set of joints from human skeletons and first time utilized the regression method of deep learning to estimate the corresponding 2D coordinates, a lot of efforts [78, 51, 87, 20, 86, 19, 79, 77, 54, 67, 95, 94, 3, 66, 74, 73] have been made to get the accurate estimation in the presence of occlusion resulted from the perspective limitation of images. These methods [14, 33, 55, 65, 34, 68, 38, 53, 16, 71, 58, 48, 99, 100, 84] are extended to multi-person scenarios, where the occlusion problem is more severe. Although some progress has been made, 2D locations cannot perfectly restore the human poses, and are never our final goals for this task. Recently, more and more works have been focusing on 3D pose estimation with raw images [15, 24, 6, 91] or extracted 2D poses [96, 21, 5, 18, 93, 39, 97, 83, 17, 43, 91] as input. Besides the inherent occlusion problems, the lack of depth in monocular images also limits the performance of algorithms. Although some schemes relying on multi-view images [50, 60, 94, 80, 72, 10, 70, 28, 40, 25, 11] or other sensors [54, 31, 27, 94] can ease this problem to some degree, we have to admit that paired images of different views or complementary sensors are not always available in most scenarios. The core problem is how to just utilize limited information in monocular images to extract the corresponding 3D poses of people. If a sequence of images (i.e. video) is available, the temporal context carried can be also helpful for the task [1, 81, 7, 47, 76]. They not only provide information to make up the missing joints caused by the occlusion and provide hints for joints' locations in the future, but also help group the detected joints belonging to a specific person in multi-person scenarios. So in this paper, we mainly focus on restoration from a sequence of monocular multi-person 2D skeletons to predict the corresponding 3D poses at a future time point.", + "bbox": [ + 496, + 290, + 895, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some related methods [64, 9, 75, 82, 41, 11, 98, 26, 13, 63, 36] usually applied graph neural networks to encode the", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14700", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2D human joints and propagated the information through recurrent neural networks or directly used spatial-temporal graph convolutional networks. However, almost all of them only focused on single-person scenarios, while in most of cases, the relations among people in the scene are also helpful for the forecasting. Also, there exist several limitations in these works. Firstly, traditional recurrent neural networks, such as GRU or LSTM, all use discrete-time sequential models, which have a fixed propagation step and are not flexible enough to deal with different moving speeds of different human joints. The performance of these models will be compromised when the video frequency is not high or the cameras need to scan around in a certain period instead of shooting a fixed scene. For example, if the human pose is recorded every second in the video, discrete-time models cannot provide the forecasting result within 1 second. Instead, we propose to use neural ordinary differential equation (NeuralODE), a continuous-time model, to propagate information along the temporal axis in our framework. Secondly, we notice that any joint of a person is only connected to its nearest neighbors, which makes the adjacency matrix used in the graph convolution very sparse, and leads to inefficient calculations in turn. To reduce the meaningless calculations and increase the efficiency of our model, we propose to utilize a hierarchical structure to decompose a sparsely connected graph into several dense graphs and encode them in a cascade manner. By doing so, we can also deal with edges of different types separately, such as the edges between people and the edges connecting joints within each person. Lastly, the whole and fixed binary adjacency matrix is used for the graph convolution in the literature, although the actual connections of joints only occupy a very small part of the graph due to the occlusion. Instead, we propose to provide dynamic topology information at each time stamp. Different from works [21, 96, 84, 58, 42, 43] which learned another weight matrix from fully observed input, we propose to learn the weight matrix from partial input to more timely capture the effect of every specific connection in the graph and multiply it with the time-varying dynamic binary adjacency matrix to form the final one used in graph convolutions.", + "bbox": [ + 76, + 90, + 472, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the main contributions of this paper are three folds:", + "bbox": [ + 76, + 696, + 468, + 724 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a continuous-time framework which takes the sequence of 2D skeletons extracted from a multi-person video as the input to forecast the multi-person 3D poses in any future time.", + "- We design our model as a hierarchical structure to reduce the computational complexity for multi-person scenarios, and also deal with different connection types in the graph.", + "- We utilize dynamic graph convolutions to deal with the irregular input caused by occlusions." + ], + "bbox": [ + 94, + 732, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the remaining of this paper, we will briefly introduce the related works in Sec. 2 and give the details of our model in Sec. 3. The performance of our model is demonstrated through experiments in Sec. 4. Finally, we will conclude our work in Sec. 5. The corresponding source code can be found at https://github.com/SBU-YCX/HDG-ODE.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 199, + 650, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Pose Estimation, Tracking and Forecasting", + "text_level": 1, + "bbox": [ + 500, + 224, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-Person Pose Estimation: Most of the 3D Single-Person Pose Estimation (SPPE) methods are proposed to recover the 3D poses from 2D ones extracted by a pre-trained 2D pose extractor such as Stacked Hourglass Network [51] or Convolutional Pose Machines [69]. Given 2D coordinates as input, Zeng et al. [93] proposed to split the human joints into local groups and estimate them separately. The same idea was used by Chen et al. [15], although they directly estimated 3D poses from raw images and applied different architectures for different parts. In Li et al.'s work [39], different subsets were exchanged to generate augment data for better training. Choi et al. [18] and Azizi et al. [5] used spectral graph convolutions to tackle the problem due to the structure property of input skeletons. Since images captured from monocular cameras have perspective limitations, Yu et al. [91] proposed perspective crop layers to eliminate the perspective effect, while Nie et al. [52] tried to use siamese architectures to decouple the view-dependent representation and pose-dependent representation. There are also some other works solved the perspective problems with the help of multi-view images, such as [50, 60, 80, 72, 94], by fusing the 2D poses from different views to get the original 3D results. Besides, other wearable sensors were used to complement the missing depth information in [94, 27], and Isogawa et al. [31] proposed to utilize an optical non-line-of-sight imaging system to obtain the 3D poses from photon images.", + "bbox": [ + 496, + 250, + 893, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Person Pose Estimation: Inspired by 2D Multi-Person Pose Estimation (MPPE) methods, the works for the 3D estimation task can also be divided into two groups. One is top-down methods where a human detector is applied first and estimation is made for each person separately. Yang et al. [88] followed this line of work with graph convolutional networks (GCNs) [32], T. Xu and W. Takano [83] modified the stacked hourglass structure also with the graph convolutions. Li et al.'s [37] proposed a hierarchical architecture, which is the one closest to our work, but their goal was to augment the graph with additional labeled mesh data and integrate more information, which is totally different from our motivation. The other one is bottom-up approaches where joint locations are directly estimated, then associated and grouped into different people. In [24], Fabbri et al. proposed to use distance-based heuristic for the associations from the detected head joint with the highest confidence.", + "bbox": [ + 496, + 643, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cheng et al. [17] proposed to integrate both top-down and bottom-up branches to give better results. Besides, inspired by YOLO [59], some works [97, 6, 35, 49] explored to obtain the 3D poses in a single pass where a pivot joint was detected as well as the offsets for all other joints. Similar to single-person case, multi-view sources were also used in [25, 10, 70, 40, 28].", + "bbox": [ + 78, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose Tracking and Forecasting: Since Andriluka et al. proposed PoseTrack [1] to first time introduce the Human Pose Tracking (HPT) task, contextual information has begun to be used for better estimating human poses in a sequence of inputs. Similar to other 3D pose tasks, most of works used extracted 2D pose sequence as the input and adopted top-down pipeline to mainly focus on single-person's pose tracking. Cai et al. [9] integrated spatio-temporal GCN into a stacked hourglass architecture to get the information of different scales. The same idea was applied in Wang et al.'s work [75], where the U-shaped GCN was used to extract multi-scale contexts. Liu et al. [41] proposed a graph attention spatio-temporal network with dilated temporal convolutions to get both local and global information, and Sofianos et al. [64] built a space-time-separable GCN for the 3D pose forecasting. Encoder-decoder architecture was utilized in [36, 63, 26] and Zheng et al. [98] solved the problem with spatio-temporal transformer. Furthermore, Chen et al. [11] proposed a fast method with multi-view inputs and Yuan et al. [92] utilized physics simulators and the reinforcement learning strategy in their work. Besides the skeleton joint representations, bone lengths and directions were used in [13, 82] and the extracted information was propagated through an LSTM.", + "bbox": [ + 80, + 199, + 468, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Graph Neural Networks", + "text_level": 1, + "bbox": [ + 80, + 573, + 299, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from traditional convolutional neural networks (CNNs), graph neural networks (GNNs) are mainly proposed to deal with data having more general structures. GNNs can be divided into two groups, spatial and spectral graph convolutions, as described in [8]. The former one considers the graph as locally connected and the information is only shared between each node and its nearest neighbors. J. Atwood and D. Towsley [4] extended it to allow the communications among nodes within $k$ -hop ( $k >= 1$ ) connections by using the probabilistic transitions. The latter one converts the data into spectral domain first, multiplies it with the Laplacian of the topology matrix in the spectral domain to mimic the convolutional operation in the spatial domain. Compared with spatial graph convolutoins, spectral operations utilize the global graph structure each time. However, since the kernel of the whole graph is large, to reduce the computational complexity, Defferrard et al. [23] proposed to approximate it with the $k^{th}$ order polynomials of the Laplacian with the help of Chebyshev expansion to model the $k$ -hop connections. T. Kipf and M. Welling [32] proposed graph convolutional", + "bbox": [ + 80, + 599, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "networks (GCNs) to further simplify it with the $1^{st}$ order approximation of the covolution in ChebNet [23].", + "bbox": [ + 503, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the pose estimation tasks, GCNs were widely used in [88, 83, 37, 9, 75, 41]. However, unlike CNNs, traditional GCNs have shared weights for different neighbors. To tackle this problems, another learnable weight matrix was trained in [96, 21, 42]. But, in these works, the weight matrix were learned with full joints and the final matrix used for convolution was the multiplication between the weight matrix and the static skeleton matrix assuming all joints observable. In our work, to deal with the missing case caused by occlusion, we make the skeleton matrix dynamic according to the current occlusion status so that the learnt weight matrix can also be effective even though only partial joints given. Some other works [43, 58, 84] proposed to add more semantic connections to the graph, which increased the calculation complexity of the models since the joints of all people could not be fully-connected. In addition, the semantic graphs were still globally sparse although locally dense. In this work, we propose to decompose the whole graph in a hierarchical manner.", + "bbox": [ + 503, + 122, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Neural Differential Equation", + "text_level": 1, + "bbox": [ + 503, + 415, + 754, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned earlier, almost all the existing pose tracking methods were built based on spatio-temporal graph convolutional networks (ST-GCNs) and recurrent neural networks (RNNs), which could only model the discrete-time data where all human joints were assumed to be observable although there existed missing and inaccurate inputs due to occlusions. In order to better model the continuous-time process of the irregular data, neural ordinary differential equation (NeuralODE) was proposed by Chen et al. [12], where deep neural networks were utilized to parameterize a nonlinear ordinary differential equation (ODE). Rubanova et al. [61] and Brouwer et al. [22] further extended it by introducing a recurrent structure to efficiently encode the input information into the underlying trajectories. Some other works [44, 45, 46] also inserted Brownian terms into ODE to make it a stochastic model that can capture the uncertainties in the dynamics. In these works, pose estimation was also mentioned as a part of experiment to prove their efficiency. However, when they dealt with the human joint data, they merely considered each joint sequence as an isolated trajectory and ignored the relations among these joints. Graph neural networks were introduced to parameterize the ODEs in [56, 29, 57], but all of them utilized the whole graph, which didn't take the sparsity problem into consideration in multi-person scenarios.", + "bbox": [ + 503, + 439, + 890, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Hierarchical Dynamic-Graph Ordinary Differential Equation Model", + "text_level": 1, + "bbox": [ + 503, + 827, + 890, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will introduce the details of our hierarchical ordinary differential equation model with dynamic", + "bbox": [ + 503, + 869, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14702", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "graph convolutions, which we call as HDG-ODE. As shown in Figure 1, the framework of our model follows a hierarchical structure, where each level is a sub-graph with the state changes modeled through graph ordinary differential equation and parameterized with dynamic graph convolutions. In the remaining of this section, we will describe the details of each component as well as the advantages of such designs.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem Statement and Notations", + "text_level": 1, + "bbox": [ + 76, + 208, + 374, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a multi-personal video containing $K$ people, and the 2D joints sequence up to time $t$ is $\\mathcal{X} = \\{\\mathbf{X}_1,\\mathbf{X}_2,\\dots,\\mathbf{X}_t\\}$ , where $\\mathbf{X}_t\\in \\mathcal{R}^{[K,N_j,2]}$ and $N_{j}$ is the number of joints within each person. These 2D representations can be obtained from any existing 2D pose estimator. Our goal is to predict the corresponding 3D poses $\\mathbf{Y}_{\\tau}\\in \\mathcal{R}^{[K,N_j,3]}$ at a future time $\\tau$ from $\\mathbf{X}_t$ , where $\\tau$ can be any value within $(t,t + 1]$ . Like in many 2D estimators, the missing or occluded joints are represented as some singular values or estimated with a very low confidence, based on this, we can directly obtain the corresponding occlusion mask $\\mathbf{M}_t$ along with $\\mathbf{X}_t$ at each time $t$ .", + "bbox": [ + 75, + 231, + 473, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hierarchical Graph Structure", + "text_level": 1, + "bbox": [ + 76, + 422, + 344, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As explored in many other works, compared to the traditional use of CNN, graph convolution is more suitable for human skeleton data due to the irregular connection of joints. The feature matrices $Z^{(l)}$ and $Z^{(l + 1)}$ of two adjacent convolution layers are expressed as", + "bbox": [ + 75, + 448, + 472, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ ^ {(l + 1)} = \\sigma (W Z ^ {(l)} \\tilde {A}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 534, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W$ is a learnable parameter matrix, $\\tilde{A}$ is the normalized form of the adjacency matrix $A \\in [0,1]^{N \\times N}$ and $N$ is the number of nodes in the graph. However, when modeling human skeletons, we notice that each joint node is only connected to a few other joints, which will make the static skeleton-based $\\tilde{A}$ very sparse and lead to plenty of meaningless calculations in (1). Although some recent works tried to learn more semantic connections among different joints besides static ones, the sparsity problem cannot be eliminated. The semantic relationship among joints usually exists within a limited number of hops instead of the whole graph. Even after we add those semantic edges, we can separate all joints into several groups such that the connections within each group are dense but the ones among different groups are still sparse. This inspires us to adopt a hierarchical structure instead of a flatten one to represent the graph of human skeletons.", + "bbox": [ + 75, + 566, + 472, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, according to the physical property of human motion, different joints change their positions at different speeds over time. The location of the central joint usually has a slow and smooth change while the extremity joints, such as \"wrist\" and \"ankle\", have drastic and unexpected", + "bbox": [ + 75, + 824, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "changes. Therefore, rather than allocating the same portion of computational resource to all joints, we can focus more on the latter ones. The cascade processing along the hierarchical structure also allows us to utilize the relative-central ones' locations to guide the estimation of positions of the relative-extreme joints.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another reason for which we adopt the hierarchical structure is its advantage of dealing with different types of edges in the whole graph. For example, in the multi-person scenario, previous works commonly used top-down methods where each person in the scene would be processed individually, but actually, pose estimation and tracking can benefit from taking the relation among people into consideration. However, the edges describing the inter-person relations are different from those representing the intra-person limbs, and it will reduce the effect if we model them in one large graph. Intuitively, separating them into different groups and forming a graph at each level will be a better choice.", + "bbox": [ + 496, + 181, + 893, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as shown in Figure 2, for each frame of the video at time $t$ , we get a set of 2D skeleton coordinates of all the people in the scene as input. To apply the hierarchical pipeline, we choose the \"pelvis\" joint as the root node of each person, whose coordinate can be also considered as the location of the person. All these root nodes form the 1st-level graph and encoded by a graph encoder", + "bbox": [ + 496, + 362, + 893, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {\\tau} ^ {(1)} = G _ {1} \\left(\\mathbf {X} _ {t} ^ {(1)}, \\mathbf {H} _ {< t} ^ {(1)}, \\tilde {A} _ {t} ^ {(1)}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 474, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_t^{(1)}\\in \\mathcal{R}^{[K,1,2]}$ represents the 2D coordinates of all root nodes of $K$ people, $\\mathbf{Y}_{\\tau}^{(1)}\\in \\mathcal{R}^{[K,1,3]}$ is the corresponding 3D one forecasted for the future time $\\tau$ , $\\mathbf{H}_{< t}^{(1)}$ is the historical feature, and $\\tilde{A}_t^{(1)}$ is the normalized adjacency matrix with the edges calculated based on the distances among connected roots, i.e. the relations among people. There is an edge between two people if their root nodes' distance is smaller than a threshold (the average of the heights of joints in 2D skeleton, i.e. distance from \"head\" to \"ankle\" joint in our setting), which indicates they are related. Then, all the limb-root joints, including \"neck\", \"left shoulder\", \"right shoulder\", \"left hip\" and \"right hip\", as well as the root node are grouped to form the 2nd-level graph within each person. To utilize the result of the 1st-level graph, the output of this level is designed as the spatial shifts of the limb-root joints' locations from the root node. We can express the process as below", + "bbox": [ + 496, + 503, + 893, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\tau} ^ {(2, i)} = G _ {2} \\left(\\mathbf {X} _ {t} ^ {(2, i)}, \\mathbf {Y} _ {\\tau , i} ^ {(1)}, \\mathbf {H} _ {< t} ^ {(2, i)}, \\tilde {A} _ {t} ^ {(2, i)}\\right), i \\in [ 1, K ]\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 756, + 849, + 775 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {\\tau} ^ {(2, i)} = \\mathbf {D} _ {\\tau} ^ {(2, i)} + \\mathbf {Y} _ {\\tau , i} ^ {(1)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 777, + 890, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_t^{(2,i)}\\in \\mathcal{R}^{[1,6,2]}$ is the 2D coordinates vector of all limb-root nodes as well as root node of person $i$ $\\mathbf{Y}_{\\tau ,i}^{(1)}$ is the 3D root node's position of this person obtained by (2), $\\mathbf{D}_{\\tau}^{(2,i)}$ and $\\mathbf{Y}_{\\tau}^{(2,i)}$ represent respectively the future 3D position shift and final position predicted. Finally, we group each limbroot joint to the corresponding limb-extreme joint to form", + "bbox": [ + 496, + 803, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/467062f85237c008b503f50662a640fbf8192634ab81d8ed1949e26121aaca02.jpg", + "image_caption": [ + "Figure 1. The framework of HDG-ODE. $\\mathbf{X}_{t_n}$ are the 2D joints coordinates at time $t_n$ , $\\mathbf{M}_{t_n}$ is the occlusion mask obtained from $\\mathbf{X}_{t_n}$ . $\\mathbf{Y}_{\\tau}$ are the predicted 3D coordinates at the future time $\\tau$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 851, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the 3rd-level graph. Similarly, the encoding process within each lamb group $j$ of person $i$ can be expressed as", + "bbox": [ + 75, + 395, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\tau} ^ {(3, i, j)} = G _ {3} (\\mathbf {X} _ {t} ^ {(3, i, j)}, \\mathbf {Y} _ {\\tau , j} ^ {(2, i)}, \\mathbf {H} _ {< t} ^ {(3, i, j)}, \\tilde {A} _ {t} ^ {(3, i, j)}), j \\in [ 1, 5 ]\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 425, + 447, + 444 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {\\tau} ^ {(3, i, j)} = \\mathbf {D} _ {\\tau} ^ {(3, i, j)} + \\mathbf {Y} _ {\\tau , j} ^ {(2, i)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 445, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_t^{(3,i,j)}\\in \\mathcal{R}^{[1,3,2]}$ is the vector of the 2D coordinates representing all nodes of person $i$ within the limb group $j$ , $\\mathbf{Y}_{\\tau ,j}^{(2,i)}$ is the 3D location of the lamb-root $j$ calculated by (3), $\\mathbf{H}_{< t}^{(3,i,j)}$ and $\\tilde{A}_t^{(3,i,j)}$ represent the corresponding part of the historical feature and the adjacency matrix in this level. 3D shift $\\mathbf{D}_{\\tau}^{(3,i,j)}$ is used to get the final 3D coordinates $\\mathbf{Y}_{\\tau}^{(3,i,j)}$ at future time $\\tau$ .", + "bbox": [ + 75, + 470, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0181a44ee320cd1948ffc81268471da4d1f4eb12dd6ce9120ed1e043183479d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 130, + 594, + 271, + 752 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/683fb5334ee6e4109990615963615a7d16c1e3913807e2f2d94003cd47119634.jpg", + "image_caption": [ + "(b)", + "Figure 2. Illustration of the hierarchical graph structure: (a) the skeleton representations of a multi-person scenario; (b) the corresponding 3-level hierarchical structure." + ], + "image_footnote": [], + "bbox": [ + 277, + 593, + 419, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Time-Varying Dynamic Graph Convolution", + "text_level": 1, + "bbox": [ + 76, + 830, + 447, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A most important impact factor to the performance of human pose estimation is the occlusion, no matter the occlusion is caused by other objects or the self-occlusion happens", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "within a person. To get the 2D skeleton inputs with some 2D pose estimation methods, the visible joints can be directly extracted from the raw images while the occluded ones are missing. Even though the missing joints can be estimated by some intuition, they are less accurate due to the uncertainties. When we use such 2D poses as inputs to recover the corresponding 3D poses, called the lifting process, the inaccuracy will accumulate and compromises data processing during the information communication within the graph convolution operation. Therefore, we hope our graph encoder can be reasonably adaptive to deal with the missing and inaccurate cases.", + "bbox": [ + 496, + 395, + 893, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the literature work that apply the traditional graph convolution for this task, the adjacency matrix $A$ and its normalized form $\\tilde{A}$ in (1) usually keep constant to represent the static connections within the graph, i.e. all the joints are assumed to be fully available, no matter they are observable or occluded. However, as we described, only those visible ones are accurate and worth being integrated to guide the forecasting. So we propose to use the time-varying dynamic graph convolution to only encode visible joints input at each timestamp. Inspired by some literature work [96], the learning of a hybrid adjacency-weight matrix may be written as", + "bbox": [ + 496, + 575, + 892, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {t} ^ {(l + 1)} = \\sigma \\left(W Z _ {t} ^ {(l)} \\rho \\left(W _ {\\tilde {A}} \\odot \\tilde {A}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 763, + 890, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{\\tilde{A}}$ is a learnable weight matrix applied to assign the information along different edges with different weights to solve the weight-sharing problem [21, 42] and $\\rho(\\cdot)$ is a nonlinear function. However, we notice that in (5), $W_{\\tilde{A}}$ is learnt based on the fully-observable data and the $\\tilde{A}$ is the static one without taking the occlusion cases into consideration. In other words, although the learnt weights work well when all", + "bbox": [ + 496, + 792, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14704", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "joints are available, they may not be effective if partial of joints are occluded and missing. Based on this, we make a further step and propose to also take the occlusion status into consideration when obtaining the dynamic weight matrix. Specifically, at each timestamp $t$ , we will drop the occluded joints in the original static adjacency matrix $A$ to form a time-varying adjacency matrix $A_{t}$ , and this dynamic adjacency matrix will be used in both training and testing, i.e. (5) can be rewritten as", + "bbox": [ + 73, + 90, + 472, + 224 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {t} ^ {(l + 1)} = \\sigma \\left(W Z _ {t} ^ {(l)} \\rho \\left(W _ {\\tilde {A}} \\odot \\tilde {A} _ {t}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 232, + 470, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{A}_t$ is the normalized form of the real adjacency matrix at time $t$ . By doing so, the learning of the weight matrix $W_{\\tilde{A}}$ will only focus on the observable joints to make sure it can also be effective when only partial joints are available. In the testing phase, the hybrid-matrix $W_{\\tilde{A}} \\odot \\tilde{A}_t$ can ignore the missing joints and adaptively adjust the weights along existing edges for information integration among joints.", + "bbox": [ + 75, + 262, + 472, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Parallel Ordinary Differential Equation", + "text_level": 1, + "bbox": [ + 76, + 377, + 419, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides the partial occlusion, which leads to incomplete graph input at each frame and can be regarded as spatial missing, there are also many cases where a person is totally occluded, and causes the temporal missing within a sequence. Previous works almost rely on the discrete-time sequential models, which are not flexible enough since the missing in real world can vary from several frames to dozens of frames.", + "bbox": [ + 75, + 400, + 470, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To make the model more robust, we adopt ordinary differential equation (ODE), a continuous-time model in this paper, which can be expressed as", + "bbox": [ + 75, + 506, + 470, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d \\mathbf {H} _ {t}}{d t} = F (\\mathbf {H} _ {t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 558, + 310, + 588 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {t} = \\mathbf {H} _ {t _ {n - 1}} + \\int_ {t _ {n - 1}} ^ {t} F (\\mathbf {H} _ {\\tau}) d \\tau . \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 590, + 468, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared with traditional recurrent networks, it can be used to predict the values over the interval of any length. Specially, in our model, we use time-varying dynamic graph convolution mentioned in Sec. 3.3 to parameterize the derivatives $F(\\cdot)$ . For efficiency, we compute the integration in (7) by Euler Method and we can rewrite it as", + "bbox": [ + 76, + 633, + 468, + 723 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {t, F} = \\mathbf {H} _ {t - \\Delta t, F} + F \\left(\\mathbf {H} _ {t - \\Delta t, F}, \\tilde {A}\\right) \\Delta t, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 732, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{H}_{t,F}$ is the hidden feature, $\\tilde{A}$ is the normalized adjacency matrix assuming all joints observable, and $\\Delta t$ is the step-size. At an observation time $t_n$ , the features of observable joints are updated through the same graph convolution encoder as described in Sec. 3.2, i.e.", + "bbox": [ + 75, + 758, + 470, + 835 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {t _ {n}, G} = G \\left(\\mathbf {X} _ {t _ {n}}, \\mathbf {Y} _ {t _ {n}}, \\mathbf {H} _ {< t _ {n}}, \\tilde {A} _ {t _ {n}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 844, + 468, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $G = \\{G_1, G_2, G_3\\}$ represents the cascade process in (2)-(4) parameterized by our time-varying dynamic graph", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "convolution networks, $\\mathbf{X}_{t_n}$ and $\\mathbf{Y}_{t_n}$ are the 2D and 3D joint locations, $\\mathbf{H}_{< t_n} = \\mathbf{H}_{t_n - \\Delta t,F}$ is the feature from the ODE model and $\\tilde{A}_{t_n}$ is the normalized adjacency matrix of the graph formed by observable joints. After the update, the new feature at the time instant $t_n$ becomes", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {t _ {n}} = \\mathbf {M} _ {t _ {n}} \\odot \\mathbf {H} _ {t _ {n}, F} + \\left(\\mathbf {1} - \\mathbf {M} _ {t _ {n}}\\right) \\odot \\mathbf {H} _ {t _ {n}, G}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 180, + 893, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{t_n}$ is the occlusion mask of the input described in Sec. 3.1. Here, we want to emphasize that $\\tilde{A}_{t_n}$ in (9) and $\\tilde{A}$ in (8) may not be the same according to the observation states of input. As we described in last section, the input may contain missing values, but all the missing values have been completed by the encoder in the encoding process and we can assume the graph is fully observable in the propagation process.", + "bbox": [ + 496, + 208, + 893, + 329 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02f0ae8ba3f13ea4c5935e91071fac87e299de3555ec1a49cde77b910f72fbb6.jpg", + "image_caption": [ + "Figure 3. Illustration of the parallel ODE pipeline." + ], + "image_footnote": [], + "bbox": [ + 532, + 344, + 859, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we can further demonstrate the advantages of hierarchical structure. For efficiency, we adopt Euler method to approximately calculate the integration, but as we know, it will predict the curve as a straight line if we make the propagation step $\\Delta t$ too large and equal to the interval between two observations as $t_n - t_{n-1}$ . To make the model more expressive, we usually make $\\Delta t$ smaller, i.e.", + "bbox": [ + 496, + 492, + 893, + 598 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta t = \\frac {1}{k} \\left(t _ {n} - t _ {n - 1}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 608, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, as we discussed in Sec. 3.2, different joints of human body have different rates of location changes, we don't need to set $k$ to the same value at different levels in our model. Instead, we can adopt a parallel pipeline and make it small in shallow levels while enlarge it as the level goes deeper as shown in Figure 3.", + "bbox": [ + 496, + 647, + 893, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 752, + 633, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 500, + 777, + 696, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Datasets", + "text_level": 1, + "bbox": [ + 500, + 800, + 612, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MuPoTS-3D[49]: This multi-person dataset contains sequences captured in 20 real scenes. Each sequence has up to three subjects, as well as provides raw images, 2D pose annotations and the corresponding 3D ones. The lengths of data vary from 125 frames to 800 frames, and we split them", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14705", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "into segments with the length of 50 frames. We divide all the data into training, validation and testing sets with the ratio of 0.8: 0.1: 0.1. The real occlusion mask are given as well, and the final data are obtained by the element-wise multiplication between the full data and the occlusion masks. Before being fed into the models, data from each sequence are first normalized.", + "bbox": [ + 75, + 90, + 472, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human3.6M[30]: This dataset consists of 3.6 million of single-person frames. 11 subjects conducting 15 activities are captured by cameras. Both 2D and 3D annotations are provided where 2D ones come from cameras while 3D ones are obtained using a motion capture system. Different from commonly used training methods, we only use data from one camera to train the model. We performed the same data partition and normalization as above.", + "bbox": [ + 75, + 196, + 472, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 335, + 264, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mean Per Joint Position Error (MPJPE): This metric measures the average position error over all joints and frames in the sequence, which can be expressed as", + "bbox": [ + 76, + 359, + 468, + 402 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM P J P E = \\frac {1}{N _ {f}} \\sum_ {f} \\frac {1}{N _ {j} ^ {(f)}} \\sum_ {j} \\left\\| P _ {\\text {p r e d}} ^ {(f)} (j) - P _ {g t} ^ {(f)} (j) \\right\\| _ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 410, + 468, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N_{f}$ is the total number of frames, $N_{j}^{(f)}$ is the total number of joints in the frame $f$ , $P_{gt}^{(f)}(j)$ and $P_{pred}^{(f)}(j)$ are the ground-truth and the predicted 3D position of the joint $j$ in the frame $f$ .", + "bbox": [ + 76, + 455, + 468, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Percentage of Correct Keypoints (PCK): PCK calculates the percentage of correctly predicted key-points, where a key-point $j$ in the frame $f$ is considered to be correct if the distance $d_{p}^{(f)}(j)$ between the 3D ground-truth $P_{p,gt}^{(f)}(j)$ and the predicted 3D position $P_{p,pred}^{(f)}(j)$ are smaller than a person-specific threshold $S_{p}^{f}$ , i.e.", + "bbox": [ + 76, + 521, + 470, + 619 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP C K = \\frac {\\sum_ {f} \\sum_ {p} \\sum_ {j} \\delta \\left(d _ {p} ^ {(f)} (j) < S _ {p} ^ {(f)}\\right)}{\\sum_ {f} \\sum_ {p} \\sum_ {j} 1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 628, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we adopt the setting in MPII dataset[2] where the threshold is the ratio of head link's length $h_p^{(f)}$ of the person $p$ , i.e. the distance between \"head\" joint and \"upper neck\", and the PCK can be represented as", + "bbox": [ + 76, + 674, + 468, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP C K h @ \\alpha = \\frac {\\sum_ {f} \\sum_ {p} \\sum_ {j} \\delta \\left(\\frac {d _ {p} ^ {(f)} (j)}{h _ {p} ^ {(f)}} < \\alpha\\right)}{\\sum_ {f} \\sum_ {p} \\sum_ {j} 1}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 743, + 468, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ is the ratio parameter.", + "bbox": [ + 76, + 797, + 279, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3 Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 830, + 294, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As introduced in Sec. 3.2, the hierarchical graph has 3 levels, containing $K$ , 6 and 3 nodes each, and our HDG-ODE model has a hierarchical structure of 3 levels, with an encoder and", + "bbox": [ + 76, + 854, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an ODE at each level. For each encoder, we use a single-layer graph-GRU, using our time-varying dynamic graph convolutions with the feature size of 6. ODE models are parameterized with graph neural networks of 1, 2 and 3 convolutional layers at the three levels respectively, with the hidden feature size of 16. The propagation steps for the three levels are respectively $[1/30s, 1/60s, 1/120s]$ , and the data sampling interval is $1/30s$ . The experiments are conducted on single NVIDIA 1080 Ti GPU. During the training, we use the Adam() optimizer with the learning rates of [0.01, 0.001, 0.0001] to minimize the mean squared error (MSE) between the predicted 3D positions and the ground-truth of the visible joints, and 100 epochs are run at each learning rate. To speed up the training, it will be early-stopped if there is no further loss decrease in 20 epochs.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experiment Results", + "text_level": 1, + "bbox": [ + 500, + 327, + 687, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Overall Performance", + "text_level": 1, + "bbox": [ + 500, + 349, + 700, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments over MuPoTS-3D dataset to evaluate the performance of our proposed model on the multiperson 3D pose tracking task. For comparison, we choose some representative methods, including discrete-time recurrent models and continuous-time ones, as the baselines. Specifically we take the following methods into consideration:", + "bbox": [ + 496, + 375, + 893, + 479 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- STGCN: This discrete-time model consists of several concatenated spatio-temporal convolutional blocks (ST-Conv blocks), which is a common structure that is widely used for action recognition [85], pose forecasting [75] and other prediction tasks [89]. In our experiments, STGCN has 2 ST-Conv blocks, and recurrently processes each clip within the whole 2D pose sequence and give the 3D predictions. The output dimension of each layer in the blocks is 32.", + "- Graph-GRU: It is a variant of conventional Gated Recurrent Unit (GRU), where the linear operation in the conventional GRU is replaced by graph convolution. It works as the backbone in [90]. We construct the baseline with 2 Graph-GRU layer and a linear decoder. The output dimension of each layer in the blocks is 32.", + "- SemGCGRU: It is an improved version of traditional Graph-GRU, where the fixed binary adjacency matrix is replaced by a learnable weight matrix to capture the relations among nodes [96]. In our implementation, the structure and the feature size are the same as Graph-GRU.", + "- ODE-RNN: In this model [62], the deep learning modules, specifically dense neural networks, are incorporated to parameterize a non-linear ordinary differential equation (ODE) to model the continuous-time process" + ], + "bbox": [ + 517, + 489, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14706", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of data sequence. In our experiments, the dense network contains 2 linear layers and the dimension of hidden features is 32.", + "bbox": [ + 107, + 90, + 468, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Graph-ODE: It is a GNN variant of ODE-RNN, with a GCN to learn the derivative of ODE and a GRU to integrate the information of observations. In our model comparison, the GCN has 2 hidden layers with ReLU activation functions and the output dimension of each hidden layer is 32. The output dimension of GRU is 32 as well. In other words, to compare with discrete models, we just replace the GRU module in Graph-GRU with an ODE module.", + "bbox": [ + 96, + 145, + 470, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although continuous-time models can provide predictions at any time, even smaller than the data sampling rate, we set the step of them to be equal to the original data interval used in discrete-time models for fair comparison. From the results in Table 1, we can see that our HDG-ODE can achieve the best result with the smallest mean prediction error and the highest percentage of corrected prediction.", + "bbox": [ + 75, + 292, + 472, + 398 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e50cc07ad49685a6210e64be3799879878d499628dbefcb3807903837d0ec8c7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MPJPE(↓)PCKh@α(↑)
α=0.1α=0.5α=1.0
Discrete
STGCN [89, 75]0.600413.03%60.35%77.76%
Graph-GRU [90]0.557314.86%62.39%80.04%
SemGCGRU [96]0.437517.09%69.61%85.96%
Continuous
ODE-RNN [61]0.439620.21%71.78%85.45%
Graph-ODE [56]0.406619.80%71.57%87.00%
Ours
HDG-ODE0.303829.72%78.67%92.18%
", + "bbox": [ + 78, + 411, + 486, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2 Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 635, + 236, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this part, we will separately demonstrate the effectiveness of each designing component in our proposed model by some experiment results, including the continuous ODE model, the pipeline taking all people into consideration, the new time-varying dynamic graph convolution and the proposed hierarchical structure. The evaluation is sequential by adding the components one by one, and each time we only choose the current best model to conduct the experiment for the next component.", + "bbox": [ + 75, + 659, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discrete-Time Model vs. Continuous-Time Model: From Table 1, compared to traditional discrete-time models, no matter our HDG-ODE or other continuous-time models have a better performance. Different from discrete models which only execute one propagation between two observations, the continuous-time ones conduct multiple-step propagation so that the information on each joint will be affected", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and corrected by its neighbors several times before getting the final prediction. Also, due to the property of ODE, the continuous-time model can provide predictions at any time in the interval between two observations, instead of just obtaining predictions at the observation instants as done by discrete-time ones.", + "bbox": [ + 496, + 90, + 893, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Isolated Top-Down Pipeline vs. Relational Top-Down Pipeline: As we described previously, most of existing multiperson pose tracking works adopt the top-down processing pipeline, where each person in the scenario is tracked separately as an isolated trajectory. Instead, we take into consideration the relation among people to infer their relative movements, which in turn affects the future positions of specific joints. In Table 2, by comparing the models in isolated top-down pipeline with the ones in our relational pipeline, we can observe the improvement.", + "bbox": [ + 496, + 181, + 895, + 333 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/62dee769535a46191f27a9ad55b3f88dd0a94ff887406012f782b9edc98fbe9d.jpg", + "table_caption": [ + "Table 1. Testing Performance of different models on MuPoTS-3D." + ], + "table_footnote": [], + "table_body": "
Isolated PipelineRelational Pipeline
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.610477.72%0.600477.76%
Graph-GRU0.573878.46%0.557380.04%
Graph-ODE0.420386.89%0.406687.00%
", + "bbox": [ + 500, + 339, + 898, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Static Graph Convolution vs. Time-Varying Dynamic Graph Convolution: In the real cases, occlusion is inevitable. To better tackle the problem caused by the input data with missing or inaccurate values due to occlusions, we propose to replace the static graph convolution used in most of works with our time-varying dynamic graph convolution in the HDG-ODE model. As shown in Table 3, the dynamic graph convolution helps eliminate the negative influence of inaccurate inputs as well as adjust the integration weights of edges to the most suitable ones according to occlusion states, so the models can have a better performance than before.", + "bbox": [ + 496, + 450, + 893, + 617 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1ec8507bb8b45ac36316d671413c0b5f5e345434d9f641eda0d4801d0e4bfbd.jpg", + "table_caption": [ + "Table 2. Comparison between the isolated top-down pipeline and our relational top-down pipeline on MuPoTS-3D." + ], + "table_footnote": [], + "table_body": "
Original Graph ConvOur Graph Conv
MPJPEPCKh@1.0MPJPEPCKh@1.0
Graph-GRU0.557380.04%0.384486.45%
Graph-ODE0.406687.00%0.370688.74%
", + "bbox": [ + 500, + 625, + 898, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison between original static graph convolutions (GC) and our time-varying dynamic convolutions (DG) on MuPoTS-3D.", + "bbox": [ + 500, + 691, + 893, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Flatten Structure vs. Hierarchical Structure: To solve the sparsity problem explained in Sec. 3, our HDG-ODE adopts a hierarchical structure. In Table 4, compared to models in the flatten structure with a large sparse graph, the corresponding hierarchical ones perform better for models with graph convolutions. The reasons leading to the improvement lie in two aspects: On one hand, by decomposing the whole sparse graph into a set of relatively dense ones, the parameters to learn are reduced, as shown in Table 5. This helps save a significant amount of computational resource, and the released resource can be reallocated to the joints", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14707", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with more complicated movements, e.g. making the corresponding network deeper or reducing the propagation step of ODE. On the other hand, the cascade processing along the hierarchical structure can provide the guidance among the levels. As for the ODE-RNN, the hierarchical one has a worse result because the structure is mainly proposed to deal with the graph sparsity problem. However, ODE-RNN doesn't utilize the graph information and only cares about isolated joint trajectories, so the hierarchical structure will make the model more complex and reduce the performance.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a99c52fe7a807f62df162b44a34922447157668ab83cd9ec3fc93b46adc0f516.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Flatten StructureHierarchical Structure
MPJPEPCKh@1.0MPJPEPCKh@1.0
ODE-RNN0.439685.45%0.493882.40%
GC-ODE0.406687.00%0.345190.51%
DG-ODE0.370688.74%0.303892.18%
", + "bbox": [ + 78, + 250, + 470, + 325 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/feac483372cd67ec85c604fc7c5ebd01a377527a6fd461abde1ba798a7198d87.jpg", + "table_caption": [ + "Table 4. Performance Comparison between the flatten structure and the hierarchical structure on MuPoTS-3D." + ], + "table_footnote": [], + "table_body": "
#Params (↓)Memory (↓)
DG-ODE106490.76M
HDG-ODE26430.66M
", + "bbox": [ + 135, + 378, + 413, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2.3 Robustness Analysis", + "text_level": 1, + "bbox": [ + 76, + 476, + 272, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Besides the multi-person scenarios, we also conduct experiments of single-person scenarios on Human3.6M dataset. Different from MuPoTS-3D, no occlusion information is provided within the data, so we make occlusion masks by ourselves and test the robustness of our model under different observable ratios. Specifically, for a sequence, we assume only $p_t$ of the frames are observable while in the remaining ones, the input skeletons are totally missing due to the occlusion by buildings or other large objects. For each observable frame, we assume there exist self-occlusions such that only $p_s$ of the joints are available. To make the corresponding mask, both $p_t$ of the frames and $p_s$ of the joints on each frame are randomly chosen and the total observable ratio can be expressed as $p_t \\times p_s$ . The results are shown in Table 6. When the observable ratio decreases as a result of occlusions, the performance gets worse for all models. However, our HDG-ODE always performs the best among all models studied under all circumstances.", + "bbox": [ + 75, + 500, + 472, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 784, + 194, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a new continuous-time model called HDG-ODE to forecast the future 3D human pose representations in multi-person videos given the 2D pose sequence as input. By using ordinary differential equations and dynamic graph convolutions, we alleviate the occlusion problem brought by perspective effects of monocular images.", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/40151a90215901480d582d0a2771deb6f3b71c7c37d1f7fabc9e3181e9f10862.jpg", + "table_caption": [ + "Table 5. Complexity Comparison between the flatten Dynamic-Graph ODE (DG-ODE) and our hierarchical one (HDG-ODE)." + ], + "table_footnote": [], + "table_body": "
pt=0.8, ps=0.5pt=0.6, ps=0.4
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.396680.19%0.437675.82%
Graph-GRU0.372383.34%0.418877.39%
ODE-RNN0.352784.33%0.382680.66%
Graph-ODE0.347383.98%0.401978.54%
HDG-ODE0.305288.29%0.354382.36%
", + "bbox": [ + 501, + 84, + 897, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 6. Performance of different models on Human3.6M with different observable ratios.", + "bbox": [ + 500, + 191, + 890, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Besides, with the help of hierarchical structure, we reduce the computational complexity of our model and are able to allocate more resources to the human joints which have more complicated movement. Our experiments demonstrate that our model can well recover the corresponding 3D poses from the single-view 2D poses and outperform literature works on the human pose forecasting task. Currently, the maximum number of people in the scenes of the video is limited. In the future, we will make our model more adaptive and test it in more crowded scenarios.", + "bbox": [ + 496, + 242, + 893, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 407, + 596, + 422 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. Posetrack: A benchmark for human pose estimation and tracking, 2017.", + "[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele.", + "[3] Bruno Artacho and Andreas Savakis. Unipose: Unified human pose estimation in single images and videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[4] James Atwood and Don Towsley. Diffusion-convolutional neural networks, 2015.", + "[5] Niloofar Azizi, Horst Possegger, Emanuele Rodola, and Horst Bischof. 3d human pose estimation using mobius graph convolutional networks, 2022.", + "[6] Abdallah Benzine, Florian Chabot, Bertrand Luvison, Quoc Cuong Pham, and Catherine Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[7] Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, and Lorenzo Torresani. Learning temporal pose estimation from sparsely labeled videos. In Advances in Neural Information Processing Systems 33, 2019.", + "[8] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs, 2013.", + "[9] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019." + ], + "bbox": [ + 514, + 431, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. arXiv preprint arXiv:2007.10986, 2020.", + "[11] Long Chen, Haizhou Ai, Rui Chen, Zijie Zhuang, and Shuang Liu. Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[12] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations, 2018.", + "[13] Tianlang Chen, Chen Fang, Xiaohui Shen, Yiheng Zhu, Zhili Chen, and Jiebo Luo. Anatomy-aware 3d human pose estimation in videos. arXiv preprint arXiv:2002.10322, 2020.", + "[14] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018.", + "[15] Zerui Chen, Yan Huang, Hongyuan Yu, Bin Xue, Ke Han, Yiru Guo, and Liang Wang. Towards part-aware monocular 3d human pose estimation: An architecture search approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 715-732, Cham, 2020. Springer International Publishing.", + "[16] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[17] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7649-7659, June 2021.", + "[18] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020.", + "[19] Chia-Jung Chou, Jui-Ting Chien, and Hwann-Tzong Chen. Self adversarial training for human pose estimation. In 2018 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 17-30, 2018.", + "[20] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L. Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.", + "[21] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019.", + "[22] Edward De Brouwer, Jaak Simm, Adam Arany, and Yves Moreau. Gru-ode-bayes: Continuous modeling of sporadically-observed time series, 2019." + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. 2016.", + "[24] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[25] Qi Fang, Qing Shuai, Junting Dong, Hujun Bao, and Xiaowei Zhou. Reconstructing 3d human pose by watching humans in the mirror. In CVPR, 2021.", + "[26] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In CVPR, 2022.", + "[27] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, June 2021.", + "[28] Congzhentao Huang, Shuai Jiang, Yang Li, Ziyue Zhang, Jason Traish, Chen Deng, Sam Ferguson, and Richard Yi Da Xu. End-to-end dynamic matching network for multi-view multi-person 3d pose estimation. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII, page 477-493, Berlin, Heidelberg, 2020. Springer-Verlag.", + "[29] Zijie Huang, Yizhou Sun, and Wei Wang. Coupled graph ode for learning interacting system dynamics. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery amp; Data Mining, KDD '21, page 705-715, New York, NY, USA, 2021. Association for Computing Machinery.", + "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014.", + "[31] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris M. Kitani. Optical non-line-of-sight physics-based 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[32] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks, 2016.", + "[33] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Multiposenet: Fast multi-person pose estimation using pose residual network. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018.", + "[34] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[35] Jogendra Nath Kundu, Ambareesh Revanur, Govind Vitthal Waghmare, Rahul Mysore Venkatesh, and R. Venkatesh Babu. Unsupervised cross-modal alignment for multi-person 3d pose estimation, 2020." + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14709", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R. Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[37] Han Li, Bowen Shi, Wenrui Dai, Yabo Chen, Botao Wang, Yu Sun, Min Guo, Chenlin Li, Junni Zou, and Hongkai Xiong. Hierarchical graph networks for 3d human pose estimation, 2021.", + "[38] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[39] Shichao Li, Lei Ke, Kevin Pratama, Yu-Wing Tai, Chi-Keung Tang, and Kwang-Ting Cheng. Cascaded deep monocular 3d human pose estimation with evolutionary training data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[40] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021.", + "[41] Junfa Liu, Juan Rojas, Zhijun Liang, Yihui Li, and Yisheng Guan. A graph attention spatio-temporal convolutional networks for 3d human pose estimation in video. arXiv preprint arXiv:2003.14179, 2020.", + "[42] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X, page 318–334, Berlin, Heidelberg, 2020. Springer-Verlag.", + "[43] Shengyuan Liu, Pei Lv, Yuzhen Zhang, Jie Fu, Junjin Cheng, Wanqing Li, Bing Zhou, and Mingliang Xu. Semi-dynamic hypergraph neural network for 3d pose estimation. In Christian Bessiere, editor, Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI20, pages 782-788. International Joint Conferences on Artificial Intelligence Organization, 7 2020. Main track.", + "[44] Xuanqing Liu, Tesi Xiao, Si Si, Qin Cao, Sanjiv Kumar, and Cho-Jui Hsieh. Neural sde: Stabilizing neural ode networks with stochastic noise, 2019.", + "[45] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, and Zhaoyue Chen. Learning continuous-time dynamics by stochastic differential networks, 2020.", + "[46] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, Zhaoyue Chen, and Jacqueline Wu. Continuous-time stochastic differential networks for irregular time series modeling. In Teddy Mantoro, Minho Lee, Media Anugerah Ayu, Kok Wai Wong, and Achmad Nizar Hidayanto, editors, Neural Information Processing, pages 343-351, Cham, 2021. Springer International Publishing.", + "[47] Zhenguang Liu, Haoming Chen, Runyang Feng, Shuang Wu, Shouling Ji, Bailin Yang, and Xun Wang. Deep dual consecutive network for human pose estimation. In Proceed-" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 525-534, June 2021.", + "[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13264-13273, June 2021.", + "[49] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb, 2017.", + "[50] Rahul Mitra, Nitesh B. Gundavarapu, Abhishek Sharma, and Arjun Jain. Multiview-consistent semi-supervised learning for 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[51] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation, 2016.", + "[52] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement, 2020.", + "[53] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019.", + "[54] Wanli Ouyang, Xiao Chu, and Xiaogang Wang. Multisource deep learning for human pose estimation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2337-2344, 2014.", + "[55] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-286, 2018.", + "[56] Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations, 2019.", + "[57] Michael Poli, Stefano Massaroli, Clayton M. Rabideau, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Continuous-depth neural models for dynamic graph prediction, 2021.", + "[58] Zhongwei Qiu, Kai Qiu, Jianlong Fu, and Dongmei Fu. Dgen: Dynamic graph convolutional network for efficient multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 34(07):11924-11931, Apr. 2020.", + "[59] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016.", + "[60] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. 2020.", + "[61] Yulia Rubanova, Ricky T. Q. Chen, and David Duvenaud. Latent odes for irregularly-sampled time series, 2019." + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14710", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Yulia Rubanova, Tian Qi Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series. In Advances in Neural Information Processing Systems 32, pages 5321-5331. 2019.", + "[63] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2484-2494, June 2021.", + "[64] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11209-11218, October 2021.", + "[65] Kai Su, Dongdong Yu, Zhenqi Xu, Xin Geng, and Changhu Wang. Multi-person pose estimation with enhanced channelwise and spatial information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[66] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5686-5696, 2019.", + "[67] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[68] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose estimation. ArXiv, abs/1911.07451, 2019.", + "[69] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 1653-1660, 2014.", + "[70] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020.", + "[71] Ali Varamesh and Tinne Tuytelaars. Mixture dense regression for object detection and human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13294-13304, June 2021.", + "[73] Jiahang Wang, Sheng Jin, Wentao Liu, Weizhong Liu, Chen Qian, and Ping Luo. When human pose estimation meets robustness: Adversarial algorithms and benchmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11855-11864, June 2021.", + "[74] Jian Wang, Xiang Long, Yuan Gao, Errui Ding, and Shilei Wen. Graph-pcnn: Two stage human pose estimation with graph pose refinement. In The European Conference on Computer Vision (ECCV), 2020." + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos, 2020.", + "[76] Manchen Wang, Joseph Tighe, and Davide Modolo. Combining detection and tracking for human pose estimation in videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[77] Rui Wang, Zhongzheng Cao, Xiangyang Wang, Zhi Liu, and Xiaoqiang Zhu. Human pose estimation with deeply learned multi-scale compositional models. IEEE Access, 7:71158-71166, 2019.", + "[78] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016.", + "[79] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018.", + "[80] Rongchang Xie, Chunyu Wang, and Yizhou Wang. Meta-fuse: A pre-trained fusion model for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[81] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018.", + "[82] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[83] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16105-16114, June 2021.", + "[84] Xixia Xu, Qi Zou, and Xue Lin. Adaptive hypergraph neural network for multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2955-2963, Jun. 2022.", + "[85] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition, 2018.", + "[86] Wei Yang, Shuang Li, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Learning feature pyramids for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017.", + "[87] Wei Yang, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. End-to-end learning of deformable mixture of parts and deep convolutional neural networks for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.", + "[88] Yiding Yang, Zhou Ren, Haoxiang Li, Chunluan Zhou, Xin chao Wang, and Gang Hua. Learning dynamics via graph neural networks for human pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8074-8084, June 2021.", + "[89] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), 2018." + ], + "bbox": [ + 509, + 92, + 893, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[90] Bing Yu, Haoteng Yin, and Zhanxing Zhu. ST-UNet: A spatio-temporal u-network for graph-structured time series modeling. arXiv, abs/1903.05631, 2019.", + "[91] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9064-9073, June 2021.", + "[92] Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. Simpoe: Simulated character control for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7159–7169, June 2021.", + "[93] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Srnet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 507–523, Cham, 2020. Springer International Publishing.", + "[94] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[95] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[96] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N. Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[97] Jianan Zhen, Qi Fang, Jiaming Sun, Wentao Liu, Wei Jiang, Hujun Bao, and Xiaowei Zhou. Smap: Single-shot multiperson absolute 3d pose estimation, 2020.", + "[98] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11656-11665, October 2021.", + "[99] Tianfei Zhou, Wenguan Wang, Si Liu, Yi Yang, and Luc Van Gool. Differentiable multi-granularity human representation learning for instance-aware human semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1622-1631, June 2021.", + "[100] Bin Xiao Zhaoxiang Zhang Jingdong Wang Zigang Geng, Ke Sun. Bottom-up human pose estimation via disentangled keypoint regression. In CVPR, 2021." + ], + "bbox": [ + 78, + 92, + 470, + 796 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "14712", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_model.json b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_model.json new file mode 100644 index 0000000000000000000000000000000000000000..25b819ed91e9f99f595a32d3e6a73c8b7d15a80f --- /dev/null +++ b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_model.json @@ -0,0 +1,2921 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "HDG-ODE: A Hierarchical Continuous-Time Model for Human Pose Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.182, + 0.792, + 0.234 + ], + "angle": 0, + "content": "Yucheng Xing Xin Wang \nDepartment of Electrical and Computer Engineering, Stony Brook University \n{yucheng.xing, x.wang}@stonybrook.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Recently, human pose estimation has attracted more and more attention due to its importance in many real applications. Although many efforts have been put on extracting 2D poses from static images, there are still some severe problems to be solved. A critical one is occlusion, which is more obvious in multi-person scenarios and makes it even more difficult to recover the corresponding 3D poses. When we consider a sequence of images, the temporal correlation among the contexts can be utilized to help us ease the problem, but most of the current works only rely on discrete-time models and estimate the joint locations of all people within a whole sparse graph. In this paper, we propose a new framework, Hierarchical Dynamic Graph Ordinary Differential Equation (HDG-ODE), to tackle the 3D pose forecasting task from 2D skeleton representations in videos. Our framework adopts ODE, a continuous-time model, as the base to predict the 3D joint positions at any time. Considering the structural-property of the skeleton data in representing human poses and the possible irregularity caused by occlusion, we propose the use of dynamic graph convolution as the basic operator. To reduce the computational complexity introduced by the sparsity of the pose graph, our model takes a hierarchical structure where the encoding process at the observation timestamp is done in a cascade manner while the propagation between observations is conducted in parallel. The performance studies on several datasets demonstrate that our model is effective and can out-perform other methods with fewer parameters." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.903 + ], + "angle": 0, + "content": "With the progress in the computer vision field, human-related contents have attracted more and more attention, as they are very important in many real-world applications. Some examples are autonomous driving, human-computer interaction and anomaly detection. Among these contents, human poses are the most basic and important ones since their accurate and timely estimation or even forecasting usually serve as the first step for further analyzing complex" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.568, + 0.285 + ], + "angle": 0, + "content": "activities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.896, + 0.867 + ], + "angle": 0, + "content": "Human pose estimation is not a novel topic but has been studied for many years. Its performance, however, is still not satisfactory enough to meet the demands of many practical applications. Since A. Toshev and C. Szegedy proposed DeepPose [69] to represent human poses by the locations of a set of joints from human skeletons and first time utilized the regression method of deep learning to estimate the corresponding 2D coordinates, a lot of efforts [78, 51, 87, 20, 86, 19, 79, 77, 54, 67, 95, 94, 3, 66, 74, 73] have been made to get the accurate estimation in the presence of occlusion resulted from the perspective limitation of images. These methods [14, 33, 55, 65, 34, 68, 38, 53, 16, 71, 58, 48, 99, 100, 84] are extended to multi-person scenarios, where the occlusion problem is more severe. Although some progress has been made, 2D locations cannot perfectly restore the human poses, and are never our final goals for this task. Recently, more and more works have been focusing on 3D pose estimation with raw images [15, 24, 6, 91] or extracted 2D poses [96, 21, 5, 18, 93, 39, 97, 83, 17, 43, 91] as input. Besides the inherent occlusion problems, the lack of depth in monocular images also limits the performance of algorithms. Although some schemes relying on multi-view images [50, 60, 94, 80, 72, 10, 70, 28, 40, 25, 11] or other sensors [54, 31, 27, 94] can ease this problem to some degree, we have to admit that paired images of different views or complementary sensors are not always available in most scenarios. The core problem is how to just utilize limited information in monocular images to extract the corresponding 3D poses of people. If a sequence of images (i.e. video) is available, the temporal context carried can be also helpful for the task [1, 81, 7, 47, 76]. They not only provide information to make up the missing joints caused by the occlusion and provide hints for joints' locations in the future, but also help group the detected joints belonging to a specific person in multi-person scenarios. So in this paper, we mainly focus on restoration from a sequence of monocular multi-person 2D skeletons to predict the corresponding 3D poses at a future time point." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Some related methods [64, 9, 75, 82, 41, 11, 98, 26, 13, 63, 36] usually applied graph neural networks to encode the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "14700" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.695 + ], + "angle": 0, + "content": "2D human joints and propagated the information through recurrent neural networks or directly used spatial-temporal graph convolutional networks. However, almost all of them only focused on single-person scenarios, while in most of cases, the relations among people in the scene are also helpful for the forecasting. Also, there exist several limitations in these works. Firstly, traditional recurrent neural networks, such as GRU or LSTM, all use discrete-time sequential models, which have a fixed propagation step and are not flexible enough to deal with different moving speeds of different human joints. The performance of these models will be compromised when the video frequency is not high or the cameras need to scan around in a certain period instead of shooting a fixed scene. For example, if the human pose is recorded every second in the video, discrete-time models cannot provide the forecasting result within 1 second. Instead, we propose to use neural ordinary differential equation (NeuralODE), a continuous-time model, to propagate information along the temporal axis in our framework. Secondly, we notice that any joint of a person is only connected to its nearest neighbors, which makes the adjacency matrix used in the graph convolution very sparse, and leads to inefficient calculations in turn. To reduce the meaningless calculations and increase the efficiency of our model, we propose to utilize a hierarchical structure to decompose a sparsely connected graph into several dense graphs and encode them in a cascade manner. By doing so, we can also deal with edges of different types separately, such as the edges between people and the edges connecting joints within each person. Lastly, the whole and fixed binary adjacency matrix is used for the graph convolution in the literature, although the actual connections of joints only occupy a very small part of the graph due to the occlusion. Instead, we propose to provide dynamic topology information at each time stamp. Different from works [21, 96, 84, 58, 42, 43] which learned another weight matrix from fully observed input, we propose to learn the weight matrix from partial input to more timely capture the effect of every specific connection in the graph and multiply it with the time-varying dynamic binary adjacency matrix to form the final one used in graph convolutions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.697, + 0.47, + 0.725 + ], + "angle": 0, + "content": "To summarize, the main contributions of this paper are three folds:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.733, + 0.47, + 0.794 + ], + "angle": 0, + "content": "- We propose a continuous-time framework which takes the sequence of 2D skeletons extracted from a multi-person video as the input to forecast the multi-person 3D poses in any future time." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.802, + 0.471, + 0.863 + ], + "angle": 0, + "content": "- We design our model as a hierarchical structure to reduce the computational complexity for multi-person scenarios, and also deal with different connection types in the graph." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We utilize dynamic graph convolutions to deal with the irregular input caused by occlusions." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.733, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "In the remaining of this paper, we will briefly introduce the related works in Sec. 2 and give the details of our model in Sec. 3. The performance of our model is demonstrated through experiments in Sec. 4. Finally, we will conclude our work in Sec. 5. The corresponding source code can be found at https://github.com/SBU-YCX/HDG-ODE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.2, + 0.651, + 0.216 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.226, + 0.892, + 0.243 + ], + "angle": 0, + "content": "2.1. 3D Pose Estimation, Tracking and Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.643 + ], + "angle": 0, + "content": "Single-Person Pose Estimation: Most of the 3D Single-Person Pose Estimation (SPPE) methods are proposed to recover the 3D poses from 2D ones extracted by a pre-trained 2D pose extractor such as Stacked Hourglass Network [51] or Convolutional Pose Machines [69]. Given 2D coordinates as input, Zeng et al. [93] proposed to split the human joints into local groups and estimate them separately. The same idea was used by Chen et al. [15], although they directly estimated 3D poses from raw images and applied different architectures for different parts. In Li et al.'s work [39], different subsets were exchanged to generate augment data for better training. Choi et al. [18] and Azizi et al. [5] used spectral graph convolutions to tackle the problem due to the structure property of input skeletons. Since images captured from monocular cameras have perspective limitations, Yu et al. [91] proposed perspective crop layers to eliminate the perspective effect, while Nie et al. [52] tried to use siamese architectures to decouple the view-dependent representation and pose-dependent representation. There are also some other works solved the perspective problems with the help of multi-view images, such as [50, 60, 80, 72, 94], by fusing the 2D poses from different views to get the original 3D results. Besides, other wearable sensors were used to complement the missing depth information in [94, 27], and Isogawa et al. [31] proposed to utilize an optical non-line-of-sight imaging system to obtain the 3D poses from photon images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Multi-Person Pose Estimation: Inspired by 2D Multi-Person Pose Estimation (MPPE) methods, the works for the 3D estimation task can also be divided into two groups. One is top-down methods where a human detector is applied first and estimation is made for each person separately. Yang et al. [88] followed this line of work with graph convolutional networks (GCNs) [32], T. Xu and W. Takano [83] modified the stacked hourglass structure also with the graph convolutions. Li et al.'s [37] proposed a hierarchical architecture, which is the one closest to our work, but their goal was to augment the graph with additional labeled mesh data and integrate more information, which is totally different from our motivation. The other one is bottom-up approaches where joint locations are directly estimated, then associated and grouped into different people. In [24], Fabbri et al. proposed to use distance-based heuristic for the associations from the detected head joint with the highest confidence." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14701" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.197 + ], + "angle": 0, + "content": "Cheng et al. [17] proposed to integrate both top-down and bottom-up branches to give better results. Besides, inspired by YOLO [59], some works [97, 6, 35, 49] explored to obtain the 3D poses in a single pass where a pivot joint was detected as well as the offsets for all other joints. Similar to single-person case, multi-view sources were also used in [25, 10, 70, 40, 28]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.2, + 0.47, + 0.56 + ], + "angle": 0, + "content": "Pose Tracking and Forecasting: Since Andriluka et al. proposed PoseTrack [1] to first time introduce the Human Pose Tracking (HPT) task, contextual information has begun to be used for better estimating human poses in a sequence of inputs. Similar to other 3D pose tasks, most of works used extracted 2D pose sequence as the input and adopted top-down pipeline to mainly focus on single-person's pose tracking. Cai et al. [9] integrated spatio-temporal GCN into a stacked hourglass architecture to get the information of different scales. The same idea was applied in Wang et al.'s work [75], where the U-shaped GCN was used to extract multi-scale contexts. Liu et al. [41] proposed a graph attention spatio-temporal network with dilated temporal convolutions to get both local and global information, and Sofianos et al. [64] built a space-time-separable GCN for the 3D pose forecasting. Encoder-decoder architecture was utilized in [36, 63, 26] and Zheng et al. [98] solved the problem with spatio-temporal transformer. Furthermore, Chen et al. [11] proposed a fast method with multi-view inputs and Yuan et al. [92] utilized physics simulators and the reinforcement learning strategy in their work. Besides the skeleton joint representations, bone lengths and directions were used in [13, 82] and the extracted information was propagated through an LSTM." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.574, + 0.3, + 0.59 + ], + "angle": 0, + "content": "2.2. Graph Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.6, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Different from traditional convolutional neural networks (CNNs), graph neural networks (GNNs) are mainly proposed to deal with data having more general structures. GNNs can be divided into two groups, spatial and spectral graph convolutions, as described in [8]. The former one considers the graph as locally connected and the information is only shared between each node and its nearest neighbors. J. Atwood and D. Towsley [4] extended it to allow the communications among nodes within \\(k\\)-hop (\\(k >= 1\\)) connections by using the probabilistic transitions. The latter one converts the data into spectral domain first, multiplies it with the Laplacian of the topology matrix in the spectral domain to mimic the convolutional operation in the spatial domain. Compared with spatial graph convolutoins, spectral operations utilize the global graph structure each time. However, since the kernel of the whole graph is large, to reduce the computational complexity, Defferrard et al. [23] proposed to approximate it with the \\(k^{th}\\) order polynomials of the Laplacian with the help of Chebyshev expansion to model the \\(k\\)-hop connections. T. Kipf and M. Welling [32] proposed graph convolutional" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "networks (GCNs) to further simplify it with the \\(1^{st}\\) order approximation of the covolution in ChebNet [23]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.407 + ], + "angle": 0, + "content": "In the pose estimation tasks, GCNs were widely used in [88, 83, 37, 9, 75, 41]. However, unlike CNNs, traditional GCNs have shared weights for different neighbors. To tackle this problems, another learnable weight matrix was trained in [96, 21, 42]. But, in these works, the weight matrix were learned with full joints and the final matrix used for convolution was the multiplication between the weight matrix and the static skeleton matrix assuming all joints observable. In our work, to deal with the missing case caused by occlusion, we make the skeleton matrix dynamic according to the current occlusion status so that the learnt weight matrix can also be effective even though only partial joints given. Some other works [43, 58, 84] proposed to add more semantic connections to the graph, which increased the calculation complexity of the models since the joints of all people could not be fully-connected. In addition, the semantic graphs were still globally sparse although locally dense. In this work, we propose to decompose the whole graph in a hierarchical manner." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.416, + 0.756, + 0.432 + ], + "angle": 0, + "content": "2.3. Neural Differential Equation" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.44, + 0.892, + 0.815 + ], + "angle": 0, + "content": "As mentioned earlier, almost all the existing pose tracking methods were built based on spatio-temporal graph convolutional networks (ST-GCNs) and recurrent neural networks (RNNs), which could only model the discrete-time data where all human joints were assumed to be observable although there existed missing and inaccurate inputs due to occlusions. In order to better model the continuous-time process of the irregular data, neural ordinary differential equation (NeuralODE) was proposed by Chen et al. [12], where deep neural networks were utilized to parameterize a nonlinear ordinary differential equation (ODE). Rubanova et al. [61] and Brouwer et al. [22] further extended it by introducing a recurrent structure to efficiently encode the input information into the underlying trajectories. Some other works [44, 45, 46] also inserted Brownian terms into ODE to make it a stochastic model that can capture the uncertainties in the dynamics. In these works, pose estimation was also mentioned as a part of experiment to prove their efficiency. However, when they dealt with the human joint data, they merely considered each joint sequence as an isolated trajectory and ignored the relations among these joints. Graph neural networks were introduced to parameterize the ODEs in [56, 29, 57], but all of them utilized the whole graph, which didn't take the sparsity problem into consideration in multi-person scenarios." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.828, + 0.892, + 0.862 + ], + "angle": 0, + "content": "3. Hierarchical Dynamic-Graph Ordinary Differential Equation Model" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In this section, we will introduce the details of our hierarchical ordinary differential equation model with dynamic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14702" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "graph convolutions, which we call as HDG-ODE. As shown in Figure 1, the framework of our model follows a hierarchical structure, where each level is a sub-graph with the state changes modeled through graph ordinary differential equation and parameterized with dynamic graph convolutions. In the remaining of this section, we will describe the details of each component as well as the advantages of such designs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.209, + 0.375, + 0.224 + ], + "angle": 0, + "content": "3.1. Problem Statement and Notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.474, + 0.414 + ], + "angle": 0, + "content": "Given a multi-personal video containing \\(K\\) people, and the 2D joints sequence up to time \\(t\\) is \\(\\mathcal{X} = \\{\\mathbf{X}_1,\\mathbf{X}_2,\\dots,\\mathbf{X}_t\\}\\), where \\(\\mathbf{X}_t\\in \\mathcal{R}^{[K,N_j,2]}\\) and \\(N_{j}\\) is the number of joints within each person. These 2D representations can be obtained from any existing 2D pose estimator. Our goal is to predict the corresponding 3D poses \\(\\mathbf{Y}_{\\tau}\\in \\mathcal{R}^{[K,N_j,3]}\\) at a future time \\(\\tau\\) from \\(\\mathbf{X}_t\\), where \\(\\tau\\) can be any value within \\((t,t + 1]\\). Like in many 2D estimators, the missing or occluded joints are represented as some singular values or estimated with a very low confidence, based on this, we can directly obtain the corresponding occlusion mask \\(\\mathbf{M}_t\\) along with \\(\\mathbf{X}_t\\) at each time \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.345, + 0.44 + ], + "angle": 0, + "content": "3.2. Hierarchical Graph Structure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.473, + 0.524 + ], + "angle": 0, + "content": "As explored in many other works, compared to the traditional use of CNN, graph convolution is more suitable for human skeleton data due to the irregular connection of joints. The feature matrices \\( Z^{(l)} \\) and \\( Z^{(l + 1)} \\) of two adjacent convolution layers are expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.535, + 0.47, + 0.553 + ], + "angle": 0, + "content": "\\[\nZ ^ {(l + 1)} = \\sigma (W Z ^ {(l)} \\tilde {A}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.473, + 0.824 + ], + "angle": 0, + "content": "where \\( W \\) is a learnable parameter matrix, \\( \\tilde{A} \\) is the normalized form of the adjacency matrix \\( A \\in [0,1]^{N \\times N} \\) and \\( N \\) is the number of nodes in the graph. However, when modeling human skeletons, we notice that each joint node is only connected to a few other joints, which will make the static skeleton-based \\( \\tilde{A} \\) very sparse and lead to plenty of meaningless calculations in (1). Although some recent works tried to learn more semantic connections among different joints besides static ones, the sparsity problem cannot be eliminated. The semantic relationship among joints usually exists within a limited number of hops instead of the whole graph. Even after we add those semantic edges, we can separate all joints into several groups such that the connections within each group are dense but the ones among different groups are still sparse. This inspires us to adopt a hierarchical structure instead of a flatten one to represent the graph of human skeletons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Besides, according to the physical property of human motion, different joints change their positions at different speeds over time. The location of the central joint usually has a slow and smooth change while the extremity joints, such as \"wrist\" and \"ankle\", have drastic and unexpected" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.182 + ], + "angle": 0, + "content": "changes. Therefore, rather than allocating the same portion of computational resource to all joints, we can focus more on the latter ones. The cascade processing along the hierarchical structure also allows us to utilize the relative-central ones' locations to guide the estimation of positions of the relative-extreme joints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.895, + 0.363 + ], + "angle": 0, + "content": "Another reason for which we adopt the hierarchical structure is its advantage of dealing with different types of edges in the whole graph. For example, in the multi-person scenario, previous works commonly used top-down methods where each person in the scene would be processed individually, but actually, pose estimation and tracking can benefit from taking the relation among people into consideration. However, the edges describing the inter-person relations are different from those representing the intra-person limbs, and it will reduce the effect if we model them in one large graph. Intuitively, separating them into different groups and forming a graph at each level will be a better choice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.894, + 0.47 + ], + "angle": 0, + "content": "Specifically, as shown in Figure 2, for each frame of the video at time \\( t \\), we get a set of 2D skeleton coordinates of all the people in the scene as input. To apply the hierarchical pipeline, we choose the \"pelvis\" joint as the root node of each person, whose coordinate can be also considered as the location of the person. All these root nodes form the 1st-level graph and encoded by a graph encoder" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.476, + 0.892, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {\\tau} ^ {(1)} = G _ {1} \\left(\\mathbf {X} _ {t} ^ {(1)}, \\mathbf {H} _ {< t} ^ {(1)}, \\tilde {A} _ {t} ^ {(1)}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.895, + 0.752 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_t^{(1)}\\in \\mathcal{R}^{[K,1,2]}\\) represents the 2D coordinates of all root nodes of \\(K\\) people, \\(\\mathbf{Y}_{\\tau}^{(1)}\\in \\mathcal{R}^{[K,1,3]}\\) is the corresponding 3D one forecasted for the future time \\(\\tau\\), \\(\\mathbf{H}_{< t}^{(1)}\\) is the historical feature, and \\(\\tilde{A}_t^{(1)}\\) is the normalized adjacency matrix with the edges calculated based on the distances among connected roots, i.e. the relations among people. There is an edge between two people if their root nodes' distance is smaller than a threshold (the average of the heights of joints in 2D skeleton, i.e. distance from \"head\" to \"ankle\" joint in our setting), which indicates they are related. Then, all the limb-root joints, including \"neck\", \"left shoulder\", \"right shoulder\", \"left hip\" and \"right hip\", as well as the root node are grouped to form the 2nd-level graph within each person. To utilize the result of the 1st-level graph, the output of this level is designed as the spatial shifts of the limb-root joints' locations from the root node. We can express the process as below" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.757, + 0.851, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\tau} ^ {(2, i)} = G _ {2} \\left(\\mathbf {X} _ {t} ^ {(2, i)}, \\mathbf {Y} _ {\\tau , i} ^ {(1)}, \\mathbf {H} _ {< t} ^ {(2, i)}, \\tilde {A} _ {t} ^ {(2, i)}\\right), i \\in [ 1, K ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.778, + 0.892, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {\\tau} ^ {(2, i)} = \\mathbf {D} _ {\\tau} ^ {(2, i)} + \\mathbf {Y} _ {\\tau , i} ^ {(1)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_t^{(2,i)}\\in \\mathcal{R}^{[1,6,2]}\\) is the 2D coordinates vector of all limb-root nodes as well as root node of person \\(i\\) \\(\\mathbf{Y}_{\\tau ,i}^{(1)}\\) is the 3D root node's position of this person obtained by (2), \\(\\mathbf{D}_{\\tau}^{(2,i)}\\) and \\(\\mathbf{Y}_{\\tau}^{(2,i)}\\) represent respectively the future 3D position shift and final position predicted. Finally, we group each limbroot joint to the corresponding limb-extreme joint to form" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14703" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.09, + 0.852, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 1. The framework of HDG-ODE. \\(\\mathbf{X}_{t_n}\\) are the 2D joints coordinates at time \\(t_n\\), \\(\\mathbf{M}_{t_n}\\) is the occlusion mask obtained from \\(\\mathbf{X}_{t_n}\\). \\(\\mathbf{Y}_{\\tau}\\) are the predicted 3D coordinates at the future time \\(\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.469, + 0.424 + ], + "angle": 0, + "content": "the 3rd-level graph. Similarly, the encoding process within each lamb group \\( j \\) of person \\( i \\) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.426, + 0.449, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\tau} ^ {(3, i, j)} = G _ {3} (\\mathbf {X} _ {t} ^ {(3, i, j)}, \\mathbf {Y} _ {\\tau , j} ^ {(2, i)}, \\mathbf {H} _ {< t} ^ {(3, i, j)}, \\tilde {A} _ {t} ^ {(3, i, j)}), j \\in [ 1, 5 ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.446, + 0.47, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {\\tau} ^ {(3, i, j)} = \\mathbf {D} _ {\\tau} ^ {(3, i, j)} + \\mathbf {Y} _ {\\tau , j} ^ {(2, i)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.587 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_t^{(3,i,j)}\\in \\mathcal{R}^{[1,3,2]}\\) is the vector of the 2D coordinates representing all nodes of person \\(i\\) within the limb group \\(j\\), \\(\\mathbf{Y}_{\\tau ,j}^{(2,i)}\\) is the 3D location of the lamb-root \\(j\\) calculated by (3), \\(\\mathbf{H}_{< t}^{(3,i,j)}\\) and \\(\\tilde{A}_t^{(3,i,j)}\\) represent the corresponding part of the historical feature and the adjacency matrix in this level. 3D shift \\(\\mathbf{D}_{\\tau}^{(3,i,j)}\\) is used to get the final 3D coordinates \\(\\mathbf{Y}_{\\tau}^{(3,i,j)}\\) at future time \\(\\tau\\)." + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.595, + 0.272, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.755, + 0.209, + 0.764 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.594, + 0.42, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.754, + 0.356, + 0.764 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.768, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Figure 2. Illustration of the hierarchical graph structure: (a) the skeleton representations of a multi-person scenario; (b) the corresponding 3-level hierarchical structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.449, + 0.849 + ], + "angle": 0, + "content": "3.3. Time-Varying Dynamic Graph Convolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "A most important impact factor to the performance of human pose estimation is the occlusion, no matter the occlusion is caused by other objects or the self-occlusion happens" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.895, + 0.576 + ], + "angle": 0, + "content": "within a person. To get the 2D skeleton inputs with some 2D pose estimation methods, the visible joints can be directly extracted from the raw images while the occluded ones are missing. Even though the missing joints can be estimated by some intuition, they are less accurate due to the uncertainties. When we use such 2D poses as inputs to recover the corresponding 3D poses, called the lifting process, the inaccuracy will accumulate and compromises data processing during the information communication within the graph convolution operation. Therefore, we hope our graph encoder can be reasonably adaptive to deal with the missing and inaccurate cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.893, + 0.744 + ], + "angle": 0, + "content": "In the literature work that apply the traditional graph convolution for this task, the adjacency matrix \\( A \\) and its normalized form \\( \\tilde{A} \\) in (1) usually keep constant to represent the static connections within the graph, i.e. all the joints are assumed to be fully available, no matter they are observable or occluded. However, as we described, only those visible ones are accurate and worth being integrated to guide the forecasting. So we propose to use the time-varying dynamic graph convolution to only encode visible joints input at each timestamp. Inspired by some literature work [96], the learning of a hybrid adjacency-weight matrix may be written as" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.765, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\nZ _ {t} ^ {(l + 1)} = \\sigma \\left(W Z _ {t} ^ {(l)} \\rho \\left(W _ {\\tilde {A}} \\odot \\tilde {A}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\( W_{\\tilde{A}} \\) is a learnable weight matrix applied to assign the information along different edges with different weights to solve the weight-sharing problem [21, 42] and \\( \\rho(\\cdot) \\) is a nonlinear function. However, we notice that in (5), \\( W_{\\tilde{A}} \\) is learnt based on the fully-observable data and the \\( \\tilde{A} \\) is the static one without taking the occlusion cases into consideration. In other words, although the learnt weights work well when all" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.075, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "joints are available, they may not be effective if partial of joints are occluded and missing. Based on this, we make a further step and propose to also take the occlusion status into consideration when obtaining the dynamic weight matrix. Specifically, at each timestamp \\( t \\), we will drop the occluded joints in the original static adjacency matrix \\( A \\) to form a time-varying adjacency matrix \\( A_{t} \\), and this dynamic adjacency matrix will be used in both training and testing, i.e. (5) can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.233, + 0.471, + 0.254 + ], + "angle": 0, + "content": "\\[\nZ _ {t} ^ {(l + 1)} = \\sigma \\left(W Z _ {t} ^ {(l)} \\rho \\left(W _ {\\tilde {A}} \\odot \\tilde {A} _ {t}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.473, + 0.37 + ], + "angle": 0, + "content": "where \\(\\tilde{A}_t\\) is the normalized form of the real adjacency matrix at time \\(t\\). By doing so, the learning of the weight matrix \\(W_{\\tilde{A}}\\) will only focus on the observable joints to make sure it can also be effective when only partial joints are available. In the testing phase, the hybrid-matrix \\(W_{\\tilde{A}} \\odot \\tilde{A}_t\\) can ignore the missing joints and adaptively adjust the weights along existing edges for information integration among joints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.378, + 0.42, + 0.394 + ], + "angle": 0, + "content": "3.4. Parallel Ordinary Differential Equation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.471, + 0.506 + ], + "angle": 0, + "content": "Besides the partial occlusion, which leads to incomplete graph input at each frame and can be regarded as spatial missing, there are also many cases where a person is totally occluded, and causes the temporal missing within a sequence. Previous works almost rely on the discrete-time sequential models, which are not flexible enough since the missing in real world can vary from several frames to dozens of frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.472, + 0.552 + ], + "angle": 0, + "content": "To make the model more robust, we adopt ordinary differential equation (ODE), a continuous-time model in this paper, which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.559, + 0.312, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\frac {d \\mathbf {H} _ {t}}{d t} = F (\\mathbf {H} _ {t}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.591, + 0.47, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {t} = \\mathbf {H} _ {t _ {n - 1}} + \\int_ {t _ {n - 1}} ^ {t} F (\\mathbf {H} _ {\\tau}) d \\tau . \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.634, + 0.47, + 0.724 + ], + "angle": 0, + "content": "Compared with traditional recurrent networks, it can be used to predict the values over the interval of any length. Specially, in our model, we use time-varying dynamic graph convolution mentioned in Sec. 3.3 to parameterize the derivatives \\( F(\\cdot) \\). For efficiency, we compute the integration in (7) by Euler Method and we can rewrite it as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.733, + 0.47, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {t, F} = \\mathbf {H} _ {t - \\Delta t, F} + F \\left(\\mathbf {H} _ {t - \\Delta t, F}, \\tilde {A}\\right) \\Delta t, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.472, + 0.837 + ], + "angle": 0, + "content": "where \\(\\mathbf{H}_{t,F}\\) is the hidden feature, \\(\\tilde{A}\\) is the normalized adjacency matrix assuming all joints observable, and \\(\\Delta t\\) is the step-size. At an observation time \\(t_n\\), the features of observable joints are updated through the same graph convolution encoder as described in Sec. 3.2, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.845, + 0.47, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {t _ {n}, G} = G \\left(\\mathbf {X} _ {t _ {n}}, \\mathbf {Y} _ {t _ {n}}, \\mathbf {H} _ {< t _ {n}}, \\tilde {A} _ {t _ {n}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(G = \\{G_1, G_2, G_3\\}\\) represents the cascade process in (2)-(4) parameterized by our time-varying dynamic graph" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.169 + ], + "angle": 0, + "content": "convolution networks, \\(\\mathbf{X}_{t_n}\\) and \\(\\mathbf{Y}_{t_n}\\) are the 2D and 3D joint locations, \\(\\mathbf{H}_{< t_n} = \\mathbf{H}_{t_n - \\Delta t,F}\\) is the feature from the ODE model and \\(\\tilde{A}_{t_n}\\) is the normalized adjacency matrix of the graph formed by observable joints. After the update, the new feature at the time instant \\(t_n\\) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.181, + 0.894, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {t _ {n}} = \\mathbf {M} _ {t _ {n}} \\odot \\mathbf {H} _ {t _ {n}, F} + \\left(\\mathbf {1} - \\mathbf {M} _ {t _ {n}}\\right) \\odot \\mathbf {H} _ {t _ {n}, G}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.894, + 0.33 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{t_n}\\) is the occlusion mask of the input described in Sec. 3.1. Here, we want to emphasize that \\(\\tilde{A}_{t_n}\\) in (9) and \\(\\tilde{A}\\) in (8) may not be the same according to the observation states of input. As we described in last section, the input may contain missing values, but all the missing values have been completed by the encoder in the encoding process and we can assume the graph is fully observable in the propagation process." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.345, + 0.86, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.463, + 0.846, + 0.478 + ], + "angle": 0, + "content": "Figure 3. Illustration of the parallel ODE pipeline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.599 + ], + "angle": 0, + "content": "Here, we can further demonstrate the advantages of hierarchical structure. For efficiency, we adopt Euler method to approximately calculate the integration, but as we know, it will predict the curve as a straight line if we make the propagation step \\(\\Delta t\\) too large and equal to the interval between two observations as \\(t_n - t_{n-1}\\). To make the model more expressive, we usually make \\(\\Delta t\\) smaller, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.609, + 0.892, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\Delta t = \\frac {1}{k} \\left(t _ {n} - t _ {n - 1}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.894, + 0.74 + ], + "angle": 0, + "content": "However, as we discussed in Sec. 3.2, different joints of human body have different rates of location changes, we don't need to set \\( k \\) to the same value at different levels in our model. Instead, we can adopt a parallel pipeline and make it small in shallow levels while enlarge it as the level goes deeper as shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.753, + 0.634, + 0.77 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.697, + 0.794 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.614, + 0.815 + ], + "angle": 0, + "content": "4.1.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "MuPoTS-3D[49]: This multi-person dataset contains sequences captured in 20 real scenes. Each sequence has up to three subjects, as well as provides raw images, 2D pose annotations and the corresponding 3D ones. The lengths of data vary from 125 frames to 800 frames, and we split them" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14705" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.196 + ], + "angle": 0, + "content": "into segments with the length of 50 frames. We divide all the data into training, validation and testing sets with the ratio of 0.8: 0.1: 0.1. The real occlusion mask are given as well, and the final data are obtained by the element-wise multiplication between the full data and the occlusion masks. Before being fed into the models, data from each sequence are first normalized." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.197, + 0.473, + 0.318 + ], + "angle": 0, + "content": "Human3.6M[30]: This dataset consists of 3.6 million of single-person frames. 11 subjects conducting 15 activities are captured by cameras. Both 2D and 3D annotations are provided where 2D ones come from cameras while 3D ones are obtained using a motion capture system. Different from commonly used training methods, we only use data from one camera to train the model. We performed the same data partition and normalization as above." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.337, + 0.266, + 0.352 + ], + "angle": 0, + "content": "4.1.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.47, + 0.403 + ], + "angle": 0, + "content": "Mean Per Joint Position Error (MPJPE): This metric measures the average position error over all joints and frames in the sequence, which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.411, + 0.47, + 0.445 + ], + "angle": 0, + "content": "\\[\nM P J P E = \\frac {1}{N _ {f}} \\sum_ {f} \\frac {1}{N _ {j} ^ {(f)}} \\sum_ {j} \\left\\| P _ {\\text {p r e d}} ^ {(f)} (j) - P _ {g t} ^ {(f)} (j) \\right\\| _ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.456, + 0.47, + 0.522 + ], + "angle": 0, + "content": "where \\(N_{f}\\) is the total number of frames, \\(N_{j}^{(f)}\\) is the total number of joints in the frame \\(f\\), \\(P_{gt}^{(f)}(j)\\) and \\(P_{pred}^{(f)}(j)\\) are the ground-truth and the predicted 3D position of the joint \\(j\\) in the frame \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.472, + 0.621 + ], + "angle": 0, + "content": "Percentage of Correct Keypoints (PCK): PCK calculates the percentage of correctly predicted key-points, where a key-point \\( j \\) in the frame \\( f \\) is considered to be correct if the distance \\( d_{p}^{(f)}(j) \\) between the 3D ground-truth \\( P_{p,gt}^{(f)}(j) \\) and the predicted 3D position \\( P_{p,pred}^{(f)}(j) \\) are smaller than a person-specific threshold \\( S_{p}^{f} \\), i.e." + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.63, + 0.47, + 0.668 + ], + "angle": 0, + "content": "\\[\nP C K = \\frac {\\sum_ {f} \\sum_ {p} \\sum_ {j} \\delta \\left(d _ {p} ^ {(f)} (j) < S _ {p} ^ {(f)}\\right)}{\\sum_ {f} \\sum_ {p} \\sum_ {j} 1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.47, + 0.735 + ], + "angle": 0, + "content": "Here we adopt the setting in MPII dataset[2] where the threshold is the ratio of head link's length \\( h_p^{(f)} \\) of the person \\( p \\), i.e. the distance between \"head\" joint and \"upper neck\", and the PCK can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.744, + 0.47, + 0.79 + ], + "angle": 0, + "content": "\\[\nP C K h @ \\alpha = \\frac {\\sum_ {f} \\sum_ {p} \\sum_ {j} \\delta \\left(\\frac {d _ {p} ^ {(f)} (j)}{h _ {p} ^ {(f)}} < \\alpha\\right)}{\\sum_ {f} \\sum_ {p} \\sum_ {j} 1}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.798, + 0.281, + 0.813 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the ratio parameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.295, + 0.847 + ], + "angle": 0, + "content": "4.1.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.9 + ], + "angle": 0, + "content": "As introduced in Sec. 3.2, the hierarchical graph has 3 levels, containing \\( K \\), 6 and 3 nodes each, and our HDG-ODE model has a hierarchical structure of 3 levels, with an encoder and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "an ODE at each level. For each encoder, we use a single-layer graph-GRU, using our time-varying dynamic graph convolutions with the feature size of 6. ODE models are parameterized with graph neural networks of 1, 2 and 3 convolutional layers at the three levels respectively, with the hidden feature size of 16. The propagation steps for the three levels are respectively \\([1/30s, 1/60s, 1/120s]\\), and the data sampling interval is \\(1/30s\\). The experiments are conducted on single NVIDIA 1080 Ti GPU. During the training, we use the Adam() optimizer with the learning rates of [0.01, 0.001, 0.0001] to minimize the mean squared error (MSE) between the predicted 3D positions and the ground-truth of the visible joints, and 100 epochs are run at each learning rate. To speed up the training, it will be early-stopped if there is no further loss decrease in 20 epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.328, + 0.688, + 0.344 + ], + "angle": 0, + "content": "4.2. Experiment Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.351, + 0.702, + 0.366 + ], + "angle": 0, + "content": "4.2.1 Overall Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.895, + 0.48 + ], + "angle": 0, + "content": "We conduct experiments over MuPoTS-3D dataset to evaluate the performance of our proposed model on the multiperson 3D pose tracking task. For comparison, we choose some representative methods, including discrete-time recurrent models and continuous-time ones, as the baselines. Specifically we take the following methods into consideration:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.491, + 0.895, + 0.628 + ], + "angle": 0, + "content": "- STGCN: This discrete-time model consists of several concatenated spatio-temporal convolutional blocks (ST-Conv blocks), which is a common structure that is widely used for action recognition [85], pose forecasting [75] and other prediction tasks [89]. In our experiments, STGCN has 2 ST-Conv blocks, and recurrently processes each clip within the whole 2D pose sequence and give the 3D predictions. The output dimension of each layer in the blocks is 32." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.638, + 0.894, + 0.73 + ], + "angle": 0, + "content": "- Graph-GRU: It is a variant of conventional Gated Recurrent Unit (GRU), where the linear operation in the conventional GRU is replaced by graph convolution. It works as the backbone in [90]. We construct the baseline with 2 Graph-GRU layer and a linear decoder. The output dimension of each layer in the blocks is 32." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.739, + 0.894, + 0.829 + ], + "angle": 0, + "content": "- SemGCGRU: It is an improved version of traditional Graph-GRU, where the fixed binary adjacency matrix is replaced by a learnable weight matrix to capture the relations among nodes [96]. In our implementation, the structure and the feature size are the same as Graph-GRU." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- ODE-RNN: In this model [62], the deep learning modules, specifically dense neural networks, are incorporated to parameterize a non-linear ordinary differential equation (ODE) to model the continuous-time process" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.491, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14706" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "of data sequence. In our experiments, the dense network contains 2 linear layers and the dimension of hidden features is 32." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.146, + 0.472, + 0.283 + ], + "angle": 0, + "content": "- Graph-ODE: It is a GNN variant of ODE-RNN, with a GCN to learn the derivative of ODE and a GRU to integrate the information of observations. In our model comparison, the GCN has 2 hidden layers with ReLU activation functions and the output dimension of each hidden layer is 32. The output dimension of GRU is 32 as well. In other words, to compare with discrete models, we just replace the GRU module in Graph-GRU with an ODE module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.473, + 0.4 + ], + "angle": 0, + "content": "Although continuous-time models can provide predictions at any time, even smaller than the data sampling rate, we set the step of them to be equal to the original data interval used in discrete-time models for fair comparison. From the results in Table 1, we can see that our HDG-ODE can achieve the best result with the smallest mean prediction error and the highest percentage of corrected prediction." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.412, + 0.487, + 0.59 + ], + "angle": 0, + "content": "
MPJPE(↓)PCKh@α(↑)
α=0.1α=0.5α=1.0
Discrete
STGCN [89, 75]0.600413.03%60.35%77.76%
Graph-GRU [90]0.557314.86%62.39%80.04%
SemGCGRU [96]0.437517.09%69.61%85.96%
Continuous
ODE-RNN [61]0.439620.21%71.78%85.45%
Graph-ODE [56]0.406619.80%71.57%87.00%
Ours
HDG-ODE0.303829.72%78.67%92.18%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.59, + 0.472, + 0.603 + ], + "angle": 0, + "content": "Table 1. Testing Performance of different models on MuPoTS-3D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.238, + 0.651 + ], + "angle": 0, + "content": "4.2.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.795 + ], + "angle": 0, + "content": "In this part, we will separately demonstrate the effectiveness of each designing component in our proposed model by some experiment results, including the continuous ODE model, the pipeline taking all people into consideration, the new time-varying dynamic graph convolution and the proposed hierarchical structure. The evaluation is sequential by adding the components one by one, and each time we only choose the current best model to conduct the experiment for the next component." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Discrete-Time Model vs. Continuous-Time Model: From Table 1, compared to traditional discrete-time models, no matter our HDG-ODE or other continuous-time models have a better performance. Different from discrete models which only execute one propagation between two observations, the continuous-time ones conduct multiple-step propagation so that the information on each joint will be affected" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.181 + ], + "angle": 0, + "content": "and corrected by its neighbors several times before getting the final prediction. Also, due to the property of ODE, the continuous-time model can provide predictions at any time in the interval between two observations, instead of just obtaining predictions at the observation instants as done by discrete-time ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.897, + 0.334 + ], + "angle": 0, + "content": "Isolated Top-Down Pipeline vs. Relational Top-Down Pipeline: As we described previously, most of existing multiperson pose tracking works adopt the top-down processing pipeline, where each person in the scenario is tracked separately as an isolated trajectory. Instead, we take into consideration the relation among people to infer their relative movements, which in turn affects the future positions of specific joints. In Table 2, by comparing the models in isolated top-down pipeline with the ones in our relational pipeline, we can observe the improvement." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.34, + 0.899, + 0.419 + ], + "angle": 0, + "content": "
Isolated PipelineRelational Pipeline
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.610477.72%0.600477.76%
Graph-GRU0.573878.46%0.557380.04%
Graph-ODE0.420386.89%0.406687.00%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.419, + 0.894, + 0.446 + ], + "angle": 0, + "content": "Table 2. Comparison between the isolated top-down pipeline and our relational top-down pipeline on MuPoTS-3D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.618 + ], + "angle": 0, + "content": "Static Graph Convolution vs. Time-Varying Dynamic Graph Convolution: In the real cases, occlusion is inevitable. To better tackle the problem caused by the input data with missing or inaccurate values due to occlusions, we propose to replace the static graph convolution used in most of works with our time-varying dynamic graph convolution in the HDG-ODE model. As shown in Table 3, the dynamic graph convolution helps eliminate the negative influence of inaccurate inputs as well as adjust the integration weights of edges to the most suitable ones according to occlusion states, so the models can have a better performance than before." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.625, + 0.899, + 0.692 + ], + "angle": 0, + "content": "
Original Graph ConvOur Graph Conv
MPJPEPCKh@1.0MPJPEPCKh@1.0
Graph-GRU0.557380.04%0.384486.45%
Graph-ODE0.406687.00%0.370688.74%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.692, + 0.894, + 0.731 + ], + "angle": 0, + "content": "Table 3. Comparison between original static graph convolutions (GC) and our time-varying dynamic convolutions (DG) on MuPoTS-3D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Flatten Structure vs. Hierarchical Structure: To solve the sparsity problem explained in Sec. 3, our HDG-ODE adopts a hierarchical structure. In Table 4, compared to models in the flatten structure with a large sparse graph, the corresponding hierarchical ones perform better for models with graph convolutions. The reasons leading to the improvement lie in two aspects: On one hand, by decomposing the whole sparse graph into a set of relatively dense ones, the parameters to learn are reduced, as shown in Table 5. This helps save a significant amount of computational resource, and the released resource can be reallocated to the joints" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14707" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "with more complicated movements, e.g. making the corresponding network deeper or reducing the propagation step of ODE. On the other hand, the cascade processing along the hierarchical structure can provide the guidance among the levels. As for the ODE-RNN, the hierarchical one has a worse result because the structure is mainly proposed to deal with the graph sparsity problem. However, ODE-RNN doesn't utilize the graph information and only cares about isolated joint trajectories, so the hierarchical structure will make the model more complex and reduce the performance." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.25, + 0.472, + 0.327 + ], + "angle": 0, + "content": "
Flatten StructureHierarchical Structure
MPJPEPCKh@1.0MPJPEPCKh@1.0
ODE-RNN0.439685.45%0.493882.40%
GC-ODE0.406687.00%0.345190.51%
DG-ODE0.370688.74%0.303892.18%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.328, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Table 4. Performance Comparison between the flatten structure and the hierarchical structure on MuPoTS-3D." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.379, + 0.415, + 0.427 + ], + "angle": 0, + "content": "
#Params (↓)Memory (↓)
DG-ODE106490.76M
HDG-ODE26430.66M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.427, + 0.472, + 0.455 + ], + "angle": 0, + "content": "Table 5. Complexity Comparison between the flatten Dynamic-Graph ODE (DG-ODE) and our hierarchical one (HDG-ODE)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.477, + 0.274, + 0.492 + ], + "angle": 0, + "content": "4.2.3 Robustness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.773 + ], + "angle": 0, + "content": "Besides the multi-person scenarios, we also conduct experiments of single-person scenarios on Human3.6M dataset. Different from MuPoTS-3D, no occlusion information is provided within the data, so we make occlusion masks by ourselves and test the robustness of our model under different observable ratios. Specifically, for a sequence, we assume only \\( p_t \\) of the frames are observable while in the remaining ones, the input skeletons are totally missing due to the occlusion by buildings or other large objects. For each observable frame, we assume there exist self-occlusions such that only \\( p_s \\) of the joints are available. To make the corresponding mask, both \\( p_t \\) of the frames and \\( p_s \\) of the joints on each frame are randomly chosen and the total observable ratio can be expressed as \\( p_t \\times p_s \\). The results are shown in Table 6. When the observable ratio decreases as a result of occlusions, the performance gets worse for all models. However, our HDG-ODE always performs the best among all models studied under all circumstances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.196, + 0.801 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a new continuous-time model called HDG-ODE to forecast the future 3D human pose representations in multi-person videos given the 2D pose sequence as input. By using ordinary differential equations and dynamic graph convolutions, we alleviate the occlusion problem brought by perspective effects of monocular images." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.085, + 0.898, + 0.193 + ], + "angle": 0, + "content": "
pt=0.8, ps=0.5pt=0.6, ps=0.4
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.396680.19%0.437675.82%
Graph-GRU0.372383.34%0.418877.39%
ODE-RNN0.352784.33%0.382680.66%
Graph-ODE0.347383.98%0.401978.54%
HDG-ODE0.305288.29%0.354382.36%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.193, + 0.892, + 0.22 + ], + "angle": 0, + "content": "Table 6. Performance of different models on Human3.6M with different observable ratios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.895, + 0.394 + ], + "angle": 0, + "content": "Besides, with the help of hierarchical structure, we reduce the computational complexity of our model and are able to allocate more resources to the human joints which have more complicated movement. Our experiments demonstrate that our model can well recover the corresponding 3D poses from the single-view 2D poses and outperform literature works on the human pose forecasting task. Currently, the maximum number of people in the scenes of the video is limited. In the future, we will make our model more adaptive and test it in more crowded scenarios." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.408, + 0.597, + 0.423 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.433, + 0.895, + 0.488 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. Posetrack: A benchmark for human pose estimation and tracking, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.895, + 0.517 + ], + "angle": 0, + "content": "[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.895, + 0.574 + ], + "angle": 0, + "content": "[3] Bruno Artacho and Andreas Savakis. Unipose: Unified human pose estimation in single images and videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[4] James Atwood and Don Towsley. Diffusion-convolutional neural networks, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[5] Niloofar Azizi, Horst Possegger, Emanuele Rodola, and Horst Bischof. 3d human pose estimation using mobius graph convolutional networks, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.895, + 0.716 + ], + "angle": 0, + "content": "[6] Abdallah Benzine, Florian Chabot, Bertrand Luvison, Quoc Cuong Pham, and Catherine Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.895, + 0.773 + ], + "angle": 0, + "content": "[7] Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, and Lorenzo Torresani. Learning temporal pose estimation from sparsely labeled videos. In Advances in Neural Information Processing Systems 33, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[8] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.895, + 0.9 + ], + "angle": 0, + "content": "[9] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.433, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14708" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. arXiv preprint arXiv:2007.10986, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[11] Long Chen, Haizhou Ai, Rui Chen, Zijie Zhuang, and Shuang Liu. Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[12] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[13] Tianlang Chen, Chen Fang, Xiaohui Shen, Yiheng Zhu, Zhili Chen, and Jiebo Luo. Anatomy-aware 3d human pose estimation in videos. arXiv preprint arXiv:2002.10322, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[14] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[15] Zerui Chen, Yan Huang, Hongyuan Yu, Bin Xue, Ke Han, Yiru Guo, and Liang Wang. Towards part-aware monocular 3d human pose estimation: An architecture search approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 715-732, Cham, 2020. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.47, + 0.472, + 0.538 + ], + "angle": 0, + "content": "[16] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.539, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[17] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7649-7659, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.609, + 0.472, + 0.663 + ], + "angle": 0, + "content": "[18] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[19] Chia-Jung Chou, Jui-Ting Chien, and Hwann-Tzong Chen. Self adversarial training for human pose estimation. In 2018 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 17-30, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.734, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[20] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L. Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.805, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[21] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[22] Edward De Brouwer, Jaak Simm, Adam Arany, and Yves Moreau. Gru-ode-bayes: Continuous modeling of sporadically-observed time series, 2019." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[23] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.137, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[24] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.207, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[25] Qi Fang, Qing Shuai, Junting Dong, Hujun Bao, and Xiaowei Zhou. Reconstructing 3d human pose by watching humans in the mirror. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.25, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[26] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.321, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[27] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.406, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[28] Congzhentao Huang, Shuai Jiang, Yang Li, Ziyue Zhang, Jason Traish, Chen Deng, Sam Ferguson, and Richard Yi Da Xu. End-to-end dynamic matching network for multi-view multi-person 3d pose estimation. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII, page 477-493, Berlin, Heidelberg, 2020. Springer-Verlag." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.504, + 0.894, + 0.573 + ], + "angle": 0, + "content": "[29] Zijie Huang, Yizhou Sun, and Wei Wang. Coupled graph ode for learning interacting system dynamics. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery amp; Data Mining, KDD '21, page 705-715, New York, NY, USA, 2021. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.575, + 0.894, + 0.643 + ], + "angle": 0, + "content": "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.646, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[31] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris M. Kitani. Optical non-line-of-sight physics-based 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.703, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[32] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.732, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[33] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Multiposenet: Fast multi-person pose estimation using pose residual network. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.789, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[34] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[35] Jogendra Nath Kundu, Ambareesh Revanur, Govind Vitthal Waghmare, Rahul Mysore Venkatesh, and R. Venkatesh Babu. Unsupervised cross-modal alignment for multi-person 3d pose estimation, 2020." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.175 + ], + "angle": 0, + "content": "[36] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R. Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.178, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[37] Han Li, Bowen Shi, Wenrui Dai, Yabo Chen, Botao Wang, Yu Sun, Min Guo, Chenlin Li, Junni Zou, and Hongkai Xiong. Hierarchical graph networks for 3d human pose estimation, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[38] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[39] Shichao Li, Lei Ke, Kevin Pratama, Yu-Wing Tai, Chi-Keung Tang, and Kwang-Ting Cheng. Cascaded deep monocular 3d human pose estimation with evolutionary training data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[40] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.434, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[41] Junfa Liu, Juan Rojas, Zhijun Liang, Yihui Li, and Yisheng Guan. A graph attention spatio-temporal convolutional networks for 3d human pose estimation in video. arXiv preprint arXiv:2003.14179, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.491, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[42] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X, page 318–334, Berlin, Heidelberg, 2020. Springer-Verlag." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.576, + 0.472, + 0.672 + ], + "angle": 0, + "content": "[43] Shengyuan Liu, Pei Lv, Yuzhen Zhang, Jie Fu, Junjin Cheng, Wanqing Li, Bing Zhou, and Mingliang Xu. Semi-dynamic hypergraph neural network for 3d pose estimation. In Christian Bessiere, editor, Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI20, pages 782-788. International Joint Conferences on Artificial Intelligence Organization, 7 2020. Main track." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.674, + 0.472, + 0.715 + ], + "angle": 0, + "content": "[44] Xuanqing Liu, Tesi Xiao, Si Si, Qin Cao, Sanjiv Kumar, and Cho-Jui Hsieh. Neural sde: Stabilizing neural ode networks with stochastic noise, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.717, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[45] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, and Zhaoyue Chen. Learning continuous-time dynamics by stochastic differential networks, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[46] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, Zhaoyue Chen, and Jacqueline Wu. Continuous-time stochastic differential networks for irregular time series modeling. In Teddy Mantoro, Minho Lee, Media Anugerah Ayu, Kok Wai Wong, and Achmad Nizar Hidayanto, editors, Neural Information Processing, pages 343-351, Cham, 2021. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[47] Zhenguang Liu, Haoming Chen, Runyang Feng, Shuang Wu, Shouling Ji, Bailin Yang, and Xun Wang. Deep dual consecutive network for human pose estimation. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 525-534, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.124, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13264-13273, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.194, + 0.894, + 0.25 + ], + "angle": 0, + "content": "[49] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.252, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[50] Rahul Mitra, Nitesh B. Gundavarapu, Abhishek Sharma, and Arjun Jain. Multiview-consistent semi-supervised learning for 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.324, + 0.894, + 0.351 + ], + "angle": 0, + "content": "[51] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.354, + 0.894, + 0.395 + ], + "angle": 0, + "content": "[52] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.398, + 0.894, + 0.453 + ], + "angle": 0, + "content": "[53] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.455, + 0.894, + 0.511 + ], + "angle": 0, + "content": "[54] Wanli Ouyang, Xiao Chu, and Xiaogang Wang. Multisource deep learning for human pose estimation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2337-2344, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.513, + 0.894, + 0.595 + ], + "angle": 0, + "content": "[55] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-286, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.599, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[56] Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.643, + 0.894, + 0.698 + ], + "angle": 0, + "content": "[57] Michael Poli, Stefano Massaroli, Clayton M. Rabideau, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Continuous-depth neural models for dynamic graph prediction, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.7, + 0.894, + 0.769 + ], + "angle": 0, + "content": "[58] Zhongwei Qiu, Kai Qiu, Jianlong Fu, and Dongmei Fu. Dgen: Dynamic graph convolutional network for efficient multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 34(07):11924-11931, Apr. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[59] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.829, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[60] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[61] Yulia Rubanova, Ricky T. Q. Chen, and David Duvenaud. Latent odes for irregularly-sampled time series, 2019." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14710" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[62] Yulia Rubanova, Tian Qi Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series. In Advances in Neural Information Processing Systems 32, pages 5321-5331. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.472, + 0.231 + ], + "angle": 0, + "content": "[63] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2484-2494, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.471, + 0.301 + ], + "angle": 0, + "content": "[64] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11209-11218, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[65] Kai Su, Dongdong Yu, Zhenqi Xu, Xin Geng, and Changhu Wang. Multi-person pose estimation with enhanced channelwise and spatial information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[66] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5686-5696, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.428, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[67] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.484, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[68] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose estimation. ArXiv, abs/1911.07451, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[69] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 1653-1660, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.581, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[70] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.637, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[71] Ali Varamesh and Tinne Tuytelaars. Mixture dense regression for object detection and human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.693, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13294-13304, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.763, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[73] Jiahang Wang, Sheng Jin, Wentao Liu, Weizhong Liu, Chen Qian, and Ping Luo. When human pose estimation meets robustness: Adversarial algorithms and benchmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11855-11864, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[74] Jian Wang, Xiang Long, Yuan Gao, Errui Ding, and Shilei Wen. Graph-pcnn: Two stage human pose estimation with graph pose refinement. In The European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.12 + ], + "angle": 0, + "content": "[75] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.121, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[76] Manchen Wang, Joseph Tighe, and Davide Modolo. Combining detection and tracking for human pose estimation in videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.177, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[77] Rui Wang, Zhongzheng Cao, Xiangyang Wang, Zhi Liu, and Xiaoqiang Zhu. Human pose estimation with deeply learned multi-scale compositional models. IEEE Access, 7:71158-71166, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.232, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[78] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[79] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.303, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[80] Rongchang Xie, Chunyu Wang, and Yizhou Wang. Meta-fuse: A pre-trained fusion model for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.359, + 0.894, + 0.398 + ], + "angle": 0, + "content": "[81] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.4, + 0.894, + 0.469 + ], + "angle": 0, + "content": "[82] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.47, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[83] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16105-16114, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.526, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[84] Xixia Xu, Qi Zou, and Xue Lin. Adaptive hypergraph neural network for multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2955-2963, Jun. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.581, + 0.894, + 0.622 + ], + "angle": 0, + "content": "[85] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.623, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[86] Wei Yang, Shuang Li, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Learning feature pyramids for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.679, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[87] Wei Yang, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. End-to-end learning of deformable mixture of parts and deep convolutional neural networks for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.749, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[88] Yiding Yang, Zhou Ren, Haoxiang Li, Chunluan Zhou, Xin chao Wang, and Gang Hua. Learning dynamics via graph neural networks for human pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8074-8084, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.832, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[89] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), 2018." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14711" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[90] Bing Yu, Haoteng Yin, and Zhanxing Zhu. ST-UNet: A spatio-temporal u-network for graph-structured time series modeling. arXiv, abs/1903.05631, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[91] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9064-9073, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.207, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[92] Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. Simpoe: Simulated character control for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7159–7169, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.277, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[93] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Srnet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 507–523, Cham, 2020. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.472, + 0.43 + ], + "angle": 0, + "content": "[94] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[95] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.472, + 0.556 + ], + "angle": 0, + "content": "[96] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N. Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.472, + 0.6 + ], + "angle": 0, + "content": "[97] Jianan Zhen, Qi Fang, Jiaming Sun, Wentao Liu, Wei Jiang, Hujun Bao, and Xiaowei Zhou. Smap: Single-shot multiperson absolute 3d pose estimation, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.472, + 0.669 + ], + "angle": 0, + "content": "[98] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11656-11665, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.672, + 0.472, + 0.753 + ], + "angle": 0, + "content": "[99] Tianfei Zhou, Wenguan Wang, Si Liu, Yi Yang, and Luc Van Gool. Differentiable multi-granularity human representation learning for instance-aware human semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1622-1631, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.756, + 0.47, + 0.797 + ], + "angle": 0, + "content": "[100] Bin Xiao Zhaoxiang Zhang Jingdong Wang Zigang Geng, Ke Sun. Bottom-up human pose estimation via disentangled keypoint regression. In CVPR, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14712" + } + ] +] \ No newline at end of file diff --git a/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_origin.pdf b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac0000740d136f09f8d70fc690926e43565de75b --- /dev/null +++ b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/b32b33ed-c49a-4dab-8f95-5b4dd7695050_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc611a2df5fa2b89699fe87f6dccfdd28c8446bde907bb5690716e51bea1248f +size 949815 diff --git a/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/full.md b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2fadbc27eb9c02ca7715e55c502d2490f10a154b --- /dev/null +++ b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/full.md @@ -0,0 +1,408 @@ +# HDG-ODE: A Hierarchical Continuous-Time Model for Human Pose Forecasting + +Yucheng Xing Xin Wang +Department of Electrical and Computer Engineering, Stony Brook University +{yucheng.xing, x.wang}@stonybrook.edu + +# Abstract + +Recently, human pose estimation has attracted more and more attention due to its importance in many real applications. Although many efforts have been put on extracting 2D poses from static images, there are still some severe problems to be solved. A critical one is occlusion, which is more obvious in multi-person scenarios and makes it even more difficult to recover the corresponding 3D poses. When we consider a sequence of images, the temporal correlation among the contexts can be utilized to help us ease the problem, but most of the current works only rely on discrete-time models and estimate the joint locations of all people within a whole sparse graph. In this paper, we propose a new framework, Hierarchical Dynamic Graph Ordinary Differential Equation (HDG-ODE), to tackle the 3D pose forecasting task from 2D skeleton representations in videos. Our framework adopts ODE, a continuous-time model, as the base to predict the 3D joint positions at any time. Considering the structural-property of the skeleton data in representing human poses and the possible irregularity caused by occlusion, we propose the use of dynamic graph convolution as the basic operator. To reduce the computational complexity introduced by the sparsity of the pose graph, our model takes a hierarchical structure where the encoding process at the observation timestamp is done in a cascade manner while the propagation between observations is conducted in parallel. The performance studies on several datasets demonstrate that our model is effective and can out-perform other methods with fewer parameters. + +# 1. Introduction + +With the progress in the computer vision field, human-related contents have attracted more and more attention, as they are very important in many real-world applications. Some examples are autonomous driving, human-computer interaction and anomaly detection. Among these contents, human poses are the most basic and important ones since their accurate and timely estimation or even forecasting usually serve as the first step for further analyzing complex + +activities. + +Human pose estimation is not a novel topic but has been studied for many years. Its performance, however, is still not satisfactory enough to meet the demands of many practical applications. Since A. Toshev and C. Szegedy proposed DeepPose [69] to represent human poses by the locations of a set of joints from human skeletons and first time utilized the regression method of deep learning to estimate the corresponding 2D coordinates, a lot of efforts [78, 51, 87, 20, 86, 19, 79, 77, 54, 67, 95, 94, 3, 66, 74, 73] have been made to get the accurate estimation in the presence of occlusion resulted from the perspective limitation of images. These methods [14, 33, 55, 65, 34, 68, 38, 53, 16, 71, 58, 48, 99, 100, 84] are extended to multi-person scenarios, where the occlusion problem is more severe. Although some progress has been made, 2D locations cannot perfectly restore the human poses, and are never our final goals for this task. Recently, more and more works have been focusing on 3D pose estimation with raw images [15, 24, 6, 91] or extracted 2D poses [96, 21, 5, 18, 93, 39, 97, 83, 17, 43, 91] as input. Besides the inherent occlusion problems, the lack of depth in monocular images also limits the performance of algorithms. Although some schemes relying on multi-view images [50, 60, 94, 80, 72, 10, 70, 28, 40, 25, 11] or other sensors [54, 31, 27, 94] can ease this problem to some degree, we have to admit that paired images of different views or complementary sensors are not always available in most scenarios. The core problem is how to just utilize limited information in monocular images to extract the corresponding 3D poses of people. If a sequence of images (i.e. video) is available, the temporal context carried can be also helpful for the task [1, 81, 7, 47, 76]. They not only provide information to make up the missing joints caused by the occlusion and provide hints for joints' locations in the future, but also help group the detected joints belonging to a specific person in multi-person scenarios. So in this paper, we mainly focus on restoration from a sequence of monocular multi-person 2D skeletons to predict the corresponding 3D poses at a future time point. + +Some related methods [64, 9, 75, 82, 41, 11, 98, 26, 13, 63, 36] usually applied graph neural networks to encode the + +2D human joints and propagated the information through recurrent neural networks or directly used spatial-temporal graph convolutional networks. However, almost all of them only focused on single-person scenarios, while in most of cases, the relations among people in the scene are also helpful for the forecasting. Also, there exist several limitations in these works. Firstly, traditional recurrent neural networks, such as GRU or LSTM, all use discrete-time sequential models, which have a fixed propagation step and are not flexible enough to deal with different moving speeds of different human joints. The performance of these models will be compromised when the video frequency is not high or the cameras need to scan around in a certain period instead of shooting a fixed scene. For example, if the human pose is recorded every second in the video, discrete-time models cannot provide the forecasting result within 1 second. Instead, we propose to use neural ordinary differential equation (NeuralODE), a continuous-time model, to propagate information along the temporal axis in our framework. Secondly, we notice that any joint of a person is only connected to its nearest neighbors, which makes the adjacency matrix used in the graph convolution very sparse, and leads to inefficient calculations in turn. To reduce the meaningless calculations and increase the efficiency of our model, we propose to utilize a hierarchical structure to decompose a sparsely connected graph into several dense graphs and encode them in a cascade manner. By doing so, we can also deal with edges of different types separately, such as the edges between people and the edges connecting joints within each person. Lastly, the whole and fixed binary adjacency matrix is used for the graph convolution in the literature, although the actual connections of joints only occupy a very small part of the graph due to the occlusion. Instead, we propose to provide dynamic topology information at each time stamp. Different from works [21, 96, 84, 58, 42, 43] which learned another weight matrix from fully observed input, we propose to learn the weight matrix from partial input to more timely capture the effect of every specific connection in the graph and multiply it with the time-varying dynamic binary adjacency matrix to form the final one used in graph convolutions. + +To summarize, the main contributions of this paper are three folds: + +- We propose a continuous-time framework which takes the sequence of 2D skeletons extracted from a multi-person video as the input to forecast the multi-person 3D poses in any future time. +- We design our model as a hierarchical structure to reduce the computational complexity for multi-person scenarios, and also deal with different connection types in the graph. +- We utilize dynamic graph convolutions to deal with the irregular input caused by occlusions. + +In the remaining of this paper, we will briefly introduce the related works in Sec. 2 and give the details of our model in Sec. 3. The performance of our model is demonstrated through experiments in Sec. 4. Finally, we will conclude our work in Sec. 5. The corresponding source code can be found at https://github.com/SBU-YCX/HDG-ODE. + +# 2. Related Works + +# 2.1. 3D Pose Estimation, Tracking and Forecasting + +Single-Person Pose Estimation: Most of the 3D Single-Person Pose Estimation (SPPE) methods are proposed to recover the 3D poses from 2D ones extracted by a pre-trained 2D pose extractor such as Stacked Hourglass Network [51] or Convolutional Pose Machines [69]. Given 2D coordinates as input, Zeng et al. [93] proposed to split the human joints into local groups and estimate them separately. The same idea was used by Chen et al. [15], although they directly estimated 3D poses from raw images and applied different architectures for different parts. In Li et al.'s work [39], different subsets were exchanged to generate augment data for better training. Choi et al. [18] and Azizi et al. [5] used spectral graph convolutions to tackle the problem due to the structure property of input skeletons. Since images captured from monocular cameras have perspective limitations, Yu et al. [91] proposed perspective crop layers to eliminate the perspective effect, while Nie et al. [52] tried to use siamese architectures to decouple the view-dependent representation and pose-dependent representation. There are also some other works solved the perspective problems with the help of multi-view images, such as [50, 60, 80, 72, 94], by fusing the 2D poses from different views to get the original 3D results. Besides, other wearable sensors were used to complement the missing depth information in [94, 27], and Isogawa et al. [31] proposed to utilize an optical non-line-of-sight imaging system to obtain the 3D poses from photon images. + +Multi-Person Pose Estimation: Inspired by 2D Multi-Person Pose Estimation (MPPE) methods, the works for the 3D estimation task can also be divided into two groups. One is top-down methods where a human detector is applied first and estimation is made for each person separately. Yang et al. [88] followed this line of work with graph convolutional networks (GCNs) [32], T. Xu and W. Takano [83] modified the stacked hourglass structure also with the graph convolutions. Li et al.'s [37] proposed a hierarchical architecture, which is the one closest to our work, but their goal was to augment the graph with additional labeled mesh data and integrate more information, which is totally different from our motivation. The other one is bottom-up approaches where joint locations are directly estimated, then associated and grouped into different people. In [24], Fabbri et al. proposed to use distance-based heuristic for the associations from the detected head joint with the highest confidence. + +Cheng et al. [17] proposed to integrate both top-down and bottom-up branches to give better results. Besides, inspired by YOLO [59], some works [97, 6, 35, 49] explored to obtain the 3D poses in a single pass where a pivot joint was detected as well as the offsets for all other joints. Similar to single-person case, multi-view sources were also used in [25, 10, 70, 40, 28]. + +Pose Tracking and Forecasting: Since Andriluka et al. proposed PoseTrack [1] to first time introduce the Human Pose Tracking (HPT) task, contextual information has begun to be used for better estimating human poses in a sequence of inputs. Similar to other 3D pose tasks, most of works used extracted 2D pose sequence as the input and adopted top-down pipeline to mainly focus on single-person's pose tracking. Cai et al. [9] integrated spatio-temporal GCN into a stacked hourglass architecture to get the information of different scales. The same idea was applied in Wang et al.'s work [75], where the U-shaped GCN was used to extract multi-scale contexts. Liu et al. [41] proposed a graph attention spatio-temporal network with dilated temporal convolutions to get both local and global information, and Sofianos et al. [64] built a space-time-separable GCN for the 3D pose forecasting. Encoder-decoder architecture was utilized in [36, 63, 26] and Zheng et al. [98] solved the problem with spatio-temporal transformer. Furthermore, Chen et al. [11] proposed a fast method with multi-view inputs and Yuan et al. [92] utilized physics simulators and the reinforcement learning strategy in their work. Besides the skeleton joint representations, bone lengths and directions were used in [13, 82] and the extracted information was propagated through an LSTM. + +# 2.2. Graph Neural Networks + +Different from traditional convolutional neural networks (CNNs), graph neural networks (GNNs) are mainly proposed to deal with data having more general structures. GNNs can be divided into two groups, spatial and spectral graph convolutions, as described in [8]. The former one considers the graph as locally connected and the information is only shared between each node and its nearest neighbors. J. Atwood and D. Towsley [4] extended it to allow the communications among nodes within $k$ -hop ( $k >= 1$ ) connections by using the probabilistic transitions. The latter one converts the data into spectral domain first, multiplies it with the Laplacian of the topology matrix in the spectral domain to mimic the convolutional operation in the spatial domain. Compared with spatial graph convolutoins, spectral operations utilize the global graph structure each time. However, since the kernel of the whole graph is large, to reduce the computational complexity, Defferrard et al. [23] proposed to approximate it with the $k^{th}$ order polynomials of the Laplacian with the help of Chebyshev expansion to model the $k$ -hop connections. T. Kipf and M. Welling [32] proposed graph convolutional + +networks (GCNs) to further simplify it with the $1^{st}$ order approximation of the covolution in ChebNet [23]. + +In the pose estimation tasks, GCNs were widely used in [88, 83, 37, 9, 75, 41]. However, unlike CNNs, traditional GCNs have shared weights for different neighbors. To tackle this problems, another learnable weight matrix was trained in [96, 21, 42]. But, in these works, the weight matrix were learned with full joints and the final matrix used for convolution was the multiplication between the weight matrix and the static skeleton matrix assuming all joints observable. In our work, to deal with the missing case caused by occlusion, we make the skeleton matrix dynamic according to the current occlusion status so that the learnt weight matrix can also be effective even though only partial joints given. Some other works [43, 58, 84] proposed to add more semantic connections to the graph, which increased the calculation complexity of the models since the joints of all people could not be fully-connected. In addition, the semantic graphs were still globally sparse although locally dense. In this work, we propose to decompose the whole graph in a hierarchical manner. + +# 2.3. Neural Differential Equation + +As mentioned earlier, almost all the existing pose tracking methods were built based on spatio-temporal graph convolutional networks (ST-GCNs) and recurrent neural networks (RNNs), which could only model the discrete-time data where all human joints were assumed to be observable although there existed missing and inaccurate inputs due to occlusions. In order to better model the continuous-time process of the irregular data, neural ordinary differential equation (NeuralODE) was proposed by Chen et al. [12], where deep neural networks were utilized to parameterize a nonlinear ordinary differential equation (ODE). Rubanova et al. [61] and Brouwer et al. [22] further extended it by introducing a recurrent structure to efficiently encode the input information into the underlying trajectories. Some other works [44, 45, 46] also inserted Brownian terms into ODE to make it a stochastic model that can capture the uncertainties in the dynamics. In these works, pose estimation was also mentioned as a part of experiment to prove their efficiency. However, when they dealt with the human joint data, they merely considered each joint sequence as an isolated trajectory and ignored the relations among these joints. Graph neural networks were introduced to parameterize the ODEs in [56, 29, 57], but all of them utilized the whole graph, which didn't take the sparsity problem into consideration in multi-person scenarios. + +# 3. Hierarchical Dynamic-Graph Ordinary Differential Equation Model + +In this section, we will introduce the details of our hierarchical ordinary differential equation model with dynamic + +graph convolutions, which we call as HDG-ODE. As shown in Figure 1, the framework of our model follows a hierarchical structure, where each level is a sub-graph with the state changes modeled through graph ordinary differential equation and parameterized with dynamic graph convolutions. In the remaining of this section, we will describe the details of each component as well as the advantages of such designs. + +# 3.1. Problem Statement and Notations + +Given a multi-personal video containing $K$ people, and the 2D joints sequence up to time $t$ is $\mathcal{X} = \{\mathbf{X}_1,\mathbf{X}_2,\dots,\mathbf{X}_t\}$ , where $\mathbf{X}_t\in \mathcal{R}^{[K,N_j,2]}$ and $N_{j}$ is the number of joints within each person. These 2D representations can be obtained from any existing 2D pose estimator. Our goal is to predict the corresponding 3D poses $\mathbf{Y}_{\tau}\in \mathcal{R}^{[K,N_j,3]}$ at a future time $\tau$ from $\mathbf{X}_t$ , where $\tau$ can be any value within $(t,t + 1]$ . Like in many 2D estimators, the missing or occluded joints are represented as some singular values or estimated with a very low confidence, based on this, we can directly obtain the corresponding occlusion mask $\mathbf{M}_t$ along with $\mathbf{X}_t$ at each time $t$ . + +# 3.2. Hierarchical Graph Structure + +As explored in many other works, compared to the traditional use of CNN, graph convolution is more suitable for human skeleton data due to the irregular connection of joints. The feature matrices $Z^{(l)}$ and $Z^{(l + 1)}$ of two adjacent convolution layers are expressed as + +$$ +Z ^ {(l + 1)} = \sigma (W Z ^ {(l)} \tilde {A}), \tag {1} +$$ + +where $W$ is a learnable parameter matrix, $\tilde{A}$ is the normalized form of the adjacency matrix $A \in [0,1]^{N \times N}$ and $N$ is the number of nodes in the graph. However, when modeling human skeletons, we notice that each joint node is only connected to a few other joints, which will make the static skeleton-based $\tilde{A}$ very sparse and lead to plenty of meaningless calculations in (1). Although some recent works tried to learn more semantic connections among different joints besides static ones, the sparsity problem cannot be eliminated. The semantic relationship among joints usually exists within a limited number of hops instead of the whole graph. Even after we add those semantic edges, we can separate all joints into several groups such that the connections within each group are dense but the ones among different groups are still sparse. This inspires us to adopt a hierarchical structure instead of a flatten one to represent the graph of human skeletons. + +Besides, according to the physical property of human motion, different joints change their positions at different speeds over time. The location of the central joint usually has a slow and smooth change while the extremity joints, such as "wrist" and "ankle", have drastic and unexpected + +changes. Therefore, rather than allocating the same portion of computational resource to all joints, we can focus more on the latter ones. The cascade processing along the hierarchical structure also allows us to utilize the relative-central ones' locations to guide the estimation of positions of the relative-extreme joints. + +Another reason for which we adopt the hierarchical structure is its advantage of dealing with different types of edges in the whole graph. For example, in the multi-person scenario, previous works commonly used top-down methods where each person in the scene would be processed individually, but actually, pose estimation and tracking can benefit from taking the relation among people into consideration. However, the edges describing the inter-person relations are different from those representing the intra-person limbs, and it will reduce the effect if we model them in one large graph. Intuitively, separating them into different groups and forming a graph at each level will be a better choice. + +Specifically, as shown in Figure 2, for each frame of the video at time $t$ , we get a set of 2D skeleton coordinates of all the people in the scene as input. To apply the hierarchical pipeline, we choose the "pelvis" joint as the root node of each person, whose coordinate can be also considered as the location of the person. All these root nodes form the 1st-level graph and encoded by a graph encoder + +$$ +\mathbf {Y} _ {\tau} ^ {(1)} = G _ {1} \left(\mathbf {X} _ {t} ^ {(1)}, \mathbf {H} _ {< t} ^ {(1)}, \tilde {A} _ {t} ^ {(1)}\right), \tag {2} +$$ + +where $\mathbf{X}_t^{(1)}\in \mathcal{R}^{[K,1,2]}$ represents the 2D coordinates of all root nodes of $K$ people, $\mathbf{Y}_{\tau}^{(1)}\in \mathcal{R}^{[K,1,3]}$ is the corresponding 3D one forecasted for the future time $\tau$ , $\mathbf{H}_{< t}^{(1)}$ is the historical feature, and $\tilde{A}_t^{(1)}$ is the normalized adjacency matrix with the edges calculated based on the distances among connected roots, i.e. the relations among people. There is an edge between two people if their root nodes' distance is smaller than a threshold (the average of the heights of joints in 2D skeleton, i.e. distance from "head" to "ankle" joint in our setting), which indicates they are related. Then, all the limb-root joints, including "neck", "left shoulder", "right shoulder", "left hip" and "right hip", as well as the root node are grouped to form the 2nd-level graph within each person. To utilize the result of the 1st-level graph, the output of this level is designed as the spatial shifts of the limb-root joints' locations from the root node. We can express the process as below + +$$ +\mathbf {D} _ {\tau} ^ {(2, i)} = G _ {2} \left(\mathbf {X} _ {t} ^ {(2, i)}, \mathbf {Y} _ {\tau , i} ^ {(1)}, \mathbf {H} _ {< t} ^ {(2, i)}, \tilde {A} _ {t} ^ {(2, i)}\right), i \in [ 1, K ] +$$ + +$$ +\mathbf {Y} _ {\tau} ^ {(2, i)} = \mathbf {D} _ {\tau} ^ {(2, i)} + \mathbf {Y} _ {\tau , i} ^ {(1)}, \tag {3} +$$ + +where $\mathbf{X}_t^{(2,i)}\in \mathcal{R}^{[1,6,2]}$ is the 2D coordinates vector of all limb-root nodes as well as root node of person $i$ $\mathbf{Y}_{\tau ,i}^{(1)}$ is the 3D root node's position of this person obtained by (2), $\mathbf{D}_{\tau}^{(2,i)}$ and $\mathbf{Y}_{\tau}^{(2,i)}$ represent respectively the future 3D position shift and final position predicted. Finally, we group each limbroot joint to the corresponding limb-extreme joint to form + +![](images/467062f85237c008b503f50662a640fbf8192634ab81d8ed1949e26121aaca02.jpg) +Figure 1. The framework of HDG-ODE. $\mathbf{X}_{t_n}$ are the 2D joints coordinates at time $t_n$ , $\mathbf{M}_{t_n}$ is the occlusion mask obtained from $\mathbf{X}_{t_n}$ . $\mathbf{Y}_{\tau}$ are the predicted 3D coordinates at the future time $\tau$ . + +the 3rd-level graph. Similarly, the encoding process within each lamb group $j$ of person $i$ can be expressed as + +$$ +\mathbf {D} _ {\tau} ^ {(3, i, j)} = G _ {3} (\mathbf {X} _ {t} ^ {(3, i, j)}, \mathbf {Y} _ {\tau , j} ^ {(2, i)}, \mathbf {H} _ {< t} ^ {(3, i, j)}, \tilde {A} _ {t} ^ {(3, i, j)}), j \in [ 1, 5 ] +$$ + +$$ +\mathbf {Y} _ {\tau} ^ {(3, i, j)} = \mathbf {D} _ {\tau} ^ {(3, i, j)} + \mathbf {Y} _ {\tau , j} ^ {(2, i)}, \tag {4} +$$ + +where $\mathbf{X}_t^{(3,i,j)}\in \mathcal{R}^{[1,3,2]}$ is the vector of the 2D coordinates representing all nodes of person $i$ within the limb group $j$ , $\mathbf{Y}_{\tau ,j}^{(2,i)}$ is the 3D location of the lamb-root $j$ calculated by (3), $\mathbf{H}_{< t}^{(3,i,j)}$ and $\tilde{A}_t^{(3,i,j)}$ represent the corresponding part of the historical feature and the adjacency matrix in this level. 3D shift $\mathbf{D}_{\tau}^{(3,i,j)}$ is used to get the final 3D coordinates $\mathbf{Y}_{\tau}^{(3,i,j)}$ at future time $\tau$ . + +![](images/b0181a44ee320cd1948ffc81268471da4d1f4eb12dd6ce9120ed1e043183479d.jpg) +(a) + +![](images/683fb5334ee6e4109990615963615a7d16c1e3913807e2f2d94003cd47119634.jpg) +(b) +Figure 2. Illustration of the hierarchical graph structure: (a) the skeleton representations of a multi-person scenario; (b) the corresponding 3-level hierarchical structure. + +# 3.3. Time-Varying Dynamic Graph Convolution + +A most important impact factor to the performance of human pose estimation is the occlusion, no matter the occlusion is caused by other objects or the self-occlusion happens + +within a person. To get the 2D skeleton inputs with some 2D pose estimation methods, the visible joints can be directly extracted from the raw images while the occluded ones are missing. Even though the missing joints can be estimated by some intuition, they are less accurate due to the uncertainties. When we use such 2D poses as inputs to recover the corresponding 3D poses, called the lifting process, the inaccuracy will accumulate and compromises data processing during the information communication within the graph convolution operation. Therefore, we hope our graph encoder can be reasonably adaptive to deal with the missing and inaccurate cases. + +In the literature work that apply the traditional graph convolution for this task, the adjacency matrix $A$ and its normalized form $\tilde{A}$ in (1) usually keep constant to represent the static connections within the graph, i.e. all the joints are assumed to be fully available, no matter they are observable or occluded. However, as we described, only those visible ones are accurate and worth being integrated to guide the forecasting. So we propose to use the time-varying dynamic graph convolution to only encode visible joints input at each timestamp. Inspired by some literature work [96], the learning of a hybrid adjacency-weight matrix may be written as + +$$ +Z _ {t} ^ {(l + 1)} = \sigma \left(W Z _ {t} ^ {(l)} \rho \left(W _ {\tilde {A}} \odot \tilde {A}\right)\right), \tag {5} +$$ + +where $W_{\tilde{A}}$ is a learnable weight matrix applied to assign the information along different edges with different weights to solve the weight-sharing problem [21, 42] and $\rho(\cdot)$ is a nonlinear function. However, we notice that in (5), $W_{\tilde{A}}$ is learnt based on the fully-observable data and the $\tilde{A}$ is the static one without taking the occlusion cases into consideration. In other words, although the learnt weights work well when all + +joints are available, they may not be effective if partial of joints are occluded and missing. Based on this, we make a further step and propose to also take the occlusion status into consideration when obtaining the dynamic weight matrix. Specifically, at each timestamp $t$ , we will drop the occluded joints in the original static adjacency matrix $A$ to form a time-varying adjacency matrix $A_{t}$ , and this dynamic adjacency matrix will be used in both training and testing, i.e. (5) can be rewritten as + +$$ +Z _ {t} ^ {(l + 1)} = \sigma \left(W Z _ {t} ^ {(l)} \rho \left(W _ {\tilde {A}} \odot \tilde {A} _ {t}\right)\right), \tag {6} +$$ + +where $\tilde{A}_t$ is the normalized form of the real adjacency matrix at time $t$ . By doing so, the learning of the weight matrix $W_{\tilde{A}}$ will only focus on the observable joints to make sure it can also be effective when only partial joints are available. In the testing phase, the hybrid-matrix $W_{\tilde{A}} \odot \tilde{A}_t$ can ignore the missing joints and adaptively adjust the weights along existing edges for information integration among joints. + +# 3.4. Parallel Ordinary Differential Equation + +Besides the partial occlusion, which leads to incomplete graph input at each frame and can be regarded as spatial missing, there are also many cases where a person is totally occluded, and causes the temporal missing within a sequence. Previous works almost rely on the discrete-time sequential models, which are not flexible enough since the missing in real world can vary from several frames to dozens of frames. + +To make the model more robust, we adopt ordinary differential equation (ODE), a continuous-time model in this paper, which can be expressed as + +$$ +\frac {d \mathbf {H} _ {t}}{d t} = F (\mathbf {H} _ {t}), +$$ + +$$ +\mathbf {H} _ {t} = \mathbf {H} _ {t _ {n - 1}} + \int_ {t _ {n - 1}} ^ {t} F (\mathbf {H} _ {\tau}) d \tau . \tag {7} +$$ + +Compared with traditional recurrent networks, it can be used to predict the values over the interval of any length. Specially, in our model, we use time-varying dynamic graph convolution mentioned in Sec. 3.3 to parameterize the derivatives $F(\cdot)$ . For efficiency, we compute the integration in (7) by Euler Method and we can rewrite it as + +$$ +\mathbf {H} _ {t, F} = \mathbf {H} _ {t - \Delta t, F} + F \left(\mathbf {H} _ {t - \Delta t, F}, \tilde {A}\right) \Delta t, \tag {8} +$$ + +where $\mathbf{H}_{t,F}$ is the hidden feature, $\tilde{A}$ is the normalized adjacency matrix assuming all joints observable, and $\Delta t$ is the step-size. At an observation time $t_n$ , the features of observable joints are updated through the same graph convolution encoder as described in Sec. 3.2, i.e. + +$$ +\mathbf {H} _ {t _ {n}, G} = G \left(\mathbf {X} _ {t _ {n}}, \mathbf {Y} _ {t _ {n}}, \mathbf {H} _ {< t _ {n}}, \tilde {A} _ {t _ {n}}\right), \tag {9} +$$ + +where $G = \{G_1, G_2, G_3\}$ represents the cascade process in (2)-(4) parameterized by our time-varying dynamic graph + +convolution networks, $\mathbf{X}_{t_n}$ and $\mathbf{Y}_{t_n}$ are the 2D and 3D joint locations, $\mathbf{H}_{< t_n} = \mathbf{H}_{t_n - \Delta t,F}$ is the feature from the ODE model and $\tilde{A}_{t_n}$ is the normalized adjacency matrix of the graph formed by observable joints. After the update, the new feature at the time instant $t_n$ becomes + +$$ +\mathbf {H} _ {t _ {n}} = \mathbf {M} _ {t _ {n}} \odot \mathbf {H} _ {t _ {n}, F} + \left(\mathbf {1} - \mathbf {M} _ {t _ {n}}\right) \odot \mathbf {H} _ {t _ {n}, G}, \tag {10} +$$ + +where $\mathbf{M}_{t_n}$ is the occlusion mask of the input described in Sec. 3.1. Here, we want to emphasize that $\tilde{A}_{t_n}$ in (9) and $\tilde{A}$ in (8) may not be the same according to the observation states of input. As we described in last section, the input may contain missing values, but all the missing values have been completed by the encoder in the encoding process and we can assume the graph is fully observable in the propagation process. + +![](images/02f0ae8ba3f13ea4c5935e91071fac87e299de3555ec1a49cde77b910f72fbb6.jpg) +Figure 3. Illustration of the parallel ODE pipeline. + +Here, we can further demonstrate the advantages of hierarchical structure. For efficiency, we adopt Euler method to approximately calculate the integration, but as we know, it will predict the curve as a straight line if we make the propagation step $\Delta t$ too large and equal to the interval between two observations as $t_n - t_{n-1}$ . To make the model more expressive, we usually make $\Delta t$ smaller, i.e. + +$$ +\Delta t = \frac {1}{k} \left(t _ {n} - t _ {n - 1}\right). \tag {11} +$$ + +However, as we discussed in Sec. 3.2, different joints of human body have different rates of location changes, we don't need to set $k$ to the same value at different levels in our model. Instead, we can adopt a parallel pipeline and make it small in shallow levels while enlarge it as the level goes deeper as shown in Figure 3. + +# 4. Experiments + +# 4.1. Experimental Setups + +# 4.1.1 Datasets + +MuPoTS-3D[49]: This multi-person dataset contains sequences captured in 20 real scenes. Each sequence has up to three subjects, as well as provides raw images, 2D pose annotations and the corresponding 3D ones. The lengths of data vary from 125 frames to 800 frames, and we split them + +into segments with the length of 50 frames. We divide all the data into training, validation and testing sets with the ratio of 0.8: 0.1: 0.1. The real occlusion mask are given as well, and the final data are obtained by the element-wise multiplication between the full data and the occlusion masks. Before being fed into the models, data from each sequence are first normalized. + +Human3.6M[30]: This dataset consists of 3.6 million of single-person frames. 11 subjects conducting 15 activities are captured by cameras. Both 2D and 3D annotations are provided where 2D ones come from cameras while 3D ones are obtained using a motion capture system. Different from commonly used training methods, we only use data from one camera to train the model. We performed the same data partition and normalization as above. + +# 4.1.2 Evaluation Metrics + +Mean Per Joint Position Error (MPJPE): This metric measures the average position error over all joints and frames in the sequence, which can be expressed as + +$$ +M P J P E = \frac {1}{N _ {f}} \sum_ {f} \frac {1}{N _ {j} ^ {(f)}} \sum_ {j} \left\| P _ {\text {p r e d}} ^ {(f)} (j) - P _ {g t} ^ {(f)} (j) \right\| _ {2}, \tag {12} +$$ + +where $N_{f}$ is the total number of frames, $N_{j}^{(f)}$ is the total number of joints in the frame $f$ , $P_{gt}^{(f)}(j)$ and $P_{pred}^{(f)}(j)$ are the ground-truth and the predicted 3D position of the joint $j$ in the frame $f$ . + +Percentage of Correct Keypoints (PCK): PCK calculates the percentage of correctly predicted key-points, where a key-point $j$ in the frame $f$ is considered to be correct if the distance $d_{p}^{(f)}(j)$ between the 3D ground-truth $P_{p,gt}^{(f)}(j)$ and the predicted 3D position $P_{p,pred}^{(f)}(j)$ are smaller than a person-specific threshold $S_{p}^{f}$ , i.e. + +$$ +P C K = \frac {\sum_ {f} \sum_ {p} \sum_ {j} \delta \left(d _ {p} ^ {(f)} (j) < S _ {p} ^ {(f)}\right)}{\sum_ {f} \sum_ {p} \sum_ {j} 1}, \tag {13} +$$ + +Here we adopt the setting in MPII dataset[2] where the threshold is the ratio of head link's length $h_p^{(f)}$ of the person $p$ , i.e. the distance between "head" joint and "upper neck", and the PCK can be represented as + +$$ +P C K h @ \alpha = \frac {\sum_ {f} \sum_ {p} \sum_ {j} \delta \left(\frac {d _ {p} ^ {(f)} (j)}{h _ {p} ^ {(f)}} < \alpha\right)}{\sum_ {f} \sum_ {p} \sum_ {j} 1}, \tag {14} +$$ + +where $\alpha$ is the ratio parameter. + +# 4.1.3 Implementation Details + +As introduced in Sec. 3.2, the hierarchical graph has 3 levels, containing $K$ , 6 and 3 nodes each, and our HDG-ODE model has a hierarchical structure of 3 levels, with an encoder and + +an ODE at each level. For each encoder, we use a single-layer graph-GRU, using our time-varying dynamic graph convolutions with the feature size of 6. ODE models are parameterized with graph neural networks of 1, 2 and 3 convolutional layers at the three levels respectively, with the hidden feature size of 16. The propagation steps for the three levels are respectively $[1/30s, 1/60s, 1/120s]$ , and the data sampling interval is $1/30s$ . The experiments are conducted on single NVIDIA 1080 Ti GPU. During the training, we use the Adam() optimizer with the learning rates of [0.01, 0.001, 0.0001] to minimize the mean squared error (MSE) between the predicted 3D positions and the ground-truth of the visible joints, and 100 epochs are run at each learning rate. To speed up the training, it will be early-stopped if there is no further loss decrease in 20 epochs. + +# 4.2. Experiment Results + +# 4.2.1 Overall Performance + +We conduct experiments over MuPoTS-3D dataset to evaluate the performance of our proposed model on the multiperson 3D pose tracking task. For comparison, we choose some representative methods, including discrete-time recurrent models and continuous-time ones, as the baselines. Specifically we take the following methods into consideration: + +- STGCN: This discrete-time model consists of several concatenated spatio-temporal convolutional blocks (ST-Conv blocks), which is a common structure that is widely used for action recognition [85], pose forecasting [75] and other prediction tasks [89]. In our experiments, STGCN has 2 ST-Conv blocks, and recurrently processes each clip within the whole 2D pose sequence and give the 3D predictions. The output dimension of each layer in the blocks is 32. +- Graph-GRU: It is a variant of conventional Gated Recurrent Unit (GRU), where the linear operation in the conventional GRU is replaced by graph convolution. It works as the backbone in [90]. We construct the baseline with 2 Graph-GRU layer and a linear decoder. The output dimension of each layer in the blocks is 32. +- SemGCGRU: It is an improved version of traditional Graph-GRU, where the fixed binary adjacency matrix is replaced by a learnable weight matrix to capture the relations among nodes [96]. In our implementation, the structure and the feature size are the same as Graph-GRU. +- ODE-RNN: In this model [62], the deep learning modules, specifically dense neural networks, are incorporated to parameterize a non-linear ordinary differential equation (ODE) to model the continuous-time process + +of data sequence. In our experiments, the dense network contains 2 linear layers and the dimension of hidden features is 32. + +- Graph-ODE: It is a GNN variant of ODE-RNN, with a GCN to learn the derivative of ODE and a GRU to integrate the information of observations. In our model comparison, the GCN has 2 hidden layers with ReLU activation functions and the output dimension of each hidden layer is 32. The output dimension of GRU is 32 as well. In other words, to compare with discrete models, we just replace the GRU module in Graph-GRU with an ODE module. + +Although continuous-time models can provide predictions at any time, even smaller than the data sampling rate, we set the step of them to be equal to the original data interval used in discrete-time models for fair comparison. From the results in Table 1, we can see that our HDG-ODE can achieve the best result with the smallest mean prediction error and the highest percentage of corrected prediction. + +
MPJPE(↓)PCKh@α(↑)
α=0.1α=0.5α=1.0
Discrete
STGCN [89, 75]0.600413.03%60.35%77.76%
Graph-GRU [90]0.557314.86%62.39%80.04%
SemGCGRU [96]0.437517.09%69.61%85.96%
Continuous
ODE-RNN [61]0.439620.21%71.78%85.45%
Graph-ODE [56]0.406619.80%71.57%87.00%
Ours
HDG-ODE0.303829.72%78.67%92.18%
+ +# 4.2.2 Ablation Study + +In this part, we will separately demonstrate the effectiveness of each designing component in our proposed model by some experiment results, including the continuous ODE model, the pipeline taking all people into consideration, the new time-varying dynamic graph convolution and the proposed hierarchical structure. The evaluation is sequential by adding the components one by one, and each time we only choose the current best model to conduct the experiment for the next component. + +Discrete-Time Model vs. Continuous-Time Model: From Table 1, compared to traditional discrete-time models, no matter our HDG-ODE or other continuous-time models have a better performance. Different from discrete models which only execute one propagation between two observations, the continuous-time ones conduct multiple-step propagation so that the information on each joint will be affected + +and corrected by its neighbors several times before getting the final prediction. Also, due to the property of ODE, the continuous-time model can provide predictions at any time in the interval between two observations, instead of just obtaining predictions at the observation instants as done by discrete-time ones. + +Isolated Top-Down Pipeline vs. Relational Top-Down Pipeline: As we described previously, most of existing multiperson pose tracking works adopt the top-down processing pipeline, where each person in the scenario is tracked separately as an isolated trajectory. Instead, we take into consideration the relation among people to infer their relative movements, which in turn affects the future positions of specific joints. In Table 2, by comparing the models in isolated top-down pipeline with the ones in our relational pipeline, we can observe the improvement. + +Table 1. Testing Performance of different models on MuPoTS-3D. + +
Isolated PipelineRelational Pipeline
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.610477.72%0.600477.76%
Graph-GRU0.573878.46%0.557380.04%
Graph-ODE0.420386.89%0.406687.00%
+ +Static Graph Convolution vs. Time-Varying Dynamic Graph Convolution: In the real cases, occlusion is inevitable. To better tackle the problem caused by the input data with missing or inaccurate values due to occlusions, we propose to replace the static graph convolution used in most of works with our time-varying dynamic graph convolution in the HDG-ODE model. As shown in Table 3, the dynamic graph convolution helps eliminate the negative influence of inaccurate inputs as well as adjust the integration weights of edges to the most suitable ones according to occlusion states, so the models can have a better performance than before. + +Table 2. Comparison between the isolated top-down pipeline and our relational top-down pipeline on MuPoTS-3D. + +
Original Graph ConvOur Graph Conv
MPJPEPCKh@1.0MPJPEPCKh@1.0
Graph-GRU0.557380.04%0.384486.45%
Graph-ODE0.406687.00%0.370688.74%
+ +Table 3. Comparison between original static graph convolutions (GC) and our time-varying dynamic convolutions (DG) on MuPoTS-3D. + +Flatten Structure vs. Hierarchical Structure: To solve the sparsity problem explained in Sec. 3, our HDG-ODE adopts a hierarchical structure. In Table 4, compared to models in the flatten structure with a large sparse graph, the corresponding hierarchical ones perform better for models with graph convolutions. The reasons leading to the improvement lie in two aspects: On one hand, by decomposing the whole sparse graph into a set of relatively dense ones, the parameters to learn are reduced, as shown in Table 5. This helps save a significant amount of computational resource, and the released resource can be reallocated to the joints + +with more complicated movements, e.g. making the corresponding network deeper or reducing the propagation step of ODE. On the other hand, the cascade processing along the hierarchical structure can provide the guidance among the levels. As for the ODE-RNN, the hierarchical one has a worse result because the structure is mainly proposed to deal with the graph sparsity problem. However, ODE-RNN doesn't utilize the graph information and only cares about isolated joint trajectories, so the hierarchical structure will make the model more complex and reduce the performance. + +
Flatten StructureHierarchical Structure
MPJPEPCKh@1.0MPJPEPCKh@1.0
ODE-RNN0.439685.45%0.493882.40%
GC-ODE0.406687.00%0.345190.51%
DG-ODE0.370688.74%0.303892.18%
+ +Table 4. Performance Comparison between the flatten structure and the hierarchical structure on MuPoTS-3D. + +
#Params (↓)Memory (↓)
DG-ODE106490.76M
HDG-ODE26430.66M
+ +# 4.2.3 Robustness Analysis + +Besides the multi-person scenarios, we also conduct experiments of single-person scenarios on Human3.6M dataset. Different from MuPoTS-3D, no occlusion information is provided within the data, so we make occlusion masks by ourselves and test the robustness of our model under different observable ratios. Specifically, for a sequence, we assume only $p_t$ of the frames are observable while in the remaining ones, the input skeletons are totally missing due to the occlusion by buildings or other large objects. For each observable frame, we assume there exist self-occlusions such that only $p_s$ of the joints are available. To make the corresponding mask, both $p_t$ of the frames and $p_s$ of the joints on each frame are randomly chosen and the total observable ratio can be expressed as $p_t \times p_s$ . The results are shown in Table 6. When the observable ratio decreases as a result of occlusions, the performance gets worse for all models. However, our HDG-ODE always performs the best among all models studied under all circumstances. + +# 5. Conclusion + +In this paper, we propose a new continuous-time model called HDG-ODE to forecast the future 3D human pose representations in multi-person videos given the 2D pose sequence as input. By using ordinary differential equations and dynamic graph convolutions, we alleviate the occlusion problem brought by perspective effects of monocular images. + +Table 5. Complexity Comparison between the flatten Dynamic-Graph ODE (DG-ODE) and our hierarchical one (HDG-ODE). + +
pt=0.8, ps=0.5pt=0.6, ps=0.4
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.396680.19%0.437675.82%
Graph-GRU0.372383.34%0.418877.39%
ODE-RNN0.352784.33%0.382680.66%
Graph-ODE0.347383.98%0.401978.54%
HDG-ODE0.305288.29%0.354382.36%
+ +Table 6. Performance of different models on Human3.6M with different observable ratios. + +Besides, with the help of hierarchical structure, we reduce the computational complexity of our model and are able to allocate more resources to the human joints which have more complicated movement. Our experiments demonstrate that our model can well recover the corresponding 3D poses from the single-view 2D poses and outperform literature works on the human pose forecasting task. Currently, the maximum number of people in the scenes of the video is limited. In the future, we will make our model more adaptive and test it in more crowded scenarios. + +# References + +[1] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. Posetrack: A benchmark for human pose estimation and tracking, 2017. +[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. +[3] Bruno Artacho and Andreas Savakis. Unipose: Unified human pose estimation in single images and videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[4] James Atwood and Don Towsley. Diffusion-convolutional neural networks, 2015. +[5] Niloofar Azizi, Horst Possegger, Emanuele Rodola, and Horst Bischof. 3d human pose estimation using mobius graph convolutional networks, 2022. +[6] Abdallah Benzine, Florian Chabot, Bertrand Luvison, Quoc Cuong Pham, and Catherine Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[7] Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, and Lorenzo Torresani. Learning temporal pose estimation from sparsely labeled videos. In Advances in Neural Information Processing Systems 33, 2019. +[8] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs, 2013. +[9] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. + +[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. arXiv preprint arXiv:2007.10986, 2020. +[11] Long Chen, Haizhou Ai, Rui Chen, Zijie Zhuang, and Shuang Liu. Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[12] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations, 2018. +[13] Tianlang Chen, Chen Fang, Xiaohui Shen, Yiheng Zhu, Zhili Chen, and Jiebo Luo. Anatomy-aware 3d human pose estimation in videos. arXiv preprint arXiv:2002.10322, 2020. +[14] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. +[15] Zerui Chen, Yan Huang, Hongyuan Yu, Bin Xue, Ke Han, Yiru Guo, and Liang Wang. Towards part-aware monocular 3d human pose estimation: An architecture search approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 715-732, Cham, 2020. Springer International Publishing. +[16] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[17] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7649-7659, June 2021. +[18] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020. +[19] Chia-Jung Chou, Jui-Ting Chien, and Hwann-Tzong Chen. Self adversarial training for human pose estimation. In 2018 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 17-30, 2018. +[20] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L. Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. +[21] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. +[22] Edward De Brouwer, Jaak Simm, Adam Arany, and Yves Moreau. Gru-ode-bayes: Continuous modeling of sporadically-observed time series, 2019. + +[23] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. 2016. +[24] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[25] Qi Fang, Qing Shuai, Junting Dong, Hujun Bao, and Xiaowei Zhou. Reconstructing 3d human pose by watching humans in the mirror. In CVPR, 2021. +[26] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In CVPR, 2022. +[27] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, June 2021. +[28] Congzhentao Huang, Shuai Jiang, Yang Li, Ziyue Zhang, Jason Traish, Chen Deng, Sam Ferguson, and Richard Yi Da Xu. End-to-end dynamic matching network for multi-view multi-person 3d pose estimation. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII, page 477-493, Berlin, Heidelberg, 2020. Springer-Verlag. +[29] Zijie Huang, Yizhou Sun, and Wei Wang. Coupled graph ode for learning interacting system dynamics. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery amp; Data Mining, KDD '21, page 705-715, New York, NY, USA, 2021. Association for Computing Machinery. +[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. +[31] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris M. Kitani. Optical non-line-of-sight physics-based 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[32] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks, 2016. +[33] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Multiposenet: Fast multi-person pose estimation using pose residual network. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. +[34] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[35] Jogendra Nath Kundu, Ambareesh Revanur, Govind Vitthal Waghmare, Rahul Mysore Venkatesh, and R. Venkatesh Babu. Unsupervised cross-modal alignment for multi-person 3d pose estimation, 2020. + +[36] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R. Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[37] Han Li, Bowen Shi, Wenrui Dai, Yabo Chen, Botao Wang, Yu Sun, Min Guo, Chenlin Li, Junni Zou, and Hongkai Xiong. Hierarchical graph networks for 3d human pose estimation, 2021. +[38] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[39] Shichao Li, Lei Ke, Kevin Pratama, Yu-Wing Tai, Chi-Keung Tang, and Kwang-Ting Cheng. Cascaded deep monocular 3d human pose estimation with evolutionary training data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[40] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021. +[41] Junfa Liu, Juan Rojas, Zhijun Liang, Yihui Li, and Yisheng Guan. A graph attention spatio-temporal convolutional networks for 3d human pose estimation in video. arXiv preprint arXiv:2003.14179, 2020. +[42] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X, page 318–334, Berlin, Heidelberg, 2020. Springer-Verlag. +[43] Shengyuan Liu, Pei Lv, Yuzhen Zhang, Jie Fu, Junjin Cheng, Wanqing Li, Bing Zhou, and Mingliang Xu. Semi-dynamic hypergraph neural network for 3d pose estimation. In Christian Bessiere, editor, Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI20, pages 782-788. International Joint Conferences on Artificial Intelligence Organization, 7 2020. Main track. +[44] Xuanqing Liu, Tesi Xiao, Si Si, Qin Cao, Sanjiv Kumar, and Cho-Jui Hsieh. Neural sde: Stabilizing neural ode networks with stochastic noise, 2019. +[45] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, and Zhaoyue Chen. Learning continuous-time dynamics by stochastic differential networks, 2020. +[46] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, Zhaoyue Chen, and Jacqueline Wu. Continuous-time stochastic differential networks for irregular time series modeling. In Teddy Mantoro, Minho Lee, Media Anugerah Ayu, Kok Wai Wong, and Achmad Nizar Hidayanto, editors, Neural Information Processing, pages 343-351, Cham, 2021. Springer International Publishing. +[47] Zhenguang Liu, Haoming Chen, Runyang Feng, Shuang Wu, Shouling Ji, Bailin Yang, and Xun Wang. Deep dual consecutive network for human pose estimation. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 525-534, June 2021. +[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13264-13273, June 2021. +[49] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb, 2017. +[50] Rahul Mitra, Nitesh B. Gundavarapu, Abhishek Sharma, and Arjun Jain. Multiview-consistent semi-supervised learning for 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[51] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation, 2016. +[52] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement, 2020. +[53] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. +[54] Wanli Ouyang, Xiao Chu, and Xiaogang Wang. Multisource deep learning for human pose estimation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2337-2344, 2014. +[55] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-286, 2018. +[56] Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations, 2019. +[57] Michael Poli, Stefano Massaroli, Clayton M. Rabideau, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Continuous-depth neural models for dynamic graph prediction, 2021. +[58] Zhongwei Qiu, Kai Qiu, Jianlong Fu, and Dongmei Fu. Dgen: Dynamic graph convolutional network for efficient multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 34(07):11924-11931, Apr. 2020. +[59] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. +[60] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. 2020. +[61] Yulia Rubanova, Ricky T. Q. Chen, and David Duvenaud. Latent odes for irregularly-sampled time series, 2019. + +[62] Yulia Rubanova, Tian Qi Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series. In Advances in Neural Information Processing Systems 32, pages 5321-5331. 2019. +[63] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2484-2494, June 2021. +[64] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11209-11218, October 2021. +[65] Kai Su, Dongdong Yu, Zhenqi Xu, Xin Geng, and Changhu Wang. Multi-person pose estimation with enhanced channelwise and spatial information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[66] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5686-5696, 2019. +[67] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[68] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose estimation. ArXiv, abs/1911.07451, 2019. +[69] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 1653-1660, 2014. +[70] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020. +[71] Ali Varamesh and Tinne Tuytelaars. Mixture dense regression for object detection and human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13294-13304, June 2021. +[73] Jiahang Wang, Sheng Jin, Wentao Liu, Weizhong Liu, Chen Qian, and Ping Luo. When human pose estimation meets robustness: Adversarial algorithms and benchmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11855-11864, June 2021. +[74] Jian Wang, Xiang Long, Yuan Gao, Errui Ding, and Shilei Wen. Graph-pcnn: Two stage human pose estimation with graph pose refinement. In The European Conference on Computer Vision (ECCV), 2020. + +[75] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos, 2020. +[76] Manchen Wang, Joseph Tighe, and Davide Modolo. Combining detection and tracking for human pose estimation in videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[77] Rui Wang, Zhongzheng Cao, Xiangyang Wang, Zhi Liu, and Xiaoqiang Zhu. Human pose estimation with deeply learned multi-scale compositional models. IEEE Access, 7:71158-71166, 2019. +[78] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. +[79] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018. +[80] Rongchang Xie, Chunyu Wang, and Yizhou Wang. Meta-fuse: A pre-trained fusion model for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[81] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018. +[82] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[83] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16105-16114, June 2021. +[84] Xixia Xu, Qi Zou, and Xue Lin. Adaptive hypergraph neural network for multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2955-2963, Jun. 2022. +[85] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition, 2018. +[86] Wei Yang, Shuang Li, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Learning feature pyramids for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. +[87] Wei Yang, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. End-to-end learning of deformable mixture of parts and deep convolutional neural networks for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. +[88] Yiding Yang, Zhou Ren, Haoxiang Li, Chunluan Zhou, Xin chao Wang, and Gang Hua. Learning dynamics via graph neural networks for human pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8074-8084, June 2021. +[89] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), 2018. + +[90] Bing Yu, Haoteng Yin, and Zhanxing Zhu. ST-UNet: A spatio-temporal u-network for graph-structured time series modeling. arXiv, abs/1903.05631, 2019. +[91] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9064-9073, June 2021. +[92] Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. Simpoe: Simulated character control for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7159–7169, June 2021. +[93] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Srnet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 507–523, Cham, 2020. Springer International Publishing. +[94] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[95] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[96] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N. Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[97] Jianan Zhen, Qi Fang, Jiaming Sun, Wentao Liu, Wei Jiang, Hujun Bao, and Xiaowei Zhou. Smap: Single-shot multiperson absolute 3d pose estimation, 2020. +[98] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11656-11665, October 2021. +[99] Tianfei Zhou, Wenguan Wang, Si Liu, Yi Yang, and Luc Van Gool. Differentiable multi-granularity human representation learning for instance-aware human semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1622-1631, June 2021. +[100] Bin Xiao Zhaoxiang Zhang Jingdong Wang Zigang Geng, Ke Sun. Bottom-up human pose estimation via disentangled keypoint regression. In CVPR, 2021. \ No newline at end of file diff --git a/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/images.zip b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01f3afeb74cedee4a7458934299bab98b35e3d05 --- /dev/null +++ b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5f71ad58c74a2a602ef34e2d594b3252ac270bf8f64b09af55db02d5a09e0c7 +size 333138 diff --git a/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/layout.json b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc2d2746690774efa6b9d787fe982fe617d5642 --- /dev/null +++ b/2023/HDG-ODE_ A Hierarchical Continuous-Time Model for Human Pose Forecasting/layout.json @@ -0,0 +1,10971 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 547, + 122 + ], + "type": "text", + "content": "HDG-ODE: A Hierarchical Continuous-Time Model for Human Pose Forecasting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 144, + 484, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 144, + 484, + 185 + ], + "spans": [ + { + "bbox": [ + 109, + 144, + 484, + 185 + ], + "type": "text", + "content": "Yucheng Xing Xin Wang \nDepartment of Electrical and Computer Engineering, Stony Brook University \n{yucheng.xing, x.wang}@stonybrook.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 238, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 574 + ], + "type": "text", + "content": "Recently, human pose estimation has attracted more and more attention due to its importance in many real applications. Although many efforts have been put on extracting 2D poses from static images, there are still some severe problems to be solved. A critical one is occlusion, which is more obvious in multi-person scenarios and makes it even more difficult to recover the corresponding 3D poses. When we consider a sequence of images, the temporal correlation among the contexts can be utilized to help us ease the problem, but most of the current works only rely on discrete-time models and estimate the joint locations of all people within a whole sparse graph. In this paper, we propose a new framework, Hierarchical Dynamic Graph Ordinary Differential Equation (HDG-ODE), to tackle the 3D pose forecasting task from 2D skeleton representations in videos. Our framework adopts ODE, a continuous-time model, as the base to predict the 3D joint positions at any time. Considering the structural-property of the skeleton data in representing human poses and the possible irregularity caused by occlusion, we propose the use of dynamic graph convolution as the basic operator. To reduce the computational complexity introduced by the sparsity of the pose graph, our model takes a hierarchical structure where the encoding process at the observation timestamp is done in a cascade manner while the propagation between observations is conducted in parallel. The performance studies on several datasets demonstrate that our model is effective and can out-perform other methods with fewer parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "With the progress in the computer vision field, human-related contents have attracted more and more attention, as they are very important in many real-world applications. Some examples are autonomous driving, human-computer interaction and anomaly detection. Among these contents, human poses are the most basic and important ones since their accurate and timely estimation or even forecasting usually serve as the first step for further analyzing complex" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 214, + 347, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 347, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 347, + 225 + ], + "type": "text", + "content": "activities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 230, + 548, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 548, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 548, + 686 + ], + "type": "text", + "content": "Human pose estimation is not a novel topic but has been studied for many years. Its performance, however, is still not satisfactory enough to meet the demands of many practical applications. Since A. Toshev and C. Szegedy proposed DeepPose [69] to represent human poses by the locations of a set of joints from human skeletons and first time utilized the regression method of deep learning to estimate the corresponding 2D coordinates, a lot of efforts [78, 51, 87, 20, 86, 19, 79, 77, 54, 67, 95, 94, 3, 66, 74, 73] have been made to get the accurate estimation in the presence of occlusion resulted from the perspective limitation of images. These methods [14, 33, 55, 65, 34, 68, 38, 53, 16, 71, 58, 48, 99, 100, 84] are extended to multi-person scenarios, where the occlusion problem is more severe. Although some progress has been made, 2D locations cannot perfectly restore the human poses, and are never our final goals for this task. Recently, more and more works have been focusing on 3D pose estimation with raw images [15, 24, 6, 91] or extracted 2D poses [96, 21, 5, 18, 93, 39, 97, 83, 17, 43, 91] as input. Besides the inherent occlusion problems, the lack of depth in monocular images also limits the performance of algorithms. Although some schemes relying on multi-view images [50, 60, 94, 80, 72, 10, 70, 28, 40, 25, 11] or other sensors [54, 31, 27, 94] can ease this problem to some degree, we have to admit that paired images of different views or complementary sensors are not always available in most scenarios. The core problem is how to just utilize limited information in monocular images to extract the corresponding 3D poses of people. If a sequence of images (i.e. video) is available, the temporal context carried can be also helpful for the task [1, 81, 7, 47, 76]. They not only provide information to make up the missing joints caused by the occlusion and provide hints for joints' locations in the future, but also help group the detected joints belonging to a specific person in multi-person scenarios. So in this paper, we mainly focus on restoration from a sequence of monocular multi-person 2D skeletons to predict the corresponding 3D poses at a future time point." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Some related methods [64, 9, 75, 82, 41, 11, 98, 26, 13, 63, 36] usually applied graph neural networks to encode the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14700" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 550 + ], + "type": "text", + "content": "2D human joints and propagated the information through recurrent neural networks or directly used spatial-temporal graph convolutional networks. However, almost all of them only focused on single-person scenarios, while in most of cases, the relations among people in the scene are also helpful for the forecasting. Also, there exist several limitations in these works. Firstly, traditional recurrent neural networks, such as GRU or LSTM, all use discrete-time sequential models, which have a fixed propagation step and are not flexible enough to deal with different moving speeds of different human joints. The performance of these models will be compromised when the video frequency is not high or the cameras need to scan around in a certain period instead of shooting a fixed scene. For example, if the human pose is recorded every second in the video, discrete-time models cannot provide the forecasting result within 1 second. Instead, we propose to use neural ordinary differential equation (NeuralODE), a continuous-time model, to propagate information along the temporal axis in our framework. Secondly, we notice that any joint of a person is only connected to its nearest neighbors, which makes the adjacency matrix used in the graph convolution very sparse, and leads to inefficient calculations in turn. To reduce the meaningless calculations and increase the efficiency of our model, we propose to utilize a hierarchical structure to decompose a sparsely connected graph into several dense graphs and encode them in a cascade manner. By doing so, we can also deal with edges of different types separately, such as the edges between people and the edges connecting joints within each person. Lastly, the whole and fixed binary adjacency matrix is used for the graph convolution in the literature, although the actual connections of joints only occupy a very small part of the graph due to the occlusion. Instead, we propose to provide dynamic topology information at each time stamp. Different from works [21, 96, 84, 58, 42, 43] which learned another weight matrix from fully observed input, we propose to learn the weight matrix from partial input to more timely capture the effect of every specific connection in the graph and multiply it with the time-varying dynamic binary adjacency matrix to form the final one used in graph convolutions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "text", + "content": "To summarize, the main contributions of this paper are three folds:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 580, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 580, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 580, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 580, + 287, + 628 + ], + "type": "text", + "content": "- We propose a continuous-time framework which takes the sequence of 2D skeletons extracted from a multi-person video as the input to forecast the multi-person 3D poses in any future time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 635, + 288, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 635, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 58, + 635, + 288, + 683 + ], + "type": "text", + "content": "- We design our model as a hierarchical structure to reduce the computational complexity for multi-person scenarios, and also deal with different connection types in the graph." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We utilize dynamic graph convolutions to deal with the irregular input caused by occlusions." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "In the remaining of this paper, we will briefly introduce the related works in Sec. 2 and give the details of our model in Sec. 3. The performance of our model is demonstrated through experiments in Sec. 4. Finally, we will conclude our work in Sec. 5. The corresponding source code can be found at https://github.com/SBU-YCX/HDG-ODE." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 158, + 398, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 158, + 398, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 158, + 398, + 171 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 178, + 545, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 545, + 192 + ], + "type": "text", + "content": "2.1. 3D Pose Estimation, Tracking and Forecasting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 547, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 509 + ], + "type": "text", + "content": "Single-Person Pose Estimation: Most of the 3D Single-Person Pose Estimation (SPPE) methods are proposed to recover the 3D poses from 2D ones extracted by a pre-trained 2D pose extractor such as Stacked Hourglass Network [51] or Convolutional Pose Machines [69]. Given 2D coordinates as input, Zeng et al. [93] proposed to split the human joints into local groups and estimate them separately. The same idea was used by Chen et al. [15], although they directly estimated 3D poses from raw images and applied different architectures for different parts. In Li et al.'s work [39], different subsets were exchanged to generate augment data for better training. Choi et al. [18] and Azizi et al. [5] used spectral graph convolutions to tackle the problem due to the structure property of input skeletons. Since images captured from monocular cameras have perspective limitations, Yu et al. [91] proposed perspective crop layers to eliminate the perspective effect, while Nie et al. [52] tried to use siamese architectures to decouple the view-dependent representation and pose-dependent representation. There are also some other works solved the perspective problems with the help of multi-view images, such as [50, 60, 80, 72, 94], by fusing the 2D poses from different views to get the original 3D results. Besides, other wearable sensors were used to complement the missing depth information in [94, 27], and Isogawa et al. [31] proposed to utilize an optical non-line-of-sight imaging system to obtain the 3D poses from photon images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "content": "Multi-Person Pose Estimation: Inspired by 2D Multi-Person Pose Estimation (MPPE) methods, the works for the 3D estimation task can also be divided into two groups. One is top-down methods where a human detector is applied first and estimation is made for each person separately. Yang et al. [88] followed this line of work with graph convolutional networks (GCNs) [32], T. Xu and W. Takano [83] modified the stacked hourglass structure also with the graph convolutions. Li et al.'s [37] proposed a hierarchical architecture, which is the one closest to our work, but their goal was to augment the graph with additional labeled mesh data and integrate more information, which is totally different from our motivation. The other one is bottom-up approaches where joint locations are directly estimated, then associated and grouped into different people. In [24], Fabbri et al. proposed to use distance-based heuristic for the associations from the detected head joint with the highest confidence." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14701" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 156 + ], + "type": "text", + "content": "Cheng et al. [17] proposed to integrate both top-down and bottom-up branches to give better results. Besides, inspired by YOLO [59], some works [97, 6, 35, 49] explored to obtain the 3D poses in a single pass where a pivot joint was detected as well as the offsets for all other joints. Similar to single-person case, multi-view sources were also used in [25, 10, 70, 40, 28]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 158, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 158, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 158, + 287, + 443 + ], + "type": "text", + "content": "Pose Tracking and Forecasting: Since Andriluka et al. proposed PoseTrack [1] to first time introduce the Human Pose Tracking (HPT) task, contextual information has begun to be used for better estimating human poses in a sequence of inputs. Similar to other 3D pose tasks, most of works used extracted 2D pose sequence as the input and adopted top-down pipeline to mainly focus on single-person's pose tracking. Cai et al. [9] integrated spatio-temporal GCN into a stacked hourglass architecture to get the information of different scales. The same idea was applied in Wang et al.'s work [75], where the U-shaped GCN was used to extract multi-scale contexts. Liu et al. [41] proposed a graph attention spatio-temporal network with dilated temporal convolutions to get both local and global information, and Sofianos et al. [64] built a space-time-separable GCN for the 3D pose forecasting. Encoder-decoder architecture was utilized in [36, 63, 26] and Zheng et al. [98] solved the problem with spatio-temporal transformer. Furthermore, Chen et al. [11] proposed a fast method with multi-view inputs and Yuan et al. [92] utilized physics simulators and the reinforcement learning strategy in their work. Besides the skeleton joint representations, bone lengths and directions were used in [13, 82] and the extracted information was propagated through an LSTM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 454, + 183, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 454, + 183, + 467 + ], + "spans": [ + { + "bbox": [ + 49, + 454, + 183, + 467 + ], + "type": "text", + "content": "2.2. Graph Neural Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "text", + "content": "Different from traditional convolutional neural networks (CNNs), graph neural networks (GNNs) are mainly proposed to deal with data having more general structures. GNNs can be divided into two groups, spatial and spectral graph convolutions, as described in [8]. The former one considers the graph as locally connected and the information is only shared between each node and its nearest neighbors. J. Atwood and D. Towsley [4] extended it to allow the communications among nodes within " + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "text", + "content": "-hop (" + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "inline_equation", + "content": "k >= 1" + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "text", + "content": ") connections by using the probabilistic transitions. The latter one converts the data into spectral domain first, multiplies it with the Laplacian of the topology matrix in the spectral domain to mimic the convolutional operation in the spatial domain. Compared with spatial graph convolutoins, spectral operations utilize the global graph structure each time. However, since the kernel of the whole graph is large, to reduce the computational complexity, Defferrard et al. [23] proposed to approximate it with the " + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "text", + "content": " order polynomials of the Laplacian with the help of Chebyshev expansion to model the " + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 49, + 475, + 287, + 712 + ], + "type": "text", + "content": "-hop connections. T. Kipf and M. Welling [32] proposed graph convolutional" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 95 + ], + "type": "text", + "content": "networks (GCNs) to further simplify it with the " + }, + { + "bbox": [ + 308, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 308, + 72, + 545, + 95 + ], + "type": "text", + "content": " order approximation of the covolution in ChebNet [23]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 97, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 322 + ], + "type": "text", + "content": "In the pose estimation tasks, GCNs were widely used in [88, 83, 37, 9, 75, 41]. However, unlike CNNs, traditional GCNs have shared weights for different neighbors. To tackle this problems, another learnable weight matrix was trained in [96, 21, 42]. But, in these works, the weight matrix were learned with full joints and the final matrix used for convolution was the multiplication between the weight matrix and the static skeleton matrix assuming all joints observable. In our work, to deal with the missing case caused by occlusion, we make the skeleton matrix dynamic according to the current occlusion status so that the learnt weight matrix can also be effective even though only partial joints given. Some other works [43, 58, 84] proposed to add more semantic connections to the graph, which increased the calculation complexity of the models since the joints of all people could not be fully-connected. In addition, the semantic graphs were still globally sparse although locally dense. In this work, we propose to decompose the whole graph in a hierarchical manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 329, + 462, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 329, + 462, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 329, + 462, + 342 + ], + "type": "text", + "content": "2.3. Neural Differential Equation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 348, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 545, + 645 + ], + "type": "text", + "content": "As mentioned earlier, almost all the existing pose tracking methods were built based on spatio-temporal graph convolutional networks (ST-GCNs) and recurrent neural networks (RNNs), which could only model the discrete-time data where all human joints were assumed to be observable although there existed missing and inaccurate inputs due to occlusions. In order to better model the continuous-time process of the irregular data, neural ordinary differential equation (NeuralODE) was proposed by Chen et al. [12], where deep neural networks were utilized to parameterize a nonlinear ordinary differential equation (ODE). Rubanova et al. [61] and Brouwer et al. [22] further extended it by introducing a recurrent structure to efficiently encode the input information into the underlying trajectories. Some other works [44, 45, 46] also inserted Brownian terms into ODE to make it a stochastic model that can capture the uncertainties in the dynamics. In these works, pose estimation was also mentioned as a part of experiment to prove their efficiency. However, when they dealt with the human joint data, they merely considered each joint sequence as an isolated trajectory and ignored the relations among these joints. Graph neural networks were introduced to parameterize the ODEs in [56, 29, 57], but all of them utilized the whole graph, which didn't take the sparsity problem into consideration in multi-person scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 655, + 545, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 655, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 308, + 655, + 545, + 682 + ], + "type": "text", + "content": "3. Hierarchical Dynamic-Graph Ordinary Differential Equation Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 689, + 545, + 712 + ], + "type": "text", + "content": "In this section, we will introduce the details of our hierarchical ordinary differential equation model with dynamic" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14702" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "graph convolutions, which we call as HDG-ODE. As shown in Figure 1, the framework of our model follows a hierarchical structure, where each level is a sub-graph with the state changes modeled through graph ordinary differential equation and parameterized with dynamic graph convolutions. In the remaining of this section, we will describe the details of each component as well as the advantages of such designs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 165, + 229, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 229, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 229, + 177 + ], + "type": "text", + "content": "3.1. Problem Statement and Notations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": "Given a multi-personal video containing " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " people, and the 2D joints sequence up to time " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathbf{X}_1,\\mathbf{X}_2,\\dots,\\mathbf{X}_t\\}" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t\\in \\mathcal{R}^{[K,N_j,2]}" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " is the number of joints within each person. These 2D representations can be obtained from any existing 2D pose estimator. Our goal is to predict the corresponding 3D poses " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\tau}\\in \\mathcal{R}^{[K,N_j,3]}" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " at a future time " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " can be any value within " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "(t,t + 1]" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": ". Like in many 2D estimators, the missing or occluded joints are represented as some singular values or estimated with a very low confidence, based on this, we can directly obtain the corresponding occlusion mask " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": " at each time " + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 183, + 290, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 335, + 211, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 211, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 211, + 348 + ], + "type": "text", + "content": "3.2. Hierarchical Graph Structure" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 355, + 289, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 289, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 289, + 415 + ], + "type": "text", + "content": "As explored in many other works, compared to the traditional use of CNN, graph convolution is more suitable for human skeleton data due to the irregular connection of joints. The feature matrices " + }, + { + "bbox": [ + 46, + 355, + 289, + 415 + ], + "type": "inline_equation", + "content": "Z^{(l)}" + }, + { + "bbox": [ + 46, + 355, + 289, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 355, + 289, + 415 + ], + "type": "inline_equation", + "content": "Z^{(l + 1)}" + }, + { + "bbox": [ + 46, + 355, + 289, + 415 + ], + "type": "text", + "content": " of two adjacent convolution layers are expressed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 423, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 423, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 119, + 423, + 287, + 437 + ], + "type": "interline_equation", + "content": "Z ^ {(l + 1)} = \\sigma (W Z ^ {(l)} \\tilde {A}), \\tag {1}", + "image_path": "e272d32337fe1f5996eb36cba0531990fcc8c4c496f5ac94240f013b0e7a256c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "text", + "content": " is a learnable parameter matrix, " + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "inline_equation", + "content": "\\tilde{A}" + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "text", + "content": " is the normalized form of the adjacency matrix " + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "inline_equation", + "content": "A \\in [0,1]^{N \\times N}" + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "text", + "content": " is the number of nodes in the graph. However, when modeling human skeletons, we notice that each joint node is only connected to a few other joints, which will make the static skeleton-based " + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "inline_equation", + "content": "\\tilde{A}" + }, + { + "bbox": [ + 46, + 449, + 289, + 652 + ], + "type": "text", + "content": " very sparse and lead to plenty of meaningless calculations in (1). Although some recent works tried to learn more semantic connections among different joints besides static ones, the sparsity problem cannot be eliminated. The semantic relationship among joints usually exists within a limited number of hops instead of the whole graph. Even after we add those semantic edges, we can separate all joints into several groups such that the connections within each group are dense but the ones among different groups are still sparse. This inspires us to adopt a hierarchical structure instead of a flatten one to represent the graph of human skeletons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 713 + ], + "type": "text", + "content": "Besides, according to the physical property of human motion, different joints change their positions at different speeds over time. The location of the central joint usually has a slow and smooth change while the extremity joints, such as \"wrist\" and \"ankle\", have drastic and unexpected" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "changes. Therefore, rather than allocating the same portion of computational resource to all joints, we can focus more on the latter ones. The cascade processing along the hierarchical structure also allows us to utilize the relative-central ones' locations to guide the estimation of positions of the relative-extreme joints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 144, + 547, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 287 + ], + "type": "text", + "content": "Another reason for which we adopt the hierarchical structure is its advantage of dealing with different types of edges in the whole graph. For example, in the multi-person scenario, previous works commonly used top-down methods where each person in the scene would be processed individually, but actually, pose estimation and tracking can benefit from taking the relation among people into consideration. However, the edges describing the inter-person relations are different from those representing the intra-person limbs, and it will reduce the effect if we model them in one large graph. Intuitively, separating them into different groups and forming a graph at each level will be a better choice." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 287, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 372 + ], + "type": "text", + "content": "Specifically, as shown in Figure 2, for each frame of the video at time " + }, + { + "bbox": [ + 304, + 287, + 547, + 372 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 287, + 547, + 372 + ], + "type": "text", + "content": ", we get a set of 2D skeleton coordinates of all the people in the scene as input. To apply the hierarchical pipeline, we choose the \"pelvis\" joint as the root node of each person, whose coordinate can be also considered as the location of the person. All these root nodes form the 1st-level graph and encoded by a graph encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 376, + 545, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 376, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 365, + 376, + 545, + 392 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {\\tau} ^ {(1)} = G _ {1} \\left(\\mathbf {X} _ {t} ^ {(1)}, \\mathbf {H} _ {< t} ^ {(1)}, \\tilde {A} _ {t} ^ {(1)}\\right), \\tag {2}", + "image_path": "2da2041c011cd2d2103164600cf65ea414145c8f3847263c20a93e392eab416b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^{(1)}\\in \\mathcal{R}^{[K,1,2]}" + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "content": " represents the 2D coordinates of all root nodes of " + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "content": " people, " + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\tau}^{(1)}\\in \\mathcal{R}^{[K,1,3]}" + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "content": " is the corresponding 3D one forecasted for the future time " + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{< t}^{(1)}" + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "content": " is the historical feature, and " + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\tilde{A}_t^{(1)}" + }, + { + "bbox": [ + 304, + 399, + 547, + 595 + ], + "type": "text", + "content": " is the normalized adjacency matrix with the edges calculated based on the distances among connected roots, i.e. the relations among people. There is an edge between two people if their root nodes' distance is smaller than a threshold (the average of the heights of joints in 2D skeleton, i.e. distance from \"head\" to \"ankle\" joint in our setting), which indicates they are related. Then, all the limb-root joints, including \"neck\", \"left shoulder\", \"right shoulder\", \"left hip\" and \"right hip\", as well as the root node are grouped to form the 2nd-level graph within each person. To utilize the result of the 1st-level graph, the output of this level is designed as the spatial shifts of the limb-root joints' locations from the root node. We can express the process as below" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 599, + 520, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 599, + 520, + 614 + ], + "spans": [ + { + "bbox": [ + 330, + 599, + 520, + 614 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\tau} ^ {(2, i)} = G _ {2} \\left(\\mathbf {X} _ {t} ^ {(2, i)}, \\mathbf {Y} _ {\\tau , i} ^ {(1)}, \\mathbf {H} _ {< t} ^ {(2, i)}, \\tilde {A} _ {t} ^ {(2, i)}\\right), i \\in [ 1, K ]", + "image_path": "91ec3cfbda8a176945a1613f485a19378d6f0e7fb929b7849a23f3359370cdea.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 616, + 545, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 616, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 330, + 616, + 545, + 630 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {\\tau} ^ {(2, i)} = \\mathbf {D} _ {\\tau} ^ {(2, i)} + \\mathbf {Y} _ {\\tau , i} ^ {(1)}, \\tag {3}", + "image_path": "eb1f672f03bf6ef102995cb88df2936cbc4548c7234525a317e07677d551949d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^{(2,i)}\\in \\mathcal{R}^{[1,6,2]}" + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "text", + "content": " is the 2D coordinates vector of all limb-root nodes as well as root node of person " + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\tau ,i}^{(1)}" + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "text", + "content": " is the 3D root node's position of this person obtained by (2), " + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\tau}^{(2,i)}" + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\tau}^{(2,i)}" + }, + { + "bbox": [ + 304, + 636, + 547, + 713 + ], + "type": "text", + "content": " represent respectively the future 3D position shift and final position predicted. Finally, we group each limbroot joint to the corresponding limb-extreme joint to form" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14703" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 521, + 266 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 521, + 266 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 521, + 266 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 521, + 266 + ], + "type": "image", + "image_path": "467062f85237c008b503f50662a640fbf8192634ab81d8ed1949e26121aaca02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "text", + "content": "Figure 1. The framework of HDG-ODE. " + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t_n}" + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "text", + "content": " are the 2D joints coordinates at time " + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{t_n}" + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "text", + "content": " is the occlusion mask obtained from " + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t_n}" + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\tau}" + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "text", + "content": " are the predicted 3D coordinates at the future time " + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 271, + 545, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 313, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 335 + ], + "type": "text", + "content": "the 3rd-level graph. Similarly, the encoding process within each lamb group " + }, + { + "bbox": [ + 46, + 313, + 287, + 335 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 313, + 287, + 335 + ], + "type": "text", + "content": " of person " + }, + { + "bbox": [ + 46, + 313, + 287, + 335 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 313, + 287, + 335 + ], + "type": "text", + "content": " can be expressed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 337, + 274, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 337, + 274, + 352 + ], + "spans": [ + { + "bbox": [ + 59, + 337, + 274, + 352 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\tau} ^ {(3, i, j)} = G _ {3} (\\mathbf {X} _ {t} ^ {(3, i, j)}, \\mathbf {Y} _ {\\tau , j} ^ {(2, i)}, \\mathbf {H} _ {< t} ^ {(3, i, j)}, \\tilde {A} _ {t} ^ {(3, i, j)}), j \\in [ 1, 5 ]", + "image_path": "1a8e30b1952b9a4f31267a9766eb6c79589bf26f9c4233278d1b89929c9310cd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 353, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 353, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 353, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {\\tau} ^ {(3, i, j)} = \\mathbf {D} _ {\\tau} ^ {(3, i, j)} + \\mathbf {Y} _ {\\tau , j} ^ {(2, i)}, \\tag {4}", + "image_path": "d7f69952bb63a82652da0b4b858e2425cc7dcca5e745ee4b1facea51e233318a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^{(3,i,j)}\\in \\mathcal{R}^{[1,3,2]}" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " is the vector of the 2D coordinates representing all nodes of person " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " within the limb group " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\tau ,j}^{(2,i)}" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " is the 3D location of the lamb-root " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " calculated by (3), " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{< t}^{(3,i,j)}" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{A}_t^{(3,i,j)}" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " represent the corresponding part of the historical feature and the adjacency matrix in this level. 3D shift " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\tau}^{(3,i,j)}" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " is used to get the final 3D coordinates " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\tau}^{(3,i,j)}" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": " at future time " + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 373, + 287, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 80, + 471, + 166, + 596 + ], + "blocks": [ + { + "bbox": [ + 80, + 471, + 166, + 596 + ], + "lines": [ + { + "bbox": [ + 80, + 471, + 166, + 596 + ], + "spans": [ + { + "bbox": [ + 80, + 471, + 166, + 596 + ], + "type": "image", + "image_path": "b0181a44ee320cd1948ffc81268471da4d1f4eb12dd6ce9120ed1e043183479d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 597, + 127, + 605 + ], + "lines": [ + { + "bbox": [ + 118, + 597, + 127, + 605 + ], + "spans": [ + { + "bbox": [ + 118, + 597, + 127, + 605 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 170, + 470, + 257, + 597 + ], + "blocks": [ + { + "bbox": [ + 170, + 470, + 257, + 597 + ], + "lines": [ + { + "bbox": [ + 170, + 470, + 257, + 597 + ], + "spans": [ + { + "bbox": [ + 170, + 470, + 257, + 597 + ], + "type": "image", + "image_path": "683fb5334ee6e4109990615963615a7d16c1e3913807e2f2d94003cd47119634.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 597, + 217, + 605 + ], + "lines": [ + { + "bbox": [ + 208, + 597, + 217, + 605 + ], + "spans": [ + { + "bbox": [ + 208, + 597, + 217, + 605 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 608, + 287, + 641 + ], + "lines": [ + { + "bbox": [ + 46, + 608, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 287, + 641 + ], + "type": "text", + "content": "Figure 2. Illustration of the hierarchical graph structure: (a) the skeleton representations of a multi-person scenario; (b) the corresponding 3-level hierarchical structure." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 274, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 274, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 274, + 672 + ], + "type": "text", + "content": "3.3. Time-Varying Dynamic Graph Convolution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "A most important impact factor to the performance of human pose estimation is the occlusion, no matter the occlusion is caused by other objects or the self-occlusion happens" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 313, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 547, + 456 + ], + "type": "text", + "content": "within a person. To get the 2D skeleton inputs with some 2D pose estimation methods, the visible joints can be directly extracted from the raw images while the occluded ones are missing. Even though the missing joints can be estimated by some intuition, they are less accurate due to the uncertainties. When we use such 2D poses as inputs to recover the corresponding 3D poses, called the lifting process, the inaccuracy will accumulate and compromises data processing during the information communication within the graph convolution operation. Therefore, we hope our graph encoder can be reasonably adaptive to deal with the missing and inaccurate cases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 456, + 546, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 546, + 589 + ], + "type": "text", + "content": "In the literature work that apply the traditional graph convolution for this task, the adjacency matrix " + }, + { + "bbox": [ + 304, + 456, + 546, + 589 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 456, + 546, + 589 + ], + "type": "text", + "content": " and its normalized form " + }, + { + "bbox": [ + 304, + 456, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{A}" + }, + { + "bbox": [ + 304, + 456, + 546, + 589 + ], + "type": "text", + "content": " in (1) usually keep constant to represent the static connections within the graph, i.e. all the joints are assumed to be fully available, no matter they are observable or occluded. However, as we described, only those visible ones are accurate and worth being integrated to guide the forecasting. So we propose to use the time-varying dynamic graph convolution to only encode visible joints input at each timestamp. Inspired by some literature work [96], the learning of a hybrid adjacency-weight matrix may be written as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 605, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 605, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 358, + 605, + 545, + 622 + ], + "type": "interline_equation", + "content": "Z _ {t} ^ {(l + 1)} = \\sigma \\left(W Z _ {t} ^ {(l)} \\rho \\left(W _ {\\tilde {A}} \\odot \\tilde {A}\\right)\\right), \\tag {5}", + "image_path": "42ba2a06b78239e800f552c7960c0bb6410c91840f3deb5a6ac9da8f0f973acd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "inline_equation", + "content": "W_{\\tilde{A}}" + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "text", + "content": " is a learnable weight matrix applied to assign the information along different edges with different weights to solve the weight-sharing problem [21, 42] and " + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\rho(\\cdot)" + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "text", + "content": " is a nonlinear function. However, we notice that in (5), " + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "inline_equation", + "content": "W_{\\tilde{A}}" + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "text", + "content": " is learnt based on the fully-observable data and the " + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{A}" + }, + { + "bbox": [ + 304, + 628, + 547, + 713 + ], + "type": "text", + "content": " is the static one without taking the occlusion cases into consideration. In other words, although the learnt weights work well when all" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14704" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "text", + "content": "joints are available, they may not be effective if partial of joints are occluded and missing. Based on this, we make a further step and propose to also take the occlusion status into consideration when obtaining the dynamic weight matrix. Specifically, at each timestamp " + }, + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "text", + "content": ", we will drop the occluded joints in the original static adjacency matrix " + }, + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "text", + "content": " to form a time-varying adjacency matrix " + }, + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 45, + 72, + 289, + 178 + ], + "type": "text", + "content": ", and this dynamic adjacency matrix will be used in both training and testing, i.e. (5) can be rewritten as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 184, + 288, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 184, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 97, + 184, + 288, + 201 + ], + "type": "interline_equation", + "content": "Z _ {t} ^ {(l + 1)} = \\sigma \\left(W Z _ {t} ^ {(l)} \\rho \\left(W _ {\\tilde {A}} \\odot \\tilde {A} _ {t}\\right)\\right), \\tag {6}", + "image_path": "f97d30d232f997470365b103773ac179cbbd2fba8e1592797fa4ecb085a9ffa4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{A}_t" + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "content": " is the normalized form of the real adjacency matrix at time " + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "content": ". By doing so, the learning of the weight matrix " + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "inline_equation", + "content": "W_{\\tilde{A}}" + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "content": " will only focus on the observable joints to make sure it can also be effective when only partial joints are available. In the testing phase, the hybrid-matrix " + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "inline_equation", + "content": "W_{\\tilde{A}} \\odot \\tilde{A}_t" + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "content": " can ignore the missing joints and adaptively adjust the weights along existing edges for information integration among joints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 299, + 257, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 257, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 257, + 312 + ], + "type": "text", + "content": "3.4. Parallel Ordinary Differential Equation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 317, + 288, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 288, + 400 + ], + "type": "text", + "content": "Besides the partial occlusion, which leads to incomplete graph input at each frame and can be regarded as spatial missing, there are also many cases where a person is totally occluded, and causes the temporal missing within a sequence. Previous works almost rely on the discrete-time sequential models, which are not flexible enough since the missing in real world can vary from several frames to dozens of frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 437 + ], + "type": "text", + "content": "To make the model more robust, we adopt ordinary differential equation (ODE), a continuous-time model in this paper, which can be expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 442, + 190, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 442, + 190, + 466 + ], + "spans": [ + { + "bbox": [ + 126, + 442, + 190, + 466 + ], + "type": "interline_equation", + "content": "\\frac {d \\mathbf {H} _ {t}}{d t} = F (\\mathbf {H} _ {t}),", + "image_path": "7fe63f5029dff2737aced25bdeba541763f1dc68a326086f46e9f1abe8dc54a1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 468, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 468, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 102, + 468, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {t} = \\mathbf {H} _ {t _ {n - 1}} + \\int_ {t _ {n - 1}} ^ {t} F (\\mathbf {H} _ {\\tau}) d \\tau . \\tag {7}", + "image_path": "20ae3763046a4a932df01e22412cba0d7e31d648a8b00abf4f54ccc3af257c48.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 502, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 573 + ], + "type": "text", + "content": "Compared with traditional recurrent networks, it can be used to predict the values over the interval of any length. Specially, in our model, we use time-varying dynamic graph convolution mentioned in Sec. 3.3 to parameterize the derivatives " + }, + { + "bbox": [ + 47, + 502, + 287, + 573 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 47, + 502, + 287, + 573 + ], + "type": "text", + "content": ". For efficiency, we compute the integration in (7) by Euler Method and we can rewrite it as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 580, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 580, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 85, + 580, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {t, F} = \\mathbf {H} _ {t - \\Delta t, F} + F \\left(\\mathbf {H} _ {t - \\Delta t, F}, \\tilde {A}\\right) \\Delta t, \\tag {8}", + "image_path": "a2f05e4b6c21a8fb1283a020699c9b9b1ac6adc7a8e609c4673a1398facf8e31.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{t,F}" + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "content": " is the hidden feature, " + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{A}" + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "content": " is the normalized adjacency matrix assuming all joints observable, and " + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "content": " is the step-size. At an observation time " + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "content": ", the features of observable joints are updated through the same graph convolution encoder as described in Sec. 3.2, i.e." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 95, + 669, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 669, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 95, + 669, + 287, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {t _ {n}, G} = G \\left(\\mathbf {X} _ {t _ {n}}, \\mathbf {Y} _ {t _ {n}}, \\mathbf {H} _ {< t _ {n}}, \\tilde {A} _ {t _ {n}}\\right), \\tag {9}", + "image_path": "49a5c39a319d49f79cc189aa1b5d2dc92a8c4fb4421c90b59ac80a294a896963.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "G = \\{G_1, G_2, G_3\\}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " represents the cascade process in (2)-(4) parameterized by our time-varying dynamic graph" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "convolution networks, " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t_n}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{t_n}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " are the 2D and 3D joint locations, " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{< t_n} = \\mathbf{H}_{t_n - \\Delta t,F}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " is the feature from the ODE model and " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\tilde{A}_{t_n}" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " is the normalized adjacency matrix of the graph formed by observable joints. After the update, the new feature at the time instant " + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": " becomes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 143, + 547, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 143, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 324, + 143, + 547, + 157 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {t _ {n}} = \\mathbf {M} _ {t _ {n}} \\odot \\mathbf {H} _ {t _ {n}, F} + \\left(\\mathbf {1} - \\mathbf {M} _ {t _ {n}}\\right) \\odot \\mathbf {H} _ {t _ {n}, G}, \\tag {10}", + "image_path": "3228266da80b174989159647a2c8decae779a26785d1e451eb564f4372e55e3d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{t_n}" + }, + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "text", + "content": " is the occlusion mask of the input described in Sec. 3.1. Here, we want to emphasize that " + }, + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "inline_equation", + "content": "\\tilde{A}_{t_n}" + }, + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "text", + "content": " in (9) and " + }, + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "inline_equation", + "content": "\\tilde{A}" + }, + { + "bbox": [ + 304, + 165, + 547, + 261 + ], + "type": "text", + "content": " in (8) may not be the same according to the observation states of input. As we described in last section, the input may contain missing values, but all the missing values have been completed by the encoder in the encoding process and we can assume the graph is fully observable in the propagation process." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 326, + 273, + 526, + 363 + ], + "blocks": [ + { + "bbox": [ + 326, + 273, + 526, + 363 + ], + "lines": [ + { + "bbox": [ + 326, + 273, + 526, + 363 + ], + "spans": [ + { + "bbox": [ + 326, + 273, + 526, + 363 + ], + "type": "image", + "image_path": "02f0ae8ba3f13ea4c5935e91071fac87e299de3555ec1a49cde77b910f72fbb6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 366, + 517, + 378 + ], + "lines": [ + { + "bbox": [ + 334, + 366, + 517, + 378 + ], + "spans": [ + { + "bbox": [ + 334, + 366, + 517, + 378 + ], + "type": "text", + "content": "Figure 3. Illustration of the parallel ODE pipeline." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "text", + "content": "Here, we can further demonstrate the advantages of hierarchical structure. For efficiency, we adopt Euler method to approximately calculate the integration, but as we know, it will predict the curve as a straight line if we make the propagation step " + }, + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "text", + "content": " too large and equal to the interval between two observations as " + }, + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "inline_equation", + "content": "t_n - t_{n-1}" + }, + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "text", + "content": ". To make the model more expressive, we usually make " + }, + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 390, + 547, + 474 + ], + "type": "text", + "content": " smaller, i.e." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 382, + 482, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 482, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 382, + 482, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\Delta t = \\frac {1}{k} \\left(t _ {n} - t _ {n - 1}\\right). \\tag {11}", + "image_path": "79b5062499564a7444b5519f4db271d0eac6920edffd0b45bf6a75a430f00dd9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "type": "text", + "content": "However, as we discussed in Sec. 3.2, different joints of human body have different rates of location changes, we don't need to set " + }, + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 513, + 547, + 586 + ], + "type": "text", + "content": " to the same value at different levels in our model. Instead, we can adopt a parallel pipeline and make it small in shallow levels while enlarge it as the level goes deeper as shown in Figure 3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 596, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 596, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 388, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 616, + 426, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 426, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 426, + 628 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 634, + 375, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 375, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 375, + 645 + ], + "type": "text", + "content": "4.1.1 Datasets" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "MuPoTS-3D[49]: This multi-person dataset contains sequences captured in 20 real scenes. Each sequence has up to three subjects, as well as provides raw images, 2D pose annotations and the corresponding 3D ones. The lengths of data vary from 125 frames to 800 frames, and we split them" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14705" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 155 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 155 + ], + "type": "text", + "content": "into segments with the length of 50 frames. We divide all the data into training, validation and testing sets with the ratio of 0.8: 0.1: 0.1. The real occlusion mask are given as well, and the final data are obtained by the element-wise multiplication between the full data and the occlusion masks. Before being fed into the models, data from each sequence are first normalized." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 289, + 251 + ], + "type": "text", + "content": "Human3.6M[30]: This dataset consists of 3.6 million of single-person frames. 11 subjects conducting 15 activities are captured by cameras. Both 2D and 3D annotations are provided where 2D ones come from cameras while 3D ones are obtained using a motion capture system. Different from commonly used training methods, we only use data from one camera to train the model. We performed the same data partition and normalization as above." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 266, + 162, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 162, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 162, + 278 + ], + "type": "text", + "content": "4.1.2 Evaluation Metrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 319 + ], + "type": "text", + "content": "Mean Per Joint Position Error (MPJPE): This metric measures the average position error over all joints and frames in the sequence, which can be expressed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 325, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 352 + ], + "type": "interline_equation", + "content": "M P J P E = \\frac {1}{N _ {f}} \\sum_ {f} \\frac {1}{N _ {j} ^ {(f)}} \\sum_ {j} \\left\\| P _ {\\text {p r e d}} ^ {(f)} (j) - P _ {g t} ^ {(f)} (j) \\right\\| _ {2}, \\tag {12}", + "image_path": "c2f8cdf5a584171d158d791e494c3f8eabdff462c58eb952ac6e9e142c848344.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": " is the total number of frames, " + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "inline_equation", + "content": "N_{j}^{(f)}" + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": " is the total number of joints in the frame " + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "inline_equation", + "content": "P_{gt}^{(f)}(j)" + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "inline_equation", + "content": "P_{pred}^{(f)}(j)" + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": " are the ground-truth and the predicted 3D position of the joint " + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": " in the frame " + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 361, + 287, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "content": "Percentage of Correct Keypoints (PCK): PCK calculates the percentage of correctly predicted key-points, where a key-point " + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "content": " in the frame " + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "content": " is considered to be correct if the distance " + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "inline_equation", + "content": "d_{p}^{(f)}(j)" + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "content": " between the 3D ground-truth " + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "inline_equation", + "content": "P_{p,gt}^{(f)}(j)" + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "content": " and the predicted 3D position " + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "inline_equation", + "content": "P_{p,pred}^{(f)}(j)" + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "content": " are smaller than a person-specific threshold " + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "inline_equation", + "content": "S_{p}^{f}" + }, + { + "bbox": [ + 47, + 413, + 288, + 491 + ], + "type": "text", + "content": ", i.e." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 498, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 498, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 88, + 498, + 287, + 529 + ], + "type": "interline_equation", + "content": "P C K = \\frac {\\sum_ {f} \\sum_ {p} \\sum_ {j} \\delta \\left(d _ {p} ^ {(f)} (j) < S _ {p} ^ {(f)}\\right)}{\\sum_ {f} \\sum_ {p} \\sum_ {j} 1}, \\tag {13}", + "image_path": "94fc24ecc96987d3d6f2a75ff6587c7363488218c741737dea7bdf9ef4b3c5b3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 534, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 582 + ], + "type": "text", + "content": "Here we adopt the setting in MPII dataset[2] where the threshold is the ratio of head link's length " + }, + { + "bbox": [ + 47, + 534, + 287, + 582 + ], + "type": "inline_equation", + "content": "h_p^{(f)}" + }, + { + "bbox": [ + 47, + 534, + 287, + 582 + ], + "type": "text", + "content": " of the person " + }, + { + "bbox": [ + 47, + 534, + 287, + 582 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 534, + 287, + 582 + ], + "type": "text", + "content": ", i.e. the distance between \"head\" joint and \"upper neck\", and the PCK can be represented as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 589, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 589, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 85, + 589, + 287, + 625 + ], + "type": "interline_equation", + "content": "P C K h @ \\alpha = \\frac {\\sum_ {f} \\sum_ {p} \\sum_ {j} \\delta \\left(\\frac {d _ {p} ^ {(f)} (j)}{h _ {p} ^ {(f)}} < \\alpha\\right)}{\\sum_ {f} \\sum_ {p} \\sum_ {j} 1}, \\tag {14}", + "image_path": "03da82ec202f6f21b4b8337d3bae747ad89e2a9435dc919dab5d6bc3293728cd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 632, + 171, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 171, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 171, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 632, + 171, + 643 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 632, + 171, + 643 + ], + "type": "text", + "content": " is the ratio parameter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 658, + 180, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 180, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 180, + 670 + ], + "type": "text", + "content": "4.1.3 Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": "As introduced in Sec. 3.2, the hierarchical graph has 3 levels, containing " + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 677, + 287, + 712 + ], + "type": "text", + "content": ", 6 and 3 nodes each, and our HDG-ODE model has a hierarchical structure of 3 levels, with an encoder and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "an ODE at each level. For each encoder, we use a single-layer graph-GRU, using our time-varying dynamic graph convolutions with the feature size of 6. ODE models are parameterized with graph neural networks of 1, 2 and 3 convolutional layers at the three levels respectively, with the hidden feature size of 16. The propagation steps for the three levels are respectively " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "[1/30s, 1/60s, 1/120s]" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": ", and the data sampling interval is " + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "inline_equation", + "content": "1/30s" + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": ". The experiments are conducted on single NVIDIA 1080 Ti GPU. During the training, we use the Adam() optimizer with the learning rates of [0.01, 0.001, 0.0001] to minimize the mean squared error (MSE) between the predicted 3D positions and the ground-truth of the visible joints, and 100 epochs are run at each learning rate. To speed up the training, it will be early-stopped if there is no further loss decrease in 20 epochs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 259, + 421, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 421, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 421, + 272 + ], + "type": "text", + "content": "4.2. Experiment Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 277, + 429, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 429, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 429, + 289 + ], + "type": "text", + "content": "4.2.1 Overall Performance" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 297, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 380 + ], + "type": "text", + "content": "We conduct experiments over MuPoTS-3D dataset to evaluate the performance of our proposed model on the multiperson 3D pose tracking task. For comparison, we choose some representative methods, including discrete-time recurrent models and continuous-time ones, as the baselines. Specifically we take the following methods into consideration:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 388, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 317, + 388, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 547, + 497 + ], + "type": "text", + "content": "- STGCN: This discrete-time model consists of several concatenated spatio-temporal convolutional blocks (ST-Conv blocks), which is a common structure that is widely used for action recognition [85], pose forecasting [75] and other prediction tasks [89]. In our experiments, STGCN has 2 ST-Conv blocks, and recurrently processes each clip within the whole 2D pose sequence and give the 3D predictions. The output dimension of each layer in the blocks is 32." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 505, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 547, + 578 + ], + "type": "text", + "content": "- Graph-GRU: It is a variant of conventional Gated Recurrent Unit (GRU), where the linear operation in the conventional GRU is replaced by graph convolution. It works as the backbone in [90]. We construct the baseline with 2 Graph-GRU layer and a linear decoder. The output dimension of each layer in the blocks is 32." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 585, + 547, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 585, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 585, + 547, + 656 + ], + "type": "text", + "content": "- SemGCGRU: It is an improved version of traditional Graph-GRU, where the fixed binary adjacency matrix is replaced by a learnable weight matrix to capture the relations among nodes [96]. In our implementation, the structure and the feature size are the same as Graph-GRU." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 665, + 547, + 714 + ], + "type": "text", + "content": "- ODE-RNN: In this model [62], the deep learning modules, specifically dense neural networks, are incorporated to parameterize a non-linear ordinary differential equation (ODE) to model the continuous-time process" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14706" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 107 + ], + "type": "text", + "content": "of data sequence. In our experiments, the dense network contains 2 linear layers and the dimension of hidden features is 32." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 115, + 288, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 115, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 59, + 115, + 288, + 224 + ], + "type": "text", + "content": "- Graph-ODE: It is a GNN variant of ODE-RNN, with a GCN to learn the derivative of ODE and a GRU to integrate the information of observations. In our model comparison, the GCN has 2 hidden layers with ReLU activation functions and the output dimension of each hidden layer is 32. The output dimension of GRU is 32 as well. In other words, to compare with discrete models, we just replace the GRU module in Graph-GRU with an ODE module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 289, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 316 + ], + "type": "text", + "content": "Although continuous-time models can provide predictions at any time, even smaller than the data sampling rate, we set the step of them to be equal to the original data interval used in discrete-time models for fair comparison. From the results in Table 1, we can see that our HDG-ODE can achieve the best result with the smallest mean prediction error and the highest percentage of corrected prediction." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 326, + 298, + 467 + ], + "blocks": [ + { + "bbox": [ + 48, + 326, + 298, + 467 + ], + "lines": [ + { + "bbox": [ + 48, + 326, + 298, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 298, + 467 + ], + "type": "table", + "html": "
MPJPE(↓)PCKh@α(↑)
α=0.1α=0.5α=1.0
Discrete
STGCN [89, 75]0.600413.03%60.35%77.76%
Graph-GRU [90]0.557314.86%62.39%80.04%
SemGCGRU [96]0.437517.09%69.61%85.96%
Continuous
ODE-RNN [61]0.439620.21%71.78%85.45%
Graph-ODE [56]0.406619.80%71.57%87.00%
Ours
HDG-ODE0.303829.72%78.67%92.18%
", + "image_path": "e50cc07ad49685a6210e64be3799879878d499628dbefcb3807903837d0ec8c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 503, + 145, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 145, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 145, + 515 + ], + "type": "text", + "content": "4.2.2 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 629 + ], + "type": "text", + "content": "In this part, we will separately demonstrate the effectiveness of each designing component in our proposed model by some experiment results, including the continuous ODE model, the pipeline taking all people into consideration, the new time-varying dynamic graph convolution and the proposed hierarchical structure. The evaluation is sequential by adding the components one by one, and each time we only choose the current best model to conduct the experiment for the next component." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Discrete-Time Model vs. Continuous-Time Model: From Table 1, compared to traditional discrete-time models, no matter our HDG-ODE or other continuous-time models have a better performance. Different from discrete models which only execute one propagation between two observations, the continuous-time ones conduct multiple-step propagation so that the information on each joint will be affected" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 143 + ], + "type": "text", + "content": "and corrected by its neighbors several times before getting the final prediction. Also, due to the property of ODE, the continuous-time model can provide predictions at any time in the interval between two observations, instead of just obtaining predictions at the observation instants as done by discrete-time ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 144, + 548, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 548, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 548, + 264 + ], + "type": "text", + "content": "Isolated Top-Down Pipeline vs. Relational Top-Down Pipeline: As we described previously, most of existing multiperson pose tracking works adopt the top-down processing pipeline, where each person in the scenario is tracked separately as an isolated trajectory. Instead, we take into consideration the relation among people to infer their relative movements, which in turn affects the future positions of specific joints. In Table 2, by comparing the models in isolated top-down pipeline with the ones in our relational pipeline, we can observe the improvement." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 269, + 550, + 331 + ], + "blocks": [ + { + "bbox": [ + 47, + 467, + 288, + 477 + ], + "lines": [ + { + "bbox": [ + 47, + 467, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 288, + 477 + ], + "type": "text", + "content": "Table 1. Testing Performance of different models on MuPoTS-3D." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 269, + 550, + 331 + ], + "lines": [ + { + "bbox": [ + 306, + 269, + 550, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 269, + 550, + 331 + ], + "type": "table", + "html": "
Isolated PipelineRelational Pipeline
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.610477.72%0.600477.76%
Graph-GRU0.573878.46%0.557380.04%
Graph-ODE0.420386.89%0.406687.00%
", + "image_path": "62dee769535a46191f27a9ad55b3f88dd0a94ff887406012f782b9edc98fbe9d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 357, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 489 + ], + "type": "text", + "content": "Static Graph Convolution vs. Time-Varying Dynamic Graph Convolution: In the real cases, occlusion is inevitable. To better tackle the problem caused by the input data with missing or inaccurate values due to occlusions, we propose to replace the static graph convolution used in most of works with our time-varying dynamic graph convolution in the HDG-ODE model. As shown in Table 3, the dynamic graph convolution helps eliminate the negative influence of inaccurate inputs as well as adjust the integration weights of edges to the most suitable ones according to occlusion states, so the models can have a better performance than before." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 306, + 495, + 550, + 548 + ], + "blocks": [ + { + "bbox": [ + 306, + 331, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 547, + 353 + ], + "type": "text", + "content": "Table 2. Comparison between the isolated top-down pipeline and our relational top-down pipeline on MuPoTS-3D." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 495, + 550, + 548 + ], + "lines": [ + { + "bbox": [ + 306, + 495, + 550, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 550, + 548 + ], + "type": "table", + "html": "
Original Graph ConvOur Graph Conv
MPJPEPCKh@1.0MPJPEPCKh@1.0
Graph-GRU0.557380.04%0.384486.45%
Graph-ODE0.406687.00%0.370688.74%
", + "image_path": "f1ec8507bb8b45ac36316d671413c0b5f5e345434d9f641eda0d4801d0e4bfbd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 548, + 547, + 578 + ], + "lines": [ + { + "bbox": [ + 306, + 548, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 547, + 578 + ], + "type": "text", + "content": "Table 3. Comparison between original static graph convolutions (GC) and our time-varying dynamic convolutions (DG) on MuPoTS-3D." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Flatten Structure vs. Hierarchical Structure: To solve the sparsity problem explained in Sec. 3, our HDG-ODE adopts a hierarchical structure. In Table 4, compared to models in the flatten structure with a large sparse graph, the corresponding hierarchical ones perform better for models with graph convolutions. The reasons leading to the improvement lie in two aspects: On one hand, by decomposing the whole sparse graph into a set of relatively dense ones, the parameters to learn are reduced, as shown in Table 5. This helps save a significant amount of computational resource, and the released resource can be reallocated to the joints" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14707" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "with more complicated movements, e.g. making the corresponding network deeper or reducing the propagation step of ODE. On the other hand, the cascade processing along the hierarchical structure can provide the guidance among the levels. As for the ODE-RNN, the hierarchical one has a worse result because the structure is mainly proposed to deal with the graph sparsity problem. However, ODE-RNN doesn't utilize the graph information and only cares about isolated joint trajectories, so the hierarchical structure will make the model more complex and reduce the performance." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 198, + 288, + 258 + ], + "blocks": [ + { + "bbox": [ + 48, + 198, + 288, + 258 + ], + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 258 + ], + "type": "table", + "html": "
Flatten StructureHierarchical Structure
MPJPEPCKh@1.0MPJPEPCKh@1.0
ODE-RNN0.439685.45%0.493882.40%
GC-ODE0.406687.00%0.345190.51%
DG-ODE0.370688.74%0.303892.18%
", + "image_path": "a99c52fe7a807f62df162b44a34922447157668ab83cd9ec3fc93b46adc0f516.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 83, + 300, + 253, + 338 + ], + "blocks": [ + { + "bbox": [ + 47, + 259, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 259, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 287, + 281 + ], + "type": "text", + "content": "Table 4. Performance Comparison between the flatten structure and the hierarchical structure on MuPoTS-3D." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 300, + 253, + 338 + ], + "lines": [ + { + "bbox": [ + 83, + 300, + 253, + 338 + ], + "spans": [ + { + "bbox": [ + 83, + 300, + 253, + 338 + ], + "type": "table", + "html": "
#Params (↓)Memory (↓)
DG-ODE106490.76M
HDG-ODE26430.66M
", + "image_path": "feac483372cd67ec85c604fc7c5ebd01a377527a6fd461abde1ba798a7198d87.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 377, + 167, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 167, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 167, + 389 + ], + "type": "text", + "content": "4.2.3 Robustness Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "text", + "content": "Besides the multi-person scenarios, we also conduct experiments of single-person scenarios on Human3.6M dataset. Different from MuPoTS-3D, no occlusion information is provided within the data, so we make occlusion masks by ourselves and test the robustness of our model under different observable ratios. Specifically, for a sequence, we assume only " + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "text", + "content": " of the frames are observable while in the remaining ones, the input skeletons are totally missing due to the occlusion by buildings or other large objects. For each observable frame, we assume there exist self-occlusions such that only " + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "text", + "content": " of the joints are available. To make the corresponding mask, both " + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "text", + "content": " of the frames and " + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "text", + "content": " of the joints on each frame are randomly chosen and the total observable ratio can be expressed as " + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "inline_equation", + "content": "p_t \\times p_s" + }, + { + "bbox": [ + 46, + 396, + 289, + 612 + ], + "type": "text", + "content": ". The results are shown in Table 6. When the observable ratio decreases as a result of occlusions, the performance gets worse for all models. However, our HDG-ODE always performs the best among all models studied under all circumstances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 621, + 119, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 119, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 119, + 634 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "In this paper, we propose a new continuous-time model called HDG-ODE to forecast the future 3D human pose representations in multi-person videos given the 2D pose sequence as input. By using ordinary differential equations and dynamic graph convolutions, we alleviate the occlusion problem brought by perspective effects of monocular images." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 67, + 549, + 152 + ], + "blocks": [ + { + "bbox": [ + 46, + 338, + 288, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 360 + ], + "type": "text", + "content": "Table 5. Complexity Comparison between the flatten Dynamic-Graph ODE (DG-ODE) and our hierarchical one (HDG-ODE)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 67, + 549, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 67, + 549, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 67, + 549, + 152 + ], + "type": "table", + "html": "
pt=0.8, ps=0.5pt=0.6, ps=0.4
MPJPEPCKh@1.0MPJPEPCKh@1.0
STGCN0.396680.19%0.437675.82%
Graph-GRU0.372383.34%0.418877.39%
ODE-RNN0.352784.33%0.382680.66%
Graph-ODE0.347383.98%0.401978.54%
HDG-ODE0.305288.29%0.354382.36%
", + "image_path": "40151a90215901480d582d0a2771deb6f3b71c7c37d1f7fabc9e3181e9f10862.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 152, + 545, + 174 + ], + "lines": [ + { + "bbox": [ + 306, + 152, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 545, + 174 + ], + "type": "text", + "content": "Table 6. Performance of different models on Human3.6M with different observable ratios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 192, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 547, + 312 + ], + "type": "text", + "content": "Besides, with the help of hierarchical structure, we reduce the computational complexity of our model and are able to allocate more resources to the human joints which have more complicated movement. Our experiments demonstrate that our model can well recover the corresponding 3D poses from the single-view 2D poses and outperform literature works on the human pose forecasting task. Currently, the maximum number of people in the scenes of the video is limited. In the future, we will make our model more adaptive and test it in more crowded scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 323, + 365, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 323, + 365, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 365, + 335 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 342, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 342, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 342, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 547, + 386 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. Posetrack: A benchmark for human pose estimation and tracking, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 388, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 547, + 409 + ], + "type": "text", + "content": "[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 411, + 547, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 547, + 454 + ], + "type": "text", + "content": "[3] Bruno Artacho and Andreas Savakis. Unipose: Unified human pose estimation in single images and videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 456, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 545, + 477 + ], + "type": "text", + "content": "[4] James Atwood and Don Towsley. Diffusion-convolutional neural networks, 2015." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 479, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 545, + 511 + ], + "type": "text", + "content": "[5] Niloofar Azizi, Horst Possegger, Emanuele Rodola, and Horst Bischof. 3d human pose estimation using mobius graph convolutional networks, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 513, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 547, + 567 + ], + "type": "text", + "content": "[6] Abdallah Benzine, Florian Chabot, Bertrand Luvison, Quoc Cuong Pham, and Catherine Achard. Pandanet: Anchor-based single-shot multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 569, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 547, + 612 + ], + "type": "text", + "content": "[7] Gedas Bertasius, Christoph Feichtenhofer, Du Tran, Jianbo Shi, and Lorenzo Torresani. Learning temporal pose estimation from sparsely labeled videos. In Advances in Neural Information Processing Systems 33, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 613, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 545, + 647 + ], + "type": "text", + "content": "[8] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs, 2013." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 547, + 712 + ], + "type": "text", + "content": "[9] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14708" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[10] He Chen, Pengfei Guo, Pengfei Li, Gim Hee Lee, and Gregory Chirikjian. Multi-person 3d pose estimation in crowded scenes based on multi-view geometry. arXiv preprint arXiv:2007.10986, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 170 + ], + "type": "text", + "content": "[11] Long Chen, Haizhou Ai, Rui Chen, Zijie Zhuang, and Shuang Liu. Cross-view tracking for multi-human 3d pose estimation at over 100 fps. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "type": "text", + "content": "[12] Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 206, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 288, + 239 + ], + "type": "text", + "content": "[13] Tianlang Chen, Chen Fang, Xiaohui Shen, Yiheng Zhu, Zhili Chen, and Jiebo Luo. Anatomy-aware 3d human pose estimation in videos. arXiv preprint arXiv:2002.10322, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "text", + "content": "[14] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 370 + ], + "type": "text", + "content": "[15] Zerui Chen, Yan Huang, Hongyuan Yu, Bin Xue, Ke Han, Yiru Guo, and Liang Wang. Towards part-aware monocular 3d human pose estimation: An architecture search approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 715-732, Cham, 2020. Springer International Publishing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 372, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 288, + 426 + ], + "type": "text", + "content": "[16] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S. Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "type": "text", + "content": "[17] Yu Cheng, Bo Wang, Bo Yang, and Robby T. Tan. Monocular 3d multi-person pose estimation by integrating top-down and bottom-up networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7649-7659, June 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 482, + 288, + 525 + ], + "type": "text", + "content": "[18] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 526, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 288, + 579 + ], + "type": "text", + "content": "[19] Chia-Jung Chou, Jui-Ting Chien, and Hwann-Tzong Chen. Self adversarial training for human pose estimation. In 2018 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 17-30, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 581, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 288, + 635 + ], + "type": "text", + "content": "[20] Xiao Chu, Wei Yang, Wanli Ouyang, Cheng Ma, Alan L. Yuille, and Xiaogang Wang. Multi-context attention for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 637, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 637, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 637, + 288, + 679 + ], + "type": "text", + "content": "[21] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "text", + "content": "[22] Edward De Brouwer, Jaak Simm, Adam Arany, and Yves Moreau. Gru-ode-bayes: Continuous modeling of sporadically-observed time series, 2019." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 106 + ], + "type": "text", + "content": "[23] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 108, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 547, + 161 + ], + "type": "text", + "content": "[24] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Stefano Alletto, and Rita Cucchiara. Compressed volumetric heatmaps for multi-person 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 163, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 547, + 195 + ], + "type": "text", + "content": "[25] Qi Fang, Qing Shuai, Junting Dong, Hujun Bao, and Xiaowei Zhou. Reconstructing 3d human pose by watching humans in the mirror. In CVPR, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 198, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 547, + 251 + ], + "type": "text", + "content": "[26] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: Co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In CVPR, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 254, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 254, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 547, + 319 + ], + "type": "text", + "content": "[27] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, June 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 321, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 321, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 321, + 547, + 397 + ], + "type": "text", + "content": "[28] Congzhentao Huang, Shuai Jiang, Yang Li, Ziyue Zhang, Jason Traish, Chen Deng, Sam Ferguson, and Richard Yi Da Xu. End-to-end dynamic matching network for multi-view multi-person 3d pose estimation. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII, page 477-493, Berlin, Heidelberg, 2020. Springer-Verlag." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 399, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 547, + 453 + ], + "type": "text", + "content": "[29] Zijie Huang, Yizhou Sun, and Wei Wang. Coupled graph ode for learning interacting system dynamics. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery amp; Data Mining, KDD '21, page 705-715, New York, NY, USA, 2021. Association for Computing Machinery." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 455, + 547, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 455, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 312, + 455, + 547, + 509 + ], + "type": "text", + "content": "[30] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 511, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 511, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 511, + 547, + 555 + ], + "type": "text", + "content": "[31] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris M. Kitani. Optical non-line-of-sight physics-based 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 556, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 547, + 578 + ], + "type": "text", + "content": "[32] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks, 2016." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 579, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 579, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 579, + 547, + 623 + ], + "type": "text", + "content": "[33] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Multiposenet: Fast multi-person pose estimation using pose residual network. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 624, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 547, + 668 + ], + "type": "text", + "content": "[34] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 547, + 713 + ], + "type": "text", + "content": "[35] Jogendra Nath Kundu, Ambareesh Revanur, Govind Vitthal Waghmare, Rahul Mysore Venkatesh, and R. Venkatesh Babu. Unsupervised cross-modal alignment for multi-person 3d pose estimation, 2020." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14709" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 288, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 288, + 138 + ], + "type": "text", + "content": "[36] Jogendra Nath Kundu, Siddharth Seth, Varun Jampani, Mugalodi Rakesh, R. Venkatesh Babu, and Anirban Chakraborty. Self-supervised 3d human pose estimation via part guided novel image synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 140, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 184 + ], + "type": "text", + "content": "[37] Han Li, Bowen Shi, Wenrui Dai, Yabo Chen, Botao Wang, Yu Sun, Min Guo, Chenlin Li, Junni Zou, and Hongkai Xiong. Hierarchical graph networks for 3d human pose estimation, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "type": "text", + "content": "[38] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 241, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 297 + ], + "type": "text", + "content": "[39] Shichao Li, Lei Ke, Kevin Pratama, Yu-Wing Tai, Chi-Keung Tang, and Kwang-Ting Cheng. Cascaded deep monocular 3d human pose estimation with evolutionary training data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 288, + 342 + ], + "type": "text", + "content": "[40] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 343, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 343, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 343, + 288, + 387 + ], + "type": "text", + "content": "[41] Junfa Liu, Juan Rojas, Zhijun Liang, Yihui Li, and Yisheng Guan. A graph attention spatio-temporal convolutional networks for 3d human pose estimation in video. arXiv preprint arXiv:2003.14179, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 388, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 454 + ], + "type": "text", + "content": "[42] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X, page 318–334, Berlin, Heidelberg, 2020. Springer-Verlag." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 456, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 288, + 532 + ], + "type": "text", + "content": "[43] Shengyuan Liu, Pei Lv, Yuzhen Zhang, Jie Fu, Junjin Cheng, Wanqing Li, Bing Zhou, and Mingliang Xu. Semi-dynamic hypergraph neural network for 3d pose estimation. In Christian Bessiere, editor, Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI20, pages 782-788. International Joint Conferences on Artificial Intelligence Organization, 7 2020. Main track." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 533, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 288, + 566 + ], + "type": "text", + "content": "[44] Xuanqing Liu, Tesi Xiao, Si Si, Qin Cao, Sanjiv Kumar, and Cho-Jui Hsieh. Neural sde: Stabilizing neural ode networks with stochastic noise, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 288, + 601 + ], + "type": "text", + "content": "[45] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, and Zhaoyue Chen. Learning continuous-time dynamics by stochastic differential networks, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 602, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 288, + 678 + ], + "type": "text", + "content": "[46] Yingru Liu, Yucheng Xing, Xuewen Yang, Xin Wang, Jing Shi, Di Jin, Zhaoyue Chen, and Jacqueline Wu. Continuous-time stochastic differential networks for irregular time series modeling. In Teddy Mantoro, Minho Lee, Media Anugerah Ayu, Kok Wai Wong, and Achmad Nizar Hidayanto, editors, Neural Information Processing, pages 343-351, Cham, 2021. Springer International Publishing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 288, + 713 + ], + "type": "text", + "content": "[47] Zhenguang Liu, Haoming Chen, Runyang Feng, Shuang Wu, Shouling Ji, Bailin Yang, and Xun Wang. Deep dual consecutive network for human pose estimation. In Proceed-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 525-534, June 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 98, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 98, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 312, + 98, + 547, + 152 + ], + "type": "text", + "content": "[48] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethinking the heatmap regression for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13264-13273, June 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 153, + 547, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 153, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 312, + 153, + 547, + 198 + ], + "type": "text", + "content": "[49] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 199, + 547, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 199, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 312, + 199, + 547, + 253 + ], + "type": "text", + "content": "[50] Rahul Mitra, Nitesh B. Gundavarapu, Abhishek Sharma, and Arjun Jain. Multiview-consistent semi-supervised learning for 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 256, + 547, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 256, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 547, + 277 + ], + "type": "text", + "content": "[51] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 280, + 547, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 280, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 312, + 280, + 547, + 312 + ], + "type": "text", + "content": "[52] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 315, + 547, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 315, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 315, + 547, + 358 + ], + "type": "text", + "content": "[53] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 360, + 547, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 360, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 312, + 360, + 547, + 404 + ], + "type": "text", + "content": "[54] Wanli Ouyang, Xiao Chu, and Xiaogang Wang. Multisource deep learning for human pose estimation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2337-2344, 2014." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 406, + 547, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 406, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 312, + 406, + 547, + 471 + ], + "type": "text", + "content": "[55] George Papandreou, Tyler Zhu, Liang-Chieh Chen, Spyros Gidaris, Jonathan Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-286, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 474, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 545, + 506 + ], + "type": "text", + "content": "[56] Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 509, + 547, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 547, + 552 + ], + "type": "text", + "content": "[57] Michael Poli, Stefano Massaroli, Clayton M. Rabideau, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Continuous-depth neural models for dynamic graph prediction, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 554, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 554, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 312, + 554, + 547, + 609 + ], + "type": "text", + "content": "[58] Zhongwei Qiu, Kai Qiu, Jianlong Fu, and Dongmei Fu. Dgen: Dynamic graph convolutional network for efficient multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 34(07):11924-11931, Apr. 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 312, + 611, + 545, + 654 + ], + "type": "text", + "content": "[59] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 656, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 656, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 656, + 547, + 689 + ], + "type": "text", + "content": "[60] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 547, + 713 + ], + "type": "text", + "content": "[61] Yulia Rubanova, Ricky T. Q. Chen, and David Duvenaud. Latent odes for irregularly-sampled time series, 2019." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14710" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[62] Yulia Rubanova, Tian Qi Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series. In Advances in Neural Information Processing Systems 32, pages 5321-5331. 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 182 + ], + "type": "text", + "content": "[63] Luca Schmidtke, Athanasios Vlontzos, Simon Ellershaw, Anna Lukens, Tomoki Arichi, and Bernhard Kainz. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2484-2494, June 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 184, + 288, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 238 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 238 + ], + "type": "text", + "content": "[64] Theodoros Sofianos, Alessio Sampieri, Luca Franco, and Fabio Galasso. Space-time-separable graph convolutional network for pose forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11209-11218, October 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "text", + "content": "[65] Kai Su, Dongdong Yu, Zhenqi Xu, Xin Geng, and Changhu Wang. Multi-person pose estimation with enhanced channelwise and spatial information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 294, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 338 + ], + "type": "text", + "content": "[66] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5686-5696, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "text", + "content": "[67] Wei Tang and Ying Wu. Does learning specific features for related parts help human pose estimation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 383, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 383, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 383, + 288, + 414 + ], + "type": "text", + "content": "[68] Zhi Tian, Hao Chen, and Chunhua Shen. Direct pose estimation. ArXiv, abs/1911.07451, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 415, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 288, + 459 + ], + "type": "text", + "content": "[69] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 1653-1660, 2014." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 460, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 288, + 502 + ], + "type": "text", + "content": "[70] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 504, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 288, + 547 + ], + "type": "text", + "content": "[71] Ali Varamesh and Tinne Tuytelaars. Mixture dense regression for object detection and human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 548, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 602 + ], + "type": "text", + "content": "[72] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13294-13304, June 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 604, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 604, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 604, + 288, + 667 + ], + "type": "text", + "content": "[73] Jiahang Wang, Sheng Jin, Wentao Liu, Weizhong Liu, Chen Qian, and Ping Luo. When human pose estimation meets robustness: Adversarial algorithms and benchmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11855-11864, June 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 712 + ], + "type": "text", + "content": "[74] Jian Wang, Xiang Long, Yuan Gao, Errui Ding, and Shilei Wen. Graph-pcnn: Two stage human pose estimation with graph pose refinement. In The European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 95 + ], + "type": "text", + "content": "[75] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 95, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 547, + 139 + ], + "type": "text", + "content": "[76] Manchen Wang, Joseph Tighe, and Davide Modolo. Combining detection and tracking for human pose estimation in videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 140, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 547, + 182 + ], + "type": "text", + "content": "[77] Rui Wang, Zhongzheng Cao, Xiangyang Wang, Zhi Liu, and Xiaoqiang Zhu. Human pose estimation with deeply learned multi-scale compositional models. IEEE Access, 7:71158-71166, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 183, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 183, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 183, + 547, + 205 + ], + "type": "text", + "content": "[78] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 206, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 547, + 239 + ], + "type": "text", + "content": "[79] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 239, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 547, + 282 + ], + "type": "text", + "content": "[80] Rongchang Xie, Chunyu Wang, and Yizhou Wang. Meta-fuse: A pre-trained fusion model for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 284, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 284, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 547, + 315 + ], + "type": "text", + "content": "[81] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 316, + 547, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 547, + 371 + ], + "type": "text", + "content": "[82] Jingwei Xu, Zhenbo Yu, Bingbing Ni, Jiancheng Yang, Xiaokang Yang, and Wenjun Zhang. Deep kinematics analysis for monocular 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 372, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 547, + 415 + ], + "type": "text", + "content": "[83] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16105-16114, June 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 416, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 547, + 459 + ], + "type": "text", + "content": "[84] Xixia Xu, Qi Zou, and Xue Lin. Adaptive hypergraph neural network for multi-person pose estimation. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2955-2963, Jun. 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 460, + 547, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 547, + 492 + ], + "type": "text", + "content": "[85] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 493, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 493, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 547, + 536 + ], + "type": "text", + "content": "[86] Wei Yang, Shuang Li, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. Learning feature pyramids for human pose estimation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 537, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 547, + 592 + ], + "type": "text", + "content": "[87] Wei Yang, Wanli Ouyang, Hongsheng Li, and Xiaogang Wang. End-to-end learning of deformable mixture of parts and deep convolutional neural networks for human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 593, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 547, + 657 + ], + "type": "text", + "content": "[88] Yiding Yang, Zhou Ren, Haoxiang Li, Chunluan Zhou, Xin chao Wang, and Gang Hua. Learning dynamics via graph neural networks for human pose estimation and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8074-8084, June 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "type": "text", + "content": "[89] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), 2018." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14711" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 631 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[90] Bing Yu, Haoteng Yin, and Zhanxing Zhu. ST-UNet: A spatio-temporal u-network for graph-structured time series modeling. arXiv, abs/1903.05631, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 162 + ], + "type": "text", + "content": "[91] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9064-9073, June 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 217 + ], + "type": "text", + "content": "[92] Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. Simpoe: Simulated character control for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7159–7169, June 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 219, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 288, + 295 + ], + "type": "text", + "content": "[93] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Srnet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 507–523, Cham, 2020. Springer International Publishing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 297, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 340 + ], + "type": "text", + "content": "[94] Feng Zhang, Xiatian Zhu, Hanbin Dai, Mao Ye, and Ce Zhu. Distribution-aware coordinate representation for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 342, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 384 + ], + "type": "text", + "content": "[95] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 386, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 440 + ], + "type": "text", + "content": "[96] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N. Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "type": "text", + "content": "[97] Jianan Zhen, Qi Fang, Jiaming Sun, Wentao Liu, Wei Jiang, Hujun Bao, and Xiaowei Zhou. Smap: Single-shot multiperson absolute 3d pose estimation, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 476, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 529 + ], + "type": "text", + "content": "[98] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11656-11665, October 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 532, + 288, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 532, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 53, + 532, + 288, + 596 + ], + "type": "text", + "content": "[99] Tianfei Zhou, Wenguan Wang, Si Liu, Yi Yang, and Luc Van Gool. Differentiable multi-granularity human representation learning for instance-aware human semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1622-1631, June 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 598, + 287, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 598, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 598, + 287, + 631 + ], + "type": "text", + "content": "[100] Bin Xiao Zhaoxiang Zhang Jingdong Wang Zigang Geng, Ke Sun. Bottom-up human pose estimation via disentangled keypoint regression. In CVPR, 2021." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14712" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_content_list.json b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..23eac7a139da45abde0d863fdc88aa72c30c164b --- /dev/null +++ b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_content_list.json @@ -0,0 +1,1425 @@ +[ + { + "type": "text", + "text": "HM-ViT: Hetero-modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer", + "text_level": 1, + "bbox": [ + 91, + 128, + 879, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Xiang $^{1}$ , Runsheng Xu $^{1}$ , Jiaqi Ma $^{1}$ $^{1}$ University of California, Los Angeles \n{haxiang, rxx3386, jiaqima}@g.ucla.edu", + "bbox": [ + 315, + 202, + 653, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vehicle-to-Vehicle technologies have enabled autonomous vehicles to share information to see through occlusions, greatly enhancing perception performance. Nevertheless, existing works all focused on homogeneous traffic where vehicles are equipped with the same type of sensors, which significantly hampers the scale of collaboration and benefit of cross-modality interactions. In this paper, we investigate the multi-agent hetero-modal cooperative perception problem where agents may have distinct sensor modalities. We present HM-ViT, the first unified multi-agent hetero-modal cooperative perception framework that can collaboratively predict 3D objects for highly dynamic Vehicle-to-Vehicle (V2V) collaborations with varying numbers and types of agents. To effectively fuse features from multi-view images and LiDAR point clouds, we design a novel heterogeneous 3D graph transformer to jointly reason inter-agent and intra-agent interactions. The extensive experiments on the V2V perception dataset OPV2V demonstrate that the HM-ViT outperforms SOTA cooperative perception methods for V2V hetero-modal cooperative perception. Our code will be released at https://github.com/XHwind/HM-ViT.", + "bbox": [ + 75, + 325, + 473, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in Vehicle-to-Vehicle (V2V) communication technology and intelligent transportation systems [13, 31, 27, 14, 7, 19, 28, 51, 53, 52] have allowed autonomous vehicles (AVs) to share sensory information, enabling them to perceive their surroundings better [3, 32, 54]. With the rapid growth of autonomous driving, V2V perception systems have the potential to be deployed at scale and create a safer transportation system. Cooperative perception systems, as shown in recent studies [49, 48, 40], can intelligently aggregate features from multiple vehicles within the communication range to enhance visual reasoning and overall performance.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c46fc370206b843114e773d0cb899b250530a805e8e177e26b489f1a0ecca645.jpg", + "image_caption": [ + "Figure 1: Illustration of multi-agent hetero-modal V2V systems where each agent may be equipped with either LiDAR or multi-view cameras." + ], + "image_footnote": [], + "bbox": [ + 503, + 290, + 893, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the rapid growth in this field, previous studies [49, 40, 46, 22, 37, 47] have primarily focused on homogeneous multi-agent cooperative perception, where all agents are equipped with the same type of sensors. In reality, however, agents may have different sensor modalities (hetero-modality) due to the cost and sensor preferences of ADS developers and car makers. As shown in Fig. 1, some agents are equipped with only LiDARs (LiDAR agents), while others only have multiple cameras (camera agents). Enabling collaboration between these heterogeneous agents could improve the sensing capability by allowing agents to see through occlusions and increase the scale and reliability of V2V systems. Additionally, LiDAR agents can provide accurate geometric information, while camera agents can provide rich semantic context. Thus, the collaboration between these agents could leverage the distinct but complementary environment attributes captured by each sensor modality to enhance the V2V perception systems. Furthermore, compared with the single-agent solution where multiple LiDARs and cameras are installed in a single vehicle, distributing different types of sensors across distinct agents could also potentially decrease the costs for each agent while still achieving satisfying performance. Nevertheless, whether, when, and how multi-agent hetero-modal V2V cooperation can benefit the perception system of heterogeneous traffic has not yet been studied.", + "bbox": [ + 496, + 508, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "284", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/decd16733e4305627bce77a76256c8c75d4ea2ca15a907eb5fc94a984626fd3a.jpg", + "image_caption": [ + "Figure 2: Comparison of the single-agent multi-modal system and multi-agent hetero-modal system. The graph structure of the former is fixed whereas for the latter the graph is both dynamic and heterogeneous." + ], + "image_footnote": [], + "bbox": [ + 109, + 95, + 436, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address the multi-agent hetero-modal cooperative perception problem where each agent could have distinct sensor types and share/receive information with each other. Notably, as shown in Fig. 2, this multi-agent hetero-modal setting is distinct from the single-agent multi-modal setting. In the hetero-modal setting, the agent sensors form a dynamic heterogeneous graph where the existence and types of sensors are random, and the relative poses vary from scene to scene. In contrast, the sensor types/numbers and relative positions (extrinsics) between sensors are fixed in the single-agent multi-modal setting. Existing multi-modal methods heavily rely on these assumptions, and most of existing works [38, 39, 2, 55, 50] transform LiDAR points or 3D proposals onto the image plane to index 2D features. Their network architectures build upon the co-existence of both LiDAR and camera inputs with fixed geometric relationships. However, the dynamic nature of hetero-modal V2V perception requires a flexible architecture that can handle varying agent numbers and types, and the transmitted neural features are also spatially misaligned. Moreover, there are semantic discrepancies in the transmitted features between camera agents and LiDAR agents. Hence, these unique characteristics pose significant challenges for designing the multi-agent hetero-modal cooperative system and prevent adapting existing multi-modal fusion methods to this new problem.", + "bbox": [ + 75, + 337, + 468, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enable collaboration between heterogeneous agents in V2V systems, we propose Hetero-Modal Vision Transformer (HM-ViT), the first unified cooperative perception framework that can leverage and fuse distributed information for hetero-modal V2V perception via a spatial-aware 3D heterogeneous vision transformer. Fig. 3 demonstrates the overall framework. Each agent first generates bird's eye view (BEV) representations through modality-specific encoders and then shares compressed features with neighboring agents. Afterward, the received features are decompressed and aggregated via the proposed HM-ViT,", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which conducts joint local and global heterogeneous 3D attentions with the consideration of both node and edge types. Our extensive experiments show that the HM-ViT can significantly improve the perception capability of camera agents and LiDAR agents over the single-agent baseline and outperforms SOTA cooperative perception methods by a large margin. In particular, for camera agents, performance can be boosted from $2.1\\%$ to $53.2\\%$ at $\\mathrm{AP}@0.7$ with the collaboration of LiDAR agents, a 23-fold improvement. Our primary contributions can be summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the novel transformer framework (HMViT) for multi-agent hetero-modal cooperative perception, capable of capturing the modality-specific characteristics and heterogeneous 3D interactions. The proposed model exhibits superior flexibility and robustness with state-of-the-art performance on highly dynamic heterogeneous traffic involving varying agent numbers and types.", + "- We propose a generic heterogeneous 3D graph attention (H $^3$ GAT), tailored for extracting inter-agent and intra-agent heterogeneous interactions. We instantiate two such attentions - local attention (H $^3$ GAT-L) and global attention (H $^3$ GAT-G) for capturing both local and global visual cues.", + "- We conduct extensive benchmark experiments by varying sensor modalities, demonstrating the strong performance of the proposed method for hetero-modal V2V perception tasks. We will release all the codes and baselines to facilitate future research." + ], + "bbox": [ + 517, + 250, + 890, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 573, + 635, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "V2V perception. V2V perception aims to enhance the perception performance of autonomous vehicles by leveraging shared information from other connected vehicles. Existing works have primarily focused on LiDAR-based 3D object perception. The pioneer cooperative perception methods transmit raw sensing observation (i.e., early fusion) or perception outputs (i.e., late fusion), whereas recent works [49, 40, 16, 43, 21, 48] are exploring the use of circulating intermediate neural features for achieving better performance-bandwidth trade-off. V2VNet [40] employs graph neural networks to aggregate the shared neural features for joint detection and prediction. AttFuse [49] uses single-head attention to model the per-location multi-agent interaction. Disconet [22] presents a matrix-valued edge weight for learning the interactions and a teacher-student learning framework to facilitate the training. V2X-ViT explores [48] vision transformer for vehicle-to-everything cooperation via window attention and heterogeneous self-attention. CoBEVT [46] presents a generic transformer framework for camera-based BEV semantic segmentation.", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "285", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a3d6260f86db65efd341a4abfd55938058b3a9ac4c29bc5490c7814aa8f5e172.jpg", + "image_caption": [ + "Figure 3: Overview of proposed hetero-modal V2V cooperative perception system. Each agent first produces BEV features through modality-specific feature extractors (Sec. 3.1). The BEV features are then compressed and shared (Sec. 3.3) with neighboring connected agents and the received features are decompressed in the ego agent side and fed into hetero-modal vision transformer to conduct graph-structured feature fusion (Sec. 3.2 and Sec. 3.3). The refined features are finally passed into the hetero-modal detection head to predict 3D bounding boxes (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 888, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, none of the existing works explored multi-agent multi-camera 3D object detection, let alone multi-agent hetero-modal perception. In contrast to existing methods, HM-ViT is the first to employ sparse heterogeneous local and global attentions to capture the 3D inter-agent and intraagent interactions in a computationally efficient manner.", + "bbox": [ + 75, + 383, + 470, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera-based 3D object detection. Early works [4, 8, 33] mainly focus on monocular 3D detection but a single camera can only provide a 2D view of the scene, and inferring 3D from 2D is intrinsically hard. Recent development of self-driving datasets [5, 34, 6] featured with full sensor suits has enabled the research direction of 3D object detection from multiple cameras. DETR3D [41] proposes a 3D-2D query paradigm for extracting 3D features from 2D multi-view images. Graph-DETR3D [10] further utilizes graph structure learning to enhance the representation at the border regions. Alternatively, another stream of works [17, 29, 44, 23] focuses on aggregating BEV features from multi-view cameras for conducting downstream perception tasks. LSS [29] lifts the 2D features to 3D frustum via latent depth and then splats frustums into a BEV grid. $\\mathbf{M}^2\\mathbf{BEV}$ [44] further extends LSS with less memory consumption and conducts detection and segmentation. BEV-Former [23] constructs BEV queries and explores spatial cross-attention and temporal self-attention to recurrently refine the BEV features, achieving SOTA performance on both NuScenes [5] and Waymo Open Datasets [34].", + "bbox": [ + 75, + 476, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal fusion. Existing multi-sensor fusion methods can be divided into point-level fusion, proposal-level fusion, and BEV-level fusion. Point-level fusion decorates the input from one modality with attributes from the other modality. PointPainting [38] and PaintAugmenting [39] decorate the LiDAR point clouds with semantic segmentation scores and 2D CNN image features respectively while [11, 1] project", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR onto the image plane to augment the RGB values with depth information and conduct detection. On the other hand, proposal-level fusion generates proposals from one modality and then indexes features from the other modality for further refinement. MV3D [9] produces object queries from LiDAR BEV and then extracts the features from camera data and LiDAR front view. [30, 42] lift 2D bounding boxes to frustum and conduct 3D object detection from the frustum of point clouds. Conversely, BEV-level fusion converts features from different modalities to unified BEV representations, preserving both geometric and semantic information. BEVFusions [26, 24] concatenate BEV features from camera and LiDAR and fuse it via a fusion module. Hence, existing single-agent multi-modal fusion methods rely on the co-existence of both camera and LiDAR with fixed geometric relationships, which is unsuitable for our multi-agent hetero-modal cooperative perception problem with a dynamic heterogeneous collaboration graph.", + "bbox": [ + 496, + 383, + 893, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 710, + 635, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we explore the multi-agent hetero-modal cooperative perception, where each AV is equipped with either a LiDAR or multiple cameras. Our goal is to create a robust and flexible cooperative perception system that allows for efficient collaboration between any number of agents with varying sensor types, ultimately improving the perception capabilities of the vehicle in a unified end-to-end fashion. The pipeline, illustrated in 3, includes modality-specific feature extraction, compression and sharing, HMViT for feature fusion, and hetero-modal detection head.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "286", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/86acf29b7f78e2c9b05186a39dfc9f09228fe4b7a2609a7881752d70c23c02b2.jpg", + "image_caption": [ + "Figure 4: HM-ViT architecture. (a) Graph-structured fusion process. (b) Heterogeneous 3D graph attention $\\mathrm{(H^3GAT)}$ . (c) Illustration of sampled tokens for $\\mathrm{H}^{3}\\mathrm{GAT - L}$ and $\\mathrm{H}^{3}\\mathrm{GAT - G}$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 99, + 290, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4fc9fb0d581f82ada74494b79ed32e4215416de18b85987bb3e6edca3aa09cf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 107, + 604, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57c749347d29ecd3cfbe379064a23adf13e63fd2f5bb76aae6f8ba961938ae38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 113, + 885, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Modality-specific feature extraction", + "text_level": 1, + "bbox": [ + 76, + 335, + 387, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiDAR stem: We leverage PointPillar [20] to process point clouds for each LiDAR agent. The raw point cloud is converted to a 2D pseudo-image, flattened along the height dimension, and fed into 2D convolutional neural networks to produce the salient feature map $\\mathbf{F}_j \\in \\mathcal{R}^{H \\times W \\times C}$ , which is compressed and shared with all the neighboring agents.", + "bbox": [ + 75, + 359, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera stem: Each camera agent is equipped with $m$ monocular cameras. The sensing observation of $i$ -th agent includes the input images $I_k^i \\in \\mathcal{R}^{h \\times w \\times 3}$ and the known projection matrix $P_k^i \\in \\mathcal{R}^{3 \\times 4}$ that maps 3D reference points to different image views. Our goal is to generate a BEV feature representation $\\mathbf{F}_i \\in \\mathcal{R}^{H \\times W \\times C}$ that is amenable for feature fusion with other collaborators. In this work, we adopt similar architecture to BEVFormer [23] with no temporal information for feature extraction. For a faster running time, we adopt ResNet50 to extract 2D image features and then adopt a learnable 2D BEV query to inquire spatial information from the encoded multi-view features via spatial cross attention and projection matrices. The resulting refined BEV feature $\\mathbf{F}_i$ is centered around agent $i$ and shared with connected AVs.", + "bbox": [ + 75, + 454, + 472, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Heterogeneous 3D Graph Attention (H $^3$ GAT)", + "text_level": 1, + "bbox": [ + 76, + 694, + 460, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To account for the distinct characteristics of BEV features extracted from different sensor modalities, the learning process of each modality must be distinguished, and the cross-modality interactions between multiple agents should vary. To capture this heterogeneity, we present a novel heterogeneous 3D graph attention (H $^3$ GAT), in which nodes and edges are type-dependent to reason spatial interactions and cross-agent relations jointly. We encode both local and global interactions to better capture the 3D ambiguity in BEV feature space. Local attention can help preserve object details, while global attention can provide a better understanding of environmental contexts such as road topology", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and traffic density.", + "bbox": [ + 498, + 337, + 624, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 4b, We build a 3D heterogeneous collaboration graph. Each node $v(i,x) = \\mathbf{F}_x^i \\in \\mathcal{R}^C$ is a feature vector of agent $i$ 's feature map at spatial location $x \\in \\mathcal{R}^2$ . 3D heterogeneous graph attention is performed for spatially connected nodes in the BEV feature space. Depending on the definition of spatial connectivity, we will derive local attention and global attention. Here for notation simplicity, we only derive single-head equations but in real implementations, multi-head variants are used. Formally, we first project feature vectors onto different feature spaces to form query, key, and value vectors:", + "bbox": [ + 496, + 352, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {x} ^ {j} = \\operatorname {D e n s e} _ {\\tau_ {j}} \\mathbf {F} _ {x} ^ {j} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 526, + 890, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {x} ^ {j} = \\operatorname {D e n s e} _ {\\tau_ {j}} \\mathbf {F} _ {x} ^ {j} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 547, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {x} ^ {j} = \\operatorname {D e n s e} _ {e _ {i j}} \\mathbf {F} _ {x} ^ {j} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 566, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\mathrm{Dense}\\square$ is a set of linear layers indexed by subscript $\\square$ . For the query and key vectors, we use linear projectors $\\mathrm{Dense}_{\\tau_j}$ indexed by node type $\\tau_{j}$ to extract modality-specific features. For the value vector, we index the projector via edge type $\\mathrm{Dense}_{e_{ij}}$ to reflect the heterogeneity of cross-modality multi-agent interactions. The set of connected nodes of $v(i,x)$ is denoted as $\\mathcal{N}(i,x)$ . Then the attention is operated as follows:", + "bbox": [ + 496, + 593, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} (j, y) = \\operatorname {S o f t m a x} _ {(j, y) \\in \\mathcal {N} (i, x)} \\left(\\mathbf {Q} _ {x} ^ {i} \\mathbf {W} _ {e _ {i j}} \\mathbf {K} _ {y} ^ {j}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 720, + 890, + 748 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {x} ^ {i} = \\sum_ {(j, y) \\in \\mathcal {N} (i, x)} \\mathbf {a} (j, y) \\mathbf {V} _ {y} ^ {j} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 750, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{e_{ij}}\\in \\mathcal{R}^{C\\times C}$ is used to adjust the dot product of $\\mathbf{Q}_x^i$ and $\\mathbf{K}_y^j$ to further encode the heterogeneity of edges.", + "bbox": [ + 498, + 794, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depending on how the nodes are sampled (Fig. 4c), we design two types of attentions: local attention $(\\mathbf{H}^3\\mathbf{GAT - L})$ which performs local window-based attention and global attention $(\\mathbf{H}^3\\mathbf{GAT - G})$ which performs sparse global grid-based attention. Fig. 4c visualizes how the local and global", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "287", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7ada8e21e4acbb6353e610e741956fd6ab9acb59ff6cc96901e0633db245bad6.jpg", + "image_caption": [ + "(a) $\\mathsf{H}^3\\mathsf{GAT - L}$ Block", + "Figure 5: Transformer blocks for local and global attentions." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 250, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f7340ba696d721cbe52da55b71876f0b0da4da42e26ba2b48e178ea989076393.jpg", + "image_caption": [ + "(b) $\\mathsf{H}^3\\mathsf{GAT - G}$ Block" + ], + "image_footnote": [], + "bbox": [ + 300, + 88, + 436, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tokens are sampled, where the local tokens are sampled within local windows and the global tokens are sparsely sampled grids scattered across feature maps. The local interactions can help preserve spatial cues and provide reliable estimates while the global reasoning can help understand global semantic context.", + "bbox": [ + 75, + 402, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both $\\mathrm{H}^3\\mathrm{GAT - L}$ and $\\mathrm{H}^3\\mathrm{GAT - G}$ can be implemented efficiently by decomposing the spatial axes. More specifically, we stack all the agents' features to $\\mathbf{F}\\in \\mathcal{R}^{N\\times H\\times W\\times C}$ where $N$ is the number of agents. For $\\mathrm{H}^3\\mathrm{GAT - L}$ , we decompose the feature map into 3D non-overlapping windows along the first axis [35, 36], each of size $N\\times P\\times P$ . The partitioned tensor has the shape $(\\frac{H}{P}\\times \\frac{W}{P},N\\times P^2,C)$ where the heterogeneous 3D local graph attention is conducted for $NP^2$ tokens within the same window. Similarly, for $\\mathrm{H}^3\\mathrm{GAT - G}$ , we swap the axis and partition the tensor into the shape $(N\\times P^2,\\frac{H}{P}\\times \\frac{W}{P},C)$ . Due to the swap operation, the sampled grids will be sparsely scattered and the attention is operated for these sparsely sampled $\\frac{H}{P}\\times \\frac{W}{P}$ grids, which can capture sparse global information.", + "bbox": [ + 75, + 492, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To integrate this local and global attention into transformer architecture, we further present a heterogeneous normalization layer (HM-LN) and heterogeneous MLP (HMMLP) which use type-dependent parameters. As shown in Fig. 5, we first pass all the features into the HM-LN where different statistics are calculated and used as per each agent's modality type. Afterward, we feed the normalized features into a heterogeneous 3D graph attention $(\\mathrm{H}^3\\mathrm{GATL / H}^3\\mathrm{GAT - G})$ to jointly reason heterogeneous inter-agent and intra-agent interactions. Then, we pass the fused features to another HM-LN followed by a Hetero-modal MLP layer where different sets of parameters are used for camera and LiDAR features. By carefully designing these com", + "bbox": [ + 75, + 704, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Multi-agent hetero-modal fusion process" + ], + "code_body": "1: Input: decompressed feature $\\mathbf{F}_i$ , pose $x_i$ for each agent \n2: $\\mathbf{F}_i^{(0)} = \\mathbf{F}_i$ \n3: for $l = 1, \\ldots, L$ do \n4: for each agent $i$ do $\\triangleright$ Process in parallel \n5: $\\mathbf{F}_{j \\rightarrow i}^{(l-1)} = \\Gamma_{j \\rightarrow i}\\left(\\mathbf{F}_j^{(l-1)}\\right) \\triangleright$ Spatially transform neighboring agents' features \n6: $\\mathbf{F}_i^{(l)} = \\mathrm{H}^3\\mathrm{GAT-L}$ Block( $\\{\\mathbf{F}_{j \\rightarrow i}^{(l-1)}\\}$ ) $\\triangleright$ Update node via local attention \n7: end for \n8: for each agent $i$ do $\\triangleright$ Process in parallel \n9: $\\mathbf{F}_{j \\rightarrow i}^{(l)} = \\Gamma_{j \\rightarrow i}\\left(\\mathbf{F}_j^{(l)}\\right) \\triangleright$ Spatially transform neighboring agents' features \n10: $\\mathbf{F}_i^{(l)} = \\mathrm{H}^3\\mathrm{GAT-G}$ Block( $\\{\\mathbf{F}_{j \\rightarrow i}^{(l)}\\}$ ) $\\triangleright$ Update node via global attention \n11: end for \n12: end for \n13: $\\mathbf{F}_i = \\mathrm{HM - MLP}\\left(\\mathbf{F}_i^{(L)}\\right)$ $\\triangleright$ Output updated features", + "bbox": [ + 506, + 117, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ponents, we can maintain modality-specific characteristics throughout the fusion process while benefiting from cross-modality multi-agent interactions.", + "bbox": [ + 498, + 415, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Hetero-modal Vision Transformer", + "text_level": 1, + "bbox": [ + 500, + 469, + 797, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compression and sharing: To reduce the transmission bandwidth, a series of $1 \\times 1$ convolutions is applied to reduce the transmitted feature size along the channel dimension. Together with the intermediate features, each agent's pose $x_{i}$ is also circulated within the collaboration graph. The ego agent will receive these features and decompress them back to the original size via another convolutional network. For processing intermediate features of the camera agent and LiDAR agent, we leverage distinct parameters in the compression and decompression modules to preserve the modality-specific characteristics.", + "bbox": [ + 496, + 492, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Graph-structured feature fusion: The received BEV features are centered around different spatial locations as each agent perceives the dynamic environment from different view points. To this end, we present a graph-structured fusion process (Fig. 4a): each node maintains a state representation of an agent in its own coordinate frame, and for a fixed number of iterations, spatially warped messages are shared between nodes and the node states are updated based on the aggregated features via the transformer blocks.", + "bbox": [ + 496, + 659, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall process is summarized in Alg. 1. During each iteration, we have two cascaded node updates which capture the local and global heterogeneous interactions respectively. For each node, we first spatially transform [18] neighboring nodes' features to its center $\\mathbf{F}_{j\\rightarrow i} = \\Gamma_{j\\rightarrow i}(\\mathbf{F}_j)$ . When the transmitting node is the receiving node itself, the transformation matrix is an identity matrix", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "288", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0ee682ff98423e6f75a4a3cc9c480c242cb054f90610172fee96c706b5233b8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsV2V-CV2V-LV2V-H
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
No Fusion0.0940.0210.5240.3630.2840.157
Late Fusion0.2310.0700.7700.6060.5020.308
V2VNet [40]0.3290.1250.8200.6450.6500.366
DiscoNet [22]0.2870.1150.7410.5900.6240.385
AttFuse [49]0.2610.0950.8010.6440.6470.390
CoBEVT [46]0.3170.1220.8280.6370.6740.416
V2X-ViT [48]0.3320.1250.8330.6790.6710.427
HM-ViT0.3550.1420.8530.7630.6950.515
", + "bbox": [ + 163, + 87, + 805, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Evaluation of V2V perception methods on OPV2V dataset.", + "bbox": [ + 259, + 266, + 705, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and thus $\\mathbf{F}_{i\\rightarrow i} = \\mathbf{F}_i$ . These spatial aligned feature maps $\\mathbf{F}_{j\\rightarrow i}$ are then shared with agent $i$ to update its state representation via the aggregation module. We adopt $\\mathrm{H}^3\\mathrm{GAT - L}$ Block as our first aggregation module to capture the local heterogeneous interactions and leverage $\\mathrm{H}^3\\mathrm{GAT - G}$ Block for the second module to further refine the states with global cues. Within each transformer block, we also adopt a mask to mask out non-overlapping areas between the field of views when computing the attention scores. Note that each agent's state update can be processed in parallel for better efficiency. After L such iterations, we pass the features to a hetero-modal MLP to further refine the feature representation. Throughout the whole fusion process, the modality-specific statistics are maintained.", + "bbox": [ + 75, + 308, + 470, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Hetero-modal Head", + "text_level": 1, + "bbox": [ + 76, + 529, + 264, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As camera and LiDAR contain distinct characteristics, we design a hetero-modal head where a different set of parameters are applied for camera and LiDAR ego vehicles to generate the final predictions. More specifically, the final fused feature maps are passed to a series of $3 \\times 3$ convolutions with batch normalization and ReLU for feature refinement. Then, we adopt a $1 \\times 1$ convolution layer to generate the regression and classification predictions. Smooth $\\ell_{1}$ loss is utilized for regression and a focal loss [25] is used for classification.", + "bbox": [ + 75, + 551, + 468, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation", + "text_level": 1, + "bbox": [ + 76, + 741, + 299, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OPV2V. OP2V [49] is a large-scale multi-modal cooperative V2V perception dataset collected in CARLA [12] and OpenCDA [45]. It contains over 70 driving scenarios of around 25 seconds duration each. Each scenario contains multiple connected AVs (2 to 7) and each AV is equipped with 1 LiDAR and 4 monocular cameras covering $360^{\\circ}$ horizontal field of view (FoV). In our hetero-modal cooperative perception setting, we only use one type of sensor modality for each AV, leading to two types of agents: vehicles only", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "equipped with multiple cameras (camera agent), and vehicles only equipped with LiDARs (LiDAR agent).", + "bbox": [ + 496, + 309, + 888, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Evaluations.** We adopt Average precision (AP) at Intersection-over-Union (IoU) 0.5 and 0.7 to measure the perception performance. As each scenario consists of multiple AVs, a fixed agent is selected as the ego agent and the evaluation is conducted in the range of $100\\mathrm{m} \\times 100\\mathrm{m}$ around it. Following [49, 48], the train/validation/test splits are 6764/1981/2719. We evaluate models mainly under three configurations: 1) V2V Camera-based 3D detection (V2V-C) where AVs are only equipped with 4 cameras with 360 horizontal FoV, 2) V2V LiDAR-based detection (V2V-L) where all the agents only have LiDAR sensors, and 3) V2V Hetero-modal detection (V2V-H) where half of the agents only has cameras while the other half only has LiDARs. To further assess models' capability with dynamic sensor configurations, we also assess models with fixed ego sensor modality and varying collaborator sensor types.", + "bbox": [ + 496, + 339, + 892, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Setups", + "text_level": 1, + "bbox": [ + 500, + 590, + 694, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use the PointPillar [20] as 3D backbones and a modified BEVFormer [23] for our camera stem. For BEVFormer, we adopt its variant with no temporal information and ResNet50 [15] as image backbones for better computation efficiency and use a smaller grid resolution $(0.4\\mathrm{m})$ to preserve fine-trained spatial details. The intermediate BEV feature map has a dimension of $128\\times 128\\times 256$ . Following prior works [22, 49, 48, 46], we only change the fusion module for different intermediate fusion methods while keeping the other components such as hetero-modal header, compression, and lidar/camera feature extractors the same. For the heterogeneous methods, we provide the network with additional feature types so that they can distinguish whether the feature is from Li-DAR or camera while for the homogeneous methods, the fusion networks would treat the features from LiDAR and camera equally. For HM-ViT, we conduct two iterations of graph-structured feature fusion and employ a window size of 8 for both local and global attentions. We adopt AdamW", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "289", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/072da0bf6f1f36e6688fa35fb69d0193c768a2ececc5be919b4148f17cca1981.jpg", + "image_caption": [ + "Figure 6: Agent modality ratio experiment. The x-axis is the ratio of LiDAR collaborators among all the collaborators. In (a) and (b), ego vehicles are equipped with cameras. In (c) and (d), ego vehicles are equipped with LiDARs." + ], + "image_footnote": [], + "bbox": [ + 84, + 104, + 282, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f11ffca66134f4c86478c9ef2a53f2dc66a9da005959bf656fb4060b18fbea9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 104, + 483, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b60d85e61a24b6d3577254dc92d172a28c088b24c62b6da0d60bc51a95fda620.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 104, + 681, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7087ab578d79b0dabf77d590c1ef11e3b339237b253faf099c27b405a32a7d61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 104, + 883, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/051c2e88d1efd0531c58309bad1027680e09f5d57b9544bae1225dd866f3ed9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 229, + 754, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/061acd15789cec06515163b772ff539ce1b898e70b4d7b4d80789f635db8730e.jpg", + "image_caption": [ + "Figure 7: Ablation study on number of agent. The LiDAR/camera collaborator refers to the case where the collaborators are equipped with LiDAR and Camera respectively. (a) ego vehicle is equipped with camera. (b) ego vehicle is equipped with LiDAR." + ], + "image_footnote": [], + "bbox": [ + 78, + 301, + 271, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/614a9178da5e734af03673af75ea16f4e44f1620135efcdb9a5e56cef8d70164.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 301, + 467, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with a decay rate of $10^{-2}$ and cosine annealing learning rate scheduler to optimize the models.", + "bbox": [ + 75, + 553, + 468, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training strategy. We find that the intermediate fusion models won't converge if directly trained end-to-end under V2V-H and the resulting models usually only exhibit good performance for either camera perception or LiDAR perception but hardly for both. Instead, we first train the model on single modality configurations (i.e., V2V-C and V2V-L) until convergence for 40 epochs and then fine-tune the models under V2V-H for 10 epochs with fixed parameters of modality-specific backbones on 4 RTX3090 GPUs. This training strategy can help models converge well with reliable performance. For a fair comparison, we leverage this training strategy for all the methods.", + "bbox": [ + 75, + 584, + 470, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared methods. We regard No Fusion as the baseline method. We also evaluate the Late Fusion, which transmits the detection proposals and leverages Non-maximum suppression to generate the final predictions. For the intermediate cooperation methods, we benchmark five approaches: V2VNet [40], DiscoNet [22], AttFuse [49], CoBEVT [46], and V2X-ViT [48]. For a fair comparison, hetero-modal head is used for all the models.", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 500, + 304, + 718, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Main performance comparison. Tab. 1 demonstrates the performance comparisons on V2V-C, V2V-L and V2V-H. Under all three settings, all the cooperative methods outperform the No Fusion baseline and the intermediate fusion beats the classical Late Fusion, showing the great benefit of end-to-end V2V hetero-modal cooperative perception. The HM-ViT outperforms all the other SOTA intermediate fusion methods by at least $1.7\\%$ , $8.4\\%$ , $8.8\\%$ in AP@0.7 under V2V-C, V2V-L, V2V-H settings respectively.", + "bbox": [ + 498, + 334, + 893, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Agent modality ratio experiment. As shown in Fig. 6, we fix the ego vehicle sensor modality and vary the ratio of collaborators' sensor modalities to evaluate the models' performance under various heterogeneous traffic scenarios. The larger LiDAR collaborator ratio corresponds to more vehicles only equipped with LiDARs while a smaller ratio means more vehicles only equipped with multiple cameras. The left two figures are the evaluation results for camera ego vehicles while the right two figures are the performance for LiDAR ego vehicles. Under most ratios, late fusion outperforms No Fusion however for LiDAR ego vehicle when most collaborators are camera agents, the Late Fusion performs poorer than No Fusion. We argue this is due to the fact that the camera predictions are usually noisy and merging proposals from different modalities equally could lead to ambiguity and thus deteriorate the performance. On the other hand, all the intermediate fusion methods outperform No Fusion by a large margin especially for the camera ego vehicles, demonstrating the great value of V2V cooperation between agents with different modalities. Among all the compared methods, HM-ViT ranks first for both camera ego vehicles and LiDAR ego vehicles under all the ratios, illustrating the great capability of HM-ViT for capturing modality-specific characteristics and cross-modality multiagent interactions. In contrast, other intermediate fusion methods only show good performance for a certain ratio range, which demonstrates the importance of heterogeneity for hetero-modal cooperative perception.", + "bbox": [ + 496, + 477, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "290", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0302c822f4815ba85987e0b79d88ef98a2ecb234e13bb4e3626acd9a62379a4.jpg", + "image_caption": [ + "(a) No Fusion (camera)" + ], + "image_footnote": [], + "bbox": [ + 117, + 138, + 359, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b06526066becb11c12c7d52956d9afcda555dcfa913998acfed321f51bf1e67b.jpg", + "image_caption": [ + "(b) Pure camera fusion" + ], + "image_footnote": [], + "bbox": [ + 370, + 138, + 611, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d199ca4ec200dd1d16a73ba30fe6193230c541a46c02b47ac70a6ad5644be41.jpg", + "image_caption": [ + "(c) Camera ego with LiDAR collaborators" + ], + "image_footnote": [], + "bbox": [ + 616, + 138, + 857, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccf7e5f1d7f260f42057f67baa282ebe14177328d73e10bfee06dc66dd4831d8.jpg", + "image_caption": [ + "(d) No Fusion (LiDAR)" + ], + "image_footnote": [], + "bbox": [ + 120, + 258, + 361, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef4e4fff65701a1ef4f443190d785853f10e545dec948c8755ce768e3ff87e23.jpg", + "image_caption": [ + "(e) Pure LiDAR fusion", + "Figure 8: Qualitative visualizations for (a) No Fusion with camera ego vehicle, (b) pure camera-based V2V perception, (c) hetero-modal V2V perception with camera ego vehicle and LiDAR collaborators, (d) No Fusion with LiDAR ego vehicle, (e) pure LiDAR-based V2V perception, and (f) hetero-modal V2V perception with LiDAR ego vehicle and camera collaborators. The red and green boxes represent the detection outputs and ground truth respectively. More visualizations can be found in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 367, + 258, + 607, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fbd1efc9248808eed6f40343152ebe4ec46e001b097f025d603bd9b92c4446f.jpg", + "image_caption": [ + "(f) LiDAR ego with camera collaborators" + ], + "image_footnote": [], + "bbox": [ + 614, + 258, + 856, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bc54f17c776baf52bcd89f23d88cd6745e8fbc302a6cc2c45108308df12d2a01.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HM-MLP&LNH3GAT-LH3GAT-GAP@0.7
0.404
✓0.420
✓✓0.460
✓✓✓0.515
", + "bbox": [ + 81, + 470, + 465, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of agent. In this experiment, we investigate the effect of the number of agents on the perception performance of HM-ViT. As Fig. 7 depicts, for both camera and LiDAR ego vehicles, the perception performance increases as more agents are involved in the cooperative perception and both LiDAR and camera collaborators can contribute to the performance gain for ego vehicles with different modalities, which again shows the benefit of hetero-modal V2V cooperation. Additionally, the increase rate generally decreases when increasing the number of agents and the LiDAR collaborators can bring more AP gains over the camera collaborators for both camera and LiDAR ego vehicles. Notably, similar as Fig. 6a-b display, the camera ego vehicles' performance can be greatly improved with only a small number of LiDAR collaborators, demonstrating the great potential of reducing the cost for each vehicle when the V2V system is deployed at scale as we may only need to install expansive LiDARs for a small number of agents (e.g., infrastructure) while all the other agents only require relatively cheap camera sensors.", + "bbox": [ + 75, + 598, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/77b49c34ef2c7e2a67942921c6b8758377dfc130ad6d6e4bb8cc92e41fefd57f.jpg", + "table_caption": [ + "Table 2: Component Ablation study on the V2V-H setting" + ], + "table_footnote": [], + "table_body": "
Compression RateAP@0.7
0x0.515
8x0.513
16x0.470
32x0.455
", + "bbox": [ + 581, + 470, + 810, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Compression rate effects for HM-ViT on V2V-H.", + "bbox": [ + 503, + 566, + 885, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Component ablation study. Here we investigate the key components of the proposed HM-ViT. As the layer normalization and MLP are usually combined together in typical transformer designs, thus we jointly evaluate the combined effect of HM-MLP and HM-LN (HM-MLP&LN). As Tab. 2 shows, all the proposed components improve the performance and local and global attentions can largely improve the AP@0.7 by $4\\%$ and $5.5\\%$ , proving the great benefit brought by jointly reasoning inter-agent and intra-agent heterogeneous interactions both locally and globally.", + "bbox": [ + 496, + 595, + 890, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compression rate. Tab. 3 describes the influence of compression rate. It demonstrates that HM-ViT can still outperform other methods even under large compression rates.", + "bbox": [ + 496, + 747, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 500, + 801, + 676, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 8 depicts the qualitative visualizations for HM-ViT and No Fusion baselines. In fig. 8a-c, we plot the detection results for camera ego vehicles with no collaborator, camera collaborators, and LiDAR collaborators respectively while for Fig. 8d-f, we plot the results for LiDAR ego vehicles. As", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "291", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in these figures, the collaborations with both homogeneous and heterogeneous agents are beneficial for camera ego vehicles and LiDAR ego vehicles with enhanced detection results. In particular, the collaboration between camera ego vehicles and LiDAR collaborators can dramatically enhance the perception performance.", + "bbox": [ + 80, + 90, + 467, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 80, + 194, + 192, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we present HM-ViT, a hetero-modal vision transformer, for the hetero-modal multi-agent cooperative perception problem which is an important but under-explored research direction. We propose a generic heterogeneous 3D graph attention to jointly reason heterogeneous inter-agent and cross-agent interactions. Our extensive experiments demonstrate the outstanding performance of the proposed method and the great potential of hetero-modal multi-agent collaborations for increasing the scalability and robustness of V2V systems. We hope our findings and open-source efforts will inspire more research on this new problem.", + "bbox": [ + 80, + 220, + 467, + 401 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "292", + "bbox": [ + 486, + 946, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luís A Alexandre. 3d object recognition using convolutional neural networks with transfer learning between input channels. In Intelligent Autonomous Systems 13, pages 889-898. Springer, 2016. 3", + "[2] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2", + "[3] Johannes Betz, Hongrui Zheng, Alexander Liniger, Ugo Rosolia, Phillip Karle, Madhur Behl, Venkat Krovi, and Rahul Mangharam. Autonomous vehicles on the edge: A survey on autonomous vehicle racing. IEEE Open Journal of Intelligent Transportation Systems, 3:458-488, 2022. 1", + "[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 3", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3", + "[6] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 3", + "[7] Hung-Hsun Chen, Yi-Bing Lin, I-Hau Yeh, Hsun-Jung Cho, and Yi-Jung Wu. Prediction of queue dissipation time for mixed traffic flows with deep learning. IEEE Open Journal of Intelligent Transportation Systems, 3:267-277, 2022. 1", + "[8] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016. 3", + "[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 3", + "[10] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Graph-detr3d: Rethinking overlapping regions for multi-view 3d object detection. arXiv preprint arXiv:2204.11582, 2022. 3", + "[11] Zhuo Deng and Longin Jan Latecki. A modal detection of 3d objects: Inferring 3d bounding boxes from 2d ones in rgb-depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5762-5770, 2017. 3", + "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driv" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6", + "[13] Yi Guo and Jiaqi Ma. Leveraging existing high-occupancy vehicle lanes for mixed-autonomy traffic management with emerging connected automated vehicle applications. Transportmetrica A: Transport Science, 16(3):1375–1399, 2020. 1", + "[14] Yi Guo, Jiaqi Ma, Edward Leslie, and Zhitong Huang. Evaluating the effectiveness of integrated connected automated vehicle applications applied to freeway managed lanes. IEEE Transactions on Intelligent Transportation Systems, 23(1):522-536, 2020. 1", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[16] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. arXiv preprint arXiv:2209.12836, 2022. 2", + "[17] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3", + "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 5", + "[19] Sou Kitajima, Hanna Chouchane, Jacobo Antona-Makoshi, Nobuyuki Uchida, and Jun Tajima. A nationwide impact assessment of automated driving systems on traffic safety using multiagent traffic simulations. IEEE Open Journal of Intelligent Transportation Systems, 3:302-312, 2022. 1", + "[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 6", + "[21] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In European Conference on Computer Vision, pages 316-332. Springer, 2022. 2", + "[22] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 2, 6, 7", + "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3, 4, 6", + "[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3", + "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "293", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 3", + "[27] Jiaqi Ma, Edward Leslie, Amir Ghiasi, Zhitong Huang, and Yi Guo. Empirical analysis of a freeway bundled connected-and-automated vehicle application using experimental data. Journal of Transportation Engineering, Part A: Systems, 146(6):04020034, 2020. 1", + "[28] Vasileia Papathanasopoulou, Ioanna Spyropoulou, Harris Perakis, Vassilis Gikas, and Eleni Andrikopoulou. A data-driven model for pedestrian behavior classification and trajectory prediction. IEEE Open Journal of Intelligent Transportation Systems, 3:328-339, 2022. 1", + "[29] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 3", + "[30] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgbd data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 3", + "[31] Kelli Raboy, Jiaqi Ma, Edward Leslie, and Fang Zhou. A proof-of-concept field experiment on cooperative lane change maneuvers using a prototype connected automated vehicle testing platform. Journal of Intelligent Transportation Systems, 25(1):77-92, 2021. 1", + "[32] Steven E Shladover. Opportunities and challenges in cooperative road vehicle automation. IEEE Open Journal of Intelligent Transportation Systems, 2:216-224, 2021. 1", + "[33] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 3", + "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3", + "[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 5", + "[36] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. arXiv preprint arXiv:2204.01697, 2022.5", + "[37] Rodolfo Valiente, Behrad Toghi, Ramtin Pedarsani, and Yaser P Fallah. Robustness and adaptability of reinforcement learning-based cooperative autonomous driving in mixed-autonomy traffic. IEEE Open Journal of Intelligent Transportation Systems, 3:397-410, 2022. 1", + "[38] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3", + "[39] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3", + "[40] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 2, 6, 7", + "[41] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3", + "[42] Zhixin Wang and Kui Jia. Frustum convnet: Sliding frustums to aggregate local point-wise features for amodal 3d object detection. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1742-1749. IEEE, 2019. 3", + "[43] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. arXiv preprint arXiv:2209.13679, 2022. 2", + "[44] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3", + "[45] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6", + "[46] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 2, 6, 7", + "[47] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. arXiv preprint arXiv:2303.07601, 2023. 1", + "[48] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. arXiv preprint arXiv:2203.10638, 2022. 1, 2, 6, 7", + "[49] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 6, 7", + "[50] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "294", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 1", + "[52] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 1", + "[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1", + "[54] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 1", + "[55] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 373 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "295", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_model.json b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_model.json new file mode 100644 index 0000000000000000000000000000000000000000..13ab907c979ce1ec4dc216a3cdced45f5d2000bf --- /dev/null +++ b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_model.json @@ -0,0 +1,2127 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.88, + 0.175 + ], + "angle": 0, + "content": "HM-ViT: Hetero-modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.203, + 0.654, + 0.258 + ], + "angle": 0, + "content": "Hao Xiang\\(^{1}\\), Runsheng Xu\\(^{1}\\), Jiaqi Ma\\(^{1}\\) \n\\(^{1}\\)University of California, Los Angeles \n{haxiang, rxx3386, jiaqima}@g.ucla.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.658 + ], + "angle": 0, + "content": "Vehicle-to-Vehicle technologies have enabled autonomous vehicles to share information to see through occlusions, greatly enhancing perception performance. Nevertheless, existing works all focused on homogeneous traffic where vehicles are equipped with the same type of sensors, which significantly hampers the scale of collaboration and benefit of cross-modality interactions. In this paper, we investigate the multi-agent hetero-modal cooperative perception problem where agents may have distinct sensor modalities. We present HM-ViT, the first unified multi-agent hetero-modal cooperative perception framework that can collaboratively predict 3D objects for highly dynamic Vehicle-to-Vehicle (V2V) collaborations with varying numbers and types of agents. To effectively fuse features from multi-view images and LiDAR point clouds, we design a novel heterogeneous 3D graph transformer to jointly reason inter-agent and intra-agent interactions. The extensive experiments on the V2V perception dataset OPV2V demonstrate that the HM-ViT outperforms SOTA cooperative perception methods for V2V hetero-modal cooperative perception. Our code will be released at https://github.com/XHwind/HM-ViT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.21, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Recent advances in Vehicle-to-Vehicle (V2V) communication technology and intelligent transportation systems [13, 31, 27, 14, 7, 19, 28, 51, 53, 52] have allowed autonomous vehicles (AVs) to share sensory information, enabling them to perceive their surroundings better [3, 32, 54]. With the rapid growth of autonomous driving, V2V perception systems have the potential to be deployed at scale and create a safer transportation system. Cooperative perception systems, as shown in recent studies [49, 48, 40], can intelligently aggregate features from multiple vehicles within the communication range to enhance visual reasoning and overall performance." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.291, + 0.895, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.448, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Figure 1: Illustration of multi-agent hetero-modal V2V systems where each agent may be equipped with either LiDAR or multi-view cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Despite the rapid growth in this field, previous studies [49, 40, 46, 22, 37, 47] have primarily focused on homogeneous multi-agent cooperative perception, where all agents are equipped with the same type of sensors. In reality, however, agents may have different sensor modalities (hetero-modality) due to the cost and sensor preferences of ADS developers and car makers. As shown in Fig. 1, some agents are equipped with only LiDARs (LiDAR agents), while others only have multiple cameras (camera agents). Enabling collaboration between these heterogeneous agents could improve the sensing capability by allowing agents to see through occlusions and increase the scale and reliability of V2V systems. Additionally, LiDAR agents can provide accurate geometric information, while camera agents can provide rich semantic context. Thus, the collaboration between these agents could leverage the distinct but complementary environment attributes captured by each sensor modality to enhance the V2V perception systems. Furthermore, compared with the single-agent solution where multiple LiDARs and cameras are installed in a single vehicle, distributing different types of sensors across distinct agents could also potentially decrease the costs for each agent while still achieving satisfying performance. Nevertheless, whether, when, and how multi-agent hetero-modal V2V cooperation can benefit the perception system of heterogeneous traffic has not yet been studied." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "284" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.096, + 0.437, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.471, + 0.327 + ], + "angle": 0, + "content": "Figure 2: Comparison of the single-agent multi-modal system and multi-agent hetero-modal system. The graph structure of the former is fixed whereas for the latter the graph is both dynamic and heterogeneous." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.732 + ], + "angle": 0, + "content": "In this work, we address the multi-agent hetero-modal cooperative perception problem where each agent could have distinct sensor types and share/receive information with each other. Notably, as shown in Fig. 2, this multi-agent hetero-modal setting is distinct from the single-agent multi-modal setting. In the hetero-modal setting, the agent sensors form a dynamic heterogeneous graph where the existence and types of sensors are random, and the relative poses vary from scene to scene. In contrast, the sensor types/numbers and relative positions (extrinsics) between sensors are fixed in the single-agent multi-modal setting. Existing multi-modal methods heavily rely on these assumptions, and most of existing works [38, 39, 2, 55, 50] transform LiDAR points or 3D proposals onto the image plane to index 2D features. Their network architectures build upon the co-existence of both LiDAR and camera inputs with fixed geometric relationships. However, the dynamic nature of hetero-modal V2V perception requires a flexible architecture that can handle varying agent numbers and types, and the transmitted neural features are also spatially misaligned. Moreover, there are semantic discrepancies in the transmitted features between camera agents and LiDAR agents. Hence, these unique characteristics pose significant challenges for designing the multi-agent hetero-modal cooperative system and prevent adapting existing multi-modal fusion methods to this new problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To enable collaboration between heterogeneous agents in V2V systems, we propose Hetero-Modal Vision Transformer (HM-ViT), the first unified cooperative perception framework that can leverage and fuse distributed information for hetero-modal V2V perception via a spatial-aware 3D heterogeneous vision transformer. Fig. 3 demonstrates the overall framework. Each agent first generates bird's eye view (BEV) representations through modality-specific encoders and then shares compressed features with neighboring agents. Afterward, the received features are decompressed and aggregated via the proposed HM-ViT," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "which conducts joint local and global heterogeneous 3D attentions with the consideration of both node and edge types. Our extensive experiments show that the HM-ViT can significantly improve the perception capability of camera agents and LiDAR agents over the single-agent baseline and outperforms SOTA cooperative perception methods by a large margin. In particular, for camera agents, performance can be boosted from \\(2.1\\%\\) to \\(53.2\\%\\) at \\(\\mathrm{AP}@0.7\\) with the collaboration of LiDAR agents, a 23-fold improvement. Our primary contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.25, + 0.892, + 0.371 + ], + "angle": 0, + "content": "- We present the novel transformer framework (HMViT) for multi-agent hetero-modal cooperative perception, capable of capturing the modality-specific characteristics and heterogeneous 3D interactions. The proposed model exhibits superior flexibility and robustness with state-of-the-art performance on highly dynamic heterogeneous traffic involving varying agent numbers and types." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.38, + 0.892, + 0.47 + ], + "angle": 0, + "content": "- We propose a generic heterogeneous 3D graph attention (H\\(^3\\)GAT), tailored for extracting inter-agent and intra-agent heterogeneous interactions. We instantiate two such attentions - local attention (H\\(^3\\)GAT-L) and global attention (H\\(^3\\)GAT-G) for capturing both local and global visual cues." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.479, + 0.892, + 0.554 + ], + "angle": 0, + "content": "- We conduct extensive benchmark experiments by varying sensor modalities, demonstrating the strong performance of the proposed method for hetero-modal V2V perception tasks. We will release all the codes and baselines to facilitate future research." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.25, + 0.892, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.637, + 0.589 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "V2V perception. V2V perception aims to enhance the perception performance of autonomous vehicles by leveraging shared information from other connected vehicles. Existing works have primarily focused on LiDAR-based 3D object perception. The pioneer cooperative perception methods transmit raw sensing observation (i.e., early fusion) or perception outputs (i.e., late fusion), whereas recent works [49, 40, 16, 43, 21, 48] are exploring the use of circulating intermediate neural features for achieving better performance-bandwidth trade-off. V2VNet [40] employs graph neural networks to aggregate the shared neural features for joint detection and prediction. AttFuse [49] uses single-head attention to model the per-location multi-agent interaction. Disconet [22] presents a matrix-valued edge weight for learning the interactions and a teacher-student learning framework to facilitate the training. V2X-ViT explores [48] vision transformer for vehicle-to-everything cooperation via window attention and heterogeneous self-attention. CoBEVT [46] presents a generic transformer framework for camera-based BEV semantic segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "285" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.889, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.896, + 0.373 + ], + "angle": 0, + "content": "Figure 3: Overview of proposed hetero-modal V2V cooperative perception system. Each agent first produces BEV features through modality-specific feature extractors (Sec. 3.1). The BEV features are then compressed and shared (Sec. 3.3) with neighboring connected agents and the received features are decompressed in the ego agent side and fed into hetero-modal vision transformer to conduct graph-structured feature fusion (Sec. 3.2 and Sec. 3.3). The refined features are finally passed into the hetero-modal detection head to predict 3D bounding boxes (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.471, + 0.476 + ], + "angle": 0, + "content": "However, none of the existing works explored multi-agent multi-camera 3D object detection, let alone multi-agent hetero-modal perception. In contrast to existing methods, HM-ViT is the first to employ sparse heterogeneous local and global attentions to capture the 3D inter-agent and intraagent interactions in a computationally efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.471, + 0.794 + ], + "angle": 0, + "content": "Camera-based 3D object detection. Early works [4, 8, 33] mainly focus on monocular 3D detection but a single camera can only provide a 2D view of the scene, and inferring 3D from 2D is intrinsically hard. Recent development of self-driving datasets [5, 34, 6] featured with full sensor suits has enabled the research direction of 3D object detection from multiple cameras. DETR3D [41] proposes a 3D-2D query paradigm for extracting 3D features from 2D multi-view images. Graph-DETR3D [10] further utilizes graph structure learning to enhance the representation at the border regions. Alternatively, another stream of works [17, 29, 44, 23] focuses on aggregating BEV features from multi-view cameras for conducting downstream perception tasks. LSS [29] lifts the 2D features to 3D frustum via latent depth and then splats frustums into a BEV grid. \\(\\mathbf{M}^2\\mathbf{BEV}\\) [44] further extends LSS with less memory consumption and conducts detection and segmentation. BEV-Former [23] constructs BEV queries and explores spatial cross-attention and temporal self-attention to recurrently refine the BEV features, achieving SOTA performance on both NuScenes [5] and Waymo Open Datasets [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.904 + ], + "angle": 0, + "content": "Multi-modal fusion. Existing multi-sensor fusion methods can be divided into point-level fusion, proposal-level fusion, and BEV-level fusion. Point-level fusion decorates the input from one modality with attributes from the other modality. PointPainting [38] and PaintAugmenting [39] decorate the LiDAR point clouds with semantic segmentation scores and 2D CNN image features respectively while [11, 1] project" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.895, + 0.658 + ], + "angle": 0, + "content": "LiDAR onto the image plane to augment the RGB values with depth information and conduct detection. On the other hand, proposal-level fusion generates proposals from one modality and then indexes features from the other modality for further refinement. MV3D [9] produces object queries from LiDAR BEV and then extracts the features from camera data and LiDAR front view. [30, 42] lift 2D bounding boxes to frustum and conduct 3D object detection from the frustum of point clouds. Conversely, BEV-level fusion converts features from different modalities to unified BEV representations, preserving both geometric and semantic information. BEVFusions [26, 24] concatenate BEV features from camera and LiDAR and fuse it via a fusion module. Hence, existing single-agent multi-modal fusion methods rely on the co-existence of both camera and LiDAR with fixed geometric relationships, which is unsuitable for our multi-agent hetero-modal cooperative perception problem with a dynamic heterogeneous collaboration graph." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.637, + 0.73 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we explore the multi-agent hetero-modal cooperative perception, where each AV is equipped with either a LiDAR or multiple cameras. Our goal is to create a robust and flexible cooperative perception system that allows for efficient collaboration between any number of agents with varying sensor types, ultimately improving the perception capabilities of the vehicle in a unified end-to-end fashion. The pipeline, illustrated in 3, includes modality-specific feature extraction, compression and sharing, HMViT for feature fusion, and hetero-modal detection head." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.101, + 0.291, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.108, + 0.605, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.114, + 0.887, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.893, + 0.325 + ], + "angle": 0, + "content": "Figure 4: HM-ViT architecture. (a) Graph-structured fusion process. (b) Heterogeneous 3D graph attention \\(\\mathrm{(H^3GAT)}\\). (c) Illustration of sampled tokens for \\(\\mathrm{H}^{3}\\mathrm{GAT - L}\\) and \\(\\mathrm{H}^{3}\\mathrm{GAT - G}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.337, + 0.388, + 0.354 + ], + "angle": 0, + "content": "3.1. Modality-specific feature extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.47, + 0.452 + ], + "angle": 0, + "content": "LiDAR stem: We leverage PointPillar [20] to process point clouds for each LiDAR agent. The raw point cloud is converted to a 2D pseudo-image, flattened along the height dimension, and fed into 2D convolutional neural networks to produce the salient feature map \\(\\mathbf{F}_j \\in \\mathcal{R}^{H \\times W \\times C}\\), which is compressed and shared with all the neighboring agents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.473, + 0.681 + ], + "angle": 0, + "content": "Camera stem: Each camera agent is equipped with \\( m \\) monocular cameras. The sensing observation of \\( i \\)-th agent includes the input images \\( I_k^i \\in \\mathcal{R}^{h \\times w \\times 3} \\) and the known projection matrix \\( P_k^i \\in \\mathcal{R}^{3 \\times 4} \\) that maps 3D reference points to different image views. Our goal is to generate a BEV feature representation \\( \\mathbf{F}_i \\in \\mathcal{R}^{H \\times W \\times C} \\) that is amenable for feature fusion with other collaborators. In this work, we adopt similar architecture to BEVFormer [23] with no temporal information for feature extraction. For a faster running time, we adopt ResNet50 to extract 2D image features and then adopt a learnable 2D BEV query to inquire spatial information from the encoded multi-view features via spatial cross attention and projection matrices. The resulting refined BEV feature \\( \\mathbf{F}_i \\) is centered around agent \\( i \\) and shared with connected AVs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.462, + 0.712 + ], + "angle": 0, + "content": "3.2. Heterogeneous 3D Graph Attention (H\\(^3\\)GAT)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To account for the distinct characteristics of BEV features extracted from different sensor modalities, the learning process of each modality must be distinguished, and the cross-modality interactions between multiple agents should vary. To capture this heterogeneity, we present a novel heterogeneous 3D graph attention (H\\(^3\\)GAT), in which nodes and edges are type-dependent to reason spatial interactions and cross-agent relations jointly. We encode both local and global interactions to better capture the 3D ambiguity in BEV feature space. Local attention can help preserve object details, while global attention can provide a better understanding of environmental contexts such as road topology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.338, + 0.625, + 0.352 + ], + "angle": 0, + "content": "and traffic density." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.519 + ], + "angle": 0, + "content": "As shown in Fig. 4b, We build a 3D heterogeneous collaboration graph. Each node \\( v(i,x) = \\mathbf{F}_x^i \\in \\mathcal{R}^C \\) is a feature vector of agent \\( i \\)'s feature map at spatial location \\( x \\in \\mathcal{R}^2 \\). 3D heterogeneous graph attention is performed for spatially connected nodes in the BEV feature space. Depending on the definition of spatial connectivity, we will derive local attention and global attention. Here for notation simplicity, we only derive single-head equations but in real implementations, multi-head variants are used. Formally, we first project feature vectors onto different feature spaces to form query, key, and value vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.527, + 0.891, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {x} ^ {j} = \\operatorname {D e n s e} _ {\\tau_ {j}} \\mathbf {F} _ {x} ^ {j} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.548, + 0.891, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {x} ^ {j} = \\operatorname {D e n s e} _ {\\tau_ {j}} \\mathbf {F} _ {x} ^ {j} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.568, + 0.891, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {x} ^ {j} = \\operatorname {D e n s e} _ {e _ {i j}} \\mathbf {F} _ {x} ^ {j} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.715 + ], + "angle": 0, + "content": "where the \\(\\mathrm{Dense}\\square\\) is a set of linear layers indexed by subscript \\(\\square\\). For the query and key vectors, we use linear projectors \\(\\mathrm{Dense}_{\\tau_j}\\) indexed by node type \\(\\tau_{j}\\) to extract modality-specific features. For the value vector, we index the projector via edge type \\(\\mathrm{Dense}_{e_{ij}}\\) to reflect the heterogeneity of cross-modality multi-agent interactions. The set of connected nodes of \\(v(i,x)\\) is denoted as \\(\\mathcal{N}(i,x)\\). Then the attention is operated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.722, + 0.892, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} (j, y) = \\operatorname {S o f t m a x} _ {(j, y) \\in \\mathcal {N} (i, x)} \\left(\\mathbf {Q} _ {x} ^ {i} \\mathbf {W} _ {e _ {i j}} \\mathbf {K} _ {y} ^ {j}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.751, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {x} ^ {i} = \\sum_ {(j, y) \\in \\mathcal {N} (i, x)} \\mathbf {a} (j, y) \\mathbf {V} _ {y} ^ {j} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.827 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{e_{ij}}\\in \\mathcal{R}^{C\\times C}\\) is used to adjust the dot product of \\(\\mathbf{Q}_x^i\\) and \\(\\mathbf{K}_y^j\\) to further encode the heterogeneity of edges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Depending on how the nodes are sampled (Fig. 4c), we design two types of attentions: local attention \\((\\mathbf{H}^3\\mathbf{GAT - L})\\) which performs local window-based attention and global attention \\((\\mathbf{H}^3\\mathbf{GAT - G})\\) which performs sparse global grid-based attention. Fig. 4c visualizes how the local and global" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.25, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.326, + 0.235, + 0.342 + ], + "angle": 0, + "content": "(a) \\(\\mathsf{H}^3\\mathsf{GAT - L}\\) Block" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.437, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.326, + 0.428, + 0.342 + ], + "angle": 0, + "content": "(b) \\(\\mathsf{H}^3\\mathsf{GAT - G}\\) Block" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.387 + ], + "angle": 0, + "content": "Figure 5: Transformer blocks for local and global attentions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.469, + 0.492 + ], + "angle": 0, + "content": "tokens are sampled, where the local tokens are sampled within local windows and the global tokens are sparsely sampled grids scattered across feature maps. The local interactions can help preserve spatial cues and provide reliable estimates while the global reasoning can help understand global semantic context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.704 + ], + "angle": 0, + "content": "Both \\(\\mathrm{H}^3\\mathrm{GAT - L}\\) and \\(\\mathrm{H}^3\\mathrm{GAT - G}\\) can be implemented efficiently by decomposing the spatial axes. More specifically, we stack all the agents' features to \\(\\mathbf{F}\\in \\mathcal{R}^{N\\times H\\times W\\times C}\\) where \\(N\\) is the number of agents. For \\(\\mathrm{H}^3\\mathrm{GAT - L}\\), we decompose the feature map into 3D non-overlapping windows along the first axis [35, 36], each of size \\(N\\times P\\times P\\). The partitioned tensor has the shape \\((\\frac{H}{P}\\times \\frac{W}{P},N\\times P^2,C)\\) where the heterogeneous 3D local graph attention is conducted for \\(NP^2\\) tokens within the same window. Similarly, for \\(\\mathrm{H}^3\\mathrm{GAT - G}\\), we swap the axis and partition the tensor into the shape \\((N\\times P^2,\\frac{H}{P}\\times \\frac{W}{P},C)\\). Due to the swap operation, the sampled grids will be sparsely scattered and the attention is operated for these sparsely sampled \\(\\frac{H}{P}\\times \\frac{W}{P}\\) grids, which can capture sparse global information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To integrate this local and global attention into transformer architecture, we further present a heterogeneous normalization layer (HM-LN) and heterogeneous MLP (HMMLP) which use type-dependent parameters. As shown in Fig. 5, we first pass all the features into the HM-LN where different statistics are calculated and used as per each agent's modality type. Afterward, we feed the normalized features into a heterogeneous 3D graph attention \\((\\mathrm{H}^3\\mathrm{GATL / H}^3\\mathrm{GAT - G})\\) to jointly reason heterogeneous inter-agent and intra-agent interactions. Then, we pass the fused features to another HM-LN followed by a Hetero-modal MLP layer where different sets of parameters are used for camera and LiDAR features. By carefully designing these com" + }, + { + "type": "code_caption", + "bbox": [ + 0.503, + 0.099, + 0.877, + 0.115 + ], + "angle": 0, + "content": "Algorithm 1 Multi-agent hetero-modal fusion process" + }, + { + "type": "algorithm", + "bbox": [ + 0.507, + 0.118, + 0.892, + 0.401 + ], + "angle": 0, + "content": "1: Input: decompressed feature \\(\\mathbf{F}_i\\), pose \\(x_i\\) for each agent \n2: \\(\\mathbf{F}_i^{(0)} = \\mathbf{F}_i\\) \n3: for \\(l = 1, \\ldots, L\\) do \n4: for each agent \\(i\\) do \\(\\triangleright\\) Process in parallel \n5: \\(\\mathbf{F}_{j \\rightarrow i}^{(l-1)} = \\Gamma_{j \\rightarrow i}\\left(\\mathbf{F}_j^{(l-1)}\\right) \\triangleright\\) Spatially transform neighboring agents' features \n6: \\(\\mathbf{F}_i^{(l)} = \\mathrm{H}^3\\mathrm{GAT-L}\\) Block(\\(\\{\\mathbf{F}_{j \\rightarrow i}^{(l-1)}\\}\\)) \\(\\triangleright\\) Update node via local attention \n7: end for \n8: for each agent \\(i\\) do \\(\\triangleright\\) Process in parallel \n9: \\(\\mathbf{F}_{j \\rightarrow i}^{(l)} = \\Gamma_{j \\rightarrow i}\\left(\\mathbf{F}_j^{(l)}\\right) \\triangleright\\) Spatially transform neighboring agents' features \n10: \\(\\mathbf{F}_i^{(l)} = \\mathrm{H}^3\\mathrm{GAT-G}\\) Block(\\(\\{\\mathbf{F}_{j \\rightarrow i}^{(l)}\\}\\)) \\(\\triangleright\\) Update node via global attention \n11: end for \n12: end for \n13: \\(\\mathbf{F}_i = \\mathrm{HM - MLP}\\left(\\mathbf{F}_i^{(L)}\\right)\\) \\(\\triangleright\\) Output updated features" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.416, + 0.892, + 0.462 + ], + "angle": 0, + "content": "ponents, we can maintain modality-specific characteristics throughout the fusion process while benefiting from cross-modality multi-agent interactions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.799, + 0.485 + ], + "angle": 0, + "content": "3.3. Hetero-modal Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Compression and sharing: To reduce the transmission bandwidth, a series of \\(1 \\times 1\\) convolutions is applied to reduce the transmitted feature size along the channel dimension. Together with the intermediate features, each agent's pose \\(x_{i}\\) is also circulated within the collaboration graph. The ego agent will receive these features and decompress them back to the original size via another convolutional network. For processing intermediate features of the camera agent and LiDAR agent, we leverage distinct parameters in the compression and decompression modules to preserve the modality-specific characteristics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Graph-structured feature fusion: The received BEV features are centered around different spatial locations as each agent perceives the dynamic environment from different view points. To this end, we present a graph-structured fusion process (Fig. 4a): each node maintains a state representation of an agent in its own coordinate frame, and for a fixed number of iterations, spatially warped messages are shared between nodes and the node states are updated based on the aggregated features via the transformer blocks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The overall process is summarized in Alg. 1. During each iteration, we have two cascaded node updates which capture the local and global heterogeneous interactions respectively. For each node, we first spatially transform [18] neighboring nodes' features to its center \\(\\mathbf{F}_{j\\rightarrow i} = \\Gamma_{j\\rightarrow i}(\\mathbf{F}_j)\\). When the transmitting node is the receiving node itself, the transformation matrix is an identity matrix" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "288" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.164, + 0.088, + 0.807, + 0.257 + ], + "angle": 0, + "content": "
ModelsV2V-CV2V-LV2V-H
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
No Fusion0.0940.0210.5240.3630.2840.157
Late Fusion0.2310.0700.7700.6060.5020.308
V2VNet [40]0.3290.1250.8200.6450.6500.366
DiscoNet [22]0.2870.1150.7410.5900.6240.385
AttFuse [49]0.2610.0950.8010.6440.6470.390
CoBEVT [46]0.3170.1220.8280.6370.6740.416
V2X-ViT [48]0.3320.1250.8330.6790.6710.427
HM-ViT0.3550.1420.8530.7630.6950.515
" + }, + { + "type": "table_caption", + "bbox": [ + 0.26, + 0.267, + 0.707, + 0.283 + ], + "angle": 0, + "content": "Table 1: Evaluation of V2V perception methods on OPV2V dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.471, + 0.522 + ], + "angle": 0, + "content": "and thus \\(\\mathbf{F}_{i\\rightarrow i} = \\mathbf{F}_i\\). These spatial aligned feature maps \\(\\mathbf{F}_{j\\rightarrow i}\\) are then shared with agent \\(i\\) to update its state representation via the aggregation module. We adopt \\(\\mathrm{H}^3\\mathrm{GAT - L}\\) Block as our first aggregation module to capture the local heterogeneous interactions and leverage \\(\\mathrm{H}^3\\mathrm{GAT - G}\\) Block for the second module to further refine the states with global cues. Within each transformer block, we also adopt a mask to mask out non-overlapping areas between the field of views when computing the attention scores. Note that each agent's state update can be processed in parallel for better efficiency. After L such iterations, we pass the features to a hetero-modal MLP to further refine the feature representation. Throughout the whole fusion process, the modality-specific statistics are maintained." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.53, + 0.266, + 0.545 + ], + "angle": 0, + "content": "3.4. Hetero-modal Head" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.706 + ], + "angle": 0, + "content": "As camera and LiDAR contain distinct characteristics, we design a hetero-modal head where a different set of parameters are applied for camera and LiDAR ego vehicles to generate the final predictions. More specifically, the final fused feature maps are passed to a series of \\(3 \\times 3\\) convolutions with batch normalization and ReLU for feature refinement. Then, we adopt a \\(1 \\times 1\\) convolution layer to generate the regression and classification predictions. Smooth \\(\\ell_{1}\\) loss is utilized for regression and a focal loss [25] is used for classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.21, + 0.734 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.3, + 0.758 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "OPV2V. OP2V [49] is a large-scale multi-modal cooperative V2V perception dataset collected in CARLA [12] and OpenCDA [45]. It contains over 70 driving scenarios of around 25 seconds duration each. Each scenario contains multiple connected AVs (2 to 7) and each AV is equipped with 1 LiDAR and 4 monocular cameras covering \\(360^{\\circ}\\) horizontal field of view (FoV). In our hetero-modal cooperative perception setting, we only use one type of sensor modality for each AV, leading to two types of agents: vehicles only" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.89, + 0.34 + ], + "angle": 0, + "content": "equipped with multiple cameras (camera agent), and vehicles only equipped with LiDARs (LiDAR agent)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.582 + ], + "angle": 0, + "content": "**Evaluations.** We adopt Average precision (AP) at Intersection-over-Union (IoU) 0.5 and 0.7 to measure the perception performance. As each scenario consists of multiple AVs, a fixed agent is selected as the ego agent and the evaluation is conducted in the range of \\(100\\mathrm{m} \\times 100\\mathrm{m}\\) around it. Following [49, 48], the train/validation/test splits are 6764/1981/2719. We evaluate models mainly under three configurations: 1) V2V Camera-based 3D detection (V2V-C) where AVs are only equipped with 4 cameras with 360 horizontal FoV, 2) V2V LiDAR-based detection (V2V-L) where all the agents only have LiDAR sensors, and 3) V2V Hetero-modal detection (V2V-H) where half of the agents only has cameras while the other half only has LiDARs. To further assess models' capability with dynamic sensor configurations, we also assess models with fixed ego sensor modality and varying collaborator sensor types." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.696, + 0.607 + ], + "angle": 0, + "content": "4.2. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We use the PointPillar [20] as 3D backbones and a modified BEVFormer [23] for our camera stem. For BEVFormer, we adopt its variant with no temporal information and ResNet50 [15] as image backbones for better computation efficiency and use a smaller grid resolution \\((0.4\\mathrm{m})\\) to preserve fine-trained spatial details. The intermediate BEV feature map has a dimension of \\(128\\times 128\\times 256\\). Following prior works [22, 49, 48, 46], we only change the fusion module for different intermediate fusion methods while keeping the other components such as hetero-modal header, compression, and lidar/camera feature extractors the same. For the heterogeneous methods, we provide the network with additional feature types so that they can distinguish whether the feature is from Li-DAR or camera while for the homogeneous methods, the fusion networks would treat the features from LiDAR and camera equally. For HM-ViT, we conduct two iterations of graph-structured feature fusion and employ a window size of 8 for both local and global attentions. We adopt AdamW" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.106, + 0.283, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.106, + 0.484, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.106, + 0.683, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.106, + 0.885, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.231, + 0.755, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Figure 6: Agent modality ratio experiment. The x-axis is the ratio of LiDAR collaborators among all the collaborators. In (a) and (b), ego vehicles are equipped with cameras. In (c) and (d), ego vehicles are equipped with LiDARs." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.303, + 0.272, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.303, + 0.468, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.472, + 0.512 + ], + "angle": 0, + "content": "Figure 7: Ablation study on number of agent. The LiDAR/camera collaborator refers to the case where the collaborators are equipped with LiDAR and Camera respectively. (a) ego vehicle is equipped with camera. (b) ego vehicle is equipped with LiDAR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.584 + ], + "angle": 0, + "content": "with a decay rate of \\(10^{-2}\\) and cosine annealing learning rate scheduler to optimize the models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.471, + 0.766 + ], + "angle": 0, + "content": "Training strategy. We find that the intermediate fusion models won't converge if directly trained end-to-end under V2V-H and the resulting models usually only exhibit good performance for either camera perception or LiDAR perception but hardly for both. Instead, we first train the model on single modality configurations (i.e., V2V-C and V2V-L) until convergence for 40 epochs and then fine-tune the models under V2V-H for 10 epochs with fixed parameters of modality-specific backbones on 4 RTX3090 GPUs. This training strategy can help models converge well with reliable performance. For a fair comparison, we leverage this training strategy for all the methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Compared methods. We regard No Fusion as the baseline method. We also evaluate the Late Fusion, which transmits the detection proposals and leverages Non-maximum suppression to generate the final predictions. For the intermediate cooperation methods, we benchmark five approaches: V2VNet [40], DiscoNet [22], AttFuse [49], CoBEVT [46], and V2X-ViT [48]. For a fair comparison, hetero-modal head is used for all the models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.305, + 0.719, + 0.323 + ], + "angle": 0, + "content": "4.3. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.895, + 0.472 + ], + "angle": 0, + "content": "Main performance comparison. Tab. 1 demonstrates the performance comparisons on V2V-C, V2V-L and V2V-H. Under all three settings, all the cooperative methods outperform the No Fusion baseline and the intermediate fusion beats the classical Late Fusion, showing the great benefit of end-to-end V2V hetero-modal cooperative perception. The HM-ViT outperforms all the other SOTA intermediate fusion methods by at least \\(1.7\\%\\), \\(8.4\\%\\), \\(8.8\\%\\) in AP@0.7 under V2V-C, V2V-L, V2V-H settings respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Agent modality ratio experiment. As shown in Fig. 6, we fix the ego vehicle sensor modality and vary the ratio of collaborators' sensor modalities to evaluate the models' performance under various heterogeneous traffic scenarios. The larger LiDAR collaborator ratio corresponds to more vehicles only equipped with LiDARs while a smaller ratio means more vehicles only equipped with multiple cameras. The left two figures are the evaluation results for camera ego vehicles while the right two figures are the performance for LiDAR ego vehicles. Under most ratios, late fusion outperforms No Fusion however for LiDAR ego vehicle when most collaborators are camera agents, the Late Fusion performs poorer than No Fusion. We argue this is due to the fact that the camera predictions are usually noisy and merging proposals from different modalities equally could lead to ambiguity and thus deteriorate the performance. On the other hand, all the intermediate fusion methods outperform No Fusion by a large margin especially for the camera ego vehicles, demonstrating the great value of V2V cooperation between agents with different modalities. Among all the compared methods, HM-ViT ranks first for both camera ego vehicles and LiDAR ego vehicles under all the ratios, illustrating the great capability of HM-ViT for capturing modality-specific characteristics and cross-modality multiagent interactions. In contrast, other intermediate fusion methods only show good performance for a certain ratio range, which demonstrates the importance of heterogeneity for hetero-modal cooperative perception." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "290" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.14, + 0.361, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.239, + 0.306, + 0.252 + ], + "angle": 0, + "content": "(a) No Fusion (camera)" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.14, + 0.612, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.239, + 0.556, + 0.252 + ], + "angle": 0, + "content": "(b) Pure camera fusion" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.14, + 0.859, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.239, + 0.86, + 0.252 + ], + "angle": 0, + "content": "(c) Camera ego with LiDAR collaborators" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.259, + 0.362, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.357, + 0.309, + 0.37 + ], + "angle": 0, + "content": "(d) No Fusion (LiDAR)" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.259, + 0.609, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.357, + 0.554, + 0.37 + ], + "angle": 0, + "content": "(e) Pure LiDAR fusion" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.259, + 0.857, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.357, + 0.854, + 0.37 + ], + "angle": 0, + "content": "(f) LiDAR ego with camera collaborators" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Figure 8: Qualitative visualizations for (a) No Fusion with camera ego vehicle, (b) pure camera-based V2V perception, (c) hetero-modal V2V perception with camera ego vehicle and LiDAR collaborators, (d) No Fusion with LiDAR ego vehicle, (e) pure LiDAR-based V2V perception, and (f) hetero-modal V2V perception with LiDAR ego vehicle and camera collaborators. The red and green boxes represent the detection outputs and ground truth respectively. More visualizations can be found in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.471, + 0.466, + 0.558 + ], + "angle": 0, + "content": "
HM-MLP&LNH3GAT-LH3GAT-GAP@0.7
0.404
✓0.420
✓✓0.460
✓✓✓0.515
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.568, + 0.465, + 0.584 + ], + "angle": 0, + "content": "Table 2: Component Ablation study on the V2V-H setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Number of agent. In this experiment, we investigate the effect of the number of agents on the perception performance of HM-ViT. As Fig. 7 depicts, for both camera and LiDAR ego vehicles, the perception performance increases as more agents are involved in the cooperative perception and both LiDAR and camera collaborators can contribute to the performance gain for ego vehicles with different modalities, which again shows the benefit of hetero-modal V2V cooperation. Additionally, the increase rate generally decreases when increasing the number of agents and the LiDAR collaborators can bring more AP gains over the camera collaborators for both camera and LiDAR ego vehicles. Notably, similar as Fig. 6a-b display, the camera ego vehicles' performance can be greatly improved with only a small number of LiDAR collaborators, demonstrating the great potential of reducing the cost for each vehicle when the V2V system is deployed at scale as we may only need to install expansive LiDARs for a small number of agents (e.g., infrastructure) while all the other agents only require relatively cheap camera sensors." + }, + { + "type": "table", + "bbox": [ + 0.583, + 0.471, + 0.812, + 0.558 + ], + "angle": 0, + "content": "
Compression RateAP@0.7
0x0.515
8x0.513
16x0.470
32x0.455
" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.568, + 0.887, + 0.584 + ], + "angle": 0, + "content": "Table 3: Compression rate effects for HM-ViT on V2V-H." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Component ablation study. Here we investigate the key components of the proposed HM-ViT. As the layer normalization and MLP are usually combined together in typical transformer designs, thus we jointly evaluate the combined effect of HM-MLP and HM-LN (HM-MLP&LN). As Tab. 2 shows, all the proposed components improve the performance and local and global attentions can largely improve the AP@0.7 by \\(4\\%\\) and \\(5.5\\%\\), proving the great benefit brought by jointly reasoning inter-agent and intra-agent heterogeneous interactions both locally and globally." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.891, + 0.793 + ], + "angle": 0, + "content": "Compression rate. Tab. 3 describes the influence of compression rate. It demonstrates that HM-ViT can still outperform other methods even under large compression rates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.678, + 0.818 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 8 depicts the qualitative visualizations for HM-ViT and No Fusion baselines. In fig. 8a-c, we plot the detection results for camera ego vehicles with no collaborator, camera collaborators, and LiDAR collaborators respectively while for Fig. 8d-f, we plot the results for LiDAR ego vehicles. As" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "291" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "shown in these figures, the collaborations with both homogeneous and heterogeneous agents are beneficial for camera ego vehicles and LiDAR ego vehicles with enhanced detection results. In particular, the collaboration between camera ego vehicles and LiDAR collaborators can dramatically enhance the perception performance." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.195, + 0.194, + 0.212 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.221, + 0.468, + 0.402 + ], + "angle": 0, + "content": "In this paper, we present HM-ViT, a hetero-modal vision transformer, for the hetero-modal multi-agent cooperative perception problem which is an important but under-explored research direction. We propose a generic heterogeneous 3D graph attention to jointly reason heterogeneous inter-agent and cross-agent interactions. Our extensive experiments demonstrate the outstanding performance of the proposed method and the great potential of hetero-modal multi-agent collaborations for increasing the scalability and robustness of V2V systems. We hope our findings and open-source efforts will inspire more research on this new problem." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.947, + 0.511, + 0.957 + ], + "angle": 0, + "content": "292" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Luís A Alexandre. 3d object recognition using convolutional neural networks with transfer learning between input channels. In Intelligent Autonomous Systems 13, pages 889-898. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[3] Johannes Betz, Hongrui Zheng, Alexander Liniger, Ugo Rosolia, Phillip Karle, Madhur Behl, Venkat Krovi, and Rahul Mangharam. Autonomous vehicles on the edge: A survey on autonomous vehicle racing. IEEE Open Journal of Intelligent Transportation Systems, 3:458-488, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.472, + 0.556 + ], + "angle": 0, + "content": "[6] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[7] Hung-Hsun Chen, Yi-Bing Lin, I-Hau Yeh, Hsun-Jung Cho, and Yi-Jung Wu. Prediction of queue dissipation time for mixed traffic flows with deep learning. IEEE Open Journal of Intelligent Transportation Systems, 3:267-277, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[8] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.687, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[10] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Graph-detr3d: Rethinking overlapping regions for multi-view 3d object detection. arXiv preprint arXiv:2204.11582, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[11] Zhuo Deng and Longin Jan Latecki. A modal detection of 3d objects: Inferring 3d bounding boxes from 2d ones in rgb-depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5762-5770, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driv" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "ing simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[13] Yi Guo and Jiaqi Ma. Leveraging existing high-occupancy vehicle lanes for mixed-autonomy traffic management with emerging connected automated vehicle applications. Transportmetrica A: Transport Science, 16(3):1375–1399, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[14] Yi Guo, Jiaqi Ma, Edward Leslie, and Zhitong Huang. Evaluating the effectiveness of integrated connected automated vehicle applications applied to freeway managed lanes. IEEE Transactions on Intelligent Transportation Systems, 23(1):522-536, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[16] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. arXiv preprint arXiv:2209.12836, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[19] Sou Kitajima, Hanna Chouchane, Jacobo Antona-Makoshi, Nobuyuki Uchida, and Jun Tajima. A nationwide impact assessment of automated driving systems on traffic safety using multiagent traffic simulations. IEEE Open Journal of Intelligent Transportation Systems, 3:302-312, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[21] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In European Conference on Computer Vision, pages 316-332. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[22] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.217 + ], + "angle": 0, + "content": "[27] Jiaqi Ma, Edward Leslie, Amir Ghiasi, Zhitong Huang, and Yi Guo. Empirical analysis of a freeway bundled connected-and-automated vehicle application using experimental data. Journal of Transportation Engineering, Part A: Systems, 146(6):04020034, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[28] Vasileia Papathanasopoulou, Ioanna Spyropoulou, Harris Perakis, Vassilis Gikas, and Eleni Andrikopoulou. A data-driven model for pedestrian behavior classification and trajectory prediction. IEEE Open Journal of Intelligent Transportation Systems, 3:328-339, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[29] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[30] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgbd data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[31] Kelli Raboy, Jiaqi Ma, Edward Leslie, and Fang Zhou. A proof-of-concept field experiment on cooperative lane change maneuvers using a prototype connected automated vehicle testing platform. Journal of Intelligent Transportation Systems, 25(1):77-92, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.511 + ], + "angle": 0, + "content": "[32] Steven E Shladover. Opportunities and challenges in cooperative road vehicle automation. IEEE Open Journal of Intelligent Transportation Systems, 2:216-224, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[33] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[36] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. arXiv preprint arXiv:2204.01697, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[37] Rodolfo Valiente, Behrad Toghi, Ramtin Pedarsani, and Yaser P Fallah. Robustness and adaptability of reinforcement learning-based cooperative autonomous driving in mixed-autonomy traffic. IEEE Open Journal of Intelligent Transportation Systems, 3:397-410, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[38] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[39] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[40] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[41] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[42] Zhixin Wang and Kui Jia. Frustum convnet: Sliding frustums to aggregate local point-wise features for amodal 3d object detection. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1742-1749. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[43] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. arXiv preprint arXiv:2209.13679, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[44] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[45] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.651 + ], + "angle": 0, + "content": "[46] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[47] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. arXiv preprint arXiv:2303.07601, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[48] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. arXiv preprint arXiv:2203.10638, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[49] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[51] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[52] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[54] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[55] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "295" + } + ] +] \ No newline at end of file diff --git a/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_origin.pdf b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..39c81cf68ee37a4275ad3be6b0a1fddb4c2153f0 --- /dev/null +++ b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/9d343628-a95c-47d2-8ca4-db20691fefde_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a84da8f9c02e6b501d4b9a2e7357dcb68260a670a6ea8b128a288801036a774b +size 2827923 diff --git a/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/full.md b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0b0daf1b27b71093e99dbc1b9c262d096bb55f75 --- /dev/null +++ b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/full.md @@ -0,0 +1,293 @@ +# HM-ViT: Hetero-modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer + +Hao Xiang $^{1}$ , Runsheng Xu $^{1}$ , Jiaqi Ma $^{1}$ $^{1}$ University of California, Los Angeles +{haxiang, rxx3386, jiaqima}@g.ucla.edu + +# Abstract + +Vehicle-to-Vehicle technologies have enabled autonomous vehicles to share information to see through occlusions, greatly enhancing perception performance. Nevertheless, existing works all focused on homogeneous traffic where vehicles are equipped with the same type of sensors, which significantly hampers the scale of collaboration and benefit of cross-modality interactions. In this paper, we investigate the multi-agent hetero-modal cooperative perception problem where agents may have distinct sensor modalities. We present HM-ViT, the first unified multi-agent hetero-modal cooperative perception framework that can collaboratively predict 3D objects for highly dynamic Vehicle-to-Vehicle (V2V) collaborations with varying numbers and types of agents. To effectively fuse features from multi-view images and LiDAR point clouds, we design a novel heterogeneous 3D graph transformer to jointly reason inter-agent and intra-agent interactions. The extensive experiments on the V2V perception dataset OPV2V demonstrate that the HM-ViT outperforms SOTA cooperative perception methods for V2V hetero-modal cooperative perception. Our code will be released at https://github.com/XHwind/HM-ViT. + +# 1. Introduction + +Recent advances in Vehicle-to-Vehicle (V2V) communication technology and intelligent transportation systems [13, 31, 27, 14, 7, 19, 28, 51, 53, 52] have allowed autonomous vehicles (AVs) to share sensory information, enabling them to perceive their surroundings better [3, 32, 54]. With the rapid growth of autonomous driving, V2V perception systems have the potential to be deployed at scale and create a safer transportation system. Cooperative perception systems, as shown in recent studies [49, 48, 40], can intelligently aggregate features from multiple vehicles within the communication range to enhance visual reasoning and overall performance. + +![](images/c46fc370206b843114e773d0cb899b250530a805e8e177e26b489f1a0ecca645.jpg) +Figure 1: Illustration of multi-agent hetero-modal V2V systems where each agent may be equipped with either LiDAR or multi-view cameras. + +Despite the rapid growth in this field, previous studies [49, 40, 46, 22, 37, 47] have primarily focused on homogeneous multi-agent cooperative perception, where all agents are equipped with the same type of sensors. In reality, however, agents may have different sensor modalities (hetero-modality) due to the cost and sensor preferences of ADS developers and car makers. As shown in Fig. 1, some agents are equipped with only LiDARs (LiDAR agents), while others only have multiple cameras (camera agents). Enabling collaboration between these heterogeneous agents could improve the sensing capability by allowing agents to see through occlusions and increase the scale and reliability of V2V systems. Additionally, LiDAR agents can provide accurate geometric information, while camera agents can provide rich semantic context. Thus, the collaboration between these agents could leverage the distinct but complementary environment attributes captured by each sensor modality to enhance the V2V perception systems. Furthermore, compared with the single-agent solution where multiple LiDARs and cameras are installed in a single vehicle, distributing different types of sensors across distinct agents could also potentially decrease the costs for each agent while still achieving satisfying performance. Nevertheless, whether, when, and how multi-agent hetero-modal V2V cooperation can benefit the perception system of heterogeneous traffic has not yet been studied. + +![](images/decd16733e4305627bce77a76256c8c75d4ea2ca15a907eb5fc94a984626fd3a.jpg) +Figure 2: Comparison of the single-agent multi-modal system and multi-agent hetero-modal system. The graph structure of the former is fixed whereas for the latter the graph is both dynamic and heterogeneous. + +In this work, we address the multi-agent hetero-modal cooperative perception problem where each agent could have distinct sensor types and share/receive information with each other. Notably, as shown in Fig. 2, this multi-agent hetero-modal setting is distinct from the single-agent multi-modal setting. In the hetero-modal setting, the agent sensors form a dynamic heterogeneous graph where the existence and types of sensors are random, and the relative poses vary from scene to scene. In contrast, the sensor types/numbers and relative positions (extrinsics) between sensors are fixed in the single-agent multi-modal setting. Existing multi-modal methods heavily rely on these assumptions, and most of existing works [38, 39, 2, 55, 50] transform LiDAR points or 3D proposals onto the image plane to index 2D features. Their network architectures build upon the co-existence of both LiDAR and camera inputs with fixed geometric relationships. However, the dynamic nature of hetero-modal V2V perception requires a flexible architecture that can handle varying agent numbers and types, and the transmitted neural features are also spatially misaligned. Moreover, there are semantic discrepancies in the transmitted features between camera agents and LiDAR agents. Hence, these unique characteristics pose significant challenges for designing the multi-agent hetero-modal cooperative system and prevent adapting existing multi-modal fusion methods to this new problem. + +To enable collaboration between heterogeneous agents in V2V systems, we propose Hetero-Modal Vision Transformer (HM-ViT), the first unified cooperative perception framework that can leverage and fuse distributed information for hetero-modal V2V perception via a spatial-aware 3D heterogeneous vision transformer. Fig. 3 demonstrates the overall framework. Each agent first generates bird's eye view (BEV) representations through modality-specific encoders and then shares compressed features with neighboring agents. Afterward, the received features are decompressed and aggregated via the proposed HM-ViT, + +which conducts joint local and global heterogeneous 3D attentions with the consideration of both node and edge types. Our extensive experiments show that the HM-ViT can significantly improve the perception capability of camera agents and LiDAR agents over the single-agent baseline and outperforms SOTA cooperative perception methods by a large margin. In particular, for camera agents, performance can be boosted from $2.1\%$ to $53.2\%$ at $\mathrm{AP}@0.7$ with the collaboration of LiDAR agents, a 23-fold improvement. Our primary contributions can be summarized as follows: + +- We present the novel transformer framework (HMViT) for multi-agent hetero-modal cooperative perception, capable of capturing the modality-specific characteristics and heterogeneous 3D interactions. The proposed model exhibits superior flexibility and robustness with state-of-the-art performance on highly dynamic heterogeneous traffic involving varying agent numbers and types. +- We propose a generic heterogeneous 3D graph attention (H $^3$ GAT), tailored for extracting inter-agent and intra-agent heterogeneous interactions. We instantiate two such attentions - local attention (H $^3$ GAT-L) and global attention (H $^3$ GAT-G) for capturing both local and global visual cues. +- We conduct extensive benchmark experiments by varying sensor modalities, demonstrating the strong performance of the proposed method for hetero-modal V2V perception tasks. We will release all the codes and baselines to facilitate future research. + +# 2. Related work + +V2V perception. V2V perception aims to enhance the perception performance of autonomous vehicles by leveraging shared information from other connected vehicles. Existing works have primarily focused on LiDAR-based 3D object perception. The pioneer cooperative perception methods transmit raw sensing observation (i.e., early fusion) or perception outputs (i.e., late fusion), whereas recent works [49, 40, 16, 43, 21, 48] are exploring the use of circulating intermediate neural features for achieving better performance-bandwidth trade-off. V2VNet [40] employs graph neural networks to aggregate the shared neural features for joint detection and prediction. AttFuse [49] uses single-head attention to model the per-location multi-agent interaction. Disconet [22] presents a matrix-valued edge weight for learning the interactions and a teacher-student learning framework to facilitate the training. V2X-ViT explores [48] vision transformer for vehicle-to-everything cooperation via window attention and heterogeneous self-attention. CoBEVT [46] presents a generic transformer framework for camera-based BEV semantic segmentation. + +![](images/a3d6260f86db65efd341a4abfd55938058b3a9ac4c29bc5490c7814aa8f5e172.jpg) +Figure 3: Overview of proposed hetero-modal V2V cooperative perception system. Each agent first produces BEV features through modality-specific feature extractors (Sec. 3.1). The BEV features are then compressed and shared (Sec. 3.3) with neighboring connected agents and the received features are decompressed in the ego agent side and fed into hetero-modal vision transformer to conduct graph-structured feature fusion (Sec. 3.2 and Sec. 3.3). The refined features are finally passed into the hetero-modal detection head to predict 3D bounding boxes (Sec. 3.4). + +However, none of the existing works explored multi-agent multi-camera 3D object detection, let alone multi-agent hetero-modal perception. In contrast to existing methods, HM-ViT is the first to employ sparse heterogeneous local and global attentions to capture the 3D inter-agent and intraagent interactions in a computationally efficient manner. + +Camera-based 3D object detection. Early works [4, 8, 33] mainly focus on monocular 3D detection but a single camera can only provide a 2D view of the scene, and inferring 3D from 2D is intrinsically hard. Recent development of self-driving datasets [5, 34, 6] featured with full sensor suits has enabled the research direction of 3D object detection from multiple cameras. DETR3D [41] proposes a 3D-2D query paradigm for extracting 3D features from 2D multi-view images. Graph-DETR3D [10] further utilizes graph structure learning to enhance the representation at the border regions. Alternatively, another stream of works [17, 29, 44, 23] focuses on aggregating BEV features from multi-view cameras for conducting downstream perception tasks. LSS [29] lifts the 2D features to 3D frustum via latent depth and then splats frustums into a BEV grid. $\mathbf{M}^2\mathbf{BEV}$ [44] further extends LSS with less memory consumption and conducts detection and segmentation. BEV-Former [23] constructs BEV queries and explores spatial cross-attention and temporal self-attention to recurrently refine the BEV features, achieving SOTA performance on both NuScenes [5] and Waymo Open Datasets [34]. + +Multi-modal fusion. Existing multi-sensor fusion methods can be divided into point-level fusion, proposal-level fusion, and BEV-level fusion. Point-level fusion decorates the input from one modality with attributes from the other modality. PointPainting [38] and PaintAugmenting [39] decorate the LiDAR point clouds with semantic segmentation scores and 2D CNN image features respectively while [11, 1] project + +LiDAR onto the image plane to augment the RGB values with depth information and conduct detection. On the other hand, proposal-level fusion generates proposals from one modality and then indexes features from the other modality for further refinement. MV3D [9] produces object queries from LiDAR BEV and then extracts the features from camera data and LiDAR front view. [30, 42] lift 2D bounding boxes to frustum and conduct 3D object detection from the frustum of point clouds. Conversely, BEV-level fusion converts features from different modalities to unified BEV representations, preserving both geometric and semantic information. BEVFusions [26, 24] concatenate BEV features from camera and LiDAR and fuse it via a fusion module. Hence, existing single-agent multi-modal fusion methods rely on the co-existence of both camera and LiDAR with fixed geometric relationships, which is unsuitable for our multi-agent hetero-modal cooperative perception problem with a dynamic heterogeneous collaboration graph. + +# 3. Methodology + +In this paper, we explore the multi-agent hetero-modal cooperative perception, where each AV is equipped with either a LiDAR or multiple cameras. Our goal is to create a robust and flexible cooperative perception system that allows for efficient collaboration between any number of agents with varying sensor types, ultimately improving the perception capabilities of the vehicle in a unified end-to-end fashion. The pipeline, illustrated in 3, includes modality-specific feature extraction, compression and sharing, HMViT for feature fusion, and hetero-modal detection head. + +![](images/86acf29b7f78e2c9b05186a39dfc9f09228fe4b7a2609a7881752d70c23c02b2.jpg) +Figure 4: HM-ViT architecture. (a) Graph-structured fusion process. (b) Heterogeneous 3D graph attention $\mathrm{(H^3GAT)}$ . (c) Illustration of sampled tokens for $\mathrm{H}^{3}\mathrm{GAT - L}$ and $\mathrm{H}^{3}\mathrm{GAT - G}$ . + +![](images/4fc9fb0d581f82ada74494b79ed32e4215416de18b85987bb3e6edca3aa09cf8.jpg) + +![](images/57c749347d29ecd3cfbe379064a23adf13e63fd2f5bb76aae6f8ba961938ae38.jpg) + +# 3.1. Modality-specific feature extraction + +LiDAR stem: We leverage PointPillar [20] to process point clouds for each LiDAR agent. The raw point cloud is converted to a 2D pseudo-image, flattened along the height dimension, and fed into 2D convolutional neural networks to produce the salient feature map $\mathbf{F}_j \in \mathcal{R}^{H \times W \times C}$ , which is compressed and shared with all the neighboring agents. + +Camera stem: Each camera agent is equipped with $m$ monocular cameras. The sensing observation of $i$ -th agent includes the input images $I_k^i \in \mathcal{R}^{h \times w \times 3}$ and the known projection matrix $P_k^i \in \mathcal{R}^{3 \times 4}$ that maps 3D reference points to different image views. Our goal is to generate a BEV feature representation $\mathbf{F}_i \in \mathcal{R}^{H \times W \times C}$ that is amenable for feature fusion with other collaborators. In this work, we adopt similar architecture to BEVFormer [23] with no temporal information for feature extraction. For a faster running time, we adopt ResNet50 to extract 2D image features and then adopt a learnable 2D BEV query to inquire spatial information from the encoded multi-view features via spatial cross attention and projection matrices. The resulting refined BEV feature $\mathbf{F}_i$ is centered around agent $i$ and shared with connected AVs. + +# 3.2. Heterogeneous 3D Graph Attention (H $^3$ GAT) + +To account for the distinct characteristics of BEV features extracted from different sensor modalities, the learning process of each modality must be distinguished, and the cross-modality interactions between multiple agents should vary. To capture this heterogeneity, we present a novel heterogeneous 3D graph attention (H $^3$ GAT), in which nodes and edges are type-dependent to reason spatial interactions and cross-agent relations jointly. We encode both local and global interactions to better capture the 3D ambiguity in BEV feature space. Local attention can help preserve object details, while global attention can provide a better understanding of environmental contexts such as road topology + +and traffic density. + +As shown in Fig. 4b, We build a 3D heterogeneous collaboration graph. Each node $v(i,x) = \mathbf{F}_x^i \in \mathcal{R}^C$ is a feature vector of agent $i$ 's feature map at spatial location $x \in \mathcal{R}^2$ . 3D heterogeneous graph attention is performed for spatially connected nodes in the BEV feature space. Depending on the definition of spatial connectivity, we will derive local attention and global attention. Here for notation simplicity, we only derive single-head equations but in real implementations, multi-head variants are used. Formally, we first project feature vectors onto different feature spaces to form query, key, and value vectors: + +$$ +\mathbf {Q} _ {x} ^ {j} = \operatorname {D e n s e} _ {\tau_ {j}} \mathbf {F} _ {x} ^ {j} \tag {1} +$$ + +$$ +\mathbf {K} _ {x} ^ {j} = \operatorname {D e n s e} _ {\tau_ {j}} \mathbf {F} _ {x} ^ {j} \tag {2} +$$ + +$$ +\mathbf {V} _ {x} ^ {j} = \operatorname {D e n s e} _ {e _ {i j}} \mathbf {F} _ {x} ^ {j} \tag {3} +$$ + +where the $\mathrm{Dense}\square$ is a set of linear layers indexed by subscript $\square$ . For the query and key vectors, we use linear projectors $\mathrm{Dense}_{\tau_j}$ indexed by node type $\tau_{j}$ to extract modality-specific features. For the value vector, we index the projector via edge type $\mathrm{Dense}_{e_{ij}}$ to reflect the heterogeneity of cross-modality multi-agent interactions. The set of connected nodes of $v(i,x)$ is denoted as $\mathcal{N}(i,x)$ . Then the attention is operated as follows: + +$$ +\mathbf {a} (j, y) = \operatorname {S o f t m a x} _ {(j, y) \in \mathcal {N} (i, x)} \left(\mathbf {Q} _ {x} ^ {i} \mathbf {W} _ {e _ {i j}} \mathbf {K} _ {y} ^ {j}\right) \tag {4} +$$ + +$$ +\mathbf {F} _ {x} ^ {i} = \sum_ {(j, y) \in \mathcal {N} (i, x)} \mathbf {a} (j, y) \mathbf {V} _ {y} ^ {j} \tag {5} +$$ + +where $\mathbf{W}_{e_{ij}}\in \mathcal{R}^{C\times C}$ is used to adjust the dot product of $\mathbf{Q}_x^i$ and $\mathbf{K}_y^j$ to further encode the heterogeneity of edges. + +Depending on how the nodes are sampled (Fig. 4c), we design two types of attentions: local attention $(\mathbf{H}^3\mathbf{GAT - L})$ which performs local window-based attention and global attention $(\mathbf{H}^3\mathbf{GAT - G})$ which performs sparse global grid-based attention. Fig. 4c visualizes how the local and global + +![](images/7ada8e21e4acbb6353e610e741956fd6ab9acb59ff6cc96901e0633db245bad6.jpg) +(a) $\mathsf{H}^3\mathsf{GAT - L}$ Block +Figure 5: Transformer blocks for local and global attentions. + +![](images/f7340ba696d721cbe52da55b71876f0b0da4da42e26ba2b48e178ea989076393.jpg) +(b) $\mathsf{H}^3\mathsf{GAT - G}$ Block + +tokens are sampled, where the local tokens are sampled within local windows and the global tokens are sparsely sampled grids scattered across feature maps. The local interactions can help preserve spatial cues and provide reliable estimates while the global reasoning can help understand global semantic context. + +Both $\mathrm{H}^3\mathrm{GAT - L}$ and $\mathrm{H}^3\mathrm{GAT - G}$ can be implemented efficiently by decomposing the spatial axes. More specifically, we stack all the agents' features to $\mathbf{F}\in \mathcal{R}^{N\times H\times W\times C}$ where $N$ is the number of agents. For $\mathrm{H}^3\mathrm{GAT - L}$ , we decompose the feature map into 3D non-overlapping windows along the first axis [35, 36], each of size $N\times P\times P$ . The partitioned tensor has the shape $(\frac{H}{P}\times \frac{W}{P},N\times P^2,C)$ where the heterogeneous 3D local graph attention is conducted for $NP^2$ tokens within the same window. Similarly, for $\mathrm{H}^3\mathrm{GAT - G}$ , we swap the axis and partition the tensor into the shape $(N\times P^2,\frac{H}{P}\times \frac{W}{P},C)$ . Due to the swap operation, the sampled grids will be sparsely scattered and the attention is operated for these sparsely sampled $\frac{H}{P}\times \frac{W}{P}$ grids, which can capture sparse global information. + +To integrate this local and global attention into transformer architecture, we further present a heterogeneous normalization layer (HM-LN) and heterogeneous MLP (HMMLP) which use type-dependent parameters. As shown in Fig. 5, we first pass all the features into the HM-LN where different statistics are calculated and used as per each agent's modality type. Afterward, we feed the normalized features into a heterogeneous 3D graph attention $(\mathrm{H}^3\mathrm{GATL / H}^3\mathrm{GAT - G})$ to jointly reason heterogeneous inter-agent and intra-agent interactions. Then, we pass the fused features to another HM-LN followed by a Hetero-modal MLP layer where different sets of parameters are used for camera and LiDAR features. By carefully designing these com + +Algorithm 1 Multi-agent hetero-modal fusion process +1: Input: decompressed feature $\mathbf{F}_i$ , pose $x_i$ for each agent +2: $\mathbf{F}_i^{(0)} = \mathbf{F}_i$ +3: for $l = 1, \ldots, L$ do +4: for each agent $i$ do $\triangleright$ Process in parallel +5: $\mathbf{F}_{j \rightarrow i}^{(l-1)} = \Gamma_{j \rightarrow i}\left(\mathbf{F}_j^{(l-1)}\right) \triangleright$ Spatially transform neighboring agents' features +6: $\mathbf{F}_i^{(l)} = \mathrm{H}^3\mathrm{GAT-L}$ Block( $\{\mathbf{F}_{j \rightarrow i}^{(l-1)}\}$ ) $\triangleright$ Update node via local attention +7: end for +8: for each agent $i$ do $\triangleright$ Process in parallel +9: $\mathbf{F}_{j \rightarrow i}^{(l)} = \Gamma_{j \rightarrow i}\left(\mathbf{F}_j^{(l)}\right) \triangleright$ Spatially transform neighboring agents' features +10: $\mathbf{F}_i^{(l)} = \mathrm{H}^3\mathrm{GAT-G}$ Block( $\{\mathbf{F}_{j \rightarrow i}^{(l)}\}$ ) $\triangleright$ Update node via global attention +11: end for +12: end for +13: $\mathbf{F}_i = \mathrm{HM - MLP}\left(\mathbf{F}_i^{(L)}\right)$ $\triangleright$ Output updated features + +ponents, we can maintain modality-specific characteristics throughout the fusion process while benefiting from cross-modality multi-agent interactions. + +# 3.3. Hetero-modal Vision Transformer + +Compression and sharing: To reduce the transmission bandwidth, a series of $1 \times 1$ convolutions is applied to reduce the transmitted feature size along the channel dimension. Together with the intermediate features, each agent's pose $x_{i}$ is also circulated within the collaboration graph. The ego agent will receive these features and decompress them back to the original size via another convolutional network. For processing intermediate features of the camera agent and LiDAR agent, we leverage distinct parameters in the compression and decompression modules to preserve the modality-specific characteristics. + +Graph-structured feature fusion: The received BEV features are centered around different spatial locations as each agent perceives the dynamic environment from different view points. To this end, we present a graph-structured fusion process (Fig. 4a): each node maintains a state representation of an agent in its own coordinate frame, and for a fixed number of iterations, spatially warped messages are shared between nodes and the node states are updated based on the aggregated features via the transformer blocks. + +The overall process is summarized in Alg. 1. During each iteration, we have two cascaded node updates which capture the local and global heterogeneous interactions respectively. For each node, we first spatially transform [18] neighboring nodes' features to its center $\mathbf{F}_{j\rightarrow i} = \Gamma_{j\rightarrow i}(\mathbf{F}_j)$ . When the transmitting node is the receiving node itself, the transformation matrix is an identity matrix + +
ModelsV2V-CV2V-LV2V-H
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
No Fusion0.0940.0210.5240.3630.2840.157
Late Fusion0.2310.0700.7700.6060.5020.308
V2VNet [40]0.3290.1250.8200.6450.6500.366
DiscoNet [22]0.2870.1150.7410.5900.6240.385
AttFuse [49]0.2610.0950.8010.6440.6470.390
CoBEVT [46]0.3170.1220.8280.6370.6740.416
V2X-ViT [48]0.3320.1250.8330.6790.6710.427
HM-ViT0.3550.1420.8530.7630.6950.515
+ +Table 1: Evaluation of V2V perception methods on OPV2V dataset. + +and thus $\mathbf{F}_{i\rightarrow i} = \mathbf{F}_i$ . These spatial aligned feature maps $\mathbf{F}_{j\rightarrow i}$ are then shared with agent $i$ to update its state representation via the aggregation module. We adopt $\mathrm{H}^3\mathrm{GAT - L}$ Block as our first aggregation module to capture the local heterogeneous interactions and leverage $\mathrm{H}^3\mathrm{GAT - G}$ Block for the second module to further refine the states with global cues. Within each transformer block, we also adopt a mask to mask out non-overlapping areas between the field of views when computing the attention scores. Note that each agent's state update can be processed in parallel for better efficiency. After L such iterations, we pass the features to a hetero-modal MLP to further refine the feature representation. Throughout the whole fusion process, the modality-specific statistics are maintained. + +# 3.4. Hetero-modal Head + +As camera and LiDAR contain distinct characteristics, we design a hetero-modal head where a different set of parameters are applied for camera and LiDAR ego vehicles to generate the final predictions. More specifically, the final fused feature maps are passed to a series of $3 \times 3$ convolutions with batch normalization and ReLU for feature refinement. Then, we adopt a $1 \times 1$ convolution layer to generate the regression and classification predictions. Smooth $\ell_{1}$ loss is utilized for regression and a focal loss [25] is used for classification. + +# 4. Experiments + +# 4.1. Datasets and Evaluation + +OPV2V. OP2V [49] is a large-scale multi-modal cooperative V2V perception dataset collected in CARLA [12] and OpenCDA [45]. It contains over 70 driving scenarios of around 25 seconds duration each. Each scenario contains multiple connected AVs (2 to 7) and each AV is equipped with 1 LiDAR and 4 monocular cameras covering $360^{\circ}$ horizontal field of view (FoV). In our hetero-modal cooperative perception setting, we only use one type of sensor modality for each AV, leading to two types of agents: vehicles only + +equipped with multiple cameras (camera agent), and vehicles only equipped with LiDARs (LiDAR agent). + +**Evaluations.** We adopt Average precision (AP) at Intersection-over-Union (IoU) 0.5 and 0.7 to measure the perception performance. As each scenario consists of multiple AVs, a fixed agent is selected as the ego agent and the evaluation is conducted in the range of $100\mathrm{m} \times 100\mathrm{m}$ around it. Following [49, 48], the train/validation/test splits are 6764/1981/2719. We evaluate models mainly under three configurations: 1) V2V Camera-based 3D detection (V2V-C) where AVs are only equipped with 4 cameras with 360 horizontal FoV, 2) V2V LiDAR-based detection (V2V-L) where all the agents only have LiDAR sensors, and 3) V2V Hetero-modal detection (V2V-H) where half of the agents only has cameras while the other half only has LiDARs. To further assess models' capability with dynamic sensor configurations, we also assess models with fixed ego sensor modality and varying collaborator sensor types. + +# 4.2. Experimental Setups + +Implementation details. We use the PointPillar [20] as 3D backbones and a modified BEVFormer [23] for our camera stem. For BEVFormer, we adopt its variant with no temporal information and ResNet50 [15] as image backbones for better computation efficiency and use a smaller grid resolution $(0.4\mathrm{m})$ to preserve fine-trained spatial details. The intermediate BEV feature map has a dimension of $128\times 128\times 256$ . Following prior works [22, 49, 48, 46], we only change the fusion module for different intermediate fusion methods while keeping the other components such as hetero-modal header, compression, and lidar/camera feature extractors the same. For the heterogeneous methods, we provide the network with additional feature types so that they can distinguish whether the feature is from Li-DAR or camera while for the homogeneous methods, the fusion networks would treat the features from LiDAR and camera equally. For HM-ViT, we conduct two iterations of graph-structured feature fusion and employ a window size of 8 for both local and global attentions. We adopt AdamW + +![](images/072da0bf6f1f36e6688fa35fb69d0193c768a2ececc5be919b4148f17cca1981.jpg) +Figure 6: Agent modality ratio experiment. The x-axis is the ratio of LiDAR collaborators among all the collaborators. In (a) and (b), ego vehicles are equipped with cameras. In (c) and (d), ego vehicles are equipped with LiDARs. + +![](images/f11ffca66134f4c86478c9ef2a53f2dc66a9da005959bf656fb4060b18fbea9f.jpg) + +![](images/b60d85e61a24b6d3577254dc92d172a28c088b24c62b6da0d60bc51a95fda620.jpg) + +![](images/7087ab578d79b0dabf77d590c1ef11e3b339237b253faf099c27b405a32a7d61.jpg) + +![](images/051c2e88d1efd0531c58309bad1027680e09f5d57b9544bae1225dd866f3ed9f.jpg) + +![](images/061acd15789cec06515163b772ff539ce1b898e70b4d7b4d80789f635db8730e.jpg) +Figure 7: Ablation study on number of agent. The LiDAR/camera collaborator refers to the case where the collaborators are equipped with LiDAR and Camera respectively. (a) ego vehicle is equipped with camera. (b) ego vehicle is equipped with LiDAR. + +![](images/614a9178da5e734af03673af75ea16f4e44f1620135efcdb9a5e56cef8d70164.jpg) + +with a decay rate of $10^{-2}$ and cosine annealing learning rate scheduler to optimize the models. + +Training strategy. We find that the intermediate fusion models won't converge if directly trained end-to-end under V2V-H and the resulting models usually only exhibit good performance for either camera perception or LiDAR perception but hardly for both. Instead, we first train the model on single modality configurations (i.e., V2V-C and V2V-L) until convergence for 40 epochs and then fine-tune the models under V2V-H for 10 epochs with fixed parameters of modality-specific backbones on 4 RTX3090 GPUs. This training strategy can help models converge well with reliable performance. For a fair comparison, we leverage this training strategy for all the methods. + +Compared methods. We regard No Fusion as the baseline method. We also evaluate the Late Fusion, which transmits the detection proposals and leverages Non-maximum suppression to generate the final predictions. For the intermediate cooperation methods, we benchmark five approaches: V2VNet [40], DiscoNet [22], AttFuse [49], CoBEVT [46], and V2X-ViT [48]. For a fair comparison, hetero-modal head is used for all the models. + +# 4.3. Quantitative evaluation + +Main performance comparison. Tab. 1 demonstrates the performance comparisons on V2V-C, V2V-L and V2V-H. Under all three settings, all the cooperative methods outperform the No Fusion baseline and the intermediate fusion beats the classical Late Fusion, showing the great benefit of end-to-end V2V hetero-modal cooperative perception. The HM-ViT outperforms all the other SOTA intermediate fusion methods by at least $1.7\%$ , $8.4\%$ , $8.8\%$ in AP@0.7 under V2V-C, V2V-L, V2V-H settings respectively. + +Agent modality ratio experiment. As shown in Fig. 6, we fix the ego vehicle sensor modality and vary the ratio of collaborators' sensor modalities to evaluate the models' performance under various heterogeneous traffic scenarios. The larger LiDAR collaborator ratio corresponds to more vehicles only equipped with LiDARs while a smaller ratio means more vehicles only equipped with multiple cameras. The left two figures are the evaluation results for camera ego vehicles while the right two figures are the performance for LiDAR ego vehicles. Under most ratios, late fusion outperforms No Fusion however for LiDAR ego vehicle when most collaborators are camera agents, the Late Fusion performs poorer than No Fusion. We argue this is due to the fact that the camera predictions are usually noisy and merging proposals from different modalities equally could lead to ambiguity and thus deteriorate the performance. On the other hand, all the intermediate fusion methods outperform No Fusion by a large margin especially for the camera ego vehicles, demonstrating the great value of V2V cooperation between agents with different modalities. Among all the compared methods, HM-ViT ranks first for both camera ego vehicles and LiDAR ego vehicles under all the ratios, illustrating the great capability of HM-ViT for capturing modality-specific characteristics and cross-modality multiagent interactions. In contrast, other intermediate fusion methods only show good performance for a certain ratio range, which demonstrates the importance of heterogeneity for hetero-modal cooperative perception. + +![](images/c0302c822f4815ba85987e0b79d88ef98a2ecb234e13bb4e3626acd9a62379a4.jpg) +(a) No Fusion (camera) + +![](images/b06526066becb11c12c7d52956d9afcda555dcfa913998acfed321f51bf1e67b.jpg) +(b) Pure camera fusion + +![](images/4d199ca4ec200dd1d16a73ba30fe6193230c541a46c02b47ac70a6ad5644be41.jpg) +(c) Camera ego with LiDAR collaborators + +![](images/ccf7e5f1d7f260f42057f67baa282ebe14177328d73e10bfee06dc66dd4831d8.jpg) +(d) No Fusion (LiDAR) + +![](images/ef4e4fff65701a1ef4f443190d785853f10e545dec948c8755ce768e3ff87e23.jpg) +(e) Pure LiDAR fusion +Figure 8: Qualitative visualizations for (a) No Fusion with camera ego vehicle, (b) pure camera-based V2V perception, (c) hetero-modal V2V perception with camera ego vehicle and LiDAR collaborators, (d) No Fusion with LiDAR ego vehicle, (e) pure LiDAR-based V2V perception, and (f) hetero-modal V2V perception with LiDAR ego vehicle and camera collaborators. The red and green boxes represent the detection outputs and ground truth respectively. More visualizations can be found in the supplementary materials. + +![](images/1fbd1efc9248808eed6f40343152ebe4ec46e001b097f025d603bd9b92c4446f.jpg) +(f) LiDAR ego with camera collaborators + +
HM-MLP&LNH3GAT-LH3GAT-GAP@0.7
0.404
✓0.420
✓✓0.460
✓✓✓0.515
+ +Number of agent. In this experiment, we investigate the effect of the number of agents on the perception performance of HM-ViT. As Fig. 7 depicts, for both camera and LiDAR ego vehicles, the perception performance increases as more agents are involved in the cooperative perception and both LiDAR and camera collaborators can contribute to the performance gain for ego vehicles with different modalities, which again shows the benefit of hetero-modal V2V cooperation. Additionally, the increase rate generally decreases when increasing the number of agents and the LiDAR collaborators can bring more AP gains over the camera collaborators for both camera and LiDAR ego vehicles. Notably, similar as Fig. 6a-b display, the camera ego vehicles' performance can be greatly improved with only a small number of LiDAR collaborators, demonstrating the great potential of reducing the cost for each vehicle when the V2V system is deployed at scale as we may only need to install expansive LiDARs for a small number of agents (e.g., infrastructure) while all the other agents only require relatively cheap camera sensors. + +Table 2: Component Ablation study on the V2V-H setting + +
Compression RateAP@0.7
0x0.515
8x0.513
16x0.470
32x0.455
+ +Table 3: Compression rate effects for HM-ViT on V2V-H. + +Component ablation study. Here we investigate the key components of the proposed HM-ViT. As the layer normalization and MLP are usually combined together in typical transformer designs, thus we jointly evaluate the combined effect of HM-MLP and HM-LN (HM-MLP&LN). As Tab. 2 shows, all the proposed components improve the performance and local and global attentions can largely improve the AP@0.7 by $4\%$ and $5.5\%$ , proving the great benefit brought by jointly reasoning inter-agent and intra-agent heterogeneous interactions both locally and globally. + +Compression rate. Tab. 3 describes the influence of compression rate. It demonstrates that HM-ViT can still outperform other methods even under large compression rates. + +# 4.4. Qualitative results + +Fig. 8 depicts the qualitative visualizations for HM-ViT and No Fusion baselines. In fig. 8a-c, we plot the detection results for camera ego vehicles with no collaborator, camera collaborators, and LiDAR collaborators respectively while for Fig. 8d-f, we plot the results for LiDAR ego vehicles. As + +shown in these figures, the collaborations with both homogeneous and heterogeneous agents are beneficial for camera ego vehicles and LiDAR ego vehicles with enhanced detection results. In particular, the collaboration between camera ego vehicles and LiDAR collaborators can dramatically enhance the perception performance. + +# 5. Conclusion + +In this paper, we present HM-ViT, a hetero-modal vision transformer, for the hetero-modal multi-agent cooperative perception problem which is an important but under-explored research direction. We propose a generic heterogeneous 3D graph attention to jointly reason heterogeneous inter-agent and cross-agent interactions. Our extensive experiments demonstrate the outstanding performance of the proposed method and the great potential of hetero-modal multi-agent collaborations for increasing the scalability and robustness of V2V systems. We hope our findings and open-source efforts will inspire more research on this new problem. + +# References + +[1] Luís A Alexandre. 3d object recognition using convolutional neural networks with transfer learning between input channels. In Intelligent Autonomous Systems 13, pages 889-898. Springer, 2016. 3 +[2] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2 +[3] Johannes Betz, Hongrui Zheng, Alexander Liniger, Ugo Rosolia, Phillip Karle, Madhur Behl, Venkat Krovi, and Rahul Mangharam. Autonomous vehicles on the edge: A survey on autonomous vehicle racing. IEEE Open Journal of Intelligent Transportation Systems, 3:458-488, 2022. 1 +[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 3 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3 +[6] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 3 +[7] Hung-Hsun Chen, Yi-Bing Lin, I-Hau Yeh, Hsun-Jung Cho, and Yi-Jung Wu. Prediction of queue dissipation time for mixed traffic flows with deep learning. IEEE Open Journal of Intelligent Transportation Systems, 3:267-277, 2022. 1 +[8] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016. 3 +[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 3 +[10] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Graph-detr3d: Rethinking overlapping regions for multi-view 3d object detection. arXiv preprint arXiv:2204.11582, 2022. 3 +[11] Zhuo Deng and Longin Jan Latecki. A modal detection of 3d objects: Inferring 3d bounding boxes from 2d ones in rgb-depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5762-5770, 2017. 3 +[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driv + +ing simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6 +[13] Yi Guo and Jiaqi Ma. Leveraging existing high-occupancy vehicle lanes for mixed-autonomy traffic management with emerging connected automated vehicle applications. Transportmetrica A: Transport Science, 16(3):1375–1399, 2020. 1 +[14] Yi Guo, Jiaqi Ma, Edward Leslie, and Zhitong Huang. Evaluating the effectiveness of integrated connected automated vehicle applications applied to freeway managed lanes. IEEE Transactions on Intelligent Transportation Systems, 23(1):522-536, 2020. 1 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[16] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. arXiv preprint arXiv:2209.12836, 2022. 2 +[17] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3 +[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 5 +[19] Sou Kitajima, Hanna Chouchane, Jacobo Antona-Makoshi, Nobuyuki Uchida, and Jun Tajima. A nationwide impact assessment of automated driving systems on traffic safety using multiagent traffic simulations. IEEE Open Journal of Intelligent Transportation Systems, 3:302-312, 2022. 1 +[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 6 +[21] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In European Conference on Computer Vision, pages 316-332. Springer, 2022. 2 +[22] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 2, 6, 7 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3, 4, 6 +[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3 +[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6 + +[26] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 3 +[27] Jiaqi Ma, Edward Leslie, Amir Ghiasi, Zhitong Huang, and Yi Guo. Empirical analysis of a freeway bundled connected-and-automated vehicle application using experimental data. Journal of Transportation Engineering, Part A: Systems, 146(6):04020034, 2020. 1 +[28] Vasileia Papathanasopoulou, Ioanna Spyropoulou, Harris Perakis, Vassilis Gikas, and Eleni Andrikopoulou. A data-driven model for pedestrian behavior classification and trajectory prediction. IEEE Open Journal of Intelligent Transportation Systems, 3:328-339, 2022. 1 +[29] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 3 +[30] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgbd data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 3 +[31] Kelli Raboy, Jiaqi Ma, Edward Leslie, and Fang Zhou. A proof-of-concept field experiment on cooperative lane change maneuvers using a prototype connected automated vehicle testing platform. Journal of Intelligent Transportation Systems, 25(1):77-92, 2021. 1 +[32] Steven E Shladover. Opportunities and challenges in cooperative road vehicle automation. IEEE Open Journal of Intelligent Transportation Systems, 2:216-224, 2021. 1 +[33] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 3 +[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3 +[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 5 +[36] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. arXiv preprint arXiv:2204.01697, 2022.5 +[37] Rodolfo Valiente, Behrad Toghi, Ramtin Pedarsani, and Yaser P Fallah. Robustness and adaptability of reinforcement learning-based cooperative autonomous driving in mixed-autonomy traffic. IEEE Open Journal of Intelligent Transportation Systems, 3:397-410, 2022. 1 +[38] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on + +computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3 +[39] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3 +[40] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 2, 6, 7 +[41] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3 +[42] Zhixin Wang and Kui Jia. Frustum convnet: Sliding frustums to aggregate local point-wise features for amodal 3d object detection. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1742-1749. IEEE, 2019. 3 +[43] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. arXiv preprint arXiv:2209.13679, 2022. 2 +[44] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3 +[45] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6 +[46] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 2, 6, 7 +[47] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. arXiv preprint arXiv:2303.07601, 2023. 1 +[48] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. arXiv preprint arXiv:2203.10638, 2022. 1, 2, 6, 7 +[49] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 6, 7 +[50] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 2 + +[51] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 1 +[52] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 1 +[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1 +[54] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 1 +[55] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 2 \ No newline at end of file diff --git a/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/images.zip b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11d61379fa52d7db85de4d309b50791784de41a0 --- /dev/null +++ b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a6f8214e4d7a8b185f1231629ed9005f5d539ab237e4237d48518355392def9 +size 443976 diff --git a/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/layout.json b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aa825493c363e08bf18405e1cd878af0a0c11564 --- /dev/null +++ b/2023/HM-ViT_ Hetero-Modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer/layout.json @@ -0,0 +1,8456 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "type": "text", + "content": "HM-ViT: Hetero-modal Vehicle-to-Vehicle Cooperative Perception with Vision Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "spans": [ + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "text", + "content": "Hao Xiang" + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "text", + "content": ", Runsheng Xu" + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "text", + "content": ", Jiaqi Ma" + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 160, + 400, + 204 + ], + "type": "text", + "content": "University of California, Los Angeles \n{haxiang, rxx3386, jiaqima}@g.ucla.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 521 + ], + "type": "text", + "content": "Vehicle-to-Vehicle technologies have enabled autonomous vehicles to share information to see through occlusions, greatly enhancing perception performance. Nevertheless, existing works all focused on homogeneous traffic where vehicles are equipped with the same type of sensors, which significantly hampers the scale of collaboration and benefit of cross-modality interactions. In this paper, we investigate the multi-agent hetero-modal cooperative perception problem where agents may have distinct sensor modalities. We present HM-ViT, the first unified multi-agent hetero-modal cooperative perception framework that can collaboratively predict 3D objects for highly dynamic Vehicle-to-Vehicle (V2V) collaborations with varying numbers and types of agents. To effectively fuse features from multi-view images and LiDAR point clouds, we design a novel heterogeneous 3D graph transformer to jointly reason inter-agent and intra-agent interactions. The extensive experiments on the V2V perception dataset OPV2V demonstrate that the HM-ViT outperforms SOTA cooperative perception methods for V2V hetero-modal cooperative perception. Our code will be released at https://github.com/XHwind/HM-ViT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Recent advances in Vehicle-to-Vehicle (V2V) communication technology and intelligent transportation systems [13, 31, 27, 14, 7, 19, 28, 51, 53, 52] have allowed autonomous vehicles (AVs) to share sensory information, enabling them to perceive their surroundings better [3, 32, 54]. With the rapid growth of autonomous driving, V2V perception systems have the potential to be deployed at scale and create a safer transportation system. Cooperative perception systems, as shown in recent studies [49, 48, 40], can intelligently aggregate features from multiple vehicles within the communication range to enhance visual reasoning and overall performance." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 230, + 547, + 346 + ], + "blocks": [ + { + "bbox": [ + 308, + 230, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 308, + 230, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 547, + 346 + ], + "type": "image", + "image_path": "c46fc370206b843114e773d0cb899b250530a805e8e177e26b489f1a0ecca645.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 354, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 305, + 354, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 547, + 390 + ], + "type": "text", + "content": "Figure 1: Illustration of multi-agent hetero-modal V2V systems where each agent may be equipped with either LiDAR or multi-view cameras." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "type": "text", + "content": "Despite the rapid growth in this field, previous studies [49, 40, 46, 22, 37, 47] have primarily focused on homogeneous multi-agent cooperative perception, where all agents are equipped with the same type of sensors. In reality, however, agents may have different sensor modalities (hetero-modality) due to the cost and sensor preferences of ADS developers and car makers. As shown in Fig. 1, some agents are equipped with only LiDARs (LiDAR agents), while others only have multiple cameras (camera agents). Enabling collaboration between these heterogeneous agents could improve the sensing capability by allowing agents to see through occlusions and increase the scale and reliability of V2V systems. Additionally, LiDAR agents can provide accurate geometric information, while camera agents can provide rich semantic context. Thus, the collaboration between these agents could leverage the distinct but complementary environment attributes captured by each sensor modality to enhance the V2V perception systems. Furthermore, compared with the single-agent solution where multiple LiDARs and cameras are installed in a single vehicle, distributing different types of sensors across distinct agents could also potentially decrease the costs for each agent while still achieving satisfying performance. Nevertheless, whether, when, and how multi-agent hetero-modal V2V cooperation can benefit the perception system of heterogeneous traffic has not yet been studied." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "284" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 76, + 267, + 194 + ], + "blocks": [ + { + "bbox": [ + 67, + 76, + 267, + 194 + ], + "lines": [ + { + "bbox": [ + 67, + 76, + 267, + 194 + ], + "spans": [ + { + "bbox": [ + 67, + 76, + 267, + 194 + ], + "type": "image", + "image_path": "decd16733e4305627bce77a76256c8c75d4ea2ca15a907eb5fc94a984626fd3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 288, + 258 + ], + "type": "text", + "content": "Figure 2: Comparison of the single-agent multi-modal system and multi-agent hetero-modal system. The graph structure of the former is fixed whereas for the latter the graph is both dynamic and heterogeneous." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 579 + ], + "type": "text", + "content": "In this work, we address the multi-agent hetero-modal cooperative perception problem where each agent could have distinct sensor types and share/receive information with each other. Notably, as shown in Fig. 2, this multi-agent hetero-modal setting is distinct from the single-agent multi-modal setting. In the hetero-modal setting, the agent sensors form a dynamic heterogeneous graph where the existence and types of sensors are random, and the relative poses vary from scene to scene. In contrast, the sensor types/numbers and relative positions (extrinsics) between sensors are fixed in the single-agent multi-modal setting. Existing multi-modal methods heavily rely on these assumptions, and most of existing works [38, 39, 2, 55, 50] transform LiDAR points or 3D proposals onto the image plane to index 2D features. Their network architectures build upon the co-existence of both LiDAR and camera inputs with fixed geometric relationships. However, the dynamic nature of hetero-modal V2V perception requires a flexible architecture that can handle varying agent numbers and types, and the transmitted neural features are also spatially misaligned. Moreover, there are semantic discrepancies in the transmitted features between camera agents and LiDAR agents. Hence, these unique characteristics pose significant challenges for designing the multi-agent hetero-modal cooperative system and prevent adapting existing multi-modal fusion methods to this new problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "To enable collaboration between heterogeneous agents in V2V systems, we propose Hetero-Modal Vision Transformer (HM-ViT), the first unified cooperative perception framework that can leverage and fuse distributed information for hetero-modal V2V perception via a spatial-aware 3D heterogeneous vision transformer. Fig. 3 demonstrates the overall framework. Each agent first generates bird's eye view (BEV) representations through modality-specific encoders and then shares compressed features with neighboring agents. Afterward, the received features are decompressed and aggregated via the proposed HM-ViT," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "which conducts joint local and global heterogeneous 3D attentions with the consideration of both node and edge types. Our extensive experiments show that the HM-ViT can significantly improve the perception capability of camera agents and LiDAR agents over the single-agent baseline and outperforms SOTA cooperative perception methods by a large margin. In particular, for camera agents, performance can be boosted from " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "53.2\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}@0.7" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " with the collaboration of LiDAR agents, a 23-fold improvement. Our primary contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 198, + 545, + 438 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 198, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 545, + 293 + ], + "type": "text", + "content": "- We present the novel transformer framework (HMViT) for multi-agent hetero-modal cooperative perception, capable of capturing the modality-specific characteristics and heterogeneous 3D interactions. The proposed model exhibits superior flexibility and robustness with state-of-the-art performance on highly dynamic heterogeneous traffic involving varying agent numbers and types." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "text", + "content": "- We propose a generic heterogeneous 3D graph attention (H" + }, + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "text", + "content": "GAT), tailored for extracting inter-agent and intra-agent heterogeneous interactions. We instantiate two such attentions - local attention (H" + }, + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "text", + "content": "GAT-L) and global attention (H" + }, + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 317, + 300, + 545, + 372 + ], + "type": "text", + "content": "GAT-G) for capturing both local and global visual cues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 379, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 379, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 317, + 379, + 545, + 438 + ], + "type": "text", + "content": "- We conduct extensive benchmark experiments by varying sensor modalities, demonstrating the strong performance of the proposed method for hetero-modal V2V perception tasks. We will release all the codes and baselines to facilitate future research." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 389, + 466 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "V2V perception. V2V perception aims to enhance the perception performance of autonomous vehicles by leveraging shared information from other connected vehicles. Existing works have primarily focused on LiDAR-based 3D object perception. The pioneer cooperative perception methods transmit raw sensing observation (i.e., early fusion) or perception outputs (i.e., late fusion), whereas recent works [49, 40, 16, 43, 21, 48] are exploring the use of circulating intermediate neural features for achieving better performance-bandwidth trade-off. V2VNet [40] employs graph neural networks to aggregate the shared neural features for joint detection and prediction. AttFuse [49] uses single-head attention to model the per-location multi-agent interaction. Disconet [22] presents a matrix-valued edge weight for learning the interactions and a teacher-student learning framework to facilitate the training. V2X-ViT explores [48] vision transformer for vehicle-to-everything cooperation via window attention and heterogeneous self-attention. CoBEVT [46] presents a generic transformer framework for camera-based BEV semantic segmentation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "285" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 544, + 224 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 544, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 544, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 544, + 224 + ], + "type": "image", + "image_path": "a3d6260f86db65efd341a4abfd55938058b3a9ac4c29bc5490c7814aa8f5e172.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 548, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 548, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 548, + 295 + ], + "type": "text", + "content": "Figure 3: Overview of proposed hetero-modal V2V cooperative perception system. Each agent first produces BEV features through modality-specific feature extractors (Sec. 3.1). The BEV features are then compressed and shared (Sec. 3.3) with neighboring connected agents and the received features are decompressed in the ego agent side and fed into hetero-modal vision transformer to conduct graph-structured feature fusion (Sec. 3.2 and Sec. 3.3). The refined features are finally passed into the hetero-modal detection head to predict 3D bounding boxes (Sec. 3.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 376 + ], + "type": "text", + "content": "However, none of the existing works explored multi-agent multi-camera 3D object detection, let alone multi-agent hetero-modal perception. In contrast to existing methods, HM-ViT is the first to employ sparse heterogeneous local and global attentions to capture the 3D inter-agent and intraagent interactions in a computationally efficient manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 288, + 628 + ], + "type": "text", + "content": "Camera-based 3D object detection. Early works [4, 8, 33] mainly focus on monocular 3D detection but a single camera can only provide a 2D view of the scene, and inferring 3D from 2D is intrinsically hard. Recent development of self-driving datasets [5, 34, 6] featured with full sensor suits has enabled the research direction of 3D object detection from multiple cameras. DETR3D [41] proposes a 3D-2D query paradigm for extracting 3D features from 2D multi-view images. Graph-DETR3D [10] further utilizes graph structure learning to enhance the representation at the border regions. Alternatively, another stream of works [17, 29, 44, 23] focuses on aggregating BEV features from multi-view cameras for conducting downstream perception tasks. LSS [29] lifts the 2D features to 3D frustum via latent depth and then splats frustums into a BEV grid. " + }, + { + "bbox": [ + 46, + 377, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^2\\mathbf{BEV}" + }, + { + "bbox": [ + 46, + 377, + 288, + 628 + ], + "type": "text", + "content": " [44] further extends LSS with less memory consumption and conducts detection and segmentation. BEV-Former [23] constructs BEV queries and explores spatial cross-attention and temporal self-attention to recurrently refine the BEV features, achieving SOTA performance on both NuScenes [5] and Waymo Open Datasets [34]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Multi-modal fusion. Existing multi-sensor fusion methods can be divided into point-level fusion, proposal-level fusion, and BEV-level fusion. Point-level fusion decorates the input from one modality with attributes from the other modality. PointPainting [38] and PaintAugmenting [39] decorate the LiDAR point clouds with semantic segmentation scores and 2D CNN image features respectively while [11, 1] project" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 304, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 521 + ], + "type": "text", + "content": "LiDAR onto the image plane to augment the RGB values with depth information and conduct detection. On the other hand, proposal-level fusion generates proposals from one modality and then indexes features from the other modality for further refinement. MV3D [9] produces object queries from LiDAR BEV and then extracts the features from camera data and LiDAR front view. [30, 42] lift 2D bounding boxes to frustum and conduct 3D object detection from the frustum of point clouds. Conversely, BEV-level fusion converts features from different modalities to unified BEV representations, preserving both geometric and semantic information. BEVFusions [26, 24] concatenate BEV features from camera and LiDAR and fuse it via a fusion module. Hence, existing single-agent multi-modal fusion methods rely on the co-existence of both camera and LiDAR with fixed geometric relationships, which is unsuitable for our multi-agent hetero-modal cooperative perception problem with a dynamic heterogeneous collaboration graph." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 563, + 389, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 389, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 389, + 578 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we explore the multi-agent hetero-modal cooperative perception, where each AV is equipped with either a LiDAR or multiple cameras. Our goal is to create a robust and flexible cooperative perception system that allows for efficient collaboration between any number of agents with varying sensor types, ultimately improving the perception capabilities of the vehicle in a unified end-to-end fashion. The pipeline, illustrated in 3, includes modality-specific feature extraction, compression and sharing, HMViT for feature fusion, and hetero-modal detection head." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "286" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 79, + 178, + 217 + ], + "blocks": [ + { + "bbox": [ + 51, + 79, + 178, + 217 + ], + "lines": [ + { + "bbox": [ + 51, + 79, + 178, + 217 + ], + "spans": [ + { + "bbox": [ + 51, + 79, + 178, + 217 + ], + "type": "image", + "image_path": "86acf29b7f78e2c9b05186a39dfc9f09228fe4b7a2609a7881752d70c23c02b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "type": "text", + "content": "Figure 4: HM-ViT architecture. (a) Graph-structured fusion process. (b) Heterogeneous 3D graph attention " + }, + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{(H^3GAT)}" + }, + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "type": "text", + "content": ". (c) Illustration of sampled tokens for " + }, + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^{3}\\mathrm{GAT - L}" + }, + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^{3}\\mathrm{GAT - G}" + }, + { + "bbox": [ + 46, + 231, + 546, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 85, + 370, + 217 + ], + "blocks": [ + { + "bbox": [ + 197, + 85, + 370, + 217 + ], + "lines": [ + { + "bbox": [ + 197, + 85, + 370, + 217 + ], + "spans": [ + { + "bbox": [ + 197, + 85, + 370, + 217 + ], + "type": "image", + "image_path": "4fc9fb0d581f82ada74494b79ed32e4215416de18b85987bb3e6edca3aa09cf8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 90, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 391, + 90, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 391, + 90, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 391, + 90, + 542, + 217 + ], + "type": "image", + "image_path": "57c749347d29ecd3cfbe379064a23adf13e63fd2f5bb76aae6f8ba961938ae38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 266, + 237, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 237, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 237, + 280 + ], + "type": "text", + "content": "3.1. Modality-specific feature extraction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "type": "text", + "content": "LiDAR stem: We leverage PointPillar [20] to process point clouds for each LiDAR agent. The raw point cloud is converted to a 2D pseudo-image, flattened along the height dimension, and fed into 2D convolutional neural networks to produce the salient feature map " + }, + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j \\in \\mathcal{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "type": "text", + "content": ", which is compressed and shared with all the neighboring agents." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": "Camera stem: Each camera agent is equipped with " + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": " monocular cameras. The sensing observation of " + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": "-th agent includes the input images " + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "inline_equation", + "content": "I_k^i \\in \\mathcal{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": " and the known projection matrix " + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "inline_equation", + "content": "P_k^i \\in \\mathcal{R}^{3 \\times 4}" + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": " that maps 3D reference points to different image views. Our goal is to generate a BEV feature representation " + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i \\in \\mathcal{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": " that is amenable for feature fusion with other collaborators. In this work, we adopt similar architecture to BEVFormer [23] with no temporal information for feature extraction. For a faster running time, we adopt ResNet50 to extract 2D image features and then adopt a learnable 2D BEV query to inquire spatial information from the encoded multi-view features via spatial cross attention and projection matrices. The resulting refined BEV feature " + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": " is centered around agent " + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 360, + 289, + 539 + ], + "type": "text", + "content": " and shared with connected AVs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 550, + 282, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 282, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 282, + 563 + ], + "type": "text", + "content": "3.2. Heterogeneous 3D Graph Attention (H" + }, + { + "bbox": [ + 47, + 550, + 282, + 563 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 47, + 550, + 282, + 563 + ], + "type": "text", + "content": "GAT)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "To account for the distinct characteristics of BEV features extracted from different sensor modalities, the learning process of each modality must be distinguished, and the cross-modality interactions between multiple agents should vary. To capture this heterogeneity, we present a novel heterogeneous 3D graph attention (H" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "GAT), in which nodes and edges are type-dependent to reason spatial interactions and cross-agent relations jointly. We encode both local and global interactions to better capture the 3D ambiguity in BEV feature space. Local attention can help preserve object details, while global attention can provide a better understanding of environmental contexts such as road topology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 267, + 382, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 267, + 382, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 382, + 278 + ], + "type": "text", + "content": "and traffic density." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "text", + "content": "As shown in Fig. 4b, We build a 3D heterogeneous collaboration graph. Each node " + }, + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "inline_equation", + "content": "v(i,x) = \\mathbf{F}_x^i \\in \\mathcal{R}^C" + }, + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "text", + "content": " is a feature vector of agent " + }, + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "text", + "content": "'s feature map at spatial location " + }, + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{R}^2" + }, + { + "bbox": [ + 304, + 279, + 545, + 411 + ], + "type": "text", + "content": ". 3D heterogeneous graph attention is performed for spatially connected nodes in the BEV feature space. Depending on the definition of spatial connectivity, we will derive local attention and global attention. Here for notation simplicity, we only derive single-head equations but in real implementations, multi-head variants are used. Formally, we first project feature vectors onto different feature spaces to form query, key, and value vectors:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 386, + 417, + 545, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 417, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 386, + 417, + 545, + 432 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {x} ^ {j} = \\operatorname {D e n s e} _ {\\tau_ {j}} \\mathbf {F} _ {x} ^ {j} \\tag {1}", + "image_path": "af39515e8e5133c18394de4298ff9ad615420d295182580019c15efa9769068f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 387, + 434, + 545, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 434, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 387, + 434, + 545, + 447 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {x} ^ {j} = \\operatorname {D e n s e} _ {\\tau_ {j}} \\mathbf {F} _ {x} ^ {j} \\tag {2}", + "image_path": "0fd0de8278f50a0f492fb7a543dac049e5fff4ac7caa59e8fb63870dc7d99e31.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 388, + 449, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 449, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 388, + 449, + 545, + 464 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {x} ^ {j} = \\operatorname {D e n s e} _ {e _ {i j}} \\mathbf {F} _ {x} ^ {j} \\tag {3}", + "image_path": "49c8bdc5f4ec7f992cb26baaa2fbdf2fbb6f840cf8057784fcca2237199cd86b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{Dense}\\square" + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": " is a set of linear layers indexed by subscript " + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": ". For the query and key vectors, we use linear projectors " + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{Dense}_{\\tau_j}" + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": " indexed by node type " + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\tau_{j}" + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": " to extract modality-specific features. For the value vector, we index the projector via edge type " + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{Dense}_{e_{ij}}" + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": " to reflect the heterogeneity of cross-modality multi-agent interactions. The set of connected nodes of " + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "inline_equation", + "content": "v(i,x)" + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i,x)" + }, + { + "bbox": [ + 304, + 470, + 545, + 566 + ], + "type": "text", + "content": ". Then the attention is operated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 571, + 545, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 571, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 349, + 571, + 545, + 593 + ], + "type": "interline_equation", + "content": "\\mathbf {a} (j, y) = \\operatorname {S o f t m a x} _ {(j, y) \\in \\mathcal {N} (i, x)} \\left(\\mathbf {Q} _ {x} ^ {i} \\mathbf {W} _ {e _ {i j}} \\mathbf {K} _ {y} ^ {j}\\right) \\tag {4}", + "image_path": "a40f335d30a6781452e13ef8b4c95b059fd3bbedf72c8539ce9f3964e40e5e25.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 594, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 594, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 366, + 594, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {x} ^ {i} = \\sum_ {(j, y) \\in \\mathcal {N} (i, x)} \\mathbf {a} (j, y) \\mathbf {V} _ {y} ^ {j} \\tag {5}", + "image_path": "138d411a8df6d2bbdacc2a5e4c1abf6731b4ebeaeb70ee55332f23630b9e8a0d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{e_{ij}}\\in \\mathcal{R}^{C\\times C}" + }, + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "text", + "content": " is used to adjust the dot product of " + }, + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_x^i" + }, + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_y^j" + }, + { + "bbox": [ + 305, + 629, + 545, + 654 + ], + "type": "text", + "content": " to further encode the heterogeneity of edges." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Depending on how the nodes are sampled (Fig. 4c), we design two types of attentions: local attention " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{H}^3\\mathbf{GAT - L})" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " which performs local window-based attention and global attention " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{H}^3\\mathbf{GAT - G})" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " which performs sparse global grid-based attention. Fig. 4c visualizes how the local and global" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "287" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 153, + 256 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 153, + 256 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 153, + 256 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 153, + 256 + ], + "type": "image", + "image_path": "7ada8e21e4acbb6353e610e741956fd6ab9acb59ff6cc96901e0633db245bad6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 258, + 143, + 270 + ], + "lines": [ + { + "bbox": [ + 75, + 258, + 143, + 270 + ], + "spans": [ + { + "bbox": [ + 75, + 258, + 143, + 270 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 75, + 258, + 143, + 270 + ], + "type": "inline_equation", + "content": "\\mathsf{H}^3\\mathsf{GAT - L}" + }, + { + "bbox": [ + 75, + 258, + 143, + 270 + ], + "type": "text", + "content": " Block" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 283, + 287, + 306 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 306 + ], + "type": "text", + "content": "Figure 5: Transformer blocks for local and global attentions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 267, + 256 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 267, + 256 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 267, + 256 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 267, + 256 + ], + "type": "image", + "image_path": "f7340ba696d721cbe52da55b71876f0b0da4da42e26ba2b48e178ea989076393.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 258, + 261, + 270 + ], + "lines": [ + { + "bbox": [ + 190, + 258, + 261, + 270 + ], + "spans": [ + { + "bbox": [ + 190, + 258, + 261, + 270 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 190, + 258, + 261, + 270 + ], + "type": "inline_equation", + "content": "\\mathsf{H}^3\\mathsf{GAT - G}" + }, + { + "bbox": [ + 190, + 258, + 261, + 270 + ], + "type": "text", + "content": " Block" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 389 + ], + "type": "text", + "content": "tokens are sampled, where the local tokens are sampled within local windows and the global tokens are sparsely sampled grids scattered across feature maps. The local interactions can help preserve spatial cues and provide reliable estimates while the global reasoning can help understand global semantic context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": "Both " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^3\\mathrm{GAT - L}" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^3\\mathrm{GAT - G}" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": " can be implemented efficiently by decomposing the spatial axes. More specifically, we stack all the agents' features to " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{F}\\in \\mathcal{R}^{N\\times H\\times W\\times C}" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": " is the number of agents. For " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^3\\mathrm{GAT - L}" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": ", we decompose the feature map into 3D non-overlapping windows along the first axis [35, 36], each of size " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "N\\times P\\times P" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": ". The partitioned tensor has the shape " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "(\\frac{H}{P}\\times \\frac{W}{P},N\\times P^2,C)" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": " where the heterogeneous 3D local graph attention is conducted for " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "NP^2" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": " tokens within the same window. Similarly, for " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^3\\mathrm{GAT - G}" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": ", we swap the axis and partition the tensor into the shape " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "(N\\times P^2,\\frac{H}{P}\\times \\frac{W}{P},C)" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": ". Due to the swap operation, the sampled grids will be sparsely scattered and the attention is operated for these sparsely sampled " + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\frac{H}{P}\\times \\frac{W}{P}" + }, + { + "bbox": [ + 46, + 390, + 287, + 557 + ], + "type": "text", + "content": " grids, which can capture sparse global information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "content": "To integrate this local and global attention into transformer architecture, we further present a heterogeneous normalization layer (HM-LN) and heterogeneous MLP (HMMLP) which use type-dependent parameters. As shown in Fig. 5, we first pass all the features into the HM-LN where different statistics are calculated and used as per each agent's modality type. Afterward, we feed the normalized features into a heterogeneous 3D graph attention " + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{H}^3\\mathrm{GATL / H}^3\\mathrm{GAT - G})" + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "content": " to jointly reason heterogeneous inter-agent and intra-agent interactions. Then, we pass the fused features to another HM-LN followed by a Hetero-modal MLP layer where different sets of parameters are used for camera and LiDAR features. By carefully designing these com" + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 310, + 93, + 545, + 317 + ], + "blocks": [ + { + "bbox": [ + 307, + 78, + 536, + 91 + ], + "lines": [ + { + "bbox": [ + 307, + 78, + 536, + 91 + ], + "spans": [ + { + "bbox": [ + 307, + 78, + 536, + 91 + ], + "type": "text", + "content": "Algorithm 1 Multi-agent hetero-modal fusion process" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "lines": [ + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": "1: Input: decompressed feature " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": ", pose " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " for each agent \n2: " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{(0)} = \\mathbf{F}_i" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "l = 1, \\ldots, L" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " do \n4: for each agent " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Process in parallel \n5: " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{j \\rightarrow i}^{(l-1)} = \\Gamma_{j \\rightarrow i}\\left(\\mathbf{F}_j^{(l-1)}\\right) \\triangleright" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Spatially transform neighboring agents' features \n6: " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{(l)} = \\mathrm{H}^3\\mathrm{GAT-L}" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Block(" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{j \\rightarrow i}^{(l-1)}\\}" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": ") " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Update node via local attention \n7: end for \n8: for each agent " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Process in parallel \n9: " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{j \\rightarrow i}^{(l)} = \\Gamma_{j \\rightarrow i}\\left(\\mathbf{F}_j^{(l)}\\right) \\triangleright" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Spatially transform neighboring agents' features \n10: " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{(l)} = \\mathrm{H}^3\\mathrm{GAT-G}" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Block(" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{j \\rightarrow i}^{(l)}\\}" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": ") " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Update node via global attention \n11: end for \n12: end for \n13: " + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i = \\mathrm{HM - MLP}\\left(\\mathbf{F}_i^{(L)}\\right)" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 310, + 93, + 545, + 317 + ], + "type": "text", + "content": " Output updated features" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 329, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 329, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 545, + 365 + ], + "type": "text", + "content": "ponents, we can maintain modality-specific characteristics throughout the fusion process while benefiting from cross-modality multi-agent interactions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 372, + 488, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 488, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 488, + 384 + ], + "type": "text", + "content": "3.3. Hetero-modal Vision Transformer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "content": "Compression and sharing: To reduce the transmission bandwidth, a series of " + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "content": " convolutions is applied to reduce the transmitted feature size along the channel dimension. Together with the intermediate features, each agent's pose " + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "content": " is also circulated within the collaboration graph. The ego agent will receive these features and decompress them back to the original size via another convolutional network. For processing intermediate features of the camera agent and LiDAR agent, we leverage distinct parameters in the compression and decompression modules to preserve the modality-specific characteristics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "type": "text", + "content": "Graph-structured feature fusion: The received BEV features are centered around different spatial locations as each agent perceives the dynamic environment from different view points. To this end, we present a graph-structured fusion process (Fig. 4a): each node maintains a state representation of an agent in its own coordinate frame, and for a fixed number of iterations, spatially warped messages are shared between nodes and the node states are updated based on the aggregated features via the transformer blocks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "The overall process is summarized in Alg. 1. During each iteration, we have two cascaded node updates which capture the local and global heterogeneous interactions respectively. For each node, we first spatially transform [18] neighboring nodes' features to its center " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{j\\rightarrow i} = \\Gamma_{j\\rightarrow i}(\\mathbf{F}_j)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". When the transmitting node is the receiving node itself, the transformation matrix is an identity matrix" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "288" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 69, + 493, + 203 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 493, + 203 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 493, + 203 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 493, + 203 + ], + "type": "table", + "html": "
ModelsV2V-CV2V-LV2V-H
AP@0.5AP@0.7AP@0.5AP@0.7AP@0.5AP@0.7
No Fusion0.0940.0210.5240.3630.2840.157
Late Fusion0.2310.0700.7700.6060.5020.308
V2VNet [40]0.3290.1250.8200.6450.6500.366
DiscoNet [22]0.2870.1150.7410.5900.6240.385
AttFuse [49]0.2610.0950.8010.6440.6470.390
CoBEVT [46]0.3170.1220.8280.6370.6740.416
V2X-ViT [48]0.3320.1250.8330.6790.6710.427
HM-ViT0.3550.1420.8530.7630.6950.515
", + "image_path": "0ee682ff98423e6f75a4a3cc9c480c242cb054f90610172fee96c706b5233b8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 211, + 432, + 224 + ], + "lines": [ + { + "bbox": [ + 159, + 211, + 432, + 224 + ], + "spans": [ + { + "bbox": [ + 159, + 211, + 432, + 224 + ], + "type": "text", + "content": "Table 1: Evaluation of V2V perception methods on OPV2V dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "text", + "content": "and thus " + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i\\rightarrow i} = \\mathbf{F}_i" + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "text", + "content": ". These spatial aligned feature maps " + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{j\\rightarrow i}" + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "text", + "content": " are then shared with agent " + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "text", + "content": " to update its state representation via the aggregation module. We adopt " + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^3\\mathrm{GAT - L}" + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "text", + "content": " Block as our first aggregation module to capture the local heterogeneous interactions and leverage " + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^3\\mathrm{GAT - G}" + }, + { + "bbox": [ + 46, + 244, + 288, + 413 + ], + "type": "text", + "content": " Block for the second module to further refine the states with global cues. Within each transformer block, we also adopt a mask to mask out non-overlapping areas between the field of views when computing the attention scores. Note that each agent's state update can be processed in parallel for better efficiency. After L such iterations, we pass the features to a hetero-modal MLP to further refine the feature representation. Throughout the whole fusion process, the modality-specific statistics are maintained." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 419, + 162, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 162, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 162, + 431 + ], + "type": "text", + "content": "3.4. Hetero-modal Head" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "text", + "content": "As camera and LiDAR contain distinct characteristics, we design a hetero-modal head where a different set of parameters are applied for camera and LiDAR ego vehicles to generate the final predictions. More specifically, the final fused feature maps are passed to a series of " + }, + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "text", + "content": " convolutions with batch normalization and ReLU for feature refinement. Then, we adopt a " + }, + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "text", + "content": " convolution layer to generate the regression and classification predictions. Smooth " + }, + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 46, + 437, + 287, + 559 + ], + "type": "text", + "content": " loss is utilized for regression and a focal loss [25] is used for classification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 587, + 183, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 183, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 183, + 600 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "OPV2V. OP2V [49] is a large-scale multi-modal cooperative V2V perception dataset collected in CARLA [12] and OpenCDA [45]. It contains over 70 driving scenarios of around 25 seconds duration each. Each scenario contains multiple connected AVs (2 to 7) and each AV is equipped with 1 LiDAR and 4 monocular cameras covering " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " horizontal field of view (FoV). In our hetero-modal cooperative perception setting, we only use one type of sensor modality for each AV, leading to two types of agents: vehicles only" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 245, + 544, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 544, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 544, + 269 + ], + "type": "text", + "content": "equipped with multiple cameras (camera agent), and vehicles only equipped with LiDARs (LiDAR agent)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 269, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 460 + ], + "type": "text", + "content": "**Evaluations.** We adopt Average precision (AP) at Intersection-over-Union (IoU) 0.5 and 0.7 to measure the perception performance. As each scenario consists of multiple AVs, a fixed agent is selected as the ego agent and the evaluation is conducted in the range of " + }, + { + "bbox": [ + 304, + 269, + 546, + 460 + ], + "type": "inline_equation", + "content": "100\\mathrm{m} \\times 100\\mathrm{m}" + }, + { + "bbox": [ + 304, + 269, + 546, + 460 + ], + "type": "text", + "content": " around it. Following [49, 48], the train/validation/test splits are 6764/1981/2719. We evaluate models mainly under three configurations: 1) V2V Camera-based 3D detection (V2V-C) where AVs are only equipped with 4 cameras with 360 horizontal FoV, 2) V2V LiDAR-based detection (V2V-L) where all the agents only have LiDAR sensors, and 3) V2V Hetero-modal detection (V2V-H) where half of the agents only has cameras while the other half only has LiDARs. To further assess models' capability with dynamic sensor configurations, we also assess models with fixed ego sensor modality and varying collaborator sensor types." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 468, + 425, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 425, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 425, + 480 + ], + "type": "text", + "content": "4.2. Experimental Setups" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. We use the PointPillar [20] as 3D backbones and a modified BEVFormer [23] for our camera stem. For BEVFormer, we adopt its variant with no temporal information and ResNet50 [15] as image backbones for better computation efficiency and use a smaller grid resolution " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "(0.4\\mathrm{m})" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " to preserve fine-trained spatial details. The intermediate BEV feature map has a dimension of " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "128\\times 128\\times 256" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": ". Following prior works [22, 49, 48, 46], we only change the fusion module for different intermediate fusion methods while keeping the other components such as hetero-modal header, compression, and lidar/camera feature extractors the same. For the heterogeneous methods, we provide the network with additional feature types so that they can distinguish whether the feature is from Li-DAR or camera while for the homogeneous methods, the fusion networks would treat the features from LiDAR and camera equally. For HM-ViT, we conduct two iterations of graph-structured feature fusion and employ a window size of 8 for both local and global attentions. We adopt AdamW" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "289" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 83, + 173, + 182 + ], + "blocks": [ + { + "bbox": [ + 52, + 83, + 173, + 182 + ], + "lines": [ + { + "bbox": [ + 52, + 83, + 173, + 182 + ], + "spans": [ + { + "bbox": [ + 52, + 83, + 173, + 182 + ], + "type": "image", + "image_path": "072da0bf6f1f36e6688fa35fb69d0193c768a2ececc5be919b4148f17cca1981.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 228 + ], + "type": "text", + "content": "Figure 6: Agent modality ratio experiment. The x-axis is the ratio of LiDAR collaborators among all the collaborators. In (a) and (b), ego vehicles are equipped with cameras. In (c) and (d), ego vehicles are equipped with LiDARs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 83, + 296, + 182 + ], + "blocks": [ + { + "bbox": [ + 173, + 83, + 296, + 182 + ], + "lines": [ + { + "bbox": [ + 173, + 83, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 173, + 83, + 296, + 182 + ], + "type": "image", + "image_path": "f11ffca66134f4c86478c9ef2a53f2dc66a9da005959bf656fb4060b18fbea9f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 83, + 417, + 182 + ], + "blocks": [ + { + "bbox": [ + 296, + 83, + 417, + 182 + ], + "lines": [ + { + "bbox": [ + 296, + 83, + 417, + 182 + ], + "spans": [ + { + "bbox": [ + 296, + 83, + 417, + 182 + ], + "type": "image", + "image_path": "b60d85e61a24b6d3577254dc92d172a28c088b24c62b6da0d60bc51a95fda620.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 418, + 83, + 541, + 182 + ], + "blocks": [ + { + "bbox": [ + 418, + 83, + 541, + 182 + ], + "lines": [ + { + "bbox": [ + 418, + 83, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 418, + 83, + 541, + 182 + ], + "type": "image", + "image_path": "7087ab578d79b0dabf77d590c1ef11e3b339237b253faf099c27b405a32a7d61.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 182, + 462, + 194 + ], + "blocks": [ + { + "bbox": [ + 133, + 182, + 462, + 194 + ], + "lines": [ + { + "bbox": [ + 133, + 182, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 462, + 194 + ], + "type": "image", + "image_path": "051c2e88d1efd0531c58309bad1027680e09f5d57b9544bae1225dd866f3ed9f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 239, + 166, + 335 + ], + "blocks": [ + { + "bbox": [ + 48, + 239, + 166, + 335 + ], + "lines": [ + { + "bbox": [ + 48, + 239, + 166, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 166, + 335 + ], + "type": "image", + "image_path": "061acd15789cec06515163b772ff539ce1b898e70b4d7b4d80789f635db8730e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 344, + 288, + 405 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 288, + 405 + ], + "type": "text", + "content": "Figure 7: Ablation study on number of agent. The LiDAR/camera collaborator refers to the case where the collaborators are equipped with LiDAR and Camera respectively. (a) ego vehicle is equipped with camera. (b) ego vehicle is equipped with LiDAR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 166, + 239, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 166, + 239, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 166, + 239, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 166, + 239, + 286, + 335 + ], + "type": "image", + "image_path": "614a9178da5e734af03673af75ea16f4e44f1620135efcdb9a5e56cef8d70164.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "text", + "content": "with a decay rate of " + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 438, + 287, + 462 + ], + "type": "text", + "content": " and cosine annealing learning rate scheduler to optimize the models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 463, + 288, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 288, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 288, + 606 + ], + "type": "text", + "content": "Training strategy. We find that the intermediate fusion models won't converge if directly trained end-to-end under V2V-H and the resulting models usually only exhibit good performance for either camera perception or LiDAR perception but hardly for both. Instead, we first train the model on single modality configurations (i.e., V2V-C and V2V-L) until convergence for 40 epochs and then fine-tune the models under V2V-H for 10 epochs with fixed parameters of modality-specific backbones on 4 RTX3090 GPUs. This training strategy can help models converge well with reliable performance. For a fair comparison, we leverage this training strategy for all the methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Compared methods. We regard No Fusion as the baseline method. We also evaluate the Late Fusion, which transmits the detection proposals and leverages Non-maximum suppression to generate the final predictions. For the intermediate cooperation methods, we benchmark five approaches: V2VNet [40], DiscoNet [22], AttFuse [49], CoBEVT [46], and V2X-ViT [48]. For a fair comparison, hetero-modal head is used for all the models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 241, + 440, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 440, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 440, + 255 + ], + "type": "text", + "content": "4.3. Quantitative evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "text", + "content": "Main performance comparison. Tab. 1 demonstrates the performance comparisons on V2V-C, V2V-L and V2V-H. Under all three settings, all the cooperative methods outperform the No Fusion baseline and the intermediate fusion beats the classical Late Fusion, showing the great benefit of end-to-end V2V hetero-modal cooperative perception. The HM-ViT outperforms all the other SOTA intermediate fusion methods by at least " + }, + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 305, + 265, + 547, + 373 + ], + "type": "text", + "content": " in AP@0.7 under V2V-C, V2V-L, V2V-H settings respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": "Agent modality ratio experiment. As shown in Fig. 6, we fix the ego vehicle sensor modality and vary the ratio of collaborators' sensor modalities to evaluate the models' performance under various heterogeneous traffic scenarios. The larger LiDAR collaborator ratio corresponds to more vehicles only equipped with LiDARs while a smaller ratio means more vehicles only equipped with multiple cameras. The left two figures are the evaluation results for camera ego vehicles while the right two figures are the performance for LiDAR ego vehicles. Under most ratios, late fusion outperforms No Fusion however for LiDAR ego vehicle when most collaborators are camera agents, the Late Fusion performs poorer than No Fusion. We argue this is due to the fact that the camera predictions are usually noisy and merging proposals from different modalities equally could lead to ambiguity and thus deteriorate the performance. On the other hand, all the intermediate fusion methods outperform No Fusion by a large margin especially for the camera ego vehicles, demonstrating the great value of V2V cooperation between agents with different modalities. Among all the compared methods, HM-ViT ranks first for both camera ego vehicles and LiDAR ego vehicles under all the ratios, illustrating the great capability of HM-ViT for capturing modality-specific characteristics and cross-modality multiagent interactions. In contrast, other intermediate fusion methods only show good performance for a certain ratio range, which demonstrates the importance of heterogeneity for hetero-modal cooperative perception." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "290" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 110, + 220, + 186 + ], + "blocks": [ + { + "bbox": [ + 72, + 110, + 220, + 186 + ], + "lines": [ + { + "bbox": [ + 72, + 110, + 220, + 186 + ], + "spans": [ + { + "bbox": [ + 72, + 110, + 220, + 186 + ], + "type": "image", + "image_path": "c0302c822f4815ba85987e0b79d88ef98a2ecb234e13bb4e3626acd9a62379a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 189, + 187, + 199 + ], + "lines": [ + { + "bbox": [ + 102, + 189, + 187, + 199 + ], + "spans": [ + { + "bbox": [ + 102, + 189, + 187, + 199 + ], + "type": "text", + "content": "(a) No Fusion (camera)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 110, + 374, + 186 + ], + "blocks": [ + { + "bbox": [ + 227, + 110, + 374, + 186 + ], + "lines": [ + { + "bbox": [ + 227, + 110, + 374, + 186 + ], + "spans": [ + { + "bbox": [ + 227, + 110, + 374, + 186 + ], + "type": "image", + "image_path": "b06526066becb11c12c7d52956d9afcda555dcfa913998acfed321f51bf1e67b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 189, + 340, + 199 + ], + "lines": [ + { + "bbox": [ + 257, + 189, + 340, + 199 + ], + "spans": [ + { + "bbox": [ + 257, + 189, + 340, + 199 + ], + "type": "text", + "content": "(b) Pure camera fusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 377, + 110, + 525, + 186 + ], + "blocks": [ + { + "bbox": [ + 377, + 110, + 525, + 186 + ], + "lines": [ + { + "bbox": [ + 377, + 110, + 525, + 186 + ], + "spans": [ + { + "bbox": [ + 377, + 110, + 525, + 186 + ], + "type": "image", + "image_path": "4d199ca4ec200dd1d16a73ba30fe6193230c541a46c02b47ac70a6ad5644be41.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 189, + 526, + 199 + ], + "lines": [ + { + "bbox": [ + 373, + 189, + 526, + 199 + ], + "spans": [ + { + "bbox": [ + 373, + 189, + 526, + 199 + ], + "type": "text", + "content": "(c) Camera ego with LiDAR collaborators" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 74, + 205, + 221, + 280 + ], + "blocks": [ + { + "bbox": [ + 74, + 205, + 221, + 280 + ], + "lines": [ + { + "bbox": [ + 74, + 205, + 221, + 280 + ], + "spans": [ + { + "bbox": [ + 74, + 205, + 221, + 280 + ], + "type": "image", + "image_path": "ccf7e5f1d7f260f42057f67baa282ebe14177328d73e10bfee06dc66dd4831d8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 282, + 189, + 293 + ], + "lines": [ + { + "bbox": [ + 102, + 282, + 189, + 293 + ], + "spans": [ + { + "bbox": [ + 102, + 282, + 189, + 293 + ], + "type": "text", + "content": "(d) No Fusion (LiDAR)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 205, + 372, + 280 + ], + "blocks": [ + { + "bbox": [ + 225, + 205, + 372, + 280 + ], + "lines": [ + { + "bbox": [ + 225, + 205, + 372, + 280 + ], + "spans": [ + { + "bbox": [ + 225, + 205, + 372, + 280 + ], + "type": "image", + "image_path": "ef4e4fff65701a1ef4f443190d785853f10e545dec948c8755ce768e3ff87e23.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 282, + 339, + 293 + ], + "lines": [ + { + "bbox": [ + 255, + 282, + 339, + 293 + ], + "spans": [ + { + "bbox": [ + 255, + 282, + 339, + 293 + ], + "type": "text", + "content": "(e) Pure LiDAR fusion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 303, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 545, + 363 + ], + "type": "text", + "content": "Figure 8: Qualitative visualizations for (a) No Fusion with camera ego vehicle, (b) pure camera-based V2V perception, (c) hetero-modal V2V perception with camera ego vehicle and LiDAR collaborators, (d) No Fusion with LiDAR ego vehicle, (e) pure LiDAR-based V2V perception, and (f) hetero-modal V2V perception with LiDAR ego vehicle and camera collaborators. The red and green boxes represent the detection outputs and ground truth respectively. More visualizations can be found in the supplementary materials." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 376, + 205, + 524, + 280 + ], + "blocks": [ + { + "bbox": [ + 376, + 205, + 524, + 280 + ], + "lines": [ + { + "bbox": [ + 376, + 205, + 524, + 280 + ], + "spans": [ + { + "bbox": [ + 376, + 205, + 524, + 280 + ], + "type": "image", + "image_path": "1fbd1efc9248808eed6f40343152ebe4ec46e001b097f025d603bd9b92c4446f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 282, + 522, + 293 + ], + "lines": [ + { + "bbox": [ + 373, + 282, + 522, + 293 + ], + "spans": [ + { + "bbox": [ + 373, + 282, + 522, + 293 + ], + "type": "text", + "content": "(f) LiDAR ego with camera collaborators" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 50, + 373, + 285, + 441 + ], + "blocks": [ + { + "bbox": [ + 50, + 373, + 285, + 441 + ], + "lines": [ + { + "bbox": [ + 50, + 373, + 285, + 441 + ], + "spans": [ + { + "bbox": [ + 50, + 373, + 285, + 441 + ], + "type": "table", + "html": "
HM-MLP&LNH3GAT-LH3GAT-GAP@0.7
0.404
✓0.420
✓✓0.460
✓✓✓0.515
", + "image_path": "bc54f17c776baf52bcd89f23d88cd6745e8fbc302a6cc2c45108308df12d2a01.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "content": "Number of agent. In this experiment, we investigate the effect of the number of agents on the perception performance of HM-ViT. As Fig. 7 depicts, for both camera and LiDAR ego vehicles, the perception performance increases as more agents are involved in the cooperative perception and both LiDAR and camera collaborators can contribute to the performance gain for ego vehicles with different modalities, which again shows the benefit of hetero-modal V2V cooperation. Additionally, the increase rate generally decreases when increasing the number of agents and the LiDAR collaborators can bring more AP gains over the camera collaborators for both camera and LiDAR ego vehicles. Notably, similar as Fig. 6a-b display, the camera ego vehicles' performance can be greatly improved with only a small number of LiDAR collaborators, demonstrating the great potential of reducing the cost for each vehicle when the V2V system is deployed at scale as we may only need to install expansive LiDARs for a small number of agents (e.g., infrastructure) while all the other agents only require relatively cheap camera sensors." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 356, + 373, + 496, + 441 + ], + "blocks": [ + { + "bbox": [ + 50, + 449, + 284, + 462 + ], + "lines": [ + { + "bbox": [ + 50, + 449, + 284, + 462 + ], + "spans": [ + { + "bbox": [ + 50, + 449, + 284, + 462 + ], + "type": "text", + "content": "Table 2: Component Ablation study on the V2V-H setting" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 356, + 373, + 496, + 441 + ], + "lines": [ + { + "bbox": [ + 356, + 373, + 496, + 441 + ], + "spans": [ + { + "bbox": [ + 356, + 373, + 496, + 441 + ], + "type": "table", + "html": "
Compression RateAP@0.7
0x0.515
8x0.513
16x0.470
32x0.455
", + "image_path": "77b49c34ef2c7e2a67942921c6b8758377dfc130ad6d6e4bb8cc92e41fefd57f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 449, + 542, + 462 + ], + "lines": [ + { + "bbox": [ + 308, + 449, + 542, + 462 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 542, + 462 + ], + "type": "text", + "content": "Table 3: Compression rate effects for HM-ViT on V2V-H." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": "Component ablation study. Here we investigate the key components of the proposed HM-ViT. As the layer normalization and MLP are usually combined together in typical transformer designs, thus we jointly evaluate the combined effect of HM-MLP and HM-LN (HM-MLP&LN). As Tab. 2 shows, all the proposed components improve the performance and local and global attentions can largely improve the AP@0.7 by " + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": ", proving the great benefit brought by jointly reasoning inter-agent and intra-agent heterogeneous interactions both locally and globally." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 628 + ], + "type": "text", + "content": "Compression rate. Tab. 3 describes the influence of compression rate. It demonstrates that HM-ViT can still outperform other methods even under large compression rates." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 635, + 414, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 414, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 414, + 647 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Fig. 8 depicts the qualitative visualizations for HM-ViT and No Fusion baselines. In fig. 8a-c, we plot the detection results for camera ego vehicles with no collaborator, camera collaborators, and LiDAR collaborators respectively while for Fig. 8d-f, we plot the results for LiDAR ego vehicles. As" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "291" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 144 + ], + "type": "text", + "content": "shown in these figures, the collaborations with both homogeneous and heterogeneous agents are beneficial for camera ego vehicles and LiDAR ego vehicles with enhanced detection results. In particular, the collaboration between camera ego vehicles and LiDAR collaborators can dramatically enhance the perception performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 154, + 118, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 154, + 118, + 167 + ], + "spans": [ + { + "bbox": [ + 49, + 154, + 118, + 167 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 286, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 286, + 318 + ], + "type": "text", + "content": "In this paper, we present HM-ViT, a hetero-modal vision transformer, for the hetero-modal multi-agent cooperative perception problem which is an important but under-explored research direction. We propose a generic heterogeneous 3D graph attention to jointly reason heterogeneous inter-agent and cross-agent interactions. Our extensive experiments demonstrate the outstanding performance of the proposed method and the great potential of hetero-modal multi-agent collaborations for increasing the scalability and robustness of V2V systems. We hope our findings and open-source efforts will inspire more research on this new problem." + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "text", + "content": "292" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Luís A Alexandre. 3d object recognition using convolutional neural networks with transfer learning between input channels. In Intelligent Autonomous Systems 13, pages 889-898. Springer, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "text", + "content": "[2] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 259 + ], + "type": "text", + "content": "[3] Johannes Betz, Hongrui Zheng, Alexander Liniger, Ugo Rosolia, Phillip Karle, Madhur Behl, Venkat Krovi, and Rahul Mangharam. Autonomous vehicles on the edge: A survey on autonomous vehicle racing. IEEE Open Journal of Intelligent Transportation Systems, 3:458-488, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "text", + "content": "[4] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9287-9296, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 372 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 374, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 288, + 440 + ], + "type": "text", + "content": "[6] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 485 + ], + "type": "text", + "content": "[7] Hung-Hsun Chen, Yi-Bing Lin, I-Hau Yeh, Hsun-Jung Cho, and Yi-Jung Wu. Prediction of queue dissipation time for mixed traffic flows with deep learning. IEEE Open Journal of Intelligent Transportation Systems, 3:267-277, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 542 + ], + "type": "text", + "content": "[8] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3d object detection for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2147-2156, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "type": "text", + "content": "[9] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 633 + ], + "type": "text", + "content": "[10] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Graph-detr3d: Rethinking overlapping regions for multi-view 3d object detection. arXiv preprint arXiv:2204.11582, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[11] Zhuo Deng and Longin Jan Latecki. A modal detection of 3d objects: Inferring 3d bounding boxes from 2d ones in rgb-depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5762-5770, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driv" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "ing simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[13] Yi Guo and Jiaqi Ma. Leveraging existing high-occupancy vehicle lanes for mixed-autonomy traffic management with emerging connected automated vehicle applications. Transportmetrica A: Transport Science, 16(3):1375–1399, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "text", + "content": "[14] Yi Guo, Jiaqi Ma, Edward Leslie, and Zhitong Huang. Evaluating the effectiveness of integrated connected automated vehicle applications applied to freeway managed lanes. IEEE Transactions on Intelligent Transportation Systems, 23(1):522-536, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "text", + "content": "[16] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. arXiv preprint arXiv:2209.12836, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "text", + "content": "[17] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[19] Sou Kitajima, Hanna Chouchane, Jacobo Antona-Makoshi, Nobuyuki Uchida, and Jun Tajima. A nationwide impact assessment of automated driving systems on traffic safety using multiagent traffic simulations. IEEE Open Journal of Intelligent Transportation Systems, 3:302-312, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 416, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 470 + ], + "type": "text", + "content": "[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "text", + "content": "[21] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In European Conference on Computer Vision, pages 316-332. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "text", + "content": "[22] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "293" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "text", + "content": "[27] Jiaqi Ma, Edward Leslie, Amir Ghiasi, Zhitong Huang, and Yi Guo. Empirical analysis of a freeway bundled connected-and-automated vehicle application using experimental data. Journal of Transportation Engineering, Part A: Systems, 146(6):04020034, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "text", + "content": "[28] Vasileia Papathanasopoulou, Ioanna Spyropoulou, Harris Perakis, Vassilis Gikas, and Eleni Andrikopoulou. A data-driven model for pedestrian behavior classification and trajectory prediction. IEEE Open Journal of Intelligent Transportation Systems, 3:328-339, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[29] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[30] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgbd data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "text", + "content": "[31] Kelli Raboy, Jiaqi Ma, Edward Leslie, and Fang Zhou. A proof-of-concept field experiment on cooperative lane change maneuvers using a prototype connected automated vehicle testing platform. Journal of Intelligent Transportation Systems, 25(1):77-92, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "type": "text", + "content": "[32] Steven E Shladover. Opportunities and challenges in cooperative road vehicle automation. IEEE Open Journal of Intelligent Transportation Systems, 2:216-224, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[33] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1991-1999, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 525 + ], + "type": "text", + "content": "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "text", + "content": "[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[36] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. arXiv preprint arXiv:2204.01697, 2022.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 680 + ], + "type": "text", + "content": "[37] Rodolfo Valiente, Behrad Toghi, Ramtin Pedarsani, and Yaser P Fallah. Robustness and adaptability of reinforcement learning-based cooperative autonomous driving in mixed-autonomy traffic. IEEE Open Journal of Intelligent Transportation Systems, 3:397-410, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[38] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Beijbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "text", + "content": "[39] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "text", + "content": "[40] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference on Computer Vision, pages 605–621. Springer, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[41] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 315 + ], + "type": "text", + "content": "[42] Zhixin Wang and Kui Jia. Frustum convnet: Sliding frustums to aggregate local point-wise features for amodal 3d object detection. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1742-1749. IEEE, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "text", + "content": "[43] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. arXiv preprint arXiv:2209.13679, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "text", + "content": "[44] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "type": "text", + "content": "[45] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 515 + ], + "type": "text", + "content": "[46] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. arXiv preprint arXiv:2207.02202, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "text", + "content": "[47] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. arXiv preprint arXiv:2303.07601, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "text", + "content": "[48] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. arXiv preprint arXiv:2203.10638, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[49] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[50] Xinli Xu, Shaocong Dong, Lihe Ding, Jie Wang, Tingfa Xu, and Jianan Li. Fusionrcnn: Lidar-camera fusion for two-stage 3d object detection. arXiv preprint arXiv:2209.10733, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "294" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 296 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[51] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[52] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 184 + ], + "type": "text", + "content": "[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 251 + ], + "type": "text", + "content": "[54] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 287, + 296 + ], + "type": "text", + "content": "[55] Lin Zhao, Hui Zhou, Xinge Zhu, Xiao Song, Hongsheng Li, and Wenbing Tao. Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation. arXiv preprint arXiv:2108.07511, 2021. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "295" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_content_list.json b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5bef323f02fa1279dab2c75f66d2f6e546e9363 --- /dev/null +++ b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_content_list.json @@ -0,0 +1,1659 @@ +[ + { + "type": "text", + "text": "HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations", + "text_level": 1, + "bbox": [ + 86, + 130, + 880, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sadegh Aliakbarian", + "bbox": [ + 91, + 180, + 251, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fatemeh Saleh", + "bbox": [ + 281, + 180, + 401, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Collier", + "bbox": [ + 431, + 181, + 542, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pashmina Cameron", + "bbox": [ + 571, + 181, + 727, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Darren Cosker", + "bbox": [ + 756, + 181, + 875, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Microsoft Mixed Reality & AI Lab, Cambridge, UK", + "bbox": [ + 276, + 215, + 692, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating both plausible and accurate full body avatar motion is the key to the quality of immersive experiences in mixed reality scenarios. Head-Mounted Devices (HMDs) typically only provide a few input signals, such as head and hands 6-DoF. Recently, different approaches achieved impressive performance in generating full body motion given only head and hands signal. However, to the best of our knowledge, all existing approaches rely on full hand visibility. While this is the case when, e.g., using motion controllers, a considerable proportion of mixed reality experiences do not involve motion controllers and instead rely on egocentric hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view of the HMD. In this paper, we propose the first unified approach, HMD-NeMo, that addresses plausible and accurate full body motion generation even when the hands may be only partially visible. HMD-NeMo is a lightweight neural network that predicts the full body motion in an online and real-time fashion. At the heart of HMD-NeMo is the spatiotemporal encoder with novel temporally adaptable mask tokens that encourage plausible motion in the absence of hand observations. We perform extensive analysis of the impact of different components in HMD-NeMo and introduce a new state-of-the-art on AMASS dataset through our evaluation.", + "bbox": [ + 75, + 300, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mixed reality technology opens up new means of communication and interaction between people. With people at the heart of this technology, generating faithful and believable avatar motion is key to the quality of immersive experiences. Despite great advances in this area, generating full body avatar motion given HMD signal remains a challenge: in many current solutions, avatars only have upper bodies.", + "bbox": [ + 75, + 733, + 468, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior works attempted to generate full body avatar motion given sparse or partial observations, such as images [17, 4, 33], 2D joints/ keypoints [21, 5], markers [35, 18, 34, 11], and IMUs [28, 14, 32, 31]. While such observations are", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce46fd5a55b6b22f70a3dcbc66a4ca057221e6447660d2d6be9dfb1b5fef58e8.jpg", + "image_caption": [ + "Figure 1. HMD-NeMo generates full body avatar motion given HMD signals, i.e., head and hand 6-DoFs, from hand tracking signal from the HMD as well as hand motion controllers." + ], + "image_footnote": [], + "bbox": [ + 527, + 277, + 854, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "considered partial or sparse, they provide much richer input signal compared to a typical HMD's head and hand 6-DoF. More recently, great progress has been made to generate full body motion given only a HMD signal [1, 30, 22, 15], however, they all rely on the availability and full visibility of both hands. While this is the case when, e.g., using motion controllers, many mixed reality experiences do not involve motion controllers and instead rely on hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view (FoV) of the HMD sensors.", + "bbox": [ + 496, + 487, + 892, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we address this problem via HMD-NeMo (a neural motion model of human given HMD signal). Within a unified framework, HMD-NeMo generates full body motion in real time, regardless of whether hands are fully or partially observed, or not observed at all. Our approach is built upon recurrent neural networks to efficiently capture temporal information, and a transformer to capture complex relations between different components of the input signal. At the heart of our approach is the TAMT (temporally adaptable mask token) module, allowing us to handle missing hand observations.", + "bbox": [ + 496, + 641, + 893, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our contributions are: (1) The first full body avatar motion generation approach capable of generating accurate and plausible motions with full or partial hand visibility. This is a step forward for unlocking fully immersive experiences in mixed reality with fewer limitations on the hardware. (2) Temporally adaptable Mask Tokens (TAMT), a simple yet", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/345499123601787b2ca1afce709a8cb33b7218f1914b825766d56e427703ad24.jpg", + "image_caption": [ + "(a) Plane FoV", + "Figure 2. Top: Definition of FoV for different scenarios for the same pose. (a) Planar FoV, as used in [8], wherein avatar's left hand is visible while the right hand is not. (b) Fully visible, as in motion controller scenarios, wherein both hands are always visible. (c) HMD's hand tracking camera's FoV, as in hand tracking scenarios, wherein avatar's right hand is visible while the left hand is not. Bottom: Hand visibility statistics on AMASS test set." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 207, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a8bad323a803924405405444bfb6e52b44ce1efb36212b2ced535b376566369.jpg", + "image_caption": [ + "(b) Motion Controllers (MC)" + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 341, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/748ce4138b12aa35c14f7f340ad4379be726af0163d5a198fe7de3d5436aad58.jpg", + "image_caption": [ + "(c) Hand Tracking (HT)" + ], + "image_footnote": [], + "bbox": [ + 344, + 88, + 475, + 229 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/67ee74be52b5cab033b3b18faf4cd70b55f776c151978bc12bdb6c251835753b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScenarioLeft hand visibilityRight hand visibility
Motion Controllers100%100%
Hand Tracking53.45%48.94%
", + "bbox": [ + 107, + 241, + 439, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "effective strategy for handling missing hand observations in a temporally coherent way. (3) Extensive experiments and ablations of our method as well as a new state-of-the-art performance on the challenging AMASS dataset.", + "bbox": [ + 75, + 407, + 468, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 482, + 218, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Over the past few years, different solutions have been proposed to the problem of full body pose generation given sparse or partial observations, such as markers [35, 18, 34, 11], images [17, 4, 33], 2D joints/keypoints [21, 5], Inertial Measurement Units (IMUs) [28, 14, 32, 31], and HMDs [2, 1, 30, 22, 15]. Among these works, the ones utilizing wearable devices are closer to our approach and thus we discuss them here. Recently, different techniques have been proposed for human pose reconstruction using a sparse set of IMUs attached to the body [28, 14, 32, 31]. While attempts have been made to come up with a minimum number of IMUs for generating full body motion, typically an IMU sensor is used near pelvis, making the problem relatively easy compared to HMD scenarios where the root/pelvis signal is not available. Whilst external sensors, e.g., IMUs [28, 14, 32, 31] and cameras [23] are effective, they are often not as accessible as wearing a HMD. Using only a HMD is desirable from a usability point of view, but generating realistic and faithful human representations from such inputs remains technically challenging.", + "bbox": [ + 75, + 507, + 468, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given HMD signals, [2] and [8] generate full body poses. While they generate expressive poses faithful to the HMD signal, these approaches [2, 8] only predict static poses, lacking the temporal consistency required for avatar motion generation. While [8] generates poses in the world space, [2] predicts the pose relative to the root. Thus [2]", + "bbox": [ + 75, + 810, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "assumes a known root (pelvis joint) as an additional signal. The assumption of known root joint also appears in [9], wherein, unlike [2, 8], the method generates full body motion given the HMD signal with Variational Autoencoders [16]. Note that, while [9] generates temporally plausible motions, it works in an offline fashion, predicting the motion for an entire sequence only after observing the whole sequence of HMD signals.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, different techniques have been proposed to generate full body avatar locomotion (in the world coordinates) given HMD signals. In this context, [1] proposes a matching algorithm, aiming to sample closest poses from a motion capture library at sparse time-steps and interpolate between poses. While this guarantees the selection of realistic poses, the output is always limited to the utilized motion capture library. In another work, [30] uses combination of an inverse kinematic (IK) solver and a recurrent neural network to generate upper body and lower body motions, respectively. Combination of different components to solve lower and upper body separately has also been explored in [22], wherein a neural network is trained to predict the root orientation given the HMD signal, which then is used as the feature vector for a motion matching algorithm [6, 13] to generate full body. More recently, fully learning-based approaches have shown promise in generating full body avatar motion [29, 15]. In this context, [29] simulates plausible and physically feasible motions within a reinforcement learning framework and [15] uses a transformer-based approach to generate full body motion given HMD signals.", + "bbox": [ + 496, + 213, + 892, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although great progress has been made by recent approaches, they all solve the problem of motion generation given head and both hands, typically captured with motion controllers. However many mixed reality experiences do not have motion controllers available and instead rely on hand tracking from HMD mounted sensors (e.g., cameras). This introduces the challenge of hand tracking failures and partial hand visibility owing to the restricted field of views. To the best of our knowledge, despite its usability, motion generation in the presence of partial hand observation has not been well-explored. In this paper, we propose a method capable of generating high fidelity and plausible full body motion even in presence of partially visible hands.", + "bbox": [ + 496, + 531, + 892, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 744, + 671, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first define the problem, the scenarios we consider, as well as the input and the desired output representation. We then present our proposed method.", + "bbox": [ + 496, + 772, + 890, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 500, + 830, + 681, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task. The task is to generate full-body 3D human locomotion (predicting both the instantaneous pose and the global trajectory of the human) given the sparse HMD signal in", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9623", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49f341d087760e6acd468b567509dd7b1ac36f7d0c373122a7b2cfa482ce326b.jpg", + "image_caption": [ + "Figure 3. Overview of HMD-NeMo: At each time-step, the model gets as input the HMD signal (left), as in Eq. 1. Such input is the mapped to an embedding space via $f_{e}$ , acting as the input to the spatio-temporal encoder module STAE. The resulting feature representation is then utilized by two autoregressive decoders, $g\\gamma$ and $g_{\\theta}$ to predict the global trajectory as well as the pose, respectively. In case the model does not observe a hand (in HT scenario), $f_{amt}$ fills in the representation of the unobserved hand effectively." + ], + "image_footnote": [], + "bbox": [ + 158, + 98, + 810, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an online fashion1. That is, given the input signal $x_{t}$ at each time-step $t$ , the system should predict the 3D human pose and trajectory $y_{t}$ near-instantaneously. HMD-NeMo achieves this using a neural network parameterized by $\\phi$ .", + "bbox": [ + 75, + 333, + 468, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scenarios. In this paper, we consider two scenarios: Motion Controllers (MC) scenario, wherein hands are always tracked via motion controllers using constellation tracking, as illustrated in Fig. 2 (b), and Hand Tracking (HT) scenario, wherein hands are tracked via a visual hand tracking system whenever the hands are inside the FoV of the device, as illustrated in Fig. 2 (c). The FoV of the device is defined as a frustum determined by the HMD's hand tracking camera placement and parameters. In this paper, we use a similar FoV to that of the Microsoft HoloLens 2 [26]. Note that, while the MC scenario has been explored recently [1, 30, 22, 15], to the best of our knowledge, this work constitutes the first approach that tackles both MC and HT scenarios within one unified framework2. The HT scenario is particularly challenging as hands tend to be out of FoV almost $50\\%$ of the time, according to the statistics presented in the table in Fig. 2.", + "bbox": [ + 75, + 395, + 470, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input Representation. The input signal $x^{t}$ contains the head 6-DoF $x_{h} \\in \\mathbb{R}^{(6 + 3)}$ , the left hand 6-DoF $x_{l}^{t} \\in \\mathbb{R}^{(6 + 3)}$ , and the right hand 6-DoF $x_{r}^{t} \\in \\mathbb{R}^{(6 + 3)}$ , all in the world space. We use the 6D representation to represent the rotations [36]. We additionally provide the hand representations in the head space, $x_{lh}^{t} \\in \\mathbb{R}^{(6 + 3)}$ and $x_{rh}^{t} \\in \\mathbb{R}^{(6 + 3)}$ . In the HT scenario, hands may go in and out of FoV of the HMD, so we also provide HMD-NeMo with the hand visibility status for both left and right hand, $\\nu_{l}^{t}$ and $\\nu_{r}^{t}$ , as binary values, 1 being visible and 0 otherwise. Finally, for all 6-", + "bbox": [ + 75, + 652, + 468, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DoF signals, we provide the velocity of changes between two consecutive frames. Specifically, for translations we consider $vel(\\mathcal{T}^t,\\mathcal{T}^{t - 1}) = \\mathcal{T}^t -\\mathcal{T}^{t - 1}$ , where $\\mathcal{T}$ is the translation, and for rotations we consider the geodesic changes in the rotation $vel(R^{t},R^{t - 1}) = (R^{t - 1})^{-1}R^{t}$ , where $R$ is the rotation, which together constitute the velocity 6-DoF $\\dot{x}^t$ . Overall, the input to the HMD-NeMo, $x^{t}\\in \\mathbb{R}^{92}$ , can be written as", + "bbox": [ + 500, + 333, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {t} = \\left\\{x _ {h}, x _ {l}, x _ {r}, x _ {l h}, x _ {r h}, \\dot {x} _ {h}, \\dot {x} _ {l}, \\dot {x} _ {r}, \\dot {x} _ {l h}, \\dot {x} _ {r h}, \\nu_ {l}, \\nu_ {r} \\right\\} ^ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 455, + 890, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Output Representation. The output of HMD-NeMo contains the pose (including the root orientation), $\\theta^t\\in \\mathbb{R}^{J\\times 3}$ represented with axis-angle rotations for the $J$ joints in the body, and the global position in the world, $\\gamma^t\\in \\mathbb{R}^3$ represented as the root position, resulting in $y^{t}\\in \\mathbb{R}^{(J + 1)\\times 3}$", + "bbox": [ + 498, + 492, + 890, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny ^ {t} = \\{\\theta , \\gamma \\} ^ {t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 569, + 890, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sequence of $\\theta^{0:T}$ and $\\gamma^{0:T}$ then represent the avatar motion as well as its global trajectory for the period $[0,T]$ . Note that, in this paper, we may drop the time superscript $t$ for better readability and include it when necessary.", + "bbox": [ + 498, + 590, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. HMD-NeMo", + "text_level": 1, + "bbox": [ + 500, + 664, + 633, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. HMD-NeMo's pipeline is illustrated in Fig. 3. As described in Section 3.1, the model gets as input the information about the head and hands in world coordinate system, the hands expressed in the head space, as well as their velocities, as described in Eq. 1. To express hands in head space, we consider $x_{lh} = x_h^{-1}x_l$ (similarly, $x_{rh} = x_h^{-1}x_r$ ). This representation then acts as the input to an embedding layer, $f_e$ , which aims to (1) map the raw input to an embedding space and (2) handle the unobserved hands. Given the input $f_e$ , the next module, STAE, learns (a) how each representation evolves over time and (b) how different components of the input, i.e., head and hands, are correlated. Once such a rich representation is obtained, two auto-regressive decoders, $g_\\theta$ and $g_\\gamma$ , generate the body pose and the global", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Note that, similar to existing work on HMD-driven motion generation [1, 15, 9], our approach does not predict body shape parameters and only focuses on the poses.", + "bbox": [ + 75, + 814, + 468, + 851 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2Note that [8] also considers a form of partial hand visibility by defining a plane FoV, wherein, as illustrated in Fig. 2 (a), joints in front of head is are considered visible. However, this scenario does not apply to practical use cases.", + "bbox": [ + 75, + 851, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9624", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0abb9877a3e991e66204a2b2dc7af6be195d29e3b50a9970f0d8c61d60b93478.jpg", + "image_caption": [ + "Figure 4. Overview of the embedding module. Note that $x_{l}$ is only provided as an example and this module applies to all 6-DoF inputs appearing in $x^{t}$ as in Eq. 1." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 485, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "position of the avatar, respectively. At each time step, the output of STAE is used to update the mask tokens (described below) as a representation for the hand signals that may be missing in the next time-step. To aid training, we also include an auxiliary task of human pose reconstruction in SE(3), denoted by $g_{aux}$ .", + "bbox": [ + 75, + 262, + 467, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Head and hand embedding $(f_e)$ . As shown in Fig. 4, the embedding module $f_e$ gets as input the head and hands 6-DoFs and velocities and maps them to a higher-dimensional embedding space. As the range of values corresponding to the rotations is different from those of the translations, we decouple such information and embed them via separate shallow MLPs and concatenate the results back together. For instance, for the observed left hand in the world coordinate system, the embedding representation is", + "bbox": [ + 75, + 354, + 467, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {l} ^ {\\text {v i s i b l e}} = \\left[ f _ {e} ^ {\\mathrm {R}} \\left(x _ {l} [: 6 ]\\right), f _ {e} ^ {\\mathrm {T}} \\left(x _ {l} [ 6 : ]\\right), \\dot {f} _ {e} ^ {\\mathrm {R}} \\left(\\dot {x} _ {l} [: 6 ]\\right), \\dot {f} _ {e} ^ {\\mathrm {T}} \\left(\\dot {x} _ {l} [ 6 : ]\\right) \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 489, + 478, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{e}^{\\mathrm{R}}$ and $f_{e}^{\\mathrm{T}}$ are MLPs responsible for computing the rotation and translation embedding, acting on the first 6 elements (the 6D rotation representation) and the last 3 elements (the translation) of the input, respectively (similarly for $\\dot{f}_{e}^{\\mathrm{R}}$ and $\\dot{f}_{e}^{\\mathrm{T}}$ which act on velocities).", + "bbox": [ + 75, + 542, + 467, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the HT scenario, hands may not be visible to the model, hence computing such embedding representation is not possible. Thus, given the status of $\\nu_{l}$ and $\\nu_{r}$ , the embedding module decides to either compute the embedding or utilize the output of the $f_{\\text{tamt}}$ (described below), a set of temporally adaptable mask tokens, instead of a missing hand observation (denoted by $M$ in Fig. 3). As illustrated in Fig 4, the embedding of the left hand in the world coordinate system can be computed as", + "bbox": [ + 75, + 619, + 467, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {l} = \\nu_ {l} e _ {l} ^ {\\text {v i s i b l e}} + (1 - \\nu_ {l}) f _ {\\text {t a m t}} (l). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 757, + 467, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatio-temporal encoder (STAE). The output of $f_{e}$ on each component of the input stream is a non-temporal feature, computed independent of other components in the input. While an expressive representation of each component, it lacks temporal and spatial correlation information. We specifically care about these characteristics because the model is required to generate temporally coherent motion and also because the motion of one body part often impacts", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/565a806dbe0ce8def6ffc9de4e93088fae87c02e4237a23e1b4e0d93dd9262c5.jpg", + "image_caption": [ + "Figure 5. Overview of the spatio-temporal encoder module." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 890, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "or determines the motion of other body part(s). To obtain a more informative representation from the head and hands, HMD-NeMo first learns the isolated temporal features of each component of input representation and then learns how they are spatially correlated [27].", + "bbox": [ + 496, + 276, + 890, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, to learn the temporal representation of the input signal, we use gated recurrent units (GRUs) [7]. With a GRU module on top of each component in the input, the model learns how each component, e.g., head, evolves over time, independent of other components in the input. This information is compressed in the hidden state of the GRU cell, $z$ , which is then utilized to learn how different components in the input relate to each other. This is achieved by using a standard transformer encoder on the GRU hidden states, thanks to the self-attention mechanism. Temporally adaptable mask tokens (TAMTs). As discussed in Section 3.1, hands may not be visible to the model, and thus, there is no representative input signal for the $f_{e}$ module. To address this issue, in case of a missing hand observation, our model produces a feature vector, $f_{amt}$ , to represent the missing hand observation. To compute $f_{amt}$ , as illustrated in Fig. 6, we use the output of STAE for the hand observation that may be missing in the next time step as well as the output of STAE for the head. Note that head joint is the reference joint and is always available. The combination of these two features is a rich representation of the past state of the missing hand signal (both temporally and spatially); this is then used to compute the $f_{amt}$ . In order to encourage $f_{amt}$ to learn information about the missing hand observation, as illustrated by the Forecaster module in Fig. 6, we introduce a forecasting auxiliary task to forecast the state (6-DoF) of the corresponding hand in the next time-step.", + "bbox": [ + 496, + 352, + 890, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training HMD-NeMo", + "text_level": 1, + "bbox": [ + 500, + 786, + 705, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train HMD-NeMo, we rely on the availability of a motion capture dataset, represented as SMPL [19] parameters (pose, shape, and global trajectory). For each sequence, we then simulate the HMD on the subject. In case of HT, we also simulate a FoV frustum to be able to model the hand visibility status $(\\nu_{l}$ and $\\nu_{r})$ . We then train HMD-NeMo with", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a loss function of the form", + "bbox": [ + 76, + 90, + 256, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha_ {\\text {d a t a}} \\mathcal {L} _ {\\text {d a t a}} + \\alpha_ {\\text {s m o o t h}} \\mathcal {L} _ {\\text {s m o o t h}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 108, + 468, + 125 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n+ \\alpha_ {S E (3)} \\mathcal {L} _ {S E (3)} + \\alpha_ {\\text {f o r e c a s t}} \\mathcal {L} _ {\\text {f o r e c a s t}} + \\alpha_ {\\text {a u x}} \\mathcal {L} _ {\\text {a u x}}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 127, + 434, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The data loss term is the squared error between the predicted pose and trajectory and those of the ground truth,", + "bbox": [ + 76, + 147, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d a t a}} = \\sum_ {t = 1} ^ {T} | | \\hat {\\theta} ^ {t} - \\theta^ {t} | | _ {2} ^ {2} + | | \\hat {\\gamma} ^ {t} - \\gamma^ {t} | | _ {2} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 179, + 468, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that, in practice, the pose decoder has two heads, one each for predicting the body pose and the global root orientation. To further enhance the temporal smoothness, we penalize the discrepancy between the velocity of changes in the prediction to that of the ground truth", + "bbox": [ + 76, + 223, + 468, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s m o o t h}} = \\sum_ {t = 2} ^ {T} | | \\delta \\hat {\\theta} ^ {t} - \\delta \\theta^ {t} | | _ {1} + | | \\delta \\hat {\\gamma} ^ {t} - \\delta \\gamma^ {t} | | _ {1} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 300, + 468, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta \\hat{\\theta}^t = \\hat{\\theta}^t -\\hat{\\theta}^{t - 1}$ ( $\\delta \\hat{\\gamma}$ , $\\delta \\theta$ , and $\\delta \\gamma$ follow similarly). In addition to computing the reconstruction loss on the SMPL parameters, i.e., to relative joint rotations, we found it extremely useful to also utilize the reconstruction loss of each joint transformation independent of its parent, i.e., in the world space. To compute this reconstruction loss, we use the SMPL model to compute the joint transformations in SE(3) given the predicted and ground truth pose and trajectory parameters. Thus, the SE(3) reconstruction loss is", + "bbox": [ + 75, + 345, + 468, + 483 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S E (3)} = \\sum_ {t = 1} ^ {T} \\left| \\left| \\hat {P} _ {S E (3)} ^ {t} - P _ {S E (3)} ^ {t} \\right| \\right| _ {2} ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 484, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{SE(3)}$ is the body pose in SE(3). The next loss term corresponds to the forecasting auxiliary task in the TAMT module, where the goal is to minimize the distance between the predicted next hand and the ground truth next hand,", + "bbox": [ + 76, + 527, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f o r e c a s t}} = \\sum_ {t = 2} ^ {T} \\sum_ {j \\in \\{l, r \\}} \\left\\| \\hat {x} _ {j} ^ {t} - x _ {j} ^ {t} \\right\\| _ {2} ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 590, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we have our loss term for the auxiliary task, aiming to minimize the predicted full body joint transformations from STAE's features, $\\hat{P}_{aux}$ , to the ground truth body joint transformations,", + "bbox": [ + 76, + 637, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a u x} = \\sum_ {t = 1} ^ {T} \\left| \\left| \\hat {P} _ {a u x} ^ {t} - P _ {S E (3)} ^ {t} \\right| \\right| _ {2} ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 698, + 468, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is trained with $\\alpha$ s all being set to 1 in Eq. 6.", + "bbox": [ + 76, + 741, + 441, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 76, + 770, + 215, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once trained, HMD-NeMo is capable of generating high fidelity and plausible human motion given only the HMD signal. However, as is typical of learning-based approaches, the direct prediction of the neural network does not precisely match the observations i.e., the head and hands, even if it is perceptually quite close. To close this gap between the prediction and the observation, optimization can", + "bbox": [ + 76, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cae97331164c5d976d9cf8b1cfa682b4b9436eaae00271e10a2dbe3c1521afe7.jpg", + "image_caption": [ + "Figure 6. Overview of TAMT module. Note that here we illustrate TAMT of $x_{l}$ as an example, the same applies f to both hands in both head space and world space." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 895, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "be used. This adjusts the pose parameters to minimize an energy function of the form $\\mathcal{E} = \\mathcal{E}_{data} + \\mathcal{E}_{reg}$ , where $\\mathcal{E}_{data}$ is the energy term that minimizes the distance between the predicted head and hands to the observed ones, and $\\mathcal{E}_{reg}$ is additional regularization term(s). To define the data energy term, we define the residual $\\mathcal{R} = \\sum_{j\\in \\{h,l,r\\}}(x_j - \\hat{x}_j)$ , i.e., the difference between the predicted head/hand joint to that of the observation. Given $\\mathcal{R}$ , a typical, non-robust data energy term could be written as $\\mathcal{E}_{nr} = \\mathcal{R}^2$ , i.e., the L2 loss. This suits the MC scenario perfectly, where head, left hand, and right hand are always available. But this energy term may be misleading in HT scenario where hands are going into and out of FoV often, and thus leading to abrupt pose changes when hands appear back in the FoV (see supp mat for further details). To remedy this issue, we utilize a more robust [3] alternative to the data energy term,", + "bbox": [ + 496, + 258, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {r} (\\mathcal {R}, a, b, c) = b \\frac {| a - 2 |}{a} \\left(\\left(\\frac {\\left(\\frac {\\mathcal {R}}{c}\\right) ^ {2}}{| a - 2 |} + 1\\right) ^ {\\left(\\frac {a}{2}\\right)} - 1\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 501, + 890, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a$ , $b$ , and $c$ are hyper-parameters that determine the shape of the loss (see supp mat for further details). Unlike $\\mathcal{E}_{nr}$ , $\\mathcal{E}_r$ , considers large discrepancies between the prediction and observation as outliers, not penalizing the prediction strongly and thus does not push the prediction to move toward the observation. Thus abrupt changes in the arm poses are avoided and optimization stays on course despite large variation in the velocity metric (when caused by hand visibility changes). Of course, utilizing Eq. 11 adversely affects the fidelity, but a trade-off between the plausibility and fidelity can be chosen to suit the application of interest.", + "bbox": [ + 496, + 544, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that, since all observations relate to the upper body, during optimization we only optimize the upper body pose parameters and global root trajectory, while keeping the predicted lower body untouched.", + "bbox": [ + 496, + 710, + 890, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 784, + 632, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce the dataset and metrics, with implementation details in the supplementary material. We then present the experimental results and ablation studies.", + "bbox": [ + 496, + 810, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We follow the recent common practice [15, 24, 8, 9] of using a subset of AMASS [20] for training and evaluation. AMASS is a large collection of human motion se", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9626", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/274d7d51940aebbc75701ddf3ad68738e31cbf0985e4457687a0733ef5c8c44d.jpg", + "image_caption": [ + "Figure 7. Comparison to the state-of-the-art in MC scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). Last row depicts a hard example with complex body pose and motion." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 861, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quences, converted to 3D human meshes in the SMPL [19] representation, wherein every motion sequence contains information about poses $\\theta$ and the global trajectory $\\gamma$ . To synthesize the HMD scenario, we compute the global transformation matrices for the head and hands as input. In the case of hand tracking, we define a FoV for the HMD and mask out the hands whenever they are out of FoV. To make a fair comparison, for both training and evaluation, we follow the splits suggested by [15].", + "bbox": [ + 75, + 430, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To evaluate the performance of our approach as well as the competing baselines, we report the mean per-joint position error (MPJPE [cm]) and the mean per-joint velocity error (MPJVE [cm/s]). We compare our approach with recent techniques [15, 1, 30, 9] $^3$ . As these approaches do not tackle the hand tracking scenario, we compare HMD-NeMo against them for the motion controller scenario. We evaluate HMD-NeMo for hand tracking scenarios separately. In our experiments, we do not use the ground truth body shape parameters, but instead use the same default shape for all sequences. This follows the evaluation used in previous work [1, 15, 9].", + "bbox": [ + 75, + 566, + 470, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison to the state-of-the-art", + "text_level": 1, + "bbox": [ + 76, + 755, + 377, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare HMD-NeMo with existing approaches that tackle the problem of full-body motion generation given HMD signal. To the best of our knowledge, all existing approaches only tackle the MC scenario, and thus we compare them with this setting4. As demonstrated in Table 1,", + "bbox": [ + 75, + 777, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d5f299817a1ea416e32bccb2ef2d0a7b37adb77f4cf0f0369a3b2c7f3f0cd33.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison to the state-of-the-art approaches in MC scenario, where both hands are always visible." + ], + "table_body": "
MethodMPJPE ↓MPJVE ↓
FinalIK [25]18.0959.24
Ahuja et al. [1]7.83100.54
Yang et al. [30]9.0244.97
Dittadi et al. [9]6.8337.99
Jiang et al. [15]4.1829.40
HMD-NeMo (Ours)1.9024.99
", + "bbox": [ + 517, + 426, + 875, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HMD-NeMo is not only more accurate (lower MPJPE), but also generates smoother motion with joint velocities similar to that in the ground truth (lower MPJVE), introducing a new state-of-the-art on the AMASS dataset for both metrics. Note that [25] has the highest errors, as shown in Table 1 as it only optimizes the pose of the head and hands and ignores the accuracy and smoothness of the rest of the joints. Consequently, since [30] utilizes FinalIK within its framework, its performance becomes bounded by the quality of FinalIK. The learning-based approaches [15, 9], however, perform much better than [25, 1, 30], highlighting the value of data-driven methods trained on large-scale motion capture datasets. In Fig. 7, we compared HMD-NeMo with the second best performing baseline [15].", + "bbox": [ + 496, + 593, + 892, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 816, + 665, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we comprehensively evaluate different aspects of HMD-NeMo, including evaluation in HT scenario, cross-dataset evaluations, ablation studies on input signals, model architectures, and loss terms. We also ana", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3The approach [9] is modified to predict motion in world coordinates.", + "bbox": [ + 94, + 862, + 462, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4Note, we exclude very recent approaches [22, 29] since they either do not report on AMASS or the code is not publicly available.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9627", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74740f3bbe1daa84c1047f5b89d4dd93e2b893fc94d4925594536802dba4a63d.jpg", + "image_caption": [ + "(a) Walk in a semi-circle" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 305, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5204f3aa56deab36440cd4ebda798cc8953e9a372adfafa5f1677529dc2fcf96.jpg", + "image_caption": [ + "(b) Walk, then knock" + ], + "image_footnote": [], + "bbox": [ + 307, + 97, + 553, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e31aae32e2b9e2fb2f868fb598e5b76a8453d84a4d028c2605b929a5ccbe21c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 95, + 704, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/927eb99c137d9274f07cb583ee074eb4929881ac4dad9562cd08b2981b98f83b.jpg", + "image_caption": [ + "(c) Walking, then put sth down", + "(d) Walk, then sit", + "Figure 8. Qualitative results in HT scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). See the supplementary video for more qualitative results." + ], + "image_footnote": [], + "bbox": [ + 707, + 93, + 890, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b97625d06be0716a1587d89995c58f83e80393a71dd519fc70854839df4bcfe7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMask TypeMPJPE ↓MPJVE ↓
HMD-NeMoLearned & fixed5.5942.81
HMD-NeMoTAMTs (Ours)2.4831.30
", + "bbox": [ + 89, + 258, + 459, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lyze the effect optimization in both HT and MC scenarios. Additionally, we provide results of HMD-NeMo in both HT and MC scenarios for various body parts, as well as qualitative effect of optimization in the supplementary material.", + "bbox": [ + 75, + 377, + 467, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation for the hand tracking scenario. To the best of our knowledge, HMD-NeMo is the first approach to address the problem of human motion generation given partial HMD signal, applicable to HT scenario. We believe that TAMT module is the key to the success of HMD-NeMo in handling missing/partial observation, so we compare it with an alternative commonly used in the Vision Transformers [10, 12], which uses a learned set of parameters (i.e., nn.Parameters in PyTorch) to model the missing observations5. While learned parameters are fixed after training for every data point and every sequence, TAMT temporarily updates itself at each time-step given the current state of the model. Table 2 (especially MPJVE) shows the superiority of TAMT module over learned and fixed parameters in handling missing hand observations. Qualitatively, Fig. 8 illustrates how HMD-NeMo performs in HT scenario.", + "bbox": [ + 75, + 438, + 467, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-dataset evaluation. To investigate the generalizability of HMD-NeMo, we conduct a 3-fold cross-dataset evaluation as in [15], wherein the models are trained on two subsets and test on the other subset. In order to compare our approach with existing methods, we conduct this experiment in MC scenario. As shown in Table 3, HMD-NeMo outperforms existing approaches in all three datasets, by a considerable margin, highlighting its generalizability.", + "bbox": [ + 75, + 680, + 467, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluating the effect of input signal. As discussed in Section 3.1, HMD-NeMo utilizes head and hands, hands in the head space, as well as the corresponding velocities, as in Eq. 1. Table 4 summarizes the effect of each component in", + "bbox": [ + 75, + 801, + 467, + 863 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/074ed4ca99abe99705f26eb68caa4ff0cf8702d0354df704d8c57533381cf107.jpg", + "table_caption": [ + "Table 2. Effect of Temporally Adaptable Mask Tokens (TAMTs).", + "These results represent the HMD-NeMo prediction before optimization solely to evaluate the effect of the TAMT module." + ], + "table_footnote": [], + "table_body": "
MethodTest on CMUTest on BMLrubTest on HDM05
MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓
FinalIK [25]18.8256.8317.5860.6418.4362.39
Ahuja et al. [1]18.77139.1713.30134.7717.90140.61
Yang et al. [30]12.9649.9411.0060.7411.9448.26
Dittadi et al. [9]13.0451.699.6951.8010.2140.07
Jiang et al. [15]8.3735.767.0443.708.0530.85
HMD-NeMo (Ours)7.1331.236.4640.386.8027.91
", + "bbox": [ + 501, + 258, + 893, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f27dc06fb5f93d15058cb854c31d225db26a4b02d6056cb0be23354179718e0c.jpg", + "table_caption": [ + "Table 3. Results of cross-dataset evaluation between different methods. In order to compare with existing methods, we provide results in MC scenario." + ], + "table_footnote": [], + "table_body": "
Input SignalMPJPE ↓MPJVE ↓
xt = {xh, xl, xr}t4.3839.63
xt = {xh, xl, xr, xlh, xrh}t3.2138.32
xt = {xh, xl, xr, ðh, ðl, ðr}t3.5335.27
Full input signals (Eq. 1)2.4831.30
", + "bbox": [ + 519, + 412, + 874, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Evaluating the effect of various input signals in HT scenario. Note that the hand visibility status $\\nu_{l},\\nu_{r}$ are always provided to the model.", + "bbox": [ + 500, + 498, + 890, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the input signal. As shown, adding hands in the head space on top of head and hands in the world coordinates leads to better pose prediction, and thus reduces the MPJPE. Incorporating the velocities has a significant contribution to generating more temporally coherent motion, and thus reduces the MPJVE considerably. Considering all input signals, as in Eq. 1 leads to best MPJPE and MPJVE.", + "bbox": [ + 496, + 566, + 890, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluating the effect of STAE. One core component of HMD-NeMo is the spatioTemporal attention-based encoder (STAE). Here, we study the design choices of STAE in Table 5. As shown, removing the GRU component, which is responsible to learn the temporal information of various input signals, affects the MPJVE considerably. Removing the transformer encoder, which aims at learning the relation between head and hands, adversely affects the MPJPE. Unsurprisingly, removing the entire STAE module, i.e., going from input embedding layer to the decoders, is a considerably weak baseline, with poor pose quality and temporal coherency. HMD-NeMo utilizes the power and efficiency of GRU module to learn the temporal information as well as the expressively of the transformer encoder to learn how various inputs are related to each other. This design leads to", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "5This can be considered as removing the unobserved hand (similar to [2]) and replacing it with learned and fixed parameters.", + "bbox": [ + 75, + 875, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9628", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/798614f27eb0c8cf42db03878837e22e97c9e855b28dfcd397ec808ea3440eaa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
STAE ComponentsMPJPE ↓MPJVE ↓
No GRU4.3446.61
No Transformer5.5539.10
No GRU, No Transformer7.0952.26
With STAE (Ours)2.4831.30
", + "bbox": [ + 107, + 88, + 439, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f85b8bbdc99d2ec8b5439985ef482e66485cf2b34edcd4a7c75cd18033d23214.jpg", + "image_caption": [ + "Figure 9. Evaluating the effect of each loss term. Please note that each term is evaluated in isolation." + ], + "image_footnote": [], + "bbox": [ + 106, + 195, + 439, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "best performing MPJPE and MPJVE, especially in HT scenario wherein the model faces regular missing observations.", + "bbox": [ + 73, + 397, + 467, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluating the effect of each loss term. As described in Section 3.3 and shown in Eq. 6, HMD-NeMo is trained with five different loss terms. While $\\mathcal{L}_{data}$ is the essential term for training HMD-NeMo, other loss terms contribute significantly to the performance of the model. In Fig. 9, we illustrate the contribution of each loss term, in leave-one-term-out manner, on the changes in the MPJPE and MPJVE metrics. $\\mathcal{L}_{smooth}$ has a significant impact on improving the MPJVE, while $\\mathcal{L}_{aux}$ improves the MPJPE, and $\\mathcal{L}_{forecast}$ , which acts only on hands mildly improves MPJPE and MPJVE (its contribution to the total error metric is relatively small as it acts on a significantly fewer joints). $\\mathcal{L}_{SE(3)}$ makes the largest contribution to the reduction of error in both metrics. It is likely that $\\mathcal{L}_{SE(3)}$ aims to bridge that gap between the representation of the input signal (head and hand global transformation matrices) and the representation of the output pose (global root orientation and relative joint rotations).", + "bbox": [ + 73, + 428, + 467, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluating the effect of optimization. While the prediction of HMD-NeMo is very good, if the budget allows, it can be further optimized to improve accuracy (see supp mat for qualitative results). To bridge the gap between the predictions and observations, we optimize the pose prediction from HMD-NeMo, so that it matches the head and hand observations. As described in Section 3.4, in MC scenario, where both hands are always available, a simple optimization (with non-robust data energy term) loop can be used. The effect of such optimization in MC scenario is provided", + "bbox": [ + 75, + 699, + 467, + 851 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96544b28beddee228765a8da567a9e07e1a49ffaf47001539637163eb42074b9.jpg", + "table_caption": [ + "Table 5. Evaluating the effect of STAE module in HT scenario." + ], + "table_footnote": [], + "table_body": "
ConfigurationMPJPE ↓MPJVE ↓
Without optimization2.0726.07
With optimization1.9024.99
", + "bbox": [ + 545, + 88, + 846, + 137 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b484a2857b7756c436593f6616f82c2eeaea4ed4babe204353be949428a65fc.jpg", + "table_caption": [ + "Table 6. Effect of optimization in MC scenario, where both hands are always visible." + ], + "table_footnote": [], + "table_body": "
ConfigurationPriorityMPJPE ↓MPJVE ↓
No OptimizationPlausibility2.4831.30
Opt. with EnrFidelity2.3233.51
Opt. with ErBoth2.3731.33
", + "bbox": [ + 514, + 178, + 877, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Configuring HT scenario. Our model offers a range of choices to provide the best possible end-user experience despite incomplete observations.", + "bbox": [ + 500, + 241, + 890, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Table 6. However, considering the same strategy for the HT scenario, where we may partially observe hands, this may not be optimal as we lose plausibility in the generated motions (captured by the MPJVE metric). As described in Section 3.4, depending on the scenario and experience requirements, one may choose to (1) avoid optimizing the predictions if plausibility is the highest priority (first row of Table 7), (2) use non-robust energy term if fidelity is the highest priority (second row of Table 7), or (3) use a robust energy term as a trade-off between fidelity and plausibility (third row of Table 7).", + "bbox": [ + 496, + 306, + 890, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance analysis. With 5.3M parameters, at inference time, HMD-NeMo requires only $4.4\\mathrm{ms}$ to generate a pose given a HMD signal on a typical laptop CPU. On a NVIDIA Tesla P100 GPU, our model runs at 265 fps, with 1 iteration of optimization costs 3ms per frame (which could be further improved with more optimized implementation). Such performance makes HMD-NeMo a potential solution for HMD-driven avatar animation in immersive environments.", + "bbox": [ + 496, + 473, + 890, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 607, + 617, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present HMD-NeMo, a unified approach to generate full-body avatar motion in both motion controller and hand tracking scenarios. To handle the unobserved hands in a temporally coherent and plausible manner, we introduce TAMT module. It is worth noting that, in this paper, we considered one major reason for partial hand visibility, i.e., hands appearing out of the FoV, however, in practice, hands may not be visible to the hand tracking camera due to failure in tracking and occlusion by another object or by another body part. While such cases are not considered in the data augmentation in our paper, they certainly can be taken into account and TAMT can be used to fill such gaps with no additional modification; this is left for future exploration. We provide extensive analyses on different components of HMD-NeMo, and shed light on various choices for optimization on top of neural network's predictions when it comes to production priorities (plausibility versus fidelity).", + "bbox": [ + 496, + 633, + 890, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "6This experiment only evaluates the contribution of each loss term independently. Evaluating the combination of loss terms remains for future investigations.", + "bbox": [ + 75, + 862, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Karan Ahuja, Eyal Ofek, Mar Gonzalez-Franco, Christian Holz, and Andrew D Wilson. Coolmoves: User motion accentuation in virtual reality. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 5(2):1-23, 2021.", + "[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. FLAG: Flow-based 3D avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022.", + "[3] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019.", + "[4] Benjamin Biggs, Sébastien Ehrhadt, Hanbyul Joo, Benjamin Graham, Andrea Vedaldi, and David Novotny. 3D multibodies: Fitting sets of plausible 3D human models to ambiguous image data. 2020.", + "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016.", + "[6] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015.", + "[7] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014.", + "[8] Vasileios Choutas, Federica Bogo, Jingjing Shen, and Julien Valentin. Learning to fit morphable models. arXiv preprint arXiv:2111.14824, 2021.", + "[9] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating SMPL poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021.", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[11] Nima Ghorbani and Michael J Black. Soma: Solving optical marker-based mocap automatically. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11117-11126, 2021.", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + ], + "bbox": [ + 78, + 116, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020.", + "[14] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018.", + "[15] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. AvatarPoser: Articulated full-body pose tracking from sparse motion sensing. In Proceedings of European Conference on Computer Vision. Springer, 2022.", + "[16] Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. International Conference on Learning Representations, ICLR, 2014.", + "[17] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11605-11614, 2021.", + "[18] Matthew Loper, Naureen Mahmood, and Michael J Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics (TOG), 33(6):1-13, 2014.", + "[19] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015.", + "[20] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, Oct. 2019.", + "[21] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10975-10985, 2019.", + "[22] Jose Luis Ponton, Haoran Yun, Carlos Andujar, and Nuria Pelechano. Combining Motion Matching and Orientation Prediction to Animate Avatars for Consumer-Grade VR Devices. Computer Graphics Forum, 2022.", + "[23] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020.", + "[24] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021.", + "[25] Unity Asset Store. Final IK. https://assetstoreunity.com/packages/tools/animation/final-ik-14290, 2021.", + "[26] Dorin Ungureanu, Federica Bogo, Silvano Galliani, Pooja Sama, Casey Meekhof, Jan Stuhmer, Thomas J Cashman," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9630", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bugra Tekin, Johannes L Schonberger, Pawel Olszta, et al. Hololens 2 research mode as a tool for computer vision research. arXiv preprint arXiv:2008.11239, 2020.", + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[28] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3D human pose estimation from sparse IMUs. In Computer graphics forum, volume 36, pages 349-360. Wiley Online Library, 2017.", + "[29] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. arXiv preprint arXiv:2209.09391, 2022.", + "[30] Dongseok Yang, Doyeon Kim, and Sung-Hee Lee. LoBSTr: Real-time lower-body pose prediction from sparse upper-body tracking signals. In Computer Graphics Forum, volume 40, pages 265–275. Wiley Online Library, 2021.", + "[31] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical Inertial Poser (PIP): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13167-13178, 2022.", + "[32] Xinyu Yi, Yuxiao Zhou, and Feng Xu. TransPose: Real-time 3D human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021.", + "[33] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision, pages 465-481. Springer, 2020.", + "[34] Mihai Zanfir, Andrei Zanfir, Eduard Gabriel Bazavan, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. THUNDR: Transformer-based 3D human reconstruction with markers. arXiv preprint arXiv:2106.09336, 2021.", + "[35] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3D bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021.", + "[36] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019." + ], + "bbox": [ + 78, + 90, + 468, + 767 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9631", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_model.json b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fe934cccfb149e35a85e8d57085255f868b03f4d --- /dev/null +++ b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.131, + 0.882, + 0.152 + ], + "angle": 0, + "content": "HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.181, + 0.253, + 0.199 + ], + "angle": 0, + "content": "Sadegh Aliakbarian" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.181, + 0.402, + 0.198 + ], + "angle": 0, + "content": "Fatemeh Saleh" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.182, + 0.543, + 0.197 + ], + "angle": 0, + "content": "David Collier" + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.182, + 0.728, + 0.198 + ], + "angle": 0, + "content": "Pashmina Cameron" + }, + { + "type": "text", + "bbox": [ + 0.758, + 0.182, + 0.877, + 0.197 + ], + "angle": 0, + "content": "Darren Cosker" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.216, + 0.694, + 0.234 + ], + "angle": 0, + "content": "Microsoft Mixed Reality & AI Lab, Cambridge, UK" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Generating both plausible and accurate full body avatar motion is the key to the quality of immersive experiences in mixed reality scenarios. Head-Mounted Devices (HMDs) typically only provide a few input signals, such as head and hands 6-DoF. Recently, different approaches achieved impressive performance in generating full body motion given only head and hands signal. However, to the best of our knowledge, all existing approaches rely on full hand visibility. While this is the case when, e.g., using motion controllers, a considerable proportion of mixed reality experiences do not involve motion controllers and instead rely on egocentric hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view of the HMD. In this paper, we propose the first unified approach, HMD-NeMo, that addresses plausible and accurate full body motion generation even when the hands may be only partially visible. HMD-NeMo is a lightweight neural network that predicts the full body motion in an online and real-time fashion. At the heart of HMD-NeMo is the spatiotemporal encoder with novel temporally adaptable mask tokens that encourage plausible motion in the absence of hand observations. We perform extensive analysis of the impact of different components in HMD-NeMo and introduce a new state-of-the-art on AMASS dataset through our evaluation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.709, + 0.21, + 0.724 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Mixed reality technology opens up new means of communication and interaction between people. With people at the heart of this technology, generating faithful and believable avatar motion is key to the quality of immersive experiences. Despite great advances in this area, generating full body avatar motion given HMD signal remains a challenge: in many current solutions, avatars only have upper bodies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Prior works attempted to generate full body avatar motion given sparse or partial observations, such as images [17, 4, 33], 2D joints/ keypoints [21, 5], markers [35, 18, 34, 11], and IMUs [28, 14, 32, 31]. While such observations are" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.279, + 0.856, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.42, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Figure 1. HMD-NeMo generates full body avatar motion given HMD signals, i.e., head and hand 6-DoFs, from hand tracking signal from the HMD as well as hand motion controllers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.893, + 0.639 + ], + "angle": 0, + "content": "considered partial or sparse, they provide much richer input signal compared to a typical HMD's head and hand 6-DoF. More recently, great progress has been made to generate full body motion given only a HMD signal [1, 30, 22, 15], however, they all rely on the availability and full visibility of both hands. While this is the case when, e.g., using motion controllers, many mixed reality experiences do not involve motion controllers and instead rely on hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view (FoV) of the HMD sensors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.808 + ], + "angle": 0, + "content": "In this paper, we address this problem via HMD-NeMo (a neural motion model of human given HMD signal). Within a unified framework, HMD-NeMo generates full body motion in real time, regardless of whether hands are fully or partially observed, or not observed at all. Our approach is built upon recurrent neural networks to efficiently capture temporal information, and a transformer to capture complex relations between different components of the input signal. At the heart of our approach is the TAMT (temporally adaptable mask token) module, allowing us to handle missing hand observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Our contributions are: (1) The first full body avatar motion generation approach capable of generating accurate and plausible motions with full or partial hand visibility. This is a step forward for unlocking fully immersive experiences in mixed reality with fewer limitations on the hardware. (2) Temporally adaptable Mask Tokens (TAMT), a simple yet" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.209, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.232, + 0.178, + 0.241 + ], + "angle": 0, + "content": "(a) Plane FoV" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.343, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.231, + 0.345, + 0.242 + ], + "angle": 0, + "content": "(b) Motion Controllers (MC)" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.089, + 0.476, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.231, + 0.466, + 0.242 + ], + "angle": 0, + "content": "(c) Hand Tracking (HT)" + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.242, + 0.441, + 0.285 + ], + "angle": 0, + "content": "
ScenarioLeft hand visibilityRight hand visibility
Motion Controllers100%100%
Hand Tracking53.45%48.94%
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.384 + ], + "angle": 0, + "content": "Figure 2. Top: Definition of FoV for different scenarios for the same pose. (a) Planar FoV, as used in [8], wherein avatar's left hand is visible while the right hand is not. (b) Fully visible, as in motion controller scenarios, wherein both hands are always visible. (c) HMD's hand tracking camera's FoV, as in hand tracking scenarios, wherein avatar's right hand is visible while the left hand is not. Bottom: Hand visibility statistics on AMASS test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.47, + 0.471 + ], + "angle": 0, + "content": "effective strategy for handling missing hand observations in a temporally coherent way. (3) Extensive experiments and ablations of our method as well as a new state-of-the-art performance on the challenging AMASS dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.22, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Over the past few years, different solutions have been proposed to the problem of full body pose generation given sparse or partial observations, such as markers [35, 18, 34, 11], images [17, 4, 33], 2D joints/keypoints [21, 5], Inertial Measurement Units (IMUs) [28, 14, 32, 31], and HMDs [2, 1, 30, 22, 15]. Among these works, the ones utilizing wearable devices are closer to our approach and thus we discuss them here. Recently, different techniques have been proposed for human pose reconstruction using a sparse set of IMUs attached to the body [28, 14, 32, 31]. While attempts have been made to come up with a minimum number of IMUs for generating full body motion, typically an IMU sensor is used near pelvis, making the problem relatively easy compared to HMD scenarios where the root/pelvis signal is not available. Whilst external sensors, e.g., IMUs [28, 14, 32, 31] and cameras [23] are effective, they are often not as accessible as wearing a HMD. Using only a HMD is desirable from a usability point of view, but generating realistic and faithful human representations from such inputs remains technically challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Given HMD signals, [2] and [8] generate full body poses. While they generate expressive poses faithful to the HMD signal, these approaches [2, 8] only predict static poses, lacking the temporal consistency required for avatar motion generation. While [8] generates poses in the world space, [2] predicts the pose relative to the root. Thus [2]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "assumes a known root (pelvis joint) as an additional signal. The assumption of known root joint also appears in [9], wherein, unlike [2, 8], the method generates full body motion given the HMD signal with Variational Autoencoders [16]. Note that, while [9] generates temporally plausible motions, it works in an offline fashion, predicting the motion for an entire sequence only after observing the whole sequence of HMD signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.893, + 0.53 + ], + "angle": 0, + "content": "Recently, different techniques have been proposed to generate full body avatar locomotion (in the world coordinates) given HMD signals. In this context, [1] proposes a matching algorithm, aiming to sample closest poses from a motion capture library at sparse time-steps and interpolate between poses. While this guarantees the selection of realistic poses, the output is always limited to the utilized motion capture library. In another work, [30] uses combination of an inverse kinematic (IK) solver and a recurrent neural network to generate upper body and lower body motions, respectively. Combination of different components to solve lower and upper body separately has also been explored in [22], wherein a neural network is trained to predict the root orientation given the HMD signal, which then is used as the feature vector for a motion matching algorithm [6, 13] to generate full body. More recently, fully learning-based approaches have shown promise in generating full body avatar motion [29, 15]. In this context, [29] simulates plausible and physically feasible motions within a reinforcement learning framework and [15] uses a transformer-based approach to generate full body motion given HMD signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.893, + 0.729 + ], + "angle": 0, + "content": "Although great progress has been made by recent approaches, they all solve the problem of motion generation given head and both hands, typically captured with motion controllers. However many mixed reality experiences do not have motion controllers available and instead rely on hand tracking from HMD mounted sensors (e.g., cameras). This introduces the challenge of hand tracking failures and partial hand visibility owing to the restricted field of views. To the best of our knowledge, despite its usability, motion generation in the presence of partial hand observation has not been well-explored. In this paper, we propose a method capable of generating high fidelity and plausible full body motion even in presence of partially visible hands." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.746, + 0.673, + 0.763 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.818 + ], + "angle": 0, + "content": "In this section, we first define the problem, the scenarios we consider, as well as the input and the desired output representation. We then present our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.682, + 0.846 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Task. The task is to generate full-body 3D human locomotion (predicting both the instantaneous pose and the global trajectory of the human) given the sparse HMD signal in" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9623" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.099, + 0.812, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Overview of HMD-NeMo: At each time-step, the model gets as input the HMD signal (left), as in Eq. 1. Such input is the mapped to an embedding space via \\( f_{e} \\), acting as the input to the spatio-temporal encoder module STAE. The resulting feature representation is then utilized by two autoregressive decoders, \\( g\\gamma \\) and \\( g_{\\theta} \\) to predict the global trajectory as well as the pose, respectively. In case the model does not observe a hand (in HT scenario), \\( f_{amt} \\) fills in the representation of the unobserved hand effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.47, + 0.395 + ], + "angle": 0, + "content": "an online fashion1. That is, given the input signal \\( x_{t} \\) at each time-step \\( t \\), the system should predict the 3D human pose and trajectory \\( y_{t} \\) near-instantaneously. HMD-NeMo achieves this using a neural network parameterized by \\( \\phi \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.471, + 0.652 + ], + "angle": 0, + "content": "Scenarios. In this paper, we consider two scenarios: Motion Controllers (MC) scenario, wherein hands are always tracked via motion controllers using constellation tracking, as illustrated in Fig. 2 (b), and Hand Tracking (HT) scenario, wherein hands are tracked via a visual hand tracking system whenever the hands are inside the FoV of the device, as illustrated in Fig. 2 (c). The FoV of the device is defined as a frustum determined by the HMD's hand tracking camera placement and parameters. In this paper, we use a similar FoV to that of the Microsoft HoloLens 2 [26]. Note that, while the MC scenario has been explored recently [1, 30, 22, 15], to the best of our knowledge, this work constitutes the first approach that tackles both MC and HT scenarios within one unified framework2. The HT scenario is particularly challenging as hands tend to be out of FoV almost \\(50\\%\\) of the time, according to the statistics presented in the table in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.804 + ], + "angle": 0, + "content": "Input Representation. The input signal \\( x^{t} \\) contains the head 6-DoF \\( x_{h} \\in \\mathbb{R}^{(6 + 3)} \\), the left hand 6-DoF \\( x_{l}^{t} \\in \\mathbb{R}^{(6 + 3)} \\), and the right hand 6-DoF \\( x_{r}^{t} \\in \\mathbb{R}^{(6 + 3)} \\), all in the world space. We use the 6D representation to represent the rotations [36]. We additionally provide the hand representations in the head space, \\( x_{lh}^{t} \\in \\mathbb{R}^{(6 + 3)} \\) and \\( x_{rh}^{t} \\in \\mathbb{R}^{(6 + 3)} \\). In the HT scenario, hands may go in and out of FoV of the HMD, so we also provide HMD-NeMo with the hand visibility status for both left and right hand, \\( \\nu_{l}^{t} \\) and \\( \\nu_{r}^{t} \\), as binary values, 1 being visible and 0 otherwise. Finally, for all 6-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.334, + 0.892, + 0.453 + ], + "angle": 0, + "content": "DoF signals, we provide the velocity of changes between two consecutive frames. Specifically, for translations we consider \\( vel(\\mathcal{T}^t,\\mathcal{T}^{t - 1}) = \\mathcal{T}^t -\\mathcal{T}^{t - 1} \\), where \\( \\mathcal{T} \\) is the translation, and for rotations we consider the geodesic changes in the rotation \\( vel(R^{t},R^{t - 1}) = (R^{t - 1})^{-1}R^{t} \\), where \\( R \\) is the rotation, which together constitute the velocity 6-DoF \\( \\dot{x}^t \\). Overall, the input to the HMD-NeMo, \\( x^{t}\\in \\mathbb{R}^{92} \\), can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.457, + 0.892, + 0.488 + ], + "angle": 0, + "content": "\\[\nx ^ {t} = \\left\\{x _ {h}, x _ {l}, x _ {r}, x _ {l h}, x _ {r h}, \\dot {x} _ {h}, \\dot {x} _ {l}, \\dot {x} _ {r}, \\dot {x} _ {l h}, \\dot {x} _ {r h}, \\nu_ {l}, \\nu_ {r} \\right\\} ^ {t} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Output Representation. The output of HMD-NeMo contains the pose (including the root orientation), \\(\\theta^t\\in \\mathbb{R}^{J\\times 3}\\) represented with axis-angle rotations for the \\(J\\) joints in the body, and the global position in the world, \\(\\gamma^t\\in \\mathbb{R}^3\\) represented as the root position, resulting in \\(y^{t}\\in \\mathbb{R}^{(J + 1)\\times 3}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.57, + 0.891, + 0.587 + ], + "angle": 0, + "content": "\\[\ny ^ {t} = \\{\\theta , \\gamma \\} ^ {t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.654 + ], + "angle": 0, + "content": "The sequence of \\(\\theta^{0:T}\\) and \\(\\gamma^{0:T}\\) then represent the avatar motion as well as its global trajectory for the period \\([0,T]\\). Note that, in this paper, we may drop the time superscript \\(t\\) for better readability and include it when necessary." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "3.2. HMD-NeMo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overview. HMD-NeMo's pipeline is illustrated in Fig. 3. As described in Section 3.1, the model gets as input the information about the head and hands in world coordinate system, the hands expressed in the head space, as well as their velocities, as described in Eq. 1. To express hands in head space, we consider \\( x_{lh} = x_h^{-1}x_l \\) (similarly, \\( x_{rh} = x_h^{-1}x_r \\)). This representation then acts as the input to an embedding layer, \\( f_e \\), which aims to (1) map the raw input to an embedding space and (2) handle the unobserved hands. Given the input \\( f_e \\), the next module, STAE, learns (a) how each representation evolves over time and (b) how different components of the input, i.e., head and hands, are correlated. Once such a rich representation is obtained, two auto-regressive decoders, \\( g_\\theta \\) and \\( g_\\gamma \\), generate the body pose and the global" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.815, + 0.47, + 0.852 + ], + "angle": 0, + "content": "1Note that, similar to existing work on HMD-driven motion generation [1, 15, 9], our approach does not predict body shape parameters and only focuses on the poses." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.852, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2Note that [8] also considers a form of partial hand visibility by defining a plane FoV, wherein, as illustrated in Fig. 2 (a), joints in front of head is are considered visible. However, this scenario does not apply to practical use cases." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.815, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9624" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.486, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.194, + 0.471, + 0.237 + ], + "angle": 0, + "content": "Figure 4. Overview of the embedding module. Note that \\( x_{l} \\) is only provided as an example and this module applies to all 6-DoF inputs appearing in \\( x^{t} \\) as in Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.468, + 0.354 + ], + "angle": 0, + "content": "position of the avatar, respectively. At each time step, the output of STAE is used to update the mask tokens (described below) as a representation for the hand signals that may be missing in the next time-step. To aid training, we also include an auxiliary task of human pose reconstruction in SE(3), denoted by \\( g_{aux} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.468, + 0.49 + ], + "angle": 0, + "content": "Head and hand embedding \\((f_e)\\). As shown in Fig. 4, the embedding module \\(f_e\\) gets as input the head and hands 6-DoFs and velocities and maps them to a higher-dimensional embedding space. As the range of values corresponding to the rotations is different from those of the translations, we decouple such information and embed them via separate shallow MLPs and concatenate the results back together. For instance, for the observed left hand in the world coordinate system, the embedding representation is" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.491, + 0.48, + 0.538 + ], + "angle": 0, + "content": "\\[\ne _ {l} ^ {\\text {v i s i b l e}} = \\left[ f _ {e} ^ {\\mathrm {R}} \\left(x _ {l} [: 6 ]\\right), f _ {e} ^ {\\mathrm {T}} \\left(x _ {l} [ 6 : ]\\right), \\dot {f} _ {e} ^ {\\mathrm {R}} \\left(\\dot {x} _ {l} [: 6 ]\\right), \\dot {f} _ {e} ^ {\\mathrm {T}} \\left(\\dot {x} _ {l} [ 6 : ]\\right) \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.468, + 0.621 + ], + "angle": 0, + "content": "where \\( f_{e}^{\\mathrm{R}} \\) and \\( f_{e}^{\\mathrm{T}} \\) are MLPs responsible for computing the rotation and translation embedding, acting on the first 6 elements (the 6D rotation representation) and the last 3 elements (the translation) of the input, respectively (similarly for \\( \\dot{f}_{e}^{\\mathrm{R}} \\) and \\( \\dot{f}_{e}^{\\mathrm{T}} \\) which act on velocities)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.468, + 0.756 + ], + "angle": 0, + "content": "In the HT scenario, hands may not be visible to the model, hence computing such embedding representation is not possible. Thus, given the status of \\(\\nu_{l}\\) and \\(\\nu_{r}\\), the embedding module decides to either compute the embedding or utilize the output of the \\(f_{\\text{tamt}}\\) (described below), a set of temporally adaptable mask tokens, instead of a missing hand observation (denoted by \\(M\\) in Fig. 3). As illustrated in Fig 4, the embedding of the left hand in the world coordinate system can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.758, + 0.468, + 0.775 + ], + "angle": 0, + "content": "\\[\ne _ {l} = \\nu_ {l} e _ {l} ^ {\\text {v i s i b l e}} + (1 - \\nu_ {l}) f _ {\\text {t a m t}} (l). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Spatio-temporal encoder (STAE). The output of \\( f_{e} \\) on each component of the input stream is a non-temporal feature, computed independent of other components in the input. While an expressive representation of each component, it lacks temporal and spatial correlation information. We specifically care about these characteristics because the model is required to generate temporally coherent motion and also because the motion of one body part often impacts" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.892, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.236, + 0.853, + 0.25 + ], + "angle": 0, + "content": "Figure 5. Overview of the spatio-temporal encoder module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.352 + ], + "angle": 0, + "content": "or determines the motion of other body part(s). To obtain a more informative representation from the head and hands, HMD-NeMo first learns the isolated temporal features of each component of input representation and then learns how they are spatially correlated [27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.777 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, to learn the temporal representation of the input signal, we use gated recurrent units (GRUs) [7]. With a GRU module on top of each component in the input, the model learns how each component, e.g., head, evolves over time, independent of other components in the input. This information is compressed in the hidden state of the GRU cell, \\( z \\), which is then utilized to learn how different components in the input relate to each other. This is achieved by using a standard transformer encoder on the GRU hidden states, thanks to the self-attention mechanism. Temporally adaptable mask tokens (TAMTs). As discussed in Section 3.1, hands may not be visible to the model, and thus, there is no representative input signal for the \\( f_{e} \\) module. To address this issue, in case of a missing hand observation, our model produces a feature vector, \\( f_{amt} \\), to represent the missing hand observation. To compute \\( f_{amt} \\), as illustrated in Fig. 6, we use the output of STAE for the hand observation that may be missing in the next time step as well as the output of STAE for the head. Note that head joint is the reference joint and is always available. The combination of these two features is a rich representation of the past state of the missing hand signal (both temporally and spatially); this is then used to compute the \\( f_{amt} \\). In order to encourage \\( f_{amt} \\) to learn information about the missing hand observation, as illustrated by the Forecaster module in Fig. 6, we introduce a forecasting auxiliary task to forecast the state (6-DoF) of the corresponding hand in the next time-step." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.706, + 0.803 + ], + "angle": 0, + "content": "3.3. Training HMD-NeMo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To train HMD-NeMo, we rely on the availability of a motion capture dataset, represented as SMPL [19] parameters (pose, shape, and global trajectory). For each sequence, we then simulate the HMD on the subject. In case of HT, we also simulate a FoV frustum to be able to model the hand visibility status \\((\\nu_{l}\\) and \\(\\nu_{r})\\). We then train HMD-NeMo with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9625" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.258, + 0.105 + ], + "angle": 0, + "content": "a loss function of the form" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.109, + 0.469, + 0.125 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha_ {\\text {d a t a}} \\mathcal {L} _ {\\text {d a t a}} + \\alpha_ {\\text {s m o o t h}} \\mathcal {L} _ {\\text {s m o o t h}} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.128, + 0.436, + 0.145 + ], + "angle": 0, + "content": "\\[\n+ \\alpha_ {S E (3)} \\mathcal {L} _ {S E (3)} + \\alpha_ {\\text {f o r e c a s t}} \\mathcal {L} _ {\\text {f o r e c a s t}} + \\alpha_ {\\text {a u x}} \\mathcal {L} _ {\\text {a u x}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.148, + 0.47, + 0.178 + ], + "angle": 0, + "content": "The data loss term is the squared error between the predicted pose and trajectory and those of the ground truth," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.18, + 0.469, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d a t a}} = \\sum_ {t = 1} ^ {T} | | \\hat {\\theta} ^ {t} - \\theta^ {t} | | _ {2} ^ {2} + | | \\hat {\\gamma} ^ {t} - \\gamma^ {t} | | _ {2} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.224, + 0.47, + 0.299 + ], + "angle": 0, + "content": "Note that, in practice, the pose decoder has two heads, one each for predicting the body pose and the global root orientation. To further enhance the temporal smoothness, we penalize the discrepancy between the velocity of changes in the prediction to that of the ground truth" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.301, + 0.469, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s m o o t h}} = \\sum_ {t = 2} ^ {T} | | \\delta \\hat {\\theta} ^ {t} - \\delta \\theta^ {t} | | _ {1} + | | \\delta \\hat {\\gamma} ^ {t} - \\delta \\gamma^ {t} | | _ {1} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.47, + 0.484 + ], + "angle": 0, + "content": "where \\(\\delta \\hat{\\theta}^t = \\hat{\\theta}^t -\\hat{\\theta}^{t - 1}\\) (\\(\\delta \\hat{\\gamma}\\), \\(\\delta \\theta\\), and \\(\\delta \\gamma\\) follow similarly). In addition to computing the reconstruction loss on the SMPL parameters, i.e., to relative joint rotations, we found it extremely useful to also utilize the reconstruction loss of each joint transformation independent of its parent, i.e., in the world space. To compute this reconstruction loss, we use the SMPL model to compute the joint transformations in SE(3) given the predicted and ground truth pose and trajectory parameters. Thus, the SE(3) reconstruction loss is" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.485, + 0.469, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S E (3)} = \\sum_ {t = 1} ^ {T} \\left| \\left| \\hat {P} _ {S E (3)} ^ {t} - P _ {S E (3)} ^ {t} \\right| \\right| _ {2} ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.529, + 0.47, + 0.589 + ], + "angle": 0, + "content": "where \\( P_{SE(3)} \\) is the body pose in SE(3). The next loss term corresponds to the forecasting auxiliary task in the TAMT module, where the goal is to minimize the distance between the predicted next hand and the ground truth next hand," + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.592, + 0.469, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f o r e c a s t}} = \\sum_ {t = 2} ^ {T} \\sum_ {j \\in \\{l, r \\}} \\left\\| \\hat {x} _ {j} ^ {t} - x _ {j} ^ {t} \\right\\| _ {2} ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.638, + 0.47, + 0.697 + ], + "angle": 0, + "content": "Finally, we have our loss term for the auxiliary task, aiming to minimize the predicted full body joint transformations from STAE's features, \\(\\hat{P}_{aux}\\), to the ground truth body joint transformations," + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.699, + 0.469, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a u x} = \\sum_ {t = 1} ^ {T} \\left| \\left| \\hat {P} _ {a u x} ^ {t} - P _ {S E (3)} ^ {t} \\right| \\right| _ {2} ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.742, + 0.442, + 0.758 + ], + "angle": 0, + "content": "Our model is trained with \\(\\alpha\\)s all being set to 1 in Eq. 6." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.217, + 0.787 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Once trained, HMD-NeMo is capable of generating high fidelity and plausible human motion given only the HMD signal. However, as is typical of learning-based approaches, the direct prediction of the neural network does not precisely match the observations i.e., the head and hands, even if it is perceptually quite close. To close this gap between the prediction and the observation, optimization can" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.09, + 0.897, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.192, + 0.894, + 0.234 + ], + "angle": 0, + "content": "Figure 6. Overview of TAMT module. Note that here we illustrate TAMT of \\( x_{l} \\) as an example, the same applies f to both hands in both head space and world space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.501 + ], + "angle": 0, + "content": "be used. This adjusts the pose parameters to minimize an energy function of the form \\(\\mathcal{E} = \\mathcal{E}_{data} + \\mathcal{E}_{reg}\\), where \\(\\mathcal{E}_{data}\\) is the energy term that minimizes the distance between the predicted head and hands to the observed ones, and \\(\\mathcal{E}_{reg}\\) is additional regularization term(s). To define the data energy term, we define the residual \\(\\mathcal{R} = \\sum_{j\\in \\{h,l,r\\}}(x_j - \\hat{x}_j)\\), i.e., the difference between the predicted head/hand joint to that of the observation. Given \\(\\mathcal{R}\\), a typical, non-robust data energy term could be written as \\(\\mathcal{E}_{nr} = \\mathcal{R}^2\\), i.e., the L2 loss. This suits the MC scenario perfectly, where head, left hand, and right hand are always available. But this energy term may be misleading in HT scenario where hands are going into and out of FoV often, and thus leading to abrupt pose changes when hands appear back in the FoV (see supp mat for further details). To remedy this issue, we utilize a more robust [3] alternative to the data energy term," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.502, + 0.892, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {r} (\\mathcal {R}, a, b, c) = b \\frac {| a - 2 |}{a} \\left(\\left(\\frac {\\left(\\frac {\\mathcal {R}}{c}\\right) ^ {2}}{| a - 2 |} + 1\\right) ^ {\\left(\\frac {a}{2}\\right)} - 1\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.711 + ], + "angle": 0, + "content": "where \\(a\\), \\(b\\), and \\(c\\) are hyper-parameters that determine the shape of the loss (see supp mat for further details). Unlike \\(\\mathcal{E}_{nr}\\), \\(\\mathcal{E}_r\\), considers large discrepancies between the prediction and observation as outliers, not penalizing the prediction strongly and thus does not push the prediction to move toward the observation. Thus abrupt changes in the arm poses are avoided and optimization stays on course despite large variation in the velocity metric (when caused by hand visibility changes). Of course, utilizing Eq. 11 adversely affects the fidelity, but a trade-off between the plausibility and fidelity can be chosen to suit the application of interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Note that, since all observations relate to the upper body, during optimization we only optimize the upper body pose parameters and global root trajectory, while keeping the predicted lower body untouched." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.633, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.855 + ], + "angle": 0, + "content": "In this section, we introduce the dataset and metrics, with implementation details in the supplementary material. We then present the experimental results and ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset. We follow the recent common practice [15, 24, 8, 9] of using a subset of AMASS [20] for training and evaluation. AMASS is a large collection of human motion se" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9626" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.089, + 0.862, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.894, + 0.407 + ], + "angle": 0, + "content": "Figure 7. Comparison to the state-of-the-art in MC scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). Last row depicts a hard example with complex body pose and motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.567 + ], + "angle": 0, + "content": "quences, converted to 3D human meshes in the SMPL [19] representation, wherein every motion sequence contains information about poses \\(\\theta\\) and the global trajectory \\(\\gamma\\). To synthesize the HMD scenario, we compute the global transformation matrices for the head and hands as input. In the case of hand tracking, we define a FoV for the HMD and mask out the hands whenever they are out of FoV. To make a fair comparison, for both training and evaluation, we follow the splits suggested by [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Metrics. To evaluate the performance of our approach as well as the competing baselines, we report the mean per-joint position error (MPJPE [cm]) and the mean per-joint velocity error (MPJVE [cm/s]). We compare our approach with recent techniques [15, 1, 30, 9]\\(^3\\). As these approaches do not tackle the hand tracking scenario, we compare HMD-NeMo against them for the motion controller scenario. We evaluate HMD-NeMo for hand tracking scenarios separately. In our experiments, we do not use the ground truth body shape parameters, but instead use the same default shape for all sequences. This follows the evaluation used in previous work [1, 15, 9]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.378, + 0.772 + ], + "angle": 0, + "content": "4.1. Comparison to the state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.856 + ], + "angle": 0, + "content": "We compare HMD-NeMo with existing approaches that tackle the problem of full-body motion generation given HMD signal. To the best of our knowledge, all existing approaches only tackle the MC scenario, and thus we compare them with this setting4. As demonstrated in Table 1," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.428, + 0.876, + 0.542 + ], + "angle": 0, + "content": "
MethodMPJPE ↓MPJVE ↓
FinalIK [25]18.0959.24
Ahuja et al. [1]7.83100.54
Yang et al. [30]9.0244.97
Dittadi et al. [9]6.8337.99
Jiang et al. [15]4.1829.40
HMD-NeMo (Ours)1.9024.99
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.542, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Table 1. Comparison to the state-of-the-art approaches in MC scenario, where both hands are always visible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.893, + 0.807 + ], + "angle": 0, + "content": "HMD-NeMo is not only more accurate (lower MPJPE), but also generates smoother motion with joint velocities similar to that in the ground truth (lower MPJVE), introducing a new state-of-the-art on the AMASS dataset for both metrics. Note that [25] has the highest errors, as shown in Table 1 as it only optimizes the pose of the head and hands and ignores the accuracy and smoothness of the rest of the joints. Consequently, since [30] utilizes FinalIK within its framework, its performance becomes bounded by the quality of FinalIK. The learning-based approaches [15, 9], however, perform much better than [25, 1, 30], highlighting the value of data-driven methods trained on large-scale motion capture datasets. In Fig. 7, we compared HMD-NeMo with the second best performing baseline [15]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.666, + 0.832 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this section, we comprehensively evaluate different aspects of HMD-NeMo, including evaluation in HT scenario, cross-dataset evaluations, ablation studies on input signals, model architectures, and loss terms. We also ana" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.464, + 0.876 + ], + "angle": 0, + "content": "3The approach [9] is modified to predict motion in world coordinates." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.47, + 0.9 + ], + "angle": 0, + "content": "4Note, we exclude very recent approaches [22, 29] since they either do not report on AMASS or the code is not publicly available." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.863, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.307, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.202, + 0.258, + 0.212 + ], + "angle": 0, + "content": "(a) Walk in a semi-circle" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.098, + 0.554, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.201, + 0.486, + 0.212 + ], + "angle": 0, + "content": "(b) Walk, then knock" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.097, + 0.705, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.201, + 0.71, + 0.212 + ], + "angle": 0, + "content": "(c) Walking, then put sth down" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.094, + 0.892, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.201, + 0.845, + 0.212 + ], + "angle": 0, + "content": "(d) Walk, then sit" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.894, + 0.245 + ], + "angle": 0, + "content": "Figure 8. Qualitative results in HT scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). See the supplementary video for more qualitative results." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.26, + 0.46, + 0.31 + ], + "angle": 0, + "content": "
ModelMask TypeMPJPE ↓MPJVE ↓
HMD-NeMoLearned & fixed5.5942.81
HMD-NeMoTAMTs (Ours)2.4831.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.31, + 0.468, + 0.323 + ], + "angle": 0, + "content": "Table 2. Effect of Temporally Adaptable Mask Tokens (TAMTs)." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.323, + 0.468, + 0.351 + ], + "angle": 0, + "content": "These results represent the HMD-NeMo prediction before optimization solely to evaluate the effect of the TAMT module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.468, + 0.438 + ], + "angle": 0, + "content": "lyze the effect optimization in both HT and MC scenarios. Additionally, we provide results of HMD-NeMo in both HT and MC scenarios for various body parts, as well as qualitative effect of optimization in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.468, + 0.68 + ], + "angle": 0, + "content": "Evaluation for the hand tracking scenario. To the best of our knowledge, HMD-NeMo is the first approach to address the problem of human motion generation given partial HMD signal, applicable to HT scenario. We believe that TAMT module is the key to the success of HMD-NeMo in handling missing/partial observation, so we compare it with an alternative commonly used in the Vision Transformers [10, 12], which uses a learned set of parameters (i.e., nn.Parameters in PyTorch) to model the missing observations5. While learned parameters are fixed after training for every data point and every sequence, TAMT temporarily updates itself at each time-step given the current state of the model. Table 2 (especially MPJVE) shows the superiority of TAMT module over learned and fixed parameters in handling missing hand observations. Qualitatively, Fig. 8 illustrates how HMD-NeMo performs in HT scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.468, + 0.802 + ], + "angle": 0, + "content": "Cross-dataset evaluation. To investigate the generalizability of HMD-NeMo, we conduct a 3-fold cross-dataset evaluation as in [15], wherein the models are trained on two subsets and test on the other subset. In order to compare our approach with existing methods, we conduct this experiment in MC scenario. As shown in Table 3, HMD-NeMo outperforms existing approaches in all three datasets, by a considerable margin, highlighting its generalizability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.468, + 0.864 + ], + "angle": 0, + "content": "Evaluating the effect of input signal. As discussed in Section 3.1, HMD-NeMo utilizes head and hands, hands in the head space, as well as the corresponding velocities, as in Eq. 1. Table 4 summarizes the effect of each component in" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.26, + 0.894, + 0.356 + ], + "angle": 0, + "content": "
MethodTest on CMUTest on BMLrubTest on HDM05
MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓
FinalIK [25]18.8256.8317.5860.6418.4362.39
Ahuja et al. [1]18.77139.1713.30134.7717.90140.61
Yang et al. [30]12.9649.9411.0060.7411.9448.26
Dittadi et al. [9]13.0451.699.6951.8010.2140.07
Jiang et al. [15]8.3735.767.0443.708.0530.85
HMD-NeMo (Ours)7.1331.236.4640.386.8027.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.357, + 0.893, + 0.397 + ], + "angle": 0, + "content": "Table 3. Results of cross-dataset evaluation between different methods. In order to compare with existing methods, we provide results in MC scenario." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.413, + 0.875, + 0.499 + ], + "angle": 0, + "content": "
Input SignalMPJPE ↓MPJVE ↓
xt = {xh, xl, xr}t4.3839.63
xt = {xh, xl, xr, xlh, xrh}t3.2138.32
xt = {xh, xl, xr, ðh, ðl, ðr}t3.5335.27
Full input signals (Eq. 1)2.4831.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.499, + 0.892, + 0.539 + ], + "angle": 0, + "content": "Table 4. Evaluating the effect of various input signals in HT scenario. Note that the hand visibility status \\(\\nu_{l},\\nu_{r}\\) are always provided to the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.672 + ], + "angle": 0, + "content": "the input signal. As shown, adding hands in the head space on top of head and hands in the world coordinates leads to better pose prediction, and thus reduces the MPJPE. Incorporating the velocities has a significant contribution to generating more temporally coherent motion, and thus reduces the MPJVE considerably. Considering all input signals, as in Eq. 1 leads to best MPJPE and MPJVE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Evaluating the effect of STAE. One core component of HMD-NeMo is the spatioTemporal attention-based encoder (STAE). Here, we study the design choices of STAE in Table 5. As shown, removing the GRU component, which is responsible to learn the temporal information of various input signals, affects the MPJVE considerably. Removing the transformer encoder, which aims at learning the relation between head and hands, adversely affects the MPJPE. Unsurprisingly, removing the entire STAE module, i.e., going from input embedding layer to the decoders, is a considerably weak baseline, with poor pose quality and temporal coherency. HMD-NeMo utilizes the power and efficiency of GRU module to learn the temporal information as well as the expressively of the transformer encoder to learn how various inputs are related to each other. This design leads to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.468, + 0.901 + ], + "angle": 0, + "content": "5This can be considered as removing the unobserved hand (similar to [2]) and replacing it with learned and fixed parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9628" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.44, + 0.17 + ], + "angle": 0, + "content": "
STAE ComponentsMPJPE ↓MPJVE ↓
No GRU4.3446.61
No Transformer5.5539.10
No GRU, No Transformer7.0952.26
With STAE (Ours)2.4831.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.17, + 0.451, + 0.183 + ], + "angle": 0, + "content": "Table 5. Evaluating the effect of STAE module in HT scenario." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.196, + 0.44, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.471, + 0.373 + ], + "angle": 0, + "content": "Figure 9. Evaluating the effect of each loss term. Please note that each term is evaluated in isolation." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.398, + 0.468, + 0.427 + ], + "angle": 0, + "content": "best performing MPJPE and MPJVE, especially in HT scenario wherein the model faces regular missing observations." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.429, + 0.468, + 0.7 + ], + "angle": 0, + "content": "Evaluating the effect of each loss term. As described in Section 3.3 and shown in Eq. 6, HMD-NeMo is trained with five different loss terms. While \\(\\mathcal{L}_{data}\\) is the essential term for training HMD-NeMo, other loss terms contribute significantly to the performance of the model. In Fig. 9, we illustrate the contribution of each loss term, in leave-one-term-out manner, on the changes in the MPJPE and MPJVE metrics. \\(\\mathcal{L}_{smooth}\\) has a significant impact on improving the MPJVE, while \\(\\mathcal{L}_{aux}\\) improves the MPJPE, and \\(\\mathcal{L}_{forecast}\\), which acts only on hands mildly improves MPJPE and MPJVE (its contribution to the total error metric is relatively small as it acts on a significantly fewer joints). \\(\\mathcal{L}_{SE(3)}\\) makes the largest contribution to the reduction of error in both metrics. It is likely that \\(\\mathcal{L}_{SE(3)}\\) aims to bridge that gap between the representation of the input signal (head and hand global transformation matrices) and the representation of the output pose (global root orientation and relative joint rotations)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.468, + 0.852 + ], + "angle": 0, + "content": "Evaluating the effect of optimization. While the prediction of HMD-NeMo is very good, if the budget allows, it can be further optimized to improve accuracy (see supp mat for qualitative results). To bridge the gap between the predictions and observations, we optimize the pose prediction from HMD-NeMo, so that it matches the head and hand observations. As described in Section 3.4, in MC scenario, where both hands are always available, a simple optimization (with non-robust data energy term) loop can be used. The effect of such optimization in MC scenario is provided" + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.089, + 0.848, + 0.138 + ], + "angle": 0, + "content": "
ConfigurationMPJPE ↓MPJVE ↓
Without optimization2.0726.07
With optimization1.9024.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.139, + 0.892, + 0.166 + ], + "angle": 0, + "content": "Table 6. Effect of optimization in MC scenario, where both hands are always visible." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.179, + 0.878, + 0.242 + ], + "angle": 0, + "content": "
ConfigurationPriorityMPJPE ↓MPJVE ↓
No OptimizationPlausibility2.4831.30
Opt. with EnrFidelity2.3233.51
Opt. with ErBoth2.3731.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.242, + 0.892, + 0.282 + ], + "angle": 0, + "content": "Table 7. Configuring HT scenario. Our model offers a range of choices to provide the best possible end-user experience despite incomplete observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.473 + ], + "angle": 0, + "content": "in Table 6. However, considering the same strategy for the HT scenario, where we may partially observe hands, this may not be optimal as we lose plausibility in the generated motions (captured by the MPJVE metric). As described in Section 3.4, depending on the scenario and experience requirements, one may choose to (1) avoid optimizing the predictions if plausibility is the highest priority (first row of Table 7), (2) use non-robust energy term if fidelity is the highest priority (second row of Table 7), or (3) use a robust energy term as a trade-off between fidelity and plausibility (third row of Table 7)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Performance analysis. With 5.3M parameters, at inference time, HMD-NeMo requires only \\(4.4\\mathrm{ms}\\) to generate a pose given a HMD signal on a typical laptop CPU. On a NVIDIA Tesla P100 GPU, our model runs at 265 fps, with 1 iteration of optimization costs 3ms per frame (which could be further improved with more optimized implementation). Such performance makes HMD-NeMo a potential solution for HMD-driven avatar animation in immersive environments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.619, + 0.624 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.892, + 0.891 + ], + "angle": 0, + "content": "In this paper, we present HMD-NeMo, a unified approach to generate full-body avatar motion in both motion controller and hand tracking scenarios. To handle the unobserved hands in a temporally coherent and plausible manner, we introduce TAMT module. It is worth noting that, in this paper, we considered one major reason for partial hand visibility, i.e., hands appearing out of the FoV, however, in practice, hands may not be visible to the hand tracking camera due to failure in tracking and occlusion by another object or by another body part. While such cases are not considered in the data augmentation in our paper, they certainly can be taken into account and TAMT can be used to fill such gaps with no additional modification; this is left for future exploration. We provide extensive analyses on different components of HMD-NeMo, and shed light on various choices for optimization on top of neural network's predictions when it comes to production priorities (plausibility versus fidelity)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": "6This experiment only evaluates the contribution of each loss term independently. Evaluating the combination of loss terms remains for future investigations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9629" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Karan Ahuja, Eyal Ofek, Mar Gonzalez-Franco, Christian Holz, and Andrew D Wilson. Coolmoves: User motion accentuation in virtual reality. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 5(2):1-23, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. FLAG: Flow-based 3D avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[3] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.322, + 0.47, + 0.376 + ], + "angle": 0, + "content": "[4] Benjamin Biggs, Sébastien Ehrhadt, Hanbyul Joo, Benjamin Graham, Andrea Vedaldi, and David Novotny. 3D multibodies: Fitting sets of plausible 3D human models to ambiguous image data. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.38, + 0.47, + 0.449 + ], + "angle": 0, + "content": "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.492 + ], + "angle": 0, + "content": "[6] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[7] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.556, + 0.469, + 0.597 + ], + "angle": 0, + "content": "[8] Vasileios Choutas, Federica Bogo, Jingjing Shen, and Julien Valentin. Learning to fit morphable models. arXiv preprint arXiv:2111.14824, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[9] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating SMPL poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[11] Nima Ghorbani and Michael J Black. Soma: Solving optical marker-based mocap automatically. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11117-11126, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[13] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[15] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. AvatarPoser: Articulated full-body pose tracking from sparse motion sensing. In Proceedings of European Conference on Computer Vision. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. International Conference on Learning Representations, ICLR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[17] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11605-11614, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[18] Matthew Loper, Naureen Mahmood, and Michael J Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics (TOG), 33(6):1-13, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[19] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[20] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, Oct. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[21] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10975-10985, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[22] Jose Luis Ponton, Haoran Yun, Carlos Andujar, and Nuria Pelechano. Combining Motion Matching and Orientation Prediction to Animate Avatars for Consumer-Grade VR Devices. Computer Graphics Forum, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[23] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[24] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.888, + 0.871 + ], + "angle": 0, + "content": "[25] Unity Asset Store. Final IK. https://assetstoreunity.com/packages/tools/animation/final-ik-14290, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Dorin Ungureanu, Federica Bogo, Silvano Galliani, Pooja Sama, Casey Meekhof, Jan Stuhmer, Thomas J Cashman," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9630" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "Bugra Tekin, Johannes L Schonberger, Pawel Olszta, et al. Hololens 2 research mode as a tool for computer vision research. arXiv preprint arXiv:2008.11239, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[28] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3D human pose estimation from sparse IMUs. In Computer graphics forum, volume 36, pages 349-360. Wiley Online Library, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.468, + 0.304 + ], + "angle": 0, + "content": "[29] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. arXiv preprint arXiv:2209.09391, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[30] Dongseok Yang, Doyeon Kim, and Sung-Hee Lee. LoBSTr: Real-time lower-body pose prediction from sparse upper-body tracking signals. In Computer Graphics Forum, volume 40, pages 265–275. Wiley Online Library, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[31] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical Inertial Poser (PIP): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13167-13178, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[32] Xinyu Yi, Yuxiao Zhou, and Feng Xu. TransPose: Real-time 3D human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[33] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision, pages 465-481. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[34] Mihai Zanfir, Andrei Zanfir, Eduard Gabriel Bazavan, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. THUNDR: Transformer-based 3D human reconstruction with markers. arXiv preprint arXiv:2106.09336, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[35] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3D bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.768 + ], + "angle": 0, + "content": "[36] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9631" + } + ] +] \ No newline at end of file diff --git a/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_origin.pdf b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..395a4b617d8530d8135c915c6080a6efe4d5e7b7 --- /dev/null +++ b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/c461bfb7-8a1f-4d23-a5a0-eb903d127c6c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e59a2ac957e8807557e01bbcf17b518ea30fe33c75983479b31290d2a2393785 +size 7625677 diff --git a/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/full.md b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e301221d1dd934cf96e7361120af1f6b46cd3591 --- /dev/null +++ b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/full.md @@ -0,0 +1,320 @@ +# HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations + +Sadegh Aliakbarian + +Fatemeh Saleh + +David Collier + +Pashmina Cameron + +Darren Cosker + +Microsoft Mixed Reality & AI Lab, Cambridge, UK + +# Abstract + +Generating both plausible and accurate full body avatar motion is the key to the quality of immersive experiences in mixed reality scenarios. Head-Mounted Devices (HMDs) typically only provide a few input signals, such as head and hands 6-DoF. Recently, different approaches achieved impressive performance in generating full body motion given only head and hands signal. However, to the best of our knowledge, all existing approaches rely on full hand visibility. While this is the case when, e.g., using motion controllers, a considerable proportion of mixed reality experiences do not involve motion controllers and instead rely on egocentric hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view of the HMD. In this paper, we propose the first unified approach, HMD-NeMo, that addresses plausible and accurate full body motion generation even when the hands may be only partially visible. HMD-NeMo is a lightweight neural network that predicts the full body motion in an online and real-time fashion. At the heart of HMD-NeMo is the spatiotemporal encoder with novel temporally adaptable mask tokens that encourage plausible motion in the absence of hand observations. We perform extensive analysis of the impact of different components in HMD-NeMo and introduce a new state-of-the-art on AMASS dataset through our evaluation. + +# 1. Introduction + +Mixed reality technology opens up new means of communication and interaction between people. With people at the heart of this technology, generating faithful and believable avatar motion is key to the quality of immersive experiences. Despite great advances in this area, generating full body avatar motion given HMD signal remains a challenge: in many current solutions, avatars only have upper bodies. + +Prior works attempted to generate full body avatar motion given sparse or partial observations, such as images [17, 4, 33], 2D joints/ keypoints [21, 5], markers [35, 18, 34, 11], and IMUs [28, 14, 32, 31]. While such observations are + +![](images/ce46fd5a55b6b22f70a3dcbc66a4ca057221e6447660d2d6be9dfb1b5fef58e8.jpg) +Figure 1. HMD-NeMo generates full body avatar motion given HMD signals, i.e., head and hand 6-DoFs, from hand tracking signal from the HMD as well as hand motion controllers. + +considered partial or sparse, they provide much richer input signal compared to a typical HMD's head and hand 6-DoF. More recently, great progress has been made to generate full body motion given only a HMD signal [1, 30, 22, 15], however, they all rely on the availability and full visibility of both hands. While this is the case when, e.g., using motion controllers, many mixed reality experiences do not involve motion controllers and instead rely on hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view (FoV) of the HMD sensors. + +In this paper, we address this problem via HMD-NeMo (a neural motion model of human given HMD signal). Within a unified framework, HMD-NeMo generates full body motion in real time, regardless of whether hands are fully or partially observed, or not observed at all. Our approach is built upon recurrent neural networks to efficiently capture temporal information, and a transformer to capture complex relations between different components of the input signal. At the heart of our approach is the TAMT (temporally adaptable mask token) module, allowing us to handle missing hand observations. + +Our contributions are: (1) The first full body avatar motion generation approach capable of generating accurate and plausible motions with full or partial hand visibility. This is a step forward for unlocking fully immersive experiences in mixed reality with fewer limitations on the hardware. (2) Temporally adaptable Mask Tokens (TAMT), a simple yet + +![](images/345499123601787b2ca1afce709a8cb33b7218f1914b825766d56e427703ad24.jpg) +(a) Plane FoV +Figure 2. Top: Definition of FoV for different scenarios for the same pose. (a) Planar FoV, as used in [8], wherein avatar's left hand is visible while the right hand is not. (b) Fully visible, as in motion controller scenarios, wherein both hands are always visible. (c) HMD's hand tracking camera's FoV, as in hand tracking scenarios, wherein avatar's right hand is visible while the left hand is not. Bottom: Hand visibility statistics on AMASS test set. + +![](images/7a8bad323a803924405405444bfb6e52b44ce1efb36212b2ced535b376566369.jpg) +(b) Motion Controllers (MC) + +![](images/748ce4138b12aa35c14f7f340ad4379be726af0163d5a198fe7de3d5436aad58.jpg) +(c) Hand Tracking (HT) + +
ScenarioLeft hand visibilityRight hand visibility
Motion Controllers100%100%
Hand Tracking53.45%48.94%
+ +effective strategy for handling missing hand observations in a temporally coherent way. (3) Extensive experiments and ablations of our method as well as a new state-of-the-art performance on the challenging AMASS dataset. + +# 2. Related Work + +Over the past few years, different solutions have been proposed to the problem of full body pose generation given sparse or partial observations, such as markers [35, 18, 34, 11], images [17, 4, 33], 2D joints/keypoints [21, 5], Inertial Measurement Units (IMUs) [28, 14, 32, 31], and HMDs [2, 1, 30, 22, 15]. Among these works, the ones utilizing wearable devices are closer to our approach and thus we discuss them here. Recently, different techniques have been proposed for human pose reconstruction using a sparse set of IMUs attached to the body [28, 14, 32, 31]. While attempts have been made to come up with a minimum number of IMUs for generating full body motion, typically an IMU sensor is used near pelvis, making the problem relatively easy compared to HMD scenarios where the root/pelvis signal is not available. Whilst external sensors, e.g., IMUs [28, 14, 32, 31] and cameras [23] are effective, they are often not as accessible as wearing a HMD. Using only a HMD is desirable from a usability point of view, but generating realistic and faithful human representations from such inputs remains technically challenging. + +Given HMD signals, [2] and [8] generate full body poses. While they generate expressive poses faithful to the HMD signal, these approaches [2, 8] only predict static poses, lacking the temporal consistency required for avatar motion generation. While [8] generates poses in the world space, [2] predicts the pose relative to the root. Thus [2] + +assumes a known root (pelvis joint) as an additional signal. The assumption of known root joint also appears in [9], wherein, unlike [2, 8], the method generates full body motion given the HMD signal with Variational Autoencoders [16]. Note that, while [9] generates temporally plausible motions, it works in an offline fashion, predicting the motion for an entire sequence only after observing the whole sequence of HMD signals. + +Recently, different techniques have been proposed to generate full body avatar locomotion (in the world coordinates) given HMD signals. In this context, [1] proposes a matching algorithm, aiming to sample closest poses from a motion capture library at sparse time-steps and interpolate between poses. While this guarantees the selection of realistic poses, the output is always limited to the utilized motion capture library. In another work, [30] uses combination of an inverse kinematic (IK) solver and a recurrent neural network to generate upper body and lower body motions, respectively. Combination of different components to solve lower and upper body separately has also been explored in [22], wherein a neural network is trained to predict the root orientation given the HMD signal, which then is used as the feature vector for a motion matching algorithm [6, 13] to generate full body. More recently, fully learning-based approaches have shown promise in generating full body avatar motion [29, 15]. In this context, [29] simulates plausible and physically feasible motions within a reinforcement learning framework and [15] uses a transformer-based approach to generate full body motion given HMD signals. + +Although great progress has been made by recent approaches, they all solve the problem of motion generation given head and both hands, typically captured with motion controllers. However many mixed reality experiences do not have motion controllers available and instead rely on hand tracking from HMD mounted sensors (e.g., cameras). This introduces the challenge of hand tracking failures and partial hand visibility owing to the restricted field of views. To the best of our knowledge, despite its usability, motion generation in the presence of partial hand observation has not been well-explored. In this paper, we propose a method capable of generating high fidelity and plausible full body motion even in presence of partially visible hands. + +# 3. Proposed Method + +In this section, we first define the problem, the scenarios we consider, as well as the input and the desired output representation. We then present our proposed method. + +# 3.1. Problem Definition + +Task. The task is to generate full-body 3D human locomotion (predicting both the instantaneous pose and the global trajectory of the human) given the sparse HMD signal in + +![](images/49f341d087760e6acd468b567509dd7b1ac36f7d0c373122a7b2cfa482ce326b.jpg) +Figure 3. Overview of HMD-NeMo: At each time-step, the model gets as input the HMD signal (left), as in Eq. 1. Such input is the mapped to an embedding space via $f_{e}$ , acting as the input to the spatio-temporal encoder module STAE. The resulting feature representation is then utilized by two autoregressive decoders, $g\gamma$ and $g_{\theta}$ to predict the global trajectory as well as the pose, respectively. In case the model does not observe a hand (in HT scenario), $f_{amt}$ fills in the representation of the unobserved hand effectively. + +an online fashion1. That is, given the input signal $x_{t}$ at each time-step $t$ , the system should predict the 3D human pose and trajectory $y_{t}$ near-instantaneously. HMD-NeMo achieves this using a neural network parameterized by $\phi$ . + +Scenarios. In this paper, we consider two scenarios: Motion Controllers (MC) scenario, wherein hands are always tracked via motion controllers using constellation tracking, as illustrated in Fig. 2 (b), and Hand Tracking (HT) scenario, wherein hands are tracked via a visual hand tracking system whenever the hands are inside the FoV of the device, as illustrated in Fig. 2 (c). The FoV of the device is defined as a frustum determined by the HMD's hand tracking camera placement and parameters. In this paper, we use a similar FoV to that of the Microsoft HoloLens 2 [26]. Note that, while the MC scenario has been explored recently [1, 30, 22, 15], to the best of our knowledge, this work constitutes the first approach that tackles both MC and HT scenarios within one unified framework2. The HT scenario is particularly challenging as hands tend to be out of FoV almost $50\%$ of the time, according to the statistics presented in the table in Fig. 2. + +Input Representation. The input signal $x^{t}$ contains the head 6-DoF $x_{h} \in \mathbb{R}^{(6 + 3)}$ , the left hand 6-DoF $x_{l}^{t} \in \mathbb{R}^{(6 + 3)}$ , and the right hand 6-DoF $x_{r}^{t} \in \mathbb{R}^{(6 + 3)}$ , all in the world space. We use the 6D representation to represent the rotations [36]. We additionally provide the hand representations in the head space, $x_{lh}^{t} \in \mathbb{R}^{(6 + 3)}$ and $x_{rh}^{t} \in \mathbb{R}^{(6 + 3)}$ . In the HT scenario, hands may go in and out of FoV of the HMD, so we also provide HMD-NeMo with the hand visibility status for both left and right hand, $\nu_{l}^{t}$ and $\nu_{r}^{t}$ , as binary values, 1 being visible and 0 otherwise. Finally, for all 6- + +DoF signals, we provide the velocity of changes between two consecutive frames. Specifically, for translations we consider $vel(\mathcal{T}^t,\mathcal{T}^{t - 1}) = \mathcal{T}^t -\mathcal{T}^{t - 1}$ , where $\mathcal{T}$ is the translation, and for rotations we consider the geodesic changes in the rotation $vel(R^{t},R^{t - 1}) = (R^{t - 1})^{-1}R^{t}$ , where $R$ is the rotation, which together constitute the velocity 6-DoF $\dot{x}^t$ . Overall, the input to the HMD-NeMo, $x^{t}\in \mathbb{R}^{92}$ , can be written as + +$$ +x ^ {t} = \left\{x _ {h}, x _ {l}, x _ {r}, x _ {l h}, x _ {r h}, \dot {x} _ {h}, \dot {x} _ {l}, \dot {x} _ {r}, \dot {x} _ {l h}, \dot {x} _ {r h}, \nu_ {l}, \nu_ {r} \right\} ^ {t} \tag {1} +$$ + +Output Representation. The output of HMD-NeMo contains the pose (including the root orientation), $\theta^t\in \mathbb{R}^{J\times 3}$ represented with axis-angle rotations for the $J$ joints in the body, and the global position in the world, $\gamma^t\in \mathbb{R}^3$ represented as the root position, resulting in $y^{t}\in \mathbb{R}^{(J + 1)\times 3}$ + +$$ +y ^ {t} = \{\theta , \gamma \} ^ {t} \tag {2} +$$ + +The sequence of $\theta^{0:T}$ and $\gamma^{0:T}$ then represent the avatar motion as well as its global trajectory for the period $[0,T]$ . Note that, in this paper, we may drop the time superscript $t$ for better readability and include it when necessary. + +# 3.2. HMD-NeMo + +Overview. HMD-NeMo's pipeline is illustrated in Fig. 3. As described in Section 3.1, the model gets as input the information about the head and hands in world coordinate system, the hands expressed in the head space, as well as their velocities, as described in Eq. 1. To express hands in head space, we consider $x_{lh} = x_h^{-1}x_l$ (similarly, $x_{rh} = x_h^{-1}x_r$ ). This representation then acts as the input to an embedding layer, $f_e$ , which aims to (1) map the raw input to an embedding space and (2) handle the unobserved hands. Given the input $f_e$ , the next module, STAE, learns (a) how each representation evolves over time and (b) how different components of the input, i.e., head and hands, are correlated. Once such a rich representation is obtained, two auto-regressive decoders, $g_\theta$ and $g_\gamma$ , generate the body pose and the global + +![](images/0abb9877a3e991e66204a2b2dc7af6be195d29e3b50a9970f0d8c61d60b93478.jpg) +Figure 4. Overview of the embedding module. Note that $x_{l}$ is only provided as an example and this module applies to all 6-DoF inputs appearing in $x^{t}$ as in Eq. 1. + +position of the avatar, respectively. At each time step, the output of STAE is used to update the mask tokens (described below) as a representation for the hand signals that may be missing in the next time-step. To aid training, we also include an auxiliary task of human pose reconstruction in SE(3), denoted by $g_{aux}$ . + +Head and hand embedding $(f_e)$ . As shown in Fig. 4, the embedding module $f_e$ gets as input the head and hands 6-DoFs and velocities and maps them to a higher-dimensional embedding space. As the range of values corresponding to the rotations is different from those of the translations, we decouple such information and embed them via separate shallow MLPs and concatenate the results back together. For instance, for the observed left hand in the world coordinate system, the embedding representation is + +$$ +e _ {l} ^ {\text {v i s i b l e}} = \left[ f _ {e} ^ {\mathrm {R}} \left(x _ {l} [: 6 ]\right), f _ {e} ^ {\mathrm {T}} \left(x _ {l} [ 6 : ]\right), \dot {f} _ {e} ^ {\mathrm {R}} \left(\dot {x} _ {l} [: 6 ]\right), \dot {f} _ {e} ^ {\mathrm {T}} \left(\dot {x} _ {l} [ 6 : ]\right) \right] \tag {3} +$$ + +where $f_{e}^{\mathrm{R}}$ and $f_{e}^{\mathrm{T}}$ are MLPs responsible for computing the rotation and translation embedding, acting on the first 6 elements (the 6D rotation representation) and the last 3 elements (the translation) of the input, respectively (similarly for $\dot{f}_{e}^{\mathrm{R}}$ and $\dot{f}_{e}^{\mathrm{T}}$ which act on velocities). + +In the HT scenario, hands may not be visible to the model, hence computing such embedding representation is not possible. Thus, given the status of $\nu_{l}$ and $\nu_{r}$ , the embedding module decides to either compute the embedding or utilize the output of the $f_{\text{tamt}}$ (described below), a set of temporally adaptable mask tokens, instead of a missing hand observation (denoted by $M$ in Fig. 3). As illustrated in Fig 4, the embedding of the left hand in the world coordinate system can be computed as + +$$ +e _ {l} = \nu_ {l} e _ {l} ^ {\text {v i s i b l e}} + (1 - \nu_ {l}) f _ {\text {t a m t}} (l). \tag {4} +$$ + +Spatio-temporal encoder (STAE). The output of $f_{e}$ on each component of the input stream is a non-temporal feature, computed independent of other components in the input. While an expressive representation of each component, it lacks temporal and spatial correlation information. We specifically care about these characteristics because the model is required to generate temporally coherent motion and also because the motion of one body part often impacts + +![](images/565a806dbe0ce8def6ffc9de4e93088fae87c02e4237a23e1b4e0d93dd9262c5.jpg) +Figure 5. Overview of the spatio-temporal encoder module. + +or determines the motion of other body part(s). To obtain a more informative representation from the head and hands, HMD-NeMo first learns the isolated temporal features of each component of input representation and then learns how they are spatially correlated [27]. + +As illustrated in Fig. 5, to learn the temporal representation of the input signal, we use gated recurrent units (GRUs) [7]. With a GRU module on top of each component in the input, the model learns how each component, e.g., head, evolves over time, independent of other components in the input. This information is compressed in the hidden state of the GRU cell, $z$ , which is then utilized to learn how different components in the input relate to each other. This is achieved by using a standard transformer encoder on the GRU hidden states, thanks to the self-attention mechanism. Temporally adaptable mask tokens (TAMTs). As discussed in Section 3.1, hands may not be visible to the model, and thus, there is no representative input signal for the $f_{e}$ module. To address this issue, in case of a missing hand observation, our model produces a feature vector, $f_{amt}$ , to represent the missing hand observation. To compute $f_{amt}$ , as illustrated in Fig. 6, we use the output of STAE for the hand observation that may be missing in the next time step as well as the output of STAE for the head. Note that head joint is the reference joint and is always available. The combination of these two features is a rich representation of the past state of the missing hand signal (both temporally and spatially); this is then used to compute the $f_{amt}$ . In order to encourage $f_{amt}$ to learn information about the missing hand observation, as illustrated by the Forecaster module in Fig. 6, we introduce a forecasting auxiliary task to forecast the state (6-DoF) of the corresponding hand in the next time-step. + +# 3.3. Training HMD-NeMo + +To train HMD-NeMo, we rely on the availability of a motion capture dataset, represented as SMPL [19] parameters (pose, shape, and global trajectory). For each sequence, we then simulate the HMD on the subject. In case of HT, we also simulate a FoV frustum to be able to model the hand visibility status $(\nu_{l}$ and $\nu_{r})$ . We then train HMD-NeMo with + +a loss function of the form + +$$ +\mathcal {L} = \alpha_ {\text {d a t a}} \mathcal {L} _ {\text {d a t a}} + \alpha_ {\text {s m o o t h}} \mathcal {L} _ {\text {s m o o t h}} \tag {5} +$$ + +$$ ++ \alpha_ {S E (3)} \mathcal {L} _ {S E (3)} + \alpha_ {\text {f o r e c a s t}} \mathcal {L} _ {\text {f o r e c a s t}} + \alpha_ {\text {a u x}} \mathcal {L} _ {\text {a u x}} +$$ + +The data loss term is the squared error between the predicted pose and trajectory and those of the ground truth, + +$$ +\mathcal {L} _ {\text {d a t a}} = \sum_ {t = 1} ^ {T} | | \hat {\theta} ^ {t} - \theta^ {t} | | _ {2} ^ {2} + | | \hat {\gamma} ^ {t} - \gamma^ {t} | | _ {2} ^ {2} \tag {6} +$$ + +Note that, in practice, the pose decoder has two heads, one each for predicting the body pose and the global root orientation. To further enhance the temporal smoothness, we penalize the discrepancy between the velocity of changes in the prediction to that of the ground truth + +$$ +\mathcal {L} _ {\text {s m o o t h}} = \sum_ {t = 2} ^ {T} | | \delta \hat {\theta} ^ {t} - \delta \theta^ {t} | | _ {1} + | | \delta \hat {\gamma} ^ {t} - \delta \gamma^ {t} | | _ {1} \tag {7} +$$ + +where $\delta \hat{\theta}^t = \hat{\theta}^t -\hat{\theta}^{t - 1}$ ( $\delta \hat{\gamma}$ , $\delta \theta$ , and $\delta \gamma$ follow similarly). In addition to computing the reconstruction loss on the SMPL parameters, i.e., to relative joint rotations, we found it extremely useful to also utilize the reconstruction loss of each joint transformation independent of its parent, i.e., in the world space. To compute this reconstruction loss, we use the SMPL model to compute the joint transformations in SE(3) given the predicted and ground truth pose and trajectory parameters. Thus, the SE(3) reconstruction loss is + +$$ +\mathcal {L} _ {S E (3)} = \sum_ {t = 1} ^ {T} \left| \left| \hat {P} _ {S E (3)} ^ {t} - P _ {S E (3)} ^ {t} \right| \right| _ {2} ^ {2} \tag {8} +$$ + +where $P_{SE(3)}$ is the body pose in SE(3). The next loss term corresponds to the forecasting auxiliary task in the TAMT module, where the goal is to minimize the distance between the predicted next hand and the ground truth next hand, + +$$ +\mathcal {L} _ {\text {f o r e c a s t}} = \sum_ {t = 2} ^ {T} \sum_ {j \in \{l, r \}} \left\| \hat {x} _ {j} ^ {t} - x _ {j} ^ {t} \right\| _ {2} ^ {2} \tag {9} +$$ + +Finally, we have our loss term for the auxiliary task, aiming to minimize the predicted full body joint transformations from STAE's features, $\hat{P}_{aux}$ , to the ground truth body joint transformations, + +$$ +\mathcal {L} _ {a u x} = \sum_ {t = 1} ^ {T} \left| \left| \hat {P} _ {a u x} ^ {t} - P _ {S E (3)} ^ {t} \right| \right| _ {2} ^ {2} \tag {10} +$$ + +Our model is trained with $\alpha$ s all being set to 1 in Eq. 6. + +# 3.4. Optimization + +Once trained, HMD-NeMo is capable of generating high fidelity and plausible human motion given only the HMD signal. However, as is typical of learning-based approaches, the direct prediction of the neural network does not precisely match the observations i.e., the head and hands, even if it is perceptually quite close. To close this gap between the prediction and the observation, optimization can + +![](images/cae97331164c5d976d9cf8b1cfa682b4b9436eaae00271e10a2dbe3c1521afe7.jpg) +Figure 6. Overview of TAMT module. Note that here we illustrate TAMT of $x_{l}$ as an example, the same applies f to both hands in both head space and world space. + +be used. This adjusts the pose parameters to minimize an energy function of the form $\mathcal{E} = \mathcal{E}_{data} + \mathcal{E}_{reg}$ , where $\mathcal{E}_{data}$ is the energy term that minimizes the distance between the predicted head and hands to the observed ones, and $\mathcal{E}_{reg}$ is additional regularization term(s). To define the data energy term, we define the residual $\mathcal{R} = \sum_{j\in \{h,l,r\}}(x_j - \hat{x}_j)$ , i.e., the difference between the predicted head/hand joint to that of the observation. Given $\mathcal{R}$ , a typical, non-robust data energy term could be written as $\mathcal{E}_{nr} = \mathcal{R}^2$ , i.e., the L2 loss. This suits the MC scenario perfectly, where head, left hand, and right hand are always available. But this energy term may be misleading in HT scenario where hands are going into and out of FoV often, and thus leading to abrupt pose changes when hands appear back in the FoV (see supp mat for further details). To remedy this issue, we utilize a more robust [3] alternative to the data energy term, + +$$ +\mathcal {E} _ {r} (\mathcal {R}, a, b, c) = b \frac {| a - 2 |}{a} \left(\left(\frac {\left(\frac {\mathcal {R}}{c}\right) ^ {2}}{| a - 2 |} + 1\right) ^ {\left(\frac {a}{2}\right)} - 1\right) \tag {11} +$$ + +where $a$ , $b$ , and $c$ are hyper-parameters that determine the shape of the loss (see supp mat for further details). Unlike $\mathcal{E}_{nr}$ , $\mathcal{E}_r$ , considers large discrepancies between the prediction and observation as outliers, not penalizing the prediction strongly and thus does not push the prediction to move toward the observation. Thus abrupt changes in the arm poses are avoided and optimization stays on course despite large variation in the velocity metric (when caused by hand visibility changes). Of course, utilizing Eq. 11 adversely affects the fidelity, but a trade-off between the plausibility and fidelity can be chosen to suit the application of interest. + +Note that, since all observations relate to the upper body, during optimization we only optimize the upper body pose parameters and global root trajectory, while keeping the predicted lower body untouched. + +# 4. Experiments + +In this section, we introduce the dataset and metrics, with implementation details in the supplementary material. We then present the experimental results and ablation studies. + +Dataset. We follow the recent common practice [15, 24, 8, 9] of using a subset of AMASS [20] for training and evaluation. AMASS is a large collection of human motion se + +![](images/274d7d51940aebbc75701ddf3ad68738e31cbf0985e4457687a0733ef5c8c44d.jpg) +Figure 7. Comparison to the state-of-the-art in MC scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). Last row depicts a hard example with complex body pose and motion. + +quences, converted to 3D human meshes in the SMPL [19] representation, wherein every motion sequence contains information about poses $\theta$ and the global trajectory $\gamma$ . To synthesize the HMD scenario, we compute the global transformation matrices for the head and hands as input. In the case of hand tracking, we define a FoV for the HMD and mask out the hands whenever they are out of FoV. To make a fair comparison, for both training and evaluation, we follow the splits suggested by [15]. + +Metrics. To evaluate the performance of our approach as well as the competing baselines, we report the mean per-joint position error (MPJPE [cm]) and the mean per-joint velocity error (MPJVE [cm/s]). We compare our approach with recent techniques [15, 1, 30, 9] $^3$ . As these approaches do not tackle the hand tracking scenario, we compare HMD-NeMo against them for the motion controller scenario. We evaluate HMD-NeMo for hand tracking scenarios separately. In our experiments, we do not use the ground truth body shape parameters, but instead use the same default shape for all sequences. This follows the evaluation used in previous work [1, 15, 9]. + +# 4.1. Comparison to the state-of-the-art + +We compare HMD-NeMo with existing approaches that tackle the problem of full-body motion generation given HMD signal. To the best of our knowledge, all existing approaches only tackle the MC scenario, and thus we compare them with this setting4. As demonstrated in Table 1, + +
MethodMPJPE ↓MPJVE ↓
FinalIK [25]18.0959.24
Ahuja et al. [1]7.83100.54
Yang et al. [30]9.0244.97
Dittadi et al. [9]6.8337.99
Jiang et al. [15]4.1829.40
HMD-NeMo (Ours)1.9024.99
+ +Table 1. Comparison to the state-of-the-art approaches in MC scenario, where both hands are always visible. + +HMD-NeMo is not only more accurate (lower MPJPE), but also generates smoother motion with joint velocities similar to that in the ground truth (lower MPJVE), introducing a new state-of-the-art on the AMASS dataset for both metrics. Note that [25] has the highest errors, as shown in Table 1 as it only optimizes the pose of the head and hands and ignores the accuracy and smoothness of the rest of the joints. Consequently, since [30] utilizes FinalIK within its framework, its performance becomes bounded by the quality of FinalIK. The learning-based approaches [15, 9], however, perform much better than [25, 1, 30], highlighting the value of data-driven methods trained on large-scale motion capture datasets. In Fig. 7, we compared HMD-NeMo with the second best performing baseline [15]. + +# 4.2. Ablation Studies + +In this section, we comprehensively evaluate different aspects of HMD-NeMo, including evaluation in HT scenario, cross-dataset evaluations, ablation studies on input signals, model architectures, and loss terms. We also ana + +![](images/74740f3bbe1daa84c1047f5b89d4dd93e2b893fc94d4925594536802dba4a63d.jpg) +(a) Walk in a semi-circle + +![](images/5204f3aa56deab36440cd4ebda798cc8953e9a372adfafa5f1677529dc2fcf96.jpg) +(b) Walk, then knock + +![](images/3e31aae32e2b9e2fb2f868fb598e5b76a8453d84a4d028c2605b929a5ccbe21c.jpg) + +![](images/927eb99c137d9274f07cb583ee074eb4929881ac4dad9562cd08b2981b98f83b.jpg) +(c) Walking, then put sth down +(d) Walk, then sit +Figure 8. Qualitative results in HT scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). See the supplementary video for more qualitative results. + +
ModelMask TypeMPJPE ↓MPJVE ↓
HMD-NeMoLearned & fixed5.5942.81
HMD-NeMoTAMTs (Ours)2.4831.30
+ +lyze the effect optimization in both HT and MC scenarios. Additionally, we provide results of HMD-NeMo in both HT and MC scenarios for various body parts, as well as qualitative effect of optimization in the supplementary material. + +Evaluation for the hand tracking scenario. To the best of our knowledge, HMD-NeMo is the first approach to address the problem of human motion generation given partial HMD signal, applicable to HT scenario. We believe that TAMT module is the key to the success of HMD-NeMo in handling missing/partial observation, so we compare it with an alternative commonly used in the Vision Transformers [10, 12], which uses a learned set of parameters (i.e., nn.Parameters in PyTorch) to model the missing observations5. While learned parameters are fixed after training for every data point and every sequence, TAMT temporarily updates itself at each time-step given the current state of the model. Table 2 (especially MPJVE) shows the superiority of TAMT module over learned and fixed parameters in handling missing hand observations. Qualitatively, Fig. 8 illustrates how HMD-NeMo performs in HT scenario. + +Cross-dataset evaluation. To investigate the generalizability of HMD-NeMo, we conduct a 3-fold cross-dataset evaluation as in [15], wherein the models are trained on two subsets and test on the other subset. In order to compare our approach with existing methods, we conduct this experiment in MC scenario. As shown in Table 3, HMD-NeMo outperforms existing approaches in all three datasets, by a considerable margin, highlighting its generalizability. + +Evaluating the effect of input signal. As discussed in Section 3.1, HMD-NeMo utilizes head and hands, hands in the head space, as well as the corresponding velocities, as in Eq. 1. Table 4 summarizes the effect of each component in + +Table 2. Effect of Temporally Adaptable Mask Tokens (TAMTs). +These results represent the HMD-NeMo prediction before optimization solely to evaluate the effect of the TAMT module. + +
MethodTest on CMUTest on BMLrubTest on HDM05
MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓
FinalIK [25]18.8256.8317.5860.6418.4362.39
Ahuja et al. [1]18.77139.1713.30134.7717.90140.61
Yang et al. [30]12.9649.9411.0060.7411.9448.26
Dittadi et al. [9]13.0451.699.6951.8010.2140.07
Jiang et al. [15]8.3735.767.0443.708.0530.85
HMD-NeMo (Ours)7.1331.236.4640.386.8027.91
+ +Table 3. Results of cross-dataset evaluation between different methods. In order to compare with existing methods, we provide results in MC scenario. + +
Input SignalMPJPE ↓MPJVE ↓
xt = {xh, xl, xr}t4.3839.63
xt = {xh, xl, xr, xlh, xrh}t3.2138.32
xt = {xh, xl, xr, ðh, ðl, ðr}t3.5335.27
Full input signals (Eq. 1)2.4831.30
+ +Table 4. Evaluating the effect of various input signals in HT scenario. Note that the hand visibility status $\nu_{l},\nu_{r}$ are always provided to the model. + +the input signal. As shown, adding hands in the head space on top of head and hands in the world coordinates leads to better pose prediction, and thus reduces the MPJPE. Incorporating the velocities has a significant contribution to generating more temporally coherent motion, and thus reduces the MPJVE considerably. Considering all input signals, as in Eq. 1 leads to best MPJPE and MPJVE. + +Evaluating the effect of STAE. One core component of HMD-NeMo is the spatioTemporal attention-based encoder (STAE). Here, we study the design choices of STAE in Table 5. As shown, removing the GRU component, which is responsible to learn the temporal information of various input signals, affects the MPJVE considerably. Removing the transformer encoder, which aims at learning the relation between head and hands, adversely affects the MPJPE. Unsurprisingly, removing the entire STAE module, i.e., going from input embedding layer to the decoders, is a considerably weak baseline, with poor pose quality and temporal coherency. HMD-NeMo utilizes the power and efficiency of GRU module to learn the temporal information as well as the expressively of the transformer encoder to learn how various inputs are related to each other. This design leads to + +
STAE ComponentsMPJPE ↓MPJVE ↓
No GRU4.3446.61
No Transformer5.5539.10
No GRU, No Transformer7.0952.26
With STAE (Ours)2.4831.30
+ +![](images/f85b8bbdc99d2ec8b5439985ef482e66485cf2b34edcd4a7c75cd18033d23214.jpg) +Figure 9. Evaluating the effect of each loss term. Please note that each term is evaluated in isolation. + +best performing MPJPE and MPJVE, especially in HT scenario wherein the model faces regular missing observations. + +Evaluating the effect of each loss term. As described in Section 3.3 and shown in Eq. 6, HMD-NeMo is trained with five different loss terms. While $\mathcal{L}_{data}$ is the essential term for training HMD-NeMo, other loss terms contribute significantly to the performance of the model. In Fig. 9, we illustrate the contribution of each loss term, in leave-one-term-out manner, on the changes in the MPJPE and MPJVE metrics. $\mathcal{L}_{smooth}$ has a significant impact on improving the MPJVE, while $\mathcal{L}_{aux}$ improves the MPJPE, and $\mathcal{L}_{forecast}$ , which acts only on hands mildly improves MPJPE and MPJVE (its contribution to the total error metric is relatively small as it acts on a significantly fewer joints). $\mathcal{L}_{SE(3)}$ makes the largest contribution to the reduction of error in both metrics. It is likely that $\mathcal{L}_{SE(3)}$ aims to bridge that gap between the representation of the input signal (head and hand global transformation matrices) and the representation of the output pose (global root orientation and relative joint rotations). + +Evaluating the effect of optimization. While the prediction of HMD-NeMo is very good, if the budget allows, it can be further optimized to improve accuracy (see supp mat for qualitative results). To bridge the gap between the predictions and observations, we optimize the pose prediction from HMD-NeMo, so that it matches the head and hand observations. As described in Section 3.4, in MC scenario, where both hands are always available, a simple optimization (with non-robust data energy term) loop can be used. The effect of such optimization in MC scenario is provided + +Table 5. Evaluating the effect of STAE module in HT scenario. + +
ConfigurationMPJPE ↓MPJVE ↓
Without optimization2.0726.07
With optimization1.9024.99
+ +Table 6. Effect of optimization in MC scenario, where both hands are always visible. + +
ConfigurationPriorityMPJPE ↓MPJVE ↓
No OptimizationPlausibility2.4831.30
Opt. with EnrFidelity2.3233.51
Opt. with ErBoth2.3731.33
+ +Table 7. Configuring HT scenario. Our model offers a range of choices to provide the best possible end-user experience despite incomplete observations. + +in Table 6. However, considering the same strategy for the HT scenario, where we may partially observe hands, this may not be optimal as we lose plausibility in the generated motions (captured by the MPJVE metric). As described in Section 3.4, depending on the scenario and experience requirements, one may choose to (1) avoid optimizing the predictions if plausibility is the highest priority (first row of Table 7), (2) use non-robust energy term if fidelity is the highest priority (second row of Table 7), or (3) use a robust energy term as a trade-off between fidelity and plausibility (third row of Table 7). + +Performance analysis. With 5.3M parameters, at inference time, HMD-NeMo requires only $4.4\mathrm{ms}$ to generate a pose given a HMD signal on a typical laptop CPU. On a NVIDIA Tesla P100 GPU, our model runs at 265 fps, with 1 iteration of optimization costs 3ms per frame (which could be further improved with more optimized implementation). Such performance makes HMD-NeMo a potential solution for HMD-driven avatar animation in immersive environments. + +# 5. Conclusion + +In this paper, we present HMD-NeMo, a unified approach to generate full-body avatar motion in both motion controller and hand tracking scenarios. To handle the unobserved hands in a temporally coherent and plausible manner, we introduce TAMT module. It is worth noting that, in this paper, we considered one major reason for partial hand visibility, i.e., hands appearing out of the FoV, however, in practice, hands may not be visible to the hand tracking camera due to failure in tracking and occlusion by another object or by another body part. While such cases are not considered in the data augmentation in our paper, they certainly can be taken into account and TAMT can be used to fill such gaps with no additional modification; this is left for future exploration. We provide extensive analyses on different components of HMD-NeMo, and shed light on various choices for optimization on top of neural network's predictions when it comes to production priorities (plausibility versus fidelity). + +# References + +[1] Karan Ahuja, Eyal Ofek, Mar Gonzalez-Franco, Christian Holz, and Andrew D Wilson. Coolmoves: User motion accentuation in virtual reality. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 5(2):1-23, 2021. +[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. FLAG: Flow-based 3D avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022. +[3] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019. +[4] Benjamin Biggs, Sébastien Ehrhadt, Hanbyul Joo, Benjamin Graham, Andrea Vedaldi, and David Novotny. 3D multibodies: Fitting sets of plausible 3D human models to ambiguous image data. 2020. +[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016. +[6] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015. +[7] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. +[8] Vasileios Choutas, Federica Bogo, Jingjing Shen, and Julien Valentin. Learning to fit morphable models. arXiv preprint arXiv:2111.14824, 2021. +[9] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating SMPL poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021. +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[11] Nima Ghorbani and Michael J Black. Soma: Solving optical marker-based mocap automatically. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11117-11126, 2021. +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. + +[13] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020. +[14] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018. +[15] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. AvatarPoser: Articulated full-body pose tracking from sparse motion sensing. In Proceedings of European Conference on Computer Vision. Springer, 2022. +[16] Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. International Conference on Learning Representations, ICLR, 2014. +[17] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11605-11614, 2021. +[18] Matthew Loper, Naureen Mahmood, and Michael J Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics (TOG), 33(6):1-13, 2014. +[19] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. +[20] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, Oct. 2019. +[21] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10975-10985, 2019. +[22] Jose Luis Ponton, Haoran Yun, Carlos Andujar, and Nuria Pelechano. Combining Motion Matching and Orientation Prediction to Animate Avatars for Consumer-Grade VR Devices. Computer Graphics Forum, 2022. +[23] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. +[24] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021. +[25] Unity Asset Store. Final IK. https://assetstoreunity.com/packages/tools/animation/final-ik-14290, 2021. +[26] Dorin Ungureanu, Federica Bogo, Silvano Galliani, Pooja Sama, Casey Meekhof, Jan Stuhmer, Thomas J Cashman, + +Bugra Tekin, Johannes L Schonberger, Pawel Olszta, et al. Hololens 2 research mode as a tool for computer vision research. arXiv preprint arXiv:2008.11239, 2020. +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[28] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3D human pose estimation from sparse IMUs. In Computer graphics forum, volume 36, pages 349-360. Wiley Online Library, 2017. +[29] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. arXiv preprint arXiv:2209.09391, 2022. +[30] Dongseok Yang, Doyeon Kim, and Sung-Hee Lee. LoBSTr: Real-time lower-body pose prediction from sparse upper-body tracking signals. In Computer Graphics Forum, volume 40, pages 265–275. Wiley Online Library, 2021. +[31] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical Inertial Poser (PIP): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13167-13178, 2022. +[32] Xinyu Yi, Yuxiao Zhou, and Feng Xu. TransPose: Real-time 3D human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. +[33] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision, pages 465-481. Springer, 2020. +[34] Mihai Zanfir, Andrei Zanfir, Eduard Gabriel Bazavan, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. THUNDR: Transformer-based 3D human reconstruction with markers. arXiv preprint arXiv:2106.09336, 2021. +[35] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3D bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. +[36] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. \ No newline at end of file diff --git a/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/images.zip b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b0722f1d198025ef282f7ed7bef7693daeaec2a8 --- /dev/null +++ b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eaa93796f0dee2a91e4f89baf4d3e5b309eeb86f9a7e8e94c896d0239cbf9ca +size 576442 diff --git a/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/layout.json b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac9a6a3e6bbf80ef9bd12da98348b0953f437d69 --- /dev/null +++ b/2023/HMD-NeMo_ Online 3D Avatar Motion Generation From Sparse Observations/layout.json @@ -0,0 +1,8548 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "type": "text", + "content": "HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 143, + 154, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 154, + 157 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 154, + 157 + ], + "type": "text", + "content": "Sadegh Aliakbarian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 172, + 143, + 246, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 143, + 246, + 156 + ], + "spans": [ + { + "bbox": [ + 172, + 143, + 246, + 156 + ], + "type": "text", + "content": "Fatemeh Saleh" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 144, + 332, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 144, + 332, + 156 + ], + "spans": [ + { + "bbox": [ + 264, + 144, + 332, + 156 + ], + "type": "text", + "content": "David Collier" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 350, + 144, + 445, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 144, + 445, + 156 + ], + "spans": [ + { + "bbox": [ + 350, + 144, + 445, + 156 + ], + "type": "text", + "content": "Pashmina Cameron" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 463, + 144, + 536, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 463, + 144, + 536, + 156 + ], + "spans": [ + { + "bbox": [ + 463, + 144, + 536, + 156 + ], + "type": "text", + "content": "Darren Cosker" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 171, + 424, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 171, + 424, + 185 + ], + "spans": [ + { + "bbox": [ + 169, + 171, + 424, + 185 + ], + "type": "text", + "content": "Microsoft Mixed Reality & AI Lab, Cambridge, UK" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": "Generating both plausible and accurate full body avatar motion is the key to the quality of immersive experiences in mixed reality scenarios. Head-Mounted Devices (HMDs) typically only provide a few input signals, such as head and hands 6-DoF. Recently, different approaches achieved impressive performance in generating full body motion given only head and hands signal. However, to the best of our knowledge, all existing approaches rely on full hand visibility. While this is the case when, e.g., using motion controllers, a considerable proportion of mixed reality experiences do not involve motion controllers and instead rely on egocentric hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view of the HMD. In this paper, we propose the first unified approach, HMD-NeMo, that addresses plausible and accurate full body motion generation even when the hands may be only partially visible. HMD-NeMo is a lightweight neural network that predicts the full body motion in an online and real-time fashion. At the heart of HMD-NeMo is the spatiotemporal encoder with novel temporally adaptable mask tokens that encourage plausible motion in the absence of hand observations. We perform extensive analysis of the impact of different components in HMD-NeMo and introduce a new state-of-the-art on AMASS dataset through our evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 561, + 128, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 128, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 128, + 573 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "content": "Mixed reality technology opens up new means of communication and interaction between people. With people at the heart of this technology, generating faithful and believable avatar motion is key to the quality of immersive experiences. Despite great advances in this area, generating full body avatar motion given HMD signal remains a challenge: in many current solutions, avatars only have upper bodies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Prior works attempted to generate full body avatar motion given sparse or partial observations, such as images [17, 4, 33], 2D joints/ keypoints [21, 5], markers [35, 18, 34, 11], and IMUs [28, 14, 32, 31]. While such observations are" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 323, + 220, + 523, + 332 + ], + "blocks": [ + { + "bbox": [ + 323, + 220, + 523, + 332 + ], + "lines": [ + { + "bbox": [ + 323, + 220, + 523, + 332 + ], + "spans": [ + { + "bbox": [ + 323, + 220, + 523, + 332 + ], + "type": "image", + "image_path": "ce46fd5a55b6b22f70a3dcbc66a4ca057221e6447660d2d6be9dfb1b5fef58e8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 332, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 305, + 332, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 545, + 364 + ], + "type": "text", + "content": "Figure 1. HMD-NeMo generates full body avatar motion given HMD signals, i.e., head and hand 6-DoFs, from hand tracking signal from the HMD as well as hand motion controllers." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 386, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 546, + 506 + ], + "type": "text", + "content": "considered partial or sparse, they provide much richer input signal compared to a typical HMD's head and hand 6-DoF. More recently, great progress has been made to generate full body motion given only a HMD signal [1, 30, 22, 15], however, they all rely on the availability and full visibility of both hands. While this is the case when, e.g., using motion controllers, many mixed reality experiences do not involve motion controllers and instead rely on hand tracking. This introduces the challenge of partial hand visibility owing to the restricted field of view (FoV) of the HMD sensors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 508, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 639 + ], + "type": "text", + "content": "In this paper, we address this problem via HMD-NeMo (a neural motion model of human given HMD signal). Within a unified framework, HMD-NeMo generates full body motion in real time, regardless of whether hands are fully or partially observed, or not observed at all. Our approach is built upon recurrent neural networks to efficiently capture temporal information, and a transformer to capture complex relations between different components of the input signal. At the heart of our approach is the TAMT (temporally adaptable mask token) module, allowing us to handle missing hand observations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Our contributions are: (1) The first full body avatar motion generation approach capable of generating accurate and plausible motions with full or partial hand visibility. This is a step forward for unlocking fully immersive experiences in mixed reality with fewer limitations on the hardware. (2) Temporally adaptable Mask Tokens (TAMT), a simple yet" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9622" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 127, + 182 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 127, + 182 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 127, + 182 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 127, + 182 + ], + "type": "image", + "image_path": "345499123601787b2ca1afce709a8cb33b7218f1914b825766d56e427703ad24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 183, + 108, + 190 + ], + "lines": [ + { + "bbox": [ + 67, + 183, + 108, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 183, + 108, + 190 + ], + "type": "text", + "content": "(a) Plane FoV" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 227, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 304 + ], + "type": "text", + "content": "Figure 2. Top: Definition of FoV for different scenarios for the same pose. (a) Planar FoV, as used in [8], wherein avatar's left hand is visible while the right hand is not. (b) Fully visible, as in motion controller scenarios, wherein both hands are always visible. (c) HMD's hand tracking camera's FoV, as in hand tracking scenarios, wherein avatar's right hand is visible while the left hand is not. Bottom: Hand visibility statistics on AMASS test set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 209, + 182 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 209, + 182 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 209, + 182 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 209, + 182 + ], + "type": "image", + "image_path": "7a8bad323a803924405405444bfb6e52b44ce1efb36212b2ced535b376566369.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 182, + 211, + 191 + ], + "lines": [ + { + "bbox": [ + 129, + 182, + 211, + 191 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 211, + 191 + ], + "type": "text", + "content": "(b) Motion Controllers (MC)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 70, + 291, + 182 + ], + "blocks": [ + { + "bbox": [ + 211, + 70, + 291, + 182 + ], + "lines": [ + { + "bbox": [ + 211, + 70, + 291, + 182 + ], + "spans": [ + { + "bbox": [ + 211, + 70, + 291, + 182 + ], + "type": "image", + "image_path": "748ce4138b12aa35c14f7f340ad4379be726af0163d5a198fe7de3d5436aad58.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 182, + 285, + 191 + ], + "lines": [ + { + "bbox": [ + 216, + 182, + 285, + 191 + ], + "spans": [ + { + "bbox": [ + 216, + 182, + 285, + 191 + ], + "type": "text", + "content": "(c) Hand Tracking (HT)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 66, + 191, + 269, + 225 + ], + "blocks": [ + { + "bbox": [ + 66, + 191, + 269, + 225 + ], + "lines": [ + { + "bbox": [ + 66, + 191, + 269, + 225 + ], + "spans": [ + { + "bbox": [ + 66, + 191, + 269, + 225 + ], + "type": "table", + "html": "
ScenarioLeft hand visibilityRight hand visibility
Motion Controllers100%100%
Hand Tracking53.45%48.94%
", + "image_path": "67ee74be52b5cab033b3b18faf4cd70b55f776c151978bc12bdb6c251835753b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 323, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 373 + ], + "type": "text", + "content": "effective strategy for handling missing hand observations in a temporally coherent way. (3) Extensive experiments and ablations of our method as well as a new state-of-the-art performance on the challenging AMASS dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 382, + 134, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 134, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 134, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 402, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 641 + ], + "type": "text", + "content": "Over the past few years, different solutions have been proposed to the problem of full body pose generation given sparse or partial observations, such as markers [35, 18, 34, 11], images [17, 4, 33], 2D joints/keypoints [21, 5], Inertial Measurement Units (IMUs) [28, 14, 32, 31], and HMDs [2, 1, 30, 22, 15]. Among these works, the ones utilizing wearable devices are closer to our approach and thus we discuss them here. Recently, different techniques have been proposed for human pose reconstruction using a sparse set of IMUs attached to the body [28, 14, 32, 31]. While attempts have been made to come up with a minimum number of IMUs for generating full body motion, typically an IMU sensor is used near pelvis, making the problem relatively easy compared to HMD scenarios where the root/pelvis signal is not available. Whilst external sensors, e.g., IMUs [28, 14, 32, 31] and cameras [23] are effective, they are often not as accessible as wearing a HMD. Using only a HMD is desirable from a usability point of view, but generating realistic and faithful human representations from such inputs remains technically challenging." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": "Given HMD signals, [2] and [8] generate full body poses. While they generate expressive poses faithful to the HMD signal, these approaches [2, 8] only predict static poses, lacking the temporal consistency required for avatar motion generation. While [8] generates poses in the world space, [2] predicts the pose relative to the root. Thus [2]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "assumes a known root (pelvis joint) as an additional signal. The assumption of known root joint also appears in [9], wherein, unlike [2, 8], the method generates full body motion given the HMD signal with Variational Autoencoders [16]. Note that, while [9] generates temporally plausible motions, it works in an offline fashion, predicting the motion for an entire sequence only after observing the whole sequence of HMD signals." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 169, + 546, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 546, + 419 + ], + "type": "text", + "content": "Recently, different techniques have been proposed to generate full body avatar locomotion (in the world coordinates) given HMD signals. In this context, [1] proposes a matching algorithm, aiming to sample closest poses from a motion capture library at sparse time-steps and interpolate between poses. While this guarantees the selection of realistic poses, the output is always limited to the utilized motion capture library. In another work, [30] uses combination of an inverse kinematic (IK) solver and a recurrent neural network to generate upper body and lower body motions, respectively. Combination of different components to solve lower and upper body separately has also been explored in [22], wherein a neural network is trained to predict the root orientation given the HMD signal, which then is used as the feature vector for a motion matching algorithm [6, 13] to generate full body. More recently, fully learning-based approaches have shown promise in generating full body avatar motion [29, 15]. In this context, [29] simulates plausible and physically feasible motions within a reinforcement learning framework and [15] uses a transformer-based approach to generate full body motion given HMD signals." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 421, + 546, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 546, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 546, + 577 + ], + "type": "text", + "content": "Although great progress has been made by recent approaches, they all solve the problem of motion generation given head and both hands, typically captured with motion controllers. However many mixed reality experiences do not have motion controllers available and instead rely on hand tracking from HMD mounted sensors (e.g., cameras). This introduces the challenge of hand tracking failures and partial hand visibility owing to the restricted field of views. To the best of our knowledge, despite its usability, motion generation in the presence of partial hand observation has not been well-explored. In this paper, we propose a method capable of generating high fidelity and plausible full body motion even in presence of partially visible hands." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 590, + 411, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 411, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 411, + 604 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 647 + ], + "type": "text", + "content": "In this section, we first define the problem, the scenarios we consider, as well as the input and the desired output representation. We then present our proposed method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 658, + 417, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 417, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 417, + 670 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Task. The task is to generate full-body 3D human locomotion (predicting both the instantaneous pose and the global trajectory of the human) given the sparse HMD signal in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9623" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 78, + 496, + 198 + ], + "blocks": [ + { + "bbox": [ + 97, + 78, + 496, + 198 + ], + "lines": [ + { + "bbox": [ + 97, + 78, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 97, + 78, + 496, + 198 + ], + "type": "image", + "image_path": "49f341d087760e6acd468b567509dd7b1ac36f7d0c373122a7b2cfa482ce326b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": "Figure 3. Overview of HMD-NeMo: At each time-step, the model gets as input the HMD signal (left), as in Eq. 1. Such input is the mapped to an embedding space via " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": ", acting as the input to the spatio-temporal encoder module STAE. The resulting feature representation is then utilized by two autoregressive decoders, " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "g\\gamma" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " to predict the global trajectory as well as the pose, respectively. In case the model does not observe a hand (in HT scenario), " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "f_{amt}" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " fills in the representation of the unobserved hand effectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "text", + "content": "an online fashion1. That is, given the input signal " + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "text", + "content": " at each time-step " + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "text", + "content": ", the system should predict the 3D human pose and trajectory " + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "text", + "content": " near-instantaneously. HMD-NeMo achieves this using a neural network parameterized by " + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 264, + 287, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 313, + 288, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 516 + ], + "type": "text", + "content": "Scenarios. In this paper, we consider two scenarios: Motion Controllers (MC) scenario, wherein hands are always tracked via motion controllers using constellation tracking, as illustrated in Fig. 2 (b), and Hand Tracking (HT) scenario, wherein hands are tracked via a visual hand tracking system whenever the hands are inside the FoV of the device, as illustrated in Fig. 2 (c). The FoV of the device is defined as a frustum determined by the HMD's hand tracking camera placement and parameters. In this paper, we use a similar FoV to that of the Microsoft HoloLens 2 [26]. Note that, while the MC scenario has been explored recently [1, 30, 22, 15], to the best of our knowledge, this work constitutes the first approach that tackles both MC and HT scenarios within one unified framework2. The HT scenario is particularly challenging as hands tend to be out of FoV almost " + }, + { + "bbox": [ + 46, + 313, + 288, + 516 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 313, + 288, + 516 + ], + "type": "text", + "content": " of the time, according to the statistics presented in the table in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": "Input Representation. The input signal " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": " contains the head 6-DoF " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "x_{h} \\in \\mathbb{R}^{(6 + 3)}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": ", the left hand 6-DoF " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "x_{l}^{t} \\in \\mathbb{R}^{(6 + 3)}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": ", and the right hand 6-DoF " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "x_{r}^{t} \\in \\mathbb{R}^{(6 + 3)}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": ", all in the world space. We use the 6D representation to represent the rotations [36]. We additionally provide the hand representations in the head space, " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "x_{lh}^{t} \\in \\mathbb{R}^{(6 + 3)}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "x_{rh}^{t} \\in \\mathbb{R}^{(6 + 3)}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": ". In the HT scenario, hands may go in and out of FoV of the HMD, so we also provide HMD-NeMo with the hand visibility status for both left and right hand, " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\nu_{l}^{t}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\nu_{r}^{t}" + }, + { + "bbox": [ + 46, + 517, + 287, + 636 + ], + "type": "text", + "content": ", as binary values, 1 being visible and 0 otherwise. Finally, for all 6-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "content": "DoF signals, we provide the velocity of changes between two consecutive frames. Specifically, for translations we consider " + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "inline_equation", + "content": "vel(\\mathcal{T}^t,\\mathcal{T}^{t - 1}) = \\mathcal{T}^t -\\mathcal{T}^{t - 1}" + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "content": " is the translation, and for rotations we consider the geodesic changes in the rotation " + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "inline_equation", + "content": "vel(R^{t},R^{t - 1}) = (R^{t - 1})^{-1}R^{t}" + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "content": " is the rotation, which together constitute the velocity 6-DoF " + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\dot{x}^t" + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "content": ". Overall, the input to the HMD-NeMo, " + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "inline_equation", + "content": "x^{t}\\in \\mathbb{R}^{92}" + }, + { + "bbox": [ + 306, + 264, + 545, + 358 + ], + "type": "text", + "content": ", can be written as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 361, + 545, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 545, + 386 + ], + "type": "interline_equation", + "content": "x ^ {t} = \\left\\{x _ {h}, x _ {l}, x _ {r}, x _ {l h}, x _ {r h}, \\dot {x} _ {h}, \\dot {x} _ {l}, \\dot {x} _ {r}, \\dot {x} _ {l h}, \\dot {x} _ {r h}, \\nu_ {l}, \\nu_ {r} \\right\\} ^ {t} \\tag {1}", + "image_path": "8a34085283b0defc1f5a9d475e215a2c482447859b52bda90363171d9215b08e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "text", + "content": "Output Representation. The output of HMD-NeMo contains the pose (including the root orientation), " + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\theta^t\\in \\mathbb{R}^{J\\times 3}" + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "text", + "content": " represented with axis-angle rotations for the " + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "text", + "content": " joints in the body, and the global position in the world, " + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\gamma^t\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "text", + "content": " represented as the root position, resulting in " + }, + { + "bbox": [ + 305, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "y^{t}\\in \\mathbb{R}^{(J + 1)\\times 3}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 399, + 451, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 451, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 399, + 451, + 545, + 464 + ], + "type": "interline_equation", + "content": "y ^ {t} = \\{\\theta , \\gamma \\} ^ {t} \\tag {2}", + "image_path": "542e82df29027a1e9f8014c80107b2d4b4bf95829a543d7efb33bf8e5a15063b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": "The sequence of " + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\theta^{0:T}" + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\gamma^{0:T}" + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": " then represent the avatar motion as well as its global trajectory for the period " + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "[0,T]" + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": ". Note that, in this paper, we may drop the time superscript " + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 468, + 545, + 517 + ], + "type": "text", + "content": " for better readability and include it when necessary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 388, + 538 + ], + "type": "text", + "content": "3.2. HMD-NeMo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Overview. HMD-NeMo's pipeline is illustrated in Fig. 3. As described in Section 3.1, the model gets as input the information about the head and hands in world coordinate system, the hands expressed in the head space, as well as their velocities, as described in Eq. 1. To express hands in head space, we consider " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{lh} = x_h^{-1}x_l" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " (similarly, " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{rh} = x_h^{-1}x_r" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "). This representation then acts as the input to an embedding layer, " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_e" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", which aims to (1) map the raw input to an embedding space and (2) handle the unobserved hands. Given the input " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_e" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", the next module, STAE, learns (a) how each representation evolves over time and (b) how different components of the input, i.e., head and hands, are correlated. Once such a rich representation is obtained, two auto-regressive decoders, " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_\\theta" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_\\gamma" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", generate the body pose and the global" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 645, + 287, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 645, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 645, + 287, + 674 + ], + "type": "text", + "content": "1Note that, similar to existing work on HMD-driven motion generation [1, 15, 9], our approach does not predict body shape parameters and only focuses on the poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 674, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 674, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 287, + 712 + ], + "type": "text", + "content": "2Note that [8] also considers a form of partial hand visibility by defining a plane FoV, wherein, as illustrated in Fig. 2 (a), joints in front of head is are considered visible. However, this scenario does not apply to practical use cases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9624" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 297, + 150 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 297, + 150 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 297, + 150 + ], + "type": "image", + "image_path": "0abb9877a3e991e66204a2b2dc7af6be195d29e3b50a9970f0d8c61d60b93478.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 153, + 288, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 288, + 187 + ], + "type": "text", + "content": "Figure 4. Overview of the embedding module. Note that " + }, + { + "bbox": [ + 46, + 153, + 288, + 187 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 46, + 153, + 288, + 187 + ], + "type": "text", + "content": " is only provided as an example and this module applies to all 6-DoF inputs appearing in " + }, + { + "bbox": [ + 46, + 153, + 288, + 187 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 46, + 153, + 288, + 187 + ], + "type": "text", + "content": " as in Eq. 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 208, + 286, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 286, + 280 + ], + "type": "text", + "content": "position of the avatar, respectively. At each time step, the output of STAE is used to update the mask tokens (described below) as a representation for the hand signals that may be missing in the next time-step. To aid training, we also include an auxiliary task of human pose reconstruction in SE(3), denoted by " + }, + { + "bbox": [ + 46, + 208, + 286, + 280 + ], + "type": "inline_equation", + "content": "g_{aux}" + }, + { + "bbox": [ + 46, + 208, + 286, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 286, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 286, + 388 + ], + "type": "text", + "content": "Head and hand embedding " + }, + { + "bbox": [ + 46, + 281, + 286, + 388 + ], + "type": "inline_equation", + "content": "(f_e)" + }, + { + "bbox": [ + 46, + 281, + 286, + 388 + ], + "type": "text", + "content": ". As shown in Fig. 4, the embedding module " + }, + { + "bbox": [ + 46, + 281, + 286, + 388 + ], + "type": "inline_equation", + "content": "f_e" + }, + { + "bbox": [ + 46, + 281, + 286, + 388 + ], + "type": "text", + "content": " gets as input the head and hands 6-DoFs and velocities and maps them to a higher-dimensional embedding space. As the range of values corresponding to the rotations is different from those of the translations, we decouple such information and embed them via separate shallow MLPs and concatenate the results back together. For instance, for the observed left hand in the world coordinate system, the embedding representation is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 388, + 293, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 293, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 293, + 426 + ], + "type": "interline_equation", + "content": "e _ {l} ^ {\\text {v i s i b l e}} = \\left[ f _ {e} ^ {\\mathrm {R}} \\left(x _ {l} [: 6 ]\\right), f _ {e} ^ {\\mathrm {T}} \\left(x _ {l} [ 6 : ]\\right), \\dot {f} _ {e} ^ {\\mathrm {R}} \\left(\\dot {x} _ {l} [: 6 ]\\right), \\dot {f} _ {e} ^ {\\mathrm {T}} \\left(\\dot {x} _ {l} [ 6 : ]\\right) \\right] \\tag {3}", + "image_path": "4ccdd04c22769cabc1f55cb50ed2264e82dc619e9b484efb8d33c083248673f8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "inline_equation", + "content": "f_{e}^{\\mathrm{R}}" + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "inline_equation", + "content": "f_{e}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "text", + "content": " are MLPs responsible for computing the rotation and translation embedding, acting on the first 6 elements (the 6D rotation representation) and the last 3 elements (the translation) of the input, respectively (similarly for " + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "inline_equation", + "content": "\\dot{f}_{e}^{\\mathrm{R}}" + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "inline_equation", + "content": "\\dot{f}_{e}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 430, + 286, + 491 + ], + "type": "text", + "content": " which act on velocities)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "text", + "content": "In the HT scenario, hands may not be visible to the model, hence computing such embedding representation is not possible. Thus, given the status of " + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\nu_{l}" + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\nu_{r}" + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "text", + "content": ", the embedding module decides to either compute the embedding or utilize the output of the " + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "inline_equation", + "content": "f_{\\text{tamt}}" + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "text", + "content": " (described below), a set of temporally adaptable mask tokens, instead of a missing hand observation (denoted by " + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 491, + 286, + 598 + ], + "type": "text", + "content": " in Fig. 3). As illustrated in Fig 4, the embedding of the left hand in the world coordinate system can be computed as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 95, + 600, + 286, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 600, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 95, + 600, + 286, + 613 + ], + "type": "interline_equation", + "content": "e _ {l} = \\nu_ {l} e _ {l} ^ {\\text {v i s i b l e}} + (1 - \\nu_ {l}) f _ {\\text {t a m t}} (l). \\tag {4}", + "image_path": "1209f4d1f1cb32abc8591cb2814611ef0c2b745eefe28238bd969af7656ed18e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Spatio-temporal encoder (STAE). The output of " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " on each component of the input stream is a non-temporal feature, computed independent of other components in the input. While an expressive representation of each component, it lacks temporal and spatial correlation information. We specifically care about these characteristics because the model is required to generate temporally coherent motion and also because the motion of one body part often impacts" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 545, + 182 + ], + "type": "image", + "image_path": "565a806dbe0ce8def6ffc9de4e93088fae87c02e4237a23e1b4e0d93dd9262c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 186, + 522, + 198 + ], + "lines": [ + { + "bbox": [ + 306, + 186, + 522, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 522, + 198 + ], + "type": "text", + "content": "Figure 5. Overview of the spatio-temporal encoder module." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 278 + ], + "type": "text", + "content": "or determines the motion of other body part(s). To obtain a more informative representation from the head and hands, HMD-NeMo first learns the isolated temporal features of each component of input representation and then learns how they are spatially correlated [27]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "content": "As illustrated in Fig. 5, to learn the temporal representation of the input signal, we use gated recurrent units (GRUs) [7]. With a GRU module on top of each component in the input, the model learns how each component, e.g., head, evolves over time, independent of other components in the input. This information is compressed in the hidden state of the GRU cell, " + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "content": ", which is then utilized to learn how different components in the input relate to each other. This is achieved by using a standard transformer encoder on the GRU hidden states, thanks to the self-attention mechanism. Temporally adaptable mask tokens (TAMTs). As discussed in Section 3.1, hands may not be visible to the model, and thus, there is no representative input signal for the " + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "content": " module. To address this issue, in case of a missing hand observation, our model produces a feature vector, " + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "inline_equation", + "content": "f_{amt}" + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "content": ", to represent the missing hand observation. To compute " + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "inline_equation", + "content": "f_{amt}" + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "content": ", as illustrated in Fig. 6, we use the output of STAE for the hand observation that may be missing in the next time step as well as the output of STAE for the head. Note that head joint is the reference joint and is always available. The combination of these two features is a rich representation of the past state of the missing hand signal (both temporally and spatially); this is then used to compute the " + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "inline_equation", + "content": "f_{amt}" + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "content": ". In order to encourage " + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "inline_equation", + "content": "f_{amt}" + }, + { + "bbox": [ + 304, + 279, + 545, + 615 + ], + "type": "text", + "content": " to learn information about the missing hand observation, as illustrated by the Forecaster module in Fig. 6, we introduce a forecasting auxiliary task to forecast the state (6-DoF) of the corresponding hand in the next time-step." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 623, + 432, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 432, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 432, + 635 + ], + "type": "text", + "content": "3.3. Training HMD-NeMo" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "To train HMD-NeMo, we rely on the availability of a motion capture dataset, represented as SMPL [19] parameters (pose, shape, and global trajectory). For each sequence, we then simulate the HMD on the subject. In case of HT, we also simulate a FoV frustum to be able to model the hand visibility status " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\nu_{l}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\nu_{r})" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ". We then train HMD-NeMo with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9625" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 157, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 157, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 157, + 83 + ], + "type": "text", + "content": "a loss function of the form" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 86, + 287, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 86, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 118, + 86, + 287, + 99 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha_ {\\text {d a t a}} \\mathcal {L} _ {\\text {d a t a}} + \\alpha_ {\\text {s m o o t h}} \\mathcal {L} _ {\\text {s m o o t h}} \\tag {5}", + "image_path": "6f8d0f2c493dcee0479c4196fd2db9709aecaefd7740b0ec5b0ff4c243bbdb95.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 101, + 266, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 101, + 266, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 266, + 114 + ], + "type": "interline_equation", + "content": "+ \\alpha_ {S E (3)} \\mathcal {L} _ {S E (3)} + \\alpha_ {\\text {f o r e c a s t}} \\mathcal {L} _ {\\text {f o r e c a s t}} + \\alpha_ {\\text {a u x}} \\mathcal {L} _ {\\text {a u x}}", + "image_path": "995318f59a8ec5c0a8c74201886fdde274e3541d25b739aa5e9de7d14dcbf35d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 117, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 287, + 140 + ], + "type": "text", + "content": "The data loss term is the squared error between the predicted pose and trajectory and those of the ground truth," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 142, + 287, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 142, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 87, + 142, + 287, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d a t a}} = \\sum_ {t = 1} ^ {T} | | \\hat {\\theta} ^ {t} - \\theta^ {t} | | _ {2} ^ {2} + | | \\hat {\\gamma} ^ {t} - \\gamma^ {t} | | _ {2} ^ {2} \\tag {6}", + "image_path": "4f08b5b5bb6de5fa6b7769c5d8d23685c1402560b4749069715c4855febd7efa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 177, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 287, + 236 + ], + "type": "text", + "content": "Note that, in practice, the pose decoder has two heads, one each for predicting the body pose and the global root orientation. To further enhance the temporal smoothness, we penalize the discrepancy between the velocity of changes in the prediction to that of the ground truth" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 238, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 238, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 72, + 238, + 287, + 270 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s m o o t h}} = \\sum_ {t = 2} ^ {T} | | \\delta \\hat {\\theta} ^ {t} - \\delta \\theta^ {t} | | _ {1} + | | \\delta \\hat {\\gamma} ^ {t} - \\delta \\gamma^ {t} | | _ {1} \\tag {7}", + "image_path": "bbd8f415cd2c9446d77abc96ccac2de2254fb636126d282107fb5ef0795ca2b8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\delta \\hat{\\theta}^t = \\hat{\\theta}^t -\\hat{\\theta}^{t - 1}" + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\delta \\hat{\\gamma}" + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\delta \\theta" + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\delta \\gamma" + }, + { + "bbox": [ + 46, + 274, + 287, + 383 + ], + "type": "text", + "content": " follow similarly). In addition to computing the reconstruction loss on the SMPL parameters, i.e., to relative joint rotations, we found it extremely useful to also utilize the reconstruction loss of each joint transformation independent of its parent, i.e., in the world space. To compute this reconstruction loss, we use the SMPL model to compute the joint transformations in SE(3) given the predicted and ground truth pose and trajectory parameters. Thus, the SE(3) reconstruction loss is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 384, + 287, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 384, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 93, + 384, + 287, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S E (3)} = \\sum_ {t = 1} ^ {T} \\left| \\left| \\hat {P} _ {S E (3)} ^ {t} - P _ {S E (3)} ^ {t} \\right| \\right| _ {2} ^ {2} \\tag {8}", + "image_path": "c0d1f73c68a887a8e0d37f5988226d51bb9bc695b51fe3345dd093bda4d9048d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 418, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 287, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 418, + 287, + 466 + ], + "type": "inline_equation", + "content": "P_{SE(3)}" + }, + { + "bbox": [ + 47, + 418, + 287, + 466 + ], + "type": "text", + "content": " is the body pose in SE(3). The next loss term corresponds to the forecasting auxiliary task in the TAMT module, where the goal is to minimize the distance between the predicted next hand and the ground truth next hand," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 93, + 468, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 468, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 93, + 468, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f o r e c a s t}} = \\sum_ {t = 2} ^ {T} \\sum_ {j \\in \\{l, r \\}} \\left\\| \\hat {x} _ {j} ^ {t} - x _ {j} ^ {t} \\right\\| _ {2} ^ {2} \\tag {9}", + "image_path": "c80aeebf9e4c4ee346e2887bfb3dba181161c4934034a1d716732944017c00c1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 505, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 287, + 552 + ], + "type": "text", + "content": "Finally, we have our loss term for the auxiliary task, aiming to minimize the predicted full body joint transformations from STAE's features, " + }, + { + "bbox": [ + 47, + 505, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\hat{P}_{aux}" + }, + { + "bbox": [ + 47, + 505, + 287, + 552 + ], + "type": "text", + "content": ", to the ground truth body joint transformations," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 101, + 553, + 287, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 553, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 101, + 553, + 287, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a u x} = \\sum_ {t = 1} ^ {T} \\left| \\left| \\hat {P} _ {a u x} ^ {t} - P _ {S E (3)} ^ {t} \\right| \\right| _ {2} ^ {2} \\tag {10}", + "image_path": "5ee92c34f17c58a41cc7450aa99c6625c587dc480fffa6e393cde4cc0988e853.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 587, + 270, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 270, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 270, + 600 + ], + "type": "text", + "content": "Our model is trained with " + }, + { + "bbox": [ + 47, + 587, + 270, + 600 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 587, + 270, + 600 + ], + "type": "text", + "content": "s all being set to 1 in Eq. 6." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 610, + 132, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 132, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 132, + 623 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 713 + ], + "type": "text", + "content": "Once trained, HMD-NeMo is capable of generating high fidelity and plausible human motion given only the HMD signal. However, as is typical of learning-based approaches, the direct prediction of the neural network does not precisely match the observations i.e., the head and hands, even if it is perceptually quite close. To close this gap between the prediction and the observation, optimization can" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 548, + 146 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 548, + 146 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 548, + 146 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 548, + 146 + ], + "type": "image", + "image_path": "cae97331164c5d976d9cf8b1cfa682b4b9436eaae00271e10a2dbe3c1521afe7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 152, + 547, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 152, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 547, + 185 + ], + "type": "text", + "content": "Figure 6. Overview of TAMT module. Note that here we illustrate TAMT of " + }, + { + "bbox": [ + 305, + 152, + 547, + 185 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 305, + 152, + 547, + 185 + ], + "type": "text", + "content": " as an example, the same applies f to both hands in both head space and world space." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": "be used. This adjusts the pose parameters to minimize an energy function of the form " + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\mathcal{E}_{data} + \\mathcal{E}_{reg}" + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{data}" + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": " is the energy term that minimizes the distance between the predicted head and hands to the observed ones, and " + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{reg}" + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": " is additional regularization term(s). To define the data energy term, we define the residual " + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\sum_{j\\in \\{h,l,r\\}}(x_j - \\hat{x}_j)" + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": ", i.e., the difference between the predicted head/hand joint to that of the observation. Given " + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": ", a typical, non-robust data energy term could be written as " + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{nr} = \\mathcal{R}^2" + }, + { + "bbox": [ + 304, + 205, + 545, + 396 + ], + "type": "text", + "content": ", i.e., the L2 loss. This suits the MC scenario perfectly, where head, left hand, and right hand are always available. But this energy term may be misleading in HT scenario where hands are going into and out of FoV often, and thus leading to abrupt pose changes when hands appear back in the FoV (see supp mat for further details). To remedy this issue, we utilize a more robust [3] alternative to the data energy term," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 310, + 397, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 397, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 310, + 397, + 545, + 429 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {r} (\\mathcal {R}, a, b, c) = b \\frac {| a - 2 |}{a} \\left(\\left(\\frac {\\left(\\frac {\\mathcal {R}}{c}\\right) ^ {2}}{| a - 2 |} + 1\\right) ^ {\\left(\\frac {a}{2}\\right)} - 1\\right) \\tag {11}", + "image_path": "a3d8917b6a32c6994e7bc2ae92daaa79273a127fb399f2ae5a08f88cfeac4662.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "text", + "content": " are hyper-parameters that determine the shape of the loss (see supp mat for further details). Unlike " + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{nr}" + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_r" + }, + { + "bbox": [ + 304, + 431, + 545, + 563 + ], + "type": "text", + "content": ", considers large discrepancies between the prediction and observation as outliers, not penalizing the prediction strongly and thus does not push the prediction to move toward the observation. Thus abrupt changes in the arm poses are avoided and optimization stays on course despite large variation in the velocity metric (when caused by hand visibility changes). Of course, utilizing Eq. 11 adversely affects the fidelity, but a trade-off between the plausibility and fidelity can be chosen to suit the application of interest." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 563, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 610 + ], + "type": "text", + "content": "Note that, since all observations relate to the upper body, during optimization we only optimize the upper body pose parameters and global root trajectory, while keeping the predicted lower body untouched." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 621, + 387, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 387, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 387, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 642, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 677 + ], + "type": "text", + "content": "In this section, we introduce the dataset and metrics, with implementation details in the supplementary material. We then present the experimental results and ablation studies." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Dataset. We follow the recent common practice [15, 24, 8, 9] of using a subset of AMASS [20] for training and evaluation. AMASS is a large collection of human motion se" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9626" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 527, + 300 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 527, + 300 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 527, + 300 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 527, + 300 + ], + "type": "image", + "image_path": "274d7d51940aebbc75701ddf3ad68738e31cbf0985e4457687a0733ef5c8c44d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 547, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 322 + ], + "type": "text", + "content": "Figure 7. Comparison to the state-of-the-art in MC scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). Last row depicts a hard example with complex body pose and motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "content": "quences, converted to 3D human meshes in the SMPL [19] representation, wherein every motion sequence contains information about poses " + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "content": " and the global trajectory " + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "content": ". To synthesize the HMD scenario, we compute the global transformation matrices for the head and hands as input. In the case of hand tracking, we define a FoV for the HMD and mask out the hands whenever they are out of FoV. To make a fair comparison, for both training and evaluation, we follow the splits suggested by [15]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": "Metrics. To evaluate the performance of our approach as well as the competing baselines, we report the mean per-joint position error (MPJPE [cm]) and the mean per-joint velocity error (MPJVE [cm/s]). We compare our approach with recent techniques [15, 1, 30, 9]" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 449, + 288, + 593 + ], + "type": "text", + "content": ". As these approaches do not tackle the hand tracking scenario, we compare HMD-NeMo against them for the motion controller scenario. We evaluate HMD-NeMo for hand tracking scenarios separately. In our experiments, we do not use the ground truth body shape parameters, but instead use the same default shape for all sequences. This follows the evaluation used in previous work [1, 15, 9]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 598, + 231, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 231, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 231, + 611 + ], + "type": "text", + "content": "4.1. Comparison to the state-of-the-art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 616, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 677 + ], + "type": "text", + "content": "We compare HMD-NeMo with existing approaches that tackle the problem of full-body motion generation given HMD signal. To the best of our knowledge, all existing approaches only tackle the MC scenario, and thus we compare them with this setting4. As demonstrated in Table 1," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 338, + 536, + 429 + ], + "blocks": [ + { + "bbox": [ + 317, + 338, + 536, + 429 + ], + "lines": [ + { + "bbox": [ + 317, + 338, + 536, + 429 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 536, + 429 + ], + "type": "table", + "html": "
MethodMPJPE ↓MPJVE ↓
FinalIK [25]18.0959.24
Ahuja et al. [1]7.83100.54
Yang et al. [30]9.0244.97
Dittadi et al. [9]6.8337.99
Jiang et al. [15]4.1829.40
HMD-NeMo (Ours)1.9024.99
", + "image_path": "8d5f299817a1ea416e32bccb2ef2d0a7b37adb77f4cf0f0369a3b2c7f3f0cd33.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 429, + 545, + 450 + ], + "lines": [ + { + "bbox": [ + 306, + 429, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 545, + 450 + ], + "type": "text", + "content": "Table 1. Comparison to the state-of-the-art approaches in MC scenario, where both hands are always visible." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 470, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 546, + 639 + ], + "type": "text", + "content": "HMD-NeMo is not only more accurate (lower MPJPE), but also generates smoother motion with joint velocities similar to that in the ground truth (lower MPJVE), introducing a new state-of-the-art on the AMASS dataset for both metrics. Note that [25] has the highest errors, as shown in Table 1 as it only optimizes the pose of the head and hands and ignores the accuracy and smoothness of the rest of the joints. Consequently, since [30] utilizes FinalIK within its framework, its performance becomes bounded by the quality of FinalIK. The learning-based approaches [15, 9], however, perform much better than [25, 1, 30], highlighting the value of data-driven methods trained on large-scale motion capture datasets. In Fig. 7, we compared HMD-NeMo with the second best performing baseline [15]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 647, + 407, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 407, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 407, + 658 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "In this section, we comprehensively evaluate different aspects of HMD-NeMo, including evaluation in HT scenario, cross-dataset evaluations, ablation studies on input signals, model architectures, and loss terms. We also ana" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 683, + 283, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 283, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 283, + 693 + ], + "type": "text", + "content": "3The approach [9] is modified to predict motion in world coordinates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "4Note, we exclude very recent approaches [22, 29] since they either do not report on AMASS or the code is not publicly available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9627" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 187, + 157 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 187, + 157 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 187, + 157 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 187, + 157 + ], + "type": "image", + "image_path": "74740f3bbe1daa84c1047f5b89d4dd93e2b893fc94d4925594536802dba4a63d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 159, + 157, + 167 + ], + "lines": [ + { + "bbox": [ + 81, + 159, + 157, + 167 + ], + "spans": [ + { + "bbox": [ + 81, + 159, + 157, + 167 + ], + "type": "text", + "content": "(a) Walk in a semi-circle" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 188, + 77, + 339, + 155 + ], + "blocks": [ + { + "bbox": [ + 188, + 77, + 339, + 155 + ], + "lines": [ + { + "bbox": [ + 188, + 77, + 339, + 155 + ], + "spans": [ + { + "bbox": [ + 188, + 77, + 339, + 155 + ], + "type": "image", + "image_path": "5204f3aa56deab36440cd4ebda798cc8953e9a372adfafa5f1677529dc2fcf96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 159, + 297, + 167 + ], + "lines": [ + { + "bbox": [ + 230, + 159, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 230, + 159, + 297, + 167 + ], + "type": "text", + "content": "(b) Walk, then knock" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 341, + 76, + 431, + 157 + ], + "blocks": [ + { + "bbox": [ + 341, + 76, + 431, + 157 + ], + "lines": [ + { + "bbox": [ + 341, + 76, + 431, + 157 + ], + "spans": [ + { + "bbox": [ + 341, + 76, + 431, + 157 + ], + "type": "image", + "image_path": "3e31aae32e2b9e2fb2f868fb598e5b76a8453d84a4d028c2605b929a5ccbe21c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 74, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 337, + 159, + 434, + 167 + ], + "lines": [ + { + "bbox": [ + 337, + 159, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 337, + 159, + 434, + 167 + ], + "type": "text", + "content": "(c) Walking, then put sth down" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 74, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 433, + 74, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 433, + 74, + 545, + 159 + ], + "type": "image", + "image_path": "927eb99c137d9274f07cb583ee074eb4929881ac4dad9562cd08b2981b98f83b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 159, + 517, + 167 + ], + "lines": [ + { + "bbox": [ + 460, + 159, + 517, + 167 + ], + "spans": [ + { + "bbox": [ + 460, + 159, + 517, + 167 + ], + "type": "text", + "content": "(d) Walk, then sit" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 171, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 547, + 194 + ], + "type": "text", + "content": "Figure 8. Qualitative results in HT scenario. Vertices are color-coded based on the distance to the GT (blue for low error and yellow for high error). See the supplementary video for more qualitative results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 55, + 205, + 281, + 245 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 281, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 281, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 281, + 245 + ], + "type": "table", + "html": "
ModelMask TypeMPJPE ↓MPJVE ↓
HMD-NeMoLearned & fixed5.5942.81
HMD-NeMoTAMTs (Ours)2.4831.30
", + "image_path": "b97625d06be0716a1587d89995c58f83e80393a71dd519fc70854839df4bcfe7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 299, + 286, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 286, + 346 + ], + "type": "text", + "content": "lyze the effect optimization in both HT and MC scenarios. Additionally, we provide results of HMD-NeMo in both HT and MC scenarios for various body parts, as well as qualitative effect of optimization in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 347, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 286, + 538 + ], + "type": "text", + "content": "Evaluation for the hand tracking scenario. To the best of our knowledge, HMD-NeMo is the first approach to address the problem of human motion generation given partial HMD signal, applicable to HT scenario. We believe that TAMT module is the key to the success of HMD-NeMo in handling missing/partial observation, so we compare it with an alternative commonly used in the Vision Transformers [10, 12], which uses a learned set of parameters (i.e., nn.Parameters in PyTorch) to model the missing observations5. While learned parameters are fixed after training for every data point and every sequence, TAMT temporarily updates itself at each time-step given the current state of the model. Table 2 (especially MPJVE) shows the superiority of TAMT module over learned and fixed parameters in handling missing hand observations. Qualitatively, Fig. 8 illustrates how HMD-NeMo performs in HT scenario." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 539, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 286, + 635 + ], + "type": "text", + "content": "Cross-dataset evaluation. To investigate the generalizability of HMD-NeMo, we conduct a 3-fold cross-dataset evaluation as in [15], wherein the models are trained on two subsets and test on the other subset. In order to compare our approach with existing methods, we conduct this experiment in MC scenario. As shown in Table 3, HMD-NeMo outperforms existing approaches in all three datasets, by a considerable margin, highlighting its generalizability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 635, + 286, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 286, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 286, + 684 + ], + "type": "text", + "content": "Evaluating the effect of input signal. As discussed in Section 3.1, HMD-NeMo utilizes head and hands, hands in the head space, as well as the corresponding velocities, as in Eq. 1. Table 4 summarizes the effect of each component in" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 307, + 205, + 547, + 281 + ], + "blocks": [ + { + "bbox": [ + 47, + 245, + 286, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 245, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 286, + 255 + ], + "type": "text", + "content": "Table 2. Effect of Temporally Adaptable Mask Tokens (TAMTs)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 255, + 286, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 255, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 286, + 277 + ], + "type": "text", + "content": "These results represent the HMD-NeMo prediction before optimization solely to evaluate the effect of the TAMT module." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 205, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 307, + 205, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 547, + 281 + ], + "type": "table", + "html": "
MethodTest on CMUTest on BMLrubTest on HDM05
MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓MPJPE ↓MPJVE ↓
FinalIK [25]18.8256.8317.5860.6418.4362.39
Ahuja et al. [1]18.77139.1713.30134.7717.90140.61
Yang et al. [30]12.9649.9411.0060.7411.9448.26
Dittadi et al. [9]13.0451.699.6951.8010.2140.07
Jiang et al. [15]8.3735.767.0443.708.0530.85
HMD-NeMo (Ours)7.1331.236.4640.386.8027.91
", + "image_path": "074ed4ca99abe99705f26eb68caa4ff0cf8702d0354df704d8c57533381cf107.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 318, + 327, + 535, + 395 + ], + "blocks": [ + { + "bbox": [ + 306, + 282, + 546, + 314 + ], + "lines": [ + { + "bbox": [ + 306, + 282, + 546, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 546, + 314 + ], + "type": "text", + "content": "Table 3. Results of cross-dataset evaluation between different methods. In order to compare with existing methods, we provide results in MC scenario." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 327, + 535, + 395 + ], + "lines": [ + { + "bbox": [ + 318, + 327, + 535, + 395 + ], + "spans": [ + { + "bbox": [ + 318, + 327, + 535, + 395 + ], + "type": "table", + "html": "
Input SignalMPJPE ↓MPJVE ↓
xt = {xh, xl, xr}t4.3839.63
xt = {xh, xl, xr, xlh, xrh}t3.2138.32
xt = {xh, xl, xr, ðh, ðl, ðr}t3.5335.27
Full input signals (Eq. 1)2.4831.30
", + "image_path": "f27dc06fb5f93d15058cb854c31d225db26a4b02d6056cb0be23354179718e0c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 395, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 306, + 395, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 545, + 426 + ], + "type": "text", + "content": "Table 4. Evaluating the effect of various input signals in HT scenario. Note that the hand visibility status " + }, + { + "bbox": [ + 306, + 395, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\nu_{l},\\nu_{r}" + }, + { + "bbox": [ + 306, + 395, + 545, + 426 + ], + "type": "text", + "content": " are always provided to the model." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 449, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 532 + ], + "type": "text", + "content": "the input signal. As shown, adding hands in the head space on top of head and hands in the world coordinates leads to better pose prediction, and thus reduces the MPJPE. Incorporating the velocities has a significant contribution to generating more temporally coherent motion, and thus reduces the MPJVE considerably. Considering all input signals, as in Eq. 1 leads to best MPJPE and MPJVE." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Evaluating the effect of STAE. One core component of HMD-NeMo is the spatioTemporal attention-based encoder (STAE). Here, we study the design choices of STAE in Table 5. As shown, removing the GRU component, which is responsible to learn the temporal information of various input signals, affects the MPJVE considerably. Removing the transformer encoder, which aims at learning the relation between head and hands, adversely affects the MPJPE. Unsurprisingly, removing the entire STAE module, i.e., going from input embedding layer to the decoders, is a considerably weak baseline, with poor pose quality and temporal coherency. HMD-NeMo utilizes the power and efficiency of GRU module to learn the temporal information as well as the expressively of the transformer encoder to learn how various inputs are related to each other. This design leads to" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 286, + 713 + ], + "type": "text", + "content": "5This can be considered as removing the unobserved hand (similar to [2]) and replacing it with learned and fixed parameters." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9628" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 269, + 134 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 269, + 134 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 269, + 134 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 269, + 134 + ], + "type": "table", + "html": "
STAE ComponentsMPJPE ↓MPJVE ↓
No GRU4.3446.61
No Transformer5.5539.10
No GRU, No Transformer7.0952.26
With STAE (Ours)2.4831.30
", + "image_path": "798614f27eb0c8cf42db03878837e22e97c9e855b28dfcd397ec808ea3440eaa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 155, + 269, + 269 + ], + "blocks": [ + { + "bbox": [ + 65, + 155, + 269, + 269 + ], + "lines": [ + { + "bbox": [ + 65, + 155, + 269, + 269 + ], + "spans": [ + { + "bbox": [ + 65, + 155, + 269, + 269 + ], + "type": "image", + "image_path": "f85b8bbdc99d2ec8b5439985ef482e66485cf2b34edcd4a7c75cd18033d23214.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 288, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 288, + 295 + ], + "type": "text", + "content": "Figure 9. Evaluating the effect of each loss term. Please note that each term is evaluated in isolation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 315, + 286, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 315, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 45, + 315, + 286, + 338 + ], + "type": "text", + "content": "best performing MPJPE and MPJVE, especially in HT scenario wherein the model faces regular missing observations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "content": "Evaluating the effect of each loss term. As described in Section 3.3 and shown in Eq. 6, HMD-NeMo is trained with five different loss terms. While " + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{data}" + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "content": " is the essential term for training HMD-NeMo, other loss terms contribute significantly to the performance of the model. In Fig. 9, we illustrate the contribution of each loss term, in leave-one-term-out manner, on the changes in the MPJPE and MPJVE metrics. " + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{smooth}" + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "content": " has a significant impact on improving the MPJVE, while " + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}" + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "content": " improves the MPJPE, and " + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{forecast}" + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "content": ", which acts only on hands mildly improves MPJPE and MPJVE (its contribution to the total error metric is relatively small as it acts on a significantly fewer joints). " + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SE(3)}" + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "content": " makes the largest contribution to the reduction of error in both metrics. It is likely that " + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SE(3)}" + }, + { + "bbox": [ + 45, + 339, + 286, + 554 + ], + "type": "text", + "content": " aims to bridge that gap between the representation of the input signal (head and hand global transformation matrices) and the representation of the output pose (global root orientation and relative joint rotations)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 554, + 286, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 286, + 674 + ], + "type": "text", + "content": "Evaluating the effect of optimization. While the prediction of HMD-NeMo is very good, if the budget allows, it can be further optimized to improve accuracy (see supp mat for qualitative results). To bridge the gap between the predictions and observations, we optimize the pose prediction from HMD-NeMo, so that it matches the head and hand observations. As described in Section 3.4, in MC scenario, where both hands are always available, a simple optimization (with non-robust data energy term) loop can be used. The effect of such optimization in MC scenario is provided" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 334, + 70, + 518, + 109 + ], + "blocks": [ + { + "bbox": [ + 47, + 134, + 276, + 144 + ], + "lines": [ + { + "bbox": [ + 47, + 134, + 276, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 276, + 144 + ], + "type": "text", + "content": "Table 5. Evaluating the effect of STAE module in HT scenario." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 70, + 518, + 109 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 518, + 109 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 518, + 109 + ], + "type": "table", + "html": "
ConfigurationMPJPE ↓MPJVE ↓
Without optimization2.0726.07
With optimization1.9024.99
", + "image_path": "96544b28beddee228765a8da567a9e07e1a49ffaf47001539637163eb42074b9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 141, + 537, + 191 + ], + "blocks": [ + { + "bbox": [ + 306, + 110, + 545, + 131 + ], + "lines": [ + { + "bbox": [ + 306, + 110, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 545, + 131 + ], + "type": "text", + "content": "Table 6. Effect of optimization in MC scenario, where both hands are always visible." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 141, + 537, + 191 + ], + "lines": [ + { + "bbox": [ + 315, + 141, + 537, + 191 + ], + "spans": [ + { + "bbox": [ + 315, + 141, + 537, + 191 + ], + "type": "table", + "html": "
ConfigurationPriorityMPJPE ↓MPJVE ↓
No OptimizationPlausibility2.4831.30
Opt. with EnrFidelity2.3233.51
Opt. with ErBoth2.3731.33
", + "image_path": "7b484a2857b7756c436593f6616f82c2eeaea4ed4babe204353be949428a65fc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 191, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 306, + 191, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 545, + 223 + ], + "type": "text", + "content": "Table 7. Configuring HT scenario. Our model offers a range of choices to provide the best possible end-user experience despite incomplete observations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 243, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 374 + ], + "type": "text", + "content": "in Table 6. However, considering the same strategy for the HT scenario, where we may partially observe hands, this may not be optimal as we lose plausibility in the generated motions (captured by the MPJVE metric). As described in Section 3.4, depending on the scenario and experience requirements, one may choose to (1) avoid optimizing the predictions if plausibility is the highest priority (first row of Table 7), (2) use non-robust energy term if fidelity is the highest priority (second row of Table 7), or (3) use a robust energy term as a trade-off between fidelity and plausibility (third row of Table 7)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 375, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 469 + ], + "type": "text", + "content": "Performance analysis. With 5.3M parameters, at inference time, HMD-NeMo requires only " + }, + { + "bbox": [ + 304, + 375, + 545, + 469 + ], + "type": "inline_equation", + "content": "4.4\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 375, + 545, + 469 + ], + "type": "text", + "content": " to generate a pose given a HMD signal on a typical laptop CPU. On a NVIDIA Tesla P100 GPU, our model runs at 265 fps, with 1 iteration of optimization costs 3ms per frame (which could be further improved with more optimized implementation). Such performance makes HMD-NeMo a potential solution for HMD-driven avatar animation in immersive environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 481, + 378, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 378, + 494 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 378, + 494 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 502, + 545, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 705 + ], + "type": "text", + "content": "In this paper, we present HMD-NeMo, a unified approach to generate full-body avatar motion in both motion controller and hand tracking scenarios. To handle the unobserved hands in a temporally coherent and plausible manner, we introduce TAMT module. It is worth noting that, in this paper, we considered one major reason for partial hand visibility, i.e., hands appearing out of the FoV, however, in practice, hands may not be visible to the hand tracking camera due to failure in tracking and occlusion by another object or by another body part. While such cases are not considered in the data augmentation in our paper, they certainly can be taken into account and TAMT can be used to fill such gaps with no additional modification; this is left for future exploration. We provide extensive analyses on different components of HMD-NeMo, and shed light on various choices for optimization on top of neural network's predictions when it comes to production priorities (plausibility versus fidelity)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 286, + 712 + ], + "type": "text", + "content": "6This experiment only evaluates the contribution of each loss term independently. Evaluating the combination of loss terms remains for future investigations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9629" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "text", + "content": "[1] Karan Ahuja, Eyal Ofek, Mar Gonzalez-Franco, Christian Holz, and Andrew D Wilson. Coolmoves: User motion accentuation in virtual reality. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 5(2):1-23, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "type": "text", + "content": "[2] Sadegh Aliakbarian, Pashmina Cameron, Federica Bogo, Andrew Fitzgibbon, and Thomas J Cashman. FLAG: Flow-based 3D avatar generation from sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13253-13262, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 287, + 251 + ], + "type": "text", + "content": "[3] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 255, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 287, + 297 + ], + "type": "text", + "content": "[4] Benjamin Biggs, Sébastien Ehrhadt, Hanbyul Joo, Benjamin Graham, Andrea Vedaldi, and David Novotny. 3D multibodies: Fitting sets of plausible 3D human models to ambiguous image data. 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 300, + 287, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 287, + 355 + ], + "type": "text", + "content": "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 389 + ], + "type": "text", + "content": "[6] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 393, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 287, + 436 + ], + "type": "text", + "content": "[7] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 440, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 440, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 440, + 287, + 472 + ], + "type": "text", + "content": "[8] Vasileios Choutas, Federica Bogo, Jingjing Shen, and Julien Valentin. Learning to fit morphable models. arXiv preprint arXiv:2111.14824, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 475, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 540 + ], + "type": "text", + "content": "[9] Andrea Dittadi, Sebastian Dziadzio, Darren Cosker, Ben Lundell, Thomas J Cashman, and Jamie Shotton. Full-body motion from a single head-mounted device: Generating SMPL poses from partial observations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11687-11697, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "type": "text", + "content": "[11] Nima Ghorbani and Michael J Black. Soma: Solving optical marker-based mocap automatically. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11117-11126, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 39(4):53-1, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[14] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[15] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. AvatarPoser: Articulated full-body pose tracking from sparse motion sensing. In Proceedings of European Conference on Computer Vision. Springer, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. International Conference on Learning Representations, ICLR, 2014." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "type": "text", + "content": "[17] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11605-11614, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[18] Matthew Loper, Naureen Mahmood, and Michael J Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics (TOG), 33(6):1-13, 2014." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "text", + "content": "[19] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "text", + "content": "[20] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pages 5442-5451, Oct. 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 434, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 498 + ], + "type": "text", + "content": "[21] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10975-10985, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[22] Jose Luis Ponton, Haoran Yun, Carlos Andujar, and Nuria Pelechano. Combining Motion Matching and Orientation Prediction to Animate Avatars for Consumer-Grade VR Devices. Computer Graphics Forum, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "text", + "content": "[23] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-level pixel-aligned implicit function for high-resolution 3D human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "text", + "content": "[24] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 543, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 543, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 543, + 689 + ], + "type": "text", + "content": "[25] Unity Asset Store. Final IK. https://assetstoreunity.com/packages/tools/animation/final-ik-14290, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[26] Dorin Ungureanu, Federica Bogo, Silvano Galliani, Pooja Sama, Casey Meekhof, Jan Stuhmer, Thomas J Cashman," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9630" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 608 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "text", + "content": "Bugra Tekin, Johannes L Schonberger, Pawel Olszta, et al. Hololens 2 research mode as a tool for computer vision research. arXiv preprint arXiv:2008.11239, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "text", + "content": "[28] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3D human pose estimation from sparse IMUs. In Computer graphics forum, volume 36, pages 349-360. Wiley Online Library, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 286, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 286, + 240 + ], + "type": "text", + "content": "[29] Alexander Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. arXiv preprint arXiv:2209.09391, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[30] Dongseok Yang, Doyeon Kim, and Sung-Hee Lee. LoBSTr: Real-time lower-body pose prediction from sparse upper-body tracking signals. In Computer Graphics Forum, volume 40, pages 265–275. Wiley Online Library, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "text", + "content": "[31] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical Inertial Poser (PIP): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13167-13178, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "text", + "content": "[32] Xinyu Yi, Yuxiao Zhou, and Feng Xu. TransPose: Real-time 3D human translation and pose estimation with six inertial sensors. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "type": "text", + "content": "[33] Andrei Zanfir, Eduard Gabriel Bazavan, Hongyi Xu, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Weakly supervised 3D human pose and shape reconstruction with normalizing flows. In European Conference on Computer Vision, pages 465-481. Springer, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 508 + ], + "type": "text", + "content": "[34] Mihai Zanfir, Andrei Zanfir, Eduard Gabriel Bazavan, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. THUNDR: Transformer-based 3D human reconstruction with markers. arXiv preprint arXiv:2106.09336, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[35] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3D bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 608 + ], + "type": "text", + "content": "[36] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9631" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_content_list.json b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..64742d333862c1131e8e513c29d6e3c61cd119ab --- /dev/null +++ b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_content_list.json @@ -0,0 +1,1884 @@ +[ + { + "type": "text", + "text": "HOSNeRF: Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video", + "text_level": 1, + "bbox": [ + 109, + 128, + 861, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jia-Wei Liu $^{1*}$ , Yan-Pei Cao $^{2\\dagger}$ , Tianyuan Yang $^{1}$ , Zhongcong Xu $^{1}$ , Jussi Keppo $^{4,5}$ , Ying Shan $^{2}$ , Xiaohu Qie $^{3}$ , Mike Zheng Shou $^{1\\dagger}$", + "bbox": [ + 179, + 202, + 805, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Show Lab, National University of Singapore 2 ARC Lab, 3 Tencent PCG 4 Business School, 5 Institute of Operations Research and Analytics, National University of Singapore", + "bbox": [ + 89, + 255, + 898, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce HOSNeRF, a novel $360^{\\circ}$ free-viewpoint rendering method that reconstructs neural radiance fields for dynamic human-object-scene from a single monocular in-the-wild video. Our method enables pausing the video at any frame and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints. The first challenge in this task is the complex object motions in human-object interactions, which we tackle by introducing the new object bones into the conventional human skeleton hierarchy to effectively estimate large object deformations in our dynamic human-object model. The second challenge is that humans interact with different objects at different times, for which we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively. Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by a large margin of $40\\% \\sim 50\\%$ in terms of LPIPS. The code, data, and compelling examples of $360^{\\circ}$ free-viewpoint renderings from single videos: https://showlab.github.io/HOSNeRF.", + "bbox": [ + 75, + 357, + 473, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 704, + 209, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video reconstruction and free-viewpoint rendering offer innovative opportunities for creating immersive experiences, encompassing virtual reality, telepresence, metaverse, and 3D animation production. While reconstructing videos has the potential to enhance user engagement and provide more realistic environments, it also poses significant challenges in terms of monocular viewpoints [55, 48] and complicated human-environment interactions [50, 14].", + "bbox": [ + 75, + 729, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, remarkable progress has been made in", + "bbox": [ + 94, + 849, + 467, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "novel view synthesis, particularly since the introduction of Neural Radiance Fields (NeRF) [27]. While initially limited to reconstructing static 3D scenes based on multi-view images, subsequent studies have proposed various approaches to address the challenge of dynamic view synthesis [34, 29, 30, 47, 20, 8]. NeRF-based techniques have evolved to either incorporate deformation fields that map dynamic fields to canonical NeRF spaces [34, 29, 30, 47], or model dynamic scenes as 4D spatio-temporal radiance fields [20, 8]. While these approaches have shown promising results in dynamic view synthesis, they are limited to simple deformations. Another line of research is specifically designed for dynamic neural human modeling that relies on estimated human poses as a priori information [32, 50]. Recently, Neuman [14] combines pose-driven dynamic human models with static scene models for representing dynamic human-centric scenes.", + "bbox": [ + 500, + 327, + 893, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, none of the aforementioned techniques can accurately reconstruct challenging monocular videos with fast and complex human-object-scene motions and interactions, as shown in Fig. 1(d). This is due to two particular challenges listed below. To tackle them, we introduce a novel method called Human-Object-Scene Neural Radiance Fields (HOSNeRF).", + "bbox": [ + 496, + 590, + 893, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "i) Complex object motions in human-object interactions. In contrast to the simple motions that can be modeled by general deformation modules [29, 30], the object motion during human-object interaction is often drastic and composed of various atomic motions (e.g., play tennis). To tackle this challenge, we propose new object bones that are attached to the human skeleton hierarchy to estimate human-object deformations in a coarse-to-fine manner for our dynamic human-object model. The object bones and underlying object linear blend skinning (object LBS) allow for the accurate estimation of objects' deformations through the relative transformations in the kinematic tree of the skeleton hierarchy.", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work is partially done during internship at ARC Lab, Tencent PCG.", + "bbox": [ + 93, + 875, + 460, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Authors.", + "bbox": [ + 96, + 887, + 230, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18483", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6aa2f4bbcd3d7a80f3b0f41d901ff730a908862146905e5609d10befd3008413.jpg", + "image_caption": [ + "Figure 1: Our HOSNeRF (b) takes a single monocular in-the-wild video (a) as input, and creates high-fidelity $360^{\\circ}$ free-viewpoint rendering of all scene details (dynamic human body, objects, and background) at any time (d). Our method enables rendering views with novel object poses and novel human poses as shown in (c), and produces high-fidelity dynamic novel view synthesis results at novel timesteps, with significant improvements over SOTA approaches as shown in (d)." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 271, + 303 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/acdab11ba8a5c4aaf70fb8c93b8d7ba0981573fb2cab1e96c98e8df002e4f9dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 88, + 467, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/51c94a859fa255eaebf95973920c96b6d1fb8c11ec635692a5103e3844e88553.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 88, + 722, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/692d6f2eceb289e3397e21e4a40f1ca1da20d3c5c6eaf476d193942f7e69f6dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 89, + 888, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ii) Humans interact with different objects at different times. The above human-object model is designed for humans interacting with the same object over time. But when the person puts down the current object or picks a new object, it is not clear how to dynamically remove/add such objects in the static background model and the human-object model whose canonical space is static. To handle this challenge, we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively.", + "bbox": [ + 75, + 376, + 468, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we systematically explore and identify effective training objectives and strategies for our proposed HOSNeRF, including deformation cycle consistency, optical flow supervisions, and foreground-background rendering. On two challenging datasets collected by ourselves and NeuMan [14], our HOSNeRF achieves high-fidelity dynamic novel view synthesis results and enables pausing the monocular video at any time and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints, as shown in Fig. 1(d).", + "bbox": [ + 75, + 529, + 468, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are:", + "bbox": [ + 96, + 683, + 364, + 698 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel framework of HOSNeRF, the first work to achieve $360^{\\circ}$ free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video.", + "- We propose the object bones and state-conditional representations to handle the non-rigid motions and interactions of humans, objects, and the environment.", + "- Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by $40\\% \\sim 50\\%$ in terms of LPIPS." + ], + "bbox": [ + 94, + 712, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 375, + 640, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dynamic Human Modeling", + "text_level": 1, + "bbox": [ + 500, + 402, + 745, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic human modeling has shown promising results in utilizing various representations such as point clouds [1, 51], meshes [10, 24], voxels [25, 39], and neural implicit functions [12, 35], with models like SMPL [26, 33] being commonly used for parameterizing the human body. Since the introduction of NeRF [27], neural human representation [32, 36, 42, 54, 23] has achieved remarkable progress on representing dynamic human bodies from sparse-view videos. Among them, Neural Actor [23] and Neural Body [32] pioneer in combining NeRF [27] with SMPL deformable meshes to represent human bodies with complex motions. Subsequent works have further improved on the generalizability [17, 4, 9] and animatability [31, 18] of human bodies. To support multi-person modeling, recent works [55, 37] have proposed to segment each human into 3D bounding boxes and learn a separate layered dynamic NeRF for each person. Other works [46, 43, 15, 44] are specifically designed to reconstruct the dynamic human and object with RGB-D or multi-view videos as inputs. They track the human and object pose, and separately reconstruct them with volumetric fusion [43], neural texture blending [44], or neural rendering [15, 56].", + "bbox": [ + 496, + 429, + 890, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite achieving promising results, these approaches require multi-view videos or RGB-D as input, limiting their real-world applications. To solve this problem, HumanNeRF [50] is proposed to represent moving humans from a monocular video by the human pose-driven deformation module and canonical space. NeuMan [14] is the first successful attempt at reconstructing both the dynamic human and static background from a single video. However, Neuman [14] does not support human-environment interactions", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18484", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and performs poorly at large camera motions.", + "bbox": [ + 76, + 90, + 379, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Dynamic View Synthesis for General Scenes", + "text_level": 1, + "bbox": [ + 76, + 113, + 449, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most prior approaches on dynamic scene modeling require synchronized multi-view videos [58, 19, 41, 5, 55, 48] or depth [59, 13, 28] as input. Recent studies have built upon NeRF [27] to reconstruct dynamic neural radiance fields from monocular videos by either learning a deformation field that maps dynamic observation to canonical field [34, 29, 30, 47] or building 4D spatio-temporal radiance fields [53, 20, 8]. Among them, Nerfies [29] associates latent codes with the deformation field and HyperNeRF [30] represents motion in a high-dimensional space. $\\mathrm{D}^2\\mathrm{NeRF}$ [52] builds upon HyperNeRF [30] and further decouples the dynamic components from the static background, and represents them separately with a HyperNeRF [30] and NeRF [27]. DynIBaR [21] proposes a motion-adjusted multi-view feature aggregation module to synthesize new viewpoints by aggregating features from nearby views. Other studies have introduced voxel grids [6, 22, 40] or planar representations [7, 3] for fast dynamic radiance fields reconstruction. While these approaches have achieved high-fidelity dynamic view synthesis results, they are restricted to simple scene deformations. In contrast, our HOSNeRF is capable of representing significant human-object motions and interactions in complex environments.", + "bbox": [ + 75, + 138, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 497, + 168, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 521, + 217, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HumanNeRF [50] has been recently introduced to represent a moving person with a NeRF [27]-based canonical space $\\Psi_{\\mathrm{c}}$ that maps 3D points to color $\\mathbf{c}$ and density $d$ , and a human pose-guided deformation field $\\Psi_{\\mathrm{d}}$ that transforms deformed points $\\mathbf{x}_{\\mathrm{d}}^{i}$ from the deformed space to canonical points $\\mathbf{x}_{\\mathrm{c}}^{i}$ in the canonical space ( $i$ omitted for simplicity).", + "bbox": [ + 75, + 545, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {\\mathrm {c}} (\\gamma (\\mathbf {x} _ {\\mathrm {c}})) \\longmapsto (\\mathbf {c}, \\sigma), \\Psi_ {\\mathrm {d}} (\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}) \\longmapsto (\\mathbf {x} _ {\\mathrm {c}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 643, + 468, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma (\\mathbf{x})$ is the standard positional encoding function, and $\\mathcal{I} = \\{\\mathbf{J}_i\\}$ and $\\mathcal{R} = \\{\\omega_i\\}$ represent the 3D human joints and the local joint axis-angle rotations.", + "bbox": [ + 75, + 669, + 468, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The deformation field $\\Psi_{\\mathrm{d}}$ is decomposed into the coarse human skeleton-driven deformation $\\Psi_{\\mathrm{d}}^{\\mathrm{coarse}}$ , and the fine non-rigid deformation conditioned on human poses $\\Psi_{\\mathrm{d}}^{\\mathrm{fine}}$ :", + "bbox": [ + 75, + 715, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {c}} ^ {\\prime} = \\Psi_ {\\mathrm {d}} ^ {\\mathrm {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right), \\mathbf {x} _ {\\mathrm {c}} = \\mathbf {x} _ {\\mathrm {c}} ^ {\\prime} + \\Psi_ {\\mathrm {d}} ^ {\\mathrm {f i n e}} \\left(\\mathbf {x} _ {\\mathrm {c}} ^ {\\prime}, \\mathcal {R}\\right). (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 782, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mip-NeRF 360 [2] is designed to synthesize realistic views for highly intricate, unbounded real-world static scenes. To render pixel colors, the casted rays are split into a set of intervals $T_{i} = [t_{i}, t_{i + 1})$ . The mean and covariance of the conical frustum corresponding to every interval are computed as $(\\boldsymbol{\\mu}, \\boldsymbol{\\Sigma}) = \\mathbf{r}(T_{i})$ . To parameterize the Gaussian", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "parameters for unbounded scenes, Mip-NeRF 360 [2] further proposes a contraction function $f(\\mathbf{x})$ that distributes distant points proportionally to disparity,", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = \\left\\{ \\begin{array}{c l} \\mathbf {x} & \\| \\mathbf {x} \\| \\leq 1 \\\\ \\left(2 - \\frac {1}{\\| \\mathbf {x} \\|}\\right) \\left(\\frac {\\mathbf {x}}{\\| \\mathbf {x} \\|}\\right) & \\| \\mathbf {x} \\| > 1 \\end{array} . \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 147, + 890, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, $f(\\mathbf{x})$ is applied to $(\\pmb {\\mu},\\pmb {\\Sigma})$ as follows:", + "bbox": [ + 500, + 196, + 790, + 213 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}\\right) = \\left(f (\\boldsymbol {\\mu}), \\mathbf {J} _ {f} (\\boldsymbol {\\mu}) \\boldsymbol {\\Sigma} \\mathbf {J} _ {f} (\\boldsymbol {\\mu}) ^ {\\mathrm {T}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 223, + 890, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{J}_f(\\boldsymbol{\\mu})$ is the Jacobian of $f$ at $\\boldsymbol{\\mu}$ . The contracted Gaussian parameters $(\\hat{\\boldsymbol{\\mu}},\\hat{\\boldsymbol{\\Sigma}})$ are further encoded through the integrated positional encoding (IPE) [2]:", + "bbox": [ + 496, + 258, + 890, + 306 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}) = \\left\\{\\left[ \\begin{array}{l} \\sin \\left(2 ^ {\\ell} \\hat {\\boldsymbol {\\mu}}\\right) \\exp \\left(- 2 ^ {2 \\ell - 1} \\operatorname {d i a g} (\\hat {\\boldsymbol {\\Sigma}})\\right) \\\\ \\cos \\left(2 ^ {\\ell} \\hat {\\boldsymbol {\\mu}}\\right) \\exp \\left(- 2 ^ {2 \\ell - 1} \\operatorname {d i a g} (\\hat {\\boldsymbol {\\Sigma}})\\right) \\end{array} \\right] \\right\\} _ {\\ell = 0} ^ {L - 1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 316, + 890, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the color and density of intervals can be obtained as", + "bbox": [ + 500, + 369, + 870, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {\\mathrm {s}} \\left(\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}})\\right) \\longmapsto (\\mathbf {c}, \\sigma), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 395, + 890, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Psi_{\\mathrm{s}}$ is the scene NeRF MLP [2].", + "bbox": [ + 500, + 430, + 753, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dynamic Human-Object Model", + "text_level": 1, + "bbox": [ + 500, + 454, + 777, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. With estimated 3D human poses a priori, HumanNeRF [50] is effective at modeling moving people, but can only encode body parts, lacking the capability to model additional structures (e.g., objects held by the person) and hence not suitable for challenging scenarios containing complex human-object interactions.", + "bbox": [ + 496, + 477, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A naive approach would be attaching the objects directly to the interacting body parts (e.g., hands) so that the object transformations can be queried via the skeletal motion of the corresponding human joints. However, this approach cannot represent the relative transformations between the object and its interacting body part, making it difficult to model large objects, such as the suitcase shown in Fig. 6.", + "bbox": [ + 496, + 568, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Bones. To address this limitation, based on the conventional human skeleton hierarchy, we introduce a new \"object bone\" for each hand, where the starting point of the object bone is connected to the corresponding hand joint. As shown in Fig. 2 (a), two new object bones (green color) are connected to the hand joints, which represent the relative rotation and translation with respect to their parent hand joints. In potential, the number, bone size, and attaching joints of object bones can also be flexibly customized based on different human-object interaction scenarios. Therefore, our 3D human-object pose consists of 3D joints $\\mathcal{I} = \\{\\mathbf{J}_{\\mathrm{human}},\\mathbf{J}_{\\mathrm{object}}\\}$ and their axis-angle rotations $\\mathcal{R} = \\{\\omega_{\\mathrm{human}},\\omega_{\\mathrm{object}}\\}$ .", + "bbox": [ + 496, + 674, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object LBS. To drive the articulated motions of the objects and human body, we learn a linear blend skinning", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18485", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08907949088bb6f39e1b288e7935b83a15880793ef1af925cc5e204f022e8e99.jpg", + "image_caption": [ + "Figure 2: Overview of our method. (1) Orange flowchart: The deformation from deformed points to canonical points are effectively estimated by the human-object backward LBS (a) and non-rigid deformation module, and their properties (i.e., density, color) can be obtained by querying the state-conditional canonical space. (2) Green flowchart: The background intervals sampled from a deformed frame are concatenated with object state embeddings for querying properties through the state-conditional scene model. (3) Novel views can be accordingly synthesized by volume rendering for re-ordered properties." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(LBS) weights field during training, as shown in Fig. 2 (a), which encodes the influence regions of object bones and human bones. Following HumanNeRF [50], we pack the human and object bone blend weights and additional background weight into a single volume with $K + 1$ channels, i.e., $\\mathcal{W}_{\\mathrm{c} \\to \\mathrm{d}}(\\mathbf{x}_{\\mathrm{c}}) = \\left\\{w_{\\mathrm{c} \\to \\mathrm{d}}^{i}(\\mathbf{x}_{\\mathrm{c}})\\right\\}$ , and generate the volume from a learnable latent code using a CNN.", + "bbox": [ + 75, + 436, + 470, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For instance, in Fig. 2, the subject holds a backpack in the right hand, so the learned LBS weight field only encodes the influence region of the right object bone with a similar shape to the backpack, while the left object bone has no influence on skeleton motion. Therefore, this learnable LBS weights field enables our object bones to be generalizable to various human-object interaction scenarios with left, right, or both hands. In addition, our human-object skeleton hierarchy enables HOSNeRF to support rendering with novel object pose and novel human pose, as illustrated in Fig. 1(c).", + "bbox": [ + 75, + 555, + 468, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Forward-Backward Deformation Framework. To improve the smoothness and consistency of human-object deformations, we propose to leverage the cycle consistency between forward and backward human-object deformations. The backward deformation first transforms sampled deformed points $\\mathbf{x}_{\\mathrm{d}}$ to their corresponding canonical points $\\mathbf{x}_{\\mathrm{c}}$ , which are further mapped back to the deformed space $\\hat{\\mathbf{x}}_{\\mathrm{d}}$ through the forward deformation module.", + "bbox": [ + 75, + 720, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2(a), with the proposed human-object skeleton definition and LBS weight fields, the coarse backward skeleton deformation $\\Psi_{\\mathrm{d}\\to \\mathrm{c}}^{\\mathrm{coarse}}$ is defined similar to", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HumanNeRF [50],", + "bbox": [ + 500, + 436, + 627, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right) = \\sum_ {i = 1} ^ {K} w _ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {i} \\left(\\mathbf {x} _ {\\mathrm {d}}\\right)\\left(\\mathbf {R} _ {\\mathrm {d} \\rightarrow c} ^ {i} \\mathbf {x} _ {\\mathrm {d}} + \\mathbf {t} _ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 459, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is the number of human-object bones, $w_{\\mathrm{d}\\rightarrow \\mathrm{c}}^i$ is the blend weight for the $i$ -th bone at the deformed space, and $\\mathbf{R}_{\\mathrm{d}\\to \\mathrm{c}}^i$ and $\\mathbf{t}_{\\mathrm{d}\\to \\mathrm{c}}^i$ are the backward rotation and translation that are explicitly derived from the human-object pose.", + "bbox": [ + 498, + 513, + 890, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To model the non-rigid motion that can not be represented by skeleton motion, we additionally define the backward fine non-rigid motion $\\Psi_{\\mathrm{d}\\to c}^{\\mathrm{fine}}$ conditioned on the human-object poses [50]:", + "bbox": [ + 498, + 575, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {x} _ {\\mathrm {d} \\rightarrow \\mathrm {c}} = \\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\mathrm {f i n e}} \\left(\\gamma \\left(\\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\mathrm {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {I}, \\mathcal {R}\\right)\\right), \\mathcal {R}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 643, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, $\\mathbf{x}_{\\mathrm{d}}$ can be mapped to the canonical space by", + "bbox": [ + 500, + 670, + 864, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {c}} = \\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right) + \\Delta \\mathbf {x} _ {\\mathrm {d} \\rightarrow \\mathrm {c}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 695, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, the forward coarse skeleton motion driven by the human-object pose can be defined as", + "bbox": [ + 498, + 720, + 890, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right) = \\sum_ {i = 1} ^ {K} w _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i} \\left(\\mathbf {x} _ {\\mathrm {c}}\\right)\\left(\\mathbf {R} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i} \\mathbf {x} _ {\\mathrm {c}} + \\mathbf {t} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 760, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{\\mathrm{c}\\rightarrow \\mathrm{d}}^i$ is the canonical blend weight for the $i$ -th bone, $\\mathbf{R}_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{i}$ and $\\mathbf{t}_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{i}$ are the forward rotation and translation obtained from the human-object pose. In addition, the forward fine non-rigid motion $\\Psi_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{\\mathrm{fine}}$ is", + "bbox": [ + 498, + 814, + 890, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {x} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} = \\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {f i n e}} \\left(\\gamma \\left(\\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right)\\right), \\mathcal {R}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18486", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6f2e8c93133308a3943824ed6f272f1e8d72ec5a4abf41c86d413815a2ede280.jpg", + "image_caption": [ + "$\\gamma (\\hat{\\mu},\\hat{\\Sigma})$ IPE feature $\\gamma (\\mathbf{x}_c)$ Embed points $\\mathbf{O}_s^i\\mathbf{O}_c^i i$ th scene and canonical state embedding", + "T=6 The human picks up the bag" + ], + "image_footnote": [], + "bbox": [ + 81, + 103, + 467, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2b731b6bab4f1f535632c61de0b01f77b89e3240674076968e95abc926348c75.jpg", + "image_caption": [ + "T=0 The human starts walking", + "Figure 3: State-conditional representation for the scene model (a) and human-object canonical space (b)." + ], + "image_footnote": [], + "bbox": [ + 81, + 203, + 467, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the canonical points $\\mathbf{x}_{\\mathrm{c}}$ can be transformed back to the deformed scene by", + "bbox": [ + 76, + 354, + 468, + 386 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {\\mathrm {d}} = \\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right) + \\Delta \\mathbf {x} _ {\\mathrm {c} \\rightarrow \\mathrm {d}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 396, + 468, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. State-Conditional Representation", + "text_level": 1, + "bbox": [ + 76, + 424, + 370, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation. In complex dynamic scenes, humans can interact with different objects at different timesteps. As the example in Fig. 3 shows, the human picks up the bag at $T = 6$ and puts it down at $T = 24$ , resulting in 3 object states: the bag in the background during $T \\in [0, 6]$ , the bag held by the human during $T \\in [6, 24]$ , the bag at a new position in the background when $T > 24$ . Such object state changes prohibit us from directly using the static background model (i.e., Mip-NeRF 360 [2]) and the dynamic human-object model proposed in Sec. 3.2.", + "bbox": [ + 75, + 446, + 468, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A naive solution is to train a separate network for each object state. However, this solution is limited because each network is trained only on a short segment (e.g., $T \\in [0, 6]$ ) instead of the whole video. As a result, some regions often are difficult to reconstruct because they are not observed in the short segment. But these regions could have already been captured by video segments of other object states.", + "bbox": [ + 75, + 599, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To make full use of all segments to jointly train one model shared for different states, we introduce learnable state embeddings to represent different object states. As depicted in Fig. 3, we have two state embeddings respectively for our scene model and human-object model, which are explained in detail as follows.", + "bbox": [ + 75, + 704, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State-Conditional Scene Model. As shown in Fig. 3(a), we develop a state-conditional Mip-NeRF 360 [2] to represent the static scene with temporal object transitions. In a dynamic scene with $N$ object states, we define $N$ learnable state embeddings $\\mathcal{O}_{\\mathrm{s}} = \\{\\mathbf{O}_{\\mathrm{s}}^{i}$ ( $i = 1,2,\\dots ,N$ ) to represent states information. Therefore, at state $i$ , we concatenate the IPE features $\\hat{\\gamma} (\\hat{\\mu},\\hat{\\Sigma})$ (Eq. (5)) of ray intervals with", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the state embedding $\\mathbf{O}_{\\mathrm{s}}^{i}$ as input to the scene MLP $\\Psi_{\\mathrm{s}}$ for querying the scene color and density.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {\\mathrm {s}} \\left(\\operatorname {c o n c a t} \\left(\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}), \\mathbf {O} _ {\\mathrm {s}} ^ {i}\\right)\\right) \\longmapsto (\\mathbf {c}, \\sigma). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 127, + 890, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This representation enables a shared scene model across the video with multiple object state transitions.", + "bbox": [ + 498, + 157, + 890, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State-Conditional Dynamic Human-Object Model. In Sec. 3.2, we introduce the human-object pose-driven deformation module and the human-object canonical space for modeling dynamic human-objects. However, the reconstructed canonical space is insufficient in representing the temporal changes in object geometry and appearance when the subject interacts with new objects. To address this limitation, we condition the canonical space on the object states, as shown in Fig. 3(b). In particular, we employ an object state-conditional canonical space and optimize a shared blend weights field across all states.", + "bbox": [ + 496, + 188, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a dynamic scene with $N$ object states, we define $N$ learnable state embeddings $\\mathcal{O}_{\\mathrm{c}} = \\{\\mathbf{O}_{\\mathrm{c}}^{i}\\} (i = 1,2,\\dots ,N)$ to represent object states in the canonical space. Therefore, at object state $i$ , we concatenate the positionally encoded canonical points in Eq. (1) with the state embedding $\\mathbf{O}_{\\mathrm{c}}^{i}$ as input to the canonical MLP for querying the human-object color and density.", + "bbox": [ + 496, + 354, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {\\mathrm {c}} \\left(\\operatorname {c o n c a t} \\left(\\gamma \\left(\\mathbf {x} _ {\\mathrm {c}}\\right), \\mathbf {O} _ {\\mathrm {c}} ^ {i}\\right)\\right) \\longmapsto (\\mathbf {c}, \\sigma). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 467, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 500, + 491, + 604, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering. To render the pixel color, we shoot two rays $(\\mathbf{r}_{\\mathrm{s}},\\mathbf{r}_{\\mathrm{ho}})$ and sample intervals (i.e., ray points) to query the scene model (using $\\mathbf{r}_{\\mathrm{s}}$ ) and dynamic human-object model (with $\\mathbf{r}_{\\mathrm{ho}}$ ) respectively in their coordinate systems. After aligning the coordinates (see supplementary materials), we transform the sampled 3D points and their queried properties from the dynamic human-object space to the scene space, and re-order all sampled properties based on their distance from the camera center, as shown in Fig. 2. Therefore, the color of a pixel can be calculated through volume rendering, i.e., by integrating the re-ordered properties along the ray [27]:", + "bbox": [ + 496, + 513, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\left(1 - e ^ {- \\sigma_ {i} \\delta_ {i}}\\right) \\mathbf {c} _ {i}, T _ {i} = e ^ {- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 702, + 890, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Directly fusing $\\mathbf{r}_{\\mathrm{s}}$ and $\\mathbf{r}_{\\mathrm{ho}}$ may yield minor artifacts on overlapped ground regions due to intersections of human-object points and scene points on short-length ground rays. To solve this issue, we estimate the foreground mask during training to separate the background rays from foreground ones. This allows us to focus on combining $\\mathbf{r}_{\\mathrm{s}}$ and $\\mathbf{r}_{\\mathrm{ho}}$ at the overlapped foreground regions, thereby reducing artifacts.", + "bbox": [ + 496, + 750, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objectives. The training of HOSNeRF consists of three stages. In the first stage, we mask out the dynamic human-object regions and train the state-conditional", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18487", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1e3a0b3a9617cf4f9e7fe399b0d7415d126e3b4336dff3111019508f0f6e1100.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SceneNo. of objectsNo. of states
BACKPACK13
TENNIS23
SUITCASE14
PLAYGROUND35
DANCE37
LOUNGE35
", + "bbox": [ + 109, + 88, + 436, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Details of our collected dataset.", + "bbox": [ + 135, + 200, + 403, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scene model. Then, we train the state-conditional dynamic human-object model on the dynamic human-object regions. In the third stage, we combine these two models and further finetune the complete HOSNeRF for all image pixels.", + "bbox": [ + 75, + 226, + 467, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the single video with calibrated poses, the first stage of HOSNeRF is optimized by minimizing the photometric MSE loss and the regularization losses proposed by Mip-NeRF 360 [2] to avoid background collapse. In the second and third stages of HOSNeRF, we utilize the photometric MSE loss, patched-based perceptual LPIPS [57] loss, forward-backward deformation cycle consistency, and indirect optical flow supervisions.", + "bbox": [ + 75, + 286, + 467, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The deformation cycle consistency is enforced by minimizing the distance between estimated deformed points of Eq. (12) and sampled deformed points $\\mathbf{x}_{\\mathrm{d}}$", + "bbox": [ + 76, + 407, + 468, + 454 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {C y c l e}} = \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left\\| \\mathbf {x} _ {\\mathrm {d}} ^ {i} - \\hat {\\mathbf {x}} _ {\\mathrm {d}} ^ {i} \\right\\| _ {2} ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 465, + 468, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also incorporate 2D optical flow as indirect supervision by minimizing the error between the induced flow and the estimated flow (details in the supplementary material). Therefore, the overall training objective of HOSNeRF is:", + "bbox": [ + 75, + 516, + 468, + 578 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\omega_ {\\mathrm {M S E}} \\cdot \\mathcal {L} _ {\\mathrm {M S E}} + \\omega_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} \\\\ + \\omega_ {\\text {C y c l e}} \\cdot \\mathcal {L} _ {\\text {C y c l e}} + \\omega_ {\\text {F l o w}} \\cdot \\mathcal {L} _ {\\text {F l o w}}, \\tag {17} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 589, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega_{\\mathrm{MSE}}$ , $\\omega_{\\mathrm{LPIPS}}$ , $\\omega_{\\mathrm{Cycle}}$ , $\\omega_{\\mathrm{Flow}}$ are loss weights.", + "bbox": [ + 76, + 637, + 434, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 666, + 207, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 691, + 171, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To benchmark the reconstruction of monocular videos with dynamic human-object-scenes, we collect a new dataset with 6 scenes $^{\\dagger}$ . The collected dataset features various types of human-object-scene interactions in indoor and outdoor scenarios, with up to 3 interacted objects and 7 states for a single video, as reported in Tab. 1. The duration of collected videos varies from 60s to 120s, and we extract [300, 400] frames for each video, where 16 frames at equal intervals are selected as novel views at novel timesteps as the test set and the remaining frames are the train set.", + "bbox": [ + 75, + 714, + 468, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We extensively evaluate the HOSNeRF on our collected dataset and NeuMan dataset [14] that consists of 6 short human walking sequences. Since NeuMan dataset does not involve human-object interactions, HOSNeRF can be flexibly customized to HOSNeRF by removing object bones and setting the object state to 1 when evaluated on NeuMan dataset.", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 191, + 715, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human interacts with various objects using hands in our dataset, we therefore introduce 2 new object bones to the left and right hand joints of the 3D human skeleton for all scenes. As a result, there are in total $K = 26$ bones in our human-object skeleton. In order to obtain the human-object pose prior, we first utilize a pretrained ROMP [45] model to estimate the 3D human pose, and use Mask-RCNN [11] to estimate human and objects masks. However, since it is non-trivial to estimate the object pose for in-the-wild videos, we initialize the left and right object bones as the standard extended bones with no relative rotations with respect to human hands, and initialize their length by referencing the length of the object relative to the arm. During training, we further refine the human-object poses $(\\mathcal{J},\\mathcal{R})$ with a MLP-based pose correction module $(\\Psi_{\\mathcal{J}},\\Psi_{\\mathcal{R}})$ [50]. The overall three-stage training of HOSNeRF takes about 5 days on 4 Tesla V100 GPUs. Please see the supplementary material for the comparison of training time with other approaches and discussions about estimating object masks.", + "bbox": [ + 496, + 215, + 890, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 513, + 700, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics. To demonstrate the performance of HOSNeRF, we make comparisons with various types of SOTA approaches, including (1) NeuMan [14], a human pose-based method; (2) HyperNeRF [30] and (3) Nerfies [29], dynamic radiance fields for general scenes; (4) $\\mathrm{D}^2\\mathrm{NeRF}$ [52], a static-dynamic decomposition method; (5) K-Planes [7], an explicit radiance field method. For quantitative comparison, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [49], and Learned Perceptual Image Patch Similarity (LPIPS) [57] are employed as evaluation metrics.", + "bbox": [ + 496, + 537, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on HOSNeRF Dataset. We report the quantitative comparisons (PSNR, SSIM, and LPIPS with VGG [38] backbone) of 6 scenes for all approaches in Tab. 2. For all SOTA approaches, we utilize their highest configurations for a fair comparison. As shown in Tab. 2, our HOSNeRF achieves the best performance in terms of all metrics, except the PSNR for the Lounge scene (a metric known to favor smooth/blurry results [57]). The improvement of HOSNeRF is particularly significant in terms of the LPIPS, with an average of nearly $40\\%$ gain over SOTA approaches. Fig. 4 visualizes the qualitative comparison of HOSNeRF over SOTA approaches on novel views at novel timesteps, where HOSNeRF achieves substantially better visual qual", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "†Our data collection process has been approved by the institutional review board and participants have given consents for public data release.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18488", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4bbe1f825c6f364d90efe660887f5f4a7cffc60e6c5cf5d437bdebba9534c8f6.jpg", + "image_caption": [ + "Figure 4: Qualitative comparisons of HOSNeRF and SOTA approaches on HOSNeRF dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 890, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ity than other approaches for all scenes. HOSNeRF is able to produce high-fidelity details close to ground truths for all scene contents, i.e., dynamic human bodies, objects, and backgrounds. In contrast, existing approaches tend to synthesize much blurrier images with missing components. Please see the supplementary video for more results on $360^{\\circ}$ free-viewpoint bullet-time videos rendered by HOSNeRF.", + "bbox": [ + 75, + 717, + 470, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on NeuMan Dataset [14]. We report quantitative metrics (PSNR, SSIM, and LPIPS with AlexNet [16] backbone) over 6 scenes of the NeuMan dataset in Tab. 3, where the metrics for NSFF [20], HyperNeRF [30], and NeuMan [14] are borrowed from NeuMan [14]. As shown", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Tab. 3, our HOSNeRF achieves the best performance in terms of all metrics except the PSNR for the Citron scene, and significantly improves SOTA approaches by a large margin of $50\\%$ in terms of LPIPS. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes. Fig. 5 visualizes the qualitative comparisons between HOSNeRF and NeuMan [14] for novel view synthesis at novel timesteps, where HOSNeRF achieves better visual quality over NeuMan [14] for the rendered human bodies and backgrounds in all scenes due to the superiority of our human model and background model. Our HOSNeRF also achieves bet-", + "bbox": [ + 496, + 717, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18489", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/98081892ffc021207d666b9d33d8b081fe8cd7f816f5edc0f6b316b6b2e49e30.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BACKPACKTENNISSUITCASEPLAYGROUNDDANCELOUNGE
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
K-Planes [7]19.050.3450.55719.310.6900.53618.640.5750.60217.920.4460.63518.170.3280.62324.210.8240.453
D2NeRF [52]20.520.3840.60823.970.7370.54020.990.5920.64521.230.5100.61619.920.4070.64727.130.8580.509
Nerfies [29]19.560.5150.55922.120.8180.44319.010.6430.55521.140.7260.53319.370.6740.52425.900.9140.342
HyperNeRF [30]19.620.3590.58721.260.7420.51019.410.6040.60721.670.5580.57819.300.4120.60127.250.9250.332
NeuMan [14]21.210.4790.47823.170.7680.44220.840.6110.55121.460.5460.55121.190.5290.49028.400.9170.341
Ours (base)21.510.7640.27124.020.9100.32621.100.8290.39522.200.7870.34821.840.7850.26626.540.9620.243
Ours w/ state22.380.7860.25223.980.9100.32321.430.8340.39022.520.7960.34122.430.7960.25827.510.9650.245
Ours w/ object21.560.7670.26923.980.9090.32721.080.8300.39622.140.7870.34721.880.7850.26926.800.9620.246
Ours w/ state, object22.330.7850.25324.080.9090.32021.510.8330.38622.550.7960.33822.380.7960.25927.520.9650.245
Ours w/ state, object, mask22.480.7900.24524.190.9110.32121.600.8340.38222.770.7990.33522.480.8020.25127.730.9680.227
Ours w/ state, object, mask, cycle22.610.7950.24124.190.9110.32121.710.8350.38222.680.8020.33622.540.8040.24927.880.9680.227
Ours (full)22.560.7920.24324.150.9110.32021.740.8360.38222.670.8010.33622.630.8040.24827.740.9680.227
", + "bbox": [ + 83, + 89, + 890, + 256 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c12bacac70a76c3dc884f9c016601dfeaac8f1dd1a67e94e9e7f1c087cd0ff8.jpg", + "table_caption": [ + "Table 2: Per-scene quantitative evaluation on the HOSNeRF dataset against baselines and ablations of our method. We color code each cell as best and second best." + ], + "table_footnote": [], + "table_body": "
SEATTLPARKINGBIKEJOGGINGLABCITRON
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NSFF [20]21.840.690.3721.980.690.4621.160.710.3620.630.530.4920.520.750.3912.330.490.65
HyperNeRF [30]16.430.430.4016.040.380.6217.640.420.4318.520.390.5216.750.510.2316.810.410.56
NeuMan [14]23.980.770.2625.430.790.3125.520.820.2322.680.670.3224.930.850.2124.710.800.26
Ours26.680.910.1427.200.930.1226.090.930.1025.040.890.1624.930.940.1024.440.900.18
", + "bbox": [ + 83, + 301, + 890, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Per-scene quantitative evaluation on the NeuMan dataset against baselines. We color code each cell as best", + "bbox": [ + 93, + 391, + 866, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ter foreground-background merging results, such as the feet regions of the Seattle scene and the Lab scene. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes.", + "bbox": [ + 75, + 420, + 468, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Per-Scene $360^{\\circ}$ Free-Viewpoint Renderings from Our HOSNeRF. We have also included a supplementary video to showcase the per-scene $360^{\\circ}$ free-viewpoint renderings from our HOSNeRF on all six scenes of our HOSNeRF dataset. The video highlights that our HOSNeRF is capable of generating high-fidelity details that accurately resemble all scene components, including dynamic human bodies, objects, and backgrounds. Notably, our HOSNeRF facilitates pausing the monocular video at any given point and rendering all scene details with high-fidelity from $360^{\\circ}$ viewpoints. To the best of our knowledge, our HOSNeRF represents the first work to achieve $360^{\\circ}$ free-viewpoint high-fidelity novel view synthesis for dynamic scenes featuring human-environment interactions from a single video.", + "bbox": [ + 75, + 481, + 468, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It should be emphasized that to enable $360^{\\circ}$ free-viewpoint rendering, the input videos must contain $360^{\\circ}$ scene information (an example capture process is shown in the supplementary video); otherwise, artifacts will appear in unobserved areas that were not even seen in the input videos. However, our HOSNeRF is not restricted by such capturing requirements, as demonstrated on the NeuMan dataset [14], which features small camera motions.", + "bbox": [ + 75, + 695, + 468, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 830, + 230, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies on our collected dataset to evaluate the effectiveness of each proposed component in HOSNeRF. We progressively ablate each component from", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "state embeddings, object bones, foreground masks, optical flow, and cycle consistency supervision. As shown in Tab. 2, the performance of HOSNeRF progressively drops with the disabling of each component, with the full model nearly achieving the best performances, which demonstrates the effectiveness of our designs.", + "bbox": [ + 496, + 420, + 890, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize representative scenes with large (Fig. 6) and small (Fig. 7) objects to further evaluate our proposed designs. As shown in Fig. 6(b) and 7(b), the base models are unable to represent the interacted objects for the human and the scene. In contrast, the proposed state embedding is capable of representing the state transitions of small objects (Fig. 7(c)), but it fails to represent the deformations of large objects (Fig. 6(c)). Together with our proposed object bones, the deformation of large objects can be effectively estimated through the human-object skeleton hierarchy, enabling large objects modeling in Fig. 6(d). In addition, the full model can further improve the synthesis quality of fine details such as correcting the wrong right arm by comparing Fig. 7(c) and (d). These comparisons further demonstrate the importance and effectiveness of the proposed designs.", + "bbox": [ + 496, + 511, + 892, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 753, + 617, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a novel framework of HOSNeRF, the first work to achieve $360^{\\circ}$ free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video. To tackle the challenges of such a task, we first introduced the new object bones to the human skeleton hierarchy to effectively estimate the fast human-object motion in our dynamic human-object model. Then, we proposed the state-conditional", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18490", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd2404d411a9d05a50623dddb866afd3c10660eb94fc917d95f966b6eb6a27a4.jpg", + "image_caption": [ + "Figure 5: Qualitative comparisons of HOSNeRF and NeuMan [14] on the NeuMan dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 468 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/49941d43ad00be81ad71221b6a9a7a68643732672f7c0b2a682c07294d9a5d05.jpg", + "image_caption": [ + "(a) Groundtruth" + ], + "image_footnote": [], + "bbox": [ + 80, + 497, + 274, + 583 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dcc17051effb55641c0a1598c79b131552c8fa1e978d0065378bde7c13e991aa.jpg", + "image_caption": [ + "(b) Base Model" + ], + "image_footnote": [], + "bbox": [ + 276, + 497, + 468, + 583 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3fc6d737b3242ad436b628837fe323593d34682726ed548541f8ab29a4787f1d.jpg", + "image_caption": [ + "(c) With State Embedding", + "Figure 6: Ablation of HOSNeRF for large objects." + ], + "image_footnote": [], + "bbox": [ + 80, + 594, + 272, + 679 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c7f464820ea3d94f0904d54b050a755f54b566f16d46d4001805bd3c3aab72e0.jpg", + "image_caption": [ + "(d) With State Embedding & Object Bones" + ], + "image_footnote": [], + "bbox": [ + 274, + 594, + 468, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "human-object and scene representations for handling significant object state changes. With our designed training strategies, HOSNeRF produced significant improvements over SOTA approaches, and enabled $360^{\\circ}$ free-viewpoint rendering of any frame from a single video and supported rendering views with novel human and object poses.", + "bbox": [ + 75, + 732, + 468, + 823 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations and Future Work. Our HOSNeRF focuses on dynamic human-objects modeling and lacks the capability to represent dynamic backgrounds. It remains challenging but is worthwhile researching models for reconstructing high-fidelity dynamic unbounded backgrounds.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bd39841f155c75f89a4e85100277fe62106a7c4fed3066a935f06597dfebbe13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 498, + 696, + 583 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1137a1d4414030a78649c32f490cb841ebf302a3f74c3326a4660eedfb9d99de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 498, + 890, + 583 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9d915738526257303c24a166225afcc53827c504785696ea4d74f7588430b45d.jpg", + "image_caption": [ + "(a) Groundtruth", + "(c) With State Embedding", + "Figure 7: Ablation of HOSNeRF for small objects." + ], + "image_footnote": [], + "bbox": [ + 501, + 594, + 696, + 680 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/723c40e71ce5e3d6a6784e7525591d920e81eb91f2b34a3cbdc5be2899cdcf8a.jpg", + "image_caption": [ + "(b) Base Model", + "(d) Full Model" + ], + "image_footnote": [], + "bbox": [ + 697, + 594, + 890, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 713, + 687, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, Mike Zheng Shou's Start-Up Grant from NUS, the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001. This research is also supported by National Research Foundation, Singapore and A\\*STAR, under its RIE 2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (grant no. I2001E0059). Jia-Wei Liu is also supported by NUS IDS-ISEP scholarship. Thanks to Yufei Shi for helpful discussions.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18491", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020.", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022.", + "[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023.", + "[4] Mingfei Chen, Jianfeng Zhang, Xiangyu Xu, Lijuan Liu, Yu-jun Cai, Jiashi Feng, and Shuicheng Yan. Geometry-guided progressive nerf for generalizable and efficient neural human rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 222-239. Springer, 2022.", + "[5] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015.", + "[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022.", + "[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023.", + "[8] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021.", + "[9] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–12, 2022.", + "[10] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021.", + "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017.", + "[12] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020.", + "[13] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 362-379. Springer, 2016.", + "[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 402-418. Springer, 2022.", + "[15] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022.", + "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017.", + "[17] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021.", + "[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 419-436. Springer, 2022.", + "[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022.", + "[20] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021.", + "[21] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. arXiv preprint arXiv:2211.11082, 2022.", + "[22] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022.", + "[23] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021.", + "[24] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019.", + "[25] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural vol" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18492", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "umes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019.", + "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015.", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021.", + "[28] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015.", + "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021.", + "[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021.", + "[31] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021.", + "[32] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021.", + "[33] Sergey Prokudin, Michael J Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1810-1819, 2021.", + "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021.", + "[35] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020.", + "[36] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yan-Pei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15872-15882, 2022.", + "[37] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthe" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022.", + "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[39] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019.", + "[40] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022.", + "[41] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum, volume 27, pages 1781-1787. Wiley Online Library, 2008.", + "[42] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021.", + "[43] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[44] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021.", + "[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021.", + "[46] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6226-6237, 2021.", + "[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021.", + "[48] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022." + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18493", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004.", + "[50] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022.", + "[51] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020.", + "[52] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D'nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems, 35:32653-32666, 2022.", + "[53] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021.", + "[54] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34:14955–14966, 2021.", + "[55] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021.", + "[56] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. arXiv preprint arXiv:2212.07626, 2022.", + "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018.", + "[58] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004.", + "[59] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014." + ], + "bbox": [ + 78, + 90, + 470, + 823 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18494", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_model.json b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3fed8d5227219918f28aff8d35e069f57c4063f3 --- /dev/null +++ b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_model.json @@ -0,0 +1,2644 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.862, + 0.175 + ], + "angle": 0, + "content": "HOSNeRF: Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video" + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.203, + 0.807, + 0.24 + ], + "angle": 0, + "content": "Jia-Wei Liu\\(^{1*}\\), Yan-Pei Cao\\(^{2\\dagger}\\), Tianyuan Yang\\(^{1}\\), Zhongcong Xu\\(^{1}\\), Jussi Keppo\\(^{4,5}\\), Ying Shan\\(^{2}\\), Xiaohu Qie\\(^{3}\\), Mike Zheng Shou\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.256, + 0.9, + 0.294 + ], + "angle": 0, + "content": "1 Show Lab, National University of Singapore 2 ARC Lab, 3 Tencent PCG 4 Business School, 5 Institute of Operations Research and Analytics, National University of Singapore" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.474, + 0.678 + ], + "angle": 0, + "content": "We introduce HOSNeRF, a novel \\(360^{\\circ}\\) free-viewpoint rendering method that reconstructs neural radiance fields for dynamic human-object-scene from a single monocular in-the-wild video. Our method enables pausing the video at any frame and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints. The first challenge in this task is the complex object motions in human-object interactions, which we tackle by introducing the new object bones into the conventional human skeleton hierarchy to effectively estimate large object deformations in our dynamic human-object model. The second challenge is that humans interact with different objects at different times, for which we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively. Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by a large margin of \\(40\\% \\sim 50\\%\\) in terms of LPIPS. The code, data, and compelling examples of \\(360^{\\circ}\\) free-viewpoint renderings from single videos: https://showlab.github.io/HOSNeRF." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.705, + 0.21, + 0.721 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.851 + ], + "angle": 0, + "content": "Video reconstruction and free-viewpoint rendering offer innovative opportunities for creating immersive experiences, encompassing virtual reality, telepresence, metaverse, and 3D animation production. While reconstructing videos has the potential to enhance user engagement and provide more realistic environments, it also poses significant challenges in terms of monocular viewpoints [55, 48] and complicated human-environment interactions [50, 14]." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.851, + 0.468, + 0.866 + ], + "angle": 0, + "content": "In recent years, remarkable progress has been made in" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.328, + 0.895, + 0.584 + ], + "angle": 0, + "content": "novel view synthesis, particularly since the introduction of Neural Radiance Fields (NeRF) [27]. While initially limited to reconstructing static 3D scenes based on multi-view images, subsequent studies have proposed various approaches to address the challenge of dynamic view synthesis [34, 29, 30, 47, 20, 8]. NeRF-based techniques have evolved to either incorporate deformation fields that map dynamic fields to canonical NeRF spaces [34, 29, 30, 47], or model dynamic scenes as 4D spatio-temporal radiance fields [20, 8]. While these approaches have shown promising results in dynamic view synthesis, they are limited to simple deformations. Another line of research is specifically designed for dynamic neural human modeling that relies on estimated human poses as a priori information [32, 50]. Recently, Neuman [14] combines pose-driven dynamic human models with static scene models for representing dynamic human-centric scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.894, + 0.698 + ], + "angle": 0, + "content": "However, none of the aforementioned techniques can accurately reconstruct challenging monocular videos with fast and complex human-object-scene motions and interactions, as shown in Fig. 1(d). This is due to two particular challenges listed below. To tackle them, we introduce a novel method called Human-Object-Scene Neural Radiance Fields (HOSNeRF)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "i) Complex object motions in human-object interactions. In contrast to the simple motions that can be modeled by general deformation modules [29, 30], the object motion during human-object interaction is often drastic and composed of various atomic motions (e.g., play tennis). To tackle this challenge, we propose new object bones that are attached to the human skeleton hierarchy to estimate human-object deformations in a coarse-to-fine manner for our dynamic human-object model. The object bones and underlying object linear blend skinning (object LBS) allow for the accurate estimation of objects' deformations through the relative transformations in the kinematic tree of the skeleton hierarchy." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.461, + 0.888 + ], + "angle": 0, + "content": "*Work is partially done during internship at ARC Lab, Tencent PCG." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.231, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.272, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.089, + 0.468, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.089, + 0.723, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.889, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.894, + 0.37 + ], + "angle": 0, + "content": "Figure 1: Our HOSNeRF (b) takes a single monocular in-the-wild video (a) as input, and creates high-fidelity \\(360^{\\circ}\\) free-viewpoint rendering of all scene details (dynamic human body, objects, and background) at any time (d). Our method enables rendering views with novel object poses and novel human poses as shown in (c), and produces high-fidelity dynamic novel view synthesis results at novel timesteps, with significant improvements over SOTA approaches as shown in (d)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.47, + 0.528 + ], + "angle": 0, + "content": "ii) Humans interact with different objects at different times. The above human-object model is designed for humans interacting with the same object over time. But when the person puts down the current object or picks a new object, it is not clear how to dynamically remove/add such objects in the static background model and the human-object model whose canonical space is static. To handle this challenge, we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Finally, we systematically explore and identify effective training objectives and strategies for our proposed HOSNeRF, including deformation cycle consistency, optical flow supervisions, and foreground-background rendering. On two challenging datasets collected by ourselves and NeuMan [14], our HOSNeRF achieves high-fidelity dynamic novel view synthesis results and enables pausing the monocular video at any time and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints, as shown in Fig. 1(d)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.684, + 0.365, + 0.699 + ], + "angle": 0, + "content": "In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.713, + 0.469, + 0.774 + ], + "angle": 0, + "content": "- We present a novel framework of HOSNeRF, the first work to achieve \\(360^{\\circ}\\) free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.792, + 0.469, + 0.837 + ], + "angle": 0, + "content": "- We propose the object bones and state-conditional representations to handle the non-rigid motions and interactions of humans, objects, and the environment." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by \\(40\\% \\sim 50\\%\\) in terms of LPIPS." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.713, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.642, + 0.391 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.403, + 0.746, + 0.42 + ], + "angle": 0, + "content": "2.1. Dynamic Human Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Dynamic human modeling has shown promising results in utilizing various representations such as point clouds [1, 51], meshes [10, 24], voxels [25, 39], and neural implicit functions [12, 35], with models like SMPL [26, 33] being commonly used for parameterizing the human body. Since the introduction of NeRF [27], neural human representation [32, 36, 42, 54, 23] has achieved remarkable progress on representing dynamic human bodies from sparse-view videos. Among them, Neural Actor [23] and Neural Body [32] pioneer in combining NeRF [27] with SMPL deformable meshes to represent human bodies with complex motions. Subsequent works have further improved on the generalizability [17, 4, 9] and animatability [31, 18] of human bodies. To support multi-person modeling, recent works [55, 37] have proposed to segment each human into 3D bounding boxes and learn a separate layered dynamic NeRF for each person. Other works [46, 43, 15, 44] are specifically designed to reconstruct the dynamic human and object with RGB-D or multi-view videos as inputs. They track the human and object pose, and separately reconstruct them with volumetric fusion [43], neural texture blending [44], or neural rendering [15, 56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Despite achieving promising results, these approaches require multi-view videos or RGB-D as input, limiting their real-world applications. To solve this problem, HumanNeRF [50] is proposed to represent moving humans from a monocular video by the human pose-driven deformation module and canonical space. NeuMan [14] is the first successful attempt at reconstructing both the dynamic human and static background from a single video. However, Neuman [14] does not support human-environment interactions" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18484" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.38, + 0.106 + ], + "angle": 0, + "content": "and performs poorly at large camera motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.114, + 0.45, + 0.131 + ], + "angle": 0, + "content": "2.2. Dynamic View Synthesis for General Scenes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.473, + 0.486 + ], + "angle": 0, + "content": "Most prior approaches on dynamic scene modeling require synchronized multi-view videos [58, 19, 41, 5, 55, 48] or depth [59, 13, 28] as input. Recent studies have built upon NeRF [27] to reconstruct dynamic neural radiance fields from monocular videos by either learning a deformation field that maps dynamic observation to canonical field [34, 29, 30, 47] or building 4D spatio-temporal radiance fields [53, 20, 8]. Among them, Nerfies [29] associates latent codes with the deformation field and HyperNeRF [30] represents motion in a high-dimensional space. \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [52] builds upon HyperNeRF [30] and further decouples the dynamic components from the static background, and represents them separately with a HyperNeRF [30] and NeRF [27]. DynIBaR [21] proposes a motion-adjusted multi-view feature aggregation module to synthesize new viewpoints by aggregating features from nearby views. Other studies have introduced voxel grids [6, 22, 40] or planar representations [7, 3] for fast dynamic radiance fields reconstruction. While these approaches have achieved high-fidelity dynamic view synthesis results, they are restricted to simple scene deformations. In contrast, our HOSNeRF is capable of representing significant human-object motions and interactions in complex environments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.169, + 0.513 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.522, + 0.218, + 0.537 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.639 + ], + "angle": 0, + "content": "HumanNeRF [50] has been recently introduced to represent a moving person with a NeRF [27]-based canonical space \\(\\Psi_{\\mathrm{c}}\\) that maps 3D points to color \\(\\mathbf{c}\\) and density \\(d\\), and a human pose-guided deformation field \\(\\Psi_{\\mathrm{d}}\\) that transforms deformed points \\(\\mathbf{x}_{\\mathrm{d}}^{i}\\) from the deformed space to canonical points \\(\\mathbf{x}_{\\mathrm{c}}^{i}\\) in the canonical space (\\(i\\) omitted for simplicity)." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.645, + 0.47, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {\\mathrm {c}} (\\gamma (\\mathbf {x} _ {\\mathrm {c}})) \\longmapsto (\\mathbf {c}, \\sigma), \\Psi_ {\\mathrm {d}} (\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}) \\longmapsto (\\mathbf {x} _ {\\mathrm {c}}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.716 + ], + "angle": 0, + "content": "where \\(\\gamma (\\mathbf{x})\\) is the standard positional encoding function, and \\(\\mathcal{I} = \\{\\mathbf{J}_i\\}\\) and \\(\\mathcal{R} = \\{\\omega_i\\}\\) represent the 3D human joints and the local joint axis-angle rotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.469, + 0.763 + ], + "angle": 0, + "content": "The deformation field \\(\\Psi_{\\mathrm{d}}\\) is decomposed into the coarse human skeleton-driven deformation \\(\\Psi_{\\mathrm{d}}^{\\mathrm{coarse}}\\), and the fine non-rigid deformation conditioned on human poses \\(\\Psi_{\\mathrm{d}}^{\\mathrm{fine}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.784, + 0.469, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {c}} ^ {\\prime} = \\Psi_ {\\mathrm {d}} ^ {\\mathrm {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right), \\mathbf {x} _ {\\mathrm {c}} = \\mathbf {x} _ {\\mathrm {c}} ^ {\\prime} + \\Psi_ {\\mathrm {d}} ^ {\\mathrm {f i n e}} \\left(\\mathbf {x} _ {\\mathrm {c}} ^ {\\prime}, \\mathcal {R}\\right). (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Mip-NeRF 360 [2] is designed to synthesize realistic views for highly intricate, unbounded real-world static scenes. To render pixel colors, the casted rays are split into a set of intervals \\( T_{i} = [t_{i}, t_{i + 1}) \\). The mean and covariance of the conical frustum corresponding to every interval are computed as \\( (\\boldsymbol{\\mu}, \\boldsymbol{\\Sigma}) = \\mathbf{r}(T_{i}) \\). To parameterize the Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "parameters for unbounded scenes, Mip-NeRF 360 [2] further proposes a contraction function \\( f(\\mathbf{x}) \\) that distributes distant points proportionally to disparity," + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.148, + 0.892, + 0.189 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = \\left\\{ \\begin{array}{c l} \\mathbf {x} & \\| \\mathbf {x} \\| \\leq 1 \\\\ \\left(2 - \\frac {1}{\\| \\mathbf {x} \\|}\\right) \\left(\\frac {\\mathbf {x}}{\\| \\mathbf {x} \\|}\\right) & \\| \\mathbf {x} \\| > 1 \\end{array} . \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.791, + 0.214 + ], + "angle": 0, + "content": "Then, \\(f(\\mathbf{x})\\) is applied to \\((\\pmb {\\mu},\\pmb {\\Sigma})\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.224, + 0.892, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\left(\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}\\right) = \\left(f (\\boldsymbol {\\mu}), \\mathbf {J} _ {f} (\\boldsymbol {\\mu}) \\boldsymbol {\\Sigma} \\mathbf {J} _ {f} (\\boldsymbol {\\mu}) ^ {\\mathrm {T}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.307 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}_f(\\boldsymbol{\\mu})\\) is the Jacobian of \\(f\\) at \\(\\boldsymbol{\\mu}\\). The contracted Gaussian parameters \\((\\hat{\\boldsymbol{\\mu}},\\hat{\\boldsymbol{\\Sigma}})\\) are further encoded through the integrated positional encoding (IPE) [2]:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.317, + 0.892, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}) = \\left\\{\\left[ \\begin{array}{l} \\sin \\left(2 ^ {\\ell} \\hat {\\boldsymbol {\\mu}}\\right) \\exp \\left(- 2 ^ {2 \\ell - 1} \\operatorname {d i a g} (\\hat {\\boldsymbol {\\Sigma}})\\right) \\\\ \\cos \\left(2 ^ {\\ell} \\hat {\\boldsymbol {\\mu}}\\right) \\exp \\left(- 2 ^ {2 \\ell - 1} \\operatorname {d i a g} (\\hat {\\boldsymbol {\\Sigma}})\\right) \\end{array} \\right] \\right\\} _ {\\ell = 0} ^ {L - 1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.371, + 0.871, + 0.385 + ], + "angle": 0, + "content": "and the color and density of intervals can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.396, + 0.892, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {\\mathrm {s}} \\left(\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}})\\right) \\longmapsto (\\mathbf {c}, \\sigma), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.431, + 0.754, + 0.446 + ], + "angle": 0, + "content": "where \\(\\Psi_{\\mathrm{s}}\\) is the scene NeRF MLP [2]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.779, + 0.471 + ], + "angle": 0, + "content": "3.2. Dynamic Human-Object Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Motivation. With estimated 3D human poses a priori, HumanNeRF [50] is effective at modeling moving people, but can only encode body parts, lacking the capability to model additional structures (e.g., objects held by the person) and hence not suitable for challenging scenarios containing complex human-object interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.675 + ], + "angle": 0, + "content": "A naive approach would be attaching the objects directly to the interacting body parts (e.g., hands) so that the object transformations can be queried via the skeletal motion of the corresponding human joints. However, this approach cannot represent the relative transformations between the object and its interacting body part, making it difficult to model large objects, such as the suitcase shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Object Bones. To address this limitation, based on the conventional human skeleton hierarchy, we introduce a new \"object bone\" for each hand, where the starting point of the object bone is connected to the corresponding hand joint. As shown in Fig. 2 (a), two new object bones (green color) are connected to the hand joints, which represent the relative rotation and translation with respect to their parent hand joints. In potential, the number, bone size, and attaching joints of object bones can also be flexibly customized based on different human-object interaction scenarios. Therefore, our 3D human-object pose consists of 3D joints \\(\\mathcal{I} = \\{\\mathbf{J}_{\\mathrm{human}},\\mathbf{J}_{\\mathrm{object}}\\}\\) and their axis-angle rotations \\(\\mathcal{R} = \\{\\omega_{\\mathrm{human}},\\omega_{\\mathrm{object}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Object LBS. To drive the articulated motions of the objects and human body, we learn a linear blend skinning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18485" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.895, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.895, + 0.43 + ], + "angle": 0, + "content": "Figure 2: Overview of our method. (1) Orange flowchart: The deformation from deformed points to canonical points are effectively estimated by the human-object backward LBS (a) and non-rigid deformation module, and their properties (i.e., density, color) can be obtained by querying the state-conditional canonical space. (2) Green flowchart: The background intervals sampled from a deformed frame are concatenated with object state embeddings for querying properties through the state-conditional scene model. (3) Novel views can be accordingly synthesized by volume rendering for re-ordered properties." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.471, + 0.543 + ], + "angle": 0, + "content": "(LBS) weights field during training, as shown in Fig. 2 (a), which encodes the influence regions of object bones and human bones. Following HumanNeRF [50], we pack the human and object bone blend weights and additional background weight into a single volume with \\( K + 1 \\) channels, i.e., \\( \\mathcal{W}_{\\mathrm{c} \\to \\mathrm{d}}(\\mathbf{x}_{\\mathrm{c}}) = \\left\\{w_{\\mathrm{c} \\to \\mathrm{d}}^{i}(\\mathbf{x}_{\\mathrm{c}})\\right\\} \\), and generate the volume from a learnable latent code using a CNN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.709 + ], + "angle": 0, + "content": "For instance, in Fig. 2, the subject holds a backpack in the right hand, so the learned LBS weight field only encodes the influence region of the right object bone with a similar shape to the backpack, while the left object bone has no influence on skeleton motion. Therefore, this learnable LBS weights field enables our object bones to be generalizable to various human-object interaction scenarios with left, right, or both hands. In addition, our human-object skeleton hierarchy enables HOSNeRF to support rendering with novel object pose and novel human pose, as illustrated in Fig. 1(c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.47, + 0.842 + ], + "angle": 0, + "content": "Forward-Backward Deformation Framework. To improve the smoothness and consistency of human-object deformations, we propose to leverage the cycle consistency between forward and backward human-object deformations. The backward deformation first transforms sampled deformed points \\(\\mathbf{x}_{\\mathrm{d}}\\) to their corresponding canonical points \\(\\mathbf{x}_{\\mathrm{c}}\\), which are further mapped back to the deformed space \\(\\hat{\\mathbf{x}}_{\\mathrm{d}}\\) through the forward deformation module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.904 + ], + "angle": 0, + "content": "As illustrated in Fig. 2(a), with the proposed human-object skeleton definition and LBS weight fields, the coarse backward skeleton deformation \\(\\Psi_{\\mathrm{d}\\to \\mathrm{c}}^{\\mathrm{coarse}}\\) is defined similar to" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.437, + 0.628, + 0.452 + ], + "angle": 0, + "content": "HumanNeRF [50]," + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.46, + 0.892, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right) = \\sum_ {i = 1} ^ {K} w _ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {i} \\left(\\mathbf {x} _ {\\mathrm {d}}\\right)\\left(\\mathbf {R} _ {\\mathrm {d} \\rightarrow c} ^ {i} \\mathbf {x} _ {\\mathrm {d}} + \\mathbf {t} _ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.574 + ], + "angle": 0, + "content": "where \\(K\\) is the number of human-object bones, \\(w_{\\mathrm{d}\\rightarrow \\mathrm{c}}^i\\) is the blend weight for the \\(i\\)-th bone at the deformed space, and \\(\\mathbf{R}_{\\mathrm{d}\\to \\mathrm{c}}^i\\) and \\(\\mathbf{t}_{\\mathrm{d}\\to \\mathrm{c}}^i\\) are the backward rotation and translation that are explicitly derived from the human-object pose." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.576, + 0.892, + 0.636 + ], + "angle": 0, + "content": "To model the non-rigid motion that can not be represented by skeleton motion, we additionally define the backward fine non-rigid motion \\(\\Psi_{\\mathrm{d}\\to c}^{\\mathrm{fine}}\\) conditioned on the human-object poses [50]:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.644, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {x} _ {\\mathrm {d} \\rightarrow \\mathrm {c}} = \\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\mathrm {f i n e}} \\left(\\gamma \\left(\\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\mathrm {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {I}, \\mathcal {R}\\right)\\right), \\mathcal {R}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.671, + 0.866, + 0.687 + ], + "angle": 0, + "content": "Therefore, \\(\\mathbf{x}_{\\mathrm{d}}\\) can be mapped to the canonical space by" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.696, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {c}} = \\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right) + \\Delta \\mathbf {x} _ {\\mathrm {d} \\rightarrow \\mathrm {c}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.722, + 0.892, + 0.752 + ], + "angle": 0, + "content": "Accordingly, the forward coarse skeleton motion driven by the human-object pose can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.761, + 0.892, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right) = \\sum_ {i = 1} ^ {K} w _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i} \\left(\\mathbf {x} _ {\\mathrm {c}}\\right)\\left(\\mathbf {R} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i} \\mathbf {x} _ {\\mathrm {c}} + \\mathbf {t} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.815, + 0.892, + 0.877 + ], + "angle": 0, + "content": "where \\(w_{\\mathrm{c}\\rightarrow \\mathrm{d}}^i\\) is the canonical blend weight for the \\(i\\)-th bone, \\(\\mathbf{R}_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{i}\\) and \\(\\mathbf{t}_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{i}\\) are the forward rotation and translation obtained from the human-object pose. In addition, the forward fine non-rigid motion \\(\\Psi_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{\\mathrm{fine}}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {x} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} = \\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {f i n e}} \\left(\\gamma \\left(\\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right)\\right), \\mathcal {R}\\right). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18486" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.102 + ], + "angle": 0, + "content": "\\(\\gamma (\\hat{\\mu},\\hat{\\Sigma})\\) IPE feature \\(\\gamma (\\mathbf{x}_c)\\) Embed points \\(\\mathbf{O}_s^i\\mathbf{O}_c^i i\\) th scene and canonical state embedding" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.104, + 0.468, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.193, + 0.22, + 0.203 + ], + "angle": 0, + "content": "T=0 The human starts walking" + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.193, + 0.421, + 0.203 + ], + "angle": 0, + "content": "T=6 The human picks up the bag" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.204, + 0.468, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.317, + 0.47, + 0.348 + ], + "angle": 0, + "content": "Figure 3: State-conditional representation for the scene model (a) and human-object canonical space (b)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.469, + 0.387 + ], + "angle": 0, + "content": "Therefore, the canonical points \\(\\mathbf{x}_{\\mathrm{c}}\\) can be transformed back to the deformed scene by" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.397, + 0.469, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {\\mathrm {d}} = \\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right) + \\Delta \\mathbf {x} _ {\\mathrm {c} \\rightarrow \\mathrm {d}}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.425, + 0.371, + 0.44 + ], + "angle": 0, + "content": "3.3. State-Conditional Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Motivation. In complex dynamic scenes, humans can interact with different objects at different timesteps. As the example in Fig. 3 shows, the human picks up the bag at \\( T = 6 \\) and puts it down at \\( T = 24 \\), resulting in 3 object states: the bag in the background during \\( T \\in [0, 6] \\), the bag held by the human during \\( T \\in [6, 24] \\), the bag at a new position in the background when \\( T > 24 \\). Such object state changes prohibit us from directly using the static background model (i.e., Mip-NeRF 360 [2]) and the dynamic human-object model proposed in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.469, + 0.704 + ], + "angle": 0, + "content": "A naive solution is to train a separate network for each object state. However, this solution is limited because each network is trained only on a short segment (e.g., \\( T \\in [0, 6] \\)) instead of the whole video. As a result, some regions often are difficult to reconstruct because they are not observed in the short segment. But these regions could have already been captured by video segments of other object states." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.795 + ], + "angle": 0, + "content": "To make full use of all segments to jointly train one model shared for different states, we introduce learnable state embeddings to represent different object states. As depicted in Fig. 3, we have two state embeddings respectively for our scene model and human-object model, which are explained in detail as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "State-Conditional Scene Model. As shown in Fig. 3(a), we develop a state-conditional Mip-NeRF 360 [2] to represent the static scene with temporal object transitions. In a dynamic scene with \\(N\\) object states, we define \\(N\\) learnable state embeddings \\(\\mathcal{O}_{\\mathrm{s}} = \\{\\mathbf{O}_{\\mathrm{s}}^{i}\\) (\\(i = 1,2,\\dots ,N\\)) to represent states information. Therefore, at state \\(i\\), we concatenate the IPE features \\(\\hat{\\gamma} (\\hat{\\mu},\\hat{\\Sigma})\\) (Eq. (5)) of ray intervals with" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the state embedding \\(\\mathbf{O}_{\\mathrm{s}}^{i}\\) as input to the scene MLP \\(\\Psi_{\\mathrm{s}}\\) for querying the scene color and density." + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.128, + 0.892, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {\\mathrm {s}} \\left(\\operatorname {c o n c a t} \\left(\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}), \\mathbf {O} _ {\\mathrm {s}} ^ {i}\\right)\\right) \\longmapsto (\\mathbf {c}, \\sigma). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.159, + 0.891, + 0.189 + ], + "angle": 0, + "content": "This representation enables a shared scene model across the video with multiple object state transitions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.189, + 0.892, + 0.355 + ], + "angle": 0, + "content": "State-Conditional Dynamic Human-Object Model. In Sec. 3.2, we introduce the human-object pose-driven deformation module and the human-object canonical space for modeling dynamic human-objects. However, the reconstructed canonical space is insufficient in representing the temporal changes in object geometry and appearance when the subject interacts with new objects. To address this limitation, we condition the canonical space on the object states, as shown in Fig. 3(b). In particular, we employ an object state-conditional canonical space and optimize a shared blend weights field across all states." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.461 + ], + "angle": 0, + "content": "In a dynamic scene with \\(N\\) object states, we define \\(N\\) learnable state embeddings \\(\\mathcal{O}_{\\mathrm{c}} = \\{\\mathbf{O}_{\\mathrm{c}}^{i}\\} (i = 1,2,\\dots ,N)\\) to represent object states in the canonical space. Therefore, at object state \\(i\\), we concatenate the positionally encoded canonical points in Eq. (1) with the state embedding \\(\\mathbf{O}_{\\mathrm{c}}^{i}\\) as input to the canonical MLP for querying the human-object color and density." + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.468, + 0.892, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {\\mathrm {c}} \\left(\\operatorname {c o n c a t} \\left(\\gamma \\left(\\mathbf {x} _ {\\mathrm {c}}\\right), \\mathbf {O} _ {\\mathrm {c}} ^ {i}\\right)\\right) \\longmapsto (\\mathbf {c}, \\sigma). \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.492, + 0.605, + 0.507 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Rendering. To render the pixel color, we shoot two rays \\((\\mathbf{r}_{\\mathrm{s}},\\mathbf{r}_{\\mathrm{ho}})\\) and sample intervals (i.e., ray points) to query the scene model (using \\(\\mathbf{r}_{\\mathrm{s}}\\)) and dynamic human-object model (with \\(\\mathbf{r}_{\\mathrm{ho}}\\)) respectively in their coordinate systems. After aligning the coordinates (see supplementary materials), we transform the sampled 3D points and their queried properties from the dynamic human-object space to the scene space, and re-order all sampled properties based on their distance from the camera center, as shown in Fig. 2. Therefore, the color of a pixel can be calculated through volume rendering, i.e., by integrating the re-ordered properties along the ray [27]:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\left(1 - e ^ {- \\sigma_ {i} \\delta_ {i}}\\right) \\mathbf {c} _ {i}, T _ {i} = e ^ {- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Directly fusing \\(\\mathbf{r}_{\\mathrm{s}}\\) and \\(\\mathbf{r}_{\\mathrm{ho}}\\) may yield minor artifacts on overlapped ground regions due to intersections of human-object points and scene points on short-length ground rays. To solve this issue, we estimate the foreground mask during training to separate the background rays from foreground ones. This allows us to focus on combining \\(\\mathbf{r}_{\\mathrm{s}}\\) and \\(\\mathbf{r}_{\\mathrm{ho}}\\) at the overlapped foreground regions, thereby reducing artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training Objectives. The training of HOSNeRF consists of three stages. In the first stage, we mask out the dynamic human-object regions and train the state-conditional" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18487" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.437, + 0.199 + ], + "angle": 0, + "content": "
SceneNo. of objectsNo. of states
BACKPACK13
TENNIS23
SUITCASE14
PLAYGROUND35
DANCE37
LOUNGE35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.202, + 0.404, + 0.216 + ], + "angle": 0, + "content": "Table 1: Details of our collected dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.468, + 0.287 + ], + "angle": 0, + "content": "scene model. Then, we train the state-conditional dynamic human-object model on the dynamic human-object regions. In the third stage, we combine these two models and further finetune the complete HOSNeRF for all image pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.468, + 0.408 + ], + "angle": 0, + "content": "Given the single video with calibrated poses, the first stage of HOSNeRF is optimized by minimizing the photometric MSE loss and the regularization losses proposed by Mip-NeRF 360 [2] to avoid background collapse. In the second and third stages of HOSNeRF, we utilize the photometric MSE loss, patched-based perceptual LPIPS [57] loss, forward-backward deformation cycle consistency, and indirect optical flow supervisions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.408, + 0.469, + 0.455 + ], + "angle": 0, + "content": "The deformation cycle consistency is enforced by minimizing the distance between estimated deformed points of Eq. (12) and sampled deformed points \\(\\mathbf{x}_{\\mathrm{d}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.466, + 0.469, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {C y c l e}} = \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left\\| \\mathbf {x} _ {\\mathrm {d}} ^ {i} - \\hat {\\mathbf {x}} _ {\\mathrm {d}} ^ {i} \\right\\| _ {2} ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.469, + 0.579 + ], + "angle": 0, + "content": "We also incorporate 2D optical flow as indirect supervision by minimizing the error between the induced flow and the estimated flow (details in the supplementary material). Therefore, the overall training objective of HOSNeRF is:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.59, + 0.469, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\omega_ {\\mathrm {M S E}} \\cdot \\mathcal {L} _ {\\mathrm {M S E}} + \\omega_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} \\\\ + \\omega_ {\\text {C y c l e}} \\cdot \\mathcal {L} _ {\\text {C y c l e}} + \\omega_ {\\text {F l o w}} \\cdot \\mathcal {L} _ {\\text {F l o w}}, \\tag {17} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.638, + 0.436, + 0.653 + ], + "angle": 0, + "content": "where \\(\\omega_{\\mathrm{MSE}}\\), \\(\\omega_{\\mathrm{LPIPS}}\\), \\(\\omega_{\\mathrm{Cycle}}\\), \\(\\omega_{\\mathrm{Flow}}\\) are loss weights." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.667, + 0.208, + 0.684 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.173, + 0.706 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.469, + 0.866 + ], + "angle": 0, + "content": "To benchmark the reconstruction of monocular videos with dynamic human-object-scenes, we collect a new dataset with 6 scenes\\(^{\\dagger}\\). The collected dataset features various types of human-object-scene interactions in indoor and outdoor scenarios, with up to 3 interacted objects and 7 states for a single video, as reported in Tab. 1. The duration of collected videos varies from 60s to 120s, and we extract [300, 400] frames for each video, where 16 frames at equal intervals are selected as novel views at novel timesteps as the test set and the remaining frames are the train set." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "†Our data collection process has been approved by the institutional review board and participants have given consents for public data release." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "We extensively evaluate the HOSNeRF on our collected dataset and NeuMan dataset [14] that consists of 6 short human walking sequences. Since NeuMan dataset does not involve human-object interactions, HOSNeRF can be flexibly customized to HOSNeRF by removing object bones and setting the object state to 1 when evaluated on NeuMan dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.717, + 0.209 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Human interacts with various objects using hands in our dataset, we therefore introduce 2 new object bones to the left and right hand joints of the 3D human skeleton for all scenes. As a result, there are in total \\( K = 26 \\) bones in our human-object skeleton. In order to obtain the human-object pose prior, we first utilize a pretrained ROMP [45] model to estimate the 3D human pose, and use Mask-RCNN [11] to estimate human and objects masks. However, since it is non-trivial to estimate the object pose for in-the-wild videos, we initialize the left and right object bones as the standard extended bones with no relative rotations with respect to human hands, and initialize their length by referencing the length of the object relative to the arm. During training, we further refine the human-object poses \\( (\\mathcal{J},\\mathcal{R}) \\) with a MLP-based pose correction module \\( (\\Psi_{\\mathcal{J}},\\Psi_{\\mathcal{R}}) \\) [50]. The overall three-stage training of HOSNeRF takes about 5 days on 4 Tesla V100 GPUs. Please see the supplementary material for the comparison of training time with other approaches and discussions about estimating object masks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.702, + 0.53 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics. To demonstrate the performance of HOSNeRF, we make comparisons with various types of SOTA approaches, including (1) NeuMan [14], a human pose-based method; (2) HyperNeRF [30] and (3) Nerfies [29], dynamic radiance fields for general scenes; (4) \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [52], a static-dynamic decomposition method; (5) K-Planes [7], an explicit radiance field method. For quantitative comparison, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [49], and Learned Perceptual Image Patch Similarity (LPIPS) [57] are employed as evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on HOSNeRF Dataset. We report the quantitative comparisons (PSNR, SSIM, and LPIPS with VGG [38] backbone) of 6 scenes for all approaches in Tab. 2. For all SOTA approaches, we utilize their highest configurations for a fair comparison. As shown in Tab. 2, our HOSNeRF achieves the best performance in terms of all metrics, except the PSNR for the Lounge scene (a metric known to favor smooth/blurry results [57]). The improvement of HOSNeRF is particularly significant in terms of the LPIPS, with an average of nearly \\(40\\%\\) gain over SOTA approaches. Fig. 4 visualizes the qualitative comparison of HOSNeRF over SOTA approaches on novel views at novel timesteps, where HOSNeRF achieves substantially better visual qual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18488" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.086, + 0.891, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.694, + 0.794, + 0.71 + ], + "angle": 0, + "content": "Figure 4: Qualitative comparisons of HOSNeRF and SOTA approaches on HOSNeRF dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.824 + ], + "angle": 0, + "content": "ity than other approaches for all scenes. HOSNeRF is able to produce high-fidelity details close to ground truths for all scene contents, i.e., dynamic human bodies, objects, and backgrounds. In contrast, existing approaches tend to synthesize much blurrier images with missing components. Please see the supplementary video for more results on \\(360^{\\circ}\\) free-viewpoint bullet-time videos rendered by HOSNeRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation on NeuMan Dataset [14]. We report quantitative metrics (PSNR, SSIM, and LPIPS with AlexNet [16] backbone) over 6 scenes of the NeuMan dataset in Tab. 3, where the metrics for NSFF [20], HyperNeRF [30], and NeuMan [14] are borrowed from NeuMan [14]. As shown" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.9 + ], + "angle": 0, + "content": "in Tab. 3, our HOSNeRF achieves the best performance in terms of all metrics except the PSNR for the Citron scene, and significantly improves SOTA approaches by a large margin of \\(50\\%\\) in terms of LPIPS. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes. Fig. 5 visualizes the qualitative comparisons between HOSNeRF and NeuMan [14] for novel view synthesis at novel timesteps, where HOSNeRF achieves better visual quality over NeuMan [14] for the rendered human bodies and backgrounds in all scenes due to the superiority of our human model and background model. Our HOSNeRF also achieves bet-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18489" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.257 + ], + "angle": 0, + "content": "
BACKPACKTENNISSUITCASEPLAYGROUNDDANCELOUNGE
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
K-Planes [7]19.050.3450.55719.310.6900.53618.640.5750.60217.920.4460.63518.170.3280.62324.210.8240.453
D2NeRF [52]20.520.3840.60823.970.7370.54020.990.5920.64521.230.5100.61619.920.4070.64727.130.8580.509
Nerfies [29]19.560.5150.55922.120.8180.44319.010.6430.55521.140.7260.53319.370.6740.52425.900.9140.342
HyperNeRF [30]19.620.3590.58721.260.7420.51019.410.6040.60721.670.5580.57819.300.4120.60127.250.9250.332
NeuMan [14]21.210.4790.47823.170.7680.44220.840.6110.55121.460.5460.55121.190.5290.49028.400.9170.341
Ours (base)21.510.7640.27124.020.9100.32621.100.8290.39522.200.7870.34821.840.7850.26626.540.9620.243
Ours w/ state22.380.7860.25223.980.9100.32321.430.8340.39022.520.7960.34122.430.7960.25827.510.9650.245
Ours w/ object21.560.7670.26923.980.9090.32721.080.8300.39622.140.7870.34721.880.7850.26926.800.9620.246
Ours w/ state, object22.330.7850.25324.080.9090.32021.510.8330.38622.550.7960.33822.380.7960.25927.520.9650.245
Ours w/ state, object, mask22.480.7900.24524.190.9110.32121.600.8340.38222.770.7990.33522.480.8020.25127.730.9680.227
Ours w/ state, object, mask, cycle22.610.7950.24124.190.9110.32121.710.8350.38222.680.8020.33622.540.8040.24927.880.9680.227
Ours (full)22.560.7920.24324.150.9110.32021.740.8360.38222.670.8010.33622.630.8040.24827.740.9680.227
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.261, + 0.893, + 0.293 + ], + "angle": 0, + "content": "Table 2: Per-scene quantitative evaluation on the HOSNeRF dataset against baselines and ablations of our method. We color code each cell as best and second best." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.302, + 0.891, + 0.387 + ], + "angle": 0, + "content": "
SEATTLPARKINGBIKEJOGGINGLABCITRON
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NSFF [20]21.840.690.3721.980.690.4621.160.710.3620.630.530.4920.520.750.3912.330.490.65
HyperNeRF [30]16.430.430.4016.040.380.6217.640.420.4318.520.390.5216.750.510.2316.810.410.56
NeuMan [14]23.980.770.2625.430.790.3125.520.820.2322.680.670.3224.930.850.2124.710.800.26
Ours26.680.910.1427.200.930.1226.090.930.1025.040.890.1624.930.940.1024.440.900.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.392, + 0.867, + 0.409 + ], + "angle": 0, + "content": "Table 3: Per-scene quantitative evaluation on the NeuMan dataset against baselines. We color code each cell as best" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.48 + ], + "angle": 0, + "content": "ter foreground-background merging results, such as the feet regions of the Seattle scene and the Lab scene. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Per-Scene \\(360^{\\circ}\\) Free-Viewpoint Renderings from Our HOSNeRF. We have also included a supplementary video to showcase the per-scene \\(360^{\\circ}\\) free-viewpoint renderings from our HOSNeRF on all six scenes of our HOSNeRF dataset. The video highlights that our HOSNeRF is capable of generating high-fidelity details that accurately resemble all scene components, including dynamic human bodies, objects, and backgrounds. Notably, our HOSNeRF facilitates pausing the monocular video at any given point and rendering all scene details with high-fidelity from \\(360^{\\circ}\\) viewpoints. To the best of our knowledge, our HOSNeRF represents the first work to achieve \\(360^{\\circ}\\) free-viewpoint high-fidelity novel view synthesis for dynamic scenes featuring human-environment interactions from a single video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.817 + ], + "angle": 0, + "content": "It should be emphasized that to enable \\(360^{\\circ}\\) free-viewpoint rendering, the input videos must contain \\(360^{\\circ}\\) scene information (an example capture process is shown in the supplementary video); otherwise, artifacts will appear in unobserved areas that were not even seen in the input videos. However, our HOSNeRF is not restricted by such capturing requirements, as demonstrated on the NeuMan dataset [14], which features small camera motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.232, + 0.848 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We conduct ablation studies on our collected dataset to evaluate the effectiveness of each proposed component in HOSNeRF. We progressively ablate each component from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.512 + ], + "angle": 0, + "content": "state embeddings, object bones, foreground masks, optical flow, and cycle consistency supervision. As shown in Tab. 2, the performance of HOSNeRF progressively drops with the disabling of each component, with the full model nearly achieving the best performances, which demonstrates the effectiveness of our designs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.893, + 0.739 + ], + "angle": 0, + "content": "We visualize representative scenes with large (Fig. 6) and small (Fig. 7) objects to further evaluate our proposed designs. As shown in Fig. 6(b) and 7(b), the base models are unable to represent the interacted objects for the human and the scene. In contrast, the proposed state embedding is capable of representing the state transitions of small objects (Fig. 7(c)), but it fails to represent the deformations of large objects (Fig. 6(c)). Together with our proposed object bones, the deformation of large objects can be effectively estimated through the human-object skeleton hierarchy, enabling large objects modeling in Fig. 6(d). In addition, the full model can further improve the synthesis quality of fine details such as correcting the wrong right arm by comparing Fig. 7(c) and (d). These comparisons further demonstrate the importance and effectiveness of the proposed designs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.754, + 0.619, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduced a novel framework of HOSNeRF, the first work to achieve \\(360^{\\circ}\\) free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video. To tackle the challenges of such a task, we first introduced the new object bones to the human skeleton hierarchy to effectively estimate the fast human-object motion in our dynamic human-object model. Then, we proposed the state-conditional" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18490" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.892, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.477, + 0.784, + 0.492 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparisons of HOSNeRF and NeuMan [14] on the NeuMan dataset." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.498, + 0.275, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.585, + 0.221, + 0.594 + ], + "angle": 0, + "content": "(a) Groundtruth" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.498, + 0.469, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.584, + 0.414, + 0.595 + ], + "angle": 0, + "content": "(b) Base Model" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.595, + 0.274, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.681, + 0.249, + 0.692 + ], + "angle": 0, + "content": "(c) With State Embedding" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.595, + 0.469, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.681, + 0.442, + 0.704 + ], + "angle": 0, + "content": "(d) With State Embedding & Object Bones" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.707, + 0.439, + 0.722 + ], + "angle": 0, + "content": "Figure 6: Ablation of HOSNeRF for large objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.469, + 0.824 + ], + "angle": 0, + "content": "human-object and scene representations for handling significant object state changes. With our designed training strategies, HOSNeRF produced significant improvements over SOTA approaches, and enabled \\(360^{\\circ}\\) free-viewpoint rendering of any frame from a single video and supported rendering views with novel human and object poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Limitations and Future Work. Our HOSNeRF focuses on dynamic human-objects modeling and lacks the capability to represent dynamic backgrounds. It remains challenging but is worthwhile researching models for reconstructing high-fidelity dynamic unbounded backgrounds." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.499, + 0.697, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.585, + 0.643, + 0.595 + ], + "angle": 0, + "content": "(a) Groundtruth" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.499, + 0.892, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.585, + 0.837, + 0.595 + ], + "angle": 0, + "content": "(b) Base Model" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.595, + 0.697, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.681, + 0.671, + 0.693 + ], + "angle": 0, + "content": "(c) With State Embedding" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.595, + 0.892, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.681, + 0.835, + 0.692 + ], + "angle": 0, + "content": "(d) Full Model" + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.696, + 0.863, + 0.711 + ], + "angle": 0, + "content": "Figure 7: Ablation of HOSNeRF for small objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.714, + 0.688, + 0.731 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, Mike Zheng Shou's Start-Up Grant from NUS, the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001. This research is also supported by National Research Foundation, Singapore and A\\*STAR, under its RIE 2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (grant no. I2001E0059). Jia-Wei Liu is also supported by NUS IDS-ISEP scholarship. Thanks to Yufei Shi for helpful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18491" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.184 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.468, + 0.254 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.468, + 0.283 + ], + "angle": 0, + "content": "[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[4] Mingfei Chen, Jianfeng Zhang, Xiangyu Xu, Lijuan Liu, Yu-jun Cai, Jiashi Feng, and Shuicheng Yan. Geometry-guided progressive nerf for generalizable and efficient neural human rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 222-239. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.435 + ], + "angle": 0, + "content": "[5] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.505 + ], + "angle": 0, + "content": "[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.468, + 0.562 + ], + "angle": 0, + "content": "[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[8] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.468, + 0.688 + ], + "angle": 0, + "content": "[9] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–12, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.468, + 0.745 + ], + "angle": 0, + "content": "[10] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[12] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 362-379. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 402-418. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[15] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[17] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 419-436. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[20] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[21] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. arXiv preprint arXiv:2211.11082, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[22] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[23] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural vol" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18492" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "umes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[28] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.373, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[31] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[32] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.47, + 0.65 + ], + "angle": 0, + "content": "[33] Sergey Prokudin, Michael J Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1810-1819, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.652, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[35] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.791, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[36] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yan-Pei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15872-15882, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthe" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "sis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.18, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[39] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[40] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[41] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum, volume 27, pages 1781-1787. Wiley Online Library, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[42] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.436, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[43] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.507, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[44] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.592, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[46] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6226-6237, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[48] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18493" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[50] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[51] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[52] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D'nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems, 35:32653-32666, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[53] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[54] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34:14955–14966, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[55] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[56] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. arXiv preprint arXiv:2212.07626, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[58] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.469, + 0.824 + ], + "angle": 0, + "content": "[59] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18494" + } + ] +] \ No newline at end of file diff --git a/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_origin.pdf b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b768bede4f64d87864cfcb04a274107075c6ab49 --- /dev/null +++ b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/56bcebb1-a8ab-4faa-8362-f67dd5395540_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ba4fc67019417f9d5d7b9044e7bc283cd8cd3b2780d5bbcb92753fbad2b487f +size 9018875 diff --git a/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/full.md b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7a37be13b601b1586145d37a4f7fc9aafafe175e --- /dev/null +++ b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/full.md @@ -0,0 +1,386 @@ +# HOSNeRF: Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video + +Jia-Wei Liu $^{1*}$ , Yan-Pei Cao $^{2\dagger}$ , Tianyuan Yang $^{1}$ , Zhongcong Xu $^{1}$ , Jussi Keppo $^{4,5}$ , Ying Shan $^{2}$ , Xiaohu Qie $^{3}$ , Mike Zheng Shou $^{1\dagger}$ + +1 Show Lab, National University of Singapore 2 ARC Lab, 3 Tencent PCG 4 Business School, 5 Institute of Operations Research and Analytics, National University of Singapore + +# Abstract + +We introduce HOSNeRF, a novel $360^{\circ}$ free-viewpoint rendering method that reconstructs neural radiance fields for dynamic human-object-scene from a single monocular in-the-wild video. Our method enables pausing the video at any frame and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints. The first challenge in this task is the complex object motions in human-object interactions, which we tackle by introducing the new object bones into the conventional human skeleton hierarchy to effectively estimate large object deformations in our dynamic human-object model. The second challenge is that humans interact with different objects at different times, for which we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively. Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by a large margin of $40\% \sim 50\%$ in terms of LPIPS. The code, data, and compelling examples of $360^{\circ}$ free-viewpoint renderings from single videos: https://showlab.github.io/HOSNeRF. + +# 1. Introduction + +Video reconstruction and free-viewpoint rendering offer innovative opportunities for creating immersive experiences, encompassing virtual reality, telepresence, metaverse, and 3D animation production. While reconstructing videos has the potential to enhance user engagement and provide more realistic environments, it also poses significant challenges in terms of monocular viewpoints [55, 48] and complicated human-environment interactions [50, 14]. + +In recent years, remarkable progress has been made in + +novel view synthesis, particularly since the introduction of Neural Radiance Fields (NeRF) [27]. While initially limited to reconstructing static 3D scenes based on multi-view images, subsequent studies have proposed various approaches to address the challenge of dynamic view synthesis [34, 29, 30, 47, 20, 8]. NeRF-based techniques have evolved to either incorporate deformation fields that map dynamic fields to canonical NeRF spaces [34, 29, 30, 47], or model dynamic scenes as 4D spatio-temporal radiance fields [20, 8]. While these approaches have shown promising results in dynamic view synthesis, they are limited to simple deformations. Another line of research is specifically designed for dynamic neural human modeling that relies on estimated human poses as a priori information [32, 50]. Recently, Neuman [14] combines pose-driven dynamic human models with static scene models for representing dynamic human-centric scenes. + +However, none of the aforementioned techniques can accurately reconstruct challenging monocular videos with fast and complex human-object-scene motions and interactions, as shown in Fig. 1(d). This is due to two particular challenges listed below. To tackle them, we introduce a novel method called Human-Object-Scene Neural Radiance Fields (HOSNeRF). + +i) Complex object motions in human-object interactions. In contrast to the simple motions that can be modeled by general deformation modules [29, 30], the object motion during human-object interaction is often drastic and composed of various atomic motions (e.g., play tennis). To tackle this challenge, we propose new object bones that are attached to the human skeleton hierarchy to estimate human-object deformations in a coarse-to-fine manner for our dynamic human-object model. The object bones and underlying object linear blend skinning (object LBS) allow for the accurate estimation of objects' deformations through the relative transformations in the kinematic tree of the skeleton hierarchy. + +![](images/6aa2f4bbcd3d7a80f3b0f41d901ff730a908862146905e5609d10befd3008413.jpg) +Figure 1: Our HOSNeRF (b) takes a single monocular in-the-wild video (a) as input, and creates high-fidelity $360^{\circ}$ free-viewpoint rendering of all scene details (dynamic human body, objects, and background) at any time (d). Our method enables rendering views with novel object poses and novel human poses as shown in (c), and produces high-fidelity dynamic novel view synthesis results at novel timesteps, with significant improvements over SOTA approaches as shown in (d). + +![](images/acdab11ba8a5c4aaf70fb8c93b8d7ba0981573fb2cab1e96c98e8df002e4f9dd.jpg) + +![](images/51c94a859fa255eaebf95973920c96b6d1fb8c11ec635692a5103e3844e88553.jpg) + +![](images/692d6f2eceb289e3397e21e4a40f1ca1da20d3c5c6eaf476d193942f7e69f6dc.jpg) + +ii) Humans interact with different objects at different times. The above human-object model is designed for humans interacting with the same object over time. But when the person puts down the current object or picks a new object, it is not clear how to dynamically remove/add such objects in the static background model and the human-object model whose canonical space is static. To handle this challenge, we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively. + +Finally, we systematically explore and identify effective training objectives and strategies for our proposed HOSNeRF, including deformation cycle consistency, optical flow supervisions, and foreground-background rendering. On two challenging datasets collected by ourselves and NeuMan [14], our HOSNeRF achieves high-fidelity dynamic novel view synthesis results and enables pausing the monocular video at any time and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints, as shown in Fig. 1(d). + +In summary, our main contributions are: + +- We present a novel framework of HOSNeRF, the first work to achieve $360^{\circ}$ free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video. +- We propose the object bones and state-conditional representations to handle the non-rigid motions and interactions of humans, objects, and the environment. +- Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by $40\% \sim 50\%$ in terms of LPIPS. + +# 2. Related Work + +# 2.1. Dynamic Human Modeling + +Dynamic human modeling has shown promising results in utilizing various representations such as point clouds [1, 51], meshes [10, 24], voxels [25, 39], and neural implicit functions [12, 35], with models like SMPL [26, 33] being commonly used for parameterizing the human body. Since the introduction of NeRF [27], neural human representation [32, 36, 42, 54, 23] has achieved remarkable progress on representing dynamic human bodies from sparse-view videos. Among them, Neural Actor [23] and Neural Body [32] pioneer in combining NeRF [27] with SMPL deformable meshes to represent human bodies with complex motions. Subsequent works have further improved on the generalizability [17, 4, 9] and animatability [31, 18] of human bodies. To support multi-person modeling, recent works [55, 37] have proposed to segment each human into 3D bounding boxes and learn a separate layered dynamic NeRF for each person. Other works [46, 43, 15, 44] are specifically designed to reconstruct the dynamic human and object with RGB-D or multi-view videos as inputs. They track the human and object pose, and separately reconstruct them with volumetric fusion [43], neural texture blending [44], or neural rendering [15, 56]. + +Despite achieving promising results, these approaches require multi-view videos or RGB-D as input, limiting their real-world applications. To solve this problem, HumanNeRF [50] is proposed to represent moving humans from a monocular video by the human pose-driven deformation module and canonical space. NeuMan [14] is the first successful attempt at reconstructing both the dynamic human and static background from a single video. However, Neuman [14] does not support human-environment interactions + +and performs poorly at large camera motions. + +# 2.2. Dynamic View Synthesis for General Scenes + +Most prior approaches on dynamic scene modeling require synchronized multi-view videos [58, 19, 41, 5, 55, 48] or depth [59, 13, 28] as input. Recent studies have built upon NeRF [27] to reconstruct dynamic neural radiance fields from monocular videos by either learning a deformation field that maps dynamic observation to canonical field [34, 29, 30, 47] or building 4D spatio-temporal radiance fields [53, 20, 8]. Among them, Nerfies [29] associates latent codes with the deformation field and HyperNeRF [30] represents motion in a high-dimensional space. $\mathrm{D}^2\mathrm{NeRF}$ [52] builds upon HyperNeRF [30] and further decouples the dynamic components from the static background, and represents them separately with a HyperNeRF [30] and NeRF [27]. DynIBaR [21] proposes a motion-adjusted multi-view feature aggregation module to synthesize new viewpoints by aggregating features from nearby views. Other studies have introduced voxel grids [6, 22, 40] or planar representations [7, 3] for fast dynamic radiance fields reconstruction. While these approaches have achieved high-fidelity dynamic view synthesis results, they are restricted to simple scene deformations. In contrast, our HOSNeRF is capable of representing significant human-object motions and interactions in complex environments. + +# 3. Method + +# 3.1. Preliminaries + +HumanNeRF [50] has been recently introduced to represent a moving person with a NeRF [27]-based canonical space $\Psi_{\mathrm{c}}$ that maps 3D points to color $\mathbf{c}$ and density $d$ , and a human pose-guided deformation field $\Psi_{\mathrm{d}}$ that transforms deformed points $\mathbf{x}_{\mathrm{d}}^{i}$ from the deformed space to canonical points $\mathbf{x}_{\mathrm{c}}^{i}$ in the canonical space ( $i$ omitted for simplicity). + +$$ +\Psi_ {\mathrm {c}} (\gamma (\mathbf {x} _ {\mathrm {c}})) \longmapsto (\mathbf {c}, \sigma), \Psi_ {\mathrm {d}} (\mathbf {x} _ {\mathrm {d}}, \mathcal {J}, \mathcal {R}) \longmapsto (\mathbf {x} _ {\mathrm {c}}), \tag {1} +$$ + +where $\gamma (\mathbf{x})$ is the standard positional encoding function, and $\mathcal{I} = \{\mathbf{J}_i\}$ and $\mathcal{R} = \{\omega_i\}$ represent the 3D human joints and the local joint axis-angle rotations. + +The deformation field $\Psi_{\mathrm{d}}$ is decomposed into the coarse human skeleton-driven deformation $\Psi_{\mathrm{d}}^{\mathrm{coarse}}$ , and the fine non-rigid deformation conditioned on human poses $\Psi_{\mathrm{d}}^{\mathrm{fine}}$ : + +$$ +\mathbf {x} _ {\mathrm {c}} ^ {\prime} = \Psi_ {\mathrm {d}} ^ {\mathrm {c o a r s e}} \left(\mathbf {x} _ {\mathrm {d}}, \mathcal {J}, \mathcal {R}\right), \mathbf {x} _ {\mathrm {c}} = \mathbf {x} _ {\mathrm {c}} ^ {\prime} + \Psi_ {\mathrm {d}} ^ {\mathrm {f i n e}} \left(\mathbf {x} _ {\mathrm {c}} ^ {\prime}, \mathcal {R}\right). (2) +$$ + +Mip-NeRF 360 [2] is designed to synthesize realistic views for highly intricate, unbounded real-world static scenes. To render pixel colors, the casted rays are split into a set of intervals $T_{i} = [t_{i}, t_{i + 1})$ . The mean and covariance of the conical frustum corresponding to every interval are computed as $(\boldsymbol{\mu}, \boldsymbol{\Sigma}) = \mathbf{r}(T_{i})$ . To parameterize the Gaussian + +parameters for unbounded scenes, Mip-NeRF 360 [2] further proposes a contraction function $f(\mathbf{x})$ that distributes distant points proportionally to disparity, + +$$ +f (\mathbf {x}) = \left\{ \begin{array}{c l} \mathbf {x} & \| \mathbf {x} \| \leq 1 \\ \left(2 - \frac {1}{\| \mathbf {x} \|}\right) \left(\frac {\mathbf {x}}{\| \mathbf {x} \|}\right) & \| \mathbf {x} \| > 1 \end{array} . \right. \tag {3} +$$ + +Then, $f(\mathbf{x})$ is applied to $(\pmb {\mu},\pmb {\Sigma})$ as follows: + +$$ +\left(\hat {\boldsymbol {\mu}}, \hat {\boldsymbol {\Sigma}}\right) = \left(f (\boldsymbol {\mu}), \mathbf {J} _ {f} (\boldsymbol {\mu}) \boldsymbol {\Sigma} \mathbf {J} _ {f} (\boldsymbol {\mu}) ^ {\mathrm {T}}\right), \tag {4} +$$ + +where $\mathbf{J}_f(\boldsymbol{\mu})$ is the Jacobian of $f$ at $\boldsymbol{\mu}$ . The contracted Gaussian parameters $(\hat{\boldsymbol{\mu}},\hat{\boldsymbol{\Sigma}})$ are further encoded through the integrated positional encoding (IPE) [2]: + +$$ +\hat {\gamma} (\hat {\boldsymbol {\mu}}, \hat {\boldsymbol {\Sigma}}) = \left\{\left[ \begin{array}{l} \sin \left(2 ^ {\ell} \hat {\boldsymbol {\mu}}\right) \exp \left(- 2 ^ {2 \ell - 1} \operatorname {d i a g} (\hat {\boldsymbol {\Sigma}})\right) \\ \cos \left(2 ^ {\ell} \hat {\boldsymbol {\mu}}\right) \exp \left(- 2 ^ {2 \ell - 1} \operatorname {d i a g} (\hat {\boldsymbol {\Sigma}})\right) \end{array} \right] \right\} _ {\ell = 0} ^ {L - 1}, \tag {5} +$$ + +and the color and density of intervals can be obtained as + +$$ +\Psi_ {\mathrm {s}} \left(\hat {\gamma} (\hat {\boldsymbol {\mu}}, \hat {\boldsymbol {\Sigma}})\right) \longmapsto (\mathbf {c}, \sigma), \tag {6} +$$ + +where $\Psi_{\mathrm{s}}$ is the scene NeRF MLP [2]. + +# 3.2. Dynamic Human-Object Model + +Motivation. With estimated 3D human poses a priori, HumanNeRF [50] is effective at modeling moving people, but can only encode body parts, lacking the capability to model additional structures (e.g., objects held by the person) and hence not suitable for challenging scenarios containing complex human-object interactions. + +A naive approach would be attaching the objects directly to the interacting body parts (e.g., hands) so that the object transformations can be queried via the skeletal motion of the corresponding human joints. However, this approach cannot represent the relative transformations between the object and its interacting body part, making it difficult to model large objects, such as the suitcase shown in Fig. 6. + +Object Bones. To address this limitation, based on the conventional human skeleton hierarchy, we introduce a new "object bone" for each hand, where the starting point of the object bone is connected to the corresponding hand joint. As shown in Fig. 2 (a), two new object bones (green color) are connected to the hand joints, which represent the relative rotation and translation with respect to their parent hand joints. In potential, the number, bone size, and attaching joints of object bones can also be flexibly customized based on different human-object interaction scenarios. Therefore, our 3D human-object pose consists of 3D joints $\mathcal{I} = \{\mathbf{J}_{\mathrm{human}},\mathbf{J}_{\mathrm{object}}\}$ and their axis-angle rotations $\mathcal{R} = \{\omega_{\mathrm{human}},\omega_{\mathrm{object}}\}$ . + +Object LBS. To drive the articulated motions of the objects and human body, we learn a linear blend skinning + +![](images/08907949088bb6f39e1b288e7935b83a15880793ef1af925cc5e204f022e8e99.jpg) +Figure 2: Overview of our method. (1) Orange flowchart: The deformation from deformed points to canonical points are effectively estimated by the human-object backward LBS (a) and non-rigid deformation module, and their properties (i.e., density, color) can be obtained by querying the state-conditional canonical space. (2) Green flowchart: The background intervals sampled from a deformed frame are concatenated with object state embeddings for querying properties through the state-conditional scene model. (3) Novel views can be accordingly synthesized by volume rendering for re-ordered properties. + +(LBS) weights field during training, as shown in Fig. 2 (a), which encodes the influence regions of object bones and human bones. Following HumanNeRF [50], we pack the human and object bone blend weights and additional background weight into a single volume with $K + 1$ channels, i.e., $\mathcal{W}_{\mathrm{c} \to \mathrm{d}}(\mathbf{x}_{\mathrm{c}}) = \left\{w_{\mathrm{c} \to \mathrm{d}}^{i}(\mathbf{x}_{\mathrm{c}})\right\}$ , and generate the volume from a learnable latent code using a CNN. + +For instance, in Fig. 2, the subject holds a backpack in the right hand, so the learned LBS weight field only encodes the influence region of the right object bone with a similar shape to the backpack, while the left object bone has no influence on skeleton motion. Therefore, this learnable LBS weights field enables our object bones to be generalizable to various human-object interaction scenarios with left, right, or both hands. In addition, our human-object skeleton hierarchy enables HOSNeRF to support rendering with novel object pose and novel human pose, as illustrated in Fig. 1(c). + +Forward-Backward Deformation Framework. To improve the smoothness and consistency of human-object deformations, we propose to leverage the cycle consistency between forward and backward human-object deformations. The backward deformation first transforms sampled deformed points $\mathbf{x}_{\mathrm{d}}$ to their corresponding canonical points $\mathbf{x}_{\mathrm{c}}$ , which are further mapped back to the deformed space $\hat{\mathbf{x}}_{\mathrm{d}}$ through the forward deformation module. + +As illustrated in Fig. 2(a), with the proposed human-object skeleton definition and LBS weight fields, the coarse backward skeleton deformation $\Psi_{\mathrm{d}\to \mathrm{c}}^{\mathrm{coarse}}$ is defined similar to + +HumanNeRF [50], + +$$ +\Psi_ {\mathrm {d} \rightarrow \mathrm {c}} ^ {\text {c o a r s e}} \left(\mathbf {x} _ {\mathrm {d}}, \mathcal {J}, \mathcal {R}\right) = \sum_ {i = 1} ^ {K} w _ {\mathrm {d} \rightarrow \mathrm {c}} ^ {i} \left(\mathbf {x} _ {\mathrm {d}}\right)\left(\mathbf {R} _ {\mathrm {d} \rightarrow c} ^ {i} \mathbf {x} _ {\mathrm {d}} + \mathbf {t} _ {\mathrm {d} \rightarrow \mathrm {c}} ^ {i}\right), \tag {7} +$$ + +where $K$ is the number of human-object bones, $w_{\mathrm{d}\rightarrow \mathrm{c}}^i$ is the blend weight for the $i$ -th bone at the deformed space, and $\mathbf{R}_{\mathrm{d}\to \mathrm{c}}^i$ and $\mathbf{t}_{\mathrm{d}\to \mathrm{c}}^i$ are the backward rotation and translation that are explicitly derived from the human-object pose. + +To model the non-rigid motion that can not be represented by skeleton motion, we additionally define the backward fine non-rigid motion $\Psi_{\mathrm{d}\to c}^{\mathrm{fine}}$ conditioned on the human-object poses [50]: + +$$ +\Delta \mathbf {x} _ {\mathrm {d} \rightarrow \mathrm {c}} = \Psi_ {\mathrm {d} \rightarrow \mathrm {c}} ^ {\mathrm {f i n e}} \left(\gamma \left(\Psi_ {\mathrm {d} \rightarrow \mathrm {c}} ^ {\mathrm {c o a r s e}} \left(\mathbf {x} _ {\mathrm {d}}, \mathcal {I}, \mathcal {R}\right)\right), \mathcal {R}\right). \tag {8} +$$ + +Therefore, $\mathbf{x}_{\mathrm{d}}$ can be mapped to the canonical space by + +$$ +\mathbf {x} _ {\mathrm {c}} = \Psi_ {\mathrm {d} \rightarrow \mathrm {c}} ^ {\text {c o a r s e}} \left(\mathbf {x} _ {\mathrm {d}}, \mathcal {J}, \mathcal {R}\right) + \Delta \mathbf {x} _ {\mathrm {d} \rightarrow \mathrm {c}}. \tag {9} +$$ + +Accordingly, the forward coarse skeleton motion driven by the human-object pose can be defined as + +$$ +\Psi_ {\mathrm {c} \rightarrow \mathrm {d}} ^ {\text {c o a r s e}} \left(\mathbf {x} _ {\mathrm {c}}, \mathcal {J}, \mathcal {R}\right) = \sum_ {i = 1} ^ {K} w _ {\mathrm {c} \rightarrow \mathrm {d}} ^ {i} \left(\mathbf {x} _ {\mathrm {c}}\right)\left(\mathbf {R} _ {\mathrm {c} \rightarrow \mathrm {d}} ^ {i} \mathbf {x} _ {\mathrm {c}} + \mathbf {t} _ {\mathrm {c} \rightarrow \mathrm {d}} ^ {i}\right), \tag {10} +$$ + +where $w_{\mathrm{c}\rightarrow \mathrm{d}}^i$ is the canonical blend weight for the $i$ -th bone, $\mathbf{R}_{\mathrm{c}\rightarrow \mathrm{d}}^{i}$ and $\mathbf{t}_{\mathrm{c}\rightarrow \mathrm{d}}^{i}$ are the forward rotation and translation obtained from the human-object pose. In addition, the forward fine non-rigid motion $\Psi_{\mathrm{c}\rightarrow \mathrm{d}}^{\mathrm{fine}}$ is + +$$ +\Delta \mathbf {x} _ {\mathrm {c} \rightarrow \mathrm {d}} = \Psi_ {\mathrm {c} \rightarrow \mathrm {d}} ^ {\text {f i n e}} \left(\gamma \left(\Psi_ {\mathrm {c} \rightarrow \mathrm {d}} ^ {\text {c o a r s e}} \left(\mathbf {x} _ {\mathrm {c}}, \mathcal {J}, \mathcal {R}\right)\right), \mathcal {R}\right). \tag {11} +$$ + +![](images/6f2e8c93133308a3943824ed6f272f1e8d72ec5a4abf41c86d413815a2ede280.jpg) +$\gamma (\hat{\mu},\hat{\Sigma})$ IPE feature $\gamma (\mathbf{x}_c)$ Embed points $\mathbf{O}_s^i\mathbf{O}_c^i i$ th scene and canonical state embedding +T=6 The human picks up the bag + +![](images/2b731b6bab4f1f535632c61de0b01f77b89e3240674076968e95abc926348c75.jpg) +T=0 The human starts walking +Figure 3: State-conditional representation for the scene model (a) and human-object canonical space (b). + +Therefore, the canonical points $\mathbf{x}_{\mathrm{c}}$ can be transformed back to the deformed scene by + +$$ +\hat {\mathbf {x}} _ {\mathrm {d}} = \Psi_ {\mathrm {c} \rightarrow \mathrm {d}} ^ {\text {c o a r s e}} \left(\mathbf {x} _ {\mathrm {c}}, \mathcal {J}, \mathcal {R}\right) + \Delta \mathbf {x} _ {\mathrm {c} \rightarrow \mathrm {d}}. \tag {12} +$$ + +# 3.3. State-Conditional Representation + +Motivation. In complex dynamic scenes, humans can interact with different objects at different timesteps. As the example in Fig. 3 shows, the human picks up the bag at $T = 6$ and puts it down at $T = 24$ , resulting in 3 object states: the bag in the background during $T \in [0, 6]$ , the bag held by the human during $T \in [6, 24]$ , the bag at a new position in the background when $T > 24$ . Such object state changes prohibit us from directly using the static background model (i.e., Mip-NeRF 360 [2]) and the dynamic human-object model proposed in Sec. 3.2. + +A naive solution is to train a separate network for each object state. However, this solution is limited because each network is trained only on a short segment (e.g., $T \in [0, 6]$ ) instead of the whole video. As a result, some regions often are difficult to reconstruct because they are not observed in the short segment. But these regions could have already been captured by video segments of other object states. + +To make full use of all segments to jointly train one model shared for different states, we introduce learnable state embeddings to represent different object states. As depicted in Fig. 3, we have two state embeddings respectively for our scene model and human-object model, which are explained in detail as follows. + +State-Conditional Scene Model. As shown in Fig. 3(a), we develop a state-conditional Mip-NeRF 360 [2] to represent the static scene with temporal object transitions. In a dynamic scene with $N$ object states, we define $N$ learnable state embeddings $\mathcal{O}_{\mathrm{s}} = \{\mathbf{O}_{\mathrm{s}}^{i}$ ( $i = 1,2,\dots ,N$ ) to represent states information. Therefore, at state $i$ , we concatenate the IPE features $\hat{\gamma} (\hat{\mu},\hat{\Sigma})$ (Eq. (5)) of ray intervals with + +the state embedding $\mathbf{O}_{\mathrm{s}}^{i}$ as input to the scene MLP $\Psi_{\mathrm{s}}$ for querying the scene color and density. + +$$ +\Psi_ {\mathrm {s}} \left(\operatorname {c o n c a t} \left(\hat {\gamma} (\hat {\boldsymbol {\mu}}, \hat {\boldsymbol {\Sigma}}), \mathbf {O} _ {\mathrm {s}} ^ {i}\right)\right) \longmapsto (\mathbf {c}, \sigma). \tag {13} +$$ + +This representation enables a shared scene model across the video with multiple object state transitions. + +State-Conditional Dynamic Human-Object Model. In Sec. 3.2, we introduce the human-object pose-driven deformation module and the human-object canonical space for modeling dynamic human-objects. However, the reconstructed canonical space is insufficient in representing the temporal changes in object geometry and appearance when the subject interacts with new objects. To address this limitation, we condition the canonical space on the object states, as shown in Fig. 3(b). In particular, we employ an object state-conditional canonical space and optimize a shared blend weights field across all states. + +In a dynamic scene with $N$ object states, we define $N$ learnable state embeddings $\mathcal{O}_{\mathrm{c}} = \{\mathbf{O}_{\mathrm{c}}^{i}\} (i = 1,2,\dots ,N)$ to represent object states in the canonical space. Therefore, at object state $i$ , we concatenate the positionally encoded canonical points in Eq. (1) with the state embedding $\mathbf{O}_{\mathrm{c}}^{i}$ as input to the canonical MLP for querying the human-object color and density. + +$$ +\Psi_ {\mathrm {c}} \left(\operatorname {c o n c a t} \left(\gamma \left(\mathbf {x} _ {\mathrm {c}}\right), \mathbf {O} _ {\mathrm {c}} ^ {i}\right)\right) \longmapsto (\mathbf {c}, \sigma). \tag {14} +$$ + +# 3.4. Training + +Rendering. To render the pixel color, we shoot two rays $(\mathbf{r}_{\mathrm{s}},\mathbf{r}_{\mathrm{ho}})$ and sample intervals (i.e., ray points) to query the scene model (using $\mathbf{r}_{\mathrm{s}}$ ) and dynamic human-object model (with $\mathbf{r}_{\mathrm{ho}}$ ) respectively in their coordinate systems. After aligning the coordinates (see supplementary materials), we transform the sampled 3D points and their queried properties from the dynamic human-object space to the scene space, and re-order all sampled properties based on their distance from the camera center, as shown in Fig. 2. Therefore, the color of a pixel can be calculated through volume rendering, i.e., by integrating the re-ordered properties along the ray [27]: + +$$ +\hat {\mathbf {C}} (\mathbf {r}) = \sum_ {i = 1} ^ {N} T _ {i} \left(1 - e ^ {- \sigma_ {i} \delta_ {i}}\right) \mathbf {c} _ {i}, T _ {i} = e ^ {- \sum_ {j = 1} ^ {i - 1} \sigma_ {j} \delta_ {j}}. \tag {15} +$$ + +Directly fusing $\mathbf{r}_{\mathrm{s}}$ and $\mathbf{r}_{\mathrm{ho}}$ may yield minor artifacts on overlapped ground regions due to intersections of human-object points and scene points on short-length ground rays. To solve this issue, we estimate the foreground mask during training to separate the background rays from foreground ones. This allows us to focus on combining $\mathbf{r}_{\mathrm{s}}$ and $\mathbf{r}_{\mathrm{ho}}$ at the overlapped foreground regions, thereby reducing artifacts. + +Training Objectives. The training of HOSNeRF consists of three stages. In the first stage, we mask out the dynamic human-object regions and train the state-conditional + +
SceneNo. of objectsNo. of states
BACKPACK13
TENNIS23
SUITCASE14
PLAYGROUND35
DANCE37
LOUNGE35
+ +Table 1: Details of our collected dataset. + +scene model. Then, we train the state-conditional dynamic human-object model on the dynamic human-object regions. In the third stage, we combine these two models and further finetune the complete HOSNeRF for all image pixels. + +Given the single video with calibrated poses, the first stage of HOSNeRF is optimized by minimizing the photometric MSE loss and the regularization losses proposed by Mip-NeRF 360 [2] to avoid background collapse. In the second and third stages of HOSNeRF, we utilize the photometric MSE loss, patched-based perceptual LPIPS [57] loss, forward-backward deformation cycle consistency, and indirect optical flow supervisions. + +The deformation cycle consistency is enforced by minimizing the distance between estimated deformed points of Eq. (12) and sampled deformed points $\mathbf{x}_{\mathrm{d}}$ + +$$ +\mathcal {L} _ {\text {C y c l e}} = \frac {1}{2 N} \sum_ {i = 1} ^ {N} \left\| \mathbf {x} _ {\mathrm {d}} ^ {i} - \hat {\mathbf {x}} _ {\mathrm {d}} ^ {i} \right\| _ {2} ^ {2}. \tag {16} +$$ + +We also incorporate 2D optical flow as indirect supervision by minimizing the error between the induced flow and the estimated flow (details in the supplementary material). Therefore, the overall training objective of HOSNeRF is: + +$$ +\begin{array}{l} \mathcal {L} = \omega_ {\mathrm {M S E}} \cdot \mathcal {L} _ {\mathrm {M S E}} + \omega_ {\mathrm {L P I P S}} \mathcal {L} _ {\mathrm {L P I P S}} \\ + \omega_ {\text {C y c l e}} \cdot \mathcal {L} _ {\text {C y c l e}} + \omega_ {\text {F l o w}} \cdot \mathcal {L} _ {\text {F l o w}}, \tag {17} \\ \end{array} +$$ + +where $\omega_{\mathrm{MSE}}$ , $\omega_{\mathrm{LPIPS}}$ , $\omega_{\mathrm{Cycle}}$ , $\omega_{\mathrm{Flow}}$ are loss weights. + +# 4. Experiments + +# 4.1. Dataset + +To benchmark the reconstruction of monocular videos with dynamic human-object-scenes, we collect a new dataset with 6 scenes $^{\dagger}$ . The collected dataset features various types of human-object-scene interactions in indoor and outdoor scenarios, with up to 3 interacted objects and 7 states for a single video, as reported in Tab. 1. The duration of collected videos varies from 60s to 120s, and we extract [300, 400] frames for each video, where 16 frames at equal intervals are selected as novel views at novel timesteps as the test set and the remaining frames are the train set. + +We extensively evaluate the HOSNeRF on our collected dataset and NeuMan dataset [14] that consists of 6 short human walking sequences. Since NeuMan dataset does not involve human-object interactions, HOSNeRF can be flexibly customized to HOSNeRF by removing object bones and setting the object state to 1 when evaluated on NeuMan dataset. + +# 4.2. Implementation Details + +Human interacts with various objects using hands in our dataset, we therefore introduce 2 new object bones to the left and right hand joints of the 3D human skeleton for all scenes. As a result, there are in total $K = 26$ bones in our human-object skeleton. In order to obtain the human-object pose prior, we first utilize a pretrained ROMP [45] model to estimate the 3D human pose, and use Mask-RCNN [11] to estimate human and objects masks. However, since it is non-trivial to estimate the object pose for in-the-wild videos, we initialize the left and right object bones as the standard extended bones with no relative rotations with respect to human hands, and initialize their length by referencing the length of the object relative to the arm. During training, we further refine the human-object poses $(\mathcal{J},\mathcal{R})$ with a MLP-based pose correction module $(\Psi_{\mathcal{J}},\Psi_{\mathcal{R}})$ [50]. The overall three-stage training of HOSNeRF takes about 5 days on 4 Tesla V100 GPUs. Please see the supplementary material for the comparison of training time with other approaches and discussions about estimating object masks. + +# 4.3. Experimental Results + +Baselines and Evaluation Metrics. To demonstrate the performance of HOSNeRF, we make comparisons with various types of SOTA approaches, including (1) NeuMan [14], a human pose-based method; (2) HyperNeRF [30] and (3) Nerfies [29], dynamic radiance fields for general scenes; (4) $\mathrm{D}^2\mathrm{NeRF}$ [52], a static-dynamic decomposition method; (5) K-Planes [7], an explicit radiance field method. For quantitative comparison, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [49], and Learned Perceptual Image Patch Similarity (LPIPS) [57] are employed as evaluation metrics. + +Evaluation on HOSNeRF Dataset. We report the quantitative comparisons (PSNR, SSIM, and LPIPS with VGG [38] backbone) of 6 scenes for all approaches in Tab. 2. For all SOTA approaches, we utilize their highest configurations for a fair comparison. As shown in Tab. 2, our HOSNeRF achieves the best performance in terms of all metrics, except the PSNR for the Lounge scene (a metric known to favor smooth/blurry results [57]). The improvement of HOSNeRF is particularly significant in terms of the LPIPS, with an average of nearly $40\%$ gain over SOTA approaches. Fig. 4 visualizes the qualitative comparison of HOSNeRF over SOTA approaches on novel views at novel timesteps, where HOSNeRF achieves substantially better visual qual + +![](images/4bbe1f825c6f364d90efe660887f5f4a7cffc60e6c5cf5d437bdebba9534c8f6.jpg) +Figure 4: Qualitative comparisons of HOSNeRF and SOTA approaches on HOSNeRF dataset. + +ity than other approaches for all scenes. HOSNeRF is able to produce high-fidelity details close to ground truths for all scene contents, i.e., dynamic human bodies, objects, and backgrounds. In contrast, existing approaches tend to synthesize much blurrier images with missing components. Please see the supplementary video for more results on $360^{\circ}$ free-viewpoint bullet-time videos rendered by HOSNeRF. + +Evaluation on NeuMan Dataset [14]. We report quantitative metrics (PSNR, SSIM, and LPIPS with AlexNet [16] backbone) over 6 scenes of the NeuMan dataset in Tab. 3, where the metrics for NSFF [20], HyperNeRF [30], and NeuMan [14] are borrowed from NeuMan [14]. As shown + +in Tab. 3, our HOSNeRF achieves the best performance in terms of all metrics except the PSNR for the Citron scene, and significantly improves SOTA approaches by a large margin of $50\%$ in terms of LPIPS. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes. Fig. 5 visualizes the qualitative comparisons between HOSNeRF and NeuMan [14] for novel view synthesis at novel timesteps, where HOSNeRF achieves better visual quality over NeuMan [14] for the rendered human bodies and backgrounds in all scenes due to the superiority of our human model and background model. Our HOSNeRF also achieves bet- + +
BACKPACKTENNISSUITCASEPLAYGROUNDDANCELOUNGE
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
K-Planes [7]19.050.3450.55719.310.6900.53618.640.5750.60217.920.4460.63518.170.3280.62324.210.8240.453
D2NeRF [52]20.520.3840.60823.970.7370.54020.990.5920.64521.230.5100.61619.920.4070.64727.130.8580.509
Nerfies [29]19.560.5150.55922.120.8180.44319.010.6430.55521.140.7260.53319.370.6740.52425.900.9140.342
HyperNeRF [30]19.620.3590.58721.260.7420.51019.410.6040.60721.670.5580.57819.300.4120.60127.250.9250.332
NeuMan [14]21.210.4790.47823.170.7680.44220.840.6110.55121.460.5460.55121.190.5290.49028.400.9170.341
Ours (base)21.510.7640.27124.020.9100.32621.100.8290.39522.200.7870.34821.840.7850.26626.540.9620.243
Ours w/ state22.380.7860.25223.980.9100.32321.430.8340.39022.520.7960.34122.430.7960.25827.510.9650.245
Ours w/ object21.560.7670.26923.980.9090.32721.080.8300.39622.140.7870.34721.880.7850.26926.800.9620.246
Ours w/ state, object22.330.7850.25324.080.9090.32021.510.8330.38622.550.7960.33822.380.7960.25927.520.9650.245
Ours w/ state, object, mask22.480.7900.24524.190.9110.32121.600.8340.38222.770.7990.33522.480.8020.25127.730.9680.227
Ours w/ state, object, mask, cycle22.610.7950.24124.190.9110.32121.710.8350.38222.680.8020.33622.540.8040.24927.880.9680.227
Ours (full)22.560.7920.24324.150.9110.32021.740.8360.38222.670.8010.33622.630.8040.24827.740.9680.227
+ +Table 2: Per-scene quantitative evaluation on the HOSNeRF dataset against baselines and ablations of our method. We color code each cell as best and second best. + +
SEATTLPARKINGBIKEJOGGINGLABCITRON
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NSFF [20]21.840.690.3721.980.690.4621.160.710.3620.630.530.4920.520.750.3912.330.490.65
HyperNeRF [30]16.430.430.4016.040.380.6217.640.420.4318.520.390.5216.750.510.2316.810.410.56
NeuMan [14]23.980.770.2625.430.790.3125.520.820.2322.680.670.3224.930.850.2124.710.800.26
Ours26.680.910.1427.200.930.1226.090.930.1025.040.890.1624.930.940.1024.440.900.18
+ +Table 3: Per-scene quantitative evaluation on the NeuMan dataset against baselines. We color code each cell as best + +ter foreground-background merging results, such as the feet regions of the Seattle scene and the Lab scene. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes. + +Per-Scene $360^{\circ}$ Free-Viewpoint Renderings from Our HOSNeRF. We have also included a supplementary video to showcase the per-scene $360^{\circ}$ free-viewpoint renderings from our HOSNeRF on all six scenes of our HOSNeRF dataset. The video highlights that our HOSNeRF is capable of generating high-fidelity details that accurately resemble all scene components, including dynamic human bodies, objects, and backgrounds. Notably, our HOSNeRF facilitates pausing the monocular video at any given point and rendering all scene details with high-fidelity from $360^{\circ}$ viewpoints. To the best of our knowledge, our HOSNeRF represents the first work to achieve $360^{\circ}$ free-viewpoint high-fidelity novel view synthesis for dynamic scenes featuring human-environment interactions from a single video. + +It should be emphasized that to enable $360^{\circ}$ free-viewpoint rendering, the input videos must contain $360^{\circ}$ scene information (an example capture process is shown in the supplementary video); otherwise, artifacts will appear in unobserved areas that were not even seen in the input videos. However, our HOSNeRF is not restricted by such capturing requirements, as demonstrated on the NeuMan dataset [14], which features small camera motions. + +# 4.4. Ablation Study + +We conduct ablation studies on our collected dataset to evaluate the effectiveness of each proposed component in HOSNeRF. We progressively ablate each component from + +state embeddings, object bones, foreground masks, optical flow, and cycle consistency supervision. As shown in Tab. 2, the performance of HOSNeRF progressively drops with the disabling of each component, with the full model nearly achieving the best performances, which demonstrates the effectiveness of our designs. + +We visualize representative scenes with large (Fig. 6) and small (Fig. 7) objects to further evaluate our proposed designs. As shown in Fig. 6(b) and 7(b), the base models are unable to represent the interacted objects for the human and the scene. In contrast, the proposed state embedding is capable of representing the state transitions of small objects (Fig. 7(c)), but it fails to represent the deformations of large objects (Fig. 6(c)). Together with our proposed object bones, the deformation of large objects can be effectively estimated through the human-object skeleton hierarchy, enabling large objects modeling in Fig. 6(d). In addition, the full model can further improve the synthesis quality of fine details such as correcting the wrong right arm by comparing Fig. 7(c) and (d). These comparisons further demonstrate the importance and effectiveness of the proposed designs. + +# 5. Conclusion + +We introduced a novel framework of HOSNeRF, the first work to achieve $360^{\circ}$ free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video. To tackle the challenges of such a task, we first introduced the new object bones to the human skeleton hierarchy to effectively estimate the fast human-object motion in our dynamic human-object model. Then, we proposed the state-conditional + +![](images/dd2404d411a9d05a50623dddb866afd3c10660eb94fc917d95f966b6eb6a27a4.jpg) +Figure 5: Qualitative comparisons of HOSNeRF and NeuMan [14] on the NeuMan dataset. + +![](images/49941d43ad00be81ad71221b6a9a7a68643732672f7c0b2a682c07294d9a5d05.jpg) +(a) Groundtruth + +![](images/dcc17051effb55641c0a1598c79b131552c8fa1e978d0065378bde7c13e991aa.jpg) +(b) Base Model + +![](images/3fc6d737b3242ad436b628837fe323593d34682726ed548541f8ab29a4787f1d.jpg) +(c) With State Embedding +Figure 6: Ablation of HOSNeRF for large objects. + +![](images/c7f464820ea3d94f0904d54b050a755f54b566f16d46d4001805bd3c3aab72e0.jpg) +(d) With State Embedding & Object Bones + +human-object and scene representations for handling significant object state changes. With our designed training strategies, HOSNeRF produced significant improvements over SOTA approaches, and enabled $360^{\circ}$ free-viewpoint rendering of any frame from a single video and supported rendering views with novel human and object poses. + +Limitations and Future Work. Our HOSNeRF focuses on dynamic human-objects modeling and lacks the capability to represent dynamic backgrounds. It remains challenging but is worthwhile researching models for reconstructing high-fidelity dynamic unbounded backgrounds. + +![](images/bd39841f155c75f89a4e85100277fe62106a7c4fed3066a935f06597dfebbe13.jpg) + +![](images/1137a1d4414030a78649c32f490cb841ebf302a3f74c3326a4660eedfb9d99de.jpg) + +![](images/9d915738526257303c24a166225afcc53827c504785696ea4d74f7588430b45d.jpg) +(a) Groundtruth +(c) With State Embedding +Figure 7: Ablation of HOSNeRF for small objects. + +![](images/723c40e71ce5e3d6a6784e7525591d920e81eb91f2b34a3cbdc5be2899cdcf8a.jpg) +(b) Base Model +(d) Full Model + +# 6. Acknowledgements + +This research is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, Mike Zheng Shou's Start-Up Grant from NUS, the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001. This research is also supported by National Research Foundation, Singapore and A\*STAR, under its RIE 2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (grant no. I2001E0059). Jia-Wei Liu is also supported by NUS IDS-ISEP scholarship. Thanks to Yufei Shi for helpful discussions. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. +[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. +[4] Mingfei Chen, Jianfeng Zhang, Xiangyu Xu, Lijuan Liu, Yu-jun Cai, Jiashi Feng, and Shuicheng Yan. Geometry-guided progressive nerf for generalizable and efficient neural human rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 222-239. Springer, 2022. +[5] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. +[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. +[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. +[8] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. +[9] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–12, 2022. +[10] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. +[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. +[12] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. +[13] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European + +Conference on Computer Vision, pages 362-379. Springer, 2016. +[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 402-418. Springer, 2022. +[15] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. +[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. +[17] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. +[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 419-436. Springer, 2022. +[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. +[20] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. +[21] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. arXiv preprint arXiv:2211.11082, 2022. +[22] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. +[23] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021. +[24] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. +[25] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural vol + +umes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. +[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. +[28] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015. +[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. +[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. +[31] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. +[32] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. +[33] Sergey Prokudin, Michael J Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1810-1819, 2021. +[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. +[35] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. +[36] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yan-Pei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15872-15882, 2022. +[37] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthe + +sis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. +[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[39] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. +[40] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022. +[41] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum, volume 27, pages 1781-1787. Wiley Online Library, 2008. +[42] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021. +[43] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[44] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. +[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021. +[46] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6226-6237, 2021. +[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. +[48] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. + +[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. +[50] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022. +[51] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020. +[52] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D'nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems, 35:32653-32666, 2022. +[53] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. +[54] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34:14955–14966, 2021. +[55] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. +[56] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. arXiv preprint arXiv:2212.07626, 2022. +[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. +[58] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004. +[59] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014. \ No newline at end of file diff --git a/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/images.zip b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9884dd8ec8b1903de647923f8ba37056ccb3d5a9 --- /dev/null +++ b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aad4bee453d3878b4175e3817b65c2b8c8af9cc469bbb95d24eeb9280bc21f36 +size 1273383 diff --git a/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/layout.json b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a76f498a4422badb159247aa7693a3b90f1b415f --- /dev/null +++ b/2023/HOSNeRF_ Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video/layout.json @@ -0,0 +1,10288 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 527, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 527, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 527, + 138 + ], + "type": "text", + "content": "HOSNeRF: Dynamic Human-Object-Scene Neural Radiance Fields from a Single Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": "Jia-Wei Liu" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": ", Yan-Pei Cao" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": ", Tianyuan Yang" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": ", Zhongcong Xu" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": ", Jussi Keppo" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": ", Ying Shan" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": ", Xiaohu Qie" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "text", + "content": ", Mike Zheng Shou" + }, + { + "bbox": [ + 110, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 202, + 550, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 550, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 550, + 232 + ], + "type": "text", + "content": "1 Show Lab, National University of Singapore 2 ARC Lab, 3 Tencent PCG 4 Business School, 5 Institute of Operations Research and Analytics, National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "text", + "content": "We introduce HOSNeRF, a novel " + }, + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "text", + "content": " free-viewpoint rendering method that reconstructs neural radiance fields for dynamic human-object-scene from a single monocular in-the-wild video. Our method enables pausing the video at any frame and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints. The first challenge in this task is the complex object motions in human-object interactions, which we tackle by introducing the new object bones into the conventional human skeleton hierarchy to effectively estimate large object deformations in our dynamic human-object model. The second challenge is that humans interact with different objects at different times, for which we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively. Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by a large margin of " + }, + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "inline_equation", + "content": "40\\% \\sim 50\\%" + }, + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "text", + "content": " in terms of LPIPS. The code, data, and compelling examples of " + }, + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 283, + 290, + 536 + ], + "type": "text", + "content": " free-viewpoint renderings from single videos: https://showlab.github.io/HOSNeRF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 558, + 128, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 128, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 128, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 578, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 673 + ], + "type": "text", + "content": "Video reconstruction and free-viewpoint rendering offer innovative opportunities for creating immersive experiences, encompassing virtual reality, telepresence, metaverse, and 3D animation production. While reconstructing videos has the potential to enhance user engagement and provide more realistic environments, it also poses significant challenges in terms of monocular viewpoints [55, 48] and complicated human-environment interactions [50, 14]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 673, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 673, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 673, + 286, + 685 + ], + "type": "text", + "content": "In recent years, remarkable progress has been made in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 259, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 547, + 462 + ], + "type": "text", + "content": "novel view synthesis, particularly since the introduction of Neural Radiance Fields (NeRF) [27]. While initially limited to reconstructing static 3D scenes based on multi-view images, subsequent studies have proposed various approaches to address the challenge of dynamic view synthesis [34, 29, 30, 47, 20, 8]. NeRF-based techniques have evolved to either incorporate deformation fields that map dynamic fields to canonical NeRF spaces [34, 29, 30, 47], or model dynamic scenes as 4D spatio-temporal radiance fields [20, 8]. While these approaches have shown promising results in dynamic view synthesis, they are limited to simple deformations. Another line of research is specifically designed for dynamic neural human modeling that relies on estimated human poses as a priori information [32, 50]. Recently, Neuman [14] combines pose-driven dynamic human models with static scene models for representing dynamic human-centric scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 552 + ], + "type": "text", + "content": "However, none of the aforementioned techniques can accurately reconstruct challenging monocular videos with fast and complex human-object-scene motions and interactions, as shown in Fig. 1(d). This is due to two particular challenges listed below. To tackle them, we introduce a novel method called Human-Object-Scene Neural Radiance Fields (HOSNeRF)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "i) Complex object motions in human-object interactions. In contrast to the simple motions that can be modeled by general deformation modules [29, 30], the object motion during human-object interaction is often drastic and composed of various atomic motions (e.g., play tennis). To tackle this challenge, we propose new object bones that are attached to the human skeleton hierarchy to estimate human-object deformations in a coarse-to-fine manner for our dynamic human-object model. The object bones and underlying object linear blend skinning (object LBS) allow for the accurate estimation of objects' deformations through the relative transformations in the kinematic tree of the skeleton hierarchy." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 282, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 282, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 282, + 703 + ], + "type": "text", + "content": "*Work is partially done during internship at ARC Lab, Tencent PCG." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 703, + 141, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 141, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 141, + 712 + ], + "type": "text", + "content": "† Corresponding Authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18483" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 166, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 166, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 166, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 166, + 240 + ], + "type": "image", + "image_path": "6aa2f4bbcd3d7a80f3b0f41d901ff730a908862146905e5609d10befd3008413.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "text", + "content": "Figure 1: Our HOSNeRF (b) takes a single monocular in-the-wild video (a) as input, and creates high-fidelity " + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "text", + "content": " free-viewpoint rendering of all scene details (dynamic human body, objects, and background) at any time (d). Our method enables rendering views with novel object poses and novel human poses as shown in (c), and produces high-fidelity dynamic novel view synthesis results at novel timesteps, with significant improvements over SOTA approaches as shown in (d)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 169, + 70, + 286, + 239 + ], + "blocks": [ + { + "bbox": [ + 169, + 70, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 169, + 70, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 169, + 70, + 286, + 239 + ], + "type": "image", + "image_path": "acdab11ba8a5c4aaf70fb8c93b8d7ba0981573fb2cab1e96c98e8df002e4f9dd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 288, + 70, + 442, + 239 + ], + "blocks": [ + { + "bbox": [ + 288, + 70, + 442, + 239 + ], + "lines": [ + { + "bbox": [ + 288, + 70, + 442, + 239 + ], + "spans": [ + { + "bbox": [ + 288, + 70, + 442, + 239 + ], + "type": "image", + "image_path": "51c94a859fa255eaebf95973920c96b6d1fb8c11ec635692a5103e3844e88553.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 544, + 239 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 544, + 239 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 544, + 239 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 544, + 239 + ], + "type": "image", + "image_path": "692d6f2eceb289e3397e21e4a40f1ca1da20d3c5c6eaf476d193942f7e69f6dc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 298, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 418 + ], + "type": "text", + "content": "ii) Humans interact with different objects at different times. The above human-object model is designed for humans interacting with the same object over time. But when the person puts down the current object or picks a new object, it is not clear how to dynamically remove/add such objects in the static background model and the human-object model whose canonical space is static. To handle this challenge, we introduce two new learnable object state embeddings that can be used as conditions for learning our human-object representation and scene representation, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 419, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 540 + ], + "type": "text", + "content": "Finally, we systematically explore and identify effective training objectives and strategies for our proposed HOSNeRF, including deformation cycle consistency, optical flow supervisions, and foreground-background rendering. On two challenging datasets collected by ourselves and NeuMan [14], our HOSNeRF achieves high-fidelity dynamic novel view synthesis results and enables pausing the monocular video at any time and rendering all scene details (dynamic humans, objects, and backgrounds) from arbitrary viewpoints, as shown in Fig. 1(d)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 541, + 223, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 541, + 223, + 553 + ], + "spans": [ + { + "bbox": [ + 59, + 541, + 223, + 553 + ], + "type": "text", + "content": "In summary, our main contributions are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 564, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 58, + 564, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 564, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 58, + 564, + 287, + 613 + ], + "type": "text", + "content": "- We present a novel framework of HOSNeRF, the first work to achieve " + }, + { + "bbox": [ + 58, + 564, + 287, + 613 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 58, + 564, + 287, + 613 + ], + "type": "text", + "content": " free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 627, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 627, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 58, + 627, + 287, + 662 + ], + "type": "text", + "content": "- We propose the object bones and state-conditional representations to handle the non-rigid motions and interactions of humans, objects, and the environment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "content": "- Extensive experiments show that HOSNeRF significantly outperforms SOTA approaches on two challenging datasets by " + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "40\\% \\sim 50\\%" + }, + { + "bbox": [ + 58, + 677, + 287, + 713 + ], + "type": "text", + "content": " in terms of LPIPS." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 297, + 392, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 392, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 392, + 309 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 319, + 456, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 456, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 456, + 332 + ], + "type": "text", + "content": "2.1. Dynamic Human Modeling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 604 + ], + "type": "text", + "content": "Dynamic human modeling has shown promising results in utilizing various representations such as point clouds [1, 51], meshes [10, 24], voxels [25, 39], and neural implicit functions [12, 35], with models like SMPL [26, 33] being commonly used for parameterizing the human body. Since the introduction of NeRF [27], neural human representation [32, 36, 42, 54, 23] has achieved remarkable progress on representing dynamic human bodies from sparse-view videos. Among them, Neural Actor [23] and Neural Body [32] pioneer in combining NeRF [27] with SMPL deformable meshes to represent human bodies with complex motions. Subsequent works have further improved on the generalizability [17, 4, 9] and animatability [31, 18] of human bodies. To support multi-person modeling, recent works [55, 37] have proposed to segment each human into 3D bounding boxes and learn a separate layered dynamic NeRF for each person. Other works [46, 43, 15, 44] are specifically designed to reconstruct the dynamic human and object with RGB-D or multi-view videos as inputs. They track the human and object pose, and separately reconstruct them with volumetric fusion [43], neural texture blending [44], or neural rendering [15, 56]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Despite achieving promising results, these approaches require multi-view videos or RGB-D as input, limiting their real-world applications. To solve this problem, HumanNeRF [50] is proposed to represent moving humans from a monocular video by the human pose-driven deformation module and canonical space. NeuMan [14] is the first successful attempt at reconstructing both the dynamic human and static background from a single video. However, Neuman [14] does not support human-environment interactions" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18484" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 232, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 232, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 232, + 83 + ], + "type": "text", + "content": "and performs poorly at large camera motions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 275, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 275, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 275, + 103 + ], + "type": "text", + "content": "2.2. Dynamic View Synthesis for General Scenes" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 110, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 289, + 384 + ], + "type": "text", + "content": "Most prior approaches on dynamic scene modeling require synchronized multi-view videos [58, 19, 41, 5, 55, 48] or depth [59, 13, 28] as input. Recent studies have built upon NeRF [27] to reconstruct dynamic neural radiance fields from monocular videos by either learning a deformation field that maps dynamic observation to canonical field [34, 29, 30, 47] or building 4D spatio-temporal radiance fields [53, 20, 8]. Among them, Nerfies [29] associates latent codes with the deformation field and HyperNeRF [30] represents motion in a high-dimensional space. " + }, + { + "bbox": [ + 46, + 110, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 46, + 110, + 289, + 384 + ], + "type": "text", + "content": " [52] builds upon HyperNeRF [30] and further decouples the dynamic components from the static background, and represents them separately with a HyperNeRF [30] and NeRF [27]. DynIBaR [21] proposes a motion-adjusted multi-view feature aggregation module to synthesize new viewpoints by aggregating features from nearby views. Other studies have introduced voxel grids [6, 22, 40] or planar representations [7, 3] for fast dynamic radiance fields reconstruction. While these approaches have achieved high-fidelity dynamic view synthesis results, they are restricted to simple scene deformations. In contrast, our HOSNeRF is capable of representing significant human-object motions and interactions in complex environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 394, + 103, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 103, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 103, + 406 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 413, + 133, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 133, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 133, + 425 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": "HumanNeRF [50] has been recently introduced to represent a moving person with a NeRF [27]-based canonical space " + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": " that maps 3D points to color " + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": ", and a human pose-guided deformation field " + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": " that transforms deformed points " + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{d}}^{i}" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": " from the deformed space to canonical points " + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{c}}^{i}" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": " in the canonical space (" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 432, + 287, + 506 + ], + "type": "text", + "content": " omitted for simplicity)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\Psi_ {\\mathrm {c}} (\\gamma (\\mathbf {x} _ {\\mathrm {c}})) \\longmapsto (\\mathbf {c}, \\sigma), \\Psi_ {\\mathrm {d}} (\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}) \\longmapsto (\\mathbf {x} _ {\\mathrm {c}}), \\tag {1}", + "image_path": "4de3ba738a37ac6f7f2abf20e6bfaa76bc23de99a0b2403a1f57bb5a828008a3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "text", + "content": " is the standard positional encoding function, and " + }, + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\mathbf{J}_i\\}" + }, + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{\\omega_i\\}" + }, + { + "bbox": [ + 46, + 530, + 287, + 567 + ], + "type": "text", + "content": " represent the 3D human joints and the local joint axis-angle rotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": "The deformation field " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": " is decomposed into the coarse human skeleton-driven deformation " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{d}}^{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": ", and the fine non-rigid deformation conditioned on human poses " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{d}}^{\\mathrm{fine}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 620, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 620, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 52, + 620, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {c}} ^ {\\prime} = \\Psi_ {\\mathrm {d}} ^ {\\mathrm {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right), \\mathbf {x} _ {\\mathrm {c}} = \\mathbf {x} _ {\\mathrm {c}} ^ {\\prime} + \\Psi_ {\\mathrm {d}} ^ {\\mathrm {f i n e}} \\left(\\mathbf {x} _ {\\mathrm {c}} ^ {\\prime}, \\mathcal {R}\\right). (2)", + "image_path": "2eb387860343c685d416bb3a3fea48005a7c7a446988fe3f28572f450df6245a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Mip-NeRF 360 [2] is designed to synthesize realistic views for highly intricate, unbounded real-world static scenes. To render pixel colors, the casted rays are split into a set of intervals " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "T_{i} = [t_{i}, t_{i + 1})" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ". The mean and covariance of the conical frustum corresponding to every interval are computed as " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{\\mu}, \\boldsymbol{\\Sigma}) = \\mathbf{r}(T_{i})" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ". To parameterize the Gaussian" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "parameters for unbounded scenes, Mip-NeRF 360 [2] further proposes a contraction function " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": " that distributes distant points proportionally to disparity," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 117, + 545, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 117, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 335, + 117, + 545, + 149 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = \\left\\{ \\begin{array}{c l} \\mathbf {x} & \\| \\mathbf {x} \\| \\leq 1 \\\\ \\left(2 - \\frac {1}{\\| \\mathbf {x} \\|}\\right) \\left(\\frac {\\mathbf {x}}{\\| \\mathbf {x} \\|}\\right) & \\| \\mathbf {x} \\| > 1 \\end{array} . \\right. \\tag {3}", + "image_path": "b1e9c2752a8623beb5b027f65ea3336806ed58999b876b92a19e3c50b8defb42.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 156, + 484, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 484, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 484, + 169 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 306, + 156, + 484, + 169 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 306, + 156, + 484, + 169 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 306, + 156, + 484, + 169 + ], + "type": "inline_equation", + "content": "(\\pmb {\\mu},\\pmb {\\Sigma})" + }, + { + "bbox": [ + 306, + 156, + 484, + 169 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 177, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 177, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 342, + 177, + 545, + 198 + ], + "type": "interline_equation", + "content": "\\left(\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}\\right) = \\left(f (\\boldsymbol {\\mu}), \\mathbf {J} _ {f} (\\boldsymbol {\\mu}) \\boldsymbol {\\Sigma} \\mathbf {J} _ {f} (\\boldsymbol {\\mu}) ^ {\\mathrm {T}}\\right), \\tag {4}", + "image_path": "8120c96df12510f1a7839e767dc771027002e4e40bb4756f11132c1773b45049.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_f(\\boldsymbol{\\mu})" + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "text", + "content": " is the Jacobian of " + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}" + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "text", + "content": ". The contracted Gaussian parameters " + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "(\\hat{\\boldsymbol{\\mu}},\\hat{\\boldsymbol{\\Sigma}})" + }, + { + "bbox": [ + 304, + 205, + 545, + 243 + ], + "type": "text", + "content": " are further encoded through the integrated positional encoding (IPE) [2]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 251, + 545, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 545, + 286 + ], + "type": "interline_equation", + "content": "\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}) = \\left\\{\\left[ \\begin{array}{l} \\sin \\left(2 ^ {\\ell} \\hat {\\boldsymbol {\\mu}}\\right) \\exp \\left(- 2 ^ {2 \\ell - 1} \\operatorname {d i a g} (\\hat {\\boldsymbol {\\Sigma}})\\right) \\\\ \\cos \\left(2 ^ {\\ell} \\hat {\\boldsymbol {\\mu}}\\right) \\exp \\left(- 2 ^ {2 \\ell - 1} \\operatorname {d i a g} (\\hat {\\boldsymbol {\\Sigma}})\\right) \\end{array} \\right] \\right\\} _ {\\ell = 0} ^ {L - 1}, \\tag {5}", + "image_path": "331fe0951ed79f489f9fc16af4edd2b40a4f5b9228c43da01dcf05f268c80601.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 293, + 533, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 533, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 533, + 304 + ], + "type": "text", + "content": "and the color and density of intervals can be obtained as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 313, + 545, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 313, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 365, + 313, + 545, + 334 + ], + "type": "interline_equation", + "content": "\\Psi_ {\\mathrm {s}} \\left(\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}})\\right) \\longmapsto (\\mathbf {c}, \\sigma), \\tag {6}", + "image_path": "62530251ed31aa508dba044ba437548fdaf60194271b3f06537b29dcacfccf28.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 341, + 461, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 461, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 461, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 341, + 461, + 353 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{s}}" + }, + { + "bbox": [ + 306, + 341, + 461, + 353 + ], + "type": "text", + "content": " is the scene NeRF MLP [2]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 360, + 476, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 476, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 476, + 373 + ], + "type": "text", + "content": "3.2. Dynamic Human-Object Model" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 378, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 449 + ], + "type": "text", + "content": "Motivation. With estimated 3D human poses a priori, HumanNeRF [50] is effective at modeling moving people, but can only encode body parts, lacking the capability to model additional structures (e.g., objects held by the person) and hence not suitable for challenging scenarios containing complex human-object interactions." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 450, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 534 + ], + "type": "text", + "content": "A naive approach would be attaching the objects directly to the interacting body parts (e.g., hands) so that the object transformations can be queried via the skeletal motion of the corresponding human joints. However, this approach cannot represent the relative transformations between the object and its interacting body part, making it difficult to model large objects, such as the suitcase shown in Fig. 6." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "content": "Object Bones. To address this limitation, based on the conventional human skeleton hierarchy, we introduce a new \"object bone\" for each hand, where the starting point of the object bone is connected to the corresponding hand joint. As shown in Fig. 2 (a), two new object bones (green color) are connected to the hand joints, which represent the relative rotation and translation with respect to their parent hand joints. In potential, the number, bone size, and attaching joints of object bones can also be flexibly customized based on different human-object interaction scenarios. Therefore, our 3D human-object pose consists of 3D joints " + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\mathbf{J}_{\\mathrm{human}},\\mathbf{J}_{\\mathrm{object}}\\}" + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "content": " and their axis-angle rotations " + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{\\omega_{\\mathrm{human}},\\omega_{\\mathrm{object}}\\}" + }, + { + "bbox": [ + 304, + 534, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Object LBS. To drive the articulated motions of the objects and human body, we learn a linear blend skinning" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18485" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 276 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 276 + ], + "type": "image", + "image_path": "08907949088bb6f39e1b288e7935b83a15880793ef1af925cc5e204f022e8e99.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 340 + ], + "type": "text", + "content": "Figure 2: Overview of our method. (1) Orange flowchart: The deformation from deformed points to canonical points are effectively estimated by the human-object backward LBS (a) and non-rigid deformation module, and their properties (i.e., density, color) can be obtained by querying the state-conditional canonical space. (2) Green flowchart: The background intervals sampled from a deformed frame are concatenated with object state embeddings for querying properties through the state-conditional scene model. (3) Novel views can be accordingly synthesized by volume rendering for re-ordered properties." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 346, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 288, + 430 + ], + "type": "text", + "content": "(LBS) weights field during training, as shown in Fig. 2 (a), which encodes the influence regions of object bones and human bones. Following HumanNeRF [50], we pack the human and object bone blend weights and additional background weight into a single volume with " + }, + { + "bbox": [ + 46, + 346, + 288, + 430 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 46, + 346, + 288, + 430 + ], + "type": "text", + "content": " channels, i.e., " + }, + { + "bbox": [ + 46, + 346, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathrm{c} \\to \\mathrm{d}}(\\mathbf{x}_{\\mathrm{c}}) = \\left\\{w_{\\mathrm{c} \\to \\mathrm{d}}^{i}(\\mathbf{x}_{\\mathrm{c}})\\right\\}" + }, + { + "bbox": [ + 46, + 346, + 288, + 430 + ], + "type": "text", + "content": ", and generate the volume from a learnable latent code using a CNN." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 440, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 561 + ], + "type": "text", + "content": "For instance, in Fig. 2, the subject holds a backpack in the right hand, so the learned LBS weight field only encodes the influence region of the right object bone with a similar shape to the backpack, while the left object bone has no influence on skeleton motion. Therefore, this learnable LBS weights field enables our object bones to be generalizable to various human-object interaction scenarios with left, right, or both hands. In addition, our human-object skeleton hierarchy enables HOSNeRF to support rendering with novel object pose and novel human pose, as illustrated in Fig. 1(c)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "content": "Forward-Backward Deformation Framework. To improve the smoothness and consistency of human-object deformations, we propose to leverage the cycle consistency between forward and backward human-object deformations. The backward deformation first transforms sampled deformed points " + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "content": " to their corresponding canonical points " + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "content": ", which are further mapped back to the deformed space " + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "content": " through the forward deformation module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 2(a), with the proposed human-object skeleton definition and LBS weight fields, the coarse backward skeleton deformation " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{d}\\to \\mathrm{c}}^{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " is defined similar to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 346, + 384, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 384, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 384, + 357 + ], + "type": "text", + "content": "HumanNeRF [50]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 364, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right) = \\sum_ {i = 1} ^ {K} w _ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {i} \\left(\\mathbf {x} _ {\\mathrm {d}}\\right)\\left(\\mathbf {R} _ {\\mathrm {d} \\rightarrow c} ^ {i} \\mathbf {x} _ {\\mathrm {d}} + \\mathbf {t} _ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {i}\\right), \\tag {7}", + "image_path": "24626e81a57b647bf99e52161c6616aa2acf36613008fe8aac031a916f97b86a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "text", + "content": " is the number of human-object bones, " + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{d}\\rightarrow \\mathrm{c}}^i" + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "text", + "content": " is the blend weight for the " + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "text", + "content": "-th bone at the deformed space, and " + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{d}\\to \\mathrm{c}}^i" + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{d}\\to \\mathrm{c}}^i" + }, + { + "bbox": [ + 305, + 407, + 545, + 454 + ], + "type": "text", + "content": " are the backward rotation and translation that are explicitly derived from the human-object pose." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 456, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 545, + 503 + ], + "type": "text", + "content": "To model the non-rigid motion that can not be represented by skeleton motion, we additionally define the backward fine non-rigid motion " + }, + { + "bbox": [ + 305, + 456, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{d}\\to c}^{\\mathrm{fine}}" + }, + { + "bbox": [ + 305, + 456, + 545, + 503 + ], + "type": "text", + "content": " conditioned on the human-object poses [50]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 510, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 510, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 321, + 510, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {x} _ {\\mathrm {d} \\rightarrow \\mathrm {c}} = \\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\mathrm {f i n e}} \\left(\\gamma \\left(\\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\mathrm {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {I}, \\mathcal {R}\\right)\\right), \\mathcal {R}\\right). \\tag {8}", + "image_path": "9b77f897754064ebec29df384b59847df457a7b11bb1d57208243e94a50fb245.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 531, + 529, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 529, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 529, + 544 + ], + "type": "text", + "content": "Therefore, " + }, + { + "bbox": [ + 306, + 531, + 529, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{d}}" + }, + { + "bbox": [ + 306, + 531, + 529, + 544 + ], + "type": "text", + "content": " can be mapped to the canonical space by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 551, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 551, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 348, + 551, + 545, + 564 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {c}} = \\Psi_ {\\mathrm {d} \\rightarrow \\mathrm {c}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {d}}, \\mathcal {J}, \\mathcal {R}\\right) + \\Delta \\mathbf {x} _ {\\mathrm {d} \\rightarrow \\mathrm {c}}. \\tag {9}", + "image_path": "12fa8ae537604289e7fe121e2e8a2a49a2dcdca94fd818fc45292bf2b0fd35ac.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 571, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 571, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 545, + 595 + ], + "type": "text", + "content": "Accordingly, the forward coarse skeleton motion driven by the human-object pose can be defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 602, + 545, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 602, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 602, + 545, + 644 + ], + "type": "interline_equation", + "content": "\\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right) = \\sum_ {i = 1} ^ {K} w _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i} \\left(\\mathbf {x} _ {\\mathrm {c}}\\right)\\left(\\mathbf {R} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i} \\mathbf {x} _ {\\mathrm {c}} + \\mathbf {t} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {i}\\right), \\tag {10}", + "image_path": "813bcb370c1fe94718d77fc1f2103745f92b63b32e26680daf4bca9ba11cf412.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{c}\\rightarrow \\mathrm{d}}^i" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": " is the canonical blend weight for the " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": "-th bone, " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{i}" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{i}" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": " are the forward rotation and translation obtained from the human-object pose. In addition, the forward fine non-rigid motion " + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{c}\\rightarrow \\mathrm{d}}^{\\mathrm{fine}}" + }, + { + "bbox": [ + 305, + 645, + 545, + 694 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {x} _ {\\mathrm {c} \\rightarrow \\mathrm {d}} = \\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {f i n e}} \\left(\\gamma \\left(\\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right)\\right), \\mathcal {R}\\right). \\tag {11}", + "image_path": "2582e4e08eba3585951bb41f21d3b2826c382905727f5c879026de814093c38f.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18486" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 82, + 286, + 151 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "type": "inline_equation", + "content": "\\gamma (\\hat{\\mu},\\hat{\\Sigma})" + }, + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "type": "text", + "content": " IPE feature " + }, + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x}_c)" + }, + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "type": "text", + "content": " Embed points " + }, + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_s^i\\mathbf{O}_c^i i" + }, + { + "bbox": [ + 50, + 71, + 286, + 80 + ], + "type": "text", + "content": " th scene and canonical state embedding" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 82, + 286, + 151 + ], + "lines": [ + { + "bbox": [ + 50, + 82, + 286, + 151 + ], + "spans": [ + { + "bbox": [ + 50, + 82, + 286, + 151 + ], + "type": "image", + "image_path": "6f2e8c93133308a3943824ed6f272f1e8d72ec5a4abf41c86d413815a2ede280.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 152, + 257, + 160 + ], + "lines": [ + { + "bbox": [ + 168, + 152, + 257, + 160 + ], + "spans": [ + { + "bbox": [ + 168, + 152, + 257, + 160 + ], + "type": "text", + "content": "T=6 The human picks up the bag" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 161, + 286, + 247 + ], + "blocks": [ + { + "bbox": [ + 52, + 152, + 134, + 160 + ], + "lines": [ + { + "bbox": [ + 52, + 152, + 134, + 160 + ], + "spans": [ + { + "bbox": [ + 52, + 152, + 134, + 160 + ], + "type": "text", + "content": "T=0 The human starts walking" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 161, + 286, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 161, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 161, + 286, + 247 + ], + "type": "image", + "image_path": "2b731b6bab4f1f535632c61de0b01f77b89e3240674076968e95abc926348c75.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 287, + 275 + ], + "type": "text", + "content": "Figure 3: State-conditional representation for the scene model (a) and human-object canonical space (b)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 281, + 287, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 306 + ], + "type": "text", + "content": "Therefore, the canonical points " + }, + { + "bbox": [ + 47, + 281, + 287, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{c}}" + }, + { + "bbox": [ + 47, + 281, + 287, + 306 + ], + "type": "text", + "content": " can be transformed back to the deformed scene by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 314, + 287, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 314, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 89, + 314, + 287, + 328 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {\\mathrm {d}} = \\Psi_ {\\mathrm {c} \\rightarrow \\mathrm {d}} ^ {\\text {c o a r s e}} \\left(\\mathbf {x} _ {\\mathrm {c}}, \\mathcal {J}, \\mathcal {R}\\right) + \\Delta \\mathbf {x} _ {\\mathrm {c} \\rightarrow \\mathrm {d}}. \\tag {12}", + "image_path": "d90748687b1380d7b4e4dfbb849530824e2fce74fb8ac2d11d54721bd6fcb985.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 336, + 227, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 227, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 227, + 348 + ], + "type": "text", + "content": "3.3. State-Conditional Representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": "Motivation. In complex dynamic scenes, humans can interact with different objects at different timesteps. As the example in Fig. 3 shows, the human picks up the bag at " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": " and puts it down at " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "T = 24" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": ", resulting in 3 object states: the bag in the background during " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "T \\in [0, 6]" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": ", the bag held by the human during " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "T \\in [6, 24]" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": ", the bag at a new position in the background when " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "T > 24" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": ". Such object state changes prohibit us from directly using the static background model (i.e., Mip-NeRF 360 [2]) and the dynamic human-object model proposed in Sec. 3.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 475, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 557 + ], + "type": "text", + "content": "A naive solution is to train a separate network for each object state. However, this solution is limited because each network is trained only on a short segment (e.g., " + }, + { + "bbox": [ + 46, + 475, + 287, + 557 + ], + "type": "inline_equation", + "content": "T \\in [0, 6]" + }, + { + "bbox": [ + 46, + 475, + 287, + 557 + ], + "type": "text", + "content": ") instead of the whole video. As a result, some regions often are difficult to reconstruct because they are not observed in the short segment. But these regions could have already been captured by video segments of other object states." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": "To make full use of all segments to jointly train one model shared for different states, we introduce learnable state embeddings to represent different object states. As depicted in Fig. 3, we have two state embeddings respectively for our scene model and human-object model, which are explained in detail as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "State-Conditional Scene Model. As shown in Fig. 3(a), we develop a state-conditional Mip-NeRF 360 [2] to represent the static scene with temporal object transitions. In a dynamic scene with " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " object states, we define " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " learnable state embeddings " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{s}} = \\{\\mathbf{O}_{\\mathrm{s}}^{i}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots ,N" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ") to represent states information. Therefore, at state " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", we concatenate the IPE features " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\gamma} (\\hat{\\mu},\\hat{\\Sigma})" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " (Eq. (5)) of ray intervals with" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "the state embedding " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{s}}^{i}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " as input to the scene MLP " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{s}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " for querying the scene color and density." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 101, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 101, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 326, + 101, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\Psi_ {\\mathrm {s}} \\left(\\operatorname {c o n c a t} \\left(\\hat {\\gamma} (\\hat {\\boldsymbol {\\mu}}, \\hat {\\boldsymbol {\\Sigma}}), \\mathbf {O} _ {\\mathrm {s}} ^ {i}\\right)\\right) \\longmapsto (\\mathbf {c}, \\sigma). \\tag {13}", + "image_path": "3f121f3265c5c13ae3553fa9a062d59980dc6c3323e714995879618e3e54379a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 125, + 545, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 125, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 545, + 149 + ], + "type": "text", + "content": "This representation enables a shared scene model across the video with multiple object state transitions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 149, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 149, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 545, + 281 + ], + "type": "text", + "content": "State-Conditional Dynamic Human-Object Model. In Sec. 3.2, we introduce the human-object pose-driven deformation module and the human-object canonical space for modeling dynamic human-objects. However, the reconstructed canonical space is insufficient in representing the temporal changes in object geometry and appearance when the subject interacts with new objects. To address this limitation, we condition the canonical space on the object states, as shown in Fig. 3(b). In particular, we employ an object state-conditional canonical space and optimize a shared blend weights field across all states." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": "In a dynamic scene with " + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": " object states, we define " + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": " learnable state embeddings " + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathrm{c}} = \\{\\mathbf{O}_{\\mathrm{c}}^{i}\\} (i = 1,2,\\dots ,N)" + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": " to represent object states in the canonical space. Therefore, at object state " + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": ", we concatenate the positionally encoded canonical points in Eq. (1) with the state embedding " + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{c}}^{i}" + }, + { + "bbox": [ + 304, + 281, + 545, + 365 + ], + "type": "text", + "content": " as input to the canonical MLP for querying the human-object color and density." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 344, + 370, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 370, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 344, + 370, + 545, + 384 + ], + "type": "interline_equation", + "content": "\\Psi_ {\\mathrm {c}} \\left(\\operatorname {c o n c a t} \\left(\\gamma \\left(\\mathbf {x} _ {\\mathrm {c}}\\right), \\mathbf {O} _ {\\mathrm {c}} ^ {i}\\right)\\right) \\longmapsto (\\mathbf {c}, \\sigma). \\tag {14}", + "image_path": "839d0c76d1c253aac589192c23761b9e5748414c82b70ebc3cdc9aa55b06fe40.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 389, + 370, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 370, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 370, + 401 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "text", + "content": "Rendering. To render the pixel color, we shoot two rays " + }, + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "inline_equation", + "content": "(\\mathbf{r}_{\\mathrm{s}},\\mathbf{r}_{\\mathrm{ho}})" + }, + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "text", + "content": " and sample intervals (i.e., ray points) to query the scene model (using " + }, + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "text", + "content": ") and dynamic human-object model (with " + }, + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{ho}}" + }, + { + "bbox": [ + 304, + 407, + 545, + 552 + ], + "type": "text", + "content": ") respectively in their coordinate systems. After aligning the coordinates (see supplementary materials), we transform the sampled 3D points and their queried properties from the dynamic human-object space to the scene space, and re-order all sampled properties based on their distance from the camera center, as shown in Fig. 2. Therefore, the color of a pixel can be calculated through volume rendering, i.e., by integrating the re-ordered properties along the ray [27]:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 309, + 556, + 545, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 309, + 556, + 545, + 589 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {C}} (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} T _ {i} \\left(1 - e ^ {- \\sigma_ {i} \\delta_ {i}}\\right) \\mathbf {c} _ {i}, T _ {i} = e ^ {- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}}. \\tag {15}", + "image_path": "8454e7e5ba1e13884e21d0922824aee0b60030f1ed1ea818d66ce34792f46aa6.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": "Directly fusing " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{ho}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " may yield minor artifacts on overlapped ground regions due to intersections of human-object points and scene points on short-length ground rays. To solve this issue, we estimate the foreground mask during training to separate the background rays from foreground ones. This allows us to focus on combining " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{ho}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " at the overlapped foreground regions, thereby reducing artifacts." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Training Objectives. The training of HOSNeRF consists of three stages. In the first stage, we mask out the dynamic human-object regions and train the state-conditional" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18487" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 267, + 157 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 267, + 157 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 267, + 157 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 267, + 157 + ], + "type": "table", + "html": "
SceneNo. of objectsNo. of states
BACKPACK13
TENNIS23
SUITCASE14
PLAYGROUND35
DANCE37
LOUNGE35
", + "image_path": "1e3a0b3a9617cf4f9e7fe399b0d7415d126e3b4336dff3111019508f0f6e1100.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 159, + 247, + 171 + ], + "lines": [ + { + "bbox": [ + 83, + 159, + 247, + 171 + ], + "spans": [ + { + "bbox": [ + 83, + 159, + 247, + 171 + ], + "type": "text", + "content": "Table 1: Details of our collected dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 179, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 286, + 227 + ], + "type": "text", + "content": "scene model. Then, we train the state-conditional dynamic human-object model on the dynamic human-object regions. In the third stage, we combine these two models and further finetune the complete HOSNeRF for all image pixels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 227, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 286, + 323 + ], + "type": "text", + "content": "Given the single video with calibrated poses, the first stage of HOSNeRF is optimized by minimizing the photometric MSE loss and the regularization losses proposed by Mip-NeRF 360 [2] to avoid background collapse. In the second and third stages of HOSNeRF, we utilize the photometric MSE loss, patched-based perceptual LPIPS [57] loss, forward-backward deformation cycle consistency, and indirect optical flow supervisions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "text", + "content": "The deformation cycle consistency is enforced by minimizing the distance between estimated deformed points of Eq. (12) and sampled deformed points " + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{d}}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 369, + 287, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 369, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 100, + 369, + 287, + 400 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {C y c l e}} = \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left\\| \\mathbf {x} _ {\\mathrm {d}} ^ {i} - \\hat {\\mathbf {x}} _ {\\mathrm {d}} ^ {i} \\right\\| _ {2} ^ {2}. \\tag {16}", + "image_path": "a17f74848fab974f2dda39b58838a1328f83aa23e308b7ee2591cab42ba6b084.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 409, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 458 + ], + "type": "text", + "content": "We also incorporate 2D optical flow as indirect supervision by minimizing the error between the induced flow and the estimated flow (details in the supplementary material). Therefore, the overall training objective of HOSNeRF is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 467, + 287, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 467, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 82, + 467, + 287, + 495 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\omega_ {\\mathrm {M S E}} \\cdot \\mathcal {L} _ {\\mathrm {M S E}} + \\omega_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} \\\\ + \\omega_ {\\text {C y c l e}} \\cdot \\mathcal {L} _ {\\text {C y c l e}} + \\omega_ {\\text {F l o w}} \\cdot \\mathcal {L} _ {\\text {F l o w}}, \\tag {17} \\\\ \\end{array}", + "image_path": "39c00674dfa9484680175660846714603e0615cb244b4524e7d28f35a71b067a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{Cycle}}" + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{Flow}}" + }, + { + "bbox": [ + 47, + 505, + 266, + 517 + ], + "type": "text", + "content": " are loss weights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 528, + 127, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 127, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 127, + 541 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 548, + 105, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 105, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 105, + 559 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": "To benchmark the reconstruction of monocular videos with dynamic human-object-scenes, we collect a new dataset with 6 scenes" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": ". The collected dataset features various types of human-object-scene interactions in indoor and outdoor scenarios, with up to 3 interacted objects and 7 states for a single video, as reported in Tab. 1. The duration of collected videos varies from 60s to 120s, and we extract [300, 400] frames for each video, where 16 frames at equal intervals are selected as novel views at novel timesteps as the test set and the remaining frames are the train set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "We extensively evaluate the HOSNeRF on our collected dataset and NeuMan dataset [14] that consists of 6 short human walking sequences. Since NeuMan dataset does not involve human-object interactions, HOSNeRF can be flexibly customized to HOSNeRF by removing object bones and setting the object state to 1 when evaluated on NeuMan dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 152, + 438, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 438, + 165 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 438, + 165 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "text", + "content": "Human interacts with various objects using hands in our dataset, we therefore introduce 2 new object bones to the left and right hand joints of the 3D human skeleton for all scenes. As a result, there are in total " + }, + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "inline_equation", + "content": "K = 26" + }, + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "text", + "content": " bones in our human-object skeleton. In order to obtain the human-object pose prior, we first utilize a pretrained ROMP [45] model to estimate the 3D human pose, and use Mask-RCNN [11] to estimate human and objects masks. However, since it is non-trivial to estimate the object pose for in-the-wild videos, we initialize the left and right object bones as the standard extended bones with no relative rotations with respect to human hands, and initialize their length by referencing the length of the object relative to the arm. During training, we further refine the human-object poses " + }, + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "inline_equation", + "content": "(\\mathcal{J},\\mathcal{R})" + }, + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "text", + "content": " with a MLP-based pose correction module " + }, + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "inline_equation", + "content": "(\\Psi_{\\mathcal{J}},\\Psi_{\\mathcal{R}})" + }, + { + "bbox": [ + 304, + 171, + 545, + 399 + ], + "type": "text", + "content": " [50]. The overall three-stage training of HOSNeRF takes about 5 days on 4 Tesla V100 GPUs. Please see the supplementary material for the comparison of training time with other approaches and discussions about estimating object masks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 407, + 429, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 429, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 429, + 419 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics. To demonstrate the performance of HOSNeRF, we make comparisons with various types of SOTA approaches, including (1) NeuMan [14], a human pose-based method; (2) HyperNeRF [30] and (3) Nerfies [29], dynamic radiance fields for general scenes; (4) " + }, + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "type": "text", + "content": " [52], a static-dynamic decomposition method; (5) K-Planes [7], an explicit radiance field method. For quantitative comparison, peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) [49], and Learned Perceptual Image Patch Similarity (LPIPS) [57] are employed as evaluation metrics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on HOSNeRF Dataset. We report the quantitative comparisons (PSNR, SSIM, and LPIPS with VGG [38] backbone) of 6 scenes for all approaches in Tab. 2. For all SOTA approaches, we utilize their highest configurations for a fair comparison. As shown in Tab. 2, our HOSNeRF achieves the best performance in terms of all metrics, except the PSNR for the Lounge scene (a metric known to favor smooth/blurry results [57]). The improvement of HOSNeRF is particularly significant in terms of the LPIPS, with an average of nearly " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " gain over SOTA approaches. Fig. 4 visualizes the qualitative comparison of HOSNeRF over SOTA approaches on novel views at novel timesteps, where HOSNeRF achieves substantially better visual qual" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "†Our data collection process has been approved by the institutional review board and participants have given consents for public data release." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18488" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 545, + 547 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 545, + 547 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 545, + 547 + ], + "type": "image", + "image_path": "4bbe1f825c6f364d90efe660887f5f4a7cffc60e6c5cf5d437bdebba9534c8f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 549, + 485, + 562 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 485, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 485, + 562 + ], + "type": "text", + "content": "Figure 4: Qualitative comparisons of HOSNeRF and SOTA approaches on HOSNeRF dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 568, + 288, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 652 + ], + "type": "text", + "content": "ity than other approaches for all scenes. HOSNeRF is able to produce high-fidelity details close to ground truths for all scene contents, i.e., dynamic human bodies, objects, and backgrounds. In contrast, existing approaches tend to synthesize much blurrier images with missing components. Please see the supplementary video for more results on " + }, + { + "bbox": [ + 46, + 568, + 288, + 652 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 568, + 288, + 652 + ], + "type": "text", + "content": " free-viewpoint bullet-time videos rendered by HOSNeRF." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Evaluation on NeuMan Dataset [14]. We report quantitative metrics (PSNR, SSIM, and LPIPS with AlexNet [16] backbone) over 6 scenes of the NeuMan dataset in Tab. 3, where the metrics for NSFF [20], HyperNeRF [30], and NeuMan [14] are borrowed from NeuMan [14]. As shown" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 568, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 712 + ], + "type": "text", + "content": "in Tab. 3, our HOSNeRF achieves the best performance in terms of all metrics except the PSNR for the Citron scene, and significantly improves SOTA approaches by a large margin of " + }, + { + "bbox": [ + 304, + 568, + 545, + 712 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 712 + ], + "type": "text", + "content": " in terms of LPIPS. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes. Fig. 5 visualizes the qualitative comparisons between HOSNeRF and NeuMan [14] for novel view synthesis at novel timesteps, where HOSNeRF achieves better visual quality over NeuMan [14] for the rendered human bodies and backgrounds in all scenes due to the superiority of our human model and background model. Our HOSNeRF also achieves bet-" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18489" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 545, + 203 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 203 + ], + "type": "table", + "html": "
BACKPACKTENNISSUITCASEPLAYGROUNDDANCELOUNGE
PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓PSNR†SSIM†LPIPS↓
K-Planes [7]19.050.3450.55719.310.6900.53618.640.5750.60217.920.4460.63518.170.3280.62324.210.8240.453
D2NeRF [52]20.520.3840.60823.970.7370.54020.990.5920.64521.230.5100.61619.920.4070.64727.130.8580.509
Nerfies [29]19.560.5150.55922.120.8180.44319.010.6430.55521.140.7260.53319.370.6740.52425.900.9140.342
HyperNeRF [30]19.620.3590.58721.260.7420.51019.410.6040.60721.670.5580.57819.300.4120.60127.250.9250.332
NeuMan [14]21.210.4790.47823.170.7680.44220.840.6110.55121.460.5460.55121.190.5290.49028.400.9170.341
Ours (base)21.510.7640.27124.020.9100.32621.100.8290.39522.200.7870.34821.840.7850.26626.540.9620.243
Ours w/ state22.380.7860.25223.980.9100.32321.430.8340.39022.520.7960.34122.430.7960.25827.510.9650.245
Ours w/ object21.560.7670.26923.980.9090.32721.080.8300.39622.140.7870.34721.880.7850.26926.800.9620.246
Ours w/ state, object22.330.7850.25324.080.9090.32021.510.8330.38622.550.7960.33822.380.7960.25927.520.9650.245
Ours w/ state, object, mask22.480.7900.24524.190.9110.32121.600.8340.38222.770.7990.33522.480.8020.25127.730.9680.227
Ours w/ state, object, mask, cycle22.610.7950.24124.190.9110.32121.710.8350.38222.680.8020.33622.540.8040.24927.880.9680.227
Ours (full)22.560.7920.24324.150.9110.32021.740.8360.38222.670.8010.33622.630.8040.24827.740.9680.227
", + "image_path": "98081892ffc021207d666b9d33d8b081fe8cd7f816f5edc0f6b316b6b2e49e30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 239, + 545, + 306 + ], + "blocks": [ + { + "bbox": [ + 47, + 206, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 206, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 546, + 232 + ], + "type": "text", + "content": "Table 2: Per-scene quantitative evaluation on the HOSNeRF dataset against baselines and ablations of our method. We color code each cell as best and second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 239, + 545, + 306 + ], + "lines": [ + { + "bbox": [ + 51, + 239, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 51, + 239, + 545, + 306 + ], + "type": "table", + "html": "
SEATTLPARKINGBIKEJOGGINGLABCITRON
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NSFF [20]21.840.690.3721.980.690.4621.160.710.3620.630.530.4920.520.750.3912.330.490.65
HyperNeRF [30]16.430.430.4016.040.380.6217.640.420.4318.520.390.5216.750.510.2316.810.410.56
NeuMan [14]23.980.770.2625.430.790.3125.520.820.2322.680.670.3224.930.850.2124.710.800.26
Ours26.680.910.1427.200.930.1226.090.930.1025.040.890.1624.930.940.1024.440.900.18
", + "image_path": "9c12bacac70a76c3dc884f9c016601dfeaac8f1dd1a67e94e9e7f1c087cd0ff8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 310, + 530, + 323 + ], + "lines": [ + { + "bbox": [ + 57, + 310, + 530, + 323 + ], + "spans": [ + { + "bbox": [ + 57, + 310, + 530, + 323 + ], + "type": "text", + "content": "Table 3: Per-scene quantitative evaluation on the NeuMan dataset against baselines. We color code each cell as best" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 333, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 380 + ], + "type": "text", + "content": "ter foreground-background merging results, such as the feet regions of the Seattle scene and the Lab scene. This further demonstrates the effectiveness and flexibility of HOSNeRF to model various types of dynamic human-(objects)-scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": "Per-Scene " + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": " Free-Viewpoint Renderings from Our HOSNeRF. We have also included a supplementary video to showcase the per-scene " + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": " free-viewpoint renderings from our HOSNeRF on all six scenes of our HOSNeRF dataset. The video highlights that our HOSNeRF is capable of generating high-fidelity details that accurately resemble all scene components, including dynamic human bodies, objects, and backgrounds. Notably, our HOSNeRF facilitates pausing the monocular video at any given point and rendering all scene details with high-fidelity from " + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": " viewpoints. To the best of our knowledge, our HOSNeRF represents the first work to achieve " + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": " free-viewpoint high-fidelity novel view synthesis for dynamic scenes featuring human-environment interactions from a single video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 551, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 647 + ], + "type": "text", + "content": "It should be emphasized that to enable " + }, + { + "bbox": [ + 46, + 551, + 287, + 647 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 551, + 287, + 647 + ], + "type": "text", + "content": " free-viewpoint rendering, the input videos must contain " + }, + { + "bbox": [ + 46, + 551, + 287, + 647 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 551, + 287, + 647 + ], + "type": "text", + "content": " scene information (an example capture process is shown in the supplementary video); otherwise, artifacts will appear in unobserved areas that were not even seen in the input videos. However, our HOSNeRF is not restricted by such capturing requirements, as demonstrated on the NeuMan dataset [14], which features small camera motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 141, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 141, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 141, + 671 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "We conduct ablation studies on our collected dataset to evaluate the effectiveness of each proposed component in HOSNeRF. We progressively ablate each component from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 333, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 405 + ], + "type": "text", + "content": "state embeddings, object bones, foreground masks, optical flow, and cycle consistency supervision. As shown in Tab. 2, the performance of HOSNeRF progressively drops with the disabling of each component, with the full model nearly achieving the best performances, which demonstrates the effectiveness of our designs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 405, + 546, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 546, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 546, + 585 + ], + "type": "text", + "content": "We visualize representative scenes with large (Fig. 6) and small (Fig. 7) objects to further evaluate our proposed designs. As shown in Fig. 6(b) and 7(b), the base models are unable to represent the interacted objects for the human and the scene. In contrast, the proposed state embedding is capable of representing the state transitions of small objects (Fig. 7(c)), but it fails to represent the deformations of large objects (Fig. 6(c)). Together with our proposed object bones, the deformation of large objects can be effectively estimated through the human-object skeleton hierarchy, enabling large objects modeling in Fig. 6(d). In addition, the full model can further improve the synthesis quality of fine details such as correcting the wrong right arm by comparing Fig. 7(c) and (d). These comparisons further demonstrate the importance and effectiveness of the proposed designs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 597, + 378, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 378, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 378, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We introduced a novel framework of HOSNeRF, the first work to achieve " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " free-viewpoint high-fidelity novel view synthesis for dynamic scenes with human-environment interactions from a single video. To tackle the challenges of such a task, we first introduced the new object bones to the human skeleton hierarchy to effectively estimate the fast human-object motion in our dynamic human-object model. Then, we proposed the state-conditional" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18490" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 371 + ], + "type": "image", + "image_path": "dd2404d411a9d05a50623dddb866afd3c10660eb94fc917d95f966b6eb6a27a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 377, + 479, + 389 + ], + "lines": [ + { + "bbox": [ + 113, + 377, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 113, + 377, + 479, + 389 + ], + "type": "text", + "content": "Figure 5: Qualitative comparisons of HOSNeRF and NeuMan [14] on the NeuMan dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 394, + 168, + 462 + ], + "blocks": [ + { + "bbox": [ + 49, + 394, + 168, + 462 + ], + "lines": [ + { + "bbox": [ + 49, + 394, + 168, + 462 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 168, + 462 + ], + "type": "image", + "image_path": "49941d43ad00be81ad71221b6a9a7a68643732672f7c0b2a682c07294d9a5d05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 463, + 135, + 470 + ], + "lines": [ + { + "bbox": [ + 82, + 463, + 135, + 470 + ], + "spans": [ + { + "bbox": [ + 82, + 463, + 135, + 470 + ], + "type": "text", + "content": "(a) Groundtruth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 394, + 287, + 462 + ], + "blocks": [ + { + "bbox": [ + 169, + 394, + 287, + 462 + ], + "lines": [ + { + "bbox": [ + 169, + 394, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 169, + 394, + 287, + 462 + ], + "type": "image", + "image_path": "dcc17051effb55641c0a1598c79b131552c8fa1e978d0065378bde7c13e991aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 462, + 253, + 471 + ], + "lines": [ + { + "bbox": [ + 200, + 462, + 253, + 471 + ], + "spans": [ + { + "bbox": [ + 200, + 462, + 253, + 471 + ], + "type": "text", + "content": "(b) Base Model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 471, + 167, + 538 + ], + "blocks": [ + { + "bbox": [ + 49, + 471, + 167, + 538 + ], + "lines": [ + { + "bbox": [ + 49, + 471, + 167, + 538 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 167, + 538 + ], + "type": "image", + "image_path": "3fc6d737b3242ad436b628837fe323593d34682726ed548541f8ab29a4787f1d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 539, + 152, + 548 + ], + "lines": [ + { + "bbox": [ + 64, + 539, + 152, + 548 + ], + "spans": [ + { + "bbox": [ + 64, + 539, + 152, + 548 + ], + "type": "text", + "content": "(c) With State Embedding" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 559, + 268, + 571 + ], + "lines": [ + { + "bbox": [ + 65, + 559, + 268, + 571 + ], + "spans": [ + { + "bbox": [ + 65, + 559, + 268, + 571 + ], + "type": "text", + "content": "Figure 6: Ablation of HOSNeRF for large objects." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 168, + 471, + 287, + 538 + ], + "blocks": [ + { + "bbox": [ + 168, + 471, + 287, + 538 + ], + "lines": [ + { + "bbox": [ + 168, + 471, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 168, + 471, + 287, + 538 + ], + "type": "image", + "image_path": "c7f464820ea3d94f0904d54b050a755f54b566f16d46d4001805bd3c3aab72e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 539, + 270, + 557 + ], + "lines": [ + { + "bbox": [ + 182, + 539, + 270, + 557 + ], + "spans": [ + { + "bbox": [ + 182, + 539, + 270, + 557 + ], + "type": "text", + "content": "(d) With State Embedding & Object Bones" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "type": "text", + "content": "human-object and scene representations for handling significant object state changes. With our designed training strategies, HOSNeRF produced significant improvements over SOTA approaches, and enabled " + }, + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "type": "text", + "content": " free-viewpoint rendering of any frame from a single video and supported rendering views with novel human and object poses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Limitations and Future Work. Our HOSNeRF focuses on dynamic human-objects modeling and lacks the capability to represent dynamic backgrounds. It remains challenging but is worthwhile researching models for reconstructing high-fidelity dynamic unbounded backgrounds." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 395, + 426, + 462 + ], + "blocks": [ + { + "bbox": [ + 307, + 395, + 426, + 462 + ], + "lines": [ + { + "bbox": [ + 307, + 395, + 426, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 426, + 462 + ], + "type": "image", + "image_path": "bd39841f155c75f89a4e85100277fe62106a7c4fed3066a935f06597dfebbe13.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 427, + 395, + 545, + 462 + ], + "blocks": [ + { + "bbox": [ + 427, + 395, + 545, + 462 + ], + "lines": [ + { + "bbox": [ + 427, + 395, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 427, + 395, + 545, + 462 + ], + "type": "image", + "image_path": "1137a1d4414030a78649c32f490cb841ebf302a3f74c3326a4660eedfb9d99de.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 471, + 426, + 539 + ], + "blocks": [ + { + "bbox": [ + 340, + 463, + 393, + 471 + ], + "lines": [ + { + "bbox": [ + 340, + 463, + 393, + 471 + ], + "spans": [ + { + "bbox": [ + 340, + 463, + 393, + 471 + ], + "type": "text", + "content": "(a) Groundtruth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 471, + 426, + 539 + ], + "lines": [ + { + "bbox": [ + 307, + 471, + 426, + 539 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 426, + 539 + ], + "type": "image", + "image_path": "9d915738526257303c24a166225afcc53827c504785696ea4d74f7588430b45d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 539, + 410, + 548 + ], + "lines": [ + { + "bbox": [ + 323, + 539, + 410, + 548 + ], + "spans": [ + { + "bbox": [ + 323, + 539, + 410, + 548 + ], + "type": "text", + "content": "(c) With State Embedding" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 551, + 528, + 563 + ], + "lines": [ + { + "bbox": [ + 322, + 551, + 528, + 563 + ], + "spans": [ + { + "bbox": [ + 322, + 551, + 528, + 563 + ], + "type": "text", + "content": "Figure 7: Ablation of HOSNeRF for small objects." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 427, + 471, + 545, + 539 + ], + "blocks": [ + { + "bbox": [ + 460, + 463, + 512, + 471 + ], + "lines": [ + { + "bbox": [ + 460, + 463, + 512, + 471 + ], + "spans": [ + { + "bbox": [ + 460, + 463, + 512, + 471 + ], + "type": "text", + "content": "(b) Base Model" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 471, + 545, + 539 + ], + "lines": [ + { + "bbox": [ + 427, + 471, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 427, + 471, + 545, + 539 + ], + "type": "image", + "image_path": "723c40e71ce5e3d6a6784e7525591d920e81eb91f2b34a3cbdc5be2899cdcf8a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 539, + 511, + 548 + ], + "lines": [ + { + "bbox": [ + 460, + 539, + 511, + 548 + ], + "spans": [ + { + "bbox": [ + 460, + 539, + 511, + 548 + ], + "type": "text", + "content": "(d) Full Model" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 565, + 421, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 565, + 421, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 565, + 421, + 578 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, Mike Zheng Shou's Start-Up Grant from NUS, the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001. This research is also supported by National Research Foundation, Singapore and A\\*STAR, under its RIE 2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) grant call (grant no. I2001E0059). Jia-Wei Liu is also supported by NUS IDS-ISEP scholarship. Thanks to Yufei Shi for helpful discussions." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18491" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 145 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 286, + 201 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 286, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 286, + 224 + ], + "type": "text", + "content": "[3] Ang Cao and Justin Johnson. Hexplane: a fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 286, + 290 + ], + "type": "text", + "content": "[4] Mingfei Chen, Jianfeng Zhang, Xiangyu Xu, Lijuan Liu, Yu-jun Cai, Jiashi Feng, and Shuicheng Yan. Geometry-guided progressive nerf for generalizable and efficient neural human rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 222-239. Springer, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 286, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 344 + ], + "type": "text", + "content": "[5] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 286, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 399 + ], + "type": "text", + "content": "[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 286, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 286, + 445 + ], + "type": "text", + "content": "[7] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 286, + 490 + ], + "type": "text", + "content": "[8] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 286, + 544 + ], + "type": "text", + "content": "[9] Xiangjun Gao, Jiaolong Yang, Jongyoo Kim, Sida Peng, Zicheng Liu, and Xin Tong. Mps-nerf: Generalizable 3d human rendering from multiview images. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–12, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "text", + "content": "[10] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "type": "text", + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 677 + ], + "type": "text", + "content": "[12] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[13] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 362-379. Springer, 2016." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 402-418. Springer, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "type": "text", + "content": "[15] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "type": "text", + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "type": "text", + "content": "[17] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 305, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 370 + ], + "type": "text", + "content": "[18] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 419-436. Springer, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 372, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 436 + ], + "type": "text", + "content": "[19] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[20] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 493, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 524 + ], + "type": "text", + "content": "[21] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. arXiv preprint arXiv:2211.11082, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 579 + ], + "type": "text", + "content": "[22] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 581, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 634 + ], + "type": "text", + "content": "[23] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[24] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[25] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural vol" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18492" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "umes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 237 + ], + "type": "text", + "content": "[28] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 295, + 287, + 348 + ], + "type": "text", + "content": "[30] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 404 + ], + "type": "text", + "content": "[31] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 470 + ], + "type": "text", + "content": "[32] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 472, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 514 + ], + "type": "text", + "content": "[33] Sergey Prokudin, Michael J Black, and Javier Romero. Smplpix: Neural avatars from 3d human models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1810-1819, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 516, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 516, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 516, + 287, + 568 + ], + "type": "text", + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "text", + "content": "[35] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 626, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 626, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 626, + 287, + 690 + ], + "type": "text", + "content": "[36] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yan-Pei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15872-15882, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[37] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthe" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "sis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "type": "text", + "content": "[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 142, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 545, + 196 + ], + "type": "text", + "content": "[39] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "type": "text", + "content": "[40] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. arXiv preprint arXiv:2210.15947, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "type": "text", + "content": "[41] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum, volume 27, pages 1781-1787. Wiley Online Library, 2008." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "text", + "content": "[42] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. Advances in Neural Information Processing Systems, 34:12278-12291, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 345, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 345, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 345, + 545, + 399 + ], + "type": "text", + "content": "[43] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 401, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 401, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 401, + 545, + 465 + ], + "type": "text", + "content": "[44] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 578 + ], + "type": "text", + "content": "[46] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6226-6237, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 581, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 645 + ], + "type": "text", + "content": "[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[48] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18493" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 652 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "text", + "content": "[50] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humaner: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16210-16220, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[51] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "text", + "content": "[52] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D'nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems, 35:32653-32666, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 339 + ], + "type": "text", + "content": "[53] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 341, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 287, + 384 + ], + "type": "text", + "content": "[54] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. Advances in Neural Information Processing Systems, 34:14955–14966, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 287, + 440 + ], + "type": "text", + "content": "[55] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "text", + "content": "[56] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. arXiv preprint arXiv:2212.07626, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 540 + ], + "type": "text", + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[58] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 587, + 287, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 287, + 652 + ], + "type": "text", + "content": "[59] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18494" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_content_list.json b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3dbef8fc31f33b3839e83fbbb71c3cb6a0b4be8c --- /dev/null +++ b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_content_list.json @@ -0,0 +1,2027 @@ +[ + { + "type": "text", + "text": "HRS-Bench: Holistic, Reliable and Scalable Benchmark for Text-to-Image Models", + "text_level": 1, + "bbox": [ + 202, + 130, + 769, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eslam Mohamed Bakr1, Pengzhan Sun2**, Xiaqian Shen1**, Faizan Farooq Khan1**, Li Erran Li3, Mohamed Elhoseiny1", + "bbox": [ + 210, + 202, + 756, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{eslam.abelrahman, xiaoqian.shen, faizan.khan, mohamed.elhoseiny}@kaust.edu.sa", + "bbox": [ + 210, + 241, + 756, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pengzhan@comp.nus.edu.sg, lilimam@amazon.com", + "bbox": [ + 223, + 276, + 745, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ King Abdullah University of Science and Technology (KAUST)", + "bbox": [ + 225, + 292, + 736, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National University of Singapore ${ }^{3} \\mathrm{AWS AI},$ Amazon", + "bbox": [ + 261, + 309, + 697, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Text-to-Image (T2I) models have been extensively studied, especially with the emergence of diffusion models that achieve state-of-the-art results on T2I synthesis tasks. However, existing benchmarks heavily rely on subjective human evaluation, limiting their ability to holistically assess the model's capabilities. Furthermore, there is a significant gap between efforts in developing new T2I architectures and those in evaluation. To address this, we introduce HRS-Bench, a concrete evaluation benchmark for T2I models that is Holistic, Reliable, and Scalable. Unlike existing benchmarks that focus on limited aspects, HRS-Bench measures 13 skills that can be categorized into five major categories: accuracy, robustness, generalization, fairness, and bias. In addition, HRS-Bench covers 50 scenarios, including fashion, animals, transportation, food, and clothes. We evaluate nine recent large-scale T2I models using metrics that cover a wide range of skills. A human evaluation aligned with $95\\%$ of our evaluations on average was conducted to probe the effectiveness of HRS-Bench. Our experiments demonstrate that existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions. We hope that our benchmark help ease future text-to-image generation research. The code and data are available at https://eslambakr.github.io/hrsbench.github.io/.", + "bbox": [ + 76, + 393, + 473, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 773, + 209, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-Image Synthesis (T2I), one of the essential multimodal tasks, witnessed remarkable progress starting from conditional GANs [58, 44, 74, 29, 78], which are shown to work on simple datasets [46, 69, 72, 36], to recently diffusion models [20, 61, 18, 10, 57, 55, 77, 45, 59], which are trained", + "bbox": [ + 76, + 799, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37206a980a5ac74d8b6214facb1f6d29db7672bbca5b6e6612114c3e25b821a9.jpg", + "image_caption": [ + "Figure 1: An overview of our proposed benchmark, HRS-Bench, measures 13 skills which could be grouped into five major categories; accuracy, robustness, generalization, fairness, and bias." + ], + "image_footnote": [], + "bbox": [ + 504, + 359, + 885, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on large-scale datasets, e.g., LAION [63, 62].", + "bbox": [ + 498, + 656, + 803, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the rapid progress, the existing models face several challenges, e.g., they cannot generate complex scenes with the desired objects and relationship composition [25, 37]. Furthermore, assessing the T2I models should include more than just fidelity, e.g., the ability to compose multiple objects and generate emotionally grounded or creative images. Therefore, some recent efforts are focusing on improving the existing metrics [23] or proposing new metrics that cover new aspects, such as bias [75], compositions [37, 25, 49]. Moreover, some other works propose new benchmarks, summarized in Table 1, that assess different aspects, e.g., counting [61, 12, 51], social-bias [12], and object fidelity [23, 51]. Even with various benchmarks available, they tend to only cover a limited range of aspects while overlooking crucial evaluation criteria such as robustness,", + "bbox": [ + 496, + 674, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 96, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20041", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/dd310ca2989cf661882aa7bee291781d5335de0a9f26efe53275367e8887a79b.jpg", + "table_caption": [ + "Table 1: Comparison of text-to-image benchmarks in terms of: 1) The number of evaluated models. 2) The number of the covered skills. 3) The number of utilized metrics. 4) The evaluation type, whether human or metric based or both. 5) The number of prompts. 6) The prompt generation type, whether a template, human-based or both. 7) Whether there are different hardness levels are included. 8) Number of annotators contributed to the evaluation." + ], + "table_footnote": [], + "table_body": "
MethodEval TypePrompt Type
# Models# Skills# MetricsHumanAuto# PromptsTemplateHumanHardness Levels# Annotators
DrawBench [61]540✓✘200✘✓✘25
DALL-EVAL [12]453✓✘7330✓✘✘6
HE-T2I [51]230✓✘90✘✓✓20
TISE [23]735✘✓N/A✘✓✘N/A
HRS-Bench (Ours)91317✓✓45000✓✓✓1000
", + "bbox": [ + 191, + 160, + 777, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fairness, and creativity.", + "bbox": [ + 75, + 285, + 233, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we propose our Holistic, Reliable, and Scalable benchmark, dubbed HRS-Bench. In contrast to existing benchmarks, we measure a wide range of different generative capabilities, precisely 13 skills which can be grouped into five major categories, as demonstrated in Figure 1; accuracy, robustness, generalization, fairness, and bias. Most of these skills have never been explored in the T2I context, such as creativity, fairness, anonymization, emotion-grounding, robustness, and visual-text generation. Even the other previously explored skills were studied from a limited perspective, for instance, DALL-EVAL [12] studied the social bias by generating limited template-based prompts; only 145 prompts. This limited evaluation scope may result in immature or sometimes misleading conclusions. In addition, to facilitate the evaluation process for existing and future architectures, we heavily rely on automatic evaluations, where a wide range of metrics are utilized in the evaluation criteria. Moreover, HRS-Bench covers 50 scenarios, e.g., fashion, animals, transportation, and food. Figure 2 demonstrates the top 15 applications and their object distribution. We evaluate nine recent large-scale T2I models, i.e., Stable-Diffusion V1 [59] and V2 [3], DALL-E 2 [55], GLIDE [45], CogView-V2 [21], Paella [57], minDALL-E [2], DALL-E-Mini [1], and Struct-Diff [25]. In addition, our benchmark is scalable with automatic evaluation, and thus can be extended for any new architectures. To probe the effectiveness of our HRS-Bench, we conduct a human assessment that aligns well with our evaluations by $95\\%$ on average. Our contributions can be summarized as follows:", + "bbox": [ + 73, + 306, + 473, + 743 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a Holistic, Reliable, and scalable T2I benchmark called HRS-Bench, depicted in Figure 1, which assess 13 skills covering 50 scenarios.", + "- Propose a new T2I alignment metric, called AC-T2I, which overcomes the composition limitations of existing large Vision-Language Models (VLMs) [31, 73].", + "- Nine T2I models are assessed based on our benchmark, including commercial and open-sourced ones.", + "- We verify the effectiveness of our HRS-Bench metric by conducting a human evaluation for $10\\%$ of our data" + ], + "bbox": [ + 94, + 750, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30a35cd2424c0cae31944cb3ddb282b3bea5fc9a86dc3c4e67008097ef79208a.jpg", + "image_caption": [ + "Figure 2: Pie chart demonstrates the wide range of covered scenarios by our proposed benchmark, termed HRS-Bench, and their object distribution." + ], + "image_footnote": [], + "bbox": [ + 517, + 282, + 875, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "per skill that shows excellent alignment.", + "bbox": [ + 529, + 560, + 799, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Driven by these holistic evaluations, several conclusions and findings are discussed. For instance, existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions.", + "bbox": [ + 517, + 575, + 893, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Revisiting Text-to-Image Benchmarks", + "text_level": 1, + "bbox": [ + 498, + 664, + 841, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, there have been rapid efforts toward designing better image-generation models [20, 61, 18, 10, 57, 55, 77, 45, 59]. However, most existing metrics suffer from several limitations that make them unreliable [23, 7, 22, 6, 75, 49, 66, 73]. Table 1 summarizes the existing T2I benchmarks. -DrawBench. Imagen [61] proposes DrawBench to evaluate the T2I models from other aspects along with the image quality. Whereas, DrawBench covers four skills; counting, compositions, conflicting, and writing, by collecting 200 prompts in total. Despite the simplicity and the limited scope of DrawBench, their efforts are appreciated as the first attempt to assess other aspects rather than the image quality. -DALL-EVAL. [12] It proposes a toolkit called PAINTSKILLS, which assesses three simple visual", + "bbox": [ + 496, + 689, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20042", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "reasoning skills; object recognition, object counting, and spatial relation understanding, alongside two social bias skills; gender and racial bias. To facilitate the automatic evaluation, they built a unity simulator including limited objects to collect the datasets, approximately 80 object classes. In contrast, we cover more than 700 object classes. DETR [9] is utilized for visual reasoning skills evaluation after being fine-tuned on the synthetic dataset collected from the simulator. However, DALL-EVAL evaluates only four models using only 7k prompts and six annotators. Therefore, there is a need for a more comprehensive benchmark that can take into account a broader range of models, prompts, and annotators to provide a more thorough evaluation.", + "- HE-T2I. [51] It proposes 32 possible aspects to benchmark T2I models. However, only three are evaluated; counting, shapes, and faces, and the rest are left unexplored. The three aspects are evaluated using human evaluations, where twenty annotators have contributed to the evaluation.", + "-TISE. [23] It introduces a bag of metrics to evaluate T2I models from three aspects; positional alignment, counting, and fidelity. In addition, three fidelity metrics are introduced, i.e., $IS^{*}$ , $OIS$ , and $OFID$ , and two alignment metrics PA for positional alignment, and CA for counting alignment." + ], + "bbox": [ + 75, + 90, + 472, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. HRS-Bench", + "text_level": 1, + "bbox": [ + 76, + 453, + 200, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first dissect the skills definition, then demonstrate the prompts collection pipeline.", + "bbox": [ + 76, + 479, + 468, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Skills", + "text_level": 1, + "bbox": [ + 76, + 520, + 156, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Accuracy", + "text_level": 1, + "bbox": [ + 76, + 542, + 197, + 558 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Counting. A reliable T2I model should be able to ground the details in the prompt. One form of these details is objects binding with a specific frequency, e.g., \"four cars are parked around five benches in a park\".", + "- Visual Text. Another essential aspect of assessing the model is generating high-quality text in wild scenes, e.g., \"a real ballroom scene with a sign written on it,\"teddy bear on the dining table!\"\". The importance of such skill comes from intervening in many scenarios, e.g., education applications, preparing illustration content, and designing billboards.", + "-Emotion. We measure to what extent the model can generate emotion-grounded images [5, 4, 41, 40], e.g., \"a rainy scene about cake, which makes us feel excitement.\"", + "-Fidelity. Image fidelity indicates how accurately an image represents the underlying source distribution [64]." + ], + "bbox": [ + 73, + 566, + 470, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Robustness", + "text_level": 1, + "bbox": [ + 76, + 815, + 210, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To assess the T2I model's robustness, we cover two types of transformations; invariance and equivariance.", + "bbox": [ + 75, + 839, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Invariance. Two skills are introduced to measure the invariance robustness: consistency and typos.", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Consistency. We measure the sensitivity of different T2I models towards prompt variations while keeping the same meaning and semantics, i.e., paraphrasing. For instance, generated images from these prompts, \"a woman is standing in front of a mirror, carefully selecting the perfect handbag\" and \"in front of a mirror, a woman is selecting the perfect handbag for the day\" should hold the same semantics.", + "-Typos. Two natural perturbations are utilized to assess the models against possible sensibleness noise that users could cause during inference, i.e., typos and wrong capitalization.", + "Equivariance. Three different compositions are explored for the equivariance robustness. Specifically, we study three types of compositions, i.e., spatial, attribute-specific, and action compositions.", + "-Spatial composition. In contrast to the counting skill which only measures the models' ability to compose multiple objects into a coherent scene, spatial composition additionally measures their ability to ground the detailed spatial relationship instructions mentioned in the input prompt, e.g., \"a person and a dog in the middle of a cat and a chair\".", + "- Attribute-specific composition. Two types of attributes are controlled to study the attribute binding ability, i.e., colors and size attributes. For instance, \"an orange cat, a red dog, and a blue chair\" and \"a banana which is smaller than a person and bigger than a car\", for colors and size attribute binding, respectively.", + "-Action composition. It incorporates different subjects that doing different actions, e.g., \"a woman is playing in the water, and an elephant is walking through woods\"." + ], + "bbox": [ + 496, + 90, + 893, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.3 Generalization", + "text_level": 1, + "bbox": [ + 500, + 571, + 658, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Creativity. In this skill, models aim to generate images that represent not only the textual description but are also imaginative and novel. The creativity skill can be regarded as an out-of-distribution generation [30]. Accordingly, we devised innovative text prompts that are conceptually plausible but may need to be more readily available in standard training data sources, detailed later in Section 3.2.", + "bbox": [ + 496, + 599, + 893, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.4 Fairness", + "text_level": 1, + "bbox": [ + 500, + 736, + 612, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define fairness as the performance disparity among different sub-groups [24, 52]. A fair model should achieve the same performance on an arbitrary metric since no correlation exists between the metric and the protected attribute. Two attributes have been studied, i.e., gender and style. Following [12], gender refers to sex [32, 54] not the gender identity [43, 17]. We use two gender categories; male and female. The styles are animation, real, sketch, black and white, and weather conditions; sunny, rainy, and cloudy.", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20043", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.5 Bias", + "text_level": 1, + "bbox": [ + 76, + 90, + 161, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assess the spurious correlation of the model towards predefined attributes, i.e., gender, race, and age. Using agnostic prompts towards a specific attribute, e.g., gender, the model should produce balanced generations of different classes of this attribute. For instance, the gender agnostic prompt could be, \"Two persons are working on their laptops\".", + "bbox": [ + 75, + 114, + 470, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Prompts Collection", + "text_level": 1, + "bbox": [ + 76, + 215, + 261, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each skill, we collect 3k prompts. To ensure our benchmark is holistic enough, we split the prompts equally into three hardness levels, i.e., easy, medium, and hard. In addition, half of the prompts are human-based, and the other half is template-based, depicted in Figure 3. We filter human prompts manually from existing datasets [70, 33, 31]. We use the foundation model GPT-3.5 [47], text-davinci-003* to facilitate prompts generation, which will be abbreviated as GPT-3.5 later for convenience.", + "bbox": [ + 75, + 239, + 468, + 375 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "-Fidelity. The human prompts are sifted from [70]. Whereas, the template-based prompts are created by defining a template that describes a styled scene that contains some objects, as shown in Figure 3. Then, we create the meta-prompt by sampling the styles from pre-defined styles and the objects from LVIS dataset [26]. Finally, GPT-3.5 [47] is utilized to generate the final prompts.", + "- Consistency and Typos. Consequently, the fidelity prompts are fed to Parrot [14] and NLU augmenter [19] to produce augmented prompts for consistency and typos, respectively. For consistency, we differentiate between the three hardness levels based on the similarity between the fidelity prompt and the augmented prompts using RoBERTa [38]. For typos, the number of introduced typos controls the three hardness levels, i.e., 1-2, 3-4, and 5-6, respectively.", + "-Counting. Given a meta-prompt, GPT-3.5 generate a realistic scenario that contains N objects. To generate the meta-prompts, we randomly samples the number of objects and the objects classes from LVIS dataset [26].", + "-Visual Text. We utilize GPT-3.5 to generate short descriptions which fit on a sign in a crowded scene. Then, we control the hardness levels by the text length and the surrounding scene complexity.", + "- Emotion. We sample random objects from LVIS [26], then append an emotion indicator word forming the meta-prompt. Finally, GPT-3.5 is utilized to generate the final prompts.", + "-Creativity. We craft text prompts that are challenging yet still within the realm of imagination. For easy level, we obtain subject, object, and relationship from Visual Genome [33] and obtain triplets by different combinations. Then sift through all the combinations from triplets extracted from LAION [62] dataset, and retain only the uncommon triplets. For the medium level, we fed the uncommon" + ], + "bbox": [ + 73, + 376, + 470, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "triplet to GPT-3.5 with the instruction: “Describe subject, relation and object in an imaginative way that will never be seen in the real world” and manually filter the undesirable sentences. Finally, to generate challenging prompts for the hard level, we experiment with various prompts to encourage GPT-3.5 to generate counterproductive sentences, as shown in Figure 3.", + "bbox": [ + 500, + 90, + 893, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Compositionality. We study three composition types, i.e., spatial, attribute-binding, and actions. The spatial prompts are collected using a pre-defined template, where a wide range of relations is utilized, e.g., \"on the right of\", \"above\", and \"between.\" For attribute-binding, two attributes are exploited, i.e., colors and size. For each hardness level, the number of objects' compositions increased, ranging from 2 to 4. For the action-level compositional generation, we design prompts with multiple combinations of actions starting from ComCLIP [31]. We combine two sentences from ComCLIP [31] for the easy level. Then, for medium and hard levels, we randomly choose one sentence and feed 'Extend text' to let the subject have at least three actions.' and 'Extend text' with other subjects doing other actions' into GPT-3.5, respectively, to obtain the final prompt. Detailed examples are demonstrated in Figure 3.", + "bbox": [ + 496, + 196, + 893, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "-Bias. Random objects are sampled from LVIS datasets [26], combined with a pre-defined template, creating a metaprompt. Then, the meta-prompt is fed to GPT-3.5 to produce the final prompt, as depicted in Figure 3. To ensure the prompts are agnostic towards the protected attributes, i.e., gender, race, and age, we manually validate them.", + "bbox": [ + 496, + 438, + 893, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "-Fairness. We adapt the bias prompts. For gender fairness, we replace gender-agnostic words, such as a person, with gender-specific words, such as man and woman. Whereas for style fairness, a style indicator is appended to the beginning of the bias prompt, as shown in Figure 3.", + "bbox": [ + 496, + 529, + 893, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Evaluation for our Benchmark", + "text_level": 1, + "bbox": [ + 498, + 617, + 782, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 4, we categorize the skills based on the evaluation criteria, one-to-many mapping. Thus the same skill may be assigned to several metrics.", + "bbox": [ + 496, + 642, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Detection-Based Metrics", + "text_level": 1, + "bbox": [ + 500, + 696, + 723, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize UniDet [76] for counting, spatial and attribute compositions because it supports a wide range of object classes, i.e., exceeding 700.", + "bbox": [ + 496, + 719, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score, where Precision assesses the accuracy of additional objects, and Recall assesses the accuracy of missing objects.", + "bbox": [ + 496, + 763, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "-Spatial Compositions. Using a simple geometry module, we use the predicted bounding boxes to validate whether the spatial relation is grounded correctly. For instance, given the prompt \"A cat above a car\", the predicted bounding boxes will be $\\{x_{min}^{1},y_{min}^{1},x_{max}^{1},y_{max}^{1}\\}$", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "* https://platform.openai.com/docs/models/gpt-3-5", + "bbox": [ + 94, + 886, + 359, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20044", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb17387a2bfdfd71fcf5785194a3985e82f28d58c32ca44c17a7111a360efce7.jpg", + "image_caption": [ + "Figure 3: Our prompt generation pipeline. First, we create a meta-prompt, which is a template-based prompt (in blue). Then, we sample the skill-related attributes (in orange). Finally, we generate the final prompt using ChatGPT (in green)." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 890, + 290 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d4adacee38b78d062ba6564271775671d0491adc52c21221508376ef6e281b1.jpg", + "image_caption": [ + "Figure 4: On the left is our evaluation taxonomy. On the right, we demonstrate our metric Augmented Captioner-based T2I alignment metric." + ], + "image_footnote": [], + "bbox": [ + 81, + 354, + 460, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\{x_{min}^2, y_{min}^2, x_{max}^2, y_{max}^2\\}$ for the cat and the car, respectively, and the grounded spatial relation is above. Then, our geometry module will assess whether the spatial relation, i.e., above, is grounded correctly based on the following condition: $(y_{min}^1 < y_{min}^2)$ or $(y_{max}^1 < y_{max}^2)$ .", + "bbox": [ + 75, + 542, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Attributes Compositions. The predicted bounding boxes' sizes are used for the size composition to validate whether the size relation is grounded correctly. Whereas for color composition, first, we convert the image to the hue color space, then calculate the average hue value within the box and compare it to the pre-defined color space.", + "bbox": [ + 73, + 622, + 468, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "-Visual Text. We adopt Textsnake [39] and SAR [34] for text detection and recognition, respectively. The recognition accuracy is measured by the Character Error Rate (CER) [42] and the Normalized Edit Distance (NED) [65].", + "bbox": [ + 73, + 715, + 470, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Alignment-Based Metrics", + "text_level": 1, + "bbox": [ + 75, + 797, + 310, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Three alignment paradigms are explored; Text-to-Image (T2I) (Sec. 4.2.1), Text-to-Image-to-Text (TIT) (Sec. 4.2.3), and Image-to-Image I2I (Sec. 4.2.4). In addition, we introduce our novel Augmented Captioner-based T2I Alignment metric, termed AC-T2I (Sec. 4.2.2).", + "bbox": [ + 75, + 824, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 T2I Alignment", + "text_level": 1, + "bbox": [ + 498, + 357, + 658, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One possible solution to assess the T2I model's grounding ability is measuring the text and image correlation, e.g., CLIPScore [27, 53]. While CLIP is widely used, its effectiveness has been repeatedly questioned, as it is not sensitive to fine-grained text-image alignment and fails to understand compositions [66, 73]. For instance, [73] shows that CLIP [53] can not distinguish between \"the horse is eating the grass\" and \"the grass is eating the horse\". This motivates us to propose our novel augmented captioner-based T2I alignment metric, termed AC-T2I, depicted in Figure 4.", + "bbox": [ + 496, + 381, + 893, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 AC-T2I Alignment Metric.", + "text_level": 1, + "bbox": [ + 498, + 554, + 740, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a new T2I alignment metric, called AC-T2I, which overcomes the compositional relationship's limitations of existing large Vision-Language Models (VLMs) [31, 73], by utilizing the n-grams based metric, e.g., CIDEr [68] and BLEU [48]. To this end, we decompose our metric into two steps; first, we transform the image embedding to text space using an image captioning model, then augment the generated caption to make the metric comprehensive enough for different perturbations.", + "bbox": [ + 496, + 579, + 893, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "-Reformatting T2I as TIT. We reformat T2I as a TIT alignment by transforming the image features to text feature space, using an arbitrary function $G(\\cdot)$ . The function $G(\\mathcal{I})$ could be interpreted as an image captioner, e.g., BLIP2 [35]. As shown in Figure 4, given a text prompt $P^{org}$ , $N_{i}$ images $\\mathcal{I} = \\{I_k\\}_{k=1}^{N_i}$ are generated, which are fed to an image captioner $G(\\mathcal{I})$ producing $N_{c_i}$ captions $\\mathcal{C} = \\{C_k\\}_{k=1}^{N_{c_i}}$ , where $N_{c_i}$ is the number of generated captions per image. Finally, the $N_{c_i}$ captions are automatically evaluated using CIDEr [68] and BLEU [48] against the input prompt $P^{org}$ .", + "bbox": [ + 496, + 715, + 893, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Comprehensive TIT. Instead of considering only the prompt $P^{org}$ as the GT caption, $N_{t}$ augmented prompts", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{P}^{aug} = \\{P_k^{aug}\\}_{k=1}^{N_t}$ are generated using GPT-3.5, to measure the similarities comprehensively. To this end, we must ensure the GT is holistic enough; therefore, the rephrased version of the prompt $P^{org}$ should be considered correct. Accordingly, the whole GT-prompt set for each image is defined as $\\mathcal{P} = \\{P^{org}, \\mathcal{P}^{aug}\\} = \\{P_k\\}_{k=1}^{N_t+1}$ , i.e., one original prompt plus $N_t$ augmented prompts.", + "bbox": [ + 75, + 89, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, for each prompt $P^{org}$ we calculate the alignment score for each generated prompt-caption pair and select the highest score as the final alignment score, Eq. 1.", + "bbox": [ + 75, + 196, + 468, + 242 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nO _ {t} = \\frac {1}{N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} \\max _ {1 \\leq j \\leq N _ {c _ {i}}, 1 \\leq k \\leq N _ {t} + 1} S _ {t} \\left(C _ {i, j}, P _ {k}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 242, + 468, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{t}(.)$ is the text similarity scoring function, e.g., CIDEr [68], BLEU [48].", + "bbox": [ + 75, + 292, + 468, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 TIT Alignment", + "text_level": 1, + "bbox": [ + 76, + 340, + 236, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "-Emotions. We explore a visual emotion classifier as illustrated in Sec. 4.3. Moreover, we apply our proposed metric, AC-T2I (Eq. 1), to avoid the aforementioned CLIP limitations, detailed in Section 4.2.1 and Section 4.2.2. The number of generated images per prompt $N_{i}$ , generated captions per image $N_{c_i}$ , and the augmented prompts $N_{t}$ are set to 3, 5, and 9, respectively.", + "bbox": [ + 73, + 364, + 472, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Creativity. We assessed the generated images to deviate from the training data while simultaneously adhering to the provided text prompts using our novel metric AC-T2I and the deviation metric (Sec. 4.2.4). We set $N_{i}$ and $N_{c_i}$ to 3 and 5, respectively. Since it is hard to rephrase our novel prompts while maintaining its creative intent correctly, there will be no augmented prompts for it ( $N_{t} = 0$ ).", + "bbox": [ + 73, + 470, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "-Gender and styles fairness and action compositions. The fairness score is defined as the disparities in subgroups' performance [24, 52], Eq. 2", + "bbox": [ + 73, + 575, + 470, + 619 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF a i r n e s s _ {s c o r e} = \\frac {1}{N _ {s} C _ {2}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = i + 1} ^ {N _ {s}} \\frac {1 0 0 \\times | A (i) - A (j) |}{\\operatorname* {m a x} (A (i) , A (j))}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 621, + 468, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\frac{100}{N_sC_2\\times max(A(i),A(j))}$ is a normalization factor, $N_{s}$ is the number of sub-groups, e.g., two for the gender and $A$ is the accuracy measure, e.g., AC-T2I or CLIP scores. The less is, the better for Fairness score. Consequently, for the action composition, we exploit the AC-T2I metric, where $N_{i},N_{c_{i}}$ , and $N_{t}$ are set to 3, 5, and 9, respectively.", + "bbox": [ + 75, + 676, + 468, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.4 I2I Alignment", + "text_level": 1, + "bbox": [ + 75, + 786, + 230, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Creativity. In addition to AC-T2I, we measure the deviation from the training data to indicate creativity. Accessing large models' training data is challenging, however, most of them are trained on LAION [63]. Accordingly, we use LAION image-text retrieval tools [8] to fetch training data, which search among the dataset using CLIP [53] and a KNN index", + "bbox": [ + 73, + 809, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to seek top-100 nearest images, denoted as $\\mathcal{I}^{train}$ for each prompt. The deviation score is calculated based on Eq. 3.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\triangle \\left(\\mathcal {I} ^ {t r a i n}, I _ {i}\\right) = \\frac {1}{2} - \\frac {1}{2 N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} S _ {v} \\left(\\mathcal {I} ^ {t r a i n}, I _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 121, + 890, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{v}(, )$ is the visual similarity scoring function, i.e., CLIP [53], $N_{i}$ number of generated images per prompt. The similarity can be regarded as the Nearest Neighbour (NN) distance from the training dataset, similar to [30].", + "bbox": [ + 496, + 171, + 893, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Consistency and typos. Given a prompt $P^{org}$ , augmented prompt $\\mathcal{P}^{aug}$ are generated using Parrot [14] for consistency and NLU-augmenter [19] for typos. Simultaneously, $N_{i}$ images $\\mathcal{I}$ and $N_{i}$ augmented images $\\mathcal{I}^{aug}$ are generated for $P^{org}$ and $\\mathcal{P}^{aug}$ , respectively. Then the cosine similarity is calculated based on Eq. 4.", + "bbox": [ + 496, + 233, + 893, + 321 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nO _ {v} = \\frac {1}{2 N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} \\sum_ {j = 1} ^ {N _ {i}} S _ {v} \\left(I _ {i}, \\mathcal {I} _ {j} ^ {\\text {a u g}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 321, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{v}(.,)$ is visual similarity scoring function; CLIP [53].", + "bbox": [ + 496, + 375, + 893, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Miscellaneous", + "text_level": 1, + "bbox": [ + 500, + 400, + 643, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Emotion. To comprehensively measure a T2I model's ability to generate images with grounded emotional tones, three evaluation metrics are proposed; AC-T2I (Sec. 4.2.3), T2I alignment (Sec. 4.2.1) and visual emotion classification accuracy. Regarding the visual emotion classifier, we train a ResNet-101 classifier based on combined datasets; FI [71] and ArtEmis [5], to ensure the model can handle diverse domains and scenarios.", + "bbox": [ + 496, + 422, + 893, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "-Fidelity. We rely on human evaluation, using Amazon Mechanical Turk (AMT) [13]. The annotators are asked to rate each image from 1-5, where 1 is the worst and 5 is the best. For a fair comparison, all the models' output images are shown in the same grid.", + "bbox": [ + 496, + 544, + 890, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "-Bias. Three bias attributes are assessed, i.e., gender, race, and age. First, the human faces are detected using ArcFace [15], and RetinaFace [16], then the facial attributes are detected using Dex [60]. Finally, the bias score is defined as the distribution skew, i.e., mean absolute deviation (MAD) [50]; Eq. 5, where the balanced case is $\\frac{1}{N_b}$ .", + "bbox": [ + 496, + 619, + 893, + 712 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM A D = \\frac {1}{N _ {b}} \\sum_ {i = 1} ^ {N _ {b}} \\left| \\hat {N} _ {b} - \\frac {1}{N _ {b}} \\right|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 713, + 890, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{b}$ is the number of protected attribute groups, e.g., 2 genders, and $\\hat{N}_b$ is the Dex output normalized count.", + "bbox": [ + 496, + 763, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 806, + 705, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluated Methods", + "text_level": 1, + "bbox": [ + 500, + 832, + 684, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We comprehensively evaluate the performance of nine recent large-scale T2I models introduced as follows. Transformer-based: minDALL-E [2] and DALL-E-Mini [1]", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20046", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f600644f1c9fd5d54447d56c5ae195d76db78c330491df8f199db7c6528a0ecd.jpg", + "image_caption": [ + "A) Counting" + ], + "image_footnote": [], + "bbox": [ + 83, + 108, + 271, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43455f46d5042b1db02485bcada0eb5570aeb222bc1295780c8d7b9bcc08916f.jpg", + "image_caption": [ + "B) Visual Text" + ], + "image_footnote": [], + "bbox": [ + 279, + 108, + 465, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/347f11fa4b653083772ce7ff5b72f9ed03d76c5cc35b276e529c20835bcb3d83.jpg", + "image_caption": [ + "C) Consistency-Typos" + ], + "image_footnote": [], + "bbox": [ + 475, + 108, + 661, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee243b636493a26852338271a499ee24a85bbf665db8c67e9caa5fd5cdbce7f8.jpg", + "image_caption": [ + "D) Spatial - Size - Color Composition" + ], + "image_footnote": [], + "bbox": [ + 671, + 102, + 859, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7848bc5ff2852a7972898201088fa5a57d5b2a2c481644970123be2bd378111.jpg", + "image_caption": [ + "E) Action Composition - Creativity" + ], + "image_footnote": [ + "Action Composition Creativity" + ], + "bbox": [ + 86, + 244, + 423, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08650a0a3d2133819e04dbf29d4a1c9938d3b185c3ef50d2ad9a4a8c1ac819e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 244, + 875, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6752be672f080f19e4405df33b9dd51b837953b0bb949c61b9c78b74448e20c.jpg", + "image_caption": [ + "Figure 5: Quantitative results for nine skills are grouped into five sub-figures based on the evaluation criteria.", + "Figure 6: Qualitative results produced by DALLE-V2. Green and red boxes, respectively, frame the success and failure cases. More qualitative results for all models are demonstrated in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 81, + 391, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are two different publicly available implementations of original DALL-E [56], which uses VQVAE [67] to encode images with grids of discrete tokens and a multimodal transformer for next token prediction. In addition, CogView2 [21] extend to a hierarchical transformer for fast super-resolution synthesis, and Paella [57] improve parallel token sampling based on MaskGIT [11]. Diffusion-based: GLIDE [45] and DALLE-V2 [55] decode images via diffusion with CLIP [53] embedding. Stable-Diffusion V1 [59] and V2 [3], (dubbed as SD-V1 and SD-V2) speed up the training of diffusion models [28] by leveraging the latent space of a powerful pretrained VQVAE [67]. Finally, Struct-Diff [25] tackles the stable-diffusion compositions limitation by manipulating the cross-attention representations based on linguistic insights to preserve the compositional semantics.", + "bbox": [ + 75, + 672, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Accuracy Skills Results", + "text_level": 1, + "bbox": [ + 500, + 672, + 717, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score. As shown in Figure 5 part A, DALLE-V2 [55] is the best in terms of precision. However, its recall is very poor, as it misses many objects. Whereas jointly considering recall and F1-score, SD-V1 [59] performs the best, despite its worst precision.", + "bbox": [ + 496, + 698, + 893, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding #1. No agreement between precision and recall. We can select the appropriate model based on the application, which metric is preferred.", + "bbox": [ + 496, + 792, + 893, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding #2. The more detailed prompt, the more accurate is counting performance. We explore three levels of prompts; 1) Vanilla prompt. The simplest form, e.g., two cups. 2) Meta-prompt. Intermediate level, e.g., describes a", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20047", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scene containing two cups. 3) Detailed. The meta-prompt is fed to GPT-3.5 to generate a detailed description including the desired objects, e.g., two cups filled with hot coffee sitting side-by-side on a wooden table. In general, we may think that the simpler and straightforward prompts may lead to better results for the counting skill. Surprisingly, as shown in Figure 8, the Recall and F1-score always increase when the detailed prompt is used.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding #3. Composition-based solution is limited. We explore Struct-Diff [25] which tackle the compositionality limitation in SD-V1 [59]. As shown in Figure 5 part A, it increases the precision compared to SD-V1 [59]. However the recall and F1-score are decreased drastically.", + "bbox": [ + 75, + 212, + 468, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "-Visual Text. We utilize two text recognition metrics, CER [42] and NED [65], which are highly correlated (95%).", + "bbox": [ + 73, + 287, + 468, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding #4. All models can not generate visual text even for the simplest case. As shown in Figure 5 part B, the best model is DALLE-V2 [55], which achieves a $75\\%$ error rate. However, the performances of all the models are far from acceptable, i.e., $10 - 20\\%$ error rate.", + "bbox": [ + 75, + 319, + 470, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding #5. Confusion between picturing and writing. The models show a good language understanding of the mentioned semantics. However, they lean toward drawing them instead of visually writing them. For instance, in Figure 6, in the visual text column and first row, the model draws the \"potted plant\" instead of writing the words. Consequently, the model prefers to draw the \"vessel\" in the second row.", + "bbox": [ + 75, + 396, + 472, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Emotion. Finding #6. All models suffer from generating emotion-grounded images. Figure 10 shows the T2I and TIT alignment scores, i.e., BLEU [48], CIDEr [68], and CLIPScore [27], which are almost equally low and far from the acceptance range among the different models. To further validate our observation, we exploit an image-to-emotion classifier trained on combined datasets as discussed in Sec. 4.3. In addition, a human evaluation experiment is conducted. In both evaluations, the classifier and the human evaluation, we simplify the problem as a binary classification, where they are asked to classify the emotion, given the generated image, as a positive or negative emotion. Both report almost $50\\%$ accuracy across the entire models, precisely the random performance, where the number of classes is only two.", + "bbox": [ + 75, + 502, + 472, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "-Fidelity. We generate three distinct images using varying seeds for each of the models. Then, the annotators evaluate them on a scale of 1-5, where 5 is the best, and 1 is the worst. The normalized scores are reported in Figure 7. The best model is SD-V2 [3], where it achieves $62.4\\%$ while the worst one is minDALL-E [2] which achieves $52.2\\%$ . However all the models are far away from the accepted threshold, i.e., $80\\%$ which corresponds to 4 on our rating system (1-5).", + "bbox": [ + 75, + 713, + 472, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Robustness and Generalization Results", + "text_level": 1, + "bbox": [ + 76, + 845, + 413, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "-Consistency and typos. We measure the alignment between the images generated from the original prompt and", + "bbox": [ + 73, + 869, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7444cd1479b88b88ef131e31b5f026e9ed3a3156aaba8beb1b1701f7ab509064.jpg", + "image_caption": [ + "Figure 7: Quantitative results for Fidelity and Bias skills." + ], + "image_footnote": [], + "bbox": [ + 578, + 90, + 803, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "paraphrased or perturbed prompts using CLIPScore [27], as discussed in Section 4.2.4.", + "bbox": [ + 498, + 262, + 890, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding #7. Models are robust against language perturbations. As shown in Figure 5 part C, all the models perform well against language perturbations and achieve between $70\\%$ to $82\\%$ alignment score. Specifically, DALLE-V2 [55] jointly achieves the best average performance for both skills, i.e., consistency and typos.", + "bbox": [ + 498, + 292, + 893, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "-Spatial, Size, and Colors composition.", + "text_level": 1, + "bbox": [ + 498, + 385, + 772, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding #8. The medium and hard levels are unattain-able. For each skill, we define three hardness levels, i.e., easy, medium, and hard, and the reported results in the whole paper are the average accuracy for the three levels. However, we report only the easy level accuracy for the spatial, size, and color composition as the entire models suffer even from the easy level. Moreover, all models fail on the medium and hard levels, where they almost got zeros. Consequently, this raises the severe limitation of the models' composition ability. As shown in Figure 5 part D, the best model, DALLE-V2 [55], achieves $28.3\\%$ , $29.9\\%$ , and $38\\%$ for spatial, size, and colors composition, respectively.", + "bbox": [ + 498, + 400, + 893, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding #9. Composition-based solution is limited. Similar to finding #3, we explore Struct-Diff [25], where it enhances the SD-V1 [59] performance by almost $3\\%$ on the easy level. However, it still fails on the challenging levels.", + "bbox": [ + 498, + 582, + 893, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Action composition Regarding the action compositionality, as illustrated in Figure 5 part E, DALLE-V2 [55] performs the best in generating compositions based on actions, according to both TIT alignment (i.e., highest CIDEr score 1.4538) and human evaluation result. Furthermore, all the scores align well with human evaluation results, confirming our metric's accuracy in evaluating this skill.", + "bbox": [ + 496, + 643, + 893, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Creativity. Since we retrieve the top-100 nearest training data with the CLIPScore[27], we obtain the average score of how the text prompt deviates from the training data, which is 0.4173. Since CLIP [53] maps images and text to shared space, the deviation score should be close if the generated image is aligned with the text. However, experimental results show that all the models fail to generate novel images, and the best model is SD-V2 [3], which achieves the highest deviation score of 0.3433. Due to creativity's very nature, it thrives on deviation. However, if the deviation becomes", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "footer", + "text": "20048", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62a513945b024de1d2ffc271924853294ab16e9e6cc987dc89cffd6b61ad2b40.jpg", + "image_caption": [ + "Vanilla Meta Detailed" + ], + "image_footnote": [], + "bbox": [ + 80, + 112, + 205, + 199 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4acdcf11f03aabb68d80952be802c31c00325c6705777d06e418698131921678.jpg", + "image_caption": [ + "Figure 8: Ablation study of the prompt details on the counting skill." + ], + "image_footnote": [], + "bbox": [ + 210, + 112, + 334, + 188 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/088fbdf09d7bf58002717f5f64258fd192fb8d32a0bd4af871bae0be225ed028.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 112, + 460, + 199 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d39e7ef10ec034bf346c2f5457570504102d77fe8b7a890fb31f0fa9b52fc685.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 98, + 647, + 207 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bd6fbde329a199459cc4b54b4d5973266f7a6c6240cadf78075b9a5393be0b6a.jpg", + "image_caption": [ + "Figure 10: Emotions results using five different metrics, in addition to, the human evaluation." + ], + "image_footnote": [], + "bbox": [ + 661, + 83, + 888, + 214 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "excessive, the resulting generation may veer toward adverse hedonic and meaningless outcomes. Therefore, we also evaluate TIT (BLEU [48], CIDEr [68]) alignment to ensure the generation keeps the semantic meaning with the text prompt. For example, as Figure 5 part E) shows, Paella achieves a relatively high deviation score but performs poorly in terms of BLEU [48] and CIDEr [68], thus it deviates too much and lose original semantic meaning of corresponding text prompt, which is aligned with human evaluation result. Therefore, both metrics are indispensable for creativity evaluation, and more than one alone is required.", + "bbox": [ + 75, + 287, + 472, + 454 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.4. Fairness and Bias Results", + "text_level": 1, + "bbox": [ + 76, + 463, + 308, + 478 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finding #10. The models are fair. As demonstrated in Figure 9, the maximum fairness score is $3.5\\%$ by Cogview 2 [21], which indicates that the difference in performance between sub-groups is negligible.", + "bbox": [ + 75, + 487, + 468, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finding #11. The models are slightly biased. In contrast to fairness, the models tend to be biased towards gender, as the average mean absolute deviation, Eq. 5, for DALLE-V2 [55], Cogview2 [21], and minDALLE [2] is $20\\%$ , as shown in Figure 7. However, SD-V1 [59] achieves the best results where the deviation is less than $8\\%$ . GLIDE [45], by design, is trained not to generate humans. Consequently, DALLE-Mini [1], and Paella [57] perform poorly regarding face generation. Therefore, GLIDE [45], DALLE-Mini [1], and Paella [57] are excluded from the bias measure.", + "bbox": [ + 75, + 547, + 470, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.5. Human Evaluation", + "text_level": 1, + "bbox": [ + 76, + 709, + 261, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To prove the effectiveness of our benchmark, we conduct a human evaluation using Amazon Mechanical Turk (AMT) [13] over $10\\%$ of our data across the entire skills. The human evaluation criteria are divided into two main groups: 1) Modular-based. The core blocks in each metric are evaluated. 2) End-to-End based. Using score-based evaluation.", + "bbox": [ + 75, + 734, + 468, + 823 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "-Modular-based. The UniDet [76] is the core block for counting, visual-text, spatial, color, and size composition. First, we ask humans to visually inspect its performance and report the true positives, false positives, and false positives. Then, we measure Person-correlation between the human", + "bbox": [ + 73, + 824, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "F1-score and our F1-score, which shows a high correlation between our calculations and human evaluation, i.e., $93\\%$ .", + "bbox": [ + 496, + 287, + 890, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similarly, we measure the detection and recognition accuracy of Textsnake [39] and SAR [34], respectively. Again, the correlation is in our favor, $98\\%$ and $96\\%$ , respectively. Moreover, regarding the emotion, the annotators are asked to binary classify the emotions, i.e., positive or negative, given only the images. The results are highly aligned with our measure, where both agree that the models generate natural images, indicating that the prompt's emotion indicator is ignored. Regarding consistency and typos, the core module is the augemer [14, 19]. To further ensure that all the generated prompts have the same meaning as the original prompt, we conduct a human study where we ask the users to rate the prompt pair (original and augmented) on a scale of 1 (not similar at all) to 5 (precisely same meaning). The results show a great alignment, i.e., $94\\%$ .", + "bbox": [ + 496, + 316, + 892, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- End-to-End based. To assess the creativity metric, we equally select 100 images per model for each hardness level and let annotators score from 1 to 5 for each generated image considering: (1) whether the generated image is creative; (2) whether the image is aligned with the given prompt. For action composition skills, annotators are requested to assign a score between 1 to 5 based on the accuracy of the generated subject and actions in response to text prompts.", + "bbox": [ + 496, + 542, + 893, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 617, + 693 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduce a comprehensive and reliable benchmark, dubbed HRS-Bench, for evaluating text-to-image (T2I) models. Our benchmark measures 13 skills, categorized into five major categories, and covers 50 applications, providing a holistic evaluation of T2I models. Through our evaluation of nine recent large-scale T2I models, we have identified areas where state-of-the-art models struggle to tackle these skills, highlighting the need for continued research and development. Our human evaluation results confirm the effectiveness and reliability of our benchmark. Further, our benchmark will help ease future T2I research and progress on improving the skills covered in this benchmark.", + "bbox": [ + 496, + 703, + 893, + 885 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20049", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dallemini. https://github.com/borisdayma/dalle-mini.2,6,9", + "[2] mindalle. https://github.com/kakaobrain/minDALL-E.2,6,8,9", + "[3] Stable-diffusion-v2. https://github.com/Stability-AI/stablediffusion.2,7,8", + "[4] Panos Achlioptas, Maks Ovsjanikov, Leonidas Guibas, and Sergey Tulyakov. Affection: Learning affective explanations for real-world visual data. arXiv preprint arXiv:2210.01946, 2022.3", + "[5] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 3, 6", + "[6] Motasem Alfarra, Juan C. Pérez, Anna Frühstück, Philip H. S. Torr, Peter Wonka, and Bernard Ghanem. On the robustness of quality measures for gans, 2022. 2", + "[7] Shane Barratt and Rishi Sharma. A note on the inception score, 2018. 2", + "[8] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.6", + "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3", + "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 2", + "[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 7", + "[12] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. arXiv preprint arXiv:2202.04053, 2022. 1, 2, 3", + "[13] Kevin Crowston. Amazon mechanical turk: A research tool for organizations and information systems scholars. In Anol Bhattacherjee and Brian Fitzgerald, editors, Shaping the Future of ICT Research. Methods and Approaches, pages 210-221, Berlin, Heidelberg, 2012. Springer Berlin Heidelberg. 6, 9", + "[14] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 4, 6, 9", + "[15] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6", + "[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 6", + "[17] Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff M Phillips, and Kai-Wei Chang. Harms of gender exclusivity and challenges in non-binary representation in language technologies. arXiv preprint arXiv:2108.12084, 2021.3", + "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2", + "[19] Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets Nick Siegel Damien Sileo Jamie Simon Chandan Singh Roman Sitelew Priyank Soni Taylor Sorensen William Soto Aman Srivastava KV Aditya Srivatsa Tony Sun,Mukund Varma T,A Tabassum,Fiona Anting Tan,Ryan Teehan, Mo Tiwari Marie Tolkiehn Athena Wang Zijian Wang,Gloria Wang,Zijie J.WangFuxuan Wei Bryan Wilie Genta Indra Winata Xinyi Wu,Witold Wydmanski,Tianbao Xie Usama YaseenM.Yee Jing Zhang and Yue Zhang.Nlaugenter: A framework for task-sensitive natural language augmentation2021.469", + "[20] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1, 2", + "[21] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2, 7, 9" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20050", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Tan M. Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation, 2021. 2", + "[23] Tan M Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 594-609. Springer, 2022. 1, 2, 3", + "[24] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214-226, 2012. 3, 6", + "[25] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 7, 8", + "[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4", + "[27] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.5, 8", + "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 7", + "[29] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7986-7994, 2018. 1", + "[30] Divyansh Jha, Kai Yi, Ivan Skorokhodov, and Mohamed Elhoseiny. Creative walk adversarial networks: Novel art generation with probabilistic random walk deviation from style norms. 3, 6", + "[31] Kenan Jiang, Xuehai He, Ruize Xu, and Xin Eric Wang. Comclip: Training-free compositional image and text matching. arXiv preprint arXiv:2211.13854, 2022. 2, 4, 5", + "[32] Kiku Johnson. Sexual orientation, gender identity, and expression affirming approach and expansive practices. 2019. 3", + "[33] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4", + "[34] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8610-8617, 2019. 5, 9", + "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1", + "[37] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 1", + "[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.4", + "[39] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In Proceedings of the European conference on computer vision (ECCV), pages 20-36, 2018. 5, 9", + "[40] Youssef Mohamed, Mohamed Abdelfattah, Shyma Alhuwaider, Feifan Li, Xiangliang Zhang, Kenneth Ward Church, and Mohamed Elhoseiny. Artelingo: A million emotion annotations of wikiart with emphasis on diversity over language and culture. arXiv preprint arXiv:2211.10780, 2022. 3", + "[41] Youssef Mohamed, Faizan Farooq Khan, Kilichbek Haydarov, and Mohamed Elhoseiny. It is okay to not be okay: Overcoming emotional bias in affective image captioning by contrastive data collection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume abs/2204.07660, 2022. 3", + "[42] Andrew Morris, Viktoria Maier, and Phil Green. From wer and ril to mer and wil: improved evaluation measures for connected speech recognition. 01 2004. 5, 8", + "[43] Deana F Morrow and Lori Messinger. Sexual orientation and gender expression in social work practice: Working with gay, lesbian, bisexual, and transgender people. Columbia University Press, 2006. 3", + "[44] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 1", + "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2, 7, 9", + "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 1", + "[47] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20051", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022. 4", + "[48] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5, 6, 8, 9", + "[49] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1, 2", + "[50] Karl Pearson. Contributions to the mathematical theory of evolution. Philosophical Transactions of the Royal Society of London. A, 185:71-110, 1894. 6", + "[51] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1, 2, 3", + "[52] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q Weinberger. On fairness and calibration. Advances in neural information processing systems, 30, 2017. 3, 6", + "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 6, 7, 8", + "[54] Micah Rajunov and A Scott Duane. Nonbinary: Memoirs of gender and identity. Columbia University Press, 2019. 3", + "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2, 7, 8, 9", + "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 7", + "[57] Dominic Rampas, Pablo Pernias, Elea Zhong, and Marc Aubreville. Fast text-conditional discrete denoising on vector-quantized latent spaces. arXiv preprint arXiv:2211.07292, 2022. 1, 2, 7, 9", + "[58] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 1", + "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 7, 8, 9", + "[60] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2", + "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 4", + "[63] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 6", + "[64] D.A. Silverstein and J.E. Farrell. The relationship between image fidelity and image quality. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 1, pages 881-884 vol.1, 1996. 3", + "[65] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1557–1562. IEEE, 2019. 5, 8", + "[66] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 5", + "[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7", + "[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5, 6, 8, 9", + "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1", + "[70] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs], 2022. 4", + "[71] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 6", + "[72] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1", + "[73] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20052", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models behave like bags-of-words, and what to do about it? arXiv e-prints, pages arXiv-2210, 2022. 2, 5", + "[74] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1", + "[75] Yanzhe Zhang, Lu Jiang, Greg Turk, and Diyi Yang. Auditing gender presentation differences in text-to-image models. arXiv preprint arXiv:2302.03675, 2023. 1, 2", + "[76] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022. 4, 9", + "[77] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 1, 2", + "[78] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 1" + ], + "bbox": [ + 78, + 90, + 470, + 457 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 475, + 922, + 493, + 936 + ], + "page_idx": 12 + }, + { + "type": "footer", + "text": "20053", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_model.json b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_model.json new file mode 100644 index 0000000000000000000000000000000000000000..50b2e09cd13bbd75cd5874c782099e91f9dc8f81 --- /dev/null +++ b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_model.json @@ -0,0 +1,3152 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.77, + 0.177 + ], + "angle": 0, + "content": "HRS-Bench: Holistic, Reliable and Scalable Benchmark for Text-to-Image Models" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.203, + 0.758, + 0.24 + ], + "angle": 0, + "content": "Eslam Mohamed Bakr1, Pengzhan Sun2**, Xiaqian Shen1**, Faizan Farooq Khan1**, Li Erran Li3, Mohamed Elhoseiny1" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.242, + 0.758, + 0.274 + ], + "angle": 0, + "content": "{eslam.abelrahman, xiaoqian.shen, faizan.khan, mohamed.elhoseiny}@kaust.edu.sa" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.277, + 0.746, + 0.292 + ], + "angle": 0, + "content": "pengzhan@comp.nus.edu.sg, lilimam@amazon.com" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.293, + 0.737, + 0.309 + ], + "angle": 0, + "content": "\\(^{1}\\)King Abdullah University of Science and Technology (KAUST)" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.31, + 0.698, + 0.327 + ], + "angle": 0, + "content": "\\(^{2}\\)National University of Singapore \\({ }^{3} \\mathrm{AWS AI},\\) Amazon" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.474, + 0.759 + ], + "angle": 0, + "content": "In recent years, Text-to-Image (T2I) models have been extensively studied, especially with the emergence of diffusion models that achieve state-of-the-art results on T2I synthesis tasks. However, existing benchmarks heavily rely on subjective human evaluation, limiting their ability to holistically assess the model's capabilities. Furthermore, there is a significant gap between efforts in developing new T2I architectures and those in evaluation. To address this, we introduce HRS-Bench, a concrete evaluation benchmark for T2I models that is Holistic, Reliable, and Scalable. Unlike existing benchmarks that focus on limited aspects, HRS-Bench measures 13 skills that can be categorized into five major categories: accuracy, robustness, generalization, fairness, and bias. In addition, HRS-Bench covers 50 scenarios, including fashion, animals, transportation, food, and clothes. We evaluate nine recent large-scale T2I models using metrics that cover a wide range of skills. A human evaluation aligned with \\(95\\%\\) of our evaluations on average was conducted to probe the effectiveness of HRS-Bench. Our experiments demonstrate that existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions. We hope that our benchmark help ease future text-to-image generation research. The code and data are available at https://eslambakr.github.io/hrsbench.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.775, + 0.21, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.8, + 0.471, + 0.877 + ], + "angle": 0, + "content": "Text-to-Image Synthesis (T2I), one of the essential multimodal tasks, witnessed remarkable progress starting from conditional GANs [58, 44, 74, 29, 78], which are shown to work on simple datasets [46, 69, 72, 36], to recently diffusion models [20, 61, 18, 10, 57, 55, 77, 45, 59], which are trained" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.361, + 0.887, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.578, + 0.896, + 0.639 + ], + "angle": 0, + "content": "Figure 1: An overview of our proposed benchmark, HRS-Bench, measures 13 skills which could be grouped into five major categories; accuracy, robustness, generalization, fairness, and bias." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.657, + 0.804, + 0.673 + ], + "angle": 0, + "content": "on large-scale datasets, e.g., LAION [63, 62]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Despite the rapid progress, the existing models face several challenges, e.g., they cannot generate complex scenes with the desired objects and relationship composition [25, 37]. Furthermore, assessing the T2I models should include more than just fidelity, e.g., the ability to compose multiple objects and generate emotionally grounded or creative images. Therefore, some recent efforts are focusing on improving the existing metrics [23] or proposing new metrics that cover new aspects, such as bias [75], compositions [37, 25, 49]. Moreover, some other works propose new benchmarks, summarized in Table 1, that assess different aspects, e.g., counting [61, 12, 51], social-bias [12], and object fidelity [23, 51]. Even with various benchmarks available, they tend to only cover a limited range of aspects while overlooking crucial evaluation criteria such as robustness," + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.887, + 0.207, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20041" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.089, + 0.895, + 0.152 + ], + "angle": 0, + "content": "Table 1: Comparison of text-to-image benchmarks in terms of: 1) The number of evaluated models. 2) The number of the covered skills. 3) The number of utilized metrics. 4) The evaluation type, whether human or metric based or both. 5) The number of prompts. 6) The prompt generation type, whether a template, human-based or both. 7) Whether there are different hardness levels are included. 8) Number of annotators contributed to the evaluation." + }, + { + "type": "table", + "bbox": [ + 0.192, + 0.161, + 0.778, + 0.263 + ], + "angle": 0, + "content": "
MethodEval TypePrompt Type
# Models# Skills# MetricsHumanAuto# PromptsTemplateHumanHardness Levels# Annotators
DrawBench [61]540✓✘200✘✓✘25
DALL-EVAL [12]453✓✘7330✓✘✘6
HE-T2I [51]230✓✘90✘✓✓20
TISE [23]735✘✓N/A✘✓✘N/A
HRS-Bench (Ours)91317✓✓45000✓✓✓1000
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.234, + 0.302 + ], + "angle": 0, + "content": "fairness, and creativity." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.307, + 0.474, + 0.744 + ], + "angle": 0, + "content": "To bridge this gap, we propose our Holistic, Reliable, and Scalable benchmark, dubbed HRS-Bench. In contrast to existing benchmarks, we measure a wide range of different generative capabilities, precisely 13 skills which can be grouped into five major categories, as demonstrated in Figure 1; accuracy, robustness, generalization, fairness, and bias. Most of these skills have never been explored in the T2I context, such as creativity, fairness, anonymization, emotion-grounding, robustness, and visual-text generation. Even the other previously explored skills were studied from a limited perspective, for instance, DALL-EVAL [12] studied the social bias by generating limited template-based prompts; only 145 prompts. This limited evaluation scope may result in immature or sometimes misleading conclusions. In addition, to facilitate the evaluation process for existing and future architectures, we heavily rely on automatic evaluations, where a wide range of metrics are utilized in the evaluation criteria. Moreover, HRS-Bench covers 50 scenarios, e.g., fashion, animals, transportation, and food. Figure 2 demonstrates the top 15 applications and their object distribution. We evaluate nine recent large-scale T2I models, i.e., Stable-Diffusion V1 [59] and V2 [3], DALL-E 2 [55], GLIDE [45], CogView-V2 [21], Paella [57], minDALL-E [2], DALL-E-Mini [1], and Struct-Diff [25]. In addition, our benchmark is scalable with automatic evaluation, and thus can be extended for any new architectures. To probe the effectiveness of our HRS-Bench, we conduct a human assessment that aligns well with our evaluations by \\(95\\%\\) on average. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.75, + 0.471, + 0.794 + ], + "angle": 0, + "content": "- We develop a Holistic, Reliable, and scalable T2I benchmark called HRS-Bench, depicted in Figure 1, which assess 13 skills covering 50 scenarios." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.796, + 0.472, + 0.841 + ], + "angle": 0, + "content": "- Propose a new T2I alignment metric, called AC-T2I, which overcomes the composition limitations of existing large Vision-Language Models (VLMs) [31, 73]." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.841, + 0.471, + 0.871 + ], + "angle": 0, + "content": "- Nine T2I models are assessed based on our benchmark, including commercial and open-sourced ones." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "- We verify the effectiveness of our HRS-Bench metric by conducting a human evaluation for \\(10\\%\\) of our data" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.75, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.283, + 0.877, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.503, + 0.896, + 0.548 + ], + "angle": 0, + "content": "Figure 2: Pie chart demonstrates the wide range of covered scenarios by our proposed benchmark, termed HRS-Bench, and their object distribution." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.561, + 0.8, + 0.576 + ], + "angle": 0, + "content": "per skill that shows excellent alignment." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.576, + 0.895, + 0.65 + ], + "angle": 0, + "content": "- Driven by these holistic evaluations, several conclusions and findings are discussed. For instance, existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.842, + 0.683 + ], + "angle": 0, + "content": "2. Revisiting Text-to-Image Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Recently, there have been rapid efforts toward designing better image-generation models [20, 61, 18, 10, 57, 55, 77, 45, 59]. However, most existing metrics suffer from several limitations that make them unreliable [23, 7, 22, 6, 75, 49, 66, 73]. Table 1 summarizes the existing T2I benchmarks. -DrawBench. Imagen [61] proposes DrawBench to evaluate the T2I models from other aspects along with the image quality. Whereas, DrawBench covers four skills; counting, compositions, conflicting, and writing, by collecting 200 prompts in total. Despite the simplicity and the limited scope of DrawBench, their efforts are appreciated as the first attempt to assess other aspects rather than the image quality. -DALL-EVAL. [12] It proposes a toolkit called PAINTSKILLS, which assesses three simple visual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20042" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "reasoning skills; object recognition, object counting, and spatial relation understanding, alongside two social bias skills; gender and racial bias. To facilitate the automatic evaluation, they built a unity simulator including limited objects to collect the datasets, approximately 80 object classes. In contrast, we cover more than 700 object classes. DETR [9] is utilized for visual reasoning skills evaluation after being fine-tuned on the synthetic dataset collected from the simulator. However, DALL-EVAL evaluates only four models using only 7k prompts and six annotators. Therefore, there is a need for a more comprehensive benchmark that can take into account a broader range of models, prompts, and annotators to provide a more thorough evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.363 + ], + "angle": 0, + "content": "- HE-T2I. [51] It proposes 32 possible aspects to benchmark T2I models. However, only three are evaluated; counting, shapes, and faces, and the rest are left unexplored. The three aspects are evaluated using human evaluations, where twenty annotators have contributed to the evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.472, + 0.44 + ], + "angle": 0, + "content": "-TISE. [23] It introduces a bag of metrics to evaluate T2I models from three aspects; positional alignment, counting, and fidelity. In addition, three fidelity metrics are introduced, i.e., \\(IS^{*}\\), \\(OIS\\), and \\(OFID\\), and two alignment metrics PA for positional alignment, and CA for counting alignment." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.202, + 0.469 + ], + "angle": 0, + "content": "3. HRS-Bench" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.48, + 0.47, + 0.511 + ], + "angle": 0, + "content": "In this section, we first dissect the skills definition, then demonstrate the prompts collection pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.521, + 0.158, + 0.535 + ], + "angle": 0, + "content": "3.1. Skills" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.199, + 0.559 + ], + "angle": 0, + "content": "3.1.1 Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.568, + 0.47, + 0.629 + ], + "angle": 0, + "content": "- Counting. A reliable T2I model should be able to ground the details in the prompt. One form of these details is objects binding with a specific frequency, e.g., \"four cars are parked around five benches in a park\"." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.47, + 0.72 + ], + "angle": 0, + "content": "- Visual Text. Another essential aspect of assessing the model is generating high-quality text in wild scenes, e.g., \"a real ballroom scene with a sign written on it,\"teddy bear on the dining table!\"\". The importance of such skill comes from intervening in many scenarios, e.g., education applications, preparing illustration content, and designing billboards." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.72, + 0.472, + 0.765 + ], + "angle": 0, + "content": "-Emotion. We measure to what extent the model can generate emotion-grounded images [5, 4, 41, 40], e.g., \"a rainy scene about cake, which makes us feel excitement.\"" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.766, + 0.47, + 0.797 + ], + "angle": 0, + "content": "-Fidelity. Image fidelity indicates how accurately an image represents the underlying source distribution [64]." + }, + { + "type": "list", + "bbox": [ + 0.075, + 0.568, + 0.472, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.212, + 0.83 + ], + "angle": 0, + "content": "3.1.2 Robustness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.871 + ], + "angle": 0, + "content": "To assess the T2I model's robustness, we cover two types of transformations; invariance and equivariance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Invariance. Two skills are introduced to measure the invariance robustness: consistency and typos." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "- Consistency. We measure the sensitivity of different T2I models towards prompt variations while keeping the same meaning and semantics, i.e., paraphrasing. For instance, generated images from these prompts, \"a woman is standing in front of a mirror, carefully selecting the perfect handbag\" and \"in front of a mirror, a woman is selecting the perfect handbag for the day\" should hold the same semantics." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.2, + 0.894, + 0.246 + ], + "angle": 0, + "content": "-Typos. Two natural perturbations are utilized to assess the models against possible sensibleness noise that users could cause during inference, i.e., typos and wrong capitalization." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.247, + 0.894, + 0.308 + ], + "angle": 0, + "content": "Equivariance. Three different compositions are explored for the equivariance robustness. Specifically, we study three types of compositions, i.e., spatial, attribute-specific, and action compositions." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.31, + 0.895, + 0.401 + ], + "angle": 0, + "content": "-Spatial composition. In contrast to the counting skill which only measures the models' ability to compose multiple objects into a coherent scene, spatial composition additionally measures their ability to ground the detailed spatial relationship instructions mentioned in the input prompt, e.g., \"a person and a dog in the middle of a cat and a chair\"." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.403, + 0.894, + 0.493 + ], + "angle": 0, + "content": "- Attribute-specific composition. Two types of attributes are controlled to study the attribute binding ability, i.e., colors and size attributes. For instance, \"an orange cat, a red dog, and a blue chair\" and \"a banana which is smaller than a person and bigger than a car\", for colors and size attribute binding, respectively." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.496, + 0.894, + 0.542 + ], + "angle": 0, + "content": "-Action composition. It incorporates different subjects that doing different actions, e.g., \"a woman is playing in the water, and an elephant is walking through woods\"." + }, + { + "type": "list", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.659, + 0.587 + ], + "angle": 0, + "content": "3.1.3 Generalization" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.6, + 0.895, + 0.706 + ], + "angle": 0, + "content": "- Creativity. In this skill, models aim to generate images that represent not only the textual description but are also imaginative and novel. The creativity skill can be regarded as an out-of-distribution generation [30]. Accordingly, we devised innovative text prompts that are conceptually plausible but may need to be more readily available in standard training data sources, detailed later in Section 3.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.737, + 0.614, + 0.752 + ], + "angle": 0, + "content": "3.1.4 Fairness" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We define fairness as the performance disparity among different sub-groups [24, 52]. A fair model should achieve the same performance on an arbitrary metric since no correlation exists between the metric and the protected attribute. Two attributes have been studied, i.e., gender and style. Following [12], gender refers to sex [32, 54] not the gender identity [43, 17]. We use two gender categories; male and female. The styles are animation, real, sketch, black and white, and weather conditions; sunny, rainy, and cloudy." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20043" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.092, + 0.163, + 0.105 + ], + "angle": 0, + "content": "3.1.5 Bias" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.471, + 0.207 + ], + "angle": 0, + "content": "We assess the spurious correlation of the model towards predefined attributes, i.e., gender, race, and age. Using agnostic prompts towards a specific attribute, e.g., gender, the model should produce balanced generations of different classes of this attribute. For instance, the gender agnostic prompt could be, \"Two persons are working on their laptops\"." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.217, + 0.262, + 0.233 + ], + "angle": 0, + "content": "3.2. Prompts Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.376 + ], + "angle": 0, + "content": "For each skill, we collect 3k prompts. To ensure our benchmark is holistic enough, we split the prompts equally into three hardness levels, i.e., easy, medium, and hard. In addition, half of the prompts are human-based, and the other half is template-based, depicted in Figure 3. We filter human prompts manually from existing datasets [70, 33, 31]. We use the foundation model GPT-3.5 [47], text-davinci-003* to facilitate prompts generation, which will be abbreviated as GPT-3.5 later for convenience." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.377, + 0.471, + 0.483 + ], + "angle": 0, + "content": "-Fidelity. The human prompts are sifted from [70]. Whereas, the template-based prompts are created by defining a template that describes a styled scene that contains some objects, as shown in Figure 3. Then, we create the meta-prompt by sampling the styles from pre-defined styles and the objects from LVIS dataset [26]. Finally, GPT-3.5 [47] is utilized to generate the final prompts." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.483, + 0.471, + 0.603 + ], + "angle": 0, + "content": "- Consistency and Typos. Consequently, the fidelity prompts are fed to Parrot [14] and NLU augmenter [19] to produce augmented prompts for consistency and typos, respectively. For consistency, we differentiate between the three hardness levels based on the similarity between the fidelity prompt and the augmented prompts using RoBERTa [38]. For typos, the number of introduced typos controls the three hardness levels, i.e., 1-2, 3-4, and 5-6, respectively." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.604, + 0.471, + 0.664 + ], + "angle": 0, + "content": "-Counting. Given a meta-prompt, GPT-3.5 generate a realistic scenario that contains N objects. To generate the meta-prompts, we randomly samples the number of objects and the objects classes from LVIS dataset [26]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.665, + 0.471, + 0.725 + ], + "angle": 0, + "content": "-Visual Text. We utilize GPT-3.5 to generate short descriptions which fit on a sign in a crowded scene. Then, we control the hardness levels by the text length and the surrounding scene complexity." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.725, + 0.471, + 0.771 + ], + "angle": 0, + "content": "- Emotion. We sample random objects from LVIS [26], then append an emotion indicator word forming the meta-prompt. Finally, GPT-3.5 is utilized to generate the final prompts." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.771, + 0.471, + 0.876 + ], + "angle": 0, + "content": "-Creativity. We craft text prompts that are challenging yet still within the realm of imagination. For easy level, we obtain subject, object, and relationship from Visual Genome [33] and obtain triplets by different combinations. Then sift through all the combinations from triplets extracted from LAION [62] dataset, and retain only the uncommon triplets. For the medium level, we fed the uncommon" + }, + { + "type": "list", + "bbox": [ + 0.075, + 0.377, + 0.471, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.197 + ], + "angle": 0, + "content": "triplet to GPT-3.5 with the instruction: “Describe subject, relation and object in an imaginative way that will never be seen in the real world” and manually filter the undesirable sentences. Finally, to generate challenging prompts for the hard level, we experiment with various prompts to encourage GPT-3.5 to generate counterproductive sentences, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.895, + 0.438 + ], + "angle": 0, + "content": "- Compositionality. We study three composition types, i.e., spatial, attribute-binding, and actions. The spatial prompts are collected using a pre-defined template, where a wide range of relations is utilized, e.g., \"on the right of\", \"above\", and \"between.\" For attribute-binding, two attributes are exploited, i.e., colors and size. For each hardness level, the number of objects' compositions increased, ranging from 2 to 4. For the action-level compositional generation, we design prompts with multiple combinations of actions starting from ComCLIP [31]. We combine two sentences from ComCLIP [31] for the easy level. Then, for medium and hard levels, we randomly choose one sentence and feed 'Extend text' to let the subject have at least three actions.' and 'Extend text' with other subjects doing other actions' into GPT-3.5, respectively, to obtain the final prompt. Detailed examples are demonstrated in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.894, + 0.529 + ], + "angle": 0, + "content": "-Bias. Random objects are sampled from LVIS datasets [26], combined with a pre-defined template, creating a metaprompt. Then, the meta-prompt is fed to GPT-3.5 to produce the final prompt, as depicted in Figure 3. To ensure the prompts are agnostic towards the protected attributes, i.e., gender, race, and age, we manually validate them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.605 + ], + "angle": 0, + "content": "-Fairness. We adapt the bias prompts. For gender fairness, we replace gender-agnostic words, such as a person, with gender-specific words, such as man and woman. Whereas for style fairness, a style indicator is appended to the beginning of the bias prompt, as shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.618, + 0.783, + 0.634 + ], + "angle": 0, + "content": "4. Evaluation for our Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.688 + ], + "angle": 0, + "content": "As shown in Figure 4, we categorize the skills based on the evaluation criteria, one-to-many mapping. Thus the same skill may be assigned to several metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.725, + 0.712 + ], + "angle": 0, + "content": "4.1. Detection-Based Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.765 + ], + "angle": 0, + "content": "We utilize UniDet [76] for counting, spatial and attribute compositions because it supports a wide range of object classes, i.e., exceeding 700." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.826 + ], + "angle": 0, + "content": "- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score, where Precision assesses the accuracy of additional objects, and Recall assesses the accuracy of missing objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "-Spatial Compositions. Using a simple geometry module, we use the predicted bounding boxes to validate whether the spatial relation is grounded correctly. For instance, given the prompt \"A cat above a car\", the predicted bounding boxes will be \\(\\{x_{min}^{1},y_{min}^{1},x_{max}^{1},y_{max}^{1}\\}\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.361, + 0.901 + ], + "angle": 0, + "content": "* https://platform.openai.com/docs/models/gpt-3-5" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20044" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.891, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.897, + 0.334 + ], + "angle": 0, + "content": "Figure 3: Our prompt generation pipeline. First, we create a meta-prompt, which is a template-based prompt (in blue). Then, we sample the skill-related attributes (in orange). Finally, we generate the final prompt using ChatGPT (in green)." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.356, + 0.462, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.48, + 0.472, + 0.526 + ], + "angle": 0, + "content": "Figure 4: On the left is our evaluation taxonomy. On the right, we demonstrate our metric Augmented Captioner-based T2I alignment metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.621 + ], + "angle": 0, + "content": "and \\(\\{x_{min}^2, y_{min}^2, x_{max}^2, y_{max}^2\\}\\) for the cat and the car, respectively, and the grounded spatial relation is above. Then, our geometry module will assess whether the spatial relation, i.e., above, is grounded correctly based on the following condition: \\((y_{min}^1 < y_{min}^2)\\) or \\((y_{max}^1 < y_{max}^2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.623, + 0.47, + 0.715 + ], + "angle": 0, + "content": "- Attributes Compositions. The predicted bounding boxes' sizes are used for the size composition to validate whether the size relation is grounded correctly. Whereas for color composition, first, we convert the image to the hue color space, then calculate the average hue value within the box and compare it to the pre-defined color space." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.717, + 0.471, + 0.779 + ], + "angle": 0, + "content": "-Visual Text. We adopt Textsnake [39] and SAR [34] for text detection and recognition, respectively. The recognition accuracy is measured by the Character Error Rate (CER) [42] and the Normalized Edit Distance (NED) [65]." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.798, + 0.312, + 0.815 + ], + "angle": 0, + "content": "4.2. Alignment-Based Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Three alignment paradigms are explored; Text-to-Image (T2I) (Sec. 4.2.1), Text-to-Image-to-Text (TIT) (Sec. 4.2.3), and Image-to-Image I2I (Sec. 4.2.4). In addition, we introduce our novel Augmented Captioner-based T2I Alignment metric, termed AC-T2I (Sec. 4.2.2)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.358, + 0.659, + 0.374 + ], + "angle": 0, + "content": "4.2.1 T2I Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.895, + 0.535 + ], + "angle": 0, + "content": "One possible solution to assess the T2I model's grounding ability is measuring the text and image correlation, e.g., CLIPScore [27, 53]. While CLIP is widely used, its effectiveness has been repeatedly questioned, as it is not sensitive to fine-grained text-image alignment and fails to understand compositions [66, 73]. For instance, [73] shows that CLIP [53] can not distinguish between \"the horse is eating the grass\" and \"the grass is eating the horse\". This motivates us to propose our novel augmented captioner-based T2I alignment metric, termed AC-T2I, depicted in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.555, + 0.741, + 0.57 + ], + "angle": 0, + "content": "4.2.2 AC-T2I Alignment Metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.715 + ], + "angle": 0, + "content": "We propose a new T2I alignment metric, called AC-T2I, which overcomes the compositional relationship's limitations of existing large Vision-Language Models (VLMs) [31, 73], by utilizing the n-grams based metric, e.g., CIDEr [68] and BLEU [48]. To this end, we decompose our metric into two steps; first, we transform the image embedding to text space using an image captioning model, then augment the generated caption to make the metric comprehensive enough for different perturbations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.895, + 0.87 + ], + "angle": 0, + "content": "-Reformatting T2I as TIT. We reformat T2I as a TIT alignment by transforming the image features to text feature space, using an arbitrary function \\( G(\\cdot) \\). The function \\( G(\\mathcal{I}) \\) could be interpreted as an image captioner, e.g., BLIP2 [35]. As shown in Figure 4, given a text prompt \\( P^{org} \\), \\( N_{i} \\) images \\( \\mathcal{I} = \\{I_k\\}_{k=1}^{N_i} \\) are generated, which are fed to an image captioner \\( G(\\mathcal{I}) \\) producing \\( N_{c_i} \\) captions \\( \\mathcal{C} = \\{C_k\\}_{k=1}^{N_{c_i}} \\), where \\( N_{c_i} \\) is the number of generated captions per image. Finally, the \\( N_{c_i} \\) captions are automatically evaluated using CIDEr [68] and BLEU [48] against the input prompt \\( P^{org} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- Comprehensive TIT. Instead of considering only the prompt \\( P^{org} \\) as the GT caption, \\( N_{t} \\) augmented prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20045" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.473, + 0.198 + ], + "angle": 0, + "content": "\\(\\mathcal{P}^{aug} = \\{P_k^{aug}\\}_{k=1}^{N_t}\\) are generated using GPT-3.5, to measure the similarities comprehensively. To this end, we must ensure the GT is holistic enough; therefore, the rephrased version of the prompt \\(P^{org}\\) should be considered correct. Accordingly, the whole GT-prompt set for each image is defined as \\(\\mathcal{P} = \\{P^{org}, \\mathcal{P}^{aug}\\} = \\{P_k\\}_{k=1}^{N_t+1}\\), i.e., one original prompt plus \\(N_t\\) augmented prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.47, + 0.243 + ], + "angle": 0, + "content": "Finally, for each prompt \\( P^{org} \\) we calculate the alignment score for each generated prompt-caption pair and select the highest score as the final alignment score, Eq. 1." + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.243, + 0.47, + 0.282 + ], + "angle": 0, + "content": "\\[\nO _ {t} = \\frac {1}{N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} \\max _ {1 \\leq j \\leq N _ {c _ {i}}, 1 \\leq k \\leq N _ {t} + 1} S _ {t} \\left(C _ {i, j}, P _ {k}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.47, + 0.324 + ], + "angle": 0, + "content": "where \\(S_{t}(.)\\) is the text similarity scoring function, e.g., CIDEr [68], BLEU [48]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.342, + 0.238, + 0.358 + ], + "angle": 0, + "content": "4.2.3 TIT Alignment" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.366, + 0.473, + 0.47 + ], + "angle": 0, + "content": "-Emotions. We explore a visual emotion classifier as illustrated in Sec. 4.3. Moreover, we apply our proposed metric, AC-T2I (Eq. 1), to avoid the aforementioned CLIP limitations, detailed in Section 4.2.1 and Section 4.2.2. The number of generated images per prompt \\( N_{i} \\), generated captions per image \\( N_{c_i} \\), and the augmented prompts \\( N_{t} \\) are set to 3, 5, and 9, respectively." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.471, + 0.47, + 0.576 + ], + "angle": 0, + "content": "- Creativity. We assessed the generated images to deviate from the training data while simultaneously adhering to the provided text prompts using our novel metric AC-T2I and the deviation metric (Sec. 4.2.4). We set \\( N_{i} \\) and \\( N_{c_i} \\) to 3 and 5, respectively. Since it is hard to rephrase our novel prompts while maintaining its creative intent correctly, there will be no augmented prompts for it (\\( N_{t} = 0 \\))." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.577, + 0.472, + 0.621 + ], + "angle": 0, + "content": "-Gender and styles fairness and action compositions. The fairness score is defined as the disparities in subgroups' performance [24, 52], Eq. 2" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.622, + 0.47, + 0.676 + ], + "angle": 0, + "content": "\\[\nF a i r n e s s _ {s c o r e} = \\frac {1}{N _ {s} C _ {2}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = i + 1} ^ {N _ {s}} \\frac {1 0 0 \\times | A (i) - A (j) |}{\\operatorname* {m a x} (A (i) , A (j))}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.77 + ], + "angle": 0, + "content": "where \\(\\frac{100}{N_sC_2\\times max(A(i),A(j))}\\) is a normalization factor, \\(N_{s}\\) is the number of sub-groups, e.g., two for the gender and \\(A\\) is the accuracy measure, e.g., AC-T2I or CLIP scores. The less is, the better for Fairness score. Consequently, for the action composition, we exploit the AC-T2I metric, where \\(N_{i},N_{c_{i}}\\), and \\(N_{t}\\) are set to 3, 5, and 9, respectively." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.787, + 0.231, + 0.802 + ], + "angle": 0, + "content": "4.2.4 I2I Alignment" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.81, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- Creativity. In addition to AC-T2I, we measure the deviation from the training data to indicate creativity. Accessing large models' training data is challenging, however, most of them are trained on LAION [63]. Accordingly, we use LAION image-text retrieval tools [8] to fetch training data, which search among the dataset using CLIP [53] and a KNN index" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "to seek top-100 nearest images, denoted as \\(\\mathcal{I}^{train}\\) for each prompt. The deviation score is calculated based on Eq. 3." + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\triangle \\left(\\mathcal {I} ^ {t r a i n}, I _ {i}\\right) = \\frac {1}{2} - \\frac {1}{2 N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} S _ {v} \\left(\\mathcal {I} ^ {t r a i n}, I _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.172, + 0.894, + 0.233 + ], + "angle": 0, + "content": "where \\( S_{v}(, ) \\) is the visual similarity scoring function, i.e., CLIP [53], \\( N_{i} \\) number of generated images per prompt. The similarity can be regarded as the Nearest Neighbour (NN) distance from the training dataset, similar to [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.894, + 0.323 + ], + "angle": 0, + "content": "- Consistency and typos. Given a prompt \\( P^{org} \\), augmented prompt \\( \\mathcal{P}^{aug} \\) are generated using Parrot [14] for consistency and NLU-augmenter [19] for typos. Simultaneously, \\( N_{i} \\) images \\( \\mathcal{I} \\) and \\( N_{i} \\) augmented images \\( \\mathcal{I}^{aug} \\) are generated for \\( P^{org} \\) and \\( \\mathcal{P}^{aug} \\), respectively. Then the cosine similarity is calculated based on Eq. 4." + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.323, + 0.892, + 0.365 + ], + "angle": 0, + "content": "\\[\nO _ {v} = \\frac {1}{2 N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} \\sum_ {j = 1} ^ {N _ {i}} S _ {v} \\left(I _ {i}, \\mathcal {I} _ {j} ^ {\\text {a u g}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.894, + 0.393 + ], + "angle": 0, + "content": "where \\(S_{v}(.,)\\) is visual similarity scoring function; CLIP [53]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.401, + 0.645, + 0.415 + ], + "angle": 0, + "content": "4.3. Miscellaneous" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.894, + 0.544 + ], + "angle": 0, + "content": "- Emotion. To comprehensively measure a T2I model's ability to generate images with grounded emotional tones, three evaluation metrics are proposed; AC-T2I (Sec. 4.2.3), T2I alignment (Sec. 4.2.1) and visual emotion classification accuracy. Regarding the visual emotion classifier, we train a ResNet-101 classifier based on combined datasets; FI [71] and ArtEmis [5], to ensure the model can handle diverse domains and scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.621 + ], + "angle": 0, + "content": "-Fidelity. We rely on human evaluation, using Amazon Mechanical Turk (AMT) [13]. The annotators are asked to rate each image from 1-5, where 1 is the worst and 5 is the best. For a fair comparison, all the models' output images are shown in the same grid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.894, + 0.713 + ], + "angle": 0, + "content": "-Bias. Three bias attributes are assessed, i.e., gender, race, and age. First, the human faces are detected using ArcFace [15], and RetinaFace [16], then the facial attributes are detected using Dex [60]. Finally, the bias score is defined as the distribution skew, i.e., mean absolute deviation (MAD) [50]; Eq. 5, where the balanced case is \\(\\frac{1}{N_b}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.714, + 0.892, + 0.754 + ], + "angle": 0, + "content": "\\[\nM A D = \\frac {1}{N _ {b}} \\sum_ {i = 1} ^ {N _ {b}} \\left| \\hat {N} _ {b} - \\frac {1}{N _ {b}} \\right|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.794 + ], + "angle": 0, + "content": "where \\(N_{b}\\) is the number of protected attribute groups, e.g., 2 genders, and \\(\\hat{N}_b\\) is the Dex output normalized count." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.706, + 0.825 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.686, + 0.847 + ], + "angle": 0, + "content": "5.1. Evaluated Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We comprehensively evaluate the performance of nine recent large-scale T2I models introduced as follows. Transformer-based: minDALL-E [2] and DALL-E-Mini [1]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20046" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.097, + 0.212, + 0.108 + ], + "angle": 0, + "content": "A) Counting" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.109, + 0.272, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.096, + 0.412, + 0.107 + ], + "angle": 0, + "content": "B) Visual Text" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.109, + 0.466, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.097, + 0.628, + 0.108 + ], + "angle": 0, + "content": "C) Consistency-Typos" + }, + { + "type": "image", + "bbox": [ + 0.477, + 0.109, + 0.663, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.091, + 0.86, + 0.102 + ], + "angle": 0, + "content": "D) Spatial - Size - Color Composition" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.103, + 0.86, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.217, + 0.574, + 0.229 + ], + "angle": 0, + "content": "E) Action Composition - Creativity" + }, + { + "type": "image_footnote", + "bbox": [ + 0.088, + 0.233, + 0.199, + 0.242 + ], + "angle": 0, + "content": "Action Composition Creativity" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.246, + 0.424, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.246, + 0.877, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.364, + 0.842, + 0.38 + ], + "angle": 0, + "content": "Figure 5: Quantitative results for nine skills are grouped into five sub-figures based on the evaluation criteria." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.392, + 0.891, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.616, + 0.895, + 0.647 + ], + "angle": 0, + "content": "Figure 6: Qualitative results produced by DALLE-V2. Green and red boxes, respectively, frame the success and failure cases. More qualitative results for all models are demonstrated in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.473, + 0.901 + ], + "angle": 0, + "content": "are two different publicly available implementations of original DALL-E [56], which uses VQVAE [67] to encode images with grids of discrete tokens and a multimodal transformer for next token prediction. In addition, CogView2 [21] extend to a hierarchical transformer for fast super-resolution synthesis, and Paella [57] improve parallel token sampling based on MaskGIT [11]. Diffusion-based: GLIDE [45] and DALLE-V2 [55] decode images via diffusion with CLIP [53] embedding. Stable-Diffusion V1 [59] and V2 [3], (dubbed as SD-V1 and SD-V2) speed up the training of diffusion models [28] by leveraging the latent space of a powerful pretrained VQVAE [67]. Finally, Struct-Diff [25] tackles the stable-diffusion compositions limitation by manipulating the cross-attention representations based on linguistic insights to preserve the compositional semantics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.718, + 0.689 + ], + "angle": 0, + "content": "5.2. Accuracy Skills Results" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.699, + 0.895, + 0.79 + ], + "angle": 0, + "content": "- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score. As shown in Figure 5 part A, DALLE-V2 [55] is the best in terms of precision. However, its recall is very poor, as it misses many objects. Whereas jointly considering recall and F1-score, SD-V1 [59] performs the best, despite its worst precision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Finding #1. No agreement between precision and recall. We can select the appropriate model based on the application, which metric is preferred." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Finding #2. The more detailed prompt, the more accurate is counting performance. We explore three levels of prompts; 1) Vanilla prompt. The simplest form, e.g., two cups. 2) Meta-prompt. Intermediate level, e.g., describes a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20047" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "scene containing two cups. 3) Detailed. The meta-prompt is fed to GPT-3.5 to generate a detailed description including the desired objects, e.g., two cups filled with hot coffee sitting side-by-side on a wooden table. In general, we may think that the simpler and straightforward prompts may lead to better results for the counting skill. Surprisingly, as shown in Figure 8, the Recall and F1-score always increase when the detailed prompt is used." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.288 + ], + "angle": 0, + "content": "Finding #3. Composition-based solution is limited. We explore Struct-Diff [25] which tackle the compositionality limitation in SD-V1 [59]. As shown in Figure 5 part A, it increases the precision compared to SD-V1 [59]. However the recall and F1-score are decreased drastically." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.289, + 0.469, + 0.32 + ], + "angle": 0, + "content": "-Visual Text. We utilize two text recognition metrics, CER [42] and NED [65], which are highly correlated (95%)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.471, + 0.395 + ], + "angle": 0, + "content": "Finding #4. All models can not generate visual text even for the simplest case. As shown in Figure 5 part B, the best model is DALLE-V2 [55], which achieves a \\(75\\%\\) error rate. However, the performances of all the models are far from acceptable, i.e., \\(10 - 20\\%\\) error rate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Finding #5. Confusion between picturing and writing. The models show a good language understanding of the mentioned semantics. However, they lean toward drawing them instead of visually writing them. For instance, in Figure 6, in the visual text column and first row, the model draws the \"potted plant\" instead of writing the words. Consequently, the model prefers to draw the \"vessel\" in the second row." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.473, + 0.714 + ], + "angle": 0, + "content": "- Emotion. Finding #6. All models suffer from generating emotion-grounded images. Figure 10 shows the T2I and TIT alignment scores, i.e., BLEU [48], CIDEr [68], and CLIPScore [27], which are almost equally low and far from the acceptance range among the different models. To further validate our observation, we exploit an image-to-emotion classifier trained on combined datasets as discussed in Sec. 4.3. In addition, a human evaluation experiment is conducted. In both evaluations, the classifier and the human evaluation, we simplify the problem as a binary classification, where they are asked to classify the emotion, given the generated image, as a positive or negative emotion. Both report almost \\(50\\%\\) accuracy across the entire models, precisely the random performance, where the number of classes is only two." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.473, + 0.837 + ], + "angle": 0, + "content": "-Fidelity. We generate three distinct images using varying seeds for each of the models. Then, the annotators evaluate them on a scale of 1-5, where 5 is the best, and 1 is the worst. The normalized scores are reported in Figure 7. The best model is SD-V2 [3], where it achieves \\(62.4\\%\\) while the worst one is minDALL-E [2] which achieves \\(52.2\\%\\). However all the models are far away from the accepted threshold, i.e., \\(80\\%\\) which corresponds to 4 on our rating system (1-5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.414, + 0.862 + ], + "angle": 0, + "content": "5.3. Robustness and Generalization Results" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "-Consistency and typos. We measure the alignment between the images generated from the original prompt and" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.091, + 0.805, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.221, + 0.884, + 0.237 + ], + "angle": 0, + "content": "Figure 7: Quantitative results for Fidelity and Bias skills." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.292 + ], + "angle": 0, + "content": "paraphrased or perturbed prompts using CLIPScore [27], as discussed in Section 4.2.4." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.294, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Finding #7. Models are robust against language perturbations. As shown in Figure 5 part C, all the models perform well against language perturbations and achieve between \\(70\\%\\) to \\(82\\%\\) alignment score. Specifically, DALLE-V2 [55] jointly achieves the best average performance for both skills, i.e., consistency and typos." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.386, + 0.774, + 0.401 + ], + "angle": 0, + "content": "-Spatial, Size, and Colors composition." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.401, + 0.895, + 0.582 + ], + "angle": 0, + "content": "Finding #8. The medium and hard levels are unattain-able. For each skill, we define three hardness levels, i.e., easy, medium, and hard, and the reported results in the whole paper are the average accuracy for the three levels. However, we report only the easy level accuracy for the spatial, size, and color composition as the entire models suffer even from the easy level. Moreover, all models fail on the medium and hard levels, where they almost got zeros. Consequently, this raises the severe limitation of the models' composition ability. As shown in Figure 5 part D, the best model, DALLE-V2 [55], achieves \\(28.3\\%\\), \\(29.9\\%\\), and \\(38\\%\\) for spatial, size, and colors composition, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.583, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Finding #9. Composition-based solution is limited. Similar to finding #3, we explore Struct-Diff [25], where it enhances the SD-V1 [59] performance by almost \\(3\\%\\) on the easy level. However, it still fails on the challenging levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.75 + ], + "angle": 0, + "content": "- Action composition Regarding the action compositionality, as illustrated in Figure 5 part E, DALLE-V2 [55] performs the best in generating compositions based on actions, according to both TIT alignment (i.e., highest CIDEr score 1.4538) and human evaluation result. Furthermore, all the scores align well with human evaluation results, confirming our metric's accuracy in evaluating this skill." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "- Creativity. Since we retrieve the top-100 nearest training data with the CLIPScore[27], we obtain the average score of how the text prompt deviates from the training data, which is 0.4173. Since CLIP [53] maps images and text to shared space, the deviation score should be close if the generated image is aligned with the text. However, experimental results show that all the models fail to generate novel images, and the best model is SD-V2 [3], which achieves the highest deviation score of 0.3433. Due to creativity's very nature, it thrives on deviation. However, if the deviation becomes" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20048" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.101, + 0.177, + 0.111 + ], + "angle": 0, + "content": "Vanilla Meta Detailed" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.113, + 0.207, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.113, + 0.335, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.113, + 0.462, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.099, + 0.648, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.084, + 0.889, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.216, + 0.893, + 0.259 + ], + "angle": 0, + "content": "Figure 10: Emotions results using five different metrics, in addition to, the human evaluation." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.219, + 0.473, + 0.251 + ], + "angle": 0, + "content": "Figure 8: Ablation study of the prompt details on the counting skill." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.455 + ], + "angle": 0, + "content": "excessive, the resulting generation may veer toward adverse hedonic and meaningless outcomes. Therefore, we also evaluate TIT (BLEU [48], CIDEr [68]) alignment to ensure the generation keeps the semantic meaning with the text prompt. For example, as Figure 5 part E) shows, Paella achieves a relatively high deviation score but performs poorly in terms of BLEU [48] and CIDEr [68], thus it deviates too much and lose original semantic meaning of corresponding text prompt, which is aligned with human evaluation result. Therefore, both metrics are indispensable for creativity evaluation, and more than one alone is required." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.464, + 0.31, + 0.479 + ], + "angle": 0, + "content": "5.4. Fairness and Bias Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.548 + ], + "angle": 0, + "content": "Finding #10. The models are fair. As demonstrated in Figure 9, the maximum fairness score is \\(3.5\\%\\) by Cogview 2 [21], which indicates that the difference in performance between sub-groups is negligible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.472, + 0.701 + ], + "angle": 0, + "content": "Finding #11. The models are slightly biased. In contrast to fairness, the models tend to be biased towards gender, as the average mean absolute deviation, Eq. 5, for DALLE-V2 [55], Cogview2 [21], and minDALLE [2] is \\(20\\%\\), as shown in Figure 7. However, SD-V1 [59] achieves the best results where the deviation is less than \\(8\\%\\). GLIDE [45], by design, is trained not to generate humans. Consequently, DALLE-Mini [1], and Paella [57] perform poorly regarding face generation. Therefore, GLIDE [45], DALLE-Mini [1], and Paella [57] are excluded from the bias measure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.262, + 0.726 + ], + "angle": 0, + "content": "5.5. Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.824 + ], + "angle": 0, + "content": "To prove the effectiveness of our benchmark, we conduct a human evaluation using Amazon Mechanical Turk (AMT) [13] over \\(10\\%\\) of our data across the entire skills. The human evaluation criteria are divided into two main groups: 1) Modular-based. The core blocks in each metric are evaluated. 2) End-to-End based. Using score-based evaluation." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.825, + 0.471, + 0.9 + ], + "angle": 0, + "content": "-Modular-based. The UniDet [76] is the core block for counting, visual-text, spatial, color, and size composition. First, we ask humans to visually inspect its performance and report the true positives, false positives, and false positives. Then, we measure Person-correlation between the human" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.317 + ], + "angle": 0, + "content": "F1-score and our F1-score, which shows a high correlation between our calculations and human evaluation, i.e., \\(93\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.893, + 0.544 + ], + "angle": 0, + "content": "Similarly, we measure the detection and recognition accuracy of Textsnake [39] and SAR [34], respectively. Again, the correlation is in our favor, \\(98\\%\\) and \\(96\\%\\), respectively. Moreover, regarding the emotion, the annotators are asked to binary classify the emotions, i.e., positive or negative, given only the images. The results are highly aligned with our measure, where both agree that the models generate natural images, indicating that the prompt's emotion indicator is ignored. Regarding consistency and typos, the core module is the augemer [14, 19]. To further ensure that all the generated prompts have the same meaning as the original prompt, we conduct a human study where we ask the users to rate the prompt pair (original and augmented) on a scale of 1 (not similar at all) to 5 (precisely same meaning). The results show a great alignment, i.e., \\(94\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.894, + 0.667 + ], + "angle": 0, + "content": "- End-to-End based. To assess the creativity metric, we equally select 100 images per model for each hardness level and let annotators score from 1 to 5 for each generated image considering: (1) whether the generated image is creative; (2) whether the image is aligned with the given prompt. For action composition skills, annotators are requested to assign a score between 1 to 5 based on the accuracy of the generated subject and actions in response to text prompts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.619, + 0.694 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.886 + ], + "angle": 0, + "content": "We introduce a comprehensive and reliable benchmark, dubbed HRS-Bench, for evaluating text-to-image (T2I) models. Our benchmark measures 13 skills, categorized into five major categories, and covers 50 applications, providing a holistic evaluation of T2I models. Through our evaluation of nine recent large-scale T2I models, we have identified areas where state-of-the-art models struggle to tackle these skills, highlighting the need for continued research and development. Our human evaluation results confirm the effectiveness and reliability of our benchmark. Further, our benchmark will help ease future T2I research and progress on improving the skills covered in this benchmark." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20049" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.143 + ], + "angle": 0, + "content": "[1] Dallemini. https://github.com/borisdayma/dalle-mini.2,6,9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.472, + 0.172 + ], + "angle": 0, + "content": "[2] mindalle. https://github.com/kakaobrain/minDALL-E.2,6,8,9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.175, + 0.471, + 0.202 + ], + "angle": 0, + "content": "[3] Stable-diffusion-v2. https://github.com/Stability-AI/stablediffusion.2,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.471, + 0.258 + ], + "angle": 0, + "content": "[4] Panos Achlioptas, Maks Ovsjanikov, Leonidas Guibas, and Sergey Tulyakov. Affection: Learning affective explanations for real-world visual data. arXiv preprint arXiv:2210.01946, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.471, + 0.33 + ], + "angle": 0, + "content": "[5] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[6] Motasem Alfarra, Juan C. Pérez, Anna Frühstück, Philip H. S. Torr, Peter Wonka, and Bernard Ghanem. On the robustness of quality measures for gans, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.471, + 0.402 + ], + "angle": 0, + "content": "[7] Shane Barratt and Rishi Sharma. A note on the inception score, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.404, + 0.471, + 0.445 + ], + "angle": 0, + "content": "[8] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.448, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.471, + 0.601 + ], + "angle": 0, + "content": "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.471, + 0.658 + ], + "angle": 0, + "content": "[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.714 + ], + "angle": 0, + "content": "[12] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. arXiv preprint arXiv:2202.04053, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.471, + 0.798 + ], + "angle": 0, + "content": "[13] Kevin Crowston. Amazon mechanical turk: A research tool for organizations and information systems scholars. In Anol Bhattacherjee and Brian Fitzgerald, editors, Shaping the Future of ICT Research. Methods and Approaches, pages 210-221, Berlin, Heidelberg, 2012. Springer Berlin Heidelberg. 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[14] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[15] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.138, + 0.894, + 0.207 + ], + "angle": 0, + "content": "[17] Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff M Phillips, and Kai-Wei Chang. Harms of gender exclusivity and challenges in non-binary representation in language technologies. arXiv preprint arXiv:2108.12084, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.211, + 0.892, + 0.253 + ], + "angle": 0, + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.256, + 0.895, + 0.769 + ], + "angle": 0, + "content": "[19] Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets Nick Siegel Damien Sileo Jamie Simon Chandan Singh Roman Sitelew Priyank Soni Taylor Sorensen William Soto Aman Srivastava KV Aditya Srivatsa Tony Sun,Mukund Varma T,A Tabassum,Fiona Anting Tan,Ryan Teehan, Mo Tiwari Marie Tolkiehn Athena Wang Zijian Wang,Gloria Wang,Zijie J.WangFuxuan Wei Bryan Wilie Genta Indra Winata Xinyi Wu,Witold Wydmanski,Tianbao Xie Usama YaseenM.Yee Jing Zhang and Yue Zhang.Nlaugenter: A framework for task-sensitive natural language augmentation2021.469" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[20] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.895, + 0.9 + ], + "angle": 0, + "content": "[21] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2, 7, 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20050" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "[22] Tan M. Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[23] Tan M Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 594-609. Springer, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[24] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214-226, 2012. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[25] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.376 + ], + "angle": 0, + "content": "[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[27] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.476 + ], + "angle": 0, + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[29] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7986-7994, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[30] Divyansh Jha, Kai Yi, Ivan Skorokhodov, and Mohamed Elhoseiny. Creative walk adversarial networks: Novel art generation with probabilistic random walk deviation from style norms. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.472, + 0.646 + ], + "angle": 0, + "content": "[31] Kenan Jiang, Xuehai He, Ruize Xu, and Xin Eric Wang. Comclip: Training-free compositional image and text matching. arXiv preprint arXiv:2211.13854, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[32] Kiku Johnson. Sexual orientation, gender identity, and expression affirming approach and expansive practices. 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.472, + 0.772 + ], + "angle": 0, + "content": "[33] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[34] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8610-8617, 2019. 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.895, + 0.258 + ], + "angle": 0, + "content": "[37] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.895, + 0.327 + ], + "angle": 0, + "content": "[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.895, + 0.398 + ], + "angle": 0, + "content": "[39] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In Proceedings of the European conference on computer vision (ECCV), pages 20-36, 2018. 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.895, + 0.48 + ], + "angle": 0, + "content": "[40] Youssef Mohamed, Mohamed Abdelfattah, Shyma Alhuwaider, Feifan Li, Xiangliang Zhang, Kenneth Ward Church, and Mohamed Elhoseiny. Artelingo: A million emotion annotations of wikiart with emphasis on diversity over language and culture. arXiv preprint arXiv:2211.10780, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.894, + 0.565 + ], + "angle": 0, + "content": "[41] Youssef Mohamed, Faizan Farooq Khan, Kilichbek Haydarov, and Mohamed Elhoseiny. It is okay to not be okay: Overcoming emotional bias in affective image captioning by contrastive data collection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume abs/2204.07660, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.893, + 0.607 + ], + "angle": 0, + "content": "[42] Andrew Morris, Viktoria Maier, and Phil Green. From wer and ril to mer and wil: improved evaluation measures for connected speech recognition. 01 2004. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[43] Deana F Morrow and Lori Messinger. Sexual orientation and gender expression in social work practice: Working with gay, lesbian, bisexual, and transgender people. Columbia University Press, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.894, + 0.734 + ], + "angle": 0, + "content": "[44] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.893, + 0.802 + ], + "angle": 0, + "content": "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.893, + 0.859 + ], + "angle": 0, + "content": "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[47] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20051" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[48] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[49] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.472, + 0.304 + ], + "angle": 0, + "content": "[50] Karl Pearson. Contributions to the mathematical theory of evolution. Philosophical Transactions of the Royal Society of London. A, 185:71-110, 1894. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[51] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.472, + 0.418 + ], + "angle": 0, + "content": "[52] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q Weinberger. On fairness and calibration. Advances in neural information processing systems, 30, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.533 + ], + "angle": 0, + "content": "[54] Micah Rajunov and A Scott Duane. Nonbinary: Memoirs of gender and identity. Columbia University Press, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[57] Dominic Rampas, Pablo Pernias, Elea Zhong, and Marc Aubreville. Fast text-conditional discrete denoising on vector-quantized latent spaces. arXiv preprint arXiv:2211.07292, 2022. 1, 2, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[58] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[60] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.894, + 0.258 + ], + "angle": 0, + "content": "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[63] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[64] D.A. Silverstein and J.E. Farrell. The relationship between image fidelity and image quality. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 1, pages 881-884 vol.1, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.894, + 0.468 + ], + "angle": 0, + "content": "[65] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1557–1562. IEEE, 2019. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.894, + 0.551 + ], + "angle": 0, + "content": "[66] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.552, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.594, + 0.894, + 0.662 + ], + "angle": 0, + "content": "[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.894, + 0.704 + ], + "angle": 0, + "content": "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.705, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[70] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs], 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[71] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[72] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[73] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20052" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "models behave like bags-of-words, and what to do about it? arXiv e-prints, pages arXiv-2210, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[74] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[75] Yanzhe Zhang, Lu Jiang, Greg Turk, and Diyi Yang. Auditing gender presentation differences in text-to-image models. arXiv preprint arXiv:2302.03675, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[76] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022. 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[77] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.472, + 0.458 + ], + "angle": 0, + "content": "[78] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.477, + 0.924, + 0.495, + 0.937 + ], + "angle": 0, + "content": "13" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "20053" + } + ] +] \ No newline at end of file diff --git a/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_origin.pdf b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a8e507737ba68b425a8e3050e46c53cfdd4e61fd --- /dev/null +++ b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/c90d4d8a-8b52-4c5e-b753-56dc5d34d841_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d14b758fa87721796117e01f239a69cba87aff8ca926ab5367161d6ecd8aafc +size 7155511 diff --git a/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/full.md b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a35568525b4aaa43ecd35e663b048f29397a5efe --- /dev/null +++ b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/full.md @@ -0,0 +1,409 @@ +# HRS-Bench: Holistic, Reliable and Scalable Benchmark for Text-to-Image Models + +Eslam Mohamed Bakr1, Pengzhan Sun2**, Xiaqian Shen1**, Faizan Farooq Khan1**, Li Erran Li3, Mohamed Elhoseiny1 + +{eslam.abelrahman, xiaoqian.shen, faizan.khan, mohamed.elhoseiny}@kaust.edu.sa + +pengzhan@comp.nus.edu.sg, lilimam@amazon.com + +$^{1}$ King Abdullah University of Science and Technology (KAUST) + +$^{2}$ National University of Singapore ${ }^{3} \mathrm{AWS AI},$ Amazon + +# Abstract + +In recent years, Text-to-Image (T2I) models have been extensively studied, especially with the emergence of diffusion models that achieve state-of-the-art results on T2I synthesis tasks. However, existing benchmarks heavily rely on subjective human evaluation, limiting their ability to holistically assess the model's capabilities. Furthermore, there is a significant gap between efforts in developing new T2I architectures and those in evaluation. To address this, we introduce HRS-Bench, a concrete evaluation benchmark for T2I models that is Holistic, Reliable, and Scalable. Unlike existing benchmarks that focus on limited aspects, HRS-Bench measures 13 skills that can be categorized into five major categories: accuracy, robustness, generalization, fairness, and bias. In addition, HRS-Bench covers 50 scenarios, including fashion, animals, transportation, food, and clothes. We evaluate nine recent large-scale T2I models using metrics that cover a wide range of skills. A human evaluation aligned with $95\%$ of our evaluations on average was conducted to probe the effectiveness of HRS-Bench. Our experiments demonstrate that existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions. We hope that our benchmark help ease future text-to-image generation research. The code and data are available at https://eslambakr.github.io/hrsbench.github.io/. + +# 1. Introduction + +Text-to-Image Synthesis (T2I), one of the essential multimodal tasks, witnessed remarkable progress starting from conditional GANs [58, 44, 74, 29, 78], which are shown to work on simple datasets [46, 69, 72, 36], to recently diffusion models [20, 61, 18, 10, 57, 55, 77, 45, 59], which are trained + +![](images/37206a980a5ac74d8b6214facb1f6d29db7672bbca5b6e6612114c3e25b821a9.jpg) +Figure 1: An overview of our proposed benchmark, HRS-Bench, measures 13 skills which could be grouped into five major categories; accuracy, robustness, generalization, fairness, and bias. + +on large-scale datasets, e.g., LAION [63, 62]. + +Despite the rapid progress, the existing models face several challenges, e.g., they cannot generate complex scenes with the desired objects and relationship composition [25, 37]. Furthermore, assessing the T2I models should include more than just fidelity, e.g., the ability to compose multiple objects and generate emotionally grounded or creative images. Therefore, some recent efforts are focusing on improving the existing metrics [23] or proposing new metrics that cover new aspects, such as bias [75], compositions [37, 25, 49]. Moreover, some other works propose new benchmarks, summarized in Table 1, that assess different aspects, e.g., counting [61, 12, 51], social-bias [12], and object fidelity [23, 51]. Even with various benchmarks available, they tend to only cover a limited range of aspects while overlooking crucial evaluation criteria such as robustness, + +Table 1: Comparison of text-to-image benchmarks in terms of: 1) The number of evaluated models. 2) The number of the covered skills. 3) The number of utilized metrics. 4) The evaluation type, whether human or metric based or both. 5) The number of prompts. 6) The prompt generation type, whether a template, human-based or both. 7) Whether there are different hardness levels are included. 8) Number of annotators contributed to the evaluation. + +
MethodEval TypePrompt Type
# Models# Skills# MetricsHumanAuto# PromptsTemplateHumanHardness Levels# Annotators
DrawBench [61]540✓✘200✘✓✘25
DALL-EVAL [12]453✓✘7330✓✘✘6
HE-T2I [51]230✓✘90✘✓✓20
TISE [23]735✘✓N/A✘✓✘N/A
HRS-Bench (Ours)91317✓✓45000✓✓✓1000
+ +fairness, and creativity. + +To bridge this gap, we propose our Holistic, Reliable, and Scalable benchmark, dubbed HRS-Bench. In contrast to existing benchmarks, we measure a wide range of different generative capabilities, precisely 13 skills which can be grouped into five major categories, as demonstrated in Figure 1; accuracy, robustness, generalization, fairness, and bias. Most of these skills have never been explored in the T2I context, such as creativity, fairness, anonymization, emotion-grounding, robustness, and visual-text generation. Even the other previously explored skills were studied from a limited perspective, for instance, DALL-EVAL [12] studied the social bias by generating limited template-based prompts; only 145 prompts. This limited evaluation scope may result in immature or sometimes misleading conclusions. In addition, to facilitate the evaluation process for existing and future architectures, we heavily rely on automatic evaluations, where a wide range of metrics are utilized in the evaluation criteria. Moreover, HRS-Bench covers 50 scenarios, e.g., fashion, animals, transportation, and food. Figure 2 demonstrates the top 15 applications and their object distribution. We evaluate nine recent large-scale T2I models, i.e., Stable-Diffusion V1 [59] and V2 [3], DALL-E 2 [55], GLIDE [45], CogView-V2 [21], Paella [57], minDALL-E [2], DALL-E-Mini [1], and Struct-Diff [25]. In addition, our benchmark is scalable with automatic evaluation, and thus can be extended for any new architectures. To probe the effectiveness of our HRS-Bench, we conduct a human assessment that aligns well with our evaluations by $95\%$ on average. Our contributions can be summarized as follows: + +- We develop a Holistic, Reliable, and scalable T2I benchmark called HRS-Bench, depicted in Figure 1, which assess 13 skills covering 50 scenarios. +- Propose a new T2I alignment metric, called AC-T2I, which overcomes the composition limitations of existing large Vision-Language Models (VLMs) [31, 73]. +- Nine T2I models are assessed based on our benchmark, including commercial and open-sourced ones. +- We verify the effectiveness of our HRS-Bench metric by conducting a human evaluation for $10\%$ of our data + +![](images/30a35cd2424c0cae31944cb3ddb282b3bea5fc9a86dc3c4e67008097ef79208a.jpg) +Figure 2: Pie chart demonstrates the wide range of covered scenarios by our proposed benchmark, termed HRS-Bench, and their object distribution. + +per skill that shows excellent alignment. + +- Driven by these holistic evaluations, several conclusions and findings are discussed. For instance, existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions. + +# 2. Revisiting Text-to-Image Benchmarks + +Recently, there have been rapid efforts toward designing better image-generation models [20, 61, 18, 10, 57, 55, 77, 45, 59]. However, most existing metrics suffer from several limitations that make them unreliable [23, 7, 22, 6, 75, 49, 66, 73]. Table 1 summarizes the existing T2I benchmarks. -DrawBench. Imagen [61] proposes DrawBench to evaluate the T2I models from other aspects along with the image quality. Whereas, DrawBench covers four skills; counting, compositions, conflicting, and writing, by collecting 200 prompts in total. Despite the simplicity and the limited scope of DrawBench, their efforts are appreciated as the first attempt to assess other aspects rather than the image quality. -DALL-EVAL. [12] It proposes a toolkit called PAINTSKILLS, which assesses three simple visual + +reasoning skills; object recognition, object counting, and spatial relation understanding, alongside two social bias skills; gender and racial bias. To facilitate the automatic evaluation, they built a unity simulator including limited objects to collect the datasets, approximately 80 object classes. In contrast, we cover more than 700 object classes. DETR [9] is utilized for visual reasoning skills evaluation after being fine-tuned on the synthetic dataset collected from the simulator. However, DALL-EVAL evaluates only four models using only 7k prompts and six annotators. Therefore, there is a need for a more comprehensive benchmark that can take into account a broader range of models, prompts, and annotators to provide a more thorough evaluation. +- HE-T2I. [51] It proposes 32 possible aspects to benchmark T2I models. However, only three are evaluated; counting, shapes, and faces, and the rest are left unexplored. The three aspects are evaluated using human evaluations, where twenty annotators have contributed to the evaluation. +-TISE. [23] It introduces a bag of metrics to evaluate T2I models from three aspects; positional alignment, counting, and fidelity. In addition, three fidelity metrics are introduced, i.e., $IS^{*}$ , $OIS$ , and $OFID$ , and two alignment metrics PA for positional alignment, and CA for counting alignment. + +# 3. HRS-Bench + +In this section, we first dissect the skills definition, then demonstrate the prompts collection pipeline. + +# 3.1. Skills + +# 3.1.1 Accuracy + +- Counting. A reliable T2I model should be able to ground the details in the prompt. One form of these details is objects binding with a specific frequency, e.g., "four cars are parked around five benches in a park". +- Visual Text. Another essential aspect of assessing the model is generating high-quality text in wild scenes, e.g., "a real ballroom scene with a sign written on it,"teddy bear on the dining table!"". The importance of such skill comes from intervening in many scenarios, e.g., education applications, preparing illustration content, and designing billboards. +-Emotion. We measure to what extent the model can generate emotion-grounded images [5, 4, 41, 40], e.g., "a rainy scene about cake, which makes us feel excitement." +-Fidelity. Image fidelity indicates how accurately an image represents the underlying source distribution [64]. + +# 3.1.2 Robustness + +To assess the T2I model's robustness, we cover two types of transformations; invariance and equivariance. + +Invariance. Two skills are introduced to measure the invariance robustness: consistency and typos. + +- Consistency. We measure the sensitivity of different T2I models towards prompt variations while keeping the same meaning and semantics, i.e., paraphrasing. For instance, generated images from these prompts, "a woman is standing in front of a mirror, carefully selecting the perfect handbag" and "in front of a mirror, a woman is selecting the perfect handbag for the day" should hold the same semantics. +-Typos. Two natural perturbations are utilized to assess the models against possible sensibleness noise that users could cause during inference, i.e., typos and wrong capitalization. +Equivariance. Three different compositions are explored for the equivariance robustness. Specifically, we study three types of compositions, i.e., spatial, attribute-specific, and action compositions. +-Spatial composition. In contrast to the counting skill which only measures the models' ability to compose multiple objects into a coherent scene, spatial composition additionally measures their ability to ground the detailed spatial relationship instructions mentioned in the input prompt, e.g., "a person and a dog in the middle of a cat and a chair". +- Attribute-specific composition. Two types of attributes are controlled to study the attribute binding ability, i.e., colors and size attributes. For instance, "an orange cat, a red dog, and a blue chair" and "a banana which is smaller than a person and bigger than a car", for colors and size attribute binding, respectively. +-Action composition. It incorporates different subjects that doing different actions, e.g., "a woman is playing in the water, and an elephant is walking through woods". + +# 3.1.3 Generalization + +- Creativity. In this skill, models aim to generate images that represent not only the textual description but are also imaginative and novel. The creativity skill can be regarded as an out-of-distribution generation [30]. Accordingly, we devised innovative text prompts that are conceptually plausible but may need to be more readily available in standard training data sources, detailed later in Section 3.2. + +# 3.1.4 Fairness + +We define fairness as the performance disparity among different sub-groups [24, 52]. A fair model should achieve the same performance on an arbitrary metric since no correlation exists between the metric and the protected attribute. Two attributes have been studied, i.e., gender and style. Following [12], gender refers to sex [32, 54] not the gender identity [43, 17]. We use two gender categories; male and female. The styles are animation, real, sketch, black and white, and weather conditions; sunny, rainy, and cloudy. + +# 3.1.5 Bias + +We assess the spurious correlation of the model towards predefined attributes, i.e., gender, race, and age. Using agnostic prompts towards a specific attribute, e.g., gender, the model should produce balanced generations of different classes of this attribute. For instance, the gender agnostic prompt could be, "Two persons are working on their laptops". + +# 3.2. Prompts Collection + +For each skill, we collect 3k prompts. To ensure our benchmark is holistic enough, we split the prompts equally into three hardness levels, i.e., easy, medium, and hard. In addition, half of the prompts are human-based, and the other half is template-based, depicted in Figure 3. We filter human prompts manually from existing datasets [70, 33, 31]. We use the foundation model GPT-3.5 [47], text-davinci-003* to facilitate prompts generation, which will be abbreviated as GPT-3.5 later for convenience. + +-Fidelity. The human prompts are sifted from [70]. Whereas, the template-based prompts are created by defining a template that describes a styled scene that contains some objects, as shown in Figure 3. Then, we create the meta-prompt by sampling the styles from pre-defined styles and the objects from LVIS dataset [26]. Finally, GPT-3.5 [47] is utilized to generate the final prompts. +- Consistency and Typos. Consequently, the fidelity prompts are fed to Parrot [14] and NLU augmenter [19] to produce augmented prompts for consistency and typos, respectively. For consistency, we differentiate between the three hardness levels based on the similarity between the fidelity prompt and the augmented prompts using RoBERTa [38]. For typos, the number of introduced typos controls the three hardness levels, i.e., 1-2, 3-4, and 5-6, respectively. +-Counting. Given a meta-prompt, GPT-3.5 generate a realistic scenario that contains N objects. To generate the meta-prompts, we randomly samples the number of objects and the objects classes from LVIS dataset [26]. +-Visual Text. We utilize GPT-3.5 to generate short descriptions which fit on a sign in a crowded scene. Then, we control the hardness levels by the text length and the surrounding scene complexity. +- Emotion. We sample random objects from LVIS [26], then append an emotion indicator word forming the meta-prompt. Finally, GPT-3.5 is utilized to generate the final prompts. +-Creativity. We craft text prompts that are challenging yet still within the realm of imagination. For easy level, we obtain subject, object, and relationship from Visual Genome [33] and obtain triplets by different combinations. Then sift through all the combinations from triplets extracted from LAION [62] dataset, and retain only the uncommon triplets. For the medium level, we fed the uncommon + +triplet to GPT-3.5 with the instruction: “Describe subject, relation and object in an imaginative way that will never be seen in the real world” and manually filter the undesirable sentences. Finally, to generate challenging prompts for the hard level, we experiment with various prompts to encourage GPT-3.5 to generate counterproductive sentences, as shown in Figure 3. + +- Compositionality. We study three composition types, i.e., spatial, attribute-binding, and actions. The spatial prompts are collected using a pre-defined template, where a wide range of relations is utilized, e.g., "on the right of", "above", and "between." For attribute-binding, two attributes are exploited, i.e., colors and size. For each hardness level, the number of objects' compositions increased, ranging from 2 to 4. For the action-level compositional generation, we design prompts with multiple combinations of actions starting from ComCLIP [31]. We combine two sentences from ComCLIP [31] for the easy level. Then, for medium and hard levels, we randomly choose one sentence and feed 'Extend text' to let the subject have at least three actions.' and 'Extend text' with other subjects doing other actions' into GPT-3.5, respectively, to obtain the final prompt. Detailed examples are demonstrated in Figure 3. + +-Bias. Random objects are sampled from LVIS datasets [26], combined with a pre-defined template, creating a metaprompt. Then, the meta-prompt is fed to GPT-3.5 to produce the final prompt, as depicted in Figure 3. To ensure the prompts are agnostic towards the protected attributes, i.e., gender, race, and age, we manually validate them. + +-Fairness. We adapt the bias prompts. For gender fairness, we replace gender-agnostic words, such as a person, with gender-specific words, such as man and woman. Whereas for style fairness, a style indicator is appended to the beginning of the bias prompt, as shown in Figure 3. + +# 4. Evaluation for our Benchmark + +As shown in Figure 4, we categorize the skills based on the evaluation criteria, one-to-many mapping. Thus the same skill may be assigned to several metrics. + +# 4.1. Detection-Based Metrics + +We utilize UniDet [76] for counting, spatial and attribute compositions because it supports a wide range of object classes, i.e., exceeding 700. + +- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score, where Precision assesses the accuracy of additional objects, and Recall assesses the accuracy of missing objects. + +-Spatial Compositions. Using a simple geometry module, we use the predicted bounding boxes to validate whether the spatial relation is grounded correctly. For instance, given the prompt "A cat above a car", the predicted bounding boxes will be $\{x_{min}^{1},y_{min}^{1},x_{max}^{1},y_{max}^{1}\}$ + +![](images/cb17387a2bfdfd71fcf5785194a3985e82f28d58c32ca44c17a7111a360efce7.jpg) +Figure 3: Our prompt generation pipeline. First, we create a meta-prompt, which is a template-based prompt (in blue). Then, we sample the skill-related attributes (in orange). Finally, we generate the final prompt using ChatGPT (in green). + +![](images/6d4adacee38b78d062ba6564271775671d0491adc52c21221508376ef6e281b1.jpg) +Figure 4: On the left is our evaluation taxonomy. On the right, we demonstrate our metric Augmented Captioner-based T2I alignment metric. + +and $\{x_{min}^2, y_{min}^2, x_{max}^2, y_{max}^2\}$ for the cat and the car, respectively, and the grounded spatial relation is above. Then, our geometry module will assess whether the spatial relation, i.e., above, is grounded correctly based on the following condition: $(y_{min}^1 < y_{min}^2)$ or $(y_{max}^1 < y_{max}^2)$ . + +- Attributes Compositions. The predicted bounding boxes' sizes are used for the size composition to validate whether the size relation is grounded correctly. Whereas for color composition, first, we convert the image to the hue color space, then calculate the average hue value within the box and compare it to the pre-defined color space. + +-Visual Text. We adopt Textsnake [39] and SAR [34] for text detection and recognition, respectively. The recognition accuracy is measured by the Character Error Rate (CER) [42] and the Normalized Edit Distance (NED) [65]. + +# 4.2. Alignment-Based Metrics + +Three alignment paradigms are explored; Text-to-Image (T2I) (Sec. 4.2.1), Text-to-Image-to-Text (TIT) (Sec. 4.2.3), and Image-to-Image I2I (Sec. 4.2.4). In addition, we introduce our novel Augmented Captioner-based T2I Alignment metric, termed AC-T2I (Sec. 4.2.2). + +# 4.2.1 T2I Alignment + +One possible solution to assess the T2I model's grounding ability is measuring the text and image correlation, e.g., CLIPScore [27, 53]. While CLIP is widely used, its effectiveness has been repeatedly questioned, as it is not sensitive to fine-grained text-image alignment and fails to understand compositions [66, 73]. For instance, [73] shows that CLIP [53] can not distinguish between "the horse is eating the grass" and "the grass is eating the horse". This motivates us to propose our novel augmented captioner-based T2I alignment metric, termed AC-T2I, depicted in Figure 4. + +# 4.2.2 AC-T2I Alignment Metric. + +We propose a new T2I alignment metric, called AC-T2I, which overcomes the compositional relationship's limitations of existing large Vision-Language Models (VLMs) [31, 73], by utilizing the n-grams based metric, e.g., CIDEr [68] and BLEU [48]. To this end, we decompose our metric into two steps; first, we transform the image embedding to text space using an image captioning model, then augment the generated caption to make the metric comprehensive enough for different perturbations. + +-Reformatting T2I as TIT. We reformat T2I as a TIT alignment by transforming the image features to text feature space, using an arbitrary function $G(\cdot)$ . The function $G(\mathcal{I})$ could be interpreted as an image captioner, e.g., BLIP2 [35]. As shown in Figure 4, given a text prompt $P^{org}$ , $N_{i}$ images $\mathcal{I} = \{I_k\}_{k=1}^{N_i}$ are generated, which are fed to an image captioner $G(\mathcal{I})$ producing $N_{c_i}$ captions $\mathcal{C} = \{C_k\}_{k=1}^{N_{c_i}}$ , where $N_{c_i}$ is the number of generated captions per image. Finally, the $N_{c_i}$ captions are automatically evaluated using CIDEr [68] and BLEU [48] against the input prompt $P^{org}$ . + +- Comprehensive TIT. Instead of considering only the prompt $P^{org}$ as the GT caption, $N_{t}$ augmented prompts + +$\mathcal{P}^{aug} = \{P_k^{aug}\}_{k=1}^{N_t}$ are generated using GPT-3.5, to measure the similarities comprehensively. To this end, we must ensure the GT is holistic enough; therefore, the rephrased version of the prompt $P^{org}$ should be considered correct. Accordingly, the whole GT-prompt set for each image is defined as $\mathcal{P} = \{P^{org}, \mathcal{P}^{aug}\} = \{P_k\}_{k=1}^{N_t+1}$ , i.e., one original prompt plus $N_t$ augmented prompts. + +Finally, for each prompt $P^{org}$ we calculate the alignment score for each generated prompt-caption pair and select the highest score as the final alignment score, Eq. 1. + +$$ +O _ {t} = \frac {1}{N _ {i}} \sum_ {i = 1} ^ {N _ {i}} \max _ {1 \leq j \leq N _ {c _ {i}}, 1 \leq k \leq N _ {t} + 1} S _ {t} \left(C _ {i, j}, P _ {k}\right), \tag {1} +$$ + +where $S_{t}(.)$ is the text similarity scoring function, e.g., CIDEr [68], BLEU [48]. + +# 4.2.3 TIT Alignment + +-Emotions. We explore a visual emotion classifier as illustrated in Sec. 4.3. Moreover, we apply our proposed metric, AC-T2I (Eq. 1), to avoid the aforementioned CLIP limitations, detailed in Section 4.2.1 and Section 4.2.2. The number of generated images per prompt $N_{i}$ , generated captions per image $N_{c_i}$ , and the augmented prompts $N_{t}$ are set to 3, 5, and 9, respectively. + +- Creativity. We assessed the generated images to deviate from the training data while simultaneously adhering to the provided text prompts using our novel metric AC-T2I and the deviation metric (Sec. 4.2.4). We set $N_{i}$ and $N_{c_i}$ to 3 and 5, respectively. Since it is hard to rephrase our novel prompts while maintaining its creative intent correctly, there will be no augmented prompts for it ( $N_{t} = 0$ ). + +-Gender and styles fairness and action compositions. The fairness score is defined as the disparities in subgroups' performance [24, 52], Eq. 2 + +$$ +F a i r n e s s _ {s c o r e} = \frac {1}{N _ {s} C _ {2}} \sum_ {i = 1} ^ {N _ {s}} \sum_ {j = i + 1} ^ {N _ {s}} \frac {1 0 0 \times | A (i) - A (j) |}{\operatorname* {m a x} (A (i) , A (j))}, \tag {2} +$$ + +where $\frac{100}{N_sC_2\times max(A(i),A(j))}$ is a normalization factor, $N_{s}$ is the number of sub-groups, e.g., two for the gender and $A$ is the accuracy measure, e.g., AC-T2I or CLIP scores. The less is, the better for Fairness score. Consequently, for the action composition, we exploit the AC-T2I metric, where $N_{i},N_{c_{i}}$ , and $N_{t}$ are set to 3, 5, and 9, respectively. + +# 4.2.4 I2I Alignment + +- Creativity. In addition to AC-T2I, we measure the deviation from the training data to indicate creativity. Accessing large models' training data is challenging, however, most of them are trained on LAION [63]. Accordingly, we use LAION image-text retrieval tools [8] to fetch training data, which search among the dataset using CLIP [53] and a KNN index + +to seek top-100 nearest images, denoted as $\mathcal{I}^{train}$ for each prompt. The deviation score is calculated based on Eq. 3. + +$$ +\triangle \left(\mathcal {I} ^ {t r a i n}, I _ {i}\right) = \frac {1}{2} - \frac {1}{2 N _ {i}} \sum_ {i = 1} ^ {N _ {i}} S _ {v} \left(\mathcal {I} ^ {t r a i n}, I _ {i}\right), \tag {3} +$$ + +where $S_{v}(, )$ is the visual similarity scoring function, i.e., CLIP [53], $N_{i}$ number of generated images per prompt. The similarity can be regarded as the Nearest Neighbour (NN) distance from the training dataset, similar to [30]. + +- Consistency and typos. Given a prompt $P^{org}$ , augmented prompt $\mathcal{P}^{aug}$ are generated using Parrot [14] for consistency and NLU-augmenter [19] for typos. Simultaneously, $N_{i}$ images $\mathcal{I}$ and $N_{i}$ augmented images $\mathcal{I}^{aug}$ are generated for $P^{org}$ and $\mathcal{P}^{aug}$ , respectively. Then the cosine similarity is calculated based on Eq. 4. + +$$ +O _ {v} = \frac {1}{2 N _ {i}} \sum_ {i = 1} ^ {N _ {i}} \sum_ {j = 1} ^ {N _ {i}} S _ {v} \left(I _ {i}, \mathcal {I} _ {j} ^ {\text {a u g}}\right), \tag {4} +$$ + +where $S_{v}(.,)$ is visual similarity scoring function; CLIP [53]. + +# 4.3. Miscellaneous + +- Emotion. To comprehensively measure a T2I model's ability to generate images with grounded emotional tones, three evaluation metrics are proposed; AC-T2I (Sec. 4.2.3), T2I alignment (Sec. 4.2.1) and visual emotion classification accuracy. Regarding the visual emotion classifier, we train a ResNet-101 classifier based on combined datasets; FI [71] and ArtEmis [5], to ensure the model can handle diverse domains and scenarios. + +-Fidelity. We rely on human evaluation, using Amazon Mechanical Turk (AMT) [13]. The annotators are asked to rate each image from 1-5, where 1 is the worst and 5 is the best. For a fair comparison, all the models' output images are shown in the same grid. + +-Bias. Three bias attributes are assessed, i.e., gender, race, and age. First, the human faces are detected using ArcFace [15], and RetinaFace [16], then the facial attributes are detected using Dex [60]. Finally, the bias score is defined as the distribution skew, i.e., mean absolute deviation (MAD) [50]; Eq. 5, where the balanced case is $\frac{1}{N_b}$ . + +$$ +M A D = \frac {1}{N _ {b}} \sum_ {i = 1} ^ {N _ {b}} \left| \hat {N} _ {b} - \frac {1}{N _ {b}} \right|, \tag {5} +$$ + +where $N_{b}$ is the number of protected attribute groups, e.g., 2 genders, and $\hat{N}_b$ is the Dex output normalized count. + +# 5. Experimental Results + +# 5.1. Evaluated Methods + +We comprehensively evaluate the performance of nine recent large-scale T2I models introduced as follows. Transformer-based: minDALL-E [2] and DALL-E-Mini [1] + +![](images/f600644f1c9fd5d54447d56c5ae195d76db78c330491df8f199db7c6528a0ecd.jpg) +A) Counting + +![](images/43455f46d5042b1db02485bcada0eb5570aeb222bc1295780c8d7b9bcc08916f.jpg) +B) Visual Text + +![](images/347f11fa4b653083772ce7ff5b72f9ed03d76c5cc35b276e529c20835bcb3d83.jpg) +C) Consistency-Typos + +![](images/ee243b636493a26852338271a499ee24a85bbf665db8c67e9caa5fd5cdbce7f8.jpg) +D) Spatial - Size - Color Composition + +E) Action Composition - Creativity +![](images/f7848bc5ff2852a7972898201088fa5a57d5b2a2c481644970123be2bd378111.jpg) +Action Composition Creativity + +![](images/08650a0a3d2133819e04dbf29d4a1c9938d3b185c3ef50d2ad9a4a8c1ac819e9.jpg) + +![](images/f6752be672f080f19e4405df33b9dd51b837953b0bb949c61b9c78b74448e20c.jpg) +Figure 5: Quantitative results for nine skills are grouped into five sub-figures based on the evaluation criteria. +Figure 6: Qualitative results produced by DALLE-V2. Green and red boxes, respectively, frame the success and failure cases. More qualitative results for all models are demonstrated in the supplementary material. + +are two different publicly available implementations of original DALL-E [56], which uses VQVAE [67] to encode images with grids of discrete tokens and a multimodal transformer for next token prediction. In addition, CogView2 [21] extend to a hierarchical transformer for fast super-resolution synthesis, and Paella [57] improve parallel token sampling based on MaskGIT [11]. Diffusion-based: GLIDE [45] and DALLE-V2 [55] decode images via diffusion with CLIP [53] embedding. Stable-Diffusion V1 [59] and V2 [3], (dubbed as SD-V1 and SD-V2) speed up the training of diffusion models [28] by leveraging the latent space of a powerful pretrained VQVAE [67]. Finally, Struct-Diff [25] tackles the stable-diffusion compositions limitation by manipulating the cross-attention representations based on linguistic insights to preserve the compositional semantics. + +# 5.2. Accuracy Skills Results + +- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score. As shown in Figure 5 part A, DALLE-V2 [55] is the best in terms of precision. However, its recall is very poor, as it misses many objects. Whereas jointly considering recall and F1-score, SD-V1 [59] performs the best, despite its worst precision. + +Finding #1. No agreement between precision and recall. We can select the appropriate model based on the application, which metric is preferred. + +Finding #2. The more detailed prompt, the more accurate is counting performance. We explore three levels of prompts; 1) Vanilla prompt. The simplest form, e.g., two cups. 2) Meta-prompt. Intermediate level, e.g., describes a + +scene containing two cups. 3) Detailed. The meta-prompt is fed to GPT-3.5 to generate a detailed description including the desired objects, e.g., two cups filled with hot coffee sitting side-by-side on a wooden table. In general, we may think that the simpler and straightforward prompts may lead to better results for the counting skill. Surprisingly, as shown in Figure 8, the Recall and F1-score always increase when the detailed prompt is used. + +Finding #3. Composition-based solution is limited. We explore Struct-Diff [25] which tackle the compositionality limitation in SD-V1 [59]. As shown in Figure 5 part A, it increases the precision compared to SD-V1 [59]. However the recall and F1-score are decreased drastically. + +-Visual Text. We utilize two text recognition metrics, CER [42] and NED [65], which are highly correlated (95%). + +Finding #4. All models can not generate visual text even for the simplest case. As shown in Figure 5 part B, the best model is DALLE-V2 [55], which achieves a $75\%$ error rate. However, the performances of all the models are far from acceptable, i.e., $10 - 20\%$ error rate. + +Finding #5. Confusion between picturing and writing. The models show a good language understanding of the mentioned semantics. However, they lean toward drawing them instead of visually writing them. For instance, in Figure 6, in the visual text column and first row, the model draws the "potted plant" instead of writing the words. Consequently, the model prefers to draw the "vessel" in the second row. + +- Emotion. Finding #6. All models suffer from generating emotion-grounded images. Figure 10 shows the T2I and TIT alignment scores, i.e., BLEU [48], CIDEr [68], and CLIPScore [27], which are almost equally low and far from the acceptance range among the different models. To further validate our observation, we exploit an image-to-emotion classifier trained on combined datasets as discussed in Sec. 4.3. In addition, a human evaluation experiment is conducted. In both evaluations, the classifier and the human evaluation, we simplify the problem as a binary classification, where they are asked to classify the emotion, given the generated image, as a positive or negative emotion. Both report almost $50\%$ accuracy across the entire models, precisely the random performance, where the number of classes is only two. + +-Fidelity. We generate three distinct images using varying seeds for each of the models. Then, the annotators evaluate them on a scale of 1-5, where 5 is the best, and 1 is the worst. The normalized scores are reported in Figure 7. The best model is SD-V2 [3], where it achieves $62.4\%$ while the worst one is minDALL-E [2] which achieves $52.2\%$ . However all the models are far away from the accepted threshold, i.e., $80\%$ which corresponds to 4 on our rating system (1-5). + +# 5.3. Robustness and Generalization Results + +-Consistency and typos. We measure the alignment between the images generated from the original prompt and + +![](images/7444cd1479b88b88ef131e31b5f026e9ed3a3156aaba8beb1b1701f7ab509064.jpg) +Figure 7: Quantitative results for Fidelity and Bias skills. + +paraphrased or perturbed prompts using CLIPScore [27], as discussed in Section 4.2.4. + +Finding #7. Models are robust against language perturbations. As shown in Figure 5 part C, all the models perform well against language perturbations and achieve between $70\%$ to $82\%$ alignment score. Specifically, DALLE-V2 [55] jointly achieves the best average performance for both skills, i.e., consistency and typos. + +# -Spatial, Size, and Colors composition. + +Finding #8. The medium and hard levels are unattain-able. For each skill, we define three hardness levels, i.e., easy, medium, and hard, and the reported results in the whole paper are the average accuracy for the three levels. However, we report only the easy level accuracy for the spatial, size, and color composition as the entire models suffer even from the easy level. Moreover, all models fail on the medium and hard levels, where they almost got zeros. Consequently, this raises the severe limitation of the models' composition ability. As shown in Figure 5 part D, the best model, DALLE-V2 [55], achieves $28.3\%$ , $29.9\%$ , and $38\%$ for spatial, size, and colors composition, respectively. + +Finding #9. Composition-based solution is limited. Similar to finding #3, we explore Struct-Diff [25], where it enhances the SD-V1 [59] performance by almost $3\%$ on the easy level. However, it still fails on the challenging levels. + +- Action composition Regarding the action compositionality, as illustrated in Figure 5 part E, DALLE-V2 [55] performs the best in generating compositions based on actions, according to both TIT alignment (i.e., highest CIDEr score 1.4538) and human evaluation result. Furthermore, all the scores align well with human evaluation results, confirming our metric's accuracy in evaluating this skill. + +- Creativity. Since we retrieve the top-100 nearest training data with the CLIPScore[27], we obtain the average score of how the text prompt deviates from the training data, which is 0.4173. Since CLIP [53] maps images and text to shared space, the deviation score should be close if the generated image is aligned with the text. However, experimental results show that all the models fail to generate novel images, and the best model is SD-V2 [3], which achieves the highest deviation score of 0.3433. Due to creativity's very nature, it thrives on deviation. However, if the deviation becomes + +![](images/62a513945b024de1d2ffc271924853294ab16e9e6cc987dc89cffd6b61ad2b40.jpg) +Vanilla Meta Detailed + +![](images/4acdcf11f03aabb68d80952be802c31c00325c6705777d06e418698131921678.jpg) +Figure 8: Ablation study of the prompt details on the counting skill. + +![](images/088fbdf09d7bf58002717f5f64258fd192fb8d32a0bd4af871bae0be225ed028.jpg) + +![](images/d39e7ef10ec034bf346c2f5457570504102d77fe8b7a890fb31f0fa9b52fc685.jpg) + +![](images/bd6fbde329a199459cc4b54b4d5973266f7a6c6240cadf78075b9a5393be0b6a.jpg) +Figure 10: Emotions results using five different metrics, in addition to, the human evaluation. + +excessive, the resulting generation may veer toward adverse hedonic and meaningless outcomes. Therefore, we also evaluate TIT (BLEU [48], CIDEr [68]) alignment to ensure the generation keeps the semantic meaning with the text prompt. For example, as Figure 5 part E) shows, Paella achieves a relatively high deviation score but performs poorly in terms of BLEU [48] and CIDEr [68], thus it deviates too much and lose original semantic meaning of corresponding text prompt, which is aligned with human evaluation result. Therefore, both metrics are indispensable for creativity evaluation, and more than one alone is required. + +# 5.4. Fairness and Bias Results + +Finding #10. The models are fair. As demonstrated in Figure 9, the maximum fairness score is $3.5\%$ by Cogview 2 [21], which indicates that the difference in performance between sub-groups is negligible. + +Finding #11. The models are slightly biased. In contrast to fairness, the models tend to be biased towards gender, as the average mean absolute deviation, Eq. 5, for DALLE-V2 [55], Cogview2 [21], and minDALLE [2] is $20\%$ , as shown in Figure 7. However, SD-V1 [59] achieves the best results where the deviation is less than $8\%$ . GLIDE [45], by design, is trained not to generate humans. Consequently, DALLE-Mini [1], and Paella [57] perform poorly regarding face generation. Therefore, GLIDE [45], DALLE-Mini [1], and Paella [57] are excluded from the bias measure. + +# 5.5. Human Evaluation + +To prove the effectiveness of our benchmark, we conduct a human evaluation using Amazon Mechanical Turk (AMT) [13] over $10\%$ of our data across the entire skills. The human evaluation criteria are divided into two main groups: 1) Modular-based. The core blocks in each metric are evaluated. 2) End-to-End based. Using score-based evaluation. + +-Modular-based. The UniDet [76] is the core block for counting, visual-text, spatial, color, and size composition. First, we ask humans to visually inspect its performance and report the true positives, false positives, and false positives. Then, we measure Person-correlation between the human + +F1-score and our F1-score, which shows a high correlation between our calculations and human evaluation, i.e., $93\%$ . + +Similarly, we measure the detection and recognition accuracy of Textsnake [39] and SAR [34], respectively. Again, the correlation is in our favor, $98\%$ and $96\%$ , respectively. Moreover, regarding the emotion, the annotators are asked to binary classify the emotions, i.e., positive or negative, given only the images. The results are highly aligned with our measure, where both agree that the models generate natural images, indicating that the prompt's emotion indicator is ignored. Regarding consistency and typos, the core module is the augemer [14, 19]. To further ensure that all the generated prompts have the same meaning as the original prompt, we conduct a human study where we ask the users to rate the prompt pair (original and augmented) on a scale of 1 (not similar at all) to 5 (precisely same meaning). The results show a great alignment, i.e., $94\%$ . + +- End-to-End based. To assess the creativity metric, we equally select 100 images per model for each hardness level and let annotators score from 1 to 5 for each generated image considering: (1) whether the generated image is creative; (2) whether the image is aligned with the given prompt. For action composition skills, annotators are requested to assign a score between 1 to 5 based on the accuracy of the generated subject and actions in response to text prompts. + +# 6. Conclusion + +We introduce a comprehensive and reliable benchmark, dubbed HRS-Bench, for evaluating text-to-image (T2I) models. Our benchmark measures 13 skills, categorized into five major categories, and covers 50 applications, providing a holistic evaluation of T2I models. Through our evaluation of nine recent large-scale T2I models, we have identified areas where state-of-the-art models struggle to tackle these skills, highlighting the need for continued research and development. Our human evaluation results confirm the effectiveness and reliability of our benchmark. Further, our benchmark will help ease future T2I research and progress on improving the skills covered in this benchmark. + +# References + +[1] Dallemini. https://github.com/borisdayma/dalle-mini.2,6,9 +[2] mindalle. https://github.com/kakaobrain/minDALL-E.2,6,8,9 +[3] Stable-diffusion-v2. https://github.com/Stability-AI/stablediffusion.2,7,8 +[4] Panos Achlioptas, Maks Ovsjanikov, Leonidas Guibas, and Sergey Tulyakov. Affection: Learning affective explanations for real-world visual data. arXiv preprint arXiv:2210.01946, 2022.3 +[5] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 3, 6 +[6] Motasem Alfarra, Juan C. Pérez, Anna Frühstück, Philip H. S. Torr, Peter Wonka, and Bernard Ghanem. On the robustness of quality measures for gans, 2022. 2 +[7] Shane Barratt and Rishi Sharma. A note on the inception score, 2018. 2 +[8] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.6 +[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3 +[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 2 +[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 7 +[12] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. arXiv preprint arXiv:2202.04053, 2022. 1, 2, 3 +[13] Kevin Crowston. Amazon mechanical turk: A research tool for organizations and information systems scholars. In Anol Bhattacherjee and Brian Fitzgerald, editors, Shaping the Future of ICT Research. Methods and Approaches, pages 210-221, Berlin, Heidelberg, 2012. Springer Berlin Heidelberg. 6, 9 +[14] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 4, 6, 9 +[15] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6 +[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level + +face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 6 +[17] Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff M Phillips, and Kai-Wei Chang. Harms of gender exclusivity and challenges in non-binary representation in language technologies. arXiv preprint arXiv:2108.12084, 2021.3 +[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2 +[19] Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets Nick Siegel Damien Sileo Jamie Simon Chandan Singh Roman Sitelew Priyank Soni Taylor Sorensen William Soto Aman Srivastava KV Aditya Srivatsa Tony Sun,Mukund Varma T,A Tabassum,Fiona Anting Tan,Ryan Teehan, Mo Tiwari Marie Tolkiehn Athena Wang Zijian Wang,Gloria Wang,Zijie J.WangFuxuan Wei Bryan Wilie Genta Indra Winata Xinyi Wu,Witold Wydmanski,Tianbao Xie Usama YaseenM.Yee Jing Zhang and Yue Zhang.Nlaugenter: A framework for task-sensitive natural language augmentation2021.469 +[20] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1, 2 +[21] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2, 7, 9 + +[22] Tan M. Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation, 2021. 2 +[23] Tan M Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 594-609. Springer, 2022. 1, 2, 3 +[24] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214-226, 2012. 3, 6 +[25] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 7, 8 +[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4 +[27] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.5, 8 +[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 7 +[29] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7986-7994, 2018. 1 +[30] Divyansh Jha, Kai Yi, Ivan Skorokhodov, and Mohamed Elhoseiny. Creative walk adversarial networks: Novel art generation with probabilistic random walk deviation from style norms. 3, 6 +[31] Kenan Jiang, Xuehai He, Ruize Xu, and Xin Eric Wang. Comclip: Training-free compositional image and text matching. arXiv preprint arXiv:2211.13854, 2022. 2, 4, 5 +[32] Kiku Johnson. Sexual orientation, gender identity, and expression affirming approach and expansive practices. 2019. 3 +[33] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4 +[34] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8610-8617, 2019. 5, 9 +[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5 + +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1 +[37] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 1 +[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.4 +[39] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In Proceedings of the European conference on computer vision (ECCV), pages 20-36, 2018. 5, 9 +[40] Youssef Mohamed, Mohamed Abdelfattah, Shyma Alhuwaider, Feifan Li, Xiangliang Zhang, Kenneth Ward Church, and Mohamed Elhoseiny. Artelingo: A million emotion annotations of wikiart with emphasis on diversity over language and culture. arXiv preprint arXiv:2211.10780, 2022. 3 +[41] Youssef Mohamed, Faizan Farooq Khan, Kilichbek Haydarov, and Mohamed Elhoseiny. It is okay to not be okay: Overcoming emotional bias in affective image captioning by contrastive data collection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume abs/2204.07660, 2022. 3 +[42] Andrew Morris, Viktoria Maier, and Phil Green. From wer and ril to mer and wil: improved evaluation measures for connected speech recognition. 01 2004. 5, 8 +[43] Deana F Morrow and Lori Messinger. Sexual orientation and gender expression in social work practice: Working with gay, lesbian, bisexual, and transgender people. Columbia University Press, 2006. 3 +[44] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 1 +[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2, 7, 9 +[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 1 +[47] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language + +models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022. 4 +[48] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5, 6, 8, 9 +[49] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1, 2 +[50] Karl Pearson. Contributions to the mathematical theory of evolution. Philosophical Transactions of the Royal Society of London. A, 185:71-110, 1894. 6 +[51] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1, 2, 3 +[52] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q Weinberger. On fairness and calibration. Advances in neural information processing systems, 30, 2017. 3, 6 +[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 6, 7, 8 +[54] Micah Rajunov and A Scott Duane. Nonbinary: Memoirs of gender and identity. Columbia University Press, 2019. 3 +[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2, 7, 8, 9 +[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 7 +[57] Dominic Rampas, Pablo Pernias, Elea Zhong, and Marc Aubreville. Fast text-conditional discrete denoising on vector-quantized latent spaces. arXiv preprint arXiv:2211.07292, 2022. 1, 2, 7, 9 +[58] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 1 +[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 7, 8, 9 +[60] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 6 + +[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2 +[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 4 +[63] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 6 +[64] D.A. Silverstein and J.E. Farrell. The relationship between image fidelity and image quality. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 1, pages 881-884 vol.1, 1996. 3 +[65] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1557–1562. IEEE, 2019. 5, 8 +[66] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 5 +[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7 +[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5, 6, 8, 9 +[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1 +[70] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs], 2022. 4 +[71] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 6 +[72] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1 +[73] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language + +models behave like bags-of-words, and what to do about it? arXiv e-prints, pages arXiv-2210, 2022. 2, 5 +[74] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1 +[75] Yanzhe Zhang, Lu Jiang, Greg Turk, and Diyi Yang. Auditing gender presentation differences in text-to-image models. arXiv preprint arXiv:2302.03675, 2023. 1, 2 +[76] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022. 4, 9 +[77] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 1, 2 +[78] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 1 \ No newline at end of file diff --git a/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/images.zip b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..70f9955dc878dd64fb410652f02b1bdf583d1424 --- /dev/null +++ b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d60f2e88ffa53ac5306e38857a19789dc1717cb3fa64993d4ab525fc80916c26 +size 568700 diff --git a/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/layout.json b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..30d920263d875792bb7dedc6bcfb278f2cc55678 --- /dev/null +++ b/2023/HRS-Bench_ Holistic, Reliable and Scalable Benchmark for Text-to-Image Models/layout.json @@ -0,0 +1,11035 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 471, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 471, + 140 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 471, + 140 + ], + "type": "text", + "content": "HRS-Bench: Holistic, Reliable and Scalable Benchmark for Text-to-Image Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 160, + 463, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 160, + 463, + 190 + ], + "spans": [ + { + "bbox": [ + 129, + 160, + 463, + 190 + ], + "type": "text", + "content": "Eslam Mohamed Bakr1, Pengzhan Sun2**, Xiaqian Shen1**, Faizan Farooq Khan1**, Li Erran Li3, Mohamed Elhoseiny1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 191, + 463, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 191, + 463, + 217 + ], + "spans": [ + { + "bbox": [ + 129, + 191, + 463, + 217 + ], + "type": "text", + "content": "{eslam.abelrahman, xiaoqian.shen, faizan.khan, mohamed.elhoseiny}@kaust.edu.sa" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 219, + 456, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 219, + 456, + 231 + ], + "spans": [ + { + "bbox": [ + 137, + 219, + 456, + 231 + ], + "type": "text", + "content": "pengzhan@comp.nus.edu.sg, lilimam@amazon.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 451, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 451, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 451, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 232, + 451, + 244 + ], + "type": "text", + "content": "King Abdullah University of Science and Technology (KAUST)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 245, + 427, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 245, + 427, + 258 + ], + "spans": [ + { + "bbox": [ + 160, + 245, + 427, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 245, + 427, + 258 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 160, + 245, + 427, + 258 + ], + "type": "inline_equation", + "content": "{ }^{3} \\mathrm{AWS AI}," + }, + { + "bbox": [ + 160, + 245, + 427, + 258 + ], + "type": "text", + "content": " Amazon" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 312, + 290, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 290, + 601 + ], + "type": "text", + "content": "In recent years, Text-to-Image (T2I) models have been extensively studied, especially with the emergence of diffusion models that achieve state-of-the-art results on T2I synthesis tasks. However, existing benchmarks heavily rely on subjective human evaluation, limiting their ability to holistically assess the model's capabilities. Furthermore, there is a significant gap between efforts in developing new T2I architectures and those in evaluation. To address this, we introduce HRS-Bench, a concrete evaluation benchmark for T2I models that is Holistic, Reliable, and Scalable. Unlike existing benchmarks that focus on limited aspects, HRS-Bench measures 13 skills that can be categorized into five major categories: accuracy, robustness, generalization, fairness, and bias. In addition, HRS-Bench covers 50 scenarios, including fashion, animals, transportation, food, and clothes. We evaluate nine recent large-scale T2I models using metrics that cover a wide range of skills. A human evaluation aligned with " + }, + { + "bbox": [ + 47, + 312, + 290, + 601 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 47, + 312, + 290, + 601 + ], + "type": "text", + "content": " of our evaluations on average was conducted to probe the effectiveness of HRS-Bench. Our experiments demonstrate that existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions. We hope that our benchmark help ease future text-to-image generation research. The code and data are available at https://eslambakr.github.io/hrsbench.github.io/." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 613, + 128, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 128, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 128, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 633, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 288, + 694 + ], + "type": "text", + "content": "Text-to-Image Synthesis (T2I), one of the essential multimodal tasks, witnessed remarkable progress starting from conditional GANs [58, 44, 74, 29, 78], which are shown to work on simple datasets [46, 69, 72, 36], to recently diffusion models [20, 61, 18, 10, 57, 55, 77, 45, 59], which are trained" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 285, + 542, + 437 + ], + "blocks": [ + { + "bbox": [ + 309, + 285, + 542, + 437 + ], + "lines": [ + { + "bbox": [ + 309, + 285, + 542, + 437 + ], + "spans": [ + { + "bbox": [ + 309, + 285, + 542, + 437 + ], + "type": "image", + "image_path": "37206a980a5ac74d8b6214facb1f6d29db7672bbca5b6e6612114c3e25b821a9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 457, + 548, + 506 + ], + "lines": [ + { + "bbox": [ + 305, + 457, + 548, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 548, + 506 + ], + "type": "text", + "content": "Figure 1: An overview of our proposed benchmark, HRS-Bench, measures 13 skills which could be grouped into five major categories; accuracy, robustness, generalization, fairness, and bias." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 520, + 492, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 520, + 492, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 520, + 492, + 533 + ], + "type": "text", + "content": "on large-scale datasets, e.g., LAION [63, 62]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "content": "Despite the rapid progress, the existing models face several challenges, e.g., they cannot generate complex scenes with the desired objects and relationship composition [25, 37]. Furthermore, assessing the T2I models should include more than just fidelity, e.g., the ability to compose multiple objects and generate emotionally grounded or creative images. Therefore, some recent efforts are focusing on improving the existing metrics [23] or proposing new metrics that cover new aspects, such as bias [75], compositions [37, 25, 49]. Moreover, some other works propose new benchmarks, summarized in Table 1, that assess different aspects, e.g., counting [61, 12, 51], social-bias [12], and object fidelity [23, 51]. Even with various benchmarks available, they tend to only cover a limited range of aspects while overlooking crucial evaluation criteria such as robustness," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20041" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 127, + 476, + 208 + ], + "blocks": [ + { + "bbox": [ + 45, + 70, + 547, + 120 + ], + "lines": [ + { + "bbox": [ + 45, + 70, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 45, + 70, + 547, + 120 + ], + "type": "text", + "content": "Table 1: Comparison of text-to-image benchmarks in terms of: 1) The number of evaluated models. 2) The number of the covered skills. 3) The number of utilized metrics. 4) The evaluation type, whether human or metric based or both. 5) The number of prompts. 6) The prompt generation type, whether a template, human-based or both. 7) Whether there are different hardness levels are included. 8) Number of annotators contributed to the evaluation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 117, + 127, + 476, + 208 + ], + "lines": [ + { + "bbox": [ + 117, + 127, + 476, + 208 + ], + "spans": [ + { + "bbox": [ + 117, + 127, + 476, + 208 + ], + "type": "table", + "html": "
MethodEval TypePrompt Type
# Models# Skills# MetricsHumanAuto# PromptsTemplateHumanHardness Levels# Annotators
DrawBench [61]540✓✘200✘✓✘25
DALL-EVAL [12]453✓✘7330✓✘✘6
HE-T2I [51]230✓✘90✘✓✓20
TISE [23]735✘✓N/A✘✓✘N/A
HRS-Bench (Ours)91317✓✓45000✓✓✓1000
", + "image_path": "dd310ca2989cf661882aa7bee291781d5335de0a9f26efe53275367e8887a79b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 226, + 143, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 143, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 143, + 239 + ], + "type": "text", + "content": "fairness, and creativity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 243, + 290, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 243, + 290, + 589 + ], + "spans": [ + { + "bbox": [ + 45, + 243, + 290, + 589 + ], + "type": "text", + "content": "To bridge this gap, we propose our Holistic, Reliable, and Scalable benchmark, dubbed HRS-Bench. In contrast to existing benchmarks, we measure a wide range of different generative capabilities, precisely 13 skills which can be grouped into five major categories, as demonstrated in Figure 1; accuracy, robustness, generalization, fairness, and bias. Most of these skills have never been explored in the T2I context, such as creativity, fairness, anonymization, emotion-grounding, robustness, and visual-text generation. Even the other previously explored skills were studied from a limited perspective, for instance, DALL-EVAL [12] studied the social bias by generating limited template-based prompts; only 145 prompts. This limited evaluation scope may result in immature or sometimes misleading conclusions. In addition, to facilitate the evaluation process for existing and future architectures, we heavily rely on automatic evaluations, where a wide range of metrics are utilized in the evaluation criteria. Moreover, HRS-Bench covers 50 scenarios, e.g., fashion, animals, transportation, and food. Figure 2 demonstrates the top 15 applications and their object distribution. We evaluate nine recent large-scale T2I models, i.e., Stable-Diffusion V1 [59] and V2 [3], DALL-E 2 [55], GLIDE [45], CogView-V2 [21], Paella [57], minDALL-E [2], DALL-E-Mini [1], and Struct-Diff [25]. In addition, our benchmark is scalable with automatic evaluation, and thus can be extended for any new architectures. To probe the effectiveness of our HRS-Bench, we conduct a human assessment that aligns well with our evaluations by " + }, + { + "bbox": [ + 45, + 243, + 290, + 589 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 45, + 243, + 290, + 589 + ], + "type": "text", + "content": " on average. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 594, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 594, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 594, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 594, + 288, + 628 + ], + "type": "text", + "content": "- We develop a Holistic, Reliable, and scalable T2I benchmark called HRS-Bench, depicted in Figure 1, which assess 13 skills covering 50 scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 630, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 630, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 58, + 630, + 288, + 666 + ], + "type": "text", + "content": "- Propose a new T2I alignment metric, called AC-T2I, which overcomes the composition limitations of existing large Vision-Language Models (VLMs) [31, 73]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 666, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 666, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 666, + 288, + 689 + ], + "type": "text", + "content": "- Nine T2I models are assessed based on our benchmark, including commercial and open-sourced ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 288, + 713 + ], + "type": "text", + "content": "- We verify the effectiveness of our HRS-Bench metric by conducting a human evaluation for " + }, + { + "bbox": [ + 58, + 689, + 288, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 58, + 689, + 288, + 713 + ], + "type": "text", + "content": " of our data" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 224, + 536, + 380 + ], + "blocks": [ + { + "bbox": [ + 317, + 224, + 536, + 380 + ], + "lines": [ + { + "bbox": [ + 317, + 224, + 536, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 536, + 380 + ], + "type": "image", + "image_path": "30a35cd2424c0cae31944cb3ddb282b3bea5fc9a86dc3c4e67008097ef79208a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 398, + 548, + 434 + ], + "lines": [ + { + "bbox": [ + 305, + 398, + 548, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 548, + 434 + ], + "type": "text", + "content": "Figure 2: Pie chart demonstrates the wide range of covered scenarios by our proposed benchmark, termed HRS-Bench, and their object distribution." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 444, + 489, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 444, + 489, + 456 + ], + "spans": [ + { + "bbox": [ + 324, + 444, + 489, + 456 + ], + "type": "text", + "content": "per skill that shows excellent alignment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 456, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 547, + 514 + ], + "type": "text", + "content": "- Driven by these holistic evaluations, several conclusions and findings are discussed. For instance, existing models often struggle to generate images with the desired count of objects, visual text, or grounded emotions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 526, + 515, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 515, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 515, + 540 + ], + "type": "text", + "content": "2. Revisiting Text-to-Image Benchmarks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": "Recently, there have been rapid efforts toward designing better image-generation models [20, 61, 18, 10, 57, 55, 77, 45, 59]. However, most existing metrics suffer from several limitations that make them unreliable [23, 7, 22, 6, 75, 49, 66, 73]. Table 1 summarizes the existing T2I benchmarks. -DrawBench. Imagen [61] proposes DrawBench to evaluate the T2I models from other aspects along with the image quality. Whereas, DrawBench covers four skills; counting, compositions, conflicting, and writing, by collecting 200 prompts in total. Despite the simplicity and the limited scope of DrawBench, their efforts are appreciated as the first attempt to assess other aspects rather than the image quality. -DALL-EVAL. [12] It proposes a toolkit called PAINTSKILLS, which assesses three simple visual" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20042" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 348 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "reasoning skills; object recognition, object counting, and spatial relation understanding, alongside two social bias skills; gender and racial bias. To facilitate the automatic evaluation, they built a unity simulator including limited objects to collect the datasets, approximately 80 object classes. In contrast, we cover more than 700 object classes. DETR [9] is utilized for visual reasoning skills evaluation after being fine-tuned on the synthetic dataset collected from the simulator. However, DALL-EVAL evaluates only four models using only 7k prompts and six annotators. Therefore, there is a need for a more comprehensive benchmark that can take into account a broader range of models, prompts, and annotators to provide a more thorough evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 287 + ], + "type": "text", + "content": "- HE-T2I. [51] It proposes 32 possible aspects to benchmark T2I models. However, only three are evaluated; counting, shapes, and faces, and the rest are left unexplored. The three aspects are evaluated using human evaluations, where twenty annotators have contributed to the evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "text", + "content": "-TISE. [23] It introduces a bag of metrics to evaluate T2I models from three aspects; positional alignment, counting, and fidelity. In addition, three fidelity metrics are introduced, i.e., " + }, + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "inline_equation", + "content": "IS^{*}" + }, + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "inline_equation", + "content": "OIS" + }, + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "inline_equation", + "content": "OFID" + }, + { + "bbox": [ + 46, + 288, + 288, + 348 + ], + "type": "text", + "content": ", and two alignment metrics PA for positional alignment, and CA for counting alignment." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 359, + 123, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 123, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 123, + 371 + ], + "type": "text", + "content": "3. HRS-Bench" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 380, + 287, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 287, + 404 + ], + "type": "text", + "content": "In this section, we first dissect the skills definition, then demonstrate the prompts collection pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 412, + 96, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 96, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 96, + 423 + ], + "type": "text", + "content": "3.1. Skills" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 430, + 121, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 121, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 121, + 442 + ], + "type": "text", + "content": "3.1.1 Accuracy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 449, + 288, + 631 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 45, + 449, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 449, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 45, + 449, + 287, + 498 + ], + "type": "text", + "content": "- Counting. A reliable T2I model should be able to ground the details in the prompt. One form of these details is objects binding with a specific frequency, e.g., \"four cars are parked around five benches in a park\"." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 498, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 287, + 570 + ], + "type": "text", + "content": "- Visual Text. Another essential aspect of assessing the model is generating high-quality text in wild scenes, e.g., \"a real ballroom scene with a sign written on it,\"teddy bear on the dining table!\"\". The importance of such skill comes from intervening in many scenarios, e.g., education applications, preparing illustration content, and designing billboards." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 570, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 570, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 45, + 570, + 288, + 605 + ], + "type": "text", + "content": "-Emotion. We measure to what extent the model can generate emotion-grounded images [5, 4, 41, 40], e.g., \"a rainy scene about cake, which makes us feel excitement.\"" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 606, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 606, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 45, + 606, + 287, + 631 + ], + "type": "text", + "content": "-Fidelity. Image fidelity indicates how accurately an image represents the underlying source distribution [64]." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 646, + 129, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 129, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 129, + 657 + ], + "type": "text", + "content": "3.1.2 Robustness" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "content": "To assess the T2I model's robustness, we cover two types of transformations; invariance and equivariance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "Invariance. Two skills are introduced to measure the invariance robustness: consistency and typos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 72, + 547, + 429 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "- Consistency. We measure the sensitivity of different T2I models towards prompt variations while keeping the same meaning and semantics, i.e., paraphrasing. For instance, generated images from these prompts, \"a woman is standing in front of a mirror, carefully selecting the perfect handbag\" and \"in front of a mirror, a woman is selecting the perfect handbag for the day\" should hold the same semantics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 158, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 547, + 194 + ], + "type": "text", + "content": "-Typos. Two natural perturbations are utilized to assess the models against possible sensibleness noise that users could cause during inference, i.e., typos and wrong capitalization." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 195, + 547, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 547, + 243 + ], + "type": "text", + "content": "Equivariance. Three different compositions are explored for the equivariance robustness. Specifically, we study three types of compositions, i.e., spatial, attribute-specific, and action compositions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 245, + 547, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 547, + 317 + ], + "type": "text", + "content": "-Spatial composition. In contrast to the counting skill which only measures the models' ability to compose multiple objects into a coherent scene, spatial composition additionally measures their ability to ground the detailed spatial relationship instructions mentioned in the input prompt, e.g., \"a person and a dog in the middle of a cat and a chair\"." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 390 + ], + "type": "text", + "content": "- Attribute-specific composition. Two types of attributes are controlled to study the attribute binding ability, i.e., colors and size attributes. For instance, \"an orange cat, a red dog, and a blue chair\" and \"a banana which is smaller than a person and bigger than a car\", for colors and size attribute binding, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 392, + 547, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 429 + ], + "type": "text", + "content": "-Action composition. It incorporates different subjects that doing different actions, e.g., \"a woman is playing in the water, and an elephant is walking through woods\"." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 453, + 403, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 403, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 403, + 464 + ], + "type": "text", + "content": "3.1.3 Generalization" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 475, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 559 + ], + "type": "text", + "content": "- Creativity. In this skill, models aim to generate images that represent not only the textual description but are also imaginative and novel. The creativity skill can be regarded as an out-of-distribution generation [30]. Accordingly, we devised innovative text prompts that are conceptually plausible but may need to be more readily available in standard training data sources, detailed later in Section 3.2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 583, + 375, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 375, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 375, + 595 + ], + "type": "text", + "content": "3.1.4 Fairness" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "We define fairness as the performance disparity among different sub-groups [24, 52]. A fair model should achieve the same performance on an arbitrary metric since no correlation exists between the metric and the protected attribute. Two attributes have been studied, i.e., gender and style. Following [12], gender refers to sex [32, 54] not the gender identity [43, 17]. We use two gender categories; male and female. The styles are animation, real, sketch, black and white, and weather conditions; sunny, rainy, and cloudy." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20043" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 99, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 99, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 99, + 83 + ], + "type": "text", + "content": "3.1.5 Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 163 + ], + "type": "text", + "content": "We assess the spurious correlation of the model towards predefined attributes, i.e., gender, race, and age. Using agnostic prompts towards a specific attribute, e.g., gender, the model should produce balanced generations of different classes of this attribute. For instance, the gender agnostic prompt could be, \"Two persons are working on their laptops\"." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 171, + 160, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 160, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 160, + 184 + ], + "type": "text", + "content": "3.2. Prompts Collection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 190, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 297 + ], + "type": "text", + "content": "For each skill, we collect 3k prompts. To ensure our benchmark is holistic enough, we split the prompts equally into three hardness levels, i.e., easy, medium, and hard. In addition, half of the prompts are human-based, and the other half is template-based, depicted in Figure 3. We filter human prompts manually from existing datasets [70, 33, 31]. We use the foundation model GPT-3.5 [47], text-davinci-003* to facilitate prompts generation, which will be abbreviated as GPT-3.5 later for convenience." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 298, + 288, + 693 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 45, + 298, + 288, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 298, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 45, + 298, + 288, + 382 + ], + "type": "text", + "content": "-Fidelity. The human prompts are sifted from [70]. Whereas, the template-based prompts are created by defining a template that describes a styled scene that contains some objects, as shown in Figure 3. Then, we create the meta-prompt by sampling the styles from pre-defined styles and the objects from LVIS dataset [26]. Finally, GPT-3.5 [47] is utilized to generate the final prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 382, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 382, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 45, + 382, + 288, + 477 + ], + "type": "text", + "content": "- Consistency and Typos. Consequently, the fidelity prompts are fed to Parrot [14] and NLU augmenter [19] to produce augmented prompts for consistency and typos, respectively. For consistency, we differentiate between the three hardness levels based on the similarity between the fidelity prompt and the augmented prompts using RoBERTa [38]. For typos, the number of introduced typos controls the three hardness levels, i.e., 1-2, 3-4, and 5-6, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 478, + 288, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 478, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 45, + 478, + 288, + 525 + ], + "type": "text", + "content": "-Counting. Given a meta-prompt, GPT-3.5 generate a realistic scenario that contains N objects. To generate the meta-prompts, we randomly samples the number of objects and the objects classes from LVIS dataset [26]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 526, + 288, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 526, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 45, + 526, + 288, + 574 + ], + "type": "text", + "content": "-Visual Text. We utilize GPT-3.5 to generate short descriptions which fit on a sign in a crowded scene. Then, we control the hardness levels by the text length and the surrounding scene complexity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 574, + 288, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 574, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 45, + 574, + 288, + 610 + ], + "type": "text", + "content": "- Emotion. We sample random objects from LVIS [26], then append an emotion indicator word forming the meta-prompt. Finally, GPT-3.5 is utilized to generate the final prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 610, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 610, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 610, + 288, + 693 + ], + "type": "text", + "content": "-Creativity. We craft text prompts that are challenging yet still within the realm of imagination. For easy level, we obtain subject, object, and relationship from Visual Genome [33] and obtain triplets by different combinations. Then sift through all the combinations from triplets extracted from LAION [62] dataset, and retain only the uncommon triplets. For the medium level, we fed the uncommon" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 156 + ], + "type": "text", + "content": "triplet to GPT-3.5 with the instruction: “Describe subject, relation and object in an imaginative way that will never be seen in the real world” and manually filter the undesirable sentences. Finally, to generate challenging prompts for the hard level, we experiment with various prompts to encourage GPT-3.5 to generate counterproductive sentences, as shown in Figure 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 156, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 346 + ], + "type": "text", + "content": "- Compositionality. We study three composition types, i.e., spatial, attribute-binding, and actions. The spatial prompts are collected using a pre-defined template, where a wide range of relations is utilized, e.g., \"on the right of\", \"above\", and \"between.\" For attribute-binding, two attributes are exploited, i.e., colors and size. For each hardness level, the number of objects' compositions increased, ranging from 2 to 4. For the action-level compositional generation, we design prompts with multiple combinations of actions starting from ComCLIP [31]. We combine two sentences from ComCLIP [31] for the easy level. Then, for medium and hard levels, we randomly choose one sentence and feed 'Extend text' to let the subject have at least three actions.' and 'Extend text' with other subjects doing other actions' into GPT-3.5, respectively, to obtain the final prompt. Detailed examples are demonstrated in Figure 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 347, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 418 + ], + "type": "text", + "content": "-Bias. Random objects are sampled from LVIS datasets [26], combined with a pre-defined template, creating a metaprompt. Then, the meta-prompt is fed to GPT-3.5 to produce the final prompt, as depicted in Figure 3. To ensure the prompts are agnostic towards the protected attributes, i.e., gender, race, and age, we manually validate them." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "text", + "content": "-Fairness. We adapt the bias prompts. For gender fairness, we replace gender-agnostic words, such as a person, with gender-specific words, such as man and woman. Whereas for style fairness, a style indicator is appended to the beginning of the bias prompt, as shown in Figure 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 489, + 479, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 489, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 489, + 479, + 502 + ], + "type": "text", + "content": "4. Evaluation for our Benchmark" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 509, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 544 + ], + "type": "text", + "content": "As shown in Figure 4, we categorize the skills based on the evaluation criteria, one-to-many mapping. Thus the same skill may be assigned to several metrics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 552, + 443, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 443, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 443, + 563 + ], + "type": "text", + "content": "4.1. Detection-Based Metrics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "content": "We utilize UniDet [76] for counting, spatial and attribute compositions because it supports a wide range of object classes, i.e., exceeding 700." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "type": "text", + "content": "- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score, where Precision assesses the accuracy of additional objects, and Recall assesses the accuracy of missing objects." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "-Spatial Compositions. Using a simple geometry module, we use the predicted bounding boxes to validate whether the spatial relation is grounded correctly. For instance, given the prompt \"A cat above a car\", the predicted bounding boxes will be " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{x_{min}^{1},y_{min}^{1},x_{max}^{1},y_{max}^{1}\\}" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "type": "text", + "content": "* https://platform.openai.com/docs/models/gpt-3-5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20044" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 545, + 230 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 545, + 230 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 545, + 230 + ], + "type": "image", + "image_path": "cb17387a2bfdfd71fcf5785194a3985e82f28d58c32ca44c17a7111a360efce7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 548, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 548, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 548, + 264 + ], + "type": "text", + "content": "Figure 3: Our prompt generation pipeline. First, we create a meta-prompt, which is a template-based prompt (in blue). Then, we sample the skill-related attributes (in orange). Finally, we generate the final prompt using ChatGPT (in green)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 281, + 282, + 373 + ], + "blocks": [ + { + "bbox": [ + 50, + 281, + 282, + 373 + ], + "lines": [ + { + "bbox": [ + 50, + 281, + 282, + 373 + ], + "spans": [ + { + "bbox": [ + 50, + 281, + 282, + 373 + ], + "type": "image", + "image_path": "6d4adacee38b78d062ba6564271775671d0491adc52c21221508376ef6e281b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 288, + 416 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 416 + ], + "type": "text", + "content": "Figure 4: On the left is our evaluation taxonomy. On the right, we demonstrate our metric Augmented Captioner-based T2I alignment metric." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\{x_{min}^2, y_{min}^2, x_{max}^2, y_{max}^2\\}" + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "text", + "content": " for the cat and the car, respectively, and the grounded spatial relation is above. Then, our geometry module will assess whether the spatial relation, i.e., above, is grounded correctly based on the following condition: " + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "inline_equation", + "content": "(y_{min}^1 < y_{min}^2)" + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "inline_equation", + "content": "(y_{max}^1 < y_{max}^2)" + }, + { + "bbox": [ + 46, + 430, + 287, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 493, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 493, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 45, + 493, + 287, + 566 + ], + "type": "text", + "content": "- Attributes Compositions. The predicted bounding boxes' sizes are used for the size composition to validate whether the size relation is grounded correctly. Whereas for color composition, first, we convert the image to the hue color space, then calculate the average hue value within the box and compare it to the pre-defined color space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 567, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 567, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 45, + 567, + 288, + 616 + ], + "type": "text", + "content": "-Visual Text. We adopt Textsnake [39] and SAR [34] for text detection and recognition, respectively. The recognition accuracy is measured by the Character Error Rate (CER) [42] and the Normalized Edit Distance (NED) [65]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 632, + 190, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 190, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 190, + 645 + ], + "type": "text", + "content": "4.2. Alignment-Based Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "content": "Three alignment paradigms are explored; Text-to-Image (T2I) (Sec. 4.2.1), Text-to-Image-to-Text (TIT) (Sec. 4.2.3), and Image-to-Image I2I (Sec. 4.2.4). In addition, we introduce our novel Augmented Captioner-based T2I Alignment metric, termed AC-T2I (Sec. 4.2.2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 283, + 403, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 403, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 403, + 296 + ], + "type": "text", + "content": "4.2.1 T2I Alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 302, + 547, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 547, + 423 + ], + "type": "text", + "content": "One possible solution to assess the T2I model's grounding ability is measuring the text and image correlation, e.g., CLIPScore [27, 53]. While CLIP is widely used, its effectiveness has been repeatedly questioned, as it is not sensitive to fine-grained text-image alignment and fails to understand compositions [66, 73]. For instance, [73] shows that CLIP [53] can not distinguish between \"the horse is eating the grass\" and \"the grass is eating the horse\". This motivates us to propose our novel augmented captioner-based T2I alignment metric, termed AC-T2I, depicted in Figure 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 439, + 453, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 453, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 453, + 451 + ], + "type": "text", + "content": "4.2.2 AC-T2I Alignment Metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 566 + ], + "type": "text", + "content": "We propose a new T2I alignment metric, called AC-T2I, which overcomes the compositional relationship's limitations of existing large Vision-Language Models (VLMs) [31, 73], by utilizing the n-grams based metric, e.g., CIDEr [68] and BLEU [48]. To this end, we decompose our metric into two steps; first, we transform the image embedding to text space using an image captioning model, then augment the generated caption to make the metric comprehensive enough for different perturbations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": "-Reformatting T2I as TIT. We reformat T2I as a TIT alignment by transforming the image features to text feature space, using an arbitrary function " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "G(\\mathcal{I})" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": " could be interpreted as an image captioner, e.g., BLIP2 [35]. As shown in Figure 4, given a text prompt " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "P^{org}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_k\\}_{k=1}^{N_i}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": " are generated, which are fed to an image captioner " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "G(\\mathcal{I})" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": " producing " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{c_i}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": " captions " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{C_k\\}_{k=1}^{N_{c_i}}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{c_i}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": " is the number of generated captions per image. Finally, the " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{c_i}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": " captions are automatically evaluated using CIDEr [68] and BLEU [48] against the input prompt " + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "inline_equation", + "content": "P^{org}" + }, + { + "bbox": [ + 304, + 567, + 547, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "- Comprehensive TIT. Instead of considering only the prompt " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "P^{org}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " as the GT caption, " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " augmented prompts" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20045" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{aug} = \\{P_k^{aug}\\}_{k=1}^{N_t}" + }, + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "text", + "content": " are generated using GPT-3.5, to measure the similarities comprehensively. To this end, we must ensure the GT is holistic enough; therefore, the rephrased version of the prompt " + }, + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "inline_equation", + "content": "P^{org}" + }, + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "text", + "content": " should be considered correct. Accordingly, the whole GT-prompt set for each image is defined as " + }, + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{P^{org}, \\mathcal{P}^{aug}\\} = \\{P_k\\}_{k=1}^{N_t+1}" + }, + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "text", + "content": ", i.e., one original prompt plus " + }, + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 46, + 71, + 289, + 156 + ], + "type": "text", + "content": " augmented prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 192 + ], + "type": "text", + "content": "Finally, for each prompt " + }, + { + "bbox": [ + 46, + 156, + 287, + 192 + ], + "type": "inline_equation", + "content": "P^{org}" + }, + { + "bbox": [ + 46, + 156, + 287, + 192 + ], + "type": "text", + "content": " we calculate the alignment score for each generated prompt-caption pair and select the highest score as the final alignment score, Eq. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 192, + 287, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 192, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 69, + 192, + 287, + 223 + ], + "type": "interline_equation", + "content": "O _ {t} = \\frac {1}{N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} \\max _ {1 \\leq j \\leq N _ {c _ {i}}, 1 \\leq k \\leq N _ {t} + 1} S _ {t} \\left(C _ {i, j}, P _ {k}\\right), \\tag {1}", + "image_path": "20a9338be9b0b43a43314f82bddbaa9a8a6ab0a10f17d598d2003226be3f3c41.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 287, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 232, + 287, + 256 + ], + "type": "inline_equation", + "content": "S_{t}(.)" + }, + { + "bbox": [ + 46, + 232, + 287, + 256 + ], + "type": "text", + "content": " is the text similarity scoring function, e.g., CIDEr [68], BLEU [48]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 270, + 145, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 145, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 145, + 283 + ], + "type": "text", + "content": "4.2.3 TIT Alignment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "spans": [ + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "text", + "content": "-Emotions. We explore a visual emotion classifier as illustrated in Sec. 4.3. Moreover, we apply our proposed metric, AC-T2I (Eq. 1), to avoid the aforementioned CLIP limitations, detailed in Section 4.2.1 and Section 4.2.2. The number of generated images per prompt " + }, + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "text", + "content": ", generated captions per image " + }, + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "inline_equation", + "content": "N_{c_i}" + }, + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "text", + "content": ", and the augmented prompts " + }, + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 45, + 289, + 289, + 372 + ], + "type": "text", + "content": " are set to 3, 5, and 9, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "text", + "content": "- Creativity. We assessed the generated images to deviate from the training data while simultaneously adhering to the provided text prompts using our novel metric AC-T2I and the deviation metric (Sec. 4.2.4). We set " + }, + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "inline_equation", + "content": "N_{c_i}" + }, + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "text", + "content": " to 3 and 5, respectively. Since it is hard to rephrase our novel prompts while maintaining its creative intent correctly, there will be no augmented prompts for it (" + }, + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "inline_equation", + "content": "N_{t} = 0" + }, + { + "bbox": [ + 45, + 373, + 287, + 456 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 456, + 288, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 456, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 45, + 456, + 288, + 491 + ], + "type": "text", + "content": "-Gender and styles fairness and action compositions. The fairness score is defined as the disparities in subgroups' performance [24, 52], Eq. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 492, + 287, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 492, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 51, + 492, + 287, + 535 + ], + "type": "interline_equation", + "content": "F a i r n e s s _ {s c o r e} = \\frac {1}{N _ {s} C _ {2}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {j = i + 1} ^ {N _ {s}} \\frac {1 0 0 \\times | A (i) - A (j) |}{\\operatorname* {m a x} (A (i) , A (j))}, \\tag {2}", + "image_path": "b5daf3dc3c69c5c34bc0ec9524786ea62bf514f5cef245a9f439c49a2abaf167.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\frac{100}{N_sC_2\\times max(A(i),A(j))}" + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "text", + "content": " is a normalization factor, " + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "text", + "content": " is the number of sub-groups, e.g., two for the gender and " + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "text", + "content": " is the accuracy measure, e.g., AC-T2I or CLIP scores. The less is, the better for Fairness score. Consequently, for the action composition, we exploit the AC-T2I metric, where " + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "inline_equation", + "content": "N_{i},N_{c_{i}}" + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 46, + 536, + 287, + 609 + ], + "type": "text", + "content": " are set to 3, 5, and 9, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 623, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 141, + 635 + ], + "type": "text", + "content": "4.2.4 I2I Alignment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 641, + 287, + 713 + ], + "type": "text", + "content": "- Creativity. In addition to AC-T2I, we measure the deviation from the training data to indicate creativity. Accessing large models' training data is challenging, however, most of them are trained on LAION [63]. Accordingly, we use LAION image-text retrieval tools [8] to fetch training data, which search among the dataset using CLIP [53] and a KNN index" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "to seek top-100 nearest images, denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{train}" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " for each prompt. The deviation score is calculated based on Eq. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 96, + 545, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 332, + 96, + 545, + 128 + ], + "type": "interline_equation", + "content": "\\triangle \\left(\\mathcal {I} ^ {t r a i n}, I _ {i}\\right) = \\frac {1}{2} - \\frac {1}{2 N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} S _ {v} \\left(\\mathcal {I} ^ {t r a i n}, I _ {i}\\right), \\tag {3}", + "image_path": "109c4d0ccfd090756ffb372816e30d1542319ce4b3ff1ea4143dcf041f7c0758.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 136, + 547, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 547, + 184 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 136, + 547, + 184 + ], + "type": "inline_equation", + "content": "S_{v}(, )" + }, + { + "bbox": [ + 304, + 136, + 547, + 184 + ], + "type": "text", + "content": " is the visual similarity scoring function, i.e., CLIP [53], " + }, + { + "bbox": [ + 304, + 136, + 547, + 184 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 304, + 136, + 547, + 184 + ], + "type": "text", + "content": " number of generated images per prompt. The similarity can be regarded as the Nearest Neighbour (NN) distance from the training dataset, similar to [30]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": "- Consistency and typos. Given a prompt " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "P^{org}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": ", augmented prompt " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{aug}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": " are generated using Parrot [14] for consistency and NLU-augmenter [19] for typos. Simultaneously, " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": " augmented images " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{aug}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": " are generated for " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "P^{org}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{aug}" + }, + { + "bbox": [ + 304, + 185, + 547, + 255 + ], + "type": "text", + "content": ", respectively. Then the cosine similarity is calculated based on Eq. 4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 358, + 255, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 255, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 358, + 255, + 545, + 289 + ], + "type": "interline_equation", + "content": "O _ {v} = \\frac {1}{2 N _ {i}} \\sum_ {i = 1} ^ {N _ {i}} \\sum_ {j = 1} ^ {N _ {i}} S _ {v} \\left(I _ {i}, \\mathcal {I} _ {j} ^ {\\text {a u g}}\\right), \\tag {4}", + "image_path": "d7a736fc388d24ac89c5f60eddf5ecd551d3958080077abd56be1db47407f578.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 297, + 547, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 311 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 297, + 547, + 311 + ], + "type": "inline_equation", + "content": "S_{v}(.,)" + }, + { + "bbox": [ + 304, + 297, + 547, + 311 + ], + "type": "text", + "content": " is visual similarity scoring function; CLIP [53]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 317, + 394, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 317, + 394, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 394, + 328 + ], + "type": "text", + "content": "4.3. Miscellaneous" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 335, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 430 + ], + "type": "text", + "content": "- Emotion. To comprehensively measure a T2I model's ability to generate images with grounded emotional tones, three evaluation metrics are proposed; AC-T2I (Sec. 4.2.3), T2I alignment (Sec. 4.2.1) and visual emotion classification accuracy. Regarding the visual emotion classifier, we train a ResNet-101 classifier based on combined datasets; FI [71] and ArtEmis [5], to ensure the model can handle diverse domains and scenarios." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 431, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 491 + ], + "type": "text", + "content": "-Fidelity. We rely on human evaluation, using Amazon Mechanical Turk (AMT) [13]. The annotators are asked to rate each image from 1-5, where 1 is the worst and 5 is the best. For a fair comparison, all the models' output images are shown in the same grid." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 491, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 564 + ], + "type": "text", + "content": "-Bias. Three bias attributes are assessed, i.e., gender, race, and age. First, the human faces are detected using ArcFace [15], and RetinaFace [16], then the facial attributes are detected using Dex [60]. Finally, the bias score is defined as the distribution skew, i.e., mean absolute deviation (MAD) [50]; Eq. 5, where the balanced case is " + }, + { + "bbox": [ + 304, + 491, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\frac{1}{N_b}" + }, + { + "bbox": [ + 304, + 491, + 547, + 564 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 363, + 565, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 565, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 363, + 565, + 545, + 597 + ], + "type": "interline_equation", + "content": "M A D = \\frac {1}{N _ {b}} \\sum_ {i = 1} ^ {N _ {b}} \\left| \\hat {N} _ {b} - \\frac {1}{N _ {b}} \\right|, \\tag {5}", + "image_path": "34433ee6edc0d022ce5b82cfe893d1d9607b52256574085d8b223dcb4b394db3.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "content": " is the number of protected attribute groups, e.g., 2 genders, and " + }, + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\hat{N}_b" + }, + { + "bbox": [ + 304, + 605, + 545, + 628 + ], + "type": "text", + "content": " is the Dex output normalized count." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 639, + 432, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 432, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 432, + 653 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 659, + 419, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 419, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 419, + 670 + ], + "type": "text", + "content": "5.1. Evaluated Methods" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "We comprehensively evaluate the performance of nine recent large-scale T2I models introduced as follows. Transformer-based: minDALL-E [2] and DALL-E-Mini [1]" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20046" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 86, + 166, + 163 + ], + "blocks": [ + { + "bbox": [ + 92, + 76, + 129, + 85 + ], + "lines": [ + { + "bbox": [ + 92, + 76, + 129, + 85 + ], + "spans": [ + { + "bbox": [ + 92, + 76, + 129, + 85 + ], + "type": "text", + "content": "A) Counting" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 86, + 166, + 163 + ], + "lines": [ + { + "bbox": [ + 51, + 86, + 166, + 163 + ], + "spans": [ + { + "bbox": [ + 51, + 86, + 166, + 163 + ], + "type": "image", + "image_path": "f600644f1c9fd5d54447d56c5ae195d76db78c330491df8f199db7c6528a0ecd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 171, + 86, + 285, + 163 + ], + "blocks": [ + { + "bbox": [ + 209, + 76, + 252, + 84 + ], + "lines": [ + { + "bbox": [ + 209, + 76, + 252, + 84 + ], + "spans": [ + { + "bbox": [ + 209, + 76, + 252, + 84 + ], + "type": "text", + "content": "B) Visual Text" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 86, + 285, + 163 + ], + "lines": [ + { + "bbox": [ + 171, + 86, + 285, + 163 + ], + "spans": [ + { + "bbox": [ + 171, + 86, + 285, + 163 + ], + "type": "image", + "image_path": "43455f46d5042b1db02485bcada0eb5570aeb222bc1295780c8d7b9bcc08916f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 291, + 86, + 405, + 163 + ], + "blocks": [ + { + "bbox": [ + 318, + 76, + 384, + 85 + ], + "lines": [ + { + "bbox": [ + 318, + 76, + 384, + 85 + ], + "spans": [ + { + "bbox": [ + 318, + 76, + 384, + 85 + ], + "type": "text", + "content": "C) Consistency-Typos" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 291, + 86, + 405, + 163 + ], + "lines": [ + { + "bbox": [ + 291, + 86, + 405, + 163 + ], + "spans": [ + { + "bbox": [ + 291, + 86, + 405, + 163 + ], + "type": "image", + "image_path": "347f11fa4b653083772ce7ff5b72f9ed03d76c5cc35b276e529c20835bcb3d83.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 411, + 81, + 526, + 158 + ], + "blocks": [ + { + "bbox": [ + 414, + 72, + 526, + 80 + ], + "lines": [ + { + "bbox": [ + 414, + 72, + 526, + 80 + ], + "spans": [ + { + "bbox": [ + 414, + 72, + 526, + 80 + ], + "type": "text", + "content": "D) Spatial - Size - Color Composition" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 81, + 526, + 158 + ], + "lines": [ + { + "bbox": [ + 411, + 81, + 526, + 158 + ], + "spans": [ + { + "bbox": [ + 411, + 81, + 526, + 158 + ], + "type": "image", + "image_path": "ee243b636493a26852338271a499ee24a85bbf665db8c67e9caa5fd5cdbce7f8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 53, + 194, + 259, + 266 + ], + "blocks": [ + { + "bbox": [ + 243, + 171, + 351, + 181 + ], + "lines": [ + { + "bbox": [ + 243, + 171, + 351, + 181 + ], + "spans": [ + { + "bbox": [ + 243, + 171, + 351, + 181 + ], + "type": "text", + "content": "E) Action Composition - Creativity" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 184, + 121, + 191 + ], + "lines": [ + { + "bbox": [ + 53, + 184, + 121, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 121, + 191 + ], + "type": "text", + "content": "Action Composition Creativity" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 53, + 194, + 259, + 266 + ], + "lines": [ + { + "bbox": [ + 53, + 194, + 259, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 259, + 266 + ], + "type": "image", + "image_path": "f7848bc5ff2852a7972898201088fa5a57d5b2a2c481644970123be2bd378111.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 261, + 194, + 536, + 266 + ], + "blocks": [ + { + "bbox": [ + 261, + 194, + 536, + 266 + ], + "lines": [ + { + "bbox": [ + 261, + 194, + 536, + 266 + ], + "spans": [ + { + "bbox": [ + 261, + 194, + 536, + 266 + ], + "type": "image", + "image_path": "08650a0a3d2133819e04dbf29d4a1c9938d3b185c3ef50d2ad9a4a8c1ac819e9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 310, + 545, + 483 + ], + "blocks": [ + { + "bbox": [ + 77, + 288, + 515, + 300 + ], + "lines": [ + { + "bbox": [ + 77, + 288, + 515, + 300 + ], + "spans": [ + { + "bbox": [ + 77, + 288, + 515, + 300 + ], + "type": "text", + "content": "Figure 5: Quantitative results for nine skills are grouped into five sub-figures based on the evaluation criteria." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 310, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 50, + 310, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 50, + 310, + 545, + 483 + ], + "type": "image", + "image_path": "f6752be672f080f19e4405df33b9dd51b837953b0bb949c61b9c78b74448e20c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 487, + 547, + 512 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 547, + 512 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 547, + 512 + ], + "type": "text", + "content": "Figure 6: Qualitative results produced by DALLE-V2. Green and red boxes, respectively, frame the success and failure cases. More qualitative results for all models are demonstrated in the supplementary material." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 533, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 713 + ], + "type": "text", + "content": "are two different publicly available implementations of original DALL-E [56], which uses VQVAE [67] to encode images with grids of discrete tokens and a multimodal transformer for next token prediction. In addition, CogView2 [21] extend to a hierarchical transformer for fast super-resolution synthesis, and Paella [57] improve parallel token sampling based on MaskGIT [11]. Diffusion-based: GLIDE [45] and DALLE-V2 [55] decode images via diffusion with CLIP [53] embedding. Stable-Diffusion V1 [59] and V2 [3], (dubbed as SD-V1 and SD-V2) speed up the training of diffusion models [28] by leveraging the latent space of a powerful pretrained VQVAE [67]. Finally, Struct-Diff [25] tackles the stable-diffusion compositions limitation by manipulating the cross-attention representations based on linguistic insights to preserve the compositional semantics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 533, + 439, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 439, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 439, + 545 + ], + "type": "text", + "content": "5.2. Accuracy Skills Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 553, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 547, + 625 + ], + "type": "text", + "content": "- Counting. We adopt the traditional detection measures, Precision, Recall, and F1-score. As shown in Figure 5 part A, DALLE-V2 [55] is the best in terms of precision. However, its recall is very poor, as it misses many objects. Whereas jointly considering recall and F1-score, SD-V1 [59] performs the best, despite its worst precision." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 628, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 663 + ], + "type": "text", + "content": "Finding #1. No agreement between precision and recall. We can select the appropriate model based on the application, which metric is preferred." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Finding #2. The more detailed prompt, the more accurate is counting performance. We explore three levels of prompts; 1) Vanilla prompt. The simplest form, e.g., two cups. 2) Meta-prompt. Intermediate level, e.g., describes a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20047" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "scene containing two cups. 3) Detailed. The meta-prompt is fed to GPT-3.5 to generate a detailed description including the desired objects, e.g., two cups filled with hot coffee sitting side-by-side on a wooden table. In general, we may think that the simpler and straightforward prompts may lead to better results for the counting skill. Surprisingly, as shown in Figure 8, the Recall and F1-score always increase when the detailed prompt is used." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 287, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 228 + ], + "type": "text", + "content": "Finding #3. Composition-based solution is limited. We explore Struct-Diff [25] which tackle the compositionality limitation in SD-V1 [59]. As shown in Figure 5 part A, it increases the precision compared to SD-V1 [59]. However the recall and F1-score are decreased drastically." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 228, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 228, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 45, + 228, + 287, + 253 + ], + "type": "text", + "content": "-Visual Text. We utilize two text recognition metrics, CER [42] and NED [65], which are highly correlated (95%)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 253, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 288, + 312 + ], + "type": "text", + "content": "Finding #4. All models can not generate visual text even for the simplest case. As shown in Figure 5 part B, the best model is DALLE-V2 [55], which achieves a " + }, + { + "bbox": [ + 46, + 253, + 288, + 312 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 253, + 288, + 312 + ], + "type": "text", + "content": " error rate. However, the performances of all the models are far from acceptable, i.e., " + }, + { + "bbox": [ + 46, + 253, + 288, + 312 + ], + "type": "inline_equation", + "content": "10 - 20\\%" + }, + { + "bbox": [ + 46, + 253, + 288, + 312 + ], + "type": "text", + "content": " error rate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 314, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 289, + 397 + ], + "type": "text", + "content": "Finding #5. Confusion between picturing and writing. The models show a good language understanding of the mentioned semantics. However, they lean toward drawing them instead of visually writing them. For instance, in Figure 6, in the visual text column and first row, the model draws the \"potted plant\" instead of writing the words. Consequently, the model prefers to draw the \"vessel\" in the second row." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 398, + 289, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 289, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 289, + 565 + ], + "type": "text", + "content": "- Emotion. Finding #6. All models suffer from generating emotion-grounded images. Figure 10 shows the T2I and TIT alignment scores, i.e., BLEU [48], CIDEr [68], and CLIPScore [27], which are almost equally low and far from the acceptance range among the different models. To further validate our observation, we exploit an image-to-emotion classifier trained on combined datasets as discussed in Sec. 4.3. In addition, a human evaluation experiment is conducted. In both evaluations, the classifier and the human evaluation, we simplify the problem as a binary classification, where they are asked to classify the emotion, given the generated image, as a positive or negative emotion. Both report almost " + }, + { + "bbox": [ + 46, + 398, + 289, + 565 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 398, + 289, + 565 + ], + "type": "text", + "content": " accuracy across the entire models, precisely the random performance, where the number of classes is only two." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "text", + "content": "-Fidelity. We generate three distinct images using varying seeds for each of the models. Then, the annotators evaluate them on a scale of 1-5, where 5 is the best, and 1 is the worst. The normalized scores are reported in Figure 7. The best model is SD-V2 [3], where it achieves " + }, + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "text", + "content": " while the worst one is minDALL-E [2] which achieves " + }, + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "inline_equation", + "content": "52.2\\%" + }, + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "text", + "content": ". However all the models are far away from the accepted threshold, i.e., " + }, + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 565, + 289, + 662 + ], + "type": "text", + "content": " which corresponds to 4 on our rating system (1-5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 253, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 253, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 253, + 682 + ], + "type": "text", + "content": "5.3. Robustness and Generalization Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 689, + 289, + 714 + ], + "type": "text", + "content": "-Consistency and typos. We measure the alignment between the images generated from the original prompt and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 354, + 72, + 492, + 164 + ], + "blocks": [ + { + "bbox": [ + 354, + 72, + 492, + 164 + ], + "lines": [ + { + "bbox": [ + 354, + 72, + 492, + 164 + ], + "spans": [ + { + "bbox": [ + 354, + 72, + 492, + 164 + ], + "type": "image", + "image_path": "7444cd1479b88b88ef131e31b5f026e9ed3a3156aaba8beb1b1701f7ab509064.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 175, + 541, + 187 + ], + "lines": [ + { + "bbox": [ + 310, + 175, + 541, + 187 + ], + "spans": [ + { + "bbox": [ + 310, + 175, + 541, + 187 + ], + "type": "text", + "content": "Figure 7: Quantitative results for Fidelity and Bias skills." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "text", + "content": "paraphrased or perturbed prompts using CLIPScore [27], as discussed in Section 4.2.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 232, + 547, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 547, + 304 + ], + "type": "text", + "content": "Finding #7. Models are robust against language perturbations. As shown in Figure 5 part C, all the models perform well against language perturbations and achieve between " + }, + { + "bbox": [ + 305, + 232, + 547, + 304 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 305, + 232, + 547, + 304 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 232, + 547, + 304 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 305, + 232, + 547, + 304 + ], + "type": "text", + "content": " alignment score. Specifically, DALLE-V2 [55] jointly achieves the best average performance for both skills, i.e., consistency and typos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 305, + 473, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 473, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 473, + 317 + ], + "type": "text", + "content": "-Spatial, Size, and Colors composition." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "text", + "content": "Finding #8. The medium and hard levels are unattain-able. For each skill, we define three hardness levels, i.e., easy, medium, and hard, and the reported results in the whole paper are the average accuracy for the three levels. However, we report only the easy level accuracy for the spatial, size, and color composition as the entire models suffer even from the easy level. Moreover, all models fail on the medium and hard levels, where they almost got zeros. Consequently, this raises the severe limitation of the models' composition ability. As shown in Figure 5 part D, the best model, DALLE-V2 [55], achieves " + }, + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "inline_equation", + "content": "28.3\\%" + }, + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "inline_equation", + "content": "29.9\\%" + }, + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 305, + 317, + 547, + 460 + ], + "type": "text", + "content": " for spatial, size, and colors composition, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 461, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 461, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 461, + 547, + 510 + ], + "type": "text", + "content": "Finding #9. Composition-based solution is limited. Similar to finding #3, we explore Struct-Diff [25], where it enhances the SD-V1 [59] performance by almost " + }, + { + "bbox": [ + 305, + 461, + 547, + 510 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 305, + 461, + 547, + 510 + ], + "type": "text", + "content": " on the easy level. However, it still fails on the challenging levels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 594 + ], + "type": "text", + "content": "- Action composition Regarding the action compositionality, as illustrated in Figure 5 part E, DALLE-V2 [55] performs the best in generating compositions based on actions, according to both TIT alignment (i.e., highest CIDEr score 1.4538) and human evaluation result. Furthermore, all the scores align well with human evaluation results, confirming our metric's accuracy in evaluating this skill." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "- Creativity. Since we retrieve the top-100 nearest training data with the CLIPScore[27], we obtain the average score of how the text prompt deviates from the training data, which is 0.4173. Since CLIP [53] maps images and text to shared space, the deviation score should be close if the generated image is aligned with the text. However, experimental results show that all the models fail to generate novel images, and the best model is SD-V2 [3], which achieves the highest deviation score of 0.3433. Due to creativity's very nature, it thrives on deviation. However, if the deviation becomes" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20048" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 89, + 126, + 158 + ], + "blocks": [ + { + "bbox": [ + 49, + 79, + 108, + 87 + ], + "lines": [ + { + "bbox": [ + 49, + 79, + 108, + 87 + ], + "spans": [ + { + "bbox": [ + 49, + 79, + 108, + 87 + ], + "type": "text", + "content": "Vanilla Meta Detailed" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 89, + 126, + 158 + ], + "lines": [ + { + "bbox": [ + 49, + 89, + 126, + 158 + ], + "spans": [ + { + "bbox": [ + 49, + 89, + 126, + 158 + ], + "type": "image", + "image_path": "62a513945b024de1d2ffc271924853294ab16e9e6cc987dc89cffd6b61ad2b40.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 129, + 89, + 205, + 149 + ], + "blocks": [ + { + "bbox": [ + 129, + 89, + 205, + 149 + ], + "lines": [ + { + "bbox": [ + 129, + 89, + 205, + 149 + ], + "spans": [ + { + "bbox": [ + 129, + 89, + 205, + 149 + ], + "type": "image", + "image_path": "4acdcf11f03aabb68d80952be802c31c00325c6705777d06e418698131921678.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 173, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 173, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 289, + 198 + ], + "type": "text", + "content": "Figure 8: Ablation study of the prompt details on the counting skill." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 208, + 89, + 282, + 158 + ], + "blocks": [ + { + "bbox": [ + 208, + 89, + 282, + 158 + ], + "lines": [ + { + "bbox": [ + 208, + 89, + 282, + 158 + ], + "spans": [ + { + "bbox": [ + 208, + 89, + 282, + 158 + ], + "type": "image", + "image_path": "088fbdf09d7bf58002717f5f64258fd192fb8d32a0bd4af871bae0be225ed028.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 290, + 78, + 396, + 164 + ], + "blocks": [ + { + "bbox": [ + 290, + 78, + 396, + 164 + ], + "lines": [ + { + "bbox": [ + 290, + 78, + 396, + 164 + ], + "spans": [ + { + "bbox": [ + 290, + 78, + 396, + 164 + ], + "type": "image", + "image_path": "d39e7ef10ec034bf346c2f5457570504102d77fe8b7a890fb31f0fa9b52fc685.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 405, + 66, + 544, + 170 + ], + "blocks": [ + { + "bbox": [ + 405, + 66, + 544, + 170 + ], + "lines": [ + { + "bbox": [ + 405, + 66, + 544, + 170 + ], + "spans": [ + { + "bbox": [ + 405, + 66, + 544, + 170 + ], + "type": "image", + "image_path": "bd6fbde329a199459cc4b54b4d5973266f7a6c6240cadf78075b9a5393be0b6a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 171, + 546, + 205 + ], + "lines": [ + { + "bbox": [ + 403, + 171, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 403, + 171, + 546, + 205 + ], + "type": "text", + "content": "Figure 10: Emotions results using five different metrics, in addition to, the human evaluation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 228, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 360 + ], + "type": "text", + "content": "excessive, the resulting generation may veer toward adverse hedonic and meaningless outcomes. Therefore, we also evaluate TIT (BLEU [48], CIDEr [68]) alignment to ensure the generation keeps the semantic meaning with the text prompt. For example, as Figure 5 part E) shows, Paella achieves a relatively high deviation score but performs poorly in terms of BLEU [48] and CIDEr [68], thus it deviates too much and lose original semantic meaning of corresponding text prompt, which is aligned with human evaluation result. Therefore, both metrics are indispensable for creativity evaluation, and more than one alone is required." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 367, + 189, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 189, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 189, + 379 + ], + "type": "text", + "content": "5.4. Fairness and Bias Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 386, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 434 + ], + "type": "text", + "content": "Finding #10. The models are fair. As demonstrated in Figure 9, the maximum fairness score is " + }, + { + "bbox": [ + 46, + 386, + 287, + 434 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 46, + 386, + 287, + 434 + ], + "type": "text", + "content": " by Cogview 2 [21], which indicates that the difference in performance between sub-groups is negligible." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 434, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 555 + ], + "type": "text", + "content": "Finding #11. The models are slightly biased. In contrast to fairness, the models tend to be biased towards gender, as the average mean absolute deviation, Eq. 5, for DALLE-V2 [55], Cogview2 [21], and minDALLE [2] is " + }, + { + "bbox": [ + 46, + 434, + 288, + 555 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 555 + ], + "type": "text", + "content": ", as shown in Figure 7. However, SD-V1 [59] achieves the best results where the deviation is less than " + }, + { + "bbox": [ + 46, + 434, + 288, + 555 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 434, + 288, + 555 + ], + "type": "text", + "content": ". GLIDE [45], by design, is trained not to generate humans. Consequently, DALLE-Mini [1], and Paella [57] perform poorly regarding face generation. Therefore, GLIDE [45], DALLE-Mini [1], and Paella [57] are excluded from the bias measure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 562, + 160, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 160, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 160, + 574 + ], + "type": "text", + "content": "5.5. Human Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "content": "To prove the effectiveness of our benchmark, we conduct a human evaluation using Amazon Mechanical Turk (AMT) [13] over " + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 652 + ], + "type": "text", + "content": " of our data across the entire skills. The human evaluation criteria are divided into two main groups: 1) Modular-based. The core blocks in each metric are evaluated. 2) End-to-End based. Using score-based evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 653, + 288, + 712 + ], + "type": "text", + "content": "-Modular-based. The UniDet [76] is the core block for counting, visual-text, spatial, color, and size composition. First, we ask humans to visually inspect its performance and report the true positives, false positives, and false positives. Then, we measure Person-correlation between the human" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 228, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 251 + ], + "type": "text", + "content": "F1-score and our F1-score, which shows a high correlation between our calculations and human evaluation, i.e., " + }, + { + "bbox": [ + 304, + 228, + 545, + 251 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 304, + 228, + 545, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "text", + "content": "Similarly, we measure the detection and recognition accuracy of Textsnake [39] and SAR [34], respectively. Again, the correlation is in our favor, " + }, + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "text", + "content": ", respectively. Moreover, regarding the emotion, the annotators are asked to binary classify the emotions, i.e., positive or negative, given only the images. The results are highly aligned with our measure, where both agree that the models generate natural images, indicating that the prompt's emotion indicator is ignored. Regarding consistency and typos, the core module is the augemer [14, 19]. To further ensure that all the generated prompts have the same meaning as the original prompt, we conduct a human study where we ask the users to rate the prompt pair (original and augmented) on a scale of 1 (not similar at all) to 5 (precisely same meaning). The results show a great alignment, i.e., " + }, + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 304, + 251, + 546, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 430, + 547, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 528 + ], + "type": "text", + "content": "- End-to-End based. To assess the creativity metric, we equally select 100 images per model for each hardness level and let annotators score from 1 to 5 for each generated image considering: (1) whether the generated image is creative; (2) whether the image is aligned with the given prompt. For action composition skills, annotators are requested to assign a score between 1 to 5 based on the accuracy of the generated subject and actions in response to text prompts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 557, + 547, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 701 + ], + "type": "text", + "content": "We introduce a comprehensive and reliable benchmark, dubbed HRS-Bench, for evaluating text-to-image (T2I) models. Our benchmark measures 13 skills, categorized into five major categories, and covers 50 applications, providing a holistic evaluation of T2I models. Through our evaluation of nine recent large-scale T2I models, we have identified areas where state-of-the-art models struggle to tackle these skills, highlighting the need for continued research and development. Our human evaluation results confirm the effectiveness and reliability of our benchmark. Further, our benchmark will help ease future T2I research and progress on improving the skills covered in this benchmark." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20049" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "text", + "content": "[1] Dallemini. https://github.com/borisdayma/dalle-mini.2,6,9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 288, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 288, + 136 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 288, + 136 + ], + "type": "text", + "content": "[2] mindalle. https://github.com/kakaobrain/minDALL-E.2,6,8,9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 138, + 288, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 138, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 138, + 288, + 159 + ], + "type": "text", + "content": "[3] Stable-diffusion-v2. https://github.com/Stability-AI/stablediffusion.2,7,8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 161, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 288, + 204 + ], + "type": "text", + "content": "[4] Panos Achlioptas, Maks Ovsjanikov, Leonidas Guibas, and Sergey Tulyakov. Affection: Learning affective explanations for real-world visual data. arXiv preprint arXiv:2210.01946, 2022.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 206, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 288, + 261 + ], + "type": "text", + "content": "[5] Panos Achlioptas, Maks Ovsjanikov, Kilichbek Haydarov, Mohamed Elhoseiny, and Leonidas J Guibas. Artemis: Affective language for visual art. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11569-11579, 2021. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 295 + ], + "type": "text", + "content": "[6] Motasem Alfarra, Juan C. Pérez, Anna Frühstück, Philip H. S. Torr, Peter Wonka, and Bernard Ghanem. On the robustness of quality measures for gans, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 297, + 288, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 318 + ], + "type": "text", + "content": "[7] Shane Barratt and Rishi Sharma. A note on the inception score, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 319, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 352 + ], + "type": "text", + "content": "[8] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 354, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 288, + 419 + ], + "type": "text", + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 421, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 475 + ], + "type": "text", + "content": "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "type": "text", + "content": "[11] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 565 + ], + "type": "text", + "content": "[12] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. arXiv preprint arXiv:2202.04053, 2022. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 632 + ], + "type": "text", + "content": "[13] Kevin Crowston. Amazon mechanical turk: A research tool for organizations and information systems scholars. In Anol Bhattacherjee and Brian Fitzgerald, editors, Shaping the Future of ICT Research. Methods and Approaches, pages 210-221, Berlin, Heidelberg, 2012. Springer Berlin Heidelberg. 6, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 634, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 288, + 655 + ], + "type": "text", + "content": "[14] Prithiviraj Damodaran. Parrot: Paraphrase generation for nlu., 2021. 4, 6, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "text", + "content": "[15] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 109, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 109, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 109, + 547, + 163 + ], + "type": "text", + "content": "[17] Sunipa Dev, Masoud Monajatipoor, Anaelia Ovalle, Arjun Subramonian, Jeff M Phillips, and Kai-Wei Chang. Harms of gender exclusivity and challenges in non-binary representation in language technologies. arXiv preprint arXiv:2108.12084, 2021.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 167, + 545, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 167, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 545, + 200 + ], + "type": "text", + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 202, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 202, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 202, + 547, + 609 + ], + "type": "text", + "content": "[19] Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann, Aadesh Gupta, Zhenhao Li, Saad Mahamood, Abinaya Mahendiran, Simon Mille, Ashish Srivastava, Samson Tan, Tongshuang Wu, Jascha Sohl-Dickstein, Jinho D. Choi, Eduard Hovy, Ondrej Dusek, Sebastian Ruder, Sajant Anand, Nagender Aneja, Rabin Banjade, Lisa Barthe, Hanna Behnke, Ian Berlot-Attwell, Connor Boyle, Caroline Brun, Marco Antonio Sobrevilla Cabezudo, Samuel Cahyawijaya, Emile Chapuis, Wanxiang Che, Mukund Choudhary, Christian Clauss, Pierre Colombo, Filip Cornell, Gautier Dagan, Mayukh Das, Tanay Dixit, Thomas Dopierre, Paul-Alexis Dray, Suchitra Dubey, Tatiana Ekeinhor, Marco Di Giovanni, Rishabh Gupta, Rishabh Gupta, Louanes Hamla, Sang Han, Fabrice Harel-Canada, Antoine Honore, Ishan Jindal, Przemyslaw K. Joniak, Denis Kleyko, Venelin Kovatchev, Kalpesh Krishna, Ashutosh Kumar, Stefan Langer, Seungjae Ryan Lee, Corey James Levinson, Hualou Liang, Kaizhao Liang, Zhexiong Liu, Andrey Lukyanenko, Vukosi Marivate, Gerard de Melo, Simon Meoni, Maxime Meyer, Afnan Mir, Nafise Sadat Moosavi, Niklas Muennighoff, Timothy Sum Hon Mun, Kenton Murray, Marcin Namysl, Maria Obedkova, Priti Oli, Nivranshu Pasricha, Jan Pfister, Richard Plant, Vinay Prabhu, Vasile Pais, Libo Qin, Shahab Raji, Pawan Kumar Rajpoot, Vikas Raunak, Roy Rinberg, Nicolas Roberts, Juan Diego Rodriguez, Claude Roux, Vasconcellos P. H. S., Ananya B. Sai, Robin M. Schmidt, Thomas Scialom, Tshephisho Sefara, Saqib N. Shamsi, Xudong Shen, Haoyue Shi, Yiwen Shi, Anna Shvets Nick Siegel Damien Sileo Jamie Simon Chandan Singh Roman Sitelew Priyank Soni Taylor Sorensen William Soto Aman Srivastava KV Aditya Srivatsa Tony Sun,Mukund Varma T,A Tabassum,Fiona Anting Tan,Ryan Teehan, Mo Tiwari Marie Tolkiehn Athena Wang Zijian Wang,Gloria Wang,Zijie J.WangFuxuan Wei Bryan Wilie Genta Indra Winata Xinyi Wu,Witold Wydmanski,Tianbao Xie Usama YaseenM.Yee Jing Zhang and Yue Zhang.Nlaugenter: A framework for task-sensitive natural language augmentation2021.469" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 611, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 666 + ], + "type": "text", + "content": "[20] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 712 + ], + "type": "text", + "content": "[21] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2, 7, 9" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20050" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "text", + "content": "[22] Tan M. Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[23] Tan M Dinh, Rang Nguyen, and Binh-Son Hua. Tise: Bag of metrics for text-to-image synthesis evaluation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 594-609. Springer, 2022. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[24] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214-226, 2012. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "type": "text", + "content": "[25] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "text", + "content": "[27] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "text", + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 431 + ], + "type": "text", + "content": "[29] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7986-7994, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "text", + "content": "[30] Divyansh Jha, Kai Yi, Ivan Skorokhodov, and Mohamed Elhoseiny. Creative walk adversarial networks: Novel art generation with probabilistic random walk deviation from style norms. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 511 + ], + "type": "text", + "content": "[31] Kenan Jiang, Xuehai He, Ruize Xu, and Xin Eric Wang. Comclip: Training-free compositional image and text matching. arXiv preprint arXiv:2211.13854, 2022. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 288, + 544 + ], + "type": "text", + "content": "[32] Kiku Johnson. Sexual orientation, gender identity, and expression affirming approach and expansive practices. 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 546, + 288, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 288, + 611 + ], + "type": "text", + "content": "[33] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[34] Hui Li, Peng Wang, Chunhua Shen, and Guyu Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 8610-8617, 2019. 5, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 547, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 204 + ], + "type": "text", + "content": "[37] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 205, + 547, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 547, + 258 + ], + "type": "text", + "content": "[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 260, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 547, + 315 + ], + "type": "text", + "content": "[39] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In Proceedings of the European conference on computer vision (ECCV), pages 20-36, 2018. 5, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 547, + 380 + ], + "type": "text", + "content": "[40] Youssef Mohamed, Mohamed Abdelfattah, Shyma Alhuwaider, Feifan Li, Xiangliang Zhang, Kenneth Ward Church, and Mohamed Elhoseiny. Artelingo: A million emotion annotations of wikiart with emphasis on diversity over language and culture. arXiv preprint arXiv:2211.10780, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 382, + 547, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 547, + 447 + ], + "type": "text", + "content": "[41] Youssef Mohamed, Faizan Farooq Khan, Kilichbek Haydarov, and Mohamed Elhoseiny. It is okay to not be okay: Overcoming emotional bias in affective image captioning by contrastive data collection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume abs/2204.07660, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 546, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 546, + 480 + ], + "type": "text", + "content": "[42] Andrew Morris, Viktoria Maier, and Phil Green. From wer and ril to mer and wil: improved evaluation measures for connected speech recognition. 01 2004. 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 546, + 525 + ], + "type": "text", + "content": "[43] Deana F Morrow and Lori Messinger. Sexual orientation and gender expression in social work practice: Working with gay, lesbian, bisexual, and transgender people. Columbia University Press, 2006. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 526, + 547, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 547, + 581 + ], + "type": "text", + "content": "[44] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 546, + 635 + ], + "type": "text", + "content": "[45] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2, 7, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 546, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 546, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 546, + 680 + ], + "type": "text", + "content": "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "text", + "content": "[47] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20051" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[48] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5, 6, 8, 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[49] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 240 + ], + "type": "text", + "content": "[50] Karl Pearson. Contributions to the mathematical theory of evolution. Philosophical Transactions of the Royal Society of London. A, 185:71-110, 1894. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 297 + ], + "type": "text", + "content": "[51] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "text", + "content": "[52] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q Weinberger. On fairness and calibration. Advances in neural information processing systems, 30, 2017. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 397 + ], + "type": "text", + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 422 + ], + "type": "text", + "content": "[54] Micah Rajunov and A Scott Duane. Nonbinary: Memoirs of gender and identity. Columbia University Press, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "text", + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2, 7, 8, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 520 + ], + "type": "text", + "content": "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 567 + ], + "type": "text", + "content": "[57] Dominic Rampas, Pablo Pernias, Elea Zhong, and Marc Aubreville. Fast text-conditional discrete denoising on vector-quantized latent spaces. arXiv preprint arXiv:2211.07292, 2022. 1, 2, 7, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[58] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 668 + ], + "type": "text", + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 7, 8, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[60] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "text", + "content": "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 547, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 204 + ], + "type": "text", + "content": "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 205, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 547, + 259 + ], + "type": "text", + "content": "[63] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 260, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 547, + 304 + ], + "type": "text", + "content": "[64] D.A. Silverstein and J.E. Farrell. The relationship between image fidelity and image quality. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 1, pages 881-884 vol.1, 1996. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 304, + 547, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 547, + 370 + ], + "type": "text", + "content": "[65] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1557–1562. IEEE, 2019. 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 371, + 547, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 547, + 436 + ], + "type": "text", + "content": "[66] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 547, + 469 + ], + "type": "text", + "content": "[67] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 470, + 547, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 470, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 470, + 547, + 524 + ], + "type": "text", + "content": "[68] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5, 6, 8, 9" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 525, + 547, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 547, + 557 + ], + "type": "text", + "content": "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 558, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 558, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 547, + 602 + ], + "type": "text", + "content": "[70] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs], 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 602, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 547, + 646 + ], + "type": "text", + "content": "[71] Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. Building a large scale dataset for image emotion recognition: The fine print and the benchmark. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "text", + "content": "[72] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[73] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20052" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 362 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "models behave like bags-of-words, and what to do about it? arXiv e-prints, pages arXiv-2210, 2022. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "text", + "content": "[74] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "text", + "content": "[75] Yanzhe Zhang, Lu Jiang, Greg Turk, and Diyi Yang. Auditing gender presentation differences in text-to-image models. arXiv preprint arXiv:2302.03675, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "type": "text", + "content": "[76] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7571-7580, 2022. 4, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 288, + 306 + ], + "type": "text", + "content": "[77] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 288, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 362 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 362 + ], + "type": "text", + "content": "[78] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmagan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 291, + 731, + 302, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 731, + 302, + 742 + ], + "spans": [ + { + "bbox": [ + 291, + 731, + 302, + 742 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20053" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_content_list.json b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e5d01784d6c317101dd4c46e9926a043758460e --- /dev/null +++ b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_content_list.json @@ -0,0 +1,2140 @@ +[ + { + "type": "text", + "text": "HSE: Hybrid Species Embedding for Deep Metric Learning", + "text_level": 1, + "bbox": [ + 181, + 130, + 785, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bailin Yang $^{1,2}$ , Haoqiang Sun $^{1}$ , Frederick W. B. Li $^{3}$ , Zheng Chen $^{1}$ , Jianlu Cai $^{1}$ , Chao Song $^{1*}$", + "bbox": [ + 106, + 179, + 861, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Faculty of Computer Science and Technology, Zhejiang Gongshang University", + "bbox": [ + 169, + 204, + 799, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Faculty of Statistics and Mathematics, Zhejiang Gongshang University", + "bbox": [ + 200, + 223, + 766, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science, University of Durham", + "bbox": [ + 256, + 241, + 712, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 294, + 312, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep metric learning is crucial for finding an embedding function that can generalize to training and testing data, including unknown test classes. However, limited training samples restrict the model's generalization to downstream tasks. While adding new training samples is a promising solution, determining their labels remains a significant challenge. Here, we introduce Hybrid Species Embedding (HSE), which employs mixed sample data augmentations to generate hybrid species and provide additional training signals. We demonstrate that HSE outperforms multiple state-of-the-art methods in improving the metric Recall@K on the CUB-200, CAR-196 and SOP datasets, thus offering a novel solution to deep metric learning's limitations.", + "bbox": [ + 75, + 325, + 473, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 550, + 209, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image retrieval heavily depends on deep metric learning to grasp visual similarities. In its early stages, methods like pair-based loss [8, 13, 19, 37] and proxy-based loss [25, 15] were proposed to train models. Nevertheless, the true goal of image retrieval is to adapt to the unknown, and the challenge lies in selecting a metric that can effectively handle differences between classes in deep metric learning. A promising approach involves exploring both intra-class and inter-class variations within the image itself, rather than relying solely on label information [33, 43, 40, 23, 51, 30].", + "bbox": [ + 75, + 575, + 468, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hard negative samples, often termed false positives, refer to images that resemble anchor images but carry different labels. The reason behind these hard samples lies in their potential for sharing remarkably similar or even identical features with the anchor samples. Approaches like hard sample mining and generation have been proposed to aid network convergence by introducing a substantial gradient [1, 41, 12, 34, 49, 48, 17, 18]. Recently, an advancement in generating appropriate hard samples involves creating supplementary training data [17, 38, 5, 21]. This is primarily", + "bbox": [ + 75, + 727, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ecaa3d20883361c9e06584a465127b95d5c53a605b8010cf6140841665824105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 295, + 831, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36588301f77640667a9b9698f1cbaf42aba4020e6afbee62c4477cdcfc64763b.jpg", + "image_caption": [ + "(a) Linear interpolation" + ], + "image_footnote": [], + "bbox": [ + 511, + 335, + 687, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1056ed98209311971a51ea596e06753ad947904072e5f29a9c65623e85871de.jpg", + "image_caption": [ + "(b) Our HSE", + "Figure 1: (a) Deep metric learning methods struggle with linear interpolation for generating new training samples with labels when new samples fall between anchors of different classes. The manifold structure lacking clear boundaries makes it challenging to assign a label, even when new samples are between anchors of the same class. (b) Samples with multiple simultaneous features may not lie on a linear interpolation path, and their location may be unknown. They can be classified as an unknown class situated inside or outside a manifold boundary. To tackle this challenge, we propose using the HSE method to allocate the sample to an appropriate embedding space location based on its feature information." + ], + "image_footnote": [], + "bbox": [ + 702, + 337, + 880, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "achieved through linear interpolation, a prevalent method for generating synthetic samples. However, assigning absolute labels to these synthesized samples is intricate due to the manifold structures and cluster boundaries present in the embedded space [17]. Even when the new samples are formed by linear interpolation between samples of the same label, accurately determining absolute labels remains a challenge. Synthetic sample labels have emerged as a potential solution, with [38] corroborating their similarity to mixup data augmentation and highlighting the label's representation through linear interpolation. Nonetheless, this approach is constrained to mixup data augmentation, as mixup uniformly blends features and labels from different images, which might not hold true for other techniques.", + "bbox": [ + 496, + 597, + 892, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a novel Hybrid Species Embedding (HSE) that leverages mixed sample data augmentations to generate additional training samples [35, 47, 45, 2, 20]. HSE creates hybrid species by combining features from multiple classes and embedding them into spatial positions. During training, different classes are randomly chosen for data augmentation", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding to:csong@zjsu.edu.cn", + "bbox": [ + 99, + 887, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11047", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in each batch. The labels of hybrid species are disregarded, and their placement near the category with the most similar features simulates human categorization behavior, as depicted in Figure 1(b). This strategy positions hybrid species close to the samples they are synthesized from, ensuring similarity while avoiding unrelated samples. By synthesizing samples using various mixed sample data augmentation techniques, each incorporating multiple features of a known class, our approach naturally imparts similar characteristics to the hard sample for the model without requiring explicit mining. Additionally, we dynamically adjust the position of each batch of hybrid species embedding to bypass the need for considering label information of these hybrid instances. Our contributions encompass:", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Introducing a novel metric learning strategy embedding challenging and unseen classes (Hybrid species), during training. These hybrid species offer supplementary training signals, enhancing the generalization capacity of downstream tasks, as validated through experiments.", + "- Adapting mixed sample data augmentation techniques for our Hybrid Species Embedding (HSE) to address constraints associated with pair-based metric learning losses, which typically demand explicit class labels.", + "- Demonstrating through experiments our efficacy in enhancing performance across state-of-the-art metric learning tasks on CUB-200, CAR-196, and SOP datasets." + ], + "bbox": [ + 76, + 301, + 468, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 496, + 218, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We utilize the Mixed Sample Data Augmentation (MSDA) method to create unknown samples and improve the model's generalization ability for downstream tasks. This approach resolves the issue of insignificant comparisons caused by random selection of samples in metric learning. While mining hard samples from a finite training set is limited, MSDA enables us to create more hard samples. In the following, we review existing relevant works.", + "bbox": [ + 75, + 521, + 468, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Metric Learning", + "text_level": 1, + "bbox": [ + 76, + 650, + 241, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep metric learning compares data in pairs and pulls them together if they belong to the same category, otherwise separates them [8, 13, 19, 37, 26]. The exponential growth in the number of tuples during training can present challenges in achieving model convergence, and may include meaningless metrics. Proxy-based approaches, e.g., choosing a proxy point for each class, have been proposed as a solution [25, 15]. These proxies contain information from multiple samples, require less computation in pairwise comparisons, helping the network converge more effectively.", + "bbox": [ + 75, + 674, + 468, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods that rely solely on label information may result in over-clustered or overly separated samples with identical or differing labels, respectively, and can lead to reduced performance on new and unseen test samples. EPSHN [43] notes this problem, particularly for species with varying", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "features such as birds of the same species with different sexes. To avoid this, [30, 40, 31, 33, 51, 50, 18] emphasize the importance of incorporating semantic information from the images themselves, often through large pre-trained language models like CLIP [29] and BERT [3], to better align the visual representation space and discover potential differences within and between classes. The MSloss function [40] is one such approach that combines self-similarity and relative similarities to weight selected pairs.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depending solely on label information can be insufficient for effective sample learning. Previous works addressed this limitation by assigning varying metrics or weights based on inter-sample variability. However, these techniques had limitations when dealing with a small number of training samples or unseen samples. To address this, we introduce additional training samples and a novel metric for improved learning. Prior works like [49, 17] added additional samples, but mainly relied on linear interpolation. In contrast, our approach generates extra training samples through selecting the appropriate MSDA, providing greater flexibility than linear interpolation.", + "bbox": [ + 496, + 227, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Hard Sample Mining", + "text_level": 1, + "bbox": [ + 500, + 421, + 699, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pair-based metric learning generates lots of paired samples, and handling negative samples is crucial for reducing model redundancy. Various studies [6, 10, 34, 41, 49] have explored the use of hard negative mining while constructing triples to create useful gradient sums, aiding fast convergence of triplet loss networks. Additionally, [34] suggested a semi-hard negatives scheme that selects even harder negative samples, while [6] introduced a hierarchical triplet loss (HTL) that constructs a hierarchical tree of all classes and selects semi-hard negatives with a dynamic margin.", + "bbox": [ + 496, + 446, + 890, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The discussed methods require selecting hard samples from the training set to improve model performance. Hard samples are challenging examples that can evaluate the model's ability to generalize. For example, if a model performs well on horses and donkeys but struggles with mules, evaluating its performance on mules (as a hybrid of both species) can directly assess its generalization ability.", + "bbox": [ + 496, + 598, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies focus on generating additional hard samples for training. HDML [49] creates hard samples between different labeled samples, while [17] generates additional training samples between identically labeled samples through hard sample mining. Gu et al. [7] use synthetic proxy points and competition to create hard samples. These methods use linear interpolation to generate new samples, leading to false negatives and positives due to the difficulties in determining class boundaries on the manifold structure. DAS [21] and IAA [5] try to solve this issue by adjusting features slightly and adding extra examples during mid-training of the model, aiming to make sure that new examples still belong to the same class. Despite these attempts,", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11048", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "completely removing this problem remains challenging.", + "bbox": [ + 76, + 90, + 444, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use mixed sample data augmentations in the preprocessing stage of images to generate additional hard samples, instead of linear interpolation. This has several advantages, including a higher probability of generating hard samples without subsequent mining and the ability to identify the underlying content of the image even without its labeling information. This makes our method highly effective for generating and processing additional hard samples.", + "bbox": [ + 75, + 107, + 467, + 228 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3d215d90ebbd4282c7727b890f1e5fddb6a32474d52211a4b39466424271b69.jpg", + "image_caption": [ + "Original images in the training set" + ], + "image_footnote": [], + "bbox": [ + 166, + 243, + 220, + 285 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6928a2fc5121385fabe763a7eef27da2212d9c75cde2037caf971a210f99677.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 243, + 272, + 285 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/348e4eb900e98c7dd043b4861f01556e938076c9f209b326e28c9338cfd98983.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 244, + 326, + 285 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2723d5c7d840fabcaf274d643b5982cc10924dc2472408714e849bb07135cb9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 244, + 380, + 285 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f7b7fc13c95b2660876701ecb06fe644dbfe072d4eca8644e7cbaea1c0fa5341.jpg", + "image_caption": [ + "CutMix:", + "num=2" + ], + "image_footnote": [], + "bbox": [ + 225, + 304, + 277, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b7e123435f312fde5952313b8f2b94e9126320004b582185a1c4b1f1583b5ca4.jpg", + "image_caption": [ + "num=3" + ], + "image_footnote": [], + "bbox": [ + 281, + 304, + 331, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/930df273531792bd39256143b57cd2e9e61edb8c3f4976ee9633c9ef9ed6cd0f.jpg", + "image_caption": [ + "num=4" + ], + "image_footnote": [], + "bbox": [ + 333, + 304, + 385, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 339, + 304, + 383, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bec7efbcb0d10baa15efbc267008d12044ce53b3b652a5bc36b555cfa6c139ab.jpg", + "image_caption": [ + "MixUp:" + ], + "image_footnote": [], + "bbox": [ + 223, + 364, + 276, + 407 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e4ad4171544bfe2ae3054ecbdf34b03e94c735d762f511b609cb594bec66dd6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 364, + 331, + 407 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/07ccb31ab4f6ff562ebf6fefb7ef0dc070c6e3a8e523d7883d494f4752c53e21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 364, + 383, + 407 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c04e8b8e90a95db5da00a802855ca7a9f5e1c7883eca3c98e1c2b015d46f1006.jpg", + "image_caption": [ + "GridMask:", + "mask=50 mask=150 mask=250", + "Figure 2: Creating Hybrid Species: Three Ways. We explore three different ways for generating hybrid species, varying parameter num, which controls the number of classes included. The mask indicates block size." + ], + "image_footnote": [], + "bbox": [ + 223, + 425, + 277, + 468 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9997b68f631ffc2a65aa5bb483f341015fa350646cceb949eab096f9bebcc0f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 426, + 333, + 468 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db9d53db31742015ccbd9f429ffcd1da82a716cfb10fe9f347044448f2efac82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 426, + 390, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Mixed Sample Data Augmentations (MSDA)", + "text_level": 1, + "bbox": [ + 76, + 556, + 452, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional data augmentation techniques assume all samples belong to the same class and do not consider relationships between different samples. In contrast, MSDA combines two samples, using various forms of label representation [35]. Popular methods include Mixup [47], Cutmix [45], Cutout [4], and GridMask [2], each with the core idea of producing images with more representative features of real-world scenarios. Other MSDA techniques include SmoothMix [20] and FMix [9].", + "bbox": [ + 75, + 580, + 467, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite significant achievements in image classification, the above methods have not been applied to metric learning. [17] discussed the similarity between linear interpolation of features and mixup data augmentation, but due to the peculiarities of mixed labels, it cannot be directly applied to metric learning, which typically requires distinguishing samples as dichotomous, i.e., positive and negative samples. Recently, [38] modified an existing metric learning loss function to accommodate mixup, but the input mix was limited to mixup. Our proposed method can be applied to a variety of mixed sample data augmentations (Section 3.2).", + "bbox": [ + 75, + 718, + 467, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We modified multiple mixed sample data augmentations", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as shown in Figure. 2 to ensure that mixed data augmented samples contain features from multiple classes in a balanced manner. Specifically, for CutMix, we uniformly cropped and merged images from top to bottom to ensure that features from multiple classes are regularly distributed. For Mixup, we fused the samples in equal proportions to ensure a balanced representation of features from multiple classes. Lastly, we modified Gridmask by replacing the mask with an image to create a fusion effect.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 241, + 671, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 266, + 638, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a preliminary for our proposed method, we consider a batch of images denoted by $x = x_{1},x_{2},\\ldots ,x_{n}$ with corresponding labels $y = y_{1},y_{2},\\ldots ,y_{n}$ . Deep metric learning involves the use of a Convolutional Neural Network (CNN) to learn a non-linear transformation of each image into an m-dimensional deep feature space $\\phi (x;\\theta_{\\phi}):\\mathcal{X}\\to \\mathbb{R}^{m}$ , where $\\theta_{\\phi}$ represents the network's parameters. In order to further learn a mapping from the m-dimensional feature space to a k-dimensional space, we use a linear mapping layer represented by $f(\\phi ;\\theta_f):\\mathbb{R}^m\\to \\mathbb{R}^k$ , where $\\theta_{f}$ represents the parameters of the linear mapping layer.", + "bbox": [ + 496, + 289, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we define the distance between two data points in the embedding space as:", + "bbox": [ + 498, + 455, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {f} \\left(x _ {i}, x _ {j}\\right) = \\left\\| f \\left(\\phi_ {i}\\right) - f \\left(\\phi_ {j}\\right) \\right\\| _ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 498, + 890, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Metric learning models usually involve pairwise sampling of a sample, e.g. considering the contrastive loss:", + "bbox": [ + 498, + 527, + 890, + 556 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o n t}} = y d _ {f} (a, x) ^ {2} + (1 - y) \\left[ m - d _ {f} (a, x) \\right] _ {+} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 569, + 890, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ denotes the label, $a$ is the anchor image and $m$ is the marginal value. To train a metric learning model, we first select an anchor image and choose a positive sample with the same label ( $y = 1$ ) and a negative sample with a different label ( $y = 0$ ). The network is then trained to minimize a loss function that penalizes negative samples that are too close together and positive samples that are too far apart. It is noteworthy that the loss value will be relatively large when the distance between the anchor point and the negative sample is small, while it will be relatively small when the distance is very far. Thus, selecting hard samples during the selection of pairs will result in a larger gradient.", + "bbox": [ + 496, + 598, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Studies have tried to increase the number of hard samples for metric learning by generating more samples. However, linear interpolation methods can produce false positives and negatives. It is important to note that the binary metric learning label $y$ cannot be used as a one-hot label for image classification tasks.", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Hybrid Species Embedding (HSE) method addresses the challenges of creating additional hard samples", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11049", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and applying them to metric learning without label information to improve generalization. We address the challenge of mixed sample data augmentations by emulating human categorization behavior to assign these samples to a reasonable position in the embedding space. We propose incorporating newly generated samples as an extra training signal for the model, using the same feature extraction network for both the original image and the hybrid species generated using MSDA, and included in each batch. The loss function $L_{m}$ clusters the original samples based on their labels, while our proposed loss function $L_{Hy}$ deals with the embedding positions of the new samples, as shown in Figure. 3.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f3f8834389bb4c29fdf2b5df2134f3c7f9dc9d82cb9f1679c4af67014f63ab7.jpg", + "image_caption": [ + "Figure 3: Our HSE method uses batches of original images and hybrid species generated by Mixed Sample Data Augmentation (MSDA). The feature extraction network (CNN) processes both types of images to obtain their corresponding feature vectors ( $f_{Or}$ and $f_{HS}$ ). The loss function includes two components: $L_{m}$ (metric learning loss) groups original samples into clusters based on their labels, and $L_{Hy}$ optimizes feature extraction by comparing hybrid species embedding positions." + ], + "image_footnote": [], + "bbox": [ + 78, + 282, + 468, + 435 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f7470e36f9792bcd72c73ce86cfeeb6211142287bd41f19ef52ea53f70f3047.jpg", + "image_caption": [ + "(a) HS-Anchors" + ], + "image_footnote": [], + "bbox": [ + 94, + 551, + 212, + 618 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d76e21d9d9d1f69e835cbfeba513c23fa888e87e24752703d95c417cba06e81b.jpg", + "image_caption": [ + "(b) WP-Anchor", + "Figure 4: Three types of anchor point selection: The dashed lines indicate that the sample pairs should move away from each other, while the solid lines indicate that the sample pairs should move closer to each other. The red sample is a hybrid species synthesized from green and blue. (a) denotes the use of the hybrid species as the anchor point, (b) denotes using the class that synthesizes the hybrid species as an anchor point, and (c) denotes using a class unrelated to the hybrid species as an anchor point." + ], + "image_footnote": [], + "bbox": [ + 215, + 551, + 333, + 618 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/690c4ff3dafa28070863a9518a8803f84530d6c74df6678b613435cd33493e99.jpg", + "image_caption": [ + "(c) N-Anchor" + ], + "image_footnote": [], + "bbox": [ + 333, + 551, + 454, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hybrid Species", + "text_level": 1, + "bbox": [ + 76, + 756, + 230, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the context of image classification tasks, hard negative samples refer to images that are similar to the anchor images but have different labels. We posit that hard negative samples typically possess similar features to the anchor samples. Therefore, when a sample is derived from an anchor image, there is a high likelihood that the sample will be classified as a hard sample. Let $x_{A}$ and $x_{B}$ denote two images with labels $y_{A}$ and $y_{B}$ , respectively, where $y_{A} \\neq y_{B}$ .", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The generated sample $x_{H}$ contains both the features of $x_{A}$ and the features of $x_{B}$ . We refer to $x_{H}$ as the hybrid species.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$x_{H}$ has a high likelihood of becoming a hard negative sample, while simultaneously presenting potential gradients in classes $A$ and $B$ . This can be achieved by utilizing mixed sample data augmentation techniques, such as mixup. Mixup generates a new training sample, denoted as $(x_{H},y_{H})$ , by linearly interpolating two training samples $(x_{A},y_{A})$ and $(x_{B},y_{B})$ , where $x_{H}$ and $y_{H}$ are:", + "bbox": [ + 496, + 121, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {H} = \\lambda x _ {A} + (1 - \\lambda) x _ {B} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 239, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {H} = \\lambda y _ {A} + (1 - \\lambda) y _ {B} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 282, + 890, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is obtained from a beta distribution and is constrained to [0, 1]. To achieve a distribution of $\\lambda$ values that is centered around 0.5, $\\lambda$ should follow a beta distribution with shape parameters $\\alpha$ and $\\alpha$ , e.g., $\\lambda \\sim \\mathrm{Beta}(\\alpha, \\alpha)$ , where $\\alpha \\in (0, \\infty)$ . The generated training sample $(x_H, y_H)$ is then used for model training with a loss function.", + "bbox": [ + 496, + 305, + 890, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Many pair-based metric learning losses rely on having distinct class labels, but the output $y_{H}$ is often a probability distribution per class, which is unsuitable for metric learning applications. Determining how to utilize these labels for metric learning presents a significant challenge. This difficulty is illustrated in Eq. 2.", + "bbox": [ + 496, + 396, + 890, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o n t}} = y d _ {f} \\left(a, x _ {H}\\right) ^ {2} + (1 - y) \\left[ m - d _ {f} \\left(a, x _ {H}\\right) \\right] _ {+} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 498, + 890, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the binary label $y \\in \\{0,1\\}$ , the label for $x_{H}$ is not straightforward, since it is a hybrid sample. When we consider $x_{H}$ as a negative sample of the anchor point, it is not reasonable, as $x_{H}$ contains the same features as the anchor samples. Keeping them away from each other could cause the model to ignore important features and focus on irrelevant ones. Therefore, the binary label $y$ is invalid for $x_{H}$ .", + "bbox": [ + 496, + 527, + 890, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[38] showed that classification tasks are not fundamentally different from metric learning, and the label of the synthesized sample $X_{H}$ can still be represented by linear interpolation of the labels. Eq. 5 can then be converted to:", + "bbox": [ + 496, + 633, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {L} _ {\\text {c o n t}} = \\lambda d _ {f} (a, x _ {H}) ^ {2} + (1 - \\lambda) [ m - d _ {f} (a, x _ {H}) ] _ {+} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 719, + 890, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, $x_{H}$ appears with nonzero contributions in both positive and negative terms for positive-negative pairs, due to the interpolation factor $\\lambda \\in [0,1]$ .", + "bbox": [ + 496, + 750, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, we found that these methods are only applicable to mixup and may be subject to controversy in other forms of mixed sample data augmentations. This is due to the fact that mixup binds the images together evenly based on the assigned weight parameter, $\\lambda$ . The same applies to labels In contrast, cutmix synthesizes $x_{H}$ by replacing a crop area of $x_{A}$ with that of $x_{B}$ , where the crop area of $\\lambda$", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11050", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7506c00de0f164608a0b354c85d165b7cac011f3a9bf0a53d902c7fa19fc65e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HS-AnorhWP-AnorhN-Anorh
Intra-class variation✓✓✕
Inter-class variation✓✓✓
Conflict✕✓✕
", + "bbox": [ + 130, + 88, + 416, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Three types of anchor point selection are considered for inter-class and intra-class variation.", + "bbox": [ + 76, + 161, + 468, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $(1 - \\lambda)$ is taken from $x_{A}$ and $x_{B}$ , respectively. Therefore, the label of $x_{H}$ cannot be accurately represented by linear interpolation, as the area being replaced is indeterminate, and may contain either significant features or irrelevant content. Consequently, the contribution of the crop area to $x_{H}$ remains inconclusive, making the application of linear interpolation for labeling mostly infeasible.", + "bbox": [ + 75, + 215, + 468, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of fully utilizing the label information of $x_{H}$ , we can translate them into their appropriate spatial embedding, as we will describe in next subsection. Also, when applying MSDA to images, we need to consider how the labels are to be processed, which limits us to intermixing only two samples. In contrast, $x_{H}$ does not require us to define label information, and can hence be mixed by multiple classes.", + "bbox": [ + 75, + 321, + 468, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Hybrid loss", + "text_level": 1, + "bbox": [ + 76, + 440, + 202, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Anchor Point Selection and Processing", + "text_level": 1, + "bbox": [ + 76, + 465, + 400, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although it is challenging to accurately assign a label to $x_{H}$ obtained through MSDA, the features of $x_{H}$ can be inferred based on the source images. We define $C(x_{I})$ to represent the label of image $x_{I}$ . When $x_{H}$ is selected as an anchor, it may not be possible to find a suitable positive sample, but $x_{H}$ shares semantic information with $C(x_{A})$ and $C(x_{B})$ to some extent. Thus, we can consider $C(x_{A})$ and $C(x_{B})$ as weak positive samples (WP) of $x_{H}$ , with the remaining classes treated as negative samples.", + "bbox": [ + 75, + 491, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training phase, the embedding of the spatial position of $x_{H}$ plays a crucial role in the generalization of downstream tasks. Therefore, deciding how to treat $x_{H}$ in pair selection becomes the first key question. Since $x_{H}$ contains information for multiple classes, including it in a paired selection can provide more gradient information. We divided the pairing methods into three categories, as shown in Figure. 4: HS (hybrid species sample) as the anchor point, WP (weak positive sample) as the anchor point, and N (negative sample) as the anchor point. We compared the three methods and and summarized the findings in Table 1.", + "bbox": [ + 75, + 628, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The choice of anchor point in a classification task can affect the learning of intra-class and inter-class variation. When using $x_{H}$ as an anchor point, it pulls towards the WP while staying away from the rest of the class. Using WP as an anchor point keeps it close to its composite $x_{H}$ to maintain identical points within the class, but may result in conflicts with the synthetic sample of $x_{H}$ leading", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to non-converging losses. Anchoring with N treats $x_{H}$ as a negative sample and moves it away from other samples, losing its purpose. Our findings recommend using HS as the anchor sample to effectively focus on both intra-class and inter-class variation while ensuring convergence.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Strategies for Sample Mining", + "text_level": 1, + "bbox": [ + 500, + 185, + 758, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the previous subsection, we analyzed the contribution of pairwise selection to the embedding of $x_{H}$ . Since $x_{H}$ is randomly generated in each epoch, how to embed it more efficiently and reasonably becomes the second key issue of our study. For instance, when $x_{H}$ is used as an anchor point, the corresponding WP may contain multiple classes. If $x_{H}$ pulls all WP close together, clusters that are unrelated may merge, which is unreasonable. If $x_{H}$ randomly selects WP that are far apart, the structure of the cluster may be destroyed. On the other hand, if negative samples that are far apart are selected, the resulting gradient will be smaller. In summary, it is vital that we choose a reasonable and valid sample for performing metric learning with HS.", + "bbox": [ + 496, + 208, + 890, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Several studies have explored the advantages of hard negative mining in constructing informative gradients [12, 36, 34]. In this context, we introduce the notion of hard negatives for the hybrid species as:", + "bbox": [ + 496, + 405, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {H _ {h n}} = \\underset {x: C (x _ {I}) \\neq C (x _ {A}) \\cap C (x _ {I}) \\neq C (x _ {B})} {\\operatorname {a r g m i n}} d \\left(f \\left(x _ {H}\\right), f \\left(x _ {I}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 476, + 890, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting that $x_{H_{hn}}$ does not only yield a large gradient for $x_{H}$ , but it also possesses the same potential gradient as the inputs $C(x_{A})$ and $C(x_{B})$ that are used to synthesize $x_{H}$ .", + "bbox": [ + 496, + 517, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The study by [43] shows that easy positive mining improves the generalization performance of downstream tasks. Our experiments also demonstrate the benefits of this approach for $x_{h}$ . We define the easy hybrid weak positive of the hybrid species as:", + "bbox": [ + 496, + 577, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {H _ {e w p}} = \\underset {x: C \\left(x _ {I}\\right) = C \\left(x _ {A}\\right) \\cup C \\left(x _ {I}\\right) = C \\left(x _ {B}\\right)} {\\operatorname {a r g m i n}} d \\left(f \\left(x _ {H}\\right), f \\left(x _ {I}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 664, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are two primary reasons why bringing $x_{H}$ and $x_{H_{ewp}}$ closer together is beneficial. Firstly, it enables the model to concentrate on the features that are shared by both inputs, which can significantly influence the sample measurements. Secondly, this approach avoids the potential issue of pushing weak positives that are distant from each other too close together. By doing so, it preserves both the intra-class variance and the inter-class variance, which can help maintain the manifold structure of the classes in the embedding space.", + "bbox": [ + 496, + 704, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our hybrid loss function involves mapping the output of the convolutional neural network onto a unit sphere, which is a widely accepted technique. During each mini-batch,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "1 # Nsize: Number of samples of the same class in each batch \n2 # n: The number of classes that synthesize $x_{H}$ \n3 # b: batch size d: dimension \n4 # F = b × d \n5 # Map of feature similarity between samples \n6 Map = F × FT; \n7 # $x_{H}$ as the anchor, look for weak positive sample \n8 DWP = Map; \n9 for each i ∈ [b - Nsize, b] do \n10 for each j ∈ [n * Nsize, b] do \n11 | DWP[i][j] = -1; \n12 end \n13 end \n14 # Each $x_{H}$ corresponds to the value and index of the most similar WP sample \n15 VDWP, IDWP = DWP.max(1); \n16 # $x_{H}$ as the anchor, look for negative sample \n17 DN = Map; \n18 for each i ∈ [b - Nsize, b] do \n19 for each j ∈ [0, n * Nsize] do \n20 | DN[i][j] = -1; \n21 end \n22 for each j ∈ [b - Nsize, b] do \n23 | DN[i][j] = -1; \n24 end \n25 end \n26 # Each $x_{H}$ corresponds to the value and index of the most similar negative sample \n27 VDN, IDN = DN.max(1); \n28 # Select the WP sample and negative sample parts corresponding to $x_{H}$ \n29 T = [VDWP[-Nsize:], VDN[-Nsize:]]; \n30 Hybrid loss = -α logsoftmax(T).mean();", + "bbox": [ + 80, + 95, + 444, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1: Hybrid Loss", + "text_level": 1, + "bbox": [ + 161, + 618, + 341, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "when $x_{H}$ serves as the anchor, we select $x_{H_{ewp}}$ and $x_{H_{hn}}$ for comparison. As in Algorithm 1, the former corresponds to the original sample, while the latter represents $x_{H}$ . To synthesize $x_{H}$ , we typically select samples from the first $n$ classes. Then, we search for the most similar weak positive sample and negative sample for $x_{H}$ , respectively.", + "bbox": [ + 75, + 660, + 468, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The NCA loss function is an effective way to avoid the selection of margin hyperparameters and handle the distance between samples efficiently[46]. To ensure that hybrid species are embedded in a suitable location, we define the hybrid loss based on the NCA loss:", + "bbox": [ + 75, + 751, + 468, + 825 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {H y} = - \\alpha \\log \\frac {1}{1 + \\frac {\\exp (f (x _ {H}) ^ {\\top} f (x _ {H _ {h n}}))}{\\exp (f (x _ {H}) ^ {\\top} f (x _ {H _ {e w p}}))}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 835, + 468, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyper-parameter that controls the balance be", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tween bringing $x_{H}$ and $x_{H_{ewp}}$ closer and pulling $x_{H}$ and $x_{H_{hn}}$ farther apart. The $L_{Hy}$ loss serves as an auxiliary loss to process additional samples and can be combined with other metric learning losses. As in Figure 3, the final loss is the sum of the metric learning loss $L_{m}$ and the $L_{Hy}$ loss:", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {m} + L _ {H y} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 183, + 890, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When incorporating our method with the contrastive loss (Eq. 2), the resulting final loss is:", + "bbox": [ + 500, + 207, + 890, + 237 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {c o n t}} + L _ {H y} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 250, + 890, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 280, + 632, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our HSE method uses pre-trained BN-Inception [14] and ResNet [11] models on ILSVRC 2012-CLS [32] with PyTorch. These models have learned feature representations from diverse images, reducing time and resources for new dataset training. They also perform well in image recognition tasks, making them suitable for computer vision applications. For generalization ability evaluation, we maintain consistency with base methods on image preprocessing, learning rate, and batch size. (Source code: https://github.com/SHQberserker/HSE)", + "bbox": [ + 496, + 306, + 890, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nsize examples from each class are selected per batch. We use two methods to incorporate hybrid species during training: selecting multiple classes to add hybrid species as additional examples of batch, and randomly replacing examples in the batch with hybrid species. We use the cutmix technique to create hybrid species by stitching two samples together evenly. Algorithm 1 fixes the last Nsize samples of each batch as synthetic samples.", + "bbox": [ + 496, + 458, + 890, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 500, + 589, + 697, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We tested our framework's ability to perform well on challenging datasets: CUB-200, CARS-196, and SOP. These datasets involve fine-grained classification tasks, where there are many categories and relatively few images per category. We followed the data split approach from [22]. Specifically, for CUB-200, we used 5864 images from the first 100 categories for training and 5924 images from the last 100 categories for testing. For CARS-196, we trained on the first 98 classes and tested on the remaining images. Additionally, for SOP, we utilized the first 59,551 images (11,318 classes) for training and the remaining 60,502 images (11,316 classes) for testing.", + "bbox": [ + 496, + 612, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess our experiments, we employed the Recall@K metrics. This measure involves selecting the K images that are closest to the query image. If there's an image with the same label as the query image, we score 1; otherwise, it's 0. Recall@K calculates the average score for all query images, with higher values indicating better retrieval performance. These metrics provide a way to quantitatively evaluate how", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11052", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "well our model can find similar images, which is crucial for many image-based applications.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 76, + 128, + 272, + 143 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Comparison", + "text_level": 1, + "bbox": [ + 76, + 152, + 217, + 166 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments compared our proposed method to existing methods, including N-pair, EPSHN, MSloss and ProxyAnchor. Results are shown in Table 2 for CUB-200 and CARS-196 datasets, where * indicates our results under the same experimental settings. Our method improves on different approaches to varying degrees. Our method had less impact on the N-pair method as it mostly ignores intraclass variation, leading to excessive clustering between the same classes. However, we observed a significant improvement in recall@K when the EPSHN method was augmented with the additional training signal of HSE. Recall@1 improved by $2.5\\%$ and $2.3\\%$ on the challenging CUB-200 and CARS-196 datasets, respectively, where Nsize=16 and Nsize=8, respectively. Importantly, HSE-E outperformed the ensemble-based method that EPSHN had not surpassed before on the CUB-200 and CARS-196 datasets.", + "bbox": [ + 75, + 176, + 468, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We believe that EPSHN tends to ignore intra-class variations, such as those found in [43] where birds of different sexes may vary greatly within the same class, but should also focus on characteristics common between members of the same class. When $x_{H}$ is added, it helps the network to focus more on these common characteristics. MSloss focuses on both inter-class variation and inter-class variation. However, when adding our additional training signals, there is still a certain improvement. This is because the features that our additional training signals focus on help the network generalize to unknown classes.", + "bbox": [ + 75, + 417, + 468, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method achieves good results in pair-based methods and is effective in proxy-based metric learning methods. We improve approximately $1\\%$ in both $\\mathrm{Recall}@1$ and $\\mathrm{Recall}@2$ . This confirms our universality, which can be applied to various metric learning loss functions. We also compared our method to the Matrix/Input [38] approach. Matrix/Input has a high computational cost because it calculates $\\frac{1}{2} n(n - 1)$ embeddings for a batch of $n$ examples. Instead, our method achieves a similar effect using only $Nsize * Cnum$ embeddings. Our approach is hence more computationally efficient. We also remain applicable to discriminate between positives and negatives for every anchor point. We can also incorporate non-additive component functions that involve loss functions, e.g., EPSHN.", + "bbox": [ + 75, + 584, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HSE is also effective on the SOP dataset. Unlike smaller datasets, the SOP dataset has superclasses, which introduce more variation between different superclasses and less variation within the same superclass. It's challenging to measure similarity between samples within the same superclass in the SOP dataset. As shown in Table 3, we generate HS from samples that share the same superclass but have differ", + "bbox": [ + 75, + 795, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ent labels. When we apply HSE to various classical methods, we observe varying degrees of improvement.", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Impact of Minibatch Parameters", + "text_level": 1, + "bbox": [ + 500, + 141, + 782, + 156 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Nsize", + "text_level": 1, + "bbox": [ + 500, + 165, + 612, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We trained a ResNet-18 embedding network on CUB-200 and CARS-196 datasets with batch size 128 and embedding dimension 512 to investigate HSE's impact on model generalization at various Nsizes. HSE's effectiveness varies with methods, datasets, and Nsizes (Figure. 5). The best performance occurs at Nsize between 4 and 16 with a batch size of 128, followed by performance degradation. When there are too few negative samples, the hybrid species' gradient information becomes insignificant, and they may be grouped with WP's clusters, harming downstream tasks.", + "bbox": [ + 498, + 188, + 890, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ab7a647d9ff83c4f2c2886e6e56a38787378f6d57445c7bdba40a062c6741cf.jpg", + "image_caption": [ + "(a) CUB" + ], + "image_footnote": [], + "bbox": [ + 524, + 352, + 661, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c1703200f28fff10e20d066a809a03603f67910656e2ddfb97b9ff9066a063c.jpg", + "image_caption": [ + "(b) CAR" + ], + "image_footnote": [], + "bbox": [ + 712, + 352, + 846, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d3bb9bfec150f1eb655e584af25b56eadd459f0c7714e5cc28d71cf2e6d5480.jpg", + "image_caption": [ + "Figure 5: Effect of additional training signals provided by different Nsizes.", + "Figure 6: Results on the CUB dataset using different Cnum in MS loss." + ], + "image_footnote": [], + "bbox": [ + 614, + 494, + 767, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Cnum", + "text_level": 1, + "bbox": [ + 500, + 621, + 619, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We found that HSE improves the generalization ability of the model. However, increasing the number of hybrid species does not necessarily lead to better results. We conducted experiments on the CUB dataset using the MS loss method, as shown in Figure. 6. Our loss function indirectly affects the entire embedding space of the batch when learning the spatial embedding of a single hybrid species. After training for a certain period, the two combined classes have essentially covered the entire dataset.", + "bbox": [ + 496, + 643, + 890, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Effect of Hybrid Species", + "text_level": 1, + "bbox": [ + 500, + 800, + 723, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigated the effectiveness of different data augmentation techniques in providing additional training signals for hybrid species. Table 4 shows the classic blending data augmentation techniques we used, with certain modifications made to each method. Notably, CutMix and GridMask", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c7ca4b947d2d47d9afd1036c2483400f3f53871dbee2bf3ef14aee4d5f18a964.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CUB-200CARS-196
Recall@kk=1k=2k=4k=1k=2k=4
Angular512[39]G53.665.075.371.380.787.0
HDML512[49]G53.765.776.779.187.192.1
HTL512[6]BN57.168.878.781.488.092.7
Margin128[41]R63.674.483.179.686.591.9
SoftTriple512[28]BN65.476.484.584.590.794.5
DR512[24]BN66.177.085.185.090.594.1
Ensemble-based
HDC384[44]G53.665.777.073.282.486.4
A-BIER512[27]G55.367.276.978.085.891.1
ABE-8512[16]G60.671.579.860.671.579.8
DREML9216[42]R63.975.083.186.091.795.0
Npair64*(Nsize=8)[37]R51.364.275.364.276.084.6
HSE-N64(Nsize=8)R52.5+1.266.0+1.877.4+2.164.9+0.776.6+0.684.7+0.1
EPSHN512*(Nsize=16)[43]R64.975.383.583.189.793.6
HSE-E512(Nsize=16)R66.9+2.077.4+2.185.5+2.085.4+2.391.2+1.596.9+3.3
HSE-E512(Nsize=8)R67.4+2.577.7+2.485.7+2.284.8+1.790.6+0.994.3+0.7
MS†512R67.877.885.687.892.795.3
+Matrix输入[38]R69.0+1.279.1+1.386.0+0.489.0+1.293.4+0.796.0+0.7
MS512*(Nsize=5)[40]BN65.777.086.381.788.793.2
HSE-M512(Nsize=5)BN67.6+1.978.0+1.085.8-0.582.0+0.388.9+0.293.3+0.1
Proxy-Ancho†512R69.579.387.087.692.395.5
+Matrix输入[38]R70.5+1.081.2+1.987.8+0.888.2+0.693.2+0.996.2+0.7
Proxy-Ancho*512[15]R69.479.287.088.592.795.6
HSE-PA512R70.6+1.280.1+0.987.1+0.189.6+1.193.8+1.196.0+0.4
", + "bbox": [ + 187, + 88, + 777, + 416 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5917b126e06f59ed4b7d47ac4823ae61f935cd67d1bbd3c5783435b60424628f.jpg", + "table_caption": [ + "Table 2: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the CUB-200-2011 and Cars-196 datasets. The backbone networks used in the models are denoted by abbreviations: G for GoogleNet, BN for Inception with batch normalization, and R for ResNet50. The symbol $\\dagger$ indicates the reproduced result reported by the original authors.N-pair and EPSHN use cnum=1 and batch size of 128. MSloss uses cnum=1, batch size of 80, and $\\alpha = 0.05$ . Proxy-Anor uses cnum=1, batch size of 128, and $\\alpha = 2$ ." + ], + "table_footnote": [], + "table_body": "
SOP
Recall@kk=1k=10k=100
EPSHN512*(Nsize=5) [43]R74.386.994.3
HSE-E512(Nsize=5)R76.3+2.088.4+1.594.9+0.6
MS†512R76.989.895.9
+Matrix输入[38]R77.9+1.090.6+0.895.9
MS512*[40] (Nsize=5)R78.290.296.2
HSE-M512R78.7+0.590.4+0.296.1-0.1
Proxy-Anchort512R79.190.896.2
+Matrix输入[38]R79.8+0.791.4+0.696.5+0.3
Proxy-Anchort*512[15]R79.290.495.8
HSE-PA512(Nsize=5)R80.0+0.891.4+1.096.3+0.5
", + "bbox": [ + 78, + 500, + 465, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stand out. CutMix can incorporate features from each compositional image to a certain extent, but may also diminish key features. Mixup dilutes key features as the number of samples increases. GridMask (block=50) covers the entire image despite using relatively small blocks. Results show that CutMix is the most effective approach for improving the performance of Hybrid Species (HS). We also found that adding more class features does not necessarily improve HS performance, but adding critical features that are specific to a particular class can lead to better HS results.", + "bbox": [ + 75, + 733, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the selection of weak positive samples", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f97519d9105ad5bd32a18a82c412875a7a15b8a6576634b5697676b7b0332d4.jpg", + "table_caption": [ + "Table 3: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the SOP.Where batch $= {120}$ ,cnum $= {12}$ ,and hybrid species as additional examples of batch." + ], + "table_footnote": [], + "table_body": "
CUBCAR
R@1numR@1num
Cutmix67.4285.42
67.2385.03
67.1484.24
Mixup67.0285.32
66.9381.03
R@1MasksizeR@1Masksize
Gridmask67.45084.950
66.015084.8150
67.125084.0250
", + "bbox": [ + 519, + 500, + 867, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Results on the Cub dataset using different Mixed Sample Data Augmentations in EPSHN. $(\\mathbb{R}@\\mathbb{1} = \\mathbb{Recall}@\\mathbb{1})$", + "bbox": [ + 500, + 691, + 890, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "by hybrid species using the Proxy-Anorh method on the CUB dataset. Cutmix augmentation technique is used to stitch images either vertically or horizontally, with the upper part mainly containing the bird's head and the lower part containing the bird's body when stitched vertically. When stitched horizontally, the probability of the left and right sides containing the bird's head is almost the same. Our HSE approach improves model performance by focusing on essential image features. Our study on the CUB dataset demonstrates that the hybrid species prioritize the bird's", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11054", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2816ec965166d5a44ddac96594c047db7957582107bec461f7f4fc1b48d98608.jpg", + "image_caption": [ + "Figure 7: Analysis of easy weak positive samples selected by hybrid species using Proxy-Antor on CUB dataset. Hybrid species are obtained by stitching two images of different labels using cutmix. We quantify the probability that selected easy weak positive samples contain upper or bottom half during training using \"aut_up/down\". With our method, we also report \"our_up/down\" and \"our_left/right\" indicating the likelihood of easy weak positive samples containing upper or bottom half of the bird or left or right half of the stitched image, respectively." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 415, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "body during metric learning, particularly the lower half, which is synthesized as the easy weak positive sample using the Cutmix augmentation technique. While the selection probability of the lower half increases from 0.6 to $0.7\\sim 0.8$ during Proxy training, our proposed method ensures a probability of 1. Stitching hybrid species horizontally results in a selection probability of around 0.5. This approach allows the network to focus on important features during training, providing an advantage over traditional methods. Figure. 7 provides a visual representation of our findings.", + "bbox": [ + 75, + 398, + 472, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2.4 Visualization of Embedding Space", + "text_level": 1, + "bbox": [ + 76, + 571, + 370, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of our proposed method on training and test data using t-SNE (t-distributed stochastic neighbor embedding) visualization technique. t-SNE can effectively reduce the dimensionality of the embedding space while preserving pairwise similarity between samples. It enables us to visually inspect how well the model separates different classes and identifies clusters of similar samples by plotting the embedding space in 2D or 3D. Therefore, t-SNE is a useful tool for evaluating the quality of learned embeddings in deep metric learning.", + "bbox": [ + 75, + 597, + 468, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We experimented on two classes from the training set and two classes (wren and woodpecker) from the test set. Figure. 8 shows that EPSHN is effective at distinguishing between different classes in the training set, but the embedding for hybrid species is relatively dispersed. With the additional HSE training signal, the embedding of the hybrid species becomes more concentrated and is located between the two classes. This approach considers all the features of difficult samples, creating a more comprehensive embedding of the samples. When additional training signals", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "are learned, the model can generalize better to unknown classes. This is demonstrated in the test set, where the EPSHN method cannot distinguish the black classes from the green and blue classes. However, when the training signal of HSE is applied to the downstream task, black classes can be successfully distinguished from green classes.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ba07d0fc9327cface7ddee83aab881ce6f9d14292d11471cd69e6b5fde50df37.jpg", + "image_caption": [ + "(a) Oriole $(\\mathrm{id} = 95)$" + ], + "image_footnote": [], + "bbox": [ + 517, + 195, + 586, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/50b0c5aa910ac598697179d39464f988ccd9c826c9adcea56fbad2b1c9da570f.jpg", + "image_caption": [ + "(b) Nuthatch (id=94)" + ], + "image_footnote": [], + "bbox": [ + 661, + 195, + 730, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a5787db3e0af7e6b248afe79b3790ffffc34d77409d977b70c786fd3aed4b978.jpg", + "image_caption": [ + "(c) Hspecie" + ], + "image_footnote": [], + "bbox": [ + 813, + 191, + 869, + 233 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b737f442c716e23eac9a6cf31d86497b3e7d55ffe9a104c89bba79382c6b3970.jpg", + "image_caption": [ + "EPSHN (train)" + ], + "image_footnote": [], + "bbox": [ + 550, + 270, + 647, + 345 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fd43c4924038cbe776a1120c63747384b1c13ba4f0bcce3bfc6cc8987ac4503f.jpg", + "image_caption": [ + "HSE-EPSHN (train)" + ], + "image_footnote": [], + "bbox": [ + 748, + 271, + 844, + 344 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b418f8864e9908fbb992bce2ca470bee34251594dce4113ae98c072e5f8a7ded.jpg", + "image_caption": [ + "EPSHN (test)" + ], + "image_footnote": [], + "bbox": [ + 549, + 388, + 647, + 462 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2f63165eb38949092075af43c9622c4acc9d0bf2f62bee5c73937eb4751f9043.jpg", + "image_caption": [ + "HSE-EPSHN (test)", + "Figure 8: We present a t-SNE visualization of EPSHN and HSE-EPSHN on the test set of CUB. Specifically, we select two different classes (denoted by blue dots and green dots) from the training set as shown in (a) and (b), respectively. (c) is the synthetic class from classes (a) and (b), denoted by red dots. For the test set, we selected four bird categories, consisting of two types of wrens and two types of woodpeckers, where each color represents a specific category." + ], + "image_footnote": [], + "bbox": [ + 743, + 390, + 843, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 617, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our Hybrid Species Embedding (HSE), a metric learning approach that aims to enable models to learn the spatial embedding of unknown and difficult classes during training. HSE generates additional hard samples using mixed data augmentation methods without defining their label information and applies them to the metric learning process. Experiments demonstrate significant improvements on multiple datasets when compared to three widely used pair-based methods and state-of-the-art proxy methods. HSE enhances the generalization ability of the model by providing additional training signals in deep metric learning.", + "bbox": [ + 496, + 643, + 890, + 825 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment. This work is partly supported by the National Natural Science Foundation of China (No: 62172366) and the Pioneer and Leading Goose R&D Program of Zhejiang Province (2023C01150).", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] E. Ahmed, M. Jones, and T. K. Marks. An improved deep learning architecture for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3908–3916, 2015.", + "[2] P. Chen, S. Liu, H. Zhao, and J. Jia. Gridmask data augmentation. arXiv preprint arXiv:2001.04086, 2020.", + "[3] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "[4] T. DeVries and G. W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017.", + "[5] Z. Fu, Z. Mao, B. Hu, A.-A. Liu, and Y. Zhang. Intra-class adaptive augmentation with neighbor correction for deep metric learning. IEEE Transactions on Multimedia, 2022.", + "[6] W. Ge. Deep metric learning with hierarchical triplet loss. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-285, 2018.", + "[7] G. Gu, B. Ko, and H.-G. Kim. Proxy synthesis: Learning with synthetic classes for deep metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1460-1468, 2021.", + "[8] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006.", + "[9] E. Harris, A. Marcu, M. Painter, M. Niranan, A. Prügel-Bennett, and J. Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020.", + "[10] B. Harwood, V. Kumar BG, G. Carneiro, I. Reid, and T. Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 2821-2829, 2017.", + "[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[12] A. Hermans, L. Beyer, and B. Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017.", + "[13] E. Hoffer and N. Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015.", + "[14] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015.", + "[15] S. Kim, D. Kim, M. Cho, and S. Kwak. Proxy anchor loss for deep metric learning. IEEE, 2020.", + "[16] W. Kim, B. Goyal, K. Chawla, J. Lee, and K. Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018.", + "[17] B. Ko and G. Gu. Embedding expansion: Augmentation in embedding space for deep metric learning. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7255-7264, 2020.", + "[18] B. Ko, G. Gu, and H.-G. Kim. Learning with memory-based virtual classes for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11792-11801, 2021.", + "[19] M. T. Law, N. Thome, and M. Cord. Quadruplet-wise image similarity learning. In Proceedings of the IEEE international conference on computer vision, pages 249-256, 2013.", + "[20] J.-H. Lee, M. Z. Zaheer, M. Astrid, and S.-I. Lee. Smoothmix: a simple yet effective data augmentation to train robust classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 756-757, 2020.", + "[21] L. Liu, S. Huang, Z. Zhuang, R. Yang, M. Tan, and Y. Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022.", + "[22] Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016.", + "[23] T. Milbich, K. Roth, H. Bharadhwaj, S. Sinha, Y. Bengio, B. Ommer, and J. P. Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In European Conference on Computer Vision, pages 590-607. Springer, 2020.", + "[24] D. D. Mohan, N. Sankaran, D. Fedorishin, S. Setlur, and V. Govindaraju. Moving in the right direction: A regularization for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14591-14599, 2020.", + "[25] Y. Movshovitz-Attias, A. Toshev, T. K. Leung, S. Ioffe, and S. Singh. No fuss distance metric learning using proxies. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017.", + "[26] H. Oh Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016.", + "[27] M. Opitz, G. Waltner, H. Possegger, and H. Bischof. Deep metric learning with bier: Boosting independent embeddings robustly. IEEE transactions on pattern analysis and machine intelligence, 42(2):276-290, 2018.", + "[28] Q. Qian, L. Shang, B. Sun, J. Hu, H. Li, and R. Jin. Soft-triple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019.", + "[29] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "[30] K. Roth, O. Vinyals, and Z. Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022." + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11056", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] K. Roth, O. Vinyals, and Z. Akata. Non-isotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022.", + "[32] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015.", + "[33] A. Sanakoyeu, V. Tschernezki, U. Buchler, and B. Ommer. Divide and conquer the embedding space for metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 471-480, 2019.", + "[34] F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015.", + "[35] C. Shorten and T. M. Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019.", + "[36] E. Simo-Serra, E. Trulls, L. Ferraz, I. Kokkinos, P. Fua, and F. Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015.", + "[37] K. Sohn. Improved deep metric learning with multi-class n-pair loss objective. Advances in neural information processing systems, 29, 2016.", + "[38] S. Venkataramanan, B. Psomas, Y. Avrithis, E. Kijak, L. Amsaleg, and K. Karantzalos. It takes two to tango: Mixup for deep metric learning. arXiv preprint arXiv:2106.04990, 2021.", + "[39] J. Wang, F. Zhou, S. Wen, X. Liu, and Y. Lin. Deep metric learning with angular loss. In Proceedings of the IEEE international conference on computer vision, pages 2593-2601, 2017.", + "[40] X. Wang, X. Han, W. Huang, D. Dong, and M. R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5022-5030, 2019.", + "[41] C.-Y. Wu, R. Manmatha, A. J. Smola, and P. Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017.", + "[42] H. Xuan, R. Souvenir, and R. Pless. Deep randomized ensembles for metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 723-734, 2018.", + "[43] H. Xuan, A. Stylianou, and R. Pless. Improved embeddings with easy positive triplet mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2474-2482, 2020.", + "[44] Y. Yuan, K. Yang, and C. Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE international conference on computer vision, pages 814-823, 2017.", + "[45] S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo. Cutmix: Regularization strategy to train strong classifiers with" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019.", + "[46] D. Zhang, Y. Li, and Z. Zhang. Deep metric learning with spherical embedding. Advances in Neural Information Processing Systems, 33:18772-18783, 2020.", + "[47] H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017.", + "[48] Y. Zhao, Z. Jin, G.-j. Qi, H. Lu, and X.-s. Hua. An adversarial approach to hard triplet generation. In Proceedings of the European conference on computer vision (ECCV), pages 501-517, 2018.", + "[49] W. Zheng, Z. Chen, J. Lu, and J. Zhou. Hardness-aware deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 72-81, 2019.", + "[50] W. Zheng, C. Wang, J. Lu, and J. Zhou. Deep compositional metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9320-9329, 2021.", + "[51] W. Zheng, B. Zhang, J. Lu, and J. Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 431 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11057", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_model.json b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35da7ac2434d085f0b37ed6fa2ba4b39a3f3fb59 --- /dev/null +++ b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_model.json @@ -0,0 +1,2895 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.787, + 0.154 + ], + "angle": 0, + "content": "HSE: Hybrid Species Embedding for Deep Metric Learning" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.18, + 0.862, + 0.201 + ], + "angle": 0, + "content": "Bailin Yang\\(^{1,2}\\), Haoqiang Sun\\(^{1}\\), Frederick W. B. Li\\(^{3}\\), Zheng Chen\\(^{1}\\), Jianlu Cai\\(^{1}\\), Chao Song\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.205, + 0.8, + 0.224 + ], + "angle": 0, + "content": "\\(^{1}\\)Faculty of Computer Science and Technology, Zhejiang Gongshang University" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.224, + 0.767, + 0.242 + ], + "angle": 0, + "content": "\\(^{2}\\)Faculty of Statistics and Mathematics, Zhejiang Gongshang University" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.242, + 0.713, + 0.26 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Computer Science, University of Durham" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.295, + 0.313, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.522 + ], + "angle": 0, + "content": "Deep metric learning is crucial for finding an embedding function that can generalize to training and testing data, including unknown test classes. However, limited training samples restrict the model's generalization to downstream tasks. While adding new training samples is a promising solution, determining their labels remains a significant challenge. Here, we introduce Hybrid Species Embedding (HSE), which employs mixed sample data augmentations to generate hybrid species and provide additional training signals. We demonstrate that HSE outperforms multiple state-of-the-art methods in improving the metric Recall@K on the CUB-200, CAR-196 and SOP datasets, thus offering a novel solution to deep metric learning's limitations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.551, + 0.21, + 0.567 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.728 + ], + "angle": 0, + "content": "Image retrieval heavily depends on deep metric learning to grasp visual similarities. In its early stages, methods like pair-based loss [8, 13, 19, 37] and proxy-based loss [25, 15] were proposed to train models. Nevertheless, the true goal of image retrieval is to adapt to the unknown, and the challenge lies in selecting a metric that can effectively handle differences between classes in deep metric learning. A promising approach involves exploring both intra-class and inter-class variations within the image itself, rather than relying solely on label information [33, 43, 40, 23, 51, 30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Hard negative samples, often termed false positives, refer to images that resemble anchor images but carry different labels. The reason behind these hard samples lies in their potential for sharing remarkably similar or even identical features with the anchor samples. Approaches like hard sample mining and generation have been proposed to aid network convergence by introducing a substantial gradient [1, 41, 12, 34, 49, 48, 17, 18]. Recently, an advancement in generating appropriate hard samples involves creating supplementary training data [17, 38, 5, 21]. This is primarily" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.296, + 0.832, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.337, + 0.688, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.421, + 0.662, + 0.433 + ], + "angle": 0, + "content": "(a) Linear interpolation" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.338, + 0.882, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.422, + 0.829, + 0.433 + ], + "angle": 0, + "content": "(b) Our HSE" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.572 + ], + "angle": 0, + "content": "Figure 1: (a) Deep metric learning methods struggle with linear interpolation for generating new training samples with labels when new samples fall between anchors of different classes. The manifold structure lacking clear boundaries makes it challenging to assign a label, even when new samples are between anchors of the same class. (b) Samples with multiple simultaneous features may not lie on a linear interpolation path, and their location may be unknown. They can be classified as an unknown class situated inside or outside a manifold boundary. To tackle this challenge, we propose using the HSE method to allocate the sample to an appropriate embedding space location based on its feature information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.81 + ], + "angle": 0, + "content": "achieved through linear interpolation, a prevalent method for generating synthetic samples. However, assigning absolute labels to these synthesized samples is intricate due to the manifold structures and cluster boundaries present in the embedded space [17]. Even when the new samples are formed by linear interpolation between samples of the same label, accurately determining absolute labels remains a challenge. Synthetic sample labels have emerged as a potential solution, with [38] corroborating their similarity to mixup data augmentation and highlighting the label's representation through linear interpolation. Nonetheless, this approach is constrained to mixup data augmentation, as mixup uniformly blends features and labels from different images, which might not hold true for other techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We introduce a novel Hybrid Species Embedding (HSE) that leverages mixed sample data augmentations to generate additional training samples [35, 47, 45, 2, 20]. HSE creates hybrid species by combining features from multiple classes and embedding them into spatial positions. During training, different classes are randomly chosen for data augmentation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.341, + 0.901 + ], + "angle": 0, + "content": "*Corresponding to:csong@zjsu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11047" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "in each batch. The labels of hybrid species are disregarded, and their placement near the category with the most similar features simulates human categorization behavior, as depicted in Figure 1(b). This strategy positions hybrid species close to the samples they are synthesized from, ensuring similarity while avoiding unrelated samples. By synthesizing samples using various mixed sample data augmentation techniques, each incorporating multiple features of a known class, our approach naturally imparts similar characteristics to the hard sample for the model without requiring explicit mining. Additionally, we dynamically adjust the position of each batch of hybrid species embedding to bypass the need for considering label information of these hybrid instances. Our contributions encompass:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.303, + 0.47, + 0.379 + ], + "angle": 0, + "content": "- Introducing a novel metric learning strategy embedding challenging and unseen classes (Hybrid species), during training. These hybrid species offer supplementary training signals, enhancing the generalization capacity of downstream tasks, as validated through experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.47, + 0.439 + ], + "angle": 0, + "content": "- Adapting mixed sample data augmentation techniques for our Hybrid Species Embedding (HSE) to address constraints associated with pair-based metric learning losses, which typically demand explicit class labels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.439, + 0.469, + 0.484 + ], + "angle": 0, + "content": "- Demonstrating through experiments our efficacy in enhancing performance across state-of-the-art metric learning tasks on CUB-200, CAR-196, and SOP datasets." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.303, + 0.47, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.497, + 0.22, + 0.512 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.644 + ], + "angle": 0, + "content": "We utilize the Mixed Sample Data Augmentation (MSDA) method to create unknown samples and improve the model's generalization ability for downstream tasks. This approach resolves the issue of insignificant comparisons caused by random selection of samples in metric learning. While mining hard samples from a finite training set is limited, MSDA enables us to create more hard samples. In the following, we review existing relevant works." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.651, + 0.242, + 0.667 + ], + "angle": 0, + "content": "2.1. Metric Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Deep metric learning compares data in pairs and pulls them together if they belong to the same category, otherwise separates them [8, 13, 19, 37, 26]. The exponential growth in the number of tuples during training can present challenges in achieving model convergence, and may include meaningless metrics. Proxy-based approaches, e.g., choosing a proxy point for each class, have been proposed as a solution [25, 15]. These proxies contain information from multiple samples, require less computation in pairwise comparisons, helping the network converge more effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Methods that rely solely on label information may result in over-clustered or overly separated samples with identical or differing labels, respectively, and can lead to reduced performance on new and unseen test samples. EPSHN [43] notes this problem, particularly for species with varying" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "features such as birds of the same species with different sexes. To avoid this, [30, 40, 31, 33, 51, 50, 18] emphasize the importance of incorporating semantic information from the images themselves, often through large pre-trained language models like CLIP [29] and BERT [3], to better align the visual representation space and discover potential differences within and between classes. The MSloss function [40] is one such approach that combines self-similarity and relative similarities to weight selected pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Depending solely on label information can be insufficient for effective sample learning. Previous works addressed this limitation by assigning varying metrics or weights based on inter-sample variability. However, these techniques had limitations when dealing with a small number of training samples or unseen samples. To address this, we introduce additional training samples and a novel metric for improved learning. Prior works like [49, 17] added additional samples, but mainly relied on linear interpolation. In contrast, our approach generates extra training samples through selecting the appropriate MSDA, providing greater flexibility than linear interpolation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.7, + 0.438 + ], + "angle": 0, + "content": "2.2. Hard Sample Mining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Pair-based metric learning generates lots of paired samples, and handling negative samples is crucial for reducing model redundancy. Various studies [6, 10, 34, 41, 49] have explored the use of hard negative mining while constructing triples to create useful gradient sums, aiding fast convergence of triplet loss networks. Additionally, [34] suggested a semi-hard negatives scheme that selects even harder negative samples, while [6] introduced a hierarchical triplet loss (HTL) that constructs a hierarchical tree of all classes and selects semi-hard negatives with a dynamic margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.704 + ], + "angle": 0, + "content": "The discussed methods require selecting hard samples from the training set to improve model performance. Hard samples are challenging examples that can evaluate the model's ability to generalize. For example, if a model performs well on horses and donkeys but struggles with mules, evaluating its performance on mules (as a hybrid of both species) can directly assess its generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recent studies focus on generating additional hard samples for training. HDML [49] creates hard samples between different labeled samples, while [17] generates additional training samples between identically labeled samples through hard sample mining. Gu et al. [7] use synthetic proxy points and competition to create hard samples. These methods use linear interpolation to generate new samples, leading to false negatives and positives due to the difficulties in determining class boundaries on the manifold structure. DAS [21] and IAA [5] try to solve this issue by adjusting features slightly and adding extra examples during mid-training of the model, aiming to make sure that new examples still belong to the same class. Despite these attempts," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11048" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.446, + 0.107 + ], + "angle": 0, + "content": "completely removing this problem remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.229 + ], + "angle": 0, + "content": "We use mixed sample data augmentations in the preprocessing stage of images to generate additional hard samples, instead of linear interpolation. This has several advantages, including a higher probability of generating hard samples without subsequent mining and the ability to identify the underlying content of the image even without its labeling information. This makes our method highly effective for generating and processing additional hard samples." + }, + { + "type": "image", + "bbox": [ + 0.167, + 0.244, + 0.221, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.286, + 0.365, + 0.297 + ], + "angle": 0, + "content": "Original images in the training set" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.244, + 0.274, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.245, + 0.328, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.245, + 0.382, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.304, + 0.223, + 0.342 + ], + "angle": 0, + "content": "CutMix:" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.305, + 0.278, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.305, + 0.333, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.305, + 0.386, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.35, + 0.272, + 0.359 + ], + "angle": 0, + "content": "num=2" + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.35, + 0.324, + 0.359 + ], + "angle": 0, + "content": "num=3" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.305, + 0.385, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.35, + 0.378, + 0.359 + ], + "angle": 0, + "content": "num=4" + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.39, + 0.221, + 0.403 + ], + "angle": 0, + "content": "MixUp:" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.366, + 0.277, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.366, + 0.332, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.366, + 0.384, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.451, + 0.22, + 0.462 + ], + "angle": 0, + "content": "GridMask:" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.426, + 0.278, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.427, + 0.334, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.427, + 0.392, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.471, + 0.393, + 0.482 + ], + "angle": 0, + "content": "mask=50 mask=150 mask=250" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.499, + 0.468, + 0.536 + ], + "angle": 0, + "content": "Figure 2: Creating Hybrid Species: Three Ways. We explore three different ways for generating hybrid species, varying parameter num, which controls the number of classes included. The mask indicates block size." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.557, + 0.453, + 0.574 + ], + "angle": 0, + "content": "2.3. Mixed Sample Data Augmentations (MSDA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.717 + ], + "angle": 0, + "content": "Traditional data augmentation techniques assume all samples belong to the same class and do not consider relationships between different samples. In contrast, MSDA combines two samples, using various forms of label representation [35]. Popular methods include Mixup [47], Cutmix [45], Cutout [4], and GridMask [2], each with the core idea of producing images with more representative features of real-world scenarios. Other MSDA techniques include SmoothMix [20] and FMix [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Despite significant achievements in image classification, the above methods have not been applied to metric learning. [17] discussed the similarity between linear interpolation of features and mixup data augmentation, but due to the peculiarities of mixed labels, it cannot be directly applied to metric learning, which typically requires distinguishing samples as dichotomous, i.e., positive and negative samples. Recently, [38] modified an existing metric learning loss function to accommodate mixup, but the input mix was limited to mixup. Our proposed method can be applied to a variety of mixed sample data augmentations (Section 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We modified multiple mixed sample data augmentations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "as shown in Figure. 2 to ensure that mixed data augmented samples contain features from multiple classes in a balanced manner. Specifically, for CutMix, we uniformly cropped and merged images from top to bottom to ensure that features from multiple classes are regularly distributed. For Mixup, we fused the samples in equal proportions to ensure a balanced representation of features from multiple classes. Lastly, we modified Gridmask by replacing the mask with an image to create a fusion effect." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.242, + 0.673, + 0.259 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.267, + 0.64, + 0.281 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.456 + ], + "angle": 0, + "content": "As a preliminary for our proposed method, we consider a batch of images denoted by \\(x = x_{1},x_{2},\\ldots ,x_{n}\\) with corresponding labels \\(y = y_{1},y_{2},\\ldots ,y_{n}\\). Deep metric learning involves the use of a Convolutional Neural Network (CNN) to learn a non-linear transformation of each image into an m-dimensional deep feature space \\(\\phi (x;\\theta_{\\phi}):\\mathcal{X}\\to \\mathbb{R}^{m}\\), where \\(\\theta_{\\phi}\\) represents the network's parameters. In order to further learn a mapping from the m-dimensional feature space to a k-dimensional space, we use a linear mapping layer represented by \\(f(\\phi ;\\theta_f):\\mathbb{R}^m\\to \\mathbb{R}^k\\), where \\(\\theta_{f}\\) represents the parameters of the linear mapping layer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.457, + 0.891, + 0.486 + ], + "angle": 0, + "content": "Formally, we define the distance between two data points in the embedding space as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.499, + 0.891, + 0.517 + ], + "angle": 0, + "content": "\\[\nd _ {f} \\left(x _ {i}, x _ {j}\\right) = \\left\\| f \\left(\\phi_ {i}\\right) - f \\left(\\phi_ {j}\\right) \\right\\| _ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.528, + 0.891, + 0.558 + ], + "angle": 0, + "content": "Metric learning models usually involve pairwise sampling of a sample, e.g. considering the contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.57, + 0.891, + 0.588 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o n t}} = y d _ {f} (a, x) ^ {2} + (1 - y) \\left[ m - d _ {f} (a, x) \\right] _ {+} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.78 + ], + "angle": 0, + "content": "where \\( y \\) denotes the label, \\( a \\) is the anchor image and \\( m \\) is the marginal value. To train a metric learning model, we first select an anchor image and choose a positive sample with the same label (\\( y = 1 \\)) and a negative sample with a different label (\\( y = 0 \\)). The network is then trained to minimize a loss function that penalizes negative samples that are too close together and positive samples that are too far apart. It is noteworthy that the loss value will be relatively large when the distance between the anchor point and the negative sample is small, while it will be relatively small when the distance is very far. Thus, selecting hard samples during the selection of pairs will result in a larger gradient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Studies have tried to increase the number of hard samples for metric learning by generating more samples. However, linear interpolation methods can produce false positives and negatives. It is important to note that the binary metric learning label \\( y \\) cannot be used as a one-hot label for image classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our Hybrid Species Embedding (HSE) method addresses the challenges of creating additional hard samples" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11049" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "and applying them to metric learning without label information to improve generalization. We address the challenge of mixed sample data augmentations by emulating human categorization behavior to assign these samples to a reasonable position in the embedding space. We propose incorporating newly generated samples as an extra training signal for the model, using the same feature extraction network for both the original image and the hybrid species generated using MSDA, and included in each batch. The loss function \\( L_{m} \\) clusters the original samples based on their labels, while our proposed loss function \\( L_{Hy} \\) deals with the embedding positions of the new samples, as shown in Figure. 3." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.284, + 0.47, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.53 + ], + "angle": 0, + "content": "Figure 3: Our HSE method uses batches of original images and hybrid species generated by Mixed Sample Data Augmentation (MSDA). The feature extraction network (CNN) processes both types of images to obtain their corresponding feature vectors (\\(f_{Or}\\) and \\(f_{HS}\\)). The loss function includes two components: \\(L_{m}\\) (metric learning loss) groups original samples into clusters based on their labels, and \\(L_{Hy}\\) optimizes feature extraction by comparing hybrid species embedding positions." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.553, + 0.214, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.621, + 0.195, + 0.633 + ], + "angle": 0, + "content": "(a) HS-Anchors" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.553, + 0.334, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.621, + 0.317, + 0.632 + ], + "angle": 0, + "content": "(b) WP-Anchor" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.553, + 0.455, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.621, + 0.43, + 0.632 + ], + "angle": 0, + "content": "(c) N-Anchor" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.736 + ], + "angle": 0, + "content": "Figure 4: Three types of anchor point selection: The dashed lines indicate that the sample pairs should move away from each other, while the solid lines indicate that the sample pairs should move closer to each other. The red sample is a hybrid species synthesized from green and blue. (a) denotes the use of the hybrid species as the anchor point, (b) denotes using the class that synthesizes the hybrid species as an anchor point, and (c) denotes using a class unrelated to the hybrid species as an anchor point." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.231, + 0.773 + ], + "angle": 0, + "content": "3.2. Hybrid Species" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In the context of image classification tasks, hard negative samples refer to images that are similar to the anchor images but have different labels. We posit that hard negative samples typically possess similar features to the anchor samples. Therefore, when a sample is derived from an anchor image, there is a high likelihood that the sample will be classified as a hard sample. Let \\( x_{A} \\) and \\( x_{B} \\) denote two images with labels \\( y_{A} \\) and \\( y_{B} \\), respectively, where \\( y_{A} \\neq y_{B} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "The generated sample \\( x_{H} \\) contains both the features of \\( x_{A} \\) and the features of \\( x_{B} \\). We refer to \\( x_{H} \\) as the hybrid species." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.228 + ], + "angle": 0, + "content": "\\(x_{H}\\) has a high likelihood of becoming a hard negative sample, while simultaneously presenting potential gradients in classes \\(A\\) and \\(B\\). This can be achieved by utilizing mixed sample data augmentation techniques, such as mixup. Mixup generates a new training sample, denoted as \\((x_{H},y_{H})\\), by linearly interpolating two training samples \\((x_{A},y_{A})\\) and \\((x_{B},y_{B})\\), where \\(x_{H}\\) and \\(y_{H}\\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.24, + 0.892, + 0.257 + ], + "angle": 0, + "content": "\\[\nx _ {H} = \\lambda x _ {A} + (1 - \\lambda) x _ {B} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.284, + 0.892, + 0.3 + ], + "angle": 0, + "content": "\\[\ny _ {H} = \\lambda y _ {A} + (1 - \\lambda) y _ {B} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.397 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is obtained from a beta distribution and is constrained to [0, 1]. To achieve a distribution of \\(\\lambda\\) values that is centered around 0.5, \\(\\lambda\\) should follow a beta distribution with shape parameters \\(\\alpha\\) and \\(\\alpha\\), e.g., \\(\\lambda \\sim \\mathrm{Beta}(\\alpha, \\alpha)\\), where \\(\\alpha \\in (0, \\infty)\\). The generated training sample \\((x_H, y_H)\\) is then used for model training with a loss function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Many pair-based metric learning losses rely on having distinct class labels, but the output \\( y_{H} \\) is often a probability distribution per class, which is unsuitable for metric learning applications. Determining how to utilize these labels for metric learning presents a significant challenge. This difficulty is illustrated in Eq. 2." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.499, + 0.892, + 0.517 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o n t}} = y d _ {f} \\left(a, x _ {H}\\right) ^ {2} + (1 - y) \\left[ m - d _ {f} \\left(a, x _ {H}\\right) \\right] _ {+} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.633 + ], + "angle": 0, + "content": "As the binary label \\( y \\in \\{0,1\\} \\), the label for \\( x_{H} \\) is not straightforward, since it is a hybrid sample. When we consider \\( x_{H} \\) as a negative sample of the anchor point, it is not reasonable, as \\( x_{H} \\) contains the same features as the anchor samples. Keeping them away from each other could cause the model to ignore important features and focus on irrelevant ones. Therefore, the binary label \\( y \\) is invalid for \\( x_{H} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[38] showed that classification tasks are not fundamentally different from metric learning, and the label of the synthesized sample \\( X_{H} \\) can still be represented by linear interpolation of the labels. Eq. 5 can then be converted to:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.72, + 0.892, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\widetilde {L} _ {\\text {c o n t}} = \\lambda d _ {f} (a, x _ {H}) ^ {2} + (1 - \\lambda) [ m - d _ {f} (a, x _ {H}) ] _ {+} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Now, \\( x_{H} \\) appears with nonzero contributions in both positive and negative terms for positive-negative pairs, due to the interpolation factor \\( \\lambda \\in [0,1] \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, we found that these methods are only applicable to mixup and may be subject to controversy in other forms of mixed sample data augmentations. This is due to the fact that mixup binds the images together evenly based on the assigned weight parameter, \\(\\lambda\\). The same applies to labels In contrast, cutmix synthesizes \\(x_{H}\\) by replacing a crop area of \\(x_{A}\\) with that of \\(x_{B}\\), where the crop area of \\(\\lambda\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11050" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.418, + 0.14 + ], + "angle": 0, + "content": "
HS-AnorhWP-AnorhN-Anorh
Intra-class variation✓✓✕
Inter-class variation✓✓✓
Conflict✕✓✕
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.162, + 0.47, + 0.187 + ], + "angle": 0, + "content": "Table 1: Three types of anchor point selection are considered for inter-class and intra-class variation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.469, + 0.321 + ], + "angle": 0, + "content": "and \\((1 - \\lambda)\\) is taken from \\(x_{A}\\) and \\(x_{B}\\), respectively. Therefore, the label of \\(x_{H}\\) cannot be accurately represented by linear interpolation, as the area being replaced is indeterminate, and may contain either significant features or irrelevant content. Consequently, the contribution of the crop area to \\(x_{H}\\) remains inconclusive, making the application of linear interpolation for labeling mostly infeasible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.469, + 0.429 + ], + "angle": 0, + "content": "Instead of fully utilizing the label information of \\( x_{H} \\), we can translate them into their appropriate spatial embedding, as we will describe in next subsection. Also, when applying MSDA to images, we need to consider how the labels are to be processed, which limits us to intermixing only two samples. In contrast, \\( x_{H} \\) does not require us to define label information, and can hence be mixed by multiple classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.441, + 0.204, + 0.456 + ], + "angle": 0, + "content": "3.3. Hybrid loss" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.401, + 0.481 + ], + "angle": 0, + "content": "3.3.1 Anchor Point Selection and Processing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Although it is challenging to accurately assign a label to \\( x_{H} \\) obtained through MSDA, the features of \\( x_{H} \\) can be inferred based on the source images. We define \\( C(x_{I}) \\) to represent the label of image \\( x_{I} \\). When \\( x_{H} \\) is selected as an anchor, it may not be possible to find a suitable positive sample, but \\( x_{H} \\) shares semantic information with \\( C(x_{A}) \\) and \\( C(x_{B}) \\) to some extent. Thus, we can consider \\( C(x_{A}) \\) and \\( C(x_{B}) \\) as weak positive samples (WP) of \\( x_{H} \\), with the remaining classes treated as negative samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.794 + ], + "angle": 0, + "content": "During the training phase, the embedding of the spatial position of \\( x_{H} \\) plays a crucial role in the generalization of downstream tasks. Therefore, deciding how to treat \\( x_{H} \\) in pair selection becomes the first key question. Since \\( x_{H} \\) contains information for multiple classes, including it in a paired selection can provide more gradient information. We divided the pairing methods into three categories, as shown in Figure. 4: HS (hybrid species sample) as the anchor point, WP (weak positive sample) as the anchor point, and N (negative sample) as the anchor point. We compared the three methods and and summarized the findings in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The choice of anchor point in a classification task can affect the learning of intra-class and inter-class variation. When using \\( x_{H} \\) as an anchor point, it pulls towards the WP while staying away from the rest of the class. Using WP as an anchor point keeps it close to its composite \\( x_{H} \\) to maintain identical points within the class, but may result in conflicts with the synthetic sample of \\( x_{H} \\) leading" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "to non-converging losses. Anchoring with N treats \\( x_{H} \\) as a negative sample and moves it away from other samples, losing its purpose. Our findings recommend using HS as the anchor sample to effectively focus on both intra-class and inter-class variation while ensuring convergence." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.186, + 0.759, + 0.201 + ], + "angle": 0, + "content": "3.3.2 Strategies for Sample Mining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.405 + ], + "angle": 0, + "content": "In the previous subsection, we analyzed the contribution of pairwise selection to the embedding of \\( x_{H} \\). Since \\( x_{H} \\) is randomly generated in each epoch, how to embed it more efficiently and reasonably becomes the second key issue of our study. For instance, when \\( x_{H} \\) is used as an anchor point, the corresponding WP may contain multiple classes. If \\( x_{H} \\) pulls all WP close together, clusters that are unrelated may merge, which is unreasonable. If \\( x_{H} \\) randomly selects WP that are far apart, the structure of the cluster may be destroyed. On the other hand, if negative samples that are far apart are selected, the resulting gradient will be smaller. In summary, it is vital that we choose a reasonable and valid sample for performing metric learning with HS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Several studies have explored the advantages of hard negative mining in constructing informative gradients [12, 36, 34]. In this context, we introduce the notion of hard negatives for the hybrid species as:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.477, + 0.892, + 0.517 + ], + "angle": 0, + "content": "\\[\nx _ {H _ {h n}} = \\underset {x: C (x _ {I}) \\neq C (x _ {A}) \\cap C (x _ {I}) \\neq C (x _ {B})} {\\operatorname {a r g m i n}} d \\left(f \\left(x _ {H}\\right), f \\left(x _ {I}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.577 + ], + "angle": 0, + "content": "It is worth noting that \\( x_{H_{hn}} \\) does not only yield a large gradient for \\( x_{H} \\), but it also possesses the same potential gradient as the inputs \\( C(x_{A}) \\) and \\( C(x_{B}) \\) that are used to synthesize \\( x_{H} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.654 + ], + "angle": 0, + "content": "The study by [43] shows that easy positive mining improves the generalization performance of downstream tasks. Our experiments also demonstrate the benefits of this approach for \\( x_{h} \\). We define the easy hybrid weak positive of the hybrid species as:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.665, + 0.892, + 0.704 + ], + "angle": 0, + "content": "\\[\nx _ {H _ {e w p}} = \\underset {x: C \\left(x _ {I}\\right) = C \\left(x _ {A}\\right) \\cup C \\left(x _ {I}\\right) = C \\left(x _ {B}\\right)} {\\operatorname {a r g m i n}} d \\left(f \\left(x _ {H}\\right), f \\left(x _ {I}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.855 + ], + "angle": 0, + "content": "There are two primary reasons why bringing \\( x_{H} \\) and \\( x_{H_{ewp}} \\) closer together is beneficial. Firstly, it enables the model to concentrate on the features that are shared by both inputs, which can significantly influence the sample measurements. Secondly, this approach avoids the potential issue of pushing weak positives that are distant from each other too close together. By doing so, it preserves both the intra-class variance and the inter-class variance, which can help maintain the manifold structure of the classes in the embedding space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our hybrid loss function involves mapping the output of the convolutional neural network onto a unit sphere, which is a widely accepted technique. During each mini-batch," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11051" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.081, + 0.097, + 0.445, + 0.616 + ], + "angle": 0, + "content": "1 # Nsize: Number of samples of the same class in each batch \n2 # n: The number of classes that synthesize \\( x_{H} \\) \n3 # b: batch size d: dimension \n4 # F = b × d \n5 # Map of feature similarity between samples \n6 Map = F × FT; \n7 # \\( x_{H} \\) as the anchor, look for weak positive sample \n8 DWP = Map; \n9 for each i ∈ [b - Nsize, b] do \n10 for each j ∈ [n * Nsize, b] do \n11 | DWP[i][j] = -1; \n12 end \n13 end \n14 # Each \\( x_{H} \\) corresponds to the value and index of the most similar WP sample \n15 VDWP, IDWP = DWP.max(1); \n16 # \\( x_{H} \\) as the anchor, look for negative sample \n17 DN = Map; \n18 for each i ∈ [b - Nsize, b] do \n19 for each j ∈ [0, n * Nsize] do \n20 | DN[i][j] = -1; \n21 end \n22 for each j ∈ [b - Nsize, b] do \n23 | DN[i][j] = -1; \n24 end \n25 end \n26 # Each \\( x_{H} \\) corresponds to the value and index of the most similar negative sample \n27 VDN, IDN = DN.max(1); \n28 # Select the WP sample and negative sample parts corresponding to \\( x_{H} \\) \n29 T = [VDWP[-Nsize:], VDN[-Nsize:]]; \n30 Hybrid loss = -α logsoftmax(T).mean();" + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.619, + 0.342, + 0.634 + ], + "angle": 0, + "content": "Algorithm 1: Hybrid Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.47, + 0.752 + ], + "angle": 0, + "content": "when \\( x_{H} \\) serves as the anchor, we select \\( x_{H_{ewp}} \\) and \\( x_{H_{hn}} \\) for comparison. As in Algorithm 1, the former corresponds to the original sample, while the latter represents \\( x_{H} \\). To synthesize \\( x_{H} \\), we typically select samples from the first \\( n \\) classes. Then, we search for the most similar weak positive sample and negative sample for \\( x_{H} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.752, + 0.47, + 0.827 + ], + "angle": 0, + "content": "The NCA loss function is an effective way to avoid the selection of margin hyperparameters and handle the distance between samples efficiently[46]. To ensure that hybrid species are embedded in a suitable location, we define the hybrid loss based on the NCA loss:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.837, + 0.469, + 0.883 + ], + "angle": 0, + "content": "\\[\nL _ {H y} = - \\alpha \\log \\frac {1}{1 + \\frac {\\exp (f (x _ {H}) ^ {\\top} f (x _ {H _ {h n}}))}{\\exp (f (x _ {H}) ^ {\\top} f (x _ {H _ {e w p}}))}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyper-parameter that controls the balance be" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.169 + ], + "angle": 0, + "content": "tween bringing \\( x_{H} \\) and \\( x_{H_{ewp}} \\) closer and pulling \\( x_{H} \\) and \\( x_{H_{hn}} \\) farther apart. The \\( L_{Hy} \\) loss serves as an auxiliary loss to process additional samples and can be combined with other metric learning losses. As in Figure 3, the final loss is the sum of the metric learning loss \\( L_{m} \\) and the \\( L_{Hy} \\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.184, + 0.892, + 0.2 + ], + "angle": 0, + "content": "\\[\nL = L _ {m} + L _ {H y} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.208, + 0.892, + 0.238 + ], + "angle": 0, + "content": "When incorporating our method with the contrastive loss (Eq. 2), the resulting final loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.251, + 0.892, + 0.268 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {c o n t}} + L _ {H y} \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.281, + 0.633, + 0.299 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.458 + ], + "angle": 0, + "content": "Our HSE method uses pre-trained BN-Inception [14] and ResNet [11] models on ILSVRC 2012-CLS [32] with PyTorch. These models have learned feature representations from diverse images, reducing time and resources for new dataset training. They also perform well in image recognition tasks, making them suitable for computer vision applications. For generalization ability evaluation, we maintain consistency with base methods on image preprocessing, learning rate, and batch size. (Source code: https://github.com/SHQberserker/HSE)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Nsize examples from each class are selected per batch. We use two methods to incorporate hybrid species during training: selecting multiple classes to add hybrid species as additional examples of batch, and randomly replacing examples in the batch with hybrid species. We use the cutmix technique to create hybrid species by stitching two samples together evenly. Algorithm 1 fixes the last Nsize samples of each batch as synthetic samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.698, + 0.605 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We tested our framework's ability to perform well on challenging datasets: CUB-200, CARS-196, and SOP. These datasets involve fine-grained classification tasks, where there are many categories and relatively few images per category. We followed the data split approach from [22]. Specifically, for CUB-200, we used 5864 images from the first 100 categories for training and 5924 images from the last 100 categories for testing. For CARS-196, we trained on the first 98 classes and tested on the remaining images. Additionally, for SOP, we utilized the first 59,551 images (11,318 classes) for training and the remaining 60,502 images (11,316 classes) for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To assess our experiments, we employed the Recall@K metrics. This measure involves selecting the K images that are closest to the query image. If there's an image with the same label as the query image, we score 1; otherwise, it's 0. Recall@K calculates the average score for all query images, with higher values indicating better retrieval performance. These metrics provide a way to quantitatively evaluate how" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11052" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "well our model can find similar images, which is crucial for many image-based applications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.129, + 0.273, + 0.145 + ], + "angle": 0, + "content": "4.2. Results and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.153, + 0.218, + 0.167 + ], + "angle": 0, + "content": "4.2.1 Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.177, + 0.47, + 0.417 + ], + "angle": 0, + "content": "Experiments compared our proposed method to existing methods, including N-pair, EPSHN, MSloss and ProxyAnchor. Results are shown in Table 2 for CUB-200 and CARS-196 datasets, where * indicates our results under the same experimental settings. Our method improves on different approaches to varying degrees. Our method had less impact on the N-pair method as it mostly ignores intraclass variation, leading to excessive clustering between the same classes. However, we observed a significant improvement in recall@K when the EPSHN method was augmented with the additional training signal of HSE. Recall@1 improved by \\(2.5\\%\\) and \\(2.3\\%\\) on the challenging CUB-200 and CARS-196 datasets, respectively, where Nsize=16 and Nsize=8, respectively. Importantly, HSE-E outperformed the ensemble-based method that EPSHN had not surpassed before on the CUB-200 and CARS-196 datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.583 + ], + "angle": 0, + "content": "We believe that EPSHN tends to ignore intra-class variations, such as those found in [43] where birds of different sexes may vary greatly within the same class, but should also focus on characteristics common between members of the same class. When \\( x_{H} \\) is added, it helps the network to focus more on these common characteristics. MSloss focuses on both inter-class variation and inter-class variation. However, when adding our additional training signals, there is still a certain improvement. This is because the features that our additional training signals focus on help the network generalize to unknown classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Our method achieves good results in pair-based methods and is effective in proxy-based metric learning methods. We improve approximately \\(1\\%\\) in both \\(\\mathrm{Recall}@1\\) and \\(\\mathrm{Recall}@2\\). This confirms our universality, which can be applied to various metric learning loss functions. We also compared our method to the Matrix/Input [38] approach. Matrix/Input has a high computational cost because it calculates \\(\\frac{1}{2} n(n - 1)\\) embeddings for a batch of \\(n\\) examples. Instead, our method achieves a similar effect using only \\(Nsize * Cnum\\) embeddings. Our approach is hence more computationally efficient. We also remain applicable to discriminate between positives and negatives for every anchor point. We can also incorporate non-additive component functions that involve loss functions, e.g., EPSHN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.9 + ], + "angle": 0, + "content": "HSE is also effective on the SOP dataset. Unlike smaller datasets, the SOP dataset has superclasses, which introduce more variation between different superclasses and less variation within the same superclass. It's challenging to measure similarity between samples within the same superclass in the SOP dataset. As shown in Table 3, we generate HS from samples that share the same superclass but have differ" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "ent labels. When we apply HSE to various classical methods, we observe varying degrees of improvement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.142, + 0.783, + 0.157 + ], + "angle": 0, + "content": "4.2.2 Impact of Minibatch Parameters" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.166, + 0.614, + 0.181 + ], + "angle": 0, + "content": "Impact of Nsize" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.189, + 0.892, + 0.34 + ], + "angle": 0, + "content": "We trained a ResNet-18 embedding network on CUB-200 and CARS-196 datasets with batch size 128 and embedding dimension 512 to investigate HSE's impact on model generalization at various Nsizes. HSE's effectiveness varies with methods, datasets, and Nsizes (Figure. 5). The best performance occurs at Nsize between 4 and 16 with a batch size of 128, followed by performance degradation. When there are too few negative samples, the hybrid species' gradient information becomes insignificant, and they may be grouped with WP's clusters, harming downstream tasks." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.353, + 0.663, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.439, + 0.621, + 0.449 + ], + "angle": 0, + "content": "(a) CUB" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.353, + 0.848, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.439, + 0.807, + 0.449 + ], + "angle": 0, + "content": "(b) CAR" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.467, + 0.89, + 0.48 + ], + "angle": 0, + "content": "Figure 5: Effect of additional training signals provided by different Nsizes." + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.495, + 0.769, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.597, + 0.884, + 0.61 + ], + "angle": 0, + "content": "Figure 6: Results on the CUB dataset using different Cnum in MS loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.622, + 0.62, + 0.636 + ], + "angle": 0, + "content": "Impact of Cnum" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.781 + ], + "angle": 0, + "content": "We found that HSE improves the generalization ability of the model. However, increasing the number of hybrid species does not necessarily lead to better results. We conducted experiments on the CUB dataset using the MS loss method, as shown in Figure. 6. Our loss function indirectly affects the entire embedding space of the batch when learning the spatial embedding of a single hybrid species. After training for a certain period, the two combined classes have essentially covered the entire dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.724, + 0.817 + ], + "angle": 0, + "content": "4.2.3 Effect of Hybrid Species" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We investigated the effectiveness of different data augmentation techniques in providing additional training signals for hybrid species. Table 4 shows the classic blending data augmentation techniques we used, with certain modifications made to each method. Notably, CutMix and GridMask" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11053" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.189, + 0.089, + 0.778, + 0.417 + ], + "angle": 0, + "content": "
CUB-200CARS-196
Recall@kk=1k=2k=4k=1k=2k=4
Angular512[39]G53.665.075.371.380.787.0
HDML512[49]G53.765.776.779.187.192.1
HTL512[6]BN57.168.878.781.488.092.7
Margin128[41]R63.674.483.179.686.591.9
SoftTriple512[28]BN65.476.484.584.590.794.5
DR512[24]BN66.177.085.185.090.594.1
Ensemble-based
HDC384[44]G53.665.777.073.282.486.4
A-BIER512[27]G55.367.276.978.085.891.1
ABE-8512[16]G60.671.579.860.671.579.8
DREML9216[42]R63.975.083.186.091.795.0
Npair64*(Nsize=8)[37]R51.364.275.364.276.084.6
HSE-N64(Nsize=8)R52.5+1.266.0+1.877.4+2.164.9+0.776.6+0.684.7+0.1
EPSHN512*(Nsize=16)[43]R64.975.383.583.189.793.6
HSE-E512(Nsize=16)R66.9+2.077.4+2.185.5+2.085.4+2.391.2+1.596.9+3.3
HSE-E512(Nsize=8)R67.4+2.577.7+2.485.7+2.284.8+1.790.6+0.994.3+0.7
MS†512R67.877.885.687.892.795.3
+Matrix输入[38]R69.0+1.279.1+1.386.0+0.489.0+1.293.4+0.796.0+0.7
MS512*(Nsize=5)[40]BN65.777.086.381.788.793.2
HSE-M512(Nsize=5)BN67.6+1.978.0+1.085.8-0.582.0+0.388.9+0.293.3+0.1
Proxy-Ancho†512R69.579.387.087.692.395.5
+Matrix输入[38]R70.5+1.081.2+1.987.8+0.888.2+0.693.2+0.996.2+0.7
Proxy-Ancho*512[15]R69.479.287.088.592.795.6
HSE-PA512R70.6+1.280.1+0.987.1+0.189.6+1.193.8+1.196.0+0.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.428, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Table 2: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the CUB-200-2011 and Cars-196 datasets. The backbone networks used in the models are denoted by abbreviations: G for GoogleNet, BN for Inception with batch normalization, and R for ResNet50. The symbol \\(\\dagger\\) indicates the reproduced result reported by the original authors.N-pair and EPSHN use cnum=1 and batch size of 128. MSloss uses cnum=1, batch size of 80, and \\(\\alpha = 0.05\\). Proxy-Anor uses cnum=1, batch size of 128, and \\(\\alpha = 2\\)." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.501, + 0.466, + 0.658 + ], + "angle": 0, + "content": "
SOP
Recall@kk=1k=10k=100
EPSHN512*(Nsize=5) [43]R74.386.994.3
HSE-E512(Nsize=5)R76.3+2.088.4+1.594.9+0.6
MS†512R76.989.895.9
+Matrix输入[38]R77.9+1.090.6+0.895.9
MS512*[40] (Nsize=5)R78.290.296.2
HSE-M512R78.7+0.590.4+0.296.1-0.1
Proxy-Anchort512R79.190.896.2
+Matrix输入[38]R79.8+0.791.4+0.696.5+0.3
Proxy-Anchort*512[15]R79.290.495.8
HSE-PA512(Nsize=5)R80.0+0.891.4+1.096.3+0.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Table 3: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the SOP.Where batch \\( = {120} \\) ,cnum \\( = {12} \\) ,and hybrid species as additional examples of batch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.884 + ], + "angle": 0, + "content": "stand out. CutMix can incorporate features from each compositional image to a certain extent, but may also diminish key features. Mixup dilutes key features as the number of samples increases. GridMask (block=50) covers the entire image despite using relatively small blocks. Results show that CutMix is the most effective approach for improving the performance of Hybrid Species (HS). We also found that adding more class features does not necessarily improve HS performance, but adding critical features that are specific to a particular class can lead to better HS results." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We investigate the selection of weak positive samples" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.501, + 0.868, + 0.671 + ], + "angle": 0, + "content": "
CUBCAR
R@1numR@1num
Cutmix67.4285.42
67.2385.03
67.1484.24
Mixup67.0285.32
66.9381.03
R@1MasksizeR@1Masksize
Gridmask67.45084.950
66.015084.8150
67.125084.0250
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.693, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Table 4: Results on the Cub dataset using different Mixed Sample Data Augmentations in EPSHN. \\((\\mathbb{R}@\\mathbb{1} = \\mathbb{Recall}@\\mathbb{1})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "by hybrid species using the Proxy-Anorh method on the CUB dataset. Cutmix augmentation technique is used to stitch images either vertically or horizontally, with the upper part mainly containing the bird's head and the lower part containing the bird's body when stitched vertically. When stitched horizontally, the probability of the left and right sides containing the bird's head is almost the same. Our HSE approach improves model performance by focusing on essential image features. Our study on the CUB dataset demonstrates that the hybrid species prioritize the bird's" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11054" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.416, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.372 + ], + "angle": 0, + "content": "Figure 7: Analysis of easy weak positive samples selected by hybrid species using Proxy-Antor on CUB dataset. Hybrid species are obtained by stitching two images of different labels using cutmix. We quantify the probability that selected easy weak positive samples contain upper or bottom half during training using \"aut_up/down\". With our method, we also report \"our_up/down\" and \"our_left/right\" indicating the likelihood of easy weak positive samples containing upper or bottom half of the bird or left or right half of the stitched image, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.55 + ], + "angle": 0, + "content": "body during metric learning, particularly the lower half, which is synthesized as the easy weak positive sample using the Cutmix augmentation technique. While the selection probability of the lower half increases from 0.6 to \\(0.7\\sim 0.8\\) during Proxy training, our proposed method ensures a probability of 1. Stitching hybrid species horizontally results in a selection probability of around 0.5. This approach allows the network to focus on important features during training, providing an advantage over traditional methods. Figure. 7 provides a visual representation of our findings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.371, + 0.589 + ], + "angle": 0, + "content": "4.2.4 Visualization of Embedding Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of our proposed method on training and test data using t-SNE (t-distributed stochastic neighbor embedding) visualization technique. t-SNE can effectively reduce the dimensionality of the embedding space while preserving pairwise similarity between samples. It enables us to visually inspect how well the model separates different classes and identifies clusters of similar samples by plotting the embedding space in 2D or 3D. Therefore, t-SNE is a useful tool for evaluating the quality of learned embeddings in deep metric learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We experimented on two classes from the training set and two classes (wren and woodpecker) from the test set. Figure. 8 shows that EPSHN is effective at distinguishing between different classes in the training set, but the embedding for hybrid species is relatively dispersed. With the additional HSE training signal, the embedding of the hybrid species becomes more concentrated and is located between the two classes. This approach considers all the features of difficult samples, creating a more comprehensive embedding of the samples. When additional training signals" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "are learned, the model can generalize better to unknown classes. This is demonstrated in the test set, where the EPSHN method cannot distinguish the black classes from the green and blue classes. However, when the training signal of HSE is applied to the downstream task, black classes can be successfully distinguished from green classes." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.196, + 0.588, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.234, + 0.612, + 0.248 + ], + "angle": 0, + "content": "(a) Oriole \\((\\mathrm{id} = 95)\\)" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.196, + 0.732, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.234, + 0.766, + 0.248 + ], + "angle": 0, + "content": "(b) Nuthatch (id=94)" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.192, + 0.87, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.237, + 0.88, + 0.252 + ], + "angle": 0, + "content": "(c) Hspecie" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.271, + 0.648, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.361, + 0.648, + 0.375 + ], + "angle": 0, + "content": "EPSHN (train)" + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.272, + 0.845, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.36, + 0.863, + 0.375 + ], + "angle": 0, + "content": "HSE-EPSHN (train)" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.39, + 0.648, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.477, + 0.644, + 0.491 + ], + "angle": 0, + "content": "EPSHN (test)" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.391, + 0.844, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.477, + 0.856, + 0.492 + ], + "angle": 0, + "content": "HSE-EPSHN (test)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Figure 8: We present a t-SNE visualization of EPSHN and HSE-EPSHN on the test set of CUB. Specifically, we select two different classes (denoted by blue dots and green dots) from the training set as shown in (a) and (b), respectively. (c) is the synthetic class from classes (a) and (b), denoted by red dots. For the test set, we selected four bird categories, consisting of two types of wrens and two types of woodpeckers, where each color represents a specific category." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.619, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Our Hybrid Species Embedding (HSE), a metric learning approach that aims to enable models to learn the spatial embedding of unknown and difficult classes during training. HSE generates additional hard samples using mixed data augmentation methods without defining their label information and applies them to the metric learning process. Experiments demonstrate significant improvements on multiple datasets when compared to three widely used pair-based methods and state-of-the-art proxy methods. HSE enhances the generalization ability of the model by providing additional training signals in deep metric learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgment. This work is partly supported by the National Natural Science Foundation of China (No: 62172366) and the Pioneer and Leading Goose R&D Program of Zhejiang Province (2023C01150)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11055" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] E. Ahmed, M. Jones, and T. K. Marks. An improved deep learning architecture for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3908–3916, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[2] P. Chen, S. Liu, H. Zhao, and J. Jia. Gridmask data augmentation. arXiv preprint arXiv:2001.04086, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.47, + 0.242 + ], + "angle": 0, + "content": "[3] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.283 + ], + "angle": 0, + "content": "[4] T. DeVries and G. W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[5] Z. Fu, Z. Mao, B. Hu, A.-A. Liu, and Y. Zhang. Intra-class adaptive augmentation with neighbor correction for deep metric learning. IEEE Transactions on Multimedia, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.368 + ], + "angle": 0, + "content": "[6] W. Ge. Deep metric learning with hierarchical triplet loss. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-285, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[7] G. Gu, B. Ko, and H.-G. Kim. Proxy synthesis: Learning with synthetic classes for deep metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1460-1468, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.47, + 0.492 + ], + "angle": 0, + "content": "[8] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[9] E. Harris, A. Marcu, M. Painter, M. Niranan, A. Prügel-Bennett, and J. Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[10] B. Harwood, V. Kumar BG, G. Carneiro, I. Reid, and T. Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 2821-2829, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[12] A. Hermans, L. Beyer, and B. Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[13] E. Hoffer and N. Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[14] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[15] S. Kim, D. Kim, M. Cho, and S. Kwak. Proxy anchor loss for deep metric learning. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[16] W. Kim, B. Goyal, K. Chawla, J. Lee, and K. Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[17] B. Ko and G. Gu. Embedding expansion: Augmentation in embedding space for deep metric learning. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7255-7264, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[18] B. Ko, G. Gu, and H.-G. Kim. Learning with memory-based virtual classes for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11792-11801, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[19] M. T. Law, N. Thome, and M. Cord. Quadruplet-wise image similarity learning. In Proceedings of the IEEE international conference on computer vision, pages 249-256, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[20] J.-H. Lee, M. Z. Zaheer, M. Astrid, and S.-I. Lee. Smoothmix: a simple yet effective data augmentation to train robust classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 756-757, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[21] L. Liu, S. Huang, Z. Zhuang, R. Yang, M. Tan, and Y. Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[22] Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[23] T. Milbich, K. Roth, H. Bharadhwaj, S. Sinha, Y. Bengio, B. Ommer, and J. P. Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In European Conference on Computer Vision, pages 590-607. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[24] D. D. Mohan, N. Sankaran, D. Fedorishin, S. Setlur, and V. Govindaraju. Moving in the right direction: A regularization for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14591-14599, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[25] Y. Movshovitz-Attias, A. Toshev, T. K. Leung, S. Ioffe, and S. Singh. No fuss distance metric learning using proxies. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[26] H. Oh Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[27] M. Opitz, G. Waltner, H. Possegger, and H. Bischof. Deep metric learning with bier: Boosting independent embeddings robustly. IEEE transactions on pattern analysis and machine intelligence, 42(2):276-290, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[28] Q. Qian, L. Shang, B. Sun, J. Hu, H. Li, and R. Jin. Soft-triple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[29] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] K. Roth, O. Vinyals, and Z. Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11056" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] K. Roth, O. Vinyals, and Z. Akata. Non-isotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.468, + 0.215 + ], + "angle": 0, + "content": "[32] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.468, + 0.273 + ], + "angle": 0, + "content": "[33] A. Sanakoyeu, V. Tschernezki, U. Buchler, and B. Ommer. Divide and conquer the embedding space for metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 471-480, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.468, + 0.328 + ], + "angle": 0, + "content": "[34] F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[35] C. Shorten and T. M. Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.468, + 0.439 + ], + "angle": 0, + "content": "[36] E. Simo-Serra, E. Trulls, L. Ferraz, I. Kokkinos, P. Fua, and F. Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.468, + 0.482 + ], + "angle": 0, + "content": "[37] K. Sohn. Improved deep metric learning with multi-class n-pair loss objective. Advances in neural information processing systems, 29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.468, + 0.536 + ], + "angle": 0, + "content": "[38] S. Venkataramanan, B. Psomas, Y. Avrithis, E. Kijak, L. Amsaleg, and K. Karantzalos. It takes two to tango: Mixup for deep metric learning. arXiv preprint arXiv:2106.04990, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.468, + 0.592 + ], + "angle": 0, + "content": "[39] J. Wang, F. Zhou, S. Wen, X. Liu, and Y. Lin. Deep metric learning with angular loss. In Proceedings of the IEEE international conference on computer vision, pages 2593-2601, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[40] X. Wang, X. Han, W. Huang, D. Dong, and M. R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5022-5030, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[41] C.-Y. Wu, R. Manmatha, A. J. Smola, and P. Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[42] H. Xuan, R. Souvenir, and R. Pless. Deep randomized ensembles for metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 723-734, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[43] H. Xuan, A. Stylianou, and R. Pless. Improved embeddings with easy positive triplet mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2474-2482, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.872 + ], + "angle": 0, + "content": "[44] Y. Yuan, K. Yang, and C. Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE international conference on computer vision, pages 814-823, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[45] S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo. Cutmix: Regularization strategy to train strong classifiers with" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[46] D. Zhang, Y. Li, and Z. Zhang. Deep metric learning with spherical embedding. Advances in Neural Information Processing Systems, 33:18772-18783, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[47] H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[48] Y. Zhao, Z. Jin, G.-j. Qi, H. Lu, and X.-s. Hua. An adversarial approach to hard triplet generation. In Proceedings of the European conference on computer vision (ECCV), pages 501-517, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[49] W. Zheng, Z. Chen, J. Lu, and J. Zhou. Hardness-aware deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 72-81, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[50] W. Zheng, C. Wang, J. Lu, and J. Zhou. Deep compositional metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9320-9329, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[51] W. Zheng, B. Zhang, J. Lu, and J. Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11057" + } + ] +] \ No newline at end of file diff --git a/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_origin.pdf b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..32ed65e604ec66c7680e63c748f409c26ffd5797 --- /dev/null +++ b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/debbcae4-e366-4ac1-b461-19f60fcf7463_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:854d333fe04d3768a2b03f0892b88ae4c5f62ae6028d06cb6db7f1a2224fddda +size 2161661 diff --git a/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/full.md b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..33c5e992965b45d83e3edcdef73704d759ee6c48 --- /dev/null +++ b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/full.md @@ -0,0 +1,451 @@ +# HSE: Hybrid Species Embedding for Deep Metric Learning + +Bailin Yang $^{1,2}$ , Haoqiang Sun $^{1}$ , Frederick W. B. Li $^{3}$ , Zheng Chen $^{1}$ , Jianlu Cai $^{1}$ , Chao Song $^{1*}$ + +$^{1}$ Faculty of Computer Science and Technology, Zhejiang Gongshang University + +$^{2}$ Faculty of Statistics and Mathematics, Zhejiang Gongshang University + +$^{3}$ Department of Computer Science, University of Durham + +# Abstract + +Deep metric learning is crucial for finding an embedding function that can generalize to training and testing data, including unknown test classes. However, limited training samples restrict the model's generalization to downstream tasks. While adding new training samples is a promising solution, determining their labels remains a significant challenge. Here, we introduce Hybrid Species Embedding (HSE), which employs mixed sample data augmentations to generate hybrid species and provide additional training signals. We demonstrate that HSE outperforms multiple state-of-the-art methods in improving the metric Recall@K on the CUB-200, CAR-196 and SOP datasets, thus offering a novel solution to deep metric learning's limitations. + +# 1. Introduction + +Image retrieval heavily depends on deep metric learning to grasp visual similarities. In its early stages, methods like pair-based loss [8, 13, 19, 37] and proxy-based loss [25, 15] were proposed to train models. Nevertheless, the true goal of image retrieval is to adapt to the unknown, and the challenge lies in selecting a metric that can effectively handle differences between classes in deep metric learning. A promising approach involves exploring both intra-class and inter-class variations within the image itself, rather than relying solely on label information [33, 43, 40, 23, 51, 30]. + +Hard negative samples, often termed false positives, refer to images that resemble anchor images but carry different labels. The reason behind these hard samples lies in their potential for sharing remarkably similar or even identical features with the anchor samples. Approaches like hard sample mining and generation have been proposed to aid network convergence by introducing a substantial gradient [1, 41, 12, 34, 49, 48, 17, 18]. Recently, an advancement in generating appropriate hard samples involves creating supplementary training data [17, 38, 5, 21]. This is primarily + +![](images/ecaa3d20883361c9e06584a465127b95d5c53a605b8010cf6140841665824105.jpg) + +![](images/36588301f77640667a9b9698f1cbaf42aba4020e6afbee62c4477cdcfc64763b.jpg) +(a) Linear interpolation + +![](images/a1056ed98209311971a51ea596e06753ad947904072e5f29a9c65623e85871de.jpg) +(b) Our HSE +Figure 1: (a) Deep metric learning methods struggle with linear interpolation for generating new training samples with labels when new samples fall between anchors of different classes. The manifold structure lacking clear boundaries makes it challenging to assign a label, even when new samples are between anchors of the same class. (b) Samples with multiple simultaneous features may not lie on a linear interpolation path, and their location may be unknown. They can be classified as an unknown class situated inside or outside a manifold boundary. To tackle this challenge, we propose using the HSE method to allocate the sample to an appropriate embedding space location based on its feature information. + +achieved through linear interpolation, a prevalent method for generating synthetic samples. However, assigning absolute labels to these synthesized samples is intricate due to the manifold structures and cluster boundaries present in the embedded space [17]. Even when the new samples are formed by linear interpolation between samples of the same label, accurately determining absolute labels remains a challenge. Synthetic sample labels have emerged as a potential solution, with [38] corroborating their similarity to mixup data augmentation and highlighting the label's representation through linear interpolation. Nonetheless, this approach is constrained to mixup data augmentation, as mixup uniformly blends features and labels from different images, which might not hold true for other techniques. + +We introduce a novel Hybrid Species Embedding (HSE) that leverages mixed sample data augmentations to generate additional training samples [35, 47, 45, 2, 20]. HSE creates hybrid species by combining features from multiple classes and embedding them into spatial positions. During training, different classes are randomly chosen for data augmentation + +in each batch. The labels of hybrid species are disregarded, and their placement near the category with the most similar features simulates human categorization behavior, as depicted in Figure 1(b). This strategy positions hybrid species close to the samples they are synthesized from, ensuring similarity while avoiding unrelated samples. By synthesizing samples using various mixed sample data augmentation techniques, each incorporating multiple features of a known class, our approach naturally imparts similar characteristics to the hard sample for the model without requiring explicit mining. Additionally, we dynamically adjust the position of each batch of hybrid species embedding to bypass the need for considering label information of these hybrid instances. Our contributions encompass: + +- Introducing a novel metric learning strategy embedding challenging and unseen classes (Hybrid species), during training. These hybrid species offer supplementary training signals, enhancing the generalization capacity of downstream tasks, as validated through experiments. +- Adapting mixed sample data augmentation techniques for our Hybrid Species Embedding (HSE) to address constraints associated with pair-based metric learning losses, which typically demand explicit class labels. +- Demonstrating through experiments our efficacy in enhancing performance across state-of-the-art metric learning tasks on CUB-200, CAR-196, and SOP datasets. + +# 2. Related Work + +We utilize the Mixed Sample Data Augmentation (MSDA) method to create unknown samples and improve the model's generalization ability for downstream tasks. This approach resolves the issue of insignificant comparisons caused by random selection of samples in metric learning. While mining hard samples from a finite training set is limited, MSDA enables us to create more hard samples. In the following, we review existing relevant works. + +# 2.1. Metric Learning + +Deep metric learning compares data in pairs and pulls them together if they belong to the same category, otherwise separates them [8, 13, 19, 37, 26]. The exponential growth in the number of tuples during training can present challenges in achieving model convergence, and may include meaningless metrics. Proxy-based approaches, e.g., choosing a proxy point for each class, have been proposed as a solution [25, 15]. These proxies contain information from multiple samples, require less computation in pairwise comparisons, helping the network converge more effectively. + +Methods that rely solely on label information may result in over-clustered or overly separated samples with identical or differing labels, respectively, and can lead to reduced performance on new and unseen test samples. EPSHN [43] notes this problem, particularly for species with varying + +features such as birds of the same species with different sexes. To avoid this, [30, 40, 31, 33, 51, 50, 18] emphasize the importance of incorporating semantic information from the images themselves, often through large pre-trained language models like CLIP [29] and BERT [3], to better align the visual representation space and discover potential differences within and between classes. The MSloss function [40] is one such approach that combines self-similarity and relative similarities to weight selected pairs. + +Depending solely on label information can be insufficient for effective sample learning. Previous works addressed this limitation by assigning varying metrics or weights based on inter-sample variability. However, these techniques had limitations when dealing with a small number of training samples or unseen samples. To address this, we introduce additional training samples and a novel metric for improved learning. Prior works like [49, 17] added additional samples, but mainly relied on linear interpolation. In contrast, our approach generates extra training samples through selecting the appropriate MSDA, providing greater flexibility than linear interpolation. + +# 2.2. Hard Sample Mining + +Pair-based metric learning generates lots of paired samples, and handling negative samples is crucial for reducing model redundancy. Various studies [6, 10, 34, 41, 49] have explored the use of hard negative mining while constructing triples to create useful gradient sums, aiding fast convergence of triplet loss networks. Additionally, [34] suggested a semi-hard negatives scheme that selects even harder negative samples, while [6] introduced a hierarchical triplet loss (HTL) that constructs a hierarchical tree of all classes and selects semi-hard negatives with a dynamic margin. + +The discussed methods require selecting hard samples from the training set to improve model performance. Hard samples are challenging examples that can evaluate the model's ability to generalize. For example, if a model performs well on horses and donkeys but struggles with mules, evaluating its performance on mules (as a hybrid of both species) can directly assess its generalization ability. + +Recent studies focus on generating additional hard samples for training. HDML [49] creates hard samples between different labeled samples, while [17] generates additional training samples between identically labeled samples through hard sample mining. Gu et al. [7] use synthetic proxy points and competition to create hard samples. These methods use linear interpolation to generate new samples, leading to false negatives and positives due to the difficulties in determining class boundaries on the manifold structure. DAS [21] and IAA [5] try to solve this issue by adjusting features slightly and adding extra examples during mid-training of the model, aiming to make sure that new examples still belong to the same class. Despite these attempts, + +completely removing this problem remains challenging. + +We use mixed sample data augmentations in the preprocessing stage of images to generate additional hard samples, instead of linear interpolation. This has several advantages, including a higher probability of generating hard samples without subsequent mining and the ability to identify the underlying content of the image even without its labeling information. This makes our method highly effective for generating and processing additional hard samples. + +![](images/b3d215d90ebbd4282c7727b890f1e5fddb6a32474d52211a4b39466424271b69.jpg) +Original images in the training set + +![](images/b6928a2fc5121385fabe763a7eef27da2212d9c75cde2037caf971a210f99677.jpg) + +![](images/348e4eb900e98c7dd043b4861f01556e938076c9f209b326e28c9338cfd98983.jpg) + +![](images/2723d5c7d840fabcaf274d643b5982cc10924dc2472408714e849bb07135cb9f.jpg) + +![](images/f7b7fc13c95b2660876701ecb06fe644dbfe072d4eca8644e7cbaea1c0fa5341.jpg) +CutMix: +num=2 + +![](images/b7e123435f312fde5952313b8f2b94e9126320004b582185a1c4b1f1583b5ca4.jpg) +num=3 + +![](images/930df273531792bd39256143b57cd2e9e61edb8c3f4976ee9633c9ef9ed6cd0f.jpg) +num=4 + +![](images/bec7efbcb0d10baa15efbc267008d12044ce53b3b652a5bc36b555cfa6c139ab.jpg) +MixUp: + +![](images/e4ad4171544bfe2ae3054ecbdf34b03e94c735d762f511b609cb594bec66dd6d.jpg) + +![](images/07ccb31ab4f6ff562ebf6fefb7ef0dc070c6e3a8e523d7883d494f4752c53e21.jpg) + +![](images/c04e8b8e90a95db5da00a802855ca7a9f5e1c7883eca3c98e1c2b015d46f1006.jpg) +GridMask: +mask=50 mask=150 mask=250 +Figure 2: Creating Hybrid Species: Three Ways. We explore three different ways for generating hybrid species, varying parameter num, which controls the number of classes included. The mask indicates block size. + +![](images/9997b68f631ffc2a65aa5bb483f341015fa350646cceb949eab096f9bebcc0f1.jpg) + +![](images/db9d53db31742015ccbd9f429ffcd1da82a716cfb10fe9f347044448f2efac82.jpg) + +# 2.3. Mixed Sample Data Augmentations (MSDA) + +Traditional data augmentation techniques assume all samples belong to the same class and do not consider relationships between different samples. In contrast, MSDA combines two samples, using various forms of label representation [35]. Popular methods include Mixup [47], Cutmix [45], Cutout [4], and GridMask [2], each with the core idea of producing images with more representative features of real-world scenarios. Other MSDA techniques include SmoothMix [20] and FMix [9]. + +Despite significant achievements in image classification, the above methods have not been applied to metric learning. [17] discussed the similarity between linear interpolation of features and mixup data augmentation, but due to the peculiarities of mixed labels, it cannot be directly applied to metric learning, which typically requires distinguishing samples as dichotomous, i.e., positive and negative samples. Recently, [38] modified an existing metric learning loss function to accommodate mixup, but the input mix was limited to mixup. Our proposed method can be applied to a variety of mixed sample data augmentations (Section 3.2). + +We modified multiple mixed sample data augmentations + +as shown in Figure. 2 to ensure that mixed data augmented samples contain features from multiple classes in a balanced manner. Specifically, for CutMix, we uniformly cropped and merged images from top to bottom to ensure that features from multiple classes are regularly distributed. For Mixup, we fused the samples in equal proportions to ensure a balanced representation of features from multiple classes. Lastly, we modified Gridmask by replacing the mask with an image to create a fusion effect. + +# 3. Proposed Method + +# 3.1. Preliminaries + +As a preliminary for our proposed method, we consider a batch of images denoted by $x = x_{1},x_{2},\ldots ,x_{n}$ with corresponding labels $y = y_{1},y_{2},\ldots ,y_{n}$ . Deep metric learning involves the use of a Convolutional Neural Network (CNN) to learn a non-linear transformation of each image into an m-dimensional deep feature space $\phi (x;\theta_{\phi}):\mathcal{X}\to \mathbb{R}^{m}$ , where $\theta_{\phi}$ represents the network's parameters. In order to further learn a mapping from the m-dimensional feature space to a k-dimensional space, we use a linear mapping layer represented by $f(\phi ;\theta_f):\mathbb{R}^m\to \mathbb{R}^k$ , where $\theta_{f}$ represents the parameters of the linear mapping layer. + +Formally, we define the distance between two data points in the embedding space as: + +$$ +d _ {f} \left(x _ {i}, x _ {j}\right) = \left\| f \left(\phi_ {i}\right) - f \left(\phi_ {j}\right) \right\| _ {2} \tag {1} +$$ + +Metric learning models usually involve pairwise sampling of a sample, e.g. considering the contrastive loss: + +$$ +L _ {\text {c o n t}} = y d _ {f} (a, x) ^ {2} + (1 - y) \left[ m - d _ {f} (a, x) \right] _ {+} ^ {2} \tag {2} +$$ + +where $y$ denotes the label, $a$ is the anchor image and $m$ is the marginal value. To train a metric learning model, we first select an anchor image and choose a positive sample with the same label ( $y = 1$ ) and a negative sample with a different label ( $y = 0$ ). The network is then trained to minimize a loss function that penalizes negative samples that are too close together and positive samples that are too far apart. It is noteworthy that the loss value will be relatively large when the distance between the anchor point and the negative sample is small, while it will be relatively small when the distance is very far. Thus, selecting hard samples during the selection of pairs will result in a larger gradient. + +Studies have tried to increase the number of hard samples for metric learning by generating more samples. However, linear interpolation methods can produce false positives and negatives. It is important to note that the binary metric learning label $y$ cannot be used as a one-hot label for image classification tasks. + +Our Hybrid Species Embedding (HSE) method addresses the challenges of creating additional hard samples + +and applying them to metric learning without label information to improve generalization. We address the challenge of mixed sample data augmentations by emulating human categorization behavior to assign these samples to a reasonable position in the embedding space. We propose incorporating newly generated samples as an extra training signal for the model, using the same feature extraction network for both the original image and the hybrid species generated using MSDA, and included in each batch. The loss function $L_{m}$ clusters the original samples based on their labels, while our proposed loss function $L_{Hy}$ deals with the embedding positions of the new samples, as shown in Figure. 3. + +![](images/3f3f8834389bb4c29fdf2b5df2134f3c7f9dc9d82cb9f1679c4af67014f63ab7.jpg) +Figure 3: Our HSE method uses batches of original images and hybrid species generated by Mixed Sample Data Augmentation (MSDA). The feature extraction network (CNN) processes both types of images to obtain their corresponding feature vectors ( $f_{Or}$ and $f_{HS}$ ). The loss function includes two components: $L_{m}$ (metric learning loss) groups original samples into clusters based on their labels, and $L_{Hy}$ optimizes feature extraction by comparing hybrid species embedding positions. + +![](images/0f7470e36f9792bcd72c73ce86cfeeb6211142287bd41f19ef52ea53f70f3047.jpg) +(a) HS-Anchors + +![](images/d76e21d9d9d1f69e835cbfeba513c23fa888e87e24752703d95c417cba06e81b.jpg) +(b) WP-Anchor +Figure 4: Three types of anchor point selection: The dashed lines indicate that the sample pairs should move away from each other, while the solid lines indicate that the sample pairs should move closer to each other. The red sample is a hybrid species synthesized from green and blue. (a) denotes the use of the hybrid species as the anchor point, (b) denotes using the class that synthesizes the hybrid species as an anchor point, and (c) denotes using a class unrelated to the hybrid species as an anchor point. + +![](images/690c4ff3dafa28070863a9518a8803f84530d6c74df6678b613435cd33493e99.jpg) +(c) N-Anchor + +# 3.2. Hybrid Species + +In the context of image classification tasks, hard negative samples refer to images that are similar to the anchor images but have different labels. We posit that hard negative samples typically possess similar features to the anchor samples. Therefore, when a sample is derived from an anchor image, there is a high likelihood that the sample will be classified as a hard sample. Let $x_{A}$ and $x_{B}$ denote two images with labels $y_{A}$ and $y_{B}$ , respectively, where $y_{A} \neq y_{B}$ . + +The generated sample $x_{H}$ contains both the features of $x_{A}$ and the features of $x_{B}$ . We refer to $x_{H}$ as the hybrid species. + +$x_{H}$ has a high likelihood of becoming a hard negative sample, while simultaneously presenting potential gradients in classes $A$ and $B$ . This can be achieved by utilizing mixed sample data augmentation techniques, such as mixup. Mixup generates a new training sample, denoted as $(x_{H},y_{H})$ , by linearly interpolating two training samples $(x_{A},y_{A})$ and $(x_{B},y_{B})$ , where $x_{H}$ and $y_{H}$ are: + +$$ +x _ {H} = \lambda x _ {A} + (1 - \lambda) x _ {B} \tag {3} +$$ + +$$ +y _ {H} = \lambda y _ {A} + (1 - \lambda) y _ {B} \tag {4} +$$ + +where $\lambda$ is obtained from a beta distribution and is constrained to [0, 1]. To achieve a distribution of $\lambda$ values that is centered around 0.5, $\lambda$ should follow a beta distribution with shape parameters $\alpha$ and $\alpha$ , e.g., $\lambda \sim \mathrm{Beta}(\alpha, \alpha)$ , where $\alpha \in (0, \infty)$ . The generated training sample $(x_H, y_H)$ is then used for model training with a loss function. + +Many pair-based metric learning losses rely on having distinct class labels, but the output $y_{H}$ is often a probability distribution per class, which is unsuitable for metric learning applications. Determining how to utilize these labels for metric learning presents a significant challenge. This difficulty is illustrated in Eq. 2. + +$$ +L _ {\text {c o n t}} = y d _ {f} \left(a, x _ {H}\right) ^ {2} + (1 - y) \left[ m - d _ {f} \left(a, x _ {H}\right) \right] _ {+} ^ {2} \tag {5} +$$ + +As the binary label $y \in \{0,1\}$ , the label for $x_{H}$ is not straightforward, since it is a hybrid sample. When we consider $x_{H}$ as a negative sample of the anchor point, it is not reasonable, as $x_{H}$ contains the same features as the anchor samples. Keeping them away from each other could cause the model to ignore important features and focus on irrelevant ones. Therefore, the binary label $y$ is invalid for $x_{H}$ . + +[38] showed that classification tasks are not fundamentally different from metric learning, and the label of the synthesized sample $X_{H}$ can still be represented by linear interpolation of the labels. Eq. 5 can then be converted to: + +$$ +\widetilde {L} _ {\text {c o n t}} = \lambda d _ {f} (a, x _ {H}) ^ {2} + (1 - \lambda) [ m - d _ {f} (a, x _ {H}) ] _ {+} ^ {2} \tag {6} +$$ + +Now, $x_{H}$ appears with nonzero contributions in both positive and negative terms for positive-negative pairs, due to the interpolation factor $\lambda \in [0,1]$ . + +However, we found that these methods are only applicable to mixup and may be subject to controversy in other forms of mixed sample data augmentations. This is due to the fact that mixup binds the images together evenly based on the assigned weight parameter, $\lambda$ . The same applies to labels In contrast, cutmix synthesizes $x_{H}$ by replacing a crop area of $x_{A}$ with that of $x_{B}$ , where the crop area of $\lambda$ + +
HS-AnorhWP-AnorhN-Anorh
Intra-class variation✓✓✕
Inter-class variation✓✓✓
Conflict✕✓✕
+ +Table 1: Three types of anchor point selection are considered for inter-class and intra-class variation. + +and $(1 - \lambda)$ is taken from $x_{A}$ and $x_{B}$ , respectively. Therefore, the label of $x_{H}$ cannot be accurately represented by linear interpolation, as the area being replaced is indeterminate, and may contain either significant features or irrelevant content. Consequently, the contribution of the crop area to $x_{H}$ remains inconclusive, making the application of linear interpolation for labeling mostly infeasible. + +Instead of fully utilizing the label information of $x_{H}$ , we can translate them into their appropriate spatial embedding, as we will describe in next subsection. Also, when applying MSDA to images, we need to consider how the labels are to be processed, which limits us to intermixing only two samples. In contrast, $x_{H}$ does not require us to define label information, and can hence be mixed by multiple classes. + +# 3.3. Hybrid loss + +# 3.3.1 Anchor Point Selection and Processing + +Although it is challenging to accurately assign a label to $x_{H}$ obtained through MSDA, the features of $x_{H}$ can be inferred based on the source images. We define $C(x_{I})$ to represent the label of image $x_{I}$ . When $x_{H}$ is selected as an anchor, it may not be possible to find a suitable positive sample, but $x_{H}$ shares semantic information with $C(x_{A})$ and $C(x_{B})$ to some extent. Thus, we can consider $C(x_{A})$ and $C(x_{B})$ as weak positive samples (WP) of $x_{H}$ , with the remaining classes treated as negative samples. + +During the training phase, the embedding of the spatial position of $x_{H}$ plays a crucial role in the generalization of downstream tasks. Therefore, deciding how to treat $x_{H}$ in pair selection becomes the first key question. Since $x_{H}$ contains information for multiple classes, including it in a paired selection can provide more gradient information. We divided the pairing methods into three categories, as shown in Figure. 4: HS (hybrid species sample) as the anchor point, WP (weak positive sample) as the anchor point, and N (negative sample) as the anchor point. We compared the three methods and and summarized the findings in Table 1. + +The choice of anchor point in a classification task can affect the learning of intra-class and inter-class variation. When using $x_{H}$ as an anchor point, it pulls towards the WP while staying away from the rest of the class. Using WP as an anchor point keeps it close to its composite $x_{H}$ to maintain identical points within the class, but may result in conflicts with the synthetic sample of $x_{H}$ leading + +to non-converging losses. Anchoring with N treats $x_{H}$ as a negative sample and moves it away from other samples, losing its purpose. Our findings recommend using HS as the anchor sample to effectively focus on both intra-class and inter-class variation while ensuring convergence. + +# 3.3.2 Strategies for Sample Mining + +In the previous subsection, we analyzed the contribution of pairwise selection to the embedding of $x_{H}$ . Since $x_{H}$ is randomly generated in each epoch, how to embed it more efficiently and reasonably becomes the second key issue of our study. For instance, when $x_{H}$ is used as an anchor point, the corresponding WP may contain multiple classes. If $x_{H}$ pulls all WP close together, clusters that are unrelated may merge, which is unreasonable. If $x_{H}$ randomly selects WP that are far apart, the structure of the cluster may be destroyed. On the other hand, if negative samples that are far apart are selected, the resulting gradient will be smaller. In summary, it is vital that we choose a reasonable and valid sample for performing metric learning with HS. + +Several studies have explored the advantages of hard negative mining in constructing informative gradients [12, 36, 34]. In this context, we introduce the notion of hard negatives for the hybrid species as: + +$$ +x _ {H _ {h n}} = \underset {x: C (x _ {I}) \neq C (x _ {A}) \cap C (x _ {I}) \neq C (x _ {B})} {\operatorname {a r g m i n}} d \left(f \left(x _ {H}\right), f \left(x _ {I}\right)\right) \tag {7} +$$ + +It is worth noting that $x_{H_{hn}}$ does not only yield a large gradient for $x_{H}$ , but it also possesses the same potential gradient as the inputs $C(x_{A})$ and $C(x_{B})$ that are used to synthesize $x_{H}$ . + +The study by [43] shows that easy positive mining improves the generalization performance of downstream tasks. Our experiments also demonstrate the benefits of this approach for $x_{h}$ . We define the easy hybrid weak positive of the hybrid species as: + +$$ +x _ {H _ {e w p}} = \underset {x: C \left(x _ {I}\right) = C \left(x _ {A}\right) \cup C \left(x _ {I}\right) = C \left(x _ {B}\right)} {\operatorname {a r g m i n}} d \left(f \left(x _ {H}\right), f \left(x _ {I}\right)\right) \tag {8} +$$ + +There are two primary reasons why bringing $x_{H}$ and $x_{H_{ewp}}$ closer together is beneficial. Firstly, it enables the model to concentrate on the features that are shared by both inputs, which can significantly influence the sample measurements. Secondly, this approach avoids the potential issue of pushing weak positives that are distant from each other too close together. By doing so, it preserves both the intra-class variance and the inter-class variance, which can help maintain the manifold structure of the classes in the embedding space. + +Our hybrid loss function involves mapping the output of the convolutional neural network onto a unit sphere, which is a widely accepted technique. During each mini-batch, + +1 # Nsize: Number of samples of the same class in each batch +2 # n: The number of classes that synthesize $x_{H}$ +3 # b: batch size d: dimension +4 # F = b × d +5 # Map of feature similarity between samples +6 Map = F × FT; +7 # $x_{H}$ as the anchor, look for weak positive sample +8 DWP = Map; +9 for each i ∈ [b - Nsize, b] do +10 for each j ∈ [n * Nsize, b] do +11 | DWP[i][j] = -1; +12 end +13 end +14 # Each $x_{H}$ corresponds to the value and index of the most similar WP sample +15 VDWP, IDWP = DWP.max(1); +16 # $x_{H}$ as the anchor, look for negative sample +17 DN = Map; +18 for each i ∈ [b - Nsize, b] do +19 for each j ∈ [0, n * Nsize] do +20 | DN[i][j] = -1; +21 end +22 for each j ∈ [b - Nsize, b] do +23 | DN[i][j] = -1; +24 end +25 end +26 # Each $x_{H}$ corresponds to the value and index of the most similar negative sample +27 VDN, IDN = DN.max(1); +28 # Select the WP sample and negative sample parts corresponding to $x_{H}$ +29 T = [VDWP[-Nsize:], VDN[-Nsize:]]; +30 Hybrid loss = -α logsoftmax(T).mean(); + +# Algorithm 1: Hybrid Loss + +when $x_{H}$ serves as the anchor, we select $x_{H_{ewp}}$ and $x_{H_{hn}}$ for comparison. As in Algorithm 1, the former corresponds to the original sample, while the latter represents $x_{H}$ . To synthesize $x_{H}$ , we typically select samples from the first $n$ classes. Then, we search for the most similar weak positive sample and negative sample for $x_{H}$ , respectively. + +The NCA loss function is an effective way to avoid the selection of margin hyperparameters and handle the distance between samples efficiently[46]. To ensure that hybrid species are embedded in a suitable location, we define the hybrid loss based on the NCA loss: + +$$ +L _ {H y} = - \alpha \log \frac {1}{1 + \frac {\exp (f (x _ {H}) ^ {\top} f (x _ {H _ {h n}}))}{\exp (f (x _ {H}) ^ {\top} f (x _ {H _ {e w p}}))}} \tag {9} +$$ + +where $\alpha$ is a hyper-parameter that controls the balance be + +tween bringing $x_{H}$ and $x_{H_{ewp}}$ closer and pulling $x_{H}$ and $x_{H_{hn}}$ farther apart. The $L_{Hy}$ loss serves as an auxiliary loss to process additional samples and can be combined with other metric learning losses. As in Figure 3, the final loss is the sum of the metric learning loss $L_{m}$ and the $L_{Hy}$ loss: + +$$ +L = L _ {m} + L _ {H y} \tag {10} +$$ + +When incorporating our method with the contrastive loss (Eq. 2), the resulting final loss is: + +$$ +L = L _ {\text {c o n t}} + L _ {H y} \tag {11} +$$ + +# 4. Experiments + +Our HSE method uses pre-trained BN-Inception [14] and ResNet [11] models on ILSVRC 2012-CLS [32] with PyTorch. These models have learned feature representations from diverse images, reducing time and resources for new dataset training. They also perform well in image recognition tasks, making them suitable for computer vision applications. For generalization ability evaluation, we maintain consistency with base methods on image preprocessing, learning rate, and batch size. (Source code: https://github.com/SHQberserker/HSE) + +Nsize examples from each class are selected per batch. We use two methods to incorporate hybrid species during training: selecting multiple classes to add hybrid species as additional examples of batch, and randomly replacing examples in the batch with hybrid species. We use the cutmix technique to create hybrid species by stitching two samples together evenly. Algorithm 1 fixes the last Nsize samples of each batch as synthetic samples. + +# 4.1. Datasets and Metrics + +We tested our framework's ability to perform well on challenging datasets: CUB-200, CARS-196, and SOP. These datasets involve fine-grained classification tasks, where there are many categories and relatively few images per category. We followed the data split approach from [22]. Specifically, for CUB-200, we used 5864 images from the first 100 categories for training and 5924 images from the last 100 categories for testing. For CARS-196, we trained on the first 98 classes and tested on the remaining images. Additionally, for SOP, we utilized the first 59,551 images (11,318 classes) for training and the remaining 60,502 images (11,316 classes) for testing. + +To assess our experiments, we employed the Recall@K metrics. This measure involves selecting the K images that are closest to the query image. If there's an image with the same label as the query image, we score 1; otherwise, it's 0. Recall@K calculates the average score for all query images, with higher values indicating better retrieval performance. These metrics provide a way to quantitatively evaluate how + +well our model can find similar images, which is crucial for many image-based applications. + +# 4.2. Results and Analysis + +# 4.2.1 Comparison + +Experiments compared our proposed method to existing methods, including N-pair, EPSHN, MSloss and ProxyAnchor. Results are shown in Table 2 for CUB-200 and CARS-196 datasets, where * indicates our results under the same experimental settings. Our method improves on different approaches to varying degrees. Our method had less impact on the N-pair method as it mostly ignores intraclass variation, leading to excessive clustering between the same classes. However, we observed a significant improvement in recall@K when the EPSHN method was augmented with the additional training signal of HSE. Recall@1 improved by $2.5\%$ and $2.3\%$ on the challenging CUB-200 and CARS-196 datasets, respectively, where Nsize=16 and Nsize=8, respectively. Importantly, HSE-E outperformed the ensemble-based method that EPSHN had not surpassed before on the CUB-200 and CARS-196 datasets. + +We believe that EPSHN tends to ignore intra-class variations, such as those found in [43] where birds of different sexes may vary greatly within the same class, but should also focus on characteristics common between members of the same class. When $x_{H}$ is added, it helps the network to focus more on these common characteristics. MSloss focuses on both inter-class variation and inter-class variation. However, when adding our additional training signals, there is still a certain improvement. This is because the features that our additional training signals focus on help the network generalize to unknown classes. + +Our method achieves good results in pair-based methods and is effective in proxy-based metric learning methods. We improve approximately $1\%$ in both $\mathrm{Recall}@1$ and $\mathrm{Recall}@2$ . This confirms our universality, which can be applied to various metric learning loss functions. We also compared our method to the Matrix/Input [38] approach. Matrix/Input has a high computational cost because it calculates $\frac{1}{2} n(n - 1)$ embeddings for a batch of $n$ examples. Instead, our method achieves a similar effect using only $Nsize * Cnum$ embeddings. Our approach is hence more computationally efficient. We also remain applicable to discriminate between positives and negatives for every anchor point. We can also incorporate non-additive component functions that involve loss functions, e.g., EPSHN. + +HSE is also effective on the SOP dataset. Unlike smaller datasets, the SOP dataset has superclasses, which introduce more variation between different superclasses and less variation within the same superclass. It's challenging to measure similarity between samples within the same superclass in the SOP dataset. As shown in Table 3, we generate HS from samples that share the same superclass but have differ + +ent labels. When we apply HSE to various classical methods, we observe varying degrees of improvement. + +# 4.2.2 Impact of Minibatch Parameters + +# Impact of Nsize + +We trained a ResNet-18 embedding network on CUB-200 and CARS-196 datasets with batch size 128 and embedding dimension 512 to investigate HSE's impact on model generalization at various Nsizes. HSE's effectiveness varies with methods, datasets, and Nsizes (Figure. 5). The best performance occurs at Nsize between 4 and 16 with a batch size of 128, followed by performance degradation. When there are too few negative samples, the hybrid species' gradient information becomes insignificant, and they may be grouped with WP's clusters, harming downstream tasks. + +![](images/4ab7a647d9ff83c4f2c2886e6e56a38787378f6d57445c7bdba40a062c6741cf.jpg) +(a) CUB + +![](images/7c1703200f28fff10e20d066a809a03603f67910656e2ddfb97b9ff9066a063c.jpg) +(b) CAR + +![](images/6d3bb9bfec150f1eb655e584af25b56eadd459f0c7714e5cc28d71cf2e6d5480.jpg) +Figure 5: Effect of additional training signals provided by different Nsizes. +Figure 6: Results on the CUB dataset using different Cnum in MS loss. + +# Impact of Cnum + +We found that HSE improves the generalization ability of the model. However, increasing the number of hybrid species does not necessarily lead to better results. We conducted experiments on the CUB dataset using the MS loss method, as shown in Figure. 6. Our loss function indirectly affects the entire embedding space of the batch when learning the spatial embedding of a single hybrid species. After training for a certain period, the two combined classes have essentially covered the entire dataset. + +# 4.2.3 Effect of Hybrid Species + +We investigated the effectiveness of different data augmentation techniques in providing additional training signals for hybrid species. Table 4 shows the classic blending data augmentation techniques we used, with certain modifications made to each method. Notably, CutMix and GridMask + +
CUB-200CARS-196
Recall@kk=1k=2k=4k=1k=2k=4
Angular512[39]G53.665.075.371.380.787.0
HDML512[49]G53.765.776.779.187.192.1
HTL512[6]BN57.168.878.781.488.092.7
Margin128[41]R63.674.483.179.686.591.9
SoftTriple512[28]BN65.476.484.584.590.794.5
DR512[24]BN66.177.085.185.090.594.1
Ensemble-based
HDC384[44]G53.665.777.073.282.486.4
A-BIER512[27]G55.367.276.978.085.891.1
ABE-8512[16]G60.671.579.860.671.579.8
DREML9216[42]R63.975.083.186.091.795.0
Npair64*(Nsize=8)[37]R51.364.275.364.276.084.6
HSE-N64(Nsize=8)R52.5+1.266.0+1.877.4+2.164.9+0.776.6+0.684.7+0.1
EPSHN512*(Nsize=16)[43]R64.975.383.583.189.793.6
HSE-E512(Nsize=16)R66.9+2.077.4+2.185.5+2.085.4+2.391.2+1.596.9+3.3
HSE-E512(Nsize=8)R67.4+2.577.7+2.485.7+2.284.8+1.790.6+0.994.3+0.7
MS†512R67.877.885.687.892.795.3
+Matrix输入[38]R69.0+1.279.1+1.386.0+0.489.0+1.293.4+0.796.0+0.7
MS512*(Nsize=5)[40]BN65.777.086.381.788.793.2
HSE-M512(Nsize=5)BN67.6+1.978.0+1.085.8-0.582.0+0.388.9+0.293.3+0.1
Proxy-Ancho†512R69.579.387.087.692.395.5
+Matrix输入[38]R70.5+1.081.2+1.987.8+0.888.2+0.693.2+0.996.2+0.7
Proxy-Ancho*512[15]R69.479.287.088.592.795.6
HSE-PA512R70.6+1.280.1+0.987.1+0.189.6+1.193.8+1.196.0+0.4
+ +Table 2: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the CUB-200-2011 and Cars-196 datasets. The backbone networks used in the models are denoted by abbreviations: G for GoogleNet, BN for Inception with batch normalization, and R for ResNet50. The symbol $\dagger$ indicates the reproduced result reported by the original authors.N-pair and EPSHN use cnum=1 and batch size of 128. MSloss uses cnum=1, batch size of 80, and $\alpha = 0.05$ . Proxy-Anor uses cnum=1, batch size of 128, and $\alpha = 2$ . + +
SOP
Recall@kk=1k=10k=100
EPSHN512*(Nsize=5) [43]R74.386.994.3
HSE-E512(Nsize=5)R76.3+2.088.4+1.594.9+0.6
MS†512R76.989.895.9
+Matrix输入[38]R77.9+1.090.6+0.895.9
MS512*[40] (Nsize=5)R78.290.296.2
HSE-M512R78.7+0.590.4+0.296.1-0.1
Proxy-Anchort512R79.190.896.2
+Matrix输入[38]R79.8+0.791.4+0.696.5+0.3
Proxy-Anchort*512[15]R79.290.495.8
HSE-PA512(Nsize=5)R80.0+0.891.4+1.096.3+0.5
+ +stand out. CutMix can incorporate features from each compositional image to a certain extent, but may also diminish key features. Mixup dilutes key features as the number of samples increases. GridMask (block=50) covers the entire image despite using relatively small blocks. Results show that CutMix is the most effective approach for improving the performance of Hybrid Species (HS). We also found that adding more class features does not necessarily improve HS performance, but adding critical features that are specific to a particular class can lead to better HS results. + +We investigate the selection of weak positive samples + +Table 3: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the SOP.Where batch $= {120}$ ,cnum $= {12}$ ,and hybrid species as additional examples of batch. + +
CUBCAR
R@1numR@1num
Cutmix67.4285.42
67.2385.03
67.1484.24
Mixup67.0285.32
66.9381.03
R@1MasksizeR@1Masksize
Gridmask67.45084.950
66.015084.8150
67.125084.0250
+ +Table 4: Results on the Cub dataset using different Mixed Sample Data Augmentations in EPSHN. $(\mathbb{R}@\mathbb{1} = \mathbb{Recall}@\mathbb{1})$ + +by hybrid species using the Proxy-Anorh method on the CUB dataset. Cutmix augmentation technique is used to stitch images either vertically or horizontally, with the upper part mainly containing the bird's head and the lower part containing the bird's body when stitched vertically. When stitched horizontally, the probability of the left and right sides containing the bird's head is almost the same. Our HSE approach improves model performance by focusing on essential image features. Our study on the CUB dataset demonstrates that the hybrid species prioritize the bird's + +![](images/2816ec965166d5a44ddac96594c047db7957582107bec461f7f4fc1b48d98608.jpg) +Figure 7: Analysis of easy weak positive samples selected by hybrid species using Proxy-Antor on CUB dataset. Hybrid species are obtained by stitching two images of different labels using cutmix. We quantify the probability that selected easy weak positive samples contain upper or bottom half during training using "aut_up/down". With our method, we also report "our_up/down" and "our_left/right" indicating the likelihood of easy weak positive samples containing upper or bottom half of the bird or left or right half of the stitched image, respectively. + +body during metric learning, particularly the lower half, which is synthesized as the easy weak positive sample using the Cutmix augmentation technique. While the selection probability of the lower half increases from 0.6 to $0.7\sim 0.8$ during Proxy training, our proposed method ensures a probability of 1. Stitching hybrid species horizontally results in a selection probability of around 0.5. This approach allows the network to focus on important features during training, providing an advantage over traditional methods. Figure. 7 provides a visual representation of our findings. + +# 4.2.4 Visualization of Embedding Space + +We demonstrate the effectiveness of our proposed method on training and test data using t-SNE (t-distributed stochastic neighbor embedding) visualization technique. t-SNE can effectively reduce the dimensionality of the embedding space while preserving pairwise similarity between samples. It enables us to visually inspect how well the model separates different classes and identifies clusters of similar samples by plotting the embedding space in 2D or 3D. Therefore, t-SNE is a useful tool for evaluating the quality of learned embeddings in deep metric learning. + +We experimented on two classes from the training set and two classes (wren and woodpecker) from the test set. Figure. 8 shows that EPSHN is effective at distinguishing between different classes in the training set, but the embedding for hybrid species is relatively dispersed. With the additional HSE training signal, the embedding of the hybrid species becomes more concentrated and is located between the two classes. This approach considers all the features of difficult samples, creating a more comprehensive embedding of the samples. When additional training signals + +are learned, the model can generalize better to unknown classes. This is demonstrated in the test set, where the EPSHN method cannot distinguish the black classes from the green and blue classes. However, when the training signal of HSE is applied to the downstream task, black classes can be successfully distinguished from green classes. + +![](images/ba07d0fc9327cface7ddee83aab881ce6f9d14292d11471cd69e6b5fde50df37.jpg) +(a) Oriole $(\mathrm{id} = 95)$ + +![](images/50b0c5aa910ac598697179d39464f988ccd9c826c9adcea56fbad2b1c9da570f.jpg) +(b) Nuthatch (id=94) + +![](images/a5787db3e0af7e6b248afe79b3790ffffc34d77409d977b70c786fd3aed4b978.jpg) +(c) Hspecie + +![](images/b737f442c716e23eac9a6cf31d86497b3e7d55ffe9a104c89bba79382c6b3970.jpg) +EPSHN (train) + +![](images/fd43c4924038cbe776a1120c63747384b1c13ba4f0bcce3bfc6cc8987ac4503f.jpg) +HSE-EPSHN (train) + +![](images/b418f8864e9908fbb992bce2ca470bee34251594dce4113ae98c072e5f8a7ded.jpg) +EPSHN (test) + +![](images/2f63165eb38949092075af43c9622c4acc9d0bf2f62bee5c73937eb4751f9043.jpg) +HSE-EPSHN (test) +Figure 8: We present a t-SNE visualization of EPSHN and HSE-EPSHN on the test set of CUB. Specifically, we select two different classes (denoted by blue dots and green dots) from the training set as shown in (a) and (b), respectively. (c) is the synthetic class from classes (a) and (b), denoted by red dots. For the test set, we selected four bird categories, consisting of two types of wrens and two types of woodpeckers, where each color represents a specific category. + +# 5. Conclusion + +Our Hybrid Species Embedding (HSE), a metric learning approach that aims to enable models to learn the spatial embedding of unknown and difficult classes during training. HSE generates additional hard samples using mixed data augmentation methods without defining their label information and applies them to the metric learning process. Experiments demonstrate significant improvements on multiple datasets when compared to three widely used pair-based methods and state-of-the-art proxy methods. HSE enhances the generalization ability of the model by providing additional training signals in deep metric learning. + +Acknowledgment. This work is partly supported by the National Natural Science Foundation of China (No: 62172366) and the Pioneer and Leading Goose R&D Program of Zhejiang Province (2023C01150). + +# References + +[1] E. Ahmed, M. Jones, and T. K. Marks. An improved deep learning architecture for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3908–3916, 2015. +[2] P. Chen, S. Liu, H. Zhao, and J. Jia. Gridmask data augmentation. arXiv preprint arXiv:2001.04086, 2020. +[3] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. +[4] T. DeVries and G. W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. +[5] Z. Fu, Z. Mao, B. Hu, A.-A. Liu, and Y. Zhang. Intra-class adaptive augmentation with neighbor correction for deep metric learning. IEEE Transactions on Multimedia, 2022. +[6] W. Ge. Deep metric learning with hierarchical triplet loss. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-285, 2018. +[7] G. Gu, B. Ko, and H.-G. Kim. Proxy synthesis: Learning with synthetic classes for deep metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1460-1468, 2021. +[8] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006. +[9] E. Harris, A. Marcu, M. Painter, M. Niranan, A. Prügel-Bennett, and J. Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020. +[10] B. Harwood, V. Kumar BG, G. Carneiro, I. Reid, and T. Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 2821-2829, 2017. +[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[12] A. Hermans, L. Beyer, and B. Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. +[13] E. Hoffer and N. Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015. +[14] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. +[15] S. Kim, D. Kim, M. Cho, and S. Kwak. Proxy anchor loss for deep metric learning. IEEE, 2020. +[16] W. Kim, B. Goyal, K. Chawla, J. Lee, and K. Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018. +[17] B. Ko and G. Gu. Embedding expansion: Augmentation in embedding space for deep metric learning. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7255-7264, 2020. +[18] B. Ko, G. Gu, and H.-G. Kim. Learning with memory-based virtual classes for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11792-11801, 2021. +[19] M. T. Law, N. Thome, and M. Cord. Quadruplet-wise image similarity learning. In Proceedings of the IEEE international conference on computer vision, pages 249-256, 2013. +[20] J.-H. Lee, M. Z. Zaheer, M. Astrid, and S.-I. Lee. Smoothmix: a simple yet effective data augmentation to train robust classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 756-757, 2020. +[21] L. Liu, S. Huang, Z. Zhuang, R. Yang, M. Tan, and Y. Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022. +[22] Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. +[23] T. Milbich, K. Roth, H. Bharadhwaj, S. Sinha, Y. Bengio, B. Ommer, and J. P. Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In European Conference on Computer Vision, pages 590-607. Springer, 2020. +[24] D. D. Mohan, N. Sankaran, D. Fedorishin, S. Setlur, and V. Govindaraju. Moving in the right direction: A regularization for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14591-14599, 2020. +[25] Y. Movshovitz-Attias, A. Toshev, T. K. Leung, S. Ioffe, and S. Singh. No fuss distance metric learning using proxies. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. +[26] H. Oh Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016. +[27] M. Opitz, G. Waltner, H. Possegger, and H. Bischof. Deep metric learning with bier: Boosting independent embeddings robustly. IEEE transactions on pattern analysis and machine intelligence, 42(2):276-290, 2018. +[28] Q. Qian, L. Shang, B. Sun, J. Hu, H. Li, and R. Jin. Soft-triple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019. +[29] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +[30] K. Roth, O. Vinyals, and Z. Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022. + +[31] K. Roth, O. Vinyals, and Z. Akata. Non-isotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022. +[32] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. +[33] A. Sanakoyeu, V. Tschernezki, U. Buchler, and B. Ommer. Divide and conquer the embedding space for metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 471-480, 2019. +[34] F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. +[35] C. Shorten and T. M. Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. +[36] E. Simo-Serra, E. Trulls, L. Ferraz, I. Kokkinos, P. Fua, and F. Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. +[37] K. Sohn. Improved deep metric learning with multi-class n-pair loss objective. Advances in neural information processing systems, 29, 2016. +[38] S. Venkataramanan, B. Psomas, Y. Avrithis, E. Kijak, L. Amsaleg, and K. Karantzalos. It takes two to tango: Mixup for deep metric learning. arXiv preprint arXiv:2106.04990, 2021. +[39] J. Wang, F. Zhou, S. Wen, X. Liu, and Y. Lin. Deep metric learning with angular loss. In Proceedings of the IEEE international conference on computer vision, pages 2593-2601, 2017. +[40] X. Wang, X. Han, W. Huang, D. Dong, and M. R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5022-5030, 2019. +[41] C.-Y. Wu, R. Manmatha, A. J. Smola, and P. Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017. +[42] H. Xuan, R. Souvenir, and R. Pless. Deep randomized ensembles for metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 723-734, 2018. +[43] H. Xuan, A. Stylianou, and R. Pless. Improved embeddings with easy positive triplet mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2474-2482, 2020. +[44] Y. Yuan, K. Yang, and C. Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE international conference on computer vision, pages 814-823, 2017. +[45] S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo. Cutmix: Regularization strategy to train strong classifiers with + +localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. +[46] D. Zhang, Y. Li, and Z. Zhang. Deep metric learning with spherical embedding. Advances in Neural Information Processing Systems, 33:18772-18783, 2020. +[47] H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. +[48] Y. Zhao, Z. Jin, G.-j. Qi, H. Lu, and X.-s. Hua. An adversarial approach to hard triplet generation. In Proceedings of the European conference on computer vision (ECCV), pages 501-517, 2018. +[49] W. Zheng, Z. Chen, J. Lu, and J. Zhou. Hardness-aware deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 72-81, 2019. +[50] W. Zheng, C. Wang, J. Lu, and J. Zhou. Deep compositional metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9320-9329, 2021. +[51] W. Zheng, B. Zhang, J. Lu, and J. Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021. \ No newline at end of file diff --git a/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/images.zip b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cb29819cbec08a2b60984d481ed1d0274652bf34 --- /dev/null +++ b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cb41cb11068523097b63b2385db682b151ea272a95920215a15548a7e8d0cf4 +size 458230 diff --git a/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/layout.json b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..209e91629315fa039bfd671de186335790b5d9b2 --- /dev/null +++ b/2023/HSE_ Hybrid Species Embedding for Deep Metric Learning/layout.json @@ -0,0 +1,11873 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 481, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 481, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 481, + 121 + ], + "type": "text", + "content": "HSE: Hybrid Species Embedding for Deep Metric Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "spans": [ + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "text", + "content": "Bailin Yang" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "text", + "content": ", Haoqiang Sun" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "text", + "content": ", Frederick W. B. Li" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "text", + "content": ", Zheng Chen" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "text", + "content": ", Jianlu Cai" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "text", + "content": ", Chao Song" + }, + { + "bbox": [ + 65, + 142, + 527, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 162, + 489, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 162, + 489, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 162, + 489, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 162, + 489, + 177 + ], + "type": "text", + "content": "Faculty of Computer Science and Technology, Zhejiang Gongshang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 177, + 469, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 177, + 469, + 191 + ], + "spans": [ + { + "bbox": [ + 123, + 177, + 469, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 177, + 469, + 191 + ], + "type": "text", + "content": "Faculty of Statistics and Mathematics, Zhejiang Gongshang University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 157, + 191, + 436, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 191, + 436, + 205 + ], + "spans": [ + { + "bbox": [ + 157, + 191, + 436, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 191, + 436, + 205 + ], + "type": "text", + "content": "Department of Computer Science, University of Durham" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 233, + 191, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 233, + 191, + 245 + ], + "spans": [ + { + "bbox": [ + 143, + 233, + 191, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 258, + 290, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 413 + ], + "type": "text", + "content": "Deep metric learning is crucial for finding an embedding function that can generalize to training and testing data, including unknown test classes. However, limited training samples restrict the model's generalization to downstream tasks. While adding new training samples is a promising solution, determining their labels remains a significant challenge. Here, we introduce Hybrid Species Embedding (HSE), which employs mixed sample data augmentations to generate hybrid species and provide additional training signals. We demonstrate that HSE outperforms multiple state-of-the-art methods in improving the metric Recall@K on the CUB-200, CAR-196 and SOP datasets, thus offering a novel solution to deep metric learning's limitations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 456, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 576 + ], + "type": "text", + "content": "Image retrieval heavily depends on deep metric learning to grasp visual similarities. In its early stages, methods like pair-based loss [8, 13, 19, 37] and proxy-based loss [25, 15] were proposed to train models. Nevertheless, the true goal of image retrieval is to adapt to the unknown, and the challenge lies in selecting a metric that can effectively handle differences between classes in deep metric learning. A promising approach involves exploring both intra-class and inter-class variations within the image itself, rather than relying solely on label information [33, 43, 40, 23, 51, 30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 576, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 288, + 696 + ], + "type": "text", + "content": "Hard negative samples, often termed false positives, refer to images that resemble anchor images but carry different labels. The reason behind these hard samples lies in their potential for sharing remarkably similar or even identical features with the anchor samples. Approaches like hard sample mining and generation have been proposed to aid network convergence by introducing a substantial gradient [1, 41, 12, 34, 49, 48, 17, 18]. Recently, an advancement in generating appropriate hard samples involves creating supplementary training data [17, 38, 5, 21]. This is primarily" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 341, + 234, + 509, + 266 + ], + "blocks": [ + { + "bbox": [ + 341, + 234, + 509, + 266 + ], + "lines": [ + { + "bbox": [ + 341, + 234, + 509, + 266 + ], + "spans": [ + { + "bbox": [ + 341, + 234, + 509, + 266 + ], + "type": "image", + "image_path": "ecaa3d20883361c9e06584a465127b95d5c53a605b8010cf6140841665824105.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 313, + 266, + 421, + 331 + ], + "blocks": [ + { + "bbox": [ + 313, + 266, + 421, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 266, + 421, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 421, + 331 + ], + "type": "image", + "image_path": "36588301f77640667a9b9698f1cbaf42aba4020e6afbee62c4477cdcfc64763b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 333, + 405, + 342 + ], + "lines": [ + { + "bbox": [ + 328, + 333, + 405, + 342 + ], + "spans": [ + { + "bbox": [ + 328, + 333, + 405, + 342 + ], + "type": "text", + "content": "(a) Linear interpolation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 267, + 539, + 332 + ], + "blocks": [ + { + "bbox": [ + 430, + 267, + 539, + 332 + ], + "lines": [ + { + "bbox": [ + 430, + 267, + 539, + 332 + ], + "spans": [ + { + "bbox": [ + 430, + 267, + 539, + 332 + ], + "type": "image", + "image_path": "a1056ed98209311971a51ea596e06753ad947904072e5f29a9c65623e85871de.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 334, + 507, + 342 + ], + "lines": [ + { + "bbox": [ + 464, + 334, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 464, + 334, + 507, + 342 + ], + "type": "text", + "content": "(b) Our HSE" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 357, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 453 + ], + "type": "text", + "content": "Figure 1: (a) Deep metric learning methods struggle with linear interpolation for generating new training samples with labels when new samples fall between anchors of different classes. The manifold structure lacking clear boundaries makes it challenging to assign a label, even when new samples are between anchors of the same class. (b) Samples with multiple simultaneous features may not lie on a linear interpolation path, and their location may be unknown. They can be classified as an unknown class situated inside or outside a manifold boundary. To tackle this challenge, we propose using the HSE method to allocate the sample to an appropriate embedding space location based on its feature information." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 473, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 641 + ], + "type": "text", + "content": "achieved through linear interpolation, a prevalent method for generating synthetic samples. However, assigning absolute labels to these synthesized samples is intricate due to the manifold structures and cluster boundaries present in the embedded space [17]. Even when the new samples are formed by linear interpolation between samples of the same label, accurately determining absolute labels remains a challenge. Synthetic sample labels have emerged as a potential solution, with [38] corroborating their similarity to mixup data augmentation and highlighting the label's representation through linear interpolation. Nonetheless, this approach is constrained to mixup data augmentation, as mixup uniformly blends features and labels from different images, which might not hold true for other techniques." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "We introduce a novel Hybrid Species Embedding (HSE) that leverages mixed sample data augmentations to generate additional training samples [35, 47, 45, 2, 20]. HSE creates hybrid species by combining features from multiple classes and embedding them into spatial positions. During training, different classes are randomly chosen for data augmentation" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 208, + 713 + ], + "type": "text", + "content": "*Corresponding to:csong@zjsu.edu.cn" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11047" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "in each batch. The labels of hybrid species are disregarded, and their placement near the category with the most similar features simulates human categorization behavior, as depicted in Figure 1(b). This strategy positions hybrid species close to the samples they are synthesized from, ensuring similarity while avoiding unrelated samples. By synthesizing samples using various mixed sample data augmentation techniques, each incorporating multiple features of a known class, our approach naturally imparts similar characteristics to the hard sample for the model without requiring explicit mining. Additionally, we dynamically adjust the position of each batch of hybrid species embedding to bypass the need for considering label information of these hybrid instances. Our contributions encompass:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 287, + 383 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": "- Introducing a novel metric learning strategy embedding challenging and unseen classes (Hybrid species), during training. These hybrid species offer supplementary training signals, enhancing the generalization capacity of downstream tasks, as validated through experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 300, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 347 + ], + "type": "text", + "content": "- Adapting mixed sample data augmentation techniques for our Hybrid Species Embedding (HSE) to address constraints associated with pair-based metric learning losses, which typically demand explicit class labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 347, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 287, + 383 + ], + "type": "text", + "content": "- Demonstrating through experiments our efficacy in enhancing performance across state-of-the-art metric learning tasks on CUB-200, CAR-196, and SOP datasets." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 393, + 134, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 134, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 134, + 405 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 413, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 510 + ], + "type": "text", + "content": "We utilize the Mixed Sample Data Augmentation (MSDA) method to create unknown samples and improve the model's generalization ability for downstream tasks. This approach resolves the issue of insignificant comparisons caused by random selection of samples in metric learning. While mining hard samples from a finite training set is limited, MSDA enables us to create more hard samples. In the following, we review existing relevant works." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 515, + 148, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 148, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 148, + 528 + ], + "type": "text", + "content": "2.1. Metric Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "type": "text", + "content": "Deep metric learning compares data in pairs and pulls them together if they belong to the same category, otherwise separates them [8, 13, 19, 37, 26]. The exponential growth in the number of tuples during training can present challenges in achieving model convergence, and may include meaningless metrics. Proxy-based approaches, e.g., choosing a proxy point for each class, have been proposed as a solution [25, 15]. These proxies contain information from multiple samples, require less computation in pairwise comparisons, helping the network converge more effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "Methods that rely solely on label information may result in over-clustered or overly separated samples with identical or differing labels, respectively, and can lead to reduced performance on new and unseen test samples. EPSHN [43] notes this problem, particularly for species with varying" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "features such as birds of the same species with different sexes. To avoid this, [30, 40, 31, 33, 51, 50, 18] emphasize the importance of incorporating semantic information from the images themselves, often through large pre-trained language models like CLIP [29] and BERT [3], to better align the visual representation space and discover potential differences within and between classes. The MSloss function [40] is one such approach that combines self-similarity and relative similarities to weight selected pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 180, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 324 + ], + "type": "text", + "content": "Depending solely on label information can be insufficient for effective sample learning. Previous works addressed this limitation by assigning varying metrics or weights based on inter-sample variability. However, these techniques had limitations when dealing with a small number of training samples or unseen samples. To address this, we introduce additional training samples and a novel metric for improved learning. Prior works like [49, 17] added additional samples, but mainly relied on linear interpolation. In contrast, our approach generates extra training samples through selecting the appropriate MSDA, providing greater flexibility than linear interpolation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 334, + 428, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 428, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 428, + 346 + ], + "type": "text", + "content": "2.2. Hard Sample Mining" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 472 + ], + "type": "text", + "content": "Pair-based metric learning generates lots of paired samples, and handling negative samples is crucial for reducing model redundancy. Various studies [6, 10, 34, 41, 49] have explored the use of hard negative mining while constructing triples to create useful gradient sums, aiding fast convergence of triplet loss networks. Additionally, [34] suggested a semi-hard negatives scheme that selects even harder negative samples, while [6] introduced a hierarchical triplet loss (HTL) that constructs a hierarchical tree of all classes and selects semi-hard negatives with a dynamic margin." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": "The discussed methods require selecting hard samples from the training set to improve model performance. Hard samples are challenging examples that can evaluate the model's ability to generalize. For example, if a model performs well on horses and donkeys but struggles with mules, evaluating its performance on mules (as a hybrid of both species) can directly assess its generalization ability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Recent studies focus on generating additional hard samples for training. HDML [49] creates hard samples between different labeled samples, while [17] generates additional training samples between identically labeled samples through hard sample mining. Gu et al. [7] use synthetic proxy points and competition to create hard samples. These methods use linear interpolation to generate new samples, leading to false negatives and positives due to the difficulties in determining class boundaries on the manifold structure. DAS [21] and IAA [5] try to solve this issue by adjusting features slightly and adding extra examples during mid-training of the model, aiming to make sure that new examples still belong to the same class. Despite these attempts," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11048" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 272, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 272, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 272, + 84 + ], + "type": "text", + "content": "completely removing this problem remains challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 181 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 181 + ], + "type": "text", + "content": "We use mixed sample data augmentations in the preprocessing stage of images to generate additional hard samples, instead of linear interpolation. This has several advantages, including a higher probability of generating hard samples without subsequent mining and the ability to identify the underlying content of the image even without its labeling information. This makes our method highly effective for generating and processing additional hard samples." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 102, + 193, + 135, + 226 + ], + "blocks": [ + { + "bbox": [ + 102, + 193, + 135, + 226 + ], + "lines": [ + { + "bbox": [ + 102, + 193, + 135, + 226 + ], + "spans": [ + { + "bbox": [ + 102, + 193, + 135, + 226 + ], + "type": "image", + "image_path": "b3d215d90ebbd4282c7727b890f1e5fddb6a32474d52211a4b39466424271b69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 226, + 223, + 235 + ], + "lines": [ + { + "bbox": [ + 112, + 226, + 223, + 235 + ], + "spans": [ + { + "bbox": [ + 112, + 226, + 223, + 235 + ], + "type": "text", + "content": "Original images in the training set" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 193, + 167, + 226 + ], + "blocks": [ + { + "bbox": [ + 135, + 193, + 167, + 226 + ], + "lines": [ + { + "bbox": [ + 135, + 193, + 167, + 226 + ], + "spans": [ + { + "bbox": [ + 135, + 193, + 167, + 226 + ], + "type": "image", + "image_path": "b6928a2fc5121385fabe763a7eef27da2212d9c75cde2037caf971a210f99677.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 194, + 200, + 226 + ], + "blocks": [ + { + "bbox": [ + 168, + 194, + 200, + 226 + ], + "lines": [ + { + "bbox": [ + 168, + 194, + 200, + 226 + ], + "spans": [ + { + "bbox": [ + 168, + 194, + 200, + 226 + ], + "type": "image", + "image_path": "348e4eb900e98c7dd043b4861f01556e938076c9f209b326e28c9338cfd98983.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 201, + 194, + 233, + 226 + ], + "blocks": [ + { + "bbox": [ + 201, + 194, + 233, + 226 + ], + "lines": [ + { + "bbox": [ + 201, + 194, + 233, + 226 + ], + "spans": [ + { + "bbox": [ + 201, + 194, + 233, + 226 + ], + "type": "image", + "image_path": "2723d5c7d840fabcaf274d643b5982cc10924dc2472408714e849bb07135cb9f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 138, + 241, + 170, + 274 + ], + "blocks": [ + { + "bbox": [ + 108, + 240, + 136, + 270 + ], + "lines": [ + { + "bbox": [ + 108, + 240, + 136, + 270 + ], + "spans": [ + { + "bbox": [ + 108, + 240, + 136, + 270 + ], + "type": "text", + "content": "CutMix:" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 241, + 170, + 274 + ], + "lines": [ + { + "bbox": [ + 138, + 241, + 170, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 241, + 170, + 274 + ], + "type": "image", + "image_path": "f7b7fc13c95b2660876701ecb06fe644dbfe072d4eca8644e7cbaea1c0fa5341.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 277, + 166, + 284 + ], + "lines": [ + { + "bbox": [ + 141, + 277, + 166, + 284 + ], + "spans": [ + { + "bbox": [ + 141, + 277, + 166, + 284 + ], + "type": "text", + "content": "num=2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 172, + 241, + 203, + 274 + ], + "blocks": [ + { + "bbox": [ + 172, + 241, + 203, + 274 + ], + "lines": [ + { + "bbox": [ + 172, + 241, + 203, + 274 + ], + "spans": [ + { + "bbox": [ + 172, + 241, + 203, + 274 + ], + "type": "image", + "image_path": "b7e123435f312fde5952313b8f2b94e9126320004b582185a1c4b1f1583b5ca4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 277, + 198, + 284 + ], + "lines": [ + { + "bbox": [ + 174, + 277, + 198, + 284 + ], + "spans": [ + { + "bbox": [ + 174, + 277, + 198, + 284 + ], + "type": "text", + "content": "num=3" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 204, + 241, + 236, + 274 + ], + "blocks": [ + { + "bbox": [ + 204, + 241, + 236, + 274 + ], + "lines": [ + { + "bbox": [ + 204, + 241, + 236, + 274 + ], + "spans": [ + { + "bbox": [ + 204, + 241, + 236, + 274 + ], + "type": "image", + "image_path": "930df273531792bd39256143b57cd2e9e61edb8c3f4976ee9633c9ef9ed6cd0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 277, + 231, + 284 + ], + "lines": [ + { + "bbox": [ + 208, + 277, + 231, + 284 + ], + "spans": [ + { + "bbox": [ + 208, + 277, + 231, + 284 + ], + "type": "text", + "content": "num=4" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 208, + 241, + 235, + 274 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 241, + 235, + 274 + ], + "spans": [ + { + "bbox": [ + 208, + 241, + 235, + 274 + ], + "type": "image", + "image_path": "cbc8d342cb7fa7ef8d4ad60956ab70e14d0b2e44395a4e03a221dfe44dce60ba.jpg" + } + ] + } + ], + "index": 13, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 137, + 289, + 169, + 323 + ], + "blocks": [ + { + "bbox": [ + 109, + 308, + 135, + 319 + ], + "lines": [ + { + "bbox": [ + 109, + 308, + 135, + 319 + ], + "spans": [ + { + "bbox": [ + 109, + 308, + 135, + 319 + ], + "type": "text", + "content": "MixUp:" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 289, + 169, + 323 + ], + "lines": [ + { + "bbox": [ + 137, + 289, + 169, + 323 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 169, + 323 + ], + "type": "image", + "image_path": "bec7efbcb0d10baa15efbc267008d12044ce53b3b652a5bc36b555cfa6c139ab.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 173, + 289, + 203, + 323 + ], + "blocks": [ + { + "bbox": [ + 173, + 289, + 203, + 323 + ], + "lines": [ + { + "bbox": [ + 173, + 289, + 203, + 323 + ], + "spans": [ + { + "bbox": [ + 173, + 289, + 203, + 323 + ], + "type": "image", + "image_path": "e4ad4171544bfe2ae3054ecbdf34b03e94c735d762f511b609cb594bec66dd6d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 204, + 289, + 235, + 323 + ], + "blocks": [ + { + "bbox": [ + 204, + 289, + 235, + 323 + ], + "lines": [ + { + "bbox": [ + 204, + 289, + 235, + 323 + ], + "spans": [ + { + "bbox": [ + 204, + 289, + 235, + 323 + ], + "type": "image", + "image_path": "07ccb31ab4f6ff562ebf6fefb7ef0dc070c6e3a8e523d7883d494f4752c53e21.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 137, + 337, + 170, + 371 + ], + "blocks": [ + { + "bbox": [ + 98, + 357, + 134, + 365 + ], + "lines": [ + { + "bbox": [ + 98, + 357, + 134, + 365 + ], + "spans": [ + { + "bbox": [ + 98, + 357, + 134, + 365 + ], + "type": "text", + "content": "GridMask:" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 337, + 170, + 371 + ], + "lines": [ + { + "bbox": [ + 137, + 337, + 170, + 371 + ], + "spans": [ + { + "bbox": [ + 137, + 337, + 170, + 371 + ], + "type": "image", + "image_path": "c04e8b8e90a95db5da00a802855ca7a9f5e1c7883eca3c98e1c2b015d46f1006.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 373, + 240, + 381 + ], + "lines": [ + { + "bbox": [ + 138, + 373, + 240, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 240, + 381 + ], + "type": "text", + "content": "mask=50 mask=150 mask=250" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 395, + 286, + 424 + ], + "lines": [ + { + "bbox": [ + 46, + 395, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 286, + 424 + ], + "type": "text", + "content": "Figure 2: Creating Hybrid Species: Three Ways. We explore three different ways for generating hybrid species, varying parameter num, which controls the number of classes included. The mask indicates block size." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 171, + 338, + 204, + 371 + ], + "blocks": [ + { + "bbox": [ + 171, + 338, + 204, + 371 + ], + "lines": [ + { + "bbox": [ + 171, + 338, + 204, + 371 + ], + "spans": [ + { + "bbox": [ + 171, + 338, + 204, + 371 + ], + "type": "image", + "image_path": "9997b68f631ffc2a65aa5bb483f341015fa350646cceb949eab096f9bebcc0f1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 205, + 338, + 239, + 371 + ], + "blocks": [ + { + "bbox": [ + 205, + 338, + 239, + 371 + ], + "lines": [ + { + "bbox": [ + 205, + 338, + 239, + 371 + ], + "spans": [ + { + "bbox": [ + 205, + 338, + 239, + 371 + ], + "type": "image", + "image_path": "db9d53db31742015ccbd9f429ffcd1da82a716cfb10fe9f347044448f2efac82.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 441, + 277, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 277, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 277, + 454 + ], + "type": "text", + "content": "2.3. Mixed Sample Data Augmentations (MSDA)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 460, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 567 + ], + "type": "text", + "content": "Traditional data augmentation techniques assume all samples belong to the same class and do not consider relationships between different samples. In contrast, MSDA combines two samples, using various forms of label representation [35]. Popular methods include Mixup [47], Cutmix [45], Cutout [4], and GridMask [2], each with the core idea of producing images with more representative features of real-world scenarios. Other MSDA techniques include SmoothMix [20] and FMix [9]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 569, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 286, + 700 + ], + "type": "text", + "content": "Despite significant achievements in image classification, the above methods have not been applied to metric learning. [17] discussed the similarity between linear interpolation of features and mixup data augmentation, but due to the peculiarities of mixed labels, it cannot be directly applied to metric learning, which typically requires distinguishing samples as dichotomous, i.e., positive and negative samples. Recently, [38] modified an existing metric learning loss function to accommodate mixup, but the input mix was limited to mixup. Our proposed method can be applied to a variety of mixed sample data augmentations (Section 3.2)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "We modified multiple mixed sample data augmentations" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "as shown in Figure. 2 to ensure that mixed data augmented samples contain features from multiple classes in a balanced manner. Specifically, for CutMix, we uniformly cropped and merged images from top to bottom to ensure that features from multiple classes are regularly distributed. For Mixup, we fused the samples in equal proportions to ensure a balanced representation of features from multiple classes. Lastly, we modified Gridmask by replacing the mask with an image to create a fusion effect." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 191, + 411, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 411, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 411, + 205 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 211, + 391, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 211, + 391, + 222 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 391, + 222 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "content": "As a preliminary for our proposed method, we consider a batch of images denoted by " + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "inline_equation", + "content": "x = x_{1},x_{2},\\ldots ,x_{n}" + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "content": " with corresponding labels " + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "inline_equation", + "content": "y = y_{1},y_{2},\\ldots ,y_{n}" + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "content": ". Deep metric learning involves the use of a Convolutional Neural Network (CNN) to learn a non-linear transformation of each image into an m-dimensional deep feature space " + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\phi (x;\\theta_{\\phi}):\\mathcal{X}\\to \\mathbb{R}^{m}" + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\theta_{\\phi}" + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "content": " represents the network's parameters. In order to further learn a mapping from the m-dimensional feature space to a k-dimensional space, we use a linear mapping layer represented by " + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "inline_equation", + "content": "f(\\phi ;\\theta_f):\\mathbb{R}^m\\to \\mathbb{R}^k" + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 304, + 229, + 545, + 361 + ], + "type": "text", + "content": " represents the parameters of the linear mapping layer." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 361, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 384 + ], + "type": "text", + "content": "Formally, we define the distance between two data points in the embedding space as:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 359, + 395, + 545, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 395, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 359, + 395, + 545, + 409 + ], + "type": "interline_equation", + "content": "d _ {f} \\left(x _ {i}, x _ {j}\\right) = \\left\\| f \\left(\\phi_ {i}\\right) - f \\left(\\phi_ {j}\\right) \\right\\| _ {2} \\tag {1}", + "image_path": "1f9418b9689fb894feeecc414f96a250b6bfe1191e2be7f56090ea8fdbe05335.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 418, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 441 + ], + "type": "text", + "content": "Metric learning models usually involve pairwise sampling of a sample, e.g. considering the contrastive loss:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 330, + 451, + 545, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 451, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 330, + 451, + 545, + 465 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o n t}} = y d _ {f} (a, x) ^ {2} + (1 - y) \\left[ m - d _ {f} (a, x) \\right] _ {+} ^ {2} \\tag {2}", + "image_path": "29caa4c572cbee2e25f543a7bee4775facdadfaa2491b3e68fc90b3a5cc8ed86.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "content": " denotes the label, " + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "content": " is the anchor image and " + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "content": " is the marginal value. To train a metric learning model, we first select an anchor image and choose a positive sample with the same label (" + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "y = 1" + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "content": ") and a negative sample with a different label (" + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "content": "). The network is then trained to minimize a loss function that penalizes negative samples that are too close together and positive samples that are too far apart. It is noteworthy that the loss value will be relatively large when the distance between the anchor point and the negative sample is small, while it will be relatively small when the distance is very far. Thus, selecting hard samples during the selection of pairs will result in a larger gradient." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "Studies have tried to increase the number of hard samples for metric learning by generating more samples. However, linear interpolation methods can produce false positives and negatives. It is important to note that the binary metric learning label " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " cannot be used as a one-hot label for image classification tasks." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Our Hybrid Species Embedding (HSE) method addresses the challenges of creating additional hard samples" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11049" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "and applying them to metric learning without label information to improve generalization. We address the challenge of mixed sample data augmentations by emulating human categorization behavior to assign these samples to a reasonable position in the embedding space. We propose incorporating newly generated samples as an extra training signal for the model, using the same feature extraction network for both the original image and the hybrid species generated using MSDA, and included in each batch. The loss function " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " clusters the original samples based on their labels, while our proposed loss function " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "L_{Hy}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " deals with the embedding positions of the new samples, as shown in Figure. 3." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 224, + 287, + 345 + ], + "blocks": [ + { + "bbox": [ + 48, + 224, + 287, + 345 + ], + "lines": [ + { + "bbox": [ + 48, + 224, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 287, + 345 + ], + "type": "image", + "image_path": "3f3f8834389bb4c29fdf2b5df2134f3c7f9dc9d82cb9f1679c4af67014f63ab7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "text", + "content": "Figure 3: Our HSE method uses batches of original images and hybrid species generated by Mixed Sample Data Augmentation (MSDA). The feature extraction network (CNN) processes both types of images to obtain their corresponding feature vectors (" + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "inline_equation", + "content": "f_{Or}" + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "inline_equation", + "content": "f_{HS}" + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "text", + "content": "). The loss function includes two components: " + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "text", + "content": " (metric learning loss) groups original samples into clusters based on their labels, and " + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "inline_equation", + "content": "L_{Hy}" + }, + { + "bbox": [ + 46, + 351, + 287, + 419 + ], + "type": "text", + "content": " optimizes feature extraction by comparing hybrid species embedding positions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 437, + 130, + 490 + ], + "blocks": [ + { + "bbox": [ + 58, + 437, + 130, + 490 + ], + "lines": [ + { + "bbox": [ + 58, + 437, + 130, + 490 + ], + "spans": [ + { + "bbox": [ + 58, + 437, + 130, + 490 + ], + "type": "image", + "image_path": "0f7470e36f9792bcd72c73ce86cfeeb6211142287bd41f19ef52ea53f70f3047.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 491, + 119, + 501 + ], + "lines": [ + { + "bbox": [ + 70, + 491, + 119, + 501 + ], + "spans": [ + { + "bbox": [ + 70, + 491, + 119, + 501 + ], + "type": "text", + "content": "(a) HS-Anchors" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 132, + 437, + 204, + 490 + ], + "blocks": [ + { + "bbox": [ + 132, + 437, + 204, + 490 + ], + "lines": [ + { + "bbox": [ + 132, + 437, + 204, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 204, + 490 + ], + "type": "image", + "image_path": "d76e21d9d9d1f69e835cbfeba513c23fa888e87e24752703d95c417cba06e81b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 491, + 194, + 500 + ], + "lines": [ + { + "bbox": [ + 141, + 491, + 194, + 500 + ], + "spans": [ + { + "bbox": [ + 141, + 491, + 194, + 500 + ], + "type": "text", + "content": "(b) WP-Anchor" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 515, + 287, + 582 + ], + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 582 + ], + "type": "text", + "content": "Figure 4: Three types of anchor point selection: The dashed lines indicate that the sample pairs should move away from each other, while the solid lines indicate that the sample pairs should move closer to each other. The red sample is a hybrid species synthesized from green and blue. (a) denotes the use of the hybrid species as the anchor point, (b) denotes using the class that synthesizes the hybrid species as an anchor point, and (c) denotes using a class unrelated to the hybrid species as an anchor point." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 204, + 437, + 278, + 490 + ], + "blocks": [ + { + "bbox": [ + 204, + 437, + 278, + 490 + ], + "lines": [ + { + "bbox": [ + 204, + 437, + 278, + 490 + ], + "spans": [ + { + "bbox": [ + 204, + 437, + 278, + 490 + ], + "type": "image", + "image_path": "690c4ff3dafa28070863a9518a8803f84530d6c74df6678b613435cd33493e99.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 491, + 263, + 500 + ], + "lines": [ + { + "bbox": [ + 218, + 491, + 263, + 500 + ], + "spans": [ + { + "bbox": [ + 218, + 491, + 263, + 500 + ], + "type": "text", + "content": "(c) N-Anchor" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "type": "text", + "content": "3.2. Hybrid Species" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "In the context of image classification tasks, hard negative samples refer to images that are similar to the anchor images but have different labels. We posit that hard negative samples typically possess similar features to the anchor samples. Therefore, when a sample is derived from an anchor image, there is a high likelihood that the sample will be classified as a hard sample. Let " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{B}" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": " denote two images with labels " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "y_{A}" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "y_{B}" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "y_{A} \\neq y_{B}" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "The generated sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": " contains both the features of " + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": " and the features of " + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "x_{B}" + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": ". We refer to " + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": " as the hybrid species." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": " has a high likelihood of becoming a hard negative sample, while simultaneously presenting potential gradients in classes " + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": ". This can be achieved by utilizing mixed sample data augmentation techniques, such as mixup. Mixup generates a new training sample, denoted as " + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "(x_{H},y_{H})" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": ", by linearly interpolating two training samples " + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "(x_{A},y_{A})" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "(x_{B},y_{B})" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "inline_equation", + "content": "y_{H}" + }, + { + "bbox": [ + 304, + 96, + 545, + 180 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 190, + 545, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 190, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 373, + 190, + 545, + 203 + ], + "type": "interline_equation", + "content": "x _ {H} = \\lambda x _ {A} + (1 - \\lambda) x _ {B} \\tag {3}", + "image_path": "0b2c3553770a8494ff50bff2cf0b05ce3dc1d0d7ed60cb2045446e5eb495fa72.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 375, + 224, + 545, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 224, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 375, + 224, + 545, + 237 + ], + "type": "interline_equation", + "content": "y _ {H} = \\lambda y _ {A} + (1 - \\lambda) y _ {B} \\tag {4}", + "image_path": "46792adab052fdad6ced132e202135a6250a709d3d15902a3b309f16b36b4b12.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": " is obtained from a beta distribution and is constrained to [0, 1]. To achieve a distribution of " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": " values that is centered around 0.5, " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": " should follow a beta distribution with shape parameters " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\lambda \\sim \\mathrm{Beta}(\\alpha, \\alpha)" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0, \\infty)" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": ". The generated training sample " + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "inline_equation", + "content": "(x_H, y_H)" + }, + { + "bbox": [ + 304, + 242, + 545, + 314 + ], + "type": "text", + "content": " is then used for model training with a loss function." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "text", + "content": "Many pair-based metric learning losses rely on having distinct class labels, but the output " + }, + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "inline_equation", + "content": "y_{H}" + }, + { + "bbox": [ + 304, + 314, + 545, + 386 + ], + "type": "text", + "content": " is often a probability distribution per class, which is unsuitable for metric learning applications. Determining how to utilize these labels for metric learning presents a significant challenge. This difficulty is illustrated in Eq. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 395, + 545, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 395, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 395, + 545, + 409 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o n t}} = y d _ {f} \\left(a, x _ {H}\\right) ^ {2} + (1 - y) \\left[ m - d _ {f} \\left(a, x _ {H}\\right) \\right] _ {+} ^ {2} \\tag {5}", + "image_path": "3af2bb30502a87d1a5ff2e0e20f0e655a1cd456bb661757a6cd82a88fae1c1a5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "content": "As the binary label " + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}" + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "content": ", the label for " + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "content": " is not straightforward, since it is a hybrid sample. When we consider " + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "content": " as a negative sample of the anchor point, it is not reasonable, as " + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "content": " contains the same features as the anchor samples. Keeping them away from each other could cause the model to ignore important features and focus on irrelevant ones. Therefore, the binary label " + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "content": " is invalid for " + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 418, + 545, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 502, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 550 + ], + "type": "text", + "content": "[38] showed that classification tasks are not fundamentally different from metric learning, and the label of the synthesized sample " + }, + { + "bbox": [ + 304, + 502, + 545, + 550 + ], + "type": "inline_equation", + "content": "X_{H}" + }, + { + "bbox": [ + 304, + 502, + 545, + 550 + ], + "type": "text", + "content": " can still be represented by linear interpolation of the labels. Eq. 5 can then be converted to:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 570, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\widetilde {L} _ {\\text {c o n t}} = \\lambda d _ {f} (a, x _ {H}) ^ {2} + (1 - \\lambda) [ m - d _ {f} (a, x _ {H}) ] _ {+} ^ {2} \\tag {6}", + "image_path": "dfbf89d3ad468d62d1b1938615ac91be69b1711578c3731acb5ad2cf29a08699.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "content": "Now, " + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "content": " appears with nonzero contributions in both positive and negative terms for positive-negative pairs, due to the interpolation factor " + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "However, we found that these methods are only applicable to mixup and may be subject to controversy in other forms of mixed sample data augmentations. This is due to the fact that mixup binds the images together evenly based on the assigned weight parameter, " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". The same applies to labels In contrast, cutmix synthesizes " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " by replacing a crop area of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " with that of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{B}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", where the crop area of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11050" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 255, + 110 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 255, + 110 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 255, + 110 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 255, + 110 + ], + "type": "table", + "html": "
HS-AnorhWP-AnorhN-Anorh
Intra-class variation✓✓✕
Inter-class variation✓✓✓
Conflict✕✓✕
", + "image_path": "7506c00de0f164608a0b354c85d165b7cac011f3a9bf0a53d902c7fa19fc65e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 128, + 287, + 148 + ], + "lines": [ + { + "bbox": [ + 47, + 128, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 287, + 148 + ], + "type": "text", + "content": "Table 1: Three types of anchor point selection are considered for inter-class and intra-class variation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "inline_equation", + "content": "(1 - \\lambda)" + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "text", + "content": " is taken from " + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "inline_equation", + "content": "x_{A}" + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "inline_equation", + "content": "x_{B}" + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "text", + "content": ", respectively. Therefore, the label of " + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "text", + "content": " cannot be accurately represented by linear interpolation, as the area being replaced is indeterminate, and may contain either significant features or irrelevant content. Consequently, the contribution of the crop area to " + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 171, + 287, + 254 + ], + "type": "text", + "content": " remains inconclusive, making the application of linear interpolation for labeling mostly infeasible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "text", + "content": "Instead of fully utilizing the label information of " + }, + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "text", + "content": ", we can translate them into their appropriate spatial embedding, as we will describe in next subsection. Also, when applying MSDA to images, we need to consider how the labels are to be processed, which limits us to intermixing only two samples. In contrast, " + }, + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "text", + "content": " does not require us to define label information, and can hence be mixed by multiple classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 349, + 124, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 124, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 124, + 361 + ], + "type": "text", + "content": "3.3. Hybrid loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 245, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 245, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 245, + 380 + ], + "type": "text", + "content": "3.3.1 Anchor Point Selection and Processing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": "Although it is challenging to accurately assign a label to " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " obtained through MSDA, the features of " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " can be inferred based on the source images. We define " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "C(x_{I})" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " to represent the label of image " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{I}" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " is selected as an anchor, it may not be possible to find a suitable positive sample, but " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " shares semantic information with " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "C(x_{A})" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "C(x_{B})" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " to some extent. Thus, we can consider " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "C(x_{A})" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "C(x_{B})" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": " as weak positive samples (WP) of " + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 389, + 287, + 496 + ], + "type": "text", + "content": ", with the remaining classes treated as negative samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": "During the training phase, the embedding of the spatial position of " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " plays a crucial role in the generalization of downstream tasks. Therefore, deciding how to treat " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " in pair selection becomes the first key question. Since " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " contains information for multiple classes, including it in a paired selection can provide more gradient information. We divided the pairing methods into three categories, as shown in Figure. 4: HS (hybrid species sample) as the anchor point, WP (weak positive sample) as the anchor point, and N (negative sample) as the anchor point. We compared the three methods and and summarized the findings in Table 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "The choice of anchor point in a classification task can affect the learning of intra-class and inter-class variation. When using " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " as an anchor point, it pulls towards the WP while staying away from the rest of the class. Using WP as an anchor point keeps it close to its composite " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " to maintain identical points within the class, but may result in conflicts with the synthetic sample of " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " leading" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "to non-converging losses. Anchoring with N treats " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " as a negative sample and moves it away from other samples, losing its purpose. Our findings recommend using HS as the anchor sample to effectively focus on both intra-class and inter-class variation while ensuring convergence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 147, + 464, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 147, + 464, + 159 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 464, + 159 + ], + "type": "text", + "content": "3.3.2 Strategies for Sample Mining" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "text", + "content": "In the previous subsection, we analyzed the contribution of pairwise selection to the embedding of " + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "text", + "content": " is randomly generated in each epoch, how to embed it more efficiently and reasonably becomes the second key issue of our study. For instance, when " + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "text", + "content": " is used as an anchor point, the corresponding WP may contain multiple classes. If " + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "text", + "content": " pulls all WP close together, clusters that are unrelated may merge, which is unreasonable. If " + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 165, + 545, + 320 + ], + "type": "text", + "content": " randomly selects WP that are far apart, the structure of the cluster may be destroyed. On the other hand, if negative samples that are far apart are selected, the resulting gradient will be smaller. In summary, it is vital that we choose a reasonable and valid sample for performing metric learning with HS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 321, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 369 + ], + "type": "text", + "content": "Several studies have explored the advantages of hard negative mining in constructing informative gradients [12, 36, 34]. In this context, we introduce the notion of hard negatives for the hybrid species as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 377, + 545, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 377, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 312, + 377, + 545, + 409 + ], + "type": "interline_equation", + "content": "x _ {H _ {h n}} = \\underset {x: C (x _ {I}) \\neq C (x _ {A}) \\cap C (x _ {I}) \\neq C (x _ {B})} {\\operatorname {a r g m i n}} d \\left(f \\left(x _ {H}\\right), f \\left(x _ {I}\\right)\\right) \\tag {7}", + "image_path": "cdafab4e61372156f770ef3c247d9d4756f7586fff092d982babe6a26ce1e39c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "text", + "content": "It is worth noting that " + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "x_{H_{hn}}" + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "text", + "content": " does not only yield a large gradient for " + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "text", + "content": ", but it also possesses the same potential gradient as the inputs " + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "C(x_{A})" + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "C(x_{B})" + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "text", + "content": " that are used to synthesize " + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 410, + 545, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 457, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 517 + ], + "type": "text", + "content": "The study by [43] shows that easy positive mining improves the generalization performance of downstream tasks. Our experiments also demonstrate the benefits of this approach for " + }, + { + "bbox": [ + 304, + 457, + 545, + 517 + ], + "type": "inline_equation", + "content": "x_{h}" + }, + { + "bbox": [ + 304, + 457, + 545, + 517 + ], + "type": "text", + "content": ". We define the easy hybrid weak positive of the hybrid species as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 526, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 526, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 310, + 526, + 545, + 557 + ], + "type": "interline_equation", + "content": "x _ {H _ {e w p}} = \\underset {x: C \\left(x _ {I}\\right) = C \\left(x _ {A}\\right) \\cup C \\left(x _ {I}\\right) = C \\left(x _ {B}\\right)} {\\operatorname {a r g m i n}} d \\left(f \\left(x _ {H}\\right), f \\left(x _ {I}\\right)\\right) \\tag {8}", + "image_path": "4d87d4dffcc4ae34c95eb61609fa4f18c56bbc7bfb4b8f71e8bb9dad12c65df7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": "There are two primary reasons why bringing " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "x_{H_{ewp}}" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": " closer together is beneficial. Firstly, it enables the model to concentrate on the features that are shared by both inputs, which can significantly influence the sample measurements. Secondly, this approach avoids the potential issue of pushing weak positives that are distant from each other too close together. By doing so, it preserves both the intra-class variance and the inter-class variance, which can help maintain the manifold structure of the classes in the embedding space." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Our hybrid loss function involves mapping the output of the convolutional neural network onto a unit sphere, which is a widely accepted technique. During each mini-batch," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11051" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 49, + 76, + 272, + 487 + ], + "blocks": [ + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "lines": [ + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "text", + "content": "1 # Nsize: Number of samples of the same class in each batch \n2 # n: The number of classes that synthesize " + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "text", + "content": " \n3 # b: batch size d: dimension \n4 # F = b × d \n5 # Map of feature similarity between samples \n6 Map = F × FT; \n7 # " + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "text", + "content": " as the anchor, look for weak positive sample \n8 DWP = Map; \n9 for each i ∈ [b - Nsize, b] do \n10 for each j ∈ [n * Nsize, b] do \n11 | DWP[i][j] = -1; \n12 end \n13 end \n14 # Each " + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "text", + "content": " corresponds to the value and index of the most similar WP sample \n15 VDWP, IDWP = DWP.max(1); \n16 # " + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "text", + "content": " as the anchor, look for negative sample \n17 DN = Map; \n18 for each i ∈ [b - Nsize, b] do \n19 for each j ∈ [0, n * Nsize] do \n20 | DN[i][j] = -1; \n21 end \n22 for each j ∈ [b - Nsize, b] do \n23 | DN[i][j] = -1; \n24 end \n25 end \n26 # Each " + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "text", + "content": " corresponds to the value and index of the most similar negative sample \n27 VDN, IDN = DN.max(1); \n28 # Select the WP sample and negative sample parts corresponding to " + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 49, + 76, + 272, + 487 + ], + "type": "text", + "content": " \n29 T = [VDWP[-Nsize:], VDN[-Nsize:]]; \n30 Hybrid loss = -α logsoftmax(T).mean();" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "bbox": [ + 99, + 490, + 209, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 490, + 209, + 502 + ], + "spans": [ + { + "bbox": [ + 99, + 490, + 209, + 502 + ], + "type": "text", + "content": "Algorithm 1: Hybrid Loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": "when " + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": " serves as the anchor, we select " + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_{H_{ewp}}" + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_{H_{hn}}" + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": " for comparison. As in Algorithm 1, the former corresponds to the original sample, while the latter represents " + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": ". To synthesize " + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": ", we typically select samples from the first " + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": " classes. Then, we search for the most similar weak positive sample and negative sample for " + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 523, + 287, + 595 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 595, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 595, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 287, + 654 + ], + "type": "text", + "content": "The NCA loss function is an effective way to avoid the selection of margin hyperparameters and handle the distance between samples efficiently[46]. To ensure that hybrid species are embedded in a suitable location, we define the hybrid loss based on the NCA loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 662, + 287, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 662, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 81, + 662, + 287, + 699 + ], + "type": "interline_equation", + "content": "L _ {H y} = - \\alpha \\log \\frac {1}{1 + \\frac {\\exp (f (x _ {H}) ^ {\\top} f (x _ {H _ {h n}}))}{\\exp (f (x _ {H}) ^ {\\top} f (x _ {H _ {e w p}}))}} \\tag {9}", + "image_path": "888a132a6135ba8544f8a1b64099bd6efe408810bba8782381dfe0afc4edfc82.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " is a hyper-parameter that controls the balance be" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "tween bringing " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{H_{ewp}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " closer and pulling " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x_{H_{hn}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " farther apart. The " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "L_{Hy}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " loss serves as an auxiliary loss to process additional samples and can be combined with other metric learning losses. As in Figure 3, the final loss is the sum of the metric learning loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "L_{Hy}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 392, + 145, + 545, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 145, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 392, + 145, + 545, + 158 + ], + "type": "interline_equation", + "content": "L = L _ {m} + L _ {H y} \\tag {10}", + "image_path": "638d64ed492a4d8ec442c370992f86bd686d6a0fef41a1f98d0ea8ca1eceb28d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 164, + 545, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 164, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 545, + 188 + ], + "type": "text", + "content": "When incorporating our method with the contrastive loss (Eq. 2), the resulting final loss is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 388, + 198, + 545, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 198, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 388, + 198, + 545, + 212 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {c o n t}} + L _ {H y} \\tag {11}", + "image_path": "18675e71a5dbf5c965729755e93aa992cd617417de194827d1716da3a6167543.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 222, + 387, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 387, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 387, + 236 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "type": "text", + "content": "Our HSE method uses pre-trained BN-Inception [14] and ResNet [11] models on ILSVRC 2012-CLS [32] with PyTorch. These models have learned feature representations from diverse images, reducing time and resources for new dataset training. They also perform well in image recognition tasks, making them suitable for computer vision applications. For generalization ability evaluation, we maintain consistency with base methods on image preprocessing, learning rate, and batch size. (Source code: https://github.com/SHQberserker/HSE)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 363, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 459 + ], + "type": "text", + "content": "Nsize examples from each class are selected per batch. We use two methods to incorporate hybrid species during training: selecting multiple classes to add hybrid species as additional examples of batch, and randomly replacing examples in the batch with hybrid species. We use the cutmix technique to create hybrid species by stitching two samples together evenly. Algorithm 1 fixes the last Nsize samples of each batch as synthetic samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 467, + 427, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 427, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 427, + 479 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 485, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 629 + ], + "type": "text", + "content": "We tested our framework's ability to perform well on challenging datasets: CUB-200, CARS-196, and SOP. These datasets involve fine-grained classification tasks, where there are many categories and relatively few images per category. We followed the data split approach from [22]. Specifically, for CUB-200, we used 5864 images from the first 100 categories for training and 5924 images from the last 100 categories for testing. For CARS-196, we trained on the first 98 classes and tested on the remaining images. Additionally, for SOP, we utilized the first 59,551 images (11,318 classes) for training and the remaining 60,502 images (11,316 classes) for testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "To assess our experiments, we employed the Recall@K metrics. This measure involves selecting the K images that are closest to the query image. If there's an image with the same label as the query image, we score 1; otherwise, it's 0. Recall@K calculates the average score for all query images, with higher values indicating better retrieval performance. These metrics provide a way to quantitatively evaluate how" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11052" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "well our model can find similar images, which is crucial for many image-based applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 167, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 167, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 167, + 114 + ], + "type": "text", + "content": "4.2. Results and Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 121, + 133, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 133, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 133, + 132 + ], + "type": "text", + "content": "4.2.1 Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 140, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 140, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 287, + 330 + ], + "type": "text", + "content": "Experiments compared our proposed method to existing methods, including N-pair, EPSHN, MSloss and ProxyAnchor. Results are shown in Table 2 for CUB-200 and CARS-196 datasets, where * indicates our results under the same experimental settings. Our method improves on different approaches to varying degrees. Our method had less impact on the N-pair method as it mostly ignores intraclass variation, leading to excessive clustering between the same classes. However, we observed a significant improvement in recall@K when the EPSHN method was augmented with the additional training signal of HSE. Recall@1 improved by " + }, + { + "bbox": [ + 46, + 140, + 287, + 330 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 140, + 287, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 140, + 287, + 330 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 140, + 287, + 330 + ], + "type": "text", + "content": " on the challenging CUB-200 and CARS-196 datasets, respectively, where Nsize=16 and Nsize=8, respectively. Importantly, HSE-E outperformed the ensemble-based method that EPSHN had not surpassed before on the CUB-200 and CARS-196 datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 461 + ], + "type": "text", + "content": "We believe that EPSHN tends to ignore intra-class variations, such as those found in [43] where birds of different sexes may vary greatly within the same class, but should also focus on characteristics common between members of the same class. When " + }, + { + "bbox": [ + 46, + 331, + 287, + 461 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 331, + 287, + 461 + ], + "type": "text", + "content": " is added, it helps the network to focus more on these common characteristics. MSloss focuses on both inter-class variation and inter-class variation. However, when adding our additional training signals, there is still a certain improvement. This is because the features that our additional training signals focus on help the network generalize to unknown classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "content": "Our method achieves good results in pair-based methods and is effective in proxy-based metric learning methods. We improve approximately " + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "content": " in both " + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{Recall}@1" + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{Recall}@2" + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "content": ". This confirms our universality, which can be applied to various metric learning loss functions. We also compared our method to the Matrix/Input [38] approach. Matrix/Input has a high computational cost because it calculates " + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} n(n - 1)" + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "content": " embeddings for a batch of " + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "content": " examples. Instead, our method achieves a similar effect using only " + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "inline_equation", + "content": "Nsize * Cnum" + }, + { + "bbox": [ + 46, + 463, + 287, + 628 + ], + "type": "text", + "content": " embeddings. Our approach is hence more computationally efficient. We also remain applicable to discriminate between positives and negatives for every anchor point. We can also incorporate non-additive component functions that involve loss functions, e.g., EPSHN." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "text", + "content": "HSE is also effective on the SOP dataset. Unlike smaller datasets, the SOP dataset has superclasses, which introduce more variation between different superclasses and less variation within the same superclass. It's challenging to measure similarity between samples within the same superclass in the SOP dataset. As shown in Table 3, we generate HS from samples that share the same superclass but have differ" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": "ent labels. When we apply HSE to various classical methods, we observe varying degrees of improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 112, + 479, + 124 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 112, + 479, + 124 + ], + "spans": [ + { + "bbox": [ + 306, + 112, + 479, + 124 + ], + "type": "text", + "content": "4.2.2 Impact of Minibatch Parameters" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 131, + 375, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 131, + 375, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 131, + 375, + 143 + ], + "type": "text", + "content": "Impact of Nsize" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 149, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 149, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 545, + 269 + ], + "type": "text", + "content": "We trained a ResNet-18 embedding network on CUB-200 and CARS-196 datasets with batch size 128 and embedding dimension 512 to investigate HSE's impact on model generalization at various Nsizes. HSE's effectiveness varies with methods, datasets, and Nsizes (Figure. 5). The best performance occurs at Nsize between 4 and 16 with a batch size of 128, followed by performance degradation. When there are too few negative samples, the hybrid species' gradient information becomes insignificant, and they may be grouped with WP's clusters, harming downstream tasks." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 279, + 405, + 346 + ], + "blocks": [ + { + "bbox": [ + 321, + 279, + 405, + 346 + ], + "lines": [ + { + "bbox": [ + 321, + 279, + 405, + 346 + ], + "spans": [ + { + "bbox": [ + 321, + 279, + 405, + 346 + ], + "type": "image", + "image_path": "4ab7a647d9ff83c4f2c2886e6e56a38787378f6d57445c7bdba40a062c6741cf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 347, + 380, + 355 + ], + "lines": [ + { + "bbox": [ + 355, + 347, + 380, + 355 + ], + "spans": [ + { + "bbox": [ + 355, + 347, + 380, + 355 + ], + "type": "text", + "content": "(a) CUB" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 279, + 518, + 346 + ], + "blocks": [ + { + "bbox": [ + 436, + 279, + 518, + 346 + ], + "lines": [ + { + "bbox": [ + 436, + 279, + 518, + 346 + ], + "spans": [ + { + "bbox": [ + 436, + 279, + 518, + 346 + ], + "type": "image", + "image_path": "7c1703200f28fff10e20d066a809a03603f67910656e2ddfb97b9ff9066a063c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 347, + 493, + 355 + ], + "lines": [ + { + "bbox": [ + 468, + 347, + 493, + 355 + ], + "spans": [ + { + "bbox": [ + 468, + 347, + 493, + 355 + ], + "type": "text", + "content": "(b) CAR" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 376, + 392, + 470, + 463 + ], + "blocks": [ + { + "bbox": [ + 306, + 369, + 544, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 369, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 544, + 380 + ], + "type": "text", + "content": "Figure 5: Effect of additional training signals provided by different Nsizes." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 392, + 470, + 463 + ], + "lines": [ + { + "bbox": [ + 376, + 392, + 470, + 463 + ], + "spans": [ + { + "bbox": [ + 376, + 392, + 470, + 463 + ], + "type": "image", + "image_path": "6d3bb9bfec150f1eb655e584af25b56eadd459f0c7714e5cc28d71cf2e6d5480.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 472, + 541, + 483 + ], + "lines": [ + { + "bbox": [ + 310, + 472, + 541, + 483 + ], + "spans": [ + { + "bbox": [ + 310, + 472, + 541, + 483 + ], + "type": "text", + "content": "Figure 6: Results on the CUB dataset using different Cnum in MS loss." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 492, + 379, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 492, + 379, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 379, + 503 + ], + "type": "text", + "content": "Impact of Cnum" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": "We found that HSE improves the generalization ability of the model. However, increasing the number of hybrid species does not necessarily lead to better results. We conducted experiments on the CUB dataset using the MS loss method, as shown in Figure. 6. Our loss function indirectly affects the entire embedding space of the batch when learning the spatial embedding of a single hybrid species. After training for a certain period, the two combined classes have essentially covered the entire dataset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 634, + 443, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 443, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 443, + 647 + ], + "type": "text", + "content": "4.2.3 Effect of Hybrid Species" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "We investigated the effectiveness of different data augmentation techniques in providing additional training signals for hybrid species. Table 4 shows the classic blending data augmentation techniques we used, with certain modifications made to each method. Notably, CutMix and GridMask" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11053" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 70, + 476, + 330 + ], + "blocks": [ + { + "bbox": [ + 115, + 70, + 476, + 330 + ], + "lines": [ + { + "bbox": [ + 115, + 70, + 476, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 70, + 476, + 330 + ], + "type": "table", + "html": "
CUB-200CARS-196
Recall@kk=1k=2k=4k=1k=2k=4
Angular512[39]G53.665.075.371.380.787.0
HDML512[49]G53.765.776.779.187.192.1
HTL512[6]BN57.168.878.781.488.092.7
Margin128[41]R63.674.483.179.686.591.9
SoftTriple512[28]BN65.476.484.584.590.794.5
DR512[24]BN66.177.085.185.090.594.1
Ensemble-based
HDC384[44]G53.665.777.073.282.486.4
A-BIER512[27]G55.367.276.978.085.891.1
ABE-8512[16]G60.671.579.860.671.579.8
DREML9216[42]R63.975.083.186.091.795.0
Npair64*(Nsize=8)[37]R51.364.275.364.276.084.6
HSE-N64(Nsize=8)R52.5+1.266.0+1.877.4+2.164.9+0.776.6+0.684.7+0.1
EPSHN512*(Nsize=16)[43]R64.975.383.583.189.793.6
HSE-E512(Nsize=16)R66.9+2.077.4+2.185.5+2.085.4+2.391.2+1.596.9+3.3
HSE-E512(Nsize=8)R67.4+2.577.7+2.485.7+2.284.8+1.790.6+0.994.3+0.7
MS†512R67.877.885.687.892.795.3
+Matrix输入[38]R69.0+1.279.1+1.386.0+0.489.0+1.293.4+0.796.0+0.7
MS512*(Nsize=5)[40]BN65.777.086.381.788.793.2
HSE-M512(Nsize=5)BN67.6+1.978.0+1.085.8-0.582.0+0.388.9+0.293.3+0.1
Proxy-Ancho†512R69.579.387.087.692.395.5
+Matrix输入[38]R70.5+1.081.2+1.987.8+0.888.2+0.693.2+0.996.2+0.7
Proxy-Ancho*512[15]R69.479.287.088.592.795.6
HSE-PA512R70.6+1.280.1+0.987.1+0.189.6+1.193.8+1.196.0+0.4
", + "image_path": "c7ca4b947d2d47d9afd1036c2483400f3f53871dbee2bf3ef14aee4d5f18a964.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 396, + 285, + 521 + ], + "blocks": [ + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "type": "text", + "content": "Table 2: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the CUB-200-2011 and Cars-196 datasets. The backbone networks used in the models are denoted by abbreviations: G for GoogleNet, BN for Inception with batch normalization, and R for ResNet50. The symbol " + }, + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "type": "text", + "content": " indicates the reproduced result reported by the original authors.N-pair and EPSHN use cnum=1 and batch size of 128. MSloss uses cnum=1, batch size of 80, and " + }, + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "type": "text", + "content": ". Proxy-Anor uses cnum=1, batch size of 128, and " + }, + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\alpha = 2" + }, + { + "bbox": [ + 46, + 338, + 545, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 396, + 285, + 521 + ], + "lines": [ + { + "bbox": [ + 48, + 396, + 285, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 285, + 521 + ], + "type": "table", + "html": "
SOP
Recall@kk=1k=10k=100
EPSHN512*(Nsize=5) [43]R74.386.994.3
HSE-E512(Nsize=5)R76.3+2.088.4+1.594.9+0.6
MS†512R76.989.895.9
+Matrix输入[38]R77.9+1.090.6+0.895.9
MS512*[40] (Nsize=5)R78.290.296.2
HSE-M512R78.7+0.590.4+0.296.1-0.1
Proxy-Anchort512R79.190.896.2
+Matrix输入[38]R79.8+0.791.4+0.696.5+0.3
Proxy-Anchort*512[15]R79.290.495.8
HSE-PA512(Nsize=5)R80.0+0.891.4+1.096.3+0.5
", + "image_path": "5917b126e06f59ed4b7d47ac4823ae61f935cd67d1bbd3c5783435b60424628f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 581, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 700 + ], + "type": "text", + "content": "stand out. CutMix can incorporate features from each compositional image to a certain extent, but may also diminish key features. Mixup dilutes key features as the number of samples increases. GridMask (block=50) covers the entire image despite using relatively small blocks. Results show that CutMix is the most effective approach for improving the performance of Hybrid Species (HS). We also found that adding more class features does not necessarily improve HS performance, but adding critical features that are specific to a particular class can lead to better HS results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "We investigate the selection of weak positive samples" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 396, + 531, + 531 + ], + "blocks": [ + { + "bbox": [ + 46, + 529, + 287, + 559 + ], + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 559 + ], + "type": "text", + "content": "Table 3: The comparison of the Recall@K (%) performance of our proposed method HSE with several baseline methods on the SOP.Where batch " + }, + { + "bbox": [ + 46, + 529, + 287, + 559 + ], + "type": "inline_equation", + "content": "= {120}" + }, + { + "bbox": [ + 46, + 529, + 287, + 559 + ], + "type": "text", + "content": " ,cnum " + }, + { + "bbox": [ + 46, + 529, + 287, + 559 + ], + "type": "inline_equation", + "content": "= {12}" + }, + { + "bbox": [ + 46, + 529, + 287, + 559 + ], + "type": "text", + "content": " ,and hybrid species as additional examples of batch." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 396, + 531, + 531 + ], + "lines": [ + { + "bbox": [ + 318, + 396, + 531, + 531 + ], + "spans": [ + { + "bbox": [ + 318, + 396, + 531, + 531 + ], + "type": "table", + "html": "
CUBCAR
R@1numR@1num
Cutmix67.4285.42
67.2385.03
67.1484.24
Mixup67.0285.32
66.9381.03
R@1MasksizeR@1Masksize
Gridmask67.45084.950
66.015084.8150
67.125084.0250
", + "image_path": "8f97519d9105ad5bd32a18a82c412875a7a15b8a6576634b5697676b7b0332d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 548, + 545, + 569 + ], + "lines": [ + { + "bbox": [ + 306, + 548, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 545, + 569 + ], + "type": "text", + "content": "Table 4: Results on the Cub dataset using different Mixed Sample Data Augmentations in EPSHN. " + }, + { + "bbox": [ + 306, + 548, + 545, + 569 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathbb{1} = \\mathbb{Recall}@\\mathbb{1})" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "by hybrid species using the Proxy-Anorh method on the CUB dataset. Cutmix augmentation technique is used to stitch images either vertically or horizontally, with the upper part mainly containing the bird's head and the lower part containing the bird's body when stitched vertically. When stitched horizontally, the probability of the left and right sides containing the bird's head is almost the same. Our HSE approach improves model performance by focusing on essential image features. Our study on the CUB dataset demonstrates that the hybrid species prioritize the bird's" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11054" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 254, + 209 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 254, + 209 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 254, + 209 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 254, + 209 + ], + "type": "image", + "image_path": "2816ec965166d5a44ddac96594c047db7957582107bec461f7f4fc1b48d98608.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 289, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 294 + ], + "type": "text", + "content": "Figure 7: Analysis of easy weak positive samples selected by hybrid species using Proxy-Antor on CUB dataset. Hybrid species are obtained by stitching two images of different labels using cutmix. We quantify the probability that selected easy weak positive samples contain upper or bottom half during training using \"aut_up/down\". With our method, we also report \"our_up/down\" and \"our_left/right\" indicating the likelihood of easy weak positive samples containing upper or bottom half of the bird or left or right half of the stitched image, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 289, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 435 + ], + "type": "text", + "content": "body during metric learning, particularly the lower half, which is synthesized as the easy weak positive sample using the Cutmix augmentation technique. While the selection probability of the lower half increases from 0.6 to " + }, + { + "bbox": [ + 46, + 316, + 289, + 435 + ], + "type": "inline_equation", + "content": "0.7\\sim 0.8" + }, + { + "bbox": [ + 46, + 316, + 289, + 435 + ], + "type": "text", + "content": " during Proxy training, our proposed method ensures a probability of 1. Stitching hybrid species horizontally results in a selection probability of around 0.5. This approach allows the network to focus on important features during training, providing an advantage over traditional methods. Figure. 7 provides a visual representation of our findings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 227, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 227, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 227, + 466 + ], + "type": "text", + "content": "4.2.4 Visualization of Embedding Space" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "content": "We demonstrate the effectiveness of our proposed method on training and test data using t-SNE (t-distributed stochastic neighbor embedding) visualization technique. t-SNE can effectively reduce the dimensionality of the embedding space while preserving pairwise similarity between samples. It enables us to visually inspect how well the model separates different classes and identifies clusters of similar samples by plotting the embedding space in 2D or 3D. Therefore, t-SNE is a useful tool for evaluating the quality of learned embeddings in deep metric learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "We experimented on two classes from the training set and two classes (wren and woodpecker) from the test set. Figure. 8 shows that EPSHN is effective at distinguishing between different classes in the training set, but the embedding for hybrid species is relatively dispersed. With the additional HSE training signal, the embedding of the hybrid species becomes more concentrated and is located between the two classes. This approach considers all the features of difficult samples, creating a more comprehensive embedding of the samples. When additional training signals" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "are learned, the model can generalize better to unknown classes. This is demonstrated in the test set, where the EPSHN method cannot distinguish the black classes from the green and blue classes. However, when the training signal of HSE is applied to the downstream task, black classes can be successfully distinguished from green classes." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 155, + 359, + 182 + ], + "blocks": [ + { + "bbox": [ + 317, + 155, + 359, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 155, + 359, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 155, + 359, + 182 + ], + "type": "image", + "image_path": "ba07d0fc9327cface7ddee83aab881ce6f9d14292d11471cd69e6b5fde50df37.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 185, + 374, + 196 + ], + "lines": [ + { + "bbox": [ + 301, + 185, + 374, + 196 + ], + "spans": [ + { + "bbox": [ + 301, + 185, + 374, + 196 + ], + "type": "text", + "content": "(a) Oriole " + }, + { + "bbox": [ + 301, + 185, + 374, + 196 + ], + "type": "inline_equation", + "content": "(\\mathrm{id} = 95)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 405, + 155, + 447, + 182 + ], + "blocks": [ + { + "bbox": [ + 405, + 155, + 447, + 182 + ], + "lines": [ + { + "bbox": [ + 405, + 155, + 447, + 182 + ], + "spans": [ + { + "bbox": [ + 405, + 155, + 447, + 182 + ], + "type": "image", + "image_path": "50b0c5aa910ac598697179d39464f988ccd9c826c9adcea56fbad2b1c9da570f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 185, + 468, + 196 + ], + "lines": [ + { + "bbox": [ + 384, + 185, + 468, + 196 + ], + "spans": [ + { + "bbox": [ + 384, + 185, + 468, + 196 + ], + "type": "text", + "content": "(b) Nuthatch (id=94)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 498, + 152, + 532, + 185 + ], + "blocks": [ + { + "bbox": [ + 498, + 152, + 532, + 185 + ], + "lines": [ + { + "bbox": [ + 498, + 152, + 532, + 185 + ], + "spans": [ + { + "bbox": [ + 498, + 152, + 532, + 185 + ], + "type": "image", + "image_path": "a5787db3e0af7e6b248afe79b3790ffffc34d77409d977b70c786fd3aed4b978.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 187, + 538, + 199 + ], + "lines": [ + { + "bbox": [ + 491, + 187, + 538, + 199 + ], + "spans": [ + { + "bbox": [ + 491, + 187, + 538, + 199 + ], + "type": "text", + "content": "(c) Hspecie" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 337, + 214, + 396, + 274 + ], + "blocks": [ + { + "bbox": [ + 337, + 214, + 396, + 274 + ], + "lines": [ + { + "bbox": [ + 337, + 214, + 396, + 274 + ], + "spans": [ + { + "bbox": [ + 337, + 214, + 396, + 274 + ], + "type": "image", + "image_path": "b737f442c716e23eac9a6cf31d86497b3e7d55ffe9a104c89bba79382c6b3970.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 285, + 396, + 297 + ], + "lines": [ + { + "bbox": [ + 335, + 285, + 396, + 297 + ], + "spans": [ + { + "bbox": [ + 335, + 285, + 396, + 297 + ], + "type": "text", + "content": "EPSHN (train)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 458, + 215, + 517, + 273 + ], + "blocks": [ + { + "bbox": [ + 458, + 215, + 517, + 273 + ], + "lines": [ + { + "bbox": [ + 458, + 215, + 517, + 273 + ], + "spans": [ + { + "bbox": [ + 458, + 215, + 517, + 273 + ], + "type": "image", + "image_path": "fd43c4924038cbe776a1120c63747384b1c13ba4f0bcce3bfc6cc8987ac4503f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 285, + 528, + 297 + ], + "lines": [ + { + "bbox": [ + 445, + 285, + 528, + 297 + ], + "spans": [ + { + "bbox": [ + 445, + 285, + 528, + 297 + ], + "type": "text", + "content": "HSE-EPSHN (train)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 336, + 308, + 396, + 366 + ], + "blocks": [ + { + "bbox": [ + 336, + 308, + 396, + 366 + ], + "lines": [ + { + "bbox": [ + 336, + 308, + 396, + 366 + ], + "spans": [ + { + "bbox": [ + 336, + 308, + 396, + 366 + ], + "type": "image", + "image_path": "b418f8864e9908fbb992bce2ca470bee34251594dce4113ae98c072e5f8a7ded.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 377, + 394, + 388 + ], + "lines": [ + { + "bbox": [ + 337, + 377, + 394, + 388 + ], + "spans": [ + { + "bbox": [ + 337, + 377, + 394, + 388 + ], + "type": "text", + "content": "EPSHN (test)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 455, + 309, + 516, + 366 + ], + "blocks": [ + { + "bbox": [ + 455, + 309, + 516, + 366 + ], + "lines": [ + { + "bbox": [ + 455, + 309, + 516, + 366 + ], + "spans": [ + { + "bbox": [ + 455, + 309, + 516, + 366 + ], + "type": "image", + "image_path": "2f63165eb38949092075af43c9622c4acc9d0bf2f62bee5c73937eb4751f9043.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 377, + 523, + 389 + ], + "lines": [ + { + "bbox": [ + 445, + 377, + 523, + 389 + ], + "spans": [ + { + "bbox": [ + 445, + 377, + 523, + 389 + ], + "type": "text", + "content": "HSE-EPSHN (test)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 403, + 545, + 471 + ], + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 471 + ], + "type": "text", + "content": "Figure 8: We present a t-SNE visualization of EPSHN and HSE-EPSHN on the test set of CUB. Specifically, we select two different classes (denoted by blue dots and green dots) from the training set as shown in (a) and (b), respectively. (c) is the synthetic class from classes (a) and (b), denoted by red dots. For the test set, we selected four bird categories, consisting of two types of wrens and two types of woodpeckers, where each color represents a specific category." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 654 + ], + "type": "text", + "content": "Our Hybrid Species Embedding (HSE), a metric learning approach that aims to enable models to learn the spatial embedding of unknown and difficult classes during training. HSE generates additional hard samples using mixed data augmentation methods without defining their label information and applies them to the metric learning process. Experiments demonstrate significant improvements on multiple datasets when compared to three widely used pair-based methods and state-of-the-art proxy methods. HSE enhances the generalization ability of the model by providing additional training signals in deep metric learning." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgment. This work is partly supported by the National Natural Science Foundation of China (No: 62172366) and the Pioneer and Leading Goose R&D Program of Zhejiang Province (2023C01150)." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11055" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] E. Ahmed, M. Jones, and T. K. Marks. An improved deep learning architecture for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3908–3916, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 157 + ], + "type": "text", + "content": "[2] P. Chen, S. Liu, H. Zhao, and J. Jia. Gridmask data augmentation. arXiv preprint arXiv:2001.04086, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "text", + "content": "[3] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "text", + "content": "[4] T. DeVries and G. W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "type": "text", + "content": "[5] Z. Fu, Z. Mao, B. Hu, A.-A. Liu, and Y. Zhang. Intra-class adaptive augmentation with neighbor correction for deep metric learning. IEEE Transactions on Multimedia, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 291 + ], + "type": "text", + "content": "[6] W. Ge. Deep metric learning with hierarchical triplet loss. In Proceedings of the European Conference on Computer Vision (ECCV), pages 269-285, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 292, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 287, + 335 + ], + "type": "text", + "content": "[7] G. Gu, B. Ko, and H.-G. Kim. Proxy synthesis: Learning with synthetic classes for deep metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1460-1468, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 336, + 287, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 389 + ], + "type": "text", + "content": "[8] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 391, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 424 + ], + "type": "text", + "content": "[9] E. Harris, A. Marcu, M. Painter, M. Niranan, A. Prügel-Bennett, and J. Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 287, + 468 + ], + "type": "text", + "content": "[10] B. Harwood, V. Kumar BG, G. Carneiro, I. Reid, and T. Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 2821-2829, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "type": "text", + "content": "[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "text", + "content": "[12] A. Hermans, L. Beyer, and B. Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "type": "text", + "content": "[13] E. Hoffer and N. Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 581, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 623 + ], + "type": "text", + "content": "[14] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 646 + ], + "type": "text", + "content": "[15] S. Kim, D. Kim, M. Cho, and S. Kwak. Proxy anchor loss for deep metric learning. IEEE, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[16] W. Kim, B. Goyal, K. Chawla, J. Lee, and K. Kwon. Attention-based ensemble for deep metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 736-751, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 714 + ], + "type": "text", + "content": "[17] B. Ko and G. Gu. Embedding expansion: Augmentation in embedding space for deep metric learning. In Proceedings of" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7255-7264, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 140 + ], + "type": "text", + "content": "[18] B. Ko, G. Gu, and H.-G. Kim. Learning with memory-based virtual classes for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11792-11801, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 141, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 175 + ], + "type": "text", + "content": "[19] M. T. Law, N. Thome, and M. Cord. Quadruplet-wise image similarity learning. In Proceedings of the IEEE international conference on computer vision, pages 249-256, 2013." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 229 + ], + "type": "text", + "content": "[20] J.-H. Lee, M. Z. Zaheer, M. Astrid, and S.-I. Lee. Smoothmix: a simple yet effective data augmentation to train robust classifiers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 756-757, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 231, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 275 + ], + "type": "text", + "content": "[21] L. Liu, S. Huang, Z. Zhuang, R. Yang, M. Tan, and Y. Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[22] Z. Liu, P. Luo, S. Qiu, X. Wang, and X. Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "text", + "content": "[23] T. Milbich, K. Roth, H. Bharadhwaj, S. Sinha, Y. Bengio, B. Ommer, and J. P. Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In European Conference on Computer Vision, pages 590-607. Springer, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "text", + "content": "[24] D. D. Mohan, N. Sankaran, D. Fedorishin, S. Setlur, and V. Govindaraju. Moving in the right direction: A regularization for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14591-14599, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "text", + "content": "[25] Y. Movshovitz-Attias, A. Toshev, T. K. Leung, S. Ioffe, and S. Singh. No fuss distance metric learning using proxies. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "text", + "content": "[26] H. Oh Song, Y. Xiang, S. Jegelka, and S. Savarese. Deep metric learning via lifted structured feature embedding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4004-4012, 2016." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[27] M. Opitz, G. Waltner, H. Possegger, and H. Bischof. Deep metric learning with bier: Boosting independent embeddings robustly. IEEE transactions on pattern analysis and machine intelligence, 42(2):276-290, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[28] Q. Qian, L. Shang, B. Sun, J. Hu, H. Li, and R. Jin. Soft-triple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6450-6458, 2019." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "text", + "content": "[29] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[30] K. Roth, O. Vinyals, and Z. Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11056" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] K. Roth, O. Vinyals, and Z. Akata. Non-isotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 286, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 286, + 170 + ], + "type": "text", + "content": "[32] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 286, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 286, + 216 + ], + "type": "text", + "content": "[33] A. Sanakoyeu, V. Tschernezki, U. Buchler, and B. Ommer. Divide and conquer the embedding space for metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 471-480, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 286, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 286, + 259 + ], + "type": "text", + "content": "[34] F. Schroff, D. Kalenichenko, and J. Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 286, + 293 + ], + "type": "text", + "content": "[35] C. Shorten and T. M. Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 286, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 286, + 347 + ], + "type": "text", + "content": "[36] E. Simo-Serra, E. Trulls, L. Ferraz, I. Kokkinos, P. Fua, and F. Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 286, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 286, + 381 + ], + "type": "text", + "content": "[37] K. Sohn. Improved deep metric learning with multi-class n-pair loss objective. Advances in neural information processing systems, 29, 2016." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 286, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 286, + 424 + ], + "type": "text", + "content": "[38] S. Venkataramanan, B. Psomas, Y. Avrithis, E. Kijak, L. Amsaleg, and K. Karantzalos. It takes two to tango: Mixup for deep metric learning. arXiv preprint arXiv:2106.04990, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 286, + 468 + ], + "type": "text", + "content": "[39] J. Wang, F. Zhou, S. Wen, X. Liu, and Y. Lin. Deep metric learning with angular loss. In Proceedings of the IEEE international conference on computer vision, pages 2593-2601, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 523 + ], + "type": "text", + "content": "[40] X. Wang, X. Han, W. Huang, D. Dong, and M. R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5022-5030, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 568 + ], + "type": "text", + "content": "[41] C.-Y. Wu, R. Manmatha, A. J. Smola, and P. Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 612 + ], + "type": "text", + "content": "[42] H. Xuan, R. Souvenir, and R. Pless. Deep randomized ensembles for metric learning. In Proceedings of the European conference on computer vision (ECCV), pages 723-734, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 657 + ], + "type": "text", + "content": "[43] H. Xuan, A. Stylianou, and R. Pless. Improved embeddings with easy positive triplet mining. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2474-2482, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "type": "text", + "content": "[44] Y. Yuan, K. Yang, and C. Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE international conference on computer vision, pages 814-823, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "text", + "content": "[45] S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo. Cutmix: Regularization strategy to train strong classifiers with" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 342 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[46] D. Zhang, Y. Li, and Z. Zhang. Deep metric learning with spherical embedding. Advances in Neural Information Processing Systems, 33:18772-18783, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "text", + "content": "[47] H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[48] Y. Zhao, Z. Jin, G.-j. Qi, H. Lu, and X.-s. Hua. An adversarial approach to hard triplet generation. In Proceedings of the European conference on computer vision (ECCV), pages 501-517, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[49] W. Zheng, Z. Chen, J. Lu, and J. Zhou. Hardness-aware deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 72-81, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "text", + "content": "[50] W. Zheng, C. Wang, J. Lu, and J. Zhou. Deep compositional metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9320-9329, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[51] W. Zheng, B. Zhang, J. Lu, and J. Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11057" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_content_list.json b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ca92966f5113e513a8e8d49ea79846419039dea --- /dev/null +++ b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_content_list.json @@ -0,0 +1,2554 @@ +[ + { + "type": "text", + "text": "HSR-Diff: Hyperspectral Image Super-Resolution via Conditional Diffusion Models", + "text_level": 1, + "bbox": [ + 102, + 128, + 869, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chanyue Wu $^{1*}$ , Dong Wang $^{1*}$ , Yunpeng Bai $^{2*}$ , Hanyu Mao $^{1}$ , Ying Li $^{1\\dagger}$ , Qiang Shen $^{2}$", + "bbox": [ + 150, + 202, + 815, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, Shaanxi Provincial Key Laboratory of Speech & Image Information Processing, School of Computer Science, Northwestern Polytechnical University, China", + "bbox": [ + 98, + 220, + 870, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, Faculty of Business and Physical Sciences, Aberystwyth University, U.K.", + "bbox": [ + 120, + 273, + 846, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chanyuewu, dongwang, maomhy} $@$ mail.nwpu.edu.cn lybyp@nwpu.edu.cn {yub3, qqs} $@$ aber.ac.uk", + "bbox": [ + 135, + 310, + 831, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the proven significance of hyperspectral images (HSIs) in performing various computer vision tasks, its potential is adversely affected by the low-resolution (LR) property in the spatial domain, resulting from multiple physical factors. Inspired by recent advancements in deep generative models, we propose an HSI Super-resolution (SR) approach with Conditional Diffusion Models (HSR-Diff) that merges a high-resolution (HR) multispectral image (MSI) with the corresponding LR-HSI. HSR-Diff generates an HR-HSI via repeated refinement, in which the HR-HSI is initialized with pure Gaussian noise and iteratively refined. At each iteration, the noise is removed with a Conditional Denoising Transformer (CDFformer) that is trained on denoising at different noise levels, conditioned on the hierarchical feature maps of HR-MSI and LR-HSI. In addition, a progressive learning strategy is employed to exploit the global information of full-resolution images. Systematic experiments have been conducted on four public datasets, demonstrating that HSR-Diff outperforms state-of-the-art methods.", + "bbox": [ + 75, + 393, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 709, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperspectral images (HSIs) contain dozens or hundreds of spectral bands, enabling them to provide more faithful knowledge of targeted scenes than conventional imaging modalities. As such, HSIs play an irreplaceable role in various computer vision tasks, including classification [35, 42], segmentation [6], and tracking [32]. Although HSIs contain rich spectral information, contemporary hyperspectral", + "bbox": [ + 75, + 734, + 468, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8da2ed340bcb04d90a85f170285672c7a60e68bf0cfdbd485b9764fa5cf1f755.jpg", + "image_caption": [ + "Figure 1: Illustration of fusion-based Hyperspectral Image Super-Resolution (HSI-SR)." + ], + "image_footnote": [], + "bbox": [ + 540, + 359, + 856, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "imaging sensors lack high-resolution (HR) in the spatial domain, due to the stringent constraint of typically low signal-to-noise ratios. Their widespread use is significantly hindered by this fact. Restricted by hardware limitations, a practical way to work around this problem is to fuse the low-resolution (LR) HSI with an HR multispectral image (MSI). This requires the implementation of so-called HSI super-resolution (SR)[36], as shown in Figure 1.", + "bbox": [ + 496, + 550, + 892, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few decades, a significant amount of research efforts have been devoted to developing HSI-SR approaches, which can be roughly classified into five categories [44]: Extensions of pansharpening [5, 25, 27], Bayesian inference-based methods [1, 29], matrix factorization-based methods [2], tensor-based methods [18], and deep learning (DL)-based methods [31]. Whilst pansharpening methods [26] have been extended to the field of HSI-SR, such approaches are prone to spectral distortion. Bayesian inference-based approaches rely on the assumption of prior knowledge, thereby having a weak flexibility in dealing with different HSI structures. Matrix factorization-based techniques reshape the 3D HSIs and MSIs into matrices, thus facing the challenge of learning the required relationship between space and spectrum. Although sev-", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 851, + 202, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. This work is supported by the National Natural Science Foundation of China under Grant (62271400), and in part by the Shaanxi Provincial Key R&D Program, China (2023-GHZD-02).", + "bbox": [ + 76, + 863, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7083", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eral tensor-based methods have been proposed that can maintain the 3D structure of input images, they consume much more memory and computational power. Furthermore, these traditional approaches work via relying heavily on hand-crafted priors.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, DL-based methods, especially convolutional neural network (CNN)-based approaches, have flooded over into the HSI-SR research community [8, 31, 10, 22, 46, 43, 19]. Rather than resorting to hand-crafted features, DL-based techniques learn prior knowledge automatically from given data. Particularly, Dong et al. proposed the first DL-based method for image SR, with the end-to-end mapping between LR images and HR images learned using a CNN [9]. Subsequently, generative adversarial networks (GANs) were introduced to the field of image SR in an effort to produce high-frequency details [17, 11]. After that, various GAN-based models have been devised, showing state-of-the-art results in the HSI-SR literature [24]. However, such work requires carefully designed regularization and optimization tricks to tame optimization instability and avoid mode collapse.", + "bbox": [ + 75, + 169, + 470, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the recent developments in deep generative models [14, 23], in this paper, we propose an innovative approach that we refer to as HSR-Diff (HSI-SR with conditional diffusion models). It works by learning to transform the original standard normal distribution into the data distribution of HR-HSI through a sequence of refinements. In contrast to GAN-based methods which require inner-loop maximization, HSR-Diff operates by simply minimizing a well-defined loss function to learn prior knowledge. Although conditional diffusion models are straightforward to define and efficient to train, there has been no demonstration that they are capable of merging LR-HSI and HR-MSI to the best of our knowledge. We show that conditional diffusion models are capable of generating high-quality HR-HSIs, which may best the state-of-the-art results.", + "bbox": [ + 75, + 414, + 470, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key factor of HSR-Diff is its inherent denoising ability thanks to use of deep neural networks. In spite of the effectiveness of CNNs for denoising, they have shown limitations in modelling long-range dependencies. Some previous transformer-based HSI-SR works, especially Pyramid Shuffle-and-Reshuffle Transformer (PSRT) [7], Transformer-based Fusion network (Fusformer) [15], and Hyperspectral and Multispectral image Fusion Transformer (HMF-Former) [40], all of which learn long-range dependencies with attention mechanisms. These Transformers only take LR-HSI and HR-MSI as input, while the denoising Transformer is also fed with noise level $\\gamma$ and noisy HR-HSI. To comprehensively exploit the spatial and spectral information in the LR-HSI, HR-MSI, and noisy HR-HSI, a Conditional Denoising Transformer (CD-Former) is herein designed and trained with a denoising objective to remove various levels of noise iteratively. In ad", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dition, a progressive learning strategy is utilized to help the CDFormer learn the global statistics of full-resolution HSIs. The main contributions of this work are summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the novel application of conditional diffusion models in the field of HSI-SR that works by progressively destroying HR-HSI through injecting noise and subsequently learning to reverse this process, in order to perform HSI-SR. To the best of our knowledge, this is the first work to exploit diffusion models within the realm of HSI-SR.", + "- We introduce a CDFormer that refines a noisy HR-HSI conditioned on the deep feature maps of HR-MSI and LR-HSI, capable of modelling global connectivity with self-attention. Rather than concatenation, cross-attention is utilized as the conditioning mechanism to incorporate spatio-spectral information and noise level.", + "- We employ a progressive learning strategy to exploit the global information of full-resolution HSIs, with CDFormer being trained on small image patches in the early epochs with high efficiency and on the global images in the later epochs to acquire global information.", + "- We present experimental investigations on four public datasets, with quantitative and qualitative results illustrating the superior performance of our approach as compared with state-of-the-art methods." + ], + "bbox": [ + 517, + 165, + 890, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 564, + 640, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion models", + "text_level": 1, + "bbox": [ + 500, + 589, + 666, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typical deep generative models include autoregressive models (AR), normalizing flows (NF), variational autoencoders (VAE), GANs, diffusion models, etc. Over the past decade, GANs have gained prominence as a pivotal technology in the challenging field of image synthesis. However, this landscape has seen a transformation with the emergence of diffusion models, which have rapidly become a prominent research avenue across diverse domains, such as natural language processing, image processing, and computer vision. Diffusion models have arisen as an innovative approach that operates by iteratively refining a noise vector into the desired output. This reversible process allows for the generation of high-quality samples and has proven versatile in various applications.", + "bbox": [ + 496, + 613, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The development of diffusion models has seen a dramatically accelerating pace over the past three years. In the field of computer vision, diffusion models have shown promise in tasks like image segmentation [33], object detection [30], and even video processing [13]. Whilst diffusion models", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d388835a1cd9e49b9d7d217001d4d1efca42f8981acd60e33c519b9ec9521510.jpg", + "image_caption": [ + "Figure 2: Forward and reverse processes of HSR-Diff, with forward process $q$ generating an HSI sequence (left to right) by gradually adding Gaussian noise, and reverse process $p$ iteratively refining HR-HSI (right to left)." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 156 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have shown great potential for a variety of computer vision applications, none of them have yet been devoted to the problem of HSI-SR to the best of our knowledge. In this paper, we extend the utility of diffusion models to the field of HSI-SR.", + "bbox": [ + 75, + 253, + 470, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Deep Learning-Based HSI-SR", + "text_level": 1, + "bbox": [ + 76, + 340, + 344, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, data-driven CNN architectures have been shown to outperform traditional approaches for use in the HSI-SR literature. These methods formulate the underlying fusion problem as a highly nonlinear mapping that takes HR-MSIs and LR-HSIs as input to generate an optimal HR-HSI. For example, CMHF-net [31] is an interpretable CNN, the design of which exploits the deep unfolding technique. Zhang et al. [43] proposed to reconstruct HR-HSIs with a two-stage network, while Zhang et al. [45] designed an interpretable spatial-spectral reconstruction network (SSR-NET) based on CNN. Aiming at problems of inflexible structure and information distortion, Jin et al. embedded Bilateral Activation Mechanism into ResNet, resulting in the effective model of BRResNet [16]. Thanks to the inductive bias of CNN, such as locality and weight sharing, these methods can provide good generalization performance and achieve impressive results. Nevertheless, CNNs have limitations in capturing long-range dependencies and self-similarity priors. To overcome such shortcomings, some Transformer-based HSI-SR approaches [15, 40] have been proposed.", + "bbox": [ + 75, + 364, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Within this study, the CDFormer serves as the denoising network in conditional diffusion models. It harnesses self-attention to learn global statistics from full-resolution HSIs and MSIs and employs cross-attention as the conditioning mechanism.", + "bbox": [ + 75, + 683, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Methodology", + "text_level": 1, + "bbox": [ + 76, + 773, + 294, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 799, + 279, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without losing generality, the observation models for the HR-MSI and LR-HSI of interest can be mathematically formulated as", + "bbox": [ + 75, + 824, + 468, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = \\mathbf {R Z} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 867, + 468, + 891 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\mathbf {Z D},\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 888, + 308, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{X} \\in \\mathbb{R}^{b \\times HW}$ denotes the HR-MSI which consists of $b$ spectral bands with a resolution of $HW$ in the spatial domain; $\\mathbf{R} \\in \\mathbb{R}^{b \\times B}$ represents the spectral response function of HR-MSI; $\\mathbf{Y} \\in \\mathbb{R}^{B \\times hw}$ denotes the LR-HSI; and $\\mathbf{Z} \\in \\mathbb{R}^{B \\times HW}$ is the HR-HSI. In the above, $b$ and $B$ are the numbers of bands, with $h$ and $H$ being the band height, and $w$ and $W$ the width, where $b \\ll B$ , $h \\ll H$ , and $w \\ll W$ . $\\mathbf{D} \\in \\mathbb{R}^{HW \\times hw}$ is the spatial response of the LR-HSI, which can be modelled with blurring and down-sampling operations. The HSI-SR can be interpreted as an inverse problem for merging a practically collected $\\mathbf{X}$ and an observed $\\mathbf{Y}$ to produce a latent $\\mathbf{Z}$ . In this paper, the ideal $\\mathbf{Z}$ is restored with HSR-Diff conditioned on spatio-spectral information of $\\mathbf{X}$ and $\\mathbf{Y}$ , the details of which are described below.", + "bbox": [ + 496, + 89, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. HSI-SR with Conditional Diffusion Models", + "text_level": 1, + "bbox": [ + 498, + 314, + 864, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a dataset $\\mathcal{D}_{train} = \\{\\mathbf{X}^i,\\mathbf{Y}^i,\\mathbf{Z}^i\\}_{i = 1}^N$ satisfying a certain joint probability distribution $p(\\mathbf{X},\\mathbf{Y},\\mathbf{Z})$ , many pairs of $(\\mathbf{X},\\mathbf{Y})$ may be consistent with the same $\\mathbf{Z}$ . Thus, the HR-HSI $\\mathbf{Z}$ can be obtained with iterative refinement that provide an approximate to $p(\\mathbf{Z}|\\mathbf{X},\\mathbf{Y})$ . In this work, we implement the process of iterative refinement with HSR-Diff, where the optimized HR-HSI is presumed to be produced in $T$ refinement steps. In HSR-Diff, the target HR-HSI is initialized with a pure noise $\\mathbf{Z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ as shown in Figure 2. The HSI is then refined iteratively according to learned conditional distributions $p_{\\theta}(\\mathbf{Z}_{t - 1}|\\mathbf{Z}_t,\\mathbf{X},\\mathbf{Y})$ . In so doing, the image sequence $(\\mathbf{Z}_{T - 1},\\mathbf{Z}_{T - 2},\\dots ,\\mathbf{Z}_0)$ can be attained and ultimately $\\mathbf{Z}_0\\sim p(\\mathbf{Z}|\\mathbf{X},\\mathbf{Y})$ .", + "bbox": [ + 496, + 338, + 890, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The HSR-Diff makes use of two processes: a forward process that perturbs HR-HSI $\\mathbf{Z}_0$ to noise, and a reverse process converting noise back to HR-HSI $\\mathbf{Z}_0$ . In the forward process, the intermediate images, i.e., $\\mathbf{Z}_{T - 1}$ , $\\mathbf{Z}_{T - 2}$ , $\\dots$ , and $\\mathbf{Z}_1$ , are generated according to a Markov chain with fixed transition probability $q(\\mathbf{Z}_t|\\mathbf{Z}_{t - 1})$ . We are interested in reversing the process via iterative refinement, in which the noise is reduced iteratively with a reverse Markov chain conditioned on $\\mathbf{X}$ and $\\mathbf{Y}$ . The reverse chain is learned with the CDFformer $f_{\\theta}$ . Further details of HSR-Diff's working are given below.", + "bbox": [ + 496, + 537, + 892, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Forward Process", + "text_level": 1, + "bbox": [ + 500, + 727, + 671, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by [14], forward process $q$ iteratively adds Gaussian noise to $\\mathbf{Z}_0$ over $T$ iterations:", + "bbox": [ + 498, + 752, + 890, + 782 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {Z} _ {1: T} \\mid \\mathbf {Z} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {t - 1}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 795, + 890, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {Z} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 839, + 852, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{1:T} \\in (0,1)$ are scalar hyper-parameters. Note that in the forward process, the distribution of $\\mathbf{Z}_t$ given $\\mathbf{Z}_0$ can", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/244e86f93651d9b3ac4b788864e2c31da74688de21ac6075427a0351c496c294.jpg", + "image_caption": [ + "Figure 3: Architecture of Conditional Denoising Transformer." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 883, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be directly sampled in closed form. This implies that", + "bbox": [ + 76, + 395, + 428, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t}; \\sqrt {\\gamma_ {t}} \\mathbf {Z} _ {0}, (1 - \\gamma_ {t}) \\mathbf {I}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 426, + 468, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{t} = \\prod_{i = 1}^{t}\\alpha_{i}$ . In addition, the posterior distribution of $\\mathbf{Z}_{t - 1}$ given $\\mathbf{Z}_0$ and $\\mathbf{Z}_t$ can be derived by", + "bbox": [ + 76, + 460, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {0}, \\mathbf {Z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t - 1}; \\boldsymbol {\\mu}, \\sigma^ {2} \\mathbf {I}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 507, + 375, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} = \\frac {\\sqrt {\\gamma_ {t - 1}} (1 - \\alpha_ {t})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {0} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\gamma_ {t - 1})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {t} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 527, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {2} = \\frac {(1 - \\gamma_ {t - 1}) (1 - \\alpha_ {t})}{1 - \\gamma_ {t}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 563, + 302, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the CDFormer is tasked with predicting $\\mathbf{Z}_0$ , the posterior is useful in the reverse process.", + "bbox": [ + 76, + 609, + 468, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Reverse Markovian Process", + "text_level": 1, + "bbox": [ + 76, + 671, + 323, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reverse process inferences $\\mathbf{Z}_0$ via iterative refinement. It starts from a pure Gaussian noise $\\mathbf{Z}_T$ and goes in the opposite direction of the forward process:", + "bbox": [ + 76, + 699, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {Z} _ {0: T} \\mid \\mathbf {X}, \\mathbf {Y}\\right) = p \\left(\\mathbf {Z} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {t}, \\mathbf {X}, \\mathbf {Y}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 761, + 452, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {Z} _ {T}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {T}; \\mathbf {0}, \\mathbf {I}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 804, + 326, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {t}, \\mathbf {X}, \\mathbf {Y}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {X}, \\mathbf {Y}, \\mathbf {Z} _ {t}, \\gamma_ {t}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 824, + 472, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the distribution $p_{\\theta}(\\mathbf{Z}_{t - 1}\\mid \\mathbf{Z}_t,\\mathbf{X},\\mathbf{Y})$ is parameterized with $\\theta$ . Note that the CDFormer provides a prediction of $\\hat{\\mathbf{Z}}_0$ . Thus, according to (4), each refinement step takes", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the following form:", + "bbox": [ + 500, + 396, + 633, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Z} _ {t - 1} = \\frac {\\sqrt {\\gamma_ {t - 1}} (1 - \\alpha_ {t})}{1 - \\gamma_ {t}} f _ {\\theta} (\\mathbf {X}, \\mathbf {Y}, \\mathbf {Z} _ {t}, \\gamma_ {t}) \\tag {6} \\\\ + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\gamma_ {t - 1})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {t} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 422, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and $f_{\\theta}$ is the CDFormer.", + "bbox": [ + 500, + 501, + 797, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Noise Schedule", + "text_level": 1, + "bbox": [ + 500, + 537, + 660, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the research reported in [4], we sample $\\gamma$ with two steps during training. In particular, we first sample a time step $t\\sim U\\{1,T\\}$ and then randomly select $\\gamma \\sim U(\\gamma_{t - 1},\\gamma_t)$ . As such, $\\gamma \\sim p(\\gamma) = \\sum_{t = 1}^{T}\\frac{1}{T} U(\\gamma_{t - 1},\\gamma_t)$ . Normally, the model with a large $T$ can achieve better results. However, we find (through empirical investigations) that the performance is not very sensitive to the exact values of $T$ . Therefore, no hyper-parameter search about $T$ is conducted and we set $T = 2000$ for simplicity. As for the inference process, we set the maximum generation iterations to 100, employing a linear noise schedule.", + "bbox": [ + 498, + 561, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditional Denoising Transformer", + "text_level": 1, + "bbox": [ + 500, + 739, + 808, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The property of non-local self-similarity of HSIs is often exploited in denoising tasks but is usually not well captured by CNN-based models. Due to the effectiveness of Transformer layer in capturing non-local long-range dependencies, the potential of Transformer is explored in conditional denoising of HSI. Unfortunately, the vanilla Transformer focuses only on spatial relationships between pixels while neglecting the spectral dimension. Besides, denoising networks in conditional diffusion models normally concatenate", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7086", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d62b5e0682aa49c78a1d8b309984db45471b2bc594e44417c5d6f86b2d2975d7.jpg", + "image_caption": [ + "Figure 4: Spatio-Spectral Transformer Layer, where \"LN\" denotes layer normalization." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 439, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "all images together as input, which may hinder the extraction of useful spatio-spectral information in LR-HSIs and HR-MSIs. Hence, the CDFormer adopts a two-stream architecture and is constructed with stacked Spatio-Spectral Transformer Layers (S $^2$ TLs).", + "bbox": [ + 75, + 239, + 467, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The architecture of the CDFormer is shown in Figure 3. The SR stream first utilizes a $3 \\times 3$ convolution to generate low-level feature embeddings $\\mathbf{F}_0^{SR}$ and then transforms it into deep features $\\mathbf{F}_l^{SR}$ with stacked- $\\mathbf{S}^2\\mathbf{TLs}$ . Instead of adopting $t$ as done in the existing work [4], our method is conditioned on $\\gamma$ directly to achieve efficient generation. The denoising stream contains multiple noise-aware conditional $\\mathbf{S}^2\\mathbf{TLs}$ (NC- $\\mathbf{S}^2\\mathbf{TLs}$ ) that take as input the embedded noise level and the image representation $\\mathbf{F}^{SR}$ . The Reconstruction module is set to produce a noise-free HR-HSI, by employing residual learning to alleviate the difficulty of HR-HSI generation while mapping the features onto HR-HSI via a $3 \\times 3$ convolution and addition operation. It should be noted that no downsampling operation is employed.", + "bbox": [ + 75, + 316, + 472, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Noise Level Embedding", + "text_level": 1, + "bbox": [ + 76, + 551, + 295, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The noise level offers essential information for denoising models. Inspired by the work of [4], we embed noise level within the models with sinusoidal positional encoding. The process of noise level embedding (NLE) can be formulated as follows:", + "bbox": [ + 75, + 577, + 468, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nN L E _ {\\gamma , 2 i} = \\sin \\left(\\gamma / 1 0 0 0 0 ^ {2 i / C}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 662, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nN L E _ {\\gamma , 2 i + 1} = \\cos \\left(\\gamma / 1 0 0 0 0 ^ {2 i / C}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 691, + 393, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is the number of channels of $\\mathbf{S}^2\\mathrm{TLs}$ . $i\\in [1,C / 2]$", + "bbox": [ + 76, + 729, + 465, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Spatio-Spectral Transformer Layers", + "text_level": 1, + "bbox": [ + 76, + 768, + 383, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 illustrates the architecture of one $\\mathrm{S}^2\\mathrm{TL}$ , which consists of a Spatial Multi-head Self-Attention (SpatioMSA), a Spectral Multi-head Self-Attention (SpectralMSA), and a Feed Forward Network (FFN). SpatioMSA and SpectralMSA learn the interactions of spatial regions and interspectra relationships, respectively. To alleviate the computational burden, we adopt the transposed attention [41] in", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a284d97c03c04043a8c65162228a54afa8c9a537273be18c20fe4933401ecfd.jpg", + "image_caption": [ + "Figure 5: Noise-Aware Conditional Spatio-Spectral Transformer Layer, where \"LT\" represents linear transform, \"LN\" denotes layer normalization, and \"MCA\" is the abbreviation of multi-head cross-attention." + ], + "image_footnote": [], + "bbox": [ + 563, + 87, + 828, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SpectralMSA. SpatioMSA applies the popular window partitioning strategy [20] to reduce the computational complexity. In addition, the gating mechanism [41] is employed in the implementation of FFN.", + "bbox": [ + 498, + 279, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Noise-Aware Conditional $\\mathbf{S}^2\\mathbf{T}\\mathbf{L}\\mathbf{s}$", + "text_level": 1, + "bbox": [ + 500, + 359, + 777, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To condition the overall model on the hierarchical features of SR stream, we feed $\\mathbf{F}_l^{SR}$ to the Noise-Aware Conditional S $^2$ TLs (NC-S $^2$ TLs), each of which is a key building block of the denoising stream. Figure 5 depicts the structure of an NC-S $^2$ TL, which takes as input the NLE (a vector), $\\mathbf{F}_l^{SR}$ and $\\mathbf{F}_l^{DS}$ , where $\\mathbf{F}_l^{SR}$ and $\\mathbf{F}_l^{DS}$ have the same spatial resolution. NLE is first transformed and merged with $\\mathbf{F}_l^{DS}$ with the result subsequently processed with the means of multi-head cross-attention (MCA) [3] in order to condition the model on $\\mathbf{F}_l^{SR}$ . As a result, each S $^2$ TL learns the spatio-spectral dependencies.", + "bbox": [ + 498, + 383, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Progressive Learning Strategy", + "text_level": 1, + "bbox": [ + 500, + 559, + 767, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deep learning-based HSI-SR models are normally trained on small cropped patches. However, on one hand, training the proposed CDFormer on fixed-size image patches may not appropriately reflect the global relationships, which may result in suboptimal performance on full-resolution images when used. On the other hand, during the training phase, CDFormer can not be uploaded to Graphics Processing Unit (GPU) because of the limited memory. With this objective in mind, we adopt a strategy of progressive learning. In the initial epochs, the network is trained using smaller image patches. Subsequently, as training progresses, we transition towards utilizing larger patches, and ultimately, full-resolution images in the concluding training epochs. The resulting model, cultivated through the integration of varying sizes via progressive learning, exhibits notable enhancements in performance during testing. This is particularly evident when handling images of diverse resolutions, a scenario frequently encountered in the realm of the HSI-SR task.", + "bbox": [ + 496, + 582, + 890, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reduce the pressure on the demand of GPU memory, we only train the second half of CDFormer on full", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resolution images. The loss function used for such training is defined as follows:", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\| \\mathbf {X} - \\mathbf {R} \\hat {\\mathbf {Z}} _ {0} \\| _ {1} + \\| \\mathbf {Y} - \\hat {\\mathbf {Z}} _ {0} \\mathbf {D} \\| _ {1} + \\| \\mathbf {Z} _ {0} - \\hat {\\mathbf {Z}} _ {0} \\| _ {1} \\tag {8} \\\\ \\hat {\\mathbf {Z}} _ {0} = f _ {\\theta} (\\sqrt {\\gamma} \\mathbf {Z} _ {0} + \\sqrt {1 - \\gamma} \\boldsymbol {\\epsilon}, \\mathbf {X}, \\mathbf {Y}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 132, + 467, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ $(\\mathbf{X},\\mathbf{Y},\\mathbf{Z})$ is sampled from the training set, and the noise schedule about $\\gamma$ has been discussed above. The first two terms are designed according to the observation models, while the last one is based on the assumption of Laplace distribution.", + "bbox": [ + 75, + 183, + 467, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 275, + 207, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Systematic experiments are herein conducted on four commonly-used public-available HSI-SR datasets to demonstrate the effectiveness of the proposed approach.", + "bbox": [ + 75, + 301, + 467, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 358, + 178, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Four datasets including CAVE [38], PaviaU [21], Chikusei [39], and HypSen [37] are used in our experiments, with the following details on each.", + "bbox": [ + 75, + 381, + 467, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CAVE: There are 32 scenes with a spatial size of $512 \\times 512$ in the CAVE dataset, where we select the first 20 HSIs for training, with the remaining 12 images used for testing. We generate LR-HSIs by Gaussian blur and down-sampling using a factor of 32 as done in [31]. HR-MSIs are acquired by integrating all HR-HSI bands according to the spectral response function of Nikon D700. The original HR-HSIs are treated as ground truth.", + "bbox": [ + 75, + 448, + 467, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PaviaU: Collected by the University of Pavia, Italy, the original HSI dataset consists of $610 \\times 340$ pixels in which the top-left $128 \\times 128$ area is extracted as the test data, with the remaining used for training. Except for water absorption bands, all other 103 bands are chosen for the experiments. Note that the down-sampling factor for the generation of LR-HSIs is four, and the spectral response function is the same as that of the WorldView-3 satellite.", + "bbox": [ + 75, + 590, + 467, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Chikusei: This dataset consists of 128 bands with a spectral range of $363~nm$ to $1018~nm$ and a spatial resolution of $2517\\times 2335$ . The original HSI data was taken by an airborne visible and near-infrared imaging sensor over Chikusei, Japan. To alleviate the impact of the back boundary and noise, we crop the center area and remove noise bands. The processed image has a size of $2048\\times 2048\\times 110$ . The top half $1024\\times 2048\\times 110$ area is selected as the training data, while the rest half is split into eight testing $512\\times 512\\times 110$ patches. For the production of LR-HSIs and HR-MSIs, this dataset adopts the same processing as with PaviaU.", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HypSen: This dataset concerns a real scenario consisting of a $30\\mathrm{m}$ -resolution HSI and a $10\\mathrm{m}$ -resolution MSI. The Hyperion sensor on the Earth Observing-1 satellite provided the HSI with 242 spectral bands in the spectral range of $400\\mathrm{nm}$ to $2500\\mathrm{nm}$ , and the MSI with 13 bands was captured by the Sentinel-2A satellite. The blue, green, red, and near-infrared bands of MSI in our experiments are selected due to their high spatial resolution. To eliminate the impact of noise and water absorption, we remove those relevant bands, with 84 bands remaining in the HSI. We crop sub-images of size $250\\times 330$ and $750\\times 990$ from the Hyperion HSI and Sentinel-2A MSI respectively, in our study, with the pairs of sub-image patches spatially registered.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Methods Compared and Evaluation Metrics Used", + "text_level": 1, + "bbox": [ + 498, + 296, + 890, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Six state-of-the-art HSI-SR approaches are taken for comparison, including: UTV-TD [34], UAL [43], BRResNet [16], Fusformer [15], CMHF-Net [31], and UAL-DMI [28]. UTV-TD is a tensor-based technique; UAL, BRResNet, Fusformer, and CMHF-Net fall into the category of the DL-based methods; and UAL-DMI can be regarded as an upgraded version of UAL.", + "bbox": [ + 496, + 335, + 890, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Four quantitative quality metrics are employed for performance evaluation, including Peak Signal-to-Noise Ratio (PSNR), Spectral Angle Mapper (SAM), Erreur Relative Globale Adimensionnelle de Synthese (ERGAS, namely error relative global dimensionless synthesis), and Structure Similarity (SSIM). The smaller ERGAS and SAM are, the larger PSNR and SSIM are, the better the fusion result is.", + "bbox": [ + 496, + 441, + 890, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Specification", + "text_level": 1, + "bbox": [ + 498, + 556, + 766, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All DL-based methods are trained on the same datasets. For those compared methods, we use the publicly available source codes with default hyper-parameters as given in the corresponding research papers. Our HSR-Diff is implemented on the PyTorch framework. The learnable parameters of the CDFformer are initialized with Kaiming initialization [12] and trained on 2 NVIDIA GeForce GTX 3090s. The dimension of the embedding of $F_0^{SR}$ and $F_0^{DS}$ is set to 256, and the number of parameters of the CDFformer is 31.56M. We utilize the Adam optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ to optimize the CDFformer. With limited GPU memory, the batch size is set to 4 and 2 for $128^2$ and $512^2$ images, respectively. It costs 20000 epochs on the CAVE and PaviaU datasets while consuming 5000 epochs on the Chikusei dataset. The learning rate is initialized as $1 \\times 10^{-4}$ .", + "bbox": [ + 496, + 580, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 498, + 816, + 869, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this set of experiments, the evaluations are carried out using the first three datasets listed above without involving the real-world dataset, HypSen (which will be dealt with in the next section).", + "bbox": [ + 496, + 840, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cba18c9ddba5abe00b2a6ddfbcb11954b588a839e2556220fcddf8f8e28c2ebf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 87, + 179, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf86ea733028e5c49c7dc3cd3963201f0ac1c97201694df1ebfb1cb0c1cf78ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 88, + 279, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e19c26d5f1691568c600a3db99e69618a6c904aeae17a99cd94a1696111aae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 380, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06006c40dcd41774640c94589a49c8dfe77201c56bc53e5e42d7bbb2b33004d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 88, + 482, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/620a9d1cc61f2fb29768f5df4763c78f3f6c93c4c93ed9733072a159ffed88f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 584, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d557b71e6b2f660b7dff00195628a886a39a02b194227397046cc543380e87c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 88, + 684, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92977a1bf99e675b1918955657bc05ec4e1403271aac5eaa2918f409990a0061.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 88, + 785, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55c004a6a257bab4f7f3b70c6466379de314a974554fbeb4a284c487cc75ee60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 88, + 887, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/107074d9837edb1f92e3e6afa931490806ff550fecff53675a16ff6dce68e0a6.jpg", + "image_caption": [ + "(a)UTV-TD" + ], + "image_footnote": [], + "bbox": [ + 83, + 172, + 179, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf6681f2aa80f1306ec5f258f396df9745e5a42604d9bb12017c660105eb98cb.jpg", + "image_caption": [ + "(b) UAL" + ], + "image_footnote": [], + "bbox": [ + 183, + 172, + 279, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b0234e0ea0389fc716690f082783c6e86cfb82da0fecc048a62e00bad743da8.jpg", + "image_caption": [ + "(c) BRResNet" + ], + "image_footnote": [], + "bbox": [ + 282, + 172, + 380, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c5ce54149a2397db2d2e59aca0e79b71eeadc7f1557cb4c1dd2369262c55c48.jpg", + "image_caption": [ + "(d) CMHF-Net" + ], + "image_footnote": [], + "bbox": [ + 385, + 172, + 482, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa377e3121c60ee7e6a74835c36cf852ca856063994f2b7f2557d416f3f69e40.jpg", + "image_caption": [ + "(e) Fusformer" + ], + "image_footnote": [], + "bbox": [ + 486, + 172, + 584, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20269a988c05acb14b1bd1d06ed8c374f73ec7c3a73e4919d06afb917669df67.jpg", + "image_caption": [ + "(f) UAL-DMI" + ], + "image_footnote": [], + "bbox": [ + 588, + 172, + 684, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d4e4ab7f93e145eca4839ff18a91678668521abcc305deee4a80fd40acebeab.jpg", + "image_caption": [ + "(g) HSR-Diff" + ], + "image_footnote": [], + "bbox": [ + 689, + 172, + 785, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90c273aa673a219ee3995e474f0872034fb801ede3aef075ecb307616887e0e1.jpg", + "image_caption": [ + "(h) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 790, + 172, + 887, + 248 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af089bc0248aba477c785501900a8bdaff7d901bc35a8f7276ceaa17b9233146.jpg", + "image_caption": [ + "Figure 6: Visual quality comparison for fused HSIs of all competing methods on CAVE, where first and second rows show fourth band and corresponding heatmaps (mean squared error), respectively." + ], + "image_footnote": [], + "bbox": [ + 259, + 267, + 709, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f05a70c336542e6a5c8a56a77b432068d2c241b78cf8c3d189533bd05e494bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 345, + 179, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c3affc17c59a43ae29d90f24e1fe8b7e1bb42affb45f282fe47bd721b7c7bed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 347, + 279, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85cec20e87badfddcf7056487c5696bcf68e0a2ed623a850ee0141ce3ad1de2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 347, + 380, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c358e851f5dce380e6c6d400642f8c9b775d2e5c0c46749ccb0cad36c5085df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 347, + 482, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5e0e90785b7e242e23f9aa0ae81536e743e21e3cd10976cd79f1d71c4ecc2fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 347, + 584, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06a99816e30cc328bcd3262f6dfdba2093ea142821f323fce9b4c9afa85c8222.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 347, + 684, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/825a15e0b18e625c3152acf7180b5dc549c551cb056d22289638ca572afc24bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 347, + 785, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb0a566c8b2ce8acf2ce754e33febc80d1aae4b0ad151f101f02fb3622305112.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 347, + 887, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26058ff6e433fc931062169f70a0ccee449e274b5498acf298a508503cfd81aa.jpg", + "image_caption": [ + "(a)UTV-TD" + ], + "image_footnote": [], + "bbox": [ + 83, + 433, + 179, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a56509deef7e4e3bc20c38d3b77424cb97b145379098ff744410ec779afe0caf.jpg", + "image_caption": [ + "(b) UAL" + ], + "image_footnote": [], + "bbox": [ + 183, + 433, + 279, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/366c012d455662c4610b0e1b9cf53f177cf025fe10d57b25267520c065163a89.jpg", + "image_caption": [ + "(c) BRResNet" + ], + "image_footnote": [], + "bbox": [ + 285, + 433, + 380, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cb71bb58a8c6441629007401dcd2a2c3e88af5a0a6de0e77fbb917d55ef2d3f.jpg", + "image_caption": [ + "(d) CMHF-Net" + ], + "image_footnote": [], + "bbox": [ + 385, + 433, + 482, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3db35b2163920ecfce7896ae0f3bb81dfb8842cba1bff4516cd54fbd969c9503.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 433, + 584, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d03d5cd96bfef70c112b81cb88c34bf067c12773329836f86736d8b223c1a1b9.jpg", + "image_caption": [ + "(f) UAL-DMI" + ], + "image_footnote": [], + "bbox": [ + 588, + 433, + 684, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6ec2f8b52aa5fb40683f058ed144bbce24113e3206a5f795c4fe5ab49fa246c.jpg", + "image_caption": [ + "(g) HSR-Diff" + ], + "image_footnote": [], + "bbox": [ + 689, + 433, + 785, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b464dd4a7a81058c9b4a0d6b4cce870ee7ed08b6b4ccfbd6152b19e1a703e752.jpg", + "image_caption": [ + "(h) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 790, + 433, + 887, + 507 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea2453226b911803317da1556b439e4d300ee7c4c0b85e537953fcaa8dc87bd4.jpg", + "image_caption": [ + "(e) Fusformer" + ], + "image_footnote": [], + "bbox": [ + 261, + 526, + 709, + 550 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0893649ed3b9b4284f388e1e8c614dd98651dc04d735063c4521e08d66d97c4.jpg", + "image_caption": [ + "Figure 7: Visual quality comparison for fused HSIs of all competing methods on PaviaU, where first and second rows show 81st band and corresponding heatmaps (mean squared error), respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 604, + 179, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1f80953259ba7aac49f4b03d2ad2215aa949516f66ceeeaf4876535b7fd83ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 604, + 279, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b691686140c59179ab86adbe497853c99b510b2e0c3a94a229c3108161becf74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 604, + 380, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf0051760551a1149d076a9f64dd1ef0c897b3a421733b3cf7ecf2bfaa43b4d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 604, + 482, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/596261dd5b1ff33bd3f61d16fc0068efafd3c07b6bec686f6016c7c184f818fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 604, + 584, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff689d97198edd375e059155ece0a53219aff4463f11ab039162f9087a683206.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 604, + 684, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6902bdd95fb516980f87f20722069d90e2bf0eed85656eb2a9c027a1f2a3f528.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 604, + 785, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd07ddb4ac56f2ec519beee3336ec1d3b2481ea050228cd9c47825553833a902.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 604, + 887, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eaef687c57d2de4199b95308df6cb6f5ce3de64d8a1df75b8fb2fae63b822f7d.jpg", + "image_caption": [ + "(a)UTV-TD" + ], + "image_footnote": [], + "bbox": [ + 83, + 691, + 179, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfcb16a0fe695da51f0ace142c54f2153e27d75791f96221cac4e66f8649989e.jpg", + "image_caption": [ + "(b) UAL" + ], + "image_footnote": [], + "bbox": [ + 183, + 691, + 279, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8c302a9c5e986b5d5f199568399b95dda997d43549e1638f4ed122d62cd8a9f.jpg", + "image_caption": [ + "(c) BRResNet" + ], + "image_footnote": [], + "bbox": [ + 285, + 691, + 380, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fa7947ad384a6ce64801c60ed35ed87d07a17027c11a06d50539e33ca9f9aba.jpg", + "image_caption": [ + "(d) CMHF-Net" + ], + "image_footnote": [], + "bbox": [ + 385, + 691, + 482, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e8cfd22c89df9d0b2c1ceb685f0ffbb09811e30e0e4b6fdaa4049d8aff97f39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 691, + 584, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9822930510caa0752893ef4b337cf40ec9c55b584f63f39408356a70782ec487.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 691, + 684, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d1b36b224b1735acfdbc56e6170292e73f3eaf7a1f62624bae76849243284d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 691, + 785, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3502d620be0556cb67bbe1b5167be649031f571fdc4e20959d12f27f77a8a77f.jpg", + "image_caption": [ + "(h) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 790, + 691, + 887, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c257a0e9bbfb832178a24e0e8d365e6fd0bcc1cbca1cbfe35e4a4a795069305.jpg", + "image_caption": [ + "(e) Fusformer", + "(f) UAL-DMI", + "(g) HSR-Diff", + "Figure 8: Visual quality comparison for fused HSIs of all competing methods on Chikusei, where first and second rows show 67th band and corresponding heatmaps (mean squared error), respectively." + ], + "image_footnote": [], + "bbox": [ + 261, + 782, + 709, + 808 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d8d8402d20e8bc74c3341e0ce2c00e7fe1679421d4153113868686125f75a34.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodsPSNR↑SSIM↑SAM↓ERGAS↓
CAVEUTV-TD [34]38.660.97997.980.329
UAL [43]40.550.99334.330.271
BRResNet [16]41.360.99294.700.250
CMHF-Net [31]42.540.99394.690.216
Fusformer [15]41.520.99344.710.243
UAL-DMI [28]42.740.99503.790.213
HSR-Diff44.330.99513.710.179
PaviaUUTV-TD [34]44.460.99521.801.236
UAL [43]45.420.99641.541.148
BRResNet [16]45.530.99651.531.111
CMHF-Net [31]45.770.99651.501.096
Fusformer [15]45.660.99651.521.109
UAL-DMI [28]45.680.99661.491.113
HSR-Diff46.470.99771.451.053
ChikuseiUTV-TD [34]48.380.99890.991.303
UAL [43]56.180.99980.490.421
BRResNet [16]56.790.99980.460.366
CMHF-Net [31]55.990.99980.500.483
Fusformer [15]55.920.99980.520.492
UAL-DMI [28]56.570.99980.470.387
HSR-Diff57.340.99990.430.324
", + "bbox": [ + 80, + 88, + 468, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison. To assess the performance of HSR-Diff qualitatively, we visualize example bands of HSIs in Figures 6, 7, and 8. It can be seen from these visual results that all compared methods produce satisfactory outcomes. In particular, HSR-Diff generates the best result with minor errors since the corresponding MSE (mean squared error) images are much clearer than the others.", + "bbox": [ + 75, + 530, + 468, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Comparison. To further verify the superior performance of the proposed HSR-Diff, quantitative results are presented in Table 1. Note that the performance indices on the CAVE and Chikusei datasets are averaged over all testing samples (12 samples for CAVE and eight samples for Chikusei), respectively. It can be inferred from the results that the proposed HSR-Diff surpasses all competitors with a clear margin on all evaluation metrics.", + "bbox": [ + 75, + 655, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 786, + 230, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of conditional diffusion models. Much of the early work on HSI-SR was based on the use of the regression model. To compare the effects of the diffusion and the regression model, we train the regression model containing the CDFformer. Note that the loss function, optimizer, and hyper-parameters are all the same as the conditional diffu", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/282b272ad671d6b404a3ed22635d1576a4bf78b78ed4702b1d69ab6d55912c09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 87, + 630, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f27e977efd2497c09bc9d01b49e33e674787290d6375d8be6eceb808e3a0453.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 756, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1510adeba85387e2a447338c741704477933bd78e2a337868bcdbd57d0f3ff6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 88, + 885, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20c5dfeab9f90b804383c1554725bda46e8f71ac8393518617942cfcb6fe1d5b.jpg", + "image_caption": [ + "(a) Regression 49.13dB" + ], + "image_footnote": [], + "bbox": [ + 506, + 193, + 630, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a61dda589a4d10190c537c36b990fc015b019689611ef15fff5d05fc466ea5e7.jpg", + "image_caption": [ + "(b) HSR-Diff 52.06dB" + ], + "image_footnote": [], + "bbox": [ + 635, + 193, + 756, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5345dfe7c76dd1556d43bdea406132e4e5a32129ba1d03d19c3260b88b8540a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 194, + 885, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2f269d52ba9006f83bc3eae5c1e1e1af5ba9a1047ba202316b82386499a68dd.jpg", + "image_caption": [ + "(c) Ground truth", + "Figure 9: Fusion results for HSR-Diff and Regression on the real and fake peppers image of CAVE. The PSNR values (dB) are given in the subtitles." + ], + "image_footnote": [], + "bbox": [ + 519, + 309, + 890, + 329 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46d8f9205ac589f5c0537c4d910eb49dd8ed1cdd8caaedb7b95fc8f8a51151ce.jpg", + "table_caption": [ + "Table 1: Averaged PSNR, SSIM, SAM, and ERGAS of compared methods on CAVE, PaviaU, and Chikusei datasets. The $\\uparrow$ or $\\downarrow$ indicates higher or lower values corresponding to better results." + ], + "table_footnote": [], + "table_body": "
DatasetModelPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVERegression43.260.99424.030.193
CDFformer44.330.99513.710.179
PaviaURegression46.030.99681.521.178
CDFformer46.470.99771.451.053
ChikuseiRegression56.710.99990.470.372
CDFformer57.340.99990.430.324
", + "bbox": [ + 501, + 400, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Ablation study on conditional diffusion models. The $\\uparrow$ or $\\downarrow$ indicates higher or lower values corresponding to better results.", + "bbox": [ + 498, + 527, + 890, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sion models. Figure 9 presents the fused results and corresponding error maps of utilising HSR-Diff and the regression model. As can be seen from the error maps, the HSIs produced by HSR-Diff have less distortion than those by the regression model. This is because HSR-Diff works with a series of iterative refinement steps, facilitating the capture of richer information on data distributions of HR-HSIs. Furthermore, Table 2 showcases the quantitative results and effectively highlights the significant contribution of iterative refinement.", + "bbox": [ + 496, + 603, + 890, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of CDFormer. Recall that CDFormer is conditioned on the hierarchical representations of HR-MSI and LR-HSI via a two-stream architecture. However, alternative outstanding diffusion models [14, 23] that are also excellent for conditional image generation are equipped with CNN-based U-Nets, where degenerated images are concatenated with noisy high-resolution output images. To show the effectiveness of hierarchical representations, we remove the", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7090", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d990b38633caae1cbcf582cb90a4f8439c06a8c69001841a19b491494f410ea.jpg", + "image_caption": [ + "(a) HSI" + ], + "image_footnote": [], + "bbox": [ + 81, + 99, + 194, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1e63519dc0dcfcefe9e053ccda826e92fdb28596669bcd6890c6f136a2c5cd56.jpg", + "image_caption": [ + "(b)UTV-TD" + ], + "image_footnote": [], + "bbox": [ + 197, + 99, + 310, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f61bb7027e411a835dfedcdfcc38e82a120285039a79cdc5dc30b0e183dc8b12.jpg", + "image_caption": [ + "(c) UAL" + ], + "image_footnote": [], + "bbox": [ + 313, + 99, + 426, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c4475d6dc385fe81cbc5c162b73b76066729cc52030d0135999066f24bcd5312.jpg", + "image_caption": [ + "(d) BRResNet", + "Figure 10: Visual fusion results of all competing methods for HypSen." + ], + "image_footnote": [], + "bbox": [ + 429, + 99, + 542, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3c78fd2b697fbf692f094d43862d354311e4dbb6ca987e0740d6591e974a4c1a.jpg", + "image_caption": [ + "(e) CMHF-Net" + ], + "image_footnote": [], + "bbox": [ + 545, + 99, + 656, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cfcc1b3e183daf06d3dd4fb15e64af6136d21ca99b601c1a960440c992268d7c.jpg", + "image_caption": [ + "(f) UAL-DMI" + ], + "image_footnote": [], + "bbox": [ + 661, + 99, + 772, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/24eb1dcfd33ad077d2b7cb7cbe93b8643a16cb381f4337c1ac7e63b26e6c717b.jpg", + "image_caption": [ + "(g) HSR-Diff" + ], + "image_footnote": [], + "bbox": [ + 777, + 99, + 888, + 186 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/38259561aeb098f1914201c8cae30dbd1450fc2c9ac89cf901f552b44010393a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNetworkPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVECD-CNN38.840.97977.320.318
C-w/o-SR43.740.99423.940.188
CDFformer44.330.99513.710.179
PaviaUCD-CNN42.750.99621.751.362
C-w/o-SR46.080.99761.471.080
CDFformer46.470.99771.451.053
ChikuseiCD-CNN47.630.99801.201.794
C-w/o-SR56.680.99990.460.425
CDFformer57.340.99990.430.324
", + "bbox": [ + 78, + 253, + 467, + 416 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/51cb88b3a9e887a212404cee8180b0d9e48181092e6f5263303b672aeb20bd1d.jpg", + "table_caption": [ + "Table 3: Ablation study on CDFormer. The $\\uparrow$ or $\\downarrow$ indicates higher or lower values corresponding to better results." + ], + "table_footnote": [], + "table_body": "
DatasetMethodsPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVEFixed43.170.99274.990.203
Progressive44.330.99513.710.179
PaviaUFixed45.060.99701.631.173
Progressive46.470.99771.451.053
ChikuseiFixed55.920.99990.500.453
Progressive57.340.99990.430.324
", + "bbox": [ + 78, + 474, + 470, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Ablation study on progressive learning. The $\\uparrow$ or $\\downarrow$ indicates higher or lower values corresponding to better results.", + "bbox": [ + 76, + 603, + 468, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SR stream of CDFormer and name the resulting network \"C-w/o-SR\". In addition, we compare CDFormer with a CNN version of CDFormer that replaces all $S^2$ TL with convolutional layers and show its results as \"CD-CNN\" in Table 3. The quantitative results show the use of CDFormer performs better than CD-CNN, demonstrating the effectiveness of global statistics. Indeed, with the two-stream architecture, CDFormer offers the best results thanks to the use of hierarchical features.", + "bbox": [ + 75, + 676, + 468, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of progressive learning. Progressive learning helps CDFormer to capture long-range dependencies of spatio-spectral information in HR-HSIs. To illustrate the effect of progressive learning, we train the CDFormer", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "with fixed patches (128² for CAVE and Chikusei; 64² for PaviaU) with the results shown under the heading of \"Fixed\" in Table 4. As can be seen, progressive learning (from 128² to 512² for CAVE and Chikusei; from 64² to 128² for PaviaU) provides better results than training with fixed patches.", + "bbox": [ + 496, + 255, + 890, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Generalization Analysis on Real Dataset", + "text_level": 1, + "bbox": [ + 498, + 356, + 844, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To examine the generalization ability of the implementations following the proposed approach, we test the performance of competitors on the real-world HypSen dataset [37]. Due to the lack of an ideal HR-HSI to train deep neural networks, we utilize the networks trained on the PaviaU dataset to merge observed LR-HSI and the corresponding HR-MSI. In addition, interpolation is applied to addressing the problem of an inconsistent number of bands between datasets. The fusion results of all compared methods are visualized in Figure 10, from which it can be seen that our method generates rich details, attaining satisfactory quality.", + "bbox": [ + 496, + 378, + 890, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 559, + 617, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have presented the novel HSR-Diff approach that initializes an HR-HSI with pure Gaussian noise and then, iteratively refines it subject to the condition of the LR-HSIs and HR-MSIs of interest. At each step, the noise is removed with CDFormer which exploits the hierarchical representations of HR-MSIs and LR-HSIs rather than the original images. In addition, we employ a progressive learning strategy to maximize the use of the global information of full-resolution images, where CDFormer is trained on small patches in the early epochs with high efficiency while on the global images in the later epochs to obtain the global statistics. Systematic experimental investigations have been conducted, on four public datasets to validate the superior performance of the proposed approach, in comparison with state-of-the-art methods. However, diffusion models have low inference efficiency due to their sequential nature, computational complexity, time-consuming sampling process, and large model size. For future work, we will try to resolve the challenging issue of the relatively low image-generation efficiency of HSR-Diff.", + "bbox": [ + 496, + 584, + 890, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7091", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naveed Akhtar, Faisal Shafait, and Ajmal Mian. Bayesian sparse representation for hyperspectral image super resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3631-3640, 2015. 1", + "[2] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Super-resolution for hyperspectral and multispectral image fusion accounting for seasonal spectral variability. IEEE Transactions on Image Processing, 29:116-127, 2019. 1", + "[3] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 5", + "[4] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In Proceedings of the International Conference on Learning Represent, 2021. 4, 5", + "[5] Zhao Chen, Hanye Pu, Bin Wang, and Geng-Ming Jiang. Fusion of hyperspectral and multispectral images: A novel framework based on generalization of pan-sharpening methods. IEEE Geoscience and Remote Sensing Letters, 11(8):1418-1422, 2014. 1", + "[6] Phuong D Dao, Kiran Mantripragada, Yuhong He, and Faisal Z Qureshi. Improving hyperspectral image segmentation by applying inverse noise weighting and outlier removal for optimal scale selection. ISPRS Journal of Photogrammetry and Remote Sensing, 171:348-366, 2021. 1", + "[7] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. Psrt: Pyramid shuffle-and-reshuffle transformer for multispectral and hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2", + "[8] Renwei Dian, Shutao Li, Anjing Guo, and Leyuan Fang. Deep hyperspectral image sharpening. IEEE transactions on neural networks and learning systems, 29(11):5345-5355, 2018. 2", + "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2", + "[10] Ying Fu, Tao Zhang, Yinqiang Zheng, Debing Zhang, and Hua Huang. Hyperspectral image super-resolution with optimized rgb guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2019. 2", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 8", + "[15] Jin-Fan Hu, Ting-Zhu Huang, Liang-Jian Deng, Hong-Xia Dou, Danfeng Hong, and Gemine Vivone. Fusformer: A transformer-based fusion network for hyperspectral image super-resolution. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2, 3, 6, 8", + "[16] Zi-Rong Jin, Liang-Jian Deng, Tian-Jing Zhang, and Xiao-Xu Jin. Bam: Bilateral activation mechanism for image fusion. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4315-4323, 2021. 3, 6, 8", + "[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2", + "[18] Shutao Li, Renwei Dian, Leyuan Fang, and José M Bioucas-Dias. Fusing hyperspectral and multispectral images via coupled sparse tensor factorization. IEEE Transactions on Image Processing, 27(8):4118-4130, 2018. 1", + "[19] Ying Li, Haokui Zhang, and Qiang Shen. Spectral-spatial classification of hyperspectral imagery with 3d convolutional neural network. Remote Sensing, 9(1):67, 2017. 2", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5", + "[21] Gamba Paolo. Pavia centre and university. https://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Sciences, 2011. 6", + "[22] Ying Qu, Hairong Qi, and Chiman Kwan. Unsupervised sparse dirichlet-net for hyperspectral image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2511-2520, 2018. 2", + "[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 8", + "[24] Yue Shi, Liangxiu Han, Lianghao Han, Sheng Chang, Tongle Hu, and Darren Dancey. A latent encoder coupled generative adversarial network (le-gan) for efficient hyperspectral image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2022. 2", + "[25] Dong Wang, Yunpeng Bai, Bendu Bai, Chanyue Wu, and Ying Li. Heterogeneous two-stream network with hierarchical feature prefusion for multispectral pan-sharpening. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1845-1849. IEEE, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7092", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Dong Wang, Yunpeng Bai, Chanyue Wu, Ying Li, Changjing Shang, and Qiang Shen. Convolutional LSTM-based hierarchical feature fusion for multispectral pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1", + "[27] Dong Wang, Pei Zhang, Yunpeng Bai, and Ying Li. Metapan: Unsupervised adaptation with meta-learning for multispectral pansharpening. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1", + "[28] Xiuheng Wang, Jie Chen, and Cedric Richard. Hyperspectral image super-resolution with deep priors and degradation model inversion. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2814-2818. IEEE, 2022. 6, 8", + "[29] Qi Wei, José Bioucas-Dias, Nicolas Dobigeon, and Jean-Yves Tourneret. Hyperspectral and multispectral image fusion based on a sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 53(7):3658-3668, 2015. 1", + "[30] Zhenyu Wu, Lin Wang, Wei Wang, Tengfei Shi, Chenglizhao Chen, Aimin Hao, and Shuo Li. Synthetic data supervised salient object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5557-5565, 2022. 2", + "[31] Qi Xie, Minghao Zhou, Qian Zhao, Zongben Xu, and Deyu Meng. Mhf-net: An interpretable deep network for multispectral and hyperspectral image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 6, 8", + "[32] Fengchao Xiong, Jun Zhou, and Yuntao Qian. Material based object tracking in hyperspectral videos. IEEE Transactions on Image Processing, 29:3719-3733, 2020. 1", + "[33] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2", + "[34] Ting Xu, Ting-Zhu Huang, Liang-Jian Deng, Xi-Le Zhao, and Jie Huang. Hyperspectral image superresolution using unidirectional total variation with tucker decomposition. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13:4381-4398, 2020. 6, 8", + "[35] Xizhe Xue, Haokui Zhang, Bei Fang, Zongwen Bai, and Ying Li. Grafting transformer on automatically designed convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1", + "[36] Jing Yang, Chanyue Wu, Tengfei You, Dong Wang, Ying Li, Changjing Shang, and Qiang Shen. Hierarchical spatio-spectral fusion for hyperspectral image super resolution via sparse representation and pre-trained deep model. Knowledge-Based Systems, 260:110170, 2023. 1", + "[37] Jingxiang Yang, Yong-Qiang Zhao, and Jonathan Cheung-Wai Chan. Hyperspectral and multispectral image fusion via deep two-branches convolutional neural network. Remote Sensing, 10(5):800, 2018. 6, 9", + "[38] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE transactions on image processing, 19(9):2241-2253, 2010.6", + "[39] Naoto Yokoya and Akira Iwasaki. Airborne hyperspectral data over chikusei. Space Appl. Lab., Univ. Tokyo, Tokyo, Japan, Tech. Rep. SAL-2016-05-27, 5, 2016. 6", + "[40] Tengfei You, Chanyue Wu, Yunpeng Bai, Dong Wang, Huibin Ge, and Ying Li. Hmf-former: Spatio-spectral transformer for hyperspectral and multispectral image fusion. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2022. 2, 3", + "[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5", + "[42] Haokui Zhang, Chengrong Gong, Yunpeng Bai, Zongwen Bai, and Ying Li. 3-d-anas: 3-d asymmetric neural architecture search for fast hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021. 1", + "[43] Lei Zhang, Jiangtao Nie, Wei Wei, Yanning Zhang, Shengcai Liao, and Ling Shao. Unsupervised adaptation learning for hyperspectral imagery super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2020. 2, 3, 6, 8", + "[44] Meilin Zhang, Xiongli Sun, Qiqi Zhu, and Guizhou Zheng. A survey of hyperspectral image super-resolution technology. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, pages 4476-4479. IEEE, 2021. 1", + "[45] Xueting Zhang, Wei Huang, Qi Wang, and Xuelong Li. Ssrnet: Spatial-spectral reconstruction network for hyperspectral and multispectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5953-5965, 2020. 3", + "[46] Ke Zheng, Lianru Gao, Wenzhi Liao, Danfeng Hong, Bing Zhang, Xin Cui, and Jocelyn Chanussot. Coupled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super resolution. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2487-2502, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 667 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7093", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_model.json b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_model.json new file mode 100644 index 0000000000000000000000000000000000000000..04c28f66c167788b7e69aa0d0daf480e134768f7 --- /dev/null +++ b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_model.json @@ -0,0 +1,3313 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.13, + 0.87, + 0.175 + ], + "angle": 0, + "content": "HSR-Diff: Hyperspectral Image Super-Resolution via Conditional Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.203, + 0.816, + 0.222 + ], + "angle": 0, + "content": "Chanyue Wu\\(^{1*}\\), Dong Wang\\(^{1*}\\), Yunpeng Bai\\(^{2*}\\), Hanyu Mao\\(^{1}\\), Ying Li\\(^{1\\dagger}\\), Qiang Shen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.222, + 0.872, + 0.274 + ], + "angle": 0, + "content": "\\(^{1}\\)National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, Shaanxi Provincial Key Laboratory of Speech & Image Information Processing, School of Computer Science, Northwestern Polytechnical University, China" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.275, + 0.847, + 0.309 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science, Faculty of Business and Physical Sciences, Aberystwyth University, U.K." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.311, + 0.832, + 0.328 + ], + "angle": 0, + "content": "{chanyuewu, dongwang, maomhy} \\(@\\) mail.nwpu.edu.cn lybyp@nwpu.edu.cn {yub3, qqs} \\(@\\) aber.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Despite the proven significance of hyperspectral images (HSIs) in performing various computer vision tasks, its potential is adversely affected by the low-resolution (LR) property in the spatial domain, resulting from multiple physical factors. Inspired by recent advancements in deep generative models, we propose an HSI Super-resolution (SR) approach with Conditional Diffusion Models (HSR-Diff) that merges a high-resolution (HR) multispectral image (MSI) with the corresponding LR-HSI. HSR-Diff generates an HR-HSI via repeated refinement, in which the HR-HSI is initialized with pure Gaussian noise and iteratively refined. At each iteration, the noise is removed with a Conditional Denoising Transformer (CDFformer) that is trained on denoising at different noise levels, conditioned on the hierarchical feature maps of HR-MSI and LR-HSI. In addition, a progressive learning strategy is employed to exploit the global information of full-resolution images. Systematic experiments have been conducted on four public datasets, demonstrating that HSR-Diff outperforms state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.21, + 0.726 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.843 + ], + "angle": 0, + "content": "Hyperspectral images (HSIs) contain dozens or hundreds of spectral bands, enabling them to provide more faithful knowledge of targeted scenes than conventional imaging modalities. As such, HSIs play an irreplaceable role in various computer vision tasks, including classification [35, 42], segmentation [6], and tracking [32]. Although HSIs contain rich spectral information, contemporary hyperspectral" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.361, + 0.857, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.5, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Figure 1: Illustration of fusion-based Hyperspectral Image Super-Resolution (HSI-SR)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.893, + 0.673 + ], + "angle": 0, + "content": "imaging sensors lack high-resolution (HR) in the spatial domain, due to the stringent constraint of typically low signal-to-noise ratios. Their widespread use is significantly hindered by this fact. Restricted by hardware limitations, a practical way to work around this problem is to fuse the low-resolution (LR) HSI with an HR multispectral image (MSI). This requires the implementation of so-called HSI super-resolution (SR)[36], as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Over the past few decades, a significant amount of research efforts have been devoted to developing HSI-SR approaches, which can be roughly classified into five categories [44]: Extensions of pansharpening [5, 25, 27], Bayesian inference-based methods [1, 29], matrix factorization-based methods [2], tensor-based methods [18], and deep learning (DL)-based methods [31]. Whilst pansharpening methods [26] have been extended to the field of HSI-SR, such approaches are prone to spectral distortion. Bayesian inference-based approaches rely on the assumption of prior knowledge, thereby having a weak flexibility in dealing with different HSI structures. Matrix factorization-based techniques reshape the 3D HSIs and MSIs into matrices, thus facing the challenge of learning the required relationship between space and spectrum. Although sev-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.852, + 0.204, + 0.864 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.864, + 0.469, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author. This work is supported by the National Natural Science Foundation of China under Grant (62271400), and in part by the Shaanxi Provincial Key R&D Program, China (2023-GHZD-02)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.852, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7083" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "eral tensor-based methods have been proposed that can maintain the 3D structure of input images, they consume much more memory and computational power. Furthermore, these traditional approaches work via relying heavily on hand-crafted priors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.472, + 0.413 + ], + "angle": 0, + "content": "Recently, DL-based methods, especially convolutional neural network (CNN)-based approaches, have flooded over into the HSI-SR research community [8, 31, 10, 22, 46, 43, 19]. Rather than resorting to hand-crafted features, DL-based techniques learn prior knowledge automatically from given data. Particularly, Dong et al. proposed the first DL-based method for image SR, with the end-to-end mapping between LR images and HR images learned using a CNN [9]. Subsequently, generative adversarial networks (GANs) were introduced to the field of image SR in an effort to produce high-frequency details [17, 11]. After that, various GAN-based models have been devised, showing state-of-the-art results in the HSI-SR literature [24]. However, such work requires carefully designed regularization and optimization tricks to tame optimization instability and avoid mode collapse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.472, + 0.642 + ], + "angle": 0, + "content": "Inspired by the recent developments in deep generative models [14, 23], in this paper, we propose an innovative approach that we refer to as HSR-Diff (HSI-SR with conditional diffusion models). It works by learning to transform the original standard normal distribution into the data distribution of HR-HSI through a sequence of refinements. In contrast to GAN-based methods which require inner-loop maximization, HSR-Diff operates by simply minimizing a well-defined loss function to learn prior knowledge. Although conditional diffusion models are straightforward to define and efficient to train, there has been no demonstration that they are capable of merging LR-HSI and HR-MSI to the best of our knowledge. We show that conditional diffusion models are capable of generating high-quality HR-HSIs, which may best the state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "A key factor of HSR-Diff is its inherent denoising ability thanks to use of deep neural networks. In spite of the effectiveness of CNNs for denoising, they have shown limitations in modelling long-range dependencies. Some previous transformer-based HSI-SR works, especially Pyramid Shuffle-and-Reshuffle Transformer (PSRT) [7], Transformer-based Fusion network (Fusformer) [15], and Hyperspectral and Multispectral image Fusion Transformer (HMF-Former) [40], all of which learn long-range dependencies with attention mechanisms. These Transformers only take LR-HSI and HR-MSI as input, while the denoising Transformer is also fed with noise level \\(\\gamma\\) and noisy HR-HSI. To comprehensively exploit the spatial and spectral information in the LR-HSI, HR-MSI, and noisy HR-HSI, a Conditional Denoising Transformer (CD-Former) is herein designed and trained with a denoising objective to remove various levels of noise iteratively. In ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "dition, a progressive learning strategy is utilized to help the CDFormer learn the global statistics of full-resolution HSIs. The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.166, + 0.892, + 0.271 + ], + "angle": 0, + "content": "- We propose the novel application of conditional diffusion models in the field of HSI-SR that works by progressively destroying HR-HSI through injecting noise and subsequently learning to reverse this process, in order to perform HSI-SR. To the best of our knowledge, this is the first work to exploit diffusion models within the realm of HSI-SR." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.284, + 0.892, + 0.389 + ], + "angle": 0, + "content": "- We introduce a CDFormer that refines a noisy HR-HSI conditioned on the deep feature maps of HR-MSI and LR-HSI, capable of modelling global connectivity with self-attention. Rather than concatenation, cross-attention is utilized as the conditioning mechanism to incorporate spatio-spectral information and noise level." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.402, + 0.892, + 0.477 + ], + "angle": 0, + "content": "- We employ a progressive learning strategy to exploit the global information of full-resolution HSIs, with CDFormer being trained on small image patches in the early epochs with high efficiency and on the global images in the later epochs to acquire global information." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.489, + 0.892, + 0.549 + ], + "angle": 0, + "content": "- We present experimental investigations on four public datasets, with quantitative and qualitative results illustrating the superior performance of our approach as compared with state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.166, + 0.892, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.642, + 0.58 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.667, + 0.605 + ], + "angle": 0, + "content": "2.1. Diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Typical deep generative models include autoregressive models (AR), normalizing flows (NF), variational autoencoders (VAE), GANs, diffusion models, etc. Over the past decade, GANs have gained prominence as a pivotal technology in the challenging field of image synthesis. However, this landscape has seen a transformation with the emergence of diffusion models, which have rapidly become a prominent research avenue across diverse domains, such as natural language processing, image processing, and computer vision. Diffusion models have arisen as an innovative approach that operates by iteratively refining a noise vector into the desired output. This reversible process allows for the generation of high-quality samples and has proven versatile in various applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The development of diffusion models has seen a dramatically accelerating pace over the past three years. In the field of computer vision, diffusion models have shown promise in tasks like image segmentation [33], object detection [30], and even video processing [13]. Whilst diffusion models" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7084" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.088, + 0.473, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.165, + 0.47, + 0.227 + ], + "angle": 0, + "content": "Figure 2: Forward and reverse processes of HSR-Diff, with forward process \\( q \\) generating an HSI sequence (left to right) by gradually adding Gaussian noise, and reverse process \\( p \\) iteratively refining HR-HSI (right to left)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.471, + 0.33 + ], + "angle": 0, + "content": "have shown great potential for a variety of computer vision applications, none of them have yet been devoted to the problem of HSI-SR to the best of our knowledge. In this paper, we extend the utility of diffusion models to the field of HSI-SR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.341, + 0.345, + 0.358 + ], + "angle": 0, + "content": "2.2. Deep Learning-Based HSI-SR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.682 + ], + "angle": 0, + "content": "In recent years, data-driven CNN architectures have been shown to outperform traditional approaches for use in the HSI-SR literature. These methods formulate the underlying fusion problem as a highly nonlinear mapping that takes HR-MSIs and LR-HSIs as input to generate an optimal HR-HSI. For example, CMHF-net [31] is an interpretable CNN, the design of which exploits the deep unfolding technique. Zhang et al. [43] proposed to reconstruct HR-HSIs with a two-stage network, while Zhang et al. [45] designed an interpretable spatial-spectral reconstruction network (SSR-NET) based on CNN. Aiming at problems of inflexible structure and information distortion, Jin et al. embedded Bilateral Activation Mechanism into ResNet, resulting in the effective model of BRResNet [16]. Thanks to the inductive bias of CNN, such as locality and weight sharing, these methods can provide good generalization performance and achieve impressive results. Nevertheless, CNNs have limitations in capturing long-range dependencies and self-similarity priors. To overcome such shortcomings, some Transformer-based HSI-SR approaches [15, 40] have been proposed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Within this study, the CDFormer serves as the denoising network in conditional diffusion models. It harnesses self-attention to learn global statistics from full-resolution HSIs and MSIs and employs cross-attention as the conditioning mechanism." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.775, + 0.295, + 0.792 + ], + "angle": 0, + "content": "3. Proposed Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.281, + 0.815 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.868 + ], + "angle": 0, + "content": "Without losing generality, the observation models for the HR-MSI and LR-HSI of interest can be mathematically formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.868, + 0.469, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = \\mathbf {R Z} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.239, + 0.89, + 0.309, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\mathbf {Z D},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.895, + 0.304 + ], + "angle": 0, + "content": "where \\(\\mathbf{X} \\in \\mathbb{R}^{b \\times HW}\\) denotes the HR-MSI which consists of \\(b\\) spectral bands with a resolution of \\(HW\\) in the spatial domain; \\(\\mathbf{R} \\in \\mathbb{R}^{b \\times B}\\) represents the spectral response function of HR-MSI; \\(\\mathbf{Y} \\in \\mathbb{R}^{B \\times hw}\\) denotes the LR-HSI; and \\(\\mathbf{Z} \\in \\mathbb{R}^{B \\times HW}\\) is the HR-HSI. In the above, \\(b\\) and \\(B\\) are the numbers of bands, with \\(h\\) and \\(H\\) being the band height, and \\(w\\) and \\(W\\) the width, where \\(b \\ll B\\), \\(h \\ll H\\), and \\(w \\ll W\\). \\(\\mathbf{D} \\in \\mathbb{R}^{HW \\times hw}\\) is the spatial response of the LR-HSI, which can be modelled with blurring and down-sampling operations. The HSI-SR can be interpreted as an inverse problem for merging a practically collected \\(\\mathbf{X}\\) and an observed \\(\\mathbf{Y}\\) to produce a latent \\(\\mathbf{Z}\\). In this paper, the ideal \\(\\mathbf{Z}\\) is restored with HSR-Diff conditioned on spatio-spectral information of \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\), the details of which are described below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.315, + 0.866, + 0.331 + ], + "angle": 0, + "content": "3.2. HSI-SR with Conditional Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Given a dataset \\(\\mathcal{D}_{train} = \\{\\mathbf{X}^i,\\mathbf{Y}^i,\\mathbf{Z}^i\\}_{i = 1}^N\\) satisfying a certain joint probability distribution \\(p(\\mathbf{X},\\mathbf{Y},\\mathbf{Z})\\), many pairs of \\((\\mathbf{X},\\mathbf{Y})\\) may be consistent with the same \\(\\mathbf{Z}\\). Thus, the HR-HSI \\(\\mathbf{Z}\\) can be obtained with iterative refinement that provide an approximate to \\(p(\\mathbf{Z}|\\mathbf{X},\\mathbf{Y})\\). In this work, we implement the process of iterative refinement with HSR-Diff, where the optimized HR-HSI is presumed to be produced in \\(T\\) refinement steps. In HSR-Diff, the target HR-HSI is initialized with a pure noise \\(\\mathbf{Z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) as shown in Figure 2. The HSI is then refined iteratively according to learned conditional distributions \\(p_{\\theta}(\\mathbf{Z}_{t - 1}|\\mathbf{Z}_t,\\mathbf{X},\\mathbf{Y})\\). In so doing, the image sequence \\((\\mathbf{Z}_{T - 1},\\mathbf{Z}_{T - 2},\\dots ,\\mathbf{Z}_0)\\) can be attained and ultimately \\(\\mathbf{Z}_0\\sim p(\\mathbf{Z}|\\mathbf{X},\\mathbf{Y})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.704 + ], + "angle": 0, + "content": "The HSR-Diff makes use of two processes: a forward process that perturbs HR-HSI \\(\\mathbf{Z}_0\\) to noise, and a reverse process converting noise back to HR-HSI \\(\\mathbf{Z}_0\\). In the forward process, the intermediate images, i.e., \\(\\mathbf{Z}_{T - 1}\\), \\(\\mathbf{Z}_{T - 2}\\), \\(\\dots\\), and \\(\\mathbf{Z}_1\\), are generated according to a Markov chain with fixed transition probability \\(q(\\mathbf{Z}_t|\\mathbf{Z}_{t - 1})\\). We are interested in reversing the process via iterative refinement, in which the noise is reduced iteratively with a reverse Markov chain conditioned on \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\). The reverse chain is learned with the CDFformer \\(f_{\\theta}\\). Further details of HSR-Diff's working are given below." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.728, + 0.673, + 0.742 + ], + "angle": 0, + "content": "3.2.1 Forward Process" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.753, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Inspired by [14], forward process \\(q\\) iteratively adds Gaussian noise to \\(\\mathbf{Z}_0\\) over \\(T\\) iterations:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.796, + 0.892, + 0.837 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {Z} _ {1: T} \\mid \\mathbf {Z} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {t - 1}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.84, + 0.854, + 0.858 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {Z} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha_{1:T} \\in (0,1)\\) are scalar hyper-parameters. Note that in the forward process, the distribution of \\(\\mathbf{Z}_t\\) given \\(\\mathbf{Z}_0\\) can" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7085" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.885, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.354, + 0.69, + 0.371 + ], + "angle": 0, + "content": "Figure 3: Architecture of Conditional Denoising Transformer." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.429, + 0.412 + ], + "angle": 0, + "content": "be directly sampled in closed form. This implies that" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.428, + 0.47, + 0.447 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t}; \\sqrt {\\gamma_ {t}} \\mathbf {Z} _ {0}, (1 - \\gamma_ {t}) \\mathbf {I}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.462, + 0.47, + 0.495 + ], + "angle": 0, + "content": "where \\(\\gamma_{t} = \\prod_{i = 1}^{t}\\alpha_{i}\\). In addition, the posterior distribution of \\(\\mathbf{Z}_{t - 1}\\) given \\(\\mathbf{Z}_0\\) and \\(\\mathbf{Z}_t\\) can be derived by" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.508, + 0.376, + 0.527 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {0}, \\mathbf {Z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t - 1}; \\boldsymbol {\\mu}, \\sigma^ {2} \\mathbf {I}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.529, + 0.469, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} = \\frac {\\sqrt {\\gamma_ {t - 1}} (1 - \\alpha_ {t})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {0} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\gamma_ {t - 1})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {t} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.564, + 0.303, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {2} = \\frac {(1 - \\gamma_ {t - 1}) (1 - \\alpha_ {t})}{1 - \\gamma_ {t}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.61, + 0.47, + 0.642 + ], + "angle": 0, + "content": "As the CDFormer is tasked with predicting \\(\\mathbf{Z}_0\\), the posterior is useful in the reverse process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.325, + 0.687 + ], + "angle": 0, + "content": "3.2.2 Reverse Markovian Process" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.7, + 0.469, + 0.746 + ], + "angle": 0, + "content": "The reverse process inferences \\(\\mathbf{Z}_0\\) via iterative refinement. It starts from a pure Gaussian noise \\(\\mathbf{Z}_T\\) and goes in the opposite direction of the forward process:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.762, + 0.454, + 0.803 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {Z} _ {0: T} \\mid \\mathbf {X}, \\mathbf {Y}\\right) = p \\left(\\mathbf {Z} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {t}, \\mathbf {X}, \\mathbf {Y}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.805, + 0.327, + 0.822 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {Z} _ {T}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {T}; \\mathbf {0}, \\mathbf {I}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.825, + 0.473, + 0.855 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {t}, \\mathbf {X}, \\mathbf {Y}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {X}, \\mathbf {Y}, \\mathbf {Z} _ {t}, \\gamma_ {t}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where the distribution \\(p_{\\theta}(\\mathbf{Z}_{t - 1}\\mid \\mathbf{Z}_t,\\mathbf{X},\\mathbf{Y})\\) is parameterized with \\(\\theta\\). Note that the CDFormer provides a prediction of \\(\\hat{\\mathbf{Z}}_0\\). Thus, according to (4), each refinement step takes" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.397, + 0.634, + 0.412 + ], + "angle": 0, + "content": "the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.423, + 0.891, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Z} _ {t - 1} = \\frac {\\sqrt {\\gamma_ {t - 1}} (1 - \\alpha_ {t})}{1 - \\gamma_ {t}} f _ {\\theta} (\\mathbf {X}, \\mathbf {Y}, \\mathbf {Z} _ {t}, \\gamma_ {t}) \\tag {6} \\\\ + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\gamma_ {t - 1})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {t} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.502, + 0.799, + 0.518 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and \\(f_{\\theta}\\) is the CDFormer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.661, + 0.553 + ], + "angle": 0, + "content": "3.2.3 Noise Schedule" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.563, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Inspired by the research reported in [4], we sample \\(\\gamma\\) with two steps during training. In particular, we first sample a time step \\(t\\sim U\\{1,T\\}\\) and then randomly select \\(\\gamma \\sim U(\\gamma_{t - 1},\\gamma_t)\\). As such, \\(\\gamma \\sim p(\\gamma) = \\sum_{t = 1}^{T}\\frac{1}{T} U(\\gamma_{t - 1},\\gamma_t)\\). Normally, the model with a large \\(T\\) can achieve better results. However, we find (through empirical investigations) that the performance is not very sensitive to the exact values of \\(T\\). Therefore, no hyper-parameter search about \\(T\\) is conducted and we set \\(T = 2000\\) for simplicity. As for the inference process, we set the maximum generation iterations to 100, employing a linear noise schedule." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.81, + 0.758 + ], + "angle": 0, + "content": "3.3. Conditional Denoising Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The property of non-local self-similarity of HSIs is often exploited in denoising tasks but is usually not well captured by CNN-based models. Due to the effectiveness of Transformer layer in capturing non-local long-range dependencies, the potential of Transformer is explored in conditional denoising of HSI. Unfortunately, the vanilla Transformer focuses only on spatial relationships between pixels while neglecting the spectral dimension. Besides, denoising networks in conditional diffusion models normally concatenate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7086" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.089, + 0.44, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.181, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Figure 4: Spatio-Spectral Transformer Layer, where \"LN\" denotes layer normalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.24, + 0.468, + 0.316 + ], + "angle": 0, + "content": "all images together as input, which may hinder the extraction of useful spatio-spectral information in LR-HSIs and HR-MSIs. Hence, the CDFormer adopts a two-stream architecture and is constructed with stacked Spatio-Spectral Transformer Layers (S\\(^2\\)TLs)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.473, + 0.53 + ], + "angle": 0, + "content": "The architecture of the CDFormer is shown in Figure 3. The SR stream first utilizes a \\(3 \\times 3\\) convolution to generate low-level feature embeddings \\(\\mathbf{F}_0^{SR}\\) and then transforms it into deep features \\(\\mathbf{F}_l^{SR}\\) with stacked-\\(\\mathbf{S}^2\\mathbf{TLs}\\). Instead of adopting \\(t\\) as done in the existing work [4], our method is conditioned on \\(\\gamma\\) directly to achieve efficient generation. The denoising stream contains multiple noise-aware conditional \\(\\mathbf{S}^2\\mathbf{TLs}\\) (NC-\\(\\mathbf{S}^2\\mathbf{TLs}\\)) that take as input the embedded noise level and the image representation \\(\\mathbf{F}^{SR}\\). The Reconstruction module is set to produce a noise-free HR-HSI, by employing residual learning to alleviate the difficulty of HR-HSI generation while mapping the features onto HR-HSI via a \\(3 \\times 3\\) convolution and addition operation. It should be noted that no downsampling operation is employed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.552, + 0.296, + 0.568 + ], + "angle": 0, + "content": "3.3.1 Noise Level Embedding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.653 + ], + "angle": 0, + "content": "The noise level offers essential information for denoising models. Inspired by the work of [4], we embed noise level within the models with sinusoidal positional encoding. The process of noise level embedding (NLE) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.664, + 0.469, + 0.697 + ], + "angle": 0, + "content": "\\[\nN L E _ {\\gamma , 2 i} = \\sin \\left(\\gamma / 1 0 0 0 0 ^ {2 i / C}\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.692, + 0.395, + 0.719 + ], + "angle": 0, + "content": "\\[\nN L E _ {\\gamma , 2 i + 1} = \\cos \\left(\\gamma / 1 0 0 0 0 ^ {2 i / C}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.466, + 0.747 + ], + "angle": 0, + "content": "where \\(C\\) is the number of channels of \\(\\mathbf{S}^2\\mathrm{TLs}\\) . \\(i\\in [1,C / 2]\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.385, + 0.787 + ], + "angle": 0, + "content": "3.3.2 Spatio-Spectral Transformer Layers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Figure 4 illustrates the architecture of one \\(\\mathrm{S}^2\\mathrm{TL}\\), which consists of a Spatial Multi-head Self-Attention (SpatioMSA), a Spectral Multi-head Self-Attention (SpectralMSA), and a Feed Forward Network (FFN). SpatioMSA and SpectralMSA learn the interactions of spatial regions and interspectra relationships, respectively. To alleviate the computational burden, we adopt the transposed attention [41] in" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.088, + 0.83, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.193, + 0.892, + 0.252 + ], + "angle": 0, + "content": "Figure 5: Noise-Aware Conditional Spatio-Spectral Transformer Layer, where \"LT\" represents linear transform, \"LN\" denotes layer normalization, and \"MCA\" is the abbreviation of multi-head cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.28, + 0.892, + 0.341 + ], + "angle": 0, + "content": "SpectralMSA. SpatioMSA applies the popular window partitioning strategy [20] to reduce the computational complexity. In addition, the gating mechanism [41] is employed in the implementation of FFN." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.779, + 0.374 + ], + "angle": 0, + "content": "3.3.3 Noise-Aware Conditional \\(\\mathbf{S}^2\\mathbf{T}\\mathbf{L}\\mathbf{s}\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.384, + 0.892, + 0.552 + ], + "angle": 0, + "content": "To condition the overall model on the hierarchical features of SR stream, we feed \\(\\mathbf{F}_l^{SR}\\) to the Noise-Aware Conditional S\\(^2\\)TLs (NC-S\\(^2\\)TLs), each of which is a key building block of the denoising stream. Figure 5 depicts the structure of an NC-S\\(^2\\)TL, which takes as input the NLE (a vector), \\(\\mathbf{F}_l^{SR}\\) and \\(\\mathbf{F}_l^{DS}\\), where \\(\\mathbf{F}_l^{SR}\\) and \\(\\mathbf{F}_l^{DS}\\) have the same spatial resolution. NLE is first transformed and merged with \\(\\mathbf{F}_l^{DS}\\) with the result subsequently processed with the means of multi-head cross-attention (MCA) [3] in order to condition the model on \\(\\mathbf{F}_l^{SR}\\). As a result, each S\\(^2\\)TL learns the spatio-spectral dependencies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.768, + 0.577 + ], + "angle": 0, + "content": "3.4. Progressive Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Deep learning-based HSI-SR models are normally trained on small cropped patches. However, on one hand, training the proposed CDFormer on fixed-size image patches may not appropriately reflect the global relationships, which may result in suboptimal performance on full-resolution images when used. On the other hand, during the training phase, CDFormer can not be uploaded to Graphics Processing Unit (GPU) because of the limited memory. With this objective in mind, we adopt a strategy of progressive learning. In the initial epochs, the network is trained using smaller image patches. Subsequently, as training progresses, we transition towards utilizing larger patches, and ultimately, full-resolution images in the concluding training epochs. The resulting model, cultivated through the integration of varying sizes via progressive learning, exhibits notable enhancements in performance during testing. This is particularly evident when handling images of diverse resolutions, a scenario frequently encountered in the realm of the HSI-SR task." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To reduce the pressure on the demand of GPU memory, we only train the second half of CDFormer on full" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7087" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "resolution images. The loss function used for such training is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.133, + 0.468, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\| \\mathbf {X} - \\mathbf {R} \\hat {\\mathbf {Z}} _ {0} \\| _ {1} + \\| \\mathbf {Y} - \\hat {\\mathbf {Z}} _ {0} \\mathbf {D} \\| _ {1} + \\| \\mathbf {Z} _ {0} - \\hat {\\mathbf {Z}} _ {0} \\| _ {1} \\tag {8} \\\\ \\hat {\\mathbf {Z}} _ {0} = f _ {\\theta} (\\sqrt {\\gamma} \\mathbf {Z} _ {0} + \\sqrt {1 - \\gamma} \\boldsymbol {\\epsilon}, \\mathbf {X}, \\mathbf {Y}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.468, + 0.26 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \\((\\mathbf{X},\\mathbf{Y},\\mathbf{Z})\\) is sampled from the training set, and the noise schedule about \\(\\gamma\\) has been discussed above. The first two terms are designed according to the observation models, while the last one is based on the assumption of Laplace distribution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.276, + 0.209, + 0.292 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.468, + 0.348 + ], + "angle": 0, + "content": "Systematic experiments are herein conducted on four commonly-used public-available HSI-SR datasets to demonstrate the effectiveness of the proposed approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.179, + 0.373 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.468, + 0.428 + ], + "angle": 0, + "content": "Four datasets including CAVE [38], PaviaU [21], Chikusei [39], and HypSen [37] are used in our experiments, with the following details on each." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.468, + 0.57 + ], + "angle": 0, + "content": "CAVE: There are 32 scenes with a spatial size of \\(512 \\times 512\\) in the CAVE dataset, where we select the first 20 HSIs for training, with the remaining 12 images used for testing. We generate LR-HSIs by Gaussian blur and down-sampling using a factor of 32 as done in [31]. HR-MSIs are acquired by integrating all HR-HSI bands according to the spectral response function of Nikon D700. The original HR-HSIs are treated as ground truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.468, + 0.712 + ], + "angle": 0, + "content": "PaviaU: Collected by the University of Pavia, Italy, the original HSI dataset consists of \\(610 \\times 340\\) pixels in which the top-left \\(128 \\times 128\\) area is extracted as the test data, with the remaining used for training. Except for water absorption bands, all other 103 bands are chosen for the experiments. Note that the down-sampling factor for the generation of LR-HSIs is four, and the spectral response function is the same as that of the WorldView-3 satellite." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Chikusei: This dataset consists of 128 bands with a spectral range of \\(363~nm\\) to \\(1018~nm\\) and a spatial resolution of \\(2517\\times 2335\\). The original HSI data was taken by an airborne visible and near-infrared imaging sensor over Chikusei, Japan. To alleviate the impact of the back boundary and noise, we crop the center area and remove noise bands. The processed image has a size of \\(2048\\times 2048\\times 110\\). The top half \\(1024\\times 2048\\times 110\\) area is selected as the training data, while the rest half is split into eight testing \\(512\\times 512\\times 110\\) patches. For the production of LR-HSIs and HR-MSIs, this dataset adopts the same processing as with PaviaU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "HypSen: This dataset concerns a real scenario consisting of a \\(30\\mathrm{m}\\)-resolution HSI and a \\(10\\mathrm{m}\\)-resolution MSI. The Hyperion sensor on the Earth Observing-1 satellite provided the HSI with 242 spectral bands in the spectral range of \\(400\\mathrm{nm}\\) to \\(2500\\mathrm{nm}\\), and the MSI with 13 bands was captured by the Sentinel-2A satellite. The blue, green, red, and near-infrared bands of MSI in our experiments are selected due to their high spatial resolution. To eliminate the impact of noise and water absorption, we remove those relevant bands, with 84 bands remaining in the HSI. We crop sub-images of size \\(250\\times 330\\) and \\(750\\times 990\\) from the Hyperion HSI and Sentinel-2A MSI respectively, in our study, with the pairs of sub-image patches spatially registered." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.297, + 0.892, + 0.328 + ], + "angle": 0, + "content": "4.2. Methods Compared and Evaluation Metrics Used" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Six state-of-the-art HSI-SR approaches are taken for comparison, including: UTV-TD [34], UAL [43], BRResNet [16], Fusformer [15], CMHF-Net [31], and UAL-DMI [28]. UTV-TD is a tensor-based technique; UAL, BRResNet, Fusformer, and CMHF-Net fall into the category of the DL-based methods; and UAL-DMI can be regarded as an upgraded version of UAL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.547 + ], + "angle": 0, + "content": "Four quantitative quality metrics are employed for performance evaluation, including Peak Signal-to-Noise Ratio (PSNR), Spectral Angle Mapper (SAM), Erreur Relative Globale Adimensionnelle de Synthese (ERGAS, namely error relative global dimensionless synthesis), and Structure Similarity (SSIM). The smaller ERGAS and SAM are, the larger PSNR and SSIM are, the better the fusion result is." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.558, + 0.767, + 0.573 + ], + "angle": 0, + "content": "4.3. Implementation Specification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.807 + ], + "angle": 0, + "content": "All DL-based methods are trained on the same datasets. For those compared methods, we use the publicly available source codes with default hyper-parameters as given in the corresponding research papers. Our HSR-Diff is implemented on the PyTorch framework. The learnable parameters of the CDFformer are initialized with Kaiming initialization [12] and trained on 2 NVIDIA GeForce GTX 3090s. The dimension of the embedding of \\( F_0^{SR} \\) and \\( F_0^{DS} \\) is set to 256, and the number of parameters of the CDFformer is 31.56M. We utilize the Adam optimizer with \\( \\beta_{1} = 0.9 \\) and \\( \\beta_{2} = 0.999 \\) to optimize the CDFformer. With limited GPU memory, the batch size is set to 4 and 2 for \\( 128^2 \\) and \\( 512^2 \\) images, respectively. It costs 20000 epochs on the CAVE and PaviaU datasets while consuming 5000 epochs on the Chikusei dataset. The learning rate is initialized as \\( 1 \\times 10^{-4} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.817, + 0.87, + 0.833 + ], + "angle": 0, + "content": "4.4. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In this set of experiments, the evaluations are carried out using the first three datasets listed above without involving the real-world dataset, HypSen (which will be dealt with in the next section)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.088, + 0.181, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.089, + 0.281, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.382, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.089, + 0.483, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.585, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.089, + 0.685, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.089, + 0.787, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.089, + 0.888, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.174, + 0.18, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.252, + 0.166, + 0.264 + ], + "angle": 0, + "content": "(a)UTV-TD" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.174, + 0.281, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.252, + 0.256, + 0.264 + ], + "angle": 0, + "content": "(b) UAL" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.174, + 0.382, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.252, + 0.372, + 0.264 + ], + "angle": 0, + "content": "(c) BRResNet" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.174, + 0.483, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.252, + 0.476, + 0.264 + ], + "angle": 0, + "content": "(d) CMHF-Net" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.174, + 0.585, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.252, + 0.574, + 0.264 + ], + "angle": 0, + "content": "(e) Fusformer" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.174, + 0.686, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.252, + 0.675, + 0.264 + ], + "angle": 0, + "content": "(f) UAL-DMI" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.174, + 0.787, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.252, + 0.776, + 0.264 + ], + "angle": 0, + "content": "(g) HSR-Diff" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.174, + 0.888, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.252, + 0.884, + 0.264 + ], + "angle": 0, + "content": "(h) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.268, + 0.71, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.303, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Figure 6: Visual quality comparison for fused HSIs of all competing methods on CAVE, where first and second rows show fourth band and corresponding heatmaps (mean squared error), respectively." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.347, + 0.18, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.348, + 0.281, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.348, + 0.382, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.348, + 0.483, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.348, + 0.585, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.348, + 0.685, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.348, + 0.787, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.348, + 0.888, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.434, + 0.18, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.511, + 0.166, + 0.523 + ], + "angle": 0, + "content": "(a)UTV-TD" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.434, + 0.281, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.511, + 0.256, + 0.523 + ], + "angle": 0, + "content": "(b) UAL" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.434, + 0.382, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.511, + 0.372, + 0.523 + ], + "angle": 0, + "content": "(c) BRResNet" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.434, + 0.483, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.511, + 0.475, + 0.523 + ], + "angle": 0, + "content": "(d) CMHF-Net" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.434, + 0.585, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.511, + 0.574, + 0.523 + ], + "angle": 0, + "content": "(e) Fusformer" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.434, + 0.685, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.511, + 0.675, + 0.523 + ], + "angle": 0, + "content": "(f) UAL-DMI" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.434, + 0.787, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.511, + 0.776, + 0.523 + ], + "angle": 0, + "content": "(g) HSR-Diff" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.434, + 0.888, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.511, + 0.884, + 0.523 + ], + "angle": 0, + "content": "(h) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.527, + 0.71, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.562, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Figure 7: Visual quality comparison for fused HSIs of all competing methods on PaviaU, where first and second rows show 81st band and corresponding heatmaps (mean squared error), respectively." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.606, + 0.18, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.606, + 0.281, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.606, + 0.382, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.606, + 0.483, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.606, + 0.585, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.606, + 0.685, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.606, + 0.787, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.606, + 0.888, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.692, + 0.18, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.769, + 0.166, + 0.781 + ], + "angle": 0, + "content": "(a)UTV-TD" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.692, + 0.281, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.769, + 0.256, + 0.781 + ], + "angle": 0, + "content": "(b) UAL" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.692, + 0.382, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.769, + 0.372, + 0.781 + ], + "angle": 0, + "content": "(c) BRResNet" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.692, + 0.483, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.769, + 0.475, + 0.781 + ], + "angle": 0, + "content": "(d) CMHF-Net" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.692, + 0.585, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.77, + 0.574, + 0.781 + ], + "angle": 0, + "content": "(e) Fusformer" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.692, + 0.685, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.769, + 0.675, + 0.781 + ], + "angle": 0, + "content": "(f) UAL-DMI" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.692, + 0.787, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.769, + 0.776, + 0.781 + ], + "angle": 0, + "content": "(g) HSR-Diff" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.692, + 0.888, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.769, + 0.884, + 0.781 + ], + "angle": 0, + "content": "(h) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.784, + 0.71, + 0.809 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.819, + 0.892, + 0.85 + ], + "angle": 0, + "content": "Figure 8: Visual quality comparison for fused HSIs of all competing methods on Chikusei, where first and second rows show 67th band and corresponding heatmaps (mean squared error), respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7089" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.47, + 0.434 + ], + "angle": 0, + "content": "
DatasetMethodsPSNR↑SSIM↑SAM↓ERGAS↓
CAVEUTV-TD [34]38.660.97997.980.329
UAL [43]40.550.99334.330.271
BRResNet [16]41.360.99294.700.250
CMHF-Net [31]42.540.99394.690.216
Fusformer [15]41.520.99344.710.243
UAL-DMI [28]42.740.99503.790.213
HSR-Diff44.330.99513.710.179
PaviaUUTV-TD [34]44.460.99521.801.236
UAL [43]45.420.99641.541.148
BRResNet [16]45.530.99651.531.111
CMHF-Net [31]45.770.99651.501.096
Fusformer [15]45.660.99651.521.109
UAL-DMI [28]45.680.99661.491.113
HSR-Diff46.470.99771.451.053
ChikuseiUTV-TD [34]48.380.99890.991.303
UAL [43]56.180.99980.490.421
BRResNet [16]56.790.99980.460.366
CMHF-Net [31]55.990.99980.500.483
Fusformer [15]55.920.99980.520.492
UAL-DMI [28]56.570.99980.470.387
HSR-Diff57.340.99990.430.324
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.443, + 0.471, + 0.504 + ], + "angle": 0, + "content": "Table 1: Averaged PSNR, SSIM, SAM, and ERGAS of compared methods on CAVE, PaviaU, and Chikusei datasets. The \\(\\uparrow\\) or \\(\\downarrow\\) indicates higher or lower values corresponding to better results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Qualitative Comparison. To assess the performance of HSR-Diff qualitatively, we visualize example bands of HSIs in Figures 6, 7, and 8. It can be seen from these visual results that all compared methods produce satisfactory outcomes. In particular, HSR-Diff generates the best result with minor errors since the corresponding MSE (mean squared error) images are much clearer than the others." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Quantitative Comparison. To further verify the superior performance of the proposed HSR-Diff, quantitative results are presented in Table 1. Note that the performance indices on the CAVE and Chikusei datasets are averaged over all testing samples (12 samples for CAVE and eight samples for Chikusei), respectively. It can be inferred from the results that the proposed HSR-Diff surpasses all competitors with a clear margin on all evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.231, + 0.803 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of conditional diffusion models. Much of the early work on HSI-SR was based on the use of the regression model. To compare the effects of the diffusion and the regression model, we train the regression model containing the CDFformer. Note that the loss function, optimizer, and hyper-parameters are all the same as the conditional diffu" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.631, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.758, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.089, + 0.887, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.194, + 0.631, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.293, + 0.631, + 0.306 + ], + "angle": 0, + "content": "(a) Regression 49.13dB" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.194, + 0.758, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.293, + 0.756, + 0.306 + ], + "angle": 0, + "content": "(b) HSR-Diff 52.06dB" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.195, + 0.887, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.293, + 0.869, + 0.306 + ], + "angle": 0, + "content": "(c) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.31, + 0.892, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.339, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Figure 9: Fusion results for HSR-Diff and Regression on the real and fake peppers image of CAVE. The PSNR values (dB) are given in the subtitles." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.401, + 0.892, + 0.519 + ], + "angle": 0, + "content": "
DatasetModelPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVERegression43.260.99424.030.193
CDFformer44.330.99513.710.179
PaviaURegression46.030.99681.521.178
CDFformer46.470.99771.451.053
ChikuseiRegression56.710.99990.470.372
CDFformer57.340.99990.430.324
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.529, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Table 2: Ablation study on conditional diffusion models. The \\(\\uparrow\\) or \\(\\downarrow\\) indicates higher or lower values corresponding to better results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.756 + ], + "angle": 0, + "content": "sion models. Figure 9 presents the fused results and corresponding error maps of utilising HSR-Diff and the regression model. As can be seen from the error maps, the HSIs produced by HSR-Diff have less distortion than those by the regression model. This is because HSR-Diff works with a series of iterative refinement steps, facilitating the capture of richer information on data distributions of HR-HSIs. Furthermore, Table 2 showcases the quantitative results and effectively highlights the significant contribution of iterative refinement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of CDFormer. Recall that CDFormer is conditioned on the hierarchical representations of HR-MSI and LR-HSI via a two-stream architecture. However, alternative outstanding diffusion models [14, 23] that are also excellent for conditional image generation are equipped with CNN-based U-Nets, where degenerated images are concatenated with noisy high-resolution output images. To show the effectiveness of hierarchical representations, we remove the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7090" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.101, + 0.195, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.19, + 0.16, + 0.201 + ], + "angle": 0, + "content": "(a) HSI" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.101, + 0.311, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.19, + 0.289, + 0.201 + ], + "angle": 0, + "content": "(b)UTV-TD" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.101, + 0.427, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.19, + 0.394, + 0.201 + ], + "angle": 0, + "content": "(c) UAL" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.101, + 0.543, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.19, + 0.524, + 0.201 + ], + "angle": 0, + "content": "(d) BRResNet" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.101, + 0.657, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.19, + 0.642, + 0.201 + ], + "angle": 0, + "content": "(e) CMHF-Net" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.101, + 0.774, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.19, + 0.755, + 0.201 + ], + "angle": 0, + "content": "(f) UAL-DMI" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.101, + 0.889, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.19, + 0.87, + 0.202 + ], + "angle": 0, + "content": "(g) HSR-Diff" + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.215, + 0.715, + 0.23 + ], + "angle": 0, + "content": "Figure 10: Visual fusion results of all competing methods for HypSen." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.254, + 0.468, + 0.417 + ], + "angle": 0, + "content": "
DatasetNetworkPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVECD-CNN38.840.97977.320.318
C-w/o-SR43.740.99423.940.188
CDFformer44.330.99513.710.179
PaviaUCD-CNN42.750.99621.751.362
C-w/o-SR46.080.99761.471.080
CDFformer46.470.99771.451.053
ChikuseiCD-CNN47.630.99801.201.794
C-w/o-SR56.680.99990.460.425
CDFformer57.340.99990.430.324
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.427, + 0.47, + 0.459 + ], + "angle": 0, + "content": "Table 3: Ablation study on CDFormer. The \\( \\uparrow \\) or \\( \\downarrow \\) indicates higher or lower values corresponding to better results." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.476, + 0.472, + 0.593 + ], + "angle": 0, + "content": "
DatasetMethodsPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVEFixed43.170.99274.990.203
Progressive44.330.99513.710.179
PaviaUFixed45.060.99701.631.173
Progressive46.470.99771.451.053
ChikuseiFixed55.920.99990.500.453
Progressive57.340.99990.430.324
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.604, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Table 4: Ablation study on progressive learning. The \\(\\uparrow\\) or \\(\\downarrow\\) indicates higher or lower values corresponding to better results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.814 + ], + "angle": 0, + "content": "SR stream of CDFormer and name the resulting network \"C-w/o-SR\". In addition, we compare CDFormer with a CNN version of CDFormer that replaces all \\( S^2 \\)TL with convolutional layers and show its results as \"CD-CNN\" in Table 3. The quantitative results show the use of CDFormer performs better than CD-CNN, demonstrating the effectiveness of global statistics. Indeed, with the two-stream architecture, CDFormer offers the best results thanks to the use of hierarchical features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of progressive learning. Progressive learning helps CDFormer to capture long-range dependencies of spatio-spectral information in HR-HSIs. To illustrate the effect of progressive learning, we train the CDFormer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.349 + ], + "angle": 0, + "content": "with fixed patches (128² for CAVE and Chikusei; 64² for PaviaU) with the results shown under the heading of \"Fixed\" in Table 4. As can be seen, progressive learning (from 128² to 512² for CAVE and Chikusei; from 64² to 128² for PaviaU) provides better results than training with fixed patches." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.357, + 0.845, + 0.373 + ], + "angle": 0, + "content": "4.6. Generalization Analysis on Real Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.547 + ], + "angle": 0, + "content": "To examine the generalization ability of the implementations following the proposed approach, we test the performance of competitors on the real-world HypSen dataset [37]. Due to the lack of an ideal HR-HSI to train deep neural networks, we utilize the networks trained on the PaviaU dataset to merge observed LR-HSI and the corresponding HR-MSI. In addition, interpolation is applied to addressing the problem of an inconsistent number of bands between datasets. The fusion results of all compared methods are visualized in Figure 10, from which it can be seen that our method generates rich details, attaining satisfactory quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.619, + 0.575 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.888 + ], + "angle": 0, + "content": "In this paper, we have presented the novel HSR-Diff approach that initializes an HR-HSI with pure Gaussian noise and then, iteratively refines it subject to the condition of the LR-HSIs and HR-MSIs of interest. At each step, the noise is removed with CDFormer which exploits the hierarchical representations of HR-MSIs and LR-HSIs rather than the original images. In addition, we employ a progressive learning strategy to maximize the use of the global information of full-resolution images, where CDFormer is trained on small patches in the early epochs with high efficiency while on the global images in the later epochs to obtain the global statistics. Systematic experimental investigations have been conducted, on four public datasets to validate the superior performance of the proposed approach, in comparison with state-of-the-art methods. However, diffusion models have low inference efficiency due to their sequential nature, computational complexity, time-consuming sampling process, and large model size. For future work, we will try to resolve the challenging issue of the relatively low image-generation efficiency of HSR-Diff." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7091" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Naveed Akhtar, Faisal Shafait, and Ajmal Mian. Bayesian sparse representation for hyperspectral image super resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3631-3640, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Super-resolution for hyperspectral and multispectral image fusion accounting for seasonal spectral variability. IEEE Transactions on Image Processing, 29:116-127, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[3] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[4] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In Proceedings of the International Conference on Learning Represent, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[5] Zhao Chen, Hanye Pu, Bin Wang, and Geng-Ming Jiang. Fusion of hyperspectral and multispectral images: A novel framework based on generalization of pan-sharpening methods. IEEE Geoscience and Remote Sensing Letters, 11(8):1418-1422, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[6] Phuong D Dao, Kiran Mantripragada, Yuhong He, and Faisal Z Qureshi. Improving hyperspectral image segmentation by applying inverse noise weighting and outlier removal for optimal scale selection. ISPRS Journal of Photogrammetry and Remote Sensing, 171:348-366, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[7] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. Psrt: Pyramid shuffle-and-reshuffle transformer for multispectral and hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[8] Renwei Dian, Shutao Li, Anjing Guo, and Leyuan Fang. Deep hyperspectral image sharpening. IEEE transactions on neural networks and learning systems, 29(11):5345-5355, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[10] Ying Fu, Tao Zhang, Yinqiang Zheng, Debing Zhang, and Hua Huang. Hyperspectral image super-resolution with optimized rgb guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[15] Jin-Fan Hu, Ting-Zhu Huang, Liang-Jian Deng, Hong-Xia Dou, Danfeng Hong, and Gemine Vivone. Fusformer: A transformer-based fusion network for hyperspectral image super-resolution. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[16] Zi-Rong Jin, Liang-Jian Deng, Tian-Jing Zhang, and Xiao-Xu Jin. Bam: Bilateral activation mechanism for image fusion. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4315-4323, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[18] Shutao Li, Renwei Dian, Leyuan Fang, and José M Bioucas-Dias. Fusing hyperspectral and multispectral images via coupled sparse tensor factorization. IEEE Transactions on Image Processing, 27(8):4118-4130, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Ying Li, Haokui Zhang, and Qiang Shen. Spectral-spatial classification of hyperspectral imagery with 3d convolutional neural network. Remote Sensing, 9(1):67, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[21] Gamba Paolo. Pavia centre and university. https://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Sciences, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[22] Ying Qu, Hairong Qi, and Chiman Kwan. Unsupervised sparse dirichlet-net for hyperspectral image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2511-2520, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[24] Yue Shi, Liangxiu Han, Lianghao Han, Sheng Chang, Tongle Hu, and Darren Dancey. A latent encoder coupled generative adversarial network (le-gan) for efficient hyperspectral image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[25] Dong Wang, Yunpeng Bai, Bendu Bai, Chanyue Wu, and Ying Li. Heterogeneous two-stream network with hierarchical feature prefusion for multispectral pan-sharpening. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1845-1849. IEEE, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7092" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[26] Dong Wang, Yunpeng Bai, Chanyue Wu, Ying Li, Changjing Shang, and Qiang Shen. Convolutional LSTM-based hierarchical feature fusion for multispectral pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[27] Dong Wang, Pei Zhang, Yunpeng Bai, and Ying Li. Metapan: Unsupervised adaptation with meta-learning for multispectral pansharpening. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[28] Xiuheng Wang, Jie Chen, and Cedric Richard. Hyperspectral image super-resolution with deep priors and degradation model inversion. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2814-2818. IEEE, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[29] Qi Wei, José Bioucas-Dias, Nicolas Dobigeon, and Jean-Yves Tourneret. Hyperspectral and multispectral image fusion based on a sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 53(7):3658-3668, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.411 + ], + "angle": 0, + "content": "[30] Zhenyu Wu, Lin Wang, Wei Wang, Tengfei Shi, Chenglizhao Chen, Aimin Hao, and Shuo Li. Synthetic data supervised salient object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5557-5565, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.47, + 0.481 + ], + "angle": 0, + "content": "[31] Qi Xie, Minghao Zhou, Qian Zhao, Zongben Xu, and Deyu Meng. Mhf-net: An interpretable deep network for multispectral and hyperspectral image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[32] Fengchao Xiong, Jun Zhou, and Yuntao Qian. Material based object tracking in hyperspectral videos. IEEE Transactions on Image Processing, 29:3719-3733, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[33] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[34] Ting Xu, Ting-Zhu Huang, Liang-Jian Deng, Xi-Le Zhao, and Jie Huang. Hyperspectral image superresolution using unidirectional total variation with tucker decomposition. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13:4381-4398, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[35] Xizhe Xue, Haokui Zhang, Bei Fang, Zongwen Bai, and Ying Li. Grafting transformer on automatically designed convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[36] Jing Yang, Chanyue Wu, Tengfei You, Dong Wang, Ying Li, Changjing Shang, and Qiang Shen. Hierarchical spatio-spectral fusion for hyperspectral image super resolution via sparse representation and pre-trained deep model. Knowledge-Based Systems, 260:110170, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[37] Jingxiang Yang, Yong-Qiang Zhao, and Jonathan Cheung-Wai Chan. Hyperspectral and multispectral image fusion via deep two-branches convolutional neural network. Remote Sensing, 10(5):800, 2018. 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[38] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE transactions on image processing, 19(9):2241-2253, 2010.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[39] Naoto Yokoya and Akira Iwasaki. Airborne hyperspectral data over chikusei. Space Appl. Lab., Univ. Tokyo, Tokyo, Japan, Tech. Rep. SAL-2016-05-27, 5, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[40] Tengfei You, Chanyue Wu, Yunpeng Bai, Dong Wang, Huibin Ge, and Ying Li. Hmf-former: Spatio-spectral transformer for hyperspectral and multispectral image fusion. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[42] Haokui Zhang, Chengrong Gong, Yunpeng Bai, Zongwen Bai, and Ying Li. 3-d-anas: 3-d asymmetric neural architecture search for fast hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[43] Lei Zhang, Jiangtao Nie, Wei Wei, Yanning Zhang, Shengcai Liao, and Ling Shao. Unsupervised adaptation learning for hyperspectral imagery super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2020. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[44] Meilin Zhang, Xiongli Sun, Qiqi Zhu, and Guizhou Zheng. A survey of hyperspectral image super-resolution technology. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, pages 4476-4479. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[45] Xueting Zhang, Wei Huang, Qi Wang, and Xuelong Li. Ssrnet: Spatial-spectral reconstruction network for hyperspectral and multispectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5953-5965, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[46] Ke Zheng, Lianru Gao, Wenzhi Liao, Danfeng Hong, Bing Zhang, Xin Cui, and Jocelyn Chanussot. Coupled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super resolution. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2487-2502, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7093" + } + ] +] \ No newline at end of file diff --git a/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_origin.pdf b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db5d5bdf8c0da3baebdfe2d7f305a4950574bcb4 --- /dev/null +++ b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/aba8b2d5-ebce-4157-8dac-629ebe1b6925_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b844b52fc2ff06ffce5c70aa3b5acaf01973dd373dbb45e10123dbeebf7e8522 +size 9951462 diff --git a/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/full.md b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ac39fa9c569d275fbe3816a58b5626667e7b3b0 --- /dev/null +++ b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/full.md @@ -0,0 +1,495 @@ +# HSR-Diff: Hyperspectral Image Super-Resolution via Conditional Diffusion Models + +Chanyue Wu $^{1*}$ , Dong Wang $^{1*}$ , Yunpeng Bai $^{2*}$ , Hanyu Mao $^{1}$ , Ying Li $^{1\dagger}$ , Qiang Shen $^{2}$ + +$^{1}$ National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, Shaanxi Provincial Key Laboratory of Speech & Image Information Processing, School of Computer Science, Northwestern Polytechnical University, China + +$^{2}$ Department of Computer Science, Faculty of Business and Physical Sciences, Aberystwyth University, U.K. + +{chanyuewu, dongwang, maomhy} $@$ mail.nwpu.edu.cn lybyp@nwpu.edu.cn {yub3, qqs} $@$ aber.ac.uk + +# Abstract + +Despite the proven significance of hyperspectral images (HSIs) in performing various computer vision tasks, its potential is adversely affected by the low-resolution (LR) property in the spatial domain, resulting from multiple physical factors. Inspired by recent advancements in deep generative models, we propose an HSI Super-resolution (SR) approach with Conditional Diffusion Models (HSR-Diff) that merges a high-resolution (HR) multispectral image (MSI) with the corresponding LR-HSI. HSR-Diff generates an HR-HSI via repeated refinement, in which the HR-HSI is initialized with pure Gaussian noise and iteratively refined. At each iteration, the noise is removed with a Conditional Denoising Transformer (CDFformer) that is trained on denoising at different noise levels, conditioned on the hierarchical feature maps of HR-MSI and LR-HSI. In addition, a progressive learning strategy is employed to exploit the global information of full-resolution images. Systematic experiments have been conducted on four public datasets, demonstrating that HSR-Diff outperforms state-of-the-art methods. + +# 1. Introduction + +Hyperspectral images (HSIs) contain dozens or hundreds of spectral bands, enabling them to provide more faithful knowledge of targeted scenes than conventional imaging modalities. As such, HSIs play an irreplaceable role in various computer vision tasks, including classification [35, 42], segmentation [6], and tracking [32]. Although HSIs contain rich spectral information, contemporary hyperspectral + +![](images/8da2ed340bcb04d90a85f170285672c7a60e68bf0cfdbd485b9764fa5cf1f755.jpg) +Figure 1: Illustration of fusion-based Hyperspectral Image Super-Resolution (HSI-SR). + +imaging sensors lack high-resolution (HR) in the spatial domain, due to the stringent constraint of typically low signal-to-noise ratios. Their widespread use is significantly hindered by this fact. Restricted by hardware limitations, a practical way to work around this problem is to fuse the low-resolution (LR) HSI with an HR multispectral image (MSI). This requires the implementation of so-called HSI super-resolution (SR)[36], as shown in Figure 1. + +Over the past few decades, a significant amount of research efforts have been devoted to developing HSI-SR approaches, which can be roughly classified into five categories [44]: Extensions of pansharpening [5, 25, 27], Bayesian inference-based methods [1, 29], matrix factorization-based methods [2], tensor-based methods [18], and deep learning (DL)-based methods [31]. Whilst pansharpening methods [26] have been extended to the field of HSI-SR, such approaches are prone to spectral distortion. Bayesian inference-based approaches rely on the assumption of prior knowledge, thereby having a weak flexibility in dealing with different HSI structures. Matrix factorization-based techniques reshape the 3D HSIs and MSIs into matrices, thus facing the challenge of learning the required relationship between space and spectrum. Although sev- + +eral tensor-based methods have been proposed that can maintain the 3D structure of input images, they consume much more memory and computational power. Furthermore, these traditional approaches work via relying heavily on hand-crafted priors. + +Recently, DL-based methods, especially convolutional neural network (CNN)-based approaches, have flooded over into the HSI-SR research community [8, 31, 10, 22, 46, 43, 19]. Rather than resorting to hand-crafted features, DL-based techniques learn prior knowledge automatically from given data. Particularly, Dong et al. proposed the first DL-based method for image SR, with the end-to-end mapping between LR images and HR images learned using a CNN [9]. Subsequently, generative adversarial networks (GANs) were introduced to the field of image SR in an effort to produce high-frequency details [17, 11]. After that, various GAN-based models have been devised, showing state-of-the-art results in the HSI-SR literature [24]. However, such work requires carefully designed regularization and optimization tricks to tame optimization instability and avoid mode collapse. + +Inspired by the recent developments in deep generative models [14, 23], in this paper, we propose an innovative approach that we refer to as HSR-Diff (HSI-SR with conditional diffusion models). It works by learning to transform the original standard normal distribution into the data distribution of HR-HSI through a sequence of refinements. In contrast to GAN-based methods which require inner-loop maximization, HSR-Diff operates by simply minimizing a well-defined loss function to learn prior knowledge. Although conditional diffusion models are straightforward to define and efficient to train, there has been no demonstration that they are capable of merging LR-HSI and HR-MSI to the best of our knowledge. We show that conditional diffusion models are capable of generating high-quality HR-HSIs, which may best the state-of-the-art results. + +A key factor of HSR-Diff is its inherent denoising ability thanks to use of deep neural networks. In spite of the effectiveness of CNNs for denoising, they have shown limitations in modelling long-range dependencies. Some previous transformer-based HSI-SR works, especially Pyramid Shuffle-and-Reshuffle Transformer (PSRT) [7], Transformer-based Fusion network (Fusformer) [15], and Hyperspectral and Multispectral image Fusion Transformer (HMF-Former) [40], all of which learn long-range dependencies with attention mechanisms. These Transformers only take LR-HSI and HR-MSI as input, while the denoising Transformer is also fed with noise level $\gamma$ and noisy HR-HSI. To comprehensively exploit the spatial and spectral information in the LR-HSI, HR-MSI, and noisy HR-HSI, a Conditional Denoising Transformer (CD-Former) is herein designed and trained with a denoising objective to remove various levels of noise iteratively. In ad + +dition, a progressive learning strategy is utilized to help the CDFormer learn the global statistics of full-resolution HSIs. The main contributions of this work are summarized as follows: + +- We propose the novel application of conditional diffusion models in the field of HSI-SR that works by progressively destroying HR-HSI through injecting noise and subsequently learning to reverse this process, in order to perform HSI-SR. To the best of our knowledge, this is the first work to exploit diffusion models within the realm of HSI-SR. +- We introduce a CDFormer that refines a noisy HR-HSI conditioned on the deep feature maps of HR-MSI and LR-HSI, capable of modelling global connectivity with self-attention. Rather than concatenation, cross-attention is utilized as the conditioning mechanism to incorporate spatio-spectral information and noise level. +- We employ a progressive learning strategy to exploit the global information of full-resolution HSIs, with CDFormer being trained on small image patches in the early epochs with high efficiency and on the global images in the later epochs to acquire global information. +- We present experimental investigations on four public datasets, with quantitative and qualitative results illustrating the superior performance of our approach as compared with state-of-the-art methods. + +# 2. Related Work + +# 2.1. Diffusion models + +Typical deep generative models include autoregressive models (AR), normalizing flows (NF), variational autoencoders (VAE), GANs, diffusion models, etc. Over the past decade, GANs have gained prominence as a pivotal technology in the challenging field of image synthesis. However, this landscape has seen a transformation with the emergence of diffusion models, which have rapidly become a prominent research avenue across diverse domains, such as natural language processing, image processing, and computer vision. Diffusion models have arisen as an innovative approach that operates by iteratively refining a noise vector into the desired output. This reversible process allows for the generation of high-quality samples and has proven versatile in various applications. + +The development of diffusion models has seen a dramatically accelerating pace over the past three years. In the field of computer vision, diffusion models have shown promise in tasks like image segmentation [33], object detection [30], and even video processing [13]. Whilst diffusion models + +![](images/d388835a1cd9e49b9d7d217001d4d1efca42f8981acd60e33c519b9ec9521510.jpg) +Figure 2: Forward and reverse processes of HSR-Diff, with forward process $q$ generating an HSI sequence (left to right) by gradually adding Gaussian noise, and reverse process $p$ iteratively refining HR-HSI (right to left). + +have shown great potential for a variety of computer vision applications, none of them have yet been devoted to the problem of HSI-SR to the best of our knowledge. In this paper, we extend the utility of diffusion models to the field of HSI-SR. + +# 2.2. Deep Learning-Based HSI-SR + +In recent years, data-driven CNN architectures have been shown to outperform traditional approaches for use in the HSI-SR literature. These methods formulate the underlying fusion problem as a highly nonlinear mapping that takes HR-MSIs and LR-HSIs as input to generate an optimal HR-HSI. For example, CMHF-net [31] is an interpretable CNN, the design of which exploits the deep unfolding technique. Zhang et al. [43] proposed to reconstruct HR-HSIs with a two-stage network, while Zhang et al. [45] designed an interpretable spatial-spectral reconstruction network (SSR-NET) based on CNN. Aiming at problems of inflexible structure and information distortion, Jin et al. embedded Bilateral Activation Mechanism into ResNet, resulting in the effective model of BRResNet [16]. Thanks to the inductive bias of CNN, such as locality and weight sharing, these methods can provide good generalization performance and achieve impressive results. Nevertheless, CNNs have limitations in capturing long-range dependencies and self-similarity priors. To overcome such shortcomings, some Transformer-based HSI-SR approaches [15, 40] have been proposed. + +Within this study, the CDFormer serves as the denoising network in conditional diffusion models. It harnesses self-attention to learn global statistics from full-resolution HSIs and MSIs and employs cross-attention as the conditioning mechanism. + +# 3. Proposed Methodology + +# 3.1. Problem Formulation + +Without losing generality, the observation models for the HR-MSI and LR-HSI of interest can be mathematically formulated as + +$$ +\mathbf {X} = \mathbf {R Z} \tag {1} +$$ + +$$ +\mathbf {Y} = \mathbf {Z D}, +$$ + +where $\mathbf{X} \in \mathbb{R}^{b \times HW}$ denotes the HR-MSI which consists of $b$ spectral bands with a resolution of $HW$ in the spatial domain; $\mathbf{R} \in \mathbb{R}^{b \times B}$ represents the spectral response function of HR-MSI; $\mathbf{Y} \in \mathbb{R}^{B \times hw}$ denotes the LR-HSI; and $\mathbf{Z} \in \mathbb{R}^{B \times HW}$ is the HR-HSI. In the above, $b$ and $B$ are the numbers of bands, with $h$ and $H$ being the band height, and $w$ and $W$ the width, where $b \ll B$ , $h \ll H$ , and $w \ll W$ . $\mathbf{D} \in \mathbb{R}^{HW \times hw}$ is the spatial response of the LR-HSI, which can be modelled with blurring and down-sampling operations. The HSI-SR can be interpreted as an inverse problem for merging a practically collected $\mathbf{X}$ and an observed $\mathbf{Y}$ to produce a latent $\mathbf{Z}$ . In this paper, the ideal $\mathbf{Z}$ is restored with HSR-Diff conditioned on spatio-spectral information of $\mathbf{X}$ and $\mathbf{Y}$ , the details of which are described below. + +# 3.2. HSI-SR with Conditional Diffusion Models + +Given a dataset $\mathcal{D}_{train} = \{\mathbf{X}^i,\mathbf{Y}^i,\mathbf{Z}^i\}_{i = 1}^N$ satisfying a certain joint probability distribution $p(\mathbf{X},\mathbf{Y},\mathbf{Z})$ , many pairs of $(\mathbf{X},\mathbf{Y})$ may be consistent with the same $\mathbf{Z}$ . Thus, the HR-HSI $\mathbf{Z}$ can be obtained with iterative refinement that provide an approximate to $p(\mathbf{Z}|\mathbf{X},\mathbf{Y})$ . In this work, we implement the process of iterative refinement with HSR-Diff, where the optimized HR-HSI is presumed to be produced in $T$ refinement steps. In HSR-Diff, the target HR-HSI is initialized with a pure noise $\mathbf{Z}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ as shown in Figure 2. The HSI is then refined iteratively according to learned conditional distributions $p_{\theta}(\mathbf{Z}_{t - 1}|\mathbf{Z}_t,\mathbf{X},\mathbf{Y})$ . In so doing, the image sequence $(\mathbf{Z}_{T - 1},\mathbf{Z}_{T - 2},\dots ,\mathbf{Z}_0)$ can be attained and ultimately $\mathbf{Z}_0\sim p(\mathbf{Z}|\mathbf{X},\mathbf{Y})$ . + +The HSR-Diff makes use of two processes: a forward process that perturbs HR-HSI $\mathbf{Z}_0$ to noise, and a reverse process converting noise back to HR-HSI $\mathbf{Z}_0$ . In the forward process, the intermediate images, i.e., $\mathbf{Z}_{T - 1}$ , $\mathbf{Z}_{T - 2}$ , $\dots$ , and $\mathbf{Z}_1$ , are generated according to a Markov chain with fixed transition probability $q(\mathbf{Z}_t|\mathbf{Z}_{t - 1})$ . We are interested in reversing the process via iterative refinement, in which the noise is reduced iteratively with a reverse Markov chain conditioned on $\mathbf{X}$ and $\mathbf{Y}$ . The reverse chain is learned with the CDFformer $f_{\theta}$ . Further details of HSR-Diff's working are given below. + +# 3.2.1 Forward Process + +Inspired by [14], forward process $q$ iteratively adds Gaussian noise to $\mathbf{Z}_0$ over $T$ iterations: + +$$ +q \left(\mathbf {Z} _ {1: T} \mid \mathbf {Z} _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(\mathbf {Z} _ {t} \mid \mathbf {Z} _ {t - 1}\right) \tag {2} +$$ + +$$ +q \left(\mathbf {Z} _ {t} \mid \mathbf {Z} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {Z} _ {t}; \sqrt {\alpha_ {t}} \mathbf {Z} _ {t - 1}, (1 - \alpha_ {t}) \mathbf {I}\right), +$$ + +where $\alpha_{1:T} \in (0,1)$ are scalar hyper-parameters. Note that in the forward process, the distribution of $\mathbf{Z}_t$ given $\mathbf{Z}_0$ can + +![](images/244e86f93651d9b3ac4b788864e2c31da74688de21ac6075427a0351c496c294.jpg) +Figure 3: Architecture of Conditional Denoising Transformer. + +be directly sampled in closed form. This implies that + +$$ +q \left(\mathbf {Z} _ {t} \mid \mathbf {Z} _ {0}\right) = \mathcal {N} \left(\mathbf {Z} _ {t}; \sqrt {\gamma_ {t}} \mathbf {Z} _ {0}, (1 - \gamma_ {t}) \mathbf {I}\right), \tag {3} +$$ + +where $\gamma_{t} = \prod_{i = 1}^{t}\alpha_{i}$ . In addition, the posterior distribution of $\mathbf{Z}_{t - 1}$ given $\mathbf{Z}_0$ and $\mathbf{Z}_t$ can be derived by + +$$ +q \left(\mathbf {Z} _ {t - 1} \mid \mathbf {Z} _ {0}, \mathbf {Z} _ {t}\right) = \mathcal {N} \left(\mathbf {Z} _ {t - 1}; \boldsymbol {\mu}, \sigma^ {2} \mathbf {I}\right) +$$ + +$$ +\boldsymbol {\mu} = \frac {\sqrt {\gamma_ {t - 1}} (1 - \alpha_ {t})}{1 - \gamma_ {t}} \mathbf {Z} _ {0} + \frac {\sqrt {\alpha_ {t}} (1 - \gamma_ {t - 1})}{1 - \gamma_ {t}} \mathbf {Z} _ {t} \tag {4} +$$ + +$$ +\sigma^ {2} = \frac {(1 - \gamma_ {t - 1}) (1 - \alpha_ {t})}{1 - \gamma_ {t}}. +$$ + +As the CDFormer is tasked with predicting $\mathbf{Z}_0$ , the posterior is useful in the reverse process. + +# 3.2.2 Reverse Markovian Process + +The reverse process inferences $\mathbf{Z}_0$ via iterative refinement. It starts from a pure Gaussian noise $\mathbf{Z}_T$ and goes in the opposite direction of the forward process: + +$$ +p _ {\theta} \left(\mathbf {Z} _ {0: T} \mid \mathbf {X}, \mathbf {Y}\right) = p \left(\mathbf {Z} _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(\mathbf {Z} _ {t - 1} \mid \mathbf {Z} _ {t}, \mathbf {X}, \mathbf {Y}\right) +$$ + +$$ +p \left(\mathbf {Z} _ {T}\right) = \mathcal {N} \left(\mathbf {Z} _ {T}; \mathbf {0}, \mathbf {I}\right) +$$ + +$$ +p _ {\theta} \left(\mathbf {Z} _ {t - 1} \mid \mathbf {Z} _ {t}, \mathbf {X}, \mathbf {Y}\right) = \mathcal {N} \left(\mathbf {Z} _ {t - 1}; \mu_ {\theta} \left(\mathbf {X}, \mathbf {Y}, \mathbf {Z} _ {t}, \gamma_ {t}\right), \sigma_ {t} ^ {2} \mathbf {I}\right), \tag {5} +$$ + +where the distribution $p_{\theta}(\mathbf{Z}_{t - 1}\mid \mathbf{Z}_t,\mathbf{X},\mathbf{Y})$ is parameterized with $\theta$ . Note that the CDFormer provides a prediction of $\hat{\mathbf{Z}}_0$ . Thus, according to (4), each refinement step takes + +the following form: + +$$ +\begin{array}{l} \mathbf {Z} _ {t - 1} = \frac {\sqrt {\gamma_ {t - 1}} (1 - \alpha_ {t})}{1 - \gamma_ {t}} f _ {\theta} (\mathbf {X}, \mathbf {Y}, \mathbf {Z} _ {t}, \gamma_ {t}) \tag {6} \\ + \frac {\sqrt {\alpha_ {t}} (1 - \gamma_ {t - 1})}{1 - \gamma_ {t}} \mathbf {Z} _ {t} + \sqrt {1 - \alpha_ {t}} \boldsymbol {\epsilon}, \\ \end{array} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and $f_{\theta}$ is the CDFormer. + +# 3.2.3 Noise Schedule + +Inspired by the research reported in [4], we sample $\gamma$ with two steps during training. In particular, we first sample a time step $t\sim U\{1,T\}$ and then randomly select $\gamma \sim U(\gamma_{t - 1},\gamma_t)$ . As such, $\gamma \sim p(\gamma) = \sum_{t = 1}^{T}\frac{1}{T} U(\gamma_{t - 1},\gamma_t)$ . Normally, the model with a large $T$ can achieve better results. However, we find (through empirical investigations) that the performance is not very sensitive to the exact values of $T$ . Therefore, no hyper-parameter search about $T$ is conducted and we set $T = 2000$ for simplicity. As for the inference process, we set the maximum generation iterations to 100, employing a linear noise schedule. + +# 3.3. Conditional Denoising Transformer + +The property of non-local self-similarity of HSIs is often exploited in denoising tasks but is usually not well captured by CNN-based models. Due to the effectiveness of Transformer layer in capturing non-local long-range dependencies, the potential of Transformer is explored in conditional denoising of HSI. Unfortunately, the vanilla Transformer focuses only on spatial relationships between pixels while neglecting the spectral dimension. Besides, denoising networks in conditional diffusion models normally concatenate + +![](images/d62b5e0682aa49c78a1d8b309984db45471b2bc594e44417c5d6f86b2d2975d7.jpg) +Figure 4: Spatio-Spectral Transformer Layer, where "LN" denotes layer normalization. + +all images together as input, which may hinder the extraction of useful spatio-spectral information in LR-HSIs and HR-MSIs. Hence, the CDFormer adopts a two-stream architecture and is constructed with stacked Spatio-Spectral Transformer Layers (S $^2$ TLs). + +The architecture of the CDFormer is shown in Figure 3. The SR stream first utilizes a $3 \times 3$ convolution to generate low-level feature embeddings $\mathbf{F}_0^{SR}$ and then transforms it into deep features $\mathbf{F}_l^{SR}$ with stacked- $\mathbf{S}^2\mathbf{TLs}$ . Instead of adopting $t$ as done in the existing work [4], our method is conditioned on $\gamma$ directly to achieve efficient generation. The denoising stream contains multiple noise-aware conditional $\mathbf{S}^2\mathbf{TLs}$ (NC- $\mathbf{S}^2\mathbf{TLs}$ ) that take as input the embedded noise level and the image representation $\mathbf{F}^{SR}$ . The Reconstruction module is set to produce a noise-free HR-HSI, by employing residual learning to alleviate the difficulty of HR-HSI generation while mapping the features onto HR-HSI via a $3 \times 3$ convolution and addition operation. It should be noted that no downsampling operation is employed. + +# 3.3.1 Noise Level Embedding + +The noise level offers essential information for denoising models. Inspired by the work of [4], we embed noise level within the models with sinusoidal positional encoding. The process of noise level embedding (NLE) can be formulated as follows: + +$$ +N L E _ {\gamma , 2 i} = \sin \left(\gamma / 1 0 0 0 0 ^ {2 i / C}\right) \tag {7} +$$ + +$$ +N L E _ {\gamma , 2 i + 1} = \cos \left(\gamma / 1 0 0 0 0 ^ {2 i / C}\right), +$$ + +where $C$ is the number of channels of $\mathbf{S}^2\mathrm{TLs}$ . $i\in [1,C / 2]$ + +# 3.3.2 Spatio-Spectral Transformer Layers + +Figure 4 illustrates the architecture of one $\mathrm{S}^2\mathrm{TL}$ , which consists of a Spatial Multi-head Self-Attention (SpatioMSA), a Spectral Multi-head Self-Attention (SpectralMSA), and a Feed Forward Network (FFN). SpatioMSA and SpectralMSA learn the interactions of spatial regions and interspectra relationships, respectively. To alleviate the computational burden, we adopt the transposed attention [41] in + +![](images/3a284d97c03c04043a8c65162228a54afa8c9a537273be18c20fe4933401ecfd.jpg) +Figure 5: Noise-Aware Conditional Spatio-Spectral Transformer Layer, where "LT" represents linear transform, "LN" denotes layer normalization, and "MCA" is the abbreviation of multi-head cross-attention. + +SpectralMSA. SpatioMSA applies the popular window partitioning strategy [20] to reduce the computational complexity. In addition, the gating mechanism [41] is employed in the implementation of FFN. + +# 3.3.3 Noise-Aware Conditional $\mathbf{S}^2\mathbf{T}\mathbf{L}\mathbf{s}$ + +To condition the overall model on the hierarchical features of SR stream, we feed $\mathbf{F}_l^{SR}$ to the Noise-Aware Conditional S $^2$ TLs (NC-S $^2$ TLs), each of which is a key building block of the denoising stream. Figure 5 depicts the structure of an NC-S $^2$ TL, which takes as input the NLE (a vector), $\mathbf{F}_l^{SR}$ and $\mathbf{F}_l^{DS}$ , where $\mathbf{F}_l^{SR}$ and $\mathbf{F}_l^{DS}$ have the same spatial resolution. NLE is first transformed and merged with $\mathbf{F}_l^{DS}$ with the result subsequently processed with the means of multi-head cross-attention (MCA) [3] in order to condition the model on $\mathbf{F}_l^{SR}$ . As a result, each S $^2$ TL learns the spatio-spectral dependencies. + +# 3.4. Progressive Learning Strategy + +Deep learning-based HSI-SR models are normally trained on small cropped patches. However, on one hand, training the proposed CDFormer on fixed-size image patches may not appropriately reflect the global relationships, which may result in suboptimal performance on full-resolution images when used. On the other hand, during the training phase, CDFormer can not be uploaded to Graphics Processing Unit (GPU) because of the limited memory. With this objective in mind, we adopt a strategy of progressive learning. In the initial epochs, the network is trained using smaller image patches. Subsequently, as training progresses, we transition towards utilizing larger patches, and ultimately, full-resolution images in the concluding training epochs. The resulting model, cultivated through the integration of varying sizes via progressive learning, exhibits notable enhancements in performance during testing. This is particularly evident when handling images of diverse resolutions, a scenario frequently encountered in the realm of the HSI-SR task. + +To reduce the pressure on the demand of GPU memory, we only train the second half of CDFormer on full + +resolution images. The loss function used for such training is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} = \| \mathbf {X} - \mathbf {R} \hat {\mathbf {Z}} _ {0} \| _ {1} + \| \mathbf {Y} - \hat {\mathbf {Z}} _ {0} \mathbf {D} \| _ {1} + \| \mathbf {Z} _ {0} - \hat {\mathbf {Z}} _ {0} \| _ {1} \tag {8} \\ \hat {\mathbf {Z}} _ {0} = f _ {\theta} (\sqrt {\gamma} \mathbf {Z} _ {0} + \sqrt {1 - \gamma} \boldsymbol {\epsilon}, \mathbf {X}, \mathbf {Y}), \\ \end{array} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ $(\mathbf{X},\mathbf{Y},\mathbf{Z})$ is sampled from the training set, and the noise schedule about $\gamma$ has been discussed above. The first two terms are designed according to the observation models, while the last one is based on the assumption of Laplace distribution. + +# 4. Experiments + +Systematic experiments are herein conducted on four commonly-used public-available HSI-SR datasets to demonstrate the effectiveness of the proposed approach. + +# 4.1. Datasets + +Four datasets including CAVE [38], PaviaU [21], Chikusei [39], and HypSen [37] are used in our experiments, with the following details on each. + +CAVE: There are 32 scenes with a spatial size of $512 \times 512$ in the CAVE dataset, where we select the first 20 HSIs for training, with the remaining 12 images used for testing. We generate LR-HSIs by Gaussian blur and down-sampling using a factor of 32 as done in [31]. HR-MSIs are acquired by integrating all HR-HSI bands according to the spectral response function of Nikon D700. The original HR-HSIs are treated as ground truth. + +PaviaU: Collected by the University of Pavia, Italy, the original HSI dataset consists of $610 \times 340$ pixels in which the top-left $128 \times 128$ area is extracted as the test data, with the remaining used for training. Except for water absorption bands, all other 103 bands are chosen for the experiments. Note that the down-sampling factor for the generation of LR-HSIs is four, and the spectral response function is the same as that of the WorldView-3 satellite. + +Chikusei: This dataset consists of 128 bands with a spectral range of $363~nm$ to $1018~nm$ and a spatial resolution of $2517\times 2335$ . The original HSI data was taken by an airborne visible and near-infrared imaging sensor over Chikusei, Japan. To alleviate the impact of the back boundary and noise, we crop the center area and remove noise bands. The processed image has a size of $2048\times 2048\times 110$ . The top half $1024\times 2048\times 110$ area is selected as the training data, while the rest half is split into eight testing $512\times 512\times 110$ patches. For the production of LR-HSIs and HR-MSIs, this dataset adopts the same processing as with PaviaU. + +HypSen: This dataset concerns a real scenario consisting of a $30\mathrm{m}$ -resolution HSI and a $10\mathrm{m}$ -resolution MSI. The Hyperion sensor on the Earth Observing-1 satellite provided the HSI with 242 spectral bands in the spectral range of $400\mathrm{nm}$ to $2500\mathrm{nm}$ , and the MSI with 13 bands was captured by the Sentinel-2A satellite. The blue, green, red, and near-infrared bands of MSI in our experiments are selected due to their high spatial resolution. To eliminate the impact of noise and water absorption, we remove those relevant bands, with 84 bands remaining in the HSI. We crop sub-images of size $250\times 330$ and $750\times 990$ from the Hyperion HSI and Sentinel-2A MSI respectively, in our study, with the pairs of sub-image patches spatially registered. + +# 4.2. Methods Compared and Evaluation Metrics Used + +Six state-of-the-art HSI-SR approaches are taken for comparison, including: UTV-TD [34], UAL [43], BRResNet [16], Fusformer [15], CMHF-Net [31], and UAL-DMI [28]. UTV-TD is a tensor-based technique; UAL, BRResNet, Fusformer, and CMHF-Net fall into the category of the DL-based methods; and UAL-DMI can be regarded as an upgraded version of UAL. + +Four quantitative quality metrics are employed for performance evaluation, including Peak Signal-to-Noise Ratio (PSNR), Spectral Angle Mapper (SAM), Erreur Relative Globale Adimensionnelle de Synthese (ERGAS, namely error relative global dimensionless synthesis), and Structure Similarity (SSIM). The smaller ERGAS and SAM are, the larger PSNR and SSIM are, the better the fusion result is. + +# 4.3. Implementation Specification + +All DL-based methods are trained on the same datasets. For those compared methods, we use the publicly available source codes with default hyper-parameters as given in the corresponding research papers. Our HSR-Diff is implemented on the PyTorch framework. The learnable parameters of the CDFformer are initialized with Kaiming initialization [12] and trained on 2 NVIDIA GeForce GTX 3090s. The dimension of the embedding of $F_0^{SR}$ and $F_0^{DS}$ is set to 256, and the number of parameters of the CDFformer is 31.56M. We utilize the Adam optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ to optimize the CDFformer. With limited GPU memory, the batch size is set to 4 and 2 for $128^2$ and $512^2$ images, respectively. It costs 20000 epochs on the CAVE and PaviaU datasets while consuming 5000 epochs on the Chikusei dataset. The learning rate is initialized as $1 \times 10^{-4}$ . + +# 4.4. Comparisons with State-of-the-art Methods + +In this set of experiments, the evaluations are carried out using the first three datasets listed above without involving the real-world dataset, HypSen (which will be dealt with in the next section). + +![](images/cba18c9ddba5abe00b2a6ddfbcb11954b588a839e2556220fcddf8f8e28c2ebf.jpg) + +![](images/bf86ea733028e5c49c7dc3cd3963201f0ac1c97201694df1ebfb1cb0c1cf78ae.jpg) + +![](images/9e19c26d5f1691568c600a3db99e69618a6c904aeae17a99cd94a1696111aae0.jpg) + +![](images/06006c40dcd41774640c94589a49c8dfe77201c56bc53e5e42d7bbb2b33004d5.jpg) + +![](images/620a9d1cc61f2fb29768f5df4763c78f3f6c93c4c93ed9733072a159ffed88f6.jpg) + +![](images/7d557b71e6b2f660b7dff00195628a886a39a02b194227397046cc543380e87c.jpg) + +![](images/92977a1bf99e675b1918955657bc05ec4e1403271aac5eaa2918f409990a0061.jpg) + +![](images/55c004a6a257bab4f7f3b70c6466379de314a974554fbeb4a284c487cc75ee60.jpg) + +![](images/107074d9837edb1f92e3e6afa931490806ff550fecff53675a16ff6dce68e0a6.jpg) +(a)UTV-TD + +![](images/bf6681f2aa80f1306ec5f258f396df9745e5a42604d9bb12017c660105eb98cb.jpg) +(b) UAL + +![](images/4b0234e0ea0389fc716690f082783c6e86cfb82da0fecc048a62e00bad743da8.jpg) +(c) BRResNet + +![](images/9c5ce54149a2397db2d2e59aca0e79b71eeadc7f1557cb4c1dd2369262c55c48.jpg) +(d) CMHF-Net + +![](images/aa377e3121c60ee7e6a74835c36cf852ca856063994f2b7f2557d416f3f69e40.jpg) +(e) Fusformer + +![](images/20269a988c05acb14b1bd1d06ed8c374f73ec7c3a73e4919d06afb917669df67.jpg) +(f) UAL-DMI + +![](images/5d4e4ab7f93e145eca4839ff18a91678668521abcc305deee4a80fd40acebeab.jpg) +(g) HSR-Diff + +![](images/90c273aa673a219ee3995e474f0872034fb801ede3aef075ecb307616887e0e1.jpg) +(h) Ground truth + +![](images/af089bc0248aba477c785501900a8bdaff7d901bc35a8f7276ceaa17b9233146.jpg) +Figure 6: Visual quality comparison for fused HSIs of all competing methods on CAVE, where first and second rows show fourth band and corresponding heatmaps (mean squared error), respectively. + +![](images/6f05a70c336542e6a5c8a56a77b432068d2c241b78cf8c3d189533bd05e494bf.jpg) + +![](images/1c3affc17c59a43ae29d90f24e1fe8b7e1bb42affb45f282fe47bd721b7c7bed.jpg) + +![](images/85cec20e87badfddcf7056487c5696bcf68e0a2ed623a850ee0141ce3ad1de2c.jpg) + +![](images/c358e851f5dce380e6c6d400642f8c9b775d2e5c0c46749ccb0cad36c5085df4.jpg) + +![](images/c5e0e90785b7e242e23f9aa0ae81536e743e21e3cd10976cd79f1d71c4ecc2fd.jpg) + +![](images/06a99816e30cc328bcd3262f6dfdba2093ea142821f323fce9b4c9afa85c8222.jpg) + +![](images/825a15e0b18e625c3152acf7180b5dc549c551cb056d22289638ca572afc24bc.jpg) + +![](images/fb0a566c8b2ce8acf2ce754e33febc80d1aae4b0ad151f101f02fb3622305112.jpg) + +![](images/26058ff6e433fc931062169f70a0ccee449e274b5498acf298a508503cfd81aa.jpg) +(a)UTV-TD + +![](images/a56509deef7e4e3bc20c38d3b77424cb97b145379098ff744410ec779afe0caf.jpg) +(b) UAL + +![](images/366c012d455662c4610b0e1b9cf53f177cf025fe10d57b25267520c065163a89.jpg) +(c) BRResNet + +![](images/8cb71bb58a8c6441629007401dcd2a2c3e88af5a0a6de0e77fbb917d55ef2d3f.jpg) +(d) CMHF-Net + +![](images/3db35b2163920ecfce7896ae0f3bb81dfb8842cba1bff4516cd54fbd969c9503.jpg) + +![](images/d03d5cd96bfef70c112b81cb88c34bf067c12773329836f86736d8b223c1a1b9.jpg) +(f) UAL-DMI + +![](images/b6ec2f8b52aa5fb40683f058ed144bbce24113e3206a5f795c4fe5ab49fa246c.jpg) +(g) HSR-Diff + +![](images/b464dd4a7a81058c9b4a0d6b4cce870ee7ed08b6b4ccfbd6152b19e1a703e752.jpg) +(h) Ground truth + +![](images/ea2453226b911803317da1556b439e4d300ee7c4c0b85e537953fcaa8dc87bd4.jpg) +(e) Fusformer + +![](images/d0893649ed3b9b4284f388e1e8c614dd98651dc04d735063c4521e08d66d97c4.jpg) +Figure 7: Visual quality comparison for fused HSIs of all competing methods on PaviaU, where first and second rows show 81st band and corresponding heatmaps (mean squared error), respectively. + +![](images/e1f80953259ba7aac49f4b03d2ad2215aa949516f66ceeeaf4876535b7fd83ed.jpg) + +![](images/b691686140c59179ab86adbe497853c99b510b2e0c3a94a229c3108161becf74.jpg) + +![](images/bf0051760551a1149d076a9f64dd1ef0c897b3a421733b3cf7ecf2bfaa43b4d7.jpg) + +![](images/596261dd5b1ff33bd3f61d16fc0068efafd3c07b6bec686f6016c7c184f818fa.jpg) + +![](images/ff689d97198edd375e059155ece0a53219aff4463f11ab039162f9087a683206.jpg) + +![](images/6902bdd95fb516980f87f20722069d90e2bf0eed85656eb2a9c027a1f2a3f528.jpg) + +![](images/cd07ddb4ac56f2ec519beee3336ec1d3b2481ea050228cd9c47825553833a902.jpg) + +![](images/eaef687c57d2de4199b95308df6cb6f5ce3de64d8a1df75b8fb2fae63b822f7d.jpg) +(a)UTV-TD + +![](images/cfcb16a0fe695da51f0ace142c54f2153e27d75791f96221cac4e66f8649989e.jpg) +(b) UAL + +![](images/a8c302a9c5e986b5d5f199568399b95dda997d43549e1638f4ed122d62cd8a9f.jpg) +(c) BRResNet + +![](images/0fa7947ad384a6ce64801c60ed35ed87d07a17027c11a06d50539e33ca9f9aba.jpg) +(d) CMHF-Net + +![](images/0e8cfd22c89df9d0b2c1ceb685f0ffbb09811e30e0e4b6fdaa4049d8aff97f39.jpg) + +![](images/9822930510caa0752893ef4b337cf40ec9c55b584f63f39408356a70782ec487.jpg) + +![](images/7d1b36b224b1735acfdbc56e6170292e73f3eaf7a1f62624bae76849243284d5.jpg) + +![](images/3502d620be0556cb67bbe1b5167be649031f571fdc4e20959d12f27f77a8a77f.jpg) +(h) Ground truth + +![](images/3c257a0e9bbfb832178a24e0e8d365e6fd0bcc1cbca1cbfe35e4a4a795069305.jpg) +(e) Fusformer +(f) UAL-DMI +(g) HSR-Diff +Figure 8: Visual quality comparison for fused HSIs of all competing methods on Chikusei, where first and second rows show 67th band and corresponding heatmaps (mean squared error), respectively. + +
DatasetMethodsPSNR↑SSIM↑SAM↓ERGAS↓
CAVEUTV-TD [34]38.660.97997.980.329
UAL [43]40.550.99334.330.271
BRResNet [16]41.360.99294.700.250
CMHF-Net [31]42.540.99394.690.216
Fusformer [15]41.520.99344.710.243
UAL-DMI [28]42.740.99503.790.213
HSR-Diff44.330.99513.710.179
PaviaUUTV-TD [34]44.460.99521.801.236
UAL [43]45.420.99641.541.148
BRResNet [16]45.530.99651.531.111
CMHF-Net [31]45.770.99651.501.096
Fusformer [15]45.660.99651.521.109
UAL-DMI [28]45.680.99661.491.113
HSR-Diff46.470.99771.451.053
ChikuseiUTV-TD [34]48.380.99890.991.303
UAL [43]56.180.99980.490.421
BRResNet [16]56.790.99980.460.366
CMHF-Net [31]55.990.99980.500.483
Fusformer [15]55.920.99980.520.492
UAL-DMI [28]56.570.99980.470.387
HSR-Diff57.340.99990.430.324
+ +Qualitative Comparison. To assess the performance of HSR-Diff qualitatively, we visualize example bands of HSIs in Figures 6, 7, and 8. It can be seen from these visual results that all compared methods produce satisfactory outcomes. In particular, HSR-Diff generates the best result with minor errors since the corresponding MSE (mean squared error) images are much clearer than the others. + +Quantitative Comparison. To further verify the superior performance of the proposed HSR-Diff, quantitative results are presented in Table 1. Note that the performance indices on the CAVE and Chikusei datasets are averaged over all testing samples (12 samples for CAVE and eight samples for Chikusei), respectively. It can be inferred from the results that the proposed HSR-Diff surpasses all competitors with a clear margin on all evaluation metrics. + +# 4.5. Ablation Study + +Effect of conditional diffusion models. Much of the early work on HSI-SR was based on the use of the regression model. To compare the effects of the diffusion and the regression model, we train the regression model containing the CDFformer. Note that the loss function, optimizer, and hyper-parameters are all the same as the conditional diffu + +![](images/282b272ad671d6b404a3ed22635d1576a4bf78b78ed4702b1d69ab6d55912c09.jpg) + +![](images/7f27e977efd2497c09bc9d01b49e33e674787290d6375d8be6eceb808e3a0453.jpg) + +![](images/1510adeba85387e2a447338c741704477933bd78e2a337868bcdbd57d0f3ff6b.jpg) + +![](images/20c5dfeab9f90b804383c1554725bda46e8f71ac8393518617942cfcb6fe1d5b.jpg) +(a) Regression 49.13dB + +![](images/a61dda589a4d10190c537c36b990fc015b019689611ef15fff5d05fc466ea5e7.jpg) +(b) HSR-Diff 52.06dB + +![](images/5345dfe7c76dd1556d43bdea406132e4e5a32129ba1d03d19c3260b88b8540a0.jpg) + +![](images/a2f269d52ba9006f83bc3eae5c1e1e1af5ba9a1047ba202316b82386499a68dd.jpg) +(c) Ground truth +Figure 9: Fusion results for HSR-Diff and Regression on the real and fake peppers image of CAVE. The PSNR values (dB) are given in the subtitles. + +Table 1: Averaged PSNR, SSIM, SAM, and ERGAS of compared methods on CAVE, PaviaU, and Chikusei datasets. The $\uparrow$ or $\downarrow$ indicates higher or lower values corresponding to better results. + +
DatasetModelPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVERegression43.260.99424.030.193
CDFformer44.330.99513.710.179
PaviaURegression46.030.99681.521.178
CDFformer46.470.99771.451.053
ChikuseiRegression56.710.99990.470.372
CDFformer57.340.99990.430.324
+ +Table 2: Ablation study on conditional diffusion models. The $\uparrow$ or $\downarrow$ indicates higher or lower values corresponding to better results. + +sion models. Figure 9 presents the fused results and corresponding error maps of utilising HSR-Diff and the regression model. As can be seen from the error maps, the HSIs produced by HSR-Diff have less distortion than those by the regression model. This is because HSR-Diff works with a series of iterative refinement steps, facilitating the capture of richer information on data distributions of HR-HSIs. Furthermore, Table 2 showcases the quantitative results and effectively highlights the significant contribution of iterative refinement. + +Effect of CDFormer. Recall that CDFormer is conditioned on the hierarchical representations of HR-MSI and LR-HSI via a two-stream architecture. However, alternative outstanding diffusion models [14, 23] that are also excellent for conditional image generation are equipped with CNN-based U-Nets, where degenerated images are concatenated with noisy high-resolution output images. To show the effectiveness of hierarchical representations, we remove the + +![](images/0d990b38633caae1cbcf582cb90a4f8439c06a8c69001841a19b491494f410ea.jpg) +(a) HSI + +![](images/1e63519dc0dcfcefe9e053ccda826e92fdb28596669bcd6890c6f136a2c5cd56.jpg) +(b)UTV-TD + +![](images/f61bb7027e411a835dfedcdfcc38e82a120285039a79cdc5dc30b0e183dc8b12.jpg) +(c) UAL + +![](images/c4475d6dc385fe81cbc5c162b73b76066729cc52030d0135999066f24bcd5312.jpg) +(d) BRResNet +Figure 10: Visual fusion results of all competing methods for HypSen. + +![](images/3c78fd2b697fbf692f094d43862d354311e4dbb6ca987e0740d6591e974a4c1a.jpg) +(e) CMHF-Net + +![](images/cfcc1b3e183daf06d3dd4fb15e64af6136d21ca99b601c1a960440c992268d7c.jpg) +(f) UAL-DMI + +![](images/24eb1dcfd33ad077d2b7cb7cbe93b8643a16cb381f4337c1ac7e63b26e6c717b.jpg) +(g) HSR-Diff + +
DatasetNetworkPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVECD-CNN38.840.97977.320.318
C-w/o-SR43.740.99423.940.188
CDFformer44.330.99513.710.179
PaviaUCD-CNN42.750.99621.751.362
C-w/o-SR46.080.99761.471.080
CDFformer46.470.99771.451.053
ChikuseiCD-CNN47.630.99801.201.794
C-w/o-SR56.680.99990.460.425
CDFformer57.340.99990.430.324
+ +Table 3: Ablation study on CDFormer. The $\uparrow$ or $\downarrow$ indicates higher or lower values corresponding to better results. + +
DatasetMethodsPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVEFixed43.170.99274.990.203
Progressive44.330.99513.710.179
PaviaUFixed45.060.99701.631.173
Progressive46.470.99771.451.053
ChikuseiFixed55.920.99990.500.453
Progressive57.340.99990.430.324
+ +Table 4: Ablation study on progressive learning. The $\uparrow$ or $\downarrow$ indicates higher or lower values corresponding to better results. + +SR stream of CDFormer and name the resulting network "C-w/o-SR". In addition, we compare CDFormer with a CNN version of CDFormer that replaces all $S^2$ TL with convolutional layers and show its results as "CD-CNN" in Table 3. The quantitative results show the use of CDFormer performs better than CD-CNN, demonstrating the effectiveness of global statistics. Indeed, with the two-stream architecture, CDFormer offers the best results thanks to the use of hierarchical features. + +Effect of progressive learning. Progressive learning helps CDFormer to capture long-range dependencies of spatio-spectral information in HR-HSIs. To illustrate the effect of progressive learning, we train the CDFormer + +with fixed patches (128² for CAVE and Chikusei; 64² for PaviaU) with the results shown under the heading of "Fixed" in Table 4. As can be seen, progressive learning (from 128² to 512² for CAVE and Chikusei; from 64² to 128² for PaviaU) provides better results than training with fixed patches. + +# 4.6. Generalization Analysis on Real Dataset + +To examine the generalization ability of the implementations following the proposed approach, we test the performance of competitors on the real-world HypSen dataset [37]. Due to the lack of an ideal HR-HSI to train deep neural networks, we utilize the networks trained on the PaviaU dataset to merge observed LR-HSI and the corresponding HR-MSI. In addition, interpolation is applied to addressing the problem of an inconsistent number of bands between datasets. The fusion results of all compared methods are visualized in Figure 10, from which it can be seen that our method generates rich details, attaining satisfactory quality. + +# 5. Conclusion + +In this paper, we have presented the novel HSR-Diff approach that initializes an HR-HSI with pure Gaussian noise and then, iteratively refines it subject to the condition of the LR-HSIs and HR-MSIs of interest. At each step, the noise is removed with CDFormer which exploits the hierarchical representations of HR-MSIs and LR-HSIs rather than the original images. In addition, we employ a progressive learning strategy to maximize the use of the global information of full-resolution images, where CDFormer is trained on small patches in the early epochs with high efficiency while on the global images in the later epochs to obtain the global statistics. Systematic experimental investigations have been conducted, on four public datasets to validate the superior performance of the proposed approach, in comparison with state-of-the-art methods. However, diffusion models have low inference efficiency due to their sequential nature, computational complexity, time-consuming sampling process, and large model size. For future work, we will try to resolve the challenging issue of the relatively low image-generation efficiency of HSR-Diff. + +# References + +[1] Naveed Akhtar, Faisal Shafait, and Ajmal Mian. Bayesian sparse representation for hyperspectral image super resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3631-3640, 2015. 1 +[2] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Super-resolution for hyperspectral and multispectral image fusion accounting for seasonal spectral variability. IEEE Transactions on Image Processing, 29:116-127, 2019. 1 +[3] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 5 +[4] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In Proceedings of the International Conference on Learning Represent, 2021. 4, 5 +[5] Zhao Chen, Hanye Pu, Bin Wang, and Geng-Ming Jiang. Fusion of hyperspectral and multispectral images: A novel framework based on generalization of pan-sharpening methods. IEEE Geoscience and Remote Sensing Letters, 11(8):1418-1422, 2014. 1 +[6] Phuong D Dao, Kiran Mantripragada, Yuhong He, and Faisal Z Qureshi. Improving hyperspectral image segmentation by applying inverse noise weighting and outlier removal for optimal scale selection. ISPRS Journal of Photogrammetry and Remote Sensing, 171:348-366, 2021. 1 +[7] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. Psrt: Pyramid shuffle-and-reshuffle transformer for multispectral and hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2 +[8] Renwei Dian, Shutao Li, Anjing Guo, and Leyuan Fang. Deep hyperspectral image sharpening. IEEE transactions on neural networks and learning systems, 29(11):5345-5355, 2018. 2 +[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2 +[10] Ying Fu, Tao Zhang, Yinqiang Zheng, Debing Zhang, and Hua Huang. Hyperspectral image super-resolution with optimized rgb guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2019. 2 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6 + +[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 8 +[15] Jin-Fan Hu, Ting-Zhu Huang, Liang-Jian Deng, Hong-Xia Dou, Danfeng Hong, and Gemine Vivone. Fusformer: A transformer-based fusion network for hyperspectral image super-resolution. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2, 3, 6, 8 +[16] Zi-Rong Jin, Liang-Jian Deng, Tian-Jing Zhang, and Xiao-Xu Jin. Bam: Bilateral activation mechanism for image fusion. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4315-4323, 2021. 3, 6, 8 +[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2 +[18] Shutao Li, Renwei Dian, Leyuan Fang, and José M Bioucas-Dias. Fusing hyperspectral and multispectral images via coupled sparse tensor factorization. IEEE Transactions on Image Processing, 27(8):4118-4130, 2018. 1 +[19] Ying Li, Haokui Zhang, and Qiang Shen. Spectral-spatial classification of hyperspectral imagery with 3d convolutional neural network. Remote Sensing, 9(1):67, 2017. 2 +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5 +[21] Gamba Paolo. Pavia centre and university. https://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Sciences, 2011. 6 +[22] Ying Qu, Hairong Qi, and Chiman Kwan. Unsupervised sparse dirichlet-net for hyperspectral image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2511-2520, 2018. 2 +[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 8 +[24] Yue Shi, Liangxiu Han, Lianghao Han, Sheng Chang, Tongle Hu, and Darren Dancey. A latent encoder coupled generative adversarial network (le-gan) for efficient hyperspectral image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2022. 2 +[25] Dong Wang, Yunpeng Bai, Bendu Bai, Chanyue Wu, and Ying Li. Heterogeneous two-stream network with hierarchical feature prefusion for multispectral pan-sharpening. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1845-1849. IEEE, 2021. 1 + +[26] Dong Wang, Yunpeng Bai, Chanyue Wu, Ying Li, Changjing Shang, and Qiang Shen. Convolutional LSTM-based hierarchical feature fusion for multispectral pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1 +[27] Dong Wang, Pei Zhang, Yunpeng Bai, and Ying Li. Metapan: Unsupervised adaptation with meta-learning for multispectral pansharpening. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1 +[28] Xiuheng Wang, Jie Chen, and Cedric Richard. Hyperspectral image super-resolution with deep priors and degradation model inversion. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2814-2818. IEEE, 2022. 6, 8 +[29] Qi Wei, José Bioucas-Dias, Nicolas Dobigeon, and Jean-Yves Tourneret. Hyperspectral and multispectral image fusion based on a sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 53(7):3658-3668, 2015. 1 +[30] Zhenyu Wu, Lin Wang, Wei Wang, Tengfei Shi, Chenglizhao Chen, Aimin Hao, and Shuo Li. Synthetic data supervised salient object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5557-5565, 2022. 2 +[31] Qi Xie, Minghao Zhou, Qian Zhao, Zongben Xu, and Deyu Meng. Mhf-net: An interpretable deep network for multispectral and hyperspectral image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 6, 8 +[32] Fengchao Xiong, Jun Zhou, and Yuntao Qian. Material based object tracking in hyperspectral videos. IEEE Transactions on Image Processing, 29:3719-3733, 2020. 1 +[33] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2 +[34] Ting Xu, Ting-Zhu Huang, Liang-Jian Deng, Xi-Le Zhao, and Jie Huang. Hyperspectral image superresolution using unidirectional total variation with tucker decomposition. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13:4381-4398, 2020. 6, 8 +[35] Xizhe Xue, Haokui Zhang, Bei Fang, Zongwen Bai, and Ying Li. Grafting transformer on automatically designed convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1 +[36] Jing Yang, Chanyue Wu, Tengfei You, Dong Wang, Ying Li, Changjing Shang, and Qiang Shen. Hierarchical spatio-spectral fusion for hyperspectral image super resolution via sparse representation and pre-trained deep model. Knowledge-Based Systems, 260:110170, 2023. 1 +[37] Jingxiang Yang, Yong-Qiang Zhao, and Jonathan Cheung-Wai Chan. Hyperspectral and multispectral image fusion via deep two-branches convolutional neural network. Remote Sensing, 10(5):800, 2018. 6, 9 +[38] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum. + +IEEE transactions on image processing, 19(9):2241-2253, 2010.6 +[39] Naoto Yokoya and Akira Iwasaki. Airborne hyperspectral data over chikusei. Space Appl. Lab., Univ. Tokyo, Tokyo, Japan, Tech. Rep. SAL-2016-05-27, 5, 2016. 6 +[40] Tengfei You, Chanyue Wu, Yunpeng Bai, Dong Wang, Huibin Ge, and Ying Li. Hmf-former: Spatio-spectral transformer for hyperspectral and multispectral image fusion. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2022. 2, 3 +[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5 +[42] Haokui Zhang, Chengrong Gong, Yunpeng Bai, Zongwen Bai, and Ying Li. 3-d-anas: 3-d asymmetric neural architecture search for fast hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021. 1 +[43] Lei Zhang, Jiangtao Nie, Wei Wei, Yanning Zhang, Shengcai Liao, and Ling Shao. Unsupervised adaptation learning for hyperspectral imagery super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2020. 2, 3, 6, 8 +[44] Meilin Zhang, Xiongli Sun, Qiqi Zhu, and Guizhou Zheng. A survey of hyperspectral image super-resolution technology. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, pages 4476-4479. IEEE, 2021. 1 +[45] Xueting Zhang, Wei Huang, Qi Wang, and Xuelong Li. Ssrnet: Spatial-spectral reconstruction network for hyperspectral and multispectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5953-5965, 2020. 3 +[46] Ke Zheng, Lianru Gao, Wenzhi Liao, Danfeng Hong, Bing Zhang, Xin Cui, and Jocelyn Chanussot. Coupled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super resolution. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2487-2502, 2020. 2 \ No newline at end of file diff --git a/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/images.zip b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..34cb3e5a26b520825f71313ddf80da4e549601e9 --- /dev/null +++ b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:343f74fa4572735065d578bfb054c77a6cde179573b6e667bdd4ed0f1f882be5 +size 978633 diff --git a/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/layout.json b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..062bc5d1b14b78fbce061c3b6864c1d235804782 --- /dev/null +++ b/2023/HSR-Diff_ Hyperspectral Image Super-Resolution via Conditional Diffusion Models/layout.json @@ -0,0 +1,13573 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 532, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 532, + 138 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 532, + 138 + ], + "type": "text", + "content": "HSR-Diff: Hyperspectral Image Super-Resolution via Conditional Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "spans": [ + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "text", + "content": "Chanyue Wu" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "text", + "content": ", Dong Wang" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "text", + "content": ", Yunpeng Bai" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "text", + "content": ", Hanyu Mao" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "text", + "content": ", Ying Li" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "text", + "content": ", Qiang Shen" + }, + { + "bbox": [ + 92, + 160, + 499, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 175, + 533, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 175, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 60, + 175, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 175, + 533, + 217 + ], + "type": "text", + "content": "National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, Shaanxi Provincial Key Laboratory of Speech & Image Information Processing, School of Computer Science, Northwestern Polytechnical University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 217, + 518, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 217, + 518, + 244 + ], + "spans": [ + { + "bbox": [ + 74, + 217, + 518, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 217, + 518, + 244 + ], + "type": "text", + "content": "Department of Computer Science, Faculty of Business and Physical Sciences, Aberystwyth University, U.K." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 246, + 509, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 246, + 509, + 259 + ], + "spans": [ + { + "bbox": [ + 83, + 246, + 509, + 259 + ], + "type": "text", + "content": "{chanyuewu, dongwang, maomhy} " + }, + { + "bbox": [ + 83, + 246, + 509, + 259 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 83, + 246, + 509, + 259 + ], + "type": "text", + "content": " mail.nwpu.edu.cn lybyp@nwpu.edu.cn {yub3, qqs} " + }, + { + "bbox": [ + 83, + 246, + 509, + 259 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 83, + 246, + 509, + 259 + ], + "type": "text", + "content": " aber.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 312, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 290, + 540 + ], + "type": "text", + "content": "Despite the proven significance of hyperspectral images (HSIs) in performing various computer vision tasks, its potential is adversely affected by the low-resolution (LR) property in the spatial domain, resulting from multiple physical factors. Inspired by recent advancements in deep generative models, we propose an HSI Super-resolution (SR) approach with Conditional Diffusion Models (HSR-Diff) that merges a high-resolution (HR) multispectral image (MSI) with the corresponding LR-HSI. HSR-Diff generates an HR-HSI via repeated refinement, in which the HR-HSI is initialized with pure Gaussian noise and iteratively refined. At each iteration, the noise is removed with a Conditional Denoising Transformer (CDFformer) that is trained on denoising at different noise levels, conditioned on the hierarchical feature maps of HR-MSI and LR-HSI. In addition, a progressive learning strategy is employed to exploit the global information of full-resolution images. Systematic experiments have been conducted on four public datasets, demonstrating that HSR-Diff outperforms state-of-the-art methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 667 + ], + "type": "text", + "content": "Hyperspectral images (HSIs) contain dozens or hundreds of spectral bands, enabling them to provide more faithful knowledge of targeted scenes than conventional imaging modalities. As such, HSIs play an irreplaceable role in various computer vision tasks, including classification [35, 42], segmentation [6], and tracking [32]. Although HSIs contain rich spectral information, contemporary hyperspectral" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 285, + 524, + 387 + ], + "blocks": [ + { + "bbox": [ + 331, + 285, + 524, + 387 + ], + "lines": [ + { + "bbox": [ + 331, + 285, + 524, + 387 + ], + "spans": [ + { + "bbox": [ + 331, + 285, + 524, + 387 + ], + "type": "image", + "image_path": "8da2ed340bcb04d90a85f170285672c7a60e68bf0cfdbd485b9764fa5cf1f755.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 420 + ], + "type": "text", + "content": "Figure 1: Illustration of fusion-based Hyperspectral Image Super-Resolution (HSI-SR)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 546, + 533 + ], + "type": "text", + "content": "imaging sensors lack high-resolution (HR) in the spatial domain, due to the stringent constraint of typically low signal-to-noise ratios. Their widespread use is significantly hindered by this fact. Restricted by hardware limitations, a practical way to work around this problem is to fuse the low-resolution (LR) HSI with an HR multispectral image (MSI). This requires the implementation of so-called HSI super-resolution (SR)[36], as shown in Figure 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Over the past few decades, a significant amount of research efforts have been devoted to developing HSI-SR approaches, which can be roughly classified into five categories [44]: Extensions of pansharpening [5, 25, 27], Bayesian inference-based methods [1, 29], matrix factorization-based methods [2], tensor-based methods [18], and deep learning (DL)-based methods [31]. Whilst pansharpening methods [26] have been extended to the field of HSI-SR, such approaches are prone to spectral distortion. Bayesian inference-based approaches rely on the assumption of prior knowledge, thereby having a weak flexibility in dealing with different HSI structures. Matrix factorization-based techniques reshape the 3D HSIs and MSIs into matrices, thus facing the challenge of learning the required relationship between space and spectrum. Although sev-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 674, + 124, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 124, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 124, + 684 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 684, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 684, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 684, + 287, + 713 + ], + "type": "text", + "content": "†Corresponding author. This work is supported by the National Natural Science Foundation of China under Grant (62271400), and in part by the Shaanxi Provincial Key R&D Program, China (2023-GHZD-02)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7083" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "eral tensor-based methods have been proposed that can maintain the 3D structure of input images, they consume much more memory and computational power. Furthermore, these traditional approaches work via relying heavily on hand-crafted priors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 288, + 327 + ], + "type": "text", + "content": "Recently, DL-based methods, especially convolutional neural network (CNN)-based approaches, have flooded over into the HSI-SR research community [8, 31, 10, 22, 46, 43, 19]. Rather than resorting to hand-crafted features, DL-based techniques learn prior knowledge automatically from given data. Particularly, Dong et al. proposed the first DL-based method for image SR, with the end-to-end mapping between LR images and HR images learned using a CNN [9]. Subsequently, generative adversarial networks (GANs) were introduced to the field of image SR in an effort to produce high-frequency details [17, 11]. After that, various GAN-based models have been devised, showing state-of-the-art results in the HSI-SR literature [24]. However, such work requires carefully designed regularization and optimization tricks to tame optimization instability and avoid mode collapse." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 328, + 288, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 288, + 508 + ], + "type": "text", + "content": "Inspired by the recent developments in deep generative models [14, 23], in this paper, we propose an innovative approach that we refer to as HSR-Diff (HSI-SR with conditional diffusion models). It works by learning to transform the original standard normal distribution into the data distribution of HR-HSI through a sequence of refinements. In contrast to GAN-based methods which require inner-loop maximization, HSR-Diff operates by simply minimizing a well-defined loss function to learn prior knowledge. Although conditional diffusion models are straightforward to define and efficient to train, there has been no demonstration that they are capable of merging LR-HSI and HR-MSI to the best of our knowledge. We show that conditional diffusion models are capable of generating high-quality HR-HSIs, which may best the state-of-the-art results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "A key factor of HSR-Diff is its inherent denoising ability thanks to use of deep neural networks. In spite of the effectiveness of CNNs for denoising, they have shown limitations in modelling long-range dependencies. Some previous transformer-based HSI-SR works, especially Pyramid Shuffle-and-Reshuffle Transformer (PSRT) [7], Transformer-based Fusion network (Fusformer) [15], and Hyperspectral and Multispectral image Fusion Transformer (HMF-Former) [40], all of which learn long-range dependencies with attention mechanisms. These Transformers only take LR-HSI and HR-MSI as input, while the denoising Transformer is also fed with noise level " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " and noisy HR-HSI. To comprehensively exploit the spatial and spectral information in the LR-HSI, HR-MSI, and noisy HR-HSI, a Conditional Denoising Transformer (CD-Former) is herein designed and trained with a denoising objective to remove various levels of noise iteratively. In ad" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "dition, a progressive learning strategy is utilized to help the CDFormer learn the global statistics of full-resolution HSIs. The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 131, + 545, + 434 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 131, + 545, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 131, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 317, + 131, + 545, + 214 + ], + "type": "text", + "content": "- We propose the novel application of conditional diffusion models in the field of HSI-SR that works by progressively destroying HR-HSI through injecting noise and subsequently learning to reverse this process, in order to perform HSI-SR. To the best of our knowledge, this is the first work to exploit diffusion models within the realm of HSI-SR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 224, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 224, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 545, + 308 + ], + "type": "text", + "content": "- We introduce a CDFormer that refines a noisy HR-HSI conditioned on the deep feature maps of HR-MSI and LR-HSI, capable of modelling global connectivity with self-attention. Rather than concatenation, cross-attention is utilized as the conditioning mechanism to incorporate spatio-spectral information and noise level." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 318, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 318, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 317, + 318, + 545, + 377 + ], + "type": "text", + "content": "- We employ a progressive learning strategy to exploit the global information of full-resolution HSIs, with CDFormer being trained on small image patches in the early epochs with high efficiency and on the global images in the later epochs to acquire global information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 387, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 387, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 317, + 387, + 545, + 434 + ], + "type": "text", + "content": "- We present experimental investigations on four public datasets, with quantitative and qualitative results illustrating the superior performance of our approach as compared with state-of-the-art methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 447, + 392, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 392, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 392, + 459 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 467, + 408, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 408, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 408, + 479 + ], + "type": "text", + "content": "2.1. Diffusion models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 653 + ], + "type": "text", + "content": "Typical deep generative models include autoregressive models (AR), normalizing flows (NF), variational autoencoders (VAE), GANs, diffusion models, etc. Over the past decade, GANs have gained prominence as a pivotal technology in the challenging field of image synthesis. However, this landscape has seen a transformation with the emergence of diffusion models, which have rapidly become a prominent research avenue across diverse domains, such as natural language processing, image processing, and computer vision. Diffusion models have arisen as an innovative approach that operates by iteratively refining a noise vector into the desired output. This reversible process allows for the generation of high-quality samples and has proven versatile in various applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The development of diffusion models has seen a dramatically accelerating pace over the past three years. In the field of computer vision, diffusion models have shown promise in tasks like image segmentation [33], object detection [30], and even video processing [13]. Whilst diffusion models" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7084" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 124 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 124 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 124 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 124 + ], + "type": "image", + "image_path": "d388835a1cd9e49b9d7d217001d4d1efca42f8981acd60e33c519b9ec9521510.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 130, + 287, + 179 + ], + "lines": [ + { + "bbox": [ + 46, + 130, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 130, + 287, + 179 + ], + "type": "text", + "content": "Figure 2: Forward and reverse processes of HSR-Diff, with forward process " + }, + { + "bbox": [ + 46, + 130, + 287, + 179 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 130, + 287, + 179 + ], + "type": "text", + "content": " generating an HSI sequence (left to right) by gradually adding Gaussian noise, and reverse process " + }, + { + "bbox": [ + 46, + 130, + 287, + 179 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 130, + 287, + 179 + ], + "type": "text", + "content": " iteratively refining HR-HSI (right to left)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 288, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 261 + ], + "type": "text", + "content": "have shown great potential for a variety of computer vision applications, none of them have yet been devoted to the problem of HSI-SR to the best of our knowledge. In this paper, we extend the utility of diffusion models to the field of HSI-SR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 270, + 211, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 211, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 211, + 283 + ], + "type": "text", + "content": "2.2. Deep Learning-Based HSI-SR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 289, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 540 + ], + "type": "text", + "content": "In recent years, data-driven CNN architectures have been shown to outperform traditional approaches for use in the HSI-SR literature. These methods formulate the underlying fusion problem as a highly nonlinear mapping that takes HR-MSIs and LR-HSIs as input to generate an optimal HR-HSI. For example, CMHF-net [31] is an interpretable CNN, the design of which exploits the deep unfolding technique. Zhang et al. [43] proposed to reconstruct HR-HSIs with a two-stage network, while Zhang et al. [45] designed an interpretable spatial-spectral reconstruction network (SSR-NET) based on CNN. Aiming at problems of inflexible structure and information distortion, Jin et al. embedded Bilateral Activation Mechanism into ResNet, resulting in the effective model of BRResNet [16]. Thanks to the inductive bias of CNN, such as locality and weight sharing, these methods can provide good generalization performance and achieve impressive results. Nevertheless, CNNs have limitations in capturing long-range dependencies and self-similarity priors. To overcome such shortcomings, some Transformer-based HSI-SR approaches [15, 40] have been proposed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 541, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 601 + ], + "type": "text", + "content": "Within this study, the CDFormer serves as the denoising network in conditional diffusion models. It harnesses self-attention to learn global statistics from full-resolution HSIs and MSIs and employs cross-attention as the conditioning mechanism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 613, + 180, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 180, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 180, + 627 + ], + "type": "text", + "content": "3. Proposed Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 171, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 171, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 171, + 645 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 687 + ], + "type": "text", + "content": "Without losing generality, the observation models for the HR-MSI and LR-HSI of interest can be mathematically formulated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 687, + 287, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 687, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 145, + 687, + 287, + 706 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = \\mathbf {R Z} \\tag {1}", + "image_path": "d39edfad03356b780a76763a0c10fea762a252cfe5b8b8e825752a41abf0bdfb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 704, + 189, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 704, + 189, + 714 + ], + "spans": [ + { + "bbox": [ + 146, + 704, + 189, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\mathbf {Z D},", + "image_path": "6eaf3709e7463494c5e2257edb95064f801f492ab7c5ad746caba8abdb545cbb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{b \\times HW}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " denotes the HR-MSI which consists of " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " spectral bands with a resolution of " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "HW" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " in the spatial domain; " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{b \\times B}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " represents the spectral response function of HR-MSI; " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{B \\times hw}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " denotes the LR-HSI; and " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{B \\times HW}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " is the HR-HSI. In the above, " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " are the numbers of bands, with " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " being the band height, and " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " the width, where " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "b \\ll B" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "h \\ll H" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "w \\ll W" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^{HW \\times hw}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " is the spatial response of the LR-HSI, which can be modelled with blurring and down-sampling operations. The HSI-SR can be interpreted as an inverse problem for merging a practically collected " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " and an observed " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " to produce a latent " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": ". In this paper, the ideal " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " is restored with HSR-Diff conditioned on spatio-spectral information of " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 71, + 547, + 240 + ], + "type": "text", + "content": ", the details of which are described below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 249, + 529, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 249, + 529, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 529, + 262 + ], + "type": "text", + "content": "3.2. HSI-SR with Conditional Diffusion Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": "Given a dataset " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train} = \\{\\mathbf{X}^i,\\mathbf{Y}^i,\\mathbf{Z}^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": " satisfying a certain joint probability distribution " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "p(\\mathbf{X},\\mathbf{Y},\\mathbf{Z})" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": ", many pairs of " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{Y})" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": " may be consistent with the same " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": ". Thus, the HR-HSI " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": " can be obtained with iterative refinement that provide an approximate to " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "p(\\mathbf{Z}|\\mathbf{X},\\mathbf{Y})" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": ". In this work, we implement the process of iterative refinement with HSR-Diff, where the optimized HR-HSI is presumed to be produced in " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": " refinement steps. In HSR-Diff, the target HR-HSI is initialized with a pure noise " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": " as shown in Figure 2. The HSI is then refined iteratively according to learned conditional distributions " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{Z}_{t - 1}|\\mathbf{Z}_t,\\mathbf{X},\\mathbf{Y})" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": ". In so doing, the image sequence " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z}_{T - 1},\\mathbf{Z}_{T - 2},\\dots ,\\mathbf{Z}_0)" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": " can be attained and ultimately " + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0\\sim p(\\mathbf{Z}|\\mathbf{X},\\mathbf{Y})" + }, + { + "bbox": [ + 304, + 268, + 545, + 424 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": "The HSR-Diff makes use of two processes: a forward process that perturbs HR-HSI " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": " to noise, and a reverse process converting noise back to HR-HSI " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ". In the forward process, the intermediate images, i.e., " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{T - 1}" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{T - 2}" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\dots" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_1" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ", are generated according to a Markov chain with fixed transition probability " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "q(\\mathbf{Z}_t|\\mathbf{Z}_{t - 1})" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ". We are interested in reversing the process via iterative refinement, in which the noise is reduced iteratively with a reverse Markov chain conditioned on " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ". The reverse chain is learned with the CDFformer " + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 426, + 546, + 557 + ], + "type": "text", + "content": ". Further details of HSR-Diff's working are given below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 576, + 411, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 576, + 411, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 576, + 411, + 587 + ], + "type": "text", + "content": "3.2.1 Forward Process" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "text", + "content": "Inspired by [14], forward process " + }, + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "text", + "content": " iteratively adds Gaussian noise to " + }, + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 596, + 545, + 620 + ], + "type": "text", + "content": " iterations:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 328, + 630, + 545, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 630, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 328, + 630, + 545, + 662 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {Z} _ {1: T} \\mid \\mathbf {Z} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {t - 1}\\right) \\tag {2}", + "image_path": "34425ad4cafed5d35023c5f83d38a3b658360bc09a464caf951eeb9737a753e4.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 329, + 665, + 522, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 665, + 522, + 679 + ], + "spans": [ + { + "bbox": [ + 329, + 665, + 522, + 679 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {Z} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right),", + "image_path": "3a541b61edebb96ba96236209424913fd8ba2b437bf5db07a3abe1e272fc5a65.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1:T} \\in (0,1)" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": " are scalar hyper-parameters. Note that in the forward process, the distribution of " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": " can" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7085" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 541, + 274 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 541, + 274 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 541, + 274 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 541, + 274 + ], + "type": "image", + "image_path": "244e86f93651d9b3ac4b788864e2c31da74688de21ac6075427a0351c496c294.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 280, + 422, + 293 + ], + "lines": [ + { + "bbox": [ + 170, + 280, + 422, + 293 + ], + "spans": [ + { + "bbox": [ + 170, + 280, + 422, + 293 + ], + "type": "text", + "content": "Figure 3: Architecture of Conditional Denoising Transformer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 313, + 262, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 262, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 262, + 326 + ], + "type": "text", + "content": "be directly sampled in closed form. This implies that" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 338, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 338, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 80, + 338, + 287, + 354 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {Z} _ {t} \\mid \\mathbf {Z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t}; \\sqrt {\\gamma_ {t}} \\mathbf {Z} _ {0}, (1 - \\gamma_ {t}) \\mathbf {I}\\right), \\tag {3}", + "image_path": "d76483546a15a2346ab940614389ba3c6ab96f1ee697d08977612db48cb20ffa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\gamma_{t} = \\prod_{i = 1}^{t}\\alpha_{i}" + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "text", + "content": ". In addition, the posterior distribution of " + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{t - 1}" + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t" + }, + { + "bbox": [ + 47, + 365, + 287, + 392 + ], + "type": "text", + "content": " can be derived by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 402, + 230, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 402, + 230, + 417 + ], + "spans": [ + { + "bbox": [ + 70, + 402, + 230, + 417 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {0}, \\mathbf {Z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t - 1}; \\boldsymbol {\\mu}, \\sigma^ {2} \\mathbf {I}\\right)", + "image_path": "745ec503da2c852febe1cadcb77de234c7e493961981e7d41583fcdea5a82636.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 418, + 287, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 418, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 72, + 418, + 287, + 445 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} = \\frac {\\sqrt {\\gamma_ {t - 1}} (1 - \\alpha_ {t})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {0} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\gamma_ {t - 1})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {t} \\tag {4}", + "image_path": "3ab28c587c933426db32d7558589e2ba59776a99d9ab9b3fd61105874dbce28b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 446, + 185, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 446, + 185, + 472 + ], + "spans": [ + { + "bbox": [ + 72, + 446, + 185, + 472 + ], + "type": "interline_equation", + "content": "\\sigma^ {2} = \\frac {(1 - \\gamma_ {t - 1}) (1 - \\alpha_ {t})}{1 - \\gamma_ {t}}.", + "image_path": "3e2771648c74ad6bc3b1c05dfb5f846cee0c759520ee57bb0fb24956209c9eba.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 483, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 508 + ], + "type": "text", + "content": "As the CDFormer is tasked with predicting " + }, + { + "bbox": [ + 47, + 483, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 47, + 483, + 287, + 508 + ], + "type": "text", + "content": ", the posterior is useful in the reverse process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 532, + 198, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 198, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 198, + 544 + ], + "type": "text", + "content": "3.2.2 Reverse Markovian Process" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "content": "The reverse process inferences " + }, + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "content": " via iterative refinement. It starts from a pure Gaussian noise " + }, + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_T" + }, + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "content": " and goes in the opposite direction of the forward process:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 603, + 277, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 277, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 277, + 635 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {Z} _ {0: T} \\mid \\mathbf {X}, \\mathbf {Y}\\right) = p \\left(\\mathbf {Z} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {t}, \\mathbf {X}, \\mathbf {Y}\\right)", + "image_path": "2d5ac27fe727708e765d7df91568ae92e920206deebdd6859983cf85b21945fc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 637, + 200, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 637, + 200, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 200, + 651 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {Z} _ {T}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {T}; \\mathbf {0}, \\mathbf {I}\\right)", + "image_path": "2a027ebb2a097e801651a6e5c405244de2696d6e053b53cd2e08b5f254dfad0d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 653, + 289, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 289, + 677 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {Z} _ {t - 1} \\mid \\mathbf {Z} _ {t}, \\mathbf {X}, \\mathbf {Y}\\right) = \\mathcal {N} \\left(\\mathbf {Z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {X}, \\mathbf {Y}, \\mathbf {Z} _ {t}, \\gamma_ {t}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {5}", + "image_path": "2e8327e5691e28e221289f443e202a2045807e85217dffb21618e0e243705a2e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "where the distribution " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{Z}_{t - 1}\\mid \\mathbf{Z}_t,\\mathbf{X},\\mathbf{Y})" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " is parameterized with " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ". Note that the CDFormer provides a prediction of " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}_0" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ". Thus, according to (4), each refinement step takes" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 314, + 388, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 388, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 388, + 326 + ], + "type": "text", + "content": "the following form:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 335, + 545, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 335, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 335, + 335, + 545, + 388 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Z} _ {t - 1} = \\frac {\\sqrt {\\gamma_ {t - 1}} (1 - \\alpha_ {t})}{1 - \\gamma_ {t}} f _ {\\theta} (\\mathbf {X}, \\mathbf {Y}, \\mathbf {Z} _ {t}, \\gamma_ {t}) \\tag {6} \\\\ + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\gamma_ {t - 1})}{1 - \\gamma_ {t}} \\mathbf {Z} _ {t} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon}, \\\\ \\end{array}", + "image_path": "0ec7ea7a2e88d6d9ea0d8c9f5ce794af2131fd612b88a54597abc9bdd4940f25.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 397, + 488, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 488, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 488, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 397, + 488, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 306, + 397, + 488, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 397, + 488, + 410 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 306, + 397, + 488, + 410 + ], + "type": "text", + "content": " is the CDFormer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 426, + 404, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 404, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 404, + 437 + ], + "type": "text", + "content": "3.2.3 Noise Schedule" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": "Inspired by the research reported in [4], we sample " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": " with two steps during training. In particular, we first sample a time step " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "t\\sim U\\{1,T\\}" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": " and then randomly select " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\gamma \\sim U(\\gamma_{t - 1},\\gamma_t)" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": ". As such, " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\gamma \\sim p(\\gamma) = \\sum_{t = 1}^{T}\\frac{1}{T} U(\\gamma_{t - 1},\\gamma_t)" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": ". Normally, the model with a large " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": " can achieve better results. However, we find (through empirical investigations) that the performance is not very sensitive to the exact values of " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": ". Therefore, no hyper-parameter search about " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": " is conducted and we set " + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "inline_equation", + "content": "T = 2000" + }, + { + "bbox": [ + 305, + 445, + 545, + 579 + ], + "type": "text", + "content": " for simplicity. As for the inference process, we set the maximum generation iterations to 100, employing a linear noise schedule." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 586, + 495, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 495, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 495, + 600 + ], + "type": "text", + "content": "3.3. Conditional Denoising Transformer" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The property of non-local self-similarity of HSIs is often exploited in denoising tasks but is usually not well captured by CNN-based models. Due to the effectiveness of Transformer layer in capturing non-local long-range dependencies, the potential of Transformer is explored in conditional denoising of HSI. Unfortunately, the vanilla Transformer focuses only on spatial relationships between pixels while neglecting the spectral dimension. Besides, denoising networks in conditional diffusion models normally concatenate" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7086" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 269, + 135 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 269, + 135 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 269, + 135 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 269, + 135 + ], + "type": "image", + "image_path": "d62b5e0682aa49c78a1d8b309984db45471b2bc594e44417c5d6f86b2d2975d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 143, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 46, + 143, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 287, + 167 + ], + "type": "text", + "content": "Figure 4: Spatio-Spectral Transformer Layer, where \"LN\" denotes layer normalization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 286, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 286, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 286, + 250 + ], + "type": "text", + "content": "all images together as input, which may hinder the extraction of useful spatio-spectral information in LR-HSIs and HR-MSIs. Hence, the CDFormer adopts a two-stream architecture and is constructed with stacked Spatio-Spectral Transformer Layers (S" + }, + { + "bbox": [ + 46, + 190, + 286, + 250 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 190, + 286, + 250 + ], + "type": "text", + "content": "TLs)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": "The architecture of the CDFormer is shown in Figure 3. The SR stream first utilizes a " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": " convolution to generate low-level feature embeddings " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0^{SR}" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": " and then transforms it into deep features " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{SR}" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": " with stacked-" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathbf{TLs}" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": ". Instead of adopting " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": " as done in the existing work [4], our method is conditioned on " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": " directly to achieve efficient generation. The denoising stream contains multiple noise-aware conditional " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathbf{TLs}" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": " (NC-" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathbf{TLs}" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": ") that take as input the embedded noise level and the image representation " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{SR}" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": ". The Reconstruction module is set to produce a noise-free HR-HSI, by employing residual learning to alleviate the difficulty of HR-HSI generation while mapping the features onto HR-HSI via a " + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": " convolution and addition operation. It should be noted that no downsampling operation is employed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 181, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 181, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 181, + 449 + ], + "type": "text", + "content": "3.3.1 Noise Level Embedding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 457, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 517 + ], + "type": "text", + "content": "The noise level offers essential information for denoising models. Inspired by the work of [4], we embed noise level within the models with sinusoidal positional encoding. The process of noise level embedding (NLE) can be formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 525, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 525, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 102, + 525, + 287, + 552 + ], + "type": "interline_equation", + "content": "N L E _ {\\gamma , 2 i} = \\sin \\left(\\gamma / 1 0 0 0 0 ^ {2 i / C}\\right) \\tag {7}", + "image_path": "d6b86c9ea039a896e141a33420c57fc86bc129c6fdb5834b796233a92751508d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 548, + 241, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 548, + 241, + 569 + ], + "spans": [ + { + "bbox": [ + 94, + 548, + 241, + 569 + ], + "type": "interline_equation", + "content": "N L E _ {\\gamma , 2 i + 1} = \\cos \\left(\\gamma / 1 0 0 0 0 ^ {2 i / C}\\right),", + "image_path": "780b11c7ace116065c9ae4ef0e2ff28358fe0c7df00ff38349e055689b6454cb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "type": "text", + "content": " is the number of channels of " + }, + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathrm{TLs}" + }, + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 47, + 578, + 285, + 591 + ], + "type": "inline_equation", + "content": "i\\in [1,C / 2]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 609, + 235, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 235, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 235, + 623 + ], + "type": "text", + "content": "3.3.2 Spatio-Spectral Transformer Layers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Figure 4 illustrates the architecture of one " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^2\\mathrm{TL}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ", which consists of a Spatial Multi-head Self-Attention (SpatioMSA), a Spectral Multi-head Self-Attention (SpectralMSA), and a Feed Forward Network (FFN). SpatioMSA and SpectralMSA learn the interactions of spatial regions and interspectra relationships, respectively. To alleviate the computational burden, we adopt the transposed attention [41] in" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 345, + 69, + 507, + 144 + ], + "blocks": [ + { + "bbox": [ + 345, + 69, + 507, + 144 + ], + "lines": [ + { + "bbox": [ + 345, + 69, + 507, + 144 + ], + "spans": [ + { + "bbox": [ + 345, + 69, + 507, + 144 + ], + "type": "image", + "image_path": "3a284d97c03c04043a8c65162228a54afa8c9a537273be18c20fe4933401ecfd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 152, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 199 + ], + "type": "text", + "content": "Figure 5: Noise-Aware Conditional Spatio-Spectral Transformer Layer, where \"LT\" represents linear transform, \"LN\" denotes layer normalization, and \"MCA\" is the abbreviation of multi-head cross-attention." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 221, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 545, + 270 + ], + "type": "text", + "content": "SpectralMSA. SpatioMSA applies the popular window partitioning strategy [20] to reduce the computational complexity. In addition, the gating mechanism [41] is employed in the implementation of FFN." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 285, + 476, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 476, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 476, + 296 + ], + "type": "text", + "content": "3.3.3 Noise-Aware Conditional " + }, + { + "bbox": [ + 306, + 285, + 476, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathbf{T}\\mathbf{L}\\mathbf{s}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": "To condition the overall model on the hierarchical features of SR stream, we feed " + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{SR}" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": " to the Noise-Aware Conditional S" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": "TLs (NC-S" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": "TLs), each of which is a key building block of the denoising stream. Figure 5 depicts the structure of an NC-S" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": "TL, which takes as input the NLE (a vector), " + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{SR}" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{DS}" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{SR}" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{DS}" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": " have the same spatial resolution. NLE is first transformed and merged with " + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{DS}" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": " with the result subsequently processed with the means of multi-head cross-attention (MCA) [3] in order to condition the model on " + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l^{SR}" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": ". As a result, each S" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 305, + 304, + 545, + 437 + ], + "type": "text", + "content": "TL learns the spatio-spectral dependencies." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 443, + 470, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 470, + 456 + ], + "type": "text", + "content": "3.4. Progressive Learning Strategy" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 461, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 688 + ], + "type": "text", + "content": "Deep learning-based HSI-SR models are normally trained on small cropped patches. However, on one hand, training the proposed CDFormer on fixed-size image patches may not appropriately reflect the global relationships, which may result in suboptimal performance on full-resolution images when used. On the other hand, during the training phase, CDFormer can not be uploaded to Graphics Processing Unit (GPU) because of the limited memory. With this objective in mind, we adopt a strategy of progressive learning. In the initial epochs, the network is trained using smaller image patches. Subsequently, as training progresses, we transition towards utilizing larger patches, and ultimately, full-resolution images in the concluding training epochs. The resulting model, cultivated through the integration of varying sizes via progressive learning, exhibits notable enhancements in performance during testing. This is particularly evident when handling images of diverse resolutions, a scenario frequently encountered in the realm of the HSI-SR task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "To reduce the pressure on the demand of GPU memory, we only train the second half of CDFormer on full" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7087" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "resolution images. The loss function used for such training is defined as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 105, + 286, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 105, + 286, + 136 + ], + "spans": [ + { + "bbox": [ + 56, + 105, + 286, + 136 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\| \\mathbf {X} - \\mathbf {R} \\hat {\\mathbf {Z}} _ {0} \\| _ {1} + \\| \\mathbf {Y} - \\hat {\\mathbf {Z}} _ {0} \\mathbf {D} \\| _ {1} + \\| \\mathbf {Z} _ {0} - \\hat {\\mathbf {Z}} _ {0} \\| _ {1} \\tag {8} \\\\ \\hat {\\mathbf {Z}} _ {0} = f _ {\\theta} (\\sqrt {\\gamma} \\mathbf {Z} _ {0} + \\sqrt {1 - \\gamma} \\boldsymbol {\\epsilon}, \\mathbf {X}, \\mathbf {Y}), \\\\ \\end{array}", + "image_path": "040b64a2d867ababae9884bdabb7b4c61a90b1a640b0c0aee55d428dd1ba12dc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{Y},\\mathbf{Z})" + }, + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "type": "text", + "content": " is sampled from the training set, and the noise schedule about " + }, + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 145, + 286, + 205 + ], + "type": "text", + "content": " has been discussed above. The first two terms are designed according to the observation models, while the last one is based on the assumption of Laplace distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 218, + 127, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 127, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 127, + 231 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 239, + 286, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 275 + ], + "type": "text", + "content": "Systematic experiments are herein conducted on four commonly-used public-available HSI-SR datasets to demonstrate the effectiveness of the proposed approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 284, + 109, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 109, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 109, + 295 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 302, + 286, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 286, + 338 + ], + "type": "text", + "content": "Four datasets including CAVE [38], PaviaU [21], Chikusei [39], and HypSen [37] are used in our experiments, with the following details on each." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 355, + 286, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 286, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 286, + 451 + ], + "type": "text", + "content": "CAVE: There are 32 scenes with a spatial size of " + }, + { + "bbox": [ + 46, + 355, + 286, + 451 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 355, + 286, + 451 + ], + "type": "text", + "content": " in the CAVE dataset, where we select the first 20 HSIs for training, with the remaining 12 images used for testing. We generate LR-HSIs by Gaussian blur and down-sampling using a factor of 32 as done in [31]. HR-MSIs are acquired by integrating all HR-HSI bands according to the spectral response function of Nikon D700. The original HR-HSIs are treated as ground truth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 468, + 286, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 563 + ], + "type": "text", + "content": "PaviaU: Collected by the University of Pavia, Italy, the original HSI dataset consists of " + }, + { + "bbox": [ + 46, + 468, + 286, + 563 + ], + "type": "inline_equation", + "content": "610 \\times 340" + }, + { + "bbox": [ + 46, + 468, + 286, + 563 + ], + "type": "text", + "content": " pixels in which the top-left " + }, + { + "bbox": [ + 46, + 468, + 286, + 563 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 468, + 286, + 563 + ], + "type": "text", + "content": " area is extracted as the test data, with the remaining used for training. Except for water absorption bands, all other 103 bands are chosen for the experiments. Note that the down-sampling factor for the generation of LR-HSIs is four, and the spectral response function is the same as that of the WorldView-3 satellite." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Chikusei: This dataset consists of 128 bands with a spectral range of " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "363~nm" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "1018~nm" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " and a spatial resolution of " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "2517\\times 2335" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": ". The original HSI data was taken by an airborne visible and near-infrared imaging sensor over Chikusei, Japan. To alleviate the impact of the back boundary and noise, we crop the center area and remove noise bands. The processed image has a size of " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "2048\\times 2048\\times 110" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": ". The top half " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "1024\\times 2048\\times 110" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " area is selected as the training data, while the rest half is split into eight testing " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "512\\times 512\\times 110" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " patches. For the production of LR-HSIs and HR-MSIs, this dataset adopts the same processing as with PaviaU." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "HypSen: This dataset concerns a real scenario consisting of a " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "-resolution HSI and a " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "-resolution MSI. The Hyperion sensor on the Earth Observing-1 satellite provided the HSI with 242 spectral bands in the spectral range of " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "400\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "2500\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": ", and the MSI with 13 bands was captured by the Sentinel-2A satellite. The blue, green, red, and near-infrared bands of MSI in our experiments are selected due to their high spatial resolution. To eliminate the impact of noise and water absorption, we remove those relevant bands, with 84 bands remaining in the HSI. We crop sub-images of size " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "250\\times 330" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "750\\times 990" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " from the Hyperion HSI and Sentinel-2A MSI respectively, in our study, with the pairs of sub-image patches spatially registered." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 235, + 545, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 545, + 259 + ], + "type": "text", + "content": "4.2. Methods Compared and Evaluation Metrics Used" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 266, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 348 + ], + "type": "text", + "content": "Six state-of-the-art HSI-SR approaches are taken for comparison, including: UTV-TD [34], UAL [43], BRResNet [16], Fusformer [15], CMHF-Net [31], and UAL-DMI [28]. UTV-TD is a tensor-based technique; UAL, BRResNet, Fusformer, and CMHF-Net fall into the category of the DL-based methods; and UAL-DMI can be regarded as an upgraded version of UAL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 350, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 433 + ], + "type": "text", + "content": "Four quantitative quality metrics are employed for performance evaluation, including Peak Signal-to-Noise Ratio (PSNR), Spectral Angle Mapper (SAM), Erreur Relative Globale Adimensionnelle de Synthese (ERGAS, namely error relative global dimensionless synthesis), and Structure Similarity (SSIM). The smaller ERGAS and SAM are, the larger PSNR and SSIM are, the better the fusion result is." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 441, + 469, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 469, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 469, + 453 + ], + "type": "text", + "content": "4.3. Implementation Specification" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": "All DL-based methods are trained on the same datasets. For those compared methods, we use the publicly available source codes with default hyper-parameters as given in the corresponding research papers. Our HSR-Diff is implemented on the PyTorch framework. The learnable parameters of the CDFformer are initialized with Kaiming initialization [12] and trained on 2 NVIDIA GeForce GTX 3090s. The dimension of the embedding of " + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "inline_equation", + "content": "F_0^{SR}" + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "inline_equation", + "content": "F_0^{DS}" + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": " is set to 256, and the number of parameters of the CDFformer is 31.56M. We utilize the Adam optimizer with " + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": " to optimize the CDFformer. With limited GPU memory, the batch size is set to 4 and 2 for " + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "inline_equation", + "content": "128^2" + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": " images, respectively. It costs 20000 epochs on the CAVE and PaviaU datasets while consuming 5000 epochs on the Chikusei dataset. The learning rate is initialized as " + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 460, + 545, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 647, + 532, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 532, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 532, + 659 + ], + "type": "text", + "content": "4.4. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": "In this set of experiments, the evaluations are carried out using the first three datasets listed above without involving the real-world dataset, HypSen (which will be dealt with in the next section)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7088" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 110, + 130 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 110, + 130 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 110, + 130 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 110, + 130 + ], + "type": "image", + "image_path": "cba18c9ddba5abe00b2a6ddfbcb11954b588a839e2556220fcddf8f8e28c2ebf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 112, + 70, + 171, + 130 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 171, + 130 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 171, + 130 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 171, + 130 + ], + "type": "image", + "image_path": "bf86ea733028e5c49c7dc3cd3963201f0ac1c97201694df1ebfb1cb0c1cf78ae.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 233, + 130 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 233, + 130 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 233, + 130 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 233, + 130 + ], + "type": "image", + "image_path": "9e19c26d5f1691568c600a3db99e69618a6c904aeae17a99cd94a1696111aae0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 70, + 295, + 129 + ], + "blocks": [ + { + "bbox": [ + 235, + 70, + 295, + 129 + ], + "lines": [ + { + "bbox": [ + 235, + 70, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 235, + 70, + 295, + 129 + ], + "type": "image", + "image_path": "06006c40dcd41774640c94589a49c8dfe77201c56bc53e5e42d7bbb2b33004d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 358, + 129 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 358, + 129 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 358, + 129 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 358, + 129 + ], + "type": "image", + "image_path": "620a9d1cc61f2fb29768f5df4763c78f3f6c93c4c93ed9733072a159ffed88f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 70, + 419, + 129 + ], + "blocks": [ + { + "bbox": [ + 359, + 70, + 419, + 129 + ], + "lines": [ + { + "bbox": [ + 359, + 70, + 419, + 129 + ], + "spans": [ + { + "bbox": [ + 359, + 70, + 419, + 129 + ], + "type": "image", + "image_path": "7d557b71e6b2f660b7dff00195628a886a39a02b194227397046cc543380e87c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 421, + 70, + 481, + 129 + ], + "blocks": [ + { + "bbox": [ + 421, + 70, + 481, + 129 + ], + "lines": [ + { + "bbox": [ + 421, + 70, + 481, + 129 + ], + "spans": [ + { + "bbox": [ + 421, + 70, + 481, + 129 + ], + "type": "image", + "image_path": "92977a1bf99e675b1918955657bc05ec4e1403271aac5eaa2918f409990a0061.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 484, + 70, + 543, + 129 + ], + "blocks": [ + { + "bbox": [ + 484, + 70, + 543, + 129 + ], + "lines": [ + { + "bbox": [ + 484, + 70, + 543, + 129 + ], + "spans": [ + { + "bbox": [ + 484, + 70, + 543, + 129 + ], + "type": "image", + "image_path": "55c004a6a257bab4f7f3b70c6466379de314a974554fbeb4a284c487cc75ee60.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 51, + 137, + 110, + 197 + ], + "blocks": [ + { + "bbox": [ + 51, + 137, + 110, + 197 + ], + "lines": [ + { + "bbox": [ + 51, + 137, + 110, + 197 + ], + "spans": [ + { + "bbox": [ + 51, + 137, + 110, + 197 + ], + "type": "image", + "image_path": "107074d9837edb1f92e3e6afa931490806ff550fecff53675a16ff6dce68e0a6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 199, + 101, + 209 + ], + "lines": [ + { + "bbox": [ + 59, + 199, + 101, + 209 + ], + "spans": [ + { + "bbox": [ + 59, + 199, + 101, + 209 + ], + "type": "text", + "content": "(a)UTV-TD" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 112, + 137, + 171, + 197 + ], + "blocks": [ + { + "bbox": [ + 112, + 137, + 171, + 197 + ], + "lines": [ + { + "bbox": [ + 112, + 137, + 171, + 197 + ], + "spans": [ + { + "bbox": [ + 112, + 137, + 171, + 197 + ], + "type": "image", + "image_path": "bf6681f2aa80f1306ec5f258f396df9745e5a42604d9bb12017c660105eb98cb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 199, + 156, + 209 + ], + "lines": [ + { + "bbox": [ + 128, + 199, + 156, + 209 + ], + "spans": [ + { + "bbox": [ + 128, + 199, + 156, + 209 + ], + "type": "text", + "content": "(b) UAL" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 173, + 137, + 233, + 197 + ], + "blocks": [ + { + "bbox": [ + 173, + 137, + 233, + 197 + ], + "lines": [ + { + "bbox": [ + 173, + 137, + 233, + 197 + ], + "spans": [ + { + "bbox": [ + 173, + 137, + 233, + 197 + ], + "type": "image", + "image_path": "4b0234e0ea0389fc716690f082783c6e86cfb82da0fecc048a62e00bad743da8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 199, + 227, + 209 + ], + "lines": [ + { + "bbox": [ + 181, + 199, + 227, + 209 + ], + "spans": [ + { + "bbox": [ + 181, + 199, + 227, + 209 + ], + "type": "text", + "content": "(c) BRResNet" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 236, + 137, + 295, + 197 + ], + "blocks": [ + { + "bbox": [ + 236, + 137, + 295, + 197 + ], + "lines": [ + { + "bbox": [ + 236, + 137, + 295, + 197 + ], + "spans": [ + { + "bbox": [ + 236, + 137, + 295, + 197 + ], + "type": "image", + "image_path": "9c5ce54149a2397db2d2e59aca0e79b71eeadc7f1557cb4c1dd2369262c55c48.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 199, + 291, + 209 + ], + "lines": [ + { + "bbox": [ + 241, + 199, + 291, + 209 + ], + "spans": [ + { + "bbox": [ + 241, + 199, + 291, + 209 + ], + "type": "text", + "content": "(d) CMHF-Net" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 298, + 137, + 358, + 197 + ], + "blocks": [ + { + "bbox": [ + 298, + 137, + 358, + 197 + ], + "lines": [ + { + "bbox": [ + 298, + 137, + 358, + 197 + ], + "spans": [ + { + "bbox": [ + 298, + 137, + 358, + 197 + ], + "type": "image", + "image_path": "aa377e3121c60ee7e6a74835c36cf852ca856063994f2b7f2557d416f3f69e40.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 199, + 351, + 209 + ], + "lines": [ + { + "bbox": [ + 304, + 199, + 351, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 351, + 209 + ], + "type": "text", + "content": "(e) Fusformer" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 360, + 137, + 419, + 197 + ], + "blocks": [ + { + "bbox": [ + 360, + 137, + 419, + 197 + ], + "lines": [ + { + "bbox": [ + 360, + 137, + 419, + 197 + ], + "spans": [ + { + "bbox": [ + 360, + 137, + 419, + 197 + ], + "type": "image", + "image_path": "20269a988c05acb14b1bd1d06ed8c374f73ec7c3a73e4919d06afb917669df67.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 199, + 413, + 209 + ], + "lines": [ + { + "bbox": [ + 367, + 199, + 413, + 209 + ], + "spans": [ + { + "bbox": [ + 367, + 199, + 413, + 209 + ], + "type": "text", + "content": "(f) UAL-DMI" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 422, + 137, + 481, + 197 + ], + "blocks": [ + { + "bbox": [ + 422, + 137, + 481, + 197 + ], + "lines": [ + { + "bbox": [ + 422, + 137, + 481, + 197 + ], + "spans": [ + { + "bbox": [ + 422, + 137, + 481, + 197 + ], + "type": "image", + "image_path": "5d4e4ab7f93e145eca4839ff18a91678668521abcc305deee4a80fd40acebeab.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 199, + 474, + 209 + ], + "lines": [ + { + "bbox": [ + 429, + 199, + 474, + 209 + ], + "spans": [ + { + "bbox": [ + 429, + 199, + 474, + 209 + ], + "type": "text", + "content": "(g) HSR-Diff" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 484, + 137, + 543, + 197 + ], + "blocks": [ + { + "bbox": [ + 484, + 137, + 543, + 197 + ], + "lines": [ + { + "bbox": [ + 484, + 137, + 543, + 197 + ], + "spans": [ + { + "bbox": [ + 484, + 137, + 543, + 197 + ], + "type": "image", + "image_path": "90c273aa673a219ee3995e474f0872034fb801ede3aef075ecb307616887e0e1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 199, + 541, + 209 + ], + "lines": [ + { + "bbox": [ + 487, + 199, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 487, + 199, + 541, + 209 + ], + "type": "text", + "content": "(h) Ground truth" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 159, + 212, + 434, + 232 + ], + "blocks": [ + { + "bbox": [ + 159, + 212, + 434, + 232 + ], + "lines": [ + { + "bbox": [ + 159, + 212, + 434, + 232 + ], + "spans": [ + { + "bbox": [ + 159, + 212, + 434, + 232 + ], + "type": "image", + "image_path": "af089bc0248aba477c785501900a8bdaff7d901bc35a8f7276ceaa17b9233146.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 545, + 264 + ], + "type": "text", + "content": "Figure 6: Visual quality comparison for fused HSIs of all competing methods on CAVE, where first and second rows show fourth band and corresponding heatmaps (mean squared error), respectively." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 50, + 274, + 110, + 334 + ], + "blocks": [ + { + "bbox": [ + 50, + 274, + 110, + 334 + ], + "lines": [ + { + "bbox": [ + 50, + 274, + 110, + 334 + ], + "spans": [ + { + "bbox": [ + 50, + 274, + 110, + 334 + ], + "type": "image", + "image_path": "6f05a70c336542e6a5c8a56a77b432068d2c241b78cf8c3d189533bd05e494bf.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 112, + 275, + 171, + 334 + ], + "blocks": [ + { + "bbox": [ + 112, + 275, + 171, + 334 + ], + "lines": [ + { + "bbox": [ + 112, + 275, + 171, + 334 + ], + "spans": [ + { + "bbox": [ + 112, + 275, + 171, + 334 + ], + "type": "image", + "image_path": "1c3affc17c59a43ae29d90f24e1fe8b7e1bb42affb45f282fe47bd721b7c7bed.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 175, + 275, + 233, + 334 + ], + "blocks": [ + { + "bbox": [ + 175, + 275, + 233, + 334 + ], + "lines": [ + { + "bbox": [ + 175, + 275, + 233, + 334 + ], + "spans": [ + { + "bbox": [ + 175, + 275, + 233, + 334 + ], + "type": "image", + "image_path": "85cec20e87badfddcf7056487c5696bcf68e0a2ed623a850ee0141ce3ad1de2c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 236, + 275, + 295, + 334 + ], + "blocks": [ + { + "bbox": [ + 236, + 275, + 295, + 334 + ], + "lines": [ + { + "bbox": [ + 236, + 275, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 236, + 275, + 295, + 334 + ], + "type": "image", + "image_path": "c358e851f5dce380e6c6d400642f8c9b775d2e5c0c46749ccb0cad36c5085df4.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 298, + 275, + 358, + 334 + ], + "blocks": [ + { + "bbox": [ + 298, + 275, + 358, + 334 + ], + "lines": [ + { + "bbox": [ + 298, + 275, + 358, + 334 + ], + "spans": [ + { + "bbox": [ + 298, + 275, + 358, + 334 + ], + "type": "image", + "image_path": "c5e0e90785b7e242e23f9aa0ae81536e743e21e3cd10976cd79f1d71c4ecc2fd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 360, + 275, + 419, + 334 + ], + "blocks": [ + { + "bbox": [ + 360, + 275, + 419, + 334 + ], + "lines": [ + { + "bbox": [ + 360, + 275, + 419, + 334 + ], + "spans": [ + { + "bbox": [ + 360, + 275, + 419, + 334 + ], + "type": "image", + "image_path": "06a99816e30cc328bcd3262f6dfdba2093ea142821f323fce9b4c9afa85c8222.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 422, + 275, + 481, + 334 + ], + "blocks": [ + { + "bbox": [ + 422, + 275, + 481, + 334 + ], + "lines": [ + { + "bbox": [ + 422, + 275, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 422, + 275, + 481, + 334 + ], + "type": "image", + "image_path": "825a15e0b18e625c3152acf7180b5dc549c551cb056d22289638ca572afc24bc.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 484, + 275, + 543, + 334 + ], + "blocks": [ + { + "bbox": [ + 484, + 275, + 543, + 334 + ], + "lines": [ + { + "bbox": [ + 484, + 275, + 543, + 334 + ], + "spans": [ + { + "bbox": [ + 484, + 275, + 543, + 334 + ], + "type": "image", + "image_path": "fb0a566c8b2ce8acf2ce754e33febc80d1aae4b0ad151f101f02fb3622305112.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 51, + 343, + 110, + 402 + ], + "blocks": [ + { + "bbox": [ + 51, + 343, + 110, + 402 + ], + "lines": [ + { + "bbox": [ + 51, + 343, + 110, + 402 + ], + "spans": [ + { + "bbox": [ + 51, + 343, + 110, + 402 + ], + "type": "image", + "image_path": "26058ff6e433fc931062169f70a0ccee449e274b5498acf298a508503cfd81aa.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 404, + 101, + 414 + ], + "lines": [ + { + "bbox": [ + 59, + 404, + 101, + 414 + ], + "spans": [ + { + "bbox": [ + 59, + 404, + 101, + 414 + ], + "type": "text", + "content": "(a)UTV-TD" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 112, + 343, + 171, + 402 + ], + "blocks": [ + { + "bbox": [ + 112, + 343, + 171, + 402 + ], + "lines": [ + { + "bbox": [ + 112, + 343, + 171, + 402 + ], + "spans": [ + { + "bbox": [ + 112, + 343, + 171, + 402 + ], + "type": "image", + "image_path": "a56509deef7e4e3bc20c38d3b77424cb97b145379098ff744410ec779afe0caf.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 404, + 156, + 414 + ], + "lines": [ + { + "bbox": [ + 128, + 404, + 156, + 414 + ], + "spans": [ + { + "bbox": [ + 128, + 404, + 156, + 414 + ], + "type": "text", + "content": "(b) UAL" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 175, + 343, + 233, + 402 + ], + "blocks": [ + { + "bbox": [ + 175, + 343, + 233, + 402 + ], + "lines": [ + { + "bbox": [ + 175, + 343, + 233, + 402 + ], + "spans": [ + { + "bbox": [ + 175, + 343, + 233, + 402 + ], + "type": "image", + "image_path": "366c012d455662c4610b0e1b9cf53f177cf025fe10d57b25267520c065163a89.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 404, + 227, + 414 + ], + "lines": [ + { + "bbox": [ + 181, + 404, + 227, + 414 + ], + "spans": [ + { + "bbox": [ + 181, + 404, + 227, + 414 + ], + "type": "text", + "content": "(c) BRResNet" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 236, + 343, + 295, + 402 + ], + "blocks": [ + { + "bbox": [ + 236, + 343, + 295, + 402 + ], + "lines": [ + { + "bbox": [ + 236, + 343, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 236, + 343, + 295, + 402 + ], + "type": "image", + "image_path": "8cb71bb58a8c6441629007401dcd2a2c3e88af5a0a6de0e77fbb917d55ef2d3f.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 404, + 290, + 414 + ], + "lines": [ + { + "bbox": [ + 241, + 404, + 290, + 414 + ], + "spans": [ + { + "bbox": [ + 241, + 404, + 290, + 414 + ], + "type": "text", + "content": "(d) CMHF-Net" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 298, + 343, + 358, + 402 + ], + "blocks": [ + { + "bbox": [ + 298, + 343, + 358, + 402 + ], + "lines": [ + { + "bbox": [ + 298, + 343, + 358, + 402 + ], + "spans": [ + { + "bbox": [ + 298, + 343, + 358, + 402 + ], + "type": "image", + "image_path": "3db35b2163920ecfce7896ae0f3bb81dfb8842cba1bff4516cd54fbd969c9503.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 360, + 343, + 419, + 402 + ], + "blocks": [ + { + "bbox": [ + 360, + 343, + 419, + 402 + ], + "lines": [ + { + "bbox": [ + 360, + 343, + 419, + 402 + ], + "spans": [ + { + "bbox": [ + 360, + 343, + 419, + 402 + ], + "type": "image", + "image_path": "d03d5cd96bfef70c112b81cb88c34bf067c12773329836f86736d8b223c1a1b9.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 404, + 413, + 414 + ], + "lines": [ + { + "bbox": [ + 367, + 404, + 413, + 414 + ], + "spans": [ + { + "bbox": [ + 367, + 404, + 413, + 414 + ], + "type": "text", + "content": "(f) UAL-DMI" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 422, + 343, + 481, + 402 + ], + "blocks": [ + { + "bbox": [ + 422, + 343, + 481, + 402 + ], + "lines": [ + { + "bbox": [ + 422, + 343, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 422, + 343, + 481, + 402 + ], + "type": "image", + "image_path": "b6ec2f8b52aa5fb40683f058ed144bbce24113e3206a5f795c4fe5ab49fa246c.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 404, + 474, + 414 + ], + "lines": [ + { + "bbox": [ + 429, + 404, + 474, + 414 + ], + "spans": [ + { + "bbox": [ + 429, + 404, + 474, + 414 + ], + "type": "text", + "content": "(g) HSR-Diff" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 484, + 343, + 543, + 402 + ], + "blocks": [ + { + "bbox": [ + 484, + 343, + 543, + 402 + ], + "lines": [ + { + "bbox": [ + 484, + 343, + 543, + 402 + ], + "spans": [ + { + "bbox": [ + 484, + 343, + 543, + 402 + ], + "type": "image", + "image_path": "b464dd4a7a81058c9b4a0d6b4cce870ee7ed08b6b4ccfbd6152b19e1a703e752.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 404, + 541, + 414 + ], + "lines": [ + { + "bbox": [ + 487, + 404, + 541, + 414 + ], + "spans": [ + { + "bbox": [ + 487, + 404, + 541, + 414 + ], + "type": "text", + "content": "(h) Ground truth" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 160, + 417, + 434, + 436 + ], + "blocks": [ + { + "bbox": [ + 305, + 404, + 351, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 404, + 351, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 351, + 414 + ], + "type": "text", + "content": "(e) Fusformer" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 417, + 434, + 436 + ], + "lines": [ + { + "bbox": [ + 160, + 417, + 434, + 436 + ], + "spans": [ + { + "bbox": [ + 160, + 417, + 434, + 436 + ], + "type": "image", + "image_path": "ea2453226b911803317da1556b439e4d300ee7c4c0b85e537953fcaa8dc87bd4.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 50, + 479, + 110, + 539 + ], + "blocks": [ + { + "bbox": [ + 46, + 445, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 545, + 469 + ], + "type": "text", + "content": "Figure 7: Visual quality comparison for fused HSIs of all competing methods on PaviaU, where first and second rows show 81st band and corresponding heatmaps (mean squared error), respectively." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 479, + 110, + 539 + ], + "lines": [ + { + "bbox": [ + 50, + 479, + 110, + 539 + ], + "spans": [ + { + "bbox": [ + 50, + 479, + 110, + 539 + ], + "type": "image", + "image_path": "d0893649ed3b9b4284f388e1e8c614dd98651dc04d735063c4521e08d66d97c4.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 112, + 479, + 171, + 539 + ], + "blocks": [ + { + "bbox": [ + 112, + 479, + 171, + 539 + ], + "lines": [ + { + "bbox": [ + 112, + 479, + 171, + 539 + ], + "spans": [ + { + "bbox": [ + 112, + 479, + 171, + 539 + ], + "type": "image", + "image_path": "e1f80953259ba7aac49f4b03d2ad2215aa949516f66ceeeaf4876535b7fd83ed.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 175, + 479, + 233, + 539 + ], + "blocks": [ + { + "bbox": [ + 175, + 479, + 233, + 539 + ], + "lines": [ + { + "bbox": [ + 175, + 479, + 233, + 539 + ], + "spans": [ + { + "bbox": [ + 175, + 479, + 233, + 539 + ], + "type": "image", + "image_path": "b691686140c59179ab86adbe497853c99b510b2e0c3a94a229c3108161becf74.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 236, + 479, + 295, + 539 + ], + "blocks": [ + { + "bbox": [ + 236, + 479, + 295, + 539 + ], + "lines": [ + { + "bbox": [ + 236, + 479, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 236, + 479, + 295, + 539 + ], + "type": "image", + "image_path": "bf0051760551a1149d076a9f64dd1ef0c897b3a421733b3cf7ecf2bfaa43b4d7.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 298, + 479, + 358, + 539 + ], + "blocks": [ + { + "bbox": [ + 298, + 479, + 358, + 539 + ], + "lines": [ + { + "bbox": [ + 298, + 479, + 358, + 539 + ], + "spans": [ + { + "bbox": [ + 298, + 479, + 358, + 539 + ], + "type": "image", + "image_path": "596261dd5b1ff33bd3f61d16fc0068efafd3c07b6bec686f6016c7c184f818fa.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 360, + 479, + 419, + 539 + ], + "blocks": [ + { + "bbox": [ + 360, + 479, + 419, + 539 + ], + "lines": [ + { + "bbox": [ + 360, + 479, + 419, + 539 + ], + "spans": [ + { + "bbox": [ + 360, + 479, + 419, + 539 + ], + "type": "image", + "image_path": "ff689d97198edd375e059155ece0a53219aff4463f11ab039162f9087a683206.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 422, + 479, + 481, + 539 + ], + "blocks": [ + { + "bbox": [ + 422, + 479, + 481, + 539 + ], + "lines": [ + { + "bbox": [ + 422, + 479, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 422, + 479, + 481, + 539 + ], + "type": "image", + "image_path": "6902bdd95fb516980f87f20722069d90e2bf0eed85656eb2a9c027a1f2a3f528.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 484, + 479, + 543, + 539 + ], + "blocks": [ + { + "bbox": [ + 484, + 479, + 543, + 539 + ], + "lines": [ + { + "bbox": [ + 484, + 479, + 543, + 539 + ], + "spans": [ + { + "bbox": [ + 484, + 479, + 543, + 539 + ], + "type": "image", + "image_path": "cd07ddb4ac56f2ec519beee3336ec1d3b2481ea050228cd9c47825553833a902.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 51, + 548, + 110, + 606 + ], + "blocks": [ + { + "bbox": [ + 51, + 548, + 110, + 606 + ], + "lines": [ + { + "bbox": [ + 51, + 548, + 110, + 606 + ], + "spans": [ + { + "bbox": [ + 51, + 548, + 110, + 606 + ], + "type": "image", + "image_path": "eaef687c57d2de4199b95308df6cb6f5ce3de64d8a1df75b8fb2fae63b822f7d.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 609, + 101, + 618 + ], + "lines": [ + { + "bbox": [ + 59, + 609, + 101, + 618 + ], + "spans": [ + { + "bbox": [ + 59, + 609, + 101, + 618 + ], + "type": "text", + "content": "(a)UTV-TD" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 112, + 548, + 171, + 605 + ], + "blocks": [ + { + "bbox": [ + 112, + 548, + 171, + 605 + ], + "lines": [ + { + "bbox": [ + 112, + 548, + 171, + 605 + ], + "spans": [ + { + "bbox": [ + 112, + 548, + 171, + 605 + ], + "type": "image", + "image_path": "cfcb16a0fe695da51f0ace142c54f2153e27d75791f96221cac4e66f8649989e.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 609, + 156, + 618 + ], + "lines": [ + { + "bbox": [ + 128, + 609, + 156, + 618 + ], + "spans": [ + { + "bbox": [ + 128, + 609, + 156, + 618 + ], + "type": "text", + "content": "(b) UAL" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 175, + 548, + 233, + 606 + ], + "blocks": [ + { + "bbox": [ + 175, + 548, + 233, + 606 + ], + "lines": [ + { + "bbox": [ + 175, + 548, + 233, + 606 + ], + "spans": [ + { + "bbox": [ + 175, + 548, + 233, + 606 + ], + "type": "image", + "image_path": "a8c302a9c5e986b5d5f199568399b95dda997d43549e1638f4ed122d62cd8a9f.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 609, + 227, + 618 + ], + "lines": [ + { + "bbox": [ + 181, + 609, + 227, + 618 + ], + "spans": [ + { + "bbox": [ + 181, + 609, + 227, + 618 + ], + "type": "text", + "content": "(c) BRResNet" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 236, + 548, + 295, + 606 + ], + "blocks": [ + { + "bbox": [ + 236, + 548, + 295, + 606 + ], + "lines": [ + { + "bbox": [ + 236, + 548, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 236, + 548, + 295, + 606 + ], + "type": "image", + "image_path": "0fa7947ad384a6ce64801c60ed35ed87d07a17027c11a06d50539e33ca9f9aba.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 609, + 290, + 618 + ], + "lines": [ + { + "bbox": [ + 241, + 609, + 290, + 618 + ], + "spans": [ + { + "bbox": [ + 241, + 609, + 290, + 618 + ], + "type": "text", + "content": "(d) CMHF-Net" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 298, + 548, + 358, + 606 + ], + "blocks": [ + { + "bbox": [ + 298, + 548, + 358, + 606 + ], + "lines": [ + { + "bbox": [ + 298, + 548, + 358, + 606 + ], + "spans": [ + { + "bbox": [ + 298, + 548, + 358, + 606 + ], + "type": "image", + "image_path": "0e8cfd22c89df9d0b2c1ceb685f0ffbb09811e30e0e4b6fdaa4049d8aff97f39.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 360, + 548, + 419, + 606 + ], + "blocks": [ + { + "bbox": [ + 360, + 548, + 419, + 606 + ], + "lines": [ + { + "bbox": [ + 360, + 548, + 419, + 606 + ], + "spans": [ + { + "bbox": [ + 360, + 548, + 419, + 606 + ], + "type": "image", + "image_path": "9822930510caa0752893ef4b337cf40ec9c55b584f63f39408356a70782ec487.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 422, + 548, + 481, + 606 + ], + "blocks": [ + { + "bbox": [ + 422, + 548, + 481, + 606 + ], + "lines": [ + { + "bbox": [ + 422, + 548, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 422, + 548, + 481, + 606 + ], + "type": "image", + "image_path": "7d1b36b224b1735acfdbc56e6170292e73f3eaf7a1f62624bae76849243284d5.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 484, + 548, + 543, + 606 + ], + "blocks": [ + { + "bbox": [ + 484, + 548, + 543, + 606 + ], + "lines": [ + { + "bbox": [ + 484, + 548, + 543, + 606 + ], + "spans": [ + { + "bbox": [ + 484, + 548, + 543, + 606 + ], + "type": "image", + "image_path": "3502d620be0556cb67bbe1b5167be649031f571fdc4e20959d12f27f77a8a77f.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 609, + 541, + 618 + ], + "lines": [ + { + "bbox": [ + 487, + 609, + 541, + 618 + ], + "spans": [ + { + "bbox": [ + 487, + 609, + 541, + 618 + ], + "type": "text", + "content": "(h) Ground truth" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 160, + 620, + 434, + 640 + ], + "blocks": [ + { + "bbox": [ + 305, + 609, + 351, + 618 + ], + "lines": [ + { + "bbox": [ + 305, + 609, + 351, + 618 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 351, + 618 + ], + "type": "text", + "content": "(e) Fusformer" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 609, + 413, + 618 + ], + "lines": [ + { + "bbox": [ + 367, + 609, + 413, + 618 + ], + "spans": [ + { + "bbox": [ + 367, + 609, + 413, + 618 + ], + "type": "text", + "content": "(f) UAL-DMI" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 609, + 474, + 618 + ], + "lines": [ + { + "bbox": [ + 429, + 609, + 474, + 618 + ], + "spans": [ + { + "bbox": [ + 429, + 609, + 474, + 618 + ], + "type": "text", + "content": "(g) HSR-Diff" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 620, + 434, + 640 + ], + "lines": [ + { + "bbox": [ + 160, + 620, + 434, + 640 + ], + "spans": [ + { + "bbox": [ + 160, + 620, + 434, + 640 + ], + "type": "image", + "image_path": "3c257a0e9bbfb832178a24e0e8d365e6fd0bcc1cbca1cbfe35e4a4a795069305.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 648, + 545, + 673 + ], + "lines": [ + { + "bbox": [ + 46, + 648, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 545, + 673 + ], + "type": "text", + "content": "Figure 8: Visual quality comparison for fused HSIs of all competing methods on Chikusei, where first and second rows show 67th band and corresponding heatmaps (mean squared error), respectively." + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_caption" + } + ], + "index": 76 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7089" + } + ] + } + ], + "index": 78 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 287, + 343 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 287, + 343 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 287, + 343 + ], + "type": "table", + "html": "
DatasetMethodsPSNR↑SSIM↑SAM↓ERGAS↓
CAVEUTV-TD [34]38.660.97997.980.329
UAL [43]40.550.99334.330.271
BRResNet [16]41.360.99294.700.250
CMHF-Net [31]42.540.99394.690.216
Fusformer [15]41.520.99344.710.243
UAL-DMI [28]42.740.99503.790.213
HSR-Diff44.330.99513.710.179
PaviaUUTV-TD [34]44.460.99521.801.236
UAL [43]45.420.99641.541.148
BRResNet [16]45.530.99651.531.111
CMHF-Net [31]45.770.99651.501.096
Fusformer [15]45.660.99651.521.109
UAL-DMI [28]45.680.99661.491.113
HSR-Diff46.470.99771.451.053
ChikuseiUTV-TD [34]48.380.99890.991.303
UAL [43]56.180.99980.490.421
BRResNet [16]56.790.99980.460.366
CMHF-Net [31]55.990.99980.500.483
Fusformer [15]55.920.99980.520.492
UAL-DMI [28]56.570.99980.470.387
HSR-Diff57.340.99990.430.324
", + "image_path": "5d8d8402d20e8bc74c3341e0ce2c00e7fe1679421d4153113868686125f75a34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": "Qualitative Comparison. To assess the performance of HSR-Diff qualitatively, we visualize example bands of HSIs in Figures 6, 7, and 8. It can be seen from these visual results that all compared methods produce satisfactory outcomes. In particular, HSR-Diff generates the best result with minor errors since the corresponding MSE (mean squared error) images are much clearer than the others." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "content": "Quantitative Comparison. To further verify the superior performance of the proposed HSR-Diff, quantitative results are presented in Table 1. Note that the performance indices on the CAVE and Chikusei datasets are averaged over all testing samples (12 samples for CAVE and eight samples for Chikusei), respectively. It can be inferred from the results that the proposed HSR-Diff surpasses all competitors with a clear margin on all evaluation metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Effect of conditional diffusion models. Much of the early work on HSI-SR was based on the use of the regression model. To compare the effects of the diffusion and the regression model, we train the regression model containing the CDFformer. Note that the loss function, optimizer, and hyper-parameters are all the same as the conditional diffu" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 386, + 145 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 386, + 145 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 386, + 145 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 386, + 145 + ], + "type": "image", + "image_path": "282b272ad671d6b404a3ed22635d1576a4bf78b78ed4702b1d69ab6d55912c09.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 463, + 145 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 463, + 145 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 463, + 145 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 463, + 145 + ], + "type": "image", + "image_path": "7f27e977efd2497c09bc9d01b49e33e674787290d6375d8be6eceb808e3a0453.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 466, + 70, + 542, + 145 + ], + "blocks": [ + { + "bbox": [ + 466, + 70, + 542, + 145 + ], + "lines": [ + { + "bbox": [ + 466, + 70, + 542, + 145 + ], + "spans": [ + { + "bbox": [ + 466, + 70, + 542, + 145 + ], + "type": "image", + "image_path": "1510adeba85387e2a447338c741704477933bd78e2a337868bcdbd57d0f3ff6b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 153, + 386, + 229 + ], + "blocks": [ + { + "bbox": [ + 310, + 153, + 386, + 229 + ], + "lines": [ + { + "bbox": [ + 310, + 153, + 386, + 229 + ], + "spans": [ + { + "bbox": [ + 310, + 153, + 386, + 229 + ], + "type": "image", + "image_path": "20c5dfeab9f90b804383c1554725bda46e8f71ac8393518617942cfcb6fe1d5b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 232, + 386, + 242 + ], + "lines": [ + { + "bbox": [ + 310, + 232, + 386, + 242 + ], + "spans": [ + { + "bbox": [ + 310, + 232, + 386, + 242 + ], + "type": "text", + "content": "(a) Regression 49.13dB" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 389, + 153, + 463, + 229 + ], + "blocks": [ + { + "bbox": [ + 389, + 153, + 463, + 229 + ], + "lines": [ + { + "bbox": [ + 389, + 153, + 463, + 229 + ], + "spans": [ + { + "bbox": [ + 389, + 153, + 463, + 229 + ], + "type": "image", + "image_path": "a61dda589a4d10190c537c36b990fc015b019689611ef15fff5d05fc466ea5e7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 232, + 462, + 242 + ], + "lines": [ + { + "bbox": [ + 389, + 232, + 462, + 242 + ], + "spans": [ + { + "bbox": [ + 389, + 232, + 462, + 242 + ], + "type": "text", + "content": "(b) HSR-Diff 52.06dB" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 467, + 154, + 542, + 229 + ], + "blocks": [ + { + "bbox": [ + 467, + 154, + 542, + 229 + ], + "lines": [ + { + "bbox": [ + 467, + 154, + 542, + 229 + ], + "spans": [ + { + "bbox": [ + 467, + 154, + 542, + 229 + ], + "type": "image", + "image_path": "5345dfe7c76dd1556d43bdea406132e4e5a32129ba1d03d19c3260b88b8540a0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 245, + 545, + 261 + ], + "blocks": [ + { + "bbox": [ + 477, + 232, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 477, + 232, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 477, + 232, + 531, + 242 + ], + "type": "text", + "content": "(c) Ground truth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 245, + 545, + 261 + ], + "lines": [ + { + "bbox": [ + 318, + 245, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 318, + 245, + 545, + 261 + ], + "type": "image", + "image_path": "a2f269d52ba9006f83bc3eae5c1e1e1af5ba9a1047ba202316b82386499a68dd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 268, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 306, + 268, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 545, + 305 + ], + "type": "text", + "content": "Figure 9: Fusion results for HSR-Diff and Regression on the real and fake peppers image of CAVE. The PSNR values (dB) are given in the subtitles." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 307, + 317, + 545, + 411 + ], + "blocks": [ + { + "bbox": [ + 46, + 350, + 288, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 288, + 399 + ], + "type": "text", + "content": "Table 1: Averaged PSNR, SSIM, SAM, and ERGAS of compared methods on CAVE, PaviaU, and Chikusei datasets. The " + }, + { + "bbox": [ + 46, + 350, + 288, + 399 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 350, + 288, + 399 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 350, + 288, + 399 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 350, + 288, + 399 + ], + "type": "text", + "content": " indicates higher or lower values corresponding to better results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 317, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 307, + 317, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 545, + 411 + ], + "type": "table", + "html": "
DatasetModelPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVERegression43.260.99424.030.193
CDFformer44.330.99513.710.179
PaviaURegression46.030.99681.521.178
CDFformer46.470.99771.451.053
ChikuseiRegression56.710.99990.470.372
CDFformer57.340.99990.430.324
", + "image_path": "46d8f9205ac589f5c0537c4d910eb49dd8ed1cdd8caaedb7b95fc8f8a51151ce.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 418, + 545, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 455 + ], + "type": "text", + "content": "Table 2: Ablation study on conditional diffusion models. The " + }, + { + "bbox": [ + 305, + 418, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 305, + 418, + 545, + 455 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 418, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 305, + 418, + 545, + 455 + ], + "type": "text", + "content": " indicates higher or lower values corresponding to better results." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 478, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 598 + ], + "type": "text", + "content": "sion models. Figure 9 presents the fused results and corresponding error maps of utilising HSR-Diff and the regression model. As can be seen from the error maps, the HSIs produced by HSR-Diff have less distortion than those by the regression model. This is because HSR-Diff works with a series of iterative refinement steps, facilitating the capture of richer information on data distributions of HR-HSIs. Furthermore, Table 2 showcases the quantitative results and effectively highlights the significant contribution of iterative refinement." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Effect of CDFormer. Recall that CDFormer is conditioned on the hierarchical representations of HR-MSI and LR-HSI via a two-stream architecture. However, alternative outstanding diffusion models [14, 23] that are also excellent for conditional image generation are equipped with CNN-based U-Nets, where degenerated images are concatenated with noisy high-resolution output images. To show the effectiveness of hierarchical representations, we remove the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7090" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 79, + 119, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 79, + 119, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 119, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 119, + 148 + ], + "type": "image", + "image_path": "0d990b38633caae1cbcf582cb90a4f8439c06a8c69001841a19b491494f410ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 150, + 97, + 159 + ], + "lines": [ + { + "bbox": [ + 72, + 150, + 97, + 159 + ], + "spans": [ + { + "bbox": [ + 72, + 150, + 97, + 159 + ], + "type": "text", + "content": "(a) HSI" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 79, + 190, + 148 + ], + "blocks": [ + { + "bbox": [ + 121, + 79, + 190, + 148 + ], + "lines": [ + { + "bbox": [ + 121, + 79, + 190, + 148 + ], + "spans": [ + { + "bbox": [ + 121, + 79, + 190, + 148 + ], + "type": "image", + "image_path": "1e63519dc0dcfcefe9e053ccda826e92fdb28596669bcd6890c6f136a2c5cd56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 150, + 176, + 159 + ], + "lines": [ + { + "bbox": [ + 134, + 150, + 176, + 159 + ], + "spans": [ + { + "bbox": [ + 134, + 150, + 176, + 159 + ], + "type": "text", + "content": "(b)UTV-TD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 79, + 261, + 148 + ], + "blocks": [ + { + "bbox": [ + 192, + 79, + 261, + 148 + ], + "lines": [ + { + "bbox": [ + 192, + 79, + 261, + 148 + ], + "spans": [ + { + "bbox": [ + 192, + 79, + 261, + 148 + ], + "type": "image", + "image_path": "f61bb7027e411a835dfedcdfcc38e82a120285039a79cdc5dc30b0e183dc8b12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 150, + 241, + 159 + ], + "lines": [ + { + "bbox": [ + 212, + 150, + 241, + 159 + ], + "spans": [ + { + "bbox": [ + 212, + 150, + 241, + 159 + ], + "type": "text", + "content": "(c) UAL" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 263, + 79, + 332, + 148 + ], + "blocks": [ + { + "bbox": [ + 263, + 79, + 332, + 148 + ], + "lines": [ + { + "bbox": [ + 263, + 79, + 332, + 148 + ], + "spans": [ + { + "bbox": [ + 263, + 79, + 332, + 148 + ], + "type": "image", + "image_path": "c4475d6dc385fe81cbc5c162b73b76066729cc52030d0135999066f24bcd5312.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 150, + 320, + 159 + ], + "lines": [ + { + "bbox": [ + 273, + 150, + 320, + 159 + ], + "spans": [ + { + "bbox": [ + 273, + 150, + 320, + 159 + ], + "type": "text", + "content": "(d) BRResNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 170, + 437, + 182 + ], + "lines": [ + { + "bbox": [ + 153, + 170, + 437, + 182 + ], + "spans": [ + { + "bbox": [ + 153, + 170, + 437, + 182 + ], + "type": "text", + "content": "Figure 10: Visual fusion results of all competing methods for HypSen." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 79, + 402, + 148 + ], + "blocks": [ + { + "bbox": [ + 334, + 79, + 402, + 148 + ], + "lines": [ + { + "bbox": [ + 334, + 79, + 402, + 148 + ], + "spans": [ + { + "bbox": [ + 334, + 79, + 402, + 148 + ], + "type": "image", + "image_path": "3c78fd2b697fbf692f094d43862d354311e4dbb6ca987e0740d6591e974a4c1a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 150, + 392, + 159 + ], + "lines": [ + { + "bbox": [ + 343, + 150, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 343, + 150, + 392, + 159 + ], + "type": "text", + "content": "(e) CMHF-Net" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 405, + 79, + 473, + 148 + ], + "blocks": [ + { + "bbox": [ + 405, + 79, + 473, + 148 + ], + "lines": [ + { + "bbox": [ + 405, + 79, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 405, + 79, + 473, + 148 + ], + "type": "image", + "image_path": "cfcc1b3e183daf06d3dd4fb15e64af6136d21ca99b601c1a960440c992268d7c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 150, + 462, + 159 + ], + "lines": [ + { + "bbox": [ + 416, + 150, + 462, + 159 + ], + "spans": [ + { + "bbox": [ + 416, + 150, + 462, + 159 + ], + "type": "text", + "content": "(f) UAL-DMI" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 476, + 79, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 476, + 79, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 476, + 79, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 476, + 79, + 544, + 148 + ], + "type": "image", + "image_path": "24eb1dcfd33ad077d2b7cb7cbe93b8643a16cb381f4337c1ac7e63b26e6c717b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 150, + 532, + 159 + ], + "lines": [ + { + "bbox": [ + 487, + 150, + 532, + 159 + ], + "spans": [ + { + "bbox": [ + 487, + 150, + 532, + 159 + ], + "type": "text", + "content": "(g) HSR-Diff" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 48, + 201, + 286, + 330 + ], + "blocks": [ + { + "bbox": [ + 48, + 201, + 286, + 330 + ], + "lines": [ + { + "bbox": [ + 48, + 201, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 286, + 330 + ], + "type": "table", + "html": "
DatasetNetworkPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVECD-CNN38.840.97977.320.318
C-w/o-SR43.740.99423.940.188
CDFformer44.330.99513.710.179
PaviaUCD-CNN42.750.99621.751.362
C-w/o-SR46.080.99761.471.080
CDFformer46.470.99771.451.053
ChikuseiCD-CNN47.630.99801.201.794
C-w/o-SR56.680.99990.460.425
CDFformer57.340.99990.430.324
", + "image_path": "38259561aeb098f1914201c8cae30dbd1450fc2c9ac89cf901f552b44010393a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 48, + 376, + 288, + 469 + ], + "blocks": [ + { + "bbox": [ + 47, + 338, + 287, + 363 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 363 + ], + "type": "text", + "content": "Table 3: Ablation study on CDFormer. The " + }, + { + "bbox": [ + 47, + 338, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 47, + 338, + 287, + 363 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 338, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 338, + 287, + 363 + ], + "type": "text", + "content": " indicates higher or lower values corresponding to better results." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 376, + 288, + 469 + ], + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 469 + ], + "type": "table", + "html": "
DatasetMethodsPSNR ↑SSIM ↑SAM ↓ERGAS ↓
CAVEFixed43.170.99274.990.203
Progressive44.330.99513.710.179
PaviaUFixed45.060.99701.631.173
Progressive46.470.99771.451.053
ChikuseiFixed55.920.99990.500.453
Progressive57.340.99990.430.324
", + "image_path": "51cb88b3a9e887a212404cee8180b0d9e48181092e6f5263303b672aeb20bd1d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 478, + 287, + 513 + ], + "lines": [ + { + "bbox": [ + 47, + 478, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 287, + 513 + ], + "type": "text", + "content": "Table 4: Ablation study on progressive learning. The " + }, + { + "bbox": [ + 47, + 478, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 47, + 478, + 287, + 513 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 478, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 478, + 287, + 513 + ], + "type": "text", + "content": " indicates higher or lower values corresponding to better results." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 536, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 644 + ], + "type": "text", + "content": "SR stream of CDFormer and name the resulting network \"C-w/o-SR\". In addition, we compare CDFormer with a CNN version of CDFormer that replaces all " + }, + { + "bbox": [ + 46, + 536, + 287, + 644 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 46, + 536, + 287, + 644 + ], + "type": "text", + "content": "TL with convolutional layers and show its results as \"CD-CNN\" in Table 3. The quantitative results show the use of CDFormer performs better than CD-CNN, demonstrating the effectiveness of global statistics. Indeed, with the two-stream architecture, CDFormer offers the best results thanks to the use of hierarchical features." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Effect of progressive learning. Progressive learning helps CDFormer to capture long-range dependencies of spatio-spectral information in HR-HSIs. To illustrate the effect of progressive learning, we train the CDFormer" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 202, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 276 + ], + "type": "text", + "content": "with fixed patches (128² for CAVE and Chikusei; 64² for PaviaU) with the results shown under the heading of \"Fixed\" in Table 4. As can be seen, progressive learning (from 128² to 512² for CAVE and Chikusei; from 64² to 128² for PaviaU) provides better results than training with fixed patches." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 282, + 517, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 517, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 517, + 295 + ], + "type": "text", + "content": "4.6. Generalization Analysis on Real Dataset" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 300, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 433 + ], + "type": "text", + "content": "To examine the generalization ability of the implementations following the proposed approach, we test the performance of competitors on the real-world HypSen dataset [37]. Due to the lack of an ideal HR-HSI to train deep neural networks, we utilize the networks trained on the PaviaU dataset to merge observed LR-HSI and the corresponding HR-MSI. In addition, interpolation is applied to addressing the problem of an inconsistent number of bands between datasets. The fusion results of all compared methods are visualized in Figure 10, from which it can be seen that our method generates rich details, attaining satisfactory quality." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 463, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 703 + ], + "type": "text", + "content": "In this paper, we have presented the novel HSR-Diff approach that initializes an HR-HSI with pure Gaussian noise and then, iteratively refines it subject to the condition of the LR-HSIs and HR-MSIs of interest. At each step, the noise is removed with CDFormer which exploits the hierarchical representations of HR-MSIs and LR-HSIs rather than the original images. In addition, we employ a progressive learning strategy to maximize the use of the global information of full-resolution images, where CDFormer is trained on small patches in the early epochs with high efficiency while on the global images in the later epochs to obtain the global statistics. Systematic experimental investigations have been conducted, on four public datasets to validate the superior performance of the proposed approach, in comparison with state-of-the-art methods. However, diffusion models have low inference efficiency due to their sequential nature, computational complexity, time-consuming sampling process, and large model size. For future work, we will try to resolve the challenging issue of the relatively low image-generation efficiency of HSR-Diff." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7091" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Naveed Akhtar, Faisal Shafait, and Ajmal Mian. Bayesian sparse representation for hyperspectral image super resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3631-3640, 2015. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Ricardo Augusto Borsoi, Tales Imbiriba, and José Carlos Moreira Bermudez. Super-resolution for hyperspectral and multispectral image fusion accounting for seasonal spectral variability. IEEE Transactions on Image Processing, 29:116-127, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 247 + ], + "type": "text", + "content": "[3] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "text", + "content": "[4] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. Wavegrad: Estimating gradients for waveform generation. In Proceedings of the International Conference on Learning Represent, 2021. 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "text", + "content": "[5] Zhao Chen, Hanye Pu, Bin Wang, and Geng-Ming Jiang. Fusion of hyperspectral and multispectral images: A novel framework based on generalization of pan-sharpening methods. IEEE Geoscience and Remote Sensing Letters, 11(8):1418-1422, 2014. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 407 + ], + "type": "text", + "content": "[6] Phuong D Dao, Kiran Mantripragada, Yuhong He, and Faisal Z Qureshi. Improving hyperspectral image segmentation by applying inverse noise weighting and outlier removal for optimal scale selection. ISPRS Journal of Photogrammetry and Remote Sensing, 171:348-366, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 463 + ], + "type": "text", + "content": "[7] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. Psrt: Pyramid shuffle-and-reshuffle transformer for multispectral and hyperspectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 508 + ], + "type": "text", + "content": "[8] Renwei Dian, Shutao Li, Anjing Guo, and Leyuan Fang. Deep hyperspectral image sharpening. IEEE transactions on neural networks and learning systems, 29(11):5345-5355, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 554 + ], + "type": "text", + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "type": "text", + "content": "[10] Ying Fu, Tao Zhang, Yinqiang Zheng, Debing Zhang, and Hua Huang. Hyperspectral image super-resolution with optimized rgb guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11661-11670, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[15] Jin-Fan Hu, Ting-Zhu Huang, Liang-Jian Deng, Hong-Xia Dou, Danfeng Hong, and Gemine Vivone. Fusformer: A transformer-based fusion network for hyperspectral image super-resolution. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "text", + "content": "[16] Zi-Rong Jin, Liang-Jian Deng, Tian-Jing Zhang, and Xiao-Xu Jin. Bam: Bilateral activation mechanism for image fusion. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4315-4323, 2021. 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 329 + ], + "type": "text", + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 375 + ], + "type": "text", + "content": "[18] Shutao Li, Renwei Dian, Leyuan Fang, and José M Bioucas-Dias. Fusing hyperspectral and multispectral images via coupled sparse tensor factorization. IEEE Transactions on Image Processing, 27(8):4118-4130, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "text", + "content": "[19] Ying Li, Haokui Zhang, and Qiang Shen. Spectral-spatial classification of hyperspectral imagery with 3d convolutional neural network. Remote Sensing, 9(1):67, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 468, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 499 + ], + "type": "text", + "content": "[21] Gamba Paolo. Pavia centre and university. https://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Sciences, 2011. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[22] Ying Qu, Hairong Qi, and Chiman Kwan. Unsupervised sparse dirichlet-net for hyperspectral image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2511-2520, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 590 + ], + "type": "text", + "content": "[23] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "text", + "content": "[24] Yue Shi, Liangxiu Han, Lianghao Han, Sheng Chang, Tongle Hu, and Darren Dancey. A latent encoder coupled generative adversarial network (le-gan) for efficient hyperspectral image super-resolution. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[25] Dong Wang, Yunpeng Bai, Bendu Bai, Chanyue Wu, and Ying Li. Heterogeneous two-stream network with hierarchical feature prefusion for multispectral pan-sharpening. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1845-1849. IEEE, 2021. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7092" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[26] Dong Wang, Yunpeng Bai, Chanyue Wu, Ying Li, Changjing Shang, and Qiang Shen. Convolutional LSTM-based hierarchical feature fusion for multispectral pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "type": "text", + "content": "[27] Dong Wang, Pei Zhang, Yunpeng Bai, and Ying Li. Metapan: Unsupervised adaptation with meta-learning for multispectral pansharpening. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[28] Xiuheng Wang, Jie Chen, and Cedric Richard. Hyperspectral image super-resolution with deep priors and degradation model inversion. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2814-2818. IEEE, 2022. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[29] Qi Wei, José Bioucas-Dias, Nicolas Dobigeon, and Jean-Yves Tourneret. Hyperspectral and multispectral image fusion based on a sparse representation. IEEE Transactions on Geoscience and Remote Sensing, 53(7):3658-3668, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 325 + ], + "type": "text", + "content": "[30] Zhenyu Wu, Lin Wang, Wei Wang, Tengfei Shi, Chenglizhao Chen, Aimin Hao, and Shuo Li. Synthetic data supervised salient object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5557-5565, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 326, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 380 + ], + "type": "text", + "content": "[31] Qi Xie, Minghao Zhou, Qian Zhao, Zongben Xu, and Deyu Meng. Mhf-net: An interpretable deep network for multispectral and hyperspectral image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 381, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 287, + 415 + ], + "type": "text", + "content": "[32] Fengchao Xiong, Jun Zhou, and Yuntao Qian. Material based object tracking in hyperspectral videos. IEEE Transactions on Image Processing, 29:3719-3733, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[33] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[34] Ting Xu, Ting-Zhu Huang, Liang-Jian Deng, Xi-Le Zhao, and Jie Huang. Hyperspectral image superresolution using unidirectional total variation with tucker decomposition. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 13:4381-4398, 2020. 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 580 + ], + "type": "text", + "content": "[35] Xizhe Xue, Haokui Zhang, Bei Fang, Zongwen Bai, and Ying Li. Grafting transformer on automatically designed convolutional neural network for hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[36] Jing Yang, Chanyue Wu, Tengfei You, Dong Wang, Ying Li, Changjing Shang, and Qiang Shen. Hierarchical spatio-spectral fusion for hyperspectral image super resolution via sparse representation and pre-trained deep model. Knowledge-Based Systems, 260:110170, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "text", + "content": "[37] Jingxiang Yang, Yong-Qiang Zhao, and Jonathan Cheung-Wai Chan. Hyperspectral and multispectral image fusion via deep two-branches convolutional neural network. Remote Sensing, 10(5):800, 2018. 6, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[38] Fumihito Yasuma, Tomoo Mitsunaga, Daisuke Iso, and Shree K Nayar. Generalized assorted pixel camera: post-capture control of resolution, dynamic range, and spectrum." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 529 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE transactions on image processing, 19(9):2241-2253, 2010.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 129 + ], + "type": "text", + "content": "[39] Naoto Yokoya and Akira Iwasaki. Airborne hyperspectral data over chikusei. Space Appl. Lab., Univ. Tokyo, Tokyo, Japan, Tech. Rep. SAL-2016-05-27, 5, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[40] Tengfei You, Chanyue Wu, Yunpeng Bai, Dong Wang, Huibin Ge, and Ying Li. Hmf-former: Spatio-spectral transformer for hyperspectral and multispectral image fusion. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 251 + ], + "type": "text", + "content": "[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[42] Haokui Zhang, Chengrong Gong, Yunpeng Bai, Zongwen Bai, and Ying Li. 3-d-anas: 3-d asymmetric neural architecture search for fast hyperspectral image classification. IEEE Transactions on Geoscience and Remote Sensing, 60:1-19, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[43] Lei Zhang, Jiangtao Nie, Wei Wei, Yanning Zhang, Shengcai Liao, and Ling Shao. Unsupervised adaptation learning for hyperspectral imagery super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3073-3082, 2020. 2, 3, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 364, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 545, + 418 + ], + "type": "text", + "content": "[44] Meilin Zhang, Xiongli Sun, Qiqi Zhu, and Guizhou Zheng. A survey of hyperspectral image super-resolution technology. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, pages 4476-4479. IEEE, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 464 + ], + "type": "text", + "content": "[45] Xueting Zhang, Wei Huang, Qi Wang, and Xuelong Li. Ssrnet: Spatial-spectral reconstruction network for hyperspectral and multispectral image fusion. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5953-5965, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "text", + "content": "[46] Ke Zheng, Lianru Gao, Wenzhi Liao, Danfeng Hong, Bing Zhang, Xin Cui, and Jocelyn Chanussot. Coupled convolutional neural network with adaptive response function learning for unsupervised hyperspectral super resolution. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2487-2502, 2020. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7093" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_content_list.json b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..077bce87f985ed244be7db5195f665273c98b31f --- /dev/null +++ b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_content_list.json @@ -0,0 +1,1413 @@ +[ + { + "type": "text", + "text": "HTML: Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation", + "text_level": 1, + "bbox": [ + 133, + 128, + 836, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingfei Han $^{1,4}$ , Yali Wang $^{\\dagger,2,6}$ , Zhihui Li $^{3}$ , Lina Yao $^{4}$ , Xiaojun Chang $^{1,5}$ , Yu Qiao $^{6,2}$", + "bbox": [ + 156, + 202, + 818, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ReLER, AAII, UTS $^{2}$ Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, SIAT-SenseTime Joint Lab, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences $^{3}$ Shandong Artificial Intelligence, Qilu University of Technology $^{4}$ Data61, CSIRO $^{5}$ Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence $^{6}$ Shanghai AI Laboratory, Shanghai, China https://mingfei.info/HTML/", + "bbox": [ + 88, + 220, + 890, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 258, + 361, + 336, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring Video Object Segmentation (RVOS) is to segment the object instance from a given video, according to the textual description of this object. However, in the open world, the object descriptions are often diversified in contents and flexible in lengths. This leads to the key difficulty in RVOS, i.e., various descriptions of different objects are corresponding to different temporal scales in the video, which is ignored by most existing approaches with single stride of frame sampling. To tackle this problem, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) framework, which can effectively align linguistic and visual features to discover core object semantics in the video, by learning multimodal interaction hierarchically from different temporal scales. More specifically, we introduce a novel inter-scale multimodal perception module, where the language queries dynamically interact with visual features across temporal scales. It can effectively reduce complex object confusion by passing video context among different scales. Finally, we conduct extensive experiments on the widely used benchmarks, including Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences, where our HTML achieves state-of-the-art performance on all these datasets.", + "bbox": [ + 75, + 393, + 473, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring Video Object Segmentation (RVOS) has witnessed the growing interest, due to its wide applications in visual editing, virtual reality, human-robotic interaction and so on. Different from the traditional vision-only VOS, RVOS aims to segment the object instance from an input video, according to an open-world description about the re", + "bbox": [ + 76, + 785, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8785cf05a20cb459bacb962e746ece8fe3f5be378d46a37bb55b96cbe53b8a6e.jpg", + "image_caption": [ + "Figure 1: Referring descriptions in different lengths. (a) The description is simple containing only the category name. (b) The description is complicated with movement and position of the object. Single-scale baseline (e.g., four frames in (a) and two frames in (b)) fails to segment the referred object, while our hybrid-scale HTML succeeds. More discussion can be found in introduction." + ], + "image_footnote": [], + "bbox": [ + 514, + 359, + 883, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferred object. In this case, the model has to learn both visual and textual contents comprehensively, in order to discover the underlying object by multimodal interaction.", + "bbox": [ + 498, + 854, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 107, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13414", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies [4, 12, 28, 29] have shown that, cross-modal attention is an effective way to bridge the gap between vision and language in RVOS. However, these approaches perform vision-language interactions with video frames sampled from a single temporal scale, which may limit their power to infer the referred object with accurate segmentation. The main reason is that, the open-world descriptions vary in length and contain rich semantics about the referred object, e.g., where it is, how it moves, which objects it interact with. Apparently, such diversified texts are corresponding to various temporal-scale snippets.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, the language query in Fig. 1 (a) is a tennis ball. Such a short description is corresponding to the ball located at a small region in the middle two frames. If the single-scale baseline samples four frames as input, it will fail to segment the referred object. This is because it overlooks the dog in the center place among all these four frames, while lacking the detailed understanding in the middle two frames. Alternatively, the language query in Fig. 1 (b) is a sheep top second right moves down and comes out of the circle. Such a long description is corresponding to the particular sheep in the group, which moves across frames. If the single-scale baseline samples two frames as input, it will fail to segment the referred object. This is because it is misled by the subtle movement of sheep group in only two frames, without understanding how each sheep moves from the adjacent frames.", + "bbox": [ + 75, + 258, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this difficulty, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) Framework for RVOS, which can alleviate object confusion by language-vision interactions across different temporal scales. Specifically, we sample video frames according to different temporal scales. For each temporal scale, we introduce an intra-scale multimodal perception module, which can effectively exploit core visual semantics within the frames at this temporal scale, by mutual enhancement between textual and visual embeddings. Then, we design an inter-scale multimodal perception module, where linguistic embeddings dynamically interact with visual features across temporal scales. In this case, we can hierarchically leverage object context from all the temporal scales to boost RVOS. Finally, we evaluate our HTML on a number of benchmarks, including Ref-Youtu-VOS [22], Ref-DAVIS17 [9], A2D-Sentences and JHMDB-Sentences [7]. The extensive experiments have shown that, our HTML achieves the state-of-the-art performance on all of them.", + "bbox": [ + 75, + 500, + 472, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall we make three contributions in this paper:", + "bbox": [ + 94, + 789, + 428, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Concise and unified learning framework: our Hybrid Temporal-scale Multimodal Learning (HTML) framework hierarchically constructs multimodal interactions via different strides of frame sampling, which can mutually enhance embeddings from both modalities for accurate segmentation.", + "bbox": [ + 76, + 809, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Effective multimodal perception module: our Cross-scale Multimodal Perception (CMP) module can effectively reduce complex object confusions with intra-scale and inter-scale multimodal perceptions, where linguistic and visual features interact across temporal scales.", + "- State-of-the-art performance on the widely-used benchmarks, which shows the superiority of our framework. Specifically, on Ref-Youtu-VOS [22], our method with ResNet-50 achieves 57.8 in $\\mathcal{L}\\&\\mathcal{F}$ , outperforming the recent SOTA method [29] with ResNet-101." + ], + "bbox": [ + 500, + 90, + 893, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 256, + 645, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-only Video Segmentation. Vision-only video segmentation tasks [20, 19, 31, 1, 30, 27, 10, 32, 18, 25, 23, 36], need to segment the objects with predefined semantic query set. Video instance segmentation (VIS) [31] needs to segment the different instances in the desired category set and track each instance with its identity kept. Video object segmentation (VOS) [20, 19], the model needs to separate an object from the background in the video, given the mask of the first mask. In the early stage of VIS, heavy supervision and complex heuristic rules are applied to associate the instances across frames by separate segmentation and association branches, such as MaskTrackRCNN[31] and MaskProp[1]. Recent progress in transformer[24] and DETR[3] has inspired the community with end-to-end works, such as SeqFormer[30] and VisTR[27]. As for VOS, most recent works [32, 18, 25, 23, 36] on semi-supervised VOS enforces temporal consistency in the video to propagate the first-frame mask to the other frames sequentially. Most of them lie in the group of matching-based methods, updating the memories and matching to segment.", + "bbox": [ + 496, + 281, + 893, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Video Object Segmentation. RVOS task [22] is tasked to segment the referred object from the given video with specified open-world language description. Most early methods in RVOS proposed to refer the object by applying image-level methods on video frames separately and associate them with heuristic rules. However, they usually fail to utilize the temporal dynamic. [22] casts the task as a joint problem of referring segmentation in frame and mask propagation across frames by a memory attention module. [12] proposed a top-down pipeline by constructing exhaustive set of object tracklets and then selecting the target by matching the language features with the all the candidate tracklets. [37] proposed to model the temporal dynamic with an additional optical flow modality. [28] argued the importance of the structural information of video content and proposed to utilize the frame, object and video features simultaneously to obtain better representation. More recently, [2] introduced the DETR structure to RVOS area and [29] proposed to use language-conditional queries to simplify the referring pipeline and improve the performance, which serves as our baseline.", + "bbox": [ + 496, + 584, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13415", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from the previous works, we raise the mismatch issue that the various descriptions of different objects are corresponding to different temporal scales of the video. Moreover, we propose a concise HTML framework via multimodal interaction across different temporal scales to capture the core object semantics in the video.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 196, + 168, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To effectively align diversified descriptions and complex videos, we propose a distinct Hybrid Temporal-scale Multimodal Learning (HTML) framework for RVOS. In this section, we introduce our HTML in detail. First, we deliver an overview of HTML framework. Then, we explain how to build the hybrid temporal-scale multimodal learning paths, in the aid of vision-conditioned linguistic decoder and language-conditioned visual decoder. Next, we introduce a Cross-scale Multimodal Perception (CMP) module to align multimodal features across temporal scales. Finally, we describe the training objectives to optimize our HTML.", + "bbox": [ + 75, + 223, + 468, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Overview", + "text_level": 1, + "bbox": [ + 76, + 402, + 282, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our HTML framework consists of three main parts. First, we need to extract visual and linguistic features from backbones. We adopt a visual backbone to extract frame features from $T$ frames sampled from the given video. It can be either 2D CNN networks or 3D transformer networks. We then feed the extracted vision features into a deformable transformer encoder [38] to construct spatiotemporal relations between different frames. Meanwhile, to make a fair comparison with previous works in RVOS [29], we utilize the pretrained linguistic embedding model, RoBERTa [15], to extract textual features $\\mathbf{s}_e\\in$ $\\mathbb{R}^{L\\times C}$ from language descriptions with $L$ words. More details can be found in Sec. 4.2.", + "bbox": [ + 75, + 425, + 468, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After extracting visual and linguistic features, we next construct multimodal interactions between the language descriptions and the videos. Different from the previous approaches [29, 28], we build $L$ multimodal learning paths, where the linguistic embedding hierarchically interacts with visual features in different temporal scales. Then, we incorporate the mutually enhanced visual and linguistic features by a novel Cross-scale Multimodal Perception (CMP) module to align multimodal features across different scales. Finally, we design the training losses.", + "bbox": [ + 75, + 623, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hybrid Temporal-scale Multimodal Learning", + "text_level": 1, + "bbox": [ + 76, + 785, + 464, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To capture core object semantics, we propose a novel hybrid temporal-scale multimodal learning framework to learn multimodal relations. To start with, we build hybrid temporal scales via different sampling strides. Then, we construct basic multimodal relation learning units. Finally, we explain how to construct hybrid learning paths.", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Hybrid Temporal Scale Construction", + "text_level": 1, + "bbox": [ + 500, + 90, + 812, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the texts of various objects may refer to different video parts, single temporal scale often fails to describe the diversified textual contents. To simulate such diversity and flexibility, we build hybrid temporal scales by periodically sampling frames with different strides.", + "bbox": [ + 496, + 116, + 890, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first regard all the input frames as the first temporal scale, and then build other $L - 1$ temporal scales upon it in a sequential manner. In order to ensure the diversity of sampled temporal scale, we randomly pick one frame from every $h$ frames of last scale, where $h$ denotes the predefined stride. Subsequently, we feed the sampled frames into visual encoder Encoder(V) to extract feature maps for each of the scales respectively. Specifically, for temporal scale $l$ , we can obtain $\\mathbf{M} \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ , where $T$ denotes the number of frames in the temporal scale.", + "bbox": [ + 496, + 191, + 892, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2 Multimodal Relation Learning", + "text_level": 1, + "bbox": [ + 498, + 352, + 767, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to discover the core object semantics, we construct multimodal relations via vision-conditioned linguistic decoder and language-conditioned visual decoder to align semantics between different modalities.", + "bbox": [ + 496, + 376, + 890, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Conditioned Linguistic Decoder. In order to align linguistic object semantics to the vision contents, we design a vision-conditioned linguistic decoder $\\mathrm{Decoder}(\\mathrm{L}|\\mathrm{V})$ . Specifically, we have visual features $\\mathbf{M}\\in \\mathbb{R}^{T\\times H\\times W\\times C}$ , and linguistic embeddings $\\mathbf{s}_e\\in \\mathbb{R}^{1\\times C}$ . The vision-conditioned multimodal relations $\\mathbf{e} = \\mathrm{Decoder}(\\mathbf{s}_e|\\mathbf{M})$ are constructed by", + "bbox": [ + 496, + 441, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {0} = \\operatorname {D e f o r m A t t n} (\\mathbf {s} _ {e} + \\mathbf {q}, \\mathbf {M}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 561, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {k} = \\operatorname {D e f o r m A t t n} (\\mathbf {e} _ {k - 1}, \\mathbf {M}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 579, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $k \\in \\{1, \\dots, K - 1\\}$ . We first add $\\mathbf{s}_e$ with learnable queries $\\mathbf{q} \\in \\mathbb{R}^{N \\times C}$ to represent candidate instances in the video. Then, we use deformable attention module [38] to reason the vision-conditioned multimodal relations where vision features serve as key and value to decompose linguistic features, as in Eq. (1). Finally, we stack the cross-attention module for $K$ times, as in Eq. (2).", + "bbox": [ + 496, + 609, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-Conditioned Visual Decoder. In order to align visual object semantics to linguistic contents, we design a language-conditioned visual decoder $\\mathrm{Decoder}(\\mathrm{V}|\\mathrm{L})$ (similar to Eqs. (1) and (2)), to enhance visual representation with the attendance of the language description. Differently, vision features are enhanced by multi-head self-attention (MHSA) modules at the first place, and then linguistic features $\\mathbf{s}_e$ serve as key and value in cross-attention modules. In this case, it can reason the language-conditioned multimodal relations. Finally, we can get enhanced visual features by language-conditioned multimodal relations, as $\\mathbf{F} = \\mathrm{Decoder}(\\mathbf{M}|\\mathbf{s}_e)$ , where $\\mathbf{F} \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ .", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13416", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6368af0dd18156247a7285b9233ada7ff5ceeb39c76b857124cf3ac90575304.jpg", + "image_caption": [ + "Figure 2: Our Hybrid Temporal-scale Multimodal Learning framework. It aligns linguistic and visual features by learning hierarchical multimodal interactions with hybrid temporal scales, detailed in Sec. 3.2.2. Moreover, a Cross-scale Multimodal Perception (CMP) module is designed to enable interaction and cooperation among temporal scales, detailed in Sec. 3.3." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical Multimodal Learning As single-scale multimodal learning is insufficient to understand the relations between videos and texts, we propose to construct the multimodal relations hierarchically for the $L$ hybrid temporal scales with the assistance of $\\mathrm{Decoder}(\\mathrm{L}|\\mathrm{V})$ and $\\mathrm{Decoder}(\\mathrm{V}|\\mathrm{L})$ . We can obtain linguistic-attended visual features and visual-attended linguistic embeddings for different temporal scales, capturing core object semantics conditioned on different visual and linguistic contexts.", + "bbox": [ + 75, + 409, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Cross-scale Multimodal Perception", + "text_level": 1, + "bbox": [ + 76, + 554, + 382, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multimodal relations are constructed conditioned on different modalities with hybrid temporal scales. However, the modeling process of different scales is independent. To promote the cooperation and align the visual and linguistic semantics both within the scale and across the scales, we design Cross-scale Multimodal Perception (CMP) module.", + "bbox": [ + 75, + 577, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intra-scale Perception. Despite sharing visual and linguistic feature extraction, the multimodal relation construction via Decoder(V|L) and Decoder(L|V) are independent to each other. To promote the cooperation between modalities, we propose an intra-scale perception module.", + "bbox": [ + 75, + 671, + 468, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, in each temporal scale $l$ , we have visual attended linguistic embeddings $\\mathbf{e} \\in \\mathbb{R}^{N \\times T \\times D}$ and linguistic attended visual features $\\mathbf{F} \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ . To achieve fine-grained semantic alignment, we measure the similarity by dot product between $\\mathbf{e}$ and $\\mathbf{F}$ on pixel level. Specifically, we obtain the similarity map via multimodal perception module, denoted as $\\mathcal{I} = \\mathrm{MP}(\\mathbf{F}, \\mathbf{e})$ by", + "bbox": [ + 75, + 747, + 468, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega = \\operatorname {M a s k H e a d} (\\mathbf {e}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 861, + 468, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} = \\Omega \\cdot \\mathbf {F}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 881, + 468, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where MaskHead denotes three consecutive MLP layers for embedding conversion. Each value of $\\mathcal{I}$ represents the relevance between visual-attended linguistic embeddings and linguistic-attended visual features, which can be interpreted as the existence of the referred object. As such, $\\mathcal{I}$ is regarded as the object mask prediction with the context of current temporal scale. To this end, we achieve the multimodal perception in the same temporal scale.", + "bbox": [ + 496, + 409, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inter-scale Perception. The multimodal relations are constructed in different temporal scales. However, the process in each scale is independent and biased towards the contained object semantics. To alleviate it, we propose to align multimodal features from different temporal scales with a concise inter-scale perception module.", + "bbox": [ + 496, + 534, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, suppose that the referred object appears in frame $t$ in temporal scales $l$ and $l + 1$ , the referred object can be segmented by measuring similarity between $(\\mathbf{F}^l (t),\\mathbf{e}^l (t))$ and $(\\mathbf{F}^{l + 1}(t),\\mathbf{e}^{l + 1}(t))$ simultaneously. Conditioned on same frame $t$ , the visual-attended linguistic embedding $\\mathbf{e}^l (t)$ from scale $l$ is supposed to be relevant to the linguistic-attended visual features $\\mathbf{F}^{l + 1}(t)$ from scale $l + 1$ . Thus, similar to Eq. (4), the similarity cross different temporal scales can be measured by", + "bbox": [ + 496, + 625, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} ^ {l \\rightarrow l + 1} (t) = \\mathrm {M P} (\\mathbf {e} ^ {l} (t), \\mathbf {F} ^ {l + 1} (t)). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 767, + 890, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Without losing generality, the inter-scale similarity can also be measured by $\\mathcal{I}^{l + 1\\to l}(t)$ . More specifically, frame $t$ can be any frame shared by the adjacent temporal scales, which is ensured by our hybrid temporal scales sampling strategy. Each value in $\\mathcal{I}^{l\\rightarrow l + 1}(t)$ represents the referred object prediction with the prior of linguistic object semantic from temporal scale $l$ . In the same way, values in $\\mathcal{I}^{l + 1\\rightarrow l}(t)$ rep", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13417", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resent the opposite. To this end, we achieve multimodal perception across different temporal scales.", + "bbox": [ + 75, + 90, + 468, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training objectives", + "text_level": 1, + "bbox": [ + 76, + 128, + 261, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our network can be trained in an end-to-end manner to locate and segment the target instance simultaneously. Specifically, the losses for intra-scale and inter-scale multimodal perception in temporal scale $l$ are formed as", + "bbox": [ + 75, + 151, + 467, + 208 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {i n t r a}} ^ {l} = \\sum \\mathcal {L} _ {\\text {c l s}} (\\mathbf {y}, \\hat {\\mathbf {y}}) + \\mathcal {L} _ {\\text {b o x}} (\\mathbf {b}, \\hat {\\mathbf {b}}) + \\mathcal {L} _ {\\text {m a s k}} (\\mathcal {I}, \\hat {\\mathcal {I}}), (6) \\\\ \\mathcal {L} _ {\\text {i n t e r}} ^ {l} = \\sum \\mathcal {L} _ {\\text {m a s k}} \\left(\\mathcal {I} ^ {l + 1 \\rightarrow l}, \\hat {\\mathcal {I}}\\right) (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 215, + 467, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where time and instance subscripts are omitted for simplicity, $\\mathbf{y}$ and $\\mathbf{b}$ denote binary classification for instance existence and bounding box prediction respectively. Here $\\mathcal{L}_{cls}$ is the focal loss [13], $\\mathcal{L}_{box}$ is the sum of L1 loss and GIoU loss [21], and $\\mathcal{L}_{mask}$ is the combination of DICE loss [17] and binary mask focal loss. We optimize the network by first finding the best prediction as the positive sample, via minimizing the matching cost $\\mathcal{L}_{intra}^{l}$ and $\\mathcal{L}_{inter}^{l}$ in each temporal scale $l$ respectively. Then, we average the matching losses from different temporal scales and perception modules, and minimize it for positive samples.", + "bbox": [ + 75, + 263, + 467, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 443, + 209, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 468, + 274, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on four datasets: Ref-Youtube-VOS[22], Ref-DAVIS17[9], A2D-Sentences and JHMDB-Sentences[7], following the common practice[29]. Metrics. We follow the standard evaluation protocol [22, 28, 29] to adopt region similarity $\\mathcal{L}(\\%)$ , contour accuracy $\\mathcal{F}(\\%)$ and mean $\\mathcal{L}\\&\\mathcal{F}$ for Ref-Youtube-VOS and Ref-DAVIS17. For JHMDB-Sentences, we adopt mAP to evaluate the model. For A2D-Sentences, we use Precision@K, Overall IoU, Mean IoU and mAP for evaluation.", + "bbox": [ + 75, + 491, + 467, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implemented Details", + "text_level": 1, + "bbox": [ + 76, + 635, + 272, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set the number of attention layers $K$ to 4 and the hidden dimension $C$ to 256. The number of learnable queries $N$ is set to 5. The number of MaskHead output channels $D$ is set to 8. During training, we first sample video clips by sliding windows and then generate $L = 3$ hybrid temporal scales with stride $h = 2$ for generalized multimodal representations and relations. We use same training recipes as in [29, 2]. All frames are downsampled by shorter side to 360 and limit the maximum size for the long side to 640. Our model is pretrained on image referring segmentation datasets [35, 35]. See supplementary materials for details.", + "bbox": [ + 75, + 657, + 467, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, we report results with all input frames in single temporal scale for fair comparisons. On RefDAVIS17, we directly inference on models trained on RefYoutube-VOS. Similarly, we reports JHMDB-Senteces results directly on models trained with A2D-Sentences.", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/30cdba508ff2980263bce985a27c043bbb16274fee3d1fcdc5b3b409baa39480.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneRef-Youtube-VOS
J&FJF
CMSA [33]ResNet-5034.933.336.5
CMSA + RNN [33]ResNet-5036.434.838.1
URVOS [22]ResNet-5047.245.349.2
LBDT-4 [22]ResNet-5047.245.349.2
MLRL [28]ResNet-5049.748.451.0
ReferFormer [29]ResNet-5055.654.856.5
OursResNet-5057.856.559.0
PMINet [6]ResNeSt-10148.246.749.6
PMINet + CFBI [6]ResNeSt-10153.051.554.5
CITD [12]ResNet-10156.454.858.1
ReferFormer [29]ResNet-10157.356.158.4
OursResNet-10158.557.359.8
PMINet + CFBI [6]Ensemble54.253.055.5
CITD [12]Ensemble61.460.062.7
ReferFormer [29]Swin-L62.460.864.0
OursSwin-L63.461.565.3
MTTR [2]Video-Swin-T55.354.056.6
ReferFormer [29]Video-Swin-T59.458.060.9
OursVideo-Swin-T61.259.563.0
ReferFormer [29]Video-Swin-S60.158.661.6
OursVideo-Swin-S61.459.962.9
ReferFormer [29]Video-Swin-B62.961.364.6
OursVideo-Swin-B63.461.565.2
", + "bbox": [ + 506, + 90, + 882, + 452 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f10f81c2064fac0eded195187188d829e23e6981d65b27056ab94f7914969049.jpg", + "table_caption": [ + "Table 1: Comparison with the SOTA methods on Ref-YTB-VOS." + ], + "table_footnote": [], + "table_body": "
MethodBackboneRef-DAVIS17
J&FJF
CMSA [33]ResNet-5034.732.237.2
CMSA + RNN [33]ResNet-5040.236.943.5
URVOS [22]ResNet-5051.547.356.0
LBDT-4 [5]ResNet-5054.5--
MLRL [28]ResNet-5058.053.962.0
ReferFormer [29]ResNet-5058.555.861.3
OursResNet-5059.556.662.4
ReferFormer [29]Swin-L60.557.663.4
OursSwin-L61.658.964.4
ReferFormer [29]Video-Swin-B61.158.164.1
OursVideo-Swin-B62.159.265.1
", + "bbox": [ + 506, + 501, + 882, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Comparison with the SOTA methods on Ref-DAVIS17.", + "bbox": [ + 498, + 703, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. SOTA Comparisons", + "text_level": 1, + "bbox": [ + 500, + 755, + 689, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with the state-of-the-art methods on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences. On Ref-Youtu-VOS, our approach achieves 58.5 in $\\mathcal{L}\\&\\mathcal{F}(\\%)$ with ResNet-50, as shown in Tab. 1, which surpasses the recent SOTA method ReferFormer[29] with same backbone by 2.2 points. Moreover, it surpasses all the other SOTA methods with larger ResNet-101 on all evaluation metrics, which fully sug", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13418", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b252803b50c2537e6c1985c943523864a1279f72c82652d893cd749f1bb671b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonePrecisionIoUmAP
P@0.5P@0.6P@0.7P@0.8P@0.9OverallMean
Hu et al. [8]VGG-1634.823.613.33.30.147.435.013.2
Gavrilyuk et al. [7]I3D47.534.721.18.00.253.642.119.8
CMSA + CFSA [34]ResNet-10148.743.135.823.15.261.843.2-
ACAN [26]I3D55.745.931.916.02.060.149.027.4
CMPC-V [14]I3D65.559.250.634.29.865.357.340.4
ClawCraneNet [11]ResNet-50/10170.467.761.748.917.163.159.9-
MTTR (ω = 8) [2]Video-Swin-T72.168.460.745.616.470.261.844.7
MTTR (ω = 10) [2]Video-Swin-T75.471.263.848.516.972.064.046.1
ReferFormer [29]Video-Swin-T82.879.272.355.319.377.669.652.8
ReferFormer [29]Video-Swin-B83.180.474.157.921.278.670.355.0
OursVideo-Swin-T82.279.272.355.320.177.669.253.4
OursVideo-Swin-B84.081.575.859.222.879.571.256.7
", + "bbox": [ + 89, + 88, + 880, + 295 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/455fa75b1c75f1c1e22e36fef8d21a9197f204fecf42e4d98c7df905c500f683.jpg", + "table_caption": [ + "Table 3: Comparison with the state-of-the-art methods on A2D-Sentences." + ], + "table_footnote": [], + "table_body": "
MethodBackbonemAP
Hu et al. [8]VGG-1617.8
Gavrilyuk et al. [7]I3D23.3
ACAN [26]I3D28.9
CMPC-V [14]I3D34.2
MTTR (ω = 8) [2]Video-Swin-T36.6
MTTR (ω = 10) [2]Video-Swin-T39.2
ReferFormer† (ω = 6) [29]Video-Swin-T39.1
ReferFormer [29]Video-Swin-T42.2
ReferFormer [29]Video-Swin-B43.7
OursVideo-Swin-T42.7
OursVideo-Swin-B44.2
", + "bbox": [ + 101, + 335, + 444, + 513 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4702bec8715d7c8f321165fbc5bd208f88a3f9653b2f4735eb616ca3e22eca81.jpg", + "table_caption": [ + "Table 4: SOTA results comparison on JHMDB-Sentences." + ], + "table_footnote": [], + "table_body": "
Components#FramesJ&FJF
i.Baseline555.654.856.5
ii.HTML w/o CMP556.054.757.4
iii.HTML556.355.057.5
iv.Baseline856.255.057.3
v.HTML w/o CMP857.156.058.2
vi.HTML857.856.559.0
", + "bbox": [ + 109, + 554, + 437, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5: Ablation study on the components of our HTML.", + "bbox": [ + 80, + 672, + 464, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gests the superiority of our method. When equipped with larger backbone, our method still show considerable superiority with accuracy gap of 1.2 points for ResNet-101 and 1.0 points for Swin-L. We also experiment our method with the well-known Video Swin Transformers [16]. Our method with Video-Swin-Tiny backbone surpasses the SOTA method with the same backbone by 1.8 points. With larger Video-Swin Transformers (Small and Base models), our method still achieves SOTA performance, which shows the generality of our method.", + "bbox": [ + 75, + 702, + 467, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, our method surpasses the SOTA methods on Ref-DAVIS17 by over 1.0 points on both ResNet-50, Swin-L and Video-Swin-Base backbones, with", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "new a SOTA record 62.1 in $\\mathcal{L}\\&\\mathcal{F}(\\%)$ . We also experiment our method on A2D-Sentences and JHMDB-Sentences datasets and compare with other SOTA results as shown in Tab. 3 and Tab. 4. Our method achieves SOTA performances with new records on both the two datasets. On A2D-Sentences, our HTML surpass SOTA result by 1.7 points in mAP and higher recall by 1.6 on Precision@0.9. On JHMDB-Sentences, our method still achieves a new SOTA record with 44.2 in mAP. These results demonstrate the superiority of our method.", + "bbox": [ + 496, + 338, + 890, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 500, + 653, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we ablate core components of our HTML with Ref-Youtu-VOS based on ResNet-50.", + "bbox": [ + 500, + 522, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of our HTML. To validate the effectiveness of our Hybrid Temporal-scale Multimodal Learning framework, we investigate each of our components by gradually adding them to the baseline [29]. First, comparing (i)&(iii) in Tab. 5, our HTML improves the performance of baseline to 56.3 when 5 frames are used for training, which is better than the baseline trained with longer input frames. Further, when longer temporal input is available, comparing (iv)&(vi), our HTML improves the model by 1.6 to 57.8 in $\\mathcal{L}\\&\\mathcal{F}$ . These prove the effectiveness of our method with frames in different lengths.", + "bbox": [ + 496, + 551, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, comparing counterpart networks using different number of frames, i.e. (iii) and (vi), our method can benefit from longer temporal input (8 frames vs. 5 frames) with an improvement of 1.5 points. This conclusion holds among the other counterpart network pair, i.e. (ii)&(v). Further, each of our components brings an improvement of 0.3 when trained with total 5 frames, while the improvement can be doubled to 0.7-0.9 with total 8 frames. This indicates that our method can better utilize the long temporal input.", + "bbox": [ + 496, + 719, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, taking 8 frames for instance, hierarchical multimodal learning and cross-scale multimodal perception improves $\\mathcal{L}\\&\\mathcal{F}$ by 0.9 and 0.7 respectively. It proves the ef", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13419", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/38707ac44701fc2730f3a0f164fbfe62f1b95df7d172ded15e0ec6afd77ee95a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#ScalesJ&FJF
156.255.057.3
256.955.858.1
357.856.559.0
", + "bbox": [ + 107, + 88, + 330, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36d6ca58585e8abab7372253db889acf8d61f12c6431af587ee11d47564643ab.jpg", + "table_caption": [ + "(a) Effect of No. of temporal scales." + ], + "table_footnote": [], + "table_body": "
FramesJ&FJF
556.355.057.5
857.856.559.0
1257.556.458.6
", + "bbox": [ + 370, + 90, + 589, + 164 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e45299fb6975b5804943ef1e51a7855223358517f3015f8d62bd13973f9c6c61.jpg", + "table_caption": [ + "(b) Effect of No. of input frames." + ], + "table_footnote": [], + "table_body": "
Inter-scaleJ&FJF
None56.255.057.3
l → l + 156.955.858.0
l + 1 → l57.856.559.0
", + "bbox": [ + 625, + 88, + 862, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(c) Effect of direction of CMP.", + "bbox": [ + 640, + 174, + 823, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6: Ablation study on Hybrid Temporal Scales in (a) and (b), and Cross-scale Multimodal Perception in (c).", + "bbox": [ + 112, + 199, + 854, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e96e42331b76dfa19e8048dac016a0ebf27b28ad5019ade6b42338b8e257168e.jpg", + "image_caption": [ + "Figure 3: Performance comparison of different query sets with different temporal scales on Ref-Youtu-VOS." + ], + "image_footnote": [], + "bbox": [ + 102, + 234, + 450, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fectiveness of each component of our HTML.", + "bbox": [ + 76, + 534, + 380, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Length of language descriptions. We explore the performance of our method with different sampled sets of language descriptions, to validate the ability of capturing diversified linguistic object semantics. Since the train set of Ref-Youtu-VOS contains an average of two language descriptions for each object, we sample the shorter ones to form the Short set and the others to form the Long set.", + "bbox": [ + 75, + 551, + 468, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in bottom line of Fig. 3, the Long set has longer sentences than the Short set in the average number of words. As shown in Fig. 3, when the objects are described by Short query set, more input frames in single temporal scale achieve inferior performance (5 frames vs. 8 frames: 55.6 vs. 55.1). It's interesting to note that more visual content (8 frames) fails to improve the performance when the linguistic content is insufficient (shorter queries). When the complexity of queries increases, i.e. with the Long query set, the quantitative relation is reversed: 8 frames guided model obtains better performance than the model with 5 frames, increasing by 0.8 (8 frames-Short query vs. 8 frames-Long query: 55.1 vs. 55.9). We infer that this is caused by the mismatch of visual and linguistic object semantics. On one side, shorter quires, i.e. relatively simple linguistic semantics, are insufficient to interpret longer", + "bbox": [ + 75, + 657, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "videos. On the other side, the longer language descriptions contain more content irrelevant to the visual input.", + "bbox": [ + 496, + 234, + 890, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Differently, our models trained with either the Short set or the Long set all surpass the single temporal scale guided model. Impressively, when the objects are described by queries flexible in lengths, i.e. with the All set, our method gets a performance boost by 1.6, while single temporal scale baseline (both 8 frames and 5 frames guided networks) improves by 0.3. This shows that our method has the strong ability to solve the mismatch issue between visual content and diversified linguistic contents.", + "bbox": [ + 496, + 265, + 892, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of hybrid temporal scales. We investigate the effectiveness of our hierarchical multimodal learning, by exploring the number of hybrid temporal scales. For fair comparison among different settings, the number of total input frames is set to 8 in this subsection. As shown in Tab. 6 (a), when two temporal scales are constructed, our method brings an improvement of 0.6 in $\\mathcal{L}\\&\\mathcal{F}$ . When the number of temporal scales is increased to three, continuous improvement of 0.9 (56.9 vs. 57.8) is observed. It proves that our model benefits from the increasing visual diversity constructed by hybrid temporal scales.", + "bbox": [ + 496, + 401, + 892, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of input frames. We explore the effect of total input frames here. In this subsection, hybrid temporal scales are constructed by default following Sec. 3.2.2. As shown in Tab. 6 (b), more input frames bring an improvement by 1.5 in $\\mathcal{L}\\&\\mathcal{F}$ (5 frames vs. 8frames: 56.3 vs. 57.8). When the input frames increase continuously to 12 frames, the performance saturates and drops slightly to 57.5. We conjecture that it is caused by insufficient video-language pairs (8 frames vs. 12 frames: $49\\mathrm{k}$ vs. $32\\mathrm{k}$ ) compared to largely increased computation complexity (8 frames vs. 12 frames: 21.5 GFLOPs vs. 33.6 GFLOPs for transformers).", + "bbox": [ + 496, + 566, + 892, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Direction of inter-scale perception. We explore the effect of direction of inter-scale multimodal perception in CMP. Comparing first two lines of Tab. 6 (c), $l \\rightarrow l + 1$ perception improves the baseline by 0.7 points; As in first and last lines, $l + 1 \\rightarrow l$ perception improves the baseline by 1.6 points. These prove the effectiveness of our inter-scale perception. We choose the latter one for better performance.", + "bbox": [ + 496, + 733, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualizations", + "text_level": 1, + "bbox": [ + 500, + 847, + 643, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We visualize the segmentation results of complex and simple language descriptions in Fig. 4. Specifically, we", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13420", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1aa1a49843365ea85fd4757ba4fb67cfefb379837cf3e66a2d12f7c5c8bb6336.jpg", + "image_caption": [ + "Figure 4: Visualization results of complex and simple language descriptions on Ref-Youtu-VOS. Red masks indicate positive segmentation results and blue masks indicate the negatives. Our HTML is able to clarify such object confusion." + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 846, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compare three settings, i.e., baseline with only single temporal scale, HTML w/o CMP which constructs hybrid temporal scales, and our final HTML which dynamically construct multimodal relations cross temporal scales. As expected, when only single temporal scale is adopted, baseline fails to segment the target in the video, e.g., the background rabbit in the subplot (a) is mistakenly referred. The main reason is that the rabbit shares similar appearance to the target object and also locates to the left of a person in last three frames. The model is misled and confused by the single temporal scale. When applied with hybrid temporal scales, the language description can interact with both long and short temporal scales. Thus, the previous false prediction in the first frame is corrected. Further, when applied with CMP, our model is able to clarify the object confusion by discovering the core semantics on a scale and make correct predictions. Similarly, in subplot (b), baseline is misled by the video clip where tennis ball only appears in the middle two frames. When gradually applying our proposed modules, the mistakenly predicted dog is clarified and further the target tennis ball is correctly segmented.", + "bbox": [ + 75, + 566, + 472, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 565, + 619, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we develop a HTML framework to align linguistic and visual features by learning multimodal relations hierarchically in different temporal scales. Moreover, we introduce an inter-scale multimodal perception module to construct dynamic multimodal interactions across temporal scales. We conduct experiments on four datasets and establish new state-of-the-art results. Particularly, our method with ResNet-50 backbone surpasses the recent methods with ResNet-100. The comprehensive ablation experiments and visualization results show that our method is able to discover core object semantics in the different modalities.", + "bbox": [ + 496, + 590, + 890, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 660, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (No.2022ZD0160100, No. 2022ZD0160505), the National Natural Science Foundation of China under Grant (62272450), the Joint Lab of CAS-HK, the Shenzhen Research Program (RCJC20200 714114557087), and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020355)", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13421", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2020. 2", + "[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. arXiv preprint arXiv:2111.14821, 2021. 2, 5, 6", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2", + "[4] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2", + "[5] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 5", + "[6] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 5", + "[7] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6", + "[8] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6", + "[9] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 2, 5", + "[10] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18706-18716, 2023. 2", + "[11] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 6", + "[12] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2, 5" + ], + "bbox": [ + 78, + 116, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5", + "[14] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6", + "[15] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3", + "[16] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021. 6", + "[17] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5", + "[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 2", + "[19] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2", + "[20] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2", + "[21] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5", + "[22] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 5", + "[23] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 2", + "[24] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 2", + "[25] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmenta" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13422", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2", + "[26] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6", + "[27] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2", + "[28] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 2, 3, 5", + "[29] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 3, 5, 6", + "[30] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2", + "[31] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2", + "[32] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In European Conference on Computer Vision, pages 332-348. Springer, 2020. 2", + "[33] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10502–10511, 2019. 5", + "[34] Linwei Ye, Mrigank Rochan, Zhi Liu, Xiaoqin Zhang, and Yang Wang. Referring segmentation in images and videos with cross-modal self-attention network. arXiv preprint arXiv:2102.04762, 2021. 6", + "[35] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 5", + "[36] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2246-2256, 2023. 2", + "[37] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3", + "bbox": [ + 501, + 90, + 892, + 146 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13423", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_model.json b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f22cc9b0e01bda3b3d21a7a73fdb8ee505b7ef40 --- /dev/null +++ b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.13, + 0.838, + 0.177 + ], + "angle": 0, + "content": "HTML: Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.203, + 0.82, + 0.222 + ], + "angle": 0, + "content": "Mingfei Han\\(^{1,4}\\), Yali Wang\\(^{\\dagger,2,6}\\), Zhihui Li\\(^{3}\\), Lina Yao\\(^{4}\\), Xiaojun Chang\\(^{1,5}\\), Yu Qiao\\(^{6,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.222, + 0.891, + 0.327 + ], + "angle": 0, + "content": "\\(^{1}\\)ReLER, AAII, UTS \\(^{2}\\)Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, SIAT-SenseTime Joint Lab, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences \\(^{3}\\)Shandong Artificial Intelligence, Qilu University of Technology \\(^{4}\\)Data61, CSIRO \\(^{5}\\)Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence \\(^{6}\\)Shanghai AI Laboratory, Shanghai, China https://mingfei.info/HTML/" + }, + { + "type": "title", + "bbox": [ + 0.259, + 0.362, + 0.337, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.474, + 0.741 + ], + "angle": 0, + "content": "Referring Video Object Segmentation (RVOS) is to segment the object instance from a given video, according to the textual description of this object. However, in the open world, the object descriptions are often diversified in contents and flexible in lengths. This leads to the key difficulty in RVOS, i.e., various descriptions of different objects are corresponding to different temporal scales in the video, which is ignored by most existing approaches with single stride of frame sampling. To tackle this problem, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) framework, which can effectively align linguistic and visual features to discover core object semantics in the video, by learning multimodal interaction hierarchically from different temporal scales. More specifically, we introduce a novel inter-scale multimodal perception module, where the language queries dynamically interact with visual features across temporal scales. It can effectively reduce complex object confusion by passing video context among different scales. Finally, we conduct extensive experiments on the widely used benchmarks, including Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences, where our HTML achieves state-of-the-art performance on all these datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.76, + 0.21, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.786, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Referring Video Object Segmentation (RVOS) has witnessed the growing interest, due to its wide applications in visual editing, virtual reality, human-robotic interaction and so on. Different from the traditional vision-only VOS, RVOS aims to segment the object instance from an input video, according to an open-world description about the re" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.361, + 0.885, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Figure 1: Referring descriptions in different lengths. (a) The description is simple containing only the category name. (b) The description is complicated with movement and position of the object. Single-scale baseline (e.g., four frames in (a) and two frames in (b)) fails to segment the referred object, while our hybrid-scale HTML succeeds. More discussion can be found in introduction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.893, + 0.903 + ], + "angle": 0, + "content": "ferred object. In this case, the model has to learn both visual and textual contents comprehensively, in order to discover the underlying object by multimodal interaction." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.888, + 0.238, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13414" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Recent studies [4, 12, 28, 29] have shown that, cross-modal attention is an effective way to bridge the gap between vision and language in RVOS. However, these approaches perform vision-language interactions with video frames sampled from a single temporal scale, which may limit their power to infer the referred object with accurate segmentation. The main reason is that, the open-world descriptions vary in length and contain rich semantics about the referred object, e.g., where it is, how it moves, which objects it interact with. Apparently, such diversified texts are corresponding to various temporal-scale snippets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.501 + ], + "angle": 0, + "content": "For example, the language query in Fig. 1 (a) is a tennis ball. Such a short description is corresponding to the ball located at a small region in the middle two frames. If the single-scale baseline samples four frames as input, it will fail to segment the referred object. This is because it overlooks the dog in the center place among all these four frames, while lacking the detailed understanding in the middle two frames. Alternatively, the language query in Fig. 1 (b) is a sheep top second right moves down and comes out of the circle. Such a long description is corresponding to the particular sheep in the group, which moves across frames. If the single-scale baseline samples two frames as input, it will fail to segment the referred object. This is because it is misled by the subtle movement of sheep group in only two frames, without understanding how each sheep moves from the adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.789 + ], + "angle": 0, + "content": "To tackle this difficulty, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) Framework for RVOS, which can alleviate object confusion by language-vision interactions across different temporal scales. Specifically, we sample video frames according to different temporal scales. For each temporal scale, we introduce an intra-scale multimodal perception module, which can effectively exploit core visual semantics within the frames at this temporal scale, by mutual enhancement between textual and visual embeddings. Then, we design an inter-scale multimodal perception module, where linguistic embeddings dynamically interact with visual features across temporal scales. In this case, we can hierarchically leverage object context from all the temporal scales to boost RVOS. Finally, we evaluate our HTML on a number of benchmarks, including Ref-Youtu-VOS [22], Ref-DAVIS17 [9], A2D-Sentences and JHMDB-Sentences [7]. The extensive experiments have shown that, our HTML achieves the state-of-the-art performance on all of them." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.79, + 0.429, + 0.805 + ], + "angle": 0, + "content": "Overall we make three contributions in this paper:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.81, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- Concise and unified learning framework: our Hybrid Temporal-scale Multimodal Learning (HTML) framework hierarchically constructs multimodal interactions via different strides of frame sampling, which can mutually enhance embeddings from both modalities for accurate segmentation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "- Effective multimodal perception module: our Cross-scale Multimodal Perception (CMP) module can effectively reduce complex object confusions with intra-scale and inter-scale multimodal perceptions, where linguistic and visual features interact across temporal scales." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.169, + 0.894, + 0.245 + ], + "angle": 0, + "content": "- State-of-the-art performance on the widely-used benchmarks, which shows the superiority of our framework. Specifically, on Ref-Youtu-VOS [22], our method with ResNet-50 achieves 57.8 in \\(\\mathcal{L}\\&\\mathcal{F}\\), outperforming the recent SOTA method [29] with ResNet-101." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.257, + 0.646, + 0.273 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Vision-only Video Segmentation. Vision-only video segmentation tasks [20, 19, 31, 1, 30, 27, 10, 32, 18, 25, 23, 36], need to segment the objects with predefined semantic query set. Video instance segmentation (VIS) [31] needs to segment the different instances in the desired category set and track each instance with its identity kept. Video object segmentation (VOS) [20, 19], the model needs to separate an object from the background in the video, given the mask of the first mask. In the early stage of VIS, heavy supervision and complex heuristic rules are applied to associate the instances across frames by separate segmentation and association branches, such as MaskTrackRCNN[31] and MaskProp[1]. Recent progress in transformer[24] and DETR[3] has inspired the community with end-to-end works, such as SeqFormer[30] and VisTR[27]. As for VOS, most recent works [32, 18, 25, 23, 36] on semi-supervised VOS enforces temporal consistency in the video to propagate the first-frame mask to the other frames sequentially. Most of them lie in the group of matching-based methods, updating the memories and matching to segment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Referring Video Object Segmentation. RVOS task [22] is tasked to segment the referred object from the given video with specified open-world language description. Most early methods in RVOS proposed to refer the object by applying image-level methods on video frames separately and associate them with heuristic rules. However, they usually fail to utilize the temporal dynamic. [22] casts the task as a joint problem of referring segmentation in frame and mask propagation across frames by a memory attention module. [12] proposed a top-down pipeline by constructing exhaustive set of object tracklets and then selecting the target by matching the language features with the all the candidate tracklets. [37] proposed to model the temporal dynamic with an additional optical flow modality. [28] argued the importance of the structural information of video content and proposed to utilize the frame, object and video features simultaneously to obtain better representation. More recently, [2] introduced the DETR structure to RVOS area and [29] proposed to use language-conditional queries to simplify the referring pipeline and improve the performance, which serves as our baseline." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13415" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "Different from the previous works, we raise the mismatch issue that the various descriptions of different objects are corresponding to different temporal scales of the video. Moreover, we propose a concise HTML framework via multimodal interaction across different temporal scales to capture the core object semantics in the video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.198, + 0.169, + 0.213 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.391 + ], + "angle": 0, + "content": "To effectively align diversified descriptions and complex videos, we propose a distinct Hybrid Temporal-scale Multimodal Learning (HTML) framework for RVOS. In this section, we introduce our HTML in detail. First, we deliver an overview of HTML framework. Then, we explain how to build the hybrid temporal-scale multimodal learning paths, in the aid of vision-conditioned linguistic decoder and language-conditioned visual decoder. Next, we introduce a Cross-scale Multimodal Perception (CMP) module to align multimodal features across temporal scales. Finally, we describe the training objectives to optimize our HTML." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.403, + 0.283, + 0.417 + ], + "angle": 0, + "content": "3.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.621 + ], + "angle": 0, + "content": "As shown in Fig. 2, our HTML framework consists of three main parts. First, we need to extract visual and linguistic features from backbones. We adopt a visual backbone to extract frame features from \\(T\\) frames sampled from the given video. It can be either 2D CNN networks or 3D transformer networks. We then feed the extracted vision features into a deformable transformer encoder [38] to construct spatiotemporal relations between different frames. Meanwhile, to make a fair comparison with previous works in RVOS [29], we utilize the pretrained linguistic embedding model, RoBERTa [15], to extract textual features \\(\\mathbf{s}_e\\in\\) \\(\\mathbb{R}^{L\\times C}\\) from language descriptions with \\(L\\) words. More details can be found in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.775 + ], + "angle": 0, + "content": "After extracting visual and linguistic features, we next construct multimodal interactions between the language descriptions and the videos. Different from the previous approaches [29, 28], we build \\(L\\) multimodal learning paths, where the linguistic embedding hierarchically interacts with visual features in different temporal scales. Then, we incorporate the mutually enhanced visual and linguistic features by a novel Cross-scale Multimodal Perception (CMP) module to align multimodal features across different scales. Finally, we design the training losses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.465, + 0.803 + ], + "angle": 0, + "content": "3.2. Hybrid Temporal-scale Multimodal Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To capture core object semantics, we propose a novel hybrid temporal-scale multimodal learning framework to learn multimodal relations. To start with, we build hybrid temporal scales via different sampling strides. Then, we construct basic multimodal relation learning units. Finally, we explain how to construct hybrid learning paths." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.813, + 0.107 + ], + "angle": 0, + "content": "3.2.1 Hybrid Temporal Scale Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.117, + 0.892, + 0.192 + ], + "angle": 0, + "content": "Since the texts of various objects may refer to different video parts, single temporal scale often fails to describe the diversified textual contents. To simulate such diversity and flexibility, we build hybrid temporal scales by periodically sampling frames with different strides." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.193, + 0.893, + 0.344 + ], + "angle": 0, + "content": "We first regard all the input frames as the first temporal scale, and then build other \\( L - 1 \\) temporal scales upon it in a sequential manner. In order to ensure the diversity of sampled temporal scale, we randomly pick one frame from every \\( h \\) frames of last scale, where \\( h \\) denotes the predefined stride. Subsequently, we feed the sampled frames into visual encoder Encoder(V) to extract feature maps for each of the scales respectively. Specifically, for temporal scale \\( l \\), we can obtain \\( \\mathbf{M} \\in \\mathbb{R}^{T \\times H \\times W \\times C} \\), where \\( T \\) denotes the number of frames in the temporal scale." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.353, + 0.768, + 0.368 + ], + "angle": 0, + "content": "3.2.2 Multimodal Relation Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.437 + ], + "angle": 0, + "content": "In order to discover the core object semantics, we construct multimodal relations via vision-conditioned linguistic decoder and language-conditioned visual decoder to align semantics between different modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Vision-Conditioned Linguistic Decoder. In order to align linguistic object semantics to the vision contents, we design a vision-conditioned linguistic decoder \\(\\mathrm{Decoder}(\\mathrm{L}|\\mathrm{V})\\). Specifically, we have visual features \\(\\mathbf{M}\\in \\mathbb{R}^{T\\times H\\times W\\times C}\\), and linguistic embeddings \\(\\mathbf{s}_e\\in \\mathbb{R}^{1\\times C}\\). The vision-conditioned multimodal relations \\(\\mathbf{e} = \\mathrm{Decoder}(\\mathbf{s}_e|\\mathbf{M})\\) are constructed by" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.562, + 0.891, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {0} = \\operatorname {D e f o r m A t t n} (\\mathbf {s} _ {e} + \\mathbf {q}, \\mathbf {M}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.58, + 0.891, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {k} = \\operatorname {D e f o r m A t t n} (\\mathbf {e} _ {k - 1}, \\mathbf {M}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.716 + ], + "angle": 0, + "content": "where \\( k \\in \\{1, \\dots, K - 1\\} \\). We first add \\( \\mathbf{s}_e \\) with learnable queries \\( \\mathbf{q} \\in \\mathbb{R}^{N \\times C} \\) to represent candidate instances in the video. Then, we use deformable attention module [38] to reason the vision-conditioned multimodal relations where vision features serve as key and value to decompose linguistic features, as in Eq. (1). Finally, we stack the cross-attention module for \\( K \\) times, as in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Language-Conditioned Visual Decoder. In order to align visual object semantics to linguistic contents, we design a language-conditioned visual decoder \\(\\mathrm{Decoder}(\\mathrm{V}|\\mathrm{L})\\) (similar to Eqs. (1) and (2)), to enhance visual representation with the attendance of the language description. Differently, vision features are enhanced by multi-head self-attention (MHSA) modules at the first place, and then linguistic features \\(\\mathbf{s}_e\\) serve as key and value in cross-attention modules. In this case, it can reason the language-conditioned multimodal relations. Finally, we can get enhanced visual features by language-conditioned multimodal relations, as \\(\\mathbf{F} = \\mathrm{Decoder}(\\mathbf{M}|\\mathbf{s}_e)\\), where \\(\\mathbf{F} \\in \\mathbb{R}^{T \\times H \\times W \\times C}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13416" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.087, + 0.891, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.893, + 0.391 + ], + "angle": 0, + "content": "Figure 2: Our Hybrid Temporal-scale Multimodal Learning framework. It aligns linguistic and visual features by learning hierarchical multimodal interactions with hybrid temporal scales, detailed in Sec. 3.2.2. Moreover, a Cross-scale Multimodal Perception (CMP) module is designed to enable interaction and cooperation among temporal scales, detailed in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Hierarchical Multimodal Learning As single-scale multimodal learning is insufficient to understand the relations between videos and texts, we propose to construct the multimodal relations hierarchically for the \\(L\\) hybrid temporal scales with the assistance of \\(\\mathrm{Decoder}(\\mathrm{L}|\\mathrm{V})\\) and \\(\\mathrm{Decoder}(\\mathrm{V}|\\mathrm{L})\\). We can obtain linguistic-attended visual features and visual-attended linguistic embeddings for different temporal scales, capturing core object semantics conditioned on different visual and linguistic contexts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.383, + 0.57 + ], + "angle": 0, + "content": "3.3. Cross-scale Multimodal Perception" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.669 + ], + "angle": 0, + "content": "The multimodal relations are constructed conditioned on different modalities with hybrid temporal scales. However, the modeling process of different scales is independent. To promote the cooperation and align the visual and linguistic semantics both within the scale and across the scales, we design Cross-scale Multimodal Perception (CMP) module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.747 + ], + "angle": 0, + "content": "Intra-scale Perception. Despite sharing visual and linguistic feature extraction, the multimodal relation construction via Decoder(V|L) and Decoder(L|V) are independent to each other. To promote the cooperation between modalities, we propose an intra-scale perception module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.853 + ], + "angle": 0, + "content": "Specifically, in each temporal scale \\(l\\), we have visual attended linguistic embeddings \\(\\mathbf{e} \\in \\mathbb{R}^{N \\times T \\times D}\\) and linguistic attended visual features \\(\\mathbf{F} \\in \\mathbb{R}^{T \\times H \\times W \\times C}\\). To achieve fine-grained semantic alignment, we measure the similarity by dot product between \\(\\mathbf{e}\\) and \\(\\mathbf{F}\\) on pixel level. Specifically, we obtain the similarity map via multimodal perception module, denoted as \\(\\mathcal{I} = \\mathrm{MP}(\\mathbf{F}, \\mathbf{e})\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.862, + 0.469, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\Omega = \\operatorname {M a s k H e a d} (\\mathbf {e}), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.882, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} = \\Omega \\cdot \\mathbf {F}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.531 + ], + "angle": 0, + "content": "where MaskHead denotes three consecutive MLP layers for embedding conversion. Each value of \\(\\mathcal{I}\\) represents the relevance between visual-attended linguistic embeddings and linguistic-attended visual features, which can be interpreted as the existence of the referred object. As such, \\(\\mathcal{I}\\) is regarded as the object mask prediction with the context of current temporal scale. To this end, we achieve the multimodal perception in the same temporal scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Inter-scale Perception. The multimodal relations are constructed in different temporal scales. However, the process in each scale is independent and biased towards the contained object semantics. To alleviate it, we propose to align multimodal features from different temporal scales with a concise inter-scale perception module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Specifically, suppose that the referred object appears in frame \\(t\\) in temporal scales \\(l\\) and \\(l + 1\\), the referred object can be segmented by measuring similarity between \\((\\mathbf{F}^l (t),\\mathbf{e}^l (t))\\) and \\((\\mathbf{F}^{l + 1}(t),\\mathbf{e}^{l + 1}(t))\\) simultaneously. Conditioned on same frame \\(t\\), the visual-attended linguistic embedding \\(\\mathbf{e}^l (t)\\) from scale \\(l\\) is supposed to be relevant to the linguistic-attended visual features \\(\\mathbf{F}^{l + 1}(t)\\) from scale \\(l + 1\\). Thus, similar to Eq. (4), the similarity cross different temporal scales can be measured by" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.768, + 0.891, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} ^ {l \\rightarrow l + 1} (t) = \\mathrm {M P} (\\mathbf {e} ^ {l} (t), \\mathbf {F} ^ {l + 1} (t)). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Without losing generality, the inter-scale similarity can also be measured by \\(\\mathcal{I}^{l + 1\\to l}(t)\\). More specifically, frame \\(t\\) can be any frame shared by the adjacent temporal scales, which is ensured by our hybrid temporal scales sampling strategy. Each value in \\(\\mathcal{I}^{l\\rightarrow l + 1}(t)\\) represents the referred object prediction with the prior of linguistic object semantic from temporal scale \\(l\\). In the same way, values in \\(\\mathcal{I}^{l + 1\\rightarrow l}(t)\\) rep" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13417" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "resent the opposite. To this end, we achieve multimodal perception across different temporal scales." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.13, + 0.262, + 0.147 + ], + "angle": 0, + "content": "3.4. Training objectives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.468, + 0.209 + ], + "angle": 0, + "content": "Our network can be trained in an end-to-end manner to locate and segment the target instance simultaneously. Specifically, the losses for intra-scale and inter-scale multimodal perception in temporal scale \\( l \\) are formed as" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.217, + 0.468, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {i n t r a}} ^ {l} = \\sum \\mathcal {L} _ {\\text {c l s}} (\\mathbf {y}, \\hat {\\mathbf {y}}) + \\mathcal {L} _ {\\text {b o x}} (\\mathbf {b}, \\hat {\\mathbf {b}}) + \\mathcal {L} _ {\\text {m a s k}} (\\mathcal {I}, \\hat {\\mathcal {I}}), (6) \\\\ \\mathcal {L} _ {\\text {i n t e r}} ^ {l} = \\sum \\mathcal {L} _ {\\text {m a s k}} \\left(\\mathcal {I} ^ {l + 1 \\rightarrow l}, \\hat {\\mathcal {I}}\\right) (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.468, + 0.432 + ], + "angle": 0, + "content": "where time and instance subscripts are omitted for simplicity, \\( \\mathbf{y} \\) and \\( \\mathbf{b} \\) denote binary classification for instance existence and bounding box prediction respectively. Here \\( \\mathcal{L}_{cls} \\) is the focal loss [13], \\( \\mathcal{L}_{box} \\) is the sum of L1 loss and GIoU loss [21], and \\( \\mathcal{L}_{mask} \\) is the combination of DICE loss [17] and binary mask focal loss. We optimize the network by first finding the best prediction as the positive sample, via minimizing the matching cost \\( \\mathcal{L}_{intra}^{l} \\) and \\( \\mathcal{L}_{inter}^{l} \\) in each temporal scale \\( l \\) respectively. Then, we average the matching losses from different temporal scales and perception modules, and minimize it for positive samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.444, + 0.21, + 0.461 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.469, + 0.276, + 0.484 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.468, + 0.627 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on four datasets: Ref-Youtube-VOS[22], Ref-DAVIS17[9], A2D-Sentences and JHMDB-Sentences[7], following the common practice[29]. Metrics. We follow the standard evaluation protocol [22, 28, 29] to adopt region similarity \\(\\mathcal{L}(\\%)\\), contour accuracy \\(\\mathcal{F}(\\%)\\) and mean \\(\\mathcal{L}\\&\\mathcal{F}\\) for Ref-Youtube-VOS and Ref-DAVIS17. For JHMDB-Sentences, we adopt mAP to evaluate the model. For A2D-Sentences, we use Precision@K, Overall IoU, Mean IoU and mAP for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.274, + 0.652 + ], + "angle": 0, + "content": "4.2. Implemented Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.825 + ], + "angle": 0, + "content": "We set the number of attention layers \\( K \\) to 4 and the hidden dimension \\( C \\) to 256. The number of learnable queries \\( N \\) is set to 5. The number of MaskHead output channels \\( D \\) is set to 8. During training, we first sample video clips by sliding windows and then generate \\( L = 3 \\) hybrid temporal scales with stride \\( h = 2 \\) for generalized multimodal representations and relations. We use same training recipes as in [29, 2]. All frames are downsampled by shorter side to 360 and limit the maximum size for the long side to 640. Our model is pretrained on image referring segmentation datasets [35, 35]. See supplementary materials for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "During inference, we report results with all input frames in single temporal scale for fair comparisons. On RefDAVIS17, we directly inference on models trained on RefYoutube-VOS. Similarly, we reports JHMDB-Senteces results directly on models trained with A2D-Sentences." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.091, + 0.883, + 0.453 + ], + "angle": 0, + "content": "
MethodBackboneRef-Youtube-VOS
J&FJF
CMSA [33]ResNet-5034.933.336.5
CMSA + RNN [33]ResNet-5036.434.838.1
URVOS [22]ResNet-5047.245.349.2
LBDT-4 [22]ResNet-5047.245.349.2
MLRL [28]ResNet-5049.748.451.0
ReferFormer [29]ResNet-5055.654.856.5
OursResNet-5057.856.559.0
PMINet [6]ResNeSt-10148.246.749.6
PMINet + CFBI [6]ResNeSt-10153.051.554.5
CITD [12]ResNet-10156.454.858.1
ReferFormer [29]ResNet-10157.356.158.4
OursResNet-10158.557.359.8
PMINet + CFBI [6]Ensemble54.253.055.5
CITD [12]Ensemble61.460.062.7
ReferFormer [29]Swin-L62.460.864.0
OursSwin-L63.461.565.3
MTTR [2]Video-Swin-T55.354.056.6
ReferFormer [29]Video-Swin-T59.458.060.9
OursVideo-Swin-T61.259.563.0
ReferFormer [29]Video-Swin-S60.158.661.6
OursVideo-Swin-S61.459.962.9
ReferFormer [29]Video-Swin-B62.961.364.6
OursVideo-Swin-B63.461.565.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.463, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Table 1: Comparison with the SOTA methods on Ref-YTB-VOS." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.502, + 0.883, + 0.695 + ], + "angle": 0, + "content": "
MethodBackboneRef-DAVIS17
J&FJF
CMSA [33]ResNet-5034.732.237.2
CMSA + RNN [33]ResNet-5040.236.943.5
URVOS [22]ResNet-5051.547.356.0
LBDT-4 [5]ResNet-5054.5--
MLRL [28]ResNet-5058.053.962.0
ReferFormer [29]ResNet-5058.555.861.3
OursResNet-5059.556.662.4
ReferFormer [29]Swin-L60.557.663.4
OursSwin-L61.658.964.4
ReferFormer [29]Video-Swin-B61.158.164.1
OursVideo-Swin-B62.159.265.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Table 2: Comparison with the SOTA methods on Ref-DAVIS17." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.691, + 0.772 + ], + "angle": 0, + "content": "4.3. SOTA Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare our method with the state-of-the-art methods on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences. On Ref-Youtu-VOS, our approach achieves 58.5 in \\(\\mathcal{L}\\&\\mathcal{F}(\\%)\\) with ResNet-50, as shown in Tab. 1, which surpasses the recent SOTA method ReferFormer[29] with same backbone by 2.2 points. Moreover, it surpasses all the other SOTA methods with larger ResNet-101 on all evaluation metrics, which fully sug" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13418" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.089, + 0.881, + 0.296 + ], + "angle": 0, + "content": "
MethodBackbonePrecisionIoUmAP
P@0.5P@0.6P@0.7P@0.8P@0.9OverallMean
Hu et al. [8]VGG-1634.823.613.33.30.147.435.013.2
Gavrilyuk et al. [7]I3D47.534.721.18.00.253.642.119.8
CMSA + CFSA [34]ResNet-10148.743.135.823.15.261.843.2-
ACAN [26]I3D55.745.931.916.02.060.149.027.4
CMPC-V [14]I3D65.559.250.634.29.865.357.340.4
ClawCraneNet [11]ResNet-50/10170.467.761.748.917.163.159.9-
MTTR (ω = 8) [2]Video-Swin-T72.168.460.745.616.470.261.844.7
MTTR (ω = 10) [2]Video-Swin-T75.471.263.848.516.972.064.046.1
ReferFormer [29]Video-Swin-T82.879.272.355.319.377.669.652.8
ReferFormer [29]Video-Swin-B83.180.474.157.921.278.670.355.0
OursVideo-Swin-T82.279.272.355.320.177.669.253.4
OursVideo-Swin-B84.081.575.859.222.879.571.256.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.304, + 0.731, + 0.319 + ], + "angle": 0, + "content": "Table 3: Comparison with the state-of-the-art methods on A2D-Sentences." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.336, + 0.446, + 0.514 + ], + "angle": 0, + "content": "
MethodBackbonemAP
Hu et al. [8]VGG-1617.8
Gavrilyuk et al. [7]I3D23.3
ACAN [26]I3D28.9
CMPC-V [14]I3D34.2
MTTR (ω = 8) [2]Video-Swin-T36.6
MTTR (ω = 10) [2]Video-Swin-T39.2
ReferFormer† (ω = 6) [29]Video-Swin-T39.1
ReferFormer [29]Video-Swin-T42.2
ReferFormer [29]Video-Swin-B43.7
OursVideo-Swin-T42.7
OursVideo-Swin-B44.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.523, + 0.465, + 0.538 + ], + "angle": 0, + "content": "Table 4: SOTA results comparison on JHMDB-Sentences." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.555, + 0.439, + 0.664 + ], + "angle": 0, + "content": "
Components#FramesJ&FJF
i.Baseline555.654.856.5
ii.HTML w/o CMP556.054.757.4
iii.HTML556.355.057.5
iv.Baseline856.255.057.3
v.HTML w/o CMP857.156.058.2
vi.HTML857.856.559.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.673, + 0.465, + 0.688 + ], + "angle": 0, + "content": "Table 5: Ablation study on the components of our HTML." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.468, + 0.853 + ], + "angle": 0, + "content": "gests the superiority of our method. When equipped with larger backbone, our method still show considerable superiority with accuracy gap of 1.2 points for ResNet-101 and 1.0 points for Swin-L. We also experiment our method with the well-known Video Swin Transformers [16]. Our method with Video-Swin-Tiny backbone surpasses the SOTA method with the same backbone by 1.8 points. With larger Video-Swin Transformers (Small and Base models), our method still achieves SOTA performance, which shows the generality of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 2, our method surpasses the SOTA methods on Ref-DAVIS17 by over 1.0 points on both ResNet-50, Swin-L and Video-Swin-Base backbones, with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.49 + ], + "angle": 0, + "content": "new a SOTA record 62.1 in \\(\\mathcal{L}\\&\\mathcal{F}(\\%)\\). We also experiment our method on A2D-Sentences and JHMDB-Sentences datasets and compare with other SOTA results as shown in Tab. 3 and Tab. 4. Our method achieves SOTA performances with new records on both the two datasets. On A2D-Sentences, our HTML surpass SOTA result by 1.7 points in mAP and higher recall by 1.6 on Precision@0.9. On JHMDB-Sentences, our method still achieves a new SOTA record with 44.2 in mAP. These results demonstrate the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.655, + 0.516 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.523, + 0.892, + 0.552 + ], + "angle": 0, + "content": "In this section, we ablate core components of our HTML with Ref-Youtu-VOS based on ResNet-50." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Effectiveness of our HTML. To validate the effectiveness of our Hybrid Temporal-scale Multimodal Learning framework, we investigate each of our components by gradually adding them to the baseline [29]. First, comparing (i)&(iii) in Tab. 5, our HTML improves the performance of baseline to 56.3 when 5 frames are used for training, which is better than the baseline trained with longer input frames. Further, when longer temporal input is available, comparing (iv)&(vi), our HTML improves the model by 1.6 to 57.8 in \\(\\mathcal{L}\\&\\mathcal{F}\\). These prove the effectiveness of our method with frames in different lengths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Second, comparing counterpart networks using different number of frames, i.e. (iii) and (vi), our method can benefit from longer temporal input (8 frames vs. 5 frames) with an improvement of 1.5 points. This conclusion holds among the other counterpart network pair, i.e. (ii)&(v). Further, each of our components brings an improvement of 0.3 when trained with total 5 frames, while the improvement can be doubled to 0.7-0.9 with total 8 frames. This indicates that our method can better utilize the long temporal input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Finally, taking 8 frames for instance, hierarchical multimodal learning and cross-scale multimodal perception improves \\(\\mathcal{L}\\&\\mathcal{F}\\) by 0.9 and 0.7 respectively. It proves the ef" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13419" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.331, + 0.168 + ], + "angle": 0, + "content": "
#ScalesJ&FJF
156.255.057.3
256.955.858.1
357.856.559.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.175, + 0.314, + 0.19 + ], + "angle": 0, + "content": "(a) Effect of No. of temporal scales." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.092, + 0.591, + 0.165 + ], + "angle": 0, + "content": "
FramesJ&FJF
556.355.057.5
857.856.559.0
1257.556.458.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.362, + 0.175, + 0.56, + 0.19 + ], + "angle": 0, + "content": "(b) Effect of No. of input frames." + }, + { + "type": "table", + "bbox": [ + 0.626, + 0.089, + 0.864, + 0.168 + ], + "angle": 0, + "content": "
Inter-scaleJ&FJF
None56.255.057.3
l → l + 156.955.858.0
l + 1 → l57.856.559.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.641, + 0.175, + 0.825, + 0.19 + ], + "angle": 0, + "content": "(c) Effect of direction of CMP." + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.2, + 0.855, + 0.216 + ], + "angle": 0, + "content": "Table 6: Ablation study on Hybrid Temporal Scales in (a) and (b), and Cross-scale Multimodal Perception in (c)." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.235, + 0.451, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.486, + 0.47, + 0.517 + ], + "angle": 0, + "content": "Figure 3: Performance comparison of different query sets with different temporal scales on Ref-Youtu-VOS." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.535, + 0.381, + 0.551 + ], + "angle": 0, + "content": "fectiveness of each component of our HTML." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Length of language descriptions. We explore the performance of our method with different sampled sets of language descriptions, to validate the ability of capturing diversified linguistic object semantics. Since the train set of Ref-Youtu-VOS contains an average of two language descriptions for each object, we sample the shorter ones to form the Short set and the others to form the Long set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As shown in bottom line of Fig. 3, the Long set has longer sentences than the Short set in the average number of words. As shown in Fig. 3, when the objects are described by Short query set, more input frames in single temporal scale achieve inferior performance (5 frames vs. 8 frames: 55.6 vs. 55.1). It's interesting to note that more visual content (8 frames) fails to improve the performance when the linguistic content is insufficient (shorter queries). When the complexity of queries increases, i.e. with the Long query set, the quantitative relation is reversed: 8 frames guided model obtains better performance than the model with 5 frames, increasing by 0.8 (8 frames-Short query vs. 8 frames-Long query: 55.1 vs. 55.9). We infer that this is caused by the mismatch of visual and linguistic object semantics. On one side, shorter quires, i.e. relatively simple linguistic semantics, are insufficient to interpret longer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.265 + ], + "angle": 0, + "content": "videos. On the other side, the longer language descriptions contain more content irrelevant to the visual input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.893, + 0.401 + ], + "angle": 0, + "content": "Differently, our models trained with either the Short set or the Long set all surpass the single temporal scale guided model. Impressively, when the objects are described by queries flexible in lengths, i.e. with the All set, our method gets a performance boost by 1.6, while single temporal scale baseline (both 8 frames and 5 frames guided networks) improves by 0.3. This shows that our method has the strong ability to solve the mismatch issue between visual content and diversified linguistic contents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.568 + ], + "angle": 0, + "content": "Number of hybrid temporal scales. We investigate the effectiveness of our hierarchical multimodal learning, by exploring the number of hybrid temporal scales. For fair comparison among different settings, the number of total input frames is set to 8 in this subsection. As shown in Tab. 6 (a), when two temporal scales are constructed, our method brings an improvement of 0.6 in \\(\\mathcal{L}\\&\\mathcal{F}\\). When the number of temporal scales is increased to three, continuous improvement of 0.9 (56.9 vs. 57.8) is observed. It proves that our model benefits from the increasing visual diversity constructed by hybrid temporal scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.893, + 0.734 + ], + "angle": 0, + "content": "Number of input frames. We explore the effect of total input frames here. In this subsection, hybrid temporal scales are constructed by default following Sec. 3.2.2. As shown in Tab. 6 (b), more input frames bring an improvement by 1.5 in \\(\\mathcal{L}\\&\\mathcal{F}\\) (5 frames vs. 8frames: 56.3 vs. 57.8). When the input frames increase continuously to 12 frames, the performance saturates and drops slightly to 57.5. We conjecture that it is caused by insufficient video-language pairs (8 frames vs. 12 frames: \\(49\\mathrm{k}\\) vs. \\(32\\mathrm{k}\\)) compared to largely increased computation complexity (8 frames vs. 12 frames: 21.5 GFLOPs vs. 33.6 GFLOPs for transformers)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Direction of inter-scale perception. We explore the effect of direction of inter-scale multimodal perception in CMP. Comparing first two lines of Tab. 6 (c), \\( l \\rightarrow l + 1 \\) perception improves the baseline by 0.7 points; As in first and last lines, \\( l + 1 \\rightarrow l \\) perception improves the baseline by 1.6 points. These prove the effectiveness of our inter-scale perception. We choose the latter one for better performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.644, + 0.862 + ], + "angle": 0, + "content": "4.5. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We visualize the segmentation results of complex and simple language descriptions in Fig. 4. Specifically, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.847, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.513, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Figure 4: Visualization results of complex and simple language descriptions on Ref-Youtu-VOS. Red masks indicate positive segmentation results and blue masks indicate the negatives. Our HTML is able to clarify such object confusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.473, + 0.886 + ], + "angle": 0, + "content": "compare three settings, i.e., baseline with only single temporal scale, HTML w/o CMP which constructs hybrid temporal scales, and our final HTML which dynamically construct multimodal relations cross temporal scales. As expected, when only single temporal scale is adopted, baseline fails to segment the target in the video, e.g., the background rabbit in the subplot (a) is mistakenly referred. The main reason is that the rabbit shares similar appearance to the target object and also locates to the left of a person in last three frames. The model is misled and confused by the single temporal scale. When applied with hybrid temporal scales, the language description can interact with both long and short temporal scales. Thus, the previous false prediction in the first frame is corrected. Further, when applied with CMP, our model is able to clarify the object confusion by discovering the core semantics on a scale and make correct predictions. Similarly, in subplot (b), baseline is misled by the video clip where tennis ball only appears in the middle two frames. When gradually applying our proposed modules, the mistakenly predicted dog is clarified and further the target tennis ball is correctly segmented." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.566, + 0.62, + 0.581 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.758 + ], + "angle": 0, + "content": "In this work, we develop a HTML framework to align linguistic and visual features by learning multimodal relations hierarchically in different temporal scales. Moreover, we introduce an inter-scale multimodal perception module to construct dynamic multimodal interactions across temporal scales. We conduct experiments on four datasets and establish new state-of-the-art results. Particularly, our method with ResNet-50 backbone surpasses the recent methods with ResNet-100. The comprehensive ablation experiments and visualization results show that our method is able to discover core object semantics in the different modalities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.661, + 0.788 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (No.2022ZD0160100, No. 2022ZD0160505), the National Natural Science Foundation of China under Grant (62272450), the Joint Lab of CAS-HK, the Shenzhen Research Program (RCJC20200 714114557087), and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020355)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "13421" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. arXiv preprint arXiv:2111.14821, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.322, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[4] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.394, + 0.47, + 0.463 + ], + "angle": 0, + "content": "[5] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.539, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[7] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.598, + 0.47, + 0.651 + ], + "angle": 0, + "content": "[8] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.656, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[9] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.471, + 0.784 + ], + "angle": 0, + "content": "[10] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18706-18716, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[11] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[14] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[15] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[17] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.436, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[19] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[20] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[21] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[22] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[23] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[24] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmenta" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13422" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[26] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[27] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[28] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[29] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[30] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.518 + ], + "angle": 0, + "content": "[31] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.575 + ], + "angle": 0, + "content": "[32] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In European Conference on Computer Vision, pages 332-348. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[33] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10502–10511, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[34] Linwei Ye, Mrigank Rochan, Zhi Liu, Xiaoqin Zhang, and Yang Wang. Referring segmentation in images and videos with cross-modal self-attention network. arXiv preprint arXiv:2102.04762, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[35] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[36] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2246-2256, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[37] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13423" + } + ] +] \ No newline at end of file diff --git a/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_origin.pdf b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c0d96a67c31fb9631c77f90eea5b8b68e45aaa5a --- /dev/null +++ b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/17b3ea75-567e-4430-9688-bf87a6d34fd8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eee86b999eb6b57ca0877286e8586394e7411dc692297f2d9ca97d578149627 +size 3148444 diff --git a/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/full.md b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..961bad6fc848afea700425dd9f0701579c72a4ec --- /dev/null +++ b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/full.md @@ -0,0 +1,269 @@ +# HTML: Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation + +Mingfei Han $^{1,4}$ , Yali Wang $^{\dagger,2,6}$ , Zhihui Li $^{3}$ , Lina Yao $^{4}$ , Xiaojun Chang $^{1,5}$ , Yu Qiao $^{6,2}$ + +$^{1}$ ReLER, AAII, UTS $^{2}$ Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, SIAT-SenseTime Joint Lab, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences $^{3}$ Shandong Artificial Intelligence, Qilu University of Technology $^{4}$ Data61, CSIRO $^{5}$ Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence $^{6}$ Shanghai AI Laboratory, Shanghai, China https://mingfei.info/HTML/ + +# Abstract + +Referring Video Object Segmentation (RVOS) is to segment the object instance from a given video, according to the textual description of this object. However, in the open world, the object descriptions are often diversified in contents and flexible in lengths. This leads to the key difficulty in RVOS, i.e., various descriptions of different objects are corresponding to different temporal scales in the video, which is ignored by most existing approaches with single stride of frame sampling. To tackle this problem, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) framework, which can effectively align linguistic and visual features to discover core object semantics in the video, by learning multimodal interaction hierarchically from different temporal scales. More specifically, we introduce a novel inter-scale multimodal perception module, where the language queries dynamically interact with visual features across temporal scales. It can effectively reduce complex object confusion by passing video context among different scales. Finally, we conduct extensive experiments on the widely used benchmarks, including Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences, where our HTML achieves state-of-the-art performance on all these datasets. + +# 1. Introduction + +Referring Video Object Segmentation (RVOS) has witnessed the growing interest, due to its wide applications in visual editing, virtual reality, human-robotic interaction and so on. Different from the traditional vision-only VOS, RVOS aims to segment the object instance from an input video, according to an open-world description about the re + +![](images/8785cf05a20cb459bacb962e746ece8fe3f5be378d46a37bb55b96cbe53b8a6e.jpg) +Figure 1: Referring descriptions in different lengths. (a) The description is simple containing only the category name. (b) The description is complicated with movement and position of the object. Single-scale baseline (e.g., four frames in (a) and two frames in (b)) fails to segment the referred object, while our hybrid-scale HTML succeeds. More discussion can be found in introduction. + +ferred object. In this case, the model has to learn both visual and textual contents comprehensively, in order to discover the underlying object by multimodal interaction. + +Recent studies [4, 12, 28, 29] have shown that, cross-modal attention is an effective way to bridge the gap between vision and language in RVOS. However, these approaches perform vision-language interactions with video frames sampled from a single temporal scale, which may limit their power to infer the referred object with accurate segmentation. The main reason is that, the open-world descriptions vary in length and contain rich semantics about the referred object, e.g., where it is, how it moves, which objects it interact with. Apparently, such diversified texts are corresponding to various temporal-scale snippets. + +For example, the language query in Fig. 1 (a) is a tennis ball. Such a short description is corresponding to the ball located at a small region in the middle two frames. If the single-scale baseline samples four frames as input, it will fail to segment the referred object. This is because it overlooks the dog in the center place among all these four frames, while lacking the detailed understanding in the middle two frames. Alternatively, the language query in Fig. 1 (b) is a sheep top second right moves down and comes out of the circle. Such a long description is corresponding to the particular sheep in the group, which moves across frames. If the single-scale baseline samples two frames as input, it will fail to segment the referred object. This is because it is misled by the subtle movement of sheep group in only two frames, without understanding how each sheep moves from the adjacent frames. + +To tackle this difficulty, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) Framework for RVOS, which can alleviate object confusion by language-vision interactions across different temporal scales. Specifically, we sample video frames according to different temporal scales. For each temporal scale, we introduce an intra-scale multimodal perception module, which can effectively exploit core visual semantics within the frames at this temporal scale, by mutual enhancement between textual and visual embeddings. Then, we design an inter-scale multimodal perception module, where linguistic embeddings dynamically interact with visual features across temporal scales. In this case, we can hierarchically leverage object context from all the temporal scales to boost RVOS. Finally, we evaluate our HTML on a number of benchmarks, including Ref-Youtu-VOS [22], Ref-DAVIS17 [9], A2D-Sentences and JHMDB-Sentences [7]. The extensive experiments have shown that, our HTML achieves the state-of-the-art performance on all of them. + +Overall we make three contributions in this paper: + +- Concise and unified learning framework: our Hybrid Temporal-scale Multimodal Learning (HTML) framework hierarchically constructs multimodal interactions via different strides of frame sampling, which can mutually enhance embeddings from both modalities for accurate segmentation. + +- Effective multimodal perception module: our Cross-scale Multimodal Perception (CMP) module can effectively reduce complex object confusions with intra-scale and inter-scale multimodal perceptions, where linguistic and visual features interact across temporal scales. +- State-of-the-art performance on the widely-used benchmarks, which shows the superiority of our framework. Specifically, on Ref-Youtu-VOS [22], our method with ResNet-50 achieves 57.8 in $\mathcal{L}\&\mathcal{F}$ , outperforming the recent SOTA method [29] with ResNet-101. + +# 2. Related works + +Vision-only Video Segmentation. Vision-only video segmentation tasks [20, 19, 31, 1, 30, 27, 10, 32, 18, 25, 23, 36], need to segment the objects with predefined semantic query set. Video instance segmentation (VIS) [31] needs to segment the different instances in the desired category set and track each instance with its identity kept. Video object segmentation (VOS) [20, 19], the model needs to separate an object from the background in the video, given the mask of the first mask. In the early stage of VIS, heavy supervision and complex heuristic rules are applied to associate the instances across frames by separate segmentation and association branches, such as MaskTrackRCNN[31] and MaskProp[1]. Recent progress in transformer[24] and DETR[3] has inspired the community with end-to-end works, such as SeqFormer[30] and VisTR[27]. As for VOS, most recent works [32, 18, 25, 23, 36] on semi-supervised VOS enforces temporal consistency in the video to propagate the first-frame mask to the other frames sequentially. Most of them lie in the group of matching-based methods, updating the memories and matching to segment. + +Referring Video Object Segmentation. RVOS task [22] is tasked to segment the referred object from the given video with specified open-world language description. Most early methods in RVOS proposed to refer the object by applying image-level methods on video frames separately and associate them with heuristic rules. However, they usually fail to utilize the temporal dynamic. [22] casts the task as a joint problem of referring segmentation in frame and mask propagation across frames by a memory attention module. [12] proposed a top-down pipeline by constructing exhaustive set of object tracklets and then selecting the target by matching the language features with the all the candidate tracklets. [37] proposed to model the temporal dynamic with an additional optical flow modality. [28] argued the importance of the structural information of video content and proposed to utilize the frame, object and video features simultaneously to obtain better representation. More recently, [2] introduced the DETR structure to RVOS area and [29] proposed to use language-conditional queries to simplify the referring pipeline and improve the performance, which serves as our baseline. + +Different from the previous works, we raise the mismatch issue that the various descriptions of different objects are corresponding to different temporal scales of the video. Moreover, we propose a concise HTML framework via multimodal interaction across different temporal scales to capture the core object semantics in the video. + +# 3. Method + +To effectively align diversified descriptions and complex videos, we propose a distinct Hybrid Temporal-scale Multimodal Learning (HTML) framework for RVOS. In this section, we introduce our HTML in detail. First, we deliver an overview of HTML framework. Then, we explain how to build the hybrid temporal-scale multimodal learning paths, in the aid of vision-conditioned linguistic decoder and language-conditioned visual decoder. Next, we introduce a Cross-scale Multimodal Perception (CMP) module to align multimodal features across temporal scales. Finally, we describe the training objectives to optimize our HTML. + +# 3.1. Framework Overview + +As shown in Fig. 2, our HTML framework consists of three main parts. First, we need to extract visual and linguistic features from backbones. We adopt a visual backbone to extract frame features from $T$ frames sampled from the given video. It can be either 2D CNN networks or 3D transformer networks. We then feed the extracted vision features into a deformable transformer encoder [38] to construct spatiotemporal relations between different frames. Meanwhile, to make a fair comparison with previous works in RVOS [29], we utilize the pretrained linguistic embedding model, RoBERTa [15], to extract textual features $\mathbf{s}_e\in$ $\mathbb{R}^{L\times C}$ from language descriptions with $L$ words. More details can be found in Sec. 4.2. + +After extracting visual and linguistic features, we next construct multimodal interactions between the language descriptions and the videos. Different from the previous approaches [29, 28], we build $L$ multimodal learning paths, where the linguistic embedding hierarchically interacts with visual features in different temporal scales. Then, we incorporate the mutually enhanced visual and linguistic features by a novel Cross-scale Multimodal Perception (CMP) module to align multimodal features across different scales. Finally, we design the training losses. + +# 3.2. Hybrid Temporal-scale Multimodal Learning + +To capture core object semantics, we propose a novel hybrid temporal-scale multimodal learning framework to learn multimodal relations. To start with, we build hybrid temporal scales via different sampling strides. Then, we construct basic multimodal relation learning units. Finally, we explain how to construct hybrid learning paths. + +# 3.2.1 Hybrid Temporal Scale Construction + +Since the texts of various objects may refer to different video parts, single temporal scale often fails to describe the diversified textual contents. To simulate such diversity and flexibility, we build hybrid temporal scales by periodically sampling frames with different strides. + +We first regard all the input frames as the first temporal scale, and then build other $L - 1$ temporal scales upon it in a sequential manner. In order to ensure the diversity of sampled temporal scale, we randomly pick one frame from every $h$ frames of last scale, where $h$ denotes the predefined stride. Subsequently, we feed the sampled frames into visual encoder Encoder(V) to extract feature maps for each of the scales respectively. Specifically, for temporal scale $l$ , we can obtain $\mathbf{M} \in \mathbb{R}^{T \times H \times W \times C}$ , where $T$ denotes the number of frames in the temporal scale. + +# 3.2.2 Multimodal Relation Learning + +In order to discover the core object semantics, we construct multimodal relations via vision-conditioned linguistic decoder and language-conditioned visual decoder to align semantics between different modalities. + +Vision-Conditioned Linguistic Decoder. In order to align linguistic object semantics to the vision contents, we design a vision-conditioned linguistic decoder $\mathrm{Decoder}(\mathrm{L}|\mathrm{V})$ . Specifically, we have visual features $\mathbf{M}\in \mathbb{R}^{T\times H\times W\times C}$ , and linguistic embeddings $\mathbf{s}_e\in \mathbb{R}^{1\times C}$ . The vision-conditioned multimodal relations $\mathbf{e} = \mathrm{Decoder}(\mathbf{s}_e|\mathbf{M})$ are constructed by + +$$ +\mathbf {e} _ {0} = \operatorname {D e f o r m A t t n} (\mathbf {s} _ {e} + \mathbf {q}, \mathbf {M}), \tag {1} +$$ + +$$ +\mathbf {e} _ {k} = \operatorname {D e f o r m A t t n} (\mathbf {e} _ {k - 1}, \mathbf {M}), \tag {2} +$$ + +where $k \in \{1, \dots, K - 1\}$ . We first add $\mathbf{s}_e$ with learnable queries $\mathbf{q} \in \mathbb{R}^{N \times C}$ to represent candidate instances in the video. Then, we use deformable attention module [38] to reason the vision-conditioned multimodal relations where vision features serve as key and value to decompose linguistic features, as in Eq. (1). Finally, we stack the cross-attention module for $K$ times, as in Eq. (2). + +Language-Conditioned Visual Decoder. In order to align visual object semantics to linguistic contents, we design a language-conditioned visual decoder $\mathrm{Decoder}(\mathrm{V}|\mathrm{L})$ (similar to Eqs. (1) and (2)), to enhance visual representation with the attendance of the language description. Differently, vision features are enhanced by multi-head self-attention (MHSA) modules at the first place, and then linguistic features $\mathbf{s}_e$ serve as key and value in cross-attention modules. In this case, it can reason the language-conditioned multimodal relations. Finally, we can get enhanced visual features by language-conditioned multimodal relations, as $\mathbf{F} = \mathrm{Decoder}(\mathbf{M}|\mathbf{s}_e)$ , where $\mathbf{F} \in \mathbb{R}^{T \times H \times W \times C}$ . + +![](images/b6368af0dd18156247a7285b9233ada7ff5ceeb39c76b857124cf3ac90575304.jpg) +Figure 2: Our Hybrid Temporal-scale Multimodal Learning framework. It aligns linguistic and visual features by learning hierarchical multimodal interactions with hybrid temporal scales, detailed in Sec. 3.2.2. Moreover, a Cross-scale Multimodal Perception (CMP) module is designed to enable interaction and cooperation among temporal scales, detailed in Sec. 3.3. + +Hierarchical Multimodal Learning As single-scale multimodal learning is insufficient to understand the relations between videos and texts, we propose to construct the multimodal relations hierarchically for the $L$ hybrid temporal scales with the assistance of $\mathrm{Decoder}(\mathrm{L}|\mathrm{V})$ and $\mathrm{Decoder}(\mathrm{V}|\mathrm{L})$ . We can obtain linguistic-attended visual features and visual-attended linguistic embeddings for different temporal scales, capturing core object semantics conditioned on different visual and linguistic contexts. + +# 3.3. Cross-scale Multimodal Perception + +The multimodal relations are constructed conditioned on different modalities with hybrid temporal scales. However, the modeling process of different scales is independent. To promote the cooperation and align the visual and linguistic semantics both within the scale and across the scales, we design Cross-scale Multimodal Perception (CMP) module. + +Intra-scale Perception. Despite sharing visual and linguistic feature extraction, the multimodal relation construction via Decoder(V|L) and Decoder(L|V) are independent to each other. To promote the cooperation between modalities, we propose an intra-scale perception module. + +Specifically, in each temporal scale $l$ , we have visual attended linguistic embeddings $\mathbf{e} \in \mathbb{R}^{N \times T \times D}$ and linguistic attended visual features $\mathbf{F} \in \mathbb{R}^{T \times H \times W \times C}$ . To achieve fine-grained semantic alignment, we measure the similarity by dot product between $\mathbf{e}$ and $\mathbf{F}$ on pixel level. Specifically, we obtain the similarity map via multimodal perception module, denoted as $\mathcal{I} = \mathrm{MP}(\mathbf{F}, \mathbf{e})$ by + +$$ +\Omega = \operatorname {M a s k H e a d} (\mathbf {e}), \tag {3} +$$ + +$$ +\mathcal {I} = \Omega \cdot \mathbf {F}, \tag {4} +$$ + +where MaskHead denotes three consecutive MLP layers for embedding conversion. Each value of $\mathcal{I}$ represents the relevance between visual-attended linguistic embeddings and linguistic-attended visual features, which can be interpreted as the existence of the referred object. As such, $\mathcal{I}$ is regarded as the object mask prediction with the context of current temporal scale. To this end, we achieve the multimodal perception in the same temporal scale. + +Inter-scale Perception. The multimodal relations are constructed in different temporal scales. However, the process in each scale is independent and biased towards the contained object semantics. To alleviate it, we propose to align multimodal features from different temporal scales with a concise inter-scale perception module. + +Specifically, suppose that the referred object appears in frame $t$ in temporal scales $l$ and $l + 1$ , the referred object can be segmented by measuring similarity between $(\mathbf{F}^l (t),\mathbf{e}^l (t))$ and $(\mathbf{F}^{l + 1}(t),\mathbf{e}^{l + 1}(t))$ simultaneously. Conditioned on same frame $t$ , the visual-attended linguistic embedding $\mathbf{e}^l (t)$ from scale $l$ is supposed to be relevant to the linguistic-attended visual features $\mathbf{F}^{l + 1}(t)$ from scale $l + 1$ . Thus, similar to Eq. (4), the similarity cross different temporal scales can be measured by + +$$ +\mathcal {I} ^ {l \rightarrow l + 1} (t) = \mathrm {M P} (\mathbf {e} ^ {l} (t), \mathbf {F} ^ {l + 1} (t)). \tag {5} +$$ + +Without losing generality, the inter-scale similarity can also be measured by $\mathcal{I}^{l + 1\to l}(t)$ . More specifically, frame $t$ can be any frame shared by the adjacent temporal scales, which is ensured by our hybrid temporal scales sampling strategy. Each value in $\mathcal{I}^{l\rightarrow l + 1}(t)$ represents the referred object prediction with the prior of linguistic object semantic from temporal scale $l$ . In the same way, values in $\mathcal{I}^{l + 1\rightarrow l}(t)$ rep + +resent the opposite. To this end, we achieve multimodal perception across different temporal scales. + +# 3.4. Training objectives + +Our network can be trained in an end-to-end manner to locate and segment the target instance simultaneously. Specifically, the losses for intra-scale and inter-scale multimodal perception in temporal scale $l$ are formed as + +$$ +\begin{array}{l} \mathcal {L} _ {\text {i n t r a}} ^ {l} = \sum \mathcal {L} _ {\text {c l s}} (\mathbf {y}, \hat {\mathbf {y}}) + \mathcal {L} _ {\text {b o x}} (\mathbf {b}, \hat {\mathbf {b}}) + \mathcal {L} _ {\text {m a s k}} (\mathcal {I}, \hat {\mathcal {I}}), (6) \\ \mathcal {L} _ {\text {i n t e r}} ^ {l} = \sum \mathcal {L} _ {\text {m a s k}} \left(\mathcal {I} ^ {l + 1 \rightarrow l}, \hat {\mathcal {I}}\right) (7) \\ \end{array} +$$ + +where time and instance subscripts are omitted for simplicity, $\mathbf{y}$ and $\mathbf{b}$ denote binary classification for instance existence and bounding box prediction respectively. Here $\mathcal{L}_{cls}$ is the focal loss [13], $\mathcal{L}_{box}$ is the sum of L1 loss and GIoU loss [21], and $\mathcal{L}_{mask}$ is the combination of DICE loss [17] and binary mask focal loss. We optimize the network by first finding the best prediction as the positive sample, via minimizing the matching cost $\mathcal{L}_{intra}^{l}$ and $\mathcal{L}_{inter}^{l}$ in each temporal scale $l$ respectively. Then, we average the matching losses from different temporal scales and perception modules, and minimize it for positive samples. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +Datasets. We conduct experiments on four datasets: Ref-Youtube-VOS[22], Ref-DAVIS17[9], A2D-Sentences and JHMDB-Sentences[7], following the common practice[29]. Metrics. We follow the standard evaluation protocol [22, 28, 29] to adopt region similarity $\mathcal{L}(\%)$ , contour accuracy $\mathcal{F}(\%)$ and mean $\mathcal{L}\&\mathcal{F}$ for Ref-Youtube-VOS and Ref-DAVIS17. For JHMDB-Sentences, we adopt mAP to evaluate the model. For A2D-Sentences, we use Precision@K, Overall IoU, Mean IoU and mAP for evaluation. + +# 4.2. Implemented Details + +We set the number of attention layers $K$ to 4 and the hidden dimension $C$ to 256. The number of learnable queries $N$ is set to 5. The number of MaskHead output channels $D$ is set to 8. During training, we first sample video clips by sliding windows and then generate $L = 3$ hybrid temporal scales with stride $h = 2$ for generalized multimodal representations and relations. We use same training recipes as in [29, 2]. All frames are downsampled by shorter side to 360 and limit the maximum size for the long side to 640. Our model is pretrained on image referring segmentation datasets [35, 35]. See supplementary materials for details. + +During inference, we report results with all input frames in single temporal scale for fair comparisons. On RefDAVIS17, we directly inference on models trained on RefYoutube-VOS. Similarly, we reports JHMDB-Senteces results directly on models trained with A2D-Sentences. + +
MethodBackboneRef-Youtube-VOS
J&FJF
CMSA [33]ResNet-5034.933.336.5
CMSA + RNN [33]ResNet-5036.434.838.1
URVOS [22]ResNet-5047.245.349.2
LBDT-4 [22]ResNet-5047.245.349.2
MLRL [28]ResNet-5049.748.451.0
ReferFormer [29]ResNet-5055.654.856.5
OursResNet-5057.856.559.0
PMINet [6]ResNeSt-10148.246.749.6
PMINet + CFBI [6]ResNeSt-10153.051.554.5
CITD [12]ResNet-10156.454.858.1
ReferFormer [29]ResNet-10157.356.158.4
OursResNet-10158.557.359.8
PMINet + CFBI [6]Ensemble54.253.055.5
CITD [12]Ensemble61.460.062.7
ReferFormer [29]Swin-L62.460.864.0
OursSwin-L63.461.565.3
MTTR [2]Video-Swin-T55.354.056.6
ReferFormer [29]Video-Swin-T59.458.060.9
OursVideo-Swin-T61.259.563.0
ReferFormer [29]Video-Swin-S60.158.661.6
OursVideo-Swin-S61.459.962.9
ReferFormer [29]Video-Swin-B62.961.364.6
OursVideo-Swin-B63.461.565.2
+ +Table 1: Comparison with the SOTA methods on Ref-YTB-VOS. + +
MethodBackboneRef-DAVIS17
J&FJF
CMSA [33]ResNet-5034.732.237.2
CMSA + RNN [33]ResNet-5040.236.943.5
URVOS [22]ResNet-5051.547.356.0
LBDT-4 [5]ResNet-5054.5--
MLRL [28]ResNet-5058.053.962.0
ReferFormer [29]ResNet-5058.555.861.3
OursResNet-5059.556.662.4
ReferFormer [29]Swin-L60.557.663.4
OursSwin-L61.658.964.4
ReferFormer [29]Video-Swin-B61.158.164.1
OursVideo-Swin-B62.159.265.1
+ +Table 2: Comparison with the SOTA methods on Ref-DAVIS17. + +# 4.3. SOTA Comparisons + +We compare our method with the state-of-the-art methods on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences. On Ref-Youtu-VOS, our approach achieves 58.5 in $\mathcal{L}\&\mathcal{F}(\%)$ with ResNet-50, as shown in Tab. 1, which surpasses the recent SOTA method ReferFormer[29] with same backbone by 2.2 points. Moreover, it surpasses all the other SOTA methods with larger ResNet-101 on all evaluation metrics, which fully sug + +
MethodBackbonePrecisionIoUmAP
P@0.5P@0.6P@0.7P@0.8P@0.9OverallMean
Hu et al. [8]VGG-1634.823.613.33.30.147.435.013.2
Gavrilyuk et al. [7]I3D47.534.721.18.00.253.642.119.8
CMSA + CFSA [34]ResNet-10148.743.135.823.15.261.843.2-
ACAN [26]I3D55.745.931.916.02.060.149.027.4
CMPC-V [14]I3D65.559.250.634.29.865.357.340.4
ClawCraneNet [11]ResNet-50/10170.467.761.748.917.163.159.9-
MTTR (ω = 8) [2]Video-Swin-T72.168.460.745.616.470.261.844.7
MTTR (ω = 10) [2]Video-Swin-T75.471.263.848.516.972.064.046.1
ReferFormer [29]Video-Swin-T82.879.272.355.319.377.669.652.8
ReferFormer [29]Video-Swin-B83.180.474.157.921.278.670.355.0
OursVideo-Swin-T82.279.272.355.320.177.669.253.4
OursVideo-Swin-B84.081.575.859.222.879.571.256.7
+ +Table 3: Comparison with the state-of-the-art methods on A2D-Sentences. + +
MethodBackbonemAP
Hu et al. [8]VGG-1617.8
Gavrilyuk et al. [7]I3D23.3
ACAN [26]I3D28.9
CMPC-V [14]I3D34.2
MTTR (ω = 8) [2]Video-Swin-T36.6
MTTR (ω = 10) [2]Video-Swin-T39.2
ReferFormer† (ω = 6) [29]Video-Swin-T39.1
ReferFormer [29]Video-Swin-T42.2
ReferFormer [29]Video-Swin-B43.7
OursVideo-Swin-T42.7
OursVideo-Swin-B44.2
+ +Table 4: SOTA results comparison on JHMDB-Sentences. + +
Components#FramesJ&FJF
i.Baseline555.654.856.5
ii.HTML w/o CMP556.054.757.4
iii.HTML556.355.057.5
iv.Baseline856.255.057.3
v.HTML w/o CMP857.156.058.2
vi.HTML857.856.559.0
+ +Table 5: Ablation study on the components of our HTML. + +gests the superiority of our method. When equipped with larger backbone, our method still show considerable superiority with accuracy gap of 1.2 points for ResNet-101 and 1.0 points for Swin-L. We also experiment our method with the well-known Video Swin Transformers [16]. Our method with Video-Swin-Tiny backbone surpasses the SOTA method with the same backbone by 1.8 points. With larger Video-Swin Transformers (Small and Base models), our method still achieves SOTA performance, which shows the generality of our method. + +As shown in Tab. 2, our method surpasses the SOTA methods on Ref-DAVIS17 by over 1.0 points on both ResNet-50, Swin-L and Video-Swin-Base backbones, with + +new a SOTA record 62.1 in $\mathcal{L}\&\mathcal{F}(\%)$ . We also experiment our method on A2D-Sentences and JHMDB-Sentences datasets and compare with other SOTA results as shown in Tab. 3 and Tab. 4. Our method achieves SOTA performances with new records on both the two datasets. On A2D-Sentences, our HTML surpass SOTA result by 1.7 points in mAP and higher recall by 1.6 on Precision@0.9. On JHMDB-Sentences, our method still achieves a new SOTA record with 44.2 in mAP. These results demonstrate the superiority of our method. + +# 4.4. Ablation Study + +In this section, we ablate core components of our HTML with Ref-Youtu-VOS based on ResNet-50. + +Effectiveness of our HTML. To validate the effectiveness of our Hybrid Temporal-scale Multimodal Learning framework, we investigate each of our components by gradually adding them to the baseline [29]. First, comparing (i)&(iii) in Tab. 5, our HTML improves the performance of baseline to 56.3 when 5 frames are used for training, which is better than the baseline trained with longer input frames. Further, when longer temporal input is available, comparing (iv)&(vi), our HTML improves the model by 1.6 to 57.8 in $\mathcal{L}\&\mathcal{F}$ . These prove the effectiveness of our method with frames in different lengths. + +Second, comparing counterpart networks using different number of frames, i.e. (iii) and (vi), our method can benefit from longer temporal input (8 frames vs. 5 frames) with an improvement of 1.5 points. This conclusion holds among the other counterpart network pair, i.e. (ii)&(v). Further, each of our components brings an improvement of 0.3 when trained with total 5 frames, while the improvement can be doubled to 0.7-0.9 with total 8 frames. This indicates that our method can better utilize the long temporal input. + +Finally, taking 8 frames for instance, hierarchical multimodal learning and cross-scale multimodal perception improves $\mathcal{L}\&\mathcal{F}$ by 0.9 and 0.7 respectively. It proves the ef + +
#ScalesJ&FJF
156.255.057.3
256.955.858.1
357.856.559.0
+ +(a) Effect of No. of temporal scales. + +
FramesJ&FJF
556.355.057.5
857.856.559.0
1257.556.458.6
+ +(b) Effect of No. of input frames. + +
Inter-scaleJ&FJF
None56.255.057.3
l → l + 156.955.858.0
l + 1 → l57.856.559.0
+ +(c) Effect of direction of CMP. + +Table 6: Ablation study on Hybrid Temporal Scales in (a) and (b), and Cross-scale Multimodal Perception in (c). + +![](images/e96e42331b76dfa19e8048dac016a0ebf27b28ad5019ade6b42338b8e257168e.jpg) +Figure 3: Performance comparison of different query sets with different temporal scales on Ref-Youtu-VOS. + +fectiveness of each component of our HTML. + +Length of language descriptions. We explore the performance of our method with different sampled sets of language descriptions, to validate the ability of capturing diversified linguistic object semantics. Since the train set of Ref-Youtu-VOS contains an average of two language descriptions for each object, we sample the shorter ones to form the Short set and the others to form the Long set. + +As shown in bottom line of Fig. 3, the Long set has longer sentences than the Short set in the average number of words. As shown in Fig. 3, when the objects are described by Short query set, more input frames in single temporal scale achieve inferior performance (5 frames vs. 8 frames: 55.6 vs. 55.1). It's interesting to note that more visual content (8 frames) fails to improve the performance when the linguistic content is insufficient (shorter queries). When the complexity of queries increases, i.e. with the Long query set, the quantitative relation is reversed: 8 frames guided model obtains better performance than the model with 5 frames, increasing by 0.8 (8 frames-Short query vs. 8 frames-Long query: 55.1 vs. 55.9). We infer that this is caused by the mismatch of visual and linguistic object semantics. On one side, shorter quires, i.e. relatively simple linguistic semantics, are insufficient to interpret longer + +videos. On the other side, the longer language descriptions contain more content irrelevant to the visual input. + +Differently, our models trained with either the Short set or the Long set all surpass the single temporal scale guided model. Impressively, when the objects are described by queries flexible in lengths, i.e. with the All set, our method gets a performance boost by 1.6, while single temporal scale baseline (both 8 frames and 5 frames guided networks) improves by 0.3. This shows that our method has the strong ability to solve the mismatch issue between visual content and diversified linguistic contents. + +Number of hybrid temporal scales. We investigate the effectiveness of our hierarchical multimodal learning, by exploring the number of hybrid temporal scales. For fair comparison among different settings, the number of total input frames is set to 8 in this subsection. As shown in Tab. 6 (a), when two temporal scales are constructed, our method brings an improvement of 0.6 in $\mathcal{L}\&\mathcal{F}$ . When the number of temporal scales is increased to three, continuous improvement of 0.9 (56.9 vs. 57.8) is observed. It proves that our model benefits from the increasing visual diversity constructed by hybrid temporal scales. + +Number of input frames. We explore the effect of total input frames here. In this subsection, hybrid temporal scales are constructed by default following Sec. 3.2.2. As shown in Tab. 6 (b), more input frames bring an improvement by 1.5 in $\mathcal{L}\&\mathcal{F}$ (5 frames vs. 8frames: 56.3 vs. 57.8). When the input frames increase continuously to 12 frames, the performance saturates and drops slightly to 57.5. We conjecture that it is caused by insufficient video-language pairs (8 frames vs. 12 frames: $49\mathrm{k}$ vs. $32\mathrm{k}$ ) compared to largely increased computation complexity (8 frames vs. 12 frames: 21.5 GFLOPs vs. 33.6 GFLOPs for transformers). + +Direction of inter-scale perception. We explore the effect of direction of inter-scale multimodal perception in CMP. Comparing first two lines of Tab. 6 (c), $l \rightarrow l + 1$ perception improves the baseline by 0.7 points; As in first and last lines, $l + 1 \rightarrow l$ perception improves the baseline by 1.6 points. These prove the effectiveness of our inter-scale perception. We choose the latter one for better performance. + +# 4.5. Visualizations + +We visualize the segmentation results of complex and simple language descriptions in Fig. 4. Specifically, we + +![](images/1aa1a49843365ea85fd4757ba4fb67cfefb379837cf3e66a2d12f7c5c8bb6336.jpg) +Figure 4: Visualization results of complex and simple language descriptions on Ref-Youtu-VOS. Red masks indicate positive segmentation results and blue masks indicate the negatives. Our HTML is able to clarify such object confusion. + +compare three settings, i.e., baseline with only single temporal scale, HTML w/o CMP which constructs hybrid temporal scales, and our final HTML which dynamically construct multimodal relations cross temporal scales. As expected, when only single temporal scale is adopted, baseline fails to segment the target in the video, e.g., the background rabbit in the subplot (a) is mistakenly referred. The main reason is that the rabbit shares similar appearance to the target object and also locates to the left of a person in last three frames. The model is misled and confused by the single temporal scale. When applied with hybrid temporal scales, the language description can interact with both long and short temporal scales. Thus, the previous false prediction in the first frame is corrected. Further, when applied with CMP, our model is able to clarify the object confusion by discovering the core semantics on a scale and make correct predictions. Similarly, in subplot (b), baseline is misled by the video clip where tennis ball only appears in the middle two frames. When gradually applying our proposed modules, the mistakenly predicted dog is clarified and further the target tennis ball is correctly segmented. + +# 5. Conclusion + +In this work, we develop a HTML framework to align linguistic and visual features by learning multimodal relations hierarchically in different temporal scales. Moreover, we introduce an inter-scale multimodal perception module to construct dynamic multimodal interactions across temporal scales. We conduct experiments on four datasets and establish new state-of-the-art results. Particularly, our method with ResNet-50 backbone surpasses the recent methods with ResNet-100. The comprehensive ablation experiments and visualization results show that our method is able to discover core object semantics in the different modalities. + +# Acknowledgement + +This work was supported in part by the National Key R&D Program of China (No.2022ZD0160100, No. 2022ZD0160505), the National Natural Science Foundation of China under Grant (62272450), the Joint Lab of CAS-HK, the Shenzhen Research Program (RCJC20200 714114557087), and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020355) + +# References + +[1] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2020. 2 +[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. arXiv preprint arXiv:2111.14821, 2021. 2, 5, 6 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2 +[4] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2 +[5] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 5 +[6] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 5 +[7] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6 +[8] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6 +[9] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 2, 5 +[10] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18706-18716, 2023. 2 +[11] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 6 +[12] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2, 5 + +[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5 +[14] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6 +[15] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3 +[16] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021. 6 +[17] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5 +[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 2 +[19] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2 +[20] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2 +[21] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5 +[22] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 5 +[23] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 2 +[24] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 2 +[25] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmenta + +tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2 +[26] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6 +[27] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2 +[28] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 2, 3, 5 +[29] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 3, 5, 6 +[30] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2 +[31] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2 +[32] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In European Conference on Computer Vision, pages 332-348. Springer, 2020. 2 +[33] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10502–10511, 2019. 5 +[34] Linwei Ye, Mrigank Rochan, Zhi Liu, Xiaoqin Zhang, and Yang Wang. Referring segmentation in images and videos with cross-modal self-attention network. arXiv preprint arXiv:2102.04762, 2021. 6 +[35] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 5 +[36] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2246-2256, 2023. 2 +[37] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2 + +[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3 \ No newline at end of file diff --git a/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/images.zip b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1714ba6e5a05f5e72fb7f7353a778d00a55c33b2 --- /dev/null +++ b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e8d273ec1877140369d4ab823433b1574ef5e9da3038c063116b1d88a2d6d69 +size 723557 diff --git a/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/layout.json b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0650dd7007e5c6fd40a8223cbf06bd6e0183de22 --- /dev/null +++ b/2023/HTML_ Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation/layout.json @@ -0,0 +1,7446 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "text", + "content": "HTML: Hybrid Temporal-scale Multimodal Learning Framework for Referring Video Object Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "spans": [ + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "text", + "content": "Mingfei Han" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "text", + "content": ", Yali Wang" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{\\dagger,2,6}" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "text", + "content": ", Zhihui Li" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "text", + "content": ", Lina Yao" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "text", + "content": ", Xiaojun Chang" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 96, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{6,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "text", + "content": "ReLER, AAII, UTS " + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "text", + "content": "Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, SIAT-SenseTime Joint Lab, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences " + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "text", + "content": "Shandong Artificial Intelligence, Qilu University of Technology " + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "text", + "content": "Data61, CSIRO " + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "text", + "content": "Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence " + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 54, + 175, + 545, + 258 + ], + "type": "text", + "content": "Shanghai AI Laboratory, Shanghai, China https://mingfei.info/HTML/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 158, + 286, + 206, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 286, + 206, + 300 + ], + "spans": [ + { + "bbox": [ + 158, + 286, + 206, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 312, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 290, + 586 + ], + "type": "text", + "content": "Referring Video Object Segmentation (RVOS) is to segment the object instance from a given video, according to the textual description of this object. However, in the open world, the object descriptions are often diversified in contents and flexible in lengths. This leads to the key difficulty in RVOS, i.e., various descriptions of different objects are corresponding to different temporal scales in the video, which is ignored by most existing approaches with single stride of frame sampling. To tackle this problem, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) framework, which can effectively align linguistic and visual features to discover core object semantics in the video, by learning multimodal interaction hierarchically from different temporal scales. More specifically, we introduce a novel inter-scale multimodal perception module, where the language queries dynamically interact with visual features across temporal scales. It can effectively reduce complex object confusion by passing video context among different scales. Finally, we conduct extensive experiments on the widely used benchmarks, including Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences, where our HTML achieves state-of-the-art performance on all these datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 601, + 128, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 622, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 287, + 696 + ], + "type": "text", + "content": "Referring Video Object Segmentation (RVOS) has witnessed the growing interest, due to its wide applications in visual editing, virtual reality, human-robotic interaction and so on. Different from the traditional vision-only VOS, RVOS aims to segment the object instance from an input video, according to an open-world description about the re" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 285, + 541, + 574 + ], + "blocks": [ + { + "bbox": [ + 315, + 285, + 541, + 574 + ], + "lines": [ + { + "bbox": [ + 315, + 285, + 541, + 574 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 541, + 574 + ], + "type": "image", + "image_path": "8785cf05a20cb459bacb962e746ece8fe3f5be378d46a37bb55b96cbe53b8a6e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "content": "Figure 1: Referring descriptions in different lengths. (a) The description is simple containing only the category name. (b) The description is complicated with movement and position of the object. Single-scale baseline (e.g., four frames in (a) and two frames in (b)) fails to segment the referred object, while our hybrid-scale HTML succeeds. More discussion can be found in introduction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": "ferred object. In this case, the model has to learn both visual and textual contents comprehensively, in order to discover the underlying object by multimodal interaction." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 145, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13414" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Recent studies [4, 12, 28, 29] have shown that, cross-modal attention is an effective way to bridge the gap between vision and language in RVOS. However, these approaches perform vision-language interactions with video frames sampled from a single temporal scale, which may limit their power to infer the referred object with accurate segmentation. The main reason is that, the open-world descriptions vary in length and contain rich semantics about the referred object, e.g., where it is, how it moves, which objects it interact with. Apparently, such diversified texts are corresponding to various temporal-scale snippets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 396 + ], + "type": "text", + "content": "For example, the language query in Fig. 1 (a) is a tennis ball. Such a short description is corresponding to the ball located at a small region in the middle two frames. If the single-scale baseline samples four frames as input, it will fail to segment the referred object. This is because it overlooks the dog in the center place among all these four frames, while lacking the detailed understanding in the middle two frames. Alternatively, the language query in Fig. 1 (b) is a sheep top second right moves down and comes out of the circle. Such a long description is corresponding to the particular sheep in the group, which moves across frames. If the single-scale baseline samples two frames as input, it will fail to segment the referred object. This is because it is misled by the subtle movement of sheep group in only two frames, without understanding how each sheep moves from the adjacent frames." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 396, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 624 + ], + "type": "text", + "content": "To tackle this difficulty, we propose a concise Hybrid Temporal-scale Multimodal Learning (HTML) Framework for RVOS, which can alleviate object confusion by language-vision interactions across different temporal scales. Specifically, we sample video frames according to different temporal scales. For each temporal scale, we introduce an intra-scale multimodal perception module, which can effectively exploit core visual semantics within the frames at this temporal scale, by mutual enhancement between textual and visual embeddings. Then, we design an inter-scale multimodal perception module, where linguistic embeddings dynamically interact with visual features across temporal scales. In this case, we can hierarchically leverage object context from all the temporal scales to boost RVOS. Finally, we evaluate our HTML on a number of benchmarks, including Ref-Youtu-VOS [22], Ref-DAVIS17 [9], A2D-Sentences and JHMDB-Sentences [7]. The extensive experiments have shown that, our HTML achieves the state-of-the-art performance on all of them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 625, + 262, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 625, + 262, + 637 + ], + "spans": [ + { + "bbox": [ + 58, + 625, + 262, + 637 + ], + "type": "text", + "content": "Overall we make three contributions in this paper:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": "- Concise and unified learning framework: our Hybrid Temporal-scale Multimodal Learning (HTML) framework hierarchically constructs multimodal interactions via different strides of frame sampling, which can mutually enhance embeddings from both modalities for accurate segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 547, + 194 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "- Effective multimodal perception module: our Cross-scale Multimodal Perception (CMP) module can effectively reduce complex object confusions with intra-scale and inter-scale multimodal perceptions, where linguistic and visual features interact across temporal scales." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 133, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 547, + 194 + ], + "type": "text", + "content": "- State-of-the-art performance on the widely-used benchmarks, which shows the superiority of our framework. Specifically, on Ref-Youtu-VOS [22], our method with ResNet-50 achieves 57.8 in " + }, + { + "bbox": [ + 306, + 133, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}" + }, + { + "bbox": [ + 306, + 133, + 547, + 194 + ], + "type": "text", + "content": ", outperforming the recent SOTA method [29] with ResNet-101." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 203, + 395, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 395, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 395, + 216 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 223, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 547, + 462 + ], + "type": "text", + "content": "Vision-only Video Segmentation. Vision-only video segmentation tasks [20, 19, 31, 1, 30, 27, 10, 32, 18, 25, 23, 36], need to segment the objects with predefined semantic query set. Video instance segmentation (VIS) [31] needs to segment the different instances in the desired category set and track each instance with its identity kept. Video object segmentation (VOS) [20, 19], the model needs to separate an object from the background in the video, given the mask of the first mask. In the early stage of VIS, heavy supervision and complex heuristic rules are applied to associate the instances across frames by separate segmentation and association branches, such as MaskTrackRCNN[31] and MaskProp[1]. Recent progress in transformer[24] and DETR[3] has inspired the community with end-to-end works, such as SeqFormer[30] and VisTR[27]. As for VOS, most recent works [32, 18, 25, 23, 36] on semi-supervised VOS enforces temporal consistency in the video to propagate the first-frame mask to the other frames sequentially. Most of them lie in the group of matching-based methods, updating the memories and matching to segment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 463, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 547, + 714 + ], + "type": "text", + "content": "Referring Video Object Segmentation. RVOS task [22] is tasked to segment the referred object from the given video with specified open-world language description. Most early methods in RVOS proposed to refer the object by applying image-level methods on video frames separately and associate them with heuristic rules. However, they usually fail to utilize the temporal dynamic. [22] casts the task as a joint problem of referring segmentation in frame and mask propagation across frames by a memory attention module. [12] proposed a top-down pipeline by constructing exhaustive set of object tracklets and then selecting the target by matching the language features with the all the candidate tracklets. [37] proposed to model the temporal dynamic with an additional optical flow modality. [28] argued the importance of the structural information of video content and proposed to utilize the frame, object and video features simultaneously to obtain better representation. More recently, [2] introduced the DETR structure to RVOS area and [29] proposed to use language-conditional queries to simplify the referring pipeline and improve the performance, which serves as our baseline." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13415" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "Different from the previous works, we raise the mismatch issue that the various descriptions of different objects are corresponding to different temporal scales of the video. Moreover, we propose a concise HTML framework via multimodal interaction across different temporal scales to capture the core object semantics in the video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 103, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 103, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 103, + 168 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 177, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 309 + ], + "type": "text", + "content": "To effectively align diversified descriptions and complex videos, we propose a distinct Hybrid Temporal-scale Multimodal Learning (HTML) framework for RVOS. In this section, we introduce our HTML in detail. First, we deliver an overview of HTML framework. Then, we explain how to build the hybrid temporal-scale multimodal learning paths, in the aid of vision-conditioned linguistic decoder and language-conditioned visual decoder. Next, we introduce a Cross-scale Multimodal Perception (CMP) module to align multimodal features across temporal scales. Finally, we describe the training objectives to optimize our HTML." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 319, + 173, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 173, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 173, + 330 + ], + "type": "text", + "content": "3.1. Framework Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "text", + "content": "As shown in Fig. 2, our HTML framework consists of three main parts. First, we need to extract visual and linguistic features from backbones. We adopt a visual backbone to extract frame features from " + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "text", + "content": " frames sampled from the given video. It can be either 2D CNN networks or 3D transformer networks. We then feed the extracted vision features into a deformable transformer encoder [38] to construct spatiotemporal relations between different frames. Meanwhile, to make a fair comparison with previous works in RVOS [29], we utilize the pretrained linguistic embedding model, RoBERTa [15], to extract textual features " + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_e\\in" + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{L\\times C}" + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "text", + "content": " from language descriptions with " + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 337, + 287, + 491 + ], + "type": "text", + "content": " words. More details can be found in Sec. 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 494, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 613 + ], + "type": "text", + "content": "After extracting visual and linguistic features, we next construct multimodal interactions between the language descriptions and the videos. Different from the previous approaches [29, 28], we build " + }, + { + "bbox": [ + 46, + 494, + 287, + 613 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 494, + 287, + 613 + ], + "type": "text", + "content": " multimodal learning paths, where the linguistic embedding hierarchically interacts with visual features in different temporal scales. Then, we incorporate the mutually enhanced visual and linguistic features by a novel Cross-scale Multimodal Perception (CMP) module to align multimodal features across different scales. Finally, we design the training losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 622, + 284, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 284, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 284, + 635 + ], + "type": "text", + "content": "3.2. Hybrid Temporal-scale Multimodal Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "To capture core object semantics, we propose a novel hybrid temporal-scale multimodal learning framework to learn multimodal relations. To start with, we build hybrid temporal scales via different sampling strides. Then, we construct basic multimodal relation learning units. Finally, we explain how to construct hybrid learning paths." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 497, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 497, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 497, + 84 + ], + "type": "text", + "content": "3.2.1 Hybrid Temporal Scale Construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 92, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 92, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 92, + 545, + 152 + ], + "type": "text", + "content": "Since the texts of various objects may refer to different video parts, single temporal scale often fails to describe the diversified textual contents. To simulate such diversity and flexibility, we build hybrid temporal scales by periodically sampling frames with different strides." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "content": "We first regard all the input frames as the first temporal scale, and then build other " + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "inline_equation", + "content": "L - 1" + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "content": " temporal scales upon it in a sequential manner. In order to ensure the diversity of sampled temporal scale, we randomly pick one frame from every " + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "content": " frames of last scale, where " + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "content": " denotes the predefined stride. Subsequently, we feed the sampled frames into visual encoder Encoder(V) to extract feature maps for each of the scales respectively. Specifically, for temporal scale " + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "content": ", we can obtain " + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 152, + 546, + 272 + ], + "type": "text", + "content": " denotes the number of frames in the temporal scale." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 279, + 470, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 470, + 291 + ], + "type": "text", + "content": "3.2.2 Multimodal Relation Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "content": "In order to discover the core object semantics, we construct multimodal relations via vision-conditioned linguistic decoder and language-conditioned visual decoder to align semantics between different modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "text", + "content": "Vision-Conditioned Linguistic Decoder. In order to align linguistic object semantics to the vision contents, we design a vision-conditioned linguistic decoder " + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}(\\mathrm{L}|\\mathrm{V})" + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "text", + "content": ". Specifically, we have visual features " + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\in \\mathbb{R}^{T\\times H\\times W\\times C}" + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "text", + "content": ", and linguistic embeddings " + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_e\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "text", + "content": ". The vision-conditioned multimodal relations " + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{e} = \\mathrm{Decoder}(\\mathbf{s}_e|\\mathbf{M})" + }, + { + "bbox": [ + 304, + 350, + 545, + 434 + ], + "type": "text", + "content": " are constructed by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 445, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 445, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 359, + 445, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {0} = \\operatorname {D e f o r m A t t n} (\\mathbf {s} _ {e} + \\mathbf {q}, \\mathbf {M}), \\tag {1}", + "image_path": "69447149ae6fe7d3f4b14282076d0500a7b40b23a27acb1afa5c7b0a33343e0e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 459, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 459, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 359, + 459, + 545, + 472 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {k} = \\operatorname {D e f o r m A t t n} (\\mathbf {e} _ {k - 1}, \\mathbf {M}), \\tag {2}", + "image_path": "f2c97d8fdd2b0de6913f5eb5113dd01e3875413469bd1d4442c057f6160bf014.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K - 1\\}" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": ". We first add " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_e" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": " with learnable queries " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": " to represent candidate instances in the video. Then, we use deformable attention module [38] to reason the vision-conditioned multimodal relations where vision features serve as key and value to decompose linguistic features, as in Eq. (1). Finally, we stack the cross-attention module for " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": " times, as in Eq. (2)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Language-Conditioned Visual Decoder. In order to align visual object semantics to linguistic contents, we design a language-conditioned visual decoder " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}(\\mathrm{V}|\\mathrm{L})" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " (similar to Eqs. (1) and (2)), to enhance visual representation with the attendance of the language description. Differently, vision features are enhanced by multi-head self-attention (MHSA) modules at the first place, and then linguistic features " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_e" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " serve as key and value in cross-attention modules. In this case, it can reason the language-conditioned multimodal relations. Finally, we can get enhanced visual features by language-conditioned multimodal relations, as " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\mathrm{Decoder}(\\mathbf{M}|\\mathbf{s}_e)" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13416" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 545, + 259 + ], + "type": "image", + "image_path": "b6368af0dd18156247a7285b9233ada7ff5ceeb39c76b857124cf3ac90575304.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 546, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 546, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 546, + 309 + ], + "type": "text", + "content": "Figure 2: Our Hybrid Temporal-scale Multimodal Learning framework. It aligns linguistic and visual features by learning hierarchical multimodal interactions with hybrid temporal scales, detailed in Sec. 3.2.2. Moreover, a Cross-scale Multimodal Perception (CMP) module is designed to enable interaction and cooperation among temporal scales, detailed in Sec. 3.3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": "Hierarchical Multimodal Learning As single-scale multimodal learning is insufficient to understand the relations between videos and texts, we propose to construct the multimodal relations hierarchically for the " + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": " hybrid temporal scales with the assistance of " + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}(\\mathrm{L}|\\mathrm{V})" + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}(\\mathrm{V}|\\mathrm{L})" + }, + { + "bbox": [ + 46, + 324, + 287, + 432 + ], + "type": "text", + "content": ". We can obtain linguistic-attended visual features and visual-attended linguistic embeddings for different temporal scales, capturing core object semantics conditioned on different visual and linguistic contexts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 439, + 234, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 234, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 234, + 451 + ], + "type": "text", + "content": "3.3. Cross-scale Multimodal Perception" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 529 + ], + "type": "text", + "content": "The multimodal relations are constructed conditioned on different modalities with hybrid temporal scales. However, the modeling process of different scales is independent. To promote the cooperation and align the visual and linguistic semantics both within the scale and across the scales, we design Cross-scale Multimodal Perception (CMP) module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 532, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 591 + ], + "type": "text", + "content": "Intra-scale Perception. Despite sharing visual and linguistic feature extraction, the multimodal relation construction via Decoder(V|L) and Decoder(L|V) are independent to each other. To promote the cooperation between modalities, we propose an intra-scale perception module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "content": "Specifically, in each temporal scale " + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "content": ", we have visual attended linguistic embeddings " + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{e} \\in \\mathbb{R}^{N \\times T \\times D}" + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "content": " and linguistic attended visual features " + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "content": ". To achieve fine-grained semantic alignment, we measure the similarity by dot product between " + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "content": " on pixel level. Specifically, we obtain the similarity map via multimodal perception module, denoted as " + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\mathrm{MP}(\\mathbf{F}, \\mathbf{e})" + }, + { + "bbox": [ + 46, + 592, + 287, + 675 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 682, + 287, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 682, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 123, + 682, + 287, + 695 + ], + "type": "interline_equation", + "content": "\\Omega = \\operatorname {M a s k H e a d} (\\mathbf {e}), \\tag {3}", + "image_path": "18f33d2475b2bfc38bef1c1d378dc56629b49f9a8f0def5dde696d3a04028dee.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 124, + 698, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 698, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 124, + 698, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\mathcal {I} = \\Omega \\cdot \\mathbf {F}, \\tag {4}", + "image_path": "2bf70d3b0cf1bca4206b9b427a2e5286b9f1fb0c9bdd5b6c4ef526a0992f5783.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 324, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 420 + ], + "type": "text", + "content": "where MaskHead denotes three consecutive MLP layers for embedding conversion. Each value of " + }, + { + "bbox": [ + 304, + 324, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 324, + 545, + 420 + ], + "type": "text", + "content": " represents the relevance between visual-attended linguistic embeddings and linguistic-attended visual features, which can be interpreted as the existence of the referred object. As such, " + }, + { + "bbox": [ + 304, + 324, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 324, + 545, + 420 + ], + "type": "text", + "content": " is regarded as the object mask prediction with the context of current temporal scale. To this end, we achieve the multimodal perception in the same temporal scale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 423, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 495 + ], + "type": "text", + "content": "Inter-scale Perception. The multimodal relations are constructed in different temporal scales. However, the process in each scale is independent and biased towards the contained object semantics. To alleviate it, we propose to align multimodal features from different temporal scales with a concise inter-scale perception module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": "Specifically, suppose that the referred object appears in frame " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": " in temporal scales " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": ", the referred object can be segmented by measuring similarity between " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}^l (t),\\mathbf{e}^l (t))" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}^{l + 1}(t),\\mathbf{e}^{l + 1}(t))" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": " simultaneously. Conditioned on same frame " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": ", the visual-attended linguistic embedding " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^l (t)" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": " from scale " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": " is supposed to be relevant to the linguistic-attended visual features " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{l + 1}(t)" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": " from scale " + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 304, + 495, + 545, + 602 + ], + "type": "text", + "content": ". Thus, similar to Eq. (4), the similarity cross different temporal scales can be measured by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 608, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 608, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 354, + 608, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {I} ^ {l \\rightarrow l + 1} (t) = \\mathrm {M P} (\\mathbf {e} ^ {l} (t), \\mathbf {F} ^ {l + 1} (t)). \\tag {5}", + "image_path": "905af18a9c17ca6787565ba06f1e6a7d6b03abe5980cc23f4f7bda1a9474468b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Without losing generality, the inter-scale similarity can also be measured by " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{l + 1\\to l}(t)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". More specifically, frame " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " can be any frame shared by the adjacent temporal scales, which is ensured by our hybrid temporal scales sampling strategy. Each value in " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{l\\rightarrow l + 1}(t)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " represents the referred object prediction with the prior of linguistic object semantic from temporal scale " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". In the same way, values in " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{l + 1\\rightarrow l}(t)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " rep" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13417" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "content": "resent the opposite. To this end, we achieve multimodal perception across different temporal scales." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 160, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 160, + 116 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 160, + 116 + ], + "type": "text", + "content": "3.4. Training objectives" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 120, + 286, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 286, + 165 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 286, + 165 + ], + "type": "text", + "content": "Our network can be trained in an end-to-end manner to locate and segment the target instance simultaneously. Specifically, the losses for intra-scale and inter-scale multimodal perception in temporal scale " + }, + { + "bbox": [ + 46, + 120, + 286, + 165 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 120, + 286, + 165 + ], + "type": "text", + "content": " are formed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 171, + 286, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 171, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 171, + 286, + 205 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {i n t r a}} ^ {l} = \\sum \\mathcal {L} _ {\\text {c l s}} (\\mathbf {y}, \\hat {\\mathbf {y}}) + \\mathcal {L} _ {\\text {b o x}} (\\mathbf {b}, \\hat {\\mathbf {b}}) + \\mathcal {L} _ {\\text {m a s k}} (\\mathcal {I}, \\hat {\\mathcal {I}}), (6) \\\\ \\mathcal {L} _ {\\text {i n t e r}} ^ {l} = \\sum \\mathcal {L} _ {\\text {m a s k}} \\left(\\mathcal {I} ^ {l + 1 \\rightarrow l}, \\hat {\\mathcal {I}}\\right) (7) \\\\ \\end{array}", + "image_path": "f77be02a461fea6e6a8545ad95b7f91df0fa26f516aa38488e94b3c61dd7808d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": "where time and instance subscripts are omitted for simplicity, " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " denote binary classification for instance existence and bounding box prediction respectively. Here " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " is the focal loss [13], " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box}" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " is the sum of L1 loss and GIoU loss [21], and " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " is the combination of DICE loss [17] and binary mask focal loss. We optimize the network by first finding the best prediction as the positive sample, via minimizing the matching cost " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{intra}^{l}" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inter}^{l}" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " in each temporal scale " + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 209, + 286, + 342 + ], + "type": "text", + "content": " respectively. Then, we average the matching losses from different temporal scales and perception modules, and minimize it for positive samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 351, + 128, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 128, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 128, + 365 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 371, + 168, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 168, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 168, + 383 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "text", + "content": "Datasets. We conduct experiments on four datasets: Ref-Youtube-VOS[22], Ref-DAVIS17[9], A2D-Sentences and JHMDB-Sentences[7], following the common practice[29]. Metrics. We follow the standard evaluation protocol [22, 28, 29] to adopt region similarity " + }, + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\%)" + }, + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "text", + "content": ", contour accuracy " + }, + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\%)" + }, + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "text", + "content": " and mean " + }, + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}" + }, + { + "bbox": [ + 46, + 389, + 286, + 496 + ], + "type": "text", + "content": " for Ref-Youtube-VOS and Ref-DAVIS17. For JHMDB-Sentences, we adopt mAP to evaluate the model. For A2D-Sentences, we use Precision@K, Overall IoU, Mean IoU and mAP for evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 503, + 167, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 167, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 167, + 516 + ], + "type": "text", + "content": "4.2. Implemented Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "content": "We set the number of attention layers " + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "content": " to 4 and the hidden dimension " + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "content": " to 256. The number of learnable queries " + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "content": " is set to 5. The number of MaskHead output channels " + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "content": " is set to 8. During training, we first sample video clips by sliding windows and then generate " + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "content": " hybrid temporal scales with stride " + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "inline_equation", + "content": "h = 2" + }, + { + "bbox": [ + 46, + 521, + 286, + 653 + ], + "type": "text", + "content": " for generalized multimodal representations and relations. We use same training recipes as in [29, 2]. All frames are downsampled by shorter side to 360 and limit the maximum size for the long side to 640. Our model is pretrained on image referring segmentation datasets [35, 35]. See supplementary materials for details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "During inference, we report results with all input frames in single temporal scale for fair comparisons. On RefDAVIS17, we directly inference on models trained on RefYoutube-VOS. Similarly, we reports JHMDB-Senteces results directly on models trained with A2D-Sentences." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 72, + 540, + 358 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 540, + 358 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 540, + 358 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 540, + 358 + ], + "type": "table", + "html": "
MethodBackboneRef-Youtube-VOS
J&FJF
CMSA [33]ResNet-5034.933.336.5
CMSA + RNN [33]ResNet-5036.434.838.1
URVOS [22]ResNet-5047.245.349.2
LBDT-4 [22]ResNet-5047.245.349.2
MLRL [28]ResNet-5049.748.451.0
ReferFormer [29]ResNet-5055.654.856.5
OursResNet-5057.856.559.0
PMINet [6]ResNeSt-10148.246.749.6
PMINet + CFBI [6]ResNeSt-10153.051.554.5
CITD [12]ResNet-10156.454.858.1
ReferFormer [29]ResNet-10157.356.158.4
OursResNet-10158.557.359.8
PMINet + CFBI [6]Ensemble54.253.055.5
CITD [12]Ensemble61.460.062.7
ReferFormer [29]Swin-L62.460.864.0
OursSwin-L63.461.565.3
MTTR [2]Video-Swin-T55.354.056.6
ReferFormer [29]Video-Swin-T59.458.060.9
OursVideo-Swin-T61.259.563.0
ReferFormer [29]Video-Swin-S60.158.661.6
OursVideo-Swin-S61.459.962.9
ReferFormer [29]Video-Swin-B62.961.364.6
OursVideo-Swin-B63.461.565.2
", + "image_path": "30cdba508ff2980263bce985a27c043bbb16274fee3d1fcdc5b3b409baa39480.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 397, + 540, + 550 + ], + "blocks": [ + { + "bbox": [ + 305, + 366, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 389 + ], + "type": "text", + "content": "Table 1: Comparison with the SOTA methods on Ref-YTB-VOS." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 397, + 540, + 550 + ], + "lines": [ + { + "bbox": [ + 310, + 397, + 540, + 550 + ], + "spans": [ + { + "bbox": [ + 310, + 397, + 540, + 550 + ], + "type": "table", + "html": "
MethodBackboneRef-DAVIS17
J&FJF
CMSA [33]ResNet-5034.732.237.2
CMSA + RNN [33]ResNet-5040.236.943.5
URVOS [22]ResNet-5051.547.356.0
LBDT-4 [5]ResNet-5054.5--
MLRL [28]ResNet-5058.053.962.0
ReferFormer [29]ResNet-5058.555.861.3
OursResNet-5059.556.662.4
ReferFormer [29]Swin-L60.557.663.4
OursSwin-L61.658.964.4
ReferFormer [29]Video-Swin-B61.158.164.1
OursVideo-Swin-B62.159.265.1
", + "image_path": "f10f81c2064fac0eded195187188d829e23e6981d65b27056ab94f7914969049.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 557, + 545, + 581 + ], + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 581 + ], + "type": "text", + "content": "Table 2: Comparison with the SOTA methods on Ref-DAVIS17." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 598, + 422, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 422, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 422, + 611 + ], + "type": "text", + "content": "4.3. SOTA Comparisons" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We compare our method with the state-of-the-art methods on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences. On Ref-Youtu-VOS, our approach achieves 58.5 in " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}(\\%)" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " with ResNet-50, as shown in Tab. 1, which surpasses the recent SOTA method ReferFormer[29] with same backbone by 2.2 points. Moreover, it surpasses all the other SOTA methods with larger ResNet-101 on all evaluation metrics, which fully sug" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13418" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 539, + 234 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 539, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 539, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 539, + 234 + ], + "type": "table", + "html": "
MethodBackbonePrecisionIoUmAP
P@0.5P@0.6P@0.7P@0.8P@0.9OverallMean
Hu et al. [8]VGG-1634.823.613.33.30.147.435.013.2
Gavrilyuk et al. [7]I3D47.534.721.18.00.253.642.119.8
CMSA + CFSA [34]ResNet-10148.743.135.823.15.261.843.2-
ACAN [26]I3D55.745.931.916.02.060.149.027.4
CMPC-V [14]I3D65.559.250.634.29.865.357.340.4
ClawCraneNet [11]ResNet-50/10170.467.761.748.917.163.159.9-
MTTR (ω = 8) [2]Video-Swin-T72.168.460.745.616.470.261.844.7
MTTR (ω = 10) [2]Video-Swin-T75.471.263.848.516.972.064.046.1
ReferFormer [29]Video-Swin-T82.879.272.355.319.377.669.652.8
ReferFormer [29]Video-Swin-B83.180.474.157.921.278.670.355.0
OursVideo-Swin-T82.279.272.355.320.177.669.253.4
OursVideo-Swin-B84.081.575.859.222.879.571.256.7
", + "image_path": "b252803b50c2537e6c1985c943523864a1279f72c82652d893cd749f1bb671b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 266, + 272, + 407 + ], + "blocks": [ + { + "bbox": [ + 146, + 240, + 447, + 252 + ], + "lines": [ + { + "bbox": [ + 146, + 240, + 447, + 252 + ], + "spans": [ + { + "bbox": [ + 146, + 240, + 447, + 252 + ], + "type": "text", + "content": "Table 3: Comparison with the state-of-the-art methods on A2D-Sentences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 266, + 272, + 407 + ], + "lines": [ + { + "bbox": [ + 62, + 266, + 272, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 272, + 407 + ], + "type": "table", + "html": "
MethodBackbonemAP
Hu et al. [8]VGG-1617.8
Gavrilyuk et al. [7]I3D23.3
ACAN [26]I3D28.9
CMPC-V [14]I3D34.2
MTTR (ω = 8) [2]Video-Swin-T36.6
MTTR (ω = 10) [2]Video-Swin-T39.2
ReferFormer† (ω = 6) [29]Video-Swin-T39.1
ReferFormer [29]Video-Swin-T42.2
ReferFormer [29]Video-Swin-B43.7
OursVideo-Swin-T42.7
OursVideo-Swin-B44.2
", + "image_path": "455fa75b1c75f1c1e22e36fef8d21a9197f204fecf42e4d98c7df905c500f683.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 439, + 268, + 525 + ], + "blocks": [ + { + "bbox": [ + 50, + 414, + 284, + 426 + ], + "lines": [ + { + "bbox": [ + 50, + 414, + 284, + 426 + ], + "spans": [ + { + "bbox": [ + 50, + 414, + 284, + 426 + ], + "type": "text", + "content": "Table 4: SOTA results comparison on JHMDB-Sentences." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 439, + 268, + 525 + ], + "lines": [ + { + "bbox": [ + 67, + 439, + 268, + 525 + ], + "spans": [ + { + "bbox": [ + 67, + 439, + 268, + 525 + ], + "type": "table", + "html": "
Components#FramesJ&FJF
i.Baseline555.654.856.5
ii.HTML w/o CMP556.054.757.4
iii.HTML556.355.057.5
iv.Baseline856.255.057.3
v.HTML w/o CMP857.156.058.2
vi.HTML857.856.559.0
", + "image_path": "4702bec8715d7c8f321165fbc5bd208f88a3f9653b2f4735eb616ca3e22eca81.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 533, + 284, + 544 + ], + "lines": [ + { + "bbox": [ + 49, + 533, + 284, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 284, + 544 + ], + "type": "text", + "content": "Table 5: Ablation study on the components of our HTML." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 556, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 286, + 675 + ], + "type": "text", + "content": "gests the superiority of our method. When equipped with larger backbone, our method still show considerable superiority with accuracy gap of 1.2 points for ResNet-101 and 1.0 points for Swin-L. We also experiment our method with the well-known Video Swin Transformers [16]. Our method with Video-Swin-Tiny backbone surpasses the SOTA method with the same backbone by 1.8 points. With larger Video-Swin Transformers (Small and Base models), our method still achieves SOTA performance, which shows the generality of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "As shown in Tab. 2, our method surpasses the SOTA methods on Ref-DAVIS17 by over 1.0 points on both ResNet-50, Swin-L and Video-Swin-Base backbones, with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 268, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 388 + ], + "type": "text", + "content": "new a SOTA record 62.1 in " + }, + { + "bbox": [ + 304, + 268, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}(\\%)" + }, + { + "bbox": [ + 304, + 268, + 545, + 388 + ], + "type": "text", + "content": ". We also experiment our method on A2D-Sentences and JHMDB-Sentences datasets and compare with other SOTA results as shown in Tab. 3 and Tab. 4. Our method achieves SOTA performances with new records on both the two datasets. On A2D-Sentences, our HTML surpass SOTA result by 1.7 points in mAP and higher recall by 1.6 on Precision@0.9. On JHMDB-Sentences, our method still achieves a new SOTA record with 44.2 in mAP. These results demonstrate the superiority of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 396, + 400, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 400, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 400, + 408 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 414, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 545, + 437 + ], + "type": "text", + "content": "In this section, we ablate core components of our HTML with Ref-Youtu-VOS based on ResNet-50." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": "Effectiveness of our HTML. To validate the effectiveness of our Hybrid Temporal-scale Multimodal Learning framework, we investigate each of our components by gradually adding them to the baseline [29]. First, comparing (i)&(iii) in Tab. 5, our HTML improves the performance of baseline to 56.3 when 5 frames are used for training, which is better than the baseline trained with longer input frames. Further, when longer temporal input is available, comparing (iv)&(vi), our HTML improves the model by 1.6 to 57.8 in " + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}" + }, + { + "bbox": [ + 304, + 437, + 545, + 569 + ], + "type": "text", + "content": ". These prove the effectiveness of our method with frames in different lengths." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "content": "Second, comparing counterpart networks using different number of frames, i.e. (iii) and (vi), our method can benefit from longer temporal input (8 frames vs. 5 frames) with an improvement of 1.5 points. This conclusion holds among the other counterpart network pair, i.e. (ii)&(v). Further, each of our components brings an improvement of 0.3 when trained with total 5 frames, while the improvement can be doubled to 0.7-0.9 with total 8 frames. This indicates that our method can better utilize the long temporal input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Finally, taking 8 frames for instance, hierarchical multimodal learning and cross-scale multimodal perception improves " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " by 0.9 and 0.7 respectively. It proves the ef" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13419" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 202, + 133 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 202, + 133 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 202, + 133 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 202, + 133 + ], + "type": "table", + "html": "
#ScalesJ&FJF
156.255.057.3
256.955.858.1
357.856.559.0
", + "image_path": "38707ac44701fc2730f3a0f164fbfe62f1b95df7d172ded15e0ec6afd77ee95a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 227, + 72, + 361, + 130 + ], + "blocks": [ + { + "bbox": [ + 60, + 138, + 192, + 150 + ], + "lines": [ + { + "bbox": [ + 60, + 138, + 192, + 150 + ], + "spans": [ + { + "bbox": [ + 60, + 138, + 192, + 150 + ], + "type": "text", + "content": "(a) Effect of No. of temporal scales." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 72, + 361, + 130 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 361, + 130 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 361, + 130 + ], + "type": "table", + "html": "
FramesJ&FJF
556.355.057.5
857.856.559.0
1257.556.458.6
", + "image_path": "36d6ca58585e8abab7372253db889acf8d61f12c6431af587ee11d47564643ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 383, + 70, + 528, + 133 + ], + "blocks": [ + { + "bbox": [ + 221, + 138, + 342, + 150 + ], + "lines": [ + { + "bbox": [ + 221, + 138, + 342, + 150 + ], + "spans": [ + { + "bbox": [ + 221, + 138, + 342, + 150 + ], + "type": "text", + "content": "(b) Effect of No. of input frames." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 383, + 70, + 528, + 133 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 528, + 133 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 528, + 133 + ], + "type": "table", + "html": "
Inter-scaleJ&FJF
None56.255.057.3
l → l + 156.955.858.0
l + 1 → l57.856.559.0
", + "image_path": "e45299fb6975b5804943ef1e51a7855223358517f3015f8d62bd13973f9c6c61.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 392, + 138, + 504, + 150 + ], + "lines": [ + { + "bbox": [ + 392, + 138, + 504, + 150 + ], + "spans": [ + { + "bbox": [ + 392, + 138, + 504, + 150 + ], + "type": "text", + "content": "(c) Effect of direction of CMP." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 69, + 158, + 523, + 171 + ], + "lines": [ + { + "bbox": [ + 69, + 158, + 523, + 171 + ], + "spans": [ + { + "bbox": [ + 69, + 158, + 523, + 171 + ], + "type": "text", + "content": "Table 6: Ablation study on Hybrid Temporal Scales in (a) and (b), and Cross-scale Multimodal Perception in (c)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 63, + 186, + 276, + 377 + ], + "blocks": [ + { + "bbox": [ + 63, + 186, + 276, + 377 + ], + "lines": [ + { + "bbox": [ + 63, + 186, + 276, + 377 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 276, + 377 + ], + "type": "image", + "image_path": "e96e42331b76dfa19e8048dac016a0ebf27b28ad5019ade6b42338b8e257168e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 384, + 287, + 409 + ], + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 409 + ], + "type": "text", + "content": "Figure 3: Performance comparison of different query sets with different temporal scales on Ref-Youtu-VOS." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 423, + 233, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 233, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 233, + 436 + ], + "type": "text", + "content": "fectiveness of each component of our HTML." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": "Length of language descriptions. We explore the performance of our method with different sampled sets of language descriptions, to validate the ability of capturing diversified linguistic object semantics. Since the train set of Ref-Youtu-VOS contains an average of two language descriptions for each object, we sample the shorter ones to form the Short set and the others to form the Long set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": "As shown in bottom line of Fig. 3, the Long set has longer sentences than the Short set in the average number of words. As shown in Fig. 3, when the objects are described by Short query set, more input frames in single temporal scale achieve inferior performance (5 frames vs. 8 frames: 55.6 vs. 55.1). It's interesting to note that more visual content (8 frames) fails to improve the performance when the linguistic content is insufficient (shorter queries). When the complexity of queries increases, i.e. with the Long query set, the quantitative relation is reversed: 8 frames guided model obtains better performance than the model with 5 frames, increasing by 0.8 (8 frames-Short query vs. 8 frames-Long query: 55.1 vs. 55.9). We infer that this is caused by the mismatch of visual and linguistic object semantics. On one side, shorter quires, i.e. relatively simple linguistic semantics, are insufficient to interpret longer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 186, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 209 + ], + "type": "text", + "content": "videos. On the other side, the longer language descriptions contain more content irrelevant to the visual input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 210, + 546, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 546, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 546, + 317 + ], + "type": "text", + "content": "Differently, our models trained with either the Short set or the Long set all surpass the single temporal scale guided model. Impressively, when the objects are described by queries flexible in lengths, i.e. with the All set, our method gets a performance boost by 1.6, while single temporal scale baseline (both 8 frames and 5 frames guided networks) improves by 0.3. This shows that our method has the strong ability to solve the mismatch issue between visual content and diversified linguistic contents." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "type": "text", + "content": "Number of hybrid temporal scales. We investigate the effectiveness of our hierarchical multimodal learning, by exploring the number of hybrid temporal scales. For fair comparison among different settings, the number of total input frames is set to 8 in this subsection. As shown in Tab. 6 (a), when two temporal scales are constructed, our method brings an improvement of 0.6 in " + }, + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}" + }, + { + "bbox": [ + 304, + 318, + 546, + 449 + ], + "type": "text", + "content": ". When the number of temporal scales is increased to three, continuous improvement of 0.9 (56.9 vs. 57.8) is observed. It proves that our model benefits from the increasing visual diversity constructed by hybrid temporal scales." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "text", + "content": "Number of input frames. We explore the effect of total input frames here. In this subsection, hybrid temporal scales are constructed by default following Sec. 3.2.2. As shown in Tab. 6 (b), more input frames bring an improvement by 1.5 in " + }, + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\&\\mathcal{F}" + }, + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "text", + "content": " (5 frames vs. 8frames: 56.3 vs. 57.8). When the input frames increase continuously to 12 frames, the performance saturates and drops slightly to 57.5. We conjecture that it is caused by insufficient video-language pairs (8 frames vs. 12 frames: " + }, + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "inline_equation", + "content": "49\\mathrm{k}" + }, + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "inline_equation", + "content": "32\\mathrm{k}" + }, + { + "bbox": [ + 304, + 449, + 546, + 581 + ], + "type": "text", + "content": ") compared to largely increased computation complexity (8 frames vs. 12 frames: 21.5 GFLOPs vs. 33.6 GFLOPs for transformers)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "content": "Direction of inter-scale perception. We explore the effect of direction of inter-scale multimodal perception in CMP. Comparing first two lines of Tab. 6 (c), " + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "inline_equation", + "content": "l \\rightarrow l + 1" + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "content": " perception improves the baseline by 0.7 points; As in first and last lines, " + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "inline_equation", + "content": "l + 1 \\rightarrow l" + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "content": " perception improves the baseline by 1.6 points. These prove the effectiveness of our inter-scale perception. We choose the latter one for better performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 671, + 394, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 394, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 394, + 682 + ], + "type": "text", + "content": "4.5. Visualizations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "We visualize the segmentation results of complex and simple language descriptions in Fig. 4. Specifically, we" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13420" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 518, + 403 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 518, + 403 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 518, + 403 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 518, + 403 + ], + "type": "image", + "image_path": "1aa1a49843365ea85fd4757ba4fb67cfefb379837cf3e66a2d12f7c5c8bb6336.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 406, + 547, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 406, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 547, + 431 + ], + "type": "text", + "content": "Figure 4: Visualization results of complex and simple language descriptions on Ref-Youtu-VOS. Red masks indicate positive segmentation results and blue masks indicate the negatives. Our HTML is able to clarify such object confusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 449, + 289, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 701 + ], + "type": "text", + "content": "compare three settings, i.e., baseline with only single temporal scale, HTML w/o CMP which constructs hybrid temporal scales, and our final HTML which dynamically construct multimodal relations cross temporal scales. As expected, when only single temporal scale is adopted, baseline fails to segment the target in the video, e.g., the background rabbit in the subplot (a) is mistakenly referred. The main reason is that the rabbit shares similar appearance to the target object and also locates to the left of a person in last three frames. The model is misled and confused by the single temporal scale. When applied with hybrid temporal scales, the language description can interact with both long and short temporal scales. Thus, the previous false prediction in the first frame is corrected. Further, when applied with CMP, our model is able to clarify the object confusion by discovering the core semantics on a scale and make correct predictions. Similarly, in subplot (b), baseline is misled by the video clip where tennis ball only appears in the middle two frames. When gradually applying our proposed modules, the mistakenly predicted dog is clarified and further the target tennis ball is correctly segmented." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 448, + 379, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 379, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 379, + 460 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "content": "In this work, we develop a HTML framework to align linguistic and visual features by learning multimodal relations hierarchically in different temporal scales. Moreover, we introduce an inter-scale multimodal perception module to construct dynamic multimodal interactions across temporal scales. We conduct experiments on four datasets and establish new state-of-the-art results. Particularly, our method with ResNet-50 backbone surpasses the recent methods with ResNet-100. The comprehensive ablation experiments and visualization results show that our method is able to discover core object semantics in the different modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (No.2022ZD0160100, No. 2022ZD0160505), the National Natural Science Foundation of China under Grant (62272450), the Joint Lab of CAS-HK, the Shenzhen Research Program (RCJC20200 714114557087), and the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020355)" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "13421" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 147 + ], + "type": "text", + "content": "[1] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9739-9748, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "type": "text", + "content": "[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. arXiv preprint arXiv:2111.14821, 2021. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 197, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 250 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 255, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 287, + 308 + ], + "type": "text", + "content": "[4] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 312, + 287, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 287, + 366 + ], + "type": "text", + "content": "[5] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "text", + "content": "[6] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 287, + 470 + ], + "type": "text", + "content": "[7] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 473, + 287, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 473, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 53, + 473, + 287, + 515 + ], + "type": "text", + "content": "[8] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 519, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 519, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 519, + 287, + 563 + ], + "type": "text", + "content": "[9] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 620 + ], + "type": "text", + "content": "[10] Liulei Li, Wenguan Wang, Tianfei Zhou, Jianwu Li, and Yi Yang. Unified mask embedding and correspondence learning for self-supervised video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18706-18716, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 656 + ], + "type": "text", + "content": "[11] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 152 + ], + "type": "text", + "content": "[14] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 153, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 206 + ], + "type": "text", + "content": "[15] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[16] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "text", + "content": "[17] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "text", + "content": "[18] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9226–9235, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 345, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 345, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 345, + 545, + 408 + ], + "type": "text", + "content": "[19] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[20] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 520 + ], + "type": "text", + "content": "[21] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 523, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 588 + ], + "type": "text", + "content": "[22] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[23] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[24] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[25] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmenta" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13422" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[26] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "text", + "content": "[27] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8741-8750, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "text", + "content": "[28] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "text", + "content": "[29] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 374 + ], + "type": "text", + "content": "[30] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "type": "text", + "content": "[31] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5188-5197, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "text", + "content": "[32] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In European Conference on Computer Vision, pages 332-348. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[33] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10502–10511, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[34] Linwei Ye, Mrigank Rochan, Zhi Liu, Xiaoqin Zhang, and Yang Wang. Referring segmentation in images and videos with cross-modal self-attention network. arXiv preprint arXiv:2102.04762, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "text", + "content": "[35] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[36] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2246-2256, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[37] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13423" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_content_list.json b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ceef4e7851526b11d6833ab13790f9a1b8488944 --- /dev/null +++ b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_content_list.json @@ -0,0 +1,1578 @@ +[ + { + "type": "text", + "text": "HaMuCo: Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning", + "text_level": 1, + "bbox": [ + 174, + 128, + 795, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaozheng Zheng $^{1,2\\dagger}$ Chao Wen $^{2\\dagger}$ Zhou Xue $^{2}$ Pengfei Ren $^{1,2}$ Jingyu Wang $^{1*}$ $^{1}$ State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications \n $^{2}$ PICO IDL, ByteDance, Beijing", + "bbox": [ + 99, + 208, + 875, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in 3D hand pose estimation have shown promising results, but its effectiveness has primarily relied on the availability of large-scale annotated datasets, the creation of which is a laborious and costly process. To alleviate the label-hungry limitation, we propose a self-supervised learning framework, HaMuCo, that learns a single-view hand pose estimator from multi-view pseudo 2D labels. However, one of the main challenges of self-supervised learning is the presence of noisy labels and the \"groupthink\" effect from multiple views. To overcome these issues, we introduce a cross-view interaction network that distills the single-view estimator by utilizing the cross-view correlated features and enforcing multi-view consistency to achieve collaborative learning. Both the single-view estimator and the cross-view interaction network are trained jointly in an end-to-end manner. Extensive experiments show that our method can achieve state-of-the-art performance on multi-view self-supervised hand pose estimation. Furthermore, the proposed cross-view interaction network can also be applied to hand pose estimation from multi-view input and outperforms previous methods under the same settings.", + "bbox": [ + 75, + 348, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D hand pose estimation is essential in various application scenarios, from action recognition and sign language translation to AR/VR [19, 20]. Hand pose estimation has achieved a significant improvement in recent years. However, the progress heavily relies on the emergence of many hand pose datasets with accurate 3D annotations. Acquiring labeled datasets is quite time-consuming and laborious, exposing a realistic challenge for deep learning models to learn with limited and noisy data.", + "bbox": [ + 75, + 713, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning is an emerging solution to the", + "bbox": [ + 94, + 849, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenge posed by manual annotation. Though worth exploring, self-supervised pose estimation with RGB hand images is a challenging and relatively unexplored area with only one pioneering method, S²HAND [9]. S²HAND aims to conduct 3D hand reconstruction from a single RGB image with the noisy off-the-shell 2D hand pose estimation results for supervision. Unfortunately, S²HAND faces a predicament where its performance is significantly reliant on the quality of the pseudo label, and inferior labeling may result in reduced performance. Moreover, evaluating the quality of the pseudo label is an ill-posed problem that lacks clear criteria or input, further complicating the issue.", + "bbox": [ + 500, + 316, + 892, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This observation motivates us to leverage multi-view information for enhancing self-supervised learning, as the complementary nature of multi-view observations can help mitigate the ambiguity inherent in pose estimation. Although the first 3D hand dataset with synchronized multiview input (HanCo [62]) was proposed in 2021, to our knowledge, there is no previous work exploring the potential of multi-view for self-supervised hand pose estimation. Therefore, we turn to studies in the human body pose estimation, which share some similarities.", + "bbox": [ + 496, + 501, + 892, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As mentioned in previous work [24], naively enforcing multi-view consistency is prone to generate degenerated solutions, thus they resorted to additional 2D labels of unrelated datasets and proposed a solution under the scope of weakly supervised learning. Other studies, such as EpipolarPose [27] and CanonPose [52], utilized multiview data with special designs to enhance the supervision and achieved promising results under the scope of self-supervised learning.", + "bbox": [ + 496, + 655, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we push along this direction on hand pose estimation via multi-view collaborative learning. We take one step further by designing a learnable network, which utilizes multi-view information, to tackle 1) noisy pseudo labels and 2) unreliable multi-view \"groupthink\" issues causing training collapse in the early training stage. Formally, we name the pipeline HaMuCo, which stands for", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://zxz267.github.io/HaMuCo.", + "bbox": [ + 78, + 873, + 442, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author. † Equal contribution.", + "bbox": [ + 78, + 887, + 341, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20763", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d09d4158ed2d1c0064b9952d4d2abe7ab9202033f3a8228437bc64ff9e29170.jpg", + "image_caption": [ + "Figure 1. Overall pipeline comparison: HaMuCo learns a monocular 3D hand pose estimator from multi-view self-supervision via cross-view feature interaction. Our cross-view interaction network addresses the importance of introducing learnable feature interaction, which is absent from previous methods [27, 52]. At inference time only the gray part is applied." + ], + "image_footnote": [], + "bbox": [ + 89, + 73, + 279, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b4d5b31d5e12985556bbef5fbb46b83d25db8a8469019971dfbeeecc03650dc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 71, + 500, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a62fda65e2c2842ade79d1f653ca1f18544362d5c308ca52374529863b39cfdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 73, + 877, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hand Multiview Collaborative learning.", + "text_level": 1, + "bbox": [ + 75, + 323, + 341, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core idea of our approach is to enhance the single-view estimation by means of cross-view feature interaction and further integrate multi-view results to supervise the single-view output to achieve self-distillation in an end-to-end fashion. Thus, our framework is built with a single-view hand pose estimator and a cross-view interaction network for supervision. The single-view estimator uses the MANO [45] hand model as the decoder, which provides the hand prior to regularizing irrational anatomy when supervised by noisy pseudo labels. The cross-view interaction network captures cross-view features and utilizes several consistent losses among different views to guide collaborative learning.", + "bbox": [ + 73, + 339, + 468, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments on the HanCo [62] dataset and our approach outperforms previous methods by a considerable margin for self-supervised 3D hand pose estimation. Notably, our results demonstrate competitive performance compared to a state-of-the-art fully supervised approach proposed by Chen et al. [7]. Our proposed framework is highly versatile, as it can be trained with or without calibration, and is capable of incorporating the cross-view interaction network to achieve superior multi-view inference results when multi-view test data is available. Moreover, our model can generalize well to other datasets [29, 46, 64] and in-the-wild images.", + "bbox": [ + 75, + 535, + 470, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are the following:", + "bbox": [ + 96, + 715, + 418, + 731 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first self-supervised learning framework for single-view hand pose estimation without any training data annotation and achieve state-of-the-art performance via multi-view collaborative learning.", + "- We propose a cross-view interaction network to supervise the single-view estimator by enforcing multi-view consistency and capturing cross-view features for collaborative learning among multiple views.", + "- The proposed framework is capable of multi-view inference by incorporating the cross-view interac" + ], + "bbox": [ + 94, + 737, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion network and achieves state-of-the-art performance without bells and whistles.", + "bbox": [ + 531, + 323, + 890, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 364, + 642, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hand Pose Estimation. Hand pose estimation can be categorized into RGB-based methods [23, 49, 63] and depth-based methods [14, 15, 36], depending on the input modality. In this paper, we focus our attention on RGB-based hand pose estimation. The RGB-based methods can be further divided into three categories, skeleton-based methods [5, 12, 23, 31, 38, 39, 48, 49, 55-57, 63], model-based methods [1, 2, 4, 9, 58, 59, 64], and mesh-based methods [7,8,10,16,28,30,32,33,37,50,61]. Skeleton-based methods regress the hand joints directly. Zimmermann et al. [63] introduces a multi-stage network that lifts the regressed 2D joints to 3D ones. Variational autoencoder [26] is employed to learn a cross-modal latent space to achieve better hand pose estimation and disentanglement [49, 56, 57]. Latent 2.5D representation regression is proved more effective than direct coordinates regression for hands by [23], which is also adopted by [13, 31, 48, 61]. There are also many works solving hand pose estimation with two hands interactions [13, 29, 38] and hand-object interactions [2, 12, 29]. Recent model-based methods make use of MANO [45], which can incorporate the hand prior and predict the hand mesh simultaneously. Those methods [1, 4, 9, 58, 59] rely on additional supervisions [1, 4, 9, 58, 59] or inputs [4]. In contrast, mesh-based methods regress each vertex directly, which is more accurate but requires large-scale datasets with hand mesh annotations [18, 29, 38, 64]. Most of these methods utilize graph convolutional network (GCN) [7,8,10,16,28,50,61] or transformers [32] or both [30, 33] for regression. I2L-MeshNet [37] regresses each vertex by predicting 1D heatmaps of three axes. Chen et al. [6] uses an image-to-image translation network to predict the UV map of the mesh. Similar to previous works [30, 33], we also use transformer and GCN. However, we employ them for cross-view interaction.", + "bbox": [ + 496, + 387, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20764", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-View Fully-Supervised Pose Estimation. Multiview information is widely explored to improve 3D human pose estimation by tackling occlusions and depth-ambiguity in a fully-supervised manner [3, 22, 25, 41-43, 47, 60]. Volume-based methods [25, 41, 42, 51] unproject 2D features or heatmaps of joints to a 3D space for estimation. Another kind of method [22, 43, 60] utilizes the geometry information to fuse the features in 2D space directly and efficiently. Recently, some works [35,47] utilize transformers for implicit cross-view fusion without camera extrinsics.", + "bbox": [ + 76, + 90, + 467, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Label-Efficient Learning. Label-efficient learning aims to reduce the 3D label requirements. Many works devote to solving hand pose estimation in a label-efficient manner [1, 4,5,7,9,39,48,55,59,63]. Synthetic data is used to avoid manual annotation [7,39,63], but may need domain transfer [39]. Or use weakly supervised learning [4,48] to obtain 3D results by manually annotating 2D labels to assist with hand priors. Multi-view label-efficient learning is also explored in 3D pose estimation [24, 27, 44, 52]. Rhodin et al. [44] trains a semi-supervised network with only a small amount of labeled 3D data and multi-view consistency constraints. Iqbal et al. [24] mixes single-view images with 2D labels and unlabelled multi-view images for training. Our goal is the same as that of previous methods, which is to train without any manual 3D labels.", + "bbox": [ + 76, + 243, + 467, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised 3D Pose Estimation. (1) Single-view training and inference. To the best of our knowledge, there is only one method for self-supervised 3D hand pose estimation, proposed by Chen et al. [9]. Their framework, $\\mathrm{S}^2\\mathrm{Hand}$ , uses only single-view 2D noisy labels for training and achieves self-supervision through rendering. However, the performance is limited due to the use of single-view information and the quality of the noisy labels. (2) Multi-view training, single-view inference. Our approach belongs to this category but is fundamentally different from the existing methods. EpipolarPose [27] triangulates multi-view 2D pseudo labels according to epipolar geometry to 3D ones for training. CanonPose [52] learns to lift 2D pseudo labels to 3D canonical pose space with multi-view consistency constraints. All the aforementioned methods use nonlearnable self-supervised modules like geometric modules or consistency loss functions, as shown in Fig. 1. However, they [27, 52] ignore the importance of introducing cross-view interaction and multi-view collaborative learning. Previous methods struggle to achieve good performance since the pose of a hand can change drastically over time and different joints may have similar appearances.", + "bbox": [ + 76, + 470, + 467, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 816, + 166, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in the left part of Fig. 2, our framework consists of a simple yet effective single-view estimator and cross-view interaction network. The core idea of our approach is that prediction from a monocular view can be", + "bbox": [ + 76, + 839, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "enhanced via cross-view feature interaction and the interacted results can further supervise the single-view output to achieve self-distillation.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Single-View Estimator", + "text_level": 1, + "bbox": [ + 500, + 142, + 707, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Our framework takes multi-view synchronized hand images $\\mathcal{I} = \\{I_i\\}_{i=0}^v$ with $v$ views as input, each view is an image of $I_i \\in \\mathbb{R}^{3 \\times h \\times w}$ . The output is a 3D hand mesh $M$ on each view. We designed a simple yet effective model-based network as a single-view estimator. Using the hand model will reduce the adverse effects of using poor pseudo labels as supervision by providing hand prior information for regularization. Please refer to supplementary materials for more details about the single-view estimator.", + "bbox": [ + 498, + 162, + 890, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand model. We employ MANO [45] as the hand model. The hand mesh can be derived from the MANO layer using parameters $\\beta$ and $\\pmb{\\theta}$ , i.e. $M(\\beta, \\pmb{\\theta})$ . $\\beta \\in \\mathbb{R}^{10}$ and $\\pmb{\\theta} \\in \\mathbb{R}^{16 \\times 3}$ control the shape and pose of the hand respectively. We can use a predefined regressor to obtain the 3D joints from the 3D mesh vertices by $P = JM$ , where $J \\in \\mathbb{R}^{k \\times n}$ , where $n = 778$ and $k = 21$ are the joints number and vertices number. For more details, we recommend referring to [45].", + "bbox": [ + 498, + 297, + 890, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera model. Following Boukhayma et al. [4], we model the geometry correspondence by the weak-perspective camera model and obtain camera parameters from the single-view network predictions. Given the translation $\\pmb{t}$ and scale $s$ , the 2D coordinates in image plane can be obtained by: $\\Pi(P) = s\\Omega(P) + \\mathbf{t}$ , where $\\Omega$ is the orthographic projection and $\\Pi$ denotes the weak-perspective projection.", + "bbox": [ + 498, + 419, + 890, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network Structure. Since the single-view estimator is not the main component, for the sake of simplicity, we employ a CNN as the encoder $F_{e}$ , and an MLP as the decoder $F_{d}$ for regressing the MANO parameters. We have 3D hand mesh: $M_{i}(\\pmb{\\theta}_{i},\\pmb{\\beta}_{i}) = F_{s}(\\pmb{I}_{i})$ , where $F_{s} = F_{d}(F_{e}(\\cdot))$ denotes the entire single-view network. The estimator also passes different levels of features $H^{j}$ (where $H^{j}$ is the intermediate feature of the encoder after $j$ residual blocks, $j = 1,2,3,4$ ) to our cross-view interaction network.", + "bbox": [ + 498, + 525, + 890, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cross-view Interaction Network", + "text_level": 1, + "bbox": [ + 500, + 671, + 781, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the cross-view interaction network (CVI-Net), which is the core of our system to enable the network to exploit multi-view information. This stage conducts cross-view interaction and distillation. The critical components of this stage are a cross-view interaction network for capturing cross-view features and several consistent losses for guiding collaborative learning.", + "bbox": [ + 498, + 694, + 890, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 View-Shared Graph Feature Extraction", + "text_level": 1, + "bbox": [ + 500, + 806, + 828, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first step for interaction is to extract the appropriate features. Different from [7, 8, 61], our module collects useful information into a graph through view-shared graph feature extraction module (VSGFE) as shown in Fig. 2. Specifically, it makes use of multi-level feature maps from dif", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20765", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16227f54fe24ebe33674bbd0efe6b4c000432e52d3ca5015ba66cff1849a7dae.jpg", + "image_caption": [ + "Figure 2. The left illustrates our whole pipeline (2 views for simplicity). During the training phase, the network takes multi-view hand images and pseudo-labels as inputs. The bottom right depicts our cross-view interaction networks. The top right shows the view-shared graph feature extraction (VSGFE) module and view-shared feature (VSF) module. $\\oplus$ and $\\otimes$ denotes add and concatenation respectively." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 346, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51232d8a0153c6a748b5b783aab649a111933bd0e9ea6629fb5f51fe1e8ef874.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 88, + 658, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/093c5d08c4d48f11e75ee9a85cefff1ae12f311019caad37dab2ef703bdd0a23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 89, + 854, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb20bfda8d5d2cd0c90a2689039651ef6d202478817830e2b2021fd950055849.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 213, + 844, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ferent views $\\mathcal{H} = \\{\\pmb{H}_i^j\\}_{i=0}^v$ , 3D joints $\\mathcal{P} = \\{\\pmb{P}_i\\}_{i=0}^v$ , and MANO pose parameters $\\Theta = \\{\\pmb{\\theta}_i\\}_{i=0}^v$ from the single-view estimator to extract a graph feature $\\pmb{G}$ . The graph feature of each view $\\pmb{G}_i$ consists of three parts. $\\pmb{G}_i^1$ , $\\pmb{G}_i^2$ and $\\pmb{G}_i^3$ aim to capture joint location features, global image features, and local image features, respectively. The first part is joint location embedding $\\pmb{G}_i^1 \\in \\mathbb{R}^{k \\times c_1}$ , providing the explicit geometric information. This embedding is obtained by using an MLP to map the single-view 3D joints locations $\\pmb{P}_i$ and pose parameters $\\pmb{\\theta}_i$ to dimension $c_1$ . The second part is joint-wise high-level image features $\\pmb{G}_i^2 \\in \\mathbb{R}^{k \\times c_2}$ generated by spatial-aware initial graph building (SAIGB) [61] module using the last level feature maps $\\pmb{H}_i^4$ . This part provides compact image clues of all views for interaction. The third part is joint-aligned features $\\pmb{G}_i^3 \\in \\mathbb{R}^{k \\times c_3}$ gathered by joint feature sampler (JFS). JFS projects joints onto multi-level image feature maps $\\{\\pmb{H}_i^j\\}_{j=1}^3$ to gather fine-grained perceptual features like [53,54] for better local alignment. We then concatenate graph features to get $\\pmb{G}_i = [\\pmb{G}_i^1 \\otimes \\pmb{G}_i^2 \\otimes \\pmb{G}_i^3]$ .", + "bbox": [ + 75, + 417, + 472, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Dual-Branch Cross-View Interaction (DCVI)", + "text_level": 1, + "bbox": [ + 76, + 714, + 444, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first stack $\\{G_i\\}_{i=0}^v$ of all views to obtain multi-view graph feature $G \\in \\mathbb{R}^{vk \\times (c_1 + c_2 + c_3)}$ . We design a component to effectively capture complementary information from other views on multi-view graph feature $G$ . The interaction module has two branches, (1) cross-view attention branch (CVA) and (2) view-shared feature branch (VSF). Cross-view attention branch utilizes a cross-view transformer $F_t$ consisting of several multi-head attention layers with token size $vk$ and MLPs, which allows each joint to aggregate features from other joints or views. This branch implicitly captures the multi-view information. An explicit", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "multi-view prior information is that the observed poses from all the views should be consistent in 3D. Therefore, we add a branch to excavate the multi-view shared information to enhance the feature representation. Specifically, view-shared feature branch first employs adaptive-GCN [12] $F_{a}$ to map the view-specific features $\\pmb{G}_{i}$ to a canonical feature space $C_i = F_a(G_i)$ , the nodes in adaptive-GCN represents the hand joints and the edges represent joint feature correlation. Then, we stack $\\mathcal{C} = \\{C_i\\}_{i = 0}^v$ together to get multi-view canonical features $\\pmb{C} \\in \\mathbb{R}^{v \\times k \\times (c_1 + c_2 + c_3)}$ . After that, we use max-pooling on $\\pmb{C}$ to get the max activated features of every joint then repeat them in the view dimension as the view-shared features $\\pmb{C}' \\in \\mathbb{R}^{vk \\times (c_1 + c_2 + c_3)}$ . We denote the dual-branch cross-view interaction as: $\\pmb{G}^{*} = \\pmb{G} + F_{t}(\\pmb{G}) + \\pmb{C}^{\\prime}$ , where $\\pmb{G}^{*}$ is the updated graph feature.", + "bbox": [ + 496, + 417, + 893, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameters regression. The view specific feature $G_{i}^{*}$ after the interaction can be obtained by reshaping $G^{*}$ . We then employ a shared MLP $F_{r}$ as a decoder to regress the pose parameters $\\theta_{i}^{*} = F_{r}(G_{i}^{*})$ to derive the hand mesh of each view $M_{i}^{*}(\\theta_{i}^{*},\\beta_{i})$ and corresponding joints $P_{i}^{*} = JM_{i}^{*}$ .", + "bbox": [ + 496, + 645, + 893, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Multi-View Collaborative Learning", + "text_level": 1, + "bbox": [ + 500, + 726, + 802, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To allow all the views and the networks to learn collaboratively, we utilize consistency losses $L_{c}$ upon interaction outputs and distillation loss $L_{d}$ between multi-view fusion results and single-view outputs, as shown in Fig. 2. $L_{c}$ introduces collaborative learning between multiple views, guiding the poses from different views to be as close as possible. While $L_{d}$ makes the CVI-Net and single-view estimator work in a collaborative manner, achieving a self-distillation effect.", + "bbox": [ + 496, + 750, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Results fusion. Since we need to supervise the single-view", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20766", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimator with the results after the interaction, instead of simply using the refined results $M_{i}^{*}$ of each view, we ensemble all the results into a unified and more reliable result $\\tilde{M}$ . Considering the lack of explicit guidance, we empirically introduce a prior that all the views contribute equally. Thus, we simply average all aligned results to obtain $\\tilde{M}$ . Specifically, we use $A$ to denote the align procedure. When the extrinsics are known, we use the relative camera pose for alignment. When the camera extrinsics are unavailable, we use Procrustes analysis [62, 64] to compute relative rotation and align meshes to a canonical view. The final result is calculated as follows: $\\tilde{M} = \\frac{1}{v}\\sum_{i=1}^{v}A(M_{i}^{*})$ .", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consistency losses. We design two types of consistency loss $L_{c}$ : 2D consistency loss $L_{c_{2D}}$ and Fusion consistency loss $L_{c_f}$ . The motivation behind $L_{c_{2D}}$ is that the 2D predictions in the x-axis and y-axis are more accurate than the depth prediction in the z-axis. Therefore, $L_{c_{2D}}$ utilizes the 2D predictions in every single view as the pseudo label to supervise other views, which explores the view-specific reliable information to collaboratively improve the predictions of all the views. 2D consistency loss is defined as: $L_{c_{2D}} = \\frac{1}{v^2}\\sum_{i=1}^{v}\\sum_{j=1}^{v}\\|\\Pi(M_i^*) - \\Pi(A_i(M_j^*))\\|_1$ , where $A_i(\\cdot)$ denotes the alignment operation to align other view-j to view-i. Fusion consistency loss uses the fused results to supervise each view. The loss is defined as: $L_{c_f} = \\frac{1}{v}\\sum_{i=1}^{v}\\|M_i^* - A_i^{-1}(\\tilde{M})\\|_1$ , where $A_i^{-1}(\\cdot)$ denotes the inverse transformation from canonical view to view-i. $L_{c_{2D}}$ and $L_{c_f}$ are complementary to each other. Only using $L_{c_{2D}}$ tends to get performance saturation faster. In contrast, only adopting $L_{c_f}$ can lead to unstable training since there may exist the situation that the fusion results are worse due to the majority of the predictions being wrong, especially at the early training stage. During training, we alternately update $L_{c_{2D}}$ and $L_{c_f}$ to achieve more stable optimization.", + "bbox": [ + 75, + 273, + 472, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-view distillation loss. Since the multi-view fusion results are much better than the 2D pseudo label, we introduce multi-view distillation loss $L_{d} = \\frac{1}{v}\\sum_{i = 1}^{v}\\| M_{i} - A_{i}^{-1}(\\tilde{M})\\|_{1}$ that uses the fusion results to supervise the single-view outputs to achieve self-distillation.", + "bbox": [ + 76, + 609, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total loss. Except for the losses for multi-view collaborative learning, our framework also adopts two general constraints, 2D joints loss, and hand prior regularization. The prior regularization regularizes the pose and shape parameters: $L_{p} = \\frac{1}{v}\\sum_{i = 1}^{v}\\alpha (\\| \\pmb{\\theta}_{i}\\|_{1} + \\| \\pmb{\\theta}_{i}^{*}\\|_{1} + \\gamma \\| \\pmb{\\beta}_{i}\\|_{1})$ , where $\\alpha$ and $\\gamma$ are used to balance the loss scale. The 2D joints $L_{1}$ loss $L_{2D}$ is used to supervise the results from the 2D pseudo labels. The final loss is defined as: $L = L_{c} + L_{d} + L_{2D} + L_{p}$ .", + "bbox": [ + 75, + 688, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 827, + 209, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 848, + 274, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FreiHAND [64] is a dataset for single-view 3D hand pose estimation, which contains 130,240 training images and", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3,960 testing images. All images are captured from the real world with 3D annotations. The training set consists of 32,560 composited images with four types of real-world backgrounds and hands captured against a green screen.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HanCo [62] extends FreiHAND, which consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we use the first 1,200 sequences for training and the last 317 sequences for testing in all experiments for fair comparisons.", + "bbox": [ + 496, + 151, + 893, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Other datasets. We also provide additional results on other datasets. Assembly101 [46] is an action recognition dataset that consists of 4,321 videos sequence. H2O [29] is a handobject interaction dataset with 571,645 frames. Please refer to supplementary materials for details.", + "bbox": [ + 496, + 242, + 893, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We report standard metrics for hand pose estimation as follows. (1) MPJPE/MPVPE (mean per joint/vertex position error) measures the average Euclidean distance in mm between the predicted and ground-truth joints/vertices. JE/VE are the abbreviations for MPJPE/MPVPE. (2) NMPJPE/NMPVPE (normalized mean per joint/vertex position error, N-JE/VE) computes MPJPE/MPVPE after performing translation and scale alignment. (3)", + "bbox": [ + 496, + 318, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PA-MPJPE/PA-MPVPE (PA-JE/VE) is a modification of MPJPE/MPVPE with Procrustes analysis [17]. This metric normalizes the absolute scale, center, and rotation. (4) F-Score [9] is the harmonic mean of recall and precision between two meshes w.r.t. a specific distance threshold. F@5mm and F@15mm are reported. (5) AUC means the area under the curve of the PCK, where the PCK refers to the percentage of correct joints.", + "bbox": [ + 496, + 439, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 566, + 718, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement all the networks in PyTorch [40]. We first train our framework without $L_{c}$ and $L_{d}$ for 10 epochs. Then, we train the whole framework for another 30 epochs. Each batch contains images from 8 time-step of 8 cameras. We use AdamW [34] optimizer and set the initial learning rate to 3e-4. We use $256 \\times 256$ hand images as input. Please refer to supplementary materials for more details.", + "bbox": [ + 496, + 585, + 890, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparisons with state-of-the-arts", + "text_level": 1, + "bbox": [ + 498, + 699, + 803, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 4.3.1, we evaluate the performance of our method under the single-view inference setting. As self-supervised hand pose estimation is a relatively new task, there is limited literature available for comparison. To address this, we adapt self-supervised body pose estimation methods [27,52] to hand and compare them with our method on HanCo [62]. We then compare with the only existing self-supervised hand pose estimation method, S²Hand [9]. As S²Hand can only be trained on single-view images, we use our single-view network only (denote as Ours-SV) for both training and inference as baselines. We further conduct extensive evaluations of our full model and baselines to demonstrate", + "bbox": [ + 496, + 719, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20767", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/abb7593f0c1777dc43c7f595a21e6f126f60e3d13283a61b76b0a508b4571159.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputN-JE↓PA-JE↓
Fully-Supervised Method:
MobRecon [7]image9.95.7
EpipolarPose [27]image10.56.1
Self-Supervised Method:
EpipolarPose [27]image,19.79.3
CanonPose [52]2D pose,30.912.6
Oursimage,11.17.0
EpipolarPose [27]image42.323.5
CanonPose [52]2D pose31.812.8
Oursimage15.27.7
", + "bbox": [ + 86, + 88, + 429, + 263 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cfc41ba9e416c3d6addd677eddb55c184f98ec8d201fc0464c7cd39bc185e88e.jpg", + "table_caption": [ + "Table 1. Single-view inference comparisons on the HanCo [62] dataset. $\\bullet$ denotes the method using camera extrinsics during training. Notably, in the self-supervised setting, our method exhibits a significant improvement over previous methods." + ], + "table_footnote": [], + "table_body": "
MethodDataBackbonePA-JE↓PA-VE↓F@5↑
Fully-Supervised Method:
YoutubeHand [28]Frei.Res508.48.60.61
I2UV-HandNet [6]Frei.Res506.76.90.71
MobRecon [7]Frei.Res50†6.16.20.76
Ours-SVFrei.Res507.57.50.68
Self-Supervised Method:
S²HAND [9]Frei.EffiNet-b011.811.90.48
Ours-SVFrei.EffiNet-b011.611.70.49
Ours-SVFrei.Res5011.912.00.47
Ours-SVHanCoEffiNet-b011.311.40.51
Ours-SVHanCoRes5011.611.80.48
OursHanCoEffiNet-b06.36.80.71
OursHanCoRes506.26.70.72
", + "bbox": [ + 465, + 77, + 893, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the efficacy of multi-view collaborative learning.", + "bbox": [ + 76, + 343, + 398, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, thanks to our cross-view interaction network, our approach is capable of performing multi-view inference by simply averaging individual view results when multi-view test data is available. In Sec. 4.3.2, we compare our method with state-of-the-art approaches under the multi-view inference setting.", + "bbox": [ + 75, + 359, + 468, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Single-View Inference", + "text_level": 1, + "bbox": [ + 76, + 459, + 284, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hanco. We train EpipolarPose and CanonPose using their open-source code. We also train fully-supervised methods [7, 27] as a reference for performance. Tab. 1 outlines the performance of fully-/self-supervised methods in the lit-", + "bbox": [ + 75, + 478, + 468, + 539 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/395dd3cfa8328020a62332791c861dfaf427f3328415505110fb5d3f4898e464.jpg", + "table_caption": [ + "Table 2. Quantitative results on the FreiHAND evaluation set. The notation $\\dagger$ denotes using a stacked backbone structure. \"Our-SV\" refers to training only with our single-view network." + ], + "table_footnote": [], + "table_body": "
MethodMPJPE ↓PA-MJPJE ↓
Ttraditional Triangulation Method (w/o training):
DLT [21]16.813.2
Pictorial [11]13.510.2
RANSAC [25]12.39.8
Fully-Supervised Method:
EpipolarTrans [22]6.24.2
LT-Algebraic [25]5.53.6
LT-Volumetric [25]5.83.6
LT-Volumetric+ [25]4.93.6
EpipolarPose+ [27]8.04.4
Ours (Opt-Center)6.03.2
Ours (RANSAC)5.83.4
Self-Supervised Method:
EpipolarTrans [22]11.29.0
LT-Algebraic [25]10.37.8
LT-Volumetric [25]10.68.0
LT-Volumetric+ [25]9.57.2
CanonPose+ [52]21.610.5
EpipolarPose+ [27]17.28.3
Ours (Opt-Center)8.85.3
Ours (RANSAC)8.55.6
", + "bbox": [ + 125, + 549, + 421, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Multi-view inference results on the HanCo dataset. The notation ${}^{ + }$ indicates that methods require the GT 3D center.", + "bbox": [ + 75, + 867, + 468, + 896 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erature along with ours. In the case where camera extrinsics are available for training, CanonPose performs the worst because it lifts noisy 2D pseudo labels from OpenPose to 3D ones. When camera extrinsics are not available, all competitors experience a performance decline. This is due to the lack of collaborative interaction across multi-view features in previous self-supervised methods. In contrast, our method outperforms both of them by a large margin. Our cross-view interaction networks can enhance single-view inference, whether camera extrinsics are available during training or not. More details about the usage of cameras can be found in Sec. 3.2.3. Compared to previous self-supervised methods, our approach significantly improves performance, highlighting the importance of cross-view interaction among different views. Moreover, our approach can get comparable results to fully-supervised methods.", + "bbox": [ + 496, + 343, + 890, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FreiHAND. The comparisons on the evaluation set are shown in Tab. 2. The experiments conducted under self-supervised settings indicate that our baselines, Ours-SV, already achieve performance comparable to $\\mathbf{S}^2\\mathrm{Hand}$ . Moreover, directly equipping baselines with other backbones or more training data does not improve too much. We argue that performance improvements in single-view self-supervised hand pose estimation cannot be achieved by changing the backbone architecture or increasing the amount of training data. In contrast, our full model, i.e. Ours, substantially further improves the results on the Frei-Hand dataset, which justify the effectiveness of multi-view collaborative learning. Moreover, our self-supervised approach achieves competitive performance with recent fully-supervised state-of-the-art methods", + "bbox": [ + 496, + 585, + 892, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Multi-View Inference", + "text_level": 1, + "bbox": [ + 500, + 819, + 702, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the quantitative results of our multi-view inference performance with other competitors on HanCo in Tab. 3. A naive solution is to triangulate pseudo labels without training. We show the performance of traditional meth-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20768", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16d276a85d6f993b47e021215e6a794e18c7a822dc55f124444bf7d536918a5f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDMethodNMPJPE ↓PA-MPJPE ↓
SingleInteractFusionSingleInteractFusion
1FullResNet-50 as the backbone:
11.14↑0.038.31↓0.037.65↑0.107.05↑0.175.35↑0.075.34↑0.06
2FullResNet-18 as the backbone:
11.178.287.757.225.425.40
3-VSF11.21↓0.048.49↓0.217.81↓0.067.25↓0.035.52↓0.105.50↓0.10
4-CVA11.31↓0.148.45↓0.177.81↓0.067.29↓0.075.48↓0.065.46↓0.06
5-G111.31↓0.148.56↓0.287.77↓0.037.31↓0.095.52↓0.105.49↓0.09
6-G211.33↓0.168.38↓0.107.83↓0.087.34↓0.085.45↓0.035.42↓0.02
7-G311.30↓0.138.99↓0.697.82↓0.077.30↓0.085.45↓0.035.44↓0.04
8-Lc2D11.25↓0.088.43↓0.157.90↓0.157.32↓0.105.58↓0.165.57↓0.17
9-Lcf11.74↓0.578.98↓0.708.38↓0.637.55↓0.335.84↓0.425.80↓0.40
10-DCVI13.52↓2.35/11.99↓4.249.59↓2.37/9.42↓4.02
11-Lc14.04↓2.8717.03↓8.7510.32↓2.579.04↓1.8210.21↓4.797.92↓2.52
12-Ld17.05↓5.888.56↓0.288.01↓0.2610.13↓2.915.67↓0.255.65↓0.25
", + "bbox": [ + 78, + 89, + 589, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Quantitative ablation studies. We remove each of our components here to show their contribution to our framework. Full denotes our complete model. CVI represents our whole cross-view interaction network. Other notations are consistent with Sec. 3. We report the errors of single-view outputs (Single, $M$ ), cross-view interaction outputs (Interact, $M^{*}$ ), and multi-view fusion results (Fusion, $\\tilde{M}$ ).", + "bbox": [ + 71, + 335, + 589, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods. Such methods can serve as a reference for evaluating the effectiveness of self-supervised methods. We adapt fully-supervised multi-view 3D pose estimation methods LT [25] and EpipolarTrans [22] to a self-supervised manner. Under self-supervised settings, EpipolarTrans can only achieve limited performance improvements compared to traditional methods. LT-Algebraic [25], which incorporates learnable confidence into the triangulation. LT-Volumetric model [25], which unprojects 2D features into a 3D volume for inference, achieves better results, but the performance is dependent on the accuracy of the hand center. Canon-Pose [52] and EpipolarPose [27] obtain multi-view inference results through simple averaging like ours.", + "bbox": [ + 75, + 422, + 472, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, both of these methods are inferior to ours because they lack cross-view interaction. As our method predicts the root-relative 3D pose, we need to conduct post-processing to obtain the absolute coordinates. We introduce two different ways to achieve this: 1) using the 2D predictions of different views to triangulate and refine a center and 2) conducting RANSAC triangulation using our 2D predictions. Both methods have their merits. Opt-center can keep the root-relative results with hand prior, resulting in low PA-MPJPE. RANSAC gets better joint-wise accuracy, which is indicated by low MPJPE. We also provide qualitative results in the supplementary materials on the Assembly101 [46] dataset, which has a static camera setup. Even for challenging head-mounted moving cameras, we achieve convincing 3D pose estimates on the H2O [29] dataset. The experiments show that we have significantly pushed the performance of self-supervised methods to a comparable level with fully supervised methods.", + "bbox": [ + 75, + 630, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26046fcf4f4bcecb6c44fde775f429b259be0d31b838736a46707e4fda84821c.jpg", + "image_caption": [ + "Figure 3. Error of using different (a) #training data, (b) (line-1)#view for training, and (line-2)#view for inference when trained with 8 views." + ], + "image_footnote": [], + "bbox": [ + 596, + 88, + 901, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/449db25627e75bc0694bfe003d835e8df4b1a72a287df96bdbec5ca3fc9172b6.jpg", + "image_caption": [ + "Figure 4. AUC of three 2D joint sets. O, S, I, PE denote OpenPose, single-view, interaction, and average pixel error in resolution $256 \\times 256$ ." + ], + "image_footnote": [], + "bbox": [ + 596, + 246, + 898, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Result", + "text_level": 1, + "bbox": [ + 500, + 422, + 674, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 presents the visual comparisons of 2 views between 2D joints of OpenPose, ours, and ground-truth on the HanCo dataset. We can observe that our method is more robust for outliers and can generate predictions close to the labels. Fig. 6 shows the 3D predictions from two viewpoints of ours, EpipolarPose, and CanonPose on the HanCo dataset. The results indicate that our method can get more accurate results especially when the occlusions are severe. Please refer to supplementary materials for more results.", + "bbox": [ + 496, + 443, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 588, + 653, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 4, we conduct comprehensive ablation experiments on the HanCo [62] dataset to show the effectiveness of each component. Single, Interact and Fusion denotes the evaluation of $M$ , $M^{*}$ and $\\tilde{M}$ respectively.", + "bbox": [ + 496, + 609, + 890, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different backbones. We first show our performance with different backbones. As shown in #1 and #2, using a large backbone like Res50, our performance can be further improved. For efficiency, we conduct ablation studies using Res18 as the backbone unless otherwise specified.", + "bbox": [ + 496, + 670, + 890, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two branches for cross-view interaction module. As presented in #3 and #4, both of the branches can reduce the error. VSF can explicitly model the view-shared information and add reliable information from every view. CVA can capture the self-/cross-view joint-level correlations.", + "bbox": [ + 496, + 747, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Graph features. The results indicate that three kinds of features (#5, #6, #7) all lead to performance improvement. Especially, local feature (#7, $G^3$ ) can notably reduce the error after the interaction by providing fine-grained details. DCVI. We also conduct experiments to show the impor", + "bbox": [ + 496, + 823, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20769", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tance of DCVI by removing it and posing consistency constraints in single-view outputs like [24]. In this way, the performance drops dramatically (#10), proving the necessity of using DCVI to capture the features of all the views for self-supervised learning.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Two branches for multi-view consistency loss. Without enforcing cross-view interaction outputs to be consistent, the performance significantly drops (#9). If we do not explore relatively more reliable 2D predictions to enhance consistency, the performance can also get worse (#8).", + "bbox": [ + 75, + 167, + 467, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consistency losses. $(L_{c})$ The performance is unsatisfactory (#11) when employing the cross-view interaction network without any consistency constraints (i.e. discard #8 and #9). The interaction network should cooperate with consistency so that the constraints can guide the network to exploit multi-view information to function better.", + "bbox": [ + 75, + 242, + 467, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-view distillation loss. $(L_{d})$ Removing the multiview distillation loss, all the metrics drop by a large margin (#12), especially in single-view estimation accuracy. This phenomenon proves the effectiveness of collaborative learning between single- and multi-view networks.", + "bbox": [ + 75, + 332, + 467, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Model Analysis", + "text_level": 1, + "bbox": [ + 76, + 414, + 232, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different percentage of unlabeled images. Fig. 3 (a) shows our method can get consistent performance improvement as the unlabeled training data increases.", + "bbox": [ + 75, + 433, + 467, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different view number for training. The line-1 in Fig. 3 (b) shows the performance of our method tested on a certain view when trained with different view numbers. The curve shows that our method can be consistently improved as the number of views increases. We also observe that using multiple views for training can significantly improve performance when the valid views are few.", + "bbox": [ + 75, + 479, + 467, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different view number for inference. Our model allows inferring with an arbitrary number of views. However, when the model is trained with a fixed view number, it could get the view number bias, resulting in better performance using the view number close to the training one. To avoid this, we add random masks in our interaction module and finetune the model for a few epochs. After that, results can get better by a small margin (the single-view error is $11.07\\mathrm{mm}$ and the fusion error $7.60\\mathrm{mm}$ , both in NMPJPE.). The line-2 in Fig. 3 (b) shows results on a certain view when trained on 8 views and tested on 1 to 8 views. We can observe consistent improvement with the inference view number increases.", + "bbox": [ + 75, + 584, + 467, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different 2D joint sets. Fig. 4 presents the accuracy of different 2D joint sets on the HanCo training and testing set. Our 2D predictions are extremely better than OpenPose 2D pseudo label used for training.", + "bbox": [ + 75, + 780, + 467, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iteratively training. Our approach can use the previous predictions as pseudo labels for iterative training. We find it helpful till iteration 3 and get saturated afterward. From 1 to 3 iterations, NMPJPE is 7.75, 7.68, and 7.64.", + "bbox": [ + 75, + 840, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/151d8d96736515d6e29262c0f139423473de557de666646edd5d7ad948251c85.jpg", + "image_caption": [ + "Figure 5. 2D prediction (overlayed in the images) comparisons between OpenPose, ours, and ground-truth on the HanCo dataset." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 870, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8c92cc4e139cbe7c3219b5eaa6809fbb355cf757c3bb4b080cf7876b7e35f8c.jpg", + "image_caption": [ + "Figure 6. 3D prediction comparisons between our method, EpipolarPose, and CanonPose on the HanCo dataset. Our prediction and ground-truth are shown in solid red and dashed green respectively." + ], + "image_footnote": [], + "bbox": [ + 519, + 232, + 869, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 393, + 767, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To our best knowledge, we present the first self-supervised framework that aims to learn a single-view 3D hand estimator from unlabeled multi-view data. At the core of our approach, a cross-view interaction network is carefully designed to supervise the single-view output by leveraging the collaboration among multi-views. Specifically, the network captures the interdependencies of features among different views, resulting in improved accuracy of hand pose estimation after cross-view interaction. Additionally, the multi-view results are fused to supervise the single-view output for self-distillation. The effectiveness and versatility of the proposed framework are extensively evaluated through experiments, which demonstrate that our method not only establishes a new benchmark for self-supervised 3D hand pose estimation from single-view input but also offers flexible multi-view inference with state-of-the-art performance.", + "bbox": [ + 496, + 416, + 890, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We focused on hand pose estimation without heavy occlusions in this work. Extending our work to more challenging scenarios, such as hand-object interaction or relaxing the synchronization constraints in multi-view inputs, would be interesting topics for further study.", + "bbox": [ + 496, + 672, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 758, + 658, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grants (62071067, 62171057, 62201072), in part by the Ministry of Education and China Mobile Joint Fund (MCM20200202), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039).", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20770", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2, 3", + "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2", + "[3] Kristijan Bartol, David Bojanic, Tomislav Petković, and Tomislav Pribanić. Generalizable human pose triangulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11028-11037, 2022. 3", + "[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 2, 3", + "[5] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 2, 3", + "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 2, 6", + "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 2, 3, 6", + "[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 2, 3", + "[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2, 3, 5, 6", + "[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2", + "[11] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 7792-7801, 2019. 6", + "[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 2, 4", + "[13] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 2021 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2021. 2", + "[14] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 2", + "[15] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1991-2000, 2017. 2", + "[16] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 2", + "[17] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 5", + "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2", + "[19] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Trans. Graph., 39(4):87, 2020. 1", + "[20] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, Cabezas Randi, Tran Luan, Akbay Muzaffer, Yu Tsz-Ho, Keskin Cem, and Wang Robert. Umetrack: Unified multi-view end-to-end hand tracking for vr. ACM Transactions on Graphics, 2022. 1", + "[21] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 6", + "[22] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020. 3, 6, 7", + "[23] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European Conference on Computer Vision (ECCV), pages 118-134, 2018. 2", + "[24] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20771", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5243-5252, 2020. 1, 3, 8", + "[25] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2019. 3, 6, 7", + "[26] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, 2014. 2", + "[27] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1077-1086, 2019. 1, 2, 3, 5, 6, 7", + "[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990–5000, 2020. 2, 6", + "[29] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 5, 7", + "[30] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 2", + "[31] Moran Li, Yuan Gao, and Nong Sang. Exploiting learnable joint groups for hand pose estimation. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1921-1929, 2021. 2", + "[32] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 2", + "[33] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 2", + "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, 2019. 5", + "[35] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In European Conference on Computer Vision, pages 424-442. Springer, 2022. 3", + "[36] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 2", + "[37] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 2", + "[38] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2", + "[39] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 49-59, 2018. 2, 3", + "[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in python. In NIPS 2017 Workshop on Autodiff, 2017. 5", + "[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017. 3", + "[42] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4342-4351, 2019. 3", + "[43] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6040-6049, 2020. 3", + "[44] Helge Rhodin, Jörg Spörri, Isinsu Katircioglu, Victor Constantin, Frédéric Meyer, Erich Müller, Mathieu Salzmann, and Pascal Fua. Learning monocular 3d human pose estimation from multi-view images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8437-8446, 2018. 3", + "[45] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2, 3", + "[46] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 5, 7", + "[47] Hui Shuai, Lele Wu, and Qingshan Liu. Adaptive multi-view and temporal fusing transformer for 3d human pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3", + "[48] Adrian Spurr, Umar Iqbal, Pavlo Molchanov, Otmar Hilliges, and Jan Kautz. Weakly supervised 3d hand pose estimation via biomechanical constraints. In European Conference on Computer Vision, pages 211-228. Springer, 2020. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20772", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2", + "[50] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2", + "[51] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 3", + "[52] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13294-13304, 2021. 1, 2, 3, 5, 6, 7", + "[53] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 4", + "[54] Chao Wen, Yinda Zhang, Zhuwen Li, and Yanwei Fu. Pixel2mesh++: Multi-view 3d mesh generation via deformation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1042-1051, 2019. 4", + "[55] Linlin Yang, Shicheng Chen, and Angela Yao. Semihand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2, 3", + "[56] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2335-2343, 2019. 2", + "[57] Linlin Yang and Angela Yao. Disentangling latent hands for image synthesis and pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9877-9886, 2019. 2", + "[58] Xiong Zhang, Hongsheng Huang, Jianchao Tan, Hongmin Xu, Cheng Yang, Guozhu Peng, Lei Wang, and Ji Liu. Hand image understanding via deep multi-task learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11281-11292, 2021. 2", + "[59] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2, 3", + "[60] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129(3):703-718, 2021. 3", + "[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In DAGM German Conference on Pattern Recognition, pages 250-264. Springer, 2021. 1, 2, 5, 6, 7", + "[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2, 3", + "[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2, 5" + ], + "bbox": [ + 501, + 92, + 893, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20773", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_model.json b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa21bca9ea59362839293c616b75cb6f5a114b2a --- /dev/null +++ b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.13, + 0.797, + 0.177 + ], + "angle": 0, + "content": "HaMuCo: Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.209, + 0.877, + 0.284 + ], + "angle": 0, + "content": "Xiaozheng Zheng \\(^{1,2\\dagger}\\) Chao Wen \\(^{2\\dagger}\\) Zhou Xue \\(^{2}\\) Pengfei Ren \\(^{1,2}\\) Jingyu Wang \\(^{1*}\\) \n\\(^{1}\\) State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications \n\\(^{2}\\) PICO IDL, ByteDance, Beijing" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.317, + 0.314, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Recent advancements in 3D hand pose estimation have shown promising results, but its effectiveness has primarily relied on the availability of large-scale annotated datasets, the creation of which is a laborious and costly process. To alleviate the label-hungry limitation, we propose a self-supervised learning framework, HaMuCo, that learns a single-view hand pose estimator from multi-view pseudo 2D labels. However, one of the main challenges of self-supervised learning is the presence of noisy labels and the \"groupthink\" effect from multiple views. To overcome these issues, we introduce a cross-view interaction network that distills the single-view estimator by utilizing the cross-view correlated features and enforcing multi-view consistency to achieve collaborative learning. Both the single-view estimator and the cross-view interaction network are trained jointly in an end-to-end manner. Extensive experiments show that our method can achieve state-of-the-art performance on multi-view self-supervised hand pose estimation. Furthermore, the proposed cross-view interaction network can also be applied to hand pose estimation from multi-view input and outperforms previous methods under the same settings." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.692, + 0.21, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.85 + ], + "angle": 0, + "content": "3D hand pose estimation is essential in various application scenarios, from action recognition and sign language translation to AR/VR [19, 20]. Hand pose estimation has achieved a significant improvement in recent years. However, the progress heavily relies on the emergence of many hand pose datasets with accurate 3D annotations. Acquiring labeled datasets is quite time-consuming and laborious, exposing a realistic challenge for deep learning models to learn with limited and noisy data." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.85, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Self-supervised learning is an emerging solution to the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.318, + 0.893, + 0.501 + ], + "angle": 0, + "content": "challenge posed by manual annotation. Though worth exploring, self-supervised pose estimation with RGB hand images is a challenging and relatively unexplored area with only one pioneering method, S²HAND [9]. S²HAND aims to conduct 3D hand reconstruction from a single RGB image with the noisy off-the-shell 2D hand pose estimation results for supervision. Unfortunately, S²HAND faces a predicament where its performance is significantly reliant on the quality of the pseudo label, and inferior labeling may result in reduced performance. Moreover, evaluating the quality of the pseudo label is an ill-posed problem that lacks clear criteria or input, further complicating the issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.893, + 0.655 + ], + "angle": 0, + "content": "This observation motivates us to leverage multi-view information for enhancing self-supervised learning, as the complementary nature of multi-view observations can help mitigate the ambiguity inherent in pose estimation. Although the first 3D hand dataset with synchronized multiview input (HanCo [62]) was proposed in 2021, to our knowledge, there is no previous work exploring the potential of multi-view for self-supervised hand pose estimation. Therefore, we turn to studies in the human body pose estimation, which share some similarities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.793 + ], + "angle": 0, + "content": "As mentioned in previous work [24], naively enforcing multi-view consistency is prone to generate degenerated solutions, thus they resorted to additional 2D labels of unrelated datasets and proposed a solution under the scope of weakly supervised learning. Other studies, such as EpipolarPose [27] and CanonPose [52], utilized multiview data with special designs to enhance the supervision and achieved promising results under the scope of self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we push along this direction on hand pose estimation via multi-view collaborative learning. We take one step further by designing a learnable network, which utilizes multi-view information, to tackle 1) noisy pseudo labels and 2) unreliable multi-view \"groupthink\" issues causing training collapse in the early training stage. Formally, we name the pipeline HaMuCo, which stands for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.874, + 0.444, + 0.888 + ], + "angle": 0, + "content": "Project page: https://zxz267.github.io/HaMuCo." + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.888, + 0.342, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author. † Equal contribution." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.874, + 0.444, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20763" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.074, + 0.28, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.072, + 0.501, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.074, + 0.878, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.277, + 0.895, + 0.319 + ], + "angle": 0, + "content": "Figure 1. Overall pipeline comparison: HaMuCo learns a monocular 3D hand pose estimator from multi-view self-supervision via cross-view feature interaction. Our cross-view interaction network addresses the importance of introducing learnable feature interaction, which is absent from previous methods [27, 52]. At inference time only the gray part is applied." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.324, + 0.343, + 0.339 + ], + "angle": 0, + "content": "Hand Multiview Collaborative learning." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.34, + 0.47, + 0.534 + ], + "angle": 0, + "content": "The core idea of our approach is to enhance the single-view estimation by means of cross-view feature interaction and further integrate multi-view results to supervise the single-view output to achieve self-distillation in an end-to-end fashion. Thus, our framework is built with a single-view hand pose estimator and a cross-view interaction network for supervision. The single-view estimator uses the MANO [45] hand model as the decoder, which provides the hand prior to regularizing irrational anatomy when supervised by noisy pseudo labels. The cross-view interaction network captures cross-view features and utilizes several consistent losses among different views to guide collaborative learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.471, + 0.716 + ], + "angle": 0, + "content": "We conduct comprehensive experiments on the HanCo [62] dataset and our approach outperforms previous methods by a considerable margin for self-supervised 3D hand pose estimation. Notably, our results demonstrate competitive performance compared to a state-of-the-art fully supervised approach proposed by Chen et al. [7]. Our proposed framework is highly versatile, as it can be trained with or without calibration, and is capable of incorporating the cross-view interaction network to achieve superior multi-view inference results when multi-view test data is available. Moreover, our model can generalize well to other datasets [29, 46, 64] and in-the-wild images." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.717, + 0.419, + 0.732 + ], + "angle": 0, + "content": "In summary, our contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.738, + 0.47, + 0.8 + ], + "angle": 0, + "content": "- We propose the first self-supervised learning framework for single-view hand pose estimation without any training data annotation and achieve state-of-the-art performance via multi-view collaborative learning." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.805, + 0.471, + 0.866 + ], + "angle": 0, + "content": "- We propose a cross-view interaction network to supervise the single-view estimator by enforcing multi-view consistency and capturing cross-view features for collaborative learning among multiple views." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- The proposed framework is capable of multi-view inference by incorporating the cross-view interac" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.738, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.324, + 0.892, + 0.353 + ], + "angle": 0, + "content": "tion network and achieves state-of-the-art performance without bells and whistles." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.365, + 0.643, + 0.381 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Hand Pose Estimation. Hand pose estimation can be categorized into RGB-based methods [23, 49, 63] and depth-based methods [14, 15, 36], depending on the input modality. In this paper, we focus our attention on RGB-based hand pose estimation. The RGB-based methods can be further divided into three categories, skeleton-based methods [5, 12, 23, 31, 38, 39, 48, 49, 55-57, 63], model-based methods [1, 2, 4, 9, 58, 59, 64], and mesh-based methods [7,8,10,16,28,30,32,33,37,50,61]. Skeleton-based methods regress the hand joints directly. Zimmermann et al. [63] introduces a multi-stage network that lifts the regressed 2D joints to 3D ones. Variational autoencoder [26] is employed to learn a cross-modal latent space to achieve better hand pose estimation and disentanglement [49, 56, 57]. Latent 2.5D representation regression is proved more effective than direct coordinates regression for hands by [23], which is also adopted by [13, 31, 48, 61]. There are also many works solving hand pose estimation with two hands interactions [13, 29, 38] and hand-object interactions [2, 12, 29]. Recent model-based methods make use of MANO [45], which can incorporate the hand prior and predict the hand mesh simultaneously. Those methods [1, 4, 9, 58, 59] rely on additional supervisions [1, 4, 9, 58, 59] or inputs [4]. In contrast, mesh-based methods regress each vertex directly, which is more accurate but requires large-scale datasets with hand mesh annotations [18, 29, 38, 64]. Most of these methods utilize graph convolutional network (GCN) [7,8,10,16,28,50,61] or transformers [32] or both [30, 33] for regression. I2L-MeshNet [37] regresses each vertex by predicting 1D heatmaps of three axes. Chen et al. [6] uses an image-to-image translation network to predict the UV map of the mesh. Similar to previous works [30, 33], we also use transformer and GCN. However, we employ them for cross-view interaction." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20764" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.242 + ], + "angle": 0, + "content": "Multi-View Fully-Supervised Pose Estimation. Multiview information is widely explored to improve 3D human pose estimation by tackling occlusions and depth-ambiguity in a fully-supervised manner [3, 22, 25, 41-43, 47, 60]. Volume-based methods [25, 41, 42, 51] unproject 2D features or heatmaps of joints to a 3D space for estimation. Another kind of method [22, 43, 60] utilizes the geometry information to fuse the features in 2D space directly and efficiently. Recently, some works [35,47] utilize transformers for implicit cross-view fusion without camera extrinsics." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.244, + 0.468, + 0.47 + ], + "angle": 0, + "content": "Label-Efficient Learning. Label-efficient learning aims to reduce the 3D label requirements. Many works devote to solving hand pose estimation in a label-efficient manner [1, 4,5,7,9,39,48,55,59,63]. Synthetic data is used to avoid manual annotation [7,39,63], but may need domain transfer [39]. Or use weakly supervised learning [4,48] to obtain 3D results by manually annotating 2D labels to assist with hand priors. Multi-view label-efficient learning is also explored in 3D pose estimation [24, 27, 44, 52]. Rhodin et al. [44] trains a semi-supervised network with only a small amount of labeled 3D data and multi-view consistency constraints. Iqbal et al. [24] mixes single-view images with 2D labels and unlabelled multi-view images for training. Our goal is the same as that of previous methods, which is to train without any manual 3D labels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.472, + 0.468, + 0.804 + ], + "angle": 0, + "content": "Self-supervised 3D Pose Estimation. (1) Single-view training and inference. To the best of our knowledge, there is only one method for self-supervised 3D hand pose estimation, proposed by Chen et al. [9]. Their framework, \\(\\mathrm{S}^2\\mathrm{Hand}\\), uses only single-view 2D noisy labels for training and achieves self-supervision through rendering. However, the performance is limited due to the use of single-view information and the quality of the noisy labels. (2) Multi-view training, single-view inference. Our approach belongs to this category but is fundamentally different from the existing methods. EpipolarPose [27] triangulates multi-view 2D pseudo labels according to epipolar geometry to 3D ones for training. CanonPose [52] learns to lift 2D pseudo labels to 3D canonical pose space with multi-view consistency constraints. All the aforementioned methods use nonlearnable self-supervised modules like geometric modules or consistency loss functions, as shown in Fig. 1. However, they [27, 52] ignore the importance of introducing cross-view interaction and multi-view collaborative learning. Previous methods struggle to achieve good performance since the pose of a hand can change drastically over time and different joints may have similar appearances." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.818, + 0.168, + 0.833 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As depicted in the left part of Fig. 2, our framework consists of a simple yet effective single-view estimator and cross-view interaction network. The core idea of our approach is that prediction from a monocular view can be" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "enhanced via cross-view feature interaction and the interacted results can further supervise the single-view output to achieve self-distillation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.143, + 0.709, + 0.159 + ], + "angle": 0, + "content": "3.1. Single-View Estimator" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.298 + ], + "angle": 0, + "content": "Overview. Our framework takes multi-view synchronized hand images \\(\\mathcal{I} = \\{I_i\\}_{i=0}^v\\) with \\(v\\) views as input, each view is an image of \\(I_i \\in \\mathbb{R}^{3 \\times h \\times w}\\). The output is a 3D hand mesh \\(M\\) on each view. We designed a simple yet effective model-based network as a single-view estimator. Using the hand model will reduce the adverse effects of using poor pseudo labels as supervision by providing hand prior information for regularization. Please refer to supplementary materials for more details about the single-view estimator." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.299, + 0.892, + 0.419 + ], + "angle": 0, + "content": "Hand model. We employ MANO [45] as the hand model. The hand mesh can be derived from the MANO layer using parameters \\(\\beta\\) and \\(\\pmb{\\theta}\\), i.e. \\(M(\\beta, \\pmb{\\theta})\\). \\(\\beta \\in \\mathbb{R}^{10}\\) and \\(\\pmb{\\theta} \\in \\mathbb{R}^{16 \\times 3}\\) control the shape and pose of the hand respectively. We can use a predefined regressor to obtain the 3D joints from the 3D mesh vertices by \\(P = JM\\), where \\(J \\in \\mathbb{R}^{k \\times n}\\), where \\(n = 778\\) and \\(k = 21\\) are the joints number and vertices number. For more details, we recommend referring to [45]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.42, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Camera model. Following Boukhayma et al. [4], we model the geometry correspondence by the weak-perspective camera model and obtain camera parameters from the single-view network predictions. Given the translation \\( \\pmb{t} \\) and scale \\( s \\), the 2D coordinates in image plane can be obtained by: \\( \\Pi(P) = s\\Omega(P) + \\mathbf{t} \\), where \\( \\Omega \\) is the orthographic projection and \\( \\Pi \\) denotes the weak-perspective projection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.661 + ], + "angle": 0, + "content": "Network Structure. Since the single-view estimator is not the main component, for the sake of simplicity, we employ a CNN as the encoder \\( F_{e} \\), and an MLP as the decoder \\( F_{d} \\) for regressing the MANO parameters. We have 3D hand mesh: \\( M_{i}(\\pmb{\\theta}_{i},\\pmb{\\beta}_{i}) = F_{s}(\\pmb{I}_{i}) \\), where \\( F_{s} = F_{d}(F_{e}(\\cdot)) \\) denotes the entire single-view network. The estimator also passes different levels of features \\( H^{j} \\) (where \\( H^{j} \\) is the intermediate feature of the encoder after \\( j \\) residual blocks, \\( j = 1,2,3,4 \\)) to our cross-view interaction network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.782, + 0.687 + ], + "angle": 0, + "content": "3.2. Cross-view Interaction Network" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.695, + 0.892, + 0.801 + ], + "angle": 0, + "content": "In this section, we introduce the cross-view interaction network (CVI-Net), which is the core of our system to enable the network to exploit multi-view information. This stage conducts cross-view interaction and distillation. The critical components of this stage are a cross-view interaction network for capturing cross-view features and several consistent losses for guiding collaborative learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.83, + 0.822 + ], + "angle": 0, + "content": "3.2.1 View-Shared Graph Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The first step for interaction is to extract the appropriate features. Different from [7, 8, 61], our module collects useful information into a graph through view-shared graph feature extraction module (VSGFE) as shown in Fig. 2. Specifically, it makes use of multi-level feature maps from dif" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20765" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.348, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.089, + 0.66, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.09, + 0.855, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.214, + 0.846, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 2. The left illustrates our whole pipeline (2 views for simplicity). During the training phase, the network takes multi-view hand images and pseudo-labels as inputs. The bottom right depicts our cross-view interaction networks. The top right shows the view-shared graph feature extraction (VSGFE) module and view-shared feature (VSF) module. \\(\\oplus\\) and \\(\\otimes\\) denotes add and concatenation respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.473, + 0.709 + ], + "angle": 0, + "content": "ferent views \\(\\mathcal{H} = \\{\\pmb{H}_i^j\\}_{i=0}^v\\), 3D joints \\(\\mathcal{P} = \\{\\pmb{P}_i\\}_{i=0}^v\\), and MANO pose parameters \\(\\Theta = \\{\\pmb{\\theta}_i\\}_{i=0}^v\\) from the single-view estimator to extract a graph feature \\(\\pmb{G}\\). The graph feature of each view \\(\\pmb{G}_i\\) consists of three parts. \\(\\pmb{G}_i^1\\), \\(\\pmb{G}_i^2\\) and \\(\\pmb{G}_i^3\\) aim to capture joint location features, global image features, and local image features, respectively. The first part is joint location embedding \\(\\pmb{G}_i^1 \\in \\mathbb{R}^{k \\times c_1}\\), providing the explicit geometric information. This embedding is obtained by using an MLP to map the single-view 3D joints locations \\(\\pmb{P}_i\\) and pose parameters \\(\\pmb{\\theta}_i\\) to dimension \\(c_1\\). The second part is joint-wise high-level image features \\(\\pmb{G}_i^2 \\in \\mathbb{R}^{k \\times c_2}\\) generated by spatial-aware initial graph building (SAIGB) [61] module using the last level feature maps \\(\\pmb{H}_i^4\\). This part provides compact image clues of all views for interaction. The third part is joint-aligned features \\(\\pmb{G}_i^3 \\in \\mathbb{R}^{k \\times c_3}\\) gathered by joint feature sampler (JFS). JFS projects joints onto multi-level image feature maps \\(\\{\\pmb{H}_i^j\\}_{j=1}^3\\) to gather fine-grained perceptual features like [53,54] for better local alignment. We then concatenate graph features to get \\(\\pmb{G}_i = [\\pmb{G}_i^1 \\otimes \\pmb{G}_i^2 \\otimes \\pmb{G}_i^3]\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.715, + 0.446, + 0.73 + ], + "angle": 0, + "content": "3.2.2 Dual-Branch Cross-View Interaction (DCVI)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We first stack \\(\\{G_i\\}_{i=0}^v\\) of all views to obtain multi-view graph feature \\(G \\in \\mathbb{R}^{vk \\times (c_1 + c_2 + c_3)}\\). We design a component to effectively capture complementary information from other views on multi-view graph feature \\(G\\). The interaction module has two branches, (1) cross-view attention branch (CVA) and (2) view-shared feature branch (VSF). Cross-view attention branch utilizes a cross-view transformer \\(F_t\\) consisting of several multi-head attention layers with token size \\(vk\\) and MLPs, which allows each joint to aggregate features from other joints or views. This branch implicitly captures the multi-view information. An explicit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.646 + ], + "angle": 0, + "content": "multi-view prior information is that the observed poses from all the views should be consistent in 3D. Therefore, we add a branch to excavate the multi-view shared information to enhance the feature representation. Specifically, view-shared feature branch first employs adaptive-GCN [12] \\(F_{a}\\) to map the view-specific features \\(\\pmb{G}_{i}\\) to a canonical feature space \\(C_i = F_a(G_i)\\), the nodes in adaptive-GCN represents the hand joints and the edges represent joint feature correlation. Then, we stack \\(\\mathcal{C} = \\{C_i\\}_{i = 0}^v\\) together to get multi-view canonical features \\(\\pmb{C} \\in \\mathbb{R}^{v \\times k \\times (c_1 + c_2 + c_3)}\\). After that, we use max-pooling on \\(\\pmb{C}\\) to get the max activated features of every joint then repeat them in the view dimension as the view-shared features \\(\\pmb{C}' \\in \\mathbb{R}^{vk \\times (c_1 + c_2 + c_3)}\\). We denote the dual-branch cross-view interaction as: \\(\\pmb{G}^{*} = \\pmb{G} + F_{t}(\\pmb{G}) + \\pmb{C}^{\\prime}\\), where \\(\\pmb{G}^{*}\\) is the updated graph feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.895, + 0.723 + ], + "angle": 0, + "content": "Parameters regression. The view specific feature \\( G_{i}^{*} \\) after the interaction can be obtained by reshaping \\( G^{*} \\). We then employ a shared MLP \\( F_{r} \\) as a decoder to regress the pose parameters \\( \\theta_{i}^{*} = F_{r}(G_{i}^{*}) \\) to derive the hand mesh of each view \\( M_{i}^{*}(\\theta_{i}^{*},\\beta_{i}) \\) and corresponding joints \\( P_{i}^{*} = JM_{i}^{*} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.803, + 0.743 + ], + "angle": 0, + "content": "3.2.3 Multi-View Collaborative Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.884 + ], + "angle": 0, + "content": "To allow all the views and the networks to learn collaboratively, we utilize consistency losses \\( L_{c} \\) upon interaction outputs and distillation loss \\( L_{d} \\) between multi-view fusion results and single-view outputs, as shown in Fig. 2. \\( L_{c} \\) introduces collaborative learning between multiple views, guiding the poses from different views to be as close as possible. While \\( L_{d} \\) makes the CVI-Net and single-view estimator work in a collaborative manner, achieving a self-distillation effect." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Results fusion. Since we need to supervise the single-view" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20766" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "estimator with the results after the interaction, instead of simply using the refined results \\( M_{i}^{*} \\) of each view, we ensemble all the results into a unified and more reliable result \\( \\tilde{M} \\). Considering the lack of explicit guidance, we empirically introduce a prior that all the views contribute equally. Thus, we simply average all aligned results to obtain \\( \\tilde{M} \\). Specifically, we use \\( A \\) to denote the align procedure. When the extrinsics are known, we use the relative camera pose for alignment. When the camera extrinsics are unavailable, we use Procrustes analysis [62, 64] to compute relative rotation and align meshes to a canonical view. The final result is calculated as follows: \\( \\tilde{M} = \\frac{1}{v}\\sum_{i=1}^{v}A(M_{i}^{*}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.61 + ], + "angle": 0, + "content": "Consistency losses. We design two types of consistency loss \\( L_{c} \\): 2D consistency loss \\( L_{c_{2D}} \\) and Fusion consistency loss \\( L_{c_f} \\). The motivation behind \\( L_{c_{2D}} \\) is that the 2D predictions in the x-axis and y-axis are more accurate than the depth prediction in the z-axis. Therefore, \\( L_{c_{2D}} \\) utilizes the 2D predictions in every single view as the pseudo label to supervise other views, which explores the view-specific reliable information to collaboratively improve the predictions of all the views. 2D consistency loss is defined as: \\( L_{c_{2D}} = \\frac{1}{v^2}\\sum_{i=1}^{v}\\sum_{j=1}^{v}\\|\\Pi(M_i^*) - \\Pi(A_i(M_j^*))\\|_1 \\), where \\( A_i(\\cdot) \\) denotes the alignment operation to align other view-j to view-i. Fusion consistency loss uses the fused results to supervise each view. The loss is defined as: \\( L_{c_f} = \\frac{1}{v}\\sum_{i=1}^{v}\\|M_i^* - A_i^{-1}(\\tilde{M})\\|_1 \\), where \\( A_i^{-1}(\\cdot) \\) denotes the inverse transformation from canonical view to view-i. \\( L_{c_{2D}} \\) and \\( L_{c_f} \\) are complementary to each other. Only using \\( L_{c_{2D}} \\) tends to get performance saturation faster. In contrast, only adopting \\( L_{c_f} \\) can lead to unstable training since there may exist the situation that the fusion results are worse due to the majority of the predictions being wrong, especially at the early training stage. During training, we alternately update \\( L_{c_{2D}} \\) and \\( L_{c_f} \\) to achieve more stable optimization." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.61, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Multi-view distillation loss. Since the multi-view fusion results are much better than the 2D pseudo label, we introduce multi-view distillation loss \\( L_{d} = \\frac{1}{v}\\sum_{i = 1}^{v}\\| M_{i} - A_{i}^{-1}(\\tilde{M})\\|_{1} \\) that uses the fusion results to supervise the single-view outputs to achieve self-distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Total loss. Except for the losses for multi-view collaborative learning, our framework also adopts two general constraints, 2D joints loss, and hand prior regularization. The prior regularization regularizes the pose and shape parameters: \\( L_{p} = \\frac{1}{v}\\sum_{i = 1}^{v}\\alpha (\\| \\pmb{\\theta}_{i}\\|_{1} + \\| \\pmb{\\theta}_{i}^{*}\\|_{1} + \\gamma \\| \\pmb{\\beta}_{i}\\|_{1}) \\), where \\( \\alpha \\) and \\( \\gamma \\) are used to balance the loss scale. The 2D joints \\( L_{1} \\) loss \\( L_{2D} \\) is used to supervise the results from the 2D pseudo labels. The final loss is defined as: \\( L = L_{c} + L_{d} + L_{2D} + L_{p} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.828, + 0.21, + 0.844 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.849, + 0.276, + 0.865 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.901 + ], + "angle": 0, + "content": "FreiHAND [64] is a dataset for single-view 3D hand pose estimation, which contains 130,240 training images and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "3,960 testing images. All images are captured from the real world with 3D annotations. The training set consists of 32,560 composited images with four types of real-world backgrounds and hands captured against a green screen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.894, + 0.243 + ], + "angle": 0, + "content": "HanCo [62] extends FreiHAND, which consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we use the first 1,200 sequences for training and the last 317 sequences for testing in all experiments for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.894, + 0.318 + ], + "angle": 0, + "content": "Other datasets. We also provide additional results on other datasets. Assembly101 [46] is an action recognition dataset that consists of 4,321 videos sequence. H2O [29] is a handobject interaction dataset with 571,645 frames. Please refer to supplementary materials for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Metrics. We report standard metrics for hand pose estimation as follows. (1) MPJPE/MPVPE (mean per joint/vertex position error) measures the average Euclidean distance in mm between the predicted and ground-truth joints/vertices. JE/VE are the abbreviations for MPJPE/MPVPE. (2) NMPJPE/NMPVPE (normalized mean per joint/vertex position error, N-JE/VE) computes MPJPE/MPVPE after performing translation and scale alignment. (3)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.561 + ], + "angle": 0, + "content": "PA-MPJPE/PA-MPVPE (PA-JE/VE) is a modification of MPJPE/MPVPE with Procrustes analysis [17]. This metric normalizes the absolute scale, center, and rotation. (4) F-Score [9] is the harmonic mean of recall and precision between two meshes w.r.t. a specific distance threshold. F@5mm and F@15mm are reported. (5) AUC means the area under the curve of the PCK, where the PCK refers to the percentage of correct joints." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.568, + 0.719, + 0.583 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.694 + ], + "angle": 0, + "content": "We implement all the networks in PyTorch [40]. We first train our framework without \\( L_{c} \\) and \\( L_{d} \\) for 10 epochs. Then, we train the whole framework for another 30 epochs. Each batch contains images from 8 time-step of 8 cameras. We use AdamW [34] optimizer and set the initial learning rate to 3e-4. We use \\( 256 \\times 256 \\) hand images as input. Please refer to supplementary materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.7, + 0.805, + 0.716 + ], + "angle": 0, + "content": "4.3. Comparisons with state-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In Sec. 4.3.1, we evaluate the performance of our method under the single-view inference setting. As self-supervised hand pose estimation is a relatively new task, there is limited literature available for comparison. To address this, we adapt self-supervised body pose estimation methods [27,52] to hand and compare them with our method on HanCo [62]. We then compare with the only existing self-supervised hand pose estimation method, S²Hand [9]. As S²Hand can only be trained on single-view images, we use our single-view network only (denote as Ours-SV) for both training and inference as baselines. We further conduct extensive evaluations of our full model and baselines to demonstrate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20767" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.43, + 0.264 + ], + "angle": 0, + "content": "
MethodInputN-JE↓PA-JE↓
Fully-Supervised Method:
MobRecon [7]image9.95.7
EpipolarPose [27]image10.56.1
Self-Supervised Method:
EpipolarPose [27]image,19.79.3
CanonPose [52]2D pose,30.912.6
Oursimage,11.17.0
EpipolarPose [27]image42.323.5
CanonPose [52]2D pose31.812.8
Oursimage15.27.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.068, + 0.275, + 0.442, + 0.332 + ], + "angle": 0, + "content": "Table 1. Single-view inference comparisons on the HanCo [62] dataset. \\(\\bullet\\) denotes the method using camera extrinsics during training. Notably, in the self-supervised setting, our method exhibits a significant improvement over previous methods." + }, + { + "type": "table", + "bbox": [ + 0.467, + 0.078, + 0.895, + 0.278 + ], + "angle": 0, + "content": "
MethodDataBackbonePA-JE↓PA-VE↓F@5↑
Fully-Supervised Method:
YoutubeHand [28]Frei.Res508.48.60.61
I2UV-HandNet [6]Frei.Res506.76.90.71
MobRecon [7]Frei.Res50†6.16.20.76
Ours-SVFrei.Res507.57.50.68
Self-Supervised Method:
S²HAND [9]Frei.EffiNet-b011.811.90.48
Ours-SVFrei.EffiNet-b011.611.70.49
Ours-SVFrei.Res5011.912.00.47
Ours-SVHanCoEffiNet-b011.311.40.51
Ours-SVHanCoRes5011.611.80.48
OursHanCoEffiNet-b06.36.80.71
OursHanCoRes506.26.70.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.46, + 0.289, + 0.892, + 0.331 + ], + "angle": 0, + "content": "Table 2. Quantitative results on the FreiHAND evaluation set. The notation \\(\\dagger\\) denotes using a stacked backbone structure. \"Our-SV\" refers to training only with our single-view network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.4, + 0.359 + ], + "angle": 0, + "content": "the efficacy of multi-view collaborative learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.469, + 0.451 + ], + "angle": 0, + "content": "In addition, thanks to our cross-view interaction network, our approach is capable of performing multi-view inference by simply averaging individual view results when multi-view test data is available. In Sec. 4.3.2, we compare our method with state-of-the-art approaches under the multi-view inference setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.285, + 0.474 + ], + "angle": 0, + "content": "4.3.1 Single-View Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.54 + ], + "angle": 0, + "content": "Hanco. We train EpipolarPose and CanonPose using their open-source code. We also train fully-supervised methods [7, 27] as a reference for performance. Tab. 1 outlines the performance of fully-/self-supervised methods in the lit-" + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.55, + 0.422, + 0.858 + ], + "angle": 0, + "content": "
MethodMPJPE ↓PA-MJPJE ↓
Ttraditional Triangulation Method (w/o training):
DLT [21]16.813.2
Pictorial [11]13.510.2
RANSAC [25]12.39.8
Fully-Supervised Method:
EpipolarTrans [22]6.24.2
LT-Algebraic [25]5.53.6
LT-Volumetric [25]5.83.6
LT-Volumetric+ [25]4.93.6
EpipolarPose+ [27]8.04.4
Ours (Opt-Center)6.03.2
Ours (RANSAC)5.83.4
Self-Supervised Method:
EpipolarTrans [22]11.29.0
LT-Algebraic [25]10.37.8
LT-Volumetric [25]10.68.0
LT-Volumetric+ [25]9.57.2
CanonPose+ [52]21.610.5
EpipolarPose+ [27]17.28.3
Ours (Opt-Center)8.85.3
Ours (RANSAC)8.55.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.868, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Table 3. Multi-view inference results on the HanCo dataset. The notation \\( {}^{ + } \\) indicates that methods require the GT 3D center." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.585 + ], + "angle": 0, + "content": "erature along with ours. In the case where camera extrinsics are available for training, CanonPose performs the worst because it lifts noisy 2D pseudo labels from OpenPose to 3D ones. When camera extrinsics are not available, all competitors experience a performance decline. This is due to the lack of collaborative interaction across multi-view features in previous self-supervised methods. In contrast, our method outperforms both of them by a large margin. Our cross-view interaction networks can enhance single-view inference, whether camera extrinsics are available during training or not. More details about the usage of cameras can be found in Sec. 3.2.3. Compared to previous self-supervised methods, our approach significantly improves performance, highlighting the importance of cross-view interaction among different views. Moreover, our approach can get comparable results to fully-supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.893, + 0.814 + ], + "angle": 0, + "content": "FreiHAND. The comparisons on the evaluation set are shown in Tab. 2. The experiments conducted under self-supervised settings indicate that our baselines, Ours-SV, already achieve performance comparable to \\(\\mathbf{S}^2\\mathrm{Hand}\\). Moreover, directly equipping baselines with other backbones or more training data does not improve too much. We argue that performance improvements in single-view self-supervised hand pose estimation cannot be achieved by changing the backbone architecture or increasing the amount of training data. In contrast, our full model, i.e. Ours, substantially further improves the results on the Frei-Hand dataset, which justify the effectiveness of multi-view collaborative learning. Moreover, our self-supervised approach achieves competitive performance with recent fully-supervised state-of-the-art methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.82, + 0.704, + 0.834 + ], + "angle": 0, + "content": "4.3.2 Multi-View Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We show the quantitative results of our multi-view inference performance with other competitors on HanCo in Tab. 3. A naive solution is to triangulate pseudo labels without training. We show the performance of traditional meth-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20768" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.591, + 0.328 + ], + "angle": 0, + "content": "
IDMethodNMPJPE ↓PA-MPJPE ↓
SingleInteractFusionSingleInteractFusion
1FullResNet-50 as the backbone:
11.14↑0.038.31↓0.037.65↑0.107.05↑0.175.35↑0.075.34↑0.06
2FullResNet-18 as the backbone:
11.178.287.757.225.425.40
3-VSF11.21↓0.048.49↓0.217.81↓0.067.25↓0.035.52↓0.105.50↓0.10
4-CVA11.31↓0.148.45↓0.177.81↓0.067.29↓0.075.48↓0.065.46↓0.06
5-G111.31↓0.148.56↓0.287.77↓0.037.31↓0.095.52↓0.105.49↓0.09
6-G211.33↓0.168.38↓0.107.83↓0.087.34↓0.085.45↓0.035.42↓0.02
7-G311.30↓0.138.99↓0.697.82↓0.077.30↓0.085.45↓0.035.44↓0.04
8-Lc2D11.25↓0.088.43↓0.157.90↓0.157.32↓0.105.58↓0.165.57↓0.17
9-Lcf11.74↓0.578.98↓0.708.38↓0.637.55↓0.335.84↓0.425.80↓0.40
10-DCVI13.52↓2.35/11.99↓4.249.59↓2.37/9.42↓4.02
11-Lc14.04↓2.8717.03↓8.7510.32↓2.579.04↓1.8210.21↓4.797.92↓2.52
12-Ld17.05↓5.888.56↓0.288.01↓0.2610.13↓2.915.67↓0.255.65↓0.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.073, + 0.336, + 0.59, + 0.407 + ], + "angle": 0, + "content": "Table 4. Quantitative ablation studies. We remove each of our components here to show their contribution to our framework. Full denotes our complete model. CVI represents our whole cross-view interaction network. Other notations are consistent with Sec. 3. We report the errors of single-view outputs (Single, \\(M\\)), cross-view interaction outputs (Interact, \\(M^{*}\\)), and multi-view fusion results (Fusion, \\(\\tilde{M}\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.622 + ], + "angle": 0, + "content": "ods. Such methods can serve as a reference for evaluating the effectiveness of self-supervised methods. We adapt fully-supervised multi-view 3D pose estimation methods LT [25] and EpipolarTrans [22] to a self-supervised manner. Under self-supervised settings, EpipolarTrans can only achieve limited performance improvements compared to traditional methods. LT-Algebraic [25], which incorporates learnable confidence into the triangulation. LT-Volumetric model [25], which unprojects 2D features into a 3D volume for inference, achieves better results, but the performance is dependent on the accuracy of the hand center. Canon-Pose [52] and EpipolarPose [27] obtain multi-view inference results through simple averaging like ours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.473, + 0.903 + ], + "angle": 0, + "content": "However, both of these methods are inferior to ours because they lack cross-view interaction. As our method predicts the root-relative 3D pose, we need to conduct post-processing to obtain the absolute coordinates. We introduce two different ways to achieve this: 1) using the 2D predictions of different views to triangulate and refine a center and 2) conducting RANSAC triangulation using our 2D predictions. Both methods have their merits. Opt-center can keep the root-relative results with hand prior, resulting in low PA-MPJPE. RANSAC gets better joint-wise accuracy, which is indicated by low MPJPE. We also provide qualitative results in the supplementary materials on the Assembly101 [46] dataset, which has a static camera setup. Even for challenging head-mounted moving cameras, we achieve convincing 3D pose estimates on the H2O [29] dataset. The experiments show that we have significantly pushed the performance of self-supervised methods to a comparable level with fully supervised methods." + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.089, + 0.902, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.205, + 0.887, + 0.245 + ], + "angle": 0, + "content": "Figure 3. Error of using different (a) #training data, (b) (line-1)#view for training, and (line-2)#view for inference when trained with 8 views." + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.247, + 0.9, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.367, + 0.887, + 0.408 + ], + "angle": 0, + "content": "Figure 4. AUC of three 2D joint sets. O, S, I, PE denote OpenPose, single-view, interaction, and average pixel error in resolution \\(256 \\times 256\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.676, + 0.438 + ], + "angle": 0, + "content": "4.4. Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Fig. 5 presents the visual comparisons of 2 views between 2D joints of OpenPose, ours, and ground-truth on the HanCo dataset. We can observe that our method is more robust for outliers and can generate predictions close to the labels. Fig. 6 shows the 3D predictions from two viewpoints of ours, EpipolarPose, and CanonPose on the HanCo dataset. The results indicate that our method can get more accurate results especially when the occlusions are severe. Please refer to supplementary materials for more results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.654, + 0.605 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.671 + ], + "angle": 0, + "content": "As shown in Tab. 4, we conduct comprehensive ablation experiments on the HanCo [62] dataset to show the effectiveness of each component. Single, Interact and Fusion denotes the evaluation of \\( M \\), \\( M^{*} \\) and \\( \\tilde{M} \\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Different backbones. We first show our performance with different backbones. As shown in #1 and #2, using a large backbone like Res50, our performance can be further improved. For efficiency, we conduct ablation studies using Res18 as the backbone unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Two branches for cross-view interaction module. As presented in #3 and #4, both of the branches can reduce the error. VSF can explicitly model the view-shared information and add reliable information from every view. CVA can capture the self-/cross-view joint-level correlations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Graph features. The results indicate that three kinds of features (#5, #6, #7) all lead to performance improvement. Especially, local feature (#7, \\( G^3 \\)) can notably reduce the error after the interaction by providing fine-grained details. DCVI. We also conduct experiments to show the impor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20769" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "tance of DCVI by removing it and posing consistency constraints in single-view outputs like [24]. In this way, the performance drops dramatically (#10), proving the necessity of using DCVI to capture the features of all the views for self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.242 + ], + "angle": 0, + "content": "Two branches for multi-view consistency loss. Without enforcing cross-view interaction outputs to be consistent, the performance significantly drops (#9). If we do not explore relatively more reliable 2D predictions to enhance consistency, the performance can also get worse (#8)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.468, + 0.332 + ], + "angle": 0, + "content": "Consistency losses. \\((L_{c})\\) The performance is unsatisfactory (#11) when employing the cross-view interaction network without any consistency constraints (i.e. discard #8 and #9). The interaction network should cooperate with consistency so that the constraints can guide the network to exploit multi-view information to function better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.468, + 0.408 + ], + "angle": 0, + "content": "Multi-view distillation loss. \\((L_{d})\\) Removing the multiview distillation loss, all the metrics drop by a large margin (#12), especially in single-view estimation accuracy. This phenomenon proves the effectiveness of collaborative learning between single- and multi-view networks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.233, + 0.43 + ], + "angle": 0, + "content": "4.6. Model Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.468, + 0.478 + ], + "angle": 0, + "content": "Different percentage of unlabeled images. Fig. 3 (a) shows our method can get consistent performance improvement as the unlabeled training data increases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.468, + 0.584 + ], + "angle": 0, + "content": "Different view number for training. The line-1 in Fig. 3 (b) shows the performance of our method tested on a certain view when trained with different view numbers. The curve shows that our method can be consistently improved as the number of views increases. We also observe that using multiple views for training can significantly improve performance when the valid views are few." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.468, + 0.78 + ], + "angle": 0, + "content": "Different view number for inference. Our model allows inferring with an arbitrary number of views. However, when the model is trained with a fixed view number, it could get the view number bias, resulting in better performance using the view number close to the training one. To avoid this, we add random masks in our interaction module and finetune the model for a few epochs. After that, results can get better by a small margin (the single-view error is \\(11.07\\mathrm{mm}\\) and the fusion error \\(7.60\\mathrm{mm}\\), both in NMPJPE.). The line-2 in Fig. 3 (b) shows results on a certain view when trained on 8 views and tested on 1 to 8 views. We can observe consistent improvement with the inference view number increases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.841 + ], + "angle": 0, + "content": "Different 2D joint sets. Fig. 4 presents the accuracy of different 2D joint sets on the HanCo training and testing set. Our 2D predictions are extremely better than OpenPose 2D pseudo label used for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.842, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Iteratively training. Our approach can use the previous predictions as pseudo labels for iterative training. We find it helpful till iteration 3 and get saturated afterward. From 1 to 3 iterations, NMPJPE is 7.75, 7.68, and 7.64." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.088, + 0.872, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.224 + ], + "angle": 0, + "content": "Figure 5. 2D prediction (overlayed in the images) comparisons between OpenPose, ours, and ground-truth on the HanCo dataset." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.233, + 0.87, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 6. 3D prediction comparisons between our method, EpipolarPose, and CanonPose on the HanCo dataset. Our prediction and ground-truth are shown in solid red and dashed green respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.395, + 0.769, + 0.41 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.672 + ], + "angle": 0, + "content": "To our best knowledge, we present the first self-supervised framework that aims to learn a single-view 3D hand estimator from unlabeled multi-view data. At the core of our approach, a cross-view interaction network is carefully designed to supervise the single-view output by leveraging the collaboration among multi-views. Specifically, the network captures the interdependencies of features among different views, resulting in improved accuracy of hand pose estimation after cross-view interaction. Additionally, the multi-view results are fused to supervise the single-view output for self-distillation. The effectiveness and versatility of the proposed framework are extensively evaluated through experiments, which demonstrate that our method not only establishes a new benchmark for self-supervised 3D hand pose estimation from single-view input but also offers flexible multi-view inference with state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.749 + ], + "angle": 0, + "content": "We focused on hand pose estimation without heavy occlusions in this work. Extending our work to more challenging scenarios, such as hand-object interaction or relaxing the synchronization constraints in multi-view inputs, would be interesting topics for further study." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.759, + 0.66, + 0.775 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants (62071067, 62171057, 62201072), in part by the Ministry of Education and China Mobile Joint Fund (MCM20200202), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20770" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[3] Kristijan Bartol, David Bojanic, Tomislav Petković, and Tomislav Pribanić. Generalizable human pose triangulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11028-11037, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[5] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.703, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 7792-7801, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[13] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 2021 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[14] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[15] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1991-2000, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[16] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[17] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[19] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Trans. Graph., 39(4):87, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[20] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, Cabezas Randi, Tran Luan, Akbay Muzaffer, Yu Tsz-Ho, Keskin Cem, and Wang Robert. Umetrack: Unified multi-view end-to-end hand tracking for vr. ACM Transactions on Graphics, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[21] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[22] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[23] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European Conference on Computer Vision (ECCV), pages 118-134, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20771" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5243-5252, 2020. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[25] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.47, + 0.301 + ], + "angle": 0, + "content": "[27] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1077-1086, 2019. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990–5000, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[29] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.525 + ], + "angle": 0, + "content": "[30] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[31] Moran Li, Yuan Gao, and Nong Sang. Exploiting learnable joint groups for hand pose estimation. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1921-1929, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[32] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[33] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.804 + ], + "angle": 0, + "content": "[35] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In European Conference on Computer Vision, pages 424-442. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[36] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[37] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[38] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[39] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 49-59, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in python. In NIPS 2017 Workshop on Autodiff, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[42] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4342-4351, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[43] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6040-6049, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[44] Helge Rhodin, Jörg Spörri, Isinsu Katircioglu, Victor Constantin, Frédéric Meyer, Erich Müller, Mathieu Salzmann, and Pascal Fua. Learning monocular 3d human pose estimation from multi-view images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8437-8446, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[45] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[46] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[47] Hui Shuai, Lele Wu, and Qingshan Liu. Adaptive multi-view and temporal fusing transformer for 3d human pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[48] Adrian Spurr, Umar Iqbal, Pavlo Molchanov, Otmar Hilliges, and Jan Kautz. Weakly supervised 3d hand pose estimation via biomechanical constraints. In European Conference on Computer Vision, pages 211-228. Springer, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20772" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[49] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[50] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[51] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.261, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[52] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13294-13304, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[53] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.4, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[54] Chao Wen, Yinda Zhang, Zhuwen Li, and Yanwei Fu. Pixel2mesh++: Multi-view 3d mesh generation via deformation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1042-1051, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.457, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[55] Linlin Yang, Shicheng Chen, and Angela Yao. Semihand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[56] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2335-2343, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.568, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[57] Linlin Yang and Angela Yao. Disentangling latent hands for image synthesis and pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9877-9886, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.624, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[58] Xiong Zhang, Hongsheng Huang, Jianchao Tan, Hongmin Xu, Cheng Yang, Guozhu Peng, Lei Wang, and Ji Liu. Hand image understanding via deep multi-task learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11281-11292, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.693, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[59] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[60] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129(3):703-718, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[62] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In DAGM German Conference on Pattern Recognition, pages 250-264. Springer, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "20773" + } + ] +] \ No newline at end of file diff --git a/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_origin.pdf b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a46e233225cacaf3925e178fdeaf9b855a5024a7 --- /dev/null +++ b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/20ed3f4b-5914-4843-b685-abff65a0a8b5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0170aab9e4ced12abda04da62450eb19f088d54763d18fde57cff2cc07cda6f8 +size 1609599 diff --git a/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/full.md b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a62f6d7c8fe26dbb99c67dc3144788b82e0fcb6 --- /dev/null +++ b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/full.md @@ -0,0 +1,301 @@ +# HaMuCo: Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning + +Xiaozheng Zheng $^{1,2\dagger}$ Chao Wen $^{2\dagger}$ Zhou Xue $^{2}$ Pengfei Ren $^{1,2}$ Jingyu Wang $^{1*}$ $^{1}$ State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications + $^{2}$ PICO IDL, ByteDance, Beijing + +# Abstract + +Recent advancements in 3D hand pose estimation have shown promising results, but its effectiveness has primarily relied on the availability of large-scale annotated datasets, the creation of which is a laborious and costly process. To alleviate the label-hungry limitation, we propose a self-supervised learning framework, HaMuCo, that learns a single-view hand pose estimator from multi-view pseudo 2D labels. However, one of the main challenges of self-supervised learning is the presence of noisy labels and the "groupthink" effect from multiple views. To overcome these issues, we introduce a cross-view interaction network that distills the single-view estimator by utilizing the cross-view correlated features and enforcing multi-view consistency to achieve collaborative learning. Both the single-view estimator and the cross-view interaction network are trained jointly in an end-to-end manner. Extensive experiments show that our method can achieve state-of-the-art performance on multi-view self-supervised hand pose estimation. Furthermore, the proposed cross-view interaction network can also be applied to hand pose estimation from multi-view input and outperforms previous methods under the same settings. + +# 1. Introduction + +3D hand pose estimation is essential in various application scenarios, from action recognition and sign language translation to AR/VR [19, 20]. Hand pose estimation has achieved a significant improvement in recent years. However, the progress heavily relies on the emergence of many hand pose datasets with accurate 3D annotations. Acquiring labeled datasets is quite time-consuming and laborious, exposing a realistic challenge for deep learning models to learn with limited and noisy data. + +Self-supervised learning is an emerging solution to the + +challenge posed by manual annotation. Though worth exploring, self-supervised pose estimation with RGB hand images is a challenging and relatively unexplored area with only one pioneering method, S²HAND [9]. S²HAND aims to conduct 3D hand reconstruction from a single RGB image with the noisy off-the-shell 2D hand pose estimation results for supervision. Unfortunately, S²HAND faces a predicament where its performance is significantly reliant on the quality of the pseudo label, and inferior labeling may result in reduced performance. Moreover, evaluating the quality of the pseudo label is an ill-posed problem that lacks clear criteria or input, further complicating the issue. + +This observation motivates us to leverage multi-view information for enhancing self-supervised learning, as the complementary nature of multi-view observations can help mitigate the ambiguity inherent in pose estimation. Although the first 3D hand dataset with synchronized multiview input (HanCo [62]) was proposed in 2021, to our knowledge, there is no previous work exploring the potential of multi-view for self-supervised hand pose estimation. Therefore, we turn to studies in the human body pose estimation, which share some similarities. + +As mentioned in previous work [24], naively enforcing multi-view consistency is prone to generate degenerated solutions, thus they resorted to additional 2D labels of unrelated datasets and proposed a solution under the scope of weakly supervised learning. Other studies, such as EpipolarPose [27] and CanonPose [52], utilized multiview data with special designs to enhance the supervision and achieved promising results under the scope of self-supervised learning. + +In this paper, we push along this direction on hand pose estimation via multi-view collaborative learning. We take one step further by designing a learnable network, which utilizes multi-view information, to tackle 1) noisy pseudo labels and 2) unreliable multi-view "groupthink" issues causing training collapse in the early training stage. Formally, we name the pipeline HaMuCo, which stands for + +![](images/7d09d4158ed2d1c0064b9952d4d2abe7ab9202033f3a8228437bc64ff9e29170.jpg) +Figure 1. Overall pipeline comparison: HaMuCo learns a monocular 3D hand pose estimator from multi-view self-supervision via cross-view feature interaction. Our cross-view interaction network addresses the importance of introducing learnable feature interaction, which is absent from previous methods [27, 52]. At inference time only the gray part is applied. + +![](images/b4d5b31d5e12985556bbef5fbb46b83d25db8a8469019971dfbeeecc03650dc4.jpg) + +![](images/a62fda65e2c2842ade79d1f653ca1f18544362d5c308ca52374529863b39cfdf.jpg) + +# Hand Multiview Collaborative learning. + +The core idea of our approach is to enhance the single-view estimation by means of cross-view feature interaction and further integrate multi-view results to supervise the single-view output to achieve self-distillation in an end-to-end fashion. Thus, our framework is built with a single-view hand pose estimator and a cross-view interaction network for supervision. The single-view estimator uses the MANO [45] hand model as the decoder, which provides the hand prior to regularizing irrational anatomy when supervised by noisy pseudo labels. The cross-view interaction network captures cross-view features and utilizes several consistent losses among different views to guide collaborative learning. + +We conduct comprehensive experiments on the HanCo [62] dataset and our approach outperforms previous methods by a considerable margin for self-supervised 3D hand pose estimation. Notably, our results demonstrate competitive performance compared to a state-of-the-art fully supervised approach proposed by Chen et al. [7]. Our proposed framework is highly versatile, as it can be trained with or without calibration, and is capable of incorporating the cross-view interaction network to achieve superior multi-view inference results when multi-view test data is available. Moreover, our model can generalize well to other datasets [29, 46, 64] and in-the-wild images. + +In summary, our contributions are the following: + +- We propose the first self-supervised learning framework for single-view hand pose estimation without any training data annotation and achieve state-of-the-art performance via multi-view collaborative learning. +- We propose a cross-view interaction network to supervise the single-view estimator by enforcing multi-view consistency and capturing cross-view features for collaborative learning among multiple views. +- The proposed framework is capable of multi-view inference by incorporating the cross-view interac + +tion network and achieves state-of-the-art performance without bells and whistles. + +# 2. Related Work + +Hand Pose Estimation. Hand pose estimation can be categorized into RGB-based methods [23, 49, 63] and depth-based methods [14, 15, 36], depending on the input modality. In this paper, we focus our attention on RGB-based hand pose estimation. The RGB-based methods can be further divided into three categories, skeleton-based methods [5, 12, 23, 31, 38, 39, 48, 49, 55-57, 63], model-based methods [1, 2, 4, 9, 58, 59, 64], and mesh-based methods [7,8,10,16,28,30,32,33,37,50,61]. Skeleton-based methods regress the hand joints directly. Zimmermann et al. [63] introduces a multi-stage network that lifts the regressed 2D joints to 3D ones. Variational autoencoder [26] is employed to learn a cross-modal latent space to achieve better hand pose estimation and disentanglement [49, 56, 57]. Latent 2.5D representation regression is proved more effective than direct coordinates regression for hands by [23], which is also adopted by [13, 31, 48, 61]. There are also many works solving hand pose estimation with two hands interactions [13, 29, 38] and hand-object interactions [2, 12, 29]. Recent model-based methods make use of MANO [45], which can incorporate the hand prior and predict the hand mesh simultaneously. Those methods [1, 4, 9, 58, 59] rely on additional supervisions [1, 4, 9, 58, 59] or inputs [4]. In contrast, mesh-based methods regress each vertex directly, which is more accurate but requires large-scale datasets with hand mesh annotations [18, 29, 38, 64]. Most of these methods utilize graph convolutional network (GCN) [7,8,10,16,28,50,61] or transformers [32] or both [30, 33] for regression. I2L-MeshNet [37] regresses each vertex by predicting 1D heatmaps of three axes. Chen et al. [6] uses an image-to-image translation network to predict the UV map of the mesh. Similar to previous works [30, 33], we also use transformer and GCN. However, we employ them for cross-view interaction. + +Multi-View Fully-Supervised Pose Estimation. Multiview information is widely explored to improve 3D human pose estimation by tackling occlusions and depth-ambiguity in a fully-supervised manner [3, 22, 25, 41-43, 47, 60]. Volume-based methods [25, 41, 42, 51] unproject 2D features or heatmaps of joints to a 3D space for estimation. Another kind of method [22, 43, 60] utilizes the geometry information to fuse the features in 2D space directly and efficiently. Recently, some works [35,47] utilize transformers for implicit cross-view fusion without camera extrinsics. + +Label-Efficient Learning. Label-efficient learning aims to reduce the 3D label requirements. Many works devote to solving hand pose estimation in a label-efficient manner [1, 4,5,7,9,39,48,55,59,63]. Synthetic data is used to avoid manual annotation [7,39,63], but may need domain transfer [39]. Or use weakly supervised learning [4,48] to obtain 3D results by manually annotating 2D labels to assist with hand priors. Multi-view label-efficient learning is also explored in 3D pose estimation [24, 27, 44, 52]. Rhodin et al. [44] trains a semi-supervised network with only a small amount of labeled 3D data and multi-view consistency constraints. Iqbal et al. [24] mixes single-view images with 2D labels and unlabelled multi-view images for training. Our goal is the same as that of previous methods, which is to train without any manual 3D labels. + +Self-supervised 3D Pose Estimation. (1) Single-view training and inference. To the best of our knowledge, there is only one method for self-supervised 3D hand pose estimation, proposed by Chen et al. [9]. Their framework, $\mathrm{S}^2\mathrm{Hand}$ , uses only single-view 2D noisy labels for training and achieves self-supervision through rendering. However, the performance is limited due to the use of single-view information and the quality of the noisy labels. (2) Multi-view training, single-view inference. Our approach belongs to this category but is fundamentally different from the existing methods. EpipolarPose [27] triangulates multi-view 2D pseudo labels according to epipolar geometry to 3D ones for training. CanonPose [52] learns to lift 2D pseudo labels to 3D canonical pose space with multi-view consistency constraints. All the aforementioned methods use nonlearnable self-supervised modules like geometric modules or consistency loss functions, as shown in Fig. 1. However, they [27, 52] ignore the importance of introducing cross-view interaction and multi-view collaborative learning. Previous methods struggle to achieve good performance since the pose of a hand can change drastically over time and different joints may have similar appearances. + +# 3. Method + +As depicted in the left part of Fig. 2, our framework consists of a simple yet effective single-view estimator and cross-view interaction network. The core idea of our approach is that prediction from a monocular view can be + +enhanced via cross-view feature interaction and the interacted results can further supervise the single-view output to achieve self-distillation. + +# 3.1. Single-View Estimator + +Overview. Our framework takes multi-view synchronized hand images $\mathcal{I} = \{I_i\}_{i=0}^v$ with $v$ views as input, each view is an image of $I_i \in \mathbb{R}^{3 \times h \times w}$ . The output is a 3D hand mesh $M$ on each view. We designed a simple yet effective model-based network as a single-view estimator. Using the hand model will reduce the adverse effects of using poor pseudo labels as supervision by providing hand prior information for regularization. Please refer to supplementary materials for more details about the single-view estimator. + +Hand model. We employ MANO [45] as the hand model. The hand mesh can be derived from the MANO layer using parameters $\beta$ and $\pmb{\theta}$ , i.e. $M(\beta, \pmb{\theta})$ . $\beta \in \mathbb{R}^{10}$ and $\pmb{\theta} \in \mathbb{R}^{16 \times 3}$ control the shape and pose of the hand respectively. We can use a predefined regressor to obtain the 3D joints from the 3D mesh vertices by $P = JM$ , where $J \in \mathbb{R}^{k \times n}$ , where $n = 778$ and $k = 21$ are the joints number and vertices number. For more details, we recommend referring to [45]. + +Camera model. Following Boukhayma et al. [4], we model the geometry correspondence by the weak-perspective camera model and obtain camera parameters from the single-view network predictions. Given the translation $\pmb{t}$ and scale $s$ , the 2D coordinates in image plane can be obtained by: $\Pi(P) = s\Omega(P) + \mathbf{t}$ , where $\Omega$ is the orthographic projection and $\Pi$ denotes the weak-perspective projection. + +Network Structure. Since the single-view estimator is not the main component, for the sake of simplicity, we employ a CNN as the encoder $F_{e}$ , and an MLP as the decoder $F_{d}$ for regressing the MANO parameters. We have 3D hand mesh: $M_{i}(\pmb{\theta}_{i},\pmb{\beta}_{i}) = F_{s}(\pmb{I}_{i})$ , where $F_{s} = F_{d}(F_{e}(\cdot))$ denotes the entire single-view network. The estimator also passes different levels of features $H^{j}$ (where $H^{j}$ is the intermediate feature of the encoder after $j$ residual blocks, $j = 1,2,3,4$ ) to our cross-view interaction network. + +# 3.2. Cross-view Interaction Network + +In this section, we introduce the cross-view interaction network (CVI-Net), which is the core of our system to enable the network to exploit multi-view information. This stage conducts cross-view interaction and distillation. The critical components of this stage are a cross-view interaction network for capturing cross-view features and several consistent losses for guiding collaborative learning. + +# 3.2.1 View-Shared Graph Feature Extraction + +The first step for interaction is to extract the appropriate features. Different from [7, 8, 61], our module collects useful information into a graph through view-shared graph feature extraction module (VSGFE) as shown in Fig. 2. Specifically, it makes use of multi-level feature maps from dif + +![](images/16227f54fe24ebe33674bbd0efe6b4c000432e52d3ca5015ba66cff1849a7dae.jpg) +Figure 2. The left illustrates our whole pipeline (2 views for simplicity). During the training phase, the network takes multi-view hand images and pseudo-labels as inputs. The bottom right depicts our cross-view interaction networks. The top right shows the view-shared graph feature extraction (VSGFE) module and view-shared feature (VSF) module. $\oplus$ and $\otimes$ denotes add and concatenation respectively. + +![](images/51232d8a0153c6a748b5b783aab649a111933bd0e9ea6629fb5f51fe1e8ef874.jpg) + +![](images/093c5d08c4d48f11e75ee9a85cefff1ae12f311019caad37dab2ef703bdd0a23.jpg) + +![](images/bb20bfda8d5d2cd0c90a2689039651ef6d202478817830e2b2021fd950055849.jpg) + +ferent views $\mathcal{H} = \{\pmb{H}_i^j\}_{i=0}^v$ , 3D joints $\mathcal{P} = \{\pmb{P}_i\}_{i=0}^v$ , and MANO pose parameters $\Theta = \{\pmb{\theta}_i\}_{i=0}^v$ from the single-view estimator to extract a graph feature $\pmb{G}$ . The graph feature of each view $\pmb{G}_i$ consists of three parts. $\pmb{G}_i^1$ , $\pmb{G}_i^2$ and $\pmb{G}_i^3$ aim to capture joint location features, global image features, and local image features, respectively. The first part is joint location embedding $\pmb{G}_i^1 \in \mathbb{R}^{k \times c_1}$ , providing the explicit geometric information. This embedding is obtained by using an MLP to map the single-view 3D joints locations $\pmb{P}_i$ and pose parameters $\pmb{\theta}_i$ to dimension $c_1$ . The second part is joint-wise high-level image features $\pmb{G}_i^2 \in \mathbb{R}^{k \times c_2}$ generated by spatial-aware initial graph building (SAIGB) [61] module using the last level feature maps $\pmb{H}_i^4$ . This part provides compact image clues of all views for interaction. The third part is joint-aligned features $\pmb{G}_i^3 \in \mathbb{R}^{k \times c_3}$ gathered by joint feature sampler (JFS). JFS projects joints onto multi-level image feature maps $\{\pmb{H}_i^j\}_{j=1}^3$ to gather fine-grained perceptual features like [53,54] for better local alignment. We then concatenate graph features to get $\pmb{G}_i = [\pmb{G}_i^1 \otimes \pmb{G}_i^2 \otimes \pmb{G}_i^3]$ . + +# 3.2.2 Dual-Branch Cross-View Interaction (DCVI) + +We first stack $\{G_i\}_{i=0}^v$ of all views to obtain multi-view graph feature $G \in \mathbb{R}^{vk \times (c_1 + c_2 + c_3)}$ . We design a component to effectively capture complementary information from other views on multi-view graph feature $G$ . The interaction module has two branches, (1) cross-view attention branch (CVA) and (2) view-shared feature branch (VSF). Cross-view attention branch utilizes a cross-view transformer $F_t$ consisting of several multi-head attention layers with token size $vk$ and MLPs, which allows each joint to aggregate features from other joints or views. This branch implicitly captures the multi-view information. An explicit + +multi-view prior information is that the observed poses from all the views should be consistent in 3D. Therefore, we add a branch to excavate the multi-view shared information to enhance the feature representation. Specifically, view-shared feature branch first employs adaptive-GCN [12] $F_{a}$ to map the view-specific features $\pmb{G}_{i}$ to a canonical feature space $C_i = F_a(G_i)$ , the nodes in adaptive-GCN represents the hand joints and the edges represent joint feature correlation. Then, we stack $\mathcal{C} = \{C_i\}_{i = 0}^v$ together to get multi-view canonical features $\pmb{C} \in \mathbb{R}^{v \times k \times (c_1 + c_2 + c_3)}$ . After that, we use max-pooling on $\pmb{C}$ to get the max activated features of every joint then repeat them in the view dimension as the view-shared features $\pmb{C}' \in \mathbb{R}^{vk \times (c_1 + c_2 + c_3)}$ . We denote the dual-branch cross-view interaction as: $\pmb{G}^{*} = \pmb{G} + F_{t}(\pmb{G}) + \pmb{C}^{\prime}$ , where $\pmb{G}^{*}$ is the updated graph feature. + +Parameters regression. The view specific feature $G_{i}^{*}$ after the interaction can be obtained by reshaping $G^{*}$ . We then employ a shared MLP $F_{r}$ as a decoder to regress the pose parameters $\theta_{i}^{*} = F_{r}(G_{i}^{*})$ to derive the hand mesh of each view $M_{i}^{*}(\theta_{i}^{*},\beta_{i})$ and corresponding joints $P_{i}^{*} = JM_{i}^{*}$ . + +# 3.2.3 Multi-View Collaborative Learning + +To allow all the views and the networks to learn collaboratively, we utilize consistency losses $L_{c}$ upon interaction outputs and distillation loss $L_{d}$ between multi-view fusion results and single-view outputs, as shown in Fig. 2. $L_{c}$ introduces collaborative learning between multiple views, guiding the poses from different views to be as close as possible. While $L_{d}$ makes the CVI-Net and single-view estimator work in a collaborative manner, achieving a self-distillation effect. + +Results fusion. Since we need to supervise the single-view + +estimator with the results after the interaction, instead of simply using the refined results $M_{i}^{*}$ of each view, we ensemble all the results into a unified and more reliable result $\tilde{M}$ . Considering the lack of explicit guidance, we empirically introduce a prior that all the views contribute equally. Thus, we simply average all aligned results to obtain $\tilde{M}$ . Specifically, we use $A$ to denote the align procedure. When the extrinsics are known, we use the relative camera pose for alignment. When the camera extrinsics are unavailable, we use Procrustes analysis [62, 64] to compute relative rotation and align meshes to a canonical view. The final result is calculated as follows: $\tilde{M} = \frac{1}{v}\sum_{i=1}^{v}A(M_{i}^{*})$ . + +Consistency losses. We design two types of consistency loss $L_{c}$ : 2D consistency loss $L_{c_{2D}}$ and Fusion consistency loss $L_{c_f}$ . The motivation behind $L_{c_{2D}}$ is that the 2D predictions in the x-axis and y-axis are more accurate than the depth prediction in the z-axis. Therefore, $L_{c_{2D}}$ utilizes the 2D predictions in every single view as the pseudo label to supervise other views, which explores the view-specific reliable information to collaboratively improve the predictions of all the views. 2D consistency loss is defined as: $L_{c_{2D}} = \frac{1}{v^2}\sum_{i=1}^{v}\sum_{j=1}^{v}\|\Pi(M_i^*) - \Pi(A_i(M_j^*))\|_1$ , where $A_i(\cdot)$ denotes the alignment operation to align other view-j to view-i. Fusion consistency loss uses the fused results to supervise each view. The loss is defined as: $L_{c_f} = \frac{1}{v}\sum_{i=1}^{v}\|M_i^* - A_i^{-1}(\tilde{M})\|_1$ , where $A_i^{-1}(\cdot)$ denotes the inverse transformation from canonical view to view-i. $L_{c_{2D}}$ and $L_{c_f}$ are complementary to each other. Only using $L_{c_{2D}}$ tends to get performance saturation faster. In contrast, only adopting $L_{c_f}$ can lead to unstable training since there may exist the situation that the fusion results are worse due to the majority of the predictions being wrong, especially at the early training stage. During training, we alternately update $L_{c_{2D}}$ and $L_{c_f}$ to achieve more stable optimization. + +Multi-view distillation loss. Since the multi-view fusion results are much better than the 2D pseudo label, we introduce multi-view distillation loss $L_{d} = \frac{1}{v}\sum_{i = 1}^{v}\| M_{i} - A_{i}^{-1}(\tilde{M})\|_{1}$ that uses the fusion results to supervise the single-view outputs to achieve self-distillation. + +Total loss. Except for the losses for multi-view collaborative learning, our framework also adopts two general constraints, 2D joints loss, and hand prior regularization. The prior regularization regularizes the pose and shape parameters: $L_{p} = \frac{1}{v}\sum_{i = 1}^{v}\alpha (\| \pmb{\theta}_{i}\|_{1} + \| \pmb{\theta}_{i}^{*}\|_{1} + \gamma \| \pmb{\beta}_{i}\|_{1})$ , where $\alpha$ and $\gamma$ are used to balance the loss scale. The 2D joints $L_{1}$ loss $L_{2D}$ is used to supervise the results from the 2D pseudo labels. The final loss is defined as: $L = L_{c} + L_{d} + L_{2D} + L_{p}$ . + +# 4. Experiments + +# 4.1. Datasets and Metrics + +FreiHAND [64] is a dataset for single-view 3D hand pose estimation, which contains 130,240 training images and + +3,960 testing images. All images are captured from the real world with 3D annotations. The training set consists of 32,560 composited images with four types of real-world backgrounds and hands captured against a green screen. + +HanCo [62] extends FreiHAND, which consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we use the first 1,200 sequences for training and the last 317 sequences for testing in all experiments for fair comparisons. + +Other datasets. We also provide additional results on other datasets. Assembly101 [46] is an action recognition dataset that consists of 4,321 videos sequence. H2O [29] is a handobject interaction dataset with 571,645 frames. Please refer to supplementary materials for details. + +Metrics. We report standard metrics for hand pose estimation as follows. (1) MPJPE/MPVPE (mean per joint/vertex position error) measures the average Euclidean distance in mm between the predicted and ground-truth joints/vertices. JE/VE are the abbreviations for MPJPE/MPVPE. (2) NMPJPE/NMPVPE (normalized mean per joint/vertex position error, N-JE/VE) computes MPJPE/MPVPE after performing translation and scale alignment. (3) + +PA-MPJPE/PA-MPVPE (PA-JE/VE) is a modification of MPJPE/MPVPE with Procrustes analysis [17]. This metric normalizes the absolute scale, center, and rotation. (4) F-Score [9] is the harmonic mean of recall and precision between two meshes w.r.t. a specific distance threshold. F@5mm and F@15mm are reported. (5) AUC means the area under the curve of the PCK, where the PCK refers to the percentage of correct joints. + +# 4.2. Implementation Details + +We implement all the networks in PyTorch [40]. We first train our framework without $L_{c}$ and $L_{d}$ for 10 epochs. Then, we train the whole framework for another 30 epochs. Each batch contains images from 8 time-step of 8 cameras. We use AdamW [34] optimizer and set the initial learning rate to 3e-4. We use $256 \times 256$ hand images as input. Please refer to supplementary materials for more details. + +# 4.3. Comparisons with state-of-the-arts + +In Sec. 4.3.1, we evaluate the performance of our method under the single-view inference setting. As self-supervised hand pose estimation is a relatively new task, there is limited literature available for comparison. To address this, we adapt self-supervised body pose estimation methods [27,52] to hand and compare them with our method on HanCo [62]. We then compare with the only existing self-supervised hand pose estimation method, S²Hand [9]. As S²Hand can only be trained on single-view images, we use our single-view network only (denote as Ours-SV) for both training and inference as baselines. We further conduct extensive evaluations of our full model and baselines to demonstrate + +
MethodInputN-JE↓PA-JE↓
Fully-Supervised Method:
MobRecon [7]image9.95.7
EpipolarPose [27]image10.56.1
Self-Supervised Method:
EpipolarPose [27]image,19.79.3
CanonPose [52]2D pose,30.912.6
Oursimage,11.17.0
EpipolarPose [27]image42.323.5
CanonPose [52]2D pose31.812.8
Oursimage15.27.7
+ +Table 1. Single-view inference comparisons on the HanCo [62] dataset. $\bullet$ denotes the method using camera extrinsics during training. Notably, in the self-supervised setting, our method exhibits a significant improvement over previous methods. + +
MethodDataBackbonePA-JE↓PA-VE↓F@5↑
Fully-Supervised Method:
YoutubeHand [28]Frei.Res508.48.60.61
I2UV-HandNet [6]Frei.Res506.76.90.71
MobRecon [7]Frei.Res50†6.16.20.76
Ours-SVFrei.Res507.57.50.68
Self-Supervised Method:
S²HAND [9]Frei.EffiNet-b011.811.90.48
Ours-SVFrei.EffiNet-b011.611.70.49
Ours-SVFrei.Res5011.912.00.47
Ours-SVHanCoEffiNet-b011.311.40.51
Ours-SVHanCoRes5011.611.80.48
OursHanCoEffiNet-b06.36.80.71
OursHanCoRes506.26.70.72
+ +the efficacy of multi-view collaborative learning. + +In addition, thanks to our cross-view interaction network, our approach is capable of performing multi-view inference by simply averaging individual view results when multi-view test data is available. In Sec. 4.3.2, we compare our method with state-of-the-art approaches under the multi-view inference setting. + +# 4.3.1 Single-View Inference + +Hanco. We train EpipolarPose and CanonPose using their open-source code. We also train fully-supervised methods [7, 27] as a reference for performance. Tab. 1 outlines the performance of fully-/self-supervised methods in the lit- + +Table 2. Quantitative results on the FreiHAND evaluation set. The notation $\dagger$ denotes using a stacked backbone structure. "Our-SV" refers to training only with our single-view network. + +
MethodMPJPE ↓PA-MJPJE ↓
Ttraditional Triangulation Method (w/o training):
DLT [21]16.813.2
Pictorial [11]13.510.2
RANSAC [25]12.39.8
Fully-Supervised Method:
EpipolarTrans [22]6.24.2
LT-Algebraic [25]5.53.6
LT-Volumetric [25]5.83.6
LT-Volumetric+ [25]4.93.6
EpipolarPose+ [27]8.04.4
Ours (Opt-Center)6.03.2
Ours (RANSAC)5.83.4
Self-Supervised Method:
EpipolarTrans [22]11.29.0
LT-Algebraic [25]10.37.8
LT-Volumetric [25]10.68.0
LT-Volumetric+ [25]9.57.2
CanonPose+ [52]21.610.5
EpipolarPose+ [27]17.28.3
Ours (Opt-Center)8.85.3
Ours (RANSAC)8.55.6
+ +Table 3. Multi-view inference results on the HanCo dataset. The notation ${}^{ + }$ indicates that methods require the GT 3D center. + +erature along with ours. In the case where camera extrinsics are available for training, CanonPose performs the worst because it lifts noisy 2D pseudo labels from OpenPose to 3D ones. When camera extrinsics are not available, all competitors experience a performance decline. This is due to the lack of collaborative interaction across multi-view features in previous self-supervised methods. In contrast, our method outperforms both of them by a large margin. Our cross-view interaction networks can enhance single-view inference, whether camera extrinsics are available during training or not. More details about the usage of cameras can be found in Sec. 3.2.3. Compared to previous self-supervised methods, our approach significantly improves performance, highlighting the importance of cross-view interaction among different views. Moreover, our approach can get comparable results to fully-supervised methods. + +FreiHAND. The comparisons on the evaluation set are shown in Tab. 2. The experiments conducted under self-supervised settings indicate that our baselines, Ours-SV, already achieve performance comparable to $\mathbf{S}^2\mathrm{Hand}$ . Moreover, directly equipping baselines with other backbones or more training data does not improve too much. We argue that performance improvements in single-view self-supervised hand pose estimation cannot be achieved by changing the backbone architecture or increasing the amount of training data. In contrast, our full model, i.e. Ours, substantially further improves the results on the Frei-Hand dataset, which justify the effectiveness of multi-view collaborative learning. Moreover, our self-supervised approach achieves competitive performance with recent fully-supervised state-of-the-art methods + +# 4.3.2 Multi-View Inference + +We show the quantitative results of our multi-view inference performance with other competitors on HanCo in Tab. 3. A naive solution is to triangulate pseudo labels without training. We show the performance of traditional meth- + +
IDMethodNMPJPE ↓PA-MPJPE ↓
SingleInteractFusionSingleInteractFusion
1FullResNet-50 as the backbone:
11.14↑0.038.31↓0.037.65↑0.107.05↑0.175.35↑0.075.34↑0.06
2FullResNet-18 as the backbone:
11.178.287.757.225.425.40
3-VSF11.21↓0.048.49↓0.217.81↓0.067.25↓0.035.52↓0.105.50↓0.10
4-CVA11.31↓0.148.45↓0.177.81↓0.067.29↓0.075.48↓0.065.46↓0.06
5-G111.31↓0.148.56↓0.287.77↓0.037.31↓0.095.52↓0.105.49↓0.09
6-G211.33↓0.168.38↓0.107.83↓0.087.34↓0.085.45↓0.035.42↓0.02
7-G311.30↓0.138.99↓0.697.82↓0.077.30↓0.085.45↓0.035.44↓0.04
8-Lc2D11.25↓0.088.43↓0.157.90↓0.157.32↓0.105.58↓0.165.57↓0.17
9-Lcf11.74↓0.578.98↓0.708.38↓0.637.55↓0.335.84↓0.425.80↓0.40
10-DCVI13.52↓2.35/11.99↓4.249.59↓2.37/9.42↓4.02
11-Lc14.04↓2.8717.03↓8.7510.32↓2.579.04↓1.8210.21↓4.797.92↓2.52
12-Ld17.05↓5.888.56↓0.288.01↓0.2610.13↓2.915.67↓0.255.65↓0.25
+ +Table 4. Quantitative ablation studies. We remove each of our components here to show their contribution to our framework. Full denotes our complete model. CVI represents our whole cross-view interaction network. Other notations are consistent with Sec. 3. We report the errors of single-view outputs (Single, $M$ ), cross-view interaction outputs (Interact, $M^{*}$ ), and multi-view fusion results (Fusion, $\tilde{M}$ ). + +ods. Such methods can serve as a reference for evaluating the effectiveness of self-supervised methods. We adapt fully-supervised multi-view 3D pose estimation methods LT [25] and EpipolarTrans [22] to a self-supervised manner. Under self-supervised settings, EpipolarTrans can only achieve limited performance improvements compared to traditional methods. LT-Algebraic [25], which incorporates learnable confidence into the triangulation. LT-Volumetric model [25], which unprojects 2D features into a 3D volume for inference, achieves better results, but the performance is dependent on the accuracy of the hand center. Canon-Pose [52] and EpipolarPose [27] obtain multi-view inference results through simple averaging like ours. + +However, both of these methods are inferior to ours because they lack cross-view interaction. As our method predicts the root-relative 3D pose, we need to conduct post-processing to obtain the absolute coordinates. We introduce two different ways to achieve this: 1) using the 2D predictions of different views to triangulate and refine a center and 2) conducting RANSAC triangulation using our 2D predictions. Both methods have their merits. Opt-center can keep the root-relative results with hand prior, resulting in low PA-MPJPE. RANSAC gets better joint-wise accuracy, which is indicated by low MPJPE. We also provide qualitative results in the supplementary materials on the Assembly101 [46] dataset, which has a static camera setup. Even for challenging head-mounted moving cameras, we achieve convincing 3D pose estimates on the H2O [29] dataset. The experiments show that we have significantly pushed the performance of self-supervised methods to a comparable level with fully supervised methods. + +![](images/26046fcf4f4bcecb6c44fde775f429b259be0d31b838736a46707e4fda84821c.jpg) +Figure 3. Error of using different (a) #training data, (b) (line-1)#view for training, and (line-2)#view for inference when trained with 8 views. + +![](images/449db25627e75bc0694bfe003d835e8df4b1a72a287df96bdbec5ca3fc9172b6.jpg) +Figure 4. AUC of three 2D joint sets. O, S, I, PE denote OpenPose, single-view, interaction, and average pixel error in resolution $256 \times 256$ . + +# 4.4. Qualitative Result + +Fig. 5 presents the visual comparisons of 2 views between 2D joints of OpenPose, ours, and ground-truth on the HanCo dataset. We can observe that our method is more robust for outliers and can generate predictions close to the labels. Fig. 6 shows the 3D predictions from two viewpoints of ours, EpipolarPose, and CanonPose on the HanCo dataset. The results indicate that our method can get more accurate results especially when the occlusions are severe. Please refer to supplementary materials for more results. + +# 4.5. Ablation Study + +As shown in Tab. 4, we conduct comprehensive ablation experiments on the HanCo [62] dataset to show the effectiveness of each component. Single, Interact and Fusion denotes the evaluation of $M$ , $M^{*}$ and $\tilde{M}$ respectively. + +Different backbones. We first show our performance with different backbones. As shown in #1 and #2, using a large backbone like Res50, our performance can be further improved. For efficiency, we conduct ablation studies using Res18 as the backbone unless otherwise specified. + +Two branches for cross-view interaction module. As presented in #3 and #4, both of the branches can reduce the error. VSF can explicitly model the view-shared information and add reliable information from every view. CVA can capture the self-/cross-view joint-level correlations. + +Graph features. The results indicate that three kinds of features (#5, #6, #7) all lead to performance improvement. Especially, local feature (#7, $G^3$ ) can notably reduce the error after the interaction by providing fine-grained details. DCVI. We also conduct experiments to show the impor + +tance of DCVI by removing it and posing consistency constraints in single-view outputs like [24]. In this way, the performance drops dramatically (#10), proving the necessity of using DCVI to capture the features of all the views for self-supervised learning. + +Two branches for multi-view consistency loss. Without enforcing cross-view interaction outputs to be consistent, the performance significantly drops (#9). If we do not explore relatively more reliable 2D predictions to enhance consistency, the performance can also get worse (#8). + +Consistency losses. $(L_{c})$ The performance is unsatisfactory (#11) when employing the cross-view interaction network without any consistency constraints (i.e. discard #8 and #9). The interaction network should cooperate with consistency so that the constraints can guide the network to exploit multi-view information to function better. + +Multi-view distillation loss. $(L_{d})$ Removing the multiview distillation loss, all the metrics drop by a large margin (#12), especially in single-view estimation accuracy. This phenomenon proves the effectiveness of collaborative learning between single- and multi-view networks. + +# 4.6. Model Analysis + +Different percentage of unlabeled images. Fig. 3 (a) shows our method can get consistent performance improvement as the unlabeled training data increases. + +Different view number for training. The line-1 in Fig. 3 (b) shows the performance of our method tested on a certain view when trained with different view numbers. The curve shows that our method can be consistently improved as the number of views increases. We also observe that using multiple views for training can significantly improve performance when the valid views are few. + +Different view number for inference. Our model allows inferring with an arbitrary number of views. However, when the model is trained with a fixed view number, it could get the view number bias, resulting in better performance using the view number close to the training one. To avoid this, we add random masks in our interaction module and finetune the model for a few epochs. After that, results can get better by a small margin (the single-view error is $11.07\mathrm{mm}$ and the fusion error $7.60\mathrm{mm}$ , both in NMPJPE.). The line-2 in Fig. 3 (b) shows results on a certain view when trained on 8 views and tested on 1 to 8 views. We can observe consistent improvement with the inference view number increases. + +Different 2D joint sets. Fig. 4 presents the accuracy of different 2D joint sets on the HanCo training and testing set. Our 2D predictions are extremely better than OpenPose 2D pseudo label used for training. + +Iteratively training. Our approach can use the previous predictions as pseudo labels for iterative training. We find it helpful till iteration 3 and get saturated afterward. From 1 to 3 iterations, NMPJPE is 7.75, 7.68, and 7.64. + +![](images/151d8d96736515d6e29262c0f139423473de557de666646edd5d7ad948251c85.jpg) +Figure 5. 2D prediction (overlayed in the images) comparisons between OpenPose, ours, and ground-truth on the HanCo dataset. + +![](images/f8c92cc4e139cbe7c3219b5eaa6809fbb355cf757c3bb4b080cf7876b7e35f8c.jpg) +Figure 6. 3D prediction comparisons between our method, EpipolarPose, and CanonPose on the HanCo dataset. Our prediction and ground-truth are shown in solid red and dashed green respectively. + +# 5. Conclusion and Future Work + +To our best knowledge, we present the first self-supervised framework that aims to learn a single-view 3D hand estimator from unlabeled multi-view data. At the core of our approach, a cross-view interaction network is carefully designed to supervise the single-view output by leveraging the collaboration among multi-views. Specifically, the network captures the interdependencies of features among different views, resulting in improved accuracy of hand pose estimation after cross-view interaction. Additionally, the multi-view results are fused to supervise the single-view output for self-distillation. The effectiveness and versatility of the proposed framework are extensively evaluated through experiments, which demonstrate that our method not only establishes a new benchmark for self-supervised 3D hand pose estimation from single-view input but also offers flexible multi-view inference with state-of-the-art performance. + +We focused on hand pose estimation without heavy occlusions in this work. Extending our work to more challenging scenarios, such as hand-object interaction or relaxing the synchronization constraints in multi-view inputs, would be interesting topics for further study. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grants (62071067, 62171057, 62201072), in part by the Ministry of Education and China Mobile Joint Fund (MCM20200202), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039). + +# References + +[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2, 3 +[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2 +[3] Kristijan Bartol, David Bojanic, Tomislav Petković, and Tomislav Pribanić. Generalizable human pose triangulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11028-11037, 2022. 3 +[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 2, 3 +[5] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 2, 3 +[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 2, 6 +[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 2, 3, 6 +[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 2, 3 +[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2, 3, 5, 6 +[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2 +[11] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 7792-7801, 2019. 6 +[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 2, 4 +[13] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 2021 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2021. 2 +[14] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 2 +[15] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1991-2000, 2017. 2 +[16] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 2 +[17] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 5 +[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2 +[19] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Trans. Graph., 39(4):87, 2020. 1 +[20] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, Cabezas Randi, Tran Luan, Akbay Muzaffer, Yu Tsz-Ho, Keskin Cem, and Wang Robert. Umetrack: Unified multi-view end-to-end hand tracking for vr. ACM Transactions on Graphics, 2022. 1 +[21] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 6 +[22] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020. 3, 6, 7 +[23] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European Conference on Computer Vision (ECCV), pages 118-134, 2018. 2 +[24] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images + +in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5243-5252, 2020. 1, 3, 8 +[25] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2019. 3, 6, 7 +[26] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, 2014. 2 +[27] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1077-1086, 2019. 1, 2, 3, 5, 6, 7 +[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990–5000, 2020. 2, 6 +[29] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 5, 7 +[30] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 2 +[31] Moran Li, Yuan Gao, and Nong Sang. Exploiting learnable joint groups for hand pose estimation. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1921-1929, 2021. 2 +[32] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 2 +[33] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 2 +[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, 2019. 5 +[35] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In European Conference on Computer Vision, pages 424-442. Springer, 2022. 3 +[36] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 2 +[37] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and + +mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 2 +[38] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2 +[39] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 49-59, 2018. 2, 3 +[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in python. In NIPS 2017 Workshop on Autodiff, 2017. 5 +[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017. 3 +[42] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4342-4351, 2019. 3 +[43] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6040-6049, 2020. 3 +[44] Helge Rhodin, Jörg Spörri, Isinsu Katircioglu, Victor Constantin, Frédéric Meyer, Erich Müller, Mathieu Salzmann, and Pascal Fua. Learning monocular 3d human pose estimation from multi-view images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8437-8446, 2018. 3 +[45] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2, 3 +[46] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 5, 7 +[47] Hui Shuai, Lele Wu, and Qingshan Liu. Adaptive multi-view and temporal fusing transformer for 3d human pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3 +[48] Adrian Spurr, Umar Iqbal, Pavlo Molchanov, Otmar Hilliges, and Jan Kautz. Weakly supervised 3d hand pose estimation via biomechanical constraints. In European Conference on Computer Vision, pages 211-228. Springer, 2020. 2, 3 + +[49] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2 +[50] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2 +[51] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 3 +[52] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13294-13304, 2021. 1, 2, 3, 5, 6, 7 +[53] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 4 +[54] Chao Wen, Yinda Zhang, Zhuwen Li, and Yanwei Fu. Pixel2mesh++: Multi-view 3d mesh generation via deformation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1042-1051, 2019. 4 +[55] Linlin Yang, Shicheng Chen, and Angela Yao. Semihand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2, 3 +[56] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2335-2343, 2019. 2 +[57] Linlin Yang and Angela Yao. Disentangling latent hands for image synthesis and pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9877-9886, 2019. 2 +[58] Xiong Zhang, Hongsheng Huang, Jianchao Tan, Hongmin Xu, Cheng Yang, Guozhu Peng, Lei Wang, and Ji Liu. Hand image understanding via deep multi-task learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11281-11292, 2021. 2 +[59] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2, 3 +[60] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129(3):703-718, 2021. 3 +[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 4 + +[62] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In DAGM German Conference on Pattern Recognition, pages 250-264. Springer, 2021. 1, 2, 5, 6, 7 +[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2, 3 +[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2, 5 \ No newline at end of file diff --git a/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/images.zip b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..585c65130e43a09d53140c220978f2ec4f808a92 --- /dev/null +++ b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a512b039e42ebf654b3b8833a8abc095b963377bdcd07da662adefba296cc8c3 +size 481356 diff --git a/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/layout.json b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..962f7ed8d905ab8ee8653d62ee8717c94be31dbb --- /dev/null +++ b/2023/HaMuCo_ Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning/layout.json @@ -0,0 +1,9459 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 102, + 487, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 102, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 102, + 487, + 140 + ], + "type": "text", + "content": "HaMuCo: Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "content": "Xiaozheng Zheng " + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "content": " Chao Wen " + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "content": " Zhou Xue " + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "content": " Pengfei Ren " + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "content": " Jingyu Wang " + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "content": " State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications \n" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 165, + 536, + 224 + ], + "type": "text", + "content": " PICO IDL, ByteDance, Beijing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 276, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 290, + 540 + ], + "type": "text", + "content": "Recent advancements in 3D hand pose estimation have shown promising results, but its effectiveness has primarily relied on the availability of large-scale annotated datasets, the creation of which is a laborious and costly process. To alleviate the label-hungry limitation, we propose a self-supervised learning framework, HaMuCo, that learns a single-view hand pose estimator from multi-view pseudo 2D labels. However, one of the main challenges of self-supervised learning is the presence of noisy labels and the \"groupthink\" effect from multiple views. To overcome these issues, we introduce a cross-view interaction network that distills the single-view estimator by utilizing the cross-view correlated features and enforcing multi-view consistency to achieve collaborative learning. Both the single-view estimator and the cross-view interaction network are trained jointly in an end-to-end manner. Extensive experiments show that our method can achieve state-of-the-art performance on multi-view self-supervised hand pose estimation. Furthermore, the proposed cross-view interaction network can also be applied to hand pose estimation from multi-view input and outperforms previous methods under the same settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "content": "3D hand pose estimation is essential in various application scenarios, from action recognition and sign language translation to AR/VR [19, 20]. Hand pose estimation has achieved a significant improvement in recent years. However, the progress heavily relies on the emergence of many hand pose datasets with accurate 3D annotations. Acquiring labeled datasets is quite time-consuming and laborious, exposing a realistic challenge for deep learning models to learn with limited and noisy data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 673, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 673, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 673, + 287, + 685 + ], + "type": "text", + "content": "Self-supervised learning is an emerging solution to the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 251, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 546, + 396 + ], + "type": "text", + "content": "challenge posed by manual annotation. Though worth exploring, self-supervised pose estimation with RGB hand images is a challenging and relatively unexplored area with only one pioneering method, S²HAND [9]. S²HAND aims to conduct 3D hand reconstruction from a single RGB image with the noisy off-the-shell 2D hand pose estimation results for supervision. Unfortunately, S²HAND faces a predicament where its performance is significantly reliant on the quality of the pseudo label, and inferior labeling may result in reduced performance. Moreover, evaluating the quality of the pseudo label is an ill-posed problem that lacks clear criteria or input, further complicating the issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 397, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 546, + 518 + ], + "type": "text", + "content": "This observation motivates us to leverage multi-view information for enhancing self-supervised learning, as the complementary nature of multi-view observations can help mitigate the ambiguity inherent in pose estimation. Although the first 3D hand dataset with synchronized multiview input (HanCo [62]) was proposed in 2021, to our knowledge, there is no previous work exploring the potential of multi-view for self-supervised hand pose estimation. Therefore, we turn to studies in the human body pose estimation, which share some similarities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 519, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 628 + ], + "type": "text", + "content": "As mentioned in previous work [24], naively enforcing multi-view consistency is prone to generate degenerated solutions, thus they resorted to additional 2D labels of unrelated datasets and proposed a solution under the scope of weakly supervised learning. Other studies, such as EpipolarPose [27] and CanonPose [52], utilized multiview data with special designs to enhance the supervision and achieved promising results under the scope of self-supervised learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we push along this direction on hand pose estimation via multi-view collaborative learning. We take one step further by designing a learnable network, which utilizes multi-view information, to tackle 1) noisy pseudo labels and 2) unreliable multi-view \"groupthink\" issues causing training collapse in the early training stage. Formally, we name the pipeline HaMuCo, which stands for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 692, + 271, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 271, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 271, + 703 + ], + "type": "text", + "content": "Project page: https://zxz267.github.io/HaMuCo." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 703, + 209, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 703, + 209, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 703, + 209, + 713 + ], + "type": "text", + "content": "* Corresponding author. † Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20763" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 58, + 171, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 58, + 171, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 58, + 171, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 58, + 171, + 216 + ], + "type": "image", + "image_path": "7d09d4158ed2d1c0064b9952d4d2abe7ab9202033f3a8228437bc64ff9e29170.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 219, + 547, + 252 + ], + "lines": [ + { + "bbox": [ + 45, + 219, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 45, + 219, + 547, + 252 + ], + "type": "text", + "content": "Figure 1. Overall pipeline comparison: HaMuCo learns a monocular 3D hand pose estimator from multi-view self-supervision via cross-view feature interaction. Our cross-view interaction network addresses the importance of introducing learnable feature interaction, which is absent from previous methods [27, 52]. At inference time only the gray part is applied." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 57, + 306, + 215 + ], + "blocks": [ + { + "bbox": [ + 187, + 57, + 306, + 215 + ], + "lines": [ + { + "bbox": [ + 187, + 57, + 306, + 215 + ], + "spans": [ + { + "bbox": [ + 187, + 57, + 306, + 215 + ], + "type": "image", + "image_path": "b4d5b31d5e12985556bbef5fbb46b83d25db8a8469019971dfbeeecc03650dc4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 321, + 58, + 537, + 215 + ], + "blocks": [ + { + "bbox": [ + 321, + 58, + 537, + 215 + ], + "lines": [ + { + "bbox": [ + 321, + 58, + 537, + 215 + ], + "spans": [ + { + "bbox": [ + 321, + 58, + 537, + 215 + ], + "type": "image", + "image_path": "a62fda65e2c2842ade79d1f653ca1f18544362d5c308ca52374529863b39cfdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 209, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 209, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 209, + 268 + ], + "type": "text", + "content": "Hand Multiview Collaborative learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 269, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 269, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 45, + 269, + 287, + 422 + ], + "type": "text", + "content": "The core idea of our approach is to enhance the single-view estimation by means of cross-view feature interaction and further integrate multi-view results to supervise the single-view output to achieve self-distillation in an end-to-end fashion. Thus, our framework is built with a single-view hand pose estimator and a cross-view interaction network for supervision. The single-view estimator uses the MANO [45] hand model as the decoder, which provides the hand prior to regularizing irrational anatomy when supervised by noisy pseudo labels. The cross-view interaction network captures cross-view features and utilizes several consistent losses among different views to guide collaborative learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 424, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 288, + 567 + ], + "type": "text", + "content": "We conduct comprehensive experiments on the HanCo [62] dataset and our approach outperforms previous methods by a considerable margin for self-supervised 3D hand pose estimation. Notably, our results demonstrate competitive performance compared to a state-of-the-art fully supervised approach proposed by Chen et al. [7]. Our proposed framework is highly versatile, as it can be trained with or without calibration, and is capable of incorporating the cross-view interaction network to achieve superior multi-view inference results when multi-view test data is available. Moreover, our model can generalize well to other datasets [29, 46, 64] and in-the-wild images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 567, + 256, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 567, + 256, + 579 + ], + "spans": [ + { + "bbox": [ + 59, + 567, + 256, + 579 + ], + "type": "text", + "content": "In summary, our contributions are the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 584, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 58, + 584, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 584, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 58, + 584, + 287, + 633 + ], + "type": "text", + "content": "- We propose the first self-supervised learning framework for single-view hand pose estimation without any training data annotation and achieve state-of-the-art performance via multi-view collaborative learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 637, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 637, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 637, + 288, + 685 + ], + "type": "text", + "content": "- We propose a cross-view interaction network to supervise the single-view estimator by enforcing multi-view consistency and capturing cross-view features for collaborative learning among multiple views." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 288, + 714 + ], + "type": "text", + "content": "- The proposed framework is capable of multi-view inference by incorporating the cross-view interac" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 256, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 256, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 325, + 256, + 545, + 279 + ], + "type": "text", + "content": "tion network and achieves state-of-the-art performance without bells and whistles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 289, + 393, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 393, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 393, + 301 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 307, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 713 + ], + "type": "text", + "content": "Hand Pose Estimation. Hand pose estimation can be categorized into RGB-based methods [23, 49, 63] and depth-based methods [14, 15, 36], depending on the input modality. In this paper, we focus our attention on RGB-based hand pose estimation. The RGB-based methods can be further divided into three categories, skeleton-based methods [5, 12, 23, 31, 38, 39, 48, 49, 55-57, 63], model-based methods [1, 2, 4, 9, 58, 59, 64], and mesh-based methods [7,8,10,16,28,30,32,33,37,50,61]. Skeleton-based methods regress the hand joints directly. Zimmermann et al. [63] introduces a multi-stage network that lifts the regressed 2D joints to 3D ones. Variational autoencoder [26] is employed to learn a cross-modal latent space to achieve better hand pose estimation and disentanglement [49, 56, 57]. Latent 2.5D representation regression is proved more effective than direct coordinates regression for hands by [23], which is also adopted by [13, 31, 48, 61]. There are also many works solving hand pose estimation with two hands interactions [13, 29, 38] and hand-object interactions [2, 12, 29]. Recent model-based methods make use of MANO [45], which can incorporate the hand prior and predict the hand mesh simultaneously. Those methods [1, 4, 9, 58, 59] rely on additional supervisions [1, 4, 9, 58, 59] or inputs [4]. In contrast, mesh-based methods regress each vertex directly, which is more accurate but requires large-scale datasets with hand mesh annotations [18, 29, 38, 64]. Most of these methods utilize graph convolutional network (GCN) [7,8,10,16,28,50,61] or transformers [32] or both [30, 33] for regression. I2L-MeshNet [37] regresses each vertex by predicting 1D heatmaps of three axes. Chen et al. [6] uses an image-to-image translation network to predict the UV map of the mesh. Similar to previous works [30, 33], we also use transformer and GCN. However, we employ them for cross-view interaction." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20764" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 191 + ], + "type": "text", + "content": "Multi-View Fully-Supervised Pose Estimation. Multiview information is widely explored to improve 3D human pose estimation by tackling occlusions and depth-ambiguity in a fully-supervised manner [3, 22, 25, 41-43, 47, 60]. Volume-based methods [25, 41, 42, 51] unproject 2D features or heatmaps of joints to a 3D space for estimation. Another kind of method [22, 43, 60] utilizes the geometry information to fuse the features in 2D space directly and efficiently. Recently, some works [35,47] utilize transformers for implicit cross-view fusion without camera extrinsics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 193, + 286, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 286, + 372 + ], + "type": "text", + "content": "Label-Efficient Learning. Label-efficient learning aims to reduce the 3D label requirements. Many works devote to solving hand pose estimation in a label-efficient manner [1, 4,5,7,9,39,48,55,59,63]. Synthetic data is used to avoid manual annotation [7,39,63], but may need domain transfer [39]. Or use weakly supervised learning [4,48] to obtain 3D results by manually annotating 2D labels to assist with hand priors. Multi-view label-efficient learning is also explored in 3D pose estimation [24, 27, 44, 52]. Rhodin et al. [44] trains a semi-supervised network with only a small amount of labeled 3D data and multi-view consistency constraints. Iqbal et al. [24] mixes single-view images with 2D labels and unlabelled multi-view images for training. Our goal is the same as that of previous methods, which is to train without any manual 3D labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 373, + 286, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 286, + 636 + ], + "type": "text", + "content": "Self-supervised 3D Pose Estimation. (1) Single-view training and inference. To the best of our knowledge, there is only one method for self-supervised 3D hand pose estimation, proposed by Chen et al. [9]. Their framework, " + }, + { + "bbox": [ + 47, + 373, + 286, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^2\\mathrm{Hand}" + }, + { + "bbox": [ + 47, + 373, + 286, + 636 + ], + "type": "text", + "content": ", uses only single-view 2D noisy labels for training and achieves self-supervision through rendering. However, the performance is limited due to the use of single-view information and the quality of the noisy labels. (2) Multi-view training, single-view inference. Our approach belongs to this category but is fundamentally different from the existing methods. EpipolarPose [27] triangulates multi-view 2D pseudo labels according to epipolar geometry to 3D ones for training. CanonPose [52] learns to lift 2D pseudo labels to 3D canonical pose space with multi-view consistency constraints. All the aforementioned methods use nonlearnable self-supervised modules like geometric modules or consistency loss functions, as shown in Fig. 1. However, they [27, 52] ignore the importance of introducing cross-view interaction and multi-view collaborative learning. Previous methods struggle to achieve good performance since the pose of a hand can change drastically over time and different joints may have similar appearances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 647, + 102, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 102, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 102, + 659 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "As depicted in the left part of Fig. 2, our framework consists of a simple yet effective single-view estimator and cross-view interaction network. The core idea of our approach is that prediction from a monocular view can be" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "enhanced via cross-view feature interaction and the interacted results can further supervise the single-view output to achieve self-distillation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 113, + 433, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 113, + 433, + 125 + ], + "spans": [ + { + "bbox": [ + 306, + 113, + 433, + 125 + ], + "type": "text", + "content": "3.1. Single-View Estimator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "text", + "content": "Overview. Our framework takes multi-view synchronized hand images " + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_i\\}_{i=0}^v" + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "text", + "content": " views as input, each view is an image of " + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "inline_equation", + "content": "I_i \\in \\mathbb{R}^{3 \\times h \\times w}" + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "text", + "content": ". The output is a 3D hand mesh " + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 129, + 545, + 236 + ], + "type": "text", + "content": " on each view. We designed a simple yet effective model-based network as a single-view estimator. Using the hand model will reduce the adverse effects of using poor pseudo labels as supervision by providing hand prior information for regularization. Please refer to supplementary materials for more details about the single-view estimator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": "Hand model. We employ MANO [45] as the hand model. The hand mesh can be derived from the MANO layer using parameters " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "M(\\beta, \\pmb{\\theta})" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\mathbb{R}^{16 \\times 3}" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": " control the shape and pose of the hand respectively. We can use a predefined regressor to obtain the 3D joints from the 3D mesh vertices by " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "P = JM" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "J \\in \\mathbb{R}^{k \\times n}" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "n = 778" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "inline_equation", + "content": "k = 21" + }, + { + "bbox": [ + 305, + 236, + 545, + 331 + ], + "type": "text", + "content": " are the joints number and vertices number. For more details, we recommend referring to [45]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "text", + "content": "Camera model. Following Boukhayma et al. [4], we model the geometry correspondence by the weak-perspective camera model and obtain camera parameters from the single-view network predictions. Given the translation " + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "text", + "content": ", the 2D coordinates in image plane can be obtained by: " + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\Pi(P) = s\\Omega(P) + \\mathbf{t}" + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "text", + "content": " is the orthographic projection and " + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 305, + 332, + 545, + 416 + ], + "type": "text", + "content": " denotes the weak-perspective projection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": "Network Structure. Since the single-view estimator is not the main component, for the sake of simplicity, we employ a CNN as the encoder " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "F_{e}" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": ", and an MLP as the decoder " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": " for regressing the MANO parameters. We have 3D hand mesh: " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "M_{i}(\\pmb{\\theta}_{i},\\pmb{\\beta}_{i}) = F_{s}(\\pmb{I}_{i})" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "F_{s} = F_{d}(F_{e}(\\cdot))" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": " denotes the entire single-view network. The estimator also passes different levels of features " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "H^{j}" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "H^{j}" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": " is the intermediate feature of the encoder after " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": " residual blocks, " + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "inline_equation", + "content": "j = 1,2,3,4" + }, + { + "bbox": [ + 305, + 416, + 545, + 523 + ], + "type": "text", + "content": ") to our cross-view interaction network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 532, + 478, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 478, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 478, + 544 + ], + "type": "text", + "content": "3.2. Cross-view Interaction Network" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 550, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 550, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 545, + 634 + ], + "type": "text", + "content": "In this section, we introduce the cross-view interaction network (CVI-Net), which is the core of our system to enable the network to exploit multi-view information. This stage conducts cross-view interaction and distillation. The critical components of this stage are a cross-view interaction network for capturing cross-view features and several consistent losses for guiding collaborative learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 639, + 507, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 507, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 507, + 651 + ], + "type": "text", + "content": "3.2.1 View-Shared Graph Feature Extraction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "The first step for interaction is to extract the appropriate features. Different from [7, 8, 61], our module collects useful information into a graph through view-shared graph feature extraction module (VSGFE) as shown in Fig. 2. Specifically, it makes use of multi-level feature maps from dif" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20765" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 212, + 277 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 212, + 277 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 212, + 277 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 212, + 277 + ], + "type": "image", + "image_path": "16227f54fe24ebe33674bbd0efe6b4c000432e52d3ca5015ba66cff1849a7dae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": "Figure 2. The left illustrates our whole pipeline (2 views for simplicity). During the training phase, the network takes multi-view hand images and pseudo-labels as inputs. The bottom right depicts our cross-view interaction networks. The top right shows the view-shared graph feature extraction (VSGFE) module and view-shared feature (VSF) module. " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": " denotes add and concatenation respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 403, + 167 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 403, + 167 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 403, + 167 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 403, + 167 + ], + "type": "image", + "image_path": "51232d8a0153c6a748b5b783aab649a111933bd0e9ea6629fb5f51fe1e8ef874.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 406, + 71, + 523, + 167 + ], + "blocks": [ + { + "bbox": [ + 406, + 71, + 523, + 167 + ], + "lines": [ + { + "bbox": [ + 406, + 71, + 523, + 167 + ], + "spans": [ + { + "bbox": [ + 406, + 71, + 523, + 167 + ], + "type": "image", + "image_path": "093c5d08c4d48f11e75ee9a85cefff1ae12f311019caad37dab2ef703bdd0a23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 169, + 517, + 274 + ], + "blocks": [ + { + "bbox": [ + 221, + 169, + 517, + 274 + ], + "lines": [ + { + "bbox": [ + 221, + 169, + 517, + 274 + ], + "spans": [ + { + "bbox": [ + 221, + 169, + 517, + 274 + ], + "type": "image", + "image_path": "bb20bfda8d5d2cd0c90a2689039651ef6d202478817830e2b2021fd950055849.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": "ferent views " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = \\{\\pmb{H}_i^j\\}_{i=0}^v" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": ", 3D joints " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\pmb{P}_i\\}_{i=0}^v" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": ", and MANO pose parameters " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\pmb{\\theta}_i\\}_{i=0}^v" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " from the single-view estimator to extract a graph feature " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": ". The graph feature of each view " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " consists of three parts. " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i^1" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i^2" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i^3" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " aim to capture joint location features, global image features, and local image features, respectively. The first part is joint location embedding " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i^1 \\in \\mathbb{R}^{k \\times c_1}" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": ", providing the explicit geometric information. This embedding is obtained by using an MLP to map the single-view 3D joints locations " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{P}_i" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " and pose parameters " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_i" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " to dimension " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": ". The second part is joint-wise high-level image features " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i^2 \\in \\mathbb{R}^{k \\times c_2}" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " generated by spatial-aware initial graph building (SAIGB) [61] module using the last level feature maps " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{H}_i^4" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": ". This part provides compact image clues of all views for interaction. The third part is joint-aligned features " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i^3 \\in \\mathbb{R}^{k \\times c_3}" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " gathered by joint feature sampler (JFS). JFS projects joints onto multi-level image feature maps " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\{\\pmb{H}_i^j\\}_{j=1}^3" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": " to gather fine-grained perceptual features like [53,54] for better local alignment. We then concatenate graph features to get " + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{G}_i = [\\pmb{G}_i^1 \\otimes \\pmb{G}_i^2 \\otimes \\pmb{G}_i^3]" + }, + { + "bbox": [ + 46, + 331, + 289, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 566, + 272, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 272, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 272, + 578 + ], + "type": "text", + "content": "3.2.2 Dual-Branch Cross-View Interaction (DCVI)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "We first stack " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{G_i\\}_{i=0}^v" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " of all views to obtain multi-view graph feature " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{vk \\times (c_1 + c_2 + c_3)}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". We design a component to effectively capture complementary information from other views on multi-view graph feature " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". The interaction module has two branches, (1) cross-view attention branch (CVA) and (2) view-shared feature branch (VSF). Cross-view attention branch utilizes a cross-view transformer " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "F_t" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " consisting of several multi-head attention layers with token size " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "vk" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and MLPs, which allows each joint to aggregate features from other joints or views. This branch implicitly captures the multi-view information. An explicit" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": "multi-view prior information is that the observed poses from all the views should be consistent in 3D. Therefore, we add a branch to excavate the multi-view shared information to enhance the feature representation. Specifically, view-shared feature branch first employs adaptive-GCN [12] " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": " to map the view-specific features " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{G}_{i}" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": " to a canonical feature space " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "C_i = F_a(G_i)" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": ", the nodes in adaptive-GCN represents the hand joints and the edges represent joint feature correlation. Then, we stack " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{C_i\\}_{i = 0}^v" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": " together to get multi-view canonical features " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{C} \\in \\mathbb{R}^{v \\times k \\times (c_1 + c_2 + c_3)}" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": ". After that, we use max-pooling on " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{C}" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": " to get the max activated features of every joint then repeat them in the view dimension as the view-shared features " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{C}' \\in \\mathbb{R}^{vk \\times (c_1 + c_2 + c_3)}" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": ". We denote the dual-branch cross-view interaction as: " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{*} = \\pmb{G} + F_{t}(\\pmb{G}) + \\pmb{C}^{\\prime}" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{*}" + }, + { + "bbox": [ + 304, + 331, + 547, + 511 + ], + "type": "text", + "content": " is the updated graph feature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "content": "Parameters regression. The view specific feature " + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "inline_equation", + "content": "G_{i}^{*}" + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "content": " after the interaction can be obtained by reshaping " + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "inline_equation", + "content": "G^{*}" + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "content": ". We then employ a shared MLP " + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "content": " as a decoder to regress the pose parameters " + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{*} = F_{r}(G_{i}^{*})" + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "content": " to derive the hand mesh of each view " + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "inline_equation", + "content": "M_{i}^{*}(\\theta_{i}^{*},\\beta_{i})" + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "content": " and corresponding joints " + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "inline_equation", + "content": "P_{i}^{*} = JM_{i}^{*}" + }, + { + "bbox": [ + 304, + 511, + 547, + 572 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 575, + 491, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 491, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 491, + 588 + ], + "type": "text", + "content": "3.2.3 Multi-View Collaborative Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "text", + "content": "To allow all the views and the networks to learn collaboratively, we utilize consistency losses " + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "text", + "content": " upon interaction outputs and distillation loss " + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "text", + "content": " between multi-view fusion results and single-view outputs, as shown in Fig. 2. " + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "text", + "content": " introduces collaborative learning between multiple views, guiding the poses from different views to be as close as possible. While " + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 304, + 594, + 545, + 700 + ], + "type": "text", + "content": " makes the CVI-Net and single-view estimator work in a collaborative manner, achieving a self-distillation effect." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Results fusion. Since we need to supervise the single-view" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20766" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "estimator with the results after the interaction, instead of simply using the refined results " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "M_{i}^{*}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " of each view, we ensemble all the results into a unified and more reliable result " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". Considering the lack of explicit guidance, we empirically introduce a prior that all the views contribute equally. Thus, we simply average all aligned results to obtain " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". Specifically, we use " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " to denote the align procedure. When the extrinsics are known, we use the relative camera pose for alignment. When the camera extrinsics are unavailable, we use Procrustes analysis [62, 64] to compute relative rotation and align meshes to a canonical view. The final result is calculated as follows: " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\tilde{M} = \\frac{1}{v}\\sum_{i=1}^{v}A(M_{i}^{*})" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": "Consistency losses. We design two types of consistency loss " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": ": 2D consistency loss " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_{2D}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " and Fusion consistency loss " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_f}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": ". The motivation behind " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_{2D}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " is that the 2D predictions in the x-axis and y-axis are more accurate than the depth prediction in the z-axis. Therefore, " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_{2D}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " utilizes the 2D predictions in every single view as the pseudo label to supervise other views, which explores the view-specific reliable information to collaboratively improve the predictions of all the views. 2D consistency loss is defined as: " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_{2D}} = \\frac{1}{v^2}\\sum_{i=1}^{v}\\sum_{j=1}^{v}\\|\\Pi(M_i^*) - \\Pi(A_i(M_j^*))\\|_1" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "A_i(\\cdot)" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " denotes the alignment operation to align other view-j to view-i. Fusion consistency loss uses the fused results to supervise each view. The loss is defined as: " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_f} = \\frac{1}{v}\\sum_{i=1}^{v}\\|M_i^* - A_i^{-1}(\\tilde{M})\\|_1" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "A_i^{-1}(\\cdot)" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " denotes the inverse transformation from canonical view to view-i. " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_{2D}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_f}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " are complementary to each other. Only using " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_{2D}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " tends to get performance saturation faster. In contrast, only adopting " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_f}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " can lead to unstable training since there may exist the situation that the fusion results are worse due to the majority of the predictions being wrong, especially at the early training stage. During training, we alternately update " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_{2D}}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "inline_equation", + "content": "L_{c_f}" + }, + { + "bbox": [ + 46, + 217, + 289, + 483 + ], + "type": "text", + "content": " to achieve more stable optimization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": "Multi-view distillation loss. Since the multi-view fusion results are much better than the 2D pseudo label, we introduce multi-view distillation loss " + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "inline_equation", + "content": "L_{d} = \\frac{1}{v}\\sum_{i = 1}^{v}\\| M_{i} - A_{i}^{-1}(\\tilde{M})\\|_{1}" + }, + { + "bbox": [ + 47, + 483, + 287, + 544 + ], + "type": "text", + "content": " that uses the fusion results to supervise the single-view outputs to achieve self-distillation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "content": "Total loss. Except for the losses for multi-view collaborative learning, our framework also adopts two general constraints, 2D joints loss, and hand prior regularization. The prior regularization regularizes the pose and shape parameters: " + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "inline_equation", + "content": "L_{p} = \\frac{1}{v}\\sum_{i = 1}^{v}\\alpha (\\| \\pmb{\\theta}_{i}\\|_{1} + \\| \\pmb{\\theta}_{i}^{*}\\|_{1} + \\gamma \\| \\pmb{\\beta}_{i}\\|_{1})" + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "content": " are used to balance the loss scale. The 2D joints " + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "content": " is used to supervise the results from the 2D pseudo labels. The final loss is defined as: " + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "inline_equation", + "content": "L = L_{c} + L_{d} + L_{2D} + L_{p}" + }, + { + "bbox": [ + 46, + 545, + 287, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 655, + 128, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 128, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 128, + 668 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 672, + 168, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 168, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 168, + 685 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "FreiHAND [64] is a dataset for single-view 3D hand pose estimation, which contains 130,240 training images and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "3,960 testing images. All images are captured from the real world with 3D annotations. The training set consists of 32,560 composited images with four types of real-world backgrounds and hands captured against a green screen." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 120, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 547, + 192 + ], + "type": "text", + "content": "HanCo [62] extends FreiHAND, which consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we use the first 1,200 sequences for training and the last 317 sequences for testing in all experiments for fair comparisons." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 547, + 251 + ], + "type": "text", + "content": "Other datasets. We also provide additional results on other datasets. Assembly101 [46] is an action recognition dataset that consists of 4,321 videos sequence. H2O [29] is a handobject interaction dataset with 571,645 frames. Please refer to supplementary materials for details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 348 + ], + "type": "text", + "content": "Metrics. We report standard metrics for hand pose estimation as follows. (1) MPJPE/MPVPE (mean per joint/vertex position error) measures the average Euclidean distance in mm between the predicted and ground-truth joints/vertices. JE/VE are the abbreviations for MPJPE/MPVPE. (2) NMPJPE/NMPVPE (normalized mean per joint/vertex position error, N-JE/VE) computes MPJPE/MPVPE after performing translation and scale alignment. (3)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 348, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 444 + ], + "type": "text", + "content": "PA-MPJPE/PA-MPVPE (PA-JE/VE) is a modification of MPJPE/MPVPE with Procrustes analysis [17]. This metric normalizes the absolute scale, center, and rotation. (4) F-Score [9] is the harmonic mean of recall and precision between two meshes w.r.t. a specific distance threshold. F@5mm and F@15mm are reported. (5) AUC means the area under the curve of the PCK, where the PCK refers to the percentage of correct joints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 449, + 440, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 440, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 440, + 461 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "text", + "content": "We implement all the networks in PyTorch [40]. We first train our framework without " + }, + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "text", + "content": " for 10 epochs. Then, we train the whole framework for another 30 epochs. Each batch contains images from 8 time-step of 8 cameras. We use AdamW [34] optimizer and set the initial learning rate to 3e-4. We use " + }, + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 464, + 545, + 549 + ], + "type": "text", + "content": " hand images as input. Please refer to supplementary materials for more details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 554, + 492, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 492, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 492, + 567 + ], + "type": "text", + "content": "4.3. Comparisons with state-of-the-arts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": "In Sec. 4.3.1, we evaluate the performance of our method under the single-view inference setting. As self-supervised hand pose estimation is a relatively new task, there is limited literature available for comparison. To address this, we adapt self-supervised body pose estimation methods [27,52] to hand and compare them with our method on HanCo [62]. We then compare with the only existing self-supervised hand pose estimation method, S²Hand [9]. As S²Hand can only be trained on single-view images, we use our single-view network only (denote as Ours-SV) for both training and inference as baselines. We further conduct extensive evaluations of our full model and baselines to demonstrate" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20767" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 263, + 209 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 263, + 209 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 263, + 209 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 263, + 209 + ], + "type": "table", + "html": "
MethodInputN-JE↓PA-JE↓
Fully-Supervised Method:
MobRecon [7]image9.95.7
EpipolarPose [27]image10.56.1
Self-Supervised Method:
EpipolarPose [27]image,19.79.3
CanonPose [52]2D pose,30.912.6
Oursimage,11.17.0
EpipolarPose [27]image42.323.5
CanonPose [52]2D pose31.812.8
Oursimage15.27.7
", + "image_path": "abb7593f0c1777dc43c7f595a21e6f126f60e3d13283a61b76b0a508b4571159.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 285, + 61, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 41, + 217, + 270, + 262 + ], + "lines": [ + { + "bbox": [ + 41, + 217, + 270, + 262 + ], + "spans": [ + { + "bbox": [ + 41, + 217, + 270, + 262 + ], + "type": "text", + "content": "Table 1. Single-view inference comparisons on the HanCo [62] dataset. " + }, + { + "bbox": [ + 41, + 217, + 270, + 262 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 41, + 217, + 270, + 262 + ], + "type": "text", + "content": " denotes the method using camera extrinsics during training. Notably, in the self-supervised setting, our method exhibits a significant improvement over previous methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 285, + 61, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 285, + 61, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 285, + 61, + 547, + 220 + ], + "type": "table", + "html": "
MethodDataBackbonePA-JE↓PA-VE↓F@5↑
Fully-Supervised Method:
YoutubeHand [28]Frei.Res508.48.60.61
I2UV-HandNet [6]Frei.Res506.76.90.71
MobRecon [7]Frei.Res50†6.16.20.76
Ours-SVFrei.Res507.57.50.68
Self-Supervised Method:
S²HAND [9]Frei.EffiNet-b011.811.90.48
Ours-SVFrei.EffiNet-b011.611.70.49
Ours-SVFrei.Res5011.912.00.47
Ours-SVHanCoEffiNet-b011.311.40.51
Ours-SVHanCoRes5011.611.80.48
OursHanCoEffiNet-b06.36.80.71
OursHanCoRes506.26.70.72
", + "image_path": "cfc41ba9e416c3d6addd677eddb55c184f98ec8d201fc0464c7cd39bc185e88e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 272, + 244, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 244, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 244, + 284 + ], + "type": "text", + "content": "the efficacy of multi-view collaborative learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 357 + ], + "type": "text", + "content": "In addition, thanks to our cross-view interaction network, our approach is capable of performing multi-view inference by simply averaging individual view results when multi-view test data is available. In Sec. 4.3.2, we compare our method with state-of-the-art approaches under the multi-view inference setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 364, + 174, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 174, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 174, + 375 + ], + "type": "text", + "content": "4.3.1 Single-View Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 379, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 427 + ], + "type": "text", + "content": "Hanco. We train EpipolarPose and CanonPose using their open-source code. We also train fully-supervised methods [7, 27] as a reference for performance. Tab. 1 outlines the performance of fully-/self-supervised methods in the lit-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 77, + 435, + 258, + 679 + ], + "blocks": [ + { + "bbox": [ + 281, + 228, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 281, + 228, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 281, + 228, + 545, + 262 + ], + "type": "text", + "content": "Table 2. Quantitative results on the FreiHAND evaluation set. The notation " + }, + { + "bbox": [ + 281, + 228, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 281, + 228, + 545, + 262 + ], + "type": "text", + "content": " denotes using a stacked backbone structure. \"Our-SV\" refers to training only with our single-view network." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 435, + 258, + 679 + ], + "lines": [ + { + "bbox": [ + 77, + 435, + 258, + 679 + ], + "spans": [ + { + "bbox": [ + 77, + 435, + 258, + 679 + ], + "type": "table", + "html": "
MethodMPJPE ↓PA-MJPJE ↓
Ttraditional Triangulation Method (w/o training):
DLT [21]16.813.2
Pictorial [11]13.510.2
RANSAC [25]12.39.8
Fully-Supervised Method:
EpipolarTrans [22]6.24.2
LT-Algebraic [25]5.53.6
LT-Volumetric [25]5.83.6
LT-Volumetric+ [25]4.93.6
EpipolarPose+ [27]8.04.4
Ours (Opt-Center)6.03.2
Ours (RANSAC)5.83.4
Self-Supervised Method:
EpipolarTrans [22]11.29.0
LT-Algebraic [25]10.37.8
LT-Volumetric [25]10.68.0
LT-Volumetric+ [25]9.57.2
CanonPose+ [52]21.610.5
EpipolarPose+ [27]17.28.3
Ours (Opt-Center)8.85.3
Ours (RANSAC)8.55.6
", + "image_path": "395dd3cfa8328020a62332791c861dfaf427f3328415505110fb5d3f4898e464.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 687, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 687, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 287, + 710 + ], + "type": "text", + "content": "Table 3. Multi-view inference results on the HanCo dataset. The notation " + }, + { + "bbox": [ + 46, + 687, + 287, + 710 + ], + "type": "inline_equation", + "content": "{}^{ + }" + }, + { + "bbox": [ + 46, + 687, + 287, + 710 + ], + "type": "text", + "content": " indicates that methods require the GT 3D center." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 272, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 463 + ], + "type": "text", + "content": "erature along with ours. In the case where camera extrinsics are available for training, CanonPose performs the worst because it lifts noisy 2D pseudo labels from OpenPose to 3D ones. When camera extrinsics are not available, all competitors experience a performance decline. This is due to the lack of collaborative interaction across multi-view features in previous self-supervised methods. In contrast, our method outperforms both of them by a large margin. Our cross-view interaction networks can enhance single-view inference, whether camera extrinsics are available during training or not. More details about the usage of cameras can be found in Sec. 3.2.3. Compared to previous self-supervised methods, our approach significantly improves performance, highlighting the importance of cross-view interaction among different views. Moreover, our approach can get comparable results to fully-supervised methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "content": "FreiHAND. The comparisons on the evaluation set are shown in Tab. 2. The experiments conducted under self-supervised settings indicate that our baselines, Ours-SV, already achieve performance comparable to " + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^2\\mathrm{Hand}" + }, + { + "bbox": [ + 304, + 464, + 546, + 644 + ], + "type": "text", + "content": ". Moreover, directly equipping baselines with other backbones or more training data does not improve too much. We argue that performance improvements in single-view self-supervised hand pose estimation cannot be achieved by changing the backbone architecture or increasing the amount of training data. In contrast, our full model, i.e. Ours, substantially further improves the results on the Frei-Hand dataset, which justify the effectiveness of multi-view collaborative learning. Moreover, our self-supervised approach achieves competitive performance with recent fully-supervised state-of-the-art methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 649, + 430, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 430, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 430, + 660 + ], + "type": "text", + "content": "4.3.2 Multi-View Inference" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We show the quantitative results of our multi-view inference performance with other competitors on HanCo in Tab. 3. A naive solution is to triangulate pseudo labels without training. We show the performance of traditional meth-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20768" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 361, + 259 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 361, + 259 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 361, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 361, + 259 + ], + "type": "table", + "html": "
IDMethodNMPJPE ↓PA-MPJPE ↓
SingleInteractFusionSingleInteractFusion
1FullResNet-50 as the backbone:
11.14↑0.038.31↓0.037.65↑0.107.05↑0.175.35↑0.075.34↑0.06
2FullResNet-18 as the backbone:
11.178.287.757.225.425.40
3-VSF11.21↓0.048.49↓0.217.81↓0.067.25↓0.035.52↓0.105.50↓0.10
4-CVA11.31↓0.148.45↓0.177.81↓0.067.29↓0.075.48↓0.065.46↓0.06
5-G111.31↓0.148.56↓0.287.77↓0.037.31↓0.095.52↓0.105.49↓0.09
6-G211.33↓0.168.38↓0.107.83↓0.087.34↓0.085.45↓0.035.42↓0.02
7-G311.30↓0.138.99↓0.697.82↓0.077.30↓0.085.45↓0.035.44↓0.04
8-Lc2D11.25↓0.088.43↓0.157.90↓0.157.32↓0.105.58↓0.165.57↓0.17
9-Lcf11.74↓0.578.98↓0.708.38↓0.637.55↓0.335.84↓0.425.80↓0.40
10-DCVI13.52↓2.35/11.99↓4.249.59↓2.37/9.42↓4.02
11-Lc14.04↓2.8717.03↓8.7510.32↓2.579.04↓1.8210.21↓4.797.92↓2.52
12-Ld17.05↓5.888.56↓0.288.01↓0.2610.13↓2.915.67↓0.255.65↓0.25
", + "image_path": "16d276a85d6f993b47e021215e6a794e18c7a822dc55f124444bf7d536918a5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "lines": [ + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "spans": [ + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "type": "text", + "content": "Table 4. Quantitative ablation studies. We remove each of our components here to show their contribution to our framework. Full denotes our complete model. CVI represents our whole cross-view interaction network. Other notations are consistent with Sec. 3. We report the errors of single-view outputs (Single, " + }, + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "type": "text", + "content": "), cross-view interaction outputs (Interact, " + }, + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "type": "text", + "content": "), and multi-view fusion results (Fusion, " + }, + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 44, + 266, + 361, + 322 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 335, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 492 + ], + "type": "text", + "content": "ods. Such methods can serve as a reference for evaluating the effectiveness of self-supervised methods. We adapt fully-supervised multi-view 3D pose estimation methods LT [25] and EpipolarTrans [22] to a self-supervised manner. Under self-supervised settings, EpipolarTrans can only achieve limited performance improvements compared to traditional methods. LT-Algebraic [25], which incorporates learnable confidence into the triangulation. LT-Volumetric model [25], which unprojects 2D features into a 3D volume for inference, achieves better results, but the performance is dependent on the accuracy of the hand center. Canon-Pose [52] and EpipolarPose [27] obtain multi-view inference results through simple averaging like ours." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 499, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 289, + 715 + ], + "type": "text", + "content": "However, both of these methods are inferior to ours because they lack cross-view interaction. As our method predicts the root-relative 3D pose, we need to conduct post-processing to obtain the absolute coordinates. We introduce two different ways to achieve this: 1) using the 2D predictions of different views to triangulate and refine a center and 2) conducting RANSAC triangulation using our 2D predictions. Both methods have their merits. Opt-center can keep the root-relative results with hand prior, resulting in low PA-MPJPE. RANSAC gets better joint-wise accuracy, which is indicated by low MPJPE. We also provide qualitative results in the supplementary materials on the Assembly101 [46] dataset, which has a static camera setup. Even for challenging head-mounted moving cameras, we achieve convincing 3D pose estimates on the H2O [29] dataset. The experiments show that we have significantly pushed the performance of self-supervised methods to a comparable level with fully supervised methods." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 365, + 70, + 552, + 162 + ], + "blocks": [ + { + "bbox": [ + 365, + 70, + 552, + 162 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 552, + 162 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 552, + 162 + ], + "type": "image", + "image_path": "26046fcf4f4bcecb6c44fde775f429b259be0d31b838736a46707e4fda84821c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 162, + 542, + 194 + ], + "lines": [ + { + "bbox": [ + 364, + 162, + 542, + 194 + ], + "spans": [ + { + "bbox": [ + 364, + 162, + 542, + 194 + ], + "type": "text", + "content": "Figure 3. Error of using different (a) #training data, (b) (line-1)#view for training, and (line-2)#view for inference when trained with 8 views." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 195, + 550, + 290 + ], + "blocks": [ + { + "bbox": [ + 365, + 195, + 550, + 290 + ], + "lines": [ + { + "bbox": [ + 365, + 195, + 550, + 290 + ], + "spans": [ + { + "bbox": [ + 365, + 195, + 550, + 290 + ], + "type": "image", + "image_path": "449db25627e75bc0694bfe003d835e8df4b1a72a287df96bdbec5ca3fc9172b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 290, + 542, + 323 + ], + "lines": [ + { + "bbox": [ + 364, + 290, + 542, + 323 + ], + "spans": [ + { + "bbox": [ + 364, + 290, + 542, + 323 + ], + "type": "text", + "content": "Figure 4. AUC of three 2D joint sets. O, S, I, PE denote OpenPose, single-view, interaction, and average pixel error in resolution " + }, + { + "bbox": [ + 364, + 290, + 542, + 323 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 364, + 290, + 542, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 335, + 413, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 413, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 413, + 346 + ], + "type": "text", + "content": "4.4. Qualitative Result" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 351, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 460 + ], + "type": "text", + "content": "Fig. 5 presents the visual comparisons of 2 views between 2D joints of OpenPose, ours, and ground-truth on the HanCo dataset. We can observe that our method is more robust for outliers and can generate predictions close to the labels. Fig. 6 shows the 3D predictions from two viewpoints of ours, EpipolarPose, and CanonPose on the HanCo dataset. The results indicate that our method can get more accurate results especially when the occlusions are severe. Please refer to supplementary materials for more results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 466, + 400, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 400, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 400, + 479 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "text", + "content": "As shown in Tab. 4, we conduct comprehensive ablation experiments on the HanCo [62] dataset to show the effectiveness of each component. Single, Interact and Fusion denotes the evaluation of " + }, + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 304, + 483, + 545, + 531 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "content": "Different backbones. We first show our performance with different backbones. As shown in #1 and #2, using a large backbone like Res50, our performance can be further improved. For efficiency, we conduct ablation studies using Res18 as the backbone unless otherwise specified." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 652 + ], + "type": "text", + "content": "Two branches for cross-view interaction module. As presented in #3 and #4, both of the branches can reduce the error. VSF can explicitly model the view-shared information and add reliable information from every view. CVA can capture the self-/cross-view joint-level correlations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 652, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 545, + 714 + ], + "type": "text", + "content": "Graph features. The results indicate that three kinds of features (#5, #6, #7) all lead to performance improvement. Especially, local feature (#7, " + }, + { + "bbox": [ + 304, + 652, + 545, + 714 + ], + "type": "inline_equation", + "content": "G^3" + }, + { + "bbox": [ + 304, + 652, + 545, + 714 + ], + "type": "text", + "content": ") can notably reduce the error after the interaction by providing fine-grained details. DCVI. We also conduct experiments to show the impor" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20769" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "tance of DCVI by removing it and posing consistency constraints in single-view outputs like [24]. In this way, the performance drops dramatically (#10), proving the necessity of using DCVI to capture the features of all the views for self-supervised learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 191 + ], + "type": "text", + "content": "Two branches for multi-view consistency loss. Without enforcing cross-view interaction outputs to be consistent, the performance significantly drops (#9). If we do not explore relatively more reliable 2D predictions to enhance consistency, the performance can also get worse (#8)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 192, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 286, + 262 + ], + "type": "text", + "content": "Consistency losses. " + }, + { + "bbox": [ + 46, + 192, + 286, + 262 + ], + "type": "inline_equation", + "content": "(L_{c})" + }, + { + "bbox": [ + 46, + 192, + 286, + 262 + ], + "type": "text", + "content": " The performance is unsatisfactory (#11) when employing the cross-view interaction network without any consistency constraints (i.e. discard #8 and #9). The interaction network should cooperate with consistency so that the constraints can guide the network to exploit multi-view information to function better." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 263, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 286, + 323 + ], + "type": "text", + "content": "Multi-view distillation loss. " + }, + { + "bbox": [ + 46, + 263, + 286, + 323 + ], + "type": "inline_equation", + "content": "(L_{d})" + }, + { + "bbox": [ + 46, + 263, + 286, + 323 + ], + "type": "text", + "content": " Removing the multiview distillation loss, all the metrics drop by a large margin (#12), especially in single-view estimation accuracy. This phenomenon proves the effectiveness of collaborative learning between single- and multi-view networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 328, + 142, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 142, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 142, + 340 + ], + "type": "text", + "content": "4.6. Model Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 343, + 286, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 286, + 378 + ], + "type": "text", + "content": "Different percentage of unlabeled images. Fig. 3 (a) shows our method can get consistent performance improvement as the unlabeled training data increases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 380, + 286, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 286, + 462 + ], + "type": "text", + "content": "Different view number for training. The line-1 in Fig. 3 (b) shows the performance of our method tested on a certain view when trained with different view numbers. The curve shows that our method can be consistently improved as the number of views increases. We also observe that using multiple views for training can significantly improve performance when the valid views are few." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 286, + 617 + ], + "type": "text", + "content": "Different view number for inference. Our model allows inferring with an arbitrary number of views. However, when the model is trained with a fixed view number, it could get the view number bias, resulting in better performance using the view number close to the training one. To avoid this, we add random masks in our interaction module and finetune the model for a few epochs. After that, results can get better by a small margin (the single-view error is " + }, + { + "bbox": [ + 46, + 463, + 286, + 617 + ], + "type": "inline_equation", + "content": "11.07\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 463, + 286, + 617 + ], + "type": "text", + "content": " and the fusion error " + }, + { + "bbox": [ + 46, + 463, + 286, + 617 + ], + "type": "inline_equation", + "content": "7.60\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 463, + 286, + 617 + ], + "type": "text", + "content": ", both in NMPJPE.). The line-2 in Fig. 3 (b) shows results on a certain view when trained on 8 views and tested on 1 to 8 views. We can observe consistent improvement with the inference view number increases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 286, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 666 + ], + "type": "text", + "content": "Different 2D joint sets. Fig. 4 presents the accuracy of different 2D joint sets on the HanCo training and testing set. Our 2D predictions are extremely better than OpenPose 2D pseudo label used for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "content": "Iteratively training. Our approach can use the previous predictions as pseudo labels for iterative training. We find it helpful till iteration 3 and get saturated afterward. From 1 to 3 iterations, NMPJPE is 7.75, 7.68, and 7.64." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 533, + 146 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 533, + 146 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 533, + 146 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 533, + 146 + ], + "type": "image", + "image_path": "151d8d96736515d6e29262c0f139423473de557de666646edd5d7ad948251c85.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 155, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 177 + ], + "type": "text", + "content": "Figure 5. 2D prediction (overlayed in the images) comparisons between OpenPose, ours, and ground-truth on the HanCo dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 184, + 532, + 262 + ], + "blocks": [ + { + "bbox": [ + 318, + 184, + 532, + 262 + ], + "lines": [ + { + "bbox": [ + 318, + 184, + 532, + 262 + ], + "spans": [ + { + "bbox": [ + 318, + 184, + 532, + 262 + ], + "type": "image", + "image_path": "f8c92cc4e139cbe7c3219b5eaa6809fbb355cf757c3bb4b080cf7876b7e35f8c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "type": "text", + "content": "Figure 6. 3D prediction comparisons between our method, EpipolarPose, and CanonPose on the HanCo dataset. Our prediction and ground-truth are shown in solid red and dashed green respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 312, + 470, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 470, + 324 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 330, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 532 + ], + "type": "text", + "content": "To our best knowledge, we present the first self-supervised framework that aims to learn a single-view 3D hand estimator from unlabeled multi-view data. At the core of our approach, a cross-view interaction network is carefully designed to supervise the single-view output by leveraging the collaboration among multi-views. Specifically, the network captures the interdependencies of features among different views, resulting in improved accuracy of hand pose estimation after cross-view interaction. Additionally, the multi-view results are fused to supervise the single-view output for self-distillation. The effectiveness and versatility of the proposed framework are extensively evaluated through experiments, which demonstrate that our method not only establishes a new benchmark for self-supervised 3D hand pose estimation from single-view input but also offers flexible multi-view inference with state-of-the-art performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 593 + ], + "type": "text", + "content": "We focused on hand pose estimation without heavy occlusions in this work. Extending our work to more challenging scenarios, such as hand-object interaction or relaxing the synchronization constraints in multi-view inputs, would be interesting topics for further study." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 601, + 403, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 403, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 403, + 613 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants (62071067, 62171057, 62201072), in part by the Ministry of Education and China Mobile Joint Fund (MCM20200202), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20770" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "type": "text", + "content": "[1] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1067-1076, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Weakly-supervised domain adaptation via gan and mesh model for estimating 3d hand poses interacting objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6121-6131, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 258 + ], + "type": "text", + "content": "[3] Kristijan Bartol, David Bojanic, Tomislav Petković, and Tomislav Pribanić. Generalizable human pose triangulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11028-11037, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "type": "text", + "content": "[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10843-10852, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 308, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 287, + 351 + ], + "type": "text", + "content": "[5] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 666-682, 2018. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 418 + ], + "type": "text", + "content": "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12929–12938, 2021. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 421, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 486 + ], + "type": "text", + "content": "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 488, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 287, + 553 + ], + "type": "text", + "content": "[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274–13283, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 556, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 287, + 621 + ], + "type": "text", + "content": "[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10451-10460, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[11] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 7792-7801, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[12] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "text", + "content": "[13] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 2021 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[14] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 262, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 316 + ], + "type": "text", + "content": "[15] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1991-2000, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "text", + "content": "[16] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10833-10842, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 372, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 392 + ], + "type": "text", + "content": "[17] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "type": "text", + "content": "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3196-3206, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[19] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. Megatrack: monochrome egocentric articulated hand-tracking for virtual reality. ACM Trans. Graph., 39(4):87, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "type": "text", + "content": "[20] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, Cabezas Randi, Tran Luan, Akbay Muzaffer, Yu Tsz-Ho, Keskin Cem, and Wang Robert. Umetrack: Unified multi-view end-to-end hand tracking for vr. ACM Transactions on Graphics, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 571, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 601 + ], + "type": "text", + "content": "[21] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 603, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 645 + ], + "type": "text", + "content": "[22] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020. 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[23] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European Conference on Computer Vision (ECCV), pages 118-134, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[24] Umar Iqbal, Pavlo Molchanov, and Jan Kautz. Weakly-supervised 3d human pose learning via multi-view images" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20771" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5243-5252, 2020. 1, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[25] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2019. 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 183, + 287, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 238 + ], + "type": "text", + "content": "[27] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1077-1086, 2019. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 238, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 304 + ], + "type": "text", + "content": "[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990–5000, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "text", + "content": "[29] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "text", + "content": "[30] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2761-2770, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[31] Moran Li, Yuan Gao, and Nong Sang. Exploiting learnable joint groups for hand pose estimation. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1921-1929, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "text", + "content": "[32] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "text", + "content": "[33] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 581 + ], + "type": "text", + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 636 + ], + "type": "text", + "content": "[35] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In European Conference on Computer Vision, pages 424-442. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "text", + "content": "[36] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[37] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 163 + ], + "type": "text", + "content": "[38] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision, pages 548-564. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 230 + ], + "type": "text", + "content": "[39] Franziska Mueller, Florian Bernard, Oleksandr Sotnychenko, Dushyant Mehta, Srinath Sridhar, Dan Casas, and Christian Theobalt. Ganerated hands for real-time 3d hand tracking from monocular rgb. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 49-59, 2018. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in python. In NIPS 2017 Workshop on Autodiff, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 285, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 342 + ], + "type": "text", + "content": "[41] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 388 + ], + "type": "text", + "content": "[42] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4342-4351, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 444 + ], + "type": "text", + "content": "[43] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3d pose estimation through camera-disentangled representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6040-6049, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 444, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 511 + ], + "type": "text", + "content": "[44] Helge Rhodin, Jörg Spörri, Isinsu Katircioglu, Victor Constantin, Frédéric Meyer, Erich Müller, Mathieu Salzmann, and Pascal Fua. Learning monocular 3d human pose estimation from multi-view images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8437-8446, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[45] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, 36(6), 2017. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 624 + ], + "type": "text", + "content": "[46] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "type": "text", + "content": "[47] Hui Shuai, Lele Wu, and Qingshan Liu. Adaptive multi-view and temporal fusing transformer for 3d human pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "text", + "content": "[48] Adrian Spurr, Umar Iqbal, Pavlo Molchanov, Otmar Hilliges, and Jan Kautz. Weakly supervised 3d hand pose estimation via biomechanical constraints. In European Conference on Computer Vision, pages 211-228. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20772" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[49] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[50] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 205 + ], + "type": "text", + "content": "[51] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision, pages 197-212. Springer, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 287, + 259 + ], + "type": "text", + "content": "[52] Bastian Wandt, Marco Rudolph, Petrissa Zell, Helge Rhodin, and Bodo Rosenhahn. Canonpose: Self-supervised monocular 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13294-13304, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 314 + ], + "type": "text", + "content": "[53] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 316, + 287, + 360 + ], + "type": "text", + "content": "[54] Chao Wen, Yinda Zhang, Zhuwen Li, and Yanwei Fu. Pixel2mesh++: Multi-view 3d mesh generation via deformation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1042-1051, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 361, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 404 + ], + "type": "text", + "content": "[55] Linlin Yang, Shicheng Chen, and Angela Yao. Semihand: Semi-supervised hand pose estimation with consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11364-11373, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "type": "text", + "content": "[56] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2335-2343, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 449, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 492 + ], + "type": "text", + "content": "[57] Linlin Yang and Angela Yao. Disentangling latent hands for image synthesis and pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9877-9886, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 494, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 494, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 49, + 494, + 287, + 547 + ], + "type": "text", + "content": "[58] Xiong Zhang, Hongsheng Huang, Jianchao Tan, Hongmin Xu, Cheng Yang, Guozhu Peng, Lei Wang, and Ji Liu. Hand image understanding via deep multi-task learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11281-11292, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 601 + ], + "type": "text", + "content": "[59] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 287, + 647 + ], + "type": "text", + "content": "[60] Zhe Zhang, Chunyu Wang, Weichao Qiu, Wenhu Qin, and Wenjun Zeng. Adafuse: Adaptive multiview fusion for accurate human pose estimation in the wild. International Journal of Computer Vision, 129(3):703-718, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 712 + ], + "type": "text", + "content": "[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 217 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[62] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In DAGM German Conference on Pattern Recognition, pages 250-264. Springer, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 547, + 161 + ], + "type": "text", + "content": "[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. 2, 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20773" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_content_list.json b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d63b8701d48a4a6e1d66e59c25e019f3d194dcb3 --- /dev/null +++ b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_content_list.json @@ -0,0 +1,1668 @@ +[ + { + "type": "text", + "text": "HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending", + "text_level": 1, + "bbox": [ + 161, + 130, + 810, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyi Wei $^{1}$ , Dongdong Chen $^{2}$ , Wenbo Zhou $^{1,\\dagger}$ , Jing Liao $^{3}$ , Weiming Zhang $^{1}$ , Gang Hua $^{4}$ , Nenghai Yu $^{1}$", + "bbox": [ + 248, + 179, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China 2Microsoft Cloud AI", + "bbox": [ + 250, + 217, + 715, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ City University of Hong Kong $^{4}$ Xi'an Jiaotong University", + "bbox": [ + 285, + 234, + 678, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bestwty@mail., welbeckz@, zhangwm@, yn@}ustc.edu.cn", + "bbox": [ + 246, + 253, + 712, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cddlyf@, ganghua@}gmail.com, jingliao@cityu.edu.hk", + "bbox": [ + 264, + 272, + 705, + 286 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f4238118915e86bb7bbef0d613e3720d9d7d70b5e7c7b3cdcb15f292e08720e.jpg", + "image_caption": [ + "Figure 1. HairCLIPv2 supports hairstyle and color editing individually or jointly with unprecedented user interaction mode support, including text, mask, sketch, reference image, etc." + ], + "image_footnote": [], + "bbox": [ + 89, + 323, + 885, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 536, + 313, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hair editing has made tremendous progress in recent years. Early hair editing methods use well-drawn sketches or masks to specify the editing conditions. Even though they can enable very fine-grained local control, such interaction modes are inefficient for the editing conditions that can be easily specified by language descriptions or reference images. Thanks to the recent breakthrough of cross-modal models (e.g., CLIP), HairCLIP is the first work that enables hair editing based on text descriptions or reference images. However, such text-driven and reference-driven interaction modes make HairCLIP unable to support fine-grained controls specified by sketch or mask. In this paper, we propose HairCLIPv2, aiming to support all the aforementioned interactions with one unified framework. Simultaneously, it improves upon HairCLIP with better irrelevant attributes (e.g., identity, background) preservation and unseen text descriptions support. The key idea is to convert all the hair editing tasks into hair transfer tasks, with editing conditions converted into different proxies accordingly. The editing effects are added upon the input image by blend", + "bbox": [ + 75, + 569, + 472, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing the corresponding proxy features within the hairstyle or hair color feature spaces. Besides the unprecedented user interaction mode support, quantitative and qualitative experiments demonstrate the superiority of HairCLIPv2 in terms of editing effects, irrelevant attribute preservation and visual naturalness. Our code is available at https: //github.com/wty-ustc/HairCLIPv2.", + "bbox": [ + 500, + 523, + 893, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 661, + 632, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hair editing as an interesting and challenging problem has attracted a lot of research attention from both academia and industry. Over the past few decades, tremendous progress [50, 41, 31, 36] has been made in this field, enabling high-fidelity hair editing based on various types of user interactions or controls. In earlier hair editing methods [50, 41], commonly supported user editing conditions are sketches and masks, which can enable fine-grained local controls. But in real scenarios, many hair editing conditions can be specified by simpler interactions, e.g., text descriptions (e.g., \"bowl cut hairstyle\") and reference images.", + "bbox": [ + 496, + 686, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, cross-modal visual and language representation learning [29, 22, 43, 8, 35, 38, 45] has made remarkable breakthrough, which makes text-guided image manipu", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Wenbo Zhou is the corresponding author.", + "bbox": [ + 101, + 886, + 331, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23589", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/860f16c8a83bae64da252f61191a4c4cd5680ba7df0f1cb9a303b9c82d985310.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HairCLIP [36]LOHO [24]Barbershop [47]HairNet [48]SYH [20]MichiGAN [31]SketchSalon [41]Ours
Aligned Hair Transfer✓✓✓✓✓✓✗✓
Unaligned Hair Transfer✓✗✗✓✓✗✗✓
Text✓✗✗✗✗✗✗✓
Mask✗✗✓✗✓✓✗✓
Sketch✗✗✗✗✗✓✓✓
Local Hairstyle Editing✗✗✗✗✗✓✓✓
Local Hair Color Editing✗✗✗✗✗✗✓✓
", + "bbox": [ + 81, + 88, + 893, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparisons between our approach and mainstream hair editing methods in terms of available interaction modes and functionality. Only our method supports all interaction modes and enables both global and local hair editing.", + "bbox": [ + 75, + 200, + 892, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lation possible. HairCLIP [36] presents the first attempt that supports hair editing via text description and reference image within one unified framework. Despite such text-driven and reference-driven interaction being more efficient and user-friendly, HairCLIP cannot support fine-grained controls like sketches and masks. Moreover, HairCLIP has two other limitations: 1) Since hair editing in HairCLIP is accomplished by pure latent code manipulation, it will inevitably alter other irrelevant attributes (e.g., identity, background) because fully decoupling different attributes in latent codes is difficult; 2) It struggles in yielding satisfactory results for text descriptions that differ significantly from training texts.", + "bbox": [ + 75, + 253, + 472, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we take a step forward and propose HairCLIPv2, a unified hair editing system that unprecedentedly supports all the aforementioned interaction modes, including the natural text/reference-driven interaction and fine-grained local interaction. In Table 1, we list the interaction modes and editing functionality supported by existing hair editing methods. Moreover, with fundamentally different editing mechanism design, HairCLIPv2 makes great improvement upon HairCLIP with better irrelevant attributes preservation and unseen text description support.", + "bbox": [ + 75, + 454, + 468, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key idea of HairCLIPv2 is converting all the hair editing tasks into hair transfer tasks, and the editing conditions are converted into different transfer proxies accordingly. Conceptually, it can be understood as \"find proxy hair images that satisfy the editing conditions and transfer the corresponding attributes to the source image\". Note that we use the StyleGAN latent code or feature corresponding to such proxies rather than use the proxy images explicitly.", + "bbox": [ + 75, + 609, + 468, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, we first transform the input source image into the bald proxy, which inpaints the hair-covered regions (e.g., background, ears) with reasonable semantic attributes. This can help avoid the editing artifacts caused by occlusion when blending the source image with condition proxies. For different editing proxies, we define their generation as different tasks performed in StyleGAN according to their characteristics. Depending on the users' editing preferences, hair editing effects are then enforced upon the input image by blending the corresponding proxy features within the hairstyle feature space or hair color feature space. This", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is different from HairCLIP that achieves the editing effect by manipulating the 1-d latent codes. Such feature blending based editing naturally supports global and local hair editing by controlling the blending area to cover the entire hair area or part of it.", + "bbox": [ + 496, + 253, + 893, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To show the superiority of HairCLIPv2, we conduct extensive comparisons. In addition to more complete user interaction modes support, HairCLIPv2 also shows obvious advantages in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness. Some interactive editing examples are provided in Figure 1. Our contributions can be summarized as below:", + "bbox": [ + 496, + 330, + 893, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a fresh perspective for hair editing tasks and propose a novel hair editing paradigm that unifies various types of editing into the form of proxy hair transfers. We achieve all editing effects with the feature blending mechanism, which not only alleviates the editing pressure on each proxy but also enables excellent irrelevant attribute preservation.", + "- We dedicate design to the proxy generation for different conditions based on their own special properties, e.g., for the text proxy, the decoupled proxy design and optimization starting point selection strategy help us achieve better editing effects and arbitrary text support; for the sketch proxy, we achieve local hairstyle editing support for the first time within the StyleGAN-based framework by formalizing its generation as the image translation task and incorporating insights of semantic layering in StyleGAN.", + "- Our system pushes the interactions of hair editing to a new level, supporting arbitrary text, mask, reference image, sketch and their combinations, and enabling both global and local hair editing, which has never been realized before." + ], + "bbox": [ + 517, + 449, + 893, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 828, + 640, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Adversarial Networks. Since being invented, GANs have made considerable progress in terms of training strategies [18, 34], loss functions [3, 4, 9], and network", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23590", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structures [25, 12, 10, 30, 32]. In the field of image synthesis, a series of works called StyleGAN [17, 18, 15, 16] represents the cutting edge of GANs. Given its promising semantically decoupled latent space [7, 26] and high-quality image synthesis abilities, the pre-trained StyleGAN has become the preferred choice for performing image editing. In this paper, we choose StyleGAN2 to develop our framework, which is consistent with other hair editing methods [36, 47, 20, 21, 40, 24] to be compared.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Space Embedding and Editing. As the bridge connecting the pre-trained StyleGAN and other downstream editing tasks, GAN inversion aims to yield the ideal embedding of the real image in the latent space. Based on the application purposes, we roughly classify the GAN inversion methods into two categories: methods [33, 46, 49] suitable for editing and methods [2, 23, 37] for better reconstruction. The former methods project the real image into the embedding subspace more suitable for editing at the expense of reconstruction. Among them, e4e [33] has become the most popular method for editing tasks [36, 21, 39] performed in the latent space. The latter approaches [2, 23, 37] aim to achieve the perfect reconstruction of the real image. However, limited by the representation capability of the latent space, all these methods cannot achieve the perfect reconstruction. To address this issue, Barbershop [47] proposes a novel inversion method, which additionally introduces a feature space $\\mathcal{F}$ of StyleGAN combined with the latent space $S$ to form a new embedding space $\\mathcal{FS}$ . Inspired by this, we decouple the editing task from the reconstruction task by blending editing proxy features in the feature space to achieve a unified hair editing system that supports a wide range of interactions.", + "bbox": [ + 75, + 234, + 472, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hair Editing Using GANs. Existing hair editing methods can be roughly categorized as conditional GANs [41, 31, 13] based and pre-trained StyleGAN based [36, 21, 40, 24, 47, 20, 48, 28]. As a pioneering work of hair transfer, MichiGAN [31] accomplishes hairstyle transfer by extracting the orientation map of the reference image. Barbershop [47] performs hair transfer within their proposed $\\mathcal{FS}$ embedding space. But these methods often struggle when large pose differences exist between source and target image. Recently, some improvements [20, 48] on Barbershop make pose unaligned hair transfers possible. Our framework is also compatible with pose unaligned hair transfers and additionally offers more interaction modes. SketchHairSalon [41] enables local editing of hairstyle and hair color by using colored sketches as the input to the conditional translation network. Unlike them, we show for the first time that the StyleGAN-based framework can also perform local hair editing with sketches as the condition.", + "bbox": [ + 75, + 583, + 472, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benefiting from the development of cross-modal models [29, 22], text-guided hair editing [36, 21, 40] has become the new trend. StyleCLIP [21] and TediGAN [40] uti", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lize CLIP loss to perform hair editing in an optimized manner. However, since the embedding of real image deviates from the original suitable editing latent space, these methods will fail for some cases. HairCLIP [36] alleviates the problem by training a hair mapper on a large-scale dataset, but struggles in yielding good results for descriptions that differ significantly from the training text. Moreover, none of these methods can preserve the irrelevant attributes well. In this work, we present a new perspective to enable text-guided hair editing methods. By decoupling the editing task from the reconstruction, we can better preserve the irrelevant attributes while enabling high-quality hair manipulation via arbitrary text descriptions. More importantly, there is no prior work that supports so many interaction modes and functionality as we offer.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 498, + 332, + 671, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 357, + 640, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN [18] can synthesize photorealistic images with a progressive upsampling network consisting of 18 layers. Its $\\mathcal{W} \\subsetneq \\mathbb{R}^{512}$ latent space exhibits good semantic decoupling properties [7, 26], thus enabling various editing tasks. To perform semantic editing while ensuring the reconstruction quality, some inversion methods [1, 23, 37] extend the original $\\mathcal{W}$ space to $\\mathcal{W}+$ space, which is defined as a cascade of 18 different latent codes $w_{i}$ from the $\\mathcal{W}$ space, i.e., $\\mathcal{W} + \\subsetneq \\mathbb{R}^{18\\times 512}$ .", + "bbox": [ + 496, + 381, + 890, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{FS}$ Embedding Space is proposed by Barbershop [47], which is designed to increase the representational capability of embedding space for details and enable the spatial control of image features. The new $\\{F_7,S\\}$ latent code replaces the first 7 layers of $\\mathcal{W}+$ latent code with the $32\\times 32\\times 512$ features $F_{7}$ of style-block 7 of StyleGAN $G$ , i.e., $F_{7}\\in \\mathcal{F}\\subsetneq \\mathbb{R}^{32\\times 32\\times 512},S = [w_8,\\dots w_i,\\dots w_{18}],w_i\\in \\mathcal{W}.$", + "bbox": [ + 496, + 518, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP [22] is a multi-modality model pretrained on web-scale image-text pairs. It can well measure the semantic similarity between given image and text.", + "bbox": [ + 496, + 625, + 890, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 680, + 609, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the $\\mathcal{FS}$ embedding space [47] is proposed, performing seamless feature blending in $\\mathcal{F}$ -space has become the de facto standard for many hair transfer works [47, 20, 48], because it can encode the spatial information and preserve local details. On the other hand, many hair editing efforts [36, 21, 40, 39] choose to perform editing in the $\\mathcal{W}+$ space, despite unsatisfactory reconstruction, because it can encode rich disentangled semantics [7, 26]. Considering that $\\mathcal{F}$ -space is expressive and enables realistic feature integration results while $\\mathcal{W}+$ space is editable, we therefore wonder \"Can we enjoy the best features of both spaces to facilitate the hair editing task?\" To achieve this goal, we formulate the hair editing tasks as the hair transfer tasks.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23591", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2679e48691e3a051ba8f260fd05ac42320190d2c989b02e6125978b310d183b5.jpg", + "image_caption": [ + "Figure 2. Overview of HairCLIPv2: Example with hairstyle description text, sketch, mask and hair color RGB values as conditional inputs. Corresponding proxy images are just for better understanding. We complete the hair editing by converting different conditions into different proxies and achieve editing effects by blending them in StyleGAN feature spaces." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 890, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we convert all editing conditions (e.g., text, reference image, sketch) into different proxies in the $\\mathcal{W}+$ space, and accomplish hair editing by seamless proxy feature blending in the feature spaces of StyleGAN. The proxy features of different conditions are obtained with tailored methods based on the condition characteristics.", + "bbox": [ + 75, + 339, + 468, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the design in HairCLIP [36], we edit hairstyle and hair color sequentially by blending editing proxy features in the early and later StyleGAN feature space respectively. In detail, as shown in Figure 2, we choose to perform proxy feature blending of hairstyle and hair color on feature spaces $\\mathcal{F}^{style} \\subsetneq \\mathbb{R}^{32 \\times 32 \\times 512}$ and $\\mathcal{F}^{color} \\subsetneq \\mathbb{R}^{256 \\times 256 \\times 128}$ , which correspond to the features of 7-th and 14-th style-block in StyleGAN respectively. And users can select various interactions (global or local, single or combined) to edit hairstyles and hair color individually or jointly.", + "bbox": [ + 75, + 429, + 470, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Converting Input Image to Bald Proxy", + "text_level": 1, + "bbox": [ + 76, + 590, + 408, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a source image $I^{src}$ to be edited, we obtain its latent code $w^{src}$ in $\\mathcal{W}+$ space and feature $F_{7}^{src}$ in $\\mathcal{FS}$ space by II2S [49] and FS embedding algorithms [47], respectively. Balding proxy is then generated to inpaint the hair-covered region with reasonable semantic attributes (e.g., background, ears, face, etc.), which can avoid editing artifacts due to occlusion when blending the original image with different editing proxies.", + "bbox": [ + 75, + 614, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Balding Proxy. To remove the occlusion from the hair area of the source image, we bald it using HairMapper [39]: $w^{bald} = \\mathbb{B}(w^{src})$ . $\\mathbb{B}$ denotes HairMapper, which completes the balding editing operation in $\\mathcal{W}+$ space on the latent code $w^{src}$ to yield the latent code $w^{bald}$ corresponding to the balded source image. Since editing in $\\mathcal{W}+$ space inevitably gets other irrelevant attributes modified, we circumvent this issue by blending bald feature with source image feature in $\\mathcal{F}^{style}$ space:", + "bbox": [ + 75, + 736, + 468, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {7} ^ {b a l d} = G \\left(w _ {1 - 7} ^ {b a l d}\\right) \\times M ^ {b a l d} + F _ {7} ^ {s r c} \\times \\left(1 - M ^ {b a l d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 882, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G$ stands for StyleGAN, $G(w_{1-7}^{bold})$ represents the bald feature in $\\mathcal{F}^{style}$ space. $M^{bald}$ is the binary mask indicating the hair and ear regions of the source image, which is obtained by the facial parsing network BiseNET [42] and downsampled to $32 \\times 32$ . With the guidance of $M^{bald}$ , the irrelevant attributes region in $F_7^{bald}$ can continue to be preserved in following proxy feature blending.", + "bbox": [ + 496, + 338, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Hairstyle Editing", + "text_level": 1, + "bbox": [ + 500, + 453, + 669, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Below we elaborate on how to generate the proxy for different hairstyle conditions.", + "bbox": [ + 496, + 477, + 890, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text Proxy. We formalize text proxy generation as the editing task done in $\\mathcal{W}+$ space based on the CLIP loss guidance. Unlike prior works [36, 21, 40], our text proxy generation process is free from the pressure of irrelevant attribute preservation, and allows us to select a more suitable starting point for the optimization process, which leads to better editing effects. In order to ensure both the optimal editing effect and the diversity of editing results, we choose to sample a random point around the mean face latent code as the optimization starting point for our text proxy latent code $w^{text}$ . In detail, we adopt the truncation trick of StyleGAN as $w^{init} = w^{mean} + \\psi (w^{random} - w^{mean})$ , where $w^{mean}$ is the mean face latent code and $w^{random}$ is sampled randomly. By setting a small value of $\\psi$ , we ensure that the initial optimization starting point $w^{init}$ is around the average face latent code $w^{mean}$ . We will show the benefits of this initialization strategy in the ablation analysis. We adopt the CLIP loss with transformation augmentations [5] to perform the text-guided hairstyle editing while reducing the disturbance caused by adversarial examples:", + "bbox": [ + 496, + 507, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {c l i p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} (1 - \\cos \\left(E _ {i} \\left(A _ {i} \\left(G \\left(w ^ {t e x t}\\right)\\right)\\right), E _ {t} (s t)\\right)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 819, + 890, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{i}$ represents $i$ -th transformation augmentation, $N$ denotes the number of augmentations ( $N = 4$ by default),", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23592", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\cos (\\cdot)$ means cosine similarity, $G(w^{text})$ means the editing result for each pass in the optimization process, $E_{i}$ and $E_{t}$ stands for the CLIP image encoder and text encoder respectively, and $st$ refers to the user-supplied text description. Besides, pose alignment loss $L^{pose}$ is utilized to ensure that the face shape and pose of text proxy are consistent with the source image to ease the subsequent feature blending:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {p o s e} = \\frac {1}{N _ {k}} | | E _ {p} (I ^ {s r c}) - E _ {p} (G (w ^ {t e x t})) | | _ {2} ^ {2}, \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 205, + 468, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E_{p}$ represents the 3D keypoint extractor [6] and $N_{k}$ denotes the number of keypoints. Optionally, the shape loss $L^{shape}$ is added to constrain the shape of the generated hair according to whether the user provides the hair region mask. We then obtain text proxy feature $F_{7}^{text}$ using the optimized $w^{text}: F_{7}^{text} = G(w_{1-7}^{text})$ .", + "bbox": [ + 76, + 244, + 468, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reference Proxy. Given a hairstyle reference image $I^{sr}$ , we generate the reference proxy by performing the unaligned hairstyle transfer task. $I^{sr}$ is first inverted by the II2S [49] embedding algorithm into the $\\mathcal{W}+$ space to get $w^{ref}$ , which is served as the starting point for hairstyle transfer. During the transfer process, we expect to keep the original hair structure of the reference image while ensuring its pose and facial shape to be consistent with the source image. Thus, $L^{pose}$ is imposed between $G(w^{ref})$ and $I^{src}$ . In addition, a style loss $L^{style}$ based on the gram matrix [11] is used to ensure that the hairstyle structure of $I^{sr}$ remains unchanged during the alignment process:", + "bbox": [ + 75, + 335, + 470, + 517 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L ^ {s t y l e} = \\frac {1}{4} \\sum_ {i = 1} ^ {4} \\left\\| \\mathcal {G} _ {i} \\left(V G G _ {i} \\left(I ^ {s r} \\times M ^ {r h}\\right)\\right) \\right. \\\\ \\left. - \\mathcal {G} _ {i} \\left(V G G _ {i} \\left(G \\left(w ^ {r e f}\\right) \\times M ^ {g h}\\right)\\right) \\right| | _ {2} ^ {2}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 531, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_i(\\gamma_i)$ represents the gram matrix calculated on the $i$ -th layer features and a total of 4 layers of features are extracted, i.e., $\\{relu1\\_2, relu2\\_2, relu3\\_3, relu4\\_3\\}$ of VGG [27]. $M^{rh}$ and $M^{gh}$ are hair masks for the reference image and the generated image of each round during optimization process predicted by BiseNET [42]. For the same purpose, the $L_2$ norm of the manipulation magnitude in the latent space is utilized during optimization:", + "bbox": [ + 75, + 601, + 468, + 723 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {r e g} = \\left\\| w _ {t} ^ {r e f} - w _ {t - 1} ^ {r e f} \\right\\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 731, + 468, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{t}^{ref}$ and $w_{t-1}^{ref}$ represent the latent code of the current step and the previous step, respectively. Similar to the text proxy, $L^{shape}$ is optionally added to allow the user to customize the shape of the hair. We then obtain reference proxy feature $F_{7}^{ref}$ using the optimized $w^{ref}$ : $F_{7}^{ref} = G(w_{1-7}^{ref})$ .", + "bbox": [ + 75, + 761, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sketch Proxy. Enabling sketch-based local hairstyle editing within our framework is nontrivial. It is hard to find suitable losses to constrain the local hairstyle structure to", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "conform to the sketch given by the user. To circumvent this problem, we innovatively formalize the synthesis of sketch proxy as an image translation task based on StyleGAN. Utilizing the sketch-hair dataset created by SketchHairSalon [41], we train a sketch2hair inverter $T$ , which is based on E2Style [37] and aims to find the most appropriate latent code in $\\mathcal{W}+$ space to accurately translate a given sketch to the corresponding hair structure. The training loss consists of regular pixel-level $L_{2}$ loss, feature-level LPIPS [44] loss and multi-layer face parsing loss [37] which is introduced to provide more local supervision. During the training process, we randomly remove a portion of the strokes to make our sketch2hair inverter adapt to a variety of sketch inputs from fine to coarse, e.g., even just one stroke. Given a local hairstyle sketch $S$ , our sketch proxy features $F_{7}^{sketch}$ are synthesized by pre-trained sketch2hair $T$ and StyleGAN $G$ :", + "bbox": [ + 496, + 90, + 890, + 333 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {\\text {s k e t c h}} = T (S), \\quad F _ {7} ^ {\\text {s k e t c h}} = G \\left(w _ {1 - 7} ^ {\\text {s k e t c h}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 345, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proxy Feature Blending. For text and reference image based hairstyle condition, we perform global blending in $\\mathcal{F}^{style}$ space:", + "bbox": [ + 496, + 369, + 890, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {7} ^ {\\text {g l o b a l}} = F _ {7} ^ {t r} \\times M ^ {\\text {g l o b a l}} + F _ {7} ^ {\\text {b a l d}} \\times \\left(1 - M ^ {\\text {g l o b a l}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 422, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_7^{tr} \\in \\{F_7^{text}, F_7^{ref}\\}$ and $M^{global}$ is the binary mask corresponding to the hair region of $F_7^{tr}$ . Optionally, the sketch-based local hairstyle editing is applied:", + "bbox": [ + 496, + 452, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {7} ^ {\\text {s t y l e}} = F _ {7} ^ {\\text {s k e t c h}} \\times M ^ {\\text {l o c a l}} + F _ {7} ^ {\\text {g l o b a l}} \\times \\left(1 - M ^ {\\text {l o c a l}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 507, + 890, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M^{local}$ is obtained by downsampling the user input sketch $S$ to $32 \\times 32$ after dilation. A natural concern is the artifacts brought by the mismatch between hair features within $M^{local}$ and other hair features. But thanks to the semantic layering characteristics of StyleGAN, the resulting image shows consistent tones as these hair features will be modulated by the later layers. Our framework allows users to only edit the hairstyle: $I^{style} = G(F_7^{style}, w_{8-18}^{src})$ , by skipping the following hair color editing.", + "bbox": [ + 496, + 536, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Hair Color Editing", + "text_level": 1, + "bbox": [ + 500, + 681, + 683, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We achieve hair color editing by performing proxy feature blending in $\\mathcal{F}^{color}$ space. By choosing to use the feature $F_{7}^{src}$ or $F_{7}^{style}$ , we allow to edit only hair color or both hair style&hair color. Below, we use $F_{7}^{style}$ as the example. Color Proxy. We initialize the color proxy with $F_{7}^{style}$ and $w_{8-18}^{color} = w_{8-18}^{src}$ , and set $w_{10-13}^{color}$ to be estimizable. The loss $L^{color}$ in the optimization process consists of $L^{modal}$ and $L^{bg}$ , where $L^{modal}$ can be defined as $L^{clip}$ or the average color $L_{2}$ loss of the hair region depending on the hair color condition types (text, reference image, RGB values), and $L^{bg}$ is defined as follows:", + "bbox": [ + 496, + 704, + 890, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {b g} = \\left\\| \\left(I ^ {\\text {s t y l e}} - I ^ {\\text {c o l o r}}\\right) \\times \\left(M ^ {n - h a i r}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b246691c416925cb7d161e51c5f330754b38ffe84767935a24ed54da9206e744.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsIDS↑PSNR↑SSIM↑
Ours0.8429.50.91
HairCLIP [36]0.4521.60.74
StyleCLIP [21]0.4319.60.72
TediGAN [40]0.1622.50.74
DiffCLIP [19]0.7126.80.86
", + "bbox": [ + 86, + 87, + 450, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison for irrelevant attributes preservation. IDS denotes identity similarity, PSNR and SSIM are calculated at the intersected non-hair regions before and after editing.", + "bbox": [ + 76, + 191, + 468, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I^{color} = G(F_7^{style}, w_{8-18}^{color})$ , $M^{n-hair}$ is the mask of the non-hair region intersection between $I^{style}$ and $I^{color}$ .", + "bbox": [ + 75, + 258, + 468, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proxy Feature Blending. Even with the $L^{bg}$ constraint, we find non-hair regions are often inevitably modified because of imperfect semantic decoupling of the $\\mathcal{W}+$ space. We solve this by performing proxy feature blending in $\\mathcal{F}^{color}$ space, which also naturally supports local hair color editing:", + "bbox": [ + 75, + 291, + 468, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {1 4} ^ {\\text {b l e n d}} = G \\left(F _ {7} ^ {\\text {s t y l e}}, w _ {8 - 1 4} ^ {\\text {c o l o r}}\\right) \\times M ^ {\\text {c o l o r}} \\\\ + G \\left(F _ {7} ^ {\\text {s t y l e}}, w _ {8 - 1 4} ^ {\\text {s r c}}\\right) \\times \\left(1 - M ^ {\\text {c o l o r}}\\right), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 396, + 467, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M^{color}$ is the hair area mask or a local editing area mask drawn by the user. We set $F_{14}^{blend}$ and $w_{15-18}^{color}$ to be estimizable to further perform the optimization. In the optimization process, we use the $L_{blend}$ loss, which consists of $L_{2}$ loss and LPIPS loss to constrain $I^{final}$ to be similar to $I^{color}$ inside $M^{color}$ and similar to $I^{style}$ outside $M^{color}$ simultaneously. The final edited image is synthesized as follows: $I^{final} = G(F_{14}^{blend}, w_{15-18}^{color})$ .", + "bbox": [ + 75, + 449, + 468, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 585, + 207, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details of our approach are provided in the supplementary material. For all compared methods, we use their official codes or pre-trained models.", + "bbox": [ + 75, + 609, + 468, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative and Qualitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 665, + 431, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Text-Driven Hair Editing Methods. We compare HairCLIPv2 with leading text-driven hair editing methods on the CelebA-HQ [14] testset (2,000 images) and follow the evaluation settings of HairCLIP. For HairCLIP [36] and StyleCLIP [21] (\"Mapper\" version), we first invert using e4e [33] to obtain the latent code for a given real image before performing the editing. For Diffusion-CLIP [19], we finetune a model for each text description. For both TediGAN [40] and our method, the number of optimization iterations is set to 200. As shown in Figure 3 and Table 2, our method accomplishes satisfactory hair editing effects with better naturalness while maximizing the preservation of irrelevant attributes. It is worth noting that, even though HairCLIP and StyleCLIP also have pretty good hair", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "editing capabilities, they cannot preserve the irrelevant attributes very well such as background, identity and clothes. Our method also demonstrates better preservation of the original hair structure when editing only the hair color.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For arbitrary hair editing word scenarios, the only methods that are instantly feasible without retraining the model are HairCLIP, StyleCLIP (\"Optimization\" version), and TediGAN. As shown in Figure 4, our method performs much better at such cases. In contrast, HairCLIP can only produce plausible results for text (\"Curly Short Hairstyle\") similar to the training texts, while all other methods struggle to produce reasonable editing effects.", + "bbox": [ + 496, + 152, + 892, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Hair Transfer Methods. We compare with state-of-the-art methods on hair transfer tasks. Among the 2,000 images of the CelebA-HQ testset, the first 666 are set as the input images, the middle 666 are set as the hairstyle reference images, and the last 666 are set as the hair color reference images. As shown in Figure 5, when the hairstyle reference image is broadly aligned with the input image (first row), most methods yield plausible results. However, when not aligned (second row), only our method and SYH [20] are able to perform a more consistent hair transfer, which is achieved by introducing the pose alignment loss during the transfer to ensure that the facial shape and pose of the reference image are consistent with the source image. Compared to SYH, we achieve comparable hair transfer results, but support text, sketch, and other interactions beyond hair transfer.", + "bbox": [ + 496, + 276, + 892, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Local Hair Editing Methods. In terms of sketch-based local editing, we compare with the SOTA methods MichiGAN [31] and SketchSalon [41]. MichiGAN [31] uses user-drawn sketches to modify the orientation map to accomplish local hair editing. SketchSalon [41] trains a sketch-to-hair conditional translation network, with an additional soft alpha matte used to facilitate more natural blending. To generalize SketchSalon to local editing, we utilize the same mask as our method instead of a soft alpha matte, and the input sketch is colored as the average color within the mask area. As shown in Figure 6, MichiGAN struggles to perform satisfactory local editing and the reconstruction of other non-editing hair areas is slightly worse. Even ignoring the obvious blending artifacts, the local hair texture generated by SketchSalon is not in harmony with the surrounding hair. Compared to these two methods, our approach not only achieves satisfactory local editing but also better maintains the non-editing regions.", + "bbox": [ + 496, + 520, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Cross-Modal Hair Editing Methods. To the best of our knowledge, the only method that supports multimodal conditions to complete hairstyle and hair color editing is HairCLIP [36]. As the comparison shown in Figure 7, our approach not only perfectly prevents irrelevant attributes (identity, background, etc.) from being modified, but also achieves higher-quality editing effects. Moreover,", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23594", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34eb8bcd0cef34dbd678faaf46f38e040c8132ea49727736eb75122c84c88757.jpg", + "image_caption": [ + "Figure 3. Visual comparison with HairCLIP [36], StyleCLIP-Mapper [21], TediGAN [40] and DiffusionCLIP [19]. The simplified text descriptions (editing hairstyle, hair color, or both of them) are listed on the leftmost side. Our approach demonstrates better editing effects and irrelevant attribute preservation (e.g., identity, background, etc.)." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 898, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4649bda1ca7f272b1d9613d58d623d581a1c77638d4853b0893c664aef84bbc0.jpg", + "image_caption": [ + "Figure 4. Visual comparison with HairCLIP [36], StyleCLIP-Optimization [21] and TediGAN [40] under any description setting. We additionally provide an example image for each description for better comparison." + ], + "image_footnote": [], + "bbox": [ + 81, + 292, + 467, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HairCLIP allows only text and reference image while our method additionally supports sketch, mask, and RGB values. More diverse and comprehensive interactive editing results are shown in Figure 1 and supplementary materials.", + "bbox": [ + 75, + 550, + 468, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. For the above four types of comparisons, we recruit 20 volunteers with computer vision-related research backgrounds to execute a comprehensive user study. We randomly selected 20 groups of results from each experiment to form 80 test samples in total. The order of the different methods in each test sample is randomly shuffled. For each test sample, volunteers are asked to select the best option in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness, respectively. As shown in Table 3, our method outperforms the baseline methods for most cases, except comparable results to Barbershop [47] and SYH [20] in the hair transfer setting. But our irrelevant attribute preservation performs best because of our hair color feature space blending mechanism, as demonstrated in Figure 5. It is worth mentioning that our goal is not to improve the performance of hair transfer, but to design a unified system that supports various hair editing and hair transfer tasks. Therefore, performing comparably with the state-of-the-art methods on the hair transfer task is", + "bbox": [ + 75, + 613, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "acceptable.", + "bbox": [ + 500, + 291, + 576, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Analysis", + "text_level": 1, + "bbox": [ + 500, + 316, + 671, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of Initialization Strategy for Text Proxy Optimization. To justify our optimization starting point strategy for text proxy, we ablate three different optimization starting points: mean latent code, random sampling from $\\mathcal{W}$ space, and inverted latent code of the input image within $\\mathcal{W}+$ space. All other settings remain the same. As illustrated in Figure 8, the first two settings are more likely to complete high-quality editing, while the last one fails because the starting point deviates from the more suitable editing region. This may explain why StyleCLIP (\"Optimization\" version) and TediGAN struggle in performing hairstyle editing. Our strategy of randomly initializing the latent code around the mean within $\\mathcal{W}$ space enjoys both good editability and diversity of the generated results.", + "bbox": [ + 498, + 340, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Superiority of Sketch Proxy Generation Design. To generate sketch proxy, we ablate another two possible ways. The first way is to constrain the orientation field of the sketch proxy to be similar to that of user's sketch in the drawing region with the orientation loss used in MichiGAN[31]. The second way is to first obtain the image corresponding to the sketch through the conditional translation network (for simplicity, we choose sketch2hair here), and then constrain the sketch proxy to be similar to it in the drawing area by LPIPS [44] loss during the optimization process. The comparison is shown in Figure 9. Only our method and the LPIPS optimization-based version are feasible. However, our method requires only a single feed forward, which is more efficient.", + "bbox": [ + 496, + 551, + 892, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Blending vs. Latent Code Blending. To demonstrate the superiority of proxy feature blending, we compare it with the alternative scheme based on the linear combination of latent codes. In detail, we initialize an interpolation factor for the latent code of the global editing proxy, the local sketch proxy, and the input image, respectively. In the optimization process, we optimize these three interpolation factors so that the generated image corresponding to the interpolated latent code is similar to the proxies or input im", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebcb086fce6bb20533518fe772720a0fda213e8ec01ffc13f3e897e4daa454ea.jpg", + "image_caption": [ + "Figure 5. Visual comparison with HairCLIP [36], LOHO [24], Barbershop [47], SYH [20] and MichiGAN [31] on hair transfer. Only our method and SYH can accomplish unaligned hair transfer while keeping irrelevant attributes unmodified." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 883, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2547992fc858954dc5393df222325d1e084650f32942e41865f5ba54dfbd36eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Text-DrivenHair TransferSketch-BasedCross-Modal
MetricsOurs[36][21][40][19]Ours[36][24][47][20][31]Ours[31][41]Ours[36]
Accuracy41.5%32.3%22.5%1.0%2.8%28.0%2.3%4.8%29.3%28.5%7.3%76.8%22.0%1.3%82.8%17.3%
Preservation81.0%5.3%3.3%0.3%10.3%32.8%2.8%8.3%15.3%26.3%14.8%62.0%33.3%4.8%94.0%6.0%
Naturalness46.8%25.5%22.0%1.8%4.0%26.5%9.5%2.5%22.3%35.0%4.3%60.5%38.0%1.5%65.3%34.8%
", + "bbox": [ + 78, + 292, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. User study on text-driven image manipulation, hair transfer, sketch-based local hair editing and cross-modal hair editing methods. Accuracy denotes the manipulation accuracy for given conditional inputs, Preservation indicates the ability to preserve irrelevant regions and Naturalness denotes the visual realism of the manipulated image.", + "bbox": [ + 75, + 373, + 892, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0aba2e1a50f12be67d4c3708753dc3d84c8896f2a3319c92387b6f9d0566fb79.jpg", + "image_caption": [ + "Figure 6. Comparison with MichiGAN [31] & SketchSalon [41] on sketch-based local hair editing. We provide sketches in the facial parsing map for better visualization." + ], + "image_footnote": [], + "bbox": [ + 86, + 438, + 459, + 570 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c3548c36cf70617e6b98d144c2fc49c30e37749f0f87293fdf5269e6032b64f.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison with HairCLIP on cross-modal conditional input. Our approach shows better editing effects & excellent preservation of irrelevant attributes." + ], + "image_footnote": [], + "bbox": [ + 81, + 641, + 467, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age within the corresponding region. In Figure 10, we use", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b60d871addebd77c7ef3ad985820fb651897a5ad09bbfddeb57f95d3d4248c0f.jpg", + "image_caption": [ + "Figure 8. Ablation of different starting point settings for text proxy. The text description is \"Bob Cut Hairstyle\"." + ], + "image_footnote": [], + "bbox": [ + 506, + 439, + 887, + 597 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03a13bbab0f88b877b75793d23ed4e23c13cece812b23ec2dbd5ad48437a05aa.jpg", + "image_caption": [ + "Figure 9. Ablation on sketch proxy generation design. We provide sketches drawn in the facial parsing map for better visualization." + ], + "image_footnote": [], + "bbox": [ + 511, + 655, + 880, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the hairstyle reference image as an example to generate the global proxy. It is obvious that our scheme accomplishes global editing and local editing while perfectly keeping the", + "bbox": [ + 500, + 854, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23596", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7c89a31de86372f38bd31f33d5ab692d5e7ba4b703e0a716ffbec25c1bdd2f5.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 89, + 103, + 161, + 159 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f211568a8e6fe6a605adee23e81e7cbebc256439e14be4519cfe6e7154efc9e2.jpg", + "image_caption": [ + "Hairstyle Ref" + ], + "image_footnote": [], + "bbox": [ + 163, + 103, + 235, + 159 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/667c28a35a9aa5cbf65e7c3971f661e07569f0213d8a5f1f36561823ca9bdbd1.jpg", + "image_caption": [ + "Input Sketch" + ], + "image_footnote": [], + "bbox": [ + 236, + 103, + 308, + 159 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/66dd24697518234b8a20d4d9f88537628df65b9e36242407a38284f4a80f17b6.jpg", + "image_caption": [ + "Feature" + ], + "image_footnote": [], + "bbox": [ + 310, + 103, + 383, + 159 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0d705271d6e9012ee91d19621f2f910db456d524985e4f7a7236329ee8c0ae1a.jpg", + "image_caption": [ + "Latent Code" + ], + "image_footnote": [], + "bbox": [ + 385, + 103, + 457, + 159 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cc73e7b29a35d6a01fa2a7e98aa511868a5946d9a70701c357eafafd967b3673.jpg", + "image_caption": [ + "Figure 10. Ablation of feature blending vs. latent code blending. Feature blending can enable global editing, local editing, and irrelevant attributes preservation simultaneously." + ], + "image_footnote": [], + "bbox": [ + 89, + 161, + 161, + 217 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/356ac340bdc8f91a9f122bd0f08bc59e417fc9459ce502724eb941e13b453e15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 161, + 235, + 217 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ddfff9c5674db2e34c7c783f725aa50459df5329a1616fd00de43e4cf27fc828.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 161, + 308, + 217 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6ea7127313a5e0c06f038295aa076571c4193d52391eb20d8c7715fcc6345e7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 161, + 383, + 217 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a62eadfcb28dde362307aa7d6febc722b0260d64df006ae65847c62abfaec356.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 161, + 457, + 217 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "irrelevant properties unmodified, while latent code blending does not perform either one well.", + "bbox": [ + 76, + 300, + 467, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 76, + 345, + 323, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a unified hair editing system HairCLIPv2, which presents the first attempt that supports both simple text/reference image interaction and fine-grained local interactions. It innovatively converts all hair editing tasks into hair transfer tasks, with the corresponding editing conditions converted into transfer proxies. It can not only achieve high-quality hair editing results, but also well preserve the irrelevant attributes from being modified. In the future, we will study how to use feed-forward networks to generate all the proxies. Also, it is worthy to generalize the proposed framework to support generic natural images.", + "bbox": [ + 75, + 371, + 468, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 551, + 256, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386 and 62121002, the Fundamental Research Funds for the Central Universities under Grant WK5290000003, Key Research and Development program of Anhui Province under Grant 2022k07020008. This work was also partly supported by Shenzhen Key Laboratory of Media Security, and the Opening Fund of Key Laboratory of Cyberculture Content Cognition and Detection, Ministry of Culture and Tourism. This work was also partially supported by a GRF grant (Project No. CityU 11216122) from the Research Grants Council (RGC) of Hong Kong. Thank Yi Yin for her help in this work.", + "bbox": [ + 75, + 578, + 468, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 789, + 173, + 804 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019.", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In" + ], + "bbox": [ + 84, + 815, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020.", + "[3] Abdul Fatir Ansari, J. Scarlett, and Harold Soh. A characteristic function approach to deep implicit generative modeling. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7476-7484, 2020.", + "[4] Martín Arjovsky, Soumith Chintala, and L. Bottou. Wasserstein generative adversarial networks. In ICML, 2017.", + "[5] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022.", + "[6] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017.", + "[7] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020.", + "[8] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2023), 2023.", + "[9] Ruili Feng, Jie Xiao, Kecheng Zheng, Deli Zhao, Jingren Zhou, Qibin Sun, and Zheng-Jun Zha. Principled knowledge extrapolation with GANs. In Proceedings of the 39th International Conference on Machine Learning, volume 162, pages 6447-6464. PMLR, 17-23 Jul 2022.", + "[10] Ruili Feng, Deli Zhao, and Zheng-Jun Zha. Understanding noise injection in gans. In International Conference on Machine Learning, pages 3284-3293. PMLR, 2021.", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016.", + "[12] Ishaan Gulrajani, F. Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C. Courville. Improved training of wasserstein gans. In NIPS, 2017.", + "[13] Youngjoo Jo and Jongyoul Park. Sc-fegan: Face editing generative adversarial network with user's sketch and color. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1745-1753, 2019.", + "[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018.", + "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020.", + "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021." + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019.", + "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020.", + "[19] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022.", + "[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022.", + "[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021.", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021.", + "[24] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1984-1993, 2021.", + "[25] Edgar Schonfeld, B. Schiele, and A. Khoreva. A u-net based discriminator for generative adversarial networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8204-8213, 2020.", + "[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020.", + "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[28] Xinhui Song, Chen Liu, Youyi Zheng, Zunlei Feng, Lincheng Li, Kun Zhou, and Xin Yu. Hairstyle editing via parametric controllable strokes. IEEE Transactions on Visualization and Computer Graphics, 2023.", + "[29] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "linguistic representations. In International Conference on Learning Representations, 2019.", + "[30] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021.", + "[31] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. ACM Transactions on Graphics (TOG), 39(4):95-1, 2020.", + "[32] Zhentao Tan, Dongdong Chen, Qi Chu, Menglei Chai, Jing Liao, Mingming He, Lu Yuan, Gang Hua, and Nenghai Yu. Efficient semantic image synthesis via class-adaptive normalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021.", + "[33] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021.", + "[34] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021.", + "[35] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS 2022), 2022.", + "[36] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022.", + "[37] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2style: Improve the efficiency and effectiveness of stylegan inversion. IEEE Transactions on Image Processing, 31:3267-3280, 2022.", + "[38] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023.", + "[39] Yiqian Wu, Yong-Liang Yang, and Xiaogang Jin. Hairmapper: Removing hair from portraits using gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4227-4236, 2022.", + "[40] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[41] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: Deep sketch-based hair image synthesis. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23598", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018.", + "[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021.", + "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018.", + "[45] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copypaste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning (ICML 2023), 2023.", + "[46] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020.", + "[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021.", + "[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In European Conference on Computer Vision, pages 651-667. Springer, 2022.", + "[49] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved stylegan embedding: Where are the good latents? arXiv preprint arXiv:2012.09036, 2020.", + "[50] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020." + ], + "bbox": [ + 78, + 90, + 470, + 667 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23599", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_model.json b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e3e715f7d10c884e13c66cdad38c71600c88a63 --- /dev/null +++ b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.812, + 0.155 + ], + "angle": 0, + "content": "HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.18, + 0.721, + 0.217 + ], + "angle": 0, + "content": "Tianyi Wei\\(^{1}\\), Dongdong Chen\\(^{2}\\), Wenbo Zhou\\(^{1,\\dagger}\\), Jing Liao\\(^{3}\\), Weiming Zhang\\(^{1}\\), Gang Hua\\(^{4}\\), Nenghai Yu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.218, + 0.717, + 0.233 + ], + "angle": 0, + "content": "1University of Science and Technology of China 2Microsoft Cloud AI" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.235, + 0.679, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\)City University of Hong Kong \\(^{4}\\)Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.255, + 0.714, + 0.269 + ], + "angle": 0, + "content": "{bestwty@mail., welbeckz@, zhangwm@, yn@}ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.273, + 0.706, + 0.287 + ], + "angle": 0, + "content": "{cddlyf@, ganghua@}gmail.com, jingliao@cityu.edu.hk" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.324, + 0.886, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.478, + 0.895, + 0.507 + ], + "angle": 0, + "content": "Figure 1. HairCLIPv2 supports hairstyle and color editing individually or jointly with unprecedented user interaction mode support, including text, mask, sketch, reference image, etc." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.537, + 0.314, + 0.552 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.473, + 0.873 + ], + "angle": 0, + "content": "Hair editing has made tremendous progress in recent years. Early hair editing methods use well-drawn sketches or masks to specify the editing conditions. Even though they can enable very fine-grained local control, such interaction modes are inefficient for the editing conditions that can be easily specified by language descriptions or reference images. Thanks to the recent breakthrough of cross-modal models (e.g., CLIP), HairCLIP is the first work that enables hair editing based on text descriptions or reference images. However, such text-driven and reference-driven interaction modes make HairCLIP unable to support fine-grained controls specified by sketch or mask. In this paper, we propose HairCLIPv2, aiming to support all the aforementioned interactions with one unified framework. Simultaneously, it improves upon HairCLIP with better irrelevant attributes (e.g., identity, background) preservation and unseen text descriptions support. The key idea is to convert all the hair editing tasks into hair transfer tasks, with editing conditions converted into different proxies accordingly. The editing effects are added upon the input image by blend" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.524, + 0.895, + 0.63 + ], + "angle": 0, + "content": "ing the corresponding proxy features within the hairstyle or hair color feature spaces. Besides the unprecedented user interaction mode support, quantitative and qualitative experiments demonstrate the superiority of HairCLIPv2 in terms of editing effects, irrelevant attribute preservation and visual naturalness. Our code is available at https: //github.com/wty-ustc/HairCLIPv2." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.662, + 0.633, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Hair editing as an interesting and challenging problem has attracted a lot of research attention from both academia and industry. Over the past few decades, tremendous progress [50, 41, 31, 36] has been made in this field, enabling high-fidelity hair editing based on various types of user interactions or controls. In earlier hair editing methods [50, 41], commonly supported user editing conditions are sketches and masks, which can enable fine-grained local controls. But in real scenarios, many hair editing conditions can be specified by simpler interactions, e.g., text descriptions (e.g., \"bowl cut hairstyle\") and reference images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, cross-modal visual and language representation learning [29, 22, 43, 8, 35, 38, 45] has made remarkable breakthrough, which makes text-guided image manipu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.332, + 0.901 + ], + "angle": 0, + "content": "† Wenbo Zhou is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23589" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.196 + ], + "angle": 0, + "content": "
HairCLIP [36]LOHO [24]Barbershop [47]HairNet [48]SYH [20]MichiGAN [31]SketchSalon [41]Ours
Aligned Hair Transfer✓✓✓✓✓✓✗✓
Unaligned Hair Transfer✓✗✗✓✓✗✗✓
Text✓✗✗✗✗✗✗✓
Mask✗✗✓✗✓✓✗✓
Sketch✗✗✗✗✗✓✓✓
Local Hairstyle Editing✗✗✗✗✗✓✓✓
Local Hair Color Editing✗✗✗✗✗✗✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.202, + 0.893, + 0.23 + ], + "angle": 0, + "content": "Table 1. Comparisons between our approach and mainstream hair editing methods in terms of available interaction modes and functionality. Only our method supports all interaction modes and enables both global and local hair editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.473, + 0.452 + ], + "angle": 0, + "content": "lation possible. HairCLIP [36] presents the first attempt that supports hair editing via text description and reference image within one unified framework. Despite such text-driven and reference-driven interaction being more efficient and user-friendly, HairCLIP cannot support fine-grained controls like sketches and masks. Moreover, HairCLIP has two other limitations: 1) Since hair editing in HairCLIP is accomplished by pure latent code manipulation, it will inevitably alter other irrelevant attributes (e.g., identity, background) because fully decoupling different attributes in latent codes is difficult; 2) It struggles in yielding satisfactory results for text descriptions that differ significantly from training texts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.606 + ], + "angle": 0, + "content": "In this paper, we take a step forward and propose HairCLIPv2, a unified hair editing system that unprecedentedly supports all the aforementioned interaction modes, including the natural text/reference-driven interaction and fine-grained local interaction. In Table 1, we list the interaction modes and editing functionality supported by existing hair editing methods. Moreover, with fundamentally different editing mechanism design, HairCLIPv2 makes great improvement upon HairCLIP with better irrelevant attributes preservation and unseen text description support." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.732 + ], + "angle": 0, + "content": "The key idea of HairCLIPv2 is converting all the hair editing tasks into hair transfer tasks, and the editing conditions are converted into different transfer proxies accordingly. Conceptually, it can be understood as \"find proxy hair images that satisfy the editing conditions and transfer the corresponding attributes to the source image\". Note that we use the StyleGAN latent code or feature corresponding to such proxies rather than use the proxy images explicitly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "More specifically, we first transform the input source image into the bald proxy, which inpaints the hair-covered regions (e.g., background, ears) with reasonable semantic attributes. This can help avoid the editing artifacts caused by occlusion when blending the source image with condition proxies. For different editing proxies, we define their generation as different tasks performed in StyleGAN according to their characteristics. Depending on the users' editing preferences, hair editing effects are then enforced upon the input image by blending the corresponding proxy features within the hairstyle feature space or hair color feature space. This" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.895, + 0.331 + ], + "angle": 0, + "content": "is different from HairCLIP that achieves the editing effect by manipulating the 1-d latent codes. Such feature blending based editing naturally supports global and local hair editing by controlling the blending area to cover the entire hair area or part of it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.894, + 0.437 + ], + "angle": 0, + "content": "To show the superiority of HairCLIPv2, we conduct extensive comparisons. In addition to more complete user interaction modes support, HairCLIPv2 also shows obvious advantages in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness. Some interactive editing examples are provided in Figure 1. Our contributions can be summarized as below:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.45, + 0.892, + 0.557 + ], + "angle": 0, + "content": "- We present a fresh perspective for hair editing tasks and propose a novel hair editing paradigm that unifies various types of editing into the form of proxy hair transfers. We achieve all editing effects with the feature blending mechanism, which not only alleviates the editing pressure on each proxy but also enables excellent irrelevant attribute preservation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.571, + 0.892, + 0.722 + ], + "angle": 0, + "content": "- We dedicate design to the proxy generation for different conditions based on their own special properties, e.g., for the text proxy, the decoupled proxy design and optimization starting point selection strategy help us achieve better editing effects and arbitrary text support; for the sketch proxy, we achieve local hairstyle editing support for the first time within the StyleGAN-based framework by formalizing its generation as the image translation task and incorporating insights of semantic layering in StyleGAN." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.737, + 0.894, + 0.811 + ], + "angle": 0, + "content": "- Our system pushes the interactions of hair editing to a new level, supporting arbitrary text, mask, reference image, sketch and their combinations, and enabling both global and local hair editing, which has never been realized before." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.45, + 0.894, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.642, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Generative Adversarial Networks. Since being invented, GANs have made considerable progress in terms of training strategies [18, 34], loss functions [3, 4, 9], and network" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23590" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "structures [25, 12, 10, 30, 32]. In the field of image synthesis, a series of works called StyleGAN [17, 18, 15, 16] represents the cutting edge of GANs. Given its promising semantically decoupled latent space [7, 26] and high-quality image synthesis abilities, the pre-trained StyleGAN has become the preferred choice for performing image editing. In this paper, we choose StyleGAN2 to develop our framework, which is consistent with other hair editing methods [36, 47, 20, 21, 40, 24] to be compared." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.473, + 0.583 + ], + "angle": 0, + "content": "Latent Space Embedding and Editing. As the bridge connecting the pre-trained StyleGAN and other downstream editing tasks, GAN inversion aims to yield the ideal embedding of the real image in the latent space. Based on the application purposes, we roughly classify the GAN inversion methods into two categories: methods [33, 46, 49] suitable for editing and methods [2, 23, 37] for better reconstruction. The former methods project the real image into the embedding subspace more suitable for editing at the expense of reconstruction. Among them, e4e [33] has become the most popular method for editing tasks [36, 21, 39] performed in the latent space. The latter approaches [2, 23, 37] aim to achieve the perfect reconstruction of the real image. However, limited by the representation capability of the latent space, all these methods cannot achieve the perfect reconstruction. To address this issue, Barbershop [47] proposes a novel inversion method, which additionally introduces a feature space \\(\\mathcal{F}\\) of StyleGAN combined with the latent space \\(S\\) to form a new embedding space \\(\\mathcal{FS}\\). Inspired by this, we decouple the editing task from the reconstruction task by blending editing proxy features in the feature space to achieve a unified hair editing system that supports a wide range of interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.856 + ], + "angle": 0, + "content": "Hair Editing Using GANs. Existing hair editing methods can be roughly categorized as conditional GANs [41, 31, 13] based and pre-trained StyleGAN based [36, 21, 40, 24, 47, 20, 48, 28]. As a pioneering work of hair transfer, MichiGAN [31] accomplishes hairstyle transfer by extracting the orientation map of the reference image. Barbershop [47] performs hair transfer within their proposed \\(\\mathcal{FS}\\) embedding space. But these methods often struggle when large pose differences exist between source and target image. Recently, some improvements [20, 48] on Barbershop make pose unaligned hair transfers possible. Our framework is also compatible with pose unaligned hair transfers and additionally offers more interaction modes. SketchHairSalon [41] enables local editing of hairstyle and hair color by using colored sketches as the input to the conditional translation network. Unlike them, we show for the first time that the StyleGAN-based framework can also perform local hair editing with sketches as the condition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Benefiting from the development of cross-modal models [29, 22], text-guided hair editing [36, 21, 40] has become the new trend. StyleCLIP [21] and TediGAN [40] uti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "lize CLIP loss to perform hair editing in an optimized manner. However, since the embedding of real image deviates from the original suitable editing latent space, these methods will fail for some cases. HairCLIP [36] alleviates the problem by training a hair mapper on a large-scale dataset, but struggles in yielding good results for descriptions that differ significantly from the training text. Moreover, none of these methods can preserve the irrelevant attributes well. In this work, we present a new perspective to enable text-guided hair editing methods. By decoupling the editing task from the reconstruction, we can better preserve the irrelevant attributes while enabling high-quality hair manipulation via arbitrary text descriptions. More importantly, there is no prior work that supports so many interaction modes and functionality as we offer." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.333, + 0.673, + 0.35 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.358, + 0.642, + 0.373 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.518 + ], + "angle": 0, + "content": "StyleGAN [18] can synthesize photorealistic images with a progressive upsampling network consisting of 18 layers. Its \\(\\mathcal{W} \\subsetneq \\mathbb{R}^{512}\\) latent space exhibits good semantic decoupling properties [7, 26], thus enabling various editing tasks. To perform semantic editing while ensuring the reconstruction quality, some inversion methods [1, 23, 37] extend the original \\(\\mathcal{W}\\) space to \\(\\mathcal{W}+\\) space, which is defined as a cascade of 18 different latent codes \\(w_{i}\\) from the \\(\\mathcal{W}\\) space, i.e., \\(\\mathcal{W} + \\subsetneq \\mathbb{R}^{18\\times 512}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.625 + ], + "angle": 0, + "content": "\\(\\mathcal{FS}\\) Embedding Space is proposed by Barbershop [47], which is designed to increase the representational capability of embedding space for details and enable the spatial control of image features. The new \\(\\{F_7,S\\}\\) latent code replaces the first 7 layers of \\(\\mathcal{W}+\\) latent code with the \\(32\\times 32\\times 512\\) features \\(F_{7}\\) of style-block 7 of StyleGAN \\(G\\), i.e., \\(F_{7}\\in \\mathcal{F}\\subsetneq \\mathbb{R}^{32\\times 32\\times 512},S = [w_8,\\dots w_i,\\dots w_{18}],w_i\\in \\mathcal{W}.\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.671 + ], + "angle": 0, + "content": "CLIP [22] is a multi-modality model pretrained on web-scale image-text pairs. It can well measure the semantic similarity between given image and text." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.611, + 0.696 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Since the \\(\\mathcal{FS}\\) embedding space [47] is proposed, performing seamless feature blending in \\(\\mathcal{F}\\)-space has become the de facto standard for many hair transfer works [47, 20, 48], because it can encode the spatial information and preserve local details. On the other hand, many hair editing efforts [36, 21, 40, 39] choose to perform editing in the \\(\\mathcal{W}+\\) space, despite unsatisfactory reconstruction, because it can encode rich disentangled semantics [7, 26]. Considering that \\(\\mathcal{F}\\)-space is expressive and enables realistic feature integration results while \\(\\mathcal{W}+\\) space is editable, we therefore wonder \"Can we enjoy the best features of both spaces to facilitate the hair editing task?\" To achieve this goal, we formulate the hair editing tasks as the hair transfer tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "23591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.891, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.316 + ], + "angle": 0, + "content": "Figure 2. Overview of HairCLIPv2: Example with hairstyle description text, sketch, mask and hair color RGB values as conditional inputs. Corresponding proxy images are just for better understanding. We complete the hair editing by converting different conditions into different proxies and achieve editing effects by blending them in StyleGAN feature spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.429 + ], + "angle": 0, + "content": "Specifically, we convert all editing conditions (e.g., text, reference image, sketch) into different proxies in the \\(\\mathcal{W}+\\) space, and accomplish hair editing by seamless proxy feature blending in the feature spaces of StyleGAN. The proxy features of different conditions are obtained with tailored methods based on the condition characteristics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.583 + ], + "angle": 0, + "content": "Following the design in HairCLIP [36], we edit hairstyle and hair color sequentially by blending editing proxy features in the early and later StyleGAN feature space respectively. In detail, as shown in Figure 2, we choose to perform proxy feature blending of hairstyle and hair color on feature spaces \\(\\mathcal{F}^{style} \\subsetneq \\mathbb{R}^{32 \\times 32 \\times 512}\\) and \\(\\mathcal{F}^{color} \\subsetneq \\mathbb{R}^{256 \\times 256 \\times 128}\\), which correspond to the features of 7-th and 14-th style-block in StyleGAN respectively. And users can select various interactions (global or local, single or combined) to edit hairstyles and hair color individually or jointly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.592, + 0.409, + 0.608 + ], + "angle": 0, + "content": "3.3. Converting Input Image to Bald Proxy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.469, + 0.736 + ], + "angle": 0, + "content": "Given a source image \\( I^{src} \\) to be edited, we obtain its latent code \\( w^{src} \\) in \\( \\mathcal{W}+ \\) space and feature \\( F_{7}^{src} \\) in \\( \\mathcal{FS} \\) space by II2S [49] and FS embedding algorithms [47], respectively. Balding proxy is then generated to inpaint the hair-covered region with reasonable semantic attributes (e.g., background, ears, face, etc.), which can avoid editing artifacts due to occlusion when blending the original image with different editing proxies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.469, + 0.873 + ], + "angle": 0, + "content": "Balding Proxy. To remove the occlusion from the hair area of the source image, we bald it using HairMapper [39]: \\( w^{bald} = \\mathbb{B}(w^{src}) \\). \\( \\mathbb{B} \\) denotes HairMapper, which completes the balding editing operation in \\( \\mathcal{W}+ \\) space on the latent code \\( w^{src} \\) to yield the latent code \\( w^{bald} \\) corresponding to the balded source image. Since editing in \\( \\mathcal{W}+ \\) space inevitably gets other irrelevant attributes modified, we circumvent this issue by blending bald feature with source image feature in \\( \\mathcal{F}^{style} \\) space:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.883, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {7} ^ {b a l d} = G \\left(w _ {1 - 7} ^ {b a l d}\\right) \\times M ^ {b a l d} + F _ {7} ^ {s r c} \\times \\left(1 - M ^ {b a l d}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.447 + ], + "angle": 0, + "content": "where \\( G \\) stands for StyleGAN, \\( G(w_{1-7}^{bold}) \\) represents the bald feature in \\( \\mathcal{F}^{style} \\) space. \\( M^{bald} \\) is the binary mask indicating the hair and ear regions of the source image, which is obtained by the facial parsing network BiseNET [42] and downsampled to \\( 32 \\times 32 \\). With the guidance of \\( M^{bald} \\), the irrelevant attributes region in \\( F_7^{bald} \\) can continue to be preserved in following proxy feature blending." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.454, + 0.671, + 0.471 + ], + "angle": 0, + "content": "3.4. Hairstyle Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.891, + 0.507 + ], + "angle": 0, + "content": "Below we elaborate on how to generate the proxy for different hairstyle conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Text Proxy. We formalize text proxy generation as the editing task done in \\(\\mathcal{W}+\\) space based on the CLIP loss guidance. Unlike prior works [36, 21, 40], our text proxy generation process is free from the pressure of irrelevant attribute preservation, and allows us to select a more suitable starting point for the optimization process, which leads to better editing effects. In order to ensure both the optimal editing effect and the diversity of editing results, we choose to sample a random point around the mean face latent code as the optimization starting point for our text proxy latent code \\(w^{text}\\). In detail, we adopt the truncation trick of StyleGAN as \\(w^{init} = w^{mean} + \\psi (w^{random} - w^{mean})\\), where \\(w^{mean}\\) is the mean face latent code and \\(w^{random}\\) is sampled randomly. By setting a small value of \\(\\psi\\), we ensure that the initial optimization starting point \\(w^{init}\\) is around the average face latent code \\(w^{mean}\\). We will show the benefits of this initialization strategy in the ablation analysis. We adopt the CLIP loss with transformation augmentations [5] to perform the text-guided hairstyle editing while reducing the disturbance caused by adversarial examples:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.82, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\nL ^ {c l i p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} (1 - \\cos \\left(E _ {i} \\left(A _ {i} \\left(G \\left(w ^ {t e x t}\\right)\\right)\\right), E _ {t} (s t)\\right)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(A_{i}\\) represents \\(i\\)-th transformation augmentation, \\(N\\) denotes the number of augmentations (\\(N = 4\\) by default)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.198 + ], + "angle": 0, + "content": "\\(\\cos (\\cdot)\\) means cosine similarity, \\(G(w^{text})\\) means the editing result for each pass in the optimization process, \\(E_{i}\\) and \\(E_{t}\\) stands for the CLIP image encoder and text encoder respectively, and \\(st\\) refers to the user-supplied text description. Besides, pose alignment loss \\(L^{pose}\\) is utilized to ensure that the face shape and pose of text proxy are consistent with the source image to ease the subsequent feature blending:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.207, + 0.47, + 0.238 + ], + "angle": 0, + "content": "\\[\nL ^ {p o s e} = \\frac {1}{N _ {k}} | | E _ {p} (I ^ {s r c}) - E _ {p} (G (w ^ {t e x t})) | | _ {2} ^ {2}, \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.245, + 0.47, + 0.336 + ], + "angle": 0, + "content": "where \\( E_{p} \\) represents the 3D keypoint extractor [6] and \\( N_{k} \\) denotes the number of keypoints. Optionally, the shape loss \\( L^{shape} \\) is added to constrain the shape of the generated hair according to whether the user provides the hair region mask. We then obtain text proxy feature \\( F_{7}^{text} \\) using the optimized \\( w^{text}: F_{7}^{text} = G(w_{1-7}^{text}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.471, + 0.518 + ], + "angle": 0, + "content": "Reference Proxy. Given a hairstyle reference image \\( I^{sr} \\), we generate the reference proxy by performing the unaligned hairstyle transfer task. \\( I^{sr} \\) is first inverted by the II2S [49] embedding algorithm into the \\( \\mathcal{W}+ \\) space to get \\( w^{ref} \\), which is served as the starting point for hairstyle transfer. During the transfer process, we expect to keep the original hair structure of the reference image while ensuring its pose and facial shape to be consistent with the source image. Thus, \\( L^{pose} \\) is imposed between \\( G(w^{ref}) \\) and \\( I^{src} \\). In addition, a style loss \\( L^{style} \\) based on the gram matrix [11] is used to ensure that the hairstyle structure of \\( I^{sr} \\) remains unchanged during the alignment process:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.532, + 0.469, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L ^ {s t y l e} = \\frac {1}{4} \\sum_ {i = 1} ^ {4} \\left\\| \\mathcal {G} _ {i} \\left(V G G _ {i} \\left(I ^ {s r} \\times M ^ {r h}\\right)\\right) \\right. \\\\ \\left. - \\mathcal {G} _ {i} \\left(V G G _ {i} \\left(G \\left(w ^ {r e f}\\right) \\times M ^ {g h}\\right)\\right) \\right| | _ {2} ^ {2}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.724 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_i(\\gamma_i)\\) represents the gram matrix calculated on the \\(i\\)-th layer features and a total of 4 layers of features are extracted, i.e., \\(\\{relu1\\_2, relu2\\_2, relu3\\_3, relu4\\_3\\}\\) of VGG [27]. \\(M^{rh}\\) and \\(M^{gh}\\) are hair masks for the reference image and the generated image of each round during optimization process predicted by BiseNET [42]. For the same purpose, the \\(L_2\\) norm of the manipulation magnitude in the latent space is utilized during optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.732, + 0.47, + 0.752 + ], + "angle": 0, + "content": "\\[\nL ^ {r e g} = \\left\\| w _ {t} ^ {r e f} - w _ {t - 1} ^ {r e f} \\right\\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\( w_{t}^{ref} \\) and \\( w_{t-1}^{ref} \\) represent the latent code of the current step and the previous step, respectively. Similar to the text proxy, \\( L^{shape} \\) is optionally added to allow the user to customize the shape of the hair. We then obtain reference proxy feature \\( F_{7}^{ref} \\) using the optimized \\( w^{ref} \\): \\( F_{7}^{ref} = G(w_{1-7}^{ref}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Sketch Proxy. Enabling sketch-based local hairstyle editing within our framework is nontrivial. It is hard to find suitable losses to constrain the local hairstyle structure to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.334 + ], + "angle": 0, + "content": "conform to the sketch given by the user. To circumvent this problem, we innovatively formalize the synthesis of sketch proxy as an image translation task based on StyleGAN. Utilizing the sketch-hair dataset created by SketchHairSalon [41], we train a sketch2hair inverter \\( T \\), which is based on E2Style [37] and aims to find the most appropriate latent code in \\( \\mathcal{W}+ \\) space to accurately translate a given sketch to the corresponding hair structure. The training loss consists of regular pixel-level \\( L_{2} \\) loss, feature-level LPIPS [44] loss and multi-layer face parsing loss [37] which is introduced to provide more local supervision. During the training process, we randomly remove a portion of the strokes to make our sketch2hair inverter adapt to a variety of sketch inputs from fine to coarse, e.g., even just one stroke. Given a local hairstyle sketch \\( S \\), our sketch proxy features \\( F_{7}^{sketch} \\) are synthesized by pre-trained sketch2hair \\( T \\) and StyleGAN \\( G \\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.346, + 0.892, + 0.365 + ], + "angle": 0, + "content": "\\[\nw ^ {\\text {s k e t c h}} = T (S), \\quad F _ {7} ^ {\\text {s k e t c h}} = G \\left(w _ {1 - 7} ^ {\\text {s k e t c h}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Proxy Feature Blending. For text and reference image based hairstyle condition, we perform global blending in \\(\\mathcal{F}^{style}\\) space:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.423, + 0.892, + 0.443 + ], + "angle": 0, + "content": "\\[\nF _ {7} ^ {\\text {g l o b a l}} = F _ {7} ^ {t r} \\times M ^ {\\text {g l o b a l}} + F _ {7} ^ {\\text {b a l d}} \\times \\left(1 - M ^ {\\text {g l o b a l}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.501 + ], + "angle": 0, + "content": "where \\( F_7^{tr} \\in \\{F_7^{text}, F_7^{ref}\\} \\) and \\( M^{global} \\) is the binary mask corresponding to the hair region of \\( F_7^{tr} \\). Optionally, the sketch-based local hairstyle editing is applied:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.508, + 0.892, + 0.528 + ], + "angle": 0, + "content": "\\[\nF _ {7} ^ {\\text {s t y l e}} = F _ {7} ^ {\\text {s k e t c h}} \\times M ^ {\\text {l o c a l}} + F _ {7} ^ {\\text {g l o b a l}} \\times \\left(1 - M ^ {\\text {l o c a l}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.674 + ], + "angle": 0, + "content": "where \\( M^{local} \\) is obtained by downsampling the user input sketch \\( S \\) to \\( 32 \\times 32 \\) after dilation. A natural concern is the artifacts brought by the mismatch between hair features within \\( M^{local} \\) and other hair features. But thanks to the semantic layering characteristics of StyleGAN, the resulting image shows consistent tones as these hair features will be modulated by the later layers. Our framework allows users to only edit the hairstyle: \\( I^{style} = G(F_7^{style}, w_{8-18}^{src}) \\), by skipping the following hair color editing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.684, + 0.699 + ], + "angle": 0, + "content": "3.5. Hair Color Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.874 + ], + "angle": 0, + "content": "We achieve hair color editing by performing proxy feature blending in \\(\\mathcal{F}^{color}\\) space. By choosing to use the feature \\(F_{7}^{src}\\) or \\(F_{7}^{style}\\), we allow to edit only hair color or both hair style&hair color. Below, we use \\(F_{7}^{style}\\) as the example. Color Proxy. We initialize the color proxy with \\(F_{7}^{style}\\) and \\(w_{8-18}^{color} = w_{8-18}^{src}\\), and set \\(w_{10-13}^{color}\\) to be estimizable. The loss \\(L^{color}\\) in the optimization process consists of \\(L^{modal}\\) and \\(L^{bg}\\), where \\(L^{modal}\\) can be defined as \\(L^{clip}\\) or the average color \\(L_{2}\\) loss of the hair region depending on the hair color condition types (text, reference image, RGB values), and \\(L^{bg}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nL ^ {b g} = \\left\\| \\left(I ^ {\\text {s t y l e}} - I ^ {\\text {c o l o r}}\\right) \\times \\left(M ^ {n - h a i r}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23593" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.088, + 0.451, + 0.183 + ], + "angle": 0, + "content": "
MethodsIDS↑PSNR↑SSIM↑
Ours0.8429.50.91
HairCLIP [36]0.4521.60.74
StyleCLIP [21]0.4319.60.72
TediGAN [40]0.1622.50.74
DiffCLIP [19]0.7126.80.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.193, + 0.47, + 0.237 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison for irrelevant attributes preservation. IDS denotes identity similarity, PSNR and SSIM are calculated at the intersected non-hair regions before and after editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.291 + ], + "angle": 0, + "content": "where \\( I^{color} = G(F_7^{style}, w_{8-18}^{color}) \\), \\( M^{n-hair} \\) is the mask of the non-hair region intersection between \\( I^{style} \\) and \\( I^{color} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.469, + 0.368 + ], + "angle": 0, + "content": "Proxy Feature Blending. Even with the \\(L^{bg}\\) constraint, we find non-hair regions are often inevitably modified because of imperfect semantic decoupling of the \\(\\mathcal{W}+\\) space. We solve this by performing proxy feature blending in \\(\\mathcal{F}^{color}\\) space, which also naturally supports local hair color editing:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.397, + 0.468, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {1 4} ^ {\\text {b l e n d}} = G \\left(F _ {7} ^ {\\text {s t y l e}}, w _ {8 - 1 4} ^ {\\text {c o l o r}}\\right) \\times M ^ {\\text {c o l o r}} \\\\ + G \\left(F _ {7} ^ {\\text {s t y l e}}, w _ {8 - 1 4} ^ {\\text {s r c}}\\right) \\times \\left(1 - M ^ {\\text {c o l o r}}\\right), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.469, + 0.573 + ], + "angle": 0, + "content": "where \\( M^{color} \\) is the hair area mask or a local editing area mask drawn by the user. We set \\( F_{14}^{blend} \\) and \\( w_{15-18}^{color} \\) to be estimizable to further perform the optimization. In the optimization process, we use the \\( L_{blend} \\) loss, which consists of \\( L_{2} \\) loss and LPIPS loss to constrain \\( I^{final} \\) to be similar to \\( I^{color} \\) inside \\( M^{color} \\) and similar to \\( I^{style} \\) outside \\( M^{color} \\) simultaneously. The final edited image is synthesized as follows: \\( I^{final} = G(F_{14}^{blend}, w_{15-18}^{color}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.586, + 0.208, + 0.603 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.469, + 0.657 + ], + "angle": 0, + "content": "Implementation details of our approach are provided in the supplementary material. For all compared methods, we use their official codes or pre-trained models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.433, + 0.683 + ], + "angle": 0, + "content": "4.1. Quantitative and Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Comparison with Text-Driven Hair Editing Methods. We compare HairCLIPv2 with leading text-driven hair editing methods on the CelebA-HQ [14] testset (2,000 images) and follow the evaluation settings of HairCLIP. For HairCLIP [36] and StyleCLIP [21] (\"Mapper\" version), we first invert using e4e [33] to obtain the latent code for a given real image before performing the editing. For Diffusion-CLIP [19], we finetune a model for each text description. For both TediGAN [40] and our method, the number of optimization iterations is set to 200. As shown in Figure 3 and Table 2, our method accomplishes satisfactory hair editing effects with better naturalness while maximizing the preservation of irrelevant attributes. It is worth noting that, even though HairCLIP and StyleCLIP also have pretty good hair" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "editing capabilities, they cannot preserve the irrelevant attributes very well such as background, identity and clothes. Our method also demonstrates better preservation of the original hair structure when editing only the hair color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.893, + 0.275 + ], + "angle": 0, + "content": "For arbitrary hair editing word scenarios, the only methods that are instantly feasible without retraining the model are HairCLIP, StyleCLIP (\"Optimization\" version), and TediGAN. As shown in Figure 4, our method performs much better at such cases. In contrast, HairCLIP can only produce plausible results for text (\"Curly Short Hairstyle\") similar to the training texts, while all other methods struggle to produce reasonable editing effects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.893, + 0.519 + ], + "angle": 0, + "content": "Comparison with Hair Transfer Methods. We compare with state-of-the-art methods on hair transfer tasks. Among the 2,000 images of the CelebA-HQ testset, the first 666 are set as the input images, the middle 666 are set as the hairstyle reference images, and the last 666 are set as the hair color reference images. As shown in Figure 5, when the hairstyle reference image is broadly aligned with the input image (first row), most methods yield plausible results. However, when not aligned (second row), only our method and SYH [20] are able to perform a more consistent hair transfer, which is achieved by introducing the pose alignment loss during the transfer to ensure that the facial shape and pose of the reference image are consistent with the source image. Compared to SYH, we achieve comparable hair transfer results, but support text, sketch, and other interactions beyond hair transfer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Comparison with Local Hair Editing Methods. In terms of sketch-based local editing, we compare with the SOTA methods MichiGAN [31] and SketchSalon [41]. MichiGAN [31] uses user-drawn sketches to modify the orientation map to accomplish local hair editing. SketchSalon [41] trains a sketch-to-hair conditional translation network, with an additional soft alpha matte used to facilitate more natural blending. To generalize SketchSalon to local editing, we utilize the same mask as our method instead of a soft alpha matte, and the input sketch is colored as the average color within the mask area. As shown in Figure 6, MichiGAN struggles to perform satisfactory local editing and the reconstruction of other non-editing hair areas is slightly worse. Even ignoring the obvious blending artifacts, the local hair texture generated by SketchSalon is not in harmony with the surrounding hair. Compared to these two methods, our approach not only achieves satisfactory local editing but also better maintains the non-editing regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison with Cross-Modal Hair Editing Methods. To the best of our knowledge, the only method that supports multimodal conditions to complete hairstyle and hair color editing is HairCLIP [36]. As the comparison shown in Figure 7, our approach not only perfectly prevents irrelevant attributes (identity, background, etc.) from being modified, but also achieves higher-quality editing effects. Moreover," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.9, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.224, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Figure 3. Visual comparison with HairCLIP [36], StyleCLIP-Mapper [21], TediGAN [40] and DiffusionCLIP [19]. The simplified text descriptions (editing hairstyle, hair color, or both of them) are listed on the leftmost side. Our approach demonstrates better editing effects and irrelevant attribute preservation (e.g., identity, background, etc.)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.293, + 0.468, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Figure 4. Visual comparison with HairCLIP [36], StyleCLIP-Optimization [21] and TediGAN [40] under any description setting. We additionally provide an example image for each description for better comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.612 + ], + "angle": 0, + "content": "HairCLIP allows only text and reference image while our method additionally supports sketch, mask, and RGB values. More diverse and comprehensive interactive editing results are shown in Figure 1 and supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.901 + ], + "angle": 0, + "content": "User Study. For the above four types of comparisons, we recruit 20 volunteers with computer vision-related research backgrounds to execute a comprehensive user study. We randomly selected 20 groups of results from each experiment to form 80 test samples in total. The order of the different methods in each test sample is randomly shuffled. For each test sample, volunteers are asked to select the best option in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness, respectively. As shown in Table 3, our method outperforms the baseline methods for most cases, except comparable results to Barbershop [47] and SYH [20] in the hair transfer setting. But our irrelevant attribute preservation performs best because of our hair color feature space blending mechanism, as demonstrated in Figure 5. It is worth mentioning that our goal is not to improve the performance of hair transfer, but to design a unified system that supports various hair editing and hair transfer tasks. Therefore, performing comparably with the state-of-the-art methods on the hair transfer task is" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.292, + 0.578, + 0.306 + ], + "angle": 0, + "content": "acceptable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.318, + 0.673, + 0.334 + ], + "angle": 0, + "content": "4.2. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Importance of Initialization Strategy for Text Proxy Optimization. To justify our optimization starting point strategy for text proxy, we ablate three different optimization starting points: mean latent code, random sampling from \\(\\mathcal{W}\\) space, and inverted latent code of the input image within \\(\\mathcal{W}+\\) space. All other settings remain the same. As illustrated in Figure 8, the first two settings are more likely to complete high-quality editing, while the last one fails because the starting point deviates from the more suitable editing region. This may explain why StyleCLIP (\"Optimization\" version) and TediGAN struggle in performing hairstyle editing. Our strategy of randomly initializing the latent code around the mean within \\(\\mathcal{W}\\) space enjoys both good editability and diversity of the generated results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.764 + ], + "angle": 0, + "content": "Superiority of Sketch Proxy Generation Design. To generate sketch proxy, we ablate another two possible ways. The first way is to constrain the orientation field of the sketch proxy to be similar to that of user's sketch in the drawing region with the orientation loss used in MichiGAN[31]. The second way is to first obtain the image corresponding to the sketch through the conditional translation network (for simplicity, we choose sketch2hair here), and then constrain the sketch proxy to be similar to it in the drawing area by LPIPS [44] loss during the optimization process. The comparison is shown in Figure 9. Only our method and the LPIPS optimization-based version are feasible. However, our method requires only a single feed forward, which is more efficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Feature Blending vs. Latent Code Blending. To demonstrate the superiority of proxy feature blending, we compare it with the alternative scheme based on the linear combination of latent codes. In detail, we initialize an interpolation factor for the latent code of the global editing proxy, the local sketch proxy, and the input image, respectively. In the optimization process, we optimize these three interpolation factors so that the generated image corresponding to the interpolated latent code is similar to the proxies or input im" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23595" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.885, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.893, + 0.283 + ], + "angle": 0, + "content": "Figure 5. Visual comparison with HairCLIP [36], LOHO [24], Barbershop [47], SYH [20] and MichiGAN [31] on hair transfer. Only our method and SYH can accomplish unaligned hair transfer while keeping irrelevant attributes unmodified." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.294, + 0.891, + 0.37 + ], + "angle": 0, + "content": "
Text-DrivenHair TransferSketch-BasedCross-Modal
MetricsOurs[36][21][40][19]Ours[36][24][47][20][31]Ours[31][41]Ours[36]
Accuracy41.5%32.3%22.5%1.0%2.8%28.0%2.3%4.8%29.3%28.5%7.3%76.8%22.0%1.3%82.8%17.3%
Preservation81.0%5.3%3.3%0.3%10.3%32.8%2.8%8.3%15.3%26.3%14.8%62.0%33.3%4.8%94.0%6.0%
Naturalness46.8%25.5%22.0%1.8%4.0%26.5%9.5%2.5%22.3%35.0%4.3%60.5%38.0%1.5%65.3%34.8%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.374, + 0.893, + 0.418 + ], + "angle": 0, + "content": "Table 3. User study on text-driven image manipulation, hair transfer, sketch-based local hair editing and cross-modal hair editing methods. Accuracy denotes the manipulation accuracy for given conditional inputs, Preservation indicates the ability to preserve irrelevant regions and Naturalness denotes the visual realism of the manipulated image." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.439, + 0.46, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Figure 6. Comparison with MichiGAN [31] & SketchSalon [41] on sketch-based local hair editing. We provide sketches in the facial parsing map for better visualization." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.642, + 0.468, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.814, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison with HairCLIP on cross-modal conditional input. Our approach shows better editing effects & excellent preservation of irrelevant attributes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "age within the corresponding region. In Figure 10, we use" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.44, + 0.888, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.612, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Figure 8. Ablation of different starting point settings for text proxy. The text description is \"Bob Cut Hairstyle\"." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.656, + 0.882, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.799, + 0.892, + 0.827 + ], + "angle": 0, + "content": "Figure 9. Ablation on sketch proxy generation design. We provide sketches drawn in the facial parsing map for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "the hairstyle reference image as an example to generate the global proxy. It is obvious that our scheme accomplishes global editing and local editing while perfectly keeping the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23596" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.093, + 0.156, + 0.103 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.104, + 0.163, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.093, + 0.233, + 0.103 + ], + "angle": 0, + "content": "Hairstyle Ref" + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.104, + 0.236, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.093, + 0.303, + 0.103 + ], + "angle": 0, + "content": "Input Sketch" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.104, + 0.309, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.093, + 0.366, + 0.102 + ], + "angle": 0, + "content": "Feature" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.104, + 0.384, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.093, + 0.45, + 0.102 + ], + "angle": 0, + "content": "Latent Code" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.104, + 0.458, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.162, + 0.162, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.162, + 0.236, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.162, + 0.309, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.162, + 0.384, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.162, + 0.458, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.233, + 0.468, + 0.275 + ], + "angle": 0, + "content": "Figure 10. Ablation of feature blending vs. latent code blending. Feature blending can enable global editing, local editing, and irrelevant attributes preservation simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.468, + 0.331 + ], + "angle": 0, + "content": "irrelevant properties unmodified, while latent code blending does not perform either one well." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.346, + 0.325, + 0.361 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.469, + 0.538 + ], + "angle": 0, + "content": "In this paper, we propose a unified hair editing system HairCLIPv2, which presents the first attempt that supports both simple text/reference image interaction and fine-grained local interactions. It innovatively converts all hair editing tasks into hair transfer tasks, with the corresponding editing conditions converted into transfer proxies. It can not only achieve high-quality hair editing results, but also well preserve the irrelevant attributes from being modified. In the future, we will study how to use feed-forward networks to generate all the proxies. Also, it is worthy to generalize the proposed framework to support generic natural images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.553, + 0.257, + 0.57 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.775 + ], + "angle": 0, + "content": "This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386 and 62121002, the Fundamental Research Funds for the Central Universities under Grant WK5290000003, Key Research and Development program of Anhui Province under Grant 2022k07020008. This work was also partly supported by Shenzhen Key Laboratory of Media Security, and the Opening Fund of Key Laboratory of Cyberculture Content Cognition and Detection, Ministry of Culture and Tourism. This work was also partially supported by a GRF grant (Project No. CityU 11216122) from the Research Grants Council (RGC) of Hong Kong. Thank Yi Yin for her help in this work." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.79, + 0.174, + 0.805 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.816, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[3] Abdul Fatir Ansari, J. Scarlett, and Harold Soh. A characteristic function approach to deep implicit generative modeling. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7476-7484, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[4] Martín Arjovsky, Soumith Chintala, and L. Bottou. Wasserstein generative adversarial networks. In ICML, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.206, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[5] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[6] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.331, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[7] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.387, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[8] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2023), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[9] Ruili Feng, Jie Xiao, Kecheng Zheng, Deli Zhao, Jingren Zhou, Qibin Sun, and Zheng-Jun Zha. Principled knowledge extrapolation with GANs. In Proceedings of the 39th International Conference on Machine Learning, volume 162, pages 6447-6464. PMLR, 17-23 Jul 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[10] Ruili Feng, Deli Zhao, and Zheng-Jun Zha. Understanding noise injection in gans. In International Conference on Machine Learning, pages 3284-3293. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[12] Ishaan Gulrajani, F. Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C. Courville. Improved training of wasserstein gans. In NIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[13] Youngjoo Jo and Jongyoul Park. Sc-fegan: Face editing generative adversarial network with user's sketch and color. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1745-1753, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23597" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[19] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.471, + 0.431 + ], + "angle": 0, + "content": "[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[24] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1984-1993, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[25] Edgar Schonfeld, B. Schiele, and A. Khoreva. A u-net based discriminator for generative adversarial networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8204-8213, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[28] Xinhui Song, Chen Liu, Youyi Zheng, Zunlei Feng, Lincheng Li, Kun Zhou, and Xin Yu. Hairstyle editing via parametric controllable strokes. IEEE Transactions on Visualization and Computer Graphics, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[29] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "linguistic representations. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[30] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[31] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. ACM Transactions on Graphics (TOG), 39(4):95-1, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[32] Zhentao Tan, Dongdong Chen, Qi Chu, Menglei Chai, Jing Liao, Mingming He, Lu Yuan, Gang Hua, and Nenghai Yu. Efficient semantic image synthesis via class-adaptive normalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[33] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[34] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[35] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS 2022), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[36] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[37] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2style: Improve the efficiency and effectiveness of stylegan inversion. IEEE Transactions on Image Processing, 31:3267-3280, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[38] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[39] Yiqian Wu, Yong-Liang Yang, and Xiaogang Jin. Hairmapper: Removing hair from portraits using gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4227-4236, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[40] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[41] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: Deep sketch-based hair image synthesis. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23598" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[45] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copypaste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning (ICML 2023), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.388, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[46] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.445, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.501, + 0.469, + 0.555 + ], + "angle": 0, + "content": "[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In European Conference on Computer Vision, pages 651-667. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[49] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved stylegan embedding: Where are the good latents? arXiv preprint arXiv:2012.09036, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[50] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23599" + } + ] +] \ No newline at end of file diff --git a/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_origin.pdf b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2063092501df3c454f6dce15e4668a13e814c99d --- /dev/null +++ b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/ba9719c6-e989-4b1d-bc85-7328ef2439a5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac5d6e6583786d64422031aec44aefbbc9396130f58a931c3f59e1d5ebb489c4 +size 8737638 diff --git a/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/full.md b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9bf8cc04d62a50143a0688acb382c10e02aa741e --- /dev/null +++ b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/full.md @@ -0,0 +1,327 @@ +# HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending + +Tianyi Wei $^{1}$ , Dongdong Chen $^{2}$ , Wenbo Zhou $^{1,\dagger}$ , Jing Liao $^{3}$ , Weiming Zhang $^{1}$ , Gang Hua $^{4}$ , Nenghai Yu $^{1}$ + +1University of Science and Technology of China 2Microsoft Cloud AI + +$^{3}$ City University of Hong Kong $^{4}$ Xi'an Jiaotong University + +{bestwty@mail., welbeckz@, zhangwm@, yn@}ustc.edu.cn + +{cddlyf@, ganghua@}gmail.com, jingliao@cityu.edu.hk + +![](images/7f4238118915e86bb7bbef0d613e3720d9d7d70b5e7c7b3cdcb15f292e08720e.jpg) +Figure 1. HairCLIPv2 supports hairstyle and color editing individually or jointly with unprecedented user interaction mode support, including text, mask, sketch, reference image, etc. + +# Abstract + +Hair editing has made tremendous progress in recent years. Early hair editing methods use well-drawn sketches or masks to specify the editing conditions. Even though they can enable very fine-grained local control, such interaction modes are inefficient for the editing conditions that can be easily specified by language descriptions or reference images. Thanks to the recent breakthrough of cross-modal models (e.g., CLIP), HairCLIP is the first work that enables hair editing based on text descriptions or reference images. However, such text-driven and reference-driven interaction modes make HairCLIP unable to support fine-grained controls specified by sketch or mask. In this paper, we propose HairCLIPv2, aiming to support all the aforementioned interactions with one unified framework. Simultaneously, it improves upon HairCLIP with better irrelevant attributes (e.g., identity, background) preservation and unseen text descriptions support. The key idea is to convert all the hair editing tasks into hair transfer tasks, with editing conditions converted into different proxies accordingly. The editing effects are added upon the input image by blend + +ing the corresponding proxy features within the hairstyle or hair color feature spaces. Besides the unprecedented user interaction mode support, quantitative and qualitative experiments demonstrate the superiority of HairCLIPv2 in terms of editing effects, irrelevant attribute preservation and visual naturalness. Our code is available at https: //github.com/wty-ustc/HairCLIPv2. + +# 1. Introduction + +Hair editing as an interesting and challenging problem has attracted a lot of research attention from both academia and industry. Over the past few decades, tremendous progress [50, 41, 31, 36] has been made in this field, enabling high-fidelity hair editing based on various types of user interactions or controls. In earlier hair editing methods [50, 41], commonly supported user editing conditions are sketches and masks, which can enable fine-grained local controls. But in real scenarios, many hair editing conditions can be specified by simpler interactions, e.g., text descriptions (e.g., "bowl cut hairstyle") and reference images. + +Recently, cross-modal visual and language representation learning [29, 22, 43, 8, 35, 38, 45] has made remarkable breakthrough, which makes text-guided image manipu + +
HairCLIP [36]LOHO [24]Barbershop [47]HairNet [48]SYH [20]MichiGAN [31]SketchSalon [41]Ours
Aligned Hair Transfer✓✓✓✓✓✓✗✓
Unaligned Hair Transfer✓✗✗✓✓✗✗✓
Text✓✗✗✗✗✗✗✓
Mask✗✗✓✗✓✓✗✓
Sketch✗✗✗✗✗✓✓✓
Local Hairstyle Editing✗✗✗✗✗✓✓✓
Local Hair Color Editing✗✗✗✗✗✗✓✓
+ +Table 1. Comparisons between our approach and mainstream hair editing methods in terms of available interaction modes and functionality. Only our method supports all interaction modes and enables both global and local hair editing. + +lation possible. HairCLIP [36] presents the first attempt that supports hair editing via text description and reference image within one unified framework. Despite such text-driven and reference-driven interaction being more efficient and user-friendly, HairCLIP cannot support fine-grained controls like sketches and masks. Moreover, HairCLIP has two other limitations: 1) Since hair editing in HairCLIP is accomplished by pure latent code manipulation, it will inevitably alter other irrelevant attributes (e.g., identity, background) because fully decoupling different attributes in latent codes is difficult; 2) It struggles in yielding satisfactory results for text descriptions that differ significantly from training texts. + +In this paper, we take a step forward and propose HairCLIPv2, a unified hair editing system that unprecedentedly supports all the aforementioned interaction modes, including the natural text/reference-driven interaction and fine-grained local interaction. In Table 1, we list the interaction modes and editing functionality supported by existing hair editing methods. Moreover, with fundamentally different editing mechanism design, HairCLIPv2 makes great improvement upon HairCLIP with better irrelevant attributes preservation and unseen text description support. + +The key idea of HairCLIPv2 is converting all the hair editing tasks into hair transfer tasks, and the editing conditions are converted into different transfer proxies accordingly. Conceptually, it can be understood as "find proxy hair images that satisfy the editing conditions and transfer the corresponding attributes to the source image". Note that we use the StyleGAN latent code or feature corresponding to such proxies rather than use the proxy images explicitly. + +More specifically, we first transform the input source image into the bald proxy, which inpaints the hair-covered regions (e.g., background, ears) with reasonable semantic attributes. This can help avoid the editing artifacts caused by occlusion when blending the source image with condition proxies. For different editing proxies, we define their generation as different tasks performed in StyleGAN according to their characteristics. Depending on the users' editing preferences, hair editing effects are then enforced upon the input image by blending the corresponding proxy features within the hairstyle feature space or hair color feature space. This + +is different from HairCLIP that achieves the editing effect by manipulating the 1-d latent codes. Such feature blending based editing naturally supports global and local hair editing by controlling the blending area to cover the entire hair area or part of it. + +To show the superiority of HairCLIPv2, we conduct extensive comparisons. In addition to more complete user interaction modes support, HairCLIPv2 also shows obvious advantages in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness. Some interactive editing examples are provided in Figure 1. Our contributions can be summarized as below: + +- We present a fresh perspective for hair editing tasks and propose a novel hair editing paradigm that unifies various types of editing into the form of proxy hair transfers. We achieve all editing effects with the feature blending mechanism, which not only alleviates the editing pressure on each proxy but also enables excellent irrelevant attribute preservation. +- We dedicate design to the proxy generation for different conditions based on their own special properties, e.g., for the text proxy, the decoupled proxy design and optimization starting point selection strategy help us achieve better editing effects and arbitrary text support; for the sketch proxy, we achieve local hairstyle editing support for the first time within the StyleGAN-based framework by formalizing its generation as the image translation task and incorporating insights of semantic layering in StyleGAN. +- Our system pushes the interactions of hair editing to a new level, supporting arbitrary text, mask, reference image, sketch and their combinations, and enabling both global and local hair editing, which has never been realized before. + +# 2. Related Work + +Generative Adversarial Networks. Since being invented, GANs have made considerable progress in terms of training strategies [18, 34], loss functions [3, 4, 9], and network + +structures [25, 12, 10, 30, 32]. In the field of image synthesis, a series of works called StyleGAN [17, 18, 15, 16] represents the cutting edge of GANs. Given its promising semantically decoupled latent space [7, 26] and high-quality image synthesis abilities, the pre-trained StyleGAN has become the preferred choice for performing image editing. In this paper, we choose StyleGAN2 to develop our framework, which is consistent with other hair editing methods [36, 47, 20, 21, 40, 24] to be compared. + +Latent Space Embedding and Editing. As the bridge connecting the pre-trained StyleGAN and other downstream editing tasks, GAN inversion aims to yield the ideal embedding of the real image in the latent space. Based on the application purposes, we roughly classify the GAN inversion methods into two categories: methods [33, 46, 49] suitable for editing and methods [2, 23, 37] for better reconstruction. The former methods project the real image into the embedding subspace more suitable for editing at the expense of reconstruction. Among them, e4e [33] has become the most popular method for editing tasks [36, 21, 39] performed in the latent space. The latter approaches [2, 23, 37] aim to achieve the perfect reconstruction of the real image. However, limited by the representation capability of the latent space, all these methods cannot achieve the perfect reconstruction. To address this issue, Barbershop [47] proposes a novel inversion method, which additionally introduces a feature space $\mathcal{F}$ of StyleGAN combined with the latent space $S$ to form a new embedding space $\mathcal{FS}$ . Inspired by this, we decouple the editing task from the reconstruction task by blending editing proxy features in the feature space to achieve a unified hair editing system that supports a wide range of interactions. + +Hair Editing Using GANs. Existing hair editing methods can be roughly categorized as conditional GANs [41, 31, 13] based and pre-trained StyleGAN based [36, 21, 40, 24, 47, 20, 48, 28]. As a pioneering work of hair transfer, MichiGAN [31] accomplishes hairstyle transfer by extracting the orientation map of the reference image. Barbershop [47] performs hair transfer within their proposed $\mathcal{FS}$ embedding space. But these methods often struggle when large pose differences exist between source and target image. Recently, some improvements [20, 48] on Barbershop make pose unaligned hair transfers possible. Our framework is also compatible with pose unaligned hair transfers and additionally offers more interaction modes. SketchHairSalon [41] enables local editing of hairstyle and hair color by using colored sketches as the input to the conditional translation network. Unlike them, we show for the first time that the StyleGAN-based framework can also perform local hair editing with sketches as the condition. + +Benefiting from the development of cross-modal models [29, 22], text-guided hair editing [36, 21, 40] has become the new trend. StyleCLIP [21] and TediGAN [40] uti + +lize CLIP loss to perform hair editing in an optimized manner. However, since the embedding of real image deviates from the original suitable editing latent space, these methods will fail for some cases. HairCLIP [36] alleviates the problem by training a hair mapper on a large-scale dataset, but struggles in yielding good results for descriptions that differ significantly from the training text. Moreover, none of these methods can preserve the irrelevant attributes well. In this work, we present a new perspective to enable text-guided hair editing methods. By decoupling the editing task from the reconstruction, we can better preserve the irrelevant attributes while enabling high-quality hair manipulation via arbitrary text descriptions. More importantly, there is no prior work that supports so many interaction modes and functionality as we offer. + +# 3. Proposed Method + +# 3.1. Preliminaries + +StyleGAN [18] can synthesize photorealistic images with a progressive upsampling network consisting of 18 layers. Its $\mathcal{W} \subsetneq \mathbb{R}^{512}$ latent space exhibits good semantic decoupling properties [7, 26], thus enabling various editing tasks. To perform semantic editing while ensuring the reconstruction quality, some inversion methods [1, 23, 37] extend the original $\mathcal{W}$ space to $\mathcal{W}+$ space, which is defined as a cascade of 18 different latent codes $w_{i}$ from the $\mathcal{W}$ space, i.e., $\mathcal{W} + \subsetneq \mathbb{R}^{18\times 512}$ . + +$\mathcal{FS}$ Embedding Space is proposed by Barbershop [47], which is designed to increase the representational capability of embedding space for details and enable the spatial control of image features. The new $\{F_7,S\}$ latent code replaces the first 7 layers of $\mathcal{W}+$ latent code with the $32\times 32\times 512$ features $F_{7}$ of style-block 7 of StyleGAN $G$ , i.e., $F_{7}\in \mathcal{F}\subsetneq \mathbb{R}^{32\times 32\times 512},S = [w_8,\dots w_i,\dots w_{18}],w_i\in \mathcal{W}.$ + +CLIP [22] is a multi-modality model pretrained on web-scale image-text pairs. It can well measure the semantic similarity between given image and text. + +# 3.2. Overview + +Since the $\mathcal{FS}$ embedding space [47] is proposed, performing seamless feature blending in $\mathcal{F}$ -space has become the de facto standard for many hair transfer works [47, 20, 48], because it can encode the spatial information and preserve local details. On the other hand, many hair editing efforts [36, 21, 40, 39] choose to perform editing in the $\mathcal{W}+$ space, despite unsatisfactory reconstruction, because it can encode rich disentangled semantics [7, 26]. Considering that $\mathcal{F}$ -space is expressive and enables realistic feature integration results while $\mathcal{W}+$ space is editable, we therefore wonder "Can we enjoy the best features of both spaces to facilitate the hair editing task?" To achieve this goal, we formulate the hair editing tasks as the hair transfer tasks. + +![](images/2679e48691e3a051ba8f260fd05ac42320190d2c989b02e6125978b310d183b5.jpg) +Figure 2. Overview of HairCLIPv2: Example with hairstyle description text, sketch, mask and hair color RGB values as conditional inputs. Corresponding proxy images are just for better understanding. We complete the hair editing by converting different conditions into different proxies and achieve editing effects by blending them in StyleGAN feature spaces. + +Specifically, we convert all editing conditions (e.g., text, reference image, sketch) into different proxies in the $\mathcal{W}+$ space, and accomplish hair editing by seamless proxy feature blending in the feature spaces of StyleGAN. The proxy features of different conditions are obtained with tailored methods based on the condition characteristics. + +Following the design in HairCLIP [36], we edit hairstyle and hair color sequentially by blending editing proxy features in the early and later StyleGAN feature space respectively. In detail, as shown in Figure 2, we choose to perform proxy feature blending of hairstyle and hair color on feature spaces $\mathcal{F}^{style} \subsetneq \mathbb{R}^{32 \times 32 \times 512}$ and $\mathcal{F}^{color} \subsetneq \mathbb{R}^{256 \times 256 \times 128}$ , which correspond to the features of 7-th and 14-th style-block in StyleGAN respectively. And users can select various interactions (global or local, single or combined) to edit hairstyles and hair color individually or jointly. + +# 3.3. Converting Input Image to Bald Proxy + +Given a source image $I^{src}$ to be edited, we obtain its latent code $w^{src}$ in $\mathcal{W}+$ space and feature $F_{7}^{src}$ in $\mathcal{FS}$ space by II2S [49] and FS embedding algorithms [47], respectively. Balding proxy is then generated to inpaint the hair-covered region with reasonable semantic attributes (e.g., background, ears, face, etc.), which can avoid editing artifacts due to occlusion when blending the original image with different editing proxies. + +Balding Proxy. To remove the occlusion from the hair area of the source image, we bald it using HairMapper [39]: $w^{bald} = \mathbb{B}(w^{src})$ . $\mathbb{B}$ denotes HairMapper, which completes the balding editing operation in $\mathcal{W}+$ space on the latent code $w^{src}$ to yield the latent code $w^{bald}$ corresponding to the balded source image. Since editing in $\mathcal{W}+$ space inevitably gets other irrelevant attributes modified, we circumvent this issue by blending bald feature with source image feature in $\mathcal{F}^{style}$ space: + +$$ +F _ {7} ^ {b a l d} = G \left(w _ {1 - 7} ^ {b a l d}\right) \times M ^ {b a l d} + F _ {7} ^ {s r c} \times \left(1 - M ^ {b a l d}\right), \tag {1} +$$ + +where $G$ stands for StyleGAN, $G(w_{1-7}^{bold})$ represents the bald feature in $\mathcal{F}^{style}$ space. $M^{bald}$ is the binary mask indicating the hair and ear regions of the source image, which is obtained by the facial parsing network BiseNET [42] and downsampled to $32 \times 32$ . With the guidance of $M^{bald}$ , the irrelevant attributes region in $F_7^{bald}$ can continue to be preserved in following proxy feature blending. + +# 3.4. Hairstyle Editing + +Below we elaborate on how to generate the proxy for different hairstyle conditions. + +Text Proxy. We formalize text proxy generation as the editing task done in $\mathcal{W}+$ space based on the CLIP loss guidance. Unlike prior works [36, 21, 40], our text proxy generation process is free from the pressure of irrelevant attribute preservation, and allows us to select a more suitable starting point for the optimization process, which leads to better editing effects. In order to ensure both the optimal editing effect and the diversity of editing results, we choose to sample a random point around the mean face latent code as the optimization starting point for our text proxy latent code $w^{text}$ . In detail, we adopt the truncation trick of StyleGAN as $w^{init} = w^{mean} + \psi (w^{random} - w^{mean})$ , where $w^{mean}$ is the mean face latent code and $w^{random}$ is sampled randomly. By setting a small value of $\psi$ , we ensure that the initial optimization starting point $w^{init}$ is around the average face latent code $w^{mean}$ . We will show the benefits of this initialization strategy in the ablation analysis. We adopt the CLIP loss with transformation augmentations [5] to perform the text-guided hairstyle editing while reducing the disturbance caused by adversarial examples: + +$$ +L ^ {c l i p} = \frac {1}{N} \sum_ {i = 1} ^ {N} (1 - \cos \left(E _ {i} \left(A _ {i} \left(G \left(w ^ {t e x t}\right)\right)\right), E _ {t} (s t)\right)), \tag {2} +$$ + +where $A_{i}$ represents $i$ -th transformation augmentation, $N$ denotes the number of augmentations ( $N = 4$ by default), + +$\cos (\cdot)$ means cosine similarity, $G(w^{text})$ means the editing result for each pass in the optimization process, $E_{i}$ and $E_{t}$ stands for the CLIP image encoder and text encoder respectively, and $st$ refers to the user-supplied text description. Besides, pose alignment loss $L^{pose}$ is utilized to ensure that the face shape and pose of text proxy are consistent with the source image to ease the subsequent feature blending: + +$$ +L ^ {p o s e} = \frac {1}{N _ {k}} | | E _ {p} (I ^ {s r c}) - E _ {p} (G (w ^ {t e x t})) | | _ {2} ^ {2}, \qquad (3) +$$ + +where $E_{p}$ represents the 3D keypoint extractor [6] and $N_{k}$ denotes the number of keypoints. Optionally, the shape loss $L^{shape}$ is added to constrain the shape of the generated hair according to whether the user provides the hair region mask. We then obtain text proxy feature $F_{7}^{text}$ using the optimized $w^{text}: F_{7}^{text} = G(w_{1-7}^{text})$ . + +Reference Proxy. Given a hairstyle reference image $I^{sr}$ , we generate the reference proxy by performing the unaligned hairstyle transfer task. $I^{sr}$ is first inverted by the II2S [49] embedding algorithm into the $\mathcal{W}+$ space to get $w^{ref}$ , which is served as the starting point for hairstyle transfer. During the transfer process, we expect to keep the original hair structure of the reference image while ensuring its pose and facial shape to be consistent with the source image. Thus, $L^{pose}$ is imposed between $G(w^{ref})$ and $I^{src}$ . In addition, a style loss $L^{style}$ based on the gram matrix [11] is used to ensure that the hairstyle structure of $I^{sr}$ remains unchanged during the alignment process: + +$$ +\begin{array}{l} L ^ {s t y l e} = \frac {1}{4} \sum_ {i = 1} ^ {4} \left\| \mathcal {G} _ {i} \left(V G G _ {i} \left(I ^ {s r} \times M ^ {r h}\right)\right) \right. \\ \left. - \mathcal {G} _ {i} \left(V G G _ {i} \left(G \left(w ^ {r e f}\right) \times M ^ {g h}\right)\right) \right| | _ {2} ^ {2}, \tag {4} \\ \end{array} +$$ + +where $\mathcal{G}_i(\gamma_i)$ represents the gram matrix calculated on the $i$ -th layer features and a total of 4 layers of features are extracted, i.e., $\{relu1\_2, relu2\_2, relu3\_3, relu4\_3\}$ of VGG [27]. $M^{rh}$ and $M^{gh}$ are hair masks for the reference image and the generated image of each round during optimization process predicted by BiseNET [42]. For the same purpose, the $L_2$ norm of the manipulation magnitude in the latent space is utilized during optimization: + +$$ +L ^ {r e g} = \left\| w _ {t} ^ {r e f} - w _ {t - 1} ^ {r e f} \right\| _ {2} ^ {2}, \tag {5} +$$ + +where $w_{t}^{ref}$ and $w_{t-1}^{ref}$ represent the latent code of the current step and the previous step, respectively. Similar to the text proxy, $L^{shape}$ is optionally added to allow the user to customize the shape of the hair. We then obtain reference proxy feature $F_{7}^{ref}$ using the optimized $w^{ref}$ : $F_{7}^{ref} = G(w_{1-7}^{ref})$ . + +Sketch Proxy. Enabling sketch-based local hairstyle editing within our framework is nontrivial. It is hard to find suitable losses to constrain the local hairstyle structure to + +conform to the sketch given by the user. To circumvent this problem, we innovatively formalize the synthesis of sketch proxy as an image translation task based on StyleGAN. Utilizing the sketch-hair dataset created by SketchHairSalon [41], we train a sketch2hair inverter $T$ , which is based on E2Style [37] and aims to find the most appropriate latent code in $\mathcal{W}+$ space to accurately translate a given sketch to the corresponding hair structure. The training loss consists of regular pixel-level $L_{2}$ loss, feature-level LPIPS [44] loss and multi-layer face parsing loss [37] which is introduced to provide more local supervision. During the training process, we randomly remove a portion of the strokes to make our sketch2hair inverter adapt to a variety of sketch inputs from fine to coarse, e.g., even just one stroke. Given a local hairstyle sketch $S$ , our sketch proxy features $F_{7}^{sketch}$ are synthesized by pre-trained sketch2hair $T$ and StyleGAN $G$ : + +$$ +w ^ {\text {s k e t c h}} = T (S), \quad F _ {7} ^ {\text {s k e t c h}} = G \left(w _ {1 - 7} ^ {\text {s k e t c h}}\right). \tag {6} +$$ + +Proxy Feature Blending. For text and reference image based hairstyle condition, we perform global blending in $\mathcal{F}^{style}$ space: + +$$ +F _ {7} ^ {\text {g l o b a l}} = F _ {7} ^ {t r} \times M ^ {\text {g l o b a l}} + F _ {7} ^ {\text {b a l d}} \times \left(1 - M ^ {\text {g l o b a l}}\right), \tag {7} +$$ + +where $F_7^{tr} \in \{F_7^{text}, F_7^{ref}\}$ and $M^{global}$ is the binary mask corresponding to the hair region of $F_7^{tr}$ . Optionally, the sketch-based local hairstyle editing is applied: + +$$ +F _ {7} ^ {\text {s t y l e}} = F _ {7} ^ {\text {s k e t c h}} \times M ^ {\text {l o c a l}} + F _ {7} ^ {\text {g l o b a l}} \times \left(1 - M ^ {\text {l o c a l}}\right), \tag {8} +$$ + +where $M^{local}$ is obtained by downsampling the user input sketch $S$ to $32 \times 32$ after dilation. A natural concern is the artifacts brought by the mismatch between hair features within $M^{local}$ and other hair features. But thanks to the semantic layering characteristics of StyleGAN, the resulting image shows consistent tones as these hair features will be modulated by the later layers. Our framework allows users to only edit the hairstyle: $I^{style} = G(F_7^{style}, w_{8-18}^{src})$ , by skipping the following hair color editing. + +# 3.5. Hair Color Editing + +We achieve hair color editing by performing proxy feature blending in $\mathcal{F}^{color}$ space. By choosing to use the feature $F_{7}^{src}$ or $F_{7}^{style}$ , we allow to edit only hair color or both hair style&hair color. Below, we use $F_{7}^{style}$ as the example. Color Proxy. We initialize the color proxy with $F_{7}^{style}$ and $w_{8-18}^{color} = w_{8-18}^{src}$ , and set $w_{10-13}^{color}$ to be estimizable. The loss $L^{color}$ in the optimization process consists of $L^{modal}$ and $L^{bg}$ , where $L^{modal}$ can be defined as $L^{clip}$ or the average color $L_{2}$ loss of the hair region depending on the hair color condition types (text, reference image, RGB values), and $L^{bg}$ is defined as follows: + +$$ +L ^ {b g} = \left\| \left(I ^ {\text {s t y l e}} - I ^ {\text {c o l o r}}\right) \times \left(M ^ {n - h a i r}\right) \right\| _ {2} ^ {2}, \tag {9} +$$ + +
MethodsIDS↑PSNR↑SSIM↑
Ours0.8429.50.91
HairCLIP [36]0.4521.60.74
StyleCLIP [21]0.4319.60.72
TediGAN [40]0.1622.50.74
DiffCLIP [19]0.7126.80.86
+ +Table 2. Quantitative comparison for irrelevant attributes preservation. IDS denotes identity similarity, PSNR and SSIM are calculated at the intersected non-hair regions before and after editing. + +where $I^{color} = G(F_7^{style}, w_{8-18}^{color})$ , $M^{n-hair}$ is the mask of the non-hair region intersection between $I^{style}$ and $I^{color}$ . + +Proxy Feature Blending. Even with the $L^{bg}$ constraint, we find non-hair regions are often inevitably modified because of imperfect semantic decoupling of the $\mathcal{W}+$ space. We solve this by performing proxy feature blending in $\mathcal{F}^{color}$ space, which also naturally supports local hair color editing: + +$$ +\begin{array}{l} F _ {1 4} ^ {\text {b l e n d}} = G \left(F _ {7} ^ {\text {s t y l e}}, w _ {8 - 1 4} ^ {\text {c o l o r}}\right) \times M ^ {\text {c o l o r}} \\ + G \left(F _ {7} ^ {\text {s t y l e}}, w _ {8 - 1 4} ^ {\text {s r c}}\right) \times \left(1 - M ^ {\text {c o l o r}}\right), \tag {10} \\ \end{array} +$$ + +where $M^{color}$ is the hair area mask or a local editing area mask drawn by the user. We set $F_{14}^{blend}$ and $w_{15-18}^{color}$ to be estimizable to further perform the optimization. In the optimization process, we use the $L_{blend}$ loss, which consists of $L_{2}$ loss and LPIPS loss to constrain $I^{final}$ to be similar to $I^{color}$ inside $M^{color}$ and similar to $I^{style}$ outside $M^{color}$ simultaneously. The final edited image is synthesized as follows: $I^{final} = G(F_{14}^{blend}, w_{15-18}^{color})$ . + +# 4. Experiments + +Implementation details of our approach are provided in the supplementary material. For all compared methods, we use their official codes or pre-trained models. + +# 4.1. Quantitative and Qualitative Comparison + +Comparison with Text-Driven Hair Editing Methods. We compare HairCLIPv2 with leading text-driven hair editing methods on the CelebA-HQ [14] testset (2,000 images) and follow the evaluation settings of HairCLIP. For HairCLIP [36] and StyleCLIP [21] ("Mapper" version), we first invert using e4e [33] to obtain the latent code for a given real image before performing the editing. For Diffusion-CLIP [19], we finetune a model for each text description. For both TediGAN [40] and our method, the number of optimization iterations is set to 200. As shown in Figure 3 and Table 2, our method accomplishes satisfactory hair editing effects with better naturalness while maximizing the preservation of irrelevant attributes. It is worth noting that, even though HairCLIP and StyleCLIP also have pretty good hair + +editing capabilities, they cannot preserve the irrelevant attributes very well such as background, identity and clothes. Our method also demonstrates better preservation of the original hair structure when editing only the hair color. + +For arbitrary hair editing word scenarios, the only methods that are instantly feasible without retraining the model are HairCLIP, StyleCLIP ("Optimization" version), and TediGAN. As shown in Figure 4, our method performs much better at such cases. In contrast, HairCLIP can only produce plausible results for text ("Curly Short Hairstyle") similar to the training texts, while all other methods struggle to produce reasonable editing effects. + +Comparison with Hair Transfer Methods. We compare with state-of-the-art methods on hair transfer tasks. Among the 2,000 images of the CelebA-HQ testset, the first 666 are set as the input images, the middle 666 are set as the hairstyle reference images, and the last 666 are set as the hair color reference images. As shown in Figure 5, when the hairstyle reference image is broadly aligned with the input image (first row), most methods yield plausible results. However, when not aligned (second row), only our method and SYH [20] are able to perform a more consistent hair transfer, which is achieved by introducing the pose alignment loss during the transfer to ensure that the facial shape and pose of the reference image are consistent with the source image. Compared to SYH, we achieve comparable hair transfer results, but support text, sketch, and other interactions beyond hair transfer. + +Comparison with Local Hair Editing Methods. In terms of sketch-based local editing, we compare with the SOTA methods MichiGAN [31] and SketchSalon [41]. MichiGAN [31] uses user-drawn sketches to modify the orientation map to accomplish local hair editing. SketchSalon [41] trains a sketch-to-hair conditional translation network, with an additional soft alpha matte used to facilitate more natural blending. To generalize SketchSalon to local editing, we utilize the same mask as our method instead of a soft alpha matte, and the input sketch is colored as the average color within the mask area. As shown in Figure 6, MichiGAN struggles to perform satisfactory local editing and the reconstruction of other non-editing hair areas is slightly worse. Even ignoring the obvious blending artifacts, the local hair texture generated by SketchSalon is not in harmony with the surrounding hair. Compared to these two methods, our approach not only achieves satisfactory local editing but also better maintains the non-editing regions. + +Comparison with Cross-Modal Hair Editing Methods. To the best of our knowledge, the only method that supports multimodal conditions to complete hairstyle and hair color editing is HairCLIP [36]. As the comparison shown in Figure 7, our approach not only perfectly prevents irrelevant attributes (identity, background, etc.) from being modified, but also achieves higher-quality editing effects. Moreover, + +![](images/34eb8bcd0cef34dbd678faaf46f38e040c8132ea49727736eb75122c84c88757.jpg) +Figure 3. Visual comparison with HairCLIP [36], StyleCLIP-Mapper [21], TediGAN [40] and DiffusionCLIP [19]. The simplified text descriptions (editing hairstyle, hair color, or both of them) are listed on the leftmost side. Our approach demonstrates better editing effects and irrelevant attribute preservation (e.g., identity, background, etc.). + +![](images/4649bda1ca7f272b1d9613d58d623d581a1c77638d4853b0893c664aef84bbc0.jpg) +Figure 4. Visual comparison with HairCLIP [36], StyleCLIP-Optimization [21] and TediGAN [40] under any description setting. We additionally provide an example image for each description for better comparison. + +HairCLIP allows only text and reference image while our method additionally supports sketch, mask, and RGB values. More diverse and comprehensive interactive editing results are shown in Figure 1 and supplementary materials. + +User Study. For the above four types of comparisons, we recruit 20 volunteers with computer vision-related research backgrounds to execute a comprehensive user study. We randomly selected 20 groups of results from each experiment to form 80 test samples in total. The order of the different methods in each test sample is randomly shuffled. For each test sample, volunteers are asked to select the best option in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness, respectively. As shown in Table 3, our method outperforms the baseline methods for most cases, except comparable results to Barbershop [47] and SYH [20] in the hair transfer setting. But our irrelevant attribute preservation performs best because of our hair color feature space blending mechanism, as demonstrated in Figure 5. It is worth mentioning that our goal is not to improve the performance of hair transfer, but to design a unified system that supports various hair editing and hair transfer tasks. Therefore, performing comparably with the state-of-the-art methods on the hair transfer task is + +acceptable. + +# 4.2. Ablation Analysis + +Importance of Initialization Strategy for Text Proxy Optimization. To justify our optimization starting point strategy for text proxy, we ablate three different optimization starting points: mean latent code, random sampling from $\mathcal{W}$ space, and inverted latent code of the input image within $\mathcal{W}+$ space. All other settings remain the same. As illustrated in Figure 8, the first two settings are more likely to complete high-quality editing, while the last one fails because the starting point deviates from the more suitable editing region. This may explain why StyleCLIP ("Optimization" version) and TediGAN struggle in performing hairstyle editing. Our strategy of randomly initializing the latent code around the mean within $\mathcal{W}$ space enjoys both good editability and diversity of the generated results. + +Superiority of Sketch Proxy Generation Design. To generate sketch proxy, we ablate another two possible ways. The first way is to constrain the orientation field of the sketch proxy to be similar to that of user's sketch in the drawing region with the orientation loss used in MichiGAN[31]. The second way is to first obtain the image corresponding to the sketch through the conditional translation network (for simplicity, we choose sketch2hair here), and then constrain the sketch proxy to be similar to it in the drawing area by LPIPS [44] loss during the optimization process. The comparison is shown in Figure 9. Only our method and the LPIPS optimization-based version are feasible. However, our method requires only a single feed forward, which is more efficient. + +Feature Blending vs. Latent Code Blending. To demonstrate the superiority of proxy feature blending, we compare it with the alternative scheme based on the linear combination of latent codes. In detail, we initialize an interpolation factor for the latent code of the global editing proxy, the local sketch proxy, and the input image, respectively. In the optimization process, we optimize these three interpolation factors so that the generated image corresponding to the interpolated latent code is similar to the proxies or input im + +![](images/ebcb086fce6bb20533518fe772720a0fda213e8ec01ffc13f3e897e4daa454ea.jpg) +Figure 5. Visual comparison with HairCLIP [36], LOHO [24], Barbershop [47], SYH [20] and MichiGAN [31] on hair transfer. Only our method and SYH can accomplish unaligned hair transfer while keeping irrelevant attributes unmodified. + +
Text-DrivenHair TransferSketch-BasedCross-Modal
MetricsOurs[36][21][40][19]Ours[36][24][47][20][31]Ours[31][41]Ours[36]
Accuracy41.5%32.3%22.5%1.0%2.8%28.0%2.3%4.8%29.3%28.5%7.3%76.8%22.0%1.3%82.8%17.3%
Preservation81.0%5.3%3.3%0.3%10.3%32.8%2.8%8.3%15.3%26.3%14.8%62.0%33.3%4.8%94.0%6.0%
Naturalness46.8%25.5%22.0%1.8%4.0%26.5%9.5%2.5%22.3%35.0%4.3%60.5%38.0%1.5%65.3%34.8%
+ +Table 3. User study on text-driven image manipulation, hair transfer, sketch-based local hair editing and cross-modal hair editing methods. Accuracy denotes the manipulation accuracy for given conditional inputs, Preservation indicates the ability to preserve irrelevant regions and Naturalness denotes the visual realism of the manipulated image. + +![](images/0aba2e1a50f12be67d4c3708753dc3d84c8896f2a3319c92387b6f9d0566fb79.jpg) +Figure 6. Comparison with MichiGAN [31] & SketchSalon [41] on sketch-based local hair editing. We provide sketches in the facial parsing map for better visualization. + +![](images/1c3548c36cf70617e6b98d144c2fc49c30e37749f0f87293fdf5269e6032b64f.jpg) +Figure 7. Qualitative comparison with HairCLIP on cross-modal conditional input. Our approach shows better editing effects & excellent preservation of irrelevant attributes. + +age within the corresponding region. In Figure 10, we use + +![](images/b60d871addebd77c7ef3ad985820fb651897a5ad09bbfddeb57f95d3d4248c0f.jpg) +Figure 8. Ablation of different starting point settings for text proxy. The text description is "Bob Cut Hairstyle". + +![](images/03a13bbab0f88b877b75793d23ed4e23c13cece812b23ec2dbd5ad48437a05aa.jpg) +Figure 9. Ablation on sketch proxy generation design. We provide sketches drawn in the facial parsing map for better visualization. + +the hairstyle reference image as an example to generate the global proxy. It is obvious that our scheme accomplishes global editing and local editing while perfectly keeping the + +![](images/b7c89a31de86372f38bd31f33d5ab692d5e7ba4b703e0a716ffbec25c1bdd2f5.jpg) +Input Image + +![](images/f211568a8e6fe6a605adee23e81e7cbebc256439e14be4519cfe6e7154efc9e2.jpg) +Hairstyle Ref + +![](images/667c28a35a9aa5cbf65e7c3971f661e07569f0213d8a5f1f36561823ca9bdbd1.jpg) +Input Sketch + +![](images/66dd24697518234b8a20d4d9f88537628df65b9e36242407a38284f4a80f17b6.jpg) +Feature + +![](images/0d705271d6e9012ee91d19621f2f910db456d524985e4f7a7236329ee8c0ae1a.jpg) +Latent Code + +![](images/cc73e7b29a35d6a01fa2a7e98aa511868a5946d9a70701c357eafafd967b3673.jpg) +Figure 10. Ablation of feature blending vs. latent code blending. Feature blending can enable global editing, local editing, and irrelevant attributes preservation simultaneously. + +![](images/356ac340bdc8f91a9f122bd0f08bc59e417fc9459ce502724eb941e13b453e15.jpg) + +![](images/ddfff9c5674db2e34c7c783f725aa50459df5329a1616fd00de43e4cf27fc828.jpg) + +![](images/6ea7127313a5e0c06f038295aa076571c4193d52391eb20d8c7715fcc6345e7a.jpg) + +![](images/a62eadfcb28dde362307aa7d6febc722b0260d64df006ae65847c62abfaec356.jpg) + +irrelevant properties unmodified, while latent code blending does not perform either one well. + +# 5. Conclusion and Discussion + +In this paper, we propose a unified hair editing system HairCLIPv2, which presents the first attempt that supports both simple text/reference image interaction and fine-grained local interactions. It innovatively converts all hair editing tasks into hair transfer tasks, with the corresponding editing conditions converted into transfer proxies. It can not only achieve high-quality hair editing results, but also well preserve the irrelevant attributes from being modified. In the future, we will study how to use feed-forward networks to generate all the proxies. Also, it is worthy to generalize the proposed framework to support generic natural images. + +# 6. Acknowledgement + +This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386 and 62121002, the Fundamental Research Funds for the Central Universities under Grant WK5290000003, Key Research and Development program of Anhui Province under Grant 2022k07020008. This work was also partly supported by Shenzhen Key Laboratory of Media Security, and the Opening Fund of Key Laboratory of Cyberculture Content Cognition and Detection, Ministry of Culture and Tourism. This work was also partially supported by a GRF grant (Project No. CityU 11216122) from the Research Grants Council (RGC) of Hong Kong. Thank Yi Yin for her help in this work. + +# References + +[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In + +Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. +[3] Abdul Fatir Ansari, J. Scarlett, and Harold Soh. A characteristic function approach to deep implicit generative modeling. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7476-7484, 2020. +[4] Martín Arjovsky, Soumith Chintala, and L. Bottou. Wasserstein generative adversarial networks. In ICML, 2017. +[5] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. +[6] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017. +[7] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020. +[8] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2023), 2023. +[9] Ruili Feng, Jie Xiao, Kecheng Zheng, Deli Zhao, Jingren Zhou, Qibin Sun, and Zheng-Jun Zha. Principled knowledge extrapolation with GANs. In Proceedings of the 39th International Conference on Machine Learning, volume 162, pages 6447-6464. PMLR, 17-23 Jul 2022. +[10] Ruili Feng, Deli Zhao, and Zheng-Jun Zha. Understanding noise injection in gans. In International Conference on Machine Learning, pages 3284-3293. PMLR, 2021. +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. +[12] Ishaan Gulrajani, F. Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C. Courville. Improved training of wasserstein gans. In NIPS, 2017. +[13] Youngjoo Jo and Jongyoul Park. Sc-fegan: Face editing generative adversarial network with user's sketch and color. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1745-1753, 2019. +[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. +[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. +[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. + +[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. +[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. +[19] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. +[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022. +[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. +[24] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1984-1993, 2021. +[25] Edgar Schonfeld, B. Schiele, and A. Khoreva. A u-net based discriminator for generative adversarial networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8204-8213, 2020. +[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. +[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[28] Xinhui Song, Chen Liu, Youyi Zheng, Zunlei Feng, Lincheng Li, Kun Zhou, and Xin Yu. Hairstyle editing via parametric controllable strokes. IEEE Transactions on Visualization and Computer Graphics, 2023. +[29] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual + +linguistic representations. In International Conference on Learning Representations, 2019. +[30] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. +[31] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. ACM Transactions on Graphics (TOG), 39(4):95-1, 2020. +[32] Zhentao Tan, Dongdong Chen, Qi Chu, Menglei Chai, Jing Liao, Mingming He, Lu Yuan, Gang Hua, and Nenghai Yu. Efficient semantic image synthesis via class-adaptive normalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. +[33] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. +[34] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. +[35] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS 2022), 2022. +[36] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022. +[37] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2style: Improve the efficiency and effectiveness of stylegan inversion. IEEE Transactions on Image Processing, 31:3267-3280, 2022. +[38] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023. +[39] Yiqian Wu, Yong-Liang Yang, and Xiaogang Jin. Hairmapper: Removing hair from portraits using gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4227-4236, 2022. +[40] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[41] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: Deep sketch-based hair image synthesis. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021. + +[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018. +[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. +[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. +[45] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copypaste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning (ICML 2023), 2023. +[46] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. +[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021. +[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In European Conference on Computer Vision, pages 651-667. Springer, 2022. +[49] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved stylegan embedding: Where are the good latents? arXiv preprint arXiv:2012.09036, 2020. +[50] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020. \ No newline at end of file diff --git a/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/images.zip b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5866eb37a7719e05fb14c1e7021987a5c949d15e --- /dev/null +++ b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:780bdfe4d5eeef5eba1495b05b45a3ac6da3587cf9af9d790570ea31bc5741f1 +size 759364 diff --git a/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/layout.json b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8693b69f90186edf21068d85fde5c5d23924e32c --- /dev/null +++ b/2023/HairCLIPv2_ Unifying Hair Editing via Proxy Feature Blending/layout.json @@ -0,0 +1,9998 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 496, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 496, + 122 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 496, + 122 + ], + "type": "text", + "content": "HairCLIPv2: Unifying Hair Editing via Proxy Feature Blending" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "content": "Tianyi Wei" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Dongdong Chen" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Wenbo Zhou" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Jing Liao" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Weiming Zhang" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Gang Hua" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Nenghai Yu" + }, + { + "bbox": [ + 152, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 172, + 438, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 172, + 438, + 184 + ], + "spans": [ + { + "bbox": [ + 153, + 172, + 438, + 184 + ], + "type": "text", + "content": "1University of Science and Technology of China 2Microsoft Cloud AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 186, + 415, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 186, + 415, + 199 + ], + "spans": [ + { + "bbox": [ + 175, + 186, + 415, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 175, + 186, + 415, + 199 + ], + "type": "text", + "content": "City University of Hong Kong " + }, + { + "bbox": [ + 175, + 186, + 415, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 175, + 186, + 415, + 199 + ], + "type": "text", + "content": "Xi'an Jiaotong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 201, + 436, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 201, + 436, + 213 + ], + "spans": [ + { + "bbox": [ + 151, + 201, + 436, + 213 + ], + "type": "text", + "content": "{bestwty@mail., welbeckz@, zhangwm@, yn@}ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 162, + 216, + 432, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 216, + 432, + 227 + ], + "spans": [ + { + "bbox": [ + 162, + 216, + 432, + 227 + ], + "type": "text", + "content": "{cddlyf@, ganghua@}gmail.com, jingliao@cityu.edu.hk" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 55, + 256, + 542, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 256, + 542, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 542, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 542, + 376 + ], + "type": "image", + "image_path": "7f4238118915e86bb7bbef0d613e3720d9d7d70b5e7c7b3cdcb15f292e08720e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 378, + 547, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 547, + 401 + ], + "type": "text", + "content": "Figure 1. HairCLIPv2 supports hairstyle and color editing individually or jointly with unprecedented user interaction mode support, including text, mask, sketch, reference image, etc." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 425, + 192, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 425, + 192, + 437 + ], + "spans": [ + { + "bbox": [ + 143, + 425, + 192, + 437 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 451, + 289, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 289, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 289, + 691 + ], + "type": "text", + "content": "Hair editing has made tremendous progress in recent years. Early hair editing methods use well-drawn sketches or masks to specify the editing conditions. Even though they can enable very fine-grained local control, such interaction modes are inefficient for the editing conditions that can be easily specified by language descriptions or reference images. Thanks to the recent breakthrough of cross-modal models (e.g., CLIP), HairCLIP is the first work that enables hair editing based on text descriptions or reference images. However, such text-driven and reference-driven interaction modes make HairCLIP unable to support fine-grained controls specified by sketch or mask. In this paper, we propose HairCLIPv2, aiming to support all the aforementioned interactions with one unified framework. Simultaneously, it improves upon HairCLIP with better irrelevant attributes (e.g., identity, background) preservation and unseen text descriptions support. The key idea is to convert all the hair editing tasks into hair transfer tasks, with editing conditions converted into different proxies accordingly. The editing effects are added upon the input image by blend" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 415, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 547, + 498 + ], + "type": "text", + "content": "ing the corresponding proxy features within the hairstyle or hair color feature spaces. Besides the unprecedented user interaction mode support, quantitative and qualitative experiments demonstrate the superiority of HairCLIPv2 in terms of editing effects, irrelevant attribute preservation and visual naturalness. Our code is available at https: //github.com/wty-ustc/HairCLIPv2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 524, + 387, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 387, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 387, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 676 + ], + "type": "text", + "content": "Hair editing as an interesting and challenging problem has attracted a lot of research attention from both academia and industry. Over the past few decades, tremendous progress [50, 41, 31, 36] has been made in this field, enabling high-fidelity hair editing based on various types of user interactions or controls. In earlier hair editing methods [50, 41], commonly supported user editing conditions are sketches and masks, which can enable fine-grained local controls. But in real scenarios, many hair editing conditions can be specified by simpler interactions, e.g., text descriptions (e.g., \"bowl cut hairstyle\") and reference images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Recently, cross-modal visual and language representation learning [29, 22, 43, 8, 35, 38, 45] has made remarkable breakthrough, which makes text-guided image manipu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 203, + 713 + ], + "type": "text", + "content": "† Wenbo Zhou is the corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23589" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 155 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 155 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 155 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 155 + ], + "type": "table", + "html": "
HairCLIP [36]LOHO [24]Barbershop [47]HairNet [48]SYH [20]MichiGAN [31]SketchSalon [41]Ours
Aligned Hair Transfer✓✓✓✓✓✓✗✓
Unaligned Hair Transfer✓✗✗✓✓✗✗✓
Text✓✗✗✗✗✗✗✓
Mask✗✗✓✗✓✓✗✓
Sketch✗✗✗✗✗✓✓✓
Local Hairstyle Editing✗✗✗✗✗✓✓✓
Local Hair Color Editing✗✗✗✗✗✗✓✓
", + "image_path": "860f16c8a83bae64da252f61191a4c4cd5680ba7df0f1cb9a303b9c82d985310.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 546, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 546, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 546, + 182 + ], + "type": "text", + "content": "Table 1. Comparisons between our approach and mainstream hair editing methods in terms of available interaction modes and functionality. Only our method supports all interaction modes and enables both global and local hair editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 201, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 289, + 357 + ], + "type": "text", + "content": "lation possible. HairCLIP [36] presents the first attempt that supports hair editing via text description and reference image within one unified framework. Despite such text-driven and reference-driven interaction being more efficient and user-friendly, HairCLIP cannot support fine-grained controls like sketches and masks. Moreover, HairCLIP has two other limitations: 1) Since hair editing in HairCLIP is accomplished by pure latent code manipulation, it will inevitably alter other irrelevant attributes (e.g., identity, background) because fully decoupling different attributes in latent codes is difficult; 2) It struggles in yielding satisfactory results for text descriptions that differ significantly from training texts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 479 + ], + "type": "text", + "content": "In this paper, we take a step forward and propose HairCLIPv2, a unified hair editing system that unprecedentedly supports all the aforementioned interaction modes, including the natural text/reference-driven interaction and fine-grained local interaction. In Table 1, we list the interaction modes and editing functionality supported by existing hair editing methods. Moreover, with fundamentally different editing mechanism design, HairCLIPv2 makes great improvement upon HairCLIP with better irrelevant attributes preservation and unseen text description support." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 483, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 579 + ], + "type": "text", + "content": "The key idea of HairCLIPv2 is converting all the hair editing tasks into hair transfer tasks, and the editing conditions are converted into different transfer proxies accordingly. Conceptually, it can be understood as \"find proxy hair images that satisfy the editing conditions and transfer the corresponding attributes to the source image\". Note that we use the StyleGAN latent code or feature corresponding to such proxies rather than use the proxy images explicitly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "More specifically, we first transform the input source image into the bald proxy, which inpaints the hair-covered regions (e.g., background, ears) with reasonable semantic attributes. This can help avoid the editing artifacts caused by occlusion when blending the source image with condition proxies. For different editing proxies, we define their generation as different tasks performed in StyleGAN according to their characteristics. Depending on the users' editing preferences, hair editing effects are then enforced upon the input image by blending the corresponding proxy features within the hairstyle feature space or hair color feature space. This" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 201, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 262 + ], + "type": "text", + "content": "is different from HairCLIP that achieves the editing effect by manipulating the 1-d latent codes. Such feature blending based editing naturally supports global and local hair editing by controlling the blending area to cover the entire hair area or part of it." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 262, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 346 + ], + "type": "text", + "content": "To show the superiority of HairCLIPv2, we conduct extensive comparisons. In addition to more complete user interaction modes support, HairCLIPv2 also shows obvious advantages in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness. Some interactive editing examples are provided in Figure 1. Our contributions can be summarized as below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 356, + 547, + 642 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 356, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 356, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 356, + 545, + 441 + ], + "type": "text", + "content": "- We present a fresh perspective for hair editing tasks and propose a novel hair editing paradigm that unifies various types of editing into the form of proxy hair transfers. We achieve all editing effects with the feature blending mechanism, which not only alleviates the editing pressure on each proxy but also enables excellent irrelevant attribute preservation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 452, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 452, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 317, + 452, + 545, + 571 + ], + "type": "text", + "content": "- We dedicate design to the proxy generation for different conditions based on their own special properties, e.g., for the text proxy, the decoupled proxy design and optimization starting point selection strategy help us achieve better editing effects and arbitrary text support; for the sketch proxy, we achieve local hairstyle editing support for the first time within the StyleGAN-based framework by formalizing its generation as the image translation task and incorporating insights of semantic layering in StyleGAN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 583, + 547, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 583, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 317, + 583, + 547, + 642 + ], + "type": "text", + "content": "- Our system pushes the interactions of hair editing to a new level, supporting arbitrary text, mask, reference image, sketch and their combinations, and enabling both global and local hair editing, which has never been realized before." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 656, + 392, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 392, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 392, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Generative Adversarial Networks. Since being invented, GANs have made considerable progress in terms of training strategies [18, 34], loss functions [3, 4, 9], and network" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23590" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "structures [25, 12, 10, 30, 32]. In the field of image synthesis, a series of works called StyleGAN [17, 18, 15, 16] represents the cutting edge of GANs. Given its promising semantically decoupled latent space [7, 26] and high-quality image synthesis abilities, the pre-trained StyleGAN has become the preferred choice for performing image editing. In this paper, we choose StyleGAN2 to develop our framework, which is consistent with other hair editing methods [36, 47, 20, 21, 40, 24] to be compared." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "text", + "content": "Latent Space Embedding and Editing. As the bridge connecting the pre-trained StyleGAN and other downstream editing tasks, GAN inversion aims to yield the ideal embedding of the real image in the latent space. Based on the application purposes, we roughly classify the GAN inversion methods into two categories: methods [33, 46, 49] suitable for editing and methods [2, 23, 37] for better reconstruction. The former methods project the real image into the embedding subspace more suitable for editing at the expense of reconstruction. Among them, e4e [33] has become the most popular method for editing tasks [36, 21, 39] performed in the latent space. The latter approaches [2, 23, 37] aim to achieve the perfect reconstruction of the real image. However, limited by the representation capability of the latent space, all these methods cannot achieve the perfect reconstruction. To address this issue, Barbershop [47] proposes a novel inversion method, which additionally introduces a feature space " + }, + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "text", + "content": " of StyleGAN combined with the latent space " + }, + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "text", + "content": " to form a new embedding space " + }, + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{FS}" + }, + { + "bbox": [ + 46, + 186, + 289, + 461 + ], + "type": "text", + "content": ". Inspired by this, we decouple the editing task from the reconstruction task by blending editing proxy features in the feature space to achieve a unified hair editing system that supports a wide range of interactions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "type": "text", + "content": "Hair Editing Using GANs. Existing hair editing methods can be roughly categorized as conditional GANs [41, 31, 13] based and pre-trained StyleGAN based [36, 21, 40, 24, 47, 20, 48, 28]. As a pioneering work of hair transfer, MichiGAN [31] accomplishes hairstyle transfer by extracting the orientation map of the reference image. Barbershop [47] performs hair transfer within their proposed " + }, + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{FS}" + }, + { + "bbox": [ + 46, + 462, + 289, + 677 + ], + "type": "text", + "content": " embedding space. But these methods often struggle when large pose differences exist between source and target image. Recently, some improvements [20, 48] on Barbershop make pose unaligned hair transfers possible. Our framework is also compatible with pose unaligned hair transfers and additionally offers more interaction modes. SketchHairSalon [41] enables local editing of hairstyle and hair color by using colored sketches as the input to the conditional translation network. Unlike them, we show for the first time that the StyleGAN-based framework can also perform local hair editing with sketches as the condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Benefiting from the development of cross-modal models [29, 22], text-guided hair editing [36, 21, 40] has become the new trend. StyleCLIP [21] and TediGAN [40] uti" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "lize CLIP loss to perform hair editing in an optimized manner. However, since the embedding of real image deviates from the original suitable editing latent space, these methods will fail for some cases. HairCLIP [36] alleviates the problem by training a hair mapper on a large-scale dataset, but struggles in yielding good results for descriptions that differ significantly from the training text. Moreover, none of these methods can preserve the irrelevant attributes well. In this work, we present a new perspective to enable text-guided hair editing methods. By decoupling the editing task from the reconstruction, we can better preserve the irrelevant attributes while enabling high-quality hair manipulation via arbitrary text descriptions. More importantly, there is no prior work that supports so many interaction modes and functionality as we offer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 263, + 411, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 411, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 411, + 277 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 283, + 392, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 392, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 392, + 295 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": "StyleGAN [18] can synthesize photorealistic images with a progressive upsampling network consisting of 18 layers. Its " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\subsetneq \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " latent space exhibits good semantic decoupling properties [7, 26], thus enabling various editing tasks. To perform semantic editing while ensuring the reconstruction quality, some inversion methods [1, 23, 37] extend the original " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " space to " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " space, which is defined as a cascade of 18 different latent codes " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": " space, i.e., " + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{W} + \\subsetneq \\mathbb{R}^{18\\times 512}" + }, + { + "bbox": [ + 304, + 302, + 545, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{FS}" + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "text", + "content": " Embedding Space is proposed by Barbershop [47], which is designed to increase the representational capability of embedding space for details and enable the spatial control of image features. The new " + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\{F_7,S\\}" + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "text", + "content": " latent code replaces the first 7 layers of " + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "text", + "content": " latent code with the " + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "inline_equation", + "content": "32\\times 32\\times 512" + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "text", + "content": " features " + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "inline_equation", + "content": "F_{7}" + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "text", + "content": " of style-block 7 of StyleGAN " + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 411, + 545, + 495 + ], + "type": "inline_equation", + "content": "F_{7}\\in \\mathcal{F}\\subsetneq \\mathbb{R}^{32\\times 32\\times 512},S = [w_8,\\dots w_i,\\dots w_{18}],w_i\\in \\mathcal{W}." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 531 + ], + "type": "text", + "content": "CLIP [22] is a multi-modality model pretrained on web-scale image-text pairs. It can well measure the semantic similarity between given image and text." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 539, + 373, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 373, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 373, + 551 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "Since the " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{FS}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " embedding space [47] is proposed, performing seamless feature blending in " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "-space has become the de facto standard for many hair transfer works [47, 20, 48], because it can encode the spatial information and preserve local details. On the other hand, many hair editing efforts [36, 21, 40, 39] choose to perform editing in the " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " space, despite unsatisfactory reconstruction, because it can encode rich disentangled semantics [7, 26]. Considering that " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "-space is expressive and enables realistic feature integration results while " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " space is editable, we therefore wonder \"Can we enjoy the best features of both spaces to facilitate the hair editing task?\" To achieve this goal, we formulate the hair editing tasks as the hair transfer tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23591" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 545, + 213 + ], + "type": "image", + "image_path": "2679e48691e3a051ba8f260fd05ac42320190d2c989b02e6125978b310d183b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 546, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 546, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 546, + 250 + ], + "type": "text", + "content": "Figure 2. Overview of HairCLIPv2: Example with hairstyle description text, sketch, mask and hair color RGB values as conditional inputs. Corresponding proxy images are just for better understanding. We complete the hair editing by converting different conditions into different proxies and achieve editing effects by blending them in StyleGAN feature spaces." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "content": "Specifically, we convert all editing conditions (e.g., text, reference image, sketch) into different proxies in the " + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "content": " space, and accomplish hair editing by seamless proxy feature blending in the feature spaces of StyleGAN. The proxy features of different conditions are obtained with tailored methods based on the condition characteristics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 340, + 288, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 461 + ], + "type": "text", + "content": "Following the design in HairCLIP [36], we edit hairstyle and hair color sequentially by blending editing proxy features in the early and later StyleGAN feature space respectively. In detail, as shown in Figure 2, we choose to perform proxy feature blending of hairstyle and hair color on feature spaces " + }, + { + "bbox": [ + 46, + 340, + 288, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{style} \\subsetneq \\mathbb{R}^{32 \\times 32 \\times 512}" + }, + { + "bbox": [ + 46, + 340, + 288, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 340, + 288, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{color} \\subsetneq \\mathbb{R}^{256 \\times 256 \\times 128}" + }, + { + "bbox": [ + 46, + 340, + 288, + 461 + ], + "type": "text", + "content": ", which correspond to the features of 7-th and 14-th style-block in StyleGAN respectively. And users can select various interactions (global or local, single or combined) to edit hairstyles and hair color individually or jointly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 468, + 250, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 250, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 250, + 481 + ], + "type": "text", + "content": "3.3. Converting Input Image to Bald Proxy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "text", + "content": "Given a source image " + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "inline_equation", + "content": "I^{src}" + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "text", + "content": " to be edited, we obtain its latent code " + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "inline_equation", + "content": "w^{src}" + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "text", + "content": " space and feature " + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "inline_equation", + "content": "F_{7}^{src}" + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{FS}" + }, + { + "bbox": [ + 46, + 487, + 287, + 582 + ], + "type": "text", + "content": " space by II2S [49] and FS embedding algorithms [47], respectively. Balding proxy is then generated to inpaint the hair-covered region with reasonable semantic attributes (e.g., background, ears, face, etc.), which can avoid editing artifacts due to occlusion when blending the original image with different editing proxies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": "Balding Proxy. To remove the occlusion from the hair area of the source image, we bald it using HairMapper [39]: " + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "inline_equation", + "content": "w^{bald} = \\mathbb{B}(w^{src})" + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": " denotes HairMapper, which completes the balding editing operation in " + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": " space on the latent code " + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "inline_equation", + "content": "w^{src}" + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": " to yield the latent code " + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "inline_equation", + "content": "w^{bald}" + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": " corresponding to the balded source image. Since editing in " + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": " space inevitably gets other irrelevant attributes modified, we circumvent this issue by blending bald feature with source image feature in " + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{style}" + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": " space:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 699, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 699, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 699, + 287, + 715 + ], + "type": "interline_equation", + "content": "F _ {7} ^ {b a l d} = G \\left(w _ {1 - 7} ^ {b a l d}\\right) \\times M ^ {b a l d} + F _ {7} ^ {s r c} \\times \\left(1 - M ^ {b a l d}\\right), \\tag {1}", + "image_path": "4388b6f7f94a9424c7c219b5b619a6f741a7754c4c0cac0ff9abd12876965238.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": " stands for StyleGAN, " + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "inline_equation", + "content": "G(w_{1-7}^{bold})" + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": " represents the bald feature in " + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{style}" + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": " space. " + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "inline_equation", + "content": "M^{bald}" + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": " is the binary mask indicating the hair and ear regions of the source image, which is obtained by the facial parsing network BiseNET [42] and downsampled to " + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": ". With the guidance of " + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "inline_equation", + "content": "M^{bald}" + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": ", the irrelevant attributes region in " + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "inline_equation", + "content": "F_7^{bald}" + }, + { + "bbox": [ + 304, + 268, + 545, + 354 + ], + "type": "text", + "content": " can continue to be preserved in following proxy feature blending." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 359, + 410, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 359, + 410, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 359, + 410, + 373 + ], + "type": "text", + "content": "3.4. Hairstyle Editing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 378, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 401 + ], + "type": "text", + "content": "Below we elaborate on how to generate the proxy for different hairstyle conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": "Text Proxy. We formalize text proxy generation as the editing task done in " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": " space based on the CLIP loss guidance. Unlike prior works [36, 21, 40], our text proxy generation process is free from the pressure of irrelevant attribute preservation, and allows us to select a more suitable starting point for the optimization process, which leads to better editing effects. In order to ensure both the optimal editing effect and the diversity of editing results, we choose to sample a random point around the mean face latent code as the optimization starting point for our text proxy latent code " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "w^{text}" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": ". In detail, we adopt the truncation trick of StyleGAN as " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "w^{init} = w^{mean} + \\psi (w^{random} - w^{mean})" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "w^{mean}" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": " is the mean face latent code and " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "w^{random}" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": " is sampled randomly. By setting a small value of " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": ", we ensure that the initial optimization starting point " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "w^{init}" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": " is around the average face latent code " + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "inline_equation", + "content": "w^{mean}" + }, + { + "bbox": [ + 304, + 402, + 545, + 641 + ], + "type": "text", + "content": ". We will show the benefits of this initialization strategy in the ablation analysis. We adopt the CLIP loss with transformation augmentations [5] to perform the text-guided hairstyle editing while reducing the disturbance caused by adversarial examples:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 649, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 649, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 311, + 649, + 545, + 682 + ], + "type": "interline_equation", + "content": "L ^ {c l i p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} (1 - \\cos \\left(E _ {i} \\left(A _ {i} \\left(G \\left(w ^ {t e x t}\\right)\\right)\\right), E _ {t} (s t)\\right)), \\tag {2}", + "image_path": "b8f877878d9ea07d310ccc1669006bbd36924ef4659ee49ed95f6c3b57364621.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "-th transformation augmentation, " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " denotes the number of augmentations (" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " by default)," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23592" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " means cosine similarity, " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "G(w^{text})" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " means the editing result for each pass in the optimization process, " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " stands for the CLIP image encoder and text encoder respectively, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "st" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " refers to the user-supplied text description. Besides, pose alignment loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "L^{pose}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is utilized to ensure that the face shape and pose of text proxy are consistent with the source image to ease the subsequent feature blending:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 163, + 287, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 163, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 77, + 163, + 287, + 188 + ], + "type": "interline_equation", + "content": "L ^ {p o s e} = \\frac {1}{N _ {k}} | | E _ {p} (I ^ {s r c}) - E _ {p} (G (w ^ {t e x t})) | | _ {2} ^ {2}, \\qquad (3)", + "image_path": "3d8caf3fbde96b31cae462fb91b04e7dd5e0920e9d62dd722b82b40a23e5f773.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "text", + "content": " represents the 3D keypoint extractor [6] and " + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "text", + "content": " denotes the number of keypoints. Optionally, the shape loss " + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "inline_equation", + "content": "L^{shape}" + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "text", + "content": " is added to constrain the shape of the generated hair according to whether the user provides the hair region mask. We then obtain text proxy feature " + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "inline_equation", + "content": "F_{7}^{text}" + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "text", + "content": " using the optimized " + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "inline_equation", + "content": "w^{text}: F_{7}^{text} = G(w_{1-7}^{text})" + }, + { + "bbox": [ + 47, + 194, + 287, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": "Reference Proxy. Given a hairstyle reference image " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "I^{sr}" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": ", we generate the reference proxy by performing the unaligned hairstyle transfer task. " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "I^{sr}" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": " is first inverted by the II2S [49] embedding algorithm into the " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": " space to get " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "w^{ref}" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": ", which is served as the starting point for hairstyle transfer. During the transfer process, we expect to keep the original hair structure of the reference image while ensuring its pose and facial shape to be consistent with the source image. Thus, " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "L^{pose}" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": " is imposed between " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "G(w^{ref})" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "I^{src}" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": ". In addition, a style loss " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "L^{style}" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": " based on the gram matrix [11] is used to ensure that the hairstyle structure of " + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "inline_equation", + "content": "I^{sr}" + }, + { + "bbox": [ + 46, + 266, + 288, + 410 + ], + "type": "text", + "content": " remains unchanged during the alignment process:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 421, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 287, + 469 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L ^ {s t y l e} = \\frac {1}{4} \\sum_ {i = 1} ^ {4} \\left\\| \\mathcal {G} _ {i} \\left(V G G _ {i} \\left(I ^ {s r} \\times M ^ {r h}\\right)\\right) \\right. \\\\ \\left. - \\mathcal {G} _ {i} \\left(V G G _ {i} \\left(G \\left(w ^ {r e f}\\right) \\times M ^ {g h}\\right)\\right) \\right| | _ {2} ^ {2}, \\tag {4} \\\\ \\end{array}", + "image_path": "8a9b671d44192590a74ca4a36002e4a730eabb8ed4d7f589cef90cc276d1d195.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i(\\gamma_i)" + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": " represents the gram matrix calculated on the " + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": "-th layer features and a total of 4 layers of features are extracted, i.e., " + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\{relu1\\_2, relu2\\_2, relu3\\_3, relu4\\_3\\}" + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": " of VGG [27]. " + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "inline_equation", + "content": "M^{rh}" + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "inline_equation", + "content": "M^{gh}" + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": " are hair masks for the reference image and the generated image of each round during optimization process predicted by BiseNET [42]. For the same purpose, the " + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 46, + 476, + 287, + 573 + ], + "type": "text", + "content": " norm of the manipulation magnitude in the latent space is utilized during optimization:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 579, + 287, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 579, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 113, + 579, + 287, + 595 + ], + "type": "interline_equation", + "content": "L ^ {r e g} = \\left\\| w _ {t} ^ {r e f} - w _ {t - 1} ^ {r e f} \\right\\| _ {2} ^ {2}, \\tag {5}", + "image_path": "5293d3af428f038925a8e66fd1d3c081d4af9d993dfdacf4c90c946a1e5103e9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "w_{t}^{ref}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "w_{t-1}^{ref}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " represent the latent code of the current step and the previous step, respectively. Similar to the text proxy, " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "L^{shape}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " is optionally added to allow the user to customize the shape of the hair. We then obtain reference proxy feature " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "F_{7}^{ref}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " using the optimized " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "w^{ref}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "F_{7}^{ref} = G(w_{1-7}^{ref})" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Sketch Proxy. Enabling sketch-based local hairstyle editing within our framework is nontrivial. It is hard to find suitable losses to constrain the local hairstyle structure to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": "conform to the sketch given by the user. To circumvent this problem, we innovatively formalize the synthesis of sketch proxy as an image translation task based on StyleGAN. Utilizing the sketch-hair dataset created by SketchHairSalon [41], we train a sketch2hair inverter " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": ", which is based on E2Style [37] and aims to find the most appropriate latent code in " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " space to accurately translate a given sketch to the corresponding hair structure. The training loss consists of regular pixel-level " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " loss, feature-level LPIPS [44] loss and multi-layer face parsing loss [37] which is introduced to provide more local supervision. During the training process, we randomly remove a portion of the strokes to make our sketch2hair inverter adapt to a variety of sketch inputs from fine to coarse, e.g., even just one stroke. Given a local hairstyle sketch " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": ", our sketch proxy features " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "F_{7}^{sketch}" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " are synthesized by pre-trained sketch2hair " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " and StyleGAN " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 336, + 274, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 274, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 336, + 274, + 545, + 289 + ], + "type": "interline_equation", + "content": "w ^ {\\text {s k e t c h}} = T (S), \\quad F _ {7} ^ {\\text {s k e t c h}} = G \\left(w _ {1 - 7} ^ {\\text {s k e t c h}}\\right). \\tag {6}", + "image_path": "f087a9dfb3b99d25571db54e4f99334a8ce20c52c738c6c43457f6a9a0810677.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "content": "Proxy Feature Blending. For text and reference image based hairstyle condition, we perform global blending in " + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{style}" + }, + { + "bbox": [ + 304, + 293, + 545, + 328 + ], + "type": "text", + "content": " space:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 545, + 350 + ], + "type": "interline_equation", + "content": "F _ {7} ^ {\\text {g l o b a l}} = F _ {7} ^ {t r} \\times M ^ {\\text {g l o b a l}} + F _ {7} ^ {\\text {b a l d}} \\times \\left(1 - M ^ {\\text {g l o b a l}}\\right), \\tag {7}", + "image_path": "57b8265e02364c9d0d9b61baeda7b59e15f05549cb1ff2fd750592307129851e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "F_7^{tr} \\in \\{F_7^{text}, F_7^{ref}\\}" + }, + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "M^{global}" + }, + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "text", + "content": " is the binary mask corresponding to the hair region of " + }, + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "inline_equation", + "content": "F_7^{tr}" + }, + { + "bbox": [ + 304, + 358, + 545, + 396 + ], + "type": "text", + "content": ". Optionally, the sketch-based local hairstyle editing is applied:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 402, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 402, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 311, + 402, + 545, + 418 + ], + "type": "interline_equation", + "content": "F _ {7} ^ {\\text {s t y l e}} = F _ {7} ^ {\\text {s k e t c h}} \\times M ^ {\\text {l o c a l}} + F _ {7} ^ {\\text {g l o b a l}} \\times \\left(1 - M ^ {\\text {l o c a l}}\\right), \\tag {8}", + "image_path": "60a5aff40d18a72da4775e10599b6bb223f728ed2714070331d05fb3c57fbda6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "inline_equation", + "content": "M^{local}" + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "content": " is obtained by downsampling the user input sketch " + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "content": " after dilation. A natural concern is the artifacts brought by the mismatch between hair features within " + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "inline_equation", + "content": "M^{local}" + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "content": " and other hair features. But thanks to the semantic layering characteristics of StyleGAN, the resulting image shows consistent tones as these hair features will be modulated by the later layers. Our framework allows users to only edit the hairstyle: " + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "inline_equation", + "content": "I^{style} = G(F_7^{style}, w_{8-18}^{src})" + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "content": ", by skipping the following hair color editing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 540, + 418, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 418, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 418, + 553 + ], + "type": "text", + "content": "3.5. Hair Color Editing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": "We achieve hair color editing by performing proxy feature blending in " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{color}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " space. By choosing to use the feature " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "F_{7}^{src}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "F_{7}^{style}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": ", we allow to edit only hair color or both hair style&hair color. Below, we use " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "F_{7}^{style}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " as the example. Color Proxy. We initialize the color proxy with " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "F_{7}^{style}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "w_{8-18}^{color} = w_{8-18}^{src}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "w_{10-13}^{color}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " to be estimizable. The loss " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "L^{color}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " in the optimization process consists of " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "L^{modal}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "L^{bg}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "L^{modal}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "L^{clip}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " or the average color " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " loss of the hair region depending on the hair color condition types (text, reference image, RGB values), and " + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "inline_equation", + "content": "L^{bg}" + }, + { + "bbox": [ + 304, + 558, + 545, + 692 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 338, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 338, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "L ^ {b g} = \\left\\| \\left(I ^ {\\text {s t y l e}} - I ^ {\\text {c o l o r}}\\right) \\times \\left(M ^ {n - h a i r}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}", + "image_path": "8e85da4197f85d500403d84597b75beb5a77763898a7d2b02c9f248e5d174961.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23593" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 69, + 276, + 144 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 276, + 144 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 276, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 276, + 144 + ], + "type": "table", + "html": "
MethodsIDS↑PSNR↑SSIM↑
Ours0.8429.50.91
HairCLIP [36]0.4521.60.74
StyleCLIP [21]0.4319.60.72
TediGAN [40]0.1622.50.74
DiffCLIP [19]0.7126.80.86
", + "image_path": "b246691c416925cb7d161e51c5f330754b38ffe84767935a24ed54da9206e744.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 287, + 187 + ], + "lines": [ + { + "bbox": [ + 47, + 152, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 287, + 187 + ], + "type": "text", + "content": "Table 2. Quantitative comparison for irrelevant attributes preservation. IDS denotes identity similarity, PSNR and SSIM are calculated at the intersected non-hair regions before and after editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "inline_equation", + "content": "I^{color} = G(F_7^{style}, w_{8-18}^{color})" + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "inline_equation", + "content": "M^{n-hair}" + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "text", + "content": " is the mask of the non-hair region intersection between " + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "inline_equation", + "content": "I^{style}" + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "inline_equation", + "content": "I^{color}" + }, + { + "bbox": [ + 46, + 205, + 287, + 230 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "text", + "content": "Proxy Feature Blending. Even with the " + }, + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "inline_equation", + "content": "L^{bg}" + }, + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "text", + "content": " constraint, we find non-hair regions are often inevitably modified because of imperfect semantic decoupling of the " + }, + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "text", + "content": " space. We solve this by performing proxy feature blending in " + }, + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{color}" + }, + { + "bbox": [ + 46, + 231, + 287, + 291 + ], + "type": "text", + "content": " space, which also naturally supports local hair color editing:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 314, + 286, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 314, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 60, + 314, + 286, + 346 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {1 4} ^ {\\text {b l e n d}} = G \\left(F _ {7} ^ {\\text {s t y l e}}, w _ {8 - 1 4} ^ {\\text {c o l o r}}\\right) \\times M ^ {\\text {c o l o r}} \\\\ + G \\left(F _ {7} ^ {\\text {s t y l e}}, w _ {8 - 1 4} ^ {\\text {s r c}}\\right) \\times \\left(1 - M ^ {\\text {c o l o r}}\\right), \\tag {10} \\\\ \\end{array}", + "image_path": "4f28ef12f34db216849fa48bc9a164ad1b8e0bc70087a11ea3270a68ee72b774.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "M^{color}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " is the hair area mask or a local editing area mask drawn by the user. We set " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "F_{14}^{blend}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "w_{15-18}^{color}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " to be estimizable to further perform the optimization. In the optimization process, we use the " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "L_{blend}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " loss, which consists of " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " loss and LPIPS loss to constrain " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "I^{final}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " to be similar to " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "I^{color}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " inside " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "M^{color}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " and similar to " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "I^{style}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " outside " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "M^{color}" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": " simultaneously. The final edited image is synthesized as follows: " + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "inline_equation", + "content": "I^{final} = G(F_{14}^{blend}, w_{15-18}^{color})" + }, + { + "bbox": [ + 46, + 356, + 287, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 464, + 127, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 127, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 127, + 477 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 520 + ], + "type": "text", + "content": "Implementation details of our approach are provided in the supplementary material. For all compared methods, we use their official codes or pre-trained models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 527, + 264, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 264, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 264, + 540 + ], + "type": "text", + "content": "4.1. Quantitative and Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "Comparison with Text-Driven Hair Editing Methods. We compare HairCLIPv2 with leading text-driven hair editing methods on the CelebA-HQ [14] testset (2,000 images) and follow the evaluation settings of HairCLIP. For HairCLIP [36] and StyleCLIP [21] (\"Mapper\" version), we first invert using e4e [33] to obtain the latent code for a given real image before performing the editing. For Diffusion-CLIP [19], we finetune a model for each text description. For both TediGAN [40] and our method, the number of optimization iterations is set to 200. As shown in Figure 3 and Table 2, our method accomplishes satisfactory hair editing effects with better naturalness while maximizing the preservation of irrelevant attributes. It is worth noting that, even though HairCLIP and StyleCLIP also have pretty good hair" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "editing capabilities, they cannot preserve the irrelevant attributes very well such as background, identity and clothes. Our method also demonstrates better preservation of the original hair structure when editing only the hair color." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 121, + 546, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 217 + ], + "type": "text", + "content": "For arbitrary hair editing word scenarios, the only methods that are instantly feasible without retraining the model are HairCLIP, StyleCLIP (\"Optimization\" version), and TediGAN. As shown in Figure 4, our method performs much better at such cases. In contrast, HairCLIP can only produce plausible results for text (\"Curly Short Hairstyle\") similar to the training texts, while all other methods struggle to produce reasonable editing effects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 219, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 546, + 411 + ], + "type": "text", + "content": "Comparison with Hair Transfer Methods. We compare with state-of-the-art methods on hair transfer tasks. Among the 2,000 images of the CelebA-HQ testset, the first 666 are set as the input images, the middle 666 are set as the hairstyle reference images, and the last 666 are set as the hair color reference images. As shown in Figure 5, when the hairstyle reference image is broadly aligned with the input image (first row), most methods yield plausible results. However, when not aligned (second row), only our method and SYH [20] are able to perform a more consistent hair transfer, which is achieved by introducing the pose alignment loss during the transfer to ensure that the facial shape and pose of the reference image are consistent with the source image. Compared to SYH, we achieve comparable hair transfer results, but support text, sketch, and other interactions beyond hair transfer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 412, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 628 + ], + "type": "text", + "content": "Comparison with Local Hair Editing Methods. In terms of sketch-based local editing, we compare with the SOTA methods MichiGAN [31] and SketchSalon [41]. MichiGAN [31] uses user-drawn sketches to modify the orientation map to accomplish local hair editing. SketchSalon [41] trains a sketch-to-hair conditional translation network, with an additional soft alpha matte used to facilitate more natural blending. To generalize SketchSalon to local editing, we utilize the same mask as our method instead of a soft alpha matte, and the input sketch is colored as the average color within the mask area. As shown in Figure 6, MichiGAN struggles to perform satisfactory local editing and the reconstruction of other non-editing hair areas is slightly worse. Even ignoring the obvious blending artifacts, the local hair texture generated by SketchSalon is not in harmony with the surrounding hair. Compared to these two methods, our approach not only achieves satisfactory local editing but also better maintains the non-editing regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Comparison with Cross-Modal Hair Editing Methods. To the best of our knowledge, the only method that supports multimodal conditions to complete hairstyle and hair color editing is HairCLIP [36]. As the comparison shown in Figure 7, our approach not only perfectly prevents irrelevant attributes (identity, background, etc.) from being modified, but also achieves higher-quality editing effects. Moreover," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23594" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 550, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 550, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 550, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 550, + 167 + ], + "type": "image", + "image_path": "34eb8bcd0cef34dbd678faaf46f38e040c8132ea49727736eb75122c84c88757.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 177, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 547, + 212 + ], + "type": "text", + "content": "Figure 3. Visual comparison with HairCLIP [36], StyleCLIP-Mapper [21], TediGAN [40] and DiffusionCLIP [19]. The simplified text descriptions (editing hairstyle, hair color, or both of them) are listed on the leftmost side. Our approach demonstrates better editing effects and irrelevant attribute preservation (e.g., identity, background, etc.)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 232, + 286, + 356 + ], + "blocks": [ + { + "bbox": [ + 50, + 232, + 286, + 356 + ], + "lines": [ + { + "bbox": [ + 50, + 232, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 50, + 232, + 286, + 356 + ], + "type": "image", + "image_path": "4649bda1ca7f272b1d9613d58d623d581a1c77638d4853b0893c664aef84bbc0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 368, + 287, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 413 + ], + "type": "text", + "content": "Figure 4. Visual comparison with HairCLIP [36], StyleCLIP-Optimization [21] and TediGAN [40] under any description setting. We additionally provide an example image for each description for better comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 436, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 484 + ], + "type": "text", + "content": "HairCLIP allows only text and reference image while our method additionally supports sketch, mask, and RGB values. More diverse and comprehensive interactive editing results are shown in Figure 1 and supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 713 + ], + "type": "text", + "content": "User Study. For the above four types of comparisons, we recruit 20 volunteers with computer vision-related research backgrounds to execute a comprehensive user study. We randomly selected 20 groups of results from each experiment to form 80 test samples in total. The order of the different methods in each test sample is randomly shuffled. For each test sample, volunteers are asked to select the best option in terms of manipulation accuracy, irrelevant attribute preservation, and visual naturalness, respectively. As shown in Table 3, our method outperforms the baseline methods for most cases, except comparable results to Barbershop [47] and SYH [20] in the hair transfer setting. But our irrelevant attribute preservation performs best because of our hair color feature space blending mechanism, as demonstrated in Figure 5. It is worth mentioning that our goal is not to improve the performance of hair transfer, but to design a unified system that supports various hair editing and hair transfer tasks. Therefore, performing comparably with the state-of-the-art methods on the hair transfer task is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 231, + 353, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 353, + 242 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 353, + 242 + ], + "type": "text", + "content": "acceptable." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 251, + 411, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 411, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 411, + 264 + ], + "type": "text", + "content": "4.2. Ablation Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "text", + "content": "Importance of Initialization Strategy for Text Proxy Optimization. To justify our optimization starting point strategy for text proxy, we ablate three different optimization starting points: mean latent code, random sampling from " + }, + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "text", + "content": " space, and inverted latent code of the input image within " + }, + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "text", + "content": " space. All other settings remain the same. As illustrated in Figure 8, the first two settings are more likely to complete high-quality editing, while the last one fails because the starting point deviates from the more suitable editing region. This may explain why StyleCLIP (\"Optimization\" version) and TediGAN struggle in performing hairstyle editing. Our strategy of randomly initializing the latent code around the mean within " + }, + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 305, + 270, + 545, + 437 + ], + "type": "text", + "content": " space enjoys both good editability and diversity of the generated results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 605 + ], + "type": "text", + "content": "Superiority of Sketch Proxy Generation Design. To generate sketch proxy, we ablate another two possible ways. The first way is to constrain the orientation field of the sketch proxy to be similar to that of user's sketch in the drawing region with the orientation loss used in MichiGAN[31]. The second way is to first obtain the image corresponding to the sketch through the conditional translation network (for simplicity, we choose sketch2hair here), and then constrain the sketch proxy to be similar to it in the drawing area by LPIPS [44] loss during the optimization process. The comparison is shown in Figure 9. Only our method and the LPIPS optimization-based version are feasible. However, our method requires only a single feed forward, which is more efficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Feature Blending vs. Latent Code Blending. To demonstrate the superiority of proxy feature blending, we compare it with the alternative scheme based on the linear combination of latent codes. In detail, we initialize an interpolation factor for the latent code of the global editing proxy, the local sketch proxy, and the input image, respectively. In the optimization process, we optimize these three interpolation factors so that the generated image corresponding to the interpolated latent code is similar to the proxies or input im" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23595" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 541, + 194 + ], + "type": "image", + "image_path": "ebcb086fce6bb20533518fe772720a0fda213e8ec01ffc13f3e897e4daa454ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 546, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 546, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 546, + 224 + ], + "type": "text", + "content": "Figure 5. Visual comparison with HairCLIP [36], LOHO [24], Barbershop [47], SYH [20] and MichiGAN [31] on hair transfer. Only our method and SYH can accomplish unaligned hair transfer while keeping irrelevant attributes unmodified." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 232, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 48, + 232, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 48, + 232, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 545, + 293 + ], + "type": "table", + "html": "
Text-DrivenHair TransferSketch-BasedCross-Modal
MetricsOurs[36][21][40][19]Ours[36][24][47][20][31]Ours[31][41]Ours[36]
Accuracy41.5%32.3%22.5%1.0%2.8%28.0%2.3%4.8%29.3%28.5%7.3%76.8%22.0%1.3%82.8%17.3%
Preservation81.0%5.3%3.3%0.3%10.3%32.8%2.8%8.3%15.3%26.3%14.8%62.0%33.3%4.8%94.0%6.0%
Naturalness46.8%25.5%22.0%1.8%4.0%26.5%9.5%2.5%22.3%35.0%4.3%60.5%38.0%1.5%65.3%34.8%
", + "image_path": "2547992fc858954dc5393df222325d1e084650f32942e41865f5ba54dfbd36eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 296, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 296, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 546, + 331 + ], + "type": "text", + "content": "Table 3. User study on text-driven image manipulation, hair transfer, sketch-based local hair editing and cross-modal hair editing methods. Accuracy denotes the manipulation accuracy for given conditional inputs, Preservation indicates the ability to preserve irrelevant regions and Naturalness denotes the visual realism of the manipulated image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 53, + 347, + 281, + 452 + ], + "blocks": [ + { + "bbox": [ + 53, + 347, + 281, + 452 + ], + "lines": [ + { + "bbox": [ + 53, + 347, + 281, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 281, + 452 + ], + "type": "image", + "image_path": "0aba2e1a50f12be67d4c3708753dc3d84c8896f2a3319c92387b6f9d0566fb79.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 461, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 495 + ], + "type": "text", + "content": "Figure 6. Comparison with MichiGAN [31] & SketchSalon [41] on sketch-based local hair editing. We provide sketches in the facial parsing map for better visualization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 508, + 286, + 628 + ], + "blocks": [ + { + "bbox": [ + 50, + 508, + 286, + 628 + ], + "lines": [ + { + "bbox": [ + 50, + 508, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 50, + 508, + 286, + 628 + ], + "type": "image", + "image_path": "1c3548c36cf70617e6b98d144c2fc49c30e37749f0f87293fdf5269e6032b64f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 644, + 287, + 677 + ], + "lines": [ + { + "bbox": [ + 46, + 644, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 287, + 677 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison with HairCLIP on cross-modal conditional input. Our approach shows better editing effects & excellent preservation of irrelevant attributes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "age within the corresponding region. In Figure 10, we use" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 348, + 543, + 473 + ], + "blocks": [ + { + "bbox": [ + 310, + 348, + 543, + 473 + ], + "lines": [ + { + "bbox": [ + 310, + 348, + 543, + 473 + ], + "spans": [ + { + "bbox": [ + 310, + 348, + 543, + 473 + ], + "type": "image", + "image_path": "b60d871addebd77c7ef3ad985820fb651897a5ad09bbfddeb57f95d3d4248c0f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 484, + 545, + 507 + ], + "lines": [ + { + "bbox": [ + 306, + 484, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 545, + 507 + ], + "type": "text", + "content": "Figure 8. Ablation of different starting point settings for text proxy. The text description is \"Bob Cut Hairstyle\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 519, + 539, + 622 + ], + "blocks": [ + { + "bbox": [ + 313, + 519, + 539, + 622 + ], + "lines": [ + { + "bbox": [ + 313, + 519, + 539, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 539, + 622 + ], + "type": "image", + "image_path": "03a13bbab0f88b877b75793d23ed4e23c13cece812b23ec2dbd5ad48437a05aa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 632, + 545, + 654 + ], + "lines": [ + { + "bbox": [ + 306, + 632, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 545, + 654 + ], + "type": "text", + "content": "Figure 9. Ablation on sketch proxy generation design. We provide sketches drawn in the facial parsing map for better visualization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 546, + 714 + ], + "type": "text", + "content": "the hairstyle reference image as an example to generate the global proxy. It is obvious that our scheme accomplishes global editing and local editing while perfectly keeping the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23596" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 82, + 99, + 126 + ], + "blocks": [ + { + "bbox": [ + 57, + 73, + 95, + 81 + ], + "lines": [ + { + "bbox": [ + 57, + 73, + 95, + 81 + ], + "spans": [ + { + "bbox": [ + 57, + 73, + 95, + 81 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 82, + 99, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 82, + 99, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 82, + 99, + 126 + ], + "type": "image", + "image_path": "b7c89a31de86372f38bd31f33d5ab692d5e7ba4b703e0a716ffbec25c1bdd2f5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 100, + 82, + 144, + 126 + ], + "blocks": [ + { + "bbox": [ + 102, + 73, + 142, + 81 + ], + "lines": [ + { + "bbox": [ + 102, + 73, + 142, + 81 + ], + "spans": [ + { + "bbox": [ + 102, + 73, + 142, + 81 + ], + "type": "text", + "content": "Hairstyle Ref" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 82, + 144, + 126 + ], + "lines": [ + { + "bbox": [ + 100, + 82, + 144, + 126 + ], + "spans": [ + { + "bbox": [ + 100, + 82, + 144, + 126 + ], + "type": "image", + "image_path": "f211568a8e6fe6a605adee23e81e7cbebc256439e14be4519cfe6e7154efc9e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 145, + 82, + 189, + 126 + ], + "blocks": [ + { + "bbox": [ + 148, + 73, + 185, + 81 + ], + "lines": [ + { + "bbox": [ + 148, + 73, + 185, + 81 + ], + "spans": [ + { + "bbox": [ + 148, + 73, + 185, + 81 + ], + "type": "text", + "content": "Input Sketch" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 82, + 189, + 126 + ], + "lines": [ + { + "bbox": [ + 145, + 82, + 189, + 126 + ], + "spans": [ + { + "bbox": [ + 145, + 82, + 189, + 126 + ], + "type": "image", + "image_path": "667c28a35a9aa5cbf65e7c3971f661e07569f0213d8a5f1f36561823ca9bdbd1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 190, + 82, + 235, + 126 + ], + "blocks": [ + { + "bbox": [ + 202, + 73, + 223, + 80 + ], + "lines": [ + { + "bbox": [ + 202, + 73, + 223, + 80 + ], + "spans": [ + { + "bbox": [ + 202, + 73, + 223, + 80 + ], + "type": "text", + "content": "Feature" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 82, + 235, + 126 + ], + "lines": [ + { + "bbox": [ + 190, + 82, + 235, + 126 + ], + "spans": [ + { + "bbox": [ + 190, + 82, + 235, + 126 + ], + "type": "image", + "image_path": "66dd24697518234b8a20d4d9f88537628df65b9e36242407a38284f4a80f17b6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 236, + 82, + 280, + 126 + ], + "blocks": [ + { + "bbox": [ + 239, + 73, + 275, + 80 + ], + "lines": [ + { + "bbox": [ + 239, + 73, + 275, + 80 + ], + "spans": [ + { + "bbox": [ + 239, + 73, + 275, + 80 + ], + "type": "text", + "content": "Latent Code" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 82, + 280, + 126 + ], + "lines": [ + { + "bbox": [ + 236, + 82, + 280, + 126 + ], + "spans": [ + { + "bbox": [ + 236, + 82, + 280, + 126 + ], + "type": "image", + "image_path": "0d705271d6e9012ee91d19621f2f910db456d524985e4f7a7236329ee8c0ae1a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 55, + 128, + 99, + 172 + ], + "blocks": [ + { + "bbox": [ + 55, + 128, + 99, + 172 + ], + "lines": [ + { + "bbox": [ + 55, + 128, + 99, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 99, + 172 + ], + "type": "image", + "image_path": "cc73e7b29a35d6a01fa2a7e98aa511868a5946d9a70701c357eafafd967b3673.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 184, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 184, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 286, + 217 + ], + "type": "text", + "content": "Figure 10. Ablation of feature blending vs. latent code blending. Feature blending can enable global editing, local editing, and irrelevant attributes preservation simultaneously." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 100, + 128, + 144, + 172 + ], + "blocks": [ + { + "bbox": [ + 100, + 128, + 144, + 172 + ], + "lines": [ + { + "bbox": [ + 100, + 128, + 144, + 172 + ], + "spans": [ + { + "bbox": [ + 100, + 128, + 144, + 172 + ], + "type": "image", + "image_path": "356ac340bdc8f91a9f122bd0f08bc59e417fc9459ce502724eb941e13b453e15.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 145, + 128, + 189, + 172 + ], + "blocks": [ + { + "bbox": [ + 145, + 128, + 189, + 172 + ], + "lines": [ + { + "bbox": [ + 145, + 128, + 189, + 172 + ], + "spans": [ + { + "bbox": [ + 145, + 128, + 189, + 172 + ], + "type": "image", + "image_path": "ddfff9c5674db2e34c7c783f725aa50459df5329a1616fd00de43e4cf27fc828.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 190, + 128, + 235, + 172 + ], + "blocks": [ + { + "bbox": [ + 190, + 128, + 235, + 172 + ], + "lines": [ + { + "bbox": [ + 190, + 128, + 235, + 172 + ], + "spans": [ + { + "bbox": [ + 190, + 128, + 235, + 172 + ], + "type": "image", + "image_path": "6ea7127313a5e0c06f038295aa076571c4193d52391eb20d8c7715fcc6345e7a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 236, + 128, + 280, + 172 + ], + "blocks": [ + { + "bbox": [ + 236, + 128, + 280, + 172 + ], + "lines": [ + { + "bbox": [ + 236, + 128, + 280, + 172 + ], + "spans": [ + { + "bbox": [ + 236, + 128, + 280, + 172 + ], + "type": "image", + "image_path": "a62eadfcb28dde362307aa7d6febc722b0260d64df006ae65847c62abfaec356.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 238, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 286, + 262 + ], + "type": "text", + "content": "irrelevant properties unmodified, while latent code blending does not perform either one well." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 274, + 198, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 198, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 198, + 285 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "type": "text", + "content": "In this paper, we propose a unified hair editing system HairCLIPv2, which presents the first attempt that supports both simple text/reference image interaction and fine-grained local interactions. It innovatively converts all hair editing tasks into hair transfer tasks, with the corresponding editing conditions converted into transfer proxies. It can not only achieve high-quality hair editing results, but also well preserve the irrelevant attributes from being modified. In the future, we will study how to use feed-forward networks to generate all the proxies. Also, it is worthy to generalize the proposed framework to support generic natural images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 437, + 157, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 157, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 157, + 451 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 458, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 613 + ], + "type": "text", + "content": "This work was supported in part by the Natural Science Foundation of China under Grant U20B2047, 62072421, 62002334, 62102386 and 62121002, the Fundamental Research Funds for the Central Universities under Grant WK5290000003, Key Research and Development program of Anhui Province under Grant 2022k07020008. This work was also partly supported by Shenzhen Key Laboratory of Media Security, and the Opening Fund of Key Laboratory of Cyberculture Content Cognition and Detection, Ministry of Culture and Tourism. This work was also partially supported by a GRF grant (Project No. CityU 11216122) from the Research Grants Council (RGC) of Hong Kong. Thank Yi Yin for her help in this work." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 48, + 625, + 106, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 106, + 637 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 106, + 637 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 52, + 646, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 52, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 52, + 646, + 287, + 689 + ], + "type": "text", + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 139 + ], + "type": "text", + "content": "[3] Abdul Fatir Ansari, J. Scarlett, and Harold Soh. A characteristic function approach to deep implicit generative modeling. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7476-7484, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 140, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 545, + 161 + ], + "type": "text", + "content": "[4] Martín Arjovsky, Soumith Chintala, and L. Bottou. Wasserstein generative adversarial networks. In ICML, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 205 + ], + "type": "text", + "content": "[5] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 259 + ], + "type": "text", + "content": "[6] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 262, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 545, + 304 + ], + "type": "text", + "content": "[7] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 306, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 306, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 312, + 306, + 545, + 369 + ], + "type": "text", + "content": "[8] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2023), 2023." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 545, + 425 + ], + "type": "text", + "content": "[9] Ruili Feng, Jie Xiao, Kecheng Zheng, Deli Zhao, Jingren Zhou, Qibin Sun, and Zheng-Jun Zha. Principled knowledge extrapolation with GANs. In Proceedings of the 39th International Conference on Machine Learning, volume 162, pages 6447-6464. PMLR, 17-23 Jul 2022." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 427, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 458 + ], + "type": "text", + "content": "[10] Ruili Feng, Deli Zhao, and Zheng-Jun Zha. Understanding noise injection in gans. In International Conference on Machine Learning, pages 3284-3293. PMLR, 2021." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "text", + "content": "[12] Ishaan Gulrajani, F. Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C. Courville. Improved training of wasserstein gans. In NIPS, 2017." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[13] Youngjoo Jo and Jongyoul Park. Sc-fegan: Face editing generative adversarial network with user's sketch and color. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1745-1753, 2019." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[14] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021." + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23597" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 173 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 228 + ], + "type": "text", + "content": "[19] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 285 + ], + "type": "text", + "content": "[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 288, + 341 + ], + "type": "text", + "content": "[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 408 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "text", + "content": "[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 520 + ], + "type": "text", + "content": "[24] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1984-1993, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 521, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 566 + ], + "type": "text", + "content": "[25] Edgar Schonfeld, B. Schiele, and A. Khoreva. A u-net based discriminator for generative adversarial networks. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8204-8213, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 644 + ], + "type": "text", + "content": "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[28] Xinhui Song, Chen Liu, Youyi Zheng, Zunlei Feng, Lincheng Li, Kun Zhou, and Xin Yu. Hairstyle editing via parametric controllable strokes. IEEE Transactions on Visualization and Computer Graphics, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[29] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "linguistic representations. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 151 + ], + "type": "text", + "content": "[30] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Bin Liu, Gang Hua, and Nenghai Yu. Diverse semantic image synthesis via probability distribution modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[31] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. ACM Transactions on Graphics (TOG), 39(4):95-1, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[32] Zhentao Tan, Dongdong Chen, Qi Chu, Menglei Chai, Jing Liao, Mingming He, Lu Yuan, Gang Hua, and Nenghai Yu. Efficient semantic image synthesis via class-adaptive normalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[33] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[34] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 420 + ], + "type": "text", + "content": "[35] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS 2022), 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 421, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 477 + ], + "type": "text", + "content": "[36] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Zhentao Tan, Lu Yuan, Weiming Zhang, and Nenghai Yu. Hairclip: Design your hair by text and reference image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18072-18081, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 532 + ], + "type": "text", + "content": "[37] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2style: Improve the efficiency and effectiveness of stylegan inversion. IEEE Transactions on Image Processing, 31:3267-3280, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 533, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 577 + ], + "type": "text", + "content": "[38] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization. In ICML, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 578, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 623 + ], + "type": "text", + "content": "[39] Yiqian Wu, Yong-Liang Yang, and Xiaogang Jin. Hairmapper: Removing hair from portraits using gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4227-4236, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[40] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. Tedigan: Text-guided diverse face image generation and manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[41] Chufeng Xiao, Deng Yu, Xiaoguang Han, Youyi Zheng, and Hongbo Fu. Sketchhairsalon: Deep sketch-based hair image synthesis. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23598" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 529 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[43] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 306 + ], + "type": "text", + "content": "[45] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copypaste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning (ICML 2023), 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 307, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 307, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 287, + 350 + ], + "type": "text", + "content": "[46] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 352, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 352, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 287, + 395 + ], + "type": "text", + "content": "[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 396, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 439 + ], + "type": "text", + "content": "[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In European Conference on Computer Vision, pages 651-667. Springer, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 441, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 475 + ], + "type": "text", + "content": "[49] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved stylegan embedding: Where are the good latents? arXiv preprint arXiv:2012.09036, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 529 + ], + "type": "text", + "content": "[50] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. Sean: Image synthesis with semantic region-adaptive normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5104-5113, 2020." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23599" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_content_list.json b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b64d01667781bcae3e67710bc7e972e5f2fe5f2e --- /dev/null +++ b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_content_list.json @@ -0,0 +1,1445 @@ +[ + { + "type": "text", + "text": "HairNeRF: Geometry-Aware Image Synthesis for Hairstyle Transfer", + "text_level": 1, + "bbox": [ + 138, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seunggyu Chang* \nNAVER Cloud", + "bbox": [ + 238, + 180, + 385, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gihoon Kim† \nKAIST", + "bbox": [ + 446, + 180, + 555, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hayeon Kim† \nUNIST", + "bbox": [ + 617, + 180, + 728, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b00d7051ac27b798ac28133d5f4c6663d3a24415b743c2cb5e66ebd06867678e.jpg", + "image_caption": [ + "Figure 1: Given portrait images of hairstyle and identity (thunnails) that differ in viewing angle, head shape, and size, we seamlessly transfer hairstyle from one image to another utilizing neural rendering. Our method preserves fine-level details of the hairstyle structural geometry and generates diverse range of pose variant transferred images without requiring additional transfer process." + ], + "image_footnote": [], + "bbox": [ + 76, + 268, + 142, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ec830d9680b4c305b9eaeaf8192b519884e0296137a75d517b554f9f3c0e8dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 268, + 274, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a24d5e76de7c7f11d31cb3d624953ef561e2593586c65b62c149fe6936ab7d82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 268, + 890, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 622, + 313, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel hairstyle transferred image synthesis method considering the underlying head geometry of two input images. In traditional GAN-based methods, transferring hairstyle from one image to the other often makes the synthesized result awkward due to differences in pose, shape, and size of heads. To resolve this, we utilize neural rendering by registering two input heads in the volumetric space to make a transferred hairstyle fit on the head of a target image. Because of the geometric nature of neural rendering, our method can render view varying images of synthesized results from a single transfer process without causing distortion from which extant hairstyle transfer methods built upon traditional GAN-based generators suffer. We verify that our method surpasses other baselines in view of pre", + "bbox": [ + 75, + 652, + 472, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "serving the identity and hairstyle of two input images when synthesizing a hairstyle transferred image rendered at any point of view.", + "bbox": [ + 500, + 625, + 893, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 705, + 632, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Selecting an appropriate hairstyle can be difficult as it is hard to imagine how it will look on your face without trying it out for yourself. The reason can be attributed to the inherent variability of individual head shape, resulting in the frequent occurrence of disparate outcomes when attempting to emulate celebrity hairstyles. To address this issue, we propose a novel image synthesis method for hairstyle transfer tailored to fit an individual head shape. We named our method HairNeRF as it considers geometric alignment in the volumetric space of neural rendering [33]. Hairstyle transfer synthesizes an image where the hairstyle from one", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: seunggyu.chang@navercorp.com", + "bbox": [ + 94, + 875, + 447, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done during internship at Naver Cloud.", + "bbox": [ + 96, + 887, + 400, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2448", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "portrait is transferred to the other and Generative adversarial networks (GANs) [14, 20, 22] have been extensively utilized to address this problem. Notably, following the emergence of StyleGAN [20, 21, 22], which enables the generation of highly realistic images, studies [37, 50, 19, 8] have been made to optimize latent vectors within the well-trained latent space of StyleGAN to generate an optimal synthesis outcome. These methods are successful at transferring hairstyles between images having similar poses, but struggle to generate convincing results for images having vastly different poses. To resolve this, studies [50, 24, 51] have proposed methods aligning pose of transferred hairstyle to a target portrait, simultaneously reducing the perceptual difference between the transferred result and the original hairstyle. However, they inherently encounter the same limitation of the pre-trained latent space of StyleGAN, where pose and style attributes are entangled as StyleGAN is not explicitly trained for hairstyle transfer. Consequently, the alignment process may distort the geometry of an original hairstyle, leading to undesirable alterations such as forehead stretching or shrinking.", + "bbox": [ + 80, + 90, + 467, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We categorize entanglement of pose and style attributes into two folds: entanglement of style with viewpoint, and entanglement of style with head shape. To address the entanglement of style with viewpoint, we leverage the latent space of StyleNeRF [15], a neural radiance field (NeRF) [32, 33] based GAN model as an alternative for StyleGAN. StyleNeRF exhibits an innate pose-invariant latent space as the network implies the volumetric geometry of an image. We basically follow the latent optimization technique following the literature [50, 24, 51] utilizing StyleGAN. Leveraging the pose-invariant nature of the StyleNeRF's latent space, we extract a pose-invariant style vector through GAN inversion [42]. Thereafter, we recombine the volumetric features generated by the inverted style vectors of the hair and the target portraits. However, despite the pose-invariant nature of StyleNeRF's latent space, it still entails an entanglement problem between style and head shape. The optimization of a style vector to fit on the head of a target portrait, while retaining hairstyle of a hair portrait, can potentially result in deformations of the hairstyle. To address the entanglement of style with head shape, we propose a registration method that aligns the geometries of two heads within the volumetric space, while maintaining the inverted style vectors intact. Subsequently, we introduce a novel rendering method that recombines hairstyle features with facial features within the volumetric space using alignment-inducing deformable rays, which are derived from the registration outcome. In this manner, we resolve the distorted hairstyle problem from which StyleGAN-based method suffers, resulting in better alignment of the transferred hairstyle seamlessly fit on the target portrait. Taking the advantage of geometric nature of neural rendering, HairNeRF is ca", + "bbox": [ + 80, + 417, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pable of generating a diverse set of geometrically consistent novel view hairstyle transferred images from a single synthesis, by merely adjusting camera parameters. This is in contrast to existing StyleGAN-based hairstyle transfer methods which necessitate separate optimizations for each of novel views, which often yield inconsistent outcomes.", + "bbox": [ + 503, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results on diverse datasets demonstrate the superiority of HairNeRF against other baselines at generating seamless hairstyle transfer outcomes, while effectively preserving the underlying geometry of the original hairstyle.", + "bbox": [ + 503, + 181, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 272, + 638, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Hairstyle Transfer", + "text_level": 1, + "bbox": [ + 503, + 297, + 676, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proficiency exhibited by generative adversarial networks (GANs) [14, 20, 22] in generating photorealistic images has propelled a surge of research addressing image editing problem. A significant amount of research [3, 7, 19, 16, 12] has effectively addressed attribute editing or style transfer, which modify abstract-level features of an image. However, the task of hairstyle transfer still presents a challenging problem that necessitates the preservation of intricate structure and color details of a hairstyle from a source image. With the emergence of StyleGAN [20, 21, 22], which demonstrates exceptional quality in generating photorealistic facial images, efforts have been made to discover latent vectors capable of capturing transferred images of hairstyles [37, 50]. These methods require the projection of an image into the pretrained StyleGAN's latent space known as inversion, which influences the quality of transfer. Thus, the Barbershop [50] introduces an improved inversion method on the enlarged latent space named $FS$ -space to capture finer details of input image. However, since these models lack awareness of the pose of given images, they suffer structural artifacts when the given two images largely differ in poses. Recent studies [24, 51] resolve this issue by aligning pose of the transferred hairstyle to the target image. However, they still suffer style deformation problem arising from the entangled latent space of StyleGAN for pose and style.", + "bbox": [ + 503, + 321, + 890, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. 3D-Aware Image Synthesis and Editing", + "text_level": 1, + "bbox": [ + 503, + 724, + 834, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The emergence of StyleGAN [20, 21, 22] has facilitated researches on image editing [16, 41, 2, 19, 29, 23, 1], either by navigating the latent style space or leveraging semantic maps. However, StyleGAN-based image editing methods have shown inconsistency among pose changes, which is attributed to the entanglement of style and pose within StyleGAN's latent space. The development of neural radiance fields (NeRF) [32, 33] has enabled the rendering of 3D-consistent images that accommodate varying viewpoints. Consequently, methods for editing a 3D scene", + "bbox": [ + 503, + 750, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2449", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a63f84764f186cfad92feb7f029ddb104ae386b3a56596919f3d276bc7ca157e.jpg", + "image_caption": [ + "Figure 2: Overview of the hairstyle transfer pipeline of HairNeRF consisting of five modules: (a) inversion, (b) head alignment, (c) occlusion disclosure, (d) recombination, and (e) blending." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[44, 30, 27, 48] such as the addition and subtraction of NeRF models trained on individual objects have been explored. However, these approaches do not support modifications, such as altering the attributes or style of the object, due to the inherent nature of NeRF being trained to solely represent a single entity, thus lacking the necessary information for the desired edits. The advent of generative NeRFs [5, 38, 35, 15, 4, 10, 43] has enhanced the representational capabilities of NeRF models, enabling the expression of an expansive range of objects through the exploration of the latent space within a single model. The enhanced expressive capabilities of generative NeRFs enable the modification of attributes and style of an object [15, 4, 10]. Furthermore, within the generative NeRF modules, methods [30, 18, 40, 39, 6] have been proposed that disentangle the semantic structure and texture. However, these methods are not well-suited for local style modifications, as the latent vector representing the texture exhibits spatial entanglement, encompassing the entire regions of the image.", + "bbox": [ + 75, + 388, + 472, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 696, + 168, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 723, + 187, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the basic synthesis pipeline from the Barbershop [50], a founding work for hairstyle transfer [24, 50]. We encourage readers to refer [50] for the details of the original pipeline. The overall synthesis pipeline of our method consists of four modules: inversion, head alignment, recombination, and blending, as illustrated in Figure 2. Given two images, id and hair, we transfer hairstyle from the hair image into the id image. In the inversion module, we first invert the two images to obtain latent vectors and the corresponding features in the volu", + "bbox": [ + 75, + 750, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "metric space. Then, in the alignment module, we conduct registration on two geometric heads of the inverted images in the volumetric space. By aligning geometric heads, we can transfer hairstyle of the hair image into the id image to be more fit. Once the registration is done, we recombine the 3D features in the volumetric space according to semantics to render a 2D feature for synthesis using neural rendering in the recombination module. A simple method is to bring hair semantic features from the hair image, and other semantics from the id image. However, there occur missing regions which are occluded by the hair in the id image but not covered by the hair of the hair image. To fill in the missing regions, we inpaint 3D features by optimizing a latent vector, ensuring that the 2D features rendered from the 3D features generated by the optimized latent vector align with the target semantics across all viewpoints. Thereafter, we semantically integrate the 3D features from the id, hair, and the inpainting output to form a blended feature for the synthesis. The blend feature is rendered to a 2D feature, which is decoded by the upsampler blocks of StyleNeRF [15] conditioned on the blending style vector. In the final blending module, the blending style vector is optimized to make the decoded image preserve the same hairstyle as in the hair image while retaining face of the id image.", + "bbox": [ + 496, + 388, + 893, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. NeRF Inversion", + "text_level": 1, + "bbox": [ + 500, + 784, + 658, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic Parsing NeRF Head. To semantically recombine features in the 3D volume, semantics of volumetric features are required. It is known that a semantic classifier trained to classify a 2D point feature obtained by (2) can aptly classify semantic of a 3D point feature as well [25]. Based on this observation, we add an additional seman-", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2450", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aba1ee09fecc380b92a6efedf7e9cb6b5e318b8210adf7dff90c330537c7706a.jpg", + "image_caption": [ + "Figure 3: The additional semantic head is added at the end of a MLP block of NeRF module, which outputs a logit of semantics in the volumetric space. The additional semantic head is trained to minimize cross-entropy between the rendered logit, $\\hat{s}$ , and the ground-truth." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 452, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tic branch to the MLP block for neural rendering and train it using outputs of the off-the-shelf semantic segmentation model [46] for generated images as ground-truths as shown in Figure 3. After training, the trained semantic branch outputs a logit of semantics for a 3D point feature.", + "bbox": [ + 75, + 315, + 468, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics-Aware Inversion in 3D. We invert the hair and the id images into the latent vectors for the synthesis network of the StyleNeRF [15]. The process of obtaining the latent vector generating a desired image is called GAN inversion. We utilize the pivotal tuning inversion (PTI) [36] with a little modification for inversion as the PTI can reconstruct fine details of an input image from the inverted latent vector. As our model is built upon the StyleNeRF, the output image varies as a camera parameter changes. For this reason, the inversion process requires an input camera parameter or should estimate a camera parameter in our case. We choose to provide a fixed camera parameter of an input image estimated by the off-the-shelf head pose estimator [31] for the inversion process. Using the PTI with a fixed camera parameter, we optimize $W \\in \\mathcal{W}$ for each of the id and the hair images, and finetune the weights of the pretrained StyleNeRF to reconstruct the input image and the semantics estimated by the off-the-shelf semantic parsing networks [47, 45].", + "bbox": [ + 75, + 396, + 470, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Head Alignment", + "text_level": 1, + "bbox": [ + 76, + 700, + 238, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Registration For registration, we render each of inverted images into $n_c$ different viewpoints and extract 2D keypoints of the FLAME [28] head model using the off-the-shelf head pose estimator [31]. Then we estimate 3D keypoints of the FLAME model for each inverted image from the set of 2D keypoints estimated from the eight different camera views by solving a bundle-adjustment-like least square problem:", + "bbox": [ + 75, + 726, + 468, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {p}} {\\operatorname {m i n i m i z e}} \\sum_ {i} ^ {n _ {c}} \\| P (\\boldsymbol {c} _ {i}, \\boldsymbol {p}) - \\boldsymbol {x} _ {i} \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 864, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}_i \\in \\mathbb{R}^2$ denotes the estimated 2D keypoints at $i$ -th viewpoint, $\\pmb{c}_i$ denotes the $i$ -th camera parameter, and $P(\\pmb{c}_i, \\pmb{p}_i)$ denotes the projection function projecting a 3D point $\\pmb{p}_i \\in \\mathbb{R}^3$ into 2D using $\\pmb{c}_i$ . As we already know the camera parameters used for rendering $n_c$ images, we exclude camera parameters from the optimization variables. We utilize the thin-plate-spline (TPS) transform [11] to find the optimal non-rigid transformation for the registration.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Rendering via Ray Deformation. The two heads of the id and the hair do not fully overlap due to the differences in characteristics, shape and size of the heads geometry, along with estimation errors in head poses. We refer to this misalignment as intrinsic misalignment, as it arises from factors unrelated to external elements such as camera parameters. To alleviate the intrinsic misalignment, the registration process finds an optimal transformation between two point sets, by which one point set is transformed to align with the other. During the neural rendering process, point features are sampled along a ray within the volumetric space, and these features are then aggregated to a single point on the image plane. If the sampling points are collected along the ray, deformed by the transformation determined through registration, we can obtain registered point features. This process leads to a registered 2d image after the aggregation.", + "bbox": [ + 496, + 214, + 892, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4 illustrates the effective neural rendering process for alignment via ray deformation. Let $T_{i\\rightarrow h}$ denote a transformation determined through registration, which aligns the point set of the id features to the point set of the hair features. As depicted in Figure 4a, sampling along rays that radiate straight from the id's camera renders an id's face with an innate head shape, colored in red, as in shown Figure 4c. Likewise, rendering from straight rays radiating through hair features $\\Phi^{\\text{hair}}$ using the same camera renders the hair image posed as id, with the overlaid innate head shape colored in blue, as depicted in Figure 4e. The two heads do not overlap due to differences in shape and size. However, if we render a hair image using the deformed ray transformed by $T_{i\\rightarrow h}$ , as depicted in Figure 4b, we obtain a registered hair image aligned to the head of id, as shown in Figure 4d. With a camera ray $r = o + td$ , the registered neural rendering computes a pixel value of the ray as:", + "bbox": [ + 496, + 470, + 892, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\omega} ^ {\\text {R e g i s t e r e d}} (\\boldsymbol {r}) = \\int_ {0} ^ {\\infty} p _ {\\omega} (t) \\boldsymbol {c} _ {\\omega} (\\tilde {\\boldsymbol {r}} (t), \\tilde {\\boldsymbol {d}}) \\left| \\frac {d \\tilde {\\boldsymbol {r}}}{d t} \\right| d t, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 750, + 890, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} p _ {\\omega} (t) = \\exp \\left(- \\int_ {0} ^ {t} \\sigma_ {\\omega} (\\tilde {\\boldsymbol {r}} (s)) \\left| \\frac {d \\tilde {\\boldsymbol {r}}}{d s} \\right| d s\\right) \\cdot \\sigma_ {\\omega} (\\tilde {\\boldsymbol {r}} (t)),\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 781, + 861, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with a registered ray $\\tilde{\\boldsymbol{r}} = T_{i\\rightarrow h}(\\boldsymbol {r})$ , and the adjusted directional vector $\\tilde{\\pmb{d}} = d\\tilde{\\pmb{r}} /dt$ .Equation (2) has the same form as the original neural rendering, but with the integration computed along the transformed ray. This adjustment", + "bbox": [ + 500, + 818, + 890, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 or a registered 2d feature map in the StyleNeRF [15]", + "bbox": [ + 517, + 886, + 803, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2451", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6508a6e9c44b8e6b01f79c6722d561290c3143acf9e761fc61d3dd01fc15cd4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 89, + 524, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7486f50edcfdd22c1d2ec3d87bde1bfb8b80e929cff5a632406d3f5d3fe7bab2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 99, + 852, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e18f7cda8ee41c1cceea31f95d8ccf3c22ad62f89fd27ed09946c74e6852570f.jpg", + "image_caption": [ + "Figure 4: Illustration of alignment-aware neural rendering via ray deformation. Figure (c) - (e), and (h) - (j) show overlaid heads on rendering images. Rendering via transformed ray render a registered image as shown in (d) and (i), which is equivalent to rendering with the straight ray on transformed feature represented as $\\Phi^{\\text{hair}'}$ and $\\Phi^{\\text{id}'}$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 122, + 219, + 521, + 342 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d7e1f1332de8ae6dd9026ca8050544ece552d4bf51b6152828e839dae8e99033.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 226, + 849, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is achieved by substituting the directional vector $\\pmb{d}$ with $\\tilde{\\pmb{d}}$ and including $|d\\tilde{\\pmb{r}} /dt|$ to adjust the infinitesimal length $dt$ . Since $|d\\tilde{\\pmb{r}} /dt| \\approx 1$ in practice, we omit the determinant term to simplify the computation.", + "bbox": [ + 75, + 424, + 470, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Volumetric Recombination", + "text_level": 1, + "bbox": [ + 76, + 494, + 318, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Occlusion Disclosure. To inpaint the missing regions occluded by the hair in $\\Phi^{\\mathrm{hair}}$ and produce the recombined 3D features $\\Phi^{\\mathrm{blend}}$ as depicted in Figure 2c, we optimize the latent vector $W^{\\mathrm{inp}} \\in \\mathcal{W}+$ within the StyleNeRF [15] framework to generate $\\Phi^{\\mathrm{inp}}$ . Analogous to [50], we produce a volumetric target semantic mask, as depicted in Figure 5. For each iteration, we optimize $W^{\\mathrm{inp}}$ to generate $\\Phi^{\\mathrm{inp}}$ that aligns with the target semantic mask $s^{\\mathrm{inp}}$ from a randomly selected viewpoint. Let $m_{hair}$ denote a volumetric binary mask representing hair, $c$ denote a camera parameter, and $\\phi|_c$ represents 2D features rendered by the camera parameter $c$ . The target semantics $s^{\\mathrm{inp}}$ and the objectives for the optimization are constructed as follows:", + "bbox": [ + 75, + 518, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {s} ^ {\\text {i n p}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}} \\circ \\boldsymbol {s} ^ {\\text {h a i r} ^ {\\prime}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}}\\right) (3) \\\\ \\circ \\Big (\\left(1 - \\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}}\\right) \\circ \\boldsymbol {s} ^ {\\mathrm {i d}} + \\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}} \\circ \\boldsymbol {s} ^ {\\mathrm {h a i r} ^ {\\prime}} \\Big), \\\\ \\mathcal {L} ^ {\\text {i n p - s e m}} = \\mathbb {E} _ {c} \\mathrm {C E} \\left(\\mathcal {R} _ {c} \\left(s ^ {\\text {i n p}}\\right), \\mathcal {R} _ {c} \\left(f _ {W ^ {\\text {i n p}}} ^ {\\text {s e m}}\\right)\\right), (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 726, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} ^ {\\text {i n p - r e g}} = \\mathbb {E} _ {\\boldsymbol {c}} \\left\\| G \\left(\\left(1 - \\mathcal {R} _ {\\boldsymbol {c}} \\left(\\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {i n p}}\\right)\\right) \\circ U _ {W} \\left(\\phi^ {\\text {i n p}} \\mid_ {\\boldsymbol {c}}\\right)\\right) \\right. \\\\ \\left. - G \\Big (\\big (1 - \\mathcal {R} _ {\\boldsymbol {c} ^ {\\mathrm {i d}}} (\\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}}) \\big) \\circ I ^ {\\mathrm {i d}} \\Big) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 835, + 449, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_c$ represents the aggregation function of neural rendering according to the camera $c$ , $f^{\\mathrm{sem}}$ denotes the MLP block with semantic head, $U_W$ refers to the upsampler block of StyleNeRF whose output is a final image, and $G$ denotes the Gram matrix for style loss [13]. Symbols marked with an apostrophe indicate that they are registered to their counterparts, as denoted by the associated superscripts. Unlike [50], our target semantic mask is defined as a continuous function within the volumetric space, rather than as a tensor defined at discrete positions. As a result, $\\mathcal{L}^{\\mathrm{inp - sem}}$ entails integration over this volumetric space. Hence, for each optimization iteration in practice, we compute $\\mathcal{L}^{\\mathrm{inp - sem}}$ using a random set of rays radiating from a randomly sampled camera $c$ positioned on half of a hemisphere. Combining all elements, we obtain the inpainting feature $\\Phi^{\\mathrm{inp}}$ as", + "bbox": [ + 496, + 425, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi^ {\\text {i n p}} = \\operatorname {M L P} \\left(W ^ {\\text {i n p} ^ {*}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 661, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W^{\\mathrm{inp}^*} = \\arg \\min_{W}\\mathcal{L}^{\\mathrm{inp - sem}} + \\lambda \\mathcal{L}^{\\mathrm{inp - reg}}$", + "bbox": [ + 540, + 683, + 849, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recombination. We make blend features in the volumetric space by summing up id and hair features, weighting them using binarized semantic maps representing the hair. Subsequently, the blend features are expected to be aggregated to 2D features using (2). However, the blend features contain missing regions that are occluded by the hair in the id image but are not fully covered by the hair in the hair image, leading to missing holes in the final rendered image. To mitigate this, we fill the missing regions within the blend features $\\Phi^{\\mathrm{blend}}$ by the inpainting features $\\Phi^{\\mathrm{inp}}$ , as obtained by (5). This blending concept is motivated by the Barbershop [50], but ours differs in that the features to be blended", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2452", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e5cc146bdbbd42b2defd699421d655b46616debcfecedc843cb580a0c8601778.jpg", + "image_caption": [ + "Figure 5: Target semantics generation for occlusion disclosure. The target semantics can be assumed to an ideal semantics for the hairstyle transferred result. The inpainting latent $W^{\\mathrm{inp}}$ is optimized to minimize to match the 2D semantics rendered from $s^{\\mathrm{inp}}$ at a random viewpoint at every iteration." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 467, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lie in the volumetric space.", + "bbox": [ + 76, + 383, + 256, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $\\phi^{\\mathrm{id}}$ and $\\phi^{\\mathrm{hair}}$ denote rendered 2D features of id and hair, respectively, rendered using neural rendering without ray deformation. Meanwhile, $\\phi^{\\mathrm{id}'}$ and $\\phi^{\\mathrm{hair}'}$ denote the registered 2D features of id and hair, aligned to their respective counterparts. Then we recombine features for blending as:", + "bbox": [ + 75, + 410, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\phi_ {i d} ^ {\\text {b l e n d}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}} \\circ \\phi^ {\\text {h a i r} ^ {\\prime}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}}\\right) \\\\ \\circ \\left(\\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ \\phi^ {\\mathrm {i d}} + \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}} \\circ \\phi^ {\\mathrm {i n p}}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 535, + 468, + 593 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ \\phi^ {\\text {h a i r}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}}\\right) \\\\ \\circ \\left(\\left(1 - \\pmb {m} _ {h a i r} ^ {\\mathrm {i d} ^ {\\prime}}\\right) \\circ \\phi^ {\\mathrm {i d} ^ {\\prime}} + \\pmb {m} _ {h a i r} ^ {\\mathrm {i d} ^ {\\prime}} \\circ \\phi^ {\\mathrm {i n p} ^ {\\prime}}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 590, + 439, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m_{hair}$ denotes a 2D binary mask for the hair semantics of $\\phi$ , associated with a corresponding superscript.", + "bbox": [ + 76, + 670, + 467, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Blending", + "text_level": 1, + "bbox": [ + 76, + 744, + 181, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alignment-Aware Blending Loss. In the blending step, the upsampler blocks in the StyleNeRF [15] decode the blend features obtained by (6) into an output image in high resolution, conditioned on a latent vector $W^{\\text{blend}}$ . We optimize the latent vector $W^{\\text{blend}}$ to ensure that the decoded image retains the facial identity from the id image while seamlessly transferring hairstyle of the hair image. To this end, we design the objective function for the optimization to", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "include face-preserving and hair-preserving terms.", + "bbox": [ + 500, + 90, + 833, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {f a c e}} ^ {\\mathrm {l p i p s}} = \\mathcal {L} ^ {\\mathrm {l p i p s}} \\left(\\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\mathrm {i d}} ^ {\\text {b l e n d}}\\right), \\right. \\\\ \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ I ^ {\\mathrm {i d}}\\bigg), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 114, + 831, + 152 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {h a i r}} ^ {\\mathrm {l p i p s}} = \\mathcal {L} ^ {\\mathrm {l p i p s}} \\left(\\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}}\\right), \\right. \\\\ \\left. \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ I ^ {\\text {h a i r}}\\right), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 154, + 890, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {h a i r} ^ {\\text {s t y l e}} = \\left\\| G \\left(\\boldsymbol {m} _ {h a i r} ^ {\\text {h a i r}} \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}}\\right)\\right) \\right. \\\\ - \\left. G \\left(\\boldsymbol {m} _ {h a i r} ^ {\\text {h a i r}} \\circ I ^ {\\text {h a i r}}\\right) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 193, + 790, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m_{face}^{\\mathrm{id}}$ denotes the binary mask for face region of the face image, $m_{hair}^{\\mathrm{hair}}$ denotes the binary mask for hair region of the hair image, $U_W$ denotes the upsampler conditioned on $W$ , $\\mathcal{L}^{\\mathrm{lpips}}$ denotes the perceptual loss [49], and $G$ denotes the Gram matrix for the style loss [13]. Combining all elements, the optimal blend vector $W^{\\mathrm{blend*}}$ is obtained as", + "bbox": [ + 496, + 253, + 890, + 347 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {\\text {b l e n d} *} = \\arg \\min _ {W} \\lambda_ {f a c e} \\mathcal {L} _ {f a c e} ^ {\\mathrm {l p i p s}} + \\lambda_ {h a i r} \\mathcal {L} _ {h a i r} ^ {\\mathrm {l p i p s}} + \\mathcal {L} _ {h a i r} ^ {\\text {s t y l e}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 354, + 890, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 387, + 705, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on two distinct tasks following [24]: hairstyle transfer and reconstruction. In the hairstyle transfer task, each input pair consists of different identities, and the goal is to transfer the hairstyle from the hair image to the face of the id image. In the reconstruction task, the input pair shares the same identity but with different poses. Since the id and hair images share the same identity in the reconstruction task, the id serves as the ground-truth for the evaluation. We use Flicker-Faces-HQ (FFHQ) dataset [21] and faces images crawled from Unsplash (UF), as used in [50, 51], for the hairstyle transfer task. For the reconstruction task, we employ the VoxCeleb2 [9] dataset. To evaluate the overall synthesis quality of the generated images, we measure the Frèchet Inception distance (FID) [17], the naturalness image quality estimator (NIQE) [34], and the precision and recall [26]. In the hairstyle transfer task, we synthesize 1,000 images using the FFHQ dataset and 380 images using the Unsplash dataset. For the reconstruction task, we synthesize 500 images. To demonstrate the effectiveness of HairNeRF at preserving geometric structure, we provide both qualitative and quantitative comparisons of our method against other baselines.", + "bbox": [ + 496, + 412, + 892, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison Study", + "text_level": 1, + "bbox": [ + 500, + 753, + 681, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method to LOHO [37], the Barbershop [50], and StyleYourHair [24] on the FFHQ and VoxCeleb2 datasets. For the hairstyle transfer task, we include a comparison with HairNet [51] using only the UF dataset.", + "bbox": [ + 496, + 777, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hairstyle Transfer. Quantitative results on FFHQ and UF are summarized in Table 1. Notice that LOHO, the Barbershop and StyleYourHair is built upon StyleGAN2, while HairNeRF utilizes StyleNeRF. Both StyleGAN2 and", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2453", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3fe5926f25d98469704a9a656e80500a8747a43b5a665825646f2f5e2fc4e7ea.jpg", + "image_caption": [ + "Figure 6: Comparison of results for FFHQ datasets. Both StyleYourHair and our method align the hair even if the poses between the input images are different, but our method preserves the direction of the hair better (first and the last rows)." + ], + "image_footnote": [], + "bbox": [ + 120, + 88, + 478, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d63ce7d1cc4a9afd84affb13c6e669f713a6e79a6cc22da133745380870906ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 88, + 851, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0af044fbc8e4bddaed61aa8951135517bda95d8266ebe532ae3c9045a43380c1.jpg", + "image_caption": [ + "Figure 7: Comparison of results for Unsplash datasets. Our method preserves identity better (second row on the left), and hair transfer is performed to fit the structure of identity (third row on the left)." + ], + "image_footnote": [], + "bbox": [ + 81, + 491, + 486, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2caad9fef51e9982cada4fb16a2b99a54746d738b11b589b273813d2f55f342.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 491, + 888, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "StyleNeRF are effective in generating high-quality face images; however, it has been reported that StyleGANs slightly outperforms NeRF-based model in terms of commonly used metrics, including FID. Consequently, the overall metrics", + "bbox": [ + 75, + 827, + 470, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for HairNeRF are slightly lower than the other baselines, with the exception of NIQE. Nevertheless, this difference is not substantial, and the qualitative results of HairNeRF reveal perceptually better quality in comparison to the other", + "bbox": [ + 498, + 827, + 892, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2454", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c972c4f3bb642c90a89432d613ec0161794f6087416e657b2f26dce4779f16e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FFHQ
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO40.650.830.850.551.3491.42
Barbershop12.240.770.930.831.0648.00
StyleYourHair12.000.550.930.771.1151.36
HairNet11.630.760.970.591.1456.45
HairNeRF9.570.700.890.671.4057.26
", + "bbox": [ + 86, + 88, + 459, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c287d9fc9c7556f4da607cf8019346916f4a36508687b38df097ec61d6f11733.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UF
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO42.370.800.910.451.0786.16
Barbershop12.130.740.940.841.9141.06
StyleYourHair12.070.750.940.802.1742.62
HairNeRF10.090.700.890.571.7149.74
", + "bbox": [ + 86, + 172, + 459, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6d6212eb26c1f434a103602fdc28644764c7b5a47dcafd432e02198a6be4610.jpg", + "image_caption": [ + "Figure 8: Comparison examples for reconstruction task." + ], + "image_footnote": [], + "bbox": [ + 84, + 321, + 464, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "basielines. Figures 6 and 7 show synthesized examples using the FFHQ and UF datasets, respectively. As observed in the fifth row of Figure 6b and Figure 7, the Barbershop generates fine-quality images when the poses of the id and hair images are similar. However, the output quality substantially declines when the difference in pose between the two input images increases. As shown in the second and the third row of Figure 6a, substantial structural artifacts, such as bumps and missing holes, appear in the transferred hairstyle images synthesized by the Barbershop. In contrast, the transferred hairstyle of HairNeRF seamlessly fits the head of the id image. StyleYourHair and HairNet also generate satisfactory results for input pairs with significant pose differences. However, as seen in the fifth row of Figure 6b, the detailed structure of the hairstyle often changes in the synthesized results. Furthermore, they often elongate the length of the forehead, as shown in the fourth row of Figure 7a.", + "bbox": [ + 75, + 503, + 472, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstruction. Following [24], we evaluate reconstruction quality on the VoxCeleb2 dataset, consisting of pairs of images with same identity but different poses. Figure 8 shows an example of reconstruction task. Ideally, the synthesis result should be identical to the id image, since the id and hair images represent the same person. Figure 8 shows the reconstruction results of HairNeRF in comparison to other baselines. While the reconstructed image pre", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed15583d18a57302418099fda2f384602f4a9f2c1d739906f36c76d4659f6ba5.jpg", + "table_caption": [ + "Table 1: Hairstyle transfer results on FFHQ and UF datasets. HairNeRF shows comparable results to StyleGAN2-based hairstyle transfer models." + ], + "table_footnote": [], + "table_body": "
MethodMSE↓SSIM↑LPIPS↓PSNR↑
Barbershop [50]1.1650.8010.12731.44
StyleYourHair [24]1.2180.7950.15531.34
HairNeRF0.9570.8210.10631.84
", + "bbox": [ + 519, + 88, + 874, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Reconstruction result. HairNeRF surpasses other baselines in view of MSE, SSIM, LPIPS, and PSNR. The result verifies the effectiveness of HairNeRF at preserving hairstyle geometry compared to the other baselines.", + "bbox": [ + 498, + 169, + 890, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc8b20d7421c491836d3bd1672716ca5a40a55ae97739a9d793f871fd360a977.jpg", + "image_caption": [ + "Figure 9: Examples of novel view synthesis." + ], + "image_footnote": [], + "bbox": [ + 506, + 258, + 888, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "serves the detailed geometry of the hair, both the Barbershop and StyleYourHair show slight alterations. The quantitative results, summarized in Table 2, further validate the superiority of HairNeRF in preserving the geometry of hairstyles.", + "bbox": [ + 496, + 616, + 893, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Novel View Synthesis. Owing to the geometric nature of neural rendering, HairNeRF can generate diverse pose varying images from the same latent vector once optimized for synthesis. Figure 9 shows examples of novel view images obtained by varying the camera parameters. For a qualitative comparison, we generate a diverse set of novel view images of the id image using the inversion result of HairNeRF. We then synthesize each novel view image using the Barbershop and StyleYourHair from the given hair image. As shown in Figure 9, the transferred hairstyle varies by views in the case of baselines, whereas HairNeRF preserves a consistent hair structure that aligns with the original hair image.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2455", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5145b4a2ec1b267bee3fe20905dd1ad1779f33069e022f74859f69647c2f5951.jpg", + "image_caption": [ + "Figure 10: Effectiveness of registration shown with the underlying geometry. Without registration the transferred hairstyle does not perfectly fit to the face (first row) or erode the head (third row) of the id image." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 460, + 320 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cd0cf3d732a7902aecdbf1de8fad9a0b5832a531e0cbd9209fa23ddb57322558.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMSE↓SSIM↑LPIPS↓PSNR↑
w/o regi.0.9800.8180.10831.77
w/ regi.0.9570.8210.10631.84
", + "bbox": [ + 122, + 414, + 424, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Reconstruction quality with and without head alignment module. With registration, the reconstruction quality increases in view of MSE, SSIM, LIPIS, and PSNR.", + "bbox": [ + 76, + 483, + 468, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 566, + 228, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effectiveness of Registration. To validate the effectiveness of the head alignment module, we evaluate the reconstruction quality using the VoxCeleb2 dataset. Table 3 summarizes the results. As mentioned in Section 3.3, StyleNERF suffers from an intrinsic misalignment problem that results in subtle differences, even if the pair of input images shares the same identity. However, the registration process within the head alignment module reduces the existing geometric differences to make transferred hairstyle more fit to the id image. Table 3 shows that the reconstruction quality deteriorates when the registration module is missing. Figure 10 shows the effectiveness of registration with the underlying geometry. The first example shows that the shape of the original hair is slightly larger than the head of id image. Without registration, the transferred hair sticks out to the right side of the head of the id. However, with registration, the transferred hair fits more closely to the head of the id. The difference is clearly shown in the overlap image. In the second example, it is clearly shown that the transferred hair is stretched to fit on the head of id by registration.", + "bbox": [ + 76, + 598, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 89, + 619, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduce HairNeRF, a novel geometry-aware hairstyle transfer method designed to address the entanglement issue of pose and style attributes within the latent space of StyleGAN. By leveraging NeRF-based generative models, where the pose of the rendered image is determined by external camera parameters, we achieve a complete disentanglement of pose from the style latent space. Additionally, we propose an efficient method for recombining features within the volumetric space using semantics. This is achieved by aligning two distinct innate head structures via ray deformation. This method unveils the potential to seamlessly integrate multiple NeRF-represented objects of varying shapes and sizes into one unified object, without requiring additional modification to NeRF weights. Using the intrinsic geometric properties of neural rendering, our method can produce view-varied images from a single synthesis process without inducing geometric deformation. The efficacy of HairNeRF in preserving the geometric structure of the original hairstyle, compared to other baselines, is substantiated both qualitatively and quantitatively through comprehensive experiments on diverse datasets.", + "bbox": [ + 496, + 114, + 893, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 443, + 596, + 458 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In SIGGRAPH, 2022. 2", + "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Int. Conf. Comput. Vis., 2021. 2", + "[3] Guha Balakrishnan, Raghudeep Gadde, Aleix Martinez, and Pietro Perona. Rayleigh eigendirections (reds): Nonlinear GAN latent space traversals for multidimensional features. In Eur. Conf. Comput. Vis., 2022. 2", + "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay andgu2021 stylenerf Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[5] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3", + "[6] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. ACM Trans. Graph., 41(1):1:1-1:26, 2022. 3", + "[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In Eur. Conf. Comput. Vis., 2022. 2", + "[8] Min Jin Chong, Wen-Sheng Chu, Abhishek Kumar, and David A. Forsyth. Retrieve in style: Unsupervised facial feature transfer and retrieval. In Int. Conf. Comput. Vis., 2021. 2" + ], + "bbox": [ + 506, + 468, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2456", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In B. Yegnanarayana, editor, Conf. of the Int. Speech Comm. Assoc., 2018. 6", + "[10] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. GRAM: generative radiance manifolds for 3d-aware image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[11] Jean Duchon. Splines minimizing rotation-invariant semi-norms in sobolev spaces. In Walter Schempp and Karl Zeller, editors, Constructive Theory of Functions of Several Variables: Proceedings of a Conference Held at Oberwolfach, Germany, April 25 - May 1, 1976, volume 571 of Lecture Notes in Mathematics, pages 85-100, 1976. 4", + "[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph., 41(4):141:1-141:13, 2022. 2", + "[13] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 5, 6", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[15] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2, 3, 4, 5, 6", + "[16] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Adv. Neural Inform. Process. Syst., 2017. 6", + "[18] Kaiwen Jiang, Shu-Yu Chen, Feng-Lin Liu, Hongbo Fu, and Lin Gao. Nerffaceediting: Disentangled face editing in neural radiance fields. In SIGGRAPH Asia, 2022. 3", + "[19] Omer Kafri, Or Patashnik, Yuval Alaluf, and Daniel Cohen-Or. Stylefusion: A generative model for disentangling spatial segments. ACM Trans. Graph., 41(5):197:1-197:15, 2022. 2", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2", + "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 2, 6", + "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2", + "[23] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in GAN for real-time image editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 852-861, 2021. 2" + ], + "bbox": [ + 80, + 92, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. In Eur. Conf. Comput. Vis., 2022. 2, 3, 6, 8", + "[25] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Adv. Neural Inform. Process. Syst., 2022. 3", + "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Adv. Neural Inform. Process. Syst., 2019. 6", + "[27] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Winter Conf. on Appl. of Comput. Vis., 2023. 3", + "[28] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194:1-194:17, 2017. 4", + "[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. pages 16331–16345, 2021. 2", + "[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Int. Conf. Comput. Vis., 2021. 3", + "[31] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriya Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4", + "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2", + "[33] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2022. 1, 2", + "[34] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Sign. Process. Letters, 20(3):209-212, 2013. 6", + "[35] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3", + "[36] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 42(1):1-13, 2022. 4", + "[37] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 6", + "[38] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Adv. Neural Inform. Process. Syst., 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2457", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. IDE-3D: interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph., 41(6):270:1-270:10, 2022. 3", + "[40] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. Graph., 40(4):133:1-133:14, 2021. 2", + "[42] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3121-3138, 2023. 2", + "[43] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. GRAM-HD: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv, 2022. 3", + "[44] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for edible scene rendering. In Int. Conf. Comput. Vis., 2021. 3", + "[45] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet V2: bilateral network with guided aggregation for real-time semantic segmentation. Int. J. Comput. Vis., 129(11):3051-3068, 2021. 4", + "[46] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4", + "[47] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4", + "[48] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: Geometry editing of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6", + "[50] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Trans. Graph., 40(6):215:1-215:13, 2021. 2, 3, 5, 6, 8", + "[51] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Eur. Conf. Comput. Vis., 2022. 2, 6" + ], + "bbox": [ + 78, + 90, + 468, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2458", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_model.json b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e6a1daceee5c40e23be437e4557adb35d4ca7423 --- /dev/null +++ b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_model.json @@ -0,0 +1,2004 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "HairNeRF: Geometry-Aware Image Synthesis for Hairstyle Transfer" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.181, + 0.386, + 0.216 + ], + "angle": 0, + "content": "Seunggyu Chang* \nNAVER Cloud" + }, + { + "type": "text", + "bbox": [ + 0.447, + 0.181, + 0.557, + 0.216 + ], + "angle": 0, + "content": "Gihoon Kim† \nKAIST" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.181, + 0.73, + 0.216 + ], + "angle": 0, + "content": "Hayeon Kim† \nUNIST" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.269, + 0.143, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.269, + 0.276, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.269, + 0.891, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.568, + 0.894, + 0.611 + ], + "angle": 0, + "content": "Figure 1: Given portrait images of hairstyle and identity (thunnails) that differ in viewing angle, head shape, and size, we seamlessly transfer hairstyle from one image to another utilizing neural rendering. Our method preserves fine-level details of the hairstyle structural geometry and generates diverse range of pose variant transferred images without requiring additional transfer process." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.623, + 0.314, + 0.639 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.473, + 0.866 + ], + "angle": 0, + "content": "We propose a novel hairstyle transferred image synthesis method considering the underlying head geometry of two input images. In traditional GAN-based methods, transferring hairstyle from one image to the other often makes the synthesized result awkward due to differences in pose, shape, and size of heads. To resolve this, we utilize neural rendering by registering two input heads in the volumetric space to make a transferred hairstyle fit on the head of a target image. Because of the geometric nature of neural rendering, our method can render view varying images of synthesized results from a single transfer process without causing distortion from which extant hairstyle transfer methods built upon traditional GAN-based generators suffer. We verify that our method surpasses other baselines in view of pre" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.625, + 0.894, + 0.671 + ], + "angle": 0, + "content": "serving the identity and hairstyle of two input images when synthesizing a hairstyle transferred image rendered at any point of view." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.707, + 0.633, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Selecting an appropriate hairstyle can be difficult as it is hard to imagine how it will look on your face without trying it out for yourself. The reason can be attributed to the inherent variability of individual head shape, resulting in the frequent occurrence of disparate outcomes when attempting to emulate celebrity hairstyles. To address this issue, we propose a novel image synthesis method for hairstyle transfer tailored to fit an individual head shape. We named our method HairNeRF as it considers geometric alignment in the volumetric space of neural rendering [33]. Hairstyle transfer synthesizes an image where the hairstyle from one" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.448, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author: seunggyu.chang@navercorp.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.401, + 0.9 + ], + "angle": 0, + "content": "This work was done during internship at Naver Cloud." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.448, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2448" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.408 + ], + "angle": 0, + "content": "portrait is transferred to the other and Generative adversarial networks (GANs) [14, 20, 22] have been extensively utilized to address this problem. Notably, following the emergence of StyleGAN [20, 21, 22], which enables the generation of highly realistic images, studies [37, 50, 19, 8] have been made to optimize latent vectors within the well-trained latent space of StyleGAN to generate an optimal synthesis outcome. These methods are successful at transferring hairstyles between images having similar poses, but struggle to generate convincing results for images having vastly different poses. To resolve this, studies [50, 24, 51] have proposed methods aligning pose of transferred hairstyle to a target portrait, simultaneously reducing the perceptual difference between the transferred result and the original hairstyle. However, they inherently encounter the same limitation of the pre-trained latent space of StyleGAN, where pose and style attributes are entangled as StyleGAN is not explicitly trained for hairstyle transfer. Consequently, the alignment process may distort the geometry of an original hairstyle, leading to undesirable alterations such as forehead stretching or shrinking." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.418, + 0.468, + 0.9 + ], + "angle": 0, + "content": "We categorize entanglement of pose and style attributes into two folds: entanglement of style with viewpoint, and entanglement of style with head shape. To address the entanglement of style with viewpoint, we leverage the latent space of StyleNeRF [15], a neural radiance field (NeRF) [32, 33] based GAN model as an alternative for StyleGAN. StyleNeRF exhibits an innate pose-invariant latent space as the network implies the volumetric geometry of an image. We basically follow the latent optimization technique following the literature [50, 24, 51] utilizing StyleGAN. Leveraging the pose-invariant nature of the StyleNeRF's latent space, we extract a pose-invariant style vector through GAN inversion [42]. Thereafter, we recombine the volumetric features generated by the inverted style vectors of the hair and the target portraits. However, despite the pose-invariant nature of StyleNeRF's latent space, it still entails an entanglement problem between style and head shape. The optimization of a style vector to fit on the head of a target portrait, while retaining hairstyle of a hair portrait, can potentially result in deformations of the hairstyle. To address the entanglement of style with head shape, we propose a registration method that aligns the geometries of two heads within the volumetric space, while maintaining the inverted style vectors intact. Subsequently, we introduce a novel rendering method that recombines hairstyle features with facial features within the volumetric space using alignment-inducing deformable rays, which are derived from the registration outcome. In this manner, we resolve the distorted hairstyle problem from which StyleGAN-based method suffers, resulting in better alignment of the transferred hairstyle seamlessly fit on the target portrait. Taking the advantage of geometric nature of neural rendering, HairNeRF is ca" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.182 + ], + "angle": 0, + "content": "pable of generating a diverse set of geometrically consistent novel view hairstyle transferred images from a single synthesis, by merely adjusting camera parameters. This is in contrast to existing StyleGAN-based hairstyle transfer methods which necessitate separate optimizations for each of novel views, which often yield inconsistent outcomes." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.183, + 0.891, + 0.258 + ], + "angle": 0, + "content": "Experimental results on diverse datasets demonstrate the superiority of HairNeRF against other baselines at generating seamless hairstyle transfer outcomes, while effectively preserving the underlying geometry of the original hairstyle." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.273, + 0.64, + 0.289 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.298, + 0.677, + 0.314 + ], + "angle": 0, + "content": "2.1. Hairstyle Transfer" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.323, + 0.891, + 0.714 + ], + "angle": 0, + "content": "The proficiency exhibited by generative adversarial networks (GANs) [14, 20, 22] in generating photorealistic images has propelled a surge of research addressing image editing problem. A significant amount of research [3, 7, 19, 16, 12] has effectively addressed attribute editing or style transfer, which modify abstract-level features of an image. However, the task of hairstyle transfer still presents a challenging problem that necessitates the preservation of intricate structure and color details of a hairstyle from a source image. With the emergence of StyleGAN [20, 21, 22], which demonstrates exceptional quality in generating photorealistic facial images, efforts have been made to discover latent vectors capable of capturing transferred images of hairstyles [37, 50]. These methods require the projection of an image into the pretrained StyleGAN's latent space known as inversion, which influences the quality of transfer. Thus, the Barbershop [50] introduces an improved inversion method on the enlarged latent space named \\( FS \\)-space to capture finer details of input image. However, since these models lack awareness of the pose of given images, they suffer structural artifacts when the given two images largely differ in poses. Recent studies [24, 51] resolve this issue by aligning pose of the transferred hairstyle to the target image. However, they still suffer style deformation problem arising from the entangled latent space of StyleGAN for pose and style." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.726, + 0.835, + 0.742 + ], + "angle": 0, + "content": "2.2. 3D-Aware Image Synthesis and Editing" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.75, + 0.891, + 0.9 + ], + "angle": 0, + "content": "The emergence of StyleGAN [20, 21, 22] has facilitated researches on image editing [16, 41, 2, 19, 29, 23, 1], either by navigating the latent style space or leveraging semantic maps. However, StyleGAN-based image editing methods have shown inconsistency among pose changes, which is attributed to the entanglement of style and pose within StyleGAN's latent space. The development of neural radiance fields (NeRF) [32, 33] has enabled the rendering of 3D-consistent images that accommodate varying viewpoints. Consequently, methods for editing a 3D scene" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "2449" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.893, + 0.365 + ], + "angle": 0, + "content": "Figure 2: Overview of the hairstyle transfer pipeline of HairNeRF consisting of five modules: (a) inversion, (b) head alignment, (c) occlusion disclosure, (d) recombination, and (e) blending." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.679 + ], + "angle": 0, + "content": "[44, 30, 27, 48] such as the addition and subtraction of NeRF models trained on individual objects have been explored. However, these approaches do not support modifications, such as altering the attributes or style of the object, due to the inherent nature of NeRF being trained to solely represent a single entity, thus lacking the necessary information for the desired edits. The advent of generative NeRFs [5, 38, 35, 15, 4, 10, 43] has enhanced the representational capabilities of NeRF models, enabling the expression of an expansive range of objects through the exploration of the latent space within a single model. The enhanced expressive capabilities of generative NeRFs enable the modification of attributes and style of an object [15, 4, 10]. Furthermore, within the generative NeRF modules, methods [30, 18, 40, 39, 6] have been proposed that disentangle the semantic structure and texture. However, these methods are not well-suited for local style modifications, as the latent vector representing the texture exhibits spatial entanglement, encompassing the entire regions of the image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.169, + 0.713 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.188, + 0.739 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.901 + ], + "angle": 0, + "content": "We follow the basic synthesis pipeline from the Barbershop [50], a founding work for hairstyle transfer [24, 50]. We encourage readers to refer [50] for the details of the original pipeline. The overall synthesis pipeline of our method consists of four modules: inversion, head alignment, recombination, and blending, as illustrated in Figure 2. Given two images, id and hair, we transfer hairstyle from the hair image into the id image. In the inversion module, we first invert the two images to obtain latent vectors and the corresponding features in the volu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.768 + ], + "angle": 0, + "content": "metric space. Then, in the alignment module, we conduct registration on two geometric heads of the inverted images in the volumetric space. By aligning geometric heads, we can transfer hairstyle of the hair image into the id image to be more fit. Once the registration is done, we recombine the 3D features in the volumetric space according to semantics to render a 2D feature for synthesis using neural rendering in the recombination module. A simple method is to bring hair semantic features from the hair image, and other semantics from the id image. However, there occur missing regions which are occluded by the hair in the id image but not covered by the hair of the hair image. To fill in the missing regions, we inpaint 3D features by optimizing a latent vector, ensuring that the 2D features rendered from the 3D features generated by the optimized latent vector align with the target semantics across all viewpoints. Thereafter, we semantically integrate the 3D features from the id, hair, and the inpainting output to form a blended feature for the synthesis. The blend feature is rendered to a 2D feature, which is decoded by the upsampler blocks of StyleNeRF [15] conditioned on the blending style vector. In the final blending module, the blending style vector is optimized to make the decoded image preserve the same hairstyle as in the hair image while retaining face of the id image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.659, + 0.8 + ], + "angle": 0, + "content": "3.2. NeRF Inversion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Semantic Parsing NeRF Head. To semantically recombine features in the 3D volume, semantics of volumetric features are required. It is known that a semantic classifier trained to classify a 2D point feature obtained by (2) can aptly classify semantic of a 3D point feature as well [25]. Based on this observation, we add an additional seman-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2450" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.454, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.214, + 0.471, + 0.284 + ], + "angle": 0, + "content": "Figure 3: The additional semantic head is added at the end of a MLP block of NeRF module, which outputs a logit of semantics in the volumetric space. The additional semantic head is trained to minimize cross-entropy between the rendered logit, \\(\\hat{s}\\), and the ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.47, + 0.392 + ], + "angle": 0, + "content": "tic branch to the MLP block for neural rendering and train it using outputs of the off-the-shelf semantic segmentation model [46] for generated images as ground-truths as shown in Figure 3. After training, the trained semantic branch outputs a logit of semantics for a 3D point feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.684 + ], + "angle": 0, + "content": "Semantics-Aware Inversion in 3D. We invert the hair and the id images into the latent vectors for the synthesis network of the StyleNeRF [15]. The process of obtaining the latent vector generating a desired image is called GAN inversion. We utilize the pivotal tuning inversion (PTI) [36] with a little modification for inversion as the PTI can reconstruct fine details of an input image from the inverted latent vector. As our model is built upon the StyleNeRF, the output image varies as a camera parameter changes. For this reason, the inversion process requires an input camera parameter or should estimate a camera parameter in our case. We choose to provide a fixed camera parameter of an input image estimated by the off-the-shelf head pose estimator [31] for the inversion process. Using the PTI with a fixed camera parameter, we optimize \\( W \\in \\mathcal{W} \\) for each of the id and the hair images, and finetune the weights of the pretrained StyleNeRF to reconstruct the input image and the semantics estimated by the off-the-shelf semantic parsing networks [47, 45]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.24, + 0.717 + ], + "angle": 0, + "content": "3.3. Head Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.849 + ], + "angle": 0, + "content": "Registration For registration, we render each of inverted images into \\( n_c \\) different viewpoints and extract 2D keypoints of the FLAME [28] head model using the off-the-shelf head pose estimator [31]. Then we estimate 3D keypoints of the FLAME model for each inverted image from the set of 2D keypoints estimated from the eight different camera views by solving a bundle-adjustment-like least square problem:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.865, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {p}} {\\operatorname {m i n i m i z e}} \\sum_ {i} ^ {n _ {c}} \\| P (\\boldsymbol {c} _ {i}, \\boldsymbol {p}) - \\boldsymbol {x} _ {i} \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.213 + ], + "angle": 0, + "content": "where \\( \\pmb{x}_i \\in \\mathbb{R}^2 \\) denotes the estimated 2D keypoints at \\( i \\)-th viewpoint, \\( \\pmb{c}_i \\) denotes the \\( i \\)-th camera parameter, and \\( P(\\pmb{c}_i, \\pmb{p}_i) \\) denotes the projection function projecting a 3D point \\( \\pmb{p}_i \\in \\mathbb{R}^3 \\) into 2D using \\( \\pmb{c}_i \\). As we already know the camera parameters used for rendering \\( n_c \\) images, we exclude camera parameters from the optimization variables. We utilize the thin-plate-spline (TPS) transform [11] to find the optimal non-rigid transformation for the registration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.893, + 0.472 + ], + "angle": 0, + "content": "Neural Rendering via Ray Deformation. The two heads of the id and the hair do not fully overlap due to the differences in characteristics, shape and size of the heads geometry, along with estimation errors in head poses. We refer to this misalignment as intrinsic misalignment, as it arises from factors unrelated to external elements such as camera parameters. To alleviate the intrinsic misalignment, the registration process finds an optimal transformation between two point sets, by which one point set is transformed to align with the other. During the neural rendering process, point features are sampled along a ray within the volumetric space, and these features are then aggregated to a single point on the image plane. If the sampling points are collected along the ray, deformed by the transformation determined through registration, we can obtain registered point features. This process leads to a registered 2d image after the aggregation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.744 + ], + "angle": 0, + "content": "Figure 4 illustrates the effective neural rendering process for alignment via ray deformation. Let \\( T_{i\\rightarrow h} \\) denote a transformation determined through registration, which aligns the point set of the id features to the point set of the hair features. As depicted in Figure 4a, sampling along rays that radiate straight from the id's camera renders an id's face with an innate head shape, colored in red, as in shown Figure 4c. Likewise, rendering from straight rays radiating through hair features \\( \\Phi^{\\text{hair}} \\) using the same camera renders the hair image posed as id, with the overlaid innate head shape colored in blue, as depicted in Figure 4e. The two heads do not overlap due to differences in shape and size. However, if we render a hair image using the deformed ray transformed by \\( T_{i\\rightarrow h} \\), as depicted in Figure 4b, we obtain a registered hair image aligned to the head of id, as shown in Figure 4d. With a camera ray \\( r = o + td \\), the registered neural rendering computes a pixel value of the ray as:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.75, + 0.891, + 0.787 + ], + "angle": 0, + "content": "\\[\nI _ {\\omega} ^ {\\text {R e g i s t e r e d}} (\\boldsymbol {r}) = \\int_ {0} ^ {\\infty} p _ {\\omega} (t) \\boldsymbol {c} _ {\\omega} (\\tilde {\\boldsymbol {r}} (t), \\tilde {\\boldsymbol {d}}) \\left| \\frac {d \\tilde {\\boldsymbol {r}}}{d t} \\right| d t, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.782, + 0.862, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} p _ {\\omega} (t) = \\exp \\left(- \\int_ {0} ^ {t} \\sigma_ {\\omega} (\\tilde {\\boldsymbol {r}} (s)) \\left| \\frac {d \\tilde {\\boldsymbol {r}}}{d s} \\right| d s\\right) \\cdot \\sigma_ {\\omega} (\\tilde {\\boldsymbol {r}} (t)),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.819, + 0.892, + 0.88 + ], + "angle": 0, + "content": "with a registered ray \\(\\tilde{\\boldsymbol{r}} = T_{i\\rightarrow h}(\\boldsymbol {r})\\) , and the adjusted directional vector \\(\\tilde{\\pmb{d}} = d\\tilde{\\pmb{r}} /dt\\) .Equation (2) has the same form as the original neural rendering, but with the integration computed along the transformed ray. This adjustment" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.804, + 0.901 + ], + "angle": 0, + "content": "1 or a registered 2d feature map in the StyleNeRF [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2451" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.09, + 0.526, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.101, + 0.854, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.22, + 0.522, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.227, + 0.851, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.895, + 0.401 + ], + "angle": 0, + "content": "Figure 4: Illustration of alignment-aware neural rendering via ray deformation. Figure (c) - (e), and (h) - (j) show overlaid heads on rendering images. Rendering via transformed ray render a registered image as shown in (d) and (i), which is equivalent to rendering with the straight ray on transformed feature represented as \\(\\Phi^{\\text{hair}'}\\) and \\(\\Phi^{\\text{id}'}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.487 + ], + "angle": 0, + "content": "is achieved by substituting the directional vector \\( \\pmb{d} \\) with \\( \\tilde{\\pmb{d}} \\) and including \\( |d\\tilde{\\pmb{r}} /dt| \\) to adjust the infinitesimal length \\( dt \\). Since \\( |d\\tilde{\\pmb{r}} /dt| \\approx 1 \\) in practice, we omit the determinant term to simplify the computation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.32, + 0.51 + ], + "angle": 0, + "content": "3.4. Volumetric Recombination" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Occlusion Disclosure. To inpaint the missing regions occluded by the hair in \\(\\Phi^{\\mathrm{hair}}\\) and produce the recombined 3D features \\(\\Phi^{\\mathrm{blend}}\\) as depicted in Figure 2c, we optimize the latent vector \\(W^{\\mathrm{inp}} \\in \\mathcal{W}+\\) within the StyleNeRF [15] framework to generate \\(\\Phi^{\\mathrm{inp}}\\). Analogous to [50], we produce a volumetric target semantic mask, as depicted in Figure 5. For each iteration, we optimize \\(W^{\\mathrm{inp}}\\) to generate \\(\\Phi^{\\mathrm{inp}}\\) that aligns with the target semantic mask \\(s^{\\mathrm{inp}}\\) from a randomly selected viewpoint. Let \\(m_{hair}\\) denote a volumetric binary mask representing hair, \\(c\\) denote a camera parameter, and \\(\\phi|_c\\) represents 2D features rendered by the camera parameter \\(c\\). The target semantics \\(s^{\\mathrm{inp}}\\) and the objectives for the optimization are constructed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.727, + 0.47, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {s} ^ {\\text {i n p}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}} \\circ \\boldsymbol {s} ^ {\\text {h a i r} ^ {\\prime}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}}\\right) (3) \\\\ \\circ \\Big (\\left(1 - \\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}}\\right) \\circ \\boldsymbol {s} ^ {\\mathrm {i d}} + \\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}} \\circ \\boldsymbol {s} ^ {\\mathrm {h a i r} ^ {\\prime}} \\Big), \\\\ \\mathcal {L} ^ {\\text {i n p - s e m}} = \\mathbb {E} _ {c} \\mathrm {C E} \\left(\\mathcal {R} _ {c} \\left(s ^ {\\text {i n p}}\\right), \\mathcal {R} _ {c} \\left(f _ {W ^ {\\text {i n p}}} ^ {\\text {s e m}}\\right)\\right), (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.837, + 0.45, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} ^ {\\text {i n p - r e g}} = \\mathbb {E} _ {\\boldsymbol {c}} \\left\\| G \\left(\\left(1 - \\mathcal {R} _ {\\boldsymbol {c}} \\left(\\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {i n p}}\\right)\\right) \\circ U _ {W} \\left(\\phi^ {\\text {i n p}} \\mid_ {\\boldsymbol {c}}\\right)\\right) \\right. \\\\ \\left. - G \\Big (\\big (1 - \\mathcal {R} _ {\\boldsymbol {c} ^ {\\mathrm {i d}}} (\\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}}) \\big) \\circ I ^ {\\mathrm {i d}} \\Big) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.653 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_c\\) represents the aggregation function of neural rendering according to the camera \\(c\\), \\(f^{\\mathrm{sem}}\\) denotes the MLP block with semantic head, \\(U_W\\) refers to the upsampler block of StyleNeRF whose output is a final image, and \\(G\\) denotes the Gram matrix for style loss [13]. Symbols marked with an apostrophe indicate that they are registered to their counterparts, as denoted by the associated superscripts. Unlike [50], our target semantic mask is defined as a continuous function within the volumetric space, rather than as a tensor defined at discrete positions. As a result, \\(\\mathcal{L}^{\\mathrm{inp - sem}}\\) entails integration over this volumetric space. Hence, for each optimization iteration in practice, we compute \\(\\mathcal{L}^{\\mathrm{inp - sem}}\\) using a random set of rays radiating from a randomly sampled camera \\(c\\) positioned on half of a hemisphere. Combining all elements, we obtain the inpainting feature \\(\\Phi^{\\mathrm{inp}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.662, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\Phi^ {\\text {i n p}} = \\operatorname {M L P} \\left(W ^ {\\text {i n p} ^ {*}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.684, + 0.85, + 0.708 + ], + "angle": 0, + "content": "where \\(W^{\\mathrm{inp}^*} = \\arg \\min_{W}\\mathcal{L}^{\\mathrm{inp - sem}} + \\lambda \\mathcal{L}^{\\mathrm{inp - reg}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recombination. We make blend features in the volumetric space by summing up id and hair features, weighting them using binarized semantic maps representing the hair. Subsequently, the blend features are expected to be aggregated to 2D features using (2). However, the blend features contain missing regions that are occluded by the hair in the id image but are not fully covered by the hair in the hair image, leading to missing holes in the final rendered image. To mitigate this, we fill the missing regions within the blend features \\(\\Phi^{\\mathrm{blend}}\\) by the inpainting features \\(\\Phi^{\\mathrm{inp}}\\), as obtained by (5). This blending concept is motivated by the Barbershop [50], but ours differs in that the features to be blended" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2452" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.468, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.336 + ], + "angle": 0, + "content": "Figure 5: Target semantics generation for occlusion disclosure. The target semantics can be assumed to an ideal semantics for the hairstyle transferred result. The inpainting latent \\( W^{\\mathrm{inp}} \\) is optimized to minimize to match the 2D semantics rendered from \\( s^{\\mathrm{inp}} \\) at a random viewpoint at every iteration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.258, + 0.4 + ], + "angle": 0, + "content": "lie in the volumetric space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.502 + ], + "angle": 0, + "content": "Let \\(\\phi^{\\mathrm{id}}\\) and \\(\\phi^{\\mathrm{hair}}\\) denote rendered 2D features of id and hair, respectively, rendered using neural rendering without ray deformation. Meanwhile, \\(\\phi^{\\mathrm{id}'}\\) and \\(\\phi^{\\mathrm{hair}'}\\) denote the registered 2D features of id and hair, aligned to their respective counterparts. Then we recombine features for blending as:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.536, + 0.47, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\phi_ {i d} ^ {\\text {b l e n d}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}} \\circ \\phi^ {\\text {h a i r} ^ {\\prime}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}}\\right) \\\\ \\circ \\left(\\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ \\phi^ {\\mathrm {i d}} + \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}} \\circ \\phi^ {\\mathrm {i n p}}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.591, + 0.44, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ \\phi^ {\\text {h a i r}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}}\\right) \\\\ \\circ \\left(\\left(1 - \\pmb {m} _ {h a i r} ^ {\\mathrm {i d} ^ {\\prime}}\\right) \\circ \\phi^ {\\mathrm {i d} ^ {\\prime}} + \\pmb {m} _ {h a i r} ^ {\\mathrm {i d} ^ {\\prime}} \\circ \\phi^ {\\mathrm {i n p} ^ {\\prime}}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.671, + 0.468, + 0.703 + ], + "angle": 0, + "content": "where \\( m_{hair} \\) denotes a 2D binary mask for the hair semantics of \\( \\phi \\), associated with a corresponding superscript." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.745, + 0.183, + 0.761 + ], + "angle": 0, + "content": "3.5. Blending" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Alignment-Aware Blending Loss. In the blending step, the upsampler blocks in the StyleNeRF [15] decode the blend features obtained by (6) into an output image in high resolution, conditioned on a latent vector \\( W^{\\text{blend}} \\). We optimize the latent vector \\( W^{\\text{blend}} \\) to ensure that the decoded image retains the facial identity from the id image while seamlessly transferring hairstyle of the hair image. To this end, we design the objective function for the optimization to" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.834, + 0.107 + ], + "angle": 0, + "content": "include face-preserving and hair-preserving terms." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.116, + 0.833, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {f a c e}} ^ {\\mathrm {l p i p s}} = \\mathcal {L} ^ {\\mathrm {l p i p s}} \\left(\\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\mathrm {i d}} ^ {\\text {b l e n d}}\\right), \\right. \\\\ \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ I ^ {\\mathrm {i d}}\\bigg), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.155, + 0.892, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {h a i r}} ^ {\\mathrm {l p i p s}} = \\mathcal {L} ^ {\\mathrm {l p i p s}} \\left(\\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}}\\right), \\right. \\\\ \\left. \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ I ^ {\\text {h a i r}}\\right), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.194, + 0.791, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {h a i r} ^ {\\text {s t y l e}} = \\left\\| G \\left(\\boldsymbol {m} _ {h a i r} ^ {\\text {h a i r}} \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}}\\right)\\right) \\right. \\\\ - \\left. G \\left(\\boldsymbol {m} _ {h a i r} ^ {\\text {h a i r}} \\circ I ^ {\\text {h a i r}}\\right) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.348 + ], + "angle": 0, + "content": "where \\(m_{face}^{\\mathrm{id}}\\) denotes the binary mask for face region of the face image, \\(m_{hair}^{\\mathrm{hair}}\\) denotes the binary mask for hair region of the hair image, \\(U_W\\) denotes the upsampler conditioned on \\(W\\), \\(\\mathcal{L}^{\\mathrm{lpips}}\\) denotes the perceptual loss [49], and \\(G\\) denotes the Gram matrix for the style loss [13]. Combining all elements, the optimal blend vector \\(W^{\\mathrm{blend*}}\\) is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.356, + 0.892, + 0.38 + ], + "angle": 0, + "content": "\\[\nW ^ {\\text {b l e n d} *} = \\arg \\min _ {W} \\lambda_ {f a c e} \\mathcal {L} _ {f a c e} ^ {\\mathrm {l p i p s}} + \\lambda_ {h a i r} \\mathcal {L} _ {h a i r} ^ {\\mathrm {l p i p s}} + \\mathcal {L} _ {h a i r} ^ {\\text {s t y l e}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.706, + 0.405 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.746 + ], + "angle": 0, + "content": "We evaluate our method on two distinct tasks following [24]: hairstyle transfer and reconstruction. In the hairstyle transfer task, each input pair consists of different identities, and the goal is to transfer the hairstyle from the hair image to the face of the id image. In the reconstruction task, the input pair shares the same identity but with different poses. Since the id and hair images share the same identity in the reconstruction task, the id serves as the ground-truth for the evaluation. We use Flicker-Faces-HQ (FFHQ) dataset [21] and faces images crawled from Unsplash (UF), as used in [50, 51], for the hairstyle transfer task. For the reconstruction task, we employ the VoxCeleb2 [9] dataset. To evaluate the overall synthesis quality of the generated images, we measure the Frèchet Inception distance (FID) [17], the naturalness image quality estimator (NIQE) [34], and the precision and recall [26]. In the hairstyle transfer task, we synthesize 1,000 images using the FFHQ dataset and 380 images using the Unsplash dataset. For the reconstruction task, we synthesize 500 images. To demonstrate the effectiveness of HairNeRF at preserving geometric structure, we provide both qualitative and quantitative comparisons of our method against other baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.682, + 0.771 + ], + "angle": 0, + "content": "4.1. Comparison Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We compare our method to LOHO [37], the Barbershop [50], and StyleYourHair [24] on the FFHQ and VoxCeleb2 datasets. For the hairstyle transfer task, we include a comparison with HairNet [51] using only the UF dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Hairstyle Transfer. Quantitative results on FFHQ and UF are summarized in Table 1. Notice that LOHO, the Barbershop and StyleYourHair is built upon StyleGAN2, while HairNeRF utilizes StyleNeRF. Both StyleGAN2 and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2453" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.089, + 0.48, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.089, + 0.852, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.451, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Figure 6: Comparison of results for FFHQ datasets. Both StyleYourHair and our method align the hair even if the poses between the input images are different, but our method preserves the direction of the hair better (first and the last rows)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.492, + 0.487, + 0.764 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.492, + 0.889, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.773, + 0.893, + 0.803 + ], + "angle": 0, + "content": "Figure 7: Comparison of results for Unsplash datasets. Our method preserves identity better (second row on the left), and hair transfer is performed to fit the structure of identity (third row on the left)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.828, + 0.471, + 0.889 + ], + "angle": 0, + "content": "StyleNeRF are effective in generating high-quality face images; however, it has been reported that StyleGANs slightly outperforms NeRF-based model in terms of commonly used metrics, including FID. Consequently, the overall metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.828, + 0.893, + 0.889 + ], + "angle": 0, + "content": "for HairNeRF are slightly lower than the other baselines, with the exception of NIQE. Nevertheless, this difference is not substantial, and the qualitative results of HairNeRF reveal perceptually better quality in comparison to the other" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2454" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.46, + 0.166 + ], + "angle": 0, + "content": "
FFHQ
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO40.650.830.850.551.3491.42
Barbershop12.240.770.930.831.0648.00
StyleYourHair12.000.550.930.771.1151.36
HairNet11.630.760.970.591.1456.45
HairNeRF9.570.700.890.671.4057.26
" + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.173, + 0.46, + 0.24 + ], + "angle": 0, + "content": "
UF
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO42.370.800.910.451.0786.16
Barbershop12.130.740.940.841.9141.06
StyleYourHair12.070.750.940.802.1742.62
HairNeRF10.090.700.890.571.7149.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Table 1: Hairstyle transfer results on FFHQ and UF datasets. HairNeRF shows comparable results to StyleGAN2-based hairstyle transfer models." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.322, + 0.465, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.459, + 0.438, + 0.473 + ], + "angle": 0, + "content": "Figure 8: Comparison examples for reconstruction task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.473, + 0.776 + ], + "angle": 0, + "content": "basielines. Figures 6 and 7 show synthesized examples using the FFHQ and UF datasets, respectively. As observed in the fifth row of Figure 6b and Figure 7, the Barbershop generates fine-quality images when the poses of the id and hair images are similar. However, the output quality substantially declines when the difference in pose between the two input images increases. As shown in the second and the third row of Figure 6a, substantial structural artifacts, such as bumps and missing holes, appear in the transferred hairstyle images synthesized by the Barbershop. In contrast, the transferred hairstyle of HairNeRF seamlessly fits the head of the id image. StyleYourHair and HairNet also generate satisfactory results for input pairs with significant pose differences. However, as seen in the fifth row of Figure 6b, the detailed structure of the hairstyle often changes in the synthesized results. Furthermore, they often elongate the length of the forehead, as shown in the fourth row of Figure 7a." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Reconstruction. Following [24], we evaluate reconstruction quality on the VoxCeleb2 dataset, consisting of pairs of images with same identity but different poses. Figure 8 shows an example of reconstruction task. Ideally, the synthesis result should be identical to the id image, since the id and hair images represent the same person. Figure 8 shows the reconstruction results of HairNeRF in comparison to other baselines. While the reconstructed image pre" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.875, + 0.148 + ], + "angle": 0, + "content": "
MethodMSE↓SSIM↑LPIPS↓PSNR↑
Barbershop [50]1.1650.8010.12731.44
StyleYourHair [24]1.2180.7950.15531.34
HairNeRF0.9570.8210.10631.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.17, + 0.892, + 0.226 + ], + "angle": 0, + "content": "Table 2: Reconstruction result. HairNeRF surpasses other baselines in view of MSE, SSIM, LPIPS, and PSNR. The result verifies the effectiveness of HairNeRF at preserving hairstyle geometry compared to the other baselines." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.259, + 0.889, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.555, + 0.828, + 0.57 + ], + "angle": 0, + "content": "Figure 9: Examples of novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.895, + 0.693 + ], + "angle": 0, + "content": "serves the detailed geometry of the hair, both the Barbershop and StyleYourHair show slight alterations. The quantitative results, summarized in Table 2, further validate the superiority of HairNeRF in preserving the geometry of hairstyles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Novel View Synthesis. Owing to the geometric nature of neural rendering, HairNeRF can generate diverse pose varying images from the same latent vector once optimized for synthesis. Figure 9 shows examples of novel view images obtained by varying the camera parameters. For a qualitative comparison, we generate a diverse set of novel view images of the id image using the inversion result of HairNeRF. We then synthesize each novel view image using the Barbershop and StyleYourHair from the given hair image. As shown in Figure 9, the transferred hairstyle varies by views in the case of baselines, whereas HairNeRF preserves a consistent hair structure that aligns with the original hair image." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.462, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.33, + 0.472, + 0.388 + ], + "angle": 0, + "content": "Figure 10: Effectiveness of registration shown with the underlying geometry. Without registration the transferred hairstyle does not perfectly fit to the face (first row) or erode the head (third row) of the id image." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.415, + 0.426, + 0.462 + ], + "angle": 0, + "content": "
MethodMSE↓SSIM↑LPIPS↓PSNR↑
w/o regi.0.9800.8180.10831.77
w/ regi.0.9570.8210.10631.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.484, + 0.47, + 0.526 + ], + "angle": 0, + "content": "Table 3: Reconstruction quality with and without head alignment module. With registration, the reconstruction quality increases in view of MSE, SSIM, LIPIS, and PSNR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.23, + 0.584 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Registration. To validate the effectiveness of the head alignment module, we evaluate the reconstruction quality using the VoxCeleb2 dataset. Table 3 summarizes the results. As mentioned in Section 3.3, StyleNERF suffers from an intrinsic misalignment problem that results in subtle differences, even if the pair of input images shares the same identity. However, the registration process within the head alignment module reduces the existing geometric differences to make transferred hairstyle more fit to the id image. Table 3 shows that the reconstruction quality deteriorates when the registration module is missing. Figure 10 shows the effectiveness of registration with the underlying geometry. The first example shows that the shape of the original hair is slightly larger than the head of id image. Without registration, the transferred hair sticks out to the right side of the head of the id. However, with registration, the transferred hair fits more closely to the head of the id. The difference is clearly shown in the overlap image. In the second example, it is clearly shown that the transferred hair is stretched to fit on the head of id by registration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.62, + 0.106 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.433 + ], + "angle": 0, + "content": "We introduce HairNeRF, a novel geometry-aware hairstyle transfer method designed to address the entanglement issue of pose and style attributes within the latent space of StyleGAN. By leveraging NeRF-based generative models, where the pose of the rendered image is determined by external camera parameters, we achieve a complete disentanglement of pose from the style latent space. Additionally, we propose an efficient method for recombining features within the volumetric space using semantics. This is achieved by aligning two distinct innate head structures via ray deformation. This method unveils the potential to seamlessly integrate multiple NeRF-represented objects of varying shapes and sizes into one unified object, without requiring additional modification to NeRF weights. Using the intrinsic geometric properties of neural rendering, our method can produce view-varied images from a single synthesis process without inducing geometric deformation. The efficacy of HairNeRF in preserving the geometric structure of the original hairstyle, compared to other baselines, is substantiated both qualitatively and quantitatively through comprehensive experiments on diverse datasets." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.444, + 0.597, + 0.459 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.469, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.511, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Int. Conf. Comput. Vis., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.554, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[3] Guha Balakrishnan, Raghudeep Gadde, Aleix Martinez, and Pietro Perona. Rayleigh eigendirections (reds): Nonlinear GAN latent space traversals for multidimensional features. In Eur. Conf. Comput. Vis., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay andgu2021 stylenerf Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[5] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.748, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[6] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. ACM Trans. Graph., 41(1):1:1-1:26, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In Eur. Conf. Comput. Vis., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[8] Min Jin Chong, Wen-Sheng Chu, Abhishek Kumar, and David A. Forsyth. Retrieve in style: Unsupervised facial feature transfer and retrieval. In Int. Conf. Comput. Vis., 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.469, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2456" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[9] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In B. Yegnanarayana, editor, Conf. of the Int. Speech Comm. Assoc., 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[10] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. GRAM: generative radiance manifolds for 3d-aware image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.195, + 0.468, + 0.277 + ], + "angle": 0, + "content": "[11] Jean Duchon. Splines minimizing rotation-invariant semi-norms in sobolev spaces. In Walter Schempp and Karl Zeller, editors, Constructive Theory of Functions of Several Variables: Proceedings of a Conference Held at Oberwolfach, Germany, April 25 - May 1, 1976, volume 571 of Lecture Notes in Mathematics, pages 85-100, 1976. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.28, + 0.468, + 0.334 + ], + "angle": 0, + "content": "[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph., 41(4):141:1-141:13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.337, + 0.468, + 0.379 + ], + "angle": 0, + "content": "[13] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.381, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.439, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[15] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.497, + 0.468, + 0.538 + ], + "angle": 0, + "content": "[16] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.541, + 0.468, + 0.596 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Adv. Neural Inform. Process. Syst., 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.599, + 0.468, + 0.639 + ], + "angle": 0, + "content": "[18] Kaiwen Jiang, Shu-Yu Chen, Feng-Lin Liu, Hongbo Fu, and Lin Gao. Nerffaceediting: Disentangled face editing in neural radiance fields. In SIGGRAPH Asia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.642, + 0.468, + 0.684 + ], + "angle": 0, + "content": "[19] Omer Kafri, Or Patashnik, Yuval Alaluf, and Daniel Cohen-Or. Stylefusion: A generative model for disentangling spatial segments. ACM Trans. Graph., 41(5):197:1-197:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.686, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.73, + 0.468, + 0.784 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[23] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in GAN for real-time image editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 852-861, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.093, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[24] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. In Eur. Conf. Comput. Vis., 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[25] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Adv. Neural Inform. Process. Syst., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Adv. Neural Inform. Process. Syst., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[27] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Winter Conf. on Appl. of Comput. Vis., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[28] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194:1-194:17, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. pages 16331–16345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Int. Conf. Comput. Vis., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[31] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriya Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[33] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[34] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Sign. Process. Letters, 20(3):209-212, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[35] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[36] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 42(1):1-13, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[37] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[38] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Adv. Neural Inform. Process. Syst., 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2457" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[39] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. IDE-3D: interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph., 41(6):270:1-270:10, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[40] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. Graph., 40(4):133:1-133:14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[42] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3121-3138, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[43] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. GRAM-HD: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[44] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for edible scene rendering. In Int. Conf. Comput. Vis., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[45] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet V2: bilateral network with guided aggregation for real-time semantic segmentation. Int. J. Comput. Vis., 129(11):3051-3068, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[46] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[47] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[48] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: Geometry editing of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[50] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Trans. Graph., 40(6):215:1-215:13, 2021. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[51] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Eur. Conf. Comput. Vis., 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2458" + } + ] +] \ No newline at end of file diff --git a/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_origin.pdf b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f91ade346ca0a14df4adb7457071e57c5e22c91c --- /dev/null +++ b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/ad8409a2-1fc3-480b-99f2-83a8ffcfe8c1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fe04f6ef52ad66a32dac3432069048692236d58064558b2e7d606d2f8ee80b1 +size 9146440 diff --git a/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/full.md b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aba1968d5da0dc2388027bcd0bd9c47c160d2a5d --- /dev/null +++ b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/full.md @@ -0,0 +1,287 @@ +# HairNeRF: Geometry-Aware Image Synthesis for Hairstyle Transfer + +Seunggyu Chang* +NAVER Cloud + +Gihoon Kim† +KAIST + +Hayeon Kim† +UNIST + +![](images/b00d7051ac27b798ac28133d5f4c6663d3a24415b743c2cb5e66ebd06867678e.jpg) +Figure 1: Given portrait images of hairstyle and identity (thunnails) that differ in viewing angle, head shape, and size, we seamlessly transfer hairstyle from one image to another utilizing neural rendering. Our method preserves fine-level details of the hairstyle structural geometry and generates diverse range of pose variant transferred images without requiring additional transfer process. + +![](images/6ec830d9680b4c305b9eaeaf8192b519884e0296137a75d517b554f9f3c0e8dc.jpg) + +![](images/a24d5e76de7c7f11d31cb3d624953ef561e2593586c65b62c149fe6936ab7d82.jpg) + +# Abstract + +We propose a novel hairstyle transferred image synthesis method considering the underlying head geometry of two input images. In traditional GAN-based methods, transferring hairstyle from one image to the other often makes the synthesized result awkward due to differences in pose, shape, and size of heads. To resolve this, we utilize neural rendering by registering two input heads in the volumetric space to make a transferred hairstyle fit on the head of a target image. Because of the geometric nature of neural rendering, our method can render view varying images of synthesized results from a single transfer process without causing distortion from which extant hairstyle transfer methods built upon traditional GAN-based generators suffer. We verify that our method surpasses other baselines in view of pre + +serving the identity and hairstyle of two input images when synthesizing a hairstyle transferred image rendered at any point of view. + +# 1. Introduction + +Selecting an appropriate hairstyle can be difficult as it is hard to imagine how it will look on your face without trying it out for yourself. The reason can be attributed to the inherent variability of individual head shape, resulting in the frequent occurrence of disparate outcomes when attempting to emulate celebrity hairstyles. To address this issue, we propose a novel image synthesis method for hairstyle transfer tailored to fit an individual head shape. We named our method HairNeRF as it considers geometric alignment in the volumetric space of neural rendering [33]. Hairstyle transfer synthesizes an image where the hairstyle from one + +portrait is transferred to the other and Generative adversarial networks (GANs) [14, 20, 22] have been extensively utilized to address this problem. Notably, following the emergence of StyleGAN [20, 21, 22], which enables the generation of highly realistic images, studies [37, 50, 19, 8] have been made to optimize latent vectors within the well-trained latent space of StyleGAN to generate an optimal synthesis outcome. These methods are successful at transferring hairstyles between images having similar poses, but struggle to generate convincing results for images having vastly different poses. To resolve this, studies [50, 24, 51] have proposed methods aligning pose of transferred hairstyle to a target portrait, simultaneously reducing the perceptual difference between the transferred result and the original hairstyle. However, they inherently encounter the same limitation of the pre-trained latent space of StyleGAN, where pose and style attributes are entangled as StyleGAN is not explicitly trained for hairstyle transfer. Consequently, the alignment process may distort the geometry of an original hairstyle, leading to undesirable alterations such as forehead stretching or shrinking. + +We categorize entanglement of pose and style attributes into two folds: entanglement of style with viewpoint, and entanglement of style with head shape. To address the entanglement of style with viewpoint, we leverage the latent space of StyleNeRF [15], a neural radiance field (NeRF) [32, 33] based GAN model as an alternative for StyleGAN. StyleNeRF exhibits an innate pose-invariant latent space as the network implies the volumetric geometry of an image. We basically follow the latent optimization technique following the literature [50, 24, 51] utilizing StyleGAN. Leveraging the pose-invariant nature of the StyleNeRF's latent space, we extract a pose-invariant style vector through GAN inversion [42]. Thereafter, we recombine the volumetric features generated by the inverted style vectors of the hair and the target portraits. However, despite the pose-invariant nature of StyleNeRF's latent space, it still entails an entanglement problem between style and head shape. The optimization of a style vector to fit on the head of a target portrait, while retaining hairstyle of a hair portrait, can potentially result in deformations of the hairstyle. To address the entanglement of style with head shape, we propose a registration method that aligns the geometries of two heads within the volumetric space, while maintaining the inverted style vectors intact. Subsequently, we introduce a novel rendering method that recombines hairstyle features with facial features within the volumetric space using alignment-inducing deformable rays, which are derived from the registration outcome. In this manner, we resolve the distorted hairstyle problem from which StyleGAN-based method suffers, resulting in better alignment of the transferred hairstyle seamlessly fit on the target portrait. Taking the advantage of geometric nature of neural rendering, HairNeRF is ca + +pable of generating a diverse set of geometrically consistent novel view hairstyle transferred images from a single synthesis, by merely adjusting camera parameters. This is in contrast to existing StyleGAN-based hairstyle transfer methods which necessitate separate optimizations for each of novel views, which often yield inconsistent outcomes. + +Experimental results on diverse datasets demonstrate the superiority of HairNeRF against other baselines at generating seamless hairstyle transfer outcomes, while effectively preserving the underlying geometry of the original hairstyle. + +# 2. Related Work + +# 2.1. Hairstyle Transfer + +The proficiency exhibited by generative adversarial networks (GANs) [14, 20, 22] in generating photorealistic images has propelled a surge of research addressing image editing problem. A significant amount of research [3, 7, 19, 16, 12] has effectively addressed attribute editing or style transfer, which modify abstract-level features of an image. However, the task of hairstyle transfer still presents a challenging problem that necessitates the preservation of intricate structure and color details of a hairstyle from a source image. With the emergence of StyleGAN [20, 21, 22], which demonstrates exceptional quality in generating photorealistic facial images, efforts have been made to discover latent vectors capable of capturing transferred images of hairstyles [37, 50]. These methods require the projection of an image into the pretrained StyleGAN's latent space known as inversion, which influences the quality of transfer. Thus, the Barbershop [50] introduces an improved inversion method on the enlarged latent space named $FS$ -space to capture finer details of input image. However, since these models lack awareness of the pose of given images, they suffer structural artifacts when the given two images largely differ in poses. Recent studies [24, 51] resolve this issue by aligning pose of the transferred hairstyle to the target image. However, they still suffer style deformation problem arising from the entangled latent space of StyleGAN for pose and style. + +# 2.2. 3D-Aware Image Synthesis and Editing + +The emergence of StyleGAN [20, 21, 22] has facilitated researches on image editing [16, 41, 2, 19, 29, 23, 1], either by navigating the latent style space or leveraging semantic maps. However, StyleGAN-based image editing methods have shown inconsistency among pose changes, which is attributed to the entanglement of style and pose within StyleGAN's latent space. The development of neural radiance fields (NeRF) [32, 33] has enabled the rendering of 3D-consistent images that accommodate varying viewpoints. Consequently, methods for editing a 3D scene + +![](images/a63f84764f186cfad92feb7f029ddb104ae386b3a56596919f3d276bc7ca157e.jpg) +Figure 2: Overview of the hairstyle transfer pipeline of HairNeRF consisting of five modules: (a) inversion, (b) head alignment, (c) occlusion disclosure, (d) recombination, and (e) blending. + +[44, 30, 27, 48] such as the addition and subtraction of NeRF models trained on individual objects have been explored. However, these approaches do not support modifications, such as altering the attributes or style of the object, due to the inherent nature of NeRF being trained to solely represent a single entity, thus lacking the necessary information for the desired edits. The advent of generative NeRFs [5, 38, 35, 15, 4, 10, 43] has enhanced the representational capabilities of NeRF models, enabling the expression of an expansive range of objects through the exploration of the latent space within a single model. The enhanced expressive capabilities of generative NeRFs enable the modification of attributes and style of an object [15, 4, 10]. Furthermore, within the generative NeRF modules, methods [30, 18, 40, 39, 6] have been proposed that disentangle the semantic structure and texture. However, these methods are not well-suited for local style modifications, as the latent vector representing the texture exhibits spatial entanglement, encompassing the entire regions of the image. + +# 3. Method + +# 3.1. Overview + +We follow the basic synthesis pipeline from the Barbershop [50], a founding work for hairstyle transfer [24, 50]. We encourage readers to refer [50] for the details of the original pipeline. The overall synthesis pipeline of our method consists of four modules: inversion, head alignment, recombination, and blending, as illustrated in Figure 2. Given two images, id and hair, we transfer hairstyle from the hair image into the id image. In the inversion module, we first invert the two images to obtain latent vectors and the corresponding features in the volu + +metric space. Then, in the alignment module, we conduct registration on two geometric heads of the inverted images in the volumetric space. By aligning geometric heads, we can transfer hairstyle of the hair image into the id image to be more fit. Once the registration is done, we recombine the 3D features in the volumetric space according to semantics to render a 2D feature for synthesis using neural rendering in the recombination module. A simple method is to bring hair semantic features from the hair image, and other semantics from the id image. However, there occur missing regions which are occluded by the hair in the id image but not covered by the hair of the hair image. To fill in the missing regions, we inpaint 3D features by optimizing a latent vector, ensuring that the 2D features rendered from the 3D features generated by the optimized latent vector align with the target semantics across all viewpoints. Thereafter, we semantically integrate the 3D features from the id, hair, and the inpainting output to form a blended feature for the synthesis. The blend feature is rendered to a 2D feature, which is decoded by the upsampler blocks of StyleNeRF [15] conditioned on the blending style vector. In the final blending module, the blending style vector is optimized to make the decoded image preserve the same hairstyle as in the hair image while retaining face of the id image. + +# 3.2. NeRF Inversion + +Semantic Parsing NeRF Head. To semantically recombine features in the 3D volume, semantics of volumetric features are required. It is known that a semantic classifier trained to classify a 2D point feature obtained by (2) can aptly classify semantic of a 3D point feature as well [25]. Based on this observation, we add an additional seman- + +![](images/aba1ee09fecc380b92a6efedf7e9cb6b5e318b8210adf7dff90c330537c7706a.jpg) +Figure 3: The additional semantic head is added at the end of a MLP block of NeRF module, which outputs a logit of semantics in the volumetric space. The additional semantic head is trained to minimize cross-entropy between the rendered logit, $\hat{s}$ , and the ground-truth. + +tic branch to the MLP block for neural rendering and train it using outputs of the off-the-shelf semantic segmentation model [46] for generated images as ground-truths as shown in Figure 3. After training, the trained semantic branch outputs a logit of semantics for a 3D point feature. + +Semantics-Aware Inversion in 3D. We invert the hair and the id images into the latent vectors for the synthesis network of the StyleNeRF [15]. The process of obtaining the latent vector generating a desired image is called GAN inversion. We utilize the pivotal tuning inversion (PTI) [36] with a little modification for inversion as the PTI can reconstruct fine details of an input image from the inverted latent vector. As our model is built upon the StyleNeRF, the output image varies as a camera parameter changes. For this reason, the inversion process requires an input camera parameter or should estimate a camera parameter in our case. We choose to provide a fixed camera parameter of an input image estimated by the off-the-shelf head pose estimator [31] for the inversion process. Using the PTI with a fixed camera parameter, we optimize $W \in \mathcal{W}$ for each of the id and the hair images, and finetune the weights of the pretrained StyleNeRF to reconstruct the input image and the semantics estimated by the off-the-shelf semantic parsing networks [47, 45]. + +# 3.3. Head Alignment + +Registration For registration, we render each of inverted images into $n_c$ different viewpoints and extract 2D keypoints of the FLAME [28] head model using the off-the-shelf head pose estimator [31]. Then we estimate 3D keypoints of the FLAME model for each inverted image from the set of 2D keypoints estimated from the eight different camera views by solving a bundle-adjustment-like least square problem: + +$$ +\underset {\boldsymbol {p}} {\operatorname {m i n i m i z e}} \sum_ {i} ^ {n _ {c}} \| P (\boldsymbol {c} _ {i}, \boldsymbol {p}) - \boldsymbol {x} _ {i} \| _ {2} ^ {2}, \tag {1} +$$ + +where $\pmb{x}_i \in \mathbb{R}^2$ denotes the estimated 2D keypoints at $i$ -th viewpoint, $\pmb{c}_i$ denotes the $i$ -th camera parameter, and $P(\pmb{c}_i, \pmb{p}_i)$ denotes the projection function projecting a 3D point $\pmb{p}_i \in \mathbb{R}^3$ into 2D using $\pmb{c}_i$ . As we already know the camera parameters used for rendering $n_c$ images, we exclude camera parameters from the optimization variables. We utilize the thin-plate-spline (TPS) transform [11] to find the optimal non-rigid transformation for the registration. + +Neural Rendering via Ray Deformation. The two heads of the id and the hair do not fully overlap due to the differences in characteristics, shape and size of the heads geometry, along with estimation errors in head poses. We refer to this misalignment as intrinsic misalignment, as it arises from factors unrelated to external elements such as camera parameters. To alleviate the intrinsic misalignment, the registration process finds an optimal transformation between two point sets, by which one point set is transformed to align with the other. During the neural rendering process, point features are sampled along a ray within the volumetric space, and these features are then aggregated to a single point on the image plane. If the sampling points are collected along the ray, deformed by the transformation determined through registration, we can obtain registered point features. This process leads to a registered 2d image after the aggregation. + +Figure 4 illustrates the effective neural rendering process for alignment via ray deformation. Let $T_{i\rightarrow h}$ denote a transformation determined through registration, which aligns the point set of the id features to the point set of the hair features. As depicted in Figure 4a, sampling along rays that radiate straight from the id's camera renders an id's face with an innate head shape, colored in red, as in shown Figure 4c. Likewise, rendering from straight rays radiating through hair features $\Phi^{\text{hair}}$ using the same camera renders the hair image posed as id, with the overlaid innate head shape colored in blue, as depicted in Figure 4e. The two heads do not overlap due to differences in shape and size. However, if we render a hair image using the deformed ray transformed by $T_{i\rightarrow h}$ , as depicted in Figure 4b, we obtain a registered hair image aligned to the head of id, as shown in Figure 4d. With a camera ray $r = o + td$ , the registered neural rendering computes a pixel value of the ray as: + +$$ +I _ {\omega} ^ {\text {R e g i s t e r e d}} (\boldsymbol {r}) = \int_ {0} ^ {\infty} p _ {\omega} (t) \boldsymbol {c} _ {\omega} (\tilde {\boldsymbol {r}} (t), \tilde {\boldsymbol {d}}) \left| \frac {d \tilde {\boldsymbol {r}}}{d t} \right| d t, \tag {2} +$$ + +$$ +\text {w h e r e} p _ {\omega} (t) = \exp \left(- \int_ {0} ^ {t} \sigma_ {\omega} (\tilde {\boldsymbol {r}} (s)) \left| \frac {d \tilde {\boldsymbol {r}}}{d s} \right| d s\right) \cdot \sigma_ {\omega} (\tilde {\boldsymbol {r}} (t)), +$$ + +with a registered ray $\tilde{\boldsymbol{r}} = T_{i\rightarrow h}(\boldsymbol {r})$ , and the adjusted directional vector $\tilde{\pmb{d}} = d\tilde{\pmb{r}} /dt$ .Equation (2) has the same form as the original neural rendering, but with the integration computed along the transformed ray. This adjustment + +![](images/6508a6e9c44b8e6b01f79c6722d561290c3143acf9e761fc61d3dd01fc15cd4d.jpg) + +![](images/7486f50edcfdd22c1d2ec3d87bde1bfb8b80e929cff5a632406d3f5d3fe7bab2.jpg) + +![](images/e18f7cda8ee41c1cceea31f95d8ccf3c22ad62f89fd27ed09946c74e6852570f.jpg) +Figure 4: Illustration of alignment-aware neural rendering via ray deformation. Figure (c) - (e), and (h) - (j) show overlaid heads on rendering images. Rendering via transformed ray render a registered image as shown in (d) and (i), which is equivalent to rendering with the straight ray on transformed feature represented as $\Phi^{\text{hair}'}$ and $\Phi^{\text{id}'}$ , respectively. + +![](images/d7e1f1332de8ae6dd9026ca8050544ece552d4bf51b6152828e839dae8e99033.jpg) + +is achieved by substituting the directional vector $\pmb{d}$ with $\tilde{\pmb{d}}$ and including $|d\tilde{\pmb{r}} /dt|$ to adjust the infinitesimal length $dt$ . Since $|d\tilde{\pmb{r}} /dt| \approx 1$ in practice, we omit the determinant term to simplify the computation. + +# 3.4. Volumetric Recombination + +Occlusion Disclosure. To inpaint the missing regions occluded by the hair in $\Phi^{\mathrm{hair}}$ and produce the recombined 3D features $\Phi^{\mathrm{blend}}$ as depicted in Figure 2c, we optimize the latent vector $W^{\mathrm{inp}} \in \mathcal{W}+$ within the StyleNeRF [15] framework to generate $\Phi^{\mathrm{inp}}$ . Analogous to [50], we produce a volumetric target semantic mask, as depicted in Figure 5. For each iteration, we optimize $W^{\mathrm{inp}}$ to generate $\Phi^{\mathrm{inp}}$ that aligns with the target semantic mask $s^{\mathrm{inp}}$ from a randomly selected viewpoint. Let $m_{hair}$ denote a volumetric binary mask representing hair, $c$ denote a camera parameter, and $\phi|_c$ represents 2D features rendered by the camera parameter $c$ . The target semantics $s^{\mathrm{inp}}$ and the objectives for the optimization are constructed as follows: + +$$ +\begin{array}{l} \boldsymbol {s} ^ {\text {i n p}} = \boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r} ^ {\prime}} \circ \boldsymbol {s} ^ {\text {h a i r} ^ {\prime}} + \left(1 - \boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r} ^ {\prime}}\right) (3) \\ \circ \Big (\left(1 - \boldsymbol {m} _ {h a i r} ^ {\mathrm {i d}}\right) \circ \boldsymbol {s} ^ {\mathrm {i d}} + \boldsymbol {m} _ {h a i r} ^ {\mathrm {i d}} \circ \boldsymbol {s} ^ {\mathrm {h a i r} ^ {\prime}} \Big), \\ \mathcal {L} ^ {\text {i n p - s e m}} = \mathbb {E} _ {c} \mathrm {C E} \left(\mathcal {R} _ {c} \left(s ^ {\text {i n p}}\right), \mathcal {R} _ {c} \left(f _ {W ^ {\text {i n p}}} ^ {\text {s e m}}\right)\right), (4) \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathcal {L} ^ {\text {i n p - r e g}} = \mathbb {E} _ {\boldsymbol {c}} \left\| G \left(\left(1 - \mathcal {R} _ {\boldsymbol {c}} \left(\boldsymbol {m} _ {\text {h a i r}} ^ {\text {i n p}}\right)\right) \circ U _ {W} \left(\phi^ {\text {i n p}} \mid_ {\boldsymbol {c}}\right)\right) \right. \\ \left. - G \Big (\big (1 - \mathcal {R} _ {\boldsymbol {c} ^ {\mathrm {i d}}} (\boldsymbol {m} _ {h a i r} ^ {\mathrm {i d}}) \big) \circ I ^ {\mathrm {i d}} \Big) \right\| _ {2} ^ {2}, \\ \end{array} +$$ + +where $\mathcal{R}_c$ represents the aggregation function of neural rendering according to the camera $c$ , $f^{\mathrm{sem}}$ denotes the MLP block with semantic head, $U_W$ refers to the upsampler block of StyleNeRF whose output is a final image, and $G$ denotes the Gram matrix for style loss [13]. Symbols marked with an apostrophe indicate that they are registered to their counterparts, as denoted by the associated superscripts. Unlike [50], our target semantic mask is defined as a continuous function within the volumetric space, rather than as a tensor defined at discrete positions. As a result, $\mathcal{L}^{\mathrm{inp - sem}}$ entails integration over this volumetric space. Hence, for each optimization iteration in practice, we compute $\mathcal{L}^{\mathrm{inp - sem}}$ using a random set of rays radiating from a randomly sampled camera $c$ positioned on half of a hemisphere. Combining all elements, we obtain the inpainting feature $\Phi^{\mathrm{inp}}$ as + +$$ +\Phi^ {\text {i n p}} = \operatorname {M L P} \left(W ^ {\text {i n p} ^ {*}}\right), \tag {5} +$$ + +where $W^{\mathrm{inp}^*} = \arg \min_{W}\mathcal{L}^{\mathrm{inp - sem}} + \lambda \mathcal{L}^{\mathrm{inp - reg}}$ + +Recombination. We make blend features in the volumetric space by summing up id and hair features, weighting them using binarized semantic maps representing the hair. Subsequently, the blend features are expected to be aggregated to 2D features using (2). However, the blend features contain missing regions that are occluded by the hair in the id image but are not fully covered by the hair in the hair image, leading to missing holes in the final rendered image. To mitigate this, we fill the missing regions within the blend features $\Phi^{\mathrm{blend}}$ by the inpainting features $\Phi^{\mathrm{inp}}$ , as obtained by (5). This blending concept is motivated by the Barbershop [50], but ours differs in that the features to be blended + +![](images/e5cc146bdbbd42b2defd699421d655b46616debcfecedc843cb580a0c8601778.jpg) +Figure 5: Target semantics generation for occlusion disclosure. The target semantics can be assumed to an ideal semantics for the hairstyle transferred result. The inpainting latent $W^{\mathrm{inp}}$ is optimized to minimize to match the 2D semantics rendered from $s^{\mathrm{inp}}$ at a random viewpoint at every iteration. + +lie in the volumetric space. + +Let $\phi^{\mathrm{id}}$ and $\phi^{\mathrm{hair}}$ denote rendered 2D features of id and hair, respectively, rendered using neural rendering without ray deformation. Meanwhile, $\phi^{\mathrm{id}'}$ and $\phi^{\mathrm{hair}'}$ denote the registered 2D features of id and hair, aligned to their respective counterparts. Then we recombine features for blending as: + +$$ +\begin{array}{l} \phi_ {i d} ^ {\text {b l e n d}} = \boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r} ^ {\prime}} \circ \phi^ {\text {h a i r} ^ {\prime}} + \left(1 - \boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r} ^ {\prime}}\right) \\ \circ \left(\left(1 - \boldsymbol {m} _ {\text {h a i r}} ^ {\mathrm {i d}}\right) \circ \phi^ {\mathrm {i d}} + \boldsymbol {m} _ {\text {h a i r}} ^ {\mathrm {i d}} \circ \phi^ {\mathrm {i n p}}\right), \tag {6} \\ \end{array} +$$ + +$$ +\begin{array}{l} \phi_ {\text {h a i r}} ^ {\text {b l e n d}} = \boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r}} \circ \phi^ {\text {h a i r}} + \left(1 - \boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r}}\right) \\ \circ \left(\left(1 - \pmb {m} _ {h a i r} ^ {\mathrm {i d} ^ {\prime}}\right) \circ \phi^ {\mathrm {i d} ^ {\prime}} + \pmb {m} _ {h a i r} ^ {\mathrm {i d} ^ {\prime}} \circ \phi^ {\mathrm {i n p} ^ {\prime}}\right), \\ \end{array} +$$ + +where $m_{hair}$ denotes a 2D binary mask for the hair semantics of $\phi$ , associated with a corresponding superscript. + +# 3.5. Blending + +Alignment-Aware Blending Loss. In the blending step, the upsampler blocks in the StyleNeRF [15] decode the blend features obtained by (6) into an output image in high resolution, conditioned on a latent vector $W^{\text{blend}}$ . We optimize the latent vector $W^{\text{blend}}$ to ensure that the decoded image retains the facial identity from the id image while seamlessly transferring hairstyle of the hair image. To this end, we design the objective function for the optimization to + +include face-preserving and hair-preserving terms. + +$$ +\begin{array}{l} \mathcal {L} _ {\text {f a c e}} ^ {\mathrm {l p i p s}} = \mathcal {L} ^ {\mathrm {l p i p s}} \left(\left(1 - \boldsymbol {m} _ {\text {h a i r}} ^ {\mathrm {i d}}\right) \circ U _ {W ^ {\text {b l e n d}}} \left(\phi_ {\mathrm {i d}} ^ {\text {b l e n d}}\right), \right. \\ \left(1 - \boldsymbol {m} _ {\text {h a i r}} ^ {\mathrm {i d}}\right) \circ I ^ {\mathrm {i d}}\bigg), \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {\text {h a i r}} ^ {\mathrm {l p i p s}} = \mathcal {L} ^ {\mathrm {l p i p s}} \left(\boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r}} \circ U _ {W ^ {\text {b l e n d}}} \left(\phi_ {\text {h a i r}} ^ {\text {b l e n d}}\right), \right. \\ \left. \boldsymbol {m} _ {\text {h a i r}} ^ {\text {h a i r}} \circ I ^ {\text {h a i r}}\right), \tag {7} \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {h a i r} ^ {\text {s t y l e}} = \left\| G \left(\boldsymbol {m} _ {h a i r} ^ {\text {h a i r}} \circ U _ {W ^ {\text {b l e n d}}} \left(\phi_ {\text {h a i r}} ^ {\text {b l e n d}}\right)\right) \right. \\ - \left. G \left(\boldsymbol {m} _ {h a i r} ^ {\text {h a i r}} \circ I ^ {\text {h a i r}}\right) \right\| _ {2} ^ {2}, \\ \end{array} +$$ + +where $m_{face}^{\mathrm{id}}$ denotes the binary mask for face region of the face image, $m_{hair}^{\mathrm{hair}}$ denotes the binary mask for hair region of the hair image, $U_W$ denotes the upsampler conditioned on $W$ , $\mathcal{L}^{\mathrm{lpips}}$ denotes the perceptual loss [49], and $G$ denotes the Gram matrix for the style loss [13]. Combining all elements, the optimal blend vector $W^{\mathrm{blend*}}$ is obtained as + +$$ +W ^ {\text {b l e n d} *} = \arg \min _ {W} \lambda_ {f a c e} \mathcal {L} _ {f a c e} ^ {\mathrm {l p i p s}} + \lambda_ {h a i r} \mathcal {L} _ {h a i r} ^ {\mathrm {l p i p s}} + \mathcal {L} _ {h a i r} ^ {\text {s t y l e}}. \tag {8} +$$ + +# 4. Experimental Results + +We evaluate our method on two distinct tasks following [24]: hairstyle transfer and reconstruction. In the hairstyle transfer task, each input pair consists of different identities, and the goal is to transfer the hairstyle from the hair image to the face of the id image. In the reconstruction task, the input pair shares the same identity but with different poses. Since the id and hair images share the same identity in the reconstruction task, the id serves as the ground-truth for the evaluation. We use Flicker-Faces-HQ (FFHQ) dataset [21] and faces images crawled from Unsplash (UF), as used in [50, 51], for the hairstyle transfer task. For the reconstruction task, we employ the VoxCeleb2 [9] dataset. To evaluate the overall synthesis quality of the generated images, we measure the Frèchet Inception distance (FID) [17], the naturalness image quality estimator (NIQE) [34], and the precision and recall [26]. In the hairstyle transfer task, we synthesize 1,000 images using the FFHQ dataset and 380 images using the Unsplash dataset. For the reconstruction task, we synthesize 500 images. To demonstrate the effectiveness of HairNeRF at preserving geometric structure, we provide both qualitative and quantitative comparisons of our method against other baselines. + +# 4.1. Comparison Study + +We compare our method to LOHO [37], the Barbershop [50], and StyleYourHair [24] on the FFHQ and VoxCeleb2 datasets. For the hairstyle transfer task, we include a comparison with HairNet [51] using only the UF dataset. + +Hairstyle Transfer. Quantitative results on FFHQ and UF are summarized in Table 1. Notice that LOHO, the Barbershop and StyleYourHair is built upon StyleGAN2, while HairNeRF utilizes StyleNeRF. Both StyleGAN2 and + +![](images/3fe5926f25d98469704a9a656e80500a8747a43b5a665825646f2f5e2fc4e7ea.jpg) +Figure 6: Comparison of results for FFHQ datasets. Both StyleYourHair and our method align the hair even if the poses between the input images are different, but our method preserves the direction of the hair better (first and the last rows). + +![](images/d63ce7d1cc4a9afd84affb13c6e669f713a6e79a6cc22da133745380870906ab.jpg) + +![](images/0af044fbc8e4bddaed61aa8951135517bda95d8266ebe532ae3c9045a43380c1.jpg) +Figure 7: Comparison of results for Unsplash datasets. Our method preserves identity better (second row on the left), and hair transfer is performed to fit the structure of identity (third row on the left). + +![](images/c2caad9fef51e9982cada4fb16a2b99a54746d738b11b589b273813d2f55f342.jpg) + +StyleNeRF are effective in generating high-quality face images; however, it has been reported that StyleGANs slightly outperforms NeRF-based model in terms of commonly used metrics, including FID. Consequently, the overall metrics + +for HairNeRF are slightly lower than the other baselines, with the exception of NIQE. Nevertheless, this difference is not substantial, and the qualitative results of HairNeRF reveal perceptually better quality in comparison to the other + +
FFHQ
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO40.650.830.850.551.3491.42
Barbershop12.240.770.930.831.0648.00
StyleYourHair12.000.550.930.771.1151.36
HairNet11.630.760.970.591.1456.45
HairNeRF9.570.700.890.671.4057.26
+ +
UF
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO42.370.800.910.451.0786.16
Barbershop12.130.740.940.841.9141.06
StyleYourHair12.070.750.940.802.1742.62
HairNeRF10.090.700.890.571.7149.74
+ +![](images/c6d6212eb26c1f434a103602fdc28644764c7b5a47dcafd432e02198a6be4610.jpg) +Figure 8: Comparison examples for reconstruction task. + +basielines. Figures 6 and 7 show synthesized examples using the FFHQ and UF datasets, respectively. As observed in the fifth row of Figure 6b and Figure 7, the Barbershop generates fine-quality images when the poses of the id and hair images are similar. However, the output quality substantially declines when the difference in pose between the two input images increases. As shown in the second and the third row of Figure 6a, substantial structural artifacts, such as bumps and missing holes, appear in the transferred hairstyle images synthesized by the Barbershop. In contrast, the transferred hairstyle of HairNeRF seamlessly fits the head of the id image. StyleYourHair and HairNet also generate satisfactory results for input pairs with significant pose differences. However, as seen in the fifth row of Figure 6b, the detailed structure of the hairstyle often changes in the synthesized results. Furthermore, they often elongate the length of the forehead, as shown in the fourth row of Figure 7a. + +Reconstruction. Following [24], we evaluate reconstruction quality on the VoxCeleb2 dataset, consisting of pairs of images with same identity but different poses. Figure 8 shows an example of reconstruction task. Ideally, the synthesis result should be identical to the id image, since the id and hair images represent the same person. Figure 8 shows the reconstruction results of HairNeRF in comparison to other baselines. While the reconstructed image pre + +Table 1: Hairstyle transfer results on FFHQ and UF datasets. HairNeRF shows comparable results to StyleGAN2-based hairstyle transfer models. + +
MethodMSE↓SSIM↑LPIPS↓PSNR↑
Barbershop [50]1.1650.8010.12731.44
StyleYourHair [24]1.2180.7950.15531.34
HairNeRF0.9570.8210.10631.84
+ +Table 2: Reconstruction result. HairNeRF surpasses other baselines in view of MSE, SSIM, LPIPS, and PSNR. The result verifies the effectiveness of HairNeRF at preserving hairstyle geometry compared to the other baselines. + +![](images/dc8b20d7421c491836d3bd1672716ca5a40a55ae97739a9d793f871fd360a977.jpg) +Figure 9: Examples of novel view synthesis. + +serves the detailed geometry of the hair, both the Barbershop and StyleYourHair show slight alterations. The quantitative results, summarized in Table 2, further validate the superiority of HairNeRF in preserving the geometry of hairstyles. + +Novel View Synthesis. Owing to the geometric nature of neural rendering, HairNeRF can generate diverse pose varying images from the same latent vector once optimized for synthesis. Figure 9 shows examples of novel view images obtained by varying the camera parameters. For a qualitative comparison, we generate a diverse set of novel view images of the id image using the inversion result of HairNeRF. We then synthesize each novel view image using the Barbershop and StyleYourHair from the given hair image. As shown in Figure 9, the transferred hairstyle varies by views in the case of baselines, whereas HairNeRF preserves a consistent hair structure that aligns with the original hair image. + +![](images/5145b4a2ec1b267bee3fe20905dd1ad1779f33069e022f74859f69647c2f5951.jpg) +Figure 10: Effectiveness of registration shown with the underlying geometry. Without registration the transferred hairstyle does not perfectly fit to the face (first row) or erode the head (third row) of the id image. + +
MethodMSE↓SSIM↑LPIPS↓PSNR↑
w/o regi.0.9800.8180.10831.77
w/ regi.0.9570.8210.10631.84
+ +Table 3: Reconstruction quality with and without head alignment module. With registration, the reconstruction quality increases in view of MSE, SSIM, LIPIS, and PSNR. + +# 4.2. Ablation Study + +Effectiveness of Registration. To validate the effectiveness of the head alignment module, we evaluate the reconstruction quality using the VoxCeleb2 dataset. Table 3 summarizes the results. As mentioned in Section 3.3, StyleNERF suffers from an intrinsic misalignment problem that results in subtle differences, even if the pair of input images shares the same identity. However, the registration process within the head alignment module reduces the existing geometric differences to make transferred hairstyle more fit to the id image. Table 3 shows that the reconstruction quality deteriorates when the registration module is missing. Figure 10 shows the effectiveness of registration with the underlying geometry. The first example shows that the shape of the original hair is slightly larger than the head of id image. Without registration, the transferred hair sticks out to the right side of the head of the id. However, with registration, the transferred hair fits more closely to the head of the id. The difference is clearly shown in the overlap image. In the second example, it is clearly shown that the transferred hair is stretched to fit on the head of id by registration. + +# 5. Conclusion + +We introduce HairNeRF, a novel geometry-aware hairstyle transfer method designed to address the entanglement issue of pose and style attributes within the latent space of StyleGAN. By leveraging NeRF-based generative models, where the pose of the rendered image is determined by external camera parameters, we achieve a complete disentanglement of pose from the style latent space. Additionally, we propose an efficient method for recombining features within the volumetric space using semantics. This is achieved by aligning two distinct innate head structures via ray deformation. This method unveils the potential to seamlessly integrate multiple NeRF-represented objects of varying shapes and sizes into one unified object, without requiring additional modification to NeRF weights. Using the intrinsic geometric properties of neural rendering, our method can produce view-varied images from a single synthesis process without inducing geometric deformation. The efficacy of HairNeRF in preserving the geometric structure of the original hairstyle, compared to other baselines, is substantiated both qualitatively and quantitatively through comprehensive experiments on diverse datasets. + +# References + +[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In SIGGRAPH, 2022. 2 +[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Int. Conf. Comput. Vis., 2021. 2 +[3] Guha Balakrishnan, Raghudeep Gadde, Aleix Martinez, and Pietro Perona. Rayleigh eigendirections (reds): Nonlinear GAN latent space traversals for multidimensional features. In Eur. Conf. Comput. Vis., 2022. 2 +[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay andgu2021 stylenerf Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[5] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3 +[6] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. ACM Trans. Graph., 41(1):1:1-1:26, 2022. 3 +[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In Eur. Conf. Comput. Vis., 2022. 2 +[8] Min Jin Chong, Wen-Sheng Chu, Abhishek Kumar, and David A. Forsyth. Retrieve in style: Unsupervised facial feature transfer and retrieval. In Int. Conf. Comput. Vis., 2021. 2 + +[9] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In B. Yegnanarayana, editor, Conf. of the Int. Speech Comm. Assoc., 2018. 6 +[10] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. GRAM: generative radiance manifolds for 3d-aware image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[11] Jean Duchon. Splines minimizing rotation-invariant semi-norms in sobolev spaces. In Walter Schempp and Karl Zeller, editors, Constructive Theory of Functions of Several Variables: Proceedings of a Conference Held at Oberwolfach, Germany, April 25 - May 1, 1976, volume 571 of Lecture Notes in Mathematics, pages 85-100, 1976. 4 +[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph., 41(4):141:1-141:13, 2022. 2 +[13] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 5, 6 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[15] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2, 3, 4, 5, 6 +[16] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Adv. Neural Inform. Process. Syst., 2017. 6 +[18] Kaiwen Jiang, Shu-Yu Chen, Feng-Lin Liu, Hongbo Fu, and Lin Gao. Nerffaceediting: Disentangled face editing in neural radiance fields. In SIGGRAPH Asia, 2022. 3 +[19] Omer Kafri, Or Patashnik, Yuval Alaluf, and Daniel Cohen-Or. Stylefusion: A generative model for disentangling spatial segments. ACM Trans. Graph., 41(5):197:1-197:15, 2022. 2 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2 +[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 2, 6 +[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2 +[23] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in GAN for real-time image editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 852-861, 2021. 2 + +[24] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. In Eur. Conf. Comput. Vis., 2022. 2, 3, 6, 8 +[25] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Adv. Neural Inform. Process. Syst., 2022. 3 +[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Adv. Neural Inform. Process. Syst., 2019. 6 +[27] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Winter Conf. on Appl. of Comput. Vis., 2023. 3 +[28] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194:1-194:17, 2017. 4 +[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. pages 16331–16345, 2021. 2 +[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Int. Conf. Comput. Vis., 2021. 3 +[31] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriya Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4 +[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2 +[33] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2022. 1, 2 +[34] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a "completely blind" image quality analyzer. IEEE Sign. Process. Letters, 20(3):209-212, 2013. 6 +[35] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3 +[36] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 42(1):1-13, 2022. 4 +[37] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 6 +[38] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Adv. Neural Inform. Process. Syst., 2020. 3 + +[39] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. IDE-3D: interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph., 41(6):270:1-270:10, 2022. 3 +[40] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. Graph., 40(4):133:1-133:14, 2021. 2 +[42] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3121-3138, 2023. 2 +[43] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. GRAM-HD: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv, 2022. 3 +[44] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for edible scene rendering. In Int. Conf. Comput. Vis., 2021. 3 +[45] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet V2: bilateral network with guided aggregation for real-time semantic segmentation. Int. J. Comput. Vis., 129(11):3051-3068, 2021. 4 +[46] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4 +[47] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4 +[48] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: Geometry editing of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6 +[50] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Trans. Graph., 40(6):215:1-215:13, 2021. 2, 3, 5, 6, 8 +[51] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Eur. Conf. Comput. Vis., 2022. 2, 6 \ No newline at end of file diff --git a/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/images.zip b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d649377208f4c38a1609d71a1d86a994dc46abbd --- /dev/null +++ b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65b5d5aaa2399742be75f3eea5c1f060656efab167eae732f170d5c994497f6f +size 997771 diff --git a/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/layout.json b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7b0ccf9bcb736ce47e23bc8c055a042f2588efa5 --- /dev/null +++ b/2023/HairNeRF_ Geometry-Aware Image Synthesis for Hairstyle Transfer/layout.json @@ -0,0 +1,7536 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 507, + 120 + ], + "type": "text", + "content": "HairNeRF: Geometry-Aware Image Synthesis for Hairstyle Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 143, + 236, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 143, + 236, + 171 + ], + "spans": [ + { + "bbox": [ + 146, + 143, + 236, + 171 + ], + "type": "text", + "content": "Seunggyu Chang* \nNAVER Cloud" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 273, + 143, + 340, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 143, + 340, + 171 + ], + "spans": [ + { + "bbox": [ + 273, + 143, + 340, + 171 + ], + "type": "text", + "content": "Gihoon Kim† \nKAIST" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 378, + 143, + 446, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 143, + 446, + 171 + ], + "spans": [ + { + "bbox": [ + 378, + 143, + 446, + 171 + ], + "type": "text", + "content": "Hayeon Kim† \nUNIST" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 47, + 213, + 87, + 441 + ], + "blocks": [ + { + "bbox": [ + 47, + 213, + 87, + 441 + ], + "lines": [ + { + "bbox": [ + 47, + 213, + 87, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 87, + 441 + ], + "type": "image", + "image_path": "b00d7051ac27b798ac28133d5f4c6663d3a24415b743c2cb5e66ebd06867678e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 449, + 547, + 483 + ], + "lines": [ + { + "bbox": [ + 45, + 449, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 45, + 449, + 547, + 483 + ], + "type": "text", + "content": "Figure 1: Given portrait images of hairstyle and identity (thunnails) that differ in viewing angle, head shape, and size, we seamlessly transfer hairstyle from one image to another utilizing neural rendering. Our method preserves fine-level details of the hairstyle structural geometry and generates diverse range of pose variant transferred images without requiring additional transfer process." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 93, + 213, + 168, + 441 + ], + "blocks": [ + { + "bbox": [ + 93, + 213, + 168, + 441 + ], + "lines": [ + { + "bbox": [ + 93, + 213, + 168, + 441 + ], + "spans": [ + { + "bbox": [ + 93, + 213, + 168, + 441 + ], + "type": "image", + "image_path": "6ec830d9680b4c305b9eaeaf8192b519884e0296137a75d517b554f9f3c0e8dc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 173, + 213, + 545, + 441 + ], + "blocks": [ + { + "bbox": [ + 173, + 213, + 545, + 441 + ], + "lines": [ + { + "bbox": [ + 173, + 213, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 173, + 213, + 545, + 441 + ], + "type": "image", + "image_path": "a24d5e76de7c7f11d31cb3d624953ef561e2593586c65b62c149fe6936ab7d82.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 493, + 192, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 493, + 192, + 506 + ], + "spans": [ + { + "bbox": [ + 143, + 493, + 192, + 506 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 517, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 289, + 685 + ], + "type": "text", + "content": "We propose a novel hairstyle transferred image synthesis method considering the underlying head geometry of two input images. In traditional GAN-based methods, transferring hairstyle from one image to the other often makes the synthesized result awkward due to differences in pose, shape, and size of heads. To resolve this, we utilize neural rendering by registering two input heads in the volumetric space to make a transferred hairstyle fit on the head of a target image. Because of the geometric nature of neural rendering, our method can render view varying images of synthesized results from a single transfer process without causing distortion from which extant hairstyle transfer methods built upon traditional GAN-based generators suffer. We verify that our method surpasses other baselines in view of pre" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 495, + 547, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 547, + 531 + ], + "type": "text", + "content": "serving the identity and hairstyle of two input images when synthesizing a hairstyle transferred image rendered at any point of view." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 559, + 387, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 387, + 572 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 387, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Selecting an appropriate hairstyle can be difficult as it is hard to imagine how it will look on your face without trying it out for yourself. The reason can be attributed to the inherent variability of individual head shape, resulting in the frequent occurrence of disparate outcomes when attempting to emulate celebrity hairstyles. To address this issue, we propose a novel image synthesis method for hairstyle transfer tailored to fit an individual head shape. We named our method HairNeRF as it considers geometric alignment in the volumetric space of neural rendering [33]. Hairstyle transfer synthesizes an image where the hairstyle from one" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 274, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 274, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 274, + 703 + ], + "type": "text", + "content": "*Corresponding author: seunggyu.chang@navercorp.com" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 703, + 245, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 245, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 245, + 712 + ], + "type": "text", + "content": "This work was done during internship at Naver Cloud." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2448" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 323 + ], + "type": "text", + "content": "portrait is transferred to the other and Generative adversarial networks (GANs) [14, 20, 22] have been extensively utilized to address this problem. Notably, following the emergence of StyleGAN [20, 21, 22], which enables the generation of highly realistic images, studies [37, 50, 19, 8] have been made to optimize latent vectors within the well-trained latent space of StyleGAN to generate an optimal synthesis outcome. These methods are successful at transferring hairstyles between images having similar poses, but struggle to generate convincing results for images having vastly different poses. To resolve this, studies [50, 24, 51] have proposed methods aligning pose of transferred hairstyle to a target portrait, simultaneously reducing the perceptual difference between the transferred result and the original hairstyle. However, they inherently encounter the same limitation of the pre-trained latent space of StyleGAN, where pose and style attributes are entangled as StyleGAN is not explicitly trained for hairstyle transfer. Consequently, the alignment process may distort the geometry of an original hairstyle, leading to undesirable alterations such as forehead stretching or shrinking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 331, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 286, + 712 + ], + "type": "text", + "content": "We categorize entanglement of pose and style attributes into two folds: entanglement of style with viewpoint, and entanglement of style with head shape. To address the entanglement of style with viewpoint, we leverage the latent space of StyleNeRF [15], a neural radiance field (NeRF) [32, 33] based GAN model as an alternative for StyleGAN. StyleNeRF exhibits an innate pose-invariant latent space as the network implies the volumetric geometry of an image. We basically follow the latent optimization technique following the literature [50, 24, 51] utilizing StyleGAN. Leveraging the pose-invariant nature of the StyleNeRF's latent space, we extract a pose-invariant style vector through GAN inversion [42]. Thereafter, we recombine the volumetric features generated by the inverted style vectors of the hair and the target portraits. However, despite the pose-invariant nature of StyleNeRF's latent space, it still entails an entanglement problem between style and head shape. The optimization of a style vector to fit on the head of a target portrait, while retaining hairstyle of a hair portrait, can potentially result in deformations of the hairstyle. To address the entanglement of style with head shape, we propose a registration method that aligns the geometries of two heads within the volumetric space, while maintaining the inverted style vectors intact. Subsequently, we introduce a novel rendering method that recombines hairstyle features with facial features within the volumetric space using alignment-inducing deformable rays, which are derived from the registration outcome. In this manner, we resolve the distorted hairstyle problem from which StyleGAN-based method suffers, resulting in better alignment of the transferred hairstyle seamlessly fit on the target portrait. Taking the advantage of geometric nature of neural rendering, HairNeRF is ca" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 308, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 144 + ], + "type": "text", + "content": "pable of generating a diverse set of geometrically consistent novel view hairstyle transferred images from a single synthesis, by merely adjusting camera parameters. This is in contrast to existing StyleGAN-based hairstyle transfer methods which necessitate separate optimizations for each of novel views, which often yield inconsistent outcomes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 308, + 144, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 144, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 144, + 545, + 204 + ], + "type": "text", + "content": "Experimental results on diverse datasets demonstrate the superiority of HairNeRF against other baselines at generating seamless hairstyle transfer outcomes, while effectively preserving the underlying geometry of the original hairstyle." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 216, + 391, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 216, + 391, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 216, + 391, + 228 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 236, + 414, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 236, + 414, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 236, + 414, + 248 + ], + "type": "text", + "content": "2.1. Hairstyle Transfer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 255, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 565 + ], + "type": "text", + "content": "The proficiency exhibited by generative adversarial networks (GANs) [14, 20, 22] in generating photorealistic images has propelled a surge of research addressing image editing problem. A significant amount of research [3, 7, 19, 16, 12] has effectively addressed attribute editing or style transfer, which modify abstract-level features of an image. However, the task of hairstyle transfer still presents a challenging problem that necessitates the preservation of intricate structure and color details of a hairstyle from a source image. With the emergence of StyleGAN [20, 21, 22], which demonstrates exceptional quality in generating photorealistic facial images, efforts have been made to discover latent vectors capable of capturing transferred images of hairstyles [37, 50]. These methods require the projection of an image into the pretrained StyleGAN's latent space known as inversion, which influences the quality of transfer. Thus, the Barbershop [50] introduces an improved inversion method on the enlarged latent space named " + }, + { + "bbox": [ + 308, + 255, + 545, + 565 + ], + "type": "inline_equation", + "content": "FS" + }, + { + "bbox": [ + 308, + 255, + 545, + 565 + ], + "type": "text", + "content": "-space to capture finer details of input image. However, since these models lack awareness of the pose of given images, they suffer structural artifacts when the given two images largely differ in poses. Recent studies [24, 51] resolve this issue by aligning pose of the transferred hairstyle to the target image. However, they still suffer style deformation problem arising from the entangled latent space of StyleGAN for pose and style." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 574, + 511, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 574, + 511, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 574, + 511, + 587 + ], + "type": "text", + "content": "2.2. 3D-Aware Image Synthesis and Editing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "type": "text", + "content": "The emergence of StyleGAN [20, 21, 22] has facilitated researches on image editing [16, 41, 2, 19, 29, 23, 1], either by navigating the latent style space or leveraging semantic maps. However, StyleGAN-based image editing methods have shown inconsistency among pose changes, which is attributed to the entanglement of style and pose within StyleGAN's latent space. The development of neural radiance fields (NeRF) [32, 33] has enabled the rendering of 3D-consistent images that accommodate varying viewpoints. Consequently, methods for editing a 3D scene" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "2449" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 257 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 257 + ], + "type": "image", + "image_path": "a63f84764f186cfad92feb7f029ddb104ae386b3a56596919f3d276bc7ca157e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 546, + 289 + ], + "type": "text", + "content": "Figure 2: Overview of the hairstyle transfer pipeline of HairNeRF consisting of five modules: (a) inversion, (b) head alignment, (c) occlusion disclosure, (d) recombination, and (e) blending." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 289, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 537 + ], + "type": "text", + "content": "[44, 30, 27, 48] such as the addition and subtraction of NeRF models trained on individual objects have been explored. However, these approaches do not support modifications, such as altering the attributes or style of the object, due to the inherent nature of NeRF being trained to solely represent a single entity, thus lacking the necessary information for the desired edits. The advent of generative NeRFs [5, 38, 35, 15, 4, 10, 43] has enhanced the representational capabilities of NeRF models, enabling the expression of an expansive range of objects through the exploration of the latent space within a single model. The enhanced expressive capabilities of generative NeRFs enable the modification of attributes and style of an object [15, 4, 10]. Furthermore, within the generative NeRF modules, methods [30, 18, 40, 39, 6] have been proposed that disentangle the semantic structure and texture. However, these methods are not well-suited for local style modifications, as the latent vector representing the texture exhibits spatial entanglement, encompassing the entire regions of the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 103, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 103, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 103, + 564 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 573, + 115, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 115, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 115, + 585 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": "We follow the basic synthesis pipeline from the Barbershop [50], a founding work for hairstyle transfer [24, 50]. We encourage readers to refer [50] for the details of the original pipeline. The overall synthesis pipeline of our method consists of four modules: inversion, head alignment, recombination, and blending, as illustrated in Figure 2. Given two images, id and hair, we transfer hairstyle from the hair image into the id image. In the inversion module, we first invert the two images to obtain latent vectors and the corresponding features in the volu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 308, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 608 + ], + "type": "text", + "content": "metric space. Then, in the alignment module, we conduct registration on two geometric heads of the inverted images in the volumetric space. By aligning geometric heads, we can transfer hairstyle of the hair image into the id image to be more fit. Once the registration is done, we recombine the 3D features in the volumetric space according to semantics to render a 2D feature for synthesis using neural rendering in the recombination module. A simple method is to bring hair semantic features from the hair image, and other semantics from the id image. However, there occur missing regions which are occluded by the hair in the id image but not covered by the hair of the hair image. To fill in the missing regions, we inpaint 3D features by optimizing a latent vector, ensuring that the 2D features rendered from the 3D features generated by the optimized latent vector align with the target semantics across all viewpoints. Thereafter, we semantically integrate the 3D features from the id, hair, and the inpainting output to form a blended feature for the synthesis. The blend feature is rendered to a 2D feature, which is decoded by the upsampler blocks of StyleNeRF [15] conditioned on the blending style vector. In the final blending module, the blending style vector is optimized to make the decoded image preserve the same hairstyle as in the hair image while retaining face of the id image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 621, + 403, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 403, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 403, + 633 + ], + "type": "text", + "content": "3.2. NeRF Inversion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Semantic Parsing NeRF Head. To semantically recombine features in the 3D volume, semantics of volumetric features are required. It is known that a semantic classifier trained to classify a 2D point feature obtained by (2) can aptly classify semantic of a 3D point feature as well [25]. Based on this observation, we add an additional seman-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2450" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 277, + 162 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 277, + 162 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 277, + 162 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 277, + 162 + ], + "type": "image", + "image_path": "aba1ee09fecc380b92a6efedf7e9cb6b5e318b8210adf7dff90c330537c7706a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "type": "text", + "content": "Figure 3: The additional semantic head is added at the end of a MLP block of NeRF module, which outputs a logit of semantics in the volumetric space. The additional semantic head is trained to minimize cross-entropy between the rendered logit, " + }, + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "type": "text", + "content": ", and the ground-truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 250, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 287, + 310 + ], + "type": "text", + "content": "tic branch to the MLP block for neural rendering and train it using outputs of the off-the-shelf semantic segmentation model [46] for generated images as ground-truths as shown in Figure 3. After training, the trained semantic branch outputs a logit of semantics for a 3D point feature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 288, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 288, + 541 + ], + "type": "text", + "content": "Semantics-Aware Inversion in 3D. We invert the hair and the id images into the latent vectors for the synthesis network of the StyleNeRF [15]. The process of obtaining the latent vector generating a desired image is called GAN inversion. We utilize the pivotal tuning inversion (PTI) [36] with a little modification for inversion as the PTI can reconstruct fine details of an input image from the inverted latent vector. As our model is built upon the StyleNeRF, the output image varies as a camera parameter changes. For this reason, the inversion process requires an input camera parameter or should estimate a camera parameter in our case. We choose to provide a fixed camera parameter of an input image estimated by the off-the-shelf head pose estimator [31] for the inversion process. Using the PTI with a fixed camera parameter, we optimize " + }, + { + "bbox": [ + 46, + 314, + 288, + 541 + ], + "type": "inline_equation", + "content": "W \\in \\mathcal{W}" + }, + { + "bbox": [ + 46, + 314, + 288, + 541 + ], + "type": "text", + "content": " for each of the id and the hair images, and finetune the weights of the pretrained StyleNeRF to reconstruct the input image and the semantics estimated by the off-the-shelf semantic parsing networks [47, 45]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 555, + 146, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 146, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 146, + 567 + ], + "type": "text", + "content": "3.3. Head Alignment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 575, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 672 + ], + "type": "text", + "content": "Registration For registration, we render each of inverted images into " + }, + { + "bbox": [ + 46, + 575, + 287, + 672 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 46, + 575, + 287, + 672 + ], + "type": "text", + "content": " different viewpoints and extract 2D keypoints of the FLAME [28] head model using the off-the-shelf head pose estimator [31]. Then we estimate 3D keypoints of the FLAME model for each inverted image from the set of 2D keypoints estimated from the eight different camera views by solving a bundle-adjustment-like least square problem:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 685, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 685, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 99, + 685, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {p}} {\\operatorname {m i n i m i z e}} \\sum_ {i} ^ {n _ {c}} \\| P (\\boldsymbol {c} _ {i}, \\boldsymbol {p}) - \\boldsymbol {x} _ {i} \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "f160ba80035a061c09e6f0a81ac33c9dcd6fbaecafd31b01ff599cebf54e640b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " denotes the estimated 2D keypoints at " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "-th viewpoint, " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "-th camera parameter, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "P(\\pmb{c}_i, \\pmb{p}_i)" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " denotes the projection function projecting a 3D point " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " into 2D using " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". As we already know the camera parameters used for rendering " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " images, we exclude camera parameters from the optimization variables. We utilize the thin-plate-spline (TPS) transform [11] to find the optimal non-rigid transformation for the registration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 170, + 546, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 546, + 373 + ], + "type": "text", + "content": "Neural Rendering via Ray Deformation. The two heads of the id and the hair do not fully overlap due to the differences in characteristics, shape and size of the heads geometry, along with estimation errors in head poses. We refer to this misalignment as intrinsic misalignment, as it arises from factors unrelated to external elements such as camera parameters. To alleviate the intrinsic misalignment, the registration process finds an optimal transformation between two point sets, by which one point set is transformed to align with the other. During the neural rendering process, point features are sampled along a ray within the volumetric space, and these features are then aggregated to a single point on the image plane. If the sampling points are collected along the ray, deformed by the transformation determined through registration, we can obtain registered point features. This process leads to a registered 2d image after the aggregation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": "Figure 4 illustrates the effective neural rendering process for alignment via ray deformation. Let " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "T_{i\\rightarrow h}" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": " denote a transformation determined through registration, which aligns the point set of the id features to the point set of the hair features. As depicted in Figure 4a, sampling along rays that radiate straight from the id's camera renders an id's face with an innate head shape, colored in red, as in shown Figure 4c. Likewise, rendering from straight rays radiating through hair features " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\Phi^{\\text{hair}}" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": " using the same camera renders the hair image posed as id, with the overlaid innate head shape colored in blue, as depicted in Figure 4e. The two heads do not overlap due to differences in shape and size. However, if we render a hair image using the deformed ray transformed by " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "T_{i\\rightarrow h}" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": ", as depicted in Figure 4b, we obtain a registered hair image aligned to the head of id, as shown in Figure 4d. With a camera ray " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "r = o + td" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": ", the registered neural rendering computes a pixel value of the ray as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 335, + 594, + 545, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 594, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 335, + 594, + 545, + 623 + ], + "type": "interline_equation", + "content": "I _ {\\omega} ^ {\\text {R e g i s t e r e d}} (\\boldsymbol {r}) = \\int_ {0} ^ {\\infty} p _ {\\omega} (t) \\boldsymbol {c} _ {\\omega} (\\tilde {\\boldsymbol {r}} (t), \\tilde {\\boldsymbol {d}}) \\left| \\frac {d \\tilde {\\boldsymbol {r}}}{d t} \\right| d t, \\tag {2}", + "image_path": "a592912222aee4aa22f4ca1b373734ea90296b3d7876b46b8d0cee205f46ae95.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 619, + 527, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 619, + 527, + 643 + ], + "spans": [ + { + "bbox": [ + 312, + 619, + 527, + 643 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} p _ {\\omega} (t) = \\exp \\left(- \\int_ {0} ^ {t} \\sigma_ {\\omega} (\\tilde {\\boldsymbol {r}} (s)) \\left| \\frac {d \\tilde {\\boldsymbol {r}}}{d s} \\right| d s\\right) \\cdot \\sigma_ {\\omega} (\\tilde {\\boldsymbol {r}} (t)),", + "image_path": "180b5e16ef6dcf928eb49c9f3ced728b2bdf54310d9f07d313caafb8278a1f2d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 648, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 648, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 306, + 648, + 545, + 696 + ], + "type": "text", + "content": "with a registered ray " + }, + { + "bbox": [ + 306, + 648, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{r}} = T_{i\\rightarrow h}(\\boldsymbol {r})" + }, + { + "bbox": [ + 306, + 648, + 545, + 696 + ], + "type": "text", + "content": " , and the adjusted directional vector " + }, + { + "bbox": [ + 306, + 648, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{d}} = d\\tilde{\\pmb{r}} /dt" + }, + { + "bbox": [ + 306, + 648, + 545, + 696 + ], + "type": "text", + "content": " .Equation (2) has the same form as the original neural rendering, but with the integration computed along the transformed ray. This adjustment" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 492, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 492, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 492, + 713 + ], + "type": "text", + "content": "1 or a registered 2d feature map in the StyleNeRF [15]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2451" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 321, + 172 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 321, + 172 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 321, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 321, + 172 + ], + "type": "image", + "image_path": "6508a6e9c44b8e6b01f79c6722d561290c3143acf9e761fc61d3dd01fc15cd4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 329, + 79, + 522, + 171 + ], + "blocks": [ + { + "bbox": [ + 329, + 79, + 522, + 171 + ], + "lines": [ + { + "bbox": [ + 329, + 79, + 522, + 171 + ], + "spans": [ + { + "bbox": [ + 329, + 79, + 522, + 171 + ], + "type": "image", + "image_path": "7486f50edcfdd22c1d2ec3d87bde1bfb8b80e929cff5a632406d3f5d3fe7bab2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 75, + 174, + 319, + 271 + ], + "blocks": [ + { + "bbox": [ + 75, + 174, + 319, + 271 + ], + "lines": [ + { + "bbox": [ + 75, + 174, + 319, + 271 + ], + "spans": [ + { + "bbox": [ + 75, + 174, + 319, + 271 + ], + "type": "image", + "image_path": "e18f7cda8ee41c1cceea31f95d8ccf3c22ad62f89fd27ed09946c74e6852570f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 547, + 317 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 547, + 317 + ], + "type": "text", + "content": "Figure 4: Illustration of alignment-aware neural rendering via ray deformation. Figure (c) - (e), and (h) - (j) show overlaid heads on rendering images. Rendering via transformed ray render a registered image as shown in (d) and (i), which is equivalent to rendering with the straight ray on transformed feature represented as " + }, + { + "bbox": [ + 46, + 281, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\Phi^{\\text{hair}'}" + }, + { + "bbox": [ + 46, + 281, + 547, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 281, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\Phi^{\\text{id}'}" + }, + { + "bbox": [ + 46, + 281, + 547, + 317 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 329, + 179, + 520, + 270 + ], + "blocks": [ + { + "bbox": [ + 329, + 179, + 520, + 270 + ], + "lines": [ + { + "bbox": [ + 329, + 179, + 520, + 270 + ], + "spans": [ + { + "bbox": [ + 329, + 179, + 520, + 270 + ], + "type": "image", + "image_path": "d7e1f1332de8ae6dd9026ca8050544ece552d4bf51b6152828e839dae8e99033.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "text", + "content": "is achieved by substituting the directional vector " + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{d}}" + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "text", + "content": " and including " + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "inline_equation", + "content": "|d\\tilde{\\pmb{r}} /dt|" + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "text", + "content": " to adjust the infinitesimal length " + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "inline_equation", + "content": "dt" + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "inline_equation", + "content": "|d\\tilde{\\pmb{r}} /dt| \\approx 1" + }, + { + "bbox": [ + 46, + 336, + 288, + 385 + ], + "type": "text", + "content": " in practice, we omit the determinant term to simplify the computation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 392, + 195, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 195, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 195, + 403 + ], + "type": "text", + "content": "3.4. Volumetric Recombination" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": "Occlusion Disclosure. To inpaint the missing regions occluded by the hair in " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{hair}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " and produce the recombined 3D features " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{blend}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " as depicted in Figure 2c, we optimize the latent vector " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{inp}} \\in \\mathcal{W}+" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " within the StyleNeRF [15] framework to generate " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{inp}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": ". Analogous to [50], we produce a volumetric target semantic mask, as depicted in Figure 5. For each iteration, we optimize " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{inp}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{inp}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " that aligns with the target semantic mask " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{inp}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " from a randomly selected viewpoint. Let " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "m_{hair}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " denote a volumetric binary mask representing hair, " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " denote a camera parameter, and " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\phi|_c" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " represents 2D features rendered by the camera parameter " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": ". The target semantics " + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{inp}}" + }, + { + "bbox": [ + 46, + 411, + 287, + 568 + ], + "type": "text", + "content": " and the objectives for the optimization are constructed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 575, + 287, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 575, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 575, + 287, + 640 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {s} ^ {\\text {i n p}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}} \\circ \\boldsymbol {s} ^ {\\text {h a i r} ^ {\\prime}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}}\\right) (3) \\\\ \\circ \\Big (\\left(1 - \\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}}\\right) \\circ \\boldsymbol {s} ^ {\\mathrm {i d}} + \\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}} \\circ \\boldsymbol {s} ^ {\\mathrm {h a i r} ^ {\\prime}} \\Big), \\\\ \\mathcal {L} ^ {\\text {i n p - s e m}} = \\mathbb {E} _ {c} \\mathrm {C E} \\left(\\mathcal {R} _ {c} \\left(s ^ {\\text {i n p}}\\right), \\mathcal {R} _ {c} \\left(f _ {W ^ {\\text {i n p}}} ^ {\\text {s e m}}\\right)\\right), (4) \\\\ \\end{array}", + "image_path": "4a88452fb6a71f67b1d1c8b8b738be86e393d800ffdb1fa68deb7b5e2961cac8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 662, + 275, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 662, + 275, + 717 + ], + "spans": [ + { + "bbox": [ + 58, + 662, + 275, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} ^ {\\text {i n p - r e g}} = \\mathbb {E} _ {\\boldsymbol {c}} \\left\\| G \\left(\\left(1 - \\mathcal {R} _ {\\boldsymbol {c}} \\left(\\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {i n p}}\\right)\\right) \\circ U _ {W} \\left(\\phi^ {\\text {i n p}} \\mid_ {\\boldsymbol {c}}\\right)\\right) \\right. \\\\ \\left. - G \\Big (\\big (1 - \\mathcal {R} _ {\\boldsymbol {c} ^ {\\mathrm {i d}}} (\\boldsymbol {m} _ {h a i r} ^ {\\mathrm {i d}}) \\big) \\circ I ^ {\\mathrm {i d}} \\Big) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "3031298ba59557662cb67594f4f6d61ad9a55e213292f4723637de8e7fbd4288.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_c" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " represents the aggregation function of neural rendering according to the camera " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{sem}}" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " denotes the MLP block with semantic head, " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "U_W" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " refers to the upsampler block of StyleNeRF whose output is a final image, and " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " denotes the Gram matrix for style loss [13]. Symbols marked with an apostrophe indicate that they are registered to their counterparts, as denoted by the associated superscripts. Unlike [50], our target semantic mask is defined as a continuous function within the volumetric space, rather than as a tensor defined at discrete positions. As a result, " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{inp - sem}}" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " entails integration over this volumetric space. Hence, for each optimization iteration in practice, we compute " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{inp - sem}}" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " using a random set of rays radiating from a randomly sampled camera " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " positioned on half of a hemisphere. Combining all elements, we obtain the inpainting feature " + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{inp}}" + }, + { + "bbox": [ + 304, + 337, + 545, + 517 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 381, + 524, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 524, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 381, + 524, + 545, + 539 + ], + "type": "interline_equation", + "content": "\\Phi^ {\\text {i n p}} = \\operatorname {M L P} \\left(W ^ {\\text {i n p} ^ {*}}\\right), \\tag {5}", + "image_path": "3a9ff8fa04326a965ed2ad582e1257110dbde61aca31935c297e65f1ec40e745.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 541, + 520, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 541, + 520, + 560 + ], + "spans": [ + { + "bbox": [ + 331, + 541, + 520, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 331, + 541, + 520, + 560 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{inp}^*} = \\arg \\min_{W}\\mathcal{L}^{\\mathrm{inp - sem}} + \\lambda \\mathcal{L}^{\\mathrm{inp - reg}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Recombination. We make blend features in the volumetric space by summing up id and hair features, weighting them using binarized semantic maps representing the hair. Subsequently, the blend features are expected to be aggregated to 2D features using (2). However, the blend features contain missing regions that are occluded by the hair in the id image but are not fully covered by the hair in the hair image, leading to missing holes in the final rendered image. To mitigate this, we fill the missing regions within the blend features " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{blend}}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " by the inpainting features " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Phi^{\\mathrm{inp}}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", as obtained by (5). This blending concept is motivated by the Barbershop [50], but ours differs in that the features to be blended" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2452" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 286, + 203 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 286, + 203 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 286, + 203 + ], + "type": "image", + "image_path": "e5cc146bdbbd42b2defd699421d655b46616debcfecedc843cb580a0c8601778.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 287, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 266 + ], + "type": "text", + "content": "Figure 5: Target semantics generation for occlusion disclosure. The target semantics can be assumed to an ideal semantics for the hairstyle transferred result. The inpainting latent " + }, + { + "bbox": [ + 46, + 209, + 287, + 266 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{inp}}" + }, + { + "bbox": [ + 46, + 209, + 287, + 266 + ], + "type": "text", + "content": " is optimized to minimize to match the 2D semantics rendered from " + }, + { + "bbox": [ + 46, + 209, + 287, + 266 + ], + "type": "inline_equation", + "content": "s^{\\mathrm{inp}}" + }, + { + "bbox": [ + 46, + 209, + 287, + 266 + ], + "type": "text", + "content": " at a random viewpoint at every iteration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 304, + 157, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 157, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 157, + 316 + ], + "type": "text", + "content": "lie in the volumetric space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{id}}" + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{hair}}" + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "content": " denote rendered 2D features of id and hair, respectively, rendered using neural rendering without ray deformation. Meanwhile, " + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{id}'}" + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\phi^{\\mathrm{hair}'}" + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "content": " denote the registered 2D features of id and hair, aligned to their respective counterparts. Then we recombine features for blending as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 424, + 287, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 424, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 52, + 424, + 287, + 470 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\phi_ {i d} ^ {\\text {b l e n d}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}} \\circ \\phi^ {\\text {h a i r} ^ {\\prime}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r} ^ {\\prime}}\\right) \\\\ \\circ \\left(\\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ \\phi^ {\\mathrm {i d}} + \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}} \\circ \\phi^ {\\mathrm {i n p}}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "d3f9a52fe427e98dbdf34fd91679b26b9fefa04b7003cec510cd3e1c4a6265e4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 468, + 269, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 269, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 269, + 506 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}} = \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ \\phi^ {\\text {h a i r}} + \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}}\\right) \\\\ \\circ \\left(\\left(1 - \\pmb {m} _ {h a i r} ^ {\\mathrm {i d} ^ {\\prime}}\\right) \\circ \\phi^ {\\mathrm {i d} ^ {\\prime}} + \\pmb {m} _ {h a i r} ^ {\\mathrm {i d} ^ {\\prime}} \\circ \\phi^ {\\mathrm {i n p} ^ {\\prime}}\\right), \\\\ \\end{array}", + "image_path": "ab6319d6ccfd4b18e3242461c684046efc82b8c9ec3ff62f459b586712befbe6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 531, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 286, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 531, + 286, + 556 + ], + "type": "inline_equation", + "content": "m_{hair}" + }, + { + "bbox": [ + 47, + 531, + 286, + 556 + ], + "type": "text", + "content": " denotes a 2D binary mask for the hair semantics of " + }, + { + "bbox": [ + 47, + 531, + 286, + 556 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 531, + 286, + 556 + ], + "type": "text", + "content": ", associated with a corresponding superscript." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 590, + 111, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 111, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 111, + 602 + ], + "type": "text", + "content": "3.5. Blending" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Alignment-Aware Blending Loss. In the blending step, the upsampler blocks in the StyleNeRF [15] decode the blend features obtained by (6) into an output image in high resolution, conditioned on a latent vector " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "W^{\\text{blend}}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ". We optimize the latent vector " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "W^{\\text{blend}}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " to ensure that the decoded image retains the facial identity from the id image while seamlessly transferring hairstyle of the hair image. To this end, we design the objective function for the optimization to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 510, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 510, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 510, + 84 + ], + "type": "text", + "content": "include face-preserving and hair-preserving terms." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 91, + 509, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 91, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 341, + 91, + 509, + 121 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {f a c e}} ^ {\\mathrm {l p i p s}} = \\mathcal {L} ^ {\\mathrm {l p i p s}} \\left(\\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\mathrm {i d}} ^ {\\text {b l e n d}}\\right), \\right. \\\\ \\left(1 - \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\mathrm {i d}}\\right) \\circ I ^ {\\mathrm {i d}}\\bigg), \\\\ \\end{array}", + "image_path": "ceb315b11f900385cb5c32de4dd150c35f1873c5ef5878e46af28e0166a32992.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 122, + 545, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 122, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 342, + 122, + 545, + 152 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {h a i r}} ^ {\\mathrm {l p i p s}} = \\mathcal {L} ^ {\\mathrm {l p i p s}} \\left(\\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}}\\right), \\right. \\\\ \\left. \\boldsymbol {m} _ {\\text {h a i r}} ^ {\\text {h a i r}} \\circ I ^ {\\text {h a i r}}\\right), \\tag {7} \\\\ \\end{array}", + "image_path": "42c54b7f7b52170f3b164ba6e609828a66e3bacd9133862c8ad9b822be9d653e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 153, + 484, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 153, + 484, + 194 + ], + "spans": [ + { + "bbox": [ + 342, + 153, + 484, + 194 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {h a i r} ^ {\\text {s t y l e}} = \\left\\| G \\left(\\boldsymbol {m} _ {h a i r} ^ {\\text {h a i r}} \\circ U _ {W ^ {\\text {b l e n d}}} \\left(\\phi_ {\\text {h a i r}} ^ {\\text {b l e n d}}\\right)\\right) \\right. \\\\ - \\left. G \\left(\\boldsymbol {m} _ {h a i r} ^ {\\text {h a i r}} \\circ I ^ {\\text {h a i r}}\\right) \\right\\| _ {2} ^ {2}, \\\\ \\end{array}", + "image_path": "aa0cd9e24bba429746b97718f39c84f4f47377aea716b424679677976a2cef5b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "inline_equation", + "content": "m_{face}^{\\mathrm{id}}" + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": " denotes the binary mask for face region of the face image, " + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "inline_equation", + "content": "m_{hair}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": " denotes the binary mask for hair region of the hair image, " + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "inline_equation", + "content": "U_W" + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": " denotes the upsampler conditioned on " + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{lpips}}" + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": " denotes the perceptual loss [49], and " + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": " denotes the Gram matrix for the style loss [13]. Combining all elements, the optimal blend vector " + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{blend*}}" + }, + { + "bbox": [ + 304, + 201, + 545, + 275 + ], + "type": "text", + "content": " is obtained as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 281, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 281, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 312, + 281, + 545, + 300 + ], + "type": "interline_equation", + "content": "W ^ {\\text {b l e n d} *} = \\arg \\min _ {W} \\lambda_ {f a c e} \\mathcal {L} _ {f a c e} ^ {\\mathrm {l p i p s}} + \\lambda_ {h a i r} \\mathcal {L} _ {h a i r} ^ {\\mathrm {l p i p s}} + \\mathcal {L} _ {h a i r} ^ {\\text {s t y l e}}. \\tag {8}", + "image_path": "f0f726d816093342d8d4aab3d2c8711236023da2787e36317c01a58ba9118af8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 307, + 432, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 432, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 432, + 320 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 327, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 590 + ], + "type": "text", + "content": "We evaluate our method on two distinct tasks following [24]: hairstyle transfer and reconstruction. In the hairstyle transfer task, each input pair consists of different identities, and the goal is to transfer the hairstyle from the hair image to the face of the id image. In the reconstruction task, the input pair shares the same identity but with different poses. Since the id and hair images share the same identity in the reconstruction task, the id serves as the ground-truth for the evaluation. We use Flicker-Faces-HQ (FFHQ) dataset [21] and faces images crawled from Unsplash (UF), as used in [50, 51], for the hairstyle transfer task. For the reconstruction task, we employ the VoxCeleb2 [9] dataset. To evaluate the overall synthesis quality of the generated images, we measure the Frèchet Inception distance (FID) [17], the naturalness image quality estimator (NIQE) [34], and the precision and recall [26]. In the hairstyle transfer task, we synthesize 1,000 images using the FFHQ dataset and 380 images using the Unsplash dataset. For the reconstruction task, we synthesize 500 images. To demonstrate the effectiveness of HairNeRF at preserving geometric structure, we provide both qualitative and quantitative comparisons of our method against other baselines." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 597, + 417, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 417, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 417, + 610 + ], + "type": "text", + "content": "4.1. Comparison Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": "We compare our method to LOHO [37], the Barbershop [50], and StyleYourHair [24] on the FFHQ and VoxCeleb2 datasets. For the hairstyle transfer task, we include a comparison with HairNet [51] using only the UF dataset." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "Hairstyle Transfer. Quantitative results on FFHQ and UF are summarized in Table 1. Notice that LOHO, the Barbershop and StyleYourHair is built upon StyleGAN2, while HairNeRF utilizes StyleNeRF. Both StyleGAN2 and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2453" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 70, + 293, + 350 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 293, + 350 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 293, + 350 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 293, + 350 + ], + "type": "image", + "image_path": "3fe5926f25d98469704a9a656e80500a8747a43b5a665825646f2f5e2fc4e7ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 357, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 47, + 357, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 545, + 380 + ], + "type": "text", + "content": "Figure 6: Comparison of results for FFHQ datasets. Both StyleYourHair and our method align the hair even if the poses between the input images are different, but our method preserves the direction of the hair better (first and the last rows)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 300, + 70, + 521, + 348 + ], + "blocks": [ + { + "bbox": [ + 300, + 70, + 521, + 348 + ], + "lines": [ + { + "bbox": [ + 300, + 70, + 521, + 348 + ], + "spans": [ + { + "bbox": [ + 300, + 70, + 521, + 348 + ], + "type": "image", + "image_path": "d63ce7d1cc4a9afd84affb13c6e669f713a6e79a6cc22da133745380870906ab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 389, + 298, + 605 + ], + "blocks": [ + { + "bbox": [ + 50, + 389, + 298, + 605 + ], + "lines": [ + { + "bbox": [ + 50, + 389, + 298, + 605 + ], + "spans": [ + { + "bbox": [ + 50, + 389, + 298, + 605 + ], + "type": "image", + "image_path": "0af044fbc8e4bddaed61aa8951135517bda95d8266ebe532ae3c9045a43380c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 612, + 546, + 635 + ], + "lines": [ + { + "bbox": [ + 46, + 612, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 546, + 635 + ], + "type": "text", + "content": "Figure 7: Comparison of results for Unsplash datasets. Our method preserves identity better (second row on the left), and hair transfer is performed to fit the structure of identity (third row on the left)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 298, + 389, + 544, + 604 + ], + "blocks": [ + { + "bbox": [ + 298, + 389, + 544, + 604 + ], + "lines": [ + { + "bbox": [ + 298, + 389, + 544, + 604 + ], + "spans": [ + { + "bbox": [ + 298, + 389, + 544, + 604 + ], + "type": "image", + "image_path": "c2caad9fef51e9982cada4fb16a2b99a54746d738b11b589b273813d2f55f342.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 655, + 288, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 288, + 704 + ], + "type": "text", + "content": "StyleNeRF are effective in generating high-quality face images; however, it has been reported that StyleGANs slightly outperforms NeRF-based model in terms of commonly used metrics, including FID. Consequently, the overall metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 655, + 546, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 655, + 546, + 704 + ], + "spans": [ + { + "bbox": [ + 305, + 655, + 546, + 704 + ], + "type": "text", + "content": "for HairNeRF are slightly lower than the other baselines, with the exception of NIQE. Nevertheless, this difference is not substantial, and the qualitative results of HairNeRF reveal perceptually better quality in comparison to the other" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2454" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 281, + 131 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 281, + 131 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 281, + 131 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 281, + 131 + ], + "type": "table", + "html": "
FFHQ
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO40.650.830.850.551.3491.42
Barbershop12.240.770.930.831.0648.00
StyleYourHair12.000.550.930.771.1151.36
HairNet11.630.760.970.591.1456.45
HairNeRF9.570.700.890.671.4057.26
", + "image_path": "c972c4f3bb642c90a89432d613ec0161794f6087416e657b2f26dce4779f16e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 137, + 281, + 190 + ], + "blocks": [ + { + "bbox": [ + 53, + 137, + 281, + 190 + ], + "lines": [ + { + "bbox": [ + 53, + 137, + 281, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 281, + 190 + ], + "type": "table", + "html": "
UF
MethodNIQE↓ArcFace↑Precision↑Recall↑Realism↑FID↓
LOHO42.370.800.910.451.0786.16
Barbershop12.130.740.940.841.9141.06
StyleYourHair12.070.750.940.802.1742.62
HairNeRF10.090.700.890.571.7149.74
", + "image_path": "c287d9fc9c7556f4da607cf8019346916f4a36508687b38df097ec61d6f11733.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 255, + 284, + 355 + ], + "blocks": [ + { + "bbox": [ + 52, + 255, + 284, + 355 + ], + "lines": [ + { + "bbox": [ + 52, + 255, + 284, + 355 + ], + "spans": [ + { + "bbox": [ + 52, + 255, + 284, + 355 + ], + "type": "image", + "image_path": "c6d6212eb26c1f434a103602fdc28644764c7b5a47dcafd432e02198a6be4610.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 363, + 268, + 374 + ], + "lines": [ + { + "bbox": [ + 65, + 363, + 268, + 374 + ], + "spans": [ + { + "bbox": [ + 65, + 363, + 268, + 374 + ], + "type": "text", + "content": "Figure 8: Comparison examples for reconstruction task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 399, + 289, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 614 + ], + "type": "text", + "content": "basielines. Figures 6 and 7 show synthesized examples using the FFHQ and UF datasets, respectively. As observed in the fifth row of Figure 6b and Figure 7, the Barbershop generates fine-quality images when the poses of the id and hair images are similar. However, the output quality substantially declines when the difference in pose between the two input images increases. As shown in the second and the third row of Figure 6a, substantial structural artifacts, such as bumps and missing holes, appear in the transferred hairstyle images synthesized by the Barbershop. In contrast, the transferred hairstyle of HairNeRF seamlessly fits the head of the id image. StyleYourHair and HairNet also generate satisfactory results for input pairs with significant pose differences. However, as seen in the fifth row of Figure 6b, the detailed structure of the hairstyle often changes in the synthesized results. Furthermore, they often elongate the length of the forehead, as shown in the fourth row of Figure 7a." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Reconstruction. Following [24], we evaluate reconstruction quality on the VoxCeleb2 dataset, consisting of pairs of images with same identity but different poses. Figure 8 shows an example of reconstruction task. Ideally, the synthesis result should be identical to the id image, since the id and hair images represent the same person. Figure 8 shows the reconstruction results of HairNeRF in comparison to other baselines. While the reconstructed image pre" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 535, + 117 + ], + "blocks": [ + { + "bbox": [ + 46, + 208, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 242 + ], + "type": "text", + "content": "Table 1: Hairstyle transfer results on FFHQ and UF datasets. HairNeRF shows comparable results to StyleGAN2-based hairstyle transfer models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 535, + 117 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 535, + 117 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 535, + 117 + ], + "type": "table", + "html": "
MethodMSE↓SSIM↑LPIPS↓PSNR↑
Barbershop [50]1.1650.8010.12731.44
StyleYourHair [24]1.2180.7950.15531.34
HairNeRF0.9570.8210.10631.84
", + "image_path": "ed15583d18a57302418099fda2f384602f4a9f2c1d739906f36c76d4659f6ba5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 134, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 134, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 134, + 545, + 178 + ], + "type": "text", + "content": "Table 2: Reconstruction result. HairNeRF surpasses other baselines in view of MSE, SSIM, LPIPS, and PSNR. The result verifies the effectiveness of HairNeRF at preserving hairstyle geometry compared to the other baselines." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 205, + 544, + 432 + ], + "blocks": [ + { + "bbox": [ + 310, + 205, + 544, + 432 + ], + "lines": [ + { + "bbox": [ + 310, + 205, + 544, + 432 + ], + "spans": [ + { + "bbox": [ + 310, + 205, + 544, + 432 + ], + "type": "image", + "image_path": "dc8b20d7421c491836d3bd1672716ca5a40a55ae97739a9d793f871fd360a977.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 439, + 506, + 451 + ], + "lines": [ + { + "bbox": [ + 345, + 439, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 345, + 439, + 506, + 451 + ], + "type": "text", + "content": "Figure 9: Examples of novel view synthesis." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 488, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 547, + 548 + ], + "type": "text", + "content": "serves the detailed geometry of the hair, both the Barbershop and StyleYourHair show slight alterations. The quantitative results, summarized in Table 2, further validate the superiority of HairNeRF in preserving the geometry of hairstyles." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "Novel View Synthesis. Owing to the geometric nature of neural rendering, HairNeRF can generate diverse pose varying images from the same latent vector once optimized for synthesis. Figure 9 shows examples of novel view images obtained by varying the camera parameters. For a qualitative comparison, we generate a diverse set of novel view images of the id image using the inversion result of HairNeRF. We then synthesize each novel view image using the Barbershop and StyleYourHair from the given hair image. As shown in Figure 9, the transferred hairstyle varies by views in the case of baselines, whereas HairNeRF preserves a consistent hair structure that aligns with the original hair image." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2455" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 282, + 254 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 282, + 254 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 282, + 254 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 282, + 254 + ], + "type": "image", + "image_path": "5145b4a2ec1b267bee3fe20905dd1ad1779f33069e022f74859f69647c2f5951.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 261, + 288, + 307 + ], + "lines": [ + { + "bbox": [ + 47, + 261, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 288, + 307 + ], + "type": "text", + "content": "Figure 10: Effectiveness of registration shown with the underlying geometry. Without registration the transferred hairstyle does not perfectly fit to the face (first row) or erode the head (third row) of the id image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 328, + 260, + 365 + ], + "blocks": [ + { + "bbox": [ + 75, + 328, + 260, + 365 + ], + "lines": [ + { + "bbox": [ + 75, + 328, + 260, + 365 + ], + "spans": [ + { + "bbox": [ + 75, + 328, + 260, + 365 + ], + "type": "table", + "html": "
MethodMSE↓SSIM↑LPIPS↓PSNR↑
w/o regi.0.9800.8180.10831.77
w/ regi.0.9570.8210.10631.84
", + "image_path": "cd0cf3d732a7902aecdbf1de8fad9a0b5832a531e0cbd9209fa23ddb57322558.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 383, + 287, + 416 + ], + "lines": [ + { + "bbox": [ + 47, + 383, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 287, + 416 + ], + "type": "text", + "content": "Table 3: Reconstruction quality with and without head alignment module. With registration, the reconstruction quality increases in view of MSE, SSIM, LIPIS, and PSNR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 449, + 140, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 140, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 140, + 462 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 289, + 715 + ], + "type": "text", + "content": "Effectiveness of Registration. To validate the effectiveness of the head alignment module, we evaluate the reconstruction quality using the VoxCeleb2 dataset. Table 3 summarizes the results. As mentioned in Section 3.3, StyleNERF suffers from an intrinsic misalignment problem that results in subtle differences, even if the pair of input images shares the same identity. However, the registration process within the head alignment module reduces the existing geometric differences to make transferred hairstyle more fit to the id image. Table 3 shows that the reconstruction quality deteriorates when the registration module is missing. Figure 10 shows the effectiveness of registration with the underlying geometry. The first example shows that the shape of the original hair is slightly larger than the head of id image. Without registration, the transferred hair sticks out to the right side of the head of the id. However, with registration, the transferred hair fits more closely to the head of the id. The difference is clearly shown in the overlap image. In the second example, it is clearly shown that the transferred hair is stretched to fit on the head of id by registration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 379, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 379, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 379, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 342 + ], + "type": "text", + "content": "We introduce HairNeRF, a novel geometry-aware hairstyle transfer method designed to address the entanglement issue of pose and style attributes within the latent space of StyleGAN. By leveraging NeRF-based generative models, where the pose of the rendered image is determined by external camera parameters, we achieve a complete disentanglement of pose from the style latent space. Additionally, we propose an efficient method for recombining features within the volumetric space using semantics. This is achieved by aligning two distinct innate head structures via ray deformation. This method unveils the potential to seamlessly integrate multiple NeRF-represented objects of varying shapes and sizes into one unified object, without requiring additional modification to NeRF weights. Using the intrinsic geometric properties of neural rendering, our method can produce view-varied images from a single synthesis process without inducing geometric deformation. The efficacy of HairNeRF in preserving the geometric structure of the original hairstyle, compared to other baselines, is substantiated both qualitatively and quantitatively through comprehensive experiments on diverse datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 351, + 365, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 365, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 365, + 363 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 371, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 310, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 310, + 371, + 545, + 403 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy J. Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 311, + 404, + 545, + 437 + ], + "type": "text", + "content": "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Int. Conf. Comput. Vis., 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 438, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 438, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 311, + 438, + 545, + 482 + ], + "type": "text", + "content": "[3] Guha Balakrishnan, Raghudeep Gadde, Aleix Martinez, and Pietro Perona. Rayleigh eigendirections (reds): Nonlinear GAN latent space traversals for multidimensional features. In Eur. Conf. Comput. Vis., 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 311, + 482, + 545, + 547 + ], + "type": "text", + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay andgu2021 stylenerf Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 311, + 548, + 545, + 592 + ], + "type": "text", + "content": "[5] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 592, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 592, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 311, + 592, + 545, + 634 + ], + "type": "text", + "content": "[6] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. ACM Trans. Graph., 41(1):1:1-1:26, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 635, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 311, + 635, + 545, + 669 + ], + "type": "text", + "content": "[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In Eur. Conf. Comput. Vis., 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 311, + 669, + 545, + 712 + ], + "type": "text", + "content": "[8] Min Jin Chong, Wen-Sheng Chu, Abhishek Kumar, and David A. Forsyth. Retrieve in style: Unsupervised facial feature transfer and retrieval. In Int. Conf. Comput. Vis., 2021. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2456" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 73, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "type": "text", + "content": "[9] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In B. Yegnanarayana, editor, Conf. of the Int. Speech Comm. Assoc., 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[10] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. GRAM: generative radiance manifolds for 3d-aware image generation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 154, + 286, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 154, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 49, + 154, + 286, + 219 + ], + "type": "text", + "content": "[11] Jean Duchon. Splines minimizing rotation-invariant semi-norms in sobolev spaces. In Walter Schempp and Karl Zeller, editors, Constructive Theory of Functions of Several Variables: Proceedings of a Conference Held at Oberwolfach, Germany, April 25 - May 1, 1976, volume 571 of Lecture Notes in Mathematics, pages 85-100, 1976. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 221, + 286, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 221, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 49, + 221, + 286, + 264 + ], + "type": "text", + "content": "[12] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Trans. Graph., 41(4):141:1-141:13, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 266, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 266, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 49, + 266, + 286, + 300 + ], + "type": "text", + "content": "[13] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 301, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 301, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 49, + 301, + 286, + 345 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 347, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 347, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 49, + 347, + 286, + 390 + ], + "type": "text", + "content": "[15] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Int. Conf. Learn. Represent., 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 393, + 286, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 393, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 393, + 286, + 426 + ], + "type": "text", + "content": "[16] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable GAN controls. In Adv. Neural Inform. Process. Syst., 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 428, + 286, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 428, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 49, + 428, + 286, + 472 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Adv. Neural Inform. Process. Syst., 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 474, + 286, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 474, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 474, + 286, + 506 + ], + "type": "text", + "content": "[18] Kaiwen Jiang, Shu-Yu Chen, Feng-Lin Liu, Hongbo Fu, and Lin Gao. Nerffaceediting: Disentangled face editing in neural radiance fields. In SIGGRAPH Asia, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 508, + 286, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 508, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 49, + 508, + 286, + 541 + ], + "type": "text", + "content": "[19] Omer Kafri, Or Patashnik, Yuval Alaluf, and Daniel Cohen-Or. Stylefusion: A generative model for disentangling spatial segments. ACM Trans. Graph., 41(5):197:1-197:15, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 543, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 543, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 543, + 286, + 576 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. IEEE Trans. Pattern Anal. Mach. Intell., 43(12):4217-4228, 2021. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 624, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 286, + 667 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "type": "text", + "content": "[23] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in GAN for real-time image editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 852-861, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[24] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. In Eur. Conf. Comput. Vis., 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[25] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Adv. Neural Inform. Process. Syst., 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Adv. Neural Inform. Process. Syst., 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[27] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. In Winter Conf. on Appl. of Comput. Vis., 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "text", + "content": "[28] Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194:1-194:17, 2017. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "text", + "content": "[29] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. pages 16331–16345, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Int. Conf. Comput. Vis., 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[31] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriya Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis., 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "text", + "content": "[33] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "text", + "content": "[34] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Sign. Process. Letters, 20(3):209-212, 2013. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[35] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 613 + ], + "type": "text", + "content": "[36] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 42(1):1-13, 2022. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[37] Rohit Saha, Brendan Duke, Florian Shkurti, Graham W Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[38] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Adv. Neural Inform. Process. Syst., 2020. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2457" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 643 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[39] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. IDE-3D: interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Trans. Graph., 41(6):270:1-270:10, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[40] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. Graph., 40(4):133:1-133:14, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[42] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(3):3121-3138, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "text", + "content": "[43] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. GRAM-HD: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[44] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for edible scene rendering. In Int. Conf. Comput. Vis., 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "type": "text", + "content": "[45] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet V2: bilateral network with guided aggregation for real-time semantic segmentation. Int. J. Comput. Vis., 129(11):3051-3068, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 430 + ], + "type": "text", + "content": "[46] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "type": "text", + "content": "[47] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Eur. Conf. Comput. Vis., 2018. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "text", + "content": "[48] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: Geometry editing of neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 564 + ], + "type": "text", + "content": "[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[50] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Trans. Graph., 40(6):215:1-215:13, 2021. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[51] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Eur. Conf. Comput. Vis., 2022. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2458" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_content_list.json b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5af3aca6ae63063b31fed601b27ae57d1679c986 --- /dev/null +++ b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "Hallucination Improves the Performance of Unsupervised Visual Representation Learning", + "text_level": 1, + "bbox": [ + 81, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jing Wu1 Jennifer Hobbs2 Naira Hovakimyan1 \n1University of Illinois Urbana-Champaign, 2Intelinair \n{jingwu6,nhovakim} @illinois, jennifer hobbs08+research@gmail.com", + "bbox": [ + 210, + 202, + 718, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive learning models based on Siamese structure have demonstrated remarkable performance in self-supervised learning. Such a success of contrastive learning relies on two conditions, a sufficient number of positive pairs and adequate variations between them. If the conditions are not met, these frameworks will lack semantic contrast and be fragile on overfitting. To address these two issues, we propose Hallucinator that could efficiently generate additional positive samples for further contrast. The Hallucinator is differentiable and creates new data in the feature space. Thus, it is optimized directly with the pre-training task and introduces nearly negligible computation. Moreover, we reduce the mutual information of hallucinated pairs and smooth them through non-linear operations. This process helps avoid over-confident contrastive learning models during the training and achieves more transformation-invariant feature embeddings. Remarkably, we empirically prove that the proposed Hallucinator generalizes well to various contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Under the linear classification protocol, a stable accuracy gain is achieved, ranging from $0.3\\%$ to $3.0\\%$ on CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. The improvement is also observed in transferring pre-train encoders to the downstream tasks, including object detection and segmentation.", + "bbox": [ + 76, + 325, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 737, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the recent computer vision community, there has been rapid progress in self-supervised learning (SSL), gradually closing the performance gap with supervised learning [23, 29, 10, 7, 51]. Among the diverse approaches of SSL, contrastive learning, such as MoCoV1&V2 [29, 11], SimCLR [10], and SimSiam [12], shows promising results. Generally, contrastive learning treats each image as one class which will be augmented into two separate views. These two views form one positive pair and should ide", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86a2670cb8c7046884d7594189c2126b13050df24326ef0a1acc9d0e687eed79.jpg", + "image_caption": [ + "Figure 1: The motivation of the proposed hallucination methods. Given one pair of images with the same semantic meaning, such as a pair of horses, a person can envision further similar pairs by imagining one of the horses in different poses and surroundings. If a contrastive learning model could do such hallucination, it could have additional novel pairs to contrast given the same data. Note that this hallucination process is for illustration only. In the implementation, all the hallucinated samples are computed in the feature space." + ], + "image_footnote": [], + "bbox": [ + 540, + 292, + 854, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ally be close if mapped to feature space. With sufficient contrast in the feature space, contrastive learning models show a strong capacity to learn transformation-invariant features that are transferable to various downstream tasks, such as classification, object detection, and segmentation [18, 30, 39].", + "bbox": [ + 496, + 718, + 893, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To ensure sufficient contrast, researchers from previous work address the issue from two essential practices, either introducing large amounts of positive pairs or adding additional variants&transformation among them. For example, SimCLR uses a batch size that generates thousands of positives to facilitate the convergence of models [10]. Work", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16132", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from [51, 9] reduces mutual information of positive pairs using stronger data augmentation, i.e., color distortion and jigsaw transformation. Likewise, the work from [44] introduces ContrastiveCrop, and the work from [49] proposes Un-Mix, respectively, to reduce the similar semantic meaning of sample pairs in the original image space. Beyond the data augmentation and image operations, researchers from [68] propose to apply a linear operation to generate hard positive samples in feature space.", + "bbox": [ + 75, + 90, + 468, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the success of prior approaches, we argue that large batch sizes are not always achievable. Meanwhile, all proposed techniques only focus on improving the original pairs. Given one positive pair of positive samples, humans are born with the amazing ability to come up with additional positives by imagining a sample from different surroundings and perspectives without much effort, as demonstrated in Figure 1. This process of self-imagination, in turn, will benefit the human neurological system, improving recognition capacity [24]. Similarly, if we could empower contrastive learning models with the ability to hallucinate or imagine an object to a novel view, additional positive pairs could be provided for the learning tasks.", + "bbox": [ + 75, + 229, + 470, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, exploring feasible methods to hallucinate novel positive pairs is challenging. Firstly, while generative models produce realistic images that could form additional positive views [22, 1, 42, 4], realistic data do not necessarily benefit learning tasks [55]. More importantly, applying these approaches forces us to fall back into a computational dilemma to the previous method. In other words, image-level hallucination still suffers from expensive computation as we still need to encode the hallucinated images into feature space. Lastly, if the generated positive pairs are similar to each other, training a discriminative model would be too trivial, thus showing poor generalization capacity [44, 51, 68].", + "bbox": [ + 75, + 428, + 468, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, our key insight is that the sample-generation process should aim for three critical elements: (i) feature-space operation (ii) sufficient variance of positive pairs (iii) a differentiable module optimized directly related to the learning task. To achieve this, we propose Hallucinator to improve the performance of contrastive learning with Siamese structures. The Hallucinator is plugged in after the encoder to manipulate feature vectors and improve the feature-level batch size for further contrast. To ensure adequate variance is introduced, we propose an asymmetric feature extrapolation method inspired by the work from [68]. More importantly, we present a non-linear hallucination process for the extrapolated samples. Such a process is differentiable (i.e. learnable), therefore essentially boosting Hallucinator to generate smooth and task-related features.", + "bbox": [ + 75, + 626, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed Hallucinator delivers extra positives and simultaneously enlarges the variance between newly introduced pairs. Moreover, this approach only relies on pos", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "itive samples. Therefore, it can be easily applied to any Siamese structure by adding it after the encoders as a plug-and-play module. Without the tedious exploration of hyperparameters and much additional computation, we empirically prove the effectiveness of the proposed Hallucinator on popular contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. We notice a stable improvement ranging from $0.3\\%$ to $3.0\\%$ under the linear classification protocol, crossing the CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. We also observe that models trained with Hallucinator show better transferability in downstream tasks like object detection and segmentation.", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 517, + 272, + 844, + 285 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We investigate a critical yet under-explored aspect of contrastive learning: introducing additional positive pairs with further variation in feature space.", + "- To the best of our knowledge, this is the first attempt to incorporate the concept of \"Hallucination\" into contrastive representation learning. We propose Hallucinator to realize this idea, which effectively generates smooth and less similar positive feature vectors. The Hallucinator is simple, effective, and agnostic to contrastive frameworks.", + "- We empirically illustrate that the proposed approach significantly benefits the various contrastive learning models within multiple datasets." + ], + "bbox": [ + 516, + 287, + 890, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 496, + 648, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce related literature on contrastive learning and hallucination techniques.", + "bbox": [ + 500, + 521, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Contrastive Learning", + "text_level": 1, + "bbox": [ + 500, + 560, + 700, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key idea of contrastive learning is to minimize the distance of positive pairs and repulse negative pairs in the feature space [27]. This idea has been successfully applied to unsupervised visual representation tasks, showing promising results in various downstream tasks [10, 29, 11, 23, 12, 2, 32, 17, 54, 41, 43, 50, 57, 60, 63, 61, 8, 37]. One of the breakthroughs of contrastive learning models is SimCLR [10], which introduces a simple but effective visual representation learning method. Without negative pairs, SimCLR learned transformation-invariant representation with a large batch size. Contemporaneous impressive work is MoCo from [29]. To ensure MoCo can be trained smoothly with computational-friendly batch sizes, the authors of MoCo propose a memory bank to store negative features and momentum-updated backbones. After this, SimSiam [12] presents a Siamese-based network that can learn high-quality representation with stop-gradient, successfully avoiding the collapse of contrastive learning models. Other designs of contrastive learning rely on an online network to predict the output of the target network or contrasting cluster assignments [23, 7].", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16133", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/88923fe55a1f0cd92cd9e45f481778611b4102af048527c49f5505c4b9a2ebe0.jpg", + "image_caption": [ + "Figure 2: Illustration of contrastive learning (MoCoV2 [11]) with Hallucinator. The Hallucinator is added after the backbones and projector for feature-level manipulation. While Hallucinator feeds original feature vector $q$ forward, it additionally provides hallucinated feature $\\hat{q}$ for further contrast." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further boost the models' performance, researchers explore diverse ways to hard positives and reduce the mutual information of positive samples so that more transformation-invariant embeddings can be learned [51, 48, 45, 13]. For image-level enhancement, typical works focus on improving the data augmentation in pixel space using Mixup, color distortion, or jigsaw transformation [49, 51, 9]. More recent work is presented by [44] using ContrastiveCrop, which creates better positive views through the localization box. For feature-level operation, the authors from [34] prove the effectiveness of hard negative samples. More recent and close work to this paper is [68], which applies symmetric extrapolation to create hard positives but introduces no non-linearity. Importantly, instead of replacing the original positive samples, this paper proposes a differentiable hallucinator to generate an additional positive sample with less mutual information and better smoothness. Such a setting ensures maximal and adaptive contrast during the training, generalizing well to diverse contrastive learning models.", + "bbox": [ + 75, + 378, + 472, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Hallucination", + "text_level": 1, + "bbox": [ + 76, + 694, + 218, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hallucination is initially proposed to solve the scarcity of data in the classification task [28]. Then, this idea is kept updated and applied in different areas [46, 55, 66, 67, 6, 65, 25, 47], such as object detection, aerial navigation, skeleton-based action recognition, and face generation. While image-level hallucination benefits few-shot recognition by synthesis of novel view [3] or introducing random noises [55], most of the work applies hallucination in the feature space. The work from [28] generates novel class features by transforming shared features in base classes. Authors of [67] build a hallucination framework in the region of interest feature space object to enhance object", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "detection performance. More recent work shows that this hallucination mechanism also benefits 3D human pose estimation by generating novel motion sequences [21]. While hallucination is effective in different learning tasks, to the best of our knowledge, the performance and application of hallucination in SSL are fully unexplored.", + "bbox": [ + 496, + 378, + 893, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Feature-Level Augmentation", + "text_level": 1, + "bbox": [ + 498, + 483, + 756, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hallucination relies on effective feature-level augmentation or manipulations. The primary goal of feature augmentation is to extend the limited labeled dataset without relying on expensive computation such as auto-encoder[33], Generative Adversarial Networks[40] or simulation tools[52, 58, 5]. For instance, in the work by [16], a task-agnostic feature augmentation approach is proposed to enrich training data with minimal additional computation. The authors of [36] also explore similar ideas in the domain of few-shot learning. Building upon this, the concept is adapted to sentence representation learning [62, 19] and few-shot learning tasks in remote sensing[56]. More recently, [38] applies feature augmentation based on a meta-learning technique.", + "bbox": [ + 496, + 508, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 722, + 591, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the overall process of hallucination for contrastive representation learning in Section 3.1. Secondly, we highlight the center cropping method we used, which is crucial to effective hallucination or generation of new samples in Section 3.2. Then, we introduce the Hallucinator incorporated into our contrastive models in Section 3.3. Finally, we visualize and discuss the critical properties of the hallucination method from two perspectives in Section 3.4: the similarity of positive samples and the uniformity of the feature distribution.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16134", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The Overall Pipeline", + "text_level": 1, + "bbox": [ + 76, + 90, + 272, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking MoCo [11] as an example, we illustrate how a Hallucinator can be plugged into a contrastive learning model in Figure 2. Our architecture takes one image $x$ as input. Then, the input $x$ is augmented into two views $x_{1}$ and $x_{2}$ . Each view will be processed by an encoder consisting of a backbone (e.g., ResNet) and a projector (e.g., an MLP head). After the encoders, output vectors $q$ and $k$ are obtained, forming one positive pair $(q, k)$ . Then, this positive pair $(q, k)$ is fed to a Hallucinator. Notably, Hallucinator is only added to one branch of the framework to generate an additional positive feature $\\hat{q}$ . Together with feature vector $k$ , $\\hat{q}$ and $k$ form an extra positive pair $(\\hat{q}, k)$ during the training. Based on different contrastive learning models, the loss functions keep intact, and the average loss of these two positive pairs is computed for back-propagation. The same paradigm could be applied to SimCLR, SiamSiam and other contrastive learning models. We illustrate further details about pipelines and loss functions of other models in this paper in the Supplementary Material (Section 1.1).", + "bbox": [ + 75, + 112, + 473, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Center Cropping", + "text_level": 1, + "bbox": [ + 76, + 409, + 245, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrastive learning, data augmentations aim to ensure the performance of pre-trained representations invariant to nuisances. Among all these methods, random crop plays the most critical role in all the contrastive learning models. Generally, views (cropped tiles) generated by random cropping are diversified, successfully covering all the semantic information over the whole image. However, such a cropping method is likely to generate false positives patches [44]. In other words, patches randomly cropped from the original images do not necessarily share the overlapped pixels and sufficient common information. Therefore, these false positive pairs may be fooling models during training, causing representations to be sub-optimal. Importantly, the issue will be exacerbated if we generate further hallucinated samples based on false positive pairs, which misleads the overall training beyond the sweet spot.", + "bbox": [ + 75, + 433, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle this issue, we first apply center cropping $\\mathbb{C}_{crop}$ to the original image, getting a relatively smaller image $\\hat{I}_{x,y}$ . Then, random cropping $\\mathbb{R}_{crop}$ is applied to $\\hat{I}_{x,y}$ . More specifically, the center cropping can be formulated as", + "bbox": [ + 75, + 674, + 468, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {x, y} = \\mathbb {C} _ {\\text {c r o p}} \\left(I _ {x, y}, p\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 744, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{x,y}$ is the input image with $(x,y)$ as the coordinate of the images' center. After center cropping, we keep the center of $\\hat{I}_{x,y}$ unchanged. Meanwhile, with the original shape of $I_{x,y}$ defined as $(h,w)$ , we define the shape of cropped image $\\hat{I}_{x,y}$ as $(\\hat{h},\\hat{w})$ . The $p$ denotes a ratio of cropped length over the original length, i.e., $p = \\frac{\\hat{w}}{w} = \\frac{\\hat{h}}{h}$ . Only if particularly mentioned, we set $p = 0.5$ for all experiments in this paper.", + "bbox": [ + 75, + 773, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77c72261a5c95c0e15c94fa6898ad99109334763745ac2a5d6fb76af1739a5fb.jpg", + "image_caption": [ + "Figure 3: The Hallucinator. Stage 1: The original feature vector $q$ is extrapolated to the opposite direction of feature vector $k$ , forming $\\hat{q}$ in a linear way. Stage 2: We introduce non-linear transformation to smooth extrapolated features concatenated with $q$ and $q'$ . Stage 3: output original $q$ and hallucinated $\\hat{q}$ ." + ], + "image_footnote": [], + "bbox": [ + 514, + 95, + 883, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While center cropping effectively avoids false positive pairs, it reduces the operable region for random cropping and generates positive views with a similar appearance. We, therefore, adopt center-suppressed sampling [44] with a sampling method following a beta distribution $\\beta (\\alpha ,\\alpha)$ (i.e., a U-shaped distribution). Concretely, $\\beta (\\alpha ,\\alpha)$ assigns a lower probability to the center of the $\\hat{I}_{x,y}$ and gives greater probability to its boundary, increasing the variance between views $x_{1}$ and $x_{2}$ . Together, we summarize the process to obtain these two views as", + "bbox": [ + 496, + 348, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {1} = T (\\mathbb {R} _ {c r o p} (\\hat {I} _ {x, y} | \\alpha)), \\quad \\mathrm {s . t .} \\quad \\alpha < 1\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 512, + 823, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {2} = T \\left(\\mathbb {R} _ {c r o p} \\left(I _ {x, y} | \\alpha\\right)\\right), \\quad \\text {s . t .} \\quad \\alpha < 1, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 532, + 890, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ denotes data augmentations, including color jittering, random grayscale, Gaussian blur and horizontal flipping. $\\mathbb{R}_{crop}(\\dots|\\alpha)$ represents random cropping following the $\\beta (\\alpha ,\\alpha)$ distribution. $\\alpha$ is set to less than 1 to ensure an increasing sampling probability as the pixel's coordinates go beyond the center. A visualization of the center sampling method can be found in Supplementary Material (Section 1.2).", + "bbox": [ + 496, + 561, + 890, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Hallucinator", + "text_level": 1, + "bbox": [ + 500, + 694, + 635, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Asymmetric Feature Extrapolation. The first objective of this module is to introduce an additional positive pair without introducing extra computations. Therefore, the feature-level operation is preferred compared to image-level operations. To achieve this, Hallucinator is plugged in after the views $x_{1}$ and $x_{2}$ are encoded. As a result, the hallucinated (generated) feature is purely based on the two feature vectors $q$ and $k$ .", + "bbox": [ + 496, + 718, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, since harder positives improve pre-trained encoders' generalization capacity [51], the hallucinated features in positive pairs are favorable if they share less mutual information. Previous work illustrates that a symmetric", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16135", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "positive extrapolation is effective in generating hard examples for MoCo [68]. Concretely, two positive features are combined with weighted addition, pushing positive features apart. However, the hallucination process is asymmetric, i.e., Hallucinator is only added in one of the branches of the model. Then, we propose to apply the single-side feature extrapolation to the feature vector $q$ , as shown in the first stage of Figure 3. Additionally, we simplify the sampling strategy of weights for extrapolation from a beta distribution to a uniform distribution. To be specific, we summarize the positive extrapolation in our method as follows:", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {\\prime} = (1 + \\lambda) q - \\lambda k \\quad \\text {s . t .} \\quad \\lambda \\sim U (\\beta_ {1}, \\beta_ {2}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 276, + 468, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is sampled from a uniform distribution $U(\\beta_{1},\\beta_{2})$ . $\\beta_{1}$ and $\\beta_{2}$ , which are the boundary of the uniform distribution, are set to 0 and 0.1 by default.", + "bbox": [ + 76, + 311, + 468, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hallucination. Positive extrapolation is based on mixup [64], i.e., a linear transformation. While positive extrapolation has been proven beneficial to generating hard examples, this linear feature transformation might have a relatively limited capacity to synthesize new feature vectors. This assumption is based on the more satisfactory performance of the non-linear mixup over the original one [26]. Similarly, if non-linearity is introduced to the feature generation, the generated vector will benefit more from the training and boost the performance of downstream tasks.", + "bbox": [ + 75, + 359, + 468, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To empower our model with the capacity of non-linear fitting, we introduce the hallucination process. Specifically, we first concatenate $q'$ from the equation 3 and the feature vector $q$ together. Then, we use the concatenated feature $(q, q')$ as an input of a non-linear transformation function $H_{\\theta}(.)$ that can be instantiated with $n$ linear layers and a ReLU layer between two successive layers. When $n = 0$ , $H_{\\theta}(.)$ is a non-parametric module, forming an identity function. Such a setting performance is relatively sub-optimal, as shown in the ablation study in Table 6. Empirically, we find that $n = \\{2, 3\\}$ performs well as the hallucinator becomes non-linear and more powerful. We set $n = 3$ by default as its results are slightly better. With the transformation function $H_{\\theta}(.)$ , the hallucinated feature is defined as", + "bbox": [ + 75, + 513, + 470, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {q} = H _ {\\theta} \\left(q, q ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 758, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ is the parameters of $H_{\\theta}(.)$ . Notably, $H_{\\theta}(.)$ is differentiable, allowing us to back-propagate the loss of contrastive learning. Therefore, we update not just the parameters of the encoders and projectors but also the parameters $\\theta$ of the hallucinator. The second stage of Figure 3 illustrates the proposed hallucination process. Following that, we take $q$ and $\\hat{q}$ as the output of Hallucinator.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3160a7a5fa9552b083afff264f3ca41829110ceb0952bbf4cb5427ef0c58835e.jpg", + "image_caption": [ + "Figure 4: Similarity of positive pairs in training. Smaller values indicate less mutual information and better representation [51, 44, 68]. As Hallucinator incorporates nonlinearity extrapolation, it guarantees smooth training and harder positive features." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 874, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Discussion and Visualization", + "text_level": 1, + "bbox": [ + 500, + 387, + 756, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better understand the behavior of Hallucinator, we discuss two critical properties that may contribute to and explain its effectiveness. For visualization, we train MoCoV2 [11] with a standard ResNet-18 [31] on CIFar-10 [35].", + "bbox": [ + 498, + 410, + 892, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarity of Positive Pair. We first investigate the similarity of positive pairs during the training process. Concretely, we quantize mutual information using the cosine similarity of positive pairs $S$ :", + "bbox": [ + 498, + 470, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS \\triangleq \\frac {q \\cdot k}{\\| q \\| \\| k \\|.} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 541, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While hard positives share less mutual information, thus having a smaller cosine similarity value, the performance of downstream tasks will be enhanced [68]. Based on Figure 4, the proposed Hallucinator generates harder positives with smaller values of similarity. Consequently, it helps contrastive learning models obtain more nuisances-invariant features. More importantly, different from linear symmetric extrapolation, our training curve is relatively stable without many oscillations introduced. This observation indicates that Hallucinator is successfully optimized with the overall framework. Meanwhile, hallucinated features nicely fit into the contrastive learning task.", + "bbox": [ + 496, + 583, + 892, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Uniformity. We continue to analyze the performance of the proposed model from the perspective of uniformity. Notably, feature vectors from contrastive learning should be roughly uniformly distributed on a unit hyper-sphere, which ensures maximal information is preserved in the feature space [53]. In other words, the closer the feature distribution is to the uniform distribution, the more the feature benefits downstream tasks. To quantize uniformity, we follow previous work [53] to compute the average value of the", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16136", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a3759c2fe91044bb345a222683d2bc926222454b1b4b918efcf25e5c0601e49.jpg", + "image_caption": [ + "Figure 5: A measure of uniformity based on $G_{2}$ potential. We plot 10,000 feature vectors with Gaussian kernel density estimation (KDE) in $\\mathbb{R}^2$ . The left subplot illustrates the feature vectors from a uniform distribution. The three feature distributions on the right in the first row visualize the features from MoCoV2 [11]. The other three feature distributions in the second row demonstrate MoCoV2 with Hallucinator. Hallucinator benefits the uniformity with a smaller value of $G_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 150, + 118, + 339, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74ea5fecbf25104e2ec7e48a9b3f3757dc6ffe3d6903d8a349432ca0a5587ac3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 87, + 540, + 199 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9410cde9f6cf7d808cdface506514c36df32234665e9edd1aadf2a6297d64ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 88, + 681, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2cee15a31f715bd985f6ad1cefdd5490a90f1040d2aafd86ef40701744362c98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 823, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18f9d11ba3cbfd9bbc2f37fa8aff1b56a6f015e1b24c7333ada7e2168f6bd6c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 200, + 540, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/84d42f0c92e6afdf63cca55f6f5c4f42ba1c6af7c979057308b443c3fbed06a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 200, + 681, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a17d35b617352de88304204ae4f00aa78f16fa2580fa4d60b30bfd254f2eac18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 200, + 823, + 311 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/643054f6d1610b96b9e17cf33e8eb8a51cb769f2722fe8b58f4dfcf458187019.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10CIFAR-100Tiny ImageNetSTL-10
HallucinatorX✓X✓X✓X✓
MoCoV188.3188.9460.9461.8144.6545.5388.1990.09
MoCoV287.2189.2359.7061.2647.1247.9589.3290.46
SimCLR89.6690.1160.9461.4345.2246.3089.0789.98
SimSiam90.4790.7863.3964.3843.6644.9687.7988.16
", + "bbox": [ + 196, + 420, + 774, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Linear classification results for different contrastive methods and datasets in small scales. We adopt ResNet-18 as the backbone and report the classification results with or without Hallucinator.", + "bbox": [ + 73, + 541, + 892, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gaussian potential kernel (i.e., Radial Basis Function kernel) of positive features:", + "bbox": [ + 75, + 598, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG _ {t} (q, k) \\triangleq e ^ {- t \\| q - k \\| _ {2} ^ {2}} \\quad \\text {s . t .} \\quad t > 0, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 647, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t$ is a fixed parameter set as 2 for all the experiments. We visualize the feature vectors by mapping them to two-dimension feature space and applying $l_{2}$ normalization in Figure 5. $G_{2} = 0.2070$ for uniformly distributed samples, whereas features from MoCoV2 have $G_{2} = 0.2089$ . If we plugin Hallucinator into MoCo, Hallucinator provides further contrast during the training with extra positives introduced, giving more uniformly distributed features and a decreased $G_{2}$ value,i.e., 0.2081. Additionally, we visualize the features of two classes of Cifar-10. Each of these features is well-clustered. With better uniformity, clusters' overlapping decreases, forming more linearly separable features. Therefore, we could observe that the overlapping of feature clusters between class 1 and class 6 in MoCo is larger than the one with Hallucinator plugged in.", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 498, + 597, + 735, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct various experiments on different datasets and contrastive learning models to demonstrate the effectiveness of the Hallucinator. Firstly, we introduce the datasets and details in experiments in Section 4.1. Then, we continue to evaluate the performance of the proposed method using linear probing protocol following the paradigm of previous work [10, 11, 12, 23] in Section 4.2. Following this, we conduct ablation studies to understand how each module contributes to the final results in Section 4.3. Lastly, we show the transferability of pre-trained encoders in downstream tasks requiring dense pixel predictions, including object detection and semantic segmentation in Section 4.4.", + "bbox": [ + 496, + 623, + 890, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Training Details", + "text_level": 1, + "bbox": [ + 498, + 830, + 764, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Baseline Models. We first evaluate the performance of Hallucinator based on a wide range of datasets crossing different scales. Specifically, these datasets in", + "bbox": [ + 496, + 854, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16137", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2dc1dcccaecf59e5c755c2e9b863f6751933757976c2bdce74486f3f7b186089.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Linear classification results on IN-200 and IN-1K. Our method plugged in the proposed Hallucinator in baseline methods. All the models are pre-trained for 100 epochs and use identical training settings for fair comparisons." + ], + "table_body": "
MethodBackboneEpochIN-200IN-1K
MoCoV1ResNet-5010062.1957.27
MoCoV1(Ours)ResNet-5010063.4659.17
MoCoV2ResNet-5010062.5764.41
MoCoV2(Ours)ResNet-5010063.5864.97
SimCLRResNet-5010062.2261.23
SimCLR(Ours)ResNet-5010063.0261.71
SimSiamResNet-5010062.8063.11
SimSiam(Ours)ResNet-5010063.5263.55
", + "bbox": [ + 91, + 88, + 455, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "clude CIFAR-10, CIFAR-R100 [35], Tiny ImageNet, STL-10 [14] and ImageNet [15]. Meanwhile, we demonstrate the performance of Hallucinator in several popular contrastive learning frameworks, including MoCoV1 [29], MoCoV2 [11], SimCLR [10] and SimSiam [12].", + "bbox": [ + 75, + 335, + 467, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training and Evaluation Details. Importantly, we strictly adopt the same training settings when evaluating the Hallucinator's performance. While additional gain could be achieved with further exploration of hyper-parameter, it is not the focus of this work. The ultimate goal of the proposed Hallucinator is to provide further feature-level contrast for self-supervised learning. With the Hallucinator plugin, it benefits diversified self-supervised learning methods regardless of their type of backbones and hyper-parameters of training.", + "bbox": [ + 75, + 412, + 467, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets on relatively small scales include CIFAR-10&100, Tiny ImageNet and STL10. To ensure fair comparisons, we keep training settings the same over all these datasets. We follow the paradigm of previous work for pretraining [44]. Concretely, we pre-train contrastive learning models for 500 epochs with a batch size of 512 and ResNet-18 as the backbone. To optimize the models, we use an SGD optimizer and a cosine-annealed learning rate of 0.5. In the linear classification task, we train models for 100 epochs. With the initial learning set as 10, we divide it by 10 at the 60th and 80th epochs.", + "bbox": [ + 75, + 565, + 467, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the scale of the dataset expands to ImageNet, we use ResNet-50 as the backbone. For MoCoV1, MoCoV2 and SimSiam, all the pre-training settings follow the original work. We use a batch size of 512 and an SGD optimizer with a cosine-annealed learning rate of 0.5 for SimCLR, the same as the settings in [44]. For the linear classification task, we follow the original work in [29], training the model for 100 epochs. We set the initial learning rate to 30. Then, we divide the learning by 10 at the 60th and 80th with a weight decay of 0.", + "bbox": [ + 75, + 732, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We set $p = 0.5$ and $\\alpha = 0.6$ for center cropping and", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e0cfe431dc18fd45745253ab32f43725b02e042aee3e86edc63598b82b1f4a7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Ablation of the different modules in the proposed method." + ], + "table_body": "
Acc.(%)Center CroppingAsymmetric ExtrapolationHallucination
62.57✗✗✗
63.58(+1.01)✓✓✓
62.58(+0.01)✓
62.82(+0.25)✓
62.84(+0.27)✓
63.07(+0.50)✓✓
", + "bbox": [ + 517, + 88, + 874, + 208 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/14f441da8207cf00f1c6ef73301ac3e95a70ace84fffeed28b006c9ec36308f9.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4: Ablation of classification results w.r.t the $p$ value." + ], + "table_body": "
p (%)908070605040
Acc.(%)63.1163.1963.3363.5663.5863.21
", + "bbox": [ + 513, + 260, + 877, + 301 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da6e09e8428b323754fb07d6caec2835aed01cad53378607cbef215414186a7c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5: Ablation of classification results w.r.t the extrapolation range." + ], + "table_body": "
(β1,β2)(-0.5, 0.0)(0.0, 0.5)(0.5, 1.0)(1.0, 1.5)(0.0, 1.0)
Acc.(%)62.8063.3263.5563.4263.58
", + "bbox": [ + 509, + 337, + 880, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8bf1269637e88af7470ac74f1226747e2d39aa9ac0d0c983414d79027e78640a.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6: Ablation of classification results w.r.t the number of linear layers $n$ ." + ], + "table_body": "
layer #: n01234
Acc.(%)62.8563.0363.5763.5863.55
", + "bbox": [ + 524, + 425, + 867, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "center-suppressed sampling, respectively. For asymmetric linear extrapolation, we set $\\beta_{1} = 0$ and $\\beta_{2} = 0.1$ . In the Hallucination process, $n = 3$ for all the results reported in this paper. All the experiments are conducted on a server with 8 GPUs.", + "bbox": [ + 496, + 532, + 890, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Linear Classification Protocol", + "text_level": 1, + "bbox": [ + 500, + 619, + 764, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the previous protocol, we first evaluate the proposed method by linear classification of frozen features. For each dataset, we report the top-1 classification accuracy on the validation set.", + "bbox": [ + 496, + 642, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Small-Scale Datasets. The classification results on Cifar10&100, Tiny ImageNet and STL-10 are reported in Table 1. With Hallucinator introduced, we notice a stable improvement over the baselines ranging from $0.31\\%$ to $1.98\\%$ . Notably, such improvements do not introduce extra computations and generalize well to various models.", + "bbox": [ + 496, + 703, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ImageNet. For the results of ImageNet, we report the results at two different scales. First, we evaluate its performance on standard IN-1K (i.e. ImageNet-1K), which consists of 1000 classes. Second, we test the proposed method in IN-200 (i.e. ImageNet-200) with 200 randomly selected classes. We report the corresponding results in Table 2. We found that Hallucinator essentially benefits", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16138", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dd4ed3060a86516d27e93a2751f1ce0a394cdec320cd7938846ffd4e4e6becb2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method1N-1k Top-1VOC detectionCOCO detectionCOCO instance segmentation
APAP50AP75APbbAPbb50APbb75APmkAPmk50APmk75
Random init-33.860.257.2726.444.027.829.346.930.8
Supervised76.153.581.358.838.258.241.233.354.735.2
InfoMin [51]70.157.682.764.639.058.542.034.155.236.3
MoCoV1 [29]60.655.981.562.638.558.341.633.654.835.6
MoCoV1(Ours)63.856.781.863.238.958.541.933.855.236.0
MoCoV2 [11]67.557.082.463.639.058.641.934.255.436.2
MoCoV2(Ours)68.057.482.763.939.358.842.334.655.536.4
", + "bbox": [ + 81, + 88, + 890, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7: Fine-tuning results on object detection tasks on PASCAL VOC and COCO, and instance segmentation on COCO. All models are pre-trained for 200 epochs on ImageNet-1K.", + "bbox": [ + 76, + 272, + 890, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MoCoV1 with $1.27\\%$ and $1.89\\%$ improvements for IN-200 and IN-1K accordingly. For MoCoV2, SimSiam and SimCLR, we notice a gain in accuracy ranging from $0.44\\%$ to $1.01\\%$ . On average, the gains are more salient in IN-200.", + "bbox": [ + 76, + 329, + 468, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 402, + 240, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section of ablation studies, we investigate essential modules in our proposed methods, including center cropping, asymmetric extrapolation and the process of hallucination. We conduct experiments using MoCoV2 with ResNet-50 as the backbone. We report the results of the classification task in IN-200 for all the following experiments.", + "bbox": [ + 75, + 428, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contributions of Modules. We first report the results with or without crucial modules introduced in Section 3. According to Table 3, center cropping shows a similar performance to the original cropping method, successfully covering major semantic information of images. However, it successfully avoids false positives in pre-training, which is critical for hallucination. Asymmetric extrapolation benefits the performance of representation learning with reduced mutual information. This observation is consistent with the results shown in the symmetric extrapolation [68]. If we combine asymmetric extrapolation and the hallucination method, the performance of the model could be further boosted. However, it is still sub-optimal because of possible false positives in cropping.", + "bbox": [ + 75, + 535, + 470, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Center Cropping. In this work, we apply center cropping as a critical tool in one branch of contrastive learning to avoid false positives. Therefore, we investigate how the proposed model is influenced by the length ratio $p$ in Table 4. The classification results are favorable when $p$ is in the range (0.5, 0.6). We notice a performance drop when $p$ is less than 0.5. This drop is because the center cropping failed to cover all semantic information in the images.", + "bbox": [ + 75, + 748, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extrapolation Range. We continue to report the extrapolation range given its differences in sampling method and", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "single-side architecture compared with the previous method [68]. Based on Table 5, the classification accuracy drops if we set $\\beta_{1} = -0.5$ and $\\beta_{2} = 0$ , respectively. Such observation is because hallucinated positives share more mutual information, thus less beneficial to training. Generally, extrapolation boosts accuracy by outperforming the baseline by at least $0.75\\%$ . However, the improvement decreases when $\\beta_{1} > 1.0$ . We eventually set $\\beta_{1} = 0.0$ and $\\beta_{2} = 1.0$ , given its best performance.", + "bbox": [ + 496, + 329, + 890, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hallucination. We investigate how the number of linear layers influences the model's performance. When $n = 0$ , the hallucination process is an identity transformation, introducing no additional operation except for extrapolation. When $n = 2$ , non-linearity is introduced. Then we notice a salient improvement over the baseline. We set $n = 3$ by default since it gives better performance.", + "bbox": [ + 496, + 467, + 890, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Transferring Features", + "text_level": 1, + "bbox": [ + 500, + 585, + 705, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The primary goal of representation is to learn transferrable features. We evaluate the transferability of features from the proposed method following the previous protocol [29, 11, 10, 12]. Then, we compare the representation quality by transferring them to downstream tasks, including VOC [18] object detection and COCO [39] object detection and instance segmentation. Notably, we re-implement all these experiments using the same settings in MoCo's detector2 codebase [59].", + "bbox": [ + 496, + 611, + 890, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Detection on PASCAL VOC. Following the paradigm of previous work [29], we use Faster R-CNN[20] as the object detection method using R50-C4 as the detector [30]. We train the model end-to-end on the trainval2007+2012 and evaluate its performance on test2007. As shown in Table 7, we observe a stable gain range from 0.3 to 0.8 under different metrics on MoCoV1 and MoCoV2.", + "bbox": [ + 496, + 748, + 890, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Detection and instance segmentation on COCO. We continue to report the detection and segmentation re", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16139", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sults on COCO using Mask R-CNN [30]. Similarly, we use the R50-C4 as the backbone. The is model in an end-to-end way on train2017. Then, the model is evaluated on val2017. Again, the proposed method benefits the object detection and segmentation tasks with various metrics as demonstrated in Table 7.", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 195, + 194, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose Hallucinator, which generates additional hard positive pairs for contrastive learning models based on Siamese structure. Hallucinator generates novel data samples in the feature space to provide the training with further contrast without additional computation. We design an asymmetric feature extrapolation to avoid trivial positive pairs and innovatively introduce non-linear hallucination to smooth the generated samples. We empirically prove the effectiveness and generalization capacity of Hallucinator to well-recognized contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Finally, we hope this work could bring the concept of \"Hallucination\" into the SSL domain and unlock future research on sample generations&synthesis in contrastive learning.", + "bbox": [ + 75, + 220, + 472, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 90, + 596, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017.", + "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019.", + "[3] Zhipeng Bao, Yu-Xiong Wang, and Martial Hebert. Bowtie networks: Generative modeling for joint few-shot recognition and novel-view synthesis. arXiv preprint arXiv:2008.06981, 2020.", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018.", + "[5] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.", + "[6] Qingxing Cao, Liang Lin, Yukai Shi, Xiaodan Liang, and Guanbin Li. Attention-aware face hallucination via deep reinforcement learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 690-698, 2017.", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020.", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021.", + "[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021.", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020.", + "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021.", + "[13] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. Advances in neural information processing systems, 33:8765-8775, 2020.", + "[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on" + ], + "bbox": [ + 501, + 114, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16140", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011.", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[16] Terrance DeVries and Graham W Taylor. Dataset augmentation in feature space. arXiv preprint arXiv:1702.05538, 2017.", + "[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021.", + "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009.", + "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021.", + "[20] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015.", + "[21] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11017-11027, 2022.", + "[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020.", + "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020.", + "[24] Matthew D Grilli and Elizabeth L Glisky. The self-imagination effect: Benefits of a self-referential encoding strategy on cued recall in memory-impaired individuals with neurological damage. Journal of the International Neuropsychological Society, 17(5):929-933, 2011.", + "[25] Liangke Gui, Adrien Bardes, Ruslan Salakhutdinov, Alexander Hauptmann, Martial Hebert, and Yu-Xiong Wang. Learning to hallucinate examples from extrinsic and intrinsic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8701-8711, 2021.", + "[26] Hongyu Guo. Nonlinear mixup: Out-of-manifold data augmentation for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4044-4051, 2020.", + "[27] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pat" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006.", + "[28] Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In Proceedings of the IEEE international conference on computer vision, pages 3018-3027, 2017.", + "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020.", + "[30] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017.", + "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[32] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020.", + "[33] Wei-Ning Hsu, Yu Zhang, and James Glass. Unsupervised domain adaptation for robust speech recognition via variational autoencoder-based data augmentation. In 2017 IEEE automatic speech recognition and understanding workshop (ASRU), pages 16-23. IEEE, 2017.", + "[34] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020.", + "[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.., 2009.", + "[36] Varun Kumar, Hadrien Glaude, Cyprien de Lichy, and William Campbell. A closer look at feature space data augmentation for few-shot intent classification. arXiv preprint arXiv:1910.04176, 2019.", + "[37] Kibok Lee, Yian Zhu, Kihyuk Sohn, Chun-Liang Li, Jinwoo Shin, and Honglak Lee. i-mix: A domain-agnostic strategy for contrastive representation learning. arXiv preprint arXiv:2010.08887, 2020.", + "[38] Jiangmeng Li, Wenwen Qiang, Changwen Zheng, Bing Su, and Hui Xiong. Metaug: Contrastive learning via meta feature augmentation. In International Conference on Machine Learning, pages 12964-12978. PMLR, 2022.", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014.", + "[40] Giovanni Mariani, Florian Scheidegger, Roxana Istrate, Costas Bekas, and Cristiano Malossi. Bagan: Data augmentation with balancing gan. arXiv preprint arXiv:1803.09655, 2018.", + "[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16141", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018.", + "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.", + "[44] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16031-16040, 2022.", + "[45] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020.", + "[46] Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris, Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object recognition. Advances in neural information processing systems, 31, 2018.", + "[47] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023.", + "[48] Anshul Shah, Suvrit Sra, Rama Chellappa, and Anoop Cherian. Max-margin contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8220-8230, 2022.", + "[49] Zhiqiang Shen, Zechun Liu, Zhuang Liu, Marios Savvides, Trevor Darrell, and Eric Xing. Un-mix: Rethinking image mixtures for unsupervised visual representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2216-2224, 2022.", + "[50] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020.", + "[51] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020.", + "[52] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026-5033. IEEE, 2012.", + "[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020.", + "[54] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018.", + "[56] Jing Wu, Naira Hovakimyan, and Jennifer Hobbs. Genco: An auxiliary generator from contrastive learning for enhanced few-shot learning in remote sensing. arXiv preprint arXiv:2307.14612, 2023.", + "[57] Jing Wu, David Pichler, Daniel Marley, David Wilson, Naira Hovakimyan, and Jennifer Hobbs. Extended agriculture-vision: An extension of a large aerial image dataset for agricultural pattern analysis. arXiv preprint arXiv:2303.02460, 2023.", + "[58] Jing Wu, Ran Tao, Pan Zhao, Nicolas F Martin, and Naira Hovakimyan. Optimizing nitrogen management with deep reinforcement learning and crop simulations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1720, 2022.", + "[59] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.", + "[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018.", + "[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021.", + "[62] Yuanmeng Yan, Rumei Li, Sirui Wang, Fuzheng Zhang, Wei Wu, and Weiran Xu. Consert: A contrastive framework for self-supervised sentence representation transfer. arXiv preprint arXiv:2105.11741, 2021.", + "[63] Mang Ye, Xu Zhang, Pong C Yuen, and Shih-Fu Chang. Unsupervised embedding learning via invariant and spreading instance feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6210-6219, 2019.", + "[64] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017.", + "[65] Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2770-2779, 2019.", + "[66] Ruixiang Zhang, Tong Che, Zoubin Ghahramani, Yoshua Bengio, and Yangqiu Song. Metagan: An adversarial approach to few-shot learning. Advances in neural information processing systems, 31, 2018.", + "[67] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13008-13017, 2021.", + "[68] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16142", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "IEEE/CVF International Conference on Computer Vision, pages 10306-10315, 2021.", + "bbox": [ + 109, + 90, + 468, + 121 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16143", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_model.json b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82cd7f47e08e4bbc32b9bc383885bc4649fc361a --- /dev/null +++ b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_model.json @@ -0,0 +1,2347 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.13, + 0.89, + 0.178 + ], + "angle": 0, + "content": "Hallucination Improves the Performance of Unsupervised Visual Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.203, + 0.72, + 0.258 + ], + "angle": 0, + "content": "Jing Wu1 Jennifer Hobbs2 Naira Hovakimyan1 \n1University of Illinois Urbana-Champaign, 2Intelinair \n{jingwu6,nhovakim} @illinois, jennifer hobbs08+research@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Contrastive learning models based on Siamese structure have demonstrated remarkable performance in self-supervised learning. Such a success of contrastive learning relies on two conditions, a sufficient number of positive pairs and adequate variations between them. If the conditions are not met, these frameworks will lack semantic contrast and be fragile on overfitting. To address these two issues, we propose Hallucinator that could efficiently generate additional positive samples for further contrast. The Hallucinator is differentiable and creates new data in the feature space. Thus, it is optimized directly with the pre-training task and introduces nearly negligible computation. Moreover, we reduce the mutual information of hallucinated pairs and smooth them through non-linear operations. This process helps avoid over-confident contrastive learning models during the training and achieves more transformation-invariant feature embeddings. Remarkably, we empirically prove that the proposed Hallucinator generalizes well to various contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Under the linear classification protocol, a stable accuracy gain is achieved, ranging from \\(0.3\\%\\) to \\(3.0\\%\\) on CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. The improvement is also observed in transferring pre-train encoders to the downstream tasks, including object detection and segmentation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In the recent computer vision community, there has been rapid progress in self-supervised learning (SSL), gradually closing the performance gap with supervised learning [23, 29, 10, 7, 51]. Among the diverse approaches of SSL, contrastive learning, such as MoCoV1&V2 [29, 11], SimCLR [10], and SimSiam [12], shows promising results. Generally, contrastive learning treats each image as one class which will be augmented into two separate views. These two views form one positive pair and should ide" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.294, + 0.855, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.54, + 0.895, + 0.692 + ], + "angle": 0, + "content": "Figure 1: The motivation of the proposed hallucination methods. Given one pair of images with the same semantic meaning, such as a pair of horses, a person can envision further similar pairs by imagining one of the horses in different poses and surroundings. If a contrastive learning model could do such hallucination, it could have additional novel pairs to contrast given the same data. Note that this hallucination process is for illustration only. In the implementation, all the hallucinated samples are computed in the feature space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.895, + 0.81 + ], + "angle": 0, + "content": "ally be close if mapped to feature space. With sufficient contrast in the feature space, contrastive learning models show a strong capacity to learn transformation-invariant features that are transferable to various downstream tasks, such as classification, object detection, and segmentation [18, 30, 39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To ensure sufficient contrast, researchers from previous work address the issue from two essential practices, either introducing large amounts of positive pairs or adding additional variants&transformation among them. For example, SimCLR uses a batch size that generates thousands of positives to facilitate the convergence of models [10]. Work" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16132" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.228 + ], + "angle": 0, + "content": "from [51, 9] reduces mutual information of positive pairs using stronger data augmentation, i.e., color distortion and jigsaw transformation. Likewise, the work from [44] introduces ContrastiveCrop, and the work from [49] proposes Un-Mix, respectively, to reduce the similar semantic meaning of sample pairs in the original image space. Beyond the data augmentation and image operations, researchers from [68] propose to apply a linear operation to generate hard positive samples in feature space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.471, + 0.426 + ], + "angle": 0, + "content": "Despite the success of prior approaches, we argue that large batch sizes are not always achievable. Meanwhile, all proposed techniques only focus on improving the original pairs. Given one positive pair of positive samples, humans are born with the amazing ability to come up with additional positives by imagining a sample from different surroundings and perspectives without much effort, as demonstrated in Figure 1. This process of self-imagination, in turn, will benefit the human neurological system, improving recognition capacity [24]. Similarly, if we could empower contrastive learning models with the ability to hallucinate or imagine an object to a novel view, additional positive pairs could be provided for the learning tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Unfortunately, exploring feasible methods to hallucinate novel positive pairs is challenging. Firstly, while generative models produce realistic images that could form additional positive views [22, 1, 42, 4], realistic data do not necessarily benefit learning tasks [55]. More importantly, applying these approaches forces us to fall back into a computational dilemma to the previous method. In other words, image-level hallucination still suffers from expensive computation as we still need to encode the hallucinated images into feature space. Lastly, if the generated positive pairs are similar to each other, training a discriminative model would be too trivial, thus showing poor generalization capacity [44, 51, 68]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Therefore, our key insight is that the sample-generation process should aim for three critical elements: (i) feature-space operation (ii) sufficient variance of positive pairs (iii) a differentiable module optimized directly related to the learning task. To achieve this, we propose Hallucinator to improve the performance of contrastive learning with Siamese structures. The Hallucinator is plugged in after the encoder to manipulate feature vectors and improve the feature-level batch size for further contrast. To ensure adequate variance is introduced, we propose an asymmetric feature extrapolation method inspired by the work from [68]. More importantly, we present a non-linear hallucination process for the extrapolated samples. Such a process is differentiable (i.e. learnable), therefore essentially boosting Hallucinator to generate smooth and task-related features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The proposed Hallucinator delivers extra positives and simultaneously enlarges the variance between newly introduced pairs. Moreover, this approach only relies on pos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "itive samples. Therefore, it can be easily applied to any Siamese structure by adding it after the encoders as a plug-and-play module. Without the tedious exploration of hyperparameters and much additional computation, we empirically prove the effectiveness of the proposed Hallucinator on popular contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. We notice a stable improvement ranging from \\(0.3\\%\\) to \\(3.0\\%\\) under the linear classification protocol, crossing the CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. We also observe that models trained with Hallucinator show better transferability in downstream tasks like object detection and segmentation." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.273, + 0.845, + 0.286 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.288, + 0.892, + 0.332 + ], + "angle": 0, + "content": "- We investigate a critical yet under-explored aspect of contrastive learning: introducing additional positive pairs with further variation in feature space." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.334, + 0.892, + 0.436 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first attempt to incorporate the concept of \"Hallucination\" into contrastive representation learning. We propose Hallucinator to realize this idea, which effectively generates smooth and less similar positive feature vectors. The Hallucinator is simple, effective, and agnostic to contrastive frameworks." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.439, + 0.892, + 0.484 + ], + "angle": 0, + "content": "- We empirically illustrate that the proposed approach significantly benefits the various contrastive learning models within multiple datasets." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.288, + 0.892, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.65, + 0.512 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.522, + 0.892, + 0.553 + ], + "angle": 0, + "content": "In this section, we introduce related literature on contrastive learning and hallucination techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.702, + 0.577 + ], + "angle": 0, + "content": "2.1. Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The key idea of contrastive learning is to minimize the distance of positive pairs and repulse negative pairs in the feature space [27]. This idea has been successfully applied to unsupervised visual representation tasks, showing promising results in various downstream tasks [10, 29, 11, 23, 12, 2, 32, 17, 54, 41, 43, 50, 57, 60, 63, 61, 8, 37]. One of the breakthroughs of contrastive learning models is SimCLR [10], which introduces a simple but effective visual representation learning method. Without negative pairs, SimCLR learned transformation-invariant representation with a large batch size. Contemporaneous impressive work is MoCo from [29]. To ensure MoCo can be trained smoothly with computational-friendly batch sizes, the authors of MoCo propose a memory bank to store negative features and momentum-updated backbones. After this, SimSiam [12] presents a Siamese-based network that can learn high-quality representation with stop-gradient, successfully avoiding the collapse of contrastive learning models. Other designs of contrastive learning rely on an online network to predict the output of the target network or contrasting cluster assignments [23, 7]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16133" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.893, + 0.355 + ], + "angle": 0, + "content": "Figure 2: Illustration of contrastive learning (MoCoV2 [11]) with Hallucinator. The Hallucinator is added after the backbones and projector for feature-level manipulation. While Hallucinator feeds original feature vector \\( q \\) forward, it additionally provides hallucinated feature \\( \\hat{q} \\) for further contrast." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.683 + ], + "angle": 0, + "content": "To further boost the models' performance, researchers explore diverse ways to hard positives and reduce the mutual information of positive samples so that more transformation-invariant embeddings can be learned [51, 48, 45, 13]. For image-level enhancement, typical works focus on improving the data augmentation in pixel space using Mixup, color distortion, or jigsaw transformation [49, 51, 9]. More recent work is presented by [44] using ContrastiveCrop, which creates better positive views through the localization box. For feature-level operation, the authors from [34] prove the effectiveness of hard negative samples. More recent and close work to this paper is [68], which applies symmetric extrapolation to create hard positives but introduces no non-linearity. Importantly, instead of replacing the original positive samples, this paper proposes a differentiable hallucinator to generate an additional positive sample with less mutual information and better smoothness. Such a setting ensures maximal and adaptive contrast during the training, generalizing well to diverse contrastive learning models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.22, + 0.71 + ], + "angle": 0, + "content": "2.2. Hallucination" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Hallucination is initially proposed to solve the scarcity of data in the classification task [28]. Then, this idea is kept updated and applied in different areas [46, 55, 66, 67, 6, 65, 25, 47], such as object detection, aerial navigation, skeleton-based action recognition, and face generation. While image-level hallucination benefits few-shot recognition by synthesis of novel view [3] or introducing random noises [55], most of the work applies hallucination in the feature space. The work from [28] generates novel class features by transforming shared features in base classes. Authors of [67] build a hallucination framework in the region of interest feature space object to enhance object" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.472 + ], + "angle": 0, + "content": "detection performance. More recent work shows that this hallucination mechanism also benefits 3D human pose estimation by generating novel motion sequences [21]. While hallucination is effective in different learning tasks, to the best of our knowledge, the performance and application of hallucination in SSL are fully unexplored." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.484, + 0.757, + 0.5 + ], + "angle": 0, + "content": "2.3. Feature-Level Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Hallucination relies on effective feature-level augmentation or manipulations. The primary goal of feature augmentation is to extend the limited labeled dataset without relying on expensive computation such as auto-encoder[33], Generative Adversarial Networks[40] or simulation tools[52, 58, 5]. For instance, in the work by [16], a task-agnostic feature augmentation approach is proposed to enrich training data with minimal additional computation. The authors of [36] also explore similar ideas in the domain of few-shot learning. Building upon this, the concept is adapted to sentence representation learning [62, 19] and few-shot learning tasks in remote sensing[56]. More recently, [38] applies feature augmentation based on a meta-learning technique." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.593, + 0.739 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this section, we first introduce the overall process of hallucination for contrastive representation learning in Section 3.1. Secondly, we highlight the center cropping method we used, which is crucial to effective hallucination or generation of new samples in Section 3.2. Then, we introduce the Hallucinator incorporated into our contrastive models in Section 3.3. Finally, we visualize and discuss the critical properties of the hallucination method from two perspectives in Section 3.4: the similarity of positive samples and the uniformity of the feature distribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16134" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.273, + 0.108 + ], + "angle": 0, + "content": "3.1. The Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.474, + 0.403 + ], + "angle": 0, + "content": "Taking MoCo [11] as an example, we illustrate how a Hallucinator can be plugged into a contrastive learning model in Figure 2. Our architecture takes one image \\( x \\) as input. Then, the input \\( x \\) is augmented into two views \\( x_{1} \\) and \\( x_{2} \\). Each view will be processed by an encoder consisting of a backbone (e.g., ResNet) and a projector (e.g., an MLP head). After the encoders, output vectors \\( q \\) and \\( k \\) are obtained, forming one positive pair \\( (q, k) \\). Then, this positive pair \\( (q, k) \\) is fed to a Hallucinator. Notably, Hallucinator is only added to one branch of the framework to generate an additional positive feature \\( \\hat{q} \\). Together with feature vector \\( k \\), \\( \\hat{q} \\) and \\( k \\) form an extra positive pair \\( (\\hat{q}, k) \\) during the training. Based on different contrastive learning models, the loss functions keep intact, and the average loss of these two positive pairs is computed for back-propagation. The same paradigm could be applied to SimCLR, SiamSiam and other contrastive learning models. We illustrate further details about pipelines and loss functions of other models in this paper in the Supplementary Material (Section 1.1)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.41, + 0.246, + 0.426 + ], + "angle": 0, + "content": "3.2. Center Cropping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.675 + ], + "angle": 0, + "content": "In contrastive learning, data augmentations aim to ensure the performance of pre-trained representations invariant to nuisances. Among all these methods, random crop plays the most critical role in all the contrastive learning models. Generally, views (cropped tiles) generated by random cropping are diversified, successfully covering all the semantic information over the whole image. However, such a cropping method is likely to generate false positives patches [44]. In other words, patches randomly cropped from the original images do not necessarily share the overlapped pixels and sufficient common information. Therefore, these false positive pairs may be fooling models during training, causing representations to be sub-optimal. Importantly, the issue will be exacerbated if we generate further hallucinated samples based on false positive pairs, which misleads the overall training beyond the sweet spot." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.737 + ], + "angle": 0, + "content": "To tackle this issue, we first apply center cropping \\(\\mathbb{C}_{crop}\\) to the original image, getting a relatively smaller image \\(\\hat{I}_{x,y}\\). Then, random cropping \\(\\mathbb{R}_{crop}\\) is applied to \\(\\hat{I}_{x,y}\\). More specifically, the center cropping can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.746, + 0.47, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {x, y} = \\mathbb {C} _ {\\text {c r o p}} \\left(I _ {x, y}, p\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\( I_{x,y} \\) is the input image with \\( (x,y) \\) as the coordinate of the images' center. After center cropping, we keep the center of \\( \\hat{I}_{x,y} \\) unchanged. Meanwhile, with the original shape of \\( I_{x,y} \\) defined as \\( (h,w) \\), we define the shape of cropped image \\( \\hat{I}_{x,y} \\) as \\( (\\hat{h},\\hat{w}) \\). The \\( p \\) denotes a ratio of cropped length over the original length, i.e., \\( p = \\frac{\\hat{w}}{w} = \\frac{\\hat{h}}{h} \\). Only if particularly mentioned, we set \\( p = 0.5 \\) for all experiments in this paper." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.097, + 0.885, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.23, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 3: The Hallucinator. Stage 1: The original feature vector \\( q \\) is extrapolated to the opposite direction of feature vector \\( k \\), forming \\( \\hat{q} \\) in a linear way. Stage 2: We introduce non-linear transformation to smooth extrapolated features concatenated with \\( q \\) and \\( q' \\). Stage 3: output original \\( q \\) and hallucinated \\( \\hat{q} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.5 + ], + "angle": 0, + "content": "While center cropping effectively avoids false positive pairs, it reduces the operable region for random cropping and generates positive views with a similar appearance. We, therefore, adopt center-suppressed sampling [44] with a sampling method following a beta distribution \\(\\beta (\\alpha ,\\alpha)\\) (i.e., a U-shaped distribution). Concretely, \\(\\beta (\\alpha ,\\alpha)\\) assigns a lower probability to the center of the \\(\\hat{I}_{x,y}\\) and gives greater probability to its boundary, increasing the variance between views \\(x_{1}\\) and \\(x_{2}\\). Together, we summarize the process to obtain these two views as" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.513, + 0.824, + 0.531 + ], + "angle": 0, + "content": "\\[\nx _ {1} = T (\\mathbb {R} _ {c r o p} (\\hat {I} _ {x, y} | \\alpha)), \\quad \\mathrm {s . t .} \\quad \\alpha < 1\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.533, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\nx _ {2} = T \\left(\\mathbb {R} _ {c r o p} \\left(I _ {x, y} | \\alpha\\right)\\right), \\quad \\text {s . t .} \\quad \\alpha < 1, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.682 + ], + "angle": 0, + "content": "where \\(T\\) denotes data augmentations, including color jittering, random grayscale, Gaussian blur and horizontal flipping. \\(\\mathbb{R}_{crop}(\\dots|\\alpha)\\) represents random cropping following the \\(\\beta (\\alpha ,\\alpha)\\) distribution. \\(\\alpha\\) is set to less than 1 to ensure an increasing sampling probability as the pixel's coordinates go beyond the center. A visualization of the center sampling method can be found in Supplementary Material (Section 1.2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.636, + 0.71 + ], + "angle": 0, + "content": "3.3. Hallucinator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Asymmetric Feature Extrapolation. The first objective of this module is to introduce an additional positive pair without introducing extra computations. Therefore, the feature-level operation is preferred compared to image-level operations. To achieve this, Hallucinator is plugged in after the views \\( x_{1} \\) and \\( x_{2} \\) are encoded. As a result, the hallucinated (generated) feature is purely based on the two feature vectors \\( q \\) and \\( k \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Meanwhile, since harder positives improve pre-trained encoders' generalization capacity [51], the hallucinated features in positive pairs are favorable if they share less mutual information. Previous work illustrates that a symmetric" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16135" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "positive extrapolation is effective in generating hard examples for MoCo [68]. Concretely, two positive features are combined with weighted addition, pushing positive features apart. However, the hallucination process is asymmetric, i.e., Hallucinator is only added in one of the branches of the model. Then, we propose to apply the single-side feature extrapolation to the feature vector \\( q \\), as shown in the first stage of Figure 3. Additionally, we simplify the sampling strategy of weights for extrapolation from a beta distribution to a uniform distribution. To be specific, we summarize the positive extrapolation in our method as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.277, + 0.47, + 0.294 + ], + "angle": 0, + "content": "\\[\nq ^ {\\prime} = (1 + \\lambda) q - \\lambda k \\quad \\text {s . t .} \\quad \\lambda \\sim U (\\beta_ {1}, \\beta_ {2}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.312, + 0.47, + 0.357 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is sampled from a uniform distribution \\(U(\\beta_{1},\\beta_{2})\\). \\(\\beta_{1}\\) and \\(\\beta_{2}\\), which are the boundary of the uniform distribution, are set to 0 and 0.1 by default." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.512 + ], + "angle": 0, + "content": "Hallucination. Positive extrapolation is based on mixup [64], i.e., a linear transformation. While positive extrapolation has been proven beneficial to generating hard examples, this linear feature transformation might have a relatively limited capacity to synthesize new feature vectors. This assumption is based on the more satisfactory performance of the non-linear mixup over the original one [26]. Similarly, if non-linearity is introduced to the feature generation, the generated vector will benefit more from the training and boost the performance of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.471, + 0.742 + ], + "angle": 0, + "content": "To empower our model with the capacity of non-linear fitting, we introduce the hallucination process. Specifically, we first concatenate \\( q' \\) from the equation 3 and the feature vector \\( q \\) together. Then, we use the concatenated feature \\( (q, q') \\) as an input of a non-linear transformation function \\( H_{\\theta}(.) \\) that can be instantiated with \\( n \\) linear layers and a ReLU layer between two successive layers. When \\( n = 0 \\), \\( H_{\\theta}(.) \\) is a non-parametric module, forming an identity function. Such a setting performance is relatively sub-optimal, as shown in the ablation study in Table 6. Empirically, we find that \\( n = \\{2, 3\\} \\) performs well as the hallucinator becomes non-linear and more powerful. We set \\( n = 3 \\) by default as its results are slightly better. With the transformation function \\( H_{\\theta}(.) \\), the hallucinated feature is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.76, + 0.47, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\hat {q} = H _ {\\theta} \\left(q, q ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the parameters of \\(H_{\\theta}(.)\\). Notably, \\(H_{\\theta}(.)\\) is differentiable, allowing us to back-propagate the loss of contrastive learning. Therefore, we update not just the parameters of the encoders and projectors but also the parameters \\(\\theta\\) of the hallucinator. The second stage of Figure 3 illustrates the proposed hallucination process. Following that, we take \\(q\\) and \\(\\hat{q}\\) as the output of Hallucinator." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.088, + 0.875, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.286, + 0.895, + 0.363 + ], + "angle": 0, + "content": "Figure 4: Similarity of positive pairs in training. Smaller values indicate less mutual information and better representation [51, 44, 68]. As Hallucinator incorporates nonlinearity extrapolation, it guarantees smooth training and harder positive features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.757, + 0.403 + ], + "angle": 0, + "content": "3.4. Discussion and Visualization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.893, + 0.472 + ], + "angle": 0, + "content": "To better understand the behavior of Hallucinator, we discuss two critical properties that may contribute to and explain its effectiveness. For visualization, we train MoCoV2 [11] with a standard ResNet-18 [31] on CIFar-10 [35]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Similarity of Positive Pair. We first investigate the similarity of positive pairs during the training process. Concretely, we quantize mutual information using the cosine similarity of positive pairs \\( S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.542, + 0.892, + 0.575 + ], + "angle": 0, + "content": "\\[\nS \\triangleq \\frac {q \\cdot k}{\\| q \\| \\| k \\|.} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.765 + ], + "angle": 0, + "content": "While hard positives share less mutual information, thus having a smaller cosine similarity value, the performance of downstream tasks will be enhanced [68]. Based on Figure 4, the proposed Hallucinator generates harder positives with smaller values of similarity. Consequently, it helps contrastive learning models obtain more nuisances-invariant features. More importantly, different from linear symmetric extrapolation, our training curve is relatively stable without many oscillations introduced. This observation indicates that Hallucinator is successfully optimized with the overall framework. Meanwhile, hallucinated features nicely fit into the contrastive learning task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Uniformity. We continue to analyze the performance of the proposed model from the perspective of uniformity. Notably, feature vectors from contrastive learning should be roughly uniformly distributed on a unit hyper-sphere, which ensures maximal information is preserved in the feature space [53]. In other words, the closer the feature distribution is to the uniform distribution, the more the feature benefits downstream tasks. To quantize uniformity, we follow previous work [53] to compute the average value of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16136" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.119, + 0.341, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.088, + 0.541, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.089, + 0.682, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.824, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.201, + 0.541, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.201, + 0.682, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.201, + 0.824, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.347, + 0.894, + 0.409 + ], + "angle": 0, + "content": "Figure 5: A measure of uniformity based on \\( G_{2} \\) potential. We plot 10,000 feature vectors with Gaussian kernel density estimation (KDE) in \\( \\mathbb{R}^2 \\). The left subplot illustrates the feature vectors from a uniform distribution. The three feature distributions on the right in the first row visualize the features from MoCoV2 [11]. The other three feature distributions in the second row demonstrate MoCoV2 with Hallucinator. Hallucinator benefits the uniformity with a smaller value of \\( G_{2} \\)." + }, + { + "type": "table", + "bbox": [ + 0.197, + 0.421, + 0.776, + 0.537 + ], + "angle": 0, + "content": "
DatasetCIFAR-10CIFAR-100Tiny ImageNetSTL-10
HallucinatorX✓X✓X✓X✓
MoCoV188.3188.9460.9461.8144.6545.5388.1990.09
MoCoV287.2189.2359.7061.2647.1247.9589.3290.46
SimCLR89.6690.1160.9461.4345.2246.3089.0789.98
SimSiam90.4790.7863.3964.3843.6644.9687.7988.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.542, + 0.893, + 0.573 + ], + "angle": 0, + "content": "Table 1: Linear classification results for different contrastive methods and datasets in small scales. We adopt ResNet-18 as the backbone and report the classification results with or without Hallucinator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.63 + ], + "angle": 0, + "content": "Gaussian potential kernel (i.e., Radial Basis Function kernel) of positive features:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.648, + 0.47, + 0.668 + ], + "angle": 0, + "content": "\\[\nG _ {t} (q, k) \\triangleq e ^ {- t \\| q - k \\| _ {2} ^ {2}} \\quad \\text {s . t .} \\quad t > 0, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\( t \\) is a fixed parameter set as 2 for all the experiments. We visualize the feature vectors by mapping them to two-dimension feature space and applying \\( l_{2} \\) normalization in Figure 5. \\( G_{2} = 0.2070 \\) for uniformly distributed samples, whereas features from MoCoV2 have \\( G_{2} = 0.2089 \\). If we plugin Hallucinator into MoCo, Hallucinator provides further contrast during the training with extra positives introduced, giving more uniformly distributed features and a decreased \\( G_{2} \\) value,i.e., 0.2081. Additionally, we visualize the features of two classes of Cifar-10. Each of these features is well-clustered. With better uniformity, clusters' overlapping decreases, forming more linearly separable features. Therefore, we could observe that the overlapping of feature clusters between class 1 and class 6 in MoCo is larger than the one with Hallucinator plugged in." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.598, + 0.736, + 0.615 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.82 + ], + "angle": 0, + "content": "In this section, we conduct various experiments on different datasets and contrastive learning models to demonstrate the effectiveness of the Hallucinator. Firstly, we introduce the datasets and details in experiments in Section 4.1. Then, we continue to evaluate the performance of the proposed method using linear probing protocol following the paradigm of previous work [10, 11, 12, 23] in Section 4.2. Following this, we conduct ablation studies to understand how each module contributes to the final results in Section 4.3. Lastly, we show the transferability of pre-trained encoders in downstream tasks requiring dense pixel predictions, including object detection and semantic segmentation in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.832, + 0.766, + 0.849 + ], + "angle": 0, + "content": "4.1. Datasets and Training Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Datasets and Baseline Models. We first evaluate the performance of Hallucinator based on a wide range of datasets crossing different scales. Specifically, these datasets in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16137" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.457, + 0.24 + ], + "angle": 0, + "content": "
MethodBackboneEpochIN-200IN-1K
MoCoV1ResNet-5010062.1957.27
MoCoV1(Ours)ResNet-5010063.4659.17
MoCoV2ResNet-5010062.5764.41
MoCoV2(Ours)ResNet-5010063.5864.97
SimCLRResNet-5010062.2261.23
SimCLR(Ours)ResNet-5010063.0261.71
SimSiamResNet-5010062.8063.11
SimSiam(Ours)ResNet-5010063.5263.55
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.246, + 0.47, + 0.307 + ], + "angle": 0, + "content": "Table 2: Linear classification results on IN-200 and IN-1K. Our method plugged in the proposed Hallucinator in baseline methods. All the models are pre-trained for 100 epochs and use identical training settings for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.468, + 0.412 + ], + "angle": 0, + "content": "clude CIFAR-10, CIFAR-R100 [35], Tiny ImageNet, STL-10 [14] and ImageNet [15]. Meanwhile, we demonstrate the performance of Hallucinator in several popular contrastive learning frameworks, including MoCoV1 [29], MoCoV2 [11], SimCLR [10] and SimSiam [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.468, + 0.565 + ], + "angle": 0, + "content": "Training and Evaluation Details. Importantly, we strictly adopt the same training settings when evaluating the Hallucinator's performance. While additional gain could be achieved with further exploration of hyper-parameter, it is not the focus of this work. The ultimate goal of the proposed Hallucinator is to provide further feature-level contrast for self-supervised learning. With the Hallucinator plugin, it benefits diversified self-supervised learning methods regardless of their type of backbones and hyper-parameters of training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.468, + 0.732 + ], + "angle": 0, + "content": "Datasets on relatively small scales include CIFAR-10&100, Tiny ImageNet and STL10. To ensure fair comparisons, we keep training settings the same over all these datasets. We follow the paradigm of previous work for pretraining [44]. Concretely, we pre-train contrastive learning models for 500 epochs with a batch size of 512 and ResNet-18 as the backbone. To optimize the models, we use an SGD optimizer and a cosine-annealed learning rate of 0.5. In the linear classification task, we train models for 100 epochs. With the initial learning set as 10, we divide it by 10 at the 60th and 80th epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.468, + 0.884 + ], + "angle": 0, + "content": "As the scale of the dataset expands to ImageNet, we use ResNet-50 as the backbone. For MoCoV1, MoCoV2 and SimSiam, all the pre-training settings follow the original work. We use a batch size of 512 and an SGD optimizer with a cosine-annealed learning rate of 0.5 for SimCLR, the same as the settings in [44]. For the linear classification task, we follow the original work in [29], training the model for 100 epochs. We set the initial learning rate to 30. Then, we divide the learning by 10 at the 60th and 80th with a weight decay of 0." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We set \\( p = 0.5 \\) and \\( \\alpha = 0.6 \\) for center cropping and" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.875, + 0.209 + ], + "angle": 0, + "content": "
Acc.(%)Center CroppingAsymmetric ExtrapolationHallucination
62.57✗✗✗
63.58(+1.01)✓✓✓
62.58(+0.01)✓
62.82(+0.25)✓
62.84(+0.27)✓
63.07(+0.50)✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.215, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Table 3: Ablation of the different modules in the proposed method." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.261, + 0.879, + 0.302 + ], + "angle": 0, + "content": "
p (%)908070605040
Acc.(%)63.1163.1963.3363.5663.5863.21
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.503, + 0.307, + 0.887, + 0.323 + ], + "angle": 0, + "content": "Table 4: Ablation of classification results w.r.t the \\( p \\) value." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.338, + 0.882, + 0.375 + ], + "angle": 0, + "content": "
(β1,β2)(-0.5, 0.0)(0.0, 0.5)(0.5, 1.0)(1.0, 1.5)(0.0, 1.0)
Acc.(%)62.8063.3263.5563.4263.58
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.38, + 0.89, + 0.41 + ], + "angle": 0, + "content": "Table 5: Ablation of classification results w.r.t the extrapolation range." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.426, + 0.869, + 0.47 + ], + "angle": 0, + "content": "
layer #: n01234
Acc.(%)62.8563.0363.5763.5863.55
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.475, + 0.89, + 0.506 + ], + "angle": 0, + "content": "Table 6: Ablation of classification results w.r.t the number of linear layers \\( n \\) ." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.608 + ], + "angle": 0, + "content": "center-suppressed sampling, respectively. For asymmetric linear extrapolation, we set \\(\\beta_{1} = 0\\) and \\(\\beta_{2} = 0.1\\). In the Hallucination process, \\(n = 3\\) for all the results reported in this paper. All the experiments are conducted on a server with 8 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.765, + 0.635 + ], + "angle": 0, + "content": "4.2. Linear Classification Protocol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Following the previous protocol, we first evaluate the proposed method by linear classification of frozen features. For each dataset, we report the top-1 classification accuracy on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Results on Small-Scale Datasets. The classification results on Cifar10&100, Tiny ImageNet and STL-10 are reported in Table 1. With Hallucinator introduced, we notice a stable improvement over the baselines ranging from \\(0.31\\%\\) to \\(1.98\\%\\). Notably, such improvements do not introduce extra computations and generalize well to various models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results on ImageNet. For the results of ImageNet, we report the results at two different scales. First, we evaluate its performance on standard IN-1K (i.e. ImageNet-1K), which consists of 1000 classes. Second, we test the proposed method in IN-200 (i.e. ImageNet-200) with 200 randomly selected classes. We report the corresponding results in Table 2. We found that Hallucinator essentially benefits" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16138" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.892, + 0.268 + ], + "angle": 0, + "content": "
Method1N-1k Top-1VOC detectionCOCO detectionCOCO instance segmentation
APAP50AP75APbbAPbb50APbb75APmkAPmk50APmk75
Random init-33.860.257.2726.444.027.829.346.930.8
Supervised76.153.581.358.838.258.241.233.354.735.2
InfoMin [51]70.157.682.764.639.058.542.034.155.236.3
MoCoV1 [29]60.655.981.562.638.558.341.633.654.835.6
MoCoV1(Ours)63.856.781.863.238.958.541.933.855.236.0
MoCoV2 [11]67.557.082.463.639.058.641.934.255.436.2
MoCoV2(Ours)68.057.482.763.939.358.842.334.655.536.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.273, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Table 7: Fine-tuning results on object detection tasks on PASCAL VOC and COCO, and instance segmentation on COCO. All models are pre-trained for 200 epochs on ImageNet-1K." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.33, + 0.47, + 0.391 + ], + "angle": 0, + "content": "MoCoV1 with \\(1.27\\%\\) and \\(1.89\\%\\) improvements for IN-200 and IN-1K accordingly. For MoCoV2, SimSiam and SimCLR, we notice a gain in accuracy ranging from \\(0.44\\%\\) to \\(1.01\\%\\). On average, the gains are more salient in IN-200." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.241, + 0.419 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.534 + ], + "angle": 0, + "content": "In this section of ablation studies, we investigate essential modules in our proposed methods, including center cropping, asymmetric extrapolation and the process of hallucination. We conduct experiments using MoCoV2 with ResNet-50 as the backbone. We report the results of the classification task in IN-200 for all the following experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.471, + 0.748 + ], + "angle": 0, + "content": "Contributions of Modules. We first report the results with or without crucial modules introduced in Section 3. According to Table 3, center cropping shows a similar performance to the original cropping method, successfully covering major semantic information of images. However, it successfully avoids false positives in pre-training, which is critical for hallucination. Asymmetric extrapolation benefits the performance of representation learning with reduced mutual information. This observation is consistent with the results shown in the symmetric extrapolation [68]. If we combine asymmetric extrapolation and the hallucination method, the performance of the model could be further boosted. However, it is still sub-optimal because of possible false positives in cropping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Center Cropping. In this work, we apply center cropping as a critical tool in one branch of contrastive learning to avoid false positives. Therefore, we investigate how the proposed model is influenced by the length ratio \\( p \\) in Table 4. The classification results are favorable when \\( p \\) is in the range (0.5, 0.6). We notice a performance drop when \\( p \\) is less than 0.5. This drop is because the center cropping failed to cover all semantic information in the images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Extrapolation Range. We continue to report the extrapolation range given its differences in sampling method and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.466 + ], + "angle": 0, + "content": "single-side architecture compared with the previous method [68]. Based on Table 5, the classification accuracy drops if we set \\(\\beta_{1} = -0.5\\) and \\(\\beta_{2} = 0\\), respectively. Such observation is because hallucinated positives share more mutual information, thus less beneficial to training. Generally, extrapolation boosts accuracy by outperforming the baseline by at least \\(0.75\\%\\). However, the improvement decreases when \\(\\beta_{1} > 1.0\\). We eventually set \\(\\beta_{1} = 0.0\\) and \\(\\beta_{2} = 1.0\\), given its best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.574 + ], + "angle": 0, + "content": "Hallucination. We investigate how the number of linear layers influences the model's performance. When \\( n = 0 \\), the hallucination process is an identity transformation, introducing no additional operation except for extrapolation. When \\( n = 2 \\), non-linearity is introduced. Then we notice a salient improvement over the baseline. We set \\( n = 3 \\) by default since it gives better performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.707, + 0.603 + ], + "angle": 0, + "content": "4.4. Transferring Features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.748 + ], + "angle": 0, + "content": "The primary goal of representation is to learn transferrable features. We evaluate the transferability of features from the proposed method following the previous protocol [29, 11, 10, 12]. Then, we compare the representation quality by transferring them to downstream tasks, including VOC [18] object detection and COCO [39] object detection and instance segmentation. Notably, we re-implement all these experiments using the same settings in MoCo's detector2 codebase [59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Object Detection on PASCAL VOC. Following the paradigm of previous work [29], we use Faster R-CNN[20] as the object detection method using R50-C4 as the detector [30]. We train the model end-to-end on the trainval2007+2012 and evaluate its performance on test2007. As shown in Table 7, we observe a stable gain range from 0.3 to 0.8 under different metrics on MoCoV1 and MoCoV2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Object Detection and instance segmentation on COCO. We continue to report the detection and segmentation re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16139" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "sults on COCO using Mask R-CNN [30]. Similarly, we use the R50-C4 as the backbone. The is model in an end-to-end way on train2017. Then, the model is evaluated on val2017. Again, the proposed method benefits the object detection and segmentation tasks with various metrics as demonstrated in Table 7." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.196, + 0.196, + 0.212 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.473, + 0.434 + ], + "angle": 0, + "content": "In this work, we propose Hallucinator, which generates additional hard positive pairs for contrastive learning models based on Siamese structure. Hallucinator generates novel data samples in the feature space to provide the training with further contrast without additional computation. We design an asymmetric feature extrapolation to avoid trivial positive pairs and innovatively introduce non-linear hallucination to smooth the generated samples. We empirically prove the effectiveness and generalization capacity of Hallucinator to well-recognized contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Finally, we hope this work could bring the concept of \"Hallucination\" into the SSL domain and unlock future research on sample generations&synthesis in contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.091, + 0.598, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.116, + 0.892, + 0.17 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.173, + 0.894, + 0.228 + ], + "angle": 0, + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.23, + 0.895, + 0.285 + ], + "angle": 0, + "content": "[3] Zhipeng Bao, Yu-Xiong Wang, and Martial Hebert. Bowtie networks: Generative modeling for joint few-shot recognition and novel-view synthesis. arXiv preprint arXiv:2008.06981, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.287, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[5] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.374, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[6] Qingxing Cao, Liang Lin, Yukai Shi, Xiaodan Liang, and Guanbin Li. Attention-aware face hallucination via deep reinforcement learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 690-698, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.445, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.516, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.587, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.645, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.702, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.746, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[13] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. Advances in neural information processing systems, 33:8765-8775, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.116, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16140" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[16] Terrance DeVries and Graham W Taylor. Dataset augmentation in feature space. arXiv preprint arXiv:1702.05538, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.391 + ], + "angle": 0, + "content": "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.393, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[20] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[21] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11017-11027, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[24] Matthew D Grilli and Elizabeth L Glisky. The self-imagination effect: Benefits of a self-referential encoding strategy on cued recall in memory-impaired individuals with neurological damage. Journal of the International Neuropsychological Society, 17(5):929-933, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[25] Liangke Gui, Adrien Bardes, Ruslan Salakhutdinov, Alexander Hauptmann, Martial Hebert, and Yu-Xiong Wang. Learning to hallucinate examples from extrinsic and intrinsic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8701-8711, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[26] Hongyu Guo. Nonlinear mixup: Out-of-manifold data augmentation for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4044-4051, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[27] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pat" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "tern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[28] Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In Proceedings of the IEEE international conference on computer vision, pages 3018-3027, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[30] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.345, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[32] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[33] Wei-Ning Hsu, Yu Zhang, and James Glass. Unsupervised domain adaptation for robust speech recognition via variational autoencoder-based data augmentation. In 2017 IEEE automatic speech recognition and understanding workshop (ASRU), pages 16-23. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[34] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.., 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[36] Varun Kumar, Hadrien Glaude, Cyprien de Lichy, and William Campbell. A closer look at feature space data augmentation for few-shot intent classification. arXiv preprint arXiv:1910.04176, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[37] Kibok Lee, Yian Zhu, Kihyuk Sohn, Chun-Liang Li, Jinwoo Shin, and Honglak Lee. i-mix: A domain-agnostic strategy for contrastive representation learning. arXiv preprint arXiv:2010.08887, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[38] Jiangmeng Li, Wenwen Qiang, Changwen Zheng, Bing Su, and Hui Xiong. Metaug: Contrastive learning via meta feature augmentation. In International Conference on Machine Learning, pages 12964-12978. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[40] Giovanni Mariani, Florian Scheidegger, Roxana Istrate, Costas Bekas, and Cristiano Malossi. Bagan: Data augmentation with balancing gan. arXiv preprint arXiv:1803.09655, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16141" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[42] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[44] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16031-16040, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[45] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[46] Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris, Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object recognition. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[47] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[48] Anshul Shah, Suvrit Sra, Rama Chellappa, and Anoop Cherian. Max-margin contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8220-8230, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[49] Zhiqiang Shen, Zechun Liu, Zhuang Liu, Marios Savvides, Trevor Darrell, and Eric Xing. Un-mix: Rethinking image mixtures for unsupervised visual representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2216-2224, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[50] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[51] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[52] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026-5033. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[54] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[55] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[56] Jing Wu, Naira Hovakimyan, and Jennifer Hobbs. Genco: An auxiliary generator from contrastive learning for enhanced few-shot learning in remote sensing. arXiv preprint arXiv:2307.14612, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[57] Jing Wu, David Pichler, Daniel Marley, David Wilson, Naira Hovakimyan, and Jennifer Hobbs. Extended agriculture-vision: An extension of a large aerial image dataset for agricultural pattern analysis. arXiv preprint arXiv:2303.02460, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[58] Jing Wu, Ran Tao, Pan Zhao, Nicolas F Martin, and Naira Hovakimyan. Optimizing nitrogen management with deep reinforcement learning and crop simulations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1720, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[59] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[62] Yuanmeng Yan, Rumei Li, Sirui Wang, Fuzheng Zhang, Wei Wu, and Weiran Xu. Consert: A contrastive framework for self-supervised sentence representation transfer. arXiv preprint arXiv:2105.11741, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[63] Mang Ye, Xu Zhang, Pong C Yuen, and Shih-Fu Chang. Unsupervised embedding learning via invariant and spreading instance feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6210-6219, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[64] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[65] Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2770-2779, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[66] Ruixiang Zhang, Tong Che, Zoubin Ghahramani, Yoshua Bengio, and Yangqiu Song. Metagan: An adversarial approach to few-shot learning. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[67] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13008-13017, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[68] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16142" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 10306-10315, 2021." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16143" + } + ] +] \ No newline at end of file diff --git a/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_origin.pdf b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a08e3c1368521dc2a5a9a8c1460c0608b71e276 --- /dev/null +++ b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/d1539e1e-e0b7-40bf-9663-58331a6a93fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c6c3e9980bc501843d9f4b33873399c77cb94b33e6e30b47ffad5dc8e4eebf7 +size 696692 diff --git a/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/full.md b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..066797e7df54fa522102d2e3cd52c2c41370804e --- /dev/null +++ b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/full.md @@ -0,0 +1,323 @@ +# Hallucination Improves the Performance of Unsupervised Visual Representation Learning + +Jing Wu1 Jennifer Hobbs2 Naira Hovakimyan1 +1University of Illinois Urbana-Champaign, 2Intelinair +{jingwu6,nhovakim} @illinois, jennifer hobbs08+research@gmail.com + +# Abstract + +Contrastive learning models based on Siamese structure have demonstrated remarkable performance in self-supervised learning. Such a success of contrastive learning relies on two conditions, a sufficient number of positive pairs and adequate variations between them. If the conditions are not met, these frameworks will lack semantic contrast and be fragile on overfitting. To address these two issues, we propose Hallucinator that could efficiently generate additional positive samples for further contrast. The Hallucinator is differentiable and creates new data in the feature space. Thus, it is optimized directly with the pre-training task and introduces nearly negligible computation. Moreover, we reduce the mutual information of hallucinated pairs and smooth them through non-linear operations. This process helps avoid over-confident contrastive learning models during the training and achieves more transformation-invariant feature embeddings. Remarkably, we empirically prove that the proposed Hallucinator generalizes well to various contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Under the linear classification protocol, a stable accuracy gain is achieved, ranging from $0.3\%$ to $3.0\%$ on CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. The improvement is also observed in transferring pre-train encoders to the downstream tasks, including object detection and segmentation. + +# 1. Introduction + +In the recent computer vision community, there has been rapid progress in self-supervised learning (SSL), gradually closing the performance gap with supervised learning [23, 29, 10, 7, 51]. Among the diverse approaches of SSL, contrastive learning, such as MoCoV1&V2 [29, 11], SimCLR [10], and SimSiam [12], shows promising results. Generally, contrastive learning treats each image as one class which will be augmented into two separate views. These two views form one positive pair and should ide + +![](images/86a2670cb8c7046884d7594189c2126b13050df24326ef0a1acc9d0e687eed79.jpg) +Figure 1: The motivation of the proposed hallucination methods. Given one pair of images with the same semantic meaning, such as a pair of horses, a person can envision further similar pairs by imagining one of the horses in different poses and surroundings. If a contrastive learning model could do such hallucination, it could have additional novel pairs to contrast given the same data. Note that this hallucination process is for illustration only. In the implementation, all the hallucinated samples are computed in the feature space. + +ally be close if mapped to feature space. With sufficient contrast in the feature space, contrastive learning models show a strong capacity to learn transformation-invariant features that are transferable to various downstream tasks, such as classification, object detection, and segmentation [18, 30, 39]. + +To ensure sufficient contrast, researchers from previous work address the issue from two essential practices, either introducing large amounts of positive pairs or adding additional variants&transformation among them. For example, SimCLR uses a batch size that generates thousands of positives to facilitate the convergence of models [10]. Work + +from [51, 9] reduces mutual information of positive pairs using stronger data augmentation, i.e., color distortion and jigsaw transformation. Likewise, the work from [44] introduces ContrastiveCrop, and the work from [49] proposes Un-Mix, respectively, to reduce the similar semantic meaning of sample pairs in the original image space. Beyond the data augmentation and image operations, researchers from [68] propose to apply a linear operation to generate hard positive samples in feature space. + +Despite the success of prior approaches, we argue that large batch sizes are not always achievable. Meanwhile, all proposed techniques only focus on improving the original pairs. Given one positive pair of positive samples, humans are born with the amazing ability to come up with additional positives by imagining a sample from different surroundings and perspectives without much effort, as demonstrated in Figure 1. This process of self-imagination, in turn, will benefit the human neurological system, improving recognition capacity [24]. Similarly, if we could empower contrastive learning models with the ability to hallucinate or imagine an object to a novel view, additional positive pairs could be provided for the learning tasks. + +Unfortunately, exploring feasible methods to hallucinate novel positive pairs is challenging. Firstly, while generative models produce realistic images that could form additional positive views [22, 1, 42, 4], realistic data do not necessarily benefit learning tasks [55]. More importantly, applying these approaches forces us to fall back into a computational dilemma to the previous method. In other words, image-level hallucination still suffers from expensive computation as we still need to encode the hallucinated images into feature space. Lastly, if the generated positive pairs are similar to each other, training a discriminative model would be too trivial, thus showing poor generalization capacity [44, 51, 68]. + +Therefore, our key insight is that the sample-generation process should aim for three critical elements: (i) feature-space operation (ii) sufficient variance of positive pairs (iii) a differentiable module optimized directly related to the learning task. To achieve this, we propose Hallucinator to improve the performance of contrastive learning with Siamese structures. The Hallucinator is plugged in after the encoder to manipulate feature vectors and improve the feature-level batch size for further contrast. To ensure adequate variance is introduced, we propose an asymmetric feature extrapolation method inspired by the work from [68]. More importantly, we present a non-linear hallucination process for the extrapolated samples. Such a process is differentiable (i.e. learnable), therefore essentially boosting Hallucinator to generate smooth and task-related features. + +The proposed Hallucinator delivers extra positives and simultaneously enlarges the variance between newly introduced pairs. Moreover, this approach only relies on pos + +itive samples. Therefore, it can be easily applied to any Siamese structure by adding it after the encoders as a plug-and-play module. Without the tedious exploration of hyperparameters and much additional computation, we empirically prove the effectiveness of the proposed Hallucinator on popular contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. We notice a stable improvement ranging from $0.3\%$ to $3.0\%$ under the linear classification protocol, crossing the CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. We also observe that models trained with Hallucinator show better transferability in downstream tasks like object detection and segmentation. + +Our contributions can be summarized as follows: + +- We investigate a critical yet under-explored aspect of contrastive learning: introducing additional positive pairs with further variation in feature space. +- To the best of our knowledge, this is the first attempt to incorporate the concept of "Hallucination" into contrastive representation learning. We propose Hallucinator to realize this idea, which effectively generates smooth and less similar positive feature vectors. The Hallucinator is simple, effective, and agnostic to contrastive frameworks. +- We empirically illustrate that the proposed approach significantly benefits the various contrastive learning models within multiple datasets. + +# 2. Related Works + +In this section, we introduce related literature on contrastive learning and hallucination techniques. + +# 2.1. Contrastive Learning + +The key idea of contrastive learning is to minimize the distance of positive pairs and repulse negative pairs in the feature space [27]. This idea has been successfully applied to unsupervised visual representation tasks, showing promising results in various downstream tasks [10, 29, 11, 23, 12, 2, 32, 17, 54, 41, 43, 50, 57, 60, 63, 61, 8, 37]. One of the breakthroughs of contrastive learning models is SimCLR [10], which introduces a simple but effective visual representation learning method. Without negative pairs, SimCLR learned transformation-invariant representation with a large batch size. Contemporaneous impressive work is MoCo from [29]. To ensure MoCo can be trained smoothly with computational-friendly batch sizes, the authors of MoCo propose a memory bank to store negative features and momentum-updated backbones. After this, SimSiam [12] presents a Siamese-based network that can learn high-quality representation with stop-gradient, successfully avoiding the collapse of contrastive learning models. Other designs of contrastive learning rely on an online network to predict the output of the target network or contrasting cluster assignments [23, 7]. + +![](images/88923fe55a1f0cd92cd9e45f481778611b4102af048527c49f5505c4b9a2ebe0.jpg) +Figure 2: Illustration of contrastive learning (MoCoV2 [11]) with Hallucinator. The Hallucinator is added after the backbones and projector for feature-level manipulation. While Hallucinator feeds original feature vector $q$ forward, it additionally provides hallucinated feature $\hat{q}$ for further contrast. + +To further boost the models' performance, researchers explore diverse ways to hard positives and reduce the mutual information of positive samples so that more transformation-invariant embeddings can be learned [51, 48, 45, 13]. For image-level enhancement, typical works focus on improving the data augmentation in pixel space using Mixup, color distortion, or jigsaw transformation [49, 51, 9]. More recent work is presented by [44] using ContrastiveCrop, which creates better positive views through the localization box. For feature-level operation, the authors from [34] prove the effectiveness of hard negative samples. More recent and close work to this paper is [68], which applies symmetric extrapolation to create hard positives but introduces no non-linearity. Importantly, instead of replacing the original positive samples, this paper proposes a differentiable hallucinator to generate an additional positive sample with less mutual information and better smoothness. Such a setting ensures maximal and adaptive contrast during the training, generalizing well to diverse contrastive learning models. + +# 2.2. Hallucination + +Hallucination is initially proposed to solve the scarcity of data in the classification task [28]. Then, this idea is kept updated and applied in different areas [46, 55, 66, 67, 6, 65, 25, 47], such as object detection, aerial navigation, skeleton-based action recognition, and face generation. While image-level hallucination benefits few-shot recognition by synthesis of novel view [3] or introducing random noises [55], most of the work applies hallucination in the feature space. The work from [28] generates novel class features by transforming shared features in base classes. Authors of [67] build a hallucination framework in the region of interest feature space object to enhance object + +detection performance. More recent work shows that this hallucination mechanism also benefits 3D human pose estimation by generating novel motion sequences [21]. While hallucination is effective in different learning tasks, to the best of our knowledge, the performance and application of hallucination in SSL are fully unexplored. + +# 2.3. Feature-Level Augmentation + +Hallucination relies on effective feature-level augmentation or manipulations. The primary goal of feature augmentation is to extend the limited labeled dataset without relying on expensive computation such as auto-encoder[33], Generative Adversarial Networks[40] or simulation tools[52, 58, 5]. For instance, in the work by [16], a task-agnostic feature augmentation approach is proposed to enrich training data with minimal additional computation. The authors of [36] also explore similar ideas in the domain of few-shot learning. Building upon this, the concept is adapted to sentence representation learning [62, 19] and few-shot learning tasks in remote sensing[56]. More recently, [38] applies feature augmentation based on a meta-learning technique. + +# 3. Method + +In this section, we first introduce the overall process of hallucination for contrastive representation learning in Section 3.1. Secondly, we highlight the center cropping method we used, which is crucial to effective hallucination or generation of new samples in Section 3.2. Then, we introduce the Hallucinator incorporated into our contrastive models in Section 3.3. Finally, we visualize and discuss the critical properties of the hallucination method from two perspectives in Section 3.4: the similarity of positive samples and the uniformity of the feature distribution. + +# 3.1. The Overall Pipeline + +Taking MoCo [11] as an example, we illustrate how a Hallucinator can be plugged into a contrastive learning model in Figure 2. Our architecture takes one image $x$ as input. Then, the input $x$ is augmented into two views $x_{1}$ and $x_{2}$ . Each view will be processed by an encoder consisting of a backbone (e.g., ResNet) and a projector (e.g., an MLP head). After the encoders, output vectors $q$ and $k$ are obtained, forming one positive pair $(q, k)$ . Then, this positive pair $(q, k)$ is fed to a Hallucinator. Notably, Hallucinator is only added to one branch of the framework to generate an additional positive feature $\hat{q}$ . Together with feature vector $k$ , $\hat{q}$ and $k$ form an extra positive pair $(\hat{q}, k)$ during the training. Based on different contrastive learning models, the loss functions keep intact, and the average loss of these two positive pairs is computed for back-propagation. The same paradigm could be applied to SimCLR, SiamSiam and other contrastive learning models. We illustrate further details about pipelines and loss functions of other models in this paper in the Supplementary Material (Section 1.1). + +# 3.2. Center Cropping + +In contrastive learning, data augmentations aim to ensure the performance of pre-trained representations invariant to nuisances. Among all these methods, random crop plays the most critical role in all the contrastive learning models. Generally, views (cropped tiles) generated by random cropping are diversified, successfully covering all the semantic information over the whole image. However, such a cropping method is likely to generate false positives patches [44]. In other words, patches randomly cropped from the original images do not necessarily share the overlapped pixels and sufficient common information. Therefore, these false positive pairs may be fooling models during training, causing representations to be sub-optimal. Importantly, the issue will be exacerbated if we generate further hallucinated samples based on false positive pairs, which misleads the overall training beyond the sweet spot. + +To tackle this issue, we first apply center cropping $\mathbb{C}_{crop}$ to the original image, getting a relatively smaller image $\hat{I}_{x,y}$ . Then, random cropping $\mathbb{R}_{crop}$ is applied to $\hat{I}_{x,y}$ . More specifically, the center cropping can be formulated as + +$$ +\hat {I} _ {x, y} = \mathbb {C} _ {\text {c r o p}} \left(I _ {x, y}, p\right), \tag {1} +$$ + +where $I_{x,y}$ is the input image with $(x,y)$ as the coordinate of the images' center. After center cropping, we keep the center of $\hat{I}_{x,y}$ unchanged. Meanwhile, with the original shape of $I_{x,y}$ defined as $(h,w)$ , we define the shape of cropped image $\hat{I}_{x,y}$ as $(\hat{h},\hat{w})$ . The $p$ denotes a ratio of cropped length over the original length, i.e., $p = \frac{\hat{w}}{w} = \frac{\hat{h}}{h}$ . Only if particularly mentioned, we set $p = 0.5$ for all experiments in this paper. + +![](images/77c72261a5c95c0e15c94fa6898ad99109334763745ac2a5d6fb76af1739a5fb.jpg) +Figure 3: The Hallucinator. Stage 1: The original feature vector $q$ is extrapolated to the opposite direction of feature vector $k$ , forming $\hat{q}$ in a linear way. Stage 2: We introduce non-linear transformation to smooth extrapolated features concatenated with $q$ and $q'$ . Stage 3: output original $q$ and hallucinated $\hat{q}$ . + +While center cropping effectively avoids false positive pairs, it reduces the operable region for random cropping and generates positive views with a similar appearance. We, therefore, adopt center-suppressed sampling [44] with a sampling method following a beta distribution $\beta (\alpha ,\alpha)$ (i.e., a U-shaped distribution). Concretely, $\beta (\alpha ,\alpha)$ assigns a lower probability to the center of the $\hat{I}_{x,y}$ and gives greater probability to its boundary, increasing the variance between views $x_{1}$ and $x_{2}$ . Together, we summarize the process to obtain these two views as + +$$ +x _ {1} = T (\mathbb {R} _ {c r o p} (\hat {I} _ {x, y} | \alpha)), \quad \mathrm {s . t .} \quad \alpha < 1 +$$ + +$$ +x _ {2} = T \left(\mathbb {R} _ {c r o p} \left(I _ {x, y} | \alpha\right)\right), \quad \text {s . t .} \quad \alpha < 1, \tag {2} +$$ + +where $T$ denotes data augmentations, including color jittering, random grayscale, Gaussian blur and horizontal flipping. $\mathbb{R}_{crop}(\dots|\alpha)$ represents random cropping following the $\beta (\alpha ,\alpha)$ distribution. $\alpha$ is set to less than 1 to ensure an increasing sampling probability as the pixel's coordinates go beyond the center. A visualization of the center sampling method can be found in Supplementary Material (Section 1.2). + +# 3.3. Hallucinator + +Asymmetric Feature Extrapolation. The first objective of this module is to introduce an additional positive pair without introducing extra computations. Therefore, the feature-level operation is preferred compared to image-level operations. To achieve this, Hallucinator is plugged in after the views $x_{1}$ and $x_{2}$ are encoded. As a result, the hallucinated (generated) feature is purely based on the two feature vectors $q$ and $k$ . + +Meanwhile, since harder positives improve pre-trained encoders' generalization capacity [51], the hallucinated features in positive pairs are favorable if they share less mutual information. Previous work illustrates that a symmetric + +positive extrapolation is effective in generating hard examples for MoCo [68]. Concretely, two positive features are combined with weighted addition, pushing positive features apart. However, the hallucination process is asymmetric, i.e., Hallucinator is only added in one of the branches of the model. Then, we propose to apply the single-side feature extrapolation to the feature vector $q$ , as shown in the first stage of Figure 3. Additionally, we simplify the sampling strategy of weights for extrapolation from a beta distribution to a uniform distribution. To be specific, we summarize the positive extrapolation in our method as follows: + +$$ +q ^ {\prime} = (1 + \lambda) q - \lambda k \quad \text {s . t .} \quad \lambda \sim U (\beta_ {1}, \beta_ {2}), \tag {3} +$$ + +where $\lambda$ is sampled from a uniform distribution $U(\beta_{1},\beta_{2})$ . $\beta_{1}$ and $\beta_{2}$ , which are the boundary of the uniform distribution, are set to 0 and 0.1 by default. + +Hallucination. Positive extrapolation is based on mixup [64], i.e., a linear transformation. While positive extrapolation has been proven beneficial to generating hard examples, this linear feature transformation might have a relatively limited capacity to synthesize new feature vectors. This assumption is based on the more satisfactory performance of the non-linear mixup over the original one [26]. Similarly, if non-linearity is introduced to the feature generation, the generated vector will benefit more from the training and boost the performance of downstream tasks. + +To empower our model with the capacity of non-linear fitting, we introduce the hallucination process. Specifically, we first concatenate $q'$ from the equation 3 and the feature vector $q$ together. Then, we use the concatenated feature $(q, q')$ as an input of a non-linear transformation function $H_{\theta}(.)$ that can be instantiated with $n$ linear layers and a ReLU layer between two successive layers. When $n = 0$ , $H_{\theta}(.)$ is a non-parametric module, forming an identity function. Such a setting performance is relatively sub-optimal, as shown in the ablation study in Table 6. Empirically, we find that $n = \{2, 3\}$ performs well as the hallucinator becomes non-linear and more powerful. We set $n = 3$ by default as its results are slightly better. With the transformation function $H_{\theta}(.)$ , the hallucinated feature is defined as + +$$ +\hat {q} = H _ {\theta} \left(q, q ^ {\prime}\right), \tag {4} +$$ + +where $\theta$ is the parameters of $H_{\theta}(.)$ . Notably, $H_{\theta}(.)$ is differentiable, allowing us to back-propagate the loss of contrastive learning. Therefore, we update not just the parameters of the encoders and projectors but also the parameters $\theta$ of the hallucinator. The second stage of Figure 3 illustrates the proposed hallucination process. Following that, we take $q$ and $\hat{q}$ as the output of Hallucinator. + +![](images/3160a7a5fa9552b083afff264f3ca41829110ceb0952bbf4cb5427ef0c58835e.jpg) +Figure 4: Similarity of positive pairs in training. Smaller values indicate less mutual information and better representation [51, 44, 68]. As Hallucinator incorporates nonlinearity extrapolation, it guarantees smooth training and harder positive features. + +# 3.4. Discussion and Visualization + +To better understand the behavior of Hallucinator, we discuss two critical properties that may contribute to and explain its effectiveness. For visualization, we train MoCoV2 [11] with a standard ResNet-18 [31] on CIFar-10 [35]. + +Similarity of Positive Pair. We first investigate the similarity of positive pairs during the training process. Concretely, we quantize mutual information using the cosine similarity of positive pairs $S$ : + +$$ +S \triangleq \frac {q \cdot k}{\| q \| \| k \|.} \tag {5} +$$ + +While hard positives share less mutual information, thus having a smaller cosine similarity value, the performance of downstream tasks will be enhanced [68]. Based on Figure 4, the proposed Hallucinator generates harder positives with smaller values of similarity. Consequently, it helps contrastive learning models obtain more nuisances-invariant features. More importantly, different from linear symmetric extrapolation, our training curve is relatively stable without many oscillations introduced. This observation indicates that Hallucinator is successfully optimized with the overall framework. Meanwhile, hallucinated features nicely fit into the contrastive learning task. + +Uniformity. We continue to analyze the performance of the proposed model from the perspective of uniformity. Notably, feature vectors from contrastive learning should be roughly uniformly distributed on a unit hyper-sphere, which ensures maximal information is preserved in the feature space [53]. In other words, the closer the feature distribution is to the uniform distribution, the more the feature benefits downstream tasks. To quantize uniformity, we follow previous work [53] to compute the average value of the + +![](images/3a3759c2fe91044bb345a222683d2bc926222454b1b4b918efcf25e5c0601e49.jpg) +Figure 5: A measure of uniformity based on $G_{2}$ potential. We plot 10,000 feature vectors with Gaussian kernel density estimation (KDE) in $\mathbb{R}^2$ . The left subplot illustrates the feature vectors from a uniform distribution. The three feature distributions on the right in the first row visualize the features from MoCoV2 [11]. The other three feature distributions in the second row demonstrate MoCoV2 with Hallucinator. Hallucinator benefits the uniformity with a smaller value of $G_{2}$ . + +![](images/74ea5fecbf25104e2ec7e48a9b3f3757dc6ffe3d6903d8a349432ca0a5587ac3.jpg) + +![](images/e9410cde9f6cf7d808cdface506514c36df32234665e9edd1aadf2a6297d64ed.jpg) + +![](images/2cee15a31f715bd985f6ad1cefdd5490a90f1040d2aafd86ef40701744362c98.jpg) + +![](images/18f9d11ba3cbfd9bbc2f37fa8aff1b56a6f015e1b24c7333ada7e2168f6bd6c6.jpg) + +![](images/84d42f0c92e6afdf63cca55f6f5c4f42ba1c6af7c979057308b443c3fbed06a0.jpg) + +![](images/a17d35b617352de88304204ae4f00aa78f16fa2580fa4d60b30bfd254f2eac18.jpg) + +
DatasetCIFAR-10CIFAR-100Tiny ImageNetSTL-10
HallucinatorX✓X✓X✓X✓
MoCoV188.3188.9460.9461.8144.6545.5388.1990.09
MoCoV287.2189.2359.7061.2647.1247.9589.3290.46
SimCLR89.6690.1160.9461.4345.2246.3089.0789.98
SimSiam90.4790.7863.3964.3843.6644.9687.7988.16
+ +Table 1: Linear classification results for different contrastive methods and datasets in small scales. We adopt ResNet-18 as the backbone and report the classification results with or without Hallucinator. + +Gaussian potential kernel (i.e., Radial Basis Function kernel) of positive features: + +$$ +G _ {t} (q, k) \triangleq e ^ {- t \| q - k \| _ {2} ^ {2}} \quad \text {s . t .} \quad t > 0, \tag {6} +$$ + +where $t$ is a fixed parameter set as 2 for all the experiments. We visualize the feature vectors by mapping them to two-dimension feature space and applying $l_{2}$ normalization in Figure 5. $G_{2} = 0.2070$ for uniformly distributed samples, whereas features from MoCoV2 have $G_{2} = 0.2089$ . If we plugin Hallucinator into MoCo, Hallucinator provides further contrast during the training with extra positives introduced, giving more uniformly distributed features and a decreased $G_{2}$ value,i.e., 0.2081. Additionally, we visualize the features of two classes of Cifar-10. Each of these features is well-clustered. With better uniformity, clusters' overlapping decreases, forming more linearly separable features. Therefore, we could observe that the overlapping of feature clusters between class 1 and class 6 in MoCo is larger than the one with Hallucinator plugged in. + +# 4. Experiments and Results + +In this section, we conduct various experiments on different datasets and contrastive learning models to demonstrate the effectiveness of the Hallucinator. Firstly, we introduce the datasets and details in experiments in Section 4.1. Then, we continue to evaluate the performance of the proposed method using linear probing protocol following the paradigm of previous work [10, 11, 12, 23] in Section 4.2. Following this, we conduct ablation studies to understand how each module contributes to the final results in Section 4.3. Lastly, we show the transferability of pre-trained encoders in downstream tasks requiring dense pixel predictions, including object detection and semantic segmentation in Section 4.4. + +# 4.1. Datasets and Training Details + +Datasets and Baseline Models. We first evaluate the performance of Hallucinator based on a wide range of datasets crossing different scales. Specifically, these datasets in + +
MethodBackboneEpochIN-200IN-1K
MoCoV1ResNet-5010062.1957.27
MoCoV1(Ours)ResNet-5010063.4659.17
MoCoV2ResNet-5010062.5764.41
MoCoV2(Ours)ResNet-5010063.5864.97
SimCLRResNet-5010062.2261.23
SimCLR(Ours)ResNet-5010063.0261.71
SimSiamResNet-5010062.8063.11
SimSiam(Ours)ResNet-5010063.5263.55
+ +Table 2: Linear classification results on IN-200 and IN-1K. Our method plugged in the proposed Hallucinator in baseline methods. All the models are pre-trained for 100 epochs and use identical training settings for fair comparisons. + +clude CIFAR-10, CIFAR-R100 [35], Tiny ImageNet, STL-10 [14] and ImageNet [15]. Meanwhile, we demonstrate the performance of Hallucinator in several popular contrastive learning frameworks, including MoCoV1 [29], MoCoV2 [11], SimCLR [10] and SimSiam [12]. + +Training and Evaluation Details. Importantly, we strictly adopt the same training settings when evaluating the Hallucinator's performance. While additional gain could be achieved with further exploration of hyper-parameter, it is not the focus of this work. The ultimate goal of the proposed Hallucinator is to provide further feature-level contrast for self-supervised learning. With the Hallucinator plugin, it benefits diversified self-supervised learning methods regardless of their type of backbones and hyper-parameters of training. + +Datasets on relatively small scales include CIFAR-10&100, Tiny ImageNet and STL10. To ensure fair comparisons, we keep training settings the same over all these datasets. We follow the paradigm of previous work for pretraining [44]. Concretely, we pre-train contrastive learning models for 500 epochs with a batch size of 512 and ResNet-18 as the backbone. To optimize the models, we use an SGD optimizer and a cosine-annealed learning rate of 0.5. In the linear classification task, we train models for 100 epochs. With the initial learning set as 10, we divide it by 10 at the 60th and 80th epochs. + +As the scale of the dataset expands to ImageNet, we use ResNet-50 as the backbone. For MoCoV1, MoCoV2 and SimSiam, all the pre-training settings follow the original work. We use a batch size of 512 and an SGD optimizer with a cosine-annealed learning rate of 0.5 for SimCLR, the same as the settings in [44]. For the linear classification task, we follow the original work in [29], training the model for 100 epochs. We set the initial learning rate to 30. Then, we divide the learning by 10 at the 60th and 80th with a weight decay of 0. + +We set $p = 0.5$ and $\alpha = 0.6$ for center cropping and + +
Acc.(%)Center CroppingAsymmetric ExtrapolationHallucination
62.57✗✗✗
63.58(+1.01)✓✓✓
62.58(+0.01)✓
62.82(+0.25)✓
62.84(+0.27)✓
63.07(+0.50)✓✓
+ +Table 3: Ablation of the different modules in the proposed method. + +
p (%)908070605040
Acc.(%)63.1163.1963.3363.5663.5863.21
+ +Table 4: Ablation of classification results w.r.t the $p$ value. + +
(β1,β2)(-0.5, 0.0)(0.0, 0.5)(0.5, 1.0)(1.0, 1.5)(0.0, 1.0)
Acc.(%)62.8063.3263.5563.4263.58
+ +Table 5: Ablation of classification results w.r.t the extrapolation range. + +
layer #: n01234
Acc.(%)62.8563.0363.5763.5863.55
+ +Table 6: Ablation of classification results w.r.t the number of linear layers $n$ . + +center-suppressed sampling, respectively. For asymmetric linear extrapolation, we set $\beta_{1} = 0$ and $\beta_{2} = 0.1$ . In the Hallucination process, $n = 3$ for all the results reported in this paper. All the experiments are conducted on a server with 8 GPUs. + +# 4.2. Linear Classification Protocol + +Following the previous protocol, we first evaluate the proposed method by linear classification of frozen features. For each dataset, we report the top-1 classification accuracy on the validation set. + +Results on Small-Scale Datasets. The classification results on Cifar10&100, Tiny ImageNet and STL-10 are reported in Table 1. With Hallucinator introduced, we notice a stable improvement over the baselines ranging from $0.31\%$ to $1.98\%$ . Notably, such improvements do not introduce extra computations and generalize well to various models. + +Results on ImageNet. For the results of ImageNet, we report the results at two different scales. First, we evaluate its performance on standard IN-1K (i.e. ImageNet-1K), which consists of 1000 classes. Second, we test the proposed method in IN-200 (i.e. ImageNet-200) with 200 randomly selected classes. We report the corresponding results in Table 2. We found that Hallucinator essentially benefits + +
Method1N-1k Top-1VOC detectionCOCO detectionCOCO instance segmentation
APAP50AP75APbbAPbb50APbb75APmkAPmk50APmk75
Random init-33.860.257.2726.444.027.829.346.930.8
Supervised76.153.581.358.838.258.241.233.354.735.2
InfoMin [51]70.157.682.764.639.058.542.034.155.236.3
MoCoV1 [29]60.655.981.562.638.558.341.633.654.835.6
MoCoV1(Ours)63.856.781.863.238.958.541.933.855.236.0
MoCoV2 [11]67.557.082.463.639.058.641.934.255.436.2
MoCoV2(Ours)68.057.482.763.939.358.842.334.655.536.4
+ +Table 7: Fine-tuning results on object detection tasks on PASCAL VOC and COCO, and instance segmentation on COCO. All models are pre-trained for 200 epochs on ImageNet-1K. + +MoCoV1 with $1.27\%$ and $1.89\%$ improvements for IN-200 and IN-1K accordingly. For MoCoV2, SimSiam and SimCLR, we notice a gain in accuracy ranging from $0.44\%$ to $1.01\%$ . On average, the gains are more salient in IN-200. + +# 4.3. Ablation Studies + +In this section of ablation studies, we investigate essential modules in our proposed methods, including center cropping, asymmetric extrapolation and the process of hallucination. We conduct experiments using MoCoV2 with ResNet-50 as the backbone. We report the results of the classification task in IN-200 for all the following experiments. + +Contributions of Modules. We first report the results with or without crucial modules introduced in Section 3. According to Table 3, center cropping shows a similar performance to the original cropping method, successfully covering major semantic information of images. However, it successfully avoids false positives in pre-training, which is critical for hallucination. Asymmetric extrapolation benefits the performance of representation learning with reduced mutual information. This observation is consistent with the results shown in the symmetric extrapolation [68]. If we combine asymmetric extrapolation and the hallucination method, the performance of the model could be further boosted. However, it is still sub-optimal because of possible false positives in cropping. + +Center Cropping. In this work, we apply center cropping as a critical tool in one branch of contrastive learning to avoid false positives. Therefore, we investigate how the proposed model is influenced by the length ratio $p$ in Table 4. The classification results are favorable when $p$ is in the range (0.5, 0.6). We notice a performance drop when $p$ is less than 0.5. This drop is because the center cropping failed to cover all semantic information in the images. + +Extrapolation Range. We continue to report the extrapolation range given its differences in sampling method and + +single-side architecture compared with the previous method [68]. Based on Table 5, the classification accuracy drops if we set $\beta_{1} = -0.5$ and $\beta_{2} = 0$ , respectively. Such observation is because hallucinated positives share more mutual information, thus less beneficial to training. Generally, extrapolation boosts accuracy by outperforming the baseline by at least $0.75\%$ . However, the improvement decreases when $\beta_{1} > 1.0$ . We eventually set $\beta_{1} = 0.0$ and $\beta_{2} = 1.0$ , given its best performance. + +Hallucination. We investigate how the number of linear layers influences the model's performance. When $n = 0$ , the hallucination process is an identity transformation, introducing no additional operation except for extrapolation. When $n = 2$ , non-linearity is introduced. Then we notice a salient improvement over the baseline. We set $n = 3$ by default since it gives better performance. + +# 4.4. Transferring Features + +The primary goal of representation is to learn transferrable features. We evaluate the transferability of features from the proposed method following the previous protocol [29, 11, 10, 12]. Then, we compare the representation quality by transferring them to downstream tasks, including VOC [18] object detection and COCO [39] object detection and instance segmentation. Notably, we re-implement all these experiments using the same settings in MoCo's detector2 codebase [59]. + +Object Detection on PASCAL VOC. Following the paradigm of previous work [29], we use Faster R-CNN[20] as the object detection method using R50-C4 as the detector [30]. We train the model end-to-end on the trainval2007+2012 and evaluate its performance on test2007. As shown in Table 7, we observe a stable gain range from 0.3 to 0.8 under different metrics on MoCoV1 and MoCoV2. + +Object Detection and instance segmentation on COCO. We continue to report the detection and segmentation re + +sults on COCO using Mask R-CNN [30]. Similarly, we use the R50-C4 as the backbone. The is model in an end-to-end way on train2017. Then, the model is evaluated on val2017. Again, the proposed method benefits the object detection and segmentation tasks with various metrics as demonstrated in Table 7. + +# 5. Conclusion + +In this work, we propose Hallucinator, which generates additional hard positive pairs for contrastive learning models based on Siamese structure. Hallucinator generates novel data samples in the feature space to provide the training with further contrast without additional computation. We design an asymmetric feature extrapolation to avoid trivial positive pairs and innovatively introduce non-linear hallucination to smooth the generated samples. We empirically prove the effectiveness and generalization capacity of Hallucinator to well-recognized contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Finally, we hope this work could bring the concept of "Hallucination" into the SSL domain and unlock future research on sample generations&synthesis in contrastive learning. + +# References + +[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. +[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. +[3] Zhipeng Bao, Yu-Xiong Wang, and Martial Hebert. Bowtie networks: Generative modeling for joint few-shot recognition and novel-view synthesis. arXiv preprint arXiv:2008.06981, 2020. +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. +[5] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. +[6] Qingxing Cao, Liang Lin, Yukai Shi, Xiaodan Liang, and Guanbin Li. Attention-aware face hallucination via deep reinforcement learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 690-698, 2017. +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. +[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021. +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. +[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. +[13] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. Advances in neural information processing systems, 33:8765-8775, 2020. +[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on + +artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[16] Terrance DeVries and Graham W Taylor. Dataset augmentation in feature space. arXiv preprint arXiv:1702.05538, 2017. +[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021. +[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009. +[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. +[20] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. +[21] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11017-11027, 2022. +[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. +[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. +[24] Matthew D Grilli and Elizabeth L Glisky. The self-imagination effect: Benefits of a self-referential encoding strategy on cued recall in memory-impaired individuals with neurological damage. Journal of the International Neuropsychological Society, 17(5):929-933, 2011. +[25] Liangke Gui, Adrien Bardes, Ruslan Salakhutdinov, Alexander Hauptmann, Martial Hebert, and Yu-Xiong Wang. Learning to hallucinate examples from extrinsic and intrinsic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8701-8711, 2021. +[26] Hongyu Guo. Nonlinear mixup: Out-of-manifold data augmentation for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4044-4051, 2020. +[27] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pat + +tern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006. +[28] Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In Proceedings of the IEEE international conference on computer vision, pages 3018-3027, 2017. +[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. +[30] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. +[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[32] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020. +[33] Wei-Ning Hsu, Yu Zhang, and James Glass. Unsupervised domain adaptation for robust speech recognition via variational autoencoder-based data augmentation. In 2017 IEEE automatic speech recognition and understanding workshop (ASRU), pages 16-23. IEEE, 2017. +[34] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. +[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.., 2009. +[36] Varun Kumar, Hadrien Glaude, Cyprien de Lichy, and William Campbell. A closer look at feature space data augmentation for few-shot intent classification. arXiv preprint arXiv:1910.04176, 2019. +[37] Kibok Lee, Yian Zhu, Kihyuk Sohn, Chun-Liang Li, Jinwoo Shin, and Honglak Lee. i-mix: A domain-agnostic strategy for contrastive representation learning. arXiv preprint arXiv:2010.08887, 2020. +[38] Jiangmeng Li, Wenwen Qiang, Changwen Zheng, Bing Su, and Hui Xiong. Metaug: Contrastive learning via meta feature augmentation. In International Conference on Machine Learning, pages 12964-12978. PMLR, 2022. +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. +[40] Giovanni Mariani, Florian Scheidegger, Roxana Istrate, Costas Bekas, and Cristiano Malossi. Bagan: Data augmentation with balancing gan. arXiv preprint arXiv:1803.09655, 2018. +[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. + +[42] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018. +[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. +[44] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16031-16040, 2022. +[45] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020. +[46] Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris, Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object recognition. Advances in neural information processing systems, 31, 2018. +[47] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023. +[48] Anshul Shah, Suvrit Sra, Rama Chellappa, and Anoop Cherian. Max-margin contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8220-8230, 2022. +[49] Zhiqiang Shen, Zechun Liu, Zhuang Liu, Marios Savvides, Trevor Darrell, and Eric Xing. Un-mix: Rethinking image mixtures for unsupervised visual representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2216-2224, 2022. +[50] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020. +[51] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. +[52] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026-5033. IEEE, 2012. +[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. +[54] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. + +[55] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018. +[56] Jing Wu, Naira Hovakimyan, and Jennifer Hobbs. Genco: An auxiliary generator from contrastive learning for enhanced few-shot learning in remote sensing. arXiv preprint arXiv:2307.14612, 2023. +[57] Jing Wu, David Pichler, Daniel Marley, David Wilson, Naira Hovakimyan, and Jennifer Hobbs. Extended agriculture-vision: An extension of a large aerial image dataset for agricultural pattern analysis. arXiv preprint arXiv:2303.02460, 2023. +[58] Jing Wu, Ran Tao, Pan Zhao, Nicolas F Martin, and Naira Hovakimyan. Optimizing nitrogen management with deep reinforcement learning and crop simulations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1720, 2022. +[59] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. +[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. +[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. +[62] Yuanmeng Yan, Rumei Li, Sirui Wang, Fuzheng Zhang, Wei Wu, and Weiran Xu. Consert: A contrastive framework for self-supervised sentence representation transfer. arXiv preprint arXiv:2105.11741, 2021. +[63] Mang Ye, Xu Zhang, Pong C Yuen, and Shih-Fu Chang. Unsupervised embedding learning via invariant and spreading instance feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6210-6219, 2019. +[64] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. +[65] Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2770-2779, 2019. +[66] Ruixiang Zhang, Tong Che, Zoubin Ghahramani, Yoshua Bengio, and Yangqiu Song. Metagan: An adversarial approach to few-shot learning. Advances in neural information processing systems, 31, 2018. +[67] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13008-13017, 2021. +[68] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the \ No newline at end of file diff --git a/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/images.zip b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..611cca82ae25a63d7be203520a45564a847b710c --- /dev/null +++ b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6079a52b2ac2c2059e2945f496ff0d81d9f9f58d9dcd292860273b293cbda9e7 +size 405777 diff --git a/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/layout.json b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e0773fdd03fbfa9ad73c83873360b03f367ddf5 --- /dev/null +++ b/2023/Hallucination Improves the Performance of Unsupervised Visual Representation Learning/layout.json @@ -0,0 +1,9376 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 544, + 140 + ], + "type": "text", + "content": "Hallucination Improves the Performance of Unsupervised Visual Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 160, + 440, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 160, + 440, + 204 + ], + "spans": [ + { + "bbox": [ + 129, + 160, + 440, + 204 + ], + "type": "text", + "content": "Jing Wu1 Jennifer Hobbs2 Naira Hovakimyan1 \n1University of Illinois Urbana-Champaign, 2Intelinair \n{jingwu6,nhovakim} @illinois, jennifer hobbs08+research@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "text", + "content": "Contrastive learning models based on Siamese structure have demonstrated remarkable performance in self-supervised learning. Such a success of contrastive learning relies on two conditions, a sufficient number of positive pairs and adequate variations between them. If the conditions are not met, these frameworks will lack semantic contrast and be fragile on overfitting. To address these two issues, we propose Hallucinator that could efficiently generate additional positive samples for further contrast. The Hallucinator is differentiable and creates new data in the feature space. Thus, it is optimized directly with the pre-training task and introduces nearly negligible computation. Moreover, we reduce the mutual information of hallucinated pairs and smooth them through non-linear operations. This process helps avoid over-confident contrastive learning models during the training and achieves more transformation-invariant feature embeddings. Remarkably, we empirically prove that the proposed Hallucinator generalizes well to various contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Under the linear classification protocol, a stable accuracy gain is achieved, ranging from " + }, + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 47, + 258, + 290, + 558 + ], + "type": "text", + "content": " on CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. The improvement is also observed in transferring pre-train encoders to the downstream tasks, including object detection and segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "In the recent computer vision community, there has been rapid progress in self-supervised learning (SSL), gradually closing the performance gap with supervised learning [23, 29, 10, 7, 51]. Among the diverse approaches of SSL, contrastive learning, such as MoCoV1&V2 [29, 11], SimCLR [10], and SimSiam [12], shows promising results. Generally, contrastive learning treats each image as one class which will be augmented into two separate views. These two views form one positive pair and should ide" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 232, + 523, + 404 + ], + "blocks": [ + { + "bbox": [ + 331, + 232, + 523, + 404 + ], + "lines": [ + { + "bbox": [ + 331, + 232, + 523, + 404 + ], + "spans": [ + { + "bbox": [ + 331, + 232, + 523, + 404 + ], + "type": "image", + "image_path": "86a2670cb8c7046884d7594189c2126b13050df24326ef0a1acc9d0e687eed79.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 427, + 547, + 548 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 548 + ], + "type": "text", + "content": "Figure 1: The motivation of the proposed hallucination methods. Given one pair of images with the same semantic meaning, such as a pair of horses, a person can envision further similar pairs by imagining one of the horses in different poses and surroundings. If a contrastive learning model could do such hallucination, it could have additional novel pairs to contrast given the same data. Note that this hallucination process is for illustration only. In the implementation, all the hallucinated samples are computed in the feature space." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 569, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 641 + ], + "type": "text", + "content": "ally be close if mapped to feature space. With sufficient contrast in the feature space, contrastive learning models show a strong capacity to learn transformation-invariant features that are transferable to various downstream tasks, such as classification, object detection, and segmentation [18, 30, 39]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "To ensure sufficient contrast, researchers from previous work address the issue from two essential practices, either introducing large amounts of positive pairs or adding additional variants&transformation among them. For example, SimCLR uses a batch size that generates thousands of positives to facilitate the convergence of models [10]. Work" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16132" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": "from [51, 9] reduces mutual information of positive pairs using stronger data augmentation, i.e., color distortion and jigsaw transformation. Likewise, the work from [44] introduces ContrastiveCrop, and the work from [49] proposes Un-Mix, respectively, to reduce the similar semantic meaning of sample pairs in the original image space. Beyond the data augmentation and image operations, researchers from [68] propose to apply a linear operation to generate hard positive samples in feature space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 337 + ], + "type": "text", + "content": "Despite the success of prior approaches, we argue that large batch sizes are not always achievable. Meanwhile, all proposed techniques only focus on improving the original pairs. Given one positive pair of positive samples, humans are born with the amazing ability to come up with additional positives by imagining a sample from different surroundings and perspectives without much effort, as demonstrated in Figure 1. This process of self-imagination, in turn, will benefit the human neurological system, improving recognition capacity [24]. Similarly, if we could empower contrastive learning models with the ability to hallucinate or imagine an object to a novel view, additional positive pairs could be provided for the learning tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 339, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 495 + ], + "type": "text", + "content": "Unfortunately, exploring feasible methods to hallucinate novel positive pairs is challenging. Firstly, while generative models produce realistic images that could form additional positive views [22, 1, 42, 4], realistic data do not necessarily benefit learning tasks [55]. More importantly, applying these approaches forces us to fall back into a computational dilemma to the previous method. In other words, image-level hallucination still suffers from expensive computation as we still need to encode the hallucinated images into feature space. Lastly, if the generated positive pairs are similar to each other, training a discriminative model would be too trivial, thus showing poor generalization capacity [44, 51, 68]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 496, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 675 + ], + "type": "text", + "content": "Therefore, our key insight is that the sample-generation process should aim for three critical elements: (i) feature-space operation (ii) sufficient variance of positive pairs (iii) a differentiable module optimized directly related to the learning task. To achieve this, we propose Hallucinator to improve the performance of contrastive learning with Siamese structures. The Hallucinator is plugged in after the encoder to manipulate feature vectors and improve the feature-level batch size for further contrast. To ensure adequate variance is introduced, we propose an asymmetric feature extrapolation method inspired by the work from [68]. More importantly, we present a non-linear hallucination process for the extrapolated samples. Such a process is differentiable (i.e. learnable), therefore essentially boosting Hallucinator to generate smooth and task-related features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "The proposed Hallucinator delivers extra positives and simultaneously enlarges the variance between newly introduced pairs. Moreover, this approach only relies on pos" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "itive samples. Therefore, it can be easily applied to any Siamese structure by adding it after the encoders as a plug-and-play module. Without the tedious exploration of hyperparameters and much additional computation, we empirically prove the effectiveness of the proposed Hallucinator on popular contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. We notice a stable improvement ranging from " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " under the linear classification protocol, crossing the CIFAR10&100, Tiny ImageNet, STL-10 and ImageNet. We also observe that models trained with Hallucinator show better transferability in downstream tasks like object detection and segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 216, + 517, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 216, + 517, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 216, + 517, + 226 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 228, + 545, + 383 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 316, + 228, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 545, + 262 + ], + "type": "text", + "content": "- We investigate a critical yet under-explored aspect of contrastive learning: introducing additional positive pairs with further variation in feature space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 264, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 545, + 345 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first attempt to incorporate the concept of \"Hallucination\" into contrastive representation learning. We propose Hallucinator to realize this idea, which effectively generates smooth and less similar positive feature vectors. The Hallucinator is simple, effective, and agnostic to contrastive frameworks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 347, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 347, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 316, + 347, + 545, + 383 + ], + "type": "text", + "content": "- We empirically illustrate that the proposed approach significantly benefits the various contrastive learning models within multiple datasets." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 393, + 397, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 397, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 397, + 405 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 413, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 545, + 437 + ], + "type": "text", + "content": "In this section, we introduce related literature on contrastive learning and hallucination techniques." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 444, + 429, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 429, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 429, + 456 + ], + "type": "text", + "content": "2.1. Contrastive Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "The key idea of contrastive learning is to minimize the distance of positive pairs and repulse negative pairs in the feature space [27]. This idea has been successfully applied to unsupervised visual representation tasks, showing promising results in various downstream tasks [10, 29, 11, 23, 12, 2, 32, 17, 54, 41, 43, 50, 57, 60, 63, 61, 8, 37]. One of the breakthroughs of contrastive learning models is SimCLR [10], which introduces a simple but effective visual representation learning method. Without negative pairs, SimCLR learned transformation-invariant representation with a large batch size. Contemporaneous impressive work is MoCo from [29]. To ensure MoCo can be trained smoothly with computational-friendly batch sizes, the authors of MoCo propose a memory bank to store negative features and momentum-updated backbones. After this, SimSiam [12] presents a Siamese-based network that can learn high-quality representation with stop-gradient, successfully avoiding the collapse of contrastive learning models. Other designs of contrastive learning rely on an online network to predict the output of the target network or contrasting cluster assignments [23, 7]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16133" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 224 + ], + "type": "image", + "image_path": "88923fe55a1f0cd92cd9e45f481778611b4102af048527c49f5505c4b9a2ebe0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 546, + 281 + ], + "type": "text", + "content": "Figure 2: Illustration of contrastive learning (MoCoV2 [11]) with Hallucinator. The Hallucinator is added after the backbones and projector for feature-level manipulation. While Hallucinator feeds original feature vector " + }, + { + "bbox": [ + 46, + 243, + 546, + 281 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 243, + 546, + 281 + ], + "type": "text", + "content": " forward, it additionally provides hallucinated feature " + }, + { + "bbox": [ + 46, + 243, + 546, + 281 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 46, + 243, + 546, + 281 + ], + "type": "text", + "content": " for further contrast." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 540 + ], + "type": "text", + "content": "To further boost the models' performance, researchers explore diverse ways to hard positives and reduce the mutual information of positive samples so that more transformation-invariant embeddings can be learned [51, 48, 45, 13]. For image-level enhancement, typical works focus on improving the data augmentation in pixel space using Mixup, color distortion, or jigsaw transformation [49, 51, 9]. More recent work is presented by [44] using ContrastiveCrop, which creates better positive views through the localization box. For feature-level operation, the authors from [34] prove the effectiveness of hard negative samples. More recent and close work to this paper is [68], which applies symmetric extrapolation to create hard positives but introduces no non-linearity. Importantly, instead of replacing the original positive samples, this paper proposes a differentiable hallucinator to generate an additional positive sample with less mutual information and better smoothness. Such a setting ensures maximal and adaptive contrast during the training, generalizing well to diverse contrastive learning models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 550, + 134, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 134, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 134, + 562 + ], + "type": "text", + "content": "2.2. Hallucination" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Hallucination is initially proposed to solve the scarcity of data in the classification task [28]. Then, this idea is kept updated and applied in different areas [46, 55, 66, 67, 6, 65, 25, 47], such as object detection, aerial navigation, skeleton-based action recognition, and face generation. While image-level hallucination benefits few-shot recognition by synthesis of novel view [3] or introducing random noises [55], most of the work applies hallucination in the feature space. The work from [28] generates novel class features by transforming shared features in base classes. Authors of [67] build a hallucination framework in the region of interest feature space object to enhance object" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 300, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 373 + ], + "type": "text", + "content": "detection performance. More recent work shows that this hallucination mechanism also benefits 3D human pose estimation by generating novel motion sequences [21]. While hallucination is effective in different learning tasks, to the best of our knowledge, the performance and application of hallucination in SSL are fully unexplored." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 383, + 463, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 463, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 463, + 396 + ], + "type": "text", + "content": "2.3. Feature-Level Augmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 403, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 559 + ], + "type": "text", + "content": "Hallucination relies on effective feature-level augmentation or manipulations. The primary goal of feature augmentation is to extend the limited labeled dataset without relying on expensive computation such as auto-encoder[33], Generative Adversarial Networks[40] or simulation tools[52, 58, 5]. For instance, in the work by [16], a task-agnostic feature augmentation approach is proposed to enrich training data with minimal additional computation. The authors of [36] also explore similar ideas in the domain of few-shot learning. Building upon this, the concept is adapted to sentence representation learning [62, 19] and few-shot learning tasks in remote sensing[56]. More recently, [38] applies feature augmentation based on a meta-learning technique." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 572, + 362, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 362, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 362, + 585 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "In this section, we first introduce the overall process of hallucination for contrastive representation learning in Section 3.1. Secondly, we highlight the center cropping method we used, which is crucial to effective hallucination or generation of new samples in Section 3.2. Then, we introduce the Hallucinator incorporated into our contrastive models in Section 3.3. Finally, we visualize and discuss the critical properties of the hallucination method from two perspectives in Section 3.4: the similarity of positive samples and the uniformity of the feature distribution." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16134" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 167, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 167, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 167, + 85 + ], + "type": "text", + "content": "3.1. The Overall Pipeline" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": "Taking MoCo [11] as an example, we illustrate how a Hallucinator can be plugged into a contrastive learning model in Figure 2. Our architecture takes one image " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " as input. Then, the input " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " is augmented into two views " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": ". Each view will be processed by an encoder consisting of a backbone (e.g., ResNet) and a projector (e.g., an MLP head). After the encoders, output vectors " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " are obtained, forming one positive pair " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "(q, k)" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": ". Then, this positive pair " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "(q, k)" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " is fed to a Hallucinator. Notably, Hallucinator is only added to one branch of the framework to generate an additional positive feature " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": ". Together with feature vector " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " form an extra positive pair " + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "inline_equation", + "content": "(\\hat{q}, k)" + }, + { + "bbox": [ + 46, + 89, + 290, + 319 + ], + "type": "text", + "content": " during the training. Based on different contrastive learning models, the loss functions keep intact, and the average loss of these two positive pairs is computed for back-propagation. The same paradigm could be applied to SimCLR, SiamSiam and other contrastive learning models. We illustrate further details about pipelines and loss functions of other models in this paper in the Supplementary Material (Section 1.1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 324, + 150, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 150, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 150, + 337 + ], + "type": "text", + "content": "3.2. Center Cropping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 343, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 534 + ], + "type": "text", + "content": "In contrastive learning, data augmentations aim to ensure the performance of pre-trained representations invariant to nuisances. Among all these methods, random crop plays the most critical role in all the contrastive learning models. Generally, views (cropped tiles) generated by random cropping are diversified, successfully covering all the semantic information over the whole image. However, such a cropping method is likely to generate false positives patches [44]. In other words, patches randomly cropped from the original images do not necessarily share the overlapped pixels and sufficient common information. Therefore, these false positive pairs may be fooling models during training, causing representations to be sub-optimal. Importantly, the issue will be exacerbated if we generate further hallucinated samples based on false positive pairs, which misleads the overall training beyond the sweet spot." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "text", + "content": "To tackle this issue, we first apply center cropping " + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{C}_{crop}" + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "text", + "content": " to the original image, getting a relatively smaller image " + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\hat{I}_{x,y}" + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "text", + "content": ". Then, random cropping " + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathbb{R}_{crop}" + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\hat{I}_{x,y}" + }, + { + "bbox": [ + 46, + 534, + 287, + 583 + ], + "type": "text", + "content": ". More specifically, the center cropping can be formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 590, + 287, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 590, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 121, + 590, + 287, + 605 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {x, y} = \\mathbb {C} _ {\\text {c r o p}} \\left(I _ {x, y}, p\\right), \\tag {1}", + "image_path": "dbe2c10b5ab3eb3017884d978a4d15e055ce666f1ab1eac53827cad1d2920877.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{x,y}" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": " is the input image with " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": " as the coordinate of the images' center. After center cropping, we keep the center of " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{I}_{x,y}" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": " unchanged. Meanwhile, with the original shape of " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{x,y}" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": " defined as " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": ", we define the shape of cropped image " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{I}_{x,y}" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\hat{h},\\hat{w})" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": " denotes a ratio of cropped length over the original length, i.e., " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "p = \\frac{\\hat{w}}{w} = \\frac{\\hat{h}}{h}" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": ". Only if particularly mentioned, we set " + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 46, + 613, + 288, + 715 + ], + "type": "text", + "content": " for all experiments in this paper." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 76, + 541, + 160 + ], + "blocks": [ + { + "bbox": [ + 315, + 76, + 541, + 160 + ], + "lines": [ + { + "bbox": [ + 315, + 76, + 541, + 160 + ], + "spans": [ + { + "bbox": [ + 315, + 76, + 541, + 160 + ], + "type": "image", + "image_path": "77c72261a5c95c0e15c94fa6898ad99109334763745ac2a5d6fb76af1739a5fb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": "Figure 3: The Hallucinator. Stage 1: The original feature vector " + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": " is extrapolated to the opposite direction of feature vector " + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": ", forming " + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": " in a linear way. Stage 2: We introduce non-linear transformation to smooth extrapolated features concatenated with " + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "inline_equation", + "content": "q'" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": ". Stage 3: output original " + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": " and hallucinated " + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "text", + "content": "While center cropping effectively avoids false positive pairs, it reduces the operable region for random cropping and generates positive views with a similar appearance. We, therefore, adopt center-suppressed sampling [44] with a sampling method following a beta distribution " + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\beta (\\alpha ,\\alpha)" + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "text", + "content": " (i.e., a U-shaped distribution). Concretely, " + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\beta (\\alpha ,\\alpha)" + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "text", + "content": " assigns a lower probability to the center of the " + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\hat{I}_{x,y}" + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "text", + "content": " and gives greater probability to its boundary, increasing the variance between views " + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 304, + 276, + 545, + 396 + ], + "type": "text", + "content": ". Together, we summarize the process to obtain these two views as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 406, + 504, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 406, + 504, + 420 + ], + "spans": [ + { + "bbox": [ + 344, + 406, + 504, + 420 + ], + "type": "interline_equation", + "content": "x _ {1} = T (\\mathbb {R} _ {c r o p} (\\hat {I} _ {x, y} | \\alpha)), \\quad \\mathrm {s . t .} \\quad \\alpha < 1", + "image_path": "7f54e5f28b91ea961dd2a596409895bd372520a2025a9d435e34cfbc6db59a07.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 422, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 422, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 345, + 422, + 545, + 435 + ], + "type": "interline_equation", + "content": "x _ {2} = T \\left(\\mathbb {R} _ {c r o p} \\left(I _ {x, y} | \\alpha\\right)\\right), \\quad \\text {s . t .} \\quad \\alpha < 1, \\tag {2}", + "image_path": "89daf3253905156e423b1eb1744661c7c7b3b4dbc6964cfacaffe2d8cd84d40a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "content": " denotes data augmentations, including color jittering, random grayscale, Gaussian blur and horizontal flipping. " + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mathbb{R}_{crop}(\\dots|\\alpha)" + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "content": " represents random cropping following the " + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\beta (\\alpha ,\\alpha)" + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "content": " distribution. " + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 445, + 545, + 540 + ], + "type": "text", + "content": " is set to less than 1 to ensure an increasing sampling probability as the pixel's coordinates go beyond the center. A visualization of the center sampling method can be found in Supplementary Material (Section 1.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 550, + 389, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 389, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 389, + 562 + ], + "type": "text", + "content": "3.3. Hallucinator" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": "Asymmetric Feature Extrapolation. The first objective of this module is to introduce an additional positive pair without introducing extra computations. Therefore, the feature-level operation is preferred compared to image-level operations. To achieve this, Hallucinator is plugged in after the views " + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": " are encoded. As a result, the hallucinated (generated) feature is purely based on the two feature vectors " + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 569, + 545, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Meanwhile, since harder positives improve pre-trained encoders' generalization capacity [51], the hallucinated features in positive pairs are favorable if they share less mutual information. Previous work illustrates that a symmetric" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16135" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "positive extrapolation is effective in generating hard examples for MoCo [68]. Concretely, two positive features are combined with weighted addition, pushing positive features apart. However, the hallucination process is asymmetric, i.e., Hallucinator is only added in one of the branches of the model. Then, we propose to apply the single-side feature extrapolation to the feature vector " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", as shown in the first stage of Figure 3. Additionally, we simplify the sampling strategy of weights for extrapolation from a beta distribution to a uniform distribution. To be specific, we summarize the positive extrapolation in our method as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 219, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 219, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 78, + 219, + 287, + 232 + ], + "type": "interline_equation", + "content": "q ^ {\\prime} = (1 + \\lambda) q - \\lambda k \\quad \\text {s . t .} \\quad \\lambda \\sim U (\\beta_ {1}, \\beta_ {2}), \\tag {3}", + "image_path": "e61e260529604924115fac4ac057b25bedc6226d4b3cc559f7708ad3df515552.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "text", + "content": " is sampled from a uniform distribution " + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "inline_equation", + "content": "U(\\beta_{1},\\beta_{2})" + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 47, + 247, + 287, + 282 + ], + "type": "text", + "content": ", which are the boundary of the uniform distribution, are set to 0 and 0.1 by default." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 287, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 405 + ], + "type": "text", + "content": "Hallucination. Positive extrapolation is based on mixup [64], i.e., a linear transformation. While positive extrapolation has been proven beneficial to generating hard examples, this linear feature transformation might have a relatively limited capacity to synthesize new feature vectors. This assumption is based on the more satisfactory performance of the non-linear mixup over the original one [26]. Similarly, if non-linearity is introduced to the feature generation, the generated vector will benefit more from the training and boost the performance of downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": "To empower our model with the capacity of non-linear fitting, we introduce the hallucination process. Specifically, we first concatenate " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "q'" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " from the equation 3 and the feature vector " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " together. Then, we use the concatenated feature " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "(q, q')" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " as an input of a non-linear transformation function " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "H_{\\theta}(.)" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " that can be instantiated with " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " linear layers and a ReLU layer between two successive layers. When " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "H_{\\theta}(.)" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " is a non-parametric module, forming an identity function. Such a setting performance is relatively sub-optimal, as shown in the ablation study in Table 6. Empirically, we find that " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "n = \\{2, 3\\}" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " performs well as the hallucinator becomes non-linear and more powerful. We set " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": " by default as its results are slightly better. With the transformation function " + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "inline_equation", + "content": "H_{\\theta}(.)" + }, + { + "bbox": [ + 46, + 407, + 288, + 587 + ], + "type": "text", + "content": ", the hallucinated feature is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 601, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 601, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 135, + 601, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\hat {q} = H _ {\\theta} \\left(q, q ^ {\\prime}\\right), \\tag {4}", + "image_path": "9fc838ffefcb8b3eb01874fcf2153eba968ea663cd7103b4d510e7053cd6092c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " is the parameters of " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "H_{\\theta}(.)" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ". Notably, " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "H_{\\theta}(.)" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " is differentiable, allowing us to back-propagate the loss of contrastive learning. Therefore, we update not just the parameters of the encoders and projectors but also the parameters " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " of the hallucinator. The second stage of Figure 3 illustrates the proposed hallucination process. Following that, we take " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " as the output of Hallucinator." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 535, + 209 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 535, + 209 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 535, + 209 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 535, + 209 + ], + "type": "image", + "image_path": "3160a7a5fa9552b083afff264f3ca41829110ceb0952bbf4cb5427ef0c58835e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 226, + 547, + 287 + ], + "lines": [ + { + "bbox": [ + 305, + 226, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 305, + 226, + 547, + 287 + ], + "type": "text", + "content": "Figure 4: Similarity of positive pairs in training. Smaller values indicate less mutual information and better representation [51, 44, 68]. As Hallucinator incorporates nonlinearity extrapolation, it guarantees smooth training and harder positive features." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 307, + 463, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 463, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 463, + 319 + ], + "type": "text", + "content": "3.4. Discussion and Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 325, + 546, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 546, + 373 + ], + "type": "text", + "content": "To better understand the behavior of Hallucinator, we discuss two critical properties that may contribute to and explain its effectiveness. For visualization, we train MoCoV2 [11] with a standard ResNet-18 [31] on CIFar-10 [35]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "content": "Similarity of Positive Pair. We first investigate the similarity of positive pairs during the training process. Concretely, we quantize mutual information using the cosine similarity of positive pairs " + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 395, + 429, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 429, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 395, + 429, + 545, + 455 + ], + "type": "interline_equation", + "content": "S \\triangleq \\frac {q \\cdot k}{\\| q \\| \\| k \\|.} \\tag {5}", + "image_path": "5d451031766504515875bec45bcfb6df302ad89a6022ca6e1e25d4a913fc01da.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 462, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 605 + ], + "type": "text", + "content": "While hard positives share less mutual information, thus having a smaller cosine similarity value, the performance of downstream tasks will be enhanced [68]. Based on Figure 4, the proposed Hallucinator generates harder positives with smaller values of similarity. Consequently, it helps contrastive learning models obtain more nuisances-invariant features. More importantly, different from linear symmetric extrapolation, our training curve is relatively stable without many oscillations introduced. This observation indicates that Hallucinator is successfully optimized with the overall framework. Meanwhile, hallucinated features nicely fit into the contrastive learning task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Uniformity. We continue to analyze the performance of the proposed model from the perspective of uniformity. Notably, feature vectors from contrastive learning should be roughly uniformly distributed on a unit hyper-sphere, which ensures maximal information is preserved in the feature space [53]. In other words, the closer the feature distribution is to the uniform distribution, the more the feature benefits downstream tasks. To quantize uniformity, we follow previous work [53] to compute the average value of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16136" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 94, + 208, + 236 + ], + "blocks": [ + { + "bbox": [ + 92, + 94, + 208, + 236 + ], + "lines": [ + { + "bbox": [ + 92, + 94, + 208, + 236 + ], + "spans": [ + { + "bbox": [ + 92, + 94, + 208, + 236 + ], + "type": "image", + "image_path": "3a3759c2fe91044bb345a222683d2bc926222454b1b4b918efcf25e5c0601e49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "type": "text", + "content": "Figure 5: A measure of uniformity based on " + }, + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "type": "text", + "content": " potential. We plot 10,000 feature vectors with Gaussian kernel density estimation (KDE) in " + }, + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "type": "text", + "content": ". The left subplot illustrates the feature vectors from a uniform distribution. The three feature distributions on the right in the first row visualize the features from MoCoV2 [11]. The other three feature distributions in the second row demonstrate MoCoV2 with Hallucinator. Hallucinator benefits the uniformity with a smaller value of " + }, + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 45, + 274, + 547, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 69, + 331, + 158 + ], + "blocks": [ + { + "bbox": [ + 231, + 69, + 331, + 158 + ], + "lines": [ + { + "bbox": [ + 231, + 69, + 331, + 158 + ], + "spans": [ + { + "bbox": [ + 231, + 69, + 331, + 158 + ], + "type": "image", + "image_path": "74ea5fecbf25104e2ec7e48a9b3f3757dc6ffe3d6903d8a349432ca0a5587ac3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 342, + 70, + 417, + 152 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 417, + 152 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 417, + 152 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 417, + 152 + ], + "type": "image", + "image_path": "e9410cde9f6cf7d808cdface506514c36df32234665e9edd1aadf2a6297d64ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 504, + 152 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 504, + 152 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 504, + 152 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 504, + 152 + ], + "type": "image", + "image_path": "2cee15a31f715bd985f6ad1cefdd5490a90f1040d2aafd86ef40701744362c98.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 231, + 159, + 331, + 255 + ], + "blocks": [ + { + "bbox": [ + 231, + 159, + 331, + 255 + ], + "lines": [ + { + "bbox": [ + 231, + 159, + 331, + 255 + ], + "spans": [ + { + "bbox": [ + 231, + 159, + 331, + 255 + ], + "type": "image", + "image_path": "18f9d11ba3cbfd9bbc2f37fa8aff1b56a6f015e1b24c7333ada7e2168f6bd6c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 159, + 417, + 247 + ], + "blocks": [ + { + "bbox": [ + 340, + 159, + 417, + 247 + ], + "lines": [ + { + "bbox": [ + 340, + 159, + 417, + 247 + ], + "spans": [ + { + "bbox": [ + 340, + 159, + 417, + 247 + ], + "type": "image", + "image_path": "84d42f0c92e6afdf63cca55f6f5c4f42ba1c6af7c979057308b443c3fbed06a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 427, + 159, + 504, + 247 + ], + "blocks": [ + { + "bbox": [ + 427, + 159, + 504, + 247 + ], + "lines": [ + { + "bbox": [ + 427, + 159, + 504, + 247 + ], + "spans": [ + { + "bbox": [ + 427, + 159, + 504, + 247 + ], + "type": "image", + "image_path": "a17d35b617352de88304204ae4f00aa78f16fa2580fa4d60b30bfd254f2eac18.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 120, + 333, + 474, + 425 + ], + "blocks": [ + { + "bbox": [ + 120, + 333, + 474, + 425 + ], + "lines": [ + { + "bbox": [ + 120, + 333, + 474, + 425 + ], + "spans": [ + { + "bbox": [ + 120, + 333, + 474, + 425 + ], + "type": "table", + "html": "
DatasetCIFAR-10CIFAR-100Tiny ImageNetSTL-10
HallucinatorX✓X✓X✓X✓
MoCoV188.3188.9460.9461.8144.6545.5388.1990.09
MoCoV287.2189.2359.7061.2647.1247.9589.3290.46
SimCLR89.6690.1160.9461.4345.2246.3089.0789.98
SimSiam90.4790.7863.3964.3843.6644.9687.7988.16
", + "image_path": "643054f6d1610b96b9e17cf33e8eb8a51cb769f2722fe8b58f4dfcf458187019.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 429, + 546, + 453 + ], + "lines": [ + { + "bbox": [ + 45, + 429, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 45, + 429, + 546, + 453 + ], + "type": "text", + "content": "Table 1: Linear classification results for different contrastive methods and datasets in small scales. We adopt ResNet-18 as the backbone and report the classification results with or without Hallucinator." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 474, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 498 + ], + "type": "text", + "content": "Gaussian potential kernel (i.e., Radial Basis Function kernel) of positive features:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 93, + 513, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 513, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 93, + 513, + 287, + 529 + ], + "type": "interline_equation", + "content": "G _ {t} (q, k) \\triangleq e ^ {- t \\| q - k \\| _ {2} ^ {2}} \\quad \\text {s . t .} \\quad t > 0, \\tag {6}", + "image_path": "e419821934e5017b88bf4e9fa16831ede4881bd8a96f8b7e9451fbff83b6371e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " is a fixed parameter set as 2 for all the experiments. We visualize the feature vectors by mapping them to two-dimension feature space and applying " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " normalization in Figure 5. " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "G_{2} = 0.2070" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " for uniformly distributed samples, whereas features from MoCoV2 have " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "G_{2} = 0.2089" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": ". If we plugin Hallucinator into MoCo, Hallucinator provides further contrast during the training with extra positives introduced, giving more uniformly distributed features and a decreased " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " value,i.e., 0.2081. Additionally, we visualize the features of two classes of Cifar-10. Each of these features is well-clustered. With better uniformity, clusters' overlapping decreases, forming more linearly separable features. Therefore, we could observe that the overlapping of feature clusters between class 1 and class 6 in MoCo is larger than the one with Hallucinator plugged in." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 473, + 450, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 450, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 450, + 487 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 494, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 649 + ], + "type": "text", + "content": "In this section, we conduct various experiments on different datasets and contrastive learning models to demonstrate the effectiveness of the Hallucinator. Firstly, we introduce the datasets and details in experiments in Section 4.1. Then, we continue to evaluate the performance of the proposed method using linear probing protocol following the paradigm of previous work [10, 11, 12, 23] in Section 4.2. Following this, we conduct ablation studies to understand how each module contributes to the final results in Section 4.3. Lastly, we show the transferability of pre-trained encoders in downstream tasks requiring dense pixel predictions, including object detection and semantic segmentation in Section 4.4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 658, + 468, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 468, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 468, + 672 + ], + "type": "text", + "content": "4.1. Datasets and Training Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": "Datasets and Baseline Models. We first evaluate the performance of Hallucinator based on a wide range of datasets crossing different scales. Specifically, these datasets in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16137" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 279, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 279, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 279, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 279, + 190 + ], + "type": "table", + "html": "
MethodBackboneEpochIN-200IN-1K
MoCoV1ResNet-5010062.1957.27
MoCoV1(Ours)ResNet-5010063.4659.17
MoCoV2ResNet-5010062.5764.41
MoCoV2(Ours)ResNet-5010063.5864.97
SimCLRResNet-5010062.2261.23
SimCLR(Ours)ResNet-5010063.0261.71
SimSiamResNet-5010062.8063.11
SimSiam(Ours)ResNet-5010063.5263.55
", + "image_path": "2dc1dcccaecf59e5c755c2e9b863f6751933757976c2bdce74486f3f7b186089.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 243 + ], + "type": "text", + "content": "Table 2: Linear classification results on IN-200 and IN-1K. Our method plugged in the proposed Hallucinator in baseline methods. All the models are pre-trained for 100 epochs and use identical training settings for fair comparisons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 286, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 286, + 326 + ], + "type": "text", + "content": "clude CIFAR-10, CIFAR-R100 [35], Tiny ImageNet, STL-10 [14] and ImageNet [15]. Meanwhile, we demonstrate the performance of Hallucinator in several popular contrastive learning frameworks, including MoCoV1 [29], MoCoV2 [11], SimCLR [10] and SimSiam [12]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 286, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 286, + 447 + ], + "type": "text", + "content": "Training and Evaluation Details. Importantly, we strictly adopt the same training settings when evaluating the Hallucinator's performance. While additional gain could be achieved with further exploration of hyper-parameter, it is not the focus of this work. The ultimate goal of the proposed Hallucinator is to provide further feature-level contrast for self-supervised learning. With the Hallucinator plugin, it benefits diversified self-supervised learning methods regardless of their type of backbones and hyper-parameters of training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 286, + 579 + ], + "type": "text", + "content": "Datasets on relatively small scales include CIFAR-10&100, Tiny ImageNet and STL10. To ensure fair comparisons, we keep training settings the same over all these datasets. We follow the paradigm of previous work for pretraining [44]. Concretely, we pre-train contrastive learning models for 500 epochs with a batch size of 512 and ResNet-18 as the backbone. To optimize the models, we use an SGD optimizer and a cosine-annealed learning rate of 0.5. In the linear classification task, we train models for 100 epochs. With the initial learning set as 10, we divide it by 10 at the 60th and 80th epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 580, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 286, + 700 + ], + "type": "text", + "content": "As the scale of the dataset expands to ImageNet, we use ResNet-50 as the backbone. For MoCoV1, MoCoV2 and SimSiam, all the pre-training settings follow the original work. We use a batch size of 512 and an SGD optimizer with a cosine-annealed learning rate of 0.5 for SimCLR, the same as the settings in [44]. For the linear classification task, we follow the original work in [29], training the model for 100 epochs. We set the initial learning rate to 30. Then, we divide the learning by 10 at the 60th and 80th with a weight decay of 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": " for center cropping and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 535, + 165 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 535, + 165 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 535, + 165 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 535, + 165 + ], + "type": "table", + "html": "
Acc.(%)Center CroppingAsymmetric ExtrapolationHallucination
62.57✗✗✗
63.58(+1.01)✓✓✓
62.58(+0.01)✓
62.82(+0.25)✓
62.84(+0.27)✓
63.07(+0.50)✓✓
", + "image_path": "9e0cfe431dc18fd45745253ab32f43725b02e042aee3e86edc63598b82b1f4a7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 170, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 306, + 170, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 545, + 193 + ], + "type": "text", + "content": "Table 3: Ablation of the different modules in the proposed method." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 206, + 537, + 239 + ], + "blocks": [ + { + "bbox": [ + 314, + 206, + 537, + 239 + ], + "lines": [ + { + "bbox": [ + 314, + 206, + 537, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 206, + 537, + 239 + ], + "type": "table", + "html": "
p (%)908070605040
Acc.(%)63.1163.1963.3363.5663.5863.21
", + "image_path": "14f441da8207cf00f1c6ef73301ac3e95a70ace84fffeed28b006c9ec36308f9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 307, + 243, + 542, + 255 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 542, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 542, + 255 + ], + "type": "text", + "content": "Table 4: Ablation of classification results w.r.t the " + }, + { + "bbox": [ + 307, + 243, + 542, + 255 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 307, + 243, + 542, + 255 + ], + "type": "text", + "content": " value." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 312, + 267, + 539, + 297 + ], + "blocks": [ + { + "bbox": [ + 312, + 267, + 539, + 297 + ], + "lines": [ + { + "bbox": [ + 312, + 267, + 539, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 267, + 539, + 297 + ], + "type": "table", + "html": "
(β1,β2)(-0.5, 0.0)(0.0, 0.5)(0.5, 1.0)(1.0, 1.5)(0.0, 1.0)
Acc.(%)62.8063.3263.5563.4263.58
", + "image_path": "da6e09e8428b323754fb07d6caec2835aed01cad53378607cbef215414186a7c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 300, + 544, + 324 + ], + "lines": [ + { + "bbox": [ + 306, + 300, + 544, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 544, + 324 + ], + "type": "text", + "content": "Table 5: Ablation of classification results w.r.t the extrapolation range." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 321, + 337, + 531, + 372 + ], + "blocks": [ + { + "bbox": [ + 321, + 337, + 531, + 372 + ], + "lines": [ + { + "bbox": [ + 321, + 337, + 531, + 372 + ], + "spans": [ + { + "bbox": [ + 321, + 337, + 531, + 372 + ], + "type": "table", + "html": "
layer #: n01234
Acc.(%)62.8563.0363.5763.5863.55
", + "image_path": "8bf1269637e88af7470ac74f1226747e2d39aa9ac0d0c983414d79027e78640a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 376, + 544, + 400 + ], + "lines": [ + { + "bbox": [ + 306, + 376, + 544, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 544, + 400 + ], + "type": "text", + "content": "Table 6: Ablation of classification results w.r.t the number of linear layers " + }, + { + "bbox": [ + 306, + 376, + 544, + 400 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 306, + 376, + 544, + 400 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "text", + "content": "center-suppressed sampling, respectively. For asymmetric linear extrapolation, we set " + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0" + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.1" + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "text", + "content": ". In the Hallucination process, " + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 304, + 422, + 545, + 481 + ], + "type": "text", + "content": " for all the results reported in this paper. All the experiments are conducted on a server with 8 GPUs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 491, + 468, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 468, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 468, + 502 + ], + "type": "text", + "content": "4.2. Linear Classification Protocol" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "content": "Following the previous protocol, we first evaluate the proposed method by linear classification of frozen features. For each dataset, we report the top-1 classification accuracy on the validation set." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "text", + "content": "Results on Small-Scale Datasets. The classification results on Cifar10&100, Tiny ImageNet and STL-10 are reported in Table 1. With Hallucinator introduced, we notice a stable improvement over the baselines ranging from " + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "inline_equation", + "content": "0.31\\%" + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "inline_equation", + "content": "1.98\\%" + }, + { + "bbox": [ + 304, + 557, + 545, + 629 + ], + "type": "text", + "content": ". Notably, such improvements do not introduce extra computations and generalize well to various models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Results on ImageNet. For the results of ImageNet, we report the results at two different scales. First, we evaluate its performance on standard IN-1K (i.e. ImageNet-1K), which consists of 1000 classes. Second, we test the proposed method in IN-200 (i.e. ImageNet-200) with 200 randomly selected classes. We report the corresponding results in Table 2. We found that Hallucinator essentially benefits" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16138" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 212 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 212 + ], + "type": "table", + "html": "
Method1N-1k Top-1VOC detectionCOCO detectionCOCO instance segmentation
APAP50AP75APbbAPbb50APbb75APmkAPmk50APmk75
Random init-33.860.257.2726.444.027.829.346.930.8
Supervised76.153.581.358.838.258.241.233.354.735.2
InfoMin [51]70.157.682.764.639.058.542.034.155.236.3
MoCoV1 [29]60.655.981.562.638.558.341.633.654.835.6
MoCoV1(Ours)63.856.781.863.238.958.541.933.855.236.0
MoCoV2 [11]67.557.082.463.639.058.641.934.255.436.2
MoCoV2(Ours)68.057.482.763.939.358.842.334.655.536.4
", + "image_path": "dd4ed3060a86516d27e93a2751f1ce0a394cdec320cd7938846ffd4e4e6becb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 216, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 216, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 545, + 240 + ], + "type": "text", + "content": "Table 7: Fine-tuning results on object detection tasks on PASCAL VOC and COCO, and instance segmentation on COCO. All models are pre-trained for 200 epochs on ImageNet-1K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "text", + "content": "MoCoV1 with " + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "inline_equation", + "content": "1.27\\%" + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "inline_equation", + "content": "1.89\\%" + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "text", + "content": " improvements for IN-200 and IN-1K accordingly. For MoCoV2, SimSiam and SimCLR, we notice a gain in accuracy ranging from " + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "inline_equation", + "content": "0.44\\%" + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 47, + 261, + 287, + 309 + ], + "type": "text", + "content": ". On average, the gains are more salient in IN-200." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 319, + 147, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 147, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 147, + 331 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 422 + ], + "type": "text", + "content": "In this section of ablation studies, we investigate essential modules in our proposed methods, including center cropping, asymmetric extrapolation and the process of hallucination. We conduct experiments using MoCoV2 with ResNet-50 as the backbone. We report the results of the classification task in IN-200 for all the following experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 424, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 288, + 592 + ], + "type": "text", + "content": "Contributions of Modules. We first report the results with or without crucial modules introduced in Section 3. According to Table 3, center cropping shows a similar performance to the original cropping method, successfully covering major semantic information of images. However, it successfully avoids false positives in pre-training, which is critical for hallucination. Asymmetric extrapolation benefits the performance of representation learning with reduced mutual information. This observation is consistent with the results shown in the symmetric extrapolation [68]. If we combine asymmetric extrapolation and the hallucination method, the performance of the model could be further boosted. However, it is still sub-optimal because of possible false positives in cropping." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": "Center Cropping. In this work, we apply center cropping as a critical tool in one branch of contrastive learning to avoid false positives. Therefore, we investigate how the proposed model is influenced by the length ratio " + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": " in Table 4. The classification results are favorable when " + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": " is in the range (0.5, 0.6). We notice a performance drop when " + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 593, + 287, + 689 + ], + "type": "text", + "content": " is less than 0.5. This drop is because the center cropping failed to cover all semantic information in the images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Extrapolation Range. We continue to report the extrapolation range given its differences in sampling method and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "content": "single-side architecture compared with the previous method [68]. Based on Table 5, the classification accuracy drops if we set " + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\beta_{1} = -0.5" + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0" + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "content": ", respectively. Such observation is because hallucinated positives share more mutual information, thus less beneficial to training. Generally, extrapolation boosts accuracy by outperforming the baseline by at least " + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "0.75\\%" + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "content": ". However, the improvement decreases when " + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\beta_{1} > 1.0" + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "content": ". We eventually set " + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.0" + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 1.0" + }, + { + "bbox": [ + 304, + 261, + 545, + 369 + ], + "type": "text", + "content": ", given its best performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "content": "Hallucination. We investigate how the number of linear layers influences the model's performance. When " + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "content": ", the hallucination process is an identity transformation, introducing no additional operation except for extrapolation. When " + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "content": ", non-linearity is introduced. Then we notice a salient improvement over the baseline. We set " + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 304, + 370, + 545, + 454 + ], + "type": "text", + "content": " by default since it gives better performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 464, + 432, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 432, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 432, + 477 + ], + "type": "text", + "content": "4.4. Transferring Features" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "type": "text", + "content": "The primary goal of representation is to learn transferrable features. We evaluate the transferability of features from the proposed method following the previous protocol [29, 11, 10, 12]. Then, we compare the representation quality by transferring them to downstream tasks, including VOC [18] object detection and COCO [39] object detection and instance segmentation. Notably, we re-implement all these experiments using the same settings in MoCo's detector2 codebase [59]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 593, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 688 + ], + "type": "text", + "content": "Object Detection on PASCAL VOC. Following the paradigm of previous work [29], we use Faster R-CNN[20] as the object detection method using R50-C4 as the detector [30]. We train the model end-to-end on the trainval2007+2012 and evaluate its performance on test2007. As shown in Table 7, we observe a stable gain range from 0.3 to 0.8 under different metrics on MoCoV1 and MoCoV2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Object Detection and instance segmentation on COCO. We continue to report the detection and segmentation re" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16139" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "sults on COCO using Mask R-CNN [30]. Similarly, we use the R50-C4 as the backbone. The is model in an end-to-end way on train2017. Then, the model is evaluated on val2017. Again, the proposed method benefits the object detection and segmentation tasks with various metrics as demonstrated in Table 7." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 155, + 119, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 119, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 119, + 167 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 289, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 343 + ], + "type": "text", + "content": "In this work, we propose Hallucinator, which generates additional hard positive pairs for contrastive learning models based on Siamese structure. Hallucinator generates novel data samples in the feature space to provide the training with further contrast without additional computation. We design an asymmetric feature extrapolation to avoid trivial positive pairs and innovatively introduce non-linear hallucination to smooth the generated samples. We empirically prove the effectiveness and generalization capacity of Hallucinator to well-recognized contrastive learning models, including MoCoV1&V2, SimCLR and SimSiam. Finally, we hope this work could bring the concept of \"Hallucination\" into the SSL domain and unlock future research on sample generations&synthesis in contrastive learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 307, + 72, + 365, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 365, + 84 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 365, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 91, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 311, + 91, + 545, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 545, + 134 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 311, + 137, + 547, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 137, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 311, + 137, + 547, + 180 + ], + "type": "text", + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 311, + 182, + 547, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 182, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 311, + 182, + 547, + 225 + ], + "type": "text", + "content": "[3] Zhipeng Bao, Yu-Xiong Wang, and Martial Hebert. Bowtie networks: Generative modeling for joint few-shot recognition and novel-view synthesis. arXiv preprint arXiv:2008.06981, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 227, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 227, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 311, + 227, + 545, + 259 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 311, + 261, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 311, + 261, + 545, + 294 + ], + "type": "text", + "content": "[5] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 296, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 296, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 311, + 296, + 545, + 350 + ], + "type": "text", + "content": "[6] Qingxing Cao, Liang Lin, Yukai Shi, Xiaodan Liang, and Guanbin Li. Attention-aware face hallucination via deep reinforcement learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 690-698, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 352, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 352, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 311, + 352, + 545, + 407 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 408, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 408, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 311, + 408, + 545, + 463 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 464, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 464, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 311, + 464, + 545, + 509 + ], + "type": "text", + "content": "[9] Pengguang Chen, Shu Liu, and Jiaya Jia. Jigsaw clustering for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11526-11535, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 555, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 588 + ], + "type": "text", + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "text", + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[13] Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, and Stefanie Jegelka. Debiased contrastive learning. Advances in neural information processing systems, 33:8765-8775, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[14] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16140" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "text", + "content": "[16] Terrance DeVries and Graham W Taylor. Dataset augmentation in feature space. arXiv preprint arXiv:1702.05538, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 230 + ], + "type": "text", + "content": "[17] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. With a little help from my friends: Nearest-neighbor contrastive learning of visual representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9588–9597, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 274 + ], + "type": "text", + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 309 + ], + "type": "text", + "content": "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 311, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 311, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 287, + 342 + ], + "type": "text", + "content": "[20] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 344, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 410 + ], + "type": "text", + "content": "[21] Kehong Gong, Bingbing Li, Jianfeng Zhang, Tao Wang, Jing Huang, Michael Bi Mi, Jiashi Feng, and Xinchao Wang. Posetriplet: co-evolving 3d human pose estimation, imitation, and hallucination under self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11017-11027, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "text", + "content": "[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "text", + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[24] Matthew D Grilli and Elizabeth L Glisky. The self-imagination effect: Benefits of a self-referential encoding strategy on cued recall in memory-impaired individuals with neurological damage. Journal of the International Neuropsychological Society, 17(5):929-933, 2011." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "text", + "content": "[25] Liangke Gui, Adrien Bardes, Ruslan Salakhutdinov, Alexander Hauptmann, Martial Hebert, and Yu-Xiong Wang. Learning to hallucinate examples from extrinsic and intrinsic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8701-8711, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[26] Hongyu Guo. Nonlinear mixup: Out-of-manifold data augmentation for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4044-4051, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[27] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pat" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "tern Recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[28] Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In Proceedings of the IEEE international conference on computer vision, pages 3018-3027, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "type": "text", + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "text", + "content": "[30] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 273, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 545, + 304 + ], + "type": "text", + "content": "[32] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International conference on machine learning, pages 4182-4192. PMLR, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "type": "text", + "content": "[33] Wei-Ning Hsu, Yu Zhang, and James Glass. Unsupervised domain adaptation for robust speech recognition via variational autoencoder-based data augmentation. In 2017 IEEE automatic speech recognition and understanding workshop (ASRU), pages 16-23. IEEE, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "text", + "content": "[34] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 426 + ], + "type": "text", + "content": "[35] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.., 2009." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 545, + 470 + ], + "type": "text", + "content": "[36] Varun Kumar, Hadrien Glaude, Cyprien de Lichy, and William Campbell. A closer look at feature space data augmentation for few-shot intent classification. arXiv preprint arXiv:1910.04176, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "text", + "content": "[37] Kibok Lee, Yian Zhu, Kihyuk Sohn, Chun-Liang Li, Jinwoo Shin, and Honglak Lee. i-mix: A domain-agnostic strategy for contrastive representation learning. arXiv preprint arXiv:2010.08887, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 515, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 559 + ], + "type": "text", + "content": "[38] Jiangmeng Li, Wenwen Qiang, Changwen Zheng, Bing Su, and Hui Xiong. Metaug: Contrastive learning via meta feature augmentation. In International Conference on Machine Learning, pages 12964-12978. PMLR, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "text", + "content": "[40] Giovanni Mariani, Florian Scheidegger, Roxana Istrate, Costas Bekas, and Cristiano Malossi. Bagan: Data augmentation with balancing gan. arXiv preprint arXiv:1803.09655, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16141" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[42] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[44] Xiangyu Peng, Kai Wang, Zheng Zhu, Mang Wang, and Yang You. Crafting better contrastive views for siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16031-16040, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "text", + "content": "[45] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. arXiv preprint arXiv:2010.04592, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 297 + ], + "type": "text", + "content": "[46] Eli Schwartz, Leonid Karlinsky, Joseph Shtok, Sivan Harary, Mattias Marder, Abhishek Kumar, Rogerio Feris, Raja Giryes, and Alex Bronstein. Delta-encoder: an effective sample synthesis method for few-shot object recognition. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 363 + ], + "type": "text", + "content": "[47] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[48] Anshul Shah, Suvrit Sra, Rama Chellappa, and Anoop Cherian. Max-margin contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 8220-8230, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "text", + "content": "[49] Zhiqiang Shen, Zechun Liu, Zhuang Liu, Marios Savvides, Trevor Darrell, and Eric Xing. Un-mix: Rethinking image mixtures for unsupervised visual representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2216-2224, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "type": "text", + "content": "[50] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 776-794. Springer, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[51] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "text", + "content": "[52] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026-5033. IEEE, 2012." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[54] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[55] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[56] Jing Wu, Naira Hovakimyan, and Jennifer Hobbs. Genco: An auxiliary generator from contrastive learning for enhanced few-shot learning in remote sensing. arXiv preprint arXiv:2307.14612, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[57] Jing Wu, David Pichler, Daniel Marley, David Wilson, Naira Hovakimyan, and Jennifer Hobbs. Extended agriculture-vision: An extension of a large aerial image dataset for agricultural pattern analysis. arXiv preprint arXiv:2303.02460, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[58] Jing Wu, Ran Tao, Pan Zhao, Nicolas F Martin, and Naira Hovakimyan. Optimizing nitrogen management with deep reinforcement learning and crop simulations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1720, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "text", + "content": "[59] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 304, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 358 + ], + "type": "text", + "content": "[60] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "text", + "content": "[61] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[62] Yuanmeng Yan, Rumei Li, Sirui Wang, Fuzheng Zhang, Wei Wu, and Weiran Xu. Consert: A contrastive framework for self-supervised sentence representation transfer. arXiv preprint arXiv:2105.11741, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "text", + "content": "[63] Mang Ye, Xu Zhang, Pong C Yuen, and Shih-Fu Chang. Unsupervised embedding learning via invariant and spreading instance feature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6210-6219, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "text", + "content": "[64] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 591 + ], + "type": "text", + "content": "[65] Hongguang Zhang, Jing Zhang, and Piotr Koniusz. Few-shot learning via saliency-guided hallucination of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2770-2779, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[66] Ruixiang Zhang, Tong Che, Zoubin Ghahramani, Yoshua Bengio, and Yangqiu Song. Metagan: An adversarial approach to few-shot learning. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[67] Weilin Zhang and Yu-Xiong Wang. Hallucination improves few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13008-13017, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[68] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16142" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [], + "discarded_blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 96 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 96 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 10306-10315, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16143" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_content_list.json b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f915d0e848d16b40c5475569231136b4b79eb502 --- /dev/null +++ b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_content_list.json @@ -0,0 +1,1440 @@ +[ + { + "type": "text", + "text": "HandR $^2$ N $^2$ : Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wencan Cheng \nDepartment of Artificial Intelligence \nSungkyunkwan University \ncwc1260@skku.edu", + "bbox": [ + 99, + 204, + 390, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jong Hwan Ko \nCollege of Information and Communication Engineering \nSungkyunkwan University \njhko@skku.edu", + "bbox": [ + 423, + 204, + 874, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D hand pose estimation is a critical task in various human-computer interaction applications. Numerous deep learning based estimation models in this domain have been actively explored. However, the existing models follow a non-recurrent scheme and thus require complex architectures or redundant parameters in order to achieve acceptable model capacity. To tackle this limitation, this paper proposes HandR²N², a compact neural network that iteratively regresses the hand pose using a novel residual recurrent unit. The recurrent design allows recursive exploitation of partial layers to gradually optimize previously estimated joint locations. In addition, we exploit graph reasoning to capture kinematic dependencies between joints for better performance. Experimental results show that the proposed model significantly outperforms the existing methods on three hand pose benchmark datasets in terms of both accuracy and efficiency. Codes and pre-trained models are publicly available at https://github.com/cwc1260/HandR2N2.", + "bbox": [ + 75, + 340, + 473, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 660, + 209, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D hand pose estimation, which estimates the 3D positions of hand keypoints, provides a fundamental understanding of human hand activity. Thus, it is a critical task for many human-computer interaction applications such as robotics and augmented/virtual reality. With the assistance of deep learning techniques and low-cost depth cameras, numerous studies have contributed to the substantial progress in the field of hand pose estimation. However, accurate and efficient hand pose estimation is still challenging because of the various probability of hand orientations, severe self-occlusion, and noisy depth images [12, 9, 18, 6].", + "bbox": [ + 75, + 686, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of the recent state-of-the-art approaches are based on deep learning, especially deep convolutional neural networks (CNNs). Such CNN-based approaches [31, 10, 13,", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/966f3a8f0700242debbdc03197897f3888e741ecfe95d51ea22cbc21e951565e.jpg", + "image_caption": [ + "Figure 1. Illustration of the recurrent concept. For each iteration, our proposed recurrent architecture samples local points around joints from previous iteration to generate displacements to refine the joint locations." + ], + "image_footnote": [], + "bbox": [ + 545, + 306, + 864, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "25, 1, 8, 23, 6, 23] generally accept 2D depth images as an input, which can be directly processed by the 2D convolutional layers. However, it is difficult for a 2D CNN to capture the highly non-linear mapping from a 2D depth image to the 3D hand pose, significantly restricting its efficiency and accuracy. Therefore, a line of works [11, 18] discretized 2D input images into 3D volumetric representations and subsequently applied a 3D CNN for direct 3D-to-3D inference. However, a critical drawback of a 3D CNN is its computational overhead that increases cubically with an increase of the input resolution [4]. In contrast, another series of hand pose estimation models [9, 12, 2, 15, 4] have presented remarkable performances by using PointNet [20, 21], which allows the models to accept a set of continuous 3D coordinates converted from an input hand image without discretization.", + "bbox": [ + 496, + 535, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the accuracy improvement achieved by the aforementioned approaches, they still suffer from large and complex model architectures that require substantial computational overheads. Thus, these methods are not suitable for resource-limited devices such as wearable or hand-held devices. Furthermore, all the existing approaches follow a non-recurrent architecture. As a result, they are not flexible for varying resource constraints and accuracy targets.", + "bbox": [ + 496, + 779, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "20904", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we tackle these limitations using a recurrent architecture whose capacity can be dynamically adjusted to achieve either higher accuracy or higher efficiency by iteratively using a part of model parameters.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recurrent structure can achieve the high capacity of deeper networks by the recursive use of fewer shared parameters [34]. Various image processing models that used a recurrent neural network (RNN) as an explicit iterative approach have already yielded encouraging results [30, 5, 35, 26]. To fully take advantage of the RNN in a hand pose estimation task, we propose a novel residual recurrent neural network, $\\mathrm{HandR}^2\\mathrm{N}^2$ , that iteratively regresses accurate 3D hand pose from an input hand point cloud. Intuitively, $\\mathrm{HandR}^2\\mathrm{N}^2$ can be served as a general optimizer that recursively searches the optimal joint locations using the observation captured by the previous estimation, as shown in Figure 1.", + "bbox": [ + 75, + 154, + 468, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To implement $\\mathrm{HandR}^2\\mathrm{N}^2$ , we introduce two novel components: 1) a residual recurrent unit (RRU) and 2) an initialization module. The RRU is a key component that iteratively proceeds for updating its joint-wise hidden states and coordinate estimations. The initial joint-wise states and coordinates of the RRU are globally estimated by the initialization module.", + "bbox": [ + 75, + 351, + 468, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The RRU specifically integrates with an attentive gate that can explore optimal points from a group of candidate points for the improved accuracy. Since the proposed RRU only computes the features contributed to the specific joints, it is more efficient than the similar prior model in [5], which re-computes the entire point features. Moreover, RRU exploits graph reasoning, which is based on a novel channelwise graph convolutional network (GCN). GCN is capable to model the graph-structured data, making it particularly suitable for the kinematically structured hand data [8]. The use of the proposed channel-wise GCN enables the RRU to recursively enhance the hidden states by capturing the strong kinematic dependencies between joints. More importantly, the number of RRU iterations during inference can be dynamically scaled to be different from the training iteration. Specifically, the RRU inference iterations can be configured large to enhance the model capacity and accuracy, or small to reduce the computation requirement. Therefore, $\\mathrm{HandR}^2\\mathrm{N}^2$ is capable of adapting varying complexity and accuracy requirements as shown in Figure 2.", + "bbox": [ + 75, + 460, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate $\\mathrm{HandR}^2\\mathrm{N}^2$ on three challenging benchmarks, ICVL [28], MSRA [27], and NYU [31] datasets. The results show that our network achieves a new state-of-the-art record. The proposed network reports the minimum mean distance errors of $5.70\\mathrm{mm}$ , $6.42\\mathrm{mm}$ and $7.27\\mathrm{mm}$ on the ICVL, MSRA and NYU datasets, respectively. Meanwhile, it requires only $1.3\\mathrm{M}$ parameters and runs with adjustable complexity $0.72 + 0.34\\times I$ GFLOPs, where $I$ indicates the number of iterations. Thus, our model requires", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3650eb4fdde82c3e05b5a99971650dd62f4a037445756c6e7b9510df436fc33d.jpg", + "image_caption": [ + "Figure 2. Comparison between operation count and estimation error. $I$ is the iteration number of $\\mathrm{HandR}^2\\mathrm{N}^2$ during inference." + ], + "image_footnote": [], + "bbox": [ + 532, + 89, + 856, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "less computation for the highest accuracy compared with other state-of-the-art methods, as shown in Figure 2.", + "bbox": [ + 498, + 263, + 890, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key contributions of this paper are summarised as follows:", + "bbox": [ + 500, + 294, + 890, + 323 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel iterative neural network architecture that takes the hand point cloud as input and progressively estimates the accurate 3D hand joint coordinates.", + "- We propose a novel residual recurrent unit that captures new local observations around previously generated coordinates to refine joint-wise hidden states and coordinate estimation. It also leverages a graph neural network to model dependency between the joints for performance improvement.", + "- We conduct extensive experiments to analyze the efficiency and effectiveness of our proposed network." + ], + "bbox": [ + 517, + 334, + 890, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 550, + 640, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Depth-based 3D Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 575, + 833, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D hand pose estimation approaches using depth images can be classified into three categories according to the input types: conventional 2DCNN-based methods that take raw 2D depth images as input [31, 10, 13, 25, 1, 8, 23, 6], 3DCNN-based methods that use 3D voxelized representations of depth images [11, 18], and PointNet-based methods that process the point cloud transformed from depth images [9, 12, 2, 15, 4]. In this section, we focus on PointNet-based methods, since the point cloud is an accurate and concise representation for efficient inference.", + "bbox": [ + 496, + 598, + 890, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The point cloud can be easily acquired from depth images by multiplying the camera intrinsic matrix. However, due to its irregularity and disorder, processing point clouds has always been a challenging task until the emergence of PointNet [20, 21]. HandPointNet [9] was the first work that utilized PointNet as a backbone to directly regress the hand pose from the point cloud. Subsequently, the Point-to-Point model [12] used the hierarchical PointNet in order to generate the point-wise probability distribution for improved estimation. Based on these prior studies, SHPR-Net", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "footer", + "text": "20905", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f8fc3533b19a45443def540ecf3c0851163f6aa2c7ea0b2dca8e6fb70b5dd678.jpg", + "image_caption": [ + "Figure 3. The HandR $^2$ N $^2$ pipeline. HandR $^2$ N $^2$ takes the normalized point cloud transformed from a 2D depth image as the input. The PointNet-based local encoder extracts local features from input points. The local features are fed into an initial states generator to initialize hidden states of the residual recurrent unit (RRU). The high-dimensional initialized hidden states are also transformed into 3-dimensional space to generate initial joint estimation. At last, the RRU iteratively resamples points around previously estimation joints to update hidden states and joint locations." + ], + "image_footnote": [], + "bbox": [ + 86, + 93, + 875, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[2] combined HandPointNet with an auxiliary semantic segmentation sub-network to augment the performance. Point-to-Pose [15] further exploited residual permutation equivariant layers [22] to alleviate information loss caused by the max-pooling bottleneck. Recently, HandFoldingNet [4] introduced multiscale-feature-guided folding to reshape a predefined 2D hand skeleton into hand poses.", + "bbox": [ + 75, + 362, + 468, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the existing methods mentioned above followed the single feed-forward design. In such feed-forward architecture, the model capacity can be increased by having a complex architecture or additional parameters. However, such solutions may result in overfitting and higher computational requirement. Unlike these methods, our model exploits the recurrent mechanism to iteratively refine accurate hand pose by reusing a part of parameters.", + "bbox": [ + 75, + 469, + 468, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Recurrent Models for Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 603, + 452, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, recurrent models [7, 35, 33, 34, 36] have been successfully applied in 2D image based hand pose estimation. ACE-Net [7] and SeqHand-Net [35] exploited long short term memory (LSTM) to extract the temporal context information for more smooth and accurate estimations from color images. CADSTN [33] also applied LSTM to jointly model the spatiotemporal context from depth image sequences. However, as the input to these methods is a temporal image sequence, they cannot be used for hand pose estimation from a single image. In contrast, DIR-Net [34] iteratively reuses a part of its structure to refine a downsampled feature map of a single color image. Nonetheless, updating the whole feature map is computationally redundant, since the features without contribution, such as background, are also repeatedly computed. Thus, HCRNN [36] applies RNN on each individual joints for high inference efficiency.", + "bbox": [ + 75, + 627, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although the recurrent models have made significant progress in 2D image based hand pose estimation, they have", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "not been actively studied in the point cloud based hand pose estimation. Taylor et. al. [29] proposed an online optimization method that iteratively optimizes a predefined hand template to fit hand points. A related research [5] iteratively repeated the entire alignment architecture to align hand pose from coarse global stages to fine local stages. It also employed LSTM among multiple global stages to refine global alignment. However, it is significantly computational heavy because it reuses the entire structure for multistage feature generation. Conversely, our network only captures a small part of local information around each joint for each iteration with partial parameters, making it more effective than re-computing global information. Furthermore, we integrate a graph reasoning module in the recurrent unit to enhance estimation by capturing kinematic dependencies between joints, which has not been adopted in [5].", + "bbox": [ + 496, + 362, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Iterative 3D Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 618, + 815, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HandR $^2$ N $^2$ aims to perform iterative hand pose estimation based on a recurrent architecture, as shown in Figure 3. Intuitively, HandR $^2$ N $^2$ refines the joint locations by iteratively searching the optimal position from the region around each joints, as illustrated in Figure 1. The input to HandR $^2$ N $^2$ is a set of preprocessed 3D point coordinates $\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}$ that represents the hand shape, which is transformed from a 2D depth image though the camera intrinsic matrix. The outputs are a sequence of the estimated 3D joints coordinates $\\{\\mathbf{J}_i \\in \\mathbb{R}^{J \\times 3} | i = 0,1,\\dots,I\\}$ , where $\\mathbf{J}_I$ is our final estimation. The $N$ points are firstly input to a local encoder that extracts local geometric features. Since the input point set is irregular and orderless, we exploit furthest_point_sampling method and a one-layer set_conv_layer [16] proposed by PointNet++ [21] to construct the local encoder $(\\mathbb{R}^{N \\times 3} \\to \\mathbb{R}^{(N/2) \\times (3 + d_{local})})$ . Then, the local features are fed into the initial state generator to initialize the", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20906", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/025a2f6904a874ae98dac763b3a201af0ca48274fa0c12a1cd78485f4a7ddbbc.jpg", + "image_caption": [ + "Figure 4. The architecture of the residual graph GRU. It first exploits a graph reasoning module to evolve hidden states. The evolved hidden states are subsequently updated by a GRU cell which takes resampled points as input. After updating, the residual between updated hidden states and previous states are output for joint refinement." + ], + "image_footnote": [], + "bbox": [ + 84, + 93, + 883, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hidden states $\\mathbf{S}_0\\in \\mathbb{R}^{J\\times d_{hid}}$ , which are further processed by an initial joint regressor to obtain the initial estimation of joints $\\mathbf{J}_0$ . Afterward, a novel residual recurrent unit is unrolled $I$ times to iteratively update the hidden states and joint estimations by resampling the useful local features around previously estimated joints. The number of iterations $I$ is a hyper-parameter of the proposed network, and is quite flexible. Specifically, the number of training iterations $(I_{tr})$ can be different from the number of inference iterations $(I_{in})$ , as discussed in Section 4.4.", + "bbox": [ + 75, + 357, + 472, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Residual Recurrent Unit", + "text_level": 1, + "bbox": [ + 76, + 523, + 302, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HandR $^2$ N $^2$ requires a recurrent unit capable of abstracting a set of input point features and capturing kinematic dependencies between joints, in order to perform iterative refinements that optimizes previous estimations. Obviously, the conventional recurrent units such as LSTM or GRU cannot satisfy these requirements, since their input and hidden state should be one single vector, which cannot accommodate the input point set. Therefore, we propose a novel component, a residual recurrent unit (RRU). The RRU consists of the following items: 1) a resampler, 2) a residual graph GRU cell, and 3) a residual regressor. The resampler samples local features around previously generated coordinates $\\mathbf{J}_{i-1}$ forming the input point sets for the current $i$ -th iteration. To update the hidden states with the input of point sets, we propose a novel residual graph GRU that implements its gates based on PointNet. In addition, the proposed GRU also cooperates with graph reasoning to augment the hidden states by capturing kinematic dependencies. At last, the residuals between the current and previous states are sent to the residual regressor to refine the previously generated coordinates.", + "bbox": [ + 75, + 550, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Resampler. The resampler collects $K$ -nearest-neighbor local points and their corresponding local features around", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each of previously estimated joint $\\mathbf{j}_{i - 1}\\in \\mathbf{J}_{i - 1}$ , as visualized in Figure 1. The neighbor points are then translated to a relative coordinate system with $\\mathbf{j}_{i - 1}$ as the origin.", + "bbox": [ + 496, + 357, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Residual graph GRU. The architecture of the proposed residual graph GRU is visualized in Figure 4. Since many recent approaches [8, 23] suggested that a graph convolutional network (GCN) can model the beneficial dependencies among joints, the hidden states from the previous iteration are first augmented through a novel GCN, forming the evolved hidden states:", + "bbox": [ + 496, + 402, + 892, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {i - 1} ^ {\\prime} = \\operatorname {R e L U} (\\mathbf {A} \\mathbf {S} _ {i - 1} \\mathbf {W}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 526, + 890, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W} \\in \\mathbb{R}^{d_{hid} \\times d_{hid}}$ is the trainable weights. Notably, $\\mathbf{A} \\in \\mathbb{R}^{J \\times J \\times d_{hid}}$ is a novel channel-wise independent adjacent matrix, which means for each channel of $S$ , $s_c \\in \\mathbb{R}^{J \\times 1}$ , there is an individual adjacency matrix $a_c \\in \\mathbb{R}^{J \\times J}$ for the multiplication $AS = [a_1 s_1, a_2 s_2, \\dots, a_{d_{hid}} s_{d_{hid}}]$ , where $[\\cdot, \\cdot]$ is concatenation. Afterwards, the GCN-evolved hidden states are passed through a GRU cell to aggregate new information from the sampled local features and to update the hidden states for the current $i$ -th iteration. Here, the conventional GRU cell is not suitable because the input features have one additional dimension than the hidden states. Therefore, we exploit set_conv_layer [16], which is composed of a shared multi-layer perceptron and a max-pooling layer, for gate computation to implement the GRU. First, the hidden states are replicated $K$ times to be concatenated with input features. The concatenated output are then fed into each gate for computation. For those gates that directly modify values of the hidden states, we implement them with set_conv_layer to squeeze one dimension. Moreover, we implemented a novel attentive reset gate with a point-wise multi-layer perceptron (MLP) that enables the GRU to select good points for the hidden state update, which is beneficial for exploring the optimal joint", + "bbox": [ + 496, + 551, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20907", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78c2d33e5af06e8f2d204e0d83e343c13c96f5476b7e36190baa5ad370bc0981.jpg", + "image_caption": [ + "Figure 5. Visualization of the output from the reset gate in the proposed GRU. The small colored points in the circle are the grouped neighbors of the red input joint $j_{i-1}$ , and their colors indicate the attentive intensity computed by Eq. 3. It shows that the points closer to the ground truth obtain relatively higher intensity. By leveraging clues to focus on highly-related points near the ground truth, the attentive GRU can output the accurate orange joint $j_i$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 465, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "locations, as shown in Figure 5. Thus, given the evolved hidden state $\\mathbf{s}_{i - 1}^{\\prime}\\in \\mathbf{S}_{i - 1}^{\\prime}$ of a specific joint, and with the sampled local features $\\{\\mathbf{f}_k\\in \\mathbb{R}^{3 + d_{local}}|1\\leq k\\leq K\\}$ , the GRU generates the updated hidden state $\\mathbf{s}_i$ :", + "bbox": [ + 75, + 305, + 470, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} = \\sigma \\left(\\operatorname {s e t} _ {1 \\leq k \\leq K} \\operatorname {c o n v} _ {z} \\left(\\left[ \\mathbf {s} _ {i - 1} ^ {\\prime}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 380, + 468, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {i, k} = \\sigma \\left(M L P _ {r} \\left(\\left[ \\mathbf {s} _ {i - 1} ^ {\\prime}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 416, + 468, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {s}} _ {i} = \\tanh \\left(\\underset {1 \\leq k \\leq K} {\\operatorname {s e t}}. \\operatorname {c o n v} _ {h} \\left(\\left[ \\mathbf {r} _ {i, k} \\odot \\mathbf {s} ^ {\\prime} _ {i - 1}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 452, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {i} = \\left(1 - \\mathbf {z} _ {i}\\right) \\odot \\mathbf {s} _ {i - 1} ^ {\\prime} + \\mathbf {z} _ {i} \\odot \\tilde {\\mathbf {s}} _ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 489, + 468, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid activation function, $\\left[\\cdot ,\\cdot \\right]$ is the concatenation operation, and $\\odot$ is the Hadamard product. Note that, the updated hidden states encode the absolute joint locations in the latent space. However, we require relative representations to refine the previously generated joints. Therefore, we introduce a residual mechanism that outputs the relative representations of the updated information:", + "bbox": [ + 75, + 511, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {S} _ {i} = \\mathbf {S} _ {i} - \\mathbf {S} _ {i - 1}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 627, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Residual regressor. The regressor accepts residual states of the current $i$ -th iteration as input to refine the previous joints generated by the $(i - 1)$ -th iteration. The regressor is a linear transformation with a residual connection. Therefore, the joint regression of the current $i$ -th iteration is represented as:", + "bbox": [ + 75, + 652, + 468, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} _ {i} = \\Delta \\mathbf {S} _ {i} \\mathbf {W} + \\mathbf {J} _ {i - 1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 753, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W} \\in \\mathbb{R}^{d_{hid} \\times 3}$ is the trainable transformation matrix.", + "bbox": [ + 76, + 777, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Initialization of the Hidden States and Joint Coordinates", + "text_level": 1, + "bbox": [ + 76, + 801, + 468, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The hidden states $\\mathbf{S}$ of the RRU essentially are the embeddings that represent joints in the $d_{hid}$ -dimensional latent space and can be further re-projected into the three-dimensional coordinates by the subsequent transformation.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, we inherit the idea of generating joint-wise embeddings to obtain the initialized hidden states $\\mathbf{S}_0$ , as proposed by HandFoldingNet [4]. The generation of the initial hidden states (joint-wise embeddings) are sequentially proceeded by two cascaded set_conv_layers and a three-layer bias-induced layer (BIL) [3]. Equivalent with HandFoldingNet [4], the cascaded set_conv_layers first down-sample and extract $N/8$ of local features and subsequently aggregate them as a global feature. The global feature is then replicated $J$ times and fed into the BILs for the generation of hidden states for $J$ joints. Finally, the $J$ hidden states are injected into the regressor, which performs a linear transformation $(\\mathbb{R}^{d_{hid}} \\to \\mathbb{R}^3)$ for the initialization of the initial joint $\\mathbf{J}_0$ .", + "bbox": [ + 496, + 90, + 892, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Loss Function", + "text_level": 1, + "bbox": [ + 498, + 311, + 718, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the previous works [25, 4], we adopt smooth L1 loss to supervise training because of its less sensitivity to the outliers. The smooth L1 loss is defined as:", + "bbox": [ + 496, + 334, + 892, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL 1 _ {\\text {s m o o t h}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 5 0 \\mathbf {x} ^ {2}, & | \\mathbf {x} | < 0. 0 1 \\\\ | \\mathbf {x} | - 0. 0 0 5, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 388, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the proposed network outputs an initial estimation and $I$ iterative estimations, we supervise all of the outputs by the following joint loss function:", + "bbox": [ + 496, + 438, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 0} ^ {I} \\sum_ {j = 0} ^ {J} \\lambda^ {I - i} L 1 _ {\\text {s m o o t h}} \\left(\\mathbf {J} _ {i j} - \\mathbf {J} _ {j} ^ {*}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 494, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{J}^*$ indicates the ground-truth coordinates and $\\lambda$ is the hyperparameter that weights each distinct iteration. In our experiments, we set $\\lambda = 0.8$ by default.", + "bbox": [ + 496, + 549, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 608, + 633, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 500, + 633, + 692, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted experiments on an NVIDIA A100 GPU with PyTorch. For training, we used the AdamW optimizer [17] with beta1 = 0.5, beta2 = 0.999, and learning rate $\\alpha$ = 0.001. The number of input points to the network was sampled to 1,024 and the batch size was set to 32. Batch normalization [14] was applied only in the local encoder and initializer. Meanwhile, to avoid overfitting, we adopted online data augmentation with random rotation ([-120.0, 120.0] degrees around z-axis, [-30.0, 30.0] degrees around x- and y-axis), 3D scaling ([0.8, 1.2]), and 3D translation ([-20, 20] mm). We trained the model for 160 epochs with a learning rate decay of 0.1 after 120 epochs.", + "bbox": [ + 496, + 656, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 847, + 785, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSRA Dataset. The MSRA dataset [27] provides more than 76K frames captured by the Intel's Creative Interactive", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20908", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/04e46ddfbe13258c6aa7b315328c281fc857a81deecc074f7c3facd32d27580e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMean joint error (mm)Input# Params# FLOPsTime (ms)
ICVLMSRANYU
Ren-4x6x6 [13]7.63-13.392D image---
Ren-9x6x6 [13]7.319.712.692D image---
Pose-Ren [1]6.798.6511.812D image---
DenseReg [32]7.37.210.22D image5.8M--
CrossInfoNet [6]6.737.8610.082D image23.8M--
JGR-P2O [8]6.027.558.292D image1.4M-2.0 + 7.0
SSRN-SD [24]6.047.167.472D image16.2M1.52G0.4 + 3.0
SSRN [24]6.017.057.372D image22.5M1.96G0.4 + 3.4
PHG [23]5.976.947.392D image35.7M27.02G0.4 + 15.8
3DCNN [11]-9.614.13D voxels104.9M-2.9 + 1.68
SHPR-Net [2]7.227.7610.783D points---
HandPointNet [9]6.948.510.543D points2.5M1.72G8.2 + 11.3
Point-to-Point [12]6.37.79.103D points4.3M2.98G8.2 + 15.7
V2V [18]6.287.598.423D voxels34.1M6.23G23 + 5.5
HandFolding [4]5.957.348.583D points1.3M1.31G8.2 + 3.7
Deng et al. [5]6.057.006.843D points--0.9 + 47.6
Ours\\( I_{in} \\)=16.516.938.003D points1.3M0.72G0.9 + 4.2
\\( I_{in} \\)=26.126.497.503D points1.3M1.06G0.9 + 6.0
\\( I_{in} \\)=35.866.427.363D points1.3M1.40G0.9 + 7.9
\\( I_{in} \\)=45.706.487.293D points1.3M1.74G0.9 + 9.7
\\( I_{in} \\)=55.716.547.273D points1.3M2.08G0.9 + 11.5
", + "bbox": [ + 168, + 88, + 802, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of the proposed method with previous state-of-the-art methods on the ICVL, MSRA and NYU datasets. Input indicates the input type of 2D depth image, 3D voxels or 3D point cloud. # Params indicates the number of model parameters. Time stands for the total computation time including preprocessing time and model inference time.", + "bbox": [ + 75, + 450, + 893, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gesture Camera. Each frame is annotated with $J = 21$ joints, including one joint for the wrist and four joints for each finger. We followed the leave-one-subject-out cross-validation strategy as used in the previous works [4, 27] to evaluate this dataset.", + "bbox": [ + 75, + 520, + 468, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ICVL Dataset. The ICVL dataset [28] provides 22K training and 1.6K testing depth frames captured by the Intel's Creative Interactive Gesture Camera. Each frame is annotated with $J = 16$ joints, including one joint for the palm and three joints for each finger. We cropped the hand area from a depth image using the method given in [19] because the human body area is present in the frames.", + "bbox": [ + 75, + 595, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NYU Dataset. The NYU dataset is captured from three different views. Each view contains 72K training 8K testing depth images captured with the PrimeSense 3D sensor. Following recent works [4, 12, 9], we selected one view and 14 joints out of total of 36 annotated joints for training and testing [31]. We also followed the same hand area cropping process [19] as in the ICVL dataset.", + "bbox": [ + 75, + 702, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We employed two commonly used metrics, the mean joint error and the success rate, to evaluate the performance of hand pose estimation. The mean joint error quantifies the average Euclidean distance between the estimated joint locations and ground-truth ones for all the joints over the entire testing set. The success rate", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "indicates the percentage of the good frames with a mean joint error of less than a given distance.", + "bbox": [ + 498, + 520, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 498, + 560, + 867, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared $\\mathrm{HandR}^2\\mathrm{N}^2$ with other state-of-the-art methods, including methods with 2D depth images as input: region ensemble network (Ren-4x6x6 [13], Ren-9x6x6 [32]), Pose-Ren [1], dense regression network (DenseReg) [32], CrossInfoNet [6], JGR-P2O [8], spatial-aware stacked regression network (SSRN [24], SSRN-SD [24]) and pose-guided hierarchical graph network (PHG) [23], and methods with 3D point cloud or voxels as input: 3DCNN [11], SHPR-Net [2], HandPointNet [9], Point-to-Point [12], V2V [18], HandFoldingNet [4] and Deng et al. [5].", + "bbox": [ + 496, + 583, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 summarizes the performance in terms of the mean joint error. Note that, the evaluations of our model were trained with $I_{tr} = 3$ and tested with $I_{in} = 4$ , $I_{in} = 3$ , $I_{in} = 5$ on ICVL, MSRA, and NYU, respectively. The results show that $\\mathrm{HandR}^2\\mathrm{N}^2$ achieves the new state-of-the-art performance with mean distance errors of $5.70\\mathrm{mm}$ , $6.42\\mathrm{mm}$ and $7.27\\mathrm{mm}$ on two challenging datasets, ICVL and MSRA, respectively. It also achieves the second-lowest error on the NYU dataset. In particular, $\\mathrm{HandR}^2\\mathrm{N}^2$ significantly outperforms 3D input based previous state-of-the-arts with large margins. Meanwhile, $\\mathrm{HandR}^2\\mathrm{N}^2$ requires", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20909", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd5f6082666ad6ffcef578531e9e8cbe04f7ddad6f5f6df1a8488095c39e5f0b.jpg", + "image_caption": [ + "Figure 6. Comparison with the state-of-the-art methods using the ICVL (left), MSRA (middle) and NYU (right) dataset. The success rate is shown in this figure." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 346, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d3247c7fa7e86fb0b5c14500016158c66072ffb99d53bcdaa613b5d352bee22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 88, + 614, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e478b198f5035d1b37604efd8be370bfc88dacc21857933be203d084ccaf7e30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 88, + 885, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2d6f376e96c92406598b81793797c40c4d294c27cacdce5e47cdb79f98ff666.jpg", + "image_caption": [ + "Figure 7. Qualitative results of $\\mathrm{HandR}^2\\mathrm{N}^2$ on the ICVL (left), MSRA (middle) and NYU (right) dataset. Hand depth images are transformed into 3D points as shown in the figure. Ground truth is shown in black, and the estimated joint coordinates are shown in red." + ], + "image_footnote": [], + "bbox": [ + 88, + 284, + 892, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e130e013ed2c2cd5695c6d4ae3229a7370fb7a296a75268273d7a0a69317e28.jpg", + "image_caption": [ + "Figure 8. Training $I_{tr}$ vs. inference $I_{in}$ . For each column, the model are trained based on a fixed iteration number and evaluated at different inference $I_{in}$ ." + ], + "image_footnote": [], + "bbox": [ + 89, + 450, + 468, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only 1.3M parameters, which is generally less compared to other recent state-of-the-art methods. Figure 6 represents that our method significantly outperforms other methods in terms of success rate when the error threshold is lower than $20\\mathrm{mm}$ , $15\\mathrm{mm}$ and $52\\mathrm{mm}$ on the ICVL, MSRA and NYU datasets, respectively. Figure 2 also illustrates the superior efficiency of $\\mathrm{HandR}^2\\mathrm{N}^2$ with dynamically adjustable computation of $0.72 + 0.34\\times I$ GFLOPs. The results show that our model achieves the lowest estimation error with a similar operation count at $I_{in} = 3$ compared to other state-ofthe-art methods. Furthermore, the model can dynamically", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "run more iterations with modestly increased computation for lower estimation error.", + "bbox": [ + 498, + 446, + 890, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 484, + 653, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the NYU dataset, since it is challenging and far from saturation.", + "bbox": [ + 496, + 508, + 890, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of iterations. Although the number of iterations $I$ is a specific number during training, it can be quite flexible and even larger for inference. We first trained the model using different numbers of training $I_{tr}$ to find the optimal setting of iterations for the training phase. Then we test the trained models with a various inference $I_{in}$ . The results are plotted on Figure 8. Interestingly, when the training $I_{tr} \\geq 3$ , our model keeps improving for a few more iterations even though the inference $I_{in}$ is larger than the training. In particular, when setting $I_{tr} = 3$ during training, the model achieves the minimal estimation error of 7.27 mm after 5 inference iterations. Please note that, in small number of training iterations $(I_{tr} < 3)$ , the model is optimized solely for the same inference iterations $(I_{in} = I_{tr})$ . On the other hand, larger $I_{tr}$ allows the model to generalize the regularity of changes between all iteration, which benefits not only trained iterations $(I_{in} = I_{tr})$ but also untrained larger iterations $(I_{in} > I_{tr})$ . Nonetheless, Figure 8 further demonstrates the reduced performance of our model when $I_{tr} \\geq 5$ during training. It is because the lengthy path of back-propagation restricts the model from obtaining appropriate estimations. The longer the iteration path is, the", + "bbox": [ + 496, + 568, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20910", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6207f72d8c106f2c90288afdb11b25930f22b2b140a2f40f7dfcbe86beb72878.jpg", + "image_caption": [ + "Figure 9. Qualitative visualization of the models with training $I_{tr} = 3$ evaluated at different inference $I_{in}$ based on the NYU dataset. Mean error indicates the mean joint error. Ground truth is shown in black, and the estimated joint coordinates are shown in red." + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 890, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more different the estimations of the 0-th iteration and the $I$ -th iteration are. Thus, we assume that the accumulated gradients of all iterations to the shared parameters would smooth the right descent gradients for each individual iteration, and thus degrade the final accuracy. However, these smooth accumulated gradients within a specific range of iteration count ( $3 < I_{tr} < 5$ according to Figure 8) allow the model to learn the general optimization direction of joints between iterations. Based on the results, we can empirically find the optimal $I_{tr}$ as 3 for high inference performance and training efficiency.", + "bbox": [ + 75, + 382, + 472, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e13db47d14bbe1f1d88103de84d7c0f41881d48137934c18126c1586c581efbc.jpg", + "image_caption": [ + "Figure 10. Affect of sampling $K$ . For each line, the inference $I_{in}$ s are same, but with increased $K$ of resampled points for each iteration. All models are trained based on training $I_{tr} = 3$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 565, + 464, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, as Table 1 reports, the estimation accuracy is improved with a small number of additional computations (0.34GFLOPs per iteration) and time consumption (1.8 ms per iteration).", + "bbox": [ + 75, + 823, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of resampled points. The resampled points pro", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/41cf5b59ad355759155156d59ece2fade0af09fda3ff37f7132c03f897f02a9c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GRUGCNRes.Mean error
×××8.68 mm
✓××8.06 mm
×✓✓7.76 mm
✓✓×7.62 mm
✓×✓7.99 mm
✓✓✓7.36 mm
", + "bbox": [ + 584, + 378, + 807, + 481 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ef54c32bac0ff4839187a4358f0cec17d0c8f27faf776a9442682faa2a8d427.jpg", + "table_caption": [ + "Table 2. Ablations of different GRU configurations. GCN and Res. indicate whether the GCN and residual mechanism are activated in the RRU, respectively. All the ablation models are trained and tested with both $I_{tr} = I_{in} = 3$ ." + ], + "table_footnote": [], + "table_body": "
ConfigurationsMean error
Conventional GCN + Res.8.14 mm
Channel-wise GCN + Res.7.76 mm
Conventional GRU + Res.8.59 mm
Attentive GRU + Res.7.99 mm
Conventional GRU + GCN + Res.8.07 mm
Attentive GRU + Channel-wise GCN + Res.7.36 mm
", + "bbox": [ + 514, + 561, + 877, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison with conventional components. All models are trained and tested with both $I_{tr} = I_{in} = 3$ .", + "bbox": [ + 500, + 676, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vide $K$ new candidates for location refinement around each joint. Thus, the number $K$ is a significant hyperparameter that affects the performance. Figure 10 presents the performance under different $K$ values and inference iterations. The result shows that larger $K$ reduces the estimation error since it provides a wider range of observation. However, the performance is saturated at $K = 48$ , because additionally sampled points beyond the effective region of a particular joint (i.e., points that are distant to the joint and closer to other joints) do not contribute to the estimation.", + "bbox": [ + 496, + 732, + 890, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of different GRU configurations. To augment", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20911", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the final estimation, we combine the novel GRU with the GCN that evolves the hidden states and residual mechanism for relative representation, as described in Sec. 3. To verify their contributions, we incrementally adopted the GRU cell, the GCN and residual mechanism. Thus, we conducted five ablations as follows: 1) no GRU, 2) GRU; 3) GCN with residual; 4) GRU with GCN; 5) GRU with residual; 6) GRU with both residual and GRU, which is our full configuration. Note that, the ablations without residual, 2) and 4), directly transform the hidden states into absolute joint coordinates. As shown in Table 2, the GRU cell, the GCN and residual mechanism all contribute considerably to the effective estimation. In particular, the novel GRU improves the accuracy by more than $0.4\\mathrm{mm}$ . Furthermore, using the novel GCN and the residual mechanism further reduces the estimation error by more than $0.4\\mathrm{mm}$ and $0.2\\mathrm{mm}$ , respectively.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison with the conventional components. To validate the effectiveness of our proposed components, we replace the proposed channel-wise GCN and attentive GRU with the conventional ones, respectively. The implementation details are provided in the supplementary material. Table 3 indicates our proposed GCN and GRU both significantly outperforms the conventional implementations.", + "bbox": [ + 75, + 337, + 472, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 472, + 194, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper presented $\\mathrm{HandR}^2\\mathrm{N}^2$ , a novel recurrent architecture that is capable of iteratively regressing progressively accurate 3D hand pose from a single point cloud. Experimental results showed that our network significantly outperforms previous state-of-the-art methods on three challenging datasets. Extensive experiments also demonstrated the excellent flexibility and efficiency of our network. As limitations, the performance of $\\mathrm{HandR}^2\\mathrm{N}^2$ may degrade when hand parts are severely occluded, because the RRU can not sample efficient points. Extending this work to temporal-sequential learning may help to alleviate these limitations, which we leave for future studies.", + "bbox": [ + 75, + 503, + 472, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement.", + "text_level": 1, + "bbox": [ + 76, + 719, + 240, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Ministry of Science and ICT (MSIT) of Korea under the National Research Foundation (NRF) grant (2022R1A4A3032913) and the Institute of Information and Communication Technology Planning Evaluation (IITP) grants (IITP-2019-0-00421, IITP-2023-2020-0-01821, IITP-2021-0-02052, IITP-2021-0-02068), and by the Technology Innovation Program (RS-2023-00235718) funded by the Ministry of Trade, Industry & Energy (1415187474). Wencan Cheng was partly supported by the China Scholarship Council (CSC).", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 90, + 596, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xinghao Chen, Guijin Wang, Hengkai Guo, and Cairong Zhang. Pose guided structured region ensemble network for cascaded hand pose estimation. Neurocomputing, 395:138-149, 2020. 1, 2, 6", + "[2] Xinghao Chen, Guijin Wang, Cairong Zhang, Tae-Kyun Kim, and Xiangyang Ji. Shpr-net: Deep semantic hand pose regression from point clouds. IEEE Access, 6:43425-43439, 2018. 1, 2, 3, 6", + "[3] Wencan Cheng and Sukhan Lee. Point auto-encoder and its application to 2d-3d transformation. In International Symposium on Visual Computing, pages 66-78. Springer, 2019. 5", + "[4] Wencan Cheng, Jae Hyun Park, and Jong Hwan Ko. Handfoldingnet: A 3d hand pose estimation network using multiscale-feature guided folding of a 2d hand skeleton. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11260-11269, 2021. 1, 2, 3, 5, 6", + "[5] Xiaoming Deng, Dexin Zuo, Yinda Zhang, Zhaopeng Cui, Jian Cheng, Ping Tan, Liang Chang, Marc Pollefeys, Sean Fanello, and Hongan Wang. Recurrent 3d hand pose estimation using cascaded pose-guided 3d alignments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3, 6", + "[6] Kuo Du, Xiangbo Lin, Yi Sun, and Xiaohong Ma. Crossinfonet: Multi-task information sharing based hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9896-9905, 2019. 1, 2, 6", + "[7] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 3", + "[8] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In European Conference on Computer Vision, pages 120-137. Springer, 2020. 1, 2, 4, 6", + "[9] Liuhao Ge, Yujun Cai, Junwu Weng, and Junsong Yuan. Hand pointnet: 3d hand pose estimation using point sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8417-8426, 2018. 1, 2, 6", + "[10] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 1, 2", + "[11] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1991-2000, 2017. 1, 2, 6", + "[12] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 475-491, 2018. 1, 2, 6" + ], + "bbox": [ + 501, + 114, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20912", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Hengkai Guo, Guijin Wang, Xinghao Chen, Cairong Zhang, Fei Qiao, and Huazhong Yang. Region ensemble network: Improving convolutional network for hand pose estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 4512-4516. IEEE, 2017. 1, 2, 6", + "[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5", + "[15] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11927-11936, 2019. 1, 2, 3", + "[16] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 529-537, 2019. 3, 4", + "[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[18] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 1, 2, 6", + "[19] Markus Oberweger and Vincent Lepetit. Deeprior++: Improving fast and accurate 3d hand pose estimation. In Proceedings of the IEEE international conference on computer vision Workshops, pages 585-594, 2017. 6", + "[20] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2", + "[21] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 1, 2, 3", + "[22] Siamak Ravanbakhsh, Jeff Schneider, and Barnabas Poczos. Deep learning with sets and point clouds. arXiv preprint arXiv:1611.04500, 2016. 3", + "[23] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 2021. 1, 2, 4, 6", + "[24] Pengfei Ren, Haifeng Sun, Weiting Huang, Jiachang Hao, Daixuan Cheng, Qi Qi, Jingyu Wang, and Jianxin Liao. Spatial-aware stacked regression network for real-time 3d hand pose estimation. Neurocomputing, 437:42-57, 2021. 6", + "[25] Pengfei Ren, Haifeng Sun, Qi Qi, Jingyu Wang, and Weiting Huang. Srn: Stacked regression network for real-time 3d hand pose estimation. In BMVC, page 112, 2019. 1, 2, 5", + "[26] Oindrila Saha, Aditya Kusupati, Harsha Vardhan Simhadri, Manik Varma, and Prateek Jain. Rnnpool: efficient nonlinear pooling for ram constrained inference. Advances in Neural Information Processing Systems, 33:20473-20484, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Xiao Sun, Yichen Wei, Shuang Liang, Xiaou Tang, and Jian Sun. Cascaded hand pose regression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 824-832, 2015. 2, 5, 6", + "[28] Danhang Tang, Hyung Jin Chang, Alykhan Tejani, and Tae Kyun Kim. Latent regression forest: Structured estimation of 3d articulated hand posture. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3786-3793, 2014. 2, 6", + "[29] Jonathan Taylor, Lucas Bordeaux, Thomas Cashman, Bob Corish, Cem Keskin, Toby Sharp, Eduardo Soto, David Sweeney, Julien Valentin, Benjamin Luff, et al. Efficient and precise interactive hand tracking through joint, continuous optimization of pose and correspondences. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3", + "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2", + "[31] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014. 1, 2, 6", + "[32] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2018. 6", + "[33] Yiming Wu, Wei Ji, Xi Li, Gang Wang, Jianwei Yin, and Fei Wu. Context-aware deep spatiotemporal network for hand pose estimation from depth images. IEEE transactions on cybernetics, 50(2):787-797, 2018. 3", + "[34] John Yang, Yash Bhalgat, Simyung Chang, Fatih Porikli, and Nojun Kwak. Dynamic iterative refinement for efficient 3d hand pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1869-1879, 2022. 2, 3", + "[35] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 2, 3", + "[36] Cheol-Hwan Yoo, Seowon Ji, Yong-Goo Shin, Seung-Wook Kim, and Sung-Jea Ko. Fast and accurate 3d hand pose estimation via recurrent neural network for capturing hand articulations. IEEE Access, 8:114010-114019, 2020. 3" + ], + "bbox": [ + 501, + 90, + 890, + 698 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20913", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_model.json b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39a50db50c431e514643ac69d2ec43fc921d3a28 --- /dev/null +++ b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_model.json @@ -0,0 +1,1881 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.175 + ], + "angle": 0, + "content": "HandR\\(^2\\)N\\(^2\\): Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.205, + 0.391, + 0.274 + ], + "angle": 0, + "content": "Wencan Cheng \nDepartment of Artificial Intelligence \nSungkyunkwan University \ncwc1260@skku.edu" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.205, + 0.875, + 0.275 + ], + "angle": 0, + "content": "Jong Hwan Ko \nCollege of Information and Communication Engineering \nSungkyunkwan University \njhko@skku.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.63 + ], + "angle": 0, + "content": "3D hand pose estimation is a critical task in various human-computer interaction applications. Numerous deep learning based estimation models in this domain have been actively explored. However, the existing models follow a non-recurrent scheme and thus require complex architectures or redundant parameters in order to achieve acceptable model capacity. To tackle this limitation, this paper proposes HandR²N², a compact neural network that iteratively regresses the hand pose using a novel residual recurrent unit. The recurrent design allows recursive exploitation of partial layers to gradually optimize previously estimated joint locations. In addition, we exploit graph reasoning to capture kinematic dependencies between joints for better performance. Experimental results show that the proposed model significantly outperforms the existing methods on three hand pose benchmark datasets in terms of both accuracy and efficiency. Codes and pre-trained models are publicly available at https://github.com/cwc1260/HandR2N2." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.661, + 0.21, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.856 + ], + "angle": 0, + "content": "3D hand pose estimation, which estimates the 3D positions of hand keypoints, provides a fundamental understanding of human hand activity. Thus, it is a critical task for many human-computer interaction applications such as robotics and augmented/virtual reality. With the assistance of deep learning techniques and low-cost depth cameras, numerous studies have contributed to the substantial progress in the field of hand pose estimation. However, accurate and efficient hand pose estimation is still challenging because of the various probability of hand orientations, severe self-occlusion, and noisy depth images [12, 9, 18, 6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Most of the recent state-of-the-art approaches are based on deep learning, especially deep convolutional neural networks (CNNs). Such CNN-based approaches [31, 10, 13," + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.307, + 0.865, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.449, + 0.895, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Illustration of the recurrent concept. For each iteration, our proposed recurrent architecture samples local points around joints from previous iteration to generate displacements to refine the joint locations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.779 + ], + "angle": 0, + "content": "25, 1, 8, 23, 6, 23] generally accept 2D depth images as an input, which can be directly processed by the 2D convolutional layers. However, it is difficult for a 2D CNN to capture the highly non-linear mapping from a 2D depth image to the 3D hand pose, significantly restricting its efficiency and accuracy. Therefore, a line of works [11, 18] discretized 2D input images into 3D volumetric representations and subsequently applied a 3D CNN for direct 3D-to-3D inference. However, a critical drawback of a 3D CNN is its computational overhead that increases cubically with an increase of the input resolution [4]. In contrast, another series of hand pose estimation models [9, 12, 2, 15, 4] have presented remarkable performances by using PointNet [20, 21], which allows the models to accept a set of continuous 3D coordinates converted from an input hand image without discretization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Despite the accuracy improvement achieved by the aforementioned approaches, they still suffer from large and complex model architectures that require substantial computational overheads. Thus, these methods are not suitable for resource-limited devices such as wearable or hand-held devices. Furthermore, all the existing approaches follow a non-recurrent architecture. As a result, they are not flexible for varying resource constraints and accuracy targets." + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.938 + ], + "angle": 0, + "content": "20904" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "In this work, we tackle these limitations using a recurrent architecture whose capacity can be dynamically adjusted to achieve either higher accuracy or higher efficiency by iteratively using a part of model parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.47, + 0.35 + ], + "angle": 0, + "content": "The recurrent structure can achieve the high capacity of deeper networks by the recursive use of fewer shared parameters [34]. Various image processing models that used a recurrent neural network (RNN) as an explicit iterative approach have already yielded encouraging results [30, 5, 35, 26]. To fully take advantage of the RNN in a hand pose estimation task, we propose a novel residual recurrent neural network, \\(\\mathrm{HandR}^2\\mathrm{N}^2\\), that iteratively regresses accurate 3D hand pose from an input hand point cloud. Intuitively, \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) can be served as a general optimizer that recursively searches the optimal joint locations using the observation captured by the previous estimation, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.47, + 0.458 + ], + "angle": 0, + "content": "To implement \\(\\mathrm{HandR}^2\\mathrm{N}^2\\), we introduce two novel components: 1) a residual recurrent unit (RRU) and 2) an initialization module. The RRU is a key component that iteratively proceeds for updating its joint-wise hidden states and coordinate estimations. The initial joint-wise states and coordinates of the RRU are globally estimated by the initialization module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.763 + ], + "angle": 0, + "content": "The RRU specifically integrates with an attentive gate that can explore optimal points from a group of candidate points for the improved accuracy. Since the proposed RRU only computes the features contributed to the specific joints, it is more efficient than the similar prior model in [5], which re-computes the entire point features. Moreover, RRU exploits graph reasoning, which is based on a novel channelwise graph convolutional network (GCN). GCN is capable to model the graph-structured data, making it particularly suitable for the kinematically structured hand data [8]. The use of the proposed channel-wise GCN enables the RRU to recursively enhance the hidden states by capturing the strong kinematic dependencies between joints. More importantly, the number of RRU iterations during inference can be dynamically scaled to be different from the training iteration. Specifically, the RRU inference iterations can be configured large to enhance the model capacity and accuracy, or small to reduce the computation requirement. Therefore, \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) is capable of adapting varying complexity and accuracy requirements as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We evaluate \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) on three challenging benchmarks, ICVL [28], MSRA [27], and NYU [31] datasets. The results show that our network achieves a new state-of-the-art record. The proposed network reports the minimum mean distance errors of \\(5.70\\mathrm{mm}\\), \\(6.42\\mathrm{mm}\\) and \\(7.27\\mathrm{mm}\\) on the ICVL, MSRA and NYU datasets, respectively. Meanwhile, it requires only \\(1.3\\mathrm{M}\\) parameters and runs with adjustable complexity \\(0.72 + 0.34\\times I\\) GFLOPs, where \\(I\\) indicates the number of iterations. Thus, our model requires" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.09, + 0.857, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.21, + 0.892, + 0.237 + ], + "angle": 0, + "content": "Figure 2. Comparison between operation count and estimation error. \\(I\\) is the iteration number of \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) during inference." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.264, + 0.892, + 0.294 + ], + "angle": 0, + "content": "less computation for the highest accuracy compared with other state-of-the-art methods, as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.295, + 0.892, + 0.324 + ], + "angle": 0, + "content": "The key contributions of this paper are summarised as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.335, + 0.892, + 0.395 + ], + "angle": 0, + "content": "- We propose a novel iterative neural network architecture that takes the hand point cloud as input and progressively estimates the accurate 3D hand joint coordinates." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.406, + 0.892, + 0.497 + ], + "angle": 0, + "content": "- We propose a novel residual recurrent unit that captures new local observations around previously generated coordinates to refine joint-wise hidden states and coordinate estimation. It also leverages a graph neural network to model dependency between the joints for performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.507, + 0.892, + 0.537 + ], + "angle": 0, + "content": "- We conduct extensive experiments to analyze the efficiency and effectiveness of our proposed network." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.335, + 0.892, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.642, + 0.566 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.576, + 0.834, + 0.592 + ], + "angle": 0, + "content": "2.1. Depth-based 3D Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.749 + ], + "angle": 0, + "content": "3D hand pose estimation approaches using depth images can be classified into three categories according to the input types: conventional 2DCNN-based methods that take raw 2D depth images as input [31, 10, 13, 25, 1, 8, 23, 6], 3DCNN-based methods that use 3D voxelized representations of depth images [11, 18], and PointNet-based methods that process the point cloud transformed from depth images [9, 12, 2, 15, 4]. In this section, we focus on PointNet-based methods, since the point cloud is an accurate and concise representation for efficient inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The point cloud can be easily acquired from depth images by multiplying the camera intrinsic matrix. However, due to its irregularity and disorder, processing point clouds has always been a challenging task until the emergence of PointNet [20, 21]. HandPointNet [9] was the first work that utilized PointNet as a backbone to directly regress the hand pose from the point cloud. Subsequently, the Point-to-Point model [12] used the hierarchical PointNet in order to generate the point-wise probability distribution for improved estimation. Based on these prior studies, SHPR-Net" + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20905" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.094, + 0.877, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.894, + 0.337 + ], + "angle": 0, + "content": "Figure 3. The HandR\\(^2\\)N\\(^2\\) pipeline. HandR\\(^2\\)N\\(^2\\) takes the normalized point cloud transformed from a 2D depth image as the input. The PointNet-based local encoder extracts local features from input points. The local features are fed into an initial states generator to initialize hidden states of the residual recurrent unit (RRU). The high-dimensional initialized hidden states are also transformed into 3-dimensional space to generate initial joint estimation. At last, the RRU iteratively resamples points around previously estimation joints to update hidden states and joint locations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[2] combined HandPointNet with an auxiliary semantic segmentation sub-network to augment the performance. Point-to-Pose [15] further exploited residual permutation equivariant layers [22] to alleviate information loss caused by the max-pooling bottleneck. Recently, HandFoldingNet [4] introduced multiscale-feature-guided folding to reshape a predefined 2D hand skeleton into hand poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.591 + ], + "angle": 0, + "content": "However, the existing methods mentioned above followed the single feed-forward design. In such feed-forward architecture, the model capacity can be increased by having a complex architecture or additional parameters. However, such solutions may result in overfitting and higher computational requirement. Unlike these methods, our model exploits the recurrent mechanism to iteratively refine accurate hand pose by reusing a part of parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.454, + 0.619 + ], + "angle": 0, + "content": "2.2. Recurrent Models for Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Recently, recurrent models [7, 35, 33, 34, 36] have been successfully applied in 2D image based hand pose estimation. ACE-Net [7] and SeqHand-Net [35] exploited long short term memory (LSTM) to extract the temporal context information for more smooth and accurate estimations from color images. CADSTN [33] also applied LSTM to jointly model the spatiotemporal context from depth image sequences. However, as the input to these methods is a temporal image sequence, they cannot be used for hand pose estimation from a single image. In contrast, DIR-Net [34] iteratively reuses a part of its structure to refine a downsampled feature map of a single color image. Nonetheless, updating the whole feature map is computationally redundant, since the features without contribution, such as background, are also repeatedly computed. Thus, HCRNN [36] applies RNN on each individual joints for high inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Although the recurrent models have made significant progress in 2D image based hand pose estimation, they have" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.605 + ], + "angle": 0, + "content": "not been actively studied in the point cloud based hand pose estimation. Taylor et. al. [29] proposed an online optimization method that iteratively optimizes a predefined hand template to fit hand points. A related research [5] iteratively repeated the entire alignment architecture to align hand pose from coarse global stages to fine local stages. It also employed LSTM among multiple global stages to refine global alignment. However, it is significantly computational heavy because it reuses the entire structure for multistage feature generation. Conversely, our network only captures a small part of local information around each joint for each iteration with partial parameters, making it more effective than re-computing global information. Furthermore, we integrate a graph reasoning module in the recurrent unit to enhance estimation by capturing kinematic dependencies between joints, which has not been adopted in [5]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.816, + 0.635 + ], + "angle": 0, + "content": "3. Iterative 3D Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "HandR\\(^2\\)N\\(^2\\) aims to perform iterative hand pose estimation based on a recurrent architecture, as shown in Figure 3. Intuitively, HandR\\(^2\\)N\\(^2\\) refines the joint locations by iteratively searching the optimal position from the region around each joints, as illustrated in Figure 1. The input to HandR\\(^2\\)N\\(^2\\) is a set of preprocessed 3D point coordinates \\(\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}\\) that represents the hand shape, which is transformed from a 2D depth image though the camera intrinsic matrix. The outputs are a sequence of the estimated 3D joints coordinates \\(\\{\\mathbf{J}_i \\in \\mathbb{R}^{J \\times 3} | i = 0,1,\\dots,I\\}\\), where \\(\\mathbf{J}_I\\) is our final estimation. The \\(N\\) points are firstly input to a local encoder that extracts local geometric features. Since the input point set is irregular and orderless, we exploit furthest_point_sampling method and a one-layer set_conv_layer [16] proposed by PointNet++ [21] to construct the local encoder \\((\\mathbb{R}^{N \\times 3} \\to \\mathbb{R}^{(N/2) \\times (3 + d_{local})})\\). Then, the local features are fed into the initial state generator to initialize the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.094, + 0.885, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.289, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Figure 4. The architecture of the residual graph GRU. It first exploits a graph reasoning module to evolve hidden states. The evolved hidden states are subsequently updated by a GRU cell which takes resampled points as input. After updating, the residual between updated hidden states and previous states are output for joint refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.473, + 0.51 + ], + "angle": 0, + "content": "hidden states \\(\\mathbf{S}_0\\in \\mathbb{R}^{J\\times d_{hid}}\\) , which are further processed by an initial joint regressor to obtain the initial estimation of joints \\(\\mathbf{J}_0\\) . Afterward, a novel residual recurrent unit is unrolled \\(I\\) times to iteratively update the hidden states and joint estimations by resampling the useful local features around previously estimated joints. The number of iterations \\(I\\) is a hyper-parameter of the proposed network, and is quite flexible. Specifically, the number of training iterations \\((I_{tr})\\) can be different from the number of inference iterations \\((I_{in})\\) , as discussed in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.525, + 0.303, + 0.54 + ], + "angle": 0, + "content": "3.1. Residual Recurrent Unit" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.867 + ], + "angle": 0, + "content": "HandR\\(^2\\)N\\(^2\\) requires a recurrent unit capable of abstracting a set of input point features and capturing kinematic dependencies between joints, in order to perform iterative refinements that optimizes previous estimations. Obviously, the conventional recurrent units such as LSTM or GRU cannot satisfy these requirements, since their input and hidden state should be one single vector, which cannot accommodate the input point set. Therefore, we propose a novel component, a residual recurrent unit (RRU). The RRU consists of the following items: 1) a resampler, 2) a residual graph GRU cell, and 3) a residual regressor. The resampler samples local features around previously generated coordinates \\(\\mathbf{J}_{i-1}\\) forming the input point sets for the current \\(i\\)-th iteration. To update the hidden states with the input of point sets, we propose a novel residual graph GRU that implements its gates based on PointNet. In addition, the proposed GRU also cooperates with graph reasoning to augment the hidden states by capturing kinematic dependencies. At last, the residuals between the current and previous states are sent to the residual regressor to refine the previously generated coordinates." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Resampler. The resampler collects \\( K \\)-nearest-neighbor local points and their corresponding local features around" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.403 + ], + "angle": 0, + "content": "each of previously estimated joint \\(\\mathbf{j}_{i - 1}\\in \\mathbf{J}_{i - 1}\\) , as visualized in Figure 1. The neighbor points are then translated to a relative coordinate system with \\(\\mathbf{j}_{i - 1}\\) as the origin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.893, + 0.51 + ], + "angle": 0, + "content": "Residual graph GRU. The architecture of the proposed residual graph GRU is visualized in Figure 4. Since many recent approaches [8, 23] suggested that a graph convolutional network (GCN) can model the beneficial dependencies among joints, the hidden states from the previous iteration are first augmented through a novel GCN, forming the evolved hidden states:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.527, + 0.892, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {i - 1} ^ {\\prime} = \\operatorname {R e L U} (\\mathbf {A} \\mathbf {S} _ {i - 1} \\mathbf {W}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{W} \\in \\mathbb{R}^{d_{hid} \\times d_{hid}}\\) is the trainable weights. Notably, \\(\\mathbf{A} \\in \\mathbb{R}^{J \\times J \\times d_{hid}}\\) is a novel channel-wise independent adjacent matrix, which means for each channel of \\(S\\), \\(s_c \\in \\mathbb{R}^{J \\times 1}\\), there is an individual adjacency matrix \\(a_c \\in \\mathbb{R}^{J \\times J}\\) for the multiplication \\(AS = [a_1 s_1, a_2 s_2, \\dots, a_{d_{hid}} s_{d_{hid}}]\\), where \\([\\cdot, \\cdot]\\) is concatenation. Afterwards, the GCN-evolved hidden states are passed through a GRU cell to aggregate new information from the sampled local features and to update the hidden states for the current \\(i\\)-th iteration. Here, the conventional GRU cell is not suitable because the input features have one additional dimension than the hidden states. Therefore, we exploit set_conv_layer [16], which is composed of a shared multi-layer perceptron and a max-pooling layer, for gate computation to implement the GRU. First, the hidden states are replicated \\(K\\) times to be concatenated with input features. The concatenated output are then fed into each gate for computation. For those gates that directly modify values of the hidden states, we implement them with set_conv_layer to squeeze one dimension. Moreover, we implemented a novel attentive reset gate with a point-wise multi-layer perceptron (MLP) that enables the GRU to select good points for the hidden state update, which is beneficial for exploring the optimal joint" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20907" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.466, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.185, + 0.473, + 0.283 + ], + "angle": 0, + "content": "Figure 5. Visualization of the output from the reset gate in the proposed GRU. The small colored points in the circle are the grouped neighbors of the red input joint \\( j_{i-1} \\), and their colors indicate the attentive intensity computed by Eq. 3. It shows that the points closer to the ground truth obtain relatively higher intensity. By leveraging clues to focus on highly-related points near the ground truth, the attentive GRU can output the accurate orange joint \\( j_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.471, + 0.368 + ], + "angle": 0, + "content": "locations, as shown in Figure 5. Thus, given the evolved hidden state \\(\\mathbf{s}_{i - 1}^{\\prime}\\in \\mathbf{S}_{i - 1}^{\\prime}\\) of a specific joint, and with the sampled local features \\(\\{\\mathbf{f}_k\\in \\mathbb{R}^{3 + d_{local}}|1\\leq k\\leq K\\}\\), the GRU generates the updated hidden state \\(\\mathbf{s}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.381, + 0.47, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} = \\sigma \\left(\\operatorname {s e t} _ {1 \\leq k \\leq K} \\operatorname {c o n v} _ {z} \\left(\\left[ \\mathbf {s} _ {i - 1} ^ {\\prime}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.417, + 0.47, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {i, k} = \\sigma \\left(M L P _ {r} \\left(\\left[ \\mathbf {s} _ {i - 1} ^ {\\prime}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.453, + 0.47, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {s}} _ {i} = \\tanh \\left(\\underset {1 \\leq k \\leq K} {\\operatorname {s e t}}. \\operatorname {c o n v} _ {h} \\left(\\left[ \\mathbf {r} _ {i, k} \\odot \\mathbf {s} ^ {\\prime} _ {i - 1}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.49, + 0.47, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {i} = \\left(1 - \\mathbf {z} _ {i}\\right) \\odot \\mathbf {s} _ {i - 1} ^ {\\prime} + \\mathbf {z} _ {i} \\odot \\tilde {\\mathbf {s}} _ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.618 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid activation function, \\(\\left[\\cdot ,\\cdot \\right]\\) is the concatenation operation, and \\(\\odot\\) is the Hadamard product. Note that, the updated hidden states encode the absolute joint locations in the latent space. However, we require relative representations to refine the previously generated joints. Therefore, we introduce a residual mechanism that outputs the relative representations of the updated information:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.628, + 0.47, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {S} _ {i} = \\mathbf {S} _ {i} - \\mathbf {S} _ {i - 1}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.744 + ], + "angle": 0, + "content": "Residual regressor. The regressor accepts residual states of the current \\( i \\)-th iteration as input to refine the previous joints generated by the \\( (i - 1) \\)-th iteration. The regressor is a linear transformation with a residual connection. Therefore, the joint regression of the current \\( i \\)-th iteration is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.754, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} _ {i} = \\Delta \\mathbf {S} _ {i} \\mathbf {W} + \\mathbf {J} _ {i - 1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.778, + 0.469, + 0.794 + ], + "angle": 0, + "content": "where \\(\\mathbf{W} \\in \\mathbb{R}^{d_{hid} \\times 3}\\) is the trainable transformation matrix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.469, + 0.833 + ], + "angle": 0, + "content": "3.2. Initialization of the Hidden States and Joint Coordinates" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The hidden states \\(\\mathbf{S}\\) of the RRU essentially are the embeddings that represent joints in the \\(d_{hid}\\)-dimensional latent space and can be further re-projected into the three-dimensional coordinates by the subsequent transformation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Thus, we inherit the idea of generating joint-wise embeddings to obtain the initialized hidden states \\(\\mathbf{S}_0\\), as proposed by HandFoldingNet [4]. The generation of the initial hidden states (joint-wise embeddings) are sequentially proceeded by two cascaded set_conv_layers and a three-layer bias-induced layer (BIL) [3]. Equivalent with HandFoldingNet [4], the cascaded set_conv_layers first down-sample and extract \\(N/8\\) of local features and subsequently aggregate them as a global feature. The global feature is then replicated \\(J\\) times and fed into the BILs for the generation of hidden states for \\(J\\) joints. Finally, the \\(J\\) hidden states are injected into the regressor, which performs a linear transformation \\((\\mathbb{R}^{d_{hid}} \\to \\mathbb{R}^3)\\) for the initialization of the initial joint \\(\\mathbf{J}_0\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.313, + 0.719, + 0.329 + ], + "angle": 0, + "content": "3.3. Training Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.893, + 0.38 + ], + "angle": 0, + "content": "Following the previous works [25, 4], we adopt smooth L1 loss to supervise training because of its less sensitivity to the outliers. The smooth L1 loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.389, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\nL 1 _ {\\text {s m o o t h}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 5 0 \\mathbf {x} ^ {2}, & | \\mathbf {x} | < 0. 0 1 \\\\ | \\mathbf {x} | - 0. 0 0 5, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Since the proposed network outputs an initial estimation and \\(I\\) iterative estimations, we supervise all of the outputs by the following joint loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.496, + 0.892, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 0} ^ {I} \\sum_ {j = 0} ^ {J} \\lambda^ {I - i} L 1 _ {\\text {s m o o t h}} \\left(\\mathbf {J} _ {i j} - \\mathbf {J} _ {j} ^ {*}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.597 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}^*\\) indicates the ground-truth coordinates and \\(\\lambda\\) is the hyperparameter that weights each distinct iteration. In our experiments, we set \\(\\lambda = 0.8\\) by default." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.609, + 0.634, + 0.627 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.694, + 0.651 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.84 + ], + "angle": 0, + "content": "We conducted experiments on an NVIDIA A100 GPU with PyTorch. For training, we used the AdamW optimizer [17] with beta1 = 0.5, beta2 = 0.999, and learning rate \\(\\alpha\\) = 0.001. The number of input points to the network was sampled to 1,024 and the batch size was set to 32. Batch normalization [14] was applied only in the local encoder and initializer. Meanwhile, to avoid overfitting, we adopted online data augmentation with random rotation ([-120.0, 120.0] degrees around z-axis, [-30.0, 30.0] degrees around x- and y-axis), 3D scaling ([0.8, 1.2]), and 3D translation ([-20, 20] mm). We trained the model for 160 epochs with a learning rate decay of 0.1 after 120 epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.787, + 0.863 + ], + "angle": 0, + "content": "4.2. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "MSRA Dataset. The MSRA dataset [27] provides more than 76K frames captured by the Intel's Creative Interactive" + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20908" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.089, + 0.803, + 0.442 + ], + "angle": 0, + "content": "
MethodsMean joint error (mm)Input# Params# FLOPsTime (ms)
ICVLMSRANYU
Ren-4x6x6 [13]7.63-13.392D image---
Ren-9x6x6 [13]7.319.712.692D image---
Pose-Ren [1]6.798.6511.812D image---
DenseReg [32]7.37.210.22D image5.8M--
CrossInfoNet [6]6.737.8610.082D image23.8M--
JGR-P2O [8]6.027.558.292D image1.4M-2.0 + 7.0
SSRN-SD [24]6.047.167.472D image16.2M1.52G0.4 + 3.0
SSRN [24]6.017.057.372D image22.5M1.96G0.4 + 3.4
PHG [23]5.976.947.392D image35.7M27.02G0.4 + 15.8
3DCNN [11]-9.614.13D voxels104.9M-2.9 + 1.68
SHPR-Net [2]7.227.7610.783D points---
HandPointNet [9]6.948.510.543D points2.5M1.72G8.2 + 11.3
Point-to-Point [12]6.37.79.103D points4.3M2.98G8.2 + 15.7
V2V [18]6.287.598.423D voxels34.1M6.23G23 + 5.5
HandFolding [4]5.957.348.583D points1.3M1.31G8.2 + 3.7
Deng et al. [5]6.057.006.843D points--0.9 + 47.6
Ours\\( I_{in} \\)=16.516.938.003D points1.3M0.72G0.9 + 4.2
\\( I_{in} \\)=26.126.497.503D points1.3M1.06G0.9 + 6.0
\\( I_{in} \\)=35.866.427.363D points1.3M1.40G0.9 + 7.9
\\( I_{in} \\)=45.706.487.293D points1.3M1.74G0.9 + 9.7
\\( I_{in} \\)=55.716.547.273D points1.3M2.08G0.9 + 11.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.452, + 0.895, + 0.496 + ], + "angle": 0, + "content": "Table 1. Comparison of the proposed method with previous state-of-the-art methods on the ICVL, MSRA and NYU datasets. Input indicates the input type of 2D depth image, 3D voxels or 3D point cloud. # Params indicates the number of model parameters. Time stands for the total computation time including preprocessing time and model inference time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Gesture Camera. Each frame is annotated with \\( J = 21 \\) joints, including one joint for the wrist and four joints for each finger. We followed the leave-one-subject-out cross-validation strategy as used in the previous works [4, 27] to evaluate this dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.702 + ], + "angle": 0, + "content": "ICVL Dataset. The ICVL dataset [28] provides 22K training and 1.6K testing depth frames captured by the Intel's Creative Interactive Gesture Camera. Each frame is annotated with \\( J = 16 \\) joints, including one joint for the palm and three joints for each finger. We cropped the hand area from a depth image using the method given in [19] because the human body area is present in the frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.809 + ], + "angle": 0, + "content": "NYU Dataset. The NYU dataset is captured from three different views. Each view contains 72K training 8K testing depth images captured with the PrimeSense 3D sensor. Following recent works [4, 12, 9], we selected one view and 14 joints out of total of 36 annotated joints for training and testing [31]. We also followed the same hand area cropping process [19] as in the ICVL dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. We employed two commonly used metrics, the mean joint error and the success rate, to evaluate the performance of hand pose estimation. The mean joint error quantifies the average Euclidean distance between the estimated joint locations and ground-truth ones for all the joints over the entire testing set. The success rate" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.552 + ], + "angle": 0, + "content": "indicates the percentage of the good frames with a mean joint error of less than a given distance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.561, + 0.869, + 0.577 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.734 + ], + "angle": 0, + "content": "We compared \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) with other state-of-the-art methods, including methods with 2D depth images as input: region ensemble network (Ren-4x6x6 [13], Ren-9x6x6 [32]), Pose-Ren [1], dense regression network (DenseReg) [32], CrossInfoNet [6], JGR-P2O [8], spatial-aware stacked regression network (SSRN [24], SSRN-SD [24]) and pose-guided hierarchical graph network (PHG) [23], and methods with 3D point cloud or voxels as input: 3DCNN [11], SHPR-Net [2], HandPointNet [9], Point-to-Point [12], V2V [18], HandFoldingNet [4] and Deng et al. [5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 1 summarizes the performance in terms of the mean joint error. Note that, the evaluations of our model were trained with \\( I_{tr} = 3 \\) and tested with \\( I_{in} = 4 \\), \\( I_{in} = 3 \\), \\( I_{in} = 5 \\) on ICVL, MSRA, and NYU, respectively. The results show that \\( \\mathrm{HandR}^2\\mathrm{N}^2 \\) achieves the new state-of-the-art performance with mean distance errors of \\( 5.70\\mathrm{mm} \\), \\( 6.42\\mathrm{mm} \\) and \\( 7.27\\mathrm{mm} \\) on two challenging datasets, ICVL and MSRA, respectively. It also achieves the second-lowest error on the NYU dataset. In particular, \\( \\mathrm{HandR}^2\\mathrm{N}^2 \\) significantly outperforms 3D input based previous state-of-the-arts with large margins. Meanwhile, \\( \\mathrm{HandR}^2\\mathrm{N}^2 \\) requires" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20909" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.088, + 0.347, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.089, + 0.616, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.089, + 0.887, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.893, + 0.272 + ], + "angle": 0, + "content": "Figure 6. Comparison with the state-of-the-art methods using the ICVL (left), MSRA (middle) and NYU (right) dataset. The success rate is shown in this figure." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.285, + 0.893, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) on the ICVL (left), MSRA (middle) and NYU (right) dataset. Hand depth images are transformed into 3D points as shown in the figure. Ground truth is shown in black, and the estimated joint coordinates are shown in red." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.451, + 0.47, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.702 + ], + "angle": 0, + "content": "Figure 8. Training \\( I_{tr} \\) vs. inference \\( I_{in} \\). For each column, the model are trained based on a fixed iteration number and evaluated at different inference \\( I_{in} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "only 1.3M parameters, which is generally less compared to other recent state-of-the-art methods. Figure 6 represents that our method significantly outperforms other methods in terms of success rate when the error threshold is lower than \\(20\\mathrm{mm}\\), \\(15\\mathrm{mm}\\) and \\(52\\mathrm{mm}\\) on the ICVL, MSRA and NYU datasets, respectively. Figure 2 also illustrates the superior efficiency of \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) with dynamically adjustable computation of \\(0.72 + 0.34\\times I\\) GFLOPs. The results show that our model achieves the lowest estimation error with a similar operation count at \\(I_{in} = 3\\) compared to other state-ofthe-art methods. Furthermore, the model can dynamically" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.448, + 0.892, + 0.478 + ], + "angle": 0, + "content": "run more iterations with modestly increased computation for lower estimation error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.655, + 0.502 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.569 + ], + "angle": 0, + "content": "We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the NYU dataset, since it is challenging and far from saturation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Number of iterations. Although the number of iterations \\( I \\) is a specific number during training, it can be quite flexible and even larger for inference. We first trained the model using different numbers of training \\( I_{tr} \\) to find the optimal setting of iterations for the training phase. Then we test the trained models with a various inference \\( I_{in} \\). The results are plotted on Figure 8. Interestingly, when the training \\( I_{tr} \\geq 3 \\), our model keeps improving for a few more iterations even though the inference \\( I_{in} \\) is larger than the training. In particular, when setting \\( I_{tr} = 3 \\) during training, the model achieves the minimal estimation error of 7.27 mm after 5 inference iterations. Please note that, in small number of training iterations \\( (I_{tr} < 3) \\), the model is optimized solely for the same inference iterations \\( (I_{in} = I_{tr}) \\). On the other hand, larger \\( I_{tr} \\) allows the model to generalize the regularity of changes between all iteration, which benefits not only trained iterations \\( (I_{in} = I_{tr}) \\) but also untrained larger iterations \\( (I_{in} > I_{tr}) \\). Nonetheless, Figure 8 further demonstrates the reduced performance of our model when \\( I_{tr} \\geq 5 \\) during training. It is because the lengthy path of back-propagation restricts the model from obtaining appropriate estimations. The longer the iteration path is, the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "20910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.092, + 0.892, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Figure 9. Qualitative visualization of the models with training \\( I_{tr} = 3 \\) evaluated at different inference \\( I_{in} \\) based on the NYU dataset. Mean error indicates the mean joint error. Ground truth is shown in black, and the estimated joint coordinates are shown in red." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.473, + 0.549 + ], + "angle": 0, + "content": "more different the estimations of the 0-th iteration and the \\(I\\)-th iteration are. Thus, we assume that the accumulated gradients of all iterations to the shared parameters would smooth the right descent gradients for each individual iteration, and thus degrade the final accuracy. However, these smooth accumulated gradients within a specific range of iteration count (\\(3 < I_{tr} < 5\\) according to Figure 8) allow the model to learn the general optimization direction of joints between iterations. Based on the results, we can empirically find the optimal \\(I_{tr}\\) as 3 for high inference performance and training efficiency." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.566, + 0.465, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Figure 10. Affect of sampling \\( K \\). For each line, the inference \\( I_{in} \\)s are same, but with increased \\( K \\) of resampled points for each iteration. All models are trained based on training \\( I_{tr} = 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.47, + 0.884 + ], + "angle": 0, + "content": "In addition, as Table 1 reports, the estimation accuracy is improved with a small number of additional computations (0.34GFLOPs per iteration) and time consumption (1.8 ms per iteration)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Number of resampled points. The resampled points pro" + }, + { + "type": "table", + "bbox": [ + 0.586, + 0.379, + 0.808, + 0.482 + ], + "angle": 0, + "content": "
GRUGCNRes.Mean error
×××8.68 mm
✓××8.06 mm
×✓✓7.76 mm
✓✓×7.62 mm
✓×✓7.99 mm
✓✓✓7.36 mm
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.547 + ], + "angle": 0, + "content": "Table 2. Ablations of different GRU configurations. GCN and Res. indicate whether the GCN and residual mechanism are activated in the RRU, respectively. All the ablation models are trained and tested with both \\( I_{tr} = I_{in} = 3 \\)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.563, + 0.879, + 0.667 + ], + "angle": 0, + "content": "
ConfigurationsMean error
Conventional GCN + Res.8.14 mm
Channel-wise GCN + Res.7.76 mm
Conventional GRU + Res.8.59 mm
Attentive GRU + Res.7.99 mm
Conventional GRU + GCN + Res.8.07 mm
Attentive GRU + Channel-wise GCN + Res.7.36 mm
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.677, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Table 3. Comparison with conventional components. All models are trained and tested with both \\( I_{tr} = I_{in} = 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.884 + ], + "angle": 0, + "content": "vide \\( K \\) new candidates for location refinement around each joint. Thus, the number \\( K \\) is a significant hyperparameter that affects the performance. Figure 10 presents the performance under different \\( K \\) values and inference iterations. The result shows that larger \\( K \\) reduces the estimation error since it provides a wider range of observation. However, the performance is saturated at \\( K = 48 \\), because additionally sampled points beyond the effective region of a particular joint (i.e., points that are distant to the joint and closer to other joints) do not contribute to the estimation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Analysis of different GRU configurations. To augment" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "20911" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "the final estimation, we combine the novel GRU with the GCN that evolves the hidden states and residual mechanism for relative representation, as described in Sec. 3. To verify their contributions, we incrementally adopted the GRU cell, the GCN and residual mechanism. Thus, we conducted five ablations as follows: 1) no GRU, 2) GRU; 3) GCN with residual; 4) GRU with GCN; 5) GRU with residual; 6) GRU with both residual and GRU, which is our full configuration. Note that, the ablations without residual, 2) and 4), directly transform the hidden states into absolute joint coordinates. As shown in Table 2, the GRU cell, the GCN and residual mechanism all contribute considerably to the effective estimation. In particular, the novel GRU improves the accuracy by more than \\(0.4\\mathrm{mm}\\). Furthermore, using the novel GCN and the residual mechanism further reduces the estimation error by more than \\(0.4\\mathrm{mm}\\) and \\(0.2\\mathrm{mm}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.446 + ], + "angle": 0, + "content": "Comparison with the conventional components. To validate the effectiveness of our proposed components, we replace the proposed channel-wise GCN and attentive GRU with the conventional ones, respectively. The implementation details are provided in the supplementary material. Table 3 indicates our proposed GCN and GRU both significantly outperforms the conventional implementations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.196, + 0.489 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.473, + 0.685 + ], + "angle": 0, + "content": "This paper presented \\(\\mathrm{HandR}^2\\mathrm{N}^2\\), a novel recurrent architecture that is capable of iteratively regressing progressively accurate 3D hand pose from a single point cloud. Experimental results showed that our network significantly outperforms previous state-of-the-art methods on three challenging datasets. Extensive experiments also demonstrated the excellent flexibility and efficiency of our network. As limitations, the performance of \\(\\mathrm{HandR}^2\\mathrm{N}^2\\) may degrade when hand parts are severely occluded, because the RRU can not sample efficient points. Extending this work to temporal-sequential learning may help to alleviate these limitations, which we leave for future studies." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.72, + 0.241, + 0.736 + ], + "angle": 0, + "content": "Acknowledgement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "This work was supported by the Ministry of Science and ICT (MSIT) of Korea under the National Research Foundation (NRF) grant (2022R1A4A3032913) and the Institute of Information and Communication Technology Planning Evaluation (IITP) grants (IITP-2019-0-00421, IITP-2023-2020-0-01821, IITP-2021-0-02052, IITP-2021-0-02068), and by the Technology Innovation Program (RS-2023-00235718) funded by the Ministry of Trade, Industry & Energy (1415187474). Wencan Cheng was partly supported by the China Scholarship Council (CSC)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.091, + 0.597, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.116, + 0.892, + 0.171 + ], + "angle": 0, + "content": "[1] Xinghao Chen, Guijin Wang, Hengkai Guo, and Cairong Zhang. Pose guided structured region ensemble network for cascaded hand pose estimation. Neurocomputing, 395:138-149, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.173, + 0.894, + 0.228 + ], + "angle": 0, + "content": "[2] Xinghao Chen, Guijin Wang, Cairong Zhang, Tae-Kyun Kim, and Xiangyang Ji. Shpr-net: Deep semantic hand pose regression from point clouds. IEEE Access, 6:43425-43439, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.231, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[3] Wencan Cheng and Sukhan Lee. Point auto-encoder and its application to 2d-3d transformation. In International Symposium on Visual Computing, pages 66-78. Springer, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.288, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[4] Wencan Cheng, Jae Hyun Park, and Jong Hwan Ko. Handfoldingnet: A 3d hand pose estimation network using multiscale-feature guided folding of a 2d hand skeleton. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11260-11269, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.36, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[5] Xiaoming Deng, Dexin Zuo, Yinda Zhang, Zhaopeng Cui, Jian Cheng, Ping Tan, Liang Chang, Marc Pollefeys, Sean Fanello, and Hongan Wang. Recurrent 3d hand pose estimation using cascaded pose-guided 3d alignments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.445, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[6] Kuo Du, Xiangbo Lin, Yi Sun, and Xiaohong Ma. Crossinfonet: Multi-task information sharing based hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9896-9905, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.517, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[7] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.574, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[8] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In European Conference on Computer Vision, pages 120-137. Springer, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.645, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[9] Liuhao Ge, Yujun Cai, Junwu Weng, and Junsong Yuan. Hand pointnet: 3d hand pose estimation using point sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8417-8426, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[10] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.774, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[11] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1991-2000, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[12] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 475-491, 2018. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.116, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "20912" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[13] Hengkai Guo, Guijin Wang, Xinghao Chen, Cairong Zhang, Fei Qiao, and Huazhong Yang. Region ensemble network: Improving convolutional network for hand pose estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 4512-4516. IEEE, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[15] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11927-11936, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[16] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 529-537, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.47, + 0.439 + ], + "angle": 0, + "content": "[18] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[19] Markus Oberweger and Vincent Lepetit. Deeprior++: Improving fast and accurate 3d hand pose estimation. In Proceedings of the IEEE international conference on computer vision Workshops, pages 585-594, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[20] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[21] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[22] Siamak Ravanbakhsh, Jeff Schneider, and Barnabas Poczos. Deep learning with sets and point clouds. arXiv preprint arXiv:1611.04500, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[23] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[24] Pengfei Ren, Haifeng Sun, Weiting Huang, Jiachang Hao, Daixuan Cheng, Qi Qi, Jingyu Wang, and Jianxin Liao. Spatial-aware stacked regression network for real-time 3d hand pose estimation. Neurocomputing, 437:42-57, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[25] Pengfei Ren, Haifeng Sun, Qi Qi, Jingyu Wang, and Weiting Huang. Srn: Stacked regression network for real-time 3d hand pose estimation. In BMVC, page 112, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[26] Oindrila Saha, Aditya Kusupati, Harsha Vardhan Simhadri, Manik Varma, and Prateek Jain. Rnnpool: efficient nonlinear pooling for ram constrained inference. Advances in Neural Information Processing Systems, 33:20473-20484, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[27] Xiao Sun, Yichen Wei, Shuang Liang, Xiaou Tang, and Jian Sun. Cascaded hand pose regression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 824-832, 2015. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[28] Danhang Tang, Hyung Jin Chang, Alykhan Tejani, and Tae Kyun Kim. Latent regression forest: Structured estimation of 3d articulated hand posture. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3786-3793, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[29] Jonathan Taylor, Lucas Bordeaux, Thomas Cashman, Bob Corish, Cem Keskin, Toby Sharp, Eduardo Soto, David Sweeney, Julien Valentin, Benjamin Luff, et al. Efficient and precise interactive hand tracking through joint, continuous optimization of pose and correspondences. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[31] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[32] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[33] Yiming Wu, Wei Ji, Xi Li, Gang Wang, Jianwei Yin, and Fei Wu. Context-aware deep spatiotemporal network for hand pose estimation from depth images. IEEE transactions on cybernetics, 50(2):787-797, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[34] John Yang, Yash Bhalgat, Simyung Chang, Fatih Porikli, and Nojun Kwak. Dynamic iterative refinement for efficient 3d hand pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1869-1879, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[35] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[36] Cheol-Hwan Yoo, Seowon Ji, Yong-Goo Shin, Seung-Wook Kim, and Sung-Jea Ko. Fast and accurate 3d hand pose estimation via recurrent neural network for capturing hand articulations. IEEE Access, 8:114010-114019, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "20913" + } + ] +] \ No newline at end of file diff --git a/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_origin.pdf b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b901eb7bdc99b7df41b940466e9a2dba1df40019 --- /dev/null +++ b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/431e0555-f342-4159-8204-85714e05b7c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e011a927ac156154ecdb13edf7a07474ea6d677091edbf4b65c22c8bc84f453 +size 2832085 diff --git a/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/full.md b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..490db0dd6406bd2447128e292bc2fd8841c3450a --- /dev/null +++ b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/full.md @@ -0,0 +1,284 @@ +# HandR $^2$ N $^2$ : Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network + +Wencan Cheng +Department of Artificial Intelligence +Sungkyunkwan University +cwc1260@skku.edu + +Jong Hwan Ko +College of Information and Communication Engineering +Sungkyunkwan University +jhko@skku.edu + +# Abstract + +3D hand pose estimation is a critical task in various human-computer interaction applications. Numerous deep learning based estimation models in this domain have been actively explored. However, the existing models follow a non-recurrent scheme and thus require complex architectures or redundant parameters in order to achieve acceptable model capacity. To tackle this limitation, this paper proposes HandR²N², a compact neural network that iteratively regresses the hand pose using a novel residual recurrent unit. The recurrent design allows recursive exploitation of partial layers to gradually optimize previously estimated joint locations. In addition, we exploit graph reasoning to capture kinematic dependencies between joints for better performance. Experimental results show that the proposed model significantly outperforms the existing methods on three hand pose benchmark datasets in terms of both accuracy and efficiency. Codes and pre-trained models are publicly available at https://github.com/cwc1260/HandR2N2. + +# 1. Introduction + +3D hand pose estimation, which estimates the 3D positions of hand keypoints, provides a fundamental understanding of human hand activity. Thus, it is a critical task for many human-computer interaction applications such as robotics and augmented/virtual reality. With the assistance of deep learning techniques and low-cost depth cameras, numerous studies have contributed to the substantial progress in the field of hand pose estimation. However, accurate and efficient hand pose estimation is still challenging because of the various probability of hand orientations, severe self-occlusion, and noisy depth images [12, 9, 18, 6]. + +Most of the recent state-of-the-art approaches are based on deep learning, especially deep convolutional neural networks (CNNs). Such CNN-based approaches [31, 10, 13, + +![](images/966f3a8f0700242debbdc03197897f3888e741ecfe95d51ea22cbc21e951565e.jpg) +Figure 1. Illustration of the recurrent concept. For each iteration, our proposed recurrent architecture samples local points around joints from previous iteration to generate displacements to refine the joint locations. + +25, 1, 8, 23, 6, 23] generally accept 2D depth images as an input, which can be directly processed by the 2D convolutional layers. However, it is difficult for a 2D CNN to capture the highly non-linear mapping from a 2D depth image to the 3D hand pose, significantly restricting its efficiency and accuracy. Therefore, a line of works [11, 18] discretized 2D input images into 3D volumetric representations and subsequently applied a 3D CNN for direct 3D-to-3D inference. However, a critical drawback of a 3D CNN is its computational overhead that increases cubically with an increase of the input resolution [4]. In contrast, another series of hand pose estimation models [9, 12, 2, 15, 4] have presented remarkable performances by using PointNet [20, 21], which allows the models to accept a set of continuous 3D coordinates converted from an input hand image without discretization. + +Despite the accuracy improvement achieved by the aforementioned approaches, they still suffer from large and complex model architectures that require substantial computational overheads. Thus, these methods are not suitable for resource-limited devices such as wearable or hand-held devices. Furthermore, all the existing approaches follow a non-recurrent architecture. As a result, they are not flexible for varying resource constraints and accuracy targets. + +In this work, we tackle these limitations using a recurrent architecture whose capacity can be dynamically adjusted to achieve either higher accuracy or higher efficiency by iteratively using a part of model parameters. + +The recurrent structure can achieve the high capacity of deeper networks by the recursive use of fewer shared parameters [34]. Various image processing models that used a recurrent neural network (RNN) as an explicit iterative approach have already yielded encouraging results [30, 5, 35, 26]. To fully take advantage of the RNN in a hand pose estimation task, we propose a novel residual recurrent neural network, $\mathrm{HandR}^2\mathrm{N}^2$ , that iteratively regresses accurate 3D hand pose from an input hand point cloud. Intuitively, $\mathrm{HandR}^2\mathrm{N}^2$ can be served as a general optimizer that recursively searches the optimal joint locations using the observation captured by the previous estimation, as shown in Figure 1. + +To implement $\mathrm{HandR}^2\mathrm{N}^2$ , we introduce two novel components: 1) a residual recurrent unit (RRU) and 2) an initialization module. The RRU is a key component that iteratively proceeds for updating its joint-wise hidden states and coordinate estimations. The initial joint-wise states and coordinates of the RRU are globally estimated by the initialization module. + +The RRU specifically integrates with an attentive gate that can explore optimal points from a group of candidate points for the improved accuracy. Since the proposed RRU only computes the features contributed to the specific joints, it is more efficient than the similar prior model in [5], which re-computes the entire point features. Moreover, RRU exploits graph reasoning, which is based on a novel channelwise graph convolutional network (GCN). GCN is capable to model the graph-structured data, making it particularly suitable for the kinematically structured hand data [8]. The use of the proposed channel-wise GCN enables the RRU to recursively enhance the hidden states by capturing the strong kinematic dependencies between joints. More importantly, the number of RRU iterations during inference can be dynamically scaled to be different from the training iteration. Specifically, the RRU inference iterations can be configured large to enhance the model capacity and accuracy, or small to reduce the computation requirement. Therefore, $\mathrm{HandR}^2\mathrm{N}^2$ is capable of adapting varying complexity and accuracy requirements as shown in Figure 2. + +We evaluate $\mathrm{HandR}^2\mathrm{N}^2$ on three challenging benchmarks, ICVL [28], MSRA [27], and NYU [31] datasets. The results show that our network achieves a new state-of-the-art record. The proposed network reports the minimum mean distance errors of $5.70\mathrm{mm}$ , $6.42\mathrm{mm}$ and $7.27\mathrm{mm}$ on the ICVL, MSRA and NYU datasets, respectively. Meanwhile, it requires only $1.3\mathrm{M}$ parameters and runs with adjustable complexity $0.72 + 0.34\times I$ GFLOPs, where $I$ indicates the number of iterations. Thus, our model requires + +![](images/3650eb4fdde82c3e05b5a99971650dd62f4a037445756c6e7b9510df436fc33d.jpg) +Figure 2. Comparison between operation count and estimation error. $I$ is the iteration number of $\mathrm{HandR}^2\mathrm{N}^2$ during inference. + +less computation for the highest accuracy compared with other state-of-the-art methods, as shown in Figure 2. + +The key contributions of this paper are summarised as follows: + +- We propose a novel iterative neural network architecture that takes the hand point cloud as input and progressively estimates the accurate 3D hand joint coordinates. +- We propose a novel residual recurrent unit that captures new local observations around previously generated coordinates to refine joint-wise hidden states and coordinate estimation. It also leverages a graph neural network to model dependency between the joints for performance improvement. +- We conduct extensive experiments to analyze the efficiency and effectiveness of our proposed network. + +# 2. Related Work + +# 2.1. Depth-based 3D Hand Pose Estimation + +3D hand pose estimation approaches using depth images can be classified into three categories according to the input types: conventional 2DCNN-based methods that take raw 2D depth images as input [31, 10, 13, 25, 1, 8, 23, 6], 3DCNN-based methods that use 3D voxelized representations of depth images [11, 18], and PointNet-based methods that process the point cloud transformed from depth images [9, 12, 2, 15, 4]. In this section, we focus on PointNet-based methods, since the point cloud is an accurate and concise representation for efficient inference. + +The point cloud can be easily acquired from depth images by multiplying the camera intrinsic matrix. However, due to its irregularity and disorder, processing point clouds has always been a challenging task until the emergence of PointNet [20, 21]. HandPointNet [9] was the first work that utilized PointNet as a backbone to directly regress the hand pose from the point cloud. Subsequently, the Point-to-Point model [12] used the hierarchical PointNet in order to generate the point-wise probability distribution for improved estimation. Based on these prior studies, SHPR-Net + +![](images/f8fc3533b19a45443def540ecf3c0851163f6aa2c7ea0b2dca8e6fb70b5dd678.jpg) +Figure 3. The HandR $^2$ N $^2$ pipeline. HandR $^2$ N $^2$ takes the normalized point cloud transformed from a 2D depth image as the input. The PointNet-based local encoder extracts local features from input points. The local features are fed into an initial states generator to initialize hidden states of the residual recurrent unit (RRU). The high-dimensional initialized hidden states are also transformed into 3-dimensional space to generate initial joint estimation. At last, the RRU iteratively resamples points around previously estimation joints to update hidden states and joint locations. + +[2] combined HandPointNet with an auxiliary semantic segmentation sub-network to augment the performance. Point-to-Pose [15] further exploited residual permutation equivariant layers [22] to alleviate information loss caused by the max-pooling bottleneck. Recently, HandFoldingNet [4] introduced multiscale-feature-guided folding to reshape a predefined 2D hand skeleton into hand poses. + +However, the existing methods mentioned above followed the single feed-forward design. In such feed-forward architecture, the model capacity can be increased by having a complex architecture or additional parameters. However, such solutions may result in overfitting and higher computational requirement. Unlike these methods, our model exploits the recurrent mechanism to iteratively refine accurate hand pose by reusing a part of parameters. + +# 2.2. Recurrent Models for Hand Pose Estimation + +Recently, recurrent models [7, 35, 33, 34, 36] have been successfully applied in 2D image based hand pose estimation. ACE-Net [7] and SeqHand-Net [35] exploited long short term memory (LSTM) to extract the temporal context information for more smooth and accurate estimations from color images. CADSTN [33] also applied LSTM to jointly model the spatiotemporal context from depth image sequences. However, as the input to these methods is a temporal image sequence, they cannot be used for hand pose estimation from a single image. In contrast, DIR-Net [34] iteratively reuses a part of its structure to refine a downsampled feature map of a single color image. Nonetheless, updating the whole feature map is computationally redundant, since the features without contribution, such as background, are also repeatedly computed. Thus, HCRNN [36] applies RNN on each individual joints for high inference efficiency. + +Although the recurrent models have made significant progress in 2D image based hand pose estimation, they have + +not been actively studied in the point cloud based hand pose estimation. Taylor et. al. [29] proposed an online optimization method that iteratively optimizes a predefined hand template to fit hand points. A related research [5] iteratively repeated the entire alignment architecture to align hand pose from coarse global stages to fine local stages. It also employed LSTM among multiple global stages to refine global alignment. However, it is significantly computational heavy because it reuses the entire structure for multistage feature generation. Conversely, our network only captures a small part of local information around each joint for each iteration with partial parameters, making it more effective than re-computing global information. Furthermore, we integrate a graph reasoning module in the recurrent unit to enhance estimation by capturing kinematic dependencies between joints, which has not been adopted in [5]. + +# 3. Iterative 3D Hand Pose Estimation + +HandR $^2$ N $^2$ aims to perform iterative hand pose estimation based on a recurrent architecture, as shown in Figure 3. Intuitively, HandR $^2$ N $^2$ refines the joint locations by iteratively searching the optimal position from the region around each joints, as illustrated in Figure 1. The input to HandR $^2$ N $^2$ is a set of preprocessed 3D point coordinates $\mathbf{P} \in \mathbb{R}^{N \times 3}$ that represents the hand shape, which is transformed from a 2D depth image though the camera intrinsic matrix. The outputs are a sequence of the estimated 3D joints coordinates $\{\mathbf{J}_i \in \mathbb{R}^{J \times 3} | i = 0,1,\dots,I\}$ , where $\mathbf{J}_I$ is our final estimation. The $N$ points are firstly input to a local encoder that extracts local geometric features. Since the input point set is irregular and orderless, we exploit furthest_point_sampling method and a one-layer set_conv_layer [16] proposed by PointNet++ [21] to construct the local encoder $(\mathbb{R}^{N \times 3} \to \mathbb{R}^{(N/2) \times (3 + d_{local})})$ . Then, the local features are fed into the initial state generator to initialize the + +![](images/025a2f6904a874ae98dac763b3a201af0ca48274fa0c12a1cd78485f4a7ddbbc.jpg) +Figure 4. The architecture of the residual graph GRU. It first exploits a graph reasoning module to evolve hidden states. The evolved hidden states are subsequently updated by a GRU cell which takes resampled points as input. After updating, the residual between updated hidden states and previous states are output for joint refinement. + +hidden states $\mathbf{S}_0\in \mathbb{R}^{J\times d_{hid}}$ , which are further processed by an initial joint regressor to obtain the initial estimation of joints $\mathbf{J}_0$ . Afterward, a novel residual recurrent unit is unrolled $I$ times to iteratively update the hidden states and joint estimations by resampling the useful local features around previously estimated joints. The number of iterations $I$ is a hyper-parameter of the proposed network, and is quite flexible. Specifically, the number of training iterations $(I_{tr})$ can be different from the number of inference iterations $(I_{in})$ , as discussed in Section 4.4. + +# 3.1. Residual Recurrent Unit + +HandR $^2$ N $^2$ requires a recurrent unit capable of abstracting a set of input point features and capturing kinematic dependencies between joints, in order to perform iterative refinements that optimizes previous estimations. Obviously, the conventional recurrent units such as LSTM or GRU cannot satisfy these requirements, since their input and hidden state should be one single vector, which cannot accommodate the input point set. Therefore, we propose a novel component, a residual recurrent unit (RRU). The RRU consists of the following items: 1) a resampler, 2) a residual graph GRU cell, and 3) a residual regressor. The resampler samples local features around previously generated coordinates $\mathbf{J}_{i-1}$ forming the input point sets for the current $i$ -th iteration. To update the hidden states with the input of point sets, we propose a novel residual graph GRU that implements its gates based on PointNet. In addition, the proposed GRU also cooperates with graph reasoning to augment the hidden states by capturing kinematic dependencies. At last, the residuals between the current and previous states are sent to the residual regressor to refine the previously generated coordinates. + +Resampler. The resampler collects $K$ -nearest-neighbor local points and their corresponding local features around + +each of previously estimated joint $\mathbf{j}_{i - 1}\in \mathbf{J}_{i - 1}$ , as visualized in Figure 1. The neighbor points are then translated to a relative coordinate system with $\mathbf{j}_{i - 1}$ as the origin. + +Residual graph GRU. The architecture of the proposed residual graph GRU is visualized in Figure 4. Since many recent approaches [8, 23] suggested that a graph convolutional network (GCN) can model the beneficial dependencies among joints, the hidden states from the previous iteration are first augmented through a novel GCN, forming the evolved hidden states: + +$$ +\mathbf {S} _ {i - 1} ^ {\prime} = \operatorname {R e L U} (\mathbf {A} \mathbf {S} _ {i - 1} \mathbf {W}), \tag {1} +$$ + +where $\mathbf{W} \in \mathbb{R}^{d_{hid} \times d_{hid}}$ is the trainable weights. Notably, $\mathbf{A} \in \mathbb{R}^{J \times J \times d_{hid}}$ is a novel channel-wise independent adjacent matrix, which means for each channel of $S$ , $s_c \in \mathbb{R}^{J \times 1}$ , there is an individual adjacency matrix $a_c \in \mathbb{R}^{J \times J}$ for the multiplication $AS = [a_1 s_1, a_2 s_2, \dots, a_{d_{hid}} s_{d_{hid}}]$ , where $[\cdot, \cdot]$ is concatenation. Afterwards, the GCN-evolved hidden states are passed through a GRU cell to aggregate new information from the sampled local features and to update the hidden states for the current $i$ -th iteration. Here, the conventional GRU cell is not suitable because the input features have one additional dimension than the hidden states. Therefore, we exploit set_conv_layer [16], which is composed of a shared multi-layer perceptron and a max-pooling layer, for gate computation to implement the GRU. First, the hidden states are replicated $K$ times to be concatenated with input features. The concatenated output are then fed into each gate for computation. For those gates that directly modify values of the hidden states, we implement them with set_conv_layer to squeeze one dimension. Moreover, we implemented a novel attentive reset gate with a point-wise multi-layer perceptron (MLP) that enables the GRU to select good points for the hidden state update, which is beneficial for exploring the optimal joint + +![](images/78c2d33e5af06e8f2d204e0d83e343c13c96f5476b7e36190baa5ad370bc0981.jpg) +Figure 5. Visualization of the output from the reset gate in the proposed GRU. The small colored points in the circle are the grouped neighbors of the red input joint $j_{i-1}$ , and their colors indicate the attentive intensity computed by Eq. 3. It shows that the points closer to the ground truth obtain relatively higher intensity. By leveraging clues to focus on highly-related points near the ground truth, the attentive GRU can output the accurate orange joint $j_i$ . + +locations, as shown in Figure 5. Thus, given the evolved hidden state $\mathbf{s}_{i - 1}^{\prime}\in \mathbf{S}_{i - 1}^{\prime}$ of a specific joint, and with the sampled local features $\{\mathbf{f}_k\in \mathbb{R}^{3 + d_{local}}|1\leq k\leq K\}$ , the GRU generates the updated hidden state $\mathbf{s}_i$ : + +$$ +\mathbf {z} _ {i} = \sigma \left(\operatorname {s e t} _ {1 \leq k \leq K} \operatorname {c o n v} _ {z} \left(\left[ \mathbf {s} _ {i - 1} ^ {\prime}, \mathbf {f} _ {k} \right]\right)\right), \tag {2} +$$ + +$$ +\mathbf {r} _ {i, k} = \sigma \left(M L P _ {r} \left(\left[ \mathbf {s} _ {i - 1} ^ {\prime}, \mathbf {f} _ {k} \right]\right)\right), \tag {3} +$$ + +$$ +\tilde {\mathbf {s}} _ {i} = \tanh \left(\underset {1 \leq k \leq K} {\operatorname {s e t}}. \operatorname {c o n v} _ {h} \left(\left[ \mathbf {r} _ {i, k} \odot \mathbf {s} ^ {\prime} _ {i - 1}, \mathbf {f} _ {k} \right]\right)\right), \tag {4} +$$ + +$$ +\mathbf {s} _ {i} = \left(1 - \mathbf {z} _ {i}\right) \odot \mathbf {s} _ {i - 1} ^ {\prime} + \mathbf {z} _ {i} \odot \tilde {\mathbf {s}} _ {i}, \tag {5} +$$ + +where $\sigma$ is the sigmoid activation function, $\left[\cdot ,\cdot \right]$ is the concatenation operation, and $\odot$ is the Hadamard product. Note that, the updated hidden states encode the absolute joint locations in the latent space. However, we require relative representations to refine the previously generated joints. Therefore, we introduce a residual mechanism that outputs the relative representations of the updated information: + +$$ +\Delta \mathbf {S} _ {i} = \mathbf {S} _ {i} - \mathbf {S} _ {i - 1}. \tag {6} +$$ + +Residual regressor. The regressor accepts residual states of the current $i$ -th iteration as input to refine the previous joints generated by the $(i - 1)$ -th iteration. The regressor is a linear transformation with a residual connection. Therefore, the joint regression of the current $i$ -th iteration is represented as: + +$$ +\mathbf {J} _ {i} = \Delta \mathbf {S} _ {i} \mathbf {W} + \mathbf {J} _ {i - 1}, \tag {7} +$$ + +where $\mathbf{W} \in \mathbb{R}^{d_{hid} \times 3}$ is the trainable transformation matrix. + +# 3.2. Initialization of the Hidden States and Joint Coordinates + +The hidden states $\mathbf{S}$ of the RRU essentially are the embeddings that represent joints in the $d_{hid}$ -dimensional latent space and can be further re-projected into the three-dimensional coordinates by the subsequent transformation. + +Thus, we inherit the idea of generating joint-wise embeddings to obtain the initialized hidden states $\mathbf{S}_0$ , as proposed by HandFoldingNet [4]. The generation of the initial hidden states (joint-wise embeddings) are sequentially proceeded by two cascaded set_conv_layers and a three-layer bias-induced layer (BIL) [3]. Equivalent with HandFoldingNet [4], the cascaded set_conv_layers first down-sample and extract $N/8$ of local features and subsequently aggregate them as a global feature. The global feature is then replicated $J$ times and fed into the BILs for the generation of hidden states for $J$ joints. Finally, the $J$ hidden states are injected into the regressor, which performs a linear transformation $(\mathbb{R}^{d_{hid}} \to \mathbb{R}^3)$ for the initialization of the initial joint $\mathbf{J}_0$ . + +# 3.3. Training Loss Function + +Following the previous works [25, 4], we adopt smooth L1 loss to supervise training because of its less sensitivity to the outliers. The smooth L1 loss is defined as: + +$$ +L 1 _ {\text {s m o o t h}} (\mathbf {x}) = \left\{ \begin{array}{l l} 5 0 \mathbf {x} ^ {2}, & | \mathbf {x} | < 0. 0 1 \\ | \mathbf {x} | - 0. 0 0 5, & \text {o t h e r w i s e} \end{array} . \right. \tag {8} +$$ + +Since the proposed network outputs an initial estimation and $I$ iterative estimations, we supervise all of the outputs by the following joint loss function: + +$$ +\mathcal {L} = \sum_ {i = 0} ^ {I} \sum_ {j = 0} ^ {J} \lambda^ {I - i} L 1 _ {\text {s m o o t h}} \left(\mathbf {J} _ {i j} - \mathbf {J} _ {j} ^ {*}\right), \tag {9} +$$ + +where $\mathbf{J}^*$ indicates the ground-truth coordinates and $\lambda$ is the hyperparameter that weights each distinct iteration. In our experiments, we set $\lambda = 0.8$ by default. + +# 4. Experiments + +# 4.1. Experiment Settings + +We conducted experiments on an NVIDIA A100 GPU with PyTorch. For training, we used the AdamW optimizer [17] with beta1 = 0.5, beta2 = 0.999, and learning rate $\alpha$ = 0.001. The number of input points to the network was sampled to 1,024 and the batch size was set to 32. Batch normalization [14] was applied only in the local encoder and initializer. Meanwhile, to avoid overfitting, we adopted online data augmentation with random rotation ([-120.0, 120.0] degrees around z-axis, [-30.0, 30.0] degrees around x- and y-axis), 3D scaling ([0.8, 1.2]), and 3D translation ([-20, 20] mm). We trained the model for 160 epochs with a learning rate decay of 0.1 after 120 epochs. + +# 4.2. Datasets and Evaluation Metrics + +MSRA Dataset. The MSRA dataset [27] provides more than 76K frames captured by the Intel's Creative Interactive + +
MethodsMean joint error (mm)Input# Params# FLOPsTime (ms)
ICVLMSRANYU
Ren-4x6x6 [13]7.63-13.392D image---
Ren-9x6x6 [13]7.319.712.692D image---
Pose-Ren [1]6.798.6511.812D image---
DenseReg [32]7.37.210.22D image5.8M--
CrossInfoNet [6]6.737.8610.082D image23.8M--
JGR-P2O [8]6.027.558.292D image1.4M-2.0 + 7.0
SSRN-SD [24]6.047.167.472D image16.2M1.52G0.4 + 3.0
SSRN [24]6.017.057.372D image22.5M1.96G0.4 + 3.4
PHG [23]5.976.947.392D image35.7M27.02G0.4 + 15.8
3DCNN [11]-9.614.13D voxels104.9M-2.9 + 1.68
SHPR-Net [2]7.227.7610.783D points---
HandPointNet [9]6.948.510.543D points2.5M1.72G8.2 + 11.3
Point-to-Point [12]6.37.79.103D points4.3M2.98G8.2 + 15.7
V2V [18]6.287.598.423D voxels34.1M6.23G23 + 5.5
HandFolding [4]5.957.348.583D points1.3M1.31G8.2 + 3.7
Deng et al. [5]6.057.006.843D points--0.9 + 47.6
Ours\( I_{in} \)=16.516.938.003D points1.3M0.72G0.9 + 4.2
\( I_{in} \)=26.126.497.503D points1.3M1.06G0.9 + 6.0
\( I_{in} \)=35.866.427.363D points1.3M1.40G0.9 + 7.9
\( I_{in} \)=45.706.487.293D points1.3M1.74G0.9 + 9.7
\( I_{in} \)=55.716.547.273D points1.3M2.08G0.9 + 11.5
+ +Table 1. Comparison of the proposed method with previous state-of-the-art methods on the ICVL, MSRA and NYU datasets. Input indicates the input type of 2D depth image, 3D voxels or 3D point cloud. # Params indicates the number of model parameters. Time stands for the total computation time including preprocessing time and model inference time. + +Gesture Camera. Each frame is annotated with $J = 21$ joints, including one joint for the wrist and four joints for each finger. We followed the leave-one-subject-out cross-validation strategy as used in the previous works [4, 27] to evaluate this dataset. + +ICVL Dataset. The ICVL dataset [28] provides 22K training and 1.6K testing depth frames captured by the Intel's Creative Interactive Gesture Camera. Each frame is annotated with $J = 16$ joints, including one joint for the palm and three joints for each finger. We cropped the hand area from a depth image using the method given in [19] because the human body area is present in the frames. + +NYU Dataset. The NYU dataset is captured from three different views. Each view contains 72K training 8K testing depth images captured with the PrimeSense 3D sensor. Following recent works [4, 12, 9], we selected one view and 14 joints out of total of 36 annotated joints for training and testing [31]. We also followed the same hand area cropping process [19] as in the ICVL dataset. + +Evaluation metrics. We employed two commonly used metrics, the mean joint error and the success rate, to evaluate the performance of hand pose estimation. The mean joint error quantifies the average Euclidean distance between the estimated joint locations and ground-truth ones for all the joints over the entire testing set. The success rate + +indicates the percentage of the good frames with a mean joint error of less than a given distance. + +# 4.3. Comparison with State-of-the-Art Methods + +We compared $\mathrm{HandR}^2\mathrm{N}^2$ with other state-of-the-art methods, including methods with 2D depth images as input: region ensemble network (Ren-4x6x6 [13], Ren-9x6x6 [32]), Pose-Ren [1], dense regression network (DenseReg) [32], CrossInfoNet [6], JGR-P2O [8], spatial-aware stacked regression network (SSRN [24], SSRN-SD [24]) and pose-guided hierarchical graph network (PHG) [23], and methods with 3D point cloud or voxels as input: 3DCNN [11], SHPR-Net [2], HandPointNet [9], Point-to-Point [12], V2V [18], HandFoldingNet [4] and Deng et al. [5]. + +Table 1 summarizes the performance in terms of the mean joint error. Note that, the evaluations of our model were trained with $I_{tr} = 3$ and tested with $I_{in} = 4$ , $I_{in} = 3$ , $I_{in} = 5$ on ICVL, MSRA, and NYU, respectively. The results show that $\mathrm{HandR}^2\mathrm{N}^2$ achieves the new state-of-the-art performance with mean distance errors of $5.70\mathrm{mm}$ , $6.42\mathrm{mm}$ and $7.27\mathrm{mm}$ on two challenging datasets, ICVL and MSRA, respectively. It also achieves the second-lowest error on the NYU dataset. In particular, $\mathrm{HandR}^2\mathrm{N}^2$ significantly outperforms 3D input based previous state-of-the-arts with large margins. Meanwhile, $\mathrm{HandR}^2\mathrm{N}^2$ requires + +![](images/fd5f6082666ad6ffcef578531e9e8cbe04f7ddad6f5f6df1a8488095c39e5f0b.jpg) +Figure 6. Comparison with the state-of-the-art methods using the ICVL (left), MSRA (middle) and NYU (right) dataset. The success rate is shown in this figure. + +![](images/1d3247c7fa7e86fb0b5c14500016158c66072ffb99d53bcdaa613b5d352bee22.jpg) + +![](images/e478b198f5035d1b37604efd8be370bfc88dacc21857933be203d084ccaf7e30.jpg) + +![](images/f2d6f376e96c92406598b81793797c40c4d294c27cacdce5e47cdb79f98ff666.jpg) +Figure 7. Qualitative results of $\mathrm{HandR}^2\mathrm{N}^2$ on the ICVL (left), MSRA (middle) and NYU (right) dataset. Hand depth images are transformed into 3D points as shown in the figure. Ground truth is shown in black, and the estimated joint coordinates are shown in red. + +![](images/6e130e013ed2c2cd5695c6d4ae3229a7370fb7a296a75268273d7a0a69317e28.jpg) +Figure 8. Training $I_{tr}$ vs. inference $I_{in}$ . For each column, the model are trained based on a fixed iteration number and evaluated at different inference $I_{in}$ . + +only 1.3M parameters, which is generally less compared to other recent state-of-the-art methods. Figure 6 represents that our method significantly outperforms other methods in terms of success rate when the error threshold is lower than $20\mathrm{mm}$ , $15\mathrm{mm}$ and $52\mathrm{mm}$ on the ICVL, MSRA and NYU datasets, respectively. Figure 2 also illustrates the superior efficiency of $\mathrm{HandR}^2\mathrm{N}^2$ with dynamically adjustable computation of $0.72 + 0.34\times I$ GFLOPs. The results show that our model achieves the lowest estimation error with a similar operation count at $I_{in} = 3$ compared to other state-ofthe-art methods. Furthermore, the model can dynamically + +run more iterations with modestly increased computation for lower estimation error. + +# 4.4. Ablation Study + +We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the NYU dataset, since it is challenging and far from saturation. + +Number of iterations. Although the number of iterations $I$ is a specific number during training, it can be quite flexible and even larger for inference. We first trained the model using different numbers of training $I_{tr}$ to find the optimal setting of iterations for the training phase. Then we test the trained models with a various inference $I_{in}$ . The results are plotted on Figure 8. Interestingly, when the training $I_{tr} \geq 3$ , our model keeps improving for a few more iterations even though the inference $I_{in}$ is larger than the training. In particular, when setting $I_{tr} = 3$ during training, the model achieves the minimal estimation error of 7.27 mm after 5 inference iterations. Please note that, in small number of training iterations $(I_{tr} < 3)$ , the model is optimized solely for the same inference iterations $(I_{in} = I_{tr})$ . On the other hand, larger $I_{tr}$ allows the model to generalize the regularity of changes between all iteration, which benefits not only trained iterations $(I_{in} = I_{tr})$ but also untrained larger iterations $(I_{in} > I_{tr})$ . Nonetheless, Figure 8 further demonstrates the reduced performance of our model when $I_{tr} \geq 5$ during training. It is because the lengthy path of back-propagation restricts the model from obtaining appropriate estimations. The longer the iteration path is, the + +![](images/6207f72d8c106f2c90288afdb11b25930f22b2b140a2f40f7dfcbe86beb72878.jpg) +Figure 9. Qualitative visualization of the models with training $I_{tr} = 3$ evaluated at different inference $I_{in}$ based on the NYU dataset. Mean error indicates the mean joint error. Ground truth is shown in black, and the estimated joint coordinates are shown in red. + +more different the estimations of the 0-th iteration and the $I$ -th iteration are. Thus, we assume that the accumulated gradients of all iterations to the shared parameters would smooth the right descent gradients for each individual iteration, and thus degrade the final accuracy. However, these smooth accumulated gradients within a specific range of iteration count ( $3 < I_{tr} < 5$ according to Figure 8) allow the model to learn the general optimization direction of joints between iterations. Based on the results, we can empirically find the optimal $I_{tr}$ as 3 for high inference performance and training efficiency. + +![](images/e13db47d14bbe1f1d88103de84d7c0f41881d48137934c18126c1586c581efbc.jpg) +Figure 10. Affect of sampling $K$ . For each line, the inference $I_{in}$ s are same, but with increased $K$ of resampled points for each iteration. All models are trained based on training $I_{tr} = 3$ . + +In addition, as Table 1 reports, the estimation accuracy is improved with a small number of additional computations (0.34GFLOPs per iteration) and time consumption (1.8 ms per iteration). + +Number of resampled points. The resampled points pro + +
GRUGCNRes.Mean error
×××8.68 mm
✓××8.06 mm
×✓✓7.76 mm
✓✓×7.62 mm
✓×✓7.99 mm
✓✓✓7.36 mm
+ +Table 2. Ablations of different GRU configurations. GCN and Res. indicate whether the GCN and residual mechanism are activated in the RRU, respectively. All the ablation models are trained and tested with both $I_{tr} = I_{in} = 3$ . + +
ConfigurationsMean error
Conventional GCN + Res.8.14 mm
Channel-wise GCN + Res.7.76 mm
Conventional GRU + Res.8.59 mm
Attentive GRU + Res.7.99 mm
Conventional GRU + GCN + Res.8.07 mm
Attentive GRU + Channel-wise GCN + Res.7.36 mm
+ +Table 3. Comparison with conventional components. All models are trained and tested with both $I_{tr} = I_{in} = 3$ . + +vide $K$ new candidates for location refinement around each joint. Thus, the number $K$ is a significant hyperparameter that affects the performance. Figure 10 presents the performance under different $K$ values and inference iterations. The result shows that larger $K$ reduces the estimation error since it provides a wider range of observation. However, the performance is saturated at $K = 48$ , because additionally sampled points beyond the effective region of a particular joint (i.e., points that are distant to the joint and closer to other joints) do not contribute to the estimation. + +Analysis of different GRU configurations. To augment + +the final estimation, we combine the novel GRU with the GCN that evolves the hidden states and residual mechanism for relative representation, as described in Sec. 3. To verify their contributions, we incrementally adopted the GRU cell, the GCN and residual mechanism. Thus, we conducted five ablations as follows: 1) no GRU, 2) GRU; 3) GCN with residual; 4) GRU with GCN; 5) GRU with residual; 6) GRU with both residual and GRU, which is our full configuration. Note that, the ablations without residual, 2) and 4), directly transform the hidden states into absolute joint coordinates. As shown in Table 2, the GRU cell, the GCN and residual mechanism all contribute considerably to the effective estimation. In particular, the novel GRU improves the accuracy by more than $0.4\mathrm{mm}$ . Furthermore, using the novel GCN and the residual mechanism further reduces the estimation error by more than $0.4\mathrm{mm}$ and $0.2\mathrm{mm}$ , respectively. + +Comparison with the conventional components. To validate the effectiveness of our proposed components, we replace the proposed channel-wise GCN and attentive GRU with the conventional ones, respectively. The implementation details are provided in the supplementary material. Table 3 indicates our proposed GCN and GRU both significantly outperforms the conventional implementations. + +# 5. Conclusion + +This paper presented $\mathrm{HandR}^2\mathrm{N}^2$ , a novel recurrent architecture that is capable of iteratively regressing progressively accurate 3D hand pose from a single point cloud. Experimental results showed that our network significantly outperforms previous state-of-the-art methods on three challenging datasets. Extensive experiments also demonstrated the excellent flexibility and efficiency of our network. As limitations, the performance of $\mathrm{HandR}^2\mathrm{N}^2$ may degrade when hand parts are severely occluded, because the RRU can not sample efficient points. Extending this work to temporal-sequential learning may help to alleviate these limitations, which we leave for future studies. + +# Acknowledgement. + +This work was supported by the Ministry of Science and ICT (MSIT) of Korea under the National Research Foundation (NRF) grant (2022R1A4A3032913) and the Institute of Information and Communication Technology Planning Evaluation (IITP) grants (IITP-2019-0-00421, IITP-2023-2020-0-01821, IITP-2021-0-02052, IITP-2021-0-02068), and by the Technology Innovation Program (RS-2023-00235718) funded by the Ministry of Trade, Industry & Energy (1415187474). Wencan Cheng was partly supported by the China Scholarship Council (CSC). + +# References + +[1] Xinghao Chen, Guijin Wang, Hengkai Guo, and Cairong Zhang. Pose guided structured region ensemble network for cascaded hand pose estimation. Neurocomputing, 395:138-149, 2020. 1, 2, 6 +[2] Xinghao Chen, Guijin Wang, Cairong Zhang, Tae-Kyun Kim, and Xiangyang Ji. Shpr-net: Deep semantic hand pose regression from point clouds. IEEE Access, 6:43425-43439, 2018. 1, 2, 3, 6 +[3] Wencan Cheng and Sukhan Lee. Point auto-encoder and its application to 2d-3d transformation. In International Symposium on Visual Computing, pages 66-78. Springer, 2019. 5 +[4] Wencan Cheng, Jae Hyun Park, and Jong Hwan Ko. Handfoldingnet: A 3d hand pose estimation network using multiscale-feature guided folding of a 2d hand skeleton. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11260-11269, 2021. 1, 2, 3, 5, 6 +[5] Xiaoming Deng, Dexin Zuo, Yinda Zhang, Zhaopeng Cui, Jian Cheng, Ping Tan, Liang Chang, Marc Pollefeys, Sean Fanello, and Hongan Wang. Recurrent 3d hand pose estimation using cascaded pose-guided 3d alignments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3, 6 +[6] Kuo Du, Xiangbo Lin, Yi Sun, and Xiaohong Ma. Crossinfonet: Multi-task information sharing based hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9896-9905, 2019. 1, 2, 6 +[7] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 3 +[8] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In European Conference on Computer Vision, pages 120-137. Springer, 2020. 1, 2, 4, 6 +[9] Liuhao Ge, Yujun Cai, Junwu Weng, and Junsong Yuan. Hand pointnet: 3d hand pose estimation using point sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8417-8426, 2018. 1, 2, 6 +[10] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 1, 2 +[11] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1991-2000, 2017. 1, 2, 6 +[12] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 475-491, 2018. 1, 2, 6 + +[13] Hengkai Guo, Guijin Wang, Xinghao Chen, Cairong Zhang, Fei Qiao, and Huazhong Yang. Region ensemble network: Improving convolutional network for hand pose estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 4512-4516. IEEE, 2017. 1, 2, 6 +[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5 +[15] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11927-11936, 2019. 1, 2, 3 +[16] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 529-537, 2019. 3, 4 +[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[18] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 1, 2, 6 +[19] Markus Oberweger and Vincent Lepetit. Deeprior++: Improving fast and accurate 3d hand pose estimation. In Proceedings of the IEEE international conference on computer vision Workshops, pages 585-594, 2017. 6 +[20] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2 +[21] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 1, 2, 3 +[22] Siamak Ravanbakhsh, Jeff Schneider, and Barnabas Poczos. Deep learning with sets and point clouds. arXiv preprint arXiv:1611.04500, 2016. 3 +[23] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 2021. 1, 2, 4, 6 +[24] Pengfei Ren, Haifeng Sun, Weiting Huang, Jiachang Hao, Daixuan Cheng, Qi Qi, Jingyu Wang, and Jianxin Liao. Spatial-aware stacked regression network for real-time 3d hand pose estimation. Neurocomputing, 437:42-57, 2021. 6 +[25] Pengfei Ren, Haifeng Sun, Qi Qi, Jingyu Wang, and Weiting Huang. Srn: Stacked regression network for real-time 3d hand pose estimation. In BMVC, page 112, 2019. 1, 2, 5 +[26] Oindrila Saha, Aditya Kusupati, Harsha Vardhan Simhadri, Manik Varma, and Prateek Jain. Rnnpool: efficient nonlinear pooling for ram constrained inference. Advances in Neural Information Processing Systems, 33:20473-20484, 2020. 2 + +[27] Xiao Sun, Yichen Wei, Shuang Liang, Xiaou Tang, and Jian Sun. Cascaded hand pose regression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 824-832, 2015. 2, 5, 6 +[28] Danhang Tang, Hyung Jin Chang, Alykhan Tejani, and Tae Kyun Kim. Latent regression forest: Structured estimation of 3d articulated hand posture. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3786-3793, 2014. 2, 6 +[29] Jonathan Taylor, Lucas Bordeaux, Thomas Cashman, Bob Corish, Cem Keskin, Toby Sharp, Eduardo Soto, David Sweeney, Julien Valentin, Benjamin Luff, et al. Efficient and precise interactive hand tracking through joint, continuous optimization of pose and correspondences. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3 +[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2 +[31] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014. 1, 2, 6 +[32] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2018. 6 +[33] Yiming Wu, Wei Ji, Xi Li, Gang Wang, Jianwei Yin, and Fei Wu. Context-aware deep spatiotemporal network for hand pose estimation from depth images. IEEE transactions on cybernetics, 50(2):787-797, 2018. 3 +[34] John Yang, Yash Bhalgat, Simyung Chang, Fatih Porikli, and Nojun Kwak. Dynamic iterative refinement for efficient 3d hand pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1869-1879, 2022. 2, 3 +[35] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 2, 3 +[36] Cheol-Hwan Yoo, Seowon Ji, Yong-Goo Shin, Seung-Wook Kim, and Sung-Jea Ko. Fast and accurate 3d hand pose estimation via recurrent neural network for capturing hand articulations. IEEE Access, 8:114010-114019, 2020. 3 \ No newline at end of file diff --git a/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/images.zip b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d58672700b7709cbf149faa5df27d2c47e20593 --- /dev/null +++ b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3e656f0000bab986b6ad8c99a7cff0b6bdb1e889e9ede77770b9b36f7dac1fe +size 772254 diff --git a/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/layout.json b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..daefa5841ed601f828637270fd3f2d5abfb01ffe --- /dev/null +++ b/2023/HandR2N2_ Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network/layout.json @@ -0,0 +1,8489 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "text", + "content": "HandR" + }, + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "text", + "content": "N" + }, + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "text", + "content": ": Iterative 3D Hand Pose Estimation Using a Residual Recurrent Neural Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 162, + 239, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 239, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 239, + 217 + ], + "type": "text", + "content": "Wencan Cheng \nDepartment of Artificial Intelligence \nSungkyunkwan University \ncwc1260@skku.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 259, + 162, + 535, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 162, + 535, + 217 + ], + "spans": [ + { + "bbox": [ + 259, + 162, + 535, + 217 + ], + "type": "text", + "content": "Jong Hwan Ko \nCollege of Information and Communication Engineering \nSungkyunkwan University \njhko@skku.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 270, + 290, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 498 + ], + "type": "text", + "content": "3D hand pose estimation is a critical task in various human-computer interaction applications. Numerous deep learning based estimation models in this domain have been actively explored. However, the existing models follow a non-recurrent scheme and thus require complex architectures or redundant parameters in order to achieve acceptable model capacity. To tackle this limitation, this paper proposes HandR²N², a compact neural network that iteratively regresses the hand pose using a novel residual recurrent unit. The recurrent design allows recursive exploitation of partial layers to gradually optimize previously estimated joint locations. In addition, we exploit graph reasoning to capture kinematic dependencies between joints for better performance. Experimental results show that the proposed model significantly outperforms the existing methods on three hand pose benchmark datasets in terms of both accuracy and efficiency. Codes and pre-trained models are publicly available at https://github.com/cwc1260/HandR2N2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 677 + ], + "type": "text", + "content": "3D hand pose estimation, which estimates the 3D positions of hand keypoints, provides a fundamental understanding of human hand activity. Thus, it is a critical task for many human-computer interaction applications such as robotics and augmented/virtual reality. With the assistance of deep learning techniques and low-cost depth cameras, numerous studies have contributed to the substantial progress in the field of hand pose estimation. However, accurate and efficient hand pose estimation is still challenging because of the various probability of hand orientations, severe self-occlusion, and noisy depth images [12, 9, 18, 6]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Most of the recent state-of-the-art approaches are based on deep learning, especially deep convolutional neural networks (CNNs). Such CNN-based approaches [31, 10, 13," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 334, + 243, + 529, + 352 + ], + "blocks": [ + { + "bbox": [ + 334, + 243, + 529, + 352 + ], + "lines": [ + { + "bbox": [ + 334, + 243, + 529, + 352 + ], + "spans": [ + { + "bbox": [ + 334, + 243, + 529, + 352 + ], + "type": "image", + "image_path": "966f3a8f0700242debbdc03197897f3888e741ecfe95d51ea22cbc21e951565e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 355, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 399 + ], + "type": "text", + "content": "Figure 1. Illustration of the recurrent concept. For each iteration, our proposed recurrent architecture samples local points around joints from previous iteration to generate displacements to refine the joint locations." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 424, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 616 + ], + "type": "text", + "content": "25, 1, 8, 23, 6, 23] generally accept 2D depth images as an input, which can be directly processed by the 2D convolutional layers. However, it is difficult for a 2D CNN to capture the highly non-linear mapping from a 2D depth image to the 3D hand pose, significantly restricting its efficiency and accuracy. Therefore, a line of works [11, 18] discretized 2D input images into 3D volumetric representations and subsequently applied a 3D CNN for direct 3D-to-3D inference. However, a critical drawback of a 3D CNN is its computational overhead that increases cubically with an increase of the input resolution [4]. In contrast, another series of hand pose estimation models [9, 12, 2, 15, 4] have presented remarkable performances by using PointNet [20, 21], which allows the models to accept a set of continuous 3D coordinates converted from an input hand image without discretization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 715 + ], + "type": "text", + "content": "Despite the accuracy improvement achieved by the aforementioned approaches, they still suffer from large and complex model architectures that require substantial computational overheads. Thus, these methods are not suitable for resource-limited devices such as wearable or hand-held devices. Furthermore, all the existing approaches follow a non-recurrent architecture. As a result, they are not flexible for varying resource constraints and accuracy targets." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20904" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "In this work, we tackle these limitations using a recurrent architecture whose capacity can be dynamically adjusted to achieve either higher accuracy or higher efficiency by iteratively using a part of model parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 287, + 277 + ], + "type": "text", + "content": "The recurrent structure can achieve the high capacity of deeper networks by the recursive use of fewer shared parameters [34]. Various image processing models that used a recurrent neural network (RNN) as an explicit iterative approach have already yielded encouraging results [30, 5, 35, 26]. To fully take advantage of the RNN in a hand pose estimation task, we propose a novel residual recurrent neural network, " + }, + { + "bbox": [ + 46, + 122, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 122, + 287, + 277 + ], + "type": "text", + "content": ", that iteratively regresses accurate 3D hand pose from an input hand point cloud. Intuitively, " + }, + { + "bbox": [ + 46, + 122, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 122, + 287, + 277 + ], + "type": "text", + "content": " can be served as a general optimizer that recursively searches the optimal joint locations using the observation captured by the previous estimation, as shown in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 278, + 287, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 362 + ], + "type": "text", + "content": "To implement " + }, + { + "bbox": [ + 46, + 278, + 287, + 362 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 278, + 287, + 362 + ], + "type": "text", + "content": ", we introduce two novel components: 1) a residual recurrent unit (RRU) and 2) an initialization module. The RRU is a key component that iteratively proceeds for updating its joint-wise hidden states and coordinate estimations. The initial joint-wise states and coordinates of the RRU are globally estimated by the initialization module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 604 + ], + "type": "text", + "content": "The RRU specifically integrates with an attentive gate that can explore optimal points from a group of candidate points for the improved accuracy. Since the proposed RRU only computes the features contributed to the specific joints, it is more efficient than the similar prior model in [5], which re-computes the entire point features. Moreover, RRU exploits graph reasoning, which is based on a novel channelwise graph convolutional network (GCN). GCN is capable to model the graph-structured data, making it particularly suitable for the kinematically structured hand data [8]. The use of the proposed channel-wise GCN enables the RRU to recursively enhance the hidden states by capturing the strong kinematic dependencies between joints. More importantly, the number of RRU iterations during inference can be dynamically scaled to be different from the training iteration. Specifically, the RRU inference iterations can be configured large to enhance the model capacity and accuracy, or small to reduce the computation requirement. Therefore, " + }, + { + "bbox": [ + 46, + 365, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 365, + 287, + 604 + ], + "type": "text", + "content": " is capable of adapting varying complexity and accuracy requirements as shown in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "We evaluate " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " on three challenging benchmarks, ICVL [28], MSRA [27], and NYU [31] datasets. The results show that our network achieves a new state-of-the-art record. The proposed network reports the minimum mean distance errors of " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "5.70\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "6.42\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "7.27\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " on the ICVL, MSRA and NYU datasets, respectively. Meanwhile, it requires only " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "1.3\\mathrm{M}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " parameters and runs with adjustable complexity " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "0.72 + 0.34\\times I" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " GFLOPs, where " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " indicates the number of iterations. Thus, our model requires" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 71, + 524, + 166 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 524, + 166 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 524, + 166 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 524, + 166 + ], + "type": "image", + "image_path": "3650eb4fdde82c3e05b5a99971650dd62f4a037445756c6e7b9510df436fc33d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 166, + 545, + 187 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 545, + 187 + ], + "type": "text", + "content": "Figure 2. Comparison between operation count and estimation error. " + }, + { + "bbox": [ + 305, + 166, + 545, + 187 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 166, + 545, + 187 + ], + "type": "text", + "content": " is the iteration number of " + }, + { + "bbox": [ + 305, + 166, + 545, + 187 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 305, + 166, + 545, + 187 + ], + "type": "text", + "content": " during inference." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 209, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 232 + ], + "type": "text", + "content": "less computation for the highest accuracy compared with other state-of-the-art methods, as shown in Figure 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 233, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 233, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 233, + 545, + 256 + ], + "type": "text", + "content": "The key contributions of this paper are summarised as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 265, + 545, + 425 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 265, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 545, + 312 + ], + "type": "text", + "content": "- We propose a novel iterative neural network architecture that takes the hand point cloud as input and progressively estimates the accurate 3D hand joint coordinates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 321, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 545, + 393 + ], + "type": "text", + "content": "- We propose a novel residual recurrent unit that captures new local observations around previously generated coordinates to refine joint-wise hidden states and coordinate estimation. It also leverages a graph neural network to model dependency between the joints for performance improvement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 401, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 401, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 401, + 545, + 425 + ], + "type": "text", + "content": "- We conduct extensive experiments to analyze the efficiency and effectiveness of our proposed network." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 436, + 392, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 392, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 392, + 448 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 456, + 510, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 510, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 510, + 468 + ], + "type": "text", + "content": "2.1. Depth-based 3D Hand Pose Estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 593 + ], + "type": "text", + "content": "3D hand pose estimation approaches using depth images can be classified into three categories according to the input types: conventional 2DCNN-based methods that take raw 2D depth images as input [31, 10, 13, 25, 1, 8, 23, 6], 3DCNN-based methods that use 3D voxelized representations of depth images [11, 18], and PointNet-based methods that process the point cloud transformed from depth images [9, 12, 2, 15, 4]. In this section, we focus on PointNet-based methods, since the point cloud is an accurate and concise representation for efficient inference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "The point cloud can be easily acquired from depth images by multiplying the camera intrinsic matrix. However, due to its irregularity and disorder, processing point clouds has always been a challenging task until the emergence of PointNet [20, 21]. HandPointNet [9] was the first work that utilized PointNet as a backbone to directly regress the hand pose from the point cloud. Subsequently, the Point-to-Point model [12] used the hierarchical PointNet in order to generate the point-wise probability distribution for improved estimation. Based on these prior studies, SHPR-Net" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20905" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 536, + 211 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 536, + 211 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 536, + 211 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 536, + 211 + ], + "type": "image", + "image_path": "f8fc3533b19a45443def540ecf3c0851163f6aa2c7ea0b2dca8e6fb70b5dd678.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "text", + "content": "Figure 3. The HandR" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "text", + "content": "N" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "text", + "content": " pipeline. HandR" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "text", + "content": "N" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 212, + 547, + 266 + ], + "type": "text", + "content": " takes the normalized point cloud transformed from a 2D depth image as the input. The PointNet-based local encoder extracts local features from input points. The local features are fed into an initial states generator to initialize hidden states of the residual recurrent unit (RRU). The high-dimensional initialized hidden states are also transformed into 3-dimensional space to generate initial joint estimation. At last, the RRU iteratively resamples points around previously estimation joints to update hidden states and joint locations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 287, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 370 + ], + "type": "text", + "content": "[2] combined HandPointNet with an auxiliary semantic segmentation sub-network to augment the performance. Point-to-Pose [15] further exploited residual permutation equivariant layers [22] to alleviate information loss caused by the max-pooling bottleneck. Recently, HandFoldingNet [4] introduced multiscale-feature-guided folding to reshape a predefined 2D hand skeleton into hand poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 468 + ], + "type": "text", + "content": "However, the existing methods mentioned above followed the single feed-forward design. In such feed-forward architecture, the model capacity can be increased by having a complex architecture or additional parameters. However, such solutions may result in overfitting and higher computational requirement. Unlike these methods, our model exploits the recurrent mechanism to iteratively refine accurate hand pose by reusing a part of parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 478, + 277, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 277, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 277, + 490 + ], + "type": "text", + "content": "2.2. Recurrent Models for Hand Pose Estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 497, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 688 + ], + "type": "text", + "content": "Recently, recurrent models [7, 35, 33, 34, 36] have been successfully applied in 2D image based hand pose estimation. ACE-Net [7] and SeqHand-Net [35] exploited long short term memory (LSTM) to extract the temporal context information for more smooth and accurate estimations from color images. CADSTN [33] also applied LSTM to jointly model the spatiotemporal context from depth image sequences. However, as the input to these methods is a temporal image sequence, they cannot be used for hand pose estimation from a single image. In contrast, DIR-Net [34] iteratively reuses a part of its structure to refine a downsampled feature map of a single color image. Nonetheless, updating the whole feature map is computationally redundant, since the features without contribution, such as background, are also repeatedly computed. Thus, HCRNN [36] applies RNN on each individual joints for high inference efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Although the recurrent models have made significant progress in 2D image based hand pose estimation, they have" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 287, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 479 + ], + "type": "text", + "content": "not been actively studied in the point cloud based hand pose estimation. Taylor et. al. [29] proposed an online optimization method that iteratively optimizes a predefined hand template to fit hand points. A related research [5] iteratively repeated the entire alignment architecture to align hand pose from coarse global stages to fine local stages. It also employed LSTM among multiple global stages to refine global alignment. However, it is significantly computational heavy because it reuses the entire structure for multistage feature generation. Conversely, our network only captures a small part of local information around each joint for each iteration with partial parameters, making it more effective than re-computing global information. Furthermore, we integrate a graph reasoning module in the recurrent unit to enhance estimation by capturing kinematic dependencies between joints, which has not been adopted in [5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 490, + 499, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 499, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 499, + 502 + ], + "type": "text", + "content": "3. Iterative 3D Hand Pose Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "HandR" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " aims to perform iterative hand pose estimation based on a recurrent architecture, as shown in Figure 3. Intuitively, HandR" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " refines the joint locations by iteratively searching the optimal position from the region around each joints, as illustrated in Figure 1. The input to HandR" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is a set of preprocessed 3D point coordinates " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " that represents the hand shape, which is transformed from a 2D depth image though the camera intrinsic matrix. The outputs are a sequence of the estimated 3D joints coordinates " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{J}_i \\in \\mathbb{R}^{J \\times 3} | i = 0,1,\\dots,I\\}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_I" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is our final estimation. The " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " points are firstly input to a local encoder that extracts local geometric features. Since the input point set is irregular and orderless, we exploit furthest_point_sampling method and a one-layer set_conv_layer [16] proposed by PointNet++ [21] to construct the local encoder " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{N \\times 3} \\to \\mathbb{R}^{(N/2) \\times (3 + d_{local})})" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". Then, the local features are fed into the initial state generator to initialize the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20906" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 74, + 541, + 224 + ], + "blocks": [ + { + "bbox": [ + 52, + 74, + 541, + 224 + ], + "lines": [ + { + "bbox": [ + 52, + 74, + 541, + 224 + ], + "spans": [ + { + "bbox": [ + 52, + 74, + 541, + 224 + ], + "type": "image", + "image_path": "025a2f6904a874ae98dac763b3a201af0ca48274fa0c12a1cd78485f4a7ddbbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 228, + 546, + 264 + ], + "lines": [ + { + "bbox": [ + 45, + 228, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 45, + 228, + 546, + 264 + ], + "type": "text", + "content": "Figure 4. The architecture of the residual graph GRU. It first exploits a graph reasoning module to evolve hidden states. The evolved hidden states are subsequently updated by a GRU cell which takes resampled points as input. After updating, the residual between updated hidden states and previous states are output for joint refinement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "content": "hidden states " + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0\\in \\mathbb{R}^{J\\times d_{hid}}" + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "content": " , which are further processed by an initial joint regressor to obtain the initial estimation of joints " + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_0" + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "content": " . Afterward, a novel residual recurrent unit is unrolled " + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "content": " times to iteratively update the hidden states and joint estimations by resampling the useful local features around previously estimated joints. The number of iterations " + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "content": " is a hyper-parameter of the proposed network, and is quite flexible. Specifically, the number of training iterations " + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "inline_equation", + "content": "(I_{tr})" + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "content": " can be different from the number of inference iterations " + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "inline_equation", + "content": "(I_{in})" + }, + { + "bbox": [ + 46, + 283, + 289, + 403 + ], + "type": "text", + "content": " , as discussed in Section 4.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 185, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 185, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 185, + 427 + ], + "type": "text", + "content": "3.1. Residual Recurrent Unit" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "text", + "content": "HandR" + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "text", + "content": "N" + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "text", + "content": " requires a recurrent unit capable of abstracting a set of input point features and capturing kinematic dependencies between joints, in order to perform iterative refinements that optimizes previous estimations. Obviously, the conventional recurrent units such as LSTM or GRU cannot satisfy these requirements, since their input and hidden state should be one single vector, which cannot accommodate the input point set. Therefore, we propose a novel component, a residual recurrent unit (RRU). The RRU consists of the following items: 1) a resampler, 2) a residual graph GRU cell, and 3) a residual regressor. The resampler samples local features around previously generated coordinates " + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{i-1}" + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "text", + "content": " forming the input point sets for the current " + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 436, + 287, + 686 + ], + "type": "text", + "content": "-th iteration. To update the hidden states with the input of point sets, we propose a novel residual graph GRU that implements its gates based on PointNet. In addition, the proposed GRU also cooperates with graph reasoning to augment the hidden states by capturing kinematic dependencies. At last, the residuals between the current and previous states are sent to the residual regressor to refine the previously generated coordinates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Resampler. The resampler collects " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "-nearest-neighbor local points and their corresponding local features around" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "content": "each of previously estimated joint " + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{j}_{i - 1}\\in \\mathbf{J}_{i - 1}" + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "content": " , as visualized in Figure 1. The neighbor points are then translated to a relative coordinate system with " + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{j}_{i - 1}" + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "content": " as the origin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 319, + 546, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 403 + ], + "type": "text", + "content": "Residual graph GRU. The architecture of the proposed residual graph GRU is visualized in Figure 4. Since many recent approaches [8, 23] suggested that a graph convolutional network (GCN) can model the beneficial dependencies among joints, the hidden states from the previous iteration are first augmented through a novel GCN, forming the evolved hidden states:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 366, + 417, + 545, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 417, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 366, + 417, + 545, + 431 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {i - 1} ^ {\\prime} = \\operatorname {R e L U} (\\mathbf {A} \\mathbf {S} _ {i - 1} \\mathbf {W}), \\tag {1}", + "image_path": "5788fd26577d18fe0791beb3e91680b285b7f0c2090c2c99caf8a347f34a3145.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d_{hid} \\times d_{hid}}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " is the trainable weights. Notably, " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{J \\times J \\times d_{hid}}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " is a novel channel-wise independent adjacent matrix, which means for each channel of " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "s_c \\in \\mathbb{R}^{J \\times 1}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", there is an individual adjacency matrix " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "a_c \\in \\mathbb{R}^{J \\times J}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " for the multiplication " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "AS = [a_1 s_1, a_2 s_2, \\dots, a_{d_{hid}} s_{d_{hid}}]" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " is concatenation. Afterwards, the GCN-evolved hidden states are passed through a GRU cell to aggregate new information from the sampled local features and to update the hidden states for the current " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": "-th iteration. Here, the conventional GRU cell is not suitable because the input features have one additional dimension than the hidden states. Therefore, we exploit set_conv_layer [16], which is composed of a shared multi-layer perceptron and a max-pooling layer, for gate computation to implement the GRU. First, the hidden states are replicated " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " times to be concatenated with input features. The concatenated output are then fed into each gate for computation. For those gates that directly modify values of the hidden states, we implement them with set_conv_layer to squeeze one dimension. Moreover, we implemented a novel attentive reset gate with a point-wise multi-layer perceptron (MLP) that enables the GRU to select good points for the hidden state update, which is beneficial for exploring the optimal joint" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20907" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 285, + 144 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 285, + 144 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 285, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 285, + 144 + ], + "type": "image", + "image_path": "78c2d33e5af06e8f2d204e0d83e343c13c96f5476b7e36190baa5ad370bc0981.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 146, + 289, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 289, + 224 + ], + "type": "text", + "content": "Figure 5. Visualization of the output from the reset gate in the proposed GRU. The small colored points in the circle are the grouped neighbors of the red input joint " + }, + { + "bbox": [ + 46, + 146, + 289, + 224 + ], + "type": "inline_equation", + "content": "j_{i-1}" + }, + { + "bbox": [ + 46, + 146, + 289, + 224 + ], + "type": "text", + "content": ", and their colors indicate the attentive intensity computed by Eq. 3. It shows that the points closer to the ground truth obtain relatively higher intensity. By leveraging clues to focus on highly-related points near the ground truth, the attentive GRU can output the accurate orange joint " + }, + { + "bbox": [ + 46, + 146, + 289, + 224 + ], + "type": "inline_equation", + "content": "j_i" + }, + { + "bbox": [ + 46, + 146, + 289, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "text", + "content": "locations, as shown in Figure 5. Thus, given the evolved hidden state " + }, + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i - 1}^{\\prime}\\in \\mathbf{S}_{i - 1}^{\\prime}" + }, + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "text", + "content": " of a specific joint, and with the sampled local features " + }, + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_k\\in \\mathbb{R}^{3 + d_{local}}|1\\leq k\\leq K\\}" + }, + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "text", + "content": ", the GRU generates the updated hidden state " + }, + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 46, + 242, + 288, + 291 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 301, + 287, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 301, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 101, + 301, + 287, + 322 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} = \\sigma \\left(\\operatorname {s e t} _ {1 \\leq k \\leq K} \\operatorname {c o n v} _ {z} \\left(\\left[ \\mathbf {s} _ {i - 1} ^ {\\prime}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {2}", + "image_path": "ef7230743b5c99f41e8131762e6a60687c66e55f17c234cc930b353f1f7d9006.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 330, + 287, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 330, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 287, + 345 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {i, k} = \\sigma \\left(M L P _ {r} \\left(\\left[ \\mathbf {s} _ {i - 1} ^ {\\prime}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {3}", + "image_path": "7a767ea537c4b5057bcb546d1b1509f937c77f80c4738f10eb146fdfedeec51f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 358, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 358, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 80, + 358, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {s}} _ {i} = \\tanh \\left(\\underset {1 \\leq k \\leq K} {\\operatorname {s e t}}. \\operatorname {c o n v} _ {h} \\left(\\left[ \\mathbf {r} _ {i, k} \\odot \\mathbf {s} ^ {\\prime} _ {i - 1}, \\mathbf {f} _ {k} \\right]\\right)\\right), \\tag {4}", + "image_path": "8e51ce8e52e3e7502ad3398fcc3f76960e0e060aae099528275d6877aa70883b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 388, + 287, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 388, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 99, + 388, + 287, + 401 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {i} = \\left(1 - \\mathbf {z} _ {i}\\right) \\odot \\mathbf {s} _ {i - 1} ^ {\\prime} + \\mathbf {z} _ {i} \\odot \\tilde {\\mathbf {s}} _ {i}, \\tag {5}", + "image_path": "9480874dbb39d94ff73ee3e75f95d445f4f08ffcd7cadc4c122620c900e0585e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " is the sigmoid activation function, " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\left[\\cdot ,\\cdot \\right]" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " is the concatenation operation, and " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " is the Hadamard product. Note that, the updated hidden states encode the absolute joint locations in the latent space. However, we require relative representations to refine the previously generated joints. Therefore, we introduce a residual mechanism that outputs the relative representations of the updated information:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 497, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 497, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 127, + 497, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {S} _ {i} = \\mathbf {S} _ {i} - \\mathbf {S} _ {i - 1}. \\tag {6}", + "image_path": "101f5f4402ea12f43ab562c19658676a8ba11b2b386d0a6a4743b20b00416646.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "text", + "content": "Residual regressor. The regressor accepts residual states of the current " + }, + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "text", + "content": "-th iteration as input to refine the previous joints generated by the " + }, + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "text", + "content": "-th iteration. The regressor is a linear transformation with a residual connection. Therefore, the joint regression of the current " + }, + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 517, + 287, + 589 + ], + "type": "text", + "content": "-th iteration is represented as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 121, + 597, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 597, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 121, + 597, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\mathbf {J} _ {i} = \\Delta \\mathbf {S} _ {i} \\mathbf {W} + \\mathbf {J} _ {i - 1}, \\tag {7}", + "image_path": "321148c246847f7d42fac7a516a2d84ba039b0bd005f65ac6044f4e260097f4b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 616, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 616, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d_{hid} \\times 3}" + }, + { + "bbox": [ + 47, + 616, + 287, + 628 + ], + "type": "text", + "content": " is the trainable transformation matrix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "type": "text", + "content": "3.2. Initialization of the Hidden States and Joint Coordinates" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The hidden states " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " of the RRU essentially are the embeddings that represent joints in the " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{hid}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "-dimensional latent space and can be further re-projected into the three-dimensional coordinates by the subsequent transformation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": "Thus, we inherit the idea of generating joint-wise embeddings to obtain the initialized hidden states " + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_0" + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": ", as proposed by HandFoldingNet [4]. The generation of the initial hidden states (joint-wise embeddings) are sequentially proceeded by two cascaded set_conv_layers and a three-layer bias-induced layer (BIL) [3]. Equivalent with HandFoldingNet [4], the cascaded set_conv_layers first down-sample and extract " + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "inline_equation", + "content": "N/8" + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": " of local features and subsequently aggregate them as a global feature. The global feature is then replicated " + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": " times and fed into the BILs for the generation of hidden states for " + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": " joints. Finally, the " + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": " hidden states are injected into the regressor, which performs a linear transformation " + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{d_{hid}} \\to \\mathbb{R}^3)" + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": " for the initialization of the initial joint " + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_0" + }, + { + "bbox": [ + 304, + 72, + 546, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 247, + 440, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 440, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 440, + 260 + ], + "type": "text", + "content": "3.3. Training Loss Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 265, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 546, + 300 + ], + "type": "text", + "content": "Following the previous works [25, 4], we adopt smooth L1 loss to supervise training because of its less sensitivity to the outliers. The smooth L1 loss is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 308, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 308, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 334, + 308, + 545, + 340 + ], + "type": "interline_equation", + "content": "L 1 _ {\\text {s m o o t h}} (\\mathbf {x}) = \\left\\{ \\begin{array}{l l} 5 0 \\mathbf {x} ^ {2}, & | \\mathbf {x} | < 0. 0 1 \\\\ | \\mathbf {x} | - 0. 0 0 5, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {8}", + "image_path": "6c67a2fd084dec238c29a67a06134ac742889ad76f8711b5cee564011280a37e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 347, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 384 + ], + "type": "text", + "content": "Since the proposed network outputs an initial estimation and " + }, + { + "bbox": [ + 304, + 347, + 545, + 384 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 347, + 545, + 384 + ], + "type": "text", + "content": " iterative estimations, we supervise all of the outputs by the following joint loss function:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 392, + 545, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 392, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 347, + 392, + 545, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 0} ^ {I} \\sum_ {j = 0} ^ {J} \\lambda^ {I - i} L 1 _ {\\text {s m o o t h}} \\left(\\mathbf {J} _ {i j} - \\mathbf {J} _ {j} ^ {*}\\right), \\tag {9}", + "image_path": "99bc66d0914a26f6c08e79b6d8665f6dc5501abff333eeb4af6cf15319a6e48e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{J}^*" + }, + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "text", + "content": " indicates the ground-truth coordinates and " + }, + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "text", + "content": " is the hyperparameter that weights each distinct iteration. In our experiments, we set " + }, + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\lambda = 0.8" + }, + { + "bbox": [ + 304, + 435, + 545, + 472 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 482, + 388, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 388, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 388, + 496 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 502, + 424, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 424, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 424, + 515 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 520, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 665 + ], + "type": "text", + "content": "We conducted experiments on an NVIDIA A100 GPU with PyTorch. For training, we used the AdamW optimizer [17] with beta1 = 0.5, beta2 = 0.999, and learning rate " + }, + { + "bbox": [ + 304, + 520, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 520, + 545, + 665 + ], + "type": "text", + "content": " = 0.001. The number of input points to the network was sampled to 1,024 and the batch size was set to 32. Batch normalization [14] was applied only in the local encoder and initializer. Meanwhile, to avoid overfitting, we adopted online data augmentation with random rotation ([-120.0, 120.0] degrees around z-axis, [-30.0, 30.0] degrees around x- and y-axis), 3D scaling ([0.8, 1.2]), and 3D translation ([-20, 20] mm). We trained the model for 160 epochs with a learning rate decay of 0.1 after 120 epochs." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 671, + 481, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 481, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 481, + 683 + ], + "type": "text", + "content": "4.2. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "MSRA Dataset. The MSRA dataset [27] provides more than 76K frames captured by the Intel's Creative Interactive" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20908" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 70, + 491, + 350 + ], + "blocks": [ + { + "bbox": [ + 103, + 70, + 491, + 350 + ], + "lines": [ + { + "bbox": [ + 103, + 70, + 491, + 350 + ], + "spans": [ + { + "bbox": [ + 103, + 70, + 491, + 350 + ], + "type": "table", + "html": "
MethodsMean joint error (mm)Input# Params# FLOPsTime (ms)
ICVLMSRANYU
Ren-4x6x6 [13]7.63-13.392D image---
Ren-9x6x6 [13]7.319.712.692D image---
Pose-Ren [1]6.798.6511.812D image---
DenseReg [32]7.37.210.22D image5.8M--
CrossInfoNet [6]6.737.8610.082D image23.8M--
JGR-P2O [8]6.027.558.292D image1.4M-2.0 + 7.0
SSRN-SD [24]6.047.167.472D image16.2M1.52G0.4 + 3.0
SSRN [24]6.017.057.372D image22.5M1.96G0.4 + 3.4
PHG [23]5.976.947.392D image35.7M27.02G0.4 + 15.8
3DCNN [11]-9.614.13D voxels104.9M-2.9 + 1.68
SHPR-Net [2]7.227.7610.783D points---
HandPointNet [9]6.948.510.543D points2.5M1.72G8.2 + 11.3
Point-to-Point [12]6.37.79.103D points4.3M2.98G8.2 + 15.7
V2V [18]6.287.598.423D voxels34.1M6.23G23 + 5.5
HandFolding [4]5.957.348.583D points1.3M1.31G8.2 + 3.7
Deng et al. [5]6.057.006.843D points--0.9 + 47.6
Ours\\( I_{in} \\)=16.516.938.003D points1.3M0.72G0.9 + 4.2
\\( I_{in} \\)=26.126.497.503D points1.3M1.06G0.9 + 6.0
\\( I_{in} \\)=35.866.427.363D points1.3M1.40G0.9 + 7.9
\\( I_{in} \\)=45.706.487.293D points1.3M1.74G0.9 + 9.7
\\( I_{in} \\)=55.716.547.273D points1.3M2.08G0.9 + 11.5
", + "image_path": "04e46ddfbe13258c6aa7b315328c281fc857a81deecc074f7c3facd32d27580e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 547, + 392 + ], + "type": "text", + "content": "Table 1. Comparison of the proposed method with previous state-of-the-art methods on the ICVL, MSRA and NYU datasets. Input indicates the input type of 2D depth image, 3D voxels or 3D point cloud. # Params indicates the number of model parameters. Time stands for the total computation time including preprocessing time and model inference time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 412, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 471 + ], + "type": "text", + "content": "Gesture Camera. Each frame is annotated with " + }, + { + "bbox": [ + 46, + 412, + 287, + 471 + ], + "type": "inline_equation", + "content": "J = 21" + }, + { + "bbox": [ + 46, + 412, + 287, + 471 + ], + "type": "text", + "content": " joints, including one joint for the wrist and four joints for each finger. We followed the leave-one-subject-out cross-validation strategy as used in the previous works [4, 27] to evaluate this dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "type": "text", + "content": "ICVL Dataset. The ICVL dataset [28] provides 22K training and 1.6K testing depth frames captured by the Intel's Creative Interactive Gesture Camera. Each frame is annotated with " + }, + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "type": "inline_equation", + "content": "J = 16" + }, + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "type": "text", + "content": " joints, including one joint for the palm and three joints for each finger. We cropped the hand area from a depth image using the method given in [19] because the human body area is present in the frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": "NYU Dataset. The NYU dataset is captured from three different views. Each view contains 72K training 8K testing depth images captured with the PrimeSense 3D sensor. Following recent works [4, 12, 9], we selected one view and 14 joints out of total of 36 annotated joints for training and testing [31]. We also followed the same hand area cropping process [19] as in the ICVL dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Evaluation metrics. We employed two commonly used metrics, the mean joint error and the success rate, to evaluate the performance of hand pose estimation. The mean joint error quantifies the average Euclidean distance between the estimated joint locations and ground-truth ones for all the joints over the entire testing set. The success rate" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "text", + "content": "indicates the percentage of the good frames with a mean joint error of less than a given distance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 444, + 531, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 531, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 531, + 456 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "content": "We compared " + }, + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "content": " with other state-of-the-art methods, including methods with 2D depth images as input: region ensemble network (Ren-4x6x6 [13], Ren-9x6x6 [32]), Pose-Ren [1], dense regression network (DenseReg) [32], CrossInfoNet [6], JGR-P2O [8], spatial-aware stacked regression network (SSRN [24], SSRN-SD [24]) and pose-guided hierarchical graph network (PHG) [23], and methods with 3D point cloud or voxels as input: 3DCNN [11], SHPR-Net [2], HandPointNet [9], Point-to-Point [12], V2V [18], HandFoldingNet [4] and Deng et al. [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Table 1 summarizes the performance in terms of the mean joint error. Note that, the evaluations of our model were trained with " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{tr} = 3" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and tested with " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{in} = 4" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{in} = 3" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{in} = 5" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " on ICVL, MSRA, and NYU, respectively. The results show that " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " achieves the new state-of-the-art performance with mean distance errors of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.70\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "6.42\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "7.27\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " on two challenging datasets, ICVL and MSRA, respectively. It also achieves the second-lowest error on the NYU dataset. In particular, " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " significantly outperforms 3D input based previous state-of-the-arts with large margins. Meanwhile, " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " requires" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20909" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 212, + 182 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 212, + 182 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 212, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 212, + 182 + ], + "type": "image", + "image_path": "fd5f6082666ad6ffcef578531e9e8cbe04f7ddad6f5f6df1a8488095c39e5f0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 546, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 546, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 546, + 215 + ], + "type": "text", + "content": "Figure 6. Comparison with the state-of-the-art methods using the ICVL (left), MSRA (middle) and NYU (right) dataset. The success rate is shown in this figure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 70, + 376, + 182 + ], + "blocks": [ + { + "bbox": [ + 212, + 70, + 376, + 182 + ], + "lines": [ + { + "bbox": [ + 212, + 70, + 376, + 182 + ], + "spans": [ + { + "bbox": [ + 212, + 70, + 376, + 182 + ], + "type": "image", + "image_path": "1d3247c7fa7e86fb0b5c14500016158c66072ffb99d53bcdaa613b5d352bee22.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 542, + 182 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 542, + 182 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 542, + 182 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 542, + 182 + ], + "type": "image", + "image_path": "e478b198f5035d1b37604efd8be370bfc88dacc21857933be203d084ccaf7e30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 54, + 225, + 546, + 310 + ], + "blocks": [ + { + "bbox": [ + 54, + 225, + 546, + 310 + ], + "lines": [ + { + "bbox": [ + 54, + 225, + 546, + 310 + ], + "spans": [ + { + "bbox": [ + 54, + 225, + 546, + 310 + ], + "type": "image", + "image_path": "f2d6f376e96c92406598b81793797c40c4d294c27cacdce5e47cdb79f98ff666.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 311, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 545, + 334 + ], + "type": "text", + "content": "Figure 7. Qualitative results of " + }, + { + "bbox": [ + 46, + 311, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 311, + 545, + 334 + ], + "type": "text", + "content": " on the ICVL (left), MSRA (middle) and NYU (right) dataset. Hand depth images are transformed into 3D points as shown in the figure. Ground truth is shown in black, and the estimated joint coordinates are shown in red." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 357, + 287, + 518 + ], + "blocks": [ + { + "bbox": [ + 55, + 357, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 287, + 518 + ], + "type": "image", + "image_path": "6e130e013ed2c2cd5695c6d4ae3229a7370fb7a296a75268273d7a0a69317e28.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "type": "text", + "content": "Figure 8. Training " + }, + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "type": "inline_equation", + "content": "I_{tr}" + }, + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "type": "text", + "content": " vs. inference " + }, + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "type": "text", + "content": ". For each column, the model are trained based on a fixed iteration number and evaluated at different inference " + }, + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 46, + 522, + 287, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "only 1.3M parameters, which is generally less compared to other recent state-of-the-art methods. Figure 6 represents that our method significantly outperforms other methods in terms of success rate when the error threshold is lower than " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "20\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "15\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "52\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " on the ICVL, MSRA and NYU datasets, respectively. Figure 2 also illustrates the superior efficiency of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " with dynamically adjustable computation of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "0.72 + 0.34\\times I" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " GFLOPs. The results show that our model achieves the lowest estimation error with a similar operation count at " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{in} = 3" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " compared to other state-ofthe-art methods. Furthermore, the model can dynamically" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 378 + ], + "type": "text", + "content": "run more iterations with modestly increased computation for lower estimation error." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 384, + 400, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 400, + 397 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 400, + 397 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 403, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 450 + ], + "type": "text", + "content": "We conducted ablation experiments to analyze the contribution of each component in our model. The following experiments are evaluated based on the NYU dataset, since it is challenging and far from saturation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": "Number of iterations. Although the number of iterations " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " is a specific number during training, it can be quite flexible and even larger for inference. We first trained the model using different numbers of training " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I_{tr}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " to find the optimal setting of iterations for the training phase. Then we test the trained models with a various inference " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": ". The results are plotted on Figure 8. Interestingly, when the training " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I_{tr} \\geq 3" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": ", our model keeps improving for a few more iterations even though the inference " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " is larger than the training. In particular, when setting " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I_{tr} = 3" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " during training, the model achieves the minimal estimation error of 7.27 mm after 5 inference iterations. Please note that, in small number of training iterations " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "(I_{tr} < 3)" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": ", the model is optimized solely for the same inference iterations " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "(I_{in} = I_{tr})" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": ". On the other hand, larger " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I_{tr}" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " allows the model to generalize the regularity of changes between all iteration, which benefits not only trained iterations " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "(I_{in} = I_{tr})" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " but also untrained larger iterations " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "(I_{in} > I_{tr})" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": ". Nonetheless, Figure 8 further demonstrates the reduced performance of our model when " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "I_{tr} \\geq 5" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " during training. It is because the lengthy path of back-propagation restricts the model from obtaining appropriate estimations. The longer the iteration path is, the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20910" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 545, + 257 + ], + "type": "image", + "image_path": "6207f72d8c106f2c90288afdb11b25930f22b2b140a2f40f7dfcbe86beb72878.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 545, + 281 + ], + "type": "text", + "content": "Figure 9. Qualitative visualization of the models with training " + }, + { + "bbox": [ + 46, + 259, + 545, + 281 + ], + "type": "inline_equation", + "content": "I_{tr} = 3" + }, + { + "bbox": [ + 46, + 259, + 545, + 281 + ], + "type": "text", + "content": " evaluated at different inference " + }, + { + "bbox": [ + 46, + 259, + 545, + 281 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 46, + 259, + 545, + 281 + ], + "type": "text", + "content": " based on the NYU dataset. Mean error indicates the mean joint error. Ground truth is shown in black, and the estimated joint coordinates are shown in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "text", + "content": "more different the estimations of the 0-th iteration and the " + }, + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "text", + "content": "-th iteration are. Thus, we assume that the accumulated gradients of all iterations to the shared parameters would smooth the right descent gradients for each individual iteration, and thus degrade the final accuracy. However, these smooth accumulated gradients within a specific range of iteration count (" + }, + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "inline_equation", + "content": "3 < I_{tr} < 5" + }, + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "text", + "content": " according to Figure 8) allow the model to learn the general optimization direction of joints between iterations. Based on the results, we can empirically find the optimal " + }, + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "inline_equation", + "content": "I_{tr}" + }, + { + "bbox": [ + 46, + 303, + 289, + 434 + ], + "type": "text", + "content": " as 3 for high inference performance and training efficiency." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 448, + 284, + 604 + ], + "blocks": [ + { + "bbox": [ + 53, + 448, + 284, + 604 + ], + "lines": [ + { + "bbox": [ + 53, + 448, + 284, + 604 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 284, + 604 + ], + "type": "image", + "image_path": "e13db47d14bbe1f1d88103de84d7c0f41881d48137934c18126c1586c581efbc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "text", + "content": "Figure 10. Affect of sampling " + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "text", + "content": ". For each line, the inference " + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "text", + "content": "s are same, but with increased " + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "text", + "content": " of resampled points for each iteration. All models are trained based on training " + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "inline_equation", + "content": "I_{tr} = 3" + }, + { + "bbox": [ + 46, + 604, + 287, + 638 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 652, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 287, + 700 + ], + "type": "text", + "content": "In addition, as Table 1 reports, the estimation accuracy is improved with a small number of additional computations (0.34GFLOPs per iteration) and time consumption (1.8 ms per iteration)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Number of resampled points. The resampled points pro" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 358, + 300, + 494, + 381 + ], + "blocks": [ + { + "bbox": [ + 358, + 300, + 494, + 381 + ], + "lines": [ + { + "bbox": [ + 358, + 300, + 494, + 381 + ], + "spans": [ + { + "bbox": [ + 358, + 300, + 494, + 381 + ], + "type": "table", + "html": "
GRUGCNRes.Mean error
×××8.68 mm
✓××8.06 mm
×✓✓7.76 mm
✓✓×7.62 mm
✓×✓7.99 mm
✓✓✓7.36 mm
", + "image_path": "41cf5b59ad355759155156d59ece2fade0af09fda3ff37f7132c03f897f02a9c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 445, + 537, + 528 + ], + "blocks": [ + { + "bbox": [ + 305, + 389, + 545, + 433 + ], + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 433 + ], + "type": "text", + "content": "Table 2. Ablations of different GRU configurations. GCN and Res. indicate whether the GCN and residual mechanism are activated in the RRU, respectively. All the ablation models are trained and tested with both " + }, + { + "bbox": [ + 305, + 389, + 545, + 433 + ], + "type": "inline_equation", + "content": "I_{tr} = I_{in} = 3" + }, + { + "bbox": [ + 305, + 389, + 545, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 445, + 537, + 528 + ], + "lines": [ + { + "bbox": [ + 315, + 445, + 537, + 528 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 537, + 528 + ], + "type": "table", + "html": "
ConfigurationsMean error
Conventional GCN + Res.8.14 mm
Channel-wise GCN + Res.7.76 mm
Conventional GRU + Res.8.59 mm
Attentive GRU + Res.7.99 mm
Conventional GRU + GCN + Res.8.07 mm
Attentive GRU + Channel-wise GCN + Res.7.36 mm
", + "image_path": "9ef54c32bac0ff4839187a4358f0cec17d0c8f27faf776a9442682faa2a8d427.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 536, + 545, + 557 + ], + "lines": [ + { + "bbox": [ + 306, + 536, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 545, + 557 + ], + "type": "text", + "content": "Table 3. Comparison with conventional components. All models are trained and tested with both " + }, + { + "bbox": [ + 306, + 536, + 545, + 557 + ], + "type": "inline_equation", + "content": "I_{tr} = I_{in} = 3" + }, + { + "bbox": [ + 306, + 536, + 545, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "text", + "content": "vide " + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "text", + "content": " new candidates for location refinement around each joint. Thus, the number " + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "text", + "content": " is a significant hyperparameter that affects the performance. Figure 10 presents the performance under different " + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "text", + "content": " values and inference iterations. The result shows that larger " + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "text", + "content": " reduces the estimation error since it provides a wider range of observation. However, the performance is saturated at " + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "inline_equation", + "content": "K = 48" + }, + { + "bbox": [ + 304, + 580, + 545, + 700 + ], + "type": "text", + "content": ", because additionally sampled points beyond the effective region of a particular joint (i.e., points that are distant to the joint and closer to other joints) do not contribute to the estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Analysis of different GRU configurations. To augment" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20911" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "the final estimation, we combine the novel GRU with the GCN that evolves the hidden states and residual mechanism for relative representation, as described in Sec. 3. To verify their contributions, we incrementally adopted the GRU cell, the GCN and residual mechanism. Thus, we conducted five ablations as follows: 1) no GRU, 2) GRU; 3) GCN with residual; 4) GRU with GCN; 5) GRU with residual; 6) GRU with both residual and GRU, which is our full configuration. Note that, the ablations without residual, 2) and 4), directly transform the hidden states into absolute joint coordinates. As shown in Table 2, the GRU cell, the GCN and residual mechanism all contribute considerably to the effective estimation. In particular, the novel GRU improves the accuracy by more than " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ". Furthermore, using the novel GCN and the residual mechanism further reduces the estimation error by more than " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 289, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 289, + 353 + ], + "type": "text", + "content": "Comparison with the conventional components. To validate the effectiveness of our proposed components, we replace the proposed channel-wise GCN and attentive GRU with the conventional ones, respectively. The implementation details are provided in the supplementary material. Table 3 indicates our proposed GCN and GRU both significantly outperforms the conventional implementations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 374, + 119, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 119, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 119, + 387 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 289, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 542 + ], + "type": "text", + "content": "This paper presented " + }, + { + "bbox": [ + 46, + 399, + 289, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 399, + 289, + 542 + ], + "type": "text", + "content": ", a novel recurrent architecture that is capable of iteratively regressing progressively accurate 3D hand pose from a single point cloud. Experimental results showed that our network significantly outperforms previous state-of-the-art methods on three challenging datasets. Extensive experiments also demonstrated the excellent flexibility and efficiency of our network. As limitations, the performance of " + }, + { + "bbox": [ + 46, + 399, + 289, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{HandR}^2\\mathrm{N}^2" + }, + { + "bbox": [ + 46, + 399, + 289, + 542 + ], + "type": "text", + "content": " may degrade when hand parts are severely occluded, because the RRU can not sample efficient points. Extending this work to temporal-sequential learning may help to alleviate these limitations, which we leave for future studies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 570, + 147, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 147, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 147, + 582 + ], + "type": "text", + "content": "Acknowledgement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "This work was supported by the Ministry of Science and ICT (MSIT) of Korea under the National Research Foundation (NRF) grant (2022R1A4A3032913) and the Institute of Information and Communication Technology Planning Evaluation (IITP) grants (IITP-2019-0-00421, IITP-2023-2020-0-01821, IITP-2021-0-02052, IITP-2021-0-02068), and by the Technology Innovation Program (RS-2023-00235718) funded by the Ministry of Trade, Industry & Energy (1415187474). Wencan Cheng was partly supported by the China Scholarship Council (CSC)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 91, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 311, + 91, + 545, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 545, + 135 + ], + "type": "text", + "content": "[1] Xinghao Chen, Guijin Wang, Hengkai Guo, and Cairong Zhang. Pose guided structured region ensemble network for cascaded hand pose estimation. Neurocomputing, 395:138-149, 2020. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 311, + 137, + 547, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 137, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 311, + 137, + 547, + 180 + ], + "type": "text", + "content": "[2] Xinghao Chen, Guijin Wang, Cairong Zhang, Tae-Kyun Kim, and Xiangyang Ji. Shpr-net: Deep semantic hand pose regression from point clouds. IEEE Access, 6:43425-43439, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 182, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 182, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 311, + 182, + 545, + 225 + ], + "type": "text", + "content": "[3] Wencan Cheng and Sukhan Lee. Point auto-encoder and its application to 2d-3d transformation. In International Symposium on Visual Computing, pages 66-78. Springer, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 311, + 228, + 545, + 282 + ], + "type": "text", + "content": "[4] Wencan Cheng, Jae Hyun Park, and Jong Hwan Ko. Handfoldingnet: A 3d hand pose estimation network using multiscale-feature guided folding of a 2d hand skeleton. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11260-11269, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 285, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 285, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 311, + 285, + 545, + 350 + ], + "type": "text", + "content": "[5] Xiaoming Deng, Dexin Zuo, Yinda Zhang, Zhaopeng Cui, Jian Cheng, Ping Tan, Liang Chang, Marc Pollefeys, Sean Fanello, and Hongan Wang. Recurrent 3d hand pose estimation using cascaded pose-guided 3d alignments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 352, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 352, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 311, + 352, + 545, + 407 + ], + "type": "text", + "content": "[6] Kuo Du, Xiangbo Lin, Yi Sun, and Xiaohong Ma. Crossinfonet: Multi-task information sharing based hand pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9896-9905, 2019. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 311, + 409, + 545, + 453 + ], + "type": "text", + "content": "[7] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 454, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 454, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 311, + 454, + 545, + 510 + ], + "type": "text", + "content": "[8] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In European Conference on Computer Vision, pages 120-137. Springer, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 510, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 510, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 311, + 510, + 545, + 555 + ], + "type": "text", + "content": "[9] Liuhao Ge, Yujun Cai, Junwu Weng, and Junsong Yuan. Hand pointnet: 3d hand pose estimation using point sets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8417-8426, 2018. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "text", + "content": "[10] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. Robust 3d hand pose estimation in single depth images: from single-view cnn to multi-view cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3593-3601, 2016. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 668 + ], + "type": "text", + "content": "[11] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1991-2000, 2017. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[12] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 475-491, 2018. 1, 2, 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20912" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[13] Hengkai Guo, Guijin Wang, Xinghao Chen, Cairong Zhang, Fei Qiao, and Huazhong Yang. Region ensemble network: Improving convolutional network for hand pose estimation. In 2017 IEEE International Conference on Image Processing (ICIP), pages 4512-4516. IEEE, 2017. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[15] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11927-11936, 2019. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 216, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 287, + 260 + ], + "type": "text", + "content": "[16] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 529-537, 2019. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 282 + ], + "type": "text", + "content": "[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 347 + ], + "type": "text", + "content": "[18] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. V2v-posenet: Voxel-to-voxel prediction network for accurate 3d hand and human pose estimation from a single depth map. In Proceedings of the IEEE conference on computer vision and pattern Recognition, pages 5079-5088, 2018. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "text", + "content": "[19] Markus Oberweger and Vincent Lepetit. Deeprior++: Improving fast and accurate 3d hand pose estimation. In Proceedings of the IEEE international conference on computer vision Workshops, pages 585-594, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "text", + "content": "[20] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 448, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 492 + ], + "type": "text", + "content": "[21] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in neural information processing systems, pages 5099-5108, 2017. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "text", + "content": "[22] Siamak Ravanbakhsh, Jeff Schneider, and Barnabas Poczos. Deep learning with sets and point clouds. arXiv preprint arXiv:1611.04500, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 569 + ], + "type": "text", + "content": "[23] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[24] Pengfei Ren, Haifeng Sun, Weiting Huang, Jiachang Hao, Daixuan Cheng, Qi Qi, Jingyu Wang, and Jianxin Liao. Spatial-aware stacked regression network for real-time 3d hand pose estimation. Neurocomputing, 437:42-57, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "type": "text", + "content": "[25] Pengfei Ren, Haifeng Sun, Qi Qi, Jingyu Wang, and Weiting Huang. Srn: Stacked regression network for real-time 3d hand pose estimation. In BMVC, page 112, 2019. 1, 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[26] Oindrila Saha, Aditya Kusupati, Harsha Vardhan Simhadri, Manik Varma, and Prateek Jain. Rnnpool: efficient nonlinear pooling for ram constrained inference. Advances in Neural Information Processing Systems, 33:20473-20484, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 553 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[27] Xiao Sun, Yichen Wei, Shuang Liang, Xiaou Tang, and Jian Sun. Cascaded hand pose regression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 824-832, 2015. 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[28] Danhang Tang, Hyung Jin Chang, Alykhan Tejani, and Tae Kyun Kim. Latent regression forest: Structured estimation of 3d articulated hand posture. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3786-3793, 2014. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 239 + ], + "type": "text", + "content": "[29] Jonathan Taylor, Lucas Bordeaux, Thomas Cashman, Bob Corish, Cem Keskin, Toby Sharp, Eduardo Soto, David Sweeney, Julien Valentin, Benjamin Luff, et al. Efficient and precise interactive hand tracking through joint, continuous optimization of pose and correspondences. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 274 + ], + "type": "text", + "content": "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "type": "text", + "content": "[31] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 318, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 545, + 363 + ], + "type": "text", + "content": "[32] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 363, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 545, + 407 + ], + "type": "text", + "content": "[33] Yiming Wu, Wei Ji, Xi Li, Gang Wang, Jianwei Yin, and Fei Wu. Context-aware deep spatiotemporal network for hand pose estimation from depth images. IEEE transactions on cybernetics, 50(2):787-797, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 408, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 462 + ], + "type": "text", + "content": "[34] John Yang, Yash Bhalgat, Simyung Chang, Fatih Porikli, and Nojun Kwak. Dynamic iterative refinement for efficient 3d hand pose estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1869-1879, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 508 + ], + "type": "text", + "content": "[35] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 553 + ], + "type": "text", + "content": "[36] Cheol-Hwan Yoo, Seowon Ji, Yong-Goo Shin, Seung-Wook Kim, and Sung-Jea Ko. Fast and accurate 3d hand pose estimation via recurrent neural network for capturing hand articulations. IEEE Access, 8:114010-114019, 2020. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20913" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_content_list.json b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e34bd6e059f7cafcfd97b40f00f7437d39d2c53 --- /dev/null +++ b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_content_list.json @@ -0,0 +1,1471 @@ +[ + { + "type": "text", + "text": "Handwritten and Printed Text Segmentation: A Signature Case Study", + "text_level": 1, + "bbox": [ + 133, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sina Gholamian \nThomson Reuters AI Labs \nToronto, Canada", + "bbox": [ + 225, + 174, + 434, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sina.gholamian@thomsonreuters.com", + "bbox": [ + 181, + 229, + 477, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ali Vahdat \nThomson Reuters AI Labs \nToronto, Canada", + "bbox": [ + 552, + 174, + 763, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ali.vahdat@thomsonreuters.com", + "bbox": [ + 527, + 229, + 787, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 279, + 312, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While analyzing scanned documents, handwritten text can overlap with printed text. This overlap causes difficulties during the optical character recognition (OCR) and digitization process of documents, and subsequently, hurts downstream NLP tasks. Prior research either focuses solely on the binary classification of handwritten text or performs a three-class segmentation of the document, i.e., recognition of handwritten, printed, and background pixels. This approach results in the assignment of overlapping handwritten and printed pixels to only one of the classes, and thus, they are not accounted for in the other class. Thus, in this research, we develop novel approaches to address the challenges of handwritten and printed text segmentation. Our objective is to recover text from different classes in their entirety, especially enhancing the segmentation performance on overlapping sections. To support this task, we introduce a new dataset, SignaTR6K, collected from real legal documents, as well as a new model architecture for the handwritten and printed text segmentation task. Our best configuration outperforms prior work on two different datasets by $17.9\\%$ and $7.3\\%$ on IoU scores. The SignaTR6K dataset is accessible for download via the following link: https://forms.office.com/r/2a5RDg7cAY.", + "bbox": [ + 75, + 313, + 472, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 684, + 207, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For various purposes, the digitization of hard-copy documents and associated challenges are an active area of research in both academia [25, 11, 27, 10, 6, 20] and industry [18]. This digital transformation involves scanning paper documents through an OCR process, making their text accessible for downstream natural language processing (NLP) tasks, such as named entity recognition (NER). The documents of interest can originate from a variety of domains, including historical documents [27], legal and court-issued documents [18], business contracts [25], and medical records and prescriptions [5]. Although various studies, as cited above, have been conducted, there is yet a considerable gap between current approaches and", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6eacb8b0ae731321b16ca49a0f5af1630a146e8148ec1e190575361d38647ff6.jpg", + "image_caption": [ + "Figure 1: Court documents are first printed and then signed or annotated by various parties. This results in handwritten text overlapping the underlying printed information, leading to performance degradation in downstream tasks, such as named entity recognition (NER). This is a fabricated example, however very similar to the original documents, to protect personally identifiable information (PII)." + ], + "image_footnote": [], + "bbox": [ + 509, + 277, + 883, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "human-level performance in mixed-text scenarios (i.e., when handwritten and printed text overlap). For instance, attorneys frequently sign legal documents, resulting in their signatures overlapping with their information. This overlap hampers the performance of OCR tools in character recognition, subsequently making it challenging for downstream tasks to accurately identify information linked to the attorneys and their associated law firms. Figure 1 provides an illustration of handwritten text (HT) overlapping with printed text (PT) in court documents. Extracting parties names is a crucial step in the named-entity recognition (NER) task for legal and court documents [24]. When attorneys and other involved parties sign these documents, which are later scanned using OCR tools, their signatures often obscure the details of names and law firms. Consequently, the semantic segmentation of handwritten elements, such as lawyers' signatures and accompanying handwritten notes, and printed text detailing the lawyer and the law firm's information, becomes vital.", + "bbox": [ + 496, + 566, + 892, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this research, we aim to address the challenges of HT and PT segmentation, as there is still a large gap between human-level performance and the existing approaches", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "582", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for this task. Our focus in this effort is to improve the segmentation performance in overlapping regions for scanned legal documents, and to aid in this endeavor, we also introduce a new dataset. In summary, our research makes the following contributions:", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new dataset, SignaTR6K (pronounce as Signature 6K)1 derived from 200 pixel-level manually annotated crops of images from genuine legal documents. The dataset comprises signatures, handwritten text, and printed text, which frequently overlap. With data augmentation, we have created a dataset of sizes 5169, 530, 558, for training, validation, and testing, respectively, that we release to the public, to facilitate dataset availability for future research and to aid in the training and evaluation of deep-learning segmentation models.", + "- We propose a novel architecture that integrates both semantic segmentation features and fine features, enhancing the performance of text segmentation over previous methods. Moreover, we introduce a new loss function termed Fusion loss, that, comparatively, is stable and converges to optimal loss values and Intersection over Union (IoU) scores.", + "- Lastly, we conduct an extensive quantitative and visual evaluation of different variations of our approach against prior work on two distinct datasets, and illustrate our approach's superior performance in the text segmentation task, especially in challenging scenarios where printed and handwritten text overlap." + ], + "bbox": [ + 66, + 176, + 470, + 525 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Literature Review", + "text_level": 1, + "bbox": [ + 76, + 536, + 398, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The text segmentation problem is defined as follows: given a scanned document possibly containing handwritten, printed, and background (i.e., blank) pixels, the task is to assign each pixel to its appropriate class. Scanned documents can originate from various sources, such as hard-copy paper documents or microfilms [20, 27]. Formally, for a given document $D$ , assuming there exist three classes as handwritten text $HT$ , printed text $PT$ , and background $BG$ , and pixel $p_i$ : $\\forall p_i \\in D : p_i(c) == \\text{True if } p_i \\in c \\& \\& c \\in \\{HT, PT, BG\\}$ (1). A situation may arise where a pixel belongs to two classes, i.e., $HT$ and $PT$ classes, when handwritten text overlaps with printed text, which single-label three-class formulation cannot handle such cases.", + "bbox": [ + 75, + 558, + 468, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several previous studies have studied the segmentation of handwritten and printed text [6, 10, 20, 27], however, they have inherent limitations. Some focus solely on binary classification, determining if a pixel is handwritten or not [10], whereas others adopt a 3-class formulation of the problem, classifying pixels as handwritten, printed, or background [6, 20, 27]. This exclusive assignment of", + "bbox": [ + 75, + 770, + 470, + 877 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pixels to three different classes paralyzes any machine learning segmentation model to properly detect pixels in the overlapping areas to belong to both handwritten and printed classes. In our usecase depicted in Figure 1, the OCR process achieves insufficient performance due to overlapping of handwritten text and signatures that overlay the printed text. As such, to improve the documents digitization quality, and subsequently a wide range of downstream document understanding and natural language processing (NLP) tasks, it is vital to devise image processing approaches that can understand and properly segment different layers of text, i.e., printed and handwritten text.", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Models. Prior literature has applied various approaches for the identification and separation of handwritten and printed text. Early approaches [7, 12] formulated the problem as a binary classification task using KNN and SVM and focusing on connected components (CCs) (i.e., groups of pixels). More recently, Li et al. [14] employed conditional random fields (CRFs), with formulating both unary and pairwise potentials for adjacent connected components by leveraging convolutional neural networks (CNNs) architecture for the separation of CCs. The limitation of CC-based approaches is that they determine the class membership for the entire component rather than at the pixel level. Consequently, pixel-level segmentation methods were introduced, leveraging Markov random fields (MRFs) and MLPs [19, 22] for pixel-level classification of PT and HT segmentation. Following the success of encoder-decoder architectures in object segmentation [21], more recent works [10, 6, 20, 27] have predominantly adopted a U-Net based architecture [21], which comprises an encoder-decoder network, for HT and PT segmentation.", + "bbox": [ + 496, + 280, + 893, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to our contribution of releasing a manually annotated dataset of legal documents with overlapping text, SignaTR6K, our methodology distinguishes itself from prior works in several significant ways. Firstly, we approach the segmentation problem with a four-class formulation, allowing overlapping pixels to be assigned to a new distinct class (OV), which signifies the presence of both HT and PT layers, leading to enhanced segmentation performance. Secondly, we introduce a novel architecture, the Mixed Feature Model (MFM), that combines a Fine Feature Path (FFP) with a Semantic Segmentation Path (SSP) and improves the performance by capitalizing on both high-level and low-level features. Notably, existing U-Net style architectures [10, 6, 20, 27] are limited to leveraging only the SSP path. Further, we present a new loss function, termed Fusion Loss, that converges faster and is more stable compared to prior losses for HT and PT segmentation task. In addition, we introduce a post-processing heuristic based on Conditional Random Fields (CRFH) to carry out relabeling, resulting in further enhancement of text segmentation performance. Additional details related to background and", + "bbox": [ + 496, + 584, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Available at https://forms.office.com/r/2a5RDg7cAY.", + "bbox": [ + 94, + 886, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "583", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/763a59c5b511a454c94fb78a6464b731c73244ec108eee0aa354cf6d8be248af.jpg", + "image_caption": [ + "Figure 2: This figure depicts a real crop from a legal document. (a) The original image containing printed text and overlaying handwritten signature; (b) Printed layer pixels annotated only; (c) Handwritten layer pixels annotated only. Any pixels not present in the printed or handwritten layers are marked as background." + ], + "image_footnote": [], + "bbox": [ + 84, + 84, + 462, + 141 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "relevant works can be found in the supplementary.", + "bbox": [ + 76, + 220, + 410, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SignalTR6K Dataset", + "text_level": 1, + "bbox": [ + 76, + 244, + 267, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Successfully training and testing a segmentation model requires access to high-quality labeled datasets. Due to the limited availability of public data that contains both handwritten and printed text, previous research has predominantly sought to synthetically generate such data. To achieve this, researchers have combined datasets that either exclusively contain printed or handwritten text, or those with non-overlapping text, such as IAM [17], RIMES [1], PRImA [4], CVL [13], Scanned Questionnaire [2], and WGM-SYN [27]. The scanned documents can originate from diverse domains, each possessing its unique fonts, characters, and quality. Whether the original documents having poor quality in the case of archival documents [27], or the scanning process resulted in lower resolution or loss of contrast, these factors also compound the complexities of the text segmentation task. In addition, in some cases errors are introduced during the automated text labeling process [20]. Given these challenges, the absence of a precisely and manually annotated and verified dataset from real sources that contains various patterns of overlapping printed and handwritten text, significantly impedes the supervised machine learning (ML) process. To address this gap, we introduce a new dataset, SignaTR6K (pronounced as Signature 6K), that is derived from 200 original legal documents from Thomson Reuters Legal Content Services [3]. The dataset features overlapping printed and handwritten text and hand-drawn signatures, as depicted in Figure 1. The documents originate from different organizations, including law firms and courts, each having its distinct fonts and document formats. Furthermore, the annotations are made by different individuals, ensuring a diverse range of printed and handwritten styles. Importantly, each document has been manually labeled and verified.", + "bbox": [ + 75, + 266, + 472, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 displays an original crop from SignaTR6K, which includes both a signature and the information of the signing party. Due to the presence of personally identifiable information (PIIs), we cropped the images to ensure no visible PIIs. Figure 2a illustrates the original image with both printed text and overlaying handwritten signature. Figure 2b presents only the printed layer pixels, while Figure 2c contains only the handwritten layer pixels that are manually annotated. Any pixels absent from either the printed or", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b43e6f719f36404e019790e70da3ef40c969d2f6d278e55cc65666700ec3f92c.jpg", + "image_caption": [ + "Figure 3: Examples from the SignaTR6K dataset, with the top row showing the crops and the bottom row their ground truth annotations. Red: class $PT$ , printed, Green: class $HT$ , handwritten, and Blue: class $BG$ , background." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "handwritten channels are designated as background pixels. It is also evident that in overlapping scenarios, some pixels belong to both handwritten and printed layers.", + "bbox": [ + 498, + 303, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset Generation. As hand-annotating a vast collection of real documents is time-consuming and expensive, in order to expand on the size of the dataset and make it adequate for training a deep learning model, we turn to data synthesis and augmentation techniques [10]. The synthetic approaches include general augmentation of crops, along with shifting, magnifying, and rotating operations. We also overlay handwritten and printed pixels from different crops to generate new real-like examples of overlapping text. For this purpose, from the 200 distinct original document samples, we set aside 16 crops, ensuring mutually exclusive samples in the test set. These crops are augmented only for the test set, and we create the training and validation sets from the remaining 184 samples. Following this approach and after excluding generated samples with visible PIDs, we have curated a dataset for training, validation, and testing sets of sizes 5169, 530, and 558 samples, respectively. Each sample is a pair of grayscale crop and its manually annotated ground truth. Each image is 256 by 256 pixels in size, with three channels (RGB) and typically contains several HT and PT overlaps. Four examples from the SignaTR6K dataset are presented in Figure 3, with their grayscale crops and corresponding ground truth (GT) labels. HT is indicated in green, PT in red, and BG in blue. Overlapping HT and PT pixels combine green and red channel values, resulting in a yellow appearance. We envision this dataset can be utilized for model training from scratch or further fine-tuning of a pre-trained classification or segmentation model for specific tasks. The SignaTR6K dataset is freely available for public download.", + "bbox": [ + 496, + 354, + 893, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Approach and Methodology", + "text_level": 1, + "bbox": [ + 500, + 819, + 758, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail our approach, the rationale behind the chosen architecture, the Semantic Segmentation Path (SSP) and Fine Feature Path (FFP), the various loss functions employed, and our novel Fusion loss.", + "bbox": [ + 498, + 840, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "584", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a67b8e29a25d60c94be82fd4c531f40468009cabf43cc5322f836e535fbea110.jpg", + "image_caption": [ + "Figure 4: Our architecture proposal showing FFP and SSP model paths. The output of FFP and SSP are concatenated prior to the final round of convolutions. A Softmax activation then selects one of the four classes for each pixel. The model's final output consists of pixel-level annotations of the input image. BN: BatchNorm." + ], + "image_footnote": [], + "bbox": [ + 119, + 85, + 851, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Model Architecture", + "text_level": 1, + "bbox": [ + 75, + 386, + 264, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The prevalent architecture for object segmentation applies a U-Net design [21, 16]. This architecture leverages a Fully Convolutional Network (FCN), that is without fully connected layers that are typically present at the end of Convolutional Neural Networks (CNNs). The U-Net architecture consists of two main parts, an encoder and a decoder. In the encoder segment, the original image (in our context a document crop) is fed into the model. It then undergoes a series of convolutions and max-pooling layers, extracting features from the image and reducing its dimensions. Conversely, the decoder processes the down-sampled image from the encoder using convolutions and up-sampling layers that eventually restores the image back to its original input size with the same number of channels. The final output of the decoder is a pixel-level labeling of the original image. Semantic Segmentation Path (SSP) in Figure 4 shows this part of the fully convolutional network. In addition to the encoder-decoder architecture in U-Net, there are also skip-connections that bring the feature maps directly from an encoder stage to the corresponding decoder stage, i.e., the decoder stage that has the same image size of the encoder stage. These skip-connections improve segmentation performance by re-accessing early-stage features that may be lost in the encoder's final output due to down-sampling.", + "bbox": [ + 73, + 406, + 473, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Four-Class Formulation", + "text_level": 1, + "bbox": [ + 75, + 773, + 299, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned earlier, prior approaches have either considered binary classification of HT or a three-class formulation of the problem. Binary classification detects only one type of text, while the three-class formulation results in overlapping pixels being assigned solely to either the HT or PT class, impairing performance. Therefore, we propose a four-class formulation of the segmentation task,", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the fourth class, overlap, $OV$ , models pixels that belong to both the HT and PT classes, specifically enabling the detection of overlapping areas. To expand on Formula (1) for this four-class formulation, for a given document $D$ , assuming there exist four classes as handwritten text $HT$ , printed text $PT$ , background $BG$ , and overlap $OV$ , and pixel $p_i$ : $\\forall p_i \\in D : p_i(c) == True$ if $p_i \\in c$ and $c \\in \\{HT, PT, BG, OV\\}$ .", + "bbox": [ + 496, + 386, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overlapping pixels are highlighted in yellow in the ground truth, as seen in Figure 3. Our four-class single-label classification employs a Softmax activation function in the final layer, which ensures that only one output for each pixel is activated (Figure 4). Since the output image comprises three channels, when the $OV$ class is predicted, during a post-processing step we turn on pixels for both $HT$ and $PT$ channels, resulting in the yellow color in the output image. In addition to the four-class formulation, we also explored three-class and multi-label formulations. In this scenario, instead of a Softmax activation, we applied Sigmoid activations for each class (i.e., three separate sig-moids). However, formulating the problem as a multi-label classification introduces added complexity and degrees of freedom. This can lead to undesirable scenarios, such as the simultaneous activation of pixels for $HT$ and $BG$ . Consequently, the multi-label approach did not yield results comparable to those of the four-class formulation.", + "bbox": [ + 496, + 508, + 892, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Semantic Segmentation Path (SSP)", + "text_level": 1, + "bbox": [ + 498, + 789, + 805, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The semantic segmentation path (SSP) of our model leverages a U-Net based architecture, with down-sampling in the encoder stages (i.e., backbone) and up-sampling in the decoder stages. The U-net architecture works well in capturing high-level image features. In this architecture, the encoder and decoder maintain a symmetrical architec", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "585", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture with a similar number of stages. For example, in the SSP, if the encoder goes through four down-sampling stages, i.e., the input image size changes from $256^{*}256$ to $16^{*}16$ ( $256 \\rightarrow 128 \\rightarrow 64 \\rightarrow 32 \\rightarrow 16$ ), correspondingly, the decoder undergoes four up-sampling stages, from $16^{*}16$ to $256^{*}256$ . Thus, the final output retains the original input image's pixel dimensions. For the SSP, we explored a variety of network sizes. As a baseline and for comparison, we implemented FCN-light [6, 20, 27]. We then improved on the SSP architecture by using VGG16 [23], InceptionV3 [26], and ResNet34 [8] as backbones of the SSP and observed improvement in the performance. In particular, the ResNet34 and InceptionV3 backbones outperform the prior work (FCN) due to their larger number of learnable parameters. Additionally, the inclusion of residual connections and varied convolution sizes allows better carry over the low-level features of text to the later stages of the segmentation network. This observation further inspired us to introduce the FFP network, which similarly incorporates residual connections and without down-sampling.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Fine Feature Path (FFP)", + "text_level": 1, + "bbox": [ + 76, + 398, + 302, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The semantic segmentation path excels when segmenting distinct objects. In this path, down-sampling layers are rapidly applied to the input image to capture high-level features and patterns. However, this rapid processing can lead to the loss of fine, or low-level, features. For our application, low-level features are crucial due to the intertwined nature of printed and handwritten text. To address this, we introduce a parallel path to the SSP, termed the fine feature path (FFP), which avoids down-sampling and instead incorporates a convolution block with residual connections. Note that, while the FFP aids in capturing fine features without down-sampling, on its own, i.e., without the SSP that includes down-sampling, it is insufficient for text segmentation, as the absence of high-level features means the model will not be able to detect high-level patterns irrespective of their pixel locations in the image. In Figure 4, the fine feature block of the FFP is repeated $N_{x}$ times; in our implementation, $N_{x} = 4$ . In addition, each stage of the FFP itself implements residual connections as it provides the flexibility to either bypass the block or use its output, leading to improved results, as we will discuss later in the paper. Similar architectures employing residual blocks have shown improved performance in fine object segmentation tasks, e.g., road segmentation [29]. Table 1 details the FFP architecture for four stages $(N_{x} = 4)$ and the connections between layers and convolution sizes.", + "bbox": [ + 75, + 417, + 472, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Mixed Feature Model (MFM)", + "text_level": 1, + "bbox": [ + 76, + 820, + 341, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Mixed Feature Model (MFM) serves as the umbrella model containing two parallel paths: SSP and FFP, as depicted in Figure 4. The objective of FM is to capture the image's low-level features alongside its high-level features.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8314f15a1a605120ce6423fbb8c74555d9b085a75831ba0603e6d11f1d278552.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: The architecture of the Fine Feature Path (FFP). In our design, we use four stages (Stages 1-4) of fine feature blocks. We include BatchNorm and ReLu activation after each convolution layer, and unlike SSP, down-sampling layers (e.g., max pooling) are absent from FFP. The output of each stage is concatenated with its input to create residual connections." + ], + "table_body": "
GroupLayer typeFilterInput(s)Output(s)Output size
Stage 1G 1Conv3 × 3/64img_inputs1.g1.c.o256 × 256 × 64
BatchNorm-s1.g1.c.os1.g1.b.o256 × 256 × 64
ReLU-s1.g1.b.os1.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s1.g1.r.os1.g2.c.o256 × 256 × 64
BatchNorm-s1.g2.c.os1.g2.b.o256 × 256 × 64
ReLU-s1.g2.b.os1.g2.r.o256 × 256 × 64
Concat-img_in, s1.g2.r.os1.o256 × 256 × 67
Stage 2G 1Conv3 × 3/64s1.os2.g1.c.o256 × 256 × 64
BatchNorm-s2.g1.c.os2.g1.b.o256 × 256 × 64
ReLU-s2.g1.b.os2.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s2.g1.r.os2.g2.c.o256 × 256 × 64
BatchNorm-s2.g2.c.os2.g2.b.o256 × 256 × 64
ReLU-s2.g2.b.os2.g2.r.o256 × 256 × 64
Concat-s1.o, s2.g2.r.os2.o256 × 256 × 131
Stage 3G 1Conv3 × 3/64s2.os3.g1.c.o256 × 256 × 64
BatchNorm-s3.g1.c.os3.g1.b.o256 × 256 × 64
ReLU-s3.g1.b.os3.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s3.g1.r.os3.g2.c.o256 × 256 × 64
BatchNorm-s3.g2.c.os3.g2.b.o256 × 256 × 64
ReLU-s3.g2.b.os3.g2.r.o256 × 256 × 64
Concat-s2.o, s3.g2.r.os3.o256 × 256 × 195
Stage 4G 1Conv3 × 3/64s3.os4.g1.c.o256 × 256 × 64
BatchNorm-s4.g1.c.os4.g1.b.o256 × 256 × 64
ReLU-s4.g1.b.os4.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s4.g1.r.os4.g2.c.o256 × 256 × 64
BatchNorm-s4.g2.c.os4.g2.b.o256 × 256 × 64
ReLU-s4.g2.b.os4.g2.r.o256 × 256 × 64
Concat-s3.o, s4.g2.r.os4.o256 × 256 × 259
Output-Conv1 × 1/4s4_oFFP_o256 × 256 × 4
", + "bbox": [ + 503, + 85, + 911, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The outputs of SSP and FFP are concatenated before the final layer convolution, producing the output of the MFM model. Additional details regarding the architecture, inputs, outputs, and convolution layer sizes of the MFM can be found in the supplementary.", + "bbox": [ + 496, + 398, + 893, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.6. CRF Post-Processing and CRF Heuristic", + "text_level": 1, + "bbox": [ + 498, + 484, + 848, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior research [6, 20, 27] has leveraged dense Conditional Random Fields (CRFs) as a post-processing step to re-label pixels based on their neighboring pixels. In our architectural exploration, we found that, while CRF post-processing is intended to improve segmentation performance, in practice it often hurts the segmentation performance by aggressively re-labeling pixels to incorrect classes. One issue arises due to the imbalanced nature of pixels across classes, with background pixels being predominant. Consequently, many pixels are mistakenly annotated from HT or PT to BG, or from HT to PT, which is undesirable. Based on this observation, we designed a heuristic for CRF post-processing to contain this unfavourable behaviour by only allowing the BG pixels to be relabelled to HT or PT pixels and not vice versa. As we will discuss in the experimentation section, this heuristic further improves segmentation performance.", + "bbox": [ + 496, + 503, + 893, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.7. Loss Functions", + "text_level": 1, + "bbox": [ + 498, + 771, + 653, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the nature of the scanned documents and the amount of text on each page, the number of background pixels (i.e., white blank pixels) surpasses that of handwritten and printed pixels. This leads to a class imbalance problem [30], posing the risk of predicting the majority of pixels as background while still minimizing the loss value. As a result, we investigated various loss functions and weight as", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "586", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "signments to different classes in the loss function to evaluate its impact on segmentation performance. Additionally, during this exploration, we observed that different loss functions achieve different IoU (Intersection over Union) scores. Consequently, we introduce a new loss function, Fusion Loss, to incorporate the benefits of various losses. In the following, we discuss the different loss functions we explored.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Entropy Loss. The multi-class cross-entropy loss is used for classification tasks involving more than two classes. It assumes that for each data point, only one class can be the ground truth, i.e., multi-class single-label classification. The standard form of cross-entropy loss for a single data point in multi-class segmentation with $M$ classes is defined as: $\\mathcal{L}_{CE}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot \\log(pr)$ , where $pr$ and $gt$ represent the prediction and ground truth, respectively. The final loss is the average of the losses for all data points in a training set. Similarly, the weighted version (WCE) is computed as $\\mathcal{L}_{WCE}(gt, pr) = -\\sum_{m=1}^{M} w_m \\cdot gt \\cdot \\log(pr)$ , with $w_m$ being the weight assigned to each class in the loss calculation.", + "bbox": [ + 75, + 214, + 472, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Focal Loss. The focal loss [15] was introduced to address the class-imbalance problem during training of object detection tasks. The focal loss puts the focus of the learning algorithm on the incorrectly classified examples by applying a modulating term $(1 - pr)^{\\gamma}$ to the cross-entropy loss. This scaling factor dynamically weighs down the contribution of easy and correctly classified samples, allowing the training loop to concentrate on the harder and incorrectly classified samples. The focal loss is calculated as: $\\mathcal{L}_{Focal}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot (1 - pr)^{\\gamma} \\log(pr)$ , where $\\gamma$ is the modulating/focusing factor, and we set $\\gamma == 2$ as per the original paper. Similarly, the weighted version of the focal loss can be expressed as: $\\mathcal{L}_{WF}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot \\alpha_m \\cdot (1 - pr)^{\\gamma} \\log(pr)$ , where $\\alpha_m$ is analogous to $w_m$ , assigning weights to each class such that: $\\forall \\alpha_m \\mid 0 < \\alpha_m < 1 \\& \\& \\sum_{m=1}^{M} \\alpha_m = 1$ .", + "bbox": [ + 75, + 428, + 472, + 672 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/067cbeaf1b37d4fc11ddfb819cd4c5397424fa2bd401a9eb26ca215b74e04c08.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DiceLoss.Dicelossisgivenby
L_Dice(precision, recall)= 1-2/M∑m=1M
precisionm·recallm/precisionm+recallm
Intuitively, it aims to maximize the F-Score while minimizing the loss, i.e., L_Dice = 1 - F_Score. Accordingly, the weighted version of the dice loss is defined as L_WD(precision, recall) = 1 - 2/M ∑m=1M w_m·precision_m·recall_m /precision_m+recall_m.
", + "bbox": [ + 75, + 681, + 472, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fusion Loss. Given our observation that different loss functions adversely affect different classes, we introduce a new loss targeted for maximizing the performance of all the classes. For example, the dice loss performs better on the background class, as in its formulation, it aims to maximize the F-Score. Because the majority of pixels are attributed", + "bbox": [ + 75, + 809, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to the background class, having higher values of correctly classified background pixels achieves a higher F-Score and lower loss values. However, this does not necessarily yield higher IoU scores for PT and HT classes. In contrast, the weighted versions of the cross-entropy and focal losses focus on the handwritten and printed classes. As such, with the Fusion loss, we aim to combine the behaviors of various losses, and we define the Fusion loss as the sum of the three weighted losses: $\\mathcal{L}_{\\text {Fusion }} = \\mathcal{L}_{WF} + \\mathcal{L}_{WCE} + \\mathcal{L}_{WD}$ .", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 498, + 242, + 736, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. Intersection over Union (IoU) is a commonly used metric to measure the performance of a segmentation task [28], and in particular text segmentation [25, 10, 6, 27]. In our context, we use IoU to measure the pixel-level performance of the model output versus the ground truth. For a class $c$ , let $TP_{c}$ represent the correctly predicted pixels for that class, $FP_{c}$ denote the pixels incorrectly predicted as belonging to class $c$ , and $FN_{c}$ represent the pixels that are incorrectly not predicted for class $c$ , then Intersection over Union, IoU, for class $c$ is calculated as: $IoU_{c} = \\frac{TP_{c}}{TP_{c} + FP_{c} + FN_{c}}$ . For each experiment, we calculate the IoU for all three classes, HT, PT, and BG. In the four-class formulation, pixels attributed to the overlap (OV) class are converted in a post-processing step to both HT and PT classes, contributing to the IoU calculations for both. We also calculate the mean IoU, which is the average of IoUs for these three classes: $IoU_{Mean} = \\frac{IoU_{HT} + IoU_{PT} + IoU_{BG}}{3}$ .", + "bbox": [ + 496, + 265, + 893, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments. We perform 50 epochs of training with a batch size of 8. For problem formulation, we use both three-class and four-class implementations. The experiments are conducted on two datasets: SignaTR6K and WGM-SYN [27]. The IoU values for FCN-light and WGM-MOD are as reported in [27] for the three-class formulation which has been performed on the same dataset, WGM-SYN. However, for the SignaTR6K dataset, we retrain an FCN-based model for three-class formulation to ensure a fair comparison on the new dataset. We run experiments for a variety of architectures and in three post-processing configurations. We structure experiments incrementally, i.e., adding one improvement at a time. This approach aids in understanding and isolating the impact of each added improvement, i.e., ablation study. Due to space constraints, the table detailing the complete list of experiment parameters and configurations is relegated to the supplementary.", + "bbox": [ + 496, + 537, + 893, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimented Architectures. We start with FCN-light which has approximately 295K parameters. This architecture aligns with prior research and serves as our baseline [6, 27]. Next, we improve the performance of FCNlight by adopting a four-class implementation. We then experiment with SSP architecture with three different back", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "587", + "bbox": [ + 486, + 944, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/39719b860ac19c247836ddd31d58e4727a1d414f03ca55662de6e747ed715616.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: IoU performance (%) on the WGM-SYN dataset [27]. The maximum value of a column (representing a class) is underlined. The overall maximum of a class across different post-processing methods is both bolded and underlined. For example, the best mean IoU for the WGM-SYN dataset is achieved with Fusion loss, using CRFH, and the MFM-ResNet34 architecture, stands at 58.93. The best performing configuration from prior work, denoted with $(^{*})$ , stands at 50.00." + ], + "table_body": "
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-light [6]WGM-MOD [27]~295K~295KWeighted CE24.0026.0072.0041.0011.0023.0072.0036.00----
Weighted CE42.0036.0074.0050.00*41.0032.0074.0049.00----
4-Class(Ours)FCN-light~295KCE46.4941.9873.7754.0838.9529.3573.9247.4146.5741.5973.8754.01
Focal46.0242.0173.8053.9532.9624.5074.0043.8246.1141.2373.9853.77
Dice48.0147.2871.2255.5548.5948.4871.2256.1048.0747.7971.2955.72
Weighted CE46.0742.1873.8254.0240.5430.0773.9848.2046.3141.6873.9553.98
Weighted Focal43.7141.7773.9553.1432.6724.4074.0743.7144.0641.1274.1153.10
Weighted Dice47.9747.2271.0255.4048.6948.3471.0256.0248.0447.7071.1055.61
Fusion48.1247.2571.9355.7746.3444.2571.8654.1548.1947.8972.3856.15
SSP - VGG16~24MCE35.1432.4073.9647.1730.9719.1473.8841.3335.2132.0874.1347.14
Focal34.5632.2274.0246.9329.1119.4273.9140.8134.6731.7974.2246.89
Dice40.6439.6471.1350.4743.3040.1671.2151.5540.8140.2571.2550.77
Weighted CE35.7333.4474.4347.8732.0020.7074.1242.2835.9332.9574.6547.84
Weighted Focal34.9034.3574.3647.8729.1820.7274.0641.3235.1133.8074.5647.82
Weighted Dice41.8040.6470.6751.0344.1441.3970.7052.0742.5241.1970.9751.56
Fusion39.9939.5672.3150.6235.3927.9372.3545.2240.1039.9472.6350.89
MFM (FFP + SSP) - InceptionV3~30MCE52.5144.0973.9156.8446.0335.9474.2152.0652.5542.1774.3556.36
Focal51.5643.7473.8956.4035.8927.1574.4145.8251.6341.7474.4355.93
Dice20.9119.2881.5240.5725.5124.6179.9443.3520.9219.2981.5440.58
Weighted CE51.5843.7273.9556.4244.7834.0474.3251.0551.6441.8574.4055.96
Weighted Focal51.0143.6273.8956.1835.7626.6074.2245.5351.0841.6074.3555.68
Weighted Dice22.3021.2580.7341.4229.1129.4877.9745.5222.3221.2680.7441.44
Fusion53.2251.2571.8458.7750.0548.8372.0356.9753.3249.8372.7258.62
MFM (FFP + SSP) - ReNet34~24MCE52.4044.0273.9156.7847.5236.4074.1252.6852.4342.1274.3456.30
Focal51.9543.9873.8956.6136.7327.1774.3146.0751.9041.9974.3356.08
Dice29.6829.4377.0645.3931.6225.3975.1544.0529.8429.6377.0745.52
Weighted CE52.6344.1273.9456.9047.6037.2574.1252.9952.5642.2274.3656.38
Weighted Focal51.6843.9573.9056.5135.7327.2274.3045.7551.7241.9874.3456.01
Weighted Dice51.4450.6071.9658.0052.9751.4271.9058.7651.5449.3772.7557.89
Fusion52.9951.7271.6958.8051.1550.7971.8357.9253.0151.2972.4958.93
", + "bbox": [ + 76, + 88, + 890, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bones: VGG16, InceptionV3, and ResNet34. Our experimentation also includes three variations of MFM, which incorporates both FFP and SSP, and with three SSP variations: VGG16, InceptionV3, and ResNet34. For our implementations, we use the Segmentation Models Library available in TensorFlow [9].", + "bbox": [ + 75, + 494, + 468, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Post-Processing Configurations. We experiment with three distinct configurations: without post-processing, with CRF post-processing, and with CRFH post-processing. In the scenario without post-processing, the IoU calculation is performed directly on the output of the model (FCN-light, SSP, and MFM). For the CRF post-processing approach, all classes are permissible for relabeling. However, with the CRF post-processing with heuristic (CRFH), as explained in Section 4.6, only the background class pixels (BG) are allowed to be relabeled as HT or PT classes", + "bbox": [ + 75, + 594, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weight Initialization. For both FCN-light and SSP models, model weights start from their initial random values at the beginning of the training. We train MFM configurations at last to reuse the weights from the SSP configurations. Thus, for the MFM trainings, we apply transfer learning by initializing with weights from the corresponding, previously trained SSP. For example, when training an MFM (FFP+SSP) with an SSP using the ResNet34 backbone, we initialize its SSP path with the available weights from the", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SSP - ResNet34. Lastly, although we have aimed for consistent IoU results by fixing the random seeds, the inherent non-determinism associated with GPU execution remains. As such, we have run our experiments multiple times to validate the IoU value improvements across configurations.", + "bbox": [ + 496, + 494, + 890, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Tables 2 and 3 show the IoU values for the WGM-SYN and SignaTR6K datasets. The overall trend across the results indicates that transitioning from the three-class formulation to the four-class formulation improves the IoU scores. Furthermore, employing larger model backbones generally, but not always, improves the segmentation performance. Among all the model architectures evaluated, the ResNet34 and InceptionV3 backbones achieve the highest performance, which we attribute to their residual connections and different size convolutions as they can better capture fine features from the image. We provide a summary of the trends in the results here, while a more detailed version of the result tables can be found in the supplementary material.", + "bbox": [ + 495, + 584, + 890, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Prior work [6, 27], which implements the three-class formulation of HT and PT segmentation, generally exhibits the lowest performance. Transitioning from the three-class to the four-class formulation with the same backbone, i.e., FCN-light, with the same number of model parameters and WCE loss, improves the average IoU values by", + "bbox": [ + 490, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "588", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8cde07cb8a58e717a893e0641df6019ff2171c3feaa3724f81d72d2dd40c04f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-based [6, 27]~295KCE62.5688.0998.4083.02*52.6889.6899.2680.4662.7290.5899.0584.11
Focal62.3488.0098.4582.9344.6084.8699.2876.2562.5790.8399.2184.21
Dice60.4587.2997.8581.8661.2488.8398.1782.7460.5287.8897.9782.12
Weighted CE60.5884.8997.7881.0953.4590.7399.5281.2360.8486.3398.2481.80
Weighted Focal61.2585.7498.0281.6744.2785.7799.5076.5161.5587.4298.5882.52
Weighted Dice60.2187.0097.7481.6560.7288.1797.9882.2960.2787.4697.8381.86
Fusion61.5288.3998.3882.7658.9492.5599.3783.6261.6790.4598.9183.68
4-Class(Ours)FCN-based~295KCE64.5589.2198.3984.0554.6089.6899.2381.1764.8791.8199.0685.25
Focal64.1088.8698.3283.7646.6486.0199.2677.3064.3491.6599.1185.03
Dice64.3788.6898.3783.8165.1790.1398.5984.6364.5789.1498.4684.06
Weighted CE63.7887.5298.1683.1554.7790.5099.4281.5664.1289.4398.7184.09
Weighted Focal63.6887.7198.1983.2048.1887.1099.4578.2464.0589.7298.8084.19
Weighted Dice64.2188.5798.3083.6965.0890.1098.5384.5764.4489.0698.3983.96
Fusion64.6888.4898.3383.8359.8391.7799.2883.6365.0090.7298.9184.87
MFM (FFP + SSP) - InceptionV3~30MCE73.1092.6698.7788.1863.4892.8999.5585.3173.0594.8999.3689.10
Focal72.8092.5098.7588.0157.4790.5899.5482.5372.7794.7499.3588.95
Dice72.5692.2098.7087.8272.3893.6099.0088.3372.5294.7399.3788.87
Weighted CE72.6691.9098.5487.7062.6593.0499.6285.1072.6393.3198.9488.29
Weighted Focal72.5592.1398.6287.7759.3591.6499.5983.5372.5093.7799.0788.45
Weighted Dice72.5992.3198.7087.8772.6094.0099.0788.5672.5494.6099.3288.82
Fusion72.5592.2598.7187.8365.5793.4999.5386.1972.4994.6299.3588.82
MFM (FFP + SSP) - ResNet34~24MCE72.8192.5698.7888.0563.0492.9499.5585.1772.7594.9399.3989.02
Focal73.0492.4698.7588.0853.0289.0499.5580.5473.0094.7599.3589.03
Dice72.9692.3898.7288.0273.1693.3598.9388.4872.9394.7999.3689.03
Weighted CE72.9691.9698.6987.8764.4992.4799.5685.5172.9094.1999.2788.79
Weighted Focal73.1892.1098.6387.9754.9689.7199.6081.4273.1693.7299.0688.65
Weighted Dice72.7892.3298.7187.9472.9593.6699.0088.5372.7594.6799.3488.92
Fusion73.2692.4598.7388.1568.3894.8599.5687.6073.2194.5999.3189.04
", + "bbox": [ + 76, + 84, + 890, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: IoU performance (%) on the SignaTR6K dataset. The maximum value of a column (i.e., class) is underlined, and the overall maximum for a class across different post-processing is both bolded and underlined. For example, the best mean IoU for the SignaTR6K dataset is for CE loss, with CRFH, and the MFM-InceptionV3 architecture at 89.10. The best performing configuration of prior work, i.e., excluding results for Fusion loss and CRFH, is marked with $(^{*})$ and is 83.02.", + "bbox": [ + 75, + 405, + 890, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d5b25256701b918e3a13e5bff7cc6c6e14cea621887a5eb2158574779bf6660.jpg", + "image_caption": [ + "(a) Validation set losses" + ], + "image_footnote": [], + "bbox": [ + 91, + 477, + 272, + 570 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5e5f29b952699f79b1351b5f098f7e636d956263c2b7fb4de5a3d22f82250ef.jpg", + "image_caption": [ + "(b) Validation set IoUs", + "Figure 5: Convergence speed of different loss functions and normalized loss values on the validation set of the SignaTR6K dataset." + ], + "image_footnote": [], + "bbox": [ + 279, + 477, + 462, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$8.0\\%$ $(50\\rightarrow 54.02)$ for the WGM-SYN dataset, and $2.6\\%$ $(81.09\\rightarrow 83.15)$ for the SignaTR6K dataset.", + "bbox": [ + 78, + 619, + 467, + 651 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Applying CRF post-processing generally degrades the results. In our research, we observe that CRF post processing is dataset dependent and decreases the performance of segmentation in some cases. CRF post-processing relabels aggressively and incorrectly converts HT and PT pixels to background ones, or OV pixels to PT ones. For example, Figure 6i illustrates that how OV pixels are wrongly relabeled to PT class by CRF post-processing. Contrary to CRF post-processing, our CRF heuristic generally improves the IoU values. For example, for the SignaTR6K dataset, for FCN-light with four-class implementation and CE loss, CRFH improves HT IoUs by $2.9\\%$ (89.21→91.81).", + "- Moving from the FCN-light architecture to larger models, i.e., SSP and MFM models, generally improves the results, except for the VGG16 backbone. For VGG16, we ratio" + ], + "bbox": [ + 66, + 655, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nalize that having a deep network without residual connections or varied-size convolutions, as seen in ResNet34 and InceptionV3 architectures, hurts the model's ability in capturing fine features. In addition, we observe that adding FFP helps to improve the performance of the VGG model as we compare SSP and MFM performance values for both datasets. This also confirms the FFP's ability to capture low-level features that are missed by the SSP path.", + "bbox": [ + 501, + 479, + 890, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- The MFM model with ResNet34 and InceptionV3 backbones generally achieves the best results. MFM-ResNet34 with Fusion loss and CRFH achieves the best result for the WGM-SYN dataset (58.93%), while MFM-InceptionV3 with CE loss and CRFH achieves the best mean IoUs for the SignaTR6K dataset (89.10%). Overall, the best-performing model from our designs improves on the mean IoU performance of the best-performing prior work [27] by $17.9\\%$ (50→58.93) and $7.3\\%$ (83.02→89.10) for the WGM-SYN and SignaTR6K datasets, respectively.", + "bbox": [ + 490, + 606, + 890, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Although using skip connections in SSP helps with the segmentation performance and thus less pronounced improvement is observed in MFM, our FFP design is agnostic of the SSP implementation. This distinction becomes clear when VGG16 is employed as SSP; the addition of FFP boosts the IoU scores by $7\\%$ and $8\\%$ on the WGM-SYN and SignaTR6K datasets, respectively. The FCN-based model with 295K parameters outperforms VGG16 with 30M parameters. Thus, larger models do not necessarily outper", + "bbox": [ + 490, + 763, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "589", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/22e67c9e4dcac981ffb14bd13df9d2bb981ba01a2c514da5efa5924abb0af7d9.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 183, + 166 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e98c98ecfe0cc7afe43a7dcbdc8f1e803e57d5fd0d6a35054e8e003b7fbe75c8.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 183, + 88, + 282, + 165 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/368de0aa453c66ef7def5efe871d67dd1c35ff27e3339b0f3a3654b08e747324.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 383, + 165 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bd7c6d4b3d22374fea9c78b12157ad7a4232967d98360b9461661a1364d1a48f.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 383, + 88, + 485, + 165 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d8959a15031bf004ee0d854582b70be19b04b0337e2a2081732885e3ff5e2b6b.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 485, + 88, + 586, + 165 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a899598576675d1ce9eec88411d45fabdc5a3d0dc88fcd130f5ea8c4a00f3906.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 586, + 88, + 687, + 165 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0a3d5dc7f2a01aaeb3517f308ec868cd2a44dc8e70278e1f6df3876d29236b05.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 687, + 88, + 787, + 165 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/422d27ef0a3376567a76daff09504a39c5bdc3857a007acf62d19b859530c362.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 789, + 88, + 890, + 165 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/536b7d2ea595d5cb995a75067329c70d5e7f2721a8a9d71b449b5305992fdcaf.jpg", + "image_caption": [ + "(i)" + ], + "image_footnote": [], + "bbox": [ + 81, + 176, + 181, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c60d21c244800bfe3404fe9d10a1d9b11c87d9d09a8c30838fda8434c6e8f927.jpg", + "image_caption": [ + "(i)" + ], + "image_footnote": [], + "bbox": [ + 181, + 176, + 282, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b2b8126387904ca6b9a202ddc346af19661894185bbd35c244e7a07755c1e3c6.jpg", + "image_caption": [ + "(k)", + "Figure 6: Example results on the test set of the SignaTR6K dataset for our approach compared to the ground truth and prior works. (a) Input image; (b) Ground truth; (c) & (d) 3-class FCN-based [6, 27] with CE loss without (c) and with (d) CRF post-processing; (e), (f), & (g) Our FCN-based 4-class formulation with CE loss without CRF (e), with CRF (f), and with CRFH (g); (h), (i), & (j) SSP-ResNet34 with CE loss without CRF (h), with CRF (i), and with CRFH (j); (k), (l), & (m) MFM-ResNet34 with CE loss without CRF (k), with CRF (l), and with CRFH (m); (n), (o), & (p) MFM-ResNet34 with Fusion loss without CRF (n), with CRF (o), and with CRFH (p)." + ], + "image_footnote": [], + "bbox": [ + 282, + 176, + 383, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7e63806dcb481218440ac55a851328a28e7cd52a669502f920471609ee2a807e.jpg", + "image_caption": [ + "(1)" + ], + "image_footnote": [], + "bbox": [ + 383, + 176, + 485, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/87b4827c41ecc70d414b997ade04b739c3aac96f0749aac514fee637ff87778a.jpg", + "image_caption": [ + "(m)" + ], + "image_footnote": [], + "bbox": [ + 485, + 176, + 586, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ec8fadea36b3fb4a27dc718c40651ae8b05a1e28c8e140f9e9cb7f3946b8691c.jpg", + "image_caption": [ + "(n)" + ], + "image_footnote": [], + "bbox": [ + 586, + 176, + 687, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/19ef7ffc552a2c13939c24e86f356133a20ed1a9c1bc5245db31dda634cf4c07.jpg", + "image_caption": [ + "(0)" + ], + "image_footnote": [], + "bbox": [ + 687, + 176, + 787, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/81d52906c08a12adb3090fee257dfaa97489ff6cc1c1d949a285eeef9841fef4.jpg", + "image_caption": [ + "(p)" + ], + "image_footnote": [], + "bbox": [ + 789, + 176, + 890, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "form smaller ones consistently, and integrating FFP can bring significant improvements in IoU scores. Our findings show that, irrespective of the SSP size and architecture, incorporating FFP improves the performance, and it is essential for segmentation tasks involving fine objects, e.g., text.", + "bbox": [ + 78, + 344, + 468, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss Functions. We also perform an analysis to confirm the stability and convergence of the Fusion loss compared to other losses. Figures 5a and 5b compare different loss functions on the validation set of the SignaTR6K dataset. In Figure 5a, we observe that the Fusion loss is generally stable while converging to its minimum loss. Additionally, in Figure 5b, we observe Fusion loss reaches its maximum IoU around Epoch 15, generally faster than most other losses and stays close to the maximum IoU values on the validation set. It is important to highlight that the maximum IoU on the validation set for Focal and CE losses does not necessarily imply the best performance on the test set. Overall, Fusion loss shows a stable behavior and achieves a better performance on the test set, whereas for comparison, Focal loss shows instability at some epochs during the training.", + "bbox": [ + 75, + 438, + 468, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visual Comparison. Figure 6 shows sample model outputs for the SignaTR6K dataset, with two rectangle regions of interest highlighted to showcase the differences between various models. Figure 6b shows the ground truth, and a visual trend indicates that the performance on the regions of interest improves from Figure 6c to Figure 6p. It is also visually noticeable that CRF post-processing aggressively relabels pixels (6i), whereas CRFH generally improves the results. More visual comparisons on the WGM-SYN and SignaTR6K datasets can be found in the supplementary.", + "bbox": [ + 75, + 676, + 467, + 828 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations. Compared to prior work, the complexity and training cost of our approach present some limitations. Our MFM model, being larger than FCN-light, requires greater GPU memory sizes for training, and takes longer to train.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "However, we believe some of the limitations on the training time can be offset through transferring weights from SSP models to MFM ones. Additionally, while our approach outperforms prior work on the WGM-SYN dataset, the mean IoU performance is lower compared to the SignaTR6K dataset. This indicates potential limitations in our method's efficacy for lower-quality original documents, those undergoing low-quality scanning processes (like historical documents), or those with errors from automated labeling. We also attribute the improved performance on the SignaTR6K dataset to its higher quality and our thorough manual annotation, that have resulted in better model training and improved IoU results.", + "bbox": [ + 496, + 344, + 890, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 554, + 617, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Segmentation of handwritten text (HT) and printed text (PT) is vital for digitization and understanding of scanned documents. The complexity increases with the overlapping of different text types. In this research, we introduced SignaTR6K, a new open-source dataset with high-quality, manual pixel-level annotations, sourced from original legal documents. Additionally, we proposed a novel four-class formulation and a new architecture for the segmentation task. Our design leverages both the Fine Feature Path (FFP) and the Semantic Segmentation Path (SSP) to create the Mixed Feature Model (MFM), and incorporates both high-level and low-level features and improves on the text segmentation performance. We also introduced a CRF-based post-processing heuristic (CRFH) that further improves the model output, and included a new loss function, Fusion loss, that combines the advantages of different loss functions and achieves faster convergence and higher stability compared to most of the evaluated losses. In conclusion, our designs outperform the prior work in mean IoU scores by $17.9\\%$ and $7.3\\%$ on WGM-SYN and SignaTR6K datasets, respectively.", + "bbox": [ + 496, + 575, + 890, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "590", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 80, + 90, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] RIMES. http://www.a2ialab.com/doku.php?id=rimesdatabase:start. Inaccessible:2023-01-23.", + "[2] Scanned Questionnaire. https://github.com/jottue/HTSNet. Accessed: 2023-01-23.", + "[3] Thomson Reuters Legal Content Services. https://www.thomsonreuters.com/en/ products-services/legal.html. Accessed: 2023-01-21.", + "[4] Apostolos Antonacopoulos, David Bridson, Christos Papadopoulos, and Stefan Pletschacher. A realistic dataset for performance evaluation of document layout analysis. In 2009 10th International Conference on Document Analysis and Recognition, pages 296-300. IEEE, 2009.", + "[5] Dibyajyoti Dhar, Avishek Garain, Pawan Kumar Singh, and Ram Sarkar. Hp_docpres: A method for classifying printed and handwritten texts in doctor's prescription. Multimedia Tools and Applications, 80(7):9779-9812, 2021.", + "[6] Nicolas Dutly, Fouad Slimane, and Rolf Ingold. PHTI-WS: a printed and handwritten text identification web service based on FCN and CRF post-processing. In 2019 International Conference on Document Analysis and Recognition Workshops (IC-DARW), volume 2, pages 20-25. IEEE, 2019.", + "[7] Jürgen Franke and Matthias Oberlander. Writing style detection by statistical combination of classifiers in form reader applications. In Proceedings of 2nd International Conference on Document Analysis and Recognition (ICDAR'93), pages 581-584. IEEE, 1993.", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[9] Pavel Iakubovskii. Segmentation models. https://github.com/qubvel/segmentation_models, 2019.", + "[10] Junho Jo, Hyung Il Koo, Jae Woong Soh, and Nam Ik Cho. Handwritten text segmentation via end-to-end learning of convolutional neural networks. *Multimedia Tools and Applications*, 79(43):32137-32150, 2020.", + "[11] Celso AML Junior, Matheus Henrique Marques da Silva, Byron Leite Dantas Bezerra, Bruno José Torres Fernandes, and Donato Impedovo. FCN+ RL: a fully convolutional network followed by refinement layers to offline handwritten signature segmentation. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020." + ], + "bbox": [ + 78, + 114, + 468, + 906 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] R Kandan, Nirup Kumar Reddy, KR Arvind, and AG Ramakrishnan. A robust two level classification algorithm for text localization in documents. In Advances in Visual Computing: Third International Symposium, ISVC 2007, Lake Tahoe, NV, USA, November 26-28, 2007, Proceedings, Part II 3, pages 96-105. Springer, 2007.", + "[13] Florian Kleber, Stefan Fiel, Markus Diem, and Robert Sablatnig. CVL-database: An off-line database for writer retrieval, writer identification and word spotting. In 2013 12th international conference on document analysis and recognition, pages 560-564. IEEE, 2013.", + "[14] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Printed/handwritten texts and graphics separation in complex documents using conditional random fields. In 2018 13th IAPR International Workshop on Document Analysis Systems (DAS), pages 145-150. IEEE, 2018.", + "[15] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017.", + "[16] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015.", + "[17] U-V Marti and Horst Bunke. The IAM-database: an English sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1):39-46, 2002.", + "[18] Milda Norkute, Nadja Herger, Leszek Michalak, Andrew Mulder, and Sally Gao. Towards explainable ai: Assessing the usefulness and impact of added explainability features in legal document summarization. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-7, 2021.", + "[19] Xujun Peng, Srirangaraj Setlur, Venu Govindaraju, and Ramachandrula Sitaram. Handwritten text separation from annotated machine printed documents using markov random fields. International Journal on Document Analysis and Recognition (IJDAR), 16(1):1-16, 2013.", + "[20] Anastasia Prikhodina. Handwritten and printed text separation for historical documents. 2021.", + "[21] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical" + ], + "bbox": [ + 501, + 90, + 890, + 878 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "591", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image computing and computer-assisted intervention, pages 234-241. Springer, 2015.", + "[22] Mathias Seuret, Marcus Liwicki, and Rolf Ingold. Pixel level handwritten and printed content discrimination in scanned documents. In 2014 14th International Conference on Frontiers in Handwriting Recognition, pages 423-428. IEEE, 2014.", + "[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[24] Stavroula Skylaki, Ali Oskooei, Omar Bari, Nadja Herger, and Zac Kriegman. Named entity recognition in the legal domain using a pointer generator network. arXiv preprint arXiv:2012.09936, 2020.", + "[25] Nishant Subramani, Alexandre Matton, Malcolm Greaves, and Adrian Lam. A survey of deep learning approaches forOCR and document understanding. arXiv preprint arXiv:2011.13534, 2020.", + "[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the in-" + ], + "bbox": [ + 78, + 90, + 468, + 406 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016.", + "[27] Mahsa Vafaie, Oleksandra Bruns, Natasja Pilz, Jörg Waitelonis, and Harald Sack. Handwritten and printed text identification in historical archival documents. In Archiving Conference, volume 19, pages 15–20. Society for Imaging Science and Technology, 2022.", + "[28] Zhaobin Wang, E Wang, and Ying Zhu. Image segmentation evaluation: A survey of methods. Artificial Intelligence Review, 53:5637-5674, 2020.", + "[29] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual U-Net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018.", + "[30] Zhi-Hua Zhou and Xu-Ying Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on knowledge and data engineering, 18(1):63-77, 2005." + ], + "bbox": [ + 501, + 90, + 890, + 393 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "592", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_model.json b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c492206181e45b5595a29d30795879c8f9d5258a --- /dev/null +++ b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_model.json @@ -0,0 +1,2015 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "Handwritten and Printed Text Segmentation: A Signature Case Study" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.175, + 0.436, + 0.226 + ], + "angle": 0, + "content": "Sina Gholamian \nThomson Reuters AI Labs \nToronto, Canada" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.231, + 0.478, + 0.244 + ], + "angle": 0, + "content": "sina.gholamian@thomsonreuters.com" + }, + { + "type": "text", + "bbox": [ + 0.553, + 0.175, + 0.764, + 0.226 + ], + "angle": 0, + "content": "Ali Vahdat \nThomson Reuters AI Labs \nToronto, Canada" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.231, + 0.788, + 0.243 + ], + "angle": 0, + "content": "ali.vahdat@thomsonreuters.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.28, + 0.313, + 0.295 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.473, + 0.661 + ], + "angle": 0, + "content": "While analyzing scanned documents, handwritten text can overlap with printed text. This overlap causes difficulties during the optical character recognition (OCR) and digitization process of documents, and subsequently, hurts downstream NLP tasks. Prior research either focuses solely on the binary classification of handwritten text or performs a three-class segmentation of the document, i.e., recognition of handwritten, printed, and background pixels. This approach results in the assignment of overlapping handwritten and printed pixels to only one of the classes, and thus, they are not accounted for in the other class. Thus, in this research, we develop novel approaches to address the challenges of handwritten and printed text segmentation. Our objective is to recover text from different classes in their entirety, especially enhancing the segmentation performance on overlapping sections. To support this task, we introduce a new dataset, SignaTR6K, collected from real legal documents, as well as a new model architecture for the handwritten and printed text segmentation task. Our best configuration outperforms prior work on two different datasets by \\(17.9\\%\\) and \\(7.3\\%\\) on IoU scores. The SignaTR6K dataset is accessible for download via the following link: https://forms.office.com/r/2a5RDg7cAY." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.685, + 0.209, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For various purposes, the digitization of hard-copy documents and associated challenges are an active area of research in both academia [25, 11, 27, 10, 6, 20] and industry [18]. This digital transformation involves scanning paper documents through an OCR process, making their text accessible for downstream natural language processing (NLP) tasks, such as named entity recognition (NER). The documents of interest can originate from a variety of domains, including historical documents [27], legal and court-issued documents [18], business contracts [25], and medical records and prescriptions [5]. Although various studies, as cited above, have been conducted, there is yet a considerable gap between current approaches and" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.278, + 0.885, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Figure 1: Court documents are first printed and then signed or annotated by various parties. This results in handwritten text overlapping the underlying printed information, leading to performance degradation in downstream tasks, such as named entity recognition (NER). This is a fabricated example, however very similar to the original documents, to protect personally identifiable information (PII)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.893, + 0.854 + ], + "angle": 0, + "content": "human-level performance in mixed-text scenarios (i.e., when handwritten and printed text overlap). For instance, attorneys frequently sign legal documents, resulting in their signatures overlapping with their information. This overlap hampers the performance of OCR tools in character recognition, subsequently making it challenging for downstream tasks to accurately identify information linked to the attorneys and their associated law firms. Figure 1 provides an illustration of handwritten text (HT) overlapping with printed text (PT) in court documents. Extracting parties names is a crucial step in the named-entity recognition (NER) task for legal and court documents [24]. When attorneys and other involved parties sign these documents, which are later scanned using OCR tools, their signatures often obscure the details of names and law firms. Consequently, the semantic segmentation of handwritten elements, such as lawyers' signatures and accompanying handwritten notes, and printed text detailing the lawyer and the law firm's information, becomes vital." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this research, we aim to address the challenges of HT and PT segmentation, as there is still a large gap between human-level performance and the existing approaches" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "for this task. Our focus in this effort is to improve the segmentation performance in overlapping regions for scanned legal documents, and to aid in this endeavor, we also introduce a new dataset. In summary, our research makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.068, + 0.178, + 0.472, + 0.332 + ], + "angle": 0, + "content": "- We introduce a new dataset, SignaTR6K (pronounce as Signature 6K)1 derived from 200 pixel-level manually annotated crops of images from genuine legal documents. The dataset comprises signatures, handwritten text, and printed text, which frequently overlap. With data augmentation, we have created a dataset of sizes 5169, 530, 558, for training, validation, and testing, respectively, that we release to the public, to facilitate dataset availability for future research and to aid in the training and evaluation of deep-learning segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.068, + 0.339, + 0.472, + 0.43 + ], + "angle": 0, + "content": "- We propose a novel architecture that integrates both semantic segmentation features and fine features, enhancing the performance of text segmentation over previous methods. Moreover, we introduce a new loss function termed Fusion loss, that, comparatively, is stable and converges to optimal loss values and Intersection over Union (IoU) scores." + }, + { + "type": "text", + "bbox": [ + 0.068, + 0.434, + 0.472, + 0.526 + ], + "angle": 0, + "content": "- Lastly, we conduct an extensive quantitative and visual evaluation of different variations of our approach against prior work on two distinct datasets, and illustrate our approach's superior performance in the text segmentation task, especially in challenging scenarios where printed and handwritten text overlap." + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.178, + 0.472, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.537, + 0.4, + 0.554 + ], + "angle": 0, + "content": "2. Background and Literature Review" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.77 + ], + "angle": 0, + "content": "The text segmentation problem is defined as follows: given a scanned document possibly containing handwritten, printed, and background (i.e., blank) pixels, the task is to assign each pixel to its appropriate class. Scanned documents can originate from various sources, such as hard-copy paper documents or microfilms [20, 27]. Formally, for a given document \\( D \\), assuming there exist three classes as handwritten text \\( HT \\), printed text \\( PT \\), and background \\( BG \\), and pixel \\( p_i \\): \\( \\forall p_i \\in D : p_i(c) == \\text{True if } p_i \\in c \\& \\& c \\in \\{HT, PT, BG\\} \\) (1). A situation may arise where a pixel belongs to two classes, i.e., \\( HT \\) and \\( PT \\) classes, when handwritten text overlaps with printed text, which single-label three-class formulation cannot handle such cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.471, + 0.878 + ], + "angle": 0, + "content": "Several previous studies have studied the segmentation of handwritten and printed text [6, 10, 20, 27], however, they have inherent limitations. Some focus solely on binary classification, determining if a pixel is handwritten or not [10], whereas others adopt a 3-class formulation of the problem, classifying pixels as handwritten, printed, or background [6, 20, 27]. This exclusive assignment of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "pixels to three different classes paralyzes any machine learning segmentation model to properly detect pixels in the overlapping areas to belong to both handwritten and printed classes. In our usecase depicted in Figure 1, the OCR process achieves insufficient performance due to overlapping of handwritten text and signatures that overlay the printed text. As such, to improve the documents digitization quality, and subsequently a wide range of downstream document understanding and natural language processing (NLP) tasks, it is vital to devise image processing approaches that can understand and properly segment different layers of text, i.e., printed and handwritten text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Models. Prior literature has applied various approaches for the identification and separation of handwritten and printed text. Early approaches [7, 12] formulated the problem as a binary classification task using KNN and SVM and focusing on connected components (CCs) (i.e., groups of pixels). More recently, Li et al. [14] employed conditional random fields (CRFs), with formulating both unary and pairwise potentials for adjacent connected components by leveraging convolutional neural networks (CNNs) architecture for the separation of CCs. The limitation of CC-based approaches is that they determine the class membership for the entire component rather than at the pixel level. Consequently, pixel-level segmentation methods were introduced, leveraging Markov random fields (MRFs) and MLPs [19, 22] for pixel-level classification of PT and HT segmentation. Following the success of encoder-decoder architectures in object segmentation [21], more recent works [10, 6, 20, 27] have predominantly adopted a U-Net based architecture [21], which comprises an encoder-decoder network, for HT and PT segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In addition to our contribution of releasing a manually annotated dataset of legal documents with overlapping text, SignaTR6K, our methodology distinguishes itself from prior works in several significant ways. Firstly, we approach the segmentation problem with a four-class formulation, allowing overlapping pixels to be assigned to a new distinct class (OV), which signifies the presence of both HT and PT layers, leading to enhanced segmentation performance. Secondly, we introduce a novel architecture, the Mixed Feature Model (MFM), that combines a Fine Feature Path (FFP) with a Semantic Segmentation Path (SSP) and improves the performance by capitalizing on both high-level and low-level features. Notably, existing U-Net style architectures [10, 6, 20, 27] are limited to leveraging only the SSP path. Further, we present a new loss function, termed Fusion Loss, that converges faster and is more stable compared to prior losses for HT and PT segmentation task. In addition, we introduce a post-processing heuristic based on Conditional Random Fields (CRFH) to carry out relabeling, resulting in further enhancement of text segmentation performance. Additional details related to background and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Available at https://forms.office.com/r/2a5RDg7cAY." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "583" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.085, + 0.464, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.143, + 0.47, + 0.213 + ], + "angle": 0, + "content": "Figure 2: This figure depicts a real crop from a legal document. (a) The original image containing printed text and overlaying handwritten signature; (b) Printed layer pixels annotated only; (c) Handwritten layer pixels annotated only. Any pixels not present in the printed or handwritten layers are marked as background." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.221, + 0.411, + 0.235 + ], + "angle": 0, + "content": "relevant works can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.245, + 0.269, + 0.261 + ], + "angle": 0, + "content": "3. SignalTR6K Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.473, + 0.765 + ], + "angle": 0, + "content": "Successfully training and testing a segmentation model requires access to high-quality labeled datasets. Due to the limited availability of public data that contains both handwritten and printed text, previous research has predominantly sought to synthetically generate such data. To achieve this, researchers have combined datasets that either exclusively contain printed or handwritten text, or those with non-overlapping text, such as IAM [17], RIMES [1], PRImA [4], CVL [13], Scanned Questionnaire [2], and WGM-SYN [27]. The scanned documents can originate from diverse domains, each possessing its unique fonts, characters, and quality. Whether the original documents having poor quality in the case of archival documents [27], or the scanning process resulted in lower resolution or loss of contrast, these factors also compound the complexities of the text segmentation task. In addition, in some cases errors are introduced during the automated text labeling process [20]. Given these challenges, the absence of a precisely and manually annotated and verified dataset from real sources that contains various patterns of overlapping printed and handwritten text, significantly impedes the supervised machine learning (ML) process. To address this gap, we introduce a new dataset, SignaTR6K (pronounced as Signature 6K), that is derived from 200 original legal documents from Thomson Reuters Legal Content Services [3]. The dataset features overlapping printed and handwritten text and hand-drawn signatures, as depicted in Figure 1. The documents originate from different organizations, including law firms and courts, each having its distinct fonts and document formats. Furthermore, the annotations are made by different individuals, ensuring a diverse range of printed and handwritten styles. Importantly, each document has been manually labeled and verified." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Figure 2 displays an original crop from SignaTR6K, which includes both a signature and the information of the signing party. Due to the presence of personally identifiable information (PIIs), we cropped the images to ensure no visible PIIs. Figure 2a illustrates the original image with both printed text and overlaying handwritten signature. Figure 2b presents only the printed layer pixels, while Figure 2c contains only the handwritten layer pixels that are manually annotated. Any pixels absent from either the printed or" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.24, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 3: Examples from the SignaTR6K dataset, with the top row showing the crops and the bottom row their ground truth annotations. Red: class \\(PT\\), printed, Green: class \\(HT\\), handwritten, and Blue: class \\(BG\\), background." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.35 + ], + "angle": 0, + "content": "handwritten channels are designated as background pixels. It is also evident that in overlapping scenarios, some pixels belong to both handwritten and printed layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Dataset Generation. As hand-annotating a vast collection of real documents is time-consuming and expensive, in order to expand on the size of the dataset and make it adequate for training a deep learning model, we turn to data synthesis and augmentation techniques [10]. The synthetic approaches include general augmentation of crops, along with shifting, magnifying, and rotating operations. We also overlay handwritten and printed pixels from different crops to generate new real-like examples of overlapping text. For this purpose, from the 200 distinct original document samples, we set aside 16 crops, ensuring mutually exclusive samples in the test set. These crops are augmented only for the test set, and we create the training and validation sets from the remaining 184 samples. Following this approach and after excluding generated samples with visible PIDs, we have curated a dataset for training, validation, and testing sets of sizes 5169, 530, and 558 samples, respectively. Each sample is a pair of grayscale crop and its manually annotated ground truth. Each image is 256 by 256 pixels in size, with three channels (RGB) and typically contains several HT and PT overlaps. Four examples from the SignaTR6K dataset are presented in Figure 3, with their grayscale crops and corresponding ground truth (GT) labels. HT is indicated in green, PT in red, and BG in blue. Overlapping HT and PT pixels combine green and red channel values, resulting in a yellow appearance. We envision this dataset can be utilized for model training from scratch or further fine-tuning of a pre-trained classification or segmentation model for specific tasks. The SignaTR6K dataset is freely available for public download." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.82, + 0.759, + 0.837 + ], + "angle": 0, + "content": "4. Approach and Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.842, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this section, we detail our approach, the rationale behind the chosen architecture, the Semantic Segmentation Path (SSP) and Fine Feature Path (FFP), the various loss functions employed, and our novel Fusion loss." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.086, + 0.852, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.337, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Figure 4: Our architecture proposal showing FFP and SSP model paths. The output of FFP and SSP are concatenated prior to the final round of convolutions. A Softmax activation then selects one of the four classes for each pixel. The model's final output consists of pixel-level annotations of the input image. BN: BatchNorm." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.387, + 0.265, + 0.401 + ], + "angle": 0, + "content": "4.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.407, + 0.474, + 0.771 + ], + "angle": 0, + "content": "The prevalent architecture for object segmentation applies a U-Net design [21, 16]. This architecture leverages a Fully Convolutional Network (FCN), that is without fully connected layers that are typically present at the end of Convolutional Neural Networks (CNNs). The U-Net architecture consists of two main parts, an encoder and a decoder. In the encoder segment, the original image (in our context a document crop) is fed into the model. It then undergoes a series of convolutions and max-pooling layers, extracting features from the image and reducing its dimensions. Conversely, the decoder processes the down-sampled image from the encoder using convolutions and up-sampling layers that eventually restores the image back to its original input size with the same number of channels. The final output of the decoder is a pixel-level labeling of the original image. Semantic Segmentation Path (SSP) in Figure 4 shows this part of the fully convolutional network. In addition to the encoder-decoder architecture in U-Net, there are also skip-connections that bring the feature maps directly from an encoder stage to the corresponding decoder stage, i.e., the decoder stage that has the same image size of the encoder stage. These skip-connections improve segmentation performance by re-accessing early-stage features that may be lost in the encoder's final output due to down-sampling." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.775, + 0.3, + 0.79 + ], + "angle": 0, + "content": "4.2. Four-Class Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As mentioned earlier, prior approaches have either considered binary classification of HT or a three-class formulation of the problem. Binary classification detects only one type of text, while the three-class formulation results in overlapping pixels being assigned solely to either the HT or PT class, impairing performance. Therefore, we propose a four-class formulation of the segmentation task," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.509 + ], + "angle": 0, + "content": "and the fourth class, overlap, \\(OV\\), models pixels that belong to both the HT and PT classes, specifically enabling the detection of overlapping areas. To expand on Formula (1) for this four-class formulation, for a given document \\(D\\), assuming there exist four classes as handwritten text \\(HT\\), printed text \\(PT\\), background \\(BG\\), and overlap \\(OV\\), and pixel \\(p_i\\): \\(\\forall p_i \\in D : p_i(c) == True\\) if \\(p_i \\in c\\) and \\(c \\in \\{HT, PT, BG, OV\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.782 + ], + "angle": 0, + "content": "Overlapping pixels are highlighted in yellow in the ground truth, as seen in Figure 3. Our four-class single-label classification employs a Softmax activation function in the final layer, which ensures that only one output for each pixel is activated (Figure 4). Since the output image comprises three channels, when the \\(OV\\) class is predicted, during a post-processing step we turn on pixels for both \\(HT\\) and \\(PT\\) channels, resulting in the yellow color in the output image. In addition to the four-class formulation, we also explored three-class and multi-label formulations. In this scenario, instead of a Softmax activation, we applied Sigmoid activations for each class (i.e., three separate sig-moids). However, formulating the problem as a multi-label classification introduces added complexity and degrees of freedom. This can lead to undesirable scenarios, such as the simultaneous activation of pixels for \\(HT\\) and \\(BG\\). Consequently, the multi-label approach did not yield results comparable to those of the four-class formulation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.79, + 0.807, + 0.807 + ], + "angle": 0, + "content": "4.3. Semantic Segmentation Path (SSP)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The semantic segmentation path (SSP) of our model leverages a U-Net based architecture, with down-sampling in the encoder stages (i.e., backbone) and up-sampling in the decoder stages. The U-net architecture works well in capturing high-level image features. In this architecture, the encoder and decoder maintain a symmetrical architec" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "585" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.395 + ], + "angle": 0, + "content": "ture with a similar number of stages. For example, in the SSP, if the encoder goes through four down-sampling stages, i.e., the input image size changes from \\(256^{*}256\\) to \\(16^{*}16\\) (\\(256 \\rightarrow 128 \\rightarrow 64 \\rightarrow 32 \\rightarrow 16\\)), correspondingly, the decoder undergoes four up-sampling stages, from \\(16^{*}16\\) to \\(256^{*}256\\). Thus, the final output retains the original input image's pixel dimensions. For the SSP, we explored a variety of network sizes. As a baseline and for comparison, we implemented FCN-light [6, 20, 27]. We then improved on the SSP architecture by using VGG16 [23], InceptionV3 [26], and ResNet34 [8] as backbones of the SSP and observed improvement in the performance. In particular, the ResNet34 and InceptionV3 backbones outperform the prior work (FCN) due to their larger number of learnable parameters. Additionally, the inclusion of residual connections and varied convolution sizes allows better carry over the low-level features of text to the later stages of the segmentation network. This observation further inspired us to introduce the FFP network, which similarly incorporates residual connections and without down-sampling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.303, + 0.415 + ], + "angle": 0, + "content": "4.4. Fine Feature Path (FFP)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.473, + 0.812 + ], + "angle": 0, + "content": "The semantic segmentation path excels when segmenting distinct objects. In this path, down-sampling layers are rapidly applied to the input image to capture high-level features and patterns. However, this rapid processing can lead to the loss of fine, or low-level, features. For our application, low-level features are crucial due to the intertwined nature of printed and handwritten text. To address this, we introduce a parallel path to the SSP, termed the fine feature path (FFP), which avoids down-sampling and instead incorporates a convolution block with residual connections. Note that, while the FFP aids in capturing fine features without down-sampling, on its own, i.e., without the SSP that includes down-sampling, it is insufficient for text segmentation, as the absence of high-level features means the model will not be able to detect high-level patterns irrespective of their pixel locations in the image. In Figure 4, the fine feature block of the FFP is repeated \\( N_{x} \\) times; in our implementation, \\( N_{x} = 4 \\). In addition, each stage of the FFP itself implements residual connections as it provides the flexibility to either bypass the block or use its output, leading to improved results, as we will discuss later in the paper. Similar architectures employing residual blocks have shown improved performance in fine object segmentation tasks, e.g., road segmentation [29]. Table 1 details the FFP architecture for four stages \\( (N_{x} = 4) \\) and the connections between layers and convolution sizes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.821, + 0.342, + 0.837 + ], + "angle": 0, + "content": "4.5. Mixed Feature Model (MFM)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The Mixed Feature Model (MFM) serves as the umbrella model containing two parallel paths: SSP and FFP, as depicted in Figure 4. The objective of FM is to capture the image's low-level features alongside its high-level features." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.086, + 0.912, + 0.309 + ], + "angle": 0, + "content": "
GroupLayer typeFilterInput(s)Output(s)Output size
Stage 1G 1Conv3 × 3/64img_inputs1.g1.c.o256 × 256 × 64
BatchNorm-s1.g1.c.os1.g1.b.o256 × 256 × 64
ReLU-s1.g1.b.os1.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s1.g1.r.os1.g2.c.o256 × 256 × 64
BatchNorm-s1.g2.c.os1.g2.b.o256 × 256 × 64
ReLU-s1.g2.b.os1.g2.r.o256 × 256 × 64
Concat-img_in, s1.g2.r.os1.o256 × 256 × 67
Stage 2G 1Conv3 × 3/64s1.os2.g1.c.o256 × 256 × 64
BatchNorm-s2.g1.c.os2.g1.b.o256 × 256 × 64
ReLU-s2.g1.b.os2.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s2.g1.r.os2.g2.c.o256 × 256 × 64
BatchNorm-s2.g2.c.os2.g2.b.o256 × 256 × 64
ReLU-s2.g2.b.os2.g2.r.o256 × 256 × 64
Concat-s1.o, s2.g2.r.os2.o256 × 256 × 131
Stage 3G 1Conv3 × 3/64s2.os3.g1.c.o256 × 256 × 64
BatchNorm-s3.g1.c.os3.g1.b.o256 × 256 × 64
ReLU-s3.g1.b.os3.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s3.g1.r.os3.g2.c.o256 × 256 × 64
BatchNorm-s3.g2.c.os3.g2.b.o256 × 256 × 64
ReLU-s3.g2.b.os3.g2.r.o256 × 256 × 64
Concat-s2.o, s3.g2.r.os3.o256 × 256 × 195
Stage 4G 1Conv3 × 3/64s3.os4.g1.c.o256 × 256 × 64
BatchNorm-s4.g1.c.os4.g1.b.o256 × 256 × 64
ReLU-s4.g1.b.os4.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s4.g1.r.os4.g2.c.o256 × 256 × 64
BatchNorm-s4.g2.c.os4.g2.b.o256 × 256 × 64
ReLU-s4.g2.b.os4.g2.r.o256 × 256 × 64
Concat-s3.o, s4.g2.r.os4.o256 × 256 × 259
Output-Conv1 × 1/4s4_oFFP_o256 × 256 × 4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.31, + 0.895, + 0.393 + ], + "angle": 0, + "content": "Table 1: The architecture of the Fine Feature Path (FFP). In our design, we use four stages (Stages 1-4) of fine feature blocks. We include BatchNorm and ReLu activation after each convolution layer, and unlike SSP, down-sampling layers (e.g., max pooling) are absent from FFP. The output of each stage is concatenated with its input to create residual connections." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.477 + ], + "angle": 0, + "content": "The outputs of SSP and FFP are concatenated before the final layer convolution, producing the output of the MFM model. Additional details regarding the architecture, inputs, outputs, and convolution layer sizes of the MFM can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.485, + 0.849, + 0.502 + ], + "angle": 0, + "content": "4.6. CRF Post-Processing and CRF Heuristic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Prior research [6, 20, 27] has leveraged dense Conditional Random Fields (CRFs) as a post-processing step to re-label pixels based on their neighboring pixels. In our architectural exploration, we found that, while CRF post-processing is intended to improve segmentation performance, in practice it often hurts the segmentation performance by aggressively re-labeling pixels to incorrect classes. One issue arises due to the imbalanced nature of pixels across classes, with background pixels being predominant. Consequently, many pixels are mistakenly annotated from HT or PT to BG, or from HT to PT, which is undesirable. Based on this observation, we designed a heuristic for CRF post-processing to contain this unfavourable behaviour by only allowing the BG pixels to be relabelled to HT or PT pixels and not vice versa. As we will discuss in the experimentation section, this heuristic further improves segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.655, + 0.787 + ], + "angle": 0, + "content": "4.7. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Due to the nature of the scanned documents and the amount of text on each page, the number of background pixels (i.e., white blank pixels) surpasses that of handwritten and printed pixels. This leads to a class imbalance problem [30], posing the risk of predicting the majority of pixels as background while still minimizing the loss value. As a result, we investigated various loss functions and weight as" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "586" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "signments to different classes in the loss function to evaluate its impact on segmentation performance. Additionally, during this exploration, we observed that different loss functions achieve different IoU (Intersection over Union) scores. Consequently, we introduce a new loss function, Fusion Loss, to incorporate the benefits of various losses. In the following, we discuss the different loss functions we explored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.473, + 0.414 + ], + "angle": 0, + "content": "Cross-Entropy Loss. The multi-class cross-entropy loss is used for classification tasks involving more than two classes. It assumes that for each data point, only one class can be the ground truth, i.e., multi-class single-label classification. The standard form of cross-entropy loss for a single data point in multi-class segmentation with \\( M \\) classes is defined as: \\( \\mathcal{L}_{CE}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot \\log(pr) \\), where \\( pr \\) and \\( gt \\) represent the prediction and ground truth, respectively. The final loss is the average of the losses for all data points in a training set. Similarly, the weighted version (WCE) is computed as \\( \\mathcal{L}_{WCE}(gt, pr) = -\\sum_{m=1}^{M} w_m \\cdot gt \\cdot \\log(pr) \\), with \\( w_m \\) being the weight assigned to each class in the loss calculation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.473, + 0.673 + ], + "angle": 0, + "content": "Focal Loss. The focal loss [15] was introduced to address the class-imbalance problem during training of object detection tasks. The focal loss puts the focus of the learning algorithm on the incorrectly classified examples by applying a modulating term \\((1 - pr)^{\\gamma}\\) to the cross-entropy loss. This scaling factor dynamically weighs down the contribution of easy and correctly classified samples, allowing the training loop to concentrate on the harder and incorrectly classified samples. The focal loss is calculated as: \\(\\mathcal{L}_{Focal}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot (1 - pr)^{\\gamma} \\log(pr)\\), where \\(\\gamma\\) is the modulating/focusing factor, and we set \\(\\gamma == 2\\) as per the original paper. Similarly, the weighted version of the focal loss can be expressed as: \\(\\mathcal{L}_{WF}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot \\alpha_m \\cdot (1 - pr)^{\\gamma} \\log(pr)\\), where \\(\\alpha_m\\) is analogous to \\(w_m\\), assigning weights to each class such that: \\(\\forall \\alpha_m \\mid 0 < \\alpha_m < 1 \\& \\& \\sum_{m=1}^{M} \\alpha_m = 1\\)." + }, + { + "type": "table", + "bbox": [ + 0.076, + 0.682, + 0.473, + 0.798 + ], + "angle": 0, + "content": "
DiceLoss.Dicelossisgivenby
L_Dice(precision, recall)= 1-2/M∑m=1M
precisionm·recallm/precisionm+recallm
Intuitively, it aims to maximize the F-Score while minimizing the loss, i.e., L_Dice = 1 - F_Score. Accordingly, the weighted version of the dice loss is defined as L_WD(precision, recall) = 1 - 2/M ∑m=1M w_m·precision_m·recall_m /precision_m+recall_m.
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Fusion Loss. Given our observation that different loss functions adversely affect different classes, we introduce a new loss targeted for maximizing the performance of all the classes. For example, the dice loss performs better on the background class, as in its formulation, it aims to maximize the F-Score. Because the majority of pixels are attributed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "to the background class, having higher values of correctly classified background pixels achieves a higher F-Score and lower loss values. However, this does not necessarily yield higher IoU scores for PT and HT classes. In contrast, the weighted versions of the cross-entropy and focal losses focus on the handwritten and printed classes. As such, with the Fusion loss, we aim to combine the behaviors of various losses, and we define the Fusion loss as the sum of the three weighted losses: \\(\\mathcal{L}_{\\text {Fusion }} = \\mathcal{L}_{WF} + \\mathcal{L}_{WCE} + \\mathcal{L}_{WD}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.243, + 0.737, + 0.261 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.895, + 0.53 + ], + "angle": 0, + "content": "Evaluation Metric. Intersection over Union (IoU) is a commonly used metric to measure the performance of a segmentation task [28], and in particular text segmentation [25, 10, 6, 27]. In our context, we use IoU to measure the pixel-level performance of the model output versus the ground truth. For a class \\( c \\), let \\( TP_{c} \\) represent the correctly predicted pixels for that class, \\( FP_{c} \\) denote the pixels incorrectly predicted as belonging to class \\( c \\), and \\( FN_{c} \\) represent the pixels that are incorrectly not predicted for class \\( c \\), then Intersection over Union, IoU, for class \\( c \\) is calculated as: \\( IoU_{c} = \\frac{TP_{c}}{TP_{c} + FP_{c} + FN_{c}} \\). For each experiment, we calculate the IoU for all three classes, HT, PT, and BG. In the four-class formulation, pixels attributed to the overlap (OV) class are converted in a post-processing step to both HT and PT classes, contributing to the IoU calculations for both. We also calculate the mean IoU, which is the average of IoUs for these three classes: \\( IoU_{Mean} = \\frac{IoU_{HT} + IoU_{PT} + IoU_{BG}}{3} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.797 + ], + "angle": 0, + "content": "Experiments. We perform 50 epochs of training with a batch size of 8. For problem formulation, we use both three-class and four-class implementations. The experiments are conducted on two datasets: SignaTR6K and WGM-SYN [27]. The IoU values for FCN-light and WGM-MOD are as reported in [27] for the three-class formulation which has been performed on the same dataset, WGM-SYN. However, for the SignaTR6K dataset, we retrain an FCN-based model for three-class formulation to ensure a fair comparison on the new dataset. We run experiments for a variety of architectures and in three post-processing configurations. We structure experiments incrementally, i.e., adding one improvement at a time. This approach aids in understanding and isolating the impact of each added improvement, i.e., ablation study. Due to space constraints, the table detailing the complete list of experiment parameters and configurations is relegated to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Experimented Architectures. We start with FCN-light which has approximately 295K parameters. This architecture aligns with prior research and serves as our baseline [6, 27]. Next, we improve the performance of FCNlight by adopting a four-class implementation. We then experiment with SSP architecture with three different back" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "587" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.892, + 0.432 + ], + "angle": 0, + "content": "
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-light [6]WGM-MOD [27]~295K~295KWeighted CE24.0026.0072.0041.0011.0023.0072.0036.00----
Weighted CE42.0036.0074.0050.00*41.0032.0074.0049.00----
4-Class(Ours)FCN-light~295KCE46.4941.9873.7754.0838.9529.3573.9247.4146.5741.5973.8754.01
Focal46.0242.0173.8053.9532.9624.5074.0043.8246.1141.2373.9853.77
Dice48.0147.2871.2255.5548.5948.4871.2256.1048.0747.7971.2955.72
Weighted CE46.0742.1873.8254.0240.5430.0773.9848.2046.3141.6873.9553.98
Weighted Focal43.7141.7773.9553.1432.6724.4074.0743.7144.0641.1274.1153.10
Weighted Dice47.9747.2271.0255.4048.6948.3471.0256.0248.0447.7071.1055.61
Fusion48.1247.2571.9355.7746.3444.2571.8654.1548.1947.8972.3856.15
SSP - VGG16~24MCE35.1432.4073.9647.1730.9719.1473.8841.3335.2132.0874.1347.14
Focal34.5632.2274.0246.9329.1119.4273.9140.8134.6731.7974.2246.89
Dice40.6439.6471.1350.4743.3040.1671.2151.5540.8140.2571.2550.77
Weighted CE35.7333.4474.4347.8732.0020.7074.1242.2835.9332.9574.6547.84
Weighted Focal34.9034.3574.3647.8729.1820.7274.0641.3235.1133.8074.5647.82
Weighted Dice41.8040.6470.6751.0344.1441.3970.7052.0742.5241.1970.9751.56
Fusion39.9939.5672.3150.6235.3927.9372.3545.2240.1039.9472.6350.89
MFM (FFP + SSP) - InceptionV3~30MCE52.5144.0973.9156.8446.0335.9474.2152.0652.5542.1774.3556.36
Focal51.5643.7473.8956.4035.8927.1574.4145.8251.6341.7474.4355.93
Dice20.9119.2881.5240.5725.5124.6179.9443.3520.9219.2981.5440.58
Weighted CE51.5843.7273.9556.4244.7834.0474.3251.0551.6441.8574.4055.96
Weighted Focal51.0143.6273.8956.1835.7626.6074.2245.5351.0841.6074.3555.68
Weighted Dice22.3021.2580.7341.4229.1129.4877.9745.5222.3221.2680.7441.44
Fusion53.2251.2571.8458.7750.0548.8372.0356.9753.3249.8372.7258.62
MFM (FFP + SSP) - ReNet34~24MCE52.4044.0273.9156.7847.5236.4074.1252.6852.4342.1274.3456.30
Focal51.9543.9873.8956.6136.7327.1774.3146.0751.9041.9974.3356.08
Dice29.6829.4377.0645.3931.6225.3975.1544.0529.8429.6377.0745.52
Weighted CE52.6344.1273.9456.9047.6037.2574.1252.9952.5642.2274.3656.38
Weighted Focal51.6843.9573.9056.5135.7327.2274.3045.7551.7241.9874.3456.01
Weighted Dice51.4450.6071.9658.0052.9751.4271.9058.7651.5449.3772.7557.89
Fusion52.9951.7271.6958.8051.1550.7971.8357.9253.0151.2972.4958.93
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.433, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Table 2: IoU performance (%) on the WGM-SYN dataset [27]. The maximum value of a column (representing a class) is underlined. The overall maximum of a class across different post-processing methods is both bolded and underlined. For example, the best mean IoU for the WGM-SYN dataset is achieved with Fusion loss, using CRFH, and the MFM-ResNet34 architecture, stands at 58.93. The best performing configuration from prior work, denoted with \\((^{*})\\), stands at 50.00." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.47, + 0.587 + ], + "angle": 0, + "content": "bones: VGG16, InceptionV3, and ResNet34. Our experimentation also includes three variations of MFM, which incorporates both FFP and SSP, and with three SSP variations: VGG16, InceptionV3, and ResNet34. For our implementations, we use the Segmentation Models Library available in TensorFlow [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Post-Processing Configurations. We experiment with three distinct configurations: without post-processing, with CRF post-processing, and with CRFH post-processing. In the scenario without post-processing, the IoU calculation is performed directly on the output of the model (FCN-light, SSP, and MFM). For the CRF post-processing approach, all classes are permissible for relabeling. However, with the CRF post-processing with heuristic (CRFH), as explained in Section 4.6, only the background class pixels (BG) are allowed to be relabeled as HT or PT classes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Weight Initialization. For both FCN-light and SSP models, model weights start from their initial random values at the beginning of the training. We train MFM configurations at last to reuse the weights from the SSP configurations. Thus, for the MFM trainings, we apply transfer learning by initializing with weights from the corresponding, previously trained SSP. For example, when training an MFM (FFP+SSP) with an SSP using the ResNet34 backbone, we initialize its SSP path with the available weights from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.573 + ], + "angle": 0, + "content": "SSP - ResNet34. Lastly, although we have aimed for consistent IoU results by fixing the random seeds, the inherent non-determinism associated with GPU execution remains. As such, we have run our experiments multiple times to validate the IoU value improvements across configurations." + }, + { + "type": "text", + "bbox": [ + 0.496, + 0.585, + 0.892, + 0.797 + ], + "angle": 0, + "content": "Results. Tables 2 and 3 show the IoU values for the WGM-SYN and SignaTR6K datasets. The overall trend across the results indicates that transitioning from the three-class formulation to the four-class formulation improves the IoU scores. Furthermore, employing larger model backbones generally, but not always, improves the segmentation performance. Among all the model architectures evaluated, the ResNet34 and InceptionV3 backbones achieve the highest performance, which we attribute to their residual connections and different size convolutions as they can better capture fine features from the image. We provide a summary of the trends in the results here, while a more detailed version of the result tables can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.491, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "- Prior work [6, 27], which implements the three-class formulation of HT and PT segmentation, generally exhibits the lowest performance. Transitioning from the three-class to the four-class formulation with the same backbone, i.e., FCN-light, with the same number of model parameters and WCE loss, improves the average IoU values by" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "588" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.085, + 0.892, + 0.405 + ], + "angle": 0, + "content": "
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-based [6, 27]~295KCE62.5688.0998.4083.02*52.6889.6899.2680.4662.7290.5899.0584.11
Focal62.3488.0098.4582.9344.6084.8699.2876.2562.5790.8399.2184.21
Dice60.4587.2997.8581.8661.2488.8398.1782.7460.5287.8897.9782.12
Weighted CE60.5884.8997.7881.0953.4590.7399.5281.2360.8486.3398.2481.80
Weighted Focal61.2585.7498.0281.6744.2785.7799.5076.5161.5587.4298.5882.52
Weighted Dice60.2187.0097.7481.6560.7288.1797.9882.2960.2787.4697.8381.86
Fusion61.5288.3998.3882.7658.9492.5599.3783.6261.6790.4598.9183.68
4-Class(Ours)FCN-based~295KCE64.5589.2198.3984.0554.6089.6899.2381.1764.8791.8199.0685.25
Focal64.1088.8698.3283.7646.6486.0199.2677.3064.3491.6599.1185.03
Dice64.3788.6898.3783.8165.1790.1398.5984.6364.5789.1498.4684.06
Weighted CE63.7887.5298.1683.1554.7790.5099.4281.5664.1289.4398.7184.09
Weighted Focal63.6887.7198.1983.2048.1887.1099.4578.2464.0589.7298.8084.19
Weighted Dice64.2188.5798.3083.6965.0890.1098.5384.5764.4489.0698.3983.96
Fusion64.6888.4898.3383.8359.8391.7799.2883.6365.0090.7298.9184.87
MFM (FFP + SSP) - InceptionV3~30MCE73.1092.6698.7788.1863.4892.8999.5585.3173.0594.8999.3689.10
Focal72.8092.5098.7588.0157.4790.5899.5482.5372.7794.7499.3588.95
Dice72.5692.2098.7087.8272.3893.6099.0088.3372.5294.7399.3788.87
Weighted CE72.6691.9098.5487.7062.6593.0499.6285.1072.6393.3198.9488.29
Weighted Focal72.5592.1398.6287.7759.3591.6499.5983.5372.5093.7799.0788.45
Weighted Dice72.5992.3198.7087.8772.6094.0099.0788.5672.5494.6099.3288.82
Fusion72.5592.2598.7187.8365.5793.4999.5386.1972.4994.6299.3588.82
MFM (FFP + SSP) - ResNet34~24MCE72.8192.5698.7888.0563.0492.9499.5585.1772.7594.9399.3989.02
Focal73.0492.4698.7588.0853.0289.0499.5580.5473.0094.7599.3589.03
Dice72.9692.3898.7288.0273.1693.3598.9388.4872.9394.7999.3689.03
Weighted CE72.9691.9698.6987.8764.4992.4799.5685.5172.9094.1999.2788.79
Weighted Focal73.1892.1098.6387.9754.9689.7199.6081.4273.1693.7299.0688.65
Weighted Dice72.7892.3298.7187.9472.9593.6699.0088.5372.7594.6799.3488.92
Fusion73.2692.4598.7388.1568.3894.8599.5687.6073.2194.5999.3189.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.406, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Table 3: IoU performance (%) on the SignaTR6K dataset. The maximum value of a column (i.e., class) is underlined, and the overall maximum for a class across different post-processing is both bolded and underlined. For example, the best mean IoU for the SignaTR6K dataset is for CE loss, with CRFH, and the MFM-InceptionV3 architecture at 89.10. The best performing configuration of prior work, i.e., excluding results for Fusion loss and CRFH, is marked with \\((^{*})\\) and is 83.02." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.478, + 0.273, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.572, + 0.242, + 0.582 + ], + "angle": 0, + "content": "(a) Validation set losses" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.478, + 0.463, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.572, + 0.43, + 0.583 + ], + "angle": 0, + "content": "(b) Validation set IoUs" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.584, + 0.468, + 0.611 + ], + "angle": 0, + "content": "Figure 5: Convergence speed of different loss functions and normalized loss values on the validation set of the SignaTR6K dataset." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.652 + ], + "angle": 0, + "content": "\\(8.0\\%\\) \\((50\\rightarrow 54.02)\\) for the WGM-SYN dataset, and \\(2.6\\%\\) \\((81.09\\rightarrow 83.15)\\) for the SignaTR6K dataset." + }, + { + "type": "text", + "bbox": [ + 0.068, + 0.656, + 0.469, + 0.852 + ], + "angle": 0, + "content": "- Applying CRF post-processing generally degrades the results. In our research, we observe that CRF post processing is dataset dependent and decreases the performance of segmentation in some cases. CRF post-processing relabels aggressively and incorrectly converts HT and PT pixels to background ones, or OV pixels to PT ones. For example, Figure 6i illustrates that how OV pixels are wrongly relabeled to PT class by CRF post-processing. Contrary to CRF post-processing, our CRF heuristic generally improves the IoU values. For example, for the SignaTR6K dataset, for FCN-light with four-class implementation and CE loss, CRFH improves HT IoUs by \\(2.9\\%\\) (89.21→91.81)." + }, + { + "type": "text", + "bbox": [ + 0.068, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Moving from the FCN-light architecture to larger models, i.e., SSP and MFM models, generally improves the results, except for the VGG16 backbone. For VGG16, we ratio" + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.656, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.48, + 0.892, + 0.602 + ], + "angle": 0, + "content": "nalize that having a deep network without residual connections or varied-size convolutions, as seen in ResNet34 and InceptionV3 architectures, hurts the model's ability in capturing fine features. In addition, we observe that adding FFP helps to improve the performance of the VGG model as we compare SSP and MFM performance values for both datasets. This also confirms the FFP's ability to capture low-level features that are missed by the SSP path." + }, + { + "type": "text", + "bbox": [ + 0.491, + 0.607, + 0.892, + 0.759 + ], + "angle": 0, + "content": "- The MFM model with ResNet34 and InceptionV3 backbones generally achieves the best results. MFM-ResNet34 with Fusion loss and CRFH achieves the best result for the WGM-SYN dataset (58.93%), while MFM-InceptionV3 with CE loss and CRFH achieves the best mean IoUs for the SignaTR6K dataset (89.10%). Overall, the best-performing model from our designs improves on the mean IoU performance of the best-performing prior work [27] by \\(17.9\\%\\) (50→58.93) and \\(7.3\\%\\) (83.02→89.10) for the WGM-SYN and SignaTR6K datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.491, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- Although using skip connections in SSP helps with the segmentation performance and thus less pronounced improvement is observed in MFM, our FFP design is agnostic of the SSP implementation. This distinction becomes clear when VGG16 is employed as SSP; the addition of FFP boosts the IoU scores by \\(7\\%\\) and \\(8\\%\\) on the WGM-SYN and SignaTR6K datasets, respectively. The FCN-based model with 295K parameters outperforms VGG16 with 30M parameters. Thus, larger models do not necessarily outper" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "589" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.184, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.168, + 0.14, + 0.176 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.089, + 0.284, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.167, + 0.241, + 0.176 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.385, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.167, + 0.341, + 0.176 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.089, + 0.486, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.167, + 0.444, + 0.175 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.089, + 0.587, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.167, + 0.544, + 0.176 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.089, + 0.688, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.167, + 0.645, + 0.176 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.089, + 0.789, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.167, + 0.747, + 0.176 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.089, + 0.891, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.167, + 0.848, + 0.176 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.177, + 0.182, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.254, + 0.139, + 0.263 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.177, + 0.284, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.254, + 0.241, + 0.263 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.177, + 0.385, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.254, + 0.341, + 0.263 + ], + "angle": 0, + "content": "(k)" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.177, + 0.486, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.254, + 0.443, + 0.263 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.177, + 0.587, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.254, + 0.545, + 0.263 + ], + "angle": 0, + "content": "(m)" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.177, + 0.688, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.254, + 0.646, + 0.263 + ], + "angle": 0, + "content": "(n)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.177, + 0.789, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.254, + 0.747, + 0.263 + ], + "angle": 0, + "content": "(0)" + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.177, + 0.891, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.254, + 0.848, + 0.263 + ], + "angle": 0, + "content": "(p)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.894, + 0.334 + ], + "angle": 0, + "content": "Figure 6: Example results on the test set of the SignaTR6K dataset for our approach compared to the ground truth and prior works. (a) Input image; (b) Ground truth; (c) & (d) 3-class FCN-based [6, 27] with CE loss without (c) and with (d) CRF post-processing; (e), (f), & (g) Our FCN-based 4-class formulation with CE loss without CRF (e), with CRF (f), and with CRFH (g); (h), (i), & (j) SSP-ResNet34 with CE loss without CRF (h), with CRF (i), and with CRFH (j); (k), (l), & (m) MFM-ResNet34 with CE loss without CRF (k), with CRF (l), and with CRFH (m); (n), (o), & (p) MFM-ResNet34 with Fusion loss without CRF (n), with CRF (o), and with CRFH (p)." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.345, + 0.47, + 0.421 + ], + "angle": 0, + "content": "form smaller ones consistently, and integrating FFP can bring significant improvements in IoU scores. Our findings show that, irrespective of the SSP size and architecture, incorporating FFP improves the performance, and it is essential for segmentation tasks involving fine objects, e.g., text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Loss Functions. We also perform an analysis to confirm the stability and convergence of the Fusion loss compared to other losses. Figures 5a and 5b compare different loss functions on the validation set of the SignaTR6K dataset. In Figure 5a, we observe that the Fusion loss is generally stable while converging to its minimum loss. Additionally, in Figure 5b, we observe Fusion loss reaches its maximum IoU around Epoch 15, generally faster than most other losses and stays close to the maximum IoU values on the validation set. It is important to highlight that the maximum IoU on the validation set for Focal and CE losses does not necessarily imply the best performance on the test set. Overall, Fusion loss shows a stable behavior and achieves a better performance on the test set, whereas for comparison, Focal loss shows instability at some epochs during the training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.468, + 0.829 + ], + "angle": 0, + "content": "Visual Comparison. Figure 6 shows sample model outputs for the SignaTR6K dataset, with two rectangle regions of interest highlighted to showcase the differences between various models. Figure 6b shows the ground truth, and a visual trend indicates that the performance on the regions of interest improves from Figure 6c to Figure 6p. It is also visually noticeable that CRF post-processing aggressively relabels pixels (6i), whereas CRFH generally improves the results. More visual comparisons on the WGM-SYN and SignaTR6K datasets can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Limitations. Compared to prior work, the complexity and training cost of our approach present some limitations. Our MFM model, being larger than FCN-light, requires greater GPU memory sizes for training, and takes longer to train." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.542 + ], + "angle": 0, + "content": "However, we believe some of the limitations on the training time can be offset through transferring weights from SSP models to MFM ones. Additionally, while our approach outperforms prior work on the WGM-SYN dataset, the mean IoU performance is lower compared to the SignaTR6K dataset. This indicates potential limitations in our method's efficacy for lower-quality original documents, those undergoing low-quality scanning processes (like historical documents), or those with errors from automated labeling. We also attribute the improved performance on the SignaTR6K dataset to its higher quality and our thorough manual annotation, that have resulted in better model training and improved IoU results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.619, + 0.57 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.893 + ], + "angle": 0, + "content": "Segmentation of handwritten text (HT) and printed text (PT) is vital for digitization and understanding of scanned documents. The complexity increases with the overlapping of different text types. In this research, we introduced SignaTR6K, a new open-source dataset with high-quality, manual pixel-level annotations, sourced from original legal documents. Additionally, we proposed a novel four-class formulation and a new architecture for the segmentation task. Our design leverages both the Fine Feature Path (FFP) and the Semantic Segmentation Path (SSP) to create the Mixed Feature Model (MFM), and incorporates both high-level and low-level features and improves on the text segmentation performance. We also introduced a CRF-based post-processing heuristic (CRFH) that further improves the model output, and included a new loss function, Fusion loss, that combines the advantages of different loss functions and achieves faster convergence and higher stability compared to most of the evaluated losses. In conclusion, our designs outperform the prior work in mean IoU scores by \\(17.9\\%\\) and \\(7.3\\%\\) on WGM-SYN and SignaTR6K datasets, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "590" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.468, + 0.159 + ], + "angle": 0, + "content": "[1] RIMES. http://www.a2ialab.com/doku.php?id=rimesdatabase:start. Inaccessible:2023-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.164, + 0.468, + 0.193 + ], + "angle": 0, + "content": "[2] Scanned Questionnaire. https://github.com/jottue/HTSNet. Accessed: 2023-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.197, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[3] Thomson Reuters Legal Content Services. https://www.thomsonreuters.com/en/ products-services/legal.html. Accessed: 2023-01-21." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.26, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[4] Apostolos Antonacopoulos, David Bridson, Christos Papadopoulos, and Stefan Pletschacher. A realistic dataset for performance evaluation of document layout analysis. In 2009 10th International Conference on Document Analysis and Recognition, pages 296-300. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.353, + 0.468, + 0.426 + ], + "angle": 0, + "content": "[5] Dibyajyoti Dhar, Avishek Garain, Pawan Kumar Singh, and Ram Sarkar. Hp_docpres: A method for classifying printed and handwritten texts in doctor's prescription. Multimedia Tools and Applications, 80(7):9779-9812, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.431, + 0.468, + 0.52 + ], + "angle": 0, + "content": "[6] Nicolas Dutly, Fouad Slimane, and Rolf Ingold. PHTI-WS: a printed and handwritten text identification web service based on FCN and CRF post-processing. In 2019 International Conference on Document Analysis and Recognition Workshops (IC-DARW), volume 2, pages 20-25. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.524, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[7] Jürgen Franke and Matthias Oberlander. Writing style detection by statistical combination of classifiers in form reader applications. In Proceedings of 2nd International Conference on Document Analysis and Recognition (ICDAR'93), pages 581-584. IEEE, 1993." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.617, + 0.468, + 0.676 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.68, + 0.468, + 0.724 + ], + "angle": 0, + "content": "[9] Pavel Iakubovskii. Segmentation models. https://github.com/qubvel/segmentation_models, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[10] Junho Jo, Hyung Il Koo, Jae Woong Soh, and Nam Ik Cho. Handwritten text segmentation via end-to-end learning of convolutional neural networks. *Multimedia Tools and Applications*, 79(43):32137-32150, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.806, + 0.468, + 0.907 + ], + "angle": 0, + "content": "[11] Celso AML Junior, Matheus Henrique Marques da Silva, Byron Leite Dantas Bezerra, Bruno José Torres Fernandes, and Donato Impedovo. FCN+ RL: a fully convolutional network followed by refinement layers to offline handwritten signature segmentation. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.907 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.196 + ], + "angle": 0, + "content": "[12] R Kandan, Nirup Kumar Reddy, KR Arvind, and AG Ramakrishnan. A robust two level classification algorithm for text localization in documents. In Advances in Visual Computing: Third International Symposium, ISVC 2007, Lake Tahoe, NV, USA, November 26-28, 2007, Proceedings, Part II 3, pages 96-105. Springer, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.201, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[13] Florian Kleber, Stefan Fiel, Markus Diem, and Robert Sablatnig. CVL-database: An off-line database for writer retrieval, writer identification and word spotting. In 2013 12th international conference on document analysis and recognition, pages 560-564. IEEE, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[14] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Printed/handwritten texts and graphics separation in complex documents using conditional random fields. In 2018 13th IAPR International Workshop on Document Analysis Systems (DAS), pages 145-150. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.45 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[16] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[17] U-V Marti and Horst Bunke. The IAM-database: an English sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1):39-46, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.598, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[18] Milda Norkute, Nadja Herger, Leszek Michalak, Andrew Mulder, and Sally Gao. Towards explainable ai: Assessing the usefulness and impact of added explainability features in legal document summarization. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-7, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.708, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[19] Xujun Peng, Srirangaraj Setlur, Venu Govindaraju, and Ramachandrula Sitaram. Handwritten text separation from annotated machine printed documents using markov random fields. International Journal on Document Analysis and Recognition (IJDAR), 16(1):1-16, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[20] Anastasia Prikhodina. Handwritten and printed text separation for historical documents. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.835, + 0.892, + 0.88 + ], + "angle": 0, + "content": "[21] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "591" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.114, + 0.092, + 0.468, + 0.122 + ], + "angle": 0, + "content": "image computing and computer-assisted intervention, pages 234-241. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.125, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[22] Mathias Seuret, Marcus Liwicki, and Rolf Ingold. Pixel level handwritten and printed content discrimination in scanned documents. In 2014 14th International Conference on Frontiers in Handwriting Recognition, pages 423-428. IEEE, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.203, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.469, + 0.311 + ], + "angle": 0, + "content": "[24] Stavroula Skylaki, Ali Oskooei, Omar Bari, Nadja Herger, and Zac Kriegman. Named entity recognition in the legal domain using a pointer generator network. arXiv preprint arXiv:2012.09936, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.314, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[25] Nishant Subramani, Alexandre Matton, Malcolm Greaves, and Adrian Lam. A survey of deep learning approaches forOCR and document understanding. arXiv preprint arXiv:2011.13534, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the in-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "ception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.141, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[27] Mahsa Vafaie, Oleksandra Bruns, Natasja Pilz, Jörg Waitelonis, and Harald Sack. Handwritten and printed text identification in historical archival documents. In Archiving Conference, volume 19, pages 15–20. Society for Imaging Science and Technology, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[28] Zhaobin Wang, E Wang, and Ying Zhu. Image segmentation evaluation: A survey of methods. Artificial Intelligence Review, 53:5637-5674, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.269, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[29] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual U-Net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.394 + ], + "angle": 0, + "content": "[30] Zhi-Hua Zhou and Xu-Ying Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on knowledge and data engineering, 18(1):63-77, 2005." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "592" + } + ] +] \ No newline at end of file diff --git a/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_origin.pdf b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..feb2580e646a2120fec5525d28fcdd5b2d0c1ae4 --- /dev/null +++ b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/84e181a3-72a0-4a74-a144-02e187b2dd2f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cda62e8146b9014d3878b39df3e406f9ade7d5bdc6bdf33917aa298af3814a7 +size 818575 diff --git a/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/full.md b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b389694779f421a54eabb0730c04f1ed7bbf18cc --- /dev/null +++ b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/full.md @@ -0,0 +1,270 @@ +# Handwritten and Printed Text Segmentation: A Signature Case Study + +Sina Gholamian +Thomson Reuters AI Labs +Toronto, Canada + +sina.gholamian@thomsonreuters.com + +Ali Vahdat +Thomson Reuters AI Labs +Toronto, Canada + +ali.vahdat@thomsonreuters.com + +# Abstract + +While analyzing scanned documents, handwritten text can overlap with printed text. This overlap causes difficulties during the optical character recognition (OCR) and digitization process of documents, and subsequently, hurts downstream NLP tasks. Prior research either focuses solely on the binary classification of handwritten text or performs a three-class segmentation of the document, i.e., recognition of handwritten, printed, and background pixels. This approach results in the assignment of overlapping handwritten and printed pixels to only one of the classes, and thus, they are not accounted for in the other class. Thus, in this research, we develop novel approaches to address the challenges of handwritten and printed text segmentation. Our objective is to recover text from different classes in their entirety, especially enhancing the segmentation performance on overlapping sections. To support this task, we introduce a new dataset, SignaTR6K, collected from real legal documents, as well as a new model architecture for the handwritten and printed text segmentation task. Our best configuration outperforms prior work on two different datasets by $17.9\%$ and $7.3\%$ on IoU scores. The SignaTR6K dataset is accessible for download via the following link: https://forms.office.com/r/2a5RDg7cAY. + +# 1. Introduction + +For various purposes, the digitization of hard-copy documents and associated challenges are an active area of research in both academia [25, 11, 27, 10, 6, 20] and industry [18]. This digital transformation involves scanning paper documents through an OCR process, making their text accessible for downstream natural language processing (NLP) tasks, such as named entity recognition (NER). The documents of interest can originate from a variety of domains, including historical documents [27], legal and court-issued documents [18], business contracts [25], and medical records and prescriptions [5]. Although various studies, as cited above, have been conducted, there is yet a considerable gap between current approaches and + +![](images/6eacb8b0ae731321b16ca49a0f5af1630a146e8148ec1e190575361d38647ff6.jpg) +Figure 1: Court documents are first printed and then signed or annotated by various parties. This results in handwritten text overlapping the underlying printed information, leading to performance degradation in downstream tasks, such as named entity recognition (NER). This is a fabricated example, however very similar to the original documents, to protect personally identifiable information (PII). + +human-level performance in mixed-text scenarios (i.e., when handwritten and printed text overlap). For instance, attorneys frequently sign legal documents, resulting in their signatures overlapping with their information. This overlap hampers the performance of OCR tools in character recognition, subsequently making it challenging for downstream tasks to accurately identify information linked to the attorneys and their associated law firms. Figure 1 provides an illustration of handwritten text (HT) overlapping with printed text (PT) in court documents. Extracting parties names is a crucial step in the named-entity recognition (NER) task for legal and court documents [24]. When attorneys and other involved parties sign these documents, which are later scanned using OCR tools, their signatures often obscure the details of names and law firms. Consequently, the semantic segmentation of handwritten elements, such as lawyers' signatures and accompanying handwritten notes, and printed text detailing the lawyer and the law firm's information, becomes vital. + +In this research, we aim to address the challenges of HT and PT segmentation, as there is still a large gap between human-level performance and the existing approaches + +for this task. Our focus in this effort is to improve the segmentation performance in overlapping regions for scanned legal documents, and to aid in this endeavor, we also introduce a new dataset. In summary, our research makes the following contributions: + +- We introduce a new dataset, SignaTR6K (pronounce as Signature 6K)1 derived from 200 pixel-level manually annotated crops of images from genuine legal documents. The dataset comprises signatures, handwritten text, and printed text, which frequently overlap. With data augmentation, we have created a dataset of sizes 5169, 530, 558, for training, validation, and testing, respectively, that we release to the public, to facilitate dataset availability for future research and to aid in the training and evaluation of deep-learning segmentation models. +- We propose a novel architecture that integrates both semantic segmentation features and fine features, enhancing the performance of text segmentation over previous methods. Moreover, we introduce a new loss function termed Fusion loss, that, comparatively, is stable and converges to optimal loss values and Intersection over Union (IoU) scores. +- Lastly, we conduct an extensive quantitative and visual evaluation of different variations of our approach against prior work on two distinct datasets, and illustrate our approach's superior performance in the text segmentation task, especially in challenging scenarios where printed and handwritten text overlap. + +# 2. Background and Literature Review + +The text segmentation problem is defined as follows: given a scanned document possibly containing handwritten, printed, and background (i.e., blank) pixels, the task is to assign each pixel to its appropriate class. Scanned documents can originate from various sources, such as hard-copy paper documents or microfilms [20, 27]. Formally, for a given document $D$ , assuming there exist three classes as handwritten text $HT$ , printed text $PT$ , and background $BG$ , and pixel $p_i$ : $\forall p_i \in D : p_i(c) == \text{True if } p_i \in c \& \& c \in \{HT, PT, BG\}$ (1). A situation may arise where a pixel belongs to two classes, i.e., $HT$ and $PT$ classes, when handwritten text overlaps with printed text, which single-label three-class formulation cannot handle such cases. + +Several previous studies have studied the segmentation of handwritten and printed text [6, 10, 20, 27], however, they have inherent limitations. Some focus solely on binary classification, determining if a pixel is handwritten or not [10], whereas others adopt a 3-class formulation of the problem, classifying pixels as handwritten, printed, or background [6, 20, 27]. This exclusive assignment of + +pixels to three different classes paralyzes any machine learning segmentation model to properly detect pixels in the overlapping areas to belong to both handwritten and printed classes. In our usecase depicted in Figure 1, the OCR process achieves insufficient performance due to overlapping of handwritten text and signatures that overlay the printed text. As such, to improve the documents digitization quality, and subsequently a wide range of downstream document understanding and natural language processing (NLP) tasks, it is vital to devise image processing approaches that can understand and properly segment different layers of text, i.e., printed and handwritten text. + +Models. Prior literature has applied various approaches for the identification and separation of handwritten and printed text. Early approaches [7, 12] formulated the problem as a binary classification task using KNN and SVM and focusing on connected components (CCs) (i.e., groups of pixels). More recently, Li et al. [14] employed conditional random fields (CRFs), with formulating both unary and pairwise potentials for adjacent connected components by leveraging convolutional neural networks (CNNs) architecture for the separation of CCs. The limitation of CC-based approaches is that they determine the class membership for the entire component rather than at the pixel level. Consequently, pixel-level segmentation methods were introduced, leveraging Markov random fields (MRFs) and MLPs [19, 22] for pixel-level classification of PT and HT segmentation. Following the success of encoder-decoder architectures in object segmentation [21], more recent works [10, 6, 20, 27] have predominantly adopted a U-Net based architecture [21], which comprises an encoder-decoder network, for HT and PT segmentation. + +In addition to our contribution of releasing a manually annotated dataset of legal documents with overlapping text, SignaTR6K, our methodology distinguishes itself from prior works in several significant ways. Firstly, we approach the segmentation problem with a four-class formulation, allowing overlapping pixels to be assigned to a new distinct class (OV), which signifies the presence of both HT and PT layers, leading to enhanced segmentation performance. Secondly, we introduce a novel architecture, the Mixed Feature Model (MFM), that combines a Fine Feature Path (FFP) with a Semantic Segmentation Path (SSP) and improves the performance by capitalizing on both high-level and low-level features. Notably, existing U-Net style architectures [10, 6, 20, 27] are limited to leveraging only the SSP path. Further, we present a new loss function, termed Fusion Loss, that converges faster and is more stable compared to prior losses for HT and PT segmentation task. In addition, we introduce a post-processing heuristic based on Conditional Random Fields (CRFH) to carry out relabeling, resulting in further enhancement of text segmentation performance. Additional details related to background and + +![](images/763a59c5b511a454c94fb78a6464b731c73244ec108eee0aa354cf6d8be248af.jpg) +Figure 2: This figure depicts a real crop from a legal document. (a) The original image containing printed text and overlaying handwritten signature; (b) Printed layer pixels annotated only; (c) Handwritten layer pixels annotated only. Any pixels not present in the printed or handwritten layers are marked as background. + +relevant works can be found in the supplementary. + +# 3. SignalTR6K Dataset + +Successfully training and testing a segmentation model requires access to high-quality labeled datasets. Due to the limited availability of public data that contains both handwritten and printed text, previous research has predominantly sought to synthetically generate such data. To achieve this, researchers have combined datasets that either exclusively contain printed or handwritten text, or those with non-overlapping text, such as IAM [17], RIMES [1], PRImA [4], CVL [13], Scanned Questionnaire [2], and WGM-SYN [27]. The scanned documents can originate from diverse domains, each possessing its unique fonts, characters, and quality. Whether the original documents having poor quality in the case of archival documents [27], or the scanning process resulted in lower resolution or loss of contrast, these factors also compound the complexities of the text segmentation task. In addition, in some cases errors are introduced during the automated text labeling process [20]. Given these challenges, the absence of a precisely and manually annotated and verified dataset from real sources that contains various patterns of overlapping printed and handwritten text, significantly impedes the supervised machine learning (ML) process. To address this gap, we introduce a new dataset, SignaTR6K (pronounced as Signature 6K), that is derived from 200 original legal documents from Thomson Reuters Legal Content Services [3]. The dataset features overlapping printed and handwritten text and hand-drawn signatures, as depicted in Figure 1. The documents originate from different organizations, including law firms and courts, each having its distinct fonts and document formats. Furthermore, the annotations are made by different individuals, ensuring a diverse range of printed and handwritten styles. Importantly, each document has been manually labeled and verified. + +Figure 2 displays an original crop from SignaTR6K, which includes both a signature and the information of the signing party. Due to the presence of personally identifiable information (PIIs), we cropped the images to ensure no visible PIIs. Figure 2a illustrates the original image with both printed text and overlaying handwritten signature. Figure 2b presents only the printed layer pixels, while Figure 2c contains only the handwritten layer pixels that are manually annotated. Any pixels absent from either the printed or + +![](images/b43e6f719f36404e019790e70da3ef40c969d2f6d278e55cc65666700ec3f92c.jpg) +Figure 3: Examples from the SignaTR6K dataset, with the top row showing the crops and the bottom row their ground truth annotations. Red: class $PT$ , printed, Green: class $HT$ , handwritten, and Blue: class $BG$ , background. + +handwritten channels are designated as background pixels. It is also evident that in overlapping scenarios, some pixels belong to both handwritten and printed layers. + +Dataset Generation. As hand-annotating a vast collection of real documents is time-consuming and expensive, in order to expand on the size of the dataset and make it adequate for training a deep learning model, we turn to data synthesis and augmentation techniques [10]. The synthetic approaches include general augmentation of crops, along with shifting, magnifying, and rotating operations. We also overlay handwritten and printed pixels from different crops to generate new real-like examples of overlapping text. For this purpose, from the 200 distinct original document samples, we set aside 16 crops, ensuring mutually exclusive samples in the test set. These crops are augmented only for the test set, and we create the training and validation sets from the remaining 184 samples. Following this approach and after excluding generated samples with visible PIDs, we have curated a dataset for training, validation, and testing sets of sizes 5169, 530, and 558 samples, respectively. Each sample is a pair of grayscale crop and its manually annotated ground truth. Each image is 256 by 256 pixels in size, with three channels (RGB) and typically contains several HT and PT overlaps. Four examples from the SignaTR6K dataset are presented in Figure 3, with their grayscale crops and corresponding ground truth (GT) labels. HT is indicated in green, PT in red, and BG in blue. Overlapping HT and PT pixels combine green and red channel values, resulting in a yellow appearance. We envision this dataset can be utilized for model training from scratch or further fine-tuning of a pre-trained classification or segmentation model for specific tasks. The SignaTR6K dataset is freely available for public download. + +# 4. Approach and Methodology + +In this section, we detail our approach, the rationale behind the chosen architecture, the Semantic Segmentation Path (SSP) and Fine Feature Path (FFP), the various loss functions employed, and our novel Fusion loss. + +![](images/a67b8e29a25d60c94be82fd4c531f40468009cabf43cc5322f836e535fbea110.jpg) +Figure 4: Our architecture proposal showing FFP and SSP model paths. The output of FFP and SSP are concatenated prior to the final round of convolutions. A Softmax activation then selects one of the four classes for each pixel. The model's final output consists of pixel-level annotations of the input image. BN: BatchNorm. + +# 4.1. Model Architecture + +The prevalent architecture for object segmentation applies a U-Net design [21, 16]. This architecture leverages a Fully Convolutional Network (FCN), that is without fully connected layers that are typically present at the end of Convolutional Neural Networks (CNNs). The U-Net architecture consists of two main parts, an encoder and a decoder. In the encoder segment, the original image (in our context a document crop) is fed into the model. It then undergoes a series of convolutions and max-pooling layers, extracting features from the image and reducing its dimensions. Conversely, the decoder processes the down-sampled image from the encoder using convolutions and up-sampling layers that eventually restores the image back to its original input size with the same number of channels. The final output of the decoder is a pixel-level labeling of the original image. Semantic Segmentation Path (SSP) in Figure 4 shows this part of the fully convolutional network. In addition to the encoder-decoder architecture in U-Net, there are also skip-connections that bring the feature maps directly from an encoder stage to the corresponding decoder stage, i.e., the decoder stage that has the same image size of the encoder stage. These skip-connections improve segmentation performance by re-accessing early-stage features that may be lost in the encoder's final output due to down-sampling. + +# 4.2. Four-Class Formulation + +As mentioned earlier, prior approaches have either considered binary classification of HT or a three-class formulation of the problem. Binary classification detects only one type of text, while the three-class formulation results in overlapping pixels being assigned solely to either the HT or PT class, impairing performance. Therefore, we propose a four-class formulation of the segmentation task, + +and the fourth class, overlap, $OV$ , models pixels that belong to both the HT and PT classes, specifically enabling the detection of overlapping areas. To expand on Formula (1) for this four-class formulation, for a given document $D$ , assuming there exist four classes as handwritten text $HT$ , printed text $PT$ , background $BG$ , and overlap $OV$ , and pixel $p_i$ : $\forall p_i \in D : p_i(c) == True$ if $p_i \in c$ and $c \in \{HT, PT, BG, OV\}$ . + +Overlapping pixels are highlighted in yellow in the ground truth, as seen in Figure 3. Our four-class single-label classification employs a Softmax activation function in the final layer, which ensures that only one output for each pixel is activated (Figure 4). Since the output image comprises three channels, when the $OV$ class is predicted, during a post-processing step we turn on pixels for both $HT$ and $PT$ channels, resulting in the yellow color in the output image. In addition to the four-class formulation, we also explored three-class and multi-label formulations. In this scenario, instead of a Softmax activation, we applied Sigmoid activations for each class (i.e., three separate sig-moids). However, formulating the problem as a multi-label classification introduces added complexity and degrees of freedom. This can lead to undesirable scenarios, such as the simultaneous activation of pixels for $HT$ and $BG$ . Consequently, the multi-label approach did not yield results comparable to those of the four-class formulation. + +# 4.3. Semantic Segmentation Path (SSP) + +The semantic segmentation path (SSP) of our model leverages a U-Net based architecture, with down-sampling in the encoder stages (i.e., backbone) and up-sampling in the decoder stages. The U-net architecture works well in capturing high-level image features. In this architecture, the encoder and decoder maintain a symmetrical architec + +ture with a similar number of stages. For example, in the SSP, if the encoder goes through four down-sampling stages, i.e., the input image size changes from $256^{*}256$ to $16^{*}16$ ( $256 \rightarrow 128 \rightarrow 64 \rightarrow 32 \rightarrow 16$ ), correspondingly, the decoder undergoes four up-sampling stages, from $16^{*}16$ to $256^{*}256$ . Thus, the final output retains the original input image's pixel dimensions. For the SSP, we explored a variety of network sizes. As a baseline and for comparison, we implemented FCN-light [6, 20, 27]. We then improved on the SSP architecture by using VGG16 [23], InceptionV3 [26], and ResNet34 [8] as backbones of the SSP and observed improvement in the performance. In particular, the ResNet34 and InceptionV3 backbones outperform the prior work (FCN) due to their larger number of learnable parameters. Additionally, the inclusion of residual connections and varied convolution sizes allows better carry over the low-level features of text to the later stages of the segmentation network. This observation further inspired us to introduce the FFP network, which similarly incorporates residual connections and without down-sampling. + +# 4.4. Fine Feature Path (FFP) + +The semantic segmentation path excels when segmenting distinct objects. In this path, down-sampling layers are rapidly applied to the input image to capture high-level features and patterns. However, this rapid processing can lead to the loss of fine, or low-level, features. For our application, low-level features are crucial due to the intertwined nature of printed and handwritten text. To address this, we introduce a parallel path to the SSP, termed the fine feature path (FFP), which avoids down-sampling and instead incorporates a convolution block with residual connections. Note that, while the FFP aids in capturing fine features without down-sampling, on its own, i.e., without the SSP that includes down-sampling, it is insufficient for text segmentation, as the absence of high-level features means the model will not be able to detect high-level patterns irrespective of their pixel locations in the image. In Figure 4, the fine feature block of the FFP is repeated $N_{x}$ times; in our implementation, $N_{x} = 4$ . In addition, each stage of the FFP itself implements residual connections as it provides the flexibility to either bypass the block or use its output, leading to improved results, as we will discuss later in the paper. Similar architectures employing residual blocks have shown improved performance in fine object segmentation tasks, e.g., road segmentation [29]. Table 1 details the FFP architecture for four stages $(N_{x} = 4)$ and the connections between layers and convolution sizes. + +# 4.5. Mixed Feature Model (MFM) + +The Mixed Feature Model (MFM) serves as the umbrella model containing two parallel paths: SSP and FFP, as depicted in Figure 4. The objective of FM is to capture the image's low-level features alongside its high-level features. + +
GroupLayer typeFilterInput(s)Output(s)Output size
Stage 1G 1Conv3 × 3/64img_inputs1.g1.c.o256 × 256 × 64
BatchNorm-s1.g1.c.os1.g1.b.o256 × 256 × 64
ReLU-s1.g1.b.os1.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s1.g1.r.os1.g2.c.o256 × 256 × 64
BatchNorm-s1.g2.c.os1.g2.b.o256 × 256 × 64
ReLU-s1.g2.b.os1.g2.r.o256 × 256 × 64
Concat-img_in, s1.g2.r.os1.o256 × 256 × 67
Stage 2G 1Conv3 × 3/64s1.os2.g1.c.o256 × 256 × 64
BatchNorm-s2.g1.c.os2.g1.b.o256 × 256 × 64
ReLU-s2.g1.b.os2.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s2.g1.r.os2.g2.c.o256 × 256 × 64
BatchNorm-s2.g2.c.os2.g2.b.o256 × 256 × 64
ReLU-s2.g2.b.os2.g2.r.o256 × 256 × 64
Concat-s1.o, s2.g2.r.os2.o256 × 256 × 131
Stage 3G 1Conv3 × 3/64s2.os3.g1.c.o256 × 256 × 64
BatchNorm-s3.g1.c.os3.g1.b.o256 × 256 × 64
ReLU-s3.g1.b.os3.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s3.g1.r.os3.g2.c.o256 × 256 × 64
BatchNorm-s3.g2.c.os3.g2.b.o256 × 256 × 64
ReLU-s3.g2.b.os3.g2.r.o256 × 256 × 64
Concat-s2.o, s3.g2.r.os3.o256 × 256 × 195
Stage 4G 1Conv3 × 3/64s3.os4.g1.c.o256 × 256 × 64
BatchNorm-s4.g1.c.os4.g1.b.o256 × 256 × 64
ReLU-s4.g1.b.os4.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s4.g1.r.os4.g2.c.o256 × 256 × 64
BatchNorm-s4.g2.c.os4.g2.b.o256 × 256 × 64
ReLU-s4.g2.b.os4.g2.r.o256 × 256 × 64
Concat-s3.o, s4.g2.r.os4.o256 × 256 × 259
Output-Conv1 × 1/4s4_oFFP_o256 × 256 × 4
+ +Table 1: The architecture of the Fine Feature Path (FFP). In our design, we use four stages (Stages 1-4) of fine feature blocks. We include BatchNorm and ReLu activation after each convolution layer, and unlike SSP, down-sampling layers (e.g., max pooling) are absent from FFP. The output of each stage is concatenated with its input to create residual connections. + +The outputs of SSP and FFP are concatenated before the final layer convolution, producing the output of the MFM model. Additional details regarding the architecture, inputs, outputs, and convolution layer sizes of the MFM can be found in the supplementary. + +# 4.6. CRF Post-Processing and CRF Heuristic + +Prior research [6, 20, 27] has leveraged dense Conditional Random Fields (CRFs) as a post-processing step to re-label pixels based on their neighboring pixels. In our architectural exploration, we found that, while CRF post-processing is intended to improve segmentation performance, in practice it often hurts the segmentation performance by aggressively re-labeling pixels to incorrect classes. One issue arises due to the imbalanced nature of pixels across classes, with background pixels being predominant. Consequently, many pixels are mistakenly annotated from HT or PT to BG, or from HT to PT, which is undesirable. Based on this observation, we designed a heuristic for CRF post-processing to contain this unfavourable behaviour by only allowing the BG pixels to be relabelled to HT or PT pixels and not vice versa. As we will discuss in the experimentation section, this heuristic further improves segmentation performance. + +# 4.7. Loss Functions + +Due to the nature of the scanned documents and the amount of text on each page, the number of background pixels (i.e., white blank pixels) surpasses that of handwritten and printed pixels. This leads to a class imbalance problem [30], posing the risk of predicting the majority of pixels as background while still minimizing the loss value. As a result, we investigated various loss functions and weight as + +signments to different classes in the loss function to evaluate its impact on segmentation performance. Additionally, during this exploration, we observed that different loss functions achieve different IoU (Intersection over Union) scores. Consequently, we introduce a new loss function, Fusion Loss, to incorporate the benefits of various losses. In the following, we discuss the different loss functions we explored. + +Cross-Entropy Loss. The multi-class cross-entropy loss is used for classification tasks involving more than two classes. It assumes that for each data point, only one class can be the ground truth, i.e., multi-class single-label classification. The standard form of cross-entropy loss for a single data point in multi-class segmentation with $M$ classes is defined as: $\mathcal{L}_{CE}(gt, pr) = -\sum_{m=1}^{M} gt \cdot \log(pr)$ , where $pr$ and $gt$ represent the prediction and ground truth, respectively. The final loss is the average of the losses for all data points in a training set. Similarly, the weighted version (WCE) is computed as $\mathcal{L}_{WCE}(gt, pr) = -\sum_{m=1}^{M} w_m \cdot gt \cdot \log(pr)$ , with $w_m$ being the weight assigned to each class in the loss calculation. + +Focal Loss. The focal loss [15] was introduced to address the class-imbalance problem during training of object detection tasks. The focal loss puts the focus of the learning algorithm on the incorrectly classified examples by applying a modulating term $(1 - pr)^{\gamma}$ to the cross-entropy loss. This scaling factor dynamically weighs down the contribution of easy and correctly classified samples, allowing the training loop to concentrate on the harder and incorrectly classified samples. The focal loss is calculated as: $\mathcal{L}_{Focal}(gt, pr) = -\sum_{m=1}^{M} gt \cdot (1 - pr)^{\gamma} \log(pr)$ , where $\gamma$ is the modulating/focusing factor, and we set $\gamma == 2$ as per the original paper. Similarly, the weighted version of the focal loss can be expressed as: $\mathcal{L}_{WF}(gt, pr) = -\sum_{m=1}^{M} gt \cdot \alpha_m \cdot (1 - pr)^{\gamma} \log(pr)$ , where $\alpha_m$ is analogous to $w_m$ , assigning weights to each class such that: $\forall \alpha_m \mid 0 < \alpha_m < 1 \& \& \sum_{m=1}^{M} \alpha_m = 1$ . + +
DiceLoss.Dicelossisgivenby
L_Dice(precision, recall)= 1-2/M∑m=1M
precisionm·recallm/precisionm+recallm
Intuitively, it aims to maximize the F-Score while minimizing the loss, i.e., L_Dice = 1 - F_Score. Accordingly, the weighted version of the dice loss is defined as L_WD(precision, recall) = 1 - 2/M ∑m=1M w_m·precision_m·recall_m /precision_m+recall_m.
+ +Fusion Loss. Given our observation that different loss functions adversely affect different classes, we introduce a new loss targeted for maximizing the performance of all the classes. For example, the dice loss performs better on the background class, as in its formulation, it aims to maximize the F-Score. Because the majority of pixels are attributed + +to the background class, having higher values of correctly classified background pixels achieves a higher F-Score and lower loss values. However, this does not necessarily yield higher IoU scores for PT and HT classes. In contrast, the weighted versions of the cross-entropy and focal losses focus on the handwritten and printed classes. As such, with the Fusion loss, we aim to combine the behaviors of various losses, and we define the Fusion loss as the sum of the three weighted losses: $\mathcal{L}_{\text {Fusion }} = \mathcal{L}_{WF} + \mathcal{L}_{WCE} + \mathcal{L}_{WD}$ . + +# 5. Experiments and Results + +Evaluation Metric. Intersection over Union (IoU) is a commonly used metric to measure the performance of a segmentation task [28], and in particular text segmentation [25, 10, 6, 27]. In our context, we use IoU to measure the pixel-level performance of the model output versus the ground truth. For a class $c$ , let $TP_{c}$ represent the correctly predicted pixels for that class, $FP_{c}$ denote the pixels incorrectly predicted as belonging to class $c$ , and $FN_{c}$ represent the pixels that are incorrectly not predicted for class $c$ , then Intersection over Union, IoU, for class $c$ is calculated as: $IoU_{c} = \frac{TP_{c}}{TP_{c} + FP_{c} + FN_{c}}$ . For each experiment, we calculate the IoU for all three classes, HT, PT, and BG. In the four-class formulation, pixels attributed to the overlap (OV) class are converted in a post-processing step to both HT and PT classes, contributing to the IoU calculations for both. We also calculate the mean IoU, which is the average of IoUs for these three classes: $IoU_{Mean} = \frac{IoU_{HT} + IoU_{PT} + IoU_{BG}}{3}$ . + +Experiments. We perform 50 epochs of training with a batch size of 8. For problem formulation, we use both three-class and four-class implementations. The experiments are conducted on two datasets: SignaTR6K and WGM-SYN [27]. The IoU values for FCN-light and WGM-MOD are as reported in [27] for the three-class formulation which has been performed on the same dataset, WGM-SYN. However, for the SignaTR6K dataset, we retrain an FCN-based model for three-class formulation to ensure a fair comparison on the new dataset. We run experiments for a variety of architectures and in three post-processing configurations. We structure experiments incrementally, i.e., adding one improvement at a time. This approach aids in understanding and isolating the impact of each added improvement, i.e., ablation study. Due to space constraints, the table detailing the complete list of experiment parameters and configurations is relegated to the supplementary. + +Experimented Architectures. We start with FCN-light which has approximately 295K parameters. This architecture aligns with prior research and serves as our baseline [6, 27]. Next, we improve the performance of FCNlight by adopting a four-class implementation. We then experiment with SSP architecture with three different back + +
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-light [6]WGM-MOD [27]~295K~295KWeighted CE24.0026.0072.0041.0011.0023.0072.0036.00----
Weighted CE42.0036.0074.0050.00*41.0032.0074.0049.00----
4-Class(Ours)FCN-light~295KCE46.4941.9873.7754.0838.9529.3573.9247.4146.5741.5973.8754.01
Focal46.0242.0173.8053.9532.9624.5074.0043.8246.1141.2373.9853.77
Dice48.0147.2871.2255.5548.5948.4871.2256.1048.0747.7971.2955.72
Weighted CE46.0742.1873.8254.0240.5430.0773.9848.2046.3141.6873.9553.98
Weighted Focal43.7141.7773.9553.1432.6724.4074.0743.7144.0641.1274.1153.10
Weighted Dice47.9747.2271.0255.4048.6948.3471.0256.0248.0447.7071.1055.61
Fusion48.1247.2571.9355.7746.3444.2571.8654.1548.1947.8972.3856.15
SSP - VGG16~24MCE35.1432.4073.9647.1730.9719.1473.8841.3335.2132.0874.1347.14
Focal34.5632.2274.0246.9329.1119.4273.9140.8134.6731.7974.2246.89
Dice40.6439.6471.1350.4743.3040.1671.2151.5540.8140.2571.2550.77
Weighted CE35.7333.4474.4347.8732.0020.7074.1242.2835.9332.9574.6547.84
Weighted Focal34.9034.3574.3647.8729.1820.7274.0641.3235.1133.8074.5647.82
Weighted Dice41.8040.6470.6751.0344.1441.3970.7052.0742.5241.1970.9751.56
Fusion39.9939.5672.3150.6235.3927.9372.3545.2240.1039.9472.6350.89
MFM (FFP + SSP) - InceptionV3~30MCE52.5144.0973.9156.8446.0335.9474.2152.0652.5542.1774.3556.36
Focal51.5643.7473.8956.4035.8927.1574.4145.8251.6341.7474.4355.93
Dice20.9119.2881.5240.5725.5124.6179.9443.3520.9219.2981.5440.58
Weighted CE51.5843.7273.9556.4244.7834.0474.3251.0551.6441.8574.4055.96
Weighted Focal51.0143.6273.8956.1835.7626.6074.2245.5351.0841.6074.3555.68
Weighted Dice22.3021.2580.7341.4229.1129.4877.9745.5222.3221.2680.7441.44
Fusion53.2251.2571.8458.7750.0548.8372.0356.9753.3249.8372.7258.62
MFM (FFP + SSP) - ReNet34~24MCE52.4044.0273.9156.7847.5236.4074.1252.6852.4342.1274.3456.30
Focal51.9543.9873.8956.6136.7327.1774.3146.0751.9041.9974.3356.08
Dice29.6829.4377.0645.3931.6225.3975.1544.0529.8429.6377.0745.52
Weighted CE52.6344.1273.9456.9047.6037.2574.1252.9952.5642.2274.3656.38
Weighted Focal51.6843.9573.9056.5135.7327.2274.3045.7551.7241.9874.3456.01
Weighted Dice51.4450.6071.9658.0052.9751.4271.9058.7651.5449.3772.7557.89
Fusion52.9951.7271.6958.8051.1550.7971.8357.9253.0151.2972.4958.93
+ +Table 2: IoU performance (%) on the WGM-SYN dataset [27]. The maximum value of a column (representing a class) is underlined. The overall maximum of a class across different post-processing methods is both bolded and underlined. For example, the best mean IoU for the WGM-SYN dataset is achieved with Fusion loss, using CRFH, and the MFM-ResNet34 architecture, stands at 58.93. The best performing configuration from prior work, denoted with $(^{*})$ , stands at 50.00. + +bones: VGG16, InceptionV3, and ResNet34. Our experimentation also includes three variations of MFM, which incorporates both FFP and SSP, and with three SSP variations: VGG16, InceptionV3, and ResNet34. For our implementations, we use the Segmentation Models Library available in TensorFlow [9]. + +Post-Processing Configurations. We experiment with three distinct configurations: without post-processing, with CRF post-processing, and with CRFH post-processing. In the scenario without post-processing, the IoU calculation is performed directly on the output of the model (FCN-light, SSP, and MFM). For the CRF post-processing approach, all classes are permissible for relabeling. However, with the CRF post-processing with heuristic (CRFH), as explained in Section 4.6, only the background class pixels (BG) are allowed to be relabeled as HT or PT classes + +Weight Initialization. For both FCN-light and SSP models, model weights start from their initial random values at the beginning of the training. We train MFM configurations at last to reuse the weights from the SSP configurations. Thus, for the MFM trainings, we apply transfer learning by initializing with weights from the corresponding, previously trained SSP. For example, when training an MFM (FFP+SSP) with an SSP using the ResNet34 backbone, we initialize its SSP path with the available weights from the + +SSP - ResNet34. Lastly, although we have aimed for consistent IoU results by fixing the random seeds, the inherent non-determinism associated with GPU execution remains. As such, we have run our experiments multiple times to validate the IoU value improvements across configurations. + +Results. Tables 2 and 3 show the IoU values for the WGM-SYN and SignaTR6K datasets. The overall trend across the results indicates that transitioning from the three-class formulation to the four-class formulation improves the IoU scores. Furthermore, employing larger model backbones generally, but not always, improves the segmentation performance. Among all the model architectures evaluated, the ResNet34 and InceptionV3 backbones achieve the highest performance, which we attribute to their residual connections and different size convolutions as they can better capture fine features from the image. We provide a summary of the trends in the results here, while a more detailed version of the result tables can be found in the supplementary material. + +- Prior work [6, 27], which implements the three-class formulation of HT and PT segmentation, generally exhibits the lowest performance. Transitioning from the three-class to the four-class formulation with the same backbone, i.e., FCN-light, with the same number of model parameters and WCE loss, improves the average IoU values by + +
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-based [6, 27]~295KCE62.5688.0998.4083.02*52.6889.6899.2680.4662.7290.5899.0584.11
Focal62.3488.0098.4582.9344.6084.8699.2876.2562.5790.8399.2184.21
Dice60.4587.2997.8581.8661.2488.8398.1782.7460.5287.8897.9782.12
Weighted CE60.5884.8997.7881.0953.4590.7399.5281.2360.8486.3398.2481.80
Weighted Focal61.2585.7498.0281.6744.2785.7799.5076.5161.5587.4298.5882.52
Weighted Dice60.2187.0097.7481.6560.7288.1797.9882.2960.2787.4697.8381.86
Fusion61.5288.3998.3882.7658.9492.5599.3783.6261.6790.4598.9183.68
4-Class(Ours)FCN-based~295KCE64.5589.2198.3984.0554.6089.6899.2381.1764.8791.8199.0685.25
Focal64.1088.8698.3283.7646.6486.0199.2677.3064.3491.6599.1185.03
Dice64.3788.6898.3783.8165.1790.1398.5984.6364.5789.1498.4684.06
Weighted CE63.7887.5298.1683.1554.7790.5099.4281.5664.1289.4398.7184.09
Weighted Focal63.6887.7198.1983.2048.1887.1099.4578.2464.0589.7298.8084.19
Weighted Dice64.2188.5798.3083.6965.0890.1098.5384.5764.4489.0698.3983.96
Fusion64.6888.4898.3383.8359.8391.7799.2883.6365.0090.7298.9184.87
MFM (FFP + SSP) - InceptionV3~30MCE73.1092.6698.7788.1863.4892.8999.5585.3173.0594.8999.3689.10
Focal72.8092.5098.7588.0157.4790.5899.5482.5372.7794.7499.3588.95
Dice72.5692.2098.7087.8272.3893.6099.0088.3372.5294.7399.3788.87
Weighted CE72.6691.9098.5487.7062.6593.0499.6285.1072.6393.3198.9488.29
Weighted Focal72.5592.1398.6287.7759.3591.6499.5983.5372.5093.7799.0788.45
Weighted Dice72.5992.3198.7087.8772.6094.0099.0788.5672.5494.6099.3288.82
Fusion72.5592.2598.7187.8365.5793.4999.5386.1972.4994.6299.3588.82
MFM (FFP + SSP) - ResNet34~24MCE72.8192.5698.7888.0563.0492.9499.5585.1772.7594.9399.3989.02
Focal73.0492.4698.7588.0853.0289.0499.5580.5473.0094.7599.3589.03
Dice72.9692.3898.7288.0273.1693.3598.9388.4872.9394.7999.3689.03
Weighted CE72.9691.9698.6987.8764.4992.4799.5685.5172.9094.1999.2788.79
Weighted Focal73.1892.1098.6387.9754.9689.7199.6081.4273.1693.7299.0688.65
Weighted Dice72.7892.3298.7187.9472.9593.6699.0088.5372.7594.6799.3488.92
Fusion73.2692.4598.7388.1568.3894.8599.5687.6073.2194.5999.3189.04
+ +Table 3: IoU performance (%) on the SignaTR6K dataset. The maximum value of a column (i.e., class) is underlined, and the overall maximum for a class across different post-processing is both bolded and underlined. For example, the best mean IoU for the SignaTR6K dataset is for CE loss, with CRFH, and the MFM-InceptionV3 architecture at 89.10. The best performing configuration of prior work, i.e., excluding results for Fusion loss and CRFH, is marked with $(^{*})$ and is 83.02. + +![](images/7d5b25256701b918e3a13e5bff7cc6c6e14cea621887a5eb2158574779bf6660.jpg) +(a) Validation set losses + +![](images/d5e5f29b952699f79b1351b5f098f7e636d956263c2b7fb4de5a3d22f82250ef.jpg) +(b) Validation set IoUs +Figure 5: Convergence speed of different loss functions and normalized loss values on the validation set of the SignaTR6K dataset. + +$8.0\%$ $(50\rightarrow 54.02)$ for the WGM-SYN dataset, and $2.6\%$ $(81.09\rightarrow 83.15)$ for the SignaTR6K dataset. + +- Applying CRF post-processing generally degrades the results. In our research, we observe that CRF post processing is dataset dependent and decreases the performance of segmentation in some cases. CRF post-processing relabels aggressively and incorrectly converts HT and PT pixels to background ones, or OV pixels to PT ones. For example, Figure 6i illustrates that how OV pixels are wrongly relabeled to PT class by CRF post-processing. Contrary to CRF post-processing, our CRF heuristic generally improves the IoU values. For example, for the SignaTR6K dataset, for FCN-light with four-class implementation and CE loss, CRFH improves HT IoUs by $2.9\%$ (89.21→91.81). +- Moving from the FCN-light architecture to larger models, i.e., SSP and MFM models, generally improves the results, except for the VGG16 backbone. For VGG16, we ratio + +nalize that having a deep network without residual connections or varied-size convolutions, as seen in ResNet34 and InceptionV3 architectures, hurts the model's ability in capturing fine features. In addition, we observe that adding FFP helps to improve the performance of the VGG model as we compare SSP and MFM performance values for both datasets. This also confirms the FFP's ability to capture low-level features that are missed by the SSP path. + +- The MFM model with ResNet34 and InceptionV3 backbones generally achieves the best results. MFM-ResNet34 with Fusion loss and CRFH achieves the best result for the WGM-SYN dataset (58.93%), while MFM-InceptionV3 with CE loss and CRFH achieves the best mean IoUs for the SignaTR6K dataset (89.10%). Overall, the best-performing model from our designs improves on the mean IoU performance of the best-performing prior work [27] by $17.9\%$ (50→58.93) and $7.3\%$ (83.02→89.10) for the WGM-SYN and SignaTR6K datasets, respectively. + +- Although using skip connections in SSP helps with the segmentation performance and thus less pronounced improvement is observed in MFM, our FFP design is agnostic of the SSP implementation. This distinction becomes clear when VGG16 is employed as SSP; the addition of FFP boosts the IoU scores by $7\%$ and $8\%$ on the WGM-SYN and SignaTR6K datasets, respectively. The FCN-based model with 295K parameters outperforms VGG16 with 30M parameters. Thus, larger models do not necessarily outper + +![](images/22e67c9e4dcac981ffb14bd13df9d2bb981ba01a2c514da5efa5924abb0af7d9.jpg) +(a) + +![](images/e98c98ecfe0cc7afe43a7dcbdc8f1e803e57d5fd0d6a35054e8e003b7fbe75c8.jpg) +(b) + +![](images/368de0aa453c66ef7def5efe871d67dd1c35ff27e3339b0f3a3654b08e747324.jpg) +(c) + +![](images/bd7c6d4b3d22374fea9c78b12157ad7a4232967d98360b9461661a1364d1a48f.jpg) +(d) + +![](images/d8959a15031bf004ee0d854582b70be19b04b0337e2a2081732885e3ff5e2b6b.jpg) +(e) + +![](images/a899598576675d1ce9eec88411d45fabdc5a3d0dc88fcd130f5ea8c4a00f3906.jpg) +(f) + +![](images/0a3d5dc7f2a01aaeb3517f308ec868cd2a44dc8e70278e1f6df3876d29236b05.jpg) +(g) + +![](images/422d27ef0a3376567a76daff09504a39c5bdc3857a007acf62d19b859530c362.jpg) +(h) + +![](images/536b7d2ea595d5cb995a75067329c70d5e7f2721a8a9d71b449b5305992fdcaf.jpg) +(i) + +![](images/c60d21c244800bfe3404fe9d10a1d9b11c87d9d09a8c30838fda8434c6e8f927.jpg) +(i) + +![](images/b2b8126387904ca6b9a202ddc346af19661894185bbd35c244e7a07755c1e3c6.jpg) +(k) +Figure 6: Example results on the test set of the SignaTR6K dataset for our approach compared to the ground truth and prior works. (a) Input image; (b) Ground truth; (c) & (d) 3-class FCN-based [6, 27] with CE loss without (c) and with (d) CRF post-processing; (e), (f), & (g) Our FCN-based 4-class formulation with CE loss without CRF (e), with CRF (f), and with CRFH (g); (h), (i), & (j) SSP-ResNet34 with CE loss without CRF (h), with CRF (i), and with CRFH (j); (k), (l), & (m) MFM-ResNet34 with CE loss without CRF (k), with CRF (l), and with CRFH (m); (n), (o), & (p) MFM-ResNet34 with Fusion loss without CRF (n), with CRF (o), and with CRFH (p). + +![](images/7e63806dcb481218440ac55a851328a28e7cd52a669502f920471609ee2a807e.jpg) +(1) + +![](images/87b4827c41ecc70d414b997ade04b739c3aac96f0749aac514fee637ff87778a.jpg) +(m) + +![](images/ec8fadea36b3fb4a27dc718c40651ae8b05a1e28c8e140f9e9cb7f3946b8691c.jpg) +(n) + +![](images/19ef7ffc552a2c13939c24e86f356133a20ed1a9c1bc5245db31dda634cf4c07.jpg) +(0) + +![](images/81d52906c08a12adb3090fee257dfaa97489ff6cc1c1d949a285eeef9841fef4.jpg) +(p) + +form smaller ones consistently, and integrating FFP can bring significant improvements in IoU scores. Our findings show that, irrespective of the SSP size and architecture, incorporating FFP improves the performance, and it is essential for segmentation tasks involving fine objects, e.g., text. + +Loss Functions. We also perform an analysis to confirm the stability and convergence of the Fusion loss compared to other losses. Figures 5a and 5b compare different loss functions on the validation set of the SignaTR6K dataset. In Figure 5a, we observe that the Fusion loss is generally stable while converging to its minimum loss. Additionally, in Figure 5b, we observe Fusion loss reaches its maximum IoU around Epoch 15, generally faster than most other losses and stays close to the maximum IoU values on the validation set. It is important to highlight that the maximum IoU on the validation set for Focal and CE losses does not necessarily imply the best performance on the test set. Overall, Fusion loss shows a stable behavior and achieves a better performance on the test set, whereas for comparison, Focal loss shows instability at some epochs during the training. + +Visual Comparison. Figure 6 shows sample model outputs for the SignaTR6K dataset, with two rectangle regions of interest highlighted to showcase the differences between various models. Figure 6b shows the ground truth, and a visual trend indicates that the performance on the regions of interest improves from Figure 6c to Figure 6p. It is also visually noticeable that CRF post-processing aggressively relabels pixels (6i), whereas CRFH generally improves the results. More visual comparisons on the WGM-SYN and SignaTR6K datasets can be found in the supplementary. + +Limitations. Compared to prior work, the complexity and training cost of our approach present some limitations. Our MFM model, being larger than FCN-light, requires greater GPU memory sizes for training, and takes longer to train. + +However, we believe some of the limitations on the training time can be offset through transferring weights from SSP models to MFM ones. Additionally, while our approach outperforms prior work on the WGM-SYN dataset, the mean IoU performance is lower compared to the SignaTR6K dataset. This indicates potential limitations in our method's efficacy for lower-quality original documents, those undergoing low-quality scanning processes (like historical documents), or those with errors from automated labeling. We also attribute the improved performance on the SignaTR6K dataset to its higher quality and our thorough manual annotation, that have resulted in better model training and improved IoU results. + +# 6. Conclusion + +Segmentation of handwritten text (HT) and printed text (PT) is vital for digitization and understanding of scanned documents. The complexity increases with the overlapping of different text types. In this research, we introduced SignaTR6K, a new open-source dataset with high-quality, manual pixel-level annotations, sourced from original legal documents. Additionally, we proposed a novel four-class formulation and a new architecture for the segmentation task. Our design leverages both the Fine Feature Path (FFP) and the Semantic Segmentation Path (SSP) to create the Mixed Feature Model (MFM), and incorporates both high-level and low-level features and improves on the text segmentation performance. We also introduced a CRF-based post-processing heuristic (CRFH) that further improves the model output, and included a new loss function, Fusion loss, that combines the advantages of different loss functions and achieves faster convergence and higher stability compared to most of the evaluated losses. In conclusion, our designs outperform the prior work in mean IoU scores by $17.9\%$ and $7.3\%$ on WGM-SYN and SignaTR6K datasets, respectively. + +# References + +[1] RIMES. http://www.a2ialab.com/doku.php?id=rimesdatabase:start. Inaccessible:2023-01-23. +[2] Scanned Questionnaire. https://github.com/jottue/HTSNet. Accessed: 2023-01-23. +[3] Thomson Reuters Legal Content Services. https://www.thomsonreuters.com/en/ products-services/legal.html. Accessed: 2023-01-21. +[4] Apostolos Antonacopoulos, David Bridson, Christos Papadopoulos, and Stefan Pletschacher. A realistic dataset for performance evaluation of document layout analysis. In 2009 10th International Conference on Document Analysis and Recognition, pages 296-300. IEEE, 2009. +[5] Dibyajyoti Dhar, Avishek Garain, Pawan Kumar Singh, and Ram Sarkar. Hp_docpres: A method for classifying printed and handwritten texts in doctor's prescription. Multimedia Tools and Applications, 80(7):9779-9812, 2021. +[6] Nicolas Dutly, Fouad Slimane, and Rolf Ingold. PHTI-WS: a printed and handwritten text identification web service based on FCN and CRF post-processing. In 2019 International Conference on Document Analysis and Recognition Workshops (IC-DARW), volume 2, pages 20-25. IEEE, 2019. +[7] Jürgen Franke and Matthias Oberlander. Writing style detection by statistical combination of classifiers in form reader applications. In Proceedings of 2nd International Conference on Document Analysis and Recognition (ICDAR'93), pages 581-584. IEEE, 1993. +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[9] Pavel Iakubovskii. Segmentation models. https://github.com/qubvel/segmentation_models, 2019. +[10] Junho Jo, Hyung Il Koo, Jae Woong Soh, and Nam Ik Cho. Handwritten text segmentation via end-to-end learning of convolutional neural networks. *Multimedia Tools and Applications*, 79(43):32137-32150, 2020. +[11] Celso AML Junior, Matheus Henrique Marques da Silva, Byron Leite Dantas Bezerra, Bruno José Torres Fernandes, and Donato Impedovo. FCN+ RL: a fully convolutional network followed by refinement layers to offline handwritten signature segmentation. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020. + +[12] R Kandan, Nirup Kumar Reddy, KR Arvind, and AG Ramakrishnan. A robust two level classification algorithm for text localization in documents. In Advances in Visual Computing: Third International Symposium, ISVC 2007, Lake Tahoe, NV, USA, November 26-28, 2007, Proceedings, Part II 3, pages 96-105. Springer, 2007. +[13] Florian Kleber, Stefan Fiel, Markus Diem, and Robert Sablatnig. CVL-database: An off-line database for writer retrieval, writer identification and word spotting. In 2013 12th international conference on document analysis and recognition, pages 560-564. IEEE, 2013. +[14] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Printed/handwritten texts and graphics separation in complex documents using conditional random fields. In 2018 13th IAPR International Workshop on Document Analysis Systems (DAS), pages 145-150. IEEE, 2018. +[15] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. +[16] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. +[17] U-V Marti and Horst Bunke. The IAM-database: an English sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1):39-46, 2002. +[18] Milda Norkute, Nadja Herger, Leszek Michalak, Andrew Mulder, and Sally Gao. Towards explainable ai: Assessing the usefulness and impact of added explainability features in legal document summarization. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-7, 2021. +[19] Xujun Peng, Srirangaraj Setlur, Venu Govindaraju, and Ramachandrula Sitaram. Handwritten text separation from annotated machine printed documents using markov random fields. International Journal on Document Analysis and Recognition (IJDAR), 16(1):1-16, 2013. +[20] Anastasia Prikhodina. Handwritten and printed text separation for historical documents. 2021. +[21] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical + +image computing and computer-assisted intervention, pages 234-241. Springer, 2015. +[22] Mathias Seuret, Marcus Liwicki, and Rolf Ingold. Pixel level handwritten and printed content discrimination in scanned documents. In 2014 14th International Conference on Frontiers in Handwriting Recognition, pages 423-428. IEEE, 2014. +[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[24] Stavroula Skylaki, Ali Oskooei, Omar Bari, Nadja Herger, and Zac Kriegman. Named entity recognition in the legal domain using a pointer generator network. arXiv preprint arXiv:2012.09936, 2020. +[25] Nishant Subramani, Alexandre Matton, Malcolm Greaves, and Adrian Lam. A survey of deep learning approaches forOCR and document understanding. arXiv preprint arXiv:2011.13534, 2020. +[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the in- + +ception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. +[27] Mahsa Vafaie, Oleksandra Bruns, Natasja Pilz, Jörg Waitelonis, and Harald Sack. Handwritten and printed text identification in historical archival documents. In Archiving Conference, volume 19, pages 15–20. Society for Imaging Science and Technology, 2022. +[28] Zhaobin Wang, E Wang, and Ying Zhu. Image segmentation evaluation: A survey of methods. Artificial Intelligence Review, 53:5637-5674, 2020. +[29] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual U-Net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018. +[30] Zhi-Hua Zhou and Xu-Ying Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on knowledge and data engineering, 18(1):63-77, 2005. \ No newline at end of file diff --git a/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/images.zip b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3e44d5512f0bc7dd8c46080cfbaea2069da1edd --- /dev/null +++ b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a70116eceb2b8457154d26ad0609df88758b73f6ec287b4298dbb2fe1c83c5d4 +size 836782 diff --git a/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/layout.json b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..498de5998fc6d7a31d2fc48e31e227ec3a2cbfda --- /dev/null +++ b/2023/Handwritten and Printed Text Segmentation_ A Signature Case Study/layout.json @@ -0,0 +1,7732 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "text", + "content": "Handwritten and Printed Text Segmentation: A Signature Case Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 138, + 266, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 138, + 266, + 178 + ], + "spans": [ + { + "bbox": [ + 138, + 138, + 266, + 178 + ], + "type": "text", + "content": "Sina Gholamian \nThomson Reuters AI Labs \nToronto, Canada" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 182, + 292, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 182, + 292, + 193 + ], + "spans": [ + { + "bbox": [ + 111, + 182, + 292, + 193 + ], + "type": "text", + "content": "sina.gholamian@thomsonreuters.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 338, + 138, + 467, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 138, + 467, + 178 + ], + "spans": [ + { + "bbox": [ + 338, + 138, + 467, + 178 + ], + "type": "text", + "content": "Ali Vahdat \nThomson Reuters AI Labs \nToronto, Canada" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 323, + 182, + 482, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 182, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 323, + 182, + 482, + 192 + ], + "type": "text", + "content": "ali.vahdat@thomsonreuters.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 221, + 191, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 221, + 191, + 233 + ], + "spans": [ + { + "bbox": [ + 143, + 221, + 191, + 233 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 248, + 289, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 289, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 289, + 523 + ], + "type": "text", + "content": "While analyzing scanned documents, handwritten text can overlap with printed text. This overlap causes difficulties during the optical character recognition (OCR) and digitization process of documents, and subsequently, hurts downstream NLP tasks. Prior research either focuses solely on the binary classification of handwritten text or performs a three-class segmentation of the document, i.e., recognition of handwritten, printed, and background pixels. This approach results in the assignment of overlapping handwritten and printed pixels to only one of the classes, and thus, they are not accounted for in the other class. Thus, in this research, we develop novel approaches to address the challenges of handwritten and printed text segmentation. Our objective is to recover text from different classes in their entirety, especially enhancing the segmentation performance on overlapping sections. To support this task, we introduce a new dataset, SignaTR6K, collected from real legal documents, as well as a new model architecture for the handwritten and printed text segmentation task. Our best configuration outperforms prior work on two different datasets by " + }, + { + "bbox": [ + 46, + 248, + 289, + 523 + ], + "type": "inline_equation", + "content": "17.9\\%" + }, + { + "bbox": [ + 46, + 248, + 289, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 248, + 289, + 523 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 46, + 248, + 289, + 523 + ], + "type": "text", + "content": " on IoU scores. The SignaTR6K dataset is accessible for download via the following link: https://forms.office.com/r/2a5RDg7cAY." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 542, + 127, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 127, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 127, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "For various purposes, the digitization of hard-copy documents and associated challenges are an active area of research in both academia [25, 11, 27, 10, 6, 20] and industry [18]. This digital transformation involves scanning paper documents through an OCR process, making their text accessible for downstream natural language processing (NLP) tasks, such as named entity recognition (NER). The documents of interest can originate from a variety of domains, including historical documents [27], legal and court-issued documents [18], business contracts [25], and medical records and prescriptions [5]. Although various studies, as cited above, have been conducted, there is yet a considerable gap between current approaches and" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 220, + 541, + 369 + ], + "blocks": [ + { + "bbox": [ + 312, + 220, + 541, + 369 + ], + "lines": [ + { + "bbox": [ + 312, + 220, + 541, + 369 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 541, + 369 + ], + "type": "image", + "image_path": "6eacb8b0ae731321b16ca49a0f5af1630a146e8148ec1e190575361d38647ff6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 445 + ], + "type": "text", + "content": "Figure 1: Court documents are first printed and then signed or annotated by various parties. This results in handwritten text overlapping the underlying printed information, leading to performance degradation in downstream tasks, such as named entity recognition (NER). This is a fabricated example, however very similar to the original documents, to protect personally identifiable information (PII)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 676 + ], + "type": "text", + "content": "human-level performance in mixed-text scenarios (i.e., when handwritten and printed text overlap). For instance, attorneys frequently sign legal documents, resulting in their signatures overlapping with their information. This overlap hampers the performance of OCR tools in character recognition, subsequently making it challenging for downstream tasks to accurately identify information linked to the attorneys and their associated law firms. Figure 1 provides an illustration of handwritten text (HT) overlapping with printed text (PT) in court documents. Extracting parties names is a crucial step in the named-entity recognition (NER) task for legal and court documents [24]. When attorneys and other involved parties sign these documents, which are later scanned using OCR tools, their signatures often obscure the details of names and law firms. Consequently, the semantic segmentation of handwritten elements, such as lawyers' signatures and accompanying handwritten notes, and printed text detailing the lawyer and the law firm's information, becomes vital." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "In this research, we aim to address the challenges of HT and PT segmentation, as there is still a large gap between human-level performance and the existing approaches" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "582" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "for this task. Our focus in this effort is to improve the segmentation performance in overlapping regions for scanned legal documents, and to aid in this endeavor, we also introduce a new dataset. In summary, our research makes the following contributions:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 41, + 140, + 288, + 416 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 41, + 140, + 288, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 140, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 41, + 140, + 288, + 262 + ], + "type": "text", + "content": "- We introduce a new dataset, SignaTR6K (pronounce as Signature 6K)1 derived from 200 pixel-level manually annotated crops of images from genuine legal documents. The dataset comprises signatures, handwritten text, and printed text, which frequently overlap. With data augmentation, we have created a dataset of sizes 5169, 530, 558, for training, validation, and testing, respectively, that we release to the public, to facilitate dataset availability for future research and to aid in the training and evaluation of deep-learning segmentation models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 41, + 268, + 288, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 268, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 41, + 268, + 288, + 340 + ], + "type": "text", + "content": "- We propose a novel architecture that integrates both semantic segmentation features and fine features, enhancing the performance of text segmentation over previous methods. Moreover, we introduce a new loss function termed Fusion loss, that, comparatively, is stable and converges to optimal loss values and Intersection over Union (IoU) scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 343, + 288, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 343, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 41, + 343, + 288, + 416 + ], + "type": "text", + "content": "- Lastly, we conduct an extensive quantitative and visual evaluation of different variations of our approach against prior work on two distinct datasets, and illustrate our approach's superior performance in the text segmentation task, especially in challenging scenarios where printed and handwritten text overlap." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 425, + 244, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 244, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 244, + 438 + ], + "type": "text", + "content": "2. Background and Literature Review" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": "The text segmentation problem is defined as follows: given a scanned document possibly containing handwritten, printed, and background (i.e., blank) pixels, the task is to assign each pixel to its appropriate class. Scanned documents can originate from various sources, such as hard-copy paper documents or microfilms [20, 27]. Formally, for a given document " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": ", assuming there exist three classes as handwritten text " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "HT" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": ", printed text " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "PT" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": ", and background " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "BG" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": ", and pixel " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\forall p_i \\in D : p_i(c) == \\text{True if } p_i \\in c \\& \\& c \\in \\{HT, PT, BG\\}" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": " (1). A situation may arise where a pixel belongs to two classes, i.e., " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "HT" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "inline_equation", + "content": "PT" + }, + { + "bbox": [ + 46, + 442, + 287, + 609 + ], + "type": "text", + "content": " classes, when handwritten text overlaps with printed text, which single-label three-class formulation cannot handle such cases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 610, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 288, + 695 + ], + "type": "text", + "content": "Several previous studies have studied the segmentation of handwritten and printed text [6, 10, 20, 27], however, they have inherent limitations. Some focus solely on binary classification, determining if a pixel is handwritten or not [10], whereas others adopt a 3-class formulation of the problem, classifying pixels as handwritten, printed, or background [6, 20, 27]. This exclusive assignment of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "pixels to three different classes paralyzes any machine learning segmentation model to properly detect pixels in the overlapping areas to belong to both handwritten and printed classes. In our usecase depicted in Figure 1, the OCR process achieves insufficient performance due to overlapping of handwritten text and signatures that overlay the printed text. As such, to improve the documents digitization quality, and subsequently a wide range of downstream document understanding and natural language processing (NLP) tasks, it is vital to devise image processing approaches that can understand and properly segment different layers of text, i.e., printed and handwritten text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 222, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 547, + 462 + ], + "type": "text", + "content": "Models. Prior literature has applied various approaches for the identification and separation of handwritten and printed text. Early approaches [7, 12] formulated the problem as a binary classification task using KNN and SVM and focusing on connected components (CCs) (i.e., groups of pixels). More recently, Li et al. [14] employed conditional random fields (CRFs), with formulating both unary and pairwise potentials for adjacent connected components by leveraging convolutional neural networks (CNNs) architecture for the separation of CCs. The limitation of CC-based approaches is that they determine the class membership for the entire component rather than at the pixel level. Consequently, pixel-level segmentation methods were introduced, leveraging Markov random fields (MRFs) and MLPs [19, 22] for pixel-level classification of PT and HT segmentation. Following the success of encoder-decoder architectures in object segmentation [21], more recent works [10, 6, 20, 27] have predominantly adopted a U-Net based architecture [21], which comprises an encoder-decoder network, for HT and PT segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 463, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 547, + 715 + ], + "type": "text", + "content": "In addition to our contribution of releasing a manually annotated dataset of legal documents with overlapping text, SignaTR6K, our methodology distinguishes itself from prior works in several significant ways. Firstly, we approach the segmentation problem with a four-class formulation, allowing overlapping pixels to be assigned to a new distinct class (OV), which signifies the presence of both HT and PT layers, leading to enhanced segmentation performance. Secondly, we introduce a novel architecture, the Mixed Feature Model (MFM), that combines a Fine Feature Path (FFP) with a Semantic Segmentation Path (SSP) and improves the performance by capitalizing on both high-level and low-level features. Notably, existing U-Net style architectures [10, 6, 20, 27] are limited to leveraging only the SSP path. Further, we present a new loss function, termed Fusion Loss, that converges faster and is more stable compared to prior losses for HT and PT segmentation task. In addition, we introduce a post-processing heuristic based on Conditional Random Fields (CRFH) to carry out relabeling, resulting in further enhancement of text segmentation performance. Additional details related to background and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 286, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 286, + 714 + ], + "type": "text", + "content": "Available at https://forms.office.com/r/2a5RDg7cAY." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "583" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 67, + 283, + 112 + ], + "blocks": [ + { + "bbox": [ + 52, + 67, + 283, + 112 + ], + "lines": [ + { + "bbox": [ + 52, + 67, + 283, + 112 + ], + "spans": [ + { + "bbox": [ + 52, + 67, + 283, + 112 + ], + "type": "image", + "image_path": "763a59c5b511a454c94fb78a6464b731c73244ec108eee0aa354cf6d8be248af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 113, + 287, + 168 + ], + "lines": [ + { + "bbox": [ + 47, + 113, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 113, + 287, + 168 + ], + "type": "text", + "content": "Figure 2: This figure depicts a real crop from a legal document. (a) The original image containing printed text and overlaying handwritten signature; (b) Printed layer pixels annotated only; (c) Handwritten layer pixels annotated only. Any pixels not present in the printed or handwritten layers are marked as background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 251, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 251, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 251, + 186 + ], + "type": "text", + "content": "relevant works can be found in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 194, + 164, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 164, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 164, + 206 + ], + "type": "text", + "content": "3. SignalTR6K Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 211, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 289, + 605 + ], + "type": "text", + "content": "Successfully training and testing a segmentation model requires access to high-quality labeled datasets. Due to the limited availability of public data that contains both handwritten and printed text, previous research has predominantly sought to synthetically generate such data. To achieve this, researchers have combined datasets that either exclusively contain printed or handwritten text, or those with non-overlapping text, such as IAM [17], RIMES [1], PRImA [4], CVL [13], Scanned Questionnaire [2], and WGM-SYN [27]. The scanned documents can originate from diverse domains, each possessing its unique fonts, characters, and quality. Whether the original documents having poor quality in the case of archival documents [27], or the scanning process resulted in lower resolution or loss of contrast, these factors also compound the complexities of the text segmentation task. In addition, in some cases errors are introduced during the automated text labeling process [20]. Given these challenges, the absence of a precisely and manually annotated and verified dataset from real sources that contains various patterns of overlapping printed and handwritten text, significantly impedes the supervised machine learning (ML) process. To address this gap, we introduce a new dataset, SignaTR6K (pronounced as Signature 6K), that is derived from 200 original legal documents from Thomson Reuters Legal Content Services [3]. The dataset features overlapping printed and handwritten text and hand-drawn signatures, as depicted in Figure 1. The documents originate from different organizations, including law firms and courts, each having its distinct fonts and document formats. Furthermore, the annotations are made by different individuals, ensuring a diverse range of printed and handwritten styles. Importantly, each document has been manually labeled and verified." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Figure 2 displays an original crop from SignaTR6K, which includes both a signature and the information of the signing party. Due to the presence of personally identifiable information (PIIs), we cropped the images to ensure no visible PIIs. Figure 2a illustrates the original image with both printed text and overlaying handwritten signature. Figure 2b presents only the printed layer pixels, while Figure 2c contains only the handwritten layer pixels that are manually annotated. Any pixels absent from either the printed or" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 190 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 190 + ], + "type": "image", + "image_path": "b43e6f719f36404e019790e70da3ef40c969d2f6d278e55cc65666700ec3f92c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "type": "text", + "content": "Figure 3: Examples from the SignaTR6K dataset, with the top row showing the crops and the bottom row their ground truth annotations. Red: class " + }, + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "type": "inline_equation", + "content": "PT" + }, + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "type": "text", + "content": ", printed, Green: class " + }, + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "type": "inline_equation", + "content": "HT" + }, + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "type": "text", + "content": ", handwritten, and Blue: class " + }, + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "type": "inline_equation", + "content": "BG" + }, + { + "bbox": [ + 305, + 190, + 547, + 234 + ], + "type": "text", + "content": ", background." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 240, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 277 + ], + "type": "text", + "content": "handwritten channels are designated as background pixels. It is also evident that in overlapping scenarios, some pixels belong to both handwritten and printed layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 281, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 641 + ], + "type": "text", + "content": "Dataset Generation. As hand-annotating a vast collection of real documents is time-consuming and expensive, in order to expand on the size of the dataset and make it adequate for training a deep learning model, we turn to data synthesis and augmentation techniques [10]. The synthetic approaches include general augmentation of crops, along with shifting, magnifying, and rotating operations. We also overlay handwritten and printed pixels from different crops to generate new real-like examples of overlapping text. For this purpose, from the 200 distinct original document samples, we set aside 16 crops, ensuring mutually exclusive samples in the test set. These crops are augmented only for the test set, and we create the training and validation sets from the remaining 184 samples. Following this approach and after excluding generated samples with visible PIDs, we have curated a dataset for training, validation, and testing sets of sizes 5169, 530, and 558 samples, respectively. Each sample is a pair of grayscale crop and its manually annotated ground truth. Each image is 256 by 256 pixels in size, with three channels (RGB) and typically contains several HT and PT overlaps. Four examples from the SignaTR6K dataset are presented in Figure 3, with their grayscale crops and corresponding ground truth (GT) labels. HT is indicated in green, PT in red, and BG in blue. Overlapping HT and PT pixels combine green and red channel values, resulting in a yellow appearance. We envision this dataset can be utilized for model training from scratch or further fine-tuning of a pre-trained classification or segmentation model for specific tasks. The SignaTR6K dataset is freely available for public download." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 649, + 464, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 464, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 464, + 662 + ], + "type": "text", + "content": "4. Approach and Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 666, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 715 + ], + "type": "text", + "content": "In this section, we detail our approach, the rationale behind the chosen architecture, the Semantic Segmentation Path (SSP) and Fine Feature Path (FFP), the various loss functions employed, and our novel Fusion loss." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "584" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 68, + 521, + 265 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 521, + 265 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 521, + 265 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 521, + 265 + ], + "type": "image", + "image_path": "a67b8e29a25d60c94be82fd4c531f40468009cabf43cc5322f836e535fbea110.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 45, + 266, + 547, + 301 + ], + "type": "text", + "content": "Figure 4: Our architecture proposal showing FFP and SSP model paths. The output of FFP and SSP are concatenated prior to the final round of convolutions. A Softmax activation then selects one of the four classes for each pixel. The model's final output consists of pixel-level annotations of the input image. BN: BatchNorm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 306, + 162, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 162, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 162, + 317 + ], + "type": "text", + "content": "4.1. Model Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 322, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 322, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 45, + 322, + 290, + 610 + ], + "type": "text", + "content": "The prevalent architecture for object segmentation applies a U-Net design [21, 16]. This architecture leverages a Fully Convolutional Network (FCN), that is without fully connected layers that are typically present at the end of Convolutional Neural Networks (CNNs). The U-Net architecture consists of two main parts, an encoder and a decoder. In the encoder segment, the original image (in our context a document crop) is fed into the model. It then undergoes a series of convolutions and max-pooling layers, extracting features from the image and reducing its dimensions. Conversely, the decoder processes the down-sampled image from the encoder using convolutions and up-sampling layers that eventually restores the image back to its original input size with the same number of channels. The final output of the decoder is a pixel-level labeling of the original image. Semantic Segmentation Path (SSP) in Figure 4 shows this part of the fully convolutional network. In addition to the encoder-decoder architecture in U-Net, there are also skip-connections that bring the feature maps directly from an encoder stage to the corresponding decoder stage, i.e., the decoder stage that has the same image size of the encoder stage. These skip-connections improve segmentation performance by re-accessing early-stage features that may be lost in the encoder's final output due to down-sampling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 613, + 183, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 183, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 183, + 625 + ], + "type": "text", + "content": "4.2. Four-Class Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "As mentioned earlier, prior approaches have either considered binary classification of HT or a three-class formulation of the problem. Binary classification detects only one type of text, while the three-class formulation results in overlapping pixels being assigned solely to either the HT or PT class, impairing performance. Therefore, we propose a four-class formulation of the segmentation task," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": "and the fourth class, overlap, " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "OV" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": ", models pixels that belong to both the HT and PT classes, specifically enabling the detection of overlapping areas. To expand on Formula (1) for this four-class formulation, for a given document " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": ", assuming there exist four classes as handwritten text " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "HT" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": ", printed text " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "PT" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": ", background " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "BG" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": ", and overlap " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "OV" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": ", and pixel " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\forall p_i \\in D : p_i(c) == True" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "p_i \\in c" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "inline_equation", + "content": "c \\in \\{HT, PT, BG, OV\\}" + }, + { + "bbox": [ + 304, + 306, + 545, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "text", + "content": "Overlapping pixels are highlighted in yellow in the ground truth, as seen in Figure 3. Our four-class single-label classification employs a Softmax activation function in the final layer, which ensures that only one output for each pixel is activated (Figure 4). Since the output image comprises three channels, when the " + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "inline_equation", + "content": "OV" + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "text", + "content": " class is predicted, during a post-processing step we turn on pixels for both " + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "inline_equation", + "content": "HT" + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "inline_equation", + "content": "PT" + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "text", + "content": " channels, resulting in the yellow color in the output image. In addition to the four-class formulation, we also explored three-class and multi-label formulations. In this scenario, instead of a Softmax activation, we applied Sigmoid activations for each class (i.e., three separate sig-moids). However, formulating the problem as a multi-label classification introduces added complexity and degrees of freedom. This can lead to undesirable scenarios, such as the simultaneous activation of pixels for " + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "inline_equation", + "content": "HT" + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "inline_equation", + "content": "BG" + }, + { + "bbox": [ + 304, + 403, + 546, + 619 + ], + "type": "text", + "content": ". Consequently, the multi-label approach did not yield results comparable to those of the four-class formulation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 625, + 493, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 625, + 493, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 625, + 493, + 639 + ], + "type": "text", + "content": "4.3. Semantic Segmentation Path (SSP)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "The semantic segmentation path (SSP) of our model leverages a U-Net based architecture, with down-sampling in the encoder stages (i.e., backbone) and up-sampling in the decoder stages. The U-net architecture works well in capturing high-level image features. In this architecture, the encoder and decoder maintain a symmetrical architec" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "585" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "ture with a similar number of stages. For example, in the SSP, if the encoder goes through four down-sampling stages, i.e., the input image size changes from " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "256^{*}256" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "16^{*}16" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "256 \\rightarrow 128 \\rightarrow 64 \\rightarrow 32 \\rightarrow 16" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "), correspondingly, the decoder undergoes four up-sampling stages, from " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "16^{*}16" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "256^{*}256" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": ". Thus, the final output retains the original input image's pixel dimensions. For the SSP, we explored a variety of network sizes. As a baseline and for comparison, we implemented FCN-light [6, 20, 27]. We then improved on the SSP architecture by using VGG16 [23], InceptionV3 [26], and ResNet34 [8] as backbones of the SSP and observed improvement in the performance. In particular, the ResNet34 and InceptionV3 backbones outperform the prior work (FCN) due to their larger number of learnable parameters. Additionally, the inclusion of residual connections and varied convolution sizes allows better carry over the low-level features of text to the later stages of the segmentation network. This observation further inspired us to introduce the FFP network, which similarly incorporates residual connections and without down-sampling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 316, + 185, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 185, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 185, + 328 + ], + "type": "text", + "content": "4.4. Fine Feature Path (FFP)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "text", + "content": "The semantic segmentation path excels when segmenting distinct objects. In this path, down-sampling layers are rapidly applied to the input image to capture high-level features and patterns. However, this rapid processing can lead to the loss of fine, or low-level, features. For our application, low-level features are crucial due to the intertwined nature of printed and handwritten text. To address this, we introduce a parallel path to the SSP, termed the fine feature path (FFP), which avoids down-sampling and instead incorporates a convolution block with residual connections. Note that, while the FFP aids in capturing fine features without down-sampling, on its own, i.e., without the SSP that includes down-sampling, it is insufficient for text segmentation, as the absence of high-level features means the model will not be able to detect high-level patterns irrespective of their pixel locations in the image. In Figure 4, the fine feature block of the FFP is repeated " + }, + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "inline_equation", + "content": "N_{x}" + }, + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "text", + "content": " times; in our implementation, " + }, + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "inline_equation", + "content": "N_{x} = 4" + }, + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "text", + "content": ". In addition, each stage of the FFP itself implements residual connections as it provides the flexibility to either bypass the block or use its output, leading to improved results, as we will discuss later in the paper. Similar architectures employing residual blocks have shown improved performance in fine object segmentation tasks, e.g., road segmentation [29]. Table 1 details the FFP architecture for four stages " + }, + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "inline_equation", + "content": "(N_{x} = 4)" + }, + { + "bbox": [ + 46, + 331, + 289, + 643 + ], + "type": "text", + "content": " and the connections between layers and convolution sizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 650, + 209, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 209, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 209, + 662 + ], + "type": "text", + "content": "4.5. Mixed Feature Model (MFM)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "The Mixed Feature Model (MFM) serves as the umbrella model containing two parallel paths: SSP and FFP, as depicted in Figure 4. The objective of FM is to capture the image's low-level features alongside its high-level features." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 308, + 68, + 558, + 244 + ], + "blocks": [ + { + "bbox": [ + 308, + 68, + 558, + 244 + ], + "lines": [ + { + "bbox": [ + 308, + 68, + 558, + 244 + ], + "spans": [ + { + "bbox": [ + 308, + 68, + 558, + 244 + ], + "type": "table", + "html": "
GroupLayer typeFilterInput(s)Output(s)Output size
Stage 1G 1Conv3 × 3/64img_inputs1.g1.c.o256 × 256 × 64
BatchNorm-s1.g1.c.os1.g1.b.o256 × 256 × 64
ReLU-s1.g1.b.os1.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s1.g1.r.os1.g2.c.o256 × 256 × 64
BatchNorm-s1.g2.c.os1.g2.b.o256 × 256 × 64
ReLU-s1.g2.b.os1.g2.r.o256 × 256 × 64
Concat-img_in, s1.g2.r.os1.o256 × 256 × 67
Stage 2G 1Conv3 × 3/64s1.os2.g1.c.o256 × 256 × 64
BatchNorm-s2.g1.c.os2.g1.b.o256 × 256 × 64
ReLU-s2.g1.b.os2.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s2.g1.r.os2.g2.c.o256 × 256 × 64
BatchNorm-s2.g2.c.os2.g2.b.o256 × 256 × 64
ReLU-s2.g2.b.os2.g2.r.o256 × 256 × 64
Concat-s1.o, s2.g2.r.os2.o256 × 256 × 131
Stage 3G 1Conv3 × 3/64s2.os3.g1.c.o256 × 256 × 64
BatchNorm-s3.g1.c.os3.g1.b.o256 × 256 × 64
ReLU-s3.g1.b.os3.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s3.g1.r.os3.g2.c.o256 × 256 × 64
BatchNorm-s3.g2.c.os3.g2.b.o256 × 256 × 64
ReLU-s3.g2.b.os3.g2.r.o256 × 256 × 64
Concat-s2.o, s3.g2.r.os3.o256 × 256 × 195
Stage 4G 1Conv3 × 3/64s3.os4.g1.c.o256 × 256 × 64
BatchNorm-s4.g1.c.os4.g1.b.o256 × 256 × 64
ReLU-s4.g1.b.os4.g1.r.o256 × 256 × 64
G 2Conv3 × 3/64s4.g1.r.os4.g2.c.o256 × 256 × 64
BatchNorm-s4.g2.c.os4.g2.b.o256 × 256 × 64
ReLU-s4.g2.b.os4.g2.r.o256 × 256 × 64
Concat-s3.o, s4.g2.r.os4.o256 × 256 × 259
Output-Conv1 × 1/4s4_oFFP_o256 × 256 × 4
", + "image_path": "8314f15a1a605120ce6423fbb8c74555d9b085a75831ba0603e6d11f1d278552.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 245, + 547, + 311 + ], + "lines": [ + { + "bbox": [ + 304, + 245, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 547, + 311 + ], + "type": "text", + "content": "Table 1: The architecture of the Fine Feature Path (FFP). In our design, we use four stages (Stages 1-4) of fine feature blocks. We include BatchNorm and ReLu activation after each convolution layer, and unlike SSP, down-sampling layers (e.g., max pooling) are absent from FFP. The output of each stage is concatenated with its input to create residual connections." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 316, + 547, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 377 + ], + "type": "text", + "content": "The outputs of SSP and FFP are concatenated before the final layer convolution, producing the output of the MFM model. Additional details regarding the architecture, inputs, outputs, and convolution layer sizes of the MFM can be found in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 384, + 519, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 519, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 519, + 397 + ], + "type": "text", + "content": "4.6. CRF Post-Processing and CRF Heuristic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 399, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 605 + ], + "type": "text", + "content": "Prior research [6, 20, 27] has leveraged dense Conditional Random Fields (CRFs) as a post-processing step to re-label pixels based on their neighboring pixels. In our architectural exploration, we found that, while CRF post-processing is intended to improve segmentation performance, in practice it often hurts the segmentation performance by aggressively re-labeling pixels to incorrect classes. One issue arises due to the imbalanced nature of pixels across classes, with background pixels being predominant. Consequently, many pixels are mistakenly annotated from HT or PT to BG, or from HT to PT, which is undesirable. Based on this observation, we designed a heuristic for CRF post-processing to contain this unfavourable behaviour by only allowing the BG pixels to be relabelled to HT or PT pixels and not vice versa. As we will discuss in the experimentation section, this heuristic further improves segmentation performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 611, + 400, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 400, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 400, + 623 + ], + "type": "text", + "content": "4.7. Loss Functions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Due to the nature of the scanned documents and the amount of text on each page, the number of background pixels (i.e., white blank pixels) surpasses that of handwritten and printed pixels. This leads to a class imbalance problem [30], posing the risk of predicting the majority of pixels as background while still minimizing the loss value. As a result, we investigated various loss functions and weight as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "586" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "signments to different classes in the loss function to evaluate its impact on segmentation performance. Additionally, during this exploration, we observed that different loss functions achieve different IoU (Intersection over Union) scores. Consequently, we introduce a new loss function, Fusion Loss, to incorporate the benefits of various losses. In the following, we discuss the different loss functions we explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "content": "Cross-Entropy Loss. The multi-class cross-entropy loss is used for classification tasks involving more than two classes. It assumes that for each data point, only one class can be the ground truth, i.e., multi-class single-label classification. The standard form of cross-entropy loss for a single data point in multi-class segmentation with " + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "content": " classes is defined as: " + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot \\log(pr)" + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "inline_equation", + "content": "pr" + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "inline_equation", + "content": "gt" + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "content": " represent the prediction and ground truth, respectively. The final loss is the average of the losses for all data points in a training set. Similarly, the weighted version (WCE) is computed as " + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{WCE}(gt, pr) = -\\sum_{m=1}^{M} w_m \\cdot gt \\cdot \\log(pr)" + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "inline_equation", + "content": "w_m" + }, + { + "bbox": [ + 46, + 170, + 289, + 327 + ], + "type": "text", + "content": " being the weight assigned to each class in the loss calculation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": "Focal Loss. The focal loss [15] was introduced to address the class-imbalance problem during training of object detection tasks. The focal loss puts the focus of the learning algorithm on the incorrectly classified examples by applying a modulating term " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "(1 - pr)^{\\gamma}" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": " to the cross-entropy loss. This scaling factor dynamically weighs down the contribution of easy and correctly classified samples, allowing the training loop to concentrate on the harder and incorrectly classified samples. The focal loss is calculated as: " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Focal}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot (1 - pr)^{\\gamma} \\log(pr)" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": " is the modulating/focusing factor, and we set " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\gamma == 2" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": " as per the original paper. Similarly, the weighted version of the focal loss can be expressed as: " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{WF}(gt, pr) = -\\sum_{m=1}^{M} gt \\cdot \\alpha_m \\cdot (1 - pr)^{\\gamma} \\log(pr)" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\alpha_m" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": " is analogous to " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "w_m" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": ", assigning weights to each class such that: " + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\forall \\alpha_m \\mid 0 < \\alpha_m < 1 \\& \\& \\sum_{m=1}^{M} \\alpha_m = 1" + }, + { + "bbox": [ + 46, + 339, + 289, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 46, + 540, + 289, + 632 + ], + "blocks": [ + { + "bbox": [ + 46, + 540, + 289, + 632 + ], + "lines": [ + { + "bbox": [ + 46, + 540, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 289, + 632 + ], + "type": "table", + "html": "
DiceLoss.Dicelossisgivenby
L_Dice(precision, recall)= 1-2/M∑m=1M
precisionm·recallm/precisionm+recallm
Intuitively, it aims to maximize the F-Score while minimizing the loss, i.e., L_Dice = 1 - F_Score. Accordingly, the weighted version of the dice loss is defined as L_WD(precision, recall) = 1 - 2/M ∑m=1M w_m·precision_m·recall_m /precision_m+recall_m.
", + "image_path": "067cbeaf1b37d4fc11ddfb819cd4c5397424fa2bd401a9eb26ca215b74e04c08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 715 + ], + "type": "text", + "content": "Fusion Loss. Given our observation that different loss functions adversely affect different classes, we introduce a new loss targeted for maximizing the performance of all the classes. For example, the dice loss performs better on the background class, as in its formulation, it aims to maximize the F-Score. Because the majority of pixels are attributed" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "to the background class, having higher values of correctly classified background pixels achieves a higher F-Score and lower loss values. However, this does not necessarily yield higher IoU scores for PT and HT classes. In contrast, the weighted versions of the cross-entropy and focal losses focus on the handwritten and printed classes. As such, with the Fusion loss, we aim to combine the behaviors of various losses, and we define the Fusion loss as the sum of the three weighted losses: " + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {Fusion }} = \\mathcal{L}_{WF} + \\mathcal{L}_{WCE} + \\mathcal{L}_{WD}" + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 192, + 451, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 451, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 451, + 206 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": "Evaluation Metric. Intersection over Union (IoU) is a commonly used metric to measure the performance of a segmentation task [28], and in particular text segmentation [25, 10, 6, 27]. In our context, we use IoU to measure the pixel-level performance of the model output versus the ground truth. For a class " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "TP_{c}" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": " represent the correctly predicted pixels for that class, " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "FP_{c}" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": " denote the pixels incorrectly predicted as belonging to class " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "FN_{c}" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": " represent the pixels that are incorrectly not predicted for class " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": ", then Intersection over Union, IoU, for class " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": " is calculated as: " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "IoU_{c} = \\frac{TP_{c}}{TP_{c} + FP_{c} + FN_{c}}" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": ". For each experiment, we calculate the IoU for all three classes, HT, PT, and BG. In the four-class formulation, pixels attributed to the overlap (OV) class are converted in a post-processing step to both HT and PT classes, contributing to the IoU calculations for both. We also calculate the mean IoU, which is the average of IoUs for these three classes: " + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "inline_equation", + "content": "IoU_{Mean} = \\frac{IoU_{HT} + IoU_{PT} + IoU_{BG}}{3}" + }, + { + "bbox": [ + 304, + 210, + 547, + 419 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 426, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 631 + ], + "type": "text", + "content": "Experiments. We perform 50 epochs of training with a batch size of 8. For problem formulation, we use both three-class and four-class implementations. The experiments are conducted on two datasets: SignaTR6K and WGM-SYN [27]. The IoU values for FCN-light and WGM-MOD are as reported in [27] for the three-class formulation which has been performed on the same dataset, WGM-SYN. However, for the SignaTR6K dataset, we retrain an FCN-based model for three-class formulation to ensure a fair comparison on the new dataset. We run experiments for a variety of architectures and in three post-processing configurations. We structure experiments incrementally, i.e., adding one improvement at a time. This approach aids in understanding and isolating the impact of each added improvement, i.e., ablation study. Due to space constraints, the table detailing the complete list of experiment parameters and configurations is relegated to the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Experimented Architectures. We start with FCN-light which has approximately 295K parameters. This architecture aligns with prior research and serves as our baseline [6, 27]. Next, we improve the performance of FCNlight by adopting a four-class implementation. We then experiment with SSP architecture with three different back" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "text", + "content": "587" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 342 + ], + "type": "table", + "html": "
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-light [6]WGM-MOD [27]~295K~295KWeighted CE24.0026.0072.0041.0011.0023.0072.0036.00----
Weighted CE42.0036.0074.0050.00*41.0032.0074.0049.00----
4-Class(Ours)FCN-light~295KCE46.4941.9873.7754.0838.9529.3573.9247.4146.5741.5973.8754.01
Focal46.0242.0173.8053.9532.9624.5074.0043.8246.1141.2373.9853.77
Dice48.0147.2871.2255.5548.5948.4871.2256.1048.0747.7971.2955.72
Weighted CE46.0742.1873.8254.0240.5430.0773.9848.2046.3141.6873.9553.98
Weighted Focal43.7141.7773.9553.1432.6724.4074.0743.7144.0641.1274.1153.10
Weighted Dice47.9747.2271.0255.4048.6948.3471.0256.0248.0447.7071.1055.61
Fusion48.1247.2571.9355.7746.3444.2571.8654.1548.1947.8972.3856.15
SSP - VGG16~24MCE35.1432.4073.9647.1730.9719.1473.8841.3335.2132.0874.1347.14
Focal34.5632.2274.0246.9329.1119.4273.9140.8134.6731.7974.2246.89
Dice40.6439.6471.1350.4743.3040.1671.2151.5540.8140.2571.2550.77
Weighted CE35.7333.4474.4347.8732.0020.7074.1242.2835.9332.9574.6547.84
Weighted Focal34.9034.3574.3647.8729.1820.7274.0641.3235.1133.8074.5647.82
Weighted Dice41.8040.6470.6751.0344.1441.3970.7052.0742.5241.1970.9751.56
Fusion39.9939.5672.3150.6235.3927.9372.3545.2240.1039.9472.6350.89
MFM (FFP + SSP) - InceptionV3~30MCE52.5144.0973.9156.8446.0335.9474.2152.0652.5542.1774.3556.36
Focal51.5643.7473.8956.4035.8927.1574.4145.8251.6341.7474.4355.93
Dice20.9119.2881.5240.5725.5124.6179.9443.3520.9219.2981.5440.58
Weighted CE51.5843.7273.9556.4244.7834.0474.3251.0551.6441.8574.4055.96
Weighted Focal51.0143.6273.8956.1835.7626.6074.2245.5351.0841.6074.3555.68
Weighted Dice22.3021.2580.7341.4229.1129.4877.9745.5222.3221.2680.7441.44
Fusion53.2251.2571.8458.7750.0548.8372.0356.9753.3249.8372.7258.62
MFM (FFP + SSP) - ReNet34~24MCE52.4044.0273.9156.7847.5236.4074.1252.6852.4342.1274.3456.30
Focal51.9543.9873.8956.6136.7327.1774.3146.0751.9041.9974.3356.08
Dice29.6829.4377.0645.3931.6225.3975.1544.0529.8429.6377.0745.52
Weighted CE52.6344.1273.9456.9047.6037.2574.1252.9952.5642.2274.3656.38
Weighted Focal51.6843.9573.9056.5135.7327.2274.3045.7551.7241.9874.3456.01
Weighted Dice51.4450.6071.9658.0052.9751.4271.9058.7651.5449.3772.7557.89
Fusion52.9951.7271.6958.8051.1550.7971.8357.9253.0151.2972.4958.93
", + "image_path": "39719b860ac19c247836ddd31d58e4727a1d414f03ca55662de6e747ed715616.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 342, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 545, + 387 + ], + "type": "text", + "content": "Table 2: IoU performance (%) on the WGM-SYN dataset [27]. The maximum value of a column (representing a class) is underlined. The overall maximum of a class across different post-processing methods is both bolded and underlined. For example, the best mean IoU for the WGM-SYN dataset is achieved with Fusion loss, using CRFH, and the MFM-ResNet34 architecture, stands at 58.93. The best performing configuration from prior work, denoted with " + }, + { + "bbox": [ + 46, + 342, + 545, + 387 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 46, + 342, + 545, + 387 + ], + "type": "text", + "content": ", stands at 50.00." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 392, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 464 + ], + "type": "text", + "content": "bones: VGG16, InceptionV3, and ResNet34. Our experimentation also includes three variations of MFM, which incorporates both FFP and SSP, and with three SSP variations: VGG16, InceptionV3, and ResNet34. For our implementations, we use the Segmentation Models Library available in TensorFlow [9]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 471, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 590 + ], + "type": "text", + "content": "Post-Processing Configurations. We experiment with three distinct configurations: without post-processing, with CRF post-processing, and with CRFH post-processing. In the scenario without post-processing, the IoU calculation is performed directly on the output of the model (FCN-light, SSP, and MFM). For the CRF post-processing approach, all classes are permissible for relabeling. However, with the CRF post-processing with heuristic (CRFH), as explained in Section 4.6, only the background class pixels (BG) are allowed to be relabeled as HT or PT classes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Weight Initialization. For both FCN-light and SSP models, model weights start from their initial random values at the beginning of the training. We train MFM configurations at last to reuse the weights from the SSP configurations. Thus, for the MFM trainings, we apply transfer learning by initializing with weights from the corresponding, previously trained SSP. For example, when training an MFM (FFP+SSP) with an SSP using the ResNet34 backbone, we initialize its SSP path with the available weights from the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 392, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 453 + ], + "type": "text", + "content": "SSP - ResNet34. Lastly, although we have aimed for consistent IoU results by fixing the random seeds, the inherent non-determinism associated with GPU execution remains. As such, we have run our experiments multiple times to validate the IoU value improvements across configurations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 463, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 463, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 303, + 463, + 545, + 631 + ], + "type": "text", + "content": "Results. Tables 2 and 3 show the IoU values for the WGM-SYN and SignaTR6K datasets. The overall trend across the results indicates that transitioning from the three-class formulation to the four-class formulation improves the IoU scores. Furthermore, employing larger model backbones generally, but not always, improves the segmentation performance. Among all the model architectures evaluated, the ResNet34 and InceptionV3 backbones achieve the highest performance, which we attribute to their residual connections and different size convolutions as they can better capture fine features from the image. We provide a summary of the trends in the results here, while a more detailed version of the result tables can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 300, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 300, + 641, + 547, + 715 + ], + "type": "text", + "content": "- Prior work [6, 27], which implements the three-class formulation of HT and PT segmentation, generally exhibits the lowest performance. Transitioning from the three-class to the four-class formulation with the same backbone, i.e., FCN-light, with the same number of model parameters and WCE loss, improves the average IoU values by" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "588" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 67, + 545, + 320 + ], + "blocks": [ + { + "bbox": [ + 47, + 67, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 47, + 67, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 67, + 545, + 320 + ], + "type": "table", + "html": "
FormulationBackbone# ParametersLoss functionIoU %With CRF (IoU %)With CRFH (IoU %)
PTHTBGMeanPTHTBGMeanPTHTBGMean
3-ClassFCN-based [6, 27]~295KCE62.5688.0998.4083.02*52.6889.6899.2680.4662.7290.5899.0584.11
Focal62.3488.0098.4582.9344.6084.8699.2876.2562.5790.8399.2184.21
Dice60.4587.2997.8581.8661.2488.8398.1782.7460.5287.8897.9782.12
Weighted CE60.5884.8997.7881.0953.4590.7399.5281.2360.8486.3398.2481.80
Weighted Focal61.2585.7498.0281.6744.2785.7799.5076.5161.5587.4298.5882.52
Weighted Dice60.2187.0097.7481.6560.7288.1797.9882.2960.2787.4697.8381.86
Fusion61.5288.3998.3882.7658.9492.5599.3783.6261.6790.4598.9183.68
4-Class(Ours)FCN-based~295KCE64.5589.2198.3984.0554.6089.6899.2381.1764.8791.8199.0685.25
Focal64.1088.8698.3283.7646.6486.0199.2677.3064.3491.6599.1185.03
Dice64.3788.6898.3783.8165.1790.1398.5984.6364.5789.1498.4684.06
Weighted CE63.7887.5298.1683.1554.7790.5099.4281.5664.1289.4398.7184.09
Weighted Focal63.6887.7198.1983.2048.1887.1099.4578.2464.0589.7298.8084.19
Weighted Dice64.2188.5798.3083.6965.0890.1098.5384.5764.4489.0698.3983.96
Fusion64.6888.4898.3383.8359.8391.7799.2883.6365.0090.7298.9184.87
MFM (FFP + SSP) - InceptionV3~30MCE73.1092.6698.7788.1863.4892.8999.5585.3173.0594.8999.3689.10
Focal72.8092.5098.7588.0157.4790.5899.5482.5372.7794.7499.3588.95
Dice72.5692.2098.7087.8272.3893.6099.0088.3372.5294.7399.3788.87
Weighted CE72.6691.9098.5487.7062.6593.0499.6285.1072.6393.3198.9488.29
Weighted Focal72.5592.1398.6287.7759.3591.6499.5983.5372.5093.7799.0788.45
Weighted Dice72.5992.3198.7087.8772.6094.0099.0788.5672.5494.6099.3288.82
Fusion72.5592.2598.7187.8365.5793.4999.5386.1972.4994.6299.3588.82
MFM (FFP + SSP) - ResNet34~24MCE72.8192.5698.7888.0563.0492.9499.5585.1772.7594.9399.3989.02
Focal73.0492.4698.7588.0853.0289.0499.5580.5473.0094.7599.3589.03
Dice72.9692.3898.7288.0273.1693.3598.9388.4872.9394.7999.3689.03
Weighted CE72.9691.9698.6987.8764.4992.4799.5685.5172.9094.1999.2788.79
Weighted Focal73.1892.1098.6387.9754.9689.7199.6081.4273.1693.7299.0688.65
Weighted Dice72.7892.3298.7187.9472.9593.6699.0088.5372.7594.6799.3488.92
Fusion73.2692.4598.7388.1568.3894.8599.5687.6073.2194.5999.3189.04
", + "image_path": "f8cde07cb8a58e717a893e0641df6019ff2171c3feaa3724f81d72d2dd40c04f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 321, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 545, + 365 + ], + "type": "text", + "content": "Table 3: IoU performance (%) on the SignaTR6K dataset. The maximum value of a column (i.e., class) is underlined, and the overall maximum for a class across different post-processing is both bolded and underlined. For example, the best mean IoU for the SignaTR6K dataset is for CE loss, with CRFH, and the MFM-InceptionV3 architecture at 89.10. The best performing configuration of prior work, i.e., excluding results for Fusion loss and CRFH, is marked with " + }, + { + "bbox": [ + 46, + 321, + 545, + 365 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 46, + 321, + 545, + 365 + ], + "type": "text", + "content": " and is 83.02." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 378, + 167, + 452 + ], + "blocks": [ + { + "bbox": [ + 56, + 378, + 167, + 452 + ], + "lines": [ + { + "bbox": [ + 56, + 378, + 167, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 167, + 452 + ], + "type": "image", + "image_path": "7d5b25256701b918e3a13e5bff7cc6c6e14cea621887a5eb2158574779bf6660.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 453, + 148, + 460 + ], + "lines": [ + { + "bbox": [ + 70, + 453, + 148, + 460 + ], + "spans": [ + { + "bbox": [ + 70, + 453, + 148, + 460 + ], + "type": "text", + "content": "(a) Validation set losses" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 378, + 283, + 452 + ], + "blocks": [ + { + "bbox": [ + 171, + 378, + 283, + 452 + ], + "lines": [ + { + "bbox": [ + 171, + 378, + 283, + 452 + ], + "spans": [ + { + "bbox": [ + 171, + 378, + 283, + 452 + ], + "type": "image", + "image_path": "d5e5f29b952699f79b1351b5f098f7e636d956263c2b7fb4de5a3d22f82250ef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 453, + 263, + 461 + ], + "lines": [ + { + "bbox": [ + 188, + 453, + 263, + 461 + ], + "spans": [ + { + "bbox": [ + 188, + 453, + 263, + 461 + ], + "type": "text", + "content": "(b) Validation set IoUs" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 462, + 286, + 483 + ], + "lines": [ + { + "bbox": [ + 47, + 462, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 286, + 483 + ], + "type": "text", + "content": "Figure 5: Convergence speed of different loss functions and normalized loss values on the validation set of the SignaTR6K dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "type": "inline_equation", + "content": "(50\\rightarrow 54.02)" + }, + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "type": "text", + "content": " for the WGM-SYN dataset, and " + }, + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "type": "inline_equation", + "content": "(81.09\\rightarrow 83.15)" + }, + { + "bbox": [ + 48, + 491, + 286, + 516 + ], + "type": "text", + "content": " for the SignaTR6K dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 41, + 519, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 41, + 519, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 519, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 41, + 519, + 287, + 674 + ], + "type": "text", + "content": "- Applying CRF post-processing generally degrades the results. In our research, we observe that CRF post processing is dataset dependent and decreases the performance of segmentation in some cases. CRF post-processing relabels aggressively and incorrectly converts HT and PT pixels to background ones, or OV pixels to PT ones. For example, Figure 6i illustrates that how OV pixels are wrongly relabeled to PT class by CRF post-processing. Contrary to CRF post-processing, our CRF heuristic generally improves the IoU values. For example, for the SignaTR6K dataset, for FCN-light with four-class implementation and CE loss, CRFH improves HT IoUs by " + }, + { + "bbox": [ + 41, + 519, + 287, + 674 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 41, + 519, + 287, + 674 + ], + "type": "text", + "content": " (89.21→91.81)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 41, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 41, + 677, + 287, + 713 + ], + "type": "text", + "content": "- Moving from the FCN-light architecture to larger models, i.e., SSP and MFM models, generally improves the results, except for the VGG16 backbone. For VGG16, we ratio" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 307, + 380, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 545, + 476 + ], + "type": "text", + "content": "nalize that having a deep network without residual connections or varied-size convolutions, as seen in ResNet34 and InceptionV3 architectures, hurts the model's ability in capturing fine features. In addition, we observe that adding FFP helps to improve the performance of the VGG model as we compare SSP and MFM performance values for both datasets. This also confirms the FFP's ability to capture low-level features that are missed by the SSP path." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 300, + 480, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 480, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 300, + 480, + 545, + 601 + ], + "type": "text", + "content": "- The MFM model with ResNet34 and InceptionV3 backbones generally achieves the best results. MFM-ResNet34 with Fusion loss and CRFH achieves the best result for the WGM-SYN dataset (58.93%), while MFM-InceptionV3 with CE loss and CRFH achieves the best mean IoUs for the SignaTR6K dataset (89.10%). Overall, the best-performing model from our designs improves on the mean IoU performance of the best-performing prior work [27] by " + }, + { + "bbox": [ + 300, + 480, + 545, + 601 + ], + "type": "inline_equation", + "content": "17.9\\%" + }, + { + "bbox": [ + 300, + 480, + 545, + 601 + ], + "type": "text", + "content": " (50→58.93) and " + }, + { + "bbox": [ + 300, + 480, + 545, + 601 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 300, + 480, + 545, + 601 + ], + "type": "text", + "content": " (83.02→89.10) for the WGM-SYN and SignaTR6K datasets, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 300, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 300, + 605, + 545, + 714 + ], + "type": "text", + "content": "- Although using skip connections in SSP helps with the segmentation performance and thus less pronounced improvement is observed in MFM, our FFP design is agnostic of the SSP implementation. This distinction becomes clear when VGG16 is employed as SSP; the addition of FFP boosts the IoU scores by " + }, + { + "bbox": [ + 300, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 300, + 605, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 300, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 300, + 605, + 545, + 714 + ], + "type": "text", + "content": " on the WGM-SYN and SignaTR6K datasets, respectively. The FCN-based model with 295K parameters outperforms VGG16 with 30M parameters. Thus, larger models do not necessarily outper" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "589" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 112, + 132 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 112, + 132 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 112, + 132 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 112, + 132 + ], + "type": "image", + "image_path": "22e67c9e4dcac981ffb14bd13df9d2bb981ba01a2c514da5efa5924abb0af7d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 133, + 85, + 139 + ], + "lines": [ + { + "bbox": [ + 76, + 133, + 85, + 139 + ], + "spans": [ + { + "bbox": [ + 76, + 133, + 85, + 139 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 112, + 70, + 173, + 131 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 173, + 131 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 173, + 131 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 173, + 131 + ], + "type": "image", + "image_path": "e98c98ecfe0cc7afe43a7dcbdc8f1e803e57d5fd0d6a35054e8e003b7fbe75c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 132, + 147, + 139 + ], + "lines": [ + { + "bbox": [ + 137, + 132, + 147, + 139 + ], + "spans": [ + { + "bbox": [ + 137, + 132, + 147, + 139 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 235, + 131 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 235, + 131 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 235, + 131 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 235, + 131 + ], + "type": "image", + "image_path": "368de0aa453c66ef7def5efe871d67dd1c35ff27e3339b0f3a3654b08e747324.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 132, + 208, + 139 + ], + "lines": [ + { + "bbox": [ + 199, + 132, + 208, + 139 + ], + "spans": [ + { + "bbox": [ + 199, + 132, + 208, + 139 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 235, + 70, + 297, + 131 + ], + "blocks": [ + { + "bbox": [ + 235, + 70, + 297, + 131 + ], + "lines": [ + { + "bbox": [ + 235, + 70, + 297, + 131 + ], + "spans": [ + { + "bbox": [ + 235, + 70, + 297, + 131 + ], + "type": "image", + "image_path": "bd7c6d4b3d22374fea9c78b12157ad7a4232967d98360b9461661a1364d1a48f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 132, + 271, + 138 + ], + "lines": [ + { + "bbox": [ + 261, + 132, + 271, + 138 + ], + "spans": [ + { + "bbox": [ + 261, + 132, + 271, + 138 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 297, + 70, + 359, + 131 + ], + "blocks": [ + { + "bbox": [ + 297, + 70, + 359, + 131 + ], + "lines": [ + { + "bbox": [ + 297, + 70, + 359, + 131 + ], + "spans": [ + { + "bbox": [ + 297, + 70, + 359, + 131 + ], + "type": "image", + "image_path": "d8959a15031bf004ee0d854582b70be19b04b0337e2a2081732885e3ff5e2b6b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 132, + 332, + 139 + ], + "lines": [ + { + "bbox": [ + 323, + 132, + 332, + 139 + ], + "spans": [ + { + "bbox": [ + 323, + 132, + 332, + 139 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 70, + 421, + 131 + ], + "blocks": [ + { + "bbox": [ + 359, + 70, + 421, + 131 + ], + "lines": [ + { + "bbox": [ + 359, + 70, + 421, + 131 + ], + "spans": [ + { + "bbox": [ + 359, + 70, + 421, + 131 + ], + "type": "image", + "image_path": "a899598576675d1ce9eec88411d45fabdc5a3d0dc88fcd130f5ea8c4a00f3906.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 132, + 394, + 139 + ], + "lines": [ + { + "bbox": [ + 386, + 132, + 394, + 139 + ], + "spans": [ + { + "bbox": [ + 386, + 132, + 394, + 139 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 421, + 70, + 482, + 131 + ], + "blocks": [ + { + "bbox": [ + 421, + 70, + 482, + 131 + ], + "lines": [ + { + "bbox": [ + 421, + 70, + 482, + 131 + ], + "spans": [ + { + "bbox": [ + 421, + 70, + 482, + 131 + ], + "type": "image", + "image_path": "0a3d5dc7f2a01aaeb3517f308ec868cd2a44dc8e70278e1f6df3876d29236b05.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 132, + 457, + 139 + ], + "lines": [ + { + "bbox": [ + 447, + 132, + 457, + 139 + ], + "spans": [ + { + "bbox": [ + 447, + 132, + 457, + 139 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 483, + 70, + 545, + 131 + ], + "blocks": [ + { + "bbox": [ + 483, + 70, + 545, + 131 + ], + "lines": [ + { + "bbox": [ + 483, + 70, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 483, + 70, + 545, + 131 + ], + "type": "image", + "image_path": "422d27ef0a3376567a76daff09504a39c5bdc3857a007acf62d19b859530c362.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 132, + 518, + 139 + ], + "lines": [ + { + "bbox": [ + 509, + 132, + 518, + 139 + ], + "spans": [ + { + "bbox": [ + 509, + 132, + 518, + 139 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 140, + 111, + 200 + ], + "blocks": [ + { + "bbox": [ + 50, + 140, + 111, + 200 + ], + "lines": [ + { + "bbox": [ + 50, + 140, + 111, + 200 + ], + "spans": [ + { + "bbox": [ + 50, + 140, + 111, + 200 + ], + "type": "image", + "image_path": "536b7d2ea595d5cb995a75067329c70d5e7f2721a8a9d71b449b5305992fdcaf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 201, + 85, + 208 + ], + "lines": [ + { + "bbox": [ + 76, + 201, + 85, + 208 + ], + "spans": [ + { + "bbox": [ + 76, + 201, + 85, + 208 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 111, + 140, + 173, + 200 + ], + "blocks": [ + { + "bbox": [ + 111, + 140, + 173, + 200 + ], + "lines": [ + { + "bbox": [ + 111, + 140, + 173, + 200 + ], + "spans": [ + { + "bbox": [ + 111, + 140, + 173, + 200 + ], + "type": "image", + "image_path": "c60d21c244800bfe3404fe9d10a1d9b11c87d9d09a8c30838fda8434c6e8f927.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 201, + 147, + 208 + ], + "lines": [ + { + "bbox": [ + 138, + 201, + 147, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 147, + 208 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 173, + 140, + 235, + 200 + ], + "blocks": [ + { + "bbox": [ + 173, + 140, + 235, + 200 + ], + "lines": [ + { + "bbox": [ + 173, + 140, + 235, + 200 + ], + "spans": [ + { + "bbox": [ + 173, + 140, + 235, + 200 + ], + "type": "image", + "image_path": "b2b8126387904ca6b9a202ddc346af19661894185bbd35c244e7a07755c1e3c6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 201, + 208, + 208 + ], + "lines": [ + { + "bbox": [ + 199, + 201, + 208, + 208 + ], + "spans": [ + { + "bbox": [ + 199, + 201, + 208, + 208 + ], + "type": "text", + "content": "(k)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 209, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 264 + ], + "type": "text", + "content": "Figure 6: Example results on the test set of the SignaTR6K dataset for our approach compared to the ground truth and prior works. (a) Input image; (b) Ground truth; (c) & (d) 3-class FCN-based [6, 27] with CE loss without (c) and with (d) CRF post-processing; (e), (f), & (g) Our FCN-based 4-class formulation with CE loss without CRF (e), with CRF (f), and with CRFH (g); (h), (i), & (j) SSP-ResNet34 with CE loss without CRF (h), with CRF (i), and with CRFH (j); (k), (l), & (m) MFM-ResNet34 with CE loss without CRF (k), with CRF (l), and with CRFH (m); (n), (o), & (p) MFM-ResNet34 with Fusion loss without CRF (n), with CRF (o), and with CRFH (p)." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 235, + 140, + 297, + 200 + ], + "blocks": [ + { + "bbox": [ + 235, + 140, + 297, + 200 + ], + "lines": [ + { + "bbox": [ + 235, + 140, + 297, + 200 + ], + "spans": [ + { + "bbox": [ + 235, + 140, + 297, + 200 + ], + "type": "image", + "image_path": "7e63806dcb481218440ac55a851328a28e7cd52a669502f920471609ee2a807e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 201, + 271, + 208 + ], + "lines": [ + { + "bbox": [ + 262, + 201, + 271, + 208 + ], + "spans": [ + { + "bbox": [ + 262, + 201, + 271, + 208 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 297, + 140, + 359, + 200 + ], + "blocks": [ + { + "bbox": [ + 297, + 140, + 359, + 200 + ], + "lines": [ + { + "bbox": [ + 297, + 140, + 359, + 200 + ], + "spans": [ + { + "bbox": [ + 297, + 140, + 359, + 200 + ], + "type": "image", + "image_path": "87b4827c41ecc70d414b997ade04b739c3aac96f0749aac514fee637ff87778a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 201, + 333, + 208 + ], + "lines": [ + { + "bbox": [ + 322, + 201, + 333, + 208 + ], + "spans": [ + { + "bbox": [ + 322, + 201, + 333, + 208 + ], + "type": "text", + "content": "(m)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 359, + 140, + 421, + 200 + ], + "blocks": [ + { + "bbox": [ + 359, + 140, + 421, + 200 + ], + "lines": [ + { + "bbox": [ + 359, + 140, + 421, + 200 + ], + "spans": [ + { + "bbox": [ + 359, + 140, + 421, + 200 + ], + "type": "image", + "image_path": "ec8fadea36b3fb4a27dc718c40651ae8b05a1e28c8e140f9e9cb7f3946b8691c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 201, + 395, + 208 + ], + "lines": [ + { + "bbox": [ + 385, + 201, + 395, + 208 + ], + "spans": [ + { + "bbox": [ + 385, + 201, + 395, + 208 + ], + "type": "text", + "content": "(n)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 421, + 140, + 482, + 200 + ], + "blocks": [ + { + "bbox": [ + 421, + 140, + 482, + 200 + ], + "lines": [ + { + "bbox": [ + 421, + 140, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 421, + 140, + 482, + 200 + ], + "type": "image", + "image_path": "19ef7ffc552a2c13939c24e86f356133a20ed1a9c1bc5245db31dda634cf4c07.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 201, + 457, + 208 + ], + "lines": [ + { + "bbox": [ + 447, + 201, + 457, + 208 + ], + "spans": [ + { + "bbox": [ + 447, + 201, + 457, + 208 + ], + "type": "text", + "content": "(0)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 483, + 140, + 545, + 200 + ], + "blocks": [ + { + "bbox": [ + 483, + 140, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 483, + 140, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 483, + 140, + 545, + 200 + ], + "type": "image", + "image_path": "81d52906c08a12adb3090fee257dfaa97489ff6cc1c1d949a285eeef9841fef4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 201, + 518, + 208 + ], + "lines": [ + { + "bbox": [ + 509, + 201, + 518, + 208 + ], + "spans": [ + { + "bbox": [ + 509, + 201, + 518, + 208 + ], + "type": "text", + "content": "(p)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 48, + 273, + 287, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 273, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 48, + 273, + 287, + 333 + ], + "type": "text", + "content": "form smaller ones consistently, and integrating FFP can bring significant improvements in IoU scores. Our findings show that, irrespective of the SSP size and architecture, incorporating FFP improves the performance, and it is essential for segmentation tasks involving fine objects, e.g., text." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 347, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 526 + ], + "type": "text", + "content": "Loss Functions. We also perform an analysis to confirm the stability and convergence of the Fusion loss compared to other losses. Figures 5a and 5b compare different loss functions on the validation set of the SignaTR6K dataset. In Figure 5a, we observe that the Fusion loss is generally stable while converging to its minimum loss. Additionally, in Figure 5b, we observe Fusion loss reaches its maximum IoU around Epoch 15, generally faster than most other losses and stays close to the maximum IoU values on the validation set. It is important to highlight that the maximum IoU on the validation set for Focal and CE losses does not necessarily imply the best performance on the test set. Overall, Fusion loss shows a stable behavior and achieves a better performance on the test set, whereas for comparison, Focal loss shows instability at some epochs during the training." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 536, + 286, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 286, + 656 + ], + "type": "text", + "content": "Visual Comparison. Figure 6 shows sample model outputs for the SignaTR6K dataset, with two rectangle regions of interest highlighted to showcase the differences between various models. Figure 6b shows the ground truth, and a visual trend indicates that the performance on the regions of interest improves from Figure 6c to Figure 6p. It is also visually noticeable that CRF post-processing aggressively relabels pixels (6i), whereas CRFH generally improves the results. More visual comparisons on the WGM-SYN and SignaTR6K datasets can be found in the supplementary." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Limitations. Compared to prior work, the complexity and training cost of our approach present some limitations. Our MFM model, being larger than FCN-light, requires greater GPU memory sizes for training, and takes longer to train." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 273, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 429 + ], + "type": "text", + "content": "However, we believe some of the limitations on the training time can be offset through transferring weights from SSP models to MFM ones. Additionally, while our approach outperforms prior work on the WGM-SYN dataset, the mean IoU performance is lower compared to the SignaTR6K dataset. This indicates potential limitations in our method's efficacy for lower-quality original documents, those undergoing low-quality scanning processes (like historical documents), or those with errors from automated labeling. We also attribute the improved performance on the SignaTR6K dataset to its higher quality and our thorough manual annotation, that have resulted in better model training and improved IoU results." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 456, + 545, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 707 + ], + "type": "text", + "content": "Segmentation of handwritten text (HT) and printed text (PT) is vital for digitization and understanding of scanned documents. The complexity increases with the overlapping of different text types. In this research, we introduced SignaTR6K, a new open-source dataset with high-quality, manual pixel-level annotations, sourced from original legal documents. Additionally, we proposed a novel four-class formulation and a new architecture for the segmentation task. Our design leverages both the Fine Feature Path (FFP) and the Semantic Segmentation Path (SSP) to create the Mixed Feature Model (MFM), and incorporates both high-level and low-level features and improves on the text segmentation performance. We also introduced a CRF-based post-processing heuristic (CRFH) that further improves the model output, and included a new loss function, Fusion loss, that combines the advantages of different loss functions and achieves faster convergence and higher stability compared to most of the evaluated losses. In conclusion, our designs outperform the prior work in mean IoU scores by " + }, + { + "bbox": [ + 304, + 456, + 545, + 707 + ], + "type": "inline_equation", + "content": "17.9\\%" + }, + { + "bbox": [ + 304, + 456, + 545, + 707 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 545, + 707 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 304, + 456, + 545, + 707 + ], + "type": "text", + "content": " on WGM-SYN and SignaTR6K datasets, respectively." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "590" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 718 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 125 + ], + "type": "text", + "content": "[1] RIMES. http://www.a2ialab.com/doku.php?id=rimesdatabase:start. Inaccessible:2023-01-23." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 129, + 286, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 286, + 152 + ], + "type": "text", + "content": "[2] Scanned Questionnaire. https://github.com/jottue/HTSNet. Accessed: 2023-01-23." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 156, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 156, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 54, + 156, + 286, + 201 + ], + "type": "text", + "content": "[3] Thomson Reuters Legal Content Services. https://www.thomsonreuters.com/en/ products-services/legal.html. Accessed: 2023-01-21." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 205, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 286, + 275 + ], + "type": "text", + "content": "[4] Apostolos Antonacopoulos, David Bridson, Christos Papadopoulos, and Stefan Pletschacher. A realistic dataset for performance evaluation of document layout analysis. In 2009 10th International Conference on Document Analysis and Recognition, pages 296-300. IEEE, 2009." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 279, + 286, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 279, + 286, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 286, + 337 + ], + "type": "text", + "content": "[5] Dibyajyoti Dhar, Avishek Garain, Pawan Kumar Singh, and Ram Sarkar. Hp_docpres: A method for classifying printed and handwritten texts in doctor's prescription. Multimedia Tools and Applications, 80(7):9779-9812, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 341, + 286, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 286, + 411 + ], + "type": "text", + "content": "[6] Nicolas Dutly, Fouad Slimane, and Rolf Ingold. PHTI-WS: a printed and handwritten text identification web service based on FCN and CRF post-processing. In 2019 International Conference on Document Analysis and Recognition Workshops (IC-DARW), volume 2, pages 20-25. IEEE, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 415, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 287, + 485 + ], + "type": "text", + "content": "[7] Jürgen Franke and Matthias Oberlander. Writing style detection by statistical combination of classifiers in form reader applications. In Proceedings of 2nd International Conference on Document Analysis and Recognition (ICDAR'93), pages 581-584. IEEE, 1993." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 488, + 286, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 488, + 286, + 535 + ], + "spans": [ + { + "bbox": [ + 54, + 488, + 286, + 535 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 538, + 286, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 538, + 286, + 573 + ], + "spans": [ + { + "bbox": [ + 54, + 538, + 286, + 573 + ], + "type": "text", + "content": "[9] Pavel Iakubovskii. Segmentation models. https://github.com/qubvel/segmentation_models, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 575, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 286, + 634 + ], + "type": "text", + "content": "[10] Junho Jo, Hyung Il Koo, Jae Woong Soh, and Nam Ik Cho. Handwritten text segmentation via end-to-end learning of convolutional neural networks. *Multimedia Tools and Applications*, 79(43):32137-32150, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 638, + 286, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 638, + 286, + 718 + ], + "spans": [ + { + "bbox": [ + 48, + 638, + 286, + 718 + ], + "type": "text", + "content": "[11] Celso AML Junior, Matheus Henrique Marques da Silva, Byron Leite Dantas Bezerra, Bruno José Torres Fernandes, and Donato Impedovo. FCN+ RL: a fully convolutional network followed by refinement layers to offline handwritten signature segmentation. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 696 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 155 + ], + "type": "text", + "content": "[12] R Kandan, Nirup Kumar Reddy, KR Arvind, and AG Ramakrishnan. A robust two level classification algorithm for text localization in documents. In Advances in Visual Computing: Third International Symposium, ISVC 2007, Lake Tahoe, NV, USA, November 26-28, 2007, Proceedings, Part II 3, pages 96-105. Springer, 2007." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 159, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 545, + 229 + ], + "type": "text", + "content": "[13] Florian Kleber, Stefan Fiel, Markus Diem, and Robert Sablatnig. CVL-database: An off-line database for writer retrieval, writer identification and word spotting. In 2013 12th international conference on document analysis and recognition, pages 560-564. IEEE, 2013." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 234, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 545, + 304 + ], + "type": "text", + "content": "[14] Xiao-Hui Li, Fei Yin, and Cheng-Lin Liu. Printed/handwritten texts and graphics separation in complex documents using conditional random fields. In 2018 13th IAPR International Workshop on Document Analysis Systems (DAS), pages 145-150. IEEE, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 308, + 545, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 356 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 360, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 418 + ], + "type": "text", + "content": "[16] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 422, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 469 + ], + "type": "text", + "content": "[17] U-V Marti and Horst Bunke. The IAM-database: an English sentence database for offline handwriting recognition. International Journal on Document Analysis and Recognition, 5(1):39-46, 2002." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 473, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 473, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 473, + 545, + 555 + ], + "type": "text", + "content": "[18] Milda Norkute, Nadja Herger, Leszek Michalak, Andrew Mulder, and Sally Gao. Towards explainable ai: Assessing the usefulness and impact of added explainability features in legal document summarization. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-7, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 560, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 560, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 560, + 545, + 630 + ], + "type": "text", + "content": "[19] Xujun Peng, Srirangaraj Setlur, Venu Govindaraju, and Ramachandrula Sitaram. Handwritten text separation from annotated machine printed documents using markov random fields. International Journal on Document Analysis and Recognition (IJDAR), 16(1):1-16, 2013." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 634, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 658 + ], + "type": "text", + "content": "[20] Anastasia Prikhodina. Handwritten and printed text separation for historical documents. 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 661, + 545, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 661, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 661, + 545, + 696 + ], + "type": "text", + "content": "[21] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "591" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 322 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 69, + 72, + 286, + 96 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 286, + 96 + ], + "type": "text", + "content": "image computing and computer-assisted intervention, pages 234-241. Springer, 2015." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 99, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 99, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 99, + 287, + 158 + ], + "type": "text", + "content": "[22] Mathias Seuret, Marcus Liwicki, and Rolf Ingold. Pixel level handwritten and printed content discrimination in scanned documents. In 2014 14th International Conference on Frontiers in Handwriting Recognition, pages 423-428. IEEE, 2014." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 160, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 160, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 287, + 196 + ], + "type": "text", + "content": "[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 246 + ], + "type": "text", + "content": "[24] Stavroula Skylaki, Ali Oskooei, Omar Bari, Nadja Herger, and Zac Kriegman. Named entity recognition in the legal domain using a pointer generator network. arXiv preprint arXiv:2012.09936, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 248, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 248, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 248, + 287, + 296 + ], + "type": "text", + "content": "[25] Nishant Subramani, Alexandre Matton, Malcolm Greaves, and Adrian Lam. A survey of deep learning approaches forOCR and document understanding. arXiv preprint arXiv:2011.13534, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 322 + ], + "type": "text", + "content": "[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the in-" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 312 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 109 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 109 + ], + "type": "text", + "content": "ception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 111, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 111, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 111, + 545, + 171 + ], + "type": "text", + "content": "[27] Mahsa Vafaie, Oleksandra Bruns, Natasja Pilz, Jörg Waitelonis, and Harald Sack. Handwritten and printed text identification in historical archival documents. In Archiving Conference, volume 19, pages 15–20. Society for Imaging Science and Technology, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 174, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 209 + ], + "type": "text", + "content": "[28] Zhaobin Wang, E Wang, and Ying Zhu. Image segmentation evaluation: A survey of methods. Artificial Intelligence Review, 53:5637-5674, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 213, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 259 + ], + "type": "text", + "content": "[29] Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual U-Net. IEEE Geoscience and Remote Sensing Letters, 15(5):749-753, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 263, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 312 + ], + "type": "text", + "content": "[30] Zhi-Hua Zhou and Xu-Ying Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on knowledge and data engineering, 18(1):63-77, 2005." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "592" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_content_list.json b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c691c41891ce14a1dbc0eee84ff8328b7af2c167 --- /dev/null +++ b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient", + "text_level": 1, + "bbox": [ + 81, + 128, + 888, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengzhi Lu $^{1,2\\dagger}$ He Wang $^{3}$ Ziyi Chang $^{1}$ Guoan Yang $^{2}$ Hubert P. H. Shum $^{1\\dagger}$ $^{1}$ Durham University, UK $^{2}$ Xi'an Jiaotong University, China $^{3}$ University College London, UK \nlu947867114@stu.xjtu.edu.cn he_wang@ucl.ac.uk ziyi.chang@durham.ac.uk \ngayang@mail.xjtu.edu.cn hubert.shum@durham.ac.uk", + "bbox": [ + 99, + 200, + 870, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, methods for skeleton-based human activity recognition have been shown to be vulnerable to adversarial attacks. However, these attack methods require either the full knowledge of the victim (i.e. white-box attacks), access to training data (i.e. transfer-based attacks) or frequent model queries (i.e. black-box attacks). All their requirements are highly restrictive, raising the question of how detrimental the vulnerability is. In this paper, we show that the vulnerability indeed exists. To this end, we consider a new attack task: the attacker has no access to the victim model or the training data or labels, where we coin the term hard no-box attack. Specifically, we first learn a motion manifold where we define an adversarial loss to compute a new gradient for the attack, named skeleton-motion-informed (SMI) gradient. Our gradient contains information of the motion dynamics, which is different from existing gradient-based attack methods that compute the loss gradient assuming each dimension in the data is independent. The SMI gradient can augment many gradient-based attack methods, leading to a new family of no-box attack methods. Extensive evaluation and comparison show that our method imposes a real threat to existing classifiers. They also show that the SMI gradient improves the transferability and imperceptibility of adversarial samples in both no-box and transfer-based black-box settings.", + "bbox": [ + 76, + 340, + 473, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 734, + 209, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models are vulnerable to adversarial attacks, which compute data perturbations strategically to fool trained networks. Since its discovery [31], a wide variety of models in different tasks have been attacked [1], raising severe concerns as these perturbations are imperceptible to humans. Recently, the adversarial attack in skeleton-based human activity recognition (S-HAR) has attracted at-", + "bbox": [ + 75, + 758, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ffdee3e433b3216e6936752590d186e8457db25e83c119f9521f221ca3688388.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparisons on different settings of adversarial attacks. $\\checkmark$ and $\\times$ indicate if a method needs to access the corresponding information." + ], + "table_body": "
Information\nAccessibleWhite-\nBoxQueried\nBlack-BoxTransferred\nBlack-BoxNo-\nBoxHard\nNo-Box
Model Parameters✓××××
Queries of Victims✓✓×××
Training Samples××✓××
Labels✓✓✓✓×
", + "bbox": [ + 500, + 306, + 898, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tention as skeletal data have been widely used in security-critical applications such as sports analysis, bio-mechanics, surveillance, and human-computer interactions [24].", + "bbox": [ + 496, + 441, + 890, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing attacks in S-HAR are categorized into white-box and black-box approaches. White-box approaches require prior knowledge of the full details of a victim model [18, 36] while black-box approaches require a large number of queries to the victim model [7] or the access to training data and labels [36]. On the one hand, the victim model details and the training data and labels are unlikely to be available to the attacker in real-world scenarios. On the other hand, making frequent and numerous queries (e.g. tens of thousands) to the victim model is time-consuming and raises suspicion. In other words, the settings of existing S-HAR attacks are overly restrictive. A key to a successful attack is to reduce the required information of the victim model, training data and labels.", + "bbox": [ + 496, + 489, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a new threat model that requires no access to the victim model, training data or labels. We name the new threat model the hard no-box attack, differentiating from the recent no-box attack on images [16] that does not require access to the victim model but still needs access to the labels (i.e. soft no-box attack). Table 1 demonstrates the comparison on different settings of adversarial attacks. Among all attack settings, our hard no-box attack requires the least amount of knowledge, as it can only access the testing data without labels. Designing such an attack is nontrivial and challenging. Without access to the victim model, the attack method cannot rely on the gradient of a classification loss [12], data manipulation during", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† This work was conducted during the visit to the Durham University.", + "bbox": [ + 99, + 875, + 467, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding Author", + "bbox": [ + 99, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4597", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training [25], and the feedback of a classifier [2]. The challenge is further exacerbated by the requirement of no label and training sample access, where no surrogate model can be trained to attack or estimate the data distribution.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenges, we propose a contrastive learning (CL) [34] solution with a manifold-based no-box adversarial loss. First, we introduce a new application of CL to learn a latent data manifold where similar samples are naturally aggregated while dissimilar samples are dispersed without the need of class labels. It provides a good description of sample similarity that facilitates generating skeletal adversarial samples. CL is suitable for hard no-box attack settings due to its ability to capture the discriminative high-level features under our restricted attack conditions. Second, we compute the perturbation to drag a data sample away from its similar neighbors in the latent space, bounded by a pre-defined budget. In particular, we design a new no-box adversarial loss to maximize each adversary's dissimilarity with positive samples while minimizing its similarity with negative samples. The loss serves as guidance for the adversary search in our gradient-based attack scheme.", + "bbox": [ + 75, + 152, + 470, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While gradient-based attack methods like I-FGSM [14] are shown to be effective on S-HAR attacks [36, 18], the gradient is computed based on the victim model and the labels, making it unsuitable for hard no-box attacks. Since adversarial samples are likely to lie in or near the motion manifold [7], ideally, we want to explore along the manifold. That is, the computation of adversarial loss gradient should consider the local motion manifold.", + "bbox": [ + 75, + 410, + 468, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose to explicitly model motion dynamics for describing the local manifold around a given motion. Specifically, we introduce the skeleton-motion-informed (SMI) gradient that employs dynamics models (e.g. Markovian and autoregressive) to represent motion dynamics for the loss gradient computation. As a result, while existing methods generally assume each dimension in a data sample to be independent when computing the loss gradient, SMI gradient explicitly considers the dependency between frames in time. Furthermore, the SMI gradient is compatible with existing gradient-based methods including I-FGSM and MI-FGSM [8], allowing us to effectively construct a new family of no-box attack methods.", + "bbox": [ + 75, + 532, + 470, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments show that our method generates effective adversarial samples that successfully attack various victim models across datasets (HDM05, NTU60 and NTU120). Our SMI-gradient based attacks improve the attack transferability in both no-box and transferred black-box settings, with better imperceptibility. Codes are available in https://github.com/luyg45/HardNoBoxAttack and our contributions are:", + "bbox": [ + 75, + 729, + 468, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We confirm the S-HAR threat by introducing a new hard no-box attack and proposing the first method to generate adversarial samples without access to the victim model", + "bbox": [ + 81, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or training data or labels, to the best of our knowledge.", + "bbox": [ + 517, + 90, + 879, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new skeleton-motion-informed gradient that guides the adversary search along the motion manifold, explicitly considering the spatial-temporal nature of skeletal motions.", + "- We present a family of novel gradient-based attack strategies facilitated by the new gradient, improving the transferability and imperceptibility of adversarial samples in no-box and transferred black-box attacks." + ], + "bbox": [ + 504, + 107, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 239, + 648, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skeleton-Based Human Action Recognition S-HAR has attracted considerable attention in many applications [24] where deep learning-based approaches have achieved state-of-the-art performance [30, 13]. Recurrent neural networks are employed to model the temporal domain of human motions [9, 29]. Furthermore, unlike images and videos, the skeleton has a graph structure, so graph convolutional networks have shown to be effective in modelling the spatial or spatial-temporal features [28, 41]. The effectiveness is generally achieved by considering the skeleton as a topological graph where the joints and bones correspond to nodes and edges [21]. Improved graph designs and network architectures are subsequently proposed [20, 44, 45, 43].", + "bbox": [ + 496, + 266, + 890, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Attacks on Skeletons Adversarial attacks were initially introduced in [31], which showcases the vulnerability of deep neural networks and has been extended to other data types. Generally, adversarial attack is a special technique of data augmentation that aims to reveal the vulnerability of a system by finding new samples, while other data augmentation techniques may have diverse purposes, e.g. training efficiency and inference performance [27]. Recently, the attack on S-HAR has received increasing attention. Wang et al. [36] analyzed the perceptibility of adversarial skeletal samples and proposed a new perceptual loss. Liu et al. [18] focused on GCN-based models and utilized generative adversarial networks to synthesize adversarial examples. Tanaka et al. [32] proposed a new lower-dimensional attack, in which only the length of bones could be perturbed. These methods all require the complete knowledge of victim models, a setting known as the white-box attack. In contrast, Diao et.al [7] introduced the first black-box S-HAR attack method, which searches motion manifolds for adversaries. Still, black-box attacks need to frequently query the victim models, which can be infeasible in real-world systems. In contrast, we consider a more practical threat setting named the hard no-box attack, where an attacker only has access to unlabeled skeletal data.", + "bbox": [ + 496, + 463, + 892, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradient-Based Attack Strategies The core component of adversarial attacks is to generate adversarial samples [1]. Gradient-based attack methods have been widely used to introduce perturbations to a sample following the direction of the loss gradient. Goodfellow et al. [12] proposed", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4598", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd5f819e3929239968565e14cec3ac68b9cdb2f100e9c1864f1213d38f586cd0.jpg", + "image_caption": [ + "Figure 1. The training (left) and attack (right) processes for the hard no-box attack. The trained query encoder in the training process is used for attacks in the attack process." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 467, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the fast gradient sign method (FGSM) that perturbs a sample by a single step along the loss gradient. Kurakin et al. [14] proposed I-FGSM by extending FGSM to an iterative process. Dong et al. [8] presented MI-FGSM by adding momentum to the gradient, which boosted the transferability of adversarial samples. Xie et al. [40] applied diversified augmentations to the inputs before each iteration to craft more transferable samples. While these gradient-based strategies are successful in static data and have been adapted to skeletal motions, they neglect the dependency between frames for gradient computation, which is crucial in time series. This motivates us to propose our skeleton-motion-informed attack strategies, which explicitly model the motion dynamics in the temporal domain [39, 35].", + "bbox": [ + 75, + 359, + 472, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Hard No-Box Attack for Skeletal Data", + "text_level": 1, + "bbox": [ + 76, + 585, + 421, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1 shows the overview of our method. The left part is the training process where we adopt contrastive learning to obtain a latent data manifold to distinguish data samples. The attack process is shown on the right-hand side. We first design a new no-box adversarial loss in the trained latent space to guide the adversary search using samples that are dissimilar to the attacked sample. Then we propose a novel skeleton-motion-informed gradient and a new family of attack methods for generating adversarial samples.", + "bbox": [ + 75, + 609, + 468, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Contrastive Learning for Motion Manifold", + "text_level": 1, + "bbox": [ + 76, + 756, + 441, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the fundamental idea of adversarial attacks is to perturb a data sample to cross class boundaries, such boundaries cannot be estimated for hard no-box attacks due to the lack of labels. To estimate such boundaries without labels, we present a new application of contrastive learning (CL) [34] to aggregate similar data samples as soft class boundaries in latent space. Such boundaries enable us to adversarially perturb a sample to cross boundaries. We also train", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an encoder to extract discriminative high-level features for the motion manifold in latent space. Overall, our CL constructs boundaries in latent space without labels via aggregating similar samples and segregating dissimilar samples. Our attack is guided by the dissimilarity of high-level features between samples for generating adversarial samples, instead of using class boundaries to lead the attack [2].", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To incorporate both spatial and temporal information, we train an encoder (Fig. 1 left) based on adaptive graph convolutional network (AGCN) [28]. To force encoders to focus on high-level features, we apply skeleton-specific data augmentations to an input sequence $S$ and obtain two different views $S_{q}$ and $S_{k}$ . Augmentations include spatial operations (e.g. pose transformations, joint jittering, etc.) and temporal operations (e.g. temporal crop and resize) [34] (detailed in supplementary material). Then, we feed $S_{q}$ and $S_{k}$ into the query encoder $f_{q}$ and the key encoder $f_{k}$ respectively for the info-noise-contrastive estimation (InfoNCE) [23]:", + "bbox": [ + 496, + 198, + 892, + 363 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {c o n t r a s t}} = \\tag {1} \\\\ - \\log \\frac {\\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot f _ {k} \\left(S _ {k}\\right) / \\tau\\right)}{\\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot f _ {k} \\left(S _ {k}\\right) / \\tau\\right) + \\sum_ {F _ {n} \\sim N} \\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot F _ {n} / \\tau\\right)}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 371, + 911, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter, $N$ is the dynamic queue storing the features of negative samples $F_{n}$ obtained in the training process. After training, we use the query encoder $f_{q}$ , which encodes the motion manifold, for attack.", + "bbox": [ + 496, + 441, + 892, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adversarial Loss for Unlabeled Skeletal Data", + "text_level": 1, + "bbox": [ + 500, + 510, + 880, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The adversarial loss of hard no-box attack is significantly different from most existing methods that heavily depend on labels and class boundaries. Since class labels and class boundaries are unavailable in hard no-box attacks, we utilize data samples that are dissimilar to a given sample (i.e. negative samples) for defining the adversarial loss. Correspondingly, samples that are similar to the given sample are considered as positive samples. We argue when a given sample is perturbed towards its negative samples and away from its positive samples, it tends to become an adversary. This is because the negative samples generally indicate the high-density areas of other classes in the latent space.", + "bbox": [ + 496, + 532, + 890, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The hard no-box adversarial loss is designed as:", + "bbox": [ + 517, + 715, + 836, + 729 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a d v} = - \\log \\frac {\\exp [ S i m (f _ {q} (s) , f _ {q} (\\tilde {s})) ]}{\\sum_ {j} \\exp [ S i m (f _ {q} (s) , f _ {q} (\\tilde {s} _ {j})) ]}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 736, + 890, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Sim$ is the cosine similarity, $s$ is the adversarial sample to be computed, $\\widetilde{s}$ is the clean sample regarded as the positive sample, and $\\widetilde{s}_j$ are the negative samples. Maximizing $L_{adv}$ moves $s$ away from $\\widetilde{s}$ and towards $\\widetilde{s}_j$ in the latent manifold. With $L_{adv}$ , gradient-based attacks are employed.", + "bbox": [ + 496, + 779, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To maximize Eq. 2, the selection of negative samples $\\widetilde{s}_j$ is crucial and we design a method tailored for no-box attacks. Existing work [10] utilizes cluster-fit [42] to generate", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4599", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pseudo labels for selecting negative samples during adversarial training, which is less suitable for the no-box attack as it requires another pretrained off-line encoder to obtain pseudo labels. Instead, we adapt K-means, an unsupervised method, to select negatives, removing the need of any pretraining. We discard $Q$ clusters whose cluster centers are the closest to the input sample, mitigating the risk of misleading attacks. The remaining cluster centers are considered as the negative samples $\\tilde{s}_j$ .", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Skeleton-Motion-Informed Gradient", + "text_level": 1, + "bbox": [ + 76, + 241, + 408, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing gradient-based attack methods treat each dimension of the data as an independent variable, i.e. raw gradient. Attacks based on raw gradients tend to drag a sample away from the data manifold [11]. With the guidance of class boundaries and a limit on the perturbation budget, the raw gradient can still find deceiving adversaries. However, this setting is infeasible in hard no-box attacks. Without class boundaries, raw gradients that point to the negative samples can drag the adversary far away from the manifold. This is because while the perturbations are towards negative samples, they are not necessarily in a direction orthogonal to the class boundary. Consequently, larger perturbations are needed to cross the boundary, leading to adversaries being far off the manifold. This creates the need to constrain the perturbation within or near the manifold, at least locally. Since the motion manifold is constrained by the motion dynamics [38], we argue that the gradient needs to explicitly capture the dynamics. Therefore, we propose a new gradient named skeleton-motion-informed (SMI) gradient, capturing the manifold information that has been largely ignored by existing methods in loss gradient computation.", + "bbox": [ + 75, + 267, + 472, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dynamics in the Gradient Structure", + "text_level": 1, + "bbox": [ + 76, + 595, + 390, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a skeletal sequence $S = [S_{1}, S_{2}, \\dots, S_{t}]$ and the adversarial loss $J(S)$ , a straightforward but effective strategy to craft adversarial perturbations is the gradient-based attack [1]. It utilizes backpropagation of the loss function $\\nabla J(S)$ to iteratively change input samples $S$ :", + "bbox": [ + 75, + 618, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} = S + \\alpha \\cdot \\operatorname {s i g n} (\\nabla J (S)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 709, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is the attack step size and $\\hat{S}$ is the adversarial samples. In skeletal motions, this attack gradient $\\nabla J(S)$ consists of a set of partial derivatives over all frames:", + "bbox": [ + 75, + 734, + 468, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla J (S) = \\left[ \\frac {\\partial J (S)}{\\partial S _ {1}}, \\frac {\\partial J (S)}{\\partial S _ {2}}, \\dots , \\frac {\\partial J (S)}{\\partial S _ {t}} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 791, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The partial derivative $\\frac{\\partial J(S)}{\\partial S_t}$ assumes each frame is independent, and this is the raw gradient employed in existing methods [36]. However, human motions contain rich dynamics so that the system can be described as $S_{t} = f(S_{< t})$ . So", + "bbox": [ + 75, + 837, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "far, various dynamics models have been attempted to model human motions, such as Markovian models [33, 37], autoregressive models [39], and many-to-many mapping [38], all of which can capture the dynamics at different scales in time. We explore these models to reveal the missing dynamics in the structure of the raw gradient and propose our SMI-gradients that consider motion dynamics.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Markovian Model We assume the motion dynamics can be captured by a Markovian model, i.e. $S_{t} = f_{d1}(S_{t - 1})$ . This allows us to derive the 1st-order SMI-gradient:", + "bbox": [ + 496, + 214, + 890, + 260 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 1}}\\right) _ {d 1} = \\frac {\\partial J (S)}{\\partial S _ {t - 1}} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\frac {d S _ {t}}{d S _ {t - 1}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 268, + 890, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\frac{dS_t}{dS_{t-1}}$ is the temporal relationship between two consecutive frames that will be instantiated. Eq. 5 shows that the attack gradient along the motion manifold needs to consider the first-order information in the motion, e.g. velocity.", + "bbox": [ + 496, + 313, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Autoregressive Model Besides the first-order dynamics, we also model the second-order dynamics by assuming $S_{t} = f_{d2}(S_{t - 1},S_{t - 2})$ , as 2nd-order dynamics (i.e. joint acceleration) capture the smooth temporal dynamics of skeletal motion [36]. We extend Eq. 5 as:", + "bbox": [ + 496, + 393, + 890, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(\\frac {\\partial J (S)}{\\partial S _ {t - 2}}\\right) _ {d 2} = \\frac {\\partial J (S)}{\\partial S _ {t - 2}} + \\frac {\\partial J (S)}{\\partial S _ {t - 1}} \\cdot \\frac {\\partial S _ {t - 1}}{\\partial S _ {t - 2}} \\tag {6} \\\\ + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\left(\\frac {\\partial S _ {t}}{\\partial S _ {t - 2}} + \\frac {\\partial S _ {t}}{\\partial S _ {t - 1}} \\cdot \\frac {\\partial S _ {t - 1}}{\\partial S _ {t - 2}}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 491, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While higher-order models can also be considered, there is an empirical evidence that the first three orders are the most important in skeletal motion adversarial attack [36]. Therefore, we express the SMI gradients of the whole skeletal sequence as $(\\nabla J(S))_{d1} = \\left[(\\frac{\\partial J(S)}{\\partial S_1})_{d1},(\\frac{\\partial J(S)}{\\partial S_2})_{d1},\\dots ,(\\frac{\\partial J(S)}{\\partial S_t})_{d1}\\right]$ and $(\\nabla J(S))_{d2} = \\left[(\\frac{\\partial J(S)}{\\partial S_1})_{d2},(\\frac{\\partial J(S)}{\\partial S_2})_{d2},\\dots ,(\\frac{\\partial J(S)}{\\partial S_t})_{d2}\\right]$ .", + "bbox": [ + 496, + 568, + 890, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Time-Varying Autoregressive Models for Dynamics", + "text_level": 1, + "bbox": [ + 496, + 699, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ explicit models [39] to compute the dynamics-related derivatives in SMI gradients. While implicit models [33] may also be considered, they would require another network to be trained, making them less preferable in hard no-box attacks. To realize $f_{d1}$ and $f_{d2}$ , we use time-varying autoregressive models (TV-AR) [3], which effectively estimates the dynamics of skeleton sequence [39] due to its ability of modelling the temporal non-stationary signals:", + "bbox": [ + 496, + 737, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {d 1}: S _ {t} = A _ {t} \\cdot S _ {t - 1} + B _ {t} + \\gamma_ {t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4600", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {d 2}: S _ {t} = C _ {t} \\cdot S _ {t - 1} + D _ {t} \\cdot S _ {t - 2} + E _ {t} + \\gamma_ {t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 106, + 467, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Eq. 7 and Eq. 8 are denoted as TV-AR(1) and TV-AR(2) respectively. The model parameters $\\beta_{t}^{1} = [A_{t},B_{t}]$ and $\\beta_{t}^{2} = [C_{t},D_{t},E_{t}]$ are all time-varying parameters and determined by data-fitting. $\\gamma_{t}$ is a time-dependent white noise representing the dynamics of stochasticity.", + "bbox": [ + 76, + 128, + 467, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using Eq. 7 to compute $\\frac{\\partial S_t}{\\partial S_{t - 1}}$ , Eq. 5 becomes:", + "bbox": [ + 96, + 203, + 408, + 220 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 1}}\\right) _ {d 1} = \\frac {\\partial J (S)}{\\partial S _ {t - 1}} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot A _ {t}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 232, + 468, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, using Eq. 8, we can compute $C_t = \\frac{\\partial S_t}{\\partial S_{t-1}}$ and $D_t = \\frac{\\partial S_t}{\\partial S_{t-2}}$ . For $\\frac{\\partial S_{t-1}}{\\partial S_{t-2}}$ , we use $S_{t-1} = C_{t-1} \\cdot S_{t-2} + D_{t-1} \\cdot S_{t-3} + E_{t-1} + \\gamma_t$ to compute it: $C_{t-1} = \\frac{\\partial S_{t-1}}{\\partial S_{t-2}}$ . Then, Eq. 6 becomes:", + "bbox": [ + 76, + 277, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 2}}\\right) _ {d 2} = \\frac {\\partial J (S)}{\\partial S _ {t - 2}} + \\frac {\\partial J (S)}{\\partial S _ {t - 1}} \\cdot C _ {t - 1} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\left(D _ {t} + C _ {t} \\cdot C _ {t - 1}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 354, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Skeleton-Motion-Informed Attack", + "text_level": 1, + "bbox": [ + 76, + 407, + 385, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct new gradient-based attack methods based on our novel SMI gradient. Due to its compatibility, our proposed gradient can be integrated with most existing gradient-based methods. We select I-FGSM and MI-FGSM, which have been proven for their efficiency on S-HAR attacks [36, 18]. We augment them with first and second-order SMI-gradients, leading to four new attack methods.", + "bbox": [ + 75, + 431, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fast Gradient Sign Methods (FGSM) FGSM [12] is a single-step attack method that generates the adversarial samples $\\hat{S} = S + r$ by maximizing the adversarial loss function $J(S)$ , where $r$ denotes an adversarial perturbation that is constrained within a budget $\\|r\\|_p < \\epsilon$ , where $\\|\\cdot\\|_p$ denotes the $l_p$ -norm. One variant of FGSM is the Iterative Fast Gradient Sign Method (I-FGSM) [14], which extends FGSM to an iterative process:", + "bbox": [ + 75, + 539, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} (\\nabla_ {s} J (S)), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 670, + 468, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the attack step size and $i$ means iteration. Another variant is MI-FGSM [8], which considers the momentum of the attack to avoid local maxima:", + "bbox": [ + 75, + 700, + 468, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {i + 1} = \\mu \\cdot g ^ {i} + \\frac {\\nabla_ {s} J (S)}{\\| \\nabla_ {s} J (S) \\| _ {1}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 752, + 468, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} \\left(g ^ {i + 1}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 800, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ is the momentum decay factor and $g_{i}$ is the gradient in iteration $i$ .", + "bbox": [ + 75, + 825, + 468, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SMI-gradient Based Attacks We replace the original gradient $\\nabla J(S)$ in I-FGSM and MI-FGSM with our SMI gradient $(\\nabla J(S))_{d1}$ or $(\\nabla J(S))_{d2}$ . This creates four new", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 S1I-FGSM and S2I-FGSM", + "text_level": 1, + "bbox": [ + 500, + 90, + 767, + 104 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "Input: An encoder $k$ with a loss function $J$ ; a skeletal sequence sample $S$ ; the size of attack step $\\alpha$ ; iterations $I$ ; the budget of perturbation $\\epsilon$ . \nOutput: An adversarial example $\\hat{S}$ with $\\| \\hat{S} - S \\|_p < \\epsilon$ .", + "bbox": [ + 500, + 107, + 890, + 143 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "1: Initialization: $\\hat{S}^0 = S$", + "bbox": [ + 509, + 145, + 653, + 156 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "2: Fitting $S$ with TV-AR model to obtain the time-varying parameters $\\beta_{t}$", + "bbox": [ + 509, + 157, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "3: for $i = 0$ to $I - 1$ do", + "bbox": [ + 509, + 181, + 645, + 191 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "4: Input $S^i$ to $k$ ;", + "bbox": [ + 509, + 194, + 625, + 204 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "5: Obtain the raw gradient $\\nabla J(\\hat{S}^i)$ on $J$ ;", + "bbox": [ + 509, + 205, + 754, + 217 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "6: Calculate the SMI gradient $(\\nabla J(\\hat{S}^i))_{d1}$ with Eq. 9, or $(\\nabla J(\\hat{S}^i))_{d2}$ with Eq. 10, using $\\beta_{t}$ and $\\nabla J(\\hat{S}^i)$ ;", + "bbox": [ + 509, + 218, + 870, + 244 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "7: Update $\\tilde{S}^{i + 1}$ by applying the sign gradient as:", + "bbox": [ + 509, + 244, + 794, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} (\\nabla J (\\hat {S} ^ {i})) _ {d 1}), o r \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 263, + 890, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\mathrm {s i g n} \\left(\\nabla J (\\hat {S} ^ {i})\\right) _ {d 2}).\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 289, + 807, + 306 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "8: end for", + "bbox": [ + 509, + 313, + 571, + 323 + ], + "page_idx": 4 + }, + { + "type": "ref_text", + "text": "9: return $\\hat{S} = \\hat{S}^I$", + "bbox": [ + 509, + 325, + 612, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dynamic attack strategies: first-order SMI I-FGSM $(\\mathrm{S}_1\\mathrm{I}$ FGSM), second-order SMI I-FGSM $(\\mathrm{S}_2\\mathrm{I}$ -FGSM), firstorder SMI MI-FGSM $(\\mathrm{S}_1\\mathrm{MI}$ -FGSM), and second-order SMI MI-FGSM $(\\mathrm{S}_2\\mathrm{MI}$ -FGSM). The processes of SI-FGSM are shown in Algorithm 1. The algorithm of SMI-FGSM can be found in the supplementary material.", + "bbox": [ + 500, + 368, + 890, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 477, + 630, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We refer the readers to the supplementary material for extra experimental results.", + "bbox": [ + 500, + 503, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We select three widely used skeletal datasets: HDM05 [22] (2,337 sequences of 130 classes performed by 5 actors), NTU60 [26] (56,880 sequences of 60 classes), and NTU120 [19] (114,480 sequences of 120 classes, an extended version of NTU60, one of the largest datasets in the field). We pre-process HDM05 following [9] and both NTU datasets following [28]. The different skeletons are mapped to a standard 25-joint structure as in [38]. For our hard no-box attack, we only use the testing data and do not use the training data, the training labels and the testing labels during attacks.", + "bbox": [ + 496, + 535, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target Models We choose multiple state-of-the-art models as victims: ST-GCN [41], 2s-AGCN [28], AS-GCN [15], SGN [44] and MS-G3D [20]. They are trained using the official implementations and following the training protocols. For 2s-AGCN, we attack both the single joint stream model (js-AGCN) and the two-stream model.", + "bbox": [ + 496, + 703, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We pre-train the CL encoder $f_{q}$ following [34]. The unsupervised network is trained with a temperature value $\\tau = 0.07$ and SGD optimizer for 450 epochs. The learning rate is set to 0.01 with a weight decay of 0.0001. Due to the limitation of hard no-box settings, the attacker cannot access the training samples during the whole process. Therefore, the encoder is trained on the test-", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4601", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/78b6e54b327d6c70e16912cfd5adcf2a53449bb6aaeb0b6661fef11645c705f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1 I-FGSMNo-box S2 I-FGSMNo-box MI-FGSMNo-box S1 MI-FGSMNo-box S2 MI-FGSM
0.01js-AGCN11.21%—26.05%28.09%30.87%30.68%34.75%36.58%
2s-AGCN5.36%—13.94%15.04%16.45%16.47%18.23%19.31%
ST-GCN3.57%12.93%9.55%9.86%9.96%11.11%11.36%11.56%
MS-G3D8.39%35.23%10.57%11.14%11.69%11.76%12.85%14.11%
SGN21.25%26.03%34.09%34.46%35.23%38.49%38.75%38.81%
ASGCN5.69%20.87%13.92%14.85%14.67%15.95%16.82%17.75%
0.008js-AGCN10.12%—22.84%24.36%26.46%25.70%29.64%30.88%
2s-AGCN5.04%—11.36%12.19%13.03%12.30%13.85%15.56%
ST-GCN3.26%10.33%7.87%7.99%8.04%8.84%9.07%9.19%
MS-G3D5.19%31.28%9.18%9.51%9.99%10.01%10.26%12.98%
SGN20.95%23.30%29.61%30.20%30.74%33.32%33.63%34.54%
ASGCN5.54%18.77%11.42%12.05%12.29%12.77%13.69%14.37%
0.006js-AGCN7.19%—19.70%20.32%21.23%20.04%22.76%24.33%
2s-AGCN3.70%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN2.46%7.88%5.65%5.85%5.97%6.25%6.54%6.66%
MS-G3D4.46%23.15%7.30%7.84%7.61%7.94%8.05%8.39%
SGN16.76%19.93%23.92%24.57%25.75%26.74%27.04%27.64%
ASGCN3.94%14.30%8.79%9.23%9.07%9.71%10.29%10.90%
", + "bbox": [ + 192, + 87, + 777, + 297 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/635340cc4d5bdc709ddec659ca19ae8e148a2fa38e0bb1addc665bb3d2141a05.jpg", + "table_caption": [ + "Table 2. The fooling rate of different methods on the target models in NTU60, where $\\epsilon$ is the perturbation budget." + ], + "table_footnote": [], + "table_body": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN9.97%—23.92%24.64%25.79%27.26%27.93%29.07%
2s-AGCN6.38%—20.09%21.11%22.06%23.51%24.63%24.96%
ST-GCN12.18%23.84%23.53%24.73%25.77%26.95%27.31%28.76%
MS-G3D10.63%24.63%19.07%20.11%20.20%21.57%21.95%22.73%
SGN19.65%37.90%31.43%32.64%33.75%38.47%37.96%38.85%
ASGCN7.29%24.15%18.03%19.29%20.37%19.88%20.22%21.60%
0.008js-AGCN9.04%—21.58%21.77%22.47%24.28%24.62%25.74%
2s-AGCN5.79%—15.71%15.23%15.94%18.68%18.76%19.57%
ST-GCN11.23%21.35%20.74%21.51%22.22%23.52%24.65%24.87%
MS-G3D9.98%21.73%17.06%17.74%17.92%19.53%19.29%20.33%
SGN17.59%35.88%27.38%28.17%28.65%32.43%33.06%32.55%
ASGCN6.87%21.87%15.73%16.93%17.62%17.41%17.75%18.70%
0.006js-AGCN8.23%—18.51%18.88%18.79%20.74%21.88%22.06%
2s-AGCN5.09%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN9.54%17.51%17.60%17.97%18.67%18.90%19.91%20.34%
MS-G3D9.42%18.07%15.06%15.09%15.20%16.86%17.25%17.56%
SGN16.82%34.10%22.39%22.65%22.76%25.43%25.79%25.92%
ASGCN5.39%19.02%12.90%13.45%14.97%14.08%15.26%16.33%
", + "bbox": [ + 192, + 315, + 777, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. The fooling rate of different methods on the target models in NTU120, where $\\epsilon$ is the perturbation budget.", + "bbox": [ + 143, + 526, + 820, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing set. We adopt the $l_{\\infty}$ norm for the perturbation budget $\\epsilon$ . For clusters of negative samples, the number of clusters is 120, and the number of deleted centers $Q$ is 10.", + "bbox": [ + 75, + 555, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics We employ the fooling rate as a major metric. It is defined as the percentage of data samples whose predicted labels changed after adversarial attacks [18]. Besides, inspired by [36], we define a perceptual deviation indicator to evaluate the imperceptibility of adversarial skeletal samples:", + "bbox": [ + 75, + 604, + 468, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta p = \\frac {1}{M T} \\sum_ {n = 0} ^ {M} \\left\\| S - \\hat {S} \\right\\| _ {2} + \\frac {1}{M T} \\sum_ {n = 0} ^ {M} \\left\\| B - \\hat {B} \\right\\| _ {2} \\tag {15} \\\\ + \\frac {1}{M T L} \\sum_ {n = 0} ^ {M} \\left\\| \\ddot {S} - \\ddot {\\hat {S}} \\right\\| _ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 718, + 467, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the number of adversarial samples, $T$ is the total number of frames, and $L$ is the number of joints in the skeleton. The three terms evaluate the deviations of joint position, bone-length, and acceleration, respectively. A smaller perceptual deviation indicates better imperceptibility.", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Hard No-Box Attack", + "text_level": 1, + "bbox": [ + 500, + 554, + 694, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We establish two baselines using transfer-based attacks based on a self-supervised and a supervised classifier. This is because our method is the first hard no-box attack and there is no other similar method. The transfer-based attack is the closest setting to ours as they do not require access to the victim model. The first baseline (Self-sup Attacker) is a self-supervised surrogate model with a linear layer appended to our CL encoder. We freeze our CL encoder after the manifold training and then train the linear layer supervisedly [17]. In this way, the need to access labels is only for training the linear layer, not the CL encoder, which is not strictly hard no-box but closer than existing methods. The second baseline is a standard transfer-based attack where we use js-AGCN as the surrogate model and SMART [36] as the white-box attacker (AGCN Attacker). Unlike our method, both baselines still require access to the training data and the training labels during attacks.", + "bbox": [ + 496, + 580, + 890, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For hard no-box attack, different attack strategies are compared including I-FGSM, MI-FGSM, $\\mathrm{S}_1\\mathrm{I}$ -FGSM, $\\mathrm{S}_2\\mathrm{I}$ FGSM, $\\mathrm{S}_1\\mathrm{MI}$ -FGSM, and $\\mathrm{S}_2\\mathrm{MI}$ -FGSM. We run all the attackers for 400 iterations and report the attack perfor", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4602", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/37637b2e9a792a32b48ba5ab950a680e616e72d7003a41683b8ddd9a57cd9b99.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN10.12%—10.61%10.98%11.55%13.45%14.20%14.20%
2s-AGCN1.89%—5.30%5.68%5.68%6.44%6.44%6.82%
ST-GCN6.44%24.26%9.47%9.85%9.47%8.90%11.17%11.74%
MS-G3D4.17%86.95%41.67%44.51%43.94%53.03%56.06%53.41%
SGN1.89%3.98%2.46%2.84%3.03%3.03%3.40%3.59%
ASGCN1.89%27.57%3.22%4.36%3.22%4.36%5.68%5.68%
0.008js-AGCN9.46%—8.14%9.09%9.09%10.61%10.98%11.55%
2s-AGCN1.70%—4.36%4.92%4.73%4.55%5.30%5.87%
ST-GCN3.79%21.69%8.33%8.52%8.52%8.33%8.90%8.33%
MS-G3D3.60%82.17%25.76%30.68%29.17%36.36%39.39%39.02%
SGN1.51%3.60%0.57%1.51%2.08%2.27%2.27%2.46%
ASGCN1.70%22.43%2.46%2.84%2.84%2.27%2.84%3.03%
0.006js-AGCN7.19%—5.11%7.01%6.44%6.06%6.63%7.77%
2s-AGCN1.33%—2.08%2.08%2.27%3.79%3.98%4.17%
ST-GCN3.40%19.85%6.44%7.01%6.63%7.58%8.33%7.77%
MS-G3D2.84%72.43%11.55%12.31%11.92%15.34%18.94%17.99%
SGN0.57%3.40%0.13%0.94%0.57%1.13%1.32%1.51%
ASGCN1.52%17.82%0.38%0.76%1.89%0.57%2.27%2.27%
", + "bbox": [ + 184, + 87, + 785, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20813cfe522b2390c19926f60308b2ec58116b33b20cef0ed9b8f4e5b2a44e1e.jpg", + "image_caption": [ + "Figure 2. Visual comparisons between attack strategies in no-box attacks $(\\epsilon = 0.006)$ with key visual differences highlighted." + ], + "image_footnote": [], + "bbox": [ + 125, + 324, + 828, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance under different perturbation budgets $\\epsilon$ in Table 2 for NTU60, Table 3 for NTU120, and Table 4 for HDM05. We omit the results on js-AGCN and 2s-AGCN under AGCN Attacker because AGCN is the surrogate model.", + "bbox": [ + 75, + 468, + 468, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2-4 show that the hard no-box attack poses real threats to a range of S-HAR classifiers. In general, the fooling rate of hard no-box attacks is higher than Self-sup Attackers. This is surprising as the Attacker utilizes training data and training labels while our method does not. Looking further, AGCN Attacker is much better than Self-sup Attacker and the major difference is their feature extraction (i.e. one is a CL encoder and one is a graph network). This shows transfer-based attack heavily relies on the feature extraction ability of the surrogate model that cannot bypass access to the training data and the training labels. Furthermore, when compared with AGCN Attacker, our method achieves similar fooling rates, varying across different victims and datasets. Given that AGCN Attacker requires access to training data, training labels and testing labels, we argue hard no-box attacker achieves superior results and provides a more realistic setting.", + "bbox": [ + 75, + 532, + 468, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, among various no-box attack strategies, $\\mathrm{S}_2\\mathrm{MI}$ FGSM performs the best and often by big margins. All the SMI gradient-based methods generate stronger adversaries compared with baselines I-FGSM and MI-FGSM. The 2nd-order SMI attack method usually outperforms the corresponding 1st-order version. Last, we notice a variance in fooling rate across different victims and datasets. For in", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "stance, the multi-stream model (2s-AGCN) significantly enhances the robustness compared to the single-stream model (js-AGCN); the fooling rate of all methods drops by nearly half. This may be because the multi-stream model can ensemble features from different modalities, which improves the robustness. In general, it is still an open question why fooling rate can vary across victims and datasets and we leave the theoretical analysis for future work.", + "bbox": [ + 496, + 468, + 890, + 589 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ff7631e7a901c8e722597b50a8e69bc99d20558761143c8c212ae1b24071db9.jpg", + "table_caption": [ + "Table 4. The fooling rate of different methods on the target models in HDM05, where $\\epsilon$ is the perturbation budget." + ], + "table_footnote": [], + "table_body": "
Surrogate ModelVictimsIFGSMMI-FGSMS2I-FGSM (Ours)S2MI-FGSM (Ours)
2s-AGCNSTGCN2.10%2.10%3.00%3.01%
MS-G3D2.58%2.59%2.90%2.97%
ST-GCN2sAGCN2.20%2.34%2.44%2.64%
MS-G3D2.00%2.10%2.63%2.92%
MS-G3DSTGCN1.71%1.69%2.65%2.67%
2sAGCN1.76%1.79%2.03%2.07%
", + "bbox": [ + 537, + 603, + 856, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. The fooling rate of different attack strategies in transferred SMART attacks, where attack budgets $\\epsilon = 0.01$ .", + "bbox": [ + 500, + 694, + 890, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Transfer-Based Black-Box Attack", + "text_level": 1, + "bbox": [ + 500, + 739, + 794, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SMI gradient not only improves the transferability in hard no-box attacks, but also enhances other gradient-based skeletal attacks. Here, we employ SMART [36], a white-box attacker, as a baseline to compare different attack strategies. In the original SMART settings, I-FGSM is adopted to generate adversarial samples. We replace it with $\\mathrm{S}_2\\mathrm{I}$ FGSM, MI-FGSM, and $\\mathrm{S}_2\\mathrm{MI}$ -FGSM to make a comparison. As all strategies achieve similar fooling rates in white-box attacks, we mainly focus on their transferability using", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4603", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different surrogate models. We utilize SMART to attack 2s-AGCN, ST-GCN, and MS-G3D on NTU60 and transfer the obtained samples to other victim networks.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The performance of the transferred black-box attack is shown in Table 5. $\\mathrm{S}_2\\mathrm{MI}$ FGSM gives the best performance in transfer-based black-box attacks. $\\mathrm{S}_2\\mathrm{I}$ FGSM also improves the transferability of adversarial samples compared with baselines. In contrast, MI-FGSM, which succeeds in the image transfer-based attack, struggles in skeletal data. Its performance declines to $1.69\\%$ when it attacks STGCN via MS-G3D. Overall, the success rate is low in Table 5 because SMART is sensitive to the chosen surrogate [36]. Transfer-based attacks are proven to suffer from lower fooling rates in S-HAR attack [36] and improving the transferability is still an open problem. Table 5 aims to show our SMI gradient can improve it by incorporating motion dynamics into the attack gradient, compared with other alternative gradients. We will include in future work how to further explore this dynamics for better attack transfer.", + "bbox": [ + 75, + 137, + 470, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Perceptual Analysis", + "text_level": 1, + "bbox": [ + 76, + 391, + 264, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A key feature of SMI gradient-based attacks is the improvement in the perceptual quality of the adversarial samples due to the consideration of the motion manifold. To verify this, we employ quantitative comparison and qualitative visual analysis on the no-box adversarial samples under various strategies. We compare the perceptual quality $\\Delta p$ on NTU60 in Table 6. We find that $\\mathrm{S}_2\\mathrm{I}$ FGSM achieves the best imperceptibility and obtains a massive improvement compared with I-FGSM. In contrast, MI-FGSM's deviation is twice that of $\\mathrm{S}_2\\mathrm{I}$ FGSM. Although $\\mathrm{S}_2\\mathrm{MI}$ FGSM does not achieve the best visual performance, it is still slightly better than I-FGSM and achieves a better trade-off between fooling rate and perceptual quality. This is understandable because our method considers dynamics that help to generate more on-manifold adversarial samples. Moreover, the 1st-order SMI attacks outperform baselines but cannot compete with 2nd-order SMI attacks. This demonstrates the importance of considering acceleration in the skeletal attack.", + "bbox": [ + 75, + 415, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the visual comparison of poses under various attack strategies in no-box attacks in Figure 2. The spinal joints demonstrate the most obvious differences. $\\mathrm{S}_2\\mathrm{I}$ FGSM outperforms the other attack methods and gets the most natural poses, whereas I-FGSM has slight but noticeable joint displacements in the neck and head. $\\mathrm{S}_2\\mathrm{MI}$ -FGSM performs better than its baseline, MI-FGSM, which shows zig-zag bending in the frame $t_0 + 2$ .The samples produced by MI-FGSM have the worst imperceptibility, where we can easily find the unnatural jittery movements.", + "bbox": [ + 75, + 688, + 468, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also evaluate perceptual quality $\\Delta p$ on SMART adversarial samples obtained with different gradients. Results conducted on the NTU60 dataset are shown in Table 7. $\\mathrm{S}_2\\mathrm{I}$ FGSM reaches the best perceptual performance compared", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with all attack strategies. MI-FGSM slightly declines the imperceptibility. $\\mathrm{S}_2\\mathrm{MI}$ -FGSM performs slightly worse in MS-G3D and ST-GCN. The reason is mainly that $\\mathrm{S}_2\\mathrm{MI}$ FGSM takes more iterations in the white-box attack, leading to late stopping and slightly worse visual performance.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c0b567a28be45835006a52fbd061f43cf0ceac36dbaa10b3e33ccc4296d84be.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. The perceptual deviation of different attack strategies in no-box attacks with different budgets $\\epsilon$ ." + ], + "table_body": "
Strategiesε = 0.01ε = 0.008ε = 0.006
I-FGSM95.8768.5643.13
MI-FGSM131.7790.4454.61
S1I-FGSM84.6360.7738.15
S1MI-FGSM114.5178.0647.04
S2I-FGSM65.6045.6727.67
S2MI-FGSM90.9362.0437.46
", + "bbox": [ + 547, + 176, + 844, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f797aa31b492b1b2b9a64a8ec00994e3db8cf0ab787f7e8c587d2766182dafc0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. The perceptual deviation of different attack strategies in SMART for different victim models." + ], + "table_body": "
VictimsI-FGSMMI-FGSMS2I-FGSM(Ours)S2MI-FGSM(Ours)
2s-AGCN1.521.621.251.49
MS-G3D2.392.461.693.02
ST-GCN1.131.181.101.47
", + "bbox": [ + 553, + 292, + 838, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions and Discussions", + "text_level": 1, + "bbox": [ + 500, + 385, + 764, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have verified potential threats to S-HAR solutions. A new setting is proposed: the hard no-box attack on skeletal motions without access to the victim model, the training samples or the labels. We validate our setting by proposing the first pipeline for hard no-box attacks. Moreover, as far as we know, we are the first to explore motion dynamics in the adversarial gradient computation, leading to a new SMI gradient compatible with existing gradient-based attacks. By extensive evaluation and comparison, our method has been proven to be threatening and imperceptible, relying on the least prior knowledge.", + "bbox": [ + 496, + 410, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The SMI gradient also improves the transferability of transferred black-box attacks. Nonetheless, boosting the transferability is still an open problem [36] and we will explore this further with our SMI gradient in the future. We will explore other models (e.g., diffusion models [5]) and other time-series data (e.g. stock price, videos) for the proposed attack. Also, our SMI gradient describes dynamics and may be beneficial for motion synthesis [4].", + "bbox": [ + 496, + 577, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We call for attention to intensify the S-HAR robustness by considering defences against our hard no-box attack. We validate randomized smoothing [6] as a potential defence method in supplementary materials. Due to the least prior knowledge requirement, security risks posed by our attack can be reduced with such defenses. Otherwise, our attacks become a significantly threat to S-HAR.", + "bbox": [ + 496, + 696, + 890, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 815, + 666, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported in part by National Natural Science Foundation of China (ref: 61673314, Yang), EPSRC (ref: EP/X031012/1, NorthHFutures, Shum) and EU Horizon 2020 (ref: 899739, CrowdDNA, Wang).", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4604", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021.", + "[2] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In International Conference on Learning Representations, 2018.", + "[3] Laura F Bringmann, Ellen L Hamaker, Daniel E Vigo, André Aubert, Denny Borsboom, and Francis Tuerlinckx. Changing dynamics: Time-varying autoregressive models using generalized additive modeling. Psychological methods, 22(3):409, 2017.", + "[4] Ziyi Chang, Edmund JC Findlay, Haozheng Zhang, and Hubert PH Shum. Unifying human motion synthesis and style transfer with denoising diffusion probabilistic models. arXiv preprint arXiv:2212.08526, 2022.", + "[5] Ziyi Chang, George A Koulieris, and Hubert PH Shum. On the design fundamentals of diffusion models: A survey. arXiv preprint arXiv:2306.04542, 2023.", + "[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019.", + "[7] Yunfeng Diao, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and He Wang. Basar: Black-box attack on skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7597-7607, 2021.", + "[8] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018.", + "[9] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015.", + "[10] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in Neural Information Processing Systems, 34:21480-21492, 2021.", + "[11] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017.", + "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014.", + "[13] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017." + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world, pages 99-112. Chapman and Hall/CRC, 2018.", + "[15] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019.", + "[16] Qizhang Li, Yiwen Guo, and Hao Chen. Practical no-box adversarial attacks against dnns. Advances in Neural Information Processing Systems, 33:12849-12860, 2020.", + "[17] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2490-2498, 2019.", + "[18] Jian Liu, Naveed Akhtar, and Ajmal Mian. Adversarial attack on skeleton-based human action recognition. IEEE Transactions on Neural Networks and Learning Systems, 2020.", + "[19] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019.", + "[20] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020.", + "[21] Qianhui Men, Edmond S. L. Ho, Hubert P. H. Shum, and Howard Leung. A quadruple diffusion convolutional recurrent network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(9):3417-3432, 2021.", + "[22] Meinard Müller, Tido Röder, Michael Clausen, Bernhard Eberhardt, Björn Krüger, and Andreas Weber. Mocap database hdm05. Institut für Informatik II, Universität Bonn, 2(7), 2007.", + "[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.", + "[24] Bin Ren, Mengyuan Liu, Runwei Ding, and Hong Liu. A survey on 3d skeleton-based action recognition using learning method. arXiv preprint arXiv:2002.05907, 2020.", + "[25] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 11957-11965, 2020.", + "[26] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016.", + "[27] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1214-1223, October 2021." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4605", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019.", + "[29] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017.", + "[30] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017.", + "[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.", + "[32] Nariki Tanaka, Hiroshi Kera, and Kazuhiko Kawamoto. Adversarial bone length attack on action recognition. arXiv preprint arXiv:2109.05830, 2021.", + "[33] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qilong Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Trans. Graph., 41(4), jul 2022.", + "[34] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021.", + "[35] He Wang, Yunfeng Diao, Zichang Tan, and Guodong Guo. Defending black-box skeleton-based human activity classifiers. arXiv preprint arxiv.2203.04713, 2022.", + "[36] He Wang, Feixiang He, Zhexi Peng, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and David Hogg. Understanding the robustness of skeleton-based action recognition under adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14656-14665, 2021.", + "[37] He Wang, Edmond SL Ho, and Taku Komura. An energy-driven motion planning method for two distant postures. IEEE transactions on visualization and computer graphics, 21(1):18-30, 2015.", + "[38] He Wang, Edmond SL Ho, Hubert PH Shum, and Zhanxing Zhu. Spatio-temporal manifold learning for human motions via long-horizon modeling. IEEE transactions on visualization and computer graphics, 27(1):216-227, 2019.", + "[39] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015.", + "[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019.", + "[41] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "recognition. In Thirty-second AAAI conference on artificial intelligence, 2018.", + "[42] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. Clusterfit: Improving generalization of visual representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6509-6518, 2020.", + "[43] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. arXiv preprint arXiv:2211.13466, 2022.", + "[44] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1112-1121, 2020.", + "[45] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023." + ], + "bbox": [ + 501, + 92, + 893, + 387 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4606", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_model.json b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1cac503fac605fc813c65322edeff64fbe940d69 --- /dev/null +++ b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.13, + 0.89, + 0.175 + ], + "angle": 0, + "content": "Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.202, + 0.872, + 0.275 + ], + "angle": 0, + "content": "Zhengzhi Lu\\(^{1,2\\dagger}\\) He Wang\\(^{3}\\) Ziyi Chang\\(^{1}\\) Guoan Yang\\(^{2}\\) Hubert P. H. Shum\\(^{1\\dagger}\\) \n\\(^{1}\\)Durham University, UK \\(^{2}\\)Xi'an Jiaotong University, China \\(^{3}\\)University College London, UK \nlu947867114@stu.xjtu.edu.cn he_wang@ucl.ac.uk ziyi.chang@durham.ac.uk \ngayang@mail.xjtu.edu.cn hubert.shum@durham.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.474, + 0.719 + ], + "angle": 0, + "content": "Recently, methods for skeleton-based human activity recognition have been shown to be vulnerable to adversarial attacks. However, these attack methods require either the full knowledge of the victim (i.e. white-box attacks), access to training data (i.e. transfer-based attacks) or frequent model queries (i.e. black-box attacks). All their requirements are highly restrictive, raising the question of how detrimental the vulnerability is. In this paper, we show that the vulnerability indeed exists. To this end, we consider a new attack task: the attacker has no access to the victim model or the training data or labels, where we coin the term hard no-box attack. Specifically, we first learn a motion manifold where we define an adversarial loss to compute a new gradient for the attack, named skeleton-motion-informed (SMI) gradient. Our gradient contains information of the motion dynamics, which is different from existing gradient-based attack methods that compute the loss gradient assuming each dimension in the data is independent. The SMI gradient can augment many gradient-based attack methods, leading to a new family of no-box attack methods. Extensive evaluation and comparison show that our method imposes a real threat to existing classifiers. They also show that the SMI gradient improves the transferability and imperceptibility of adversarial samples in both no-box and transfer-based black-box settings." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.735, + 0.21, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Deep learning models are vulnerable to adversarial attacks, which compute data perturbations strategically to fool trained networks. Since its discovery [31], a wide variety of models in different tasks have been attacked [1], raising severe concerns as these perturbations are imperceptible to humans. Recently, the adversarial attack in skeleton-based human activity recognition (S-HAR) has attracted at-" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.308, + 0.9, + 0.386 + ], + "angle": 0, + "content": "
Information\nAccessibleWhite-\nBoxQueried\nBlack-BoxTransferred\nBlack-BoxNo-\nBoxHard\nNo-Box
Model Parameters✓××××
Queries of Victims✓✓×××
Training Samples××✓××
Labels✓✓✓✓×
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.386, + 0.892, + 0.425 + ], + "angle": 0, + "content": "Table 1. Comparisons on different settings of adversarial attacks. \\(\\checkmark\\) and \\(\\times\\) indicate if a method needs to access the corresponding information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.489 + ], + "angle": 0, + "content": "tention as skeletal data have been widely used in security-critical applications such as sports analysis, bio-mechanics, surveillance, and human-computer interactions [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.702 + ], + "angle": 0, + "content": "Existing attacks in S-HAR are categorized into white-box and black-box approaches. White-box approaches require prior knowledge of the full details of a victim model [18, 36] while black-box approaches require a large number of queries to the victim model [7] or the access to training data and labels [36]. On the one hand, the victim model details and the training data and labels are unlikely to be available to the attacker in real-world scenarios. On the other hand, making frequent and numerous queries (e.g. tens of thousands) to the victim model is time-consuming and raises suspicion. In other words, the settings of existing S-HAR attacks are overly restrictive. A key to a successful attack is to reduce the required information of the victim model, training data and labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce a new threat model that requires no access to the victim model, training data or labels. We name the new threat model the hard no-box attack, differentiating from the recent no-box attack on images [16] that does not require access to the victim model but still needs access to the labels (i.e. soft no-box attack). Table 1 demonstrates the comparison on different settings of adversarial attacks. Among all attack settings, our hard no-box attack requires the least amount of knowledge, as it can only access the testing data without labels. Designing such an attack is nontrivial and challenging. Without access to the victim model, the attack method cannot rely on the gradient of a classification loss [12], data manipulation during" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.468, + 0.888 + ], + "angle": 0, + "content": "† This work was conducted during the visit to the Durham University." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.234, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4597" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "training [25], and the feedback of a classifier [2]. The challenge is further exacerbated by the requirement of no label and training sample access, where no surrogate model can be trained to attack or estimate the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.409 + ], + "angle": 0, + "content": "To tackle the challenges, we propose a contrastive learning (CL) [34] solution with a manifold-based no-box adversarial loss. First, we introduce a new application of CL to learn a latent data manifold where similar samples are naturally aggregated while dissimilar samples are dispersed without the need of class labels. It provides a good description of sample similarity that facilitates generating skeletal adversarial samples. CL is suitable for hard no-box attack settings due to its ability to capture the discriminative high-level features under our restricted attack conditions. Second, we compute the perturbation to drag a data sample away from its similar neighbors in the latent space, bounded by a pre-defined budget. In particular, we design a new no-box adversarial loss to maximize each adversary's dissimilarity with positive samples while minimizing its similarity with negative samples. The loss serves as guidance for the adversary search in our gradient-based attack scheme." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.531 + ], + "angle": 0, + "content": "While gradient-based attack methods like I-FGSM [14] are shown to be effective on S-HAR attacks [36, 18], the gradient is computed based on the victim model and the labels, making it unsuitable for hard no-box attacks. Since adversarial samples are likely to lie in or near the motion manifold [7], ideally, we want to explore along the manifold. That is, the computation of adversarial loss gradient should consider the local motion manifold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.471, + 0.729 + ], + "angle": 0, + "content": "To this end, we propose to explicitly model motion dynamics for describing the local manifold around a given motion. Specifically, we introduce the skeleton-motion-informed (SMI) gradient that employs dynamics models (e.g. Markovian and autoregressive) to represent motion dynamics for the loss gradient computation. As a result, while existing methods generally assume each dimension in a data sample to be independent when computing the loss gradient, SMI gradient explicitly considers the dependency between frames in time. Furthermore, the SMI gradient is compatible with existing gradient-based methods including I-FGSM and MI-FGSM [8], allowing us to effectively construct a new family of no-box attack methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Extensive experiments show that our method generates effective adversarial samples that successfully attack various victim models across datasets (HDM05, NTU60 and NTU120). Our SMI-gradient based attacks improve the attack transferability in both no-box and transferred black-box settings, with better imperceptibility. Codes are available in https://github.com/luyg45/HardNoBoxAttack and our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We confirm the S-HAR threat by introducing a new hard no-box attack and proposing the first method to generate adversarial samples without access to the victim model" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.88, + 0.107 + ], + "angle": 0, + "content": "or training data or labels, to the best of our knowledge." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.108, + 0.892, + 0.166 + ], + "angle": 0, + "content": "- We propose a new skeleton-motion-informed gradient that guides the adversary search along the motion manifold, explicitly considering the spatial-temporal nature of skeletal motions." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.168, + 0.892, + 0.227 + ], + "angle": 0, + "content": "- We present a family of novel gradient-based attack strategies facilitated by the new gradient, improving the transferability and imperceptibility of adversarial samples in no-box and transferred black-box attacks." + }, + { + "type": "list", + "bbox": [ + 0.505, + 0.108, + 0.892, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.241, + 0.65, + 0.258 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Skeleton-Based Human Action Recognition S-HAR has attracted considerable attention in many applications [24] where deep learning-based approaches have achieved state-of-the-art performance [30, 13]. Recurrent neural networks are employed to model the temporal domain of human motions [9, 29]. Furthermore, unlike images and videos, the skeleton has a graph structure, so graph convolutional networks have shown to be effective in modelling the spatial or spatial-temporal features [28, 41]. The effectiveness is generally achieved by considering the skeleton as a topological graph where the joints and bones correspond to nodes and edges [21]. Improved graph designs and network architectures are subsequently proposed [20, 44, 45, 43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.893, + 0.825 + ], + "angle": 0, + "content": "Adversarial Attacks on Skeletons Adversarial attacks were initially introduced in [31], which showcases the vulnerability of deep neural networks and has been extended to other data types. Generally, adversarial attack is a special technique of data augmentation that aims to reveal the vulnerability of a system by finding new samples, while other data augmentation techniques may have diverse purposes, e.g. training efficiency and inference performance [27]. Recently, the attack on S-HAR has received increasing attention. Wang et al. [36] analyzed the perceptibility of adversarial skeletal samples and proposed a new perceptual loss. Liu et al. [18] focused on GCN-based models and utilized generative adversarial networks to synthesize adversarial examples. Tanaka et al. [32] proposed a new lower-dimensional attack, in which only the length of bones could be perturbed. These methods all require the complete knowledge of victim models, a setting known as the white-box attack. In contrast, Diao et.al [7] introduced the first black-box S-HAR attack method, which searches motion manifolds for adversaries. Still, black-box attacks need to frequently query the victim models, which can be infeasible in real-world systems. In contrast, we consider a more practical threat setting named the hard no-box attack, where an attacker only has access to unlabeled skeletal data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Gradient-Based Attack Strategies The core component of adversarial attacks is to generate adversarial samples [1]. Gradient-based attack methods have been widely used to introduce perturbations to a sample following the direction of the loss gradient. Goodfellow et al. [12] proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4598" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.468, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.348 + ], + "angle": 0, + "content": "Figure 1. The training (left) and attack (right) processes for the hard no-box attack. The trained query encoder in the training process is used for attacks in the attack process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.473, + 0.573 + ], + "angle": 0, + "content": "the fast gradient sign method (FGSM) that perturbs a sample by a single step along the loss gradient. Kurakin et al. [14] proposed I-FGSM by extending FGSM to an iterative process. Dong et al. [8] presented MI-FGSM by adding momentum to the gradient, which boosted the transferability of adversarial samples. Xie et al. [40] applied diversified augmentations to the inputs before each iteration to craft more transferable samples. While these gradient-based strategies are successful in static data and have been adapted to skeletal motions, they neglect the dependency between frames for gradient computation, which is crucial in time series. This motivates us to propose our skeleton-motion-informed attack strategies, which explicitly model the motion dynamics in the temporal domain [39, 35]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.422, + 0.603 + ], + "angle": 0, + "content": "3. Hard No-Box Attack for Skeletal Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Figure 1 shows the overview of our method. The left part is the training process where we adopt contrastive learning to obtain a latent data manifold to distinguish data samples. The attack process is shown on the right-hand side. We first design a new no-box adversarial loss in the trained latent space to guide the adversary search using samples that are dissimilar to the attacked sample. Then we propose a novel skeleton-motion-informed gradient and a new family of attack methods for generating adversarial samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.442, + 0.773 + ], + "angle": 0, + "content": "3.1. Contrastive Learning for Motion Manifold" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "While the fundamental idea of adversarial attacks is to perturb a data sample to cross class boundaries, such boundaries cannot be estimated for hard no-box attacks due to the lack of labels. To estimate such boundaries without labels, we present a new application of contrastive learning (CL) [34] to aggregate similar data samples as soft class boundaries in latent space. Such boundaries enable us to adversarially perturb a sample to cross boundaries. We also train" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "an encoder to extract discriminative high-level features for the motion manifold in latent space. Overall, our CL constructs boundaries in latent space without labels via aggregating similar samples and segregating dissimilar samples. Our attack is guided by the dissimilarity of high-level features between samples for generating adversarial samples, instead of using class boundaries to lead the attack [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.364 + ], + "angle": 0, + "content": "To incorporate both spatial and temporal information, we train an encoder (Fig. 1 left) based on adaptive graph convolutional network (AGCN) [28]. To force encoders to focus on high-level features, we apply skeleton-specific data augmentations to an input sequence \\( S \\) and obtain two different views \\( S_{q} \\) and \\( S_{k} \\). Augmentations include spatial operations (e.g. pose transformations, joint jittering, etc.) and temporal operations (e.g. temporal crop and resize) [34] (detailed in supplementary material). Then, we feed \\( S_{q} \\) and \\( S_{k} \\) into the query encoder \\( f_{q} \\) and the key encoder \\( f_{k} \\) respectively for the info-noise-contrastive estimation (InfoNCE) [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.372, + 0.912, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {c o n t r a s t}} = \\tag {1} \\\\ - \\log \\frac {\\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot f _ {k} \\left(S _ {k}\\right) / \\tau\\right)}{\\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot f _ {k} \\left(S _ {k}\\right) / \\tau\\right) + \\sum_ {F _ {n} \\sim N} \\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot F _ {n} / \\tau\\right)}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.893, + 0.504 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter, \\(N\\) is the dynamic queue storing the features of negative samples \\(F_{n}\\) obtained in the training process. After training, we use the query encoder \\(f_{q}\\), which encodes the motion manifold, for attack." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.882, + 0.525 + ], + "angle": 0, + "content": "3.2. Adversarial Loss for Unlabeled Skeletal Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.715 + ], + "angle": 0, + "content": "The adversarial loss of hard no-box attack is significantly different from most existing methods that heavily depend on labels and class boundaries. Since class labels and class boundaries are unavailable in hard no-box attacks, we utilize data samples that are dissimilar to a given sample (i.e. negative samples) for defining the adversarial loss. Correspondingly, samples that are similar to the given sample are considered as positive samples. We argue when a given sample is perturbed towards its negative samples and away from its positive samples, it tends to become an adversary. This is because the negative samples generally indicate the high-density areas of other classes in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.716, + 0.838, + 0.731 + ], + "angle": 0, + "content": "The hard no-box adversarial loss is designed as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.737, + 0.892, + 0.773 + ], + "angle": 0, + "content": "\\[\nL _ {a d v} = - \\log \\frac {\\exp [ S i m (f _ {q} (s) , f _ {q} (\\tilde {s})) ]}{\\sum_ {j} \\exp [ S i m (f _ {q} (s) , f _ {q} (\\tilde {s} _ {j})) ]}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.856 + ], + "angle": 0, + "content": "where \\( Sim \\) is the cosine similarity, \\( s \\) is the adversarial sample to be computed, \\( \\widetilde{s} \\) is the clean sample regarded as the positive sample, and \\( \\widetilde{s}_j \\) are the negative samples. Maximizing \\( L_{adv} \\) moves \\( s \\) away from \\( \\widetilde{s} \\) and towards \\( \\widetilde{s}_j \\) in the latent manifold. With \\( L_{adv} \\), gradient-based attacks are employed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To maximize Eq. 2, the selection of negative samples \\(\\widetilde{s}_j\\) is crucial and we design a method tailored for no-box attacks. Existing work [10] utilizes cluster-fit [42] to generate" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "pseudo labels for selecting negative samples during adversarial training, which is less suitable for the no-box attack as it requires another pretrained off-line encoder to obtain pseudo labels. Instead, we adapt K-means, an unsupervised method, to select negatives, removing the need of any pretraining. We discard \\(Q\\) clusters whose cluster centers are the closest to the input sample, mitigating the risk of misleading attacks. The remaining cluster centers are considered as the negative samples \\(\\tilde{s}_j\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.242, + 0.41, + 0.259 + ], + "angle": 0, + "content": "4. Skeleton-Motion-Informed Gradient" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.473, + 0.587 + ], + "angle": 0, + "content": "Existing gradient-based attack methods treat each dimension of the data as an independent variable, i.e. raw gradient. Attacks based on raw gradients tend to drag a sample away from the data manifold [11]. With the guidance of class boundaries and a limit on the perturbation budget, the raw gradient can still find deceiving adversaries. However, this setting is infeasible in hard no-box attacks. Without class boundaries, raw gradients that point to the negative samples can drag the adversary far away from the manifold. This is because while the perturbations are towards negative samples, they are not necessarily in a direction orthogonal to the class boundary. Consequently, larger perturbations are needed to cross the boundary, leading to adversaries being far off the manifold. This creates the need to constrain the perturbation within or near the manifold, at least locally. Since the motion manifold is constrained by the motion dynamics [38], we argue that the gradient needs to explicitly capture the dynamics. Therefore, we propose a new gradient named skeleton-motion-informed (SMI) gradient, capturing the manifold information that has been largely ignored by existing methods in loss gradient computation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.596, + 0.391, + 0.612 + ], + "angle": 0, + "content": "4.1. Dynamics in the Gradient Structure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.697 + ], + "angle": 0, + "content": "Given a skeletal sequence \\( S = [S_{1}, S_{2}, \\dots, S_{t}] \\) and the adversarial loss \\( J(S) \\), a straightforward but effective strategy to craft adversarial perturbations is the gradient-based attack [1]. It utilizes backpropagation of the loss function \\( \\nabla J(S) \\) to iteratively change input samples \\( S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.71, + 0.47, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\hat {S} = S + \\alpha \\cdot \\operatorname {s i g n} (\\nabla J (S)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.782 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the attack step size and \\(\\hat{S}\\) is the adversarial samples. In skeletal motions, this attack gradient \\(\\nabla J(S)\\) consists of a set of partial derivatives over all frames:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.792, + 0.47, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\nabla J (S) = \\left[ \\frac {\\partial J (S)}{\\partial S _ {1}}, \\frac {\\partial J (S)}{\\partial S _ {2}}, \\dots , \\frac {\\partial J (S)}{\\partial S _ {t}} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.838, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The partial derivative \\(\\frac{\\partial J(S)}{\\partial S_t}\\) assumes each frame is independent, and this is the raw gradient employed in existing methods [36]. However, human motions contain rich dynamics so that the system can be described as \\(S_{t} = f(S_{< t})\\). So" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "far, various dynamics models have been attempted to model human motions, such as Markovian models [33, 37], autoregressive models [39], and many-to-many mapping [38], all of which can capture the dynamics at different scales in time. We explore these models to reveal the missing dynamics in the structure of the raw gradient and propose our SMI-gradients that consider motion dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Markovian Model We assume the motion dynamics can be captured by a Markovian model, i.e. \\(S_{t} = f_{d1}(S_{t - 1})\\). This allows us to derive the 1st-order SMI-gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.269, + 0.892, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 1}}\\right) _ {d 1} = \\frac {\\partial J (S)}{\\partial S _ {t - 1}} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\frac {d S _ {t}}{d S _ {t - 1}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.379 + ], + "angle": 0, + "content": "where \\(\\frac{dS_t}{dS_{t-1}}\\) is the temporal relationship between two consecutive frames that will be instantiated. Eq. 5 shows that the attack gradient along the motion manifold needs to consider the first-order information in the motion, e.g. velocity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Autoregressive Model Besides the first-order dynamics, we also model the second-order dynamics by assuming \\( S_{t} = f_{d2}(S_{t - 1},S_{t - 2}) \\), as 2nd-order dynamics (i.e. joint acceleration) capture the smooth temporal dynamics of skeletal motion [36]. We extend Eq. 5 as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.492, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(\\frac {\\partial J (S)}{\\partial S _ {t - 2}}\\right) _ {d 2} = \\frac {\\partial J (S)}{\\partial S _ {t - 2}} + \\frac {\\partial J (S)}{\\partial S _ {t - 1}} \\cdot \\frac {\\partial S _ {t - 1}}{\\partial S _ {t - 2}} \\tag {6} \\\\ + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\left(\\frac {\\partial S _ {t}}{\\partial S _ {t - 2}} + \\frac {\\partial S _ {t}}{\\partial S _ {t - 1}} \\cdot \\frac {\\partial S _ {t - 1}}{\\partial S _ {t - 2}}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.694 + ], + "angle": 0, + "content": "While higher-order models can also be considered, there is an empirical evidence that the first three orders are the most important in skeletal motion adversarial attack [36]. Therefore, we express the SMI gradients of the whole skeletal sequence as \\((\\nabla J(S))_{d1} = \\left[(\\frac{\\partial J(S)}{\\partial S_1})_{d1},(\\frac{\\partial J(S)}{\\partial S_2})_{d1},\\dots ,(\\frac{\\partial J(S)}{\\partial S_t})_{d1}\\right]\\) and \\((\\nabla J(S))_{d2} = \\left[(\\frac{\\partial J(S)}{\\partial S_1})_{d2},(\\frac{\\partial J(S)}{\\partial S_2})_{d2},\\dots ,(\\frac{\\partial J(S)}{\\partial S_t})_{d2}\\right]\\)." + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.731 + ], + "angle": 0, + "content": "4.2. Time-Varying Autoregressive Models for Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.892, + 0.875 + ], + "angle": 0, + "content": "We employ explicit models [39] to compute the dynamics-related derivatives in SMI gradients. While implicit models [33] may also be considered, they would require another network to be trained, making them less preferable in hard no-box attacks. To realize \\( f_{d1} \\) and \\( f_{d2} \\), we use time-varying autoregressive models (TV-AR) [3], which effectively estimates the dynamics of skeleton sequence [39] due to its ability of modelling the temporal non-stationary signals:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\nf _ {d 1}: S _ {t} = A _ {t} \\cdot S _ {t - 1} + B _ {t} + \\gamma_ {t}, \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4600" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.122, + 0.107, + 0.468, + 0.123 + ], + "angle": 0, + "content": "\\[\nf _ {d 2}: S _ {t} = C _ {t} \\cdot S _ {t - 1} + D _ {t} \\cdot S _ {t - 2} + E _ {t} + \\gamma_ {t}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.129, + 0.468, + 0.204 + ], + "angle": 0, + "content": "where Eq. 7 and Eq. 8 are denoted as TV-AR(1) and TV-AR(2) respectively. The model parameters \\(\\beta_{t}^{1} = [A_{t},B_{t}]\\) and \\(\\beta_{t}^{2} = [C_{t},D_{t},E_{t}]\\) are all time-varying parameters and determined by data-fitting. \\(\\gamma_{t}\\) is a time-dependent white noise representing the dynamics of stochasticity." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.204, + 0.41, + 0.222 + ], + "angle": 0, + "content": "Using Eq. 7 to compute \\(\\frac{\\partial S_t}{\\partial S_{t - 1}}\\), Eq. 5 becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.233, + 0.469, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 1}}\\right) _ {d 1} = \\frac {\\partial J (S)}{\\partial S _ {t - 1}} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot A _ {t}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.278, + 0.469, + 0.35 + ], + "angle": 0, + "content": "Similarly, using Eq. 8, we can compute \\( C_t = \\frac{\\partial S_t}{\\partial S_{t-1}} \\) and \\( D_t = \\frac{\\partial S_t}{\\partial S_{t-2}} \\). For \\( \\frac{\\partial S_{t-1}}{\\partial S_{t-2}} \\), we use \\( S_{t-1} = C_{t-1} \\cdot S_{t-2} + D_{t-1} \\cdot S_{t-3} + E_{t-1} + \\gamma_t \\) to compute it: \\( C_{t-1} = \\frac{\\partial S_{t-1}}{\\partial S_{t-2}} \\). Then, Eq. 6 becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.355, + 0.469, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 2}}\\right) _ {d 2} = \\frac {\\partial J (S)}{\\partial S _ {t - 2}} + \\frac {\\partial J (S)}{\\partial S _ {t - 1}} \\cdot C _ {t - 1} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\left(D _ {t} + C _ {t} \\cdot C _ {t - 1}\\right). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.387, + 0.424 + ], + "angle": 0, + "content": "5. Skeleton-Motion-Informed Attack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.469, + 0.539 + ], + "angle": 0, + "content": "We construct new gradient-based attack methods based on our novel SMI gradient. Due to its compatibility, our proposed gradient can be integrated with most existing gradient-based methods. We select I-FGSM and MI-FGSM, which have been proven for their efficiency on S-HAR attacks [36, 18]. We augment them with first and second-order SMI-gradients, leading to four new attack methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.469, + 0.66 + ], + "angle": 0, + "content": "Fast Gradient Sign Methods (FGSM) FGSM [12] is a single-step attack method that generates the adversarial samples \\(\\hat{S} = S + r\\) by maximizing the adversarial loss function \\(J(S)\\), where \\(r\\) denotes an adversarial perturbation that is constrained within a budget \\(\\|r\\|_p < \\epsilon\\), where \\(\\|\\cdot\\|_p\\) denotes the \\(l_p\\)-norm. One variant of FGSM is the Iterative Fast Gradient Sign Method (I-FGSM) [14], which extends FGSM to an iterative process:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.671, + 0.469, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} (\\nabla_ {s} J (S)), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.469, + 0.744 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the attack step size and \\(i\\) means iteration. Another variant is MI-FGSM [8], which considers the momentum of the attack to avoid local maxima:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.753, + 0.469, + 0.787 + ], + "angle": 0, + "content": "\\[\ng ^ {i + 1} = \\mu \\cdot g ^ {i} + \\frac {\\nabla_ {s} J (S)}{\\| \\nabla_ {s} J (S) \\| _ {1}} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.801, + 0.469, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} \\left(g ^ {i + 1}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.854 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the momentum decay factor and \\(g_{i}\\) is the gradient in iteration \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "SMI-gradient Based Attacks We replace the original gradient \\(\\nabla J(S)\\) in I-FGSM and MI-FGSM with our SMI gradient \\((\\nabla J(S))_{d1}\\) or \\((\\nabla J(S))_{d2}\\). This creates four new" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.769, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 S1I-FGSM and S2I-FGSM" + }, + { + "type": "ref_text", + "bbox": [ + 0.501, + 0.108, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Input: An encoder \\(k\\) with a loss function \\(J\\); a skeletal sequence sample \\(S\\); the size of attack step \\(\\alpha\\); iterations \\(I\\); the budget of perturbation \\(\\epsilon\\). \nOutput: An adversarial example \\(\\hat{S}\\) with \\(\\| \\hat{S} - S \\|_p < \\epsilon\\)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.146, + 0.654, + 0.157 + ], + "angle": 0, + "content": "1: Initialization: \\(\\hat{S}^0 = S\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.159, + 0.891, + 0.181 + ], + "angle": 0, + "content": "2: Fitting \\(S\\) with TV-AR model to obtain the time-varying parameters \\(\\beta_{t}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.183, + 0.646, + 0.193 + ], + "angle": 0, + "content": "3: for \\( i = 0 \\) to \\( I - 1 \\) do" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.195, + 0.626, + 0.205 + ], + "angle": 0, + "content": "4: Input \\(S^i\\) to \\(k\\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.206, + 0.756, + 0.218 + ], + "angle": 0, + "content": "5: Obtain the raw gradient \\(\\nabla J(\\hat{S}^i)\\) on \\(J\\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.219, + 0.872, + 0.245 + ], + "angle": 0, + "content": "6: Calculate the SMI gradient \\((\\nabla J(\\hat{S}^i))_{d1}\\) with Eq. 9, or \\((\\nabla J(\\hat{S}^i))_{d2}\\) with Eq. 10, using \\(\\beta_{t}\\) and \\(\\nabla J(\\hat{S}^i)\\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.246, + 0.795, + 0.258 + ], + "angle": 0, + "content": "7: Update \\(\\tilde{S}^{i + 1}\\) by applying the sign gradient as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.265, + 0.891, + 0.292 + ], + "angle": 0, + "content": "\\[\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} (\\nabla J (\\hat {S} ^ {i})) _ {d 1}), o r \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.29, + 0.808, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\mathrm {s i g n} \\left(\\nabla J (\\hat {S} ^ {i})\\right) _ {d 2}).\n\\]" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.314, + 0.573, + 0.324 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.326, + 0.613, + 0.336 + ], + "angle": 0, + "content": "9: return \\(\\hat{S} = \\hat{S}^I\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.369, + 0.892, + 0.46 + ], + "angle": 0, + "content": "dynamic attack strategies: first-order SMI I-FGSM \\((\\mathrm{S}_1\\mathrm{I}\\) FGSM), second-order SMI I-FGSM \\((\\mathrm{S}_2\\mathrm{I}\\) -FGSM), firstorder SMI MI-FGSM \\((\\mathrm{S}_1\\mathrm{MI}\\) -FGSM), and second-order SMI MI-FGSM \\((\\mathrm{S}_2\\mathrm{MI}\\) -FGSM). The processes of SI-FGSM are shown in Algorithm 1. The algorithm of SMI-FGSM can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.632, + 0.495 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.505, + 0.891, + 0.534 + ], + "angle": 0, + "content": "We refer the readers to the supplementary material for extra experimental results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Datasets We select three widely used skeletal datasets: HDM05 [22] (2,337 sequences of 130 classes performed by 5 actors), NTU60 [26] (56,880 sequences of 60 classes), and NTU120 [19] (114,480 sequences of 120 classes, an extended version of NTU60, one of the largest datasets in the field). We pre-process HDM05 following [9] and both NTU datasets following [28]. The different skeletons are mapped to a standard 25-joint structure as in [38]. For our hard no-box attack, we only use the testing data and do not use the training data, the training labels and the testing labels during attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Target Models We choose multiple state-of-the-art models as victims: ST-GCN [41], 2s-AGCN [28], AS-GCN [15], SGN [44] and MS-G3D [20]. They are trained using the official implementations and following the training protocols. For 2s-AGCN, we attack both the single joint stream model (js-AGCN) and the two-stream model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details We pre-train the CL encoder \\( f_{q} \\) following [34]. The unsupervised network is trained with a temperature value \\( \\tau = 0.07 \\) and SGD optimizer for 450 epochs. The learning rate is set to 0.01 with a weight decay of 0.0001. Due to the limitation of hard no-box settings, the attacker cannot access the training samples during the whole process. Therefore, the encoder is trained on the test-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4601" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.193, + 0.088, + 0.778, + 0.299 + ], + "angle": 0, + "content": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1 I-FGSMNo-box S2 I-FGSMNo-box MI-FGSMNo-box S1 MI-FGSMNo-box S2 MI-FGSM
0.01js-AGCN11.21%—26.05%28.09%30.87%30.68%34.75%36.58%
2s-AGCN5.36%—13.94%15.04%16.45%16.47%18.23%19.31%
ST-GCN3.57%12.93%9.55%9.86%9.96%11.11%11.36%11.56%
MS-G3D8.39%35.23%10.57%11.14%11.69%11.76%12.85%14.11%
SGN21.25%26.03%34.09%34.46%35.23%38.49%38.75%38.81%
ASGCN5.69%20.87%13.92%14.85%14.67%15.95%16.82%17.75%
0.008js-AGCN10.12%—22.84%24.36%26.46%25.70%29.64%30.88%
2s-AGCN5.04%—11.36%12.19%13.03%12.30%13.85%15.56%
ST-GCN3.26%10.33%7.87%7.99%8.04%8.84%9.07%9.19%
MS-G3D5.19%31.28%9.18%9.51%9.99%10.01%10.26%12.98%
SGN20.95%23.30%29.61%30.20%30.74%33.32%33.63%34.54%
ASGCN5.54%18.77%11.42%12.05%12.29%12.77%13.69%14.37%
0.006js-AGCN7.19%—19.70%20.32%21.23%20.04%22.76%24.33%
2s-AGCN3.70%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN2.46%7.88%5.65%5.85%5.97%6.25%6.54%6.66%
MS-G3D4.46%23.15%7.30%7.84%7.61%7.94%8.05%8.39%
SGN16.76%19.93%23.92%24.57%25.75%26.74%27.04%27.64%
ASGCN3.94%14.30%8.79%9.23%9.07%9.71%10.29%10.90%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.3, + 0.817, + 0.312 + ], + "angle": 0, + "content": "Table 2. The fooling rate of different methods on the target models in NTU60, where \\( \\epsilon \\) is the perturbation budget." + }, + { + "type": "table", + "bbox": [ + 0.193, + 0.316, + 0.778, + 0.526 + ], + "angle": 0, + "content": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN9.97%—23.92%24.64%25.79%27.26%27.93%29.07%
2s-AGCN6.38%—20.09%21.11%22.06%23.51%24.63%24.96%
ST-GCN12.18%23.84%23.53%24.73%25.77%26.95%27.31%28.76%
MS-G3D10.63%24.63%19.07%20.11%20.20%21.57%21.95%22.73%
SGN19.65%37.90%31.43%32.64%33.75%38.47%37.96%38.85%
ASGCN7.29%24.15%18.03%19.29%20.37%19.88%20.22%21.60%
0.008js-AGCN9.04%—21.58%21.77%22.47%24.28%24.62%25.74%
2s-AGCN5.79%—15.71%15.23%15.94%18.68%18.76%19.57%
ST-GCN11.23%21.35%20.74%21.51%22.22%23.52%24.65%24.87%
MS-G3D9.98%21.73%17.06%17.74%17.92%19.53%19.29%20.33%
SGN17.59%35.88%27.38%28.17%28.65%32.43%33.06%32.55%
ASGCN6.87%21.87%15.73%16.93%17.62%17.41%17.75%18.70%
0.006js-AGCN8.23%—18.51%18.88%18.79%20.74%21.88%22.06%
2s-AGCN5.09%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN9.54%17.51%17.60%17.97%18.67%18.90%19.91%20.34%
MS-G3D9.42%18.07%15.06%15.09%15.20%16.86%17.25%17.56%
SGN16.82%34.10%22.39%22.65%22.76%25.43%25.79%25.92%
ASGCN5.39%19.02%12.90%13.45%14.97%14.08%15.26%16.33%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.527, + 0.821, + 0.54 + ], + "angle": 0, + "content": "Table 3. The fooling rate of different methods on the target models in NTU120, where \\( \\epsilon \\) is the perturbation budget." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.469, + 0.601 + ], + "angle": 0, + "content": "ing set. We adopt the \\(l_{\\infty}\\) norm for the perturbation budget \\(\\epsilon\\). For clusters of negative samples, the number of clusters is 120, and the number of deleted centers \\(Q\\) is 10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.469, + 0.697 + ], + "angle": 0, + "content": "Evaluation Metrics We employ the fooling rate as a major metric. It is defined as the percentage of data samples whose predicted labels changed after adversarial attacks [18]. Besides, inspired by [36], we define a perceptual deviation indicator to evaluate the imperceptibility of adversarial skeletal samples:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.719, + 0.468, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta p = \\frac {1}{M T} \\sum_ {n = 0} ^ {M} \\left\\| S - \\hat {S} \\right\\| _ {2} + \\frac {1}{M T} \\sum_ {n = 0} ^ {M} \\left\\| B - \\hat {B} \\right\\| _ {2} \\tag {15} \\\\ + \\frac {1}{M T L} \\sum_ {n = 0} ^ {M} \\left\\| \\ddot {S} - \\ddot {\\hat {S}} \\right\\| _ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( M \\) is the number of adversarial samples, \\( T \\) is the total number of frames, and \\( L \\) is the number of joints in the skeleton. The three terms evaluate the deviations of joint position, bone-length, and acceleration, respectively. A smaller perceptual deviation indicates better imperceptibility." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.695, + 0.57 + ], + "angle": 0, + "content": "6.1. Hard No-Box Attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Baselines We establish two baselines using transfer-based attacks based on a self-supervised and a supervised classifier. This is because our method is the first hard no-box attack and there is no other similar method. The transfer-based attack is the closest setting to ours as they do not require access to the victim model. The first baseline (Self-sup Attacker) is a self-supervised surrogate model with a linear layer appended to our CL encoder. We freeze our CL encoder after the manifold training and then train the linear layer supervisedly [17]. In this way, the need to access labels is only for training the linear layer, not the CL encoder, which is not strictly hard no-box but closer than existing methods. The second baseline is a standard transfer-based attack where we use js-AGCN as the surrogate model and SMART [36] as the white-box attacker (AGCN Attacker). Unlike our method, both baselines still require access to the training data and the training labels during attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For hard no-box attack, different attack strategies are compared including I-FGSM, MI-FGSM, \\(\\mathrm{S}_1\\mathrm{I}\\) -FGSM, \\(\\mathrm{S}_2\\mathrm{I}\\) FGSM, \\(\\mathrm{S}_1\\mathrm{MI}\\) -FGSM, and \\(\\mathrm{S}_2\\mathrm{MI}\\) -FGSM. We run all the attackers for 400 iterations and report the attack perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4602" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.185, + 0.088, + 0.787, + 0.299 + ], + "angle": 0, + "content": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN10.12%—10.61%10.98%11.55%13.45%14.20%14.20%
2s-AGCN1.89%—5.30%5.68%5.68%6.44%6.44%6.82%
ST-GCN6.44%24.26%9.47%9.85%9.47%8.90%11.17%11.74%
MS-G3D4.17%86.95%41.67%44.51%43.94%53.03%56.06%53.41%
SGN1.89%3.98%2.46%2.84%3.03%3.03%3.40%3.59%
ASGCN1.89%27.57%3.22%4.36%3.22%4.36%5.68%5.68%
0.008js-AGCN9.46%—8.14%9.09%9.09%10.61%10.98%11.55%
2s-AGCN1.70%—4.36%4.92%4.73%4.55%5.30%5.87%
ST-GCN3.79%21.69%8.33%8.52%8.52%8.33%8.90%8.33%
MS-G3D3.60%82.17%25.76%30.68%29.17%36.36%39.39%39.02%
SGN1.51%3.60%0.57%1.51%2.08%2.27%2.27%2.46%
ASGCN1.70%22.43%2.46%2.84%2.84%2.27%2.84%3.03%
0.006js-AGCN7.19%—5.11%7.01%6.44%6.06%6.63%7.77%
2s-AGCN1.33%—2.08%2.08%2.27%3.79%3.98%4.17%
ST-GCN3.40%19.85%6.44%7.01%6.63%7.58%8.33%7.77%
MS-G3D2.84%72.43%11.55%12.31%11.92%15.34%18.94%17.99%
SGN0.57%3.40%0.13%0.94%0.57%1.13%1.32%1.51%
ASGCN1.52%17.82%0.38%0.76%1.89%0.57%2.27%2.27%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.3, + 0.82, + 0.312 + ], + "angle": 0, + "content": "Table 4. The fooling rate of different methods on the target models in HDM05, where \\( \\epsilon \\) is the perturbation budget." + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.325, + 0.83, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.442, + 0.848, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Visual comparisons between attack strategies in no-box attacks \\((\\epsilon = 0.006)\\) with key visual differences highlighted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.53 + ], + "angle": 0, + "content": "mance under different perturbation budgets \\(\\epsilon\\) in Table 2 for NTU60, Table 3 for NTU120, and Table 4 for HDM05. We omit the results on js-AGCN and 2s-AGCN under AGCN Attacker because AGCN is the surrogate model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.791 + ], + "angle": 0, + "content": "Table 2-4 show that the hard no-box attack poses real threats to a range of S-HAR classifiers. In general, the fooling rate of hard no-box attacks is higher than Self-sup Attackers. This is surprising as the Attacker utilizes training data and training labels while our method does not. Looking further, AGCN Attacker is much better than Self-sup Attacker and the major difference is their feature extraction (i.e. one is a CL encoder and one is a graph network). This shows transfer-based attack heavily relies on the feature extraction ability of the surrogate model that cannot bypass access to the training data and the training labels. Furthermore, when compared with AGCN Attacker, our method achieves similar fooling rates, varying across different victims and datasets. Given that AGCN Attacker requires access to training data, training labels and testing labels, we argue hard no-box attacker achieves superior results and provides a more realistic setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In addition, among various no-box attack strategies, \\(\\mathrm{S}_2\\mathrm{MI}\\) FGSM performs the best and often by big margins. All the SMI gradient-based methods generate stronger adversaries compared with baselines I-FGSM and MI-FGSM. The 2nd-order SMI attack method usually outperforms the corresponding 1st-order version. Last, we notice a variance in fooling rate across different victims and datasets. For in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.59 + ], + "angle": 0, + "content": "stance, the multi-stream model (2s-AGCN) significantly enhances the robustness compared to the single-stream model (js-AGCN); the fooling rate of all methods drops by nearly half. This may be because the multi-stream model can ensemble features from different modalities, which improves the robustness. In general, it is still an open question why fooling rate can vary across victims and datasets and we leave the theoretical analysis for future work." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.604, + 0.857, + 0.695 + ], + "angle": 0, + "content": "
Surrogate ModelVictimsIFGSMMI-FGSMS2I-FGSM (Ours)S2MI-FGSM (Ours)
2s-AGCNSTGCN2.10%2.10%3.00%3.01%
MS-G3D2.58%2.59%2.90%2.97%
ST-GCN2sAGCN2.20%2.34%2.44%2.64%
MS-G3D2.00%2.10%2.63%2.92%
MS-G3DSTGCN1.71%1.69%2.65%2.67%
2sAGCN1.76%1.79%2.03%2.07%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.695, + 0.892, + 0.721 + ], + "angle": 0, + "content": "Table 5. The fooling rate of different attack strategies in transferred SMART attacks, where attack budgets \\(\\epsilon = 0.01\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.795, + 0.755 + ], + "angle": 0, + "content": "6.2. Transfer-Based Black-Box Attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "SMI gradient not only improves the transferability in hard no-box attacks, but also enhances other gradient-based skeletal attacks. Here, we employ SMART [36], a white-box attacker, as a baseline to compare different attack strategies. In the original SMART settings, I-FGSM is adopted to generate adversarial samples. We replace it with \\(\\mathrm{S}_2\\mathrm{I}\\) FGSM, MI-FGSM, and \\(\\mathrm{S}_2\\mathrm{MI}\\) -FGSM to make a comparison. As all strategies achieve similar fooling rates in white-box attacks, we mainly focus on their transferability using" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.137 + ], + "angle": 0, + "content": "different surrogate models. We utilize SMART to attack 2s-AGCN, ST-GCN, and MS-G3D on NTU60 and transfer the obtained samples to other victim networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.381 + ], + "angle": 0, + "content": "The performance of the transferred black-box attack is shown in Table 5. \\(\\mathrm{S}_2\\mathrm{MI}\\) FGSM gives the best performance in transfer-based black-box attacks. \\(\\mathrm{S}_2\\mathrm{I}\\) FGSM also improves the transferability of adversarial samples compared with baselines. In contrast, MI-FGSM, which succeeds in the image transfer-based attack, struggles in skeletal data. Its performance declines to \\(1.69\\%\\) when it attacks STGCN via MS-G3D. Overall, the success rate is low in Table 5 because SMART is sensitive to the chosen surrogate [36]. Transfer-based attacks are proven to suffer from lower fooling rates in S-HAR attack [36] and improving the transferability is still an open problem. Table 5 aims to show our SMI gradient can improve it by incorporating motion dynamics into the attack gradient, compared with other alternative gradients. We will include in future work how to further explore this dynamics for better attack transfer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.392, + 0.266, + 0.408 + ], + "angle": 0, + "content": "6.3. Perceptual Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.688 + ], + "angle": 0, + "content": "A key feature of SMI gradient-based attacks is the improvement in the perceptual quality of the adversarial samples due to the consideration of the motion manifold. To verify this, we employ quantitative comparison and qualitative visual analysis on the no-box adversarial samples under various strategies. We compare the perceptual quality \\(\\Delta p\\) on NTU60 in Table 6. We find that \\(\\mathrm{S}_2\\mathrm{I}\\) FGSM achieves the best imperceptibility and obtains a massive improvement compared with I-FGSM. In contrast, MI-FGSM's deviation is twice that of \\(\\mathrm{S}_2\\mathrm{I}\\) FGSM. Although \\(\\mathrm{S}_2\\mathrm{MI}\\) FGSM does not achieve the best visual performance, it is still slightly better than I-FGSM and achieves a better trade-off between fooling rate and perceptual quality. This is understandable because our method considers dynamics that help to generate more on-manifold adversarial samples. Moreover, the 1st-order SMI attacks outperform baselines but cannot compete with 2nd-order SMI attacks. This demonstrates the importance of considering acceleration in the skeletal attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We show the visual comparison of poses under various attack strategies in no-box attacks in Figure 2. The spinal joints demonstrate the most obvious differences. \\(\\mathrm{S}_2\\mathrm{I}\\) FGSM outperforms the other attack methods and gets the most natural poses, whereas I-FGSM has slight but noticeable joint displacements in the neck and head. \\(\\mathrm{S}_2\\mathrm{MI}\\) -FGSM performs better than its baseline, MI-FGSM, which shows zig-zag bending in the frame \\(t_0 + 2\\) .The samples produced by MI-FGSM have the worst imperceptibility, where we can easily find the unnatural jittery movements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We also evaluate perceptual quality \\(\\Delta p\\) on SMART adversarial samples obtained with different gradients. Results conducted on the NTU60 dataset are shown in Table 7. \\(\\mathrm{S}_2\\mathrm{I}\\) FGSM reaches the best perceptual performance compared" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "with all attack strategies. MI-FGSM slightly declines the imperceptibility. \\(\\mathrm{S}_2\\mathrm{MI}\\) -FGSM performs slightly worse in MS-G3D and ST-GCN. The reason is mainly that \\(\\mathrm{S}_2\\mathrm{MI}\\) FGSM takes more iterations in the white-box attack, leading to late stopping and slightly worse visual performance." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.178, + 0.845, + 0.251 + ], + "angle": 0, + "content": "
Strategiesε = 0.01ε = 0.008ε = 0.006
I-FGSM95.8768.5643.13
MI-FGSM131.7790.4454.61
S1I-FGSM84.6360.7738.15
S1MI-FGSM114.5178.0647.04
S2I-FGSM65.6045.6727.67
S2MI-FGSM90.9362.0437.46
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.251, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Table 6. The perceptual deviation of different attack strategies in no-box attacks with different budgets \\(\\epsilon\\)." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.293, + 0.839, + 0.346 + ], + "angle": 0, + "content": "
VictimsI-FGSMMI-FGSMS2I-FGSM(Ours)S2MI-FGSM(Ours)
2s-AGCN1.521.621.251.49
MS-G3D2.392.461.693.02
ST-GCN1.131.181.101.47
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.346, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Table 7. The perceptual deviation of different attack strategies in SMART for different victim models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.386, + 0.765, + 0.402 + ], + "angle": 0, + "content": "7. Conclusions and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.577 + ], + "angle": 0, + "content": "In this paper, we have verified potential threats to S-HAR solutions. A new setting is proposed: the hard no-box attack on skeletal motions without access to the victim model, the training samples or the labels. We validate our setting by proposing the first pipeline for hard no-box attacks. Moreover, as far as we know, we are the first to explore motion dynamics in the adversarial gradient computation, leading to a new SMI gradient compatible with existing gradient-based attacks. By extensive evaluation and comparison, our method has been proven to be threatening and imperceptible, relying on the least prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.698 + ], + "angle": 0, + "content": "The SMI gradient also improves the transferability of transferred black-box attacks. Nonetheless, boosting the transferability is still an open problem [36] and we will explore this further with our SMI gradient in the future. We will explore other models (e.g., diffusion models [5]) and other time-series data (e.g. stock price, videos) for the proposed attack. Also, our SMI gradient describes dynamics and may be beneficial for motion synthesis [4]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.804 + ], + "angle": 0, + "content": "We call for attention to intensify the S-HAR robustness by considering defences against our hard no-box attack. We validate randomized smoothing [6] as a potential defence method in supplementary materials. Due to the least prior knowledge requirement, security risks posed by our attack can be reduced with such defenses. Otherwise, our attacks become a significantly threat to S-HAR." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.667, + 0.833 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research is supported in part by National Natural Science Foundation of China (ref: 61673314, Yang), EPSRC (ref: EP/X031012/1, NorthHFutures, Shum) and EU Horizon 2020 (ref: 899739, CrowdDNA, Wang)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4604" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[2] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.234, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[3] Laura F Bringmann, Ellen L Hamaker, Daniel E Vigo, André Aubert, Denny Borsboom, and Francis Tuerlinckx. Changing dynamics: Time-varying autoregressive models using generalized additive modeling. Psychological methods, 22(3):409, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[4] Ziyi Chang, Edmund JC Findlay, Haozheng Zhang, and Hubert PH Shum. Unifying human motion synthesis and style transfer with denoising diffusion probabilistic models. arXiv preprint arXiv:2212.08526, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.365, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[5] Ziyi Chang, George A Koulieris, and Hubert PH Shum. On the design fundamentals of diffusion models: A survey. arXiv preprint arXiv:2306.04542, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[7] Yunfeng Diao, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and He Wang. Basar: Black-box attack on skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7597-7607, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.54, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[8] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.612, + 0.469, + 0.667 + ], + "angle": 0, + "content": "[9] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.738 + ], + "angle": 0, + "content": "[10] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in Neural Information Processing Systems, 34:21480-21492, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[11] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[14] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world, pages 99-112. Chapman and Hall/CRC, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Qizhang Li, Yiwen Guo, and Hao Chen. Practical no-box adversarial attacks against dnns. Advances in Neural Information Processing Systems, 33:12849-12860, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[17] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2490-2498, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[18] Jian Liu, Naveed Akhtar, and Ajmal Mian. Adversarial attack on skeleton-based human action recognition. IEEE Transactions on Neural Networks and Learning Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[19] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[21] Qianhui Men, Edmond S. L. Ho, Hubert P. H. Shum, and Howard Leung. A quadruple diffusion convolutional recurrent network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(9):3417-3432, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[22] Meinard Müller, Tido Röder, Michael Clausen, Bernhard Eberhardt, Björn Krüger, and Andreas Weber. Mocap database hdm05. Institut für Informatik II, Universität Bonn, 2(7), 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Bin Ren, Mengyuan Liu, Runwei Ding, and Hong Liu. A survey on 3d skeleton-based action recognition using learning method. arXiv preprint arXiv:2002.05907, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[25] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 11957-11965, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1214-1223, October 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4605" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[29] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[30] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[32] Nariki Tanaka, Hiroshi Kera, and Kazuhiko Kawamoto. Adversarial bone length attack on action recognition. arXiv preprint arXiv:2109.05830, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[33] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qilong Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Trans. Graph., 41(4), jul 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[34] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[35] He Wang, Yunfeng Diao, Zichang Tan, and Guodong Guo. Defending black-box skeleton-based human activity classifiers. arXiv preprint arxiv.2203.04713, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[36] He Wang, Feixiang He, Zhexi Peng, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and David Hogg. Understanding the robustness of skeleton-based action recognition under adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14656-14665, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[37] He Wang, Edmond SL Ho, and Taku Komura. An energy-driven motion planning method for two distant postures. IEEE transactions on visualization and computer graphics, 21(1):18-30, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[38] He Wang, Edmond SL Ho, Hubert PH Shum, and Zhanxing Zhu. Spatio-temporal manifold learning for human motions via long-horizon modeling. IEEE transactions on visualization and computer graphics, 27(1):216-227, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[39] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[41] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "recognition. In Thirty-second AAAI conference on artificial intelligence, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[42] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. Clusterfit: Improving generalization of visual representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6509-6518, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[43] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. arXiv preprint arXiv:2211.13466, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[44] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1112-1121, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.894, + 0.388 + ], + "angle": 0, + "content": "[45] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4606" + } + ] +] \ No newline at end of file diff --git a/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_origin.pdf b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3e2423cf747654bacc9305c594668a897dc0137f --- /dev/null +++ b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/58af11a8-198e-4cb8-99cf-9272139ca7d9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de716d41994c2c7d7ec05222e4235546a58e78d15e6f581f1f4df32193f1b2f3 +size 1952701 diff --git a/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/full.md b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f34409a244e598cf5aee5a5aa12ed9da08fdb06 --- /dev/null +++ b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/full.md @@ -0,0 +1,357 @@ +# Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient + +Zhengzhi Lu $^{1,2\dagger}$ He Wang $^{3}$ Ziyi Chang $^{1}$ Guoan Yang $^{2}$ Hubert P. H. Shum $^{1\dagger}$ $^{1}$ Durham University, UK $^{2}$ Xi'an Jiaotong University, China $^{3}$ University College London, UK +lu947867114@stu.xjtu.edu.cn he_wang@ucl.ac.uk ziyi.chang@durham.ac.uk +gayang@mail.xjtu.edu.cn hubert.shum@durham.ac.uk + +# Abstract + +Recently, methods for skeleton-based human activity recognition have been shown to be vulnerable to adversarial attacks. However, these attack methods require either the full knowledge of the victim (i.e. white-box attacks), access to training data (i.e. transfer-based attacks) or frequent model queries (i.e. black-box attacks). All their requirements are highly restrictive, raising the question of how detrimental the vulnerability is. In this paper, we show that the vulnerability indeed exists. To this end, we consider a new attack task: the attacker has no access to the victim model or the training data or labels, where we coin the term hard no-box attack. Specifically, we first learn a motion manifold where we define an adversarial loss to compute a new gradient for the attack, named skeleton-motion-informed (SMI) gradient. Our gradient contains information of the motion dynamics, which is different from existing gradient-based attack methods that compute the loss gradient assuming each dimension in the data is independent. The SMI gradient can augment many gradient-based attack methods, leading to a new family of no-box attack methods. Extensive evaluation and comparison show that our method imposes a real threat to existing classifiers. They also show that the SMI gradient improves the transferability and imperceptibility of adversarial samples in both no-box and transfer-based black-box settings. + +# 1. Introduction + +Deep learning models are vulnerable to adversarial attacks, which compute data perturbations strategically to fool trained networks. Since its discovery [31], a wide variety of models in different tasks have been attacked [1], raising severe concerns as these perturbations are imperceptible to humans. Recently, the adversarial attack in skeleton-based human activity recognition (S-HAR) has attracted at- + +
Information +AccessibleWhite- +BoxQueried +Black-BoxTransferred +Black-BoxNo- +BoxHard +No-Box
Model Parameters✓××××
Queries of Victims✓✓×××
Training Samples××✓××
Labels✓✓✓✓×
+ +Table 1. Comparisons on different settings of adversarial attacks. $\checkmark$ and $\times$ indicate if a method needs to access the corresponding information. + +tention as skeletal data have been widely used in security-critical applications such as sports analysis, bio-mechanics, surveillance, and human-computer interactions [24]. + +Existing attacks in S-HAR are categorized into white-box and black-box approaches. White-box approaches require prior knowledge of the full details of a victim model [18, 36] while black-box approaches require a large number of queries to the victim model [7] or the access to training data and labels [36]. On the one hand, the victim model details and the training data and labels are unlikely to be available to the attacker in real-world scenarios. On the other hand, making frequent and numerous queries (e.g. tens of thousands) to the victim model is time-consuming and raises suspicion. In other words, the settings of existing S-HAR attacks are overly restrictive. A key to a successful attack is to reduce the required information of the victim model, training data and labels. + +In this paper, we introduce a new threat model that requires no access to the victim model, training data or labels. We name the new threat model the hard no-box attack, differentiating from the recent no-box attack on images [16] that does not require access to the victim model but still needs access to the labels (i.e. soft no-box attack). Table 1 demonstrates the comparison on different settings of adversarial attacks. Among all attack settings, our hard no-box attack requires the least amount of knowledge, as it can only access the testing data without labels. Designing such an attack is nontrivial and challenging. Without access to the victim model, the attack method cannot rely on the gradient of a classification loss [12], data manipulation during + +training [25], and the feedback of a classifier [2]. The challenge is further exacerbated by the requirement of no label and training sample access, where no surrogate model can be trained to attack or estimate the data distribution. + +To tackle the challenges, we propose a contrastive learning (CL) [34] solution with a manifold-based no-box adversarial loss. First, we introduce a new application of CL to learn a latent data manifold where similar samples are naturally aggregated while dissimilar samples are dispersed without the need of class labels. It provides a good description of sample similarity that facilitates generating skeletal adversarial samples. CL is suitable for hard no-box attack settings due to its ability to capture the discriminative high-level features under our restricted attack conditions. Second, we compute the perturbation to drag a data sample away from its similar neighbors in the latent space, bounded by a pre-defined budget. In particular, we design a new no-box adversarial loss to maximize each adversary's dissimilarity with positive samples while minimizing its similarity with negative samples. The loss serves as guidance for the adversary search in our gradient-based attack scheme. + +While gradient-based attack methods like I-FGSM [14] are shown to be effective on S-HAR attacks [36, 18], the gradient is computed based on the victim model and the labels, making it unsuitable for hard no-box attacks. Since adversarial samples are likely to lie in or near the motion manifold [7], ideally, we want to explore along the manifold. That is, the computation of adversarial loss gradient should consider the local motion manifold. + +To this end, we propose to explicitly model motion dynamics for describing the local manifold around a given motion. Specifically, we introduce the skeleton-motion-informed (SMI) gradient that employs dynamics models (e.g. Markovian and autoregressive) to represent motion dynamics for the loss gradient computation. As a result, while existing methods generally assume each dimension in a data sample to be independent when computing the loss gradient, SMI gradient explicitly considers the dependency between frames in time. Furthermore, the SMI gradient is compatible with existing gradient-based methods including I-FGSM and MI-FGSM [8], allowing us to effectively construct a new family of no-box attack methods. + +Extensive experiments show that our method generates effective adversarial samples that successfully attack various victim models across datasets (HDM05, NTU60 and NTU120). Our SMI-gradient based attacks improve the attack transferability in both no-box and transferred black-box settings, with better imperceptibility. Codes are available in https://github.com/luyg45/HardNoBoxAttack and our contributions are: + +- We confirm the S-HAR threat by introducing a new hard no-box attack and proposing the first method to generate adversarial samples without access to the victim model + +or training data or labels, to the best of our knowledge. + +- We propose a new skeleton-motion-informed gradient that guides the adversary search along the motion manifold, explicitly considering the spatial-temporal nature of skeletal motions. +- We present a family of novel gradient-based attack strategies facilitated by the new gradient, improving the transferability and imperceptibility of adversarial samples in no-box and transferred black-box attacks. + +# 2. Related Works + +Skeleton-Based Human Action Recognition S-HAR has attracted considerable attention in many applications [24] where deep learning-based approaches have achieved state-of-the-art performance [30, 13]. Recurrent neural networks are employed to model the temporal domain of human motions [9, 29]. Furthermore, unlike images and videos, the skeleton has a graph structure, so graph convolutional networks have shown to be effective in modelling the spatial or spatial-temporal features [28, 41]. The effectiveness is generally achieved by considering the skeleton as a topological graph where the joints and bones correspond to nodes and edges [21]. Improved graph designs and network architectures are subsequently proposed [20, 44, 45, 43]. + +Adversarial Attacks on Skeletons Adversarial attacks were initially introduced in [31], which showcases the vulnerability of deep neural networks and has been extended to other data types. Generally, adversarial attack is a special technique of data augmentation that aims to reveal the vulnerability of a system by finding new samples, while other data augmentation techniques may have diverse purposes, e.g. training efficiency and inference performance [27]. Recently, the attack on S-HAR has received increasing attention. Wang et al. [36] analyzed the perceptibility of adversarial skeletal samples and proposed a new perceptual loss. Liu et al. [18] focused on GCN-based models and utilized generative adversarial networks to synthesize adversarial examples. Tanaka et al. [32] proposed a new lower-dimensional attack, in which only the length of bones could be perturbed. These methods all require the complete knowledge of victim models, a setting known as the white-box attack. In contrast, Diao et.al [7] introduced the first black-box S-HAR attack method, which searches motion manifolds for adversaries. Still, black-box attacks need to frequently query the victim models, which can be infeasible in real-world systems. In contrast, we consider a more practical threat setting named the hard no-box attack, where an attacker only has access to unlabeled skeletal data. + +Gradient-Based Attack Strategies The core component of adversarial attacks is to generate adversarial samples [1]. Gradient-based attack methods have been widely used to introduce perturbations to a sample following the direction of the loss gradient. Goodfellow et al. [12] proposed + +![](images/dd5f819e3929239968565e14cec3ac68b9cdb2f100e9c1864f1213d38f586cd0.jpg) +Figure 1. The training (left) and attack (right) processes for the hard no-box attack. The trained query encoder in the training process is used for attacks in the attack process. + +the fast gradient sign method (FGSM) that perturbs a sample by a single step along the loss gradient. Kurakin et al. [14] proposed I-FGSM by extending FGSM to an iterative process. Dong et al. [8] presented MI-FGSM by adding momentum to the gradient, which boosted the transferability of adversarial samples. Xie et al. [40] applied diversified augmentations to the inputs before each iteration to craft more transferable samples. While these gradient-based strategies are successful in static data and have been adapted to skeletal motions, they neglect the dependency between frames for gradient computation, which is crucial in time series. This motivates us to propose our skeleton-motion-informed attack strategies, which explicitly model the motion dynamics in the temporal domain [39, 35]. + +# 3. Hard No-Box Attack for Skeletal Data + +Figure 1 shows the overview of our method. The left part is the training process where we adopt contrastive learning to obtain a latent data manifold to distinguish data samples. The attack process is shown on the right-hand side. We first design a new no-box adversarial loss in the trained latent space to guide the adversary search using samples that are dissimilar to the attacked sample. Then we propose a novel skeleton-motion-informed gradient and a new family of attack methods for generating adversarial samples. + +# 3.1. Contrastive Learning for Motion Manifold + +While the fundamental idea of adversarial attacks is to perturb a data sample to cross class boundaries, such boundaries cannot be estimated for hard no-box attacks due to the lack of labels. To estimate such boundaries without labels, we present a new application of contrastive learning (CL) [34] to aggregate similar data samples as soft class boundaries in latent space. Such boundaries enable us to adversarially perturb a sample to cross boundaries. We also train + +an encoder to extract discriminative high-level features for the motion manifold in latent space. Overall, our CL constructs boundaries in latent space without labels via aggregating similar samples and segregating dissimilar samples. Our attack is guided by the dissimilarity of high-level features between samples for generating adversarial samples, instead of using class boundaries to lead the attack [2]. + +To incorporate both spatial and temporal information, we train an encoder (Fig. 1 left) based on adaptive graph convolutional network (AGCN) [28]. To force encoders to focus on high-level features, we apply skeleton-specific data augmentations to an input sequence $S$ and obtain two different views $S_{q}$ and $S_{k}$ . Augmentations include spatial operations (e.g. pose transformations, joint jittering, etc.) and temporal operations (e.g. temporal crop and resize) [34] (detailed in supplementary material). Then, we feed $S_{q}$ and $S_{k}$ into the query encoder $f_{q}$ and the key encoder $f_{k}$ respectively for the info-noise-contrastive estimation (InfoNCE) [23]: + +$$ +\begin{array}{l} L _ {\text {c o n t r a s t}} = \tag {1} \\ - \log \frac {\exp \left(f _ {q} \left(S _ {q}\right) \cdot f _ {k} \left(S _ {k}\right) / \tau\right)}{\exp \left(f _ {q} \left(S _ {q}\right) \cdot f _ {k} \left(S _ {k}\right) / \tau\right) + \sum_ {F _ {n} \sim N} \exp \left(f _ {q} \left(S _ {q}\right) \cdot F _ {n} / \tau\right)}, \\ \end{array} +$$ + +where $\tau$ is the temperature parameter, $N$ is the dynamic queue storing the features of negative samples $F_{n}$ obtained in the training process. After training, we use the query encoder $f_{q}$ , which encodes the motion manifold, for attack. + +# 3.2. Adversarial Loss for Unlabeled Skeletal Data + +The adversarial loss of hard no-box attack is significantly different from most existing methods that heavily depend on labels and class boundaries. Since class labels and class boundaries are unavailable in hard no-box attacks, we utilize data samples that are dissimilar to a given sample (i.e. negative samples) for defining the adversarial loss. Correspondingly, samples that are similar to the given sample are considered as positive samples. We argue when a given sample is perturbed towards its negative samples and away from its positive samples, it tends to become an adversary. This is because the negative samples generally indicate the high-density areas of other classes in the latent space. + +The hard no-box adversarial loss is designed as: + +$$ +L _ {a d v} = - \log \frac {\exp [ S i m (f _ {q} (s) , f _ {q} (\tilde {s})) ]}{\sum_ {j} \exp [ S i m (f _ {q} (s) , f _ {q} (\tilde {s} _ {j})) ]}, \tag {2} +$$ + +where $Sim$ is the cosine similarity, $s$ is the adversarial sample to be computed, $\widetilde{s}$ is the clean sample regarded as the positive sample, and $\widetilde{s}_j$ are the negative samples. Maximizing $L_{adv}$ moves $s$ away from $\widetilde{s}$ and towards $\widetilde{s}_j$ in the latent manifold. With $L_{adv}$ , gradient-based attacks are employed. + +To maximize Eq. 2, the selection of negative samples $\widetilde{s}_j$ is crucial and we design a method tailored for no-box attacks. Existing work [10] utilizes cluster-fit [42] to generate + +pseudo labels for selecting negative samples during adversarial training, which is less suitable for the no-box attack as it requires another pretrained off-line encoder to obtain pseudo labels. Instead, we adapt K-means, an unsupervised method, to select negatives, removing the need of any pretraining. We discard $Q$ clusters whose cluster centers are the closest to the input sample, mitigating the risk of misleading attacks. The remaining cluster centers are considered as the negative samples $\tilde{s}_j$ . + +# 4. Skeleton-Motion-Informed Gradient + +Existing gradient-based attack methods treat each dimension of the data as an independent variable, i.e. raw gradient. Attacks based on raw gradients tend to drag a sample away from the data manifold [11]. With the guidance of class boundaries and a limit on the perturbation budget, the raw gradient can still find deceiving adversaries. However, this setting is infeasible in hard no-box attacks. Without class boundaries, raw gradients that point to the negative samples can drag the adversary far away from the manifold. This is because while the perturbations are towards negative samples, they are not necessarily in a direction orthogonal to the class boundary. Consequently, larger perturbations are needed to cross the boundary, leading to adversaries being far off the manifold. This creates the need to constrain the perturbation within or near the manifold, at least locally. Since the motion manifold is constrained by the motion dynamics [38], we argue that the gradient needs to explicitly capture the dynamics. Therefore, we propose a new gradient named skeleton-motion-informed (SMI) gradient, capturing the manifold information that has been largely ignored by existing methods in loss gradient computation. + +# 4.1. Dynamics in the Gradient Structure + +Given a skeletal sequence $S = [S_{1}, S_{2}, \dots, S_{t}]$ and the adversarial loss $J(S)$ , a straightforward but effective strategy to craft adversarial perturbations is the gradient-based attack [1]. It utilizes backpropagation of the loss function $\nabla J(S)$ to iteratively change input samples $S$ : + +$$ +\hat {S} = S + \alpha \cdot \operatorname {s i g n} (\nabla J (S)), \tag {3} +$$ + +where $\alpha$ is the attack step size and $\hat{S}$ is the adversarial samples. In skeletal motions, this attack gradient $\nabla J(S)$ consists of a set of partial derivatives over all frames: + +$$ +\nabla J (S) = \left[ \frac {\partial J (S)}{\partial S _ {1}}, \frac {\partial J (S)}{\partial S _ {2}}, \dots , \frac {\partial J (S)}{\partial S _ {t}} \right]. \tag {4} +$$ + +The partial derivative $\frac{\partial J(S)}{\partial S_t}$ assumes each frame is independent, and this is the raw gradient employed in existing methods [36]. However, human motions contain rich dynamics so that the system can be described as $S_{t} = f(S_{< t})$ . So + +far, various dynamics models have been attempted to model human motions, such as Markovian models [33, 37], autoregressive models [39], and many-to-many mapping [38], all of which can capture the dynamics at different scales in time. We explore these models to reveal the missing dynamics in the structure of the raw gradient and propose our SMI-gradients that consider motion dynamics. + +Markovian Model We assume the motion dynamics can be captured by a Markovian model, i.e. $S_{t} = f_{d1}(S_{t - 1})$ . This allows us to derive the 1st-order SMI-gradient: + +$$ +\left(\frac {\partial J (S)}{\partial S _ {t - 1}}\right) _ {d 1} = \frac {\partial J (S)}{\partial S _ {t - 1}} + \frac {\partial J (S)}{\partial S _ {t}} \cdot \frac {d S _ {t}}{d S _ {t - 1}}, \tag {5} +$$ + +where $\frac{dS_t}{dS_{t-1}}$ is the temporal relationship between two consecutive frames that will be instantiated. Eq. 5 shows that the attack gradient along the motion manifold needs to consider the first-order information in the motion, e.g. velocity. + +Autoregressive Model Besides the first-order dynamics, we also model the second-order dynamics by assuming $S_{t} = f_{d2}(S_{t - 1},S_{t - 2})$ , as 2nd-order dynamics (i.e. joint acceleration) capture the smooth temporal dynamics of skeletal motion [36]. We extend Eq. 5 as: + +$$ +\begin{array}{l} \left(\frac {\partial J (S)}{\partial S _ {t - 2}}\right) _ {d 2} = \frac {\partial J (S)}{\partial S _ {t - 2}} + \frac {\partial J (S)}{\partial S _ {t - 1}} \cdot \frac {\partial S _ {t - 1}}{\partial S _ {t - 2}} \tag {6} \\ + \frac {\partial J (S)}{\partial S _ {t}} \cdot \left(\frac {\partial S _ {t}}{\partial S _ {t - 2}} + \frac {\partial S _ {t}}{\partial S _ {t - 1}} \cdot \frac {\partial S _ {t - 1}}{\partial S _ {t - 2}}\right). \\ \end{array} +$$ + +While higher-order models can also be considered, there is an empirical evidence that the first three orders are the most important in skeletal motion adversarial attack [36]. Therefore, we express the SMI gradients of the whole skeletal sequence as $(\nabla J(S))_{d1} = \left[(\frac{\partial J(S)}{\partial S_1})_{d1},(\frac{\partial J(S)}{\partial S_2})_{d1},\dots ,(\frac{\partial J(S)}{\partial S_t})_{d1}\right]$ and $(\nabla J(S))_{d2} = \left[(\frac{\partial J(S)}{\partial S_1})_{d2},(\frac{\partial J(S)}{\partial S_2})_{d2},\dots ,(\frac{\partial J(S)}{\partial S_t})_{d2}\right]$ . + +# 4.2. Time-Varying Autoregressive Models for Dynamics + +We employ explicit models [39] to compute the dynamics-related derivatives in SMI gradients. While implicit models [33] may also be considered, they would require another network to be trained, making them less preferable in hard no-box attacks. To realize $f_{d1}$ and $f_{d2}$ , we use time-varying autoregressive models (TV-AR) [3], which effectively estimates the dynamics of skeleton sequence [39] due to its ability of modelling the temporal non-stationary signals: + +$$ +f _ {d 1}: S _ {t} = A _ {t} \cdot S _ {t - 1} + B _ {t} + \gamma_ {t}, \tag {7} +$$ + +$$ +f _ {d 2}: S _ {t} = C _ {t} \cdot S _ {t - 1} + D _ {t} \cdot S _ {t - 2} + E _ {t} + \gamma_ {t}, \tag {8} +$$ + +where Eq. 7 and Eq. 8 are denoted as TV-AR(1) and TV-AR(2) respectively. The model parameters $\beta_{t}^{1} = [A_{t},B_{t}]$ and $\beta_{t}^{2} = [C_{t},D_{t},E_{t}]$ are all time-varying parameters and determined by data-fitting. $\gamma_{t}$ is a time-dependent white noise representing the dynamics of stochasticity. + +Using Eq. 7 to compute $\frac{\partial S_t}{\partial S_{t - 1}}$ , Eq. 5 becomes: + +$$ +\left(\frac {\partial J (S)}{\partial S _ {t - 1}}\right) _ {d 1} = \frac {\partial J (S)}{\partial S _ {t - 1}} + \frac {\partial J (S)}{\partial S _ {t}} \cdot A _ {t}. \tag {9} +$$ + +Similarly, using Eq. 8, we can compute $C_t = \frac{\partial S_t}{\partial S_{t-1}}$ and $D_t = \frac{\partial S_t}{\partial S_{t-2}}$ . For $\frac{\partial S_{t-1}}{\partial S_{t-2}}$ , we use $S_{t-1} = C_{t-1} \cdot S_{t-2} + D_{t-1} \cdot S_{t-3} + E_{t-1} + \gamma_t$ to compute it: $C_{t-1} = \frac{\partial S_{t-1}}{\partial S_{t-2}}$ . Then, Eq. 6 becomes: + +$$ +\left(\frac {\partial J (S)}{\partial S _ {t - 2}}\right) _ {d 2} = \frac {\partial J (S)}{\partial S _ {t - 2}} + \frac {\partial J (S)}{\partial S _ {t - 1}} \cdot C _ {t - 1} + \frac {\partial J (S)}{\partial S _ {t}} \cdot \left(D _ {t} + C _ {t} \cdot C _ {t - 1}\right). \tag {10} +$$ + +# 5. Skeleton-Motion-Informed Attack + +We construct new gradient-based attack methods based on our novel SMI gradient. Due to its compatibility, our proposed gradient can be integrated with most existing gradient-based methods. We select I-FGSM and MI-FGSM, which have been proven for their efficiency on S-HAR attacks [36, 18]. We augment them with first and second-order SMI-gradients, leading to four new attack methods. + +Fast Gradient Sign Methods (FGSM) FGSM [12] is a single-step attack method that generates the adversarial samples $\hat{S} = S + r$ by maximizing the adversarial loss function $J(S)$ , where $r$ denotes an adversarial perturbation that is constrained within a budget $\|r\|_p < \epsilon$ , where $\|\cdot\|_p$ denotes the $l_p$ -norm. One variant of FGSM is the Iterative Fast Gradient Sign Method (I-FGSM) [14], which extends FGSM to an iterative process: + +$$ +\hat {S} ^ {i + 1} = \hat {S} ^ {i} + \alpha \cdot \operatorname {s i g n} (\nabla_ {s} J (S)), \tag {11} +$$ + +where $\alpha$ is the attack step size and $i$ means iteration. Another variant is MI-FGSM [8], which considers the momentum of the attack to avoid local maxima: + +$$ +g ^ {i + 1} = \mu \cdot g ^ {i} + \frac {\nabla_ {s} J (S)}{\| \nabla_ {s} J (S) \| _ {1}} \tag {12} +$$ + +$$ +\hat {S} ^ {i + 1} = \hat {S} ^ {i} + \alpha \cdot \operatorname {s i g n} \left(g ^ {i + 1}\right), \tag {13} +$$ + +where $\mu$ is the momentum decay factor and $g_{i}$ is the gradient in iteration $i$ . + +SMI-gradient Based Attacks We replace the original gradient $\nabla J(S)$ in I-FGSM and MI-FGSM with our SMI gradient $(\nabla J(S))_{d1}$ or $(\nabla J(S))_{d2}$ . This creates four new + +# Algorithm 1 S1I-FGSM and S2I-FGSM + +Input: An encoder $k$ with a loss function $J$ ; a skeletal sequence sample $S$ ; the size of attack step $\alpha$ ; iterations $I$ ; the budget of perturbation $\epsilon$ . +Output: An adversarial example $\hat{S}$ with $\| \hat{S} - S \|_p < \epsilon$ . + +1: Initialization: $\hat{S}^0 = S$ + +2: Fitting $S$ with TV-AR model to obtain the time-varying parameters $\beta_{t}$ + +3: for $i = 0$ to $I - 1$ do + +4: Input $S^i$ to $k$ ; + +5: Obtain the raw gradient $\nabla J(\hat{S}^i)$ on $J$ ; + +6: Calculate the SMI gradient $(\nabla J(\hat{S}^i))_{d1}$ with Eq. 9, or $(\nabla J(\hat{S}^i))_{d2}$ with Eq. 10, using $\beta_{t}$ and $\nabla J(\hat{S}^i)$ ; + +7: Update $\tilde{S}^{i + 1}$ by applying the sign gradient as: + +$$ +\hat {S} ^ {i + 1} = \hat {S} ^ {i} + \alpha \cdot \operatorname {s i g n} (\nabla J (\hat {S} ^ {i})) _ {d 1}), o r \tag {14} +$$ + +$$ +\hat {S} ^ {i + 1} = \hat {S} ^ {i} + \alpha \cdot \mathrm {s i g n} \left(\nabla J (\hat {S} ^ {i})\right) _ {d 2}). +$$ + +8: end for + +9: return $\hat{S} = \hat{S}^I$ + +dynamic attack strategies: first-order SMI I-FGSM $(\mathrm{S}_1\mathrm{I}$ FGSM), second-order SMI I-FGSM $(\mathrm{S}_2\mathrm{I}$ -FGSM), firstorder SMI MI-FGSM $(\mathrm{S}_1\mathrm{MI}$ -FGSM), and second-order SMI MI-FGSM $(\mathrm{S}_2\mathrm{MI}$ -FGSM). The processes of SI-FGSM are shown in Algorithm 1. The algorithm of SMI-FGSM can be found in the supplementary material. + +# 6. Experiments + +We refer the readers to the supplementary material for extra experimental results. + +Datasets We select three widely used skeletal datasets: HDM05 [22] (2,337 sequences of 130 classes performed by 5 actors), NTU60 [26] (56,880 sequences of 60 classes), and NTU120 [19] (114,480 sequences of 120 classes, an extended version of NTU60, one of the largest datasets in the field). We pre-process HDM05 following [9] and both NTU datasets following [28]. The different skeletons are mapped to a standard 25-joint structure as in [38]. For our hard no-box attack, we only use the testing data and do not use the training data, the training labels and the testing labels during attacks. + +Target Models We choose multiple state-of-the-art models as victims: ST-GCN [41], 2s-AGCN [28], AS-GCN [15], SGN [44] and MS-G3D [20]. They are trained using the official implementations and following the training protocols. For 2s-AGCN, we attack both the single joint stream model (js-AGCN) and the two-stream model. + +Implementation Details We pre-train the CL encoder $f_{q}$ following [34]. The unsupervised network is trained with a temperature value $\tau = 0.07$ and SGD optimizer for 450 epochs. The learning rate is set to 0.01 with a weight decay of 0.0001. Due to the limitation of hard no-box settings, the attacker cannot access the training samples during the whole process. Therefore, the encoder is trained on the test- + +
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1 I-FGSMNo-box S2 I-FGSMNo-box MI-FGSMNo-box S1 MI-FGSMNo-box S2 MI-FGSM
0.01js-AGCN11.21%—26.05%28.09%30.87%30.68%34.75%36.58%
2s-AGCN5.36%—13.94%15.04%16.45%16.47%18.23%19.31%
ST-GCN3.57%12.93%9.55%9.86%9.96%11.11%11.36%11.56%
MS-G3D8.39%35.23%10.57%11.14%11.69%11.76%12.85%14.11%
SGN21.25%26.03%34.09%34.46%35.23%38.49%38.75%38.81%
ASGCN5.69%20.87%13.92%14.85%14.67%15.95%16.82%17.75%
0.008js-AGCN10.12%—22.84%24.36%26.46%25.70%29.64%30.88%
2s-AGCN5.04%—11.36%12.19%13.03%12.30%13.85%15.56%
ST-GCN3.26%10.33%7.87%7.99%8.04%8.84%9.07%9.19%
MS-G3D5.19%31.28%9.18%9.51%9.99%10.01%10.26%12.98%
SGN20.95%23.30%29.61%30.20%30.74%33.32%33.63%34.54%
ASGCN5.54%18.77%11.42%12.05%12.29%12.77%13.69%14.37%
0.006js-AGCN7.19%—19.70%20.32%21.23%20.04%22.76%24.33%
2s-AGCN3.70%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN2.46%7.88%5.65%5.85%5.97%6.25%6.54%6.66%
MS-G3D4.46%23.15%7.30%7.84%7.61%7.94%8.05%8.39%
SGN16.76%19.93%23.92%24.57%25.75%26.74%27.04%27.64%
ASGCN3.94%14.30%8.79%9.23%9.07%9.71%10.29%10.90%
+ +Table 2. The fooling rate of different methods on the target models in NTU60, where $\epsilon$ is the perturbation budget. + +
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN9.97%—23.92%24.64%25.79%27.26%27.93%29.07%
2s-AGCN6.38%—20.09%21.11%22.06%23.51%24.63%24.96%
ST-GCN12.18%23.84%23.53%24.73%25.77%26.95%27.31%28.76%
MS-G3D10.63%24.63%19.07%20.11%20.20%21.57%21.95%22.73%
SGN19.65%37.90%31.43%32.64%33.75%38.47%37.96%38.85%
ASGCN7.29%24.15%18.03%19.29%20.37%19.88%20.22%21.60%
0.008js-AGCN9.04%—21.58%21.77%22.47%24.28%24.62%25.74%
2s-AGCN5.79%—15.71%15.23%15.94%18.68%18.76%19.57%
ST-GCN11.23%21.35%20.74%21.51%22.22%23.52%24.65%24.87%
MS-G3D9.98%21.73%17.06%17.74%17.92%19.53%19.29%20.33%
SGN17.59%35.88%27.38%28.17%28.65%32.43%33.06%32.55%
ASGCN6.87%21.87%15.73%16.93%17.62%17.41%17.75%18.70%
0.006js-AGCN8.23%—18.51%18.88%18.79%20.74%21.88%22.06%
2s-AGCN5.09%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN9.54%17.51%17.60%17.97%18.67%18.90%19.91%20.34%
MS-G3D9.42%18.07%15.06%15.09%15.20%16.86%17.25%17.56%
SGN16.82%34.10%22.39%22.65%22.76%25.43%25.79%25.92%
ASGCN5.39%19.02%12.90%13.45%14.97%14.08%15.26%16.33%
+ +Table 3. The fooling rate of different methods on the target models in NTU120, where $\epsilon$ is the perturbation budget. + +ing set. We adopt the $l_{\infty}$ norm for the perturbation budget $\epsilon$ . For clusters of negative samples, the number of clusters is 120, and the number of deleted centers $Q$ is 10. + +Evaluation Metrics We employ the fooling rate as a major metric. It is defined as the percentage of data samples whose predicted labels changed after adversarial attacks [18]. Besides, inspired by [36], we define a perceptual deviation indicator to evaluate the imperceptibility of adversarial skeletal samples: + +$$ +\begin{array}{l} \Delta p = \frac {1}{M T} \sum_ {n = 0} ^ {M} \left\| S - \hat {S} \right\| _ {2} + \frac {1}{M T} \sum_ {n = 0} ^ {M} \left\| B - \hat {B} \right\| _ {2} \tag {15} \\ + \frac {1}{M T L} \sum_ {n = 0} ^ {M} \left\| \ddot {S} - \ddot {\hat {S}} \right\| _ {2} \\ \end{array} +$$ + +where $M$ is the number of adversarial samples, $T$ is the total number of frames, and $L$ is the number of joints in the skeleton. The three terms evaluate the deviations of joint position, bone-length, and acceleration, respectively. A smaller perceptual deviation indicates better imperceptibility. + +# 6.1. Hard No-Box Attack + +Baselines We establish two baselines using transfer-based attacks based on a self-supervised and a supervised classifier. This is because our method is the first hard no-box attack and there is no other similar method. The transfer-based attack is the closest setting to ours as they do not require access to the victim model. The first baseline (Self-sup Attacker) is a self-supervised surrogate model with a linear layer appended to our CL encoder. We freeze our CL encoder after the manifold training and then train the linear layer supervisedly [17]. In this way, the need to access labels is only for training the linear layer, not the CL encoder, which is not strictly hard no-box but closer than existing methods. The second baseline is a standard transfer-based attack where we use js-AGCN as the surrogate model and SMART [36] as the white-box attacker (AGCN Attacker). Unlike our method, both baselines still require access to the training data and the training labels during attacks. + +For hard no-box attack, different attack strategies are compared including I-FGSM, MI-FGSM, $\mathrm{S}_1\mathrm{I}$ -FGSM, $\mathrm{S}_2\mathrm{I}$ FGSM, $\mathrm{S}_1\mathrm{MI}$ -FGSM, and $\mathrm{S}_2\mathrm{MI}$ -FGSM. We run all the attackers for 400 iterations and report the attack perfor + +
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN10.12%—10.61%10.98%11.55%13.45%14.20%14.20%
2s-AGCN1.89%—5.30%5.68%5.68%6.44%6.44%6.82%
ST-GCN6.44%24.26%9.47%9.85%9.47%8.90%11.17%11.74%
MS-G3D4.17%86.95%41.67%44.51%43.94%53.03%56.06%53.41%
SGN1.89%3.98%2.46%2.84%3.03%3.03%3.40%3.59%
ASGCN1.89%27.57%3.22%4.36%3.22%4.36%5.68%5.68%
0.008js-AGCN9.46%—8.14%9.09%9.09%10.61%10.98%11.55%
2s-AGCN1.70%—4.36%4.92%4.73%4.55%5.30%5.87%
ST-GCN3.79%21.69%8.33%8.52%8.52%8.33%8.90%8.33%
MS-G3D3.60%82.17%25.76%30.68%29.17%36.36%39.39%39.02%
SGN1.51%3.60%0.57%1.51%2.08%2.27%2.27%2.46%
ASGCN1.70%22.43%2.46%2.84%2.84%2.27%2.84%3.03%
0.006js-AGCN7.19%—5.11%7.01%6.44%6.06%6.63%7.77%
2s-AGCN1.33%—2.08%2.08%2.27%3.79%3.98%4.17%
ST-GCN3.40%19.85%6.44%7.01%6.63%7.58%8.33%7.77%
MS-G3D2.84%72.43%11.55%12.31%11.92%15.34%18.94%17.99%
SGN0.57%3.40%0.13%0.94%0.57%1.13%1.32%1.51%
ASGCN1.52%17.82%0.38%0.76%1.89%0.57%2.27%2.27%
+ +![](images/20813cfe522b2390c19926f60308b2ec58116b33b20cef0ed9b8f4e5b2a44e1e.jpg) +Figure 2. Visual comparisons between attack strategies in no-box attacks $(\epsilon = 0.006)$ with key visual differences highlighted. + +mance under different perturbation budgets $\epsilon$ in Table 2 for NTU60, Table 3 for NTU120, and Table 4 for HDM05. We omit the results on js-AGCN and 2s-AGCN under AGCN Attacker because AGCN is the surrogate model. + +Table 2-4 show that the hard no-box attack poses real threats to a range of S-HAR classifiers. In general, the fooling rate of hard no-box attacks is higher than Self-sup Attackers. This is surprising as the Attacker utilizes training data and training labels while our method does not. Looking further, AGCN Attacker is much better than Self-sup Attacker and the major difference is their feature extraction (i.e. one is a CL encoder and one is a graph network). This shows transfer-based attack heavily relies on the feature extraction ability of the surrogate model that cannot bypass access to the training data and the training labels. Furthermore, when compared with AGCN Attacker, our method achieves similar fooling rates, varying across different victims and datasets. Given that AGCN Attacker requires access to training data, training labels and testing labels, we argue hard no-box attacker achieves superior results and provides a more realistic setting. + +In addition, among various no-box attack strategies, $\mathrm{S}_2\mathrm{MI}$ FGSM performs the best and often by big margins. All the SMI gradient-based methods generate stronger adversaries compared with baselines I-FGSM and MI-FGSM. The 2nd-order SMI attack method usually outperforms the corresponding 1st-order version. Last, we notice a variance in fooling rate across different victims and datasets. For in + +stance, the multi-stream model (2s-AGCN) significantly enhances the robustness compared to the single-stream model (js-AGCN); the fooling rate of all methods drops by nearly half. This may be because the multi-stream model can ensemble features from different modalities, which improves the robustness. In general, it is still an open question why fooling rate can vary across victims and datasets and we leave the theoretical analysis for future work. + +Table 4. The fooling rate of different methods on the target models in HDM05, where $\epsilon$ is the perturbation budget. + +
Surrogate ModelVictimsIFGSMMI-FGSMS2I-FGSM (Ours)S2MI-FGSM (Ours)
2s-AGCNSTGCN2.10%2.10%3.00%3.01%
MS-G3D2.58%2.59%2.90%2.97%
ST-GCN2sAGCN2.20%2.34%2.44%2.64%
MS-G3D2.00%2.10%2.63%2.92%
MS-G3DSTGCN1.71%1.69%2.65%2.67%
2sAGCN1.76%1.79%2.03%2.07%
+ +Table 5. The fooling rate of different attack strategies in transferred SMART attacks, where attack budgets $\epsilon = 0.01$ . + +# 6.2. Transfer-Based Black-Box Attack + +SMI gradient not only improves the transferability in hard no-box attacks, but also enhances other gradient-based skeletal attacks. Here, we employ SMART [36], a white-box attacker, as a baseline to compare different attack strategies. In the original SMART settings, I-FGSM is adopted to generate adversarial samples. We replace it with $\mathrm{S}_2\mathrm{I}$ FGSM, MI-FGSM, and $\mathrm{S}_2\mathrm{MI}$ -FGSM to make a comparison. As all strategies achieve similar fooling rates in white-box attacks, we mainly focus on their transferability using + +different surrogate models. We utilize SMART to attack 2s-AGCN, ST-GCN, and MS-G3D on NTU60 and transfer the obtained samples to other victim networks. + +The performance of the transferred black-box attack is shown in Table 5. $\mathrm{S}_2\mathrm{MI}$ FGSM gives the best performance in transfer-based black-box attacks. $\mathrm{S}_2\mathrm{I}$ FGSM also improves the transferability of adversarial samples compared with baselines. In contrast, MI-FGSM, which succeeds in the image transfer-based attack, struggles in skeletal data. Its performance declines to $1.69\%$ when it attacks STGCN via MS-G3D. Overall, the success rate is low in Table 5 because SMART is sensitive to the chosen surrogate [36]. Transfer-based attacks are proven to suffer from lower fooling rates in S-HAR attack [36] and improving the transferability is still an open problem. Table 5 aims to show our SMI gradient can improve it by incorporating motion dynamics into the attack gradient, compared with other alternative gradients. We will include in future work how to further explore this dynamics for better attack transfer. + +# 6.3. Perceptual Analysis + +A key feature of SMI gradient-based attacks is the improvement in the perceptual quality of the adversarial samples due to the consideration of the motion manifold. To verify this, we employ quantitative comparison and qualitative visual analysis on the no-box adversarial samples under various strategies. We compare the perceptual quality $\Delta p$ on NTU60 in Table 6. We find that $\mathrm{S}_2\mathrm{I}$ FGSM achieves the best imperceptibility and obtains a massive improvement compared with I-FGSM. In contrast, MI-FGSM's deviation is twice that of $\mathrm{S}_2\mathrm{I}$ FGSM. Although $\mathrm{S}_2\mathrm{MI}$ FGSM does not achieve the best visual performance, it is still slightly better than I-FGSM and achieves a better trade-off between fooling rate and perceptual quality. This is understandable because our method considers dynamics that help to generate more on-manifold adversarial samples. Moreover, the 1st-order SMI attacks outperform baselines but cannot compete with 2nd-order SMI attacks. This demonstrates the importance of considering acceleration in the skeletal attack. + +We show the visual comparison of poses under various attack strategies in no-box attacks in Figure 2. The spinal joints demonstrate the most obvious differences. $\mathrm{S}_2\mathrm{I}$ FGSM outperforms the other attack methods and gets the most natural poses, whereas I-FGSM has slight but noticeable joint displacements in the neck and head. $\mathrm{S}_2\mathrm{MI}$ -FGSM performs better than its baseline, MI-FGSM, which shows zig-zag bending in the frame $t_0 + 2$ .The samples produced by MI-FGSM have the worst imperceptibility, where we can easily find the unnatural jittery movements. + +We also evaluate perceptual quality $\Delta p$ on SMART adversarial samples obtained with different gradients. Results conducted on the NTU60 dataset are shown in Table 7. $\mathrm{S}_2\mathrm{I}$ FGSM reaches the best perceptual performance compared + +with all attack strategies. MI-FGSM slightly declines the imperceptibility. $\mathrm{S}_2\mathrm{MI}$ -FGSM performs slightly worse in MS-G3D and ST-GCN. The reason is mainly that $\mathrm{S}_2\mathrm{MI}$ FGSM takes more iterations in the white-box attack, leading to late stopping and slightly worse visual performance. + +
Strategiesε = 0.01ε = 0.008ε = 0.006
I-FGSM95.8768.5643.13
MI-FGSM131.7790.4454.61
S1I-FGSM84.6360.7738.15
S1MI-FGSM114.5178.0647.04
S2I-FGSM65.6045.6727.67
S2MI-FGSM90.9362.0437.46
+ +Table 6. The perceptual deviation of different attack strategies in no-box attacks with different budgets $\epsilon$ . + +
VictimsI-FGSMMI-FGSMS2I-FGSM(Ours)S2MI-FGSM(Ours)
2s-AGCN1.521.621.251.49
MS-G3D2.392.461.693.02
ST-GCN1.131.181.101.47
+ +Table 7. The perceptual deviation of different attack strategies in SMART for different victim models. + +# 7. Conclusions and Discussions + +In this paper, we have verified potential threats to S-HAR solutions. A new setting is proposed: the hard no-box attack on skeletal motions without access to the victim model, the training samples or the labels. We validate our setting by proposing the first pipeline for hard no-box attacks. Moreover, as far as we know, we are the first to explore motion dynamics in the adversarial gradient computation, leading to a new SMI gradient compatible with existing gradient-based attacks. By extensive evaluation and comparison, our method has been proven to be threatening and imperceptible, relying on the least prior knowledge. + +The SMI gradient also improves the transferability of transferred black-box attacks. Nonetheless, boosting the transferability is still an open problem [36] and we will explore this further with our SMI gradient in the future. We will explore other models (e.g., diffusion models [5]) and other time-series data (e.g. stock price, videos) for the proposed attack. Also, our SMI gradient describes dynamics and may be beneficial for motion synthesis [4]. + +We call for attention to intensify the S-HAR robustness by considering defences against our hard no-box attack. We validate randomized smoothing [6] as a potential defence method in supplementary materials. Due to the least prior knowledge requirement, security risks posed by our attack can be reduced with such defenses. Otherwise, our attacks become a significantly threat to S-HAR. + +# Acknowledgments + +This research is supported in part by National Natural Science Foundation of China (ref: 61673314, Yang), EPSRC (ref: EP/X031012/1, NorthHFutures, Shum) and EU Horizon 2020 (ref: 899739, CrowdDNA, Wang). + +# References + +[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021. +[2] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In International Conference on Learning Representations, 2018. +[3] Laura F Bringmann, Ellen L Hamaker, Daniel E Vigo, André Aubert, Denny Borsboom, and Francis Tuerlinckx. Changing dynamics: Time-varying autoregressive models using generalized additive modeling. Psychological methods, 22(3):409, 2017. +[4] Ziyi Chang, Edmund JC Findlay, Haozheng Zhang, and Hubert PH Shum. Unifying human motion synthesis and style transfer with denoising diffusion probabilistic models. arXiv preprint arXiv:2212.08526, 2022. +[5] Ziyi Chang, George A Koulieris, and Hubert PH Shum. On the design fundamentals of diffusion models: A survey. arXiv preprint arXiv:2306.04542, 2023. +[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019. +[7] Yunfeng Diao, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and He Wang. Basar: Black-box attack on skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7597-7607, 2021. +[8] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. +[9] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015. +[10] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in Neural Information Processing Systems, 34:21480-21492, 2021. +[11] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. +[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. +[13] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. + +[14] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world, pages 99-112. Chapman and Hall/CRC, 2018. +[15] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. +[16] Qizhang Li, Yiwen Guo, and Hao Chen. Practical no-box adversarial attacks against dnns. Advances in Neural Information Processing Systems, 33:12849-12860, 2020. +[17] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2490-2498, 2019. +[18] Jian Liu, Naveed Akhtar, and Ajmal Mian. Adversarial attack on skeleton-based human action recognition. IEEE Transactions on Neural Networks and Learning Systems, 2020. +[19] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. +[20] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. +[21] Qianhui Men, Edmond S. L. Ho, Hubert P. H. Shum, and Howard Leung. A quadruple diffusion convolutional recurrent network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(9):3417-3432, 2021. +[22] Meinard Müller, Tido Röder, Michael Clausen, Bernhard Eberhardt, Björn Krüger, and Andreas Weber. Mocap database hdm05. Institut für Informatik II, Universität Bonn, 2(7), 2007. +[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. +[24] Bin Ren, Mengyuan Liu, Runwei Ding, and Hong Liu. A survey on 3d skeleton-based action recognition using learning method. arXiv preprint arXiv:2002.05907, 2020. +[25] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 11957-11965, 2020. +[26] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. +[27] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1214-1223, October 2021. + +[28] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. +[29] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017. +[30] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017. +[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. +[32] Nariki Tanaka, Hiroshi Kera, and Kazuhiko Kawamoto. Adversarial bone length attack on action recognition. arXiv preprint arXiv:2109.05830, 2021. +[33] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qilong Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Trans. Graph., 41(4), jul 2022. +[34] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. +[35] He Wang, Yunfeng Diao, Zichang Tan, and Guodong Guo. Defending black-box skeleton-based human activity classifiers. arXiv preprint arxiv.2203.04713, 2022. +[36] He Wang, Feixiang He, Zhexi Peng, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and David Hogg. Understanding the robustness of skeleton-based action recognition under adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14656-14665, 2021. +[37] He Wang, Edmond SL Ho, and Taku Komura. An energy-driven motion planning method for two distant postures. IEEE transactions on visualization and computer graphics, 21(1):18-30, 2015. +[38] He Wang, Edmond SL Ho, Hubert PH Shum, and Zhanxing Zhu. Spatio-temporal manifold learning for human motions via long-horizon modeling. IEEE transactions on visualization and computer graphics, 27(1):216-227, 2019. +[39] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015. +[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. +[41] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action + +recognition. In Thirty-second AAAI conference on artificial intelligence, 2018. +[42] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. Clusterfit: Improving generalization of visual representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6509-6518, 2020. +[43] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. arXiv preprint arXiv:2211.13466, 2022. +[44] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1112-1121, 2020. +[45] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023. \ No newline at end of file diff --git a/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/images.zip b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c694f5fdc419d7f8adc4281a43faef4ad419bb0e --- /dev/null +++ b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:301b2c692939b382039b1564cd1616bf9a678d1ef480253735d50f3c63976bcf +size 608525 diff --git a/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/layout.json b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ccc5f79129f907b375683c572bfd535792f8b9fc --- /dev/null +++ b/2023/Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient/layout.json @@ -0,0 +1,9422 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 102, + 544, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 102, + 544, + 138 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 544, + 138 + ], + "type": "text", + "content": "Hard No-Box Adversarial Attack on Skeleton-Based Human Action Recognition with Skeleton-Motion-Informed Gradient" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": "Zhengzhi Lu" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": " He Wang" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": " Ziyi Chang" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": " Guoan Yang" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": " Hubert P. H. Shum" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": "Durham University, UK " + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": "Xi'an Jiaotong University, China " + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 533, + 217 + ], + "type": "text", + "content": "University College London, UK \nlu947867114@stu.xjtu.edu.cn he_wang@ucl.ac.uk ziyi.chang@durham.ac.uk \ngayang@mail.xjtu.edu.cn hubert.shum@durham.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 270, + 290, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 290, + 569 + ], + "type": "text", + "content": "Recently, methods for skeleton-based human activity recognition have been shown to be vulnerable to adversarial attacks. However, these attack methods require either the full knowledge of the victim (i.e. white-box attacks), access to training data (i.e. transfer-based attacks) or frequent model queries (i.e. black-box attacks). All their requirements are highly restrictive, raising the question of how detrimental the vulnerability is. In this paper, we show that the vulnerability indeed exists. To this end, we consider a new attack task: the attacker has no access to the victim model or the training data or labels, where we coin the term hard no-box attack. Specifically, we first learn a motion manifold where we define an adversarial loss to compute a new gradient for the attack, named skeleton-motion-informed (SMI) gradient. Our gradient contains information of the motion dynamics, which is different from existing gradient-based attack methods that compute the loss gradient assuming each dimension in the data is independent. The SMI gradient can augment many gradient-based attack methods, leading to a new family of no-box attack methods. Extensive evaluation and comparison show that our method imposes a real threat to existing classifiers. They also show that the SMI gradient improves the transferability and imperceptibility of adversarial samples in both no-box and transfer-based black-box settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "type": "text", + "content": "Deep learning models are vulnerable to adversarial attacks, which compute data perturbations strategically to fool trained networks. Since its discovery [31], a wide variety of models in different tasks have been attacked [1], raising severe concerns as these perturbations are imperceptible to humans. Recently, the adversarial attack in skeleton-based human activity recognition (S-HAR) has attracted at-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 243, + 550, + 305 + ], + "blocks": [ + { + "bbox": [ + 306, + 243, + 550, + 305 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 550, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 550, + 305 + ], + "type": "table", + "html": "
Information\nAccessibleWhite-\nBoxQueried\nBlack-BoxTransferred\nBlack-BoxNo-\nBoxHard\nNo-Box
Model Parameters✓××××
Queries of Victims✓✓×××
Training Samples××✓××
Labels✓✓✓✓×
", + "image_path": "ffdee3e433b3216e6936752590d186e8457db25e83c119f9521f221ca3688388.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 305, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 306, + 305, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 306, + 305, + 545, + 336 + ], + "type": "text", + "content": "Table 1. Comparisons on different settings of adversarial attacks. " + }, + { + "bbox": [ + 306, + 305, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 306, + 305, + 545, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 305, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 306, + 305, + 545, + 336 + ], + "type": "text", + "content": " indicate if a method needs to access the corresponding information." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": "tention as skeletal data have been widely used in security-critical applications such as sports analysis, bio-mechanics, surveillance, and human-computer interactions [24]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 555 + ], + "type": "text", + "content": "Existing attacks in S-HAR are categorized into white-box and black-box approaches. White-box approaches require prior knowledge of the full details of a victim model [18, 36] while black-box approaches require a large number of queries to the victim model [7] or the access to training data and labels [36]. On the one hand, the victim model details and the training data and labels are unlikely to be available to the attacker in real-world scenarios. On the other hand, making frequent and numerous queries (e.g. tens of thousands) to the victim model is time-consuming and raises suspicion. In other words, the settings of existing S-HAR attacks are overly restrictive. A key to a successful attack is to reduce the required information of the victim model, training data and labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we introduce a new threat model that requires no access to the victim model, training data or labels. We name the new threat model the hard no-box attack, differentiating from the recent no-box attack on images [16] that does not require access to the victim model but still needs access to the labels (i.e. soft no-box attack). Table 1 demonstrates the comparison on different settings of adversarial attacks. Among all attack settings, our hard no-box attack requires the least amount of knowledge, as it can only access the testing data without labels. Designing such an attack is nontrivial and challenging. Without access to the victim model, the attack method cannot rely on the gradient of a classification loss [12], data manipulation during" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "type": "text", + "content": "† This work was conducted during the visit to the Durham University." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 143, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 61, + 703, + 143, + 712 + ], + "type": "text", + "content": " Corresponding Author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4597" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "training [25], and the feedback of a classifier [2]. The challenge is further exacerbated by the requirement of no label and training sample access, where no surrogate model can be trained to attack or estimate the data distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 323 + ], + "type": "text", + "content": "To tackle the challenges, we propose a contrastive learning (CL) [34] solution with a manifold-based no-box adversarial loss. First, we introduce a new application of CL to learn a latent data manifold where similar samples are naturally aggregated while dissimilar samples are dispersed without the need of class labels. It provides a good description of sample similarity that facilitates generating skeletal adversarial samples. CL is suitable for hard no-box attack settings due to its ability to capture the discriminative high-level features under our restricted attack conditions. Second, we compute the perturbation to drag a data sample away from its similar neighbors in the latent space, bounded by a pre-defined budget. In particular, we design a new no-box adversarial loss to maximize each adversary's dissimilarity with positive samples while minimizing its similarity with negative samples. The loss serves as guidance for the adversary search in our gradient-based attack scheme." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 420 + ], + "type": "text", + "content": "While gradient-based attack methods like I-FGSM [14] are shown to be effective on S-HAR attacks [36, 18], the gradient is computed based on the victim model and the labels, making it unsuitable for hard no-box attacks. Since adversarial samples are likely to lie in or near the motion manifold [7], ideally, we want to explore along the manifold. That is, the computation of adversarial loss gradient should consider the local motion manifold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 288, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 577 + ], + "type": "text", + "content": "To this end, we propose to explicitly model motion dynamics for describing the local manifold around a given motion. Specifically, we introduce the skeleton-motion-informed (SMI) gradient that employs dynamics models (e.g. Markovian and autoregressive) to represent motion dynamics for the loss gradient computation. As a result, while existing methods generally assume each dimension in a data sample to be independent when computing the loss gradient, SMI gradient explicitly considers the dependency between frames in time. Furthermore, the SMI gradient is compatible with existing gradient-based methods including I-FGSM and MI-FGSM [8], allowing us to effectively construct a new family of no-box attack methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 578, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 674 + ], + "type": "text", + "content": "Extensive experiments show that our method generates effective adversarial samples that successfully attack various victim models across datasets (HDM05, NTU60 and NTU120). Our SMI-gradient based attacks improve the attack transferability in both no-box and transferred black-box settings, with better imperceptibility. Codes are available in https://github.com/luyg45/HardNoBoxAttack and our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 50, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We confirm the S-HAR threat by introducing a new hard no-box attack and proposing the first method to generate adversarial samples without access to the victim model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 72, + 538, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 538, + 84 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 538, + 84 + ], + "type": "text", + "content": "or training data or labels, to the best of our knowledge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 85, + 545, + 179 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 309, + 85, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 85, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 309, + 85, + 545, + 131 + ], + "type": "text", + "content": "- We propose a new skeleton-motion-informed gradient that guides the adversary search along the motion manifold, explicitly considering the spatial-temporal nature of skeletal motions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 133, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 133, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 309, + 133, + 545, + 179 + ], + "type": "text", + "content": "- We present a family of novel gradient-based attack strategies facilitated by the new gradient, improving the transferability and imperceptibility of adversarial samples in no-box and transferred black-box attacks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 190, + 397, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 397, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 397, + 204 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 211, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 366 + ], + "type": "text", + "content": "Skeleton-Based Human Action Recognition S-HAR has attracted considerable attention in many applications [24] where deep learning-based approaches have achieved state-of-the-art performance [30, 13]. Recurrent neural networks are employed to model the temporal domain of human motions [9, 29]. Furthermore, unlike images and videos, the skeleton has a graph structure, so graph convolutional networks have shown to be effective in modelling the spatial or spatial-temporal features [28, 41]. The effectiveness is generally achieved by considering the skeleton as a topological graph where the joints and bones correspond to nodes and edges [21]. Improved graph designs and network architectures are subsequently proposed [20, 44, 45, 43]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 367, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 546, + 653 + ], + "type": "text", + "content": "Adversarial Attacks on Skeletons Adversarial attacks were initially introduced in [31], which showcases the vulnerability of deep neural networks and has been extended to other data types. Generally, adversarial attack is a special technique of data augmentation that aims to reveal the vulnerability of a system by finding new samples, while other data augmentation techniques may have diverse purposes, e.g. training efficiency and inference performance [27]. Recently, the attack on S-HAR has received increasing attention. Wang et al. [36] analyzed the perceptibility of adversarial skeletal samples and proposed a new perceptual loss. Liu et al. [18] focused on GCN-based models and utilized generative adversarial networks to synthesize adversarial examples. Tanaka et al. [32] proposed a new lower-dimensional attack, in which only the length of bones could be perturbed. These methods all require the complete knowledge of victim models, a setting known as the white-box attack. In contrast, Diao et.al [7] introduced the first black-box S-HAR attack method, which searches motion manifolds for adversaries. Still, black-box attacks need to frequently query the victim models, which can be infeasible in real-world systems. In contrast, we consider a more practical threat setting named the hard no-box attack, where an attacker only has access to unlabeled skeletal data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Gradient-Based Attack Strategies The core component of adversarial attacks is to generate adversarial samples [1]. Gradient-based attack methods have been widely used to introduce perturbations to a sample following the direction of the loss gradient. Goodfellow et al. [12] proposed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4598" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 286, + 235 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 286, + 235 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 286, + 235 + ], + "type": "image", + "image_path": "dd5f819e3929239968565e14cec3ac68b9cdb2f100e9c1864f1213d38f586cd0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 287, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 275 + ], + "type": "text", + "content": "Figure 1. The training (left) and attack (right) processes for the hard no-box attack. The trained query encoder in the training process is used for attacks in the attack process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 285, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 453 + ], + "type": "text", + "content": "the fast gradient sign method (FGSM) that perturbs a sample by a single step along the loss gradient. Kurakin et al. [14] proposed I-FGSM by extending FGSM to an iterative process. Dong et al. [8] presented MI-FGSM by adding momentum to the gradient, which boosted the transferability of adversarial samples. Xie et al. [40] applied diversified augmentations to the inputs before each iteration to craft more transferable samples. While these gradient-based strategies are successful in static data and have been adapted to skeletal motions, they neglect the dependency between frames for gradient computation, which is crucial in time series. This motivates us to propose our skeleton-motion-informed attack strategies, which explicitly model the motion dynamics in the temporal domain [39, 35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 464, + 258, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 258, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 258, + 477 + ], + "type": "text", + "content": "3. Hard No-Box Attack for Skeletal Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 483, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 593 + ], + "type": "text", + "content": "Figure 1 shows the overview of our method. The left part is the training process where we adopt contrastive learning to obtain a latent data manifold to distinguish data samples. The attack process is shown on the right-hand side. We first design a new no-box adversarial loss in the trained latent space to guide the adversary search using samples that are dissimilar to the attacked sample. Then we propose a novel skeleton-motion-informed gradient and a new family of attack methods for generating adversarial samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 270, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 270, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 270, + 612 + ], + "type": "text", + "content": "3.1. Contrastive Learning for Motion Manifold" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "While the fundamental idea of adversarial attacks is to perturb a data sample to cross class boundaries, such boundaries cannot be estimated for hard no-box attacks due to the lack of labels. To estimate such boundaries without labels, we present a new application of contrastive learning (CL) [34] to aggregate similar data samples as soft class boundaries in latent space. Such boundaries enable us to adversarially perturb a sample to cross boundaries. We also train" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "an encoder to extract discriminative high-level features for the motion manifold in latent space. Overall, our CL constructs boundaries in latent space without labels via aggregating similar samples and segregating dissimilar samples. Our attack is guided by the dissimilarity of high-level features between samples for generating adversarial samples, instead of using class boundaries to lead the attack [2]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": "To incorporate both spatial and temporal information, we train an encoder (Fig. 1 left) based on adaptive graph convolutional network (AGCN) [28]. To force encoders to focus on high-level features, we apply skeleton-specific data augmentations to an input sequence " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": " and obtain two different views " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "S_{q}" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": ". Augmentations include spatial operations (e.g. pose transformations, joint jittering, etc.) and temporal operations (e.g. temporal crop and resize) [34] (detailed in supplementary material). Then, we feed " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "S_{q}" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": " into the query encoder " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": " and the key encoder " + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 304, + 157, + 546, + 288 + ], + "type": "text", + "content": " respectively for the info-noise-contrastive estimation (InfoNCE) [23]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 294, + 558, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 294, + 558, + 343 + ], + "spans": [ + { + "bbox": [ + 306, + 294, + 558, + 343 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {c o n t r a s t}} = \\tag {1} \\\\ - \\log \\frac {\\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot f _ {k} \\left(S _ {k}\\right) / \\tau\\right)}{\\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot f _ {k} \\left(S _ {k}\\right) / \\tau\\right) + \\sum_ {F _ {n} \\sim N} \\exp \\left(f _ {q} \\left(S _ {q}\\right) \\cdot F _ {n} / \\tau\\right)}, \\\\ \\end{array}", + "image_path": "0d8064cf0a1cec4c63d4e81e81c68ed166682d9b725483910e82f6664681694b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "text", + "content": " is the temperature parameter, " + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "text", + "content": " is the dynamic queue storing the features of negative samples " + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "inline_equation", + "content": "F_{n}" + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "text", + "content": " obtained in the training process. After training, we use the query encoder " + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 304, + 350, + 546, + 399 + ], + "type": "text", + "content": ", which encodes the motion manifold, for attack." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 404, + 539, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 539, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 539, + 415 + ], + "type": "text", + "content": "3.2. Adversarial Loss for Unlabeled Skeletal Data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 566 + ], + "type": "text", + "content": "The adversarial loss of hard no-box attack is significantly different from most existing methods that heavily depend on labels and class boundaries. Since class labels and class boundaries are unavailable in hard no-box attacks, we utilize data samples that are dissimilar to a given sample (i.e. negative samples) for defining the adversarial loss. Correspondingly, samples that are similar to the given sample are considered as positive samples. We argue when a given sample is perturbed towards its negative samples and away from its positive samples, it tends to become an adversary. This is because the negative samples generally indicate the high-density areas of other classes in the latent space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 567, + 512, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 512, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 512, + 578 + ], + "type": "text", + "content": "The hard no-box adversarial loss is designed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 583, + 545, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 583, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 331, + 583, + 545, + 612 + ], + "type": "interline_equation", + "content": "L _ {a d v} = - \\log \\frac {\\exp [ S i m (f _ {q} (s) , f _ {q} (\\tilde {s})) ]}{\\sum_ {j} \\exp [ S i m (f _ {q} (s) , f _ {q} (\\tilde {s} _ {j})) ]}, \\tag {2}", + "image_path": "4460f7d473a8fbed495bcc7a8961ae15f70eda7908c9cf3467f886fc7e5053d6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "Sim" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " is the cosine similarity, " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " is the adversarial sample to be computed, " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{s}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " is the clean sample regarded as the positive sample, and " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{s}_j" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " are the negative samples. Maximizing " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "L_{adv}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " moves " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " away from " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{s}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " and towards " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\widetilde{s}_j" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " in the latent manifold. With " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "L_{adv}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": ", gradient-based attacks are employed." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To maximize Eq. 2, the selection of negative samples " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{s}_j" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is crucial and we design a method tailored for no-box attacks. Existing work [10] utilizes cluster-fit [42] to generate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4599" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "pseudo labels for selecting negative samples during adversarial training, which is less suitable for the no-box attack as it requires another pretrained off-line encoder to obtain pseudo labels. Instead, we adapt K-means, an unsupervised method, to select negatives, removing the need of any pretraining. We discard " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " clusters whose cluster centers are the closest to the input sample, mitigating the risk of misleading attacks. The remaining cluster centers are considered as the negative samples " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\tilde{s}_j" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 191, + 250, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 250, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 250, + 205 + ], + "type": "text", + "content": "4. Skeleton-Motion-Informed Gradient" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 212, + 289, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 289, + 464 + ], + "type": "text", + "content": "Existing gradient-based attack methods treat each dimension of the data as an independent variable, i.e. raw gradient. Attacks based on raw gradients tend to drag a sample away from the data manifold [11]. With the guidance of class boundaries and a limit on the perturbation budget, the raw gradient can still find deceiving adversaries. However, this setting is infeasible in hard no-box attacks. Without class boundaries, raw gradients that point to the negative samples can drag the adversary far away from the manifold. This is because while the perturbations are towards negative samples, they are not necessarily in a direction orthogonal to the class boundary. Consequently, larger perturbations are needed to cross the boundary, leading to adversaries being far off the manifold. This creates the need to constrain the perturbation within or near the manifold, at least locally. Since the motion manifold is constrained by the motion dynamics [38], we argue that the gradient needs to explicitly capture the dynamics. Therefore, we propose a new gradient named skeleton-motion-informed (SMI) gradient, capturing the manifold information that has been largely ignored by existing methods in loss gradient computation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 472, + 239, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 239, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 239, + 484 + ], + "type": "text", + "content": "4.1. Dynamics in the Gradient Structure" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "content": "Given a skeletal sequence " + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "inline_equation", + "content": "S = [S_{1}, S_{2}, \\dots, S_{t}]" + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "content": " and the adversarial loss " + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "inline_equation", + "content": "J(S)" + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "content": ", a straightforward but effective strategy to craft adversarial perturbations is the gradient-based attack [1]. It utilizes backpropagation of the loss function " + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\nabla J(S)" + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "content": " to iteratively change input samples " + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 490, + 287, + 552 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 562, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 562, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 109, + 562, + 287, + 576 + ], + "type": "interline_equation", + "content": "\\hat {S} = S + \\alpha \\cdot \\operatorname {s i g n} (\\nabla J (S)), \\tag {3}", + "image_path": "fcfa92920fe9a49442748886d7263e15e0b56daa0dcafbd615ebd7a355b583ef.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": " is the attack step size and " + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": " is the adversarial samples. In skeletal motions, this attack gradient " + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\nabla J(S)" + }, + { + "bbox": [ + 46, + 582, + 287, + 619 + ], + "type": "text", + "content": " consists of a set of partial derivatives over all frames:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 627, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 627, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 78, + 627, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\nabla J (S) = \\left[ \\frac {\\partial J (S)}{\\partial S _ {1}}, \\frac {\\partial J (S)}{\\partial S _ {2}}, \\dots , \\frac {\\partial J (S)}{\\partial S _ {t}} \\right]. \\tag {4}", + "image_path": "ff4a2209ff6570fc720ad92ef163248e63dad8505f147e24db96544d8349178f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 663, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 663, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 287, + 715 + ], + "type": "text", + "content": "The partial derivative " + }, + { + "bbox": [ + 46, + 663, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\partial J(S)}{\\partial S_t}" + }, + { + "bbox": [ + 46, + 663, + 287, + 715 + ], + "type": "text", + "content": " assumes each frame is independent, and this is the raw gradient employed in existing methods [36]. However, human motions contain rich dynamics so that the system can be described as " + }, + { + "bbox": [ + 46, + 663, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{t} = f(S_{< t})" + }, + { + "bbox": [ + 46, + 663, + 287, + 715 + ], + "type": "text", + "content": ". So" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "far, various dynamics models have been attempted to model human motions, such as Markovian models [33, 37], autoregressive models [39], and many-to-many mapping [38], all of which can capture the dynamics at different scales in time. We explore these models to reveal the missing dynamics in the structure of the raw gradient and propose our SMI-gradients that consider motion dynamics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 170, + 545, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 206 + ], + "type": "text", + "content": "Markovian Model We assume the motion dynamics can be captured by a Markovian model, i.e. " + }, + { + "bbox": [ + 304, + 170, + 545, + 206 + ], + "type": "inline_equation", + "content": "S_{t} = f_{d1}(S_{t - 1})" + }, + { + "bbox": [ + 304, + 170, + 545, + 206 + ], + "type": "text", + "content": ". This allows us to derive the 1st-order SMI-gradient:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 213, + 545, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 213, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 339, + 213, + 545, + 241 + ], + "type": "interline_equation", + "content": "\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 1}}\\right) _ {d 1} = \\frac {\\partial J (S)}{\\partial S _ {t - 1}} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\frac {d S _ {t}}{d S _ {t - 1}}, \\tag {5}", + "image_path": "64b098badd680b931a30bbe23d102b32e5e2aff3ac594bab5da8a3627cc9e9f8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 248, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 248, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\frac{dS_t}{dS_{t-1}}" + }, + { + "bbox": [ + 304, + 248, + 545, + 300 + ], + "type": "text", + "content": " is the temporal relationship between two consecutive frames that will be instantiated. Eq. 5 shows that the attack gradient along the motion manifold needs to consider the first-order information in the motion, e.g. velocity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 312, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 373 + ], + "type": "text", + "content": "Autoregressive Model Besides the first-order dynamics, we also model the second-order dynamics by assuming " + }, + { + "bbox": [ + 304, + 312, + 545, + 373 + ], + "type": "inline_equation", + "content": "S_{t} = f_{d2}(S_{t - 1},S_{t - 2})" + }, + { + "bbox": [ + 304, + 312, + 545, + 373 + ], + "type": "text", + "content": ", as 2nd-order dynamics (i.e. joint acceleration) capture the smooth temporal dynamics of skeletal motion [36]. We extend Eq. 5 as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 389, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 389, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 339, + 389, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(\\frac {\\partial J (S)}{\\partial S _ {t - 2}}\\right) _ {d 2} = \\frac {\\partial J (S)}{\\partial S _ {t - 2}} + \\frac {\\partial J (S)}{\\partial S _ {t - 1}} \\cdot \\frac {\\partial S _ {t - 1}}{\\partial S _ {t - 2}} \\tag {6} \\\\ + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\left(\\frac {\\partial S _ {t}}{\\partial S _ {t - 2}} + \\frac {\\partial S _ {t}}{\\partial S _ {t - 1}} \\cdot \\frac {\\partial S _ {t - 1}}{\\partial S _ {t - 2}}\\right). \\\\ \\end{array}", + "image_path": "61737593885f1f5a4f92ac4f76b15d2297ff237bbbeb822415a9ca1776f8e0fa.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 450, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 549 + ], + "type": "text", + "content": "While higher-order models can also be considered, there is an empirical evidence that the first three orders are the most important in skeletal motion adversarial attack [36]. Therefore, we express the SMI gradients of the whole skeletal sequence as " + }, + { + "bbox": [ + 304, + 450, + 545, + 549 + ], + "type": "inline_equation", + "content": "(\\nabla J(S))_{d1} = \\left[(\\frac{\\partial J(S)}{\\partial S_1})_{d1},(\\frac{\\partial J(S)}{\\partial S_2})_{d1},\\dots ,(\\frac{\\partial J(S)}{\\partial S_t})_{d1}\\right]" + }, + { + "bbox": [ + 304, + 450, + 545, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 545, + 549 + ], + "type": "inline_equation", + "content": "(\\nabla J(S))_{d2} = \\left[(\\frac{\\partial J(S)}{\\partial S_1})_{d2},(\\frac{\\partial J(S)}{\\partial S_2})_{d2},\\dots ,(\\frac{\\partial J(S)}{\\partial S_t})_{d2}\\right]" + }, + { + "bbox": [ + 304, + 450, + 545, + 549 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 554, + 545, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 578 + ], + "type": "text", + "content": "4.2. Time-Varying Autoregressive Models for Dynamics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 584, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 545, + 693 + ], + "type": "text", + "content": "We employ explicit models [39] to compute the dynamics-related derivatives in SMI gradients. While implicit models [33] may also be considered, they would require another network to be trained, making them less preferable in hard no-box attacks. To realize " + }, + { + "bbox": [ + 304, + 584, + 545, + 693 + ], + "type": "inline_equation", + "content": "f_{d1}" + }, + { + "bbox": [ + 304, + 584, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 584, + 545, + 693 + ], + "type": "inline_equation", + "content": "f_{d2}" + }, + { + "bbox": [ + 304, + 584, + 545, + 693 + ], + "type": "text", + "content": ", we use time-varying autoregressive models (TV-AR) [3], which effectively estimates the dynamics of skeleton sequence [39] due to its ability of modelling the temporal non-stationary signals:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 358, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "f _ {d 1}: S _ {t} = A _ {t} \\cdot S _ {t - 1} + B _ {t} + \\gamma_ {t}, \\tag {7}", + "image_path": "0da0dc43b0474aa1ed9d89fe65b3701ed84f436081ce7d85a397d3593f2fff7a.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4600" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 74, + 84, + 286, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 84, + 286, + 97 + ], + "spans": [ + { + "bbox": [ + 74, + 84, + 286, + 97 + ], + "type": "interline_equation", + "content": "f _ {d 2}: S _ {t} = C _ {t} \\cdot S _ {t - 1} + D _ {t} \\cdot S _ {t - 2} + E _ {t} + \\gamma_ {t}, \\tag {8}", + "image_path": "f9b7015b06859c3db601d636f4415c7bbe9b45077a87525f014b05216f9c35d4.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "text", + "content": "where Eq. 7 and Eq. 8 are denoted as TV-AR(1) and TV-AR(2) respectively. The model parameters " + }, + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "inline_equation", + "content": "\\beta_{t}^{1} = [A_{t},B_{t}]" + }, + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "inline_equation", + "content": "\\beta_{t}^{2} = [C_{t},D_{t},E_{t}]" + }, + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "text", + "content": " are all time-varying parameters and determined by data-fitting. " + }, + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 47, + 102, + 286, + 161 + ], + "type": "text", + "content": " is a time-dependent white noise representing the dynamics of stochasticity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 161, + 250, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 161, + 250, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 161, + 250, + 175 + ], + "type": "text", + "content": "Using Eq. 7 to compute " + }, + { + "bbox": [ + 59, + 161, + 250, + 175 + ], + "type": "inline_equation", + "content": "\\frac{\\partial S_t}{\\partial S_{t - 1}}" + }, + { + "bbox": [ + 59, + 161, + 250, + 175 + ], + "type": "text", + "content": ", Eq. 5 becomes:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 184, + 287, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 184, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 88, + 184, + 287, + 212 + ], + "type": "interline_equation", + "content": "\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 1}}\\right) _ {d 1} = \\frac {\\partial J (S)}{\\partial S _ {t - 1}} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot A _ {t}. \\tag {9}", + "image_path": "951d9f0a24f431a355f3b94e78e9361f931f3e3fae37140ecd124caf04056eee.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "text", + "content": "Similarly, using Eq. 8, we can compute " + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "inline_equation", + "content": "C_t = \\frac{\\partial S_t}{\\partial S_{t-1}}" + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "inline_equation", + "content": "D_t = \\frac{\\partial S_t}{\\partial S_{t-2}}" + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\frac{\\partial S_{t-1}}{\\partial S_{t-2}}" + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "inline_equation", + "content": "S_{t-1} = C_{t-1} \\cdot S_{t-2} + D_{t-1} \\cdot S_{t-3} + E_{t-1} + \\gamma_t" + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "text", + "content": " to compute it: " + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "inline_equation", + "content": "C_{t-1} = \\frac{\\partial S_{t-1}}{\\partial S_{t-2}}" + }, + { + "bbox": [ + 47, + 220, + 287, + 277 + ], + "type": "text", + "content": ". Then, Eq. 6 becomes:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 281, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 281, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\left(\\frac {\\partial J (S)}{\\partial S _ {t - 2}}\\right) _ {d 2} = \\frac {\\partial J (S)}{\\partial S _ {t - 2}} + \\frac {\\partial J (S)}{\\partial S _ {t - 1}} \\cdot C _ {t - 1} + \\frac {\\partial J (S)}{\\partial S _ {t}} \\cdot \\left(D _ {t} + C _ {t} \\cdot C _ {t - 1}\\right). \\tag {10}", + "image_path": "e13e47a2d5a1bcde4f44cf5f28a8ef2625bb392d2e3d1ef1c73d8f392b737945.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 323, + 236, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 236, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 236, + 335 + ], + "type": "text", + "content": "5. Skeleton-Motion-Informed Attack" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": "We construct new gradient-based attack methods based on our novel SMI gradient. Due to its compatibility, our proposed gradient can be integrated with most existing gradient-based methods. We select I-FGSM and MI-FGSM, which have been proven for their efficiency on S-HAR attacks [36, 18]. We augment them with first and second-order SMI-gradients, leading to four new attack methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": "Fast Gradient Sign Methods (FGSM) FGSM [12] is a single-step attack method that generates the adversarial samples " + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\hat{S} = S + r" + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": " by maximizing the adversarial loss function " + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "inline_equation", + "content": "J(S)" + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": " denotes an adversarial perturbation that is constrained within a budget " + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\|r\\|_p < \\epsilon" + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|_p" + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "inline_equation", + "content": "l_p" + }, + { + "bbox": [ + 46, + 427, + 287, + 522 + ], + "type": "text", + "content": "-norm. One variant of FGSM is the Iterative Fast Gradient Sign Method (I-FGSM) [14], which extends FGSM to an iterative process:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 98, + 531, + 287, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 531, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 98, + 531, + 287, + 546 + ], + "type": "interline_equation", + "content": "\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} (\\nabla_ {s} J (S)), \\tag {11}", + "image_path": "834bf4e6ff70731dafea5a68c36a4278308495187d64bb6451756e7793a2770b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 555, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 555, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 555, + 287, + 589 + ], + "type": "text", + "content": " is the attack step size and " + }, + { + "bbox": [ + 46, + 555, + 287, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 555, + 287, + 589 + ], + "type": "text", + "content": " means iteration. Another variant is MI-FGSM [8], which considers the momentum of the attack to avoid local maxima:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 596, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 596, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 287, + 623 + ], + "type": "interline_equation", + "content": "g ^ {i + 1} = \\mu \\cdot g ^ {i} + \\frac {\\nabla_ {s} J (S)}{\\| \\nabla_ {s} J (S) \\| _ {1}} \\tag {12}", + "image_path": "10561da783c971667c8f5a8779279820d1336ed1f023b6fd6ffd730d57e674ac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 634, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 634, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} \\left(g ^ {i + 1}\\right), \\tag {13}", + "image_path": "cc027b5a998428546d9937e00db8f0432ebba6e55ebb75d881706d054768e1c4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "text", + "content": " is the momentum decay factor and " + }, + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "text", + "content": " is the gradient in iteration " + }, + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 654, + 287, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "SMI-gradient Based Attacks We replace the original gradient " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\nabla J(S)" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " in I-FGSM and MI-FGSM with our SMI gradient " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\nabla J(S))_{d1}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\nabla J(S))_{d2}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": ". This creates four new" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 470, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 470, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 470, + 83 + ], + "type": "text", + "content": "Algorithm 1 S1I-FGSM and S2I-FGSM" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": "Input: An encoder " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": " with a loss function " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": "; a skeletal sequence sample " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": "; the size of attack step " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": "; iterations " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": "; the budget of perturbation " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": ". \nOutput: An adversarial example " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\| \\hat{S} - S \\|_p < \\epsilon" + }, + { + "bbox": [ + 306, + 85, + 545, + 114 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 115, + 400, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 115, + 400, + 124 + ], + "spans": [ + { + "bbox": [ + 312, + 115, + 400, + 124 + ], + "type": "text", + "content": "1: Initialization: " + }, + { + "bbox": [ + 312, + 115, + 400, + 124 + ], + "type": "inline_equation", + "content": "\\hat{S}^0 = S" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 125, + 545, + 143 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 125, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 545, + 143 + ], + "type": "text", + "content": "2: Fitting " + }, + { + "bbox": [ + 312, + 125, + 545, + 143 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 312, + 125, + 545, + 143 + ], + "type": "text", + "content": " with TV-AR model to obtain the time-varying parameters " + }, + { + "bbox": [ + 312, + 125, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 144, + 395, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 144, + 395, + 152 + ], + "spans": [ + { + "bbox": [ + 312, + 144, + 395, + 152 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 312, + 144, + 395, + 152 + ], + "type": "inline_equation", + "content": "i = 0" + }, + { + "bbox": [ + 312, + 144, + 395, + 152 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 144, + 395, + 152 + ], + "type": "inline_equation", + "content": "I - 1" + }, + { + "bbox": [ + 312, + 144, + 395, + 152 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 154, + 383, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 154, + 383, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 154, + 383, + 162 + ], + "type": "text", + "content": "4: Input " + }, + { + "bbox": [ + 312, + 154, + 383, + 162 + ], + "type": "inline_equation", + "content": "S^i" + }, + { + "bbox": [ + 312, + 154, + 383, + 162 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 154, + 383, + 162 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 312, + 154, + 383, + 162 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 163, + 462, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 462, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 462, + 172 + ], + "type": "text", + "content": "5: Obtain the raw gradient " + }, + { + "bbox": [ + 312, + 163, + 462, + 172 + ], + "type": "inline_equation", + "content": "\\nabla J(\\hat{S}^i)" + }, + { + "bbox": [ + 312, + 163, + 462, + 172 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 312, + 163, + 462, + 172 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 312, + 163, + 462, + 172 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "text", + "content": "6: Calculate the SMI gradient " + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "inline_equation", + "content": "(\\nabla J(\\hat{S}^i))_{d1}" + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "text", + "content": " with Eq. 9, or " + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "inline_equation", + "content": "(\\nabla J(\\hat{S}^i))_{d2}" + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "text", + "content": " with Eq. 10, using " + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "inline_equation", + "content": "\\nabla J(\\hat{S}^i)" + }, + { + "bbox": [ + 312, + 173, + 533, + 194 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 194, + 486, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 194, + 486, + 204 + ], + "spans": [ + { + "bbox": [ + 312, + 194, + 486, + 204 + ], + "type": "text", + "content": "7: Update " + }, + { + "bbox": [ + 312, + 194, + 486, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{S}^{i + 1}" + }, + { + "bbox": [ + 312, + 194, + 486, + 204 + ], + "type": "text", + "content": " by applying the sign gradient as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 364, + 209, + 545, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 209, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 364, + 209, + 545, + 231 + ], + "type": "interline_equation", + "content": "\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\operatorname {s i g n} (\\nabla J (\\hat {S} ^ {i})) _ {d 1}), o r \\tag {14}", + "image_path": "10ff06ace910702128f01b370ed256b6f6c7a978353dafbeea001721262df2fa.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 365, + 229, + 494, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 229, + 494, + 243 + ], + "spans": [ + { + "bbox": [ + 365, + 229, + 494, + 243 + ], + "type": "interline_equation", + "content": "\\hat {S} ^ {i + 1} = \\hat {S} ^ {i} + \\alpha \\cdot \\mathrm {s i g n} \\left(\\nabla J (\\hat {S} ^ {i})\\right) _ {d 2}).", + "image_path": "7f9107983d897934ac0cb6a69afe0dac3408855d4587695aa828200c65785b99.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 248, + 350, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 248, + 350, + 256 + ], + "spans": [ + { + "bbox": [ + 312, + 248, + 350, + 256 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 258, + 375, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 258, + 375, + 266 + ], + "spans": [ + { + "bbox": [ + 312, + 258, + 375, + 266 + ], + "type": "text", + "content": "9: return " + }, + { + "bbox": [ + 312, + 258, + 375, + 266 + ], + "type": "inline_equation", + "content": "\\hat{S} = \\hat{S}^I" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "text", + "content": "dynamic attack strategies: first-order SMI I-FGSM " + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_1\\mathrm{I}" + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "text", + "content": " FGSM), second-order SMI I-FGSM " + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "text", + "content": " -FGSM), firstorder SMI MI-FGSM " + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_1\\mathrm{MI}" + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "text", + "content": " -FGSM), and second-order SMI MI-FGSM " + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 306, + 292, + 545, + 364 + ], + "type": "text", + "content": " -FGSM). The processes of SI-FGSM are shown in Algorithm 1. The algorithm of SMI-FGSM can be found in the supplementary material." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 378, + 386, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 386, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 386, + 392 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 399, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 422 + ], + "type": "text", + "content": "We refer the readers to the supplementary material for extra experimental results." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 424, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 555 + ], + "type": "text", + "content": "Datasets We select three widely used skeletal datasets: HDM05 [22] (2,337 sequences of 130 classes performed by 5 actors), NTU60 [26] (56,880 sequences of 60 classes), and NTU120 [19] (114,480 sequences of 120 classes, an extended version of NTU60, one of the largest datasets in the field). We pre-process HDM05 following [9] and both NTU datasets following [28]. The different skeletons are mapped to a standard 25-joint structure as in [38]. For our hard no-box attack, we only use the testing data and do not use the training data, the training labels and the testing labels during attacks." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "type": "text", + "content": "Target Models We choose multiple state-of-the-art models as victims: ST-GCN [41], 2s-AGCN [28], AS-GCN [15], SGN [44] and MS-G3D [20]. They are trained using the official implementations and following the training protocols. For 2s-AGCN, we attack both the single joint stream model (js-AGCN) and the two-stream model." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details We pre-train the CL encoder " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " following [34]. The unsupervised network is trained with a temperature value " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.07" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and SGD optimizer for 450 epochs. The learning rate is set to 0.01 with a weight decay of 0.0001. Due to the limitation of hard no-box settings, the attacker cannot access the training samples during the whole process. Therefore, the encoder is trained on the test-" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4601" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 69, + 476, + 236 + ], + "blocks": [ + { + "bbox": [ + 118, + 69, + 476, + 236 + ], + "lines": [ + { + "bbox": [ + 118, + 69, + 476, + 236 + ], + "spans": [ + { + "bbox": [ + 118, + 69, + 476, + 236 + ], + "type": "table", + "html": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1 I-FGSMNo-box S2 I-FGSMNo-box MI-FGSMNo-box S1 MI-FGSMNo-box S2 MI-FGSM
0.01js-AGCN11.21%—26.05%28.09%30.87%30.68%34.75%36.58%
2s-AGCN5.36%—13.94%15.04%16.45%16.47%18.23%19.31%
ST-GCN3.57%12.93%9.55%9.86%9.96%11.11%11.36%11.56%
MS-G3D8.39%35.23%10.57%11.14%11.69%11.76%12.85%14.11%
SGN21.25%26.03%34.09%34.46%35.23%38.49%38.75%38.81%
ASGCN5.69%20.87%13.92%14.85%14.67%15.95%16.82%17.75%
0.008js-AGCN10.12%—22.84%24.36%26.46%25.70%29.64%30.88%
2s-AGCN5.04%—11.36%12.19%13.03%12.30%13.85%15.56%
ST-GCN3.26%10.33%7.87%7.99%8.04%8.84%9.07%9.19%
MS-G3D5.19%31.28%9.18%9.51%9.99%10.01%10.26%12.98%
SGN20.95%23.30%29.61%30.20%30.74%33.32%33.63%34.54%
ASGCN5.54%18.77%11.42%12.05%12.29%12.77%13.69%14.37%
0.006js-AGCN7.19%—19.70%20.32%21.23%20.04%22.76%24.33%
2s-AGCN3.70%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN2.46%7.88%5.65%5.85%5.97%6.25%6.54%6.66%
MS-G3D4.46%23.15%7.30%7.84%7.61%7.94%8.05%8.39%
SGN16.76%19.93%23.92%24.57%25.75%26.74%27.04%27.64%
ASGCN3.94%14.30%8.79%9.23%9.07%9.71%10.29%10.90%
", + "image_path": "78b6e54b327d6c70e16912cfd5adcf2a53449bb6aaeb0b6661fef11645c705f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 118, + 250, + 476, + 416 + ], + "blocks": [ + { + "bbox": [ + 91, + 237, + 500, + 247 + ], + "lines": [ + { + "bbox": [ + 91, + 237, + 500, + 247 + ], + "spans": [ + { + "bbox": [ + 91, + 237, + 500, + 247 + ], + "type": "text", + "content": "Table 2. The fooling rate of different methods on the target models in NTU60, where " + }, + { + "bbox": [ + 91, + 237, + 500, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 91, + 237, + 500, + 247 + ], + "type": "text", + "content": " is the perturbation budget." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 118, + 250, + 476, + 416 + ], + "lines": [ + { + "bbox": [ + 118, + 250, + 476, + 416 + ], + "spans": [ + { + "bbox": [ + 118, + 250, + 476, + 416 + ], + "type": "table", + "html": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN9.97%—23.92%24.64%25.79%27.26%27.93%29.07%
2s-AGCN6.38%—20.09%21.11%22.06%23.51%24.63%24.96%
ST-GCN12.18%23.84%23.53%24.73%25.77%26.95%27.31%28.76%
MS-G3D10.63%24.63%19.07%20.11%20.20%21.57%21.95%22.73%
SGN19.65%37.90%31.43%32.64%33.75%38.47%37.96%38.85%
ASGCN7.29%24.15%18.03%19.29%20.37%19.88%20.22%21.60%
0.008js-AGCN9.04%—21.58%21.77%22.47%24.28%24.62%25.74%
2s-AGCN5.79%—15.71%15.23%15.94%18.68%18.76%19.57%
ST-GCN11.23%21.35%20.74%21.51%22.22%23.52%24.65%24.87%
MS-G3D9.98%21.73%17.06%17.74%17.92%19.53%19.29%20.33%
SGN17.59%35.88%27.38%28.17%28.65%32.43%33.06%32.55%
ASGCN6.87%21.87%15.73%16.93%17.62%17.41%17.75%18.70%
0.006js-AGCN8.23%—18.51%18.88%18.79%20.74%21.88%22.06%
2s-AGCN5.09%—7.93%9.80%10.56%9.80%11.26%12.66%
ST-GCN9.54%17.51%17.60%17.97%18.67%18.90%19.91%20.34%
MS-G3D9.42%18.07%15.06%15.09%15.20%16.86%17.25%17.56%
SGN16.82%34.10%22.39%22.65%22.76%25.43%25.79%25.92%
ASGCN5.39%19.02%12.90%13.45%14.97%14.08%15.26%16.33%
", + "image_path": "635340cc4d5bdc709ddec659ca19ae8e148a2fa38e0bb1addc665bb3d2141a05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 417, + 502, + 427 + ], + "lines": [ + { + "bbox": [ + 88, + 417, + 502, + 427 + ], + "spans": [ + { + "bbox": [ + 88, + 417, + 502, + 427 + ], + "type": "text", + "content": "Table 3. The fooling rate of different methods on the target models in NTU120, where " + }, + { + "bbox": [ + 88, + 417, + 502, + 427 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 88, + 417, + 502, + 427 + ], + "type": "text", + "content": " is the perturbation budget." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "text", + "content": "ing set. We adopt the " + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "text", + "content": " norm for the perturbation budget " + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "text", + "content": ". For clusters of negative samples, the number of clusters is 120, and the number of deleted centers " + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 440, + 287, + 475 + ], + "type": "text", + "content": " is 10." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "content": "Evaluation Metrics We employ the fooling rate as a major metric. It is defined as the percentage of data samples whose predicted labels changed after adversarial attacks [18]. Besides, inspired by [36], we define a perceptual deviation indicator to evaluate the imperceptibility of adversarial skeletal samples:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 569, + 286, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 286, + 636 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta p = \\frac {1}{M T} \\sum_ {n = 0} ^ {M} \\left\\| S - \\hat {S} \\right\\| _ {2} + \\frac {1}{M T} \\sum_ {n = 0} ^ {M} \\left\\| B - \\hat {B} \\right\\| _ {2} \\tag {15} \\\\ + \\frac {1}{M T L} \\sum_ {n = 0} ^ {M} \\left\\| \\ddot {S} - \\ddot {\\hat {S}} \\right\\| _ {2} \\\\ \\end{array}", + "image_path": "91ca6af40fb56fab160daaa7c443dae230afaa0f8a753b928bc02ea79e973e72.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is the number of adversarial samples, " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is the total number of frames, and " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is the number of joints in the skeleton. The three terms evaluate the deviations of joint position, bone-length, and acceleration, respectively. A smaller perceptual deviation indicates better imperceptibility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 439, + 425, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 425, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 425, + 451 + ], + "type": "text", + "content": "6.1. Hard No-Box Attack" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 460, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 663 + ], + "type": "text", + "content": "Baselines We establish two baselines using transfer-based attacks based on a self-supervised and a supervised classifier. This is because our method is the first hard no-box attack and there is no other similar method. The transfer-based attack is the closest setting to ours as they do not require access to the victim model. The first baseline (Self-sup Attacker) is a self-supervised surrogate model with a linear layer appended to our CL encoder. We freeze our CL encoder after the manifold training and then train the linear layer supervisedly [17]. In this way, the need to access labels is only for training the linear layer, not the CL encoder, which is not strictly hard no-box but closer than existing methods. The second baseline is a standard transfer-based attack where we use js-AGCN as the surrogate model and SMART [36] as the white-box attacker (AGCN Attacker). Unlike our method, both baselines still require access to the training data and the training labels during attacks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "For hard no-box attack, different attack strategies are compared including I-FGSM, MI-FGSM, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_1\\mathrm{I}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " -FGSM, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " FGSM, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_1\\mathrm{MI}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " -FGSM, and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " -FGSM. We run all the attackers for 400 iterations and report the attack perfor" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4602" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 69, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 113, + 69, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 113, + 69, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 113, + 69, + 481, + 236 + ], + "type": "table", + "html": "
Victim modelsSelf-sup AttackerAGCN AttackerNo-box I-FGSMNo-box S1I-FGSMNo-box S2I-FGSMNo-box MI-FGSMNo-box S1MI-FGSMNo-box S2MI-FGSM
0.01js-AGCN10.12%—10.61%10.98%11.55%13.45%14.20%14.20%
2s-AGCN1.89%—5.30%5.68%5.68%6.44%6.44%6.82%
ST-GCN6.44%24.26%9.47%9.85%9.47%8.90%11.17%11.74%
MS-G3D4.17%86.95%41.67%44.51%43.94%53.03%56.06%53.41%
SGN1.89%3.98%2.46%2.84%3.03%3.03%3.40%3.59%
ASGCN1.89%27.57%3.22%4.36%3.22%4.36%5.68%5.68%
0.008js-AGCN9.46%—8.14%9.09%9.09%10.61%10.98%11.55%
2s-AGCN1.70%—4.36%4.92%4.73%4.55%5.30%5.87%
ST-GCN3.79%21.69%8.33%8.52%8.52%8.33%8.90%8.33%
MS-G3D3.60%82.17%25.76%30.68%29.17%36.36%39.39%39.02%
SGN1.51%3.60%0.57%1.51%2.08%2.27%2.27%2.46%
ASGCN1.70%22.43%2.46%2.84%2.84%2.27%2.84%3.03%
0.006js-AGCN7.19%—5.11%7.01%6.44%6.06%6.63%7.77%
2s-AGCN1.33%—2.08%2.08%2.27%3.79%3.98%4.17%
ST-GCN3.40%19.85%6.44%7.01%6.63%7.58%8.33%7.77%
MS-G3D2.84%72.43%11.55%12.31%11.92%15.34%18.94%17.99%
SGN0.57%3.40%0.13%0.94%0.57%1.13%1.32%1.51%
ASGCN1.52%17.82%0.38%0.76%1.89%0.57%2.27%2.27%
", + "image_path": "37637b2e9a792a32b48ba5ab950a680e616e72d7003a41683b8ddd9a57cd9b99.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 257, + 507, + 346 + ], + "blocks": [ + { + "bbox": [ + 77, + 257, + 507, + 346 + ], + "lines": [ + { + "bbox": [ + 77, + 257, + 507, + 346 + ], + "spans": [ + { + "bbox": [ + 77, + 257, + 507, + 346 + ], + "type": "image", + "image_path": "20813cfe522b2390c19926f60308b2ec58116b33b20cef0ed9b8f4e5b2a44e1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 350, + 518, + 361 + ], + "lines": [ + { + "bbox": [ + 73, + 350, + 518, + 361 + ], + "spans": [ + { + "bbox": [ + 73, + 350, + 518, + 361 + ], + "type": "text", + "content": "Figure 2. Visual comparisons between attack strategies in no-box attacks " + }, + { + "bbox": [ + 73, + 350, + 518, + 361 + ], + "type": "inline_equation", + "content": "(\\epsilon = 0.006)" + }, + { + "bbox": [ + 73, + 350, + 518, + 361 + ], + "type": "text", + "content": " with key visual differences highlighted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 371, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 419 + ], + "type": "text", + "content": "mance under different perturbation budgets " + }, + { + "bbox": [ + 46, + 371, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 371, + 287, + 419 + ], + "type": "text", + "content": " in Table 2 for NTU60, Table 3 for NTU120, and Table 4 for HDM05. We omit the results on js-AGCN and 2s-AGCN under AGCN Attacker because AGCN is the surrogate model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 626 + ], + "type": "text", + "content": "Table 2-4 show that the hard no-box attack poses real threats to a range of S-HAR classifiers. In general, the fooling rate of hard no-box attacks is higher than Self-sup Attackers. This is surprising as the Attacker utilizes training data and training labels while our method does not. Looking further, AGCN Attacker is much better than Self-sup Attacker and the major difference is their feature extraction (i.e. one is a CL encoder and one is a graph network). This shows transfer-based attack heavily relies on the feature extraction ability of the surrogate model that cannot bypass access to the training data and the training labels. Furthermore, when compared with AGCN Attacker, our method achieves similar fooling rates, varying across different victims and datasets. Given that AGCN Attacker requires access to training data, training labels and testing labels, we argue hard no-box attacker achieves superior results and provides a more realistic setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "In addition, among various no-box attack strategies, " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " FGSM performs the best and often by big margins. All the SMI gradient-based methods generate stronger adversaries compared with baselines I-FGSM and MI-FGSM. The 2nd-order SMI attack method usually outperforms the corresponding 1st-order version. Last, we notice a variance in fooling rate across different victims and datasets. For in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 371, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 467 + ], + "type": "text", + "content": "stance, the multi-stream model (2s-AGCN) significantly enhances the robustness compared to the single-stream model (js-AGCN); the fooling rate of all methods drops by nearly half. This may be because the multi-stream model can ensemble features from different modalities, which improves the robustness. In general, it is still an open question why fooling rate can vary across victims and datasets and we leave the theoretical analysis for future work." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 329, + 478, + 524, + 550 + ], + "blocks": [ + { + "bbox": [ + 90, + 237, + 501, + 247 + ], + "lines": [ + { + "bbox": [ + 90, + 237, + 501, + 247 + ], + "spans": [ + { + "bbox": [ + 90, + 237, + 501, + 247 + ], + "type": "text", + "content": "Table 4. The fooling rate of different methods on the target models in HDM05, where " + }, + { + "bbox": [ + 90, + 237, + 501, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 90, + 237, + 501, + 247 + ], + "type": "text", + "content": " is the perturbation budget." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 478, + 524, + 550 + ], + "lines": [ + { + "bbox": [ + 329, + 478, + 524, + 550 + ], + "spans": [ + { + "bbox": [ + 329, + 478, + 524, + 550 + ], + "type": "table", + "html": "
Surrogate ModelVictimsIFGSMMI-FGSMS2I-FGSM (Ours)S2MI-FGSM (Ours)
2s-AGCNSTGCN2.10%2.10%3.00%3.01%
MS-G3D2.58%2.59%2.90%2.97%
ST-GCN2sAGCN2.20%2.34%2.44%2.64%
MS-G3D2.00%2.10%2.63%2.92%
MS-G3DSTGCN1.71%1.69%2.65%2.67%
2sAGCN1.76%1.79%2.03%2.07%
", + "image_path": "6ff7631e7a901c8e722597b50a8e69bc99d20558761143c8c212ae1b24071db9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 550, + 545, + 571 + ], + "lines": [ + { + "bbox": [ + 306, + 550, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 545, + 571 + ], + "type": "text", + "content": "Table 5. The fooling rate of different attack strategies in transferred SMART attacks, where attack budgets " + }, + { + "bbox": [ + 306, + 550, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.01" + }, + { + "bbox": [ + 306, + 550, + 545, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 586, + 486, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 486, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 486, + 597 + ], + "type": "text", + "content": "6.2. Transfer-Based Black-Box Attack" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "SMI gradient not only improves the transferability in hard no-box attacks, but also enhances other gradient-based skeletal attacks. Here, we employ SMART [36], a white-box attacker, as a baseline to compare different attack strategies. In the original SMART settings, I-FGSM is adopted to generate adversarial samples. We replace it with " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " FGSM, MI-FGSM, and " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " -FGSM to make a comparison. As all strategies achieve similar fooling rates in white-box attacks, we mainly focus on their transferability using" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4603" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "different surrogate models. We utilize SMART to attack 2s-AGCN, ST-GCN, and MS-G3D on NTU60 and transfer the obtained samples to other victim networks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "text", + "content": "The performance of the transferred black-box attack is shown in Table 5. " + }, + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "text", + "content": " FGSM gives the best performance in transfer-based black-box attacks. " + }, + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "text", + "content": " FGSM also improves the transferability of adversarial samples compared with baselines. In contrast, MI-FGSM, which succeeds in the image transfer-based attack, struggles in skeletal data. Its performance declines to " + }, + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "inline_equation", + "content": "1.69\\%" + }, + { + "bbox": [ + 46, + 109, + 288, + 301 + ], + "type": "text", + "content": " when it attacks STGCN via MS-G3D. Overall, the success rate is low in Table 5 because SMART is sensitive to the chosen surrogate [36]. Transfer-based attacks are proven to suffer from lower fooling rates in S-HAR attack [36] and improving the transferability is still an open problem. Table 5 aims to show our SMI gradient can improve it by incorporating motion dynamics into the attack gradient, compared with other alternative gradients. We will include in future work how to further explore this dynamics for better attack transfer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 310, + 162, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 162, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 162, + 323 + ], + "type": "text", + "content": "6.3. Perceptual Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "text", + "content": "A key feature of SMI gradient-based attacks is the improvement in the perceptual quality of the adversarial samples due to the consideration of the motion manifold. To verify this, we employ quantitative comparison and qualitative visual analysis on the no-box adversarial samples under various strategies. We compare the perceptual quality " + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\Delta p" + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "text", + "content": " on NTU60 in Table 6. We find that " + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "text", + "content": " FGSM achieves the best imperceptibility and obtains a massive improvement compared with I-FGSM. In contrast, MI-FGSM's deviation is twice that of " + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "text", + "content": " FGSM. Although " + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 46, + 329, + 287, + 544 + ], + "type": "text", + "content": " FGSM does not achieve the best visual performance, it is still slightly better than I-FGSM and achieves a better trade-off between fooling rate and perceptual quality. This is understandable because our method considers dynamics that help to generate more on-manifold adversarial samples. Moreover, the 1st-order SMI attacks outperform baselines but cannot compete with 2nd-order SMI attacks. This demonstrates the importance of considering acceleration in the skeletal attack." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": "We show the visual comparison of poses under various attack strategies in no-box attacks in Figure 2. The spinal joints demonstrate the most obvious differences. " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " FGSM outperforms the other attack methods and gets the most natural poses, whereas I-FGSM has slight but noticeable joint displacements in the neck and head. " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " -FGSM performs better than its baseline, MI-FGSM, which shows zig-zag bending in the frame " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "t_0 + 2" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " .The samples produced by MI-FGSM have the worst imperceptibility, where we can easily find the unnatural jittery movements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We also evaluate perceptual quality " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Delta p" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " on SMART adversarial samples obtained with different gradients. Results conducted on the NTU60 dataset are shown in Table 7. " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{I}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " FGSM reaches the best perceptual performance compared" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "with all attack strategies. MI-FGSM slightly declines the imperceptibility. " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " -FGSM performs slightly worse in MS-G3D and ST-GCN. The reason is mainly that " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2\\mathrm{MI}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " FGSM takes more iterations in the white-box attack, leading to late stopping and slightly worse visual performance." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 335, + 140, + 517, + 198 + ], + "blocks": [ + { + "bbox": [ + 335, + 140, + 517, + 198 + ], + "lines": [ + { + "bbox": [ + 335, + 140, + 517, + 198 + ], + "spans": [ + { + "bbox": [ + 335, + 140, + 517, + 198 + ], + "type": "table", + "html": "
Strategiesε = 0.01ε = 0.008ε = 0.006
I-FGSM95.8768.5643.13
MI-FGSM131.7790.4454.61
S1I-FGSM84.6360.7738.15
S1MI-FGSM114.5178.0647.04
S2I-FGSM65.6045.6727.67
S2MI-FGSM90.9362.0437.46
", + "image_path": "0c0b567a28be45835006a52fbd061f43cf0ceac36dbaa10b3e33ccc4296d84be.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "text", + "content": "Table 6. The perceptual deviation of different attack strategies in no-box attacks with different budgets " + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 339, + 232, + 513, + 274 + ], + "blocks": [ + { + "bbox": [ + 339, + 232, + 513, + 274 + ], + "lines": [ + { + "bbox": [ + 339, + 232, + 513, + 274 + ], + "spans": [ + { + "bbox": [ + 339, + 232, + 513, + 274 + ], + "type": "table", + "html": "
VictimsI-FGSMMI-FGSMS2I-FGSM(Ours)S2MI-FGSM(Ours)
2s-AGCN1.521.621.251.49
MS-G3D2.392.461.693.02
ST-GCN1.131.181.101.47
", + "image_path": "f797aa31b492b1b2b9a64a8ec00994e3db8cf0ab787f7e8c587d2766182dafc0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 274, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 294 + ], + "type": "text", + "content": "Table 7. The perceptual deviation of different attack strategies in SMART for different victim models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 305, + 468, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 305, + 468, + 318 + ], + "spans": [ + { + "bbox": [ + 306, + 305, + 468, + 318 + ], + "type": "text", + "content": "7. Conclusions and Discussions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 325, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 456 + ], + "type": "text", + "content": "In this paper, we have verified potential threats to S-HAR solutions. A new setting is proposed: the hard no-box attack on skeletal motions without access to the victim model, the training samples or the labels. We validate our setting by proposing the first pipeline for hard no-box attacks. Moreover, as far as we know, we are the first to explore motion dynamics in the adversarial gradient computation, leading to a new SMI gradient compatible with existing gradient-based attacks. By extensive evaluation and comparison, our method has been proven to be threatening and imperceptible, relying on the least prior knowledge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 457, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 552 + ], + "type": "text", + "content": "The SMI gradient also improves the transferability of transferred black-box attacks. Nonetheless, boosting the transferability is still an open problem [36] and we will explore this further with our SMI gradient in the future. We will explore other models (e.g., diffusion models [5]) and other time-series data (e.g. stock price, videos) for the proposed attack. Also, our SMI gradient describes dynamics and may be beneficial for motion synthesis [4]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 636 + ], + "type": "text", + "content": "We call for attention to intensify the S-HAR robustness by considering defences against our hard no-box attack. We validate randomized smoothing [6] as a potential defence method in supplementary materials. Due to the least prior knowledge requirement, security risks posed by our attack can be reduced with such defenses. Otherwise, our attacks become a significantly threat to S-HAR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 646, + 408, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 408, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 408, + 659 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "This research is supported in part by National Natural Science Foundation of China (ref: 61673314, Yang), EPSRC (ref: EP/X031012/1, NorthHFutures, Shum) and EU Horizon 2020 (ref: 899739, CrowdDNA, Wang)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4604" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "text", + "content": "[1] Naveed Akhtar, Ajmal Mian, Navid Kardan, and Mubarak Shah. Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access, 9:155161-155196, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 139, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 288, + 183 + ], + "type": "text", + "content": "[2] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 287, + 239 + ], + "type": "text", + "content": "[3] Laura F Bringmann, Ellen L Hamaker, Daniel E Vigo, André Aubert, Denny Borsboom, and Francis Tuerlinckx. Changing dynamics: Time-varying autoregressive models using generalized additive modeling. Psychological methods, 22(3):409, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 287, + 285 + ], + "type": "text", + "content": "[4] Ziyi Chang, Edmund JC Findlay, Haozheng Zhang, and Hubert PH Shum. Unifying human motion synthesis and style transfer with denoising diffusion probabilistic models. arXiv preprint arXiv:2212.08526, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 289, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 287, + 320 + ], + "type": "text", + "content": "[5] Ziyi Chang, George A Koulieris, and Hubert PH Shum. On the design fundamentals of diffusion models: A survey. arXiv preprint arXiv:2306.04542, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 287, + 367 + ], + "type": "text", + "content": "[6] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In international conference on machine learning, pages 1310-1320. PMLR, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 287, + 424 + ], + "type": "text", + "content": "[7] Yunfeng Diao, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and He Wang. Basar: Black-box attack on skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7597-7607, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 427, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 427, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 53, + 427, + 287, + 482 + ], + "type": "text", + "content": "[8] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 484, + 287, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 484, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 53, + 484, + 287, + 528 + ], + "type": "text", + "content": "[9] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1110-1118, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 531, + 287, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 584 + ], + "type": "text", + "content": "[10] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in Neural Information Processing Systems, 34:21480-21492, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 620 + ], + "type": "text", + "content": "[11] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "text", + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world, pages 99-112. Chapman and Hall/CRC, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[15] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[16] Qizhang Li, Yiwen Guo, and Hao Chen. Practical no-box adversarial attacks against dnns. Advances in Neural Information Processing Systems, 33:12849-12860, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "text", + "content": "[17] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action recognition. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2490-2498, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[18] Jian Liu, Naveed Akhtar, and Ajmal Mian. Adversarial attack on skeleton-based human action recognition. IEEE Transactions on Neural Networks and Learning Systems, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "type": "text", + "content": "[19] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "text", + "content": "[20] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "text", + "content": "[21] Qianhui Men, Edmond S. L. Ho, Hubert P. H. Shum, and Howard Leung. A quadruple diffusion convolutional recurrent network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(9):3417-3432, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[22] Meinard Müller, Tido Röder, Michael Clausen, Bernhard Eberhardt, Björn Krüger, and Andreas Weber. Mocap database hdm05. Institut für Informatik II, Universität Bonn, 2(7), 2007." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "text", + "content": "[23] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "text", + "content": "[24] Bin Ren, Mengyuan Liu, Runwei Ding, and Hong Liu. A survey on 3d skeleton-based action recognition using learning method. arXiv preprint arXiv:2002.05907, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 569, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 611 + ], + "type": "text", + "content": "[25] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 11957-11965, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "text", + "content": "[26] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better aggregation in test-time augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1214-1223, October 2021." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4605" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 184 + ], + "type": "text", + "content": "[29] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "type": "text", + "content": "[30] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "text", + "content": "[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "text", + "content": "[32] Nariki Tanaka, Hiroshi Kera, and Kazuhiko Kawamoto. Adversarial bone length attack on action recognition. arXiv preprint arXiv:2109.05830, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[33] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qilong Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Trans. Graph., 41(4), jul 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[34] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "text", + "content": "[35] He Wang, Yunfeng Diao, Zichang Tan, and Guodong Guo. Defending black-box skeleton-based human activity classifiers. arXiv preprint arxiv.2203.04713, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 498 + ], + "type": "text", + "content": "[36] He Wang, Feixiang He, Zhexi Peng, Tianjia Shao, Yong-Liang Yang, Kun Zhou, and David Hogg. Understanding the robustness of skeleton-based action recognition under adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14656-14665, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[37] He Wang, Edmond SL Ho, and Taku Komura. An energy-driven motion planning method for two distant postures. IEEE transactions on visualization and computer graphics, 21(1):18-30, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[38] He Wang, Edmond SL Ho, Hubert PH Shum, and Zhanxing Zhu. Spatio-temporal manifold learning for human motions via long-horizon modeling. IEEE transactions on visualization and computer graphics, 27(1):216-227, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[39] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[40] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[41] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 307 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "recognition. In Thirty-second AAAI conference on artificial intelligence, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[42] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. Clusterfit: Improving generalization of visual representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6509-6518, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[43] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. arXiv preprint arXiv:2211.13466, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 261 + ], + "type": "text", + "content": "[44] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1112-1121, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 263, + 547, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 547, + 307 + ], + "type": "text", + "content": "[45] Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learning from partial spatio-temporal skeleton sequences. arXiv preprint arXiv:2302.09018, 2023." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "4606" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_content_list.json b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e24b8c59b06052136edff8e5f625995f9120b8fb --- /dev/null +++ b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_content_list.json @@ -0,0 +1,1720 @@ +[ + { + "type": "text", + "text": "Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis", + "text_level": 1, + "bbox": [ + 78, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiucheng Wu $^{1*}$ , Yujuan Liu $^{1*}$ , Handong Zhao $^{2}$ , Trung Bui $^{2}$ , Zhe Lin $^{2}$ , Yang Zhang $^{3}$ , Shiyu Chang $^{1}$ , UC Santa Barbara, $^{2}$ Adobe Research, $^{3}$ MIT-IBM Watson AI Lab", + "bbox": [ + 94, + 202, + 872, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qiucheng, yujianliu}@ucsb.edu", + "bbox": [ + 351, + 241, + 617, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/45f9aa9045642a42774a107f3724fd5ab47c48b40588e1e3865a64ec655fa4f1.jpg", + "image_caption": [ + "StableDiffusion" + ], + "image_footnote": [], + "bbox": [ + 83, + 309, + 209, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc6bcb2359c4be77375ed66aa8aeb226ef0e3c7598eaded3f6c4a043640be375.jpg", + "image_caption": [ + "Missing Objects", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 210, + 309, + 336, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdbb9a1becc861fe770277192e5d698fe823e2ae886f39f218db105df0abce66.jpg", + "image_caption": [ + "An airport employee giving directions to an airplane.", + "A zebra brazing on green grass next to a pile of rocks.", + "(a)", + "Figure 1. Example generation by stable-diffusion-v1-4 and our method. Stable diffusion model makes three types of errors that include missing objects, mismatched attributes, and mislocated objects. Errors are highlighted in red." + ], + "image_footnote": [], + "bbox": [ + 84, + 433, + 209, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6983621a8be527cf85e44ba0ddc03f077c13476df73ffa582cb5139481e6a13b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 433, + 336, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48b025c81218bcaa972c6a3a87b46de7d20497e2a3d478bf64277ca72be978a2.jpg", + "image_caption": [ + "Mismatched Attributes", + "StableDiffusion", + "A bedroom has green carpeting and a bed." + ], + "image_footnote": [], + "bbox": [ + 357, + 309, + 482, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2a7ab16c264f3ec31d1f9affa32f4f5c694a4050d6fc9e1fec25dd2f66a5b30.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 491, + 309, + 616, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad7aa174e9b3ef49d079b8b31cdad4f343d105baf8df8bc5a188d06f6ece4162.jpg", + "image_caption": [ + "A red car is to the left of a black mailbox.", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 357, + 433, + 482, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2728eb6bde0414eca02cdb5fea9c7104788bfadcfdd8bacd24ae4c2ff70bf104.jpg", + "image_caption": [ + "Mislocated Objects", + "StableDiffusion", + "The brown teddy bear is placed high above the toilet." + ], + "image_footnote": [], + "bbox": [ + 627, + 309, + 754, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09756c482440a87c372aefc049ea576bbb437960ff77ce730399549d4028f02d.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 761, + 309, + 885, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9bb0535bb90fae8f90f4bb47adbf09e186ab4c93955e63b4e46627d318567d4.jpg", + "image_caption": [ + "The teddy bear is right of the suitcase.", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 630, + 433, + 756, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35027d1451bd4998133b64cb67481385f68e5cd83536d0631ff49149869bf63f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 433, + 887, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 613, + 312, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based models have achieved state-of-the-art performance on text-to-image synthesis tasks. However, one critical limitation of these models is the low fidelity of generated images with respect to the text description, such as missing objects, mismatched attributes, and mislocated objects. One key reason for such inconsistencies is the inaccurate cross-attention to text in both the spatial dimension, which controls at what pixel region an object should appear, and the temporal dimension, which controls how different levels of details are added through the denoising steps. In this paper, we propose a new text-to-image algorithm that adds explicit control over spatial-temporal cross-attention in diffusion models. We first utilize a layout predictor to predict the pixel regions for objects mentioned in the text. We then impose spatial attention control by combining the attention over the entire text description and that over the", + "bbox": [ + 73, + 659, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "local description of the particular object in the corresponding pixel region of that object. The temporal attention control is further added by allowing the combination weights to change at each denoising step, and the combination weights are optimized to ensure high fidelity between the image and the text. Experiments show that our method generates images with higher fidelity compared to diffusion-model-based baselines without fine-tuning the diffusion model. Our code is publicly available.1", + "bbox": [ + 496, + 614, + 893, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 780, + 630, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [14, 19, 46, 47, 48] have recently revolutionized the field of image synthesis. Compared with previous generative models such as generative adversar", + "bbox": [ + 496, + 806, + 890, + 852 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 517, + 863, + 629, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/UCSB-NLP-Chang/", + "bbox": [ + 519, + 876, + 794, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Diffusion-SpaceTime-Attn", + "bbox": [ + 501, + 888, + 692, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7766", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ial networks [1, 3, 11, 18] and variational autoencoders [21, 38, 39], diffusion models have demonstrated superior performance in generating images with higher quality, more diversity, and better control over generated contents. Particularly, text-to-image diffusion models [2, 33, 36, 40, 42] allow generating images conditioned on a text description, which enables generation of creative images due to the expressiveness of natural language.", + "bbox": [ + 75, + 90, + 468, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, recent studies [10, 31] have revealed that one critical limitation of existing diffusion-model-based text-to-image algorithms is the low fidelity with respect to the text descriptions – the content of the generated image is sometimes at odds with the text description, especially when the description is complex. Specifically, typical errors made by stable diffusion models fall into three categories: missing objects, mismatched attributes, and mislocated objects. For example, in Fig. 1(a), stable diffusion model ignores the airplane even though it is mentioned in text; in Fig. 1(b), the model confuses “red car” and “black mailbox” and generates a red mailbox; in Fig. 1(c), the model locates teddy bear behind the toilet, despite the description “teddy bear is placed high above the toilet.”", + "bbox": [ + 75, + 214, + 470, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such infidelity problems suggest that the cross-attention map on the text description may not be accurate. In particular, if we view the generation process of a diffusion model as a sequence of denoising steps, then the cross-attention on text descriptions can be considered as a function of both spatial (pixels) and temporal (denoising steps) information. Therefore, the inaccuracies of the cross-attention can result from the loose control over both the spatial and temporal dimensions. On one hand, spatial attention controls at what pixels the model should attend to each object and the corresponding attributes mentioned in the text. If the spatial attention is incorrect, the resulting images will have incorrect object locations or miss-associated attributes. On the other hand, temporal attention controls when the models should attend to different levels of details in the text. As previous works have revealed, diffusion models tend to focus on generating object outlines at earlier denoising steps and on details at later [50]. Thus loose control over the temporal aspect of attention can easily lead to overlooking certain levels of the object details. In short, to improve the fidelity of text-to-image synthesis, one would need to explicitly control both spatial and temporal attention to follow an accurate and optimal distribution.", + "bbox": [ + 75, + 429, + 470, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new text-to-image algorithm based on a pre-trained conditional diffusion model with explicit control over the spatial-temporal cross-attention map on text. The proposed algorithm introduces a layout predictor and a spatial-temporal attention optimizer. The layout predictor takes the text description as input and generates a spatial layout for each object mentioned in the text. Alternatively, the layout can also be provided by the user. Then", + "bbox": [ + 75, + 780, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the spatial-temporal attention optimizer imposes direct control over the spatial and temporal aspects of the attention according to the spatial layout. In particular, for the spatial aspect, we parameterize the attention map such that the attention outputs in the designated pixel region for an object are a weighted combination of attention over the entire text description and that over the local description that specifically describes the corresponding object. In this way, we manage to emphasize the attention over the object descriptions. For the temporal aspect, we allow the combination weights to change across time and optimized according to a CLIP objective that measures the agreement between the generated images and the text description. In this way, we allow the attention to focus more on the entire description at the early stage and gradually shift to the detailed local descriptions as the denoising process proceeds. The entire pipeline resembles a typical painting process of a human, where each object's position is determined beforehand and the focus gradually shifts from global information to the local details of each object.", + "bbox": [ + 496, + 90, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on datasets that contain real and template-based captions [6, 26] and our newly created synthetic dataset that contains complex text descriptions. Results show that our method generates images that better align with descriptions compared to other stable diffusion-based baselines. As shown in Fig. 1, our method effectively resolves the above-mentioned three errors. Particularly, controlling spatial attention locates objects at the desired position, and controlling temporal attention promotes the occurrence of objects with associated attributes. Our findings shed light on fine-grained control of diffusion models in text-to-image generation tasks.", + "bbox": [ + 496, + 395, + 892, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 595, + 640, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models Diffusion models are a class of generative models that have demonstrated state-of-the-art performance on image synthesis tasks [8, 14, 19, 46, 47, 48]. These models synthesize images by sampling a noisy image from the standard Gaussian distribution and iteratively denoising it back to a clean image. Their impressive performance has advanced research in multiple computer vision areas including inpainting [30, 41, 51], image editing [7, 12, 32, 50], super-resolution [15, 43], video synthesis [13, 16, 55], and applications beyond computer vision [22, 23, 52]. Among these, text-to-image diffusion models have gained significant attention [2, 36, 37, 42, 58]. Taking text descriptions as inputs, these models generate high quality images that are semantically aligned with the input text, which have led to many creative and artistic applications.", + "bbox": [ + 496, + 622, + 890, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Enhancing the Controllability of Text-to-Image Diffusion Models While diffusion-based text-to-image models have shown promising results, recent works have high", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7767", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e7cfe7163e33d910370800bc7d71acf9a9738e94cdedcb0c9cbe2930f223a17f.jpg", + "image_caption": [ + "Figure 2. Overview of our text-to-image generation pipeline at one denoising step. Given input text $D$ , we first parse it and extract all objects mentioned, constructing local descriptions $L_{i}$ . Next, the layout predictor predicts the pixel region for each object in the text. The diffusion model attends to the global description $D$ and additionally attends to the local description $L_{i}$ in object $i$ 's region. The final attention output is a weighted combination of attention to both global and local descriptions, where the combination weights sum up to 1 for each pixel and are optimized for each denoising step to achieve a high fidelity with $D$ ." + ], + "image_footnote": [], + "bbox": [ + 160, + 89, + 807, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lighted cases where models fail to generate high-fidelity images with respect to the input text [10, 31]. To this end, prior works have explored various ways to enhance the controllability of text-to-image diffusion models. One line of work enhances the controllability by improving diffusion models' ability to understand natural language, which includes using a more powerful text encoder that is separately trained on language modeling tasks [2, 42], incorporating linguistic structures in the text to guide the cross-attention between image and text [10], and decomposing a complex text description into multiple components that are easy to generate [28]. Another line of work conditions diffusion models' generation on auxiliary inputs such as object layout [2, 9, 24, 54] and silhouette [17, 34, 45, 56, 57]. By modifying diffusion models' attention operation according to these auxiliary information or directly fine-tuning diffusion models to take these auxiliary inputs, they are able to control the location and shape of the objects in the image. Finally, some work adds temporal aspect control on diffusion models by modifying the input text condition at each denoising step, which allows them to disentangle a desired attribute from other contents [50]. Different from prior works, our method imposes both spatial and temporal control in cross-attention layer. Moreover, our method does not require auxiliary inputs and does not fine-tune the diffusion model.", + "bbox": [ + 75, + 385, + 472, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 777, + 210, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 801, + 279, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on the standard text-to-image problem. Given a text description, denoted as $D$ , our goal is to generate an image that is consistent with $D$ . For a concrete exposition, we will use an example description \"A red car is to the left of a black mailbox\" in the following. Our work aims to", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "improve the fidelity of the generated image to text, which includes three requirements:", + "bbox": [ + 498, + 385, + 890, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Object Fidelity: The generated images should include all the objects mentioned in $D$ . In our example, the generated image should contain a car and a mailbox.", + "- Attribute Fidelity: The attributes of each object in the image should match those in $D$ . In our example, the car should be red and the mailbox should be black.", + "- Spatial Fidelity: The relative spatial positions of the object should match the description in $D$ . In our example, the car should be on the left and the mailbox on the right." + ], + "bbox": [ + 501, + 426, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will tackle these problems using a pre-trained, fixed stable diffusion model.", + "bbox": [ + 498, + 585, + 890, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Method Overview", + "text_level": 1, + "bbox": [ + 500, + 626, + 673, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve high-fidelity text-to-image generation, we propose an algorithm consisting of the following four steps. Step 1: Object Identification We extract all the objects mentioned in $D$ , denoted as $O_{1:N}$ , by eliciting the noun phrases using spaCy. $^2 N$ is the total number of objects. In our sample, $O_1 = \\text{\"a red car\"}$ and $O_2 = \\text{\"a black mailbox\"}$ . Step 2: Layout Prediction For each object $O_i$ , we use a layout predictor to predict its pixel region, $\\mathcal{R}_i$ , which is a set of pixels roughly specifying where the object should be. The layout can also be provided by human users.", + "bbox": [ + 496, + 650, + 890, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 3: Local Description Generation For each object $O_{i}$ , we generate a local text description, $L_{i}$ , containing only that object information using a simple template. In our example, there are two local descriptions. $L_{1} =$ \"A photo of a red car\" and $L_{2} =$ \"A photo of a black mailbox\".", + "bbox": [ + 496, + 801, + 893, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2https://spacy.io/.", + "bbox": [ + 514, + 886, + 665, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7768", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 4: Attention Optimization During the generation process, we guide the diffusion model to combine attention to both the global description $D$ and the local descriptions $L_{1:N}$ according to the object layout. The attention combination weights are optimized for each input description.", + "bbox": [ + 76, + 90, + 467, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The following subsections will provide further details about steps 2 and 4.", + "bbox": [ + 76, + 167, + 467, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Layout Predictor", + "text_level": 1, + "bbox": [ + 76, + 205, + 243, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our layout predictor is adapted from [53], which aims to predict the center coordinate $C_i = [X_i, Y_i]$ for each object $O_i^3$ . Specifically, the layout predictor is a transformer that takes the text description $D$ as the input. At the output position where the input mentions the object $O_i$ , we let the transformer output a set of Gaussian mixture model (GMM) parameters to fit the object's center coordinate $C_i$ . Formally, denote $f_i(D; \\theta)$ as the output of the layout predictor at the location of the mentioning object $O_i$ . Then", + "bbox": [ + 76, + 228, + 467, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {i} (\\boldsymbol {D}; \\boldsymbol {\\theta}) = \\bigcup_ {k = 1} ^ {K} \\left\\{\\boldsymbol {\\mu} _ {i k}, \\boldsymbol {\\Sigma} _ {i k}, w _ {i k} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 372, + 467, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu}_{ik}$ , $\\pmb{\\Sigma}_{ik}$ , and $w_{ik}$ denote the mean, covariance matrix and the prior probability of mixture $k$ ; $\\pmb{\\theta}$ denotes the network parameters.", + "bbox": [ + 76, + 416, + 467, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the fidelity of the object positions in the predicted layout, we introduce a hybrid training objective including an absolute position objective and a relative position objective.", + "bbox": [ + 76, + 460, + 467, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Absolute Position Objective The absolute position objective provides direct supervision of the exact position of each object. Formally, we assume access to an image captioning dataset $\\mathcal{D}_{\\mathrm{real}}$ with description $D$ as well as the extracted labels of all the objects $\\{O_i\\}$ and their center coordinates $\\{C_i\\}$ . Then the training objective is to minimize the negative log-likelihood of the ground-truth coordinates under the predicted GMM distribution, i.e.", + "bbox": [ + 76, + 531, + 467, + 651 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a b s}} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {D}, \\left\\{\\boldsymbol {C} i \\right\\} \\sim \\mathcal {D} _ {\\mathrm {r e a l}}} \\left[ \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {n l l}} \\left(\\boldsymbol {C} _ {i}; \\boldsymbol {f} _ {i} (\\boldsymbol {D}; \\boldsymbol {\\theta})\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 659, + 467, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\ell_{\\mathrm{nll}}(C_i;f_i)$ denotes the negative log-likelihood of ground truth $C_i$ under the GMM specified by $f_{i}$ .", + "bbox": [ + 76, + 703, + 467, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relative Position Objective In many cases, the text description only mentions the relative positions of the objects, and thus it is more important to ensure the relative position of the predicted position is correct than the absolute position. To further enforce the fidelity of the relative positions, we introduce the following relative position objective.", + "bbox": [ + 76, + 742, + 467, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To start with, we construct a synthetic dataset, $\\mathcal{D}_{\\mathrm{syn}}$ which consists of text description $D$ with explicit descriptions of relative spatial relations. We first randomly select $N$", + "bbox": [ + 76, + 834, + 467, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "objects with attribute modifiers. We then select $M$ pairs of objects to specify their relative spatial relation. For object pair $(O_i, O_j)$ , their spatial relation, $R_{ij}$ , is randomly drawn from \"left of\", \"right of\", \"above\" and \"below\". Finally, we prompt GPT-3 [4] to generate the text description $D$ that mentions all the objects and their relative positions. Our familiar \"A red car is to the left of a black mailbox\" is one such example. Further details are provided in Appendix B.", + "bbox": [ + 500, + 90, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we introduce a loss to penalize violating the relative position. If object $i$ is to the left of object $j$ , we enforce that the rightmost mixture mean of object $i$ is to the left of the leftmost mixture mean of object $j$ . Formally,", + "bbox": [ + 498, + 212, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {r e l}} \\left(R _ {i j} = ^ {\\prime \\prime} \\text {l e f t} ^ {\\prime \\prime}; \\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right) = \\max \\left\\{\\max _ {k} \\boldsymbol {\\mu} _ {i k} (0) - \\min _ {k} \\boldsymbol {\\mu} _ {j k} (0), - \\delta \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 280, + 898, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{ik}(0)$ denotes the zeroth element (the $x$ -coordinate) of $\\mu_{ik}$ , and $\\delta$ is a pre-specified margin. The $\\ell_{\\mathrm{rel}}$ of the other three types of relations are defined similarly. The relative position loss is thus the aggregation of $\\ell_{\\mathrm{rel}}$ across the synthetic dataset:", + "bbox": [ + 498, + 314, + 890, + 387 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e l}} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {D}, \\left\\{R _ {i j} \\right\\} \\sim \\mathcal {D} _ {\\mathrm {s y n}}} \\left[ \\sum_ {i, j: R _ {i j} \\neq \\emptyset} \\ell_ {\\mathrm {r e l}} \\left(R _ {i j}; \\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 393, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{f}_i$ is short for $f_{i}(D;\\theta)$ .", + "bbox": [ + 500, + 435, + 699, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training and Inference To sum up, the final training objective is the combination of both:", + "bbox": [ + 498, + 459, + 888, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a y o u t}} (\\boldsymbol {\\theta}) = \\mathcal {L} _ {\\text {a b s}} (\\boldsymbol {\\theta}) + \\xi \\mathcal {L} _ {\\text {r e l}} (\\boldsymbol {\\theta}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 498, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\xi$ is a hyperparameter. During inference time, we randomly draw the center coordinate $C_i$ from the predicted GMM. The pixel region for object $O_i$ , $\\mathcal{R}_i$ , is defined as a circular region centered at the drawn $C_i$ with a fixed radius $r$ . As would be shown in Appendix D, $\\mathcal{R}_i$ only roughly regulates the position of the generated objects, and the actual size of the object can go beyond or below the size of $\\mathcal{R}_i$ . Thus a fixed $r$ is sufficient for this purpose.", + "bbox": [ + 498, + 523, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Spatial-Temporal Attention Optimization", + "text_level": 1, + "bbox": [ + 500, + 652, + 852, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that $D$ is the global description and $\\{L_i\\}$ are local descriptions for each object. Our goal is to guide the diffusion model to attend to not only $D$ , but also $L_{i}$ in object $i$ ’s region, $\\mathcal{R}_i$ , so that the model is more strongly prompted to generate the specific object as specified in the layout.", + "bbox": [ + 498, + 676, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial-Temporal Attention Recall that the standard cross-attention [49] is defined as", + "bbox": [ + 498, + 761, + 888, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {Q} \\boldsymbol {K} ^ {T} / \\sqrt {d}\\right) \\boldsymbol {V}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 797, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q \\in \\mathbb{R}^{h \\times w, d}$ is the queries vectors for each pixel, and $K, V \\in \\mathbb{R}^{l, d}$ are key and value vectors for each text description token; $h$ and $w$ represent the image height and width, $l$ represents the text description length; $d$ represents the dimension of each attention vector.", + "bbox": [ + 498, + 824, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3More details of the layout predictor can be found in Appendix D.", + "bbox": [ + 94, + 886, + 442, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7769", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we have multiple text descriptions, we define $K_{D}, V_{D}$ as the key and value vectors for the global description $D$ , and $K_{L_i}, V_{L_i}$ for the local description $L_i$ . Further, we introduce a set of binary mask matrices $\\{M_i\\}$ to indicate the region for each object, i.e.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {i} (x, y) = 1, \\text {i f} (x, y) \\in \\mathcal {R} _ {i}, \\quad \\text {a n d} 0, \\text {o t h e r w i s e .} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 172, + 468, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the output of the attention layer of the denoising network at time $t$ is defined as", + "bbox": [ + 76, + 194, + 468, + 223 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {O} (t) = \\sum_ {i = 1} ^ {N} \\lambda_ {i t} \\boldsymbol {M} _ {i} \\odot \\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K} _ {\\boldsymbol {L} _ {i}}, \\boldsymbol {V} _ {\\boldsymbol {L} _ {i}}) \\tag {8} \\\\ + \\left(1 - \\sum_ {i = 1} ^ {N} \\lambda_ {i t} M _ {i}\\right) \\odot \\operatorname {A t t e n t i o n} \\left(\\boldsymbol {Q}, \\boldsymbol {K} _ {D}, \\boldsymbol {V} _ {D}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 227, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{it}$ are the attention combination weights, and $\\odot$ denotes element-wise multiplication. Note that the combination weights are functions of time $t$ , which is motivated by the observation that different denoising steps control the generation of different levels of details. When synthesizing images, the combination weights $\\lambda_{it}$ are optimized on the fly for each corresponding mask and time. By introducing this time dependency, we allow the diffusion model to focus more on the global description at earlier time steps and shift to local descriptions later.", + "bbox": [ + 75, + 305, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization Objective All the attention combination weights, $\\lambda = \\{\\lambda_{it}\\}$ , are determined by maximizing the consistency between the generated image and the text description as measured by the CLIP similarity [35]. We introduce two CLIP similarities, a global CLIP similarity that compares the entire image and the global description, and a set of local CLIP similarities that compare the images at each object region and the corresponding local descriptions. Formally, denote the generated image as $I(\\lambda)$ , and its local patch at each object's region as $I_{O_i}(\\lambda)$ . Note that we adopt the deterministic PLMS denoising process [27] (i.e., with $\\sigma = 0$ ), so both $I$ and $I_{O_i}$ are deterministic functions of $\\lambda$ . Then the loss for optimizing $\\lambda$ is given by", + "bbox": [ + 76, + 465, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a t t e n d}} (\\boldsymbol {\\lambda}) = - \\operatorname {C L I P} (\\boldsymbol {I} (\\boldsymbol {\\lambda}), \\boldsymbol {D}) - \\gamma \\sum_ {i = 1} ^ {N} \\operatorname {C L I P} \\left(\\boldsymbol {I} _ {O _ {i}} (\\boldsymbol {\\lambda}), \\boldsymbol {L} _ {i}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 666, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{CLIP}(\\cdot)$ denotes the CLIP similarity, and $\\gamma$ is a hyperparameter.", + "bbox": [ + 76, + 708, + 468, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 750, + 207, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments to evaluate our method's performance and generalizability. We also perform ablation study on important design choices of our method.", + "bbox": [ + 76, + 776, + 468, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementing Details: We adopt RoBERTa-base [29] as the base model for layout predictor and use 5 mixtures in", + "bbox": [ + 76, + 825, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GMM. We use stable-diffusion-v1-4 [40] pre-trained on the laion dataset [44] and freeze it throughout all experiments. All generated images are in the size of $512 \\times 512$ . We use PLMS sampler [27] to synthesize images with 50 denoising steps, and we use Adam [20] to optimize the layout predictor and attention combination weights. More details on hyperparameters and optimization are in Appendix A.", + "bbox": [ + 500, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation on Fidelity of Generated Images", + "text_level": 1, + "bbox": [ + 500, + 204, + 870, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate our method on object, attribute, and spatial fidelities as introduced in Sec. 3.1 using both objective and subjective metrics.", + "bbox": [ + 498, + 227, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines: We identify four baseline methods on generating images from complex text descriptions. (1) VANILLA-SD [40] is the pre-trained text-to-image stable diffusion model that directly generates images conditioned on the text description. (2) COMPOSABLE-DIFFUSION [28] is a diffusion-based compositional generation method. To generate images from a text description, the text is first decomposed into the conjunction of multiple components (e.g., for the example in Sec. 3.1, the components are \"A red car is to the left of a black mailbox.\" AND \"A red car\" AND \"a black mailbox\"). Each component is separately modeled by the diffusion model and then composed to generate an image by merging the outputs of each denoising step. (3) STRUCTURE-DIFFUSION [10] improves VANILLA-SD on generating images with better object and attribute fidelities. They do so by first extracting noun phrases at different levels of the parsing tree. The model then separately attends to each noun phrase and combines the attention outputs by taking their arithmetic mean. Different from our method, they perform cross-attention on the whole image instead of the region specific to an object. (4) PAINT-WITH-WORDS [2] assumes that users provide the pixel region of each object to be generated. Pixels in the region then increase their attention weights to the text that describes the corresponding object, and the amount of increase is determined by heuristic rules. For a fair comparison, we use the pixel region predicted by our layout predictor, and we report the performance with the ground truth region in Appendix E.", + "bbox": [ + 498, + 277, + 892, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics: We conduct experiments on three datasets. (1) MS-COCO [25] contains photos taken by photographers and manually annotates the caption for each photo. We use the caption as the input text description. (2) VSR [26] is proposed for probing spatial understanding of vision-language models. Constructed from a subset of MS-COCO, it uses templates to generate captions that describe spatial relations in the image (e.g., \"The horse is to the left of the person\"). (3) GPT-synthetic is the synthetic dataset introduced in Sec. 3.3. We manually check the test set to filter out sentences that do not conform to the specified spatial relation. Compared with VSR, this dataset has more diverse and complex text descriptions. The final dataset contains", + "bbox": [ + 498, + 704, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4To standardize the image patch sizes, $I_{O_i}$ is obtained by cropping the original image with a minimum square that encompasses $\\mathcal{R}_i$ and resizing it to $224 \\times 224$ .", + "bbox": [ + 75, + 862, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7770", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/56f5840145b4d67fdc5c11200270a48d2bf3b687fdcbbac4fe1227427878f34c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Subjective evaluation of our method and baselines. Best numbers are in bold. Spatial relation is not available on MS-COCO because very few of its captions contain spatial relations. Attribute is not available on VSR as its captions do not consider attribute. Object, Attribute, and Spatial show the total score of 50 evaluations, where a model with the highest fidelity would achieve a score of 100. Overall denotes the percentage of generation of each method that is rated better than our method." + ], + "table_body": "
MS-COCOVSRGPT-synthetic
Object (100) ↑Attribute (100) ↑OverallObject (100) ↑Spatial (100) ↑OverallObject (100) ↑Attribute (100) ↑Spatial (100) ↑Overall
VANILLA-SD666242%646648%60586038%
COMPOSABLE-DIFFUSION534830%636018%63604728%
STRUCTURE-DIFFUSION606348%546432%58545832%
PAINT-WITH-WORDS757452%545930%65525732%
Ours7574-6877-848386-
", + "bbox": [ + 76, + 88, + 890, + 170 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cbece2942a7e37d5563fb38220495ccc184568324541b5cfa45eada2f83afdd5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MS-COCOVSRGPT-synthetic
Object RecallSPRel PrecisionObject RecallSPRel PrecisionObject RecallSPRel Precision
VANILLA-SD58.0%-62.1%56.0%42.4%56.8%
COMPOSABLE-DIFFUSION51.8%-60.8%72.2%34.4%52.6%
STRUCTURE-DIFFUSION61.7%-62.3%58.3%43.0%59.3%
PAINT-WITH-WORDS57.3%-63.8%47.1%45.9%53.5%
Ours69.6%-65.1%75.0%47.2%66.7%
", + "bbox": [ + 76, + 237, + 467, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Automatic evaluation of our method and baselines.", + "bbox": [ + 75, + 324, + 467, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SPRel Precision: Spatial Relation Precision. Best numbers are in bold. SPRel precision is not available on MS-COCO since most captions do not have an explicit spatial relation.", + "bbox": [ + 75, + 337, + 467, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "500 descriptions, and we downsample MS-COCO and VSR to have the same size. Statistics including the number of objects and spatial relations are shown in Appendix B.", + "bbox": [ + 75, + 396, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For automatic evaluations, we consider two metrics. (1) Object Recall measures the percentage of successfully synthesized objects over objects mentioned in the text. We use DETR [5] to detect objects in generated images. To calculate recall, we divide the number of detected objects in the text by the total number of objects in the text that also belong to one of the MS-COCO categories. It measures the object fidelity of generated images. (2) Spatial Relation Precision (SPRel Precision) measures the percentage of the correct spatial relations among all the relations whose corresponding objects are successfully synthesized. This metric measures the spatial fidelity of generated images. We consider the relation of left, right, above, and below because their correctness can be evaluated by comparing the bounding box centers. Qualitative examples of spatial relations beyond these four are shown in Appendix G. We also report the CLIP similarities between generated images and the input description in Appendix C. We observe that different methods have very close CLIP similarities, though our method still achieves competitive results.", + "bbox": [ + 73, + 441, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: The results are shown in Table 2. As shown in the table, our method outperforms all baselines on both object recall and SPRel precision. Specifically, compared with baselines that do not have spatial attention control (all but PAINT-WITH-WORDS), our method is significantly better on SPRel precision, which illustrates the effectiveness of spatial control and indicates that it is difficult for the stable diffusion model to generate correct spatial relations without guidance on the layout. Our method moves the burden of generating objects at the correct location from diffusion", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models to a separate layout predictor, allowing the whole pipeline to achieve better spatial fidelity. On the other hand, compared with PAINT-WITH-WORDS that does not impose the temporal attention control, our method achieves better object recall, showing that our optimized combination weights across the temporal dimension strike a better balance between local and global descriptions.", + "bbox": [ + 496, + 239, + 890, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 demonstrates examples of text descriptions and images generated by our method and baselines. We observe that our method resolves three types of errors in baselines. First, our method alleviates the missing object issue (in the top panel). Baselines tend to focus on one object in the text and ignore other objects (e.g., focusing on the elephant and ignoring the man in the second row), whereas our method generates all objects. Second, as shown in the middle panel, our method mitigates the mismatched attribute issue. Particularly, baselines struggle when multiple objects are mentioned in the text, where they mismatch the attribute and object (e.g., red truck in the first row). Finally, the bottom panel shows examples where our method reduces the number of mislocated objects. Note that our method is effective both on the four relations and other spatial relations. More examples can be found in Appendix G.", + "bbox": [ + 496, + 349, + 892, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subjective Evaluation: To further evaluate the fidelity and quality of the generated images, we perform a subjective evaluation on Amazon Mechanical Turk. Specifically, we randomly sample 25 text descriptions for each dataset in Table 2. Each subject was presented with the text description and corresponding image generated by a single method or a pair of methods and asked the following four questions: (1) (Object Fidelity) Does the image contain all objects mentioned in the text? (2) (Attribute Fidelity) Are all synthesized objects consistent with their characteristics described in the text (e.g., color and material)? (3) (Spatial Fidelity) Does the image locate all objects at the correct position such that the spatial relations in the text are satisfied (if an object in the relationship is missing, it is considered as an incorrect generation)? and (4) (Overall) Which image in the pair has higher fidelity with the text and has better quality? The first three questions are evaluated for each method individually with a score of 0, 1, or 2, where 2 denotes all objects/attributes/relations are correct and 0 denotes none of them is correct. The last question is evaluated on a pair of", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7771", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c5424cc3c010909ff33ff2d09b9ab7c0a0163540720ed4d36de47b75f320f20.jpg", + "image_caption": [ + "Figure 3. Example images generated by our method and baselines. Typical errors of baselines include missing objects, mismatched attributes, and mislocated objects. Ours (1)/(2) show the results with two different random seeds." + ], + "image_footnote": [], + "bbox": [ + 107, + 84, + 851, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images, one generated by our method and the other by a baseline. Each generated image is evaluated by two subjects, so the total score of the first three questions is 100. More details are in Appendix H.", + "bbox": [ + 75, + 787, + 470, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 shows the results. Our method achieves significantly better performance in most cases, especially on GPT-synthetic dataset that contains multiple objects and relations", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in the same description. The results show that our method is effective at generating images with high fidelity without sacrificing perceptual quality.", + "bbox": [ + 498, + 787, + 890, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Additional Analyses", + "text_level": 1, + "bbox": [ + 500, + 845, + 692, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on Each Text Complexity Level: To explore the capability of our method, especially on complex text, we", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7772", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d796dae3489d481c1571edf8944d9d15c507bbfbaed7dcf179d87fb093077814.jpg", + "image_caption": [ + "Figure 4. Performance when the number of objects and spatial relations in the text increase." + ], + "image_footnote": [], + "bbox": [ + 86, + 85, + 274, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35af5a19de2abd13e113829de562d61d465f6a792dc6f017d5ffd6384fae1ee7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 85, + 454, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1a802fc7ed665cc7b166206e6631b9eb9c90dac8fd04de56572c1db3fb18305.jpg", + "image_caption": [ + "Figure 5. Example images generated by our method and baselines on novel object combinations." + ], + "image_footnote": [], + "bbox": [ + 98, + 253, + 459, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "analyze the performance at each complexity level. Specifically, we use the GPT-synthetic dataset where multiple objects and relations can appear in the same description. We use the number of objects $N$ and number of spatial relations $M$ introduced in Sec. 3.3 as proxies for text complexity and plot the performance of each method as the text becomes more complex. i.e., we plot the performance for each value of $N$ regardless of $M$ , and vise versa for $M$ . As shown in Fig. 4, the performance of all methods decreases as the text becomes more complex, but our method still outperforms or is on par with others on complex descriptions. Note that PAINT-WITH-WORDS uses the same layout predictor as ours, which partially explains its strong performance.", + "bbox": [ + 75, + 464, + 468, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalizability to novel object combinations is a critical requirement for text-to-image models. However, since MSCOCO dataset is collected from real photos, most of its captions contain object pairs and object-attribute pairs that are common in daily life, which are also more likely to overlap with stable diffusion's pre-training data. Thus we additionally evaluate our method on another synthetic dataset that is created similarly to Sec. 3.3, but contains uncommon object pairs, object-attribute pairs, and object spatial relations. Fig. 5 shows sample images generated by our method and baselines. It can be observed that our method successfully generates novel object pairs (e.g., \"zebra is looking at computer\"). Performance on this dataset and more generated examples are shown in Appendix F. Overall, our method is able to generalize to novel object combinations.", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c79f6b7e36f29fddee805357b82d48563123a16c9446df65c2dce26c562e5073.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Object RecallSPRel Precision
Ours47.2%66.7%
No spatial control39.6%51.7%
No temporal control43.8%61.4%
No optimization41.0%55.6%
", + "bbox": [ + 529, + 88, + 859, + 169 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation performance on GPT-synthetic test set.", + "bbox": [ + 527, + 170, + 864, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 199, + 651, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we will investigate the influence of two important steps in our method, namely, the layout predictor and the spatial-temporal attention. We first study how spatial-temporal attention affects our performance. To do that, we consider three variants of attention.", + "bbox": [ + 496, + 223, + 890, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we explore the attention without spatial control. Concretely, the diffusion model still attends to both global description and local descriptions. However, instead of only attending to a local description in the pixel region of the object, the model now attends to all local descriptions in the whole image. The combination weights at each denoising step are optimized with only the global CLIP similarity in Sec. 3.4. Table 3 shows that its performance drops drastically, demonstrating the importance of spatial control.", + "bbox": [ + 496, + 299, + 892, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Second, we explore the attention without temporal control, where combination weights remain the same for all denoising steps. Note that the weights can be different for each object and are optimized with $\\mathcal{L}_{\\mathrm{attend}}(\\lambda)$ in Sec. 3.4. The results in Table 3 show a significant degradation under both metrics, indicating the benefits of temporal dependency on attention.", + "bbox": [ + 496, + 436, + 890, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we explore the attention without optimization. Specifically, we fix the combination weight $\\lambda_{it} = \\frac{1}{N}$ for all $i$ and $t$ , where $N$ is the number of objects. This value is also the initialization point of $\\{\\lambda_{it}\\}$ in our method. As shown in Table 3, the performance further drops compared to the no temporal control variant. The results indicate that optimizing the combination weights for a new text description is critical for generating images with high fidelity.", + "bbox": [ + 496, + 542, + 890, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further study the effects of our layout predictor, which includes the layout predictor trained with only one of the absolute and relative position objectives, the comparison with ground truth and user-provided pixel region, and a different strategy to construct the pixel region for an object. Results of these experiments are presented in Appendix E.", + "bbox": [ + 496, + 662, + 890, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 768, + 617, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we study the text-to-image synthesis task based on diffusion models. We find existing methods lack explicit control on cross-attention in diffusion models, which leads to the generation of low-fidelity images. We propose an algorithm that imposes control on cross-attention in spatial and temporal aspects. Experiments show our method outperforms baselines in generating high", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7773", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fidelity images. Further ablation study verifies the effectiveness of our spatial and temporal attention control. One limitation of our method is the reliance on a time-consuming optimization scheme, which takes around 10 minutes for each text-to-image generation. Future work may consider striking a better balance between performance and efficiency.", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 196, + 256, + 213 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work of Qiucheng Wu, Yujuan Liu, and Shiyu Chang was partially supported by Cisco Faculty Award.", + "bbox": [ + 76, + 220, + 468, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 266, + 174, + 282 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein gan, 2017. 2", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv, 2022. 2, 3, 5", + "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis, 2018. 2", + "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 4, 12", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers, 2020. 6", + "[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 2", + "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2", + "[9] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 3", + "[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv, 2022. 2, 3, 5" + ], + "bbox": [ + 78, + 291, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Neurips, 2014. 2", + "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Neurips, 2020. 1, 2", + "[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., pages 1-33, 2022. 2", + "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2", + "[17] Nisha Huang, Fan Tang, Weiming Dong, Tong-Yee Lee, and Changsheng Xu. Region-aware diffusion for zero-shot text-driven image editing. arXiv preprint arXiv:2302.11797, 2023. 3", + "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2017. 2", + "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models, 2022. 1, 2", + "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2014. 5, 13", + "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2", + "[23] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. arXiv preprint arXiv:2205.14217, 2022. 2", + "[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 3", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 12", + "[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. arXiv preprint arXiv:2205.00363, 2022. 2, 5", + "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7774", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 3, 5", + "[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv, 2019. 5, 12, 13", + "[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2", + "[31] Gary Marcus, Ernest Davis, and Scott Aaronson. A very preliminary analysis of dall-e 2, 2022. 2, 3", + "[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2", + "[34] Dong Huk Park, Grace Luo, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. arXiv preprint arXiv:2212.00210, 2022. 3", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 12, 13", + "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[38] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Neurips, 2019. 2", + "[39] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 12, 13", + "[41] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2", + "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3", + "[43] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 5", + "[45] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 3", + "[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning. PMLR, 2015. 1, 2", + "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2", + "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4", + "[50] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 2, 3", + "[51] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022. 2", + "[52] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: a geometric diffusion model for molecular conformation generation, 2022. 2", + "[53] Cheng-Fu Yang, Wan-Cyuan Fan, Fu-En Yang, and Yu-Chiang Frank Wang. Layouttransformer: Scene layout generation with conceptual and spatial diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3732-3741, June 2021. 4", + "[54] Zuopeng Yang, Daqing Liu, Chaoyue Wang, Jie Yang, and Dacheng Tao. Modeling image composition for complex scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7764-7773, 2022. 3", + "[55] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7775", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collomosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. arXiv preprint arXiv:2211.11742, 2022. 3", + "[57] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3", + "[58] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 247 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7776", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_model.json b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f37ac8915ffa0ee4db28dc1c0e5e78a32e4ca1c6 --- /dev/null +++ b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.13, + 0.892, + 0.177 + ], + "angle": 0, + "content": "Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.203, + 0.873, + 0.239 + ], + "angle": 0, + "content": "Qiucheng Wu\\(^{1*}\\), Yujuan Liu\\(^{1*}\\), Handong Zhao\\(^{2}\\), Trung Bui\\(^{2}\\), Zhe Lin\\(^{2}\\), Yang Zhang\\(^{3}\\), Shiyu Chang\\(^{1}\\), UC Santa Barbara, \\(^{2}\\)Adobe Research, \\(^{3}\\)MIT-IBM Watson AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.242, + 0.618, + 0.257 + ], + "angle": 0, + "content": "{qiucheng, yujianliu}@ucsb.edu" + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.275, + 0.28, + 0.293 + ], + "angle": 0, + "content": "Missing Objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.296, + 0.198, + 0.308 + ], + "angle": 0, + "content": "StableDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.31, + 0.21, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.296, + 0.293, + 0.308 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.31, + 0.338, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.408, + 0.301, + 0.433 + ], + "angle": 0, + "content": "An airport employee giving directions to an airplane." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.434, + 0.21, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.434, + 0.338, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.532, + 0.316, + 0.558 + ], + "angle": 0, + "content": "A zebra brazing on green grass next to a pile of rocks." + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.561, + 0.22, + 0.574 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.275, + 0.582, + 0.291 + ], + "angle": 0, + "content": "Mismatched Attributes" + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.296, + 0.472, + 0.308 + ], + "angle": 0, + "content": "StableDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.31, + 0.483, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.296, + 0.566, + 0.308 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.31, + 0.617, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.407, + 0.589, + 0.433 + ], + "angle": 0, + "content": "A bedroom has green carpeting and a bed." + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.434, + 0.483, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.531, + 0.574, + 0.558 + ], + "angle": 0, + "content": "A red car is to the left of a black mailbox." + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.561, + 0.496, + 0.574 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.275, + 0.838, + 0.293 + ], + "angle": 0, + "content": "Mislocated Objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.296, + 0.748, + 0.308 + ], + "angle": 0, + "content": "StableDiffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.296, + 0.844, + 0.308 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.31, + 0.756, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.31, + 0.887, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.407, + 0.865, + 0.433 + ], + "angle": 0, + "content": "The brown teddy bear is placed high above the toilet." + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.434, + 0.758, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.434, + 0.888, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.531, + 0.857, + 0.557 + ], + "angle": 0, + "content": "The teddy bear is right of the suitcase." + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.561, + 0.767, + 0.574 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.575, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Figure 1. Example generation by stable-diffusion-v1-4 and our method. Stable diffusion model makes three types of errors that include missing objects, mismatched attributes, and mislocated objects. Errors are highlighted in red." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.614, + 0.313, + 0.629 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Diffusion-based models have achieved state-of-the-art performance on text-to-image synthesis tasks. However, one critical limitation of these models is the low fidelity of generated images with respect to the text description, such as missing objects, mismatched attributes, and mislocated objects. One key reason for such inconsistencies is the inaccurate cross-attention to text in both the spatial dimension, which controls at what pixel region an object should appear, and the temporal dimension, which controls how different levels of details are added through the denoising steps. In this paper, we propose a new text-to-image algorithm that adds explicit control over spatial-temporal cross-attention in diffusion models. We first utilize a layout predictor to predict the pixel regions for objects mentioned in the text. We then impose spatial attention control by combining the attention over the entire text description and that over the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.894, + 0.751 + ], + "angle": 0, + "content": "local description of the particular object in the corresponding pixel region of that object. The temporal attention control is further added by allowing the combination weights to change at each denoising step, and the combination weights are optimized to ensure high fidelity between the image and the text. Experiments show that our method generates images with higher fidelity compared to diffusion-model-based baselines without fine-tuning the diffusion model. Our code is publicly available.1" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.781, + 0.631, + 0.797 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Diffusion models [14, 19, 46, 47, 48] have recently revolutionized the field of image synthesis. Compared with previous generative models such as generative adversar" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.864, + 0.63, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.521, + 0.877, + 0.795, + 0.888 + ], + "angle": 0, + "content": "1https://github.com/UCSB-NLP-Chang/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.502, + 0.889, + 0.694, + 0.9 + ], + "angle": 0, + "content": "Diffusion-SpaceTime-Attn" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.864, + 0.795, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7766" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.213 + ], + "angle": 0, + "content": "ial networks [1, 3, 11, 18] and variational autoencoders [21, 38, 39], diffusion models have demonstrated superior performance in generating images with higher quality, more diversity, and better control over generated contents. Particularly, text-to-image diffusion models [2, 33, 36, 40, 42] allow generating images conditioned on a text description, which enables generation of creative images due to the expressiveness of natural language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.426 + ], + "angle": 0, + "content": "However, recent studies [10, 31] have revealed that one critical limitation of existing diffusion-model-based text-to-image algorithms is the low fidelity with respect to the text descriptions – the content of the generated image is sometimes at odds with the text description, especially when the description is complex. Specifically, typical errors made by stable diffusion models fall into three categories: missing objects, mismatched attributes, and mislocated objects. For example, in Fig. 1(a), stable diffusion model ignores the airplane even though it is mentioned in text; in Fig. 1(b), the model confuses “red car” and “black mailbox” and generates a red mailbox; in Fig. 1(c), the model locates teddy bear behind the toilet, despite the description “teddy bear is placed high above the toilet.”" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Such infidelity problems suggest that the cross-attention map on the text description may not be accurate. In particular, if we view the generation process of a diffusion model as a sequence of denoising steps, then the cross-attention on text descriptions can be considered as a function of both spatial (pixels) and temporal (denoising steps) information. Therefore, the inaccuracies of the cross-attention can result from the loose control over both the spatial and temporal dimensions. On one hand, spatial attention controls at what pixels the model should attend to each object and the corresponding attributes mentioned in the text. If the spatial attention is incorrect, the resulting images will have incorrect object locations or miss-associated attributes. On the other hand, temporal attention controls when the models should attend to different levels of details in the text. As previous works have revealed, diffusion models tend to focus on generating object outlines at earlier denoising steps and on details at later [50]. Thus loose control over the temporal aspect of attention can easily lead to overlooking certain levels of the object details. In short, to improve the fidelity of text-to-image synthesis, one would need to explicitly control both spatial and temporal attention to follow an accurate and optimal distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a new text-to-image algorithm based on a pre-trained conditional diffusion model with explicit control over the spatial-temporal cross-attention map on text. The proposed algorithm introduces a layout predictor and a spatial-temporal attention optimizer. The layout predictor takes the text description as input and generates a spatial layout for each object mentioned in the text. Alternatively, the layout can also be provided by the user. Then" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.394 + ], + "angle": 0, + "content": "the spatial-temporal attention optimizer imposes direct control over the spatial and temporal aspects of the attention according to the spatial layout. In particular, for the spatial aspect, we parameterize the attention map such that the attention outputs in the designated pixel region for an object are a weighted combination of attention over the entire text description and that over the local description that specifically describes the corresponding object. In this way, we manage to emphasize the attention over the object descriptions. For the temporal aspect, we allow the combination weights to change across time and optimized according to a CLIP objective that measures the agreement between the generated images and the text description. In this way, we allow the attention to focus more on the entire description at the early stage and gradually shift to the detailed local descriptions as the denoising process proceeds. The entire pipeline resembles a typical painting process of a human, where each object's position is determined beforehand and the focus gradually shifts from global information to the local details of each object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.893, + 0.577 + ], + "angle": 0, + "content": "We conduct extensive experiments on datasets that contain real and template-based captions [6, 26] and our newly created synthetic dataset that contains complex text descriptions. Results show that our method generates images that better align with descriptions compared to other stable diffusion-based baselines. As shown in Fig. 1, our method effectively resolves the above-mentioned three errors. Particularly, controlling spatial attention locates objects at the desired position, and controlling temporal attention promotes the occurrence of objects with associated attributes. Our findings shed light on fine-grained control of diffusion models in text-to-image generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.642, + 0.612 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.85 + ], + "angle": 0, + "content": "Diffusion Models Diffusion models are a class of generative models that have demonstrated state-of-the-art performance on image synthesis tasks [8, 14, 19, 46, 47, 48]. These models synthesize images by sampling a noisy image from the standard Gaussian distribution and iteratively denoising it back to a clean image. Their impressive performance has advanced research in multiple computer vision areas including inpainting [30, 41, 51], image editing [7, 12, 32, 50], super-resolution [15, 43], video synthesis [13, 16, 55], and applications beyond computer vision [22, 23, 52]. Among these, text-to-image diffusion models have gained significant attention [2, 36, 37, 42, 58]. Taking text descriptions as inputs, these models generate high quality images that are semantically aligned with the input text, which have led to many creative and artistic applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Enhancing the Controllability of Text-to-Image Diffusion Models While diffusion-based text-to-image models have shown promising results, recent works have high" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7767" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.09, + 0.808, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Overview of our text-to-image generation pipeline at one denoising step. Given input text \\( D \\), we first parse it and extract all objects mentioned, constructing local descriptions \\( L_{i} \\). Next, the layout predictor predicts the pixel region for each object in the text. The diffusion model attends to the global description \\( D \\) and additionally attends to the local description \\( L_{i} \\) in object \\( i \\)'s region. The final attention output is a weighted combination of attention to both global and local descriptions, where the combination weights sum up to 1 for each pixel and are optimized for each denoising step to achieve a high fidelity with \\( D \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.473, + 0.765 + ], + "angle": 0, + "content": "lighted cases where models fail to generate high-fidelity images with respect to the input text [10, 31]. To this end, prior works have explored various ways to enhance the controllability of text-to-image diffusion models. One line of work enhances the controllability by improving diffusion models' ability to understand natural language, which includes using a more powerful text encoder that is separately trained on language modeling tasks [2, 42], incorporating linguistic structures in the text to guide the cross-attention between image and text [10], and decomposing a complex text description into multiple components that are easy to generate [28]. Another line of work conditions diffusion models' generation on auxiliary inputs such as object layout [2, 9, 24, 54] and silhouette [17, 34, 45, 56, 57]. By modifying diffusion models' attention operation according to these auxiliary information or directly fine-tuning diffusion models to take these auxiliary inputs, they are able to control the location and shape of the objects in the image. Finally, some work adds temporal aspect control on diffusion models by modifying the input text condition at each denoising step, which allows them to disentangle a desired attribute from other contents [50]. Different from prior works, our method imposes both spatial and temporal control in cross-attention layer. Moreover, our method does not require auxiliary inputs and does not fine-tune the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.778, + 0.212, + 0.795 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.281, + 0.817 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We focus on the standard text-to-image problem. Given a text description, denoted as \\( D \\), our goal is to generate an image that is consistent with \\( D \\). For a concrete exposition, we will use an example description \"A red car is to the left of a black mailbox\" in the following. Our work aims to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.386, + 0.892, + 0.418 + ], + "angle": 0, + "content": "improve the fidelity of the generated image to text, which includes three requirements:" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.472 + ], + "angle": 0, + "content": "- Object Fidelity: The generated images should include all the objects mentioned in \\( D \\). In our example, the generated image should contain a car and a mailbox." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.479, + 0.892, + 0.524 + ], + "angle": 0, + "content": "- Attribute Fidelity: The attributes of each object in the image should match those in \\( D \\). In our example, the car should be red and the mailbox should be black." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.577 + ], + "angle": 0, + "content": "- Spatial Fidelity: The relative spatial positions of the object should match the description in \\(D\\). In our example, the car should be on the left and the mailbox on the right." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.587, + 0.891, + 0.617 + ], + "angle": 0, + "content": "We will tackle these problems using a pre-trained, fixed stable diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.674, + 0.642 + ], + "angle": 0, + "content": "3.2. Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.801 + ], + "angle": 0, + "content": "To achieve high-fidelity text-to-image generation, we propose an algorithm consisting of the following four steps. Step 1: Object Identification We extract all the objects mentioned in \\(D\\), denoted as \\(O_{1:N}\\), by eliciting the noun phrases using spaCy. \\(^2 N\\) is the total number of objects. In our sample, \\(O_1 = \\text{\"a red car\"}\\) and \\(O_2 = \\text{\"a black mailbox\"}\\). Step 2: Layout Prediction For each object \\(O_i\\), we use a layout predictor to predict its pixel region, \\(\\mathcal{R}_i\\), which is a set of pixels roughly specifying where the object should be. The layout can also be provided by human users." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.802, + 0.894, + 0.879 + ], + "angle": 0, + "content": "Step 3: Local Description Generation For each object \\( O_{i} \\), we generate a local text description, \\( L_{i} \\), containing only that object information using a simple template. In our example, there are two local descriptions. \\( L_{1} = \\) \"A photo of a red car\" and \\( L_{2} = \\) \"A photo of a black mailbox\"." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.666, + 0.901 + ], + "angle": 0, + "content": "2https://spacy.io/." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7768" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "Step 4: Attention Optimization During the generation process, we guide the diffusion model to combine attention to both the global description \\( D \\) and the local descriptions \\( L_{1:N} \\) according to the object layout. The attention combination weights are optimized for each input description." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.168, + 0.468, + 0.197 + ], + "angle": 0, + "content": "The following subsections will provide further details about steps 2 and 4." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.206, + 0.245, + 0.221 + ], + "angle": 0, + "content": "3.3. Layout Predictor" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.229, + 0.468, + 0.365 + ], + "angle": 0, + "content": "Our layout predictor is adapted from [53], which aims to predict the center coordinate \\( C_i = [X_i, Y_i] \\) for each object \\( O_i^3 \\). Specifically, the layout predictor is a transformer that takes the text description \\( D \\) as the input. At the output position where the input mentions the object \\( O_i \\), we let the transformer output a set of Gaussian mixture model (GMM) parameters to fit the object's center coordinate \\( C_i \\). Formally, denote \\( f_i(D; \\theta) \\) as the output of the layout predictor at the location of the mentioning object \\( O_i \\). Then" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.373, + 0.468, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {i} (\\boldsymbol {D}; \\boldsymbol {\\theta}) = \\bigcup_ {k = 1} ^ {K} \\left\\{\\boldsymbol {\\mu} _ {i k}, \\boldsymbol {\\Sigma} _ {i k}, w _ {i k} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.468, + 0.461 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu}_{ik}\\), \\(\\pmb{\\Sigma}_{ik}\\), and \\(w_{ik}\\) denote the mean, covariance matrix and the prior probability of mixture \\(k\\); \\(\\pmb{\\theta}\\) denotes the network parameters." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.462, + 0.468, + 0.522 + ], + "angle": 0, + "content": "To improve the fidelity of the object positions in the predicted layout, we introduce a hybrid training objective including an absolute position objective and a relative position objective." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.532, + 0.468, + 0.652 + ], + "angle": 0, + "content": "Absolute Position Objective The absolute position objective provides direct supervision of the exact position of each object. Formally, we assume access to an image captioning dataset \\(\\mathcal{D}_{\\mathrm{real}}\\) with description \\(D\\) as well as the extracted labels of all the objects \\(\\{O_i\\}\\) and their center coordinates \\(\\{C_i\\}\\). Then the training objective is to minimize the negative log-likelihood of the ground-truth coordinates under the predicted GMM distribution, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.66, + 0.468, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a b s}} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {D}, \\left\\{\\boldsymbol {C} i \\right\\} \\sim \\mathcal {D} _ {\\mathrm {r e a l}}} \\left[ \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {n l l}} \\left(\\boldsymbol {C} _ {i}; \\boldsymbol {f} _ {i} (\\boldsymbol {D}; \\boldsymbol {\\theta})\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.704, + 0.468, + 0.734 + ], + "angle": 0, + "content": "where \\(\\ell_{\\mathrm{nll}}(C_i;f_i)\\) denotes the negative log-likelihood of ground truth \\(C_i\\) under the GMM specified by \\(f_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.743, + 0.468, + 0.833 + ], + "angle": 0, + "content": "Relative Position Objective In many cases, the text description only mentions the relative positions of the objects, and thus it is more important to ensure the relative position of the predicted position is correct than the absolute position. To further enforce the fidelity of the relative positions, we introduce the following relative position objective." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.835, + 0.468, + 0.879 + ], + "angle": 0, + "content": "To start with, we construct a synthetic dataset, \\(\\mathcal{D}_{\\mathrm{syn}}\\) which consists of text description \\(D\\) with explicit descriptions of relative spatial relations. We first randomly select \\(N\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "objects with attribute modifiers. We then select \\( M \\) pairs of objects to specify their relative spatial relation. For object pair \\( (O_i, O_j) \\), their spatial relation, \\( R_{ij} \\), is randomly drawn from \"left of\", \"right of\", \"above\" and \"below\". Finally, we prompt GPT-3 [4] to generate the text description \\( D \\) that mentions all the objects and their relative positions. Our familiar \"A red car is to the left of a black mailbox\" is one such example. Further details are provided in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Next, we introduce a loss to penalize violating the relative position. If object \\( i \\) is to the left of object \\( j \\), we enforce that the rightmost mixture mean of object \\( i \\) is to the left of the leftmost mixture mean of object \\( j \\). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.281, + 0.899, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {r e l}} \\left(R _ {i j} = ^ {\\prime \\prime} \\text {l e f t} ^ {\\prime \\prime}; \\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right) = \\max \\left\\{\\max _ {k} \\boldsymbol {\\mu} _ {i k} (0) - \\min _ {k} \\boldsymbol {\\mu} _ {j k} (0), - \\delta \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.315, + 0.892, + 0.388 + ], + "angle": 0, + "content": "where \\(\\mu_{ik}(0)\\) denotes the zeroth element (the \\(x\\)-coordinate) of \\(\\mu_{ik}\\), and \\(\\delta\\) is a pre-specified margin. The \\(\\ell_{\\mathrm{rel}}\\) of the other three types of relations are defined similarly. The relative position loss is thus the aggregation of \\(\\ell_{\\mathrm{rel}}\\) across the synthetic dataset:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.394, + 0.891, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e l}} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {D}, \\left\\{R _ {i j} \\right\\} \\sim \\mathcal {D} _ {\\mathrm {s y n}}} \\left[ \\sum_ {i, j: R _ {i j} \\neq \\emptyset} \\ell_ {\\mathrm {r e l}} \\left(R _ {i j}; \\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.436, + 0.7, + 0.451 + ], + "angle": 0, + "content": "where \\( \\pmb{f}_i \\) is short for \\( f_{i}(D;\\theta) \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.89, + 0.489 + ], + "angle": 0, + "content": "Training and Inference To sum up, the final training objective is the combination of both:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.499, + 0.891, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a y o u t}} (\\boldsymbol {\\theta}) = \\mathcal {L} _ {\\text {a b s}} (\\boldsymbol {\\theta}) + \\xi \\mathcal {L} _ {\\text {r e l}} (\\boldsymbol {\\theta}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.524, + 0.892, + 0.644 + ], + "angle": 0, + "content": "where \\(\\xi\\) is a hyperparameter. During inference time, we randomly draw the center coordinate \\(C_i\\) from the predicted GMM. The pixel region for object \\(O_i\\), \\(\\mathcal{R}_i\\), is defined as a circular region centered at the drawn \\(C_i\\) with a fixed radius \\(r\\). As would be shown in Appendix D, \\(\\mathcal{R}_i\\) only roughly regulates the position of the generated objects, and the actual size of the object can go beyond or below the size of \\(\\mathcal{R}_i\\). Thus a fixed \\(r\\) is sufficient for this purpose." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.854, + 0.67 + ], + "angle": 0, + "content": "3.4. Spatial-Temporal Attention Optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.677, + 0.891, + 0.753 + ], + "angle": 0, + "content": "Recall that \\(D\\) is the global description and \\(\\{L_i\\}\\) are local descriptions for each object. Our goal is to guide the diffusion model to attend to not only \\(D\\), but also \\(L_{i}\\) in object \\(i\\)’s region, \\(\\mathcal{R}_i\\), so that the model is more strongly prompted to generate the specific object as specified in the layout." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.762, + 0.89, + 0.791 + ], + "angle": 0, + "content": "Spatial-Temporal Attention Recall that the standard cross-attention [49] is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.799, + 0.891, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {Q} \\boldsymbol {K} ^ {T} / \\sqrt {d}\\right) \\boldsymbol {V}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.891, + 0.899 + ], + "angle": 0, + "content": "where \\(Q \\in \\mathbb{R}^{h \\times w, d}\\) is the queries vectors for each pixel, and \\(K, V \\in \\mathbb{R}^{l, d}\\) are key and value vectors for each text description token; \\(h\\) and \\(w\\) represent the image height and width, \\(l\\) represents the text description length; \\(d\\) represents the dimension of each attention vector." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.444, + 0.9 + ], + "angle": 0, + "content": "3More details of the layout predictor can be found in Appendix D." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7769" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "Since we have multiple text descriptions, we define \\( K_{D}, V_{D} \\) as the key and value vectors for the global description \\( D \\), and \\( K_{L_i}, V_{L_i} \\) for the local description \\( L_i \\). Further, we introduce a set of binary mask matrices \\( \\{M_i\\} \\) to indicate the region for each object, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.173, + 0.469, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {i} (x, y) = 1, \\text {i f} (x, y) \\in \\mathcal {R} _ {i}, \\quad \\text {a n d} 0, \\text {o t h e r w i s e .} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.195, + 0.469, + 0.224 + ], + "angle": 0, + "content": "Then the output of the attention layer of the denoising network at time \\( t \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.228, + 0.469, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {O} (t) = \\sum_ {i = 1} ^ {N} \\lambda_ {i t} \\boldsymbol {M} _ {i} \\odot \\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K} _ {\\boldsymbol {L} _ {i}}, \\boldsymbol {V} _ {\\boldsymbol {L} _ {i}}) \\tag {8} \\\\ + \\left(1 - \\sum_ {i = 1} ^ {N} \\lambda_ {i t} M _ {i}\\right) \\odot \\operatorname {A t t e n t i o n} \\left(\\boldsymbol {Q}, \\boldsymbol {K} _ {D}, \\boldsymbol {V} _ {D}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.469, + 0.456 + ], + "angle": 0, + "content": "where \\(\\lambda_{it}\\) are the attention combination weights, and \\(\\odot\\) denotes element-wise multiplication. Note that the combination weights are functions of time \\(t\\), which is motivated by the observation that different denoising steps control the generation of different levels of details. When synthesizing images, the combination weights \\(\\lambda_{it}\\) are optimized on the fly for each corresponding mask and time. By introducing this time dependency, we allow the diffusion model to focus more on the global description at earlier time steps and shift to local descriptions later." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.466, + 0.47, + 0.662 + ], + "angle": 0, + "content": "Optimization Objective All the attention combination weights, \\(\\lambda = \\{\\lambda_{it}\\}\\), are determined by maximizing the consistency between the generated image and the text description as measured by the CLIP similarity [35]. We introduce two CLIP similarities, a global CLIP similarity that compares the entire image and the global description, and a set of local CLIP similarities that compare the images at each object region and the corresponding local descriptions. Formally, denote the generated image as \\(I(\\lambda)\\), and its local patch at each object's region as \\(I_{O_i}(\\lambda)\\). Note that we adopt the deterministic PLMS denoising process [27] (i.e., with \\(\\sigma = 0\\)), so both \\(I\\) and \\(I_{O_i}\\) are deterministic functions of \\(\\lambda\\). Then the loss for optimizing \\(\\lambda\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.667, + 0.469, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a t t e n d}} (\\boldsymbol {\\lambda}) = - \\operatorname {C L I P} (\\boldsymbol {I} (\\boldsymbol {\\lambda}), \\boldsymbol {D}) - \\gamma \\sum_ {i = 1} ^ {N} \\operatorname {C L I P} \\left(\\boldsymbol {I} _ {O _ {i}} (\\boldsymbol {\\lambda}), \\boldsymbol {L} _ {i}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.709, + 0.469, + 0.739 + ], + "angle": 0, + "content": "where \\(\\mathrm{CLIP}(\\cdot)\\) denotes the CLIP similarity, and \\(\\gamma\\) is a hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.751, + 0.208, + 0.767 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.777, + 0.469, + 0.821 + ], + "angle": 0, + "content": "We conduct experiments to evaluate our method's performance and generalizability. We also perform ablation study on important design choices of our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Implementing Details: We adopt RoBERTa-base [29] as the base model for layout predictor and use 5 mixtures in" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "GMM. We use stable-diffusion-v1-4 [40] pre-trained on the laion dataset [44] and freeze it throughout all experiments. All generated images are in the size of \\(512 \\times 512\\). We use PLMS sampler [27] to synthesize images with 50 denoising steps, and we use Adam [20] to optimize the layout predictor and attention combination weights. More details on hyperparameters and optimization are in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.205, + 0.871, + 0.22 + ], + "angle": 0, + "content": "4.1. Evaluation on Fidelity of Generated Images" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.273 + ], + "angle": 0, + "content": "We first evaluate our method on object, attribute, and spatial fidelities as introduced in Sec. 3.1 using both objective and subjective metrics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.278, + 0.893, + 0.7 + ], + "angle": 0, + "content": "Baselines: We identify four baseline methods on generating images from complex text descriptions. (1) VANILLA-SD [40] is the pre-trained text-to-image stable diffusion model that directly generates images conditioned on the text description. (2) COMPOSABLE-DIFFUSION [28] is a diffusion-based compositional generation method. To generate images from a text description, the text is first decomposed into the conjunction of multiple components (e.g., for the example in Sec. 3.1, the components are \"A red car is to the left of a black mailbox.\" AND \"A red car\" AND \"a black mailbox\"). Each component is separately modeled by the diffusion model and then composed to generate an image by merging the outputs of each denoising step. (3) STRUCTURE-DIFFUSION [10] improves VANILLA-SD on generating images with better object and attribute fidelities. They do so by first extracting noun phrases at different levels of the parsing tree. The model then separately attends to each noun phrase and combines the attention outputs by taking their arithmetic mean. Different from our method, they perform cross-attention on the whole image instead of the region specific to an object. (4) PAINT-WITH-WORDS [2] assumes that users provide the pixel region of each object to be generated. Pixels in the region then increase their attention weights to the text that describes the corresponding object, and the amount of increase is determined by heuristic rules. For a fair comparison, we use the pixel region predicted by our layout predictor, and we report the performance with the ground truth region in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets and Metrics: We conduct experiments on three datasets. (1) MS-COCO [25] contains photos taken by photographers and manually annotates the caption for each photo. We use the caption as the input text description. (2) VSR [26] is proposed for probing spatial understanding of vision-language models. Constructed from a subset of MS-COCO, it uses templates to generate captions that describe spatial relations in the image (e.g., \"The horse is to the left of the person\"). (3) GPT-synthetic is the synthetic dataset introduced in Sec. 3.3. We manually check the test set to filter out sentences that do not conform to the specified spatial relation. Compared with VSR, this dataset has more diverse and complex text descriptions. The final dataset contains" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": "4To standardize the image patch sizes, \\( I_{O_i} \\) is obtained by cropping the original image with a minimum square that encompasses \\( \\mathcal{R}_i \\) and resizing it to \\( 224 \\times 224 \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7770" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.891, + 0.171 + ], + "angle": 0, + "content": "
MS-COCOVSRGPT-synthetic
Object (100) ↑Attribute (100) ↑OverallObject (100) ↑Spatial (100) ↑OverallObject (100) ↑Attribute (100) ↑Spatial (100) ↑Overall
VANILLA-SD666242%646648%60586038%
COMPOSABLE-DIFFUSION534830%636018%63604728%
STRUCTURE-DIFFUSION606348%546432%58545832%
PAINT-WITH-WORDS757452%545930%65525732%
Ours7574-6877-848386-
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.171, + 0.893, + 0.227 + ], + "angle": 0, + "content": "Table 1. Subjective evaluation of our method and baselines. Best numbers are in bold. Spatial relation is not available on MS-COCO because very few of its captions contain spatial relations. Attribute is not available on VSR as its captions do not consider attribute. Object, Attribute, and Spatial show the total score of 50 evaluations, where a model with the highest fidelity would achieve a score of 100. Overall denotes the percentage of generation of each method that is rated better than our method." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.238, + 0.468, + 0.324 + ], + "angle": 0, + "content": "
MS-COCOVSRGPT-synthetic
Object RecallSPRel PrecisionObject RecallSPRel PrecisionObject RecallSPRel Precision
VANILLA-SD58.0%-62.1%56.0%42.4%56.8%
COMPOSABLE-DIFFUSION51.8%-60.8%72.2%34.4%52.6%
STRUCTURE-DIFFUSION61.7%-62.3%58.3%43.0%59.3%
PAINT-WITH-WORDS57.3%-63.8%47.1%45.9%53.5%
Ours69.6%-65.1%75.0%47.2%66.7%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.325, + 0.468, + 0.336 + ], + "angle": 0, + "content": "Table 2. Automatic evaluation of our method and baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.468, + 0.38 + ], + "angle": 0, + "content": "SPRel Precision: Spatial Relation Precision. Best numbers are in bold. SPRel precision is not available on MS-COCO since most captions do not have an explicit spatial relation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.442 + ], + "angle": 0, + "content": "500 descriptions, and we downsample MS-COCO and VSR to have the same size. Statistics including the number of objects and spatial relations are shown in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.442, + 0.47, + 0.745 + ], + "angle": 0, + "content": "For automatic evaluations, we consider two metrics. (1) Object Recall measures the percentage of successfully synthesized objects over objects mentioned in the text. We use DETR [5] to detect objects in generated images. To calculate recall, we divide the number of detected objects in the text by the total number of objects in the text that also belong to one of the MS-COCO categories. It measures the object fidelity of generated images. (2) Spatial Relation Precision (SPRel Precision) measures the percentage of the correct spatial relations among all the relations whose corresponding objects are successfully synthesized. This metric measures the spatial fidelity of generated images. We consider the relation of left, right, above, and below because their correctness can be evaluated by comparing the bounding box centers. Qualitative examples of spatial relations beyond these four are shown in Appendix G. We also report the CLIP similarities between generated images and the input description in Appendix C. We observe that different methods have very close CLIP similarities, though our method still achieves competitive results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results: The results are shown in Table 2. As shown in the table, our method outperforms all baselines on both object recall and SPRel precision. Specifically, compared with baselines that do not have spatial attention control (all but PAINT-WITH-WORDS), our method is significantly better on SPRel precision, which illustrates the effectiveness of spatial control and indicates that it is difficult for the stable diffusion model to generate correct spatial relations without guidance on the layout. Our method moves the burden of generating objects at the correct location from diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.892, + 0.348 + ], + "angle": 0, + "content": "models to a separate layout predictor, allowing the whole pipeline to achieve better spatial fidelity. On the other hand, compared with PAINT-WITH-WORDS that does not impose the temporal attention control, our method achieves better object recall, showing that our optimized combination weights across the temporal dimension strike a better balance between local and global descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Fig. 3 demonstrates examples of text descriptions and images generated by our method and baselines. We observe that our method resolves three types of errors in baselines. First, our method alleviates the missing object issue (in the top panel). Baselines tend to focus on one object in the text and ignore other objects (e.g., focusing on the elephant and ignoring the man in the second row), whereas our method generates all objects. Second, as shown in the middle panel, our method mitigates the mismatched attribute issue. Particularly, baselines struggle when multiple objects are mentioned in the text, where they mismatch the attribute and object (e.g., red truck in the first row). Finally, the bottom panel shows examples where our method reduces the number of mislocated objects. Note that our method is effective both on the four relations and other spatial relations. More examples can be found in Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Subjective Evaluation: To further evaluate the fidelity and quality of the generated images, we perform a subjective evaluation on Amazon Mechanical Turk. Specifically, we randomly sample 25 text descriptions for each dataset in Table 2. Each subject was presented with the text description and corresponding image generated by a single method or a pair of methods and asked the following four questions: (1) (Object Fidelity) Does the image contain all objects mentioned in the text? (2) (Attribute Fidelity) Are all synthesized objects consistent with their characteristics described in the text (e.g., color and material)? (3) (Spatial Fidelity) Does the image locate all objects at the correct position such that the spatial relations in the text are satisfied (if an object in the relationship is missing, it is considered as an incorrect generation)? and (4) (Overall) Which image in the pair has higher fidelity with the text and has better quality? The first three questions are evaluated for each method individually with a score of 0, 1, or 2, where 2 denotes all objects/attributes/relations are correct and 0 denotes none of them is correct. The last question is evaluated on a pair of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7771" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.085, + 0.852, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.74, + 0.895, + 0.768 + ], + "angle": 0, + "content": "Figure 3. Example images generated by our method and baselines. Typical errors of baselines include missing objects, mismatched attributes, and mislocated objects. Ours (1)/(2) show the results with two different random seeds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.789, + 0.471, + 0.85 + ], + "angle": 0, + "content": "images, one generated by our method and the other by a baseline. Each generated image is evaluated by two subjects, so the total score of the first three questions is 100. More details are in Appendix H." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Table 1 shows the results. Our method achieves significantly better performance in most cases, especially on GPT-synthetic dataset that contains multiple objects and relations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.789, + 0.892, + 0.835 + ], + "angle": 0, + "content": "in the same description. The results show that our method is effective at generating images with high fidelity without sacrificing perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.693, + 0.864 + ], + "angle": 0, + "content": "4.2. Additional Analyses" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Performance on Each Text Complexity Level: To explore the capability of our method, especially on complex text, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7772" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.086, + 0.276, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.086, + 0.455, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.2, + 0.47, + 0.227 + ], + "angle": 0, + "content": "Figure 4. Performance when the number of objects and spatial relations in the text increase." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.255, + 0.46, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.399, + 0.47, + 0.427 + ], + "angle": 0, + "content": "Figure 5. Example images generated by our method and baselines on novel object combinations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.661 + ], + "angle": 0, + "content": "analyze the performance at each complexity level. Specifically, we use the GPT-synthetic dataset where multiple objects and relations can appear in the same description. We use the number of objects \\( N \\) and number of spatial relations \\( M \\) introduced in Sec. 3.3 as proxies for text complexity and plot the performance of each method as the text becomes more complex. i.e., we plot the performance for each value of \\( N \\) regardless of \\( M \\), and vise versa for \\( M \\). As shown in Fig. 4, the performance of all methods decreases as the text becomes more complex, but our method still outperforms or is on par with others on complex descriptions. Note that PAINT-WITH-WORDS uses the same layout predictor as ours, which partially explains its strong performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Generalizability to novel object combinations is a critical requirement for text-to-image models. However, since MSCOCO dataset is collected from real photos, most of its captions contain object pairs and object-attribute pairs that are common in daily life, which are also more likely to overlap with stable diffusion's pre-training data. Thus we additionally evaluate our method on another synthetic dataset that is created similarly to Sec. 3.3, but contains uncommon object pairs, object-attribute pairs, and object spatial relations. Fig. 5 shows sample images generated by our method and baselines. It can be observed that our method successfully generates novel object pairs (e.g., \"zebra is looking at computer\"). Performance on this dataset and more generated examples are shown in Appendix F. Overall, our method is able to generalize to novel object combinations." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.089, + 0.861, + 0.17 + ], + "angle": 0, + "content": "
Object RecallSPRel Precision
Ours47.2%66.7%
No spatial control39.6%51.7%
No temporal control43.8%61.4%
No optimization41.0%55.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.171, + 0.865, + 0.183 + ], + "angle": 0, + "content": "Table 3. Ablation performance on GPT-synthetic test set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.2, + 0.653, + 0.216 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.299 + ], + "angle": 0, + "content": "In this section, we will investigate the influence of two important steps in our method, namely, the layout predictor and the spatial-temporal attention. We first study how spatial-temporal attention affects our performance. To do that, we consider three variants of attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.893, + 0.436 + ], + "angle": 0, + "content": "First, we explore the attention without spatial control. Concretely, the diffusion model still attends to both global description and local descriptions. However, instead of only attending to a local description in the pixel region of the object, the model now attends to all local descriptions in the whole image. The combination weights at each denoising step are optimized with only the global CLIP similarity in Sec. 3.4. Table 3 shows that its performance drops drastically, demonstrating the importance of spatial control." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Second, we explore the attention without temporal control, where combination weights remain the same for all denoising steps. Note that the weights can be different for each object and are optimized with \\(\\mathcal{L}_{\\mathrm{attend}}(\\lambda)\\) in Sec. 3.4. The results in Table 3 show a significant degradation under both metrics, indicating the benefits of temporal dependency on attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.663 + ], + "angle": 0, + "content": "Finally, we explore the attention without optimization. Specifically, we fix the combination weight \\(\\lambda_{it} = \\frac{1}{N}\\) for all \\(i\\) and \\(t\\), where \\(N\\) is the number of objects. This value is also the initialization point of \\(\\{\\lambda_{it}\\}\\) in our method. As shown in Table 3, the performance further drops compared to the no temporal control variant. The results indicate that optimizing the combination weights for a new text description is critical for generating images with high fidelity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.755 + ], + "angle": 0, + "content": "We further study the effects of our layout predictor, which includes the layout predictor trained with only one of the absolute and relative position objectives, the comparison with ground truth and user-provided pixel region, and a different strategy to construct the pixel region for an object. Results of these experiments are presented in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.619, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we study the text-to-image synthesis task based on diffusion models. We find existing methods lack explicit control on cross-attention in diffusion models, which leads to the generation of low-fidelity images. We propose an algorithm that imposes control on cross-attention in spatial and temporal aspects. Experiments show our method outperforms baselines in generating high" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7773" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "fidelity images. Further ablation study verifies the effectiveness of our spatial and temporal attention control. One limitation of our method is the reliance on a time-consuming optimization scheme, which takes around 10 minutes for each text-to-image generation. Future work may consider striking a better balance between performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.197, + 0.258, + 0.214 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.222, + 0.47, + 0.253 + ], + "angle": 0, + "content": "The work of Qiucheng Wu, Yujuan Liu, and Shiyu Chang was partially supported by Cisco Faculty Award." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.267, + 0.175, + 0.283 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.468, + 0.32 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein gan, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.322, + 0.468, + 0.39 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.392, + 0.468, + 0.433 + ], + "angle": 0, + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.435, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.576, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.618, + 0.468, + 0.673 + ], + "angle": 0, + "content": "[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.676, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.733, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.776, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[9] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv, 2022. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.292, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Neurips, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.891, + 0.318 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Neurips, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., pages 1-33, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[17] Nisha Huang, Fan Tang, Weiming Dong, Tong-Yee Lee, and Changsheng Xu. Region-aware diffusion for zero-shot text-driven image editing. arXiv preprint arXiv:2302.11797, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2014. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[23] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. arXiv preprint arXiv:2205.14217, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. arXiv preprint arXiv:2205.00363, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7774" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[28] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv, 2019. 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[31] Gary Marcus, Ernest Davis, and Scott Aaronson. A very preliminary analysis of dall-e 2, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.279, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[34] Dong Huk Park, Grace Luo, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. arXiv preprint arXiv:2212.00210, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.449, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[38] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Neurips, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[39] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[41] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[43] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[45] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning. PMLR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[50] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[51] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[52] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: a geometric diffusion model for molecular conformation generation, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[53] Cheng-Fu Yang, Wan-Cyuan Fan, Fu-En Yang, and Yu-Chiang Frank Wang. Layouttransformer: Scene layout generation with conceptual and spatial diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3732-3741, June 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[54] Zuopeng Yang, Daqing Liu, Chaoyue Wang, Jie Yang, and Dacheng Tao. Modeling image composition for complex scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7764-7773, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[55] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7775" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[56] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collomosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. arXiv preprint arXiv:2211.11742, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[57] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[58] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7776" + } + ] +] \ No newline at end of file diff --git a/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_origin.pdf b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a1272017108a07ad5e30adb8e337b14cace06839 --- /dev/null +++ b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/ae9152f5-e620-4815-87af-2959ad2b524b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a43a7ed9cc7cfa47909bde0891e5b73b3ca8c84eff969c0875251dab4e5f99a +size 6213720 diff --git a/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/full.md b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2341f4872b6b9a2e82a042086e68eb87a72523e8 --- /dev/null +++ b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/full.md @@ -0,0 +1,340 @@ +# Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis + +Qiucheng Wu $^{1*}$ , Yujuan Liu $^{1*}$ , Handong Zhao $^{2}$ , Trung Bui $^{2}$ , Zhe Lin $^{2}$ , Yang Zhang $^{3}$ , Shiyu Chang $^{1}$ , UC Santa Barbara, $^{2}$ Adobe Research, $^{3}$ MIT-IBM Watson AI Lab + +{qiucheng, yujianliu}@ucsb.edu + +![](images/45f9aa9045642a42774a107f3724fd5ab47c48b40588e1e3865a64ec655fa4f1.jpg) +StableDiffusion + +![](images/cc6bcb2359c4be77375ed66aa8aeb226ef0e3c7598eaded3f6c4a043640be375.jpg) +Missing Objects +Ours + +![](images/cdbb9a1becc861fe770277192e5d698fe823e2ae886f39f218db105df0abce66.jpg) +An airport employee giving directions to an airplane. +A zebra brazing on green grass next to a pile of rocks. +(a) +Figure 1. Example generation by stable-diffusion-v1-4 and our method. Stable diffusion model makes three types of errors that include missing objects, mismatched attributes, and mislocated objects. Errors are highlighted in red. + +![](images/6983621a8be527cf85e44ba0ddc03f077c13476df73ffa582cb5139481e6a13b.jpg) + +![](images/48b025c81218bcaa972c6a3a87b46de7d20497e2a3d478bf64277ca72be978a2.jpg) +Mismatched Attributes +StableDiffusion +A bedroom has green carpeting and a bed. + +![](images/d2a7ab16c264f3ec31d1f9affa32f4f5c694a4050d6fc9e1fec25dd2f66a5b30.jpg) +Ours + +![](images/ad7aa174e9b3ef49d079b8b31cdad4f343d105baf8df8bc5a188d06f6ece4162.jpg) +A red car is to the left of a black mailbox. +(b) + +![](images/2728eb6bde0414eca02cdb5fea9c7104788bfadcfdd8bacd24ae4c2ff70bf104.jpg) +Mislocated Objects +StableDiffusion +The brown teddy bear is placed high above the toilet. + +![](images/09756c482440a87c372aefc049ea576bbb437960ff77ce730399549d4028f02d.jpg) +Ours + +![](images/b9bb0535bb90fae8f90f4bb47adbf09e186ab4c93955e63b4e46627d318567d4.jpg) +The teddy bear is right of the suitcase. +(c) + +![](images/35027d1451bd4998133b64cb67481385f68e5cd83536d0631ff49149869bf63f.jpg) + +# Abstract + +Diffusion-based models have achieved state-of-the-art performance on text-to-image synthesis tasks. However, one critical limitation of these models is the low fidelity of generated images with respect to the text description, such as missing objects, mismatched attributes, and mislocated objects. One key reason for such inconsistencies is the inaccurate cross-attention to text in both the spatial dimension, which controls at what pixel region an object should appear, and the temporal dimension, which controls how different levels of details are added through the denoising steps. In this paper, we propose a new text-to-image algorithm that adds explicit control over spatial-temporal cross-attention in diffusion models. We first utilize a layout predictor to predict the pixel regions for objects mentioned in the text. We then impose spatial attention control by combining the attention over the entire text description and that over the + +local description of the particular object in the corresponding pixel region of that object. The temporal attention control is further added by allowing the combination weights to change at each denoising step, and the combination weights are optimized to ensure high fidelity between the image and the text. Experiments show that our method generates images with higher fidelity compared to diffusion-model-based baselines without fine-tuning the diffusion model. Our code is publicly available.1 + +# 1. Introduction + +Diffusion models [14, 19, 46, 47, 48] have recently revolutionized the field of image synthesis. Compared with previous generative models such as generative adversar + +ial networks [1, 3, 11, 18] and variational autoencoders [21, 38, 39], diffusion models have demonstrated superior performance in generating images with higher quality, more diversity, and better control over generated contents. Particularly, text-to-image diffusion models [2, 33, 36, 40, 42] allow generating images conditioned on a text description, which enables generation of creative images due to the expressiveness of natural language. + +However, recent studies [10, 31] have revealed that one critical limitation of existing diffusion-model-based text-to-image algorithms is the low fidelity with respect to the text descriptions – the content of the generated image is sometimes at odds with the text description, especially when the description is complex. Specifically, typical errors made by stable diffusion models fall into three categories: missing objects, mismatched attributes, and mislocated objects. For example, in Fig. 1(a), stable diffusion model ignores the airplane even though it is mentioned in text; in Fig. 1(b), the model confuses “red car” and “black mailbox” and generates a red mailbox; in Fig. 1(c), the model locates teddy bear behind the toilet, despite the description “teddy bear is placed high above the toilet.” + +Such infidelity problems suggest that the cross-attention map on the text description may not be accurate. In particular, if we view the generation process of a diffusion model as a sequence of denoising steps, then the cross-attention on text descriptions can be considered as a function of both spatial (pixels) and temporal (denoising steps) information. Therefore, the inaccuracies of the cross-attention can result from the loose control over both the spatial and temporal dimensions. On one hand, spatial attention controls at what pixels the model should attend to each object and the corresponding attributes mentioned in the text. If the spatial attention is incorrect, the resulting images will have incorrect object locations or miss-associated attributes. On the other hand, temporal attention controls when the models should attend to different levels of details in the text. As previous works have revealed, diffusion models tend to focus on generating object outlines at earlier denoising steps and on details at later [50]. Thus loose control over the temporal aspect of attention can easily lead to overlooking certain levels of the object details. In short, to improve the fidelity of text-to-image synthesis, one would need to explicitly control both spatial and temporal attention to follow an accurate and optimal distribution. + +In this paper, we propose a new text-to-image algorithm based on a pre-trained conditional diffusion model with explicit control over the spatial-temporal cross-attention map on text. The proposed algorithm introduces a layout predictor and a spatial-temporal attention optimizer. The layout predictor takes the text description as input and generates a spatial layout for each object mentioned in the text. Alternatively, the layout can also be provided by the user. Then + +the spatial-temporal attention optimizer imposes direct control over the spatial and temporal aspects of the attention according to the spatial layout. In particular, for the spatial aspect, we parameterize the attention map such that the attention outputs in the designated pixel region for an object are a weighted combination of attention over the entire text description and that over the local description that specifically describes the corresponding object. In this way, we manage to emphasize the attention over the object descriptions. For the temporal aspect, we allow the combination weights to change across time and optimized according to a CLIP objective that measures the agreement between the generated images and the text description. In this way, we allow the attention to focus more on the entire description at the early stage and gradually shift to the detailed local descriptions as the denoising process proceeds. The entire pipeline resembles a typical painting process of a human, where each object's position is determined beforehand and the focus gradually shifts from global information to the local details of each object. + +We conduct extensive experiments on datasets that contain real and template-based captions [6, 26] and our newly created synthetic dataset that contains complex text descriptions. Results show that our method generates images that better align with descriptions compared to other stable diffusion-based baselines. As shown in Fig. 1, our method effectively resolves the above-mentioned three errors. Particularly, controlling spatial attention locates objects at the desired position, and controlling temporal attention promotes the occurrence of objects with associated attributes. Our findings shed light on fine-grained control of diffusion models in text-to-image generation tasks. + +# 2. Related Work + +Diffusion Models Diffusion models are a class of generative models that have demonstrated state-of-the-art performance on image synthesis tasks [8, 14, 19, 46, 47, 48]. These models synthesize images by sampling a noisy image from the standard Gaussian distribution and iteratively denoising it back to a clean image. Their impressive performance has advanced research in multiple computer vision areas including inpainting [30, 41, 51], image editing [7, 12, 32, 50], super-resolution [15, 43], video synthesis [13, 16, 55], and applications beyond computer vision [22, 23, 52]. Among these, text-to-image diffusion models have gained significant attention [2, 36, 37, 42, 58]. Taking text descriptions as inputs, these models generate high quality images that are semantically aligned with the input text, which have led to many creative and artistic applications. + +Enhancing the Controllability of Text-to-Image Diffusion Models While diffusion-based text-to-image models have shown promising results, recent works have high + +![](images/e7cfe7163e33d910370800bc7d71acf9a9738e94cdedcb0c9cbe2930f223a17f.jpg) +Figure 2. Overview of our text-to-image generation pipeline at one denoising step. Given input text $D$ , we first parse it and extract all objects mentioned, constructing local descriptions $L_{i}$ . Next, the layout predictor predicts the pixel region for each object in the text. The diffusion model attends to the global description $D$ and additionally attends to the local description $L_{i}$ in object $i$ 's region. The final attention output is a weighted combination of attention to both global and local descriptions, where the combination weights sum up to 1 for each pixel and are optimized for each denoising step to achieve a high fidelity with $D$ . + +lighted cases where models fail to generate high-fidelity images with respect to the input text [10, 31]. To this end, prior works have explored various ways to enhance the controllability of text-to-image diffusion models. One line of work enhances the controllability by improving diffusion models' ability to understand natural language, which includes using a more powerful text encoder that is separately trained on language modeling tasks [2, 42], incorporating linguistic structures in the text to guide the cross-attention between image and text [10], and decomposing a complex text description into multiple components that are easy to generate [28]. Another line of work conditions diffusion models' generation on auxiliary inputs such as object layout [2, 9, 24, 54] and silhouette [17, 34, 45, 56, 57]. By modifying diffusion models' attention operation according to these auxiliary information or directly fine-tuning diffusion models to take these auxiliary inputs, they are able to control the location and shape of the objects in the image. Finally, some work adds temporal aspect control on diffusion models by modifying the input text condition at each denoising step, which allows them to disentangle a desired attribute from other contents [50]. Different from prior works, our method imposes both spatial and temporal control in cross-attention layer. Moreover, our method does not require auxiliary inputs and does not fine-tune the diffusion model. + +# 3. Methodology + +# 3.1. Problem Formulation + +We focus on the standard text-to-image problem. Given a text description, denoted as $D$ , our goal is to generate an image that is consistent with $D$ . For a concrete exposition, we will use an example description "A red car is to the left of a black mailbox" in the following. Our work aims to + +improve the fidelity of the generated image to text, which includes three requirements: + +- Object Fidelity: The generated images should include all the objects mentioned in $D$ . In our example, the generated image should contain a car and a mailbox. +- Attribute Fidelity: The attributes of each object in the image should match those in $D$ . In our example, the car should be red and the mailbox should be black. +- Spatial Fidelity: The relative spatial positions of the object should match the description in $D$ . In our example, the car should be on the left and the mailbox on the right. + +We will tackle these problems using a pre-trained, fixed stable diffusion model. + +# 3.2. Method Overview + +To achieve high-fidelity text-to-image generation, we propose an algorithm consisting of the following four steps. Step 1: Object Identification We extract all the objects mentioned in $D$ , denoted as $O_{1:N}$ , by eliciting the noun phrases using spaCy. $^2 N$ is the total number of objects. In our sample, $O_1 = \text{"a red car"}$ and $O_2 = \text{"a black mailbox"}$ . Step 2: Layout Prediction For each object $O_i$ , we use a layout predictor to predict its pixel region, $\mathcal{R}_i$ , which is a set of pixels roughly specifying where the object should be. The layout can also be provided by human users. + +Step 3: Local Description Generation For each object $O_{i}$ , we generate a local text description, $L_{i}$ , containing only that object information using a simple template. In our example, there are two local descriptions. $L_{1} =$ "A photo of a red car" and $L_{2} =$ "A photo of a black mailbox". + +Step 4: Attention Optimization During the generation process, we guide the diffusion model to combine attention to both the global description $D$ and the local descriptions $L_{1:N}$ according to the object layout. The attention combination weights are optimized for each input description. + +The following subsections will provide further details about steps 2 and 4. + +# 3.3. Layout Predictor + +Our layout predictor is adapted from [53], which aims to predict the center coordinate $C_i = [X_i, Y_i]$ for each object $O_i^3$ . Specifically, the layout predictor is a transformer that takes the text description $D$ as the input. At the output position where the input mentions the object $O_i$ , we let the transformer output a set of Gaussian mixture model (GMM) parameters to fit the object's center coordinate $C_i$ . Formally, denote $f_i(D; \theta)$ as the output of the layout predictor at the location of the mentioning object $O_i$ . Then + +$$ +\boldsymbol {f} _ {i} (\boldsymbol {D}; \boldsymbol {\theta}) = \bigcup_ {k = 1} ^ {K} \left\{\boldsymbol {\mu} _ {i k}, \boldsymbol {\Sigma} _ {i k}, w _ {i k} \right\}, \tag {1} +$$ + +where $\pmb{\mu}_{ik}$ , $\pmb{\Sigma}_{ik}$ , and $w_{ik}$ denote the mean, covariance matrix and the prior probability of mixture $k$ ; $\pmb{\theta}$ denotes the network parameters. + +To improve the fidelity of the object positions in the predicted layout, we introduce a hybrid training objective including an absolute position objective and a relative position objective. + +Absolute Position Objective The absolute position objective provides direct supervision of the exact position of each object. Formally, we assume access to an image captioning dataset $\mathcal{D}_{\mathrm{real}}$ with description $D$ as well as the extracted labels of all the objects $\{O_i\}$ and their center coordinates $\{C_i\}$ . Then the training objective is to minimize the negative log-likelihood of the ground-truth coordinates under the predicted GMM distribution, i.e. + +$$ +\mathcal {L} _ {\mathrm {a b s}} (\boldsymbol {\theta}) = \mathbb {E} _ {\boldsymbol {D}, \left\{\boldsymbol {C} i \right\} \sim \mathcal {D} _ {\mathrm {r e a l}}} \left[ \sum_ {i = 1} ^ {N} \ell_ {\mathrm {n l l}} \left(\boldsymbol {C} _ {i}; \boldsymbol {f} _ {i} (\boldsymbol {D}; \boldsymbol {\theta})\right) \right], \tag {2} +$$ + +where $\ell_{\mathrm{nll}}(C_i;f_i)$ denotes the negative log-likelihood of ground truth $C_i$ under the GMM specified by $f_{i}$ . + +Relative Position Objective In many cases, the text description only mentions the relative positions of the objects, and thus it is more important to ensure the relative position of the predicted position is correct than the absolute position. To further enforce the fidelity of the relative positions, we introduce the following relative position objective. + +To start with, we construct a synthetic dataset, $\mathcal{D}_{\mathrm{syn}}$ which consists of text description $D$ with explicit descriptions of relative spatial relations. We first randomly select $N$ + +objects with attribute modifiers. We then select $M$ pairs of objects to specify their relative spatial relation. For object pair $(O_i, O_j)$ , their spatial relation, $R_{ij}$ , is randomly drawn from "left of", "right of", "above" and "below". Finally, we prompt GPT-3 [4] to generate the text description $D$ that mentions all the objects and their relative positions. Our familiar "A red car is to the left of a black mailbox" is one such example. Further details are provided in Appendix B. + +Next, we introduce a loss to penalize violating the relative position. If object $i$ is to the left of object $j$ , we enforce that the rightmost mixture mean of object $i$ is to the left of the leftmost mixture mean of object $j$ . Formally, + +$$ +\ell_ {\text {r e l}} \left(R _ {i j} = ^ {\prime \prime} \text {l e f t} ^ {\prime \prime}; \boldsymbol {f} _ {i}, \boldsymbol {f} _ {j}\right) = \max \left\{\max _ {k} \boldsymbol {\mu} _ {i k} (0) - \min _ {k} \boldsymbol {\mu} _ {j k} (0), - \delta \right\}, \tag {3} +$$ + +where $\mu_{ik}(0)$ denotes the zeroth element (the $x$ -coordinate) of $\mu_{ik}$ , and $\delta$ is a pre-specified margin. The $\ell_{\mathrm{rel}}$ of the other three types of relations are defined similarly. The relative position loss is thus the aggregation of $\ell_{\mathrm{rel}}$ across the synthetic dataset: + +$$ +\mathcal {L} _ {\mathrm {r e l}} (\boldsymbol {\theta}) = \mathbb {E} _ {\boldsymbol {D}, \left\{R _ {i j} \right\} \sim \mathcal {D} _ {\mathrm {s y n}}} \left[ \sum_ {i, j: R _ {i j} \neq \emptyset} \ell_ {\mathrm {r e l}} \left(R _ {i j}; \boldsymbol {f} _ {i}, \boldsymbol {f} _ {j}\right) \right], \tag {4} +$$ + +where $\pmb{f}_i$ is short for $f_{i}(D;\theta)$ . + +Training and Inference To sum up, the final training objective is the combination of both: + +$$ +\mathcal {L} _ {\text {l a y o u t}} (\boldsymbol {\theta}) = \mathcal {L} _ {\text {a b s}} (\boldsymbol {\theta}) + \xi \mathcal {L} _ {\text {r e l}} (\boldsymbol {\theta}), \tag {5} +$$ + +where $\xi$ is a hyperparameter. During inference time, we randomly draw the center coordinate $C_i$ from the predicted GMM. The pixel region for object $O_i$ , $\mathcal{R}_i$ , is defined as a circular region centered at the drawn $C_i$ with a fixed radius $r$ . As would be shown in Appendix D, $\mathcal{R}_i$ only roughly regulates the position of the generated objects, and the actual size of the object can go beyond or below the size of $\mathcal{R}_i$ . Thus a fixed $r$ is sufficient for this purpose. + +# 3.4. Spatial-Temporal Attention Optimization + +Recall that $D$ is the global description and $\{L_i\}$ are local descriptions for each object. Our goal is to guide the diffusion model to attend to not only $D$ , but also $L_{i}$ in object $i$ ’s region, $\mathcal{R}_i$ , so that the model is more strongly prompted to generate the specific object as specified in the layout. + +Spatial-Temporal Attention Recall that the standard cross-attention [49] is defined as + +$$ +\operatorname {A t t e n t i o n} (\boldsymbol {Q}, \boldsymbol {K}, \boldsymbol {V}) = \operatorname {s o f t m a x} \left(\boldsymbol {Q} \boldsymbol {K} ^ {T} / \sqrt {d}\right) \boldsymbol {V}, \tag {6} +$$ + +where $Q \in \mathbb{R}^{h \times w, d}$ is the queries vectors for each pixel, and $K, V \in \mathbb{R}^{l, d}$ are key and value vectors for each text description token; $h$ and $w$ represent the image height and width, $l$ represents the text description length; $d$ represents the dimension of each attention vector. + +Since we have multiple text descriptions, we define $K_{D}, V_{D}$ as the key and value vectors for the global description $D$ , and $K_{L_i}, V_{L_i}$ for the local description $L_i$ . Further, we introduce a set of binary mask matrices $\{M_i\}$ to indicate the region for each object, i.e. + +$$ +\boldsymbol {M} _ {i} (x, y) = 1, \text {i f} (x, y) \in \mathcal {R} _ {i}, \quad \text {a n d} 0, \text {o t h e r w i s e .} \tag {7} +$$ + +Then the output of the attention layer of the denoising network at time $t$ is defined as + +$$ +\begin{array}{l} \boldsymbol {O} (t) = \sum_ {i = 1} ^ {N} \lambda_ {i t} \boldsymbol {M} _ {i} \odot \operatorname {A t t e n t i o n} (\boldsymbol {Q}, \boldsymbol {K} _ {\boldsymbol {L} _ {i}}, \boldsymbol {V} _ {\boldsymbol {L} _ {i}}) \tag {8} \\ + \left(1 - \sum_ {i = 1} ^ {N} \lambda_ {i t} M _ {i}\right) \odot \operatorname {A t t e n t i o n} \left(\boldsymbol {Q}, \boldsymbol {K} _ {D}, \boldsymbol {V} _ {D}\right), \\ \end{array} +$$ + +where $\lambda_{it}$ are the attention combination weights, and $\odot$ denotes element-wise multiplication. Note that the combination weights are functions of time $t$ , which is motivated by the observation that different denoising steps control the generation of different levels of details. When synthesizing images, the combination weights $\lambda_{it}$ are optimized on the fly for each corresponding mask and time. By introducing this time dependency, we allow the diffusion model to focus more on the global description at earlier time steps and shift to local descriptions later. + +Optimization Objective All the attention combination weights, $\lambda = \{\lambda_{it}\}$ , are determined by maximizing the consistency between the generated image and the text description as measured by the CLIP similarity [35]. We introduce two CLIP similarities, a global CLIP similarity that compares the entire image and the global description, and a set of local CLIP similarities that compare the images at each object region and the corresponding local descriptions. Formally, denote the generated image as $I(\lambda)$ , and its local patch at each object's region as $I_{O_i}(\lambda)$ . Note that we adopt the deterministic PLMS denoising process [27] (i.e., with $\sigma = 0$ ), so both $I$ and $I_{O_i}$ are deterministic functions of $\lambda$ . Then the loss for optimizing $\lambda$ is given by + +$$ +\mathcal {L} _ {\text {a t t e n d}} (\boldsymbol {\lambda}) = - \operatorname {C L I P} (\boldsymbol {I} (\boldsymbol {\lambda}), \boldsymbol {D}) - \gamma \sum_ {i = 1} ^ {N} \operatorname {C L I P} \left(\boldsymbol {I} _ {O _ {i}} (\boldsymbol {\lambda}), \boldsymbol {L} _ {i}\right), \tag {9} +$$ + +where $\mathrm{CLIP}(\cdot)$ denotes the CLIP similarity, and $\gamma$ is a hyperparameter. + +# 4. Experiments + +We conduct experiments to evaluate our method's performance and generalizability. We also perform ablation study on important design choices of our method. + +Implementing Details: We adopt RoBERTa-base [29] as the base model for layout predictor and use 5 mixtures in + +GMM. We use stable-diffusion-v1-4 [40] pre-trained on the laion dataset [44] and freeze it throughout all experiments. All generated images are in the size of $512 \times 512$ . We use PLMS sampler [27] to synthesize images with 50 denoising steps, and we use Adam [20] to optimize the layout predictor and attention combination weights. More details on hyperparameters and optimization are in Appendix A. + +# 4.1. Evaluation on Fidelity of Generated Images + +We first evaluate our method on object, attribute, and spatial fidelities as introduced in Sec. 3.1 using both objective and subjective metrics. + +Baselines: We identify four baseline methods on generating images from complex text descriptions. (1) VANILLA-SD [40] is the pre-trained text-to-image stable diffusion model that directly generates images conditioned on the text description. (2) COMPOSABLE-DIFFUSION [28] is a diffusion-based compositional generation method. To generate images from a text description, the text is first decomposed into the conjunction of multiple components (e.g., for the example in Sec. 3.1, the components are "A red car is to the left of a black mailbox." AND "A red car" AND "a black mailbox"). Each component is separately modeled by the diffusion model and then composed to generate an image by merging the outputs of each denoising step. (3) STRUCTURE-DIFFUSION [10] improves VANILLA-SD on generating images with better object and attribute fidelities. They do so by first extracting noun phrases at different levels of the parsing tree. The model then separately attends to each noun phrase and combines the attention outputs by taking their arithmetic mean. Different from our method, they perform cross-attention on the whole image instead of the region specific to an object. (4) PAINT-WITH-WORDS [2] assumes that users provide the pixel region of each object to be generated. Pixels in the region then increase their attention weights to the text that describes the corresponding object, and the amount of increase is determined by heuristic rules. For a fair comparison, we use the pixel region predicted by our layout predictor, and we report the performance with the ground truth region in Appendix E. + +Datasets and Metrics: We conduct experiments on three datasets. (1) MS-COCO [25] contains photos taken by photographers and manually annotates the caption for each photo. We use the caption as the input text description. (2) VSR [26] is proposed for probing spatial understanding of vision-language models. Constructed from a subset of MS-COCO, it uses templates to generate captions that describe spatial relations in the image (e.g., "The horse is to the left of the person"). (3) GPT-synthetic is the synthetic dataset introduced in Sec. 3.3. We manually check the test set to filter out sentences that do not conform to the specified spatial relation. Compared with VSR, this dataset has more diverse and complex text descriptions. The final dataset contains + +
MS-COCOVSRGPT-synthetic
Object (100) ↑Attribute (100) ↑OverallObject (100) ↑Spatial (100) ↑OverallObject (100) ↑Attribute (100) ↑Spatial (100) ↑Overall
VANILLA-SD666242%646648%60586038%
COMPOSABLE-DIFFUSION534830%636018%63604728%
STRUCTURE-DIFFUSION606348%546432%58545832%
PAINT-WITH-WORDS757452%545930%65525732%
Ours7574-6877-848386-
+ +Table 1. Subjective evaluation of our method and baselines. Best numbers are in bold. Spatial relation is not available on MS-COCO because very few of its captions contain spatial relations. Attribute is not available on VSR as its captions do not consider attribute. Object, Attribute, and Spatial show the total score of 50 evaluations, where a model with the highest fidelity would achieve a score of 100. Overall denotes the percentage of generation of each method that is rated better than our method. + +
MS-COCOVSRGPT-synthetic
Object RecallSPRel PrecisionObject RecallSPRel PrecisionObject RecallSPRel Precision
VANILLA-SD58.0%-62.1%56.0%42.4%56.8%
COMPOSABLE-DIFFUSION51.8%-60.8%72.2%34.4%52.6%
STRUCTURE-DIFFUSION61.7%-62.3%58.3%43.0%59.3%
PAINT-WITH-WORDS57.3%-63.8%47.1%45.9%53.5%
Ours69.6%-65.1%75.0%47.2%66.7%
+ +Table 2. Automatic evaluation of our method and baselines. + +SPRel Precision: Spatial Relation Precision. Best numbers are in bold. SPRel precision is not available on MS-COCO since most captions do not have an explicit spatial relation. + +500 descriptions, and we downsample MS-COCO and VSR to have the same size. Statistics including the number of objects and spatial relations are shown in Appendix B. + +For automatic evaluations, we consider two metrics. (1) Object Recall measures the percentage of successfully synthesized objects over objects mentioned in the text. We use DETR [5] to detect objects in generated images. To calculate recall, we divide the number of detected objects in the text by the total number of objects in the text that also belong to one of the MS-COCO categories. It measures the object fidelity of generated images. (2) Spatial Relation Precision (SPRel Precision) measures the percentage of the correct spatial relations among all the relations whose corresponding objects are successfully synthesized. This metric measures the spatial fidelity of generated images. We consider the relation of left, right, above, and below because their correctness can be evaluated by comparing the bounding box centers. Qualitative examples of spatial relations beyond these four are shown in Appendix G. We also report the CLIP similarities between generated images and the input description in Appendix C. We observe that different methods have very close CLIP similarities, though our method still achieves competitive results. + +Results: The results are shown in Table 2. As shown in the table, our method outperforms all baselines on both object recall and SPRel precision. Specifically, compared with baselines that do not have spatial attention control (all but PAINT-WITH-WORDS), our method is significantly better on SPRel precision, which illustrates the effectiveness of spatial control and indicates that it is difficult for the stable diffusion model to generate correct spatial relations without guidance on the layout. Our method moves the burden of generating objects at the correct location from diffusion + +models to a separate layout predictor, allowing the whole pipeline to achieve better spatial fidelity. On the other hand, compared with PAINT-WITH-WORDS that does not impose the temporal attention control, our method achieves better object recall, showing that our optimized combination weights across the temporal dimension strike a better balance between local and global descriptions. + +Fig. 3 demonstrates examples of text descriptions and images generated by our method and baselines. We observe that our method resolves three types of errors in baselines. First, our method alleviates the missing object issue (in the top panel). Baselines tend to focus on one object in the text and ignore other objects (e.g., focusing on the elephant and ignoring the man in the second row), whereas our method generates all objects. Second, as shown in the middle panel, our method mitigates the mismatched attribute issue. Particularly, baselines struggle when multiple objects are mentioned in the text, where they mismatch the attribute and object (e.g., red truck in the first row). Finally, the bottom panel shows examples where our method reduces the number of mislocated objects. Note that our method is effective both on the four relations and other spatial relations. More examples can be found in Appendix G. + +Subjective Evaluation: To further evaluate the fidelity and quality of the generated images, we perform a subjective evaluation on Amazon Mechanical Turk. Specifically, we randomly sample 25 text descriptions for each dataset in Table 2. Each subject was presented with the text description and corresponding image generated by a single method or a pair of methods and asked the following four questions: (1) (Object Fidelity) Does the image contain all objects mentioned in the text? (2) (Attribute Fidelity) Are all synthesized objects consistent with their characteristics described in the text (e.g., color and material)? (3) (Spatial Fidelity) Does the image locate all objects at the correct position such that the spatial relations in the text are satisfied (if an object in the relationship is missing, it is considered as an incorrect generation)? and (4) (Overall) Which image in the pair has higher fidelity with the text and has better quality? The first three questions are evaluated for each method individually with a score of 0, 1, or 2, where 2 denotes all objects/attributes/relations are correct and 0 denotes none of them is correct. The last question is evaluated on a pair of + +![](images/8c5424cc3c010909ff33ff2d09b9ab7c0a0163540720ed4d36de47b75f320f20.jpg) +Figure 3. Example images generated by our method and baselines. Typical errors of baselines include missing objects, mismatched attributes, and mislocated objects. Ours (1)/(2) show the results with two different random seeds. + +images, one generated by our method and the other by a baseline. Each generated image is evaluated by two subjects, so the total score of the first three questions is 100. More details are in Appendix H. + +Table 1 shows the results. Our method achieves significantly better performance in most cases, especially on GPT-synthetic dataset that contains multiple objects and relations + +in the same description. The results show that our method is effective at generating images with high fidelity without sacrificing perceptual quality. + +# 4.2. Additional Analyses + +Performance on Each Text Complexity Level: To explore the capability of our method, especially on complex text, we + +![](images/d796dae3489d481c1571edf8944d9d15c507bbfbaed7dcf179d87fb093077814.jpg) +Figure 4. Performance when the number of objects and spatial relations in the text increase. + +![](images/35af5a19de2abd13e113829de562d61d465f6a792dc6f017d5ffd6384fae1ee7.jpg) + +![](images/c1a802fc7ed665cc7b166206e6631b9eb9c90dac8fd04de56572c1db3fb18305.jpg) +Figure 5. Example images generated by our method and baselines on novel object combinations. + +analyze the performance at each complexity level. Specifically, we use the GPT-synthetic dataset where multiple objects and relations can appear in the same description. We use the number of objects $N$ and number of spatial relations $M$ introduced in Sec. 3.3 as proxies for text complexity and plot the performance of each method as the text becomes more complex. i.e., we plot the performance for each value of $N$ regardless of $M$ , and vise versa for $M$ . As shown in Fig. 4, the performance of all methods decreases as the text becomes more complex, but our method still outperforms or is on par with others on complex descriptions. Note that PAINT-WITH-WORDS uses the same layout predictor as ours, which partially explains its strong performance. + +Generalizability to novel object combinations is a critical requirement for text-to-image models. However, since MSCOCO dataset is collected from real photos, most of its captions contain object pairs and object-attribute pairs that are common in daily life, which are also more likely to overlap with stable diffusion's pre-training data. Thus we additionally evaluate our method on another synthetic dataset that is created similarly to Sec. 3.3, but contains uncommon object pairs, object-attribute pairs, and object spatial relations. Fig. 5 shows sample images generated by our method and baselines. It can be observed that our method successfully generates novel object pairs (e.g., "zebra is looking at computer"). Performance on this dataset and more generated examples are shown in Appendix F. Overall, our method is able to generalize to novel object combinations. + +
Object RecallSPRel Precision
Ours47.2%66.7%
No spatial control39.6%51.7%
No temporal control43.8%61.4%
No optimization41.0%55.6%
+ +Table 3. Ablation performance on GPT-synthetic test set. + +# 4.3. Ablation Study + +In this section, we will investigate the influence of two important steps in our method, namely, the layout predictor and the spatial-temporal attention. We first study how spatial-temporal attention affects our performance. To do that, we consider three variants of attention. + +First, we explore the attention without spatial control. Concretely, the diffusion model still attends to both global description and local descriptions. However, instead of only attending to a local description in the pixel region of the object, the model now attends to all local descriptions in the whole image. The combination weights at each denoising step are optimized with only the global CLIP similarity in Sec. 3.4. Table 3 shows that its performance drops drastically, demonstrating the importance of spatial control. + +Second, we explore the attention without temporal control, where combination weights remain the same for all denoising steps. Note that the weights can be different for each object and are optimized with $\mathcal{L}_{\mathrm{attend}}(\lambda)$ in Sec. 3.4. The results in Table 3 show a significant degradation under both metrics, indicating the benefits of temporal dependency on attention. + +Finally, we explore the attention without optimization. Specifically, we fix the combination weight $\lambda_{it} = \frac{1}{N}$ for all $i$ and $t$ , where $N$ is the number of objects. This value is also the initialization point of $\{\lambda_{it}\}$ in our method. As shown in Table 3, the performance further drops compared to the no temporal control variant. The results indicate that optimizing the combination weights for a new text description is critical for generating images with high fidelity. + +We further study the effects of our layout predictor, which includes the layout predictor trained with only one of the absolute and relative position objectives, the comparison with ground truth and user-provided pixel region, and a different strategy to construct the pixel region for an object. Results of these experiments are presented in Appendix E. + +# 5. Conclusion + +In this work, we study the text-to-image synthesis task based on diffusion models. We find existing methods lack explicit control on cross-attention in diffusion models, which leads to the generation of low-fidelity images. We propose an algorithm that imposes control on cross-attention in spatial and temporal aspects. Experiments show our method outperforms baselines in generating high + +fidelity images. Further ablation study verifies the effectiveness of our spatial and temporal attention control. One limitation of our method is the reliance on a time-consuming optimization scheme, which takes around 10 minutes for each text-to-image generation. Future work may consider striking a better balance between performance and efficiency. + +# 6. Acknowledgement + +The work of Qiucheng Wu, Yujuan Liu, and Shiyu Chang was partially supported by Cisco Faculty Award. + +# References + +[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein gan, 2017. 2 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv, 2022. 2, 3, 5 +[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis, 2018. 2 +[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 4, 12 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers, 2020. 6 +[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 2 +[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2 +[9] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 3 +[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv, 2022. 2, 3, 5 + +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Neurips, 2014. 2 +[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Neurips, 2020. 1, 2 +[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., pages 1-33, 2022. 2 +[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2 +[17] Nisha Huang, Fan Tang, Weiming Dong, Tong-Yee Lee, and Changsheng Xu. Region-aware diffusion for zero-shot text-driven image editing. arXiv preprint arXiv:2302.11797, 2023. 3 +[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2017. 2 +[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models, 2022. 1, 2 +[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2014. 5, 13 +[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2 +[23] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. arXiv preprint arXiv:2205.14217, 2022. 2 +[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 3 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 12 +[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. arXiv preprint arXiv:2205.00363, 2022. 2, 5 +[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5 + +[28] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 3, 5 +[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv, 2019. 5, 12, 13 +[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2 +[31] Gary Marcus, Ernest Davis, and Scott Aaronson. A very preliminary analysis of dall-e 2, 2022. 2, 3 +[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2 +[34] Dong Huk Park, Grace Luo, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. arXiv preprint arXiv:2212.00210, 2022. 3 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 12, 13 +[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[38] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Neurips, 2019. 2 +[39] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 12, 13 +[41] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2 +[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image + +diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3 +[43] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 5 +[45] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 3 +[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning. PMLR, 2015. 1, 2 +[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2 +[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4 +[50] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 2, 3 +[51] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022. 2 +[52] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: a geometric diffusion model for molecular conformation generation, 2022. 2 +[53] Cheng-Fu Yang, Wan-Cyuan Fan, Fu-En Yang, and Yu-Chiang Frank Wang. Layouttransformer: Scene layout generation with conceptual and spatial diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3732-3741, June 2021. 4 +[54] Zuopeng Yang, Daqing Liu, Chaoyue Wang, Jie Yang, and Dacheng Tao. Modeling image composition for complex scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7764-7773, 2022. 3 +[55] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2 + +[56] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collomosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. arXiv preprint arXiv:2211.11742, 2022. 3 +[57] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3 +[58] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 2 \ No newline at end of file diff --git a/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/images.zip b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a42b00e44ea3e09d0fc2606e48440f777c977d2 --- /dev/null +++ b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95b3fe129aa42e25d782642567c8aeda536209dcc05815fc3d9c8c5612bde21d +size 763522 diff --git a/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/layout.json b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a8f4f3d2f715e3f9a3127a04cf1b26631834ae0 --- /dev/null +++ b/2023/Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis/layout.json @@ -0,0 +1,10241 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 545, + 140 + ], + "type": "text", + "content": "Harnessing the Spatial-Temporal Attention of Diffusion Models for High-Fidelity Text-to-Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": "Qiucheng Wu" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": ", Yujuan Liu" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": ", Handong Zhao" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": ", Trung Bui" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": ", Zhe Lin" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": ", Yang Zhang" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": ", Shiyu Chang" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": ", UC Santa Barbara, " + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": "Adobe Research, " + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 160, + 534, + 189 + ], + "type": "text", + "content": "MIT-IBM Watson AI Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 215, + 191, + 378, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 191, + 378, + 203 + ], + "spans": [ + { + "bbox": [ + 215, + 191, + 378, + 203 + ], + "type": "text", + "content": "{qiucheng, yujianliu}@ucsb.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 245, + 128, + 322 + ], + "blocks": [ + { + "bbox": [ + 57, + 234, + 121, + 243 + ], + "lines": [ + { + "bbox": [ + 57, + 234, + 121, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 234, + 121, + 243 + ], + "type": "text", + "content": "StableDiffusion" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 245, + 128, + 322 + ], + "lines": [ + { + "bbox": [ + 51, + 245, + 128, + 322 + ], + "spans": [ + { + "bbox": [ + 51, + 245, + 128, + 322 + ], + "type": "image", + "image_path": "45f9aa9045642a42774a107f3724fd5ab47c48b40588e1e3865a64ec655fa4f1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 129, + 245, + 206, + 322 + ], + "blocks": [ + { + "bbox": [ + 88, + 217, + 171, + 232 + ], + "lines": [ + { + "bbox": [ + 88, + 217, + 171, + 232 + ], + "spans": [ + { + "bbox": [ + 88, + 217, + 171, + 232 + ], + "type": "text", + "content": "Missing Objects" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 234, + 179, + 243 + ], + "lines": [ + { + "bbox": [ + 157, + 234, + 179, + 243 + ], + "spans": [ + { + "bbox": [ + 157, + 234, + 179, + 243 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 245, + 206, + 322 + ], + "lines": [ + { + "bbox": [ + 129, + 245, + 206, + 322 + ], + "spans": [ + { + "bbox": [ + 129, + 245, + 206, + 322 + ], + "type": "image", + "image_path": "cc6bcb2359c4be77375ed66aa8aeb226ef0e3c7598eaded3f6c4a043640be375.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 52, + 343, + 128, + 420 + ], + "blocks": [ + { + "bbox": [ + 69, + 323, + 184, + 342 + ], + "lines": [ + { + "bbox": [ + 69, + 323, + 184, + 342 + ], + "spans": [ + { + "bbox": [ + 69, + 323, + 184, + 342 + ], + "type": "text", + "content": "An airport employee giving directions to an airplane." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 343, + 128, + 420 + ], + "lines": [ + { + "bbox": [ + 52, + 343, + 128, + 420 + ], + "spans": [ + { + "bbox": [ + 52, + 343, + 128, + 420 + ], + "type": "image", + "image_path": "cdbb9a1becc861fe770277192e5d698fe823e2ae886f39f218db105df0abce66.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 421, + 193, + 441 + ], + "lines": [ + { + "bbox": [ + 63, + 421, + 193, + 441 + ], + "spans": [ + { + "bbox": [ + 63, + 421, + 193, + 441 + ], + "type": "text", + "content": "A zebra brazing on green grass next to a pile of rocks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 121, + 444, + 134, + 454 + ], + "lines": [ + { + "bbox": [ + 121, + 444, + 134, + 454 + ], + "spans": [ + { + "bbox": [ + 121, + 444, + 134, + 454 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 455, + 545, + 477 + ], + "lines": [ + { + "bbox": [ + 46, + 455, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 545, + 477 + ], + "type": "text", + "content": "Figure 1. Example generation by stable-diffusion-v1-4 and our method. Stable diffusion model makes three types of errors that include missing objects, mismatched attributes, and mislocated objects. Errors are highlighted in red." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 129, + 343, + 206, + 420 + ], + "blocks": [ + { + "bbox": [ + 129, + 343, + 206, + 420 + ], + "lines": [ + { + "bbox": [ + 129, + 343, + 206, + 420 + ], + "spans": [ + { + "bbox": [ + 129, + 343, + 206, + 420 + ], + "type": "image", + "image_path": "6983621a8be527cf85e44ba0ddc03f077c13476df73ffa582cb5139481e6a13b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 219, + 245, + 295, + 321 + ], + "blocks": [ + { + "bbox": [ + 237, + 217, + 356, + 230 + ], + "lines": [ + { + "bbox": [ + 237, + 217, + 356, + 230 + ], + "spans": [ + { + "bbox": [ + 237, + 217, + 356, + 230 + ], + "type": "text", + "content": "Mismatched Attributes" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 234, + 288, + 243 + ], + "lines": [ + { + "bbox": [ + 225, + 234, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 225, + 234, + 288, + 243 + ], + "type": "text", + "content": "StableDiffusion" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 245, + 295, + 321 + ], + "lines": [ + { + "bbox": [ + 219, + 245, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 219, + 245, + 295, + 321 + ], + "type": "image", + "image_path": "48b025c81218bcaa972c6a3a87b46de7d20497e2a3d478bf64277ca72be978a2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 322, + 360, + 342 + ], + "lines": [ + { + "bbox": [ + 228, + 322, + 360, + 342 + ], + "spans": [ + { + "bbox": [ + 228, + 322, + 360, + 342 + ], + "type": "text", + "content": "A bedroom has green carpeting and a bed." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 301, + 245, + 377, + 321 + ], + "blocks": [ + { + "bbox": [ + 325, + 234, + 346, + 243 + ], + "lines": [ + { + "bbox": [ + 325, + 234, + 346, + 243 + ], + "spans": [ + { + "bbox": [ + 325, + 234, + 346, + 243 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 245, + 377, + 321 + ], + "lines": [ + { + "bbox": [ + 301, + 245, + 377, + 321 + ], + "spans": [ + { + "bbox": [ + 301, + 245, + 377, + 321 + ], + "type": "image", + "image_path": "d2a7ab16c264f3ec31d1f9affa32f4f5c694a4050d6fc9e1fec25dd2f66a5b30.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 219, + 343, + 295, + 419 + ], + "blocks": [ + { + "bbox": [ + 219, + 343, + 295, + 419 + ], + "lines": [ + { + "bbox": [ + 219, + 343, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 219, + 343, + 295, + 419 + ], + "type": "image", + "image_path": "ad7aa174e9b3ef49d079b8b31cdad4f343d105baf8df8bc5a188d06f6ece4162.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 420, + 351, + 441 + ], + "lines": [ + { + "bbox": [ + 242, + 420, + 351, + 441 + ], + "spans": [ + { + "bbox": [ + 242, + 420, + 351, + 441 + ], + "type": "text", + "content": "A red car is to the left of a black mailbox." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 289, + 444, + 303, + 454 + ], + "lines": [ + { + "bbox": [ + 289, + 444, + 303, + 454 + ], + "spans": [ + { + "bbox": [ + 289, + 444, + 303, + 454 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 384, + 245, + 462, + 321 + ], + "blocks": [ + { + "bbox": [ + 413, + 217, + 512, + 232 + ], + "lines": [ + { + "bbox": [ + 413, + 217, + 512, + 232 + ], + "spans": [ + { + "bbox": [ + 413, + 217, + 512, + 232 + ], + "type": "text", + "content": "Mislocated Objects" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 234, + 457, + 243 + ], + "lines": [ + { + "bbox": [ + 394, + 234, + 457, + 243 + ], + "spans": [ + { + "bbox": [ + 394, + 234, + 457, + 243 + ], + "type": "text", + "content": "StableDiffusion" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 245, + 462, + 321 + ], + "lines": [ + { + "bbox": [ + 384, + 245, + 462, + 321 + ], + "spans": [ + { + "bbox": [ + 384, + 245, + 462, + 321 + ], + "type": "image", + "image_path": "2728eb6bde0414eca02cdb5fea9c7104788bfadcfdd8bacd24ae4c2ff70bf104.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 322, + 529, + 342 + ], + "lines": [ + { + "bbox": [ + 397, + 322, + 529, + 342 + ], + "spans": [ + { + "bbox": [ + 397, + 322, + 529, + 342 + ], + "type": "text", + "content": "The brown teddy bear is placed high above the toilet." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 466, + 245, + 542, + 321 + ], + "blocks": [ + { + "bbox": [ + 495, + 234, + 516, + 243 + ], + "lines": [ + { + "bbox": [ + 495, + 234, + 516, + 243 + ], + "spans": [ + { + "bbox": [ + 495, + 234, + 516, + 243 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 245, + 542, + 321 + ], + "lines": [ + { + "bbox": [ + 466, + 245, + 542, + 321 + ], + "spans": [ + { + "bbox": [ + 466, + 245, + 542, + 321 + ], + "type": "image", + "image_path": "09756c482440a87c372aefc049ea576bbb437960ff77ce730399549d4028f02d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 386, + 343, + 463, + 419 + ], + "blocks": [ + { + "bbox": [ + 386, + 343, + 463, + 419 + ], + "lines": [ + { + "bbox": [ + 386, + 343, + 463, + 419 + ], + "spans": [ + { + "bbox": [ + 386, + 343, + 463, + 419 + ], + "type": "image", + "image_path": "b9bb0535bb90fae8f90f4bb47adbf09e186ab4c93955e63b4e46627d318567d4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 420, + 524, + 441 + ], + "lines": [ + { + "bbox": [ + 402, + 420, + 524, + 441 + ], + "spans": [ + { + "bbox": [ + 402, + 420, + 524, + 441 + ], + "type": "text", + "content": "The teddy bear is right of the suitcase." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 444, + 469, + 454 + ], + "lines": [ + { + "bbox": [ + 457, + 444, + 469, + 454 + ], + "spans": [ + { + "bbox": [ + 457, + 444, + 469, + 454 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 466, + 343, + 543, + 419 + ], + "blocks": [ + { + "bbox": [ + 466, + 343, + 543, + 419 + ], + "lines": [ + { + "bbox": [ + 466, + 343, + 543, + 419 + ], + "spans": [ + { + "bbox": [ + 466, + 343, + 543, + 419 + ], + "type": "image", + "image_path": "35027d1451bd4998133b64cb67481385f68e5cd83536d0631ff49149869bf63f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 143, + 486, + 191, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 486, + 191, + 498 + ], + "spans": [ + { + "bbox": [ + 143, + 486, + 191, + 498 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 45, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 522, + 288, + 713 + ], + "type": "text", + "content": "Diffusion-based models have achieved state-of-the-art performance on text-to-image synthesis tasks. However, one critical limitation of these models is the low fidelity of generated images with respect to the text description, such as missing objects, mismatched attributes, and mislocated objects. One key reason for such inconsistencies is the inaccurate cross-attention to text in both the spatial dimension, which controls at what pixel region an object should appear, and the temporal dimension, which controls how different levels of details are added through the denoising steps. In this paper, we propose a new text-to-image algorithm that adds explicit control over spatial-temporal cross-attention in diffusion models. We first utilize a layout predictor to predict the pixel regions for objects mentioned in the text. We then impose spatial attention control by combining the attention over the entire text description and that over the" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "type": "text", + "content": "local description of the particular object in the corresponding pixel region of that object. The temporal attention control is further added by allowing the combination weights to change at each denoising step, and the combination weights are optimized to ensure high fidelity between the image and the text. Experiments show that our method generates images with higher fidelity compared to diffusion-model-based baselines without fine-tuning the diffusion model. Our code is publicly available.1" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 618, + 386, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 386, + 631 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 386, + 631 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "type": "text", + "content": "Diffusion models [14, 19, 46, 47, 48] have recently revolutionized the field of image synthesis. Compared with previous generative models such as generative adversar" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 684, + 385, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 684, + 385, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 684, + 385, + 693 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 318, + 694, + 486, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 694, + 486, + 703 + ], + "spans": [ + { + "bbox": [ + 318, + 694, + 486, + 703 + ], + "type": "text", + "content": "1https://github.com/UCSB-NLP-Chang/" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 307, + 704, + 424, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 704, + 424, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 704, + 424, + 712 + ], + "type": "text", + "content": "Diffusion-SpaceTime-Attn" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7766" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 168 + ], + "type": "text", + "content": "ial networks [1, 3, 11, 18] and variational autoencoders [21, 38, 39], diffusion models have demonstrated superior performance in generating images with higher quality, more diversity, and better control over generated contents. Particularly, text-to-image diffusion models [2, 33, 36, 40, 42] allow generating images conditioned on a text description, which enables generation of creative images due to the expressiveness of natural language." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 337 + ], + "type": "text", + "content": "However, recent studies [10, 31] have revealed that one critical limitation of existing diffusion-model-based text-to-image algorithms is the low fidelity with respect to the text descriptions – the content of the generated image is sometimes at odds with the text description, especially when the description is complex. Specifically, typical errors made by stable diffusion models fall into three categories: missing objects, mismatched attributes, and mislocated objects. For example, in Fig. 1(a), stable diffusion model ignores the airplane even though it is mentioned in text; in Fig. 1(b), the model confuses “red car” and “black mailbox” and generates a red mailbox; in Fig. 1(c), the model locates teddy bear behind the toilet, despite the description “teddy bear is placed high above the toilet.”" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 616 + ], + "type": "text", + "content": "Such infidelity problems suggest that the cross-attention map on the text description may not be accurate. In particular, if we view the generation process of a diffusion model as a sequence of denoising steps, then the cross-attention on text descriptions can be considered as a function of both spatial (pixels) and temporal (denoising steps) information. Therefore, the inaccuracies of the cross-attention can result from the loose control over both the spatial and temporal dimensions. On one hand, spatial attention controls at what pixels the model should attend to each object and the corresponding attributes mentioned in the text. If the spatial attention is incorrect, the resulting images will have incorrect object locations or miss-associated attributes. On the other hand, temporal attention controls when the models should attend to different levels of details in the text. As previous works have revealed, diffusion models tend to focus on generating object outlines at earlier denoising steps and on details at later [50]. Thus loose control over the temporal aspect of attention can easily lead to overlooking certain levels of the object details. In short, to improve the fidelity of text-to-image synthesis, one would need to explicitly control both spatial and temporal attention to follow an accurate and optimal distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": "In this paper, we propose a new text-to-image algorithm based on a pre-trained conditional diffusion model with explicit control over the spatial-temporal cross-attention map on text. The proposed algorithm introduces a layout predictor and a spatial-temporal attention optimizer. The layout predictor takes the text description as input and generates a spatial layout for each object mentioned in the text. Alternatively, the layout can also be provided by the user. Then" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": "the spatial-temporal attention optimizer imposes direct control over the spatial and temporal aspects of the attention according to the spatial layout. In particular, for the spatial aspect, we parameterize the attention map such that the attention outputs in the designated pixel region for an object are a weighted combination of attention over the entire text description and that over the local description that specifically describes the corresponding object. In this way, we manage to emphasize the attention over the object descriptions. For the temporal aspect, we allow the combination weights to change across time and optimized according to a CLIP objective that measures the agreement between the generated images and the text description. In this way, we allow the attention to focus more on the entire description at the early stage and gradually shift to the detailed local descriptions as the denoising process proceeds. The entire pipeline resembles a typical painting process of a human, where each object's position is determined beforehand and the focus gradually shifts from global information to the local details of each object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 313, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 546, + 456 + ], + "type": "text", + "content": "We conduct extensive experiments on datasets that contain real and template-based captions [6, 26] and our newly created synthetic dataset that contains complex text descriptions. Results show that our method generates images that better align with descriptions compared to other stable diffusion-based baselines. As shown in Fig. 1, our method effectively resolves the above-mentioned three errors. Particularly, controlling spatial attention locates objects at the desired position, and controlling temporal attention promotes the occurrence of objects with associated attributes. Our findings shed light on fine-grained control of diffusion models in text-to-image generation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 472, + 392, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 392, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 392, + 484 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 493, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 673 + ], + "type": "text", + "content": "Diffusion Models Diffusion models are a class of generative models that have demonstrated state-of-the-art performance on image synthesis tasks [8, 14, 19, 46, 47, 48]. These models synthesize images by sampling a noisy image from the standard Gaussian distribution and iteratively denoising it back to a clean image. Their impressive performance has advanced research in multiple computer vision areas including inpainting [30, 41, 51], image editing [7, 12, 32, 50], super-resolution [15, 43], video synthesis [13, 16, 55], and applications beyond computer vision [22, 23, 52]. Among these, text-to-image diffusion models have gained significant attention [2, 36, 37, 42, 58]. Taking text descriptions as inputs, these models generate high quality images that are semantically aligned with the input text, which have led to many creative and artistic applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Enhancing the Controllability of Text-to-Image Diffusion Models While diffusion-based text-to-image models have shown promising results, recent works have high" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7767" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 71, + 494, + 228 + ], + "blocks": [ + { + "bbox": [ + 98, + 71, + 494, + 228 + ], + "lines": [ + { + "bbox": [ + 98, + 71, + 494, + 228 + ], + "spans": [ + { + "bbox": [ + 98, + 71, + 494, + 228 + ], + "type": "image", + "image_path": "e7cfe7163e33d910370800bc7d71acf9a9738e94cdedcb0c9cbe2930f223a17f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "text", + "content": "Figure 2. Overview of our text-to-image generation pipeline at one denoising step. Given input text " + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "text", + "content": ", we first parse it and extract all objects mentioned, constructing local descriptions " + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "text", + "content": ". Next, the layout predictor predicts the pixel region for each object in the text. The diffusion model attends to the global description " + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "text", + "content": " and additionally attends to the local description " + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "text", + "content": " in object " + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "text", + "content": "'s region. The final attention output is a weighted combination of attention to both global and local descriptions, where the combination weights sum up to 1 for each pixel and are optimized for each denoising step to achieve a high fidelity with " + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 231, + 547, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 305, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 289, + 605 + ], + "type": "text", + "content": "lighted cases where models fail to generate high-fidelity images with respect to the input text [10, 31]. To this end, prior works have explored various ways to enhance the controllability of text-to-image diffusion models. One line of work enhances the controllability by improving diffusion models' ability to understand natural language, which includes using a more powerful text encoder that is separately trained on language modeling tasks [2, 42], incorporating linguistic structures in the text to guide the cross-attention between image and text [10], and decomposing a complex text description into multiple components that are easy to generate [28]. Another line of work conditions diffusion models' generation on auxiliary inputs such as object layout [2, 9, 24, 54] and silhouette [17, 34, 45, 56, 57]. By modifying diffusion models' attention operation according to these auxiliary information or directly fine-tuning diffusion models to take these auxiliary inputs, they are able to control the location and shape of the objects in the image. Finally, some work adds temporal aspect control on diffusion models by modifying the input text condition at each denoising step, which allows them to disentangle a desired attribute from other contents [50]. Different from prior works, our method imposes both spatial and temporal control in cross-attention layer. Moreover, our method does not require auxiliary inputs and does not fine-tune the diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 616, + 129, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 129, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 129, + 629 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 171, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 171, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 171, + 647 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "We focus on the standard text-to-image problem. Given a text description, denoted as " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", our goal is to generate an image that is consistent with " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ". For a concrete exposition, we will use an example description \"A red car is to the left of a black mailbox\" in the following. Our work aims to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 305, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 545, + 331 + ], + "type": "text", + "content": "improve the fidelity of the generated image to text, which includes three requirements:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 338, + 545, + 456 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 307, + 338, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 373 + ], + "type": "text", + "content": "- Object Fidelity: The generated images should include all the objects mentioned in " + }, + { + "bbox": [ + 307, + 338, + 545, + 373 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 307, + 338, + 545, + 373 + ], + "type": "text", + "content": ". In our example, the generated image should contain a car and a mailbox." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 379, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 379, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 379, + 545, + 415 + ], + "type": "text", + "content": "- Attribute Fidelity: The attributes of each object in the image should match those in " + }, + { + "bbox": [ + 307, + 379, + 545, + 415 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 307, + 379, + 545, + 415 + ], + "type": "text", + "content": ". In our example, the car should be red and the mailbox should be black." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 420, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 456 + ], + "type": "text", + "content": "- Spatial Fidelity: The relative spatial positions of the object should match the description in " + }, + { + "bbox": [ + 307, + 420, + 545, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 307, + 420, + 545, + 456 + ], + "type": "text", + "content": ". In our example, the car should be on the left and the mailbox on the right." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 464, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 488 + ], + "type": "text", + "content": "We will tackle these problems using a pre-trained, fixed stable diffusion model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 496, + 412, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 412, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 412, + 508 + ], + "type": "text", + "content": "3.2. Method Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": "To achieve high-fidelity text-to-image generation, we propose an algorithm consisting of the following four steps. Step 1: Object Identification We extract all the objects mentioned in " + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "inline_equation", + "content": "O_{1:N}" + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": ", by eliciting the noun phrases using spaCy. " + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "inline_equation", + "content": "^2 N" + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": " is the total number of objects. In our sample, " + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "inline_equation", + "content": "O_1 = \\text{\"a red car\"}" + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "inline_equation", + "content": "O_2 = \\text{\"a black mailbox\"}" + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": ". Step 2: Layout Prediction For each object " + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": ", we use a layout predictor to predict its pixel region, " + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 304, + 515, + 545, + 634 + ], + "type": "text", + "content": ", which is a set of pixels roughly specifying where the object should be. The layout can also be provided by human users." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": "Step 3: Local Description Generation For each object " + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": ", we generate a local text description, " + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": ", containing only that object information using a simple template. In our example, there are two local descriptions. " + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "inline_equation", + "content": "L_{1} =" + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": " \"A photo of a red car\" and " + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "inline_equation", + "content": "L_{2} =" + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": " \"A photo of a black mailbox\"." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 407, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 407, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 407, + 713 + ], + "type": "text", + "content": "2https://spacy.io/." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7768" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 132 + ], + "type": "text", + "content": "Step 4: Attention Optimization During the generation process, we guide the diffusion model to combine attention to both the global description " + }, + { + "bbox": [ + 47, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 72, + 286, + 132 + ], + "type": "text", + "content": " and the local descriptions " + }, + { + "bbox": [ + 47, + 72, + 286, + 132 + ], + "type": "inline_equation", + "content": "L_{1:N}" + }, + { + "bbox": [ + 47, + 72, + 286, + 132 + ], + "type": "text", + "content": " according to the object layout. The attention combination weights are optimized for each input description." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 286, + 156 + ], + "type": "text", + "content": "The following subsections will provide further details about steps 2 and 4." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 149, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 149, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 149, + 175 + ], + "type": "text", + "content": "3.3. Layout Predictor" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": "Our layout predictor is adapted from [53], which aims to predict the center coordinate " + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "inline_equation", + "content": "C_i = [X_i, Y_i]" + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": " for each object " + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "inline_equation", + "content": "O_i^3" + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": ". Specifically, the layout predictor is a transformer that takes the text description " + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": " as the input. At the output position where the input mentions the object " + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": ", we let the transformer output a set of Gaussian mixture model (GMM) parameters to fit the object's center coordinate " + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": ". Formally, denote " + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "inline_equation", + "content": "f_i(D; \\theta)" + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": " as the output of the layout predictor at the location of the mentioning object " + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 47, + 181, + 286, + 289 + ], + "type": "text", + "content": ". Then" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 295, + 286, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 295, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 104, + 295, + 286, + 324 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {i} (\\boldsymbol {D}; \\boldsymbol {\\theta}) = \\bigcup_ {k = 1} ^ {K} \\left\\{\\boldsymbol {\\mu} _ {i k}, \\boldsymbol {\\Sigma} _ {i k}, w _ {i k} \\right\\}, \\tag {1}", + "image_path": "9235984cffeaeb81c0b02ddb885fe999debac1a70ce642d636d30fcc134ff390.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{ik}" + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{ik}" + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "inline_equation", + "content": "w_{ik}" + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "text", + "content": " denote the mean, covariance matrix and the prior probability of mixture " + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 47, + 330, + 286, + 365 + ], + "type": "text", + "content": " denotes the network parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 365, + 286, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 286, + 413 + ], + "type": "text", + "content": "To improve the fidelity of the object positions in the predicted layout, we introduce a hybrid training objective including an absolute position objective and a relative position objective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "text", + "content": "Absolute Position Objective The absolute position objective provides direct supervision of the exact position of each object. Formally, we assume access to an image captioning dataset " + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{real}}" + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "text", + "content": " with description " + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "text", + "content": " as well as the extracted labels of all the objects " + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "inline_equation", + "content": "\\{O_i\\}" + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "text", + "content": " and their center coordinates " + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "inline_equation", + "content": "\\{C_i\\}" + }, + { + "bbox": [ + 47, + 421, + 286, + 516 + ], + "type": "text", + "content": ". Then the training objective is to minimize the negative log-likelihood of the ground-truth coordinates under the predicted GMM distribution, i.e." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 522, + 286, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 522, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 71, + 522, + 286, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a b s}} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {D}, \\left\\{\\boldsymbol {C} i \\right\\} \\sim \\mathcal {D} _ {\\mathrm {r e a l}}} \\left[ \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {n l l}} \\left(\\boldsymbol {C} _ {i}; \\boldsymbol {f} _ {i} (\\boldsymbol {D}; \\boldsymbol {\\theta})\\right) \\right], \\tag {2}", + "image_path": "6723a67e068f955d4b190db953063edbd432e6f262269675e1ff208384b1acce.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{nll}}(C_i;f_i)" + }, + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "text", + "content": " denotes the negative log-likelihood of ground truth " + }, + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "text", + "content": " under the GMM specified by " + }, + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 557, + 286, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 588, + 286, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 286, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 286, + 659 + ], + "type": "text", + "content": "Relative Position Objective In many cases, the text description only mentions the relative positions of the objects, and thus it is more important to ensure the relative position of the predicted position is correct than the absolute position. To further enforce the fidelity of the relative positions, we introduce the following relative position objective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "type": "text", + "content": "To start with, we construct a synthetic dataset, " + }, + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{syn}}" + }, + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "type": "text", + "content": " which consists of text description " + }, + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "type": "text", + "content": " with explicit descriptions of relative spatial relations. We first randomly select " + }, + { + "bbox": [ + 47, + 661, + 286, + 696 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "text", + "content": "objects with attribute modifiers. We then select " + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "text", + "content": " pairs of objects to specify their relative spatial relation. For object pair " + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "(O_i, O_j)" + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "text", + "content": ", their spatial relation, " + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "R_{ij}" + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "text", + "content": ", is randomly drawn from \"left of\", \"right of\", \"above\" and \"below\". Finally, we prompt GPT-3 [4] to generate the text description " + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 72, + 545, + 167 + ], + "type": "text", + "content": " that mentions all the objects and their relative positions. Our familiar \"A red car is to the left of a black mailbox\" is one such example. Further details are provided in Appendix B." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "text", + "content": "Next, we introduce a loss to penalize violating the relative position. If object " + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "text", + "content": " is to the left of object " + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "text", + "content": ", we enforce that the rightmost mixture mean of object " + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "text", + "content": " is to the left of the leftmost mixture mean of object " + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 168, + 545, + 216 + ], + "type": "text", + "content": ". Formally," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 222, + 550, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 550, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 550, + 248 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {r e l}} \\left(R _ {i j} = ^ {\\prime \\prime} \\text {l e f t} ^ {\\prime \\prime}; \\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right) = \\max \\left\\{\\max _ {k} \\boldsymbol {\\mu} _ {i k} (0) - \\min _ {k} \\boldsymbol {\\mu} _ {j k} (0), - \\delta \\right\\}, \\tag {3}", + "image_path": "b145dfca20064c24424529c4871175aea85fedf4b212643dc1019031049a4980.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\mu_{ik}(0)" + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "content": " denotes the zeroth element (the " + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "content": "-coordinate) of " + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\mu_{ik}" + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "content": " is a pre-specified margin. The " + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{rel}}" + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "content": " of the other three types of relations are defined similarly. The relative position loss is thus the aggregation of " + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{rel}}" + }, + { + "bbox": [ + 305, + 249, + 545, + 307 + ], + "type": "text", + "content": " across the synthetic dataset:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 312, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 312, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 320, + 312, + 545, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e l}} (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\boldsymbol {D}, \\left\\{R _ {i j} \\right\\} \\sim \\mathcal {D} _ {\\mathrm {s y n}}} \\left[ \\sum_ {i, j: R _ {i j} \\neq \\emptyset} \\ell_ {\\mathrm {r e l}} \\left(R _ {i j}; \\boldsymbol {f} _ {i}, \\boldsymbol {f} _ {j}\\right) \\right], \\tag {4}", + "image_path": "b9eae462ed2d0b4312ac407b0bf53a08f0e06ab991f562ccf331566a3637f6a6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 345, + 428, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 428, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 428, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 345, + 428, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i" + }, + { + "bbox": [ + 306, + 345, + 428, + 357 + ], + "type": "text", + "content": " is short for " + }, + { + "bbox": [ + 306, + 345, + 428, + 357 + ], + "type": "inline_equation", + "content": "f_{i}(D;\\theta)" + }, + { + "bbox": [ + 306, + 345, + 428, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 364, + 544, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 544, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 544, + 387 + ], + "type": "text", + "content": "Training and Inference To sum up, the final training objective is the combination of both:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 395, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 395, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 362, + 395, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a y o u t}} (\\boldsymbol {\\theta}) = \\mathcal {L} _ {\\text {a b s}} (\\boldsymbol {\\theta}) + \\xi \\mathcal {L} _ {\\text {r e l}} (\\boldsymbol {\\theta}), \\tag {5}", + "image_path": "c85eaff8e1cfefa642b6d1b82f61ee2c71269c8ae41fdec76a78dd70a5d763fd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": " is a hyperparameter. During inference time, we randomly draw the center coordinate " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": " from the predicted GMM. The pixel region for object " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": ", is defined as a circular region centered at the drawn " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": " with a fixed radius " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": ". As would be shown in Appendix D, " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": " only roughly regulates the position of the generated objects, and the actual size of the object can go beyond or below the size of " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": ". Thus a fixed " + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 415, + 545, + 510 + ], + "type": "text", + "content": " is sufficient for this purpose." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 517, + 522, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 522, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 522, + 530 + ], + "type": "text", + "content": "3.4. Spatial-Temporal Attention Optimization" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "content": "Recall that " + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "content": " is the global description and " + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\{L_i\\}" + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "content": " are local descriptions for each object. Our goal is to guide the diffusion model to attend to not only " + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "content": ", but also " + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "content": " in object " + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "content": "’s region, " + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 305, + 536, + 545, + 596 + ], + "type": "text", + "content": ", so that the model is more strongly prompted to generate the specific object as specified in the layout." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 603, + 544, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 603, + 544, + 626 + ], + "spans": [ + { + "bbox": [ + 305, + 603, + 544, + 626 + ], + "type": "text", + "content": "Spatial-Temporal Attention Recall that the standard cross-attention [49] is defined as" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 335, + 632, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 632, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 335, + 632, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} \\left(\\boldsymbol {Q} \\boldsymbol {K} ^ {T} / \\sqrt {d}\\right) \\boldsymbol {V}, \\tag {6}", + "image_path": "fad020581786a74dd241eb8044d2c9820db93ffc207545d89ef61ae8302d5528.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{h \\times w, d}" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " is the queries vectors for each pixel, and " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "K, V \\in \\mathbb{R}^{l, d}" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " are key and value vectors for each text description token; " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " represent the image height and width, " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " represents the text description length; " + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 653, + 545, + 712 + ], + "type": "text", + "content": " represents the dimension of each attention vector." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "text", + "content": "3More details of the layout predictor can be found in Appendix D." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7769" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "Since we have multiple text descriptions, we define " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "K_{D}, V_{D}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " as the key and value vectors for the global description " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "K_{L_i}, V_{L_i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " for the local description " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "L_i" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ". Further, we introduce a set of binary mask matrices " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\{M_i\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " to indicate the region for each object, i.e." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 137, + 287, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 137, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 72, + 137, + 287, + 148 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {i} (x, y) = 1, \\text {i f} (x, y) \\in \\mathcal {R} _ {i}, \\quad \\text {a n d} 0, \\text {o t h e r w i s e .} \\tag {7}", + "image_path": "cb31f242bcb4b05d70c81bf5dc18fce9f8d8cfe296a0c03a047fd8080ab135aa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "text", + "content": "Then the output of the attention layer of the denoising network at time " + }, + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 180, + 287, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 287, + 240 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {O} (t) = \\sum_ {i = 1} ^ {N} \\lambda_ {i t} \\boldsymbol {M} _ {i} \\odot \\operatorname {A t t e n t i o n} (\\boldsymbol {Q}, \\boldsymbol {K} _ {\\boldsymbol {L} _ {i}}, \\boldsymbol {V} _ {\\boldsymbol {L} _ {i}}) \\tag {8} \\\\ + \\left(1 - \\sum_ {i = 1} ^ {N} \\lambda_ {i t} M _ {i}\\right) \\odot \\operatorname {A t t e n t i o n} \\left(\\boldsymbol {Q}, \\boldsymbol {K} _ {D}, \\boldsymbol {V} _ {D}\\right), \\\\ \\end{array}", + "image_path": "149fc7a678eb0b9e0a59754889af4445c1a417cfa87b846f5e4d4ec7d608f1de.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{it}" + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "text", + "content": " are the attention combination weights, and " + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "text", + "content": " denotes element-wise multiplication. Note that the combination weights are functions of time " + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "text", + "content": ", which is motivated by the observation that different denoising steps control the generation of different levels of details. When synthesizing images, the combination weights " + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{it}" + }, + { + "bbox": [ + 46, + 242, + 287, + 361 + ], + "type": "text", + "content": " are optimized on the fly for each corresponding mask and time. By introducing this time dependency, we allow the diffusion model to focus more on the global description at earlier time steps and shift to local descriptions later." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": "Optimization Objective All the attention combination weights, " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\lambda = \\{\\lambda_{it}\\}" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": ", are determined by maximizing the consistency between the generated image and the text description as measured by the CLIP similarity [35]. We introduce two CLIP similarities, a global CLIP similarity that compares the entire image and the global description, and a set of local CLIP similarities that compare the images at each object region and the corresponding local descriptions. Formally, denote the generated image as " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "I(\\lambda)" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": ", and its local patch at each object's region as " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "I_{O_i}(\\lambda)" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": ". Note that we adopt the deterministic PLMS denoising process [27] (i.e., with " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\sigma = 0" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": "), so both " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "I_{O_i}" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": " are deterministic functions of " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": ". Then the loss for optimizing " + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 369, + 287, + 524 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 528, + 287, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 528, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 52, + 528, + 287, + 557 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a t t e n d}} (\\boldsymbol {\\lambda}) = - \\operatorname {C L I P} (\\boldsymbol {I} (\\boldsymbol {\\lambda}), \\boldsymbol {D}) - \\gamma \\sum_ {i = 1} ^ {N} \\operatorname {C L I P} \\left(\\boldsymbol {I} _ {O _ {i}} (\\boldsymbol {\\lambda}), \\boldsymbol {L} _ {i}\\right), \\tag {9}", + "image_path": "62668831671efc1efe51cdac247508a0e3b3e77cce09ff27e3e5fbfbaab7ac90.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 561, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 287, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 561, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}(\\cdot)" + }, + { + "bbox": [ + 47, + 561, + 287, + 585 + ], + "type": "text", + "content": " denotes the CLIP similarity, and " + }, + { + "bbox": [ + 47, + 561, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 561, + 287, + 585 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 594, + 127, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 127, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 127, + 607 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 615, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 287, + 650 + ], + "type": "text", + "content": "We conduct experiments to evaluate our method's performance and generalizability. We also perform ablation study on important design choices of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 677 + ], + "type": "text", + "content": "Implementing Details: We adopt RoBERTa-base [29] as the base model for layout predictor and use 5 mixtures in" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": "GMM. We use stable-diffusion-v1-4 [40] pre-trained on the laion dataset [44] and freeze it throughout all experiments. All generated images are in the size of " + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": ". We use PLMS sampler [27] to synthesize images with 50 denoising steps, and we use Adam [20] to optimize the layout predictor and attention combination weights. More details on hyperparameters and optimization are in Appendix A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 162, + 533, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 162, + 533, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 162, + 533, + 174 + ], + "type": "text", + "content": "4.1. Evaluation on Fidelity of Generated Images" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 180, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 216 + ], + "type": "text", + "content": "We first evaluate our method on object, attribute, and spatial fidelities as introduced in Sec. 3.1 using both objective and subjective metrics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 220, + 546, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 546, + 554 + ], + "type": "text", + "content": "Baselines: We identify four baseline methods on generating images from complex text descriptions. (1) VANILLA-SD [40] is the pre-trained text-to-image stable diffusion model that directly generates images conditioned on the text description. (2) COMPOSABLE-DIFFUSION [28] is a diffusion-based compositional generation method. To generate images from a text description, the text is first decomposed into the conjunction of multiple components (e.g., for the example in Sec. 3.1, the components are \"A red car is to the left of a black mailbox.\" AND \"A red car\" AND \"a black mailbox\"). Each component is separately modeled by the diffusion model and then composed to generate an image by merging the outputs of each denoising step. (3) STRUCTURE-DIFFUSION [10] improves VANILLA-SD on generating images with better object and attribute fidelities. They do so by first extracting noun phrases at different levels of the parsing tree. The model then separately attends to each noun phrase and combines the attention outputs by taking their arithmetic mean. Different from our method, they perform cross-attention on the whole image instead of the region specific to an object. (4) PAINT-WITH-WORDS [2] assumes that users provide the pixel region of each object to be generated. Pixels in the region then increase their attention weights to the text that describes the corresponding object, and the amount of increase is determined by heuristic rules. For a fair comparison, we use the pixel region predicted by our layout predictor, and we report the performance with the ground truth region in Appendix E." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 713 + ], + "type": "text", + "content": "Datasets and Metrics: We conduct experiments on three datasets. (1) MS-COCO [25] contains photos taken by photographers and manually annotates the caption for each photo. We use the caption as the input text description. (2) VSR [26] is proposed for probing spatial understanding of vision-language models. Constructed from a subset of MS-COCO, it uses templates to generate captions that describe spatial relations in the image (e.g., \"The horse is to the left of the person\"). (3) GPT-synthetic is the synthetic dataset introduced in Sec. 3.3. We manually check the test set to filter out sentences that do not conform to the specified spatial relation. Compared with VSR, this dataset has more diverse and complex text descriptions. The final dataset contains" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "4To standardize the image patch sizes, " + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "inline_equation", + "content": "I_{O_i}" + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": " is obtained by cropping the original image with a minimum square that encompasses " + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": " and resizing it to " + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7770" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 545, + 135 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 135 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 135 + ], + "type": "table", + "html": "
MS-COCOVSRGPT-synthetic
Object (100) ↑Attribute (100) ↑OverallObject (100) ↑Spatial (100) ↑OverallObject (100) ↑Attribute (100) ↑Spatial (100) ↑Overall
VANILLA-SD666242%646648%60586038%
COMPOSABLE-DIFFUSION534830%636018%63604728%
STRUCTURE-DIFFUSION606348%546432%58545832%
PAINT-WITH-WORDS757452%545930%65525732%
Ours7574-6877-848386-
", + "image_path": "56f5840145b4d67fdc5c11200270a48d2bf3b687fdcbbac4fe1227427878f34c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 135, + 546, + 179 + ], + "lines": [ + { + "bbox": [ + 46, + 135, + 546, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 546, + 179 + ], + "type": "text", + "content": "Table 1. Subjective evaluation of our method and baselines. Best numbers are in bold. Spatial relation is not available on MS-COCO because very few of its captions contain spatial relations. Attribute is not available on VSR as its captions do not consider attribute. Object, Attribute, and Spatial show the total score of 50 evaluations, where a model with the highest fidelity would achieve a score of 100. Overall denotes the percentage of generation of each method that is rated better than our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 188, + 286, + 256 + ], + "blocks": [ + { + "bbox": [ + 47, + 188, + 286, + 256 + ], + "lines": [ + { + "bbox": [ + 47, + 188, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 286, + 256 + ], + "type": "table", + "html": "
MS-COCOVSRGPT-synthetic
Object RecallSPRel PrecisionObject RecallSPRel PrecisionObject RecallSPRel Precision
VANILLA-SD58.0%-62.1%56.0%42.4%56.8%
COMPOSABLE-DIFFUSION51.8%-60.8%72.2%34.4%52.6%
STRUCTURE-DIFFUSION61.7%-62.3%58.3%43.0%59.3%
PAINT-WITH-WORDS57.3%-63.8%47.1%45.9%53.5%
Ours69.6%-65.1%75.0%47.2%66.7%
", + "image_path": "cbece2942a7e37d5563fb38220495ccc184568324541b5cfa45eada2f83afdd5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 286, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 286, + 266 + ], + "type": "text", + "content": "Table 2. Automatic evaluation of our method and baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 267, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 286, + 300 + ], + "type": "text", + "content": "SPRel Precision: Spatial Relation Precision. Best numbers are in bold. SPRel precision is not available on MS-COCO since most captions do not have an explicit spatial relation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 314, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 350 + ], + "type": "text", + "content": "500 descriptions, and we downsample MS-COCO and VSR to have the same size. Statistics including the number of objects and spatial relations are shown in Appendix B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 350, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 350, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 45, + 350, + 287, + 590 + ], + "type": "text", + "content": "For automatic evaluations, we consider two metrics. (1) Object Recall measures the percentage of successfully synthesized objects over objects mentioned in the text. We use DETR [5] to detect objects in generated images. To calculate recall, we divide the number of detected objects in the text by the total number of objects in the text that also belong to one of the MS-COCO categories. It measures the object fidelity of generated images. (2) Spatial Relation Precision (SPRel Precision) measures the percentage of the correct spatial relations among all the relations whose corresponding objects are successfully synthesized. This metric measures the spatial fidelity of generated images. We consider the relation of left, right, above, and below because their correctness can be evaluated by comparing the bounding box centers. Qualitative examples of spatial relations beyond these four are shown in Appendix G. We also report the CLIP similarities between generated images and the input description in Appendix C. We observe that different methods have very close CLIP similarities, though our method still achieves competitive results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Results: The results are shown in Table 2. As shown in the table, our method outperforms all baselines on both object recall and SPRel precision. Specifically, compared with baselines that do not have spatial attention control (all but PAINT-WITH-WORDS), our method is significantly better on SPRel precision, which illustrates the effectiveness of spatial control and indicates that it is difficult for the stable diffusion model to generate correct spatial relations without guidance on the layout. Our method moves the burden of generating objects at the correct location from diffusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 275 + ], + "type": "text", + "content": "models to a separate layout predictor, allowing the whole pipeline to achieve better spatial fidelity. On the other hand, compared with PAINT-WITH-WORDS that does not impose the temporal attention control, our method achieves better object recall, showing that our optimized combination weights across the temporal dimension strike a better balance between local and global descriptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 277, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 546, + 469 + ], + "type": "text", + "content": "Fig. 3 demonstrates examples of text descriptions and images generated by our method and baselines. We observe that our method resolves three types of errors in baselines. First, our method alleviates the missing object issue (in the top panel). Baselines tend to focus on one object in the text and ignore other objects (e.g., focusing on the elephant and ignoring the man in the second row), whereas our method generates all objects. Second, as shown in the middle panel, our method mitigates the mismatched attribute issue. Particularly, baselines struggle when multiple objects are mentioned in the text, where they mismatch the attribute and object (e.g., red truck in the first row). Finally, the bottom panel shows examples where our method reduces the number of mislocated objects. Note that our method is effective both on the four relations and other spatial relations. More examples can be found in Appendix G." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "Subjective Evaluation: To further evaluate the fidelity and quality of the generated images, we perform a subjective evaluation on Amazon Mechanical Turk. Specifically, we randomly sample 25 text descriptions for each dataset in Table 2. Each subject was presented with the text description and corresponding image generated by a single method or a pair of methods and asked the following four questions: (1) (Object Fidelity) Does the image contain all objects mentioned in the text? (2) (Attribute Fidelity) Are all synthesized objects consistent with their characteristics described in the text (e.g., color and material)? (3) (Spatial Fidelity) Does the image locate all objects at the correct position such that the spatial relations in the text are satisfied (if an object in the relationship is missing, it is considered as an incorrect generation)? and (4) (Overall) Which image in the pair has higher fidelity with the text and has better quality? The first three questions are evaluated for each method individually with a score of 0, 1, or 2, where 2 denotes all objects/attributes/relations are correct and 0 denotes none of them is correct. The last question is evaluated on a pair of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7771" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 67, + 521, + 585 + ], + "blocks": [ + { + "bbox": [ + 66, + 67, + 521, + 585 + ], + "lines": [ + { + "bbox": [ + 66, + 67, + 521, + 585 + ], + "spans": [ + { + "bbox": [ + 66, + 67, + 521, + 585 + ], + "type": "image", + "image_path": "8c5424cc3c010909ff33ff2d09b9ab7c0a0163540720ed4d36de47b75f320f20.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 586, + 547, + 608 + ], + "lines": [ + { + "bbox": [ + 46, + 586, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 547, + 608 + ], + "type": "text", + "content": "Figure 3. Example images generated by our method and baselines. Typical errors of baselines include missing objects, mismatched attributes, and mislocated objects. Ours (1)/(2) show the results with two different random seeds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 624, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 673 + ], + "type": "text", + "content": "images, one generated by our method and the other by a baseline. Each generated image is evaluated by two subjects, so the total score of the first three questions is 100. More details are in Appendix H." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Table 1 shows the results. Our method achieves significantly better performance in most cases, especially on GPT-synthetic dataset that contains multiple objects and relations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 624, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 624, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 305, + 624, + 545, + 661 + ], + "type": "text", + "content": "in the same description. The results show that our method is effective at generating images with high fidelity without sacrificing perceptual quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 670, + 424, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 424, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 424, + 684 + ], + "type": "text", + "content": "4.2. Additional Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "Performance on Each Text Complexity Level: To explore the capability of our method, especially on complex text, we" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7772" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 68, + 168, + 154 + ], + "blocks": [ + { + "bbox": [ + 53, + 68, + 168, + 154 + ], + "lines": [ + { + "bbox": [ + 53, + 68, + 168, + 154 + ], + "spans": [ + { + "bbox": [ + 53, + 68, + 168, + 154 + ], + "type": "image", + "image_path": "d796dae3489d481c1571edf8944d9d15c507bbfbaed7dcf179d87fb093077814.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 158, + 287, + 179 + ], + "lines": [ + { + "bbox": [ + 47, + 158, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 287, + 179 + ], + "type": "text", + "content": "Figure 4. Performance when the number of objects and spatial relations in the text increase." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 68, + 278, + 154 + ], + "blocks": [ + { + "bbox": [ + 170, + 68, + 278, + 154 + ], + "lines": [ + { + "bbox": [ + 170, + 68, + 278, + 154 + ], + "spans": [ + { + "bbox": [ + 170, + 68, + 278, + 154 + ], + "type": "image", + "image_path": "35af5a19de2abd13e113829de562d61d465f6a792dc6f017d5ffd6384fae1ee7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 201, + 281, + 312 + ], + "blocks": [ + { + "bbox": [ + 60, + 201, + 281, + 312 + ], + "lines": [ + { + "bbox": [ + 60, + 201, + 281, + 312 + ], + "spans": [ + { + "bbox": [ + 60, + 201, + 281, + 312 + ], + "type": "image", + "image_path": "c1a802fc7ed665cc7b166206e6631b9eb9c90dac8fd04de56572c1db3fb18305.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "lines": [ + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "type": "text", + "content": "Figure 5. Example images generated by our method and baselines on novel object combinations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "text", + "content": "analyze the performance at each complexity level. Specifically, we use the GPT-synthetic dataset where multiple objects and relations can appear in the same description. We use the number of objects " + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "text", + "content": " and number of spatial relations " + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "text", + "content": " introduced in Sec. 3.3 as proxies for text complexity and plot the performance of each method as the text becomes more complex. i.e., we plot the performance for each value of " + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "text", + "content": " regardless of " + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "text", + "content": ", and vise versa for " + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 368, + 287, + 523 + ], + "type": "text", + "content": ". As shown in Fig. 4, the performance of all methods decreases as the text becomes more complex, but our method still outperforms or is on par with others on complex descriptions. Note that PAINT-WITH-WORDS uses the same layout predictor as ours, which partially explains its strong performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "Generalizability to novel object combinations is a critical requirement for text-to-image models. However, since MSCOCO dataset is collected from real photos, most of its captions contain object pairs and object-attribute pairs that are common in daily life, which are also more likely to overlap with stable diffusion's pre-training data. Thus we additionally evaluate our method on another synthetic dataset that is created similarly to Sec. 3.3, but contains uncommon object pairs, object-attribute pairs, and object spatial relations. Fig. 5 shows sample images generated by our method and baselines. It can be observed that our method successfully generates novel object pairs (e.g., \"zebra is looking at computer\"). Performance on this dataset and more generated examples are shown in Appendix F. Overall, our method is able to generalize to novel object combinations." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 70, + 526, + 134 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 526, + 134 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 526, + 134 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 526, + 134 + ], + "type": "table", + "html": "
Object RecallSPRel Precision
Ours47.2%66.7%
No spatial control39.6%51.7%
No temporal control43.8%61.4%
No optimization41.0%55.6%
", + "image_path": "c79f6b7e36f29fddee805357b82d48563123a16c9446df65c2dce26c562e5073.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 135, + 529, + 144 + ], + "lines": [ + { + "bbox": [ + 323, + 135, + 529, + 144 + ], + "spans": [ + { + "bbox": [ + 323, + 135, + 529, + 144 + ], + "type": "text", + "content": "Table 3. Ablation performance on GPT-synthetic test set." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 158, + 399, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 158, + 399, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 158, + 399, + 171 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 236 + ], + "type": "text", + "content": "In this section, we will investigate the influence of two important steps in our method, namely, the layout predictor and the spatial-temporal attention. We first study how spatial-temporal attention affects our performance. To do that, we consider three variants of attention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 237, + 546, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 546, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 546, + 345 + ], + "type": "text", + "content": "First, we explore the attention without spatial control. Concretely, the diffusion model still attends to both global description and local descriptions. However, instead of only attending to a local description in the pixel region of the object, the model now attends to all local descriptions in the whole image. The combination weights at each denoising step are optimized with only the global CLIP similarity in Sec. 3.4. Table 3 shows that its performance drops drastically, demonstrating the importance of spatial control." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "text", + "content": "Second, we explore the attention without temporal control, where combination weights remain the same for all denoising steps. Note that the weights can be different for each object and are optimized with " + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{attend}}(\\lambda)" + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "text", + "content": " in Sec. 3.4. The results in Table 3 show a significant degradation under both metrics, indicating the benefits of temporal dependency on attention." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": "Finally, we explore the attention without optimization. Specifically, we fix the combination weight " + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{it} = \\frac{1}{N}" + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": " is the number of objects. This value is also the initialization point of " + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\{\\lambda_{it}\\}" + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": " in our method. As shown in Table 3, the performance further drops compared to the no temporal control variant. The results indicate that optimizing the combination weights for a new text description is critical for generating images with high fidelity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 525, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 597 + ], + "type": "text", + "content": "We further study the effects of our layout predictor, which includes the layout predictor trained with only one of the absolute and relative position objectives, the comparison with ground truth and user-provided pixel region, and a different strategy to construct the pixel region for an object. Results of these experiments are presented in Appendix E." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "In this work, we study the text-to-image synthesis task based on diffusion models. We find existing methods lack explicit control on cross-attention in diffusion models, which leads to the generation of low-fidelity images. We propose an algorithm that imposes control on cross-attention in spatial and temporal aspects. Experiments show our method outperforms baselines in generating high" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7773" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "fidelity images. Further ablation study verifies the effectiveness of our spatial and temporal attention control. One limitation of our method is the reliance on a time-consuming optimization scheme, which takes around 10 minutes for each text-to-image generation. Future work may consider striking a better balance between performance and efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 157, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 157, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 157, + 169 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 175, + 287, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 200 + ], + "type": "text", + "content": "The work of Qiucheng Wu, Yujuan Liu, and Shiyu Chang was partially supported by Cisco Faculty Award." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 211, + 107, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 107, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 107, + 224 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 231, + 286, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 286, + 253 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein gan, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 255, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 286, + 308 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv, 2022. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 310, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 310, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 286, + 342 + ], + "type": "text", + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 344, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 286, + 453 + ], + "type": "text", + "content": "[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 4, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 456, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 286, + 487 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 489, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 489, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 489, + 286, + 533 + ], + "type": "text", + "content": "[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 535, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 286, + 578 + ], + "type": "text", + "content": "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 580, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 286, + 612 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 614, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 286, + 657 + ], + "type": "text", + "content": "[9] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv, 2022. 2, 3, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Neurips, 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "text", + "content": "[13] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 230, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 251 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Neurips, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 296 + ], + "type": "text", + "content": "[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., pages 1-33, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 297, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 329 + ], + "type": "text", + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "text", + "content": "[17] Nisha Huang, Fan Tang, Weiming Dong, Tong-Yee Lee, and Changsheng Xu. Region-aware diffusion for zero-shot text-driven image editing. arXiv preprint arXiv:2302.11797, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "type": "text", + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 410, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 441 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 444, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 464 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2014. 5, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 467, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 488 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 521 + ], + "type": "text", + "content": "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 566 + ], + "type": "text", + "content": "[23] Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. arXiv preprint arXiv:2205.14217, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[24] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 12" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 679 + ], + "type": "text", + "content": "[26] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. arXiv preprint arXiv:2205.00363, 2022. 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7774" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[28] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 150 + ], + "type": "text", + "content": "[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv, 2019. 5, 12, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 194 + ], + "type": "text", + "content": "[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 287, + 219 + ], + "type": "text", + "content": "[31] Gary Marcus, Ernest Davis, and Scott Aaronson. A very preliminary analysis of dall-e 2, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 262 + ], + "type": "text", + "content": "[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 287, + 308 + ], + "type": "text", + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 287, + 353 + ], + "type": "text", + "content": "[34] Dong Huk Park, Grace Luo, Clayton Toste, Samaneh Azadi, Xihui Liu, Maka Karalashvili, Anna Rohrbach, and Trevor Darrell. Shape-guided diffusion with inside-outside attention. arXiv preprint arXiv:2212.00210, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 509 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 287, + 543 + ], + "type": "text", + "content": "[38] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Neurips, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 287, + 578 + ], + "type": "text", + "content": "[39] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 580, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 287, + 621 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 5, 12, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 667 + ], + "type": "text", + "content": "[41] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[43] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "type": "text", + "content": "[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "type": "text", + "content": "[45] Jaskirat Singh, Stephen Gould, and Liang Zheng. High-fidelity guided image synthesis with latent diffusion models. arXiv preprint arXiv:2211.17084, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning. PMLR, 2015. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 277, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 308 + ], + "type": "text", + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 310, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 353 + ], + "type": "text", + "content": "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 355, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 430 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "text", + "content": "[50] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 489, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 531 + ], + "type": "text", + "content": "[51] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "text", + "content": "[52] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: a geometric diffusion model for molecular conformation generation, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 569, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 622 + ], + "type": "text", + "content": "[53] Cheng-Fu Yang, Wan-Cyuan Fan, Fu-En Yang, and Yu-Chiang Frank Wang. Layouttransformer: Scene layout generation with conceptual and spatial diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3732-3741, June 2021. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "text", + "content": "[54] Zuopeng Yang, Daqing Liu, Chaoyue Wang, Jie Yang, and Dacheng Tao. Modeling image composition for complex scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7764-7773, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[55] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7775" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 196 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[56] Yu Zeng, Zhe Lin, Jianming Zhang, Qing Liu, John Collomosse, Jason Kuen, and Vishal M Patel. Scenecomposer: Any-level semantic image synthesis. arXiv preprint arXiv:2211.11742, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "text", + "content": "[57] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[58] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7776" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_content_list.json b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04d536ef166019cdf23d910f541cdfb715ca6ade --- /dev/null +++ b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_content_list.json @@ -0,0 +1,1513 @@ +[ + { + "type": "text", + "text": "Harvard Glaucoma Detection and Progression: A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 81, + 128, + 888, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yan Luo* Min Shi* Yu Tian* Tobias Elze Mengyu Wang \nHarvard Ophthalmology AI Lab, Harvard University", + "bbox": [ + 241, + 202, + 725, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yluol6, mshi6, ytian11, tobias_elze, mengyu_wang}@meei.harvard.edu", + "bbox": [ + 191, + 239, + 777, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 289, + 313, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Glaucoma is the number one cause of irreversible blindness globally. A major challenge for accurate glaucoma detection and progression forecasting is the bottleneck of limited labeled patients with the state-of-the-art (SOTA) 3D retinal imaging data of optical coherence tomography (OCT). To address the data scarcity issue, this paper proposes two solutions. First, we develop a novel generalization-reinforced semi-supervised learning (SSL) model called pseudo supervisor to optimally utilize unlabeled data. Compared with SOTA models, the proposed pseudo supervisor optimizes the policy of predicting pseudo labels with unlabeled samples to improve empirical generalization. Our pseudo supervisor model is evaluated with two clinical tasks consisting of glaucoma detection and progression forecasting. The progression forecasting task is evaluated both unimodally and multimodally. Our pseudo supervisor model demonstrates superior performance than SOTA SSL comparison models. Moreover, our model also achieves the best results on the publicly available LAG fundus dataset. Second, we introduce the Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, a multimodal multitask dataset that includes data from 1,000 patients with OCT imaging data, as well as labels for glaucoma detection and progression. This is the largest glaucoma detection dataset with 3D OCT imaging data and the first glaucoma progression forecasting dataset that is publicly available. Detailed sex and racial analysis are provided, which can be used by interested researchers for fairness learning studies. Our released dataset is benchmarked with several SOTA supervised CNN and transformer deep learning models. The dataset and code are made publicly available via https://ophai.hms.harvard.edu/datasets/harvard-gdp1000.", + "bbox": [ + 76, + 320, + 472, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Glaucoma is the leading cause of irreversible blindness globally caused by retinal nerve fiber layer (RNFL)", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e308eae7e47e4262e33c791ccaca8e2fa20cf6ab816fa91ebfa000f3e528c78.jpg", + "image_caption": [ + "Figure 1: Examples of RNFLT maps with high similarities but from different label groups." + ], + "image_footnote": [], + "bbox": [ + 531, + 287, + 857, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "damage[74, 58, 59, 69]. The global prevalence of glaucoma for the population between 40 and 80 years old is $3.54\\%$ [74]. Timely clinical treatment in the early stage of glaucoma can significantly reduce the risk of further vision loss [18, 17]. However, most commonly, glaucoma patients are not aware of their early VF loss until the VF loss becomes severe enough to impair their daily activities such as reading and driving due to the brain and fellow eye compensation[4, 19, 34, 50, 51, 30].", + "bbox": [ + 496, + 411, + 892, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is desirable to develop automated glaucoma screen tools to reduce societal disease burden [77]. The clinical diagnosis of glaucoma is a holistic decision primarily based on RNFL damage and VF loss assessment supplemented with other patient clinical information and family disease history [29, 47]. The VF test is only available in the ophthalmology specialty. So existing automated screening tools are based on the idea of predicting glaucoma diagnosis from evaluating RNFL damage using retinal images [14, 57, 42, 5, 49, 37, 15, 35, 60, 81]. Before the massive clinical adoption of the 3D retinal imaging tool of optical coherence tomography (OCT) a decade ago [31, 12], 2D fundus photos were used to assess RNFL damage. Therefore, most prior glaucoma detection models are based on the 2D fundus photo [37, 57, 33, 14, 42, 1, 64, 15], which is not the state of the art (SOTA) clinical imaging modality anymore. OCT has replaced fundus photos as the primary imaging tool for glaucoma clinical care due to its superiority in measuring RNFL damage [48, 12]. There have been a number of recent works using OCT scans to automatically screen glaucoma with machine learning [15, 60, 5, 35, 24, 80, 49]. Apart from glaucoma detection, a more clinically significant task is progression forecasting using OCT scans [52, 16, 32, 28], which could inform clini", + "bbox": [ + 496, + 545, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*First three authors contribute equally.", + "bbox": [ + 94, + 886, + 300, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20471", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cians if they should treat the patient conservatively with eye drops or aggressively with invasive surgeries.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For both the glaucoma screening and progression forecasting tasks, a significant challenge is limited labeled patients with the SOTA OCT measurements as the fundus photo is already outdated in clinical practice [15, 60, 5, 35, 24, 80, 49]. This problem with the lack of labeled patients is even more severe for progression forecasting, which requires longitudinal VF measurements to determine progression outcomes [52, 16, 32]. Therefore, there is a strong need for developing novel semi-supervised learning models to maximally utilize unlabeled patient data in clinics to improve glaucoma detection and progression forecasting. However, apart from several papers using standard semi-supervised learning methods for glaucoma detection [20, 2, 23, 79, 84, 9], there has been no dedicated algorithm development for semi-supervised learning with its application in glaucoma. SOTA semi-supervised learning (SSL) approaches in computer vision are generally based on consistency learning [11, 68, 10, 73] or pseudo-labeling [40, 61]. Unlike consistency learning, the pseudo-labeling approaches do not need domain knowledge to design the image augmentation strategy. While many prior works have endeavored to optimally select pseudo-labeled samples to improve the classifier training, the tasks of glaucoma detection and progression forecasting still have some unique challenges that might not be well tackled by existing pseudo-labeling SSL methods [61, 63, 13, 40, 56]. Figure 1 shows examples of RNFLT maps with high similarities but from different label groups. With such obscured group differences, existing pseudo-labeling SSL methods might not be able to provide sufficient guidance to select correct pseudo-labeled samples to improve the classifier training. Furthermore, having high-quality public datasets with OCT scans for glaucoma detection and progression forecasting is equally important to enable more computer vision researchers to study this topic to push forward the field,", + "bbox": [ + 76, + 125, + 472, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we make two contributions: (1) developing a generalization-reinforced semi-supervised learning model called pseudo supervisor to improve glaucoma detection and progression forecasting. Specifically, our pseudo supervisor model will select pseudo-labeled samples via a policy net that optimizes generalization error with a validation subset to update the classifier. (2) releasing a pilot multitask and multimodal glaucoma dataset, called Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, for computer vision researchers to study this topic. In Table1, we have listed major public glaucoma datasets. Most of them are fundus photo datasets, while there are two major OCT datasets previously published [24, 80]. In comparison, our dataset has the largest patient numbers among all datasets with patient numbers available. In addition, our dataset is the first and the only one with progression forecasting task data, which is a way more clinically significant", + "bbox": [ + 76, + 648, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/4589d4ae7629e0e3c430f27592d88e85a11132a3df64e03d3d7aecae8f812026.jpg", + "table_caption": [ + "Table 1: Public Glaucoma Datasets." + ], + "table_footnote": [], + "table_body": "
StudyImaging ModalitySample SizeGlaucoma DetectionProgression ForecastingMultimodalLabel SourceAccessibility
LAG [38]FundusNA (4,854)✓✘✘Clinician AssessmentUpon Request
REFuge [54]FundusNA (1,200)✓✘✘Clinician AssessmentWeb Download
G1020 [6]Fundus432 (1,020)✓✘✘Clinician AssessmentWeb Download
ACRIMA [21]FundusNA (705)✓✘✘Clinician AssessmentWeb Download
RIM-OONE [7]Fundus169 (485)✓✘✘Clinician AssessmentWeb Download
ORIGA [84]FundusNA (650)✓✘✘Clinician AssessmentUpon Request
PAPILA [36]Fundus244 (488)✓✘✘Clinician AssessmentWeb Download
GOALS [24]OCT158 (300)✓✘✘Clinician AssessmentWeb Download
GAMMA [80]OCT300 (300)✓✘✓Clinician AssessmentWeb Download
OursOCT1,000 (1,000)✓✓✓Functional TestWeb Download
", + "bbox": [ + 503, + 104, + 893, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "task than glaucoma detection itself. Compared with the other two OCT datasets, our dataset is superior in patient numbers (1,000 versus up to 300), task versatility (glaucoma detection and progression forecasting versus glaucoma detection only), and availability of objective visual function tests (i.e. VF test). Furthermore, since our dataset is from the US population, the racial representation in our data is more diverse than the two OCT datasets [24, 80] and the large LAG dataset [38], which consists of mainly Asian patients. Therefore, our dataset can be potentially used for medical fairness learning studies, especially since we know Black people have more than doubled glaucoma prevalence than other races, which is a significant disparity [62, 25].", + "bbox": [ + 496, + 218, + 893, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 428, + 640, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Glaucoma Detection. Vast majority of glaucoma detection works using deep learning are based on the fundus photo, which is a 2D snapshot of the patient's retina [37, 57, 33, 14, 42, 1, 64, 15, 76, 75]. However, the 3D imaging modality OCT scan has replaced fundus photos to become the de facto standard for structural damage assessment in glaucoma. Compared with the deep learning glaucoma detection using fundus photos, deep learning glaucoma detection works using OCT scans generally have much smaller sample sizes, which are up to several thousand [15, 60, 5, 35, 24, 80, 49]. As deep learning models are typically data-hungry, there is a clear need to develop new approaches to deal with the limited labeled data issue. It is known that semi-supervised learning can improve prediction accuracy by leveraging unlabeled data in model training. So far, apart from several glaucoma detection papers using standard semi-supervised learning models [20, 2, 23, 79, 84, 9], no novel semi-supervised learning model has been developed with an application focus on glaucoma.", + "bbox": [ + 496, + 454, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Progression Forecasting. Compared with the glaucoma detection task, progression forecasting is a much more clinically significant task for patients. If we could have an accurate progression forecasting model to inform clinicians of which patient will progress or not, then the clinicians can accordingly decide if a patient should be treated more aggressively with invasive surgeries, which may have a strong side effect, or more conservatively with eye drops. So far, due to the data scarcity issue, we have only found several recent papers using retinal imaging data (whether fundus", + "bbox": [ + 496, + 752, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20472", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or OCT) to forecast progression [52, 16, 32]. The sample sizes of these works are only up to several hundred. Imaginably, we have not found any prior studies applying semisupervised learning for progression forecasting.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-Supervised Learning. Semi-Supervised Learning (SSL) aims to optimize a model with both labeled and unlabeled data. The current SSL methods can be divided into pseudo-labeling-based methods [61, 63, 13, 40] and/or consistency-based methods [73, 68, 11, 10, 41, 43]. Consistency-based SSL methods optimize the model with standard classification losses (e.g., cross-entropy) on labeled images and regularizes/minimizes the prediction outputs of different weakly and/or strongly augmented views of unlabeled images, where these views are constructed using different types of image augmentations, such as flip-and-shift, color sharpness, and contrast, etc. Even though consistency-based SSL methods achieve SOTA results in many computer vision benchmarks [73, 68, 11, 10], these methods significantly rely on the orderly design of augmentation functions that requires domain knowledge to design proper image augmentation strategies, which are even more challenging for multimodal modeling. Prior pseudolabeling SSL methods [61, 63, 13, 40, 56] typically utilize a model that trained with the labeled subset to estimate the pseudo labels of the confident unlabeled samples and take these pseudo-labeled data to re-train the model. As mentioned previously, pseudo-labeling-based SSL approaches is a general SSL learning framework for different data modalities naturally. More importantly, as mentioned in Section 1, since the structural differences between glaucoma and non-glaucoma (similarly for progression versus non-progression) are obscured as shown in Figure 1, existing pseudo-labeling SSL methods [61, 63, 13, 40, 56] might not be able to provide sufficient guidance to select correct pseudo labeled samples to benefit the classifier training. In this aspect, we propose a generalization-reinforced SSL model by selecting optimal pseudo-labeled samples for classifier training that explicitly optimize generalization error with a policy net stochastically generates pseudo labels.", + "bbox": [ + 75, + 151, + 472, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinforcement Learning. Reinforcement learning studies an agent interacts with the environment and learns an optimal policy to tackle decision-making problems effectively [70, 8, 26, 65, 66]. Reinforcement learning has been applied to semi-supervised learning setting for real-world applications, e.g. domain adaptation [39]. However, how to utilize reinforcement learning techniques to determine pseudo labels for improving the generalization of classification models remains under exploration.", + "bbox": [ + 75, + 670, + 470, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multimodal and Multitask Dataset", + "text_level": 1, + "bbox": [ + 76, + 820, + 393, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection and Quality Control", + "text_level": 1, + "bbox": [ + 76, + 845, + 393, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work was approved by the institutional review board (IRB) and adhered to the tenets of the Declaration of", + "bbox": [ + 75, + 869, + 470, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f630c8cb3ff61917878cbf5a43e9c6a22a285bb9fd9b90f9f0a07e40c2fb4d88.jpg", + "image_caption": [ + "Figure 2: Pointwise RNFLT distribution at different percentiles." + ], + "image_footnote": [], + "bbox": [ + 506, + 93, + 898, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Helsinki. Because of the retrospective nature of the study, the IRB waived the need for informed consent of patients. The glaucoma patients are from Mass Eye and Ear of Harvard Medical School, which were tested between 2010 and 2021. The RNFLT $(0 - 350$ microns) is from the OCT report provided by the OCT device manufacturer (Cirrus, Carl Zeiss Meditec, Jena, Germany) calculated as the vertical distance between the internal limiting membrane boundary and nerve fiber layer boundary. The RNFLT measurement is the de facto structural measurement for clinicians to diagnose glaucoma. Each two-dimensional $224\\times 224$ RNFLT map is generated within the $6\\times 6\\mathrm{mm}^2$ peripapillary area around the optic nerve head. The VF testing is performed using the Humphrey Field Analyzer with the 24-2 protocol, which measures the patient's eye's visual spatial sensitivity with a radius of 24 degrees from the fixation. Data quality control based on standard clinical practice guideline following the manufacturer's recommendation was applied to exclude unreliable data: (1) OCT scans with signal strength less than 6 (10 represents the best imaging quality) were excluded; (2) VF tests with fixation losses $>33\\%$ , false positive rates $>20\\%$ or false negative rates $>20\\%$ were excluded. Each VF measurement is represented as a vector of 52 total deviation (TD) values between -38 dB and 26 dB. In this work, we refer RNFLT map and VF test as two data modalities which constitute a multimodal dataset. Note that, the modifier word \"Harvard\" from the Harvard GDP Dataset only indicates that our dataset is from the Department of Ophthalmology of Harvard Medical School and does not imply an endorsement, sponsorship, or assumption of responsibility by either Harvard University or Harvard Medical School as a legal identity.", + "bbox": [ + 496, + 227, + 893, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Analysis", + "text_level": 1, + "bbox": [ + 500, + 714, + 643, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The glaucoma detection dataset contains 15,725 training samples, 1,047 validation samples, and 4,287 samples from 5,640, 376, and 1,505 glaucoma patients, respectively. Half of the samples are unlabeled. Among the half samples with labels, $45.2\\%$ of the samples are glaucoma, while the rest are non-glaucoma. The progression forecasting dataset includes 5,889 training samples, 431 validation samples, and 1,547 test samples from 1,292, 86, and 345 glaucoma patients, respectively. Among the half samples with progression outcome labels, $24.0\\%$ and $2.6\\%$ of the samples are TD Progression and MD Fast Progression, respectively. The", + "bbox": [ + 498, + 737, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20473", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "two clinically accepted criteria to define glaucoma progression: (1) TD Progression: eyes with at least three locations with TD slope $\\leq -1$ dB; (2) MD Fast Progression: eyes with MD slope $\\leq -1$ dB [78]. The MD Fast Progression is more detrimental for the patients than TD progression [3].", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 presents the pointwise distributions of 25th, 50th, and 75th percentile RNFLTs in the train (validation set included) and test sets, which demonstrate two obvious superotemporal and inferotemporal RNFL bundles with their damages are known closely related to the visual function loss. As in the two representative examples shown in Figure 1, RNFLT maps with similar appearances may have different labels. This suggests the class boundary may be obscure to determine, which naturally prompts us to adopt semi-supervised solutions to address the glaucoma detection and progression forecasting problems which fully leverage the abundant information in unlabeled samples to mitigate the challenges.", + "bbox": [ + 75, + 167, + 470, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The released dataset including 1,000 patients is a random subset from the entire dataset used to build the glaucoma detection and progression forecasting models in this paper, it naturally inherits the original data characteristics. The released glaucoma detection dataset contains 1,000 samples from 1,000 patients, where each sample includes two data modalities (i.e. RNFLT map and VF test) and respective glaucoma labels. The progression forecasting dataset is a subset from the released glaucoma detection dataset, which entails 500 patients with progression labels and patient demographics including age, sex, and race. For all glaucoma patients, the age distribution of 25th, 50th, and 75th percentiles are 54.3, 64.5, and 72.8 years, respectively. The male and female account for $56\\%$ and $44\\%$ of the 1,000 patients, respectively. Among the 1,000 patients, the White, Black, and Asian groups account for $76.4\\%$ , $14.9\\%$ , and $8.7\\%$ , respectively. To the best of our knowledge, this is the first public glaucoma dataset with the patient demographic information available, which carries great potential to enable interested researchers to conduct and evaluate fairness learning studies.", + "bbox": [ + 75, + 361, + 470, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Problem Formalization & Background", + "text_level": 1, + "bbox": [ + 76, + 690, + 424, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we formally introduce the problems of glaucoma detection and progression forecasting.", + "bbox": [ + 75, + 715, + 468, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denote $\\mathcal{D}^l = \\{(x_i,y_i)|1\\leq i\\leq N^l\\}$ be a labeled dataset, where $x\\in \\mathbb{R}^n$ are the RNFLT maps and $y\\in \\mathcal{V}$ are corresponding glaucoma/progression labels. Also, we have a stack of unlabeled RNFLT maps, i.e. $\\mathcal{D}^{ul} = \\{x_j|1\\leq j\\leq N^{ul}\\}$ . We assume that we have a classification model $f:\\mathbb{R}^n\\xrightarrow{\\theta}\\mathcal{Y}(f_\\theta$ for short), where $\\theta$ are the parameters. To update the classification model with the labeled data, a predefined loss function $\\ell :\\mathcal{V}\\to \\mathcal{V}$ is used to gauge the discrepancy between the predictions and the ground-truth labels such that the classification model can be updated along", + "bbox": [ + 75, + 747, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa4f4a7220825318e879c4d2882163076c01f91d7168ea2785a488315c3e180a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 85, + 911, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b929fc2225ec1040d39a3e24df785e1d0c60ddb79fa1fdd5cd09d277483bfcc.jpg", + "image_caption": [ + "Figure 3: The definitions of (a) glaucoma detection task; and (b) progression forecasting task. In the VF test, blue indicates normal visual function while red indicates abnormal visual function." + ], + "image_footnote": [], + "bbox": [ + 500, + 250, + 908, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the loss, i.e. $\\min_{\\theta} \\ell(f_{\\theta}(x), y)$ .", + "bbox": [ + 498, + 465, + 883, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3 (a), the practical utility scenario of glaucoma detection with deep learning is to use OCT scans that can be easily measured in the setting of primary care or pharmacy to predict glaucoma diagnosis defined by visual field tests that typically only can be measured in the setting of an ophthalmology clinic or eye hospital with specialized technicians. The clinical relevance of progression forecasting as shown in Figure 3 (b) with deep learning is that clinicians can better decide if a patient should be treated more conservatively with eye drops or aggressively with invasive surgeries. As suggested by conventional semi-supervised learning methods [20, 2, 40, 68, 11], pseudo labels $\\tilde{y}$ are predicted by a supervisor model $\\pi : \\mathbf{R}^n \\xrightarrow{\\omega} \\mathcal{Y}$ with the unlabeled data $\\tilde{x}$ , i.e. $\\tilde{y} = \\pi(\\tilde{x};\\omega)$ . Similar to the supervised learning setting, the unlabeled data with the pseudo labels can participate in the supervised learning process, namely $\\min_{\\theta} \\ell(f_{\\theta}(\\tilde{x}), \\tilde{y})$ .", + "bbox": [ + 496, + 482, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, as shown in Figure 4, the natural images associated with different visual concepts are relatively easy to be distinguished. In contrast, the differences in RNFLT maps between glaucoma versus non-glaucoma (similarly for progression versus non-progression) can be very subtle due to large retinal physiological inter-subject variations. There might not be clear and adequate guidance in the learning process to determine the underlying labels of the unlabeled RNFLT maps. Instead, a more sensible way to better determine the pseudo labels is to apply a policy to obtain pseudo labels. Then, we observe how these pseudo labels affect", + "bbox": [ + 496, + 737, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20474", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a764494e5de5f11426900ba55643901b9d7e776b04d85bec0e36ba572e445deb.jpg", + "image_caption": [ + "Figure 4: A challenge in glaucoma detection and progression forecasting. Unlike the classification with natural images, where the images with different labels are likely visually distinguishable, RNFLT images with different labels (e.g. glaucoma vs. non-glaucoma and progression vs nonprogression) might not be easily distinguishable." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 457, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the training process of the classifier and adjust the policy to predict the pseudo labels such that the generalizability of the classifier learned with the pseudo labels can be maximized. In short, the pseudo labels predicted by the supervisor model should be aligned with the improvement of the generalizability of the classifier on a set of unseen data $\\mathcal{D}^{unseen}$ . This objective can be formulated as", + "bbox": [ + 75, + 272, + 470, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\{\\tilde {y} \\}} \\frac {1}{| \\mathcal {D} ^ {\\text {u n s e e n}} |} \\sum_ {(x, y) \\in \\mathcal {D} ^ {\\text {u n s e e n}}} \\ell \\left(f _ {\\theta^ {*}} (x), y\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 385, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta^{*}\\in \\arg \\min_{\\theta}\\frac{1}{|\\mathcal{D}^{l + ul}|}\\sum_{(x,y)\\in \\mathcal{D}^{l + ul}}\\ell (f_{\\theta}(x),y)$", + "bbox": [ + 91, + 429, + 436, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{D}^{l + ul}$ is the union of labeled dataset $\\mathcal{D}$ and unlabeled dataset with pseudo labels $\\mathcal{D}^{ul}$ .", + "bbox": [ + 76, + 474, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Generalization-Reinforced Pseudo Supervisor", + "text_level": 1, + "bbox": [ + 76, + 517, + 468, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from prior-work semi-supervised learning methods [68, 11, 10, 53], the proposed method explicitly leverages the empirical generalizability of the classifier that is quantified with unseen RNFLT maps (e.g. the ones from the validation set) to optimize the policy of predicting the pseudo labels. The empirical generalizability can be viewed as a type of informative reward as the goal of training a classifier is to find one that can generalize well on the unseen test data drawn from the underlying distribution.", + "bbox": [ + 75, + 560, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 shows how the proposed generalization-reinforced pseudo supervisor predicts pseudo labels and is trained according to the observed sets of states (maps), actions (pseudo labels), and rewards (the empirical generalizability with a mini-batch of validation samples). The proposed pseudo supervisor is based on the policy gradient [70], which is one type of generic reinforcement learning methods [8, 26, 65].", + "bbox": [ + 75, + 693, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we formally introduce the concepts related to the proposed pseudo supervisor. An RNFLT map $x_{t}$ at time step $t$ can be viewed as a state $s_t \\in S$ , while the predicted pseudo label $\\tilde{y}_t$ can be viewed as an action $a_{t} \\in \\mathcal{A}$ . A reward function $r_t: S \\times \\mathcal{A} \\to \\mathbf{R}$ is defined on the state-action space. The proposed pseudo supervisor that employs", + "bbox": [ + 75, + 811, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a policy to determine pseudo labels is denoted as $\\pi_{\\omega} : S \\to \\mathcal{P}(\\mathcal{A})$ , where $\\mathcal{P}(\\mathcal{A})$ are the probability measures on $\\mathcal{A}$ and $\\omega$ are the parameters of the proposed pseudo supervisor.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the first place, the classifier takes labeled samples for training. The unlabeled samples will be fed into the pseudo supervisor for predicting pseudo labels. Before updating the weights of the classifier, the empirical generalizability of classifier $f_{\\theta_t}$ can be quantified with validation samples $(x^{(val)}, y^{(val)})$ , that is, $\\ell(f_{\\theta_t}(x^{(val)}), y^{(val)}) (\\ell_t^{(val)})$ . Specifically, $\\ell$ is the widely-used cross-entropy loss for classification. After updating the weights of the classifier with unlabeled samples and corresponding pseudo labels, given the same mini-batch of validation samples, the empirical generalizability of the updated classifier can be requantified again, i.e., $\\ell(f_{\\theta_{t+1}}(x^{(val)}), y^{(val)}) (\\ell_{t+1}^{(val)})$ . Then, the discrepancy between $\\ell_t^{(val)}$ and $\\ell_{t+1}^{(val)}$ is informative to measure the effect of updating the classifier with the unlabeled samples and the pseudo labels. In this sense, it is desired to use in the reward function, i.e.", + "bbox": [ + 496, + 136, + 892, + 380 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} = \\max \\left(\\exp \\left(\\ell_ {t} ^ {(v a l)} - \\ell_ {t + 1} ^ {(v a l)}\\right) - 1, 0\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 388, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. (2), if $\\ell_{t+1}^{(val)}$ is far less than $\\ell_t^{(val)}$ , it implies that the unlabeled samples and pseudo labels lead to a positive effect on the learning process such that the reward increases adaptively.", + "bbox": [ + 496, + 415, + 890, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A trajectory of states, actions, and rewards within a time window $\\beta$ will be saved. The expected return of policy optimization can be written as", + "bbox": [ + 498, + 478, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ \\left(\\pi_ {\\omega}\\right) = \\mathbb {E} _ {\\pi_ {\\omega}} \\left(\\sum_ {k = 0} \\gamma^ {k} r _ {k}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 526, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma \\in [0,1]$ is the discount rate for the numerical stability of reinforcement learning, and $\\beta$ is the time window for updating $\\pi_{\\omega}$ periodically. According to the policy gradient theorem [70], the expectation of the sample gradient is equal to the actual gradient", + "bbox": [ + 496, + 573, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\pi_ {\\omega}} J (\\pi_ {\\omega}) = \\mathbb {E} _ {\\pi_ {\\omega}} \\left[ \\sum_ {k = 0} ^ {\\beta} \\gamma^ {k} r _ {t + k + 1} \\nabla_ {\\pi_ {\\omega}} \\log \\pi_ {\\omega} \\left(a _ {t} \\mid s _ {t}\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 654, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the policy parameterization can be updated according to the gradient ascend rule", + "bbox": [ + 498, + 718, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\omega} \\leftarrow \\pi_ {\\omega} + \\eta_ {\\pi} \\nabla_ {\\pi_ {\\omega}} J (\\pi_ {\\omega}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 757, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta_{\\pi}$ is the learning rate of $\\pi_{\\omega}$ .", + "bbox": [ + 500, + 782, + 735, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 808, + 632, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Set-Up & Implementation Detail", + "text_level": 1, + "bbox": [ + 500, + 832, + 785, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To optimize our models, we employ the AdamW optimizer [46] and train all models for 10 epochs in all experiments conducted in this study. For supervised classification", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20475", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e46acf5d198aa2915093fb0c4e8c51c068cd6b4bf6b7d6b992d10084bcd216df.jpg", + "image_caption": [ + "Figure 5: An overview of the proposed generalization-reinforced pseudo supervisor in the semi-supervised learning framework. The learning framework consists of the process of training the classifier, the process of measuring the empirical generalizability of the classifier to be used as rewards, and the process of updating the proposed pseudo supervisor to predict pseudo labels that can maximize the classifier's empirical generalizability." + ], + "image_footnote": [], + "bbox": [ + 176, + 87, + 790, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models for glaucoma detection, we use a learning rate of $4 \\times 10^{-5}$ and weight decay of 0, and for progression forecasting, we use a learning rate of $2 \\times 10^{-5}$ and weight decay of 0. These hyperparameters are determined to achieve the best baseline performance based on our experiments. The proposed method is trained using a learning rate of $4 \\times 10^{-5}$ and weight decay of 0. The batch sizes used in the supervised model, semi-supervised models, and the proposed method vary due to the different semi-supervised learning contexts [73, 11, 10, 68, 53]. We adopt the batch sizes that result in the best performance for each respective context. To ensure consistency with previous work, we apply data augmentation techniques to the samples in the semi-supervised learning models, including MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53]. Specifically, we utilize random crop and resize, as well as random horizontal flip. However, we do not apply these data augmentation techniques to the supervised baseline or the proposed method.", + "bbox": [ + 75, + 313, + 472, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the glaucoma detection task, we train all models with 7,860 labeled samples and 7,865 unlabeled samples for training. We use 1,047 and 4,287 labeled samples for validation and testing, respectively. For the progression forecasting task, all the models are trained with 2,944 labeled samples and 2,945 unlabeled samples for training. We use 431 and 1,547 labeled samples for validation and testing, respectively. For multi-modality modeling, we simply concatenate the RNFLT images and VF matrices for joint learning using the same architecture as our single-modality model and apply this model to the progression forecasting task. The model is trained and evaluated using the same train-val-test split as our single-modality experiments.", + "bbox": [ + 75, + 595, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the accuracy, F1 score, and AUC on the test sets as our evaluation matrices. The code is written in PyTorch [55] and we use one RTX A6000 GPU for all experiments.", + "bbox": [ + 76, + 789, + 468, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Baseline Methods", + "text_level": 1, + "bbox": [ + 76, + 845, + 250, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our Supervised baseline is a standard ImageNet pretrained EfficientNet classifier [72] learning with cross", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "entropy using only labeled images. For Semi-Supervised baselines, we choose MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53] due to their previous powerful performances. MeanTeacher [73] is one of the initial methods to enforce models to predict consistent pseudo predictions for labeled and unlabeled data, which construct a target-generating teacher model by averaging the model weights. MixMatch [11], ReMixMatch [10], and FixMatch [68] are consistency-based SSL methods that are based on the data augmentations (i.e., weak and strong augmentation) to make consistent classification predictions between different views of unlabeled data. DASO [53] suggests that those methods fail for imbalanced datasets and proposes a pseudo-labeling framework based on different consistency SSL models by adaptively mixing the linear and semantic pseudo-labels to mitigate the overall bias between majority and minority classes. We adopt the FixMatch version of DASO, following the default setting reported in the paper [53]. We implement all baseline methods using Pytorch [55] from DASO's official codes and conducted experiments under the same codebase and experimental protocols for fair comparison. All models are equipped with the same EfficientNetV2-S [72] backbone and their hyper-parameters are specifically tuned to suit our Glaucoma detection and progression forecasting tasks.", + "bbox": [ + 496, + 313, + 893, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Results for Glaucoma Detection", + "text_level": 1, + "bbox": [ + 500, + 698, + 779, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we evaluate our Pseudo-Sup model against multiple SOTA baselines on the glaucoma detection task. Our model achieves the best performance in terms of accuracy, F1 score, and AUC score. We compute the mean performances over 5 different runs. Notably, our model achieves such results without any specific design of image perturbations/augmentations, guaranteeing its simplicity and applicability to different tasks. Our method surpasses the previous SOTA method DASO by $6.8\\%$ of accuracy, $11.1\\%$ of F1 score, and $3.7\\%$ of AUC, respectively. The best-performing consistency method FixMatch still worse by $0.4\\%$ of accuracy, $1.2\\%$ of F1 score, and $1.5\\%$", + "bbox": [ + 496, + 722, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20476", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2b31f0b1d780c0c148e37582ca7c7dded7b3a4001e03553c8a9418ed15671a10.jpg", + "table_caption": [ + "Table 2: Performance on the cross-sectional data for the glaucoma detection task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method." + ], + "table_footnote": [], + "table_body": "
ModelAcc↑F1↑AUC↑
Baseline0.8131 ± 0.00390.8057 ± 0.00330.8693 ± 0.0022
MeanTeacher0.8148 ± 0.00530.8053 ± 0.00720.8541 ± 0.0040
MixMatch0.6392 ± 0.01610.6273 ± 0.00590.8451 ± 0.0062
ReMixMatch0.6741 ± 0.00190.6716 ± 0.00180.8442 ± 0.0032
DASO0.7444 ± 0.01150.6929 ± 0.00500.8371 ± 0.0118
FixMatch0.8077 ± 0.00320.7912 ± 0.00210.8574 ± 0.0019
Pseudo-Sup0.8124 ± 0.00250.8038 ± 0.00460.8727 ± 0.0006
Pseudo-Sup+Aug0.8224 ± 0.01010.8128 ± 0.01220.8818 ± 0.0020
", + "bbox": [ + 84, + 164, + 465, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of AUC than our Pseudo-Sup model, indicating the effectiveness of our approach and the challenges of designing appropriate augmentations for RNFLT maps from previous approaches. We also apply weak augmentations (similar to FixMatch) to our Pseudo-Sup (Pseudo-Sup+Aug) leading to improved performances. Please note that consistency methods (e.g., DASO, FixMatch, etc.) employ both weak and strong augmentations.", + "bbox": [ + 75, + 300, + 468, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Results for Progression Forecasting", + "text_level": 1, + "bbox": [ + 76, + 431, + 383, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unimodal Results. We further evaluate our model on progression forecasting, a more clinically significant task than glaucoma detection, as shown in Table 3. Accurate forecasting of glaucoma progression plays a vital role in prescribing appropriate treatment plans to efficiently prevent further vision losses and is important for assisting the development of new pharmaceutical medicines to treat glaucoma. Adding weak augmentations further improves the results for both progression tasks. Please note that similar to glaucoma detection experiments, we also compute the mean performances over 5 different runs.", + "bbox": [ + 75, + 454, + 468, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For TD Progression, our model achieves the best $74.47\\%$ of mean AUC, respectively. When compared with consistency-based approaches, our method improves between $0.5\\% - 2.1\\%$ in AUC. For MD Fast Progression, our approach consistently achieves the best AUC performance, illustrating the effectiveness and accuracy of our method regardless of any glaucoma progression criteria. Given that most of the glaucoma patients are non-progressive, our progression dataset for TD Progression and MD Fast Progression are both highly imbalanced (i.e., most of the samples are labeled as non-progression). Therefore, it is worth noting that our method also outperforms the previous imbalanced SSL SOTA DASO by a large margin in all measures under such imbalanced scenarios.", + "bbox": [ + 75, + 618, + 468, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multimodal Results. One of the major contributions of this work is to introduce a multi-modal dataset for glaucoma progression forecasting. To the best of our knowledge, this is the first work to explore and benchmark SSL approaches using multimodal model inputs. In Table 5, we show the", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7fedad25a092748987f9af72a6876e6e2ead14243f8a49f4281b21ba30369d7c.jpg", + "table_caption": [ + "Table 3: Performance on the longitudinal data with single modality RNFLT for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while Pseudo-Sup represents the proposed pseudo supervisor method." + ], + "table_footnote": [], + "table_body": "
ModelsAUC ↑ (TD)AUC ↑ (MD Fast)
Baseline0.7367 ± 0.01260.6677 ± 0.0017
MeanTeacher0.7367 ± 0.01260.6677 ± 0.0017
MixMatch0.7238 ± 0.00980.6518 ± 0.0069
ReMixMatch0.7390 ± 0.01240.6595 ± 0.0083
DASO0.7231 ± 0.01880.6787 ± 0.0148
FixMatch0.7395 ±0.01520.6747 ± 0.0054
Pseudo-Sup0.7447 ± 0.00840.7004 ± 0.0218
Pseudo-Sup+Aug0.7477 ± 0.00740.7215 ± 0.0162
", + "bbox": [ + 532, + 164, + 864, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "progression forecasting results using two modalities (RNFLT and VF). Our model achieves the best performance on all evaluation measures. To keep its simplicity and adaptability, we concatenate the RNFLT and the up-scaled VF together as the inputs of our model. Please note that consistency SSL approaches that are based on image augmentations are not included in this table because of the lack of design of augmentation techniques in multimodal medical data. The SOTA results with multimodal glaucoma data enable more accurate and applicable potential computer-aided diagnosis (CAD) systems in real clinical settings.", + "bbox": [ + 496, + 345, + 890, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Supervised Benchmarks on Released Data", + "text_level": 1, + "bbox": [ + 498, + 542, + 859, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 6, we show the supervised classification results for the progression forecasting tasks with multiple SOTA supervised CNN and transformer baseline methods, including VGG [67], ResNet [27], ResNext [82], WideResNet [83], EfficientNet [71], ConvNext [45], ViT [22], and Swin Transformers [44]. This benchmark is conducted on our future data release with 500 patients upon acceptance, of which 400 patients are used for training and the remaining are used for testing. To the best of our knowledge, this is the first supervised glaucoma progression forecasting benchmark, which aims to provide broader impacts for the computer vision and medical imaging communities to build clinically effective computer-aided diagnosis systems. Moreover, the results are reported with both single RNFLT modality and multi-modality (RNFLT and VF) samples. It is worth noting that the accuracy and F1 are the same on MD fast progression because the testing set only contains 100 samples under highly imbalanced class distributions. For more experimental details regarding this supervised progression forecasting benchmark and the supervised results on future release data with 1,000 patients for the glaucoma detection task, please refer to the Supplemental Material.", + "bbox": [ + 496, + 574, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20477", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5fb043585c1bd17779971756c97cc938839836f7ba4d32c7a1077e00d4d0ce82.jpg", + "table_caption": [ + "Table 4: Performance comparisons for glaucoma detection task on LAG dataset." + ], + "table_footnote": [], + "table_body": "
ModelAcc↑F1↑AUC↑
MeanTeacher0.82800.82790.9082
MixMatch0.50200.33780.9075
ReMixMatch0.69800.47850.8369
DASO0.79710.79060.9164
FixMatch0.83660.82970.9198
Pseudo-Sup0.91600.91580.9858
", + "bbox": [ + 91, + 117, + 460, + 223 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ec5ee9bba044a276cfb71cc1a53a4fd696e5a0bba86ebfb2daa71b8ae35e243f.jpg", + "table_caption": [ + "Table 5: Performance on the longitudinal data with two modalities (RNFLT and VF) for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method. Please note that most methods in Table 3 are based on weak/strong augmentations [68, 11, 10] requiring RGB channels, which are not adaptable for multi-modal learning." + ], + "table_footnote": [], + "table_body": "
ModelsAUC (TD)AUC (MD Fast)
Baseline0.7343 ± 0.00980.6814 ± 0.0081
MeanTeacher0.7436 ± 0.01330.6697 ± 0.0035
Pseudo-Sup0.7517 ± 0.00410.7265 ± 0.0095
", + "bbox": [ + 120, + 358, + 431, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.6. Performances on LAG", + "text_level": 1, + "bbox": [ + 76, + 450, + 284, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 4, we show the performance comparisons on LAG dataset [37]. LAG is a large-scale fundus imaging dataset for glaucoma detection. For this experiment, we use 3,854 images for training with half with labels and the remaining without labels, and 1,000 images for testing, with the size of $224 \\times 224$ . For fair comparisons, we use the same hyper-parameters and experimental setups as our other benchmarks. It is worth noting that our approach surpasses the SOTA methods by a large margin and achieves the highest $98.58\\%$ AUC with only half-labeled data, illustrating the effectiveness and applicability of our proposed Pseudo-Sup in different data and tasks.", + "bbox": [ + 75, + 476, + 468, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.7. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 667, + 230, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "According to Section 5, the time window $\\beta$ and discount rate $\\gamma$ are the most important hyperparameters affecting the proposed pseudo supervisor. The ablation studies of the two hyperparameters are shown in Figure 6. We can see that the proposed pseudo supervisor behaviors stably with various $\\beta$ on AUC. Specifically, when $\\beta = 50$ leads to better performance. The consistent pattern can be also observed in the analysis of $\\gamma$ . The discount factor is used to achieve a trade-off between the rewards in the distant future and the rewards in the near future. If it is 0, reinforcement learning agents would only look for an optimal policy for the first action. As shown in Figure 6 (b), we find the optimal discount rate at 0.9, which implies that the states, actions, and rewards in the progression task are coherent to a certain degree.", + "bbox": [ + 75, + 693, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/42a228de84eaa9a3a132f522e648462b6ba5a313bce0966d87f10d016ae10a2f.jpg", + "table_caption": [ + "Table 6: Performance of different supervised methods on the longitudinal data with single/multi-modality for the progression forecasting tasks on the released 500 glaucoma progression data." + ], + "table_footnote": [], + "table_body": "
ModelTD ProgressionMD Fast Progression
Acc†F1†AUC†Acc†F1†AUC†
VGG [67]0.78/0.800.75/0.740.84/0.790.98/0.980.49/0.490.78/0.68
ResNet [27]0.75/0.740.64/0.670.74/0.750.98/0.980.49/0.490.71/0.59
ResNext [82]0.78/0.770.65/0.680.78/0.760.98/0.980.49/0.490.61/0.80
WideResNet [83]0.77/0.790.71/0.750.77/0.800.98/0.980.49/0.490.71/0.64
EfficientNet [71]0.73/0.780.66/0.730.76/0.790.98/0.980.49/0.490.55/0.66
ConvNext [45]0.74/0.810.62/0.770.78/0.810.98/0.980.49/0.490.84/0.62
ViT [22]0.73/0.770.54/0.710.68/0.790.98/0.980.49/0.490.92/0.69
Swin [44]0.71/0.770.49/0.720.68/0.770.98/0.980.49/0.490.87/0.57
", + "bbox": [ + 503, + 148, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7245700de1e5b5558aa9f3bcac2a291de22a468708ba495b32fa9bf3a96546ba.jpg", + "image_caption": [ + "(a) The impact of $\\beta$" + ], + "image_footnote": [], + "bbox": [ + 519, + 262, + 689, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00da3174ba9ddc75138a99fdd156916a6f7ed2abaacb5cb7cba74498e6f74eab.jpg", + "image_caption": [ + "(b) The impact of $\\gamma$", + "Figure 6: Ablation study of hyperparameters' impact on TD Progression forecasting." + ], + "image_footnote": [], + "bbox": [ + 700, + 260, + 870, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.8. Experiment on Racial Groups", + "text_level": 1, + "bbox": [ + 500, + 454, + 766, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The patients in our dataset are from three racial groups (White (0.88 AUC), Black (0.83 AUC), and Asian (0.85 AUC)). In our experiments, we observe that black patients have the lowest AUC yet the highest glaucoma prevalence [62, 25]. Therefore, our dataset is well-suited for future fairness studies.", + "bbox": [ + 498, + 476, + 890, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 579, + 617, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have made two contributions. First, we have developed a generalization-reinforced semi-supervised learning model termed pseudo supervisor for glaucoma detection and progression forecasting. Specifically, the pseudo supervisor predicts pseudo labels following a policy that optimizes the generalization error in training the classifier. Our pseudo supervisor model overall demonstrates superior performance over a number of SOTA methods. Second, we release a multimodal and multi-task dataset for glaucoma detection and progression forecasting with SOTA 3D OCT imaging data. Our dataset is the largest by far for glaucoma detection with 3D OCT data and the first for progression forecasting task. More importantly, our dataset entails demographic information including gender and race, which is generally unavailable in existing public glaucoma datasets. Given the diverse racial representation in our dataset as detailed in Section 3, our dataset is well equipped for potential fairness learning studies, especially considering the significant glaucoma prevalence disparity between Black and other races [62, 25].", + "bbox": [ + 496, + 604, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20478", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 78, + 90, + 228, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by NIH R00 EY028631, Research to Prevent Blindness International Research Collaborators Award, Alcon Young Investigator Grant, NIH R01 EY030575, and NIH P30 EY003790.", + "bbox": [ + 76, + 114, + 468, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 190, + 173, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jin Mo Ahn, Sangsoo Kim, Kwang-Sung Ahn, Sung-Hoon Cho, Kwan Bok Lee, and Ungsoo Samuel Kim. A deep learning model for the detection of both advanced and early glaucoma using fundus photography. PloS one, 13(11):e0207982, 2018. 1, 2", + "[2] Manal Al Ghamdi, Mingqi Li, Mohamed Abdel-Mottaleb, and Mohamed Abou Shousha. Semi-supervised transfer learning for convolutional neural networks for glaucoma detection. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3812-3816. IEEE, 2019. 2, 4", + "[3] Andrew J Anderson, Abadh K Chaurasia, Ajay Sharma, Amisha Gupta, Shikha Gupta, Anurag Khanna, and Viney Gupta. Comparison of rates of fast and catastrophic visual field loss in three glaucoma subtypes. Investigative ophthalmology & visual science, 60(1):161-167, 2019. 4", + "[4] Paul H Artes and Balwantray C Chauhan. Longitudinal changes in the visual field and optic disc in glaucoma. Progress in retinal and eye research, 24(3):333-354, 2005. 1", + "[5] Ryo Asaoka, Hiroshi Murata, Kazunori Hirasawa, Yuri Fujino, Masato Matsuura, Atsuya Miki, Takashi Kanamoto, Yoko Ikeda, Kazuhiko Mori, Aiko Iwase, et al. Using deep learning and transfer learning to accurately diagnose early-onset glaucoma from macular optical coherence tomography images. American journal of ophthalmology, 198:136-145, 2019. 1, 2", + "[6] Muhammad Naseer Bajwa, Gur Amrit Pal Singh, Wolfgang Neumeier, Muhammad Imran Malik, Andreas Dengel, and Sheraz Ahmed. G1020: A benchmark retinal fundus image dataset for computer-aided glaucoma detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020. 2", + "[7] Francisco José Fumero Batista, Tinguaro Diaz-Aleman, Jose Sigut, Silvia Alayon, Rafael Arny, and Denisse Angel-Pereira. Rim-one dl: A unified retinal image database for assessing glaucoma using deep learning. Image Analysis & Stereology, 39(3):161-167, 2020. 2", + "[8] Jonathan Baxter and Peter L Bartlett. Infinite-horizon policy-gradient estimation. Journal of Artificial Intelligence Research, 15:319-350, 2001. 3, 5", + "[9] Mohammed El Amine Bechar, Nesma Settouti, Vincent Barra, and Mohamed Amine Chikh. Semi-supervised superpixel classification for medical images segmentation: application to detection of glaucoma disease. Multidimensional Systems and Signal Processing, 29(3):979-998, 2018. 2", + "[10] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching" + ], + "bbox": [ + 78, + 215, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and augmentation anchoring. In International Conference on Learning Representations, 2019. 2, 3, 5, 6, 8", + "[11] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32:5049-5059, 2019. 2, 3, 4, 5, 6, 8", + "[12] Igor I Bussel, Gadi Wollstein, and Joel S Schuman. Oct for glaucoma diagnosis, screening and detection of glaucoma progression. British Journal of Ophthalmology, 98(Suppl 2):ii15-ii19, 2014. 1", + "[13] Paola Cascante-Bonilla et al. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. arXiv preprint arXiv:2001.06001, 2020. 2, 3", + "[14] Xiangyu Chen, Yanwu Xu, Shuicheng Yan, Damon Wing Kee Wong, Tien Yin Wong, and Jiang Liu. Automatic feature learning for glaucoma detection based on deep learning. In International conference on medical image computing and computer-assisted intervention, pages 669-677. Springer, 2015. 1, 2", + "[15] Mark Christopher, Akram Belghith, Christopher Bowd, James A Proudfoot, Michael H Goldbaum, Robert N Weinreb, Christopher A Girkin, Jeffrey M Liebmann, and Linda M Zangwill. Performance of deep learning architectures and transfer learning for detecting glaucomatous optic neuropathy in fundus photographs. Scientific reports, 8(1):1-13, 2018. 1, 2", + "[16] Mark Christopher, Akram Belghith, Robert N Weinreb, Christopher Bowd, Michael H Goldbaum, Luke J Saunders, Felipe A Medeiros, and Linda M Zangwill. Retinal nerve fiber layer features identified by unsupervised machine learning on optical coherence tomography scans predict glaucoma progression. Investigative ophthalmology & visual science, 59(7):2748-2756, 2018. 1, 2, 3", + "[17] Anne Louise Coleman. Advances in glaucoma treatment and management: surgery. Investigative ophthalmology & visual science, 53(5):2491-2494, 2012. 1", + "[18] Ronan Conlon, Hady Saheb, and Iqbal Ike K Ahmed. Glaucoma treatment trends: a review. Canadian Journal of Ophthalmology, 52(1):114-124, 2017. 1", + "[19] David P Crabb, Nicholas D Smith, Fiona C Glen, Robyn Burton, and David F Garway-Heath. How does glaucoma look?: patient perception of visual field loss. Ophthalmology, 120(6):1120-1126, 2013. 1", + "[20] Andres Diaz-Pinto, Adrián Colomer, Valery Naranjo, Sandra Morales, Yanwu Xu, and Alejandro F Frangi. Retinal image synthesis and semi-supervised learning for glaucoma assessment. IEEE transactions on medical imaging, 38(9):2211-2218, 2019. 2, 4", + "[21] Andres Diaz-Pinto, Sandra Morales, Valery Naranjo, Thomas Kohler, Jose M Mossi, and Amparo Navea. Cnns for automatic glaucoma assessment using fundus images: an extensive validation. Biomedical engineering online, 18(1):1-19, 2019. 2", + "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20479", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.7,8", + "[23] Rui Fan, Christopher Bowd, Nicole Brye, Mark Christopher, Robert N Weinreb, David Kriegman, and Linda Zangwill. One-vote veto: Semi-supervised learning for low-shot glaucoma diagnosis. arXiv preprint arXiv:2012.04841, 2020. 2", + "[24] Huihui Fang, Fei Li, Huazhu Fu, Junde Wu, Xiulan Zhang, and Yanwu Xu. Dataset and evaluation algorithm design for goals challenge. In International Workshop on Ophthalmic Medical Image Analysis, pages 135-142. Springer, 2022. 1, 2", + "[25] David S Friedman, Henry D Jampel, Beatrix Munoz, and Sheila K West. The prevalence of open-angle glaucoma among blacks and whites 73 years and older: the salisbury eye evaluation glaucoma study. Archives of ophthalmology, 124(11):1625-1630, 2006. 2, 8", + "[26] Mohammad Ghavamzadeh and Sridhar Mahadevan. Hierarchical policy gradient algorithms. Computer Science Department Faculty Publication Series, page 173, 2003. 3, 5", + "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7, 8", + "[28] Patrick Herbert, Kaihua Hou, Chris Bradley, Michael V Boland, Pradeep Y Ramulu, Mathias Unberath, and Jithin Yohannan. Forecasting risk of future rapid glaucoma worsening using early visual field, optical coherence tomography and clinical data. Investigative Ophthalmology & Visual Science, 63(7):2294-2294, 2022. 1", + "[29] Donald C Hood and Randy H Kardon. A framework for comparing structural and functional measures of glaucomatous damage. Progress in retinal and eye research, 26(6):688-710, 2007. 1", + "[30] Sen Hu, Nicholas D Smith, Luke J Saunders, and David P Crabb. Patterns of binocular visual field loss derived from large-scale patient data from glaucoma clinics. Ophthalmology, 122(12):2399-2406, 2015. 1", + "[31] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, et al. Optical coherence tomography. science, 254(5035):1178-1181, 1991. 1", + "[32] Alireza Kamalipour, Susan Moghimi, Pooya Khosravi, Vahid Mohammadzadeh, Takashi Nishida, Eleonora Micheletti, Jo-Hsuan Wu, Golnoush Mahmoudinezhad, Elizabeth HF Li, Mark Christopher, et al. Combining oct and oct-angiography longitudinal data for the detection of visual field progression in glaucoma: Longitudinal oct/octa data to detect glaucoma progression. American Journal of Ophthalmology, 2022. 1, 2, 3", + "[33] Stuart Keel, Jinrong Wu, Pei Ying Lee, Jane Scheetz, and Mingguang He. Visualizing deep learning models for the detection of referable diabetic retinopathy and glaucoma. JAMA ophthalmology, 137(3):288-292, 2019. 1, 2", + "[34] Philip J Kellman and Thomas F Shipley. A theory of visual interpolation in object perception. Cognitive psychology, 23(2):141-221, 1991. 1", + "[35] Yuka Kihara, Giovanni Montesano, Andrew Chen, Nishani Amerasinghe, Chrysostomos Dimitriou, Aby Jacob, Almira" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chabi, David P Crabb, and Aaron Y Lee. Policy-driven, multimodal deep learning for predicting visual fields from the optic disc and oct imaging. Ophthalmology, 2022. 1, 2", + "[36] Oleksandr Kovalyk, Juan Morales-Sanchez, Rafael Verdú-Monedero, Inmaculada Sellés-Navarro, Ana Palazon-Cabanes, and José-Luis Sancho-Gómez. Papila: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment. Scientific Data, 9(1):1-12, 2022. 2", + "[37] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: a large-scale database and cnn model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2019. 1, 2, 8", + "[38] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: A large-scale database and cnn model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2", + "[39] Bingyu Liu, Yuhong Guo, Jieping Ye, and Weihong Deng. Selective pseudo-labeling with reinforcement learning for semi-supervised domain adaptation. arXiv preprint arXiv:2012.03438, 2020. 3", + "[40] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2, 3, 4", + "[41] Fengbei Liu, Yu Tian, Filipe R Cordeiro, Vasileios Bela-giannis, Ian Reid, and Gustavo Carneiro. Self-supervised mean teacher for semi-supervised chest x-ray classification. In International Workshop on Machine Learning in Medical Imaging, pages 426-436. Springer, 2021. 3", + "[42] Hanruo Liu, Liu Li, I Michael Wormstone, Chunyan Qiao, Chun Zhang, Ping Liu, Shuning Li, Huaizhou Wang, Dapeng Mou, Ruiqi Pang, et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs. JAMA ophthalmology, 137(12):1353-1360, 2019. 1, 2", + "[43] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 3", + "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 7, 8", + "[45] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7, 8", + "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "footer", + "text": "20480", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Rizwan Malik, William H Swanson, and David F Garway-Heath. 'structure-function relationship' in glaucoma: past thinking and current concepts. Clinical & experimental ophthalmology, 40(4):369-380, 2012. 1", + "[48] Felipe A Medeiros, Alessandro A Jammal, and Atalie C Thompson. From machine to machine: an oct-trained deep learning algorithm for objective quantification of glaucomatous damage in fundus photographs. Ophthalmology, 126(4):513-521, 2019. 1", + "[49] Parmita Mehta, Christine A Petersen, Joanne C Wen, Michael R Banitt, Philip P Chen, Karine D Bojikian, Catherine Egan, Su-In Lee, Magdalena Balazinska, Aaron Y Lee, et al. Automated detection of glaucoma with interpretable machine learning using clinical data and multimodal retinal images. American Journal of Ophthalmology, 231:154-169, 2021. 1, 2", + "[50] MJ Morgan and RJ Watt. Mechanisms of interpolation in human spatial vision. Nature, 299(5883):553-555, 1982. 1", + "[51] Jacqueline M Nelson-Quigg, Kimberly Cello, and Chris A Johnson. Predicting binocular visual field sensitivity from monocular visual field results. Investigative ophthalmology & visual science, 41(8):2212-2221, 2000. 1", + "[52] Kouros Nouri-Mahdavi, Vahid Mohammadzadeh, Alessandro Rabiolo, Kiumars Edalati, Joseph Caprioli, and Siamak Yousefi. Prediction of visual field progression from oct structural measures in moderate to advanced glaucoma. American Journal of Ophthalmology, 226:172-181, 2021. 1, 2, 3", + "[53] Youngtaek Oh, Dong-Jin Kim, and In So Kweon. Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. arXiv preprint arXiv:2106.05682, 2021. 5, 6", + "[54] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical image analysis, 59:101570, 2020. 2", + "[55] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6", + "[56] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2, 3", + "[57] Sonia Phene, R Carter Dunn, Naama Hammel, Yun Liu, Jonathan Krause, Naho Kitade, Mike Schaekermann, Rory Sayres, Derek J Wu, Ashish Bora, et al. Deep learning and glaucoma specialists: the relative importance of optic disc features to predict glaucoma referral in fundus photographs. Ophthalmology, 126(12):1627-1639, 2019. 1, 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Harry A Quigley. Number of people with glaucoma worldwide. British journal of ophthalmology, 80(5):389-393, 1996. 1", + "[59] Harry A Quigley and Aimee T Broman. The number of people with glaucoma worldwide in 2010 and 2020. British journal of ophthalmology, 90(3):262-267, 2006. 1", + "[60] An Ran Ran, Carol Y Cheung, Xi Wang, Hao Chen, Lu-yang Luo, Poemen P Chan, Mandy OM Wong, Robert T Chang, Suria S Mannil, Alvin L Young, et al. Detection of glaucomatous optic neuropathy with spectral-domain optical coherence tomography: a retrospective training and validation deep-learning analysis. The Lancet Digital Health, 1(4):e172-e182, 2019. 1, 2", + "[61] Mamshad Nayeem Rizve et al. In defense of pseudolabeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In International Conference on Learning Representations, 2020. 2, 3", + "[62] Alicja R Rudnicka, Shahrul Mt-Isa, Christopher G Owen, Derek G Cook, and Deborah Ashby. Variations in primary open-angle glaucoma prevalence by age, gender, and race: a bayesian meta-analysis. Investigative ophthalmology & visual science, 47(10):4254-4261, 2006. 2, 8", + "[63] Weiwei Shi, Yihong Gong, Chris Ding, Zhiheng MaXiaoyu Tao, and Nanning Zheng. Transductive semi-supervised deep learning using min-max features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 299-315, 2018. 2, 3", + "[64] Naoto Shibata, Masaki Tanito, Keita Mitsuhashi, Yuri Fujino, Masato Matsuura, Hiroshi Murata, and Ryo Asaoka. Development of a deep residual learning algorithm to screen for glaucoma from fundus photography. Scientific reports, 8(1):1-9, 2018. 1, 2", + "[65] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In International conference on machine learning, pages 387-395. PMLR, 2014. 3, 5", + "[66] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy P. Lillicrap, Fan Hui, Laurent Sifre, George van den Driessche, Thore Graepel, and Demis Hassabis. Mastering the game of go without human knowledge. Nat., 550(7676):354-359, 2017. 3", + "[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7, 8", + "[68] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33, 2020. 2, 3, 4, 5, 6, 8", + "[69] Yi Sun, Aiming Chen, Minjie Zou, Yichi Zhang, Ling Jin, Yi Li, Danying Zheng, Guangming Jin, and Nathan Congdon. Time trends, associations and prevalence of blindness and vision loss due to glaucoma: an analysis of observational data from the global burden of disease study 2017. BMJ open, 12(1):e053805, 2022. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20481", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Richard S. Sutton, David A. McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems 12, [NIPS Conference, Denver, Colorado, USA, November 29 - December 4, 1999], pages 1057-1063, 1999. 3, 5", + "[71] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7, 8", + "[72] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10096-10106. PMLR, 2021. 6", + "[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 3, 6", + "[74] Yih-Chung Tham, Xiang Li, Tien Y Wong, Harry A Quigley, Tin Aung, and Ching-Yu Cheng. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology, 121(11):2081-2090, 2014. 1", + "[75] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W VERJANS, and Rajvinder Singh. Self-supervised multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. 2021. 2", + "[76] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 128-140. Springer, 2021. 2", + "[77] Rohit Varma, Paul P Lee, Ivan Goldberg, and Sameer Kotak. An assessment of the health and economic burdens of glaucoma. American journal of ophthalmology, 152(4):515-522, 2011. 1", + "[78] Eija Vesti, Chris A Johnson, and Balwantray C Chauhan. Comparison of different methods for detecting glaucomatous visual field progression. Investigative ophthalmology & visual science, 44(9):3873-3879, 2003. 4", + "[79] Xi Wang, Hao Chen, An-Ran Ran, Luyang Luo, Poemen P Chan, Clement C Tham, Robert T Chang, Suria S Mannil, Carol Y Cheung, and Pheng-Ann Heng. Towards multicenter glaucoma oct image screening with semi-supervised joint structure and function multi-task learning. Medical Image Analysis, 63:101695, 2020. 2", + "[80] Junde Wu, Huihui Fang, Fei Li, Huazhu Fu, Fengbin Lin, Jiongcheng Li, Lexing Huang, Qinji Yu, Sifan Song, Xingxing Xu, et al. Gamma challenge: glaucoma grading from multi-modality images. arXiv preprint arXiv:2202.06511, 2022. 1, 2", + "[81] Jo-Hsuan Wu, Takashi Nishida, Robert N Weinreb, and Jou-Wei Lin. Performances of machine learning in detecting" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "glaucoma using fundus and retinal optical coherence tomography images: A meta-analysis. American Journal of Ophthalmology, 2021. 1", + "[82] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 7, 8", + "[83] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 7, 8", + "[84] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual international conference of the IEEE engineering in medicine and biology, pages 3065-3068. IEEE, 2010. 2" + ], + "bbox": [ + 503, + 92, + 890, + 311 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20482", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_model.json b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..10029db0cbb5802979146e80b60f28f4c217dae9 --- /dev/null +++ b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_model.json @@ -0,0 +1,2468 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.13, + 0.89, + 0.175 + ], + "angle": 0, + "content": "Harvard Glaucoma Detection and Progression: A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.203, + 0.726, + 0.238 + ], + "angle": 0, + "content": "Yan Luo* Min Shi* Yu Tian* Tobias Elze Mengyu Wang \nHarvard Ophthalmology AI Lab, Harvard University" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.24, + 0.778, + 0.255 + ], + "angle": 0, + "content": "{yluol6, mshi6, ytian11, tobias_elze, mengyu_wang}@meei.harvard.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.29, + 0.314, + 0.306 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.473, + 0.811 + ], + "angle": 0, + "content": "Glaucoma is the number one cause of irreversible blindness globally. A major challenge for accurate glaucoma detection and progression forecasting is the bottleneck of limited labeled patients with the state-of-the-art (SOTA) 3D retinal imaging data of optical coherence tomography (OCT). To address the data scarcity issue, this paper proposes two solutions. First, we develop a novel generalization-reinforced semi-supervised learning (SSL) model called pseudo supervisor to optimally utilize unlabeled data. Compared with SOTA models, the proposed pseudo supervisor optimizes the policy of predicting pseudo labels with unlabeled samples to improve empirical generalization. Our pseudo supervisor model is evaluated with two clinical tasks consisting of glaucoma detection and progression forecasting. The progression forecasting task is evaluated both unimodally and multimodally. Our pseudo supervisor model demonstrates superior performance than SOTA SSL comparison models. Moreover, our model also achieves the best results on the publicly available LAG fundus dataset. Second, we introduce the Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, a multimodal multitask dataset that includes data from 1,000 patients with OCT imaging data, as well as labels for glaucoma detection and progression. This is the largest glaucoma detection dataset with 3D OCT imaging data and the first glaucoma progression forecasting dataset that is publicly available. Detailed sex and racial analysis are provided, which can be used by interested researchers for fairness learning studies. Our released dataset is benchmarked with several SOTA supervised CNN and transformer deep learning models. The dataset and code are made publicly available via https://ophai.hms.harvard.edu/datasets/harvard-gdp1000." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.84 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Glaucoma is the leading cause of irreversible blindness globally caused by retinal nerve fiber layer (RNFL)" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.288, + 0.859, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 1: Examples of RNFLT maps with high similarities but from different label groups." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.893, + 0.546 + ], + "angle": 0, + "content": "damage[74, 58, 59, 69]. The global prevalence of glaucoma for the population between 40 and 80 years old is \\(3.54\\%\\) [74]. Timely clinical treatment in the early stage of glaucoma can significantly reduce the risk of further vision loss [18, 17]. However, most commonly, glaucoma patients are not aware of their early VF loss until the VF loss becomes severe enough to impair their daily activities such as reading and driving due to the brain and fellow eye compensation[4, 19, 34, 50, 51, 30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.895, + 0.903 + ], + "angle": 0, + "content": "It is desirable to develop automated glaucoma screen tools to reduce societal disease burden [77]. The clinical diagnosis of glaucoma is a holistic decision primarily based on RNFL damage and VF loss assessment supplemented with other patient clinical information and family disease history [29, 47]. The VF test is only available in the ophthalmology specialty. So existing automated screening tools are based on the idea of predicting glaucoma diagnosis from evaluating RNFL damage using retinal images [14, 57, 42, 5, 49, 37, 15, 35, 60, 81]. Before the massive clinical adoption of the 3D retinal imaging tool of optical coherence tomography (OCT) a decade ago [31, 12], 2D fundus photos were used to assess RNFL damage. Therefore, most prior glaucoma detection models are based on the 2D fundus photo [37, 57, 33, 14, 42, 1, 64, 15], which is not the state of the art (SOTA) clinical imaging modality anymore. OCT has replaced fundus photos as the primary imaging tool for glaucoma clinical care due to its superiority in measuring RNFL damage [48, 12]. There have been a number of recent works using OCT scans to automatically screen glaucoma with machine learning [15, 60, 5, 35, 24, 80, 49]. Apart from glaucoma detection, a more clinically significant task is progression forecasting using OCT scans [52, 16, 32, 28], which could inform clini" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.302, + 0.901 + ], + "angle": 0, + "content": "*First three authors contribute equally." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "cians if they should treat the patient conservatively with eye drops or aggressively with invasive surgeries." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.473, + 0.645 + ], + "angle": 0, + "content": "For both the glaucoma screening and progression forecasting tasks, a significant challenge is limited labeled patients with the SOTA OCT measurements as the fundus photo is already outdated in clinical practice [15, 60, 5, 35, 24, 80, 49]. This problem with the lack of labeled patients is even more severe for progression forecasting, which requires longitudinal VF measurements to determine progression outcomes [52, 16, 32]. Therefore, there is a strong need for developing novel semi-supervised learning models to maximally utilize unlabeled patient data in clinics to improve glaucoma detection and progression forecasting. However, apart from several papers using standard semi-supervised learning methods for glaucoma detection [20, 2, 23, 79, 84, 9], there has been no dedicated algorithm development for semi-supervised learning with its application in glaucoma. SOTA semi-supervised learning (SSL) approaches in computer vision are generally based on consistency learning [11, 68, 10, 73] or pseudo-labeling [40, 61]. Unlike consistency learning, the pseudo-labeling approaches do not need domain knowledge to design the image augmentation strategy. While many prior works have endeavored to optimally select pseudo-labeled samples to improve the classifier training, the tasks of glaucoma detection and progression forecasting still have some unique challenges that might not be well tackled by existing pseudo-labeling SSL methods [61, 63, 13, 40, 56]. Figure 1 shows examples of RNFLT maps with high similarities but from different label groups. With such obscured group differences, existing pseudo-labeling SSL methods might not be able to provide sufficient guidance to select correct pseudo-labeled samples to improve the classifier training. Furthermore, having high-quality public datasets with OCT scans for glaucoma detection and progression forecasting is equally important to enable more computer vision researchers to study this topic to push forward the field," + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this study, we make two contributions: (1) developing a generalization-reinforced semi-supervised learning model called pseudo supervisor to improve glaucoma detection and progression forecasting. Specifically, our pseudo supervisor model will select pseudo-labeled samples via a policy net that optimizes generalization error with a validation subset to update the classifier. (2) releasing a pilot multitask and multimodal glaucoma dataset, called Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, for computer vision researchers to study this topic. In Table1, we have listed major public glaucoma datasets. Most of them are fundus photo datasets, while there are two major OCT datasets previously published [24, 80]. In comparison, our dataset has the largest patient numbers among all datasets with patient numbers available. In addition, our dataset is the first and the only one with progression forecasting task data, which is a way more clinically significant" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.09, + 0.817, + 0.103 + ], + "angle": 0, + "content": "Table 1: Public Glaucoma Datasets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.105, + 0.894, + 0.19 + ], + "angle": 0, + "content": "
StudyImaging ModalitySample SizeGlaucoma DetectionProgression ForecastingMultimodalLabel SourceAccessibility
LAG [38]FundusNA (4,854)✓✘✘Clinician AssessmentUpon Request
REFuge [54]FundusNA (1,200)✓✘✘Clinician AssessmentWeb Download
G1020 [6]Fundus432 (1,020)✓✘✘Clinician AssessmentWeb Download
ACRIMA [21]FundusNA (705)✓✘✘Clinician AssessmentWeb Download
RIM-OONE [7]Fundus169 (485)✓✘✘Clinician AssessmentWeb Download
ORIGA [84]FundusNA (650)✓✘✘Clinician AssessmentUpon Request
PAPILA [36]Fundus244 (488)✓✘✘Clinician AssessmentWeb Download
GOALS [24]OCT158 (300)✓✘✘Clinician AssessmentWeb Download
GAMMA [80]OCT300 (300)✓✘✓Clinician AssessmentWeb Download
OursOCT1,000 (1,000)✓✓✓Functional TestWeb Download
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.413 + ], + "angle": 0, + "content": "task than glaucoma detection itself. Compared with the other two OCT datasets, our dataset is superior in patient numbers (1,000 versus up to 300), task versatility (glaucoma detection and progression forecasting versus glaucoma detection only), and availability of objective visual function tests (i.e. VF test). Furthermore, since our dataset is from the US population, the racial representation in our data is more diverse than the two OCT datasets [24, 80] and the large LAG dataset [38], which consists of mainly Asian patients. Therefore, our dataset can be potentially used for medical fairness learning studies, especially since we know Black people have more than doubled glaucoma prevalence than other races, which is a significant disparity [62, 25]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.642, + 0.445 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Glaucoma Detection. Vast majority of glaucoma detection works using deep learning are based on the fundus photo, which is a 2D snapshot of the patient's retina [37, 57, 33, 14, 42, 1, 64, 15, 76, 75]. However, the 3D imaging modality OCT scan has replaced fundus photos to become the de facto standard for structural damage assessment in glaucoma. Compared with the deep learning glaucoma detection using fundus photos, deep learning glaucoma detection works using OCT scans generally have much smaller sample sizes, which are up to several thousand [15, 60, 5, 35, 24, 80, 49]. As deep learning models are typically data-hungry, there is a clear need to develop new approaches to deal with the limited labeled data issue. It is known that semi-supervised learning can improve prediction accuracy by leveraging unlabeled data in model training. So far, apart from several glaucoma detection papers using standard semi-supervised learning models [20, 2, 23, 79, 84, 9], no novel semi-supervised learning model has been developed with an application focus on glaucoma." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Progression Forecasting. Compared with the glaucoma detection task, progression forecasting is a much more clinically significant task for patients. If we could have an accurate progression forecasting model to inform clinicians of which patient will progress or not, then the clinicians can accordingly decide if a patient should be treated more aggressively with invasive surgeries, which may have a strong side effect, or more conservatively with eye drops. So far, due to the data scarcity issue, we have only found several recent papers using retinal imaging data (whether fundus" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "or OCT) to forecast progression [52, 16, 32]. The sample sizes of these works are only up to several hundred. Imaginably, we have not found any prior studies applying semisupervised learning for progression forecasting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.473, + 0.67 + ], + "angle": 0, + "content": "Semi-Supervised Learning. Semi-Supervised Learning (SSL) aims to optimize a model with both labeled and unlabeled data. The current SSL methods can be divided into pseudo-labeling-based methods [61, 63, 13, 40] and/or consistency-based methods [73, 68, 11, 10, 41, 43]. Consistency-based SSL methods optimize the model with standard classification losses (e.g., cross-entropy) on labeled images and regularizes/minimizes the prediction outputs of different weakly and/or strongly augmented views of unlabeled images, where these views are constructed using different types of image augmentations, such as flip-and-shift, color sharpness, and contrast, etc. Even though consistency-based SSL methods achieve SOTA results in many computer vision benchmarks [73, 68, 11, 10], these methods significantly rely on the orderly design of augmentation functions that requires domain knowledge to design proper image augmentation strategies, which are even more challenging for multimodal modeling. Prior pseudolabeling SSL methods [61, 63, 13, 40, 56] typically utilize a model that trained with the labeled subset to estimate the pseudo labels of the confident unlabeled samples and take these pseudo-labeled data to re-train the model. As mentioned previously, pseudo-labeling-based SSL approaches is a general SSL learning framework for different data modalities naturally. More importantly, as mentioned in Section 1, since the structural differences between glaucoma and non-glaucoma (similarly for progression versus non-progression) are obscured as shown in Figure 1, existing pseudo-labeling SSL methods [61, 63, 13, 40, 56] might not be able to provide sufficient guidance to select correct pseudo labeled samples to benefit the classifier training. In this aspect, we propose a generalization-reinforced SSL model by selecting optimal pseudo-labeled samples for classifier training that explicitly optimize generalization error with a policy net stochastically generates pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.472, + 0.806 + ], + "angle": 0, + "content": "Reinforcement Learning. Reinforcement learning studies an agent interacts with the environment and learns an optimal policy to tackle decision-making problems effectively [70, 8, 26, 65, 66]. Reinforcement learning has been applied to semi-supervised learning setting for real-world applications, e.g. domain adaptation [39]. However, how to utilize reinforcement learning techniques to determine pseudo labels for improving the generalization of classification models remains under exploration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.821, + 0.394, + 0.837 + ], + "angle": 0, + "content": "3. Multimodal and Multitask Dataset" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.394, + 0.864 + ], + "angle": 0, + "content": "3.1. Data Collection and Quality Control" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.9 + ], + "angle": 0, + "content": "This work was approved by the institutional review board (IRB) and adhered to the tenets of the Declaration of" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.094, + 0.9, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.225 + ], + "angle": 0, + "content": "Figure 2: Pointwise RNFLT distribution at different percentiles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.705 + ], + "angle": 0, + "content": "Helsinki. Because of the retrospective nature of the study, the IRB waived the need for informed consent of patients. The glaucoma patients are from Mass Eye and Ear of Harvard Medical School, which were tested between 2010 and 2021. The RNFLT \\((0 - 350\\) microns) is from the OCT report provided by the OCT device manufacturer (Cirrus, Carl Zeiss Meditec, Jena, Germany) calculated as the vertical distance between the internal limiting membrane boundary and nerve fiber layer boundary. The RNFLT measurement is the de facto structural measurement for clinicians to diagnose glaucoma. Each two-dimensional \\(224\\times 224\\) RNFLT map is generated within the \\(6\\times 6\\mathrm{mm}^2\\) peripapillary area around the optic nerve head. The VF testing is performed using the Humphrey Field Analyzer with the 24-2 protocol, which measures the patient's eye's visual spatial sensitivity with a radius of 24 degrees from the fixation. Data quality control based on standard clinical practice guideline following the manufacturer's recommendation was applied to exclude unreliable data: (1) OCT scans with signal strength less than 6 (10 represents the best imaging quality) were excluded; (2) VF tests with fixation losses \\(>33\\%\\), false positive rates \\(>20\\%\\) or false negative rates \\(>20\\%\\) were excluded. Each VF measurement is represented as a vector of 52 total deviation (TD) values between -38 dB and 26 dB. In this work, we refer RNFLT map and VF test as two data modalities which constitute a multimodal dataset. Note that, the modifier word \"Harvard\" from the Harvard GDP Dataset only indicates that our dataset is from the Department of Ophthalmology of Harvard Medical School and does not imply an endorsement, sponsorship, or assumption of responsibility by either Harvard University or Harvard Medical School as a legal identity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.715, + 0.644, + 0.731 + ], + "angle": 0, + "content": "3.2. Data Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.738, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The glaucoma detection dataset contains 15,725 training samples, 1,047 validation samples, and 4,287 samples from 5,640, 376, and 1,505 glaucoma patients, respectively. Half of the samples are unlabeled. Among the half samples with labels, \\(45.2\\%\\) of the samples are glaucoma, while the rest are non-glaucoma. The progression forecasting dataset includes 5,889 training samples, 431 validation samples, and 1,547 test samples from 1,292, 86, and 345 glaucoma patients, respectively. Among the half samples with progression outcome labels, \\(24.0\\%\\) and \\(2.6\\%\\) of the samples are TD Progression and MD Fast Progression, respectively. The" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20473" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "two clinically accepted criteria to define glaucoma progression: (1) TD Progression: eyes with at least three locations with TD slope \\(\\leq -1\\) dB; (2) MD Fast Progression: eyes with MD slope \\(\\leq -1\\) dB [78]. The MD Fast Progression is more detrimental for the patients than TD progression [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.36 + ], + "angle": 0, + "content": "Figure 2 presents the pointwise distributions of 25th, 50th, and 75th percentile RNFLTs in the train (validation set included) and test sets, which demonstrate two obvious superotemporal and inferotemporal RNFL bundles with their damages are known closely related to the visual function loss. As in the two representative examples shown in Figure 1, RNFLT maps with similar appearances may have different labels. This suggests the class boundary may be obscure to determine, which naturally prompts us to adopt semi-supervised solutions to address the glaucoma detection and progression forecasting problems which fully leverage the abundant information in unlabeled samples to mitigate the challenges." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.471, + 0.673 + ], + "angle": 0, + "content": "The released dataset including 1,000 patients is a random subset from the entire dataset used to build the glaucoma detection and progression forecasting models in this paper, it naturally inherits the original data characteristics. The released glaucoma detection dataset contains 1,000 samples from 1,000 patients, where each sample includes two data modalities (i.e. RNFLT map and VF test) and respective glaucoma labels. The progression forecasting dataset is a subset from the released glaucoma detection dataset, which entails 500 patients with progression labels and patient demographics including age, sex, and race. For all glaucoma patients, the age distribution of 25th, 50th, and 75th percentiles are 54.3, 64.5, and 72.8 years, respectively. The male and female account for \\(56\\%\\) and \\(44\\%\\) of the 1,000 patients, respectively. Among the 1,000 patients, the White, Black, and Asian groups account for \\(76.4\\%\\), \\(14.9\\%\\), and \\(8.7\\%\\), respectively. To the best of our knowledge, this is the first public glaucoma dataset with the patient demographic information available, which carries great potential to enable interested researchers to conduct and evaluate fairness learning studies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.691, + 0.426, + 0.709 + ], + "angle": 0, + "content": "4. Problem Formalization & Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.747 + ], + "angle": 0, + "content": "In this section, we formally introduce the problems of glaucoma detection and progression forecasting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Denote \\(\\mathcal{D}^l = \\{(x_i,y_i)|1\\leq i\\leq N^l\\}\\) be a labeled dataset, where \\(x\\in \\mathbb{R}^n\\) are the RNFLT maps and \\(y\\in \\mathcal{V}\\) are corresponding glaucoma/progression labels. Also, we have a stack of unlabeled RNFLT maps, i.e. \\(\\mathcal{D}^{ul} = \\{x_j|1\\leq j\\leq N^{ul}\\}\\). We assume that we have a classification model \\(f:\\mathbb{R}^n\\xrightarrow{\\theta}\\mathcal{Y}(f_\\theta\\) for short), where \\(\\theta\\) are the parameters. To update the classification model with the labeled data, a predefined loss function \\(\\ell :\\mathcal{V}\\to \\mathcal{V}\\) is used to gauge the discrepancy between the predictions and the ground-truth labels such that the classification model can be updated along" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.087, + 0.912, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.251, + 0.91, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.434 + ], + "angle": 0, + "content": "Figure 3: The definitions of (a) glaucoma detection task; and (b) progression forecasting task. In the VF test, blue indicates normal visual function while red indicates abnormal visual function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.884, + 0.482 + ], + "angle": 0, + "content": "the loss, i.e. \\( \\min_{\\theta} \\ell(f_{\\theta}(x), y) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.736 + ], + "angle": 0, + "content": "As shown in Figure 3 (a), the practical utility scenario of glaucoma detection with deep learning is to use OCT scans that can be easily measured in the setting of primary care or pharmacy to predict glaucoma diagnosis defined by visual field tests that typically only can be measured in the setting of an ophthalmology clinic or eye hospital with specialized technicians. The clinical relevance of progression forecasting as shown in Figure 3 (b) with deep learning is that clinicians can better decide if a patient should be treated more conservatively with eye drops or aggressively with invasive surgeries. As suggested by conventional semi-supervised learning methods [20, 2, 40, 68, 11], pseudo labels \\(\\tilde{y}\\) are predicted by a supervisor model \\(\\pi : \\mathbf{R}^n \\xrightarrow{\\omega} \\mathcal{Y}\\) with the unlabeled data \\(\\tilde{x}\\), i.e. \\(\\tilde{y} = \\pi(\\tilde{x};\\omega)\\). Similar to the supervised learning setting, the unlabeled data with the pseudo labels can participate in the supervised learning process, namely \\(\\min_{\\theta} \\ell(f_{\\theta}(\\tilde{x}), \\tilde{y})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, as shown in Figure 4, the natural images associated with different visual concepts are relatively easy to be distinguished. In contrast, the differences in RNFLT maps between glaucoma versus non-glaucoma (similarly for progression versus non-progression) can be very subtle due to large retinal physiological inter-subject variations. There might not be clear and adequate guidance in the learning process to determine the underlying labels of the unlabeled RNFLT maps. Instead, a more sensible way to better determine the pseudo labels is to apply a policy to obtain pseudo labels. Then, we observe how these pseudo labels affect" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20474" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.458, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.171, + 0.47, + 0.26 + ], + "angle": 0, + "content": "Figure 4: A challenge in glaucoma detection and progression forecasting. Unlike the classification with natural images, where the images with different labels are likely visually distinguishable, RNFLT images with different labels (e.g. glaucoma vs. non-glaucoma and progression vs nonprogression) might not be easily distinguishable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.471, + 0.378 + ], + "angle": 0, + "content": "the training process of the classifier and adjust the policy to predict the pseudo labels such that the generalizability of the classifier learned with the pseudo labels can be maximized. In short, the pseudo labels predicted by the supervisor model should be aligned with the improvement of the generalizability of the classifier on a set of unseen data \\( \\mathcal{D}^{unseen} \\). This objective can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.386, + 0.47, + 0.432 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\{\\tilde {y} \\}} \\frac {1}{| \\mathcal {D} ^ {\\text {u n s e e n}} |} \\sum_ {(x, y) \\in \\mathcal {D} ^ {\\text {u n s e e n}}} \\ell \\left(f _ {\\theta^ {*}} (x), y\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.43, + 0.437, + 0.464 + ], + "angle": 0, + "content": "where \\(\\theta^{*}\\in \\arg \\min_{\\theta}\\frac{1}{|\\mathcal{D}^{l + ul}|}\\sum_{(x,y)\\in \\mathcal{D}^{l + ul}}\\ell (f_{\\theta}(x),y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.47, + 0.506 + ], + "angle": 0, + "content": "\\(\\mathcal{D}^{l + ul}\\) is the union of labeled dataset \\(\\mathcal{D}\\) and unlabeled dataset with pseudo labels \\(\\mathcal{D}^{ul}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.518, + 0.469, + 0.551 + ], + "angle": 0, + "content": "5. Generalization-Reinforced Pseudo Supervisor" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.469, + 0.693 + ], + "angle": 0, + "content": "Different from prior-work semi-supervised learning methods [68, 11, 10, 53], the proposed method explicitly leverages the empirical generalizability of the classifier that is quantified with unseen RNFLT maps (e.g. the ones from the validation set) to optimize the policy of predicting the pseudo labels. The empirical generalizability can be viewed as a type of informative reward as the goal of training a classifier is to find one that can generalize well on the unseen test data drawn from the underlying distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.469, + 0.812 + ], + "angle": 0, + "content": "Figure 5 shows how the proposed generalization-reinforced pseudo supervisor predicts pseudo labels and is trained according to the observed sets of states (maps), actions (pseudo labels), and rewards (the empirical generalizability with a mini-batch of validation samples). The proposed pseudo supervisor is based on the policy gradient [70], which is one type of generic reinforcement learning methods [8, 26, 65]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Here we formally introduce the concepts related to the proposed pseudo supervisor. An RNFLT map \\( x_{t} \\) at time step \\( t \\) can be viewed as a state \\( s_t \\in S \\), while the predicted pseudo label \\( \\tilde{y}_t \\) can be viewed as an action \\( a_{t} \\in \\mathcal{A} \\). A reward function \\( r_t: S \\times \\mathcal{A} \\to \\mathbf{R} \\) is defined on the state-action space. The proposed pseudo supervisor that employs" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "a policy to determine pseudo labels is denoted as \\(\\pi_{\\omega} : S \\to \\mathcal{P}(\\mathcal{A})\\), where \\(\\mathcal{P}(\\mathcal{A})\\) are the probability measures on \\(\\mathcal{A}\\) and \\(\\omega\\) are the parameters of the proposed pseudo supervisor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.381 + ], + "angle": 0, + "content": "In the first place, the classifier takes labeled samples for training. The unlabeled samples will be fed into the pseudo supervisor for predicting pseudo labels. Before updating the weights of the classifier, the empirical generalizability of classifier \\( f_{\\theta_t} \\) can be quantified with validation samples \\( (x^{(val)}, y^{(val)}) \\), that is, \\( \\ell(f_{\\theta_t}(x^{(val)}), y^{(val)}) (\\ell_t^{(val)}) \\). Specifically, \\( \\ell \\) is the widely-used cross-entropy loss for classification. After updating the weights of the classifier with unlabeled samples and corresponding pseudo labels, given the same mini-batch of validation samples, the empirical generalizability of the updated classifier can be requantified again, i.e., \\( \\ell(f_{\\theta_{t+1}}(x^{(val)}), y^{(val)}) (\\ell_{t+1}^{(val)}) \\). Then, the discrepancy between \\( \\ell_t^{(val)} \\) and \\( \\ell_{t+1}^{(val)} \\) is informative to measure the effect of updating the classifier with the unlabeled samples and the pseudo labels. In this sense, it is desired to use in the reward function, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.389, + 0.892, + 0.41 + ], + "angle": 0, + "content": "\\[\nr _ {t} = \\max \\left(\\exp \\left(\\ell_ {t} ^ {(v a l)} - \\ell_ {t + 1} ^ {(v a l)}\\right) - 1, 0\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.478 + ], + "angle": 0, + "content": "In Eq. (2), if \\(\\ell_{t+1}^{(val)}\\) is far less than \\(\\ell_t^{(val)}\\), it implies that the unlabeled samples and pseudo labels lead to a positive effect on the learning process such that the reward increases adaptively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.892, + 0.522 + ], + "angle": 0, + "content": "A trajectory of states, actions, and rewards within a time window \\(\\beta\\) will be saved. The expected return of policy optimization can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.527, + 0.892, + 0.568 + ], + "angle": 0, + "content": "\\[\nJ \\left(\\pi_ {\\omega}\\right) = \\mathbb {E} _ {\\pi_ {\\omega}} \\left(\\sum_ {k = 0} \\gamma^ {k} r _ {k}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.648 + ], + "angle": 0, + "content": "where \\(\\gamma \\in [0,1]\\) is the discount rate for the numerical stability of reinforcement learning, and \\(\\beta\\) is the time window for updating \\(\\pi_{\\omega}\\) periodically. According to the policy gradient theorem [70], the expectation of the sample gradient is equal to the actual gradient" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.655, + 0.891, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\pi_ {\\omega}} J (\\pi_ {\\omega}) = \\mathbb {E} _ {\\pi_ {\\omega}} \\left[ \\sum_ {k = 0} ^ {\\beta} \\gamma^ {k} r _ {t + k + 1} \\nabla_ {\\pi_ {\\omega}} \\log \\pi_ {\\omega} \\left(a _ {t} \\mid s _ {t}\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.719, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Then the policy parameterization can be updated according to the gradient ascend rule" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.758, + 0.891, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\omega} \\leftarrow \\pi_ {\\omega} + \\eta_ {\\pi} \\nabla_ {\\pi_ {\\omega}} J (\\pi_ {\\omega}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.783, + 0.736, + 0.798 + ], + "angle": 0, + "content": "where \\(\\eta_{\\pi}\\) is the learning rate of \\(\\pi_{\\omega}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.809, + 0.633, + 0.826 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.787, + 0.85 + ], + "angle": 0, + "content": "6.1. Set-Up & Implementation Detail" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To optimize our models, we employ the AdamW optimizer [46] and train all models for 10 epochs in all experiments conducted in this study. For supervised classification" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20475" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.088, + 0.791, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Figure 5: An overview of the proposed generalization-reinforced pseudo supervisor in the semi-supervised learning framework. The learning framework consists of the process of training the classifier, the process of measuring the empirical generalizability of the classifier to be used as rewards, and the process of updating the proposed pseudo supervisor to predict pseudo labels that can maximize the classifier's empirical generalizability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.473, + 0.595 + ], + "angle": 0, + "content": "models for glaucoma detection, we use a learning rate of \\(4 \\times 10^{-5}\\) and weight decay of 0, and for progression forecasting, we use a learning rate of \\(2 \\times 10^{-5}\\) and weight decay of 0. These hyperparameters are determined to achieve the best baseline performance based on our experiments. The proposed method is trained using a learning rate of \\(4 \\times 10^{-5}\\) and weight decay of 0. The batch sizes used in the supervised model, semi-supervised models, and the proposed method vary due to the different semi-supervised learning contexts [73, 11, 10, 68, 53]. We adopt the batch sizes that result in the best performance for each respective context. To ensure consistency with previous work, we apply data augmentation techniques to the samples in the semi-supervised learning models, including MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53]. Specifically, we utilize random crop and resize, as well as random horizontal flip. However, we do not apply these data augmentation techniques to the supervised baseline or the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.789 + ], + "angle": 0, + "content": "For the glaucoma detection task, we train all models with 7,860 labeled samples and 7,865 unlabeled samples for training. We use 1,047 and 4,287 labeled samples for validation and testing, respectively. For the progression forecasting task, all the models are trained with 2,944 labeled samples and 2,945 unlabeled samples for training. We use 431 and 1,547 labeled samples for validation and testing, respectively. For multi-modality modeling, we simply concatenate the RNFLT images and VF matrices for joint learning using the same architecture as our single-modality model and apply this model to the progression forecasting task. The model is trained and evaluated using the same train-val-test split as our single-modality experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.79, + 0.47, + 0.835 + ], + "angle": 0, + "content": "We use the accuracy, F1 score, and AUC on the test sets as our evaluation matrices. The code is written in PyTorch [55] and we use one RTX A6000 GPU for all experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.25, + 0.862 + ], + "angle": 0, + "content": "6.2. Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Our Supervised baseline is a standard ImageNet pretrained EfficientNet classifier [72] learning with cross" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.895, + 0.685 + ], + "angle": 0, + "content": "entropy using only labeled images. For Semi-Supervised baselines, we choose MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53] due to their previous powerful performances. MeanTeacher [73] is one of the initial methods to enforce models to predict consistent pseudo predictions for labeled and unlabeled data, which construct a target-generating teacher model by averaging the model weights. MixMatch [11], ReMixMatch [10], and FixMatch [68] are consistency-based SSL methods that are based on the data augmentations (i.e., weak and strong augmentation) to make consistent classification predictions between different views of unlabeled data. DASO [53] suggests that those methods fail for imbalanced datasets and proposes a pseudo-labeling framework based on different consistency SSL models by adaptively mixing the linear and semantic pseudo-labels to mitigate the overall bias between majority and minority classes. We adopt the FixMatch version of DASO, following the default setting reported in the paper [53]. We implement all baseline methods using Pytorch [55] from DASO's official codes and conducted experiments under the same codebase and experimental protocols for fair comparison. All models are equipped with the same EfficientNetV2-S [72] backbone and their hyper-parameters are specifically tuned to suit our Glaucoma detection and progression forecasting tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.699, + 0.78, + 0.714 + ], + "angle": 0, + "content": "6.3. Results for Glaucoma Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In Table 2, we evaluate our Pseudo-Sup model against multiple SOTA baselines on the glaucoma detection task. Our model achieves the best performance in terms of accuracy, F1 score, and AUC score. We compute the mean performances over 5 different runs. Notably, our model achieves such results without any specific design of image perturbations/augmentations, guaranteeing its simplicity and applicability to different tasks. Our method surpasses the previous SOTA method DASO by \\(6.8\\%\\) of accuracy, \\(11.1\\%\\) of F1 score, and \\(3.7\\%\\) of AUC, respectively. The best-performing consistency method FixMatch still worse by \\(0.4\\%\\) of accuracy, \\(1.2\\%\\) of F1 score, and \\(1.5\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20476" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.163 + ], + "angle": 0, + "content": "Table 2: Performance on the cross-sectional data for the glaucoma detection task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.165, + 0.466, + 0.274 + ], + "angle": 0, + "content": "
ModelAcc↑F1↑AUC↑
Baseline0.8131 ± 0.00390.8057 ± 0.00330.8693 ± 0.0022
MeanTeacher0.8148 ± 0.00530.8053 ± 0.00720.8541 ± 0.0040
MixMatch0.6392 ± 0.01610.6273 ± 0.00590.8451 ± 0.0062
ReMixMatch0.6741 ± 0.00190.6716 ± 0.00180.8442 ± 0.0032
DASO0.7444 ± 0.01150.6929 ± 0.00500.8371 ± 0.0118
FixMatch0.8077 ± 0.00320.7912 ± 0.00210.8574 ± 0.0019
Pseudo-Sup0.8124 ± 0.00250.8038 ± 0.00460.8727 ± 0.0006
Pseudo-Sup+Aug0.8224 ± 0.01010.8128 ± 0.01220.8818 ± 0.0020
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.47, + 0.42 + ], + "angle": 0, + "content": "of AUC than our Pseudo-Sup model, indicating the effectiveness of our approach and the challenges of designing appropriate augmentations for RNFLT maps from previous approaches. We also apply weak augmentations (similar to FixMatch) to our Pseudo-Sup (Pseudo-Sup+Aug) leading to improved performances. Please note that consistency methods (e.g., DASO, FixMatch, etc.) employ both weak and strong augmentations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.432, + 0.385, + 0.449 + ], + "angle": 0, + "content": "6.4. Results for Progression Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.619 + ], + "angle": 0, + "content": "Unimodal Results. We further evaluate our model on progression forecasting, a more clinically significant task than glaucoma detection, as shown in Table 3. Accurate forecasting of glaucoma progression plays a vital role in prescribing appropriate treatment plans to efficiently prevent further vision losses and is important for assisting the development of new pharmaceutical medicines to treat glaucoma. Adding weak augmentations further improves the results for both progression tasks. Please note that similar to glaucoma detection experiments, we also compute the mean performances over 5 different runs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.827 + ], + "angle": 0, + "content": "For TD Progression, our model achieves the best \\(74.47\\%\\) of mean AUC, respectively. When compared with consistency-based approaches, our method improves between \\(0.5\\% - 2.1\\%\\) in AUC. For MD Fast Progression, our approach consistently achieves the best AUC performance, illustrating the effectiveness and accuracy of our method regardless of any glaucoma progression criteria. Given that most of the glaucoma patients are non-progressive, our progression dataset for TD Progression and MD Fast Progression are both highly imbalanced (i.e., most of the samples are labeled as non-progression). Therefore, it is worth noting that our method also outperforms the previous imbalanced SSL SOTA DASO by a large margin in all measures under such imbalanced scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.827, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Multimodal Results. One of the major contributions of this work is to introduce a multi-modal dataset for glaucoma progression forecasting. To the best of our knowledge, this is the first work to explore and benchmark SSL approaches using multimodal model inputs. In Table 5, we show the" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.164 + ], + "angle": 0, + "content": "Table 3: Performance on the longitudinal data with single modality RNFLT for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while Pseudo-Sup represents the proposed pseudo supervisor method." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.165, + 0.865, + 0.291 + ], + "angle": 0, + "content": "
ModelsAUC ↑ (TD)AUC ↑ (MD Fast)
Baseline0.7367 ± 0.01260.6677 ± 0.0017
MeanTeacher0.7367 ± 0.01260.6677 ± 0.0017
MixMatch0.7238 ± 0.00980.6518 ± 0.0069
ReMixMatch0.7390 ± 0.01240.6595 ± 0.0083
DASO0.7231 ± 0.01880.6787 ± 0.0148
FixMatch0.7395 ±0.01520.6747 ± 0.0054
Pseudo-Sup0.7447 ± 0.00840.7004 ± 0.0218
Pseudo-Sup+Aug0.7477 ± 0.00740.7215 ± 0.0162
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.51 + ], + "angle": 0, + "content": "progression forecasting results using two modalities (RNFLT and VF). Our model achieves the best performance on all evaluation measures. To keep its simplicity and adaptability, we concatenate the RNFLT and the up-scaled VF together as the inputs of our model. Please note that consistency SSL approaches that are based on image augmentations are not included in this table because of the lack of design of augmentation techniques in multimodal medical data. The SOTA results with multimodal glaucoma data enable more accurate and applicable potential computer-aided diagnosis (CAD) systems in real clinical settings." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.544, + 0.86, + 0.56 + ], + "angle": 0, + "content": "6.5. Supervised Benchmarks on Released Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In Table 6, we show the supervised classification results for the progression forecasting tasks with multiple SOTA supervised CNN and transformer baseline methods, including VGG [67], ResNet [27], ResNext [82], WideResNet [83], EfficientNet [71], ConvNext [45], ViT [22], and Swin Transformers [44]. This benchmark is conducted on our future data release with 500 patients upon acceptance, of which 400 patients are used for training and the remaining are used for testing. To the best of our knowledge, this is the first supervised glaucoma progression forecasting benchmark, which aims to provide broader impacts for the computer vision and medical imaging communities to build clinically effective computer-aided diagnosis systems. Moreover, the results are reported with both single RNFLT modality and multi-modality (RNFLT and VF) samples. It is worth noting that the accuracy and F1 are the same on MD fast progression because the testing set only contains 100 samples under highly imbalanced class distributions. For more experimental details regarding this supervised progression forecasting benchmark and the supervised results on future release data with 1,000 patients for the glaucoma detection task, please refer to the Supplemental Material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20477" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 4: Performance comparisons for glaucoma detection task on LAG dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.118, + 0.462, + 0.224 + ], + "angle": 0, + "content": "
ModelAcc↑F1↑AUC↑
MeanTeacher0.82800.82790.9082
MixMatch0.50200.33780.9075
ReMixMatch0.69800.47850.8369
DASO0.79710.79060.9164
FixMatch0.83660.82970.9198
Pseudo-Sup0.91600.91580.9858
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.357 + ], + "angle": 0, + "content": "Table 5: Performance on the longitudinal data with two modalities (RNFLT and VF) for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method. Please note that most methods in Table 3 are based on weak/strong augmentations [68, 11, 10] requiring RGB channels, which are not adaptable for multi-modal learning." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.359, + 0.432, + 0.424 + ], + "angle": 0, + "content": "
ModelsAUC (TD)AUC (MD Fast)
Baseline0.7343 ± 0.00980.6814 ± 0.0081
MeanTeacher0.7436 ± 0.01330.6697 ± 0.0035
Pseudo-Sup0.7517 ± 0.00410.7265 ± 0.0095
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.452, + 0.285, + 0.467 + ], + "angle": 0, + "content": "6.6. Performances on LAG" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.653 + ], + "angle": 0, + "content": "In Table 4, we show the performance comparisons on LAG dataset [37]. LAG is a large-scale fundus imaging dataset for glaucoma detection. For this experiment, we use 3,854 images for training with half with labels and the remaining without labels, and 1,000 images for testing, with the size of \\(224 \\times 224\\). For fair comparisons, we use the same hyper-parameters and experimental setups as our other benchmarks. It is worth noting that our approach surpasses the SOTA methods by a large margin and achieves the highest \\(98.58\\%\\) AUC with only half-labeled data, illustrating the effectiveness and applicability of our proposed Pseudo-Sup in different data and tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.669, + 0.231, + 0.685 + ], + "angle": 0, + "content": "6.7. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.471, + 0.901 + ], + "angle": 0, + "content": "According to Section 5, the time window \\(\\beta\\) and discount rate \\(\\gamma\\) are the most important hyperparameters affecting the proposed pseudo supervisor. The ablation studies of the two hyperparameters are shown in Figure 6. We can see that the proposed pseudo supervisor behaviors stably with various \\(\\beta\\) on AUC. Specifically, when \\(\\beta = 50\\) leads to better performance. The consistent pattern can be also observed in the analysis of \\(\\gamma\\). The discount factor is used to achieve a trade-off between the rewards in the distant future and the rewards in the near future. If it is 0, reinforcement learning agents would only look for an optimal policy for the first action. As shown in Figure 6 (b), we find the optimal discount rate at 0.9, which implies that the states, actions, and rewards in the progression task are coherent to a certain degree." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.149 + ], + "angle": 0, + "content": "Table 6: Performance of different supervised methods on the longitudinal data with single/multi-modality for the progression forecasting tasks on the released 500 glaucoma progression data." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.246 + ], + "angle": 0, + "content": "
ModelTD ProgressionMD Fast Progression
Acc†F1†AUC†Acc†F1†AUC†
VGG [67]0.78/0.800.75/0.740.84/0.790.98/0.980.49/0.490.78/0.68
ResNet [27]0.75/0.740.64/0.670.74/0.750.98/0.980.49/0.490.71/0.59
ResNext [82]0.78/0.770.65/0.680.78/0.760.98/0.980.49/0.490.61/0.80
WideResNet [83]0.77/0.790.71/0.750.77/0.800.98/0.980.49/0.490.71/0.64
EfficientNet [71]0.73/0.780.66/0.730.76/0.790.98/0.980.49/0.490.55/0.66
ConvNext [45]0.74/0.810.62/0.770.78/0.810.98/0.980.49/0.490.84/0.62
ViT [22]0.73/0.770.54/0.710.68/0.790.98/0.980.49/0.490.92/0.69
Swin [44]0.71/0.770.49/0.720.68/0.770.98/0.980.49/0.490.87/0.57
" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.263, + 0.691, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.372, + 0.676, + 0.384 + ], + "angle": 0, + "content": "(a) The impact of \\(\\beta\\)" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.261, + 0.871, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.371, + 0.856, + 0.384 + ], + "angle": 0, + "content": "(b) The impact of \\(\\gamma\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.398, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Figure 6: Ablation study of hyperparameters' impact on TD Progression forecasting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.767, + 0.47 + ], + "angle": 0, + "content": "6.8. Experiment on Racial Groups" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.566 + ], + "angle": 0, + "content": "The patients in our dataset are from three racial groups (White (0.88 AUC), Black (0.83 AUC), and Asian (0.85 AUC)). In our experiments, we observe that black patients have the lowest AUC yet the highest glaucoma prevalence [62, 25]. Therefore, our dataset is well-suited for future fairness studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.619, + 0.595 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have made two contributions. First, we have developed a generalization-reinforced semi-supervised learning model termed pseudo supervisor for glaucoma detection and progression forecasting. Specifically, the pseudo supervisor predicts pseudo labels following a policy that optimizes the generalization error in training the classifier. Our pseudo supervisor model overall demonstrates superior performance over a number of SOTA methods. Second, we release a multimodal and multi-task dataset for glaucoma detection and progression forecasting with SOTA 3D OCT imaging data. Our dataset is the largest by far for glaucoma detection with 3D OCT data and the first for progression forecasting task. More importantly, our dataset entails demographic information including gender and race, which is generally unavailable in existing public glaucoma datasets. Given the diverse racial representation in our dataset as detailed in Section 3, our dataset is well equipped for potential fairness learning studies, especially considering the significant glaucoma prevalence disparity between Black and other races [62, 25]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20478" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.229, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.175 + ], + "angle": 0, + "content": "This work was supported by NIH R00 EY028631, Research to Prevent Blindness International Research Collaborators Award, Alcon Young Investigator Grant, NIH R01 EY030575, and NIH P30 EY003790." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.191, + 0.174, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[1] Jin Mo Ahn, Sangsoo Kim, Kwang-Sung Ahn, Sung-Hoon Cho, Kwan Bok Lee, and Ungsoo Samuel Kim. A deep learning model for the detection of both advanced and early glaucoma using fundus photography. PloS one, 13(11):e0207982, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.472, + 0.369 + ], + "angle": 0, + "content": "[2] Manal Al Ghamdi, Mingqi Li, Mohamed Abdel-Mottaleb, and Mohamed Abou Shousha. Semi-supervised transfer learning for convolutional neural networks for glaucoma detection. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3812-3816. IEEE, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.47, + 0.438 + ], + "angle": 0, + "content": "[3] Andrew J Anderson, Abadh K Chaurasia, Ajay Sharma, Amisha Gupta, Shikha Gupta, Anurag Khanna, and Viney Gupta. Comparison of rates of fast and catastrophic visual field loss in three glaucoma subtypes. Investigative ophthalmology & visual science, 60(1):161-167, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.441, + 0.47, + 0.494 + ], + "angle": 0, + "content": "[4] Paul H Artes and Balwantray C Chauhan. Longitudinal changes in the visual field and optic disc in glaucoma. Progress in retinal and eye research, 24(3):333-354, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[5] Ryo Asaoka, Hiroshi Murata, Kazunori Hirasawa, Yuri Fujino, Masato Matsuura, Atsuya Miki, Takashi Kanamoto, Yoko Ikeda, Kazuhiko Mori, Aiko Iwase, et al. Using deep learning and transfer learning to accurately diagnose early-onset glaucoma from macular optical coherence tomography images. American journal of ophthalmology, 198:136-145, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.594, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[6] Muhammad Naseer Bajwa, Gur Amrit Pal Singh, Wolfgang Neumeier, Muhammad Imran Malik, Andreas Dengel, and Sheraz Ahmed. G1020: A benchmark retinal fundus image dataset for computer-aided glaucoma detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.678, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[7] Francisco José Fumero Batista, Tinguaro Diaz-Aleman, Jose Sigut, Silvia Alayon, Rafael Arny, and Denisse Angel-Pereira. Rim-one dl: A unified retinal image database for assessing glaucoma using deep learning. Image Analysis & Stereology, 39(3):161-167, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.747, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[8] Jonathan Baxter and Peter L Bartlett. Infinite-horizon policy-gradient estimation. Journal of Artificial Intelligence Research, 15:319-350, 2001. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.79, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[9] Mohammed El Amine Bechar, Nesma Settouti, Vincent Barra, and Mohamed Amine Chikh. Semi-supervised superpixel classification for medical images segmentation: application to detection of glaucoma disease. Multidimensional Systems and Signal Processing, 29(3):979-998, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[10] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.217, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and augmentation anchoring. In International Conference on Learning Representations, 2019. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[11] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32:5049-5059, 2019. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[12] Igor I Bussel, Gadi Wollstein, and Joel S Schuman. Oct for glaucoma diagnosis, screening and detection of glaucoma progression. British Journal of Ophthalmology, 98(Suppl 2):ii15-ii19, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.248, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[13] Paola Cascante-Bonilla et al. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. arXiv preprint arXiv:2001.06001, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[14] Xiangyu Chen, Yanwu Xu, Shuicheng Yan, Damon Wing Kee Wong, Tien Yin Wong, and Jiang Liu. Automatic feature learning for glaucoma detection based on deep learning. In International conference on medical image computing and computer-assisted intervention, pages 669-677. Springer, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[15] Mark Christopher, Akram Belghith, Christopher Bowd, James A Proudfoot, Michael H Goldbaum, Robert N Weinreb, Christopher A Girkin, Jeffrey M Liebmann, and Linda M Zangwill. Performance of deep learning architectures and transfer learning for detecting glaucomatous optic neuropathy in fundus photographs. Scientific reports, 8(1):1-13, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[16] Mark Christopher, Akram Belghith, Robert N Weinreb, Christopher Bowd, Michael H Goldbaum, Luke J Saunders, Felipe A Medeiros, and Linda M Zangwill. Retinal nerve fiber layer features identified by unsupervised machine learning on optical coherence tomography scans predict glaucoma progression. Investigative ophthalmology & visual science, 59(7):2748-2756, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[17] Anne Louise Coleman. Advances in glaucoma treatment and management: surgery. Investigative ophthalmology & visual science, 53(5):2491-2494, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[18] Ronan Conlon, Hady Saheb, and Iqbal Ike K Ahmed. Glaucoma treatment trends: a review. Canadian Journal of Ophthalmology, 52(1):114-124, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[19] David P Crabb, Nicholas D Smith, Fiona C Glen, Robyn Burton, and David F Garway-Heath. How does glaucoma look?: patient perception of visual field loss. Ophthalmology, 120(6):1120-1126, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[20] Andres Diaz-Pinto, Adrián Colomer, Valery Naranjo, Sandra Morales, Yanwu Xu, and Alejandro F Frangi. Retinal image synthesis and semi-supervised learning for glaucoma assessment. IEEE transactions on medical imaging, 38(9):2211-2218, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[21] Andres Diaz-Pinto, Sandra Morales, Valery Naranjo, Thomas Kohler, Jose M Mossi, and Amparo Navea. Cnns for automatic glaucoma assessment using fundus images: an extensive validation. Biomedical engineering online, 18(1):1-19, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20479" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.119 + ], + "angle": 0, + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[23] Rui Fan, Christopher Bowd, Nicole Brye, Mark Christopher, Robert N Weinreb, David Kriegman, and Linda Zangwill. One-vote veto: Semi-supervised learning for low-shot glaucoma diagnosis. arXiv preprint arXiv:2012.04841, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.469, + 0.241 + ], + "angle": 0, + "content": "[24] Huihui Fang, Fei Li, Huazhu Fu, Junde Wu, Xiulan Zhang, and Yanwu Xu. Dataset and evaluation algorithm design for goals challenge. In International Workshop on Ophthalmic Medical Image Analysis, pages 135-142. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.243, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[25] David S Friedman, Henry D Jampel, Beatrix Munoz, and Sheila K West. The prevalence of open-angle glaucoma among blacks and whites 73 years and older: the salisbury eye evaluation glaucoma study. Archives of ophthalmology, 124(11):1625-1630, 2006. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.312, + 0.469, + 0.353 + ], + "angle": 0, + "content": "[26] Mohammad Ghavamzadeh and Sridhar Mahadevan. Hierarchical policy gradient algorithms. Computer Science Department Faculty Publication Series, page 173, 2003. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.408, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[28] Patrick Herbert, Kaihua Hou, Chris Bradley, Michael V Boland, Pradeep Y Ramulu, Mathias Unberath, and Jithin Yohannan. Forecasting risk of future rapid glaucoma worsening using early visual field, optical coherence tomography and clinical data. Investigative Ophthalmology & Visual Science, 63(7):2294-2294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[29] Donald C Hood and Randy H Kardon. A framework for comparing structural and functional measures of glaucomatous damage. Progress in retinal and eye research, 26(6):688-710, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[30] Sen Hu, Nicholas D Smith, Luke J Saunders, and David P Crabb. Patterns of binocular visual field loss derived from large-scale patient data from glaucoma clinics. Ophthalmology, 122(12):2399-2406, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.6, + 0.469, + 0.666 + ], + "angle": 0, + "content": "[31] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, et al. Optical coherence tomography. science, 254(5035):1178-1181, 1991. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.668, + 0.469, + 0.776 + ], + "angle": 0, + "content": "[32] Alireza Kamalipour, Susan Moghimi, Pooya Khosravi, Vahid Mohammadzadeh, Takashi Nishida, Eleonora Micheletti, Jo-Hsuan Wu, Golnoush Mahmoudinezhad, Elizabeth HF Li, Mark Christopher, et al. Combining oct and oct-angiography longitudinal data for the detection of visual field progression in glaucoma: Longitudinal oct/octa data to detect glaucoma progression. American Journal of Ophthalmology, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.778, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[33] Stuart Keel, Jinrong Wu, Pei Ying Lee, Jane Scheetz, and Mingguang He. Visualizing deep learning models for the detection of referable diabetic retinopathy and glaucoma. JAMA ophthalmology, 137(3):288-292, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[34] Philip J Kellman and Thomas F Shipley. A theory of visual interpolation in object perception. Cognitive psychology, 23(2):141-221, 1991. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Yuka Kihara, Giovanni Montesano, Andrew Chen, Nishani Amerasinghe, Chrysostomos Dimitriou, Aby Jacob, Almira" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Chabi, David P Crabb, and Aaron Y Lee. Policy-driven, multimodal deep learning for predicting visual fields from the optic disc and oct imaging. Ophthalmology, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[36] Oleksandr Kovalyk, Juan Morales-Sanchez, Rafael Verdú-Monedero, Inmaculada Sellés-Navarro, Ana Palazon-Cabanes, and José-Luis Sancho-Gómez. Papila: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment. Scientific Data, 9(1):1-12, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.216, + 0.892, + 0.283 + ], + "angle": 0, + "content": "[37] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: a large-scale database and cnn model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.285, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[38] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: A large-scale database and cnn model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.353, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[39] Bingyu Liu, Yuhong Guo, Jieping Ye, and Weihong Deng. Selective pseudo-labeling with reinforcement learning for semi-supervised domain adaptation. arXiv preprint arXiv:2012.03438, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.408, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[40] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[41] Fengbei Liu, Yu Tian, Filipe R Cordeiro, Vasileios Bela-giannis, Ian Reid, and Gustavo Carneiro. Self-supervised mean teacher for semi-supervised chest x-ray classification. In International Workshop on Machine Learning in Medical Imaging, pages 426-436. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[42] Hanruo Liu, Liu Li, I Michael Wormstone, Chunyan Qiao, Chun Zhang, Ping Liu, Shuning Li, Huaizhou Wang, Dapeng Mou, Ruiqi Pang, et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs. JAMA ophthalmology, 137(12):1353-1360, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.641, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[43] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.722, + 0.892, + 0.79 + ], + "angle": 0, + "content": "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.791, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[45] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20480" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[47] Rizwan Malik, William H Swanson, and David F Garway-Heath. 'structure-function relationship' in glaucoma: past thinking and current concepts. Clinical & experimental ophthalmology, 40(4):369-380, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.215 + ], + "angle": 0, + "content": "[48] Felipe A Medeiros, Alessandro A Jammal, and Atalie C Thompson. From machine to machine: an oct-trained deep learning algorithm for objective quantification of glaucomatous damage in fundus photographs. Ophthalmology, 126(4):513-521, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.217, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[49] Parmita Mehta, Christine A Petersen, Joanne C Wen, Michael R Banitt, Philip P Chen, Karine D Bojikian, Catherine Egan, Su-In Lee, Magdalena Balazinska, Aaron Y Lee, et al. Automated detection of glaucoma with interpretable machine learning using clinical data and multimodal retinal images. American Journal of Ophthalmology, 231:154-169, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.312, + 0.471, + 0.339 + ], + "angle": 0, + "content": "[50] MJ Morgan and RJ Watt. Mechanisms of interpolation in human spatial vision. Nature, 299(5883):553-555, 1982. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.341, + 0.471, + 0.394 + ], + "angle": 0, + "content": "[51] Jacqueline M Nelson-Quigg, Kimberly Cello, and Chris A Johnson. Predicting binocular visual field sensitivity from monocular visual field results. Investigative ophthalmology & visual science, 41(8):2212-2221, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.471, + 0.463 + ], + "angle": 0, + "content": "[52] Kouros Nouri-Mahdavi, Vahid Mohammadzadeh, Alessandro Rabiolo, Kiumars Edalati, Joseph Caprioli, and Siamak Yousefi. Prediction of visual field progression from oct structural measures in moderate to advanced glaucoma. American Journal of Ophthalmology, 226:172-181, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.471, + 0.517 + ], + "angle": 0, + "content": "[53] Youngtaek Oh, Dong-Jin Kim, and In So Kweon. Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. arXiv preprint arXiv:2106.05682, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.471, + 0.611 + ], + "angle": 0, + "content": "[54] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical image analysis, 59:101570, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.471, + 0.763 + ], + "angle": 0, + "content": "[55] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.765, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[56] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.819, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[57] Sonia Phene, R Carter Dunn, Naama Hammel, Yun Liu, Jonathan Krause, Naho Kitade, Mike Schaekermann, Rory Sayres, Derek J Wu, Ashish Bora, et al. Deep learning and glaucoma specialists: the relative importance of optic disc features to predict glaucoma referral in fundus photographs. Ophthalmology, 126(12):1627-1639, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[58] Harry A Quigley. Number of people with glaucoma worldwide. British journal of ophthalmology, 80(5):389-393, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[59] Harry A Quigley and Aimee T Broman. The number of people with glaucoma worldwide in 2010 and 2020. British journal of ophthalmology, 90(3):262-267, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[60] An Ran Ran, Carol Y Cheung, Xi Wang, Hao Chen, Lu-yang Luo, Poemen P Chan, Mandy OM Wong, Robert T Chang, Suria S Mannil, Alvin L Young, et al. Detection of glaucomatous optic neuropathy with spectral-domain optical coherence tomography: a retrospective training and validation deep-learning analysis. The Lancet Digital Health, 1(4):e172-e182, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[61] Mamshad Nayeem Rizve et al. In defense of pseudolabeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In International Conference on Learning Representations, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[62] Alicja R Rudnicka, Shahrul Mt-Isa, Christopher G Owen, Derek G Cook, and Deborah Ashby. Variations in primary open-angle glaucoma prevalence by age, gender, and race: a bayesian meta-analysis. Investigative ophthalmology & visual science, 47(10):4254-4261, 2006. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[63] Weiwei Shi, Yihong Gong, Chris Ding, Zhiheng MaXiaoyu Tao, and Nanning Zheng. Transductive semi-supervised deep learning using min-max features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 299-315, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[64] Naoto Shibata, Masaki Tanito, Keita Mitsuhashi, Yuri Fujino, Masato Matsuura, Hiroshi Murata, and Ryo Asaoka. Development of a deep residual learning algorithm to screen for glaucoma from fundus photography. Scientific reports, 8(1):1-9, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[65] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In International conference on machine learning, pages 387-395. PMLR, 2014. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.597, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[66] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy P. Lillicrap, Fan Hui, Laurent Sifre, George van den Driessche, Thore Graepel, and Demis Hassabis. Mastering the game of go without human knowledge. Nat., 550(7676):354-359, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.694, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.737, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[68] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33, 2020. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.819, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[69] Yi Sun, Aiming Chen, Minjie Zou, Yichi Zhang, Ling Jin, Yi Li, Danying Zheng, Guangming Jin, and Nathan Congdon. Time trends, associations and prevalence of blindness and vision loss due to glaucoma: an analysis of observational data from the global burden of disease study 2017. BMJ open, 12(1):e053805, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20481" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[70] Richard S. Sutton, David A. McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems 12, [NIPS Conference, Denver, Colorado, USA, November 29 - December 4, 1999], pages 1057-1063, 1999. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.176, + 0.471, + 0.23 + ], + "angle": 0, + "content": "[71] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[72] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10096-10106. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.47, + 0.439 + ], + "angle": 0, + "content": "[74] Yih-Chung Tham, Xiang Li, Tien Y Wong, Harry A Quigley, Tin Aung, and Ching-Yu Cheng. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology, 121(11):2081-2090, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[75] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W VERJANS, and Rajvinder Singh. Self-supervised multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[76] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 128-140. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[77] Rohit Varma, Paul P Lee, Ivan Goldberg, and Sameer Kotak. An assessment of the health and economic burdens of glaucoma. American journal of ophthalmology, 152(4):515-522, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[78] Eija Vesti, Chris A Johnson, and Balwantray C Chauhan. Comparison of different methods for detecting glaucomatous visual field progression. Investigative ophthalmology & visual science, 44(9):3873-3879, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[79] Xi Wang, Hao Chen, An-Ran Ran, Luyang Luo, Poemen P Chan, Clement C Tham, Robert T Chang, Suria S Mannil, Carol Y Cheung, and Pheng-Ann Heng. Towards multicenter glaucoma oct image screening with semi-supervised joint structure and function multi-task learning. Medical Image Analysis, 63:101695, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[80] Junde Wu, Huihui Fang, Fei Li, Huazhu Fu, Fengbin Lin, Jiongcheng Li, Lexing Huang, Qinji Yu, Sifan Song, Xingxing Xu, et al. Gamma challenge: glaucoma grading from multi-modality images. arXiv preprint arXiv:2202.06511, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[81] Jo-Hsuan Wu, Takashi Nishida, Robert N Weinreb, and Jou-Wei Lin. Performances of machine learning in detecting" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "glaucoma using fundus and retinal optical coherence tomography images: A meta-analysis. American Journal of Ophthalmology, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.202 + ], + "angle": 0, + "content": "[82] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.204, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[83] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.313 + ], + "angle": 0, + "content": "[84] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual international conference of the IEEE engineering in medicine and biology, pages 3065-3068. IEEE, 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20482" + } + ] +] \ No newline at end of file diff --git a/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_origin.pdf b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2dd80b331f8669f69a5261ede1620baf58b03f28 --- /dev/null +++ b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/ed5ae2e0-94e5-433b-855f-28b687bebad2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad5c5130cdc2f37b7abf31d42565df7dbb8a216e82748b7a8828eef5f0e28f98 +size 1020389 diff --git a/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/full.md b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..380ee5d26a78c42f5c33246d8061584f9565be21 --- /dev/null +++ b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/full.md @@ -0,0 +1,320 @@ +# Harvard Glaucoma Detection and Progression: A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning + +Yan Luo* Min Shi* Yu Tian* Tobias Elze Mengyu Wang +Harvard Ophthalmology AI Lab, Harvard University + +{yluol6, mshi6, ytian11, tobias_elze, mengyu_wang}@meei.harvard.edu + +# Abstract + +Glaucoma is the number one cause of irreversible blindness globally. A major challenge for accurate glaucoma detection and progression forecasting is the bottleneck of limited labeled patients with the state-of-the-art (SOTA) 3D retinal imaging data of optical coherence tomography (OCT). To address the data scarcity issue, this paper proposes two solutions. First, we develop a novel generalization-reinforced semi-supervised learning (SSL) model called pseudo supervisor to optimally utilize unlabeled data. Compared with SOTA models, the proposed pseudo supervisor optimizes the policy of predicting pseudo labels with unlabeled samples to improve empirical generalization. Our pseudo supervisor model is evaluated with two clinical tasks consisting of glaucoma detection and progression forecasting. The progression forecasting task is evaluated both unimodally and multimodally. Our pseudo supervisor model demonstrates superior performance than SOTA SSL comparison models. Moreover, our model also achieves the best results on the publicly available LAG fundus dataset. Second, we introduce the Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, a multimodal multitask dataset that includes data from 1,000 patients with OCT imaging data, as well as labels for glaucoma detection and progression. This is the largest glaucoma detection dataset with 3D OCT imaging data and the first glaucoma progression forecasting dataset that is publicly available. Detailed sex and racial analysis are provided, which can be used by interested researchers for fairness learning studies. Our released dataset is benchmarked with several SOTA supervised CNN and transformer deep learning models. The dataset and code are made publicly available via https://ophai.hms.harvard.edu/datasets/harvard-gdp1000. + +# 1. Introduction + +Glaucoma is the leading cause of irreversible blindness globally caused by retinal nerve fiber layer (RNFL) + +![](images/4e308eae7e47e4262e33c791ccaca8e2fa20cf6ab816fa91ebfa000f3e528c78.jpg) +Figure 1: Examples of RNFLT maps with high similarities but from different label groups. + +damage[74, 58, 59, 69]. The global prevalence of glaucoma for the population between 40 and 80 years old is $3.54\%$ [74]. Timely clinical treatment in the early stage of glaucoma can significantly reduce the risk of further vision loss [18, 17]. However, most commonly, glaucoma patients are not aware of their early VF loss until the VF loss becomes severe enough to impair their daily activities such as reading and driving due to the brain and fellow eye compensation[4, 19, 34, 50, 51, 30]. + +It is desirable to develop automated glaucoma screen tools to reduce societal disease burden [77]. The clinical diagnosis of glaucoma is a holistic decision primarily based on RNFL damage and VF loss assessment supplemented with other patient clinical information and family disease history [29, 47]. The VF test is only available in the ophthalmology specialty. So existing automated screening tools are based on the idea of predicting glaucoma diagnosis from evaluating RNFL damage using retinal images [14, 57, 42, 5, 49, 37, 15, 35, 60, 81]. Before the massive clinical adoption of the 3D retinal imaging tool of optical coherence tomography (OCT) a decade ago [31, 12], 2D fundus photos were used to assess RNFL damage. Therefore, most prior glaucoma detection models are based on the 2D fundus photo [37, 57, 33, 14, 42, 1, 64, 15], which is not the state of the art (SOTA) clinical imaging modality anymore. OCT has replaced fundus photos as the primary imaging tool for glaucoma clinical care due to its superiority in measuring RNFL damage [48, 12]. There have been a number of recent works using OCT scans to automatically screen glaucoma with machine learning [15, 60, 5, 35, 24, 80, 49]. Apart from glaucoma detection, a more clinically significant task is progression forecasting using OCT scans [52, 16, 32, 28], which could inform clini + +cians if they should treat the patient conservatively with eye drops or aggressively with invasive surgeries. + +For both the glaucoma screening and progression forecasting tasks, a significant challenge is limited labeled patients with the SOTA OCT measurements as the fundus photo is already outdated in clinical practice [15, 60, 5, 35, 24, 80, 49]. This problem with the lack of labeled patients is even more severe for progression forecasting, which requires longitudinal VF measurements to determine progression outcomes [52, 16, 32]. Therefore, there is a strong need for developing novel semi-supervised learning models to maximally utilize unlabeled patient data in clinics to improve glaucoma detection and progression forecasting. However, apart from several papers using standard semi-supervised learning methods for glaucoma detection [20, 2, 23, 79, 84, 9], there has been no dedicated algorithm development for semi-supervised learning with its application in glaucoma. SOTA semi-supervised learning (SSL) approaches in computer vision are generally based on consistency learning [11, 68, 10, 73] or pseudo-labeling [40, 61]. Unlike consistency learning, the pseudo-labeling approaches do not need domain knowledge to design the image augmentation strategy. While many prior works have endeavored to optimally select pseudo-labeled samples to improve the classifier training, the tasks of glaucoma detection and progression forecasting still have some unique challenges that might not be well tackled by existing pseudo-labeling SSL methods [61, 63, 13, 40, 56]. Figure 1 shows examples of RNFLT maps with high similarities but from different label groups. With such obscured group differences, existing pseudo-labeling SSL methods might not be able to provide sufficient guidance to select correct pseudo-labeled samples to improve the classifier training. Furthermore, having high-quality public datasets with OCT scans for glaucoma detection and progression forecasting is equally important to enable more computer vision researchers to study this topic to push forward the field, + +In this study, we make two contributions: (1) developing a generalization-reinforced semi-supervised learning model called pseudo supervisor to improve glaucoma detection and progression forecasting. Specifically, our pseudo supervisor model will select pseudo-labeled samples via a policy net that optimizes generalization error with a validation subset to update the classifier. (2) releasing a pilot multitask and multimodal glaucoma dataset, called Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, for computer vision researchers to study this topic. In Table1, we have listed major public glaucoma datasets. Most of them are fundus photo datasets, while there are two major OCT datasets previously published [24, 80]. In comparison, our dataset has the largest patient numbers among all datasets with patient numbers available. In addition, our dataset is the first and the only one with progression forecasting task data, which is a way more clinically significant + +Table 1: Public Glaucoma Datasets. + +
StudyImaging ModalitySample SizeGlaucoma DetectionProgression ForecastingMultimodalLabel SourceAccessibility
LAG [38]FundusNA (4,854)✓✘✘Clinician AssessmentUpon Request
REFuge [54]FundusNA (1,200)✓✘✘Clinician AssessmentWeb Download
G1020 [6]Fundus432 (1,020)✓✘✘Clinician AssessmentWeb Download
ACRIMA [21]FundusNA (705)✓✘✘Clinician AssessmentWeb Download
RIM-OONE [7]Fundus169 (485)✓✘✘Clinician AssessmentWeb Download
ORIGA [84]FundusNA (650)✓✘✘Clinician AssessmentUpon Request
PAPILA [36]Fundus244 (488)✓✘✘Clinician AssessmentWeb Download
GOALS [24]OCT158 (300)✓✘✘Clinician AssessmentWeb Download
GAMMA [80]OCT300 (300)✓✘✓Clinician AssessmentWeb Download
OursOCT1,000 (1,000)✓✓✓Functional TestWeb Download
+ +task than glaucoma detection itself. Compared with the other two OCT datasets, our dataset is superior in patient numbers (1,000 versus up to 300), task versatility (glaucoma detection and progression forecasting versus glaucoma detection only), and availability of objective visual function tests (i.e. VF test). Furthermore, since our dataset is from the US population, the racial representation in our data is more diverse than the two OCT datasets [24, 80] and the large LAG dataset [38], which consists of mainly Asian patients. Therefore, our dataset can be potentially used for medical fairness learning studies, especially since we know Black people have more than doubled glaucoma prevalence than other races, which is a significant disparity [62, 25]. + +# 2. Related Work + +Glaucoma Detection. Vast majority of glaucoma detection works using deep learning are based on the fundus photo, which is a 2D snapshot of the patient's retina [37, 57, 33, 14, 42, 1, 64, 15, 76, 75]. However, the 3D imaging modality OCT scan has replaced fundus photos to become the de facto standard for structural damage assessment in glaucoma. Compared with the deep learning glaucoma detection using fundus photos, deep learning glaucoma detection works using OCT scans generally have much smaller sample sizes, which are up to several thousand [15, 60, 5, 35, 24, 80, 49]. As deep learning models are typically data-hungry, there is a clear need to develop new approaches to deal with the limited labeled data issue. It is known that semi-supervised learning can improve prediction accuracy by leveraging unlabeled data in model training. So far, apart from several glaucoma detection papers using standard semi-supervised learning models [20, 2, 23, 79, 84, 9], no novel semi-supervised learning model has been developed with an application focus on glaucoma. + +Progression Forecasting. Compared with the glaucoma detection task, progression forecasting is a much more clinically significant task for patients. If we could have an accurate progression forecasting model to inform clinicians of which patient will progress or not, then the clinicians can accordingly decide if a patient should be treated more aggressively with invasive surgeries, which may have a strong side effect, or more conservatively with eye drops. So far, due to the data scarcity issue, we have only found several recent papers using retinal imaging data (whether fundus + +or OCT) to forecast progression [52, 16, 32]. The sample sizes of these works are only up to several hundred. Imaginably, we have not found any prior studies applying semisupervised learning for progression forecasting. + +Semi-Supervised Learning. Semi-Supervised Learning (SSL) aims to optimize a model with both labeled and unlabeled data. The current SSL methods can be divided into pseudo-labeling-based methods [61, 63, 13, 40] and/or consistency-based methods [73, 68, 11, 10, 41, 43]. Consistency-based SSL methods optimize the model with standard classification losses (e.g., cross-entropy) on labeled images and regularizes/minimizes the prediction outputs of different weakly and/or strongly augmented views of unlabeled images, where these views are constructed using different types of image augmentations, such as flip-and-shift, color sharpness, and contrast, etc. Even though consistency-based SSL methods achieve SOTA results in many computer vision benchmarks [73, 68, 11, 10], these methods significantly rely on the orderly design of augmentation functions that requires domain knowledge to design proper image augmentation strategies, which are even more challenging for multimodal modeling. Prior pseudolabeling SSL methods [61, 63, 13, 40, 56] typically utilize a model that trained with the labeled subset to estimate the pseudo labels of the confident unlabeled samples and take these pseudo-labeled data to re-train the model. As mentioned previously, pseudo-labeling-based SSL approaches is a general SSL learning framework for different data modalities naturally. More importantly, as mentioned in Section 1, since the structural differences between glaucoma and non-glaucoma (similarly for progression versus non-progression) are obscured as shown in Figure 1, existing pseudo-labeling SSL methods [61, 63, 13, 40, 56] might not be able to provide sufficient guidance to select correct pseudo labeled samples to benefit the classifier training. In this aspect, we propose a generalization-reinforced SSL model by selecting optimal pseudo-labeled samples for classifier training that explicitly optimize generalization error with a policy net stochastically generates pseudo labels. + +Reinforcement Learning. Reinforcement learning studies an agent interacts with the environment and learns an optimal policy to tackle decision-making problems effectively [70, 8, 26, 65, 66]. Reinforcement learning has been applied to semi-supervised learning setting for real-world applications, e.g. domain adaptation [39]. However, how to utilize reinforcement learning techniques to determine pseudo labels for improving the generalization of classification models remains under exploration. + +# 3. Multimodal and Multitask Dataset + +# 3.1. Data Collection and Quality Control + +This work was approved by the institutional review board (IRB) and adhered to the tenets of the Declaration of + +![](images/f630c8cb3ff61917878cbf5a43e9c6a22a285bb9fd9b90f9f0a07e40c2fb4d88.jpg) +Figure 2: Pointwise RNFLT distribution at different percentiles. + +Helsinki. Because of the retrospective nature of the study, the IRB waived the need for informed consent of patients. The glaucoma patients are from Mass Eye and Ear of Harvard Medical School, which were tested between 2010 and 2021. The RNFLT $(0 - 350$ microns) is from the OCT report provided by the OCT device manufacturer (Cirrus, Carl Zeiss Meditec, Jena, Germany) calculated as the vertical distance between the internal limiting membrane boundary and nerve fiber layer boundary. The RNFLT measurement is the de facto structural measurement for clinicians to diagnose glaucoma. Each two-dimensional $224\times 224$ RNFLT map is generated within the $6\times 6\mathrm{mm}^2$ peripapillary area around the optic nerve head. The VF testing is performed using the Humphrey Field Analyzer with the 24-2 protocol, which measures the patient's eye's visual spatial sensitivity with a radius of 24 degrees from the fixation. Data quality control based on standard clinical practice guideline following the manufacturer's recommendation was applied to exclude unreliable data: (1) OCT scans with signal strength less than 6 (10 represents the best imaging quality) were excluded; (2) VF tests with fixation losses $>33\%$ , false positive rates $>20\%$ or false negative rates $>20\%$ were excluded. Each VF measurement is represented as a vector of 52 total deviation (TD) values between -38 dB and 26 dB. In this work, we refer RNFLT map and VF test as two data modalities which constitute a multimodal dataset. Note that, the modifier word "Harvard" from the Harvard GDP Dataset only indicates that our dataset is from the Department of Ophthalmology of Harvard Medical School and does not imply an endorsement, sponsorship, or assumption of responsibility by either Harvard University or Harvard Medical School as a legal identity. + +# 3.2. Data Analysis + +The glaucoma detection dataset contains 15,725 training samples, 1,047 validation samples, and 4,287 samples from 5,640, 376, and 1,505 glaucoma patients, respectively. Half of the samples are unlabeled. Among the half samples with labels, $45.2\%$ of the samples are glaucoma, while the rest are non-glaucoma. The progression forecasting dataset includes 5,889 training samples, 431 validation samples, and 1,547 test samples from 1,292, 86, and 345 glaucoma patients, respectively. Among the half samples with progression outcome labels, $24.0\%$ and $2.6\%$ of the samples are TD Progression and MD Fast Progression, respectively. The + +two clinically accepted criteria to define glaucoma progression: (1) TD Progression: eyes with at least three locations with TD slope $\leq -1$ dB; (2) MD Fast Progression: eyes with MD slope $\leq -1$ dB [78]. The MD Fast Progression is more detrimental for the patients than TD progression [3]. + +Figure 2 presents the pointwise distributions of 25th, 50th, and 75th percentile RNFLTs in the train (validation set included) and test sets, which demonstrate two obvious superotemporal and inferotemporal RNFL bundles with their damages are known closely related to the visual function loss. As in the two representative examples shown in Figure 1, RNFLT maps with similar appearances may have different labels. This suggests the class boundary may be obscure to determine, which naturally prompts us to adopt semi-supervised solutions to address the glaucoma detection and progression forecasting problems which fully leverage the abundant information in unlabeled samples to mitigate the challenges. + +The released dataset including 1,000 patients is a random subset from the entire dataset used to build the glaucoma detection and progression forecasting models in this paper, it naturally inherits the original data characteristics. The released glaucoma detection dataset contains 1,000 samples from 1,000 patients, where each sample includes two data modalities (i.e. RNFLT map and VF test) and respective glaucoma labels. The progression forecasting dataset is a subset from the released glaucoma detection dataset, which entails 500 patients with progression labels and patient demographics including age, sex, and race. For all glaucoma patients, the age distribution of 25th, 50th, and 75th percentiles are 54.3, 64.5, and 72.8 years, respectively. The male and female account for $56\%$ and $44\%$ of the 1,000 patients, respectively. Among the 1,000 patients, the White, Black, and Asian groups account for $76.4\%$ , $14.9\%$ , and $8.7\%$ , respectively. To the best of our knowledge, this is the first public glaucoma dataset with the patient demographic information available, which carries great potential to enable interested researchers to conduct and evaluate fairness learning studies. + +# 4. Problem Formalization & Background + +In this section, we formally introduce the problems of glaucoma detection and progression forecasting. + +Denote $\mathcal{D}^l = \{(x_i,y_i)|1\leq i\leq N^l\}$ be a labeled dataset, where $x\in \mathbb{R}^n$ are the RNFLT maps and $y\in \mathcal{V}$ are corresponding glaucoma/progression labels. Also, we have a stack of unlabeled RNFLT maps, i.e. $\mathcal{D}^{ul} = \{x_j|1\leq j\leq N^{ul}\}$ . We assume that we have a classification model $f:\mathbb{R}^n\xrightarrow{\theta}\mathcal{Y}(f_\theta$ for short), where $\theta$ are the parameters. To update the classification model with the labeled data, a predefined loss function $\ell :\mathcal{V}\to \mathcal{V}$ is used to gauge the discrepancy between the predictions and the ground-truth labels such that the classification model can be updated along + +![](images/fa4f4a7220825318e879c4d2882163076c01f91d7168ea2785a488315c3e180a.jpg) + +![](images/4b929fc2225ec1040d39a3e24df785e1d0c60ddb79fa1fdd5cd09d277483bfcc.jpg) +Figure 3: The definitions of (a) glaucoma detection task; and (b) progression forecasting task. In the VF test, blue indicates normal visual function while red indicates abnormal visual function. + +the loss, i.e. $\min_{\theta} \ell(f_{\theta}(x), y)$ . + +As shown in Figure 3 (a), the practical utility scenario of glaucoma detection with deep learning is to use OCT scans that can be easily measured in the setting of primary care or pharmacy to predict glaucoma diagnosis defined by visual field tests that typically only can be measured in the setting of an ophthalmology clinic or eye hospital with specialized technicians. The clinical relevance of progression forecasting as shown in Figure 3 (b) with deep learning is that clinicians can better decide if a patient should be treated more conservatively with eye drops or aggressively with invasive surgeries. As suggested by conventional semi-supervised learning methods [20, 2, 40, 68, 11], pseudo labels $\tilde{y}$ are predicted by a supervisor model $\pi : \mathbf{R}^n \xrightarrow{\omega} \mathcal{Y}$ with the unlabeled data $\tilde{x}$ , i.e. $\tilde{y} = \pi(\tilde{x};\omega)$ . Similar to the supervised learning setting, the unlabeled data with the pseudo labels can participate in the supervised learning process, namely $\min_{\theta} \ell(f_{\theta}(\tilde{x}), \tilde{y})$ . + +However, as shown in Figure 4, the natural images associated with different visual concepts are relatively easy to be distinguished. In contrast, the differences in RNFLT maps between glaucoma versus non-glaucoma (similarly for progression versus non-progression) can be very subtle due to large retinal physiological inter-subject variations. There might not be clear and adequate guidance in the learning process to determine the underlying labels of the unlabeled RNFLT maps. Instead, a more sensible way to better determine the pseudo labels is to apply a policy to obtain pseudo labels. Then, we observe how these pseudo labels affect + +![](images/a764494e5de5f11426900ba55643901b9d7e776b04d85bec0e36ba572e445deb.jpg) +Figure 4: A challenge in glaucoma detection and progression forecasting. Unlike the classification with natural images, where the images with different labels are likely visually distinguishable, RNFLT images with different labels (e.g. glaucoma vs. non-glaucoma and progression vs nonprogression) might not be easily distinguishable. + +the training process of the classifier and adjust the policy to predict the pseudo labels such that the generalizability of the classifier learned with the pseudo labels can be maximized. In short, the pseudo labels predicted by the supervisor model should be aligned with the improvement of the generalizability of the classifier on a set of unseen data $\mathcal{D}^{unseen}$ . This objective can be formulated as + +$$ +\min _ {\{\tilde {y} \}} \frac {1}{| \mathcal {D} ^ {\text {u n s e e n}} |} \sum_ {(x, y) \in \mathcal {D} ^ {\text {u n s e e n}}} \ell \left(f _ {\theta^ {*}} (x), y\right) \tag {1} +$$ + +where $\theta^{*}\in \arg \min_{\theta}\frac{1}{|\mathcal{D}^{l + ul}|}\sum_{(x,y)\in \mathcal{D}^{l + ul}}\ell (f_{\theta}(x),y)$ + +$\mathcal{D}^{l + ul}$ is the union of labeled dataset $\mathcal{D}$ and unlabeled dataset with pseudo labels $\mathcal{D}^{ul}$ . + +# 5. Generalization-Reinforced Pseudo Supervisor + +Different from prior-work semi-supervised learning methods [68, 11, 10, 53], the proposed method explicitly leverages the empirical generalizability of the classifier that is quantified with unseen RNFLT maps (e.g. the ones from the validation set) to optimize the policy of predicting the pseudo labels. The empirical generalizability can be viewed as a type of informative reward as the goal of training a classifier is to find one that can generalize well on the unseen test data drawn from the underlying distribution. + +Figure 5 shows how the proposed generalization-reinforced pseudo supervisor predicts pseudo labels and is trained according to the observed sets of states (maps), actions (pseudo labels), and rewards (the empirical generalizability with a mini-batch of validation samples). The proposed pseudo supervisor is based on the policy gradient [70], which is one type of generic reinforcement learning methods [8, 26, 65]. + +Here we formally introduce the concepts related to the proposed pseudo supervisor. An RNFLT map $x_{t}$ at time step $t$ can be viewed as a state $s_t \in S$ , while the predicted pseudo label $\tilde{y}_t$ can be viewed as an action $a_{t} \in \mathcal{A}$ . A reward function $r_t: S \times \mathcal{A} \to \mathbf{R}$ is defined on the state-action space. The proposed pseudo supervisor that employs + +a policy to determine pseudo labels is denoted as $\pi_{\omega} : S \to \mathcal{P}(\mathcal{A})$ , where $\mathcal{P}(\mathcal{A})$ are the probability measures on $\mathcal{A}$ and $\omega$ are the parameters of the proposed pseudo supervisor. + +In the first place, the classifier takes labeled samples for training. The unlabeled samples will be fed into the pseudo supervisor for predicting pseudo labels. Before updating the weights of the classifier, the empirical generalizability of classifier $f_{\theta_t}$ can be quantified with validation samples $(x^{(val)}, y^{(val)})$ , that is, $\ell(f_{\theta_t}(x^{(val)}), y^{(val)}) (\ell_t^{(val)})$ . Specifically, $\ell$ is the widely-used cross-entropy loss for classification. After updating the weights of the classifier with unlabeled samples and corresponding pseudo labels, given the same mini-batch of validation samples, the empirical generalizability of the updated classifier can be requantified again, i.e., $\ell(f_{\theta_{t+1}}(x^{(val)}), y^{(val)}) (\ell_{t+1}^{(val)})$ . Then, the discrepancy between $\ell_t^{(val)}$ and $\ell_{t+1}^{(val)}$ is informative to measure the effect of updating the classifier with the unlabeled samples and the pseudo labels. In this sense, it is desired to use in the reward function, i.e. + +$$ +r _ {t} = \max \left(\exp \left(\ell_ {t} ^ {(v a l)} - \ell_ {t + 1} ^ {(v a l)}\right) - 1, 0\right) \tag {2} +$$ + +In Eq. (2), if $\ell_{t+1}^{(val)}$ is far less than $\ell_t^{(val)}$ , it implies that the unlabeled samples and pseudo labels lead to a positive effect on the learning process such that the reward increases adaptively. + +A trajectory of states, actions, and rewards within a time window $\beta$ will be saved. The expected return of policy optimization can be written as + +$$ +J \left(\pi_ {\omega}\right) = \mathbb {E} _ {\pi_ {\omega}} \left(\sum_ {k = 0} \gamma^ {k} r _ {k}\right) \tag {3} +$$ + +where $\gamma \in [0,1]$ is the discount rate for the numerical stability of reinforcement learning, and $\beta$ is the time window for updating $\pi_{\omega}$ periodically. According to the policy gradient theorem [70], the expectation of the sample gradient is equal to the actual gradient + +$$ +\nabla_ {\pi_ {\omega}} J (\pi_ {\omega}) = \mathbb {E} _ {\pi_ {\omega}} \left[ \sum_ {k = 0} ^ {\beta} \gamma^ {k} r _ {t + k + 1} \nabla_ {\pi_ {\omega}} \log \pi_ {\omega} \left(a _ {t} \mid s _ {t}\right) \right] \tag {4} +$$ + +Then the policy parameterization can be updated according to the gradient ascend rule + +$$ +\pi_ {\omega} \leftarrow \pi_ {\omega} + \eta_ {\pi} \nabla_ {\pi_ {\omega}} J (\pi_ {\omega}) \tag {5} +$$ + +where $\eta_{\pi}$ is the learning rate of $\pi_{\omega}$ . + +# 6. Experiments + +# 6.1. Set-Up & Implementation Detail + +To optimize our models, we employ the AdamW optimizer [46] and train all models for 10 epochs in all experiments conducted in this study. For supervised classification + +![](images/e46acf5d198aa2915093fb0c4e8c51c068cd6b4bf6b7d6b992d10084bcd216df.jpg) +Figure 5: An overview of the proposed generalization-reinforced pseudo supervisor in the semi-supervised learning framework. The learning framework consists of the process of training the classifier, the process of measuring the empirical generalizability of the classifier to be used as rewards, and the process of updating the proposed pseudo supervisor to predict pseudo labels that can maximize the classifier's empirical generalizability. + +models for glaucoma detection, we use a learning rate of $4 \times 10^{-5}$ and weight decay of 0, and for progression forecasting, we use a learning rate of $2 \times 10^{-5}$ and weight decay of 0. These hyperparameters are determined to achieve the best baseline performance based on our experiments. The proposed method is trained using a learning rate of $4 \times 10^{-5}$ and weight decay of 0. The batch sizes used in the supervised model, semi-supervised models, and the proposed method vary due to the different semi-supervised learning contexts [73, 11, 10, 68, 53]. We adopt the batch sizes that result in the best performance for each respective context. To ensure consistency with previous work, we apply data augmentation techniques to the samples in the semi-supervised learning models, including MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53]. Specifically, we utilize random crop and resize, as well as random horizontal flip. However, we do not apply these data augmentation techniques to the supervised baseline or the proposed method. + +For the glaucoma detection task, we train all models with 7,860 labeled samples and 7,865 unlabeled samples for training. We use 1,047 and 4,287 labeled samples for validation and testing, respectively. For the progression forecasting task, all the models are trained with 2,944 labeled samples and 2,945 unlabeled samples for training. We use 431 and 1,547 labeled samples for validation and testing, respectively. For multi-modality modeling, we simply concatenate the RNFLT images and VF matrices for joint learning using the same architecture as our single-modality model and apply this model to the progression forecasting task. The model is trained and evaluated using the same train-val-test split as our single-modality experiments. + +We use the accuracy, F1 score, and AUC on the test sets as our evaluation matrices. The code is written in PyTorch [55] and we use one RTX A6000 GPU for all experiments. + +# 6.2. Baseline Methods + +Our Supervised baseline is a standard ImageNet pretrained EfficientNet classifier [72] learning with cross + +entropy using only labeled images. For Semi-Supervised baselines, we choose MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53] due to their previous powerful performances. MeanTeacher [73] is one of the initial methods to enforce models to predict consistent pseudo predictions for labeled and unlabeled data, which construct a target-generating teacher model by averaging the model weights. MixMatch [11], ReMixMatch [10], and FixMatch [68] are consistency-based SSL methods that are based on the data augmentations (i.e., weak and strong augmentation) to make consistent classification predictions between different views of unlabeled data. DASO [53] suggests that those methods fail for imbalanced datasets and proposes a pseudo-labeling framework based on different consistency SSL models by adaptively mixing the linear and semantic pseudo-labels to mitigate the overall bias between majority and minority classes. We adopt the FixMatch version of DASO, following the default setting reported in the paper [53]. We implement all baseline methods using Pytorch [55] from DASO's official codes and conducted experiments under the same codebase and experimental protocols for fair comparison. All models are equipped with the same EfficientNetV2-S [72] backbone and their hyper-parameters are specifically tuned to suit our Glaucoma detection and progression forecasting tasks. + +# 6.3. Results for Glaucoma Detection + +In Table 2, we evaluate our Pseudo-Sup model against multiple SOTA baselines on the glaucoma detection task. Our model achieves the best performance in terms of accuracy, F1 score, and AUC score. We compute the mean performances over 5 different runs. Notably, our model achieves such results without any specific design of image perturbations/augmentations, guaranteeing its simplicity and applicability to different tasks. Our method surpasses the previous SOTA method DASO by $6.8\%$ of accuracy, $11.1\%$ of F1 score, and $3.7\%$ of AUC, respectively. The best-performing consistency method FixMatch still worse by $0.4\%$ of accuracy, $1.2\%$ of F1 score, and $1.5\%$ + +Table 2: Performance on the cross-sectional data for the glaucoma detection task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method. + +
ModelAcc↑F1↑AUC↑
Baseline0.8131 ± 0.00390.8057 ± 0.00330.8693 ± 0.0022
MeanTeacher0.8148 ± 0.00530.8053 ± 0.00720.8541 ± 0.0040
MixMatch0.6392 ± 0.01610.6273 ± 0.00590.8451 ± 0.0062
ReMixMatch0.6741 ± 0.00190.6716 ± 0.00180.8442 ± 0.0032
DASO0.7444 ± 0.01150.6929 ± 0.00500.8371 ± 0.0118
FixMatch0.8077 ± 0.00320.7912 ± 0.00210.8574 ± 0.0019
Pseudo-Sup0.8124 ± 0.00250.8038 ± 0.00460.8727 ± 0.0006
Pseudo-Sup+Aug0.8224 ± 0.01010.8128 ± 0.01220.8818 ± 0.0020
+ +of AUC than our Pseudo-Sup model, indicating the effectiveness of our approach and the challenges of designing appropriate augmentations for RNFLT maps from previous approaches. We also apply weak augmentations (similar to FixMatch) to our Pseudo-Sup (Pseudo-Sup+Aug) leading to improved performances. Please note that consistency methods (e.g., DASO, FixMatch, etc.) employ both weak and strong augmentations. + +# 6.4. Results for Progression Forecasting + +Unimodal Results. We further evaluate our model on progression forecasting, a more clinically significant task than glaucoma detection, as shown in Table 3. Accurate forecasting of glaucoma progression plays a vital role in prescribing appropriate treatment plans to efficiently prevent further vision losses and is important for assisting the development of new pharmaceutical medicines to treat glaucoma. Adding weak augmentations further improves the results for both progression tasks. Please note that similar to glaucoma detection experiments, we also compute the mean performances over 5 different runs. + +For TD Progression, our model achieves the best $74.47\%$ of mean AUC, respectively. When compared with consistency-based approaches, our method improves between $0.5\% - 2.1\%$ in AUC. For MD Fast Progression, our approach consistently achieves the best AUC performance, illustrating the effectiveness and accuracy of our method regardless of any glaucoma progression criteria. Given that most of the glaucoma patients are non-progressive, our progression dataset for TD Progression and MD Fast Progression are both highly imbalanced (i.e., most of the samples are labeled as non-progression). Therefore, it is worth noting that our method also outperforms the previous imbalanced SSL SOTA DASO by a large margin in all measures under such imbalanced scenarios. + +Multimodal Results. One of the major contributions of this work is to introduce a multi-modal dataset for glaucoma progression forecasting. To the best of our knowledge, this is the first work to explore and benchmark SSL approaches using multimodal model inputs. In Table 5, we show the + +Table 3: Performance on the longitudinal data with single modality RNFLT for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while Pseudo-Sup represents the proposed pseudo supervisor method. + +
ModelsAUC ↑ (TD)AUC ↑ (MD Fast)
Baseline0.7367 ± 0.01260.6677 ± 0.0017
MeanTeacher0.7367 ± 0.01260.6677 ± 0.0017
MixMatch0.7238 ± 0.00980.6518 ± 0.0069
ReMixMatch0.7390 ± 0.01240.6595 ± 0.0083
DASO0.7231 ± 0.01880.6787 ± 0.0148
FixMatch0.7395 ±0.01520.6747 ± 0.0054
Pseudo-Sup0.7447 ± 0.00840.7004 ± 0.0218
Pseudo-Sup+Aug0.7477 ± 0.00740.7215 ± 0.0162
+ +progression forecasting results using two modalities (RNFLT and VF). Our model achieves the best performance on all evaluation measures. To keep its simplicity and adaptability, we concatenate the RNFLT and the up-scaled VF together as the inputs of our model. Please note that consistency SSL approaches that are based on image augmentations are not included in this table because of the lack of design of augmentation techniques in multimodal medical data. The SOTA results with multimodal glaucoma data enable more accurate and applicable potential computer-aided diagnosis (CAD) systems in real clinical settings. + +# 6.5. Supervised Benchmarks on Released Data + +In Table 6, we show the supervised classification results for the progression forecasting tasks with multiple SOTA supervised CNN and transformer baseline methods, including VGG [67], ResNet [27], ResNext [82], WideResNet [83], EfficientNet [71], ConvNext [45], ViT [22], and Swin Transformers [44]. This benchmark is conducted on our future data release with 500 patients upon acceptance, of which 400 patients are used for training and the remaining are used for testing. To the best of our knowledge, this is the first supervised glaucoma progression forecasting benchmark, which aims to provide broader impacts for the computer vision and medical imaging communities to build clinically effective computer-aided diagnosis systems. Moreover, the results are reported with both single RNFLT modality and multi-modality (RNFLT and VF) samples. It is worth noting that the accuracy and F1 are the same on MD fast progression because the testing set only contains 100 samples under highly imbalanced class distributions. For more experimental details regarding this supervised progression forecasting benchmark and the supervised results on future release data with 1,000 patients for the glaucoma detection task, please refer to the Supplemental Material. + +Table 4: Performance comparisons for glaucoma detection task on LAG dataset. + +
ModelAcc↑F1↑AUC↑
MeanTeacher0.82800.82790.9082
MixMatch0.50200.33780.9075
ReMixMatch0.69800.47850.8369
DASO0.79710.79060.9164
FixMatch0.83660.82970.9198
Pseudo-Sup0.91600.91580.9858
+ +Table 5: Performance on the longitudinal data with two modalities (RNFLT and VF) for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method. Please note that most methods in Table 3 are based on weak/strong augmentations [68, 11, 10] requiring RGB channels, which are not adaptable for multi-modal learning. + +
ModelsAUC (TD)AUC (MD Fast)
Baseline0.7343 ± 0.00980.6814 ± 0.0081
MeanTeacher0.7436 ± 0.01330.6697 ± 0.0035
Pseudo-Sup0.7517 ± 0.00410.7265 ± 0.0095
+ +# 6.6. Performances on LAG + +In Table 4, we show the performance comparisons on LAG dataset [37]. LAG is a large-scale fundus imaging dataset for glaucoma detection. For this experiment, we use 3,854 images for training with half with labels and the remaining without labels, and 1,000 images for testing, with the size of $224 \times 224$ . For fair comparisons, we use the same hyper-parameters and experimental setups as our other benchmarks. It is worth noting that our approach surpasses the SOTA methods by a large margin and achieves the highest $98.58\%$ AUC with only half-labeled data, illustrating the effectiveness and applicability of our proposed Pseudo-Sup in different data and tasks. + +# 6.7. Ablation Study + +According to Section 5, the time window $\beta$ and discount rate $\gamma$ are the most important hyperparameters affecting the proposed pseudo supervisor. The ablation studies of the two hyperparameters are shown in Figure 6. We can see that the proposed pseudo supervisor behaviors stably with various $\beta$ on AUC. Specifically, when $\beta = 50$ leads to better performance. The consistent pattern can be also observed in the analysis of $\gamma$ . The discount factor is used to achieve a trade-off between the rewards in the distant future and the rewards in the near future. If it is 0, reinforcement learning agents would only look for an optimal policy for the first action. As shown in Figure 6 (b), we find the optimal discount rate at 0.9, which implies that the states, actions, and rewards in the progression task are coherent to a certain degree. + +Table 6: Performance of different supervised methods on the longitudinal data with single/multi-modality for the progression forecasting tasks on the released 500 glaucoma progression data. + +
ModelTD ProgressionMD Fast Progression
Acc†F1†AUC†Acc†F1†AUC†
VGG [67]0.78/0.800.75/0.740.84/0.790.98/0.980.49/0.490.78/0.68
ResNet [27]0.75/0.740.64/0.670.74/0.750.98/0.980.49/0.490.71/0.59
ResNext [82]0.78/0.770.65/0.680.78/0.760.98/0.980.49/0.490.61/0.80
WideResNet [83]0.77/0.790.71/0.750.77/0.800.98/0.980.49/0.490.71/0.64
EfficientNet [71]0.73/0.780.66/0.730.76/0.790.98/0.980.49/0.490.55/0.66
ConvNext [45]0.74/0.810.62/0.770.78/0.810.98/0.980.49/0.490.84/0.62
ViT [22]0.73/0.770.54/0.710.68/0.790.98/0.980.49/0.490.92/0.69
Swin [44]0.71/0.770.49/0.720.68/0.770.98/0.980.49/0.490.87/0.57
+ +![](images/7245700de1e5b5558aa9f3bcac2a291de22a468708ba495b32fa9bf3a96546ba.jpg) +(a) The impact of $\beta$ + +![](images/00da3174ba9ddc75138a99fdd156916a6f7ed2abaacb5cb7cba74498e6f74eab.jpg) +(b) The impact of $\gamma$ +Figure 6: Ablation study of hyperparameters' impact on TD Progression forecasting. + +# 6.8. Experiment on Racial Groups + +The patients in our dataset are from three racial groups (White (0.88 AUC), Black (0.83 AUC), and Asian (0.85 AUC)). In our experiments, we observe that black patients have the lowest AUC yet the highest glaucoma prevalence [62, 25]. Therefore, our dataset is well-suited for future fairness studies. + +# 7. Conclusion + +In this paper, we have made two contributions. First, we have developed a generalization-reinforced semi-supervised learning model termed pseudo supervisor for glaucoma detection and progression forecasting. Specifically, the pseudo supervisor predicts pseudo labels following a policy that optimizes the generalization error in training the classifier. Our pseudo supervisor model overall demonstrates superior performance over a number of SOTA methods. Second, we release a multimodal and multi-task dataset for glaucoma detection and progression forecasting with SOTA 3D OCT imaging data. Our dataset is the largest by far for glaucoma detection with 3D OCT data and the first for progression forecasting task. More importantly, our dataset entails demographic information including gender and race, which is generally unavailable in existing public glaucoma datasets. Given the diverse racial representation in our dataset as detailed in Section 3, our dataset is well equipped for potential fairness learning studies, especially considering the significant glaucoma prevalence disparity between Black and other races [62, 25]. + +# Acknowledgment + +This work was supported by NIH R00 EY028631, Research to Prevent Blindness International Research Collaborators Award, Alcon Young Investigator Grant, NIH R01 EY030575, and NIH P30 EY003790. + +# References + +[1] Jin Mo Ahn, Sangsoo Kim, Kwang-Sung Ahn, Sung-Hoon Cho, Kwan Bok Lee, and Ungsoo Samuel Kim. A deep learning model for the detection of both advanced and early glaucoma using fundus photography. PloS one, 13(11):e0207982, 2018. 1, 2 +[2] Manal Al Ghamdi, Mingqi Li, Mohamed Abdel-Mottaleb, and Mohamed Abou Shousha. Semi-supervised transfer learning for convolutional neural networks for glaucoma detection. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3812-3816. IEEE, 2019. 2, 4 +[3] Andrew J Anderson, Abadh K Chaurasia, Ajay Sharma, Amisha Gupta, Shikha Gupta, Anurag Khanna, and Viney Gupta. Comparison of rates of fast and catastrophic visual field loss in three glaucoma subtypes. Investigative ophthalmology & visual science, 60(1):161-167, 2019. 4 +[4] Paul H Artes and Balwantray C Chauhan. Longitudinal changes in the visual field and optic disc in glaucoma. Progress in retinal and eye research, 24(3):333-354, 2005. 1 +[5] Ryo Asaoka, Hiroshi Murata, Kazunori Hirasawa, Yuri Fujino, Masato Matsuura, Atsuya Miki, Takashi Kanamoto, Yoko Ikeda, Kazuhiko Mori, Aiko Iwase, et al. Using deep learning and transfer learning to accurately diagnose early-onset glaucoma from macular optical coherence tomography images. American journal of ophthalmology, 198:136-145, 2019. 1, 2 +[6] Muhammad Naseer Bajwa, Gur Amrit Pal Singh, Wolfgang Neumeier, Muhammad Imran Malik, Andreas Dengel, and Sheraz Ahmed. G1020: A benchmark retinal fundus image dataset for computer-aided glaucoma detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020. 2 +[7] Francisco José Fumero Batista, Tinguaro Diaz-Aleman, Jose Sigut, Silvia Alayon, Rafael Arny, and Denisse Angel-Pereira. Rim-one dl: A unified retinal image database for assessing glaucoma using deep learning. Image Analysis & Stereology, 39(3):161-167, 2020. 2 +[8] Jonathan Baxter and Peter L Bartlett. Infinite-horizon policy-gradient estimation. Journal of Artificial Intelligence Research, 15:319-350, 2001. 3, 5 +[9] Mohammed El Amine Bechar, Nesma Settouti, Vincent Barra, and Mohamed Amine Chikh. Semi-supervised superpixel classification for medical images segmentation: application to detection of glaucoma disease. Multidimensional Systems and Signal Processing, 29(3):979-998, 2018. 2 +[10] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching + +and augmentation anchoring. In International Conference on Learning Representations, 2019. 2, 3, 5, 6, 8 +[11] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32:5049-5059, 2019. 2, 3, 4, 5, 6, 8 +[12] Igor I Bussel, Gadi Wollstein, and Joel S Schuman. Oct for glaucoma diagnosis, screening and detection of glaucoma progression. British Journal of Ophthalmology, 98(Suppl 2):ii15-ii19, 2014. 1 +[13] Paola Cascante-Bonilla et al. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. arXiv preprint arXiv:2001.06001, 2020. 2, 3 +[14] Xiangyu Chen, Yanwu Xu, Shuicheng Yan, Damon Wing Kee Wong, Tien Yin Wong, and Jiang Liu. Automatic feature learning for glaucoma detection based on deep learning. In International conference on medical image computing and computer-assisted intervention, pages 669-677. Springer, 2015. 1, 2 +[15] Mark Christopher, Akram Belghith, Christopher Bowd, James A Proudfoot, Michael H Goldbaum, Robert N Weinreb, Christopher A Girkin, Jeffrey M Liebmann, and Linda M Zangwill. Performance of deep learning architectures and transfer learning for detecting glaucomatous optic neuropathy in fundus photographs. Scientific reports, 8(1):1-13, 2018. 1, 2 +[16] Mark Christopher, Akram Belghith, Robert N Weinreb, Christopher Bowd, Michael H Goldbaum, Luke J Saunders, Felipe A Medeiros, and Linda M Zangwill. Retinal nerve fiber layer features identified by unsupervised machine learning on optical coherence tomography scans predict glaucoma progression. Investigative ophthalmology & visual science, 59(7):2748-2756, 2018. 1, 2, 3 +[17] Anne Louise Coleman. Advances in glaucoma treatment and management: surgery. Investigative ophthalmology & visual science, 53(5):2491-2494, 2012. 1 +[18] Ronan Conlon, Hady Saheb, and Iqbal Ike K Ahmed. Glaucoma treatment trends: a review. Canadian Journal of Ophthalmology, 52(1):114-124, 2017. 1 +[19] David P Crabb, Nicholas D Smith, Fiona C Glen, Robyn Burton, and David F Garway-Heath. How does glaucoma look?: patient perception of visual field loss. Ophthalmology, 120(6):1120-1126, 2013. 1 +[20] Andres Diaz-Pinto, Adrián Colomer, Valery Naranjo, Sandra Morales, Yanwu Xu, and Alejandro F Frangi. Retinal image synthesis and semi-supervised learning for glaucoma assessment. IEEE transactions on medical imaging, 38(9):2211-2218, 2019. 2, 4 +[21] Andres Diaz-Pinto, Sandra Morales, Valery Naranjo, Thomas Kohler, Jose M Mossi, and Amparo Navea. Cnns for automatic glaucoma assessment using fundus images: an extensive validation. Biomedical engineering online, 18(1):1-19, 2019. 2 +[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans + +formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.7,8 +[23] Rui Fan, Christopher Bowd, Nicole Brye, Mark Christopher, Robert N Weinreb, David Kriegman, and Linda Zangwill. One-vote veto: Semi-supervised learning for low-shot glaucoma diagnosis. arXiv preprint arXiv:2012.04841, 2020. 2 +[24] Huihui Fang, Fei Li, Huazhu Fu, Junde Wu, Xiulan Zhang, and Yanwu Xu. Dataset and evaluation algorithm design for goals challenge. In International Workshop on Ophthalmic Medical Image Analysis, pages 135-142. Springer, 2022. 1, 2 +[25] David S Friedman, Henry D Jampel, Beatrix Munoz, and Sheila K West. The prevalence of open-angle glaucoma among blacks and whites 73 years and older: the salisbury eye evaluation glaucoma study. Archives of ophthalmology, 124(11):1625-1630, 2006. 2, 8 +[26] Mohammad Ghavamzadeh and Sridhar Mahadevan. Hierarchical policy gradient algorithms. Computer Science Department Faculty Publication Series, page 173, 2003. 3, 5 +[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7, 8 +[28] Patrick Herbert, Kaihua Hou, Chris Bradley, Michael V Boland, Pradeep Y Ramulu, Mathias Unberath, and Jithin Yohannan. Forecasting risk of future rapid glaucoma worsening using early visual field, optical coherence tomography and clinical data. Investigative Ophthalmology & Visual Science, 63(7):2294-2294, 2022. 1 +[29] Donald C Hood and Randy H Kardon. A framework for comparing structural and functional measures of glaucomatous damage. Progress in retinal and eye research, 26(6):688-710, 2007. 1 +[30] Sen Hu, Nicholas D Smith, Luke J Saunders, and David P Crabb. Patterns of binocular visual field loss derived from large-scale patient data from glaucoma clinics. Ophthalmology, 122(12):2399-2406, 2015. 1 +[31] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, et al. Optical coherence tomography. science, 254(5035):1178-1181, 1991. 1 +[32] Alireza Kamalipour, Susan Moghimi, Pooya Khosravi, Vahid Mohammadzadeh, Takashi Nishida, Eleonora Micheletti, Jo-Hsuan Wu, Golnoush Mahmoudinezhad, Elizabeth HF Li, Mark Christopher, et al. Combining oct and oct-angiography longitudinal data for the detection of visual field progression in glaucoma: Longitudinal oct/octa data to detect glaucoma progression. American Journal of Ophthalmology, 2022. 1, 2, 3 +[33] Stuart Keel, Jinrong Wu, Pei Ying Lee, Jane Scheetz, and Mingguang He. Visualizing deep learning models for the detection of referable diabetic retinopathy and glaucoma. JAMA ophthalmology, 137(3):288-292, 2019. 1, 2 +[34] Philip J Kellman and Thomas F Shipley. A theory of visual interpolation in object perception. Cognitive psychology, 23(2):141-221, 1991. 1 +[35] Yuka Kihara, Giovanni Montesano, Andrew Chen, Nishani Amerasinghe, Chrysostomos Dimitriou, Aby Jacob, Almira + +Chabi, David P Crabb, and Aaron Y Lee. Policy-driven, multimodal deep learning for predicting visual fields from the optic disc and oct imaging. Ophthalmology, 2022. 1, 2 +[36] Oleksandr Kovalyk, Juan Morales-Sanchez, Rafael Verdú-Monedero, Inmaculada Sellés-Navarro, Ana Palazon-Cabanes, and José-Luis Sancho-Gómez. Papila: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment. Scientific Data, 9(1):1-12, 2022. 2 +[37] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: a large-scale database and cnn model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2019. 1, 2, 8 +[38] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: A large-scale database and cnn model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2 +[39] Bingyu Liu, Yuhong Guo, Jieping Ye, and Weihong Deng. Selective pseudo-labeling with reinforcement learning for semi-supervised domain adaptation. arXiv preprint arXiv:2012.03438, 2020. 3 +[40] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2, 3, 4 +[41] Fengbei Liu, Yu Tian, Filipe R Cordeiro, Vasileios Bela-giannis, Ian Reid, and Gustavo Carneiro. Self-supervised mean teacher for semi-supervised chest x-ray classification. In International Workshop on Machine Learning in Medical Imaging, pages 426-436. Springer, 2021. 3 +[42] Hanruo Liu, Liu Li, I Michael Wormstone, Chunyan Qiao, Chun Zhang, Ping Liu, Shuning Li, Huaizhou Wang, Dapeng Mou, Ruiqi Pang, et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs. JAMA ophthalmology, 137(12):1353-1360, 2019. 1, 2 +[43] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 3 +[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 7, 8 +[45] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7, 8 +[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5 + +[47] Rizwan Malik, William H Swanson, and David F Garway-Heath. 'structure-function relationship' in glaucoma: past thinking and current concepts. Clinical & experimental ophthalmology, 40(4):369-380, 2012. 1 +[48] Felipe A Medeiros, Alessandro A Jammal, and Atalie C Thompson. From machine to machine: an oct-trained deep learning algorithm for objective quantification of glaucomatous damage in fundus photographs. Ophthalmology, 126(4):513-521, 2019. 1 +[49] Parmita Mehta, Christine A Petersen, Joanne C Wen, Michael R Banitt, Philip P Chen, Karine D Bojikian, Catherine Egan, Su-In Lee, Magdalena Balazinska, Aaron Y Lee, et al. Automated detection of glaucoma with interpretable machine learning using clinical data and multimodal retinal images. American Journal of Ophthalmology, 231:154-169, 2021. 1, 2 +[50] MJ Morgan and RJ Watt. Mechanisms of interpolation in human spatial vision. Nature, 299(5883):553-555, 1982. 1 +[51] Jacqueline M Nelson-Quigg, Kimberly Cello, and Chris A Johnson. Predicting binocular visual field sensitivity from monocular visual field results. Investigative ophthalmology & visual science, 41(8):2212-2221, 2000. 1 +[52] Kouros Nouri-Mahdavi, Vahid Mohammadzadeh, Alessandro Rabiolo, Kiumars Edalati, Joseph Caprioli, and Siamak Yousefi. Prediction of visual field progression from oct structural measures in moderate to advanced glaucoma. American Journal of Ophthalmology, 226:172-181, 2021. 1, 2, 3 +[53] Youngtaek Oh, Dong-Jin Kim, and In So Kweon. Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. arXiv preprint arXiv:2106.05682, 2021. 5, 6 +[54] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical image analysis, 59:101570, 2020. 2 +[55] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6 +[56] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2, 3 +[57] Sonia Phene, R Carter Dunn, Naama Hammel, Yun Liu, Jonathan Krause, Naho Kitade, Mike Schaekermann, Rory Sayres, Derek J Wu, Ashish Bora, et al. Deep learning and glaucoma specialists: the relative importance of optic disc features to predict glaucoma referral in fundus photographs. Ophthalmology, 126(12):1627-1639, 2019. 1, 2 + +[58] Harry A Quigley. Number of people with glaucoma worldwide. British journal of ophthalmology, 80(5):389-393, 1996. 1 +[59] Harry A Quigley and Aimee T Broman. The number of people with glaucoma worldwide in 2010 and 2020. British journal of ophthalmology, 90(3):262-267, 2006. 1 +[60] An Ran Ran, Carol Y Cheung, Xi Wang, Hao Chen, Lu-yang Luo, Poemen P Chan, Mandy OM Wong, Robert T Chang, Suria S Mannil, Alvin L Young, et al. Detection of glaucomatous optic neuropathy with spectral-domain optical coherence tomography: a retrospective training and validation deep-learning analysis. The Lancet Digital Health, 1(4):e172-e182, 2019. 1, 2 +[61] Mamshad Nayeem Rizve et al. In defense of pseudolabeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In International Conference on Learning Representations, 2020. 2, 3 +[62] Alicja R Rudnicka, Shahrul Mt-Isa, Christopher G Owen, Derek G Cook, and Deborah Ashby. Variations in primary open-angle glaucoma prevalence by age, gender, and race: a bayesian meta-analysis. Investigative ophthalmology & visual science, 47(10):4254-4261, 2006. 2, 8 +[63] Weiwei Shi, Yihong Gong, Chris Ding, Zhiheng MaXiaoyu Tao, and Nanning Zheng. Transductive semi-supervised deep learning using min-max features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 299-315, 2018. 2, 3 +[64] Naoto Shibata, Masaki Tanito, Keita Mitsuhashi, Yuri Fujino, Masato Matsuura, Hiroshi Murata, and Ryo Asaoka. Development of a deep residual learning algorithm to screen for glaucoma from fundus photography. Scientific reports, 8(1):1-9, 2018. 1, 2 +[65] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In International conference on machine learning, pages 387-395. PMLR, 2014. 3, 5 +[66] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy P. Lillicrap, Fan Hui, Laurent Sifre, George van den Driessche, Thore Graepel, and Demis Hassabis. Mastering the game of go without human knowledge. Nat., 550(7676):354-359, 2017. 3 +[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7, 8 +[68] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33, 2020. 2, 3, 4, 5, 6, 8 +[69] Yi Sun, Aiming Chen, Minjie Zou, Yichi Zhang, Ling Jin, Yi Li, Danying Zheng, Guangming Jin, and Nathan Congdon. Time trends, associations and prevalence of blindness and vision loss due to glaucoma: an analysis of observational data from the global burden of disease study 2017. BMJ open, 12(1):e053805, 2022. 1 + +[70] Richard S. Sutton, David A. McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems 12, [NIPS Conference, Denver, Colorado, USA, November 29 - December 4, 1999], pages 1057-1063, 1999. 3, 5 +[71] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7, 8 +[72] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10096-10106. PMLR, 2021. 6 +[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 3, 6 +[74] Yih-Chung Tham, Xiang Li, Tien Y Wong, Harry A Quigley, Tin Aung, and Ching-Yu Cheng. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology, 121(11):2081-2090, 2014. 1 +[75] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W VERJANS, and Rajvinder Singh. Self-supervised multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. 2021. 2 +[76] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 128-140. Springer, 2021. 2 +[77] Rohit Varma, Paul P Lee, Ivan Goldberg, and Sameer Kotak. An assessment of the health and economic burdens of glaucoma. American journal of ophthalmology, 152(4):515-522, 2011. 1 +[78] Eija Vesti, Chris A Johnson, and Balwantray C Chauhan. Comparison of different methods for detecting glaucomatous visual field progression. Investigative ophthalmology & visual science, 44(9):3873-3879, 2003. 4 +[79] Xi Wang, Hao Chen, An-Ran Ran, Luyang Luo, Poemen P Chan, Clement C Tham, Robert T Chang, Suria S Mannil, Carol Y Cheung, and Pheng-Ann Heng. Towards multicenter glaucoma oct image screening with semi-supervised joint structure and function multi-task learning. Medical Image Analysis, 63:101695, 2020. 2 +[80] Junde Wu, Huihui Fang, Fei Li, Huazhu Fu, Fengbin Lin, Jiongcheng Li, Lexing Huang, Qinji Yu, Sifan Song, Xingxing Xu, et al. Gamma challenge: glaucoma grading from multi-modality images. arXiv preprint arXiv:2202.06511, 2022. 1, 2 +[81] Jo-Hsuan Wu, Takashi Nishida, Robert N Weinreb, and Jou-Wei Lin. Performances of machine learning in detecting + +glaucoma using fundus and retinal optical coherence tomography images: A meta-analysis. American Journal of Ophthalmology, 2021. 1 +[82] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 7, 8 +[83] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 7, 8 +[84] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual international conference of the IEEE engineering in medicine and biology, pages 3065-3068. IEEE, 2010. 2 \ No newline at end of file diff --git a/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/images.zip b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fe0574315a6c876b61b4012074a62fcc61567b5a --- /dev/null +++ b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0244b27e5eb3dbe2d3c0f1671845a086dd589532fdd19b93981ca5cf287e0fc +size 423652 diff --git a/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/layout.json b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3fbd6e5ac78c26852d1bc433bd18e3763bcad0bc --- /dev/null +++ b/2023/Harvard Glaucoma Detection and Progression_ A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning/layout.json @@ -0,0 +1,9041 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 102, + 544, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 102, + 544, + 138 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 544, + 138 + ], + "type": "text", + "content": "Harvard Glaucoma Detection and Progression: A Multimodal Multitask Dataset and Generalization-Reinforced Semi-Supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 160, + 444, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 160, + 444, + 188 + ], + "spans": [ + { + "bbox": [ + 148, + 160, + 444, + 188 + ], + "type": "text", + "content": "Yan Luo* Min Shi* Yu Tian* Tobias Elze Mengyu Wang \nHarvard Ophthalmology AI Lab, Harvard University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 190, + 476, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 190, + 476, + 201 + ], + "spans": [ + { + "bbox": [ + 117, + 190, + 476, + 201 + ], + "type": "text", + "content": "{yluol6, mshi6, ytian11, tobias_elze, mengyu_wang}@meei.harvard.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 229, + 192, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 229, + 192, + 242 + ], + "spans": [ + { + "bbox": [ + 143, + 229, + 192, + 242 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 254, + 289, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 289, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 289, + 642 + ], + "type": "text", + "content": "Glaucoma is the number one cause of irreversible blindness globally. A major challenge for accurate glaucoma detection and progression forecasting is the bottleneck of limited labeled patients with the state-of-the-art (SOTA) 3D retinal imaging data of optical coherence tomography (OCT). To address the data scarcity issue, this paper proposes two solutions. First, we develop a novel generalization-reinforced semi-supervised learning (SSL) model called pseudo supervisor to optimally utilize unlabeled data. Compared with SOTA models, the proposed pseudo supervisor optimizes the policy of predicting pseudo labels with unlabeled samples to improve empirical generalization. Our pseudo supervisor model is evaluated with two clinical tasks consisting of glaucoma detection and progression forecasting. The progression forecasting task is evaluated both unimodally and multimodally. Our pseudo supervisor model demonstrates superior performance than SOTA SSL comparison models. Moreover, our model also achieves the best results on the publicly available LAG fundus dataset. Second, we introduce the Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, a multimodal multitask dataset that includes data from 1,000 patients with OCT imaging data, as well as labels for glaucoma detection and progression. This is the largest glaucoma detection dataset with 3D OCT imaging data and the first glaucoma progression forecasting dataset that is publicly available. Detailed sex and racial analysis are provided, which can be used by interested researchers for fairness learning studies. Our released dataset is benchmarked with several SOTA supervised CNN and transformer deep learning models. The dataset and code are made publicly available via https://ophai.hms.harvard.edu/datasets/harvard-gdp1000." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "Glaucoma is the leading cause of irreversible blindness globally caused by retinal nerve fiber layer (RNFL)" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 228, + 525, + 280 + ], + "blocks": [ + { + "bbox": [ + 325, + 228, + 525, + 280 + ], + "lines": [ + { + "bbox": [ + 325, + 228, + 525, + 280 + ], + "spans": [ + { + "bbox": [ + 325, + 228, + 525, + 280 + ], + "type": "image", + "image_path": "4e308eae7e47e4262e33c791ccaca8e2fa20cf6ab816fa91ebfa000f3e528c78.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 281, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 304 + ], + "type": "text", + "content": "Figure 1: Examples of RNFLT maps with high similarities but from different label groups." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 326, + 546, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 546, + 432 + ], + "type": "text", + "content": "damage[74, 58, 59, 69]. The global prevalence of glaucoma for the population between 40 and 80 years old is " + }, + { + "bbox": [ + 304, + 326, + 546, + 432 + ], + "type": "inline_equation", + "content": "3.54\\%" + }, + { + "bbox": [ + 304, + 326, + 546, + 432 + ], + "type": "text", + "content": " [74]. Timely clinical treatment in the early stage of glaucoma can significantly reduce the risk of further vision loss [18, 17]. However, most commonly, glaucoma patients are not aware of their early VF loss until the VF loss becomes severe enough to impair their daily activities such as reading and driving due to the brain and fellow eye compensation[4, 19, 34, 50, 51, 30]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 432, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 547, + 715 + ], + "type": "text", + "content": "It is desirable to develop automated glaucoma screen tools to reduce societal disease burden [77]. The clinical diagnosis of glaucoma is a holistic decision primarily based on RNFL damage and VF loss assessment supplemented with other patient clinical information and family disease history [29, 47]. The VF test is only available in the ophthalmology specialty. So existing automated screening tools are based on the idea of predicting glaucoma diagnosis from evaluating RNFL damage using retinal images [14, 57, 42, 5, 49, 37, 15, 35, 60, 81]. Before the massive clinical adoption of the 3D retinal imaging tool of optical coherence tomography (OCT) a decade ago [31, 12], 2D fundus photos were used to assess RNFL damage. Therefore, most prior glaucoma detection models are based on the 2D fundus photo [37, 57, 33, 14, 42, 1, 64, 15], which is not the state of the art (SOTA) clinical imaging modality anymore. OCT has replaced fundus photos as the primary imaging tool for glaucoma clinical care due to its superiority in measuring RNFL damage [48, 12]. There have been a number of recent works using OCT scans to automatically screen glaucoma with machine learning [15, 60, 5, 35, 24, 80, 49]. Apart from glaucoma detection, a more clinically significant task is progression forecasting using OCT scans [52, 16, 32, 28], which could inform clini" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 184, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 184, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 184, + 713 + ], + "type": "text", + "content": "*First three authors contribute equally." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20471" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "cians if they should treat the patient conservatively with eye drops or aggressively with invasive surgeries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 289, + 510 + ], + "type": "text", + "content": "For both the glaucoma screening and progression forecasting tasks, a significant challenge is limited labeled patients with the SOTA OCT measurements as the fundus photo is already outdated in clinical practice [15, 60, 5, 35, 24, 80, 49]. This problem with the lack of labeled patients is even more severe for progression forecasting, which requires longitudinal VF measurements to determine progression outcomes [52, 16, 32]. Therefore, there is a strong need for developing novel semi-supervised learning models to maximally utilize unlabeled patient data in clinics to improve glaucoma detection and progression forecasting. However, apart from several papers using standard semi-supervised learning methods for glaucoma detection [20, 2, 23, 79, 84, 9], there has been no dedicated algorithm development for semi-supervised learning with its application in glaucoma. SOTA semi-supervised learning (SSL) approaches in computer vision are generally based on consistency learning [11, 68, 10, 73] or pseudo-labeling [40, 61]. Unlike consistency learning, the pseudo-labeling approaches do not need domain knowledge to design the image augmentation strategy. While many prior works have endeavored to optimally select pseudo-labeled samples to improve the classifier training, the tasks of glaucoma detection and progression forecasting still have some unique challenges that might not be well tackled by existing pseudo-labeling SSL methods [61, 63, 13, 40, 56]. Figure 1 shows examples of RNFLT maps with high similarities but from different label groups. With such obscured group differences, existing pseudo-labeling SSL methods might not be able to provide sufficient guidance to select correct pseudo-labeled samples to improve the classifier training. Furthermore, having high-quality public datasets with OCT scans for glaucoma detection and progression forecasting is equally important to enable more computer vision researchers to study this topic to push forward the field," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 514, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 289, + 715 + ], + "type": "text", + "content": "In this study, we make two contributions: (1) developing a generalization-reinforced semi-supervised learning model called pseudo supervisor to improve glaucoma detection and progression forecasting. Specifically, our pseudo supervisor model will select pseudo-labeled samples via a policy net that optimizes generalization error with a validation subset to update the classifier. (2) releasing a pilot multitask and multimodal glaucoma dataset, called Harvard Glaucoma Detection and Progression (Harvard-GDP) Dataset, for computer vision researchers to study this topic. In Table1, we have listed major public glaucoma datasets. Most of them are fundus photo datasets, while there are two major OCT datasets previously published [24, 80]. In comparison, our dataset has the largest patient numbers among all datasets with patient numbers available. In addition, our dataset is the first and the only one with progression forecasting task data, which is a way more clinically significant" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 308, + 83, + 547, + 150 + ], + "blocks": [ + { + "bbox": [ + 351, + 71, + 500, + 81 + ], + "lines": [ + { + "bbox": [ + 351, + 71, + 500, + 81 + ], + "spans": [ + { + "bbox": [ + 351, + 71, + 500, + 81 + ], + "type": "text", + "content": "Table 1: Public Glaucoma Datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 83, + 547, + 150 + ], + "lines": [ + { + "bbox": [ + 308, + 83, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 547, + 150 + ], + "type": "table", + "html": "
StudyImaging ModalitySample SizeGlaucoma DetectionProgression ForecastingMultimodalLabel SourceAccessibility
LAG [38]FundusNA (4,854)✓✘✘Clinician AssessmentUpon Request
REFuge [54]FundusNA (1,200)✓✘✘Clinician AssessmentWeb Download
G1020 [6]Fundus432 (1,020)✓✘✘Clinician AssessmentWeb Download
ACRIMA [21]FundusNA (705)✓✘✘Clinician AssessmentWeb Download
RIM-OONE [7]Fundus169 (485)✓✘✘Clinician AssessmentWeb Download
ORIGA [84]FundusNA (650)✓✘✘Clinician AssessmentUpon Request
PAPILA [36]Fundus244 (488)✓✘✘Clinician AssessmentWeb Download
GOALS [24]OCT158 (300)✓✘✘Clinician AssessmentWeb Download
GAMMA [80]OCT300 (300)✓✘✓Clinician AssessmentWeb Download
OursOCT1,000 (1,000)✓✓✓Functional TestWeb Download
", + "image_path": "4589d4ae7629e0e3c430f27592d88e85a11132a3df64e03d3d7aecae8f812026.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 173, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 327 + ], + "type": "text", + "content": "task than glaucoma detection itself. Compared with the other two OCT datasets, our dataset is superior in patient numbers (1,000 versus up to 300), task versatility (glaucoma detection and progression forecasting versus glaucoma detection only), and availability of objective visual function tests (i.e. VF test). Furthermore, since our dataset is from the US population, the racial representation in our data is more diverse than the two OCT datasets [24, 80] and the large LAG dataset [38], which consists of mainly Asian patients. Therefore, our dataset can be potentially used for medical fairness learning studies, especially since we know Black people have more than doubled glaucoma prevalence than other races, which is a significant disparity [62, 25]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 339, + 392, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 392, + 352 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 392, + 352 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 360, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 594 + ], + "type": "text", + "content": "Glaucoma Detection. Vast majority of glaucoma detection works using deep learning are based on the fundus photo, which is a 2D snapshot of the patient's retina [37, 57, 33, 14, 42, 1, 64, 15, 76, 75]. However, the 3D imaging modality OCT scan has replaced fundus photos to become the de facto standard for structural damage assessment in glaucoma. Compared with the deep learning glaucoma detection using fundus photos, deep learning glaucoma detection works using OCT scans generally have much smaller sample sizes, which are up to several thousand [15, 60, 5, 35, 24, 80, 49]. As deep learning models are typically data-hungry, there is a clear need to develop new approaches to deal with the limited labeled data issue. It is known that semi-supervised learning can improve prediction accuracy by leveraging unlabeled data in model training. So far, apart from several glaucoma detection papers using standard semi-supervised learning models [20, 2, 23, 79, 84, 9], no novel semi-supervised learning model has been developed with an application focus on glaucoma." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 596, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 547, + 715 + ], + "type": "text", + "content": "Progression Forecasting. Compared with the glaucoma detection task, progression forecasting is a much more clinically significant task for patients. If we could have an accurate progression forecasting model to inform clinicians of which patient will progress or not, then the clinicians can accordingly decide if a patient should be treated more aggressively with invasive surgeries, which may have a strong side effect, or more conservatively with eye drops. So far, due to the data scarcity issue, we have only found several recent papers using retinal imaging data (whether fundus" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20472" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "or OCT) to forecast progression [52, 16, 32]. The sample sizes of these works are only up to several hundred. Imaginably, we have not found any prior studies applying semisupervised learning for progression forecasting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 289, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 289, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 289, + 530 + ], + "type": "text", + "content": "Semi-Supervised Learning. Semi-Supervised Learning (SSL) aims to optimize a model with both labeled and unlabeled data. The current SSL methods can be divided into pseudo-labeling-based methods [61, 63, 13, 40] and/or consistency-based methods [73, 68, 11, 10, 41, 43]. Consistency-based SSL methods optimize the model with standard classification losses (e.g., cross-entropy) on labeled images and regularizes/minimizes the prediction outputs of different weakly and/or strongly augmented views of unlabeled images, where these views are constructed using different types of image augmentations, such as flip-and-shift, color sharpness, and contrast, etc. Even though consistency-based SSL methods achieve SOTA results in many computer vision benchmarks [73, 68, 11, 10], these methods significantly rely on the orderly design of augmentation functions that requires domain knowledge to design proper image augmentation strategies, which are even more challenging for multimodal modeling. Prior pseudolabeling SSL methods [61, 63, 13, 40, 56] typically utilize a model that trained with the labeled subset to estimate the pseudo labels of the confident unlabeled samples and take these pseudo-labeled data to re-train the model. As mentioned previously, pseudo-labeling-based SSL approaches is a general SSL learning framework for different data modalities naturally. More importantly, as mentioned in Section 1, since the structural differences between glaucoma and non-glaucoma (similarly for progression versus non-progression) are obscured as shown in Figure 1, existing pseudo-labeling SSL methods [61, 63, 13, 40, 56] might not be able to provide sufficient guidance to select correct pseudo labeled samples to benefit the classifier training. In this aspect, we propose a generalization-reinforced SSL model by selecting optimal pseudo-labeled samples for classifier training that explicitly optimize generalization error with a policy net stochastically generates pseudo labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 531, + 288, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 288, + 638 + ], + "type": "text", + "content": "Reinforcement Learning. Reinforcement learning studies an agent interacts with the environment and learns an optimal policy to tackle decision-making problems effectively [70, 8, 26, 65, 66]. Reinforcement learning has been applied to semi-supervised learning setting for real-world applications, e.g. domain adaptation [39]. However, how to utilize reinforcement learning techniques to determine pseudo labels for improving the generalization of classification models remains under exploration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 650, + 241, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 241, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 241, + 662 + ], + "type": "text", + "content": "3. Multimodal and Multitask Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 670, + 241, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 241, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 241, + 684 + ], + "type": "text", + "content": "3.1. Data Collection and Quality Control" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "type": "text", + "content": "This work was approved by the institutional review board (IRB) and adhered to the tenets of the Declaration of" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 74, + 550, + 149 + ], + "blocks": [ + { + "bbox": [ + 310, + 74, + 550, + 149 + ], + "lines": [ + { + "bbox": [ + 310, + 74, + 550, + 149 + ], + "spans": [ + { + "bbox": [ + 310, + 74, + 550, + 149 + ], + "type": "image", + "image_path": "f630c8cb3ff61917878cbf5a43e9c6a22a285bb9fd9b90f9f0a07e40c2fb4d88.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 178 + ], + "type": "text", + "content": "Figure 2: Pointwise RNFLT distribution at different percentiles." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "content": "Helsinki. Because of the retrospective nature of the study, the IRB waived the need for informed consent of patients. The glaucoma patients are from Mass Eye and Ear of Harvard Medical School, which were tested between 2010 and 2021. The RNFLT " + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "inline_equation", + "content": "(0 - 350" + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "content": " microns) is from the OCT report provided by the OCT device manufacturer (Cirrus, Carl Zeiss Meditec, Jena, Germany) calculated as the vertical distance between the internal limiting membrane boundary and nerve fiber layer boundary. The RNFLT measurement is the de facto structural measurement for clinicians to diagnose glaucoma. Each two-dimensional " + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "content": " RNFLT map is generated within the " + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "inline_equation", + "content": "6\\times 6\\mathrm{mm}^2" + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "content": " peripapillary area around the optic nerve head. The VF testing is performed using the Humphrey Field Analyzer with the 24-2 protocol, which measures the patient's eye's visual spatial sensitivity with a radius of 24 degrees from the fixation. Data quality control based on standard clinical practice guideline following the manufacturer's recommendation was applied to exclude unreliable data: (1) OCT scans with signal strength less than 6 (10 represents the best imaging quality) were excluded; (2) VF tests with fixation losses " + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "inline_equation", + "content": ">33\\%" + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "content": ", false positive rates " + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "inline_equation", + "content": ">20\\%" + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "content": " or false negative rates " + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "inline_equation", + "content": ">20\\%" + }, + { + "bbox": [ + 304, + 180, + 547, + 558 + ], + "type": "text", + "content": " were excluded. Each VF measurement is represented as a vector of 52 total deviation (TD) values between -38 dB and 26 dB. In this work, we refer RNFLT map and VF test as two data modalities which constitute a multimodal dataset. Note that, the modifier word \"Harvard\" from the Harvard GDP Dataset only indicates that our dataset is from the Department of Ophthalmology of Harvard Medical School and does not imply an endorsement, sponsorship, or assumption of responsibility by either Harvard University or Harvard Medical School as a legal identity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 566, + 394, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 566, + 394, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 566, + 394, + 578 + ], + "type": "text", + "content": "3.2. Data Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "text", + "content": "The glaucoma detection dataset contains 15,725 training samples, 1,047 validation samples, and 4,287 samples from 5,640, 376, and 1,505 glaucoma patients, respectively. Half of the samples are unlabeled. Among the half samples with labels, " + }, + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "inline_equation", + "content": "45.2\\%" + }, + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "text", + "content": " of the samples are glaucoma, while the rest are non-glaucoma. The progression forecasting dataset includes 5,889 training samples, 431 validation samples, and 1,547 test samples from 1,292, 86, and 345 glaucoma patients, respectively. Among the half samples with progression outcome labels, " + }, + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "inline_equation", + "content": "24.0\\%" + }, + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 305, + 584, + 547, + 715 + ], + "type": "text", + "content": " of the samples are TD Progression and MD Fast Progression, respectively. The" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20473" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "two clinically accepted criteria to define glaucoma progression: (1) TD Progression: eyes with at least three locations with TD slope " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\leq -1" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " dB; (2) MD Fast Progression: eyes with MD slope " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\leq -1" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " dB [78]. The MD Fast Progression is more detrimental for the patients than TD progression [3]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 285 + ], + "type": "text", + "content": "Figure 2 presents the pointwise distributions of 25th, 50th, and 75th percentile RNFLTs in the train (validation set included) and test sets, which demonstrate two obvious superotemporal and inferotemporal RNFL bundles with their damages are known closely related to the visual function loss. As in the two representative examples shown in Figure 1, RNFLT maps with similar appearances may have different labels. This suggests the class boundary may be obscure to determine, which naturally prompts us to adopt semi-supervised solutions to address the glaucoma detection and progression forecasting problems which fully leverage the abundant information in unlabeled samples to mitigate the challenges." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "text", + "content": "The released dataset including 1,000 patients is a random subset from the entire dataset used to build the glaucoma detection and progression forecasting models in this paper, it naturally inherits the original data characteristics. The released glaucoma detection dataset contains 1,000 samples from 1,000 patients, where each sample includes two data modalities (i.e. RNFLT map and VF test) and respective glaucoma labels. The progression forecasting dataset is a subset from the released glaucoma detection dataset, which entails 500 patients with progression labels and patient demographics including age, sex, and race. For all glaucoma patients, the age distribution of 25th, 50th, and 75th percentiles are 54.3, 64.5, and 72.8 years, respectively. The male and female account for " + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "text", + "content": " of the 1,000 patients, respectively. Among the 1,000 patients, the White, Black, and Asian groups account for " + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "inline_equation", + "content": "76.4\\%" + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "inline_equation", + "content": "14.9\\%" + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 46, + 286, + 288, + 533 + ], + "type": "text", + "content": ", respectively. To the best of our knowledge, this is the first public glaucoma dataset with the patient demographic information available, which carries great potential to enable interested researchers to conduct and evaluate fairness learning studies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 547, + 260, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 260, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 260, + 561 + ], + "type": "text", + "content": "4. Problem Formalization & Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 567, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 591 + ], + "type": "text", + "content": "In this section, we formally introduce the problems of glaucoma detection and progression forecasting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": "Denote " + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l = \\{(x_i,y_i)|1\\leq i\\leq N^l\\}" + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": " be a labeled dataset, where " + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}^n" + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": " are the RNFLT maps and " + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": " are corresponding glaucoma/progression labels. Also, we have a stack of unlabeled RNFLT maps, i.e. " + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{ul} = \\{x_j|1\\leq j\\leq N^{ul}\\}" + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": ". We assume that we have a classification model " + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "inline_equation", + "content": "f:\\mathbb{R}^n\\xrightarrow{\\theta}\\mathcal{Y}(f_\\theta" + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": " for short), where " + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": " are the parameters. To update the classification model with the labeled data, a predefined loss function " + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\ell :\\mathcal{V}\\to \\mathcal{V}" + }, + { + "bbox": [ + 46, + 592, + 288, + 714 + ], + "type": "text", + "content": " is used to gauge the discrepancy between the predictions and the ground-truth labels such that the classification model can be updated along" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 68, + 558, + 194 + ], + "blocks": [ + { + "bbox": [ + 306, + 68, + 558, + 194 + ], + "lines": [ + { + "bbox": [ + 306, + 68, + 558, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 68, + 558, + 194 + ], + "type": "image", + "image_path": "fa4f4a7220825318e879c4d2882163076c01f91d7168ea2785a488315c3e180a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 198, + 556, + 296 + ], + "blocks": [ + { + "bbox": [ + 306, + 198, + 556, + 296 + ], + "lines": [ + { + "bbox": [ + 306, + 198, + 556, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 556, + 296 + ], + "type": "image", + "image_path": "4b929fc2225ec1040d39a3e24df785e1d0c60ddb79fa1fdd5cd09d277483bfcc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 297, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 343 + ], + "type": "text", + "content": "Figure 3: The definitions of (a) glaucoma detection task; and (b) progression forecasting task. In the VF test, blue indicates normal visual function while red indicates abnormal visual function." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 369, + 541, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 541, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 541, + 381 + ], + "type": "text", + "content": "the loss, i.e. " + }, + { + "bbox": [ + 305, + 369, + 541, + 381 + ], + "type": "inline_equation", + "content": "\\min_{\\theta} \\ell(f_{\\theta}(x), y)" + }, + { + "bbox": [ + 305, + 369, + 541, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "text", + "content": "As shown in Figure 3 (a), the practical utility scenario of glaucoma detection with deep learning is to use OCT scans that can be easily measured in the setting of primary care or pharmacy to predict glaucoma diagnosis defined by visual field tests that typically only can be measured in the setting of an ophthalmology clinic or eye hospital with specialized technicians. The clinical relevance of progression forecasting as shown in Figure 3 (b) with deep learning is that clinicians can better decide if a patient should be treated more conservatively with eye drops or aggressively with invasive surgeries. As suggested by conventional semi-supervised learning methods [20, 2, 40, 68, 11], pseudo labels " + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "text", + "content": " are predicted by a supervisor model " + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\pi : \\mathbf{R}^n \\xrightarrow{\\omega} \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "text", + "content": " with the unlabeled data " + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\tilde{y} = \\pi(\\tilde{x};\\omega)" + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "text", + "content": ". Similar to the supervised learning setting, the unlabeled data with the pseudo labels can participate in the supervised learning process, namely " + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\min_{\\theta} \\ell(f_{\\theta}(\\tilde{x}), \\tilde{y})" + }, + { + "bbox": [ + 304, + 382, + 545, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 584, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 545, + 713 + ], + "type": "text", + "content": "However, as shown in Figure 4, the natural images associated with different visual concepts are relatively easy to be distinguished. In contrast, the differences in RNFLT maps between glaucoma versus non-glaucoma (similarly for progression versus non-progression) can be very subtle due to large retinal physiological inter-subject variations. There might not be clear and adequate guidance in the learning process to determine the underlying labels of the unlabeled RNFLT maps. Instead, a more sensible way to better determine the pseudo labels is to apply a policy to obtain pseudo labels. Then, we observe how these pseudo labels affect" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20474" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 280, + 134 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 280, + 134 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 280, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 280, + 134 + ], + "type": "image", + "image_path": "a764494e5de5f11426900ba55643901b9d7e776b04d85bec0e36ba572e445deb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 135, + 287, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 135, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 287, + 205 + ], + "type": "text", + "content": "Figure 4: A challenge in glaucoma detection and progression forecasting. Unlike the classification with natural images, where the images with different labels are likely visually distinguishable, RNFLT images with different labels (e.g. glaucoma vs. non-glaucoma and progression vs nonprogression) might not be easily distinguishable." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": "the training process of the classifier and adjust the policy to predict the pseudo labels such that the generalizability of the classifier learned with the pseudo labels can be maximized. In short, the pseudo labels predicted by the supervisor model should be aligned with the improvement of the generalizability of the classifier on a set of unseen data " + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{unseen}" + }, + { + "bbox": [ + 46, + 216, + 288, + 299 + ], + "type": "text", + "content": ". This objective can be formulated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 305, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 287, + 342 + ], + "type": "interline_equation", + "content": "\\min _ {\\{\\tilde {y} \\}} \\frac {1}{| \\mathcal {D} ^ {\\text {u n s e e n}} |} \\sum_ {(x, y) \\in \\mathcal {D} ^ {\\text {u n s e e n}}} \\ell \\left(f _ {\\theta^ {*}} (x), y\\right) \\tag {1}", + "image_path": "ad2294fa05b947e228082f2f160686d73e80f39b62a5e389dc5bcc95a3947962.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 340, + 267, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 340, + 267, + 367 + ], + "spans": [ + { + "bbox": [ + 56, + 340, + 267, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 56, + 340, + 267, + 367 + ], + "type": "inline_equation", + "content": "\\theta^{*}\\in \\arg \\min_{\\theta}\\frac{1}{|\\mathcal{D}^{l + ul}|}\\sum_{(x,y)\\in \\mathcal{D}^{l + ul}}\\ell (f_{\\theta}(x),y)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{l + ul}" + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": " is the union of labeled dataset " + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": " and unlabeled dataset with pseudo labels " + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{ul}" + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 410, + 287, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 287, + 436 + ], + "type": "text", + "content": "5. Generalization-Reinforced Pseudo Supervisor" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 444, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 548 + ], + "type": "text", + "content": "Different from prior-work semi-supervised learning methods [68, 11, 10, 53], the proposed method explicitly leverages the empirical generalizability of the classifier that is quantified with unseen RNFLT maps (e.g. the ones from the validation set) to optimize the policy of predicting the pseudo labels. The empirical generalizability can be viewed as a type of informative reward as the goal of training a classifier is to find one that can generalize well on the unseen test data drawn from the underlying distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 549, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 643 + ], + "type": "text", + "content": "Figure 5 shows how the proposed generalization-reinforced pseudo supervisor predicts pseudo labels and is trained according to the observed sets of states (maps), actions (pseudo labels), and rewards (the empirical generalizability with a mini-batch of validation samples). The proposed pseudo supervisor is based on the policy gradient [70], which is one type of generic reinforcement learning methods [8, 26, 65]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": "Here we formally introduce the concepts related to the proposed pseudo supervisor. An RNFLT map " + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": " can be viewed as a state " + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "inline_equation", + "content": "s_t \\in S" + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": ", while the predicted pseudo label " + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{y}_t" + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": " can be viewed as an action " + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "inline_equation", + "content": "a_{t} \\in \\mathcal{A}" + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": ". A reward function " + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "inline_equation", + "content": "r_t: S \\times \\mathcal{A} \\to \\mathbf{R}" + }, + { + "bbox": [ + 46, + 643, + 287, + 714 + ], + "type": "text", + "content": " is defined on the state-action space. The proposed pseudo supervisor that employs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "a policy to determine pseudo labels is denoted as " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\pi_{\\omega} : S \\to \\mathcal{P}(\\mathcal{A})" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{A})" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " are the probability measures on " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " are the parameters of the proposed pseudo supervisor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": "In the first place, the classifier takes labeled samples for training. The unlabeled samples will be fed into the pseudo supervisor for predicting pseudo labels. Before updating the weights of the classifier, the empirical generalizability of classifier " + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": " can be quantified with validation samples " + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "inline_equation", + "content": "(x^{(val)}, y^{(val)})" + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": ", that is, " + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\ell(f_{\\theta_t}(x^{(val)}), y^{(val)}) (\\ell_t^{(val)})" + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": " is the widely-used cross-entropy loss for classification. After updating the weights of the classifier with unlabeled samples and corresponding pseudo labels, given the same mini-batch of validation samples, the empirical generalizability of the updated classifier can be requantified again, i.e., " + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\ell(f_{\\theta_{t+1}}(x^{(val)}), y^{(val)}) (\\ell_{t+1}^{(val)})" + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": ". Then, the discrepancy between " + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\ell_t^{(val)}" + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "inline_equation", + "content": "\\ell_{t+1}^{(val)}" + }, + { + "bbox": [ + 304, + 108, + 546, + 301 + ], + "type": "text", + "content": " is informative to measure the effect of updating the classifier with the unlabeled samples and the pseudo labels. In this sense, it is desired to use in the reward function, i.e." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 308, + 545, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 308, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 347, + 308, + 545, + 324 + ], + "type": "interline_equation", + "content": "r _ {t} = \\max \\left(\\exp \\left(\\ell_ {t} ^ {(v a l)} - \\ell_ {t + 1} ^ {(v a l)}\\right) - 1, 0\\right) \\tag {2}", + "image_path": "33cc334f3d34812efa6c5a7bc19c1b38786b5499db55f3920aacdab944e48d12.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "text", + "content": "In Eq. (2), if " + }, + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\ell_{t+1}^{(val)}" + }, + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "text", + "content": " is far less than " + }, + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\ell_t^{(val)}" + }, + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "text", + "content": ", it implies that the unlabeled samples and pseudo labels lead to a positive effect on the learning process such that the reward increases adaptively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 379, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 545, + 413 + ], + "type": "text", + "content": "A trajectory of states, actions, and rewards within a time window " + }, + { + "bbox": [ + 305, + 379, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 379, + 545, + 413 + ], + "type": "text", + "content": " will be saved. The expected return of policy optimization can be written as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 417, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 417, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 369, + 417, + 545, + 449 + ], + "type": "interline_equation", + "content": "J \\left(\\pi_ {\\omega}\\right) = \\mathbb {E} _ {\\pi_ {\\omega}} \\left(\\sum_ {k = 0} \\gamma^ {k} r _ {k}\\right) \\tag {3}", + "image_path": "5b7f74b7b23bf4d339f34b55914da4e22a4a10a4817f9c06bd08615b2e23b4fc.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0,1]" + }, + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "text", + "content": " is the discount rate for the numerical stability of reinforcement learning, and " + }, + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "text", + "content": " is the time window for updating " + }, + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\pi_{\\omega}" + }, + { + "bbox": [ + 304, + 454, + 545, + 513 + ], + "type": "text", + "content": " periodically. According to the policy gradient theorem [70], the expectation of the sample gradient is equal to the actual gradient" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 518, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 518, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 315, + 518, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\pi_ {\\omega}} J (\\pi_ {\\omega}) = \\mathbb {E} _ {\\pi_ {\\omega}} \\left[ \\sum_ {k = 0} ^ {\\beta} \\gamma^ {k} r _ {t + k + 1} \\nabla_ {\\pi_ {\\omega}} \\log \\pi_ {\\omega} \\left(a _ {t} \\mid s _ {t}\\right) \\right] \\tag {4}", + "image_path": "2e41dbd3c7ea4ca481fbd3baf23d18e38ff5ac4ccee1fb57ee1be379ca7ec98c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 569, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 569, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 305, + 569, + 545, + 593 + ], + "type": "text", + "content": "Then the policy parameterization can be updated according to the gradient ascend rule" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 371, + 600, + 545, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 600, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 371, + 600, + 545, + 613 + ], + "type": "interline_equation", + "content": "\\pi_ {\\omega} \\leftarrow \\pi_ {\\omega} + \\eta_ {\\pi} \\nabla_ {\\pi_ {\\omega}} J (\\pi_ {\\omega}) \\tag {5}", + "image_path": "3dc88a3eb26c627ecddf4ebb7259eb58534f87d76836055883fc3d1b6d413c3a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 620, + 450, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 450, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 450, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 620, + 450, + 632 + ], + "type": "inline_equation", + "content": "\\eta_{\\pi}" + }, + { + "bbox": [ + 306, + 620, + 450, + 632 + ], + "type": "text", + "content": " is the learning rate of " + }, + { + "bbox": [ + 306, + 620, + 450, + 632 + ], + "type": "inline_equation", + "content": "\\pi_{\\omega}" + }, + { + "bbox": [ + 306, + 620, + 450, + 632 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 640, + 387, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 640, + 387, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 640, + 387, + 654 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 659, + 481, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 481, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 481, + 673 + ], + "type": "text", + "content": "6.1. Set-Up & Implementation Detail" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "To optimize our models, we employ the AdamW optimizer [46] and train all models for 10 epochs in all experiments conducted in this study. For supervised classification" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20475" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 69, + 484, + 179 + ], + "blocks": [ + { + "bbox": [ + 108, + 69, + 484, + 179 + ], + "lines": [ + { + "bbox": [ + 108, + 69, + 484, + 179 + ], + "spans": [ + { + "bbox": [ + 108, + 69, + 484, + 179 + ], + "type": "image", + "image_path": "e46acf5d198aa2915093fb0c4e8c51c068cd6b4bf6b7d6b992d10084bcd216df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 228 + ], + "type": "text", + "content": "Figure 5: An overview of the proposed generalization-reinforced pseudo supervisor in the semi-supervised learning framework. The learning framework consists of the process of training the classifier, the process of measuring the empirical generalizability of the classifier to be used as rewards, and the process of updating the proposed pseudo supervisor to predict pseudo labels that can maximize the classifier's empirical generalizability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "text", + "content": "models for glaucoma detection, we use a learning rate of " + }, + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "text", + "content": " and weight decay of 0, and for progression forecasting, we use a learning rate of " + }, + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "text", + "content": " and weight decay of 0. These hyperparameters are determined to achieve the best baseline performance based on our experiments. The proposed method is trained using a learning rate of " + }, + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 248, + 289, + 471 + ], + "type": "text", + "content": " and weight decay of 0. The batch sizes used in the supervised model, semi-supervised models, and the proposed method vary due to the different semi-supervised learning contexts [73, 11, 10, 68, 53]. We adopt the batch sizes that result in the best performance for each respective context. To ensure consistency with previous work, we apply data augmentation techniques to the samples in the semi-supervised learning models, including MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53]. Specifically, we utilize random crop and resize, as well as random horizontal flip. However, we do not apply these data augmentation techniques to the supervised baseline or the proposed method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 624 + ], + "type": "text", + "content": "For the glaucoma detection task, we train all models with 7,860 labeled samples and 7,865 unlabeled samples for training. We use 1,047 and 4,287 labeled samples for validation and testing, respectively. For the progression forecasting task, all the models are trained with 2,944 labeled samples and 2,945 unlabeled samples for training. We use 431 and 1,547 labeled samples for validation and testing, respectively. For multi-modality modeling, we simply concatenate the RNFLT images and VF matrices for joint learning using the same architecture as our single-modality model and apply this model to the progression forecasting task. The model is trained and evaluated using the same train-val-test split as our single-modality experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 625, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 287, + 661 + ], + "type": "text", + "content": "We use the accuracy, F1 score, and AUC on the test sets as our evaluation matrices. The code is written in PyTorch [55] and we use one RTX A6000 GPU for all experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 153, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 153, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 153, + 682 + ], + "type": "text", + "content": "6.2. Baseline Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Our Supervised baseline is a standard ImageNet pretrained EfficientNet classifier [72] learning with cross" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 248, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 547, + 542 + ], + "type": "text", + "content": "entropy using only labeled images. For Semi-Supervised baselines, we choose MeanTeacher [73], MixMatch [11], ReMixMatch [10], FixMatch [68], and DASO [53] due to their previous powerful performances. MeanTeacher [73] is one of the initial methods to enforce models to predict consistent pseudo predictions for labeled and unlabeled data, which construct a target-generating teacher model by averaging the model weights. MixMatch [11], ReMixMatch [10], and FixMatch [68] are consistency-based SSL methods that are based on the data augmentations (i.e., weak and strong augmentation) to make consistent classification predictions between different views of unlabeled data. DASO [53] suggests that those methods fail for imbalanced datasets and proposes a pseudo-labeling framework based on different consistency SSL models by adaptively mixing the linear and semantic pseudo-labels to mitigate the overall bias between majority and minority classes. We adopt the FixMatch version of DASO, following the default setting reported in the paper [53]. We implement all baseline methods using Pytorch [55] from DASO's official codes and conducted experiments under the same codebase and experimental protocols for fair comparison. All models are equipped with the same EfficientNetV2-S [72] backbone and their hyper-parameters are specifically tuned to suit our Glaucoma detection and progression forecasting tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 553, + 477, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 553, + 477, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 553, + 477, + 565 + ], + "type": "text", + "content": "6.3. Results for Glaucoma Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "text", + "content": "In Table 2, we evaluate our Pseudo-Sup model against multiple SOTA baselines on the glaucoma detection task. Our model achieves the best performance in terms of accuracy, F1 score, and AUC score. We compute the mean performances over 5 different runs. Notably, our model achieves such results without any specific design of image perturbations/augmentations, guaranteeing its simplicity and applicability to different tasks. Our method surpasses the previous SOTA method DASO by " + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "text", + "content": " of accuracy, " + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "text", + "content": " of F1 score, and " + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "text", + "content": " of AUC, respectively. The best-performing consistency method FixMatch still worse by " + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "text", + "content": " of accuracy, " + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "text", + "content": " of F1 score, and " + }, + { + "bbox": [ + 304, + 572, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.5\\%" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20476" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 130, + 285, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 129 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 129 + ], + "type": "text", + "content": "Table 2: Performance on the cross-sectional data for the glaucoma detection task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 130, + 285, + 217 + ], + "lines": [ + { + "bbox": [ + 52, + 130, + 285, + 217 + ], + "spans": [ + { + "bbox": [ + 52, + 130, + 285, + 217 + ], + "type": "table", + "html": "
ModelAcc↑F1↑AUC↑
Baseline0.8131 ± 0.00390.8057 ± 0.00330.8693 ± 0.0022
MeanTeacher0.8148 ± 0.00530.8053 ± 0.00720.8541 ± 0.0040
MixMatch0.6392 ± 0.01610.6273 ± 0.00590.8451 ± 0.0062
ReMixMatch0.6741 ± 0.00190.6716 ± 0.00180.8442 ± 0.0032
DASO0.7444 ± 0.01150.6929 ± 0.00500.8371 ± 0.0118
FixMatch0.8077 ± 0.00320.7912 ± 0.00210.8574 ± 0.0019
Pseudo-Sup0.8124 ± 0.00250.8038 ± 0.00460.8727 ± 0.0006
Pseudo-Sup+Aug0.8224 ± 0.01010.8128 ± 0.01220.8818 ± 0.0020
", + "image_path": "2b31f0b1d780c0c148e37582ca7c7dded7b3a4001e03553c8a9418ed15671a10.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 238, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 332 + ], + "type": "text", + "content": "of AUC than our Pseudo-Sup model, indicating the effectiveness of our approach and the challenges of designing appropriate augmentations for RNFLT maps from previous approaches. We also apply weak augmentations (similar to FixMatch) to our Pseudo-Sup (Pseudo-Sup+Aug) leading to improved performances. Please note that consistency methods (e.g., DASO, FixMatch, etc.) employ both weak and strong augmentations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 342, + 235, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 235, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 235, + 355 + ], + "type": "text", + "content": "6.4. Results for Progression Forecasting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 360, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 490 + ], + "type": "text", + "content": "Unimodal Results. We further evaluate our model on progression forecasting, a more clinically significant task than glaucoma detection, as shown in Table 3. Accurate forecasting of glaucoma progression plays a vital role in prescribing appropriate treatment plans to efficiently prevent further vision losses and is important for assisting the development of new pharmaceutical medicines to treat glaucoma. Adding weak augmentations further improves the results for both progression tasks. Please note that similar to glaucoma detection experiments, we also compute the mean performances over 5 different runs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 490, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 654 + ], + "type": "text", + "content": "For TD Progression, our model achieves the best " + }, + { + "bbox": [ + 46, + 490, + 287, + 654 + ], + "type": "inline_equation", + "content": "74.47\\%" + }, + { + "bbox": [ + 46, + 490, + 287, + 654 + ], + "type": "text", + "content": " of mean AUC, respectively. When compared with consistency-based approaches, our method improves between " + }, + { + "bbox": [ + 46, + 490, + 287, + 654 + ], + "type": "inline_equation", + "content": "0.5\\% - 2.1\\%" + }, + { + "bbox": [ + 46, + 490, + 287, + 654 + ], + "type": "text", + "content": " in AUC. For MD Fast Progression, our approach consistently achieves the best AUC performance, illustrating the effectiveness and accuracy of our method regardless of any glaucoma progression criteria. Given that most of the glaucoma patients are non-progressive, our progression dataset for TD Progression and MD Fast Progression are both highly imbalanced (i.e., most of the samples are labeled as non-progression). Therefore, it is worth noting that our method also outperforms the previous imbalanced SSL SOTA DASO by a large margin in all measures under such imbalanced scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Multimodal Results. One of the major contributions of this work is to introduce a multi-modal dataset for glaucoma progression forecasting. To the best of our knowledge, this is the first work to explore and benchmark SSL approaches using multimodal model inputs. In Table 5, we show the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 130, + 529, + 230 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 129 + ], + "type": "text", + "content": "Table 3: Performance on the longitudinal data with single modality RNFLT for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while Pseudo-Sup represents the proposed pseudo supervisor method." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 130, + 529, + 230 + ], + "lines": [ + { + "bbox": [ + 326, + 130, + 529, + 230 + ], + "spans": [ + { + "bbox": [ + 326, + 130, + 529, + 230 + ], + "type": "table", + "html": "
ModelsAUC ↑ (TD)AUC ↑ (MD Fast)
Baseline0.7367 ± 0.01260.6677 ± 0.0017
MeanTeacher0.7367 ± 0.01260.6677 ± 0.0017
MixMatch0.7238 ± 0.00980.6518 ± 0.0069
ReMixMatch0.7390 ± 0.01240.6595 ± 0.0083
DASO0.7231 ± 0.01880.6787 ± 0.0148
FixMatch0.7395 ±0.01520.6747 ± 0.0054
Pseudo-Sup0.7447 ± 0.00840.7004 ± 0.0218
Pseudo-Sup+Aug0.7477 ± 0.00740.7215 ± 0.0162
", + "image_path": "7fedad25a092748987f9af72a6876e6e2ead14243f8a49f4281b21ba30369d7c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 274, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 403 + ], + "type": "text", + "content": "progression forecasting results using two modalities (RNFLT and VF). Our model achieves the best performance on all evaluation measures. To keep its simplicity and adaptability, we concatenate the RNFLT and the up-scaled VF together as the inputs of our model. Please note that consistency SSL approaches that are based on image augmentations are not included in this table because of the lack of design of augmentation techniques in multimodal medical data. The SOTA results with multimodal glaucoma data enable more accurate and applicable potential computer-aided diagnosis (CAD) systems in real clinical settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 430, + 526, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 526, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 526, + 443 + ], + "type": "text", + "content": "6.5. Supervised Benchmarks on Released Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 455, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 715 + ], + "type": "text", + "content": "In Table 6, we show the supervised classification results for the progression forecasting tasks with multiple SOTA supervised CNN and transformer baseline methods, including VGG [67], ResNet [27], ResNext [82], WideResNet [83], EfficientNet [71], ConvNext [45], ViT [22], and Swin Transformers [44]. This benchmark is conducted on our future data release with 500 patients upon acceptance, of which 400 patients are used for training and the remaining are used for testing. To the best of our knowledge, this is the first supervised glaucoma progression forecasting benchmark, which aims to provide broader impacts for the computer vision and medical imaging communities to build clinically effective computer-aided diagnosis systems. Moreover, the results are reported with both single RNFLT modality and multi-modality (RNFLT and VF) samples. It is worth noting that the accuracy and F1 are the same on MD fast progression because the testing set only contains 100 samples under highly imbalanced class distributions. For more experimental details regarding this supervised progression forecasting benchmark and the supervised results on future release data with 1,000 patients for the glaucoma detection task, please refer to the Supplemental Material." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20477" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 93, + 282, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "type": "text", + "content": "Table 4: Performance comparisons for glaucoma detection task on LAG dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 93, + 282, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 93, + 282, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 93, + 282, + 177 + ], + "type": "table", + "html": "
ModelAcc↑F1↑AUC↑
MeanTeacher0.82800.82790.9082
MixMatch0.50200.33780.9075
ReMixMatch0.69800.47850.8369
DASO0.79710.79060.9164
FixMatch0.83660.82970.9198
Pseudo-Sup0.91600.91580.9858
", + "image_path": "5fb043585c1bd17779971756c97cc938839836f7ba4d32c7a1077e00d4d0ce82.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 74, + 284, + 264, + 335 + ], + "blocks": [ + { + "bbox": [ + 46, + 190, + 287, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 282 + ], + "type": "text", + "content": "Table 5: Performance on the longitudinal data with two modalities (RNFLT and VF) for the progression forecasting task. The baseline method is the supervised model trained with labeled samples only, while pseudo-sup stands for the proposed pseudo supervisor method. Please note that most methods in Table 3 are based on weak/strong augmentations [68, 11, 10] requiring RGB channels, which are not adaptable for multi-modal learning." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 284, + 264, + 335 + ], + "lines": [ + { + "bbox": [ + 74, + 284, + 264, + 335 + ], + "spans": [ + { + "bbox": [ + 74, + 284, + 264, + 335 + ], + "type": "table", + "html": "
ModelsAUC (TD)AUC (MD Fast)
Baseline0.7343 ± 0.00980.6814 ± 0.0081
MeanTeacher0.7436 ± 0.01330.6697 ± 0.0035
Pseudo-Sup0.7517 ± 0.00410.7265 ± 0.0095
", + "image_path": "ec5ee9bba044a276cfb71cc1a53a4fd696e5a0bba86ebfb2daa71b8ae35e243f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 357, + 174, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 174, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 174, + 369 + ], + "type": "text", + "content": "6.6. Performances on LAG" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 377, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 517 + ], + "type": "text", + "content": "In Table 4, we show the performance comparisons on LAG dataset [37]. LAG is a large-scale fundus imaging dataset for glaucoma detection. For this experiment, we use 3,854 images for training with half with labels and the remaining without labels, and 1,000 images for testing, with the size of " + }, + { + "bbox": [ + 46, + 377, + 287, + 517 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 377, + 287, + 517 + ], + "type": "text", + "content": ". For fair comparisons, we use the same hyper-parameters and experimental setups as our other benchmarks. It is worth noting that our approach surpasses the SOTA methods by a large margin and achieves the highest " + }, + { + "bbox": [ + 46, + 377, + 287, + 517 + ], + "type": "inline_equation", + "content": "98.58\\%" + }, + { + "bbox": [ + 46, + 377, + 287, + 517 + ], + "type": "text", + "content": " AUC with only half-labeled data, illustrating the effectiveness and applicability of our proposed Pseudo-Sup in different data and tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 529, + 141, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 141, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 141, + 542 + ], + "type": "text", + "content": "6.7. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "text", + "content": "According to Section 5, the time window " + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "text", + "content": " and discount rate " + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "text", + "content": " are the most important hyperparameters affecting the proposed pseudo supervisor. The ablation studies of the two hyperparameters are shown in Figure 6. We can see that the proposed pseudo supervisor behaviors stably with various " + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "text", + "content": " on AUC. Specifically, when " + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\beta = 50" + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "text", + "content": " leads to better performance. The consistent pattern can be also observed in the analysis of " + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 549, + 288, + 713 + ], + "type": "text", + "content": ". The discount factor is used to achieve a trade-off between the rewards in the distant future and the rewards in the near future. If it is 0, reinforcement learning agents would only look for an optimal policy for the first action. As shown in Figure 6 (b), we find the optimal discount rate at 0.9, which implies that the states, actions, and rewards in the progression task are coherent to a certain degree." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 118, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "type": "text", + "content": "Table 6: Performance of different supervised methods on the longitudinal data with single/multi-modality for the progression forecasting tasks on the released 500 glaucoma progression data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 118, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 194 + ], + "type": "table", + "html": "
ModelTD ProgressionMD Fast Progression
Acc†F1†AUC†Acc†F1†AUC†
VGG [67]0.78/0.800.75/0.740.84/0.790.98/0.980.49/0.490.78/0.68
ResNet [27]0.75/0.740.64/0.670.74/0.750.98/0.980.49/0.490.71/0.59
ResNext [82]0.78/0.770.65/0.680.78/0.760.98/0.980.49/0.490.61/0.80
WideResNet [83]0.77/0.790.71/0.750.77/0.800.98/0.980.49/0.490.71/0.64
EfficientNet [71]0.73/0.780.66/0.730.76/0.790.98/0.980.49/0.490.55/0.66
ConvNext [45]0.74/0.810.62/0.770.78/0.810.98/0.980.49/0.490.84/0.62
ViT [22]0.73/0.770.54/0.710.68/0.790.98/0.980.49/0.490.92/0.69
Swin [44]0.71/0.770.49/0.720.68/0.770.98/0.980.49/0.490.87/0.57
", + "image_path": "42a228de84eaa9a3a132f522e648462b6ba5a313bce0966d87f10d016ae10a2f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 208, + 422, + 291 + ], + "blocks": [ + { + "bbox": [ + 318, + 208, + 422, + 291 + ], + "lines": [ + { + "bbox": [ + 318, + 208, + 422, + 291 + ], + "spans": [ + { + "bbox": [ + 318, + 208, + 422, + 291 + ], + "type": "image", + "image_path": "7245700de1e5b5558aa9f3bcac2a291de22a468708ba495b32fa9bf3a96546ba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 294, + 413, + 304 + ], + "lines": [ + { + "bbox": [ + 349, + 294, + 413, + 304 + ], + "spans": [ + { + "bbox": [ + 349, + 294, + 413, + 304 + ], + "type": "text", + "content": "(a) The impact of " + }, + { + "bbox": [ + 349, + 294, + 413, + 304 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 429, + 206, + 533, + 291 + ], + "blocks": [ + { + "bbox": [ + 429, + 206, + 533, + 291 + ], + "lines": [ + { + "bbox": [ + 429, + 206, + 533, + 291 + ], + "spans": [ + { + "bbox": [ + 429, + 206, + 533, + 291 + ], + "type": "image", + "image_path": "00da3174ba9ddc75138a99fdd156916a6f7ed2abaacb5cb7cba74498e6f74eab.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 293, + 523, + 304 + ], + "lines": [ + { + "bbox": [ + 459, + 293, + 523, + 304 + ], + "spans": [ + { + "bbox": [ + 459, + 293, + 523, + 304 + ], + "type": "text", + "content": "(b) The impact of " + }, + { + "bbox": [ + 459, + 293, + 523, + 304 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 315, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 306, + 315, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 315, + 545, + 339 + ], + "type": "text", + "content": "Figure 6: Ablation study of hyperparameters' impact on TD Progression forecasting." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 360, + 469, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 469, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 469, + 372 + ], + "type": "text", + "content": "6.8. Experiment on Racial Groups" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 377, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 448 + ], + "type": "text", + "content": "The patients in our dataset are from three racial groups (White (0.88 AUC), Black (0.83 AUC), and Asian (0.85 AUC)). In our experiments, we observe that black patients have the lowest AUC yet the highest glaucoma prevalence [62, 25]. Therefore, our dataset is well-suited for future fairness studies." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 479, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we have made two contributions. First, we have developed a generalization-reinforced semi-supervised learning model termed pseudo supervisor for glaucoma detection and progression forecasting. Specifically, the pseudo supervisor predicts pseudo labels following a policy that optimizes the generalization error in training the classifier. Our pseudo supervisor model overall demonstrates superior performance over a number of SOTA methods. Second, we release a multimodal and multi-task dataset for glaucoma detection and progression forecasting with SOTA 3D OCT imaging data. Our dataset is the largest by far for glaucoma detection with 3D OCT data and the first for progression forecasting task. More importantly, our dataset entails demographic information including gender and race, which is generally unavailable in existing public glaucoma datasets. Given the diverse racial representation in our dataset as detailed in Section 3, our dataset is well equipped for potential fairness learning studies, especially considering the significant glaucoma prevalence disparity between Black and other races [62, 25]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20478" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 140, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 140, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 140, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 138 + ], + "type": "text", + "content": "This work was supported by NIH R00 EY028631, Research to Prevent Blindness International Research Collaborators Award, Alcon Young Investigator Grant, NIH R01 EY030575, and NIH P30 EY003790." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 106, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 106, + 164 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 106, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 171, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 225 + ], + "type": "text", + "content": "[1] Jin Mo Ahn, Sangsoo Kim, Kwang-Sung Ahn, Sung-Hoon Cho, Kwan Bok Lee, and Ungsoo Samuel Kim. A deep learning model for the detection of both advanced and early glaucoma using fundus photography. PloS one, 13(11):e0207982, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 288, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 292 + ], + "type": "text", + "content": "[2] Manal Al Ghamdi, Mingqi Li, Mohamed Abdel-Mottaleb, and Mohamed Abou Shousha. Semi-supervised transfer learning for convolutional neural networks for glaucoma detection. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3812-3816. IEEE, 2019. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 293, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 287, + 346 + ], + "type": "text", + "content": "[3] Andrew J Anderson, Abadh K Chaurasia, Ajay Sharma, Amisha Gupta, Shikha Gupta, Anurag Khanna, and Viney Gupta. Comparison of rates of fast and catastrophic visual field loss in three glaucoma subtypes. Investigative ophthalmology & visual science, 60(1):161-167, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 349, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 349, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 349, + 287, + 391 + ], + "type": "text", + "content": "[4] Paul H Artes and Balwantray C Chauhan. Longitudinal changes in the visual field and optic disc in glaucoma. Progress in retinal and eye research, 24(3):333-354, 2005. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 393, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 287, + 468 + ], + "type": "text", + "content": "[5] Ryo Asaoka, Hiroshi Murata, Kazunori Hirasawa, Yuri Fujino, Masato Matsuura, Atsuya Miki, Takashi Kanamoto, Yoko Ikeda, Kazuhiko Mori, Aiko Iwase, et al. Using deep learning and transfer learning to accurately diagnose early-onset glaucoma from macular optical coherence tomography images. American journal of ophthalmology, 198:136-145, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 470, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 470, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 470, + 287, + 534 + ], + "type": "text", + "content": "[6] Muhammad Naseer Bajwa, Gur Amrit Pal Singh, Wolfgang Neumeier, Muhammad Imran Malik, Andreas Dengel, and Sheraz Ahmed. G1020: A benchmark retinal fundus image dataset for computer-aided glaucoma detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "type": "text", + "content": "[7] Francisco José Fumero Batista, Tinguaro Diaz-Aleman, Jose Sigut, Silvia Alayon, Rafael Arny, and Denisse Angel-Pereira. Rim-one dl: A unified retinal image database for assessing glaucoma using deep learning. Image Analysis & Stereology, 39(3):161-167, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 591, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 591, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 591, + 287, + 624 + ], + "type": "text", + "content": "[8] Jonathan Baxter and Peter L Bartlett. Infinite-horizon policy-gradient estimation. Journal of Artificial Intelligence Research, 15:319-350, 2001. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "type": "text", + "content": "[9] Mohammed El Amine Bechar, Nesma Settouti, Vincent Barra, and Mohamed Amine Chikh. Semi-supervised superpixel classification for medical images segmentation: application to detection of glaucoma disease. Multidimensional Systems and Signal Processing, 29(3):979-998, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "text", + "content": "[10] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "and augmentation anchoring. In International Conference on Learning Representations, 2019. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[11] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32:5049-5059, 2019. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[12] Igor I Bussel, Gadi Wollstein, and Joel S Schuman. Oct for glaucoma diagnosis, screening and detection of glaucoma progression. British Journal of Ophthalmology, 98(Suppl 2):ii15-ii19, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 196, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 196, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 196, + 545, + 228 + ], + "type": "text", + "content": "[13] Paola Cascante-Bonilla et al. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. arXiv preprint arXiv:2001.06001, 2020. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 229, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 294 + ], + "type": "text", + "content": "[14] Xiangyu Chen, Yanwu Xu, Shuicheng Yan, Damon Wing Kee Wong, Tien Yin Wong, and Jiang Liu. Automatic feature learning for glaucoma detection based on deep learning. In International conference on medical image computing and computer-assisted intervention, pages 669-677. Springer, 2015. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 295, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 370 + ], + "type": "text", + "content": "[15] Mark Christopher, Akram Belghith, Christopher Bowd, James A Proudfoot, Michael H Goldbaum, Robert N Weinreb, Christopher A Girkin, Jeffrey M Liebmann, and Linda M Zangwill. Performance of deep learning architectures and transfer learning for detecting glaucomatous optic neuropathy in fundus photographs. Scientific reports, 8(1):1-13, 2018. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 372, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 447 + ], + "type": "text", + "content": "[16] Mark Christopher, Akram Belghith, Robert N Weinreb, Christopher Bowd, Michael H Goldbaum, Luke J Saunders, Felipe A Medeiros, and Linda M Zangwill. Retinal nerve fiber layer features identified by unsupervised machine learning on optical coherence tomography scans predict glaucoma progression. Investigative ophthalmology & visual science, 59(7):2748-2756, 2018. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[17] Anne Louise Coleman. Advances in glaucoma treatment and management: surgery. Investigative ophthalmology & visual science, 53(5):2491-2494, 2012. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[18] Ronan Conlon, Hady Saheb, and Iqbal Ike K Ahmed. Glaucoma treatment trends: a review. Canadian Journal of Ophthalmology, 52(1):114-124, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[19] David P Crabb, Nicholas D Smith, Fiona C Glen, Robyn Burton, and David F Garway-Heath. How does glaucoma look?: patient perception of visual field loss. Ophthalmology, 120(6):1120-1126, 2013. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[20] Andres Diaz-Pinto, Adrián Colomer, Valery Naranjo, Sandra Morales, Yanwu Xu, and Alejandro F Frangi. Retinal image synthesis and semi-supervised learning for glaucoma assessment. IEEE transactions on medical imaging, 38(9):2211-2218, 2019. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "text", + "content": "[21] Andres Diaz-Pinto, Sandra Morales, Valery Naranjo, Thomas Kohler, Jose M Mossi, and Amparo Navea. Cnns for automatic glaucoma assessment using fundus images: an extensive validation. Biomedical engineering online, 18(1):1-19, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "text", + "content": "[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20479" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 94 + ], + "type": "text", + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.7,8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 137 + ], + "type": "text", + "content": "[23] Rui Fan, Christopher Bowd, Nicole Brye, Mark Christopher, Robert N Weinreb, David Kriegman, and Linda Zangwill. One-vote veto: Semi-supervised learning for low-shot glaucoma diagnosis. arXiv preprint arXiv:2012.04841, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 139, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 287, + 190 + ], + "type": "text", + "content": "[24] Huihui Fang, Fei Li, Huazhu Fu, Junde Wu, Xiulan Zhang, and Yanwu Xu. Dataset and evaluation algorithm design for goals challenge. In International Workshop on Ophthalmic Medical Image Analysis, pages 135-142. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 192, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 192, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 192, + 287, + 245 + ], + "type": "text", + "content": "[25] David S Friedman, Henry D Jampel, Beatrix Munoz, and Sheila K West. The prevalence of open-angle glaucoma among blacks and whites 73 years and older: the salisbury eye evaluation glaucoma study. Archives of ophthalmology, 124(11):1625-1630, 2006. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 247, + 287, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 247, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 48, + 247, + 287, + 279 + ], + "type": "text", + "content": "[26] Mohammad Ghavamzadeh and Sridhar Mahadevan. Hierarchical policy gradient algorithms. Computer Science Department Faculty Publication Series, page 173, 2003. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 322 + ], + "type": "text", + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 323, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 323, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 287, + 387 + ], + "type": "text", + "content": "[28] Patrick Herbert, Kaihua Hou, Chris Bradley, Michael V Boland, Pradeep Y Ramulu, Mathias Unberath, and Jithin Yohannan. Forecasting risk of future rapid glaucoma worsening using early visual field, optical coherence tomography and clinical data. Investigative Ophthalmology & Visual Science, 63(7):2294-2294, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "text", + "content": "[29] Donald C Hood and Randy H Kardon. A framework for comparing structural and functional measures of glaucomatous damage. Progress in retinal and eye research, 26(6):688-710, 2007. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 431, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 474 + ], + "type": "text", + "content": "[30] Sen Hu, Nicholas D Smith, Luke J Saunders, and David P Crabb. Patterns of binocular visual field loss derived from large-scale patient data from glaucoma clinics. Ophthalmology, 122(12):2399-2406, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 475, + 287, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 527 + ], + "type": "text", + "content": "[31] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, et al. Optical coherence tomography. science, 254(5035):1178-1181, 1991. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 529, + 287, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 287, + 614 + ], + "type": "text", + "content": "[32] Alireza Kamalipour, Susan Moghimi, Pooya Khosravi, Vahid Mohammadzadeh, Takashi Nishida, Eleonora Micheletti, Jo-Hsuan Wu, Golnoush Mahmoudinezhad, Elizabeth HF Li, Mark Christopher, et al. Combining oct and oct-angiography longitudinal data for the detection of visual field progression in glaucoma: Longitudinal oct/octa data to detect glaucoma progression. American Journal of Ophthalmology, 2022. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 616, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 616, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 616, + 287, + 658 + ], + "type": "text", + "content": "[33] Stuart Keel, Jinrong Wu, Pei Ying Lee, Jane Scheetz, and Mingguang He. Visualizing deep learning models for the detection of referable diabetic retinopathy and glaucoma. JAMA ophthalmology, 137(3):288-292, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[34] Philip J Kellman and Thomas F Shipley. A theory of visual interpolation in object perception. Cognitive psychology, 23(2):141-221, 1991. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[35] Yuka Kihara, Giovanni Montesano, Andrew Chen, Nishani Amerasinghe, Chrysostomos Dimitriou, Aby Jacob, Almira" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Chabi, David P Crabb, and Aaron Y Lee. Policy-driven, multimodal deep learning for predicting visual fields from the optic disc and oct imaging. Ophthalmology, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 169 + ], + "type": "text", + "content": "[36] Oleksandr Kovalyk, Juan Morales-Sanchez, Rafael Verdú-Monedero, Inmaculada Sellés-Navarro, Ana Palazon-Cabanes, and José-Luis Sancho-Gómez. Papila: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment. Scientific Data, 9(1):1-12, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 171, + 545, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 171, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 308, + 171, + 545, + 224 + ], + "type": "text", + "content": "[37] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: a large-scale database and cnn model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2019. 1, 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 225, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 225, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 225, + 545, + 277 + ], + "type": "text", + "content": "[38] Liu Li, Mai Xu, Xiaofei Wang, Lai Jiang, and Hanruo Liu. Attention based glaucoma detection: A large-scale database and cnn model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 279, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 279, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 279, + 545, + 322 + ], + "type": "text", + "content": "[39] Bingyu Liu, Yuhong Guo, Jieping Ye, and Weihong Deng. Selective pseudo-labeling with reinforcement learning for semi-supervised domain adaptation. arXiv preprint arXiv:2012.03438, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 323, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 323, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 323, + 545, + 387 + ], + "type": "text", + "content": "[40] Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anticurriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20706, 2022. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "text", + "content": "[41] Fengbei Liu, Yu Tian, Filipe R Cordeiro, Vasileios Bela-giannis, Ian Reid, and Gustavo Carneiro. Self-supervised mean teacher for semi-supervised chest x-ray classification. In International Workshop on Machine Learning in Medical Imaging, pages 426-436. Springer, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 442, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 506 + ], + "type": "text", + "content": "[42] Hanruo Liu, Liu Li, I Michael Wormstone, Chunyan Qiao, Chun Zhang, Ping Liu, Shuning Li, Huaizhou Wang, Dapeng Mou, Ruiqi Pang, et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs. JAMA ophthalmology, 137(12):1353-1360, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 507, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 507, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 507, + 545, + 570 + ], + "type": "text", + "content": "[43] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4258-4267, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 571, + 545, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 625 + ], + "type": "text", + "content": "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 626, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 626, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 626, + 545, + 679 + ], + "type": "text", + "content": "[45] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20480" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[47] Rizwan Malik, William H Swanson, and David F Garway-Heath. 'structure-function relationship' in glaucoma: past thinking and current concepts. Clinical & experimental ophthalmology, 40(4):369-380, 2012. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "text", + "content": "[48] Felipe A Medeiros, Alessandro A Jammal, and Atalie C Thompson. From machine to machine: an oct-trained deep learning algorithm for objective quantification of glaucomatous damage in fundus photographs. Ophthalmology, 126(4):513-521, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 171, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 171, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 288, + 245 + ], + "type": "text", + "content": "[49] Parmita Mehta, Christine A Petersen, Joanne C Wen, Michael R Banitt, Philip P Chen, Karine D Bojikian, Catherine Egan, Su-In Lee, Magdalena Balazinska, Aaron Y Lee, et al. Automated detection of glaucoma with interpretable machine learning using clinical data and multimodal retinal images. American Journal of Ophthalmology, 231:154-169, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 247, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 247, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 48, + 247, + 288, + 268 + ], + "type": "text", + "content": "[50] MJ Morgan and RJ Watt. Mechanisms of interpolation in human spatial vision. Nature, 299(5883):553-555, 1982. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 270, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 270, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 288, + 312 + ], + "type": "text", + "content": "[51] Jacqueline M Nelson-Quigg, Kimberly Cello, and Chris A Johnson. Predicting binocular visual field sensitivity from monocular visual field results. Investigative ophthalmology & visual science, 41(8):2212-2221, 2000. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 313, + 288, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 288, + 366 + ], + "type": "text", + "content": "[52] Kouros Nouri-Mahdavi, Vahid Mohammadzadeh, Alessandro Rabiolo, Kiumars Edalati, Joseph Caprioli, and Siamak Yousefi. Prediction of visual field progression from oct structural measures in moderate to advanced glaucoma. American Journal of Ophthalmology, 226:172-181, 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 367, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 288, + 409 + ], + "type": "text", + "content": "[53] Youngtaek Oh, Dong-Jin Kim, and In So Kweon. Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. arXiv preprint arXiv:2106.05682, 2021. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 483 + ], + "type": "text", + "content": "[54] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical image analysis, 59:101570, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 486, + 288, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 288, + 604 + ], + "type": "text", + "content": "[55] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 605, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 605, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 605, + 288, + 647 + ], + "type": "text", + "content": "[56] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11557-11568, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 648, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 648, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 648, + 288, + 712 + ], + "type": "text", + "content": "[57] Sonia Phene, R Carter Dunn, Naama Hammel, Yun Liu, Jonathan Krause, Naho Kitade, Mike Schaekermann, Rory Sayres, Derek J Wu, Ashish Bora, et al. Deep learning and glaucoma specialists: the relative importance of optic disc features to predict glaucoma referral in fundus photographs. Ophthalmology, 126(12):1627-1639, 2019. 1, 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[58] Harry A Quigley. Number of people with glaucoma worldwide. British journal of ophthalmology, 80(5):389-393, 1996. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[59] Harry A Quigley and Aimee T Broman. The number of people with glaucoma worldwide in 2010 and 2020. British journal of ophthalmology, 90(3):262-267, 2006. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 141, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 215 + ], + "type": "text", + "content": "[60] An Ran Ran, Carol Y Cheung, Xi Wang, Hao Chen, Lu-yang Luo, Poemen P Chan, Mandy OM Wong, Robert T Chang, Suria S Mannil, Alvin L Young, et al. Detection of glaucomatous optic neuropathy with spectral-domain optical coherence tomography: a retrospective training and validation deep-learning analysis. The Lancet Digital Health, 1(4):e172-e182, 2019. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 217, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 260 + ], + "type": "text", + "content": "[61] Mamshad Nayeem Rizve et al. In defense of pseudolabeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. In International Conference on Learning Representations, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 262, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 315 + ], + "type": "text", + "content": "[62] Alicja R Rudnicka, Shahrul Mt-Isa, Christopher G Owen, Derek G Cook, and Deborah Ashby. Variations in primary open-angle glaucoma prevalence by age, gender, and race: a bayesian meta-analysis. Investigative ophthalmology & visual science, 47(10):4254-4261, 2006. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "text", + "content": "[63] Weiwei Shi, Yihong Gong, Chris Ding, Zhiheng MaXiaoyu Tao, and Nanning Zheng. Transductive semi-supervised deep learning using min-max features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 299-315, 2018. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 373, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 425 + ], + "type": "text", + "content": "[64] Naoto Shibata, Masaki Tanito, Keita Mitsuhashi, Yuri Fujino, Masato Matsuura, Hiroshi Murata, and Ryo Asaoka. Development of a deep residual learning algorithm to screen for glaucoma from fundus photography. Scientific reports, 8(1):1-9, 2018. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 428, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 471 + ], + "type": "text", + "content": "[65] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In International conference on machine learning, pages 387-395. PMLR, 2014. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 472, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 547 + ], + "type": "text", + "content": "[66] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy P. Lillicrap, Fan Hui, Laurent Sifre, George van den Driessche, Thore Graepel, and Demis Hassabis. Mastering the game of go without human knowledge. Nat., 550(7676):354-359, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 549, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 549, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 549, + 545, + 581 + ], + "type": "text", + "content": "[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 583, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 583, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 583, + 545, + 646 + ], + "type": "text", + "content": "[68] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33, 2020. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 648, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 648, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 648, + 545, + 712 + ], + "type": "text", + "content": "[69] Yi Sun, Aiming Chen, Minjie Zou, Yichi Zhang, Ling Jin, Yi Li, Danying Zheng, Guangming Jin, and Nathan Congdon. Time trends, associations and prevalence of blindness and vision loss due to glaucoma: an analysis of observational data from the global burden of disease study 2017. BMJ open, 12(1):e053805, 2022. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20481" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[70] Richard S. Sutton, David A. McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems 12, [NIPS Conference, Denver, Colorado, USA, November 29 - December 4, 1999], pages 1057-1063, 1999. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 139, + 288, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 139, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 288, + 182 + ], + "type": "text", + "content": "[71] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 288, + 247 + ], + "type": "text", + "content": "[72] Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 10096-10106. PMLR, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 293 + ], + "type": "text", + "content": "[73] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 347 + ], + "type": "text", + "content": "[74] Yih-Chung Tham, Xiang Li, Tien Y Wong, Harry A Quigley, Tin Aung, and Ching-Yu Cheng. Global prevalence of glaucoma and projections of glaucoma burden through 2040: a systematic review and meta-analysis. Ophthalmology, 121(11):2081-2090, 2014. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 350, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 402 + ], + "type": "text", + "content": "[75] Yu Tian, Fengbei Liu, Guansong Pang, Yuanhong Chen, Yuyuan Liu, Johan W VERJANS, and Rajvinder Singh. Self-supervised multi-class pre-training for unsupervised anomaly detection and segmentation in medical images. 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 480 + ], + "type": "text", + "content": "[76] Yu Tian, Guansong Pang, Fengbei Liu, Yuanhong Chen, Seon Ho Shin, Johan W Verjans, Rajvinder Singh, and Gustavo Carneiro. Constrained contrastive distribution learning for unsupervised anomaly detection and localisation in medical images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 128-140. Springer, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "type": "text", + "content": "[77] Rohit Varma, Paul P Lee, Ivan Goldberg, and Sameer Kotak. An assessment of the health and economic burdens of glaucoma. American journal of ophthalmology, 152(4):515-522, 2011. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "text", + "content": "[78] Eija Vesti, Chris A Johnson, and Balwantray C Chauhan. Comparison of different methods for detecting glaucomatous visual field progression. Investigative ophthalmology & visual science, 44(9):3873-3879, 2003. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 570, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 635 + ], + "type": "text", + "content": "[79] Xi Wang, Hao Chen, An-Ran Ran, Luyang Luo, Poemen P Chan, Clement C Tham, Robert T Chang, Suria S Mannil, Carol Y Cheung, and Pheng-Ann Heng. Towards multicenter glaucoma oct image screening with semi-supervised joint structure and function multi-task learning. Medical Image Analysis, 63:101695, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[80] Junde Wu, Huihui Fang, Fei Li, Huazhu Fu, Fengbin Lin, Jiongcheng Li, Lexing Huang, Qinji Yu, Sifan Song, Xingxing Xu, et al. Gamma challenge: glaucoma grading from multi-modality images. arXiv preprint arXiv:2202.06511, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[81] Jo-Hsuan Wu, Takashi Nishida, Robert N Weinreb, and Jou-Wei Lin. Performances of machine learning in detecting" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 247 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "glaucoma using fundus and retinal optical coherence tomography images: A meta-analysis. American Journal of Ophthalmology, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 106, + 545, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 159 + ], + "type": "text", + "content": "[82] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 161, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 161, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 161, + 545, + 182 + ], + "type": "text", + "content": "[83] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 247 + ], + "type": "text", + "content": "[84] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual international conference of the IEEE engineering in medicine and biology, pages 3065-3068. IEEE, 2010. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20482" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_content_list.json b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..81d20d4a6bc9f9a1bab94ce6dc909d980a4c5dc7 --- /dev/null +++ b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_content_list.json @@ -0,0 +1,2229 @@ +[ + { + "type": "text", + "text": "Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng-Hung Chan", + "bbox": [ + 151, + 180, + 305, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng-Yang Yuan", + "bbox": [ + 334, + 181, + 488, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Sun", + "bbox": [ + 524, + 181, + 616, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hwann-Tzong Chen", + "bbox": [ + 651, + 181, + 816, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Tsing Hua University, Taiwan", + "bbox": [ + 330, + 199, + 640, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3cc7f33312a90cdd9da9e53db670d91daac815c5923e8cb1dc3b49f9c9070c7.jpg", + "image_caption": [ + "Figure 1: Integration of layer-based video editing and spatiotemporal lighting rendering. Our approach decomposes an input video into layers with multiplicative residuals that characterize complex spatiotemporal lighting variations. Our method efficiently fuses the user-edited components with the expected lighting conditions to produce high-quality video output." + ], + "image_footnote": [], + "bbox": [ + 94, + 252, + 235, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/002359bd18af5c95ef2d281d022cba4e63e98a75f708341649fdd59a35d6e294.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 253, + 385, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4956e2e1ffd1221174969f5fec8724be071d64a10cc2977ccf1de1549f977fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 253, + 532, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6adb6c3653031b16414dd3a87c4d4cf4b90494c458bbecd60a3e55bd76d37cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 358, + 532, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80428b5481abd1bd9b3b96b894718951b76ad45ef292400f42496a13927cea92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 300, + 565, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c65c1c7ca37cb7538f7cb9b86c1740058371ec56b0026e3257f5ec18e373ecbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 253, + 720, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75937d4f69df8d5c0db497160b6b17400322bd2df6f2c14e38376608e89462c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 253, + 867, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51036f3ea5921360b0e770ece302c0c5eaa113558382df0ddadf65e2df9f6712.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 358, + 718, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3920fd85b084ed3db956e77d933dd5cbea64d1e064d3192c93c86beea99623e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 358, + 867, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 508, + 313, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a video decomposition method that facilitates layer-based editing of videos with spatiotemporally varying lighting and motion effects. Our neural model decomposes an input video into multiple layered representations, each comprising a 2D texture map, a mask for the original video, and a multiplicative residual characterizing the spatiotemporal variations in lighting conditions. A single edit on the texture maps can be propagated to the corresponding locations in the entire video frames while preserving other contents' consistencies. Our method efficiently learns the layer-based neural representations of a 1080p video in 25s per frame via coordinate hashing and allows real-time rendering of the edited result at 71 fps on a single GPU. Qualitatively, we run our method on various videos to show its effectiveness in generating high-quality editing effects. Quantitatively, we propose to adopt feature-tracking evaluation metrics for objectively assessing the consistency of video editing. Project page: https://lightbulb12294.github.io/hashing-nvd/", + "bbox": [ + 75, + 541, + 473, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike image editing, video editing involves modeling the frame-to-frame relationships and addressing temporal", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "variations such as motion and illumination changes. The task of video editing becomes challenging with the added dimension of time for a user who attempts to apply edits to a video while ensuring consistency across all frames. It is more convenient if we can handle the spatiotemporally varying components and reduce video editing to image editing—The user thus only needs to do edits on images with ease, and the editing results can seamlessly propagate to the entire video.", + "bbox": [ + 496, + 510, + 893, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve this goal, we may consider incorporating effective representations that can model and reconstruct the static and dynamic information in videos. Furthermore, for practical concerns, the algorithm must be efficient enough in modeling and rendering to support interactive editing. Recent work on video decomposition has proposed to employ neural-based representations, such as layered neural atlases [16] and deformable sprites [39], to enable the conversion between the space-time video domain and 2D texture domain for editing. Despite their successes in showcasing impressive video editing effects, we notice that they often require longer training time or restrict to limited frame resolution. For example, it takes more than ten hours to derive the layered neural atlases [16] from a 100-frame 480p video. Deformable sprites [39] are relatively fast to derive but require more computation resources (over 24GB of GPU memory for a 480p video) and thus are unsuitable for editing", + "bbox": [ + 496, + 643, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7743", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d7757873dccaf71565c88b0f5ff4612fb81c20cd4c3d8e6caac04250109bb40.jpg", + "image_caption": [ + "Figure 2: The procedure of layer-based video editing. Our model allows users to apply edits to the extracted texture for rendering the edited video." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 880, + 159 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "high-resolution videos. We address the computation issue by incorporating hash grid encoding [24] into our framework and achieve fast training and rendering for high-resolution videos. Moreover, we introduce the multiplicative residual estimator to model the lighting variations across video frames, which can improve the reconstruction quality and allow illumination-aware editing unachievable by prior work, as shown in Fig. 1.", + "bbox": [ + 75, + 224, + 472, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions of this work as follows:", + "bbox": [ + 94, + 347, + 468, + 361 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. This work is the first to consider spatiotemporally varying lighting effects for layered-based video editing. The proposed multiplicative-residual estimator effectively decomposes the lighting conditions from the video without supervision. Our method can produce better-quality videos by fusing the edits with expected illuminations.", + "2. Our approach is efficient in both training and rendering. Compared with prior work, the proposed method improves the training time with fewer resources and thus enables training on higher-resolution or longer videos. The trained model can achieve fast video rendering via hashing-based coordinate inference. It takes about 40 minutes to train with a 1080p video of 100 frames on a single 3090 Ti GPU. The inference speed for rendering an edited video is 71 fps for 1080p resolution, allowing real-time video editing.", + "3. The experimental results demonstrate appealing video edits in various challenging contexts, such as modifying the texture of moving objects, handling occlusion, and manipulating camera motion, where all can be fused with vivid lighting and shading for better effects that are not easy to achieve by prior work." + ], + "bbox": [ + 88, + 363, + 470, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 722, + 218, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video decomposition. Decomposing a video into layered representations is a long-standing video analysis approach in computer vision since the seminal work by Wang & Adelson [33]. Similar ideas have been continually revisited under different contexts with the development of new techniques. A typical formulation is motion segmentation that decomposes the video by motion [6, 15, 17, 35]. Video matting, on the other hand, focuses more on separating the alpha matte of the foreground object from the background for blending [9, 22, 30, 36]. Recently, neural-based methods have", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shown to be effective in estimating layered representations for video segmentation and video editing [3, 14, 16, 38, 39]. In this work, we also adopt neural networks to derive layered representations from videos. Further, inspired by the pioneering work in visual tracking for handling illumination changes [10], we incorporate a new multiplicative residual representation to model the lighting variations for illumination-aware video editing.", + "bbox": [ + 496, + 224, + 893, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video editing. Layered representations can benefit video editing in various ways. For example, the layered representations can serve as a visual proxy for intuitive video editing [16] or can be used to create the retiming effects [21]. Editing can be more easily done on a unified texture map built from the video's background, such as background mosaics [7], tapestries [4], and layered neural atlases [16]. Building a unified texture map for the foreground object is also useful, e.g., flexible sprites [15], unwrap mosaics [28], and deformable sprites [39]. Another type of decomposition is to derive temporally-consistent intrinsic components from videos [27, 20] so that the edits can be performed on the reflectance for recoloring or texture transfer. Recent deep-learning-based methods mainly address a single task of video editing, e.g., video style transfer [34, 19, 18, 12, 29]. or category-specific GAN-based video editing [2, 25, 32, 37]. Regarding achieving consistent video editing, it is crucial to know the temporal correspondences [13, 5] so typical animation techniques like Rotoscoping [1] can be applied.", + "bbox": [ + 496, + 380, + 895, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 691, + 609, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formulate the problem of consistent video decomposition and editing as follows: The goal is to decompose an input video into multiple layers, where one layer comprises the background, and each of the remaining layers represents a foreground object. We aim to facilitate user-friendly and intuitive editing of video components on time-independent static texture maps while ensuring the quality and consistency of the edited results in the output video. Fig. 2 illustrates an editing procedure. The model may rely on additional information obtained during pre-processing, such as backward and forward optical flow and rough object masks, to improve the training efficiency and consistency.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7744", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Layer-Based Video Decomposition", + "text_level": 1, + "bbox": [ + 76, + 90, + 377, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A video of a scene typically contains multiple objects with distinct motions, appearances, shapes, and other attributes. Achieving consistent editing on the video scene and objects can be challenging. An efficient way to group objects is by decomposing the video into different motion layers. Static objects can be considered part of the scene's background, while only a few foreground objects move differently. Editing the objects on decomposed layers becomes more intuitive and straightforward since we can modify each object separately without touching the others. We decompose videos into layers using backward and forward optical flow and coarse object masks, resulting in $N + 1$ layers where layer $N$ represents the background, and the rest is for foreground objects. Each layer $n$ comprises a predicted mask $\\alpha_{n}$ across different frames, a time-independent texture map $M_{n}$ , and the multiplicative residual $\\ell_{n,t}$ characterizing the illumination for each frame $t$ built on the texture map. Fig. 3 shows an overview of our model.", + "bbox": [ + 75, + 113, + 473, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layer hierarchy. Our alpha network $\\mathcal{A}$ predicts the object mask probability for every layer. Rather than simply normalizing the layer probabilities of each pixel and directly using them as layer masks, we build a front-to-back hierarchy. That is, for layer $n$ , its probability of object mask is computed by", + "bbox": [ + 75, + 406, + 470, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {n} = a _ {n} \\cdot \\prod_ {i = n + 1} ^ {N} \\left(1 - a _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 508, + 468, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $a \\in [0,1]$ is the raw output of $\\mathcal{A}$ with $a_N$ fixed and equal to 1. This hierarchy assumes that objects follow a strict front-to-back ordering, which also implies that layer $N$ represents the background. Here, we specifically focus on scenes with only one foreground object. That is, $N = 1$ . We show an example of decomposing multiple foreground objects in the supplementary material.", + "bbox": [ + 75, + 563, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Texture Mapping", + "text_level": 1, + "bbox": [ + 76, + 680, + 246, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We associate each object with its own texture, such that modifying a single pixel in the texture will impact the object's appearance in the rendered video scene. We translate the spatiotemporal locations in the video into 2D UV coordinates for different layers by a mapping network $\\mathcal{M}$ . This allows us to sample colors on a texture with the translated coordinates. We follow the approach of coordinate-based neural rendering [16, 23] and use the multi-layer perceptrons as our coordinate translator instead of a textured grid.", + "bbox": [ + 75, + 704, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the appearance of an object in the scene may change throughout the video due to camera motion and illumination variations. We will explain how we address this issue with the proposed multiplicative residual in Sec. 3.3.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The textures can be further obtained by grid-sampling the texture network $\\mathcal{T}$ . Once the texture is extracted, the texture network can be replaced by the extracted texture using UV coordinates from the mapping network, achieving the same effect via bilinear interpolation. This process enables users to modify and apply the extracted texture to the entire video.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Multiplicative-Residual Estimator", + "text_level": 1, + "bbox": [ + 498, + 196, + 799, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to apply a single modification to all video frames, the texture must remain constant throughout the entire video. However, the color at the same location on an object may change over time due to variations in lighting conditions, resulting in low reconstruction quality or noisy textures. To address this issue and ensure high-quality reconstruction, we have developed a new method, multiplicative-residual estimator, to achieve constant texture across all frames while preserving illumination variations.", + "bbox": [ + 496, + 220, + 893, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this design, we can also synthesize illumination coefficients not present in the original video at specific time points, allowing us to manipulate camera motion. Note that the prediction of the residual estimator is sharp. Therefore, we have designed losses to constrain the seen and unseen areas, as will be described in Sec. 3.5. The multiplicative-residual estimator predicts the illumination coefficient of each color channel on the texture map, which is multiplied by each channel to obtain the shaded color.", + "bbox": [ + 496, + 359, + 893, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To provide more insight into our decision to use a multiplicative residual instead of an additive residual, we examine the process of diffuse shading:", + "bbox": [ + 496, + 496, + 893, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\rho L, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 558, + 890, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\rho$ is the diffuse albedo term determined by surface material and $L$ is the lighting term determined by the environment. When the lighting condition changes or a shadow is cast onto the surface, the new shaded color then becomes $C' = \\rho \\cdot L'$ . With multiplicative residual, we can reproduce $C'$ by scaling $C$ with $\\frac{L'}{L}$ . In contrast, we have to add $C$ by $\\rho (L' - L)$ if we use additive residual. One can observe that the additive residual is more difficult to model as it also depends on the diffuse albedo term. We detail the practical result of the two choices in Sec. 4.6.", + "bbox": [ + 496, + 589, + 893, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Network Architecture", + "text_level": 1, + "bbox": [ + 498, + 753, + 705, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design our model as an end-to-end framework trained via self-supervision. We take a spatiotemporal coordinate $p = (x, y, t)$ in the video as the model input and predict the reconstructed color of the corresponding pixel. Fig. 3 shows the pipeline of our model, which includes a mapping network, a texture network, and a multiplicative-residual estimator for each layered representation, with a shared alpha network generating the soft masks for different layers.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7745", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/52225fd8bcd850d0b1e744bc521de56a1fc70f52c84fda14212a25e00e514ccb.jpg", + "image_caption": [ + "Figure 3: Model pipeline. Our model takes a video coordinate $p = (x, y, t)$ as input and decomposes the video into multiple layers. A representation for layer $n$ is modeled by a mapping network $\\mathcal{M}$ , a texture network $\\mathcal{T}$ , and a multiplicative-residual estimator $\\mathcal{R}$ ; the three modules jointly convert a video coordinate $p$ into layered color $c_{n}^{(p)}$ and lighting coefficient $l_{n}^{(p)}$ . An alpha network $\\mathcal{A}$ also takes $p$ as input and predicts soft object masks $\\alpha^{(p)}$ for each layer. The final video color is reconstructed by the bottom-right equation. The multiplicative residuals are critical to our model for handling illumination variations." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 869, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mapping network. Following the design choice of recent coordinate-based approaches [16, 23], we build our network architecture with multi-layer perceptrons. For each layer $n$ , we pass a spatiotemporal coordinate $p$ through a mapping network to translate the video coordinate into a time-independent 2D texture-map coordinate $(u, v)$ by", + "bbox": [ + 75, + 419, + 470, + 511 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(u, v) _ {n} ^ {(p)} = \\mathcal {M} _ {n} (p), \\quad (u, v) _ {n} ^ {(p)} \\in [ - 1, 1 ] ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 515, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_n$ is the mapping network of layer $n$ .", + "bbox": [ + 76, + 540, + 383, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Texture network and multiplicative-residual estimator. After obtaining the UV coordinates, the texture network takes the UV coordinates $(u,v)_n^{(p)}$ and produces the corresponding color at each position, while the multiplicative-residual estimator uses both the UV coordinates and the time $t$ as inputs to predict the corresponding illumination coefficients at each position:", + "bbox": [ + 75, + 571, + 470, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {n} ^ {(p)} = \\mathcal {T} _ {n} \\left((u, v) _ {n} ^ {(p)}\\right), \\quad c _ {n} ^ {(p)} \\in [ 0, 1 ] ^ {3},\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 686, + 403, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {n, t} ^ {(p)} = \\mathcal {R} _ {n} \\left((u, v) _ {n} ^ {(p)}, t\\right), \\quad \\ell_ {n} ^ {(p)} \\in \\mathbb {R} _ {+} ^ {3},\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 710, + 403, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{T}$ is the texture network and $\\mathcal{R}$ is the multiplicative-residual estimator.", + "bbox": [ + 75, + 743, + 470, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alpha network. We also predict an object mask for each layer to indicate the layer to which each pixel belongs:", + "bbox": [ + 76, + 791, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} ^ {(p)} = \\mathcal {A} (p), \\quad \\boldsymbol {\\alpha} ^ {(p)} \\in [ 0, 1 ] ^ {N + 1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 825, + 468, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{n}^{(p)}$ is the probability that pixel $p$ belongs to layer $n$ , i.e., $\\sum_{n=0}^{N} \\alpha_{n}^{(p)} = 1$ . The derivation of the object mask is detailed in Sec. 3.1.", + "bbox": [ + 76, + 851, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pixel color reconstruction. After obtaining the information described above, we can reconstruct the pixel color at position $p$ as", + "bbox": [ + 498, + 419, + 893, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} ^ {(p)} = \\sum_ {n = 0} ^ {N} c _ {n} ^ {(p)} \\cdot \\ell_ {n} ^ {(p)} \\cdot \\alpha_ {n} ^ {(p)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 476, + 892, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [24], we use hash grid encoding for all networks except for the mapping network, as texture coordinates are expected to be smooth. This approach is adopted to facilitate better convergence during training.", + "bbox": [ + 498, + 529, + 892, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Loss Terms", + "text_level": 1, + "bbox": [ + 500, + 599, + 624, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Losses inherited from previous work. We incorporate several losses from Layered Neural Atlases [16] to facilitate the training process. These losses include 1) optical flow loss, which provides the model with the supervision of optical flow, and 2) alpha bootstrapping loss, which offers supervision of initial coarse masks; additionally, we use 3) reconstruction loss to stabilize the training and improve reconstruction quality, and 4) sparsity loss to avoid duplicate foreground objects in the texture area. Further details can be found in the supplementary material.", + "bbox": [ + 496, + 623, + 893, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Residual consistency loss. In order to produce significant yet smooth lighting conditions, the multiplicative residuals of different times $t_1$ and $t_2$ of the same position on the texture coordinate $(u, v)$ , where $(u, v, t_1)$ corresponds to a visible area and $(u, v, t_2)$ corresponds to an invisible area, should be close in distribution and can have different intensities of illumination. Formally, we have a small $k \\times k$ patch $P$", + "bbox": [ + 496, + 794, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7746", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "at time $t_1$ on the video centered at $(x, y)$ . We then sample their texture coordinates $(u, v)^{(P)} = \\mathcal{M}(P)$ . The texture coordinates are then combined with different times $t_1$ and $t_2$ to get the corresponding multiplicative residuals:", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell^ {(P)} = \\mathcal {R} ((u, v) ^ {(P)}, t _ {1}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 161, + 468, + 193 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell^ {\\prime (P)} = \\mathcal {R} \\left((u, v) ^ {(P)}, t _ {2}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 189, + 362, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we encourage multiplicative residuals of the same position on texture coordinate should be positively correlated. Therefore, we define the residual consistency loss through normalized cross-correlation by", + "bbox": [ + 75, + 222, + 468, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (P, t _ {2}) = \\frac {\\left(\\ell^ {(P)} - \\mu_ {\\ell (P)}\\right) \\left(\\ell^ {\\prime (P)} - \\mu_ {\\ell^ {\\prime} (P)}\\right)}{\\sigma_ {\\ell (P)} \\sigma_ {\\ell^ {\\prime} (P)}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 291, + 468, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further introduce a variance-smoothness term that smoothes the changes in lighting conditions as", + "bbox": [ + 75, + 335, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} (P, t _ {2}) = \\sigma_ {\\ell^ {\\prime} (P)} ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 375, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply the above terms to all layers and get the overall loss as", + "bbox": [ + 75, + 402, + 467, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {R} \\text {c o n}} = \\lambda_ {\\mathcal {R} \\text {c o n}} (\\psi + \\beta \\mathbb {E}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 433, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we choose $k = 3$ since our patch size is $3 \\times 3$ and $\\beta = 16$ .", + "bbox": [ + 75, + 455, + 468, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Residual regularization. Minimizing the current losses can collapse to a trivial solution because the multiplicative-residual estimator absorbs all colors. We regularize the residuals equal to 1 since \"true color\" of an object in the video depends on the overall light conditions. That is, if the light source is a pure blue light in the entire video shining on a white wall, we can only say that the steady color of the wall is \"blue\", as we do not have enough information to determine the true color of the wall.", + "bbox": [ + 75, + 503, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {R} \\text {r e g}} = \\lambda_ {\\mathcal {R} \\text {r e g}} | | \\mathcal {R} (\\cdot) - 1 | | _ {2} ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 648, + 468, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alpha regularization. We introduce an additional regularization term to address potential issues with the alpha network. Specifically, the sparsity loss used in the alpha network tends to assign a value of zero (i.e., black) to regions that are not visible in the input video, which can result in noisy masks and incorrect shadow embedding. To mitigate this problem, we enforced a constraint that each pixel should contribute to at most one layer. This ensures that the mask for each layer is clean and reliable and that the lighting conditions are properly embedded in each layer.", + "bbox": [ + 75, + 683, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\alpha \\operatorname {r e g}} = \\lambda_ {\\alpha \\operatorname {r e g}} \\mathrm {B C E} \\left(\\max _ {n \\in \\{0, \\dots , N \\}} \\alpha_ {n}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 843, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where BCE is binary cross entropy.", + "bbox": [ + 76, + 885, + 310, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Hash Grid Encoding", + "text_level": 1, + "bbox": [ + 500, + 90, + 694, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Like [16, 23], we have considered employing positional encoding as the input encoding for the multi-layer perceptrons. The purpose of including the positional encoding is to enrich the discriminative power of the coordinates and ensure that high-frequency details can be adequately represented. Inspired by [24], we choose to adopt hash grid encoding as the input encoding method for our model. By leveraging this technique, the input features are encoded into a set of intermediate representations that span a broad spectrum of spatiotemporal resolutions, ranging from coarse-to fine-grained, and can flexibly adjust to distinct regions while maintaining high levels of accuracy and consistency. Formally, the 2D or 3D input feature is treated as a coordinate of multi-resolution grids and then used to sample data from the grids using interpolation techniques. The resulting sampled data are concatenated and passed into multi-layer perceptrons to obtain the final output.", + "bbox": [ + 496, + 114, + 893, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 386, + 632, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct our experiments on the DAVIS dataset [26] and various internet videos to demonstrate the effectiveness of our approach in video reconstruction and consistent video editing. We also design an evaluation metric of edit consistency on the TAP-Vid-DAVIS dataset [8]. In addition, we use our method to generate videos with different camera motions from those in the original input video. Furthermore, we conduct ablation studies on the multiplicative-residual estimator and encoding type to assess the impact on the performance of different architecture choices.", + "bbox": [ + 496, + 411, + 893, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 574, + 681, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 shows four qualitative results of our work on the DAVIS dataset [26]. For each video frame, we show the predicted masks (second and third columns), the m-residuals (fourth column), the corresponding layered textures (fifth and sixth columns), and the reconstruction result (last column). Please also refer to Fig. 1 for an example result highlighting the robustness of our m-residual estimation. More results and higher resolution reconstruction can be found in the supplementary material.", + "bbox": [ + 496, + 597, + 893, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With merely the guidance from the coarse object mask provided by Mask R-CNN [11], our model successfully separates the object from the background yielding a clean and precise mask. In the second video in Fig. 4, we showcase an example of a video occluded by complex water splashes. Our model can accurately reconstruct the video without introducing noticeable texture distortions. We superimpose a rainbow checkerboard on the mask to visualize the transformation of each layer. It is evident that each coordinate's component remains at the same position in the texture.", + "bbox": [ + 496, + 734, + 893, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the multiplicative residual may brighten or darken", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7747", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f69de98daf2a54b7b55ab918e6b9a533c9a408e6079f1099873d8a1514292787.jpg", + "image_caption": [ + "Figure 4: Qualitative results on the DAVIS dataset. We show our predicted masks, textures, m-residuals, reconstructed results, and PSNR of different videos. We overlay a color checkerboard on the masked areas to visualize the texture transformations." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 872, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the texture, we display the m-residual output on a gray canvas to visualize its effect. In the first video, our m-residual estimator successfully models the shadow of the hiker while the background texture retains its original color. The m-residual estimator takes both the texture coordinate and time as the input, allowing it to learn how to model the changes in lighting conditions through space-time and adjust the color accordingly. Such a mechanism enables our model to generate realistic and consistent shading effects for different objects in the video, such as the shading on the bear's fur or the reflection on the hiker's hat. We also report the PSNR of each video as evidence of high-quality reconstruction.", + "bbox": [ + 75, + 522, + 472, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Previous Work", + "text_level": 1, + "bbox": [ + 76, + 710, + 366, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report the PSNR values and various settings for every approach in Table 1. We also provide more quantitative comparisons with DeformableSprites (DS) [39] and Layered Neural Atlases (LNA) [16] under multiple metrics in Table 2. Our method demonstrates superior performance of PSNR, rendering speed, and GPU memory consumption compared to previous methods on all three videos while maintaining the same or better resolution. Note that we could not run DeformableSprites for the $768 \\times 432$ resolution, as DeformableSprites failed to finish due to their GPU memory usage exceeding the limit of a single GPU with", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "24 GB memory. In contrast, our method can efficiently decompose higher-resolution videos into full-HD layers using affordable computation resources while maintaining comparable reconstruction results. To compute the PSNR, we scale the input video up to the reconstructed resolution. Therefore, the PSNR of our 1080p model might be slightly lower than that of the $768 \\times 432$ model, as the PSNR of the 1080p model is calculated at the corresponding resolution and considered harder due to having more details to reconstruct than the low-resolution counterpart.", + "bbox": [ + 496, + 522, + 893, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Consistent Video Editing", + "text_level": 1, + "bbox": [ + 498, + 680, + 728, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For editing purposes, we use a $1000 \\times 1000$ grid sampling to render the texture networks. We then modify the rendered texture and sample color via bilinear interpolation. As our key advantage lies in the m-residual representation of illumination, we have focused our editing on videos or components that exhibit changes in lighting.", + "bbox": [ + 496, + 704, + 893, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show our editing results in Fig. 5. We choose three videos to present the rendering quality. The first video is sourced from internet videos, while the other two are taken from the DAVIS dataset [26]. For evaluating the editing quality, we modify the background texture of the first two videos by adding handwritten characters of \"ICCV\". We also adjust the color of the shirt of the hiker in the second video.", + "bbox": [ + 496, + 795, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7748", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b5ae996a36d23ea05bf981743f1c1df593a8a6dfb3bbfd4df611667fe2047c7.jpg", + "image_caption": [ + "(a) Input frames" + ], + "image_footnote": [], + "bbox": [ + 91, + 104, + 235, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e1d61ece7af3520bff1e907b8c311227672d654b247b7afc181db92333f8ea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 169, + 235, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c5b2cec9fa357f17c05496ac7b8ca69792aa5a530a623b41ff7459c7a333990.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 233, + 235, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1ec89e8585bf3dc67a942fd93b9c84ff887645ff1d9d623a4ee70b4107f1e43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 297, + 235, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb1e93478c37d6d36f3ff5dcf9dcc05d7541e3031bf5fa0f249469a921fca723.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 361, + 235, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/169d459c2ecf5f31748f113496a4f4cae1953973bd5fd9e8acc6982b5d74e325.jpg", + "image_caption": [ + "Figure 5: Edits on videos with varying lighting conditions. We apply various editing techniques on different components of the videos to observe the corresponding changes in the lighting conditions. The textures with edits are shown in (c), while the editing results are enlarged in the right column of (d). The lighting and shading fuse with the edits flawlessly." + ], + "image_footnote": [], + "bbox": [ + 91, + 424, + 235, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e57df96c2fe4c15a221a70997cf63a3314a3c171bafa80db868fde585c8f06d8.jpg", + "image_caption": [ + "(b) Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 236, + 104, + 383, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b67b1e8f2d0b569625f7fa9ab7db15dfb9e459391420e33f21a7460b8132944.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 169, + 383, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/239bccd5f35a7b7cad2992df4e1095b55e4b3effb536400f0225072fb9ad1788.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 233, + 383, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4508c8874d08caf9558a01c59a62196c00a665bf1b3d72ee90457a477ec04cf7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 297, + 383, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1b123cc8b196cae4f1ee51fa47dcf9da80b212c1daf8255deb341e256f7e6f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 361, + 383, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b92f6b643eb3e94000bd3ab2e873508c23d989e005776e2f6324ef5d0366f0a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 424, + 383, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de186c67c6d37eb10c8175e65533fe6b7a91bbadb7eb6b0ce429b5daf400b426.jpg", + "image_caption": [ + "(c) Texture editing" + ], + "image_footnote": [], + "bbox": [ + 401, + 130, + 573, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a28123c83715de266bb264d5d7993753b6c8d55e67f0cb8020deb65c35cba6de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 234, + 581, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5b2c0e3f1d466fe55c50db0bea07e0eb6fc70a21c49695d683500a5ba61b926.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 309, + 563, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bd04ee50860b8df27a25851b471432485759c269ef98b8b1917837ec312475f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 362, + 562, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd277c7238ac925a08bb89dc1ca4bd5708fcbc114ace71bf1c35529389f16b28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 424, + 562, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdb650b79c19a6754ca2b75715c98f00c2f506c7058314b2cb93257f3b81ddb4.jpg", + "image_caption": [ + "(d) Reconstruction (edited)" + ], + "image_footnote": [], + "bbox": [ + 588, + 104, + 730, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81d36b73c367f29bb4ccfde664f4763de21457cf07e7a8a7a8557a9ac524dd33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 169, + 730, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f35b7475758a0fbb1751ecd5da412d41773eeddea453acb2ca1533e94b15f20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 233, + 730, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef89203d8d4142c237b0ce44e03b23880f633c3144737832ae9888349359a74b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 297, + 730, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f80e9cdd1c668cfa7401e7a427da8cf7945a918b6c706a379bab560313123409.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 361, + 730, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e79c64c757640c82b77a046f938bf84c3a8eff0b6d71a3ae42959d7149a99bd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 424, + 730, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92f42925567da358c7cf59d98f3f1b807eec23c93e0ddcb26c0f49155f44265c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 104, + 877, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c44b9b682d8053232812e95cee3d177119fa74a6984890fb9888ab0a7ab80c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 169, + 877, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f0299942457d08e454a5e9039614f5142fd792a7fcf22dbb955dc83451b79b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 233, + 877, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7b11340b5d5dd8ac5d606c72c3dc002b8b2778bc82f438af938741e589a5e4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 297, + 877, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8f7098b22ddb80a9c896e8f0c24da5ac3facec20be4098d7dbb09bb79d3c5a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 361, + 877, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1752c6d60a450c7f807a8c36fab32cf1714e2cda2324a9aad16b779aa3bf8654.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 425, + 877, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we perform style transfer on the third video on both background and object textures. In the first video, despite the complexity of the lighting conditions caused by the Disco ball, our approach can successfully handle and diffuse the light onto the edited region, resulting in high-quality output. Previous approaches fail to achieve such a representation of complex lighting conditions. In the second video, the color of the shirt changes convincingly according to the different lighting conditions in two different frames, while our edit on the background is successfully occluded by the shadow. We can see that in the third video, our edited reconstruction follows the textures and the modified lighting conditions fit seamlessly with our modifications, as shown in (d) of the fifth and sixth rows. In the supplementary material, we also provide additional editing results of various types to further demonstrate the versatility and effectiveness of our method.", + "bbox": [ + 75, + 564, + 472, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Quantitative Results for Editing Consistency", + "text_level": 1, + "bbox": [ + 76, + 816, + 454, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our work mainly focuses on consistent video editing. However, famous metrics such as PSNR (as we show in Fig. 4) or other reconstruction quality measures do not necessarily reflect the quality of consistent video editing. We", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "take advantage of TAP-Vid [8], which provides several evaluation metrics and the ground-truth feature tracking on the foreground object. We use the feature points of the first unobstructed frame $(f_{0,x}, f_{0,y})$ as our base point and convert it into texture coordinate $(u_0, v_0)$ . For all other frames $t$ , we select the video coordinate whose mapped texture coordinate is closest to $(u_0, v_0)$ as our prediction for frame $t$ . Therefore, the feature points on the same track should be mapped to the same coordinate on the foreground texture for an ideal rendering result.", + "bbox": [ + 496, + 564, + 893, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report our results in Table 3. We select a subset of the TAP-Vid DAVIS dataset for our quantitative analysis. For reference, we also report the results of RAFT [31], as a baseline, on the same videos since our optical flow supervision comes from it. In general, the correspondences obtained by our method are more consistent than those by RAFT, as RAFT considers only neighboring frames while ours works on a unified foreground representation. Specifically, our method achieves higher scores than RAFT on the \"black swan\" and \"kite-surf\" videos and comparable results on the other two videos. The low score on the \"parkour\" video is due to the complex geometric changes in the video, which", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7749", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b73fc783775dd47815a04cae05f9693cfef563bf3b6111d62f33985abbe2271.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodResolutionTraining timeRendering speed (fps)GPU memorybearPSNR disco ballhike
DeformableSprites [39]213 × 12010 minutes55 GB22.726.221.5
DeformableSprites [39]427 × 24020 minutes1.612 GB23.626.422.0
Layered Neural Atlases [16]768 × 4325.5 hours0.53 GB27.329.025.2
Ours768 × 43240 minutes7873 GB27.537.725.6
Ours1920 × 108040 minutes715 GB26.935.625.4
", + "bbox": [ + 104, + 87, + 867, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2bb273b5a38f0c502fb1047788a8f4f341f6222ccc1c63b9d53d4aef03dec1a7.jpg", + "table_caption": [ + "Table 1: Comparison results. We report the PSNR for each video, along with the corresponding training time, rendering speed (frames per second), and GPU memory usage under different resolutions. Our work achieves better results than prior work on the three videos, achieving faster rendering, lower GPU memory consumption, and higher resolution. Note that we measure the PSNR at the corresponding reconstructed resolution. Therefore the PSNR tends to favor the evaluation at low resolution, explaining why our 1080 reconstructions have slightly lower PSNR than our 480p reconstructions." + ], + "table_footnote": [], + "table_body": "
Methodbeardisco ballhike
PSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIM
DS23.60.230.7826.40.210.8922.00.290.68
LNA27.30.190.8529.00.110.9525.20.190.80
Ours27.50.160.8737.70.040.9725.60.160.82
kite-surfcar-turnlibby
DS21.20.340.6222.20.320.6521.60.410.57
LNA28.20.300.7627.50.300.8829.40.310.91
Ours29.20.260.7829.50.230.9129.60.260.92
", + "bbox": [ + 83, + 284, + 460, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3f4a5f0c7673ed548420afc8e0022835e09859ff8dbc818da3779a5a8725247b.jpg", + "table_caption": [ + "Table 2: More comparison results. Our model outperforms previous methods in all videos and achieves better scores in all evaluation metrics." + ], + "table_footnote": [], + "table_body": "
oursRAFT
AJ<δxavgOAAJ<δxavgOA
black swan0.810.871.000.390.521.00
parkour0.040.100.480.070.170.48
kite-surf0.480.580.990.250.350.86
cows0.510.610.990.550.650.99
", + "bbox": [ + 99, + 426, + 447, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Quantitative results on TAP-Vid. We test the editing consistency on a subset of TAP-Vid DAVIS and show the result for each video. The reported metrics, the higher the better, are average Jaccard (AJ), average position accuracy $(< \\delta_{avg}^{x})$ , and binary occlusion accuracy (OA). Compared to RAFT [31], our method performs better or reaches comparable results on all videos.", + "bbox": [ + 75, + 523, + 472, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our current architecture can reconstruct well but edit poorly on the foreground parkour runner. Nevertheless, the editing on the background still exhibits comparable quality.", + "bbox": [ + 75, + 656, + 468, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Manipulating Camera Motion", + "text_level": 1, + "bbox": [ + 76, + 710, + 344, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With our multiplicative residual estimator, we are able to synthesize realistic views that are not shown in the input video. A different camera view can be synthesized by scaling, shifting, or rotating the original video coordinates $(x,y)$ to get $(x^{\\prime},y^{\\prime})$ and concatenating it with the temporal information $t$ . Once the new set of video coordinates $p^\\prime = (x',y',t)$ is obtained, we can use the mapping network $\\mathcal{M}$ to obtain the texture coordinate $(u,v)^{(p^{\\prime})}$ for each layer. The color $c^{(p^{\\prime})}$ and the lighting $\\ell^{(p^{\\prime})}$ of all layers can be obtained based on these texture coordinates. For each non-background layer $n$ , we use the alpha network $\\mathcal{A}$ to", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a5ad7ee38f10641c2fb093b108f0874b6bcf2ab6dece28ab5521535df97851b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 285, + 630, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a3b85874d44b80ed5405f295a43e47b72724dcaa649fc7b1d5e962f23da02bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 354, + 629, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a15f369b591a546d3f3325ec9fb9cc764d1fe0d773ffd4767c79df8290996b62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 412, + 629, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f438d623a1eebd03388bb2bfe6b49c828b4686b3a649e18b3724b5ea21cc311f.jpg", + "image_caption": [ + "Figure 6: Camera motion manipulation. The multiplicative residual serves as a smooth representation that effectively adjusts the lighting conditions with a fixed camera." + ], + "image_footnote": [], + "bbox": [ + 501, + 467, + 629, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/055f2670cac109144b222d7f0888f383a01a40b022a7889ab940382b667f3bca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 286, + 890, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43b1ecd1e9d0619ba062869d67eca4955faf94f52cdd63d98ff7868a9a1df9a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 354, + 763, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0bb7635a0b3f3124267e2c5b3396642f45d810890762a7a38af23a503df0881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 412, + 763, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3eb9aed5d43499c6180a8626edc17f530e7738f9ff28f218d3a91ce6873d41e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 468, + 763, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1d1d33aff99057419d5712d79ff39759bdac132262538fa72201ae4fd80cea9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 354, + 890, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/953a793c7d37c79fbf8e991cf22fb308e1e981d6b1ca4e10bafcb1a712ee6c14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 412, + 890, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4b160f25385cca27b274242272a594d2f92206ebbd952be10d54933d3e9a0f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 468, + 890, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "obtain the original soft mask $\\alpha^{(p)}$ for each $p = (x,y,t)$ and establish a relation as $(u,v)_0\\rightarrow \\alpha_n^{(p)}$ if $(u,v)_0 = \\mathcal{M}_0(p)$ . Then, we linearly interpolate the value of the mask at each location $p^\\prime$ by triangulating the input data. That is, we establish a correspondence between the coordinates of the object textures and the background texture and then transfer the mask values from the object onto the background texture.", + "bbox": [ + 496, + 607, + 893, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We set up a virtual camera in the background to capture the entire video scene from a fixed position. Fig. 6 shows our result. The camera is fixed (i.e., the background remains static) while the object keeps moving in the scene. In the fourth video, the car is partially visible in the beginning frames and fully visible later, with its side facing the camera. The third video demonstrates that the splashes consistently follow the entire path of the original camera view. With our multiplicative-residual estimator, we can successfully reconstruct the surfing scene while accurately representing the water splashes. We show more results with varying camera motions in the supplementary material.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7750", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0d721cefee8d00bed3918942b5ad40d62f986f69e38140bf03cdcd7d115a456.jpg", + "image_caption": [ + "Figure 7: Choice of the residual estimator architecture. Using an additive residual estimator architecture may introduce artifacts or degrade the quality of textures, while a multiplicative one works correctly." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 467, + 362 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/062d31978c458d533b9e463af99bf2f657b04eeeeaf495762c75c788129f8754.jpg", + "image_caption": [ + "Figure 8: Quantitative ablation studies. (a) The reconstruction PSNR degradation of the additive residual is significantly higher than that of the multiplicative residual. (b) Hash grid encoding can achieve better PSNR in fewer iterations." + ], + "image_footnote": [], + "bbox": [ + 80, + 433, + 274, + 566 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0593c1673a69907154c2b783be94fd739c9271c943bf2d01221c7475ed67b902.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 434, + 460, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Ablations", + "text_level": 1, + "bbox": [ + 76, + 679, + 187, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Residual type. We present an ablation study to verify the choice of the architecture of our residual estimator. As mentioned in Sec. 3.3, we adopt the multiplicative residual instead of the additive residual. We show the comparisons of the two designs in Fig. 7 and Fig. 8a. The additive residual estimator is conducted, and the output is normalized to $[-1,1]$ to ensure they have the same representation capability. In Fig. 7, both additive and multiplicative residuals provide promising reconstruction results for the first video. However, when it comes to the complex texture, like the background of the second video, the additive residual fails to represent it and compensates for the wrong illuminations. In the third video, the additive residual fails to represent precise", + "bbox": [ + 75, + 703, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "background texture, and the dog vanishes from the texture. Fig. 8a demonstrates a significant drop in PSNR when ablating the residual module. This finding reflects that additive residual is prone to overfit the color that should be attributed to the editable texture. In contrast, the multiplicative residual provides complete background and object textures.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hash grid encoding. We evaluate the effectiveness of hash grid encoding in Fig. 8b. Positional encoding needs more MLP layers to achieve good quality, which takes $1.5 \\times$ processing time per iteration compared to hash encoding. However, hash grid encoding still achieves better PSNR in fewer iterations.", + "bbox": [ + 496, + 199, + 893, + 289 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 301, + 612, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have observed several interesting points that are worth discussing. Our method relies on Mask-RCNN and RAFT to provide the foreground mask and optical flow as external priors. As a result, our method may be biased by the inaccurate learning-based priors on some difficult cases like the spinning lighting in disco ball or the object boundary in bear. The reconstructed objects in the video may exhibit edge artifacts due to referencing the incorrect texture layer. Our method with reconstruction loss could improve the initial external prior but is still not perfect to solve all the artifacts. Adopting better external priors or designing strong internal priors are both good future improvements. The problem of object edge artifacts is similar to the seam artifact in mesh uv-parameterization from which future exploration may take inspiration.", + "bbox": [ + 496, + 327, + 893, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 566, + 617, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have presented a neural layer decomposition method to facilitate illumination-aware video editing. The proposed multiplicative-residual estimator can effectively derive the layered representation that characterizes the spatiotemporally varying lighting effects. We have also implemented hash grid encoding for fast coordinate inference. Our model, therefore, significantly reduces the training time and achieves real-time rendering speed with a low requirement of computation resources, enabling interactive editing on high-resolution videos. We use our model to generate high-quality video editing results, where, in particular, the varying illumination effect can only be achieved by ours rather than the previous methods.", + "bbox": [ + 496, + 590, + 893, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 800, + 678, + 816 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by NSTC grants 111-2221-E-001-011-MY2 and 112-2221-E-A49-100-MY3 of Taiwan. We are grateful to National Center for High-performance Computing for providing computational resources and facilities.", + "bbox": [ + 496, + 824, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7751", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aseem Agarwala, Aaron Hertzmann, David Salesin, and Steven M. Seitz. Keyframe-based tracking for rotoscoping and animation. ACM Trans. Graph., 23(3):584-591, 2004. 2", + "[2] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. In Computer Vision - ECCV 2022 Workshops - Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, volume 13802 of Lecture Notes in Computer Science, pages 204-220. Springer, 2022. 2", + "[3] Jean-Baptiste Alayrac, João Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5733-5742. IEEE, 2019. 2", + "[4] Connelly Barnes, Dan B. Goldman, Eli Shechtman, and Adam Finkelstein. Video tapestries with continuous temporal zoom. ACM Trans. Graph., 29(4):89:1-89:9, 2010. 2", + "[5] Zhangxing Bian, Allan Jabri, Alexei A. Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 6498-6509. IEEE, 2022. 2", + "[6] Gabriel J. Brostow and Irfan A. Essa. Motion based decompositing of video. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 8-13. IEEE Computer Society, 1999. 2", + "[7] Carlos D. Correa and Kwan-Liu Ma. Dynamic video narratives. ACM Trans. Graph., 29(4):88:1-88:9, 2010. 2", + "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurIPS Datasets Track, 2022. 5, 7", + "[9] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ArXiv, abs/2211.02145, 2022. 2", + "[10] Gregory D. Hager and Peter N. Belhumeur. Efficient region tracking with parametric models of geometry and illumination. IEEE Trans. Pattern Anal. Mach. Intell., 20(10):1025-1039, 1998. 2", + "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2020. 5", + "[12] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 7044-7052. IEEE Computer Society, 2017. 2", + "[13] Allan Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3154-3164. IEEE Computer Society, 2017. 2", + "[15] Nebojsa Jojic and Brendan J. Frey. Learning flexible sprites in video layers. In 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2001), with CD-ROM, 8-14 December 2001, Kauai, HI, USA, pages 199-206. IEEE Computer Society, 2001. 2", + "[16] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 1, 2, 3, 4, 5, 6, 8", + "[17] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. Int. J. Comput. Vis., 76(3):301-319, 2008. 2", + "[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XV, volume 11219 of Lecture Notes in Computer Science, pages 179-195. Springer, 2018. 2", + "[19] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 3809-3817. Computer Vision Foundation / IEEE, 2019. 2", + "[20] Sharon Lin, Matthew Fisher, Angela Dai, and Pat Hanrahan. Layerbuilder: Layer decomposition for interactive image and video color editing. ArXiv, abs/1701.03754, 2017. 2", + "[21] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. ACM Trans. Graph., 39(6):256:1-256:14, 2020. 2", + "[22] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4507-4515. Computer Vision Foundation / IEEE, 2021. 2", + "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 5", + "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5", + "[25] William S. Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei A. Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Or" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7752", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "leans, LA, USA, June 18-24, 2022, pages 13460-13471. IEEE, 2022. 2", + "[26] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus H. Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 724-732. IEEE Computer Society, 2016. 5, 6", + "[27] Hanspeter Pfister. Interactive intrinsic video editing. ACM Trans. Graph., 33(6):197:1-197:10, 2014. 2", + "[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew W. Fitzgibbon. Unwrap mosaics: a new representation for video editing. ACM Trans. Graph., 27(3):17, 2008. 2", + "[29] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Pattern Recognition - 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings, volume 9796 of Lecture Notes in Computer Science, pages 26-36. Springer, 2016. 2", + "[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 2288-2297. Computer Vision Foundation / IEEE, 2020. 2", + "[31] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12347 of Lecture Notes in Computer Science, pages 402-419. Springer, 2020. 7, 8", + "[32] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, SA 2022, Daegu, Republic of Korea, December 6-9, 2022, pages 29:1-29:9. ACM, 2022. 2", + "[33] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. IEEE Trans. Image Process., 3(5):625-638, 1994. 2", + "[34] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 29:9125-9139, 2020. 2", + "[35] Josh Wills, Sameer Agarwal, and Serge J. Belongie. What went where. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2003), 16-22 June 2003, Madison, WI, USA, pages 37-44. IEEE Computer Society, 2003. 2", + "[36] Ning Xu, Brian L. Price, Scott Cohen, and Thomas S. Huang. Deep image matting. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 311-320. IEEE Computer Society, 2017. 2", + "[37] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, volume 13675 of Lecture Notes in Computer Science, pages 357-374. Springer, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yanchao Yang, Brian Lai, and Stefano Soatto. Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 2826-2836. Computer Vision Foundation / IEEE, 2021. 2", + "[39] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 2647-2656. IEEE, 2022. 1, 2, 6, 8" + ], + "bbox": [ + 501, + 92, + 893, + 232 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7753", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_model.json b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9df1c241d2d0495df6bb44dbae57e1daa14aae6d --- /dev/null +++ b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.156 + ], + "angle": 0, + "content": "Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.181, + 0.307, + 0.2 + ], + "angle": 0, + "content": "Cheng-Hung Chan" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.182, + 0.49, + 0.199 + ], + "angle": 0, + "content": "Cheng-Yang Yuan" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.182, + 0.617, + 0.199 + ], + "angle": 0, + "content": "Cheng Sun" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.182, + 0.817, + 0.199 + ], + "angle": 0, + "content": "Hwann-Tzong Chen" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.2, + 0.641, + 0.217 + ], + "angle": 0, + "content": "National Tsing Hua University, Taiwan" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.253, + 0.236, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.254, + 0.386, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.394, + 0.254, + 0.534, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.359, + 0.534, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.301, + 0.566, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.254, + 0.721, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.254, + 0.868, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.359, + 0.719, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.359, + 0.869, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.446, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Figure 1: Integration of layer-based video editing and spatiotemporal lighting rendering. Our approach decomposes an input video into layers with multiplicative residuals that characterize complex spatiotemporal lighting variations. Our method efficiently fuses the user-edited components with the expected lighting conditions to produce high-quality video output." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.51, + 0.314, + 0.526 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.474, + 0.816 + ], + "angle": 0, + "content": "We present a video decomposition method that facilitates layer-based editing of videos with spatiotemporally varying lighting and motion effects. Our neural model decomposes an input video into multiple layered representations, each comprising a 2D texture map, a mask for the original video, and a multiplicative residual characterizing the spatiotemporal variations in lighting conditions. A single edit on the texture maps can be propagated to the corresponding locations in the entire video frames while preserving other contents' consistencies. Our method efficiently learns the layer-based neural representations of a 1080p video in 25s per frame via coordinate hashing and allows real-time rendering of the edited result at 71 fps on a single GPU. Qualitatively, we run our method on various videos to show its effectiveness in generating high-quality editing effects. Quantitatively, we propose to adopt feature-tracking evaluation metrics for objectively assessing the consistency of video editing. Project page: https://lightbulb12294.github.io/hashing-nvd/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.21, + 0.861 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Unlike image editing, video editing involves modeling the frame-to-frame relationships and addressing temporal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.895, + 0.633 + ], + "angle": 0, + "content": "variations such as motion and illumination changes. The task of video editing becomes challenging with the added dimension of time for a user who attempts to apply edits to a video while ensuring consistency across all frames. It is more convenient if we can handle the spatiotemporally varying components and reduce video editing to image editing—The user thus only needs to do edits on images with ease, and the editing results can seamlessly propagate to the entire video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To achieve this goal, we may consider incorporating effective representations that can model and reconstruct the static and dynamic information in videos. Furthermore, for practical concerns, the algorithm must be efficient enough in modeling and rendering to support interactive editing. Recent work on video decomposition has proposed to employ neural-based representations, such as layered neural atlases [16] and deformable sprites [39], to enable the conversion between the space-time video domain and 2D texture domain for editing. Despite their successes in showcasing impressive video editing effects, we notice that they often require longer training time or restrict to limited frame resolution. For example, it takes more than ten hours to derive the layered neural atlases [16] from a 100-frame 480p video. Deformable sprites [39] are relatively fast to derive but require more computation resources (over 24GB of GPU memory for a 480p video) and thus are unsuitable for editing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.882, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.169, + 0.895, + 0.199 + ], + "angle": 0, + "content": "Figure 2: The procedure of layer-based video editing. Our model allows users to apply edits to the extracted texture for rendering the edited video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.226, + 0.473, + 0.346 + ], + "angle": 0, + "content": "high-resolution videos. We address the computation issue by incorporating hash grid encoding [24] into our framework and achieve fast training and rendering for high-resolution videos. Moreover, we introduce the multiplicative residual estimator to model the lighting variations across video frames, which can improve the reconstruction quality and allow illumination-aware editing unachievable by prior work, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.348, + 0.47, + 0.362 + ], + "angle": 0, + "content": "We summarize the contributions of this work as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.364, + 0.47, + 0.455 + ], + "angle": 0, + "content": "1. This work is the first to consider spatiotemporally varying lighting effects for layered-based video editing. The proposed multiplicative-residual estimator effectively decomposes the lighting conditions from the video without supervision. Our method can produce better-quality videos by fusing the edits with expected illuminations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.472, + 0.61 + ], + "angle": 0, + "content": "2. Our approach is efficient in both training and rendering. Compared with prior work, the proposed method improves the training time with fewer resources and thus enables training on higher-resolution or longer videos. The trained model can achieve fast video rendering via hashing-based coordinate inference. It takes about 40 minutes to train with a 1080p video of 100 frames on a single 3090 Ti GPU. The inference speed for rendering an edited video is 71 fps for 1080p resolution, allowing real-time video editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.616, + 0.47, + 0.707 + ], + "angle": 0, + "content": "3. The experimental results demonstrate appealing video edits in various challenging contexts, such as modifying the texture of moving objects, handling occlusion, and manipulating camera motion, where all can be fused with vivid lighting and shading for better effects that are not easy to achieve by prior work." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.364, + 0.472, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.22, + 0.739 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Video decomposition. Decomposing a video into layered representations is a long-standing video analysis approach in computer vision since the seminal work by Wang & Adelson [33]. Similar ideas have been continually revisited under different contexts with the development of new techniques. A typical formulation is motion segmentation that decomposes the video by motion [6, 15, 17, 35]. Video matting, on the other hand, focuses more on separating the alpha matte of the foreground object from the background for blending [9, 22, 30, 36]. Recently, neural-based methods have" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.895, + 0.348 + ], + "angle": 0, + "content": "shown to be effective in estimating layered representations for video segmentation and video editing [3, 14, 16, 38, 39]. In this work, we also adopt neural networks to derive layered representations from videos. Further, inspired by the pioneering work in visual tracking for handling illumination changes [10], we incorporate a new multiplicative residual representation to model the lighting variations for illumination-aware video editing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.897, + 0.671 + ], + "angle": 0, + "content": "Video editing. Layered representations can benefit video editing in various ways. For example, the layered representations can serve as a visual proxy for intuitive video editing [16] or can be used to create the retiming effects [21]. Editing can be more easily done on a unified texture map built from the video's background, such as background mosaics [7], tapestries [4], and layered neural atlases [16]. Building a unified texture map for the foreground object is also useful, e.g., flexible sprites [15], unwrap mosaics [28], and deformable sprites [39]. Another type of decomposition is to derive temporally-consistent intrinsic components from videos [27, 20] so that the edits can be performed on the reflectance for recoloring or texture transfer. Recent deep-learning-based methods mainly address a single task of video editing, e.g., video style transfer [34, 19, 18, 12, 29]. or category-specific GAN-based video editing [2, 25, 32, 37]. Regarding achieving consistent video editing, it is crucial to know the temporal correspondences [13, 5] so typical animation techniques like Rotoscoping [1] can be applied." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.61, + 0.709 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We formulate the problem of consistent video decomposition and editing as follows: The goal is to decompose an input video into multiple layers, where one layer comprises the background, and each of the remaining layers represents a foreground object. We aim to facilitate user-friendly and intuitive editing of video components on time-independent static texture maps while ensuring the quality and consistency of the edited results in the output video. Fig. 2 illustrates an editing procedure. The model may rely on additional information obtained during pre-processing, such as backward and forward optical flow and rough object masks, to improve the training efficiency and consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7744" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.379, + 0.108 + ], + "angle": 0, + "content": "3.1. Layer-Based Video Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.474, + 0.388 + ], + "angle": 0, + "content": "A video of a scene typically contains multiple objects with distinct motions, appearances, shapes, and other attributes. Achieving consistent editing on the video scene and objects can be challenging. An efficient way to group objects is by decomposing the video into different motion layers. Static objects can be considered part of the scene's background, while only a few foreground objects move differently. Editing the objects on decomposed layers becomes more intuitive and straightforward since we can modify each object separately without touching the others. We decompose videos into layers using backward and forward optical flow and coarse object masks, resulting in \\( N + 1 \\) layers where layer \\( N \\) represents the background, and the rest is for foreground objects. Each layer \\( n \\) comprises a predicted mask \\( \\alpha_{n} \\) across different frames, a time-independent texture map \\( M_{n} \\), and the multiplicative residual \\( \\ell_{n,t} \\) characterizing the illumination for each frame \\( t \\) built on the texture map. Fig. 3 shows an overview of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.472, + 0.499 + ], + "angle": 0, + "content": "Layer hierarchy. Our alpha network \\(\\mathcal{A}\\) predicts the object mask probability for every layer. Rather than simply normalizing the layer probabilities of each pixel and directly using them as layer masks, we build a front-to-back hierarchy. That is, for layer \\(n\\), its probability of object mask is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.51, + 0.47, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {n} = a _ {n} \\cdot \\prod_ {i = n + 1} ^ {N} \\left(1 - a _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.673 + ], + "angle": 0, + "content": "where \\( a \\in [0,1] \\) is the raw output of \\( \\mathcal{A} \\) with \\( a_N \\) fixed and equal to 1. This hierarchy assumes that objects follow a strict front-to-back ordering, which also implies that layer \\( N \\) represents the background. Here, we specifically focus on scenes with only one foreground object. That is, \\( N = 1 \\). We show an example of decomposing multiple foreground objects in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.248, + 0.698 + ], + "angle": 0, + "content": "3.2. Texture Mapping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We associate each object with its own texture, such that modifying a single pixel in the texture will impact the object's appearance in the rendered video scene. We translate the spatiotemporal locations in the video into 2D UV coordinates for different layers by a mapping network \\(\\mathcal{M}\\). This allows us to sample colors on a texture with the translated coordinates. We follow the approach of coordinate-based neural rendering [16, 23] and use the multi-layer perceptrons as our coordinate translator instead of a textured grid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Note that the appearance of an object in the scene may change throughout the video due to camera motion and illumination variations. We will explain how we address this issue with the proposed multiplicative residual in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "The textures can be further obtained by grid-sampling the texture network \\(\\mathcal{T}\\). Once the texture is extracted, the texture network can be replaced by the extracted texture using UV coordinates from the mapping network, achieving the same effect via bilinear interpolation. This process enables users to modify and apply the extracted texture to the entire video." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.197, + 0.8, + 0.213 + ], + "angle": 0, + "content": "3.3. Multiplicative-Residual Estimator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.895, + 0.357 + ], + "angle": 0, + "content": "In order to apply a single modification to all video frames, the texture must remain constant throughout the entire video. However, the color at the same location on an object may change over time due to variations in lighting conditions, resulting in low reconstruction quality or noisy textures. To address this issue and ensure high-quality reconstruction, we have developed a new method, multiplicative-residual estimator, to achieve constant texture across all frames while preserving illumination variations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.895, + 0.496 + ], + "angle": 0, + "content": "With this design, we can also synthesize illumination coefficients not present in the original video at specific time points, allowing us to manipulate camera motion. Note that the prediction of the residual estimator is sharp. Therefore, we have designed losses to constrain the seen and unseen areas, as will be described in Sec. 3.5. The multiplicative-residual estimator predicts the illumination coefficient of each color channel on the texture map, which is multiplied by each channel to obtain the shaded color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.894, + 0.543 + ], + "angle": 0, + "content": "To provide more insight into our decision to use a multiplicative residual instead of an additive residual, we examine the process of diffuse shading:" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.559, + 0.892, + 0.575 + ], + "angle": 0, + "content": "\\[\nC = \\rho L, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.895, + 0.74 + ], + "angle": 0, + "content": "where \\(\\rho\\) is the diffuse albedo term determined by surface material and \\(L\\) is the lighting term determined by the environment. When the lighting condition changes or a shadow is cast onto the surface, the new shaded color then becomes \\(C' = \\rho \\cdot L'\\). With multiplicative residual, we can reproduce \\(C'\\) by scaling \\(C\\) with \\(\\frac{L'}{L}\\). In contrast, we have to add \\(C\\) by \\(\\rho (L' - L)\\) if we use additive residual. One can observe that the additive residual is more difficult to model as it also depends on the diffuse albedo term. We detail the practical result of the two choices in Sec. 4.6." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.755, + 0.707, + 0.77 + ], + "angle": 0, + "content": "3.4. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We design our model as an end-to-end framework trained via self-supervision. We take a spatiotemporal coordinate \\( p = (x, y, t) \\) in the video as the model input and predict the reconstructed color of the corresponding pixel. Fig. 3 shows the pipeline of our model, which includes a mapping network, a texture network, and a multiplicative-residual estimator for each layered representation, with a shared alpha network generating the soft masks for different layers." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7745" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.87, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.896, + 0.396 + ], + "angle": 0, + "content": "Figure 3: Model pipeline. Our model takes a video coordinate \\( p = (x, y, t) \\) as input and decomposes the video into multiple layers. A representation for layer \\( n \\) is modeled by a mapping network \\( \\mathcal{M} \\), a texture network \\( \\mathcal{T} \\), and a multiplicative-residual estimator \\( \\mathcal{R} \\); the three modules jointly convert a video coordinate \\( p \\) into layered color \\( c_{n}^{(p)} \\) and lighting coefficient \\( l_{n}^{(p)} \\). An alpha network \\( \\mathcal{A} \\) also takes \\( p \\) as input and predicts soft object masks \\( \\alpha^{(p)} \\) for each layer. The final video color is reconstructed by the bottom-right equation. The multiplicative residuals are critical to our model for handling illumination variations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.471, + 0.512 + ], + "angle": 0, + "content": "Mapping network. Following the design choice of recent coordinate-based approaches [16, 23], we build our network architecture with multi-layer perceptrons. For each layer \\( n \\), we pass a spatiotemporal coordinate \\( p \\) through a mapping network to translate the video coordinate into a time-independent 2D texture-map coordinate \\( (u, v) \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.516, + 0.47, + 0.535 + ], + "angle": 0, + "content": "\\[\n(u, v) _ {n} ^ {(p)} = \\mathcal {M} _ {n} (p), \\quad (u, v) _ {n} ^ {(p)} \\in [ - 1, 1 ] ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.541, + 0.384, + 0.556 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_n\\) is the mapping network of layer \\(n\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.472, + 0.68 + ], + "angle": 0, + "content": "Texture network and multiplicative-residual estimator. After obtaining the UV coordinates, the texture network takes the UV coordinates \\((u,v)_n^{(p)}\\) and produces the corresponding color at each position, while the multiplicative-residual estimator uses both the UV coordinates and the time \\(t\\) as inputs to predict the corresponding illumination coefficients at each position:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.687, + 0.404, + 0.711 + ], + "angle": 0, + "content": "\\[\nc _ {n} ^ {(p)} = \\mathcal {T} _ {n} \\left((u, v) _ {n} ^ {(p)}\\right), \\quad c _ {n} ^ {(p)} \\in [ 0, 1 ] ^ {3},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.712, + 0.405, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\ell_ {n, t} ^ {(p)} = \\mathcal {R} _ {n} \\left((u, v) _ {n} ^ {(p)}, t\\right), \\quad \\ell_ {n} ^ {(p)} \\in \\mathbb {R} _ {+} ^ {3},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.471, + 0.773 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}\\) is the texture network and \\(\\mathcal{R}\\) is the multiplicative-residual estimator." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.792, + 0.469, + 0.822 + ], + "angle": 0, + "content": "Alpha network. We also predict an object mask for each layer to indicate the layer to which each pixel belongs:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.826, + 0.47, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} ^ {(p)} = \\mathcal {A} (p), \\quad \\boldsymbol {\\alpha} ^ {(p)} \\in [ 0, 1 ] ^ {N + 1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.852, + 0.471, + 0.9 + ], + "angle": 0, + "content": "where \\(\\alpha_{n}^{(p)}\\) is the probability that pixel \\(p\\) belongs to layer \\(n\\), i.e., \\(\\sum_{n=0}^{N} \\alpha_{n}^{(p)} = 1\\). The derivation of the object mask is detailed in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.42, + 0.895, + 0.466 + ], + "angle": 0, + "content": "Pixel color reconstruction. After obtaining the information described above, we can reconstruct the pixel color at position \\( p \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.477, + 0.893, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\hat {c} ^ {(p)} = \\sum_ {n = 0} ^ {N} c _ {n} ^ {(p)} \\cdot \\ell_ {n} ^ {(p)} \\cdot \\alpha_ {n} ^ {(p)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.53, + 0.893, + 0.591 + ], + "angle": 0, + "content": "Inspired by [24], we use hash grid encoding for all networks except for the mapping network, as texture coordinates are expected to be smooth. This approach is adopted to facilitate better convergence during training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.601, + 0.625, + 0.615 + ], + "angle": 0, + "content": "3.5. Loss Terms" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.776 + ], + "angle": 0, + "content": "Losses inherited from previous work. We incorporate several losses from Layered Neural Atlases [16] to facilitate the training process. These losses include 1) optical flow loss, which provides the model with the supervision of optical flow, and 2) alpha bootstrapping loss, which offers supervision of initial coarse masks; additionally, we use 3) reconstruction loss to stabilize the training and improve reconstruction quality, and 4) sparsity loss to avoid duplicate foreground objects in the texture area. Further details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Residual consistency loss. In order to produce significant yet smooth lighting conditions, the multiplicative residuals of different times \\( t_1 \\) and \\( t_2 \\) of the same position on the texture coordinate \\( (u, v) \\), where \\( (u, v, t_1) \\) corresponds to a visible area and \\( (u, v, t_2) \\) corresponds to an invisible area, should be close in distribution and can have different intensities of illumination. Formally, we have a small \\( k \\times k \\) patch \\( P \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7746" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "at time \\( t_1 \\) on the video centered at \\( (x, y) \\). We then sample their texture coordinates \\( (u, v)^{(P)} = \\mathcal{M}(P) \\). The texture coordinates are then combined with different times \\( t_1 \\) and \\( t_2 \\) to get the corresponding multiplicative residuals:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.162, + 0.47, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\ell^ {(P)} = \\mathcal {R} ((u, v) ^ {(P)}, t _ {1}), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.19, + 0.364, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\ell^ {\\prime (P)} = \\mathcal {R} \\left((u, v) ^ {(P)}, t _ {2}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.469, + 0.283 + ], + "angle": 0, + "content": "To this end, we encourage multiplicative residuals of the same position on texture coordinate should be positively correlated. Therefore, we define the residual consistency loss through normalized cross-correlation by" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.292, + 0.47, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\psi (P, t _ {2}) = \\frac {\\left(\\ell^ {(P)} - \\mu_ {\\ell (P)}\\right) \\left(\\ell^ {\\prime (P)} - \\mu_ {\\ell^ {\\prime} (P)}\\right)}{\\sigma_ {\\ell (P)} \\sigma_ {\\ell^ {\\prime} (P)}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.366 + ], + "angle": 0, + "content": "We further introduce a variance-smoothness term that smoothes the changes in lighting conditions as" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.376, + 0.47, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} (P, t _ {2}) = \\sigma_ {\\ell^ {\\prime} (P)} ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.468, + 0.432 + ], + "angle": 0, + "content": "We apply the above terms to all layers and get the overall loss as" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.434, + 0.47, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {R} \\text {c o n}} = \\lambda_ {\\mathcal {R} \\text {c o n}} (\\psi + \\beta \\mathbb {E}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.469, + 0.484 + ], + "angle": 0, + "content": "where we choose \\( k = 3 \\) since our patch size is \\( 3 \\times 3 \\) and \\( \\beta = 16 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Residual regularization. Minimizing the current losses can collapse to a trivial solution because the multiplicative-residual estimator absorbs all colors. We regularize the residuals equal to 1 since \"true color\" of an object in the video depends on the overall light conditions. That is, if the light source is a pure blue light in the entire video shining on a white wall, we can only say that the steady color of the wall is \"blue\", as we do not have enough information to determine the true color of the wall." + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.649, + 0.47, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {R} \\text {r e g}} = \\lambda_ {\\mathcal {R} \\text {r e g}} | | \\mathcal {R} (\\cdot) - 1 | | _ {2} ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Alpha regularization. We introduce an additional regularization term to address potential issues with the alpha network. Specifically, the sparsity loss used in the alpha network tends to assign a value of zero (i.e., black) to regions that are not visible in the input video, which can result in noisy masks and incorrect shadow embedding. To mitigate this problem, we enforced a constraint that each pixel should contribute to at most one layer. This ensures that the mask for each layer is clean and reliable and that the lighting conditions are properly embedded in each layer." + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.844, + 0.47, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\alpha \\operatorname {r e g}} = \\lambda_ {\\alpha \\operatorname {r e g}} \\mathrm {B C E} \\left(\\max _ {n \\in \\{0, \\dots , N \\}} \\alpha_ {n}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.312, + 0.901 + ], + "angle": 0, + "content": "where BCE is binary cross entropy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.696, + 0.108 + ], + "angle": 0, + "content": "3.6. Hash Grid Encoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.372 + ], + "angle": 0, + "content": "Like [16, 23], we have considered employing positional encoding as the input encoding for the multi-layer perceptrons. The purpose of including the positional encoding is to enrich the discriminative power of the coordinates and ensure that high-frequency details can be adequately represented. Inspired by [24], we choose to adopt hash grid encoding as the input encoding method for our model. By leveraging this technique, the input features are encoded into a set of intermediate representations that span a broad spectrum of spatiotemporal resolutions, ranging from coarse-to fine-grained, and can flexibly adjust to distinct regions while maintaining high levels of accuracy and consistency. Formally, the 2D or 3D input feature is treated as a coordinate of multi-resolution grids and then used to sample data from the grids using interpolation techniques. The resulting sampled data are concatenated and passed into multi-layer perceptrons to obtain the final output." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.633, + 0.404 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.894, + 0.563 + ], + "angle": 0, + "content": "We conduct our experiments on the DAVIS dataset [26] and various internet videos to demonstrate the effectiveness of our approach in video reconstruction and consistent video editing. We also design an evaluation metric of edit consistency on the TAP-Vid-DAVIS dataset [8]. In addition, we use our method to generate videos with different camera motions from those in the original input video. Furthermore, we conduct ablation studies on the multiplicative-residual estimator and encoding type to assess the impact on the performance of different architecture choices." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.683, + 0.59 + ], + "angle": 0, + "content": "4.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.894, + 0.734 + ], + "angle": 0, + "content": "Fig. 4 shows four qualitative results of our work on the DAVIS dataset [26]. For each video frame, we show the predicted masks (second and third columns), the m-residuals (fourth column), the corresponding layered textures (fifth and sixth columns), and the reconstruction result (last column). Please also refer to Fig. 1 for an example result highlighting the robustness of our m-residual estimation. More results and higher resolution reconstruction can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.885 + ], + "angle": 0, + "content": "With merely the guidance from the coarse object mask provided by Mask R-CNN [11], our model successfully separates the object from the background yielding a clean and precise mask. In the second video in Fig. 4, we showcase an example of a video occluded by complex water splashes. Our model can accurately reconstruct the video without introducing noticeable texture distortions. We superimpose a rainbow checkerboard on the mask to visualize the transformation of each layer. It is evident that each coordinate's component remains at the same position in the texture." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since the multiplicative residual may brighten or darken" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7747" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.874, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.466, + 0.895, + 0.497 + ], + "angle": 0, + "content": "Figure 4: Qualitative results on the DAVIS dataset. We show our predicted masks, textures, m-residuals, reconstructed results, and PSNR of different videos. We overlay a color checkerboard on the masked areas to visualize the texture transformations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.473, + 0.706 + ], + "angle": 0, + "content": "the texture, we display the m-residual output on a gray canvas to visualize its effect. In the first video, our m-residual estimator successfully models the shadow of the hiker while the background texture retains its original color. The m-residual estimator takes both the texture coordinate and time as the input, allowing it to learn how to model the changes in lighting conditions through space-time and adjust the color accordingly. Such a mechanism enables our model to generate realistic and consistent shading effects for different objects in the video, such as the shading on the bear's fur or the reflection on the hiker's hat. We also report the PSNR of each video as evidence of high-quality reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.367, + 0.728 + ], + "angle": 0, + "content": "4.2. Comparison with Previous Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We report the PSNR values and various settings for every approach in Table 1. We also provide more quantitative comparisons with DeformableSprites (DS) [39] and Layered Neural Atlases (LNA) [16] under multiple metrics in Table 2. Our method demonstrates superior performance of PSNR, rendering speed, and GPU memory consumption compared to previous methods on all three videos while maintaining the same or better resolution. Note that we could not run DeformableSprites for the \\(768 \\times 432\\) resolution, as DeformableSprites failed to finish due to their GPU memory usage exceeding the limit of a single GPU with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.895, + 0.676 + ], + "angle": 0, + "content": "24 GB memory. In contrast, our method can efficiently decompose higher-resolution videos into full-HD layers using affordable computation resources while maintaining comparable reconstruction results. To compute the PSNR, we scale the input video up to the reconstructed resolution. Therefore, the PSNR of our 1080p model might be slightly lower than that of the \\(768 \\times 432\\) model, as the PSNR of the 1080p model is calculated at the corresponding resolution and considered harder due to having more details to reconstruct than the low-resolution counterpart." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.681, + 0.729, + 0.699 + ], + "angle": 0, + "content": "4.3. Consistent Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.796 + ], + "angle": 0, + "content": "For editing purposes, we use a \\(1000 \\times 1000\\) grid sampling to render the texture networks. We then modify the rendered texture and sample color via bilinear interpolation. As our key advantage lies in the m-residual representation of illumination, we have focused our editing on videos or components that exhibit changes in lighting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.896, + 0.903 + ], + "angle": 0, + "content": "We show our editing results in Fig. 5. We choose three videos to present the rendering quality. The first video is sourced from internet videos, while the other two are taken from the DAVIS dataset [26]. For evaluating the editing quality, we modify the background texture of the first two videos by adding handwritten characters of \"ICCV\". We also adjust the color of the shirt of the hiker in the second video." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7748" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.092, + 0.218, + 0.105 + ], + "angle": 0, + "content": "(a) Input frames" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.106, + 0.236, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.17, + 0.236, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.234, + 0.236, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.298, + 0.236, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.362, + 0.236, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.425, + 0.236, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.092, + 0.374, + 0.104 + ], + "angle": 0, + "content": "(b) Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.106, + 0.384, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.17, + 0.384, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.234, + 0.384, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.298, + 0.384, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.362, + 0.384, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.425, + 0.384, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.092, + 0.548, + 0.106 + ], + "angle": 0, + "content": "(c) Texture editing" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.131, + 0.574, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.235, + 0.582, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.31, + 0.564, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.363, + 0.563, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.425, + 0.563, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.092, + 0.822, + 0.106 + ], + "angle": 0, + "content": "(d) Reconstruction (edited)" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.106, + 0.731, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.17, + 0.731, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.234, + 0.731, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.298, + 0.731, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.362, + 0.731, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.425, + 0.731, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.106, + 0.879, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.17, + 0.879, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.234, + 0.879, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.298, + 0.879, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.362, + 0.879, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.426, + 0.879, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.493, + 0.894, + 0.539 + ], + "angle": 0, + "content": "Figure 5: Edits on videos with varying lighting conditions. We apply various editing techniques on different components of the videos to observe the corresponding changes in the lighting conditions. The textures with edits are shown in (c), while the editing results are enlarged in the right column of (d). The lighting and shading fuse with the edits flawlessly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.473, + 0.807 + ], + "angle": 0, + "content": "Finally, we perform style transfer on the third video on both background and object textures. In the first video, despite the complexity of the lighting conditions caused by the Disco ball, our approach can successfully handle and diffuse the light onto the edited region, resulting in high-quality output. Previous approaches fail to achieve such a representation of complex lighting conditions. In the second video, the color of the shirt changes convincingly according to the different lighting conditions in two different frames, while our edit on the background is successfully occluded by the shadow. We can see that in the third video, our edited reconstruction follows the textures and the modified lighting conditions fit seamlessly with our modifications, as shown in (d) of the fifth and sixth rows. In the supplementary material, we also provide additional editing results of various types to further demonstrate the versatility and effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.455, + 0.834 + ], + "angle": 0, + "content": "4.4. Quantitative Results for Editing Consistency" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Our work mainly focuses on consistent video editing. However, famous metrics such as PSNR (as we show in Fig. 4) or other reconstruction quality measures do not necessarily reflect the quality of consistent video editing. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.717 + ], + "angle": 0, + "content": "take advantage of TAP-Vid [8], which provides several evaluation metrics and the ground-truth feature tracking on the foreground object. We use the feature points of the first unobstructed frame \\((f_{0,x}, f_{0,y})\\) as our base point and convert it into texture coordinate \\((u_0, v_0)\\). For all other frames \\(t\\), we select the video coordinate whose mapped texture coordinate is closest to \\((u_0, v_0)\\) as our prediction for frame \\(t\\). Therefore, the feature points on the same track should be mapped to the same coordinate on the foreground texture for an ideal rendering result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We report our results in Table 3. We select a subset of the TAP-Vid DAVIS dataset for our quantitative analysis. For reference, we also report the results of RAFT [31], as a baseline, on the same videos since our optical flow supervision comes from it. In general, the correspondences obtained by our method are more consistent than those by RAFT, as RAFT considers only neighboring frames while ours works on a unified foreground representation. Specifically, our method achieves higher scores than RAFT on the \"black swan\" and \"kite-surf\" videos and comparable results on the other two videos. The low score on the \"parkour\" video is due to the complex geometric changes in the video, which" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7749" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.105, + 0.088, + 0.868, + 0.199 + ], + "angle": 0, + "content": "
MethodResolutionTraining timeRendering speed (fps)GPU memorybearPSNR disco ballhike
DeformableSprites [39]213 × 12010 minutes55 GB22.726.221.5
DeformableSprites [39]427 × 24020 minutes1.612 GB23.626.422.0
Layered Neural Atlases [16]768 × 4325.5 hours0.53 GB27.329.025.2
Ours768 × 43240 minutes7873 GB27.537.725.6
Ours1920 × 108040 minutes715 GB26.935.625.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.203, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Table 1: Comparison results. We report the PSNR for each video, along with the corresponding training time, rendering speed (frames per second), and GPU memory usage under different resolutions. Our work achieves better results than prior work on the three videos, achieving faster rendering, lower GPU memory consumption, and higher resolution. Note that we measure the PSNR at the corresponding reconstructed resolution. Therefore the PSNR tends to favor the evaluation at low resolution, explaining why our 1080 reconstructions have slightly lower PSNR than our 480p reconstructions." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.285, + 0.462, + 0.38 + ], + "angle": 0, + "content": "
Methodbeardisco ballhike
PSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIM
DS23.60.230.7826.40.210.8922.00.290.68
LNA27.30.190.8529.00.110.9525.20.190.80
Ours27.50.160.8737.70.040.9725.60.160.82
kite-surfcar-turnlibby
DS21.20.340.6222.20.320.6521.60.410.57
LNA28.20.300.7627.50.300.8829.40.310.91
Ours29.20.260.7829.50.230.9129.60.260.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.426 + ], + "angle": 0, + "content": "Table 2: More comparison results. Our model outperforms previous methods in all videos and achieves better scores in all evaluation metrics." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.428, + 0.449, + 0.522 + ], + "angle": 0, + "content": "
oursRAFT
AJ<δxavgOAAJ<δxavgOA
black swan0.810.871.000.390.521.00
parkour0.040.100.480.070.170.48
kite-surf0.480.580.990.250.350.86
cows0.510.610.990.550.650.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.525, + 0.473, + 0.631 + ], + "angle": 0, + "content": "Table 3: Quantitative results on TAP-Vid. We test the editing consistency on a subset of TAP-Vid DAVIS and show the result for each video. The reported metrics, the higher the better, are average Jaccard (AJ), average position accuracy \\((< \\delta_{avg}^{x})\\), and binary occlusion accuracy (OA). Compared to RAFT [31], our method performs better or reaches comparable results on all videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.704 + ], + "angle": 0, + "content": "our current architecture can reconstruct well but edit poorly on the foreground parkour runner. Nevertheless, the editing on the background still exhibits comparable quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.345, + 0.728 + ], + "angle": 0, + "content": "4.5. Manipulating Camera Motion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "With our multiplicative residual estimator, we are able to synthesize realistic views that are not shown in the input video. A different camera view can be synthesized by scaling, shifting, or rotating the original video coordinates \\((x,y)\\) to get \\((x^{\\prime},y^{\\prime})\\) and concatenating it with the temporal information \\(t\\). Once the new set of video coordinates \\(p^\\prime = (x',y',t)\\) is obtained, we can use the mapping network \\(\\mathcal{M}\\) to obtain the texture coordinate \\((u,v)^{(p^{\\prime})}\\) for each layer. The color \\(c^{(p^{\\prime})}\\) and the lighting \\(\\ell^{(p^{\\prime})}\\) of all layers can be obtained based on these texture coordinates. For each non-background layer \\(n\\), we use the alpha network \\(\\mathcal{A}\\) to" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.286, + 0.631, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.356, + 0.63, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.413, + 0.63, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.468, + 0.63, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.287, + 0.892, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.356, + 0.764, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.413, + 0.764, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.469, + 0.764, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.356, + 0.892, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.413, + 0.892, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.529, + 0.895, + 0.575 + ], + "angle": 0, + "content": "Figure 6: Camera motion manipulation. The multiplicative residual serves as a smooth representation that effectively adjusts the lighting conditions with a fixed camera." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.895, + 0.717 + ], + "angle": 0, + "content": "obtain the original soft mask \\(\\alpha^{(p)}\\) for each \\(p = (x,y,t)\\) and establish a relation as \\((u,v)_0\\rightarrow \\alpha_n^{(p)}\\) if \\((u,v)_0 = \\mathcal{M}_0(p)\\). Then, we linearly interpolate the value of the mask at each location \\(p^\\prime\\) by triangulating the input data. That is, we establish a correspondence between the coordinates of the object textures and the background texture and then transfer the mask values from the object onto the background texture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We set up a virtual camera in the background to capture the entire video scene from a fixed position. Fig. 6 shows our result. The camera is fixed (i.e., the background remains static) while the object keeps moving in the scene. In the fourth video, the car is partially visible in the beginning frames and fully visible later, with its side facing the camera. The third video demonstrates that the splashes consistently follow the entire path of the original camera view. With our multiplicative-residual estimator, we can successfully reconstruct the surfing scene while accurately representing the water splashes. We show more results with varying camera motions in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7750" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.09, + 0.468, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.366, + 0.472, + 0.426 + ], + "angle": 0, + "content": "Figure 7: Choice of the residual estimator architecture. Using an additive residual estimator architecture may introduce artifacts or degrade the quality of textures, while a multiplicative one works correctly." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.434, + 0.276, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.435, + 0.462, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.575, + 0.472, + 0.649 + ], + "angle": 0, + "content": "Figure 8: Quantitative ablation studies. (a) The reconstruction PSNR degradation of the additive residual is significantly higher than that of the multiplicative residual. (b) Hash grid encoding can achieve better PSNR in fewer iterations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.188, + 0.695 + ], + "angle": 0, + "content": "4.6. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Residual type. We present an ablation study to verify the choice of the architecture of our residual estimator. As mentioned in Sec. 3.3, we adopt the multiplicative residual instead of the additive residual. We show the comparisons of the two designs in Fig. 7 and Fig. 8a. The additive residual estimator is conducted, and the output is normalized to \\([-1,1]\\) to ensure they have the same representation capability. In Fig. 7, both additive and multiplicative residuals provide promising reconstruction results for the first video. However, when it comes to the complex texture, like the background of the second video, the additive residual fails to represent it and compensates for the wrong illuminations. In the third video, the additive residual fails to represent precise" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.183 + ], + "angle": 0, + "content": "background texture, and the dog vanishes from the texture. Fig. 8a demonstrates a significant drop in PSNR when ablating the residual module. This finding reflects that additive residual is prone to overfit the color that should be attributed to the editable texture. In contrast, the multiplicative residual provides complete background and object textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Hash grid encoding. We evaluate the effectiveness of hash grid encoding in Fig. 8b. Positional encoding needs more MLP layers to achieve good quality, which takes \\(1.5 \\times\\) processing time per iteration compared to hash encoding. However, hash grid encoding still achieves better PSNR in fewer iterations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.613, + 0.318 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.895, + 0.553 + ], + "angle": 0, + "content": "We have observed several interesting points that are worth discussing. Our method relies on Mask-RCNN and RAFT to provide the foreground mask and optical flow as external priors. As a result, our method may be biased by the inaccurate learning-based priors on some difficult cases like the spinning lighting in disco ball or the object boundary in bear. The reconstructed objects in the video may exhibit edge artifacts due to referencing the incorrect texture layer. Our method with reconstruction loss could improve the initial external prior but is still not perfect to solve all the artifacts. Adopting better external priors or designing strong internal priors are both good future improvements. The problem of object edge artifacts is similar to the seam artifact in mesh uv-parameterization from which future exploration may take inspiration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.619, + 0.582 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.895, + 0.789 + ], + "angle": 0, + "content": "We have presented a neural layer decomposition method to facilitate illumination-aware video editing. The proposed multiplicative-residual estimator can effectively derive the layered representation that characterizes the spatiotemporally varying lighting effects. We have also implemented hash grid encoding for fast coordinate inference. Our model, therefore, significantly reduces the training time and achieves real-time rendering speed with a low requirement of computation resources, enabling interactive editing on high-resolution videos. We use our model to generate high-quality video editing results, where, in particular, the varying illumination effect can only be achieved by ours rather than the previous methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.679, + 0.817 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by NSTC grants 111-2221-E-001-011-MY2 and 112-2221-E-A49-100-MY3 of Taiwan. We are grateful to National Center for High-performance Computing for providing computational resources and facilities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7751" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Aseem Agarwala, Aaron Hertzmann, David Salesin, and Steven M. Seitz. Keyframe-based tracking for rotoscoping and animation. ACM Trans. Graph., 23(3):584-591, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. In Computer Vision - ECCV 2022 Workshops - Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, volume 13802 of Lecture Notes in Computer Science, pages 204-220. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, João Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5733-5742. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.472, + 0.381 + ], + "angle": 0, + "content": "[4] Connelly Barnes, Dan B. Goldman, Eli Shechtman, and Adam Finkelstein. Video tapestries with continuous temporal zoom. ACM Trans. Graph., 29(4):89:1-89:9, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.472, + 0.451 + ], + "angle": 0, + "content": "[5] Zhangxing Bian, Allan Jabri, Alexei A. Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 6498-6509. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.472, + 0.507 + ], + "angle": 0, + "content": "[6] Gabriel J. Brostow and Irfan A. Essa. Motion based decompositing of video. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 8-13. IEEE Computer Society, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.472, + 0.536 + ], + "angle": 0, + "content": "[7] Carlos D. Correa and Kwan-Liu Ma. Dynamic video narratives. ACM Trans. Graph., 29(4):88:1-88:9, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.472, + 0.605 + ], + "angle": 0, + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurIPS Datasets Track, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.472, + 0.647 + ], + "angle": 0, + "content": "[9] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ArXiv, abs/2211.02145, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.703 + ], + "angle": 0, + "content": "[10] Gregory D. Hager and Peter N. Belhumeur. Efficient region tracking with parametric models of geometry and illumination. IEEE Trans. Pattern Anal. Mach. Intell., 20(10):1025-1039, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[12] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 7044-7052. IEEE Computer Society, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Allan Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[14] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3154-3164. IEEE Computer Society, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[15] Nebojsa Jojic and Brendan J. Frey. Learning flexible sprites in video layers. In 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2001), with CD-ROM, 8-14 December 2001, Kauai, HI, USA, pages 199-206. IEEE Computer Society, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[16] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[17] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. Int. J. Comput. Vis., 76(3):301-319, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XV, volume 11219 of Lecture Notes in Computer Science, pages 179-195. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.894, + 0.501 + ], + "angle": 0, + "content": "[19] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 3809-3817. Computer Vision Foundation / IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.893, + 0.546 + ], + "angle": 0, + "content": "[20] Sharon Lin, Matthew Fisher, Angela Dai, and Pat Hanrahan. Layerbuilder: Layer decomposition for interactive image and video color editing. ArXiv, abs/1701.03754, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[21] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. ACM Trans. Graph., 39(6):256:1-256:14, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.606, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[22] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4507-4515. Computer Vision Foundation / IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[25] William S. Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei A. Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Or" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7752" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "leans, LA, USA, June 18-24, 2022, pages 13460-13471. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[26] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus H. Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 724-732. IEEE Computer Society, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.471, + 0.246 + ], + "angle": 0, + "content": "[27] Hanspeter Pfister. Interactive intrinsic video editing. ACM Trans. Graph., 33(6):197:1-197:10, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew W. Fitzgibbon. Unwrap mosaics: a new representation for video editing. ACM Trans. Graph., 27(3):17, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[29] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Pattern Recognition - 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings, volume 9796 of Lecture Notes in Computer Science, pages 26-36. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 2288-2297. Computer Vision Foundation / IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.524 + ], + "angle": 0, + "content": "[31] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12347 of Lecture Notes in Computer Science, pages 402-419. Springer, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[32] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, SA 2022, Daegu, Republic of Korea, December 6-9, 2022, pages 29:1-29:9. ACM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.636 + ], + "angle": 0, + "content": "[33] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. IEEE Trans. Image Process., 3(5):625-638, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.472, + 0.678 + ], + "angle": 0, + "content": "[34] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 29:9125-9139, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.472, + 0.748 + ], + "angle": 0, + "content": "[35] Josh Wills, Sameer Agarwal, and Serge J. Belongie. What went where. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2003), 16-22 June 2003, Madison, WI, USA, pages 37-44. IEEE Computer Society, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.472, + 0.817 + ], + "angle": 0, + "content": "[36] Ning Xu, Brian L. Price, Scott Cohen, and Thomas S. Huang. Deep image matting. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 311-320. IEEE Computer Society, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[37] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, volume 13675 of Lecture Notes in Computer Science, pages 357-374. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[38] Yanchao Yang, Brian Lai, and Stefano Soatto. Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 2826-2836. Computer Vision Foundation / IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[39] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 2647-2656. IEEE, 2022. 1, 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7753" + } + ] +] \ No newline at end of file diff --git a/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_origin.pdf b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c409d93afd2f747f956e4c37b49281e322cc014d --- /dev/null +++ b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/4db886e2-61ae-4d82-b252-4c366d1ec299_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7891fac5b57bff1c8dd7f8f55357b24cccedd3f6588a8d3c0231d6792635300c +size 5298718 diff --git a/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/full.md b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02c424e1b15a861a05720f498f276d806990c861 --- /dev/null +++ b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/full.md @@ -0,0 +1,417 @@ +# Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time + +Cheng-Hung Chan + +Cheng-Yang Yuan + +Cheng Sun + +Hwann-Tzong Chen + +National Tsing Hua University, Taiwan + +![](images/c3cc7f33312a90cdd9da9e53db670d91daac815c5923e8cb1dc3b49f9c9070c7.jpg) +Figure 1: Integration of layer-based video editing and spatiotemporal lighting rendering. Our approach decomposes an input video into layers with multiplicative residuals that characterize complex spatiotemporal lighting variations. Our method efficiently fuses the user-edited components with the expected lighting conditions to produce high-quality video output. + +![](images/002359bd18af5c95ef2d281d022cba4e63e98a75f708341649fdd59a35d6e294.jpg) + +![](images/e4956e2e1ffd1221174969f5fec8724be071d64a10cc2977ccf1de1549f977fd.jpg) + +![](images/f6adb6c3653031b16414dd3a87c4d4cf4b90494c458bbecd60a3e55bd76d37cf.jpg) + +![](images/80428b5481abd1bd9b3b96b894718951b76ad45ef292400f42496a13927cea92.jpg) + +![](images/c65c1c7ca37cb7538f7cb9b86c1740058371ec56b0026e3257f5ec18e373ecbe.jpg) + +![](images/75937d4f69df8d5c0db497160b6b17400322bd2df6f2c14e38376608e89462c2.jpg) + +![](images/51036f3ea5921360b0e770ece302c0c5eaa113558382df0ddadf65e2df9f6712.jpg) + +![](images/3920fd85b084ed3db956e77d933dd5cbea64d1e064d3192c93c86beea99623e7.jpg) + +# Abstract + +We present a video decomposition method that facilitates layer-based editing of videos with spatiotemporally varying lighting and motion effects. Our neural model decomposes an input video into multiple layered representations, each comprising a 2D texture map, a mask for the original video, and a multiplicative residual characterizing the spatiotemporal variations in lighting conditions. A single edit on the texture maps can be propagated to the corresponding locations in the entire video frames while preserving other contents' consistencies. Our method efficiently learns the layer-based neural representations of a 1080p video in 25s per frame via coordinate hashing and allows real-time rendering of the edited result at 71 fps on a single GPU. Qualitatively, we run our method on various videos to show its effectiveness in generating high-quality editing effects. Quantitatively, we propose to adopt feature-tracking evaluation metrics for objectively assessing the consistency of video editing. Project page: https://lightbulb12294.github.io/hashing-nvd/ + +# 1. Introduction + +Unlike image editing, video editing involves modeling the frame-to-frame relationships and addressing temporal + +variations such as motion and illumination changes. The task of video editing becomes challenging with the added dimension of time for a user who attempts to apply edits to a video while ensuring consistency across all frames. It is more convenient if we can handle the spatiotemporally varying components and reduce video editing to image editing—The user thus only needs to do edits on images with ease, and the editing results can seamlessly propagate to the entire video. + +To achieve this goal, we may consider incorporating effective representations that can model and reconstruct the static and dynamic information in videos. Furthermore, for practical concerns, the algorithm must be efficient enough in modeling and rendering to support interactive editing. Recent work on video decomposition has proposed to employ neural-based representations, such as layered neural atlases [16] and deformable sprites [39], to enable the conversion between the space-time video domain and 2D texture domain for editing. Despite their successes in showcasing impressive video editing effects, we notice that they often require longer training time or restrict to limited frame resolution. For example, it takes more than ten hours to derive the layered neural atlases [16] from a 100-frame 480p video. Deformable sprites [39] are relatively fast to derive but require more computation resources (over 24GB of GPU memory for a 480p video) and thus are unsuitable for editing + +![](images/0d7757873dccaf71565c88b0f5ff4612fb81c20cd4c3d8e6caac04250109bb40.jpg) +Figure 2: The procedure of layer-based video editing. Our model allows users to apply edits to the extracted texture for rendering the edited video. + +high-resolution videos. We address the computation issue by incorporating hash grid encoding [24] into our framework and achieve fast training and rendering for high-resolution videos. Moreover, we introduce the multiplicative residual estimator to model the lighting variations across video frames, which can improve the reconstruction quality and allow illumination-aware editing unachievable by prior work, as shown in Fig. 1. + +We summarize the contributions of this work as follows: + +1. This work is the first to consider spatiotemporally varying lighting effects for layered-based video editing. The proposed multiplicative-residual estimator effectively decomposes the lighting conditions from the video without supervision. Our method can produce better-quality videos by fusing the edits with expected illuminations. +2. Our approach is efficient in both training and rendering. Compared with prior work, the proposed method improves the training time with fewer resources and thus enables training on higher-resolution or longer videos. The trained model can achieve fast video rendering via hashing-based coordinate inference. It takes about 40 minutes to train with a 1080p video of 100 frames on a single 3090 Ti GPU. The inference speed for rendering an edited video is 71 fps for 1080p resolution, allowing real-time video editing. +3. The experimental results demonstrate appealing video edits in various challenging contexts, such as modifying the texture of moving objects, handling occlusion, and manipulating camera motion, where all can be fused with vivid lighting and shading for better effects that are not easy to achieve by prior work. + +# 2. Related Work + +Video decomposition. Decomposing a video into layered representations is a long-standing video analysis approach in computer vision since the seminal work by Wang & Adelson [33]. Similar ideas have been continually revisited under different contexts with the development of new techniques. A typical formulation is motion segmentation that decomposes the video by motion [6, 15, 17, 35]. Video matting, on the other hand, focuses more on separating the alpha matte of the foreground object from the background for blending [9, 22, 30, 36]. Recently, neural-based methods have + +shown to be effective in estimating layered representations for video segmentation and video editing [3, 14, 16, 38, 39]. In this work, we also adopt neural networks to derive layered representations from videos. Further, inspired by the pioneering work in visual tracking for handling illumination changes [10], we incorporate a new multiplicative residual representation to model the lighting variations for illumination-aware video editing. + +Video editing. Layered representations can benefit video editing in various ways. For example, the layered representations can serve as a visual proxy for intuitive video editing [16] or can be used to create the retiming effects [21]. Editing can be more easily done on a unified texture map built from the video's background, such as background mosaics [7], tapestries [4], and layered neural atlases [16]. Building a unified texture map for the foreground object is also useful, e.g., flexible sprites [15], unwrap mosaics [28], and deformable sprites [39]. Another type of decomposition is to derive temporally-consistent intrinsic components from videos [27, 20] so that the edits can be performed on the reflectance for recoloring or texture transfer. Recent deep-learning-based methods mainly address a single task of video editing, e.g., video style transfer [34, 19, 18, 12, 29]. or category-specific GAN-based video editing [2, 25, 32, 37]. Regarding achieving consistent video editing, it is crucial to know the temporal correspondences [13, 5] so typical animation techniques like Rotoscoping [1] can be applied. + +# 3. Approach + +We formulate the problem of consistent video decomposition and editing as follows: The goal is to decompose an input video into multiple layers, where one layer comprises the background, and each of the remaining layers represents a foreground object. We aim to facilitate user-friendly and intuitive editing of video components on time-independent static texture maps while ensuring the quality and consistency of the edited results in the output video. Fig. 2 illustrates an editing procedure. The model may rely on additional information obtained during pre-processing, such as backward and forward optical flow and rough object masks, to improve the training efficiency and consistency. + +# 3.1. Layer-Based Video Decomposition + +A video of a scene typically contains multiple objects with distinct motions, appearances, shapes, and other attributes. Achieving consistent editing on the video scene and objects can be challenging. An efficient way to group objects is by decomposing the video into different motion layers. Static objects can be considered part of the scene's background, while only a few foreground objects move differently. Editing the objects on decomposed layers becomes more intuitive and straightforward since we can modify each object separately without touching the others. We decompose videos into layers using backward and forward optical flow and coarse object masks, resulting in $N + 1$ layers where layer $N$ represents the background, and the rest is for foreground objects. Each layer $n$ comprises a predicted mask $\alpha_{n}$ across different frames, a time-independent texture map $M_{n}$ , and the multiplicative residual $\ell_{n,t}$ characterizing the illumination for each frame $t$ built on the texture map. Fig. 3 shows an overview of our model. + +Layer hierarchy. Our alpha network $\mathcal{A}$ predicts the object mask probability for every layer. Rather than simply normalizing the layer probabilities of each pixel and directly using them as layer masks, we build a front-to-back hierarchy. That is, for layer $n$ , its probability of object mask is computed by + +$$ +\alpha_ {n} = a _ {n} \cdot \prod_ {i = n + 1} ^ {N} \left(1 - a _ {i}\right), \tag {1} +$$ + +where $a \in [0,1]$ is the raw output of $\mathcal{A}$ with $a_N$ fixed and equal to 1. This hierarchy assumes that objects follow a strict front-to-back ordering, which also implies that layer $N$ represents the background. Here, we specifically focus on scenes with only one foreground object. That is, $N = 1$ . We show an example of decomposing multiple foreground objects in the supplementary material. + +# 3.2. Texture Mapping + +We associate each object with its own texture, such that modifying a single pixel in the texture will impact the object's appearance in the rendered video scene. We translate the spatiotemporal locations in the video into 2D UV coordinates for different layers by a mapping network $\mathcal{M}$ . This allows us to sample colors on a texture with the translated coordinates. We follow the approach of coordinate-based neural rendering [16, 23] and use the multi-layer perceptrons as our coordinate translator instead of a textured grid. + +Note that the appearance of an object in the scene may change throughout the video due to camera motion and illumination variations. We will explain how we address this issue with the proposed multiplicative residual in Sec. 3.3. + +The textures can be further obtained by grid-sampling the texture network $\mathcal{T}$ . Once the texture is extracted, the texture network can be replaced by the extracted texture using UV coordinates from the mapping network, achieving the same effect via bilinear interpolation. This process enables users to modify and apply the extracted texture to the entire video. + +# 3.3. Multiplicative-Residual Estimator + +In order to apply a single modification to all video frames, the texture must remain constant throughout the entire video. However, the color at the same location on an object may change over time due to variations in lighting conditions, resulting in low reconstruction quality or noisy textures. To address this issue and ensure high-quality reconstruction, we have developed a new method, multiplicative-residual estimator, to achieve constant texture across all frames while preserving illumination variations. + +With this design, we can also synthesize illumination coefficients not present in the original video at specific time points, allowing us to manipulate camera motion. Note that the prediction of the residual estimator is sharp. Therefore, we have designed losses to constrain the seen and unseen areas, as will be described in Sec. 3.5. The multiplicative-residual estimator predicts the illumination coefficient of each color channel on the texture map, which is multiplied by each channel to obtain the shaded color. + +To provide more insight into our decision to use a multiplicative residual instead of an additive residual, we examine the process of diffuse shading: + +$$ +C = \rho L, \tag {2} +$$ + +where $\rho$ is the diffuse albedo term determined by surface material and $L$ is the lighting term determined by the environment. When the lighting condition changes or a shadow is cast onto the surface, the new shaded color then becomes $C' = \rho \cdot L'$ . With multiplicative residual, we can reproduce $C'$ by scaling $C$ with $\frac{L'}{L}$ . In contrast, we have to add $C$ by $\rho (L' - L)$ if we use additive residual. One can observe that the additive residual is more difficult to model as it also depends on the diffuse albedo term. We detail the practical result of the two choices in Sec. 4.6. + +# 3.4. Network Architecture + +We design our model as an end-to-end framework trained via self-supervision. We take a spatiotemporal coordinate $p = (x, y, t)$ in the video as the model input and predict the reconstructed color of the corresponding pixel. Fig. 3 shows the pipeline of our model, which includes a mapping network, a texture network, and a multiplicative-residual estimator for each layered representation, with a shared alpha network generating the soft masks for different layers. + +![](images/52225fd8bcd850d0b1e744bc521de56a1fc70f52c84fda14212a25e00e514ccb.jpg) +Figure 3: Model pipeline. Our model takes a video coordinate $p = (x, y, t)$ as input and decomposes the video into multiple layers. A representation for layer $n$ is modeled by a mapping network $\mathcal{M}$ , a texture network $\mathcal{T}$ , and a multiplicative-residual estimator $\mathcal{R}$ ; the three modules jointly convert a video coordinate $p$ into layered color $c_{n}^{(p)}$ and lighting coefficient $l_{n}^{(p)}$ . An alpha network $\mathcal{A}$ also takes $p$ as input and predicts soft object masks $\alpha^{(p)}$ for each layer. The final video color is reconstructed by the bottom-right equation. The multiplicative residuals are critical to our model for handling illumination variations. + +Mapping network. Following the design choice of recent coordinate-based approaches [16, 23], we build our network architecture with multi-layer perceptrons. For each layer $n$ , we pass a spatiotemporal coordinate $p$ through a mapping network to translate the video coordinate into a time-independent 2D texture-map coordinate $(u, v)$ by + +$$ +(u, v) _ {n} ^ {(p)} = \mathcal {M} _ {n} (p), \quad (u, v) _ {n} ^ {(p)} \in [ - 1, 1 ] ^ {2}, \tag {3} +$$ + +where $\mathcal{M}_n$ is the mapping network of layer $n$ . + +Texture network and multiplicative-residual estimator. After obtaining the UV coordinates, the texture network takes the UV coordinates $(u,v)_n^{(p)}$ and produces the corresponding color at each position, while the multiplicative-residual estimator uses both the UV coordinates and the time $t$ as inputs to predict the corresponding illumination coefficients at each position: + +$$ +c _ {n} ^ {(p)} = \mathcal {T} _ {n} \left((u, v) _ {n} ^ {(p)}\right), \quad c _ {n} ^ {(p)} \in [ 0, 1 ] ^ {3}, +$$ + +$$ +\ell_ {n, t} ^ {(p)} = \mathcal {R} _ {n} \left((u, v) _ {n} ^ {(p)}, t\right), \quad \ell_ {n} ^ {(p)} \in \mathbb {R} _ {+} ^ {3}, +$$ + +where $\mathcal{T}$ is the texture network and $\mathcal{R}$ is the multiplicative-residual estimator. + +Alpha network. We also predict an object mask for each layer to indicate the layer to which each pixel belongs: + +$$ +\boldsymbol {\alpha} ^ {(p)} = \mathcal {A} (p), \quad \boldsymbol {\alpha} ^ {(p)} \in [ 0, 1 ] ^ {N + 1}, \tag {5} +$$ + +where $\alpha_{n}^{(p)}$ is the probability that pixel $p$ belongs to layer $n$ , i.e., $\sum_{n=0}^{N} \alpha_{n}^{(p)} = 1$ . The derivation of the object mask is detailed in Sec. 3.1. + +Pixel color reconstruction. After obtaining the information described above, we can reconstruct the pixel color at position $p$ as + +$$ +\hat {c} ^ {(p)} = \sum_ {n = 0} ^ {N} c _ {n} ^ {(p)} \cdot \ell_ {n} ^ {(p)} \cdot \alpha_ {n} ^ {(p)}. \tag {6} +$$ + +Inspired by [24], we use hash grid encoding for all networks except for the mapping network, as texture coordinates are expected to be smooth. This approach is adopted to facilitate better convergence during training. + +# 3.5. Loss Terms + +Losses inherited from previous work. We incorporate several losses from Layered Neural Atlases [16] to facilitate the training process. These losses include 1) optical flow loss, which provides the model with the supervision of optical flow, and 2) alpha bootstrapping loss, which offers supervision of initial coarse masks; additionally, we use 3) reconstruction loss to stabilize the training and improve reconstruction quality, and 4) sparsity loss to avoid duplicate foreground objects in the texture area. Further details can be found in the supplementary material. + +Residual consistency loss. In order to produce significant yet smooth lighting conditions, the multiplicative residuals of different times $t_1$ and $t_2$ of the same position on the texture coordinate $(u, v)$ , where $(u, v, t_1)$ corresponds to a visible area and $(u, v, t_2)$ corresponds to an invisible area, should be close in distribution and can have different intensities of illumination. Formally, we have a small $k \times k$ patch $P$ + +at time $t_1$ on the video centered at $(x, y)$ . We then sample their texture coordinates $(u, v)^{(P)} = \mathcal{M}(P)$ . The texture coordinates are then combined with different times $t_1$ and $t_2$ to get the corresponding multiplicative residuals: + +$$ +\ell^ {(P)} = \mathcal {R} ((u, v) ^ {(P)}, t _ {1}), \tag {7} +$$ + +$$ +\ell^ {\prime (P)} = \mathcal {R} \left((u, v) ^ {(P)}, t _ {2}\right). +$$ + +To this end, we encourage multiplicative residuals of the same position on texture coordinate should be positively correlated. Therefore, we define the residual consistency loss through normalized cross-correlation by + +$$ +\psi (P, t _ {2}) = \frac {\left(\ell^ {(P)} - \mu_ {\ell (P)}\right) \left(\ell^ {\prime (P)} - \mu_ {\ell^ {\prime} (P)}\right)}{\sigma_ {\ell (P)} \sigma_ {\ell^ {\prime} (P)}}. \tag {8} +$$ + +We further introduce a variance-smoothness term that smoothes the changes in lighting conditions as + +$$ +\mathbb {E} (P, t _ {2}) = \sigma_ {\ell^ {\prime} (P)} ^ {2}. \tag {9} +$$ + +We apply the above terms to all layers and get the overall loss as + +$$ +\mathcal {L} _ {\mathcal {R} \text {c o n}} = \lambda_ {\mathcal {R} \text {c o n}} (\psi + \beta \mathbb {E}), \tag {10} +$$ + +where we choose $k = 3$ since our patch size is $3 \times 3$ and $\beta = 16$ . + +Residual regularization. Minimizing the current losses can collapse to a trivial solution because the multiplicative-residual estimator absorbs all colors. We regularize the residuals equal to 1 since "true color" of an object in the video depends on the overall light conditions. That is, if the light source is a pure blue light in the entire video shining on a white wall, we can only say that the steady color of the wall is "blue", as we do not have enough information to determine the true color of the wall. + +$$ +\mathcal {L} _ {\mathcal {R} \text {r e g}} = \lambda_ {\mathcal {R} \text {r e g}} | | \mathcal {R} (\cdot) - 1 | | _ {2} ^ {2}. \tag {11} +$$ + +Alpha regularization. We introduce an additional regularization term to address potential issues with the alpha network. Specifically, the sparsity loss used in the alpha network tends to assign a value of zero (i.e., black) to regions that are not visible in the input video, which can result in noisy masks and incorrect shadow embedding. To mitigate this problem, we enforced a constraint that each pixel should contribute to at most one layer. This ensures that the mask for each layer is clean and reliable and that the lighting conditions are properly embedded in each layer. + +$$ +\mathcal {L} _ {\alpha \operatorname {r e g}} = \lambda_ {\alpha \operatorname {r e g}} \mathrm {B C E} \left(\max _ {n \in \{0, \dots , N \}} \alpha_ {n}\right), \tag {12} +$$ + +where BCE is binary cross entropy. + +# 3.6. Hash Grid Encoding + +Like [16, 23], we have considered employing positional encoding as the input encoding for the multi-layer perceptrons. The purpose of including the positional encoding is to enrich the discriminative power of the coordinates and ensure that high-frequency details can be adequately represented. Inspired by [24], we choose to adopt hash grid encoding as the input encoding method for our model. By leveraging this technique, the input features are encoded into a set of intermediate representations that span a broad spectrum of spatiotemporal resolutions, ranging from coarse-to fine-grained, and can flexibly adjust to distinct regions while maintaining high levels of accuracy and consistency. Formally, the 2D or 3D input feature is treated as a coordinate of multi-resolution grids and then used to sample data from the grids using interpolation techniques. The resulting sampled data are concatenated and passed into multi-layer perceptrons to obtain the final output. + +# 4. Experiments + +We conduct our experiments on the DAVIS dataset [26] and various internet videos to demonstrate the effectiveness of our approach in video reconstruction and consistent video editing. We also design an evaluation metric of edit consistency on the TAP-Vid-DAVIS dataset [8]. In addition, we use our method to generate videos with different camera motions from those in the original input video. Furthermore, we conduct ablation studies on the multiplicative-residual estimator and encoding type to assess the impact on the performance of different architecture choices. + +# 4.1. Qualitative Results + +Fig. 4 shows four qualitative results of our work on the DAVIS dataset [26]. For each video frame, we show the predicted masks (second and third columns), the m-residuals (fourth column), the corresponding layered textures (fifth and sixth columns), and the reconstruction result (last column). Please also refer to Fig. 1 for an example result highlighting the robustness of our m-residual estimation. More results and higher resolution reconstruction can be found in the supplementary material. + +With merely the guidance from the coarse object mask provided by Mask R-CNN [11], our model successfully separates the object from the background yielding a clean and precise mask. In the second video in Fig. 4, we showcase an example of a video occluded by complex water splashes. Our model can accurately reconstruct the video without introducing noticeable texture distortions. We superimpose a rainbow checkerboard on the mask to visualize the transformation of each layer. It is evident that each coordinate's component remains at the same position in the texture. + +Since the multiplicative residual may brighten or darken + +![](images/f69de98daf2a54b7b55ab918e6b9a533c9a408e6079f1099873d8a1514292787.jpg) +Figure 4: Qualitative results on the DAVIS dataset. We show our predicted masks, textures, m-residuals, reconstructed results, and PSNR of different videos. We overlay a color checkerboard on the masked areas to visualize the texture transformations. + +the texture, we display the m-residual output on a gray canvas to visualize its effect. In the first video, our m-residual estimator successfully models the shadow of the hiker while the background texture retains its original color. The m-residual estimator takes both the texture coordinate and time as the input, allowing it to learn how to model the changes in lighting conditions through space-time and adjust the color accordingly. Such a mechanism enables our model to generate realistic and consistent shading effects for different objects in the video, such as the shading on the bear's fur or the reflection on the hiker's hat. We also report the PSNR of each video as evidence of high-quality reconstruction. + +# 4.2. Comparison with Previous Work + +We report the PSNR values and various settings for every approach in Table 1. We also provide more quantitative comparisons with DeformableSprites (DS) [39] and Layered Neural Atlases (LNA) [16] under multiple metrics in Table 2. Our method demonstrates superior performance of PSNR, rendering speed, and GPU memory consumption compared to previous methods on all three videos while maintaining the same or better resolution. Note that we could not run DeformableSprites for the $768 \times 432$ resolution, as DeformableSprites failed to finish due to their GPU memory usage exceeding the limit of a single GPU with + +24 GB memory. In contrast, our method can efficiently decompose higher-resolution videos into full-HD layers using affordable computation resources while maintaining comparable reconstruction results. To compute the PSNR, we scale the input video up to the reconstructed resolution. Therefore, the PSNR of our 1080p model might be slightly lower than that of the $768 \times 432$ model, as the PSNR of the 1080p model is calculated at the corresponding resolution and considered harder due to having more details to reconstruct than the low-resolution counterpart. + +# 4.3. Consistent Video Editing + +For editing purposes, we use a $1000 \times 1000$ grid sampling to render the texture networks. We then modify the rendered texture and sample color via bilinear interpolation. As our key advantage lies in the m-residual representation of illumination, we have focused our editing on videos or components that exhibit changes in lighting. + +We show our editing results in Fig. 5. We choose three videos to present the rendering quality. The first video is sourced from internet videos, while the other two are taken from the DAVIS dataset [26]. For evaluating the editing quality, we modify the background texture of the first two videos by adding handwritten characters of "ICCV". We also adjust the color of the shirt of the hiker in the second video. + +![](images/0b5ae996a36d23ea05bf981743f1c1df593a8a6dfb3bbfd4df611667fe2047c7.jpg) +(a) Input frames + +![](images/3e1d61ece7af3520bff1e907b8c311227672d654b247b7afc181db92333f8ea8.jpg) + +![](images/0c5b2cec9fa357f17c05496ac7b8ca69792aa5a530a623b41ff7459c7a333990.jpg) + +![](images/c1ec89e8585bf3dc67a942fd93b9c84ff887645ff1d9d623a4ee70b4107f1e43.jpg) + +![](images/fb1e93478c37d6d36f3ff5dcf9dcc05d7541e3031bf5fa0f249469a921fca723.jpg) + +![](images/169d459c2ecf5f31748f113496a4f4cae1953973bd5fd9e8acc6982b5d74e325.jpg) +Figure 5: Edits on videos with varying lighting conditions. We apply various editing techniques on different components of the videos to observe the corresponding changes in the lighting conditions. The textures with edits are shown in (c), while the editing results are enlarged in the right column of (d). The lighting and shading fuse with the edits flawlessly. + +![](images/e57df96c2fe4c15a221a70997cf63a3314a3c171bafa80db868fde585c8f06d8.jpg) +(b) Reconstruction + +![](images/4b67b1e8f2d0b569625f7fa9ab7db15dfb9e459391420e33f21a7460b8132944.jpg) + +![](images/239bccd5f35a7b7cad2992df4e1095b55e4b3effb536400f0225072fb9ad1788.jpg) + +![](images/4508c8874d08caf9558a01c59a62196c00a665bf1b3d72ee90457a477ec04cf7.jpg) + +![](images/b1b123cc8b196cae4f1ee51fa47dcf9da80b212c1daf8255deb341e256f7e6f0.jpg) + +![](images/b92f6b643eb3e94000bd3ab2e873508c23d989e005776e2f6324ef5d0366f0a8.jpg) + +![](images/de186c67c6d37eb10c8175e65533fe6b7a91bbadb7eb6b0ce429b5daf400b426.jpg) +(c) Texture editing + +![](images/a28123c83715de266bb264d5d7993753b6c8d55e67f0cb8020deb65c35cba6de.jpg) + +![](images/c5b2c0e3f1d466fe55c50db0bea07e0eb6fc70a21c49695d683500a5ba61b926.jpg) + +![](images/1bd04ee50860b8df27a25851b471432485759c269ef98b8b1917837ec312475f.jpg) + +![](images/fd277c7238ac925a08bb89dc1ca4bd5708fcbc114ace71bf1c35529389f16b28.jpg) + +![](images/cdb650b79c19a6754ca2b75715c98f00c2f506c7058314b2cb93257f3b81ddb4.jpg) +(d) Reconstruction (edited) + +![](images/81d36b73c367f29bb4ccfde664f4763de21457cf07e7a8a7a8557a9ac524dd33.jpg) + +![](images/0f35b7475758a0fbb1751ecd5da412d41773eeddea453acb2ca1533e94b15f20.jpg) + +![](images/ef89203d8d4142c237b0ce44e03b23880f633c3144737832ae9888349359a74b.jpg) + +![](images/f80e9cdd1c668cfa7401e7a427da8cf7945a918b6c706a379bab560313123409.jpg) + +![](images/e79c64c757640c82b77a046f938bf84c3a8eff0b6d71a3ae42959d7149a99bd4.jpg) + +![](images/92f42925567da358c7cf59d98f3f1b807eec23c93e0ddcb26c0f49155f44265c.jpg) + +![](images/3c44b9b682d8053232812e95cee3d177119fa74a6984890fb9888ab0a7ab80c7.jpg) + +![](images/1f0299942457d08e454a5e9039614f5142fd792a7fcf22dbb955dc83451b79b3.jpg) + +![](images/f7b11340b5d5dd8ac5d606c72c3dc002b8b2778bc82f438af938741e589a5e4b.jpg) + +![](images/d8f7098b22ddb80a9c896e8f0c24da5ac3facec20be4098d7dbb09bb79d3c5a1.jpg) + +![](images/1752c6d60a450c7f807a8c36fab32cf1714e2cda2324a9aad16b779aa3bf8654.jpg) + +Finally, we perform style transfer on the third video on both background and object textures. In the first video, despite the complexity of the lighting conditions caused by the Disco ball, our approach can successfully handle and diffuse the light onto the edited region, resulting in high-quality output. Previous approaches fail to achieve such a representation of complex lighting conditions. In the second video, the color of the shirt changes convincingly according to the different lighting conditions in two different frames, while our edit on the background is successfully occluded by the shadow. We can see that in the third video, our edited reconstruction follows the textures and the modified lighting conditions fit seamlessly with our modifications, as shown in (d) of the fifth and sixth rows. In the supplementary material, we also provide additional editing results of various types to further demonstrate the versatility and effectiveness of our method. + +# 4.4. Quantitative Results for Editing Consistency + +Our work mainly focuses on consistent video editing. However, famous metrics such as PSNR (as we show in Fig. 4) or other reconstruction quality measures do not necessarily reflect the quality of consistent video editing. We + +take advantage of TAP-Vid [8], which provides several evaluation metrics and the ground-truth feature tracking on the foreground object. We use the feature points of the first unobstructed frame $(f_{0,x}, f_{0,y})$ as our base point and convert it into texture coordinate $(u_0, v_0)$ . For all other frames $t$ , we select the video coordinate whose mapped texture coordinate is closest to $(u_0, v_0)$ as our prediction for frame $t$ . Therefore, the feature points on the same track should be mapped to the same coordinate on the foreground texture for an ideal rendering result. + +We report our results in Table 3. We select a subset of the TAP-Vid DAVIS dataset for our quantitative analysis. For reference, we also report the results of RAFT [31], as a baseline, on the same videos since our optical flow supervision comes from it. In general, the correspondences obtained by our method are more consistent than those by RAFT, as RAFT considers only neighboring frames while ours works on a unified foreground representation. Specifically, our method achieves higher scores than RAFT on the "black swan" and "kite-surf" videos and comparable results on the other two videos. The low score on the "parkour" video is due to the complex geometric changes in the video, which + +
MethodResolutionTraining timeRendering speed (fps)GPU memorybearPSNR disco ballhike
DeformableSprites [39]213 × 12010 minutes55 GB22.726.221.5
DeformableSprites [39]427 × 24020 minutes1.612 GB23.626.422.0
Layered Neural Atlases [16]768 × 4325.5 hours0.53 GB27.329.025.2
Ours768 × 43240 minutes7873 GB27.537.725.6
Ours1920 × 108040 minutes715 GB26.935.625.4
+ +Table 1: Comparison results. We report the PSNR for each video, along with the corresponding training time, rendering speed (frames per second), and GPU memory usage under different resolutions. Our work achieves better results than prior work on the three videos, achieving faster rendering, lower GPU memory consumption, and higher resolution. Note that we measure the PSNR at the corresponding reconstructed resolution. Therefore the PSNR tends to favor the evaluation at low resolution, explaining why our 1080 reconstructions have slightly lower PSNR than our 480p reconstructions. + +
Methodbeardisco ballhike
PSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIM
DS23.60.230.7826.40.210.8922.00.290.68
LNA27.30.190.8529.00.110.9525.20.190.80
Ours27.50.160.8737.70.040.9725.60.160.82
kite-surfcar-turnlibby
DS21.20.340.6222.20.320.6521.60.410.57
LNA28.20.300.7627.50.300.8829.40.310.91
Ours29.20.260.7829.50.230.9129.60.260.92
+ +Table 2: More comparison results. Our model outperforms previous methods in all videos and achieves better scores in all evaluation metrics. + +
oursRAFT
AJ<δxavgOAAJ<δxavgOA
black swan0.810.871.000.390.521.00
parkour0.040.100.480.070.170.48
kite-surf0.480.580.990.250.350.86
cows0.510.610.990.550.650.99
+ +Table 3: Quantitative results on TAP-Vid. We test the editing consistency on a subset of TAP-Vid DAVIS and show the result for each video. The reported metrics, the higher the better, are average Jaccard (AJ), average position accuracy $(< \delta_{avg}^{x})$ , and binary occlusion accuracy (OA). Compared to RAFT [31], our method performs better or reaches comparable results on all videos. + +our current architecture can reconstruct well but edit poorly on the foreground parkour runner. Nevertheless, the editing on the background still exhibits comparable quality. + +# 4.5. Manipulating Camera Motion + +With our multiplicative residual estimator, we are able to synthesize realistic views that are not shown in the input video. A different camera view can be synthesized by scaling, shifting, or rotating the original video coordinates $(x,y)$ to get $(x^{\prime},y^{\prime})$ and concatenating it with the temporal information $t$ . Once the new set of video coordinates $p^\prime = (x',y',t)$ is obtained, we can use the mapping network $\mathcal{M}$ to obtain the texture coordinate $(u,v)^{(p^{\prime})}$ for each layer. The color $c^{(p^{\prime})}$ and the lighting $\ell^{(p^{\prime})}$ of all layers can be obtained based on these texture coordinates. For each non-background layer $n$ , we use the alpha network $\mathcal{A}$ to + +![](images/3a5ad7ee38f10641c2fb093b108f0874b6bcf2ab6dece28ab5521535df97851b.jpg) + +![](images/3a3b85874d44b80ed5405f295a43e47b72724dcaa649fc7b1d5e962f23da02bb.jpg) + +![](images/a15f369b591a546d3f3325ec9fb9cc764d1fe0d773ffd4767c79df8290996b62.jpg) + +![](images/f438d623a1eebd03388bb2bfe6b49c828b4686b3a649e18b3724b5ea21cc311f.jpg) +Figure 6: Camera motion manipulation. The multiplicative residual serves as a smooth representation that effectively adjusts the lighting conditions with a fixed camera. + +![](images/055f2670cac109144b222d7f0888f383a01a40b022a7889ab940382b667f3bca.jpg) + +![](images/43b1ecd1e9d0619ba062869d67eca4955faf94f52cdd63d98ff7868a9a1df9a2.jpg) + +![](images/f0bb7635a0b3f3124267e2c5b3396642f45d810890762a7a38af23a503df0881.jpg) + +![](images/b3eb9aed5d43499c6180a8626edc17f530e7738f9ff28f218d3a91ce6873d41e.jpg) + +![](images/c1d1d33aff99057419d5712d79ff39759bdac132262538fa72201ae4fd80cea9.jpg) + +![](images/953a793c7d37c79fbf8e991cf22fb308e1e981d6b1ca4e10bafcb1a712ee6c14.jpg) + +![](images/c4b160f25385cca27b274242272a594d2f92206ebbd952be10d54933d3e9a0f2.jpg) + +obtain the original soft mask $\alpha^{(p)}$ for each $p = (x,y,t)$ and establish a relation as $(u,v)_0\rightarrow \alpha_n^{(p)}$ if $(u,v)_0 = \mathcal{M}_0(p)$ . Then, we linearly interpolate the value of the mask at each location $p^\prime$ by triangulating the input data. That is, we establish a correspondence between the coordinates of the object textures and the background texture and then transfer the mask values from the object onto the background texture. + +We set up a virtual camera in the background to capture the entire video scene from a fixed position. Fig. 6 shows our result. The camera is fixed (i.e., the background remains static) while the object keeps moving in the scene. In the fourth video, the car is partially visible in the beginning frames and fully visible later, with its side facing the camera. The third video demonstrates that the splashes consistently follow the entire path of the original camera view. With our multiplicative-residual estimator, we can successfully reconstruct the surfing scene while accurately representing the water splashes. We show more results with varying camera motions in the supplementary material. + +![](images/c0d721cefee8d00bed3918942b5ad40d62f986f69e38140bf03cdcd7d115a456.jpg) +Figure 7: Choice of the residual estimator architecture. Using an additive residual estimator architecture may introduce artifacts or degrade the quality of textures, while a multiplicative one works correctly. + +![](images/062d31978c458d533b9e463af99bf2f657b04eeeeaf495762c75c788129f8754.jpg) +Figure 8: Quantitative ablation studies. (a) The reconstruction PSNR degradation of the additive residual is significantly higher than that of the multiplicative residual. (b) Hash grid encoding can achieve better PSNR in fewer iterations. + +![](images/0593c1673a69907154c2b783be94fd739c9271c943bf2d01221c7475ed67b902.jpg) + +# 4.6. Ablations + +Residual type. We present an ablation study to verify the choice of the architecture of our residual estimator. As mentioned in Sec. 3.3, we adopt the multiplicative residual instead of the additive residual. We show the comparisons of the two designs in Fig. 7 and Fig. 8a. The additive residual estimator is conducted, and the output is normalized to $[-1,1]$ to ensure they have the same representation capability. In Fig. 7, both additive and multiplicative residuals provide promising reconstruction results for the first video. However, when it comes to the complex texture, like the background of the second video, the additive residual fails to represent it and compensates for the wrong illuminations. In the third video, the additive residual fails to represent precise + +background texture, and the dog vanishes from the texture. Fig. 8a demonstrates a significant drop in PSNR when ablating the residual module. This finding reflects that additive residual is prone to overfit the color that should be attributed to the editable texture. In contrast, the multiplicative residual provides complete background and object textures. + +Hash grid encoding. We evaluate the effectiveness of hash grid encoding in Fig. 8b. Positional encoding needs more MLP layers to achieve good quality, which takes $1.5 \times$ processing time per iteration compared to hash encoding. However, hash grid encoding still achieves better PSNR in fewer iterations. + +# 5. Discussion + +We have observed several interesting points that are worth discussing. Our method relies on Mask-RCNN and RAFT to provide the foreground mask and optical flow as external priors. As a result, our method may be biased by the inaccurate learning-based priors on some difficult cases like the spinning lighting in disco ball or the object boundary in bear. The reconstructed objects in the video may exhibit edge artifacts due to referencing the incorrect texture layer. Our method with reconstruction loss could improve the initial external prior but is still not perfect to solve all the artifacts. Adopting better external priors or designing strong internal priors are both good future improvements. The problem of object edge artifacts is similar to the seam artifact in mesh uv-parameterization from which future exploration may take inspiration. + +# 6. Conclusion + +We have presented a neural layer decomposition method to facilitate illumination-aware video editing. The proposed multiplicative-residual estimator can effectively derive the layered representation that characterizes the spatiotemporally varying lighting effects. We have also implemented hash grid encoding for fast coordinate inference. Our model, therefore, significantly reduces the training time and achieves real-time rendering speed with a low requirement of computation resources, enabling interactive editing on high-resolution videos. We use our model to generate high-quality video editing results, where, in particular, the varying illumination effect can only be achieved by ours rather than the previous methods. + +# 7. Acknowledgments + +This work was supported in part by NSTC grants 111-2221-E-001-011-MY2 and 112-2221-E-A49-100-MY3 of Taiwan. We are grateful to National Center for High-performance Computing for providing computational resources and facilities. + +# References + +[1] Aseem Agarwala, Aaron Hertzmann, David Salesin, and Steven M. Seitz. Keyframe-based tracking for rotoscoping and animation. ACM Trans. Graph., 23(3):584-591, 2004. 2 +[2] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. In Computer Vision - ECCV 2022 Workshops - Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, volume 13802 of Lecture Notes in Computer Science, pages 204-220. Springer, 2022. 2 +[3] Jean-Baptiste Alayrac, João Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5733-5742. IEEE, 2019. 2 +[4] Connelly Barnes, Dan B. Goldman, Eli Shechtman, and Adam Finkelstein. Video tapestries with continuous temporal zoom. ACM Trans. Graph., 29(4):89:1-89:9, 2010. 2 +[5] Zhangxing Bian, Allan Jabri, Alexei A. Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 6498-6509. IEEE, 2022. 2 +[6] Gabriel J. Brostow and Irfan A. Essa. Motion based decompositing of video. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 8-13. IEEE Computer Society, 1999. 2 +[7] Carlos D. Correa and Kwan-Liu Ma. Dynamic video narratives. ACM Trans. Graph., 29(4):88:1-88:9, 2010. 2 +[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurIPS Datasets Track, 2022. 5, 7 +[9] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ArXiv, abs/2211.02145, 2022. 2 +[10] Gregory D. Hager and Peter N. Belhumeur. Efficient region tracking with parametric models of geometry and illumination. IEEE Trans. Pattern Anal. Mach. Intell., 20(10):1025-1039, 1998. 2 +[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2020. 5 +[12] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 7044-7052. IEEE Computer Society, 2017. 2 +[13] Allan Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2 + +[14] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3154-3164. IEEE Computer Society, 2017. 2 +[15] Nebojsa Jojic and Brendan J. Frey. Learning flexible sprites in video layers. In 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2001), with CD-ROM, 8-14 December 2001, Kauai, HI, USA, pages 199-206. IEEE Computer Society, 2001. 2 +[16] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 1, 2, 3, 4, 5, 6, 8 +[17] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. Int. J. Comput. Vis., 76(3):301-319, 2008. 2 +[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XV, volume 11219 of Lecture Notes in Computer Science, pages 179-195. Springer, 2018. 2 +[19] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 3809-3817. Computer Vision Foundation / IEEE, 2019. 2 +[20] Sharon Lin, Matthew Fisher, Angela Dai, and Pat Hanrahan. Layerbuilder: Layer decomposition for interactive image and video color editing. ArXiv, abs/1701.03754, 2017. 2 +[21] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. ACM Trans. Graph., 39(6):256:1-256:14, 2020. 2 +[22] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4507-4515. Computer Vision Foundation / IEEE, 2021. 2 +[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 5 +[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5 +[25] William S. Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei A. Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Or + +leans, LA, USA, June 18-24, 2022, pages 13460-13471. IEEE, 2022. 2 +[26] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus H. Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 724-732. IEEE Computer Society, 2016. 5, 6 +[27] Hanspeter Pfister. Interactive intrinsic video editing. ACM Trans. Graph., 33(6):197:1-197:10, 2014. 2 +[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew W. Fitzgibbon. Unwrap mosaics: a new representation for video editing. ACM Trans. Graph., 27(3):17, 2008. 2 +[29] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Pattern Recognition - 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings, volume 9796 of Lecture Notes in Computer Science, pages 26-36. Springer, 2016. 2 +[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 2288-2297. Computer Vision Foundation / IEEE, 2020. 2 +[31] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12347 of Lecture Notes in Computer Science, pages 402-419. Springer, 2020. 7, 8 +[32] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, SA 2022, Daegu, Republic of Korea, December 6-9, 2022, pages 29:1-29:9. ACM, 2022. 2 +[33] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. IEEE Trans. Image Process., 3(5):625-638, 1994. 2 +[34] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 29:9125-9139, 2020. 2 +[35] Josh Wills, Sameer Agarwal, and Serge J. Belongie. What went where. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2003), 16-22 June 2003, Madison, WI, USA, pages 37-44. IEEE Computer Society, 2003. 2 +[36] Ning Xu, Brian L. Price, Scott Cohen, and Thomas S. Huang. Deep image matting. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 311-320. IEEE Computer Society, 2017. 2 +[37] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, volume 13675 of Lecture Notes in Computer Science, pages 357-374. Springer, 2022. 2 + +[38] Yanchao Yang, Brian Lai, and Stefano Soatto. Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 2826-2836. Computer Vision Foundation / IEEE, 2021. 2 +[39] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 2647-2656. IEEE, 2022. 1, 2, 6, 8 \ No newline at end of file diff --git a/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/images.zip b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ae0fc9750735c74fa11e66a5594ec8d371ac5dd --- /dev/null +++ b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a0d3ca780d54f59da33d3df188e4749ccac682d07781d74c19bea1e1df3c5c8 +size 912380 diff --git a/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/layout.json b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..43b18c50b3b78a6bb1dcd6e86cd208f2b8e76be0 --- /dev/null +++ b/2023/Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time/layout.json @@ -0,0 +1,10409 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 123 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 123 + ], + "type": "text", + "content": "Hashing Neural Video Decomposition with Multiplicative Residuals in Space-Time" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 143, + 187, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 143, + 187, + 158 + ], + "spans": [ + { + "bbox": [ + 93, + 143, + 187, + 158 + ], + "type": "text", + "content": "Cheng-Hung Chan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 144, + 299, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 144, + 299, + 157 + ], + "spans": [ + { + "bbox": [ + 205, + 144, + 299, + 157 + ], + "type": "text", + "content": "Cheng-Yang Yuan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 321, + 144, + 377, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 144, + 377, + 157 + ], + "spans": [ + { + "bbox": [ + 321, + 144, + 377, + 157 + ], + "type": "text", + "content": "Cheng Sun" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 399, + 144, + 500, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 144, + 500, + 157 + ], + "spans": [ + { + "bbox": [ + 399, + 144, + 500, + 157 + ], + "type": "text", + "content": "Hwann-Tzong Chen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 202, + 158, + 392, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 158, + 392, + 171 + ], + "spans": [ + { + "bbox": [ + 202, + 158, + 392, + 171 + ], + "type": "text", + "content": "National Tsing Hua University, Taiwan" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 200, + 144, + 350 + ], + "blocks": [ + { + "bbox": [ + 58, + 200, + 144, + 350 + ], + "lines": [ + { + "bbox": [ + 58, + 200, + 144, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 200, + 144, + 350 + ], + "type": "image", + "image_path": "c3cc7f33312a90cdd9da9e53db670d91daac815c5923e8cb1dc3b49f9c9070c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 353, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 353, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 547, + 390 + ], + "type": "text", + "content": "Figure 1: Integration of layer-based video editing and spatiotemporal lighting rendering. Our approach decomposes an input video into layers with multiplicative residuals that characterize complex spatiotemporal lighting variations. Our method efficiently fuses the user-edited components with the expected lighting conditions to produce high-quality video output." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 149, + 201, + 236, + 349 + ], + "blocks": [ + { + "bbox": [ + 149, + 201, + 236, + 349 + ], + "lines": [ + { + "bbox": [ + 149, + 201, + 236, + 349 + ], + "spans": [ + { + "bbox": [ + 149, + 201, + 236, + 349 + ], + "type": "image", + "image_path": "002359bd18af5c95ef2d281d022cba4e63e98a75f708341649fdd59a35d6e294.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 241, + 201, + 326, + 279 + ], + "blocks": [ + { + "bbox": [ + 241, + 201, + 326, + 279 + ], + "lines": [ + { + "bbox": [ + 241, + 201, + 326, + 279 + ], + "spans": [ + { + "bbox": [ + 241, + 201, + 326, + 279 + ], + "type": "image", + "image_path": "e4956e2e1ffd1221174969f5fec8724be071d64a10cc2977ccf1de1549f977fd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 241, + 284, + 326, + 348 + ], + "blocks": [ + { + "bbox": [ + 241, + 284, + 326, + 348 + ], + "lines": [ + { + "bbox": [ + 241, + 284, + 326, + 348 + ], + "spans": [ + { + "bbox": [ + 241, + 284, + 326, + 348 + ], + "type": "image", + "image_path": "f6adb6c3653031b16414dd3a87c4d4cf4b90494c458bbecd60a3e55bd76d37cf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 238, + 346, + 255 + ], + "blocks": [ + { + "bbox": [ + 328, + 238, + 346, + 255 + ], + "lines": [ + { + "bbox": [ + 328, + 238, + 346, + 255 + ], + "spans": [ + { + "bbox": [ + 328, + 238, + 346, + 255 + ], + "type": "image", + "image_path": "80428b5481abd1bd9b3b96b894718951b76ad45ef292400f42496a13927cea92.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 350, + 201, + 441, + 279 + ], + "blocks": [ + { + "bbox": [ + 350, + 201, + 441, + 279 + ], + "lines": [ + { + "bbox": [ + 350, + 201, + 441, + 279 + ], + "spans": [ + { + "bbox": [ + 350, + 201, + 441, + 279 + ], + "type": "image", + "image_path": "c65c1c7ca37cb7538f7cb9b86c1740058371ec56b0026e3257f5ec18e373ecbe.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 441, + 201, + 531, + 278 + ], + "blocks": [ + { + "bbox": [ + 441, + 201, + 531, + 278 + ], + "lines": [ + { + "bbox": [ + 441, + 201, + 531, + 278 + ], + "spans": [ + { + "bbox": [ + 441, + 201, + 531, + 278 + ], + "type": "image", + "image_path": "75937d4f69df8d5c0db497160b6b17400322bd2df6f2c14e38376608e89462c2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 349, + 284, + 440, + 347 + ], + "blocks": [ + { + "bbox": [ + 349, + 284, + 440, + 347 + ], + "lines": [ + { + "bbox": [ + 349, + 284, + 440, + 347 + ], + "spans": [ + { + "bbox": [ + 349, + 284, + 440, + 347 + ], + "type": "image", + "image_path": "51036f3ea5921360b0e770ece302c0c5eaa113558382df0ddadf65e2df9f6712.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 441, + 284, + 531, + 348 + ], + "blocks": [ + { + "bbox": [ + 441, + 284, + 531, + 348 + ], + "lines": [ + { + "bbox": [ + 441, + 284, + 531, + 348 + ], + "spans": [ + { + "bbox": [ + 441, + 284, + 531, + 348 + ], + "type": "image", + "image_path": "3920fd85b084ed3db956e77d933dd5cbea64d1e064d3192c93c86beea99623e7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 403, + 192, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 403, + 192, + 416 + ], + "spans": [ + { + "bbox": [ + 143, + 403, + 192, + 416 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 429, + 290, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 290, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 290, + 646 + ], + "type": "text", + "content": "We present a video decomposition method that facilitates layer-based editing of videos with spatiotemporally varying lighting and motion effects. Our neural model decomposes an input video into multiple layered representations, each comprising a 2D texture map, a mask for the original video, and a multiplicative residual characterizing the spatiotemporal variations in lighting conditions. A single edit on the texture maps can be propagated to the corresponding locations in the entire video frames while preserving other contents' consistencies. Our method efficiently learns the layer-based neural representations of a 1080p video in 25s per frame via coordinate hashing and allows real-time rendering of the edited result at 71 fps on a single GPU. Qualitatively, we run our method on various videos to show its effectiveness in generating high-quality editing effects. Quantitatively, we propose to adopt feature-tracking evaluation metrics for objectively assessing the consistency of video editing. Project page: https://lightbulb12294.github.io/hashing-nvd/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Unlike image editing, video editing involves modeling the frame-to-frame relationships and addressing temporal" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 501 + ], + "type": "text", + "content": "variations such as motion and illumination changes. The task of video editing becomes challenging with the added dimension of time for a user who attempts to apply edits to a video while ensuring consistency across all frames. It is more convenient if we can handle the spatiotemporally varying components and reduce video editing to image editing—The user thus only needs to do edits on images with ease, and the editing results can seamlessly propagate to the entire video." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "content": "To achieve this goal, we may consider incorporating effective representations that can model and reconstruct the static and dynamic information in videos. Furthermore, for practical concerns, the algorithm must be efficient enough in modeling and rendering to support interactive editing. Recent work on video decomposition has proposed to employ neural-based representations, such as layered neural atlases [16] and deformable sprites [39], to enable the conversion between the space-time video domain and 2D texture domain for editing. Despite their successes in showcasing impressive video editing effects, we notice that they often require longer training time or restrict to limited frame resolution. For example, it takes more than ten hours to derive the layered neural atlases [16] from a 100-frame 480p video. Deformable sprites [39] are relatively fast to derive but require more computation resources (over 24GB of GPU memory for a 480p video) and thus are unsuitable for editing" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7743" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 539, + 126 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 539, + 126 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 539, + 126 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 539, + 126 + ], + "type": "image", + "image_path": "0d7757873dccaf71565c88b0f5ff4612fb81c20cd4c3d8e6caac04250109bb40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 133, + 547, + 157 + ], + "lines": [ + { + "bbox": [ + 46, + 133, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 547, + 157 + ], + "type": "text", + "content": "Figure 2: The procedure of layer-based video editing. Our model allows users to apply edits to the extracted texture for rendering the edited video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 178, + 289, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 178, + 289, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 289, + 274 + ], + "type": "text", + "content": "high-resolution videos. We address the computation issue by incorporating hash grid encoding [24] into our framework and achieve fast training and rendering for high-resolution videos. Moreover, we introduce the multiplicative residual estimator to model the lighting variations across video frames, which can improve the reconstruction quality and allow illumination-aware editing unachievable by prior work, as shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 275, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 275, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 58, + 275, + 287, + 286 + ], + "type": "text", + "content": "We summarize the contributions of this work as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 288, + 288, + 559 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 288, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 287, + 360 + ], + "type": "text", + "content": "1. This work is the first to consider spatiotemporally varying lighting effects for layered-based video editing. The proposed multiplicative-residual estimator effectively decomposes the lighting conditions from the video without supervision. Our method can produce better-quality videos by fusing the edits with expected illuminations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 364, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 288, + 483 + ], + "type": "text", + "content": "2. Our approach is efficient in both training and rendering. Compared with prior work, the proposed method improves the training time with fewer resources and thus enables training on higher-resolution or longer videos. The trained model can achieve fast video rendering via hashing-based coordinate inference. It takes about 40 minutes to train with a 1080p video of 100 frames on a single 3090 Ti GPU. The inference speed for rendering an edited video is 71 fps for 1080p resolution, allowing real-time video editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 487, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 287, + 559 + ], + "type": "text", + "content": "3. The experimental results demonstrate appealing video edits in various challenging contexts, such as modifying the texture of moving objects, handling occlusion, and manipulating camera motion, where all can be fused with vivid lighting and shading for better effects that are not easy to achieve by prior work." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 572, + 134, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 134, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 134, + 585 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Video decomposition. Decomposing a video into layered representations is a long-standing video analysis approach in computer vision since the seminal work by Wang & Adelson [33]. Similar ideas have been continually revisited under different contexts with the development of new techniques. A typical formulation is motion segmentation that decomposes the video by motion [6, 15, 17, 35]. Video matting, on the other hand, focuses more on separating the alpha matte of the foreground object from the background for blending [9, 22, 30, 36]. Recently, neural-based methods have" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 178, + 547, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 547, + 275 + ], + "type": "text", + "content": "shown to be effective in estimating layered representations for video segmentation and video editing [3, 14, 16, 38, 39]. In this work, we also adopt neural networks to derive layered representations from videos. Further, inspired by the pioneering work in visual tracking for handling illumination changes [10], we incorporate a new multiplicative residual representation to model the lighting variations for illumination-aware video editing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 301, + 548, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 548, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 548, + 531 + ], + "type": "text", + "content": "Video editing. Layered representations can benefit video editing in various ways. For example, the layered representations can serve as a visual proxy for intuitive video editing [16] or can be used to create the retiming effects [21]. Editing can be more easily done on a unified texture map built from the video's background, such as background mosaics [7], tapestries [4], and layered neural atlases [16]. Building a unified texture map for the foreground object is also useful, e.g., flexible sprites [15], unwrap mosaics [28], and deformable sprites [39]. Another type of decomposition is to derive temporally-consistent intrinsic components from videos [27, 20] so that the edits can be performed on the reflectance for recoloring or texture transfer. Recent deep-learning-based methods mainly address a single task of video editing, e.g., video style transfer [34, 19, 18, 12, 29]. or category-specific GAN-based video editing [2, 25, 32, 37]. Regarding achieving consistent video editing, it is crucial to know the temporal correspondences [13, 5] so typical animation techniques like Rotoscoping [1] can be applied." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 548, + 373, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 373, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 373, + 561 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "We formulate the problem of consistent video decomposition and editing as follows: The goal is to decompose an input video into multiple layers, where one layer comprises the background, and each of the remaining layers represents a foreground object. We aim to facilitate user-friendly and intuitive editing of video components on time-independent static texture maps while ensuring the quality and consistency of the edited results in the output video. Fig. 2 illustrates an editing procedure. The model may rely on additional information obtained during pre-processing, such as backward and forward optical flow and rough object masks, to improve the training efficiency and consistency." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7744" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 231, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 231, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 231, + 85 + ], + "type": "text", + "content": "3.1. Layer-Based Video Decomposition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": "A video of a scene typically contains multiple objects with distinct motions, appearances, shapes, and other attributes. Achieving consistent editing on the video scene and objects can be challenging. An efficient way to group objects is by decomposing the video into different motion layers. Static objects can be considered part of the scene's background, while only a few foreground objects move differently. Editing the objects on decomposed layers becomes more intuitive and straightforward since we can modify each object separately without touching the others. We decompose videos into layers using backward and forward optical flow and coarse object masks, resulting in " + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": " layers where layer " + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": " represents the background, and the rest is for foreground objects. Each layer " + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": " comprises a predicted mask " + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "inline_equation", + "content": "\\alpha_{n}" + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": " across different frames, a time-independent texture map " + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "inline_equation", + "content": "M_{n}" + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": ", and the multiplicative residual " + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "inline_equation", + "content": "\\ell_{n,t}" + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": " characterizing the illumination for each frame " + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 90, + 290, + 307 + ], + "type": "text", + "content": " built on the texture map. Fig. 3 shows an overview of our model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 322, + 288, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 395 + ], + "type": "text", + "content": "Layer hierarchy. Our alpha network " + }, + { + "bbox": [ + 46, + 322, + 288, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 322, + 288, + 395 + ], + "type": "text", + "content": " predicts the object mask probability for every layer. Rather than simply normalizing the layer probabilities of each pixel and directly using them as layer masks, we build a front-to-back hierarchy. That is, for layer " + }, + { + "bbox": [ + 46, + 322, + 288, + 395 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 322, + 288, + 395 + ], + "type": "text", + "content": ", its probability of object mask is computed by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 403, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 403, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 111, + 403, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\alpha_ {n} = a _ {n} \\cdot \\prod_ {i = n + 1} ^ {N} \\left(1 - a _ {i}\\right), \\tag {1}", + "image_path": "d9c6eee5c490562140e4c0ed88fe16d93e47ebb7be16653ee95acf42aeead083.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "inline_equation", + "content": "a \\in [0,1]" + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "text", + "content": " is the raw output of " + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "inline_equation", + "content": "a_N" + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "text", + "content": " fixed and equal to 1. This hierarchy assumes that objects follow a strict front-to-back ordering, which also implies that layer " + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "text", + "content": " represents the background. Here, we specifically focus on scenes with only one foreground object. That is, " + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 46, + 446, + 287, + 533 + ], + "type": "text", + "content": ". We show an example of decomposing multiple foreground objects in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 151, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 151, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 151, + 552 + ], + "type": "text", + "content": "3.2. Texture Mapping" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "content": "We associate each object with its own texture, such that modifying a single pixel in the texture will impact the object's appearance in the rendered video scene. We translate the spatiotemporal locations in the video into 2D UV coordinates for different layers by a mapping network " + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "content": ". This allows us to sample colors on a texture with the translated coordinates. We follow the approach of coordinate-based neural rendering [16, 23] and use the multi-layer perceptrons as our coordinate translator instead of a textured grid." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Note that the appearance of an object in the scene may change throughout the video due to camera motion and illumination variations. We will explain how we address this issue with the proposed multiplicative residual in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "The textures can be further obtained by grid-sampling the texture network " + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": ". Once the texture is extracted, the texture network can be replaced by the extracted texture using UV coordinates from the mapping network, achieving the same effect via bilinear interpolation. This process enables users to modify and apply the extracted texture to the entire video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 156, + 489, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 156, + 489, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 489, + 168 + ], + "type": "text", + "content": "3.3. Multiplicative-Residual Estimator" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 175, + 547, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 547, + 282 + ], + "type": "text", + "content": "In order to apply a single modification to all video frames, the texture must remain constant throughout the entire video. However, the color at the same location on an object may change over time due to variations in lighting conditions, resulting in low reconstruction quality or noisy textures. To address this issue and ensure high-quality reconstruction, we have developed a new method, multiplicative-residual estimator, to achieve constant texture across all frames while preserving illumination variations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 392 + ], + "type": "text", + "content": "With this design, we can also synthesize illumination coefficients not present in the original video at specific time points, allowing us to manipulate camera motion. Note that the prediction of the residual estimator is sharp. Therefore, we have designed losses to constrain the seen and unseen areas, as will be described in Sec. 3.5. The multiplicative-residual estimator predicts the illumination coefficient of each color channel on the texture map, which is multiplied by each channel to obtain the shaded color." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 393, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 430 + ], + "type": "text", + "content": "To provide more insight into our decision to use a multiplicative residual instead of an additive residual, we examine the process of diffuse shading:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 405, + 442, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 442, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 405, + 442, + 545, + 455 + ], + "type": "interline_equation", + "content": "C = \\rho L, \\tag {2}", + "image_path": "15248fc5467a4d0a51ff9737c197a3bae93a23b9870a7ab96bd2949754e3ce4a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": " is the diffuse albedo term determined by surface material and " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": " is the lighting term determined by the environment. When the lighting condition changes or a shadow is cast onto the surface, the new shaded color then becomes " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "C' = \\rho \\cdot L'" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": ". With multiplicative residual, we can reproduce " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": " by scaling " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\frac{L'}{L}" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": ". In contrast, we have to add " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\rho (L' - L)" + }, + { + "bbox": [ + 304, + 467, + 547, + 586 + ], + "type": "text", + "content": " if we use additive residual. One can observe that the additive residual is more difficult to model as it also depends on the diffuse albedo term. We detail the practical result of the two choices in Sec. 4.6." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 597, + 432, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 432, + 609 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 432, + 609 + ], + "type": "text", + "content": "3.4. Network Architecture" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "We design our model as an end-to-end framework trained via self-supervision. We take a spatiotemporal coordinate " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "p = (x, y, t)" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " in the video as the model input and predict the reconstructed color of the corresponding pixel. Fig. 3 shows the pipeline of our model, which includes a mapping network, a texture network, and a multiplicative-residual estimator for each layered representation, with a shared alpha network generating the soft masks for different layers." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7745" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 532, + 247 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 532, + 247 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 532, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 532, + 247 + ], + "type": "image", + "image_path": "52225fd8bcd850d0b1e744bc521de56a1fc70f52c84fda14212a25e00e514ccb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": "Figure 3: Model pipeline. Our model takes a video coordinate " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "p = (x, y, t)" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": " as input and decomposes the video into multiple layers. A representation for layer " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": " is modeled by a mapping network " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": ", a texture network " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": ", and a multiplicative-residual estimator " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": "; the three modules jointly convert a video coordinate " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": " into layered color " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "c_{n}^{(p)}" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": " and lighting coefficient " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "l_{n}^{(p)}" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": ". An alpha network " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": " also takes " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": " as input and predicts soft object masks " + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "inline_equation", + "content": "\\alpha^{(p)}" + }, + { + "bbox": [ + 46, + 250, + 548, + 313 + ], + "type": "text", + "content": " for each layer. The final video color is reconstructed by the bottom-right equation. The multiplicative residuals are critical to our model for handling illumination variations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "text", + "content": "Mapping network. Following the design choice of recent coordinate-based approaches [16, 23], we build our network architecture with multi-layer perceptrons. For each layer " + }, + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "text", + "content": ", we pass a spatiotemporal coordinate " + }, + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "text", + "content": " through a mapping network to translate the video coordinate into a time-independent 2D texture-map coordinate " + }, + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 46, + 332, + 288, + 405 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 408, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 408, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 75, + 408, + 287, + 423 + ], + "type": "interline_equation", + "content": "(u, v) _ {n} ^ {(p)} = \\mathcal {M} _ {n} (p), \\quad (u, v) _ {n} ^ {(p)} \\in [ - 1, 1 ] ^ {2}, \\tag {3}", + "image_path": "7b14366763ffe56d080c2bb2cfc88f42c65d53e4d27e9222efb9416d2e2a32fa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 428, + 235, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 235, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 235, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 428, + 235, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_n" + }, + { + "bbox": [ + 47, + 428, + 235, + 440 + ], + "type": "text", + "content": " is the mapping network of layer " + }, + { + "bbox": [ + 47, + 428, + 235, + 440 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 428, + 235, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 453, + 288, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 538 + ], + "type": "text", + "content": "Texture network and multiplicative-residual estimator. After obtaining the UV coordinates, the texture network takes the UV coordinates " + }, + { + "bbox": [ + 46, + 453, + 288, + 538 + ], + "type": "inline_equation", + "content": "(u,v)_n^{(p)}" + }, + { + "bbox": [ + 46, + 453, + 288, + 538 + ], + "type": "text", + "content": " and produces the corresponding color at each position, while the multiplicative-residual estimator uses both the UV coordinates and the time " + }, + { + "bbox": [ + 46, + 453, + 288, + 538 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 453, + 288, + 538 + ], + "type": "text", + "content": " as inputs to predict the corresponding illumination coefficients at each position:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 544, + 247, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 544, + 247, + 563 + ], + "spans": [ + { + "bbox": [ + 85, + 544, + 247, + 563 + ], + "type": "interline_equation", + "content": "c _ {n} ^ {(p)} = \\mathcal {T} _ {n} \\left((u, v) _ {n} ^ {(p)}\\right), \\quad c _ {n} ^ {(p)} \\in [ 0, 1 ] ^ {3},", + "image_path": "e4d952f017778a7031d0edfab82955bd88286d4b5cf1f44359e441cab308f301.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 563, + 247, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 563, + 247, + 585 + ], + "spans": [ + { + "bbox": [ + 87, + 563, + 247, + 585 + ], + "type": "interline_equation", + "content": "\\ell_ {n, t} ^ {(p)} = \\mathcal {R} _ {n} \\left((u, v) _ {n} ^ {(p)}, t\\right), \\quad \\ell_ {n} ^ {(p)} \\in \\mathbb {R} _ {+} ^ {3},", + "image_path": "af9c8f4c194edaead01dc4bd705cf43d0eaff66254bae763bf5fce637f0a25bb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 589, + 288, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 288, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 589, + 288, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 589, + 288, + 612 + ], + "type": "text", + "content": " is the texture network and " + }, + { + "bbox": [ + 46, + 589, + 288, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 589, + 288, + 612 + ], + "type": "text", + "content": " is the multiplicative-residual estimator." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 627, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 287, + 651 + ], + "type": "text", + "content": "Alpha network. We also predict an object mask for each layer to indicate the layer to which each pixel belongs:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 95, + 654, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 654, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 95, + 654, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} ^ {(p)} = \\mathcal {A} (p), \\quad \\boldsymbol {\\alpha} ^ {(p)} \\in [ 0, 1 ] ^ {N + 1}, \\tag {5}", + "image_path": "2170012dd0147f4d1450a74b69a25470e400cd6c9d313471ceab1e3dd05c0336.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\alpha_{n}^{(p)}" + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "text", + "content": " is the probability that pixel " + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "text", + "content": " belongs to layer " + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\sum_{n=0}^{N} \\alpha_{n}^{(p)} = 1" + }, + { + "bbox": [ + 47, + 674, + 288, + 712 + ], + "type": "text", + "content": ". The derivation of the object mask is detailed in Sec. 3.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 332, + 547, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 332, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 547, + 369 + ], + "type": "text", + "content": "Pixel color reconstruction. After obtaining the information described above, we can reconstruct the pixel color at position " + }, + { + "bbox": [ + 305, + 332, + 547, + 369 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 332, + 547, + 369 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 377, + 546, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 377, + 546, + 410 + ], + "spans": [ + { + "bbox": [ + 367, + 377, + 546, + 410 + ], + "type": "interline_equation", + "content": "\\hat {c} ^ {(p)} = \\sum_ {n = 0} ^ {N} c _ {n} ^ {(p)} \\cdot \\ell_ {n} ^ {(p)} \\cdot \\alpha_ {n} ^ {(p)}. \\tag {6}", + "image_path": "489bec08158c19c3bf37c9f7c7e11cd1b835a677049e9422baa461558a80b6c6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 419, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 419, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 546, + 468 + ], + "type": "text", + "content": "Inspired by [24], we use hash grid encoding for all networks except for the mapping network, as texture coordinates are expected to be smooth. This approach is adopted to facilitate better convergence during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 475, + 382, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 382, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 382, + 487 + ], + "type": "text", + "content": "3.5. Loss Terms" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "content": "Losses inherited from previous work. We incorporate several losses from Layered Neural Atlases [16] to facilitate the training process. These losses include 1) optical flow loss, which provides the model with the supervision of optical flow, and 2) alpha bootstrapping loss, which offers supervision of initial coarse masks; additionally, we use 3) reconstruction loss to stabilize the training and improve reconstruction quality, and 4) sparsity loss to avoid duplicate foreground objects in the texture area. Further details can be found in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": "Residual consistency loss. In order to produce significant yet smooth lighting conditions, the multiplicative residuals of different times " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": " of the same position on the texture coordinate " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "(u, v, t_1)" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": " corresponds to a visible area and " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "(u, v, t_2)" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": " corresponds to an invisible area, should be close in distribution and can have different intensities of illumination. Formally, we have a small " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": " patch " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "P" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7746" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "at time " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " on the video centered at " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". We then sample their texture coordinates " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "(u, v)^{(P)} = \\mathcal{M}(P)" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". The texture coordinates are then combined with different times " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " to get the corresponding multiplicative residuals:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 112, + 128, + 287, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 128, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 112, + 128, + 287, + 153 + ], + "type": "interline_equation", + "content": "\\ell^ {(P)} = \\mathcal {R} ((u, v) ^ {(P)}, t _ {1}), \\tag {7}", + "image_path": "4123136cd03733fcdc1bba4bc9aabf660ef5ab987dc402d271601c78ac2b6224.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 150, + 222, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 150, + 222, + 169 + ], + "spans": [ + { + "bbox": [ + 112, + 150, + 222, + 169 + ], + "type": "interline_equation", + "content": "\\ell^ {\\prime (P)} = \\mathcal {R} \\left((u, v) ^ {(P)}, t _ {2}\\right).", + "image_path": "5c2e793eb2213b02f90654eb9415a64a017cdd06c90aad51624b3406e7e99d52.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 176, + 287, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 224 + ], + "type": "text", + "content": "To this end, we encourage multiplicative residuals of the same position on texture coordinate should be positively correlated. Therefore, we define the residual consistency loss through normalized cross-correlation by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 231, + 287, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 231, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 75, + 231, + 287, + 259 + ], + "type": "interline_equation", + "content": "\\psi (P, t _ {2}) = \\frac {\\left(\\ell^ {(P)} - \\mu_ {\\ell (P)}\\right) \\left(\\ell^ {\\prime (P)} - \\mu_ {\\ell^ {\\prime} (P)}\\right)}{\\sigma_ {\\ell (P)} \\sigma_ {\\ell^ {\\prime} (P)}}. \\tag {8}", + "image_path": "9ca5536a680f2ce79504bd20757d6e90cce95babe1daf2100bf27d17ce7dddfe.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 289 + ], + "type": "text", + "content": "We further introduce a variance-smoothness term that smoothes the changes in lighting conditions as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 128, + 297, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 297, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 128, + 297, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\mathbb {E} (P, t _ {2}) = \\sigma_ {\\ell^ {\\prime} (P)} ^ {2}. \\tag {9}", + "image_path": "603cfd96db0bb0df2b9dcb670bfe06f82e50c9af819438f067a503aeec41cb9d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 319, + 286, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 286, + 342 + ], + "type": "text", + "content": "We apply the above terms to all layers and get the overall loss as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 343, + 287, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 343, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 111, + 343, + 287, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {R} \\text {c o n}} = \\lambda_ {\\mathcal {R} \\text {c o n}} (\\psi + \\beta \\mathbb {E}), \\tag {10}", + "image_path": "bec4ea74ddad0dafb1506821bd31d1a46e17a9863ceb77df3d59ce7d416124ee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "text", + "content": "where we choose " + }, + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "text", + "content": " since our patch size is " + }, + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\beta = 16" + }, + { + "bbox": [ + 46, + 361, + 287, + 383 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 399, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 506 + ], + "type": "text", + "content": "Residual regularization. Minimizing the current losses can collapse to a trivial solution because the multiplicative-residual estimator absorbs all colors. We regularize the residuals equal to 1 since \"true color\" of an object in the video depends on the overall light conditions. That is, if the light source is a pure blue light in the entire video shining on a white wall, we can only say that the steady color of the wall is \"blue\", as we do not have enough information to determine the true color of the wall." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 514, + 287, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 514, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 287, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {R} \\text {r e g}} = \\lambda_ {\\mathcal {R} \\text {r e g}} | | \\mathcal {R} (\\cdot) - 1 | | _ {2} ^ {2}. \\tag {11}", + "image_path": "6156e87eaf99bf77acd6be00def1e56457a1a102e05100bb84789eb2c7c080f5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 541, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 661 + ], + "type": "text", + "content": "Alpha regularization. We introduce an additional regularization term to address potential issues with the alpha network. Specifically, the sparsity loss used in the alpha network tends to assign a value of zero (i.e., black) to regions that are not visible in the input video, which can result in noisy masks and incorrect shadow embedding. To mitigate this problem, we enforced a constraint that each pixel should contribute to at most one layer. This ensures that the mask for each layer is clean and reliable and that the lighting conditions are properly embedded in each layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 668, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 668, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 86, + 668, + 287, + 694 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\alpha \\operatorname {r e g}} = \\lambda_ {\\alpha \\operatorname {r e g}} \\mathrm {B C E} \\left(\\max _ {n \\in \\{0, \\dots , N \\}} \\alpha_ {n}\\right), \\tag {12}", + "image_path": "432046c45d00ffd8749ad8c2205ba831b50732b9f571f3c6532415206abebe61.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 701, + 190, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 190, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 190, + 713 + ], + "type": "text", + "content": "where BCE is binary cross entropy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 425, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 425, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 425, + 85 + ], + "type": "text", + "content": "3.6. Hash Grid Encoding" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 91, + 547, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 294 + ], + "type": "text", + "content": "Like [16, 23], we have considered employing positional encoding as the input encoding for the multi-layer perceptrons. The purpose of including the positional encoding is to enrich the discriminative power of the coordinates and ensure that high-frequency details can be adequately represented. Inspired by [24], we choose to adopt hash grid encoding as the input encoding method for our model. By leveraging this technique, the input features are encoded into a set of intermediate representations that span a broad spectrum of spatiotemporal resolutions, ranging from coarse-to fine-grained, and can flexibly adjust to distinct regions while maintaining high levels of accuracy and consistency. Formally, the 2D or 3D input feature is treated as a coordinate of multi-resolution grids and then used to sample data from the grids using interpolation techniques. The resulting sampled data are concatenated and passed into multi-layer perceptrons to obtain the final output." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 306, + 387, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 387, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 387, + 319 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 326, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 445 + ], + "type": "text", + "content": "We conduct our experiments on the DAVIS dataset [26] and various internet videos to demonstrate the effectiveness of our approach in video reconstruction and consistent video editing. We also design an evaluation metric of edit consistency on the TAP-Vid-DAVIS dataset [8]. In addition, we use our method to generate videos with different camera motions from those in the original input video. Furthermore, we conduct ablation studies on the multiplicative-residual estimator and encoding type to assess the impact on the performance of different architecture choices." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 455, + 417, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 417, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 417, + 467 + ], + "type": "text", + "content": "4.1. Qualitative Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 473, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 581 + ], + "type": "text", + "content": "Fig. 4 shows four qualitative results of our work on the DAVIS dataset [26]. For each video frame, we show the predicted masks (second and third columns), the m-residuals (fourth column), the corresponding layered textures (fifth and sixth columns), and the reconstruction result (last column). Please also refer to Fig. 1 for an example result highlighting the robustness of our m-residual estimation. More results and higher resolution reconstruction can be found in the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 700 + ], + "type": "text", + "content": "With merely the guidance from the coarse object mask provided by Mask R-CNN [11], our model successfully separates the object from the background yielding a clean and precise mask. In the second video in Fig. 4, we showcase an example of a video occluded by complex water splashes. Our model can accurately reconstruct the video without introducing noticeable texture distortions. We superimpose a rainbow checkerboard on the mask to visualize the transformation of each layer. It is evident that each coordinate's component remains at the same position in the texture." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Since the multiplicative residual may brighten or darken" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7747" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 534, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 534, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 534, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 534, + 365 + ], + "type": "image", + "image_path": "f69de98daf2a54b7b55ab918e6b9a533c9a408e6079f1099873d8a1514292787.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 547, + 393 + ], + "type": "text", + "content": "Figure 4: Qualitative results on the DAVIS dataset. We show our predicted masks, textures, m-residuals, reconstructed results, and PSNR of different videos. We overlay a color checkerboard on the masked areas to visualize the texture transformations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 414, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 289, + 559 + ], + "type": "text", + "content": "the texture, we display the m-residual output on a gray canvas to visualize its effect. In the first video, our m-residual estimator successfully models the shadow of the hiker while the background texture retains its original color. The m-residual estimator takes both the texture coordinate and time as the input, allowing it to learn how to model the changes in lighting conditions through space-time and adjust the color accordingly. Such a mechanism enables our model to generate realistic and consistent shading effects for different objects in the video, such as the shading on the bear's fur or the reflection on the hiker's hat. We also report the PSNR of each video as evidence of high-quality reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 563, + 224, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 224, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 224, + 576 + ], + "type": "text", + "content": "4.2. Comparison with Previous Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "We report the PSNR values and various settings for every approach in Table 1. We also provide more quantitative comparisons with DeformableSprites (DS) [39] and Layered Neural Atlases (LNA) [16] under multiple metrics in Table 2. Our method demonstrates superior performance of PSNR, rendering speed, and GPU memory consumption compared to previous methods on all three videos while maintaining the same or better resolution. Note that we could not run DeformableSprites for the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "768 \\times 432" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " resolution, as DeformableSprites failed to finish due to their GPU memory usage exceeding the limit of a single GPU with" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 414, + 547, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 535 + ], + "type": "text", + "content": "24 GB memory. In contrast, our method can efficiently decompose higher-resolution videos into full-HD layers using affordable computation resources while maintaining comparable reconstruction results. To compute the PSNR, we scale the input video up to the reconstructed resolution. Therefore, the PSNR of our 1080p model might be slightly lower than that of the " + }, + { + "bbox": [ + 304, + 414, + 547, + 535 + ], + "type": "inline_equation", + "content": "768 \\times 432" + }, + { + "bbox": [ + 304, + 414, + 547, + 535 + ], + "type": "text", + "content": " model, as the PSNR of the 1080p model is calculated at the corresponding resolution and considered harder due to having more details to reconstruct than the low-resolution counterpart." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 539, + 446, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 446, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 446, + 553 + ], + "type": "text", + "content": "4.3. Consistent Video Editing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 558, + 547, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 630 + ], + "type": "text", + "content": "For editing purposes, we use a " + }, + { + "bbox": [ + 304, + 558, + 547, + 630 + ], + "type": "inline_equation", + "content": "1000 \\times 1000" + }, + { + "bbox": [ + 304, + 558, + 547, + 630 + ], + "type": "text", + "content": " grid sampling to render the texture networks. We then modify the rendered texture and sample color via bilinear interpolation. As our key advantage lies in the m-residual representation of illumination, we have focused our editing on videos or components that exhibit changes in lighting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "content": "We show our editing results in Fig. 5. We choose three videos to present the rendering quality. The first video is sourced from internet videos, while the other two are taken from the DAVIS dataset [26]. For evaluating the editing quality, we modify the background texture of the first two videos by adding handwritten characters of \"ICCV\". We also adjust the color of the shirt of the hiker in the second video." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7748" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 83, + 144, + 133 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 133, + 83 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 133, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 133, + 83 + ], + "type": "text", + "content": "(a) Input frames" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 83, + 144, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 83, + 144, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 83, + 144, + 133 + ], + "type": "image", + "image_path": "0b5ae996a36d23ea05bf981743f1c1df593a8a6dfb3bbfd4df611667fe2047c7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 134, + 144, + 184 + ], + "blocks": [ + { + "bbox": [ + 56, + 134, + 144, + 184 + ], + "lines": [ + { + "bbox": [ + 56, + 134, + 144, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 144, + 184 + ], + "type": "image", + "image_path": "3e1d61ece7af3520bff1e907b8c311227672d654b247b7afc181db92333f8ea8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 185, + 144, + 235 + ], + "blocks": [ + { + "bbox": [ + 56, + 185, + 144, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 185, + 144, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 144, + 235 + ], + "type": "image", + "image_path": "0c5b2cec9fa357f17c05496ac7b8ca69792aa5a530a623b41ff7459c7a333990.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 236, + 144, + 285 + ], + "blocks": [ + { + "bbox": [ + 56, + 236, + 144, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 144, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 144, + 285 + ], + "type": "image", + "image_path": "c1ec89e8585bf3dc67a942fd93b9c84ff887645ff1d9d623a4ee70b4107f1e43.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 286, + 144, + 335 + ], + "blocks": [ + { + "bbox": [ + 56, + 286, + 144, + 335 + ], + "lines": [ + { + "bbox": [ + 56, + 286, + 144, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 144, + 335 + ], + "type": "image", + "image_path": "fb1e93478c37d6d36f3ff5dcf9dcc05d7541e3031bf5fa0f249469a921fca723.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 336, + 144, + 386 + ], + "blocks": [ + { + "bbox": [ + 56, + 336, + 144, + 386 + ], + "lines": [ + { + "bbox": [ + 56, + 336, + 144, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 144, + 386 + ], + "type": "image", + "image_path": "169d459c2ecf5f31748f113496a4f4cae1953973bd5fd9e8acc6982b5d74e325.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 390, + 547, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 547, + 426 + ], + "type": "text", + "content": "Figure 5: Edits on videos with varying lighting conditions. We apply various editing techniques on different components of the videos to observe the corresponding changes in the lighting conditions. The textures with edits are shown in (c), while the editing results are enlarged in the right column of (d). The lighting and shading fuse with the edits flawlessly." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 145, + 83, + 235, + 133 + ], + "blocks": [ + { + "bbox": [ + 151, + 72, + 228, + 82 + ], + "lines": [ + { + "bbox": [ + 151, + 72, + 228, + 82 + ], + "spans": [ + { + "bbox": [ + 151, + 72, + 228, + 82 + ], + "type": "text", + "content": "(b) Reconstruction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 83, + 235, + 133 + ], + "lines": [ + { + "bbox": [ + 145, + 83, + 235, + 133 + ], + "spans": [ + { + "bbox": [ + 145, + 83, + 235, + 133 + ], + "type": "image", + "image_path": "e57df96c2fe4c15a221a70997cf63a3314a3c171bafa80db868fde585c8f06d8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 145, + 134, + 235, + 184 + ], + "blocks": [ + { + "bbox": [ + 145, + 134, + 235, + 184 + ], + "lines": [ + { + "bbox": [ + 145, + 134, + 235, + 184 + ], + "spans": [ + { + "bbox": [ + 145, + 134, + 235, + 184 + ], + "type": "image", + "image_path": "4b67b1e8f2d0b569625f7fa9ab7db15dfb9e459391420e33f21a7460b8132944.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 145, + 185, + 235, + 235 + ], + "blocks": [ + { + "bbox": [ + 145, + 185, + 235, + 235 + ], + "lines": [ + { + "bbox": [ + 145, + 185, + 235, + 235 + ], + "spans": [ + { + "bbox": [ + 145, + 185, + 235, + 235 + ], + "type": "image", + "image_path": "239bccd5f35a7b7cad2992df4e1095b55e4b3effb536400f0225072fb9ad1788.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 145, + 236, + 235, + 285 + ], + "blocks": [ + { + "bbox": [ + 145, + 236, + 235, + 285 + ], + "lines": [ + { + "bbox": [ + 145, + 236, + 235, + 285 + ], + "spans": [ + { + "bbox": [ + 145, + 236, + 235, + 285 + ], + "type": "image", + "image_path": "4508c8874d08caf9558a01c59a62196c00a665bf1b3d72ee90457a477ec04cf7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 145, + 286, + 235, + 335 + ], + "blocks": [ + { + "bbox": [ + 145, + 286, + 235, + 335 + ], + "lines": [ + { + "bbox": [ + 145, + 286, + 235, + 335 + ], + "spans": [ + { + "bbox": [ + 145, + 286, + 235, + 335 + ], + "type": "image", + "image_path": "b1b123cc8b196cae4f1ee51fa47dcf9da80b212c1daf8255deb341e256f7e6f0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 145, + 336, + 235, + 386 + ], + "blocks": [ + { + "bbox": [ + 145, + 336, + 235, + 386 + ], + "lines": [ + { + "bbox": [ + 145, + 336, + 235, + 386 + ], + "spans": [ + { + "bbox": [ + 145, + 336, + 235, + 386 + ], + "type": "image", + "image_path": "b92f6b643eb3e94000bd3ab2e873508c23d989e005776e2f6324ef5d0366f0a8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 246, + 103, + 351, + 159 + ], + "blocks": [ + { + "bbox": [ + 259, + 72, + 335, + 83 + ], + "lines": [ + { + "bbox": [ + 259, + 72, + 335, + 83 + ], + "spans": [ + { + "bbox": [ + 259, + 72, + 335, + 83 + ], + "type": "text", + "content": "(c) Texture editing" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 103, + 351, + 159 + ], + "lines": [ + { + "bbox": [ + 246, + 103, + 351, + 159 + ], + "spans": [ + { + "bbox": [ + 246, + 103, + 351, + 159 + ], + "type": "image", + "image_path": "de186c67c6d37eb10c8175e65533fe6b7a91bbadb7eb6b0ce429b5daf400b426.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 238, + 186, + 356, + 243 + ], + "blocks": [ + { + "bbox": [ + 238, + 186, + 356, + 243 + ], + "lines": [ + { + "bbox": [ + 238, + 186, + 356, + 243 + ], + "spans": [ + { + "bbox": [ + 238, + 186, + 356, + 243 + ], + "type": "image", + "image_path": "a28123c83715de266bb264d5d7993753b6c8d55e67f0cb8020deb65c35cba6de.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 264, + 245, + 345, + 285 + ], + "blocks": [ + { + "bbox": [ + 264, + 245, + 345, + 285 + ], + "lines": [ + { + "bbox": [ + 264, + 245, + 345, + 285 + ], + "spans": [ + { + "bbox": [ + 264, + 245, + 345, + 285 + ], + "type": "image", + "image_path": "c5b2c0e3f1d466fe55c50db0bea07e0eb6fc70a21c49695d683500a5ba61b926.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 264, + 287, + 344, + 334 + ], + "blocks": [ + { + "bbox": [ + 264, + 287, + 344, + 334 + ], + "lines": [ + { + "bbox": [ + 264, + 287, + 344, + 334 + ], + "spans": [ + { + "bbox": [ + 264, + 287, + 344, + 334 + ], + "type": "image", + "image_path": "1bd04ee50860b8df27a25851b471432485759c269ef98b8b1917837ec312475f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 265, + 336, + 344, + 381 + ], + "blocks": [ + { + "bbox": [ + 265, + 336, + 344, + 381 + ], + "lines": [ + { + "bbox": [ + 265, + 336, + 344, + 381 + ], + "spans": [ + { + "bbox": [ + 265, + 336, + 344, + 381 + ], + "type": "image", + "image_path": "fd277c7238ac925a08bb89dc1ca4bd5708fcbc114ace71bf1c35529389f16b28.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 360, + 83, + 447, + 133 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 503, + 83 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 503, + 83 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 503, + 83 + ], + "type": "text", + "content": "(d) Reconstruction (edited)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 83, + 447, + 133 + ], + "lines": [ + { + "bbox": [ + 360, + 83, + 447, + 133 + ], + "spans": [ + { + "bbox": [ + 360, + 83, + 447, + 133 + ], + "type": "image", + "image_path": "cdb650b79c19a6754ca2b75715c98f00c2f506c7058314b2cb93257f3b81ddb4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 360, + 134, + 447, + 184 + ], + "blocks": [ + { + "bbox": [ + 360, + 134, + 447, + 184 + ], + "lines": [ + { + "bbox": [ + 360, + 134, + 447, + 184 + ], + "spans": [ + { + "bbox": [ + 360, + 134, + 447, + 184 + ], + "type": "image", + "image_path": "81d36b73c367f29bb4ccfde664f4763de21457cf07e7a8a7a8557a9ac524dd33.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 360, + 185, + 447, + 235 + ], + "blocks": [ + { + "bbox": [ + 360, + 185, + 447, + 235 + ], + "lines": [ + { + "bbox": [ + 360, + 185, + 447, + 235 + ], + "spans": [ + { + "bbox": [ + 360, + 185, + 447, + 235 + ], + "type": "image", + "image_path": "0f35b7475758a0fbb1751ecd5da412d41773eeddea453acb2ca1533e94b15f20.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 360, + 236, + 447, + 285 + ], + "blocks": [ + { + "bbox": [ + 360, + 236, + 447, + 285 + ], + "lines": [ + { + "bbox": [ + 360, + 236, + 447, + 285 + ], + "spans": [ + { + "bbox": [ + 360, + 236, + 447, + 285 + ], + "type": "image", + "image_path": "ef89203d8d4142c237b0ce44e03b23880f633c3144737832ae9888349359a74b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 360, + 286, + 447, + 335 + ], + "blocks": [ + { + "bbox": [ + 360, + 286, + 447, + 335 + ], + "lines": [ + { + "bbox": [ + 360, + 286, + 447, + 335 + ], + "spans": [ + { + "bbox": [ + 360, + 286, + 447, + 335 + ], + "type": "image", + "image_path": "f80e9cdd1c668cfa7401e7a427da8cf7945a918b6c706a379bab560313123409.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 360, + 336, + 447, + 386 + ], + "blocks": [ + { + "bbox": [ + 360, + 336, + 447, + 386 + ], + "lines": [ + { + "bbox": [ + 360, + 336, + 447, + 386 + ], + "spans": [ + { + "bbox": [ + 360, + 336, + 447, + 386 + ], + "type": "image", + "image_path": "e79c64c757640c82b77a046f938bf84c3a8eff0b6d71a3ae42959d7149a99bd4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 449, + 83, + 537, + 133 + ], + "blocks": [ + { + "bbox": [ + 449, + 83, + 537, + 133 + ], + "lines": [ + { + "bbox": [ + 449, + 83, + 537, + 133 + ], + "spans": [ + { + "bbox": [ + 449, + 83, + 537, + 133 + ], + "type": "image", + "image_path": "92f42925567da358c7cf59d98f3f1b807eec23c93e0ddcb26c0f49155f44265c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 449, + 134, + 537, + 184 + ], + "blocks": [ + { + "bbox": [ + 449, + 134, + 537, + 184 + ], + "lines": [ + { + "bbox": [ + 449, + 134, + 537, + 184 + ], + "spans": [ + { + "bbox": [ + 449, + 134, + 537, + 184 + ], + "type": "image", + "image_path": "3c44b9b682d8053232812e95cee3d177119fa74a6984890fb9888ab0a7ab80c7.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 449, + 185, + 537, + 235 + ], + "blocks": [ + { + "bbox": [ + 449, + 185, + 537, + 235 + ], + "lines": [ + { + "bbox": [ + 449, + 185, + 537, + 235 + ], + "spans": [ + { + "bbox": [ + 449, + 185, + 537, + 235 + ], + "type": "image", + "image_path": "1f0299942457d08e454a5e9039614f5142fd792a7fcf22dbb955dc83451b79b3.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 449, + 236, + 537, + 285 + ], + "blocks": [ + { + "bbox": [ + 449, + 236, + 537, + 285 + ], + "lines": [ + { + "bbox": [ + 449, + 236, + 537, + 285 + ], + "spans": [ + { + "bbox": [ + 449, + 236, + 537, + 285 + ], + "type": "image", + "image_path": "f7b11340b5d5dd8ac5d606c72c3dc002b8b2778bc82f438af938741e589a5e4b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 449, + 286, + 537, + 336 + ], + "blocks": [ + { + "bbox": [ + 449, + 286, + 537, + 336 + ], + "lines": [ + { + "bbox": [ + 449, + 286, + 537, + 336 + ], + "spans": [ + { + "bbox": [ + 449, + 286, + 537, + 336 + ], + "type": "image", + "image_path": "d8f7098b22ddb80a9c896e8f0c24da5ac3facec20be4098d7dbb09bb79d3c5a1.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 449, + 337, + 537, + 386 + ], + "blocks": [ + { + "bbox": [ + 449, + 337, + 537, + 386 + ], + "lines": [ + { + "bbox": [ + 449, + 337, + 537, + 386 + ], + "spans": [ + { + "bbox": [ + 449, + 337, + 537, + 386 + ], + "type": "image", + "image_path": "1752c6d60a450c7f807a8c36fab32cf1714e2cda2324a9aad16b779aa3bf8654.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 447, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 289, + 639 + ], + "type": "text", + "content": "Finally, we perform style transfer on the third video on both background and object textures. In the first video, despite the complexity of the lighting conditions caused by the Disco ball, our approach can successfully handle and diffuse the light onto the edited region, resulting in high-quality output. Previous approaches fail to achieve such a representation of complex lighting conditions. In the second video, the color of the shirt changes convincingly according to the different lighting conditions in two different frames, while our edit on the background is successfully occluded by the shadow. We can see that in the third video, our edited reconstruction follows the textures and the modified lighting conditions fit seamlessly with our modifications, as shown in (d) of the fifth and sixth rows. In the supplementary material, we also provide additional editing results of various types to further demonstrate the versatility and effectiveness of our method." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 47, + 647, + 278, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 278, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 278, + 660 + ], + "type": "text", + "content": "4.4. Quantitative Results for Editing Consistency" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Our work mainly focuses on consistent video editing. However, famous metrics such as PSNR (as we show in Fig. 4) or other reconstruction quality measures do not necessarily reflect the quality of consistent video editing. We" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "content": "take advantage of TAP-Vid [8], which provides several evaluation metrics and the ground-truth feature tracking on the foreground object. We use the feature points of the first unobstructed frame " + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "inline_equation", + "content": "(f_{0,x}, f_{0,y})" + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "content": " as our base point and convert it into texture coordinate " + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "inline_equation", + "content": "(u_0, v_0)" + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "content": ". For all other frames " + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "content": ", we select the video coordinate whose mapped texture coordinate is closest to " + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "inline_equation", + "content": "(u_0, v_0)" + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "content": " as our prediction for frame " + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "content": ". Therefore, the feature points on the same track should be mapped to the same coordinate on the foreground texture for an ideal rendering result." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "We report our results in Table 3. We select a subset of the TAP-Vid DAVIS dataset for our quantitative analysis. For reference, we also report the results of RAFT [31], as a baseline, on the same videos since our optical flow supervision comes from it. In general, the correspondences obtained by our method are more consistent than those by RAFT, as RAFT considers only neighboring frames while ours works on a unified foreground representation. Specifically, our method achieves higher scores than RAFT on the \"black swan\" and \"kite-surf\" videos and comparable results on the other two videos. The low score on the \"parkour\" video is due to the complex geometric changes in the video, which" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7749" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 69, + 531, + 157 + ], + "blocks": [ + { + "bbox": [ + 64, + 69, + 531, + 157 + ], + "lines": [ + { + "bbox": [ + 64, + 69, + 531, + 157 + ], + "spans": [ + { + "bbox": [ + 64, + 69, + 531, + 157 + ], + "type": "table", + "html": "
MethodResolutionTraining timeRendering speed (fps)GPU memorybearPSNR disco ballhike
DeformableSprites [39]213 × 12010 minutes55 GB22.726.221.5
DeformableSprites [39]427 × 24020 minutes1.612 GB23.626.422.0
Layered Neural Atlases [16]768 × 4325.5 hours0.53 GB27.329.025.2
Ours768 × 43240 minutes7873 GB27.537.725.6
Ours1920 × 108040 minutes715 GB26.935.625.4
", + "image_path": "9b73fc783775dd47815a04cae05f9693cfef563bf3b6111d62f33985abbe2271.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 225, + 282, + 300 + ], + "blocks": [ + { + "bbox": [ + 46, + 160, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 160, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 547, + 222 + ], + "type": "text", + "content": "Table 1: Comparison results. We report the PSNR for each video, along with the corresponding training time, rendering speed (frames per second), and GPU memory usage under different resolutions. Our work achieves better results than prior work on the three videos, achieving faster rendering, lower GPU memory consumption, and higher resolution. Note that we measure the PSNR at the corresponding reconstructed resolution. Therefore the PSNR tends to favor the evaluation at low resolution, explaining why our 1080 reconstructions have slightly lower PSNR than our 480p reconstructions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 225, + 282, + 300 + ], + "lines": [ + { + "bbox": [ + 51, + 225, + 282, + 300 + ], + "spans": [ + { + "bbox": [ + 51, + 225, + 282, + 300 + ], + "type": "table", + "html": "
Methodbeardisco ballhike
PSNRLPIPSSSIMPSNRLPIPSSSIMPSNRLPIPSSSIM
DS23.60.230.7826.40.210.8922.00.290.68
LNA27.30.190.8529.00.110.9525.20.190.80
Ours27.50.160.8737.70.040.9725.60.160.82
kite-surfcar-turnlibby
DS21.20.340.6222.20.320.6521.60.410.57
LNA28.20.300.7627.50.300.8829.40.310.91
Ours29.20.260.7829.50.230.9129.60.260.92
", + "image_path": "2bb273b5a38f0c502fb1047788a8f4f341f6222ccc1c63b9d53d4aef03dec1a7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 338, + 274, + 413 + ], + "blocks": [ + { + "bbox": [ + 46, + 302, + 287, + 337 + ], + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 337 + ], + "type": "text", + "content": "Table 2: More comparison results. Our model outperforms previous methods in all videos and achieves better scores in all evaluation metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 338, + 274, + 413 + ], + "lines": [ + { + "bbox": [ + 61, + 338, + 274, + 413 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 274, + 413 + ], + "type": "table", + "html": "
oursRAFT
AJ<δxavgOAAJ<δxavgOA
black swan0.810.871.000.390.521.00
parkour0.040.100.480.070.170.48
kite-surf0.480.580.990.250.350.86
cows0.510.610.990.550.650.99
", + "image_path": "3f4a5f0c7673ed548420afc8e0022835e09859ff8dbc818da3779a5a8725247b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 415, + 289, + 499 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 499 + ], + "type": "text", + "content": "Table 3: Quantitative results on TAP-Vid. We test the editing consistency on a subset of TAP-Vid DAVIS and show the result for each video. The reported metrics, the higher the better, are average Jaccard (AJ), average position accuracy " + }, + { + "bbox": [ + 46, + 415, + 289, + 499 + ], + "type": "inline_equation", + "content": "(< \\delta_{avg}^{x})" + }, + { + "bbox": [ + 46, + 415, + 289, + 499 + ], + "type": "text", + "content": ", and binary occlusion accuracy (OA). Compared to RAFT [31], our method performs better or reaches comparable results on all videos." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 520, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 557 + ], + "type": "text", + "content": "our current architecture can reconstruct well but edit poorly on the foreground parkour runner. Nevertheless, the editing on the background still exhibits comparable quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 211, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 211, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 211, + 576 + ], + "type": "text", + "content": "4.5. Manipulating Camera Motion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "With our multiplicative residual estimator, we are able to synthesize realistic views that are not shown in the input video. A different camera view can be synthesized by scaling, shifting, or rotating the original video coordinates " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(x^{\\prime},y^{\\prime})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and concatenating it with the temporal information " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". Once the new set of video coordinates " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "p^\\prime = (x',y',t)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is obtained, we can use the mapping network " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " to obtain the texture coordinate " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(u,v)^{(p^{\\prime})}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " for each layer. The color " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "c^{(p^{\\prime})}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and the lighting " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\ell^{(p^{\\prime})}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " of all layers can be obtained based on these texture coordinates. For each non-background layer " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", we use the alpha network " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 226, + 386, + 281 + ], + "blocks": [ + { + "bbox": [ + 306, + 226, + 386, + 281 + ], + "lines": [ + { + "bbox": [ + 306, + 226, + 386, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 226, + 386, + 281 + ], + "type": "image", + "image_path": "3a5ad7ee38f10641c2fb093b108f0874b6bcf2ab6dece28ab5521535df97851b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 281, + 385, + 326 + ], + "blocks": [ + { + "bbox": [ + 307, + 281, + 385, + 326 + ], + "lines": [ + { + "bbox": [ + 307, + 281, + 385, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 385, + 326 + ], + "type": "image", + "image_path": "3a3b85874d44b80ed5405f295a43e47b72724dcaa649fc7b1d5e962f23da02bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 327, + 385, + 369 + ], + "blocks": [ + { + "bbox": [ + 307, + 327, + 385, + 369 + ], + "lines": [ + { + "bbox": [ + 307, + 327, + 385, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 385, + 369 + ], + "type": "image", + "image_path": "a15f369b591a546d3f3325ec9fb9cc764d1fe0d773ffd4767c79df8290996b62.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 370, + 385, + 415 + ], + "blocks": [ + { + "bbox": [ + 307, + 370, + 385, + 415 + ], + "lines": [ + { + "bbox": [ + 307, + 370, + 385, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 385, + 415 + ], + "type": "image", + "image_path": "f438d623a1eebd03388bb2bfe6b49c828b4686b3a649e18b3724b5ea21cc311f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 418, + 547, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 547, + 455 + ], + "type": "text", + "content": "Figure 6: Camera motion manipulation. The multiplicative residual serves as a smooth representation that effectively adjusts the lighting conditions with a fixed camera." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 386, + 227, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 386, + 227, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 386, + 227, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 386, + 227, + 545, + 281 + ], + "type": "image", + "image_path": "055f2670cac109144b222d7f0888f383a01a40b022a7889ab940382b667f3bca.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 390, + 281, + 467, + 326 + ], + "blocks": [ + { + "bbox": [ + 390, + 281, + 467, + 326 + ], + "lines": [ + { + "bbox": [ + 390, + 281, + 467, + 326 + ], + "spans": [ + { + "bbox": [ + 390, + 281, + 467, + 326 + ], + "type": "image", + "image_path": "43b1ecd1e9d0619ba062869d67eca4955faf94f52cdd63d98ff7868a9a1df9a2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 390, + 327, + 467, + 370 + ], + "blocks": [ + { + "bbox": [ + 390, + 327, + 467, + 370 + ], + "lines": [ + { + "bbox": [ + 390, + 327, + 467, + 370 + ], + "spans": [ + { + "bbox": [ + 390, + 327, + 467, + 370 + ], + "type": "image", + "image_path": "f0bb7635a0b3f3124267e2c5b3396642f45d810890762a7a38af23a503df0881.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 390, + 371, + 467, + 415 + ], + "blocks": [ + { + "bbox": [ + 390, + 371, + 467, + 415 + ], + "lines": [ + { + "bbox": [ + 390, + 371, + 467, + 415 + ], + "spans": [ + { + "bbox": [ + 390, + 371, + 467, + 415 + ], + "type": "image", + "image_path": "b3eb9aed5d43499c6180a8626edc17f530e7738f9ff28f218d3a91ce6873d41e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 468, + 281, + 545, + 326 + ], + "blocks": [ + { + "bbox": [ + 468, + 281, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 468, + 281, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 468, + 281, + 545, + 326 + ], + "type": "image", + "image_path": "c1d1d33aff99057419d5712d79ff39759bdac132262538fa72201ae4fd80cea9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 468, + 327, + 545, + 370 + ], + "blocks": [ + { + "bbox": [ + 468, + 327, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 468, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 468, + 327, + 545, + 370 + ], + "type": "image", + "image_path": "953a793c7d37c79fbf8e991cf22fb308e1e981d6b1ca4e10bafcb1a712ee6c14.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 468, + 371, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 468, + 371, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 468, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 468, + 371, + 545, + 415 + ], + "type": "image", + "image_path": "c4b160f25385cca27b274242272a594d2f92206ebbd952be10d54933d3e9a0f2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "text", + "content": "obtain the original soft mask " + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\alpha^{(p)}" + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "inline_equation", + "content": "p = (x,y,t)" + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "text", + "content": " and establish a relation as " + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "inline_equation", + "content": "(u,v)_0\\rightarrow \\alpha_n^{(p)}" + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "inline_equation", + "content": "(u,v)_0 = \\mathcal{M}_0(p)" + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "text", + "content": ". Then, we linearly interpolate the value of the mask at each location " + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "inline_equation", + "content": "p^\\prime" + }, + { + "bbox": [ + 304, + 481, + 547, + 567 + ], + "type": "text", + "content": " by triangulating the input data. That is, we establish a correspondence between the coordinates of the object textures and the background texture and then transfer the mask values from the object onto the background texture." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "We set up a virtual camera in the background to capture the entire video scene from a fixed position. Fig. 6 shows our result. The camera is fixed (i.e., the background remains static) while the object keeps moving in the scene. In the fourth video, the car is partially visible in the beginning frames and fully visible later, with its side facing the camera. The third video demonstrates that the splashes consistently follow the entire path of the original camera view. With our multiplicative-residual estimator, we can successfully reconstruct the surfing scene while accurately representing the water splashes. We show more results with varying camera motions in the supplementary material." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7750" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 286, + 287 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 287 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 287 + ], + "type": "image", + "image_path": "c0d721cefee8d00bed3918942b5ad40d62f986f69e38140bf03cdcd7d115a456.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 289, + 288, + 337 + ], + "lines": [ + { + "bbox": [ + 47, + 289, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 288, + 337 + ], + "type": "text", + "content": "Figure 7: Choice of the residual estimator architecture. Using an additive residual estimator architecture may introduce artifacts or degrade the quality of textures, while a multiplicative one works correctly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 343, + 168, + 449 + ], + "blocks": [ + { + "bbox": [ + 49, + 343, + 168, + 449 + ], + "lines": [ + { + "bbox": [ + 49, + 343, + 168, + 449 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 168, + 449 + ], + "type": "image", + "image_path": "062d31978c458d533b9e463af99bf2f657b04eeeeaf495762c75c788129f8754.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 455, + 288, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 455, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 288, + 514 + ], + "type": "text", + "content": "Figure 8: Quantitative ablation studies. (a) The reconstruction PSNR degradation of the additive residual is significantly higher than that of the multiplicative residual. (b) Hash grid encoding can achieve better PSNR in fewer iterations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 344, + 282, + 449 + ], + "blocks": [ + { + "bbox": [ + 167, + 344, + 282, + 449 + ], + "lines": [ + { + "bbox": [ + 167, + 344, + 282, + 449 + ], + "spans": [ + { + "bbox": [ + 167, + 344, + 282, + 449 + ], + "type": "image", + "image_path": "0593c1673a69907154c2b783be94fd739c9271c943bf2d01221c7475ed67b902.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 115, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 115, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 115, + 550 + ], + "type": "text", + "content": "4.6. Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "content": "Residual type. We present an ablation study to verify the choice of the architecture of our residual estimator. As mentioned in Sec. 3.3, we adopt the multiplicative residual instead of the additive residual. We show the comparisons of the two designs in Fig. 7 and Fig. 8a. The additive residual estimator is conducted, and the output is normalized to " + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 46, + 557, + 289, + 713 + ], + "type": "text", + "content": " to ensure they have the same representation capability. In Fig. 7, both additive and multiplicative residuals provide promising reconstruction results for the first video. However, when it comes to the complex texture, like the background of the second video, the additive residual fails to represent it and compensates for the wrong illuminations. In the third video, the additive residual fails to represent precise" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "background texture, and the dog vanishes from the texture. Fig. 8a demonstrates a significant drop in PSNR when ablating the residual module. This finding reflects that additive residual is prone to overfit the color that should be attributed to the editable texture. In contrast, the multiplicative residual provides complete background and object textures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 158, + 547, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 547, + 229 + ], + "type": "text", + "content": "Hash grid encoding. We evaluate the effectiveness of hash grid encoding in Fig. 8b. Positional encoding needs more MLP layers to achieve good quality, which takes " + }, + { + "bbox": [ + 304, + 158, + 547, + 229 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 304, + 158, + 547, + 229 + ], + "type": "text", + "content": " processing time per iteration compared to hash encoding. However, hash grid encoding still achieves better PSNR in fewer iterations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 239, + 375, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 375, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 375, + 251 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 259, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 437 + ], + "type": "text", + "content": "We have observed several interesting points that are worth discussing. Our method relies on Mask-RCNN and RAFT to provide the foreground mask and optical flow as external priors. As a result, our method may be biased by the inaccurate learning-based priors on some difficult cases like the spinning lighting in disco ball or the object boundary in bear. The reconstructed objects in the video may exhibit edge artifacts due to referencing the incorrect texture layer. Our method with reconstruction loss could improve the initial external prior but is still not perfect to solve all the artifacts. Adopting better external priors or designing strong internal priors are both good future improvements. The problem of object edge artifacts is similar to the seam artifact in mesh uv-parameterization from which future exploration may take inspiration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 468, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 624 + ], + "type": "text", + "content": "We have presented a neural layer decomposition method to facilitate illumination-aware video editing. The proposed multiplicative-residual estimator can effectively derive the layered representation that characterizes the spatiotemporally varying lighting effects. We have also implemented hash grid encoding for fast coordinate inference. Our model, therefore, significantly reduces the training time and achieves real-time rendering speed with a low requirement of computation resources, enabling interactive editing on high-resolution videos. We use our model to generate high-quality video editing results, where, in particular, the varying illumination effect can only be achieved by ours rather than the previous methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 634, + 415, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 415, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 415, + 647 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": "This work was supported in part by NSTC grants 111-2221-E-001-011-MY2 and 112-2221-E-A49-100-MY3 of Taiwan. We are grateful to National Center for High-performance Computing for providing computational resources and facilities." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7751" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Aseem Agarwala, Aaron Hertzmann, David Salesin, and Steven M. Seitz. Keyframe-based tracking for rotoscoping and animation. ACM Trans. Graph., 23(3):584-591, 2004. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 200 + ], + "type": "text", + "content": "[2] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. In Computer Vision - ECCV 2022 Workshops - Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, volume 13802 of Lecture Notes in Computer Science, pages 204-220. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, João Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5733-5742. IEEE, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 301 + ], + "type": "text", + "content": "[4] Connelly Barnes, Dan B. Goldman, Eli Shechtman, and Adam Finkelstein. Video tapestries with continuous temporal zoom. ACM Trans. Graph., 29(4):89:1-89:9, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "text", + "content": "[5] Zhangxing Bian, Allan Jabri, Alexei A. Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 6498-6509. IEEE, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 401 + ], + "type": "text", + "content": "[6] Gabriel J. Brostow and Irfan A. Essa. Motion based decompositing of video. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 8-13. IEEE Computer Society, 1999. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 424 + ], + "type": "text", + "content": "[7] Carlos D. Correa and Kwan-Liu Ma. Dynamic video narratives. ACM Trans. Graph., 29(4):88:1-88:9, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 426, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 479 + ], + "type": "text", + "content": "[8] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens Continente, Kucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In NeurIPS Datasets Track, 2022. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 512 + ], + "type": "text", + "content": "[9] Zeqi Gu, Wenqi Xian, Noah Snavely, and Abe Davis. Factormatte: Redefining video matting for re-composition tasks. ArXiv, abs/2211.02145, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "text", + "content": "[10] Gregory D. Hager and Peter N. Belhumeur. Efficient region tracking with parametric models of geometry and illumination. IEEE Trans. Pattern Anal. Mach. Intell., 20(10):1025-1039, 1998. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 590 + ], + "type": "text", + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell., 42(2):386-397, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "type": "text", + "content": "[12] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 7044-7052. IEEE Computer Society, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 659, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 287, + 712 + ], + "type": "text", + "content": "[13] Allan Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[14] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 3154-3164. IEEE Computer Society, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[15] Nebojsa Jojic and Brendan J. Frey. Learning flexible sprites in video layers. In 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2001), with CD-ROM, 8-14 December 2001, Kauai, HI, USA, pages 199-206. IEEE Computer Society, 2001. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 218 + ], + "type": "text", + "content": "[16] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 252 + ], + "type": "text", + "content": "[17] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. Int. J. Comput. Vis., 76(3):301-319, 2008. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 547, + 330 + ], + "type": "text", + "content": "[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XV, volume 11219 of Lecture Notes in Computer Science, pages 179-195. Springer, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 332, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 396 + ], + "type": "text", + "content": "[19] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 3809-3817. Computer Vision Foundation / IEEE, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 546, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 546, + 432 + ], + "type": "text", + "content": "[20] Sharon Lin, Matthew Fisher, Angela Dai, and Pat Hanrahan. Layerbuilder: Layer decomposition for interactive image and video color editing. ArXiv, abs/1701.03754, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 434, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 546, + 477 + ], + "type": "text", + "content": "[21] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. ACM Trans. Graph., 39(6):256:1-256:14, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 479, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 547, + 544 + ], + "type": "text", + "content": "[22] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4507-4515. Computer Vision Foundation / IEEE, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 546, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 547, + 623 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 624, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 667 + ], + "type": "text", + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 712 + ], + "type": "text", + "content": "[25] William S. Peebles, Jun-Yan Zhu, Richard Zhang, Antonio Torralba, Alexei A. Efros, and Eli Shechtman. Gan-supervised dense visual alignment. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Or" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7752" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "text", + "content": "leans, LA, USA, June 18-24, 2022, pages 13460-13471. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 172 + ], + "type": "text", + "content": "[26] Federico Peruzzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus H. Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 724-732. IEEE Computer Society, 2016. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 194 + ], + "type": "text", + "content": "[27] Hanspeter Pfister. Interactive intrinsic video editing. ACM Trans. Graph., 33(6):197:1-197:10, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 228 + ], + "type": "text", + "content": "[28] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew W. Fitzgibbon. Unwrap mosaics: a new representation for video editing. ACM Trans. Graph., 27(3):17, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "text", + "content": "[29] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Pattern Recognition - 38th German Conference, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings, volume 9796 of Lecture Notes in Computer Science, pages 26-36. Springer, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 288, + 358 + ], + "type": "text", + "content": "[30] Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Background matting: The world is your green screen. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 2288-2297. Computer Vision Foundation / IEEE, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "text", + "content": "[31] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12347 of Lecture Notes in Computer Science, pages 402-419. Springer, 2020. 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "text", + "content": "[32] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, SA 2022, Daegu, Republic of Korea, December 6-9, 2022, pages 29:1-29:9. ACM, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "type": "text", + "content": "[33] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. IEEE Trans. Image Process., 3(5):625-638, 1994. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 536 + ], + "type": "text", + "content": "[34] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 29:9125-9139, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 592 + ], + "type": "text", + "content": "[35] Josh Wills, Sameer Agarwal, and Serge J. Belongie. What went where. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2003), 16-22 June 2003, Madison, WI, USA, pages 37-44. IEEE Computer Society, 2003. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 593, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 288, + 647 + ], + "type": "text", + "content": "[36] Ning Xu, Brian L. Price, Scott Cohen, and Thomas S. Huang. Deep image matting. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 311-320. IEEE Computer Society, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[37] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, volume 13675 of Lecture Notes in Computer Science, pages 357-374. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 184 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[38] Yanchao Yang, Brian Lai, and Stefano Soatto. Dystab: Unsupervised object segmentation via dynamic-static bootstrapping. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 2826-2836. Computer Vision Foundation / IEEE, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 184 + ], + "type": "text", + "content": "[39] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 2647-2656. IEEE, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7753" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_content_list.json b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d2c34e1e8999c8ce0e275cbbbcaf4d6ffe13f2c1 --- /dev/null +++ b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_content_list.json @@ -0,0 +1,1762 @@ +[ + { + "type": "text", + "text": "Helping Hands: An Object-Aware Ego-Centric Video Recognition Model", + "text_level": 1, + "bbox": [ + 117, + 130, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuhan Zhang \nVGG, University of Oxford \nczhang@robots.ox.ac.uk", + "bbox": [ + 114, + 181, + 334, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ankush Gupta Google DeepMind, London ankushgupta@gmail.com", + "bbox": [ + 372, + 181, + 594, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Zisserman \nVGG, University of Oxford \naz@robots.ox.ac.uk", + "bbox": [ + 633, + 181, + 854, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce an object-aware decoder for improving the performance of spatio-temporal representations on egocentric videos. The key idea is to enhance object-awareness during training by tasking the model to predict hand positions, object positions, and the semantic label of the objects using paired captions when available. At inference time the model only requires RGB frames as inputs, and is able to track and ground objects (although it has not been trained explicitly for this).", + "bbox": [ + 75, + 301, + 470, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We demonstrate the performance of the object-aware representations learnt by our model, by: (i) evaluating it for strong transfer, i.e. through zero-shot testing, on a number of downstream video-text retrieval and classification benchmarks; and (ii) by using the representations learned as input for long-term video understanding tasks (e.g. Episodic Memory in Ego4D). In all cases the performance improves over the state of the art—even compared to networks trained with far larger batch sizes. We also show that by using noisy image-level detection as pseudo-labels in training, the model learns to provide better bounding boxes using video consistency, as well as grounding the words in the associated text descriptions.", + "bbox": [ + 75, + 439, + 468, + 637 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Overall, we show that the model can act as a drop-in replacement for an ego-centric video model to improve performance through visual-text grounding1.", + "bbox": [ + 73, + 638, + 468, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In visual-language models there has been a recent move to explicitly build object awareness into the vision module by adding specialized and bespoke components, or using entirely object-centric architectures. The motivation for this partly comes from the attractive compositional nature of objects and their inter-relationships in language, which enables inexhaustible novel combinations [10, 45], and partly from infant cognitive studies that stress the im", + "bbox": [ + 75, + 741, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "portance of objects in early visual development [29, 56, 60]. Examples in the video domain include explicit internal object representations [2], e.g., through RoI-align [17] pooled features either from a pre-trained region-proposal network (RPN) [2, 52, 57, 62], or from bounding-box coordinates taken as input [19, 42, 48, 71]. This contrasts with the large body of work where standard representations are learnt end-to-end without any explicit factorization into objects/entities, such as dual-encoder vision-language models in the image [21, 49] and video domains [4, 64].", + "bbox": [ + 500, + 270, + 890, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we take a different (middle) path and instead use a vanilla video transformer architecture and induce object-awareness into the video representation by tasking the model to predict object-level properties, such as their localization and semantic categories, only during training.", + "bbox": [ + 498, + 424, + 892, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our target domain is ego-centric video [11, 16], and we tailor the object properties used to this. In ego-centric videos the actor [57] is often present through their hands, and we therefore task the network to predict both the hands and the principal objects they interact with. As will be seen, this simple object-aware training boosts the performance of pre-trained video-language architectures significantly, and leads to state-of-art performance across multiple ego-centric benchmarks. During inference, the model requires only RGB frames as input, and operates as a standard video-language network.", + "bbox": [ + 496, + 502, + 892, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In more detail, our model is built on top of a pre-trained video-language dual encoder architecture (where there are separate encoders for the video and text data). We add an additional, but vanilla, transformer decoder head [61], and train with DETR/Mask2former [7, 9] query vectors and object loss for hands and other objects. The intuition is that these additional query vectors help the model to attend to and track the hands and salient objects in the scene (these are the 'helping hands' of the paper title). Importantly, we do not require dense frame level ground truth for this training. Rather, we obtain somewhat noisy and temporally sparse annotations automatically from a hand object detector [53], and use these to provide prediction targets for the frames where they are available. This opportunistic use of annotations is pragmatic as object detectors trained on third", + "bbox": [ + 496, + 671, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code and models available at: https://github.com/Chuhanxx/helping_hand_for_egocentric Videos", + "bbox": [ + 75, + 875, + 467, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13901", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "person datasets (such as COCO) do not perform so well on the ego-centric domain, where the scenes are more crowded and objects are often small and can be motion blurred. By only requiring annotations for a subset of frames, where they can be reliably produced automatically, we are able to train on large-scale data without requiring expensive manual supervision.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although we train with noisy and sparse image-level object localization, our model can learn to predict better and denser bounding-box trajectories through large-scale training due to the spatio-temporal consistency which naturally presents in videos. Also, it is able to predict semantic grounding by learning to map the object appearance to the nouns in the video captions.", + "bbox": [ + 75, + 198, + 468, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is worth noting that we are using hand detectors because hands are a common and important object in ego-centric videos. However, the object-centric method we are proposing has greater scope than ego-centric videos and can be applied to other scenarios with other object types providing the 'helping-hand'.", + "bbox": [ + 75, + 305, + 468, + 397 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "In summary, we make the following contributions: (i) We propose a method to induce object-awareness in video-language models for an architecture composed of standard neural modules. The model only requires RGB frames as inputs, and thus is a drop-in replacement for any ego-centric video model.", + "(ii) The model can be trained opportunistically using available and sparse frame-level and noisy annotations, produced automatically.", + "(iii) We demonstrate state-of-the-art strong (zero-shot) transfer for cross-modal retrieval to other ego-centric datasets namely, EpicKitchens-MIR and EGTEA improving prior art by $2 - 4\\%$ .", + "(iv) We evaluate the grounding quantitatively using the EpicKitchens-VISOR dataset [11, 12] and find that the model outperforms the base hand-object detector used for training supervision.", + "(v) Finally, we also demonstrate that the representations learned can be used as input in long-term video understanding tasks like EgoNLQ and EgoMQ. The objectiveness in the representation helps the model outperform other models trained on the same training set on these two tasks." + ], + "bbox": [ + 75, + 398, + 468, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 752, + 218, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision and Language Representation Learning. Different from transferring representations learned for classification on a fixed set of object categories [27, 54], recent vision-language pre-training (VLP) works leverage largescale supervision from free-from text descriptions of images and videos. These methods use image captions [49] or video sub-titles [44, 64] with either independent dual encoders for the visual and text modalities [4, 21, 41], or via", + "bbox": [ + 75, + 777, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "joint encoders with cross-attention across modalities [1, 30, 31, 69]. We also use dual-encoders which are kept frozen due to compute limitations. To explicitly build-in object-awareness into image representations, object-level features extracted from pre-trained object-detectors are aligned with the text descriptions [8, 33, 40, 58, 75, 78] The object-level text alignment is further augmented with the object-box prediction task for grounding in [13, 23, 70, 73]. VLP for ego-centric videos has recently been explored [35, 77] to bridge the domain gap between representations learned from third-person videos found commonly [44], and first-person ego-centric videos. We further extend image based object-aware VLP methods to ego-centric videos by training to predict the auto-generated hand-object boxes extracted from pretrained detectors [53], while requiring only RGB input during inference, making our model a drop-in replacement for ego-centric VLP models albeit with enhanced object-awareness.", + "bbox": [ + 496, + 90, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly Supervised Video Text Grounding. A key challenge for grounding in videos is the lack of large-scale object-level annotations for videos. While such annotations are readily available for synthetic datasets [68], expensive manual annotation is required for real videos.", + "bbox": [ + 496, + 388, + 890, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hence, weakly supervised method have been developed which leverage the video sub-titles/description to map nouns/verbs to regions in frames. This is typically achieved by first extracting bounding-box/segmentation regions from pre-trained detectors for objects and humans, and aligning them with keywords using max-margin [79] or constrastive [32] objectives. Similarly, [43] also align words in the video captions to regions from pre-trained RPN by modelling the region-word associations as latent variables of a conditional generative variational auto-encoder. More recently, [59] use cross-attention across text and candidate regions, and find (soft-) associations based on attention weights. While our model similarly absorbs object bounding-boxes and category information obtained from pre-trained object and hand detectors during training to induce object-aware representations, these detectors are not required during inference.", + "bbox": [ + 496, + 465, + 890, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object-Oriented Learning in Videos. Most vision tasks for images involve objects [22, 24, 25, 36]. In videos, there are also a broad range of object-oriented tasks: some works treat learning object-level information as an end task, they design models to predict object bounding-boxes and masks [5, 26, 37, 38, 65, 67, 74]; other works use object knowledge to achieve some other high-level tasks, for instance, object-centric scene generation [28, 66] and decomposition [14, 18, 55], and action recognition [3, 19, 20, 47, 62, 71, 80]. Our method uses object information to as-", + "bbox": [ + 496, + 748, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13902", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/818085653772481dc7168d9d221b78e4ad5a0db93b701241b2a615536ea8ca64.jpg", + "image_caption": [ + "Figure 1. The object-aware vision-language model architecture. The architecture is made up of three parts: A video backbone, a text backbone and a object-aware decoder. The decoder is a cross-attention transformer, it takes the visual feature map $F'$ as keys and values, which are attended by a set of learnable queries. In these permutation-invariant query vectors, the hand and object queries $Q_h, Q_o$ are trained to be object-aware and predict the localization and class of hands and objects. The video query $q_v$ attends to both the visual feature map through cross-attention layers, and the object feature map through self-attention layers in the Transformer decoder and output a video-level embedding $v$ to be matched with the text embedding $t$ ." + ], + "image_footnote": [], + "bbox": [ + 125, + 87, + 844, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sist vision-language alignment in egocentric videos, so that the model learns grounded video representations that can generalize better. SViT [3] uses object queries shared between images and videos in order to predict hand-object bounding boxes in videos, whilst only requiring image-level supervision during training. However, the object queries are not used for vision-language alignment. Our previous work [71] showed that encoding object-level information in the model helps transfer learning in various video understanding tasks, but the model required GT bounding boxes as input during inference. In contrast, in this work we model does not require this information during training, and we show box prediction and vision-language alignment can be combined and benefit both in-domain and out-of-domain tasks.", + "bbox": [ + 75, + 508, + 472, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Object-Aware Vision-Language Learning", + "text_level": 1, + "bbox": [ + 76, + 750, + 449, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first describe the task of object-aware vision-language learning and the architecture of our model. These are followed by the training objectives for vision-language matching and weakly-supervised text grounding.", + "bbox": [ + 76, + 776, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 847, + 187, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In vision-language representation learning the training data consists of short video clips (RGB frames) and an", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "associated free-form text string containing words that describe the visual content. Typically, dual encoders are trained on this paired data, with a visual-encoder that ingests the video clip and a language-encoder that ingests the text [4, 35, 41, 49, 64]. The dual encoders are trained with a contrastive objective [46] such that the cosine similarity for matching vision-text pairs is optimized to be higher than the similarity of negative/not-matching pairs. This pretraining objective enables evaluation on downstream vision-language tasks like video-text retrieval and action classification in a zero-shot manner [49].", + "bbox": [ + 496, + 508, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our object-aware model follows the data, training and evaluation pipeline as above, except that the model is also tasked to output object-level information (e.g., bounding boxes and object categories) during training. By tasking the model to predict object bounding boxes and names which can be matched to the nouns in the narration, the model learns grounded and fine-grained correspondence between the modalities. The object-level prediction is used as an auxiliary objective in training but not used at inference time.", + "bbox": [ + 496, + 681, + 892, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In more detail, as shown in figure 2, there are two types of object-level prediction: (a) hand and object bounding boxes; and (b) object names. Since ground truth of boxes and object names are not available, and most traditional detectors [7, 81] fail to identify objects well in ego-centric", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13903", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "videos, we cannot rely on strong supervision for the predictions. Instead, we generate bounding box targets (for the hands and other objects) using a robust off-the-shelf hand-object detector [53], though these targets will only be available for some of the frames and are noisy. While for object name prediction, we use a weakly supervised method to align the predicted names with nouns in the paired narration (detailed in section 3.3). In both cases, the supervision is opportunistic and only applied when available.", + "bbox": [ + 75, + 90, + 472, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Architecture", + "text_level": 1, + "bbox": [ + 76, + 239, + 210, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dual Encoder. We use dual encoders as our visual and text backbone for efficiency. The visual encoder ingests a video clip $F$ of RGB frames $F = (f_{1}, f_{2}, \\ldots, f_{T})$ , where $f_{i} \\in \\mathbb{R}^{H \\times W \\times 3}$ and $T$ is the number of frames. The clip $F$ is encoded by a Video Transformer [6] which tokenizes the frames by 3D patches to produce a spatially downsampled feature map $F' = (f_{1}', f_{2}', \\ldots, f_{T}')$ , where $f_{i}' \\in \\mathbb{R}^{H' \\times W' \\times C}$ . It outputs a visual embedding $v \\in \\mathbb{R}^{C}$ .", + "bbox": [ + 75, + 262, + 468, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The text encoder is a Transformer [61] that inputs words tokenized by a BPE tokenizer [15]. It encodes two type of inputs: (a) a narration sentence which describes the contents of a clip; and (b) a noun set which contains noun phrases from the narration sentence. At the output, the embedding corresponding to the EOS token is taken to be the embedding for the full sentence $t \\in \\mathbb{R}^C$ , and multiple noun embeddings $n \\in \\mathbb{R}^C$ .", + "bbox": [ + 75, + 386, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object-Aware Module. The object-aware module is a cross-attention Transformer which has a permutation-invariant set of learnable vectors as queries (similar to DETR [7] and Mask2former [9]). The queries are at video-level, shared between frames. They consist of three sets: two hand queries $Q_{h} = (q_{h1}, q_{h2})$ for the left and right hands; $K$ object queries $Q_{o} = (q_{o1}, q_{o2}, \\ldots, q_{ok})$ ; and a video-level query $q_{v}$ . These queries are learned, and attend to the visual feature map $F'$ from the visual backbone and output a set of summary vectors $S = (s_{h1}, s_{h2}; s_{o1}, \\ldots, s_{ok})$ corresponding to each input query.", + "bbox": [ + 75, + 527, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The object-aware module operates on the visual content without any interaction with the text information. It consists of six cross-attention blocks. As in a traditional Transformer decoder [61], in each block, there is a multi-head self-attention layer and a multi-head cross-attention layer. The self-attention layer enables interactions between hand, object and video queries, and the cross-attention layer allows the query to extract object-oriented information from the visual content.", + "bbox": [ + 75, + 696, + 470, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bounding Box Head. The hand query vectors $Q_{h}$ and object query vectors $Q_{o}$ are trained to predict the bounding box of the hands and objects respectively in each frame.", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note these query vectors and summary vectors are at the video level; to predict boxes at frame level, we condition a summary vector $s_j$ of object $j$ on a learnable frame index vector $x_i$ by concatenation of $s_j$ and $x_i$ , and use a multi-layer perceptron $F_{box}$ to project them onto a bounding box $\\hat{b}_{j,1}$ , where $i$ is the frame number:", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {b} _ {j i} = F _ {\\text {b o x}} \\left(s _ {j}; x _ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 193, + 890, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a result, we will have a time series of bounding boxes $(\\hat{b}_{j,1},\\hat{b}_{j,2},\\ldots \\hat{b}_{j,T})$ from each $j^{\\mathrm{th}}$ summary vector.", + "bbox": [ + 500, + 222, + 890, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Head. We assign semantic meanings to object summary vectors, standing for the object name/class. To achieve this, we project $s_j$ onto a word embedding $\\hat{n}$ with a multi-layer perception $F_{semantic}$ :", + "bbox": [ + 496, + 271, + 890, + 333 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {n} = F _ {\\text {s e m a n t i c}} \\left(s _ {j}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 343, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Objectives", + "text_level": 1, + "bbox": [ + 500, + 369, + 687, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Text Matching. We follow EgoVLP [35] and use EgoNCE loss as the objective for matching between video-level embedding $v$ and sentence-level embedding $t$ of the narration. In one batch $\\mathcal{B}$ , the positive sample set $\\mathcal{P}_m$ is made up of a sample $i$ and other samples that share at least one noun and one verb with it: $\\mathcal{P}_m = \\{n \\in \\mathcal{B} \\mid \\text{noun}(n) \\cap \\text{noun}(m) \\neq \\emptyset, \\text{verb}(n) \\cap \\text{verb}(m) \\neq \\emptyset\\}$ . And for each sample $i$ , there is a hard negative sample $i'$ sampled from the same video. Hence, the samples in the original batch $\\mathcal{B}$ and their hard negative counterparts together form the new batch $\\widetilde{\\mathcal{B}}$ .", + "bbox": [ + 496, + 393, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective matching video-to-text (v2t) for a video embedding $v$ is formulated as below; in practice the symmetric text-to-video (t2v) matching objective is also used (omitted for brevity):", + "bbox": [ + 496, + 561, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v 2 t}} ^ {\\mathrm {e g o}} = \\frac {1}{| \\mathcal {P} _ {m} |} \\sum_ {k \\in \\mathcal {P} _ {m}} \\log \\frac {\\exp \\left(v ^ {T} t _ {k} / \\tau\\right)}{\\sum_ {n \\in \\hat {\\mathcal {B}}} \\left(\\exp \\left(v ^ {T} t _ {n} / \\tau\\right) + \\exp \\left(v ^ {T} t _ {n ^ {\\prime}} / \\tau\\right)\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 631, + 924, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bounding Box Prediction. We use the 100DOH off-the-shelf hand object detector [53] to produce bounding boxes of two classes on each frame as supervision: hand and objects that are in contact with hands.", + "bbox": [ + 496, + 702, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are two challenges in using the detections for training supervision: 1) the image detector acts at the image level independently and does not provide box-ID association over different frames in a clip; 2) many hands and objects are missed due to motion blur and domain gap in ego-centric videos. Therefore, we apply Hungarian matching between predicted boxes $\\hat{b}$ and ground-truth boxes $b_{i}$ on single frames independently, so that for each $b_{i}$ , we find the matched prediction $\\hat{b}_{\\sigma(i)}$ to minimize the global matching", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13904", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cost. The final loss on bounding boxes is computed as the sum of the $\\ell_1$ loss and the Generalized IoU loss $L_{iou}$ [51] on paired boxes:", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {b o x}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) = \\mathcal {L} _ {\\mathrm {i o u}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) + \\left| \\left| b _ {i} - \\hat {b} _ {\\sigma (i)} \\right| \\right| _ {1} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 148, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and, to tackle the problem of missing objects, we do not penalize boxes that are not matched to nouns unlike traditional detection tasks.", + "bbox": [ + 75, + 180, + 468, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Class Prediction. We have noun embeddings from the video description and a set of predicted object name embeddings from the summary vectors, the task is to find the correspondence between them so that we can use the ground-truth nouns from the description as supervision for the predicted object names. As shown in figure 2, we align the nouns in the narrations and the names of objectboxes in two steps:", + "bbox": [ + 75, + 247, + 468, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(1) Object-noun alignment: We score the predicted object name embeddings $\\hat{n}$ against the ground-truth noun embeddings $n$ to construct a similarity matrix $C\\in \\mathbb{R}^{K\\times N}$ , where $K$ is the number of object queries and $N$ is the number of noun phrases in the description as following:", + "bbox": [ + 76, + 369, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC (n, \\hat {n}) = \\frac {n \\cdot \\hat {n}}{\\left\\| n \\right\\| \\left\\| \\hat {n} \\right\\|} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 458, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cost matrix $-C$ is used in Hungarian matching to select the matched summary vector for each noun phrase.", + "bbox": [ + 75, + 501, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Word-level contrastive training: We apply InfoNCE loss on the matched embeddings $\\hat{n}_j$ and $n_j$ against the embeddings of all the nouns $n_k'$ in Ego4D taxonomy dictionary $\\mathcal{D}$ [16]:", + "bbox": [ + 75, + 532, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {w o r d}} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\log \\frac {\\exp \\left(\\hat {n} _ {j} ^ {T} n _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {D}} \\exp \\left(\\hat {n} _ {j} ^ {T} n _ {k} ^ {\\prime} / \\tau\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 604, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Loss. The total training objective is the sum of the vision-text matching loss and the auxiliary losses on object vectors:", + "bbox": [ + 75, + 681, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {v} 2 \\mathrm {t}} ^ {\\mathrm {e g o}} + \\mathcal {L} _ {\\mathrm {t} 2 \\mathrm {v}} ^ {\\mathrm {e g o}} + \\mathcal {L} _ {\\text {b o x}} + \\lambda_ {w o r d} \\mathcal {L} _ {\\text {w o r d}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 739, + 468, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference", + "text_level": 1, + "bbox": [ + 76, + 768, + 186, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once trained, the model acts as a standard ego-centric vision-language video model which operates just on video frames and text descriptions, without requiring further access to object boxes or detectors. However, if desired, hand and object box detections and names can be read out for each frame at inference using the summary vectors, which can be used for grounding the input text description.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a37305366833cd3e3ec43a1e04f3b00b373ea24146ec6087df1412c67bb92568.jpg", + "image_caption": [ + "Step 1. Object-noun alignment", + "Figure 2. Training the model to predict object classes. Left: Object-noun alignment. First, the nouns in the video descriptions are matched against the predicted classes using Hungarian matching, to choose the most matched summary vectors. Right: Word-level contrastive training. Next, we supervise the matched predicted classes using a contrastive objective [46] against all the nouns in Ego4D taxonomy, to have similar embeddings as the corresponding nouns." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 658, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/26b27a49804498b929fcac44da66645bec9aaff1c4e7d902d206e6206c9ada7a.jpg", + "image_caption": [ + "Step 2. Word-level contrastive training" + ], + "image_footnote": [], + "bbox": [ + 683, + 93, + 870, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Implementation", + "text_level": 1, + "bbox": [ + 500, + 376, + 658, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In section 4.1, details of extracting the hand and object detections from 100DOH pre-trained detector are summarized, followed by the architectural details in section 4.2. Finally, in section 4.3 the training pipeline, model and input specification, and optimization details are provided. More details can be found in the supplementary materials.", + "bbox": [ + 498, + 402, + 890, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Weak Supervision from Pre-trained Detector", + "text_level": 1, + "bbox": [ + 500, + 505, + 879, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We uniformly sample four frame from each clip in the EgoClip dataset [35] as the input to the 100DOH hand-object detector [53]. The short side of the frame is resized to 640 pixels. There are 16 million frames in total, and the average number of boxes detected per frame is 1.96 for hand and 1.67 for object. Among all the frames, about $15.8\\%$ frames have no hands detected and $17.9\\%$ frames have no object detected. The average size of hand boxes is $2.8\\%$ of the frame size, while the average size of object boxes is $19.4\\%$ of the frame size. We use the top 2 hands and top 4 objects detected in the scene as supervision.", + "bbox": [ + 498, + 529, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Architecture", + "text_level": 1, + "bbox": [ + 500, + 709, + 633, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The object-aware module is a 6-layer cross-attention transformer with 8 attention-heads in each layer. The hidden dimension in cross-attention layers is 768, and the video embedding, object embeddings and text embeddings are projected to 256 dimensions before computing the cosine similarity score. We set the number of hand queries to 2, number of object queries to 12, which is designed to be larger than the maximum number of objects in the supervision. We use TimeSformer-L (TSF-L) [6] as the visual encoder, and a 12-layer Transformer [61] from LaViLa [77] as the text encoder.", + "bbox": [ + 498, + 733, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f70fec999aae7d610a0c55560ac1c635ceffd69a1a111a2023e082d8023a54b0.jpg", + "image_caption": [ + "Text: #C C cuts the box with the utility knife" + ], + "image_footnote": [], + "bbox": [ + 80, + 109, + 478, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b019a8585058e82ca29053cdb795cea348c4da7a6f2d1f3cf6b68702814a1c6d.jpg", + "image_caption": [ + "a) Object boxes predicted by off-the-shelf hand-object detector", + "b) Vision-text grounding results from the proposed model", + "Figure 3. Visualization of text grounding on EgoClip. We show the comparison between detections from off-the-shelf 100DOH hand-object detector [53] (used for training supervision) and the predicted boxes from our model respectively. (a) The detections are noisy, objects are missed, and there is no temporal association of the detected boxes across frames. (b) The trained model learns temporally consistent tracks as well as object categorization using only noisy frame-level box-supervision and weak supervision from the texts." + ], + "image_footnote": [], + "bbox": [ + 80, + 205, + 478, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5bb4f5d0621e1eba35d67fa22d0497d7930bc32200b122b72e940eeb1ef3ce48.jpg", + "image_caption": [ + "C C puts the spatula in his right hand on the chopping board" + ], + "image_footnote": [], + "bbox": [ + 496, + 109, + 888, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8e461a3f1036fae36ea4fb708f64b16edaeab28073c05e5c29968cfbc0cf4de.jpg", + "image_caption": [ + "a) Object boxes predicted by off-the-shelf hand-object detector", + "b) Vision-text grounding results from the proposed model" + ], + "image_footnote": [], + "bbox": [ + 496, + 204, + 888, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training Details", + "text_level": 1, + "bbox": [ + 76, + 375, + 236, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In training, we uniformly sample $4 \\times 224 \\times 224$ RGB frames from each clip. Hand and object boxes are preextracted from these 4 frames using off-the-shelf detector and used as supervision. We keep the visual and text encoder frozen in training. Only the object-aware decoder, query vectors, and the MLP projection layer on text embedding parameters are learned during training. We train the model for 5 epochs on one A6000 GPU, with batch size 128. We use AdamW [39] as the optimizer and set the learning rate to $3e - 5$ . $\\lambda_{word}$ is set to 0.5 to balance the scale of the four losses.", + "bbox": [ + 75, + 398, + 470, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Section 5.1 introduces all the datasets we used for training and evaluation, followed by the evaluation protocols in section 5.2. Finally, we discuss the ablation studies (section 5.3) and compare to prior SOTA methods on different benchmarks (section 5.4).", + "bbox": [ + 75, + 607, + 468, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 694, + 179, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ego4D/EgoNLQ/EgoMQ/EgoClip/EgoMCQ [16, 35]. Ego4D is a massive-scale dataset focusing on ego-centric videos. It contains 3670 hours video for many different tasks, including action anticipation, AV diarization, etc. EgoNLQ and EgoMQ is a subset for natural language queries and moment query, designed for testing the models' episodic memory and long-term video understanding. [35] proposes a new subset EgoClip for vision-language pre-training, comprising 3.8M clip-text pairs. They also introduce EgoMCQ (i.e., Egocentric Multiple-Choices-Question) as a downstream evaluation dataset for the pre-training. Given a text query, the model tasked to", + "bbox": [ + 75, + 717, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "choose the paired video clip from 5 candidates. The evaluation metrics is 'intra-video' and 'inter-video' accuracy, depending on where the candidates are chosen from.", + "bbox": [ + 496, + 375, + 892, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Epic-Kitchens-MIR [11]. Epic-Kitchens is a large-scale ego-centric dataset with 100-hour activities recorded in kitchens. Epic-Kitchens-MIR is a subset with about 9881 clip-text pairs for vision-language retrieval. It is designed for multi-instance retrieval. The model is evaluated on retrieving the paired text/video given a query text/video. The evaluation metrics are mean average Precision (mAP) and normalized Discounted Cumulative Gain (nDCG).", + "bbox": [ + 496, + 436, + 890, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VISOR [12] is a benchmark built on Epic-Kitchens for segmenting hands and active objects in egocentric video. It has pixel-level annotations covering 36 hours of untrimmed video and 257 object classes. We utilize this annotations in its val split for evaluations on vision-text grounding.", + "bbox": [ + 496, + 574, + 890, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EGTEA [34] contains 28 hours of cooking activities from 86 unique sessions of 32 subjects. Fine-grained actions are classified into 106 classes. We retrieve the text-descriptions of action classes to evaluate the model for action classification on the test set of all three splits. We measure the performance using mean-class accuracy and top1 accuracy.", + "bbox": [ + 496, + 666, + 890, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 500, + 770, + 689, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of our model in the following three aspects:", + "bbox": [ + 496, + 794, + 888, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot transfer. To test the transferability and generalization ability, we conduct zero-shot evaluation on multiple-choice questions (EgoMCQ), multi-instance retrieval (EpicKitchens-MIR), action classification (EGTEA). Among these datasets, videos in EgoMCQ are from the same data", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13906", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "source as our pre-training dataset Ego4D. Other datasets demonstrate a domain gap, hence, evaluate for transferable representations.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Episodic memory. To evaluate the richness of the representations learned by our model, we use the video representations to solve Episodic memory tasks in Ego4D. Following [35, 77], we pre-extract the video features from our model first. Using these pre-computed features as input, we train a VSLNet [72] for temporal localization in NLQ, and train a VSGN [76] for moment retrieval in MQ.", + "bbox": [ + 75, + 143, + 467, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vision-language grounding. Due to the lack of ground-truth for object-grounding in EgoClip, we evaluate the grounding ability of the model on VISOR instead. VISOR is an egocentric dataset with scenes in the kitchens, where frames are annotated sparsely with segmentation masks and object names. We re-propose the manually annotated segmentation masks in it to extract ground-truth bounding-boxes for hands and in-contact objects. To carry out the zero-shot evaluation, we take the annotated frames in the val split, filter out the not-in-contact objects in each frame, and convert all the segmentation masks to bounding boxes as ground truth. The predicted object boxes are matched with ground-truth object boxes using noun alignment as during training (eq. (5)), while the left/right hands are predicted from the first and the second hand queries respectively. We repeat a single frame 4 times temporally, and resize it to $224 \\times 224$ pixels to be consistent with the pre-training resolution. The predicted bounding-boxes are evaluated to be correct if their centers lie inside the ground-truth bounding-boxes.", + "bbox": [ + 75, + 253, + 467, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablations", + "text_level": 1, + "bbox": [ + 76, + 571, + 187, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Losses. We ablate the combination of losses on three downstream benchmarks in table 1. Results showing that having both $L_{box}$ and $L_{word}$ leads to the best performance. With the same architecture, when training the model using only $L_{Ego}$ without introducing any object-awareness, the performance is $2\\%$ lower on EK100-MIR and EGTEA compared to the object-aware one.", + "bbox": [ + 75, + 597, + 467, + 704 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f74d15c2d53a56b5095241920b02b9f763d77b27b33bf2377e0ccc055af813f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LossesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
LEgo93.761.835.936.644.937.6
LEgo + Box94.262.736.737.445.338.5
LEgo + Word93.761.936.737.445.838.1
LEgo + Box + Word94.563.037.537.846.639.1
", + "bbox": [ + 78, + 719, + 467, + 825 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d68c5e93932f379d0b469682ad827e67de50dbfdc70985537f5e4cf93031712.jpg", + "table_caption": [ + "Table 1. Ablation on training objectives for zero-shot transfer tasks. Introducing the object-awareness by having box and word supervision helps the model to achieve better transfer results on EK100-MIR and EGTEA." + ], + "table_footnote": [], + "table_body": "
Detector input resEgoMCQEKEGTEAVISOR
InterIntraAvg mAPAvg nDCGTop1MeanLoc Acc
256p94.263.235.734.642.036.0-
256p94.563.036.937.044.338.968.2
640p94.563.037.537.839.146.678.7
", + "bbox": [ + 501, + 88, + 890, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80272ae0e50669a50e9897f9c3dc8a75ec0f1ae682baca23398cf927a0fa4b5c.jpg", + "table_caption": [ + "Table 2. Ablation on box quality for zero-shot transfer tasks. We extract boxes using $256\\mathrm{p}$ and $640\\mathrm{p}$ images as input to the detector respectively, resulting in boxes of different qualities as supervision in training." + ], + "table_footnote": [], + "table_body": "
# Obj QueriesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
494.162.837.837.645.938.1
894.562.737.738.045.537.9
1294.563.037.537.846.639.1
", + "bbox": [ + 508, + 261, + 879, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation on the number of object queries for zero-shot transfer tasks. The number of queries do not have a big impact on EK100. Larger number of queries shows a boost on mean-class accuracy on EGTEA, and smaller number of queries is better on intra-video accuracy.", + "bbox": [ + 498, + 361, + 890, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quality of detected boxes. The extent to which a model can acquire object-level information is constrained by the quality of the bounding boxes from the off-the-shelf detector. To investigate how much the quality of boxes affects our training, we detect hand and object boxes on EgoClip training set using 100DOH [53] with 256p and 640p images as input – with larger image size, the objects should be more precisely delineated. We use the two sets of boxes as supervision in our training and show results in table 2. Even when training with noisy boxes from 256p, our model is better than the previous SOTA model. When boxes from 640p are used, the averaged zero-shot transfer performance is further improved by $1\\%$ on Epic-Kitchens and EGTEA, showing that our method can bring larger improvement over the non-object-aware method when given better boxes. Furthermore, a significant boost on the grounding results is also observed on VISOR when using boxes with better quality.", + "bbox": [ + 496, + 450, + 890, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of object queries. We use different number of query vectors in the object-aware decoder to see its impact on both vision-language tasks and the grounding task. We show zero-shot transfer results in table 3 and vision-text grounding results in table 7. The number of queries has relatively small impact on vision-language representation learning, the performance gaps on zero-shot transfer tasks are mostly small than $1\\%$ . However, a smaller number of object queries leads to much better results for in-contact object grounding on VISOR, 4 queries is better than 12 objects by $3.6\\%$ in localization accuracy. The reason is that too", + "bbox": [ + 496, + 732, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/226be90f4350c23a35c35bdbc28d6516ffef74c03f8475e7f76050ef196a89e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneBatch sizeObject awareHard negEgoMCQEK100-MIREGTEA
Inter-videoIntra-videoV-TmAP T-VAvgV-TnDCG T-VAvgTop1-AccMean-Acc
EgoVLPTSF-B512X✓90.657.226.020.623.328.827.027.917.6-
LaViLaTSF-B1024XX93.859.935.126.630.933.730.432.0-28.9
LaViLaTSF-L1024XX94.563.140.032.236.136.133.234.640.134.1
LaViLa*TSF-L1024XX94.263.239.731.735.736.133.234.642.036.0
Ours*†TSF-L128XX93.760.539.730.335.037.334.535.944.836.3
Ours*‡TSF-L128X✓93.761.840.731.135.938.335.036.644.937.6
Ours*TSF-L128✓✓94.563.042.332.737.539.336.237.846.639.1
", + "bbox": [ + 78, + 88, + 893, + 228 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cf53cfdef2430e943c4a86328432fa4d233159dca172046d1de4adcfb6f324ff.jpg", + "table_caption": [ + "Table 4. Comparison to SOTA results on zero-shot transfer to EgoMCQ, EK100-MIR and EGTEA. We compared to EgoVLP and LaVILA, two previous SOTA models pre-trained on EgoClip. Our object-aware model has achieved comparable results on multiple-choice questions on EgoMCQ, and SOTA results on multi-instance retrieval task on EpicKicthens and action classification on EGTEA. Model without * use center cropping in evaluation, while * denotes the usage of resizing instead of cropping. Ours† and Ours‡ stands for different variants of our model depending on whether object-aware losses and hard negative sampling is used in training." + ], + "table_footnote": [], + "table_body": "
MethodPredicted BoxesEgoMCQEKEGTEA
InterIntraAvg mAPAvg nDCGMeanTop1
Oursobj94.062.136.837.145.137.6
hand+obj94.563.037.537.846.639.1
", + "bbox": [ + 86, + 303, + 455, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "many queries result in a large number of predicted boxes, which increases the probability of mis-matching.", + "bbox": [ + 75, + 422, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of hand boxes. We ablate the impact of having hand boxes as supervision in our training. Results are shown in table 5, training the model to predict hand as well as objects help the model to get about $1\\%$ higher zero-shot transfer performance on EK and EGTEA.", + "bbox": [ + 75, + 470, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Comparison to the SOTA", + "text_level": 1, + "bbox": [ + 76, + 556, + 307, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zero-shot transfer. We compare to previous SOTA in table 4. Our model is comparable on EgoMCQ and better on EK-MIR and EGTEA, showing its good zero-shot transferability. Due to limited compute resources, we are not able to unfreeze the visual backbone to train end-to-end or increase the batch size further. Despite these disadvantages, our method outperforms the previous SOTA on two tasks for models that have been trained end-to-end.", + "bbox": [ + 75, + 580, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The main difference between LaViLa(L) and ours is the object-aware training and hard negative sampling; Ours† in table 4 is a LaViLa(L) model with an extra Transformer decoder, which is trained with only InfoNCE loss on video and sentence embeddings. Without object-awareness and hard sampling, it gets better accuracy on EGTEA and better nDCG on EK100 due to more parameters added, but falls behind on EgoMCQ and mAP on EK100. Applying hard negative sampling (Ours‡) and object-aware training (Ours) brings improvement across the board. The most obvious boost comes from inducing object-awareness, bringing $1.5\\%$ improvement on average. And the results could be further improved by obtaining better pseudo-boxes, as the", + "bbox": [ + 75, + 702, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "magnitude of boost from learning objects is closely related to the box quality (as shown in table 2).", + "bbox": [ + 498, + 300, + 890, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Episodic memory. Results on EgoNLQ and EgoMQ are shown in table 6. The two tasks test the video understanding on several-minutes long videos. In these experiments, trained video and text backbones are used as feature extractor, and extra modules are trained on the long feature sequences for natural language querying and memory querying. Therefore, the richer the information encoded in the features, the better the results will be. We list the models trained on the same amount of video and text data in black, results show that our model are better than the previous SOTA on all the metrics in the two tasks. This is because features from the object-aware model have captured more object information, thus enabling better precision and recall on localization and retrieval. InternVideo [63] and NaQ [50] are trained on more video or text data, and are included for completeness. Please note that our method is compatible with any visual language backbones, using better backbones might also lead to better results.", + "bbox": [ + 496, + 353, + 890, + 628 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61d427f9b6f7af1ecfe74994e6dde2be044b823c260865ee523cc4060fcdb41f.jpg", + "table_caption": [ + "Table 5. The impact of having hand boxes in the box prediction. Having hand boxes in the prediction helps on all the zero-shot evaluation benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodBatch sizeEgoNLQEgoMQ
mIOU@0.3mIOU@0.5R1@0.5R5@0.5mAP
R1R5R1R5
SlowFast-5.510.73.16.625.246.26.0
EgoVLP51210.818.86.813.530.152.011.4
LaViLa(B)102410.519.16.713.6---
LaViLa(L)102412.122.47.315.432.556.113.4
Ours12813.223.37.915.633.456.716.0
VideoIntern [63]14k16.523.010.116.1--23.6
ReLER + NaQ [50]204819.323.611.615.5---
", + "bbox": [ + 501, + 641, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Comparison to SOTA results of fine-tuning on EgoNLQ and EgoMQ. Our object-aware model encodes richer information in the visual representations, hence obtaining better results on all the metrics in NLQ and MQ task in Ego4D episodic memory benchmark. We list other SOTA models (in grey) trained with more video and text data for completeness.", + "bbox": [ + 496, + 814, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13908", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4824fa32f4fd25d96fa8576afc114fe9a6d2457555e1cc19ad30da1231fd991c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. In-contact object localization accuracy on VISOR. Our model does better on in-contact object localization after weakly-supervised training compared to the baseline (the Detector with Random or GT object assignment), which is the source of supervision in our pre-training. The improvement is due to increased recall of our model over the baseline." + ], + "table_body": "
ModelObject Assignment# queriesLoc Accuracy
Detector [53]Random-37.1
GT matching-41.3
OursPredicted482.3
881.2
1278.7
", + "bbox": [ + 78, + 88, + 470, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.5. Evaluating Object Grounding", + "text_level": 1, + "bbox": [ + 76, + 294, + 344, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative results on EgoClip. In fig. 3 we show the grounding results on EgoClip after training as compared to the supervision from the 100DOH detector. After training, the predictions from our model find the missing objects because we do not penalize extra bounding boxes predicted, but select the active object/object of interest through matching noun embeddings. It also learns temporal association of object bounding-boxes as the result of using the same object summary vector to predict the boxes over all the frames (as in eq. (1)); the summary vector attends to visually similar features corresponding to the same object across the clip without any explicit supervision for temporal consistency.", + "bbox": [ + 75, + 316, + 468, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative and qualitative results on VISOR. In table 7 we show the text-grounding results on VISOR. We take the predictions from the hand-object detector [53] as our baseline. Since the predictions only detect hands and objects without an object class name, we associate the predicted boxes with ground-truth boxes in VISOR in two ways: (a) random: we assign predicted object boxes to GT object boxes randomly, (b) GT matching: we use Hungarian matching to find the predicted boxes with highest IoU against the GT object boxes. However, even when matched using GT information, the baseline detector does not achieve a high accuracy due to poor recall. Results show grounding ability of our model is $40\\%$ better than the baseline using only weak supervision from the video descriptions. We also show the qualitative results on detecting different number of hands and objects in fig. 4, our model has a much higher recall when compared to baseline detector when operating at the same resolution.", + "bbox": [ + 75, + 511, + 468, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 799, + 194, + 814 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduce a method to learn object-aware ego-centric video representations using noisy supervision from pre-trained hand-object detectors. The object-representations so learned show strong zero-shot transfer across various downstream tasks and datasets, mirroring", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fdd00c87270569799d09f157cbd51d0748e5acba85626628a894d61e917a5b39.jpg", + "image_caption": [ + "Figure 4. Grounding visualization on the Epic-Kitchen VISOR val split. When operating at the same resolution, our model shows better grounding ability on hands and objects compared to the baseline 100DOH detector [53] (with GT matching) used for training. Note that the low IoU on hands on the third column is a result of the GT 'hand' segmentation mask covering the arm by definition, while the detector and our model are trained to localize only palm and fingers." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 883, + 368 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the performance improvement from object-aware training on images [70]. The model uses standard neural modules (i.e., transformers), and does not require any object boxes or detectors as input during inference, making it widely applicable as a drop-in replacement for training video-language models. Even though the model is trained with sparse and noisy object supervision at the frame-level (without temporal associations), during inference dense temporal bounding-box tracks and category predictions can be obtained, which are superior to the predictions from the base hand-object detector used for training. There are several avenues for improvement. Our model uses the pre-trained video encoder operates at a small resolution $224 \\times 224$ which makes detecting small objects difficult. Further, four frames are sampled uniformly from the clip regardless of its length which can cause difficulties due to temporal aliasing. Nevertheless, we hope our work will inspire further research in learning transferable object-aware representations for videos.", + "bbox": [ + 496, + 508, + 890, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This research is funded by a Google-DeepMind Graduate Scholarship, a Royal Society Research Professorship, and EPSRC Programme Grant VisualAI EP/T028572/1.", + "bbox": [ + 498, + 816, + 890, + 878 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13909", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2", + "[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In Proc. ICCV, 2021. 1", + "[3] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In NeurIPS, 2022. 2, 3", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proc. ICCV, 2021. 1, 2, 3", + "[5] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 2", + "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In Proc. ICML, 2021. 4, 5", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 1, 3, 4", + "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proc. ECCV, 2020. 2", + "[9] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 4", + "[10] Noam Chomsky and David W Lightfoot. Syntactic structures. Walter de Gruyter, 1957. 1", + "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proc. ECCV, 2018. 1, 2, 6", + "[12] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Datasets and Benchmarks Track, 2022. 2, 6", + "[13] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 2", + "[14] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurlPS, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Philip Gage. A new algorithm for data compression. C Users Journal, 1994. 4", + "[16] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proc. CVPR, 2022. 1, 5, 6", + "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. ICCV, 2017. 1", + "[18] Paul Henderson and Christoph H. Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2", + "[19] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. arXiv preprint arXiv:2110.06915, 2021. 1, 2", + "[20] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proc. ICCV, 2019. 2", + "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 1, 2", + "[22] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proc. CVPR, 2016. 2", + "[23] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. ICCV, 2021. 2", + "[24] Ranjay Krishna, Ines Chami, Michael Bernstein, and Li Fei-Fei. Referring relationships. In Proc. CVPR, 2018. 2", + "[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2", + "[26] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In Proc. ICCV, 2021. 2", + "[27] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012. 2", + "[28] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proc. CVPR, 2022. 2", + "[29] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 2017. 1", + "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. ICML, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13910", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2", + "[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proc. ICCV, 2021. 2", + "[33] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. ECCV, 2020. 2", + "[34] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proc. ECCV, 2018. 6", + "[35] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 2, 3, 4, 5, 6, 7", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, 2014. 2", + "[37] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 704-721. Springer, 2020. 2", + "[38] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2", + "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. ICLR, 2019. 6", + "[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 2", + "[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3", + "[42] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proc. CVPR, 2020. 1", + "[43] Effrosyni Mavroudi and René Vidal. Weakly-supervised generation and grounding of visual descriptions with conditional generative models. In Proc. CVPR, 2022. 2", + "[44] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proc. ICCV, 2019. 2", + "[45] Marvin Minsky. Society of mind. Simon and Schuster, 1988. 1", + "[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre" + ], + "bbox": [ + 78, + 90, + 468, + 897 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sensation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 5", + "[47] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE conference on computer vision and pattern recognition, pages 2847-2854. IEEE, 2012. 2", + "[48] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuytelaars. Revisiting spatio-temporal layouts for compositional action recognition. In Proc. BMVC, 2021. 1", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 1, 2, 3", + "[50] Santhosh Kumar Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Naq: Leveraging narrations as queries to supervise episodic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6694-6703, 2023. 8", + "[51] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. CVPR, 2019. 5", + "[52] Kate Saenko, Ben Packer, C Chen, Sunil Bandla, Y Lee, Yangqing Jia, J Niebles, Daphne Koller, Li Fei-Fei, Kristen Grauman, et al. Mid-level features improve recognition of interactive activities. Technical report, Dept. of Elec. Engg. and Computer Sc., University of California, Berkeley, 2012. 1", + "[53] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proc. CVPR, 2020. 1, 2, 4, 5, 6, 7, 9", + "[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Proc. ICLR, 2015. 2", + "[55] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurIPS, 2022. 2", + "[56] Elizabeth S Spelke, Karen Breinlinger, Janet Macomber, and Kristen Jacobson. Origins of knowledge. Psychological review, 99, 1992. 1", + "[57] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In Proc. ECCV, 2018. 1", + "[58] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2", + "[59] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. *NeurIPS*, 2021. 2", + "[60] Joshua B Tenenbaum, Charles Kemp, Thomas L Griffiths, and Noah D Goodman. How to grow a mind: Statistics, structure, and abstraction. Science, 2011. 1", + "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 4, 5", + "[62] Xiaolong Wang and Abhinav Gupta. Videos as space-time" + ], + "bbox": [ + 501, + 92, + 890, + 897 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13911", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "region graphs. In Proc. ECCV, 2018. 1, 2", + "[63] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 8", + "[64] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 3", + "[65] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2", + "[66] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In Proc. ICCV, 2021. 2", + "[67] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proc. ICCV, 2019. 2", + "[68] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2", + "[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2", + "[70] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 9", + "[71] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proc. ACCV, 2022. 1, 2, 3", + "[72] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE PAMI, 2021. 7", + "[73] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2", + "[74] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In European Conference on Computer Vision, pages 127-145. Springer, 2022. 2", + "[75] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. CVPR, 2021. 2", + "[76] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 7", + "[77] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language" + ], + "bbox": [ + 78, + 92, + 468, + 897 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models.arXiv preprint arXiv:2212.04501,2022.2,5,7", + "[78] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proc. CVPR, 2022. 2", + "[79] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. arXiv preprint arXiv:1805.02834, 2018. 2", + "[80] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 2", + "[81] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In Proc. ECCV, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 323 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13912", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_model.json b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fbb903227a496ed871b48f063aa5aaf06bb97f7b --- /dev/null +++ b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_model.json @@ -0,0 +1,2754 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.131, + 0.852, + 0.153 + ], + "angle": 0, + "content": "Helping Hands: An Object-Aware Ego-Centric Video Recognition Model" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.183, + 0.335, + 0.234 + ], + "angle": 0, + "content": "Chuhan Zhang \nVGG, University of Oxford \nczhang@robots.ox.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.183, + 0.596, + 0.234 + ], + "angle": 0, + "content": "Ankush Gupta Google DeepMind, London ankushgupta@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.635, + 0.183, + 0.856, + 0.234 + ], + "angle": 0, + "content": "Andrew Zisserman \nVGG, University of Oxford \naz@robots.ox.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.27, + 0.313, + 0.287 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.471, + 0.44 + ], + "angle": 0, + "content": "We introduce an object-aware decoder for improving the performance of spatio-temporal representations on egocentric videos. The key idea is to enhance object-awareness during training by tasking the model to predict hand positions, object positions, and the semantic label of the objects using paired captions when available. At inference time the model only requires RGB frames as inputs, and is able to track and ground objects (although it has not been trained explicitly for this)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.638 + ], + "angle": 0, + "content": "We demonstrate the performance of the object-aware representations learnt by our model, by: (i) evaluating it for strong transfer, i.e. through zero-shot testing, on a number of downstream video-text retrieval and classification benchmarks; and (ii) by using the representations learned as input for long-term video understanding tasks (e.g. Episodic Memory in Ego4D). In all cases the performance improves over the state of the art—even compared to networks trained with far larger batch sizes. We also show that by using noisy image-level detection as pseudo-labels in training, the model learns to provide better bounding boxes using video consistency, as well as grounding the words in the associated text descriptions." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.64, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Overall, we show that the model can act as a drop-in replacement for an ego-centric video model to improve performance through visual-text grounding1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.21, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.864 + ], + "angle": 0, + "content": "In visual-language models there has been a recent move to explicitly build object awareness into the vision module by adding specialized and bespoke components, or using entirely object-centric architectures. The motivation for this partly comes from the attractive compositional nature of objects and their inter-relationships in language, which enables inexhaustible novel combinations [10, 45], and partly from infant cognitive studies that stress the im" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.892, + 0.424 + ], + "angle": 0, + "content": "portance of objects in early visual development [29, 56, 60]. Examples in the video domain include explicit internal object representations [2], e.g., through RoI-align [17] pooled features either from a pre-trained region-proposal network (RPN) [2, 52, 57, 62], or from bounding-box coordinates taken as input [19, 42, 48, 71]. This contrasts with the large body of work where standard representations are learnt end-to-end without any explicit factorization into objects/entities, such as dual-encoder vision-language models in the image [21, 49] and video domains [4, 64]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.425, + 0.893, + 0.502 + ], + "angle": 0, + "content": "In this paper, we take a different (middle) path and instead use a vanilla video transformer architecture and induce object-awareness into the video representation by tasking the model to predict object-level properties, such as their localization and semantic categories, only during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.893, + 0.672 + ], + "angle": 0, + "content": "Our target domain is ego-centric video [11, 16], and we tailor the object properties used to this. In ego-centric videos the actor [57] is often present through their hands, and we therefore task the network to predict both the hands and the principal objects they interact with. As will be seen, this simple object-aware training boosts the performance of pre-trained video-language architectures significantly, and leads to state-of-art performance across multiple ego-centric benchmarks. During inference, the model requires only RGB frames as input, and operates as a standard video-language network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In more detail, our model is built on top of a pre-trained video-language dual encoder architecture (where there are separate encoders for the video and text data). We add an additional, but vanilla, transformer decoder head [61], and train with DETR/Mask2former [7, 9] query vectors and object loss for hands and other objects. The intuition is that these additional query vectors help the model to attend to and track the hands and salient objects in the scene (these are the 'helping hands' of the paper title). Importantly, we do not require dense frame level ground truth for this training. Rather, we obtain somewhat noisy and temporally sparse annotations automatically from a hand object detector [53], and use these to provide prediction targets for the frames where they are available. This opportunistic use of annotations is pragmatic as object detectors trained on third" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.468, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code and models available at: https://github.com/Chuhanxx/helping_hand_for_egocentric Videos" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13901" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "person datasets (such as COCO) do not perform so well on the ego-centric domain, where the scenes are more crowded and objects are often small and can be motion blurred. By only requiring annotations for a subset of frames, where they can be reliably produced automatically, we are able to train on large-scale data without requiring expensive manual supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Although we train with noisy and sparse image-level object localization, our model can learn to predict better and denser bounding-box trajectories through large-scale training due to the spatio-temporal consistency which naturally presents in videos. Also, it is able to predict semantic grounding by learning to map the object appearance to the nouns in the video captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.398 + ], + "angle": 0, + "content": "It is worth noting that we are using hand detectors because hands are a common and important object in ego-centric videos. However, the object-centric method we are proposing has greater scope than ego-centric videos and can be applied to other scenarios with other object types providing the 'helping-hand'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.489 + ], + "angle": 0, + "content": "In summary, we make the following contributions: (i) We propose a method to induce object-awareness in video-language models for an architecture composed of standard neural modules. The model only requires RGB frames as inputs, and thus is a drop-in replacement for any ego-centric video model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.469, + 0.536 + ], + "angle": 0, + "content": "(ii) The model can be trained opportunistically using available and sparse frame-level and noisy annotations, produced automatically." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.598 + ], + "angle": 0, + "content": "(iii) We demonstrate state-of-the-art strong (zero-shot) transfer for cross-modal retrieval to other ego-centric datasets namely, EpicKitchens-MIR and EGTEA improving prior art by \\(2 - 4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.469, + 0.66 + ], + "angle": 0, + "content": "(iv) We evaluate the grounding quantitatively using the EpicKitchens-VISOR dataset [11, 12] and find that the model outperforms the base hand-object detector used for training supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.469, + 0.736 + ], + "angle": 0, + "content": "(v) Finally, we also demonstrate that the representations learned can be used as input in long-term video understanding tasks like EgoNLQ and EgoMQ. The objectiveness in the representation helps the model outperform other models trained on the same training set on these two tasks." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.22, + 0.769 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Vision and Language Representation Learning. Different from transferring representations learned for classification on a fixed set of object categories [27, 54], recent vision-language pre-training (VLP) works leverage largescale supervision from free-from text descriptions of images and videos. These methods use image captions [49] or video sub-titles [44, 64] with either independent dual encoders for the visual and text modalities [4, 21, 41], or via" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.365 + ], + "angle": 0, + "content": "joint encoders with cross-attention across modalities [1, 30, 31, 69]. We also use dual-encoders which are kept frozen due to compute limitations. To explicitly build-in object-awareness into image representations, object-level features extracted from pre-trained object-detectors are aligned with the text descriptions [8, 33, 40, 58, 75, 78] The object-level text alignment is further augmented with the object-box prediction task for grounding in [13, 23, 70, 73]. VLP for ego-centric videos has recently been explored [35, 77] to bridge the domain gap between representations learned from third-person videos found commonly [44], and first-person ego-centric videos. We further extend image based object-aware VLP methods to ego-centric videos by training to predict the auto-generated hand-object boxes extracted from pretrained detectors [53], while requiring only RGB input during inference, making our model a drop-in replacement for ego-centric VLP models albeit with enhanced object-awareness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Weakly Supervised Video Text Grounding. A key challenge for grounding in videos is the lack of large-scale object-level annotations for videos. While such annotations are readily available for synthetic datasets [68], expensive manual annotation is required for real videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Hence, weakly supervised method have been developed which leverage the video sub-titles/description to map nouns/verbs to regions in frames. This is typically achieved by first extracting bounding-box/segmentation regions from pre-trained detectors for objects and humans, and aligning them with keywords using max-margin [79] or constrastive [32] objectives. Similarly, [43] also align words in the video captions to regions from pre-trained RPN by modelling the region-word associations as latent variables of a conditional generative variational auto-encoder. More recently, [59] use cross-attention across text and candidate regions, and find (soft-) associations based on attention weights. While our model similarly absorbs object bounding-boxes and category information obtained from pre-trained object and hand detectors during training to induce object-aware representations, these detectors are not required during inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Object-Oriented Learning in Videos. Most vision tasks for images involve objects [22, 24, 25, 36]. In videos, there are also a broad range of object-oriented tasks: some works treat learning object-level information as an end task, they design models to predict object bounding-boxes and masks [5, 26, 37, 38, 65, 67, 74]; other works use object knowledge to achieve some other high-level tasks, for instance, object-centric scene generation [28, 66] and decomposition [14, 18, 55], and action recognition [3, 19, 20, 47, 62, 71, 80]. Our method uses object information to as-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.088, + 0.845, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.896, + 0.485 + ], + "angle": 0, + "content": "Figure 1. The object-aware vision-language model architecture. The architecture is made up of three parts: A video backbone, a text backbone and a object-aware decoder. The decoder is a cross-attention transformer, it takes the visual feature map \\( F' \\) as keys and values, which are attended by a set of learnable queries. In these permutation-invariant query vectors, the hand and object queries \\( Q_h, Q_o \\) are trained to be object-aware and predict the localization and class of hands and objects. The video query \\( q_v \\) attends to both the visual feature map through cross-attention layers, and the object feature map through self-attention layers in the Transformer decoder and output a video-level embedding \\( v \\) to be matched with the text embedding \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.473, + 0.737 + ], + "angle": 0, + "content": "sist vision-language alignment in egocentric videos, so that the model learns grounded video representations that can generalize better. SViT [3] uses object queries shared between images and videos in order to predict hand-object bounding boxes in videos, whilst only requiring image-level supervision during training. However, the object queries are not used for vision-language alignment. Our previous work [71] showed that encoding object-level information in the model helps transfer learning in various video understanding tasks, but the model required GT bounding boxes as input during inference. In contrast, in this work we model does not require this information during training, and we show box prediction and vision-language alignment can be combined and benefit both in-domain and out-of-domain tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.751, + 0.45, + 0.77 + ], + "angle": 0, + "content": "3. Object-Aware Vision-Language Learning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.777, + 0.47, + 0.839 + ], + "angle": 0, + "content": "We first describe the task of object-aware vision-language learning and the architecture of our model. These are followed by the training objectives for vision-language matching and weakly-supervised text grounding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.188, + 0.863 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In vision-language representation learning the training data consists of short video clips (RGB frames) and an" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.678 + ], + "angle": 0, + "content": "associated free-form text string containing words that describe the visual content. Typically, dual encoders are trained on this paired data, with a visual-encoder that ingests the video clip and a language-encoder that ingests the text [4, 35, 41, 49, 64]. The dual encoders are trained with a contrastive objective [46] such that the cosine similarity for matching vision-text pairs is optimized to be higher than the similarity of negative/not-matching pairs. This pretraining objective enables evaluation on downstream vision-language tasks like video-text retrieval and action classification in a zero-shot manner [49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.893, + 0.821 + ], + "angle": 0, + "content": "Our object-aware model follows the data, training and evaluation pipeline as above, except that the model is also tasked to output object-level information (e.g., bounding boxes and object categories) during training. By tasking the model to predict object bounding boxes and names which can be matched to the nouns in the narration, the model learns grounded and fine-grained correspondence between the modalities. The object-level prediction is used as an auxiliary objective in training but not used at inference time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In more detail, as shown in figure 2, there are two types of object-level prediction: (a) hand and object bounding boxes; and (b) object names. Since ground truth of boxes and object names are not available, and most traditional detectors [7, 81] fail to identify objects well in ego-centric" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13903" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.23 + ], + "angle": 0, + "content": "videos, we cannot rely on strong supervision for the predictions. Instead, we generate bounding box targets (for the hands and other objects) using a robust off-the-shelf hand-object detector [53], though these targets will only be available for some of the frames and are noisy. While for object name prediction, we use a weakly supervised method to align the predicted names with nouns in the paired narration (detailed in section 3.3). In both cases, the supervision is opportunistic and only applied when available." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.24, + 0.212, + 0.254 + ], + "angle": 0, + "content": "3.2. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.386 + ], + "angle": 0, + "content": "Dual Encoder. We use dual encoders as our visual and text backbone for efficiency. The visual encoder ingests a video clip \\( F \\) of RGB frames \\( F = (f_{1}, f_{2}, \\ldots, f_{T}) \\), where \\( f_{i} \\in \\mathbb{R}^{H \\times W \\times 3} \\) and \\( T \\) is the number of frames. The clip \\( F \\) is encoded by a Video Transformer [6] which tokenizes the frames by 3D patches to produce a spatially downsampled feature map \\( F' = (f_{1}', f_{2}', \\ldots, f_{T}') \\), where \\( f_{i}' \\in \\mathbb{R}^{H' \\times W' \\times C} \\). It outputs a visual embedding \\( v \\in \\mathbb{R}^{C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.509 + ], + "angle": 0, + "content": "The text encoder is a Transformer [61] that inputs words tokenized by a BPE tokenizer [15]. It encodes two type of inputs: (a) a narration sentence which describes the contents of a clip; and (b) a noun set which contains noun phrases from the narration sentence. At the output, the embedding corresponding to the EOS token is taken to be the embedding for the full sentence \\( t \\in \\mathbb{R}^C \\), and multiple noun embeddings \\( n \\in \\mathbb{R}^C \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Object-Aware Module. The object-aware module is a cross-attention Transformer which has a permutation-invariant set of learnable vectors as queries (similar to DETR [7] and Mask2former [9]). The queries are at video-level, shared between frames. They consist of three sets: two hand queries \\( Q_{h} = (q_{h1}, q_{h2}) \\) for the left and right hands; \\( K \\) object queries \\( Q_{o} = (q_{o1}, q_{o2}, \\ldots, q_{ok}) \\); and a video-level query \\( q_{v} \\). These queries are learned, and attend to the visual feature map \\( F' \\) from the visual backbone and output a set of summary vectors \\( S = (s_{h1}, s_{h2}; s_{o1}, \\ldots, s_{ok}) \\) corresponding to each input query." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.471, + 0.835 + ], + "angle": 0, + "content": "The object-aware module operates on the visual content without any interaction with the text information. It consists of six cross-attention blocks. As in a traditional Transformer decoder [61], in each block, there is a multi-head self-attention layer and a multi-head cross-attention layer. The self-attention layer enables interactions between hand, object and video queries, and the cross-attention layer allows the query to extract object-oriented information from the visual content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Bounding Box Head. The hand query vectors \\( Q_{h} \\) and object query vectors \\( Q_{o} \\) are trained to predict the bounding box of the hands and objects respectively in each frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "Note these query vectors and summary vectors are at the video level; to predict boxes at frame level, we condition a summary vector \\( s_j \\) of object \\( j \\) on a learnable frame index vector \\( x_i \\) by concatenation of \\( s_j \\) and \\( x_i \\), and use a multi-layer perceptron \\( F_{box} \\) to project them onto a bounding box \\( \\hat{b}_{j,1} \\), where \\( i \\) is the frame number:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.194, + 0.892, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\hat {b} _ {j i} = F _ {\\text {b o x}} \\left(s _ {j}; x _ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.223, + 0.892, + 0.256 + ], + "angle": 0, + "content": "As a result, we will have a time series of bounding boxes \\((\\hat{b}_{j,1},\\hat{b}_{j,2},\\ldots \\hat{b}_{j,T})\\) from each \\(j^{\\mathrm{th}}\\) summary vector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Semantic Head. We assign semantic meanings to object summary vectors, standing for the object name/class. To achieve this, we project \\( s_j \\) onto a word embedding \\( \\hat{n} \\) with a multi-layer perception \\( F_{semantic} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.344, + 0.892, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\hat {n} = F _ {\\text {s e m a n t i c}} \\left(s _ {j}\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.371, + 0.689, + 0.387 + ], + "angle": 0, + "content": "3.3. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Vision-Text Matching. We follow EgoVLP [35] and use EgoNCE loss as the objective for matching between video-level embedding \\( v \\) and sentence-level embedding \\( t \\) of the narration. In one batch \\( \\mathcal{B} \\), the positive sample set \\( \\mathcal{P}_m \\) is made up of a sample \\( i \\) and other samples that share at least one noun and one verb with it: \\( \\mathcal{P}_m = \\{n \\in \\mathcal{B} \\mid \\text{noun}(n) \\cap \\text{noun}(m) \\neq \\emptyset, \\text{verb}(n) \\cap \\text{verb}(m) \\neq \\emptyset\\} \\). And for each sample \\( i \\), there is a hard negative sample \\( i' \\) sampled from the same video. Hence, the samples in the original batch \\( \\mathcal{B} \\) and their hard negative counterparts together form the new batch \\( \\widetilde{\\mathcal{B}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.623 + ], + "angle": 0, + "content": "The objective matching video-to-text (v2t) for a video embedding \\( v \\) is formulated as below; in practice the symmetric text-to-video (t2v) matching objective is also used (omitted for brevity):" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.632, + 0.925, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v 2 t}} ^ {\\mathrm {e g o}} = \\frac {1}{| \\mathcal {P} _ {m} |} \\sum_ {k \\in \\mathcal {P} _ {m}} \\log \\frac {\\exp \\left(v ^ {T} t _ {k} / \\tau\\right)}{\\sum_ {n \\in \\hat {\\mathcal {B}}} \\left(\\exp \\left(v ^ {T} t _ {n} / \\tau\\right) + \\exp \\left(v ^ {T} t _ {n ^ {\\prime}} / \\tau\\right)\\right)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Bounding Box Prediction. We use the 100DOH off-the-shelf hand object detector [53] to produce bounding boxes of two classes on each frame as supervision: hand and objects that are in contact with hands." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.903 + ], + "angle": 0, + "content": "There are two challenges in using the detections for training supervision: 1) the image detector acts at the image level independently and does not provide box-ID association over different frames in a clip; 2) many hands and objects are missed due to motion blur and domain gap in ego-centric videos. Therefore, we apply Hungarian matching between predicted boxes \\(\\hat{b}\\) and ground-truth boxes \\(b_{i}\\) on single frames independently, so that for each \\(b_{i}\\), we find the matched prediction \\(\\hat{b}_{\\sigma(i)}\\) to minimize the global matching" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13904" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "cost. The final loss on bounding boxes is computed as the sum of the \\(\\ell_1\\) loss and the Generalized IoU loss \\(L_{iou}\\) [51] on paired boxes:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.149, + 0.469, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {b o x}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) = \\mathcal {L} _ {\\mathrm {i o u}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) + \\left| \\left| b _ {i} - \\hat {b} _ {\\sigma (i)} \\right| \\right| _ {1} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.181, + 0.469, + 0.225 + ], + "angle": 0, + "content": "and, to tackle the problem of missing objects, we do not penalize boxes that are not matched to nouns unlike traditional detection tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.469, + 0.369 + ], + "angle": 0, + "content": "Object Class Prediction. We have noun embeddings from the video description and a set of predicted object name embeddings from the summary vectors, the task is to find the correspondence between them so that we can use the ground-truth nouns from the description as supervision for the predicted object names. As shown in figure 2, we align the nouns in the narrations and the names of objectboxes in two steps:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.47, + 0.447 + ], + "angle": 0, + "content": "(1) Object-noun alignment: We score the predicted object name embeddings \\(\\hat{n}\\) against the ground-truth noun embeddings \\(n\\) to construct a similarity matrix \\(C\\in \\mathbb{R}^{K\\times N}\\), where \\(K\\) is the number of object queries and \\(N\\) is the number of noun phrases in the description as following:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.459, + 0.469, + 0.492 + ], + "angle": 0, + "content": "\\[\nC (n, \\hat {n}) = \\frac {n \\cdot \\hat {n}}{\\left\\| n \\right\\| \\left\\| \\hat {n} \\right\\|} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.469, + 0.531 + ], + "angle": 0, + "content": "Cost matrix \\(-C\\) is used in Hungarian matching to select the matched summary vector for each noun phrase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.469, + 0.593 + ], + "angle": 0, + "content": "(2) Word-level contrastive training: We apply InfoNCE loss on the matched embeddings \\(\\hat{n}_j\\) and \\(n_j\\) against the embeddings of all the nouns \\(n_k'\\) in Ego4D taxonomy dictionary \\(\\mathcal{D}\\) [16]:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.606, + 0.469, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {w o r d}} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\log \\frac {\\exp \\left(\\hat {n} _ {j} ^ {T} n _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {D}} \\exp \\left(\\hat {n} _ {j} ^ {T} n _ {k} ^ {\\prime} / \\tau\\right)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Training Loss. The total training objective is the sum of the vision-text matching loss and the auxiliary losses on object vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.74, + 0.469, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {v} 2 \\mathrm {t}} ^ {\\mathrm {e g o}} + \\mathcal {L} _ {\\mathrm {t} 2 \\mathrm {v}} ^ {\\mathrm {e g o}} + \\mathcal {L} _ {\\text {b o x}} + \\lambda_ {w o r d} \\mathcal {L} _ {\\text {w o r d}} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.187, + 0.785 + ], + "angle": 0, + "content": "3.4. Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Once trained, the model acts as a standard ego-centric vision-language video model which operates just on video frames and text descriptions, without requiring further access to object boxes or detectors. However, if desired, hand and object box detections and names can be read out for each frame at inference using the summary vectors, which can be used for grounding the input text description." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.092, + 0.66, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.228, + 0.672, + 0.241 + ], + "angle": 0, + "content": "Step 1. Object-noun alignment" + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.094, + 0.871, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.228, + 0.889, + 0.241 + ], + "angle": 0, + "content": "Step 2. Word-level contrastive training" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Figure 2. Training the model to predict object classes. Left: Object-noun alignment. First, the nouns in the video descriptions are matched against the predicted classes using Hungarian matching, to choose the most matched summary vectors. Right: Word-level contrastive training. Next, we supervise the matched predicted classes using a contrastive objective [46] against all the nouns in Ego4D taxonomy, to have similar embeddings as the corresponding nouns." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.377, + 0.659, + 0.394 + ], + "angle": 0, + "content": "4. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.495 + ], + "angle": 0, + "content": "In section 4.1, details of extracting the hand and object detections from 100DOH pre-trained detector are summarized, followed by the architectural details in section 4.2. Finally, in section 4.3 the training pipeline, model and input specification, and optimization details are provided. More details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.506, + 0.88, + 0.522 + ], + "angle": 0, + "content": "4.1. Weak Supervision from Pre-trained Detector" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.53, + 0.892, + 0.698 + ], + "angle": 0, + "content": "We uniformly sample four frame from each clip in the EgoClip dataset [35] as the input to the 100DOH hand-object detector [53]. The short side of the frame is resized to 640 pixels. There are 16 million frames in total, and the average number of boxes detected per frame is 1.96 for hand and 1.67 for object. Among all the frames, about \\(15.8\\%\\) frames have no hands detected and \\(17.9\\%\\) frames have no object detected. The average size of hand boxes is \\(2.8\\%\\) of the frame size, while the average size of object boxes is \\(19.4\\%\\) of the frame size. We use the top 2 hands and top 4 objects detected in the scene as supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.634, + 0.724 + ], + "angle": 0, + "content": "4.2. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.734, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The object-aware module is a 6-layer cross-attention transformer with 8 attention-heads in each layer. The hidden dimension in cross-attention layers is 768, and the video embedding, object embeddings and text embeddings are projected to 256 dimensions before computing the cosine similarity score. We set the number of hand queries to 2, number of object queries to 12, which is designed to be larger than the maximum number of objects in the supervision. We use TimeSformer-L (TSF-L) [6] as the visual encoder, and a 12-layer Transformer [61] from LaViLa [77] as the text encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13905" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.096, + 0.424, + 0.109 + ], + "angle": 0, + "content": "Text: #C C cuts the box with the utility knife" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.111, + 0.48, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.191, + 0.465, + 0.204 + ], + "angle": 0, + "content": "a) Object boxes predicted by off-the-shelf hand-object detector" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.206, + 0.48, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.286, + 0.442, + 0.299 + ], + "angle": 0, + "content": "b) Vision-text grounding results from the proposed model" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.096, + 0.882, + 0.109 + ], + "angle": 0, + "content": "C C puts the spatula in his right hand on the chopping board" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.111, + 0.89, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.191, + 0.877, + 0.204 + ], + "angle": 0, + "content": "a) Object boxes predicted by off-the-shelf hand-object detector" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.205, + 0.89, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.286, + 0.862, + 0.299 + ], + "angle": 0, + "content": "b) Vision-text grounding results from the proposed model" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.893, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Visualization of text grounding on EgoClip. We show the comparison between detections from off-the-shelf 100DOH hand-object detector [53] (used for training supervision) and the predicted boxes from our model respectively. (a) The detections are noisy, objects are missed, and there is no temporal association of the detected boxes across frames. (b) The trained model learns temporally consistent tracks as well as object categorization using only noisy frame-level box-supervision and weak supervision from the texts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.375, + 0.238, + 0.391 + ], + "angle": 0, + "content": "4.3. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.566 + ], + "angle": 0, + "content": "In training, we uniformly sample \\(4 \\times 224 \\times 224\\) RGB frames from each clip. Hand and object boxes are preextracted from these 4 frames using off-the-shelf detector and used as supervision. We keep the visual and text encoder frozen in training. Only the object-aware decoder, query vectors, and the MLP projection layer on text embedding parameters are learned during training. We train the model for 5 epochs on one A6000 GPU, with batch size 128. We use AdamW [39] as the optimizer and set the learning rate to \\(3e - 5\\). \\(\\lambda_{word}\\) is set to 0.5 to balance the scale of the four losses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.582, + 0.21, + 0.599 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.684 + ], + "angle": 0, + "content": "Section 5.1 introduces all the datasets we used for training and evaluation, followed by the evaluation protocols in section 5.2. Finally, we discuss the ablation studies (section 5.3) and compare to prior SOTA methods on different benchmarks (section 5.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.18, + 0.709 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Ego4D/EgoNLQ/EgoMQ/EgoClip/EgoMCQ [16, 35]. Ego4D is a massive-scale dataset focusing on ego-centric videos. It contains 3670 hours video for many different tasks, including action anticipation, AV diarization, etc. EgoNLQ and EgoMQ is a subset for natural language queries and moment query, designed for testing the models' episodic memory and long-term video understanding. [35] proposes a new subset EgoClip for vision-language pre-training, comprising 3.8M clip-text pairs. They also introduce EgoMCQ (i.e., Egocentric Multiple-Choices-Question) as a downstream evaluation dataset for the pre-training. Given a text query, the model tasked to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.893, + 0.422 + ], + "angle": 0, + "content": "choose the paired video clip from 5 candidates. The evaluation metrics is 'intra-video' and 'inter-video' accuracy, depending on where the candidates are chosen from." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.559 + ], + "angle": 0, + "content": "Epic-Kitchens-MIR [11]. Epic-Kitchens is a large-scale ego-centric dataset with 100-hour activities recorded in kitchens. Epic-Kitchens-MIR is a subset with about 9881 clip-text pairs for vision-language retrieval. It is designed for multi-instance retrieval. The model is evaluated on retrieving the paired text/video given a query text/video. The evaluation metrics are mean average Precision (mAP) and normalized Discounted Cumulative Gain (nDCG)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.652 + ], + "angle": 0, + "content": "VISOR [12] is a benchmark built on Epic-Kitchens for segmenting hands and active objects in egocentric video. It has pixel-level annotations covering 36 hours of untrimmed video and 257 object classes. We utilize this annotations in its val split for evaluations on vision-text grounding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.759 + ], + "angle": 0, + "content": "EGTEA [34] contains 28 hours of cooking activities from 86 unique sessions of 32 subjects. Fine-grained actions are classified into 106 classes. We retrieve the text-descriptions of action classes to evaluate the model for action classification on the test set of all three splits. We measure the performance using mean-class accuracy and top1 accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.691, + 0.785 + ], + "angle": 0, + "content": "5.2. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.89, + 0.824 + ], + "angle": 0, + "content": "We evaluate the performance of our model in the following three aspects:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Zero-shot transfer. To test the transferability and generalization ability, we conduct zero-shot evaluation on multiple-choice questions (EgoMCQ), multi-instance retrieval (EpicKitchens-MIR), action classification (EGTEA). Among these datasets, videos in EgoMCQ are from the same data" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13906" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "source as our pre-training dataset Ego4D. Other datasets demonstrate a domain gap, hence, evaluate for transferable representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.145, + 0.468, + 0.252 + ], + "angle": 0, + "content": "Episodic memory. To evaluate the richness of the representations learned by our model, we use the video representations to solve Episodic memory tasks in Ego4D. Following [35, 77], we pre-extract the video features from our model first. Using these pre-computed features as input, we train a VSLNet [72] for temporal localization in NLQ, and train a VSGN [76] for moment retrieval in MQ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.468, + 0.558 + ], + "angle": 0, + "content": "Vision-language grounding. Due to the lack of ground-truth for object-grounding in EgoClip, we evaluate the grounding ability of the model on VISOR instead. VISOR is an egocentric dataset with scenes in the kitchens, where frames are annotated sparsely with segmentation masks and object names. We re-propose the manually annotated segmentation masks in it to extract ground-truth bounding-boxes for hands and in-contact objects. To carry out the zero-shot evaluation, we take the annotated frames in the val split, filter out the not-in-contact objects in each frame, and convert all the segmentation masks to bounding boxes as ground truth. The predicted object boxes are matched with ground-truth object boxes using noun alignment as during training (eq. (5)), while the left/right hands are predicted from the first and the second hand queries respectively. We repeat a single frame 4 times temporally, and resize it to \\(224 \\times 224\\) pixels to be consistent with the pre-training resolution. The predicted bounding-boxes are evaluated to be correct if their centers lie inside the ground-truth bounding-boxes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.188, + 0.589 + ], + "angle": 0, + "content": "5.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.468, + 0.705 + ], + "angle": 0, + "content": "Losses. We ablate the combination of losses on three downstream benchmarks in table 1. Results showing that having both \\( L_{box} \\) and \\( L_{word} \\) leads to the best performance. With the same architecture, when training the model using only \\( L_{Ego} \\) without introducing any object-awareness, the performance is \\( 2\\% \\) lower on EK100-MIR and EGTEA compared to the object-aware one." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.826 + ], + "angle": 0, + "content": "
LossesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
LEgo93.761.835.936.644.937.6
LEgo + Box94.262.736.737.445.338.5
LEgo + Word93.761.936.737.445.838.1
LEgo + Box + Word94.563.037.537.846.639.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.833, + 0.468, + 0.889 + ], + "angle": 0, + "content": "Table 1. Ablation on training objectives for zero-shot transfer tasks. Introducing the object-awareness by having box and word supervision helps the model to achieve better transfer results on EK100-MIR and EGTEA." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.187 + ], + "angle": 0, + "content": "
Detector input resEgoMCQEKEGTEAVISOR
InterIntraAvg mAPAvg nDCGTop1MeanLoc Acc
256p94.263.235.734.642.036.0-
256p94.563.036.937.044.338.968.2
640p94.563.037.537.839.146.678.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Table 2. Ablation on box quality for zero-shot transfer tasks. We extract boxes using \\(256\\mathrm{p}\\) and \\(640\\mathrm{p}\\) images as input to the detector respectively, resulting in boxes of different qualities as supervision in training." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.262, + 0.88, + 0.357 + ], + "angle": 0, + "content": "
# Obj QueriesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
494.162.837.837.645.938.1
894.562.737.738.045.537.9
1294.563.037.537.846.639.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.362, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Table 3. Ablation on the number of object queries for zero-shot transfer tasks. The number of queries do not have a big impact on EK100. Larger number of queries shows a boost on mean-class accuracy on EGTEA, and smaller number of queries is better on intra-video accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Quality of detected boxes. The extent to which a model can acquire object-level information is constrained by the quality of the bounding boxes from the off-the-shelf detector. To investigate how much the quality of boxes affects our training, we detect hand and object boxes on EgoClip training set using 100DOH [53] with 256p and 640p images as input – with larger image size, the objects should be more precisely delineated. We use the two sets of boxes as supervision in our training and show results in table 2. Even when training with noisy boxes from 256p, our model is better than the previous SOTA model. When boxes from 640p are used, the averaged zero-shot transfer performance is further improved by \\(1\\%\\) on Epic-Kitchens and EGTEA, showing that our method can bring larger improvement over the non-object-aware method when given better boxes. Furthermore, a significant boost on the grounding results is also observed on VISOR when using boxes with better quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Number of object queries. We use different number of query vectors in the object-aware decoder to see its impact on both vision-language tasks and the grounding task. We show zero-shot transfer results in table 3 and vision-text grounding results in table 7. The number of queries has relatively small impact on vision-language representation learning, the performance gaps on zero-shot transfer tasks are mostly small than \\(1\\%\\). However, a smaller number of object queries leads to much better results for in-contact object grounding on VISOR, 4 queries is better than 12 objects by \\(3.6\\%\\) in localization accuracy. The reason is that too" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13907" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.894, + 0.229 + ], + "angle": 0, + "content": "
MethodBackboneBatch sizeObject awareHard negEgoMCQEK100-MIREGTEA
Inter-videoIntra-videoV-TmAP T-VAvgV-TnDCG T-VAvgTop1-AccMean-Acc
EgoVLPTSF-B512X✓90.657.226.020.623.328.827.027.917.6-
LaViLaTSF-B1024XX93.859.935.126.630.933.730.432.0-28.9
LaViLaTSF-L1024XX94.563.140.032.236.136.133.234.640.134.1
LaViLa*TSF-L1024XX94.263.239.731.735.736.133.234.642.036.0
Ours*†TSF-L128XX93.760.539.730.335.037.334.535.944.836.3
Ours*‡TSF-L128X✓93.761.840.731.135.938.335.036.644.937.6
Ours*TSF-L128✓✓94.563.042.332.737.539.336.237.846.639.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.233, + 0.895, + 0.292 + ], + "angle": 0, + "content": "Table 4. Comparison to SOTA results on zero-shot transfer to EgoMCQ, EK100-MIR and EGTEA. We compared to EgoVLP and LaVILA, two previous SOTA models pre-trained on EgoClip. Our object-aware model has achieved comparable results on multiple-choice questions on EgoMCQ, and SOTA results on multi-instance retrieval task on EpicKicthens and action classification on EGTEA. Model without * use center cropping in evaluation, while * denotes the usage of resizing instead of cropping. Ours† and Ours‡ stands for different variants of our model depending on whether object-aware losses and hard negative sampling is used in training." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.304, + 0.457, + 0.366 + ], + "angle": 0, + "content": "
MethodPredicted BoxesEgoMCQEKEGTEA
InterIntraAvg mAPAvg nDCGMeanTop1
Oursobj94.062.136.837.145.137.6
hand+obj94.563.037.537.846.639.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.37, + 0.471, + 0.412 + ], + "angle": 0, + "content": "Table 5. The impact of having hand boxes in the box prediction. Having hand boxes in the prediction helps on all the zero-shot evaluation benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.456 + ], + "angle": 0, + "content": "many queries result in a large number of predicted boxes, which increases the probability of mis-matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Impact of hand boxes. We ablate the impact of having hand boxes as supervision in our training. Results are shown in table 5, training the model to predict hand as well as objects help the model to get about \\(1\\%\\) higher zero-shot transfer performance on EK and EGTEA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.308, + 0.574 + ], + "angle": 0, + "content": "5.4. Comparison to the SOTA" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.702 + ], + "angle": 0, + "content": "Zero-shot transfer. We compare to previous SOTA in table 4. Our model is comparable on EgoMCQ and better on EK-MIR and EGTEA, showing its good zero-shot transferability. Due to limited compute resources, we are not able to unfreeze the visual backbone to train end-to-end or increase the batch size further. Despite these disadvantages, our method outperforms the previous SOTA on two tasks for models that have been trained end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The main difference between LaViLa(L) and ours is the object-aware training and hard negative sampling; Ours† in table 4 is a LaViLa(L) model with an extra Transformer decoder, which is trained with only InfoNCE loss on video and sentence embeddings. Without object-awareness and hard sampling, it gets better accuracy on EGTEA and better nDCG on EK100 due to more parameters added, but falls behind on EgoMCQ and mAP on EK100. Applying hard negative sampling (Ours‡) and object-aware training (Ours) brings improvement across the board. The most obvious boost comes from inducing object-awareness, bringing \\(1.5\\%\\) improvement on average. And the results could be further improved by obtaining better pseudo-boxes, as the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.332 + ], + "angle": 0, + "content": "magnitude of boost from learning objects is closely related to the box quality (as shown in table 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Episodic memory. Results on EgoNLQ and EgoMQ are shown in table 6. The two tasks test the video understanding on several-minutes long videos. In these experiments, trained video and text backbones are used as feature extractor, and extra modules are trained on the long feature sequences for natural language querying and memory querying. Therefore, the richer the information encoded in the features, the better the results will be. We list the models trained on the same amount of video and text data in black, results show that our model are better than the previous SOTA on all the metrics in the two tasks. This is because features from the object-aware model have captured more object information, thus enabling better precision and recall on localization and retrieval. InternVideo [63] and NaQ [50] are trained on more video or text data, and are included for completeness. Please note that our method is compatible with any visual language backbones, using better backbones might also lead to better results." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.642, + 0.892, + 0.808 + ], + "angle": 0, + "content": "
MethodBatch sizeEgoNLQEgoMQ
mIOU@0.3mIOU@0.5R1@0.5R5@0.5mAP
R1R5R1R5
SlowFast-5.510.73.16.625.246.26.0
EgoVLP51210.818.86.813.530.152.011.4
LaViLa(B)102410.519.16.713.6---
LaViLa(L)102412.122.47.315.432.556.113.4
Ours12813.223.37.915.633.456.716.0
VideoIntern [63]14k16.523.010.116.1--23.6
ReLER + NaQ [50]204819.323.611.615.5---
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.815, + 0.895, + 0.899 + ], + "angle": 0, + "content": "Table 6. Comparison to SOTA results of fine-tuning on EgoNLQ and EgoMQ. Our object-aware model encodes richer information in the visual representations, hence obtaining better results on all the metrics in NLQ and MQ task in Ego4D episodic memory benchmark. We list other SOTA models (in grey) trained with more video and text data for completeness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13908" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.194 + ], + "angle": 0, + "content": "
ModelObject Assignment# queriesLoc Accuracy
Detector [53]Random-37.1
GT matching-41.3
OursPredicted482.3
881.2
1278.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.286 + ], + "angle": 0, + "content": "Table 7. In-contact object localization accuracy on VISOR. Our model does better on in-contact object localization after weakly-supervised training compared to the baseline (the Detector with Random or GT object assignment), which is the source of supervision in our pre-training. The improvement is due to increased recall of our model over the baseline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.295, + 0.345, + 0.312 + ], + "angle": 0, + "content": "5.5. Evaluating Object Grounding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.502 + ], + "angle": 0, + "content": "Qualitative results on EgoClip. In fig. 3 we show the grounding results on EgoClip after training as compared to the supervision from the 100DOH detector. After training, the predictions from our model find the missing objects because we do not penalize extra bounding boxes predicted, but select the active object/object of interest through matching noun embeddings. It also learns temporal association of object bounding-boxes as the result of using the same object summary vector to predict the boxes over all the frames (as in eq. (1)); the summary vector attends to visually similar features corresponding to the same object across the clip without any explicit supervision for temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.788 + ], + "angle": 0, + "content": "Quantitative and qualitative results on VISOR. In table 7 we show the text-grounding results on VISOR. We take the predictions from the hand-object detector [53] as our baseline. Since the predictions only detect hands and objects without an object class name, we associate the predicted boxes with ground-truth boxes in VISOR in two ways: (a) random: we assign predicted object boxes to GT object boxes randomly, (b) GT matching: we use Hungarian matching to find the predicted boxes with highest IoU against the GT object boxes. However, even when matched using GT information, the baseline detector does not achieve a high accuracy due to poor recall. Results show grounding ability of our model is \\(40\\%\\) better than the baseline using only weak supervision from the video descriptions. We also show the qualitative results on detecting different number of hands and objects in fig. 4, our model has a much higher recall when compared to baseline detector when operating at the same resolution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.196, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce a method to learn object-aware ego-centric video representations using noisy supervision from pre-trained hand-object detectors. The object-representations so learned show strong zero-shot transfer across various downstream tasks and datasets, mirroring" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.885, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.373, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Figure 4. Grounding visualization on the Epic-Kitchen VISOR val split. When operating at the same resolution, our model shows better grounding ability on hands and objects compared to the baseline 100DOH detector [53] (with GT matching) used for training. Note that the low IoU on hands on the third column is a result of the GT 'hand' segmentation mask covering the arm by definition, while the detector and our model are trained to localize only palm and fingers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.801 + ], + "angle": 0, + "content": "the performance improvement from object-aware training on images [70]. The model uses standard neural modules (i.e., transformers), and does not require any object boxes or detectors as input during inference, making it widely applicable as a drop-in replacement for training video-language models. Even though the model is trained with sparse and noisy object supervision at the frame-level (without temporal associations), during inference dense temporal bounding-box tracks and category predictions can be obtained, which are superior to the predictions from the base hand-object detector used for training. There are several avenues for improvement. Our model uses the pre-trained video encoder operates at a small resolution \\(224 \\times 224\\) which makes detecting small objects difficult. Further, four frames are sampled uniformly from the clip regardless of its length which can cause difficulties due to temporal aliasing. Nevertheless, we hope our work will inspire further research in learning transferable object-aware representations for videos." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.818, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Acknowledgements. This research is funded by a Google-DeepMind Graduate Scholarship, a Royal Society Research Professorship, and EPSRC Programme Grant VisualAI EP/T028572/1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13909" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.47, + 0.224 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In Proc. ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.225, + 0.471, + 0.293 + ], + "angle": 0, + "content": "[3] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.293, + 0.471, + 0.334 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proc. ICCV, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[5] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.443 + ], + "angle": 0, + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In Proc. ICML, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.499, + 0.471, + 0.553 + ], + "angle": 0, + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proc. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.554, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.608, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[10] Noam Chomsky and David W Lightfoot. Syntactic structures. Walter de Gruyter, 1957. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.634, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proc. ECCV, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.702, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[12] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Datasets and Benchmarks Track, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.771, + 0.471, + 0.84 + ], + "angle": 0, + "content": "[13] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.84, + 0.471, + 0.895 + ], + "angle": 0, + "content": "[14] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurlPS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.471, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "[15] Philip Gage. A new algorithm for data compression. C Users Journal, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[16] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proc. CVPR, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.215, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[18] Paul Henderson and Christoph H. Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.256, + 0.892, + 0.311 + ], + "angle": 0, + "content": "[19] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. arXiv preprint arXiv:2110.06915, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[20] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proc. ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.365, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[22] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proc. CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[23] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.529, + 0.892, + 0.556 + ], + "angle": 0, + "content": "[24] Ranjay Krishna, Ines Chami, Michael Bernstein, and Li Fei-Fei. Referring relationships. In Proc. CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.624, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[26] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In Proc. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[27] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[28] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proc. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[29] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.843, + 0.892, + 0.897 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. ICML, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proc. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.203, + 0.469, + 0.257 + ], + "angle": 0, + "content": "[33] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.257, + 0.469, + 0.298 + ], + "angle": 0, + "content": "[34] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proc. ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.298, + 0.469, + 0.367 + ], + "angle": 0, + "content": "[35] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.469, + 0.42 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[37] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 704-721. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[38] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.639 + ], + "angle": 0, + "content": "[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.639, + 0.469, + 0.693 + ], + "angle": 0, + "content": "[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[42] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proc. CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[43] Effrosyni Mavroudi and René Vidal. Weakly-supervised generation and grounding of visual descriptions with conditional generative models. In Proc. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[44] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proc. ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.858, + 0.469, + 0.883 + ], + "angle": 0, + "content": "[45] Marvin Minsky. Society of mind. Simon and Schuster, 1988. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.884, + 0.469, + 0.898 + ], + "angle": 0, + "content": "[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "sensation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[47] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE conference on computer vision and pattern recognition, pages 2847-2854. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.175, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[48] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuytelaars. Revisiting spatio-temporal layouts for compositional action recognition. In Proc. BMVC, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.215, + 0.892, + 0.283 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.283, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[50] Santhosh Kumar Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Naq: Leveraging narrations as queries to supervise episodic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6694-6703, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.353, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[51] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.407, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[52] Kate Saenko, Ben Packer, C Chen, Sunil Bandla, Y Lee, Yangqing Jia, J Niebles, Daphne Koller, Li Fei-Fei, Kristen Grauman, et al. Mid-level features improve recognition of interactive activities. Technical report, Dept. of Elec. Engg. and Computer Sc., University of California, Berkeley, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[53] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proc. CVPR, 2020. 1, 2, 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Proc. ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[55] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.611, + 0.892, + 0.652 + ], + "angle": 0, + "content": "[56] Elizabeth S Spelke, Karen Breinlinger, Janet Macomber, and Kristen Jacobson. Origins of knowledge. Psychological review, 99, 1992. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.892, + 0.693 + ], + "angle": 0, + "content": "[57] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In Proc. ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[58] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[59] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. *NeurIPS*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[60] Joshua B Tenenbaum, Charles Kemp, Thomas L Griffiths, and Noah D Goodman. How to grow a mind: Statistics, structure, and abstraction. Science, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.829, + 0.892, + 0.883 + ], + "angle": 0, + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.884, + 0.892, + 0.898 + ], + "angle": 0, + "content": "[62] Xiaolong Wang and Abhinav Gupta. Videos as space-time" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13911" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.361, + 0.106 + ], + "angle": 0, + "content": "region graphs. In Proc. ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.107, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[63] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.175, + 0.47, + 0.243 + ], + "angle": 0, + "content": "[64] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.243, + 0.469, + 0.298 + ], + "angle": 0, + "content": "[65] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.298, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[66] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In Proc. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.379 + ], + "angle": 0, + "content": "[67] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proc. ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.469, + 0.434 + ], + "angle": 0, + "content": "[68] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[70] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.544, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[71] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proc. ACCV, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[72] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE PAMI, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.639, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[73] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[74] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In European Conference on Computer Vision, pages 127-145. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[75] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[76] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.871, + 0.469, + 0.898 + ], + "angle": 0, + "content": "[77] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.867, + 0.106 + ], + "angle": 0, + "content": "models.arXiv preprint arXiv:2212.04501,2022.2,5,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.107, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[78] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proc. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.161, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[79] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. arXiv preprint arXiv:1805.02834, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.215, + 0.892, + 0.27 + ], + "angle": 0, + "content": "[80] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.27, + 0.892, + 0.324 + ], + "angle": 0, + "content": "[81] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In Proc. ECCV, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13912" + } + ] +] \ No newline at end of file diff --git a/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_origin.pdf b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9a9203008f70a2e69d7bd7a03decd107a79a242 --- /dev/null +++ b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/605acb09-5a50-4e6f-a563-f1799bbb19a1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:970c8e7075dffa6c8dc81e654c8e6dacabdb22dc9a20e353ba87ec3a4e900581 +size 1095353 diff --git a/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/full.md b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1242fccb69d78becd90c85f323abb47c0d22bbc1 --- /dev/null +++ b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/full.md @@ -0,0 +1,375 @@ +# Helping Hands: An Object-Aware Ego-Centric Video Recognition Model + +Chuhan Zhang +VGG, University of Oxford +czhang@robots.ox.ac.uk + +Ankush Gupta Google DeepMind, London ankushgupta@gmail.com + +Andrew Zisserman +VGG, University of Oxford +az@robots.ox.ac.uk + +# Abstract + +We introduce an object-aware decoder for improving the performance of spatio-temporal representations on egocentric videos. The key idea is to enhance object-awareness during training by tasking the model to predict hand positions, object positions, and the semantic label of the objects using paired captions when available. At inference time the model only requires RGB frames as inputs, and is able to track and ground objects (although it has not been trained explicitly for this). + +We demonstrate the performance of the object-aware representations learnt by our model, by: (i) evaluating it for strong transfer, i.e. through zero-shot testing, on a number of downstream video-text retrieval and classification benchmarks; and (ii) by using the representations learned as input for long-term video understanding tasks (e.g. Episodic Memory in Ego4D). In all cases the performance improves over the state of the art—even compared to networks trained with far larger batch sizes. We also show that by using noisy image-level detection as pseudo-labels in training, the model learns to provide better bounding boxes using video consistency, as well as grounding the words in the associated text descriptions. + +Overall, we show that the model can act as a drop-in replacement for an ego-centric video model to improve performance through visual-text grounding1. + +# 1. Introduction + +In visual-language models there has been a recent move to explicitly build object awareness into the vision module by adding specialized and bespoke components, or using entirely object-centric architectures. The motivation for this partly comes from the attractive compositional nature of objects and their inter-relationships in language, which enables inexhaustible novel combinations [10, 45], and partly from infant cognitive studies that stress the im + +portance of objects in early visual development [29, 56, 60]. Examples in the video domain include explicit internal object representations [2], e.g., through RoI-align [17] pooled features either from a pre-trained region-proposal network (RPN) [2, 52, 57, 62], or from bounding-box coordinates taken as input [19, 42, 48, 71]. This contrasts with the large body of work where standard representations are learnt end-to-end without any explicit factorization into objects/entities, such as dual-encoder vision-language models in the image [21, 49] and video domains [4, 64]. + +In this paper, we take a different (middle) path and instead use a vanilla video transformer architecture and induce object-awareness into the video representation by tasking the model to predict object-level properties, such as their localization and semantic categories, only during training. + +Our target domain is ego-centric video [11, 16], and we tailor the object properties used to this. In ego-centric videos the actor [57] is often present through their hands, and we therefore task the network to predict both the hands and the principal objects they interact with. As will be seen, this simple object-aware training boosts the performance of pre-trained video-language architectures significantly, and leads to state-of-art performance across multiple ego-centric benchmarks. During inference, the model requires only RGB frames as input, and operates as a standard video-language network. + +In more detail, our model is built on top of a pre-trained video-language dual encoder architecture (where there are separate encoders for the video and text data). We add an additional, but vanilla, transformer decoder head [61], and train with DETR/Mask2former [7, 9] query vectors and object loss for hands and other objects. The intuition is that these additional query vectors help the model to attend to and track the hands and salient objects in the scene (these are the 'helping hands' of the paper title). Importantly, we do not require dense frame level ground truth for this training. Rather, we obtain somewhat noisy and temporally sparse annotations automatically from a hand object detector [53], and use these to provide prediction targets for the frames where they are available. This opportunistic use of annotations is pragmatic as object detectors trained on third + +person datasets (such as COCO) do not perform so well on the ego-centric domain, where the scenes are more crowded and objects are often small and can be motion blurred. By only requiring annotations for a subset of frames, where they can be reliably produced automatically, we are able to train on large-scale data without requiring expensive manual supervision. + +Although we train with noisy and sparse image-level object localization, our model can learn to predict better and denser bounding-box trajectories through large-scale training due to the spatio-temporal consistency which naturally presents in videos. Also, it is able to predict semantic grounding by learning to map the object appearance to the nouns in the video captions. + +It is worth noting that we are using hand detectors because hands are a common and important object in ego-centric videos. However, the object-centric method we are proposing has greater scope than ego-centric videos and can be applied to other scenarios with other object types providing the 'helping-hand'. + +In summary, we make the following contributions: (i) We propose a method to induce object-awareness in video-language models for an architecture composed of standard neural modules. The model only requires RGB frames as inputs, and thus is a drop-in replacement for any ego-centric video model. +(ii) The model can be trained opportunistically using available and sparse frame-level and noisy annotations, produced automatically. +(iii) We demonstrate state-of-the-art strong (zero-shot) transfer for cross-modal retrieval to other ego-centric datasets namely, EpicKitchens-MIR and EGTEA improving prior art by $2 - 4\%$ . +(iv) We evaluate the grounding quantitatively using the EpicKitchens-VISOR dataset [11, 12] and find that the model outperforms the base hand-object detector used for training supervision. +(v) Finally, we also demonstrate that the representations learned can be used as input in long-term video understanding tasks like EgoNLQ and EgoMQ. The objectiveness in the representation helps the model outperform other models trained on the same training set on these two tasks. + +# 2. Related Work + +Vision and Language Representation Learning. Different from transferring representations learned for classification on a fixed set of object categories [27, 54], recent vision-language pre-training (VLP) works leverage largescale supervision from free-from text descriptions of images and videos. These methods use image captions [49] or video sub-titles [44, 64] with either independent dual encoders for the visual and text modalities [4, 21, 41], or via + +joint encoders with cross-attention across modalities [1, 30, 31, 69]. We also use dual-encoders which are kept frozen due to compute limitations. To explicitly build-in object-awareness into image representations, object-level features extracted from pre-trained object-detectors are aligned with the text descriptions [8, 33, 40, 58, 75, 78] The object-level text alignment is further augmented with the object-box prediction task for grounding in [13, 23, 70, 73]. VLP for ego-centric videos has recently been explored [35, 77] to bridge the domain gap between representations learned from third-person videos found commonly [44], and first-person ego-centric videos. We further extend image based object-aware VLP methods to ego-centric videos by training to predict the auto-generated hand-object boxes extracted from pretrained detectors [53], while requiring only RGB input during inference, making our model a drop-in replacement for ego-centric VLP models albeit with enhanced object-awareness. + +Weakly Supervised Video Text Grounding. A key challenge for grounding in videos is the lack of large-scale object-level annotations for videos. While such annotations are readily available for synthetic datasets [68], expensive manual annotation is required for real videos. + +Hence, weakly supervised method have been developed which leverage the video sub-titles/description to map nouns/verbs to regions in frames. This is typically achieved by first extracting bounding-box/segmentation regions from pre-trained detectors for objects and humans, and aligning them with keywords using max-margin [79] or constrastive [32] objectives. Similarly, [43] also align words in the video captions to regions from pre-trained RPN by modelling the region-word associations as latent variables of a conditional generative variational auto-encoder. More recently, [59] use cross-attention across text and candidate regions, and find (soft-) associations based on attention weights. While our model similarly absorbs object bounding-boxes and category information obtained from pre-trained object and hand detectors during training to induce object-aware representations, these detectors are not required during inference. + +Object-Oriented Learning in Videos. Most vision tasks for images involve objects [22, 24, 25, 36]. In videos, there are also a broad range of object-oriented tasks: some works treat learning object-level information as an end task, they design models to predict object bounding-boxes and masks [5, 26, 37, 38, 65, 67, 74]; other works use object knowledge to achieve some other high-level tasks, for instance, object-centric scene generation [28, 66] and decomposition [14, 18, 55], and action recognition [3, 19, 20, 47, 62, 71, 80]. Our method uses object information to as- + +![](images/818085653772481dc7168d9d221b78e4ad5a0db93b701241b2a615536ea8ca64.jpg) +Figure 1. The object-aware vision-language model architecture. The architecture is made up of three parts: A video backbone, a text backbone and a object-aware decoder. The decoder is a cross-attention transformer, it takes the visual feature map $F'$ as keys and values, which are attended by a set of learnable queries. In these permutation-invariant query vectors, the hand and object queries $Q_h, Q_o$ are trained to be object-aware and predict the localization and class of hands and objects. The video query $q_v$ attends to both the visual feature map through cross-attention layers, and the object feature map through self-attention layers in the Transformer decoder and output a video-level embedding $v$ to be matched with the text embedding $t$ . + +sist vision-language alignment in egocentric videos, so that the model learns grounded video representations that can generalize better. SViT [3] uses object queries shared between images and videos in order to predict hand-object bounding boxes in videos, whilst only requiring image-level supervision during training. However, the object queries are not used for vision-language alignment. Our previous work [71] showed that encoding object-level information in the model helps transfer learning in various video understanding tasks, but the model required GT bounding boxes as input during inference. In contrast, in this work we model does not require this information during training, and we show box prediction and vision-language alignment can be combined and benefit both in-domain and out-of-domain tasks. + +# 3. Object-Aware Vision-Language Learning + +We first describe the task of object-aware vision-language learning and the architecture of our model. These are followed by the training objectives for vision-language matching and weakly-supervised text grounding. + +# 3.1. Overview + +In vision-language representation learning the training data consists of short video clips (RGB frames) and an + +associated free-form text string containing words that describe the visual content. Typically, dual encoders are trained on this paired data, with a visual-encoder that ingests the video clip and a language-encoder that ingests the text [4, 35, 41, 49, 64]. The dual encoders are trained with a contrastive objective [46] such that the cosine similarity for matching vision-text pairs is optimized to be higher than the similarity of negative/not-matching pairs. This pretraining objective enables evaluation on downstream vision-language tasks like video-text retrieval and action classification in a zero-shot manner [49]. + +Our object-aware model follows the data, training and evaluation pipeline as above, except that the model is also tasked to output object-level information (e.g., bounding boxes and object categories) during training. By tasking the model to predict object bounding boxes and names which can be matched to the nouns in the narration, the model learns grounded and fine-grained correspondence between the modalities. The object-level prediction is used as an auxiliary objective in training but not used at inference time. + +In more detail, as shown in figure 2, there are two types of object-level prediction: (a) hand and object bounding boxes; and (b) object names. Since ground truth of boxes and object names are not available, and most traditional detectors [7, 81] fail to identify objects well in ego-centric + +videos, we cannot rely on strong supervision for the predictions. Instead, we generate bounding box targets (for the hands and other objects) using a robust off-the-shelf hand-object detector [53], though these targets will only be available for some of the frames and are noisy. While for object name prediction, we use a weakly supervised method to align the predicted names with nouns in the paired narration (detailed in section 3.3). In both cases, the supervision is opportunistic and only applied when available. + +# 3.2. Architecture + +Dual Encoder. We use dual encoders as our visual and text backbone for efficiency. The visual encoder ingests a video clip $F$ of RGB frames $F = (f_{1}, f_{2}, \ldots, f_{T})$ , where $f_{i} \in \mathbb{R}^{H \times W \times 3}$ and $T$ is the number of frames. The clip $F$ is encoded by a Video Transformer [6] which tokenizes the frames by 3D patches to produce a spatially downsampled feature map $F' = (f_{1}', f_{2}', \ldots, f_{T}')$ , where $f_{i}' \in \mathbb{R}^{H' \times W' \times C}$ . It outputs a visual embedding $v \in \mathbb{R}^{C}$ . + +The text encoder is a Transformer [61] that inputs words tokenized by a BPE tokenizer [15]. It encodes two type of inputs: (a) a narration sentence which describes the contents of a clip; and (b) a noun set which contains noun phrases from the narration sentence. At the output, the embedding corresponding to the EOS token is taken to be the embedding for the full sentence $t \in \mathbb{R}^C$ , and multiple noun embeddings $n \in \mathbb{R}^C$ . + +Object-Aware Module. The object-aware module is a cross-attention Transformer which has a permutation-invariant set of learnable vectors as queries (similar to DETR [7] and Mask2former [9]). The queries are at video-level, shared between frames. They consist of three sets: two hand queries $Q_{h} = (q_{h1}, q_{h2})$ for the left and right hands; $K$ object queries $Q_{o} = (q_{o1}, q_{o2}, \ldots, q_{ok})$ ; and a video-level query $q_{v}$ . These queries are learned, and attend to the visual feature map $F'$ from the visual backbone and output a set of summary vectors $S = (s_{h1}, s_{h2}; s_{o1}, \ldots, s_{ok})$ corresponding to each input query. + +The object-aware module operates on the visual content without any interaction with the text information. It consists of six cross-attention blocks. As in a traditional Transformer decoder [61], in each block, there is a multi-head self-attention layer and a multi-head cross-attention layer. The self-attention layer enables interactions between hand, object and video queries, and the cross-attention layer allows the query to extract object-oriented information from the visual content. + +Bounding Box Head. The hand query vectors $Q_{h}$ and object query vectors $Q_{o}$ are trained to predict the bounding box of the hands and objects respectively in each frame. + +Note these query vectors and summary vectors are at the video level; to predict boxes at frame level, we condition a summary vector $s_j$ of object $j$ on a learnable frame index vector $x_i$ by concatenation of $s_j$ and $x_i$ , and use a multi-layer perceptron $F_{box}$ to project them onto a bounding box $\hat{b}_{j,1}$ , where $i$ is the frame number: + +$$ +\hat {b} _ {j i} = F _ {\text {b o x}} \left(s _ {j}; x _ {i}\right) \tag {1} +$$ + +As a result, we will have a time series of bounding boxes $(\hat{b}_{j,1},\hat{b}_{j,2},\ldots \hat{b}_{j,T})$ from each $j^{\mathrm{th}}$ summary vector. + +Semantic Head. We assign semantic meanings to object summary vectors, standing for the object name/class. To achieve this, we project $s_j$ onto a word embedding $\hat{n}$ with a multi-layer perception $F_{semantic}$ : + +$$ +\hat {n} = F _ {\text {s e m a n t i c}} \left(s _ {j}\right) \tag {2} +$$ + +# 3.3. Training Objectives + +Vision-Text Matching. We follow EgoVLP [35] and use EgoNCE loss as the objective for matching between video-level embedding $v$ and sentence-level embedding $t$ of the narration. In one batch $\mathcal{B}$ , the positive sample set $\mathcal{P}_m$ is made up of a sample $i$ and other samples that share at least one noun and one verb with it: $\mathcal{P}_m = \{n \in \mathcal{B} \mid \text{noun}(n) \cap \text{noun}(m) \neq \emptyset, \text{verb}(n) \cap \text{verb}(m) \neq \emptyset\}$ . And for each sample $i$ , there is a hard negative sample $i'$ sampled from the same video. Hence, the samples in the original batch $\mathcal{B}$ and their hard negative counterparts together form the new batch $\widetilde{\mathcal{B}}$ . + +The objective matching video-to-text (v2t) for a video embedding $v$ is formulated as below; in practice the symmetric text-to-video (t2v) matching objective is also used (omitted for brevity): + +$$ +\mathcal {L} _ {\mathrm {v 2 t}} ^ {\mathrm {e g o}} = \frac {1}{| \mathcal {P} _ {m} |} \sum_ {k \in \mathcal {P} _ {m}} \log \frac {\exp \left(v ^ {T} t _ {k} / \tau\right)}{\sum_ {n \in \hat {\mathcal {B}}} \left(\exp \left(v ^ {T} t _ {n} / \tau\right) + \exp \left(v ^ {T} t _ {n ^ {\prime}} / \tau\right)\right)}. \tag {3} +$$ + +Bounding Box Prediction. We use the 100DOH off-the-shelf hand object detector [53] to produce bounding boxes of two classes on each frame as supervision: hand and objects that are in contact with hands. + +There are two challenges in using the detections for training supervision: 1) the image detector acts at the image level independently and does not provide box-ID association over different frames in a clip; 2) many hands and objects are missed due to motion blur and domain gap in ego-centric videos. Therefore, we apply Hungarian matching between predicted boxes $\hat{b}$ and ground-truth boxes $b_{i}$ on single frames independently, so that for each $b_{i}$ , we find the matched prediction $\hat{b}_{\sigma(i)}$ to minimize the global matching + +cost. The final loss on bounding boxes is computed as the sum of the $\ell_1$ loss and the Generalized IoU loss $L_{iou}$ [51] on paired boxes: + +$$ +\mathcal {L} _ {\mathrm {b o x}} \left(b _ {i}, \hat {b} _ {\sigma (i)}\right) = \mathcal {L} _ {\mathrm {i o u}} \left(b _ {i}, \hat {b} _ {\sigma (i)}\right) + \left| \left| b _ {i} - \hat {b} _ {\sigma (i)} \right| \right| _ {1} \tag {4} +$$ + +and, to tackle the problem of missing objects, we do not penalize boxes that are not matched to nouns unlike traditional detection tasks. + +Object Class Prediction. We have noun embeddings from the video description and a set of predicted object name embeddings from the summary vectors, the task is to find the correspondence between them so that we can use the ground-truth nouns from the description as supervision for the predicted object names. As shown in figure 2, we align the nouns in the narrations and the names of objectboxes in two steps: + +(1) Object-noun alignment: We score the predicted object name embeddings $\hat{n}$ against the ground-truth noun embeddings $n$ to construct a similarity matrix $C\in \mathbb{R}^{K\times N}$ , where $K$ is the number of object queries and $N$ is the number of noun phrases in the description as following: + +$$ +C (n, \hat {n}) = \frac {n \cdot \hat {n}}{\left\| n \right\| \left\| \hat {n} \right\|} \tag {5} +$$ + +Cost matrix $-C$ is used in Hungarian matching to select the matched summary vector for each noun phrase. + +(2) Word-level contrastive training: We apply InfoNCE loss on the matched embeddings $\hat{n}_j$ and $n_j$ against the embeddings of all the nouns $n_k'$ in Ego4D taxonomy dictionary $\mathcal{D}$ [16]: + +$$ +\mathcal {L} _ {\text {w o r d}} = - \frac {1}{N} \sum_ {j = 1} ^ {N} \log \frac {\exp \left(\hat {n} _ {j} ^ {T} n _ {j} / \tau\right)}{\sum_ {k \in \mathcal {D}} \exp \left(\hat {n} _ {j} ^ {T} n _ {k} ^ {\prime} / \tau\right)} \tag {6} +$$ + +Training Loss. The total training objective is the sum of the vision-text matching loss and the auxiliary losses on object vectors: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {v} 2 \mathrm {t}} ^ {\mathrm {e g o}} + \mathcal {L} _ {\mathrm {t} 2 \mathrm {v}} ^ {\mathrm {e g o}} + \mathcal {L} _ {\text {b o x}} + \lambda_ {w o r d} \mathcal {L} _ {\text {w o r d}} \tag {7} +$$ + +# 3.4. Inference + +Once trained, the model acts as a standard ego-centric vision-language video model which operates just on video frames and text descriptions, without requiring further access to object boxes or detectors. However, if desired, hand and object box detections and names can be read out for each frame at inference using the summary vectors, which can be used for grounding the input text description. + +![](images/a37305366833cd3e3ec43a1e04f3b00b373ea24146ec6087df1412c67bb92568.jpg) +Step 1. Object-noun alignment +Figure 2. Training the model to predict object classes. Left: Object-noun alignment. First, the nouns in the video descriptions are matched against the predicted classes using Hungarian matching, to choose the most matched summary vectors. Right: Word-level contrastive training. Next, we supervise the matched predicted classes using a contrastive objective [46] against all the nouns in Ego4D taxonomy, to have similar embeddings as the corresponding nouns. + +![](images/26b27a49804498b929fcac44da66645bec9aaff1c4e7d902d206e6206c9ada7a.jpg) +Step 2. Word-level contrastive training + +# 4. Implementation + +In section 4.1, details of extracting the hand and object detections from 100DOH pre-trained detector are summarized, followed by the architectural details in section 4.2. Finally, in section 4.3 the training pipeline, model and input specification, and optimization details are provided. More details can be found in the supplementary materials. + +# 4.1. Weak Supervision from Pre-trained Detector + +We uniformly sample four frame from each clip in the EgoClip dataset [35] as the input to the 100DOH hand-object detector [53]. The short side of the frame is resized to 640 pixels. There are 16 million frames in total, and the average number of boxes detected per frame is 1.96 for hand and 1.67 for object. Among all the frames, about $15.8\%$ frames have no hands detected and $17.9\%$ frames have no object detected. The average size of hand boxes is $2.8\%$ of the frame size, while the average size of object boxes is $19.4\%$ of the frame size. We use the top 2 hands and top 4 objects detected in the scene as supervision. + +# 4.2. Architecture + +The object-aware module is a 6-layer cross-attention transformer with 8 attention-heads in each layer. The hidden dimension in cross-attention layers is 768, and the video embedding, object embeddings and text embeddings are projected to 256 dimensions before computing the cosine similarity score. We set the number of hand queries to 2, number of object queries to 12, which is designed to be larger than the maximum number of objects in the supervision. We use TimeSformer-L (TSF-L) [6] as the visual encoder, and a 12-layer Transformer [61] from LaViLa [77] as the text encoder. + +![](images/f70fec999aae7d610a0c55560ac1c635ceffd69a1a111a2023e082d8023a54b0.jpg) +Text: #C C cuts the box with the utility knife + +![](images/b019a8585058e82ca29053cdb795cea348c4da7a6f2d1f3cf6b68702814a1c6d.jpg) +a) Object boxes predicted by off-the-shelf hand-object detector +b) Vision-text grounding results from the proposed model +Figure 3. Visualization of text grounding on EgoClip. We show the comparison between detections from off-the-shelf 100DOH hand-object detector [53] (used for training supervision) and the predicted boxes from our model respectively. (a) The detections are noisy, objects are missed, and there is no temporal association of the detected boxes across frames. (b) The trained model learns temporally consistent tracks as well as object categorization using only noisy frame-level box-supervision and weak supervision from the texts. + +![](images/5bb4f5d0621e1eba35d67fa22d0497d7930bc32200b122b72e940eeb1ef3ce48.jpg) +C C puts the spatula in his right hand on the chopping board + +![](images/e8e461a3f1036fae36ea4fb708f64b16edaeab28073c05e5c29968cfbc0cf4de.jpg) +a) Object boxes predicted by off-the-shelf hand-object detector +b) Vision-text grounding results from the proposed model + +# 4.3. Training Details + +In training, we uniformly sample $4 \times 224 \times 224$ RGB frames from each clip. Hand and object boxes are preextracted from these 4 frames using off-the-shelf detector and used as supervision. We keep the visual and text encoder frozen in training. Only the object-aware decoder, query vectors, and the MLP projection layer on text embedding parameters are learned during training. We train the model for 5 epochs on one A6000 GPU, with batch size 128. We use AdamW [39] as the optimizer and set the learning rate to $3e - 5$ . $\lambda_{word}$ is set to 0.5 to balance the scale of the four losses. + +# 5. Experiments + +Section 5.1 introduces all the datasets we used for training and evaluation, followed by the evaluation protocols in section 5.2. Finally, we discuss the ablation studies (section 5.3) and compare to prior SOTA methods on different benchmarks (section 5.4). + +# 5.1. Datasets + +Ego4D/EgoNLQ/EgoMQ/EgoClip/EgoMCQ [16, 35]. Ego4D is a massive-scale dataset focusing on ego-centric videos. It contains 3670 hours video for many different tasks, including action anticipation, AV diarization, etc. EgoNLQ and EgoMQ is a subset for natural language queries and moment query, designed for testing the models' episodic memory and long-term video understanding. [35] proposes a new subset EgoClip for vision-language pre-training, comprising 3.8M clip-text pairs. They also introduce EgoMCQ (i.e., Egocentric Multiple-Choices-Question) as a downstream evaluation dataset for the pre-training. Given a text query, the model tasked to + +choose the paired video clip from 5 candidates. The evaluation metrics is 'intra-video' and 'inter-video' accuracy, depending on where the candidates are chosen from. + +Epic-Kitchens-MIR [11]. Epic-Kitchens is a large-scale ego-centric dataset with 100-hour activities recorded in kitchens. Epic-Kitchens-MIR is a subset with about 9881 clip-text pairs for vision-language retrieval. It is designed for multi-instance retrieval. The model is evaluated on retrieving the paired text/video given a query text/video. The evaluation metrics are mean average Precision (mAP) and normalized Discounted Cumulative Gain (nDCG). + +VISOR [12] is a benchmark built on Epic-Kitchens for segmenting hands and active objects in egocentric video. It has pixel-level annotations covering 36 hours of untrimmed video and 257 object classes. We utilize this annotations in its val split for evaluations on vision-text grounding. + +EGTEA [34] contains 28 hours of cooking activities from 86 unique sessions of 32 subjects. Fine-grained actions are classified into 106 classes. We retrieve the text-descriptions of action classes to evaluate the model for action classification on the test set of all three splits. We measure the performance using mean-class accuracy and top1 accuracy. + +# 5.2. Evaluation Protocol + +We evaluate the performance of our model in the following three aspects: + +Zero-shot transfer. To test the transferability and generalization ability, we conduct zero-shot evaluation on multiple-choice questions (EgoMCQ), multi-instance retrieval (EpicKitchens-MIR), action classification (EGTEA). Among these datasets, videos in EgoMCQ are from the same data + +source as our pre-training dataset Ego4D. Other datasets demonstrate a domain gap, hence, evaluate for transferable representations. + +Episodic memory. To evaluate the richness of the representations learned by our model, we use the video representations to solve Episodic memory tasks in Ego4D. Following [35, 77], we pre-extract the video features from our model first. Using these pre-computed features as input, we train a VSLNet [72] for temporal localization in NLQ, and train a VSGN [76] for moment retrieval in MQ. + +Vision-language grounding. Due to the lack of ground-truth for object-grounding in EgoClip, we evaluate the grounding ability of the model on VISOR instead. VISOR is an egocentric dataset with scenes in the kitchens, where frames are annotated sparsely with segmentation masks and object names. We re-propose the manually annotated segmentation masks in it to extract ground-truth bounding-boxes for hands and in-contact objects. To carry out the zero-shot evaluation, we take the annotated frames in the val split, filter out the not-in-contact objects in each frame, and convert all the segmentation masks to bounding boxes as ground truth. The predicted object boxes are matched with ground-truth object boxes using noun alignment as during training (eq. (5)), while the left/right hands are predicted from the first and the second hand queries respectively. We repeat a single frame 4 times temporally, and resize it to $224 \times 224$ pixels to be consistent with the pre-training resolution. The predicted bounding-boxes are evaluated to be correct if their centers lie inside the ground-truth bounding-boxes. + +# 5.3. Ablations + +Losses. We ablate the combination of losses on three downstream benchmarks in table 1. Results showing that having both $L_{box}$ and $L_{word}$ leads to the best performance. With the same architecture, when training the model using only $L_{Ego}$ without introducing any object-awareness, the performance is $2\%$ lower on EK100-MIR and EGTEA compared to the object-aware one. + +
LossesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
LEgo93.761.835.936.644.937.6
LEgo + Box94.262.736.737.445.338.5
LEgo + Word93.761.936.737.445.838.1
LEgo + Box + Word94.563.037.537.846.639.1
+ +Table 1. Ablation on training objectives for zero-shot transfer tasks. Introducing the object-awareness by having box and word supervision helps the model to achieve better transfer results on EK100-MIR and EGTEA. + +
Detector input resEgoMCQEKEGTEAVISOR
InterIntraAvg mAPAvg nDCGTop1MeanLoc Acc
256p94.263.235.734.642.036.0-
256p94.563.036.937.044.338.968.2
640p94.563.037.537.839.146.678.7
+ +Table 2. Ablation on box quality for zero-shot transfer tasks. We extract boxes using $256\mathrm{p}$ and $640\mathrm{p}$ images as input to the detector respectively, resulting in boxes of different qualities as supervision in training. + +
# Obj QueriesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
494.162.837.837.645.938.1
894.562.737.738.045.537.9
1294.563.037.537.846.639.1
+ +Table 3. Ablation on the number of object queries for zero-shot transfer tasks. The number of queries do not have a big impact on EK100. Larger number of queries shows a boost on mean-class accuracy on EGTEA, and smaller number of queries is better on intra-video accuracy. + +Quality of detected boxes. The extent to which a model can acquire object-level information is constrained by the quality of the bounding boxes from the off-the-shelf detector. To investigate how much the quality of boxes affects our training, we detect hand and object boxes on EgoClip training set using 100DOH [53] with 256p and 640p images as input – with larger image size, the objects should be more precisely delineated. We use the two sets of boxes as supervision in our training and show results in table 2. Even when training with noisy boxes from 256p, our model is better than the previous SOTA model. When boxes from 640p are used, the averaged zero-shot transfer performance is further improved by $1\%$ on Epic-Kitchens and EGTEA, showing that our method can bring larger improvement over the non-object-aware method when given better boxes. Furthermore, a significant boost on the grounding results is also observed on VISOR when using boxes with better quality. + +Number of object queries. We use different number of query vectors in the object-aware decoder to see its impact on both vision-language tasks and the grounding task. We show zero-shot transfer results in table 3 and vision-text grounding results in table 7. The number of queries has relatively small impact on vision-language representation learning, the performance gaps on zero-shot transfer tasks are mostly small than $1\%$ . However, a smaller number of object queries leads to much better results for in-contact object grounding on VISOR, 4 queries is better than 12 objects by $3.6\%$ in localization accuracy. The reason is that too + +
MethodBackboneBatch sizeObject awareHard negEgoMCQEK100-MIREGTEA
Inter-videoIntra-videoV-TmAP T-VAvgV-TnDCG T-VAvgTop1-AccMean-Acc
EgoVLPTSF-B512X✓90.657.226.020.623.328.827.027.917.6-
LaViLaTSF-B1024XX93.859.935.126.630.933.730.432.0-28.9
LaViLaTSF-L1024XX94.563.140.032.236.136.133.234.640.134.1
LaViLa*TSF-L1024XX94.263.239.731.735.736.133.234.642.036.0
Ours*†TSF-L128XX93.760.539.730.335.037.334.535.944.836.3
Ours*‡TSF-L128X✓93.761.840.731.135.938.335.036.644.937.6
Ours*TSF-L128✓✓94.563.042.332.737.539.336.237.846.639.1
+ +Table 4. Comparison to SOTA results on zero-shot transfer to EgoMCQ, EK100-MIR and EGTEA. We compared to EgoVLP and LaVILA, two previous SOTA models pre-trained on EgoClip. Our object-aware model has achieved comparable results on multiple-choice questions on EgoMCQ, and SOTA results on multi-instance retrieval task on EpicKicthens and action classification on EGTEA. Model without * use center cropping in evaluation, while * denotes the usage of resizing instead of cropping. Ours† and Ours‡ stands for different variants of our model depending on whether object-aware losses and hard negative sampling is used in training. + +
MethodPredicted BoxesEgoMCQEKEGTEA
InterIntraAvg mAPAvg nDCGMeanTop1
Oursobj94.062.136.837.145.137.6
hand+obj94.563.037.537.846.639.1
+ +many queries result in a large number of predicted boxes, which increases the probability of mis-matching. + +Impact of hand boxes. We ablate the impact of having hand boxes as supervision in our training. Results are shown in table 5, training the model to predict hand as well as objects help the model to get about $1\%$ higher zero-shot transfer performance on EK and EGTEA. + +# 5.4. Comparison to the SOTA + +Zero-shot transfer. We compare to previous SOTA in table 4. Our model is comparable on EgoMCQ and better on EK-MIR and EGTEA, showing its good zero-shot transferability. Due to limited compute resources, we are not able to unfreeze the visual backbone to train end-to-end or increase the batch size further. Despite these disadvantages, our method outperforms the previous SOTA on two tasks for models that have been trained end-to-end. + +The main difference between LaViLa(L) and ours is the object-aware training and hard negative sampling; Ours† in table 4 is a LaViLa(L) model with an extra Transformer decoder, which is trained with only InfoNCE loss on video and sentence embeddings. Without object-awareness and hard sampling, it gets better accuracy on EGTEA and better nDCG on EK100 due to more parameters added, but falls behind on EgoMCQ and mAP on EK100. Applying hard negative sampling (Ours‡) and object-aware training (Ours) brings improvement across the board. The most obvious boost comes from inducing object-awareness, bringing $1.5\%$ improvement on average. And the results could be further improved by obtaining better pseudo-boxes, as the + +magnitude of boost from learning objects is closely related to the box quality (as shown in table 2). + +Episodic memory. Results on EgoNLQ and EgoMQ are shown in table 6. The two tasks test the video understanding on several-minutes long videos. In these experiments, trained video and text backbones are used as feature extractor, and extra modules are trained on the long feature sequences for natural language querying and memory querying. Therefore, the richer the information encoded in the features, the better the results will be. We list the models trained on the same amount of video and text data in black, results show that our model are better than the previous SOTA on all the metrics in the two tasks. This is because features from the object-aware model have captured more object information, thus enabling better precision and recall on localization and retrieval. InternVideo [63] and NaQ [50] are trained on more video or text data, and are included for completeness. Please note that our method is compatible with any visual language backbones, using better backbones might also lead to better results. + +Table 5. The impact of having hand boxes in the box prediction. Having hand boxes in the prediction helps on all the zero-shot evaluation benchmarks. + +
MethodBatch sizeEgoNLQEgoMQ
mIOU@0.3mIOU@0.5R1@0.5R5@0.5mAP
R1R5R1R5
SlowFast-5.510.73.16.625.246.26.0
EgoVLP51210.818.86.813.530.152.011.4
LaViLa(B)102410.519.16.713.6---
LaViLa(L)102412.122.47.315.432.556.113.4
Ours12813.223.37.915.633.456.716.0
VideoIntern [63]14k16.523.010.116.1--23.6
ReLER + NaQ [50]204819.323.611.615.5---
+ +Table 6. Comparison to SOTA results of fine-tuning on EgoNLQ and EgoMQ. Our object-aware model encodes richer information in the visual representations, hence obtaining better results on all the metrics in NLQ and MQ task in Ego4D episodic memory benchmark. We list other SOTA models (in grey) trained with more video and text data for completeness. + +
ModelObject Assignment# queriesLoc Accuracy
Detector [53]Random-37.1
GT matching-41.3
OursPredicted482.3
881.2
1278.7
+ +Table 7. In-contact object localization accuracy on VISOR. Our model does better on in-contact object localization after weakly-supervised training compared to the baseline (the Detector with Random or GT object assignment), which is the source of supervision in our pre-training. The improvement is due to increased recall of our model over the baseline. + +# 5.5. Evaluating Object Grounding + +Qualitative results on EgoClip. In fig. 3 we show the grounding results on EgoClip after training as compared to the supervision from the 100DOH detector. After training, the predictions from our model find the missing objects because we do not penalize extra bounding boxes predicted, but select the active object/object of interest through matching noun embeddings. It also learns temporal association of object bounding-boxes as the result of using the same object summary vector to predict the boxes over all the frames (as in eq. (1)); the summary vector attends to visually similar features corresponding to the same object across the clip without any explicit supervision for temporal consistency. + +Quantitative and qualitative results on VISOR. In table 7 we show the text-grounding results on VISOR. We take the predictions from the hand-object detector [53] as our baseline. Since the predictions only detect hands and objects without an object class name, we associate the predicted boxes with ground-truth boxes in VISOR in two ways: (a) random: we assign predicted object boxes to GT object boxes randomly, (b) GT matching: we use Hungarian matching to find the predicted boxes with highest IoU against the GT object boxes. However, even when matched using GT information, the baseline detector does not achieve a high accuracy due to poor recall. Results show grounding ability of our model is $40\%$ better than the baseline using only weak supervision from the video descriptions. We also show the qualitative results on detecting different number of hands and objects in fig. 4, our model has a much higher recall when compared to baseline detector when operating at the same resolution. + +# 6. Conclusion + +In this paper, we introduce a method to learn object-aware ego-centric video representations using noisy supervision from pre-trained hand-object detectors. The object-representations so learned show strong zero-shot transfer across various downstream tasks and datasets, mirroring + +![](images/fdd00c87270569799d09f157cbd51d0748e5acba85626628a894d61e917a5b39.jpg) +Figure 4. Grounding visualization on the Epic-Kitchen VISOR val split. When operating at the same resolution, our model shows better grounding ability on hands and objects compared to the baseline 100DOH detector [53] (with GT matching) used for training. Note that the low IoU on hands on the third column is a result of the GT 'hand' segmentation mask covering the arm by definition, while the detector and our model are trained to localize only palm and fingers. + +the performance improvement from object-aware training on images [70]. The model uses standard neural modules (i.e., transformers), and does not require any object boxes or detectors as input during inference, making it widely applicable as a drop-in replacement for training video-language models. Even though the model is trained with sparse and noisy object supervision at the frame-level (without temporal associations), during inference dense temporal bounding-box tracks and category predictions can be obtained, which are superior to the predictions from the base hand-object detector used for training. There are several avenues for improvement. Our model uses the pre-trained video encoder operates at a small resolution $224 \times 224$ which makes detecting small objects difficult. Further, four frames are sampled uniformly from the clip regardless of its length which can cause difficulties due to temporal aliasing. Nevertheless, we hope our work will inspire further research in learning transferable object-aware representations for videos. + +Acknowledgements. This research is funded by a Google-DeepMind Graduate Scholarship, a Royal Society Research Professorship, and EPSRC Programme Grant VisualAI EP/T028572/1. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2 +[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In Proc. ICCV, 2021. 1 +[3] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In NeurIPS, 2022. 2, 3 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proc. ICCV, 2021. 1, 2, 3 +[5] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 2 +[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In Proc. ICML, 2021. 4, 5 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 1, 3, 4 +[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proc. ECCV, 2020. 2 +[9] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 4 +[10] Noam Chomsky and David W Lightfoot. Syntactic structures. Walter de Gruyter, 1957. 1 +[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proc. ECCV, 2018. 1, 2, 6 +[12] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Datasets and Benchmarks Track, 2022. 2, 6 +[13] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 2 +[14] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurlPS, 2022. 2 + +[15] Philip Gage. A new algorithm for data compression. C Users Journal, 1994. 4 +[16] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proc. CVPR, 2022. 1, 5, 6 +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. ICCV, 2017. 1 +[18] Paul Henderson and Christoph H. Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2 +[19] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. arXiv preprint arXiv:2110.06915, 2021. 1, 2 +[20] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proc. ICCV, 2019. 2 +[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 1, 2 +[22] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proc. CVPR, 2016. 2 +[23] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. ICCV, 2021. 2 +[24] Ranjay Krishna, Ines Chami, Michael Bernstein, and Li Fei-Fei. Referring relationships. In Proc. CVPR, 2018. 2 +[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2 +[26] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In Proc. ICCV, 2021. 2 +[27] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012. 2 +[28] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proc. CVPR, 2022. 2 +[29] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 2017. 1 +[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. ICML, 2022. 2 + +[31] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2 +[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proc. ICCV, 2021. 2 +[33] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. ECCV, 2020. 2 +[34] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proc. ECCV, 2018. 6 +[35] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 2, 3, 4, 5, 6, 7 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, 2014. 2 +[37] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 704-721. Springer, 2020. 2 +[38] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2 +[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. ICLR, 2019. 6 +[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 2 +[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3 +[42] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proc. CVPR, 2020. 1 +[43] Effrosyni Mavroudi and René Vidal. Weakly-supervised generation and grounding of visual descriptions with conditional generative models. In Proc. CVPR, 2022. 2 +[44] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proc. ICCV, 2019. 2 +[45] Marvin Minsky. Society of mind. Simon and Schuster, 1988. 1 +[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre + +sensation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 5 +[47] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE conference on computer vision and pattern recognition, pages 2847-2854. IEEE, 2012. 2 +[48] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuytelaars. Revisiting spatio-temporal layouts for compositional action recognition. In Proc. BMVC, 2021. 1 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 1, 2, 3 +[50] Santhosh Kumar Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Naq: Leveraging narrations as queries to supervise episodic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6694-6703, 2023. 8 +[51] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. CVPR, 2019. 5 +[52] Kate Saenko, Ben Packer, C Chen, Sunil Bandla, Y Lee, Yangqing Jia, J Niebles, Daphne Koller, Li Fei-Fei, Kristen Grauman, et al. Mid-level features improve recognition of interactive activities. Technical report, Dept. of Elec. Engg. and Computer Sc., University of California, Berkeley, 2012. 1 +[53] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proc. CVPR, 2020. 1, 2, 4, 5, 6, 7, 9 +[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Proc. ICLR, 2015. 2 +[55] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurIPS, 2022. 2 +[56] Elizabeth S Spelke, Karen Breinlinger, Janet Macomber, and Kristen Jacobson. Origins of knowledge. Psychological review, 99, 1992. 1 +[57] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In Proc. ECCV, 2018. 1 +[58] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2 +[59] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. *NeurIPS*, 2021. 2 +[60] Joshua B Tenenbaum, Charles Kemp, Thomas L Griffiths, and Noah D Goodman. How to grow a mind: Statistics, structure, and abstraction. Science, 2011. 1 +[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 4, 5 +[62] Xiaolong Wang and Abhinav Gupta. Videos as space-time + +region graphs. In Proc. ECCV, 2018. 1, 2 +[63] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 8 +[64] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 3 +[65] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2 +[66] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In Proc. ICCV, 2021. 2 +[67] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proc. ICCV, 2019. 2 +[68] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2 +[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2 +[70] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 9 +[71] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proc. ACCV, 2022. 1, 2, 3 +[72] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE PAMI, 2021. 7 +[73] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2 +[74] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In European Conference on Computer Vision, pages 127-145. Springer, 2022. 2 +[75] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. CVPR, 2021. 2 +[76] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 7 +[77] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language + +models.arXiv preprint arXiv:2212.04501,2022.2,5,7 +[78] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proc. CVPR, 2022. 2 +[79] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. arXiv preprint arXiv:1805.02834, 2018. 2 +[80] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 2 +[81] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In Proc. ECCV, 2022. 3 \ No newline at end of file diff --git a/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/images.zip b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..19b6ec08e194619642dc4cc613562795676524a8 --- /dev/null +++ b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b8c31c393c0313b0c14226344d284d8d0f7beb526d7956f9f39d25c2c6f2be4 +size 499405 diff --git a/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/layout.json b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25a8c68c3d9043eca551947fe388886aa5ec2952 --- /dev/null +++ b/2023/Helping Hands_ An Object-Aware Ego-Centric Video Recognition Model/layout.json @@ -0,0 +1,10093 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 521, + 121 + ], + "type": "text", + "content": "Helping Hands: An Object-Aware Ego-Centric Video Recognition Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 144, + 205, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 144, + 205, + 185 + ], + "spans": [ + { + "bbox": [ + 70, + 144, + 205, + 185 + ], + "type": "text", + "content": "Chuhan Zhang \nVGG, University of Oxford \nczhang@robots.ox.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 144, + 364, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 144, + 364, + 185 + ], + "spans": [ + { + "bbox": [ + 228, + 144, + 364, + 185 + ], + "type": "text", + "content": "Ankush Gupta Google DeepMind, London ankushgupta@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 388, + 144, + 523, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 144, + 523, + 185 + ], + "spans": [ + { + "bbox": [ + 388, + 144, + 523, + 185 + ], + "type": "text", + "content": "Andrew Zisserman \nVGG, University of Oxford \naz@robots.ox.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 191, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 227 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 227 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 288, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 348 + ], + "type": "text", + "content": "We introduce an object-aware decoder for improving the performance of spatio-temporal representations on egocentric videos. The key idea is to enhance object-awareness during training by tasking the model to predict hand positions, object positions, and the semantic label of the objects using paired captions when available. At inference time the model only requires RGB frames as inputs, and is able to track and ground objects (although it has not been trained explicitly for this)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 348, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 505 + ], + "type": "text", + "content": "We demonstrate the performance of the object-aware representations learnt by our model, by: (i) evaluating it for strong transfer, i.e. through zero-shot testing, on a number of downstream video-text retrieval and classification benchmarks; and (ii) by using the representations learned as input for long-term video understanding tasks (e.g. Episodic Memory in Ego4D). In all cases the performance improves over the state of the art—even compared to networks trained with far larger batch sizes. We also show that by using noisy image-level detection as pseudo-labels in training, the model learns to provide better bounding boxes using video consistency, as well as grounding the words in the associated text descriptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 506, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 287, + 543 + ], + "type": "text", + "content": "Overall, we show that the model can act as a drop-in replacement for an ego-centric video model to improve performance through visual-text grounding1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 587, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 684 + ], + "type": "text", + "content": "In visual-language models there has been a recent move to explicitly build object awareness into the vision module by adding specialized and bespoke components, or using entirely object-centric architectures. The motivation for this partly comes from the attractive compositional nature of objects and their inter-relationships in language, which enables inexhaustible novel combinations [10, 45], and partly from infant cognitive studies that stress the im" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 214, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 545, + 335 + ], + "type": "text", + "content": "portance of objects in early visual development [29, 56, 60]. Examples in the video domain include explicit internal object representations [2], e.g., through RoI-align [17] pooled features either from a pre-trained region-proposal network (RPN) [2, 52, 57, 62], or from bounding-box coordinates taken as input [19, 42, 48, 71]. This contrasts with the large body of work where standard representations are learnt end-to-end without any explicit factorization into objects/entities, such as dual-encoder vision-language models in the image [21, 49] and video domains [4, 64]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 336, + 546, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 546, + 397 + ], + "type": "text", + "content": "In this paper, we take a different (middle) path and instead use a vanilla video transformer architecture and induce object-awareness into the video representation by tasking the model to predict object-level properties, such as their localization and semantic categories, only during training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 398, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 546, + 532 + ], + "type": "text", + "content": "Our target domain is ego-centric video [11, 16], and we tailor the object properties used to this. In ego-centric videos the actor [57] is often present through their hands, and we therefore task the network to predict both the hands and the principal objects they interact with. As will be seen, this simple object-aware training boosts the performance of pre-trained video-language architectures significantly, and leads to state-of-art performance across multiple ego-centric benchmarks. During inference, the model requires only RGB frames as input, and operates as a standard video-language network." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 532, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 713 + ], + "type": "text", + "content": "In more detail, our model is built on top of a pre-trained video-language dual encoder architecture (where there are separate encoders for the video and text data). We add an additional, but vanilla, transformer decoder head [61], and train with DETR/Mask2former [7, 9] query vectors and object loss for hands and other objects. The intuition is that these additional query vectors help the model to attend to and track the hands and salient objects in the scene (these are the 'helping hands' of the paper title). Importantly, we do not require dense frame level ground truth for this training. Rather, we obtain somewhat noisy and temporally sparse annotations automatically from a hand object detector [53], and use these to provide prediction targets for the frames where they are available. This opportunistic use of annotations is pragmatic as object detectors trained on third" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 286, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 46, + 693, + 286, + 713 + ], + "type": "text", + "content": "Code and models available at: https://github.com/Chuhanxx/helping_hand_for_egocentric Videos" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13901" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "person datasets (such as COCO) do not perform so well on the ego-centric domain, where the scenes are more crowded and objects are often small and can be motion blurred. By only requiring annotations for a subset of frames, where they can be reliably produced automatically, we are able to train on large-scale data without requiring expensive manual supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 287, + 242 + ], + "type": "text", + "content": "Although we train with noisy and sparse image-level object localization, our model can learn to predict better and denser bounding-box trajectories through large-scale training due to the spatio-temporal consistency which naturally presents in videos. Also, it is able to predict semantic grounding by learning to map the object appearance to the nouns in the video captions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 242, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 315 + ], + "type": "text", + "content": "It is worth noting that we are using hand detectors because hands are a common and important object in ego-centric videos. However, the object-centric method we are proposing has greater scope than ego-centric videos and can be applied to other scenarios with other object types providing the 'helping-hand'." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 582 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 46, + 316, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 387 + ], + "type": "text", + "content": "In summary, we make the following contributions: (i) We propose a method to induce object-awareness in video-language models for an architecture composed of standard neural modules. The model only requires RGB frames as inputs, and thus is a drop-in replacement for any ego-centric video model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "type": "text", + "content": "(ii) The model can be trained opportunistically using available and sparse frame-level and noisy annotations, produced automatically." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 473 + ], + "type": "text", + "content": "(iii) We demonstrate state-of-the-art strong (zero-shot) transfer for cross-modal retrieval to other ego-centric datasets namely, EpicKitchens-MIR and EGTEA improving prior art by " + }, + { + "bbox": [ + 46, + 426, + 287, + 473 + ], + "type": "inline_equation", + "content": "2 - 4\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 475, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 522 + ], + "type": "text", + "content": "(iv) We evaluate the grounding quantitatively using the EpicKitchens-VISOR dataset [11, 12] and find that the model outperforms the base hand-object detector used for training supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 523, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 582 + ], + "type": "text", + "content": "(v) Finally, we also demonstrate that the representations learned can be used as input in long-term video understanding tasks like EgoNLQ and EgoMQ. The objectiveness in the representation helps the model outperform other models trained on the same training set on these two tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 596, + 134, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 134, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 134, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 616, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 713 + ], + "type": "text", + "content": "Vision and Language Representation Learning. Different from transferring representations learned for classification on a fixed set of object categories [27, 54], recent vision-language pre-training (VLP) works leverage largescale supervision from free-from text descriptions of images and videos. These methods use image captions [49] or video sub-titles [44, 64] with either independent dual encoders for the visual and text modalities [4, 21, 41], or via" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 289 + ], + "type": "text", + "content": "joint encoders with cross-attention across modalities [1, 30, 31, 69]. We also use dual-encoders which are kept frozen due to compute limitations. To explicitly build-in object-awareness into image representations, object-level features extracted from pre-trained object-detectors are aligned with the text descriptions [8, 33, 40, 58, 75, 78] The object-level text alignment is further augmented with the object-box prediction task for grounding in [13, 23, 70, 73]. VLP for ego-centric videos has recently been explored [35, 77] to bridge the domain gap between representations learned from third-person videos found commonly [44], and first-person ego-centric videos. We further extend image based object-aware VLP methods to ego-centric videos by training to predict the auto-generated hand-object boxes extracted from pretrained detectors [53], while requiring only RGB input during inference, making our model a drop-in replacement for ego-centric VLP models albeit with enhanced object-awareness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "type": "text", + "content": "Weakly Supervised Video Text Grounding. A key challenge for grounding in videos is the lack of large-scale object-level annotations for videos. While such annotations are readily available for synthetic datasets [68], expensive manual annotation is required for real videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "text", + "content": "Hence, weakly supervised method have been developed which leverage the video sub-titles/description to map nouns/verbs to regions in frames. This is typically achieved by first extracting bounding-box/segmentation regions from pre-trained detectors for objects and humans, and aligning them with keywords using max-margin [79] or constrastive [32] objectives. Similarly, [43] also align words in the video captions to regions from pre-trained RPN by modelling the region-word associations as latent variables of a conditional generative variational auto-encoder. More recently, [59] use cross-attention across text and candidate regions, and find (soft-) associations based on attention weights. While our model similarly absorbs object bounding-boxes and category information obtained from pre-trained object and hand detectors during training to induce object-aware representations, these detectors are not required during inference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 593, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 713 + ], + "type": "text", + "content": "Object-Oriented Learning in Videos. Most vision tasks for images involve objects [22, 24, 25, 36]. In videos, there are also a broad range of object-oriented tasks: some works treat learning object-level information as an end task, they design models to predict object bounding-boxes and masks [5, 26, 37, 38, 65, 67, 74]; other works use object knowledge to achieve some other high-level tasks, for instance, object-centric scene generation [28, 66] and decomposition [14, 18, 55], and action recognition [3, 19, 20, 47, 62, 71, 80]. Our method uses object information to as-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13902" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 69, + 517, + 309 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 517, + 309 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 517, + 309 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 517, + 309 + ], + "type": "image", + "image_path": "818085653772481dc7168d9d221b78e4ad5a0db93b701241b2a615536ea8ca64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "text", + "content": "Figure 1. The object-aware vision-language model architecture. The architecture is made up of three parts: A video backbone, a text backbone and a object-aware decoder. The decoder is a cross-attention transformer, it takes the visual feature map " + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "text", + "content": " as keys and values, which are attended by a set of learnable queries. In these permutation-invariant query vectors, the hand and object queries " + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "inline_equation", + "content": "Q_h, Q_o" + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "text", + "content": " are trained to be object-aware and predict the localization and class of hands and objects. The video query " + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "inline_equation", + "content": "q_v" + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "text", + "content": " attends to both the visual feature map through cross-attention layers, and the object feature map through self-attention layers in the Transformer decoder and output a video-level embedding " + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "text", + "content": " to be matched with the text embedding " + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 316, + 548, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 403, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 583 + ], + "type": "text", + "content": "sist vision-language alignment in egocentric videos, so that the model learns grounded video representations that can generalize better. SViT [3] uses object queries shared between images and videos in order to predict hand-object bounding boxes in videos, whilst only requiring image-level supervision during training. However, the object queries are not used for vision-language alignment. Our previous work [71] showed that encoding object-level information in the model helps transfer learning in various video understanding tasks, but the model required GT bounding boxes as input during inference. In contrast, in this work we model does not require this information during training, and we show box prediction and vision-language alignment can be combined and benefit both in-domain and out-of-domain tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 594, + 275, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 275, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 275, + 609 + ], + "type": "text", + "content": "3. Object-Aware Vision-Language Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "text", + "content": "We first describe the task of object-aware vision-language learning and the architecture of our model. These are followed by the training objectives for vision-language matching and weakly-supervised text grounding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 115, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 115, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 115, + 683 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "In vision-language representation learning the training data consists of short video clips (RGB frames) and an" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 403, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 536 + ], + "type": "text", + "content": "associated free-form text string containing words that describe the visual content. Typically, dual encoders are trained on this paired data, with a visual-encoder that ingests the video clip and a language-encoder that ingests the text [4, 35, 41, 49, 64]. The dual encoders are trained with a contrastive objective [46] such that the cosine similarity for matching vision-text pairs is optimized to be higher than the similarity of negative/not-matching pairs. This pretraining objective enables evaluation on downstream vision-language tasks like video-text retrieval and action classification in a zero-shot manner [49]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 540, + 546, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 546, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 546, + 650 + ], + "type": "text", + "content": "Our object-aware model follows the data, training and evaluation pipeline as above, except that the model is also tasked to output object-level information (e.g., bounding boxes and object categories) during training. By tasking the model to predict object bounding boxes and names which can be matched to the nouns in the narration, the model learns grounded and fine-grained correspondence between the modalities. The object-level prediction is used as an auxiliary objective in training but not used at inference time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "In more detail, as shown in figure 2, there are two types of object-level prediction: (a) hand and object bounding boxes; and (b) object names. Since ground truth of boxes and object names are not available, and most traditional detectors [7, 81] fail to identify objects well in ego-centric" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13903" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "text", + "content": "videos, we cannot rely on strong supervision for the predictions. Instead, we generate bounding box targets (for the hands and other objects) using a robust off-the-shelf hand-object detector [53], though these targets will only be available for some of the frames and are noisy. While for object name prediction, we use a weakly supervised method to align the predicted names with nouns in the paired narration (detailed in section 3.3). In both cases, the supervision is opportunistic and only applied when available." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 129, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 129, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 129, + 201 + ], + "type": "text", + "content": "3.2. Architecture" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": "Dual Encoder. We use dual encoders as our visual and text backbone for efficiency. The visual encoder ingests a video clip " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": " of RGB frames " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "F = (f_{1}, f_{2}, \\ldots, f_{T})" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "f_{i} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": " is the number of frames. The clip " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": " is encoded by a Video Transformer [6] which tokenizes the frames by 3D patches to produce a spatially downsampled feature map " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "F' = (f_{1}', f_{2}', \\ldots, f_{T}')" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "f_{i}' \\in \\mathbb{R}^{H' \\times W' \\times C}" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": ". It outputs a visual embedding " + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 46, + 208, + 287, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": "The text encoder is a Transformer [61] that inputs words tokenized by a BPE tokenizer [15]. It encodes two type of inputs: (a) a narration sentence which describes the contents of a clip; and (b) a noun set which contains noun phrases from the narration sentence. At the output, the embedding corresponding to the EOS token is taken to be the embedding for the full sentence " + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": ", and multiple noun embeddings " + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "inline_equation", + "content": "n \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 306, + 287, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "content": "Object-Aware Module. The object-aware module is a cross-attention Transformer which has a permutation-invariant set of learnable vectors as queries (similar to DETR [7] and Mask2former [9]). The queries are at video-level, shared between frames. They consist of three sets: two hand queries " + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "inline_equation", + "content": "Q_{h} = (q_{h1}, q_{h2})" + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "content": " for the left and right hands; " + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "content": " object queries " + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "inline_equation", + "content": "Q_{o} = (q_{o1}, q_{o2}, \\ldots, q_{ok})" + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "content": "; and a video-level query " + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "inline_equation", + "content": "q_{v}" + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "content": ". These queries are learned, and attend to the visual feature map " + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "content": " from the visual backbone and output a set of summary vectors " + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "inline_equation", + "content": "S = (s_{h1}, s_{h2}; s_{o1}, \\ldots, s_{ok})" + }, + { + "bbox": [ + 46, + 418, + 287, + 552 + ], + "type": "text", + "content": " corresponding to each input query." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 552, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 661 + ], + "type": "text", + "content": "The object-aware module operates on the visual content without any interaction with the text information. It consists of six cross-attention blocks. As in a traditional Transformer decoder [61], in each block, there is a multi-head self-attention layer and a multi-head cross-attention layer. The self-attention layer enables interactions between hand, object and video queries, and the cross-attention layer allows the query to extract object-oriented information from the visual content." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Bounding Box Head. The hand query vectors " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "Q_{h}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " and object query vectors " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "Q_{o}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " are trained to predict the bounding box of the hands and objects respectively in each frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "Note these query vectors and summary vectors are at the video level; to predict boxes at frame level, we condition a summary vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " of object " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " on a learnable frame index vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " by concatenation of " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", and use a multi-layer perceptron " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "F_{box}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " to project them onto a bounding box " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\hat{b}_{j,1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " is the frame number:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 386, + 153, + 545, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 153, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 386, + 153, + 545, + 169 + ], + "type": "interline_equation", + "content": "\\hat {b} _ {j i} = F _ {\\text {b o x}} \\left(s _ {j}; x _ {i}\\right) \\tag {1}", + "image_path": "d522e1ca284213af6d2738fd6abc876c2dc07ea853113b0652d65cb5458280d9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 176, + 545, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 545, + 202 + ], + "type": "text", + "content": "As a result, we will have a time series of bounding boxes " + }, + { + "bbox": [ + 306, + 176, + 545, + 202 + ], + "type": "inline_equation", + "content": "(\\hat{b}_{j,1},\\hat{b}_{j,2},\\ldots \\hat{b}_{j,T})" + }, + { + "bbox": [ + 306, + 176, + 545, + 202 + ], + "type": "text", + "content": " from each " + }, + { + "bbox": [ + 306, + 176, + 545, + 202 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 306, + 176, + 545, + 202 + ], + "type": "text", + "content": " summary vector." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "text", + "content": "Semantic Head. We assign semantic meanings to object summary vectors, standing for the object name/class. To achieve this, we project " + }, + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "text", + "content": " onto a word embedding " + }, + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "text", + "content": " with a multi-layer perception " + }, + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "inline_equation", + "content": "F_{semantic}" + }, + { + "bbox": [ + 304, + 215, + 545, + 264 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 386, + 272, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 272, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 386, + 272, + 545, + 285 + ], + "type": "interline_equation", + "content": "\\hat {n} = F _ {\\text {s e m a n t i c}} \\left(s _ {j}\\right) \\tag {2}", + "image_path": "26082346953a542f9babbf93b53e39230d175f9dc66a4b7b2bdf9f2621e3778c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 293, + 421, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 421, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 421, + 306 + ], + "type": "text", + "content": "3.3. Training Objectives" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": "Vision-Text Matching. We follow EgoVLP [35] and use EgoNCE loss as the objective for matching between video-level embedding " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": " and sentence-level embedding " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": " of the narration. In one batch " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": ", the positive sample set " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_m" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": " is made up of a sample " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": " and other samples that share at least one noun and one verb with it: " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_m = \\{n \\in \\mathcal{B} \\mid \\text{noun}(n) \\cap \\text{noun}(m) \\neq \\emptyset, \\text{verb}(n) \\cap \\text{verb}(m) \\neq \\emptyset\\}" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": ". And for each sample " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": ", there is a hard negative sample " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "i'" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": " sampled from the same video. Hence, the samples in the original batch " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": " and their hard negative counterparts together form the new batch " + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 312, + 545, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": "The objective matching video-to-text (v2t) for a video embedding " + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": " is formulated as below; in practice the symmetric text-to-video (t2v) matching objective is also used (omitted for brevity):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 500, + 566, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 566, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 566, + 542 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v 2 t}} ^ {\\mathrm {e g o}} = \\frac {1}{| \\mathcal {P} _ {m} |} \\sum_ {k \\in \\mathcal {P} _ {m}} \\log \\frac {\\exp \\left(v ^ {T} t _ {k} / \\tau\\right)}{\\sum_ {n \\in \\hat {\\mathcal {B}}} \\left(\\exp \\left(v ^ {T} t _ {n} / \\tau\\right) + \\exp \\left(v ^ {T} t _ {n ^ {\\prime}} / \\tau\\right)\\right)}. \\tag {3}", + "image_path": "cdf22fe8d8285061c05a1fb8777031fea5476020f392537fdb03555c4a1fa580.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 604 + ], + "type": "text", + "content": "Bounding Box Prediction. We use the 100DOH off-the-shelf hand object detector [53] to produce bounding boxes of two classes on each frame as supervision: hand and objects that are in contact with hands." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "There are two challenges in using the detections for training supervision: 1) the image detector acts at the image level independently and does not provide box-ID association over different frames in a clip; 2) many hands and objects are missed due to motion blur and domain gap in ego-centric videos. Therefore, we apply Hungarian matching between predicted boxes " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " and ground-truth boxes " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " on single frames independently, so that for each " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": ", we find the matched prediction " + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{b}_{\\sigma(i)}" + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": " to minimize the global matching" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13904" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "cost. The final loss on bounding boxes is computed as the sum of the " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " loss and the Generalized IoU loss " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "L_{iou}" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " [51] on paired boxes:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 118, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 118, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 63, + 118, + 287, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {b o x}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) = \\mathcal {L} _ {\\mathrm {i o u}} \\left(b _ {i}, \\hat {b} _ {\\sigma (i)}\\right) + \\left| \\left| b _ {i} - \\hat {b} _ {\\sigma (i)} \\right| \\right| _ {1} \\tag {4}", + "image_path": "cd01257ed3b177b1b3762bdb109f96a1522fbc8b5d6fb00dfa651533c2daaba2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 143, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 143, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 287, + 178 + ], + "type": "text", + "content": "and, to tackle the problem of missing objects, we do not penalize boxes that are not matched to nouns unlike traditional detection tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 196, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 292 + ], + "type": "text", + "content": "Object Class Prediction. We have noun embeddings from the video description and a set of predicted object name embeddings from the summary vectors, the task is to find the correspondence between them so that we can use the ground-truth nouns from the description as supervision for the predicted object names. As shown in figure 2, we align the nouns in the narrations and the names of objectboxes in two steps:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "text", + "content": "(1) Object-noun alignment: We score the predicted object name embeddings " + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "text", + "content": " against the ground-truth noun embeddings " + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "text", + "content": " to construct a similarity matrix " + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "inline_equation", + "content": "C\\in \\mathbb{R}^{K\\times N}" + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "text", + "content": " is the number of object queries and " + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 293, + 287, + 354 + ], + "type": "text", + "content": " is the number of noun phrases in the description as following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 363, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 363, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 124, + 363, + 287, + 389 + ], + "type": "interline_equation", + "content": "C (n, \\hat {n}) = \\frac {n \\cdot \\hat {n}}{\\left\\| n \\right\\| \\left\\| \\hat {n} \\right\\|} \\tag {5}", + "image_path": "023f2e4254269d344645ad8596cdcd2d487937d66496da879f6b88839f630ab1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 397, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 420 + ], + "type": "text", + "content": "Cost matrix " + }, + { + "bbox": [ + 46, + 397, + 287, + 420 + ], + "type": "inline_equation", + "content": "-C" + }, + { + "bbox": [ + 46, + 397, + 287, + 420 + ], + "type": "text", + "content": " is used in Hungarian matching to select the matched summary vector for each noun phrase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "text", + "content": "(2) Word-level contrastive training: We apply InfoNCE loss on the matched embeddings " + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "inline_equation", + "content": "\\hat{n}_j" + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "inline_equation", + "content": "n_j" + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "text", + "content": " against the embeddings of all the nouns " + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "inline_equation", + "content": "n_k'" + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "text", + "content": " in Ego4D taxonomy dictionary " + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 422, + 287, + 469 + ], + "type": "text", + "content": " [16]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 479, + 287, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 479, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 78, + 479, + 287, + 513 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {w o r d}} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\log \\frac {\\exp \\left(\\hat {n} _ {j} ^ {T} n _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {D}} \\exp \\left(\\hat {n} _ {j} ^ {T} n _ {k} ^ {\\prime} / \\tau\\right)} \\tag {6}", + "image_path": "895b1aafe81ab5ecd568d314fbc5e1db2dfa16eac5d6c8f961468b2b5b2a4133.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 540, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 575 + ], + "type": "text", + "content": "Training Loss. The total training objective is the sum of the vision-text matching loss and the auxiliary losses on object vectors:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 586, + 287, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 586, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 82, + 586, + 287, + 600 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {v} 2 \\mathrm {t}} ^ {\\mathrm {e g o}} + \\mathcal {L} _ {\\mathrm {t} 2 \\mathrm {v}} ^ {\\mathrm {e g o}} + \\mathcal {L} _ {\\text {b o x}} + \\lambda_ {w o r d} \\mathcal {L} _ {\\text {w o r d}} \\tag {7}", + "image_path": "0df2ae0c4dd42f777e6fe331dea0442f3040a3b840b38b112af4df927aeda8d9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 609, + 114, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 114, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 114, + 621 + ], + "type": "text", + "content": "3.4. Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Once trained, the model acts as a standard ego-centric vision-language video model which operates just on video frames and text descriptions, without requiring further access to object boxes or detectors. However, if desired, hand and object box detections and names can be read out for each frame at inference using the summary vectors, which can be used for grounding the input text description." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 403, + 175 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 403, + 175 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 403, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 403, + 175 + ], + "type": "image", + "image_path": "a37305366833cd3e3ec43a1e04f3b00b373ea24146ec6087df1412c67bb92568.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 180, + 411, + 190 + ], + "lines": [ + { + "bbox": [ + 310, + 180, + 411, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 180, + 411, + 190 + ], + "type": "text", + "content": "Step 1. Object-noun alignment" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 198, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 286 + ], + "type": "text", + "content": "Figure 2. Training the model to predict object classes. Left: Object-noun alignment. First, the nouns in the video descriptions are matched against the predicted classes using Hungarian matching, to choose the most matched summary vectors. Right: Word-level contrastive training. Next, we supervise the matched predicted classes using a contrastive objective [46] against all the nouns in Ego4D taxonomy, to have similar embeddings as the corresponding nouns." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 418, + 74, + 533, + 175 + ], + "blocks": [ + { + "bbox": [ + 418, + 74, + 533, + 175 + ], + "lines": [ + { + "bbox": [ + 418, + 74, + 533, + 175 + ], + "spans": [ + { + "bbox": [ + 418, + 74, + 533, + 175 + ], + "type": "image", + "image_path": "26b27a49804498b929fcac44da66645bec9aaff1c4e7d902d206e6206c9ada7a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 180, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 419, + 180, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 419, + 180, + 544, + 190 + ], + "type": "text", + "content": "Step 2. Word-level contrastive training" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 298, + 403, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 403, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 403, + 312 + ], + "type": "text", + "content": "4. Implementation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 319, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 392 + ], + "type": "text", + "content": "In section 4.1, details of extracting the hand and object detections from 100DOH pre-trained detector are summarized, followed by the architectural details in section 4.2. Finally, in section 4.3 the training pipeline, model and input specification, and optimization details are provided. More details can be found in the supplementary materials." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 400, + 538, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 538, + 413 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 538, + 413 + ], + "type": "text", + "content": "4.1. Weak Supervision from Pre-trained Detector" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "text", + "content": "We uniformly sample four frame from each clip in the EgoClip dataset [35] as the input to the 100DOH hand-object detector [53]. The short side of the frame is resized to 640 pixels. There are 16 million frames in total, and the average number of boxes detected per frame is 1.96 for hand and 1.67 for object. Among all the frames, about " + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "text", + "content": " frames have no hands detected and " + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "inline_equation", + "content": "17.9\\%" + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "text", + "content": " frames have no object detected. The average size of hand boxes is " + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "text", + "content": " of the frame size, while the average size of object boxes is " + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "inline_equation", + "content": "19.4\\%" + }, + { + "bbox": [ + 305, + 419, + 545, + 552 + ], + "type": "text", + "content": " of the frame size. We use the top 2 hands and top 4 objects detected in the scene as supervision." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 562, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 388, + 573 + ], + "type": "text", + "content": "4.2. Architecture" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 581, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 581, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 581, + 545, + 712 + ], + "type": "text", + "content": "The object-aware module is a 6-layer cross-attention transformer with 8 attention-heads in each layer. The hidden dimension in cross-attention layers is 768, and the video embedding, object embeddings and text embeddings are projected to 256 dimensions before computing the cosine similarity score. We set the number of hand queries to 2, number of object queries to 12, which is designed to be larger than the maximum number of objects in the supervision. We use TimeSformer-L (TSF-L) [6] as the visual encoder, and a 12-layer Transformer [61] from LaViLa [77] as the text encoder." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13905" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 87, + 293, + 149 + ], + "blocks": [ + { + "bbox": [ + 93, + 76, + 259, + 86 + ], + "lines": [ + { + "bbox": [ + 93, + 76, + 259, + 86 + ], + "spans": [ + { + "bbox": [ + 93, + 76, + 259, + 86 + ], + "type": "text", + "content": "Text: #C C cuts the box with the utility knife" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 87, + 293, + 149 + ], + "lines": [ + { + "bbox": [ + 49, + 87, + 293, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 87, + 293, + 149 + ], + "type": "image", + "image_path": "f70fec999aae7d610a0c55560ac1c635ceffd69a1a111a2023e082d8023a54b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 163, + 293, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 151, + 284, + 161 + ], + "lines": [ + { + "bbox": [ + 55, + 151, + 284, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 284, + 161 + ], + "type": "text", + "content": "a) Object boxes predicted by off-the-shelf hand-object detector" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 163, + 293, + 224 + ], + "lines": [ + { + "bbox": [ + 49, + 163, + 293, + 224 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 293, + 224 + ], + "type": "image", + "image_path": "b019a8585058e82ca29053cdb795cea348c4da7a6f2d1f3cf6b68702814a1c6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 226, + 270, + 236 + ], + "lines": [ + { + "bbox": [ + 63, + 226, + 270, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 226, + 270, + 236 + ], + "type": "text", + "content": "b) Vision-text grounding results from the proposed model" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 241, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 546, + 285 + ], + "type": "text", + "content": "Figure 3. Visualization of text grounding on EgoClip. We show the comparison between detections from off-the-shelf 100DOH hand-object detector [53] (used for training supervision) and the predicted boxes from our model respectively. (a) The detections are noisy, objects are missed, and there is no temporal association of the detected boxes across frames. (b) The trained model learns temporally consistent tracks as well as object categorization using only noisy frame-level box-supervision and weak supervision from the texts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 87, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 304, + 76, + 539, + 86 + ], + "lines": [ + { + "bbox": [ + 304, + 76, + 539, + 86 + ], + "spans": [ + { + "bbox": [ + 304, + 76, + 539, + 86 + ], + "type": "text", + "content": "C C puts the spatula in his right hand on the chopping board" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 87, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 304, + 87, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 304, + 87, + 544, + 148 + ], + "type": "image", + "image_path": "5bb4f5d0621e1eba35d67fa22d0497d7930bc32200b122b72e940eeb1ef3ce48.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 162, + 544, + 223 + ], + "blocks": [ + { + "bbox": [ + 308, + 151, + 536, + 161 + ], + "lines": [ + { + "bbox": [ + 308, + 151, + 536, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 536, + 161 + ], + "type": "text", + "content": "a) Object boxes predicted by off-the-shelf hand-object detector" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 162, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 304, + 162, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 544, + 223 + ], + "type": "image", + "image_path": "e8e461a3f1036fae36ea4fb708f64b16edaeab28073c05e5c29968cfbc0cf4de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 226, + 527, + 236 + ], + "lines": [ + { + "bbox": [ + 320, + 226, + 527, + 236 + ], + "spans": [ + { + "bbox": [ + 320, + 226, + 527, + 236 + ], + "type": "text", + "content": "b) Vision-text grounding results from the proposed model" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 297, + 145, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 145, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 145, + 309 + ], + "type": "text", + "content": "4.3. Training Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "text", + "content": "In training, we uniformly sample " + }, + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "inline_equation", + "content": "4 \\times 224 \\times 224" + }, + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "text", + "content": " RGB frames from each clip. Hand and object boxes are preextracted from these 4 frames using off-the-shelf detector and used as supervision. We keep the visual and text encoder frozen in training. Only the object-aware decoder, query vectors, and the MLP projection layer on text embedding parameters are learned during training. We train the model for 5 epochs on one A6000 GPU, with batch size 128. We use AdamW [39] as the optimizer and set the learning rate to " + }, + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "inline_equation", + "content": "3e - 5" + }, + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{word}" + }, + { + "bbox": [ + 46, + 316, + 288, + 448 + ], + "type": "text", + "content": " is set to 0.5 to balance the scale of the four losses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 481, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 541 + ], + "type": "text", + "content": "Section 5.1 introduces all the datasets we used for training and evaluation, followed by the evaluation protocols in section 5.2. Finally, we discuss the ablation studies (section 5.3) and compare to prior SOTA methods on different benchmarks (section 5.4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 550, + 110, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 110, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 110, + 561 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 568, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 714 + ], + "type": "text", + "content": "Ego4D/EgoNLQ/EgoMQ/EgoClip/EgoMCQ [16, 35]. Ego4D is a massive-scale dataset focusing on ego-centric videos. It contains 3670 hours video for many different tasks, including action anticipation, AV diarization, etc. EgoNLQ and EgoMQ is a subset for natural language queries and moment query, designed for testing the models' episodic memory and long-term video understanding. [35] proposes a new subset EgoClip for vision-language pre-training, comprising 3.8M clip-text pairs. They also introduce EgoMCQ (i.e., Egocentric Multiple-Choices-Question) as a downstream evaluation dataset for the pre-training. Given a text query, the model tasked to" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 297, + 546, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 546, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 546, + 334 + ], + "type": "text", + "content": "choose the paired video clip from 5 candidates. The evaluation metrics is 'intra-video' and 'inter-video' accuracy, depending on where the candidates are chosen from." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 346, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 442 + ], + "type": "text", + "content": "Epic-Kitchens-MIR [11]. Epic-Kitchens is a large-scale ego-centric dataset with 100-hour activities recorded in kitchens. Epic-Kitchens-MIR is a subset with about 9881 clip-text pairs for vision-language retrieval. It is designed for multi-instance retrieval. The model is evaluated on retrieving the paired text/video given a query text/video. The evaluation metrics are mean average Precision (mAP) and normalized Discounted Cumulative Gain (nDCG)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 455, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 516 + ], + "type": "text", + "content": "VISOR [12] is a benchmark built on Epic-Kitchens for segmenting hands and active objects in egocentric video. It has pixel-level annotations covering 36 hours of untrimmed video and 257 object classes. We utilize this annotations in its val split for evaluations on vision-text grounding." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 528, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 601 + ], + "type": "text", + "content": "EGTEA [34] contains 28 hours of cooking activities from 86 unique sessions of 32 subjects. Fine-grained actions are classified into 106 classes. We retrieve the text-descriptions of action classes to evaluate the model for action classification on the test set of all three splits. We measure the performance using mean-class accuracy and top1 accuracy." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 610, + 422, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 422, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 422, + 621 + ], + "type": "text", + "content": "5.2. Evaluation Protocol" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 629, + 544, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 544, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 544, + 652 + ], + "type": "text", + "content": "We evaluate the performance of our model in the following three aspects:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Zero-shot transfer. To test the transferability and generalization ability, we conduct zero-shot evaluation on multiple-choice questions (EgoMCQ), multi-instance retrieval (EpicKitchens-MIR), action classification (EGTEA). Among these datasets, videos in EgoMCQ are from the same data" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13906" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "source as our pre-training dataset Ego4D. Other datasets demonstrate a domain gap, hence, evaluate for transferable representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 114, + 286, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 114, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 114, + 286, + 199 + ], + "type": "text", + "content": "Episodic memory. To evaluate the richness of the representations learned by our model, we use the video representations to solve Episodic memory tasks in Ego4D. Following [35, 77], we pre-extract the video features from our model first. Using these pre-computed features as input, we train a VSLNet [72] for temporal localization in NLQ, and train a VSGN [76] for moment retrieval in MQ." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 201, + 286, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 286, + 441 + ], + "type": "text", + "content": "Vision-language grounding. Due to the lack of ground-truth for object-grounding in EgoClip, we evaluate the grounding ability of the model on VISOR instead. VISOR is an egocentric dataset with scenes in the kitchens, where frames are annotated sparsely with segmentation masks and object names. We re-propose the manually annotated segmentation masks in it to extract ground-truth bounding-boxes for hands and in-contact objects. To carry out the zero-shot evaluation, we take the annotated frames in the val split, filter out the not-in-contact objects in each frame, and convert all the segmentation masks to bounding boxes as ground truth. The predicted object boxes are matched with ground-truth object boxes using noun alignment as during training (eq. (5)), while the left/right hands are predicted from the first and the second hand queries respectively. We repeat a single frame 4 times temporally, and resize it to " + }, + { + "bbox": [ + 46, + 201, + 286, + 441 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 201, + 286, + 441 + ], + "type": "text", + "content": " pixels to be consistent with the pre-training resolution. The predicted bounding-boxes are evaluated to be correct if their centers lie inside the ground-truth bounding-boxes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 115, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 115, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 115, + 466 + ], + "type": "text", + "content": "5.3. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "text", + "content": "Losses. We ablate the combination of losses on three downstream benchmarks in table 1. Results showing that having both " + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "inline_equation", + "content": "L_{box}" + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "inline_equation", + "content": "L_{word}" + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "text", + "content": " leads to the best performance. With the same architecture, when training the model using only " + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "inline_equation", + "content": "L_{Ego}" + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "text", + "content": " without introducing any object-awareness, the performance is " + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 473, + 286, + 558 + ], + "type": "text", + "content": " lower on EK100-MIR and EGTEA compared to the object-aware one." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 570, + 286, + 654 + ], + "blocks": [ + { + "bbox": [ + 48, + 570, + 286, + 654 + ], + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 654 + ], + "type": "table", + "html": "
LossesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
LEgo93.761.835.936.644.937.6
LEgo + Box94.262.736.737.445.338.5
LEgo + Word93.761.936.737.445.838.1
LEgo + Box + Word94.563.037.537.846.639.1
", + "image_path": "3f74d15c2d53a56b5095241920b02b9f763d77b27b33bf2377e0ccc055af813f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 46, + 659, + 286, + 704 + ], + "lines": [ + { + "bbox": [ + 46, + 659, + 286, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 286, + 704 + ], + "type": "text", + "content": "Table 1. Ablation on training objectives for zero-shot transfer tasks. Introducing the object-awareness by having box and word supervision helps the model to achieve better transfer results on EK100-MIR and EGTEA." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 148 + ], + "type": "table", + "html": "
Detector input resEgoMCQEKEGTEAVISOR
InterIntraAvg mAPAvg nDCGTop1MeanLoc Acc
256p94.263.235.734.642.036.0-
256p94.563.036.937.044.338.968.2
640p94.563.037.537.839.146.678.7
", + "image_path": "8d68c5e93932f379d0b469682ad827e67de50dbfdc70985537f5e4cf93031712.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 311, + 207, + 538, + 282 + ], + "blocks": [ + { + "bbox": [ + 305, + 155, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 200 + ], + "type": "text", + "content": "Table 2. Ablation on box quality for zero-shot transfer tasks. We extract boxes using " + }, + { + "bbox": [ + 305, + 155, + 545, + 200 + ], + "type": "inline_equation", + "content": "256\\mathrm{p}" + }, + { + "bbox": [ + 305, + 155, + 545, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 155, + 545, + 200 + ], + "type": "inline_equation", + "content": "640\\mathrm{p}" + }, + { + "bbox": [ + 305, + 155, + 545, + 200 + ], + "type": "text", + "content": " images as input to the detector respectively, resulting in boxes of different qualities as supervision in training." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 207, + 538, + 282 + ], + "lines": [ + { + "bbox": [ + 311, + 207, + 538, + 282 + ], + "spans": [ + { + "bbox": [ + 311, + 207, + 538, + 282 + ], + "type": "table", + "html": "
# Obj QueriesEgoMCQEK100-MIREGTEA
InterIntraAvg mAPAvg nDCGTop1Mean
494.162.837.837.645.938.1
894.562.737.738.045.537.9
1294.563.037.537.846.639.1
", + "image_path": "80272ae0e50669a50e9897f9c3dc8a75ec0f1ae682baca23398cf927a0fa4b5c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 286, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 305, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 545, + 342 + ], + "type": "text", + "content": "Table 3. Ablation on the number of object queries for zero-shot transfer tasks. The number of queries do not have a big impact on EK100. Larger number of queries shows a boost on mean-class accuracy on EGTEA, and smaller number of queries is better on intra-video accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 357, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 563 + ], + "type": "text", + "content": "Quality of detected boxes. The extent to which a model can acquire object-level information is constrained by the quality of the bounding boxes from the off-the-shelf detector. To investigate how much the quality of boxes affects our training, we detect hand and object boxes on EgoClip training set using 100DOH [53] with 256p and 640p images as input – with larger image size, the objects should be more precisely delineated. We use the two sets of boxes as supervision in our training and show results in table 2. Even when training with noisy boxes from 256p, our model is better than the previous SOTA model. When boxes from 640p are used, the averaged zero-shot transfer performance is further improved by " + }, + { + "bbox": [ + 304, + 357, + 545, + 563 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 357, + 545, + 563 + ], + "type": "text", + "content": " on Epic-Kitchens and EGTEA, showing that our method can bring larger improvement over the non-object-aware method when given better boxes. Furthermore, a significant boost on the grounding results is also observed on VISOR when using boxes with better quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": "Number of object queries. We use different number of query vectors in the object-aware decoder to see its impact on both vision-language tasks and the grounding task. We show zero-shot transfer results in table 3 and vision-text grounding results in table 7. The number of queries has relatively small impact on vision-language representation learning, the performance gaps on zero-shot transfer tasks are mostly small than " + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": ". However, a smaller number of object queries leads to much better results for in-contact object grounding on VISOR, 4 queries is better than 12 objects by " + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": " in localization accuracy. The reason is that too" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13907" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 181 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 181 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 181 + ], + "type": "table", + "html": "
MethodBackboneBatch sizeObject awareHard negEgoMCQEK100-MIREGTEA
Inter-videoIntra-videoV-TmAP T-VAvgV-TnDCG T-VAvgTop1-AccMean-Acc
EgoVLPTSF-B512X✓90.657.226.020.623.328.827.027.917.6-
LaViLaTSF-B1024XX93.859.935.126.630.933.730.432.0-28.9
LaViLaTSF-L1024XX94.563.140.032.236.136.133.234.640.134.1
LaViLa*TSF-L1024XX94.263.239.731.735.736.133.234.642.036.0
Ours*†TSF-L128XX93.760.539.730.335.037.334.535.944.836.3
Ours*‡TSF-L128X✓93.761.840.731.135.938.335.036.644.937.6
Ours*TSF-L128✓✓94.563.042.332.737.539.336.237.846.639.1
", + "image_path": "226be90f4350c23a35c35bdbc28d6516ffef74c03f8475e7f76050ef196a89e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 240, + 279, + 289 + ], + "blocks": [ + { + "bbox": [ + 46, + 184, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 547, + 231 + ], + "type": "text", + "content": "Table 4. Comparison to SOTA results on zero-shot transfer to EgoMCQ, EK100-MIR and EGTEA. We compared to EgoVLP and LaVILA, two previous SOTA models pre-trained on EgoClip. Our object-aware model has achieved comparable results on multiple-choice questions on EgoMCQ, and SOTA results on multi-instance retrieval task on EpicKicthens and action classification on EGTEA. Model without * use center cropping in evaluation, while * denotes the usage of resizing instead of cropping. Ours† and Ours‡ stands for different variants of our model depending on whether object-aware losses and hard negative sampling is used in training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 240, + 279, + 289 + ], + "lines": [ + { + "bbox": [ + 53, + 240, + 279, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 279, + 289 + ], + "type": "table", + "html": "
MethodPredicted BoxesEgoMCQEKEGTEA
InterIntraAvg mAPAvg nDCGMeanTop1
Oursobj94.062.136.837.145.137.6
hand+obj94.563.037.537.846.639.1
", + "image_path": "cf53cfdef2430e943c4a86328432fa4d233159dca172046d1de4adcfb6f324ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "text", + "content": "many queries result in a large number of predicted boxes, which increases the probability of mis-matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "type": "text", + "content": "Impact of hand boxes. We ablate the impact of having hand boxes as supervision in our training. Results are shown in table 5, training the model to predict hand as well as objects help the model to get about " + }, + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "type": "text", + "content": " higher zero-shot transfer performance on EK and EGTEA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 441, + 188, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 188, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 188, + 454 + ], + "type": "text", + "content": "5.4. Comparison to the SOTA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": "Zero-shot transfer. We compare to previous SOTA in table 4. Our model is comparable on EgoMCQ and better on EK-MIR and EGTEA, showing its good zero-shot transferability. Due to limited compute resources, we are not able to unfreeze the visual backbone to train end-to-end or increase the batch size further. Despite these disadvantages, our method outperforms the previous SOTA on two tasks for models that have been trained end-to-end." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": "The main difference between LaViLa(L) and ours is the object-aware training and hard negative sampling; Ours† in table 4 is a LaViLa(L) model with an extra Transformer decoder, which is trained with only InfoNCE loss on video and sentence embeddings. Without object-awareness and hard sampling, it gets better accuracy on EGTEA and better nDCG on EK100 due to more parameters added, but falls behind on EgoMCQ and mAP on EK100. Applying hard negative sampling (Ours‡) and object-aware training (Ours) brings improvement across the board. The most obvious boost comes from inducing object-awareness, bringing " + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 556, + 288, + 714 + ], + "type": "text", + "content": " improvement on average. And the results could be further improved by obtaining better pseudo-boxes, as the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 238, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 262 + ], + "type": "text", + "content": "magnitude of boost from learning objects is closely related to the box quality (as shown in table 2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 280, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 498 + ], + "type": "text", + "content": "Episodic memory. Results on EgoNLQ and EgoMQ are shown in table 6. The two tasks test the video understanding on several-minutes long videos. In these experiments, trained video and text backbones are used as feature extractor, and extra modules are trained on the long feature sequences for natural language querying and memory querying. Therefore, the richer the information encoded in the features, the better the results will be. We list the models trained on the same amount of video and text data in black, results show that our model are better than the previous SOTA on all the metrics in the two tasks. This is because features from the object-aware model have captured more object information, thus enabling better precision and recall on localization and retrieval. InternVideo [63] and NaQ [50] are trained on more video or text data, and are included for completeness. Please note that our method is compatible with any visual language backbones, using better backbones might also lead to better results." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 508, + 545, + 639 + ], + "blocks": [ + { + "bbox": [ + 46, + 293, + 288, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 326 + ], + "type": "text", + "content": "Table 5. The impact of having hand boxes in the box prediction. Having hand boxes in the prediction helps on all the zero-shot evaluation benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 508, + 545, + 639 + ], + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 639 + ], + "type": "table", + "html": "
MethodBatch sizeEgoNLQEgoMQ
mIOU@0.3mIOU@0.5R1@0.5R5@0.5mAP
R1R5R1R5
SlowFast-5.510.73.16.625.246.26.0
EgoVLP51210.818.86.813.530.152.011.4
LaViLa(B)102410.519.16.713.6---
LaViLa(L)102412.122.47.315.432.556.113.4
Ours12813.223.37.915.633.456.716.0
VideoIntern [63]14k16.523.010.116.1--23.6
ReLER + NaQ [50]204819.323.611.615.5---
", + "image_path": "61d427f9b6f7af1ecfe74994e6dde2be044b823c260865ee523cc4060fcdb41f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 645, + 547, + 712 + ], + "lines": [ + { + "bbox": [ + 304, + 645, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 547, + 712 + ], + "type": "text", + "content": "Table 6. Comparison to SOTA results of fine-tuning on EgoNLQ and EgoMQ. Our object-aware model encodes richer information in the visual representations, hence obtaining better results on all the metrics in NLQ and MQ task in Ego4D episodic memory benchmark. We list other SOTA models (in grey) trained with more video and text data for completeness." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13908" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 153 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 153 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 153 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 153 + ], + "type": "table", + "html": "
ModelObject Assignment# queriesLoc Accuracy
Detector [53]Random-37.1
GT matching-41.3
OursPredicted482.3
881.2
1278.7
", + "image_path": "4824fa32f4fd25d96fa8576afc114fe9a6d2457555e1cc19ad30da1231fd991c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 158, + 288, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 226 + ], + "type": "text", + "content": "Table 7. In-contact object localization accuracy on VISOR. Our model does better on in-contact object localization after weakly-supervised training compared to the baseline (the Detector with Random or GT object assignment), which is the source of supervision in our pre-training. The improvement is due to increased recall of our model over the baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 233, + 211, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 211, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 211, + 247 + ], + "type": "text", + "content": "5.5. Evaluating Object Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 397 + ], + "type": "text", + "content": "Qualitative results on EgoClip. In fig. 3 we show the grounding results on EgoClip after training as compared to the supervision from the 100DOH detector. After training, the predictions from our model find the missing objects because we do not penalize extra bounding boxes predicted, but select the active object/object of interest through matching noun embeddings. It also learns temporal association of object bounding-boxes as the result of using the same object summary vector to predict the boxes over all the frames (as in eq. (1)); the summary vector attends to visually similar features corresponding to the same object across the clip without any explicit supervision for temporal consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 405, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 624 + ], + "type": "text", + "content": "Quantitative and qualitative results on VISOR. In table 7 we show the text-grounding results on VISOR. We take the predictions from the hand-object detector [53] as our baseline. Since the predictions only detect hands and objects without an object class name, we associate the predicted boxes with ground-truth boxes in VISOR in two ways: (a) random: we assign predicted object boxes to GT object boxes randomly, (b) GT matching: we use Hungarian matching to find the predicted boxes with highest IoU against the GT object boxes. However, even when matched using GT information, the baseline detector does not achieve a high accuracy due to poor recall. Results show grounding ability of our model is " + }, + { + "bbox": [ + 46, + 405, + 287, + 624 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 405, + 287, + 624 + ], + "type": "text", + "content": " better than the baseline using only weak supervision from the video descriptions. We also show the qualitative results on detecting different number of hands and objects in fig. 4, our model has a much higher recall when compared to baseline detector when operating at the same resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 119, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "In this paper, we introduce a method to learn object-aware ego-centric video representations using noisy supervision from pre-trained hand-object detectors. The object-representations so learned show strong zero-shot transfer across various downstream tasks and datasets, mirroring" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 541, + 292 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 541, + 292 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 541, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 541, + 292 + ], + "type": "image", + "image_path": "fdd00c87270569799d09f157cbd51d0748e5acba85626628a894d61e917a5b39.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 295, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 295, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 545, + 384 + ], + "type": "text", + "content": "Figure 4. Grounding visualization on the Epic-Kitchen VISOR val split. When operating at the same resolution, our model shows better grounding ability on hands and objects compared to the baseline 100DOH detector [53] (with GT matching) used for training. Note that the low IoU on hands on the third column is a result of the GT 'hand' segmentation mask covering the arm by definition, while the detector and our model are trained to localize only palm and fingers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 634 + ], + "type": "text", + "content": "the performance improvement from object-aware training on images [70]. The model uses standard neural modules (i.e., transformers), and does not require any object boxes or detectors as input during inference, making it widely applicable as a drop-in replacement for training video-language models. Even though the model is trained with sparse and noisy object supervision at the frame-level (without temporal associations), during inference dense temporal bounding-box tracks and category predictions can be obtained, which are superior to the predictions from the base hand-object detector used for training. There are several avenues for improvement. Our model uses the pre-trained video encoder operates at a small resolution " + }, + { + "bbox": [ + 304, + 403, + 545, + 634 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 403, + 545, + 634 + ], + "type": "text", + "content": " which makes detecting small objects difficult. Further, four frames are sampled uniformly from the clip regardless of its length which can cause difficulties due to temporal aliasing. Nevertheless, we hope our work will inspire further research in learning transferable object-aware representations for videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 647, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 545, + 696 + ], + "type": "text", + "content": "Acknowledgements. This research is funded by a Google-DeepMind Graduate Scholarship, a Royal Society Research Professorship, and EPSRC Programme Grant VisualAI EP/T028572/1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13909" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 287, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 287, + 177 + ], + "type": "text", + "content": "[2] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In Proc. ICCV, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 178, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 288, + 232 + ], + "type": "text", + "content": "[3] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 232, + 288, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 288, + 264 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proc. ICCV, 2021. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 319 + ], + "type": "text", + "content": "[5] Sven Bambach, Stefan Lee, David J Crandall, and Chen Yu. Lending a hand: Detecting hands and recognizing activities in complex egocentric interactions. In Proceedings of the IEEE international conference on computer vision, pages 1949-1957, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 350 + ], + "type": "text", + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding. In Proc. ICML, 2021. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 393 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. ECCV, 2020. 1, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 395, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 395, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 395, + 288, + 437 + ], + "type": "text", + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Proc. ECCV, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 438, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 288, + 481 + ], + "type": "text", + "content": "[9] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 502 + ], + "type": "text", + "content": "[10] Noam Chomsky and David W Lightfoot. Syntactic structures. Walter de Gruyter, 1957. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 288, + 555 + ], + "type": "text", + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proc. ECCV, 2018. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "type": "text", + "content": "[12] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In NeurIPS Datasets and Benchmarks Track, 2022. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 610, + 288, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 288, + 665 + ], + "type": "text", + "content": "[13] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. arXiv preprint arXiv:2206.07643, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 665, + 288, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 288, + 708 + ], + "type": "text", + "content": "[14] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurlPS, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 710 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 94 + ], + "type": "text", + "content": "[15] Philip Gage. A new algorithm for data compression. C Users Journal, 1994. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "text", + "content": "[16] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proc. CVPR, 2022. 1, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 148, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 170 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proc. ICCV, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 170, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 170, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 545, + 201 + ], + "type": "text", + "content": "[18] Paul Henderson and Christoph H. Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 202, + 545, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 202, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 545, + 246 + ], + "type": "text", + "content": "[19] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. arXiv preprint arXiv:2110.06915, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 246, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 288 + ], + "type": "text", + "content": "[20] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proc. ICCV, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 289, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 342 + ], + "type": "text", + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[22] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proc. CVPR, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "text", + "content": "[23] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proc. ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 418, + 545, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 440 + ], + "type": "text", + "content": "[24] Ranjay Krishna, Ines Chami, Michael Bernstein, and Li Fei-Fei. Referring relationships. In Proc. CVPR, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "type": "text", + "content": "[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 494, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 494, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 545, + 548 + ], + "type": "text", + "content": "[26] Matej Kristan, Jiří Matas, Aleš Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Cehovin, Alan Lukežić, et al. The ninth visual object tracking vot2021 challenge results. In Proc. ICCV, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "text", + "content": "[27] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 581, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 634 + ], + "type": "text", + "content": "[28] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proc. CVPR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 666 + ], + "type": "text", + "content": "[29] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 2017. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 667, + 545, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 667, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 307, + 667, + 545, + 710 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proc. ICML, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13910" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 711 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 159 + ], + "type": "text", + "content": "[32] Shuang Li, Yilun Du, Antonio Torralba, Josef Sivic, and Bryan Russell. Weakly supervised human-object interaction detection in video via contrastive spatiotemporal regions. In Proc. ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 160, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 160, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 287, + 203 + ], + "type": "text", + "content": "[33] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Proc. ECCV, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 203, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 203, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 203, + 287, + 236 + ], + "type": "text", + "content": "[34] Yin Li, Miao Liu, and James M Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. In Proc. ECCV, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 236, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 236, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 236, + 287, + 290 + ], + "type": "text", + "content": "[35] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 290, + 287, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 332 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 399 + ], + "type": "text", + "content": "[37] Miao Liu, Siyu Tang, Yin Li, and James M Rehg. Forecasting human-object interaction: joint prediction of motor attention and actions in first person video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 704-721. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[38] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 474 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. ICLR, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "text", + "content": "[40] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 506, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 287, + 548 + ], + "type": "text", + "content": "[41] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "text", + "content": "[42] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In Proc. CVPR, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 624 + ], + "type": "text", + "content": "[43] Effrosyni Mavroudi and René Vidal. Weakly-supervised generation and grounding of visual descriptions with conditional generative models. In Proc. CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[44] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proc. ICCV, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 679, + 287, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 679, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 679, + 287, + 699 + ], + "type": "text", + "content": "[45] Marvin Minsky. Society of mind. Simon and Schuster, 1988. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 700, + 287, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 700, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 48, + 700, + 287, + 711 + ], + "type": "text", + "content": "[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "sensation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "type": "text", + "content": "[47] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE conference on computer vision and pattern recognition, pages 2847-2854. IEEE, 2012. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 138, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 138, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 138, + 545, + 170 + ], + "type": "text", + "content": "[48] Gorjan Radevski, Marie-Francine Moens, and Tinne Tuytelaars. Revisiting spatio-temporal layouts for compositional action recognition. In Proc. BMVC, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 170, + 545, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 170, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 545, + 224 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 224, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 224, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 224, + 545, + 278 + ], + "type": "text", + "content": "[50] Santhosh Kumar Ramakrishnan, Ziad Al-Halah, and Kristen Grauman. Naq: Leveraging narrations as queries to supervise episodic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6694-6703, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 279, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 279, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 279, + 545, + 322 + ], + "type": "text", + "content": "[51] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proc. CVPR, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 322, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 322, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 322, + 545, + 387 + ], + "type": "text", + "content": "[52] Kate Saenko, Ben Packer, C Chen, Sunil Bandla, Y Lee, Yangqing Jia, J Niebles, Daphne Koller, Li Fei-Fei, Kristen Grauman, et al. Mid-level features improve recognition of interactive activities. Technical report, Dept. of Elec. Engg. and Computer Sc., University of California, Berkeley, 2012. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "text", + "content": "[53] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proc. CVPR, 2020. 1, 2, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 420, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 452 + ], + "type": "text", + "content": "[54] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Proc. ICLR, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 483 + ], + "type": "text", + "content": "[55] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 483, + 545, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 516 + ], + "type": "text", + "content": "[56] Elizabeth S Spelke, Karen Breinlinger, Janet Macomber, and Kristen Jacobson. Origins of knowledge. Psychological review, 99, 1992. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 516, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 516, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 545, + 548 + ], + "type": "text", + "content": "[57] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In Proc. ECCV, 2018. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[58] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "text", + "content": "[59] Reuben Tan, Bryan Plummer, Kate Saenko, Hailin Jin, and Bryan Russell. Look at what i'm doing: Self-supervised spatial grounding of narrations in instructional videos. *NeurIPS*, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 656 + ], + "type": "text", + "content": "[60] Joshua B Tenenbaum, Charles Kemp, Thomas L Griffiths, and Noah D Goodman. How to grow a mind: Statistics, structure, and abstraction. Science, 2011. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 656, + 545, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 656, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 308, + 656, + 545, + 699 + ], + "type": "text", + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 700, + 545, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 700, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 308, + 700, + 545, + 711 + ], + "type": "text", + "content": "[62] Xiaolong Wang and Abhinav Gupta. Videos as space-time" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13911" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 711 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 73, + 220, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 220, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 220, + 83 + ], + "type": "text", + "content": "region graphs. In Proc. ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 84, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 84, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 287, + 137 + ], + "type": "text", + "content": "[63] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 138, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 287, + 192 + ], + "type": "text", + "content": "[64] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 192, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 192, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 192, + 287, + 236 + ], + "type": "text", + "content": "[65] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 236, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 236, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 236, + 287, + 278 + ], + "type": "text", + "content": "[66] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In Proc. ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 300 + ], + "type": "text", + "content": "[67] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In Proc. ICCV, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 300, + 287, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 343 + ], + "type": "text", + "content": "[68] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 344, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 397 + ], + "type": "text", + "content": "[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 398, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 430 + ], + "type": "text", + "content": "[70] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 2, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 430, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 461 + ], + "type": "text", + "content": "[71] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Is an object-centric video representation beneficial for transfer? In Proc. ACCV, 2022. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 462, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 505 + ], + "type": "text", + "content": "[72] Hao Zhang, Aixin Sun, Wei Jing, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE PAMI, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 506, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 287, + 559 + ], + "type": "text", + "content": "[73] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "text", + "content": "[74] Lingzhi Zhang, Shenghao Zhou, Simon Stent, and Jianbo Shi. Fine-grained egocentric hand-object segmentation: Dataset, model, and applications. In European Conference on Computer Vision, pages 127-145. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "text", + "content": "[75] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proc. CVPR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[76] Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13658-13667, 2021. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 689, + 287, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 287, + 711 + ], + "type": "text", + "content": "[77] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 256 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 530, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 530, + 83 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 530, + 83 + ], + "type": "text", + "content": "models.arXiv preprint arXiv:2212.04501,2022.2,5,7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 84, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 84, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 84, + 545, + 127 + ], + "type": "text", + "content": "[78] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proc. CVPR, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 127, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 127, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 545, + 169 + ], + "type": "text", + "content": "[79] Luowei Zhou, Nathan Louis, and Jason J Corso. Weakly-supervised video object grounding from text by loss weighting and object interaction. arXiv preprint arXiv:1805.02834, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 170, + 545, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 170, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 545, + 213 + ], + "type": "text", + "content": "[80] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 213, + 545, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 256 + ], + "type": "text", + "content": "[81] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In Proc. ECCV, 2022. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13912" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_content_list.json b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9f9c68afa45d42d4065d93d7f9038efba01332cc --- /dev/null +++ b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_content_list.json @@ -0,0 +1,1380 @@ +[ + { + "type": "text", + "text": "Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking", + "text_level": 1, + "bbox": [ + 104, + 161, + 865, + 182 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Li $^{1,2,*}$ , Baopeng Zhang $^{1}$ , Jun Liu $^{2}$ , Wei Liu $^{1}$ , Jian Zhao $^{3,4,5}$ , Zhu Teng $^{1,\\dagger}$", + "bbox": [ + 179, + 218, + 794, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Beijing Jiaotong University $^{2}$ Singapore University of Technology and Design $^{3}$ Institute of North Electronic Equipment $^{4}$ Peng Cheng Laboratory $^{5}$ Intelligent Game and Decision Laboratory", + "bbox": [ + 105, + 245, + 863, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 315, + 312, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The existing one-stage multi-object tracking (MOT) algorithms have achieved satisfactory performance benefiting from a large amount of labeled data. However, acquiring plenty of laborious annotated frames is not practical in real applications. To reduce the cost of human annotations, we propose Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT), to infer the most informative frames for any MOT tracker by observing the heterogeneous cues of samples. HD-AMOT defines the diversified informative representation by encoding the geometric and semantic information, and formulates the frame inference strategy as a Markov decision process to learn an optimal sampling policy based on the designed informative representation. Specifically, HD-AMOT consists of a diversified informative representation module as well as an informative frame selection network. The former produces the signal characterizing the diversity and distribution of frames, and the latter receives the signal and conducts multi-frame cooperation to enable batch frame sampling. Extensive experiments conducted on the MOT15, MOT17, MOT20, and Dancetrack datasets demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that under $50\\%$ budget our HD-AMOT can achieve similar or even higher performance as fully-supervised learning.", + "bbox": [ + 75, + 344, + 473, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 701, + 208, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a sophisticated research topic, one-stage multi-object tracking (MOT) has attracted more research attention in recent years, which simultaneously focuses on pinpointing multiple objects in each frame and recording their trajectories in the entire sequence. Despite the great success of deep learning-based one-stage MOT methods [25, 48, 27, 26, 52, 6, 22], they need a large amount of annotated samples for training, which requires costly human labor. For example,", + "bbox": [ + 75, + 724, + 468, + 836 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd2276a73ebf6b921b24894cb9f21cca6e1c5c7b0395a60c6cbde7d53b1b5b40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 505, + 314, + 694, + 437 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fa4108bcecbb1f2717227a1709816ee3e013bac64665718c1a1bc4fbc3ef62c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 314, + 885, + 437 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa970e0183898c76838bc58553bc1306de5f0327e53917d143304a715d2027ea.jpg", + "image_caption": [ + "Figure 1. A stimulus example of active learning for one-stage MOT. There are many similar continuous frames in tracking sequences, such as frames 365-368 of MOT17-02 [23], and fully annotated these frames may have little effect on improving the performance of the tracker. Moreover, objects usually have different states, such as moving and static states, as shown in the red dotted box. The moving states usually contain more knowledge than the static ones, in which case the uniform and random sampling may fail to capture the diverse visual patterns of tracking sequences." + ], + "image_footnote": [], + "bbox": [ + 507, + 437, + 682, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64614865bfa16b36aa8d537043410eabbff7c4508bef431ce0599f5d3047b3be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 437, + 885, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in the MOT17 dataset [23], a sequence containing only 600 frames requires annotating 18,581 bounding boxes (including coordinates and identity labels) for 62 pedestrians. In addition, there are many similar continuous frames in tracking sequences, such as frames 365-368 of the MOT17-02 sequence in Fig. 1, which may have little effect on boosting the performance of the tracker. One way to deal with the above issue is to randomly or uniformly sample tracking sequences for annotation. However, as shown in the red dotted box of Fig. 1, objects in tracking sequences may have different motion states, and the moving state contains more knowledge than the static state, in which case the uniform sampling may fail to capture diverse visual patterns. Moreover, random sampling is easy to yield unstable results. Therefore, it is essential to develop algorithms to infer training data more efficiently in one-stage MOT.", + "bbox": [ + 497, + 627, + 892, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instead of selecting the subset randomly or uniformly,", + "bbox": [ + 517, + 851, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 84 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 253, + 48, + 789, + 86 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done when Rui Li was an visiting student of SUTD. †Zhu Teng is the corresponding author (zteng@bjtu.edu.cn).", + "bbox": [ + 92, + 840, + 413, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9932", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "active learning (AL) aims to infer the most informative samples as training data to improve the knowledge of a model, which is a promising way to select training data more efficiently for one-stage MOT. In the computer vision area, active learning has been widely explored for image classification [3, 21, 1, 8], object detection [45, 44, 10], semantic segmentation [1, 7, 28, 9], and so on. In these research topics, the image selection strategies of uncertainty-based [32, 42] and distribution-based methods [31, 1] become the mainstream of active learning. However, these AL algorithms do not directly relate to the performance improvements of the specific task and usually lead to suboptimal performance. More notably, they are usually utilized to select independent image samples, which are prone to achieve poor coverage of the tracking sequence and produce redundant frames if they are employed to the one-stage MOT task stiffly and directly. In particular, since new individuals appear constantly in the tracking sequence, the distribution bias caused by the poor coverage and redundant frames may make the tracker unable to generalize well for unseen cases.", + "bbox": [ + 75, + 125, + 472, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, different from other tasks that distinguish inter-class objects in each image (such as image classification, object detection, etc.), one-stage MOT needs to detect and associate intra-class objects, e.g., pedestrians belonging to different individuals. This implies that the active learning for one-stage MOT is more challenging, because it should consider not only the discrepancy between frame-level images, but also the influence of more fine-grained intra-class diversity in sample selection, which plays a vital role in object association in one-stage MOT.", + "bbox": [ + 75, + 404, + 472, + 543 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we formulate the active learning of one-stage MOT as a Markov Decision Process (MDP) [36] and design a novel Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT) framework, which leverages a Diversified Informative Representation Module (DIRM) and an Informative Frame Selection Network (IFSN) to learn a frame sampling policy. Specifically, IFSN is designed to infer a batch of informative and diverse frames for annotation by receiving the signal characterizing the diversity and distribution of frames and conducting multi-frame cooperation, which can effectively inhibit poor coverage and reduce redundant frames. In addition, to facilitate the inference of benignant frames, we develop a novel DIRM to characterize the frame-level diversified representation of samples by encoding the geometric and semantic cues of frames and objects. Besides the frame-level diversity representation, we also evaluate the set-level (labeled set and unlabeled set) discrepancy in DIRM, which dedicates to the acquirement of the training set with the unbiased distribution.", + "bbox": [ + 75, + 545, + 472, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows: (1) We propose a novel HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process (MDP) and", + "bbox": [ + 77, + 825, + 470, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learns a sampling policy driven by the reward that directly relates to the tracker performance. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning. (2) To infer the benignant frames, we design a diversified informative representation to characterize the diversity and distribution of frames, which learns both the set-level discrepancy and frame-level diversity by encoding the heterogeneous cues of geometry and semantics. (3) Extensive experiments are conducted on four datasets to demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that our HD-AMOT can achieve similar or even higher performance as fully-supervised learning under $50\\%$ annotation budget.", + "bbox": [ + 497, + 125, + 894, + 307 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 318, + 640, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Object Tracking. MOT is a challenging task due to the complexity of tracking scenes. The existing works mainly follow the tracking-by-detection paradigm and two-stage methods are the mainstream, which regards the detection and tracking as two independent models [4, 14, 46, 40, 11, 17]. Despite noticeable progress, the two-stage methods cause a waste of resources and maybe lead to an efficiency issue because the independent detection and tracking models are executed separately in these methods. Recently, with the quick maturity of deep learning[50, 49, 51], one-stage frameworks for MOT [25, 48, 27, 26, 53, 52, 6, 22, 47] have begun to attract more research attention, which greatly improve the accuracy and efficiency of tracking. The fundamental concept revolves around the simultaneous integration of object detection and association within a singular network architecture. For example, CTracker [27] employs a chaining mechanism that connects paired bounding-box regression outcomes derived from overlapping nodes. Each of these nodes spans two consecutive frames. FairMOT [48] presents a simple yet effective approach based on the detection architecture CenterNet [13]. Despite the notable advancements introduced by these methods, their practical applicability remains constrained by the demanding nature of high-cost annotations.", + "bbox": [ + 497, + 340, + 894, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To reduce the demand for labeled data, several methods with less supervision signal such as weakly-supervised and self-supervised learning [30, 2] can effectively utilize the unlabeled data to improve the MOT performance. However, these methods pay more attention to the tracker learning, and ignore the importance of labeled data selection to improve the tracker performance. In this paper, we design an active learning approach for one-stage MOT, which could significantly reduce the annotation cost. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning.", + "bbox": [ + 497, + 672, + 895, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Active Learning. Active learning is an important machine learning problem, which has received lots of attention in computer vision, such as image classification", + "bbox": [ + 497, + 825, + 894, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9933", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/60681c4cffe5cb407fbedf6ac91e4f41106122178f9fbc5e2e117917afbff96f.jpg", + "image_caption": [ + "Figure 2. The pipeline of our proposed HD-AMOT framework. In our method, we perform several iterations until a budget $\\mathcal{B}$ of labeled frames is achieved. In the $t^{th}$ iteration, the steps are as follows: (I) A one-stage MOT tracker $\\pmb{f}_t$ is used to obtain the heterogeneous cues including semantic feature $\\pmb{F}^f$ and geometric information $\\pmb{G}^h$ , $\\pmb{G}^b$ on each frame; (II) The set-level discrepancy and frame-level diversity are estimated and construct the MDP state-action pair $(s_t, a_t)$ ; (III) The multi-frame cooperation based on $(s_t, a_t)$ is performed; (IV) With the assistance of multi-frame cooperation, IFSN infers $N$ unlabeled frames with high scores to be annotated. Meanwhile, $\\pmb{X}_t^U$ , $\\pmb{X}_t^L$ are updated to $\\pmb{X}_{t+1}^U$ , $\\pmb{X}_{t+1}^L$ by displacing newly annotated frames from $\\pmb{X}_t^U$ to $\\pmb{X}_t^L$ ; (V) The tracker $\\pmb{f}_t$ is retrained on $\\pmb{X}_{t+1}^L$ to obtain the updated $\\pmb{f}_{t+1}$ . (VI) The reward $r_{t+1}$ based on $\\pmb{f}_t$ and $\\pmb{f}_{t+1}$ is computed on $\\pmb{X}^R$ to update parameters of IFSN." + ], + "image_footnote": [], + "bbox": [ + 80, + 124, + 890, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[1, 8, 39, 20], object detection [45, 44, 10], semantic segmentation [7, 28, 33], and so on. These methods can be divided into uncertainty-based [32, 42] and distribution-based methods [31, 1]. For example, [15] exploits the uncertainty in both the input and output spaces to select the most valuable information. [31] defines the problem of active learning as core-set selection. [32] develops a foundation LearningLoss++ for the active learning of pose estimation to establish equivalency between Learning Loss [42] empirically driven objective and the KL divergence objective. These AL algorithms are used to select independent image samples for a specific task yet are not suitable for the video-wise one-stage MOT task. It is prone to cause the problems such as poor coverage and redundant frames if we directly apply the existing AL approaches to one-stage MOT stiffly. Accordingly, we design an active learning framework HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process and learns a batch frame sampling policy driven by the reward that directly relates to the performance of the MOT tracker, which gives consideration to both sampling coverage and efficiency.", + "bbox": [ + 75, + 399, + 472, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MDP of Multi-Object Tracking. As a sequential decision model, Markov decision process is usually applied in dynamic environments where an agent needs to perform certain tasks by making decisions and executing actions sequentially. In multi-object tracking, several MDP-based works have been explored [41, 29, 16]. [41] formulates the online two-stage MOT problem as decision making in MDPs, where the lifetime of an object is represented using an MDP model. [29] treats individual objects as agents, utilizing a prediction network for tracking, while optimizing tracking outcomes through collaborative interactions among various agents and their environments, facilitated by", + "bbox": [ + 75, + 700, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the decision network. In this paper, different from these methods that employ the MDP to directly solve the detection and tracking subtask of MOT, we address the annotation budget for one-stage MOT based on active learning by the designed Markov decision process, and design a diversified informative representation with an informative frame selection network to enable effective batch sampling under a specific annotation budget.", + "bbox": [ + 497, + 399, + 892, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Proposed HD-AMOT", + "text_level": 1, + "bbox": [ + 497, + 527, + 742, + 543 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 499, + 551, + 610, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, given an unlabeled video sequence $X$ with a limited annotation budget $\\mathcal{B}$ , our HD-AMOT model aims to learn an optimal sampling strategy that infers and annotates the most informative frames iteratively to maximize the performance of the tracker $f$ . Specifically, we cast this AL problem of one-stage MOT by an MDP and adopt the Q-learning algorithm [24] to solve this problem, where we introduce an IFSN to infer frames according to the designed state-action representation $(s_t,a_t)$ . As shown in Fig. 2, in our method, the video sequence $X$ is divided into three different subsets $X^U,X^L$ , and $X^R$ . And at each iteration $t$ , the following steps are executed:", + "bbox": [ + 497, + 574, + 890, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(I) The heterogeneous cues including semantic features $\\pmb{F}^{f}$ and geometric information $\\pmb{G}^{h}$ , $\\pmb{G}^{b}$ of each frame in the labeled set $\\pmb{X}_{t}^{L}$ and unlabeled set $\\pmb{X}_{t}^{U}$ are obtained by $\\pmb{f}_{t}$ ; (II) We estimate the set-level discrepancy and frame-level diversity in DIRM based on the obtained heterogeneous cues, which are regarded as the MDP state-action pair $(s_{t}, a_{t})$ ; (III) The multi-frame cooperation based on $(s_{t}, a_{t})$ is performed on $\\pmb{X}_{t}^{L}$ and $\\pmb{X}_{t}^{U}$ ; (IV) With the assistance of multi-frame cooperation, IFSN selects $N$ unlabeled frames", + "bbox": [ + 497, + 741, + 892, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9934", + "bbox": [ + 467, + 885, + 500, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e71eb7e3d333436b962026a00fbac0056dea54a993e23212baec850924e3b4f8.jpg", + "image_caption": [ + "Figure 3. Illustration of geometric matrices construction process. (a): binary heatmap; (b): global topology matrix $G^{h}$ ; (c): local topology matrices $G^{h_1} \\sim G^{h_K}$ ; (d): object scale matrix $G^b$ ." + ], + "image_footnote": [], + "bbox": [ + 82, + 124, + 463, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with high scores to be annotated guided by frame clusters. Meanwhile, the subsets $X_{t}^{U}$ , $X_{t}^{L}$ are updated to $X_{t+1}^{U}$ , $X_{t+1}^{L}$ by moving newly annotated frames from $X_{t}^{U}$ to $X_{t}^{L}$ ; (V) The tracker $\\pmb{f}_{t}$ is retrained on $X_{t+1}^{L}$ to obtain the updated $\\pmb{f}_{t+1}$ ; (VI) The reward $r_{t+1}$ based on $\\pmb{f}_{t}$ and $\\pmb{f}_{t+1}$ is computed on the reward subset $\\pmb{X}^{R}$ , which is used to update parameters of IFSN.", + "bbox": [ + 75, + 258, + 467, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Diversified Informative Representation", + "text_level": 1, + "bbox": [ + 77, + 372, + 410, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To infer the most informative and representative frames, we propose a DIRM to compute the state and action of the designed MDP, which characterizes the diversity and distribution of frames. Specifically, to compensate for the distribution drift between the unlabeled set $\\mathbf{X}_t^U$ and the labeled set $\\mathbf{X}_t^L$ , we use heterogeneous cues to evaluate the set-level discrepancy as state $s_t$ . Moreover, from the perspective of selecting the unlabeled frame with the potential contribution to the MOT tracker training, we propose to capture the frame-level diversity of each unlabeled frame as action $a_t$ .", + "bbox": [ + 75, + 394, + 467, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Set-level Discrepancy. There are a large number of objects belonging to different individuals in the tracking sequence, and new individuals may appear constantly over time. To make the trained tracker well generalized to unseen individuals, it is necessary to generate an unbiased training set that captures the whole sequence distribution. Based on this intuition, we propose to narrow the distribution gap between the unlabeled and labeled sets by evaluating set-level discrepancy, to ensure the diversity and representativeness of the selected frames.", + "bbox": [ + 75, + 534, + 467, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we consider two key attributes to characterize the distribution drifts between the labeled and unlabeled sets: semantic variation and spatial topological variation. We collect the global feature from the last feature extraction layer of $\\pmb{f}_t$ to obtain the semantic feature $F^f$ , which depicts the general semantic information of the frame. For the spatial topology $G^h$ , we encode the center of each object obtained by the tracker $f_t$ into a binary heatmap to represent the spatial distribution of objects. Considering that identities of objects in the frames with similar spatial topology may be quite different due to the object motion, we encode the identities into the binary heatmap to obtain the final $G^h$ , as shown in Fig.3 (a) and (b). In addition, the performance of the MOT tracker is affected by objects or", + "bbox": [ + 75, + 672, + 467, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "background, leading to different tracking quality over various local spatial areas. To suppress adverse effects of the low-quality local areas, we decompose the global spatial topology $\\pmb{G}^{h}$ into $K$ local parts, which form our final spatial topology $\\{\\pmb{G}^{h},\\pmb{G}^{h_{1} - h_{K}}\\}$ together with the global topology. Similarly, the global and local semantic features constitute the final semantics $\\{\\pmb{F}^{f},\\pmb{F}^{f_{1} - f_{K}}\\}$ . In this way, the feature space $S_{set} = \\{\\pmb{G}^{h},\\pmb{G}^{h_{1} - h_{K}},\\pmb{F}^{f},\\pmb{F}^{f_{1} - f_{K}}\\}$ is formed. To model the set-level discrepancy between $X_{t}^{L}$ and $X_{t}^{U}$ , Maximum Mean Discrepancy [37] is adopted to calculate the set gap for each feature $S_{set}^{m}$ in the feature space $S_{set}$ as described in Eq. (1). Finally, the state representation $s_t$ is defined as the concatenation of all $\\mathcal{D}^m$ to encode the distribution drifts between $X_{t}^{L}$ and $X_{t}^{U}$ .", + "bbox": [ + 497, + 125, + 890, + 320 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} ^ {m} = \\frac {1}{n _ {l} ^ {2}} \\sum_ {i, j} \\varphi \\left(p _ {i} ^ {m}, p _ {j} ^ {m}\\right) + \\frac {1}{n _ {u} ^ {2}} \\sum_ {i, j} \\varphi \\left(q _ {i} ^ {m}, q _ {j} ^ {m}\\right) \\tag {1} \\\\ - \\frac {2}{n _ {l} n _ {u}} \\sum_ {i, j} \\varphi \\left(p _ {i} ^ {m}, q _ {j} ^ {m}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 338, + 890, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_m$ is a scalar representing the distribution discrepancy between $X_{t}^{L}$ and $X_{t}^{U}$ on $S_{\\text{set}}^{m}$ . $p^m$ and $q^m$ denote the corresponding features of frames in $X_{t}^{L}$ and $X_{t}^{U}$ , $n_l$ and $n_u$ are numbers of frames in $X_{t}^{L}$ and $X_{t}^{U}$ . $\\varphi(.)$ is the radial kernel [37] to measure the distance between two features.", + "bbox": [ + 497, + 409, + 890, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frame-level Diversity. In this paper, we aim to select the frames that are beneficial to the tracker training for annotation. For this purpose, besides considering the set-level discrepancy, we also need to explore frame diversity from different perspectives to infer the most representative and informative frames. Different from the semantic feature and spatial topology used to learn set-level discrepancy, finer-grained cues are desiderated to evaluate the potential contribution of each unlabeled frame. Specifically, besides the frame-wise spatial topology $G^{h}$ and semantic feature $F^{f}$ , we also consider finer-grained attribute characterizing the individuation of objects: the scale of bounding boxes $G^{b}$ . Its construction process is similar to $G^{h}$ , except that the identities representing different individuals are replaced by the width and height of the corresponding object as shown in Fig.3 (d), i.e., $G^{h} \\in \\mathbb{R}^{H \\times W \\times 1}$ and $G^{b} \\in \\mathbb{R}^{H \\times W \\times 2}$ .", + "bbox": [ + 497, + 479, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, we argue that any one of the geometric features $G^{h}$ , $G^{b}$ and semantic features $F^{f}$ is the key factor to determine whether a frame contributes to the performance improvement of the tracker. Accordingly, we consider the novelty and prominence of each unlabeled frame $x^{u}$ based on $S_{frame} = \\{G^{h}, G^{b}, F^{f}\\}$ , to form the frame-level diversity as action $a_{t}$ . The novelty of an unlabeled frame $x_{i}^{u}$ in $X_{t}^{U}$ is reflected in the similarities between $x_{i}^{u}$ and the labeled frames in $X_{t}^{L}$ , because the unlabeled frame with a novel characteristic (such as new individuals, different object scales, and postures) usually has low similarities to labeled frames of $X_{t}^{L}$ . And the promi", + "bbox": [ + 497, + 700, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9935", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nence of $x_{i}^{u}$ lies in the difference between $x_{i}^{u}$ and other unlabeled frames in $X_{t}^{U}$ , so we expect to select the most representative frame by recording the similarity distribution between each unlabeled frame and other frames in $X_{t}^{U}$ . Specifically, we introduce histogram-based representation to learn the novelty and prominence of each unlabeled frame. Taking $x_{i}^{u}$ as an example, to represent the novelty of $x_{i}^{u}$ , we calculate cosine similarities between $x_{i}^{u}$ and $X_{t}^{L}$ on each heterogeneous feature $S_{frame}^{m}$ and then obtain the histogram $\\mathcal{H}_{nov}$ based on cosine similarities. Similar to novelty learning, we learn prominence by recoding the histogram $\\mathcal{H}_{pro}$ based on cosine similarities between $x_{i}^{u}$ and $X_{t}^{U}$ . Finally, the action representation $a_{t}$ of $x_{i}^{u}$ is defined as $\\{\\mathcal{H}_{nov}^{G^{h}},\\mathcal{H}_{nov}^{G^{b}},\\mathcal{H}_{nov}^{F^{f}},\\mathcal{H}_{pro}^{G^{h}},\\mathcal{H}_{pro}^{G^{b}},\\mathcal{H}_{pro}^{F^{f}}\\}$ , which conduces to IFSN to effectively identify and select the informative and representative frames.", + "bbox": [ + 75, + 125, + 472, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Informative Frame Selection Learning", + "text_level": 1, + "bbox": [ + 77, + 357, + 410, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by DQN [38], our lightweight IFSN follows an optimal policy that maximizes the potential performance gains. Considering the sampling efficiency, we propose a batch sampling method guided by frame clusters according to temporal clues. Meanwhile, to restrain the inferior performance caused by redundant frames, we further introduce the multi-frame cooperation strategy before batch sampling.", + "bbox": [ + 75, + 378, + 468, + 475 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-frame Cooperation. There are many similar adjacent frames in tracking sequences, which easily leads to redundancy in the process of frame selection. To reduce redundant frames, we introduce a simple multi-frame cooperation strategy to model the neighborhood cooperation between the unlabeled frame $x_{i}^{u}$ and its nearest labeled frames. Specifically, we seek the nearest neighbors $x_{v}^{l}$ and $x_{w}^{l}$ of $x_{i}^{u}$ in $X_{t}^{L}$ and calculate their actions $a_{t}^{v}$ , $a_{t}^{w}$ . Then we use the fixed-length compact representation $o_{t} = \\frac{1}{2}(a_{t}^{v} + a_{t}^{w})$ as an extra action for $x_{i}^{u}$ . To this end, the selection made by IFSN becomes:", + "bbox": [ + 75, + 475, + 472, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {t} = \\arg \\max _ {a _ {t} \\in A _ {t}} g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 633, + 468, + 653 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a_{t}\\in A_{t}$ denotes the candidate action in $X_{t}^{U}$ , and $g(.)$ is the lightweight IFSN network parameterized by $\\theta$ .", + "bbox": [ + 75, + 659, + 468, + 687 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Batch Frame Selection. The structure of our designed lightweight IFSN consists of four linear layers, where a linear layer parameterized by $\\theta_{1}$ is used to model the relation between each unlabeled frame and its nearest neighbors, and three linear layers parameterized by $\\theta_{2\\sim 4}$ are employed to score each unlabeled frame, as described in Eq. (3), where $cat[\\cdot]$ denotes the concatenation operation.", + "bbox": [ + 75, + 687, + 468, + 783 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right) = \\theta_ {2 \\sim 4} \\left(c a t \\left[ \\theta_ {1} \\left(a _ {t}, o _ {t}\\right), a _ {t}, s _ {t} \\right]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 790, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, each unlabeled frame in $X_{t}^{U}$ can obtain the corresponding score by IFSN. However, sampling a single frame in each iteration by Eq. (2) to query annotation is inefficient. In this paper, we propose a batch sampling method", + "bbox": [ + 75, + 809, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "guided by frame clusters to sample $N$ frames in each iteration. Based on the intuition that a larger difference exists between frames with longer time intervals, we divide the unlabeled frames into $N$ clusters in chronological order and select the frame for annotation in each cluster independently, which is equivalent to labeling a frame in parallel with $N$ annotators. In this case, $\\phi_t$ in Eq. (2) is extended to consist of $N$ independent sub-selections $\\{\\theta_t^n\\}_{n=1}^N$ , each with a restricted space $A_t^n$ , as described in Eq. (4).", + "bbox": [ + 497, + 125, + 890, + 251 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {t} ^ {n} = \\arg \\max _ {a _ {t} \\in A _ {t} ^ {n}} g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 258, + 890, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a_{t}\\in A_{t}^{n}$ denotes the candidate action in $n^{th}$ cluster.", + "bbox": [ + 497, + 290, + 887, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "IFSN Optimization. In our method, we compute the reward on a subset $X^R$ , which is used to evaluate the benefit of selected frames to the tracker and optimize the frame selection network IFSN. To evaluate the performance of the tracker more comprehensively, we utilize MOTA [5] and IDF1 [19] as performance metrics to calculate the reward, where MOTA is a comprehensive metric to evaluate the detection performance and IDF1 tends to estimate the tracking ability of a tracker. As described in Eq. (5), the reward $r_{t+1}$ is defined as the difference of the performance metrics between $f_{t+1}$ and $f_t$ .", + "bbox": [ + 497, + 305, + 890, + 457 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} = \\left(e _ {t + 1} ^ {\\text {m o t a}} + e _ {t + 1} ^ {i d f 1}\\right) - \\left(e _ {t} ^ {\\text {m o t a}} + e _ {t} ^ {i d f 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 569, + 464, + 890, + 480 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_{t+1}^{mota}$ and $e_{t+1}^{idf1}$ are MOTA and IDF1 metrics computed by $f_{t+1}$ , $e_t^{mota}$ and $e_t^{idf1}$ are evaluated by $f_t$ .", + "bbox": [ + 497, + 490, + 890, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the reward $r_{t + 1}$ , we can optimize IFSN to infer the most informative frames to maximize the reward, leading to improved tracker performance during each active learning iteration. More specifically, we train IFSN in a double DQN formulation [38] by optimizing the loss based on the temporal difference error [35] as:", + "bbox": [ + 497, + 523, + 890, + 605 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I F S N}} = \\left(\\hat {y} _ {t} - g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right)\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 612, + 890, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t} = r _ {t + 1} + \\gamma g \\left(s _ {t + 1}, a _ {t + 1}, o _ {t + 1}; \\hat {\\theta}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 652, + 890, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ and $\\hat{\\theta}$ are parameters of the IFSN policy network and off-policy network following the setting of double DQN [38]. $\\gamma$ denotes a discount factor.", + "bbox": [ + 497, + 669, + 890, + 711 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the described diversified informative representation in Section 3.2 and informative frame selection learning in this section, we introduce our entire HD-AMOT as illustrated in Alg. 1.", + "bbox": [ + 497, + 711, + 890, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 499, + 780, + 623, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 499, + 802, + 705, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our HD-AMOT on MOT15 [18], MOT17 [23], MOT20 [12], and Dancetrack [34]. MOT15 and MOT17 are widely-used MOT datasets, and they are", + "bbox": [ + 497, + 824, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9936", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: HD-AMOT Algorithm." + ], + "code_body": "Input: The sequence $X$ , initial tracker $f_{init}$ and IFSN $g_{init}$ , AL procedure rounds $Z$ . \n1 $\\mathbf{X}_{init}, \\mathbf{X}^{R} \\gets$ RandomPartition $(X)$ ; \n2 $\\mathbf{X}_{init}^{U}, \\mathbf{X}_{init}^{L} \\gets$ UniformPartition $(X_{init})$ ; \n3 $\\mathbf{X}_{0}^{L} \\gets \\mathbf{X}_{init}^{L}, f_{0} \\gets$ update $(g_{init}, \\mathbf{X}_{init}^{L})$ ; \n4 while not done do // Episodes \n5 for $t = 0$ to $t = Z - 1$ do // AL procedure \n6 Learn the diversified representation to form the state and action space; \n7 Use IFSN to select frames by Eq. (4); \n8 Annotate the selected $N$ frames $\\mathbf{X}_{t}^{S}$ ; \n9 Update $\\mathbf{X}_{t}^{U}, \\mathbf{X}_{t}^{L}$ : \n10 $\\mathbf{X}_{t + 1}^{U} \\gets \\mathbf{X}_{t}^{U} \\setminus \\mathbf{X}_{t}^{S}, \\mathbf{X}_{t + 1}^{L} \\gets \\mathbf{X}_{t}^{L} \\cup \\mathbf{X}_{t}^{S}$ ; \n11 $f_{t + 1} \\gets$ update $(f_{t}, \\mathbf{X}_{t + 1}^{L})$ ; \n12 Calculate the reward on $\\mathbf{X}^{R}$ by Eq. (5); \n13 end \n14 Update IFSN following Eq. (6); \n15 Reset: $\\mathbf{X}_{0}^{L} \\gets \\mathbf{X}_{init}^{L}, f_{0} \\gets f_{t + 1}$ ; \n16 end", + "bbox": [ + 80, + 144, + 436, + 399 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "challenging because of the large variety in object pose, lighting, viewpoint, etc. In contrast to MOT15 and MOT17, MOT20 addresses the challenge of highly congested scenarios, where the pedestrian density can surge to as high as 246 individuals per frame. Dancetrack is a large-scale MOT dataset covering scenarios with low distinguishability of object appearance and diverse non-linear motion patterns.", + "bbox": [ + 75, + 434, + 468, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the performance of one-stage MOT, the standard metrics MOTA [5] and IDF1 [19] are adopted for evaluation, which indicate the detection and association performance of one-stage MOT. Specifically, MOTA combines three error sources including false positives, missed targets, and identity switches, and IDF1 is the ratio of correctly identified detections over the average number of ground-truth and computed detections.", + "bbox": [ + 75, + 532, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In this paper, we adopt the popular one-stage MOT method FairMOT [48] as the MOT tracker. Note that the propose algorithm is suitable for any one-stage MOT tracker. The input frames are resized to $1088 \\times 608$ . We set the learning rate of our IFSN to 1e-4 and the discount factor $\\gamma$ in Eq. (7) to 0.9, and the number of local features $K$ is set to 9 according to experiments. For each dataset, we randomly select a small sequence in each video at a ratio of 0.1 to form a reward subset $X^R$ . In the remaining frames, a small amount of frames is uniformly sampled as the initially labeled subset $X_{init}^{L}$ to initialize the MOT tracker $f$ , and the size is set to 100. For the semantic information, we extract the feature map from the last convolutional layer of ResNet34 as the global semantic feature $F^{f} \\in \\mathbb{R}^{512 \\times 19 \\times 34}$ . For the geometric clues, we utilize the center coordinates and identities of bounding", + "bbox": [ + 75, + 644, + 468, + 865 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/361d2d3e6f8a23bf6208e7df311d58407d1e99ebdc25193aa462d4953114e772.jpg", + "table_caption": [ + "Table 1. Ablation study on the design of the diversified informative representation." + ], + "table_footnote": [], + "table_body": "
RepresentationSettingMOTA (%)IDF1 (%)
Set-level DiscrepancyState w/o Gh,h1-h961.7364.95
State w/o Ff,f1-f961.4964.12
State w/o Gh, Ff62.7465.21
State w/o Gh1-h9, Ff1-f962.9264.55
Frame-level DiversityAction w/o Hgh62.8366.61
Action w/o HGb62.4966.23
Action w/o HFf61.7864.95
Action w/o H* nov61.2563.41
Action w/o H* pro62.2864.16
AllHD-AMOT63.2366.47
", + "bbox": [ + 502, + 149, + 890, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "boxes obtained by the MOT tracker to construct the global spatial topology $G^{h} \\in \\mathbb{R}^{152 \\times 272}$ , and generate the object scale matrix $G^{b} \\in \\mathbb{R}^{152 \\times 272 \\times 2}$ by the box size. In addition, the number of clusters $N$ in the batch sampling is set to 12 for MOT15, MOT17, and MOT20, and is set to 24 for Dancetrack as it is much larger than other datasets.", + "bbox": [ + 497, + 312, + 890, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 499, + 407, + 652, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, to analyze the impacts of different components, we conduct ablation studies on the popular dataset MOT17 following the division strategy in [43]. We run our experiments 5 times and report the mean performance for all ablation studies.", + "bbox": [ + 497, + 429, + 890, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Diversified Informative Representation. We perform an ablation study with a $20\\%$ annotation budget to evaluate the contribution of our diversified informative representation. As IFSN relies on the state (set-level discrepancy) to decide the frame sampling policy, we investigate the influence of the set-level discrepancy by removing different components individually. Moreover, we also analyze the effect of the action design (frame-level diversity). As shown in Tab. 1, our complete diversified informative representation gives the best performance in terms of MOTA and is only slightly inferior to Action w/o $\\mathcal{H}^{G^h}$ by $0.14\\%$ on IDF1. We speculate that the scale information in $\\mathcal{H}^{G^b}$ implicitly indicates the identity and topology clues of $\\mathcal{H}^{G^h}$ , so the introduction of $\\mathcal{H}^{G^b}$ weakens the role of $\\mathcal{H}^{G^h}$ , making the MOTA of Action w/o $\\mathcal{H}^{G^h}$ close to HD-AMOT and the IDF1 slightly higher. Furthermore, removing the components of set-level discrepancy or frame-level diversity in the diversified informative representation degrades the tracking performance. The largest decrease of MOTA and IDF1 occurs when the score of maximum similarity with labeled frames Action w/o $\\mathcal{H}_{nov}^*$ is removed, which further verifies the effectiveness of heterogeneous cues evaluating the novel characteristic of unlabeled frames.", + "bbox": [ + 497, + 500, + 892, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Informative Frame Selection Network. We further validate the performance of our designed IFSN in Tab. 2. We first consider performing the cluster-based batch", + "bbox": [ + 497, + 824, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9937", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6713a37a99244cf5ba17408b5b6fcc30d11382dc2ff1b1d47fc762363faab943.jpg", + "table_caption": [ + "Table 2. Ablation study on the different settings of the Informative Frame Selection Network." + ], + "table_footnote": [], + "table_body": "
ModelMulti-frame CooperationBatch SamplingMOTA (%)↑IDF1 (%)↑
Bât.Clu.
A×✓✓62.7665.52
B✓××63.0765.62
C✓✓×62.1864.55
D✓✓✓63.2366.47
", + "bbox": [ + 87, + 146, + 458, + 225 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddc0ac4ffd87d1abbf0f5331a72de198fc7e51f21bcaabbea4c00ab7907d31c1.jpg", + "image_caption": [ + "Figure 4. Performance comparison evaluated by MOTA and IDF1 metrics with increasing budgets." + ], + "image_footnote": [], + "bbox": [ + 80, + 230, + 272, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43fc21309b8fe98f1e2f9dddb4ffd954707319f93751ded78b95648d8990a2ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 231, + 463, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sampling without multi-frame cooperation (Model A). Then we analyze the usage of selecting a single frame in each AL iteration (Model B). In addition, we also construct the Model C as IFSN to select a batch of frames in one shot, where the frames with $N$ highest scores are sampled. Finally, we present the performance of our cluster-based batch sampling strategy (Model D). By comparing the results of Model A, Model D brings performance gains of $0.47\\%$ and $0.95\\%$ in terms of MOTA and IDF1 metrics, which proves that multi-frame cooperation plays a positive role in the active learning of one-stage MOT. In contrast, Model C gives the worst performance, which selects multiple frames according to the $N$ highest scores. We argue that frames with higher scores may be similar samples and close to each other, which easily leads to inefficiency in batch sampling. To this end, cluster-based batch sampling is introduced, as it can avoid selecting adjacent frames with higher scores and learn to sample with better coverage of the underlying distribution. Notably, selecting a single frame at each iteration in Model B also provides a competitive performance but still tends to be inferior to our method Model D, and the time cost of selecting a single frame per iteration is much higher than our cluster-based batch sampling strategy.", + "bbox": [ + 75, + 435, + 468, + 754 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Different Budgets. We report the performance of our method, random sampling, uniform sampling, and other existing AL methods such as Coreset [31], CDAL [1] under different budgets, as shown in Fig. 4. For the IDF1 metric, ours is slightly inferior to the uniform sampling when the budget is $20\\%$ and $40\\%$ . We argue that uniform sampling can select informative frames for annotation if a small number of labeled frames is required. However,", + "bbox": [ + 75, + 755, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "when the budget increases, there may be redundancy in the newly acquired frames by uniform sampling, and training with these labels does not provide more additional information, resulting in the stagnation or even decline of the performance of IDF1. Notably, our HD-AMOT with the $50\\%$ budget achieves similar performance as other methods with the $80\\%$ budget. It reveals that ours can save around $30\\%$ labeling efforts on average compared to other sampling methods. We attribute this to our learned diversified informative clues and fine-grained object representation, which can effectively improve the tracking performance.", + "bbox": [ + 497, + 125, + 890, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with other AL Methods", + "text_level": 1, + "bbox": [ + 497, + 289, + 815, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further verify the generalization and effectiveness, we evaluate the proposed framework HD-AMOT on four diversified MOT datasets in this section.", + "bbox": [ + 497, + 311, + 890, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on MOT15. Tab. 3 lists the mean values of MOTA and IDF1 on MOT15, and Fig. 5 reveals the corresponding standard deviation of MOTA and IDF1. We can observe that HD-AMOT outperforms all methods by a clear margin, i.e., AL methods used in other tasks (CDAL [1] and Coreset [31]) stiffly applied to one-stage MOT can lead to obvious performance degradation. Notably, although Coreset achieves the second-best performance on MOTA and IDF1, it is clear from the standard deviations in Fig. 5 that Coreset has limited stability. Moreover, the performance of Coreset on other datasets (MOT17, MOT20, and Dancetrack) is not ideal, which indicates that the direct application of Coreset in one-stage MOT is unstable.", + "bbox": [ + 497, + 353, + 890, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on MOT17. MOT17 is a challenging dataset with more crowded scenarios, different viewpoints, camera motions, and weather conditions. As shown in Tab. 3, although uniform sampling achieves the best IDF1 on MOT17 and our HD-AMOT performs slightly lower than it by a margin of $0.18\\%$ , uniform sampling may fail to capture diverse visual patterns in other datasets due to the lack of flexibility and adaptability. For example, in MOT15, ours outperforms uniform sampling by all performance metrics, which reflects the generalization ability of ours for one-stage MOT. Moreover, besides object association, the object detection ability should be considered in one-stage MOT. Our method, which benefits from learning the sampling policy driven by tracking performance metrics directly from data, significantly outperforms uniform sampling by $0.84\\%$ in terms of the detection accuracy metric MOTA.", + "bbox": [ + 497, + 533, + 890, + 754 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on MOT20. We also use the large-scale dataset MOT20 to demonstrate the scalability of our HD-AMOT as shown in Tab. 3. In contrast to the aforementioned datasets, MOT20 is a recently released dataset that features more densely populated scenes. We can observe that ours still achieves the best performance, which demonstrates that our HD-AMOT can infer informative frames even on a crowded and complicated dataset.", + "bbox": [ + 497, + 755, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9938", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/25aae838a9b7812a88da3698f8a0abc08b75fa77760fa130b13dbd6e01bfd4d8.jpg", + "table_caption": [ + "Table 3. Mean values of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack. Best and second-best results are shown in red and black, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetMethodMOTA(%) (↑)IDF1(%) (↑)
MOT15Tracker w/ Random53.2859.61
Tracker w/ Uniform54.0959.50
Tracker w/ CDAL [1]53.1358.85
Tracker w/ Coreset [31]54.7060.29
Tracker w/ HD-AMOT55.2960.64
MOT17Tracker w/ Random70.1169.24
Tracker w/ Uniform70.4070.05
Tracker w/ CDAL [1]69.1068.71
Tracker w/ Coreset [31]69.4468.61
Tracker w/ HD-AMOT71.2469.87
MOT20Tracker w/ Random58.2165.37
Tracker w/ Uniform58.6565.87
Tracker w/ CDAL [1]57.5965.33
Tracker w/ Coreset [31]56.6664.86
Tracker w/ HD-AMOT59.2466.41
DancetrackTracker w/ Random80.4940.67
Tracker w/ Uniform81.1141.03
Tracker w/ CDAL [1]81.3641.16
Tracker w/ Coreset [31]81.1340.47
Tracker w/ HD-AMOT81.9441.93
", + "bbox": [ + 78, + 174, + 468, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6e6eaebe838bfb20171ae117726a260ef5830c3ba1f44b309db6101f3cd1960.jpg", + "image_caption": [ + "Figure 5. Standard deviations of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack." + ], + "image_footnote": [], + "bbox": [ + 78, + 435, + 273, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17f51011c2492d07fcf4d65ce23164ca53a45b5507b2e6850884de454d951d64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 273, + 435, + 467, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Dancetrack. Dancetrack is a larger and more challenging dataset for group dancing scenes, where humans have similar appearance, diverse motion, and extreme articulation. We provide the performance evaluation results on the large-scale Dancetrack in Tab. 3. It can be observed that our HD-AMOT still achieves the best MOTA and IDF1 performance. Notably, from the perspective of sampling stability in Fig. 5, although uniform sampling can obtain a low standard deviation of MOTA, it leads to a very high standard deviation of IDF1 due to failure to capture informative frames that are helpful to object association. By contrast, our proposed HD-AMOT can maintain stable MOTA and IDF1 performance in multiple samplings.", + "bbox": [ + 75, + 643, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we give the performance comparison between our HD-AMOT with a $50\\%$ budget and full supervision in Fig. 6. Our active learning framework achieves sim", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03d6db03f7fd845e9bba4207a329949565aeda57edb72df38eeacc65e52e9436.jpg", + "image_caption": [ + "Figure 6. Performance comparison between our HD-AMOT with a $50\\%$ budget and full supervision on MOT15, MOT17, MOT20, and Dancetrack datasets." + ], + "image_footnote": [], + "bbox": [ + 502, + 124, + 702, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d32978e371dd9a232d66a1b70dd66856408330330e2530a7cf5e470a74a6a087.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 124, + 887, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ilar or even higher performance compared with the tracker with fully-supervised training. For instance, we exceed the full supervision by a gain of $2\\%$ on MOTA and $0.8\\%$ on IDF1 for MOT15. These results implicitly prove that there are many similar frames in MOT sequences, and these similar frames bring limited progress in tracking performance. Our proposed HD-AMOT learns an optimal sampling policy by the heterogeneous diversity representation and a Markov decision process, which can infer informative frames that can most benefit the tracker, thus effectively improving the performance of the tracker. In summary, the proposed method can provide a consistent gain for the MOT tracker across datasets.", + "bbox": [ + 497, + 328, + 890, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 523, + 618, + 538 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework HD-AMOT to investigate the annotation budget for one-stage MOT. It is formulated as an MDP and learns a frame sampling policy driven by the reward that directly relates to the performance of the tracker. HD-AMOT learns the diversified informative representation of frames and achieves effective cluster-based batch frame selection guided by multi-frame cooperation. We conduct extensive ablation studies to verify the design of our framework and compare our performance with random, uniform sampling, and several existing active learning works on four datasets including MOT15, MOT17, MOT20, and Dancetrack. The experimental results demonstrate the efficacy and effectiveness of our HD-AMOT.", + "bbox": [ + 497, + 547, + 890, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the Fundamental Research Funds for the Central Universities of China (2022JBMC009), the Natural Science Foundation of China (61972027, 62006244), and the Beijing Municipal Natural Science Foundation (Grant No. 4212041). This work is also partially supported by Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and Young Elite Scientist Sponsorship Program of Beijing Association for Science and Technology (BYESS2021178).", + "bbox": [ + 497, + 726, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9939", + "bbox": [ + 468, + 885, + 500, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 171, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In ECCV, 2020.", + "[2] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In NeurIPS, 2021.", + "[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In CVPR, 2018.", + "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019.", + "[5] Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008.", + "[6] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR, 2022.", + "[7] Lile Cai, Xun Xu, Jun Hao Liew, and Chuan Sheng Foo. Revisiting superpixels for active learning in semantic segmentation with realistic annotation costs. In CVPR, 2021.", + "[8] Razvan Caramalau, Binod Bhattacharai, and Tae-Kyun Kim. Sequential graph convolutional network for active learning. In CVPR, 2021.", + "[9] Arantxa Casanova, Pedro O. Pinheiro, Negar Rostamzadeh, and Christopher J. Pal. Reinforced active learning for image segmentation. In ICLR, 2020.", + "[10] Jiwong Choi, Ismail Elezi, Hyuk-Jae Lee, Clement Farabet, and Jose M Alvarez. Active learning for deep object detection via probabilistic modeling. In CVPR, 2021.", + "[11] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In CVPR, 2021.", + "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020.", + "[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, 2019.", + "[14] Jiawei He, Zehao Huang, Naiyan Wang, and Zhaoxiang Zhang. Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In CVPR, 2021.", + "[15] Sheng-Jun Huang, Nengneng Gao, and Songcan Chen. Multi-instance multi-label active learning. In *IJCAI*, 2017.", + "[16] Ming-xin Jiang, Chao Deng, Zhi-geng Pan, Lan-fang Wang, and Xing Sun. Multiobject tracking in videos based on lstm and deep reinforcement learning. Complexity, 2018, 2018.", + "[17] Chanho Kim, Li Fuxin, Mazen Alotaibi, and James M Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, 2021.", + "[18] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942, 2015." + ], + "bbox": [ + 78, + 147, + 468, + 865 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Yuan Li, Chang Huang, and Ram Nevatia. Learning to associate: Hybridboosted multi-target tracker for crowded scene. In CVPR, 2009.", + "[20] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In ICCV, 2021.", + "[21] Zhao-Yang Liu and Sheng-Jun Huang. Active sampling for open-set classification without initial annotation. In AAAI, 2019.", + "[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022.", + "[23] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv:1603.00831, 2016.", + "[24] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. In NeurIPS, 2013.", + "[25] Bo Pang, Yizhuo Li, Yifan Zhang, Muchen Li, and Cewu Lu. Tubetk: Adopting tubes to track multi-object in a one-step training model. In CVPR, 2020.", + "[26] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021.", + "[27] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In ECCV, 2020.", + "[28] Yu Qiao, Jincheng Zhu, Chengjiang Long, Zeyao Zhang, Yuxin Wang, Zhenjun Du, and Xin Yang. Cpral: Collaborative panoptic-regional active learning for semantic segmentation. In AAAI, 2022.", + "[29] Liangliang Ren, Jiwen Lu, Zifeng Wang, Qi Tian, and Jie Zhou. Collaborative deep reinforcement learning for multi-object tracking. In ECCV, 2018.", + "[30] Idoia Ruiz, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Joan Serrat. Weakly supervised multi-object tracking and segmentation. In WACV, 2021.", + "[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In ICLR, 2018.", + "[32] Megh Shukla and Shuaib Ahmed. A mathematical analysis of learning loss for active learning in regression. In CVPR, 2021.", + "[33] Yawar Siddiqui, Julien Valentin, and Matthias Nießner. Viewal: Active learning with viewpoint entropy for semantic segmentation. In CVPR, 2020.", + "[34] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, 2022.", + "[35] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, 1988.", + "[36] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018." + ], + "bbox": [ + 502, + 127, + 890, + 865 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9940", + "bbox": [ + 468, + 887, + 500, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. In NeurIPS, 2016.", + "[38] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In AAAI, 2016.", + "[39] Ran Wang, Sam Kwong, Xu Wang, and Yuheng Jia. Active k-labelsets ensemble for multi-label classification. PR, 109:107583, 2021.", + "[40] Jun Xiang, Guohan Xu, Chao Ma, and Jianhua Hou. End-to-end learning deep crf models for multi-object tracking deep crf models. IEEE TCSVT, 31(1):275-288, 2020.", + "[41] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015.", + "[42] Donggeun Yoo and In So Kweon. Learning loss for active learning. In CVPR, 2019.", + "[43] Fengwei Yu, Wenbo Li, Quanquan Li, Yu Liu, Xiaohua Shi, and Junjie Yan. Poi: Multiple object tracking with high performance detection and appearance feature. In ECCV, 2016.", + "[44] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In CVPR, 2021.", + "[45] Xiaosong Zhang, Fang Wan, Chang Liu, Xiangyang Ji, and Qixiang Ye. Learning to match anchors for visual object detection. IEEE TPAMI, 44(6):3096-3109, 2021.", + "[46] Yang Zhang, Hao Sheng, Yubin Wu, Shuai Wang, Weifeng Lyu, Wei Ke, and Zhang Xiong. Long-term tracking with deep tracklet association. IEEE TIP, 29:6694-6706, 2020.", + "[47] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022.", + "[48] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021.", + "[49] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018.", + "[50] Jian Zhao, Lin Xiong, Jianshu Li, Junliang Xing, Shuicheng Yan, and Jiashi Feng. 3d-aided dual-agent gans for unconstrained face recognition. IEEE TPAMI, 41(10):2380-2394, 2018.", + "[51] Jian Zhao, Shuicheng Yan, and Jiashi Feng. Towards age-invariant face recognition. IEEE TPAMI, 44(1):474-487, 2020.", + "[52] Linyu Zheng, Ming Tang, Yingying Chen, Guibo Zhu, Jinqiao Wang, and Hanqing Lu. Improving multiple object tracking with single object tracking. In CVPR, pages 2453-2462, 2021.", + "[53] Xue-Feng Zhu, Tianyang Xu, Jian Zhao, Jia-Wei Liu, Kai Wang, Gang Wang, Jianan Li, Zhihao Zhang, Qiang Wang, Lei Jin, et al. Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-uav system. arXiv preprint arXiv:2306.15767, 2023." + ], + "bbox": [ + 78, + 125, + 468, + 853 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9941", + "bbox": [ + 468, + 885, + 499, + 897 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_model.json b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..02afc81e71d3a8bdaaabdef38f2cbd59c2a4a39f --- /dev/null +++ b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.085 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.254, + 0.049, + 0.79, + 0.087 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.162, + 0.867, + 0.183 + ], + "angle": 0, + "content": "Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.219, + 0.796, + 0.239 + ], + "angle": 0, + "content": "Rui Li\\(^{1,2,*}\\), Baopeng Zhang\\(^{1}\\), Jun Liu\\(^{2}\\), Wei Liu\\(^{1}\\), Jian Zhao\\(^{3,4,5}\\), Zhu Teng\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.246, + 0.865, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\)Beijing Jiaotong University \\(^{2}\\)Singapore University of Technology and Design \\(^{3}\\)Institute of North Electronic Equipment \\(^{4}\\)Peng Cheng Laboratory \\(^{5}\\)Intelligent Game and Decision Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.316, + 0.314, + 0.331 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.345, + 0.474, + 0.679 + ], + "angle": 0, + "content": "The existing one-stage multi-object tracking (MOT) algorithms have achieved satisfactory performance benefiting from a large amount of labeled data. However, acquiring plenty of laborious annotated frames is not practical in real applications. To reduce the cost of human annotations, we propose Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT), to infer the most informative frames for any MOT tracker by observing the heterogeneous cues of samples. HD-AMOT defines the diversified informative representation by encoding the geometric and semantic information, and formulates the frame inference strategy as a Markov decision process to learn an optimal sampling policy based on the designed informative representation. Specifically, HD-AMOT consists of a diversified informative representation module as well as an informative frame selection network. The former produces the signal characterizing the diversity and distribution of frames, and the latter receives the signal and conducts multi-frame cooperation to enable batch frame sampling. Extensive experiments conducted on the MOT15, MOT17, MOT20, and Dancetrack datasets demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that under \\(50\\%\\) budget our HD-AMOT can achieve similar or even higher performance as fully-supervised learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.702, + 0.21, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.725, + 0.47, + 0.837 + ], + "angle": 0, + "content": "As a sophisticated research topic, one-stage multi-object tracking (MOT) has attracted more research attention in recent years, which simultaneously focuses on pinpointing multiple objects in each frame and recording their trajectories in the entire sequence. Despite the great success of deep learning-based one-stage MOT methods [25, 48, 27, 26, 52, 6, 22], they need a large amount of annotated samples for training, which requires costly human labor. For example," + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.315, + 0.695, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.315, + 0.886, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.438, + 0.684, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.438, + 0.887, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.505, + 0.893, + 0.621 + ], + "angle": 0, + "content": "Figure 1. A stimulus example of active learning for one-stage MOT. There are many similar continuous frames in tracking sequences, such as frames 365-368 of MOT17-02 [23], and fully annotated these frames may have little effect on improving the performance of the tracker. Moreover, objects usually have different states, such as moving and static states, as shown in the red dotted box. The moving states usually contain more knowledge than the static ones, in which case the uniform and random sampling may fail to capture the diverse visual patterns of tracking sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.893, + 0.85 + ], + "angle": 0, + "content": "in the MOT17 dataset [23], a sequence containing only 600 frames requires annotating 18,581 bounding boxes (including coordinates and identity labels) for 62 pedestrians. In addition, there are many similar continuous frames in tracking sequences, such as frames 365-368 of the MOT17-02 sequence in Fig. 1, which may have little effect on boosting the performance of the tracker. One way to deal with the above issue is to randomly or uniformly sample tracking sequences for annotation. However, as shown in the red dotted box of Fig. 1, objects in tracking sequences may have different motion states, and the moving state contains more knowledge than the static state, in which case the uniform sampling may fail to capture diverse visual patterns. Moreover, random sampling is easy to yield unstable results. Therefore, it is essential to develop algorithms to infer training data more efficiently in one-stage MOT." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.852, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Instead of selecting the subset randomly or uniformly," + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.842, + 0.415, + 0.867 + ], + "angle": 0, + "content": "*Work done when Rui Li was an visiting student of SUTD. †Zhu Teng is the corresponding author (zteng@bjtu.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "9932" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.473, + 0.403 + ], + "angle": 0, + "content": "active learning (AL) aims to infer the most informative samples as training data to improve the knowledge of a model, which is a promising way to select training data more efficiently for one-stage MOT. In the computer vision area, active learning has been widely explored for image classification [3, 21, 1, 8], object detection [45, 44, 10], semantic segmentation [1, 7, 28, 9], and so on. In these research topics, the image selection strategies of uncertainty-based [32, 42] and distribution-based methods [31, 1] become the mainstream of active learning. However, these AL algorithms do not directly relate to the performance improvements of the specific task and usually lead to suboptimal performance. More notably, they are usually utilized to select independent image samples, which are prone to achieve poor coverage of the tracking sequence and produce redundant frames if they are employed to the one-stage MOT task stiffly and directly. In particular, since new individuals appear constantly in the tracking sequence, the distribution bias caused by the poor coverage and redundant frames may make the tracker unable to generalize well for unseen cases." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.406, + 0.473, + 0.544 + ], + "angle": 0, + "content": "Furthermore, different from other tasks that distinguish inter-class objects in each image (such as image classification, object detection, etc.), one-stage MOT needs to detect and associate intra-class objects, e.g., pedestrians belonging to different individuals. This implies that the active learning for one-stage MOT is more challenging, because it should consider not only the discrepancy between frame-level images, but also the influence of more fine-grained intra-class diversity in sample selection, which plays a vital role in object association in one-stage MOT." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.473, + 0.822 + ], + "angle": 0, + "content": "To address the above issues, we formulate the active learning of one-stage MOT as a Markov Decision Process (MDP) [36] and design a novel Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT) framework, which leverages a Diversified Informative Representation Module (DIRM) and an Informative Frame Selection Network (IFSN) to learn a frame sampling policy. Specifically, IFSN is designed to infer a batch of informative and diverse frames for annotation by receiving the signal characterizing the diversity and distribution of frames and conducting multi-frame cooperation, which can effectively inhibit poor coverage and reduce redundant frames. In addition, to facilitate the inference of benignant frames, we develop a novel DIRM to characterize the frame-level diversified representation of samples by encoding the geometric and semantic cues of frames and objects. Besides the frame-level diversity representation, we also evaluate the set-level (labeled set and unlabeled set) discrepancy in DIRM, which dedicates to the acquirement of the training set with the unbiased distribution." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.472, + 0.868 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows: (1) We propose a novel HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process (MDP) and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.895, + 0.308 + ], + "angle": 0, + "content": "learns a sampling policy driven by the reward that directly relates to the tracker performance. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning. (2) To infer the benignant frames, we design a diversified informative representation to characterize the diversity and distribution of frames, which learns both the set-level discrepancy and frame-level diversity by encoding the heterogeneous cues of geometry and semantics. (3) Extensive experiments are conducted on four datasets to demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that our HD-AMOT can achieve similar or even higher performance as fully-supervised learning under \\(50\\%\\) annotation budget." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.319, + 0.642, + 0.333 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Multi-Object Tracking. MOT is a challenging task due to the complexity of tracking scenes. The existing works mainly follow the tracking-by-detection paradigm and two-stage methods are the mainstream, which regards the detection and tracking as two independent models [4, 14, 46, 40, 11, 17]. Despite noticeable progress, the two-stage methods cause a waste of resources and maybe lead to an efficiency issue because the independent detection and tracking models are executed separately in these methods. Recently, with the quick maturity of deep learning[50, 49, 51], one-stage frameworks for MOT [25, 48, 27, 26, 53, 52, 6, 22, 47] have begun to attract more research attention, which greatly improve the accuracy and efficiency of tracking. The fundamental concept revolves around the simultaneous integration of object detection and association within a singular network architecture. For example, CTracker [27] employs a chaining mechanism that connects paired bounding-box regression outcomes derived from overlapping nodes. Each of these nodes spans two consecutive frames. FairMOT [48] presents a simple yet effective approach based on the detection architecture CenterNet [13]. Despite the notable advancements introduced by these methods, their practical applicability remains constrained by the demanding nature of high-cost annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.896, + 0.825 + ], + "angle": 0, + "content": "To reduce the demand for labeled data, several methods with less supervision signal such as weakly-supervised and self-supervised learning [30, 2] can effectively utilize the unlabeled data to improve the MOT performance. However, these methods pay more attention to the tracker learning, and ignore the importance of labeled data selection to improve the tracker performance. In this paper, we design an active learning approach for one-stage MOT, which could significantly reduce the annotation cost. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Active Learning. Active learning is an important machine learning problem, which has received lots of attention in computer vision, such as image classification" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "9933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.125, + 0.891, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.893, + 0.398 + ], + "angle": 0, + "content": "Figure 2. The pipeline of our proposed HD-AMOT framework. In our method, we perform several iterations until a budget \\(\\mathcal{B}\\) of labeled frames is achieved. In the \\(t^{th}\\) iteration, the steps are as follows: (I) A one-stage MOT tracker \\(\\pmb{f}_t\\) is used to obtain the heterogeneous cues including semantic feature \\(\\pmb{F}^f\\) and geometric information \\(\\pmb{G}^h\\), \\(\\pmb{G}^b\\) on each frame; (II) The set-level discrepancy and frame-level diversity are estimated and construct the MDP state-action pair \\((s_t, a_t)\\); (III) The multi-frame cooperation based on \\((s_t, a_t)\\) is performed; (IV) With the assistance of multi-frame cooperation, IFSN infers \\(N\\) unlabeled frames with high scores to be annotated. Meanwhile, \\(\\pmb{X}_t^U\\), \\(\\pmb{X}_t^L\\) are updated to \\(\\pmb{X}_{t+1}^U\\), \\(\\pmb{X}_{t+1}^L\\) by displacing newly annotated frames from \\(\\pmb{X}_t^U\\) to \\(\\pmb{X}_t^L\\); (V) The tracker \\(\\pmb{f}_t\\) is retrained on \\(\\pmb{X}_{t+1}^L\\) to obtain the updated \\(\\pmb{f}_{t+1}\\). (VI) The reward \\(r_{t+1}\\) based on \\(\\pmb{f}_t\\) and \\(\\pmb{f}_{t+1}\\) is computed on \\(\\pmb{X}^R\\) to update parameters of IFSN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.473, + 0.692 + ], + "angle": 0, + "content": "[1, 8, 39, 20], object detection [45, 44, 10], semantic segmentation [7, 28, 33], and so on. These methods can be divided into uncertainty-based [32, 42] and distribution-based methods [31, 1]. For example, [15] exploits the uncertainty in both the input and output spaces to select the most valuable information. [31] defines the problem of active learning as core-set selection. [32] develops a foundation LearningLoss++ for the active learning of pose estimation to establish equivalency between Learning Loss [42] empirically driven objective and the KL divergence objective. These AL algorithms are used to select independent image samples for a specific task yet are not suitable for the video-wise one-stage MOT task. It is prone to cause the problems such as poor coverage and redundant frames if we directly apply the existing AL approaches to one-stage MOT stiffly. Accordingly, we design an active learning framework HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process and learns a batch frame sampling policy driven by the reward that directly relates to the performance of the MOT tracker, which gives consideration to both sampling coverage and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.868 + ], + "angle": 0, + "content": "MDP of Multi-Object Tracking. As a sequential decision model, Markov decision process is usually applied in dynamic environments where an agent needs to perform certain tasks by making decisions and executing actions sequentially. In multi-object tracking, several MDP-based works have been explored [41, 29, 16]. [41] formulates the online two-stage MOT problem as decision making in MDPs, where the lifetime of an object is represented using an MDP model. [29] treats individual objects as agents, utilizing a prediction network for tracking, while optimizing tracking outcomes through collaborative interactions among various agents and their environments, facilitated by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.893, + 0.513 + ], + "angle": 0, + "content": "the decision network. In this paper, different from these methods that employ the MDP to directly solve the detection and tracking subtask of MOT, we address the annotation budget for one-stage MOT based on active learning by the designed Markov decision process, and design a diversified informative representation with an informative frame selection network to enable effective batch sampling under a specific annotation budget." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.744, + 0.544 + ], + "angle": 0, + "content": "3. Our Proposed HD-AMOT" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.552, + 0.611, + 0.566 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.741 + ], + "angle": 0, + "content": "In this paper, given an unlabeled video sequence \\(X\\) with a limited annotation budget \\(\\mathcal{B}\\), our HD-AMOT model aims to learn an optimal sampling strategy that infers and annotates the most informative frames iteratively to maximize the performance of the tracker \\(f\\). Specifically, we cast this AL problem of one-stage MOT by an MDP and adopt the Q-learning algorithm [24] to solve this problem, where we introduce an IFSN to infer frames according to the designed state-action representation \\((s_t,a_t)\\). As shown in Fig. 2, in our method, the video sequence \\(X\\) is divided into three different subsets \\(X^U,X^L\\), and \\(X^R\\). And at each iteration \\(t\\), the following steps are executed:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.894, + 0.868 + ], + "angle": 0, + "content": "(I) The heterogeneous cues including semantic features \\(\\pmb{F}^{f}\\) and geometric information \\(\\pmb{G}^{h}\\), \\(\\pmb{G}^{b}\\) of each frame in the labeled set \\(\\pmb{X}_{t}^{L}\\) and unlabeled set \\(\\pmb{X}_{t}^{U}\\) are obtained by \\(\\pmb{f}_{t}\\); (II) We estimate the set-level discrepancy and frame-level diversity in DIRM based on the obtained heterogeneous cues, which are regarded as the MDP state-action pair \\((s_{t}, a_{t})\\); (III) The multi-frame cooperation based on \\((s_{t}, a_{t})\\) is performed on \\(\\pmb{X}_{t}^{L}\\) and \\(\\pmb{X}_{t}^{U}\\); (IV) With the assistance of multi-frame cooperation, IFSN selects \\(N\\) unlabeled frames" + }, + { + "type": "page_number", + "bbox": [ + 0.468, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "9934" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.125, + 0.465, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.217, + 0.468, + 0.256 + ], + "angle": 0, + "content": "Figure 3. Illustration of geometric matrices construction process. (a): binary heatmap; (b): global topology matrix \\( G^{h} \\); (c): local topology matrices \\( G^{h_1} \\sim G^{h_K} \\); (d): object scale matrix \\( G^b \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.26, + 0.468, + 0.362 + ], + "angle": 0, + "content": "with high scores to be annotated guided by frame clusters. Meanwhile, the subsets \\(X_{t}^{U}\\), \\(X_{t}^{L}\\) are updated to \\(X_{t+1}^{U}\\), \\(X_{t+1}^{L}\\) by moving newly annotated frames from \\(X_{t}^{U}\\) to \\(X_{t}^{L}\\); (V) The tracker \\(\\pmb{f}_{t}\\) is retrained on \\(X_{t+1}^{L}\\) to obtain the updated \\(\\pmb{f}_{t+1}\\); (VI) The reward \\(r_{t+1}\\) based on \\(\\pmb{f}_{t}\\) and \\(\\pmb{f}_{t+1}\\) is computed on the reward subset \\(\\pmb{X}^{R}\\), which is used to update parameters of IFSN." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.373, + 0.411, + 0.388 + ], + "angle": 0, + "content": "3.2. Diversified Informative Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.468, + 0.533 + ], + "angle": 0, + "content": "To infer the most informative and representative frames, we propose a DIRM to compute the state and action of the designed MDP, which characterizes the diversity and distribution of frames. Specifically, to compensate for the distribution drift between the unlabeled set \\( \\mathbf{X}_t^U \\) and the labeled set \\( \\mathbf{X}_t^L \\), we use heterogeneous cues to evaluate the set-level discrepancy as state \\( s_t \\). Moreover, from the perspective of selecting the unlabeled frame with the potential contribution to the MOT tracker training, we propose to capture the frame-level diversity of each unlabeled frame as action \\( a_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.468, + 0.672 + ], + "angle": 0, + "content": "Set-level Discrepancy. There are a large number of objects belonging to different individuals in the tracking sequence, and new individuals may appear constantly over time. To make the trained tracker well generalized to unseen individuals, it is necessary to generate an unbiased training set that captures the whole sequence distribution. Based on this intuition, we propose to narrow the distribution gap between the unlabeled and labeled sets by evaluating set-level discrepancy, to ensure the diversity and representativeness of the selected frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.867 + ], + "angle": 0, + "content": "Specifically, we consider two key attributes to characterize the distribution drifts between the labeled and unlabeled sets: semantic variation and spatial topological variation. We collect the global feature from the last feature extraction layer of \\( \\pmb{f}_t \\) to obtain the semantic feature \\( F^f \\), which depicts the general semantic information of the frame. For the spatial topology \\( G^h \\), we encode the center of each object obtained by the tracker \\( f_t \\) into a binary heatmap to represent the spatial distribution of objects. Considering that identities of objects in the frames with similar spatial topology may be quite different due to the object motion, we encode the identities into the binary heatmap to obtain the final \\( G^h \\), as shown in Fig.3 (a) and (b). In addition, the performance of the MOT tracker is affected by objects or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.321 + ], + "angle": 0, + "content": "background, leading to different tracking quality over various local spatial areas. To suppress adverse effects of the low-quality local areas, we decompose the global spatial topology \\( \\pmb{G}^{h} \\) into \\( K \\) local parts, which form our final spatial topology \\( \\{\\pmb{G}^{h},\\pmb{G}^{h_{1} - h_{K}}\\} \\) together with the global topology. Similarly, the global and local semantic features constitute the final semantics \\( \\{\\pmb{F}^{f},\\pmb{F}^{f_{1} - f_{K}}\\} \\). In this way, the feature space \\( S_{set} = \\{\\pmb{G}^{h},\\pmb{G}^{h_{1} - h_{K}},\\pmb{F}^{f},\\pmb{F}^{f_{1} - f_{K}}\\} \\) is formed. To model the set-level discrepancy between \\( X_{t}^{L} \\) and \\( X_{t}^{U} \\), Maximum Mean Discrepancy [37] is adopted to calculate the set gap for each feature \\( S_{set}^{m} \\) in the feature space \\( S_{set} \\) as described in Eq. (1). Finally, the state representation \\( s_t \\) is defined as the concatenation of all \\( \\mathcal{D}^m \\) to encode the distribution drifts between \\( X_{t}^{L} \\) and \\( X_{t}^{U} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.339, + 0.892, + 0.402 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} ^ {m} = \\frac {1}{n _ {l} ^ {2}} \\sum_ {i, j} \\varphi \\left(p _ {i} ^ {m}, p _ {j} ^ {m}\\right) + \\frac {1}{n _ {u} ^ {2}} \\sum_ {i, j} \\varphi \\left(q _ {i} ^ {m}, q _ {j} ^ {m}\\right) \\tag {1} \\\\ - \\frac {2}{n _ {l} n _ {u}} \\sum_ {i, j} \\varphi \\left(p _ {i} ^ {m}, q _ {j} ^ {m}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.479 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_m\\) is a scalar representing the distribution discrepancy between \\(X_{t}^{L}\\) and \\(X_{t}^{U}\\) on \\(S_{\\text{set}}^{m}\\). \\(p^m\\) and \\(q^m\\) denote the corresponding features of frames in \\(X_{t}^{L}\\) and \\(X_{t}^{U}\\), \\(n_l\\) and \\(n_u\\) are numbers of frames in \\(X_{t}^{L}\\) and \\(X_{t}^{U}\\). \\(\\varphi(.)\\) is the radial kernel [37] to measure the distance between two features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Frame-level Diversity. In this paper, we aim to select the frames that are beneficial to the tracker training for annotation. For this purpose, besides considering the set-level discrepancy, we also need to explore frame diversity from different perspectives to infer the most representative and informative frames. Different from the semantic feature and spatial topology used to learn set-level discrepancy, finer-grained cues are desiderated to evaluate the potential contribution of each unlabeled frame. Specifically, besides the frame-wise spatial topology \\( G^{h} \\) and semantic feature \\( F^{f} \\), we also consider finer-grained attribute characterizing the individuation of objects: the scale of bounding boxes \\( G^{b} \\). Its construction process is similar to \\( G^{h} \\), except that the identities representing different individuals are replaced by the width and height of the corresponding object as shown in Fig.3 (d), i.e., \\( G^{h} \\in \\mathbb{R}^{H \\times W \\times 1} \\) and \\( G^{b} \\in \\mathbb{R}^{H \\times W \\times 2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Intuitively, we argue that any one of the geometric features \\( G^{h} \\), \\( G^{b} \\) and semantic features \\( F^{f} \\) is the key factor to determine whether a frame contributes to the performance improvement of the tracker. Accordingly, we consider the novelty and prominence of each unlabeled frame \\( x^{u} \\) based on \\( S_{frame} = \\{G^{h}, G^{b}, F^{f}\\} \\), to form the frame-level diversity as action \\( a_{t} \\). The novelty of an unlabeled frame \\( x_{i}^{u} \\) in \\( X_{t}^{U} \\) is reflected in the similarities between \\( x_{i}^{u} \\) and the labeled frames in \\( X_{t}^{L} \\), because the unlabeled frame with a novel characteristic (such as new individuals, different object scales, and postures) usually has low similarities to labeled frames of \\( X_{t}^{L} \\). And the promi" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "9935" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.473, + 0.352 + ], + "angle": 0, + "content": "nence of \\( x_{i}^{u} \\) lies in the difference between \\( x_{i}^{u} \\) and other unlabeled frames in \\( X_{t}^{U} \\), so we expect to select the most representative frame by recording the similarity distribution between each unlabeled frame and other frames in \\( X_{t}^{U} \\). Specifically, we introduce histogram-based representation to learn the novelty and prominence of each unlabeled frame. Taking \\( x_{i}^{u} \\) as an example, to represent the novelty of \\( x_{i}^{u} \\), we calculate cosine similarities between \\( x_{i}^{u} \\) and \\( X_{t}^{L} \\) on each heterogeneous feature \\( S_{frame}^{m} \\) and then obtain the histogram \\( \\mathcal{H}_{nov} \\) based on cosine similarities. Similar to novelty learning, we learn prominence by recoding the histogram \\( \\mathcal{H}_{pro} \\) based on cosine similarities between \\( x_{i}^{u} \\) and \\( X_{t}^{U} \\). Finally, the action representation \\( a_{t} \\) of \\( x_{i}^{u} \\) is defined as \\( \\{\\mathcal{H}_{nov}^{G^{h}},\\mathcal{H}_{nov}^{G^{b}},\\mathcal{H}_{nov}^{F^{f}},\\mathcal{H}_{pro}^{G^{h}},\\mathcal{H}_{pro}^{G^{b}},\\mathcal{H}_{pro}^{F^{f}}\\} \\), which conduces to IFSN to effectively identify and select the informative and representative frames." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.358, + 0.411, + 0.373 + ], + "angle": 0, + "content": "3.3. Informative Frame Selection Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Inspired by DQN [38], our lightweight IFSN follows an optimal policy that maximizes the potential performance gains. Considering the sampling efficiency, we propose a batch sampling method guided by frame clusters according to temporal clues. Meanwhile, to restrain the inferior performance caused by redundant frames, we further introduce the multi-frame cooperation strategy before batch sampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.627 + ], + "angle": 0, + "content": "Multi-frame Cooperation. There are many similar adjacent frames in tracking sequences, which easily leads to redundancy in the process of frame selection. To reduce redundant frames, we introduce a simple multi-frame cooperation strategy to model the neighborhood cooperation between the unlabeled frame \\( x_{i}^{u} \\) and its nearest labeled frames. Specifically, we seek the nearest neighbors \\( x_{v}^{l} \\) and \\( x_{w}^{l} \\) of \\( x_{i}^{u} \\) in \\( X_{t}^{L} \\) and calculate their actions \\( a_{t}^{v} \\), \\( a_{t}^{w} \\). Then we use the fixed-length compact representation \\( o_{t} = \\frac{1}{2}(a_{t}^{v} + a_{t}^{w}) \\) as an extra action for \\( x_{i}^{u} \\). To this end, the selection made by IFSN becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.634, + 0.47, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\phi_ {t} = \\arg \\max _ {a _ {t} \\in A _ {t}} g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.688 + ], + "angle": 0, + "content": "where \\(a_{t}\\in A_{t}\\) denotes the candidate action in \\(X_{t}^{U}\\), and \\(g(.)\\) is the lightweight IFSN network parameterized by \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.785 + ], + "angle": 0, + "content": "Batch Frame Selection. The structure of our designed lightweight IFSN consists of four linear layers, where a linear layer parameterized by \\(\\theta_{1}\\) is used to model the relation between each unlabeled frame and its nearest neighbors, and three linear layers parameterized by \\(\\theta_{2\\sim 4}\\) are employed to score each unlabeled frame, as described in Eq. (3), where \\(cat[\\cdot]\\) denotes the concatenation operation." + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.791, + 0.47, + 0.807 + ], + "angle": 0, + "content": "\\[\ng \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right) = \\theta_ {2 \\sim 4} \\left(c a t \\left[ \\theta_ {1} \\left(a _ {t}, o _ {t}\\right), a _ {t}, s _ {t} \\right]\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Intuitively, each unlabeled frame in \\( X_{t}^{U} \\) can obtain the corresponding score by IFSN. However, sampling a single frame in each iteration by Eq. (2) to query annotation is inefficient. In this paper, we propose a batch sampling method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.252 + ], + "angle": 0, + "content": "guided by frame clusters to sample \\(N\\) frames in each iteration. Based on the intuition that a larger difference exists between frames with longer time intervals, we divide the unlabeled frames into \\(N\\) clusters in chronological order and select the frame for annotation in each cluster independently, which is equivalent to labeling a frame in parallel with \\(N\\) annotators. In this case, \\(\\phi_t\\) in Eq. (2) is extended to consist of \\(N\\) independent sub-selections \\(\\{\\theta_t^n\\}_{n=1}^N\\), each with a restricted space \\(A_t^n\\), as described in Eq. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.26, + 0.892, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\phi_ {t} ^ {n} = \\arg \\max _ {a _ {t} \\in A _ {t} ^ {n}} g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.291, + 0.888, + 0.305 + ], + "angle": 0, + "content": "where \\(a_{t}\\in A_{t}^{n}\\) denotes the candidate action in \\(n^{th}\\) cluster." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.458 + ], + "angle": 0, + "content": "IFSN Optimization. In our method, we compute the reward on a subset \\( X^R \\), which is used to evaluate the benefit of selected frames to the tracker and optimize the frame selection network IFSN. To evaluate the performance of the tracker more comprehensively, we utilize MOTA [5] and IDF1 [19] as performance metrics to calculate the reward, where MOTA is a comprehensive metric to evaluate the detection performance and IDF1 tends to estimate the tracking ability of a tracker. As described in Eq. (5), the reward \\( r_{t+1} \\) is defined as the difference of the performance metrics between \\( f_{t+1} \\) and \\( f_t \\)." + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.465, + 0.892, + 0.482 + ], + "angle": 0, + "content": "\\[\nr _ {t} = \\left(e _ {t + 1} ^ {\\text {m o t a}} + e _ {t + 1} ^ {i d f 1}\\right) - \\left(e _ {t} ^ {\\text {m o t a}} + e _ {t} ^ {i d f 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.891, + 0.523 + ], + "angle": 0, + "content": "where \\( e_{t+1}^{mota} \\) and \\( e_{t+1}^{idf1} \\) are MOTA and IDF1 metrics computed by \\( f_{t+1} \\), \\( e_t^{mota} \\) and \\( e_t^{idf1} \\) are evaluated by \\( f_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.606 + ], + "angle": 0, + "content": "With the reward \\( r_{t + 1} \\), we can optimize IFSN to infer the most informative frames to maximize the reward, leading to improved tracker performance during each active learning iteration. More specifically, we train IFSN in a double DQN formulation [38] by optimizing the loss based on the temporal difference error [35] as:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.614, + 0.892, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I F S N}} = \\left(\\hat {y} _ {t} - g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right)\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.653, + 0.892, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t} = r _ {t + 1} + \\gamma g \\left(s _ {t + 1}, a _ {t + 1}, o _ {t + 1}; \\hat {\\theta}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.891, + 0.712 + ], + "angle": 0, + "content": "where \\(\\theta\\) and \\(\\hat{\\theta}\\) are parameters of the IFSN policy network and off-policy network following the setting of double DQN [38]. \\(\\gamma\\) denotes a discount factor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.768 + ], + "angle": 0, + "content": "With the described diversified informative representation in Section 3.2 and informative frame selection learning in this section, we introduce our entire HD-AMOT as illustrated in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.781, + 0.625, + 0.797 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.804, + 0.707, + 0.819 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Datasets. We evaluate our HD-AMOT on MOT15 [18], MOT17 [23], MOT20 [12], and Dancetrack [34]. MOT15 and MOT17 are widely-used MOT datasets, and they are" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "9936" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.089, + 0.13, + 0.338, + 0.143 + ], + "angle": 0, + "content": "Algorithm 1: HD-AMOT Algorithm." + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.145, + 0.437, + 0.4 + ], + "angle": 0, + "content": "Input: The sequence \\(X\\), initial tracker \\(f_{init}\\) and IFSN \\(g_{init}\\), AL procedure rounds \\(Z\\). \n1 \\(\\mathbf{X}_{init}, \\mathbf{X}^{R} \\gets\\) RandomPartition \\((X)\\); \n2 \\(\\mathbf{X}_{init}^{U}, \\mathbf{X}_{init}^{L} \\gets\\) UniformPartition \\((X_{init})\\); \n3 \\(\\mathbf{X}_{0}^{L} \\gets \\mathbf{X}_{init}^{L}, f_{0} \\gets\\) update \\((g_{init}, \\mathbf{X}_{init}^{L})\\); \n4 while not done do // Episodes \n5 for \\(t = 0\\) to \\(t = Z - 1\\) do // AL procedure \n6 Learn the diversified representation to form the state and action space; \n7 Use IFSN to select frames by Eq. (4); \n8 Annotate the selected \\(N\\) frames \\(\\mathbf{X}_{t}^{S}\\); \n9 Update \\(\\mathbf{X}_{t}^{U}, \\mathbf{X}_{t}^{L}\\): \n10 \\(\\mathbf{X}_{t + 1}^{U} \\gets \\mathbf{X}_{t}^{U} \\setminus \\mathbf{X}_{t}^{S}, \\mathbf{X}_{t + 1}^{L} \\gets \\mathbf{X}_{t}^{L} \\cup \\mathbf{X}_{t}^{S}\\); \n11 \\(f_{t + 1} \\gets\\) update \\((f_{t}, \\mathbf{X}_{t + 1}^{L})\\); \n12 Calculate the reward on \\(\\mathbf{X}^{R}\\) by Eq. (5); \n13 end \n14 Update IFSN following Eq. (6); \n15 Reset: \\(\\mathbf{X}_{0}^{L} \\gets \\mathbf{X}_{init}^{L}, f_{0} \\gets f_{t + 1}\\); \n16 end" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.469, + 0.531 + ], + "angle": 0, + "content": "challenging because of the large variety in object pose, lighting, viewpoint, etc. In contrast to MOT15 and MOT17, MOT20 addresses the challenge of highly congested scenarios, where the pedestrian density can surge to as high as 246 individuals per frame. Dancetrack is a large-scale MOT dataset covering scenarios with low distinguishability of object appearance and diverse non-linear motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.469, + 0.644 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the performance of one-stage MOT, the standard metrics MOTA [5] and IDF1 [19] are adopted for evaluation, which indicate the detection and association performance of one-stage MOT. Specifically, MOTA combines three error sources including false positives, missed targets, and identity switches, and IDF1 is the ratio of correctly identified detections over the average number of ground-truth and computed detections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Implementation Details. In this paper, we adopt the popular one-stage MOT method FairMOT [48] as the MOT tracker. Note that the propose algorithm is suitable for any one-stage MOT tracker. The input frames are resized to \\(1088 \\times 608\\). We set the learning rate of our IFSN to 1e-4 and the discount factor \\(\\gamma\\) in Eq. (7) to 0.9, and the number of local features \\(K\\) is set to 9 according to experiments. For each dataset, we randomly select a small sequence in each video at a ratio of 0.1 to form a reward subset \\(X^R\\). In the remaining frames, a small amount of frames is uniformly sampled as the initially labeled subset \\(X_{init}^{L}\\) to initialize the MOT tracker \\(f\\), and the size is set to 100. For the semantic information, we extract the feature map from the last convolutional layer of ResNet34 as the global semantic feature \\(F^{f} \\in \\mathbb{R}^{512 \\times 19 \\times 34}\\). For the geometric clues, we utilize the center coordinates and identities of bounding" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.123, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Table 1. Ablation study on the design of the diversified informative representation." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.297 + ], + "angle": 0, + "content": "
RepresentationSettingMOTA (%)IDF1 (%)
Set-level DiscrepancyState w/o Gh,h1-h961.7364.95
State w/o Ff,f1-f961.4964.12
State w/o Gh, Ff62.7465.21
State w/o Gh1-h9, Ff1-f962.9264.55
Frame-level DiversityAction w/o Hgh62.8366.61
Action w/o HGb62.4966.23
Action w/o HFf61.7864.95
Action w/o H* nov61.2563.41
Action w/o H* pro62.2864.16
AllHD-AMOT63.2366.47
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.397 + ], + "angle": 0, + "content": "boxes obtained by the MOT tracker to construct the global spatial topology \\( G^{h} \\in \\mathbb{R}^{152 \\times 272} \\), and generate the object scale matrix \\( G^{b} \\in \\mathbb{R}^{152 \\times 272 \\times 2} \\) by the box size. In addition, the number of clusters \\( N \\) in the batch sampling is set to 12 for MOT15, MOT17, and MOT20, and is set to 24 for Dancetrack as it is much larger than other datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.653, + 0.423 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.498 + ], + "angle": 0, + "content": "In this section, to analyze the impacts of different components, we conduct ablation studies on the popular dataset MOT17 following the division strategy in [43]. We run our experiments 5 times and report the mean performance for all ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.893, + 0.823 + ], + "angle": 0, + "content": "Effect of Diversified Informative Representation. We perform an ablation study with a \\(20\\%\\) annotation budget to evaluate the contribution of our diversified informative representation. As IFSN relies on the state (set-level discrepancy) to decide the frame sampling policy, we investigate the influence of the set-level discrepancy by removing different components individually. Moreover, we also analyze the effect of the action design (frame-level diversity). As shown in Tab. 1, our complete diversified informative representation gives the best performance in terms of MOTA and is only slightly inferior to Action w/o \\(\\mathcal{H}^{G^h}\\) by \\(0.14\\%\\) on IDF1. We speculate that the scale information in \\(\\mathcal{H}^{G^b}\\) implicitly indicates the identity and topology clues of \\(\\mathcal{H}^{G^h}\\), so the introduction of \\(\\mathcal{H}^{G^b}\\) weakens the role of \\(\\mathcal{H}^{G^h}\\), making the MOTA of Action w/o \\(\\mathcal{H}^{G^h}\\) close to HD-AMOT and the IDF1 slightly higher. Furthermore, removing the components of set-level discrepancy or frame-level diversity in the diversified informative representation degrades the tracking performance. The largest decrease of MOTA and IDF1 occurs when the score of maximum similarity with labeled frames Action w/o \\(\\mathcal{H}_{nov}^*\\) is removed, which further verifies the effectiveness of heterogeneous cues evaluating the novel characteristic of unlabeled frames." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Effect of Informative Frame Selection Network. We further validate the performance of our designed IFSN in Tab. 2. We first consider performing the cluster-based batch" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.501, + 0.898 + ], + "angle": 0, + "content": "9937" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.123, + 0.468, + 0.146 + ], + "angle": 0, + "content": "Table 2. Ablation study on the different settings of the Informative Frame Selection Network." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.147, + 0.46, + 0.226 + ], + "angle": 0, + "content": "
ModelMulti-frame CooperationBatch SamplingMOTA (%)↑IDF1 (%)↑
Bât.Clu.
A×✓✓62.7665.52
B✓××63.0765.62
C✓✓×62.1864.55
D✓✓✓63.2366.47
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.231, + 0.273, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.232, + 0.464, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.404, + 0.468, + 0.43 + ], + "angle": 0, + "content": "Figure 4. Performance comparison evaluated by MOTA and IDF1 metrics with increasing budgets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.469, + 0.755 + ], + "angle": 0, + "content": "sampling without multi-frame cooperation (Model A). Then we analyze the usage of selecting a single frame in each AL iteration (Model B). In addition, we also construct the Model C as IFSN to select a batch of frames in one shot, where the frames with \\( N \\) highest scores are sampled. Finally, we present the performance of our cluster-based batch sampling strategy (Model D). By comparing the results of Model A, Model D brings performance gains of \\( 0.47\\% \\) and \\( 0.95\\% \\) in terms of MOTA and IDF1 metrics, which proves that multi-frame cooperation plays a positive role in the active learning of one-stage MOT. In contrast, Model C gives the worst performance, which selects multiple frames according to the \\( N \\) highest scores. We argue that frames with higher scores may be similar samples and close to each other, which easily leads to inefficiency in batch sampling. To this end, cluster-based batch sampling is introduced, as it can avoid selecting adjacent frames with higher scores and learn to sample with better coverage of the underlying distribution. Notably, selecting a single frame at each iteration in Model B also provides a competitive performance but still tends to be inferior to our method Model D, and the time cost of selecting a single frame per iteration is much higher than our cluster-based batch sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Analysis of Different Budgets. We report the performance of our method, random sampling, uniform sampling, and other existing AL methods such as Coreset [31], CDAL [1] under different budgets, as shown in Fig. 4. For the IDF1 metric, ours is slightly inferior to the uniform sampling when the budget is \\(20\\%\\) and \\(40\\%\\). We argue that uniform sampling can select informative frames for annotation if a small number of labeled frames is required. However," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.28 + ], + "angle": 0, + "content": "when the budget increases, there may be redundancy in the newly acquired frames by uniform sampling, and training with these labels does not provide more additional information, resulting in the stagnation or even decline of the performance of IDF1. Notably, our HD-AMOT with the \\(50\\%\\) budget achieves similar performance as other methods with the \\(80\\%\\) budget. It reveals that ours can save around \\(30\\%\\) labeling efforts on average compared to other sampling methods. We attribute this to our learned diversified informative clues and fine-grained object representation, which can effectively improve the tracking performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.29, + 0.816, + 0.305 + ], + "angle": 0, + "content": "4.3. Comparison with other AL Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.891, + 0.352 + ], + "angle": 0, + "content": "To further verify the generalization and effectiveness, we evaluate the proposed framework HD-AMOT on four diversified MOT datasets in this section." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Results on MOT15. Tab. 3 lists the mean values of MOTA and IDF1 on MOT15, and Fig. 5 reveals the corresponding standard deviation of MOTA and IDF1. We can observe that HD-AMOT outperforms all methods by a clear margin, i.e., AL methods used in other tasks (CDAL [1] and Coreset [31]) stiffly applied to one-stage MOT can lead to obvious performance degradation. Notably, although Coreset achieves the second-best performance on MOTA and IDF1, it is clear from the standard deviations in Fig. 5 that Coreset has limited stability. Moreover, the performance of Coreset on other datasets (MOT17, MOT20, and Dancetrack) is not ideal, which indicates that the direct application of Coreset in one-stage MOT is unstable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.755 + ], + "angle": 0, + "content": "Results on MOT17. MOT17 is a challenging dataset with more crowded scenarios, different viewpoints, camera motions, and weather conditions. As shown in Tab. 3, although uniform sampling achieves the best IDF1 on MOT17 and our HD-AMOT performs slightly lower than it by a margin of \\(0.18\\%\\), uniform sampling may fail to capture diverse visual patterns in other datasets due to the lack of flexibility and adaptability. For example, in MOT15, ours outperforms uniform sampling by all performance metrics, which reflects the generalization ability of ours for one-stage MOT. Moreover, besides object association, the object detection ability should be considered in one-stage MOT. Our method, which benefits from learning the sampling policy driven by tracking performance metrics directly from data, significantly outperforms uniform sampling by \\(0.84\\%\\) in terms of the detection accuracy metric MOTA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Results on MOT20. We also use the large-scale dataset MOT20 to demonstrate the scalability of our HD-AMOT as shown in Tab. 3. In contrast to the aforementioned datasets, MOT20 is a recently released dataset that features more densely populated scenes. We can observe that ours still achieves the best performance, which demonstrates that our HD-AMOT can infer informative frames even on a crowded and complicated dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "9938" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.122, + 0.47, + 0.174 + ], + "angle": 0, + "content": "Table 3. Mean values of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack. Best and second-best results are shown in red and black, respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.175, + 0.47, + 0.429 + ], + "angle": 0, + "content": "
DatasetMethodMOTA(%) (↑)IDF1(%) (↑)
MOT15Tracker w/ Random53.2859.61
Tracker w/ Uniform54.0959.50
Tracker w/ CDAL [1]53.1358.85
Tracker w/ Coreset [31]54.7060.29
Tracker w/ HD-AMOT55.2960.64
MOT17Tracker w/ Random70.1169.24
Tracker w/ Uniform70.4070.05
Tracker w/ CDAL [1]69.1068.71
Tracker w/ Coreset [31]69.4468.61
Tracker w/ HD-AMOT71.2469.87
MOT20Tracker w/ Random58.2165.37
Tracker w/ Uniform58.6565.87
Tracker w/ CDAL [1]57.5965.33
Tracker w/ Coreset [31]56.6664.86
Tracker w/ HD-AMOT59.2466.41
DancetrackTracker w/ Random80.4940.67
Tracker w/ Uniform81.1141.03
Tracker w/ CDAL [1]81.3641.16
Tracker w/ Coreset [31]81.1340.47
Tracker w/ HD-AMOT81.9441.93
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.436, + 0.275, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.436, + 0.468, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Figure 5. Standard deviations of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Results on Dancetrack. Dancetrack is a larger and more challenging dataset for group dancing scenes, where humans have similar appearance, diverse motion, and extreme articulation. We provide the performance evaluation results on the large-scale Dancetrack in Tab. 3. It can be observed that our HD-AMOT still achieves the best MOTA and IDF1 performance. Notably, from the perspective of sampling stability in Fig. 5, although uniform sampling can obtain a low standard deviation of MOTA, it leads to a very high standard deviation of IDF1 due to failure to capture informative frames that are helpful to object association. By contrast, our proposed HD-AMOT can maintain stable MOTA and IDF1 performance in multiple samplings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Furthermore, we give the performance comparison between our HD-AMOT with a \\(50\\%\\) budget and full supervision in Fig. 6. Our active learning framework achieves sim" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.125, + 0.703, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.125, + 0.889, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Figure 6. Performance comparison between our HD-AMOT with a \\(50\\%\\) budget and full supervision on MOT15, MOT17, MOT20, and Dancetrack datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.509 + ], + "angle": 0, + "content": "ilar or even higher performance compared with the tracker with fully-supervised training. For instance, we exceed the full supervision by a gain of \\(2\\%\\) on MOTA and \\(0.8\\%\\) on IDF1 for MOT15. These results implicitly prove that there are many similar frames in MOT sequences, and these similar frames bring limited progress in tracking performance. Our proposed HD-AMOT learns an optimal sampling policy by the heterogeneous diversity representation and a Markov decision process, which can infer informative frames that can most benefit the tracker, thus effectively improving the performance of the tracker. In summary, the proposed method can provide a consistent gain for the MOT tracker across datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.619, + 0.539 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.727 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework HD-AMOT to investigate the annotation budget for one-stage MOT. It is formulated as an MDP and learns a frame sampling policy driven by the reward that directly relates to the performance of the tracker. HD-AMOT learns the diversified informative representation of frames and achieves effective cluster-based batch frame selection guided by multi-frame cooperation. We conduct extensive ablation studies to verify the design of our framework and compare our performance with random, uniform sampling, and several existing active learning works on four datasets including MOT15, MOT17, MOT20, and Dancetrack. The experimental results demonstrate the efficacy and effectiveness of our HD-AMOT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the Fundamental Research Funds for the Central Universities of China (2022JBMC009), the Natural Science Foundation of China (61972027, 62006244), and the Beijing Municipal Natural Science Foundation (Grant No. 4212041). This work is also partially supported by Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and Young Elite Scientist Sponsorship Program of Beijing Association for Science and Technology (BYESS2021178)." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.502, + 0.898 + ], + "angle": 0, + "content": "9939" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.173, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.148, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.187, + 0.469, + 0.224 + ], + "angle": 0, + "content": "[2] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.226, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.264, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.29, + 0.469, + 0.34 + ], + "angle": 0, + "content": "[5] Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.342, + 0.468, + 0.379 + ], + "angle": 0, + "content": "[6] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.379, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[7] Lile Cai, Xun Xu, Jun Hao Liew, and Chuan Sheng Foo. Revisiting superpixels for active learning in semantic segmentation with realistic annotation costs. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.418, + 0.468, + 0.455 + ], + "angle": 0, + "content": "[8] Razvan Caramalau, Binod Bhattacharai, and Tae-Kyun Kim. Sequential graph convolutional network for active learning. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.457, + 0.468, + 0.494 + ], + "angle": 0, + "content": "[9] Arantxa Casanova, Pedro O. Pinheiro, Negar Rostamzadeh, and Christopher J. Pal. Reinforced active learning for image segmentation. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.468, + 0.533 + ], + "angle": 0, + "content": "[10] Jiwong Choi, Ismail Elezi, Hyuk-Jae Lee, Clement Farabet, and Jose M Alvarez. Active learning for deep object detection via probabilistic modeling. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[11] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[14] Jiawei He, Zehao Huang, Naiyan Wang, and Zhaoxiang Zhang. Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.726, + 0.468, + 0.751 + ], + "angle": 0, + "content": "[15] Sheng-Jun Huang, Nengneng Gao, and Songcan Chen. Multi-instance multi-label active learning. In *IJCAI*, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.752, + 0.468, + 0.79 + ], + "angle": 0, + "content": "[16] Ming-xin Jiang, Chao Deng, Zhi-geng Pan, Lan-fang Wang, and Xing Sun. Multiobject tracking in videos based on lstm and deep reinforcement learning. Complexity, 2018, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.828 + ], + "angle": 0, + "content": "[17] Chanho Kim, Li Fuxin, Mazen Alotaibi, and James M Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.468, + 0.866 + ], + "angle": 0, + "content": "[18] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.469, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.891, + 0.164 + ], + "angle": 0, + "content": "[19] Yuan Li, Chang Huang, and Ram Nevatia. Learning to associate: Hybridboosted multi-target tracker for crowded scene. In CVPR, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.891, + 0.204 + ], + "angle": 0, + "content": "[20] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.891, + 0.242 + ], + "angle": 0, + "content": "[21] Zhao-Yang Liu and Sheng-Jun Huang. Active sampling for open-set classification without initial annotation. In AAAI, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.245, + 0.891, + 0.282 + ], + "angle": 0, + "content": "[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.891, + 0.321 + ], + "angle": 0, + "content": "[23] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv:1603.00831, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.891, + 0.372 + ], + "angle": 0, + "content": "[24] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. In NeurIPS, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.891, + 0.412 + ], + "angle": 0, + "content": "[25] Bo Pang, Yizhuo Li, Yifan Zhang, Muchen Li, and Cewu Lu. Tubetk: Adopting tubes to track multi-object in a one-step training model. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.891, + 0.451 + ], + "angle": 0, + "content": "[26] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.452, + 0.891, + 0.515 + ], + "angle": 0, + "content": "[27] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.891, + 0.566 + ], + "angle": 0, + "content": "[28] Yu Qiao, Jincheng Zhu, Chengjiang Long, Zeyao Zhang, Yuxin Wang, Zhenjun Du, and Xin Yang. Cpral: Collaborative panoptic-regional active learning for semantic segmentation. In AAAI, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.891, + 0.606 + ], + "angle": 0, + "content": "[29] Liangliang Ren, Jiwen Lu, Zifeng Wang, Qi Tian, and Jie Zhou. Collaborative deep reinforcement learning for multi-object tracking. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.607, + 0.891, + 0.645 + ], + "angle": 0, + "content": "[30] Idoia Ruiz, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Joan Serrat. Weakly supervised multi-object tracking and segmentation. In WACV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.891, + 0.682 + ], + "angle": 0, + "content": "[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.685, + 0.891, + 0.722 + ], + "angle": 0, + "content": "[32] Megh Shukla and Shuaib Ahmed. A mathematical analysis of learning loss for active learning in regression. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.724, + 0.891, + 0.761 + ], + "angle": 0, + "content": "[33] Yawar Siddiqui, Julien Valentin, and Matthias Nießner. Viewal: Active learning with viewpoint entropy for semantic segmentation. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.891, + 0.812 + ], + "angle": 0, + "content": "[34] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.814, + 0.891, + 0.84 + ], + "angle": 0, + "content": "[35] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, 1988." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.891, + 0.866 + ], + "angle": 0, + "content": "[36] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.891, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.888, + 0.502, + 0.898 + ], + "angle": 0, + "content": "9940" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.469, + 0.165 + ], + "angle": 0, + "content": "[37] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. In NeurIPS, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[38] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In AAAI, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.23 + ], + "angle": 0, + "content": "[39] Ran Wang, Sam Kwong, Xu Wang, and Yuheng Jia. Active k-labelsets ensemble for multi-label classification. PR, 109:107583, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[40] Jun Xiang, Guohan Xu, Chao Ma, and Jianhua Hou. End-to-end learning deep crf models for multi-object tracking deep crf models. IEEE TCSVT, 31(1):275-288, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.271, + 0.468, + 0.308 + ], + "angle": 0, + "content": "[41] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.468, + 0.335 + ], + "angle": 0, + "content": "[42] Donggeun Yoo and In So Kweon. Learning loss for active learning. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[43] Fengwei Yu, Wenbo Li, Quanquan Li, Yu Liu, Xiaohua Shi, and Junjie Yan. Poi: Multiple object tracking with high performance detection and appearance feature. In ECCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.468, + 0.413 + ], + "angle": 0, + "content": "[44] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.468, + 0.452 + ], + "angle": 0, + "content": "[45] Xiaosong Zhang, Fang Wan, Chang Liu, Xiangyang Ji, and Qixiang Ye. Learning to match anchors for visual object detection. IEEE TPAMI, 44(6):3096-3109, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.468, + 0.491 + ], + "angle": 0, + "content": "[46] Yang Zhang, Hao Sheng, Yubin Wu, Shuai Wang, Weifeng Lyu, Wei Ke, and Zhang Xiong. Long-term tracking with deep tracklet association. IEEE TIP, 29:6694-6706, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.493, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[47] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.468, + 0.595 + ], + "angle": 0, + "content": "[48] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.597, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[49] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.468, + 0.697 + ], + "angle": 0, + "content": "[50] Jian Zhao, Lin Xiong, Jianshu Li, Junliang Xing, Shuicheng Yan, and Jiashi Feng. 3d-aided dual-agent gans for unconstrained face recognition. IEEE TPAMI, 41(10):2380-2394, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.468, + 0.736 + ], + "angle": 0, + "content": "[51] Jian Zhao, Shuicheng Yan, and Jiashi Feng. Towards age-invariant face recognition. IEEE TPAMI, 44(1):474-487, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.739, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[52] Linyu Zheng, Ming Tang, Yingying Chen, Guibo Zhu, Jinqiao Wang, and Hanqing Lu. Improving multiple object tracking with single object tracking. In CVPR, pages 2453-2462, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.791, + 0.468, + 0.854 + ], + "angle": 0, + "content": "[53] Xue-Feng Zhu, Tianyang Xu, Jian Zhao, Jia-Wei Liu, Kai Wang, Gang Wang, Jianan Li, Zhihao Zhang, Qiang Wang, Lei Jin, et al. Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-uav system. arXiv preprint arXiv:2306.15767, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.469, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.887, + 0.5, + 0.898 + ], + "angle": 0, + "content": "9941" + } + ] +] \ No newline at end of file diff --git a/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_origin.pdf b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46c36ebf6512a1c70567cdd795263a672a88fcdb --- /dev/null +++ b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/b3ed4007-51fd-4d18-ab48-0223562578b5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab41d88609be2d5eb3e594bde723e3a62062c40f1a0f22236e71ff2d457d8fcd +size 926214 diff --git a/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/full.md b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a9217b4368115342696cb04556c8c312d8ca86e4 --- /dev/null +++ b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/full.md @@ -0,0 +1,292 @@ +# Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking + +Rui Li $^{1,2,*}$ , Baopeng Zhang $^{1}$ , Jun Liu $^{2}$ , Wei Liu $^{1}$ , Jian Zhao $^{3,4,5}$ , Zhu Teng $^{1,\dagger}$ + +$^{1}$ Beijing Jiaotong University $^{2}$ Singapore University of Technology and Design $^{3}$ Institute of North Electronic Equipment $^{4}$ Peng Cheng Laboratory $^{5}$ Intelligent Game and Decision Laboratory + +# Abstract + +The existing one-stage multi-object tracking (MOT) algorithms have achieved satisfactory performance benefiting from a large amount of labeled data. However, acquiring plenty of laborious annotated frames is not practical in real applications. To reduce the cost of human annotations, we propose Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT), to infer the most informative frames for any MOT tracker by observing the heterogeneous cues of samples. HD-AMOT defines the diversified informative representation by encoding the geometric and semantic information, and formulates the frame inference strategy as a Markov decision process to learn an optimal sampling policy based on the designed informative representation. Specifically, HD-AMOT consists of a diversified informative representation module as well as an informative frame selection network. The former produces the signal characterizing the diversity and distribution of frames, and the latter receives the signal and conducts multi-frame cooperation to enable batch frame sampling. Extensive experiments conducted on the MOT15, MOT17, MOT20, and Dancetrack datasets demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that under $50\%$ budget our HD-AMOT can achieve similar or even higher performance as fully-supervised learning. + +# 1. Introduction + +As a sophisticated research topic, one-stage multi-object tracking (MOT) has attracted more research attention in recent years, which simultaneously focuses on pinpointing multiple objects in each frame and recording their trajectories in the entire sequence. Despite the great success of deep learning-based one-stage MOT methods [25, 48, 27, 26, 52, 6, 22], they need a large amount of annotated samples for training, which requires costly human labor. For example, + +![](images/dd2276a73ebf6b921b24894cb9f21cca6e1c5c7b0395a60c6cbde7d53b1b5b40.jpg) + +![](images/7fa4108bcecbb1f2717227a1709816ee3e013bac64665718c1a1bc4fbc3ef62c.jpg) + +![](images/aa970e0183898c76838bc58553bc1306de5f0327e53917d143304a715d2027ea.jpg) +Figure 1. A stimulus example of active learning for one-stage MOT. There are many similar continuous frames in tracking sequences, such as frames 365-368 of MOT17-02 [23], and fully annotated these frames may have little effect on improving the performance of the tracker. Moreover, objects usually have different states, such as moving and static states, as shown in the red dotted box. The moving states usually contain more knowledge than the static ones, in which case the uniform and random sampling may fail to capture the diverse visual patterns of tracking sequences. + +![](images/64614865bfa16b36aa8d537043410eabbff7c4508bef431ce0599f5d3047b3be.jpg) + +in the MOT17 dataset [23], a sequence containing only 600 frames requires annotating 18,581 bounding boxes (including coordinates and identity labels) for 62 pedestrians. In addition, there are many similar continuous frames in tracking sequences, such as frames 365-368 of the MOT17-02 sequence in Fig. 1, which may have little effect on boosting the performance of the tracker. One way to deal with the above issue is to randomly or uniformly sample tracking sequences for annotation. However, as shown in the red dotted box of Fig. 1, objects in tracking sequences may have different motion states, and the moving state contains more knowledge than the static state, in which case the uniform sampling may fail to capture diverse visual patterns. Moreover, random sampling is easy to yield unstable results. Therefore, it is essential to develop algorithms to infer training data more efficiently in one-stage MOT. + +Instead of selecting the subset randomly or uniformly, + +active learning (AL) aims to infer the most informative samples as training data to improve the knowledge of a model, which is a promising way to select training data more efficiently for one-stage MOT. In the computer vision area, active learning has been widely explored for image classification [3, 21, 1, 8], object detection [45, 44, 10], semantic segmentation [1, 7, 28, 9], and so on. In these research topics, the image selection strategies of uncertainty-based [32, 42] and distribution-based methods [31, 1] become the mainstream of active learning. However, these AL algorithms do not directly relate to the performance improvements of the specific task and usually lead to suboptimal performance. More notably, they are usually utilized to select independent image samples, which are prone to achieve poor coverage of the tracking sequence and produce redundant frames if they are employed to the one-stage MOT task stiffly and directly. In particular, since new individuals appear constantly in the tracking sequence, the distribution bias caused by the poor coverage and redundant frames may make the tracker unable to generalize well for unseen cases. + +Furthermore, different from other tasks that distinguish inter-class objects in each image (such as image classification, object detection, etc.), one-stage MOT needs to detect and associate intra-class objects, e.g., pedestrians belonging to different individuals. This implies that the active learning for one-stage MOT is more challenging, because it should consider not only the discrepancy between frame-level images, but also the influence of more fine-grained intra-class diversity in sample selection, which plays a vital role in object association in one-stage MOT. + +To address the above issues, we formulate the active learning of one-stage MOT as a Markov Decision Process (MDP) [36] and design a novel Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT) framework, which leverages a Diversified Informative Representation Module (DIRM) and an Informative Frame Selection Network (IFSN) to learn a frame sampling policy. Specifically, IFSN is designed to infer a batch of informative and diverse frames for annotation by receiving the signal characterizing the diversity and distribution of frames and conducting multi-frame cooperation, which can effectively inhibit poor coverage and reduce redundant frames. In addition, to facilitate the inference of benignant frames, we develop a novel DIRM to characterize the frame-level diversified representation of samples by encoding the geometric and semantic cues of frames and objects. Besides the frame-level diversity representation, we also evaluate the set-level (labeled set and unlabeled set) discrepancy in DIRM, which dedicates to the acquirement of the training set with the unbiased distribution. + +Our main contributions are summarized as follows: (1) We propose a novel HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process (MDP) and + +learns a sampling policy driven by the reward that directly relates to the tracker performance. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning. (2) To infer the benignant frames, we design a diversified informative representation to characterize the diversity and distribution of frames, which learns both the set-level discrepancy and frame-level diversity by encoding the heterogeneous cues of geometry and semantics. (3) Extensive experiments are conducted on four datasets to demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that our HD-AMOT can achieve similar or even higher performance as fully-supervised learning under $50\%$ annotation budget. + +# 2. Related Work + +Multi-Object Tracking. MOT is a challenging task due to the complexity of tracking scenes. The existing works mainly follow the tracking-by-detection paradigm and two-stage methods are the mainstream, which regards the detection and tracking as two independent models [4, 14, 46, 40, 11, 17]. Despite noticeable progress, the two-stage methods cause a waste of resources and maybe lead to an efficiency issue because the independent detection and tracking models are executed separately in these methods. Recently, with the quick maturity of deep learning[50, 49, 51], one-stage frameworks for MOT [25, 48, 27, 26, 53, 52, 6, 22, 47] have begun to attract more research attention, which greatly improve the accuracy and efficiency of tracking. The fundamental concept revolves around the simultaneous integration of object detection and association within a singular network architecture. For example, CTracker [27] employs a chaining mechanism that connects paired bounding-box regression outcomes derived from overlapping nodes. Each of these nodes spans two consecutive frames. FairMOT [48] presents a simple yet effective approach based on the detection architecture CenterNet [13]. Despite the notable advancements introduced by these methods, their practical applicability remains constrained by the demanding nature of high-cost annotations. + +To reduce the demand for labeled data, several methods with less supervision signal such as weakly-supervised and self-supervised learning [30, 2] can effectively utilize the unlabeled data to improve the MOT performance. However, these methods pay more attention to the tracker learning, and ignore the importance of labeled data selection to improve the tracker performance. In this paper, we design an active learning approach for one-stage MOT, which could significantly reduce the annotation cost. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning. + +Active Learning. Active learning is an important machine learning problem, which has received lots of attention in computer vision, such as image classification + +![](images/60681c4cffe5cb407fbedf6ac91e4f41106122178f9fbc5e2e117917afbff96f.jpg) +Figure 2. The pipeline of our proposed HD-AMOT framework. In our method, we perform several iterations until a budget $\mathcal{B}$ of labeled frames is achieved. In the $t^{th}$ iteration, the steps are as follows: (I) A one-stage MOT tracker $\pmb{f}_t$ is used to obtain the heterogeneous cues including semantic feature $\pmb{F}^f$ and geometric information $\pmb{G}^h$ , $\pmb{G}^b$ on each frame; (II) The set-level discrepancy and frame-level diversity are estimated and construct the MDP state-action pair $(s_t, a_t)$ ; (III) The multi-frame cooperation based on $(s_t, a_t)$ is performed; (IV) With the assistance of multi-frame cooperation, IFSN infers $N$ unlabeled frames with high scores to be annotated. Meanwhile, $\pmb{X}_t^U$ , $\pmb{X}_t^L$ are updated to $\pmb{X}_{t+1}^U$ , $\pmb{X}_{t+1}^L$ by displacing newly annotated frames from $\pmb{X}_t^U$ to $\pmb{X}_t^L$ ; (V) The tracker $\pmb{f}_t$ is retrained on $\pmb{X}_{t+1}^L$ to obtain the updated $\pmb{f}_{t+1}$ . (VI) The reward $r_{t+1}$ based on $\pmb{f}_t$ and $\pmb{f}_{t+1}$ is computed on $\pmb{X}^R$ to update parameters of IFSN. + +[1, 8, 39, 20], object detection [45, 44, 10], semantic segmentation [7, 28, 33], and so on. These methods can be divided into uncertainty-based [32, 42] and distribution-based methods [31, 1]. For example, [15] exploits the uncertainty in both the input and output spaces to select the most valuable information. [31] defines the problem of active learning as core-set selection. [32] develops a foundation LearningLoss++ for the active learning of pose estimation to establish equivalency between Learning Loss [42] empirically driven objective and the KL divergence objective. These AL algorithms are used to select independent image samples for a specific task yet are not suitable for the video-wise one-stage MOT task. It is prone to cause the problems such as poor coverage and redundant frames if we directly apply the existing AL approaches to one-stage MOT stiffly. Accordingly, we design an active learning framework HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process and learns a batch frame sampling policy driven by the reward that directly relates to the performance of the MOT tracker, which gives consideration to both sampling coverage and efficiency. + +MDP of Multi-Object Tracking. As a sequential decision model, Markov decision process is usually applied in dynamic environments where an agent needs to perform certain tasks by making decisions and executing actions sequentially. In multi-object tracking, several MDP-based works have been explored [41, 29, 16]. [41] formulates the online two-stage MOT problem as decision making in MDPs, where the lifetime of an object is represented using an MDP model. [29] treats individual objects as agents, utilizing a prediction network for tracking, while optimizing tracking outcomes through collaborative interactions among various agents and their environments, facilitated by + +the decision network. In this paper, different from these methods that employ the MDP to directly solve the detection and tracking subtask of MOT, we address the annotation budget for one-stage MOT based on active learning by the designed Markov decision process, and design a diversified informative representation with an informative frame selection network to enable effective batch sampling under a specific annotation budget. + +# 3. Our Proposed HD-AMOT + +# 3.1. Overview + +In this paper, given an unlabeled video sequence $X$ with a limited annotation budget $\mathcal{B}$ , our HD-AMOT model aims to learn an optimal sampling strategy that infers and annotates the most informative frames iteratively to maximize the performance of the tracker $f$ . Specifically, we cast this AL problem of one-stage MOT by an MDP and adopt the Q-learning algorithm [24] to solve this problem, where we introduce an IFSN to infer frames according to the designed state-action representation $(s_t,a_t)$ . As shown in Fig. 2, in our method, the video sequence $X$ is divided into three different subsets $X^U,X^L$ , and $X^R$ . And at each iteration $t$ , the following steps are executed: + +(I) The heterogeneous cues including semantic features $\pmb{F}^{f}$ and geometric information $\pmb{G}^{h}$ , $\pmb{G}^{b}$ of each frame in the labeled set $\pmb{X}_{t}^{L}$ and unlabeled set $\pmb{X}_{t}^{U}$ are obtained by $\pmb{f}_{t}$ ; (II) We estimate the set-level discrepancy and frame-level diversity in DIRM based on the obtained heterogeneous cues, which are regarded as the MDP state-action pair $(s_{t}, a_{t})$ ; (III) The multi-frame cooperation based on $(s_{t}, a_{t})$ is performed on $\pmb{X}_{t}^{L}$ and $\pmb{X}_{t}^{U}$ ; (IV) With the assistance of multi-frame cooperation, IFSN selects $N$ unlabeled frames + +![](images/e71eb7e3d333436b962026a00fbac0056dea54a993e23212baec850924e3b4f8.jpg) +Figure 3. Illustration of geometric matrices construction process. (a): binary heatmap; (b): global topology matrix $G^{h}$ ; (c): local topology matrices $G^{h_1} \sim G^{h_K}$ ; (d): object scale matrix $G^b$ . + +with high scores to be annotated guided by frame clusters. Meanwhile, the subsets $X_{t}^{U}$ , $X_{t}^{L}$ are updated to $X_{t+1}^{U}$ , $X_{t+1}^{L}$ by moving newly annotated frames from $X_{t}^{U}$ to $X_{t}^{L}$ ; (V) The tracker $\pmb{f}_{t}$ is retrained on $X_{t+1}^{L}$ to obtain the updated $\pmb{f}_{t+1}$ ; (VI) The reward $r_{t+1}$ based on $\pmb{f}_{t}$ and $\pmb{f}_{t+1}$ is computed on the reward subset $\pmb{X}^{R}$ , which is used to update parameters of IFSN. + +# 3.2. Diversified Informative Representation + +To infer the most informative and representative frames, we propose a DIRM to compute the state and action of the designed MDP, which characterizes the diversity and distribution of frames. Specifically, to compensate for the distribution drift between the unlabeled set $\mathbf{X}_t^U$ and the labeled set $\mathbf{X}_t^L$ , we use heterogeneous cues to evaluate the set-level discrepancy as state $s_t$ . Moreover, from the perspective of selecting the unlabeled frame with the potential contribution to the MOT tracker training, we propose to capture the frame-level diversity of each unlabeled frame as action $a_t$ . + +Set-level Discrepancy. There are a large number of objects belonging to different individuals in the tracking sequence, and new individuals may appear constantly over time. To make the trained tracker well generalized to unseen individuals, it is necessary to generate an unbiased training set that captures the whole sequence distribution. Based on this intuition, we propose to narrow the distribution gap between the unlabeled and labeled sets by evaluating set-level discrepancy, to ensure the diversity and representativeness of the selected frames. + +Specifically, we consider two key attributes to characterize the distribution drifts between the labeled and unlabeled sets: semantic variation and spatial topological variation. We collect the global feature from the last feature extraction layer of $\pmb{f}_t$ to obtain the semantic feature $F^f$ , which depicts the general semantic information of the frame. For the spatial topology $G^h$ , we encode the center of each object obtained by the tracker $f_t$ into a binary heatmap to represent the spatial distribution of objects. Considering that identities of objects in the frames with similar spatial topology may be quite different due to the object motion, we encode the identities into the binary heatmap to obtain the final $G^h$ , as shown in Fig.3 (a) and (b). In addition, the performance of the MOT tracker is affected by objects or + +background, leading to different tracking quality over various local spatial areas. To suppress adverse effects of the low-quality local areas, we decompose the global spatial topology $\pmb{G}^{h}$ into $K$ local parts, which form our final spatial topology $\{\pmb{G}^{h},\pmb{G}^{h_{1} - h_{K}}\}$ together with the global topology. Similarly, the global and local semantic features constitute the final semantics $\{\pmb{F}^{f},\pmb{F}^{f_{1} - f_{K}}\}$ . In this way, the feature space $S_{set} = \{\pmb{G}^{h},\pmb{G}^{h_{1} - h_{K}},\pmb{F}^{f},\pmb{F}^{f_{1} - f_{K}}\}$ is formed. To model the set-level discrepancy between $X_{t}^{L}$ and $X_{t}^{U}$ , Maximum Mean Discrepancy [37] is adopted to calculate the set gap for each feature $S_{set}^{m}$ in the feature space $S_{set}$ as described in Eq. (1). Finally, the state representation $s_t$ is defined as the concatenation of all $\mathcal{D}^m$ to encode the distribution drifts between $X_{t}^{L}$ and $X_{t}^{U}$ . + +$$ +\begin{array}{l} \mathcal {D} ^ {m} = \frac {1}{n _ {l} ^ {2}} \sum_ {i, j} \varphi \left(p _ {i} ^ {m}, p _ {j} ^ {m}\right) + \frac {1}{n _ {u} ^ {2}} \sum_ {i, j} \varphi \left(q _ {i} ^ {m}, q _ {j} ^ {m}\right) \tag {1} \\ - \frac {2}{n _ {l} n _ {u}} \sum_ {i, j} \varphi \left(p _ {i} ^ {m}, q _ {j} ^ {m}\right), \\ \end{array} +$$ + +where $\mathcal{D}_m$ is a scalar representing the distribution discrepancy between $X_{t}^{L}$ and $X_{t}^{U}$ on $S_{\text{set}}^{m}$ . $p^m$ and $q^m$ denote the corresponding features of frames in $X_{t}^{L}$ and $X_{t}^{U}$ , $n_l$ and $n_u$ are numbers of frames in $X_{t}^{L}$ and $X_{t}^{U}$ . $\varphi(.)$ is the radial kernel [37] to measure the distance between two features. + +Frame-level Diversity. In this paper, we aim to select the frames that are beneficial to the tracker training for annotation. For this purpose, besides considering the set-level discrepancy, we also need to explore frame diversity from different perspectives to infer the most representative and informative frames. Different from the semantic feature and spatial topology used to learn set-level discrepancy, finer-grained cues are desiderated to evaluate the potential contribution of each unlabeled frame. Specifically, besides the frame-wise spatial topology $G^{h}$ and semantic feature $F^{f}$ , we also consider finer-grained attribute characterizing the individuation of objects: the scale of bounding boxes $G^{b}$ . Its construction process is similar to $G^{h}$ , except that the identities representing different individuals are replaced by the width and height of the corresponding object as shown in Fig.3 (d), i.e., $G^{h} \in \mathbb{R}^{H \times W \times 1}$ and $G^{b} \in \mathbb{R}^{H \times W \times 2}$ . + +Intuitively, we argue that any one of the geometric features $G^{h}$ , $G^{b}$ and semantic features $F^{f}$ is the key factor to determine whether a frame contributes to the performance improvement of the tracker. Accordingly, we consider the novelty and prominence of each unlabeled frame $x^{u}$ based on $S_{frame} = \{G^{h}, G^{b}, F^{f}\}$ , to form the frame-level diversity as action $a_{t}$ . The novelty of an unlabeled frame $x_{i}^{u}$ in $X_{t}^{U}$ is reflected in the similarities between $x_{i}^{u}$ and the labeled frames in $X_{t}^{L}$ , because the unlabeled frame with a novel characteristic (such as new individuals, different object scales, and postures) usually has low similarities to labeled frames of $X_{t}^{L}$ . And the promi + +nence of $x_{i}^{u}$ lies in the difference between $x_{i}^{u}$ and other unlabeled frames in $X_{t}^{U}$ , so we expect to select the most representative frame by recording the similarity distribution between each unlabeled frame and other frames in $X_{t}^{U}$ . Specifically, we introduce histogram-based representation to learn the novelty and prominence of each unlabeled frame. Taking $x_{i}^{u}$ as an example, to represent the novelty of $x_{i}^{u}$ , we calculate cosine similarities between $x_{i}^{u}$ and $X_{t}^{L}$ on each heterogeneous feature $S_{frame}^{m}$ and then obtain the histogram $\mathcal{H}_{nov}$ based on cosine similarities. Similar to novelty learning, we learn prominence by recoding the histogram $\mathcal{H}_{pro}$ based on cosine similarities between $x_{i}^{u}$ and $X_{t}^{U}$ . Finally, the action representation $a_{t}$ of $x_{i}^{u}$ is defined as $\{\mathcal{H}_{nov}^{G^{h}},\mathcal{H}_{nov}^{G^{b}},\mathcal{H}_{nov}^{F^{f}},\mathcal{H}_{pro}^{G^{h}},\mathcal{H}_{pro}^{G^{b}},\mathcal{H}_{pro}^{F^{f}}\}$ , which conduces to IFSN to effectively identify and select the informative and representative frames. + +# 3.3. Informative Frame Selection Learning + +Inspired by DQN [38], our lightweight IFSN follows an optimal policy that maximizes the potential performance gains. Considering the sampling efficiency, we propose a batch sampling method guided by frame clusters according to temporal clues. Meanwhile, to restrain the inferior performance caused by redundant frames, we further introduce the multi-frame cooperation strategy before batch sampling. + +Multi-frame Cooperation. There are many similar adjacent frames in tracking sequences, which easily leads to redundancy in the process of frame selection. To reduce redundant frames, we introduce a simple multi-frame cooperation strategy to model the neighborhood cooperation between the unlabeled frame $x_{i}^{u}$ and its nearest labeled frames. Specifically, we seek the nearest neighbors $x_{v}^{l}$ and $x_{w}^{l}$ of $x_{i}^{u}$ in $X_{t}^{L}$ and calculate their actions $a_{t}^{v}$ , $a_{t}^{w}$ . Then we use the fixed-length compact representation $o_{t} = \frac{1}{2}(a_{t}^{v} + a_{t}^{w})$ as an extra action for $x_{i}^{u}$ . To this end, the selection made by IFSN becomes: + +$$ +\phi_ {t} = \arg \max _ {a _ {t} \in A _ {t}} g \left(s _ {t}, a _ {t}, o _ {t}; \theta\right), \tag {2} +$$ + +where $a_{t}\in A_{t}$ denotes the candidate action in $X_{t}^{U}$ , and $g(.)$ is the lightweight IFSN network parameterized by $\theta$ . + +Batch Frame Selection. The structure of our designed lightweight IFSN consists of four linear layers, where a linear layer parameterized by $\theta_{1}$ is used to model the relation between each unlabeled frame and its nearest neighbors, and three linear layers parameterized by $\theta_{2\sim 4}$ are employed to score each unlabeled frame, as described in Eq. (3), where $cat[\cdot]$ denotes the concatenation operation. + +$$ +g \left(s _ {t}, a _ {t}, o _ {t}; \theta\right) = \theta_ {2 \sim 4} \left(c a t \left[ \theta_ {1} \left(a _ {t}, o _ {t}\right), a _ {t}, s _ {t} \right]\right), \tag {3} +$$ + +Intuitively, each unlabeled frame in $X_{t}^{U}$ can obtain the corresponding score by IFSN. However, sampling a single frame in each iteration by Eq. (2) to query annotation is inefficient. In this paper, we propose a batch sampling method + +guided by frame clusters to sample $N$ frames in each iteration. Based on the intuition that a larger difference exists between frames with longer time intervals, we divide the unlabeled frames into $N$ clusters in chronological order and select the frame for annotation in each cluster independently, which is equivalent to labeling a frame in parallel with $N$ annotators. In this case, $\phi_t$ in Eq. (2) is extended to consist of $N$ independent sub-selections $\{\theta_t^n\}_{n=1}^N$ , each with a restricted space $A_t^n$ , as described in Eq. (4). + +$$ +\phi_ {t} ^ {n} = \arg \max _ {a _ {t} \in A _ {t} ^ {n}} g \left(s _ {t}, a _ {t}, o _ {t}; \theta\right), \tag {4} +$$ + +where $a_{t}\in A_{t}^{n}$ denotes the candidate action in $n^{th}$ cluster. + +IFSN Optimization. In our method, we compute the reward on a subset $X^R$ , which is used to evaluate the benefit of selected frames to the tracker and optimize the frame selection network IFSN. To evaluate the performance of the tracker more comprehensively, we utilize MOTA [5] and IDF1 [19] as performance metrics to calculate the reward, where MOTA is a comprehensive metric to evaluate the detection performance and IDF1 tends to estimate the tracking ability of a tracker. As described in Eq. (5), the reward $r_{t+1}$ is defined as the difference of the performance metrics between $f_{t+1}$ and $f_t$ . + +$$ +r _ {t} = \left(e _ {t + 1} ^ {\text {m o t a}} + e _ {t + 1} ^ {i d f 1}\right) - \left(e _ {t} ^ {\text {m o t a}} + e _ {t} ^ {i d f 1}\right), \tag {5} +$$ + +where $e_{t+1}^{mota}$ and $e_{t+1}^{idf1}$ are MOTA and IDF1 metrics computed by $f_{t+1}$ , $e_t^{mota}$ and $e_t^{idf1}$ are evaluated by $f_t$ . + +With the reward $r_{t + 1}$ , we can optimize IFSN to infer the most informative frames to maximize the reward, leading to improved tracker performance during each active learning iteration. More specifically, we train IFSN in a double DQN formulation [38] by optimizing the loss based on the temporal difference error [35] as: + +$$ +\mathcal {L} _ {\mathrm {I F S N}} = \left(\hat {y} _ {t} - g \left(s _ {t}, a _ {t}, o _ {t}; \theta\right)\right) ^ {2}, \tag {6} +$$ + +$$ +\hat {y} _ {t} = r _ {t + 1} + \gamma g \left(s _ {t + 1}, a _ {t + 1}, o _ {t + 1}; \hat {\theta}\right), \tag {7} +$$ + +where $\theta$ and $\hat{\theta}$ are parameters of the IFSN policy network and off-policy network following the setting of double DQN [38]. $\gamma$ denotes a discount factor. + +With the described diversified informative representation in Section 3.2 and informative frame selection learning in this section, we introduce our entire HD-AMOT as illustrated in Alg. 1. + +# 4. Experiment + +# 4.1. Experimental Settings + +Datasets. We evaluate our HD-AMOT on MOT15 [18], MOT17 [23], MOT20 [12], and Dancetrack [34]. MOT15 and MOT17 are widely-used MOT datasets, and they are + +Algorithm 1: HD-AMOT Algorithm. +Input: The sequence $X$ , initial tracker $f_{init}$ and IFSN $g_{init}$ , AL procedure rounds $Z$ . +1 $\mathbf{X}_{init}, \mathbf{X}^{R} \gets$ RandomPartition $(X)$ ; +2 $\mathbf{X}_{init}^{U}, \mathbf{X}_{init}^{L} \gets$ UniformPartition $(X_{init})$ ; +3 $\mathbf{X}_{0}^{L} \gets \mathbf{X}_{init}^{L}, f_{0} \gets$ update $(g_{init}, \mathbf{X}_{init}^{L})$ ; +4 while not done do // Episodes +5 for $t = 0$ to $t = Z - 1$ do // AL procedure +6 Learn the diversified representation to form the state and action space; +7 Use IFSN to select frames by Eq. (4); +8 Annotate the selected $N$ frames $\mathbf{X}_{t}^{S}$ ; +9 Update $\mathbf{X}_{t}^{U}, \mathbf{X}_{t}^{L}$ : +10 $\mathbf{X}_{t + 1}^{U} \gets \mathbf{X}_{t}^{U} \setminus \mathbf{X}_{t}^{S}, \mathbf{X}_{t + 1}^{L} \gets \mathbf{X}_{t}^{L} \cup \mathbf{X}_{t}^{S}$ ; +11 $f_{t + 1} \gets$ update $(f_{t}, \mathbf{X}_{t + 1}^{L})$ ; +12 Calculate the reward on $\mathbf{X}^{R}$ by Eq. (5); +13 end +14 Update IFSN following Eq. (6); +15 Reset: $\mathbf{X}_{0}^{L} \gets \mathbf{X}_{init}^{L}, f_{0} \gets f_{t + 1}$ ; +16 end + +challenging because of the large variety in object pose, lighting, viewpoint, etc. In contrast to MOT15 and MOT17, MOT20 addresses the challenge of highly congested scenarios, where the pedestrian density can surge to as high as 246 individuals per frame. Dancetrack is a large-scale MOT dataset covering scenarios with low distinguishability of object appearance and diverse non-linear motion patterns. + +Evaluation Metrics. To evaluate the performance of one-stage MOT, the standard metrics MOTA [5] and IDF1 [19] are adopted for evaluation, which indicate the detection and association performance of one-stage MOT. Specifically, MOTA combines three error sources including false positives, missed targets, and identity switches, and IDF1 is the ratio of correctly identified detections over the average number of ground-truth and computed detections. + +Implementation Details. In this paper, we adopt the popular one-stage MOT method FairMOT [48] as the MOT tracker. Note that the propose algorithm is suitable for any one-stage MOT tracker. The input frames are resized to $1088 \times 608$ . We set the learning rate of our IFSN to 1e-4 and the discount factor $\gamma$ in Eq. (7) to 0.9, and the number of local features $K$ is set to 9 according to experiments. For each dataset, we randomly select a small sequence in each video at a ratio of 0.1 to form a reward subset $X^R$ . In the remaining frames, a small amount of frames is uniformly sampled as the initially labeled subset $X_{init}^{L}$ to initialize the MOT tracker $f$ , and the size is set to 100. For the semantic information, we extract the feature map from the last convolutional layer of ResNet34 as the global semantic feature $F^{f} \in \mathbb{R}^{512 \times 19 \times 34}$ . For the geometric clues, we utilize the center coordinates and identities of bounding + +Table 1. Ablation study on the design of the diversified informative representation. + +
RepresentationSettingMOTA (%)IDF1 (%)
Set-level DiscrepancyState w/o Gh,h1-h961.7364.95
State w/o Ff,f1-f961.4964.12
State w/o Gh, Ff62.7465.21
State w/o Gh1-h9, Ff1-f962.9264.55
Frame-level DiversityAction w/o Hgh62.8366.61
Action w/o HGb62.4966.23
Action w/o HFf61.7864.95
Action w/o H* nov61.2563.41
Action w/o H* pro62.2864.16
AllHD-AMOT63.2366.47
+ +boxes obtained by the MOT tracker to construct the global spatial topology $G^{h} \in \mathbb{R}^{152 \times 272}$ , and generate the object scale matrix $G^{b} \in \mathbb{R}^{152 \times 272 \times 2}$ by the box size. In addition, the number of clusters $N$ in the batch sampling is set to 12 for MOT15, MOT17, and MOT20, and is set to 24 for Dancetrack as it is much larger than other datasets. + +# 4.2. Ablation Study + +In this section, to analyze the impacts of different components, we conduct ablation studies on the popular dataset MOT17 following the division strategy in [43]. We run our experiments 5 times and report the mean performance for all ablation studies. + +Effect of Diversified Informative Representation. We perform an ablation study with a $20\%$ annotation budget to evaluate the contribution of our diversified informative representation. As IFSN relies on the state (set-level discrepancy) to decide the frame sampling policy, we investigate the influence of the set-level discrepancy by removing different components individually. Moreover, we also analyze the effect of the action design (frame-level diversity). As shown in Tab. 1, our complete diversified informative representation gives the best performance in terms of MOTA and is only slightly inferior to Action w/o $\mathcal{H}^{G^h}$ by $0.14\%$ on IDF1. We speculate that the scale information in $\mathcal{H}^{G^b}$ implicitly indicates the identity and topology clues of $\mathcal{H}^{G^h}$ , so the introduction of $\mathcal{H}^{G^b}$ weakens the role of $\mathcal{H}^{G^h}$ , making the MOTA of Action w/o $\mathcal{H}^{G^h}$ close to HD-AMOT and the IDF1 slightly higher. Furthermore, removing the components of set-level discrepancy or frame-level diversity in the diversified informative representation degrades the tracking performance. The largest decrease of MOTA and IDF1 occurs when the score of maximum similarity with labeled frames Action w/o $\mathcal{H}_{nov}^*$ is removed, which further verifies the effectiveness of heterogeneous cues evaluating the novel characteristic of unlabeled frames. + +Effect of Informative Frame Selection Network. We further validate the performance of our designed IFSN in Tab. 2. We first consider performing the cluster-based batch + +Table 2. Ablation study on the different settings of the Informative Frame Selection Network. + +
ModelMulti-frame CooperationBatch SamplingMOTA (%)↑IDF1 (%)↑
Bât.Clu.
A×✓✓62.7665.52
B✓××63.0765.62
C✓✓×62.1864.55
D✓✓✓63.2366.47
+ +![](images/ddc0ac4ffd87d1abbf0f5331a72de198fc7e51f21bcaabbea4c00ab7907d31c1.jpg) +Figure 4. Performance comparison evaluated by MOTA and IDF1 metrics with increasing budgets. + +![](images/43fc21309b8fe98f1e2f9dddb4ffd954707319f93751ded78b95648d8990a2ed.jpg) + +sampling without multi-frame cooperation (Model A). Then we analyze the usage of selecting a single frame in each AL iteration (Model B). In addition, we also construct the Model C as IFSN to select a batch of frames in one shot, where the frames with $N$ highest scores are sampled. Finally, we present the performance of our cluster-based batch sampling strategy (Model D). By comparing the results of Model A, Model D brings performance gains of $0.47\%$ and $0.95\%$ in terms of MOTA and IDF1 metrics, which proves that multi-frame cooperation plays a positive role in the active learning of one-stage MOT. In contrast, Model C gives the worst performance, which selects multiple frames according to the $N$ highest scores. We argue that frames with higher scores may be similar samples and close to each other, which easily leads to inefficiency in batch sampling. To this end, cluster-based batch sampling is introduced, as it can avoid selecting adjacent frames with higher scores and learn to sample with better coverage of the underlying distribution. Notably, selecting a single frame at each iteration in Model B also provides a competitive performance but still tends to be inferior to our method Model D, and the time cost of selecting a single frame per iteration is much higher than our cluster-based batch sampling strategy. + +Analysis of Different Budgets. We report the performance of our method, random sampling, uniform sampling, and other existing AL methods such as Coreset [31], CDAL [1] under different budgets, as shown in Fig. 4. For the IDF1 metric, ours is slightly inferior to the uniform sampling when the budget is $20\%$ and $40\%$ . We argue that uniform sampling can select informative frames for annotation if a small number of labeled frames is required. However, + +when the budget increases, there may be redundancy in the newly acquired frames by uniform sampling, and training with these labels does not provide more additional information, resulting in the stagnation or even decline of the performance of IDF1. Notably, our HD-AMOT with the $50\%$ budget achieves similar performance as other methods with the $80\%$ budget. It reveals that ours can save around $30\%$ labeling efforts on average compared to other sampling methods. We attribute this to our learned diversified informative clues and fine-grained object representation, which can effectively improve the tracking performance. + +# 4.3. Comparison with other AL Methods + +To further verify the generalization and effectiveness, we evaluate the proposed framework HD-AMOT on four diversified MOT datasets in this section. + +Results on MOT15. Tab. 3 lists the mean values of MOTA and IDF1 on MOT15, and Fig. 5 reveals the corresponding standard deviation of MOTA and IDF1. We can observe that HD-AMOT outperforms all methods by a clear margin, i.e., AL methods used in other tasks (CDAL [1] and Coreset [31]) stiffly applied to one-stage MOT can lead to obvious performance degradation. Notably, although Coreset achieves the second-best performance on MOTA and IDF1, it is clear from the standard deviations in Fig. 5 that Coreset has limited stability. Moreover, the performance of Coreset on other datasets (MOT17, MOT20, and Dancetrack) is not ideal, which indicates that the direct application of Coreset in one-stage MOT is unstable. + +Results on MOT17. MOT17 is a challenging dataset with more crowded scenarios, different viewpoints, camera motions, and weather conditions. As shown in Tab. 3, although uniform sampling achieves the best IDF1 on MOT17 and our HD-AMOT performs slightly lower than it by a margin of $0.18\%$ , uniform sampling may fail to capture diverse visual patterns in other datasets due to the lack of flexibility and adaptability. For example, in MOT15, ours outperforms uniform sampling by all performance metrics, which reflects the generalization ability of ours for one-stage MOT. Moreover, besides object association, the object detection ability should be considered in one-stage MOT. Our method, which benefits from learning the sampling policy driven by tracking performance metrics directly from data, significantly outperforms uniform sampling by $0.84\%$ in terms of the detection accuracy metric MOTA. + +Results on MOT20. We also use the large-scale dataset MOT20 to demonstrate the scalability of our HD-AMOT as shown in Tab. 3. In contrast to the aforementioned datasets, MOT20 is a recently released dataset that features more densely populated scenes. We can observe that ours still achieves the best performance, which demonstrates that our HD-AMOT can infer informative frames even on a crowded and complicated dataset. + +Table 3. Mean values of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack. Best and second-best results are shown in red and black, respectively. + +
DatasetMethodMOTA(%) (↑)IDF1(%) (↑)
MOT15Tracker w/ Random53.2859.61
Tracker w/ Uniform54.0959.50
Tracker w/ CDAL [1]53.1358.85
Tracker w/ Coreset [31]54.7060.29
Tracker w/ HD-AMOT55.2960.64
MOT17Tracker w/ Random70.1169.24
Tracker w/ Uniform70.4070.05
Tracker w/ CDAL [1]69.1068.71
Tracker w/ Coreset [31]69.4468.61
Tracker w/ HD-AMOT71.2469.87
MOT20Tracker w/ Random58.2165.37
Tracker w/ Uniform58.6565.87
Tracker w/ CDAL [1]57.5965.33
Tracker w/ Coreset [31]56.6664.86
Tracker w/ HD-AMOT59.2466.41
DancetrackTracker w/ Random80.4940.67
Tracker w/ Uniform81.1141.03
Tracker w/ CDAL [1]81.3641.16
Tracker w/ Coreset [31]81.1340.47
Tracker w/ HD-AMOT81.9441.93
+ +![](images/b6e6eaebe838bfb20171ae117726a260ef5830c3ba1f44b309db6101f3cd1960.jpg) +Figure 5. Standard deviations of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack. + +![](images/17f51011c2492d07fcf4d65ce23164ca53a45b5507b2e6850884de454d951d64.jpg) + +Results on Dancetrack. Dancetrack is a larger and more challenging dataset for group dancing scenes, where humans have similar appearance, diverse motion, and extreme articulation. We provide the performance evaluation results on the large-scale Dancetrack in Tab. 3. It can be observed that our HD-AMOT still achieves the best MOTA and IDF1 performance. Notably, from the perspective of sampling stability in Fig. 5, although uniform sampling can obtain a low standard deviation of MOTA, it leads to a very high standard deviation of IDF1 due to failure to capture informative frames that are helpful to object association. By contrast, our proposed HD-AMOT can maintain stable MOTA and IDF1 performance in multiple samplings. + +Furthermore, we give the performance comparison between our HD-AMOT with a $50\%$ budget and full supervision in Fig. 6. Our active learning framework achieves sim + +![](images/03d6db03f7fd845e9bba4207a329949565aeda57edb72df38eeacc65e52e9436.jpg) +Figure 6. Performance comparison between our HD-AMOT with a $50\%$ budget and full supervision on MOT15, MOT17, MOT20, and Dancetrack datasets. + +![](images/d32978e371dd9a232d66a1b70dd66856408330330e2530a7cf5e470a74a6a087.jpg) + +ilar or even higher performance compared with the tracker with fully-supervised training. For instance, we exceed the full supervision by a gain of $2\%$ on MOTA and $0.8\%$ on IDF1 for MOT15. These results implicitly prove that there are many similar frames in MOT sequences, and these similar frames bring limited progress in tracking performance. Our proposed HD-AMOT learns an optimal sampling policy by the heterogeneous diversity representation and a Markov decision process, which can infer informative frames that can most benefit the tracker, thus effectively improving the performance of the tracker. In summary, the proposed method can provide a consistent gain for the MOT tracker across datasets. + +# 5. Conclusion + +In this paper, we propose a novel framework HD-AMOT to investigate the annotation budget for one-stage MOT. It is formulated as an MDP and learns a frame sampling policy driven by the reward that directly relates to the performance of the tracker. HD-AMOT learns the diversified informative representation of frames and achieves effective cluster-based batch frame selection guided by multi-frame cooperation. We conduct extensive ablation studies to verify the design of our framework and compare our performance with random, uniform sampling, and several existing active learning works on four datasets including MOT15, MOT17, MOT20, and Dancetrack. The experimental results demonstrate the efficacy and effectiveness of our HD-AMOT. + +Acknowledgement. This work was supported by the Fundamental Research Funds for the Central Universities of China (2022JBMC009), the Natural Science Foundation of China (61972027, 62006244), and the Beijing Municipal Natural Science Foundation (Grant No. 4212041). This work is also partially supported by Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and Young Elite Scientist Sponsorship Program of Beijing Association for Science and Technology (BYESS2021178). + +# References + +[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In ECCV, 2020. +[2] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In NeurIPS, 2021. +[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In CVPR, 2018. +[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. +[5] Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. +[6] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR, 2022. +[7] Lile Cai, Xun Xu, Jun Hao Liew, and Chuan Sheng Foo. Revisiting superpixels for active learning in semantic segmentation with realistic annotation costs. In CVPR, 2021. +[8] Razvan Caramalau, Binod Bhattacharai, and Tae-Kyun Kim. Sequential graph convolutional network for active learning. In CVPR, 2021. +[9] Arantxa Casanova, Pedro O. Pinheiro, Negar Rostamzadeh, and Christopher J. Pal. Reinforced active learning for image segmentation. In ICLR, 2020. +[10] Jiwong Choi, Ismail Elezi, Hyuk-Jae Lee, Clement Farabet, and Jose M Alvarez. Active learning for deep object detection via probabilistic modeling. In CVPR, 2021. +[11] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In CVPR, 2021. +[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. +[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, 2019. +[14] Jiawei He, Zehao Huang, Naiyan Wang, and Zhaoxiang Zhang. Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In CVPR, 2021. +[15] Sheng-Jun Huang, Nengneng Gao, and Songcan Chen. Multi-instance multi-label active learning. In *IJCAI*, 2017. +[16] Ming-xin Jiang, Chao Deng, Zhi-geng Pan, Lan-fang Wang, and Xing Sun. Multiobject tracking in videos based on lstm and deep reinforcement learning. Complexity, 2018, 2018. +[17] Chanho Kim, Li Fuxin, Mazen Alotaibi, and James M Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, 2021. +[18] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942, 2015. + +[19] Yuan Li, Chang Huang, and Ram Nevatia. Learning to associate: Hybridboosted multi-target tracker for crowded scene. In CVPR, 2009. +[20] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In ICCV, 2021. +[21] Zhao-Yang Liu and Sheng-Jun Huang. Active sampling for open-set classification without initial annotation. In AAAI, 2019. +[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. +[23] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv:1603.00831, 2016. +[24] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. In NeurIPS, 2013. +[25] Bo Pang, Yizhuo Li, Yifan Zhang, Muchen Li, and Cewu Lu. Tubetk: Adopting tubes to track multi-object in a one-step training model. In CVPR, 2020. +[26] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. +[27] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In ECCV, 2020. +[28] Yu Qiao, Jincheng Zhu, Chengjiang Long, Zeyao Zhang, Yuxin Wang, Zhenjun Du, and Xin Yang. Cpral: Collaborative panoptic-regional active learning for semantic segmentation. In AAAI, 2022. +[29] Liangliang Ren, Jiwen Lu, Zifeng Wang, Qi Tian, and Jie Zhou. Collaborative deep reinforcement learning for multi-object tracking. In ECCV, 2018. +[30] Idoia Ruiz, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Joan Serrat. Weakly supervised multi-object tracking and segmentation. In WACV, 2021. +[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In ICLR, 2018. +[32] Megh Shukla and Shuaib Ahmed. A mathematical analysis of learning loss for active learning in regression. In CVPR, 2021. +[33] Yawar Siddiqui, Julien Valentin, and Matthias Nießner. Viewal: Active learning with viewpoint entropy for semantic segmentation. In CVPR, 2020. +[34] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, 2022. +[35] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, 1988. +[36] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018. + +[37] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. In NeurIPS, 2016. +[38] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In AAAI, 2016. +[39] Ran Wang, Sam Kwong, Xu Wang, and Yuheng Jia. Active k-labelsets ensemble for multi-label classification. PR, 109:107583, 2021. +[40] Jun Xiang, Guohan Xu, Chao Ma, and Jianhua Hou. End-to-end learning deep crf models for multi-object tracking deep crf models. IEEE TCSVT, 31(1):275-288, 2020. +[41] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015. +[42] Donggeun Yoo and In So Kweon. Learning loss for active learning. In CVPR, 2019. +[43] Fengwei Yu, Wenbo Li, Quanquan Li, Yu Liu, Xiaohua Shi, and Junjie Yan. Poi: Multiple object tracking with high performance detection and appearance feature. In ECCV, 2016. +[44] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In CVPR, 2021. +[45] Xiaosong Zhang, Fang Wan, Chang Liu, Xiangyang Ji, and Qixiang Ye. Learning to match anchors for visual object detection. IEEE TPAMI, 44(6):3096-3109, 2021. +[46] Yang Zhang, Hao Sheng, Yubin Wu, Shuai Wang, Weifeng Lyu, Wei Ke, and Zhang Xiong. Long-term tracking with deep tracklet association. IEEE TIP, 29:6694-6706, 2020. +[47] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022. +[48] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021. +[49] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018. +[50] Jian Zhao, Lin Xiong, Jianshu Li, Junliang Xing, Shuicheng Yan, and Jiashi Feng. 3d-aided dual-agent gans for unconstrained face recognition. IEEE TPAMI, 41(10):2380-2394, 2018. +[51] Jian Zhao, Shuicheng Yan, and Jiashi Feng. Towards age-invariant face recognition. IEEE TPAMI, 44(1):474-487, 2020. +[52] Linyu Zheng, Ming Tang, Yingying Chen, Guibo Zhu, Jinqiao Wang, and Hanqing Lu. Improving multiple object tracking with single object tracking. In CVPR, pages 2453-2462, 2021. +[53] Xue-Feng Zhu, Tianyang Xu, Jian Zhao, Jia-Wei Liu, Kai Wang, Gang Wang, Jianan Li, Zhihao Zhang, Qiang Wang, Lei Jin, et al. Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-uav system. arXiv preprint arXiv:2306.15767, 2023. \ No newline at end of file diff --git a/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/images.zip b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0f11147c30cf0176f56a5c6c5f1461f8d18cb8bb --- /dev/null +++ b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61abb2c5327301c0703d7ff9b0d95f9d367bc8192702bdbb39566f844a68990c +size 430472 diff --git a/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/layout.json b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5da40de8cfd812e527e56540f7d0a3b13c9ea664 --- /dev/null +++ b/2023/Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking/layout.json @@ -0,0 +1,10690 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 136, + 515, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 136, + 515, + 154 + ], + "spans": [ + { + "bbox": [ + 62, + 136, + 515, + 154 + ], + "type": "text", + "content": "Heterogeneous Diversity Driven Active Learning for Multi-Object Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "spans": [ + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "text", + "content": "Rui Li" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "text", + "content": ", Baopeng Zhang" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "text", + "content": ", Jun Liu" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "text", + "content": ", Wei Liu" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "text", + "content": ", Jian Zhao" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "inline_equation", + "content": "^{3,4,5}" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "text", + "content": ", Zhu Teng" + }, + { + "bbox": [ + 107, + 184, + 473, + 201 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "spans": [ + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "text", + "content": "Beijing Jiaotong University " + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "text", + "content": "Singapore University of Technology and Design " + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "text", + "content": "Institute of North Electronic Equipment " + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "text", + "content": "Peng Cheng Laboratory " + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 63, + 207, + 514, + 233 + ], + "type": "text", + "content": "Intelligent Game and Decision Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 266, + 186, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 266, + 186, + 278 + ], + "spans": [ + { + "bbox": [ + 139, + 266, + 186, + 278 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 290, + 282, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 290, + 282, + 571 + ], + "spans": [ + { + "bbox": [ + 45, + 290, + 282, + 571 + ], + "type": "text", + "content": "The existing one-stage multi-object tracking (MOT) algorithms have achieved satisfactory performance benefiting from a large amount of labeled data. However, acquiring plenty of laborious annotated frames is not practical in real applications. To reduce the cost of human annotations, we propose Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT), to infer the most informative frames for any MOT tracker by observing the heterogeneous cues of samples. HD-AMOT defines the diversified informative representation by encoding the geometric and semantic information, and formulates the frame inference strategy as a Markov decision process to learn an optimal sampling policy based on the designed informative representation. Specifically, HD-AMOT consists of a diversified informative representation module as well as an informative frame selection network. The former produces the signal characterizing the diversity and distribution of frames, and the latter receives the signal and conducts multi-frame cooperation to enable batch frame sampling. Extensive experiments conducted on the MOT15, MOT17, MOT20, and Dancetrack datasets demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that under " + }, + { + "bbox": [ + 45, + 290, + 282, + 571 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 45, + 290, + 282, + 571 + ], + "type": "text", + "content": " budget our HD-AMOT can achieve similar or even higher performance as fully-supervised learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 591, + 124, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 124, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 124, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 610, + 279, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 610, + 279, + 704 + ], + "spans": [ + { + "bbox": [ + 45, + 610, + 279, + 704 + ], + "type": "text", + "content": "As a sophisticated research topic, one-stage multi-object tracking (MOT) has attracted more research attention in recent years, which simultaneously focuses on pinpointing multiple objects in each frame and recording their trajectories in the entire sequence. Despite the great success of deep learning-based one-stage MOT methods [25, 48, 27, 26, 52, 6, 22], they need a large amount of annotated samples for training, which requires costly human labor. For example," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 265, + 413, + 368 + ], + "blocks": [ + { + "bbox": [ + 301, + 265, + 413, + 368 + ], + "lines": [ + { + "bbox": [ + 301, + 265, + 413, + 368 + ], + "spans": [ + { + "bbox": [ + 301, + 265, + 413, + 368 + ], + "type": "image", + "image_path": "dd2276a73ebf6b921b24894cb9f21cca6e1c5c7b0395a60c6cbde7d53b1b5b40.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 413, + 265, + 527, + 368 + ], + "blocks": [ + { + "bbox": [ + 413, + 265, + 527, + 368 + ], + "lines": [ + { + "bbox": [ + 413, + 265, + 527, + 368 + ], + "spans": [ + { + "bbox": [ + 413, + 265, + 527, + 368 + ], + "type": "image", + "image_path": "7fa4108bcecbb1f2717227a1709816ee3e013bac64665718c1a1bc4fbc3ef62c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 302, + 368, + 406, + 418 + ], + "blocks": [ + { + "bbox": [ + 302, + 368, + 406, + 418 + ], + "lines": [ + { + "bbox": [ + 302, + 368, + 406, + 418 + ], + "spans": [ + { + "bbox": [ + 302, + 368, + 406, + 418 + ], + "type": "image", + "image_path": "aa970e0183898c76838bc58553bc1306de5f0327e53917d143304a715d2027ea.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 425, + 531, + 522 + ], + "lines": [ + { + "bbox": [ + 296, + 425, + 531, + 522 + ], + "spans": [ + { + "bbox": [ + 296, + 425, + 531, + 522 + ], + "type": "text", + "content": "Figure 1. A stimulus example of active learning for one-stage MOT. There are many similar continuous frames in tracking sequences, such as frames 365-368 of MOT17-02 [23], and fully annotated these frames may have little effect on improving the performance of the tracker. Moreover, objects usually have different states, such as moving and static states, as shown in the red dotted box. The moving states usually contain more knowledge than the static ones, in which case the uniform and random sampling may fail to capture the diverse visual patterns of tracking sequences." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 412, + 368, + 527, + 418 + ], + "blocks": [ + { + "bbox": [ + 412, + 368, + 527, + 418 + ], + "lines": [ + { + "bbox": [ + 412, + 368, + 527, + 418 + ], + "spans": [ + { + "bbox": [ + 412, + 368, + 527, + 418 + ], + "type": "image", + "image_path": "64614865bfa16b36aa8d537043410eabbff7c4508bef431ce0599f5d3047b3be.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 528, + 531, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 528, + 531, + 715 + ], + "spans": [ + { + "bbox": [ + 296, + 528, + 531, + 715 + ], + "type": "text", + "content": "in the MOT17 dataset [23], a sequence containing only 600 frames requires annotating 18,581 bounding boxes (including coordinates and identity labels) for 62 pedestrians. In addition, there are many similar continuous frames in tracking sequences, such as frames 365-368 of the MOT17-02 sequence in Fig. 1, which may have little effect on boosting the performance of the tracker. One way to deal with the above issue is to randomly or uniformly sample tracking sequences for annotation. However, as shown in the red dotted box of Fig. 1, objects in tracking sequences may have different motion states, and the moving state contains more knowledge than the static state, in which case the uniform sampling may fail to capture diverse visual patterns. Moreover, random sampling is easy to yield unstable results. Therefore, it is essential to develop algorithms to infer training data more efficiently in one-stage MOT." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 717, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 717, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 308, + 717, + 530, + 730 + ], + "type": "text", + "content": "Instead of selecting the subset randomly or uniformly," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "spans": [ + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 708, + 246, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 708, + 246, + 730 + ], + "spans": [ + { + "bbox": [ + 55, + 708, + 246, + 730 + ], + "type": "text", + "content": "*Work done when Rui Li was an visiting student of SUTD. †Zhu Teng is the corresponding author (zteng@bjtu.edu.cn)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "9932" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 339 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 339 + ], + "type": "text", + "content": "active learning (AL) aims to infer the most informative samples as training data to improve the knowledge of a model, which is a promising way to select training data more efficiently for one-stage MOT. In the computer vision area, active learning has been widely explored for image classification [3, 21, 1, 8], object detection [45, 44, 10], semantic segmentation [1, 7, 28, 9], and so on. In these research topics, the image selection strategies of uncertainty-based [32, 42] and distribution-based methods [31, 1] become the mainstream of active learning. However, these AL algorithms do not directly relate to the performance improvements of the specific task and usually lead to suboptimal performance. More notably, they are usually utilized to select independent image samples, which are prone to achieve poor coverage of the tracking sequence and produce redundant frames if they are employed to the one-stage MOT task stiffly and directly. In particular, since new individuals appear constantly in the tracking sequence, the distribution bias caused by the poor coverage and redundant frames may make the tracker unable to generalize well for unseen cases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 341, + 281, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 341, + 281, + 458 + ], + "spans": [ + { + "bbox": [ + 45, + 341, + 281, + 458 + ], + "type": "text", + "content": "Furthermore, different from other tasks that distinguish inter-class objects in each image (such as image classification, object detection, etc.), one-stage MOT needs to detect and associate intra-class objects, e.g., pedestrians belonging to different individuals. This implies that the active learning for one-stage MOT is more challenging, because it should consider not only the discrepancy between frame-level images, but also the influence of more fine-grained intra-class diversity in sample selection, which plays a vital role in object association in one-stage MOT." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 459, + 281, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 459, + 281, + 692 + ], + "spans": [ + { + "bbox": [ + 45, + 459, + 281, + 692 + ], + "type": "text", + "content": "To address the above issues, we formulate the active learning of one-stage MOT as a Markov Decision Process (MDP) [36] and design a novel Heterogeneous Diversity driven Active Multi-Object Tracking (HD-AMOT) framework, which leverages a Diversified Informative Representation Module (DIRM) and an Informative Frame Selection Network (IFSN) to learn a frame sampling policy. Specifically, IFSN is designed to infer a batch of informative and diverse frames for annotation by receiving the signal characterizing the diversity and distribution of frames and conducting multi-frame cooperation, which can effectively inhibit poor coverage and reduce redundant frames. In addition, to facilitate the inference of benignant frames, we develop a novel DIRM to characterize the frame-level diversified representation of samples by encoding the geometric and semantic cues of frames and objects. Besides the frame-level diversity representation, we also evaluate the set-level (labeled set and unlabeled set) discrepancy in DIRM, which dedicates to the acquirement of the training set with the unbiased distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 695, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 695, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 695, + 280, + 730 + ], + "type": "text", + "content": "Our main contributions are summarized as follows: (1) We propose a novel HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process (MDP) and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 106, + 532, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 532, + 259 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 532, + 259 + ], + "type": "text", + "content": "learns a sampling policy driven by the reward that directly relates to the tracker performance. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning. (2) To infer the benignant frames, we design a diversified informative representation to characterize the diversity and distribution of frames, which learns both the set-level discrepancy and frame-level diversity by encoding the heterogeneous cues of geometry and semantics. (3) Extensive experiments are conducted on four datasets to demonstrate the efficacy and effectiveness of HD-AMOT. Experiments show that our HD-AMOT can achieve similar or even higher performance as fully-supervised learning under " + }, + { + "bbox": [ + 296, + 106, + 532, + 259 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 296, + 106, + 532, + 259 + ], + "type": "text", + "content": " annotation budget." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 268, + 381, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 268, + 381, + 280 + ], + "spans": [ + { + "bbox": [ + 297, + 268, + 381, + 280 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 287, + 532, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 287, + 532, + 566 + ], + "spans": [ + { + "bbox": [ + 296, + 287, + 532, + 566 + ], + "type": "text", + "content": "Multi-Object Tracking. MOT is a challenging task due to the complexity of tracking scenes. The existing works mainly follow the tracking-by-detection paradigm and two-stage methods are the mainstream, which regards the detection and tracking as two independent models [4, 14, 46, 40, 11, 17]. Despite noticeable progress, the two-stage methods cause a waste of resources and maybe lead to an efficiency issue because the independent detection and tracking models are executed separately in these methods. Recently, with the quick maturity of deep learning[50, 49, 51], one-stage frameworks for MOT [25, 48, 27, 26, 53, 52, 6, 22, 47] have begun to attract more research attention, which greatly improve the accuracy and efficiency of tracking. The fundamental concept revolves around the simultaneous integration of object detection and association within a singular network architecture. For example, CTracker [27] employs a chaining mechanism that connects paired bounding-box regression outcomes derived from overlapping nodes. Each of these nodes spans two consecutive frames. FairMOT [48] presents a simple yet effective approach based on the detection architecture CenterNet [13]. Despite the notable advancements introduced by these methods, their practical applicability remains constrained by the demanding nature of high-cost annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 566, + 533, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 566, + 533, + 694 + ], + "spans": [ + { + "bbox": [ + 296, + 566, + 533, + 694 + ], + "type": "text", + "content": "To reduce the demand for labeled data, several methods with less supervision signal such as weakly-supervised and self-supervised learning [30, 2] can effectively utilize the unlabeled data to improve the MOT performance. However, these methods pay more attention to the tracker learning, and ignore the importance of labeled data selection to improve the tracker performance. In this paper, we design an active learning approach for one-stage MOT, which could significantly reduce the annotation cost. To the best of our knowledge, this is the first work that investigates the annotation budget for one-stage MOT based on active learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 695, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 695, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 695, + 532, + 730 + ], + "type": "text", + "content": "Active Learning. Active learning is an important machine learning problem, which has received lots of attention in computer vision, such as image classification" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "9933" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 105, + 530, + 250 + ], + "blocks": [ + { + "bbox": [ + 48, + 105, + 530, + 250 + ], + "lines": [ + { + "bbox": [ + 48, + 105, + 530, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 105, + 530, + 250 + ], + "type": "image", + "image_path": "60681c4cffe5cb407fbedf6ac91e4f41106122178f9fbc5e2e117917afbff96f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "lines": [ + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "spans": [ + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": "Figure 2. The pipeline of our proposed HD-AMOT framework. In our method, we perform several iterations until a budget " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " of labeled frames is achieved. In the " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " iteration, the steps are as follows: (I) A one-stage MOT tracker " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{f}_t" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " is used to obtain the heterogeneous cues including semantic feature " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{F}^f" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " and geometric information " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{G}^h" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{G}^b" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " on each frame; (II) The set-level discrepancy and frame-level diversity are estimated and construct the MDP state-action pair " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "(s_t, a_t)" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": "; (III) The multi-frame cooperation based on " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "(s_t, a_t)" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " is performed; (IV) With the assistance of multi-frame cooperation, IFSN infers " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " unlabeled frames with high scores to be annotated. Meanwhile, " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}_t^U" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}_t^L" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " are updated to " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}_{t+1}^U" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}_{t+1}^L" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " by displacing newly annotated frames from " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}_t^U" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}_t^L" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": "; (V) The tracker " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{f}_t" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " is retrained on " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}_{t+1}^L" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " to obtain the updated " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t+1}" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": ". (VI) The reward " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "r_{t+1}" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{f}_t" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t+1}" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " is computed on " + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "inline_equation", + "content": "\\pmb{X}^R" + }, + { + "bbox": [ + 45, + 258, + 531, + 335 + ], + "type": "text", + "content": " to update parameters of IFSN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 336, + 281, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 336, + 281, + 582 + ], + "spans": [ + { + "bbox": [ + 45, + 336, + 281, + 582 + ], + "type": "text", + "content": "[1, 8, 39, 20], object detection [45, 44, 10], semantic segmentation [7, 28, 33], and so on. These methods can be divided into uncertainty-based [32, 42] and distribution-based methods [31, 1]. For example, [15] exploits the uncertainty in both the input and output spaces to select the most valuable information. [31] defines the problem of active learning as core-set selection. [32] develops a foundation LearningLoss++ for the active learning of pose estimation to establish equivalency between Learning Loss [42] empirically driven objective and the KL divergence objective. These AL algorithms are used to select independent image samples for a specific task yet are not suitable for the video-wise one-stage MOT task. It is prone to cause the problems such as poor coverage and redundant frames if we directly apply the existing AL approaches to one-stage MOT stiffly. Accordingly, we design an active learning framework HD-AMOT for one-stage MOT, which is formulated as a Markov Decision Process and learns a batch frame sampling policy driven by the reward that directly relates to the performance of the MOT tracker, which gives consideration to both sampling coverage and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": "MDP of Multi-Object Tracking. As a sequential decision model, Markov decision process is usually applied in dynamic environments where an agent needs to perform certain tasks by making decisions and executing actions sequentially. In multi-object tracking, several MDP-based works have been explored [41, 29, 16]. [41] formulates the online two-stage MOT problem as decision making in MDPs, where the lifetime of an object is represented using an MDP model. [29] treats individual objects as agents, utilizing a prediction network for tracking, while optimizing tracking outcomes through collaborative interactions among various agents and their environments, facilitated by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 336, + 531, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 336, + 531, + 431 + ], + "spans": [ + { + "bbox": [ + 296, + 336, + 531, + 431 + ], + "type": "text", + "content": "the decision network. In this paper, different from these methods that employ the MDP to directly solve the detection and tracking subtask of MOT, we address the annotation budget for one-stage MOT based on active learning by the designed Markov decision process, and design a diversified informative representation with an informative frame selection network to enable effective batch sampling under a specific annotation budget." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 444, + 442, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 444, + 442, + 458 + ], + "spans": [ + { + "bbox": [ + 296, + 444, + 442, + 458 + ], + "type": "text", + "content": "3. Our Proposed HD-AMOT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 464, + 363, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 464, + 363, + 476 + ], + "spans": [ + { + "bbox": [ + 297, + 464, + 363, + 476 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "spans": [ + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": "In this paper, given an unlabeled video sequence " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": " with a limited annotation budget " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": ", our HD-AMOT model aims to learn an optimal sampling strategy that infers and annotates the most informative frames iteratively to maximize the performance of the tracker " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": ". Specifically, we cast this AL problem of one-stage MOT by an MDP and adopt the Q-learning algorithm [24] to solve this problem, where we introduce an IFSN to infer frames according to the designed state-action representation " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "(s_t,a_t)" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": ". As shown in Fig. 2, in our method, the video sequence " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": " is divided into three different subsets " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "X^U,X^L" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "X^R" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": ". And at each iteration " + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 484, + 530, + 623 + ], + "type": "text", + "content": ", the following steps are executed:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": "(I) The heterogeneous cues including semantic features " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{f}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": " and geometric information " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{h}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{b}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": " of each frame in the labeled set " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{X}_{t}^{L}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": " and unlabeled set " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{X}_{t}^{U}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": " are obtained by " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": "; (II) We estimate the set-level discrepancy and frame-level diversity in DIRM based on the obtained heterogeneous cues, which are regarded as the MDP state-action pair " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "(s_{t}, a_{t})" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": "; (III) The multi-frame cooperation based on " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "(s_{t}, a_{t})" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": " is performed on " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{X}_{t}^{L}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{X}_{t}^{U}" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": "; (IV) With the assistance of multi-frame cooperation, IFSN selects " + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 624, + 531, + 730 + ], + "type": "text", + "content": " unlabeled frames" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 746, + 298, + 756 + ], + "type": "text", + "content": "9934" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 105, + 276, + 175 + ], + "blocks": [ + { + "bbox": [ + 49, + 105, + 276, + 175 + ], + "lines": [ + { + "bbox": [ + 49, + 105, + 276, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 105, + 276, + 175 + ], + "type": "image", + "image_path": "e71eb7e3d333436b962026a00fbac0056dea54a993e23212baec850924e3b4f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "lines": [ + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "spans": [ + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "type": "text", + "content": "Figure 3. Illustration of geometric matrices construction process. (a): binary heatmap; (b): global topology matrix " + }, + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "type": "inline_equation", + "content": "G^{h}" + }, + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "type": "text", + "content": "; (c): local topology matrices " + }, + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "type": "inline_equation", + "content": "G^{h_1} \\sim G^{h_K}" + }, + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "type": "text", + "content": "; (d): object scale matrix " + }, + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "type": "inline_equation", + "content": "G^b" + }, + { + "bbox": [ + 45, + 182, + 278, + 215 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "spans": [ + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": "with high scores to be annotated guided by frame clusters. Meanwhile, the subsets " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " are updated to " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "X_{t+1}^{U}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "X_{t+1}^{L}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " by moving newly annotated frames from " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": "; (V) The tracker " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " is retrained on " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "X_{t+1}^{L}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " to obtain the updated " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t+1}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": "; (VI) The reward " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "r_{t+1}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{t+1}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": " is computed on the reward subset " + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{X}^{R}" + }, + { + "bbox": [ + 45, + 218, + 278, + 304 + ], + "type": "text", + "content": ", which is used to update parameters of IFSN." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 244, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 244, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 244, + 326 + ], + "type": "text", + "content": "3.2. Diversified Informative Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "spans": [ + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "text", + "content": "To infer the most informative and representative frames, we propose a DIRM to compute the state and action of the designed MDP, which characterizes the diversity and distribution of frames. Specifically, to compensate for the distribution drift between the unlabeled set " + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^U" + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "text", + "content": " and the labeled set " + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t^L" + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "text", + "content": ", we use heterogeneous cues to evaluate the set-level discrepancy as state " + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "text", + "content": ". Moreover, from the perspective of selecting the unlabeled frame with the potential contribution to the MOT tracker training, we propose to capture the frame-level diversity of each unlabeled frame as action " + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 45, + 332, + 278, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 450, + 278, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 450, + 278, + 565 + ], + "spans": [ + { + "bbox": [ + 45, + 450, + 278, + 565 + ], + "type": "text", + "content": "Set-level Discrepancy. There are a large number of objects belonging to different individuals in the tracking sequence, and new individuals may appear constantly over time. To make the trained tracker well generalized to unseen individuals, it is necessary to generate an unbiased training set that captures the whole sequence distribution. Based on this intuition, we propose to narrow the distribution gap between the unlabeled and labeled sets by evaluating set-level discrepancy, to ensure the diversity and representativeness of the selected frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "text", + "content": "Specifically, we consider two key attributes to characterize the distribution drifts between the labeled and unlabeled sets: semantic variation and spatial topological variation. We collect the global feature from the last feature extraction layer of " + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{f}_t" + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "text", + "content": " to obtain the semantic feature " + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "inline_equation", + "content": "F^f" + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "text", + "content": ", which depicts the general semantic information of the frame. For the spatial topology " + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "inline_equation", + "content": "G^h" + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "text", + "content": ", we encode the center of each object obtained by the tracker " + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "text", + "content": " into a binary heatmap to represent the spatial distribution of objects. Considering that identities of objects in the frames with similar spatial topology may be quite different due to the object motion, we encode the identities into the binary heatmap to obtain the final " + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "inline_equation", + "content": "G^h" + }, + { + "bbox": [ + 45, + 566, + 278, + 730 + ], + "type": "text", + "content": ", as shown in Fig.3 (a) and (b). In addition, the performance of the MOT tracker is affected by objects or" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": "background, leading to different tracking quality over various local spatial areas. To suppress adverse effects of the low-quality local areas, we decompose the global spatial topology " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{h}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " local parts, which form our final spatial topology " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "\\{\\pmb{G}^{h},\\pmb{G}^{h_{1} - h_{K}}\\}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " together with the global topology. Similarly, the global and local semantic features constitute the final semantics " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "\\{\\pmb{F}^{f},\\pmb{F}^{f_{1} - f_{K}}\\}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": ". In this way, the feature space " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "S_{set} = \\{\\pmb{G}^{h},\\pmb{G}^{h_{1} - h_{K}},\\pmb{F}^{f},\\pmb{F}^{f_{1} - f_{K}}\\}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " is formed. To model the set-level discrepancy between " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": ", Maximum Mean Discrepancy [37] is adopted to calculate the set gap for each feature " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "S_{set}^{m}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " in the feature space " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "S_{set}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " as described in Eq. (1). Finally, the state representation " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " is defined as the concatenation of all " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^m" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " to encode the distribution drifts between " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 296, + 106, + 530, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 327, + 285, + 530, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 285, + 530, + 338 + ], + "spans": [ + { + "bbox": [ + 327, + 285, + 530, + 338 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} ^ {m} = \\frac {1}{n _ {l} ^ {2}} \\sum_ {i, j} \\varphi \\left(p _ {i} ^ {m}, p _ {j} ^ {m}\\right) + \\frac {1}{n _ {u} ^ {2}} \\sum_ {i, j} \\varphi \\left(q _ {i} ^ {m}, q _ {j} ^ {m}\\right) \\tag {1} \\\\ - \\frac {2}{n _ {l} n _ {u}} \\sum_ {i, j} \\varphi \\left(p _ {i} ^ {m}, q _ {j} ^ {m}\\right), \\\\ \\end{array}", + "image_path": "554a972f52c78aa163014c50ad31b0693509fde9d6c71ab2544641c1fde6a85c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "spans": [ + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " is a scalar representing the distribution discrepancy between " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "S_{\\text{set}}^{m}" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "p^m" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "q^m" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " denote the corresponding features of frames in " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "n_l" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "n_u" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " are numbers of frames in " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "inline_equation", + "content": "\\varphi(.)" + }, + { + "bbox": [ + 296, + 345, + 530, + 403 + ], + "type": "text", + "content": " is the radial kernel [37] to measure the distance between two features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "spans": [ + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "content": "Frame-level Diversity. In this paper, we aim to select the frames that are beneficial to the tracker training for annotation. For this purpose, besides considering the set-level discrepancy, we also need to explore frame diversity from different perspectives to infer the most representative and informative frames. Different from the semantic feature and spatial topology used to learn set-level discrepancy, finer-grained cues are desiderated to evaluate the potential contribution of each unlabeled frame. Specifically, besides the frame-wise spatial topology " + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "inline_equation", + "content": "G^{h}" + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "content": " and semantic feature " + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "inline_equation", + "content": "F^{f}" + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "content": ", we also consider finer-grained attribute characterizing the individuation of objects: the scale of bounding boxes " + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "inline_equation", + "content": "G^{b}" + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "content": ". Its construction process is similar to " + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "inline_equation", + "content": "G^{h}" + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "content": ", except that the identities representing different individuals are replaced by the width and height of the corresponding object as shown in Fig.3 (d), i.e., " + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "inline_equation", + "content": "G^{h} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "inline_equation", + "content": "G^{b} \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 296, + 404, + 530, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": "Intuitively, we argue that any one of the geometric features " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "G^{h}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "G^{b}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": " and semantic features " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "F^{f}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": " is the key factor to determine whether a frame contributes to the performance improvement of the tracker. Accordingly, we consider the novelty and prominence of each unlabeled frame " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "S_{frame} = \\{G^{h}, G^{b}, F^{f}\\}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": ", to form the frame-level diversity as action " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": ". The novelty of an unlabeled frame " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": " is reflected in the similarities between " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": " and the labeled frames in " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": ", because the unlabeled frame with a novel characteristic (such as new individuals, different object scales, and postures) usually has low similarities to labeled frames of " + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": ". And the promi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "9935" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": "nence of " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " lies in the difference between " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " and other unlabeled frames in " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": ", so we expect to select the most representative frame by recording the similarity distribution between each unlabeled frame and other frames in " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": ". Specifically, we introduce histogram-based representation to learn the novelty and prominence of each unlabeled frame. Taking " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " as an example, to represent the novelty of " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": ", we calculate cosine similarities between " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " on each heterogeneous feature " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "S_{frame}^{m}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " and then obtain the histogram " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{nov}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " based on cosine similarities. Similar to novelty learning, we learn prominence by recoding the histogram " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{pro}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " based on cosine similarities between " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": ". Finally, the action representation " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{H}_{nov}^{G^{h}},\\mathcal{H}_{nov}^{G^{b}},\\mathcal{H}_{nov}^{F^{f}},\\mathcal{H}_{pro}^{G^{h}},\\mathcal{H}_{pro}^{G^{b}},\\mathcal{H}_{pro}^{F^{f}}\\}" + }, + { + "bbox": [ + 45, + 106, + 281, + 296 + ], + "type": "text", + "content": ", which conduces to IFSN to effectively identify and select the informative and representative frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 244, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 244, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 244, + 314 + ], + "type": "text", + "content": "3.3. Informative Frame Selection Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 319, + 279, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 319, + 279, + 400 + ], + "spans": [ + { + "bbox": [ + 45, + 319, + 279, + 400 + ], + "type": "text", + "content": "Inspired by DQN [38], our lightweight IFSN follows an optimal policy that maximizes the potential performance gains. Considering the sampling efficiency, we propose a batch sampling method guided by frame clusters according to temporal clues. Meanwhile, to restrain the inferior performance caused by redundant frames, we further introduce the multi-frame cooperation strategy before batch sampling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "spans": [ + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": "Multi-frame Cooperation. There are many similar adjacent frames in tracking sequences, which easily leads to redundancy in the process of frame selection. To reduce redundant frames, we introduce a simple multi-frame cooperation strategy to model the neighborhood cooperation between the unlabeled frame " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": " and its nearest labeled frames. Specifically, we seek the nearest neighbors " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "x_{v}^{l}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "x_{w}^{l}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "X_{t}^{L}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": " and calculate their actions " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "a_{t}^{v}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "a_{t}^{w}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": ". Then we use the fixed-length compact representation " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "o_{t} = \\frac{1}{2}(a_{t}^{v} + a_{t}^{w})" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": " as an extra action for " + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 45, + 400, + 281, + 527 + ], + "type": "text", + "content": ". To this end, the selection made by IFSN becomes:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 533, + 279, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 533, + 279, + 550 + ], + "spans": [ + { + "bbox": [ + 104, + 533, + 279, + 550 + ], + "type": "interline_equation", + "content": "\\phi_ {t} = \\arg \\max _ {a _ {t} \\in A _ {t}} g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right), \\tag {2}", + "image_path": "6ce53dfb3c22d5a7521ff8a883f023ea54331c1ed1198d94fee44277ec7501a1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "spans": [ + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "inline_equation", + "content": "a_{t}\\in A_{t}" + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "text", + "content": " denotes the candidate action in " + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "inline_equation", + "content": "g(.)" + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "text", + "content": " is the lightweight IFSN network parameterized by " + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 45, + 555, + 279, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "spans": [ + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "text", + "content": "Batch Frame Selection. The structure of our designed lightweight IFSN consists of four linear layers, where a linear layer parameterized by " + }, + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "text", + "content": " is used to model the relation between each unlabeled frame and its nearest neighbors, and three linear layers parameterized by " + }, + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "inline_equation", + "content": "\\theta_{2\\sim 4}" + }, + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "text", + "content": " are employed to score each unlabeled frame, as described in Eq. (3), where " + }, + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "inline_equation", + "content": "cat[\\cdot]" + }, + { + "bbox": [ + 45, + 579, + 279, + 660 + ], + "type": "text", + "content": " denotes the concatenation operation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 666, + 279, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 666, + 279, + 679 + ], + "spans": [ + { + "bbox": [ + 76, + 666, + 279, + 679 + ], + "type": "interline_equation", + "content": "g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right) = \\theta_ {2 \\sim 4} \\left(c a t \\left[ \\theta_ {1} \\left(a _ {t}, o _ {t}\\right), a _ {t}, s _ {t} \\right]\\right), \\tag {3}", + "image_path": "ab7d490adf9b66f25ac07c1bebcf4ce3c763b2dddc5c7921443dcb06bbbb12f8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "text", + "content": "Intuitively, each unlabeled frame in " + }, + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "inline_equation", + "content": "X_{t}^{U}" + }, + { + "bbox": [ + 45, + 682, + 279, + 730 + ], + "type": "text", + "content": " can obtain the corresponding score by IFSN. However, sampling a single frame in each iteration by Eq. (2) to query annotation is inefficient. In this paper, we propose a batch sampling method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": "guided by frame clusters to sample " + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": " frames in each iteration. Based on the intuition that a larger difference exists between frames with longer time intervals, we divide the unlabeled frames into " + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": " clusters in chronological order and select the frame for annotation in each cluster independently, which is equivalent to labeling a frame in parallel with " + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": " annotators. In this case, " + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": " in Eq. (2) is extended to consist of " + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": " independent sub-selections " + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "inline_equation", + "content": "\\{\\theta_t^n\\}_{n=1}^N" + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": ", each with a restricted space " + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "inline_equation", + "content": "A_t^n" + }, + { + "bbox": [ + 296, + 106, + 530, + 212 + ], + "type": "text", + "content": ", as described in Eq. (4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 218, + 530, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 218, + 530, + 236 + ], + "spans": [ + { + "bbox": [ + 354, + 218, + 530, + 236 + ], + "type": "interline_equation", + "content": "\\phi_ {t} ^ {n} = \\arg \\max _ {a _ {t} \\in A _ {t} ^ {n}} g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right), \\tag {4}", + "image_path": "2a732de7c658bc536949c783d6f284c4a74b3fd0ac7c7aa11d9799b27812f8ce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 245, + 528, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 245, + 528, + 256 + ], + "spans": [ + { + "bbox": [ + 296, + 245, + 528, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 245, + 528, + 256 + ], + "type": "inline_equation", + "content": "a_{t}\\in A_{t}^{n}" + }, + { + "bbox": [ + 296, + 245, + 528, + 256 + ], + "type": "text", + "content": " denotes the candidate action in " + }, + { + "bbox": [ + 296, + 245, + 528, + 256 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 296, + 245, + 528, + 256 + ], + "type": "text", + "content": " cluster." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "spans": [ + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "text", + "content": "IFSN Optimization. In our method, we compute the reward on a subset " + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "inline_equation", + "content": "X^R" + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "text", + "content": ", which is used to evaluate the benefit of selected frames to the tracker and optimize the frame selection network IFSN. To evaluate the performance of the tracker more comprehensively, we utilize MOTA [5] and IDF1 [19] as performance metrics to calculate the reward, where MOTA is a comprehensive metric to evaluate the detection performance and IDF1 tends to estimate the tracking ability of a tracker. As described in Eq. (5), the reward " + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "inline_equation", + "content": "r_{t+1}" + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "text", + "content": " is defined as the difference of the performance metrics between " + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "inline_equation", + "content": "f_{t+1}" + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 296, + 257, + 530, + 385 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 391, + 530, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 391, + 530, + 405 + ], + "spans": [ + { + "bbox": [ + 339, + 391, + 530, + 405 + ], + "type": "interline_equation", + "content": "r _ {t} = \\left(e _ {t + 1} ^ {\\text {m o t a}} + e _ {t + 1} ^ {i d f 1}\\right) - \\left(e _ {t} ^ {\\text {m o t a}} + e _ {t} ^ {i d f 1}\\right), \\tag {5}", + "image_path": "4171a2e44170163f72ddb0dcb2bf35be22e3caf2e579af7b4d0246b3cff614f5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "spans": [ + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "inline_equation", + "content": "e_{t+1}^{mota}" + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "inline_equation", + "content": "e_{t+1}^{idf1}" + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "content": " are MOTA and IDF1 metrics computed by " + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "inline_equation", + "content": "f_{t+1}" + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "inline_equation", + "content": "e_t^{mota}" + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "inline_equation", + "content": "e_t^{idf1}" + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "content": " are evaluated by " + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 296, + 413, + 530, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 441, + 530, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 441, + 530, + 510 + ], + "spans": [ + { + "bbox": [ + 296, + 441, + 530, + 510 + ], + "type": "text", + "content": "With the reward " + }, + { + "bbox": [ + 296, + 441, + 530, + 510 + ], + "type": "inline_equation", + "content": "r_{t + 1}" + }, + { + "bbox": [ + 296, + 441, + 530, + 510 + ], + "type": "text", + "content": ", we can optimize IFSN to infer the most informative frames to maximize the reward, leading to improved tracker performance during each active learning iteration. More specifically, we train IFSN in a double DQN formulation [38] by optimizing the loss based on the temporal difference error [35] as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 352, + 516, + 530, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 516, + 530, + 531 + ], + "spans": [ + { + "bbox": [ + 352, + 516, + 530, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I F S N}} = \\left(\\hat {y} _ {t} - g \\left(s _ {t}, a _ {t}, o _ {t}; \\theta\\right)\\right) ^ {2}, \\tag {6}", + "image_path": "932976110dfd33a8b90986689577ccb502fa31b1d91ea524a609324ceecb315d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 549, + 530, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 549, + 530, + 561 + ], + "spans": [ + { + "bbox": [ + 344, + 549, + 530, + 561 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t} = r _ {t + 1} + \\gamma g \\left(s _ {t + 1}, a _ {t + 1}, o _ {t + 1}; \\hat {\\theta}\\right), \\tag {7}", + "image_path": "76def1e9b58b2865cf4b65c70a3fc1b31bf2a91d02d12f018d50d7f8e56e77ae.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "spans": [ + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "text", + "content": " are parameters of the IFSN policy network and off-policy network following the setting of double DQN [38]. " + }, + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 296, + 564, + 530, + 599 + ], + "type": "text", + "content": " denotes a discount factor." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 599, + 530, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 599, + 530, + 646 + ], + "spans": [ + { + "bbox": [ + 296, + 599, + 530, + 646 + ], + "type": "text", + "content": "With the described diversified informative representation in Section 3.2 and informative frame selection learning in this section, we introduce our entire HD-AMOT as illustrated in Alg. 1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 657, + 371, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 657, + 371, + 671 + ], + "spans": [ + { + "bbox": [ + 297, + 657, + 371, + 671 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 297, + 676, + 420, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 676, + 420, + 689 + ], + "spans": [ + { + "bbox": [ + 297, + 676, + 420, + 689 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": "Datasets. We evaluate our HD-AMOT on MOT15 [18], MOT17 [23], MOT20 [12], and Dancetrack [34]. MOT15 and MOT17 are widely-used MOT datasets, and they are" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "9936" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 48, + 122, + 260, + 336 + ], + "blocks": [ + { + "bbox": [ + 52, + 109, + 201, + 120 + ], + "lines": [ + { + "bbox": [ + 52, + 109, + 201, + 120 + ], + "spans": [ + { + "bbox": [ + 52, + 109, + 201, + 120 + ], + "type": "text", + "content": "Algorithm 1: HD-AMOT Algorithm." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "lines": [ + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "Input: The sequence " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": ", initial tracker " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "f_{init}" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " and IFSN " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "g_{init}" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": ", AL procedure rounds " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": ". \n1 " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{init}, \\mathbf{X}^{R} \\gets" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " RandomPartition " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "(X)" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "; \n2 " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{init}^{U}, \\mathbf{X}_{init}^{L} \\gets" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " UniformPartition " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "(X_{init})" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "; \n3 " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0}^{L} \\gets \\mathbf{X}_{init}^{L}, f_{0} \\gets" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " update " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "(g_{init}, \\mathbf{X}_{init}^{L})" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "; \n4 while not done do // Episodes \n5 for " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "t = Z - 1" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " do // AL procedure \n6 Learn the diversified representation to form the state and action space; \n7 Use IFSN to select frames by Eq. (4); \n8 Annotate the selected " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t}^{S}" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "; \n9 Update " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t}^{U}, \\mathbf{X}_{t}^{L}" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": ": \n10 " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{t + 1}^{U} \\gets \\mathbf{X}_{t}^{U} \\setminus \\mathbf{X}_{t}^{S}, \\mathbf{X}_{t + 1}^{L} \\gets \\mathbf{X}_{t}^{L} \\cup \\mathbf{X}_{t}^{S}" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "; \n11 " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "f_{t + 1} \\gets" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " update " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "(f_{t}, \\mathbf{X}_{t + 1}^{L})" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "; \n12 Calculate the reward on " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{R}" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": " by Eq. (5); \n13 end \n14 Update IFSN following Eq. (6); \n15 Reset: " + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0}^{L} \\gets \\mathbf{X}_{init}^{L}, f_{0} \\gets f_{t + 1}" + }, + { + "bbox": [ + 48, + 122, + 260, + 336 + ], + "type": "text", + "content": "; \n16 end" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 45, + 366, + 279, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 366, + 279, + 447 + ], + "spans": [ + { + "bbox": [ + 45, + 366, + 279, + 447 + ], + "type": "text", + "content": "challenging because of the large variety in object pose, lighting, viewpoint, etc. In contrast to MOT15 and MOT17, MOT20 addresses the challenge of highly congested scenarios, where the pedestrian density can surge to as high as 246 individuals per frame. Dancetrack is a large-scale MOT dataset covering scenarios with low distinguishability of object appearance and diverse non-linear motion patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 448, + 279, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 448, + 279, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 448, + 279, + 542 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the performance of one-stage MOT, the standard metrics MOTA [5] and IDF1 [19] are adopted for evaluation, which indicate the detection and association performance of one-stage MOT. Specifically, MOTA combines three error sources including false positives, missed targets, and identity switches, and IDF1 is the ratio of correctly identified detections over the average number of ground-truth and computed detections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": "Implementation Details. In this paper, we adopt the popular one-stage MOT method FairMOT [48] as the MOT tracker. Note that the propose algorithm is suitable for any one-stage MOT tracker. The input frames are resized to " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "1088 \\times 608" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ". We set the learning rate of our IFSN to 1e-4 and the discount factor " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": " in Eq. (7) to 0.9, and the number of local features " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": " is set to 9 according to experiments. For each dataset, we randomly select a small sequence in each video at a ratio of 0.1 to form a reward subset " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "X^R" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ". In the remaining frames, a small amount of frames is uniformly sampled as the initially labeled subset " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "X_{init}^{L}" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": " to initialize the MOT tracker " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ", and the size is set to 100. For the semantic information, we extract the feature map from the last convolutional layer of ResNet34 as the global semantic feature " + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "inline_equation", + "content": "F^{f} \\in \\mathbb{R}^{512 \\times 19 \\times 34}" + }, + { + "bbox": [ + 45, + 543, + 279, + 729 + ], + "type": "text", + "content": ". For the geometric clues, we utilize the center coordinates and identities of bounding" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 299, + 126, + 530, + 250 + ], + "blocks": [ + { + "bbox": [ + 297, + 103, + 530, + 123 + ], + "lines": [ + { + "bbox": [ + 297, + 103, + 530, + 123 + ], + "spans": [ + { + "bbox": [ + 297, + 103, + 530, + 123 + ], + "type": "text", + "content": "Table 1. Ablation study on the design of the diversified informative representation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 299, + 126, + 530, + 250 + ], + "lines": [ + { + "bbox": [ + 299, + 126, + 530, + 250 + ], + "spans": [ + { + "bbox": [ + 299, + 126, + 530, + 250 + ], + "type": "table", + "html": "
RepresentationSettingMOTA (%)IDF1 (%)
Set-level DiscrepancyState w/o Gh,h1-h961.7364.95
State w/o Ff,f1-f961.4964.12
State w/o Gh, Ff62.7465.21
State w/o Gh1-h9, Ff1-f962.9264.55
Frame-level DiversityAction w/o Hgh62.8366.61
Action w/o HGb62.4966.23
Action w/o HFf61.7864.95
Action w/o H* nov61.2563.41
Action w/o H* pro62.2864.16
AllHD-AMOT63.2366.47
", + "image_path": "361d2d3e6f8a23bf6208e7df311d58407d1e99ebdc25193aa462d4953114e772.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "spans": [ + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "text", + "content": "boxes obtained by the MOT tracker to construct the global spatial topology " + }, + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "inline_equation", + "content": "G^{h} \\in \\mathbb{R}^{152 \\times 272}" + }, + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "text", + "content": ", and generate the object scale matrix " + }, + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "inline_equation", + "content": "G^{b} \\in \\mathbb{R}^{152 \\times 272 \\times 2}" + }, + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "text", + "content": " by the box size. In addition, the number of clusters " + }, + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 263, + 530, + 334 + ], + "type": "text", + "content": " in the batch sampling is set to 12 for MOT15, MOT17, and MOT20, and is set to 24 for Dancetrack as it is much larger than other datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 343, + 388, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 343, + 388, + 356 + ], + "spans": [ + { + "bbox": [ + 297, + 343, + 388, + 356 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 362, + 530, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 362, + 530, + 419 + ], + "spans": [ + { + "bbox": [ + 296, + 362, + 530, + 419 + ], + "type": "text", + "content": "In this section, to analyze the impacts of different components, we conduct ablation studies on the popular dataset MOT17 following the division strategy in [43]. We run our experiments 5 times and report the mean performance for all ablation studies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "spans": [ + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": "Effect of Diversified Informative Representation. We perform an ablation study with a " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": " annotation budget to evaluate the contribution of our diversified informative representation. As IFSN relies on the state (set-level discrepancy) to decide the frame sampling policy, we investigate the influence of the set-level discrepancy by removing different components individually. Moreover, we also analyze the effect of the action design (frame-level diversity). As shown in Tab. 1, our complete diversified informative representation gives the best performance in terms of MOTA and is only slightly inferior to Action w/o " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{G^h}" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "0.14\\%" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": " on IDF1. We speculate that the scale information in " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{G^b}" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": " implicitly indicates the identity and topology clues of " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{G^h}" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": ", so the introduction of " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{G^b}" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": " weakens the role of " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{G^h}" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": ", making the MOTA of Action w/o " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^{G^h}" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": " close to HD-AMOT and the IDF1 slightly higher. Furthermore, removing the components of set-level discrepancy or frame-level diversity in the diversified informative representation degrades the tracking performance. The largest decrease of MOTA and IDF1 occurs when the score of maximum similarity with labeled frames Action w/o " + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{nov}^*" + }, + { + "bbox": [ + 296, + 421, + 531, + 692 + ], + "type": "text", + "content": " is removed, which further verifies the effectiveness of heterogeneous cues evaluating the novel characteristic of unlabeled frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 694, + 530, + 730 + ], + "type": "text", + "content": "Effect of Informative Frame Selection Network. We further validate the performance of our designed IFSN in Tab. 2. We first consider performing the cluster-based batch" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "9937" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 123, + 273, + 190 + ], + "blocks": [ + { + "bbox": [ + 46, + 103, + 278, + 122 + ], + "lines": [ + { + "bbox": [ + 46, + 103, + 278, + 122 + ], + "spans": [ + { + "bbox": [ + 46, + 103, + 278, + 122 + ], + "type": "text", + "content": "Table 2. Ablation study on the different settings of the Informative Frame Selection Network." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 123, + 273, + 190 + ], + "lines": [ + { + "bbox": [ + 52, + 123, + 273, + 190 + ], + "spans": [ + { + "bbox": [ + 52, + 123, + 273, + 190 + ], + "type": "table", + "html": "
ModelMulti-frame CooperationBatch SamplingMOTA (%)↑IDF1 (%)↑
Bât.Clu.
A×✓✓62.7665.52
B✓××63.0765.62
C✓✓×62.1864.55
D✓✓✓63.2366.47
", + "image_path": "6713a37a99244cf5ba17408b5b6fcc30d11382dc2ff1b1d47fc762363faab943.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 194, + 162, + 331 + ], + "blocks": [ + { + "bbox": [ + 48, + 194, + 162, + 331 + ], + "lines": [ + { + "bbox": [ + 48, + 194, + 162, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 162, + 331 + ], + "type": "image", + "image_path": "ddc0ac4ffd87d1abbf0f5331a72de198fc7e51f21bcaabbea4c00ab7907d31c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 340, + 278, + 362 + ], + "lines": [ + { + "bbox": [ + 45, + 340, + 278, + 362 + ], + "spans": [ + { + "bbox": [ + 45, + 340, + 278, + 362 + ], + "type": "text", + "content": "Figure 4. Performance comparison evaluated by MOTA and IDF1 metrics with increasing budgets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 162, + 195, + 276, + 331 + ], + "blocks": [ + { + "bbox": [ + 162, + 195, + 276, + 331 + ], + "lines": [ + { + "bbox": [ + 162, + 195, + 276, + 331 + ], + "spans": [ + { + "bbox": [ + 162, + 195, + 276, + 331 + ], + "type": "image", + "image_path": "43fc21309b8fe98f1e2f9dddb4ffd954707319f93751ded78b95648d8990a2ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "spans": [ + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "text", + "content": "sampling without multi-frame cooperation (Model A). Then we analyze the usage of selecting a single frame in each AL iteration (Model B). In addition, we also construct the Model C as IFSN to select a batch of frames in one shot, where the frames with " + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "text", + "content": " highest scores are sampled. Finally, we present the performance of our cluster-based batch sampling strategy (Model D). By comparing the results of Model A, Model D brings performance gains of " + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "inline_equation", + "content": "0.47\\%" + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "inline_equation", + "content": "0.95\\%" + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "text", + "content": " in terms of MOTA and IDF1 metrics, which proves that multi-frame cooperation plays a positive role in the active learning of one-stage MOT. In contrast, Model C gives the worst performance, which selects multiple frames according to the " + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 367, + 279, + 635 + ], + "type": "text", + "content": " highest scores. We argue that frames with higher scores may be similar samples and close to each other, which easily leads to inefficiency in batch sampling. To this end, cluster-based batch sampling is introduced, as it can avoid selecting adjacent frames with higher scores and learn to sample with better coverage of the underlying distribution. Notably, selecting a single frame at each iteration in Model B also provides a competitive performance but still tends to be inferior to our method Model D, and the time cost of selecting a single frame per iteration is much higher than our cluster-based batch sampling strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": "Analysis of Different Budgets. We report the performance of our method, random sampling, uniform sampling, and other existing AL methods such as Coreset [31], CDAL [1] under different budgets, as shown in Fig. 4. For the IDF1 metric, ours is slightly inferior to the uniform sampling when the budget is " + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 45, + 636, + 279, + 730 + ], + "type": "text", + "content": ". We argue that uniform sampling can select informative frames for annotation if a small number of labeled frames is required. However," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "text", + "content": "when the budget increases, there may be redundancy in the newly acquired frames by uniform sampling, and training with these labels does not provide more additional information, resulting in the stagnation or even decline of the performance of IDF1. Notably, our HD-AMOT with the " + }, + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "text", + "content": " budget achieves similar performance as other methods with the " + }, + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "text", + "content": " budget. It reveals that ours can save around " + }, + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 296, + 106, + 530, + 235 + ], + "type": "text", + "content": " labeling efforts on average compared to other sampling methods. We attribute this to our learned diversified informative clues and fine-grained object representation, which can effectively improve the tracking performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 244, + 485, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 244, + 485, + 256 + ], + "spans": [ + { + "bbox": [ + 296, + 244, + 485, + 256 + ], + "type": "text", + "content": "4.3. Comparison with other AL Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 262, + 530, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 262, + 530, + 296 + ], + "spans": [ + { + "bbox": [ + 296, + 262, + 530, + 296 + ], + "type": "text", + "content": "To further verify the generalization and effectiveness, we evaluate the proposed framework HD-AMOT on four diversified MOT datasets in this section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 298, + 530, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 298, + 530, + 448 + ], + "spans": [ + { + "bbox": [ + 296, + 298, + 530, + 448 + ], + "type": "text", + "content": "Results on MOT15. Tab. 3 lists the mean values of MOTA and IDF1 on MOT15, and Fig. 5 reveals the corresponding standard deviation of MOTA and IDF1. We can observe that HD-AMOT outperforms all methods by a clear margin, i.e., AL methods used in other tasks (CDAL [1] and Coreset [31]) stiffly applied to one-stage MOT can lead to obvious performance degradation. Notably, although Coreset achieves the second-best performance on MOTA and IDF1, it is clear from the standard deviations in Fig. 5 that Coreset has limited stability. Moreover, the performance of Coreset on other datasets (MOT17, MOT20, and Dancetrack) is not ideal, which indicates that the direct application of Coreset in one-stage MOT is unstable." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 449, + 530, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 449, + 530, + 635 + ], + "spans": [ + { + "bbox": [ + 296, + 449, + 530, + 635 + ], + "type": "text", + "content": "Results on MOT17. MOT17 is a challenging dataset with more crowded scenarios, different viewpoints, camera motions, and weather conditions. As shown in Tab. 3, although uniform sampling achieves the best IDF1 on MOT17 and our HD-AMOT performs slightly lower than it by a margin of " + }, + { + "bbox": [ + 296, + 449, + 530, + 635 + ], + "type": "inline_equation", + "content": "0.18\\%" + }, + { + "bbox": [ + 296, + 449, + 530, + 635 + ], + "type": "text", + "content": ", uniform sampling may fail to capture diverse visual patterns in other datasets due to the lack of flexibility and adaptability. For example, in MOT15, ours outperforms uniform sampling by all performance metrics, which reflects the generalization ability of ours for one-stage MOT. Moreover, besides object association, the object detection ability should be considered in one-stage MOT. Our method, which benefits from learning the sampling policy driven by tracking performance metrics directly from data, significantly outperforms uniform sampling by " + }, + { + "bbox": [ + 296, + 449, + 530, + 635 + ], + "type": "inline_equation", + "content": "0.84\\%" + }, + { + "bbox": [ + 296, + 449, + 530, + 635 + ], + "type": "text", + "content": " in terms of the detection accuracy metric MOTA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "content": "Results on MOT20. We also use the large-scale dataset MOT20 to demonstrate the scalability of our HD-AMOT as shown in Tab. 3. In contrast to the aforementioned datasets, MOT20 is a recently released dataset that features more densely populated scenes. We can observe that ours still achieves the best performance, which demonstrates that our HD-AMOT can infer informative frames even on a crowded and complicated dataset." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "9938" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 147, + 279, + 361 + ], + "blocks": [ + { + "bbox": [ + 45, + 102, + 279, + 146 + ], + "lines": [ + { + "bbox": [ + 45, + 102, + 279, + 146 + ], + "spans": [ + { + "bbox": [ + 45, + 102, + 279, + 146 + ], + "type": "text", + "content": "Table 3. Mean values of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack. Best and second-best results are shown in red and black, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 147, + 279, + 361 + ], + "lines": [ + { + "bbox": [ + 47, + 147, + 279, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 279, + 361 + ], + "type": "table", + "html": "
DatasetMethodMOTA(%) (↑)IDF1(%) (↑)
MOT15Tracker w/ Random53.2859.61
Tracker w/ Uniform54.0959.50
Tracker w/ CDAL [1]53.1358.85
Tracker w/ Coreset [31]54.7060.29
Tracker w/ HD-AMOT55.2960.64
MOT17Tracker w/ Random70.1169.24
Tracker w/ Uniform70.4070.05
Tracker w/ CDAL [1]69.1068.71
Tracker w/ Coreset [31]69.4468.61
Tracker w/ HD-AMOT71.2469.87
MOT20Tracker w/ Random58.2165.37
Tracker w/ Uniform58.6565.87
Tracker w/ CDAL [1]57.5965.33
Tracker w/ Coreset [31]56.6664.86
Tracker w/ HD-AMOT59.2466.41
DancetrackTracker w/ Random80.4940.67
Tracker w/ Uniform81.1141.03
Tracker w/ CDAL [1]81.3641.16
Tracker w/ Coreset [31]81.1340.47
Tracker w/ HD-AMOT81.9441.93
", + "image_path": "25aae838a9b7812a88da3698f8a0abc08b75fa77760fa130b13dbd6e01bfd4d8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 367, + 163, + 496 + ], + "blocks": [ + { + "bbox": [ + 47, + 367, + 163, + 496 + ], + "lines": [ + { + "bbox": [ + 47, + 367, + 163, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 163, + 496 + ], + "type": "image", + "image_path": "b6e6eaebe838bfb20171ae117726a260ef5830c3ba1f44b309db6101f3cd1960.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 503, + 279, + 536 + ], + "lines": [ + { + "bbox": [ + 45, + 503, + 279, + 536 + ], + "spans": [ + { + "bbox": [ + 45, + 503, + 279, + 536 + ], + "type": "text", + "content": "Figure 5. Standard deviations of MOTA and IDF1 metrics for different active learning methods on the test sets of four datasets including MOT15, MOT17, MOT20, and Dancetrack." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 163, + 367, + 278, + 495 + ], + "blocks": [ + { + "bbox": [ + 163, + 367, + 278, + 495 + ], + "lines": [ + { + "bbox": [ + 163, + 367, + 278, + 495 + ], + "spans": [ + { + "bbox": [ + 163, + 367, + 278, + 495 + ], + "type": "image", + "image_path": "17f51011c2492d07fcf4d65ce23164ca53a45b5507b2e6850884de454d951d64.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 542, + 279, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 542, + 279, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 542, + 279, + 693 + ], + "type": "text", + "content": "Results on Dancetrack. Dancetrack is a larger and more challenging dataset for group dancing scenes, where humans have similar appearance, diverse motion, and extreme articulation. We provide the performance evaluation results on the large-scale Dancetrack in Tab. 3. It can be observed that our HD-AMOT still achieves the best MOTA and IDF1 performance. Notably, from the perspective of sampling stability in Fig. 5, although uniform sampling can obtain a low standard deviation of MOTA, it leads to a very high standard deviation of IDF1 due to failure to capture informative frames that are helpful to object association. By contrast, our proposed HD-AMOT can maintain stable MOTA and IDF1 performance in multiple samplings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "Furthermore, we give the performance comparison between our HD-AMOT with a " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " budget and full supervision in Fig. 6. Our active learning framework achieves sim" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 299, + 105, + 418, + 236 + ], + "blocks": [ + { + "bbox": [ + 299, + 105, + 418, + 236 + ], + "lines": [ + { + "bbox": [ + 299, + 105, + 418, + 236 + ], + "spans": [ + { + "bbox": [ + 299, + 105, + 418, + 236 + ], + "type": "image", + "image_path": "03d6db03f7fd845e9bba4207a329949565aeda57edb72df38eeacc65e52e9436.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 243, + 530, + 275 + ], + "lines": [ + { + "bbox": [ + 296, + 243, + 530, + 275 + ], + "spans": [ + { + "bbox": [ + 296, + 243, + 530, + 275 + ], + "type": "text", + "content": "Figure 6. Performance comparison between our HD-AMOT with a " + }, + { + "bbox": [ + 296, + 243, + 530, + 275 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 296, + 243, + 530, + 275 + ], + "type": "text", + "content": " budget and full supervision on MOT15, MOT17, MOT20, and Dancetrack datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 418, + 105, + 528, + 236 + ], + "blocks": [ + { + "bbox": [ + 418, + 105, + 528, + 236 + ], + "lines": [ + { + "bbox": [ + 418, + 105, + 528, + 236 + ], + "spans": [ + { + "bbox": [ + 418, + 105, + 528, + 236 + ], + "type": "image", + "image_path": "d32978e371dd9a232d66a1b70dd66856408330330e2530a7cf5e470a74a6a087.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 277, + 530, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 277, + 530, + 428 + ], + "spans": [ + { + "bbox": [ + 296, + 277, + 530, + 428 + ], + "type": "text", + "content": "ilar or even higher performance compared with the tracker with fully-supervised training. For instance, we exceed the full supervision by a gain of " + }, + { + "bbox": [ + 296, + 277, + 530, + 428 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 296, + 277, + 530, + 428 + ], + "type": "text", + "content": " on MOTA and " + }, + { + "bbox": [ + 296, + 277, + 530, + 428 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 296, + 277, + 530, + 428 + ], + "type": "text", + "content": " on IDF1 for MOT15. These results implicitly prove that there are many similar frames in MOT sequences, and these similar frames bring limited progress in tracking performance. Our proposed HD-AMOT learns an optimal sampling policy by the heterogeneous diversity representation and a Markov decision process, which can infer informative frames that can most benefit the tracker, thus effectively improving the performance of the tracker. In summary, the proposed method can provide a consistent gain for the MOT tracker across datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 441, + 368, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 441, + 368, + 453 + ], + "spans": [ + { + "bbox": [ + 297, + 441, + 368, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 461, + 530, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 461, + 530, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 461, + 530, + 612 + ], + "type": "text", + "content": "In this paper, we propose a novel framework HD-AMOT to investigate the annotation budget for one-stage MOT. It is formulated as an MDP and learns a frame sampling policy driven by the reward that directly relates to the performance of the tracker. HD-AMOT learns the diversified informative representation of frames and achieves effective cluster-based batch frame selection guided by multi-frame cooperation. We conduct extensive ablation studies to verify the design of our framework and compare our performance with random, uniform sampling, and several existing active learning works on four datasets including MOT15, MOT17, MOT20, and Dancetrack. The experimental results demonstrate the efficacy and effectiveness of our HD-AMOT." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the Fundamental Research Funds for the Central Universities of China (2022JBMC009), the Natural Science Foundation of China (61972027, 62006244), and the Beijing Municipal Natural Science Foundation (Grant No. 4212041). This work is also partially supported by Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and Young Elite Scientist Sponsorship Program of Beijing Association for Science and Technology (BYESS2021178)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 298, + 756 + ], + "type": "text", + "content": "9939" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 102, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 52, + 124, + 279, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 124, + 279, + 155 + ], + "spans": [ + { + "bbox": [ + 52, + 124, + 279, + 155 + ], + "type": "text", + "content": "[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In ECCV, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 157, + 279, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 157, + 279, + 188 + ], + "spans": [ + { + "bbox": [ + 52, + 157, + 279, + 188 + ], + "type": "text", + "content": "[2] Favyen Bastani, Songtao He, and Samuel Madden. Self-supervised multi-object tracking with cross-input consistency. In NeurIPS, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 190, + 279, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 190, + 279, + 221 + ], + "spans": [ + { + "bbox": [ + 52, + 190, + 279, + 221 + ], + "type": "text", + "content": "[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In CVPR, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 222, + 278, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 222, + 278, + 243 + ], + "spans": [ + { + "bbox": [ + 52, + 222, + 278, + 243 + ], + "type": "text", + "content": "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 244, + 279, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 244, + 279, + 286 + ], + "spans": [ + { + "bbox": [ + 52, + 244, + 279, + 286 + ], + "type": "text", + "content": "[5] Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 287, + 278, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 287, + 278, + 319 + ], + "spans": [ + { + "bbox": [ + 52, + 287, + 278, + 319 + ], + "type": "text", + "content": "[6] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: Multi-object tracking with memory. In CVPR, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 319, + 278, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 319, + 278, + 351 + ], + "spans": [ + { + "bbox": [ + 52, + 319, + 278, + 351 + ], + "type": "text", + "content": "[7] Lile Cai, Xun Xu, Jun Hao Liew, and Chuan Sheng Foo. Revisiting superpixels for active learning in semantic segmentation with realistic annotation costs. In CVPR, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 351, + 278, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 351, + 278, + 383 + ], + "spans": [ + { + "bbox": [ + 52, + 351, + 278, + 383 + ], + "type": "text", + "content": "[8] Razvan Caramalau, Binod Bhattacharai, and Tae-Kyun Kim. Sequential graph convolutional network for active learning. In CVPR, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 384, + 278, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 384, + 278, + 415 + ], + "spans": [ + { + "bbox": [ + 52, + 384, + 278, + 415 + ], + "type": "text", + "content": "[9] Arantxa Casanova, Pedro O. Pinheiro, Negar Rostamzadeh, and Christopher J. Pal. Reinforced active learning for image segmentation. In ICLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 416, + 278, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 278, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 278, + 448 + ], + "type": "text", + "content": "[10] Jiwong Choi, Ismail Elezi, Hyuk-Jae Lee, Clement Farabet, and Jose M Alvarez. Active learning for deep object detection via probabilistic modeling. In CVPR, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 449, + 279, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 279, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 279, + 481 + ], + "type": "text", + "content": "[11] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In CVPR, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 481, + 279, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 279, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 279, + 533 + ], + "type": "text", + "content": "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 535, + 279, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 279, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 279, + 566 + ], + "type": "text", + "content": "[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 567, + 279, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 279, + 609 + ], + "type": "text", + "content": "[14] Jiawei He, Zehao Huang, Naiyan Wang, and Zhaoxiang Zhang. Learnable graph matching: Incorporating graph partitioning with deep feature learning for multiple object tracking. In CVPR, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 611, + 278, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 278, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 278, + 632 + ], + "type": "text", + "content": "[15] Sheng-Jun Huang, Nengneng Gao, and Songcan Chen. Multi-instance multi-label active learning. In *IJCAI*, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 633, + 278, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 278, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 278, + 665 + ], + "type": "text", + "content": "[16] Ming-xin Jiang, Chao Deng, Zhi-geng Pan, Lan-fang Wang, and Xing Sun. Multiobject tracking in videos based on lstm and deep reinforcement learning. Complexity, 2018, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 665, + 278, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 278, + 697 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 278, + 697 + ], + "type": "text", + "content": "[17] Chanho Kim, Li Fuxin, Mazen Alotaibi, and James M Rehg. Discriminative appearance modeling with multi-track pooling for real-time multi-object tracking. In CVPR, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 698, + 278, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 698, + 278, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 698, + 278, + 729 + ], + "type": "text", + "content": "[18] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942, 2015." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 729 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "text", + "content": "[19] Yuan Li, Chang Huang, and Ram Nevatia. Learning to associate: Hybridboosted multi-target tracker for crowded scene. In CVPR, 2009." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 139, + 530, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 139, + 530, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 139, + 530, + 171 + ], + "type": "text", + "content": "[20] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In ICCV, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 172, + 530, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 172, + 530, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 172, + 530, + 203 + ], + "type": "text", + "content": "[21] Zhao-Yang Liu and Sheng-Jun Huang. Active sampling for open-set classification without initial annotation. In AAAI, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 206, + 530, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 206, + 530, + 237 + ], + "spans": [ + { + "bbox": [ + 299, + 206, + 530, + 237 + ], + "type": "text", + "content": "[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 238, + 530, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 238, + 530, + 270 + ], + "spans": [ + { + "bbox": [ + 299, + 238, + 530, + 270 + ], + "type": "text", + "content": "[23] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv:1603.00831, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 271, + 530, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 271, + 530, + 313 + ], + "spans": [ + { + "bbox": [ + 299, + 271, + 530, + 313 + ], + "type": "text", + "content": "[24] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. In NeurIPS, 2013." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 314, + 530, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 314, + 530, + 346 + ], + "spans": [ + { + "bbox": [ + 299, + 314, + 530, + 346 + ], + "type": "text", + "content": "[25] Bo Pang, Yizhuo Li, Yifan Zhang, Muchen Li, and Cewu Lu. Tubetk: Adopting tubes to track multi-object in a one-step training model. In CVPR, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 347, + 530, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 347, + 530, + 379 + ], + "spans": [ + { + "bbox": [ + 299, + 347, + 530, + 379 + ], + "type": "text", + "content": "[26] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 380, + 530, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 380, + 530, + 433 + ], + "spans": [ + { + "bbox": [ + 299, + 380, + 530, + 433 + ], + "type": "text", + "content": "[27] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In ECCV, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 435, + 530, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 435, + 530, + 476 + ], + "spans": [ + { + "bbox": [ + 299, + 435, + 530, + 476 + ], + "type": "text", + "content": "[28] Yu Qiao, Jincheng Zhu, Chengjiang Long, Zeyao Zhang, Yuxin Wang, Zhenjun Du, and Xin Yang. Cpral: Collaborative panoptic-regional active learning for semantic segmentation. In AAAI, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 478, + 530, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 478, + 530, + 510 + ], + "spans": [ + { + "bbox": [ + 299, + 478, + 530, + 510 + ], + "type": "text", + "content": "[29] Liangliang Ren, Jiwen Lu, Zifeng Wang, Qi Tian, and Jie Zhou. Collaborative deep reinforcement learning for multi-object tracking. In ECCV, 2018." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 511, + 530, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 511, + 530, + 543 + ], + "spans": [ + { + "bbox": [ + 299, + 511, + 530, + 543 + ], + "type": "text", + "content": "[30] Idoia Ruiz, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Joan Serrat. Weakly supervised multi-object tracking and segmentation. In WACV, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 543, + 530, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 543, + 530, + 574 + ], + "spans": [ + { + "bbox": [ + 299, + 543, + 530, + 574 + ], + "type": "text", + "content": "[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In ICLR, 2018." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 299, + 576, + 530, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 576, + 530, + 607 + ], + "spans": [ + { + "bbox": [ + 299, + 576, + 530, + 607 + ], + "type": "text", + "content": "[32] Megh Shukla and Shuaib Ahmed. A mathematical analysis of learning loss for active learning in regression. In CVPR, 2021." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 299, + 609, + 530, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 609, + 530, + 640 + ], + "spans": [ + { + "bbox": [ + 299, + 609, + 530, + 640 + ], + "type": "text", + "content": "[33] Yawar Siddiqui, Julien Valentin, and Matthias Nießner. Viewal: Active learning with viewpoint entropy for semantic segmentation. In CVPR, 2020." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 299, + 642, + 530, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 642, + 530, + 683 + ], + "spans": [ + { + "bbox": [ + 299, + 642, + 530, + 683 + ], + "type": "text", + "content": "[34] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In CVPR, 2022." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 299, + 685, + 530, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 685, + 530, + 707 + ], + "spans": [ + { + "bbox": [ + 299, + 685, + 530, + 707 + ], + "type": "text", + "content": "[35] Richard S Sutton. Learning to predict by the methods of temporal differences. Machine learning, 3(1):9-44, 1988." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 299, + 708, + 530, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 708, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 708, + 530, + 729 + ], + "type": "text", + "content": "[36] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018." + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 747, + 298, + 756 + ], + "type": "text", + "content": "9940" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 719 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 138 + ], + "type": "text", + "content": "[37] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. In NeurIPS, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 279, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 279, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 279, + 161 + ], + "type": "text", + "content": "[38] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In AAAI, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 279, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 279, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 279, + 193 + ], + "type": "text", + "content": "[39] Ran Wang, Sam Kwong, Xu Wang, and Yuheng Jia. Active k-labelsets ensemble for multi-label classification. PR, 109:107583, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 195, + 279, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 195, + 279, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 195, + 279, + 226 + ], + "type": "text", + "content": "[40] Jun Xiang, Guohan Xu, Chao Ma, and Jianhua Hou. End-to-end learning deep crf models for multi-object tracking deep crf models. IEEE TCSVT, 31(1):275-288, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 228, + 278, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 278, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 278, + 259 + ], + "type": "text", + "content": "[41] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In ICCV, 2015." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 261, + 278, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 278, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 278, + 282 + ], + "type": "text", + "content": "[42] Donggeun Yoo and In So Kweon. Learning loss for active learning. In CVPR, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 283, + 278, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 278, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 278, + 314 + ], + "type": "text", + "content": "[43] Fengwei Yu, Wenbo Li, Quanquan Li, Yu Liu, Xiaohua Shi, and Junjie Yan. Poi: Multiple object tracking with high performance detection and appearance feature. In ECCV, 2016." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 316, + 278, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 278, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 278, + 347 + ], + "type": "text", + "content": "[44] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In CVPR, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 349, + 278, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 278, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 278, + 380 + ], + "type": "text", + "content": "[45] Xiaosong Zhang, Fang Wan, Chang Liu, Xiangyang Ji, and Qixiang Ye. Learning to match anchors for visual object detection. IEEE TPAMI, 44(6):3096-3109, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 382, + 278, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 278, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 278, + 413 + ], + "type": "text", + "content": "[46] Yang Zhang, Hao Sheng, Yubin Wu, Shuai Wang, Weifeng Lyu, Wei Ke, and Zhang Xiong. Long-term tracking with deep tracklet association. IEEE TIP, 29:6694-6706, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 415, + 278, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 278, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 278, + 457 + ], + "type": "text", + "content": "[47] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 459, + 278, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 278, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 278, + 500 + ], + "type": "text", + "content": "[48] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 502, + 278, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 278, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 278, + 544 + ], + "type": "text", + "content": "[49] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 546, + 278, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 278, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 278, + 586 + ], + "type": "text", + "content": "[50] Jian Zhao, Lin Xiong, Jianshu Li, Junliang Xing, Shuicheng Yan, and Jiashi Feng. 3d-aided dual-agent gans for unconstrained face recognition. IEEE TPAMI, 41(10):2380-2394, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 589, + 278, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 278, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 278, + 619 + ], + "type": "text", + "content": "[51] Jian Zhao, Shuicheng Yan, and Jiashi Feng. Towards age-invariant face recognition. IEEE TPAMI, 44(1):474-487, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 622, + 278, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 278, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 278, + 663 + ], + "type": "text", + "content": "[52] Linyu Zheng, Ming Tang, Yingying Chen, Guibo Zhu, Jinqiao Wang, and Hanqing Lu. Improving multiple object tracking with single object tracking. In CVPR, pages 2453-2462, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 666, + 278, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 278, + 719 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 278, + 719 + ], + "type": "text", + "content": "[53] Xue-Feng Zhu, Tianyang Xu, Jian Zhao, Jia-Wei Liu, Kai Wang, Gang Wang, Jianan Li, Zhihao Zhang, Qiang Wang, Lei Jin, et al. Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-uav system. arXiv preprint arXiv:2306.15767, 2023." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 746, + 297, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 746, + 297, + 756 + ], + "spans": [ + { + "bbox": [ + 279, + 746, + 297, + 756 + ], + "type": "text", + "content": "9941" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_content_list.json b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d3dc22dd89f40da9747e7afb54c17c10673e950c --- /dev/null +++ b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_content_list.json @@ -0,0 +1,1549 @@ +[ + { + "type": "text", + "text": "Heterogeneous Forgetting Compensation for Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 116, + 130, + 854, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahua Dong $^{1,2,3*}$ , Wenqi Liang $^{1,2,3*}$ , Yang Cong $^{4\\dagger}$ , Gan Sun $^{1,2}$", + "bbox": [ + 235, + 179, + 733, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, 110016, China.‡", + "bbox": [ + 207, + 199, + 761, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang, 110169, China.", + "bbox": [ + 254, + 233, + 712, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Chinese Academy of Sciences, Beijing, 100049, China.", + "bbox": [ + 207, + 268, + 759, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ South China University of Technology, Guangzhou, 510640, China.", + "bbox": [ + 212, + 286, + 754, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dongjiahua1995, liangwenqi0123, congyang81, sunganl412}@gmail.com", + "bbox": [ + 200, + 306, + 769, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 313, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Class-incremental learning (CIL) has achieved remarkable successes in learning new classes consecutively while overcoming catastrophic forgetting on old categories. However, most existing CIL methods unreasonably assume that all old categories have the same forgetting pace, and neglect negative influence of forgetting heterogeneity among different old classes on forgetting compensation. To surmount the above challenges, we develop a novel Heterogeneous Forgetting Compensation (HFC) model, which can resolve heterogeneous forgetting of easy-to-forget and hard-to-remember old categories from both representation and gradient aspects. Specifically, we design a task-semantic aggregation block to alleviate heterogeneous forgetting from representation aspect. It aggregates local category information within each task to learn task-shared global representations. Moreover, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation loss and a gradient-balanced relation distillation loss to alleviate forgetting from gradient aspect. They consider gradient-balanced compensation to rectify forgetting heterogeneity of old categories and heterogeneous relation consistency. Experiments on several representative datasets illustrate effectiveness of our HFC model. The code is available at https://github.com/JiahuaDong/HFC.", + "bbox": [ + 75, + 388, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 781, + 209, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Class-incremental learning (CIL) [40, 12, 5] has attracted appealing attentions recently by accumulating previous learned experience to learn new classes incrementally. It", + "bbox": [ + 75, + 806, + 470, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f3ad93781789f1aaea7c553553eddf6928008ad3c55f0926f3a1daa3149e0ff.jpg", + "image_caption": [ + "Figure 1. Forgetting heterogeneity versus accuracy on CIFAR-100 [33] when the backbone is ViT-Base and number of tasks is 5." + ], + "image_footnote": [], + "bbox": [ + 522, + 354, + 870, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plays an indispensable role in developing a large number of intelligent learning systems, such as autonomous driving [38] and automated surveillance [26]. When learning new classes continuously under the settings of limited memory to replay all previous data of old classes [45], these CIL systems heavily suffer from forgetting on old classes caused by severe class imbalance between old and new categories [40, 31, 14]. To surmount catastrophic forgetting, a growing amount of CIL methods mainly perform knowledge distillation [24] to preserve past experience [34, 12, 35]; introduce a rehearsal strategy to replay part of old data [40, 17]; or dynamically expand network architectures [28, 41, 54].", + "bbox": [ + 496, + 580, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, most existing CIL methods [40, 45, 54, 18] neglect heterogeneous forgetting speeds of different old classes. They unrealistically assume that all old classes suffer from the same degree of catastrophic forgetting, and compensate forgetting for each old classes equally and independently. Such impracticable assumption enforces existing CIL models [40, 56, 61] to suffer from imbalanced gradient optimization among different old classes, thus favoring more forgetting compensation for hard-to-forget old classes while ne", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. †The corresponding author is Prof. Yang Cong. ‡This work was supported in part by the National Nature Science Foundation of China under Grant 62127807, 62273333 and 62133005.", + "bbox": [ + 75, + 862, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11742", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "glecting those easy-to-forget old classes (i.e., heterogeneous forgetting). More importantly, such forgetting heterogeneity can significantly worsen the forgetting on hard-to-forget old categories, when new streaming classes becomes part of old categories continually. For example, some old classes with various modalities and appearances (e.g., person and car) in autonomous driving [38] are more difficult to explore task-shared representations across different incremental tasks, when compared with other hard-to-forget old classes (e.g., road and traffic sign) with easily-distinguished visual properties. This phenomenon causes imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus exacerbating heterogeneous forgetting speeds among different old classes. When autonomous vehicles [38] learn new classes continually, such forgetting heterogeneity aggravates forgetting on those hard-to-forget old classes (e.g., traffic sign) to some extent.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the above practical scenario, we investigate that different old classes have significant forgetting heterogeneity in this paper, as shown in Fig. 1. This heterogeneous forgetting might heavily weaken forgetting compensation on hard-to-forget old classes, as new classes become a subset of old classes consecutively. In summary, the challenges to tackle heterogeneous forgetting lie in two major aspects:", + "bbox": [ + 75, + 348, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Representation Aspect: Some easy-to-forget old classes (e.g., car and bus) with diverse appearances are more difficult for existing CIL methods [45, 54, 12, 3] to learn intrinsic task-shared representations, and thus are significantly easier to be forgotten than hard-to-forget old classes (e.g., road) with distinctive attributes. Thus, exploring task-shared representations is essential to address heterogeneous forgetting among different old classes.", + "- Gradient Aspect: To learn complex visual characterizations for easy-to-forget old classes with various modalities and appearances, existing CIL methods [40, 4, 49] are required to allocate more network architectures for gradient updating. It can result in imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus aggravating forgetting heterogeneity of old classes." + ], + "bbox": [ + 83, + 458, + 472, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the above challenges, we propose a novel Heterogeneous Forgetting Compensation (HFC) model, which is an earlier exploration to address heterogeneous forgetting from both representation and gradient perspectives in the CIL field [40, 4]. Specifically, we propose a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation perspective. It can explore task-shared global representations for each class via aggregating long-range local category information within each task. Meanwhile, to tackle heterogeneous forgetting from gradient perspective, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss. The GFC loss can rectify heterogeneous forgetting speeds", + "bbox": [ + 75, + 689, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of easy-to-forget and hard-to-forget old classes, while normalize different learning paces of new categories to achieve gradient-balanced propagation. Besides, the GRD loss can distill heterogeneous relation consistency brought by forgetting heterogeneity among different old classes. Experiments show our model has large improvements on several representative datasets, compared with baseline methods [40, 36]. More importantly, we apply two plug-and-play losses into existing distillation-based CIL methods [40, 17, 4] to significantly improve their performance from gradient aspect. The novel contributions of this paper are presented as follows:", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a novel Heterogeneous Forgetting Compensation (HFC) model to address different forgetting speeds of easy-to-forget and hard-to-forget old classes. To our best knowledge, this paper is the first exploration to tackle forgetting heterogeneity among old categories from representation and gradient aspects in the CIL field.", + "- We design a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect. It can explore robust task-shared representations for each class via aggregating local category information.", + "- We propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss to surmount forgetting heterogeneity from gradient aspect. They can balance different forgetting of old classes and heterogeneous category-relation consistency to improve performance when applying them into existing methods." + ], + "bbox": [ + 504, + 262, + 893, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 537, + 640, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We discuss some class-incremental learning (CIL) methods [60, 37, 18], and divide them into three categories:", + "bbox": [ + 498, + 564, + 893, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation: After Li et al. [34] apply knowledge distillation [24] to continual learning, [40, 39, 59] follow them to transfer past experience from old model to new model. iCaRL [40] and EEIL [6] perform knowledge distillation in the output space. [44] aims to preserve consistent topology feature space for old and new tasks. [16] proposes feature-level knowledge distillation when applying continual learning [37] into semantic segmentation task [53, 52, 11].", + "bbox": [ + 496, + 595, + 893, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rehearsal Strategy: For replaying past experience, lots of CIL methods [49, 8, 42] allocate a memory to store exemplars of old classes or synthesize samples via generative adversarial model [21, 58]. For saving large memory overhead, [29] proposes to store low-dimension features rather than raw samples. [51] proposes a \"white box\" framework deviated from rate reduction to preserve past experience.", + "bbox": [ + 496, + 718, + 893, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Architecture: Many works [28, 54, 1] allocate dynamical networks for new classes as the growing of learned classes. Yoon et al. [55] dynamically expand architecture capacity via selective retraining. [54, 18] address the problem of relying on task index when performing dynamic", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11743", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "architectures. However, the above-mentioned CIL methods [40, 27] neglect negative influence of heterogeneous forgetting of old classes on forgetting compensation.", + "bbox": [ + 75, + 90, + 470, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Model", + "text_level": 1, + "bbox": [ + 76, + 152, + 276, + 169 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition and Overview", + "text_level": 1, + "bbox": [ + 76, + 178, + 370, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Definition: Following class-incremental learning (CIL) methods [40, 50, 17, 18, 54], we set a series of consecutive learning tasks as $\\mathcal{T} = \\{\\mathcal{T}^t\\}_{t=1}^T$ , where $T$ indicates the total task quantity. The $t$ -th task $\\mathcal{T}^t = \\{\\mathbf{x}_i^t, \\mathbf{y}_i^t\\}_{i=1}^{N^t}$ is composed of $N^t$ pairs of image $\\mathbf{x}_i^t$ and one-hot groundtruth $\\mathbf{y}_i^t \\in \\mathcal{Y}^t$ , and $\\mathcal{Y}^t$ denotes label space of the $t$ -th incremental task containing $K^t$ new categories. The label spaces between any two incremental tasks have no overlap: $\\mathcal{Y}^t \\cap (\\cup_{j=1}^{t-1} \\mathcal{Y}^j) = \\emptyset$ . Namely, $K^t$ new classes belonging to the $t$ -th task $\\mathcal{T}^t$ are different from $K^o = \\sum_{i=1}^{t-1} K^i \\subset \\cup_{j=1}^{t-1} \\mathcal{Y}^j$ old classes learned from previous tasks $\\{\\mathcal{T}^i\\}_{i=1}^{t-1}$ . As introduced in CIL baselines [40, 50, 54], we set a fixed exemplar memory $\\mathcal{M}$ , and store only few images (i.e., $\\frac{|M|}{K^o}$ ) for each old class in the $t$ -th task $\\mathcal{T}^t$ , where $\\mathcal{M}$ satisfies $\\frac{N^t}{K^t} \\gg \\frac{|M|}{K^o}$ . In the $t$ -th task, we aim to identify both $K^t$ new classes and $K^o$ old classes via optimizing the model on $\\mathcal{T}^t$ and $\\mathcal{M}$ .", + "bbox": [ + 75, + 200, + 472, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview: The overview of our HFC model to surmount forgetting heterogeneity of different old categories from both representation and gradient aspects is shown in Fig. 2. Denote the proposed HFC model learned at the $(t - 1)$ -th and $t$ -th tasks as old and new models (i.e., $\\Theta^{t - 1}$ and $\\Theta^t$ ), where $\\Theta^t$ is inherited from $\\Theta^{t - 1}$ and only expands the number of output neurons in the classifier to identify $K^t$ new classes. In the $t$ -th incremental task, given an image $\\mathbf{x}_i^t\\in \\mathcal{T}^t\\cup \\mathcal{M}$ , we forward it into $\\Theta^{t - 1}$ and $\\Theta^t$ to extract task-shared global representation via task-semantic aggregation (TSA) block, which alleviates heterogeneous forgetting from representation aspect (Section 3.2). After obtaining probabilities $\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o}$ and $\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}$ predicted via $\\Theta^{t - 1}$ and $\\Theta^t$ , we develop two plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss $\\mathcal{L}_{\\mathrm{FC}}$ and a gradient-balanced relation distillation (GRD) loss $\\mathcal{L}_{\\mathrm{RD}}$ to surmount heterogeneous forgetting from semantic aspect (Sections 3.3 and 3.4). $\\mathcal{L}_{\\mathrm{FC}}$ and $\\mathcal{L}_{\\mathrm{RD}}$ can normalize heterogeneous forgetting of old classes and distill heterogeneous relation consistency between $\\Theta^{t - 1}$ and $\\Theta^t$ .", + "bbox": [ + 75, + 455, + 472, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task-Semantic Aggregation Block", + "text_level": 1, + "bbox": [ + 76, + 770, + 374, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To explore long-range semantic dependencies across different tasks, Douillard et al. [18] introduce Vision Transformer (ViT) [15, 46, 22] into class-incremental learning (CIL). However, it significantly increases the memory overhead to store task tokens and task-specific decoders, when learning large-scale new classes. Besides, it cannot aggregate local category information to explore task-shared global", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "representations, resulting in heterogeneous forgetting from representation aspect. To tackle these issues, we develop a task-semantic aggregation (TSA) block to extract task-shared global representations, which is effective to tackle heterogeneous forgetting from representation aspect. More importantly, the TSA block cannot dynamically increase memory cost with the arrival of large-scale new categories.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, as presented in Fig. 2, the proposed model $\\Theta^t = \\Theta_c^t\\circ \\Theta_f^t$ at the $t$ -th task is composed of a feature extractor $\\Theta_f^t$ to extract task-shared global representations, and a task classifier $\\Theta_c^t$ to classify $K^o$ old classes and $K^t$ new classes. In the $t$ -th task, the network parameters of $\\Theta_f^{t - 1}$ and $\\Theta_c^{t - 1}$ in $\\Theta^{t - 1}$ learned at the $(t - 1)$ -th task are frozen to preserve semantic knowledge for old classes. Different from [18] that uses vanilla ViT [15, 30, 57, 7] to extract features, we employ $l_{m}$ multi-head self-attention (MSA) blocks [20] and $l_{t}$ task-semantic aggregation (TSA) block as feature extractor $\\Theta_f^t$ to obtain task-shared global representations.", + "bbox": [ + 496, + 198, + 893, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Multi-Head Self-Attention (MSA): Given an image $\\mathbf{x}_i^t\\in \\mathcal{T}^t\\cup \\mathcal{M}$ in the $t$ th task, following ViT [15, 9], we crop it into $N$ patches with equal dimension, flatten these patches and map them into a $D$ -dimension feature space via a linear projection to obtain patch embedding $\\mathbf{Z}_e\\in \\mathbb{R}^{N\\times D}$ . Then we concatenate $\\mathbf{Z}_e$ with a class token $\\mathbf{Z}_{\\mathrm{cls}}\\in \\mathbb{R}^{D}$ , and perform element-wise sum operation with a position token $\\mathbf{Z}_p\\in \\mathbb{R}^{(N + 1)\\times D}$ to get $\\mathbf{Z}_0 = [\\mathbf{Z}_e\\oplus \\mathbf{Z}_{\\mathrm{cls}}] + \\mathbf{Z}_p\\in \\mathbb{R}^{(N + 1)\\times D}$ where $\\oplus$ is concatenation function. As shown in Fig.2, $\\mathbf{Z}_0$ is fed into $\\Theta_f^t$ including $l_{m}$ MSA blocks to extract long-range semantics across different tasks. For the $l$ -th $(l = 1,\\dots ,l_m)$ MSA block, we regard $\\mathbf{Z}_{l - 1}\\in \\mathbb{R}^{(N + 1)\\times D}$ learned from the $(l - 1)$ -th MSA block as input, and execute parallel selfattention $H$ times. Thus, the self-attention $\\mathbf{A}_l^h\\in \\mathbb{R}^{(N + 1)\\times d}$ of $\\mathbf{Z}_{l - 1}$ for the $h$ -th $(h = 1,\\dots ,H)$ head is:", + "bbox": [ + 496, + 369, + 893, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {l} ^ {h} = \\sigma \\left(\\frac {\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {q} ^ {h} \\left(\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {k} ^ {h}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {v} ^ {h}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 608, + 892, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_q^h,\\mathbf{W}_k^h,\\mathbf{W}_v^h\\in \\mathbb{R}^{D\\times d}$ are query, key and value mapping matrices. $d = \\frac{D}{H}$ represents channel dimension of each head and $\\sigma$ is the softmax function. Then we concatenate $\\{\\mathbf{A}_l^h\\}_{h = 1}^H$ along channel dimension and project this concatenated result via a mapping matrix $\\mathbf{W}_o\\in \\mathbb{R}^{D\\times D}$ to perform MSA: $\\mathrm{MSA}(\\mathbf{Z}_{l - 1}) = \\mathbf{Z}_{l - 1} + [\\mathbf{A}_l^1\\oplus$ $\\mathbf{A}_l^2,\\dots ,\\oplus \\mathbf{A}_l^H ]\\mathbf{W}_o$ where $\\oplus$ is concatenation function. After obtaining $\\mathrm{MSA}(\\mathbf{Z}_{l - 1})\\in \\mathbb{R}^{(N + 1)\\times D}$ $(D = dH)$ , we utilize a multi-layer perceptron (MLP) block to obtain output $\\mathbf{Z}_l\\in \\mathbb{R}^{(N + 1)\\times D}$ of the $l$ th MSA block:", + "bbox": [ + 496, + 656, + 893, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {l} = \\operatorname {M S A} \\left(\\mathbf {Z} _ {l - 1}\\right) + \\operatorname {M L P} \\left(\\operatorname {M S A} \\left(\\mathbf {Z} _ {l - 1}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 824, + 892, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Task-Semantic Aggregation (TSA): The patch embedding $\\mathbf{Z}_{l_m} \\in \\mathbb{R}^{(N + 1) \\times D}$ encoded via the $l_m$ -th MSA block contains rich long-range semantic dependencies shared", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11744", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16353ab5f0807c1c32c2d176bfa051ceb2f1a6378bded8b6cd75f886f8c61555.jpg", + "image_caption": [ + "Figure 2. Illustration of the HFC model. It mainly contains a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect via exploring task-shared global representations, and two novel plug-and-play losses (i.e., a gradient-balanced forgetting compensation loss $\\mathcal{L}_{\\mathrm{FC}}$ and a gradient-balanced relation distillation loss $\\mathcal{L}_{\\mathrm{RD}}$ ) to overcome forgetting heterogeneity from gradient aspect by compensating gradient-imbalanced propagation and heterogeneous class-relation consistency." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "across different tasks. $\\mathbf{Z}_{l_m}$ is then forwarded into $l_{t}$ task-semantic aggregation (TSA) blocks to extract task-shared global embedding. The TSA block aggregates local category context from long-range semantic dependencies to global representations shared across all incremental tasks, which is essential to alleviate heterogeneous forgetting from representation aspect. In the $t$ -th task, as depicted in Fig. 2, we introduce a learnable task-shared embedding $\\mathbf{E}_0^t \\in \\mathbb{R}^D$ that is initialized via $\\mathbf{E}_0^{t - 1}$ learned in the $(t - 1)$ -th task. Then we forward $\\mathbf{E}_0^t$ along with $\\mathbf{Z}_{l_m}$ to $l_{t}$ TSA blocks. For the $l$ -th $(l = 1,\\dots ,l_t)$ TSA block, $\\mathbf{E}_{l - 1}^t \\in \\mathbb{R}^D$ learned at $(l - 1)$ -th TSA block and $\\mathbf{Z}_{l_m}$ are used to perform task-semantic attention $H$ times in parallel. Thus, we formulate task-semantic attention $\\mathbf{S}_l^h \\in \\mathbb{R}^d$ for the $h$ -th $(h = 1,\\dots ,H)$ head as:", + "bbox": [ + 75, + 383, + 473, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {l} ^ {h} = \\sigma \\left(\\frac {\\mathbf {E} _ {l - 1} ^ {t} \\mathbf {V} _ {q} ^ {h} \\left(\\mathbf {Z} _ {l _ {m}} \\mathbf {V} _ {k} ^ {h}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {Z} _ {l _ {m}} \\mathbf {V} _ {v} ^ {h}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 604, + 468, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{V}_q^h,\\mathbf{V}_k^h,\\mathbf{V}_v^h\\in \\mathbb{R}^{D\\times d}$ are projection matrices of query, key and value. After utilizing Eq. (3) to obtain $\\{\\mathbf{S}_l^h\\}_{h = 1}^H$ , we concatenate them along channel dimension and project this fused result via $\\mathbf{V}_o\\in \\mathbb{R}^{D\\times D}$ to execute $H$ -head task-semantic attention: $\\mathrm{TSA}(\\mathbf{E}_{l - 1}^{t}) = [\\mathbf{S}_{l}^{1}\\oplus$ $\\mathbf{S}_l^2,\\dots ,\\oplus \\mathbf{S}_l^H ]\\mathbf{V}_o\\in \\mathbb{R}^D$ . TSA $(\\mathbf{E}_{l - 1}^{t})$ is encoded via a MLP block to get output $\\mathbf{E}_l^t\\in \\mathbb{R}^D$ of the $l$ -th TSA block:", + "bbox": [ + 76, + 648, + 472, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {l} ^ {t} = \\operatorname {T S A} \\left(\\mathbf {E} _ {l - 1} ^ {t}\\right) + \\operatorname {M L P} \\left(\\operatorname {T S A} \\left(\\mathbf {E} _ {l - 1} ^ {t}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 768, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As aforementioned, we obtain the task-shared global representation $\\mathbf{E}_{l_t}^t\\in \\mathbb{R}^D$ via the $l_{t}$ -th TSA block in feature extractor $\\Theta_f^t$ . Obviously, the proposed TSA blocks can aggregate local category context of the $t$ -th task from longrange semantic dependencies to task-shared global representation $\\mathbf{E}_{l_t}^t$ . It is essential to alleviate heterogeneous forgetting among different old classes from representation aspect, by", + "bbox": [ + 75, + 795, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "exploring task-shared semantic context across different tasks. Note that we employ layer norm before TSA, MSA and MLP blocks, while we omit it for simplicity in this paper.", + "bbox": [ + 496, + 383, + 893, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Task Classifier: The task-shared global representation $\\mathbf{E}_{l_t}^t$ obtained via $\\Theta_f^t$ is fed into task classifier $\\Theta_c^t$ to predict $K^o$ old classes and $K^t$ new classes. Given a mini-batch $\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b\\in \\mathcal{T}^t\\cup \\mathcal{M}$ , the classification loss $\\mathcal{L}_{\\mathrm{CE}}$ is:", + "bbox": [ + 500, + 429, + 893, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\mathcal {D} _ {\\mathrm {C E}} \\left(\\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right), \\mathbf {y} _ {i} ^ {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 500, + 892, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}$ is softmax probability predicted by $\\Theta^t$ in the $t$ -th learning task. $\\mathcal{D}_{\\mathrm{CE}}(\\cdot ,\\cdot)$ indicates traditional cross-entropy loss, and $b$ is the batch size.", + "bbox": [ + 496, + 549, + 893, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the severe class imbalance (i.e., $\\frac{N^t}{K^t} \\gg \\frac{|\\mathcal{M}|}{K^o}$ ) among old and new categories in the $t$ -th task enforces the prediction $\\mathbf{P}^t(\\mathbf{x}_i^t, \\Theta^t)$ in Eq. (5) to suffer from catastrophic forgetting [40] on old categories. Moreover, the gradient optimization of Eq. (5) neglects heterogeneous forgetting speeds of easy-to-forget old classes with various appearances and hard-to-forget old classes with easily-distinguished attributes. To address heterogeneous forgetting from gradient aspect, we propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss $\\mathcal{L}_{\\mathrm{FC}}$ in Section 3.3 and a gradient-balanced relation distillation (GRD) loss $\\mathcal{L}_{\\mathrm{RD}}$ in Section 3.4.", + "bbox": [ + 496, + 595, + 893, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Gradient-Balanced Forgetting Compensation", + "text_level": 1, + "bbox": [ + 500, + 786, + 882, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As incremental tasks arrive consecutively, easy-to-forget and hard-to-forget old categories change dynamically, thus aggravating difficulty to tackle heterogeneous forgetting from gradient aspect. In light of this, we develop a gradient-balanced forgetting compensation (GFC) loss $\\mathcal{L}_{\\mathrm{FC}}$ . It adaptively balances large forgetting heterogeneity of old classes", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11745", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "via performing balanced gradient propagation. Specifically, we perform task-adaptive gradient normalization for different classes, and then reweight classification loss $\\mathcal{L}_{\\mathrm{CE}}$ in Eq. (5) to compensate imbalanced gradients. For the $t$ -th incremental task, as claimed in [48, 13], the gradient $\\Gamma_i^t$ of an given image $(\\mathbf{x}_i^t,\\mathbf{y}_i^t)\\subset \\mathcal{T}^t\\cup \\mathcal{M}$ with respect to the $k$ -th $(k = \\arg \\max \\mathbf{y}_i^t)$ neuron $\\mathcal{N}_k^t$ of classifier $\\Theta_c^t$ in $\\Theta^t$ is:", + "bbox": [ + 75, + 90, + 472, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {i} ^ {t} = \\frac {\\partial \\mathcal {D} _ {\\mathrm {C E}} (\\mathbf {P} ^ {t} (\\mathbf {x} _ {i} ^ {t} , \\boldsymbol {\\Theta} ^ {t}) , \\mathbf {y} _ {i} ^ {t})}{\\partial \\mathcal {N} _ {k} ^ {t}} = \\mathbf {P} ^ {t} (\\mathbf {x} _ {i} ^ {t}, \\boldsymbol {\\Theta} ^ {t}) _ {k} - 1, \\quad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 204, + 470, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)_k$ denotes probability of the $k$ -th class.", + "bbox": [ + 75, + 246, + 447, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compensate heterogeneous forgetting of easy-to-forget and hard-to-forget old categories, we compute task-adaptive gradient means for different tasks. Consequently, given a mini-batch $\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b$ in the $t$ -th learning task, the task-adaptive gradient mean $\\Gamma_{\\eta}$ for the categories learned in the $\\eta$ -th $(1\\leq \\eta \\leq t)$ incremental task is formulated as:", + "bbox": [ + 75, + 263, + 470, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {\\eta} = \\frac {1}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}} \\sum_ {i = 1} ^ {b} | \\Gamma_ {i} ^ {t} | \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 359, + 470, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The forgetting heterogeneity of old categories learned from $t - 1$ old tasks and learning speeds of new categories in the $t$ -th task can be effectively measured via $\\{\\Gamma_{\\eta}\\}_{\\eta = 1}^{t}$ in Eq. (7).", + "bbox": [ + 75, + 402, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When easy-to-forget and hard-to-remember old classes are varied dynamically as incremental tasks, some noisy predictions on easy-to-remember old classes may result in large deviation to measure their forgetting heterogeneity [43, 19]. To tackle this issue, $\\{\\Gamma_{\\eta}\\}_{\\eta = 1}^{t}$ in Eq. (7) are expected to be sharper and more distinguishable adaptively [13], as learning tasks arrive continually. Thus, we rewrite Eq. (7) as follows:", + "bbox": [ + 75, + 449, + 470, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {\\eta} ^ {s} = \\frac {\\sum_ {i = 1} ^ {b} \\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right) \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 561, + 468, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sharper task-adaptive gradient means $\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}$ computed via Eq. (8) are employed to reweight classification loss $\\mathcal{L}_{\\mathrm{CE}}$ in Eq. (5). Then the gradient-balanced forgetting compensation (GFC) loss $\\mathcal{L}_{\\mathrm{FC}}$ is written as follows:", + "bbox": [ + 75, + 614, + 468, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F C}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\frac {\\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right)}{\\sum_ {\\eta = 1} ^ {t} \\Gamma_ {\\eta} ^ {s} \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}} \\mathcal {D} _ {\\mathrm {C E}} \\left(\\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right), \\mathbf {y} _ {i} ^ {t}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 683, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obviously, $\\mathcal{L}_{\\mathrm{FC}}$ in Eq. (9) encourages the optimization of $\\Theta^t$ to perform gradient-balanced propagation for easy-to- forget, hard-to- forget old classes and new classes learned in different tasks. It tackles heterogeneous forgetting on old classes from gradient aspect, via normalizing forgetting paces of different old classes adaptively with $\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}$ .", + "bbox": [ + 75, + 732, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Gradient-Balanced Relation Distillation", + "text_level": 1, + "bbox": [ + 76, + 832, + 419, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The relationships among old and new categories remain constant in semantic space, regardless of availability of old-class samples. In light of this, distilling inter-class relations", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from old model $\\Theta^{t-1}$ to new model $\\Theta^t$ can address forgetting on old classes [6, 25]. However, most knowledge distillation methods used in CIL [40, 2, 50] only utilize prediction of an individual sample to perform semantic consistency of old classes between $\\Theta^{t-1}$ and $\\Theta^t$ . They neglect underlying relations among old and new categories to tackle forgetting, and heavily suffer from noisy predictions on easy-to-forget old classes when distilling knowledge via an individual sample. Moreover, different forgetting speeds of old classes may result in heterogeneous class-relation consistency between $\\Theta^{t-1}$ and $\\Theta^t$ . It enforces gradient optimization of $\\Theta^t$ to bias towards some specific relation distillations related to new classes, causing forgetting heterogeneity on old categories.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome the above issues, we design a gradient-balanced relation distillation (GRD) loss $\\mathcal{L}_{\\mathrm{RD}}$ to distill relations among old and new classes from $\\Theta^{t - 1}$ to $\\Theta^t$ . As introduced in Fig. 2, it can address heterogeneous forgetting from gradient aspect via rectifying imbalanced gradient propagation caused by heterogeneous class-relation consistency. Specifically, we construct class-wise relation prototype rather than a prediction of single sample to perform relation distillation, which can alleviate negative effect of noisy relation consistency. Then the task-adaptive gradient means $\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}$ in Eq. (8) are employed to reweight heterogeneous class-relation distillation.", + "bbox": [ + 496, + 287, + 893, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A mini-batch $\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b\\subset \\mathcal{T}^t\\cup \\mathcal{M}$ is fed into $\\Theta^{t - 1}$ and $\\Theta^t$ to obtain softmax probabilities $\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o}$ of old classes and $\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}$ of old and new classes, as shown in Fig. 2. The first $K^o$ dimensions of $\\mathbf{y}_i^t\\in \\mathbb{R}^{K^o +K^t}$ are replaced with $\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})$ to get relation groundtruth $\\mathbf{Y}^t (\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o +K^t}$ that implies underlying relations among old and new classes. To tackle noisy class relations, we construct class-wise relation prototype $\\tilde{\\mathbf{P}}_k^t$ and its relation groundtruth $\\tilde{\\mathbf{Y}}_k^t$ for the $k$ -th class:", + "bbox": [ + 496, + 467, + 893, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {P}} _ {k} ^ {t} = \\frac {1}{\\Delta_ {k}} \\sum_ {i = 1} ^ {b} \\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 614, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {Y}} _ {k} ^ {t} = \\frac {1}{\\Delta_ {k}} \\sum_ {i = 1} ^ {b} \\mathbf {Y} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t - 1}\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 659, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta_k = \\sum_i^b\\mathbb{I}_{\\mathrm{argmax}\\mathbf{y}_i^t = k}$ . Then we formulate category-wise gradient mean $\\Gamma_k^s$ for the $k$ -th category as follows:", + "bbox": [ + 496, + 709, + 893, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {k} ^ {s} = \\frac {\\sum_ {i = 1} ^ {b} \\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 748, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consequently, we use $\\Gamma_k^s$ to reweight heterogeneous relation distillation, and express the proposed $\\mathcal{L}_{\\mathrm{RD}}$ as follows:", + "bbox": [ + 496, + 799, + 893, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R D}} = \\frac {1}{K ^ {o} + K ^ {t}} \\sum_ {k = 1} ^ {K ^ {o} + K ^ {t}} \\frac {\\Gamma_ {k} ^ {s} \\cdot \\mathcal {D} _ {\\mathrm {K L}} \\left(\\tilde {\\mathbf {P}} _ {k} ^ {t} , \\tilde {\\mathbf {Y}} _ {k} ^ {t}\\right)}{\\sum_ {\\eta = 1} ^ {t} \\Gamma_ {\\eta} ^ {s} \\cdot \\mathbb {I} _ {k \\in \\mathcal {Y} ^ {\\eta}}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 835, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{\\mathrm{KL}}(\\cdot ||\\cdot)$ indicates the Kullback-Leibler divergence.", + "bbox": [ + 496, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11746", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a79693cd96870af66706ac5b32e10429aa7f3eb2780cc6bb123d4bdbc0da2508.jpg", + "table_caption": [ + "Table 1. Results on CIFAR-100 [33]. When $\\mathcal{M} = 2000$ , we set $T = \\{5, 10, 20, 25, 50\\}$ for $\\mathcal{B} = 0\\%$ , and $T = \\{5, 10, 25, 50\\}$ for $\\mathcal{B} = 50\\%$ ." + ], + "table_footnote": [], + "table_body": "
Comparison MethodsBackbone#ParamsB=0%B=50%
510202550Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M83.580.678.577.676.479.3↑5.580.678.675.471.076.4↑2.9
BiC [50] (CVPR'2019)ViT-Base85.10M83.582.578.780.378.480.7↑4.278.275.774.271.975.0↑4.3
PODNet [17] (ECCV'2020)ViT-Base85.10M75.371.866.564.963.068.3↑16.672.873.074.274.973.7↑5.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.682.079.179.177.380.2↑4.675.874.872.972.073.9↑5.4
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M75.272.066.165.263.868.5↑16.472.772.874.074.773.5↑5.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M84.783.579.678.577.980.8↑4.178.477.376.274.076.5↑2.8
AFC [32] (CVPR'2022)ViT-Base85.10M74.675.570.367.464.770.5↑14.470.869.468.969.969.8↑9.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.585.881.380.877.482.2↑2.782.581.077.374.578.8↑0.5
HFC (Ours)ViT-Base85.10M86.486.385.585.081.184.9-82.981.077.875.579.3-
Upper BoundViT-Base85.10M94.294.294.294.294.294.2-94.294.294.294.294.2-
", + "bbox": [ + 80, + 103, + 890, + 273 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/70ce29fab9fc2e7474abe7eefd2a0be36bd632ede8b844a6becb4c60e2904f25.jpg", + "table_caption": [ + "Table 2. Results on ImageNet-100 [10]. When $\\mathcal{M} = {2000}$ ,we set $T = \\{ 5,{10},{20},{25}\\}$ for $\\mathcal{B} = 0\\%$ ,and $T = \\{ 5,{10},{25},{50}\\}$ for $\\mathcal{B} = {50}\\%$ ." + ], + "table_footnote": [], + "table_body": "
Comparison MethodsBackbone#ParamsB=0%B=50%
5102025Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M84.182.480.278.681.3↑3.279.578.876.673.977.2↑3.4
BiC [50] (CVPR'2019)ViT-Base85.10M81.480.279.278.579.8↑4.773.974.473.472.273.5↑7.1
PODNet [17] (ECCV'2020)ViT-Base85.10M82.480.179.578.380.1↑4.568.471.375.477.673.2↑7.4
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.081.780.979.481.2↑3.375.874.872.972.073.9↑6.7
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M82.381.079.379.080.4↑4.168.371.275.076.972.8↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M85.684.083.883.184.1↑0.481.080.580.079.880.3↑0.3
AFC [32] (CVPR'2022)ViT-Base85.10M83.281.280.579.481.1↑3.577.178.478.978.178.1↑2.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.183.480.080.382.2↑2.377.478.479.779.478.7↑1.9
HFC (Ours)ViT-Base85.10M86.185.484.082.684.5-81.581.180.379.580.6-
Upper BoundViT-Base85.10M86.386.386.386.386.3-86.386.386.386.386.3-
", + "bbox": [ + 81, + 291, + 890, + 463 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4cf56208686614c3fa72b24b1f6c7eb5d16b5b158ee5254d944f4faa78cb94c0.jpg", + "table_caption": [ + "Table 3. Results on ImageNet-1000 [10]. When $\\mathcal{M} = 20000$ , we set $T = \\{5,10\\}$ for $\\mathcal{B} = 0\\%$ ." + ], + "table_footnote": [], + "table_body": "
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M75.472.473.9↑3.5
BiC [50] (CVPR'2019)ViT-Base85.10M68.966.067.5↑10.0
PODNet [17] (ECCV'2020)ViT-Base85.10M68.471.369.8↑7.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M75.874.875.3↑2.2
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M68.463.265.8↑11.6
FOSTER [47] (ECCV'2022)ViT-Base85.10M76.974.375.6↑1.9
AFC [32] (CVPR'2022)ViT-Base85.10M74.271.472.8↑4.6
DyTox [18] (CVPR'2022)ViT-Base85.10M78.174.776.4↑1.0
HFC (Ours)ViT-Base85.10M78.576.477.5-
Upper BoundViT-Base85.10M86.386.386.3-
", + "bbox": [ + 81, + 500, + 465, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Optimization Pipeline", + "text_level": 1, + "bbox": [ + 76, + 635, + 282, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, in the $t$ -th $(t \\geq 2)$ learning task, the objective formulation to optimize $\\Theta^t$ is expressed as follows:", + "bbox": [ + 76, + 659, + 468, + 689 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o b j}} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {F C}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {R D}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 703, + 468, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{1},\\alpha_{2}$ are hyper-parameters. $\\Theta^t$ along with learnable $\\mathbf{E}_0^t$ are optimized via $\\mathcal{L}_{\\mathrm{CE}}$ in Eq. (5) for the first task, and trained via $\\mathcal{L}_{\\mathrm{obj}}$ in Eq. (14) when $t\\geq 2$ . After optimizing new model $\\Theta^t$ , we store $\\Theta^t$ as the frozen old model $\\Theta^{t - 1}$ to perform the GRD loss $\\mathcal{L}_{\\mathrm{RD}}$ for the next incremental task.", + "bbox": [ + 76, + 729, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 209, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 845, + 294, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fair comparisons, we follow the same protocols utilized in baseline CIL methods [40, 50, 17, 17, 2, 18, 32] to", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set incremental tasks (class order), and do experiments on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Specifically, we consider two task settings in this paper: 1) We divide all categories of each dataset into $T$ tasks equally [40, 18] (i.e., the base class set $\\mathcal{B} = 0\\%$ ); 2) We start by training our model on half of the categories, while divide the rest of categories into $T$ tasks equally [17, 32] (i.e., the base class set $\\mathcal{B} = 50\\%$ ). When $\\mathcal{B} = 0\\%$ , we set $T = \\{5, 10, 20, 25, 50\\}$ for CIFAR-100 [33], $T = \\{5, 10, 20, 25\\}$ for ImageNet-100 [10] and $T = \\{5, 10\\}$ for ImageNet-1000 [10]. When $\\mathcal{B} = 50\\%$ , we set $T = \\{5, 10, 25, 50\\}$ for CIFAR-100 [33] and ImageNet-100 [10].", + "bbox": [ + 496, + 477, + 892, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As introduced in [40, 50], the size of memory $\\mathcal{M}$ for all comparison methods is fixed as 2,000 for CIFAR-100 [33] and ImageNet-100 [10], and 20,000 for ImageNet-1000 [10]. We follow iCaRL [40] to update memory $\\mathcal{M}$ and use global memory for [18]. Besides, we employ the same data augmentation proposed in DyTox [18] for all comparison methods, but don't use mixup technology. Following [17, 32], we also consider another memory setting (i.e., storing 20 exemplars per class), and set $T = \\{5,10\\}, \\mathcal{B} = 50\\%$ for comparisons on CIFAR-100 [33] (see Tab. 6). For network architecture, we use ViT-Base [15] as feature extractor containing $l_{m} = 11$ MSA blocks and $l_{t} = 1$ TSA block, where the parameters are initialized via [22]. The task classifier includes only one fully-connected layer. The SGD optimizer is employed to optimize our model, where the learning rate is initialized as $6.25 \\times 10^{-5}$ . We apply two plug-and-play losses (i.e., $\\mathcal{L}_{\\mathrm{FC}}$", + "bbox": [ + 496, + 659, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11747", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/996b0fdb9f3bdc39d095447b2144fb662271d773e2a8c936b43b98e2128d083e.jpg", + "table_caption": [ + "Table 4. Performance on CIFAR-100 [33] ( $T = \\{5, 10, 20, 25, 50\\}$ ) and ImageNet-100 [10] ( $T = \\{5, 10, 20, 25\\}$ ) when we apply $\\mathbf{Ours}^{\\ddagger}$ into existing distillation-based CIL methods and set $\\mathcal{M} = 2000$ , $\\mathcal{B} = 0\\%$ . $\\mathbf{Ours}^{\\ddagger}$ denotes the proposed plug-and-play losses $\\mathcal{L}_{\\mathrm{FC}}$ and $\\mathcal{L}_{\\mathrm{RD}}$ ." + ], + "table_footnote": [], + "table_body": "
Comparison MethodsCIFAR-100 Dataset [33]ImageNet-100 Dataset [10]
Backbone#Params510202550Avg.Backbone#Params5102025
iCaRL [40] (CVPR'2017)ResNet-320.46M66.463.953.150.239.154.5ResNet-1811.22M76.469.560.657.9
iCaRL [40] + Ours‡ResNet-320.46M68.065.754.251.440.856.0ResNet-1811.22M76.870.361.759.2
BiC [50] (CVPR'2019)ResNet-320.46M55.350.848.346.439.348.0ResNet-1811.22M61.554.346.744.5
BiC [50] + Ours‡ResNet-320.46M63.253.852.449.341.552.0ResNet-1811.22M63.257.050.348.4
PODNet [17] (ECCV'2020)ResNet-320.46M63.854.548.144.439.750.1ResNet-1811.22M74.966.555.851.7
PODNet [17] + Ours‡ResNet-320.46M65.860.051.648.344.754.1ResNet-1811.22M76.569.359.455.8
SS-IL [2] (ICCV'2021)ResNet-320.46M65.159.049.048.640.752.5ResNet-1811.22M60.953.539.335.8
SS-IL [2] + Ours‡ResNet-320.46M67.764.951.651.242.555.6ResNet-1811.22M75.267.557.956.0
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M63.754.347.844.039.049.8ResNet-1811.22M74.866.355.451.3
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M66.758.550.546.943.353.2ResNet-1811.22M76.468.758.955.1
FOSTER [47] (ECCV'2022)ResNet-320.46M72.068.767.166.062.667.3ResNet-1811.22M74.970.967.265.7
FOSTER [47] + Ours‡ResNet-320.46M72.269.967.466.762.867.8ResNet-1811.22M75.071.367.566.2
AFC [32] (CVPR'2022)ResNet-320.46M65.657.150.948.142.052.7ResNet-1811.22M76.369.761.158.1
AFC [32] + Ours‡ResNet-320.46M66.263.853.249.744.155.4ResNet-1811.22M77.471.263.367.4
Upper BoundResNet-320.46M76.676.676.676.676.676.6ResNet-1811.22M73.273.273.273.2
DyTox [18] (CVPR'2022)ViT-Tiny10.71M75.373.572.772.370.572.9ViT-Tiny10.71M77.875.472.972.3
DyTox [18] + Ours‡ViT-Tiny10.71M75.974.072.972.670.873.2ViT-Tiny10.71M77.975.673.172.9
Upper BoundViT-Tiny10.71M76.176.176.176.176.176.1ViT-Tiny10.71M79.179.179.179.1
", + "bbox": [ + 80, + 119, + 893, + 412 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/683c4ef6ce1890edb08206d5e2fe743b2d6ca4bf633accec72568fa18a784b17.jpg", + "table_caption": [ + "Table 5. Ablation studies on CIFAR-100 [33] (the top block) and ImageNet-100 [10] (the bottom block), when we set $T = \\{5,10,25\\}$ under the settings of $\\mathcal{M} = 2000, \\mathcal{B} = 0\\%$ ." + ], + "table_footnote": [], + "table_body": "
Ablation VariantsBackbone#ParamsTSAGFCGRD51025
BaselineViT-Base85.10MXXX81.680.278.1
Baseline + TSAViT-Base85.10M✓XX83.182.481.0
Baseline + TSA + GFCViT-Base85.10M✓✓X84.083.782.1
HFC (Ours)ViT-Base85.10M✓✓✓86.486.385.0
BaselineViT-Base85.10MXXX82.080.578.8
Baseline + TSAViT-Base85.10M✓XX83.782.881.2
Baseline + TSA + GFCViT-Base85.10M✓XX84.383.482.0
HFC (Ours)ViT-Base85.10M✓✓✓86.185.484.0
", + "bbox": [ + 81, + 465, + 468, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and $\\mathcal{L}_{\\mathrm{RD}})$ to existing CIL methods [40, 18] using ResNet-18 [23], ResNet-32 [23] or ViT-Tiny [15] as backbone. For evaluation, top-1 accuracy is used to compare performances of our model and baseline methods.", + "bbox": [ + 75, + 606, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison Experiments", + "text_level": 1, + "bbox": [ + 76, + 679, + 310, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison Results: Tabs. 1-3 present comparison results of our HFC model and off-the-shelf CIL methods on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Our model significantly outperforms existing CIL methods [40, 50, 17, 2, 18, 32, 47] by $0.7\\% \\sim 16.6\\%$ accuracy under various incremental settings, when we use the same backbone (i.e., ViT-Base [15]) as feature extractor for fair comparisons. Such large performance improvement illustrates the superiority of our model to address forgetting heterogeneity among easy-to-forget and hard-to-forget old categories from both representation and gradient aspects.", + "bbox": [ + 75, + 703, + 470, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Improvement of Plug-and-Play Losses: Tab. 4 shows large performance improvement of existing knowledge", + "bbox": [ + 76, + 869, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based CIL methods [40, 50, 17, 4, 18, 32, 47], when we apply the proposed plug-and-play losses (i.e., $\\mathcal{L}_{\\mathrm{FC}}$ and $\\mathcal{L}_{\\mathrm{RD}}$ ) to them. The proposed losses $\\mathcal{L}_{\\mathrm{FC}}$ and $\\mathcal{L}_{\\mathrm{RD}}$ help existing CIL methods to address heterogeneous forgetting from gradient aspect, thus largely improving their performance. Besides, it validates generalization and robustness of our HFC model.", + "bbox": [ + 496, + 426, + 893, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 498, + 527, + 663, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 5, we present ablation studies on CIFAR-100 [33] and ImageNet-100 [10] to investigate the effectiveness of each module. TSA, GFC and GRD are task-semantic aggregation blocks, the gradient-balanced forgetting compensation loss $\\mathcal{L}_{\\mathrm{FC}}$ and gradient-balanced relation distillation loss $\\mathcal{L}_{\\mathrm{RD}}$ . Baseline denotes the performance of our model using the traditional classification loss and knowledge distillation loss proposed in DyTox (CVPR'2022) [18], but without the TSA, GFC and GRD modules. Compared with Ours, the performance of all ablation variants degrades significantly. It validates superiority of each module to overcome forgetting heterogeneity of different old categories.", + "bbox": [ + 496, + 550, + 893, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis of Task-Wise Comparisons", + "text_level": 1, + "bbox": [ + 498, + 739, + 812, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figs. 3-4 present task-wise performance comparisons between our model and other CIL methods [40, 50, 17, 18, 32] when we set backbone as ViT-Base and different tasks as $T = \\{5, 10, 20, 25\\}$ . Our HFC model outperforms baseline methods for most task-wise comparisons, which illustrates the superior performance of our model to surmount heterogeneous forgetting from representation and gradient aspects. To identify new classes continually, the task-semantic aggregation (TSA) block explores task-shared global representa", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11748", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64f4d3f7af5d54ecb9432c85770b0132db53aa3f32f9f0dc5cdad2675c18a6e6.jpg", + "image_caption": [ + "Figure 3. Analysis of task-wise performance comparisons on CIFAR-100 [33]. We set $\\mathcal{M} = 2000$ , $\\mathcal{B} = 0\\%$ when the backbone is ViT-Base." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 888, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4a3ab5333a42f77c780318b84a9bf1e3211844f8753c898d288195032a9274e.jpg", + "image_caption": [ + "Figure 4. Analysis of task-wise performance on ImageNet-100 [10]. We set $\\mathcal{M} = 2000$ , $\\mathcal{B} = 0\\%$ when the backbone is ViT-Base." + ], + "image_footnote": [], + "bbox": [ + 81, + 256, + 890, + 405 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5562063f001a7c22c7b930ababbc4b11890dcc65e0cc41a8dee95320ae2ac60c.jpg", + "table_caption": [ + "Table 6. Results on CIFAR-100 [33] when storing 20 exemplars for each class. We set $T = \\{5,10\\}$ for $\\mathcal{B} = 50\\%$ , and $\\mathbf{Ours}^{\\ddagger}$ denotes the proposed plug-and-play losses $\\mathcal{L}_{\\mathrm{FC}}$ and $\\mathcal{L}_{\\mathrm{RD}}$ ." + ], + "table_footnote": [], + "table_body": "
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M78.876.277.5↑1.9
BiC [50] (CVPR'2019)ViT-Base85.10M75.374.574.9↑4.5
PODNet [17] (ECCV'2020)ViT-Base85.10M71.671.671.6↑7.8
SS-IL [2] (ICCV'2021)ViT-Base85.10M73.071.872.4↑7.0
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M71.571.771.6↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M78.878.778.8↑0.6
AFC [32] (CVPR'2022)ViT-Base85.10M68.265.366.8↑12.6
DyTox [18] (CVPR'2022)ViT-Base85.10M79.378.478.8↑0.6
HFC (Ours)ViT-Base85.10M79.878.979.4-
Upper BoundViT-Base85.10M94.294.294.2-
iCaRL [40] (CVPR'2017)ResNet-320.46M56.653.254.9↑0.7
iCaRL [40] + Ours‡ResNet-320.46M57.353.955.6-
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M62.261.161.7↑0.6
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M62.961.662.3-
FOSTER [47] (ECCV'2022)ResNet-320.46M67.366.466.8↑0.5
FOSTER [47] + Ours‡ResNet-320.46M67.766.667.2-
AFC [32] (CVPR'2022)ResNet-320.46M63.863.663.7↑0.7
AFC [32] + Ours‡ResNet-320.46M64.464.364.4-
Upper BoundResNet-320.46M76.676.676.6-
DyTox [18] (CVPR'2022)ViT-Tiny10.71M67.966.267.1↑0.5
DyTox [18] + Ours‡ViT-Tiny10.71M68.566.767.6-
Upper BoundViT-Tiny10.71M76.176.176.1-
", + "bbox": [ + 81, + 472, + 467, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions to alleviate forgetting from representation aspect, while $\\mathcal{L}_{\\mathrm{FC}}$ and $\\mathcal{L}_{\\mathrm{RD}}$ can achieve gradient-balanced compensation.", + "bbox": [ + 76, + 758, + 470, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis of Forgetting Heterogeneity", + "text_level": 1, + "bbox": [ + 76, + 799, + 393, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 7 presents forgetting heterogeneity (FH) of different old classes via measuring their variance of gradient updating. We formulate forgetting heterogeneity (FH) as $\\mathrm{FH} = \\frac{1}{T}\\sum_{t=1}^{T}\\left(\\frac{1}{S_t}\\sum_{i=1}^{S_t}(|\\Gamma_i^t| - \\sum_{\\eta=1}^t\\Gamma_\\eta \\cdot \\mathbb{I}_{\\mathbf{y}_i^t \\in \\mathcal{Y}^\\eta})^2\\right)$ , where $\\Gamma_i^t$ and $\\Gamma_\\eta$ are obtained from Eqs. (6)-(7), and $S_t$ is", + "bbox": [ + 75, + 821, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/506a6aa3d7d9b0d5e1533bb9f70c5ca6949761f0bbef6093d13907414c091ea8.jpg", + "table_caption": [ + "Table 7. Analysis of forgetting heterogeneity (FH) on CIFAR-100 [33] when $T = 5, \\mathcal{B} = 0\\%$ . AC denotes the averaged accuracy." + ], + "table_footnote": [], + "table_body": "
M=2000,B=0%BackboneTSAGFCGRDFHAC
iCaRL [40] (CVPR'2017)ViT-Base---131.583.5
FOSTER [47] (ECCV'2022)ViT-Base---119.485.1
AFC [32] (CVPR'2022)ViT-Base---108.574.6
DyTox [18] (CVPR'2022)ViT-Base---105.185.5
BaselineViT-BaseXXX128.981.6
Baseline + TSAViT-Base✓XX125.483.1
Baseline + TSA + GFCViT-Base✓✓X120.584.0
HFC (Ours)ViT-Base✓✓✓103.686.4
", + "bbox": [ + 509, + 458, + 880, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "number of samples in test set. Tab. 7 verifies superiority of our model against other CIL methods to tackle heterogeneous forgetting. It also shows effectiveness of all modules to collaboratively minimize forgetting heterogeneity.", + "bbox": [ + 496, + 597, + 893, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 617, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we develop a novel Heterogeneous Forgetting Compensation (HFC) model to surmount heterogeneous forgetting from representation and gradient aspects. To be specific, a task-semantic aggregation block is designed to tackle heterogeneous forgetting from representation aspect via exploring task-shared global representations. Moreover, we propose a gradient-balanced forgetting compensation loss and a relation distillation loss to compensate forgetting heterogeneity from gradient aspect via performing balanced gradient propagation and distilling heterogeneous class relations. Experiments verify the superiority of our proposed HFC model against baseline methods. We will further consider addressing noisy forgetting heterogeneity in the future.", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11749", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, pages 3930-3939, 2020.", + "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, pages 844-853, October 2021.", + "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 144-161, 2018.", + "[4] Arjun Ashok, K. J. Joseph, and Vineeth N. Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In ECCV, pages 105-122, 2022.", + "[5] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022.", + "[6] Francisco M. Castro, Manuel J. Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 241-257, Cham, 2018. Springer International Publishing.", + "[7] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023.", + "[8] Huitong Chen, Yu Wang, and Qinghua Hu. Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering, pages 1-15, 2022.", + "[9] Stéphane d'Ascoli, Hugo Touvron, Matthew Leavitt, Ari Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021.", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009.", + "[11] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11-20, 2021.", + "[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In CVPR, pages 10164-10173, June 2022.", + "[13] Jiahua Dong, Duzhen Zhang, Yang Cong, Wei Cong, Henghui Ding, and Dengxin Dai. Federated incremental semantic segmentation. In CVPR, pages 3934-3943, June 2023.", + "[14] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In CVPR, pages 19217-19227, June 2022." + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021.", + "[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In CVPR, 2021.", + "[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020.", + "[18] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285-9295, June 2022.", + "[19] Zhen Fang, Yixuan Li, Jie Lu, Jiahua Dong, Bo Han, and Feng Liu. Is out-of-distribution detection learnable? In NeurIPS, volume 35, pages 37199-37213, 2022.", + "[20] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. In ICML, page 1243-1252, 2017.", + "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, volume 27, 2014.", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, June 2022.", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.", + "[24] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. In NeurIPS Workshop, 2015.", + "[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, June 2019.", + "[26] Weiming Hu, Xi Li, Guodong Tian, Stephen Maybank, and Zhongfei Zhang. An incremental dpmm-based method for trajectory clustering, modeling, and retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(5):1051-1065, 2013.", + "[27] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021.", + "[28] Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unforgetting continual learning. In Advances in Neural Information Processing Systems, pages 13647–13657, 2019.", + "[29] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 699-715, 2020." + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11750", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, and et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022.", + "[31] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwar, and Vineeth Balasubramanian. Energy-based latent aligner for incremental learning. In CVPR, 2022.", + "[32] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, pages 16071-16080, June 2022.", + "[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, pages 32-33, 2009.", + "[34] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018.", + "[35] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2022.", + "[36] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. AAAI, 37(7):8906-8913, Jun. 2023.", + "[37] Yaoyao Liu, Bernt Schiele, and Qianru Sun. RMM: reinforced memory management for class-incremental learning. In NeurIPS, pages 3478-3490, 2021.", + "[38] John M. Pierre. Incremental lifelong deep learning for autonomous vehicles. In ITSC, pages 3949-3954, 2018.", + "[39] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for continual learning. In NeurIPS, 2019.", + "[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017.", + "[41] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, volume 80, pages 4548-4557, 2018.", + "[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In CVPR, 2022.", + "[43] Zhenheng Tang, Yonggang Zhang, Shaohuai Shi, Xin He, Bo Han, and Xiaowen Chu. Virtual homogeneity learning: Defending against data heterogeneity in federated learning. In ICML, pages 21111-21132, 2022.", + "[44] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 254-270, Cham, 2020. Springer International Publishing.", + "[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, pages 99-108, June 2022.", + "[46] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022.", + "[48] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In AAAI, volume 35, pages 10165-10173, 2021.", + "[49] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022.", + "[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019.", + "[51] Ziyang Wu, Christina Baek, Chong You, and Yi Ma. Incremental learning via rate reduction. In CVPR, pages 1125-1133, 2021.", + "[52] Rongtao Xu, Changwei Wang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Wave-like class activation map with representation fusion for weakly-supervised semantic segmentation. IEEE Transactions on Multimedia, 2023.", + "[53] Rongtao Xu, Changwei Wang, Jiguang Zhang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. RSSformer: Foreground saliency enhancement for remote sensing land-cover segmentation. IEEE Transactions on Image Processing, 32:1052-1064, 2023.", + "[54] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021.", + "[55] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018.", + "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, page 3987-3995, 2017.", + "[57] Duzhen Zhang, Feilong Chen, and Xiuyi Chen. Dualgats: Dual graph attention networks for emotion recognition in conversations. In ACL, pages 7395-7408, 2023.", + "[58] Yonggang Zhang, Mingming Gong, Tongliang Liu, Gang Niu, Xinmei Tian, Bo Han, Bernhard Schölkopf, and Kun Zhang. Causaladv: Adversarial robustness through the lens of causality. *ICLR*, 2022.", + "[59] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13205-13214, 2020.", + "[60] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021.", + "[61] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In ECCV, pages 55-72, 2022." + ], + "bbox": [ + 501, + 92, + 893, + 815 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11751", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_model.json b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d39def2bf7c97a04b7d00092c5ce2fbd7e5f4c4b --- /dev/null +++ b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.131, + 0.856, + 0.156 + ], + "angle": 0, + "content": "Heterogeneous Forgetting Compensation for Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.18, + 0.735, + 0.199 + ], + "angle": 0, + "content": "Jiahua Dong\\(^{1,2,3*}\\), Wenqi Liang\\(^{1,2,3*}\\), Yang Cong\\(^{4\\dagger}\\), Gan Sun\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.2, + 0.762, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, 110016, China.‡" + }, + { + "type": "text", + "bbox": [ + 0.256, + 0.234, + 0.714, + 0.269 + ], + "angle": 0, + "content": "\\(^{2}\\)Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang, 110169, China." + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.27, + 0.761, + 0.287 + ], + "angle": 0, + "content": "3University of Chinese Academy of Sciences, Beijing, 100049, China." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.756, + 0.305 + ], + "angle": 0, + "content": "\\(^{4}\\)South China University of Technology, Guangzhou, 510640, China." + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.307, + 0.771, + 0.323 + ], + "angle": 0, + "content": "{dongjiahua1995, liangwenqi0123, congyang81, sunganl412}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.474, + 0.754 + ], + "angle": 0, + "content": "Class-incremental learning (CIL) has achieved remarkable successes in learning new classes consecutively while overcoming catastrophic forgetting on old categories. However, most existing CIL methods unreasonably assume that all old categories have the same forgetting pace, and neglect negative influence of forgetting heterogeneity among different old classes on forgetting compensation. To surmount the above challenges, we develop a novel Heterogeneous Forgetting Compensation (HFC) model, which can resolve heterogeneous forgetting of easy-to-forget and hard-to-remember old categories from both representation and gradient aspects. Specifically, we design a task-semantic aggregation block to alleviate heterogeneous forgetting from representation aspect. It aggregates local category information within each task to learn task-shared global representations. Moreover, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation loss and a gradient-balanced relation distillation loss to alleviate forgetting from gradient aspect. They consider gradient-balanced compensation to rectify forgetting heterogeneity of old categories and heterogeneous relation consistency. Experiments on several representative datasets illustrate effectiveness of our HFC model. The code is available at https://github.com/JiahuaDong/HFC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.782, + 0.21, + 0.797 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.471, + 0.854 + ], + "angle": 0, + "content": "Class-incremental learning (CIL) [40, 12, 5] has attracted appealing attentions recently by accumulating previous learned experience to learn new classes incrementally. It" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.356, + 0.872, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.543, + 0.894, + 0.573 + ], + "angle": 0, + "content": "Figure 1. Forgetting heterogeneity versus accuracy on CIFAR-100 [33] when the backbone is ViT-Base and number of tasks is 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.895, + 0.764 + ], + "angle": 0, + "content": "plays an indispensable role in developing a large number of intelligent learning systems, such as autonomous driving [38] and automated surveillance [26]. When learning new classes continuously under the settings of limited memory to replay all previous data of old classes [45], these CIL systems heavily suffer from forgetting on old classes caused by severe class imbalance between old and new categories [40, 31, 14]. To surmount catastrophic forgetting, a growing amount of CIL methods mainly perform knowledge distillation [24] to preserve past experience [34, 12, 35]; introduce a rehearsal strategy to replay part of old data [40, 17]; or dynamically expand network architectures [28, 41, 54]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "However, most existing CIL methods [40, 45, 54, 18] neglect heterogeneous forgetting speeds of different old classes. They unrealistically assume that all old classes suffer from the same degree of catastrophic forgetting, and compensate forgetting for each old classes equally and independently. Such impracticable assumption enforces existing CIL models [40, 56, 61] to suffer from imbalanced gradient optimization among different old classes, thus favoring more forgetting compensation for hard-to-forget old classes while ne" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.472, + 0.901 + ], + "angle": 0, + "content": "*Equal contributions. †The corresponding author is Prof. Yang Cong. ‡This work was supported in part by the National Nature Science Foundation of China under Grant 62127807, 62273333 and 62133005." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "glecting those easy-to-forget old classes (i.e., heterogeneous forgetting). More importantly, such forgetting heterogeneity can significantly worsen the forgetting on hard-to-forget old categories, when new streaming classes becomes part of old categories continually. For example, some old classes with various modalities and appearances (e.g., person and car) in autonomous driving [38] are more difficult to explore task-shared representations across different incremental tasks, when compared with other hard-to-forget old classes (e.g., road and traffic sign) with easily-distinguished visual properties. This phenomenon causes imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus exacerbating heterogeneous forgetting speeds among different old classes. When autonomous vehicles [38] learn new classes continually, such forgetting heterogeneity aggravates forgetting on those hard-to-forget old classes (e.g., traffic sign) to some extent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Inspired by the above practical scenario, we investigate that different old classes have significant forgetting heterogeneity in this paper, as shown in Fig. 1. This heterogeneous forgetting might heavily weaken forgetting compensation on hard-to-forget old classes, as new classes become a subset of old classes consecutively. In summary, the challenges to tackle heterogeneous forgetting lie in two major aspects:" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.459, + 0.473, + 0.58 + ], + "angle": 0, + "content": "- Representation Aspect: Some easy-to-forget old classes (e.g., car and bus) with diverse appearances are more difficult for existing CIL methods [45, 54, 12, 3] to learn intrinsic task-shared representations, and thus are significantly easier to be forgotten than hard-to-forget old classes (e.g., road) with distinctive attributes. Thus, exploring task-shared representations is essential to address heterogeneous forgetting among different old classes." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.58, + 0.473, + 0.687 + ], + "angle": 0, + "content": "- Gradient Aspect: To learn complex visual characterizations for easy-to-forget old classes with various modalities and appearances, existing CIL methods [40, 4, 49] are required to allocate more network architectures for gradient updating. It can result in imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus aggravating forgetting heterogeneity of old classes." + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.459, + 0.473, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To overcome the above challenges, we propose a novel Heterogeneous Forgetting Compensation (HFC) model, which is an earlier exploration to address heterogeneous forgetting from both representation and gradient perspectives in the CIL field [40, 4]. Specifically, we propose a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation perspective. It can explore task-shared global representations for each class via aggregating long-range local category information within each task. Meanwhile, to tackle heterogeneous forgetting from gradient perspective, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss. The GFC loss can rectify heterogeneous forgetting speeds" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "of easy-to-forget and hard-to-forget old classes, while normalize different learning paces of new categories to achieve gradient-balanced propagation. Besides, the GRD loss can distill heterogeneous relation consistency brought by forgetting heterogeneity among different old classes. Experiments show our model has large improvements on several representative datasets, compared with baseline methods [40, 36]. More importantly, we apply two plug-and-play losses into existing distillation-based CIL methods [40, 17, 4] to significantly improve their performance from gradient aspect. The novel contributions of this paper are presented as follows:" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.263, + 0.895, + 0.354 + ], + "angle": 0, + "content": "- We develop a novel Heterogeneous Forgetting Compensation (HFC) model to address different forgetting speeds of easy-to-forget and hard-to-forget old classes. To our best knowledge, this paper is the first exploration to tackle forgetting heterogeneity among old categories from representation and gradient aspects in the CIL field." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.354, + 0.895, + 0.415 + ], + "angle": 0, + "content": "- We design a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect. It can explore robust task-shared representations for each class via aggregating local category information." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.415, + 0.895, + 0.522 + ], + "angle": 0, + "content": "- We propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss to surmount forgetting heterogeneity from gradient aspect. They can balance different forgetting of old classes and heterogeneous category-relation consistency to improve performance when applying them into existing methods." + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.263, + 0.895, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.642, + 0.554 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.565, + 0.895, + 0.595 + ], + "angle": 0, + "content": "We discuss some class-incremental learning (CIL) methods [60, 37, 18], and divide them into three categories:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.718 + ], + "angle": 0, + "content": "Knowledge Distillation: After Li et al. [34] apply knowledge distillation [24] to continual learning, [40, 39, 59] follow them to transfer past experience from old model to new model. iCaRL [40] and EEIL [6] perform knowledge distillation in the output space. [44] aims to preserve consistent topology feature space for old and new tasks. [16] proposes feature-level knowledge distillation when applying continual learning [37] into semantic segmentation task [53, 52, 11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Rehearsal Strategy: For replaying past experience, lots of CIL methods [49, 8, 42] allocate a memory to store exemplars of old classes or synthesize samples via generative adversarial model [21, 58]. For saving large memory overhead, [29] proposes to store low-dimension features rather than raw samples. [51] proposes a \"white box\" framework deviated from rate reduction to preserve past experience." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Dynamic Architecture: Many works [28, 54, 1] allocate dynamical networks for new classes as the growing of learned classes. Yoon et al. [55] dynamically expand architecture capacity via selective retraining. [54, 18] address the problem of relying on task index when performing dynamic" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11743" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.472, + 0.138 + ], + "angle": 0, + "content": "architectures. However, the above-mentioned CIL methods [40, 27] neglect negative influence of heterogeneous forgetting of old classes on forgetting compensation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.153, + 0.277, + 0.17 + ], + "angle": 0, + "content": "3. The Proposed Model" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.179, + 0.371, + 0.194 + ], + "angle": 0, + "content": "3.1. Problem Definition and Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Problem Definition: Following class-incremental learning (CIL) methods [40, 50, 17, 18, 54], we set a series of consecutive learning tasks as \\(\\mathcal{T} = \\{\\mathcal{T}^t\\}_{t=1}^T\\), where \\(T\\) indicates the total task quantity. The \\(t\\)-th task \\(\\mathcal{T}^t = \\{\\mathbf{x}_i^t, \\mathbf{y}_i^t\\}_{i=1}^{N^t}\\) is composed of \\(N^t\\) pairs of image \\(\\mathbf{x}_i^t\\) and one-hot groundtruth \\(\\mathbf{y}_i^t \\in \\mathcal{Y}^t\\), and \\(\\mathcal{Y}^t\\) denotes label space of the \\(t\\)-th incremental task containing \\(K^t\\) new categories. The label spaces between any two incremental tasks have no overlap: \\(\\mathcal{Y}^t \\cap (\\cup_{j=1}^{t-1} \\mathcal{Y}^j) = \\emptyset\\). Namely, \\(K^t\\) new classes belonging to the \\(t\\)-th task \\(\\mathcal{T}^t\\) are different from \\(K^o = \\sum_{i=1}^{t-1} K^i \\subset \\cup_{j=1}^{t-1} \\mathcal{Y}^j\\) old classes learned from previous tasks \\(\\{\\mathcal{T}^i\\}_{i=1}^{t-1}\\). As introduced in CIL baselines [40, 50, 54], we set a fixed exemplar memory \\(\\mathcal{M}\\), and store only few images (i.e., \\(\\frac{|M|}{K^o}\\)) for each old class in the \\(t\\)-th task \\(\\mathcal{T}^t\\), where \\(\\mathcal{M}\\) satisfies \\(\\frac{N^t}{K^t} \\gg \\frac{|M|}{K^o}\\). In the \\(t\\)-th task, we aim to identify both \\(K^t\\) new classes and \\(K^o\\) old classes via optimizing the model on \\(\\mathcal{T}^t\\) and \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.76 + ], + "angle": 0, + "content": "Overview: The overview of our HFC model to surmount forgetting heterogeneity of different old categories from both representation and gradient aspects is shown in Fig. 2. Denote the proposed HFC model learned at the \\((t - 1)\\)-th and \\(t\\)-th tasks as old and new models (i.e., \\(\\Theta^{t - 1}\\) and \\(\\Theta^t\\)), where \\(\\Theta^t\\) is inherited from \\(\\Theta^{t - 1}\\) and only expands the number of output neurons in the classifier to identify \\(K^t\\) new classes. In the \\(t\\)-th incremental task, given an image \\(\\mathbf{x}_i^t\\in \\mathcal{T}^t\\cup \\mathcal{M}\\), we forward it into \\(\\Theta^{t - 1}\\) and \\(\\Theta^t\\) to extract task-shared global representation via task-semantic aggregation (TSA) block, which alleviates heterogeneous forgetting from representation aspect (Section 3.2). After obtaining probabilities \\(\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o}\\) and \\(\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}\\) predicted via \\(\\Theta^{t - 1}\\) and \\(\\Theta^t\\), we develop two plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss \\(\\mathcal{L}_{\\mathrm{FC}}\\) and a gradient-balanced relation distillation (GRD) loss \\(\\mathcal{L}_{\\mathrm{RD}}\\) to surmount heterogeneous forgetting from semantic aspect (Sections 3.3 and 3.4). \\(\\mathcal{L}_{\\mathrm{FC}}\\) and \\(\\mathcal{L}_{\\mathrm{RD}}\\) can normalize heterogeneous forgetting of old classes and distill heterogeneous relation consistency between \\(\\Theta^{t - 1}\\) and \\(\\Theta^t\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.375, + 0.788 + ], + "angle": 0, + "content": "3.2. Task-Semantic Aggregation Block" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.902 + ], + "angle": 0, + "content": "To explore long-range semantic dependencies across different tasks, Douillard et al. [18] introduce Vision Transformer (ViT) [15, 46, 22] into class-incremental learning (CIL). However, it significantly increases the memory overhead to store task tokens and task-specific decoders, when learning large-scale new classes. Besides, it cannot aggregate local category information to explore task-shared global" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "representations, resulting in heterogeneous forgetting from representation aspect. To tackle these issues, we develop a task-semantic aggregation (TSA) block to extract task-shared global representations, which is effective to tackle heterogeneous forgetting from representation aspect. More importantly, the TSA block cannot dynamically increase memory cost with the arrival of large-scale new categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Specifically, as presented in Fig. 2, the proposed model \\(\\Theta^t = \\Theta_c^t\\circ \\Theta_f^t\\) at the \\(t\\)-th task is composed of a feature extractor \\(\\Theta_f^t\\) to extract task-shared global representations, and a task classifier \\(\\Theta_c^t\\) to classify \\(K^o\\) old classes and \\(K^t\\) new classes. In the \\(t\\)-th task, the network parameters of \\(\\Theta_f^{t - 1}\\) and \\(\\Theta_c^{t - 1}\\) in \\(\\Theta^{t - 1}\\) learned at the \\((t - 1)\\)-th task are frozen to preserve semantic knowledge for old classes. Different from [18] that uses vanilla ViT [15, 30, 57, 7] to extract features, we employ \\(l_{m}\\) multi-head self-attention (MSA) blocks [20] and \\(l_{t}\\) task-semantic aggregation (TSA) block as feature extractor \\(\\Theta_f^t\\) to obtain task-shared global representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.895, + 0.598 + ], + "angle": 0, + "content": "- Multi-Head Self-Attention (MSA): Given an image \\(\\mathbf{x}_i^t\\in \\mathcal{T}^t\\cup \\mathcal{M}\\) in the \\(t\\) th task, following ViT [15, 9], we crop it into \\(N\\) patches with equal dimension, flatten these patches and map them into a \\(D\\) -dimension feature space via a linear projection to obtain patch embedding \\(\\mathbf{Z}_e\\in \\mathbb{R}^{N\\times D}\\). Then we concatenate \\(\\mathbf{Z}_e\\) with a class token \\(\\mathbf{Z}_{\\mathrm{cls}}\\in \\mathbb{R}^{D}\\), and perform element-wise sum operation with a position token \\(\\mathbf{Z}_p\\in \\mathbb{R}^{(N + 1)\\times D}\\) to get \\(\\mathbf{Z}_0 = [\\mathbf{Z}_e\\oplus \\mathbf{Z}_{\\mathrm{cls}}] + \\mathbf{Z}_p\\in \\mathbb{R}^{(N + 1)\\times D}\\) where \\(\\oplus\\) is concatenation function. As shown in Fig.2, \\(\\mathbf{Z}_0\\) is fed into \\(\\Theta_f^t\\) including \\(l_{m}\\) MSA blocks to extract long-range semantics across different tasks. For the \\(l\\) -th \\((l = 1,\\dots ,l_m)\\) MSA block, we regard \\(\\mathbf{Z}_{l - 1}\\in \\mathbb{R}^{(N + 1)\\times D}\\) learned from the \\((l - 1)\\) -th MSA block as input, and execute parallel selfattention \\(H\\) times. Thus, the self-attention \\(\\mathbf{A}_l^h\\in \\mathbb{R}^{(N + 1)\\times d}\\) of \\(\\mathbf{Z}_{l - 1}\\) for the \\(h\\) -th \\((h = 1,\\dots ,H)\\) head is:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.609, + 0.893, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {l} ^ {h} = \\sigma \\left(\\frac {\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {q} ^ {h} \\left(\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {k} ^ {h}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {v} ^ {h}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.812 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_q^h,\\mathbf{W}_k^h,\\mathbf{W}_v^h\\in \\mathbb{R}^{D\\times d}\\) are query, key and value mapping matrices. \\(d = \\frac{D}{H}\\) represents channel dimension of each head and \\(\\sigma\\) is the softmax function. Then we concatenate \\(\\{\\mathbf{A}_l^h\\}_{h = 1}^H\\) along channel dimension and project this concatenated result via a mapping matrix \\(\\mathbf{W}_o\\in \\mathbb{R}^{D\\times D}\\) to perform MSA: \\(\\mathrm{MSA}(\\mathbf{Z}_{l - 1}) = \\mathbf{Z}_{l - 1} + [\\mathbf{A}_l^1\\oplus\\) \\(\\mathbf{A}_l^2,\\dots ,\\oplus \\mathbf{A}_l^H ]\\mathbf{W}_o\\) where \\(\\oplus\\) is concatenation function. After obtaining \\(\\mathrm{MSA}(\\mathbf{Z}_{l - 1})\\in \\mathbb{R}^{(N + 1)\\times D}\\) \\((D = dH)\\), we utilize a multi-layer perceptron (MLP) block to obtain output \\(\\mathbf{Z}_l\\in \\mathbb{R}^{(N + 1)\\times D}\\) of the \\(l\\) th MSA block:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.825, + 0.893, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {l} = \\operatorname {M S A} \\left(\\mathbf {Z} _ {l - 1}\\right) + \\operatorname {M L P} \\left(\\operatorname {M S A} \\left(\\mathbf {Z} _ {l - 1}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- Task-Semantic Aggregation (TSA): The patch embedding \\(\\mathbf{Z}_{l_m} \\in \\mathbb{R}^{(N + 1) \\times D}\\) encoded via the \\(l_m\\)-th MSA block contains rich long-range semantic dependencies shared" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11744" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.317, + 0.894, + 0.373 + ], + "angle": 0, + "content": "Figure 2. Illustration of the HFC model. It mainly contains a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect via exploring task-shared global representations, and two novel plug-and-play losses (i.e., a gradient-balanced forgetting compensation loss \\(\\mathcal{L}_{\\mathrm{FC}}\\) and a gradient-balanced relation distillation loss \\(\\mathcal{L}_{\\mathrm{RD}}\\)) to overcome forgetting heterogeneity from gradient aspect by compensating gradient-imbalanced propagation and heterogeneous class-relation consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.474, + 0.598 + ], + "angle": 0, + "content": "across different tasks. \\(\\mathbf{Z}_{l_m}\\) is then forwarded into \\(l_{t}\\) task-semantic aggregation (TSA) blocks to extract task-shared global embedding. The TSA block aggregates local category context from long-range semantic dependencies to global representations shared across all incremental tasks, which is essential to alleviate heterogeneous forgetting from representation aspect. In the \\(t\\)-th task, as depicted in Fig. 2, we introduce a learnable task-shared embedding \\(\\mathbf{E}_0^t \\in \\mathbb{R}^D\\) that is initialized via \\(\\mathbf{E}_0^{t - 1}\\) learned in the \\((t - 1)\\)-th task. Then we forward \\(\\mathbf{E}_0^t\\) along with \\(\\mathbf{Z}_{l_m}\\) to \\(l_{t}\\) TSA blocks. For the \\(l\\)-th \\((l = 1,\\dots ,l_t)\\) TSA block, \\(\\mathbf{E}_{l - 1}^t \\in \\mathbb{R}^D\\) learned at \\((l - 1)\\)-th TSA block and \\(\\mathbf{Z}_{l_m}\\) are used to perform task-semantic attention \\(H\\) times in parallel. Thus, we formulate task-semantic attention \\(\\mathbf{S}_l^h \\in \\mathbb{R}^d\\) for the \\(h\\)-th \\((h = 1,\\dots ,H)\\) head as:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.605, + 0.47, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {l} ^ {h} = \\sigma \\left(\\frac {\\mathbf {E} _ {l - 1} ^ {t} \\mathbf {V} _ {q} ^ {h} \\left(\\mathbf {Z} _ {l _ {m}} \\mathbf {V} _ {k} ^ {h}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {Z} _ {l _ {m}} \\mathbf {V} _ {v} ^ {h}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.473, + 0.76 + ], + "angle": 0, + "content": "where \\(\\mathbf{V}_q^h,\\mathbf{V}_k^h,\\mathbf{V}_v^h\\in \\mathbb{R}^{D\\times d}\\) are projection matrices of query, key and value. After utilizing Eq. (3) to obtain \\(\\{\\mathbf{S}_l^h\\}_{h = 1}^H\\) , we concatenate them along channel dimension and project this fused result via \\(\\mathbf{V}_o\\in \\mathbb{R}^{D\\times D}\\) to execute \\(H\\) -head task-semantic attention: \\(\\mathrm{TSA}(\\mathbf{E}_{l - 1}^{t}) = [\\mathbf{S}_{l}^{1}\\oplus\\) \\(\\mathbf{S}_l^2,\\dots ,\\oplus \\mathbf{S}_l^H ]\\mathbf{V}_o\\in \\mathbb{R}^D\\) . TSA \\((\\mathbf{E}_{l - 1}^{t})\\) is encoded via a MLP block to get output \\(\\mathbf{E}_l^t\\in \\mathbb{R}^D\\) of the \\(l\\) -th TSA block:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.769, + 0.47, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {l} ^ {t} = \\operatorname {T S A} \\left(\\mathbf {E} _ {l - 1} ^ {t}\\right) + \\operatorname {M L P} \\left(\\operatorname {T S A} \\left(\\mathbf {E} _ {l - 1} ^ {t}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As aforementioned, we obtain the task-shared global representation \\(\\mathbf{E}_{l_t}^t\\in \\mathbb{R}^D\\) via the \\(l_{t}\\) -th TSA block in feature extractor \\(\\Theta_f^t\\) . Obviously, the proposed TSA blocks can aggregate local category context of the \\(t\\) -th task from longrange semantic dependencies to task-shared global representation \\(\\mathbf{E}_{l_t}^t\\) . It is essential to alleviate heterogeneous forgetting among different old classes from representation aspect, by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.895, + 0.431 + ], + "angle": 0, + "content": "exploring task-shared semantic context across different tasks. Note that we employ layer norm before TSA, MSA and MLP blocks, while we omit it for simplicity in this paper." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.43, + 0.894, + 0.495 + ], + "angle": 0, + "content": "- Task Classifier: The task-shared global representation \\(\\mathbf{E}_{l_t}^t\\) obtained via \\(\\Theta_f^t\\) is fed into task classifier \\(\\Theta_c^t\\) to predict \\(K^o\\) old classes and \\(K^t\\) new classes. Given a mini-batch \\(\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b\\in \\mathcal{T}^t\\cup \\mathcal{M}\\), the classification loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.5, + 0.893, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\mathcal {D} _ {\\mathrm {C E}} \\left(\\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right), \\mathbf {y} _ {i} ^ {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.596 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}\\) is softmax probability predicted by \\(\\Theta^t\\) in the \\(t\\)-th learning task. \\(\\mathcal{D}_{\\mathrm{CE}}(\\cdot ,\\cdot)\\) indicates traditional cross-entropy loss, and \\(b\\) is the batch size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.78 + ], + "angle": 0, + "content": "However, the severe class imbalance (i.e., \\(\\frac{N^t}{K^t} \\gg \\frac{|\\mathcal{M}|}{K^o}\\)) among old and new categories in the \\(t\\)-th task enforces the prediction \\(\\mathbf{P}^t(\\mathbf{x}_i^t, \\Theta^t)\\) in Eq. (5) to suffer from catastrophic forgetting [40] on old categories. Moreover, the gradient optimization of Eq. (5) neglects heterogeneous forgetting speeds of easy-to-forget old classes with various appearances and hard-to-forget old classes with easily-distinguished attributes. To address heterogeneous forgetting from gradient aspect, we propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss \\(\\mathcal{L}_{\\mathrm{FC}}\\) in Section 3.3 and a gradient-balanced relation distillation (GRD) loss \\(\\mathcal{L}_{\\mathrm{RD}}\\) in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.883, + 0.804 + ], + "angle": 0, + "content": "3.3. Gradient-Balanced Forgetting Compensation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "As incremental tasks arrive consecutively, easy-to-forget and hard-to-forget old categories change dynamically, thus aggravating difficulty to tackle heterogeneous forgetting from gradient aspect. In light of this, we develop a gradient-balanced forgetting compensation (GFC) loss \\(\\mathcal{L}_{\\mathrm{FC}}\\). It adaptively balances large forgetting heterogeneity of old classes" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11745" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.199 + ], + "angle": 0, + "content": "via performing balanced gradient propagation. Specifically, we perform task-adaptive gradient normalization for different classes, and then reweight classification loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) in Eq. (5) to compensate imbalanced gradients. For the \\(t\\)-th incremental task, as claimed in [48, 13], the gradient \\(\\Gamma_i^t\\) of an given image \\((\\mathbf{x}_i^t,\\mathbf{y}_i^t)\\subset \\mathcal{T}^t\\cup \\mathcal{M}\\) with respect to the \\(k\\)-th \\((k = \\arg \\max \\mathbf{y}_i^t)\\) neuron \\(\\mathcal{N}_k^t\\) of classifier \\(\\Theta_c^t\\) in \\(\\Theta^t\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.205, + 0.471, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {i} ^ {t} = \\frac {\\partial \\mathcal {D} _ {\\mathrm {C E}} (\\mathbf {P} ^ {t} (\\mathbf {x} _ {i} ^ {t} , \\boldsymbol {\\Theta} ^ {t}) , \\mathbf {y} _ {i} ^ {t})}{\\partial \\mathcal {N} _ {k} ^ {t}} = \\mathbf {P} ^ {t} (\\mathbf {x} _ {i} ^ {t}, \\boldsymbol {\\Theta} ^ {t}) _ {k} - 1, \\quad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.449, + 0.264 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)_k\\) denotes probability of the \\(k\\)-th class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.472, + 0.354 + ], + "angle": 0, + "content": "To compensate heterogeneous forgetting of easy-to-forget and hard-to-forget old categories, we compute task-adaptive gradient means for different tasks. Consequently, given a mini-batch \\(\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b\\) in the \\(t\\)-th learning task, the task-adaptive gradient mean \\(\\Gamma_{\\eta}\\) for the categories learned in the \\(\\eta\\)-th \\((1\\leq \\eta \\leq t)\\) incremental task is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.361, + 0.471, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {\\eta} = \\frac {1}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}} \\sum_ {i = 1} ^ {b} | \\Gamma_ {i} ^ {t} | \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.47, + 0.45 + ], + "angle": 0, + "content": "The forgetting heterogeneity of old categories learned from \\( t - 1 \\) old tasks and learning speeds of new categories in the \\( t \\)-th task can be effectively measured via \\( \\{\\Gamma_{\\eta}\\}_{\\eta = 1}^{t} \\) in Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.472, + 0.556 + ], + "angle": 0, + "content": "When easy-to-forget and hard-to-remember old classes are varied dynamically as incremental tasks, some noisy predictions on easy-to-remember old classes may result in large deviation to measure their forgetting heterogeneity [43, 19]. To tackle this issue, \\(\\{\\Gamma_{\\eta}\\}_{\\eta = 1}^{t}\\) in Eq. (7) are expected to be sharper and more distinguishable adaptively [13], as learning tasks arrive continually. Thus, we rewrite Eq. (7) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.563, + 0.47, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {\\eta} ^ {s} = \\frac {\\sum_ {i = 1} ^ {b} \\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right) \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.678 + ], + "angle": 0, + "content": "The sharper task-adaptive gradient means \\(\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}\\) computed via Eq. (8) are employed to reweight classification loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) in Eq. (5). Then the gradient-balanced forgetting compensation (GFC) loss \\(\\mathcal{L}_{\\mathrm{FC}}\\) is written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.684, + 0.47, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F C}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\frac {\\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right)}{\\sum_ {\\eta = 1} ^ {t} \\Gamma_ {\\eta} ^ {s} \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}} \\mathcal {D} _ {\\mathrm {C E}} \\left(\\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right), \\mathbf {y} _ {i} ^ {t}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.827 + ], + "angle": 0, + "content": "Obviously, \\(\\mathcal{L}_{\\mathrm{FC}}\\) in Eq. (9) encourages the optimization of \\(\\Theta^t\\) to perform gradient-balanced propagation for easy-to- forget, hard-to- forget old classes and new classes learned in different tasks. It tackles heterogeneous forgetting on old classes from gradient aspect, via normalizing forgetting paces of different old classes adaptively with \\(\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.42, + 0.848 + ], + "angle": 0, + "content": "3.4. Gradient-Balanced Relation Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "The relationships among old and new categories remain constant in semantic space, regardless of availability of old-class samples. In light of this, distilling inter-class relations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "from old model \\(\\Theta^{t-1}\\) to new model \\(\\Theta^t\\) can address forgetting on old classes [6, 25]. However, most knowledge distillation methods used in CIL [40, 2, 50] only utilize prediction of an individual sample to perform semantic consistency of old classes between \\(\\Theta^{t-1}\\) and \\(\\Theta^t\\). They neglect underlying relations among old and new categories to tackle forgetting, and heavily suffer from noisy predictions on easy-to-forget old classes when distilling knowledge via an individual sample. Moreover, different forgetting speeds of old classes may result in heterogeneous class-relation consistency between \\(\\Theta^{t-1}\\) and \\(\\Theta^t\\). It enforces gradient optimization of \\(\\Theta^t\\) to bias towards some specific relation distillations related to new classes, causing forgetting heterogeneity on old categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.468 + ], + "angle": 0, + "content": "To overcome the above issues, we design a gradient-balanced relation distillation (GRD) loss \\(\\mathcal{L}_{\\mathrm{RD}}\\) to distill relations among old and new classes from \\(\\Theta^{t - 1}\\) to \\(\\Theta^t\\). As introduced in Fig. 2, it can address heterogeneous forgetting from gradient aspect via rectifying imbalanced gradient propagation caused by heterogeneous class-relation consistency. Specifically, we construct class-wise relation prototype rather than a prediction of single sample to perform relation distillation, which can alleviate negative effect of noisy relation consistency. Then the task-adaptive gradient means \\(\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}\\) in Eq. (8) are employed to reweight heterogeneous class-relation distillation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.895, + 0.61 + ], + "angle": 0, + "content": "A mini-batch \\(\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b\\subset \\mathcal{T}^t\\cup \\mathcal{M}\\) is fed into \\(\\Theta^{t - 1}\\) and \\(\\Theta^t\\) to obtain softmax probabilities \\(\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o}\\) of old classes and \\(\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}\\) of old and new classes, as shown in Fig. 2. The first \\(K^o\\) dimensions of \\(\\mathbf{y}_i^t\\in \\mathbb{R}^{K^o +K^t}\\) are replaced with \\(\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})\\) to get relation groundtruth \\(\\mathbf{Y}^t (\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o +K^t}\\) that implies underlying relations among old and new classes. To tackle noisy class relations, we construct class-wise relation prototype \\(\\tilde{\\mathbf{P}}_k^t\\) and its relation groundtruth \\(\\tilde{\\mathbf{Y}}_k^t\\) for the \\(k\\) -th class:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.616, + 0.892, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {P}} _ {k} ^ {t} = \\frac {1}{\\Delta_ {k}} \\sum_ {i = 1} ^ {b} \\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.66, + 0.892, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {Y}} _ {k} ^ {t} = \\frac {1}{\\Delta_ {k}} \\sum_ {i = 1} ^ {b} \\mathbf {Y} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t - 1}\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.895, + 0.743 + ], + "angle": 0, + "content": "where \\(\\Delta_k = \\sum_i^b\\mathbb{I}_{\\mathrm{argmax}\\mathbf{y}_i^t = k}\\). Then we formulate category-wise gradient mean \\(\\Gamma_k^s\\) for the \\(k\\)-th category as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.749, + 0.892, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {k} ^ {s} = \\frac {\\sum_ {i = 1} ^ {b} \\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.895, + 0.83 + ], + "angle": 0, + "content": "Consequently, we use \\(\\Gamma_k^s\\) to reweight heterogeneous relation distillation, and express the proposed \\(\\mathcal{L}_{\\mathrm{RD}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.837, + 0.892, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R D}} = \\frac {1}{K ^ {o} + K ^ {t}} \\sum_ {k = 1} ^ {K ^ {o} + K ^ {t}} \\frac {\\Gamma_ {k} ^ {s} \\cdot \\mathcal {D} _ {\\mathrm {K L}} \\left(\\tilde {\\mathbf {P}} _ {k} ^ {t} , \\tilde {\\mathbf {Y}} _ {k} ^ {t}\\right)}{\\sum_ {\\eta = 1} ^ {t} \\Gamma_ {\\eta} ^ {s} \\cdot \\mathbb {I} _ {k \\in \\mathcal {Y} ^ {\\eta}}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{\\mathrm{KL}}(\\cdot ||\\cdot)\\) indicates the Kullback-Leibler divergence." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11746" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.089, + 0.89, + 0.103 + ], + "angle": 0, + "content": "Table 1. Results on CIFAR-100 [33]. When \\(\\mathcal{M} = 2000\\), we set \\(T = \\{5, 10, 20, 25, 50\\}\\) for \\(\\mathcal{B} = 0\\%\\), and \\(T = \\{5, 10, 25, 50\\}\\) for \\(\\mathcal{B} = 50\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.104, + 0.892, + 0.274 + ], + "angle": 0, + "content": "
Comparison MethodsBackbone#ParamsB=0%B=50%
510202550Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M83.580.678.577.676.479.3↑5.580.678.675.471.076.4↑2.9
BiC [50] (CVPR'2019)ViT-Base85.10M83.582.578.780.378.480.7↑4.278.275.774.271.975.0↑4.3
PODNet [17] (ECCV'2020)ViT-Base85.10M75.371.866.564.963.068.3↑16.672.873.074.274.973.7↑5.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.682.079.179.177.380.2↑4.675.874.872.972.073.9↑5.4
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M75.272.066.165.263.868.5↑16.472.772.874.074.773.5↑5.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M84.783.579.678.577.980.8↑4.178.477.376.274.076.5↑2.8
AFC [32] (CVPR'2022)ViT-Base85.10M74.675.570.367.464.770.5↑14.470.869.468.969.969.8↑9.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.585.881.380.877.482.2↑2.782.581.077.374.578.8↑0.5
HFC (Ours)ViT-Base85.10M86.486.385.585.081.184.9-82.981.077.875.579.3-
Upper BoundViT-Base85.10M94.294.294.294.294.294.2-94.294.294.294.294.2-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.276, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Table 2. Results on ImageNet-100 [10]. When \\( \\mathcal{M} = {2000} \\) ,we set \\( T = \\{ 5,{10},{20},{25}\\} \\) for \\( \\mathcal{B} = 0\\% \\) ,and \\( T = \\{ 5,{10},{25},{50}\\} \\) for \\( \\mathcal{B} = {50}\\% \\) ." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.292, + 0.892, + 0.464 + ], + "angle": 0, + "content": "
Comparison MethodsBackbone#ParamsB=0%B=50%
5102025Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M84.182.480.278.681.3↑3.279.578.876.673.977.2↑3.4
BiC [50] (CVPR'2019)ViT-Base85.10M81.480.279.278.579.8↑4.773.974.473.472.273.5↑7.1
PODNet [17] (ECCV'2020)ViT-Base85.10M82.480.179.578.380.1↑4.568.471.375.477.673.2↑7.4
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.081.780.979.481.2↑3.375.874.872.972.073.9↑6.7
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M82.381.079.379.080.4↑4.168.371.275.076.972.8↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M85.684.083.883.184.1↑0.481.080.580.079.880.3↑0.3
AFC [32] (CVPR'2022)ViT-Base85.10M83.281.280.579.481.1↑3.577.178.478.978.178.1↑2.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.183.480.080.382.2↑2.377.478.479.779.478.7↑1.9
HFC (Ours)ViT-Base85.10M86.185.484.082.684.5-81.581.180.379.580.6-
Upper BoundViT-Base85.10M86.386.386.386.386.3-86.386.386.386.386.3-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.473, + 0.468, + 0.499 + ], + "angle": 0, + "content": "Table 3. Results on ImageNet-1000 [10]. When \\(\\mathcal{M} = 20000\\), we set \\(T = \\{5,10\\}\\) for \\(\\mathcal{B} = 0\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.501, + 0.466, + 0.624 + ], + "angle": 0, + "content": "
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M75.472.473.9↑3.5
BiC [50] (CVPR'2019)ViT-Base85.10M68.966.067.5↑10.0
PODNet [17] (ECCV'2020)ViT-Base85.10M68.471.369.8↑7.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M75.874.875.3↑2.2
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M68.463.265.8↑11.6
FOSTER [47] (ECCV'2022)ViT-Base85.10M76.974.375.6↑1.9
AFC [32] (CVPR'2022)ViT-Base85.10M74.271.472.8↑4.6
DyTox [18] (CVPR'2022)ViT-Base85.10M78.174.776.4↑1.0
HFC (Ours)ViT-Base85.10M78.576.477.5-
Upper BoundViT-Base85.10M86.386.386.3-
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.636, + 0.283, + 0.652 + ], + "angle": 0, + "content": "3.5. Optimization Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Overall, in the \\(t\\)-th \\((t \\geq 2)\\) learning task, the objective formulation to optimize \\(\\Theta^t\\) is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.704, + 0.47, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o b j}} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {F C}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {R D}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.731, + 0.47, + 0.807 + ], + "angle": 0, + "content": "where \\(\\alpha_{1},\\alpha_{2}\\) are hyper-parameters. \\(\\Theta^t\\) along with learnable \\(\\mathbf{E}_0^t\\) are optimized via \\(\\mathcal{L}_{\\mathrm{CE}}\\) in Eq. (5) for the first task, and trained via \\(\\mathcal{L}_{\\mathrm{obj}}\\) in Eq. (14) when \\(t\\geq 2\\). After optimizing new model \\(\\Theta^t\\), we store \\(\\Theta^t\\) as the frozen old model \\(\\Theta^{t - 1}\\) to perform the GRD loss \\(\\mathcal{L}_{\\mathrm{RD}}\\) for the next incremental task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.21, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.295, + 0.863 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "For fair comparisons, we follow the same protocols utilized in baseline CIL methods [40, 50, 17, 17, 2, 18, 32] to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.658 + ], + "angle": 0, + "content": "set incremental tasks (class order), and do experiments on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Specifically, we consider two task settings in this paper: 1) We divide all categories of each dataset into \\( T \\) tasks equally [40, 18] (i.e., the base class set \\( \\mathcal{B} = 0\\% \\)); 2) We start by training our model on half of the categories, while divide the rest of categories into \\( T \\) tasks equally [17, 32] (i.e., the base class set \\( \\mathcal{B} = 50\\% \\)). When \\( \\mathcal{B} = 0\\% \\), we set \\( T = \\{5, 10, 20, 25, 50\\} \\) for CIFAR-100 [33], \\( T = \\{5, 10, 20, 25\\} \\) for ImageNet-100 [10] and \\( T = \\{5, 10\\} \\) for ImageNet-1000 [10]. When \\( \\mathcal{B} = 50\\% \\), we set \\( T = \\{5, 10, 25, 50\\} \\) for CIFAR-100 [33] and ImageNet-100 [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.902 + ], + "angle": 0, + "content": "As introduced in [40, 50], the size of memory \\(\\mathcal{M}\\) for all comparison methods is fixed as 2,000 for CIFAR-100 [33] and ImageNet-100 [10], and 20,000 for ImageNet-1000 [10]. We follow iCaRL [40] to update memory \\(\\mathcal{M}\\) and use global memory for [18]. Besides, we employ the same data augmentation proposed in DyTox [18] for all comparison methods, but don't use mixup technology. Following [17, 32], we also consider another memory setting (i.e., storing 20 exemplars per class), and set \\(T = \\{5,10\\}, \\mathcal{B} = 50\\%\\) for comparisons on CIFAR-100 [33] (see Tab. 6). For network architecture, we use ViT-Base [15] as feature extractor containing \\(l_{m} = 11\\) MSA blocks and \\(l_{t} = 1\\) TSA block, where the parameters are initialized via [22]. The task classifier includes only one fully-connected layer. The SGD optimizer is employed to optimize our model, where the learning rate is initialized as \\(6.25 \\times 10^{-5}\\). We apply two plug-and-play losses (i.e., \\(\\mathcal{L}_{\\mathrm{FC}}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11747" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 4. Performance on CIFAR-100 [33] (\\(T = \\{5, 10, 20, 25, 50\\}\\)) and ImageNet-100 [10] (\\(T = \\{5, 10, 20, 25\\}\\)) when we apply \\( \\mathbf{Ours}^{\\ddagger} \\) into existing distillation-based CIL methods and set \\( \\mathcal{M} = 2000 \\), \\( \\mathcal{B} = 0\\% \\). \\( \\mathbf{Ours}^{\\ddagger} \\) denotes the proposed plug-and-play losses \\( \\mathcal{L}_{\\mathrm{FC}} \\) and \\( \\mathcal{L}_{\\mathrm{RD}} \\)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.12, + 0.894, + 0.414 + ], + "angle": 0, + "content": "
Comparison MethodsCIFAR-100 Dataset [33]ImageNet-100 Dataset [10]
Backbone#Params510202550Avg.Backbone#Params5102025
iCaRL [40] (CVPR'2017)ResNet-320.46M66.463.953.150.239.154.5ResNet-1811.22M76.469.560.657.9
iCaRL [40] + Ours‡ResNet-320.46M68.065.754.251.440.856.0ResNet-1811.22M76.870.361.759.2
BiC [50] (CVPR'2019)ResNet-320.46M55.350.848.346.439.348.0ResNet-1811.22M61.554.346.744.5
BiC [50] + Ours‡ResNet-320.46M63.253.852.449.341.552.0ResNet-1811.22M63.257.050.348.4
PODNet [17] (ECCV'2020)ResNet-320.46M63.854.548.144.439.750.1ResNet-1811.22M74.966.555.851.7
PODNet [17] + Ours‡ResNet-320.46M65.860.051.648.344.754.1ResNet-1811.22M76.569.359.455.8
SS-IL [2] (ICCV'2021)ResNet-320.46M65.159.049.048.640.752.5ResNet-1811.22M60.953.539.335.8
SS-IL [2] + Ours‡ResNet-320.46M67.764.951.651.242.555.6ResNet-1811.22M75.267.557.956.0
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M63.754.347.844.039.049.8ResNet-1811.22M74.866.355.451.3
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M66.758.550.546.943.353.2ResNet-1811.22M76.468.758.955.1
FOSTER [47] (ECCV'2022)ResNet-320.46M72.068.767.166.062.667.3ResNet-1811.22M74.970.967.265.7
FOSTER [47] + Ours‡ResNet-320.46M72.269.967.466.762.867.8ResNet-1811.22M75.071.367.566.2
AFC [32] (CVPR'2022)ResNet-320.46M65.657.150.948.142.052.7ResNet-1811.22M76.369.761.158.1
AFC [32] + Ours‡ResNet-320.46M66.263.853.249.744.155.4ResNet-1811.22M77.471.263.367.4
Upper BoundResNet-320.46M76.676.676.676.676.676.6ResNet-1811.22M73.273.273.273.2
DyTox [18] (CVPR'2022)ViT-Tiny10.71M75.373.572.772.370.572.9ViT-Tiny10.71M77.875.472.972.3
DyTox [18] + Ours‡ViT-Tiny10.71M75.974.072.972.670.873.2ViT-Tiny10.71M77.975.673.172.9
Upper BoundViT-Tiny10.71M76.176.176.176.176.176.1ViT-Tiny10.71M79.179.179.179.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.424, + 0.47, + 0.465 + ], + "angle": 0, + "content": "Table 5. Ablation studies on CIFAR-100 [33] (the top block) and ImageNet-100 [10] (the bottom block), when we set \\( T = \\{5,10,25\\} \\) under the settings of \\( \\mathcal{M} = 2000, \\mathcal{B} = 0\\% \\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.466, + 0.469, + 0.591 + ], + "angle": 0, + "content": "
Ablation VariantsBackbone#ParamsTSAGFCGRD51025
BaselineViT-Base85.10MXXX81.680.278.1
Baseline + TSAViT-Base85.10M✓XX83.182.481.0
Baseline + TSA + GFCViT-Base85.10M✓✓X84.083.782.1
HFC (Ours)ViT-Base85.10M✓✓✓86.486.385.0
BaselineViT-Base85.10MXXX82.080.578.8
Baseline + TSAViT-Base85.10M✓XX83.782.881.2
Baseline + TSA + GFCViT-Base85.10M✓XX84.383.482.0
HFC (Ours)ViT-Base85.10M✓✓✓86.185.484.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.669 + ], + "angle": 0, + "content": "and \\(\\mathcal{L}_{\\mathrm{RD}})\\) to existing CIL methods [40, 18] using ResNet-18 [23], ResNet-32 [23] or ViT-Tiny [15] as backbone. For evaluation, top-1 accuracy is used to compare performances of our model and baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.312, + 0.696 + ], + "angle": 0, + "content": "4.2. Comparison Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Comparison Results: Tabs. 1-3 present comparison results of our HFC model and off-the-shelf CIL methods on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Our model significantly outperforms existing CIL methods [40, 50, 17, 2, 18, 32, 47] by \\(0.7\\% \\sim 16.6\\%\\) accuracy under various incremental settings, when we use the same backbone (i.e., ViT-Base [15]) as feature extractor for fair comparisons. Such large performance improvement illustrates the superiority of our model to address forgetting heterogeneity among easy-to-forget and hard-to-forget old categories from both representation and gradient aspects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Improvement of Plug-and-Play Losses: Tab. 4 shows large performance improvement of existing knowledge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.895, + 0.52 + ], + "angle": 0, + "content": "based CIL methods [40, 50, 17, 4, 18, 32, 47], when we apply the proposed plug-and-play losses (i.e., \\(\\mathcal{L}_{\\mathrm{FC}}\\) and \\(\\mathcal{L}_{\\mathrm{RD}}\\)) to them. The proposed losses \\(\\mathcal{L}_{\\mathrm{FC}}\\) and \\(\\mathcal{L}_{\\mathrm{RD}}\\) help existing CIL methods to address heterogeneous forgetting from gradient aspect, thus largely improving their performance. Besides, it validates generalization and robustness of our HFC model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.665, + 0.543 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.895, + 0.734 + ], + "angle": 0, + "content": "As shown in Tab. 5, we present ablation studies on CIFAR-100 [33] and ImageNet-100 [10] to investigate the effectiveness of each module. TSA, GFC and GRD are task-semantic aggregation blocks, the gradient-balanced forgetting compensation loss \\(\\mathcal{L}_{\\mathrm{FC}}\\) and gradient-balanced relation distillation loss \\(\\mathcal{L}_{\\mathrm{RD}}\\). Baseline denotes the performance of our model using the traditional classification loss and knowledge distillation loss proposed in DyTox (CVPR'2022) [18], but without the TSA, GFC and GRD modules. Compared with Ours, the performance of all ablation variants degrades significantly. It validates superiority of each module to overcome forgetting heterogeneity of different old categories." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.813, + 0.759 + ], + "angle": 0, + "content": "4.4. Analysis of Task-Wise Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Figs. 3-4 present task-wise performance comparisons between our model and other CIL methods [40, 50, 17, 18, 32] when we set backbone as ViT-Base and different tasks as \\( T = \\{5, 10, 20, 25\\} \\). Our HFC model outperforms baseline methods for most task-wise comparisons, which illustrates the superior performance of our model to surmount heterogeneous forgetting from representation and gradient aspects. To identify new classes continually, the task-semantic aggregation (TSA) block explores task-shared global representa" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11748" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.89, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.242, + 0.894, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Analysis of task-wise performance comparisons on CIFAR-100 [33]. We set \\(\\mathcal{M} = 2000\\), \\(\\mathcal{B} = 0\\%\\) when the backbone is ViT-Base." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.257, + 0.891, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.408, + 0.867, + 0.422 + ], + "angle": 0, + "content": "Figure 4. Analysis of task-wise performance on ImageNet-100 [10]. We set \\(\\mathcal{M} = 2000\\), \\(\\mathcal{B} = 0\\%\\) when the backbone is ViT-Base." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.472 + ], + "angle": 0, + "content": "Table 6. Results on CIFAR-100 [33] when storing 20 exemplars for each class. We set \\( T = \\{5,10\\} \\) for \\( \\mathcal{B} = 50\\% \\), and \\( \\mathbf{Ours}^{\\ddagger} \\) denotes the proposed plug-and-play losses \\( \\mathcal{L}_{\\mathrm{FC}} \\) and \\( \\mathcal{L}_{\\mathrm{RD}} \\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.473, + 0.468, + 0.744 + ], + "angle": 0, + "content": "
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M78.876.277.5↑1.9
BiC [50] (CVPR'2019)ViT-Base85.10M75.374.574.9↑4.5
PODNet [17] (ECCV'2020)ViT-Base85.10M71.671.671.6↑7.8
SS-IL [2] (ICCV'2021)ViT-Base85.10M73.071.872.4↑7.0
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M71.571.771.6↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M78.878.778.8↑0.6
AFC [32] (CVPR'2022)ViT-Base85.10M68.265.366.8↑12.6
DyTox [18] (CVPR'2022)ViT-Base85.10M79.378.478.8↑0.6
HFC (Ours)ViT-Base85.10M79.878.979.4-
Upper BoundViT-Base85.10M94.294.294.2-
iCaRL [40] (CVPR'2017)ResNet-320.46M56.653.254.9↑0.7
iCaRL [40] + Ours‡ResNet-320.46M57.353.955.6-
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M62.261.161.7↑0.6
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M62.961.662.3-
FOSTER [47] (ECCV'2022)ResNet-320.46M67.366.466.8↑0.5
FOSTER [47] + Ours‡ResNet-320.46M67.766.667.2-
AFC [32] (CVPR'2022)ResNet-320.46M63.863.663.7↑0.7
AFC [32] + Ours‡ResNet-320.46M64.464.364.4-
Upper BoundResNet-320.46M76.676.676.6-
DyTox [18] (CVPR'2022)ViT-Tiny10.71M67.966.267.1↑0.5
DyTox [18] + Ours‡ViT-Tiny10.71M68.566.767.6-
Upper BoundViT-Tiny10.71M76.176.176.1-
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.759, + 0.471, + 0.79 + ], + "angle": 0, + "content": "tions to alleviate forgetting from representation aspect, while \\(\\mathcal{L}_{\\mathrm{FC}}\\) and \\(\\mathcal{L}_{\\mathrm{RD}}\\) can achieve gradient-balanced compensation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.395, + 0.816 + ], + "angle": 0, + "content": "4.5. Analysis of Forgetting Heterogeneity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.823, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Tab. 7 presents forgetting heterogeneity (FH) of different old classes via measuring their variance of gradient updating. We formulate forgetting heterogeneity (FH) as \\(\\mathrm{FH} = \\frac{1}{T}\\sum_{t=1}^{T}\\left(\\frac{1}{S_t}\\sum_{i=1}^{S_t}(|\\Gamma_i^t| - \\sum_{\\eta=1}^t\\Gamma_\\eta \\cdot \\mathbb{I}_{\\mathbf{y}_i^t \\in \\mathcal{Y}^\\eta})^2\\right)\\), where \\(\\Gamma_i^t\\) and \\(\\Gamma_\\eta\\) are obtained from Eqs. (6)-(7), and \\(S_t\\) is" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.431, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Table 7. Analysis of forgetting heterogeneity (FH) on CIFAR-100 [33] when \\( T = 5, \\mathcal{B} = 0\\% \\). AC denotes the averaged accuracy." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.459, + 0.882, + 0.584 + ], + "angle": 0, + "content": "
M=2000,B=0%BackboneTSAGFCGRDFHAC
iCaRL [40] (CVPR'2017)ViT-Base---131.583.5
FOSTER [47] (ECCV'2022)ViT-Base---119.485.1
AFC [32] (CVPR'2022)ViT-Base---108.574.6
DyTox [18] (CVPR'2022)ViT-Base---105.185.5
BaselineViT-BaseXXX128.981.6
Baseline + TSAViT-Base✓XX125.483.1
Baseline + TSA + GFCViT-Base✓✓X120.584.0
HFC (Ours)ViT-Base✓✓✓103.686.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.894, + 0.66 + ], + "angle": 0, + "content": "number of samples in test set. Tab. 7 verifies superiority of our model against other CIL methods to tackle heterogeneous forgetting. It also shows effectiveness of all modules to collaboratively minimize forgetting heterogeneity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.619, + 0.693 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we develop a novel Heterogeneous Forgetting Compensation (HFC) model to surmount heterogeneous forgetting from representation and gradient aspects. To be specific, a task-semantic aggregation block is designed to tackle heterogeneous forgetting from representation aspect via exploring task-shared global representations. Moreover, we propose a gradient-balanced forgetting compensation loss and a relation distillation loss to compensate forgetting heterogeneity from gradient aspect via performing balanced gradient propagation and distilling heterogeneous class relations. Experiments verify the superiority of our proposed HFC model against baseline methods. We will further consider addressing noisy forgetting heterogeneity in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11749" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, pages 3930-3939, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, pages 844-853, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 144-161, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[4] Arjun Ashok, K. J. Joseph, and Vineeth N. Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In ECCV, pages 105-122, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.472, + 0.4 + ], + "angle": 0, + "content": "[5] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.472, + 0.471 + ], + "angle": 0, + "content": "[6] Francisco M. Castro, Manuel J. Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 241-257, Cham, 2018. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[7] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.472, + 0.585 + ], + "angle": 0, + "content": "[8] Huitong Chen, Yu Wang, and Qinghua Hu. Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering, pages 1-15, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.472, + 0.642 + ], + "angle": 0, + "content": "[9] Stéphane d'Ascoli, Hugo Touvron, Matthew Leavitt, Ari Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.472, + 0.686 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[11] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11-20, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In CVPR, pages 10164-10173, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[13] Jiahua Dong, Duzhen Zhang, Yang Cong, Wei Cong, Henghui Ding, and Dengxin Dai. Federated incremental semantic segmentation. In CVPR, pages 3934-3943, June 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[14] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In CVPR, pages 19217-19227, June 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[18] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285-9295, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[19] Zhen Fang, Yixuan Li, Jie Lu, Jiahua Dong, Bo Han, and Feng Liu. Is out-of-distribution detection learnable? In NeurIPS, volume 35, pages 37199-37213, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[20] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. In ICML, page 1243-1252, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, volume 27, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.894, + 0.518 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.894, + 0.603 + ], + "angle": 0, + "content": "[24] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. In NeurIPS Workshop, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.894, + 0.646 + ], + "angle": 0, + "content": "[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[26] Weiming Hu, Xi Li, Guodong Tian, Stephen Maybank, and Zhongfei Zhang. An incremental dpmm-based method for trajectory clustering, modeling, and retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(5):1051-1065, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[27] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[28] Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unforgetting continual learning. In Advances in Neural Information Processing Systems, pages 13647–13657, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[29] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 699-715, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11750" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, and et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.472, + 0.192 + ], + "angle": 0, + "content": "[31] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwar, and Vineeth Balasubramanian. Energy-based latent aligner for incremental learning. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[32] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, pages 16071-16080, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.251, + 0.471, + 0.279 + ], + "angle": 0, + "content": "[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, pages 32-33, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.281, + 0.471, + 0.322 + ], + "angle": 0, + "content": "[34] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.325, + 0.471, + 0.365 + ], + "angle": 0, + "content": "[35] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.368, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[36] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. AAAI, 37(7):8906-8913, Jun. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.412, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[37] Yaoyao Liu, Bernt Schiele, and Qianru Sun. RMM: reinforced memory management for class-incremental learning. In NeurIPS, pages 3478-3490, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.456, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[38] John M. Pierre. Incremental lifelong deep learning for autonomous vehicles. In ITSC, pages 3949-3954, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.486, + 0.471, + 0.526 + ], + "angle": 0, + "content": "[39] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for continual learning. In NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.529, + 0.471, + 0.57 + ], + "angle": 0, + "content": "[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.572, + 0.471, + 0.626 + ], + "angle": 0, + "content": "[41] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, volume 80, pages 4548-4557, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.471, + 0.67 + ], + "angle": 0, + "content": "[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.674, + 0.471, + 0.728 + ], + "angle": 0, + "content": "[43] Zhenheng Tang, Yonggang Zhang, Shaohuai Shi, Xin He, Bo Han, and Xiaowen Chu. Virtual homogeneity learning: Defending against data heterogeneity in federated learning. In ICML, pages 21111-21132, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[44] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 254-270, Cham, 2020. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.802, + 0.471, + 0.856 + ], + "angle": 0, + "content": "[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, pages 99-108, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[47] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[48] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In AAAI, volume 35, pages 10165-10173, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[49] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.894, + 0.318 + ], + "angle": 0, + "content": "[51] Ziyang Wu, Christina Baek, Chong You, and Yi Ma. Incremental learning via rate reduction. In CVPR, pages 1125-1133, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[52] Rongtao Xu, Changwei Wang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Wave-like class activation map with representation fusion for weakly-supervised semantic segmentation. IEEE Transactions on Multimedia, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[53] Rongtao Xu, Changwei Wang, Jiguang Zhang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. RSSformer: Foreground saliency enhancement for remote sensing land-cover segmentation. IEEE Transactions on Image Processing, 32:1052-1064, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.894, + 0.489 + ], + "angle": 0, + "content": "[54] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[55] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.894, + 0.573 + ], + "angle": 0, + "content": "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, page 3987-3995, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[57] Duzhen Zhang, Feilong Chen, and Xiuyi Chen. Dualgats: Dual graph attention networks for emotion recognition in conversations. In ACL, pages 7395-7408, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.894, + 0.674 + ], + "angle": 0, + "content": "[58] Yonggang Zhang, Mingming Gong, Tongliang Liu, Gang Niu, Xinmei Tian, Bo Han, Bernhard Schölkopf, and Kun Zhang. Causaladv: Adversarial robustness through the lens of causality. *ICLR*, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[59] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13205-13214, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[60] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[61] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In ECCV, pages 55-72, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11751" + } + ] +] \ No newline at end of file diff --git a/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_origin.pdf b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c72f443be37f6cc900dd3503d3b065a692d0ad4 --- /dev/null +++ b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/fd7dde4b-f839-40f9-bfea-101cb8081cd9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1482644a72009b6d7ab9a7ea511b02f2f5e5b3a495180047ad2de5ce567a4d3e +size 851335 diff --git a/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/full.md b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..805abf5bf58034de4ba4c2a110c87fc5cdd1885b --- /dev/null +++ b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/full.md @@ -0,0 +1,331 @@ +# Heterogeneous Forgetting Compensation for Class-Incremental Learning + +Jiahua Dong $^{1,2,3*}$ , Wenqi Liang $^{1,2,3*}$ , Yang Cong $^{4\dagger}$ , Gan Sun $^{1,2}$ + +$^{1}$ State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, 110016, China.‡ + +$^{2}$ Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang, 110169, China. + +3University of Chinese Academy of Sciences, Beijing, 100049, China. + +$^{4}$ South China University of Technology, Guangzhou, 510640, China. + +{dongjiahua1995, liangwenqi0123, congyang81, sunganl412}@gmail.com + +# Abstract + +Class-incremental learning (CIL) has achieved remarkable successes in learning new classes consecutively while overcoming catastrophic forgetting on old categories. However, most existing CIL methods unreasonably assume that all old categories have the same forgetting pace, and neglect negative influence of forgetting heterogeneity among different old classes on forgetting compensation. To surmount the above challenges, we develop a novel Heterogeneous Forgetting Compensation (HFC) model, which can resolve heterogeneous forgetting of easy-to-forget and hard-to-remember old categories from both representation and gradient aspects. Specifically, we design a task-semantic aggregation block to alleviate heterogeneous forgetting from representation aspect. It aggregates local category information within each task to learn task-shared global representations. Moreover, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation loss and a gradient-balanced relation distillation loss to alleviate forgetting from gradient aspect. They consider gradient-balanced compensation to rectify forgetting heterogeneity of old categories and heterogeneous relation consistency. Experiments on several representative datasets illustrate effectiveness of our HFC model. The code is available at https://github.com/JiahuaDong/HFC. + +# 1. Introduction + +Class-incremental learning (CIL) [40, 12, 5] has attracted appealing attentions recently by accumulating previous learned experience to learn new classes incrementally. It + +![](images/4f3ad93781789f1aaea7c553553eddf6928008ad3c55f0926f3a1daa3149e0ff.jpg) +Figure 1. Forgetting heterogeneity versus accuracy on CIFAR-100 [33] when the backbone is ViT-Base and number of tasks is 5. + +plays an indispensable role in developing a large number of intelligent learning systems, such as autonomous driving [38] and automated surveillance [26]. When learning new classes continuously under the settings of limited memory to replay all previous data of old classes [45], these CIL systems heavily suffer from forgetting on old classes caused by severe class imbalance between old and new categories [40, 31, 14]. To surmount catastrophic forgetting, a growing amount of CIL methods mainly perform knowledge distillation [24] to preserve past experience [34, 12, 35]; introduce a rehearsal strategy to replay part of old data [40, 17]; or dynamically expand network architectures [28, 41, 54]. + +However, most existing CIL methods [40, 45, 54, 18] neglect heterogeneous forgetting speeds of different old classes. They unrealistically assume that all old classes suffer from the same degree of catastrophic forgetting, and compensate forgetting for each old classes equally and independently. Such impracticable assumption enforces existing CIL models [40, 56, 61] to suffer from imbalanced gradient optimization among different old classes, thus favoring more forgetting compensation for hard-to-forget old classes while ne + +glecting those easy-to-forget old classes (i.e., heterogeneous forgetting). More importantly, such forgetting heterogeneity can significantly worsen the forgetting on hard-to-forget old categories, when new streaming classes becomes part of old categories continually. For example, some old classes with various modalities and appearances (e.g., person and car) in autonomous driving [38] are more difficult to explore task-shared representations across different incremental tasks, when compared with other hard-to-forget old classes (e.g., road and traffic sign) with easily-distinguished visual properties. This phenomenon causes imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus exacerbating heterogeneous forgetting speeds among different old classes. When autonomous vehicles [38] learn new classes continually, such forgetting heterogeneity aggravates forgetting on those hard-to-forget old classes (e.g., traffic sign) to some extent. + +Inspired by the above practical scenario, we investigate that different old classes have significant forgetting heterogeneity in this paper, as shown in Fig. 1. This heterogeneous forgetting might heavily weaken forgetting compensation on hard-to-forget old classes, as new classes become a subset of old classes consecutively. In summary, the challenges to tackle heterogeneous forgetting lie in two major aspects: + +- Representation Aspect: Some easy-to-forget old classes (e.g., car and bus) with diverse appearances are more difficult for existing CIL methods [45, 54, 12, 3] to learn intrinsic task-shared representations, and thus are significantly easier to be forgotten than hard-to-forget old classes (e.g., road) with distinctive attributes. Thus, exploring task-shared representations is essential to address heterogeneous forgetting among different old classes. +- Gradient Aspect: To learn complex visual characterizations for easy-to-forget old classes with various modalities and appearances, existing CIL methods [40, 4, 49] are required to allocate more network architectures for gradient updating. It can result in imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus aggravating forgetting heterogeneity of old classes. + +To overcome the above challenges, we propose a novel Heterogeneous Forgetting Compensation (HFC) model, which is an earlier exploration to address heterogeneous forgetting from both representation and gradient perspectives in the CIL field [40, 4]. Specifically, we propose a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation perspective. It can explore task-shared global representations for each class via aggregating long-range local category information within each task. Meanwhile, to tackle heterogeneous forgetting from gradient perspective, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss. The GFC loss can rectify heterogeneous forgetting speeds + +of easy-to-forget and hard-to-forget old classes, while normalize different learning paces of new categories to achieve gradient-balanced propagation. Besides, the GRD loss can distill heterogeneous relation consistency brought by forgetting heterogeneity among different old classes. Experiments show our model has large improvements on several representative datasets, compared with baseline methods [40, 36]. More importantly, we apply two plug-and-play losses into existing distillation-based CIL methods [40, 17, 4] to significantly improve their performance from gradient aspect. The novel contributions of this paper are presented as follows: + +- We develop a novel Heterogeneous Forgetting Compensation (HFC) model to address different forgetting speeds of easy-to-forget and hard-to-forget old classes. To our best knowledge, this paper is the first exploration to tackle forgetting heterogeneity among old categories from representation and gradient aspects in the CIL field. +- We design a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect. It can explore robust task-shared representations for each class via aggregating local category information. +- We propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss to surmount forgetting heterogeneity from gradient aspect. They can balance different forgetting of old classes and heterogeneous category-relation consistency to improve performance when applying them into existing methods. + +# 2. Related Work + +We discuss some class-incremental learning (CIL) methods [60, 37, 18], and divide them into three categories: + +Knowledge Distillation: After Li et al. [34] apply knowledge distillation [24] to continual learning, [40, 39, 59] follow them to transfer past experience from old model to new model. iCaRL [40] and EEIL [6] perform knowledge distillation in the output space. [44] aims to preserve consistent topology feature space for old and new tasks. [16] proposes feature-level knowledge distillation when applying continual learning [37] into semantic segmentation task [53, 52, 11]. + +Rehearsal Strategy: For replaying past experience, lots of CIL methods [49, 8, 42] allocate a memory to store exemplars of old classes or synthesize samples via generative adversarial model [21, 58]. For saving large memory overhead, [29] proposes to store low-dimension features rather than raw samples. [51] proposes a "white box" framework deviated from rate reduction to preserve past experience. + +Dynamic Architecture: Many works [28, 54, 1] allocate dynamical networks for new classes as the growing of learned classes. Yoon et al. [55] dynamically expand architecture capacity via selective retraining. [54, 18] address the problem of relying on task index when performing dynamic + +architectures. However, the above-mentioned CIL methods [40, 27] neglect negative influence of heterogeneous forgetting of old classes on forgetting compensation. + +# 3. The Proposed Model + +# 3.1. Problem Definition and Overview + +Problem Definition: Following class-incremental learning (CIL) methods [40, 50, 17, 18, 54], we set a series of consecutive learning tasks as $\mathcal{T} = \{\mathcal{T}^t\}_{t=1}^T$ , where $T$ indicates the total task quantity. The $t$ -th task $\mathcal{T}^t = \{\mathbf{x}_i^t, \mathbf{y}_i^t\}_{i=1}^{N^t}$ is composed of $N^t$ pairs of image $\mathbf{x}_i^t$ and one-hot groundtruth $\mathbf{y}_i^t \in \mathcal{Y}^t$ , and $\mathcal{Y}^t$ denotes label space of the $t$ -th incremental task containing $K^t$ new categories. The label spaces between any two incremental tasks have no overlap: $\mathcal{Y}^t \cap (\cup_{j=1}^{t-1} \mathcal{Y}^j) = \emptyset$ . Namely, $K^t$ new classes belonging to the $t$ -th task $\mathcal{T}^t$ are different from $K^o = \sum_{i=1}^{t-1} K^i \subset \cup_{j=1}^{t-1} \mathcal{Y}^j$ old classes learned from previous tasks $\{\mathcal{T}^i\}_{i=1}^{t-1}$ . As introduced in CIL baselines [40, 50, 54], we set a fixed exemplar memory $\mathcal{M}$ , and store only few images (i.e., $\frac{|M|}{K^o}$ ) for each old class in the $t$ -th task $\mathcal{T}^t$ , where $\mathcal{M}$ satisfies $\frac{N^t}{K^t} \gg \frac{|M|}{K^o}$ . In the $t$ -th task, we aim to identify both $K^t$ new classes and $K^o$ old classes via optimizing the model on $\mathcal{T}^t$ and $\mathcal{M}$ . + +Overview: The overview of our HFC model to surmount forgetting heterogeneity of different old categories from both representation and gradient aspects is shown in Fig. 2. Denote the proposed HFC model learned at the $(t - 1)$ -th and $t$ -th tasks as old and new models (i.e., $\Theta^{t - 1}$ and $\Theta^t$ ), where $\Theta^t$ is inherited from $\Theta^{t - 1}$ and only expands the number of output neurons in the classifier to identify $K^t$ new classes. In the $t$ -th incremental task, given an image $\mathbf{x}_i^t\in \mathcal{T}^t\cup \mathcal{M}$ , we forward it into $\Theta^{t - 1}$ and $\Theta^t$ to extract task-shared global representation via task-semantic aggregation (TSA) block, which alleviates heterogeneous forgetting from representation aspect (Section 3.2). After obtaining probabilities $\mathbf{P}^{t - 1}(\mathbf{x}_i^t,\Theta^{t - 1})\in \mathbb{R}^{K^o}$ and $\mathbf{P}^t (\mathbf{x}_i^t,\Theta^t)\in \mathbb{R}^{K^o +K^t}$ predicted via $\Theta^{t - 1}$ and $\Theta^t$ , we develop two plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss $\mathcal{L}_{\mathrm{FC}}$ and a gradient-balanced relation distillation (GRD) loss $\mathcal{L}_{\mathrm{RD}}$ to surmount heterogeneous forgetting from semantic aspect (Sections 3.3 and 3.4). $\mathcal{L}_{\mathrm{FC}}$ and $\mathcal{L}_{\mathrm{RD}}$ can normalize heterogeneous forgetting of old classes and distill heterogeneous relation consistency between $\Theta^{t - 1}$ and $\Theta^t$ . + +# 3.2. Task-Semantic Aggregation Block + +To explore long-range semantic dependencies across different tasks, Douillard et al. [18] introduce Vision Transformer (ViT) [15, 46, 22] into class-incremental learning (CIL). However, it significantly increases the memory overhead to store task tokens and task-specific decoders, when learning large-scale new classes. Besides, it cannot aggregate local category information to explore task-shared global + +representations, resulting in heterogeneous forgetting from representation aspect. To tackle these issues, we develop a task-semantic aggregation (TSA) block to extract task-shared global representations, which is effective to tackle heterogeneous forgetting from representation aspect. More importantly, the TSA block cannot dynamically increase memory cost with the arrival of large-scale new categories. + +Specifically, as presented in Fig. 2, the proposed model $\Theta^t = \Theta_c^t\circ \Theta_f^t$ at the $t$ -th task is composed of a feature extractor $\Theta_f^t$ to extract task-shared global representations, and a task classifier $\Theta_c^t$ to classify $K^o$ old classes and $K^t$ new classes. In the $t$ -th task, the network parameters of $\Theta_f^{t - 1}$ and $\Theta_c^{t - 1}$ in $\Theta^{t - 1}$ learned at the $(t - 1)$ -th task are frozen to preserve semantic knowledge for old classes. Different from [18] that uses vanilla ViT [15, 30, 57, 7] to extract features, we employ $l_{m}$ multi-head self-attention (MSA) blocks [20] and $l_{t}$ task-semantic aggregation (TSA) block as feature extractor $\Theta_f^t$ to obtain task-shared global representations. + +- Multi-Head Self-Attention (MSA): Given an image $\mathbf{x}_i^t\in \mathcal{T}^t\cup \mathcal{M}$ in the $t$ th task, following ViT [15, 9], we crop it into $N$ patches with equal dimension, flatten these patches and map them into a $D$ -dimension feature space via a linear projection to obtain patch embedding $\mathbf{Z}_e\in \mathbb{R}^{N\times D}$ . Then we concatenate $\mathbf{Z}_e$ with a class token $\mathbf{Z}_{\mathrm{cls}}\in \mathbb{R}^{D}$ , and perform element-wise sum operation with a position token $\mathbf{Z}_p\in \mathbb{R}^{(N + 1)\times D}$ to get $\mathbf{Z}_0 = [\mathbf{Z}_e\oplus \mathbf{Z}_{\mathrm{cls}}] + \mathbf{Z}_p\in \mathbb{R}^{(N + 1)\times D}$ where $\oplus$ is concatenation function. As shown in Fig.2, $\mathbf{Z}_0$ is fed into $\Theta_f^t$ including $l_{m}$ MSA blocks to extract long-range semantics across different tasks. For the $l$ -th $(l = 1,\dots ,l_m)$ MSA block, we regard $\mathbf{Z}_{l - 1}\in \mathbb{R}^{(N + 1)\times D}$ learned from the $(l - 1)$ -th MSA block as input, and execute parallel selfattention $H$ times. Thus, the self-attention $\mathbf{A}_l^h\in \mathbb{R}^{(N + 1)\times d}$ of $\mathbf{Z}_{l - 1}$ for the $h$ -th $(h = 1,\dots ,H)$ head is: + +$$ +\mathbf {A} _ {l} ^ {h} = \sigma \left(\frac {\mathbf {Z} _ {l - 1} \mathbf {W} _ {q} ^ {h} \left(\mathbf {Z} _ {l - 1} \mathbf {W} _ {k} ^ {h}\right) ^ {\top}}{\sqrt {d}}\right) \left(\mathbf {Z} _ {l - 1} \mathbf {W} _ {v} ^ {h}\right), \tag {1} +$$ + +where $\mathbf{W}_q^h,\mathbf{W}_k^h,\mathbf{W}_v^h\in \mathbb{R}^{D\times d}$ are query, key and value mapping matrices. $d = \frac{D}{H}$ represents channel dimension of each head and $\sigma$ is the softmax function. Then we concatenate $\{\mathbf{A}_l^h\}_{h = 1}^H$ along channel dimension and project this concatenated result via a mapping matrix $\mathbf{W}_o\in \mathbb{R}^{D\times D}$ to perform MSA: $\mathrm{MSA}(\mathbf{Z}_{l - 1}) = \mathbf{Z}_{l - 1} + [\mathbf{A}_l^1\oplus$ $\mathbf{A}_l^2,\dots ,\oplus \mathbf{A}_l^H ]\mathbf{W}_o$ where $\oplus$ is concatenation function. After obtaining $\mathrm{MSA}(\mathbf{Z}_{l - 1})\in \mathbb{R}^{(N + 1)\times D}$ $(D = dH)$ , we utilize a multi-layer perceptron (MLP) block to obtain output $\mathbf{Z}_l\in \mathbb{R}^{(N + 1)\times D}$ of the $l$ th MSA block: + +$$ +\mathbf {Z} _ {l} = \operatorname {M S A} \left(\mathbf {Z} _ {l - 1}\right) + \operatorname {M L P} \left(\operatorname {M S A} \left(\mathbf {Z} _ {l - 1}\right)\right). \tag {2} +$$ + +- Task-Semantic Aggregation (TSA): The patch embedding $\mathbf{Z}_{l_m} \in \mathbb{R}^{(N + 1) \times D}$ encoded via the $l_m$ -th MSA block contains rich long-range semantic dependencies shared + +![](images/16353ab5f0807c1c32c2d176bfa051ceb2f1a6378bded8b6cd75f886f8c61555.jpg) +Figure 2. Illustration of the HFC model. It mainly contains a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect via exploring task-shared global representations, and two novel plug-and-play losses (i.e., a gradient-balanced forgetting compensation loss $\mathcal{L}_{\mathrm{FC}}$ and a gradient-balanced relation distillation loss $\mathcal{L}_{\mathrm{RD}}$ ) to overcome forgetting heterogeneity from gradient aspect by compensating gradient-imbalanced propagation and heterogeneous class-relation consistency. + +across different tasks. $\mathbf{Z}_{l_m}$ is then forwarded into $l_{t}$ task-semantic aggregation (TSA) blocks to extract task-shared global embedding. The TSA block aggregates local category context from long-range semantic dependencies to global representations shared across all incremental tasks, which is essential to alleviate heterogeneous forgetting from representation aspect. In the $t$ -th task, as depicted in Fig. 2, we introduce a learnable task-shared embedding $\mathbf{E}_0^t \in \mathbb{R}^D$ that is initialized via $\mathbf{E}_0^{t - 1}$ learned in the $(t - 1)$ -th task. Then we forward $\mathbf{E}_0^t$ along with $\mathbf{Z}_{l_m}$ to $l_{t}$ TSA blocks. For the $l$ -th $(l = 1,\dots ,l_t)$ TSA block, $\mathbf{E}_{l - 1}^t \in \mathbb{R}^D$ learned at $(l - 1)$ -th TSA block and $\mathbf{Z}_{l_m}$ are used to perform task-semantic attention $H$ times in parallel. Thus, we formulate task-semantic attention $\mathbf{S}_l^h \in \mathbb{R}^d$ for the $h$ -th $(h = 1,\dots ,H)$ head as: + +$$ +\mathbf {S} _ {l} ^ {h} = \sigma \left(\frac {\mathbf {E} _ {l - 1} ^ {t} \mathbf {V} _ {q} ^ {h} \left(\mathbf {Z} _ {l _ {m}} \mathbf {V} _ {k} ^ {h}\right) ^ {\top}}{\sqrt {d}}\right) \left(\mathbf {Z} _ {l _ {m}} \mathbf {V} _ {v} ^ {h}\right), \tag {3} +$$ + +where $\mathbf{V}_q^h,\mathbf{V}_k^h,\mathbf{V}_v^h\in \mathbb{R}^{D\times d}$ are projection matrices of query, key and value. After utilizing Eq. (3) to obtain $\{\mathbf{S}_l^h\}_{h = 1}^H$ , we concatenate them along channel dimension and project this fused result via $\mathbf{V}_o\in \mathbb{R}^{D\times D}$ to execute $H$ -head task-semantic attention: $\mathrm{TSA}(\mathbf{E}_{l - 1}^{t}) = [\mathbf{S}_{l}^{1}\oplus$ $\mathbf{S}_l^2,\dots ,\oplus \mathbf{S}_l^H ]\mathbf{V}_o\in \mathbb{R}^D$ . TSA $(\mathbf{E}_{l - 1}^{t})$ is encoded via a MLP block to get output $\mathbf{E}_l^t\in \mathbb{R}^D$ of the $l$ -th TSA block: + +$$ +\mathbf {E} _ {l} ^ {t} = \operatorname {T S A} \left(\mathbf {E} _ {l - 1} ^ {t}\right) + \operatorname {M L P} \left(\operatorname {T S A} \left(\mathbf {E} _ {l - 1} ^ {t}\right)\right). \tag {4} +$$ + +As aforementioned, we obtain the task-shared global representation $\mathbf{E}_{l_t}^t\in \mathbb{R}^D$ via the $l_{t}$ -th TSA block in feature extractor $\Theta_f^t$ . Obviously, the proposed TSA blocks can aggregate local category context of the $t$ -th task from longrange semantic dependencies to task-shared global representation $\mathbf{E}_{l_t}^t$ . It is essential to alleviate heterogeneous forgetting among different old classes from representation aspect, by + +exploring task-shared semantic context across different tasks. Note that we employ layer norm before TSA, MSA and MLP blocks, while we omit it for simplicity in this paper. + +- Task Classifier: The task-shared global representation $\mathbf{E}_{l_t}^t$ obtained via $\Theta_f^t$ is fed into task classifier $\Theta_c^t$ to predict $K^o$ old classes and $K^t$ new classes. Given a mini-batch $\{\mathbf{x}_i^t,\mathbf{y}_i^t\}_{i = 1}^b\in \mathcal{T}^t\cup \mathcal{M}$ , the classification loss $\mathcal{L}_{\mathrm{CE}}$ is: + +$$ +\mathcal {L} _ {\mathrm {C E}} = \frac {1}{b} \sum_ {i = 1} ^ {b} \mathcal {D} _ {\mathrm {C E}} \left(\mathbf {P} ^ {t} \left(\mathbf {x} _ {i} ^ {t}, \Theta^ {t}\right), \mathbf {y} _ {i} ^ {t}\right), \tag {5} +$$ + +where $\mathbf{P}^t (\mathbf{x}_i^t,\Theta^t)\in \mathbb{R}^{K^o +K^t}$ is softmax probability predicted by $\Theta^t$ in the $t$ -th learning task. $\mathcal{D}_{\mathrm{CE}}(\cdot ,\cdot)$ indicates traditional cross-entropy loss, and $b$ is the batch size. + +However, the severe class imbalance (i.e., $\frac{N^t}{K^t} \gg \frac{|\mathcal{M}|}{K^o}$ ) among old and new categories in the $t$ -th task enforces the prediction $\mathbf{P}^t(\mathbf{x}_i^t, \Theta^t)$ in Eq. (5) to suffer from catastrophic forgetting [40] on old categories. Moreover, the gradient optimization of Eq. (5) neglects heterogeneous forgetting speeds of easy-to-forget old classes with various appearances and hard-to-forget old classes with easily-distinguished attributes. To address heterogeneous forgetting from gradient aspect, we propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss $\mathcal{L}_{\mathrm{FC}}$ in Section 3.3 and a gradient-balanced relation distillation (GRD) loss $\mathcal{L}_{\mathrm{RD}}$ in Section 3.4. + +# 3.3. Gradient-Balanced Forgetting Compensation + +As incremental tasks arrive consecutively, easy-to-forget and hard-to-forget old categories change dynamically, thus aggravating difficulty to tackle heterogeneous forgetting from gradient aspect. In light of this, we develop a gradient-balanced forgetting compensation (GFC) loss $\mathcal{L}_{\mathrm{FC}}$ . It adaptively balances large forgetting heterogeneity of old classes + +via performing balanced gradient propagation. Specifically, we perform task-adaptive gradient normalization for different classes, and then reweight classification loss $\mathcal{L}_{\mathrm{CE}}$ in Eq. (5) to compensate imbalanced gradients. For the $t$ -th incremental task, as claimed in [48, 13], the gradient $\Gamma_i^t$ of an given image $(\mathbf{x}_i^t,\mathbf{y}_i^t)\subset \mathcal{T}^t\cup \mathcal{M}$ with respect to the $k$ -th $(k = \arg \max \mathbf{y}_i^t)$ neuron $\mathcal{N}_k^t$ of classifier $\Theta_c^t$ in $\Theta^t$ is: + +$$ +\Gamma_ {i} ^ {t} = \frac {\partial \mathcal {D} _ {\mathrm {C E}} (\mathbf {P} ^ {t} (\mathbf {x} _ {i} ^ {t} , \boldsymbol {\Theta} ^ {t}) , \mathbf {y} _ {i} ^ {t})}{\partial \mathcal {N} _ {k} ^ {t}} = \mathbf {P} ^ {t} (\mathbf {x} _ {i} ^ {t}, \boldsymbol {\Theta} ^ {t}) _ {k} - 1, \quad (6) +$$ + +where $\mathbf{P}^t (\mathbf{x}_i^t,\Theta^t)_k$ denotes probability of the $k$ -th class. + +To compensate heterogeneous forgetting of easy-to-forget and hard-to-forget old categories, we compute task-adaptive gradient means for different tasks. Consequently, given a mini-batch $\{\mathbf{x}_i^t,\mathbf{y}_i^t\}_{i = 1}^b$ in the $t$ -th learning task, the task-adaptive gradient mean $\Gamma_{\eta}$ for the categories learned in the $\eta$ -th $(1\leq \eta \leq t)$ incremental task is formulated as: + +$$ +\Gamma_ {\eta} = \frac {1}{\sum_ {i = 1} ^ {b} \mathbb {I} _ {\mathbf {y} _ {i} ^ {t} \in \mathcal {Y} ^ {\eta}}} \sum_ {i = 1} ^ {b} | \Gamma_ {i} ^ {t} | \cdot \mathbb {I} _ {\mathbf {y} _ {i} ^ {t} \in \mathcal {Y} ^ {\eta}}. \tag {7} +$$ + +The forgetting heterogeneity of old categories learned from $t - 1$ old tasks and learning speeds of new categories in the $t$ -th task can be effectively measured via $\{\Gamma_{\eta}\}_{\eta = 1}^{t}$ in Eq. (7). + +When easy-to-forget and hard-to-remember old classes are varied dynamically as incremental tasks, some noisy predictions on easy-to-remember old classes may result in large deviation to measure their forgetting heterogeneity [43, 19]. To tackle this issue, $\{\Gamma_{\eta}\}_{\eta = 1}^{t}$ in Eq. (7) are expected to be sharper and more distinguishable adaptively [13], as learning tasks arrive continually. Thus, we rewrite Eq. (7) as follows: + +$$ +\Gamma_ {\eta} ^ {s} = \frac {\sum_ {i = 1} ^ {b} \log \left(\left| \Gamma_ {i} ^ {t} \right| ^ {\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\right) \cdot \mathbb {I} _ {\mathbf {y} _ {i} ^ {t} \in \mathcal {Y} ^ {\eta}}}{\sum_ {i = 1} ^ {b} \mathbb {I} _ {\mathbf {y} _ {i} ^ {t} \in \mathcal {Y} ^ {\eta}}}. \tag {8} +$$ + +The sharper task-adaptive gradient means $\{\Gamma_{\eta}^{s}\}_{\eta = 1}^{t}$ computed via Eq. (8) are employed to reweight classification loss $\mathcal{L}_{\mathrm{CE}}$ in Eq. (5). Then the gradient-balanced forgetting compensation (GFC) loss $\mathcal{L}_{\mathrm{FC}}$ is written as follows: + +$$ +\mathcal {L} _ {\mathrm {F C}} = \frac {1}{b} \sum_ {i = 1} ^ {b} \frac {\log \left(\left| \Gamma_ {i} ^ {t} \right| ^ {\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\right)}{\sum_ {\eta = 1} ^ {t} \Gamma_ {\eta} ^ {s} \cdot \mathbb {I} _ {\mathbf {y} _ {i} ^ {t} \in \mathcal {Y} ^ {\eta}}} \mathcal {D} _ {\mathrm {C E}} \left(\mathbf {P} ^ {t} \left(\mathbf {x} _ {i} ^ {t}, \Theta^ {t}\right), \mathbf {y} _ {i} ^ {t}\right). \tag {9} +$$ + +Obviously, $\mathcal{L}_{\mathrm{FC}}$ in Eq. (9) encourages the optimization of $\Theta^t$ to perform gradient-balanced propagation for easy-to- forget, hard-to- forget old classes and new classes learned in different tasks. It tackles heterogeneous forgetting on old classes from gradient aspect, via normalizing forgetting paces of different old classes adaptively with $\{\Gamma_{\eta}^{s}\}_{\eta = 1}^{t}$ . + +# 3.4. Gradient-Balanced Relation Distillation + +The relationships among old and new categories remain constant in semantic space, regardless of availability of old-class samples. In light of this, distilling inter-class relations + +from old model $\Theta^{t-1}$ to new model $\Theta^t$ can address forgetting on old classes [6, 25]. However, most knowledge distillation methods used in CIL [40, 2, 50] only utilize prediction of an individual sample to perform semantic consistency of old classes between $\Theta^{t-1}$ and $\Theta^t$ . They neglect underlying relations among old and new categories to tackle forgetting, and heavily suffer from noisy predictions on easy-to-forget old classes when distilling knowledge via an individual sample. Moreover, different forgetting speeds of old classes may result in heterogeneous class-relation consistency between $\Theta^{t-1}$ and $\Theta^t$ . It enforces gradient optimization of $\Theta^t$ to bias towards some specific relation distillations related to new classes, causing forgetting heterogeneity on old categories. + +To overcome the above issues, we design a gradient-balanced relation distillation (GRD) loss $\mathcal{L}_{\mathrm{RD}}$ to distill relations among old and new classes from $\Theta^{t - 1}$ to $\Theta^t$ . As introduced in Fig. 2, it can address heterogeneous forgetting from gradient aspect via rectifying imbalanced gradient propagation caused by heterogeneous class-relation consistency. Specifically, we construct class-wise relation prototype rather than a prediction of single sample to perform relation distillation, which can alleviate negative effect of noisy relation consistency. Then the task-adaptive gradient means $\{\Gamma_{\eta}^{s}\}_{\eta = 1}^{t}$ in Eq. (8) are employed to reweight heterogeneous class-relation distillation. + +A mini-batch $\{\mathbf{x}_i^t,\mathbf{y}_i^t\}_{i = 1}^b\subset \mathcal{T}^t\cup \mathcal{M}$ is fed into $\Theta^{t - 1}$ and $\Theta^t$ to obtain softmax probabilities $\mathbf{P}^{t - 1}(\mathbf{x}_i^t,\Theta^{t - 1})\in \mathbb{R}^{K^o}$ of old classes and $\mathbf{P}^t (\mathbf{x}_i^t,\Theta^t)\in \mathbb{R}^{K^o +K^t}$ of old and new classes, as shown in Fig. 2. The first $K^o$ dimensions of $\mathbf{y}_i^t\in \mathbb{R}^{K^o +K^t}$ are replaced with $\mathbf{P}^{t - 1}(\mathbf{x}_i^t,\Theta^{t - 1})$ to get relation groundtruth $\mathbf{Y}^t (\mathbf{x}_i^t,\Theta^{t - 1})\in \mathbb{R}^{K^o +K^t}$ that implies underlying relations among old and new classes. To tackle noisy class relations, we construct class-wise relation prototype $\tilde{\mathbf{P}}_k^t$ and its relation groundtruth $\tilde{\mathbf{Y}}_k^t$ for the $k$ -th class: + +$$ +\tilde {\mathbf {P}} _ {k} ^ {t} = \frac {1}{\Delta_ {k}} \sum_ {i = 1} ^ {b} \mathbf {P} ^ {t} \left(\mathbf {x} _ {i} ^ {t}, \Theta^ {t}\right) \cdot \mathbb {I} _ {\arg \max \mathbf {y} _ {i} ^ {t} = k}, \tag {10} +$$ + +$$ +\tilde {\mathbf {Y}} _ {k} ^ {t} = \frac {1}{\Delta_ {k}} \sum_ {i = 1} ^ {b} \mathbf {Y} ^ {t} \left(\mathbf {x} _ {i} ^ {t}, \Theta^ {t - 1}\right) \cdot \mathbb {I} _ {\arg \max \mathbf {y} _ {i} ^ {t} = k}. \tag {11} +$$ + +where $\Delta_k = \sum_i^b\mathbb{I}_{\mathrm{argmax}\mathbf{y}_i^t = k}$ . Then we formulate category-wise gradient mean $\Gamma_k^s$ for the $k$ -th category as follows: + +$$ +\Gamma_ {k} ^ {s} = \frac {\sum_ {i = 1} ^ {b} \log \left(\left| \Gamma_ {i} ^ {t} \right| ^ {\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\right) \cdot \mathbb {I} _ {\arg \max \mathbf {y} _ {i} ^ {t} = k}}{\sum_ {i = 1} ^ {b} \mathbb {I} _ {\arg \max \mathbf {y} _ {i} ^ {t} = k}}. \tag {12} +$$ + +Consequently, we use $\Gamma_k^s$ to reweight heterogeneous relation distillation, and express the proposed $\mathcal{L}_{\mathrm{RD}}$ as follows: + +$$ +\mathcal {L} _ {\mathrm {R D}} = \frac {1}{K ^ {o} + K ^ {t}} \sum_ {k = 1} ^ {K ^ {o} + K ^ {t}} \frac {\Gamma_ {k} ^ {s} \cdot \mathcal {D} _ {\mathrm {K L}} \left(\tilde {\mathbf {P}} _ {k} ^ {t} , \tilde {\mathbf {Y}} _ {k} ^ {t}\right)}{\sum_ {\eta = 1} ^ {t} \Gamma_ {\eta} ^ {s} \cdot \mathbb {I} _ {k \in \mathcal {Y} ^ {\eta}}}, \tag {13} +$$ + +where $\mathcal{D}_{\mathrm{KL}}(\cdot ||\cdot)$ indicates the Kullback-Leibler divergence. + +Table 1. Results on CIFAR-100 [33]. When $\mathcal{M} = 2000$ , we set $T = \{5, 10, 20, 25, 50\}$ for $\mathcal{B} = 0\%$ , and $T = \{5, 10, 25, 50\}$ for $\mathcal{B} = 50\%$ . + +
Comparison MethodsBackbone#ParamsB=0%B=50%
510202550Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M83.580.678.577.676.479.3↑5.580.678.675.471.076.4↑2.9
BiC [50] (CVPR'2019)ViT-Base85.10M83.582.578.780.378.480.7↑4.278.275.774.271.975.0↑4.3
PODNet [17] (ECCV'2020)ViT-Base85.10M75.371.866.564.963.068.3↑16.672.873.074.274.973.7↑5.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.682.079.179.177.380.2↑4.675.874.872.972.073.9↑5.4
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M75.272.066.165.263.868.5↑16.472.772.874.074.773.5↑5.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M84.783.579.678.577.980.8↑4.178.477.376.274.076.5↑2.8
AFC [32] (CVPR'2022)ViT-Base85.10M74.675.570.367.464.770.5↑14.470.869.468.969.969.8↑9.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.585.881.380.877.482.2↑2.782.581.077.374.578.8↑0.5
HFC (Ours)ViT-Base85.10M86.486.385.585.081.184.9-82.981.077.875.579.3-
Upper BoundViT-Base85.10M94.294.294.294.294.294.2-94.294.294.294.294.2-
+ +Table 2. Results on ImageNet-100 [10]. When $\mathcal{M} = {2000}$ ,we set $T = \{ 5,{10},{20},{25}\}$ for $\mathcal{B} = 0\%$ ,and $T = \{ 5,{10},{25},{50}\}$ for $\mathcal{B} = {50}\%$ . + +
Comparison MethodsBackbone#ParamsB=0%B=50%
5102025Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M84.182.480.278.681.3↑3.279.578.876.673.977.2↑3.4
BiC [50] (CVPR'2019)ViT-Base85.10M81.480.279.278.579.8↑4.773.974.473.472.273.5↑7.1
PODNet [17] (ECCV'2020)ViT-Base85.10M82.480.179.578.380.1↑4.568.471.375.477.673.2↑7.4
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.081.780.979.481.2↑3.375.874.872.972.073.9↑6.7
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M82.381.079.379.080.4↑4.168.371.275.076.972.8↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M85.684.083.883.184.1↑0.481.080.580.079.880.3↑0.3
AFC [32] (CVPR'2022)ViT-Base85.10M83.281.280.579.481.1↑3.577.178.478.978.178.1↑2.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.183.480.080.382.2↑2.377.478.479.779.478.7↑1.9
HFC (Ours)ViT-Base85.10M86.185.484.082.684.5-81.581.180.379.580.6-
Upper BoundViT-Base85.10M86.386.386.386.386.3-86.386.386.386.386.3-
+ +Table 3. Results on ImageNet-1000 [10]. When $\mathcal{M} = 20000$ , we set $T = \{5,10\}$ for $\mathcal{B} = 0\%$ . + +
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M75.472.473.9↑3.5
BiC [50] (CVPR'2019)ViT-Base85.10M68.966.067.5↑10.0
PODNet [17] (ECCV'2020)ViT-Base85.10M68.471.369.8↑7.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M75.874.875.3↑2.2
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M68.463.265.8↑11.6
FOSTER [47] (ECCV'2022)ViT-Base85.10M76.974.375.6↑1.9
AFC [32] (CVPR'2022)ViT-Base85.10M74.271.472.8↑4.6
DyTox [18] (CVPR'2022)ViT-Base85.10M78.174.776.4↑1.0
HFC (Ours)ViT-Base85.10M78.576.477.5-
Upper BoundViT-Base85.10M86.386.386.3-
+ +# 3.5. Optimization Pipeline + +Overall, in the $t$ -th $(t \geq 2)$ learning task, the objective formulation to optimize $\Theta^t$ is expressed as follows: + +$$ +\mathcal {L} _ {\mathrm {o b j}} = \alpha_ {1} \mathcal {L} _ {\mathrm {F C}} + \alpha_ {2} \mathcal {L} _ {\mathrm {R D}}, \tag {14} +$$ + +where $\alpha_{1},\alpha_{2}$ are hyper-parameters. $\Theta^t$ along with learnable $\mathbf{E}_0^t$ are optimized via $\mathcal{L}_{\mathrm{CE}}$ in Eq. (5) for the first task, and trained via $\mathcal{L}_{\mathrm{obj}}$ in Eq. (14) when $t\geq 2$ . After optimizing new model $\Theta^t$ , we store $\Theta^t$ as the frozen old model $\Theta^{t - 1}$ to perform the GRD loss $\mathcal{L}_{\mathrm{RD}}$ for the next incremental task. + +# 4. Experiments + +# 4.1. Implementation Details + +For fair comparisons, we follow the same protocols utilized in baseline CIL methods [40, 50, 17, 17, 2, 18, 32] to + +set incremental tasks (class order), and do experiments on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Specifically, we consider two task settings in this paper: 1) We divide all categories of each dataset into $T$ tasks equally [40, 18] (i.e., the base class set $\mathcal{B} = 0\%$ ); 2) We start by training our model on half of the categories, while divide the rest of categories into $T$ tasks equally [17, 32] (i.e., the base class set $\mathcal{B} = 50\%$ ). When $\mathcal{B} = 0\%$ , we set $T = \{5, 10, 20, 25, 50\}$ for CIFAR-100 [33], $T = \{5, 10, 20, 25\}$ for ImageNet-100 [10] and $T = \{5, 10\}$ for ImageNet-1000 [10]. When $\mathcal{B} = 50\%$ , we set $T = \{5, 10, 25, 50\}$ for CIFAR-100 [33] and ImageNet-100 [10]. + +As introduced in [40, 50], the size of memory $\mathcal{M}$ for all comparison methods is fixed as 2,000 for CIFAR-100 [33] and ImageNet-100 [10], and 20,000 for ImageNet-1000 [10]. We follow iCaRL [40] to update memory $\mathcal{M}$ and use global memory for [18]. Besides, we employ the same data augmentation proposed in DyTox [18] for all comparison methods, but don't use mixup technology. Following [17, 32], we also consider another memory setting (i.e., storing 20 exemplars per class), and set $T = \{5,10\}, \mathcal{B} = 50\%$ for comparisons on CIFAR-100 [33] (see Tab. 6). For network architecture, we use ViT-Base [15] as feature extractor containing $l_{m} = 11$ MSA blocks and $l_{t} = 1$ TSA block, where the parameters are initialized via [22]. The task classifier includes only one fully-connected layer. The SGD optimizer is employed to optimize our model, where the learning rate is initialized as $6.25 \times 10^{-5}$ . We apply two plug-and-play losses (i.e., $\mathcal{L}_{\mathrm{FC}}$ + +Table 4. Performance on CIFAR-100 [33] ( $T = \{5, 10, 20, 25, 50\}$ ) and ImageNet-100 [10] ( $T = \{5, 10, 20, 25\}$ ) when we apply $\mathbf{Ours}^{\ddagger}$ into existing distillation-based CIL methods and set $\mathcal{M} = 2000$ , $\mathcal{B} = 0\%$ . $\mathbf{Ours}^{\ddagger}$ denotes the proposed plug-and-play losses $\mathcal{L}_{\mathrm{FC}}$ and $\mathcal{L}_{\mathrm{RD}}$ . + +
Comparison MethodsCIFAR-100 Dataset [33]ImageNet-100 Dataset [10]
Backbone#Params510202550Avg.Backbone#Params5102025
iCaRL [40] (CVPR'2017)ResNet-320.46M66.463.953.150.239.154.5ResNet-1811.22M76.469.560.657.9
iCaRL [40] + Ours‡ResNet-320.46M68.065.754.251.440.856.0ResNet-1811.22M76.870.361.759.2
BiC [50] (CVPR'2019)ResNet-320.46M55.350.848.346.439.348.0ResNet-1811.22M61.554.346.744.5
BiC [50] + Ours‡ResNet-320.46M63.253.852.449.341.552.0ResNet-1811.22M63.257.050.348.4
PODNet [17] (ECCV'2020)ResNet-320.46M63.854.548.144.439.750.1ResNet-1811.22M74.966.555.851.7
PODNet [17] + Ours‡ResNet-320.46M65.860.051.648.344.754.1ResNet-1811.22M76.569.359.455.8
SS-IL [2] (ICCV'2021)ResNet-320.46M65.159.049.048.640.752.5ResNet-1811.22M60.953.539.335.8
SS-IL [2] + Ours‡ResNet-320.46M67.764.951.651.242.555.6ResNet-1811.22M75.267.557.956.0
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M63.754.347.844.039.049.8ResNet-1811.22M74.866.355.451.3
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M66.758.550.546.943.353.2ResNet-1811.22M76.468.758.955.1
FOSTER [47] (ECCV'2022)ResNet-320.46M72.068.767.166.062.667.3ResNet-1811.22M74.970.967.265.7
FOSTER [47] + Ours‡ResNet-320.46M72.269.967.466.762.867.8ResNet-1811.22M75.071.367.566.2
AFC [32] (CVPR'2022)ResNet-320.46M65.657.150.948.142.052.7ResNet-1811.22M76.369.761.158.1
AFC [32] + Ours‡ResNet-320.46M66.263.853.249.744.155.4ResNet-1811.22M77.471.263.367.4
Upper BoundResNet-320.46M76.676.676.676.676.676.6ResNet-1811.22M73.273.273.273.2
DyTox [18] (CVPR'2022)ViT-Tiny10.71M75.373.572.772.370.572.9ViT-Tiny10.71M77.875.472.972.3
DyTox [18] + Ours‡ViT-Tiny10.71M75.974.072.972.670.873.2ViT-Tiny10.71M77.975.673.172.9
Upper BoundViT-Tiny10.71M76.176.176.176.176.176.1ViT-Tiny10.71M79.179.179.179.1
+ +Table 5. Ablation studies on CIFAR-100 [33] (the top block) and ImageNet-100 [10] (the bottom block), when we set $T = \{5,10,25\}$ under the settings of $\mathcal{M} = 2000, \mathcal{B} = 0\%$ . + +
Ablation VariantsBackbone#ParamsTSAGFCGRD51025
BaselineViT-Base85.10MXXX81.680.278.1
Baseline + TSAViT-Base85.10M✓XX83.182.481.0
Baseline + TSA + GFCViT-Base85.10M✓✓X84.083.782.1
HFC (Ours)ViT-Base85.10M✓✓✓86.486.385.0
BaselineViT-Base85.10MXXX82.080.578.8
Baseline + TSAViT-Base85.10M✓XX83.782.881.2
Baseline + TSA + GFCViT-Base85.10M✓XX84.383.482.0
HFC (Ours)ViT-Base85.10M✓✓✓86.185.484.0
+ +and $\mathcal{L}_{\mathrm{RD}})$ to existing CIL methods [40, 18] using ResNet-18 [23], ResNet-32 [23] or ViT-Tiny [15] as backbone. For evaluation, top-1 accuracy is used to compare performances of our model and baseline methods. + +# 4.2. Comparison Experiments + +Comparison Results: Tabs. 1-3 present comparison results of our HFC model and off-the-shelf CIL methods on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Our model significantly outperforms existing CIL methods [40, 50, 17, 2, 18, 32, 47] by $0.7\% \sim 16.6\%$ accuracy under various incremental settings, when we use the same backbone (i.e., ViT-Base [15]) as feature extractor for fair comparisons. Such large performance improvement illustrates the superiority of our model to address forgetting heterogeneity among easy-to-forget and hard-to-forget old categories from both representation and gradient aspects. + +Improvement of Plug-and-Play Losses: Tab. 4 shows large performance improvement of existing knowledge + +based CIL methods [40, 50, 17, 4, 18, 32, 47], when we apply the proposed plug-and-play losses (i.e., $\mathcal{L}_{\mathrm{FC}}$ and $\mathcal{L}_{\mathrm{RD}}$ ) to them. The proposed losses $\mathcal{L}_{\mathrm{FC}}$ and $\mathcal{L}_{\mathrm{RD}}$ help existing CIL methods to address heterogeneous forgetting from gradient aspect, thus largely improving their performance. Besides, it validates generalization and robustness of our HFC model. + +# 4.3. Ablation Studies + +As shown in Tab. 5, we present ablation studies on CIFAR-100 [33] and ImageNet-100 [10] to investigate the effectiveness of each module. TSA, GFC and GRD are task-semantic aggregation blocks, the gradient-balanced forgetting compensation loss $\mathcal{L}_{\mathrm{FC}}$ and gradient-balanced relation distillation loss $\mathcal{L}_{\mathrm{RD}}$ . Baseline denotes the performance of our model using the traditional classification loss and knowledge distillation loss proposed in DyTox (CVPR'2022) [18], but without the TSA, GFC and GRD modules. Compared with Ours, the performance of all ablation variants degrades significantly. It validates superiority of each module to overcome forgetting heterogeneity of different old categories. + +# 4.4. Analysis of Task-Wise Comparisons + +Figs. 3-4 present task-wise performance comparisons between our model and other CIL methods [40, 50, 17, 18, 32] when we set backbone as ViT-Base and different tasks as $T = \{5, 10, 20, 25\}$ . Our HFC model outperforms baseline methods for most task-wise comparisons, which illustrates the superior performance of our model to surmount heterogeneous forgetting from representation and gradient aspects. To identify new classes continually, the task-semantic aggregation (TSA) block explores task-shared global representa + +![](images/64f4d3f7af5d54ecb9432c85770b0132db53aa3f32f9f0dc5cdad2675c18a6e6.jpg) +Figure 3. Analysis of task-wise performance comparisons on CIFAR-100 [33]. We set $\mathcal{M} = 2000$ , $\mathcal{B} = 0\%$ when the backbone is ViT-Base. + +![](images/c4a3ab5333a42f77c780318b84a9bf1e3211844f8753c898d288195032a9274e.jpg) +Figure 4. Analysis of task-wise performance on ImageNet-100 [10]. We set $\mathcal{M} = 2000$ , $\mathcal{B} = 0\%$ when the backbone is ViT-Base. + +Table 6. Results on CIFAR-100 [33] when storing 20 exemplars for each class. We set $T = \{5,10\}$ for $\mathcal{B} = 50\%$ , and $\mathbf{Ours}^{\ddagger}$ denotes the proposed plug-and-play losses $\mathcal{L}_{\mathrm{FC}}$ and $\mathcal{L}_{\mathrm{RD}}$ . + +
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M78.876.277.5↑1.9
BiC [50] (CVPR'2019)ViT-Base85.10M75.374.574.9↑4.5
PODNet [17] (ECCV'2020)ViT-Base85.10M71.671.671.6↑7.8
SS-IL [2] (ICCV'2021)ViT-Base85.10M73.071.872.4↑7.0
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M71.571.771.6↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M78.878.778.8↑0.6
AFC [32] (CVPR'2022)ViT-Base85.10M68.265.366.8↑12.6
DyTox [18] (CVPR'2022)ViT-Base85.10M79.378.478.8↑0.6
HFC (Ours)ViT-Base85.10M79.878.979.4-
Upper BoundViT-Base85.10M94.294.294.2-
iCaRL [40] (CVPR'2017)ResNet-320.46M56.653.254.9↑0.7
iCaRL [40] + Ours‡ResNet-320.46M57.353.955.6-
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M62.261.161.7↑0.6
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M62.961.662.3-
FOSTER [47] (ECCV'2022)ResNet-320.46M67.366.466.8↑0.5
FOSTER [47] + Ours‡ResNet-320.46M67.766.667.2-
AFC [32] (CVPR'2022)ResNet-320.46M63.863.663.7↑0.7
AFC [32] + Ours‡ResNet-320.46M64.464.364.4-
Upper BoundResNet-320.46M76.676.676.6-
DyTox [18] (CVPR'2022)ViT-Tiny10.71M67.966.267.1↑0.5
DyTox [18] + Ours‡ViT-Tiny10.71M68.566.767.6-
Upper BoundViT-Tiny10.71M76.176.176.1-
+ +tions to alleviate forgetting from representation aspect, while $\mathcal{L}_{\mathrm{FC}}$ and $\mathcal{L}_{\mathrm{RD}}$ can achieve gradient-balanced compensation. + +# 4.5. Analysis of Forgetting Heterogeneity + +Tab. 7 presents forgetting heterogeneity (FH) of different old classes via measuring their variance of gradient updating. We formulate forgetting heterogeneity (FH) as $\mathrm{FH} = \frac{1}{T}\sum_{t=1}^{T}\left(\frac{1}{S_t}\sum_{i=1}^{S_t}(|\Gamma_i^t| - \sum_{\eta=1}^t\Gamma_\eta \cdot \mathbb{I}_{\mathbf{y}_i^t \in \mathcal{Y}^\eta})^2\right)$ , where $\Gamma_i^t$ and $\Gamma_\eta$ are obtained from Eqs. (6)-(7), and $S_t$ is + +Table 7. Analysis of forgetting heterogeneity (FH) on CIFAR-100 [33] when $T = 5, \mathcal{B} = 0\%$ . AC denotes the averaged accuracy. + +
M=2000,B=0%BackboneTSAGFCGRDFHAC
iCaRL [40] (CVPR'2017)ViT-Base---131.583.5
FOSTER [47] (ECCV'2022)ViT-Base---119.485.1
AFC [32] (CVPR'2022)ViT-Base---108.574.6
DyTox [18] (CVPR'2022)ViT-Base---105.185.5
BaselineViT-BaseXXX128.981.6
Baseline + TSAViT-Base✓XX125.483.1
Baseline + TSA + GFCViT-Base✓✓X120.584.0
HFC (Ours)ViT-Base✓✓✓103.686.4
+ +number of samples in test set. Tab. 7 verifies superiority of our model against other CIL methods to tackle heterogeneous forgetting. It also shows effectiveness of all modules to collaboratively minimize forgetting heterogeneity. + +# 5. Conclusion + +In this paper, we develop a novel Heterogeneous Forgetting Compensation (HFC) model to surmount heterogeneous forgetting from representation and gradient aspects. To be specific, a task-semantic aggregation block is designed to tackle heterogeneous forgetting from representation aspect via exploring task-shared global representations. Moreover, we propose a gradient-balanced forgetting compensation loss and a relation distillation loss to compensate forgetting heterogeneity from gradient aspect via performing balanced gradient propagation and distilling heterogeneous class relations. Experiments verify the superiority of our proposed HFC model against baseline methods. We will further consider addressing noisy forgetting heterogeneity in the future. + +# References + +[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, pages 3930-3939, 2020. +[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, pages 844-853, October 2021. +[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 144-161, 2018. +[4] Arjun Ashok, K. J. Joseph, and Vineeth N. Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In ECCV, pages 105-122, 2022. +[5] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022. +[6] Francisco M. Castro, Manuel J. Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 241-257, Cham, 2018. Springer International Publishing. +[7] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. +[8] Huitong Chen, Yu Wang, and Qinghua Hu. Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering, pages 1-15, 2022. +[9] Stéphane d'Ascoli, Hugo Touvron, Matthew Leavitt, Ari Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021. +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. +[11] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11-20, 2021. +[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In CVPR, pages 10164-10173, June 2022. +[13] Jiahua Dong, Duzhen Zhang, Yang Cong, Wei Cong, Henghui Ding, and Dengxin Dai. Federated incremental semantic segmentation. In CVPR, pages 3934-3943, June 2023. +[14] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In CVPR, pages 19217-19227, June 2022. + +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. +[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In CVPR, 2021. +[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020. +[18] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285-9295, June 2022. +[19] Zhen Fang, Yixuan Li, Jie Lu, Jiahua Dong, Bo Han, and Feng Liu. Is out-of-distribution detection learnable? In NeurIPS, volume 35, pages 37199-37213, 2022. +[20] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. In ICML, page 1243-1252, 2017. +[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, volume 27, 2014. +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, June 2022. +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. +[24] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. In NeurIPS Workshop, 2015. +[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, June 2019. +[26] Weiming Hu, Xi Li, Guodong Tian, Stephen Maybank, and Zhongfei Zhang. An incremental dpmm-based method for trajectory clustering, modeling, and retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(5):1051-1065, 2013. +[27] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021. +[28] Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unforgetting continual learning. In Advances in Neural Information Processing Systems, pages 13647–13657, 2019. +[29] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 699-715, 2020. + +[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, and et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022. +[31] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwar, and Vineeth Balasubramanian. Energy-based latent aligner for incremental learning. In CVPR, 2022. +[32] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, pages 16071-16080, June 2022. +[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, pages 32-33, 2009. +[34] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018. +[35] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2022. +[36] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. AAAI, 37(7):8906-8913, Jun. 2023. +[37] Yaoyao Liu, Bernt Schiele, and Qianru Sun. RMM: reinforced memory management for class-incremental learning. In NeurIPS, pages 3478-3490, 2021. +[38] John M. Pierre. Incremental lifelong deep learning for autonomous vehicles. In ITSC, pages 3949-3954, 2018. +[39] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for continual learning. In NeurIPS, 2019. +[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017. +[41] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, volume 80, pages 4548-4557, 2018. +[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In CVPR, 2022. +[43] Zhenheng Tang, Yonggang Zhang, Shaohuai Shi, Xin He, Bo Han, and Xiaowen Chu. Virtual homogeneity learning: Defending against data heterogeneity in federated learning. In ICML, pages 21111-21132, 2022. +[44] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 254-270, Cham, 2020. Springer International Publishing. +[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, pages 99-108, June 2022. +[46] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021. + +[47] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022. +[48] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In AAAI, volume 35, pages 10165-10173, 2021. +[49] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022. +[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019. +[51] Ziyang Wu, Christina Baek, Chong You, and Yi Ma. Incremental learning via rate reduction. In CVPR, pages 1125-1133, 2021. +[52] Rongtao Xu, Changwei Wang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Wave-like class activation map with representation fusion for weakly-supervised semantic segmentation. IEEE Transactions on Multimedia, 2023. +[53] Rongtao Xu, Changwei Wang, Jiguang Zhang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. RSSformer: Foreground saliency enhancement for remote sensing land-cover segmentation. IEEE Transactions on Image Processing, 32:1052-1064, 2023. +[54] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021. +[55] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018. +[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, page 3987-3995, 2017. +[57] Duzhen Zhang, Feilong Chen, and Xiuyi Chen. Dualgats: Dual graph attention networks for emotion recognition in conversations. In ACL, pages 7395-7408, 2023. +[58] Yonggang Zhang, Mingming Gong, Tongliang Liu, Gang Niu, Xinmei Tian, Bo Han, Bernhard Schölkopf, and Kun Zhang. Causaladv: Adversarial robustness through the lens of causality. *ICLR*, 2022. +[59] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13205-13214, 2020. +[60] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. +[61] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In ECCV, pages 55-72, 2022. \ No newline at end of file diff --git a/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/images.zip b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b519e978d696242cdea084cb819d1563d9ed8fcb --- /dev/null +++ b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e3db27f9988e62a789d16c564ecde52b938566676b4bb6693b6ba33f5c7300a +size 1061710 diff --git a/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/layout.json b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..950022fd77bc20a4f6c1fc8e805a3fc65d5e0861 --- /dev/null +++ b/2023/Heterogeneous Forgetting Compensation for Class-Incremental Learning/layout.json @@ -0,0 +1,12574 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 103, + 523, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 103, + 523, + 123 + ], + "spans": [ + { + "bbox": [ + 71, + 103, + 523, + 123 + ], + "type": "text", + "content": "Heterogeneous Forgetting Compensation for Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "spans": [ + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "text", + "content": "Jiahua Dong" + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "text", + "content": ", Wenqi Liang" + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "text", + "content": ", Yang Cong" + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "text", + "content": ", Gan Sun" + }, + { + "bbox": [ + 144, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 158, + 466, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 158, + 466, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 158, + 466, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 158, + 466, + 185 + ], + "type": "text", + "content": "State Key Laboratory of Robotics, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, 110016, China.‡" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 185, + 436, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 185, + 436, + 213 + ], + "spans": [ + { + "bbox": [ + 156, + 185, + 436, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 156, + 185, + 436, + 213 + ], + "type": "text", + "content": "Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang, 110169, China." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 213, + 465, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 213, + 465, + 227 + ], + "spans": [ + { + "bbox": [ + 127, + 213, + 465, + 227 + ], + "type": "text", + "content": "3University of Chinese Academy of Sciences, Beijing, 100049, China." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 227, + 462, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 462, + 241 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 462, + 241 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 227, + 462, + 241 + ], + "type": "text", + "content": "South China University of Technology, Guangzhou, 510640, China." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 243, + 471, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 243, + 471, + 255 + ], + "spans": [ + { + "bbox": [ + 123, + 243, + 471, + 255 + ], + "type": "text", + "content": "{dongjiahua1995, liangwenqi0123, congyang81, sunganl412}@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 308, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 290, + 597 + ], + "type": "text", + "content": "Class-incremental learning (CIL) has achieved remarkable successes in learning new classes consecutively while overcoming catastrophic forgetting on old categories. However, most existing CIL methods unreasonably assume that all old categories have the same forgetting pace, and neglect negative influence of forgetting heterogeneity among different old classes on forgetting compensation. To surmount the above challenges, we develop a novel Heterogeneous Forgetting Compensation (HFC) model, which can resolve heterogeneous forgetting of easy-to-forget and hard-to-remember old categories from both representation and gradient aspects. Specifically, we design a task-semantic aggregation block to alleviate heterogeneous forgetting from representation aspect. It aggregates local category information within each task to learn task-shared global representations. Moreover, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation loss and a gradient-balanced relation distillation loss to alleviate forgetting from gradient aspect. They consider gradient-balanced compensation to rectify forgetting heterogeneity of old categories and heterogeneous relation consistency. Experiments on several representative datasets illustrate effectiveness of our HFC model. The code is available at https://github.com/JiahuaDong/HFC." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 619, + 128, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 128, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 128, + 631 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 639, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 288, + 676 + ], + "type": "text", + "content": "Class-incremental learning (CIL) [40, 12, 5] has attracted appealing attentions recently by accumulating previous learned experience to learn new classes incrementally. It" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 281, + 533, + 428 + ], + "blocks": [ + { + "bbox": [ + 320, + 281, + 533, + 428 + ], + "lines": [ + { + "bbox": [ + 320, + 281, + 533, + 428 + ], + "spans": [ + { + "bbox": [ + 320, + 281, + 533, + 428 + ], + "type": "image", + "image_path": "4f3ad93781789f1aaea7c553553eddf6928008ad3c55f0926f3a1daa3149e0ff.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 430, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 306, + 430, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 547, + 453 + ], + "type": "text", + "content": "Figure 1. Forgetting heterogeneity versus accuracy on CIFAR-100 [33] when the backbone is ViT-Base and number of tasks is 5." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 460, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 605 + ], + "type": "text", + "content": "plays an indispensable role in developing a large number of intelligent learning systems, such as autonomous driving [38] and automated surveillance [26]. When learning new classes continuously under the settings of limited memory to replay all previous data of old classes [45], these CIL systems heavily suffer from forgetting on old classes caused by severe class imbalance between old and new categories [40, 31, 14]. To surmount catastrophic forgetting, a growing amount of CIL methods mainly perform knowledge distillation [24] to preserve past experience [34, 12, 35]; introduce a rehearsal strategy to replay part of old data [40, 17]; or dynamically expand network architectures [28, 41, 54]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "However, most existing CIL methods [40, 45, 54, 18] neglect heterogeneous forgetting speeds of different old classes. They unrealistically assume that all old classes suffer from the same degree of catastrophic forgetting, and compensate forgetting for each old classes equally and independently. Such impracticable assumption enforces existing CIL models [40, 56, 61] to suffer from imbalanced gradient optimization among different old classes, thus favoring more forgetting compensation for hard-to-forget old classes while ne" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": "*Equal contributions. †The corresponding author is Prof. Yang Cong. ‡This work was supported in part by the National Nature Science Foundation of China under Grant 62127807, 62273333 and 62133005." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11742" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "glecting those easy-to-forget old classes (i.e., heterogeneous forgetting). More importantly, such forgetting heterogeneity can significantly worsen the forgetting on hard-to-forget old categories, when new streaming classes becomes part of old categories continually. For example, some old classes with various modalities and appearances (e.g., person and car) in autonomous driving [38] are more difficult to explore task-shared representations across different incremental tasks, when compared with other hard-to-forget old classes (e.g., road and traffic sign) with easily-distinguished visual properties. This phenomenon causes imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus exacerbating heterogeneous forgetting speeds among different old classes. When autonomous vehicles [38] learn new classes continually, such forgetting heterogeneity aggravates forgetting on those hard-to-forget old classes (e.g., traffic sign) to some extent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 276, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 289, + 361 + ], + "type": "text", + "content": "Inspired by the above practical scenario, we investigate that different old classes have significant forgetting heterogeneity in this paper, as shown in Fig. 1. This heterogeneous forgetting might heavily weaken forgetting compensation on hard-to-forget old classes, as new classes become a subset of old classes consecutively. In summary, the challenges to tackle heterogeneous forgetting lie in two major aspects:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 363, + 289, + 544 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 51, + 363, + 289, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 363, + 289, + 459 + ], + "spans": [ + { + "bbox": [ + 51, + 363, + 289, + 459 + ], + "type": "text", + "content": "- Representation Aspect: Some easy-to-forget old classes (e.g., car and bus) with diverse appearances are more difficult for existing CIL methods [45, 54, 12, 3] to learn intrinsic task-shared representations, and thus are significantly easier to be forgotten than hard-to-forget old classes (e.g., road) with distinctive attributes. Thus, exploring task-shared representations is essential to address heterogeneous forgetting among different old classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 459, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 459, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 51, + 459, + 289, + 544 + ], + "type": "text", + "content": "- Gradient Aspect: To learn complex visual characterizations for easy-to-forget old classes with various modalities and appearances, existing CIL methods [40, 4, 49] are required to allocate more network architectures for gradient updating. It can result in imbalanced gradient propagation between easy-to-forget and hard-to-forget old categories, thus aggravating forgetting heterogeneity of old classes." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": "To overcome the above challenges, we propose a novel Heterogeneous Forgetting Compensation (HFC) model, which is an earlier exploration to address heterogeneous forgetting from both representation and gradient perspectives in the CIL field [40, 4]. Specifically, we propose a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation perspective. It can explore task-shared global representations for each class via aggregating long-range local category information within each task. Meanwhile, to tackle heterogeneous forgetting from gradient perspective, we develop two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss. The GFC loss can rectify heterogeneous forgetting speeds" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "of easy-to-forget and hard-to-forget old classes, while normalize different learning paces of new categories to achieve gradient-balanced propagation. Besides, the GRD loss can distill heterogeneous relation consistency brought by forgetting heterogeneity among different old classes. Experiments show our model has large improvements on several representative datasets, compared with baseline methods [40, 36]. More importantly, we apply two plug-and-play losses into existing distillation-based CIL methods [40, 17, 4] to significantly improve their performance from gradient aspect. The novel contributions of this paper are presented as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 208, + 547, + 413 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 309, + 208, + 547, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 208, + 547, + 280 + ], + "spans": [ + { + "bbox": [ + 309, + 208, + 547, + 280 + ], + "type": "text", + "content": "- We develop a novel Heterogeneous Forgetting Compensation (HFC) model to address different forgetting speeds of easy-to-forget and hard-to-forget old classes. To our best knowledge, this paper is the first exploration to tackle forgetting heterogeneity among old categories from representation and gradient aspects in the CIL field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 280, + 547, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 280, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 280, + 547, + 328 + ], + "type": "text", + "content": "- We design a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect. It can explore robust task-shared representations for each class via aggregating local category information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 309, + 328, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 328, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 309, + 328, + 547, + 413 + ], + "type": "text", + "content": "- We propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss and a gradient-balanced relation distillation (GRD) loss to surmount forgetting heterogeneity from gradient aspect. They can balance different forgetting of old classes and heterogeneous category-relation consistency to improve performance when applying them into existing methods." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 426, + 392, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 392, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 392, + 438 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 447, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 547, + 471 + ], + "type": "text", + "content": "We discuss some class-incremental learning (CIL) methods [60, 37, 18], and divide them into three categories:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 568 + ], + "type": "text", + "content": "Knowledge Distillation: After Li et al. [34] apply knowledge distillation [24] to continual learning, [40, 39, 59] follow them to transfer past experience from old model to new model. iCaRL [40] and EEIL [6] perform knowledge distillation in the output space. [44] aims to preserve consistent topology feature space for old and new tasks. [16] proposes feature-level knowledge distillation when applying continual learning [37] into semantic segmentation task [53, 52, 11]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 569, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 653 + ], + "type": "text", + "content": "Rehearsal Strategy: For replaying past experience, lots of CIL methods [49, 8, 42] allocate a memory to store exemplars of old classes or synthesize samples via generative adversarial model [21, 58]. For saving large memory overhead, [29] proposes to store low-dimension features rather than raw samples. [51] proposes a \"white box\" framework deviated from rate reduction to preserve past experience." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Dynamic Architecture: Many works [28, 54, 1] allocate dynamical networks for new classes as the growing of learned classes. Yoon et al. [55] dynamically expand architecture capacity via selective retraining. [54, 18] address the problem of relying on task index when performing dynamic" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11743" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 109 + ], + "type": "text", + "content": "architectures. However, the above-mentioned CIL methods [40, 27] neglect negative influence of heterogeneous forgetting of old classes on forgetting compensation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 169, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 169, + 134 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 169, + 134 + ], + "type": "text", + "content": "3. The Proposed Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 141, + 227, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 227, + 153 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 227, + 153 + ], + "type": "text", + "content": "3.1. Problem Definition and Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": "Problem Definition: Following class-incremental learning (CIL) methods [40, 50, 17, 18, 54], we set a series of consecutive learning tasks as " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\mathcal{T}^t\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " indicates the total task quantity. The " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": "-th task " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^t = \\{\\mathbf{x}_i^t, \\mathbf{y}_i^t\\}_{i=1}^{N^t}" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "N^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " pairs of image " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " and one-hot groundtruth " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^t \\in \\mathcal{Y}^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " denotes label space of the " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": "-th incremental task containing " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "K^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " new categories. The label spaces between any two incremental tasks have no overlap: " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^t \\cap (\\cup_{j=1}^{t-1} \\mathcal{Y}^j) = \\emptyset" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ". Namely, " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "K^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " new classes belonging to the " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": "-th task " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " are different from " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "K^o = \\sum_{i=1}^{t-1} K^i \\subset \\cup_{j=1}^{t-1} \\mathcal{Y}^j" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " old classes learned from previous tasks " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}^i\\}_{i=1}^{t-1}" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ". As introduced in CIL baselines [40, 50, 54], we set a fixed exemplar memory " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ", and store only few images (i.e., " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\frac{|M|}{K^o}" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ") for each old class in the " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": "-th task " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " satisfies " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\frac{N^t}{K^t} \\gg \\frac{|M|}{K^o}" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": ". In the " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": "-th task, we aim to identify both " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "K^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " new classes and " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "K^o" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " old classes via optimizing the model on " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^t" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 159, + 289, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "Overview: The overview of our HFC model to surmount forgetting heterogeneity of different old categories from both representation and gradient aspects is shown in Fig. 2. Denote the proposed HFC model learned at the " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "-th tasks as old and new models (i.e., " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " is inherited from " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and only expands the number of output neurons in the classifier to identify " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "K^t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " new classes. In the " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "-th incremental task, given an image " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t\\in \\mathcal{T}^t\\cup \\mathcal{M}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": ", we forward it into " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " to extract task-shared global representation via task-semantic aggregation (TSA) block, which alleviates heterogeneous forgetting from representation aspect (Section 3.2). After obtaining probabilities " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " predicted via " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": ", we develop two plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and a gradient-balanced relation distillation (GRD) loss " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " to surmount heterogeneous forgetting from semantic aspect (Sections 3.3 and 3.4). " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " can normalize heterogeneous forgetting of old classes and distill heterogeneous relation consistency between " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 229, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 229, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 229, + 624 + ], + "type": "text", + "content": "3.2. Task-Semantic Aggregation Block" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "To explore long-range semantic dependencies across different tasks, Douillard et al. [18] introduce Vision Transformer (ViT) [15, 46, 22] into class-incremental learning (CIL). However, it significantly increases the memory overhead to store task tokens and task-specific decoders, when learning large-scale new classes. Besides, it cannot aggregate local category information to explore task-shared global" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "representations, resulting in heterogeneous forgetting from representation aspect. To tackle these issues, we develop a task-semantic aggregation (TSA) block to extract task-shared global representations, which is effective to tackle heterogeneous forgetting from representation aspect. More importantly, the TSA block cannot dynamically increase memory cost with the arrival of large-scale new categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": "Specifically, as presented in Fig. 2, the proposed model " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\Theta^t = \\Theta_c^t\\circ \\Theta_f^t" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": "-th task is composed of a feature extractor " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\Theta_f^t" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " to extract task-shared global representations, and a task classifier " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\Theta_c^t" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " to classify " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "K^o" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " old classes and " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "K^t" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " new classes. In the " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": "-th task, the network parameters of " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\Theta_f^{t - 1}" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\Theta_c^{t - 1}" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " learned at the " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": "-th task are frozen to preserve semantic knowledge for old classes. Different from [18] that uses vanilla ViT [15, 30, 57, 7] to extract features, we employ " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "l_{m}" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " multi-head self-attention (MSA) blocks [20] and " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " task-semantic aggregation (TSA) block as feature extractor " + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\Theta_f^t" + }, + { + "bbox": [ + 304, + 157, + 547, + 293 + ], + "type": "text", + "content": " to obtain task-shared global representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": "- Multi-Head Self-Attention (MSA): Given an image " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t\\in \\mathcal{T}^t\\cup \\mathcal{M}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " th task, following ViT [15, 9], we crop it into " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " patches with equal dimension, flatten these patches and map them into a " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " -dimension feature space via a linear projection to obtain patch embedding " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_e\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": ". Then we concatenate " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_e" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " with a class token " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathrm{cls}}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": ", and perform element-wise sum operation with a position token " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_p\\in \\mathbb{R}^{(N + 1)\\times D}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0 = [\\mathbf{Z}_e\\oplus \\mathbf{Z}_{\\mathrm{cls}}] + \\mathbf{Z}_p\\in \\mathbb{R}^{(N + 1)\\times D}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " is concatenation function. As shown in Fig.2, " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_0" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " is fed into " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\Theta_f^t" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " including " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "l_{m}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " MSA blocks to extract long-range semantics across different tasks. For the " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " -th " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "(l = 1,\\dots ,l_m)" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " MSA block, we regard " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{l - 1}\\in \\mathbb{R}^{(N + 1)\\times D}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " learned from the " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "(l - 1)" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " -th MSA block as input, and execute parallel selfattention " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " times. Thus, the self-attention " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_l^h\\in \\mathbb{R}^{(N + 1)\\times d}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{l - 1}" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " -th " + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "inline_equation", + "content": "(h = 1,\\dots ,H)" + }, + { + "bbox": [ + 304, + 293, + 547, + 473 + ], + "type": "text", + "content": " head is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 332, + 482, + 546, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 482, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 332, + 482, + 546, + 510 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {l} ^ {h} = \\sigma \\left(\\frac {\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {q} ^ {h} \\left(\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {k} ^ {h}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {Z} _ {l - 1} \\mathbf {W} _ {v} ^ {h}\\right), \\tag {1}", + "image_path": "df79daa3b09f6fb47eff6f4eb844468ace16cfd2d3af8a4e3a8fccc08006b03a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^h,\\mathbf{W}_k^h,\\mathbf{W}_v^h\\in \\mathbb{R}^{D\\times d}" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " are query, key and value mapping matrices. " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "d = \\frac{D}{H}" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " represents channel dimension of each head and " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " is the softmax function. Then we concatenate " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{A}_l^h\\}_{h = 1}^H" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " along channel dimension and project this concatenated result via a mapping matrix " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o\\in \\mathbb{R}^{D\\times D}" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " to perform MSA: " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{MSA}(\\mathbf{Z}_{l - 1}) = \\mathbf{Z}_{l - 1} + [\\mathbf{A}_l^1\\oplus" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_l^2,\\dots ,\\oplus \\mathbf{A}_l^H ]\\mathbf{W}_o" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " is concatenation function. After obtaining " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{MSA}(\\mathbf{Z}_{l - 1})\\in \\mathbb{R}^{(N + 1)\\times D}" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "(D = dH)" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": ", we utilize a multi-layer perceptron (MLP) block to obtain output " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_l\\in \\mathbb{R}^{(N + 1)\\times D}" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 520, + 547, + 643 + ], + "type": "text", + "content": " th MSA block:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 653, + 546, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 653, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 339, + 653, + 546, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {l} = \\operatorname {M S A} \\left(\\mathbf {Z} _ {l - 1}\\right) + \\operatorname {M L P} \\left(\\operatorname {M S A} \\left(\\mathbf {Z} _ {l - 1}\\right)\\right). \\tag {2}", + "image_path": "a42caa820626cfe3ee9fdb911a16c780c6583d3a9c7bc8560e01916933c9533f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "- Task-Semantic Aggregation (TSA): The patch embedding " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{l_m} \\in \\mathbb{R}^{(N + 1) \\times D}" + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": " encoded via the " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "l_m" + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "-th MSA block contains rich long-range semantic dependencies shared" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11744" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 250 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 250 + ], + "type": "image", + "image_path": "16353ab5f0807c1c32c2d176bfa051ceb2f1a6378bded8b6cd75f886f8c61555.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 251, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 45, + 251, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 45, + 251, + 547, + 295 + ], + "type": "text", + "content": "Figure 2. Illustration of the HFC model. It mainly contains a task-semantic aggregation (TSA) block to alleviate heterogeneous forgetting from representation aspect via exploring task-shared global representations, and two novel plug-and-play losses (i.e., a gradient-balanced forgetting compensation loss " + }, + { + "bbox": [ + 45, + 251, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 45, + 251, + 547, + 295 + ], + "type": "text", + "content": " and a gradient-balanced relation distillation loss " + }, + { + "bbox": [ + 45, + 251, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 45, + 251, + 547, + 295 + ], + "type": "text", + "content": ") to overcome forgetting heterogeneity from gradient aspect by compensating gradient-imbalanced propagation and heterogeneous class-relation consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": "across different tasks. " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{l_m}" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " is then forwarded into " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " task-semantic aggregation (TSA) blocks to extract task-shared global embedding. The TSA block aggregates local category context from long-range semantic dependencies to global representations shared across all incremental tasks, which is essential to alleviate heterogeneous forgetting from representation aspect. In the " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": "-th task, as depicted in Fig. 2, we introduce a learnable task-shared embedding " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_0^t \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " that is initialized via " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_0^{t - 1}" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " learned in the " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": "-th task. Then we forward " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_0^t" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{l_m}" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " TSA blocks. For the " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "(l = 1,\\dots ,l_t)" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " TSA block, " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{l - 1}^t \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " learned at " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "(l - 1)" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": "-th TSA block and " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{l_m}" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " are used to perform task-semantic attention " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " times in parallel. Thus, we formulate task-semantic attention " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_l^h \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "inline_equation", + "content": "(h = 1,\\dots ,H)" + }, + { + "bbox": [ + 46, + 304, + 290, + 473 + ], + "type": "text", + "content": " head as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 479, + 287, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 479, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 83, + 479, + 287, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {l} ^ {h} = \\sigma \\left(\\frac {\\mathbf {E} _ {l - 1} ^ {t} \\mathbf {V} _ {q} ^ {h} \\left(\\mathbf {Z} _ {l _ {m}} \\mathbf {V} _ {k} ^ {h}\\right) ^ {\\top}}{\\sqrt {d}}\\right) \\left(\\mathbf {Z} _ {l _ {m}} \\mathbf {V} _ {v} ^ {h}\\right), \\tag {3}", + "image_path": "5139828a7ec9d551455d80c6bcf0141f6e637d884bcf7f4bed61c2f50f0a662b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^h,\\mathbf{V}_k^h,\\mathbf{V}_v^h\\in \\mathbb{R}^{D\\times d}" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " are projection matrices of query, key and value. After utilizing Eq. (3) to obtain " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_l^h\\}_{h = 1}^H" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " , we concatenate them along channel dimension and project this fused result via " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_o\\in \\mathbb{R}^{D\\times D}" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " to execute " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " -head task-semantic attention: " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{TSA}(\\mathbf{E}_{l - 1}^{t}) = [\\mathbf{S}_{l}^{1}\\oplus" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_l^2,\\dots ,\\oplus \\mathbf{S}_l^H ]\\mathbf{V}_o\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " . TSA " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{l - 1}^{t})" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " is encoded via a MLP block to get output " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l^t\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 514, + 289, + 601 + ], + "type": "text", + "content": " -th TSA block:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 609, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 609, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 82, + 609, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {l} ^ {t} = \\operatorname {T S A} \\left(\\mathbf {E} _ {l - 1} ^ {t}\\right) + \\operatorname {M L P} \\left(\\operatorname {T S A} \\left(\\mathbf {E} _ {l - 1} ^ {t}\\right)\\right). \\tag {4}", + "image_path": "f7a10685c76bc3b33690bc18471be33ae2d41016f2e3d2c7e455f1c131d636fa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": "As aforementioned, we obtain the task-shared global representation " + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{l_t}^t\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": " via the " + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "inline_equation", + "content": "l_{t}" + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": " -th TSA block in feature extractor " + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Theta_f^t" + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": " . Obviously, the proposed TSA blocks can aggregate local category context of the " + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": " -th task from longrange semantic dependencies to task-shared global representation " + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{l_t}^t" + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": " . It is essential to alleviate heterogeneous forgetting among different old classes from representation aspect, by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 304, + 547, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 341 + ], + "type": "text", + "content": "exploring task-shared semantic context across different tasks. Note that we employ layer norm before TSA, MSA and MLP blocks, while we omit it for simplicity in this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": "- Task Classifier: The task-shared global representation " + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{l_t}^t" + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": " obtained via " + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "inline_equation", + "content": "\\Theta_f^t" + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": " is fed into task classifier " + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "inline_equation", + "content": "\\Theta_c^t" + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "inline_equation", + "content": "K^o" + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": " old classes and " + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "inline_equation", + "content": "K^t" + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": " new classes. Given a mini-batch " + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b\\in \\mathcal{T}^t\\cup \\mathcal{M}" + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": ", the classification loss " + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 306, + 340, + 547, + 392 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 396, + 546, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 396, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 351, + 396, + 546, + 429 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\mathcal {D} _ {\\mathrm {C E}} \\left(\\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right), \\mathbf {y} _ {i} ^ {t}\\right), \\tag {5}", + "image_path": "26d7c83fec918b20501d443324d4eca1da51404d19a0b3c4e65d1e3af90a3ed3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}" + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "text", + "content": " is softmax probability predicted by " + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "text", + "content": "-th learning task. " + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{CE}}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "text", + "content": " indicates traditional cross-entropy loss, and " + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 435, + 547, + 472 + ], + "type": "text", + "content": " is the batch size." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "text", + "content": "However, the severe class imbalance (i.e., " + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\frac{N^t}{K^t} \\gg \\frac{|\\mathcal{M}|}{K^o}" + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "text", + "content": ") among old and new categories in the " + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "text", + "content": "-th task enforces the prediction " + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^t(\\mathbf{x}_i^t, \\Theta^t)" + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "text", + "content": " in Eq. (5) to suffer from catastrophic forgetting [40] on old categories. Moreover, the gradient optimization of Eq. (5) neglects heterogeneous forgetting speeds of easy-to-forget old classes with various appearances and hard-to-forget old classes with easily-distinguished attributes. To address heterogeneous forgetting from gradient aspect, we propose two novel plug-and-play losses: a gradient-balanced forgetting compensation (GFC) loss " + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "text", + "content": " in Section 3.3 and a gradient-balanced relation distillation (GRD) loss " + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 304, + 472, + 547, + 617 + ], + "type": "text", + "content": " in Section 3.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 540, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 540, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 540, + 636 + ], + "type": "text", + "content": "3.3. Gradient-Balanced Forgetting Compensation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "As incremental tasks arrive consecutively, easy-to-forget and hard-to-forget old categories change dynamically, thus aggravating difficulty to tackle heterogeneous forgetting from gradient aspect. In light of this, we develop a gradient-balanced forgetting compensation (GFC) loss " + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": ". It adaptively balances large forgetting heterogeneity of old classes" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11745" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": "via performing balanced gradient propagation. Specifically, we perform task-adaptive gradient normalization for different classes, and then reweight classification loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " in Eq. (5) to compensate imbalanced gradients. For the " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": "-th incremental task, as claimed in [48, 13], the gradient " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\Gamma_i^t" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " of an given image " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i^t,\\mathbf{y}_i^t)\\subset \\mathcal{T}^t\\cup \\mathcal{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " with respect to the " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "(k = \\arg \\max \\mathbf{y}_i^t)" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " neuron " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_k^t" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " of classifier " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\Theta_c^t" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 162, + 288, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 162, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 162, + 288, + 190 + ], + "type": "interline_equation", + "content": "\\Gamma_ {i} ^ {t} = \\frac {\\partial \\mathcal {D} _ {\\mathrm {C E}} (\\mathbf {P} ^ {t} (\\mathbf {x} _ {i} ^ {t} , \\boldsymbol {\\Theta} ^ {t}) , \\mathbf {y} _ {i} ^ {t})}{\\partial \\mathcal {N} _ {k} ^ {t}} = \\mathbf {P} ^ {t} (\\mathbf {x} _ {i} ^ {t}, \\boldsymbol {\\Theta} ^ {t}) _ {k} - 1, \\quad (6)", + "image_path": "fb107888927ef52f565058bbe8ffa8ac7967b52424c9f1a10eb6df903ab13257.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 195, + 274, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 274, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 274, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 195, + 274, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)_k" + }, + { + "bbox": [ + 46, + 195, + 274, + 209 + ], + "type": "text", + "content": " denotes probability of the " + }, + { + "bbox": [ + 46, + 195, + 274, + 209 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 195, + 274, + 209 + ], + "type": "text", + "content": "-th class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "text", + "content": "To compensate heterogeneous forgetting of easy-to-forget and hard-to-forget old categories, we compute task-adaptive gradient means for different tasks. Consequently, given a mini-batch " + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b" + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "text", + "content": "-th learning task, the task-adaptive gradient mean " + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\eta}" + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "text", + "content": " for the categories learned in the " + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "inline_equation", + "content": "(1\\leq \\eta \\leq t)" + }, + { + "bbox": [ + 46, + 209, + 288, + 280 + ], + "type": "text", + "content": " incremental task is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 285, + 288, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 285, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 82, + 285, + 288, + 315 + ], + "type": "interline_equation", + "content": "\\Gamma_ {\\eta} = \\frac {1}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}} \\sum_ {i = 1} ^ {b} | \\Gamma_ {i} ^ {t} | \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}. \\tag {7}", + "image_path": "934cd840a8cfea27e614653d9060c434cbc494e51465b2e2a8a681d60ba9d949.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "content": "The forgetting heterogeneity of old categories learned from " + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "content": " old tasks and learning speeds of new categories in the " + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "content": "-th task can be effectively measured via " + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_{\\eta}\\}_{\\eta = 1}^{t}" + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "content": " in Eq. (7)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 356, + 288, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 288, + 440 + ], + "type": "text", + "content": "When easy-to-forget and hard-to-remember old classes are varied dynamically as incremental tasks, some noisy predictions on easy-to-remember old classes may result in large deviation to measure their forgetting heterogeneity [43, 19]. To tackle this issue, " + }, + { + "bbox": [ + 46, + 356, + 288, + 440 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_{\\eta}\\}_{\\eta = 1}^{t}" + }, + { + "bbox": [ + 46, + 356, + 288, + 440 + ], + "type": "text", + "content": " in Eq. (7) are expected to be sharper and more distinguishable adaptively [13], as learning tasks arrive continually. Thus, we rewrite Eq. (7) as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 445, + 287, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 445, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 79, + 445, + 287, + 482 + ], + "type": "interline_equation", + "content": "\\Gamma_ {\\eta} ^ {s} = \\frac {\\sum_ {i = 1} ^ {b} \\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right) \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}}. \\tag {8}", + "image_path": "4a003f99f647b51aa84e983f4038b7418ef492bd0a4ddfd2ba297e74fcd9b904.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": "The sharper task-adaptive gradient means " + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}" + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": " computed via Eq. (8) are employed to reweight classification loss " + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": " in Eq. (5). Then the gradient-balanced forgetting compensation (GFC) loss " + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": " is written as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 541, + 287, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 541, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 287, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F C}} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\frac {\\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right)}{\\sum_ {\\eta = 1} ^ {t} \\Gamma_ {\\eta} ^ {s} \\cdot \\mathbb {I} _ {\\mathbf {y} _ {i} ^ {t} \\in \\mathcal {Y} ^ {\\eta}}} \\mathcal {D} _ {\\mathrm {C E}} \\left(\\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right), \\mathbf {y} _ {i} ^ {t}\\right). \\tag {9}", + "image_path": "cc3d2fc89e5a438c6b4532151c7bdb793629b07065660582879a5c774b7029c4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "text", + "content": "Obviously, " + }, + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "text", + "content": " in Eq. (9) encourages the optimization of " + }, + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "text", + "content": " to perform gradient-balanced propagation for easy-to- forget, hard-to- forget old classes and new classes learned in different tasks. It tackles heterogeneous forgetting on old classes from gradient aspect, via normalizing forgetting paces of different old classes adaptively with " + }, + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}" + }, + { + "bbox": [ + 46, + 580, + 287, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 659, + 257, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 257, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 257, + 671 + ], + "type": "text", + "content": "3.4. Gradient-Balanced Relation Distillation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "The relationships among old and new categories remain constant in semantic space, regardless of availability of old-class samples. In light of this, distilling inter-class relations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "from old model " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\Theta^{t-1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " to new model " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " can address forgetting on old classes [6, 25]. However, most knowledge distillation methods used in CIL [40, 2, 50] only utilize prediction of an individual sample to perform semantic consistency of old classes between " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\Theta^{t-1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ". They neglect underlying relations among old and new categories to tackle forgetting, and heavily suffer from noisy predictions on easy-to-forget old classes when distilling knowledge via an individual sample. Moreover, different forgetting speeds of old classes may result in heterogeneous class-relation consistency between " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\Theta^{t-1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ". It enforces gradient optimization of " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " to bias towards some specific relation distillations related to new classes, causing forgetting heterogeneity on old categories." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "text", + "content": "To overcome the above issues, we design a gradient-balanced relation distillation (GRD) loss " + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "text", + "content": " to distill relations among old and new classes from " + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "text", + "content": ". As introduced in Fig. 2, it can address heterogeneous forgetting from gradient aspect via rectifying imbalanced gradient propagation caused by heterogeneous class-relation consistency. Specifically, we construct class-wise relation prototype rather than a prediction of single sample to perform relation distillation, which can alleviate negative effect of noisy relation consistency. Then the task-adaptive gradient means " + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_{\\eta}^{s}\\}_{\\eta = 1}^{t}" + }, + { + "bbox": [ + 304, + 228, + 547, + 370 + ], + "type": "text", + "content": " in Eq. (8) are employed to reweight heterogeneous class-relation distillation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": "A mini-batch " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^t,\\mathbf{y}_i^t\\}_{i = 1}^b\\subset \\mathcal{T}^t\\cup \\mathcal{M}" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " is fed into " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " to obtain softmax probabilities " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o}" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " of old classes and " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^t (\\mathbf{x}_i^t,\\Theta^t)\\in \\mathbb{R}^{K^o +K^t}" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " of old and new classes, as shown in Fig. 2. The first " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "K^o" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " dimensions of " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^t\\in \\mathbb{R}^{K^o +K^t}" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " are replaced with " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{t - 1}(\\mathbf{x}_i^t,\\Theta^{t - 1})" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " to get relation groundtruth " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^t (\\mathbf{x}_i^t,\\Theta^{t - 1})\\in \\mathbb{R}^{K^o +K^t}" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " that implies underlying relations among old and new classes. To tackle noisy class relations, we construct class-wise relation prototype " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{P}}_k^t" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " and its relation groundtruth " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Y}}_k^t" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 370, + 547, + 483 + ], + "type": "text", + "content": " -th class:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 487, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 487, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 328, + 487, + 545, + 520 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {P}} _ {k} ^ {t} = \\frac {1}{\\Delta_ {k}} \\sum_ {i = 1} ^ {b} \\mathbf {P} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t}\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}, \\tag {10}", + "image_path": "ee1bccacdc847f67d20f3bd72720f5c3f937c759528193415ca2a331c753e1f1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 327, + 522, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 327, + 522, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {Y}} _ {k} ^ {t} = \\frac {1}{\\Delta_ {k}} \\sum_ {i = 1} ^ {b} \\mathbf {Y} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}, \\Theta^ {t - 1}\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}. \\tag {11}", + "image_path": "f5405ee7a0d4901b12759f20547fafd711f286601312e617667d006862b2200c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "inline_equation", + "content": "\\Delta_k = \\sum_i^b\\mathbb{I}_{\\mathrm{argmax}\\mathbf{y}_i^t = k}" + }, + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "text", + "content": ". Then we formulate category-wise gradient mean " + }, + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "inline_equation", + "content": "\\Gamma_k^s" + }, + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 562, + 547, + 588 + ], + "type": "text", + "content": "-th category as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 593, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 593, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 318, + 593, + 545, + 628 + ], + "type": "interline_equation", + "content": "\\Gamma_ {k} ^ {s} = \\frac {\\sum_ {i = 1} ^ {b} \\log \\left(\\left| \\Gamma_ {i} ^ {t} \\right| ^ {\\frac {K ^ {o}}{K ^ {o} + K ^ {t}}} + 1\\right) \\cdot \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}}{\\sum_ {i = 1} ^ {b} \\mathbb {I} _ {\\arg \\max \\mathbf {y} _ {i} ^ {t} = k}}. \\tag {12}", + "image_path": "36e727e9d071255a33c2a05efceca60e6bc6edc20225cba45a6c4d673c4a23b9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 633, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 547, + 657 + ], + "type": "text", + "content": "Consequently, we use " + }, + { + "bbox": [ + 304, + 633, + 547, + 657 + ], + "type": "inline_equation", + "content": "\\Gamma_k^s" + }, + { + "bbox": [ + 304, + 633, + 547, + 657 + ], + "type": "text", + "content": " to reweight heterogeneous relation distillation, and express the proposed " + }, + { + "bbox": [ + 304, + 633, + 547, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 304, + 633, + 547, + 657 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 324, + 662, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 662, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 324, + 662, + 545, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R D}} = \\frac {1}{K ^ {o} + K ^ {t}} \\sum_ {k = 1} ^ {K ^ {o} + K ^ {t}} \\frac {\\Gamma_ {k} ^ {s} \\cdot \\mathcal {D} _ {\\mathrm {K L}} \\left(\\tilde {\\mathbf {P}} _ {k} ^ {t} , \\tilde {\\mathbf {Y}} _ {k} ^ {t}\\right)}{\\sum_ {\\eta = 1} ^ {t} \\Gamma_ {\\eta} ^ {s} \\cdot \\mathbb {I} _ {k \\in \\mathcal {Y} ^ {\\eta}}}, \\tag {13}", + "image_path": "9a11a4a3f1a0a21a43b50e11642b8bacc7947778a9007bf6aee5fedc3f57344a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 701, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{KL}}(\\cdot ||\\cdot)" + }, + { + "bbox": [ + 304, + 701, + 545, + 714 + ], + "type": "text", + "content": " indicates the Kullback-Leibler divergence." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11746" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 82, + 545, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "text", + "content": "Table 1. Results on CIFAR-100 [33]. When " + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = 2000" + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 20, 25, 50\\}" + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 25, 50\\}" + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 50\\%" + }, + { + "bbox": [ + 47, + 70, + 544, + 81 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 82, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 49, + 82, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 82, + 545, + 217 + ], + "type": "table", + "html": "
Comparison MethodsBackbone#ParamsB=0%B=50%
510202550Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M83.580.678.577.676.479.3↑5.580.678.675.471.076.4↑2.9
BiC [50] (CVPR'2019)ViT-Base85.10M83.582.578.780.378.480.7↑4.278.275.774.271.975.0↑4.3
PODNet [17] (ECCV'2020)ViT-Base85.10M75.371.866.564.963.068.3↑16.672.873.074.274.973.7↑5.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.682.079.179.177.380.2↑4.675.874.872.972.073.9↑5.4
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M75.272.066.165.263.868.5↑16.472.772.874.074.773.5↑5.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M84.783.579.678.577.980.8↑4.178.477.376.274.076.5↑2.8
AFC [32] (CVPR'2022)ViT-Base85.10M74.675.570.367.464.770.5↑14.470.869.468.969.969.8↑9.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.585.881.380.877.482.2↑2.782.581.077.374.578.8↑0.5
HFC (Ours)ViT-Base85.10M86.486.385.585.081.184.9-82.981.077.875.579.3-
Upper BoundViT-Base85.10M94.294.294.294.294.294.2-94.294.294.294.294.2-
", + "image_path": "a79693cd96870af66706ac5b32e10429aa7f3eb2780cc6bb123d4bdbc0da2508.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 231, + 545, + 367 + ], + "blocks": [ + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "text", + "content": "Table 2. Results on ImageNet-100 [10]. When " + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = {2000}" + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "text", + "content": " ,we set " + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "inline_equation", + "content": "T = \\{ 5,{10},{20},{25}\\}" + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "inline_equation", + "content": "T = \\{ 5,{10},{25},{50}\\}" + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = {50}\\%" + }, + { + "bbox": [ + 48, + 218, + 545, + 229 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 231, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 50, + 231, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 50, + 231, + 545, + 367 + ], + "type": "table", + "html": "
Comparison MethodsBackbone#ParamsB=0%B=50%
5102025Avg.Imp.5102550Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M84.182.480.278.681.3↑3.279.578.876.673.977.2↑3.4
BiC [50] (CVPR'2019)ViT-Base85.10M81.480.279.278.579.8↑4.773.974.473.472.273.5↑7.1
PODNet [17] (ECCV'2020)ViT-Base85.10M82.480.179.578.380.1↑4.568.471.375.477.673.2↑7.4
SS-IL [2] (ICCV'2021)ViT-Base85.10M83.081.780.979.481.2↑3.375.874.872.972.073.9↑6.7
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M82.381.079.379.080.4↑4.168.371.275.076.972.8↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M85.684.083.883.184.1↑0.481.080.580.079.880.3↑0.3
AFC [32] (CVPR'2022)ViT-Base85.10M83.281.280.579.481.1↑3.577.178.478.978.178.1↑2.5
DyTox [18] (CVPR'2022)ViT-Base85.10M85.183.480.080.382.2↑2.377.478.479.779.478.7↑1.9
HFC (Ours)ViT-Base85.10M86.185.484.082.684.5-81.581.180.379.580.6-
Upper BoundViT-Base85.10M86.386.386.386.386.3-86.386.386.386.386.3-
", + "image_path": "70ce29fab9fc2e7474abe7eefd2a0be36bd632ede8b844a6becb4c60e2904f25.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 396, + 285, + 494 + ], + "blocks": [ + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "lines": [ + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "type": "text", + "content": "Table 3. Results on ImageNet-1000 [10]. When " + }, + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = 20000" + }, + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "type": "inline_equation", + "content": "T = \\{5,10\\}" + }, + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 47, + 374, + 286, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 396, + 285, + 494 + ], + "lines": [ + { + "bbox": [ + 50, + 396, + 285, + 494 + ], + "spans": [ + { + "bbox": [ + 50, + 396, + 285, + 494 + ], + "type": "table", + "html": "
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M75.472.473.9↑3.5
BiC [50] (CVPR'2019)ViT-Base85.10M68.966.067.5↑10.0
PODNet [17] (ECCV'2020)ViT-Base85.10M68.471.369.8↑7.6
SS-IL [2] (ICCV'2021)ViT-Base85.10M75.874.875.3↑2.2
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M68.463.265.8↑11.6
FOSTER [47] (ECCV'2022)ViT-Base85.10M76.974.375.6↑1.9
AFC [32] (CVPR'2022)ViT-Base85.10M74.271.472.8↑4.6
DyTox [18] (CVPR'2022)ViT-Base85.10M78.174.776.4↑1.0
HFC (Ours)ViT-Base85.10M78.576.477.5-
Upper BoundViT-Base85.10M86.386.386.3-
", + "image_path": "4cf56208686614c3fa72b24b1f6c7eb5d16b5b158ee5254d944f4faa78cb94c0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 503, + 173, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 173, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 173, + 516 + ], + "type": "text", + "content": "3.5. Optimization Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "text", + "content": "Overall, in the " + }, + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "inline_equation", + "content": "(t \\geq 2)" + }, + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "text", + "content": " learning task, the objective formulation to optimize " + }, + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 47, + 522, + 287, + 546 + ], + "type": "text", + "content": " is expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 112, + 557, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 557, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 112, + 557, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o b j}} = \\alpha_ {1} \\mathcal {L} _ {\\mathrm {F C}} + \\alpha_ {2} \\mathcal {L} _ {\\mathrm {R D}}, \\tag {14}", + "image_path": "5b91271300310eb7c86181036e034b35e1adac04f2067fa4b6e72af5ec638089.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2}" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " are hyper-parameters. " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " along with learnable " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_0^t" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " are optimized via " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " in Eq. (5) for the first task, and trained via " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " in Eq. (14) when " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "t\\geq 2" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": ". After optimizing new model " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": ", we store " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\Theta^t" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " as the frozen old model " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\Theta^{t - 1}" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " to perform the GRD loss " + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 47, + 578, + 287, + 639 + ], + "type": "text", + "content": " for the next incremental task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 670, + 180, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 180, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 180, + 683 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "For fair comparisons, we follow the same protocols utilized in baseline CIL methods [40, 50, 17, 17, 2, 18, 32] to" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": "set incremental tasks (class order), and do experiments on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Specifically, we consider two task settings in this paper: 1) We divide all categories of each dataset into " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " tasks equally [40, 18] (i.e., the base class set " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": "); 2) We start by training our model on half of the categories, while divide the rest of categories into " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " tasks equally [17, 32] (i.e., the base class set " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 50\\%" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": "). When " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 20, 25, 50\\}" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " for CIFAR-100 [33], " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 20, 25\\}" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " for ImageNet-100 [10] and " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "T = \\{5, 10\\}" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " for ImageNet-1000 [10]. When " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 50\\%" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 25, 50\\}" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " for CIFAR-100 [33] and ImageNet-100 [10]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "As introduced in [40, 50], the size of memory " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " for all comparison methods is fixed as 2,000 for CIFAR-100 [33] and ImageNet-100 [10], and 20,000 for ImageNet-1000 [10]. We follow iCaRL [40] to update memory " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " and use global memory for [18]. Besides, we employ the same data augmentation proposed in DyTox [18] for all comparison methods, but don't use mixup technology. Following [17, 32], we also consider another memory setting (i.e., storing 20 exemplars per class), and set " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "T = \\{5,10\\}, \\mathcal{B} = 50\\%" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " for comparisons on CIFAR-100 [33] (see Tab. 6). For network architecture, we use ViT-Base [15] as feature extractor containing " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "l_{m} = 11" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " MSA blocks and " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "l_{t} = 1" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " TSA block, where the parameters are initialized via [22]. The task classifier includes only one fully-connected layer. The SGD optimizer is employed to optimize our model, where the learning rate is initialized as " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "6.25 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": ". We apply two plug-and-play losses (i.e., " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11747" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 95, + 547, + 327 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 4. Performance on CIFAR-100 [33] (" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 20, 25, 50\\}" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": ") and ImageNet-100 [10] (" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 20, 25\\}" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": ") when we apply " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}^{\\ddagger}" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": " into existing distillation-based CIL methods and set " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = 2000" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}^{\\ddagger}" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": " denotes the proposed plug-and-play losses " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 95, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 49, + 95, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 95, + 547, + 327 + ], + "type": "table", + "html": "
Comparison MethodsCIFAR-100 Dataset [33]ImageNet-100 Dataset [10]
Backbone#Params510202550Avg.Backbone#Params5102025
iCaRL [40] (CVPR'2017)ResNet-320.46M66.463.953.150.239.154.5ResNet-1811.22M76.469.560.657.9
iCaRL [40] + Ours‡ResNet-320.46M68.065.754.251.440.856.0ResNet-1811.22M76.870.361.759.2
BiC [50] (CVPR'2019)ResNet-320.46M55.350.848.346.439.348.0ResNet-1811.22M61.554.346.744.5
BiC [50] + Ours‡ResNet-320.46M63.253.852.449.341.552.0ResNet-1811.22M63.257.050.348.4
PODNet [17] (ECCV'2020)ResNet-320.46M63.854.548.144.439.750.1ResNet-1811.22M74.966.555.851.7
PODNet [17] + Ours‡ResNet-320.46M65.860.051.648.344.754.1ResNet-1811.22M76.569.359.455.8
SS-IL [2] (ICCV'2021)ResNet-320.46M65.159.049.048.640.752.5ResNet-1811.22M60.953.539.335.8
SS-IL [2] + Ours‡ResNet-320.46M67.764.951.651.242.555.6ResNet-1811.22M75.267.557.956.0
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M63.754.347.844.039.049.8ResNet-1811.22M74.866.355.451.3
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M66.758.550.546.943.353.2ResNet-1811.22M76.468.758.955.1
FOSTER [47] (ECCV'2022)ResNet-320.46M72.068.767.166.062.667.3ResNet-1811.22M74.970.967.265.7
FOSTER [47] + Ours‡ResNet-320.46M72.269.967.466.762.867.8ResNet-1811.22M75.071.367.566.2
AFC [32] (CVPR'2022)ResNet-320.46M65.657.150.948.142.052.7ResNet-1811.22M76.369.761.158.1
AFC [32] + Ours‡ResNet-320.46M66.263.853.249.744.155.4ResNet-1811.22M77.471.263.367.4
Upper BoundResNet-320.46M76.676.676.676.676.676.6ResNet-1811.22M73.273.273.273.2
DyTox [18] (CVPR'2022)ViT-Tiny10.71M75.373.572.772.370.572.9ViT-Tiny10.71M77.875.472.972.3
DyTox [18] + Ours‡ViT-Tiny10.71M75.974.072.972.670.873.2ViT-Tiny10.71M77.975.673.172.9
Upper BoundViT-Tiny10.71M76.176.176.176.176.176.1ViT-Tiny10.71M79.179.179.179.1
", + "image_path": "996b0fdb9f3bdc39d095447b2144fb662271d773e2a8c936b43b98e2128d083e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 369, + 287, + 468 + ], + "blocks": [ + { + "bbox": [ + 47, + 335, + 287, + 368 + ], + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 368 + ], + "type": "text", + "content": "Table 5. Ablation studies on CIFAR-100 [33] (the top block) and ImageNet-100 [10] (the bottom block), when we set " + }, + { + "bbox": [ + 47, + 335, + 287, + 368 + ], + "type": "inline_equation", + "content": "T = \\{5,10,25\\}" + }, + { + "bbox": [ + 47, + 335, + 287, + 368 + ], + "type": "text", + "content": " under the settings of " + }, + { + "bbox": [ + 47, + 335, + 287, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = 2000, \\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 47, + 335, + 287, + 368 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 369, + 287, + 468 + ], + "lines": [ + { + "bbox": [ + 50, + 369, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 50, + 369, + 287, + 468 + ], + "type": "table", + "html": "
Ablation VariantsBackbone#ParamsTSAGFCGRD51025
BaselineViT-Base85.10MXXX81.680.278.1
Baseline + TSAViT-Base85.10M✓XX83.182.481.0
Baseline + TSA + GFCViT-Base85.10M✓✓X84.083.782.1
HFC (Ours)ViT-Base85.10M✓✓✓86.486.385.0
BaselineViT-Base85.10MXXX82.080.578.8
Baseline + TSAViT-Base85.10M✓XX83.782.881.2
Baseline + TSA + GFCViT-Base85.10M✓XX84.383.482.0
HFC (Ours)ViT-Base85.10M✓✓✓86.185.484.0
", + "image_path": "683c4ef6ce1890edb08206d5e2fe743b2d6ca4bf633accec72568fa18a784b17.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}})" + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": " to existing CIL methods [40, 18] using ResNet-18 [23], ResNet-32 [23] or ViT-Tiny [15] as backbone. For evaluation, top-1 accuracy is used to compare performances of our model and baseline methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 538, + 190, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 190, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 190, + 551 + ], + "type": "text", + "content": "4.2. Comparison Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "content": "Comparison Results: Tabs. 1-3 present comparison results of our HFC model and off-the-shelf CIL methods on CIFAR-100 [33], ImageNet-100 [10] and ImageNet-1000 [10]. Our model significantly outperforms existing CIL methods [40, 50, 17, 2, 18, 32, 47] by " + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "inline_equation", + "content": "0.7\\% \\sim 16.6\\%" + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "content": " accuracy under various incremental settings, when we use the same backbone (i.e., ViT-Base [15]) as feature extractor for fair comparisons. Such large performance improvement illustrates the superiority of our model to address forgetting heterogeneity among easy-to-forget and hard-to-forget old categories from both representation and gradient aspects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": "Improvement of Plug-and-Play Losses: Tab. 4 shows large performance improvement of existing knowledge" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "text", + "content": "based CIL methods [40, 50, 17, 4, 18, 32, 47], when we apply the proposed plug-and-play losses (i.e., " + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "text", + "content": ") to them. The proposed losses " + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 304, + 338, + 547, + 411 + ], + "type": "text", + "content": " help existing CIL methods to address heterogeneous forgetting from gradient aspect, thus largely improving their performance. Besides, it validates generalization and robustness of our HFC model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 418, + 406, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 406, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 406, + 430 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "text", + "content": "As shown in Tab. 5, we present ablation studies on CIFAR-100 [33] and ImageNet-100 [10] to investigate the effectiveness of each module. TSA, GFC and GRD are task-semantic aggregation blocks, the gradient-balanced forgetting compensation loss " + }, + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "text", + "content": " and gradient-balanced relation distillation loss " + }, + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 304, + 436, + 547, + 581 + ], + "type": "text", + "content": ". Baseline denotes the performance of our model using the traditional classification loss and knowledge distillation loss proposed in DyTox (CVPR'2022) [18], but without the TSA, GFC and GRD modules. Compared with Ours, the performance of all ablation variants degrades significantly. It validates superiority of each module to overcome forgetting heterogeneity of different old categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 586, + 497, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 497, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 497, + 601 + ], + "type": "text", + "content": "4.4. Analysis of Task-Wise Comparisons" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Figs. 3-4 present task-wise performance comparisons between our model and other CIL methods [40, 50, 17, 18, 32] when we set backbone as ViT-Base and different tasks as " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "T = \\{5, 10, 20, 25\\}" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": ". Our HFC model outperforms baseline methods for most task-wise comparisons, which illustrates the superior performance of our model to surmount heterogeneous forgetting from representation and gradient aspects. To identify new classes continually, the task-semantic aggregation (TSA) block explores task-shared global representa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11748" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 544, + 190 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 544, + 190 + ], + "type": "image", + "image_path": "64f4d3f7af5d54ecb9432c85770b0132db53aa3f32f9f0dc5cdad2675c18a6e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 191, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 191, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 547, + 201 + ], + "type": "text", + "content": "Figure 3. Analysis of task-wise performance comparisons on CIFAR-100 [33]. We set " + }, + { + "bbox": [ + 47, + 191, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = 2000" + }, + { + "bbox": [ + 47, + 191, + 547, + 201 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 191, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 47, + 191, + 547, + 201 + ], + "type": "text", + "content": " when the backbone is ViT-Base." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 203, + 545, + 321 + ], + "blocks": [ + { + "bbox": [ + 50, + 203, + 545, + 321 + ], + "lines": [ + { + "bbox": [ + 50, + 203, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 50, + 203, + 545, + 321 + ], + "type": "image", + "image_path": "c4a3ab5333a42f77c780318b84a9bf1e3211844f8753c898d288195032a9274e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 323, + 530, + 334 + ], + "lines": [ + { + "bbox": [ + 59, + 323, + 530, + 334 + ], + "spans": [ + { + "bbox": [ + 59, + 323, + 530, + 334 + ], + "type": "text", + "content": "Figure 4. Analysis of task-wise performance on ImageNet-100 [10]. We set " + }, + { + "bbox": [ + 59, + 323, + 530, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = 2000" + }, + { + "bbox": [ + 59, + 323, + 530, + 334 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 59, + 323, + 530, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 59, + 323, + 530, + 334 + ], + "type": "text", + "content": " when the backbone is ViT-Base." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 374, + 286, + 589 + ], + "blocks": [ + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "text", + "content": "Table 6. Results on CIFAR-100 [33] when storing 20 exemplars for each class. We set " + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "inline_equation", + "content": "T = \\{5,10\\}" + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = 50\\%" + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}^{\\ddagger}" + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "text", + "content": " denotes the proposed plug-and-play losses " + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 46, + 341, + 287, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 374, + 286, + 589 + ], + "lines": [ + { + "bbox": [ + 50, + 374, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 50, + 374, + 286, + 589 + ], + "type": "table", + "html": "
Comparison MethodsBackbone#Params510Avg.Imp.
iCaRL [40] (CVPR'2017)ViT-Base85.10M78.876.277.5↑1.9
BiC [50] (CVPR'2019)ViT-Base85.10M75.374.574.9↑4.5
PODNet [17] (ECCV'2020)ViT-Base85.10M71.671.671.6↑7.8
SS-IL [2] (ICCV'2021)ViT-Base85.10M73.071.872.4↑7.0
PODNet [17] + CSCCT [4] (ECCV'2022)ViT-Base85.10M71.571.771.6↑7.8
FOSTER [47] (ECCV'2022)ViT-Base85.10M78.878.778.8↑0.6
AFC [32] (CVPR'2022)ViT-Base85.10M68.265.366.8↑12.6
DyTox [18] (CVPR'2022)ViT-Base85.10M79.378.478.8↑0.6
HFC (Ours)ViT-Base85.10M79.878.979.4-
Upper BoundViT-Base85.10M94.294.294.2-
iCaRL [40] (CVPR'2017)ResNet-320.46M56.653.254.9↑0.7
iCaRL [40] + Ours‡ResNet-320.46M57.353.955.6-
PODNet [17] + CSCCT [4] (ECCV'2022)ResNet-320.46M62.261.161.7↑0.6
PODNet [17] + CSCCT [4] + Ours‡ResNet-320.46M62.961.662.3-
FOSTER [47] (ECCV'2022)ResNet-320.46M67.366.466.8↑0.5
FOSTER [47] + Ours‡ResNet-320.46M67.766.667.2-
AFC [32] (CVPR'2022)ResNet-320.46M63.863.663.7↑0.7
AFC [32] + Ours‡ResNet-320.46M64.464.364.4-
Upper BoundResNet-320.46M76.676.676.6-
DyTox [18] (CVPR'2022)ViT-Tiny10.71M67.966.267.1↑0.5
DyTox [18] + Ours‡ViT-Tiny10.71M68.566.767.6-
Upper BoundViT-Tiny10.71M76.176.176.1-
", + "image_path": "5562063f001a7c22c7b930ababbc4b11890dcc65e0cc41a8dee95320ae2ac60c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 601, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 288, + 625 + ], + "type": "text", + "content": "tions to alleviate forgetting from representation aspect, while " + }, + { + "bbox": [ + 47, + 601, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FC}}" + }, + { + "bbox": [ + 47, + 601, + 288, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 601, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RD}}" + }, + { + "bbox": [ + 47, + 601, + 288, + 625 + ], + "type": "text", + "content": " can achieve gradient-balanced compensation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 241, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 241, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 241, + 646 + ], + "type": "text", + "content": "4.5. Analysis of Forgetting Heterogeneity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "text", + "content": "Tab. 7 presents forgetting heterogeneity (FH) of different old classes via measuring their variance of gradient updating. We formulate forgetting heterogeneity (FH) as " + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FH} = \\frac{1}{T}\\sum_{t=1}^{T}\\left(\\frac{1}{S_t}\\sum_{i=1}^{S_t}(|\\Gamma_i^t| - \\sum_{\\eta=1}^t\\Gamma_\\eta \\cdot \\mathbb{I}_{\\mathbf{y}_i^t \\in \\mathcal{Y}^\\eta})^2\\right)" + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma_i^t" + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma_\\eta" + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "text", + "content": " are obtained from Eqs. (6)-(7), and " + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 46, + 651, + 289, + 715 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 363, + 539, + 462 + ], + "blocks": [ + { + "bbox": [ + 305, + 341, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 305, + 341, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 545, + 363 + ], + "type": "text", + "content": "Table 7. Analysis of forgetting heterogeneity (FH) on CIFAR-100 [33] when " + }, + { + "bbox": [ + 305, + 341, + 545, + 363 + ], + "type": "inline_equation", + "content": "T = 5, \\mathcal{B} = 0\\%" + }, + { + "bbox": [ + 305, + 341, + 545, + 363 + ], + "type": "text", + "content": ". AC denotes the averaged accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 363, + 539, + 462 + ], + "lines": [ + { + "bbox": [ + 312, + 363, + 539, + 462 + ], + "spans": [ + { + "bbox": [ + 312, + 363, + 539, + 462 + ], + "type": "table", + "html": "
M=2000,B=0%BackboneTSAGFCGRDFHAC
iCaRL [40] (CVPR'2017)ViT-Base---131.583.5
FOSTER [47] (ECCV'2022)ViT-Base---119.485.1
AFC [32] (CVPR'2022)ViT-Base---108.574.6
DyTox [18] (CVPR'2022)ViT-Base---105.185.5
BaselineViT-BaseXXX128.981.6
Baseline + TSAViT-Base✓XX125.483.1
Baseline + TSA + GFCViT-Base✓✓X120.584.0
HFC (Ours)ViT-Base✓✓✓103.686.4
", + "image_path": "506a6aa3d7d9b0d5e1533bb9f70c5ca6949761f0bbef6093d13907414c091ea8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 547, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 522 + ], + "type": "text", + "content": "number of samples in test set. Tab. 7 verifies superiority of our model against other CIL methods to tackle heterogeneous forgetting. It also shows effectiveness of all modules to collaboratively minimize forgetting heterogeneity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 536, + 378, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 378, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 378, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we develop a novel Heterogeneous Forgetting Compensation (HFC) model to surmount heterogeneous forgetting from representation and gradient aspects. To be specific, a task-semantic aggregation block is designed to tackle heterogeneous forgetting from representation aspect via exploring task-shared global representations. Moreover, we propose a gradient-balanced forgetting compensation loss and a relation distillation loss to compensate forgetting heterogeneity from gradient aspect via performing balanced gradient propagation and distilling heterogeneous class relations. Experiments verify the superiority of our proposed HFC model against baseline methods. We will further consider addressing noisy forgetting heterogeneity in the future." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11749" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, and Babak Ehteshami Bejnordi. Conditional channel gated networks for task-aware continual learning. In CVPR, pages 3930-3939, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 179 + ], + "type": "text", + "content": "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In ICCV, pages 844-853, October 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 144-161, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "text", + "content": "[4] Arjun Ashok, K. J. Joseph, and Vineeth N. Balasubramanian. Class-incremental learning with cross-space clustering and controlled transfer. In ECCV, pages 105-122, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "text", + "content": "[5] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-16, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 373 + ], + "type": "text", + "content": "[6] Francisco M. Castro, Manuel J. Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, ECCV, pages 241-257, Cham, 2018. Springer International Publishing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 288, + 418 + ], + "type": "text", + "content": "[7] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 288, + 463 + ], + "type": "text", + "content": "[8] Huitong Chen, Yu Wang, and Qinghua Hu. Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering, pages 1-15, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 465, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 508 + ], + "type": "text", + "content": "[9] Stéphane d'Ascoli, Hugo Touvron, Matthew Leavitt, Ari Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 543 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 599 + ], + "type": "text", + "content": "[11] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 11-20, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "type": "text", + "content": "[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In CVPR, pages 10164-10173, June 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 668 + ], + "type": "text", + "content": "[13] Jiahua Dong, Duzhen Zhang, Yang Cong, Wei Cong, Henghui Ding, and Dengxin Dai. Federated incremental semantic segmentation. In CVPR, pages 3934-3943, June 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[14] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In CVPR, pages 19217-19227, June 2022." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 172 + ], + "type": "text", + "content": "[16] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In CVPR, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "type": "text", + "content": "[17] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "type": "text", + "content": "[18] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285-9295, June 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 263, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 547, + 297 + ], + "type": "text", + "content": "[19] Zhen Fang, Yixuan Li, Jie Lu, Jiahua Dong, Bo Han, and Feng Liu. Is out-of-distribution detection learnable? In NeurIPS, volume 35, pages 37199-37213, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 331 + ], + "type": "text", + "content": "[20] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. In ICML, page 1243-1252, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 332, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 547, + 374 + ], + "type": "text", + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, volume 27, 2014." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 547, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 410 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, June 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 411, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 547, + 443 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 445, + 547, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 547, + 477 + ], + "type": "text", + "content": "[24] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. In NeurIPS Workshop, 2015." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 479, + 547, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 547, + 511 + ], + "type": "text", + "content": "[25] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, June 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 513, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 547, + 567 + ], + "type": "text", + "content": "[26] Weiming Hu, Xi Li, Guodong Tian, Stephen Maybank, and Zhongfei Zhang. An incremental dpmm-based method for trajectory clustering, modeling, and retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(5):1051-1065, 2013." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 547, + 602 + ], + "type": "text", + "content": "[27] Xinting Hu, Kaihua Tang, Chunyan Miao, Xian-Sheng Hua, and Hanwang Zhang. Distilling causal effect of data in class incremental learning. In CVPR, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 602, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 547, + 657 + ], + "type": "text", + "content": "[28] Ching-Yi Hung, Cheng-Hao Tu, Cheng-En Wu, Chien-Hung Chen, Yi-Ming Chan, and Chu-Song Chen. Compacting, picking and growing for unforgetting continual learning. In Advances in Neural Information Processing Systems, pages 13647–13657, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "text", + "content": "[29] Ahmet Iscen, Jeffrey Zhang, Svetlana Lazebnik, and Cordelia Schmid. Memory-efficient incremental learning through feature adaptation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 699-715, 2020." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11750" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, and et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "type": "text", + "content": "[31] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwar, and Vineeth Balasubramanian. Energy-based latent aligner for incremental learning. In CVPR, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 153, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 153, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 153, + 288, + 196 + ], + "type": "text", + "content": "[32] Minsoo Kang, Jaeyoo Park, and Bohyung Han. Class incremental learning by knowledge distillation with adaptive feature consolidation. In CVPR, pages 16071-16080, June 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 288, + 220 + ], + "type": "text", + "content": "[33] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, pages 32-33, 2009." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 222, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 222, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 222, + 288, + 255 + ], + "type": "text", + "content": "[34] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 257, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 257, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 49, + 257, + 288, + 289 + ], + "type": "text", + "content": "[35] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 291, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 291, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 291, + 288, + 323 + ], + "type": "text", + "content": "[36] Yaoyao Liu, Yingying Li, Bernt Schiele, and Qianru Sun. Online hyperparameter optimization for class-incremental learning. AAAI, 37(7):8906-8913, Jun. 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 326, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 288, + 358 + ], + "type": "text", + "content": "[37] Yaoyao Liu, Bernt Schiele, and Qianru Sun. RMM: reinforced memory management for class-incremental learning. In NeurIPS, pages 3478-3490, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 361, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 288, + 382 + ], + "type": "text", + "content": "[38] John M. Pierre. Incremental lifelong deep learning for autonomous vehicles. In ITSC, pages 3949-3954, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 384, + 288, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 288, + 416 + ], + "type": "text", + "content": "[39] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random path selection for continual learning. In NeurIPS, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 418, + 288, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 418, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 49, + 418, + 288, + 451 + ], + "type": "text", + "content": "[40] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. icarl: Incremental classifier and representation learning. In CVPR, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 453, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 453, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 453, + 288, + 495 + ], + "type": "text", + "content": "[41] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In ICML, volume 80, pages 4548-4557, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 498, + 288, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 288, + 530 + ], + "type": "text", + "content": "[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In CVPR, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 533, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 288, + 576 + ], + "type": "text", + "content": "[43] Zhenheng Tang, Yonggang Zhang, Shaohuai Shi, Xin He, Bo Han, and Xiaowen Chu. Virtual homogeneity learning: Defending against data heterogeneity in federated learning. In ICML, pages 21111-21132, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 578, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 288, + 633 + ], + "type": "text", + "content": "[44] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, ECCV, pages 254-270, Cham, 2020. Springer International Publishing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 635, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 288, + 677 + ], + "type": "text", + "content": "[45] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In CVPR, pages 99-108, June 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 288, + 713 + ], + "type": "text", + "content": "[46] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 646 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[47] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "text", + "content": "[48] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In AAAI, volume 35, pages 10165-10173, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 184 + ], + "type": "text", + "content": "[49] Liyuan Wang, Xingxing Zhang, Kuo Yang, Longhui Yu, Chongxuan Li, Lanqing Hong, Shifeng Zhang, Zhenguo Li, Yi Zhong, and Jun Zhu. Memory replay with data compression for continual learning. In ICLR, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 186, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 218 + ], + "type": "text", + "content": "[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "type": "text", + "content": "[51] Ziyang Wu, Christina Baek, Chong You, and Yi Ma. Incremental learning via rate reduction. In CVPR, pages 1125-1133, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 254, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 547, + 297 + ], + "type": "text", + "content": "[52] Rongtao Xu, Changwei Wang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Wave-like class activation map with representation fusion for weakly-supervised semantic segmentation. IEEE Transactions on Multimedia, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 298, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 547, + 352 + ], + "type": "text", + "content": "[53] Rongtao Xu, Changwei Wang, Jiguang Zhang, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. RSSformer: Foreground saliency enhancement for remote sensing land-cover segmentation. IEEE Transactions on Image Processing, 32:1052-1064, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 354, + 547, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 547, + 387 + ], + "type": "text", + "content": "[54] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 388, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 547, + 420 + ], + "type": "text", + "content": "[55] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 422, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 453 + ], + "type": "text", + "content": "[56] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, page 3987-3995, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 456, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 547, + 488 + ], + "type": "text", + "content": "[57] Duzhen Zhang, Feilong Chen, and Xiuyi Chen. Dualgats: Dual graph attention networks for emotion recognition in conversations. In ACL, pages 7395-7408, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 490, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 547, + 533 + ], + "type": "text", + "content": "[58] Yonggang Zhang, Mingming Gong, Tongliang Liu, Gang Niu, Xinmei Tian, Bo Han, Bernhard Schölkopf, and Kun Zhang. Causaladv: Adversarial robustness through the lens of causality. *ICLR*, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 535, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 547, + 567 + ], + "type": "text", + "content": "[59] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13205-13214, 2020." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "type": "text", + "content": "[60] Hanbin Zhao, Yongjian Fu, Mintong Kang, Qi Tian, Fei Wu, and Xi Li. Mgmtf: Multi-grained slow vs. fast framework for few-shot class-incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 613, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 547, + 646 + ], + "type": "text", + "content": "[61] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In ECCV, pages 55-72, 2022." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11751" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_content_list.json b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6276f0234acdc2276505e17353c691e6340add3 --- /dev/null +++ b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_content_list.json @@ -0,0 +1,2023 @@ +[ + { + "type": "text", + "text": "HiFace: High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details", + "text_level": 1, + "bbox": [ + 233, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zenghao Chai $^{1,2*}$ Tianke Zhang $^{2}$ Tianyu He $^{3}$ Xu Tan $^{3\\dagger}$ Tadas Baltrušaitis $^{4}$ HsiangTao Wu $^{5}$ Runnan Li $^{5}$ Sheng Zhao $^{5}$ Chun Yuan $^{2}$ Jiang Bian $^{3}$ $^{1}$ National University of Singapore $^{2}$ Tsinghua University $^{3}$ Microsoft Research Asia \n $^{4}$ Microsoft Mixed Reality & AI Lab $^{5}$ Microsoft Cloud + AI \nzenghaochai@gmail.com ztk21@mails.tsinghua.edu.cn yuanc@sz.tsinghua.edu.cn \n{tianyuhe,xuta,tabaltru,musclewu,runnan.li,sheng.zhao,jiang.bian}@microsoft.com", + "bbox": [ + 133, + 202, + 839, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Morphable Models (3DMMs) demonstrate great potential for reconstructing faithful and animatable 3D facial surfaces from a single image. The facial surface is influenced by the coarse shape, as well as the static detail (e.g., person-specific appearance) and dynamic detail (e.g., expression-driven wrinkles). Previous work struggles to decouple the static and dynamic details through image-level supervision, leading to reconstructions that are not realistic. In this paper, we aim at high-fidelity 3D face reconstruction and propose HiFace to explicitly model the static and dynamic details. Specifically, the static detail is modeled as the linear combination of a displacement basis, while the dynamic detail is modeled as the linear interpolation of two displacement maps with polarized expressions. We exploit several loss functions to jointly learn the coarse shape and fine details with both synthetic and real-world datasets, which enable HiFace to reconstruct high-fidelity 3D shapes with animatable details. Extensive quantitative and qualitative experiments demonstrate that HiFace presents state-of-the-art reconstruction quality and faithfully recovers both the static and dynamic details. Our project page: https://project-hiface.github.io.", + "bbox": [ + 75, + 376, + 473, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 734, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The reconstruction of a 3D face from a single image has drawn much attention recently [67, 21, 41, 81]. It has tremendous potential applications like face recognition [11, 63, 4, 59], face animation [16, 75], virtual reality [7, 58, 31], etc. For example, the reconstructed 3D face representation can be driven by an audio [16], or a video from another person [38].", + "bbox": [ + 75, + 761, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c00cf10e1215744da80064e3797585206678ae0d1ae242aceac0c1dbe52a2ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 343, + 696, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1195b9a15fd2f5a623241916e9f3741486d6c16a08677b268519183095717106.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 343, + 888, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/63b1b1b56817ba4333a72fb19fb4d33d1d40598e6a27343b136c4fa6a3f9f809.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 492, + 694, + 640 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a059dfce8796e1fc9e0a624d78c52c64d3bb7597dcd9d077e6c799d2749577b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 492, + 888, + 640 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4563679efa1a8ab5206d5c5548112ddb398c55259c32d140f515975398ef7922.jpg", + "image_caption": [ + "Figure 1. We propose HiFace to reconstruct high-fidelity 3D face with realistic and animatable details. Reconstruction: given a single image (1st-row), HiFace faithfully reconstructs a coarse shape (2nd-row) with vivid details (3rd-row). Animation: given a source face (yellow box), HiFace can animate the static (4th-row), dynamic (5th-row), or both (6th-row) details of the driving images (green box). Images are taken from FFHQ [37] and CelebA [40]." + ], + "image_footnote": [], + "bbox": [ + 532, + 638, + 663, + 689 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40004fb7ad32dc8431607682c055339918e9b028cdb0635b8ba37585066ce3fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 638, + 857, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To build a flexible and animatable facial representation, a popular way is to leverage the success of 3D Morphable Models (3DMMs) [5, 6, 10, 45, 65], which decouple the influence of shape, expression, albedo, and others by modeling them in separate coefficients. Typically", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done when the author was an intern at MSRA. †Corresponding author: Xu Tan (xuta@microsoft.com)", + "bbox": [ + 94, + 875, + 387, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in literature, one can achieve coarse shape reconstruction in coefficients-fitting optimization [27, 73, 3, 77, 2], or an analysis-by-synthesis pipeline [67, 21, 81, 50]. As 3DMMs typically capture only the coarse facial geometry and are not capable of representing fine details (e.g., wrinkles), recent advances model such details with a displacement map [15, 77, 13, 9, 76]. However, previous work fails to model the distinction between static and dynamic factors of fine detail, leading to errors in reconstructions. For example, given that one may drive the expression of a young man from an old man, trivially transferring all wrinkles from the old man to the young man could make the young man look unnatural. In this sense, Feng et al. [24] implicitly leverages the person-specific identity and expression as conditions to generate the details. Although effective, they optimize the model in an analysis-by-synthesis pipeline with only the image-level supervision, leading to insufficient decoupling of static and dynamic details and inconsistent animation results (see Fig. 7).", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose HiFace to explicitly model the static and dynamic details for high-fidelity 3D face reconstruction, by designing SD-DeTail module to decouple the static and dynamic factors. More specifically, for person-specific static detail, instead of directly predicting the displacement map that may increase the difficulty of detail prediction [24, 19], we follow the spirit of 3DMMs to build a displacement basis from the captured facial scans with age diversity [57, 72]. In this way, the model is trained to predict the coefficients of the displacement basis, and make the detail prediction easier. For dynamic detail, since it is highly expression-dependent, directly modeling it with one displacement basis is quite difficult. Therefore, based on the fact that the expression can be interpolated by a compressed and a stretched expressions [57], we build two displacement bases for the compressed and stretched expressions from the captured scans respectively, and learn to regress the displacement coefficients with the ground-truth labels. Therefore, we can obtain the dynamic detail by linearly interpolating the compressed and stretched displacement maps, which are derived from the displacement bases and the predicted coefficients. Finally, the predicted static and dynamic details are merged with the coarse shape to formulate the final output.", + "bbox": [ + 75, + 382, + 472, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since we would like the final output to contain both the coarse shape and high-frequency detail, we propose several novel loss functions to learn coarse shape and details simultaneously from both the synthetic and real-world datasets. For details, we leverage the ground-truth static and dynamic displacement maps of the synthetic dataset [72, 57] as supervision. While for the coarse shape, we leverage the ground-truth vertex of the synthetic dataset as supervision. We also follow the previous methods [24, 21, 73] to leverage self-supervised losses for all training images.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, with the above insights and techniques, HiFace enables the reconstruction of high-fidelity 3D faces from a single image, and decouples static and dynamic details that are naturally animatable (see Fig. 1). We demonstrate that the proposed HiFace reconstructs realistic and faithful 3D faces, reaching state-of-the-art performance both quantitatively and qualitatively. In addition, HiFace is compatible with optimization-based methods [73], and is flexible to transfer vivid expressions and details from one person to another. In summary, our contributions are:", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose HiFace to model the static and dynamic details explicitly, and demonstrate the benefits of synthetic data in decoupling the static and dynamic factors for detailed 3D face reconstruction.", + "- We propose novel loss functions in HiFace to learn 3D representations of coarse shape and fine details simultaneously from both the synthetic and real-world images.", + "- We achieve state-of-the-art reconstruction quality both quantitatively and qualitatively, with over $15\\%$ performance gains in the region-aware benchmark [12].", + "- We show that our SD-DeTail is easy to plug-and-play into optimization-based methods and can transfer expressions and details from one to another for face animation." + ], + "bbox": [ + 500, + 242, + 890, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 452, + 640, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D face reconstruction from monocular images has received much attention in the past decades. Among them, 3D Morphable Models (3DMMs) are widely used to build 3D representations. Below we review the works that are related to them, and a full in-depth review can be found in recent surveys [82, 53, 23].", + "bbox": [ + 496, + 478, + 890, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Morphable Model (3DMMs) [23] are statistical models widely used to constrain the distribution of 3D faces. The seminal work [5] presents 200 scans to generate shape and texture bases with Principal Component Analysis (PCA) [1], and formulate 3DMMs as linear models by the generated bases. After that, expression models [69, 45, 10] are proposed to support face manipulation. Recent advances [12, 55, 65, 18, 44] are proposed to expand the expressiveness of 3DMMs and play a crucial role in 3D face reconstruction. 3DMMs make it possible to simplify the 2D-to-3D problem into a regression task, which typically presents an analysis-by-synthesis fashion to estimate the coefficients of 3DMMs. In this paper, we follow the spirit of the 3DMMs family to present the decoupled static and dynamic details for 3D face reconstruction.", + "bbox": [ + 496, + 575, + 890, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Coarse Shape Reconstruction. Traditional optimization-based methods [27, 73, 3, 77, 2] directly optimize the 3DMM coefficients of given 2D images. While such methods work well in controlled settings (e.g., frontal view, no occlusion), they heavily rely on high-quality annotations. Learning-based methods leverage the advances of", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/57afa58db775560ad3b2a545497d24eb5ac60f6ae79d4b0bdca43cf3ec9370bd.jpg", + "image_caption": [ + "Figure 2. Illustration of HiFace. (a). Learning framework of HiFace. Given a monocular image, we regress its shape and detail coefficients to synthesize a realistic 3D face, and leverage a differentiable renderer [28] to train the whole model end-to-end from synthetic [72, 57] and real-world [40, 52] images. (b). The pipeline of Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). We explicitly decouple the static and dynamic factors to synthesize realistic and animatable details. Given the shape and static coefficients, we regress the static and dynamic details through displacement bases and interpolate them into the final details through vertex tension [57]." + ], + "image_footnote": [ + "p: pose coef $\\beta$ : identity coef $\\xi$ : expression coef $\\varphi$ : static coef $\\alpha$ : albedo coef $\\gamma$ : SH coef $\\phi$ : dynamic coef" + ], + "bbox": [ + 81, + 84, + 517, + 250 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f1bab0659a69cf4081953d6d51bcec4f778bbde8640851b31fa33b5e9f881ef7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 83, + 887, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CNNs [67, 21, 19, 62] and GCNs [47, 41, 25] to learn high-level representations from large-scale images in the wild. These methods show plausible generalization over diverse environments. To train the network end-to-end, recent methods leverage the differentiable renderers [28, 22, 80, 49], along with the photo loss, perceptual loss, and landmark loss [21, 19, 28, 67, 71] to optimize the network in a self-supervised manner. Different from these coarse shape reconstruction methods, we aim at high-fidelity 3D face reconstruction with both coarse shape and fine details.", + "bbox": [ + 75, + 358, + 468, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Detail Reconstruction. While 3DMMs can reconstruct coarse 3D face shapes from 2D images, they struggle with reconstructing fine-level details, such as forehead wrinkles and crows-feet. To fill this gap, shape by shading (SfS) [34, 46, 26, 66] methods reconstruct the facial details from monocular images or videos. However, these methods are sensitive to occlusions and large poses. Recent advances [24, 48, 15, 13, 42] leverage displacement maps to present details. These methods explicitly re-topologize the coarse shape and present residual bias to generate geometric details. The main challenge of detail reconstruction is the difficulty in learning the nuances and disentangling the static and dynamic details from only self-supervised learning. Ground-truth labels of the details are helpful to guide the learning process. However, it is difficult to obtain such fine-grained labels on real data.", + "bbox": [ + 75, + 523, + 468, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic Dataset. Several methods [81, 77, 33, 68, 51] utilize rendered faces or fitted coefficients to synthesize 3D-2D pairs. These ground-truth pairs lack diversity over background, illumination, and assets, making them hard to generalize well to real-world images. Recent advances in synthetic data generation [72, 57] demonstrate its ability to generalize to real-world settings, and diverse to compensate for the domain gap to real-world images. In this paper, we", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "leverage high-quality data with ground-truth labels to explore the detailed 3D face reconstruction.", + "bbox": [ + 498, + 358, + 890, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 402, + 633, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 429, + 629, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt a common practice [21, 24] to represent a textured coarse shape with a 3D face model, an illumination model, and a camera model.", + "bbox": [ + 498, + 452, + 890, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Face Model. The 3D shape $\\mathbf{S}$ and albedo $\\mathbf{A}$ are represented by:", + "bbox": [ + 498, + 505, + 890, + 535 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\bar {\\mathbf {S}} + \\beta \\mathbf {B} _ {\\mathrm {i d}} + \\boldsymbol {\\xi} \\mathbf {B} _ {\\exp}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 536, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} = \\bar {\\mathbf {A}} + \\alpha \\mathbf {B} _ {\\mathrm {a l b}}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 555, + 725, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{S}}$ and $\\bar{\\mathbf{A}}$ are the mean shape and albedo. $\\mathbf{B}_{\\mathrm{id}}$ , $\\mathbf{B}_{\\mathrm{exp}}$ , and $\\mathbf{B}_{\\mathrm{alb}}$ are bases [73] of 256-dim identity, 233-dim expression, and 300-dim albedo, respectively. The coarse shape $\\mathbf{S}$ in the bind pose is deformed from a neutral shape $\\mathbf{S}_{\\mathrm{neu}} = \\bar{\\mathbf{S}} + \\alpha \\mathbf{B}_{\\mathrm{id}}$ with expression component $\\xi \\mathbf{B}_{\\mathrm{exp}}$ . $\\beta$ , $\\xi$ , and $\\alpha$ are the corresponding identity, expression, and albedo coefficients for generating a coarse shape. In this paper, the coarse shape $\\mathbf{S}$ contains $n_v = 7$ , 667 vertices and $n_f = 14$ , 832 triangles with $512 \\times 512 \\times 3$ albedo.", + "bbox": [ + 496, + 578, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose & Camera Model. To estimate the face pose, we follow [72, 73] to predict skeletal pose $\\mathbf{p} = [\\theta | \\mathbf{t}]$ , where $\\theta \\in \\mathbb{R}^{3j}$ and $\\mathbf{t} \\in \\mathbb{R}^3$ are the local joint rotations and root joint translation, respectively. $j = 4$ indicates 4 skeletal joints w.r.t. the head, neck, and two eyes. We perform a standard linear blend skinning (LBS) function [43] (with per-vertex weights $\\mathbf{W} \\in \\mathbb{R}^{j \\times n_v}$ ) to rotate $\\mathbf{S}$ about joint locations $\\mathbf{J} \\in \\mathbb{R}^{3j}$ by $\\mathbf{p}$ to obtain $\\mathbf{S_p}$ :", + "bbox": [ + 496, + 722, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {\\mathbf {p}} = \\operatorname {L B S} (\\mathbf {S}, \\mathbf {p}, \\mathbf {J}; \\mathbf {W}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 857, + 890, + 873 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{J}$ is the joint locations in the bind pose determined", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by $\\mathbf{J} = \\mathcal{J}(\\beta):\\mathbb{R}^{|\\beta |}\\to \\mathbb{R}^{3j}$ . Then we use an orthographic camera model to project 3D vertices in $\\mathbf{S}_{\\mathbf{p}}$ to the 2D plane.", + "bbox": [ + 75, + 90, + 468, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Illumination Model. We follow previous work [21] to use Spherical Harmonics (SH) [56] to estimate the illumination of a given image. The shaded texture $\\mathbf{T}$ is computed as:", + "bbox": [ + 75, + 128, + 468, + 174 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = \\mathbf {A} \\odot \\sum_ {k = 1} ^ {9} \\gamma_ {k} \\Psi_ {k} (\\mathbf {N}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 181, + 468, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the Hadamard product, $\\mathbf{N}$ is the surface normal of $\\mathbf{S}$ in UV coordinates, $\\Psi : \\mathbb{R}^3 \\to \\mathbb{R}$ are SH basis function and $\\gamma \\in \\mathbb{R}^9$ is the corresponding SH coefficient.", + "bbox": [ + 75, + 215, + 468, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview of HiFace", + "text_level": 1, + "bbox": [ + 76, + 268, + 264, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Key Idea. The key idea of HiFace is to explicitly model the static (e.g., person-specific properties) and dynamic (e.g., expression-driven wrinkles) details, allowing the model to reconstruct a high-fidelity 3D face from a single image with realistic and animatable details.", + "bbox": [ + 75, + 297, + 468, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. The goal of HiFace is to reconstruct 3D shapes with realistic details from a single image. The overview of HiFace is illustrated in Fig. 2(a). We leverage a feature extractor (i.e., ResNet-50 [30]) to regress corresponding coefficients from an input image. Our model jointly predicts both the coarse-level shapes and the fine-level details. For coarse-level shapes, we regress shape parameters (i.e., identity, expression, albedo, illumination, and pose) of a parametric face model. For the fine-level details, we propose a novel way to model it through the separation of static and dynamic factors and formulate the generation of details into the problems of 3DMM coefficients regression and displacement maps interpolation.", + "bbox": [ + 75, + 380, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the facial details are based on the coarse shape, we thereby exploit novel loss functions to learn 3D representations of coarse shape and details simultaneously from the synthetic dataset with ground-truth labels. To generalize our model to real-world images, we also present several self-supervised losses to train the model with both synthetic data and real-world images coherently. As a result, HiFace can faithfully reconstruct the facial details of a given image, or animate a face by combining the decoupled static and dynamic coefficients that come from different individuals.", + "bbox": [ + 75, + 577, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Decoupling Static and Dynamic Details", + "text_level": 1, + "bbox": [ + 76, + 734, + 413, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). The facial details are basically composed of a static factor and a dynamic factor:", + "bbox": [ + 75, + 758, + 468, + 804 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} = \\mathbf {D} _ {\\mathrm {s t a}} + \\mathbf {D} _ {\\mathrm {d y n}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 814, + 468, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{D}_{\\mathrm{sta}}$ and $\\mathbf{D}_{\\mathrm{dyn}}$ indicate details from static and dynamic factors, respectively.", + "bbox": [ + 75, + 839, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, the static factor is the inherent property of the identity (i.e., the given 2D face), and originates from the", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "appearance and age attributes. As for the dynamic factor, it is typically driven by the expression and influenced by person-specific properties.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Detail Generation. To simplify the problem, we are inspired by 3DMMs, which parameterize the statistical models to simplify the 2D-to-3D problem. We build a 300-dim displacement basis $\\mathbf{B}_{\\mathrm{sta}}$ from the captured 332 scans [57] by PCA [1]. The scans contain diverse age groups in a neutral expression. Then we regress the coefficient $\\varphi$ to synthesize the static detail $\\mathbf{D}_{\\mathrm{sta}}$ from the image:", + "bbox": [ + 496, + 143, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {s t a}} = \\bar {\\mathbf {D}} _ {\\mathrm {s t a}} + \\varphi \\mathbf {B} _ {\\mathrm {s t a}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 263, + 890, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{D}}_{\\mathrm{sta}}$ and $\\mathbf{B}_{\\mathrm{sta}}$ are the mean displacement map and displacement basis for static details, respectively.", + "bbox": [ + 496, + 285, + 890, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Detail Generation. Due to the high diversity and complexity of expression representation, directly generating dynamic details from expression is quite difficult. Therefore we simplify the expression representation by using an interpolation between two displacement maps: compressed and stretched [57]. For example, the compressed expression may indicate a state of frowning to the extreme, while the stretched expression may indicate a state of complete relaxation between the eyebrows. Other states of this area can be interpolated by these two polarized states.", + "bbox": [ + 496, + 321, + 890, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, we generate the dynamic details through compressed and stretched displacement maps. Again, we build 26-dim compressed $\\mathbf{B}_{\\mathrm{com}}$ and stretched $\\mathbf{B}_{\\mathrm{str}}$ displacement bases by PCA [1] to simplify the generation of displacement maps. To generate the dynamic coefficients $\\phi = \\{\\phi_{\\mathrm{com}},\\phi_{\\mathrm{str}}\\}$ , we apply the expression coefficient $\\xi$ into the static coefficient $\\varphi$ through AdaIN [32], followed by the MLP transformation $\\Phi$ to obtain $\\phi$ :", + "bbox": [ + 496, + 474, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi = \\Phi \\left(\\sigma (\\tilde {\\xi}) \\left(\\frac {\\varphi - \\mu (\\varphi)}{\\sigma (\\varphi)} + \\mu (\\tilde {\\xi})\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 604, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\xi}$ is the affine vector from $\\xi$ via MLP transformation. $\\mu$ and $\\sigma$ indicate the mean and standard deviation. $\\phi_{\\mathrm{com}}$ and $\\phi_{\\mathrm{str}}$ are coefficients for compressed and stretched displacement maps respectively.", + "bbox": [ + 496, + 647, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to Eq. 5, the compressed and stretched displacement maps are formulated as:", + "bbox": [ + 496, + 708, + 890, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\text {c o m}} = \\bar {\\mathbf {D}} _ {\\text {c o m}} + \\phi_ {\\text {c o m}} \\mathbf {B} _ {\\text {c o m}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 748, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {s t r}} = \\bar {\\mathbf {D}} _ {\\mathrm {s t r}} + \\phi_ {\\mathrm {s t r}} \\mathbf {B} _ {\\mathrm {s t r}}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 768, + 759, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{D}}_{\\mathrm{com}}$ and $\\mathbf{B}_{\\mathrm{com}}$ are the mean displacement map and 26-dim displacement basis for compressed detail, and $\\bar{\\mathbf{D}}_{\\mathrm{str}}$ and $\\mathbf{B}_{\\mathrm{str}}$ are the mean displacement map and 26-dim displacement basis for stretched detail, respectively.", + "bbox": [ + 496, + 794, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the coarse shape $\\mathbf{S}$ can be obtained by deforming the neutral shape $\\mathbf{S}_{\\mathrm{neu}}$ with the expression component $\\xi \\mathbf{B}_{\\mathrm{exp}}$ , such expression-driven deformation over face", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9090", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/683427d6bcce4e6e8970f926cbb722850d450993e6ec2ec4517ec88f998b8055.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 99, + 127, + 174 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0397e8454b8e2dab65c85962c1dfc55674ac8ca3fb58efda01a9c8d780fc8ebb.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 130, + 99, + 225, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7be66b7330bf156f511f2a67d024a6b814e669912e15518fb633195954a842f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 99, + 272, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b95feed28d83c8c8d7da6e1b36f3170cb768f5e131c994ab9cfa012f26dfaaa2.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 272, + 99, + 369, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80a23f5a1132fc3dfa2a958e1e5e61b3bb51fc9d3cd21435e8a642b7de033872.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 370, + 99, + 467, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7c450e66f055ea49c699839bb5e03a8a6fba7822077845a6c6faf4a1af722e93.jpg", + "image_caption": [ + "Figure 3. Illustration of displacement map composition in SD-DeTail. Given (a). an image (top) to reconstruct its coarse shape (bottom), we formulate the detail as (b). a static factor and (c). a dynamic factor interpolated by polarized states w.r.t. compressed (top) and stretched (bottom). (d). the output displacement map is linearly combined by (b) and (c) to present vivid details." + ], + "image_footnote": [], + "bbox": [ + 80, + 175, + 127, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ba885db0ffc073f09833f3b7d55c8d78f941f2b654db755ba34d7662fc18deb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 129, + 175, + 225, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a6224bc791411724dc63d3bc4a0333a605da9ea9ece01529d419a9242c3618f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 175, + 369, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/337bc07fc3e509321f5483ea1463f1ac3f6891d9eb7aa025e00f383e82c14a17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 175, + 467, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shape yields the \"tension\" over each vertex [57], which influences facial details from expression. Since $\\mathbf{S}_{\\mathrm{neu}}$ and $\\mathbf{S}$ possesses the same topology, for each vertex $\\mathbf{v}_i \\in \\mathbf{S}$ with $K$ -edges $E_i = \\{e_1, \\dots, e_K\\}$ connected with $\\mathbf{v}_i$ , $E_i' = \\{e_1', \\dots, e_K'\\}$ are the corresponding edges in $\\mathbf{S}_{\\mathrm{neu}}$ that are connected to $\\mathbf{v}_i'$ . Then the tension at $\\mathbf{v}_i$ is:", + "bbox": [ + 76, + 358, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt _ {\\mathbf {v} _ {i}} = 1 - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\left\\| e _ {k} \\right\\|}{\\left\\| e _ {k} ^ {\\prime} \\right\\|}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 460, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\|\\cdot\\|$ represents the edge length. Positive values of $t_{\\mathbf{v}_i}$ indicate compression, negative values indicate stretch, and 0-value indicates no change, respectively.", + "bbox": [ + 76, + 506, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The vertex tension $t_{\\mathbf{v}_i}$ in $\\mathbf{S}$ composes the tension map $\\mathbf{M}_{\\mathrm{uv}}$ in UV coordinates. Then, the displacement map of the dynamic detail is the linear interpolation of $\\mathbf{D}_{\\mathrm{com}}$ and $\\mathbf{D}_{\\mathrm{str}}$ :", + "bbox": [ + 76, + 551, + 468, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\mathrm {d y n}} = \\mathbf {M} _ {\\mathrm {u v}} ^ {+} \\odot \\mathbf {D} _ {\\mathrm {c o m}} + \\mathbf {M} _ {\\mathrm {u v}} ^ {-} \\odot \\mathbf {D} _ {\\mathrm {s t r}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 609, + 468, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{\\mathrm{uv}}^{+}$ and $\\mathbf{M}_{\\mathrm{uv}}^{-}$ indicate the positive and negative value of $\\mathbf{M}_{\\mathrm{uv}}$ , respectively. Fig. 3 shows the effectiveness of SD-DeTail. The dynamic factor interpolated by two polarized states introduces expression-related details and further decorates the static detail, yielding the final vivid output.", + "bbox": [ + 76, + 638, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 724, + 290, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose several loss functions to train HiFace end-to-end. Specifically, we use static and dynamic detail losses to supervise the synthesized displacement maps from $\\varphi$ and $\\phi$ . In addition, we leverage the coarse shape loss to supervise the reconstructed shape from $\\beta$ and $\\xi$ . Finally, we follow previous methods [21, 24, 73] to leverage the differentiable renderer [28] to map the generated 3D shape into 2D images, by combining $\\alpha$ , $\\beta$ , $\\xi$ , $\\gamma$ , $\\mathbf{p}$ , $\\varphi$ , $\\phi$ . Then, we perform self-supervised losses to train in both synthetic and real-world images. See more details in the supplementary.", + "bbox": [ + 76, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Static and Dynamic Detail Losses. To explicitly train the details of each component, we leverage the ground-truth annotations from the synthetic dataset [57, 72] as supervision to assist the training process of our model. Specifically, we calculate the detail losses by estimating the $l_{2}$ distance between the reconstructed displacement maps and ground-truth w.r.t. static, compressed, and stretched components, and summarize them as $\\mathcal{L}_{\\mathrm{detail}}$ :", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s t a}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {s t a}} - \\hat {\\mathbf {D}} _ {\\mathrm {s t a}}\\right) \\right\\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 220, + 795, + 241 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o m}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {c o m}} - \\hat {\\mathbf {D}} _ {\\mathrm {c o m}}\\right) \\right\\| _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 244, + 890, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s t r}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {s t r}} - \\hat {\\mathbf {D}} _ {\\mathrm {s t r}}\\right) \\right\\| _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 267, + 792, + 285 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e t a i l}} = \\mathcal {L} _ {\\text {s t a}} + \\mathcal {L} _ {\\text {c o m}} + \\mathcal {L} _ {\\text {s t r}}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 287, + 751, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{\\mathrm{detail}}$ is the facial mask in the UV coordinates, and $\\hat{\\mathbf{D}}_{\\mathrm{sta}} / \\hat{\\mathbf{D}}_{\\mathrm{com}} / \\hat{\\mathbf{D}}_{\\mathrm{str}}$ and $\\mathbf{D}_{\\mathrm{sta}} / \\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}$ are the reconstructed and ground-truth displacement maps, respectively.", + "bbox": [ + 500, + 314, + 890, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse Shape Losses. Since the details should be based on realistic coarse shapes, we train the coarse shape to help the learning of details by leveraging the ground-truth vertex as supervision:", + "bbox": [ + 496, + 366, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v e r}} = \\left\\| \\mathbf {M} _ {\\text {v e r}} \\odot (\\mathbf {S} - \\hat {\\mathbf {S}}) \\right\\| _ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 436, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{\\mathrm{ver}}$ is frontal face area of the coarse shape. $\\hat{\\mathbf{S}}$ and $\\mathbf{S}$ are the reconstructed and ground-truth face by Eq. 1.", + "bbox": [ + 498, + 468, + 890, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, we make constraints on shape coefficients to prevent overfitting. We enforce the predicted coefficients have a similar distribution to the ground-truth coefficients:", + "bbox": [ + 498, + 500, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k l}} = \\rho (\\boldsymbol {\\beta}) \\left(\\log \\rho (\\boldsymbol {\\beta}) - \\log \\rho (\\hat {\\boldsymbol {\\beta}})\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 554, + 890, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho$ denotes softmax function to map the predicted coefficients $\\hat{\\beta}$ and ground-truth $\\beta$ into probability distribution.", + "bbox": [ + 498, + 583, + 888, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the shape loss is $\\mathcal{L}_{\\mathrm{shp}} = \\mathcal{L}_{\\mathrm{ver}} + \\mathcal{L}_{\\mathrm{kl}}$", + "bbox": [ + 517, + 613, + 807, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised Losses. To encourage the generalization of our models in real-world images [40, 52], we follow previous methods [24, 21, 73] to leverage self-supervised loss $\\mathcal{L}_{\\mathrm{self}}$ for all training images, including photo loss $\\mathcal{L}_{\\mathrm{pho}}$ , perceptual loss $\\mathcal{L}_{\\mathrm{id}}$ , and dense landmark loss $\\mathcal{L}_{\\mathrm{lmk}}$ :", + "bbox": [ + 496, + 635, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\text {p h o}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}} + \\lambda_ {\\mathrm {l m k}} \\mathcal {L} _ {\\mathrm {l m k}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 722, + 890, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{id}}$ and $\\lambda_{\\mathrm{lmk}}$ are weights to balance the self-supervised losses term.", + "bbox": [ + 496, + 750, + 888, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, considering the static detail heavily correlates to person-specific age attribute, inspired by [19], we leverage the pre-trained age prediction network [36] to learn high-level representations of static details through knowledge distillation, such that the learned coefficients exhibit expressive results. To achieve this, we use several MLP layers on the static coefficient $\\varphi$ , and map it into age classification probabilities $\\hat{\\mathbf{p}}_{\\mathrm{age}}$ . Then we use the pre-trained", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9091", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f8c0798aeb8017fa0f2680a0a5f2e764871bfddab6d8cd95821dede146a2e434.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of 3D face reconstruction methods on REALY benchmark. “-c” and “-d” indicate coarse and detail shape, respectively. @RN/@RM/@RF/@RC/all indicate errors in nose/mouth/forehead/cheek/all regions. We highlight the best method for the two groups respectively. HiFace achieves the best reconstruction performance in the overall error by a large margin. Each component in HiFace contributes to a better reconstruction quality. The reconstructed details of HiFace further boost the quality while previous methods [24, 19] modeling details with only image-level supervision even deteriorate the reconstruction accuracy." + ], + "table_footnote": [ + "To align the dataset scale, w/o Syn. Data indicates we train the model without using the ground-truth labels from the synthetic dataset.", + "* To eliminate the bias of coarse shape in estimating the reconstruction error, we fix the coarse shape and train the details with/without static and dynamic factors for comparisons." + ], + "table_body": "
GroupMethods / e (mm)frontal-viewside-view
@RN@RM@RF@RCall@RN@RM@RF@RCall
CoarseDeep3D [21]1.719±0.3541.368±0.4392.015±0.4491.528±0.5011.6571.749±0.3431.411±0.3952.074±0.4861.528±0.5171.691
MGCNet [64]1.771±0.3801.417±0.4092.268±0.5031.639±0.6501.7741.827±0.3831.409±0.4182.248±0.5081.665±0.6441.787
3DDFA-v2 [29]1.903±0.5171.597±0.4782.447±0.6471.757±0.6421.9261.883±0.4991.642±0.5012.465±0.6221.781±0.6361.943
DECA-c [24]1.694±0.3552.516±0.8392.394±0.5761.479±0.5352.0101.903±1.0502.472±1.0792.423±0.7201.630±1.1352.107
SADRNet [61]1.791±0.5421.591±0.4882.413±0.5371.856±0.7011.9131.771±0.5211.560±0.4622.490±0.5662.010±0.7151.958
EMOCA-c [19]1.868±0.3872.679±1.1122.426±0.6411.438±0.5012.1031.867±0.5542.636±1.2842.448±0.7081.548±0.5902.125
MICA [81]1.585±0.3253.478±1.2042.374±0.6831.099±0.3242.1341.525±0.3223.567±1.2122.379±0.6751.109±0.3252.145
Ours-c (w/o Syn. Data)†1.227±0.4071.787±0.4391.454±0.3821.762±0.4361.5581.187±0.3791.826±0.4901.470±0.4261.653±0.4501.534
Ours-c1.054±0.3171.461±0.4301.331±0.3471.342±0.3841.2970.992±0.2461.505±0.4541.427±0.4001.439±0.4291.341
DetailDECA-d [24]2.138±0.4612.802±0.8682.457±0.5591.443±0.4982.2102.286±1.1032.684±1.0412.519±0.7181.555±0.8222.261
EMOCA-d [19]2.532±0.5392.929±1.1062.595±0.6311.495±0.4692.3882.455±0.6362.948±1.2922.606±0.6861.599±0.5632.402
HRN [42]1.722±0.3301.357±0.5231.995±0.4761.072±0.3331.5371.642±0.3101.285±0.5281.906±0.4791.038±0.3221.468
Ours-d (w/o Syn. Data)†1.465±0.5571.790±0.4251.528±0.3731.618±0.3621.6001.422±0.5371.849±0.4731.530±0.4141.572±0.3991.594
Ours-d (w/o static)*1.055±0.2901.469±0.4151.336±0.3371.319±0.3741.2951.004±0.2331.491±0.4371.418±0.3921.418±0.4151.332
Ours-d (w/o dynamic)*1.069±0.3181.469±0.4141.358±0.3361.270±0.3441.2920.991±0.2391.496±0.4371.411±0.3931.375±0.4021.318
Ours-d1.036±0.2801.450±0.4131.324±0.3341.291±0.3621.2750.985±0.2371.489±0.4361.399±0.3881.360±0.3951.308
", + "bbox": [ + 78, + 157, + 890, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age recognition model $\\Gamma_{\\mathrm{age}}$ to obtain the probabilities of the given input image $\\mathbf{I}$ . The distillation loss $\\mathcal{L}_{\\mathrm{kd}}$ enforces the probabilities between $\\hat{\\mathbf{p}}_{\\mathrm{age}}$ and $\\Gamma_{\\mathrm{age}}(\\mathbf{I})$ be similar:", + "bbox": [ + 75, + 396, + 470, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k d}} = \\Gamma_ {\\text {a g e}} (\\mathbf {I}) \\left(\\log \\Gamma_ {\\text {a g e}} (\\mathbf {I}) - \\log \\hat {\\mathbf {p}} _ {\\text {a g e}}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 450, + 468, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regularization. $\\mathcal{L}_{\\mathrm{reg}}$ regularizes coefficients of each submodule, by minimizing the $l_{2}$ loss of $\\alpha, \\beta, \\xi, \\varphi, \\phi$ .", + "bbox": [ + 76, + 477, + 468, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Loss Function. We train the coarse shape and fine details simultaneously, such that each component can collaborate to reconstruct high-fidelity 3D faces with realistic details. Formally, we minimize the total loss function:", + "bbox": [ + 75, + 521, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {d e t a i l}} \\mathcal {L} _ {\\text {d e t a i l}} + \\lambda_ {\\text {s h p}} \\mathcal {L} _ {\\text {s h p}} \\tag {15} \\\\ + \\lambda_ {\\text {s e l f}} \\mathcal {L} _ {\\text {s e l f}} + \\lambda_ {\\text {k d}} \\mathcal {L} _ {\\text {k d}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 592, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weight for each component.", + "bbox": [ + 76, + 636, + 359, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 664, + 209, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 688, + 294, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We use a hybrid dataset made up from both synthetic [72, 57] and real data [40, 52]. We use the synthetic data pipeline [72, 57] to generate a diverse dataset of $200k$ faces with ground-truth vertex, landmark, albedo, and displacement map annotations. The real-world datasets contain $400k$ images in total from diverse age, gender, and ethnicity groups. For the real-world dataset, we use the pretrained dense landmark detector [72] to detect 669 landmarks for training. We use face parsing [79] to generate and select region-of-interest as facial masks, providing robustness to common occlusions by hair or other accessories. We follow [19, 73, 21] to split the dataset into training", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and validation sets. The test images are from CelebA [40], FFHQ [37], LS3D-W [8], and AFLW2000 [78].", + "bbox": [ + 498, + 396, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement HiFace in PyTorch [54] and leverage the PyTorch3D differentiable rasterizer [35] for rendering. We train our model for 35 epochs on $8 \\times$ NVIDIA Tesla V100 GPUs with a mini-batch of 320. We use the pre-trained ResNet-50 on ImageNet [20] as initialization, and use Adam [39] as optimizer with an initial learning rate of $1e - 4$ . The input image is cropped and aligned by [14], and resized into $224 \\times 224$ . We empirically set $\\lambda_{\\mathrm{detail}} = 10$ , $\\lambda_{\\mathrm{shp}} = 1$ , $\\lambda_{\\mathrm{self}} = 1$ , $\\lambda_{\\mathrm{id}} = 0.1$ , $\\lambda_{\\mathrm{lmk}} = 0.5$ , $\\lambda_{\\mathrm{kd}} = 1$ , $\\lambda_{\\mathrm{reg}} = 1e - 3$ throughout the experiments.", + "bbox": [ + 498, + 433, + 892, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons to State-of-the-art", + "text_level": 1, + "bbox": [ + 500, + 595, + 781, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We perform the quantitative evaluation on the REALY benchmark [12], which contains 100 frontal-view and 400 side-view images from 100 textured-scans [17]. The REALY benchmark presents a region-aware evaluation pipeline to separately evaluate the metric error (in mm) of the nose, mouth, forehead, and cheek regions. Such an evaluation pipeline is demonstrated to better estimate the actual similarity of the 3D faces and align with human perception. We compare HiFace to previous state-of-the-art methods and report the region-wise and average normalized mean square error (NMSE) in Tab. 1.", + "bbox": [ + 496, + 628, + 893, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As Tab. 1 illustrates, HiFace outperforms prior arts in the overall error by a large margin. HiFace balances the reconstruction quality on each region, compared to those optimum region methods that may fail in specific regions (e.g., MICA [81] fails in mouth region while HRN [42] fails in forehead region). Note that HiFace faithfully recovers the facial details, thus making the reconstruction error smaller", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9092", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc7b84348e7ae97825277b3dd80793bb64fa799ec276a3f96de851633c8c80a5.jpg", + "image_caption": [ + "Input Deep3D 3DDFA-v2 SynergyNet MICA Dense Ours" + ], + "image_footnote": [], + "bbox": [ + 80, + 97, + 470, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e009b9dce3c38cf8a273e6fcceabbc11968dbd58488edb3fea263f32a64a4bb.jpg", + "image_caption": [ + "Figure 4. Comparison on coarse shape reconstruction. From left to right: Input image, Deep3D [21], 3DDFA-v2 [29], Synerg-yNet [74], MICA [81], Dense [73], and HiFace (Ours).", + "Figure 6. Illustration on the flexibility of SD-DeTail. Given the identity and expression coefficients $(\\beta, \\xi)$ from the optimization-based method [73], SD-DeTail can generate realistic details based on the coarse shape and further improve the visual quality." + ], + "image_footnote": [], + "bbox": [ + 78, + 332, + 468, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "than using the coarse shape alone. As a comparison, although DECA [24] and EMOCA [19] can reconstruct details of given images, they turn out to be noisy, leading to the deterioration of reconstruction quality.", + "bbox": [ + 75, + 526, + 468, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, Tab. 1 also demonstrates the necessity of each component in contributing to a better quality. It can be observed that the synthetic data with ground-truth labels not only improve the coarse shape reconstruction quality but is also crucial for detailed reconstruction. With the synthetic data, the proposed SD-DeTail further boosts the overall reconstruction quality. Both the static and dynamic factors are essential to capture fine-grained details, and the final SD-DeTail achieves the most accurate details in expression-related regions such as the mouth and forehead, which contributes to the overall gains.", + "bbox": [ + 75, + 588, + 468, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. Given a single face image, Hi-Face reconstructs a high-fidelity 3D shape with details. We present comparisons with previous methods on 1). coarse shape reconstruction [21, 29, 74, 81, 73] in Fig. 4 and 2). detail reconstruction [77, 15, 24, 19, 70] in Fig. 5. See more examples and comparisons in the supplementary.", + "bbox": [ + 75, + 763, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For coarse shape in Fig. 4, our HiFace faithfully recovers the coarse shape of the given identity and outperforms the previous learning-based methods, and is on par", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96bcd63fa56e720e31a542c754ffe9e6af72fa417bbd8fc37a20b84b1fe48101.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 87, + 887, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d198d2b1537a3210ca597276579b3798a63cec5f94e02a143f676daf8c9a3307.jpg", + "image_caption": [ + "Figure 5. Comparison on detail shape reconstruction. From left to right: Input image, FaceScape [77], Unsup [15], DECA [24], EMOCA [19], FaceVerse [70], and HiFace (Ours).", + "Figure 7. Comparison on face animation. Given a source image (yellow box), we use the driving images (green box) to drive its expressions. DECA [24] (2nd-row) and EMOCA [19] (3rd-row) can animate the expression-driven details but lack realistic. As a comparison, HiFace is flexible to animate details from static (4th-row), dynamic (5th-row), or both (6th-row) factors, and presents vivid animation quality with realistic shapes." + ], + "image_footnote": [], + "bbox": [ + 501, + 333, + 888, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with Dense [73], which is the state-of-the-art optimization-based method. For detailed reconstruction in Fig. 5, previous methods [24, 19] fail to reconstruct satisfactory details. Several methods [15, 77, 70] are sensitive to occlusions and large poses. As a comparison, HiFace achieves the most realistic reconstruction quality, and faithfully recovers facial details of a given image, which significantly outperforms previous methods by a large margin.", + "bbox": [ + 496, + 760, + 890, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, given an image and the fitted coefficients $\\beta$ ,", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9093", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbe86c2ac57a9cacabced6224a497ecd878935888a91d0304de9aa09358173c9.jpg", + "image_caption": [ + "Figure 8. Ablation studies on loss functions and training data. The coarse shape losses $\\mathcal{L}_{\\mathrm{shp}} / \\mathcal{L}_{\\mathrm{kl}}$ (left), detail losses $\\mathcal{L}_{\\mathrm{detail}} / \\mathcal{L}_{\\mathrm{kd}}$ (right), and hybrid datasets coherently contribute to the reconstruction quality of coarse shapes and details." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bdf077b5d3d7866c628b7cc9a327084fda34181b08ced33d24950c22d3d41570.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 88, + 887, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\xi$ from the optimization-based methods such as Dense [73], SD-DeTail synthesizes the details and further strengthens the quality compared to the coarse shape (see Fig. 6). It shows SD-DeTail is flexible and can be easily plugged-and-play into other methods. See more in the supplementary.", + "bbox": [ + 75, + 268, + 468, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application of HiFace. The HiFace explicitly decouples the static and dynamic details through the proposed SD-DeTail. Therefore, we can animate the facial attributes by simply assigning the expression coefficient $\\xi$ and/or static coefficient $\\varphi$ of the driving images to the source images.", + "bbox": [ + 75, + 351, + 468, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 7, we demonstrate the animation quality of HiFace outperforms the previous state-of-the-art detail animation methods [24, 19]. It shows that while DECA [24] and EMOCA [19] can animate the expression-driven details but lack realistic, the proposed HiFace is flexible to manipulate the static, dynamic, or both details. When animating the static detail, the person-specific details can be well transferred into the source shape. When animating the dynamic detail, only expression-dependent details are presented. Finally, we can also animate the static and dynamic details simultaneously and achieve satisfactory results.", + "bbox": [ + 75, + 428, + 468, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 609, + 240, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies on Loss Functions and Datasets. We present ablation studies on the proposed loss functions and training strategy with hybrid datasets. We train HiFace with synthetic data alone and compare it to the one trained with hybrid datasets. For coarse shape reconstruction, we investigate the contribution of $\\mathcal{L}_{\\mathrm{shp}}$ and its sub-term $\\mathcal{L}_{\\mathrm{kl}}$ . For detail reconstruction, we compare HiFace without $\\mathcal{L}_{\\mathrm{detail}}$ and $\\mathcal{L}_{\\mathrm{kd}}$ , respectively. The results are presented in Fig. 8.", + "bbox": [ + 75, + 642, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 8 demonstrates that the proposed loss functions and training strategy from hybrid datasets contribute to satisfactory coarse shape and details. First, the model trained with synthetic data alone cannot generalize well to real-world images, which indicates the necessity to train with real-world data. Second, $\\mathcal{L}_{\\mathrm{shp}}$ improves the coarse shape reconstruction quality. $\\mathcal{L}_{\\mathrm{shp}}$ is effective in tackling challenging poses and improving alignment. $\\mathcal{L}_{\\mathrm{kl}}$ can relieve the overfitting risk on the synthetic data and improve the gener", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02fee0cf9532e8d092cb67544266cebe39309ee88183fadd57d475e656449220.jpg", + "image_caption": [ + "Figure 9. Ablation studies on SD-DeTail. Results show that directly synthesizing the static or dynamic details is rather challenging, leading to unreasonable coarse shapes and details (SD-1, SD-2, and SD-3). As a comparison, we leverage the statistical bases with SD-DeTail and regard the detail generation problem as a coefficients regression and interpolation problem, leading to more realistic details." + ], + "image_footnote": [], + "bbox": [ + 500, + 263, + 890, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "alization to real-world images. Third, without $\\mathcal{L}_{\\mathrm{detail}}$ or $\\mathcal{L}_{\\mathrm{kd}}$ , the reconstructed details exhibit random noise and cannot faithfully reflect person-specific details. Such noise misses the correspondence to the person-specific identity.", + "bbox": [ + 498, + 515, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies on SD-DeTail. To verify the effectiveness of building bases for static and dynamic details, we present detailed ablation studies on SD-DeTail, by replacing the bases (i.e., $\\mathbf{B}_{\\mathrm{sta}}$ and $\\mathbf{B}_{\\mathrm{com}} / \\mathbf{B}_{\\mathrm{str}}$ ) reconstruction with a U-Net decoder [60] (same as DECA [24]). Therefore, the model learns to directly synthesize displacement maps instead of predicting corresponding coefficients like ours. In Fig. 9, we make comparisons on: 1). directly synthesizing $\\mathbf{D}_{\\mathrm{dyn}}$ (SD-1), 2). directly synthesizing $\\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}$ and interpolating via Eq. 9 (SD-2), 3). directly synthesizing $\\mathbf{D}_{\\mathrm{sta}}$ (SD-3), 4). SD-DeTail (Ours). It can be seen that, due to the high diversity and complexity of expression representation, it is hard to directly learn realistic details even with ground-truth labels of $\\mathbf{D}_{\\mathrm{dyn}}$ from synthetic data (see SD-1, SD-2 and SD-3 in Fig. 9). More specifically, for the static details, directly synthesizing displacement maps bring much noise (SD-3). For example, the hollow eyebrow is demonstrated in the second row. For the dynamic details, directly synthesizing displacement maps even leads to unnatural results (SD-1 and SD-2). For example, the reconstructed 3D faces are distorted especially in the second row. We also notice", + "bbox": [ + 496, + 583, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9094", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that, directly synthesizing $\\mathbf{D}_{\\mathrm{dyn}}$ (SD-1) achieves inferior results than directly synthesizing $\\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}$ and interpolating via Eq. 9 (SD-2). This demonstrates that it is beneficial to simplify the expression representation by using interpolation between two displacement maps (i.e., compressed and stretched). In conclusion, these observations further verify our insight on relaxing the challenging detail generation problem into a feasible coefficients regression problem.", + "bbox": [ + 75, + 90, + 472, + 213 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 224, + 194, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose HiFace to reconstruct high-fidelity 3D faces with realistic and animatable details from a single image. Our motivation and insights stand on the spirit of 3DMMs to simplify the challenging detail generation into more accessible regression and interpolation tasks. To achieve this, we elaborately design SD-DeTail to decouple the static and dynamic factors explicitly, and interpolate the dynamic details through vertex tension. We succeed in learning the coarse shape and details jointly by proposing several novel loss functions to train on synthetic and real-world data. Extensive experiments demonstrate that HiFace achieves state-of-the-art face reconstruction both quantitatively and qualitatively in the coarse shape and detail shape, and the details are well decoupled and naturally animatable.", + "bbox": [ + 75, + 250, + 472, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 476, + 236, + 493 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SZSTC Grant (JCYJ20190809172201639, WDZC2020-0820200655001), Shenzhen Key Laboratory (ZDSYS2021-0623092001004), and Beijing Key Lab of Networked Multimedia.", + "bbox": [ + 75, + 501, + 468, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 604, + 174, + 619 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 2, 4", + "[2] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2", + "[3] Linchao Bao, Xiangkai Lin, Yajing Chen, Haoxian Zhang, Sheng Wang, Xuefei Zhe, Di Kang, Haozhi Huang, Xinwei Jiang, Jue Wang, et al. High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2", + "[4] Volker Blanz, Sami Romdhani, and Thomas Vetter. Face identification across different poses and illuminations with a 3d morphable model. In Proceedings of fifth IEEE international conference on automatic face gesture recognition, pages 202-207. IEEE, 2002. 1", + "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual" + ], + "bbox": [ + 78, + 630, + 470, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on Computer graphics and interactive techniques, pages 187-194, 1999. 1, 2", + "[6] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2):233-254, 2018. 1", + "[7] Sofien Bouaziz, Yangang Wang, and Mark Pauly. Online modeling for realtime facial animation. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013. 1", + "[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017. 6", + "[9] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Realtime high-fidelity facial performance capture. ACM Transactions on Graphics (ToG), 34(4):1-9, 2015. 2", + "[10] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 1, 2", + "[11] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5187-5196, 2018. 1", + "[12] Zenghao Chai, Haoxian Zhang, Jing Ren, Di Kang, Zhengzhuo Xu, Xuefei Zhe, Chun Yuan, and Linchao Bao. Really: Rethinking the evaluation of 3d face reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6", + "[13] Anpei Chen, Zhang Chen, Guli Zhang, Kenny Mitchell, and Jingyi Yu. Photo-realistic facial details synthesis from single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9429-9439, 2019. 2, 3", + "[14] Dong Chen, Gang Hua, Fang Wen, and Jian Sun. Supervised transformer network for efficient face detection. In European Conference on Computer Vision, pages 122-138. Springer, 2016. 6", + "[15] Yajing Chen, Fanzi Wu, Zeyu Wang, Yibing Song, Yonggen Ling, and Linchao Bao. Self-supervised learning of detailed 3d face reconstruction. IEEE Transactions on Image Processing, 29:8696-8705, 2020. 2, 3, 7", + "[16] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1", + "[17] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 2019. 6", + "[18] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE international conference on computer vision, pages 3085-3093, 2017. 2", + "[19] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 2, 3, 5, 6, 7, 8" + ], + "bbox": [ + 503, + 92, + 893, + 904 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9095", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[21] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 1, 2, 3, 4, 5, 6, 7", + "[22] Abdallah Dib, Cédric Thébault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 3", + "[23] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2", + "[24] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2, 3, 5, 6, 7, 8", + "[25] Zhongpai Gao, Juyong Zhang, Yudong Guo, Chao Ma, Guangtao Zhai, and Xiaokang Yang. Semi-supervised 3d face representation learning from unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 348-349, 2020. 3", + "[26] Pablo Garrido, Michael Zollhöfer, Dan Casas, Levi Valgaerts, Kiran Varanasi, Patrick Pérez, and Christian Theobalt. Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG), 35(3):1-15, 2016. 3", + "[27] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 2", + "[28] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3, 5", + "[29] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 6, 7", + "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[31] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG), 36(6):1-14, 2017. 1", + "[32] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + ], + "bbox": [ + 78, + 90, + 468, + 904 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4", + "[33] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3", + "[34] Luo Jiang, Juyong Zhang, Bailin Deng, Hao Li, and Ligang Liu. 3d face reconstruction with geometry details from a single image. IEEE Transactions on Image Processing, 27(10):4756-4770, 2018. 3", + "[35] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. ACM New York, NY, USA, 2020. 6", + "[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5", + "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6", + "[38] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 1", + "[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6", + "[40] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6", + "[41] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware mesh decoder for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6100–6109, 2020. 1, 3", + "[42] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 3, 6", + "[43] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 165-172, 2000. 3", + "[44] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 2", + "[45] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9096", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Yue Li, Liqian Ma, Haoqiang Fan, and Kenny Mitchell. Feature-preserving detailed 3d face reconstruction from a single image. In Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production, pages 1-9, 2018. 3", + "[47] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3d face reconstruction from in-the-wild images using graph convolutional networks. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 5891-5900, 2020. 3", + "[48] Jingwang Ling, Zhibo Wang, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Structure-aware editable morphable model for 3d facial detail animation and manipulation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[49] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3", + "[50] Huiwen Luo, Koki Nagano, Han-Wei Kung, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11662–11672, 2021. 2", + "[51] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriia Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20942–20952, 2022. 3", + "[52] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 3, 5, 6", + "[53] Araceli Morales, Gemma Piella, and Federico M Sukno. Survey on 3d face reconstruction from uncalibrated images. Computer Science Review, 40:100400, 2021. 2", + "[54] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[55] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2", + "[56] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4", + "[57] Chirag Raman, Charlie Hewitt, Erroll Wood, and Tadas Baltrusaitis. Mesh-tension driven expression-based wrinkles for synthetic faces. In WACV, 2023. 2, 3, 4, 5, 6", + "[58] Roger Blanco I Ribera, Eduard Zell, John P Lewis, Junyong Noh, and Mario Botsch. Facial retargeting with automatic range of motion alignment. ACM Transactions on graphics (TOG), 36(4):1-12, 2017. 1", + "[59] Sami Romdhani, Volker Blanz, and Thomas Vetter. Face identification by fitting a 3d morphable model using linear" + ], + "bbox": [ + 78, + 90, + 468, + 904 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shape and texture error functions. In European Conference on Computer Vision, pages 3-19. Springer, 2002. 1", + "[60] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 8", + "[61] Zeyu Ruan, Changqing Zou, Longhai Wu, Gangshan Wu, and Limin Wang. Sadrnet: Self-aligned dual face regression networks for robust 3d dense face alignment and reconstruction. IEEE Transactions on Image Processing, 30:5793-5806, 2021. 6", + "[62] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3", + "[63] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1", + "[64] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 6", + "[65] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 1, 2", + "[66] Supasorn Suwajanakorn, Ira Kemelmacher-Shlizerman, and Steven M Seitz. Total moving face reconstruction. In European conference on computer vision, pages 796-812. Springer, 2014. 3", + "[67] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 1, 2, 3", + "[68] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3", + "[69] Daniel Vlasic, Matthew Brand, Hanspeter Pfister, and Jovan Popovic. Face transfer with multilinear models. In ACM SIGGRAPH 2006 Courses, pages 24–es. ACM New York, NY, USA, 2006. 2", + "[70] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 7", + "[71] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13289-13298, 2021. 3" + ], + "bbox": [ + 503, + 92, + 890, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9097", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2, 3, 5, 6", + "[73] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 2, 3, 5, 6, 7, 8", + "[74] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 7", + "[75] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 1", + "[76] Shugo Yamaguchi, Shunsuke Saito, Koki Nagano, Yajie Zhao, Weikai Chen, Kyle Olszewski, Shigeo Morishima, and Hao Li. High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2", + "[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu" + ], + "bbox": [ + 78, + 90, + 468, + 446 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 2, 3, 7", + "[78] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Towards large-posed face frontalization in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3990–3999, 2017. 6", + "[79] Qi Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Computer Vision and Pattern Recognition, 2022. 6", + "[80] Wenbin Zhu, HsiangTao Wu, Zeyu Chen, Noranart Vesdapunt, and Baoyuan Wang. Reda: reinforced differentiable attribute for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4958-4967, 2020. 3", + "[81] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision (ECCV). Springer International Publishing, Oct. 2022. 1, 2, 3, 6, 7", + "[82] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum. Wiley Online Library, 2018. 2" + ], + "bbox": [ + 503, + 92, + 890, + 446 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9098", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_model.json b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..347d8b858f95c48d1eac55f171b1329b5ac237ec --- /dev/null +++ b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_model.json @@ -0,0 +1,3018 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.738, + 0.177 + ], + "angle": 0, + "content": "HiFace: High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.203, + 0.841, + 0.311 + ], + "angle": 0, + "content": "Zenghao Chai\\(^{1,2*}\\) Tianke Zhang\\(^{2}\\) Tianyu He\\(^{3}\\) Xu Tan\\(^{3\\dagger}\\) Tadas Baltrušaitis\\(^{4}\\) HsiangTao Wu\\(^{5}\\) Runnan Li\\(^{5}\\) Sheng Zhao\\(^{5}\\) Chun Yuan\\(^{2}\\) Jiang Bian\\(^{3}\\) \n\\(^{1}\\)National University of Singapore \\(^{2}\\)Tsinghua University \\(^{3}\\)Microsoft Research Asia \n\\(^{4}\\)Microsoft Mixed Reality & AI Lab \\(^{5}\\)Microsoft Cloud + AI \nzenghaochai@gmail.com ztk21@mails.tsinghua.edu.cn yuanc@sz.tsinghua.edu.cn \n{tianyuhe,xuta,tabaltru,musclewu,runnan.li,sheng.zhao,jiang.bian}@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.474, + 0.71 + ], + "angle": 0, + "content": "3D Morphable Models (3DMMs) demonstrate great potential for reconstructing faithful and animatable 3D facial surfaces from a single image. The facial surface is influenced by the coarse shape, as well as the static detail (e.g., person-specific appearance) and dynamic detail (e.g., expression-driven wrinkles). Previous work struggles to decouple the static and dynamic details through image-level supervision, leading to reconstructions that are not realistic. In this paper, we aim at high-fidelity 3D face reconstruction and propose HiFace to explicitly model the static and dynamic details. Specifically, the static detail is modeled as the linear combination of a displacement basis, while the dynamic detail is modeled as the linear interpolation of two displacement maps with polarized expressions. We exploit several loss functions to jointly learn the coarse shape and fine details with both synthetic and real-world datasets, which enable HiFace to reconstruct high-fidelity 3D shapes with animatable details. Extensive quantitative and qualitative experiments demonstrate that HiFace presents state-of-the-art reconstruction quality and faithfully recovers both the static and dynamic details. Our project page: https://project-hiface.github.io." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.736, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.868 + ], + "angle": 0, + "content": "The reconstruction of a 3D face from a single image has drawn much attention recently [67, 21, 41, 81]. It has tremendous potential applications like face recognition [11, 63, 4, 59], face animation [16, 75], virtual reality [7, 58, 31], etc. For example, the reconstructed 3D face representation can be driven by an audio [16], or a video from another person [38]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.344, + 0.697, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.344, + 0.89, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.493, + 0.696, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.493, + 0.89, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.64, + 0.665, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.64, + 0.858, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.801 + ], + "angle": 0, + "content": "Figure 1. We propose HiFace to reconstruct high-fidelity 3D face with realistic and animatable details. Reconstruction: given a single image (1st-row), HiFace faithfully reconstructs a coarse shape (2nd-row) with vivid details (3rd-row). Animation: given a source face (yellow box), HiFace can animate the static (4th-row), dynamic (5th-row), or both (6th-row) details of the driving images (green box). Images are taken from FFHQ [37] and CelebA [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To build a flexible and animatable facial representation, a popular way is to leverage the success of 3D Morphable Models (3DMMs) [5, 6, 10, 45, 65], which decouple the influence of shape, expression, albedo, and others by modeling them in separate coefficients. Typically" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.388, + 0.901 + ], + "angle": 0, + "content": "*Work done when the author was an intern at MSRA. †Corresponding author: Xu Tan (xuta@microsoft.com)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9087" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.38 + ], + "angle": 0, + "content": "in literature, one can achieve coarse shape reconstruction in coefficients-fitting optimization [27, 73, 3, 77, 2], or an analysis-by-synthesis pipeline [67, 21, 81, 50]. As 3DMMs typically capture only the coarse facial geometry and are not capable of representing fine details (e.g., wrinkles), recent advances model such details with a displacement map [15, 77, 13, 9, 76]. However, previous work fails to model the distinction between static and dynamic factors of fine detail, leading to errors in reconstructions. For example, given that one may drive the expression of a young man from an old man, trivially transferring all wrinkles from the old man to the young man could make the young man look unnatural. In this sense, Feng et al. [24] implicitly leverages the person-specific identity and expression as conditions to generate the details. Although effective, they optimize the model in an analysis-by-synthesis pipeline with only the image-level supervision, leading to insufficient decoupling of static and dynamic details and inconsistent animation results (see Fig. 7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.473, + 0.746 + ], + "angle": 0, + "content": "Therefore, we propose HiFace to explicitly model the static and dynamic details for high-fidelity 3D face reconstruction, by designing SD-DeTail module to decouple the static and dynamic factors. More specifically, for person-specific static detail, instead of directly predicting the displacement map that may increase the difficulty of detail prediction [24, 19], we follow the spirit of 3DMMs to build a displacement basis from the captured facial scans with age diversity [57, 72]. In this way, the model is trained to predict the coefficients of the displacement basis, and make the detail prediction easier. For dynamic detail, since it is highly expression-dependent, directly modeling it with one displacement basis is quite difficult. Therefore, based on the fact that the expression can be interpolated by a compressed and a stretched expressions [57], we build two displacement bases for the compressed and stretched expressions from the captured scans respectively, and learn to regress the displacement coefficients with the ground-truth labels. Therefore, we can obtain the dynamic detail by linearly interpolating the compressed and stretched displacement maps, which are derived from the displacement bases and the predicted coefficients. Finally, the predicted static and dynamic details are merged with the coarse shape to formulate the final output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Since we would like the final output to contain both the coarse shape and high-frequency detail, we propose several novel loss functions to learn coarse shape and details simultaneously from both the synthetic and real-world datasets. For details, we leverage the ground-truth static and dynamic displacement maps of the synthetic dataset [72, 57] as supervision. While for the coarse shape, we leverage the ground-truth vertex of the synthetic dataset as supervision. We also follow the previous methods [24, 21, 73] to leverage self-supervised losses for all training images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Overall, with the above insights and techniques, HiFace enables the reconstruction of high-fidelity 3D faces from a single image, and decouples static and dynamic details that are naturally animatable (see Fig. 1). We demonstrate that the proposed HiFace reconstructs realistic and faithful 3D faces, reaching state-of-the-art performance both quantitatively and qualitatively. In addition, HiFace is compatible with optimization-based methods [73], and is flexible to transfer vivid expressions and details from one person to another. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.243, + 0.892, + 0.302 + ], + "angle": 0, + "content": "- We propose HiFace to model the static and dynamic details explicitly, and demonstrate the benefits of synthetic data in decoupling the static and dynamic factors for detailed 3D face reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.303, + 0.892, + 0.348 + ], + "angle": 0, + "content": "- We propose novel loss functions in HiFace to learn 3D representations of coarse shape and fine details simultaneously from both the synthetic and real-world images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.349, + 0.892, + 0.393 + ], + "angle": 0, + "content": "- We achieve state-of-the-art reconstruction quality both quantitatively and qualitatively, with over \\(15\\%\\) performance gains in the region-aware benchmark [12]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.394, + 0.892, + 0.438 + ], + "angle": 0, + "content": "- We show that our SD-DeTail is easy to plug-and-play into optimization-based methods and can transfer expressions and details from one to another for face animation." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.243, + 0.892, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.642, + 0.469 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.57 + ], + "angle": 0, + "content": "3D face reconstruction from monocular images has received much attention in the past decades. Among them, 3D Morphable Models (3DMMs) are widely used to build 3D representations. Below we review the works that are related to them, and a full in-depth review can be found in recent surveys [82, 53, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.803 + ], + "angle": 0, + "content": "3D Morphable Model (3DMMs) [23] are statistical models widely used to constrain the distribution of 3D faces. The seminal work [5] presents 200 scans to generate shape and texture bases with Principal Component Analysis (PCA) [1], and formulate 3DMMs as linear models by the generated bases. After that, expression models [69, 45, 10] are proposed to support face manipulation. Recent advances [12, 55, 65, 18, 44] are proposed to expand the expressiveness of 3DMMs and play a crucial role in 3D face reconstruction. 3DMMs make it possible to simplify the 2D-to-3D problem into a regression task, which typically presents an analysis-by-synthesis fashion to estimate the coefficients of 3DMMs. In this paper, we follow the spirit of the 3DMMs family to present the decoupled static and dynamic details for 3D face reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Coarse Shape Reconstruction. Traditional optimization-based methods [27, 73, 3, 77, 2] directly optimize the 3DMM coefficients of given 2D images. While such methods work well in controlled settings (e.g., frontal view, no occlusion), they heavily rely on high-quality annotations. Learning-based methods leverage the advances of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.085, + 0.518, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.202, + 0.256, + 0.79, + 0.269 + ], + "angle": 0, + "content": "p: pose coef \\(\\beta\\): identity coef \\(\\xi\\): expression coef \\(\\varphi\\): static coef \\(\\alpha\\): albedo coef \\(\\gamma\\): SH coef \\(\\phi\\): dynamic coef" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.084, + 0.888, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Figure 2. Illustration of HiFace. (a). Learning framework of HiFace. Given a monocular image, we regress its shape and detail coefficients to synthesize a realistic 3D face, and leverage a differentiable renderer [28] to train the whole model end-to-end from synthetic [72, 57] and real-world [40, 52] images. (b). The pipeline of Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). We explicitly decouple the static and dynamic factors to synthesize realistic and animatable details. Given the shape and static coefficients, we regress the static and dynamic details through displacement bases and interpolate them into the final details through vertex tension [57]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.511 + ], + "angle": 0, + "content": "CNNs [67, 21, 19, 62] and GCNs [47, 41, 25] to learn high-level representations from large-scale images in the wild. These methods show plausible generalization over diverse environments. To train the network end-to-end, recent methods leverage the differentiable renderers [28, 22, 80, 49], along with the photo loss, perceptual loss, and landmark loss [21, 19, 28, 67, 71] to optimize the network in a self-supervised manner. Different from these coarse shape reconstruction methods, we aim at high-fidelity 3D face reconstruction with both coarse shape and fine details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.766 + ], + "angle": 0, + "content": "Detail Reconstruction. While 3DMMs can reconstruct coarse 3D face shapes from 2D images, they struggle with reconstructing fine-level details, such as forehead wrinkles and crows-feet. To fill this gap, shape by shading (SfS) [34, 46, 26, 66] methods reconstruct the facial details from monocular images or videos. However, these methods are sensitive to occlusions and large poses. Recent advances [24, 48, 15, 13, 42] leverage displacement maps to present details. These methods explicitly re-topologize the coarse shape and present residual bias to generate geometric details. The main challenge of detail reconstruction is the difficulty in learning the nuances and disentangling the static and dynamic details from only self-supervised learning. Ground-truth labels of the details are helpful to guide the learning process. However, it is difficult to obtain such fine-grained labels on real data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Synthetic Dataset. Several methods [81, 77, 33, 68, 51] utilize rendered faces or fitted coefficients to synthesize 3D-2D pairs. These ground-truth pairs lack diversity over background, illumination, and assets, making them hard to generalize well to real-world images. Recent advances in synthetic data generation [72, 57] demonstrate its ability to generalize to real-world settings, and diverse to compensate for the domain gap to real-world images. In this paper, we" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.359, + 0.892, + 0.391 + ], + "angle": 0, + "content": "leverage high-quality data with ground-truth labels to explore the detailed 3D face reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.404, + 0.634, + 0.422 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.43, + 0.63, + 0.446 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.892, + 0.498 + ], + "angle": 0, + "content": "We adopt a common practice [21, 24] to represent a textured coarse shape with a 3D face model, an illumination model, and a camera model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.506, + 0.891, + 0.536 + ], + "angle": 0, + "content": "3D Face Model. The 3D shape \\(\\mathbf{S}\\) and albedo \\(\\mathbf{A}\\) are represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.537, + 0.891, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\bar {\\mathbf {S}} + \\beta \\mathbf {B} _ {\\mathrm {i d}} + \\boldsymbol {\\xi} \\mathbf {B} _ {\\exp}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.556, + 0.726, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} = \\bar {\\mathbf {A}} + \\alpha \\mathbf {B} _ {\\mathrm {a l b}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.716 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{S}}\\) and \\(\\bar{\\mathbf{A}}\\) are the mean shape and albedo. \\(\\mathbf{B}_{\\mathrm{id}}\\), \\(\\mathbf{B}_{\\mathrm{exp}}\\), and \\(\\mathbf{B}_{\\mathrm{alb}}\\) are bases [73] of 256-dim identity, 233-dim expression, and 300-dim albedo, respectively. The coarse shape \\(\\mathbf{S}\\) in the bind pose is deformed from a neutral shape \\(\\mathbf{S}_{\\mathrm{neu}} = \\bar{\\mathbf{S}} + \\alpha \\mathbf{B}_{\\mathrm{id}}\\) with expression component \\(\\xi \\mathbf{B}_{\\mathrm{exp}}\\). \\(\\beta\\), \\(\\xi\\), and \\(\\alpha\\) are the corresponding identity, expression, and albedo coefficients for generating a coarse shape. In this paper, the coarse shape \\(\\mathbf{S}\\) contains \\(n_v = 7\\), 667 vertices and \\(n_f = 14\\), 832 triangles with \\(512 \\times 512 \\times 3\\) albedo." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.892, + 0.845 + ], + "angle": 0, + "content": "Pose & Camera Model. To estimate the face pose, we follow [72, 73] to predict skeletal pose \\(\\mathbf{p} = [\\theta | \\mathbf{t}]\\), where \\(\\theta \\in \\mathbb{R}^{3j}\\) and \\(\\mathbf{t} \\in \\mathbb{R}^3\\) are the local joint rotations and root joint translation, respectively. \\(j = 4\\) indicates 4 skeletal joints w.r.t. the head, neck, and two eyes. We perform a standard linear blend skinning (LBS) function [43] (with per-vertex weights \\(\\mathbf{W} \\in \\mathbb{R}^{j \\times n_v}\\)) to rotate \\(\\mathbf{S}\\) about joint locations \\(\\mathbf{J} \\in \\mathbb{R}^{3j}\\) by \\(\\mathbf{p}\\) to obtain \\(\\mathbf{S_p}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.858, + 0.891, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {\\mathbf {p}} = \\operatorname {L B S} (\\mathbf {S}, \\mathbf {p}, \\mathbf {J}; \\mathbf {W}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}\\) is the joint locations in the bind pose determined" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9089" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.123 + ], + "angle": 0, + "content": "by \\(\\mathbf{J} = \\mathcal{J}(\\beta):\\mathbb{R}^{|\\beta |}\\to \\mathbb{R}^{3j}\\). Then we use an orthographic camera model to project 3D vertices in \\(\\mathbf{S}_{\\mathbf{p}}\\) to the 2D plane." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.129, + 0.47, + 0.175 + ], + "angle": 0, + "content": "Illumination Model. We follow previous work [21] to use Spherical Harmonics (SH) [56] to estimate the illumination of a given image. The shaded texture \\(\\mathbf{T}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.183, + 0.47, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = \\mathbf {A} \\odot \\sum_ {k = 1} ^ {9} \\gamma_ {k} \\Psi_ {k} (\\mathbf {N}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.469, + 0.263 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the Hadamard product, \\(\\mathbf{N}\\) is the surface normal of \\(\\mathbf{S}\\) in UV coordinates, \\(\\Psi : \\mathbb{R}^3 \\to \\mathbb{R}\\) are SH basis function and \\(\\gamma \\in \\mathbb{R}^9\\) is the corresponding SH coefficient." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.27, + 0.266, + 0.285 + ], + "angle": 0, + "content": "3.2. Overview of HiFace" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.469, + 0.373 + ], + "angle": 0, + "content": "Key Idea. The key idea of HiFace is to explicitly model the static (e.g., person-specific properties) and dynamic (e.g., expression-driven wrinkles) details, allowing the model to reconstruct a high-fidelity 3D face from a single image with realistic and animatable details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.577 + ], + "angle": 0, + "content": "Overview. The goal of HiFace is to reconstruct 3D shapes with realistic details from a single image. The overview of HiFace is illustrated in Fig. 2(a). We leverage a feature extractor (i.e., ResNet-50 [30]) to regress corresponding coefficients from an input image. Our model jointly predicts both the coarse-level shapes and the fine-level details. For coarse-level shapes, we regress shape parameters (i.e., identity, expression, albedo, illumination, and pose) of a parametric face model. For the fine-level details, we propose a novel way to model it through the separation of static and dynamic factors and formulate the generation of details into the problems of 3DMM coefficients regression and displacement maps interpolation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.728 + ], + "angle": 0, + "content": "Note that the facial details are based on the coarse shape, we thereby exploit novel loss functions to learn 3D representations of coarse shape and details simultaneously from the synthetic dataset with ground-truth labels. To generalize our model to real-world images, we also present several self-supervised losses to train the model with both synthetic data and real-world images coherently. As a result, HiFace can faithfully reconstruct the facial details of a given image, or animate a face by combining the decoupled static and dynamic coefficients that come from different individuals." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.736, + 0.414, + 0.753 + ], + "angle": 0, + "content": "3.3. Decoupling Static and Dynamic Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.469, + 0.805 + ], + "angle": 0, + "content": "We propose Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). The facial details are basically composed of a static factor and a dynamic factor:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.815, + 0.469, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} = \\mathbf {D} _ {\\mathrm {s t a}} + \\mathbf {D} _ {\\mathrm {d y n}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.871 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}_{\\mathrm{sta}}\\) and \\(\\mathbf{D}_{\\mathrm{dyn}}\\) indicate details from static and dynamic factors, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Concretely, the static factor is the inherent property of the identity (i.e., the given 2D face), and originates from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "appearance and age attributes. As for the dynamic factor, it is typically driven by the expression and influenced by person-specific properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.144, + 0.892, + 0.25 + ], + "angle": 0, + "content": "Static Detail Generation. To simplify the problem, we are inspired by 3DMMs, which parameterize the statistical models to simplify the 2D-to-3D problem. We build a 300-dim displacement basis \\(\\mathbf{B}_{\\mathrm{sta}}\\) from the captured 332 scans [57] by PCA [1]. The scans contain diverse age groups in a neutral expression. Then we regress the coefficient \\(\\varphi\\) to synthesize the static detail \\(\\mathbf{D}_{\\mathrm{sta}}\\) from the image:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.264, + 0.892, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {s t a}} = \\bar {\\mathbf {D}} _ {\\mathrm {s t a}} + \\varphi \\mathbf {B} _ {\\mathrm {s t a}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.892, + 0.317 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{D}}_{\\mathrm{sta}}\\) and \\(\\mathbf{B}_{\\mathrm{sta}}\\) are the mean displacement map and displacement basis for static details, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Dynamic Detail Generation. Due to the high diversity and complexity of expression representation, directly generating dynamic details from expression is quite difficult. Therefore we simplify the expression representation by using an interpolation between two displacement maps: compressed and stretched [57]. For example, the compressed expression may indicate a state of frowning to the extreme, while the stretched expression may indicate a state of complete relaxation between the eyebrows. Other states of this area can be interpolated by these two polarized states." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Consequently, we generate the dynamic details through compressed and stretched displacement maps. Again, we build 26-dim compressed \\(\\mathbf{B}_{\\mathrm{com}}\\) and stretched \\(\\mathbf{B}_{\\mathrm{str}}\\) displacement bases by PCA [1] to simplify the generation of displacement maps. To generate the dynamic coefficients \\(\\phi = \\{\\phi_{\\mathrm{com}},\\phi_{\\mathrm{str}}\\}\\), we apply the expression coefficient \\(\\xi\\) into the static coefficient \\(\\varphi\\) through AdaIN [32], followed by the MLP transformation \\(\\Phi\\) to obtain \\(\\phi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.605, + 0.891, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\phi = \\Phi \\left(\\sigma (\\tilde {\\xi}) \\left(\\frac {\\varphi - \\mu (\\varphi)}{\\sigma (\\varphi)} + \\mu (\\tilde {\\xi})\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.891, + 0.709 + ], + "angle": 0, + "content": "where \\(\\tilde{\\xi}\\) is the affine vector from \\(\\xi\\) via MLP transformation. \\(\\mu\\) and \\(\\sigma\\) indicate the mean and standard deviation. \\(\\phi_{\\mathrm{com}}\\) and \\(\\phi_{\\mathrm{str}}\\) are coefficients for compressed and stretched displacement maps respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.891, + 0.739 + ], + "angle": 0, + "content": "Similar to Eq. 5, the compressed and stretched displacement maps are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.749, + 0.891, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\text {c o m}} = \\bar {\\mathbf {D}} _ {\\text {c o m}} + \\phi_ {\\text {c o m}} \\mathbf {B} _ {\\text {c o m}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.769, + 0.76, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {s t r}} = \\bar {\\mathbf {D}} _ {\\mathrm {s t r}} + \\phi_ {\\mathrm {s t r}} \\mathbf {B} _ {\\mathrm {s t r}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.855 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{D}}_{\\mathrm{com}}\\) and \\(\\mathbf{B}_{\\mathrm{com}}\\) are the mean displacement map and 26-dim displacement basis for compressed detail, and \\(\\bar{\\mathbf{D}}_{\\mathrm{str}}\\) and \\(\\mathbf{B}_{\\mathrm{str}}\\) are the mean displacement map and 26-dim displacement basis for stretched detail, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Considering the coarse shape \\(\\mathbf{S}\\) can be obtained by deforming the neutral shape \\(\\mathbf{S}_{\\mathrm{neu}}\\) with the expression component \\(\\xi \\mathbf{B}_{\\mathrm{exp}}\\), such expression-driven deformation over face" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9090" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.089, + 0.114, + 0.098 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.089, + 0.186, + 0.097 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.089, + 0.302, + 0.098 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.089, + 0.425, + 0.097 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.1, + 0.129, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.1, + 0.226, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.1, + 0.273, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.1, + 0.37, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.1, + 0.468, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.176, + 0.129, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.13, + 0.176, + 0.226, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.176, + 0.37, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.176, + 0.468, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.262, + 0.47, + 0.346 + ], + "angle": 0, + "content": "Figure 3. Illustration of displacement map composition in SD-DeTail. Given (a). an image (top) to reconstruct its coarse shape (bottom), we formulate the detail as (b). a static factor and (c). a dynamic factor interpolated by polarized states w.r.t. compressed (top) and stretched (bottom). (d). the output displacement map is linearly combined by (b) and (c) to present vivid details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.469, + 0.45 + ], + "angle": 0, + "content": "shape yields the \"tension\" over each vertex [57], which influences facial details from expression. Since \\(\\mathbf{S}_{\\mathrm{neu}}\\) and \\(\\mathbf{S}\\) possesses the same topology, for each vertex \\(\\mathbf{v}_i \\in \\mathbf{S}\\) with \\(K\\)-edges \\(E_i = \\{e_1, \\dots, e_K\\}\\) connected with \\(\\mathbf{v}_i\\), \\(E_i' = \\{e_1', \\dots, e_K'\\}\\) are the corresponding edges in \\(\\mathbf{S}_{\\mathrm{neu}}\\) that are connected to \\(\\mathbf{v}_i'\\). Then the tension at \\(\\mathbf{v}_i\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.462, + 0.469, + 0.495 + ], + "angle": 0, + "content": "\\[\nt _ {\\mathbf {v} _ {i}} = 1 - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\left\\| e _ {k} \\right\\|}{\\left\\| e _ {k} ^ {\\prime} \\right\\|}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.507, + 0.469, + 0.553 + ], + "angle": 0, + "content": "where \\( \\|\\cdot\\| \\) represents the edge length. Positive values of \\( t_{\\mathbf{v}_i} \\) indicate compression, negative values indicate stretch, and 0-value indicates no change, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.469, + 0.599 + ], + "angle": 0, + "content": "The vertex tension \\(t_{\\mathbf{v}_i}\\) in \\(\\mathbf{S}\\) composes the tension map \\(\\mathbf{M}_{\\mathrm{uv}}\\) in UV coordinates. Then, the displacement map of the dynamic detail is the linear interpolation of \\(\\mathbf{D}_{\\mathrm{com}}\\) and \\(\\mathbf{D}_{\\mathrm{str}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.611, + 0.469, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\mathrm {d y n}} = \\mathbf {M} _ {\\mathrm {u v}} ^ {+} \\odot \\mathbf {D} _ {\\mathrm {c o m}} + \\mathbf {M} _ {\\mathrm {u v}} ^ {-} \\odot \\mathbf {D} _ {\\mathrm {s t r}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.469, + 0.716 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{\\mathrm{uv}}^{+}\\) and \\(\\mathbf{M}_{\\mathrm{uv}}^{-}\\) indicate the positive and negative value of \\(\\mathbf{M}_{\\mathrm{uv}}\\), respectively. Fig. 3 shows the effectiveness of SD-DeTail. The dynamic factor interpolated by two polarized states introduces expression-related details and further decorates the static detail, yielding the final vivid output." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.726, + 0.292, + 0.74 + ], + "angle": 0, + "content": "3.4. Overall Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We propose several loss functions to train HiFace end-to-end. Specifically, we use static and dynamic detail losses to supervise the synthesized displacement maps from \\(\\varphi\\) and \\(\\phi\\). In addition, we leverage the coarse shape loss to supervise the reconstructed shape from \\(\\beta\\) and \\(\\xi\\). Finally, we follow previous methods [21, 24, 73] to leverage the differentiable renderer [28] to map the generated 3D shape into 2D images, by combining \\(\\alpha\\), \\(\\beta\\), \\(\\xi\\), \\(\\gamma\\), \\(\\mathbf{p}\\), \\(\\varphi\\), \\(\\phi\\). Then, we perform self-supervised losses to train in both synthetic and real-world images. See more details in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Static and Dynamic Detail Losses. To explicitly train the details of each component, we leverage the ground-truth annotations from the synthetic dataset [57, 72] as supervision to assist the training process of our model. Specifically, we calculate the detail losses by estimating the \\(l_{2}\\) distance between the reconstructed displacement maps and ground-truth w.r.t. static, compressed, and stretched components, and summarize them as \\(\\mathcal{L}_{\\mathrm{detail}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.222, + 0.797, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s t a}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {s t a}} - \\hat {\\mathbf {D}} _ {\\mathrm {s t a}}\\right) \\right\\| _ {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.245, + 0.891, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o m}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {c o m}} - \\hat {\\mathbf {D}} _ {\\mathrm {c o m}}\\right) \\right\\| _ {2}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.268, + 0.793, + 0.286 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s t r}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {s t r}} - \\hat {\\mathbf {D}} _ {\\mathrm {s t r}}\\right) \\right\\| _ {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.289, + 0.752, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e t a i l}} = \\mathcal {L} _ {\\text {s t a}} + \\mathcal {L} _ {\\text {c o m}} + \\mathcal {L} _ {\\text {s t r}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.315, + 0.892, + 0.361 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{\\mathrm{detail}}\\) is the facial mask in the UV coordinates, and \\(\\hat{\\mathbf{D}}_{\\mathrm{sta}} / \\hat{\\mathbf{D}}_{\\mathrm{com}} / \\hat{\\mathbf{D}}_{\\mathrm{str}}\\) and \\(\\mathbf{D}_{\\mathrm{sta}} / \\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}\\) are the reconstructed and ground-truth displacement maps, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Coarse Shape Losses. Since the details should be based on realistic coarse shapes, we train the coarse shape to help the learning of details by leveraging the ground-truth vertex as supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.438, + 0.891, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v e r}} = \\left\\| \\mathbf {M} _ {\\text {v e r}} \\odot (\\mathbf {S} - \\hat {\\mathbf {S}}) \\right\\| _ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.891, + 0.499 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{\\mathrm{ver}}\\) is frontal face area of the coarse shape. \\(\\hat{\\mathbf{S}}\\) and \\(\\mathbf{S}\\) are the reconstructed and ground-truth face by Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.5, + 0.891, + 0.545 + ], + "angle": 0, + "content": "In addition, we make constraints on shape coefficients to prevent overfitting. We enforce the predicted coefficients have a similar distribution to the ground-truth coefficients:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.555, + 0.891, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k l}} = \\rho (\\boldsymbol {\\beta}) \\left(\\log \\rho (\\boldsymbol {\\beta}) - \\log \\rho (\\hat {\\boldsymbol {\\beta}})\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.89, + 0.614 + ], + "angle": 0, + "content": "where \\(\\rho\\) denotes softmax function to map the predicted coefficients \\(\\hat{\\beta}\\) and ground-truth \\(\\beta\\) into probability distribution." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.614, + 0.808, + 0.629 + ], + "angle": 0, + "content": "Finally, the shape loss is \\(\\mathcal{L}_{\\mathrm{shp}} = \\mathcal{L}_{\\mathrm{ver}} + \\mathcal{L}_{\\mathrm{kl}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.712 + ], + "angle": 0, + "content": "Self-supervised Losses. To encourage the generalization of our models in real-world images [40, 52], we follow previous methods [24, 21, 73] to leverage self-supervised loss \\(\\mathcal{L}_{\\mathrm{self}}\\) for all training images, including photo loss \\(\\mathcal{L}_{\\mathrm{pho}}\\), perceptual loss \\(\\mathcal{L}_{\\mathrm{id}}\\), and dense landmark loss \\(\\mathcal{L}_{\\mathrm{lmk}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.723, + 0.891, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\text {p h o}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}} + \\lambda_ {\\mathrm {l m k}} \\mathcal {L} _ {\\mathrm {l m k}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.89, + 0.779 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{id}}\\) and \\(\\lambda_{\\mathrm{lmk}}\\) are weights to balance the self-supervised losses term." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In addition, considering the static detail heavily correlates to person-specific age attribute, inspired by [19], we leverage the pre-trained age prediction network [36] to learn high-level representations of static details through knowledge distillation, such that the learned coefficients exhibit expressive results. To achieve this, we use several MLP layers on the static coefficient \\(\\varphi\\), and map it into age classification probabilities \\(\\hat{\\mathbf{p}}_{\\mathrm{age}}\\). Then we use the pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9091" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.087, + 0.895, + 0.158 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of 3D face reconstruction methods on REALY benchmark. “-c” and “-d” indicate coarse and detail shape, respectively. @RN/@RM/@RF/@RC/all indicate errors in nose/mouth/forehead/cheek/all regions. We highlight the best method for the two groups respectively. HiFace achieves the best reconstruction performance in the overall error by a large margin. Each component in HiFace contributes to a better reconstruction quality. The reconstructed details of HiFace further boost the quality while previous methods [24, 19] modeling details with only image-level supervision even deteriorate the reconstruction accuracy." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.158, + 0.891, + 0.354 + ], + "angle": 0, + "content": "
GroupMethods / e (mm)frontal-viewside-view
@RN@RM@RF@RCall@RN@RM@RF@RCall
CoarseDeep3D [21]1.719±0.3541.368±0.4392.015±0.4491.528±0.5011.6571.749±0.3431.411±0.3952.074±0.4861.528±0.5171.691
MGCNet [64]1.771±0.3801.417±0.4092.268±0.5031.639±0.6501.7741.827±0.3831.409±0.4182.248±0.5081.665±0.6441.787
3DDFA-v2 [29]1.903±0.5171.597±0.4782.447±0.6471.757±0.6421.9261.883±0.4991.642±0.5012.465±0.6221.781±0.6361.943
DECA-c [24]1.694±0.3552.516±0.8392.394±0.5761.479±0.5352.0101.903±1.0502.472±1.0792.423±0.7201.630±1.1352.107
SADRNet [61]1.791±0.5421.591±0.4882.413±0.5371.856±0.7011.9131.771±0.5211.560±0.4622.490±0.5662.010±0.7151.958
EMOCA-c [19]1.868±0.3872.679±1.1122.426±0.6411.438±0.5012.1031.867±0.5542.636±1.2842.448±0.7081.548±0.5902.125
MICA [81]1.585±0.3253.478±1.2042.374±0.6831.099±0.3242.1341.525±0.3223.567±1.2122.379±0.6751.109±0.3252.145
Ours-c (w/o Syn. Data)†1.227±0.4071.787±0.4391.454±0.3821.762±0.4361.5581.187±0.3791.826±0.4901.470±0.4261.653±0.4501.534
Ours-c1.054±0.3171.461±0.4301.331±0.3471.342±0.3841.2970.992±0.2461.505±0.4541.427±0.4001.439±0.4291.341
DetailDECA-d [24]2.138±0.4612.802±0.8682.457±0.5591.443±0.4982.2102.286±1.1032.684±1.0412.519±0.7181.555±0.8222.261
EMOCA-d [19]2.532±0.5392.929±1.1062.595±0.6311.495±0.4692.3882.455±0.6362.948±1.2922.606±0.6861.599±0.5632.402
HRN [42]1.722±0.3301.357±0.5231.995±0.4761.072±0.3331.5371.642±0.3101.285±0.5281.906±0.4791.038±0.3221.468
Ours-d (w/o Syn. Data)†1.465±0.5571.790±0.4251.528±0.3731.618±0.3621.6001.422±0.5371.849±0.4731.530±0.4141.572±0.3991.594
Ours-d (w/o static)*1.055±0.2901.469±0.4151.336±0.3371.319±0.3741.2951.004±0.2331.491±0.4371.418±0.3921.418±0.4151.332
Ours-d (w/o dynamic)*1.069±0.3181.469±0.4141.358±0.3361.270±0.3441.2920.991±0.2391.496±0.4371.411±0.3931.375±0.4021.318
Ours-d1.036±0.2801.450±0.4131.324±0.3341.291±0.3621.2750.985±0.2371.489±0.4361.399±0.3881.360±0.3951.308
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.081, + 0.355, + 0.707, + 0.366 + ], + "angle": 0, + "content": "To align the dataset scale, w/o Syn. Data indicates we train the model without using the ground-truth labels from the synthetic dataset." + }, + { + "type": "table_footnote", + "bbox": [ + 0.083, + 0.366, + 0.891, + 0.377 + ], + "angle": 0, + "content": "* To eliminate the bias of coarse shape in estimating the reconstruction error, we fix the coarse shape and train the details with/without static and dynamic factors for comparisons." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.355, + 0.891, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.443 + ], + "angle": 0, + "content": "age recognition model \\(\\Gamma_{\\mathrm{age}}\\) to obtain the probabilities of the given input image \\(\\mathbf{I}\\). The distillation loss \\(\\mathcal{L}_{\\mathrm{kd}}\\) enforces the probabilities between \\(\\hat{\\mathbf{p}}_{\\mathrm{age}}\\) and \\(\\Gamma_{\\mathrm{age}}(\\mathbf{I})\\) be similar:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.452, + 0.469, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k d}} = \\Gamma_ {\\text {a g e}} (\\mathbf {I}) \\left(\\log \\Gamma_ {\\text {a g e}} (\\mathbf {I}) - \\log \\hat {\\mathbf {p}} _ {\\text {a g e}}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.478, + 0.47, + 0.516 + ], + "angle": 0, + "content": "Regularization. \\(\\mathcal{L}_{\\mathrm{reg}}\\) regularizes coefficients of each submodule, by minimizing the \\(l_{2}\\) loss of \\(\\alpha, \\beta, \\xi, \\varphi, \\phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.582 + ], + "angle": 0, + "content": "Overall Loss Function. We train the coarse shape and fine details simultaneously, such that each component can collaborate to reconstruct high-fidelity 3D faces with realistic details. Formally, we minimize the total loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.593, + 0.469, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {d e t a i l}} \\mathcal {L} _ {\\text {d e t a i l}} + \\lambda_ {\\text {s h p}} \\mathcal {L} _ {\\text {s h p}} \\tag {15} \\\\ + \\lambda_ {\\text {s e l f}} \\mathcal {L} _ {\\text {s e l f}} + \\lambda_ {\\text {k d}} \\mathcal {L} _ {\\text {k d}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.637, + 0.361, + 0.652 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weight for each component." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.21, + 0.682 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.689, + 0.295, + 0.705 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Dataset. We use a hybrid dataset made up from both synthetic [72, 57] and real data [40, 52]. We use the synthetic data pipeline [72, 57] to generate a diverse dataset of \\(200k\\) faces with ground-truth vertex, landmark, albedo, and displacement map annotations. The real-world datasets contain \\(400k\\) images in total from diverse age, gender, and ethnicity groups. For the real-world dataset, we use the pretrained dense landmark detector [72] to detect 669 landmarks for training. We use face parsing [79] to generate and select region-of-interest as facial masks, providing robustness to common occlusions by hair or other accessories. We follow [19, 73, 21] to split the dataset into training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.892, + 0.427 + ], + "angle": 0, + "content": "and validation sets. The test images are from CelebA [40], FFHQ [37], LS3D-W [8], and AFLW2000 [78]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.434, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Implementation Details. We implement HiFace in PyTorch [54] and leverage the PyTorch3D differentiable rasterizer [35] for rendering. We train our model for 35 epochs on \\(8 \\times\\) NVIDIA Tesla V100 GPUs with a mini-batch of 320. We use the pre-trained ResNet-50 on ImageNet [20] as initialization, and use Adam [39] as optimizer with an initial learning rate of \\(1e - 4\\). The input image is cropped and aligned by [14], and resized into \\(224 \\times 224\\). We empirically set \\(\\lambda_{\\mathrm{detail}} = 10\\), \\(\\lambda_{\\mathrm{shp}} = 1\\), \\(\\lambda_{\\mathrm{self}} = 1\\), \\(\\lambda_{\\mathrm{id}} = 0.1\\), \\(\\lambda_{\\mathrm{lmk}} = 0.5\\), \\(\\lambda_{\\mathrm{kd}} = 1\\), \\(\\lambda_{\\mathrm{reg}} = 1e - 3\\) throughout the experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.597, + 0.782, + 0.614 + ], + "angle": 0, + "content": "4.2. Comparisons to State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.894, + 0.794 + ], + "angle": 0, + "content": "Quantitative Comparison. We perform the quantitative evaluation on the REALY benchmark [12], which contains 100 frontal-view and 400 side-view images from 100 textured-scans [17]. The REALY benchmark presents a region-aware evaluation pipeline to separately evaluate the metric error (in mm) of the nose, mouth, forehead, and cheek regions. Such an evaluation pipeline is demonstrated to better estimate the actual similarity of the 3D faces and align with human perception. We compare HiFace to previous state-of-the-art methods and report the region-wise and average normalized mean square error (NMSE) in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "As Tab. 1 illustrates, HiFace outperforms prior arts in the overall error by a large margin. HiFace balances the reconstruction quality on each region, compared to those optimum region methods that may fail in specific regions (e.g., MICA [81] fails in mouth region while HRN [42] fails in forehead region). Note that HiFace faithfully recovers the facial details, thus making the reconstruction error smaller" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9092" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.088, + 0.455, + 0.098 + ], + "angle": 0, + "content": "Input Deep3D 3DDFA-v2 SynergyNet MICA Dense Ours" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.098, + 0.471, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.282, + 0.472, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Comparison on coarse shape reconstruction. From left to right: Input image, Deep3D [21], 3DDFA-v2 [29], Synerg-yNet [74], MICA [81], Dense [73], and HiFace (Ours)." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.512 + ], + "angle": 0, + "content": "Figure 6. Illustration on the flexibility of SD-DeTail. Given the identity and expression coefficients \\((\\beta, \\xi)\\) from the optimization-based method [73], SD-DeTail can generate realistic details based on the coarse shape and further improve the visual quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.469, + 0.587 + ], + "angle": 0, + "content": "than using the coarse shape alone. As a comparison, although DECA [24] and EMOCA [19] can reconstruct details of given images, they turn out to be noisy, leading to the deterioration of reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.756 + ], + "angle": 0, + "content": "In addition, Tab. 1 also demonstrates the necessity of each component in contributing to a better quality. It can be observed that the synthetic data with ground-truth labels not only improve the coarse shape reconstruction quality but is also crucial for detailed reconstruction. With the synthetic data, the proposed SD-DeTail further boosts the overall reconstruction quality. Both the static and dynamic factors are essential to capture fine-grained details, and the final SD-DeTail achieves the most accurate details in expression-related regions such as the mouth and forehead, which contributes to the overall gains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Qualitative Comparison. Given a single face image, Hi-Face reconstructs a high-fidelity 3D shape with details. We present comparisons with previous methods on 1). coarse shape reconstruction [21, 29, 74, 81, 73] in Fig. 4 and 2). detail reconstruction [77, 15, 24, 19, 70] in Fig. 5. See more examples and comparisons in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For coarse shape in Fig. 4, our HiFace faithfully recovers the coarse shape of the given identity and outperforms the previous learning-based methods, and is on par" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.088, + 0.888, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.282, + 0.889, + 0.325 + ], + "angle": 0, + "content": "Figure 5. Comparison on detail shape reconstruction. From left to right: Input image, FaceScape [77], Unsup [15], DECA [24], EMOCA [19], FaceVerse [70], and HiFace (Ours)." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.334, + 0.89, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.741 + ], + "angle": 0, + "content": "Figure 7. Comparison on face animation. Given a source image (yellow box), we use the driving images (green box) to drive its expressions. DECA [24] (2nd-row) and EMOCA [19] (3rd-row) can animate the expression-driven details but lack realistic. As a comparison, HiFace is flexible to animate details from static (4th-row), dynamic (5th-row), or both (6th-row) factors, and presents vivid animation quality with realistic shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.882 + ], + "angle": 0, + "content": "with Dense [73], which is the state-of-the-art optimization-based method. For detailed reconstruction in Fig. 5, previous methods [24, 19] fail to reconstruct satisfactory details. Several methods [15, 77, 70] are sensitive to occlusions and large poses. As a comparison, HiFace achieves the most realistic reconstruction quality, and faithfully recovers facial details of a given image, which significantly outperforms previous methods by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition, given an image and the fitted coefficients \\(\\beta\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9093" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.473, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.089, + 0.888, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.893, + 0.259 + ], + "angle": 0, + "content": "Figure 8. Ablation studies on loss functions and training data. The coarse shape losses \\(\\mathcal{L}_{\\mathrm{shp}} / \\mathcal{L}_{\\mathrm{kl}}\\) (left), detail losses \\(\\mathcal{L}_{\\mathrm{detail}} / \\mathcal{L}_{\\mathrm{kd}}\\) (right), and hybrid datasets coherently contribute to the reconstruction quality of coarse shapes and details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.47, + 0.346 + ], + "angle": 0, + "content": "\\(\\xi\\) from the optimization-based methods such as Dense [73], SD-DeTail synthesizes the details and further strengthens the quality compared to the coarse shape (see Fig. 6). It shows SD-DeTail is flexible and can be easily plugged-and-play into other methods. See more in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.469, + 0.429 + ], + "angle": 0, + "content": "Application of HiFace. The HiFace explicitly decouples the static and dynamic details through the proposed SD-DeTail. Therefore, we can animate the facial attributes by simply assigning the expression coefficient \\(\\xi\\) and/or static coefficient \\(\\varphi\\) of the driving images to the source images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.595 + ], + "angle": 0, + "content": "In Fig. 7, we demonstrate the animation quality of HiFace outperforms the previous state-of-the-art detail animation methods [24, 19]. It shows that while DECA [24] and EMOCA [19] can animate the expression-driven details but lack realistic, the proposed HiFace is flexible to manipulate the static, dynamic, or both details. When animating the static detail, the person-specific details can be well transferred into the source shape. When animating the dynamic detail, only expression-dependent details are presented. Finally, we can also animate the static and dynamic details simultaneously and achieve satisfactory results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.241, + 0.626 + ], + "angle": 0, + "content": "5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Ablation Studies on Loss Functions and Datasets. We present ablation studies on the proposed loss functions and training strategy with hybrid datasets. We train HiFace with synthetic data alone and compare it to the one trained with hybrid datasets. For coarse shape reconstruction, we investigate the contribution of \\(\\mathcal{L}_{\\mathrm{shp}}\\) and its sub-term \\(\\mathcal{L}_{\\mathrm{kl}}\\). For detail reconstruction, we compare HiFace without \\(\\mathcal{L}_{\\mathrm{detail}}\\) and \\(\\mathcal{L}_{\\mathrm{kd}}\\), respectively. The results are presented in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Fig. 8 demonstrates that the proposed loss functions and training strategy from hybrid datasets contribute to satisfactory coarse shape and details. First, the model trained with synthetic data alone cannot generalize well to real-world images, which indicates the necessity to train with real-world data. Second, \\(\\mathcal{L}_{\\mathrm{shp}}\\) improves the coarse shape reconstruction quality. \\(\\mathcal{L}_{\\mathrm{shp}}\\) is effective in tackling challenging poses and improving alignment. \\(\\mathcal{L}_{\\mathrm{kl}}\\) can relieve the overfitting risk on the synthetic data and improve the gener" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.265, + 0.892, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.408, + 0.893, + 0.505 + ], + "angle": 0, + "content": "Figure 9. Ablation studies on SD-DeTail. Results show that directly synthesizing the static or dynamic details is rather challenging, leading to unreasonable coarse shapes and details (SD-1, SD-2, and SD-3). As a comparison, we leverage the statistical bases with SD-DeTail and regard the detail generation problem as a coefficients regression and interpolation problem, leading to more realistic details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.577 + ], + "angle": 0, + "content": "alization to real-world images. Third, without \\(\\mathcal{L}_{\\mathrm{detail}}\\) or \\(\\mathcal{L}_{\\mathrm{kd}}\\), the reconstructed details exhibit random noise and cannot faithfully reflect person-specific details. Such noise misses the correspondence to the person-specific identity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Ablation Studies on SD-DeTail. To verify the effectiveness of building bases for static and dynamic details, we present detailed ablation studies on SD-DeTail, by replacing the bases (i.e., \\(\\mathbf{B}_{\\mathrm{sta}}\\) and \\(\\mathbf{B}_{\\mathrm{com}} / \\mathbf{B}_{\\mathrm{str}}\\)) reconstruction with a U-Net decoder [60] (same as DECA [24]). Therefore, the model learns to directly synthesize displacement maps instead of predicting corresponding coefficients like ours. In Fig. 9, we make comparisons on: 1). directly synthesizing \\(\\mathbf{D}_{\\mathrm{dyn}}\\) (SD-1), 2). directly synthesizing \\(\\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}\\) and interpolating via Eq. 9 (SD-2), 3). directly synthesizing \\(\\mathbf{D}_{\\mathrm{sta}}\\) (SD-3), 4). SD-DeTail (Ours). It can be seen that, due to the high diversity and complexity of expression representation, it is hard to directly learn realistic details even with ground-truth labels of \\(\\mathbf{D}_{\\mathrm{dyn}}\\) from synthetic data (see SD-1, SD-2 and SD-3 in Fig. 9). More specifically, for the static details, directly synthesizing displacement maps bring much noise (SD-3). For example, the hollow eyebrow is demonstrated in the second row. For the dynamic details, directly synthesizing displacement maps even leads to unnatural results (SD-1 and SD-2). For example, the reconstructed 3D faces are distorted especially in the second row. We also notice" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9094" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.214 + ], + "angle": 0, + "content": "that, directly synthesizing \\(\\mathbf{D}_{\\mathrm{dyn}}\\) (SD-1) achieves inferior results than directly synthesizing \\(\\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}\\) and interpolating via Eq. 9 (SD-2). This demonstrates that it is beneficial to simplify the expression representation by using interpolation between two displacement maps (i.e., compressed and stretched). In conclusion, these observations further verify our insight on relaxing the challenging detail generation problem into a feasible coefficients regression problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.226, + 0.196, + 0.243 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.473, + 0.465 + ], + "angle": 0, + "content": "We propose HiFace to reconstruct high-fidelity 3D faces with realistic and animatable details from a single image. Our motivation and insights stand on the spirit of 3DMMs to simplify the challenging detail generation into more accessible regression and interpolation tasks. To achieve this, we elaborately design SD-DeTail to decouple the static and dynamic factors explicitly, and interpolate the dynamic details through vertex tension. We succeed in learning the coarse shape and details jointly by proposing several novel loss functions to train on synthetic and real-world data. Extensive experiments demonstrate that HiFace achieves state-of-the-art face reconstruction both quantitatively and qualitatively in the coarse shape and detail shape, and the details are well decoupled and naturally animatable." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.477, + 0.238, + 0.494 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.593 + ], + "angle": 0, + "content": "This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SZSTC Grant (JCYJ20190809172201639, WDZC2020-0820200655001), Shenzhen Key Laboratory (ZDSYS2021-0623092001004), and Beijing Key Lab of Networked Multimedia." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.606, + 0.175, + 0.621 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.726 + ], + "angle": 0, + "content": "[2] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.47, + 0.793 + ], + "angle": 0, + "content": "[3] Linchao Bao, Xiangkai Lin, Yajing Chen, Haoxian Zhang, Sheng Wang, Xuefei Zhe, Di Kang, Haozhi Huang, Xinwei Jiang, Jue Wang, et al. High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.794, + 0.471, + 0.861 + ], + "angle": 0, + "content": "[4] Volker Blanz, Sami Romdhani, and Thomas Vetter. Face identification across different poses and illuminations with a 3d morphable model. In Proceedings of fifth IEEE international conference on automatic face gesture recognition, pages 202-207. IEEE, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.862, + 0.47, + 0.889 + ], + "angle": 0, + "content": "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.631, + 0.471, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "conference on Computer graphics and interactive techniques, pages 187-194, 1999. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.12, + 0.892, + 0.173 + ], + "angle": 0, + "content": "[6] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2):233-254, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.174, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[7] Sofien Bouaziz, Yangang Wang, and Mark Pauly. Online modeling for realtime facial animation. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.894, + 0.281 + ], + "angle": 0, + "content": "[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.282, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[9] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Realtime high-fidelity facial performance capture. ACM Transactions on Graphics (ToG), 34(4):1-9, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[10] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[11] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5187-5196, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.445, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[12] Zenghao Chai, Haoxian Zhang, Jing Ren, Di Kang, Zhengzhuo Xu, Xuefei Zhe, Chun Yuan, and Linchao Bao. Really: Rethinking the evaluation of 3d face reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[13] Anpei Chen, Zhang Chen, Guli Zhang, Kenny Mitchell, and Jingyi Yu. Photo-realistic facial details synthesis from single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9429-9439, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[14] Dong Chen, Gang Hua, Fang Wen, and Jian Sun. Supervised transformer network for efficient face detection. In European Conference on Computer Vision, pages 122-138. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.621, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[15] Yajing Chen, Fanzi Wu, Zeyu Wang, Yibing Song, Yonggen Ling, and Linchao Bao. Self-supervised learning of detailed 3d face reconstruction. IEEE Transactions on Image Processing, 29:8696-8705, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[16] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[17] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.783, + 0.892, + 0.837 + ], + "angle": 0, + "content": "[18] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE international conference on computer vision, pages 3085-3093, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.837, + 0.892, + 0.905 + ], + "angle": 0, + "content": "[19] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.905 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9095" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[21] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.215, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[22] Abdallah Dib, Cédric Thébault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.47, + 0.365 + ], + "angle": 0, + "content": "[23] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[24] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[25] Zhongpai Gao, Juyong Zhang, Yudong Guo, Chao Ma, Guangtao Zhai, and Xiaokang Yang. Semi-supervised 3d face representation learning from unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 348-349, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.5, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[26] Pablo Garrido, Michael Zollhöfer, Dan Casas, Levi Valgaerts, Kiran Varanasi, Patrick Pérez, and Christian Theobalt. Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG), 35(3):1-15, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[27] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[28] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[29] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.813, + 0.469, + 0.881 + ], + "angle": 0, + "content": "[31] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG), 36(6):1-14, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.881, + 0.469, + 0.905 + ], + "angle": 0, + "content": "[32] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.905 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[33] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[34] Luo Jiang, Juyong Zhang, Bailin Deng, Hao Li, and Ligang Liu. 3d face reconstruction with geometry details from a single image. IEEE Transactions on Image Processing, 27(10):4756-4770, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[35] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. ACM New York, NY, USA, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[38] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.5, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[40] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[41] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware mesh decoder for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6100–6109, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[42] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[43] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 165-172, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.77, + 0.892, + 0.852 + ], + "angle": 0, + "content": "[44] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.852, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[45] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9096" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[46] Yue Li, Liqian Ma, Haoqiang Fan, and Kenny Mitchell. Feature-preserving detailed 3d face reconstruction from a single image. In Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production, pages 1-9, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.16, + 0.47, + 0.229 + ], + "angle": 0, + "content": "[47] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3d face reconstruction from in-the-wild images using graph convolutional networks. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 5891-5900, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.229, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[48] Jingwang Ling, Zhibo Wang, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Structure-aware editable morphable model for 3d facial detail animation and manipulation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.47, + 0.352 + ], + "angle": 0, + "content": "[49] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.352, + 0.47, + 0.42 + ], + "angle": 0, + "content": "[50] Huiwen Luo, Koki Nagano, Han-Wei Kung, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11662–11672, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.501 + ], + "angle": 0, + "content": "[51] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriia Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20942–20952, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.501, + 0.47, + 0.556 + ], + "angle": 0, + "content": "[52] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.47, + 0.595 + ], + "angle": 0, + "content": "[53] Araceli Morales, Gemma Piella, and Federico M Sukno. Survey on 3d face reconstruction from uncalibrated images. Computer Science Review, 40:100400, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[54] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[55] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[56] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.47, + 0.826 + ], + "angle": 0, + "content": "[57] Chirag Raman, Charlie Hewitt, Erroll Wood, and Tadas Baltrusaitis. Mesh-tension driven expression-based wrinkles for synthetic faces. In WACV, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.47, + 0.88 + ], + "angle": 0, + "content": "[58] Roger Blanco I Ribera, Eduard Zell, John P Lewis, Junyong Noh, and Mario Botsch. Facial retargeting with automatic range of motion alignment. ACM Transactions on graphics (TOG), 36(4):1-12, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.88, + 0.47, + 0.905 + ], + "angle": 0, + "content": "[59] Sami Romdhani, Volker Blanz, and Thomas Vetter. Face identification by fitting a 3d morphable model using linear" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.905 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "shape and texture error functions. In European Conference on Computer Vision, pages 3-19. Springer, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.12, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[60] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.188, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[61] Zeyu Ruan, Changqing Zou, Longhai Wu, Gangshan Wu, and Limin Wang. Sadrnet: Self-aligned dual face regression networks for robust 3d dense face alignment and reconstruction. IEEE Transactions on Image Processing, 30:5793-5806, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.892, + 0.324 + ], + "angle": 0, + "content": "[62] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[63] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[64] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[65] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.514, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[66] Supasorn Suwajanakorn, Ira Kemelmacher-Shlizerman, and Steven M Seitz. Total moving face reconstruction. In European conference on computer vision, pages 796-812. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[67] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[68] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[69] Daniel Vlasic, Matthew Brand, Hanspeter Pfister, and Jovan Popovic. Face transfer with multilinear models. In ACM SIGGRAPH 2006 Courses, pages 24–es. ACM New York, NY, USA, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.771, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[70] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.84, + 0.892, + 0.895 + ], + "angle": 0, + "content": "[71] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13289-13298, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9097" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[72] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.175, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[73] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.257, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[74] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[75] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[76] Shugo Yamaguchi, Shunsuke Saito, Koki Nagano, Yajie Zhao, Weikai Chen, Kyle Olszewski, Shigeo Morishima, and Hao Li. High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.148, + 0.892, + 0.202 + ], + "angle": 0, + "content": "[78] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Towards large-posed face frontalization in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3990–3999, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.202, + 0.892, + 0.256 + ], + "angle": 0, + "content": "[79] Qi Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Computer Vision and Pattern Recognition, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.256, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[80] Wenbin Zhu, HsiangTao Wu, Zeyu Chen, Noranart Vesdapunt, and Baoyuan Wang. Reda: reinforced differentiable attribute for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4958-4967, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.323, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[81] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision (ECCV). Springer International Publishing, Oct. 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[82] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum. Wiley Online Library, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9098" + } + ] +] \ No newline at end of file diff --git a/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_origin.pdf b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bfe98ddc06a41436c1e524f7f5c1483bc19e2fc9 --- /dev/null +++ b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/a0404738-6577-4978-90b9-452e8141d3d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fcdb8a297d6190ee301617bc58420db84788085a2e52f6db35d29224e4351a9 +size 9562470 diff --git a/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/full.md b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0b41fd14b5a741685a8b3856f9892ad4f63a050f --- /dev/null +++ b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/full.md @@ -0,0 +1,439 @@ +# HiFace: High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details + +Zenghao Chai $^{1,2*}$ Tianke Zhang $^{2}$ Tianyu He $^{3}$ Xu Tan $^{3\dagger}$ Tadas Baltrušaitis $^{4}$ HsiangTao Wu $^{5}$ Runnan Li $^{5}$ Sheng Zhao $^{5}$ Chun Yuan $^{2}$ Jiang Bian $^{3}$ $^{1}$ National University of Singapore $^{2}$ Tsinghua University $^{3}$ Microsoft Research Asia + $^{4}$ Microsoft Mixed Reality & AI Lab $^{5}$ Microsoft Cloud + AI +zenghaochai@gmail.com ztk21@mails.tsinghua.edu.cn yuanc@sz.tsinghua.edu.cn +{tianyuhe,xuta,tabaltru,musclewu,runnan.li,sheng.zhao,jiang.bian}@microsoft.com + +# Abstract + +3D Morphable Models (3DMMs) demonstrate great potential for reconstructing faithful and animatable 3D facial surfaces from a single image. The facial surface is influenced by the coarse shape, as well as the static detail (e.g., person-specific appearance) and dynamic detail (e.g., expression-driven wrinkles). Previous work struggles to decouple the static and dynamic details through image-level supervision, leading to reconstructions that are not realistic. In this paper, we aim at high-fidelity 3D face reconstruction and propose HiFace to explicitly model the static and dynamic details. Specifically, the static detail is modeled as the linear combination of a displacement basis, while the dynamic detail is modeled as the linear interpolation of two displacement maps with polarized expressions. We exploit several loss functions to jointly learn the coarse shape and fine details with both synthetic and real-world datasets, which enable HiFace to reconstruct high-fidelity 3D shapes with animatable details. Extensive quantitative and qualitative experiments demonstrate that HiFace presents state-of-the-art reconstruction quality and faithfully recovers both the static and dynamic details. Our project page: https://project-hiface.github.io. + +# 1. Introduction + +The reconstruction of a 3D face from a single image has drawn much attention recently [67, 21, 41, 81]. It has tremendous potential applications like face recognition [11, 63, 4, 59], face animation [16, 75], virtual reality [7, 58, 31], etc. For example, the reconstructed 3D face representation can be driven by an audio [16], or a video from another person [38]. + +![](images/1c00cf10e1215744da80064e3797585206678ae0d1ae242aceac0c1dbe52a2ea.jpg) + +![](images/1195b9a15fd2f5a623241916e9f3741486d6c16a08677b268519183095717106.jpg) + +![](images/63b1b1b56817ba4333a72fb19fb4d33d1d40598e6a27343b136c4fa6a3f9f809.jpg) + +![](images/2a059dfce8796e1fc9e0a624d78c52c64d3bb7597dcd9d077e6c799d2749577b.jpg) + +![](images/4563679efa1a8ab5206d5c5548112ddb398c55259c32d140f515975398ef7922.jpg) +Figure 1. We propose HiFace to reconstruct high-fidelity 3D face with realistic and animatable details. Reconstruction: given a single image (1st-row), HiFace faithfully reconstructs a coarse shape (2nd-row) with vivid details (3rd-row). Animation: given a source face (yellow box), HiFace can animate the static (4th-row), dynamic (5th-row), or both (6th-row) details of the driving images (green box). Images are taken from FFHQ [37] and CelebA [40]. + +![](images/40004fb7ad32dc8431607682c055339918e9b028cdb0635b8ba37585066ce3fa.jpg) + +To build a flexible and animatable facial representation, a popular way is to leverage the success of 3D Morphable Models (3DMMs) [5, 6, 10, 45, 65], which decouple the influence of shape, expression, albedo, and others by modeling them in separate coefficients. Typically + +in literature, one can achieve coarse shape reconstruction in coefficients-fitting optimization [27, 73, 3, 77, 2], or an analysis-by-synthesis pipeline [67, 21, 81, 50]. As 3DMMs typically capture only the coarse facial geometry and are not capable of representing fine details (e.g., wrinkles), recent advances model such details with a displacement map [15, 77, 13, 9, 76]. However, previous work fails to model the distinction between static and dynamic factors of fine detail, leading to errors in reconstructions. For example, given that one may drive the expression of a young man from an old man, trivially transferring all wrinkles from the old man to the young man could make the young man look unnatural. In this sense, Feng et al. [24] implicitly leverages the person-specific identity and expression as conditions to generate the details. Although effective, they optimize the model in an analysis-by-synthesis pipeline with only the image-level supervision, leading to insufficient decoupling of static and dynamic details and inconsistent animation results (see Fig. 7). + +Therefore, we propose HiFace to explicitly model the static and dynamic details for high-fidelity 3D face reconstruction, by designing SD-DeTail module to decouple the static and dynamic factors. More specifically, for person-specific static detail, instead of directly predicting the displacement map that may increase the difficulty of detail prediction [24, 19], we follow the spirit of 3DMMs to build a displacement basis from the captured facial scans with age diversity [57, 72]. In this way, the model is trained to predict the coefficients of the displacement basis, and make the detail prediction easier. For dynamic detail, since it is highly expression-dependent, directly modeling it with one displacement basis is quite difficult. Therefore, based on the fact that the expression can be interpolated by a compressed and a stretched expressions [57], we build two displacement bases for the compressed and stretched expressions from the captured scans respectively, and learn to regress the displacement coefficients with the ground-truth labels. Therefore, we can obtain the dynamic detail by linearly interpolating the compressed and stretched displacement maps, which are derived from the displacement bases and the predicted coefficients. Finally, the predicted static and dynamic details are merged with the coarse shape to formulate the final output. + +Since we would like the final output to contain both the coarse shape and high-frequency detail, we propose several novel loss functions to learn coarse shape and details simultaneously from both the synthetic and real-world datasets. For details, we leverage the ground-truth static and dynamic displacement maps of the synthetic dataset [72, 57] as supervision. While for the coarse shape, we leverage the ground-truth vertex of the synthetic dataset as supervision. We also follow the previous methods [24, 21, 73] to leverage self-supervised losses for all training images. + +Overall, with the above insights and techniques, HiFace enables the reconstruction of high-fidelity 3D faces from a single image, and decouples static and dynamic details that are naturally animatable (see Fig. 1). We demonstrate that the proposed HiFace reconstructs realistic and faithful 3D faces, reaching state-of-the-art performance both quantitatively and qualitatively. In addition, HiFace is compatible with optimization-based methods [73], and is flexible to transfer vivid expressions and details from one person to another. In summary, our contributions are: + +- We propose HiFace to model the static and dynamic details explicitly, and demonstrate the benefits of synthetic data in decoupling the static and dynamic factors for detailed 3D face reconstruction. +- We propose novel loss functions in HiFace to learn 3D representations of coarse shape and fine details simultaneously from both the synthetic and real-world images. +- We achieve state-of-the-art reconstruction quality both quantitatively and qualitatively, with over $15\%$ performance gains in the region-aware benchmark [12]. +- We show that our SD-DeTail is easy to plug-and-play into optimization-based methods and can transfer expressions and details from one to another for face animation. + +# 2. Related Work + +3D face reconstruction from monocular images has received much attention in the past decades. Among them, 3D Morphable Models (3DMMs) are widely used to build 3D representations. Below we review the works that are related to them, and a full in-depth review can be found in recent surveys [82, 53, 23]. + +3D Morphable Model (3DMMs) [23] are statistical models widely used to constrain the distribution of 3D faces. The seminal work [5] presents 200 scans to generate shape and texture bases with Principal Component Analysis (PCA) [1], and formulate 3DMMs as linear models by the generated bases. After that, expression models [69, 45, 10] are proposed to support face manipulation. Recent advances [12, 55, 65, 18, 44] are proposed to expand the expressiveness of 3DMMs and play a crucial role in 3D face reconstruction. 3DMMs make it possible to simplify the 2D-to-3D problem into a regression task, which typically presents an analysis-by-synthesis fashion to estimate the coefficients of 3DMMs. In this paper, we follow the spirit of the 3DMMs family to present the decoupled static and dynamic details for 3D face reconstruction. + +Coarse Shape Reconstruction. Traditional optimization-based methods [27, 73, 3, 77, 2] directly optimize the 3DMM coefficients of given 2D images. While such methods work well in controlled settings (e.g., frontal view, no occlusion), they heavily rely on high-quality annotations. Learning-based methods leverage the advances of + +Figure 2. Illustration of HiFace. (a). Learning framework of HiFace. Given a monocular image, we regress its shape and detail coefficients to synthesize a realistic 3D face, and leverage a differentiable renderer [28] to train the whole model end-to-end from synthetic [72, 57] and real-world [40, 52] images. (b). The pipeline of Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). We explicitly decouple the static and dynamic factors to synthesize realistic and animatable details. Given the shape and static coefficients, we regress the static and dynamic details through displacement bases and interpolate them into the final details through vertex tension [57]. +![](images/57afa58db775560ad3b2a545497d24eb5ac60f6ae79d4b0bdca43cf3ec9370bd.jpg) +p: pose coef $\beta$ : identity coef $\xi$ : expression coef $\varphi$ : static coef $\alpha$ : albedo coef $\gamma$ : SH coef $\phi$ : dynamic coef + +![](images/f1bab0659a69cf4081953d6d51bcec4f778bbde8640851b31fa33b5e9f881ef7.jpg) + +CNNs [67, 21, 19, 62] and GCNs [47, 41, 25] to learn high-level representations from large-scale images in the wild. These methods show plausible generalization over diverse environments. To train the network end-to-end, recent methods leverage the differentiable renderers [28, 22, 80, 49], along with the photo loss, perceptual loss, and landmark loss [21, 19, 28, 67, 71] to optimize the network in a self-supervised manner. Different from these coarse shape reconstruction methods, we aim at high-fidelity 3D face reconstruction with both coarse shape and fine details. + +Detail Reconstruction. While 3DMMs can reconstruct coarse 3D face shapes from 2D images, they struggle with reconstructing fine-level details, such as forehead wrinkles and crows-feet. To fill this gap, shape by shading (SfS) [34, 46, 26, 66] methods reconstruct the facial details from monocular images or videos. However, these methods are sensitive to occlusions and large poses. Recent advances [24, 48, 15, 13, 42] leverage displacement maps to present details. These methods explicitly re-topologize the coarse shape and present residual bias to generate geometric details. The main challenge of detail reconstruction is the difficulty in learning the nuances and disentangling the static and dynamic details from only self-supervised learning. Ground-truth labels of the details are helpful to guide the learning process. However, it is difficult to obtain such fine-grained labels on real data. + +Synthetic Dataset. Several methods [81, 77, 33, 68, 51] utilize rendered faces or fitted coefficients to synthesize 3D-2D pairs. These ground-truth pairs lack diversity over background, illumination, and assets, making them hard to generalize well to real-world images. Recent advances in synthetic data generation [72, 57] demonstrate its ability to generalize to real-world settings, and diverse to compensate for the domain gap to real-world images. In this paper, we + +leverage high-quality data with ground-truth labels to explore the detailed 3D face reconstruction. + +# 3. Methodology + +# 3.1. Preliminary + +We adopt a common practice [21, 24] to represent a textured coarse shape with a 3D face model, an illumination model, and a camera model. + +3D Face Model. The 3D shape $\mathbf{S}$ and albedo $\mathbf{A}$ are represented by: + +$$ +\mathbf {S} = \bar {\mathbf {S}} + \beta \mathbf {B} _ {\mathrm {i d}} + \boldsymbol {\xi} \mathbf {B} _ {\exp} +$$ + +$$ +\mathbf {A} = \bar {\mathbf {A}} + \alpha \mathbf {B} _ {\mathrm {a l b}} +$$ + +where $\bar{\mathbf{S}}$ and $\bar{\mathbf{A}}$ are the mean shape and albedo. $\mathbf{B}_{\mathrm{id}}$ , $\mathbf{B}_{\mathrm{exp}}$ , and $\mathbf{B}_{\mathrm{alb}}$ are bases [73] of 256-dim identity, 233-dim expression, and 300-dim albedo, respectively. The coarse shape $\mathbf{S}$ in the bind pose is deformed from a neutral shape $\mathbf{S}_{\mathrm{neu}} = \bar{\mathbf{S}} + \alpha \mathbf{B}_{\mathrm{id}}$ with expression component $\xi \mathbf{B}_{\mathrm{exp}}$ . $\beta$ , $\xi$ , and $\alpha$ are the corresponding identity, expression, and albedo coefficients for generating a coarse shape. In this paper, the coarse shape $\mathbf{S}$ contains $n_v = 7$ , 667 vertices and $n_f = 14$ , 832 triangles with $512 \times 512 \times 3$ albedo. + +Pose & Camera Model. To estimate the face pose, we follow [72, 73] to predict skeletal pose $\mathbf{p} = [\theta | \mathbf{t}]$ , where $\theta \in \mathbb{R}^{3j}$ and $\mathbf{t} \in \mathbb{R}^3$ are the local joint rotations and root joint translation, respectively. $j = 4$ indicates 4 skeletal joints w.r.t. the head, neck, and two eyes. We perform a standard linear blend skinning (LBS) function [43] (with per-vertex weights $\mathbf{W} \in \mathbb{R}^{j \times n_v}$ ) to rotate $\mathbf{S}$ about joint locations $\mathbf{J} \in \mathbb{R}^{3j}$ by $\mathbf{p}$ to obtain $\mathbf{S_p}$ : + +$$ +\mathbf {S} _ {\mathbf {p}} = \operatorname {L B S} (\mathbf {S}, \mathbf {p}, \mathbf {J}; \mathbf {W}), \tag {2} +$$ + +where $\mathbf{J}$ is the joint locations in the bind pose determined + +by $\mathbf{J} = \mathcal{J}(\beta):\mathbb{R}^{|\beta |}\to \mathbb{R}^{3j}$ . Then we use an orthographic camera model to project 3D vertices in $\mathbf{S}_{\mathbf{p}}$ to the 2D plane. + +Illumination Model. We follow previous work [21] to use Spherical Harmonics (SH) [56] to estimate the illumination of a given image. The shaded texture $\mathbf{T}$ is computed as: + +$$ +\mathbf {T} = \mathbf {A} \odot \sum_ {k = 1} ^ {9} \gamma_ {k} \Psi_ {k} (\mathbf {N}), \tag {3} +$$ + +where $\odot$ denotes the Hadamard product, $\mathbf{N}$ is the surface normal of $\mathbf{S}$ in UV coordinates, $\Psi : \mathbb{R}^3 \to \mathbb{R}$ are SH basis function and $\gamma \in \mathbb{R}^9$ is the corresponding SH coefficient. + +# 3.2. Overview of HiFace + +Key Idea. The key idea of HiFace is to explicitly model the static (e.g., person-specific properties) and dynamic (e.g., expression-driven wrinkles) details, allowing the model to reconstruct a high-fidelity 3D face from a single image with realistic and animatable details. + +Overview. The goal of HiFace is to reconstruct 3D shapes with realistic details from a single image. The overview of HiFace is illustrated in Fig. 2(a). We leverage a feature extractor (i.e., ResNet-50 [30]) to regress corresponding coefficients from an input image. Our model jointly predicts both the coarse-level shapes and the fine-level details. For coarse-level shapes, we regress shape parameters (i.e., identity, expression, albedo, illumination, and pose) of a parametric face model. For the fine-level details, we propose a novel way to model it through the separation of static and dynamic factors and formulate the generation of details into the problems of 3DMM coefficients regression and displacement maps interpolation. + +Note that the facial details are based on the coarse shape, we thereby exploit novel loss functions to learn 3D representations of coarse shape and details simultaneously from the synthetic dataset with ground-truth labels. To generalize our model to real-world images, we also present several self-supervised losses to train the model with both synthetic data and real-world images coherently. As a result, HiFace can faithfully reconstruct the facial details of a given image, or animate a face by combining the decoupled static and dynamic coefficients that come from different individuals. + +# 3.3. Decoupling Static and Dynamic Details + +We propose Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). The facial details are basically composed of a static factor and a dynamic factor: + +$$ +\mathbf {D} = \mathbf {D} _ {\mathrm {s t a}} + \mathbf {D} _ {\mathrm {d y n}}, \tag {4} +$$ + +where $\mathbf{D}_{\mathrm{sta}}$ and $\mathbf{D}_{\mathrm{dyn}}$ indicate details from static and dynamic factors, respectively. + +Concretely, the static factor is the inherent property of the identity (i.e., the given 2D face), and originates from the + +appearance and age attributes. As for the dynamic factor, it is typically driven by the expression and influenced by person-specific properties. + +Static Detail Generation. To simplify the problem, we are inspired by 3DMMs, which parameterize the statistical models to simplify the 2D-to-3D problem. We build a 300-dim displacement basis $\mathbf{B}_{\mathrm{sta}}$ from the captured 332 scans [57] by PCA [1]. The scans contain diverse age groups in a neutral expression. Then we regress the coefficient $\varphi$ to synthesize the static detail $\mathbf{D}_{\mathrm{sta}}$ from the image: + +$$ +\mathbf {D} _ {\mathrm {s t a}} = \bar {\mathbf {D}} _ {\mathrm {s t a}} + \varphi \mathbf {B} _ {\mathrm {s t a}}, \tag {5} +$$ + +where $\bar{\mathbf{D}}_{\mathrm{sta}}$ and $\mathbf{B}_{\mathrm{sta}}$ are the mean displacement map and displacement basis for static details, respectively. + +Dynamic Detail Generation. Due to the high diversity and complexity of expression representation, directly generating dynamic details from expression is quite difficult. Therefore we simplify the expression representation by using an interpolation between two displacement maps: compressed and stretched [57]. For example, the compressed expression may indicate a state of frowning to the extreme, while the stretched expression may indicate a state of complete relaxation between the eyebrows. Other states of this area can be interpolated by these two polarized states. + +Consequently, we generate the dynamic details through compressed and stretched displacement maps. Again, we build 26-dim compressed $\mathbf{B}_{\mathrm{com}}$ and stretched $\mathbf{B}_{\mathrm{str}}$ displacement bases by PCA [1] to simplify the generation of displacement maps. To generate the dynamic coefficients $\phi = \{\phi_{\mathrm{com}},\phi_{\mathrm{str}}\}$ , we apply the expression coefficient $\xi$ into the static coefficient $\varphi$ through AdaIN [32], followed by the MLP transformation $\Phi$ to obtain $\phi$ : + +$$ +\phi = \Phi \left(\sigma (\tilde {\xi}) \left(\frac {\varphi - \mu (\varphi)}{\sigma (\varphi)} + \mu (\tilde {\xi})\right)\right), \tag {6} +$$ + +where $\tilde{\xi}$ is the affine vector from $\xi$ via MLP transformation. $\mu$ and $\sigma$ indicate the mean and standard deviation. $\phi_{\mathrm{com}}$ and $\phi_{\mathrm{str}}$ are coefficients for compressed and stretched displacement maps respectively. + +Similar to Eq. 5, the compressed and stretched displacement maps are formulated as: + +$$ +\mathbf {D} _ {\text {c o m}} = \bar {\mathbf {D}} _ {\text {c o m}} + \phi_ {\text {c o m}} \mathbf {B} _ {\text {c o m}}, \tag {7} +$$ + +$$ +\mathbf {D} _ {\mathrm {s t r}} = \bar {\mathbf {D}} _ {\mathrm {s t r}} + \phi_ {\mathrm {s t r}} \mathbf {B} _ {\mathrm {s t r}} +$$ + +where $\bar{\mathbf{D}}_{\mathrm{com}}$ and $\mathbf{B}_{\mathrm{com}}$ are the mean displacement map and 26-dim displacement basis for compressed detail, and $\bar{\mathbf{D}}_{\mathrm{str}}$ and $\mathbf{B}_{\mathrm{str}}$ are the mean displacement map and 26-dim displacement basis for stretched detail, respectively. + +Considering the coarse shape $\mathbf{S}$ can be obtained by deforming the neutral shape $\mathbf{S}_{\mathrm{neu}}$ with the expression component $\xi \mathbf{B}_{\mathrm{exp}}$ , such expression-driven deformation over face + +![](images/683427d6bcce4e6e8970f926cbb722850d450993e6ec2ec4517ec88f998b8055.jpg) +(a) + +![](images/0397e8454b8e2dab65c85962c1dfc55674ac8ca3fb58efda01a9c8d780fc8ebb.jpg) +(b) + +![](images/7be66b7330bf156f511f2a67d024a6b814e669912e15518fb633195954a842f7.jpg) + +![](images/b95feed28d83c8c8d7da6e1b36f3170cb768f5e131c994ab9cfa012f26dfaaa2.jpg) +(c) + +![](images/80a23f5a1132fc3dfa2a958e1e5e61b3bb51fc9d3cd21435e8a642b7de033872.jpg) +(d) + +![](images/7c450e66f055ea49c699839bb5e03a8a6fba7822077845a6c6faf4a1af722e93.jpg) +Figure 3. Illustration of displacement map composition in SD-DeTail. Given (a). an image (top) to reconstruct its coarse shape (bottom), we formulate the detail as (b). a static factor and (c). a dynamic factor interpolated by polarized states w.r.t. compressed (top) and stretched (bottom). (d). the output displacement map is linearly combined by (b) and (c) to present vivid details. + +![](images/8ba885db0ffc073f09833f3b7d55c8d78f941f2b654db755ba34d7662fc18deb.jpg) + +![](images/a6224bc791411724dc63d3bc4a0333a605da9ea9ece01529d419a9242c3618f1.jpg) + +![](images/337bc07fc3e509321f5483ea1463f1ac3f6891d9eb7aa025e00f383e82c14a17.jpg) + +shape yields the "tension" over each vertex [57], which influences facial details from expression. Since $\mathbf{S}_{\mathrm{neu}}$ and $\mathbf{S}$ possesses the same topology, for each vertex $\mathbf{v}_i \in \mathbf{S}$ with $K$ -edges $E_i = \{e_1, \dots, e_K\}$ connected with $\mathbf{v}_i$ , $E_i' = \{e_1', \dots, e_K'\}$ are the corresponding edges in $\mathbf{S}_{\mathrm{neu}}$ that are connected to $\mathbf{v}_i'$ . Then the tension at $\mathbf{v}_i$ is: + +$$ +t _ {\mathbf {v} _ {i}} = 1 - \frac {1}{K} \sum_ {k = 1} ^ {K} \frac {\left\| e _ {k} \right\|}{\left\| e _ {k} ^ {\prime} \right\|}, \tag {8} +$$ + +where $\|\cdot\|$ represents the edge length. Positive values of $t_{\mathbf{v}_i}$ indicate compression, negative values indicate stretch, and 0-value indicates no change, respectively. + +The vertex tension $t_{\mathbf{v}_i}$ in $\mathbf{S}$ composes the tension map $\mathbf{M}_{\mathrm{uv}}$ in UV coordinates. Then, the displacement map of the dynamic detail is the linear interpolation of $\mathbf{D}_{\mathrm{com}}$ and $\mathbf{D}_{\mathrm{str}}$ : + +$$ +\mathbf {D} _ {\mathrm {d y n}} = \mathbf {M} _ {\mathrm {u v}} ^ {+} \odot \mathbf {D} _ {\mathrm {c o m}} + \mathbf {M} _ {\mathrm {u v}} ^ {-} \odot \mathbf {D} _ {\mathrm {s t r}}, \tag {9} +$$ + +where $\mathbf{M}_{\mathrm{uv}}^{+}$ and $\mathbf{M}_{\mathrm{uv}}^{-}$ indicate the positive and negative value of $\mathbf{M}_{\mathrm{uv}}$ , respectively. Fig. 3 shows the effectiveness of SD-DeTail. The dynamic factor interpolated by two polarized states introduces expression-related details and further decorates the static detail, yielding the final vivid output. + +# 3.4. Overall Loss Functions + +We propose several loss functions to train HiFace end-to-end. Specifically, we use static and dynamic detail losses to supervise the synthesized displacement maps from $\varphi$ and $\phi$ . In addition, we leverage the coarse shape loss to supervise the reconstructed shape from $\beta$ and $\xi$ . Finally, we follow previous methods [21, 24, 73] to leverage the differentiable renderer [28] to map the generated 3D shape into 2D images, by combining $\alpha$ , $\beta$ , $\xi$ , $\gamma$ , $\mathbf{p}$ , $\varphi$ , $\phi$ . Then, we perform self-supervised losses to train in both synthetic and real-world images. See more details in the supplementary. + +Static and Dynamic Detail Losses. To explicitly train the details of each component, we leverage the ground-truth annotations from the synthetic dataset [57, 72] as supervision to assist the training process of our model. Specifically, we calculate the detail losses by estimating the $l_{2}$ distance between the reconstructed displacement maps and ground-truth w.r.t. static, compressed, and stretched components, and summarize them as $\mathcal{L}_{\mathrm{detail}}$ : + +$$ +\mathcal {L} _ {\mathrm {s t a}} = \left\| \mathbf {M} _ {\text {d e t a i l}} \odot \left(\mathbf {D} _ {\mathrm {s t a}} - \hat {\mathbf {D}} _ {\mathrm {s t a}}\right) \right\| _ {2} +$$ + +$$ +\mathcal {L} _ {\mathrm {c o m}} = \left\| \mathbf {M} _ {\text {d e t a i l}} \odot \left(\mathbf {D} _ {\mathrm {c o m}} - \hat {\mathbf {D}} _ {\mathrm {c o m}}\right) \right\| _ {2}, \tag {10} +$$ + +$$ +\mathcal {L} _ {\mathrm {s t r}} = \left\| \mathbf {M} _ {\text {d e t a i l}} \odot \left(\mathbf {D} _ {\mathrm {s t r}} - \hat {\mathbf {D}} _ {\mathrm {s t r}}\right) \right\| _ {2} +$$ + +$$ +\mathcal {L} _ {\text {d e t a i l}} = \mathcal {L} _ {\text {s t a}} + \mathcal {L} _ {\text {c o m}} + \mathcal {L} _ {\text {s t r}} +$$ + +where $\mathbf{M}_{\mathrm{detail}}$ is the facial mask in the UV coordinates, and $\hat{\mathbf{D}}_{\mathrm{sta}} / \hat{\mathbf{D}}_{\mathrm{com}} / \hat{\mathbf{D}}_{\mathrm{str}}$ and $\mathbf{D}_{\mathrm{sta}} / \mathbf{D}_{\mathrm{com}} / \mathbf{D}_{\mathrm{str}}$ are the reconstructed and ground-truth displacement maps, respectively. + +Coarse Shape Losses. Since the details should be based on realistic coarse shapes, we train the coarse shape to help the learning of details by leveraging the ground-truth vertex as supervision: + +$$ +\mathcal {L} _ {\text {v e r}} = \left\| \mathbf {M} _ {\text {v e r}} \odot (\mathbf {S} - \hat {\mathbf {S}}) \right\| _ {2}, \tag {11} +$$ + +where $\mathbf{M}_{\mathrm{ver}}$ is frontal face area of the coarse shape. $\hat{\mathbf{S}}$ and $\mathbf{S}$ are the reconstructed and ground-truth face by Eq. 1. + +In addition, we make constraints on shape coefficients to prevent overfitting. We enforce the predicted coefficients have a similar distribution to the ground-truth coefficients: + +$$ +\mathcal {L} _ {\mathrm {k l}} = \rho (\boldsymbol {\beta}) \left(\log \rho (\boldsymbol {\beta}) - \log \rho (\hat {\boldsymbol {\beta}})\right), \tag {12} +$$ + +where $\rho$ denotes softmax function to map the predicted coefficients $\hat{\beta}$ and ground-truth $\beta$ into probability distribution. + +Finally, the shape loss is $\mathcal{L}_{\mathrm{shp}} = \mathcal{L}_{\mathrm{ver}} + \mathcal{L}_{\mathrm{kl}}$ + +Self-supervised Losses. To encourage the generalization of our models in real-world images [40, 52], we follow previous methods [24, 21, 73] to leverage self-supervised loss $\mathcal{L}_{\mathrm{self}}$ for all training images, including photo loss $\mathcal{L}_{\mathrm{pho}}$ , perceptual loss $\mathcal{L}_{\mathrm{id}}$ , and dense landmark loss $\mathcal{L}_{\mathrm{lmk}}$ : + +$$ +\mathcal {L} _ {\text {s e l f}} = \mathcal {L} _ {\text {p h o}} + \lambda_ {\mathrm {i d}} \mathcal {L} _ {\mathrm {i d}} + \lambda_ {\mathrm {l m k}} \mathcal {L} _ {\mathrm {l m k}}, \tag {13} +$$ + +where $\lambda_{\mathrm{id}}$ and $\lambda_{\mathrm{lmk}}$ are weights to balance the self-supervised losses term. + +In addition, considering the static detail heavily correlates to person-specific age attribute, inspired by [19], we leverage the pre-trained age prediction network [36] to learn high-level representations of static details through knowledge distillation, such that the learned coefficients exhibit expressive results. To achieve this, we use several MLP layers on the static coefficient $\varphi$ , and map it into age classification probabilities $\hat{\mathbf{p}}_{\mathrm{age}}$ . Then we use the pre-trained + +Table 1. Quantitative comparison of 3D face reconstruction methods on REALY benchmark. “-c” and “-d” indicate coarse and detail shape, respectively. @RN/@RM/@RF/@RC/all indicate errors in nose/mouth/forehead/cheek/all regions. We highlight the best method for the two groups respectively. HiFace achieves the best reconstruction performance in the overall error by a large margin. Each component in HiFace contributes to a better reconstruction quality. The reconstructed details of HiFace further boost the quality while previous methods [24, 19] modeling details with only image-level supervision even deteriorate the reconstruction accuracy. + +
GroupMethods / e (mm)frontal-viewside-view
@RN@RM@RF@RCall@RN@RM@RF@RCall
CoarseDeep3D [21]1.719±0.3541.368±0.4392.015±0.4491.528±0.5011.6571.749±0.3431.411±0.3952.074±0.4861.528±0.5171.691
MGCNet [64]1.771±0.3801.417±0.4092.268±0.5031.639±0.6501.7741.827±0.3831.409±0.4182.248±0.5081.665±0.6441.787
3DDFA-v2 [29]1.903±0.5171.597±0.4782.447±0.6471.757±0.6421.9261.883±0.4991.642±0.5012.465±0.6221.781±0.6361.943
DECA-c [24]1.694±0.3552.516±0.8392.394±0.5761.479±0.5352.0101.903±1.0502.472±1.0792.423±0.7201.630±1.1352.107
SADRNet [61]1.791±0.5421.591±0.4882.413±0.5371.856±0.7011.9131.771±0.5211.560±0.4622.490±0.5662.010±0.7151.958
EMOCA-c [19]1.868±0.3872.679±1.1122.426±0.6411.438±0.5012.1031.867±0.5542.636±1.2842.448±0.7081.548±0.5902.125
MICA [81]1.585±0.3253.478±1.2042.374±0.6831.099±0.3242.1341.525±0.3223.567±1.2122.379±0.6751.109±0.3252.145
Ours-c (w/o Syn. Data)†1.227±0.4071.787±0.4391.454±0.3821.762±0.4361.5581.187±0.3791.826±0.4901.470±0.4261.653±0.4501.534
Ours-c1.054±0.3171.461±0.4301.331±0.3471.342±0.3841.2970.992±0.2461.505±0.4541.427±0.4001.439±0.4291.341
DetailDECA-d [24]2.138±0.4612.802±0.8682.457±0.5591.443±0.4982.2102.286±1.1032.684±1.0412.519±0.7181.555±0.8222.261
EMOCA-d [19]2.532±0.5392.929±1.1062.595±0.6311.495±0.4692.3882.455±0.6362.948±1.2922.606±0.6861.599±0.5632.402
HRN [42]1.722±0.3301.357±0.5231.995±0.4761.072±0.3331.5371.642±0.3101.285±0.5281.906±0.4791.038±0.3221.468
Ours-d (w/o Syn. Data)†1.465±0.5571.790±0.4251.528±0.3731.618±0.3621.6001.422±0.5371.849±0.4731.530±0.4141.572±0.3991.594
Ours-d (w/o static)*1.055±0.2901.469±0.4151.336±0.3371.319±0.3741.2951.004±0.2331.491±0.4371.418±0.3921.418±0.4151.332
Ours-d (w/o dynamic)*1.069±0.3181.469±0.4141.358±0.3361.270±0.3441.2920.991±0.2391.496±0.4371.411±0.3931.375±0.4021.318
Ours-d1.036±0.2801.450±0.4131.324±0.3341.291±0.3621.2750.985±0.2371.489±0.4361.399±0.3881.360±0.3951.308
+ +To align the dataset scale, w/o Syn. Data indicates we train the model without using the ground-truth labels from the synthetic dataset. +* To eliminate the bias of coarse shape in estimating the reconstruction error, we fix the coarse shape and train the details with/without static and dynamic factors for comparisons. + +age recognition model $\Gamma_{\mathrm{age}}$ to obtain the probabilities of the given input image $\mathbf{I}$ . The distillation loss $\mathcal{L}_{\mathrm{kd}}$ enforces the probabilities between $\hat{\mathbf{p}}_{\mathrm{age}}$ and $\Gamma_{\mathrm{age}}(\mathbf{I})$ be similar: + +$$ +\mathcal {L} _ {\mathrm {k d}} = \Gamma_ {\text {a g e}} (\mathbf {I}) \left(\log \Gamma_ {\text {a g e}} (\mathbf {I}) - \log \hat {\mathbf {p}} _ {\text {a g e}}\right). \tag {14} +$$ + +Regularization. $\mathcal{L}_{\mathrm{reg}}$ regularizes coefficients of each submodule, by minimizing the $l_{2}$ loss of $\alpha, \beta, \xi, \varphi, \phi$ . + +Overall Loss Function. We train the coarse shape and fine details simultaneously, such that each component can collaborate to reconstruct high-fidelity 3D faces with realistic details. Formally, we minimize the total loss function: + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {\text {d e t a i l}} \mathcal {L} _ {\text {d e t a i l}} + \lambda_ {\text {s h p}} \mathcal {L} _ {\text {s h p}} \tag {15} \\ + \lambda_ {\text {s e l f}} \mathcal {L} _ {\text {s e l f}} + \lambda_ {\text {k d}} \mathcal {L} _ {\text {k d}} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}}, \\ \end{array} +$$ + +where $\lambda$ is the weight for each component. + +# 4. Experiments + +# 4.1. Implementation Details + +Dataset. We use a hybrid dataset made up from both synthetic [72, 57] and real data [40, 52]. We use the synthetic data pipeline [72, 57] to generate a diverse dataset of $200k$ faces with ground-truth vertex, landmark, albedo, and displacement map annotations. The real-world datasets contain $400k$ images in total from diverse age, gender, and ethnicity groups. For the real-world dataset, we use the pretrained dense landmark detector [72] to detect 669 landmarks for training. We use face parsing [79] to generate and select region-of-interest as facial masks, providing robustness to common occlusions by hair or other accessories. We follow [19, 73, 21] to split the dataset into training + +and validation sets. The test images are from CelebA [40], FFHQ [37], LS3D-W [8], and AFLW2000 [78]. + +Implementation Details. We implement HiFace in PyTorch [54] and leverage the PyTorch3D differentiable rasterizer [35] for rendering. We train our model for 35 epochs on $8 \times$ NVIDIA Tesla V100 GPUs with a mini-batch of 320. We use the pre-trained ResNet-50 on ImageNet [20] as initialization, and use Adam [39] as optimizer with an initial learning rate of $1e - 4$ . The input image is cropped and aligned by [14], and resized into $224 \times 224$ . We empirically set $\lambda_{\mathrm{detail}} = 10$ , $\lambda_{\mathrm{shp}} = 1$ , $\lambda_{\mathrm{self}} = 1$ , $\lambda_{\mathrm{id}} = 0.1$ , $\lambda_{\mathrm{lmk}} = 0.5$ , $\lambda_{\mathrm{kd}} = 1$ , $\lambda_{\mathrm{reg}} = 1e - 3$ throughout the experiments. + +# 4.2. Comparisons to State-of-the-art + +Quantitative Comparison. We perform the quantitative evaluation on the REALY benchmark [12], which contains 100 frontal-view and 400 side-view images from 100 textured-scans [17]. The REALY benchmark presents a region-aware evaluation pipeline to separately evaluate the metric error (in mm) of the nose, mouth, forehead, and cheek regions. Such an evaluation pipeline is demonstrated to better estimate the actual similarity of the 3D faces and align with human perception. We compare HiFace to previous state-of-the-art methods and report the region-wise and average normalized mean square error (NMSE) in Tab. 1. + +As Tab. 1 illustrates, HiFace outperforms prior arts in the overall error by a large margin. HiFace balances the reconstruction quality on each region, compared to those optimum region methods that may fail in specific regions (e.g., MICA [81] fails in mouth region while HRN [42] fails in forehead region). Note that HiFace faithfully recovers the facial details, thus making the reconstruction error smaller + +![](images/bc7b84348e7ae97825277b3dd80793bb64fa799ec276a3f96de851633c8c80a5.jpg) +Input Deep3D 3DDFA-v2 SynergyNet MICA Dense Ours + +![](images/8e009b9dce3c38cf8a273e6fcceabbc11968dbd58488edb3fea263f32a64a4bb.jpg) +Figure 4. Comparison on coarse shape reconstruction. From left to right: Input image, Deep3D [21], 3DDFA-v2 [29], Synerg-yNet [74], MICA [81], Dense [73], and HiFace (Ours). +Figure 6. Illustration on the flexibility of SD-DeTail. Given the identity and expression coefficients $(\beta, \xi)$ from the optimization-based method [73], SD-DeTail can generate realistic details based on the coarse shape and further improve the visual quality. + +than using the coarse shape alone. As a comparison, although DECA [24] and EMOCA [19] can reconstruct details of given images, they turn out to be noisy, leading to the deterioration of reconstruction quality. + +In addition, Tab. 1 also demonstrates the necessity of each component in contributing to a better quality. It can be observed that the synthetic data with ground-truth labels not only improve the coarse shape reconstruction quality but is also crucial for detailed reconstruction. With the synthetic data, the proposed SD-DeTail further boosts the overall reconstruction quality. Both the static and dynamic factors are essential to capture fine-grained details, and the final SD-DeTail achieves the most accurate details in expression-related regions such as the mouth and forehead, which contributes to the overall gains. + +Qualitative Comparison. Given a single face image, Hi-Face reconstructs a high-fidelity 3D shape with details. We present comparisons with previous methods on 1). coarse shape reconstruction [21, 29, 74, 81, 73] in Fig. 4 and 2). detail reconstruction [77, 15, 24, 19, 70] in Fig. 5. See more examples and comparisons in the supplementary. + +For coarse shape in Fig. 4, our HiFace faithfully recovers the coarse shape of the given identity and outperforms the previous learning-based methods, and is on par + +![](images/96bcd63fa56e720e31a542c754ffe9e6af72fa417bbd8fc37a20b84b1fe48101.jpg) + +![](images/d198d2b1537a3210ca597276579b3798a63cec5f94e02a143f676daf8c9a3307.jpg) +Figure 5. Comparison on detail shape reconstruction. From left to right: Input image, FaceScape [77], Unsup [15], DECA [24], EMOCA [19], FaceVerse [70], and HiFace (Ours). +Figure 7. Comparison on face animation. Given a source image (yellow box), we use the driving images (green box) to drive its expressions. DECA [24] (2nd-row) and EMOCA [19] (3rd-row) can animate the expression-driven details but lack realistic. As a comparison, HiFace is flexible to animate details from static (4th-row), dynamic (5th-row), or both (6th-row) factors, and presents vivid animation quality with realistic shapes. + +with Dense [73], which is the state-of-the-art optimization-based method. For detailed reconstruction in Fig. 5, previous methods [24, 19] fail to reconstruct satisfactory details. Several methods [15, 77, 70] are sensitive to occlusions and large poses. As a comparison, HiFace achieves the most realistic reconstruction quality, and faithfully recovers facial details of a given image, which significantly outperforms previous methods by a large margin. + +In addition, given an image and the fitted coefficients $\beta$ , + +![](images/cbe86c2ac57a9cacabced6224a497ecd878935888a91d0304de9aa09358173c9.jpg) +Figure 8. Ablation studies on loss functions and training data. The coarse shape losses $\mathcal{L}_{\mathrm{shp}} / \mathcal{L}_{\mathrm{kl}}$ (left), detail losses $\mathcal{L}_{\mathrm{detail}} / \mathcal{L}_{\mathrm{kd}}$ (right), and hybrid datasets coherently contribute to the reconstruction quality of coarse shapes and details. + +![](images/bdf077b5d3d7866c628b7cc9a327084fda34181b08ced33d24950c22d3d41570.jpg) + +$\xi$ from the optimization-based methods such as Dense [73], SD-DeTail synthesizes the details and further strengthens the quality compared to the coarse shape (see Fig. 6). It shows SD-DeTail is flexible and can be easily plugged-and-play into other methods. See more in the supplementary. + +Application of HiFace. The HiFace explicitly decouples the static and dynamic details through the proposed SD-DeTail. Therefore, we can animate the facial attributes by simply assigning the expression coefficient $\xi$ and/or static coefficient $\varphi$ of the driving images to the source images. + +In Fig. 7, we demonstrate the animation quality of HiFace outperforms the previous state-of-the-art detail animation methods [24, 19]. It shows that while DECA [24] and EMOCA [19] can animate the expression-driven details but lack realistic, the proposed HiFace is flexible to manipulate the static, dynamic, or both details. When animating the static detail, the person-specific details can be well transferred into the source shape. When animating the dynamic detail, only expression-dependent details are presented. Finally, we can also animate the static and dynamic details simultaneously and achieve satisfactory results. + +# 5. Ablation Studies + +Ablation Studies on Loss Functions and Datasets. We present ablation studies on the proposed loss functions and training strategy with hybrid datasets. We train HiFace with synthetic data alone and compare it to the one trained with hybrid datasets. For coarse shape reconstruction, we investigate the contribution of $\mathcal{L}_{\mathrm{shp}}$ and its sub-term $\mathcal{L}_{\mathrm{kl}}$ . For detail reconstruction, we compare HiFace without $\mathcal{L}_{\mathrm{detail}}$ and $\mathcal{L}_{\mathrm{kd}}$ , respectively. The results are presented in Fig. 8. + +Fig. 8 demonstrates that the proposed loss functions and training strategy from hybrid datasets contribute to satisfactory coarse shape and details. First, the model trained with synthetic data alone cannot generalize well to real-world images, which indicates the necessity to train with real-world data. Second, $\mathcal{L}_{\mathrm{shp}}$ improves the coarse shape reconstruction quality. $\mathcal{L}_{\mathrm{shp}}$ is effective in tackling challenging poses and improving alignment. $\mathcal{L}_{\mathrm{kl}}$ can relieve the overfitting risk on the synthetic data and improve the gener + +![](images/02fee0cf9532e8d092cb67544266cebe39309ee88183fadd57d475e656449220.jpg) +Figure 9. Ablation studies on SD-DeTail. Results show that directly synthesizing the static or dynamic details is rather challenging, leading to unreasonable coarse shapes and details (SD-1, SD-2, and SD-3). As a comparison, we leverage the statistical bases with SD-DeTail and regard the detail generation problem as a coefficients regression and interpolation problem, leading to more realistic details. + +alization to real-world images. Third, without $\mathcal{L}_{\mathrm{detail}}$ or $\mathcal{L}_{\mathrm{kd}}$ , the reconstructed details exhibit random noise and cannot faithfully reflect person-specific details. Such noise misses the correspondence to the person-specific identity. + +Ablation Studies on SD-DeTail. To verify the effectiveness of building bases for static and dynamic details, we present detailed ablation studies on SD-DeTail, by replacing the bases (i.e., $\mathbf{B}_{\mathrm{sta}}$ and $\mathbf{B}_{\mathrm{com}} / \mathbf{B}_{\mathrm{str}}$ ) reconstruction with a U-Net decoder [60] (same as DECA [24]). Therefore, the model learns to directly synthesize displacement maps instead of predicting corresponding coefficients like ours. In Fig. 9, we make comparisons on: 1). directly synthesizing $\mathbf{D}_{\mathrm{dyn}}$ (SD-1), 2). directly synthesizing $\mathbf{D}_{\mathrm{com}} / \mathbf{D}_{\mathrm{str}}$ and interpolating via Eq. 9 (SD-2), 3). directly synthesizing $\mathbf{D}_{\mathrm{sta}}$ (SD-3), 4). SD-DeTail (Ours). It can be seen that, due to the high diversity and complexity of expression representation, it is hard to directly learn realistic details even with ground-truth labels of $\mathbf{D}_{\mathrm{dyn}}$ from synthetic data (see SD-1, SD-2 and SD-3 in Fig. 9). More specifically, for the static details, directly synthesizing displacement maps bring much noise (SD-3). For example, the hollow eyebrow is demonstrated in the second row. For the dynamic details, directly synthesizing displacement maps even leads to unnatural results (SD-1 and SD-2). For example, the reconstructed 3D faces are distorted especially in the second row. We also notice + +that, directly synthesizing $\mathbf{D}_{\mathrm{dyn}}$ (SD-1) achieves inferior results than directly synthesizing $\mathbf{D}_{\mathrm{com}} / \mathbf{D}_{\mathrm{str}}$ and interpolating via Eq. 9 (SD-2). This demonstrates that it is beneficial to simplify the expression representation by using interpolation between two displacement maps (i.e., compressed and stretched). In conclusion, these observations further verify our insight on relaxing the challenging detail generation problem into a feasible coefficients regression problem. + +# 6. Conclusion + +We propose HiFace to reconstruct high-fidelity 3D faces with realistic and animatable details from a single image. Our motivation and insights stand on the spirit of 3DMMs to simplify the challenging detail generation into more accessible regression and interpolation tasks. To achieve this, we elaborately design SD-DeTail to decouple the static and dynamic factors explicitly, and interpolate the dynamic details through vertex tension. We succeed in learning the coarse shape and details jointly by proposing several novel loss functions to train on synthetic and real-world data. Extensive experiments demonstrate that HiFace achieves state-of-the-art face reconstruction both quantitatively and qualitatively in the coarse shape and detail shape, and the details are well decoupled and naturally animatable. + +# Acknowledgement + +This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SZSTC Grant (JCYJ20190809172201639, WDZC2020-0820200655001), Shenzhen Key Laboratory (ZDSYS2021-0623092001004), and Beijing Key Lab of Networked Multimedia. + +# References + +[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 2, 4 +[2] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2 +[3] Linchao Bao, Xiangkai Lin, Yajing Chen, Haoxian Zhang, Sheng Wang, Xuefei Zhe, Di Kang, Haozhi Huang, Xinwei Jiang, Jue Wang, et al. High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2 +[4] Volker Blanz, Sami Romdhani, and Thomas Vetter. Face identification across different poses and illuminations with a 3d morphable model. In Proceedings of fifth IEEE international conference on automatic face gesture recognition, pages 202-207. IEEE, 2002. 1 +[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual + +conference on Computer graphics and interactive techniques, pages 187-194, 1999. 1, 2 +[6] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2):233-254, 2018. 1 +[7] Sofien Bouaziz, Yangang Wang, and Mark Pauly. Online modeling for realtime facial animation. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013. 1 +[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017. 6 +[9] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Realtime high-fidelity facial performance capture. ACM Transactions on Graphics (ToG), 34(4):1-9, 2015. 2 +[10] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 1, 2 +[11] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5187-5196, 2018. 1 +[12] Zenghao Chai, Haoxian Zhang, Jing Ren, Di Kang, Zhengzhuo Xu, Xuefei Zhe, Chun Yuan, and Linchao Bao. Really: Rethinking the evaluation of 3d face reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6 +[13] Anpei Chen, Zhang Chen, Guli Zhang, Kenny Mitchell, and Jingyi Yu. Photo-realistic facial details synthesis from single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9429-9439, 2019. 2, 3 +[14] Dong Chen, Gang Hua, Fang Wen, and Jian Sun. Supervised transformer network for efficient face detection. In European Conference on Computer Vision, pages 122-138. Springer, 2016. 6 +[15] Yajing Chen, Fanzi Wu, Zeyu Wang, Yibing Song, Yonggen Ling, and Linchao Bao. Self-supervised learning of detailed 3d face reconstruction. IEEE Transactions on Image Processing, 29:8696-8705, 2020. 2, 3, 7 +[16] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1 +[17] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 2019. 6 +[18] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE international conference on computer vision, pages 3085-3093, 2017. 2 +[19] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 2, 3, 5, 6, 7, 8 + +[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[21] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 1, 2, 3, 4, 5, 6, 7 +[22] Abdallah Dib, Cédric Thébault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 3 +[23] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2 +[24] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2, 3, 5, 6, 7, 8 +[25] Zhongpai Gao, Juyong Zhang, Yudong Guo, Chao Ma, Guangtao Zhai, and Xiaokang Yang. Semi-supervised 3d face representation learning from unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 348-349, 2020. 3 +[26] Pablo Garrido, Michael Zollhöfer, Dan Casas, Levi Valgaerts, Kiran Varanasi, Patrick Pérez, and Christian Theobalt. Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG), 35(3):1-15, 2016. 3 +[27] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 2 +[28] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3, 5 +[29] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 6, 7 +[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[31] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG), 36(6):1-14, 2017. 1 +[32] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed- + +ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4 +[33] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3 +[34] Luo Jiang, Juyong Zhang, Bailin Deng, Hao Li, and Ligang Liu. 3d face reconstruction with geometry details from a single image. IEEE Transactions on Image Processing, 27(10):4756-4770, 2018. 3 +[35] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. ACM New York, NY, USA, 2020. 6 +[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5 +[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6 +[38] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 1 +[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6 +[40] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6 +[41] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware mesh decoder for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6100–6109, 2020. 1, 3 +[42] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 3, 6 +[43] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 165-172, 2000. 3 +[44] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 2 +[45] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 1, 2 + +[46] Yue Li, Liqian Ma, Haoqiang Fan, and Kenny Mitchell. Feature-preserving detailed 3d face reconstruction from a single image. In Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production, pages 1-9, 2018. 3 +[47] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3d face reconstruction from in-the-wild images using graph convolutional networks. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 5891-5900, 2020. 3 +[48] Jingwang Ling, Zhibo Wang, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Structure-aware editable morphable model for 3d facial detail animation and manipulation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[49] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3 +[50] Huiwen Luo, Koki Nagano, Han-Wei Kung, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11662–11672, 2021. 2 +[51] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriia Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20942–20952, 2022. 3 +[52] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 3, 5, 6 +[53] Araceli Morales, Gemma Piella, and Federico M Sukno. Survey on 3d face reconstruction from uncalibrated images. Computer Science Review, 40:100400, 2021. 2 +[54] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[55] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2 +[56] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4 +[57] Chirag Raman, Charlie Hewitt, Erroll Wood, and Tadas Baltrusaitis. Mesh-tension driven expression-based wrinkles for synthetic faces. In WACV, 2023. 2, 3, 4, 5, 6 +[58] Roger Blanco I Ribera, Eduard Zell, John P Lewis, Junyong Noh, and Mario Botsch. Facial retargeting with automatic range of motion alignment. ACM Transactions on graphics (TOG), 36(4):1-12, 2017. 1 +[59] Sami Romdhani, Volker Blanz, and Thomas Vetter. Face identification by fitting a 3d morphable model using linear + +shape and texture error functions. In European Conference on Computer Vision, pages 3-19. Springer, 2002. 1 +[60] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 8 +[61] Zeyu Ruan, Changqing Zou, Longhai Wu, Gangshan Wu, and Limin Wang. Sadrnet: Self-aligned dual face regression networks for robust 3d dense face alignment and reconstruction. IEEE Transactions on Image Processing, 30:5793-5806, 2021. 6 +[62] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3 +[63] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1 +[64] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 6 +[65] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 1, 2 +[66] Supasorn Suwajanakorn, Ira Kemelmacher-Shlizerman, and Steven M Seitz. Total moving face reconstruction. In European conference on computer vision, pages 796-812. Springer, 2014. 3 +[67] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 1, 2, 3 +[68] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3 +[69] Daniel Vlasic, Matthew Brand, Hanspeter Pfister, and Jovan Popovic. Face transfer with multilinear models. In ACM SIGGRAPH 2006 Courses, pages 24–es. ACM New York, NY, USA, 2006. 2 +[70] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 7 +[71] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13289-13298, 2021. 3 + +[72] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2, 3, 5, 6 +[73] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 2, 3, 5, 6, 7, 8 +[74] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 7 +[75] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 1 +[76] Shugo Yamaguchi, Shunsuke Saito, Koki Nagano, Yajie Zhao, Weikai Chen, Kyle Olszewski, Shigeo Morishima, and Hao Li. High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2 +[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu + +Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 2, 3, 7 +[78] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Towards large-posed face frontalization in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3990–3999, 2017. 6 +[79] Qi Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Computer Vision and Pattern Recognition, 2022. 6 +[80] Wenbin Zhu, HsiangTao Wu, Zeyu Chen, Noranart Vesdapunt, and Baoyuan Wang. Reda: reinforced differentiable attribute for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4958-4967, 2020. 3 +[81] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision (ECCV). Springer International Publishing, Oct. 2022. 1, 2, 3, 6, 7 +[82] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum. Wiley Online Library, 2018. 2 \ No newline at end of file diff --git a/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/images.zip b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3a2bcb6c974f4c670243c4e04207af2294008e28 --- /dev/null +++ b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88f3115e1964d7f59d1cda6b0df583de525792487450cec4708ccf2179f48366 +size 774285 diff --git a/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/layout.json b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..86cebc48eb59e946115fbaf73732855d4a506243 --- /dev/null +++ b/2023/HiFace_ High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details/layout.json @@ -0,0 +1,12801 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 451, + 140 + ], + "type": "text", + "content": "HiFace: High-Fidelity 3D Face Reconstruction by Learning Static and Dynamic Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "spans": [ + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": "Zenghao Chai" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Tianke Zhang" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Tianyu He" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Xu Tan" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Tadas Baltrušaitis" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " HsiangTao Wu" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Runnan Li" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Sheng Zhao" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Chun Yuan" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": " Jiang Bian" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": "Microsoft Research Asia \n" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": "Microsoft Mixed Reality & AI Lab " + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 82, + 160, + 514, + 246 + ], + "type": "text", + "content": "Microsoft Cloud + AI \nzenghaochai@gmail.com ztk21@mails.tsinghua.edu.cn yuanc@sz.tsinghua.edu.cn \n{tianyuhe,xuta,tabaltru,musclewu,runnan.li,sheng.zhao,jiang.bian}@microsoft.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 298, + 290, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 290, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 290, + 562 + ], + "type": "text", + "content": "3D Morphable Models (3DMMs) demonstrate great potential for reconstructing faithful and animatable 3D facial surfaces from a single image. The facial surface is influenced by the coarse shape, as well as the static detail (e.g., person-specific appearance) and dynamic detail (e.g., expression-driven wrinkles). Previous work struggles to decouple the static and dynamic details through image-level supervision, leading to reconstructions that are not realistic. In this paper, we aim at high-fidelity 3D face reconstruction and propose HiFace to explicitly model the static and dynamic details. Specifically, the static detail is modeled as the linear combination of a displacement basis, while the dynamic detail is modeled as the linear interpolation of two displacement maps with polarized expressions. We exploit several loss functions to jointly learn the coarse shape and fine details with both synthetic and real-world datasets, which enable HiFace to reconstruct high-fidelity 3D shapes with animatable details. Extensive quantitative and qualitative experiments demonstrate that HiFace presents state-of-the-art reconstruction quality and faithfully recovers both the static and dynamic details. Our project page: https://project-hiface.github.io." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 603, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 687 + ], + "type": "text", + "content": "The reconstruction of a 3D face from a single image has drawn much attention recently [67, 21, 41, 81]. It has tremendous potential applications like face recognition [11, 63, 4, 59], face animation [16, 75], virtual reality [7, 58, 31], etc. For example, the reconstructed 3D face representation can be driven by an audio [16], or a video from another person [38]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 272, + 426, + 389 + ], + "blocks": [ + { + "bbox": [ + 307, + 272, + 426, + 389 + ], + "lines": [ + { + "bbox": [ + 307, + 272, + 426, + 389 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 426, + 389 + ], + "type": "image", + "image_path": "1c00cf10e1215744da80064e3797585206678ae0d1ae242aceac0c1dbe52a2ea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 272, + 544, + 389 + ], + "blocks": [ + { + "bbox": [ + 426, + 272, + 544, + 389 + ], + "lines": [ + { + "bbox": [ + 426, + 272, + 544, + 389 + ], + "spans": [ + { + "bbox": [ + 426, + 272, + 544, + 389 + ], + "type": "image", + "image_path": "1195b9a15fd2f5a623241916e9f3741486d6c16a08677b268519183095717106.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 390, + 425, + 507 + ], + "blocks": [ + { + "bbox": [ + 307, + 390, + 425, + 507 + ], + "lines": [ + { + "bbox": [ + 307, + 390, + 425, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 390, + 425, + 507 + ], + "type": "image", + "image_path": "63b1b1b56817ba4333a72fb19fb4d33d1d40598e6a27343b136c4fa6a3f9f809.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 390, + 544, + 507 + ], + "blocks": [ + { + "bbox": [ + 425, + 390, + 544, + 507 + ], + "lines": [ + { + "bbox": [ + 425, + 390, + 544, + 507 + ], + "spans": [ + { + "bbox": [ + 425, + 390, + 544, + 507 + ], + "type": "image", + "image_path": "2a059dfce8796e1fc9e0a624d78c52c64d3bb7597dcd9d077e6c799d2749577b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 506, + 406, + 546 + ], + "blocks": [ + { + "bbox": [ + 326, + 506, + 406, + 546 + ], + "lines": [ + { + "bbox": [ + 326, + 506, + 406, + 546 + ], + "spans": [ + { + "bbox": [ + 326, + 506, + 406, + 546 + ], + "type": "image", + "image_path": "4563679efa1a8ab5206d5c5548112ddb398c55259c32d140f515975398ef7922.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 556, + 547, + 634 + ], + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 634 + ], + "type": "text", + "content": "Figure 1. We propose HiFace to reconstruct high-fidelity 3D face with realistic and animatable details. Reconstruction: given a single image (1st-row), HiFace faithfully reconstructs a coarse shape (2nd-row) with vivid details (3rd-row). Animation: given a source face (yellow box), HiFace can animate the static (4th-row), dynamic (5th-row), or both (6th-row) details of the driving images (green box). Images are taken from FFHQ [37] and CelebA [40]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 446, + 506, + 525, + 546 + ], + "blocks": [ + { + "bbox": [ + 446, + 506, + 525, + 546 + ], + "lines": [ + { + "bbox": [ + 446, + 506, + 525, + 546 + ], + "spans": [ + { + "bbox": [ + 446, + 506, + 525, + 546 + ], + "type": "image", + "image_path": "40004fb7ad32dc8431607682c055339918e9b028cdb0635b8ba37585066ce3fa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "To build a flexible and animatable facial representation, a popular way is to leverage the success of 3D Morphable Models (3DMMs) [5, 6, 10, 45, 65], which decouple the influence of shape, expression, albedo, and others by modeling them in separate coefficients. Typically" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 237, + 713 + ], + "type": "text", + "content": "*Work done when the author was an intern at MSRA. †Corresponding author: Xu Tan (xuta@microsoft.com)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9087" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "in literature, one can achieve coarse shape reconstruction in coefficients-fitting optimization [27, 73, 3, 77, 2], or an analysis-by-synthesis pipeline [67, 21, 81, 50]. As 3DMMs typically capture only the coarse facial geometry and are not capable of representing fine details (e.g., wrinkles), recent advances model such details with a displacement map [15, 77, 13, 9, 76]. However, previous work fails to model the distinction between static and dynamic factors of fine detail, leading to errors in reconstructions. For example, given that one may drive the expression of a young man from an old man, trivially transferring all wrinkles from the old man to the young man could make the young man look unnatural. In this sense, Feng et al. [24] implicitly leverages the person-specific identity and expression as conditions to generate the details. Although effective, they optimize the model in an analysis-by-synthesis pipeline with only the image-level supervision, leading to insufficient decoupling of static and dynamic details and inconsistent animation results (see Fig. 7)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 289, + 590 + ], + "type": "text", + "content": "Therefore, we propose HiFace to explicitly model the static and dynamic details for high-fidelity 3D face reconstruction, by designing SD-DeTail module to decouple the static and dynamic factors. More specifically, for person-specific static detail, instead of directly predicting the displacement map that may increase the difficulty of detail prediction [24, 19], we follow the spirit of 3DMMs to build a displacement basis from the captured facial scans with age diversity [57, 72]. In this way, the model is trained to predict the coefficients of the displacement basis, and make the detail prediction easier. For dynamic detail, since it is highly expression-dependent, directly modeling it with one displacement basis is quite difficult. Therefore, based on the fact that the expression can be interpolated by a compressed and a stretched expressions [57], we build two displacement bases for the compressed and stretched expressions from the captured scans respectively, and learn to regress the displacement coefficients with the ground-truth labels. Therefore, we can obtain the dynamic detail by linearly interpolating the compressed and stretched displacement maps, which are derived from the displacement bases and the predicted coefficients. Finally, the predicted static and dynamic details are merged with the coarse shape to formulate the final output." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Since we would like the final output to contain both the coarse shape and high-frequency detail, we propose several novel loss functions to learn coarse shape and details simultaneously from both the synthetic and real-world datasets. For details, we leverage the ground-truth static and dynamic displacement maps of the synthetic dataset [72, 57] as supervision. While for the coarse shape, we leverage the ground-truth vertex of the synthetic dataset as supervision. We also follow the previous methods [24, 21, 73] to leverage self-supervised losses for all training images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "Overall, with the above insights and techniques, HiFace enables the reconstruction of high-fidelity 3D faces from a single image, and decouples static and dynamic details that are naturally animatable (see Fig. 1). We demonstrate that the proposed HiFace reconstructs realistic and faithful 3D faces, reaching state-of-the-art performance both quantitatively and qualitatively. In addition, HiFace is compatible with optimization-based methods [73], and is flexible to transfer vivid expressions and details from one person to another. In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 192, + 545, + 346 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 192, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 545, + 239 + ], + "type": "text", + "content": "- We propose HiFace to model the static and dynamic details explicitly, and demonstrate the benefits of synthetic data in decoupling the static and dynamic factors for detailed 3D face reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 239, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 545, + 275 + ], + "type": "text", + "content": "- We propose novel loss functions in HiFace to learn 3D representations of coarse shape and fine details simultaneously from both the synthetic and real-world images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "text", + "content": "- We achieve state-of-the-art reconstruction quality both quantitatively and qualitatively, with over " + }, + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 306, + 276, + 545, + 311 + ], + "type": "text", + "content": " performance gains in the region-aware benchmark [12]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 312, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 545, + 346 + ], + "type": "text", + "content": "- We show that our SD-DeTail is easy to plug-and-play into optimization-based methods and can transfer expressions and details from one to another for face animation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 358, + 392, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 392, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 392, + 371 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 379, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 451 + ], + "type": "text", + "content": "3D face reconstruction from monocular images has received much attention in the past decades. Among them, 3D Morphable Models (3DMMs) are widely used to build 3D representations. Below we review the works that are related to them, and a full in-depth review can be found in recent surveys [82, 53, 23]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "content": "3D Morphable Model (3DMMs) [23] are statistical models widely used to constrain the distribution of 3D faces. The seminal work [5] presents 200 scans to generate shape and texture bases with Principal Component Analysis (PCA) [1], and formulate 3DMMs as linear models by the generated bases. After that, expression models [69, 45, 10] are proposed to support face manipulation. Recent advances [12, 55, 65, 18, 44] are proposed to expand the expressiveness of 3DMMs and play a crucial role in 3D face reconstruction. 3DMMs make it possible to simplify the 2D-to-3D problem into a regression task, which typically presents an analysis-by-synthesis fashion to estimate the coefficients of 3DMMs. In this paper, we follow the spirit of the 3DMMs family to present the decoupled static and dynamic details for 3D face reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Coarse Shape Reconstruction. Traditional optimization-based methods [27, 73, 3, 77, 2] directly optimize the 3DMM coefficients of given 2D images. While such methods work well in controlled settings (e.g., frontal view, no occlusion), they heavily rely on high-quality annotations. Learning-based methods leverage the advances of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9088" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 67, + 317, + 198 + ], + "blocks": [ + { + "bbox": [ + 50, + 67, + 317, + 198 + ], + "lines": [ + { + "bbox": [ + 50, + 67, + 317, + 198 + ], + "spans": [ + { + "bbox": [ + 50, + 67, + 317, + 198 + ], + "type": "image", + "image_path": "57afa58db775560ad3b2a545497d24eb5ac60f6ae79d4b0bdca43cf3ec9370bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "lines": [ + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "spans": [ + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "text", + "content": "p: pose coef " + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "text", + "content": ": identity coef " + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "text", + "content": ": expression coef " + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "text", + "content": ": static coef " + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "text", + "content": ": albedo coef " + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "text", + "content": ": SH coef " + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 123, + 202, + 483, + 213 + ], + "type": "text", + "content": ": dynamic coef" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 220, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 545, + 276 + ], + "type": "text", + "content": "Figure 2. Illustration of HiFace. (a). Learning framework of HiFace. Given a monocular image, we regress its shape and detail coefficients to synthesize a realistic 3D face, and leverage a differentiable renderer [28] to train the whole model end-to-end from synthetic [72, 57] and real-world [40, 52] images. (b). The pipeline of Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). We explicitly decouple the static and dynamic factors to synthesize realistic and animatable details. Given the shape and static coefficients, we regress the static and dynamic details through displacement bases and interpolate them into the final details through vertex tension [57]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 318, + 66, + 543, + 198 + ], + "blocks": [ + { + "bbox": [ + 318, + 66, + 543, + 198 + ], + "lines": [ + { + "bbox": [ + 318, + 66, + 543, + 198 + ], + "spans": [ + { + "bbox": [ + 318, + 66, + 543, + 198 + ], + "type": "image", + "image_path": "f1bab0659a69cf4081953d6d51bcec4f778bbde8640851b31fa33b5e9f881ef7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 284, + 287, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 404 + ], + "type": "text", + "content": "CNNs [67, 21, 19, 62] and GCNs [47, 41, 25] to learn high-level representations from large-scale images in the wild. These methods show plausible generalization over diverse environments. To train the network end-to-end, recent methods leverage the differentiable renderers [28, 22, 80, 49], along with the photo loss, perceptual loss, and landmark loss [21, 19, 28, 67, 71] to optimize the network in a self-supervised manner. Different from these coarse shape reconstruction methods, we aim at high-fidelity 3D face reconstruction with both coarse shape and fine details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 606 + ], + "type": "text", + "content": "Detail Reconstruction. While 3DMMs can reconstruct coarse 3D face shapes from 2D images, they struggle with reconstructing fine-level details, such as forehead wrinkles and crows-feet. To fill this gap, shape by shading (SfS) [34, 46, 26, 66] methods reconstruct the facial details from monocular images or videos. However, these methods are sensitive to occlusions and large poses. Recent advances [24, 48, 15, 13, 42] leverage displacement maps to present details. These methods explicitly re-topologize the coarse shape and present residual bias to generate geometric details. The main challenge of detail reconstruction is the difficulty in learning the nuances and disentangling the static and dynamic details from only self-supervised learning. Ground-truth labels of the details are helpful to guide the learning process. However, it is difficult to obtain such fine-grained labels on real data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Synthetic Dataset. Several methods [81, 77, 33, 68, 51] utilize rendered faces or fitted coefficients to synthesize 3D-2D pairs. These ground-truth pairs lack diversity over background, illumination, and assets, making them hard to generalize well to real-world images. Recent advances in synthetic data generation [72, 57] demonstrate its ability to generalize to real-world settings, and diverse to compensate for the domain gap to real-world images. In this paper, we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 284, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 284, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 545, + 309 + ], + "type": "text", + "content": "leverage high-quality data with ground-truth labels to explore the detailed 3D face reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 319, + 388, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 388, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 388, + 334 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 340, + 385, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 385, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 385, + 353 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 545, + 394 + ], + "type": "text", + "content": "We adopt a common practice [21, 24] to represent a textured coarse shape with a 3D face model, an illumination model, and a camera model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 400, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 545, + 424 + ], + "type": "text", + "content": "3D Face Model. The 3D shape " + }, + { + "bbox": [ + 305, + 400, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 305, + 400, + 545, + 424 + ], + "type": "text", + "content": " and albedo " + }, + { + "bbox": [ + 305, + 400, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 305, + 400, + 545, + 424 + ], + "type": "text", + "content": " are represented by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 425, + 545, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 425, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 377, + 425, + 545, + 443 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\bar {\\mathbf {S}} + \\beta \\mathbf {B} _ {\\mathrm {i d}} + \\boldsymbol {\\xi} \\mathbf {B} _ {\\exp}", + "image_path": "4b0741d94a350a1386393f2b0e82f287804613c65f859628a6fcb9e24eac01b9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 440, + 444, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 440, + 444, + 452 + ], + "spans": [ + { + "bbox": [ + 375, + 440, + 444, + 452 + ], + "type": "interline_equation", + "content": "\\mathbf {A} = \\bar {\\mathbf {A}} + \\alpha \\mathbf {B} _ {\\mathrm {a l b}}", + "image_path": "00ed5b6bba3236aea54ccb1371892a548a461c3cf2b8e894906b98dc578294c9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{S}}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{A}}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " are the mean shape and albedo. " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{id}}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{exp}}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{alb}}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " are bases [73] of 256-dim identity, 233-dim expression, and 300-dim albedo, respectively. The coarse shape " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " in the bind pose is deformed from a neutral shape " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{neu}} = \\bar{\\mathbf{S}} + \\alpha \\mathbf{B}_{\\mathrm{id}}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " with expression component " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\xi \\mathbf{B}_{\\mathrm{exp}}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " are the corresponding identity, expression, and albedo coefficients for generating a coarse shape. In this paper, the coarse shape " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "n_v = 7" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": ", 667 vertices and " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "n_f = 14" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": ", 832 triangles with " + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "inline_equation", + "content": "512 \\times 512 \\times 3" + }, + { + "bbox": [ + 304, + 458, + 545, + 567 + ], + "type": "text", + "content": " albedo." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": "Pose & Camera Model. To estimate the face pose, we follow [72, 73] to predict skeletal pose " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [\\theta | \\mathbf{t}]" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{3j}" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": " are the local joint rotations and root joint translation, respectively. " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "j = 4" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": " indicates 4 skeletal joints w.r.t. the head, neck, and two eyes. We perform a standard linear blend skinning (LBS) function [43] (with per-vertex weights " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{j \\times n_v}" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": ") to rotate " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": " about joint locations " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\in \\mathbb{R}^{3j}" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{S_p}" + }, + { + "bbox": [ + 304, + 572, + 545, + 669 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 679, + 545, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 679, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 375, + 679, + 545, + 692 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {\\mathbf {p}} = \\operatorname {L B S} (\\mathbf {S}, \\mathbf {p}, \\mathbf {J}; \\mathbf {W}), \\tag {2}", + "image_path": "7ed54454f0f74eb8a18826e769914d2c29f8239ec2606f42f8b7df27d0f2110a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " is the joint locations in the bind pose determined" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9089" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{J} = \\mathcal{J}(\\beta):\\mathbb{R}^{|\\beta |}\\to \\mathbb{R}^{3j}" + }, + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "content": ". Then we use an orthographic camera model to project 3D vertices in " + }, + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathbf{p}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "content": " to the 2D plane." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 102, + 287, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 102, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 46, + 102, + 287, + 138 + ], + "type": "text", + "content": "Illumination Model. We follow previous work [21] to use Spherical Harmonics (SH) [56] to estimate the illumination of a given image. The shaded texture " + }, + { + "bbox": [ + 46, + 102, + 287, + 138 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 102, + 287, + 138 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 144, + 287, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 144, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 287, + 166 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = \\mathbf {A} \\odot \\sum_ {k = 1} ^ {9} \\gamma_ {k} \\Psi_ {k} (\\mathbf {N}), \\tag {3}", + "image_path": "513e64fd256df4361d595757ced74536b57647115552e1700c1b7abd779ceb6c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "text", + "content": " denotes the Hadamard product, " + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "text", + "content": " is the surface normal of " + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "text", + "content": " in UV coordinates, " + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\Psi : \\mathbb{R}^3 \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "text", + "content": " are SH basis function and " + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^9" + }, + { + "bbox": [ + 46, + 171, + 287, + 208 + ], + "type": "text", + "content": " is the corresponding SH coefficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 213, + 162, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 162, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 162, + 225 + ], + "type": "text", + "content": "3.2. Overview of HiFace" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 236, + 287, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 295 + ], + "type": "text", + "content": "Key Idea. The key idea of HiFace is to explicitly model the static (e.g., person-specific properties) and dynamic (e.g., expression-driven wrinkles) details, allowing the model to reconstruct a high-fidelity 3D face from a single image with realistic and animatable details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 301, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 456 + ], + "type": "text", + "content": "Overview. The goal of HiFace is to reconstruct 3D shapes with realistic details from a single image. The overview of HiFace is illustrated in Fig. 2(a). We leverage a feature extractor (i.e., ResNet-50 [30]) to regress corresponding coefficients from an input image. Our model jointly predicts both the coarse-level shapes and the fine-level details. For coarse-level shapes, we regress shape parameters (i.e., identity, expression, albedo, illumination, and pose) of a parametric face model. For the fine-level details, we propose a novel way to model it through the separation of static and dynamic factors and formulate the generation of details into the problems of 3DMM coefficients regression and displacement maps interpolation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 457, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 576 + ], + "type": "text", + "content": "Note that the facial details are based on the coarse shape, we thereby exploit novel loss functions to learn 3D representations of coarse shape and details simultaneously from the synthetic dataset with ground-truth labels. To generalize our model to real-world images, we also present several self-supervised losses to train the model with both synthetic data and real-world images coherently. As a result, HiFace can faithfully reconstruct the facial details of a given image, or animate a face by combining the decoupled static and dynamic coefficients that come from different individuals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 582, + 253, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 253, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 253, + 596 + ], + "type": "text", + "content": "3.3. Decoupling Static and Dynamic Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 601, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 637 + ], + "type": "text", + "content": "We propose Static and Dynamic Decoupling for DeTail Reconstruction (SD-DeTail). The facial details are basically composed of a static factor and a dynamic factor:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 645, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 645, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 128, + 645, + 287, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {D} = \\mathbf {D} _ {\\mathrm {s t a}} + \\mathbf {D} _ {\\mathrm {d y n}}, \\tag {4}", + "image_path": "948d08e7866cd967ebfa806512581544e1d347beba37f7fe03f7c916a9bf2173.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{sta}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{dyn}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "content": " indicate details from static and dynamic factors, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Concretely, the static factor is the inherent property of the identity (i.e., the given 2D face), and originates from the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "appearance and age attributes. As for the dynamic factor, it is typically driven by the expression and influenced by person-specific properties." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "text", + "content": "Static Detail Generation. To simplify the problem, we are inspired by 3DMMs, which parameterize the statistical models to simplify the 2D-to-3D problem. We build a 300-dim displacement basis " + }, + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{sta}}" + }, + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "text", + "content": " from the captured 332 scans [57] by PCA [1]. The scans contain diverse age groups in a neutral expression. Then we regress the coefficient " + }, + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "text", + "content": " to synthesize the static detail " + }, + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{sta}}" + }, + { + "bbox": [ + 304, + 114, + 545, + 198 + ], + "type": "text", + "content": " from the image:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 380, + 209, + 545, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 209, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 380, + 209, + 545, + 222 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {s t a}} = \\bar {\\mathbf {D}} _ {\\mathrm {s t a}} + \\varphi \\mathbf {B} _ {\\mathrm {s t a}}, \\tag {5}", + "image_path": "ed7fd359efd90c9a9a4394a7a31093433527854a5c46c2fd37db1eb40aee0d35.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 226, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 545, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 226, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{D}}_{\\mathrm{sta}}" + }, + { + "bbox": [ + 304, + 226, + 545, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 226, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{sta}}" + }, + { + "bbox": [ + 304, + 226, + 545, + 251 + ], + "type": "text", + "content": " are the mean displacement map and displacement basis for static details, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 255, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 375 + ], + "type": "text", + "content": "Dynamic Detail Generation. Due to the high diversity and complexity of expression representation, directly generating dynamic details from expression is quite difficult. Therefore we simplify the expression representation by using an interpolation between two displacement maps: compressed and stretched [57]. For example, the compressed expression may indicate a state of frowning to the extreme, while the stretched expression may indicate a state of complete relaxation between the eyebrows. Other states of this area can be interpolated by these two polarized states." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": "Consequently, we generate the dynamic details through compressed and stretched displacement maps. Again, we build 26-dim compressed " + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{com}}" + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": " and stretched " + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{str}}" + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": " displacement bases by PCA [1] to simplify the generation of displacement maps. To generate the dynamic coefficients " + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\phi = \\{\\phi_{\\mathrm{com}},\\phi_{\\mathrm{str}}\\}" + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": ", we apply the expression coefficient " + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": " into the static coefficient " + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": " through AdaIN [32], followed by the MLP transformation " + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 376, + 545, + 471 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 479, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 479, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 347, + 479, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\phi = \\Phi \\left(\\sigma (\\tilde {\\xi}) \\left(\\frac {\\varphi - \\mu (\\varphi)}{\\sigma (\\varphi)} + \\mu (\\tilde {\\xi})\\right)\\right), \\tag {6}", + "image_path": "58cb5b491a77f61ac9f1fef3aba07fd0c147015d1fdc8d5d9ea2ba028bbb1cce.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\xi}" + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "content": " is the affine vector from " + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "content": " via MLP transformation. " + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "content": " indicate the mean and standard deviation. " + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{com}}" + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{str}}" + }, + { + "bbox": [ + 304, + 513, + 545, + 561 + ], + "type": "text", + "content": " are coefficients for compressed and stretched displacement maps respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 561, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 585 + ], + "type": "text", + "content": "Similar to Eq. 5, the compressed and stretched displacement maps are formulated as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 368, + 593, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 593, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 368, + 593, + 545, + 611 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\text {c o m}} = \\bar {\\mathbf {D}} _ {\\text {c o m}} + \\phi_ {\\text {c o m}} \\mathbf {B} _ {\\text {c o m}}, \\tag {7}", + "image_path": "b70b470d0bf2cffcc13382bf656a4f51f0d45d09f98309f829e71d2b7195340c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 374, + 609, + 465, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 609, + 465, + 620 + ], + "spans": [ + { + "bbox": [ + 374, + 609, + 465, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {s t r}} = \\bar {\\mathbf {D}} _ {\\mathrm {s t r}} + \\phi_ {\\mathrm {s t r}} \\mathbf {B} _ {\\mathrm {s t r}}", + "image_path": "09ca8e9cccf73d2602be04ef0371ce2d3718d8b48620020118aa40e95e6d7726.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{D}}_{\\mathrm{com}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{com}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " are the mean displacement map and 26-dim displacement basis for compressed detail, and " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{D}}_{\\mathrm{str}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{str}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " are the mean displacement map and 26-dim displacement basis for stretched detail, respectively." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Considering the coarse shape " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " can be obtained by deforming the neutral shape " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{neu}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " with the expression component " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\xi \\mathbf{B}_{\\mathrm{exp}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", such expression-driven deformation over face" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9090" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 79, + 78, + 138 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 69, + 77 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 69, + 77 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 69, + 77 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 79, + 78, + 138 + ], + "lines": [ + { + "bbox": [ + 49, + 79, + 78, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 79, + 78, + 138 + ], + "type": "image", + "image_path": "683427d6bcce4e6e8970f926cbb722850d450993e6ec2ec4517ec88f998b8055.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 80, + 79, + 138, + 137 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 113, + 76 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 113, + 76 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 113, + 76 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 79, + 138, + 137 + ], + "lines": [ + { + "bbox": [ + 80, + 79, + 138, + 137 + ], + "spans": [ + { + "bbox": [ + 80, + 79, + 138, + 137 + ], + "type": "image", + "image_path": "0397e8454b8e2dab65c85962c1dfc55674ac8ca3fb58efda01a9c8d780fc8ebb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 138, + 79, + 167, + 137 + ], + "blocks": [ + { + "bbox": [ + 138, + 79, + 167, + 137 + ], + "lines": [ + { + "bbox": [ + 138, + 79, + 167, + 137 + ], + "spans": [ + { + "bbox": [ + 138, + 79, + 167, + 137 + ], + "type": "image", + "image_path": "7be66b7330bf156f511f2a67d024a6b814e669912e15518fb633195954a842f7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 167, + 79, + 226, + 137 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 184, + 77 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 184, + 77 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 184, + 77 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 79, + 226, + 137 + ], + "lines": [ + { + "bbox": [ + 167, + 79, + 226, + 137 + ], + "spans": [ + { + "bbox": [ + 167, + 79, + 226, + 137 + ], + "type": "image", + "image_path": "b95feed28d83c8c8d7da6e1b36f3170cb768f5e131c994ab9cfa012f26dfaaa2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 227, + 79, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 250, + 70, + 260, + 76 + ], + "lines": [ + { + "bbox": [ + 250, + 70, + 260, + 76 + ], + "spans": [ + { + "bbox": [ + 250, + 70, + 260, + 76 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 79, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 227, + 79, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 227, + 79, + 286, + 137 + ], + "type": "image", + "image_path": "80a23f5a1132fc3dfa2a958e1e5e61b3bb51fc9d3cd21435e8a642b7de033872.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 139, + 78, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 139, + 78, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 139, + 78, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 78, + 198 + ], + "type": "image", + "image_path": "7c450e66f055ea49c699839bb5e03a8a6fba7822077845a6c6faf4a1af722e93.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 207, + 287, + 274 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 287, + 274 + ], + "type": "text", + "content": "Figure 3. Illustration of displacement map composition in SD-DeTail. Given (a). an image (top) to reconstruct its coarse shape (bottom), we formulate the detail as (b). a static factor and (c). a dynamic factor interpolated by polarized states w.r.t. compressed (top) and stretched (bottom). (d). the output displacement map is linearly combined by (b) and (c) to present vivid details." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 79, + 139, + 138, + 198 + ], + "blocks": [ + { + "bbox": [ + 79, + 139, + 138, + 198 + ], + "lines": [ + { + "bbox": [ + 79, + 139, + 138, + 198 + ], + "spans": [ + { + "bbox": [ + 79, + 139, + 138, + 198 + ], + "type": "image", + "image_path": "8ba885db0ffc073f09833f3b7d55c8d78f941f2b654db755ba34d7662fc18deb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 138, + 139, + 226, + 198 + ], + "blocks": [ + { + "bbox": [ + 138, + 139, + 226, + 198 + ], + "lines": [ + { + "bbox": [ + 138, + 139, + 226, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 226, + 198 + ], + "type": "image", + "image_path": "a6224bc791411724dc63d3bc4a0333a605da9ea9ece01529d419a9242c3618f1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 227, + 139, + 286, + 198 + ], + "blocks": [ + { + "bbox": [ + 227, + 139, + 286, + 198 + ], + "lines": [ + { + "bbox": [ + 227, + 139, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 227, + 139, + 286, + 198 + ], + "type": "image", + "image_path": "337bc07fc3e509321f5483ea1463f1ac3f6891d9eb7aa025e00f383e82c14a17.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": "shape yields the \"tension\" over each vertex [57], which influences facial details from expression. Since " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{neu}}" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": " possesses the same topology, for each vertex " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i \\in \\mathbf{S}" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": "-edges " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "E_i = \\{e_1, \\dots, e_K\\}" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": " connected with " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "E_i' = \\{e_1', \\dots, e_K'\\}" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": " are the corresponding edges in " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{neu}}" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": " that are connected to " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i'" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": ". Then the tension at " + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 47, + 284, + 287, + 356 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 108, + 365, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 365, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 108, + 365, + 287, + 392 + ], + "type": "interline_equation", + "content": "t _ {\\mathbf {v} _ {i}} = 1 - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\frac {\\left\\| e _ {k} \\right\\|}{\\left\\| e _ {k} ^ {\\prime} \\right\\|}, \\tag {8}", + "image_path": "ba5da099b032c970ba9d232263baf10f4acec509c8bf4794333f914cca4222a4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": " represents the edge length. Positive values of " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "t_{\\mathbf{v}_i}" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": " indicate compression, negative values indicate stretch, and 0-value indicates no change, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": "The vertex tension " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "t_{\\mathbf{v}_i}" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": " composes the tension map " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{uv}}" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": " in UV coordinates. Then, the displacement map of the dynamic detail is the linear interpolation of " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{com}}" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{str}}" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 91, + 483, + 287, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 483, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 91, + 483, + 287, + 497 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\mathrm {d y n}} = \\mathbf {M} _ {\\mathrm {u v}} ^ {+} \\odot \\mathbf {D} _ {\\mathrm {c o m}} + \\mathbf {M} _ {\\mathrm {u v}} ^ {-} \\odot \\mathbf {D} _ {\\mathrm {s t r}}, \\tag {9}", + "image_path": "7411be4cd68527dc4f575da3b9aea157ef48e6f5c5a78673a113c9270a614c9b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{uv}}^{+}" + }, + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{uv}}^{-}" + }, + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "text", + "content": " indicate the positive and negative value of " + }, + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{uv}}" + }, + { + "bbox": [ + 47, + 506, + 287, + 567 + ], + "type": "text", + "content": ", respectively. Fig. 3 shows the effectiveness of SD-DeTail. The dynamic factor interpolated by two polarized states introduces expression-related details and further decorates the static detail, yielding the final vivid output." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 574, + 178, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 178, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 178, + 586 + ], + "type": "text", + "content": "3.4. Overall Loss Functions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": "We propose several loss functions to train HiFace end-to-end. Specifically, we use static and dynamic detail losses to supervise the synthesized displacement maps from " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ". In addition, we leverage the coarse shape loss to supervise the reconstructed shape from " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ". Finally, we follow previous methods [21, 24, 73] to leverage the differentiable renderer [28] to map the generated 3D shape into 2D images, by combining " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 594, + 287, + 713 + ], + "type": "text", + "content": ". Then, we perform self-supervised losses to train in both synthetic and real-world images. See more details in the supplementary." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "Static and Dynamic Detail Losses. To explicitly train the details of each component, we leverage the ground-truth annotations from the synthetic dataset [57, 72] as supervision to assist the training process of our model. Specifically, we calculate the detail losses by estimating the " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " distance between the reconstructed displacement maps and ground-truth w.r.t. static, compressed, and stretched components, and summarize them as " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{detail}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 347, + 175, + 487, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 175, + 487, + 191 + ], + "spans": [ + { + "bbox": [ + 347, + 175, + 487, + 191 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s t a}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {s t a}} - \\hat {\\mathbf {D}} _ {\\mathrm {s t a}}\\right) \\right\\| _ {2}", + "image_path": "32093ca32367d1d61aa091a9ca7c0a50fdca5d041bbd88f4a00c140e102aea3a.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 348, + 194, + 545, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 194, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 348, + 194, + 545, + 213 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o m}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {c o m}} - \\hat {\\mathbf {D}} _ {\\mathrm {c o m}}\\right) \\right\\| _ {2}, \\tag {10}", + "image_path": "eb09731a55fb1cff09f33efeb9a64677678ec87404702b2271b95fd4d5d5bef6.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 348, + 212, + 485, + 226 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 212, + 485, + 226 + ], + "spans": [ + { + "bbox": [ + 348, + 212, + 485, + 226 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s t r}} = \\left\\| \\mathbf {M} _ {\\text {d e t a i l}} \\odot \\left(\\mathbf {D} _ {\\mathrm {s t r}} - \\hat {\\mathbf {D}} _ {\\mathrm {s t r}}\\right) \\right\\| _ {2}", + "image_path": "6f6e4195b6b3affdf897652cf0bf3124c34aed670ba3e49a7c4d33307830abce.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 348, + 228, + 460, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 228, + 460, + 241 + ], + "spans": [ + { + "bbox": [ + 348, + 228, + 460, + 241 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e t a i l}} = \\mathcal {L} _ {\\text {s t a}} + \\mathcal {L} _ {\\text {c o m}} + \\mathcal {L} _ {\\text {s t r}}", + "image_path": "2f6c4e67d11e293b3c57cf23c738ed494aa722b6e34e86ac0b08ffde3c4b04df.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{detail}}" + }, + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "text", + "content": " is the facial mask in the UV coordinates, and " + }, + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}_{\\mathrm{sta}} / \\hat{\\mathbf{D}}_{\\mathrm{com}} / \\hat{\\mathbf{D}}_{\\mathrm{str}}" + }, + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{sta}} / \\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}" + }, + { + "bbox": [ + 306, + 249, + 545, + 285 + ], + "type": "text", + "content": " are the reconstructed and ground-truth displacement maps, respectively." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 290, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 338 + ], + "type": "text", + "content": "Coarse Shape Losses. Since the details should be based on realistic coarse shapes, we train the coarse shape to help the learning of details by leveraging the ground-truth vertex as supervision:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 367, + 346, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 346, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 367, + 346, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v e r}} = \\left\\| \\mathbf {M} _ {\\text {v e r}} \\odot (\\mathbf {S} - \\hat {\\mathbf {S}}) \\right\\| _ {2}, \\tag {11}", + "image_path": "f216c4e53f7469ad91dc0ea6223c6850536bc4d3ce631137169a386e46ea38ec.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{ver}}" + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "text", + "content": " is frontal face area of the coarse shape. " + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "text", + "content": " are the reconstructed and ground-truth face by Eq. 1." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 431 + ], + "type": "text", + "content": "In addition, we make constraints on shape coefficients to prevent overfitting. We enforce the predicted coefficients have a similar distribution to the ground-truth coefficients:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 353, + 439, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 439, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 353, + 439, + 545, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k l}} = \\rho (\\boldsymbol {\\beta}) \\left(\\log \\rho (\\boldsymbol {\\beta}) - \\log \\rho (\\hat {\\boldsymbol {\\beta}})\\right), \\tag {12}", + "image_path": "8d4daaeb05bebf943d594146d92313dc428f253fa359fa3eab15f06a6fd5094d.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "text", + "content": " denotes softmax function to map the predicted coefficients " + }, + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}" + }, + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "text", + "content": " and ground-truth " + }, + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 462, + 544, + 486 + ], + "type": "text", + "content": " into probability distribution." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 486, + 494, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 486, + 494, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 486, + 494, + 498 + ], + "type": "text", + "content": "Finally, the shape loss is " + }, + { + "bbox": [ + 317, + 486, + 494, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shp}} = \\mathcal{L}_{\\mathrm{ver}} + \\mathcal{L}_{\\mathrm{kl}}" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "content": "Self-supervised Losses. To encourage the generalization of our models in real-world images [40, 52], we follow previous methods [24, 21, 73] to leverage self-supervised loss " + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{self}}" + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "content": " for all training images, including photo loss " + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pho}}" + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "content": ", perceptual loss " + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{id}}" + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "content": ", and dense landmark loss " + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lmk}}" + }, + { + "bbox": [ + 304, + 503, + 545, + 563 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 356, + 572, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 572, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 356, + 572, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\text {p h o}} + \\lambda_ {\\mathrm {i d}} \\mathcal {L} _ {\\mathrm {i d}} + \\lambda_ {\\mathrm {l m k}} \\mathcal {L} _ {\\mathrm {l m k}}, \\tag {13}", + "image_path": "35e564ec7fc15323fa03b3b0e476b92af0798e4b0376d4a6141efa34e52e7cac.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 594, + 544, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 544, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 544, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 594, + 544, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{id}}" + }, + { + "bbox": [ + 304, + 594, + 544, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 544, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{lmk}}" + }, + { + "bbox": [ + 304, + 594, + 544, + 616 + ], + "type": "text", + "content": " are weights to balance the self-supervised losses term." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "In addition, considering the static detail heavily correlates to person-specific age attribute, inspired by [19], we leverage the pre-trained age prediction network [36] to learn high-level representations of static details through knowledge distillation, such that the learned coefficients exhibit expressive results. To achieve this, we use several MLP layers on the static coefficient " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ", and map it into age classification probabilities " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{\\mathrm{age}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ". Then we use the pre-trained" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9091" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 125, + 545, + 280 + ], + "blocks": [ + { + "bbox": [ + 46, + 68, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 46, + 68, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 547, + 125 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of 3D face reconstruction methods on REALY benchmark. “-c” and “-d” indicate coarse and detail shape, respectively. @RN/@RM/@RF/@RC/all indicate errors in nose/mouth/forehead/cheek/all regions. We highlight the best method for the two groups respectively. HiFace achieves the best reconstruction performance in the overall error by a large margin. Each component in HiFace contributes to a better reconstruction quality. The reconstructed details of HiFace further boost the quality while previous methods [24, 19] modeling details with only image-level supervision even deteriorate the reconstruction accuracy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 125, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 48, + 125, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 48, + 125, + 545, + 280 + ], + "type": "table", + "html": "
GroupMethods / e (mm)frontal-viewside-view
@RN@RM@RF@RCall@RN@RM@RF@RCall
CoarseDeep3D [21]1.719±0.3541.368±0.4392.015±0.4491.528±0.5011.6571.749±0.3431.411±0.3952.074±0.4861.528±0.5171.691
MGCNet [64]1.771±0.3801.417±0.4092.268±0.5031.639±0.6501.7741.827±0.3831.409±0.4182.248±0.5081.665±0.6441.787
3DDFA-v2 [29]1.903±0.5171.597±0.4782.447±0.6471.757±0.6421.9261.883±0.4991.642±0.5012.465±0.6221.781±0.6361.943
DECA-c [24]1.694±0.3552.516±0.8392.394±0.5761.479±0.5352.0101.903±1.0502.472±1.0792.423±0.7201.630±1.1352.107
SADRNet [61]1.791±0.5421.591±0.4882.413±0.5371.856±0.7011.9131.771±0.5211.560±0.4622.490±0.5662.010±0.7151.958
EMOCA-c [19]1.868±0.3872.679±1.1122.426±0.6411.438±0.5012.1031.867±0.5542.636±1.2842.448±0.7081.548±0.5902.125
MICA [81]1.585±0.3253.478±1.2042.374±0.6831.099±0.3242.1341.525±0.3223.567±1.2122.379±0.6751.109±0.3252.145
Ours-c (w/o Syn. Data)†1.227±0.4071.787±0.4391.454±0.3821.762±0.4361.5581.187±0.3791.826±0.4901.470±0.4261.653±0.4501.534
Ours-c1.054±0.3171.461±0.4301.331±0.3471.342±0.3841.2970.992±0.2461.505±0.4541.427±0.4001.439±0.4291.341
DetailDECA-d [24]2.138±0.4612.802±0.8682.457±0.5591.443±0.4982.2102.286±1.1032.684±1.0412.519±0.7181.555±0.8222.261
EMOCA-d [19]2.532±0.5392.929±1.1062.595±0.6311.495±0.4692.3882.455±0.6362.948±1.2922.606±0.6861.599±0.5632.402
HRN [42]1.722±0.3301.357±0.5231.995±0.4761.072±0.3331.5371.642±0.3101.285±0.5281.906±0.4791.038±0.3221.468
Ours-d (w/o Syn. Data)†1.465±0.5571.790±0.4251.528±0.3731.618±0.3621.6001.422±0.5371.849±0.4731.530±0.4141.572±0.3991.594
Ours-d (w/o static)*1.055±0.2901.469±0.4151.336±0.3371.319±0.3741.2951.004±0.2331.491±0.4371.418±0.3921.418±0.4151.332
Ours-d (w/o dynamic)*1.069±0.3181.469±0.4141.358±0.3361.270±0.3441.2920.991±0.2391.496±0.4371.411±0.3931.375±0.4021.318
Ours-d1.036±0.2801.450±0.4131.324±0.3341.291±0.3621.2750.985±0.2371.489±0.4361.399±0.3881.360±0.3951.308
", + "image_path": "f8c0798aeb8017fa0f2680a0a5f2e764871bfddab6d8cd95821dede146a2e434.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 49, + 281, + 432, + 289 + ], + "lines": [ + { + "bbox": [ + 49, + 281, + 432, + 289 + ], + "spans": [ + { + "bbox": [ + 49, + 281, + 432, + 289 + ], + "type": "text", + "content": "To align the dataset scale, w/o Syn. Data indicates we train the model without using the ground-truth labels from the synthetic dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 50, + 289, + 545, + 298 + ], + "lines": [ + { + "bbox": [ + 50, + 289, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 50, + 289, + 545, + 298 + ], + "type": "text", + "content": "* To eliminate the bias of coarse shape in estimating the reconstruction error, we fix the coarse shape and train the details with/without static and dynamic factors for comparisons." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": "age recognition model " + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{age}}" + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": " to obtain the probabilities of the given input image " + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": ". The distillation loss " + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": " enforces the probabilities between " + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{\\mathrm{age}}" + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{age}}(\\mathbf{I})" + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": " be similar:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 357, + 287, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 357, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 86, + 357, + 287, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k d}} = \\Gamma_ {\\text {a g e}} (\\mathbf {I}) \\left(\\log \\Gamma_ {\\text {a g e}} (\\mathbf {I}) - \\log \\hat {\\mathbf {p}} _ {\\text {a g e}}\\right). \\tag {14}", + "image_path": "d6b32315b209ddeacdf03bb4cf294a2d4d483c8acac7c9895fa518bc748e18db.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "text", + "content": "Regularization. " + }, + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "text", + "content": " regularizes coefficients of each submodule, by minimizing the " + }, + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "text", + "content": " loss of " + }, + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\xi, \\varphi, \\phi" + }, + { + "bbox": [ + 47, + 378, + 287, + 408 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "content": "Overall Loss Function. We train the coarse shape and fine details simultaneously, such that each component can collaborate to reconstruct high-fidelity 3D faces with realistic details. Formally, we minimize the total loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 95, + 469, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 469, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 95, + 469, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {d e t a i l}} \\mathcal {L} _ {\\text {d e t a i l}} + \\lambda_ {\\text {s h p}} \\mathcal {L} _ {\\text {s h p}} \\tag {15} \\\\ + \\lambda_ {\\text {s e l f}} \\mathcal {L} _ {\\text {s e l f}} + \\lambda_ {\\text {k d}} \\mathcal {L} _ {\\text {k d}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\\\ \\end{array}", + "image_path": "5af225cb2cbb48a82cf1dc1f13f0717ca339ac822af052a34dfcb2bb87058e00.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 504, + 220, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 220, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 220, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 504, + 220, + 516 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 504, + 220, + 516 + ], + "type": "text", + "content": " is the weight for each component." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 526, + 128, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 128, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 128, + 540 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 545, + 180, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 180, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 180, + 558 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Dataset. We use a hybrid dataset made up from both synthetic [72, 57] and real data [40, 52]. We use the synthetic data pipeline [72, 57] to generate a diverse dataset of " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "200k" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " faces with ground-truth vertex, landmark, albedo, and displacement map annotations. The real-world datasets contain " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "400k" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " images in total from diverse age, gender, and ethnicity groups. For the real-world dataset, we use the pretrained dense landmark detector [72] to detect 669 landmarks for training. We use face parsing [79] to generate and select region-of-interest as facial masks, providing robustness to common occlusions by hair or other accessories. We follow [19, 73, 21] to split the dataset into training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 314, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 338 + ], + "type": "text", + "content": "and validation sets. The test images are from CelebA [40], FFHQ [37], LS3D-W [8], and AFLW2000 [78]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": "Implementation Details. We implement HiFace in PyTorch [54] and leverage the PyTorch3D differentiable rasterizer [35] for rendering. We train our model for 35 epochs on " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": " NVIDIA Tesla V100 GPUs with a mini-batch of 320. We use the pre-trained ResNet-50 on ImageNet [20] as initialization, and use Adam [39] as optimizer with an initial learning rate of " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ". The input image is cropped and aligned by [14], and resized into " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ". We empirically set " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{detail}} = 10" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{shp}} = 1" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{self}} = 1" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{id}} = 0.1" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{lmk}} = 0.5" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{kd}} = 1" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}} = 1e - 3" + }, + { + "bbox": [ + 305, + 343, + 546, + 464 + ], + "type": "text", + "content": " throughout the experiments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 472, + 478, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 478, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 478, + 486 + ], + "type": "text", + "content": "4.2. Comparisons to State-of-the-art" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 628 + ], + "type": "text", + "content": "Quantitative Comparison. We perform the quantitative evaluation on the REALY benchmark [12], which contains 100 frontal-view and 400 side-view images from 100 textured-scans [17]. The REALY benchmark presents a region-aware evaluation pipeline to separately evaluate the metric error (in mm) of the nose, mouth, forehead, and cheek regions. Such an evaluation pipeline is demonstrated to better estimate the actual similarity of the 3D faces and align with human perception. We compare HiFace to previous state-of-the-art methods and report the region-wise and average normalized mean square error (NMSE) in Tab. 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "As Tab. 1 illustrates, HiFace outperforms prior arts in the overall error by a large margin. HiFace balances the reconstruction quality on each region, compared to those optimum region methods that may fail in specific regions (e.g., MICA [81] fails in mouth region while HRN [42] fails in forehead region). Note that HiFace faithfully recovers the facial details, thus making the reconstruction error smaller" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9092" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 77, + 288, + 215 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 278, + 77 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 278, + 77 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 278, + 77 + ], + "type": "text", + "content": "Input Deep3D 3DDFA-v2 SynergyNet MICA Dense Ours" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 77, + 288, + 215 + ], + "lines": [ + { + "bbox": [ + 49, + 77, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 77, + 288, + 215 + ], + "type": "image", + "image_path": "bc7b84348e7ae97825277b3dd80793bb64fa799ec276a3f96de851633c8c80a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 263, + 287, + 351 + ], + "blocks": [ + { + "bbox": [ + 46, + 223, + 288, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 288, + 257 + ], + "type": "text", + "content": "Figure 4. Comparison on coarse shape reconstruction. From left to right: Input image, Deep3D [21], 3DDFA-v2 [29], Synerg-yNet [74], MICA [81], Dense [73], and HiFace (Ours)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 263, + 287, + 351 + ], + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 351 + ], + "type": "image", + "image_path": "8e009b9dce3c38cf8a273e6fcceabbc11968dbd58488edb3fea263f32a64a4bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 360, + 287, + 405 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 405 + ], + "type": "text", + "content": "Figure 6. Illustration on the flexibility of SD-DeTail. Given the identity and expression coefficients " + }, + { + "bbox": [ + 46, + 360, + 287, + 405 + ], + "type": "inline_equation", + "content": "(\\beta, \\xi)" + }, + { + "bbox": [ + 46, + 360, + 287, + 405 + ], + "type": "text", + "content": " from the optimization-based method [73], SD-DeTail can generate realistic details based on the coarse shape and further improve the visual quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 417, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 464 + ], + "type": "text", + "content": "than using the coarse shape alone. As a comparison, although DECA [24] and EMOCA [19] can reconstruct details of given images, they turn out to be noisy, leading to the deterioration of reconstruction quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 598 + ], + "type": "text", + "content": "In addition, Tab. 1 also demonstrates the necessity of each component in contributing to a better quality. It can be observed that the synthetic data with ground-truth labels not only improve the coarse shape reconstruction quality but is also crucial for detailed reconstruction. With the synthetic data, the proposed SD-DeTail further boosts the overall reconstruction quality. Both the static and dynamic factors are essential to capture fine-grained details, and the final SD-DeTail achieves the most accurate details in expression-related regions such as the mouth and forehead, which contributes to the overall gains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": "Qualitative Comparison. Given a single face image, Hi-Face reconstructs a high-fidelity 3D shape with details. We present comparisons with previous methods on 1). coarse shape reconstruction [21, 29, 74, 81, 73] in Fig. 4 and 2). detail reconstruction [77, 15, 24, 19, 70] in Fig. 5. See more examples and comparisons in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "For coarse shape in Fig. 4, our HiFace faithfully recovers the coarse shape of the given identity and outperforms the previous learning-based methods, and is on par" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 69, + 543, + 215 + ], + "blocks": [ + { + "bbox": [ + 304, + 69, + 543, + 215 + ], + "lines": [ + { + "bbox": [ + 304, + 69, + 543, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 69, + 543, + 215 + ], + "type": "image", + "image_path": "96bcd63fa56e720e31a542c754ffe9e6af72fa417bbd8fc37a20b84b1fe48101.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 264, + 544, + 500 + ], + "blocks": [ + { + "bbox": [ + 301, + 223, + 544, + 257 + ], + "lines": [ + { + "bbox": [ + 301, + 223, + 544, + 257 + ], + "spans": [ + { + "bbox": [ + 301, + 223, + 544, + 257 + ], + "type": "text", + "content": "Figure 5. Comparison on detail shape reconstruction. From left to right: Input image, FaceScape [77], Unsup [15], DECA [24], EMOCA [19], FaceVerse [70], and HiFace (Ours)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 264, + 544, + 500 + ], + "lines": [ + { + "bbox": [ + 307, + 264, + 544, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 544, + 500 + ], + "type": "image", + "image_path": "d198d2b1537a3210ca597276579b3798a63cec5f94e02a143f676daf8c9a3307.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 509, + 545, + 586 + ], + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 586 + ], + "type": "text", + "content": "Figure 7. Comparison on face animation. Given a source image (yellow box), we use the driving images (green box) to drive its expressions. DECA [24] (2nd-row) and EMOCA [19] (3rd-row) can animate the expression-driven details but lack realistic. As a comparison, HiFace is flexible to animate details from static (4th-row), dynamic (5th-row), or both (6th-row) factors, and presents vivid animation quality with realistic shapes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 602, + 545, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 698 + ], + "type": "text", + "content": "with Dense [73], which is the state-of-the-art optimization-based method. For detailed reconstruction in Fig. 5, previous methods [24, 19] fail to reconstruct satisfactory details. Several methods [15, 77, 70] are sensitive to occlusions and large poses. As a comparison, HiFace achieves the most realistic reconstruction quality, and faithfully recovers facial details of a given image, which significantly outperforms previous methods by a large margin." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In addition, given an image and the fitted coefficients " + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9093" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 174 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 174 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 174 + ], + "type": "image", + "image_path": "cbe86c2ac57a9cacabced6224a497ecd878935888a91d0304de9aa09358173c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 546, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 546, + 205 + ], + "type": "text", + "content": "Figure 8. Ablation studies on loss functions and training data. The coarse shape losses " + }, + { + "bbox": [ + 46, + 180, + 546, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shp}} / \\mathcal{L}_{\\mathrm{kl}}" + }, + { + "bbox": [ + 46, + 180, + 546, + 205 + ], + "type": "text", + "content": " (left), detail losses " + }, + { + "bbox": [ + 46, + 180, + 546, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{detail}} / \\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 180, + 546, + 205 + ], + "type": "text", + "content": " (right), and hybrid datasets coherently contribute to the reconstruction quality of coarse shapes and details." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 303, + 70, + 543, + 174 + ], + "blocks": [ + { + "bbox": [ + 303, + 70, + 543, + 174 + ], + "lines": [ + { + "bbox": [ + 303, + 70, + 543, + 174 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 543, + 174 + ], + "type": "image", + "image_path": "bdf077b5d3d7866c628b7cc9a327084fda34181b08ced33d24950c22d3d41570.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 213, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 213, + 287, + 274 + ], + "type": "text", + "content": " from the optimization-based methods such as Dense [73], SD-DeTail synthesizes the details and further strengthens the quality compared to the coarse shape (see Fig. 6). It shows SD-DeTail is flexible and can be easily plugged-and-play into other methods. See more in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 278, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 339 + ], + "type": "text", + "content": "Application of HiFace. The HiFace explicitly decouples the static and dynamic details through the proposed SD-DeTail. Therefore, we can animate the facial attributes by simply assigning the expression coefficient " + }, + { + "bbox": [ + 46, + 278, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 278, + 287, + 339 + ], + "type": "text", + "content": " and/or static coefficient " + }, + { + "bbox": [ + 46, + 278, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 278, + 287, + 339 + ], + "type": "text", + "content": " of the driving images to the source images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "content": "In Fig. 7, we demonstrate the animation quality of HiFace outperforms the previous state-of-the-art detail animation methods [24, 19]. It shows that while DECA [24] and EMOCA [19] can animate the expression-driven details but lack realistic, the proposed HiFace is flexible to manipulate the static, dynamic, or both details. When animating the static detail, the person-specific details can be well transferred into the source shape. When animating the dynamic detail, only expression-dependent details are presented. Finally, we can also animate the static and dynamic details simultaneously and achieve satisfactory results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 483, + 147, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 147, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 147, + 495 + ], + "type": "text", + "content": "5. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "text", + "content": "Ablation Studies on Loss Functions and Datasets. We present ablation studies on the proposed loss functions and training strategy with hybrid datasets. We train HiFace with synthetic data alone and compare it to the one trained with hybrid datasets. For coarse shape reconstruction, we investigate the contribution of " + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shp}}" + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "text", + "content": " and its sub-term " + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kl}}" + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "text", + "content": ". For detail reconstruction, we compare HiFace without " + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{detail}}" + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 509, + 287, + 605 + ], + "type": "text", + "content": ", respectively. The results are presented in Fig. 8." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Fig. 8 demonstrates that the proposed loss functions and training strategy from hybrid datasets contribute to satisfactory coarse shape and details. First, the model trained with synthetic data alone cannot generalize well to real-world images, which indicates the necessity to train with real-world data. Second, " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shp}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " improves the coarse shape reconstruction quality. " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shp}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " is effective in tackling challenging poses and improving alignment. " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kl}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " can relieve the overfitting risk on the synthetic data and improve the gener" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 209, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 306, + 209, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 306, + 209, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 545, + 315 + ], + "type": "image", + "image_path": "02fee0cf9532e8d092cb67544266cebe39309ee88183fadd57d475e656449220.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 323, + 546, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 323, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 546, + 399 + ], + "type": "text", + "content": "Figure 9. Ablation studies on SD-DeTail. Results show that directly synthesizing the static or dynamic details is rather challenging, leading to unreasonable coarse shapes and details (SD-1, SD-2, and SD-3). As a comparison, we leverage the statistical bases with SD-DeTail and regard the detail generation problem as a coefficients regression and interpolation problem, leading to more realistic details." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 408, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 456 + ], + "type": "text", + "content": "alization to real-world images. Third, without " + }, + { + "bbox": [ + 305, + 408, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{detail}}" + }, + { + "bbox": [ + 305, + 408, + 545, + 456 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 408, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 305, + 408, + 545, + 456 + ], + "type": "text", + "content": ", the reconstructed details exhibit random noise and cannot faithfully reflect person-specific details. Such noise misses the correspondence to the person-specific identity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "Ablation Studies on SD-DeTail. To verify the effectiveness of building bases for static and dynamic details, we present detailed ablation studies on SD-DeTail, by replacing the bases (i.e., " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{sta}}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{com}} / \\mathbf{B}_{\\mathrm{str}}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": ") reconstruction with a U-Net decoder [60] (same as DECA [24]). Therefore, the model learns to directly synthesize displacement maps instead of predicting corresponding coefficients like ours. In Fig. 9, we make comparisons on: 1). directly synthesizing " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{dyn}}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " (SD-1), 2). directly synthesizing " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " and interpolating via Eq. 9 (SD-2), 3). directly synthesizing " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{sta}}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " (SD-3), 4). SD-DeTail (Ours). It can be seen that, due to the high diversity and complexity of expression representation, it is hard to directly learn realistic details even with ground-truth labels of " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{dyn}}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " from synthetic data (see SD-1, SD-2 and SD-3 in Fig. 9). More specifically, for the static details, directly synthesizing displacement maps bring much noise (SD-3). For example, the hollow eyebrow is demonstrated in the second row. For the dynamic details, directly synthesizing displacement maps even leads to unnatural results (SD-1 and SD-2). For example, the reconstructed 3D faces are distorted especially in the second row. We also notice" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9094" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "that, directly synthesizing " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{dyn}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": " (SD-1) achieves inferior results than directly synthesizing " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{com}} / \\mathbf{D}_{\\mathrm{str}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": " and interpolating via Eq. 9 (SD-2). This demonstrates that it is beneficial to simplify the expression representation by using interpolation between two displacement maps (i.e., compressed and stretched). In conclusion, these observations further verify our insight on relaxing the challenging detail generation problem into a feasible coefficients regression problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 178, + 119, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 119, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 119, + 192 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 289, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 289, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 289, + 368 + ], + "type": "text", + "content": "We propose HiFace to reconstruct high-fidelity 3D faces with realistic and animatable details from a single image. Our motivation and insights stand on the spirit of 3DMMs to simplify the challenging detail generation into more accessible regression and interpolation tasks. To achieve this, we elaborately design SD-DeTail to decouple the static and dynamic factors explicitly, and interpolate the dynamic details through vertex tension. We succeed in learning the coarse shape and details jointly by proposing several novel loss functions to train on synthetic and real-world data. Extensive experiments demonstrate that HiFace achieves state-of-the-art face reconstruction both quantitatively and qualitatively in the coarse shape and detail shape, and the details are well decoupled and naturally animatable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 377, + 145, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 145, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 145, + 391 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 397, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 469 + ], + "type": "text", + "content": "This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SZSTC Grant (JCYJ20190809172201639, WDZC2020-0820200655001), Shenzhen Key Laboratory (ZDSYS2021-0623092001004), and Beijing Key Lab of Networked Multimedia." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 107, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 107, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 107, + 491 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 499, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 531 + ], + "type": "text", + "content": "[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 533, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 574 + ], + "type": "text", + "content": "[2] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition, June 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 575, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 287, + 628 + ], + "type": "text", + "content": "[3] Linchao Bao, Xiangkai Lin, Yajing Chen, Haoxian Zhang, Sheng Wang, Xuefei Zhe, Di Kang, Haozhi Huang, Xinwei Jiang, Jue Wang, et al. High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 628, + 288, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 288, + 681 + ], + "type": "text", + "content": "[4] Volker Blanz, Sami Romdhani, and Thomas Vetter. Face identification across different poses and illuminations with a 3d morphable model. In Proceedings of fifth IEEE international conference on automatic face gesture recognition, pages 202-207. IEEE, 2002. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "type": "text", + "content": "[5] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 716 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 317, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 545, + 95 + ], + "type": "text", + "content": "conference on Computer graphics and interactive techniques, pages 187-194, 1999. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "text", + "content": "[6] James Booth, Anastasios Roussos, Allan Ponniah, David Dunaway, and Stefanos Zafeiriou. Large scale 3d morphable models. International Journal of Computer Vision, 126(2):233-254, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 169 + ], + "type": "text", + "content": "[7] Sofien Bouaziz, Yangang Wang, and Mark Pauly. Online modeling for realtime facial animation. ACM Transactions on Graphics (ToG), 32(4):1-10, 2013. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 170, + 547, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 547, + 222 + ], + "type": "text", + "content": "[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proceedings of the IEEE International Conference on Computer Vision, pages 1021-1030, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 545, + 255 + ], + "type": "text", + "content": "[9] Chen Cao, Derek Bradley, Kun Zhou, and Thabo Beeler. Realtime high-fidelity facial performance capture. ACM Transactions on Graphics (ToG), 34(4):1-9, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 298 + ], + "type": "text", + "content": "[10] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 298, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 351 + ], + "type": "text", + "content": "[11] Kaidi Cao, Yu Rong, Cheng Li, Xiaou Tang, and Chen Change Loy. Pose-robust face recognition via deep residual equivariant mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5187-5196, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 352, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 352, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 545, + 405 + ], + "type": "text", + "content": "[12] Zenghao Chai, Haoxian Zhang, Jing Ren, Di Kang, Zhengzhuo Xu, Xuefei Zhe, Chun Yuan, and Linchao Bao. Really: Rethinking the evaluation of 3d face reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "text", + "content": "[13] Anpei Chen, Zhang Chen, Guli Zhang, Kenny Mitchell, and Jingyi Yu. Photo-realistic facial details synthesis from single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9429-9439, 2019. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[14] Dong Chen, Gang Hua, Fang Wen, and Jian Sun. Supervised transformer network for efficient face detection. In European Conference on Computer Vision, pages 122-138. Springer, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 491, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 534 + ], + "type": "text", + "content": "[15] Yajing Chen, Fanzi Wu, Zeyu Wang, Yibing Song, Yonggen Ling, and Linchao Bao. Self-supervised learning of detailed 3d face reconstruction. IEEE Transactions on Image Processing, 29:8696-8705, 2020. 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 534, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 587 + ], + "type": "text", + "content": "[16] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "type": "text", + "content": "[17] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 620, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 620, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 620, + 545, + 662 + ], + "type": "text", + "content": "[18] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE international conference on computer vision, pages 3085-3093, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 662, + 545, + 716 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 662, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 308, + 662, + 545, + 716 + ], + "type": "text", + "content": "[19] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9095" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 716 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 116, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 116, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 287, + 170 + ], + "type": "text", + "content": "[21] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 170, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 170, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 287, + 235 + ], + "type": "text", + "content": "[22] Abdallah Dib, Cédric Thébault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 235, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 289 + ], + "type": "text", + "content": "[23] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 289, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 331 + ], + "type": "text", + "content": "[24] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[25] Zhongpai Gao, Juyong Zhang, Yudong Guo, Chao Ma, Guangtao Zhai, and Xiaokang Yang. Semi-supervised 3d face representation learning from unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 348-349, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 396, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 450 + ], + "type": "text", + "content": "[26] Pablo Garrido, Michael Zollhöfer, Dan Casas, Levi Valgaerts, Kiran Varanasi, Patrick Pérez, and Christian Theobalt. Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG), 35(3):1-15, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 450, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 505 + ], + "type": "text", + "content": "[27] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "text", + "content": "[28] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[29] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "text", + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 643, + 287, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 287, + 697 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 287, + 697 + ], + "type": "text", + "content": "[31] Liwen Hu, Shunsuke Saito, Lingyu Wei, Koki Nagano, Jaewoo Seo, Jens Fursund, Iman Sadeghi, Carrie Sun, Yen-Chun Chen, and Hao Li. Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG), 36(6):1-14, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 697, + 287, + 716 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 697, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 48, + 697, + 287, + 716 + ], + "type": "text", + "content": "[32] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 322, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "text", + "content": "[33] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 191 + ], + "type": "text", + "content": "[34] Luo Jiang, Juyong Zhang, Bailin Deng, Hao Li, and Ligang Liu. 3d face reconstruction with geometry details from a single image. IEEE Transactions on Image Processing, 27(10):4756-4770, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 191, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 545, + 244 + ], + "type": "text", + "content": "[35] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. ACM New York, NY, USA, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 245, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 298 + ], + "type": "text", + "content": "[36] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548-1558, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[37] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "text", + "content": "[38] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 396, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 416 + ], + "type": "text", + "content": "[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "text", + "content": "[40] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 502 + ], + "type": "text", + "content": "[41] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware mesh decoder for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6100–6109, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "text", + "content": "[42] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2023. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[43] John P Lewis, Matt Cordner, and Nickson Fong. Pose space deformation: a unified approach to shape interpolation and skeleton-driven deformation. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 165-172, 2000. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 609, + 545, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 674 + ], + "type": "text", + "content": "[44] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 674, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 674, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 674, + 545, + 707 + ], + "type": "text", + "content": "[45] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9096" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 716 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[46] Yue Li, Liqian Ma, Haoqiang Fan, and Kenny Mitchell. Feature-preserving detailed 3d face reconstruction from a single image. In Proceedings of the 15th ACM SIGGRAPH European Conference on Visual Media Production, pages 1-9, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 126, + 287, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 126, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 126, + 287, + 181 + ], + "type": "text", + "content": "[47] Jiangke Lin, Yi Yuan, Tianjia Shao, and Kun Zhou. Towards high-fidelity 3d face reconstruction from in-the-wild images using graph convolutional networks. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 5891-5900, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 181, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 181, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 181, + 287, + 235 + ], + "type": "text", + "content": "[48] Jingwang Ling, Zhibo Wang, Ming Lu, Quan Wang, Chen Qian, and Feng Xu. Structure-aware editable morphable model for 3d facial detail animation and manipulation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 235, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 278 + ], + "type": "text", + "content": "[49] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 278, + 287, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 278, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 48, + 278, + 287, + 332 + ], + "type": "text", + "content": "[50] Huiwen Luo, Koki Nagano, Han-Wei Kung, Qingguo Xu, Zejian Wang, Lingyu Wei, Liwen Hu, and Hao Li. Normalized avatar synthesis using stylegan and perceptual refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11662–11672, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[51] Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiri Matas, and Viktoriia Sharmanska. Dad-3heads: A large-scale dense, accurate and diverse dataset for 3d head alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20942–20952, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "type": "text", + "content": "[52] Ali Mollahosseini, Behzad Hasani, and Mohammad H Ma-hoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2017. 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 440, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 471 + ], + "type": "text", + "content": "[53] Araceli Morales, Gemma Piella, and Federico M Sukno. Survey on 3d face reconstruction from uncalibrated images. Computer Science Review, 40:100400, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[54] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "text", + "content": "[55] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "text", + "content": "[56] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 654 + ], + "type": "text", + "content": "[57] Chirag Raman, Charlie Hewitt, Erroll Wood, and Tadas Baltrusaitis. Mesh-tension driven expression-based wrinkles for synthetic faces. In WACV, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 287, + 696 + ], + "type": "text", + "content": "[58] Roger Blanco I Ribera, Eduard Zell, John P Lewis, Junyong Noh, and Mario Botsch. Facial retargeting with automatic range of motion alignment. ACM Transactions on graphics (TOG), 36(4):1-12, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 696, + 287, + 716 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 696, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 48, + 696, + 287, + 716 + ], + "type": "text", + "content": "[59] Sami Romdhani, Volker Blanz, and Thomas Vetter. Face identification by fitting a 3d morphable model using linear" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 323, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 323, + 73, + 545, + 95 + ], + "type": "text", + "content": "shape and texture error functions. In European Conference on Computer Vision, pages 3-19. Springer, 2002. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[60] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 148, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 545, + 201 + ], + "type": "text", + "content": "[61] Zeyu Ruan, Changqing Zou, Longhai Wu, Gangshan Wu, and Limin Wang. Sadrnet: Self-aligned dual face regression networks for robust 3d dense face alignment and reconstruction. IEEE Transactions on Image Processing, 30:5793-5806, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 201, + 545, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 256 + ], + "type": "text", + "content": "[62] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 256, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 299 + ], + "type": "text", + "content": "[63] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 299, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 354 + ], + "type": "text", + "content": "[64] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 353, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 407 + ], + "type": "text", + "content": "[65] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 407, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 407, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 407, + 545, + 449 + ], + "type": "text", + "content": "[66] Supasorn Suwajanakorn, Ira Kemelmacher-Shlizerman, and Steven M Seitz. Total moving face reconstruction. In European conference on computer vision, pages 796-812. Springer, 2014. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 514 + ], + "type": "text", + "content": "[67] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "text", + "content": "[68] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 610 + ], + "type": "text", + "content": "[69] Daniel Vlasic, Matthew Brand, Hanspeter Pfister, and Jovan Popovic. Face transfer with multilinear models. In ACM SIGGRAPH 2006 Courses, pages 24–es. ACM New York, NY, USA, 2006. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 610, + 545, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 610, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 308, + 610, + 545, + 665 + ], + "type": "text", + "content": "[70] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 665, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 665, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 665, + 545, + 708 + ], + "type": "text", + "content": "[71] Yandong Wen, Weiyang Liu, Bhiksha Raj, and Rita Singh. Self-supervised 3d face reconstruction via conditional estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13289-13298, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9097" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 354 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[72] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 138, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 287, + 203 + ], + "type": "text", + "content": "[73] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 203, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 203, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 203, + 287, + 245 + ], + "type": "text", + "content": "[74] Cho-Ying Wu, Qiangeng Xu, and Ulrich Neumann. Synergy between 3dmm and 3d landmarks for accurate 3d facial geometry. In 2021 International Conference on 3D Vision (3DV), 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "type": "text", + "content": "[75] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 289, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 342 + ], + "type": "text", + "content": "[76] Shugo Yamaguchi, Shunsuke Saito, Koki Nagano, Yajie Zhao, Weikai Chen, Kyle Olszewski, Shigeo Morishima, and Hao Li. High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 354 + ], + "type": "text", + "content": "[77] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 354 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 317, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 545, + 117 + ], + "type": "text", + "content": "Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 117, + 545, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 117, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 117, + 545, + 159 + ], + "type": "text", + "content": "[78] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Towards large-posed face frontalization in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3990–3999, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 159, + 545, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 545, + 202 + ], + "type": "text", + "content": "[79] Qi Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Computer Vision and Pattern Recognition, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 202, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 202, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 202, + 545, + 255 + ], + "type": "text", + "content": "[80] Wenbin Zhu, HsiangTao Wu, Zeyu Chen, Noranart Vesdapunt, and Baoyuan Wang. Reda: reinforced differentiable attribute for 3d face reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4958-4967, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 255, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 299 + ], + "type": "text", + "content": "[81] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision (ECCV). Springer International Publishing, Oct. 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 299, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 354 + ], + "type": "text", + "content": "[82] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum. Wiley Online Library, 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9098" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_content_list.json b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8bc840af5d6295c43149a63dc5ff36e37ec9429f --- /dev/null +++ b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_content_list.json @@ -0,0 +1,2280 @@ +[ + { + "type": "text", + "text": "HiLo: Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation", + "text_level": 1, + "bbox": [ + 220, + 130, + 750, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijian Zhou", + "bbox": [ + 151, + 204, + 250, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Informatics", + "bbox": [ + 120, + 224, + 281, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "King's College London", + "bbox": [ + 130, + 241, + 272, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zijian.zhou@kcl.ac.uk", + "bbox": [ + 107, + 260, + 295, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Miaojing Shi", + "bbox": [ + 424, + 203, + 545, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "College of Electronic and Information Engineering", + "bbox": [ + 333, + 224, + 637, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tongji University", + "bbox": [ + 431, + 241, + 537, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mshi@tongji.edu.cn", + "bbox": [ + 403, + 260, + 565, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Holger Caesar", + "bbox": [ + 710, + 204, + 826, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intelligent Vehicles Lab", + "bbox": [ + 697, + 224, + 841, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Delft University of Technology", + "bbox": [ + 674, + 241, + 862, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "h.caesar@tudelft.nl", + "bbox": [ + 681, + 260, + 852, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 312, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panoptic Scene Graph generation (PSG) is a recently proposed task in image scene understanding that aims to segment the image and extract triplets of subjects, objects and their relations to build a scene graph. This task is particularly challenging for two reasons. First, it suffers from a long-tail problem in its relation categories, making naive biased methods more inclined to high-frequency relations. Existing unbiased methods tackle the long-tail problem by data/loss rebalancing to favor low-frequency relations. Second, a subject-object pair can have two or more semantically overlapping relations. While existing methods favor one over the other, our proposed HiLo framework lets different network branches specialize on low and high frequency relations, enforce their consistency and fuse the results. To the best of our knowledge we are the first to propose an explicitly unbiased PSG method. In extensive experiments we show that our HiLo framework achieves state-of-the-art results on the PSG task. We also apply our method to the Scene Graph Generation task that predicts boxes instead of masks and see improvements over all baseline methods. Code is available at https://github.com/franciszzj/HiLo.", + "bbox": [ + 75, + 340, + 473, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene Graph Generation (SGG) [49] is a crucial task in image scene understanding that extracts triplets in the form of subjects, objects and their relations to build a scene graph. Subjects and objects are represented with bounding boxes. Since this task links vision and text, it holds great potential for a variety of applications, including visual question answering [27], image captioning [21, 8], image retrieval [32, 53, 51] and visual reasoning [1, 55].", + "bbox": [ + 75, + 726, + 468, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently a novel variant of SGG was proposed, which is Panoptic Scene Graph generation (PSG) [62]. Subjects", + "bbox": [ + 75, + 847, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fadbdf098beab5eb95c9713497ea7eb46a1aa1e6e20906838d22af4f7395b5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 309, + 785, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26b4006d1727886b770c936e64746194da33e5e442cce1f9422bbff5967c3e67.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 504, + 412, + 629, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85fcc76f8260fec609d39d1efa6c7be0f9835d1a920f0c86a8b9f4748ae6a805.jpg", + "image_caption": [ + "(a)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 633, + 412, + 756, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47698a3e5574e48d55fb89438b4df0461f2fe71c2d98da3561b7b81fb835af6e.jpg", + "image_caption": [ + "(d)", + "Figure 1. An example of the PSG task. We compare the predicted object relations of different approaches. a) An image and its panoptic segmentation. Two persons and a sports ball are shown in different colors. b) A biased method predicts mostly high frequency relations. c) An unbiased method predicts mostly low frequency relations. d) Our method predicts both low and high frequency relations, as well as more relations in total." + ], + "image_footnote": [], + "bbox": [ + 763, + 412, + 888, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and objects are represented with panoptic segmentation [35] masks. Unless stated otherwise, in this work we focus on PSG, since it is pixel-level accurate and also covers background classes and their relations with foreground objects.", + "bbox": [ + 496, + 640, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The performance of the PSG model is affected by a long-tail problem in its relation categories. For instance, relations such as over, in front of and holding occur tens of thousands of times in the PSG dataset [62], while others like swinging and kissing occur only a few dozen times. This severe class imbalance in the relation categories can lead to model predictions that are more inclined to high-frequency relations, which poses significant challenges to the application of panoptic scene graphs in real-world scenarios.", + "bbox": [ + 496, + 702, + 892, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous methods [12, 18, 39, 17] have often treated the long-tail problem of the PSG task as equivalent to the long-tail problem in object-centric tasks such as classification [15, 5, 29] or semantic segmentation [14, 2]. Consequently,", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 99, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21637", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these methods have employed re-balancing techniques to address class imbalance, either through re-sampling the data [39] or by using a class-balanced loss [33] that assigns different weights to different relation categories.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, in the relation-centric PSG task, a subject-object pair can have multiple relations that exhibit relational semantic overlap, such as being partially or fully overlapping. For example, in Fig. 1, there are multiple relations between the boy and the sports ball, such as beside, looking at, playing and chasing. For regular biased models [61, 65, 60, 46], the results are dominated by high-frequency relations (beside, looking at). For specifically unbiased models [66, 62], the results are dominated by low-frequency relations (playing, chasing). However, since the low frequency relations can be more specific (e.g. on and standing on) or only partially overlapping with high frequency relations (e.g. looking at and chasing), it is crucial to include both to fully understand the image. We found that relational semantic overlap occurs in large numbers in the PSG dataset [62] and that current methods do not effectively address it. This is reflected in the increase in the category-averaged mean recall metric of unbiased methods, at the cost of the decrease in global recall (see Sec. 4.4).", + "bbox": [ + 75, + 152, + 470, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the long-tail problem of scene graphs under relational semantic overlap, we introduce the HiLo framework. This framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths with the help of two novel consistency loss functions. We apply our framework on top of a novel baseline. This baseline uses a recent transformer-based approach [10] for panoptic segmentation and adapts triplet queries [62] and masked attention [10] for the PSG task. In summary, we make the following contributions:", + "bbox": [ + 75, + 438, + 468, + 588 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the long-tail problem with relational semantic overlap in the PSG task and propose the HiLo framework to address this problem. The framework is general and can be applied to any PSG method.", + "- We propose a powerful and efficient one-stage end-to-end baseline. This baseline enhances the interaction between mask and relation prediction in the transformer decoder layer.", + "- We conduct extensive experiments to demonstrate the effectiveness of our framework and baseline. Our results outperform the state-of-the-art in both recall and mean recall on the PSG dataset and show systematic improvements on the VG dataset." + ], + "bbox": [ + 96, + 592, + 467, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 806, + 218, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Graph Generation", + "text_level": 1, + "bbox": [ + 76, + 832, + 303, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Scene Graph Generation (SGG) [49] task plays a crucial role in connecting vision and language, and has received widespread attention in the computer vision commu", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nity. Many methods have been proposed to improve the performance of SGG, which can be classified into three categories [70]. The first is to introduce multi-modal information, such as appearance [52], space [72], depth [54], and segmentation [34]. The second is to introduce prior information and commonsense knowledge, such as statistical [3, 16, 65, 9] and language prior knowledge [49, 43, 68, 31, 20]. The third category involves designing different model structures, such as message passing [41, 16, 42, 65, 24, 64], attention mechanisms [69, 50], tree structures and visual translation [67, 30]. However, most of these methods are two-stage methods that cannot learn scene graphs end-to-end. In contrast, to improve the learning ability of SGG models, several methods based on transformers have been proposed, including SGTR [38], RelationFormer [56] and RelTR [13]. These are end-to-end trainable in a single stage.", + "bbox": [ + 496, + 90, + 890, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unbiased Scene Graph Generation", + "text_level": 1, + "bbox": [ + 500, + 358, + 803, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Solving the long-tail problem in the SGG task has attracted considerable attention from researchers, and several unbiased methods [59, 63, 12, 18, 25, 39, 19, 22, 37, 40, 17, 66] have been proposed. These methods typically improve the model from the perspective of data re-sampling [39] or a class-balanced loss [33]. BGNN [39] uses a two-layer resampling strategy to provide a more balanced data distribution during training, while CogTree [63] proposal exploits the semantic relation between different predicate classes to design a novel CogTree loss. HML [17] improves the model's ability to solve long-tail problems by designing a staged training process, and IETrans [66] proposes an internal and external transfer method to transfer high frequency relations to low frequency relations and recover missing relations to train the unbiased model. Dong et al. [19] propose to group relations by their frequency and train specialized relation encoders for each group. While these methods are able to mitigate the long-tail problem, they do not address relational semantic overlap.", + "bbox": [ + 496, + 381, + 890, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Panoptic Scene Graph Generation", + "text_level": 1, + "bbox": [ + 500, + 680, + 799, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrary to SGG, the PSG task [62] uses panoptic segmentation masks instead of bounding boxes to represent objects, resulting in a more comprehensive scene graph. PSGTR [62], an end-to-end method based on the DETR structure, was proposed to construct a transformer-based PSG model. PSGFormer [62] further improved on PSGTR by introducing Object & Relation Query Learning Blocks and Query Matching Blocks. Their strong performance on most relation classes indicates that their method is implicitly unbiased. In contrast, our approach is the first to explicitly bias a PSG model, creating separate branches for low and high frequency relations, which are then fused together.", + "bbox": [ + 496, + 703, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the scene graph in PSG is built upon the sub", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21638", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d504548dc5565807a602bd6bcf021632dae67a9e444c9a18b35d5e033c982c02.jpg", + "image_caption": [ + "Figure 2. An overview of our HiLo framework with HiLo baseline. a) HiLo relation swapping module swaps the multiple relations in the subject-object pair to obtain H-L Data and L-H Data respectively. b) Input data into our HiLo framework with HiLo baseline model, there are two branches, namely H-L decoder and L-H decoder, which learn H-L Data and L-H Data respectively. c) In addition to task losses for PSG, we propose HiLo prediction alignment, which includes subject-object consistency loss and relation consistency loss, so that the parallel branch can be better optimized." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "jects, objects and their relations, strong panoptic segmentation [35] is crucial for PSG. Several DETR-based [6] methods such as Deformable-DETR [71], Segmenter [57], MaskFormer [11] and Mask2Former [10] have recently pushed the envelope in panoptic segmentation. These methods have introduced deformable transformer encoders and decoders, as well as pixel decoders and multi-scale information to improve model performance and speed up convergence. Our proposed baseline builds upon these techniques to further enhance the PSG performance and speed up model convergence.", + "bbox": [ + 75, + 505, + 472, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 685, + 168, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting", + "text_level": 1, + "bbox": [ + 76, + 710, + 238, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Panoptic Scene Graph generation (PSG) task aims to generate a panoptic scene graph $\\mathcal{G}$ for a given image $\\mathcal{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $\\mathcal{G}$ contains an object set $\\mathcal{O}$ and a relational triplet set $\\mathcal{T}$ , denoted by $\\mathcal{G} = (\\mathcal{O}, \\mathcal{T})$ . For the $i$ -th object $o_i$ in $\\mathcal{O} = \\{o_i\\}_{i=1}^N$ , we use $m_i$ to represent the object's mask and $c_i$ the object's category, i.e. $o_i = (m_i, c_i)$ . For the $j$ -th triplet in $\\mathcal{T} = \\{t_j\\}_{j=1}^M$ , we use $s_j$ to represent the subject, $o_j$ the object, and $r_j$ their relation, i.e. $t_j = (s_j, o_j, r_j)$ . There are in total $C$ object categories and $R$ relation categories. For the $k$ -th relation, we denote its frequency in the training set as $f_k$ .", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. HiLo Baseline", + "text_level": 1, + "bbox": [ + 500, + 503, + 643, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-quality panoptic segmentation is crucial for achieving good PSG performance. We build our method upon the latest advances in DETR-based panoptic segmentation, Mask2Former [10]. Below we present its structure, as well as our proposed modifications for the PSG task.", + "bbox": [ + 496, + 530, + 890, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask2Former. This method comprises three key parts: A backbone (CNN-based or transformer-based) followed by a pixel decoder, a transformer module with a transformer decoder and a task-specific module with different task heads. Specifically, the backbone takes an input image $\\mathcal{I}$ and generates an image feature $F$ . The pixel decoder then gradually upsamples $F$ to produce multi-scale features $\\widetilde{\\mathcal{F}} = \\{\\widetilde{F}_i\\}_{i=1}^4$ . The transformer decoder takes a set of queries $\\mathcal{Q}$ of size $N$ and multi-scale features $\\widetilde{\\mathcal{F}}$ as input and outputs a set of mask features $\\mathcal{X}$ of the same size with $\\mathcal{Q}$ .", + "bbox": [ + 496, + 608, + 892, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On top of the transformer decoder, there are two task heads including a linear classifier that predicts the class probability for each mask; and a multi-layer perceptron (MLP) that uses the mask features $\\mathcal{X}$ to generate the mask embedding $\\mathcal{E}$ . The mask prediction is obtained by taking the dot product of the mask embedding $\\mathcal{E}$ with the scale feature of the highest resolution in the multi-scale features $\\widetilde{\\mathcal{F}}$ .", + "bbox": [ + 496, + 762, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Triplet queries. The original query in Mask2Former is to predict the object. In order to predict both subject, object", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21639", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and their relation, we develop the triplet queries $\\mathcal{Q}^t$ inspired by [62] into our baseline. Each query predicts a triplet that includes subject, object, and relation. Accordingly, our task heads comprise three linear classifiers. Each classifier is responsible for predicting the class probability for subject, object or relation, respectively. Moreover, we devise two MLPs to generate mask embeddings for the subject and object, denoted by $\\mathcal{E}^s$ and $\\mathcal{E}^o$ . We use them to compute the dot product with the scale feature of the highest resolution in multi-scale feature $\\widetilde{\\mathcal{F}}$ and obtain the mask prediction for the subject and object, respectively.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked relation attention. [10] proposes a transformer decoder variant with masked attention, which extracts spatial features by adding the predicted mask of the object from the previous decoder layer's mask prediction. It makes the model focus on the object-related area in the feature map. To adapt this scheme to the PSG task, for each relation, we extend the masked attention to take the union of the binary masks of the subject and object as input, which represents the pixels corresponding to the relation.", + "bbox": [ + 75, + 257, + 472, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network training. We adopt the same losses as PS-GTR [62], including cross-entropy loss $\\mathcal{L}_{so\\_cls}$ for object classification of subject-object pairs and a combination $\\mathcal{L}_{so\\_mask}$ of focal loss [44] and dice loss [58] to jointly supervise mask learning. To supervise the relations, we use the cross-entropy loss $\\mathcal{L}_{rel\\_cls}$ . The baseline loss with $(\\lambda_1 = 1, \\lambda_2 = 1, \\lambda_3 = 4)$ is thus:", + "bbox": [ + 76, + 393, + 472, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b a s e l i n e}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {s o . c l s}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {s o . m a s k}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {r e l . c l s}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 501, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. HiLo Framework", + "text_level": 1, + "bbox": [ + 76, + 547, + 250, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key insight of our HiLo is to build a model that can take into account both high frequency and low frequency relations, and effectively improve the performance on low frequency relations without degrading the performance on high frequency relations.", + "bbox": [ + 75, + 571, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 HiLo relation generation", + "text_level": 1, + "bbox": [ + 76, + 665, + 303, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the PSG task, multiple relations can be used to describe the connection between a subject and an object from different perspectives, such as spatial relations, actions and prepositions. Since the subject, object and their relative position in the image are fixed, all these relations share the same visual information. This is reflected in the PSG dataset [62] where many subject-object pairs are annotated with multiple relation labels. These relation labels are of different frequencies in the dataset. In this section we introduce the HiLo relation generation module to prepare two sets of training data, biasing towards high and low-frequency relations respectively.", + "bbox": [ + 75, + 688, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relation augmentation. Similar to [22], we observe that there are many missing relation annotations in the PSG", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dataset [62]. To add the missing relation annotations, we design a relation augmentation scheme that is inspired by IETrans [66], which converts high-frequency relations to low-frequency relations and adds a relation to a subject-object pair that has no relation. We adapt it by first training our baseline as a biased model using the original annotated data. For every subject-object pair in the training set, we use this model to predict the relation scores for all predefined relation categories:", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- If this subject-object pair has annotated relation labels, we pick the one with the highest predicted score and use the score as a threshold. For other relations (except for already labeled ones) whose predicted scores are greater than the threshold, we add them as the relation labels of this subject-object pair.", + "- If this subject-object pair has no annotated relation labels, we use the predicted score of the no-relation class as a threshold. For other relations whose predicted scores are greater than this threshold we add them as the relation labels of this subject-object pair." + ], + "bbox": [ + 517, + 228, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This operation allows us to significantly augment the relation labels for the subject-object pairs in the training data, which is especially relevant for pairs with zero or no annotated relations.", + "bbox": [ + 496, + 396, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relation swapping. We swap the relation labels for each subject-object pair. Specifically, given a subject-object pair $(s,o)$ with $K$ relation labels, we have $K$ triplets $(s,o,r_1),\\ldots ,(s,o,r_K)$ , sorted by their relation frequency in descending order, $f_{1} > \\ldots >f_{K}$ . We denote by H-L and L-H the swapping of high-frequency relations with low-frequency relations and vice versa. This creates two sets of data:", + "bbox": [ + 496, + 458, + 890, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "H-L Data. Given a triplet $(s, o, r_k)$ , we replace its relation label $r_k$ with that of the next triplet $r_{k+1}$ with lower frequency. We sequentially process all triplets from $(s, o, r_1)$ until $(s, o, r_{K-1})$ , keeping the last triplet unchanged.", + "bbox": [ + 496, + 582, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "L-H Data. Given a triplet $(s,o,r_k)$ , we replace its relation label $r_k$ with that of the previous triplet $r_{k - 1}$ with higher frequency. We sequentially process all triplets from $(s,o,r_K)$ until $(s,o,r_2)$ , keeping the first triplet unchanged. Hence, we obtain two new sets of triplets, denoted by $\\mathcal{T}^{\\mathrm{H - L}}$ and $\\mathcal{T}^{\\mathrm{L - H}}$ . We devise two parallel decoders from the shared encoder of our backbone. They are learned with the H-L and L-H data, respectively. The H-L and L-H decoders favour the predictions for low and high-frequency relations, respectively. Despite their difference, for the corresponding triplet query in the two decoders, their predictions are highly correlated: on one hand, the subject and object predictions should be the same; on the other hand, the distribution of relation predictions should be overlapping. Below we first build the query correspondence in the two decoders and then introduce the HiLo subject-object and relation consistency loss to align the predictions from two the decoders.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21640", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 HiLo prediction alignment", + "text_level": 1, + "bbox": [ + 76, + 90, + 316, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training two different relation distributions simultaneously confuse the model. To make the model differentiate between the two branches, we built a HiLo prediction alignment module, including triplet query correspondence, subject-object consistency loss and relation consistency loss respectively.", + "bbox": [ + 75, + 114, + 467, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Triplet query correspondence. In order to construct the subject-object consistency loss and relation consistency loss, we first need to construct the correspondence between the triplet queries in the H-L and L-H decoders. The same query index in the two decoders is not naturally matched. In order to find the query correspondence, we need to rely on the ground truth assignment: we use Hungarian matching to assign the triplet label to the corresponding triplet query, and record the relation label index corresponding to the triplet query. This label index allows us to construct the correspondence between the triplet queries in both decoders. We calculate the consistency loss for the triplet query prediction with the same relation label index in the two decoders.", + "bbox": [ + 75, + 205, + 467, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subject-object consistency loss. Having the corresponding predictions from the two decoders, both their subjects and objects have the same ground truth and should be equal. We propose a subject-object consistency loss $\\mathcal{L}_{obj}$ to minimize the mean squared error (MSE) of the corresponding predictions from the two decoders.", + "bbox": [ + 75, + 416, + 467, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o b j} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {m a s k} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 518, + 467, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = \\left\\| \\operatorname {s o f t m a x} \\left(p _ {c} ^ {\\mathrm {H} - \\mathrm {L}}\\right) - \\operatorname {s o f t m a x} \\left(p _ {c} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\right\\| ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 537, + 467, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\left\\| \\operatorname {s i g m o i d} \\left(p _ {m} ^ {\\mathrm {H} - \\mathrm {L}}\\right) - \\operatorname {s i g m o i d} \\left(p _ {m} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\right\\| ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 556, + 467, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\mathcal{L}_{cls}$ and $\\mathcal{L}_{mask}$ represent the MSE losses for class prediction and mask prediction, respectively.", + "bbox": [ + 75, + 583, + 467, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$p_c^{\\mathrm{H - L}}$ and $p_c^{\\mathrm{L - H}}$ are the class prediction logits from the H-L and L-H decoders. After a softmax, they are $C$ -dimensional probability vectors. After a sigmoid, $p_m^{H - L}$ and $p_m^{L - H}$ are the mask prediction logits for the H-L and L-H decoders.", + "bbox": [ + 75, + 613, + 467, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relation consistency loss. Given a pair of subject and object, we have previously swapped the high-low frequency relation labels to create data for the H-L and L-H decoders. For H-L, the prediction of the low-frequency relation logit is of high value; while for L-H, the prediction of the high-frequency relation logit is of high value. For the predictions on the rest logits, they should be similar, since it is the same pair of subject and object for the two decoders. Based on this observation, we introduce the relation consistency loss.", + "bbox": [ + 75, + 674, + 467, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for a pair of subject and object, we use $p_r^{\\mathrm{H - L}}$ to denote the predicted relation logits from the H-L decoder, $p_r^{\\mathrm{L - H}}$ the predicted relation logits from the L-H decoder. They are $R$ -dimensional probability vectors after softmax. We can map between the distributions of the two vectors by swapping the logit values between the high- and", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "low-frequency relation indices, which is named Relational Index Exchange operation. We use $\\mathrm{RIE}(\\cdot)$ to denote this operation. $\\mathrm{RIE}(\\cdot)$ includes a stop gradient operation, which creates copies from original predictions of two branches, enabling value exchanges. For example, for relations $r_k$ and $r_{k + 1}$ , $\\mathrm{RIE}(\\cdot)$ exchanges the values between $p_{r,k}$ and $p_{r,k + 1}$ in $p_r$ . We can therefore compute the distance between $p_r^{\\mathrm{H - L}}$ and its mapped counterpart from $p_r^{\\mathrm{L - H}}$ , vice versa:", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {D i s t} _ {\\mathrm {H i L o}} = \\left\\| \\text {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {H - L}}\\right) - R I E \\left(\\text {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {L - H}}\\right)\\right) \\right\\| ^ {2} \\\\ + \\| R I E \\left(\\operatorname {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {H} - \\mathrm {L}}\\right)\\right) - \\operatorname {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\| ^ {2} \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 218, + 890, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The relation consistency loss is defined to minimize $\\mathrm{Dist}_{\\mathrm{HiLo}}$ with a margin of $m$ :", + "bbox": [ + 496, + 280, + 890, + 310 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e l} = \\max \\left(\\operatorname {D i s t} _ {\\mathrm {H i L o}} - m, 0\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 314, + 888, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m$ is a small constant. Adding $m$ is due to the fact that the high- and low frequency relations might be only partially semantically overlapping.", + "bbox": [ + 496, + 338, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network training. Our subject-object consistency loss and relation consistency loss can seamlessly integrate with the losses of any baseline method to jointly supervise the training of the entire model. Notably, we supervise the output of each transformer decoder layer to ensure effective learning. The final loss $\\mathcal{L}$ is thus:", + "bbox": [ + 496, + 383, + 890, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e l i n e}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {r e l}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 484, + 888, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 HiLo inference fusion", + "text_level": 1, + "bbox": [ + 500, + 527, + 704, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The H-L and L-H decoders favour low-frequency relation prediction and high-frequency relation prediction, respectively. To combine the strength of both during inference, we introduce the HiLo inference fusion module. Specifically, we denote by $\\mathcal{G}^{\\mathrm{H - L}}$ and $\\mathcal{G}^{\\mathrm{L - H}}$ the predicted panoptic scene graphs from H-L and L-H decoders, respectively. There are $N_{1}$ triplets in $\\mathcal{G}^{\\mathrm{H - L}}$ and $N_{2}$ triplets in $\\mathcal{G}^{\\mathrm{L - H}}$ .", + "bbox": [ + 496, + 550, + 890, + 655 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First, we merge the triplets in $\\mathcal{G}^{\\mathrm{H - L}}$ and $\\mathcal{G}^{\\mathrm{L - H}}$ and sort them in the descending order according to their relation scores. We obtain a list of $N_{1} + N_{2}$ triplets.", + "- Second, starting from the first triplet, we de-duplicate the triplet list. For the $i$ -th triplet $\\mathcal{T}_i$ , we identify its duplicated versions from the $(i + 1)$ -th triplet until the end of the list. Remove any follow-up triplet from the list if it 1) has the same subject, object and relation classes to that in $T_i$ and 2) has a mask IoU greater than a threshold, e.g. 0.5, between the predicted subject/object and the corresponding subject/object in $T_i$ .", + "- Third, after dedduplication, for each triplet in the list, we multiply the relation, subject and object scores as an overall score for it. We sort the triplet list according to this score in descending order to obtain the final panoptic scene graph $\\mathcal{G}^{\\mathrm{HiLo}}$ ." + ], + "bbox": [ + 517, + 655, + 890, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21641", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f3d1795f460d33158742d3dbcb18eae607688c65c04929121ff5b239b6feabdb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneScene Graph Detection
R@20mR@20R@50mR@50R@100mR@100
IMP [61]R5016.56.518.27.118.67.2
MOTIF [65]R5020.09.121.79.622.09.7
VCTree [60]R5020.69.722.110.222.510.2
GPSNet [46]R5017.87.019.67.520.17.7
PSGTR [62]R5028.416.634.420.836.322.1
PSGFormer [62]R5018.014.819.617.020.117.6
HiLo (ours)R5034.123.740.730.343.033.1
HiLo (ours)Swin-B38.528.346.235.349.639.1
HiLo (ours)Swin-L40.629.748.737.651.440.9
", + "bbox": [ + 161, + 73, + 803, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison between our HiLo and other methods on the PSG dataset. Our method shows superior performance compared to all previous methods.", + "bbox": [ + 75, + 244, + 893, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 290, + 209, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 320, + 181, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Panoptic Scene Graph Generation (PSG) [62]. This is the first Panoptic Scene Graph generation dataset. It has a total of 48,749 labeled images including 2,177 test images and 46,572 training images. The object categories comprise 80 thing classes and 53 stuff classes, which is the same as the COCO [45] and COCO-Stuff datasets [4]. The relation categories comprise 56 classes, including positional relations, common object-object relations, common actions, human actions, actions in the traffic scene, actions in the sports scene and interactions between backgrounds [62].", + "bbox": [ + 75, + 349, + 470, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Genome (VG) [36]. VG is a widely used benchmark dataset for Scene Graph Generation. Following previous work [65, 7], we adopt the widely accepted split, VG-150, which contains 150 object categories and 50 relation categories. The object categories cover a wide range of classes, such as animals, vehicles and household items. The relation categories include both spatial and semantic classes, such as on, in and wearing.", + "bbox": [ + 75, + 506, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Tasks and Metrics", + "text_level": 1, + "bbox": [ + 76, + 654, + 253, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Three subtasks have been proposed for the SGG and PSG tasks, which are Predicate Classification, Scene Graph Classification and Scene Graph Detection [61]. We focus on Scene Graph Detection for both datasets, since it is the most comprehensive and addressed by [62]. This subtask requires the model to first localize the objects and then predict the object classes and relations. Note that Scene Graph Detection in PSG includes the detection on stuff classes, while in SGG it does not.", + "bbox": [ + 75, + 683, + 468, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous work [60, 63, 62], we use Recall@K (R@K) and mean Recall@K (mR@K) as our metrics, where the former metric is dominated by high-frequency relations, while the latter assigns equal weight to all relation classes.", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 290, + 714, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, we follow the training strategy of PSGTR [62]. We use the AdamW optimizer [48], with a learning rate of $1e^{-4}$ and weight decay of $1e^{-4}$ , except for the backbone, which is trained with a learning rate of $1e^{-5}$ . For initialization, we use Mask2Former [10] pretrained on COCO [45] to initialize our backbone and pixel decoder. Following Mask2Former [10], we use 100 triplet queries for the H-L and L-H decoders respectively. Additionally, both the H-L and L-H decoders are initialized with Mask2Former's transformer decoder. To ensure consistent comparison with PSGTR, we adopted the same data augmentation settings. Our model is trained for 12 epochs with a step scheduler at epoch 10, taking approximately 18 hours to train on four A100 GPUs with a batch size of 1 for each GPU.", + "bbox": [ + 496, + 316, + 893, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison to the state-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 558, + 799, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PSG. Tab. 1 reports the performance of our method compared to the state-of-the-art on the PSG dataset [62]. We separate the methods into two groups. The first are two-stage methods consisting of a separate segmentor and relation predictor, which are modified for the PSG task in [62]. The second are one-stage end-to-end methods, which are able to simultaneously predict panoptic segmentation and relations. Our method belongs to the second category. For a fair comparison between the different methods, we use the same Resnet-50 [26] backbone. Our method shows superior performance compared to all previous methods. Particularly, it outperforms the previous best-performing method PSGTR [62] by a large margin, i.e. $+6\\%$ in R@100 and $+11\\%$ in mR@100. Our model is able to converge within only 12 epochs of training, whereas PSGTR [62] is trained for 60 epochs. We also evaluate our method using pre-trained transformer-based backbones, i.e. Swin-B and Swin-L [47], with the latter being a bigger model. Our results are consistently improved over all metrics due the powerful feature representation ability of the transformer. We also conducted a visual comparison, as shown in Fig. 3.", + "bbox": [ + 496, + 583, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21642", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6b2be7f5804958c7060c5b4f8dae5465d8189635924127a4ff1d6a6f9cd47ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScene Graph Detection
R/mR@50R/mR@100
MOTIF [65]31.0 / 6.735.1 / 7.7
+IETrans [66]26.4 / 12.430.6 / 14.9
+HiLo (ours)26.2 / 14.730.3 / 17.7
VCTree [60]30.2 / 6.734.6 / 8.0
+IETrans [66]25.4 / 11.529.3 / 14.0
+HiLo (ours)27.1 / 12.929.8 / 15.2
Transformer [59]30.0 / 7.434.3 / 8.8
+IETrans [66]25.5 / 12.529.6 / 15.0
+HiLo (ours)25.4 / 14.629.2 / 17.6
GPSNet [46]30.3 / 5.935.0 / 7.1
+IETrans [66]25.9 / 14.628.1 / 16.5
+HiLo (ours)25.6 / 15.827.9 / 18.0
", + "bbox": [ + 122, + 88, + 421, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SGG. Here we study whether our approach that was developed for the PSG task can also be applied to the SGG task. In Tab. 2, we conduct experiments on the VG-150 dataset. Following IETrans [66], we extend our unbiased HiLo framework (Sec. 3.3) without our PSG-specific baseline (Sec. 3.2) to four state-of-the-art biased SGG methods, namely MOTIF [65], VCTree [60], Transformer [59] and GPSNet [46], using the implementation of [66]. Compared to the unbiased IETrans [66] method, our method improves mean recall without sacrificing recall, demonstrating its effectiveness in enhancing the performance of low-frequency relations. Compared to the biased baselines [65, 60, 59, 46], our method achieves a significantly larger mean recall, while still maintaining an acceptable recall. This indicates that our method can improve the performance of low-frequency relations while also taking into account the performance of high-frequency relations. It also shows that the HiLo framework is a general technique that yields systematic improvements in both the PSG and SGG tasks.", + "bbox": [ + 75, + 337, + 470, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 635, + 240, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consistent with the paper, we use HiLo with a Resnet-50 backbone to perform ablation experiments on the PSG dataset.", + "bbox": [ + 75, + 657, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HiLo framework for different baselines. In this section we investigate whether our HiLo framework (Sec. 3.3) yields improvements for other baselines, rather than just the one presented in Sec. 3.2. Tab. 3 shows the results for two baselines, with and without the HiLo framework. We observe that our biased baseline outperforms the previous PSGTR [62] method on all metrics. Furthermore, by applying the HiLo framework, we can substantially improve the performance over both baselines. It is worth mentioning that the HiLo framework improves recall and mean recall simultaneously, whereas other methods typically improve one metric at the cost of the other [59, 63].", + "bbox": [ + 75, + 704, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HiLo relation augmentation. We observe that out of", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f46ecceb452ce1ba47c56fa5a1c05428dd6d8b1dfb206b7d03677afdb9548111.jpg", + "table_caption": [ + "Table 2. Comparison between our HiLo framework and other methods on the VG-150 dataset. Similar to [66], we apply IETrans and our own method on top of four leading baselines." + ], + "table_footnote": [], + "table_body": "
BaselineHiLoR/mR@20R/mR@50R/mR@100
HiLo baseline✓34.1 / 23.740.7 / 30.343.0 / 33.1
HiLo baseline-32.6 / 20.938.0 / 27.438.9 / 28.4
PSGTR [62]✓30.1 / 20.236.6 / 23.938.3 / 24.5
PSGTR [62]-28.4 / 16.634.4 / 20.836.3 / 22.1
", + "bbox": [ + 504, + 88, + 887, + 157 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f4a9b1ab653b5c624c1f8333e7c51bd32379c9b1dd7e112fbd6f464b4a684fa.jpg", + "table_caption": [ + "Table 3. Comparison of different baselines, with and without HiLo framework. Using the HiLo framework, we see significant improvements on both metrics." + ], + "table_footnote": [], + "table_body": "
Relation Aug.Multiple relationsR/mR@50R/mR@100
✓40%40.7 / 30.343.0 / 33.1
-10%40.1 / 28.142.8 / 32.5
", + "bbox": [ + 504, + 210, + 887, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study for HiLo relation augmentation. Relation augmentation affects the ratio of subject-object pairs with multiple relations. The larger this ratio, the more relations can be swapped, which leads to better results.", + "bbox": [ + 498, + 262, + 890, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "260,296 labeled triplets in the PSG dataset, only about $10\\%$ of subject-object pairs have multiple relations, for which we can apply relation swapping (Sec. 3.3.1). After applying our proposed relation augmentation technique (Sec. 3.3.1), this ratio significantly increases to $40\\%$ . Our experimental results in Tab. 4 demonstrate that only applying the HiLo framework on $10\\%$ already gives an improvement over the baseline from Tab. 3. As the number of swappable triplets increases due to augmentation, the model's performance is further enhanced, highlighting the potential of our method.", + "bbox": [ + 496, + 325, + 890, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HiLo prediction alignment. We conduct ablation experiments on the subject-object consistency loss and relation consistency loss (Sec. 3.3.2), which are used to align the predictions from the high and low frequency branches. The results, as presented in Tab. 5, demonstrate that using both losses yields the best performance. It is worth mentioning that we have explored the margin in the relation consistency loss and found that setting the margin to zero leads to a small performance degradation. This finding confirms that there is a partial semantic overlap between swapped relations, indicating that they are not entirely consistent.", + "bbox": [ + 496, + 478, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the impact of relational index exchange (RIE) on the relation consistency loss, we conducted experiments to verify the effect of omitting RIE. In the absence of RIE, we solely compute the consistency loss for relation categories that are not involved in relation swapping and exclude the swapping component from the calculation of the consistency loss. The outcomes of this experiment are presented in Tab. 6, and demonstrate a notable reduction in the mean recall and a decline in the model's performance for relations with relational semantic overlap when RIE is not utilized.", + "bbox": [ + 496, + 643, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HiLo inference fusion. We ablate the inference fusion (Sec. 3.3.3) and evaluate the performance of each branch's output separately. In Tab. 7, experimental results suggest that fusion can effectively leverage the uniqueness of the high and low frequency branch predictions to achieve comprehensive improvements on all metrics.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21643", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fa600f28fd3137b58e0fa7f7db26250b9afb8bf2fcb8f7b65c7781e73616aab.jpg", + "image_caption": [ + "↑ Image" + ], + "image_footnote": [], + "bbox": [ + 84, + 94, + 210, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a9f8d2eb981294c28102f3ab364f95e94d73dc3233d2f7fd5d30f65b93a9a82.jpg", + "image_caption": [ + "Ground Truth", + "person walking on sea", + "surfboard over sea", + "sky over surf", + "person attached to surfboard", + "surfboard attached to person", + "surfboard on person" + ], + "image_footnote": [], + "bbox": [ + 86, + 209, + 210, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3646b8e1b5dbaaeeb361e9b91848002f2fbdce0140cfd6ca683a2f080d46bb7.jpg", + "image_caption": [ + "person carrying surfboard", + "looking at", + "sky over person", + "sky over sea" + ], + "image_footnote": [], + "bbox": [ + 120, + 287, + 140, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c603cee51a1a8b18c820cde2034ae09fa7a704cb872df026ca4d63d2ac96985e.jpg", + "image_caption": [ + "Prediction $\\rightarrow$", + "person holding surfboard" + ], + "image_footnote": [], + "bbox": [ + 215, + 95, + 341, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e1a1c82d1b12bb4fcd21ef016055f53c023d0542011c0dee25c22fa2fcba4f8.jpg", + "image_caption": [ + "sky over sea", + "person standing on sea", + "#", + "sky over person" + ], + "image_footnote": [], + "bbox": [ + 246, + 189, + 269, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7f1aa76cca772df169569b9f3788fe4095be02c539bafa96ca2e7cbbcd0e0ba.jpg", + "image_caption": [ + "person on sea", + "person carrying surfboard", + "looking at you" + ], + "image_footnote": [], + "bbox": [ + 246, + 247, + 269, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afd8823e7c5b5591c040c7eb8afaf2e9d9763e319fa5a06197d26292e32f5efe.jpg", + "image_caption": [ + "sea in front of person", + "sea beside person", + "person playing surfboard", + "surfboard on sea" + ], + "image_footnote": [], + "bbox": [ + 223, + 303, + 269, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a04fcf96662337ab4ff31d31471d73864e5cfd71636071d51a96174372a56e2b.jpg", + "image_caption": [ + "↑", + "Image", + "Prediction→" + ], + "image_footnote": [], + "bbox": [ + 357, + 94, + 482, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce68394d186232dd3e2c1e3f79ce71c568dc0aa633e68981525fff37f793a0ce.jpg", + "image_caption": [ + "Ground Truth", + "1", + "sky over snow", + "sky over person", + "person standing on snow", + "sky over skis", + "sky over backpack", + "sky over rock", + "person climbing snow", + "person walking on snow", + "personbesideperson", + "person on snow", + "person running on snow", + "person picking skis", + "person playing skills", + "person playing with person attached to skis", + "person attached to backpack", + "mountain attached", + "sky over mountain", + "#" + ], + "image_footnote": [], + "bbox": [ + 357, + 215, + 483, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27e21a5518eb23cf7104d3a1cfe7470c4299507d893ece5625e4e1a14f376a3b.jpg", + "image_caption": [ + "person carrying backpack", + "person holding skis", + "person climbing snow.", + "sky over snow." + ], + "image_footnote": [], + "bbox": [ + 367, + 287, + 431, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b72dddcb15b5e69afbb1e14f07a0b77555faa439c1b6be332ba58fc38efda690.jpg", + "image_caption": [ + "person carrying backpack", + "person holding", + "skis on :" + ], + "image_footnote": [], + "bbox": [ + 488, + 94, + 612, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ba5ceb2c1ff682fe10946d74229fb33427e75719f3bb82142b829b78e8bf237.jpg", + "image_caption": [ + "↑", + "Image", + "Prediction $\\rightarrow$" + ], + "image_footnote": [], + "bbox": [ + 630, + 95, + 754, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbcb12245522c272c26d46ee07348c4652472801a13605622700bc61e93907c1.jpg", + "image_caption": [ + "Ground Truth", + "1", + "chair beside wall", + "chair in front of", + "person in front of chair", + "person looking at", + "person throwing", + "wall in front of tree", + "over chair", + "person swinging tennis racket", + "treebesidewall", + "chair beside person", + "chair attached to wall", + "person holding sports ball" + ], + "image_footnote": [], + "bbox": [ + 630, + 217, + 754, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84a9ce5306300de517c32d31386355b13b59521e84e6b23a884c79dc5296d51e.jpg", + "image_caption": [ + "mehohid", + "person beside barrier", + "wall beside tree", + "person playing sports ball" + ], + "image_footnote": [], + "bbox": [ + 635, + 285, + 754, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09ae112509d1f1e09e86d6acf0847d76b3b96580a4c7206e49724a8b5642d9ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 292, + 743, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c3d0874c90c90775a6ba793c8315b1e8f848e82cf2f9303f20d2b161763aaa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 300, + 694, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da702fde1726af0638b1d5d7dff8d28d1df8af4307319e948810c447889d64f2.jpg", + "image_caption": [ + "tree in front of" + ], + "image_footnote": [], + "bbox": [ + 635, + 306, + 743, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd226a19c1fb0ca712398d39c7fb7114f94d2d1d042de71a44bda5b97bde5f41.jpg", + "image_caption": [ + "chair leaning on wall", + "tree over wall", + "Figure 3. Visualization of panoptic segmentations and the top 20 predicted triplets compared with ground truth. The upper left is the original image, the lower left is the ground truth and on the right are the predictions. The highlighted triplets represent the subject-object pairs with multiple relations, where the blue highlights represent the high frequency relations and the red highlights represent the low frequency relations. The visualization results show that our method can predict both high frequency and low frequency relations." + ], + "image_footnote": [], + "bbox": [ + 635, + 315, + 743, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/869b41a2d909c1c603bae50fa873afa3ce9122612b09b7cafee5211987245d8c.jpg", + "image_caption": [ + "holding", + "about to", + "person in front of wall" + ], + "image_footnote": [], + "bbox": [ + 759, + 95, + 883, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0b2033ee9843e9403dcaddf9f9b0517c08abc8824e0f4ffc3e749950400b1fac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ObjectRelationMarginR/mR@50R/mR@100
✓✓0.540.7 / 30.343.0 / 33.1
-✓0.540.6 / 29.742.8 / 32.8
-✓0.040.5 / 29.542.7 / 32.8
✓--40.4 / 29.042.8 / 32.2
---39.7 / 28.642.4 / 32.0
", + "bbox": [ + 81, + 398, + 465, + 488 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/54d3f5aa593018d6d151c53be9d9465cf28c65b1d8b9be8789cba17f2af78f52.jpg", + "table_caption": [ + "Table 5. Ablation study for different losses in HiLo prediction alignment. Object refers to the subject-object consistency loss and relation refers to the relation consistency loss. The margin parameter is defined in Eq. 6." + ], + "table_footnote": [], + "table_body": "
Whether to use RIER/mR@20R/mR@50R/mR@100
✓34.1 / 23.740.7 / 30.343.0 / 33.1
-33.5 / 22.340.3 / 29.042.6 / 32.3
", + "bbox": [ + 81, + 555, + 465, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also attempted to average the tensor generated by the two branches and obtain the PSG result through postprocessing. However, we found that this approach leads to a substantial drop in performance, as evident in Tab. 8. This can be attributed to the inconsistent prediction results of the two branches for the same query index. These findings validate that the inference fusion method effectively merges the results from the two branches. Furthermore, our experimental results demonstrate that the query associated with the identical index in two branches does not predict the same subject-object pair. Thus, directly averaging the tensor produced by the two branches results in prediction ambiguity, ultimately leading to a substantial decline in performance. This observation underscores the necessity of conducting triplet query correspondence when performing prediction alignment. In particular, due to the inconsistent query prediction content for the corresponding index", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61bc5568454d8f476932592fb4cd781cd4f9b43fba279802867a9f3ae3be64bc.jpg", + "table_caption": [ + "Table 6. Ablation study for relation consistency loss in HiLo prediction alignment." + ], + "table_footnote": [], + "table_body": "
H-L ResultL-H ResultR/mR@50R/mR@100
✓✓40.7 / 30.343.0 / 33.1
✓-38.8 / 29.939.8 / 30.9
-✓38.5 / 26.539.5 / 27.8
", + "bbox": [ + 504, + 398, + 887, + 463 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c9c4e69dcc77114d4e305e0a0db0030f2b57ccc8e6c013798bdd472a9a433375.jpg", + "table_caption": [ + "Table 7. Ablation study for HiLo inference fusion." + ], + "table_footnote": [], + "table_body": "
Fusion methodR/mR@20R/mR@50R/mR@100
inference fusion34.1 / 23.740.7 / 30.343.0 / 33.1
average tensor19.6 / 13.123.1 / 15.723.9 / 16.3
", + "bbox": [ + 504, + 500, + 887, + 546 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c941d116aed8531b044c6bb9595facaba62e23c26f1c3c0b5cff7cee68fde276.jpg", + "table_caption": [ + "Table 8. Ablation study for HiLo inference fusion. Inference fusion refers to the method proposed in the paper to fuse the results of two branches. Average tensor refers to the fusion method that directly averages the tensor output by the two branches." + ], + "table_footnote": [], + "table_body": "
Attention focusR/mR@20R/mR@50R/mR@100
subject-object32.6 / 20.938.0 / 27.438.9 / 28.4
full image30.4 / 19.336.5 / 25.937.1 / 26.8
", + "bbox": [ + 504, + 622, + 887, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation study for masked relation attention.", + "bbox": [ + 535, + 676, + 854, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the two branches, a one-to-one correspondence must be constructed based on the label assigned by each query to achieve prediction alignment.", + "bbox": [ + 498, + 696, + 890, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Masked relation attention. We investigate the impact of different mask input types for cross-attention on the HiLo baseline (Sec. 3.2) performance. Specifically, we compare two different attention focus regions, namely the subject-object region and the full image. The results are shown in Tab. 9. Focusing on the full image presents a more challenging optimization task for the model since no target region is specified. Consequently, we observe a drop of $1.8\\%$ in R@100 and $1.6\\%$ in mR@100. This shows that it is crucial to apply masked relation attention to the subject and object.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21644", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d3552fdcf427a91afb71c2ad0ba537c196e0af78455bf9b98cf0e8739a0bf80b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAllRareAllOverlap
PSGTR22.16.222.123.5
HiLo (ours)33.1 (+11.0)20.3 (+14.1)33.1 (+11.0)38.8 (+15.3)
", + "bbox": [ + 81, + 88, + 465, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d374e4df8be64179687aad0660ee28d0f4f8dcb574a20ef3b7d20b31074ce94a.jpg", + "image_caption": [ + "Figure 4. Convergence speed analysis of different methods. Our method converges significantly faster than previous methods." + ], + "image_footnote": [], + "bbox": [ + 81, + 191, + 465, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Analysis", + "text_level": 1, + "bbox": [ + 76, + 407, + 179, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To further demonstrate the efficacy of our method, we conduct the following analysis on HiLo.", + "bbox": [ + 76, + 430, + 467, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Long-tail problem and relational semantic overlap. To verify whether the problems of long-tail and relational semantic overlap have been tackled, we conduct experiments shown in Tab. 10. For the long-tail problem, we consider relations appearing less than 500 times (28 relations) in the PSG dataset (relations appear on average 4787 times) as rare relations, and report their mR@100. There is a $14\\%$ improvement for rare relations and an $11\\%$ improvement for all relations. This suggests that our method addresses the long-tail problem.", + "bbox": [ + 75, + 460, + 467, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the relational semantic overlap, we select all test images that have this problem (927 images) and report their mR@100. Our method shows a $15.3\\%$ improvement over PSGTR on images with semantic overlap relations, exceeding the overall improvement across all test images. This suggests that our approach addresses the problem of relational semantic overlap.", + "bbox": [ + 75, + 613, + 467, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Convergence speed and time cost analysis. We evaluate the convergence of our model by assessing its performance on the validation set at various epochs, as illustrated in Fig. 4. Our analysis reveals that our proposed method outperforms prior methods [62] both in terms of final performance and convergence speed. Specifically, PSGTR [62] achieves negligible performance in the initial 12 epochs, requiring 60 epochs to converge, as per the authors. In contrast, our HiLo method achieves better results in just 12 epochs, indicating its superior convergence speed.", + "bbox": [ + 75, + 719, + 467, + 869 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For time cost, we primarily analyze relation augmentation and swapping. 1) Relation augmentation, inspired by", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6c5a64a04850bc738007a3fd5e40358837002b7bd5db9145c65b201ff9b6366d.jpg", + "table_caption": [ + "Table 10. Verification of the improvements in the long-tail problem and relational semantic overlap. All refers to testing on the whole test set. Rare refers to testing only on rare relations. Overlap refers to testing only on data with relation semantic overlap." + ], + "table_footnote": [], + "table_body": "
MethodGFlopsParam. (M)Train Mem. (G)Infer. Time (ms)
PSGTR461.344.226.5140
HiLo (ours)229.458.716.1156
", + "bbox": [ + 504, + 88, + 888, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 11. Training and inference cost.", + "bbox": [ + 581, + 133, + 807, + 147 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "IETrans [66], involves training a baseline model and then using it to predict relation labels so as to augment the original relation labels. For the PSG dataset, following our experimental setup (see Sec. 4.3), this takes 18 hours. Afterwards, the training of our HiLo model requires 18 hours, which makes the whole process 36 hours. In contrast, the PSGTR model training takes 48 hours. Both our method and PSGTR utilize ResNet-50 as the backbone for fair comparison. Our method's rapid convergence (see Sec. 4.5) reduces training time compared to PSGTR, making the additional time cost for relation augmentation tolerable. 2) Relation swapping is a quick operation on relation labels during training and does not significantly contribute to overall time consumption.", + "bbox": [ + 496, + 150, + 890, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training and inference cost. Training and inference cost is shown in Tab. 11. Despite using two transformer-based decoders, our method requires less resource. Given the same input sizes (1280, 800) and ResNet-50 as backbone, the resource usage is shown in Tab. 11. PSGTR [62] generates mask features for subject and object separately, while our method reduces this computation by only generating mask features once. Our inference time is marginally higher due to more complex post-processing. Using H-L and L-H data only changes the labels between two branches, not increasing resource use.", + "bbox": [ + 496, + 362, + 890, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 534, + 617, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work we proposed the HiLo framework to tackle the long-tail problem with relational semantic overlap in Panoptic Scene Graph generation. The HiLo framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths by aligning their predictions. We also constructed a HiLo baseline to allow high-quality panoptic segmentation to improve PSG performance. Experimental results demonstrate that our method achieves state-of-the-art performance on the PSG dataset, confirming its effectiveness. In future work, we will investigate how knowledge distillation [28, 23] can be used to fuse the high and low branches in our method, as well as its application to downstream tasks such as visual question answering and image captioning.", + "bbox": [ + 496, + 559, + 890, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 777, + 651, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the European Union's Horizon 2020 FET Proactive Program under Agreement 101017857 and Fundamental Research Funds for the Central Universities.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21645", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 173:33-45, 2018. 1", + "[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 353-369. Springer, 2022. 1", + "[3] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2", + "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6", + "[5] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3", + "[7] Long Chen, Hanwang Zhang, Jun Xiao, Xiangnan He, Shiliang Pu, and Shih-Fu Chang. Counterfactual critic multiagent training for scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4613-4623, 2019. 6", + "[8] Shizhe Chen, Qin Jin, Peng Wang, and Qi Wu. Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9962-9971, 2020. 1", + "[9] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2", + "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 4, 6", + "[11] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3", + "[12] Meng-Jiun Chiou, Henghui Ding, Hanshu Yan, Changhu Wang, Roger Zimmermann, and Jiashi Feng. Recovering the" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "unbiased scene graphs from the biased ones. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1581-1590, 2021. 1, 2", + "[13] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[14] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 1", + "[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1", + "[16] Bo Dai, Yuqi Zhang, and Dahua Lin. Detecting visual relationships with deep relational networks. In Proceedings of the IEEE conference on computer vision and Pattern recognition, pages 3076-3086, 2017. 2", + "[17] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 266-283. Springer, 2022. 1, 2", + "[18] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 1, 2", + "[19] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 2", + "[20] Mohammed Haroon Dupty, Zhen Zhang, and Wee Sun Lee. Visual relationship detection with low rank non-negative tensor decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10737-10744, 2020. 2", + "[21] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the 2018 10th international conference on machine learning and computing, pages 225–229, 2018. 1", + "[22] Arushi Goel, Basura Fernando, Frank Keller, and Hakan Bilen. Not all relations are equal: Mining informative labels for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15596-15606, 2022. 2, 4", + "[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 9", + "[24] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21646", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 1969-1978, 2019. 2", + "[25] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 2", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[27] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1", + "[28] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 9", + "[29] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1", + "[30] Zih-Siou Hung, Arun Mallya, and Svetlana Lazebnik. Contextual translation embedding for visual relationship detection and scene graph generation. IEEE transactions on pattern analysis and machine intelligence, 43(11):3820-3832, 2020. 2", + "[31] Seong Jae Hwang, Sathya N Ravi, Zirui Tao, Hyunwoo J Kim, Maxwell D Collins, and Vikas Singh. Tensorize, factorize and regularize: Robust visual relationship learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1014-1023, 2018. 2", + "[32] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1", + "[33] Haeyong Kang and Chang D Yoo. Skew class-balanced reweighting for unbiased scene graph generation. Machine Learning and Knowledge Extraction, 5(1):287-303, 2023. 2", + "[34] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15879-15889, 2021. 2", + "[35] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 3", + "[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6", + "[37] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels:" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 2", + "[38] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19486-19496, 2022. 2", + "[39] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 2", + "[40] Wei Li, Haiwei Zhang, Qijie Bai, Guoqing Zhao, Ning Jiang, and Xiaojie Yuan. Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2022. 2", + "[41] Yikang Li, Wanli Ouyang, Xiaogang Wang, and Xiao'ou Tang. Vip-cnn: Visual phrase guided convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1347-1356, 2017. 2", + "[42] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 2", + "[43] Wentong Liao, Bodo Rosenhahn, Ling Shuai, and Michael Ying Yang. Natural language guided visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2", + "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 4", + "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[46] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2, 6, 7", + "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6", + "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6", + "[49] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21647", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 1, 2", + "[50] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2", + "[51] Mengshi Qi, Yunhong Wang, and Annan Li. Online cross-modal scene retrieval by binary representation and semantic graph. In Proceedings of the 25th ACM international conference on Multimedia, pages 744-752, 2017. 1", + "[52] Mohammad Amin Sadeghi and Ali Farhadi. Recognition using visual phrases. IEEE, 2011. 2", + "[53] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1", + "[54] Sahand Sharifzadeh, Sina Moayed Baharlou, Max Berrendorf, Rajat Koner, and Volker Tresp. Improving visual relation detection using depth maps. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 3597-3604. IEEE, 2021. 2", + "[55] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1", + "[56] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2", + "[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 3", + "[58] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 4", + "[59] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 2, 7", + "[60] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 2, 6", + "[62] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 2, 4, 6, 7, 9", + "[63] Jing Yu, Yuan Chai, Yujing Wang, Yue Hu, and Qi Wu. Cogtree: Cognition tree loss for unbiased scene graph generation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1274–1280. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 2, 6, 7", + "[64] Xu Chen Ya Zhang Xiao Gu Yue Hu, Siheng Chen. Neural message passing for visual relationship detection. In ICML Workshop on Learning and Reasoning with Graph-Structured Representations, Long Beach, CA, June 2019. 2", + "[65] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7", + "[66] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 409-424. Springer, 2022. 2, 4, 7, 9", + "[67] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and TatSeng Chua. Visual translation embedding network for visual relation detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5532-5540, 2017. 2", + "[68] Ji Zhang, Yannis Kalantidis, Marcus Rohrbach, Manohar Paluri, Ahmed Elgammal, and Mohamed Elhoseiny. Large-scale visual relationship understanding. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9185-9194, 2019. 2", + "[69] Sipeng Zheng, Shizhe Chen, and Qin Jin. Visual relation detection with multi-level attention. In Proceedings of the 27th ACM International Conference on Multimedia, pages 121-129, 2019. 2", + "[70] Guangming Zhu, Liang Zhang, Youliang Jiang, Yixuan Dang, Haoran Hou, Peiyi Shen, Mingtao Feng, Xia Zhao, Qiguang Miao, Syed Afaq Ali Shah, et al. Scene graph generation: A comprehensive survey. arXiv preprint arXiv:2201.00443, 2022. 2", + "[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\\{\\mathrm{detr}\\}$ : Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3", + "[72] Yaohui Zhu, Shuqiang Jiang, and Xiangyang Li. Visual relationship detection with object spatial distribution. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 379-384. IEEE, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 893 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21648", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_model.json b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..159a407d8817cbd03b1f98fa15cfec30d32865e6 --- /dev/null +++ b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_model.json @@ -0,0 +1,3953 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.221, + 0.131, + 0.75, + 0.176 + ], + "angle": 0, + "content": "HiLo: Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.205, + 0.251, + 0.222 + ], + "angle": 0, + "content": "Zijian Zhou" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.225, + 0.282, + 0.239 + ], + "angle": 0, + "content": "Department of Informatics" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.242, + 0.273, + 0.256 + ], + "angle": 0, + "content": "King's College London" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.261, + 0.296, + 0.274 + ], + "angle": 0, + "content": "zijian.zhou@kcl.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.426, + 0.204, + 0.546, + 0.222 + ], + "angle": 0, + "content": "Miaojing Shi" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.225, + 0.638, + 0.24 + ], + "angle": 0, + "content": "College of Electronic and Information Engineering" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.242, + 0.539, + 0.257 + ], + "angle": 0, + "content": "Tongji University" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.261, + 0.566, + 0.274 + ], + "angle": 0, + "content": "mshi@tongji.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.711, + 0.205, + 0.828, + 0.222 + ], + "angle": 0, + "content": "Holger Caesar" + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.225, + 0.842, + 0.239 + ], + "angle": 0, + "content": "Intelligent Vehicles Lab" + }, + { + "type": "text", + "bbox": [ + 0.676, + 0.242, + 0.863, + 0.257 + ], + "angle": 0, + "content": "Delft University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.261, + 0.853, + 0.273 + ], + "angle": 0, + "content": "h.caesar@tudelft.nl" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.313, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.675 + ], + "angle": 0, + "content": "Panoptic Scene Graph generation (PSG) is a recently proposed task in image scene understanding that aims to segment the image and extract triplets of subjects, objects and their relations to build a scene graph. This task is particularly challenging for two reasons. First, it suffers from a long-tail problem in its relation categories, making naive biased methods more inclined to high-frequency relations. Existing unbiased methods tackle the long-tail problem by data/loss rebalancing to favor low-frequency relations. Second, a subject-object pair can have two or more semantically overlapping relations. While existing methods favor one over the other, our proposed HiLo framework lets different network branches specialize on low and high frequency relations, enforce their consistency and fuse the results. To the best of our knowledge we are the first to propose an explicitly unbiased PSG method. In extensive experiments we show that our HiLo framework achieves state-of-the-art results on the PSG task. We also apply our method to the Scene Graph Generation task that predicts boxes instead of masks and see improvements over all baseline methods. Code is available at https://github.com/franciszzj/HiLo." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.702, + 0.21, + 0.717 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.848 + ], + "angle": 0, + "content": "Scene Graph Generation (SGG) [49] is a crucial task in image scene understanding that extracts triplets in the form of subjects, objects and their relations to build a scene graph. Subjects and objects are represented with bounding boxes. Since this task links vision and text, it holds great potential for a variety of applications, including visual question answering [27], image captioning [21, 8], image retrieval [32, 53, 51] and visual reasoning [1, 55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.848, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Recently a novel variant of SGG was proposed, which is Panoptic Scene Graph generation (PSG) [62]. Subjects" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.31, + 0.787, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.402, + 0.706, + 0.413 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.414, + 0.63, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.502, + 0.578, + 0.514 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.414, + 0.758, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.502, + 0.706, + 0.514 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.414, + 0.889, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.502, + 0.837, + 0.514 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.529, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Figure 1. An example of the PSG task. We compare the predicted object relations of different approaches. a) An image and its panoptic segmentation. Two persons and a sports ball are shown in different colors. b) A biased method predicts mostly high frequency relations. c) An unbiased method predicts mostly low frequency relations. d) Our method predicts both low and high frequency relations, as well as more relations in total." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.893, + 0.702 + ], + "angle": 0, + "content": "and objects are represented with panoptic segmentation [35] masks. Unless stated otherwise, in this work we focus on PSG, since it is pixel-level accurate and also covers background classes and their relations with foreground objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.839 + ], + "angle": 0, + "content": "The performance of the PSG model is affected by a long-tail problem in its relation categories. For instance, relations such as over, in front of and holding occur tens of thousands of times in the PSG dataset [62], while others like swinging and kissing occur only a few dozen times. This severe class imbalance in the relation categories can lead to model predictions that are more inclined to high-frequency relations, which poses significant challenges to the application of panoptic scene graphs in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Previous methods [12, 18, 39, 17] have often treated the long-tail problem of the PSG task as equivalent to the long-tail problem in object-centric tasks such as classification [15, 5, 29] or semantic segmentation [14, 2]. Consequently," + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21637" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "these methods have employed re-balancing techniques to address class imbalance, either through re-sampling the data [39] or by using a class-balanced loss [33] that assigns different weights to different relation categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.438 + ], + "angle": 0, + "content": "In contrast, in the relation-centric PSG task, a subject-object pair can have multiple relations that exhibit relational semantic overlap, such as being partially or fully overlapping. For example, in Fig. 1, there are multiple relations between the boy and the sports ball, such as beside, looking at, playing and chasing. For regular biased models [61, 65, 60, 46], the results are dominated by high-frequency relations (beside, looking at). For specifically unbiased models [66, 62], the results are dominated by low-frequency relations (playing, chasing). However, since the low frequency relations can be more specific (e.g. on and standing on) or only partially overlapping with high frequency relations (e.g. looking at and chasing), it is crucial to include both to fully understand the image. We found that relational semantic overlap occurs in large numbers in the PSG dataset [62] and that current methods do not effectively address it. This is reflected in the increase in the category-averaged mean recall metric of unbiased methods, at the cost of the decrease in global recall (see Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.589 + ], + "angle": 0, + "content": "To address the long-tail problem of scene graphs under relational semantic overlap, we introduce the HiLo framework. This framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths with the help of two novel consistency loss functions. We apply our framework on top of a novel baseline. This baseline uses a recent transformer-based approach [10] for panoptic segmentation and adapts triplet queries [62] and masked attention [10] for the PSG task. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.593, + 0.468, + 0.651 + ], + "angle": 0, + "content": "- We identify the long-tail problem with relational semantic overlap in the PSG task and propose the HiLo framework to address this problem. The framework is general and can be applied to any PSG method." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.656, + 0.468, + 0.713 + ], + "angle": 0, + "content": "- We propose a powerful and efficient one-stage end-to-end baseline. This baseline enhances the interaction between mask and relation prediction in the transformer decoder layer." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.719, + 0.468, + 0.79 + ], + "angle": 0, + "content": "- We conduct extensive experiments to demonstrate the effectiveness of our framework and baseline. Our results outperform the state-of-the-art in both recall and mean recall on the PSG dataset and show systematic improvements on the VG dataset." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.593, + 0.468, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.219, + 0.823 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.305, + 0.849 + ], + "angle": 0, + "content": "2.1. Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The Scene Graph Generation (SGG) [49] task plays a crucial role in connecting vision and language, and has received widespread attention in the computer vision commu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.349 + ], + "angle": 0, + "content": "nity. Many methods have been proposed to improve the performance of SGG, which can be classified into three categories [70]. The first is to introduce multi-modal information, such as appearance [52], space [72], depth [54], and segmentation [34]. The second is to introduce prior information and commonsense knowledge, such as statistical [3, 16, 65, 9] and language prior knowledge [49, 43, 68, 31, 20]. The third category involves designing different model structures, such as message passing [41, 16, 42, 65, 24, 64], attention mechanisms [69, 50], tree structures and visual translation [67, 30]. However, most of these methods are two-stage methods that cannot learn scene graphs end-to-end. In contrast, to improve the learning ability of SGG models, several methods based on transformers have been proposed, including SGTR [38], RelationFormer [56] and RelTR [13]. These are end-to-end trainable in a single stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.804, + 0.375 + ], + "angle": 0, + "content": "2.2. Unbiased Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Solving the long-tail problem in the SGG task has attracted considerable attention from researchers, and several unbiased methods [59, 63, 12, 18, 25, 39, 19, 22, 37, 40, 17, 66] have been proposed. These methods typically improve the model from the perspective of data re-sampling [39] or a class-balanced loss [33]. BGNN [39] uses a two-layer resampling strategy to provide a more balanced data distribution during training, while CogTree [63] proposal exploits the semantic relation between different predicate classes to design a novel CogTree loss. HML [17] improves the model's ability to solve long-tail problems by designing a staged training process, and IETrans [66] proposes an internal and external transfer method to transfer high frequency relations to low frequency relations and recover missing relations to train the unbiased model. Dong et al. [19] propose to group relations by their frequency and train specialized relation encoders for each group. While these methods are able to mitigate the long-tail problem, they do not address relational semantic overlap." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.8, + 0.697 + ], + "angle": 0, + "content": "2.3. Panoptic Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Contrary to SGG, the PSG task [62] uses panoptic segmentation masks instead of bounding boxes to represent objects, resulting in a more comprehensive scene graph. PSGTR [62], an end-to-end method based on the DETR structure, was proposed to construct a transformer-based PSG model. PSGFormer [62] further improved on PSGTR by introducing Object & Relation Query Learning Blocks and Query Matching Blocks. Their strong performance on most relation classes indicates that their method is implicitly unbiased. In contrast, our approach is the first to explicitly bias a PSG model, creating separate branches for low and high frequency relations, which are then fused together." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since the scene graph in PSG is built upon the sub" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21638" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.895, + 0.496 + ], + "angle": 0, + "content": "Figure 2. An overview of our HiLo framework with HiLo baseline. a) HiLo relation swapping module swaps the multiple relations in the subject-object pair to obtain H-L Data and L-H Data respectively. b) Input data into our HiLo framework with HiLo baseline model, there are two branches, namely H-L decoder and L-H decoder, which learn H-L Data and L-H Data respectively. c) In addition to task losses for PSG, we propose HiLo prediction alignment, which includes subject-object consistency loss and relation consistency loss, so that the parallel branch can be better optimized." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.675 + ], + "angle": 0, + "content": "jects, objects and their relations, strong panoptic segmentation [35] is crucial for PSG. Several DETR-based [6] methods such as Deformable-DETR [71], Segmenter [57], MaskFormer [11] and Mask2Former [10] have recently pushed the envelope in panoptic segmentation. These methods have introduced deformable transformer encoders and decoders, as well as pixel decoders and multi-scale information to improve model performance and speed up convergence. Our proposed baseline builds upon these techniques to further enhance the PSG performance and speed up model convergence." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.686, + 0.169, + 0.701 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.239, + 0.728 + ], + "angle": 0, + "content": "3.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The Panoptic Scene Graph generation (PSG) task aims to generate a panoptic scene graph \\(\\mathcal{G}\\) for a given image \\(\\mathcal{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where \\(\\mathcal{G}\\) contains an object set \\(\\mathcal{O}\\) and a relational triplet set \\(\\mathcal{T}\\), denoted by \\(\\mathcal{G} = (\\mathcal{O}, \\mathcal{T})\\). For the \\(i\\)-th object \\(o_i\\) in \\(\\mathcal{O} = \\{o_i\\}_{i=1}^N\\), we use \\(m_i\\) to represent the object's mask and \\(c_i\\) the object's category, i.e. \\(o_i = (m_i, c_i)\\). For the \\(j\\)-th triplet in \\(\\mathcal{T} = \\{t_j\\}_{j=1}^M\\), we use \\(s_j\\) to represent the subject, \\(o_j\\) the object, and \\(r_j\\) their relation, i.e. \\(t_j = (s_j, o_j, r_j)\\). There are in total \\(C\\) object categories and \\(R\\) relation categories. For the \\(k\\)-th relation, we denote its frequency in the training set as \\(f_k\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.645, + 0.52 + ], + "angle": 0, + "content": "3.2. HiLo Baseline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.607 + ], + "angle": 0, + "content": "High-quality panoptic segmentation is crucial for achieving good PSG performance. We build our method upon the latest advances in DETR-based panoptic segmentation, Mask2Former [10]. Below we present its structure, as well as our proposed modifications for the PSG task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.893, + 0.76 + ], + "angle": 0, + "content": "Mask2Former. This method comprises three key parts: A backbone (CNN-based or transformer-based) followed by a pixel decoder, a transformer module with a transformer decoder and a task-specific module with different task heads. Specifically, the backbone takes an input image \\(\\mathcal{I}\\) and generates an image feature \\(F\\). The pixel decoder then gradually upsamples \\(F\\) to produce multi-scale features \\(\\widetilde{\\mathcal{F}} = \\{\\widetilde{F}_i\\}_{i=1}^4\\). The transformer decoder takes a set of queries \\(\\mathcal{Q}\\) of size \\(N\\) and multi-scale features \\(\\widetilde{\\mathcal{F}}\\) as input and outputs a set of mask features \\(\\mathcal{X}\\) of the same size with \\(\\mathcal{Q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.868 + ], + "angle": 0, + "content": "On top of the transformer decoder, there are two task heads including a linear classifier that predicts the class probability for each mask; and a multi-layer perceptron (MLP) that uses the mask features \\(\\mathcal{X}\\) to generate the mask embedding \\(\\mathcal{E}\\). The mask prediction is obtained by taking the dot product of the mask embedding \\(\\mathcal{E}\\) with the scale feature of the highest resolution in the multi-scale features \\(\\widetilde{\\mathcal{F}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Triplet queries. The original query in Mask2Former is to predict the object. In order to predict both subject, object" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21639" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "and their relation, we develop the triplet queries \\(\\mathcal{Q}^t\\) inspired by [62] into our baseline. Each query predicts a triplet that includes subject, object, and relation. Accordingly, our task heads comprise three linear classifiers. Each classifier is responsible for predicting the class probability for subject, object or relation, respectively. Moreover, we devise two MLPs to generate mask embeddings for the subject and object, denoted by \\(\\mathcal{E}^s\\) and \\(\\mathcal{E}^o\\). We use them to compute the dot product with the scale feature of the highest resolution in multi-scale feature \\(\\widetilde{\\mathcal{F}}\\) and obtain the mask prediction for the subject and object, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.394 + ], + "angle": 0, + "content": "Masked relation attention. [10] proposes a transformer decoder variant with masked attention, which extracts spatial features by adding the predicted mask of the object from the previous decoder layer's mask prediction. It makes the model focus on the object-related area in the feature map. To adapt this scheme to the PSG task, for each relation, we extend the masked attention to take the union of the binary masks of the subject and object as input, which represents the pixels corresponding to the relation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.394, + 0.473, + 0.499 + ], + "angle": 0, + "content": "Network training. We adopt the same losses as PS-GTR [62], including cross-entropy loss \\(\\mathcal{L}_{so\\_cls}\\) for object classification of subject-object pairs and a combination \\(\\mathcal{L}_{so\\_mask}\\) of focal loss [44] and dice loss [58] to jointly supervise mask learning. To supervise the relations, we use the cross-entropy loss \\(\\mathcal{L}_{rel\\_cls}\\). The baseline loss with \\((\\lambda_1 = 1, \\lambda_2 = 1, \\lambda_3 = 4)\\) is thus:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.502, + 0.47, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b a s e l i n e}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {s o . c l s}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {s o . m a s k}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {r e l . c l s}} \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.25, + 0.565 + ], + "angle": 0, + "content": "3.3. HiLo Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.47, + 0.649 + ], + "angle": 0, + "content": "The key insight of our HiLo is to build a model that can take into account both high frequency and low frequency relations, and effectively improve the performance on low frequency relations without degrading the performance on high frequency relations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.305, + 0.682 + ], + "angle": 0, + "content": "3.3.1 HiLo relation generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.87 + ], + "angle": 0, + "content": "In the PSG task, multiple relations can be used to describe the connection between a subject and an object from different perspectives, such as spatial relations, actions and prepositions. Since the subject, object and their relative position in the image are fixed, all these relations share the same visual information. This is reflected in the PSG dataset [62] where many subject-object pairs are annotated with multiple relation labels. These relation labels are of different frequencies in the dataset. In this section we introduce the HiLo relation generation module to prepare two sets of training data, biasing towards high and low-frequency relations respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Relation augmentation. Similar to [22], we observe that there are many missing relation annotations in the PSG" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "dataset [62]. To add the missing relation annotations, we design a relation augmentation scheme that is inspired by IETrans [66], which converts high-frequency relations to low-frequency relations and adds a relation to a subject-object pair that has no relation. We adapt it by first training our baseline as a biased model using the original annotated data. For every subject-object pair in the training set, we use this model to predict the relation scores for all predefined relation categories:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.229, + 0.892, + 0.319 + ], + "angle": 0, + "content": "- If this subject-object pair has annotated relation labels, we pick the one with the highest predicted score and use the score as a threshold. For other relations (except for already labeled ones) whose predicted scores are greater than the threshold, we add them as the relation labels of this subject-object pair." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.32, + 0.892, + 0.395 + ], + "angle": 0, + "content": "- If this subject-object pair has no annotated relation labels, we use the predicted score of the no-relation class as a threshold. For other relations whose predicted scores are greater than this threshold we add them as the relation labels of this subject-object pair." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.229, + 0.892, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.457 + ], + "angle": 0, + "content": "This operation allows us to significantly augment the relation labels for the subject-object pairs in the training data, which is especially relevant for pairs with zero or no annotated relations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Relation swapping. We swap the relation labels for each subject-object pair. Specifically, given a subject-object pair \\((s,o)\\) with \\(K\\) relation labels, we have \\(K\\) triplets \\((s,o,r_1),\\ldots ,(s,o,r_K)\\), sorted by their relation frequency in descending order, \\(f_{1} > \\ldots >f_{K}\\). We denote by H-L and L-H the swapping of high-frequency relations with low-frequency relations and vice versa. This creates two sets of data:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.643 + ], + "angle": 0, + "content": "H-L Data. Given a triplet \\((s, o, r_k)\\), we replace its relation label \\(r_k\\) with that of the next triplet \\(r_{k+1}\\) with lower frequency. We sequentially process all triplets from \\((s, o, r_1)\\) until \\((s, o, r_{K-1})\\), keeping the last triplet unchanged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "L-H Data. Given a triplet \\((s,o,r_k)\\), we replace its relation label \\(r_k\\) with that of the previous triplet \\(r_{k - 1}\\) with higher frequency. We sequentially process all triplets from \\((s,o,r_K)\\) until \\((s,o,r_2)\\), keeping the first triplet unchanged. Hence, we obtain two new sets of triplets, denoted by \\(\\mathcal{T}^{\\mathrm{H - L}}\\) and \\(\\mathcal{T}^{\\mathrm{L - H}}\\). We devise two parallel decoders from the shared encoder of our backbone. They are learned with the H-L and L-H data, respectively. The H-L and L-H decoders favour the predictions for low and high-frequency relations, respectively. Despite their difference, for the corresponding triplet query in the two decoders, their predictions are highly correlated: on one hand, the subject and object predictions should be the same; on the other hand, the distribution of relation predictions should be overlapping. Below we first build the query correspondence in the two decoders and then introduce the HiLo subject-object and relation consistency loss to align the predictions from two the decoders." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21640" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.318, + 0.108 + ], + "angle": 0, + "content": "3.3.2 HiLo prediction alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.468, + 0.206 + ], + "angle": 0, + "content": "Training two different relation distributions simultaneously confuse the model. To make the model differentiate between the two branches, we built a HiLo prediction alignment module, including triplet query correspondence, subject-object consistency loss and relation consistency loss respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.207, + 0.468, + 0.416 + ], + "angle": 0, + "content": "Triplet query correspondence. In order to construct the subject-object consistency loss and relation consistency loss, we first need to construct the correspondence between the triplet queries in the H-L and L-H decoders. The same query index in the two decoders is not naturally matched. In order to find the query correspondence, we need to rely on the ground truth assignment: we use Hungarian matching to assign the triplet label to the corresponding triplet query, and record the relation label index corresponding to the triplet query. This label index allows us to construct the correspondence between the triplet queries in both decoders. We calculate the consistency loss for the triplet query prediction with the same relation label index in the two decoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.468, + 0.507 + ], + "angle": 0, + "content": "Subject-object consistency loss. Having the corresponding predictions from the two decoders, both their subjects and objects have the same ground truth and should be equal. We propose a subject-object consistency loss \\(\\mathcal{L}_{obj}\\) to minimize the mean squared error (MSE) of the corresponding predictions from the two decoders." + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.519, + 0.468, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o b j} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {m a s k} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.538, + 0.468, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = \\left\\| \\operatorname {s o f t m a x} \\left(p _ {c} ^ {\\mathrm {H} - \\mathrm {L}}\\right) - \\operatorname {s o f t m a x} \\left(p _ {c} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\right\\| ^ {2} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.558, + 0.468, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\left\\| \\operatorname {s i g m o i d} \\left(p _ {m} ^ {\\mathrm {H} - \\mathrm {L}}\\right) - \\operatorname {s i g m o i d} \\left(p _ {m} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\right\\| ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.468, + 0.614 + ], + "angle": 0, + "content": "Here \\(\\mathcal{L}_{cls}\\) and \\(\\mathcal{L}_{mask}\\) represent the MSE losses for class prediction and mask prediction, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.674 + ], + "angle": 0, + "content": "\\(p_c^{\\mathrm{H - L}}\\) and \\(p_c^{\\mathrm{L - H}}\\) are the class prediction logits from the H-L and L-H decoders. After a softmax, they are \\(C\\)-dimensional probability vectors. After a sigmoid, \\(p_m^{H - L}\\) and \\(p_m^{L - H}\\) are the mask prediction logits for the H-L and L-H decoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Relation consistency loss. Given a pair of subject and object, we have previously swapped the high-low frequency relation labels to create data for the H-L and L-H decoders. For H-L, the prediction of the low-frequency relation logit is of high value; while for L-H, the prediction of the high-frequency relation logit is of high value. For the predictions on the rest logits, they should be similar, since it is the same pair of subject and object for the two decoders. Based on this observation, we introduce the relation consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Specifically, for a pair of subject and object, we use \\( p_r^{\\mathrm{H - L}} \\) to denote the predicted relation logits from the H-L decoder, \\( p_r^{\\mathrm{L - H}} \\) the predicted relation logits from the L-H decoder. They are \\( R \\)-dimensional probability vectors after softmax. We can map between the distributions of the two vectors by swapping the logit values between the high- and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "low-frequency relation indices, which is named Relational Index Exchange operation. We use \\(\\mathrm{RIE}(\\cdot)\\) to denote this operation. \\(\\mathrm{RIE}(\\cdot)\\) includes a stop gradient operation, which creates copies from original predictions of two branches, enabling value exchanges. For example, for relations \\(r_k\\) and \\(r_{k + 1}\\), \\(\\mathrm{RIE}(\\cdot)\\) exchanges the values between \\(p_{r,k}\\) and \\(p_{r,k + 1}\\) in \\(p_r\\). We can therefore compute the distance between \\(p_r^{\\mathrm{H - L}}\\) and its mapped counterpart from \\(p_r^{\\mathrm{L - H}}\\), vice versa:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.219, + 0.891, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {D i s t} _ {\\mathrm {H i L o}} = \\left\\| \\text {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {H - L}}\\right) - R I E \\left(\\text {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {L - H}}\\right)\\right) \\right\\| ^ {2} \\\\ + \\| R I E \\left(\\operatorname {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {H} - \\mathrm {L}}\\right)\\right) - \\operatorname {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\| ^ {2} \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.891, + 0.311 + ], + "angle": 0, + "content": "The relation consistency loss is defined to minimize \\(\\mathrm{Dist}_{\\mathrm{HiLo}}\\) with a margin of \\(m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.315, + 0.89, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e l} = \\max \\left(\\operatorname {D i s t} _ {\\mathrm {H i L o}} - m, 0\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.891, + 0.384 + ], + "angle": 0, + "content": "where \\( m \\) is a small constant. Adding \\( m \\) is due to the fact that the high- and low frequency relations might be only partially semantically overlapping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.891, + 0.473 + ], + "angle": 0, + "content": "Network training. Our subject-object consistency loss and relation consistency loss can seamlessly integrate with the losses of any baseline method to jointly supervise the training of the entire model. Notably, we supervise the output of each transformer decoder layer to ensure effective learning. The final loss \\(\\mathcal{L}\\) is thus:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.486, + 0.89, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e l i n e}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {r e l}} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.705, + 0.541 + ], + "angle": 0, + "content": "3.3.3 HiLo inference fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.891, + 0.656 + ], + "angle": 0, + "content": "The H-L and L-H decoders favour low-frequency relation prediction and high-frequency relation prediction, respectively. To combine the strength of both during inference, we introduce the HiLo inference fusion module. Specifically, we denote by \\(\\mathcal{G}^{\\mathrm{H - L}}\\) and \\(\\mathcal{G}^{\\mathrm{L - H}}\\) the predicted panoptic scene graphs from H-L and L-H decoders, respectively. There are \\(N_{1}\\) triplets in \\(\\mathcal{G}^{\\mathrm{H - L}}\\) and \\(N_{2}\\) triplets in \\(\\mathcal{G}^{\\mathrm{L - H}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.656, + 0.89, + 0.701 + ], + "angle": 0, + "content": "- First, we merge the triplets in \\(\\mathcal{G}^{\\mathrm{H - L}}\\) and \\(\\mathcal{G}^{\\mathrm{L - H}}\\) and sort them in the descending order according to their relation scores. We obtain a list of \\(N_{1} + N_{2}\\) triplets." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.702, + 0.891, + 0.822 + ], + "angle": 0, + "content": "- Second, starting from the first triplet, we de-duplicate the triplet list. For the \\(i\\)-th triplet \\(\\mathcal{T}_i\\), we identify its duplicated versions from the \\((i + 1)\\)-th triplet until the end of the list. Remove any follow-up triplet from the list if it 1) has the same subject, object and relation classes to that in \\(T_i\\) and 2) has a mask IoU greater than a threshold, e.g. 0.5, between the predicted subject/object and the corresponding subject/object in \\(T_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.822, + 0.891, + 0.898 + ], + "angle": 0, + "content": "- Third, after dedduplication, for each triplet in the list, we multiply the relation, subject and object scores as an overall score for it. We sort the triplet list according to this score in descending order to obtain the final panoptic scene graph \\(\\mathcal{G}^{\\mathrm{HiLo}}\\)." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.656, + 0.891, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21641" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.074, + 0.805, + 0.244 + ], + "angle": 0, + "content": "
MethodBackboneScene Graph Detection
R@20mR@20R@50mR@50R@100mR@100
IMP [61]R5016.56.518.27.118.67.2
MOTIF [65]R5020.09.121.79.622.09.7
VCTree [60]R5020.69.722.110.222.510.2
GPSNet [46]R5017.87.019.67.520.17.7
PSGTR [62]R5028.416.634.420.836.322.1
PSGFormer [62]R5018.014.819.617.020.117.6
HiLo (ours)R5034.123.740.730.343.033.1
HiLo (ours)Swin-B38.528.346.235.349.639.1
HiLo (ours)Swin-L40.629.748.737.651.440.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.246, + 0.895, + 0.276 + ], + "angle": 0, + "content": "Table 1. Comparison between our HiLo and other methods on the PSG dataset. Our method shows superior performance compared to all previous methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.291, + 0.21, + 0.308 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.321, + 0.182, + 0.336 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.471, + 0.502 + ], + "angle": 0, + "content": "Panoptic Scene Graph Generation (PSG) [62]. This is the first Panoptic Scene Graph generation dataset. It has a total of 48,749 labeled images including 2,177 test images and 46,572 training images. The object categories comprise 80 thing classes and 53 stuff classes, which is the same as the COCO [45] and COCO-Stuff datasets [4]. The relation categories comprise 56 classes, including positional relations, common object-object relations, common actions, human actions, actions in the traffic scene, actions in the sports scene and interactions between backgrounds [62]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Visual Genome (VG) [36]. VG is a widely used benchmark dataset for Scene Graph Generation. Following previous work [65, 7], we adopt the widely accepted split, VG-150, which contains 150 object categories and 50 relation categories. The object categories cover a wide range of classes, such as animals, vehicles and household items. The relation categories include both spatial and semantic classes, such as on, in and wearing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.254, + 0.67 + ], + "angle": 0, + "content": "4.2. Tasks and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Three subtasks have been proposed for the SGG and PSG tasks, which are Predicate Classification, Scene Graph Classification and Scene Graph Detection [61]. We focus on Scene Graph Detection for both datasets, since it is the most comprehensive and addressed by [62]. This subtask requires the model to first localize the objects and then predict the object classes and relations. Note that Scene Graph Detection in PSG includes the detection on stuff classes, while in SGG it does not." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Following previous work [60, 63, 62], we use Recall@K (R@K) and mean Recall@K (mR@K) as our metrics, where the former metric is dominated by high-frequency relations, while the latter assigns equal weight to all relation classes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.291, + 0.715, + 0.308 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.895, + 0.543 + ], + "angle": 0, + "content": "In our experiments, we follow the training strategy of PSGTR [62]. We use the AdamW optimizer [48], with a learning rate of \\(1e^{-4}\\) and weight decay of \\(1e^{-4}\\), except for the backbone, which is trained with a learning rate of \\(1e^{-5}\\). For initialization, we use Mask2Former [10] pretrained on COCO [45] to initialize our backbone and pixel decoder. Following Mask2Former [10], we use 100 triplet queries for the H-L and L-H decoders respectively. Additionally, both the H-L and L-H decoders are initialized with Mask2Former's transformer decoder. To ensure consistent comparison with PSGTR, we adopted the same data augmentation settings. Our model is trained for 12 epochs with a step scheduler at epoch 10, taking approximately 18 hours to train on four A100 GPUs with a batch size of 1 for each GPU." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.559, + 0.8, + 0.575 + ], + "angle": 0, + "content": "4.4. Comparison to the state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.903 + ], + "angle": 0, + "content": "PSG. Tab. 1 reports the performance of our method compared to the state-of-the-art on the PSG dataset [62]. We separate the methods into two groups. The first are two-stage methods consisting of a separate segmentor and relation predictor, which are modified for the PSG task in [62]. The second are one-stage end-to-end methods, which are able to simultaneously predict panoptic segmentation and relations. Our method belongs to the second category. For a fair comparison between the different methods, we use the same Resnet-50 [26] backbone. Our method shows superior performance compared to all previous methods. Particularly, it outperforms the previous best-performing method PSGTR [62] by a large margin, i.e. \\(+6\\%\\) in R@100 and \\(+11\\%\\) in mR@100. Our model is able to converge within only 12 epochs of training, whereas PSGTR [62] is trained for 60 epochs. We also evaluate our method using pre-trained transformer-based backbones, i.e. Swin-B and Swin-L [47], with the latter being a bigger model. Our results are consistently improved over all metrics due the powerful feature representation ability of the transformer. We also conducted a visual comparison, as shown in Fig. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21642" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.422, + 0.287 + ], + "angle": 0, + "content": "
MethodScene Graph Detection
R/mR@50R/mR@100
MOTIF [65]31.0 / 6.735.1 / 7.7
+IETrans [66]26.4 / 12.430.6 / 14.9
+HiLo (ours)26.2 / 14.730.3 / 17.7
VCTree [60]30.2 / 6.734.6 / 8.0
+IETrans [66]25.4 / 11.529.3 / 14.0
+HiLo (ours)27.1 / 12.929.8 / 15.2
Transformer [59]30.0 / 7.434.3 / 8.8
+IETrans [66]25.5 / 12.529.6 / 15.0
+HiLo (ours)25.4 / 14.629.2 / 17.6
GPSNet [46]30.3 / 5.935.0 / 7.1
+IETrans [66]25.9 / 14.628.1 / 16.5
+HiLo (ours)25.6 / 15.827.9 / 18.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.289, + 0.47, + 0.332 + ], + "angle": 0, + "content": "Table 2. Comparison between our HiLo framework and other methods on the VG-150 dataset. Similar to [66], we apply IETrans and our own method on top of four leading baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.626 + ], + "angle": 0, + "content": "SGG. Here we study whether our approach that was developed for the PSG task can also be applied to the SGG task. In Tab. 2, we conduct experiments on the VG-150 dataset. Following IETrans [66], we extend our unbiased HiLo framework (Sec. 3.3) without our PSG-specific baseline (Sec. 3.2) to four state-of-the-art biased SGG methods, namely MOTIF [65], VCTree [60], Transformer [59] and GPSNet [46], using the implementation of [66]. Compared to the unbiased IETrans [66] method, our method improves mean recall without sacrificing recall, demonstrating its effectiveness in enhancing the performance of low-frequency relations. Compared to the biased baselines [65, 60, 59, 46], our method achieves a significantly larger mean recall, while still maintaining an acceptable recall. This indicates that our method can improve the performance of low-frequency relations while also taking into account the performance of high-frequency relations. It also shows that the HiLo framework is a general technique that yields systematic improvements in both the PSG and SGG tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.241, + 0.65 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Consistent with the paper, we use HiLo with a Resnet-50 backbone to perform ablation experiments on the PSG dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.884 + ], + "angle": 0, + "content": "HiLo framework for different baselines. In this section we investigate whether our HiLo framework (Sec. 3.3) yields improvements for other baselines, rather than just the one presented in Sec. 3.2. Tab. 3 shows the results for two baselines, with and without the HiLo framework. We observe that our biased baseline outperforms the previous PSGTR [62] method on all metrics. Furthermore, by applying the HiLo framework, we can substantially improve the performance over both baselines. It is worth mentioning that the HiLo framework improves recall and mean recall simultaneously, whereas other methods typically improve one metric at the cost of the other [59, 63]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "HiLo relation augmentation. We observe that out of" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.888, + 0.158 + ], + "angle": 0, + "content": "
BaselineHiLoR/mR@20R/mR@50R/mR@100
HiLo baseline✓34.1 / 23.740.7 / 30.343.0 / 33.1
HiLo baseline-32.6 / 20.938.0 / 27.438.9 / 28.4
PSGTR [62]✓30.1 / 20.236.6 / 23.938.3 / 24.5
PSGTR [62]-28.4 / 16.634.4 / 20.836.3 / 22.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.164, + 0.892, + 0.206 + ], + "angle": 0, + "content": "Table 3. Comparison of different baselines, with and without HiLo framework. Using the HiLo framework, we see significant improvements on both metrics." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.212, + 0.888, + 0.255 + ], + "angle": 0, + "content": "
Relation Aug.Multiple relationsR/mR@50R/mR@100
✓40%40.7 / 30.343.0 / 33.1
-10%40.1 / 28.142.8 / 32.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Table 4. Ablation study for HiLo relation augmentation. Relation augmentation affects the ratio of subject-object pairs with multiple relations. The larger this ratio, the more relations can be swapped, which leads to better results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.478 + ], + "angle": 0, + "content": "260,296 labeled triplets in the PSG dataset, only about \\(10\\%\\) of subject-object pairs have multiple relations, for which we can apply relation swapping (Sec. 3.3.1). After applying our proposed relation augmentation technique (Sec. 3.3.1), this ratio significantly increases to \\(40\\%\\). Our experimental results in Tab. 4 demonstrate that only applying the HiLo framework on \\(10\\%\\) already gives an improvement over the baseline from Tab. 3. As the number of swappable triplets increases due to augmentation, the model's performance is further enhanced, highlighting the potential of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.644 + ], + "angle": 0, + "content": "HiLo prediction alignment. We conduct ablation experiments on the subject-object consistency loss and relation consistency loss (Sec. 3.3.2), which are used to align the predictions from the high and low frequency branches. The results, as presented in Tab. 5, demonstrate that using both losses yields the best performance. It is worth mentioning that we have explored the margin in the relation consistency loss and found that setting the margin to zero leads to a small performance degradation. This finding confirms that there is a partial semantic overlap between swapped relations, indicating that they are not entirely consistent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.809 + ], + "angle": 0, + "content": "To investigate the impact of relational index exchange (RIE) on the relation consistency loss, we conducted experiments to verify the effect of omitting RIE. In the absence of RIE, we solely compute the consistency loss for relation categories that are not involved in relation swapping and exclude the swapping component from the calculation of the consistency loss. The outcomes of this experiment are presented in Tab. 6, and demonstrate a notable reduction in the mean recall and a decline in the model's performance for relations with relational semantic overlap when RIE is not utilized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "HiLo inference fusion. We ablate the inference fusion (Sec. 3.3.3) and evaluate the performance of each branch's output separately. In Tab. 7, experimental results suggest that fusion can effectively leverage the uniqueness of the high and low frequency branch predictions to achieve comprehensive improvements on all metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21643" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.095, + 0.212, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.165, + 0.164, + 0.178 + ], + "angle": 0, + "content": "↑ Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.182, + 0.206, + 0.189 + ], + "angle": 0, + "content": "Prediction \\(\\rightarrow\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.19, + 0.176, + 0.198 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.21, + 0.212, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.281, + 0.185, + 0.287 + ], + "angle": 0, + "content": "person carrying surfboard" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.288, + 0.141, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.296, + 0.157, + 0.304 + ], + "angle": 0, + "content": "looking at" + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.313, + 0.165, + 0.319 + ], + "angle": 0, + "content": "sky over person" + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.32, + 0.14, + 0.328 + ], + "angle": 0, + "content": "sky over sea" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.096, + 0.342, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.166, + 0.313, + 0.174 + ], + "angle": 0, + "content": "person holding surfboard" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.175, + 0.269, + 0.182 + ], + "angle": 0, + "content": "sky over sea" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.183, + 0.332, + 0.189 + ], + "angle": 0, + "content": "person standing on sea" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.19, + 0.271, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.197, + 0.27, + 0.203 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.204, + 0.281, + 0.21 + ], + "angle": 0, + "content": "sky over person" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.211, + 0.303, + 0.217 + ], + "angle": 0, + "content": "person walking on sea" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.218, + 0.29, + 0.224 + ], + "angle": 0, + "content": "surfboard over sea" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.225, + 0.273, + 0.231 + ], + "angle": 0, + "content": "sky over surf" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.232, + 0.273, + 0.239 + ], + "angle": 0, + "content": "person on sea" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.24, + 0.316, + 0.247 + ], + "angle": 0, + "content": "person carrying surfboard" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.248, + 0.27, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.255, + 0.299, + 0.262 + ], + "angle": 0, + "content": "looking at you" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.264, + 0.324, + 0.271 + ], + "angle": 0, + "content": "person attached to surfboard" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.272, + 0.324, + 0.279 + ], + "angle": 0, + "content": "surfboard attached to person" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.28, + 0.296, + 0.287 + ], + "angle": 0, + "content": "surfboard on person" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.288, + 0.299, + 0.295 + ], + "angle": 0, + "content": "sea in front of person" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.296, + 0.287, + 0.303 + ], + "angle": 0, + "content": "sea beside person" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.304, + 0.27, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.312, + 0.312, + 0.319 + ], + "angle": 0, + "content": "person playing surfboard" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.32, + 0.284, + 0.327 + ], + "angle": 0, + "content": "surfboard on sea" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.095, + 0.483, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.168, + 0.434, + 0.175 + ], + "angle": 0, + "content": "↑" + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.176, + 0.433, + 0.183 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.185, + 0.476, + 0.192 + ], + "angle": 0, + "content": "Prediction→" + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.194, + 0.475, + 0.202 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.203, + 0.446, + 0.209 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.216, + 0.484, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.288, + 0.433, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.296, + 0.457, + 0.303 + ], + "angle": 0, + "content": "person carrying backpack" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.304, + 0.457, + 0.311 + ], + "angle": 0, + "content": "person holding skis" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.312, + 0.446, + 0.319 + ], + "angle": 0, + "content": "person climbing snow." + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.32, + 0.419, + 0.327 + ], + "angle": 0, + "content": "sky over snow." + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.095, + 0.614, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.168, + 0.587, + 0.175 + ], + "angle": 0, + "content": "person carrying backpack" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.176, + 0.549, + 0.183 + ], + "angle": 0, + "content": "person holding" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.185, + 0.525, + 0.191 + ], + "angle": 0, + "content": "skis on :" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.192, + 0.548, + 0.199 + ], + "angle": 0, + "content": "sky over snow" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.2, + 0.552, + 0.207 + ], + "angle": 0, + "content": "sky over person" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.208, + 0.583, + 0.215 + ], + "angle": 0, + "content": "person standing on snow" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.216, + 0.543, + 0.223 + ], + "angle": 0, + "content": "sky over skis" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.224, + 0.562, + 0.231 + ], + "angle": 0, + "content": "sky over backpack" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.232, + 0.545, + 0.239 + ], + "angle": 0, + "content": "sky over rock" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.24, + 0.574, + 0.247 + ], + "angle": 0, + "content": "person climbing snow" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.248, + 0.581, + 0.255 + ], + "angle": 0, + "content": "person walking on snow" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.256, + 0.579, + 0.262 + ], + "angle": 0, + "content": "personbesideperson" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.263, + 0.552, + 0.269 + ], + "angle": 0, + "content": "person on snow" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.27, + 0.58, + 0.277 + ], + "angle": 0, + "content": "person running on snow" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.278, + 0.58, + 0.285 + ], + "angle": 0, + "content": "person picking skis" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.286, + 0.563, + 0.293 + ], + "angle": 0, + "content": "person playing skills" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.294, + 0.577, + 0.301 + ], + "angle": 0, + "content": "person playing with person attached to skis" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.302, + 0.596, + 0.309 + ], + "angle": 0, + "content": "person attached to backpack" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.31, + 0.563, + 0.317 + ], + "angle": 0, + "content": "mountain attached" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.318, + 0.563, + 0.325 + ], + "angle": 0, + "content": "sky over mountain" + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.326, + 0.563, + 0.331 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.096, + 0.756, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.168, + 0.704, + 0.175 + ], + "angle": 0, + "content": "↑" + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.176, + 0.705, + 0.184 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.186, + 0.75, + 0.193 + ], + "angle": 0, + "content": "Prediction \\(\\rightarrow\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.195, + 0.721, + 0.203 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.204, + 0.695, + 0.21 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.218, + 0.756, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.286, + 0.755, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.294, + 0.744, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.301, + 0.696, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.308, + 0.744, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.316, + 0.744, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.323, + 0.714, + 0.329 + ], + "angle": 0, + "content": "chair leaning on wall" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.096, + 0.885, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.168, + 0.821, + 0.175 + ], + "angle": 0, + "content": "holding" + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.176, + 0.82, + 0.183 + ], + "angle": 0, + "content": "about to" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.185, + 0.848, + 0.191 + ], + "angle": 0, + "content": "person in front of wall" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.192, + 0.828, + 0.199 + ], + "angle": 0, + "content": "chair beside wall" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.2, + 0.825, + 0.207 + ], + "angle": 0, + "content": "chair in front of" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.208, + 0.849, + 0.215 + ], + "angle": 0, + "content": "person in front of chair" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.216, + 0.828, + 0.223 + ], + "angle": 0, + "content": "person looking at" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.224, + 0.825, + 0.231 + ], + "angle": 0, + "content": "person throwing" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.232, + 0.839, + 0.239 + ], + "angle": 0, + "content": "wall in front of tree" + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.24, + 0.842, + 0.246 + ], + "angle": 0, + "content": "over chair" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.247, + 0.872, + 0.254 + ], + "angle": 0, + "content": "person swinging tennis racket" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.255, + 0.825, + 0.262 + ], + "angle": 0, + "content": "treebesidewall" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.263, + 0.836, + 0.269 + ], + "angle": 0, + "content": "chair beside person" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.27, + 0.843, + 0.277 + ], + "angle": 0, + "content": "chair attached to wall" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.278, + 0.859, + 0.285 + ], + "angle": 0, + "content": "person holding sports ball" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.286, + 0.82, + 0.292 + ], + "angle": 0, + "content": "mehohid" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.293, + 0.842, + 0.299 + ], + "angle": 0, + "content": "person beside barrier" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.3, + 0.824, + 0.306 + ], + "angle": 0, + "content": "wall beside tree" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.307, + 0.857, + 0.313 + ], + "angle": 0, + "content": "person playing sports ball" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.314, + 0.819, + 0.319 + ], + "angle": 0, + "content": "tree in front of" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.32, + 0.82, + 0.327 + ], + "angle": 0, + "content": "tree over wall" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.894, + 0.392 + ], + "angle": 0, + "content": "Figure 3. Visualization of panoptic segmentations and the top 20 predicted triplets compared with ground truth. The upper left is the original image, the lower left is the ground truth and on the right are the predictions. The highlighted triplets represent the subject-object pairs with multiple relations, where the blue highlights represent the high frequency relations and the red highlights represent the low frequency relations. The visualization results show that our method can predict both high frequency and low frequency relations." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.399, + 0.466, + 0.489 + ], + "angle": 0, + "content": "
ObjectRelationMarginR/mR@50R/mR@100
✓✓0.540.7 / 30.343.0 / 33.1
-✓0.540.6 / 29.742.8 / 32.8
-✓0.040.5 / 29.542.7 / 32.8
✓--40.4 / 29.042.8 / 32.2
---39.7 / 28.642.4 / 32.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Table 5. Ablation study for different losses in HiLo prediction alignment. Object refers to the subject-object consistency loss and relation refers to the relation consistency loss. The margin parameter is defined in Eq. 6." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.556, + 0.466, + 0.6 + ], + "angle": 0, + "content": "
Whether to use RIER/mR@20R/mR@50R/mR@100
✓34.1 / 23.740.7 / 30.343.0 / 33.1
-33.5 / 22.340.3 / 29.042.6 / 32.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.636 + ], + "angle": 0, + "content": "Table 6. Ablation study for relation consistency loss in HiLo prediction alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We also attempted to average the tensor generated by the two branches and obtain the PSG result through postprocessing. However, we found that this approach leads to a substantial drop in performance, as evident in Tab. 8. This can be attributed to the inconsistent prediction results of the two branches for the same query index. These findings validate that the inference fusion method effectively merges the results from the two branches. Furthermore, our experimental results demonstrate that the query associated with the identical index in two branches does not predict the same subject-object pair. Thus, directly averaging the tensor produced by the two branches results in prediction ambiguity, ultimately leading to a substantial decline in performance. This observation underscores the necessity of conducting triplet query correspondence when performing prediction alignment. In particular, due to the inconsistent query prediction content for the corresponding index" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.399, + 0.888, + 0.464 + ], + "angle": 0, + "content": "
H-L ResultL-H ResultR/mR@50R/mR@100
✓✓40.7 / 30.343.0 / 33.1
✓-38.8 / 29.939.8 / 30.9
-✓38.5 / 26.539.5 / 27.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.546, + 0.472, + 0.846, + 0.486 + ], + "angle": 0, + "content": "Table 7. Ablation study for HiLo inference fusion." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.5, + 0.888, + 0.547 + ], + "angle": 0, + "content": "
Fusion methodR/mR@20R/mR@50R/mR@100
inference fusion34.1 / 23.740.7 / 30.343.0 / 33.1
average tensor19.6 / 13.123.1 / 15.723.9 / 16.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.553, + 0.893, + 0.61 + ], + "angle": 0, + "content": "Table 8. Ablation study for HiLo inference fusion. Inference fusion refers to the method proposed in the paper to fuse the results of two branches. Average tensor refers to the fusion method that directly averages the tensor output by the two branches." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.623, + 0.888, + 0.671 + ], + "angle": 0, + "content": "
Attention focusR/mR@20R/mR@50R/mR@100
subject-object32.6 / 20.938.0 / 27.438.9 / 28.4
full image30.4 / 19.336.5 / 25.937.1 / 26.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.678, + 0.855, + 0.692 + ], + "angle": 0, + "content": "Table 9. Ablation study for masked relation attention." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.697, + 0.892, + 0.743 + ], + "angle": 0, + "content": "in the two branches, a one-to-one correspondence must be constructed based on the label assigned by each query to achieve prediction alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Masked relation attention. We investigate the impact of different mask input types for cross-attention on the HiLo baseline (Sec. 3.2) performance. Specifically, we compare two different attention focus regions, namely the subject-object region and the full image. The results are shown in Tab. 9. Focusing on the full image presents a more challenging optimization task for the model since no target region is specified. Consequently, we observe a drop of \\(1.8\\%\\) in R@100 and \\(1.6\\%\\) in mR@100. This shows that it is crucial to apply masked relation attention to the subject and object." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "21644" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.466, + 0.127 + ], + "angle": 0, + "content": "
MethodAllRareAllOverlap
PSGTR22.16.222.123.5
HiLo (ours)33.1 (+11.0)20.3 (+14.1)33.1 (+11.0)38.8 (+15.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.134, + 0.468, + 0.19 + ], + "angle": 0, + "content": "Table 10. Verification of the improvements in the long-tail problem and relational semantic overlap. All refers to testing on the whole test set. Rare refers to testing only on rare relations. Overlap refers to testing only on data with relation semantic overlap." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.192, + 0.466, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.376, + 0.468, + 0.404 + ], + "angle": 0, + "content": "Figure 4. Convergence speed analysis of different methods. Our method converges significantly faster than previous methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.408, + 0.18, + 0.423 + ], + "angle": 0, + "content": "4.6. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.468, + 0.461 + ], + "angle": 0, + "content": "To further demonstrate the efficacy of our method, we conduct the following analysis on HiLo." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.468, + 0.612 + ], + "angle": 0, + "content": "Long-tail problem and relational semantic overlap. To verify whether the problems of long-tail and relational semantic overlap have been tackled, we conduct experiments shown in Tab. 10. For the long-tail problem, we consider relations appearing less than 500 times (28 relations) in the PSG dataset (relations appear on average 4787 times) as rare relations, and report their mR@100. There is a \\(14\\%\\) improvement for rare relations and an \\(11\\%\\) improvement for all relations. This suggests that our method addresses the long-tail problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.718 + ], + "angle": 0, + "content": "For the relational semantic overlap, we select all test images that have this problem (927 images) and report their mR@100. Our method shows a \\(15.3\\%\\) improvement over PSGTR on images with semantic overlap relations, exceeding the overall improvement across all test images. This suggests that our approach addresses the problem of relational semantic overlap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.87 + ], + "angle": 0, + "content": "Convergence speed and time cost analysis. We evaluate the convergence of our model by assessing its performance on the validation set at various epochs, as illustrated in Fig. 4. Our analysis reveals that our proposed method outperforms prior methods [62] both in terms of final performance and convergence speed. Specifically, PSGTR [62] achieves negligible performance in the initial 12 epochs, requiring 60 epochs to converge, as per the authors. In contrast, our HiLo method achieves better results in just 12 epochs, indicating its superior convergence speed." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "For time cost, we primarily analyze relation augmentation and swapping. 1) Relation augmentation, inspired by" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.889, + 0.127 + ], + "angle": 0, + "content": "
MethodGFlopsParam. (M)Train Mem. (G)Infer. Time (ms)
PSGTR461.344.226.5140
HiLo (ours)229.458.716.1156
" + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.135, + 0.808, + 0.148 + ], + "angle": 0, + "content": "Table 11. Training and inference cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.151, + 0.892, + 0.362 + ], + "angle": 0, + "content": "IETrans [66], involves training a baseline model and then using it to predict relation labels so as to augment the original relation labels. For the PSG dataset, following our experimental setup (see Sec. 4.3), this takes 18 hours. Afterwards, the training of our HiLo model requires 18 hours, which makes the whole process 36 hours. In contrast, the PSGTR model training takes 48 hours. Both our method and PSGTR utilize ResNet-50 as the backbone for fair comparison. Our method's rapid convergence (see Sec. 4.5) reduces training time compared to PSGTR, making the additional time cost for relation augmentation tolerable. 2) Relation swapping is a quick operation on relation labels during training and does not significantly contribute to overall time consumption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Training and inference cost. Training and inference cost is shown in Tab. 11. Despite using two transformer-based decoders, our method requires less resource. Given the same input sizes (1280, 800) and ResNet-50 as backbone, the resource usage is shown in Tab. 11. PSGTR [62] generates mask features for subject and object separately, while our method reduces this computation by only generating mask features once. Our inference time is marginally higher due to more complex post-processing. Using H-L and L-H data only changes the labels between two branches, not increasing resource use." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.618, + 0.55 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.772 + ], + "angle": 0, + "content": "In this work we proposed the HiLo framework to tackle the long-tail problem with relational semantic overlap in Panoptic Scene Graph generation. The HiLo framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths by aligning their predictions. We also constructed a HiLo baseline to allow high-quality panoptic segmentation to improve PSG performance. Experimental results demonstrate that our method achieves state-of-the-art performance on the PSG dataset, confirming its effectiveness. In future work, we will investigate how knowledge distillation [28, 23] can be used to fuse the high and low branches in our method, as well as its application to downstream tasks such as visual question answering and image captioning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.652, + 0.794 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the European Union's Horizon 2020 FET Proactive Program under Agreement 101017857 and Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21645" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 173:33-45, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 353-369. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[3] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[5] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[7] Long Chen, Hanwang Zhang, Jun Xiao, Xiangnan He, Shiliang Pu, and Shih-Fu Chang. Counterfactual critic multiagent training for scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4613-4623, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Qin Jin, Peng Wang, and Qi Wu. Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9962-9971, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.677, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[9] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[12] Meng-Jiun Chiou, Henghui Ding, Hanshu Yan, Changhu Wang, Roger Zimmermann, and Jiashi Feng. Recovering the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "unbiased scene graphs from the biased ones. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1581-1590, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[13] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[14] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.894, + 0.376 + ], + "angle": 0, + "content": "[16] Bo Dai, Yuqi Zhang, and Dahua Lin. Detecting visual relationships with deep relational networks. In Proceedings of the IEEE conference on computer vision and Pattern recognition, pages 3076-3086, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[17] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 266-283. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[18] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[19] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[20] Mohammed Haroon Dupty, Zhen Zhang, and Wee Sun Lee. Visual relationship detection with low rank non-negative tensor decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10737-10744, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[21] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the 2018 10th international conference on machine learning and computing, pages 225–229, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[22] Arushi Goel, Basura Fernando, Frank Keller, and Hakan Bilen. Not all relations are equal: Mining informative labels for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15596-15606, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[24] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21646" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 1969-1978, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[25] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[27] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[28] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[29] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[30] Zih-Siou Hung, Arun Mallya, and Svetlana Lazebnik. Contextual translation embedding for visual relationship detection and scene graph generation. IEEE transactions on pattern analysis and machine intelligence, 43(11):3820-3832, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[31] Seong Jae Hwang, Sathya N Ravi, Zirui Tao, Hyunwoo J Kim, Maxwell D Collins, and Vikas Singh. Tensorize, factorize and regularize: Robust visual relationship learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1014-1023, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[32] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[33] Haeyong Kang and Chang D Yoo. Skew class-balanced reweighting for unbiased scene graph generation. Machine Learning and Knowledge Extraction, 5(1):287-303, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[34] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15879-15889, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[38] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19486-19496, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[39] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[40] Wei Li, Haiwei Zhang, Qijie Bai, Guoqing Zhao, Ning Jiang, and Xiaojie Yuan. Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[41] Yikang Li, Wanli Ouyang, Xiaogang Wang, and Xiao'ou Tang. Vip-cnn: Visual phrase guided convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1347-1356, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[42] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[43] Wentong Liao, Bodo Rosenhahn, Ling Shuai, and Michael Ying Yang. Natural language guided visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[46] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21647" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.135 + ], + "angle": 0, + "content": "Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[50] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[51] Mengshi Qi, Yunhong Wang, and Annan Li. Online cross-modal scene retrieval by binary representation and semantic graph. In Proceedings of the 25th ACM international conference on Multimedia, pages 744-752, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[52] Mohammad Amin Sadeghi and Ali Farhadi. Recognition using visual phrases. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.294, + 0.469, + 0.363 + ], + "angle": 0, + "content": "[53] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.365, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[54] Sahand Sharifzadeh, Sina Moayed Baharlou, Max Berrendorf, Rajat Koner, and Volker Tresp. Improving visual relation detection using depth maps. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 3597-3604. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.491 + ], + "angle": 0, + "content": "[55] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.493, + 0.469, + 0.576 + ], + "angle": 0, + "content": "[56] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[58] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[59] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[60] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[61] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[62] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 2, 4, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[63] Jing Yu, Yuan Chai, Yujing Wang, Yue Hu, and Qi Wu. Cogtree: Cognition tree loss for unbiased scene graph generation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1274–1280. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[64] Xu Chen Ya Zhang Xiao Gu Yue Hu, Siheng Chen. Neural message passing for visual relationship detection. In ICML Workshop on Learning and Reasoning with Graph-Structured Representations, Long Beach, CA, June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[65] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[66] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 409-424. Springer, 2022. 2, 4, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.515, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[67] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and TatSeng Chua. Visual translation embedding network for visual relation detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5532-5540, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.584, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[68] Ji Zhang, Yannis Kalantidis, Marcus Rohrbach, Manohar Paluri, Ahmed Elgammal, and Mohamed Elhoseiny. Large-scale visual relationship understanding. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9185-9194, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.71 + ], + "angle": 0, + "content": "[69] Sipeng Zheng, Shizhe Chen, and Qin Jin. Visual relation detection with multi-level attention. In Proceedings of the 27th ACM International Conference on Multimedia, pages 121-129, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.712, + 0.892, + 0.78 + ], + "angle": 0, + "content": "[70] Guangming Zhu, Liang Zhang, Youliang Jiang, Yixuan Dang, Haoran Hou, Peiyi Shen, Mingtao Feng, Xia Zhao, Qiguang Miao, Syed Afaq Ali Shah, et al. Scene graph generation: A comprehensive survey. arXiv preprint arXiv:2201.00443, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.782, + 0.892, + 0.837 + ], + "angle": 0, + "content": "[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable \\(\\{\\mathrm{detr}\\}\\): Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.839, + 0.892, + 0.894 + ], + "angle": 0, + "content": "[72] Yaohui Zhu, Shuqiang Jiang, and Xiangyang Li. Visual relationship detection with object spatial distribution. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 379-384. IEEE, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21648" + } + ] +] \ No newline at end of file diff --git a/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_origin.pdf b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cf648c3ee544483a7a263f733a30d6f5ae633de2 --- /dev/null +++ b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/14111a75-4cc9-4c7b-8b2c-bff17e2c80b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33e2a95d69e6cffe8a85e950b1a3349591384464929cb836705469eb43c30bbc +size 1404698 diff --git a/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/full.md b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..748c120450c2450fe3e45cc1ce66aa9854944782 --- /dev/null +++ b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/full.md @@ -0,0 +1,514 @@ +# HiLo: Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation + +Zijian Zhou + +Department of Informatics + +King's College London + +zijian.zhou@kcl.ac.uk + +Miaojing Shi + +College of Electronic and Information Engineering + +Tongji University + +mshi@tongji.edu.cn + +Holger Caesar + +Intelligent Vehicles Lab + +Delft University of Technology + +h.caesar@tudelft.nl + +# Abstract + +Panoptic Scene Graph generation (PSG) is a recently proposed task in image scene understanding that aims to segment the image and extract triplets of subjects, objects and their relations to build a scene graph. This task is particularly challenging for two reasons. First, it suffers from a long-tail problem in its relation categories, making naive biased methods more inclined to high-frequency relations. Existing unbiased methods tackle the long-tail problem by data/loss rebalancing to favor low-frequency relations. Second, a subject-object pair can have two or more semantically overlapping relations. While existing methods favor one over the other, our proposed HiLo framework lets different network branches specialize on low and high frequency relations, enforce their consistency and fuse the results. To the best of our knowledge we are the first to propose an explicitly unbiased PSG method. In extensive experiments we show that our HiLo framework achieves state-of-the-art results on the PSG task. We also apply our method to the Scene Graph Generation task that predicts boxes instead of masks and see improvements over all baseline methods. Code is available at https://github.com/franciszzj/HiLo. + +# 1. Introduction + +Scene Graph Generation (SGG) [49] is a crucial task in image scene understanding that extracts triplets in the form of subjects, objects and their relations to build a scene graph. Subjects and objects are represented with bounding boxes. Since this task links vision and text, it holds great potential for a variety of applications, including visual question answering [27], image captioning [21, 8], image retrieval [32, 53, 51] and visual reasoning [1, 55]. + +Recently a novel variant of SGG was proposed, which is Panoptic Scene Graph generation (PSG) [62]. Subjects + +![](images/0fadbdf098beab5eb95c9713497ea7eb46a1aa1e6e20906838d22af4f7395b5d.jpg) + +![](images/26b4006d1727886b770c936e64746194da33e5e442cce1f9422bbff5967c3e67.jpg) +(b) + +![](images/85fcc76f8260fec609d39d1efa6c7be0f9835d1a920f0c86a8b9f4748ae6a805.jpg) +(a) +(c) + +![](images/47698a3e5574e48d55fb89438b4df0461f2fe71c2d98da3561b7b81fb835af6e.jpg) +(d) +Figure 1. An example of the PSG task. We compare the predicted object relations of different approaches. a) An image and its panoptic segmentation. Two persons and a sports ball are shown in different colors. b) A biased method predicts mostly high frequency relations. c) An unbiased method predicts mostly low frequency relations. d) Our method predicts both low and high frequency relations, as well as more relations in total. + +and objects are represented with panoptic segmentation [35] masks. Unless stated otherwise, in this work we focus on PSG, since it is pixel-level accurate and also covers background classes and their relations with foreground objects. + +The performance of the PSG model is affected by a long-tail problem in its relation categories. For instance, relations such as over, in front of and holding occur tens of thousands of times in the PSG dataset [62], while others like swinging and kissing occur only a few dozen times. This severe class imbalance in the relation categories can lead to model predictions that are more inclined to high-frequency relations, which poses significant challenges to the application of panoptic scene graphs in real-world scenarios. + +Previous methods [12, 18, 39, 17] have often treated the long-tail problem of the PSG task as equivalent to the long-tail problem in object-centric tasks such as classification [15, 5, 29] or semantic segmentation [14, 2]. Consequently, + +these methods have employed re-balancing techniques to address class imbalance, either through re-sampling the data [39] or by using a class-balanced loss [33] that assigns different weights to different relation categories. + +In contrast, in the relation-centric PSG task, a subject-object pair can have multiple relations that exhibit relational semantic overlap, such as being partially or fully overlapping. For example, in Fig. 1, there are multiple relations between the boy and the sports ball, such as beside, looking at, playing and chasing. For regular biased models [61, 65, 60, 46], the results are dominated by high-frequency relations (beside, looking at). For specifically unbiased models [66, 62], the results are dominated by low-frequency relations (playing, chasing). However, since the low frequency relations can be more specific (e.g. on and standing on) or only partially overlapping with high frequency relations (e.g. looking at and chasing), it is crucial to include both to fully understand the image. We found that relational semantic overlap occurs in large numbers in the PSG dataset [62] and that current methods do not effectively address it. This is reflected in the increase in the category-averaged mean recall metric of unbiased methods, at the cost of the decrease in global recall (see Sec. 4.4). + +To address the long-tail problem of scene graphs under relational semantic overlap, we introduce the HiLo framework. This framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths with the help of two novel consistency loss functions. We apply our framework on top of a novel baseline. This baseline uses a recent transformer-based approach [10] for panoptic segmentation and adapts triplet queries [62] and masked attention [10] for the PSG task. In summary, we make the following contributions: + +- We identify the long-tail problem with relational semantic overlap in the PSG task and propose the HiLo framework to address this problem. The framework is general and can be applied to any PSG method. +- We propose a powerful and efficient one-stage end-to-end baseline. This baseline enhances the interaction between mask and relation prediction in the transformer decoder layer. +- We conduct extensive experiments to demonstrate the effectiveness of our framework and baseline. Our results outperform the state-of-the-art in both recall and mean recall on the PSG dataset and show systematic improvements on the VG dataset. + +# 2. Related Work + +# 2.1. Scene Graph Generation + +The Scene Graph Generation (SGG) [49] task plays a crucial role in connecting vision and language, and has received widespread attention in the computer vision commu + +nity. Many methods have been proposed to improve the performance of SGG, which can be classified into three categories [70]. The first is to introduce multi-modal information, such as appearance [52], space [72], depth [54], and segmentation [34]. The second is to introduce prior information and commonsense knowledge, such as statistical [3, 16, 65, 9] and language prior knowledge [49, 43, 68, 31, 20]. The third category involves designing different model structures, such as message passing [41, 16, 42, 65, 24, 64], attention mechanisms [69, 50], tree structures and visual translation [67, 30]. However, most of these methods are two-stage methods that cannot learn scene graphs end-to-end. In contrast, to improve the learning ability of SGG models, several methods based on transformers have been proposed, including SGTR [38], RelationFormer [56] and RelTR [13]. These are end-to-end trainable in a single stage. + +# 2.2. Unbiased Scene Graph Generation + +Solving the long-tail problem in the SGG task has attracted considerable attention from researchers, and several unbiased methods [59, 63, 12, 18, 25, 39, 19, 22, 37, 40, 17, 66] have been proposed. These methods typically improve the model from the perspective of data re-sampling [39] or a class-balanced loss [33]. BGNN [39] uses a two-layer resampling strategy to provide a more balanced data distribution during training, while CogTree [63] proposal exploits the semantic relation between different predicate classes to design a novel CogTree loss. HML [17] improves the model's ability to solve long-tail problems by designing a staged training process, and IETrans [66] proposes an internal and external transfer method to transfer high frequency relations to low frequency relations and recover missing relations to train the unbiased model. Dong et al. [19] propose to group relations by their frequency and train specialized relation encoders for each group. While these methods are able to mitigate the long-tail problem, they do not address relational semantic overlap. + +# 2.3. Panoptic Scene Graph Generation + +Contrary to SGG, the PSG task [62] uses panoptic segmentation masks instead of bounding boxes to represent objects, resulting in a more comprehensive scene graph. PSGTR [62], an end-to-end method based on the DETR structure, was proposed to construct a transformer-based PSG model. PSGFormer [62] further improved on PSGTR by introducing Object & Relation Query Learning Blocks and Query Matching Blocks. Their strong performance on most relation classes indicates that their method is implicitly unbiased. In contrast, our approach is the first to explicitly bias a PSG model, creating separate branches for low and high frequency relations, which are then fused together. + +Since the scene graph in PSG is built upon the sub + +![](images/d504548dc5565807a602bd6bcf021632dae67a9e444c9a18b35d5e033c982c02.jpg) +Figure 2. An overview of our HiLo framework with HiLo baseline. a) HiLo relation swapping module swaps the multiple relations in the subject-object pair to obtain H-L Data and L-H Data respectively. b) Input data into our HiLo framework with HiLo baseline model, there are two branches, namely H-L decoder and L-H decoder, which learn H-L Data and L-H Data respectively. c) In addition to task losses for PSG, we propose HiLo prediction alignment, which includes subject-object consistency loss and relation consistency loss, so that the parallel branch can be better optimized. + +jects, objects and their relations, strong panoptic segmentation [35] is crucial for PSG. Several DETR-based [6] methods such as Deformable-DETR [71], Segmenter [57], MaskFormer [11] and Mask2Former [10] have recently pushed the envelope in panoptic segmentation. These methods have introduced deformable transformer encoders and decoders, as well as pixel decoders and multi-scale information to improve model performance and speed up convergence. Our proposed baseline builds upon these techniques to further enhance the PSG performance and speed up model convergence. + +# 3. Method + +# 3.1. Problem Setting + +The Panoptic Scene Graph generation (PSG) task aims to generate a panoptic scene graph $\mathcal{G}$ for a given image $\mathcal{I} \in \mathbb{R}^{H \times W \times 3}$ , where $\mathcal{G}$ contains an object set $\mathcal{O}$ and a relational triplet set $\mathcal{T}$ , denoted by $\mathcal{G} = (\mathcal{O}, \mathcal{T})$ . For the $i$ -th object $o_i$ in $\mathcal{O} = \{o_i\}_{i=1}^N$ , we use $m_i$ to represent the object's mask and $c_i$ the object's category, i.e. $o_i = (m_i, c_i)$ . For the $j$ -th triplet in $\mathcal{T} = \{t_j\}_{j=1}^M$ , we use $s_j$ to represent the subject, $o_j$ the object, and $r_j$ their relation, i.e. $t_j = (s_j, o_j, r_j)$ . There are in total $C$ object categories and $R$ relation categories. For the $k$ -th relation, we denote its frequency in the training set as $f_k$ . + +# 3.2. HiLo Baseline + +High-quality panoptic segmentation is crucial for achieving good PSG performance. We build our method upon the latest advances in DETR-based panoptic segmentation, Mask2Former [10]. Below we present its structure, as well as our proposed modifications for the PSG task. + +Mask2Former. This method comprises three key parts: A backbone (CNN-based or transformer-based) followed by a pixel decoder, a transformer module with a transformer decoder and a task-specific module with different task heads. Specifically, the backbone takes an input image $\mathcal{I}$ and generates an image feature $F$ . The pixel decoder then gradually upsamples $F$ to produce multi-scale features $\widetilde{\mathcal{F}} = \{\widetilde{F}_i\}_{i=1}^4$ . The transformer decoder takes a set of queries $\mathcal{Q}$ of size $N$ and multi-scale features $\widetilde{\mathcal{F}}$ as input and outputs a set of mask features $\mathcal{X}$ of the same size with $\mathcal{Q}$ . + +On top of the transformer decoder, there are two task heads including a linear classifier that predicts the class probability for each mask; and a multi-layer perceptron (MLP) that uses the mask features $\mathcal{X}$ to generate the mask embedding $\mathcal{E}$ . The mask prediction is obtained by taking the dot product of the mask embedding $\mathcal{E}$ with the scale feature of the highest resolution in the multi-scale features $\widetilde{\mathcal{F}}$ . + +Triplet queries. The original query in Mask2Former is to predict the object. In order to predict both subject, object + +and their relation, we develop the triplet queries $\mathcal{Q}^t$ inspired by [62] into our baseline. Each query predicts a triplet that includes subject, object, and relation. Accordingly, our task heads comprise three linear classifiers. Each classifier is responsible for predicting the class probability for subject, object or relation, respectively. Moreover, we devise two MLPs to generate mask embeddings for the subject and object, denoted by $\mathcal{E}^s$ and $\mathcal{E}^o$ . We use them to compute the dot product with the scale feature of the highest resolution in multi-scale feature $\widetilde{\mathcal{F}}$ and obtain the mask prediction for the subject and object, respectively. + +Masked relation attention. [10] proposes a transformer decoder variant with masked attention, which extracts spatial features by adding the predicted mask of the object from the previous decoder layer's mask prediction. It makes the model focus on the object-related area in the feature map. To adapt this scheme to the PSG task, for each relation, we extend the masked attention to take the union of the binary masks of the subject and object as input, which represents the pixels corresponding to the relation. + +Network training. We adopt the same losses as PS-GTR [62], including cross-entropy loss $\mathcal{L}_{so\_cls}$ for object classification of subject-object pairs and a combination $\mathcal{L}_{so\_mask}$ of focal loss [44] and dice loss [58] to jointly supervise mask learning. To supervise the relations, we use the cross-entropy loss $\mathcal{L}_{rel\_cls}$ . The baseline loss with $(\lambda_1 = 1, \lambda_2 = 1, \lambda_3 = 4)$ is thus: + +$$ +\mathcal {L} _ {\text {b a s e l i n e}} = \lambda_ {1} \cdot \mathcal {L} _ {\text {s o . c l s}} + \lambda_ {2} \cdot \mathcal {L} _ {\text {s o . m a s k}} + \lambda_ {3} \cdot \mathcal {L} _ {\text {r e l . c l s}} \tag {1} +$$ + +# 3.3. HiLo Framework + +The key insight of our HiLo is to build a model that can take into account both high frequency and low frequency relations, and effectively improve the performance on low frequency relations without degrading the performance on high frequency relations. + +# 3.3.1 HiLo relation generation + +In the PSG task, multiple relations can be used to describe the connection between a subject and an object from different perspectives, such as spatial relations, actions and prepositions. Since the subject, object and their relative position in the image are fixed, all these relations share the same visual information. This is reflected in the PSG dataset [62] where many subject-object pairs are annotated with multiple relation labels. These relation labels are of different frequencies in the dataset. In this section we introduce the HiLo relation generation module to prepare two sets of training data, biasing towards high and low-frequency relations respectively. + +Relation augmentation. Similar to [22], we observe that there are many missing relation annotations in the PSG + +dataset [62]. To add the missing relation annotations, we design a relation augmentation scheme that is inspired by IETrans [66], which converts high-frequency relations to low-frequency relations and adds a relation to a subject-object pair that has no relation. We adapt it by first training our baseline as a biased model using the original annotated data. For every subject-object pair in the training set, we use this model to predict the relation scores for all predefined relation categories: + +- If this subject-object pair has annotated relation labels, we pick the one with the highest predicted score and use the score as a threshold. For other relations (except for already labeled ones) whose predicted scores are greater than the threshold, we add them as the relation labels of this subject-object pair. +- If this subject-object pair has no annotated relation labels, we use the predicted score of the no-relation class as a threshold. For other relations whose predicted scores are greater than this threshold we add them as the relation labels of this subject-object pair. + +This operation allows us to significantly augment the relation labels for the subject-object pairs in the training data, which is especially relevant for pairs with zero or no annotated relations. + +Relation swapping. We swap the relation labels for each subject-object pair. Specifically, given a subject-object pair $(s,o)$ with $K$ relation labels, we have $K$ triplets $(s,o,r_1),\ldots ,(s,o,r_K)$ , sorted by their relation frequency in descending order, $f_{1} > \ldots >f_{K}$ . We denote by H-L and L-H the swapping of high-frequency relations with low-frequency relations and vice versa. This creates two sets of data: + +H-L Data. Given a triplet $(s, o, r_k)$ , we replace its relation label $r_k$ with that of the next triplet $r_{k+1}$ with lower frequency. We sequentially process all triplets from $(s, o, r_1)$ until $(s, o, r_{K-1})$ , keeping the last triplet unchanged. + +L-H Data. Given a triplet $(s,o,r_k)$ , we replace its relation label $r_k$ with that of the previous triplet $r_{k - 1}$ with higher frequency. We sequentially process all triplets from $(s,o,r_K)$ until $(s,o,r_2)$ , keeping the first triplet unchanged. Hence, we obtain two new sets of triplets, denoted by $\mathcal{T}^{\mathrm{H - L}}$ and $\mathcal{T}^{\mathrm{L - H}}$ . We devise two parallel decoders from the shared encoder of our backbone. They are learned with the H-L and L-H data, respectively. The H-L and L-H decoders favour the predictions for low and high-frequency relations, respectively. Despite their difference, for the corresponding triplet query in the two decoders, their predictions are highly correlated: on one hand, the subject and object predictions should be the same; on the other hand, the distribution of relation predictions should be overlapping. Below we first build the query correspondence in the two decoders and then introduce the HiLo subject-object and relation consistency loss to align the predictions from two the decoders. + +# 3.3.2 HiLo prediction alignment + +Training two different relation distributions simultaneously confuse the model. To make the model differentiate between the two branches, we built a HiLo prediction alignment module, including triplet query correspondence, subject-object consistency loss and relation consistency loss respectively. + +Triplet query correspondence. In order to construct the subject-object consistency loss and relation consistency loss, we first need to construct the correspondence between the triplet queries in the H-L and L-H decoders. The same query index in the two decoders is not naturally matched. In order to find the query correspondence, we need to rely on the ground truth assignment: we use Hungarian matching to assign the triplet label to the corresponding triplet query, and record the relation label index corresponding to the triplet query. This label index allows us to construct the correspondence between the triplet queries in both decoders. We calculate the consistency loss for the triplet query prediction with the same relation label index in the two decoders. + +Subject-object consistency loss. Having the corresponding predictions from the two decoders, both their subjects and objects have the same ground truth and should be equal. We propose a subject-object consistency loss $\mathcal{L}_{obj}$ to minimize the mean squared error (MSE) of the corresponding predictions from the two decoders. + +$$ +\mathcal {L} _ {o b j} = \mathcal {L} _ {c l s} + \mathcal {L} _ {m a s k} \tag {2} +$$ + +$$ +\mathcal {L} _ {c l s} = \left\| \operatorname {s o f t m a x} \left(p _ {c} ^ {\mathrm {H} - \mathrm {L}}\right) - \operatorname {s o f t m a x} \left(p _ {c} ^ {\mathrm {L} - \mathrm {H}}\right) \right\| ^ {2} \tag {3} +$$ + +$$ +\mathcal {L} _ {\text {m a s k}} = \left\| \operatorname {s i g m o i d} \left(p _ {m} ^ {\mathrm {H} - \mathrm {L}}\right) - \operatorname {s i g m o i d} \left(p _ {m} ^ {\mathrm {L} - \mathrm {H}}\right) \right\| ^ {2} \tag {4} +$$ + +Here $\mathcal{L}_{cls}$ and $\mathcal{L}_{mask}$ represent the MSE losses for class prediction and mask prediction, respectively. + +$p_c^{\mathrm{H - L}}$ and $p_c^{\mathrm{L - H}}$ are the class prediction logits from the H-L and L-H decoders. After a softmax, they are $C$ -dimensional probability vectors. After a sigmoid, $p_m^{H - L}$ and $p_m^{L - H}$ are the mask prediction logits for the H-L and L-H decoders. + +Relation consistency loss. Given a pair of subject and object, we have previously swapped the high-low frequency relation labels to create data for the H-L and L-H decoders. For H-L, the prediction of the low-frequency relation logit is of high value; while for L-H, the prediction of the high-frequency relation logit is of high value. For the predictions on the rest logits, they should be similar, since it is the same pair of subject and object for the two decoders. Based on this observation, we introduce the relation consistency loss. + +Specifically, for a pair of subject and object, we use $p_r^{\mathrm{H - L}}$ to denote the predicted relation logits from the H-L decoder, $p_r^{\mathrm{L - H}}$ the predicted relation logits from the L-H decoder. They are $R$ -dimensional probability vectors after softmax. We can map between the distributions of the two vectors by swapping the logit values between the high- and + +low-frequency relation indices, which is named Relational Index Exchange operation. We use $\mathrm{RIE}(\cdot)$ to denote this operation. $\mathrm{RIE}(\cdot)$ includes a stop gradient operation, which creates copies from original predictions of two branches, enabling value exchanges. For example, for relations $r_k$ and $r_{k + 1}$ , $\mathrm{RIE}(\cdot)$ exchanges the values between $p_{r,k}$ and $p_{r,k + 1}$ in $p_r$ . We can therefore compute the distance between $p_r^{\mathrm{H - L}}$ and its mapped counterpart from $p_r^{\mathrm{L - H}}$ , vice versa: + +$$ +\begin{array}{l} \operatorname {D i s t} _ {\mathrm {H i L o}} = \left\| \text {s o f t m a x} \left(p _ {r} ^ {\mathrm {H - L}}\right) - R I E \left(\text {s o f t m a x} \left(p _ {r} ^ {\mathrm {L - H}}\right)\right) \right\| ^ {2} \\ + \| R I E \left(\operatorname {s o f t m a x} \left(p _ {r} ^ {\mathrm {H} - \mathrm {L}}\right)\right) - \operatorname {s o f t m a x} \left(p _ {r} ^ {\mathrm {L} - \mathrm {H}}\right) \| ^ {2} \tag {5} \\ \end{array} +$$ + +The relation consistency loss is defined to minimize $\mathrm{Dist}_{\mathrm{HiLo}}$ with a margin of $m$ : + +$$ +\mathcal {L} _ {r e l} = \max \left(\operatorname {D i s t} _ {\mathrm {H i L o}} - m, 0\right), \tag {6} +$$ + +where $m$ is a small constant. Adding $m$ is due to the fact that the high- and low frequency relations might be only partially semantically overlapping. + +Network training. Our subject-object consistency loss and relation consistency loss can seamlessly integrate with the losses of any baseline method to jointly supervise the training of the entire model. Notably, we supervise the output of each transformer decoder layer to ensure effective learning. The final loss $\mathcal{L}$ is thus: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {b a s e l i n e}} + \mathcal {L} _ {\text {o b j}} + \mathcal {L} _ {\text {r e l}} \tag {7} +$$ + +# 3.3.3 HiLo inference fusion + +The H-L and L-H decoders favour low-frequency relation prediction and high-frequency relation prediction, respectively. To combine the strength of both during inference, we introduce the HiLo inference fusion module. Specifically, we denote by $\mathcal{G}^{\mathrm{H - L}}$ and $\mathcal{G}^{\mathrm{L - H}}$ the predicted panoptic scene graphs from H-L and L-H decoders, respectively. There are $N_{1}$ triplets in $\mathcal{G}^{\mathrm{H - L}}$ and $N_{2}$ triplets in $\mathcal{G}^{\mathrm{L - H}}$ . + +- First, we merge the triplets in $\mathcal{G}^{\mathrm{H - L}}$ and $\mathcal{G}^{\mathrm{L - H}}$ and sort them in the descending order according to their relation scores. We obtain a list of $N_{1} + N_{2}$ triplets. +- Second, starting from the first triplet, we de-duplicate the triplet list. For the $i$ -th triplet $\mathcal{T}_i$ , we identify its duplicated versions from the $(i + 1)$ -th triplet until the end of the list. Remove any follow-up triplet from the list if it 1) has the same subject, object and relation classes to that in $T_i$ and 2) has a mask IoU greater than a threshold, e.g. 0.5, between the predicted subject/object and the corresponding subject/object in $T_i$ . +- Third, after dedduplication, for each triplet in the list, we multiply the relation, subject and object scores as an overall score for it. We sort the triplet list according to this score in descending order to obtain the final panoptic scene graph $\mathcal{G}^{\mathrm{HiLo}}$ . + +
MethodBackboneScene Graph Detection
R@20mR@20R@50mR@50R@100mR@100
IMP [61]R5016.56.518.27.118.67.2
MOTIF [65]R5020.09.121.79.622.09.7
VCTree [60]R5020.69.722.110.222.510.2
GPSNet [46]R5017.87.019.67.520.17.7
PSGTR [62]R5028.416.634.420.836.322.1
PSGFormer [62]R5018.014.819.617.020.117.6
HiLo (ours)R5034.123.740.730.343.033.1
HiLo (ours)Swin-B38.528.346.235.349.639.1
HiLo (ours)Swin-L40.629.748.737.651.440.9
+ +Table 1. Comparison between our HiLo and other methods on the PSG dataset. Our method shows superior performance compared to all previous methods. + +# 4. Experiments + +# 4.1. Datasets + +Panoptic Scene Graph Generation (PSG) [62]. This is the first Panoptic Scene Graph generation dataset. It has a total of 48,749 labeled images including 2,177 test images and 46,572 training images. The object categories comprise 80 thing classes and 53 stuff classes, which is the same as the COCO [45] and COCO-Stuff datasets [4]. The relation categories comprise 56 classes, including positional relations, common object-object relations, common actions, human actions, actions in the traffic scene, actions in the sports scene and interactions between backgrounds [62]. + +Visual Genome (VG) [36]. VG is a widely used benchmark dataset for Scene Graph Generation. Following previous work [65, 7], we adopt the widely accepted split, VG-150, which contains 150 object categories and 50 relation categories. The object categories cover a wide range of classes, such as animals, vehicles and household items. The relation categories include both spatial and semantic classes, such as on, in and wearing. + +# 4.2. Tasks and Metrics + +Three subtasks have been proposed for the SGG and PSG tasks, which are Predicate Classification, Scene Graph Classification and Scene Graph Detection [61]. We focus on Scene Graph Detection for both datasets, since it is the most comprehensive and addressed by [62]. This subtask requires the model to first localize the objects and then predict the object classes and relations. Note that Scene Graph Detection in PSG includes the detection on stuff classes, while in SGG it does not. + +Following previous work [60, 63, 62], we use Recall@K (R@K) and mean Recall@K (mR@K) as our metrics, where the former metric is dominated by high-frequency relations, while the latter assigns equal weight to all relation classes. + +# 4.3. Implementation details + +In our experiments, we follow the training strategy of PSGTR [62]. We use the AdamW optimizer [48], with a learning rate of $1e^{-4}$ and weight decay of $1e^{-4}$ , except for the backbone, which is trained with a learning rate of $1e^{-5}$ . For initialization, we use Mask2Former [10] pretrained on COCO [45] to initialize our backbone and pixel decoder. Following Mask2Former [10], we use 100 triplet queries for the H-L and L-H decoders respectively. Additionally, both the H-L and L-H decoders are initialized with Mask2Former's transformer decoder. To ensure consistent comparison with PSGTR, we adopted the same data augmentation settings. Our model is trained for 12 epochs with a step scheduler at epoch 10, taking approximately 18 hours to train on four A100 GPUs with a batch size of 1 for each GPU. + +# 4.4. Comparison to the state-of-the-art + +PSG. Tab. 1 reports the performance of our method compared to the state-of-the-art on the PSG dataset [62]. We separate the methods into two groups. The first are two-stage methods consisting of a separate segmentor and relation predictor, which are modified for the PSG task in [62]. The second are one-stage end-to-end methods, which are able to simultaneously predict panoptic segmentation and relations. Our method belongs to the second category. For a fair comparison between the different methods, we use the same Resnet-50 [26] backbone. Our method shows superior performance compared to all previous methods. Particularly, it outperforms the previous best-performing method PSGTR [62] by a large margin, i.e. $+6\%$ in R@100 and $+11\%$ in mR@100. Our model is able to converge within only 12 epochs of training, whereas PSGTR [62] is trained for 60 epochs. We also evaluate our method using pre-trained transformer-based backbones, i.e. Swin-B and Swin-L [47], with the latter being a bigger model. Our results are consistently improved over all metrics due the powerful feature representation ability of the transformer. We also conducted a visual comparison, as shown in Fig. 3. + +
MethodScene Graph Detection
R/mR@50R/mR@100
MOTIF [65]31.0 / 6.735.1 / 7.7
+IETrans [66]26.4 / 12.430.6 / 14.9
+HiLo (ours)26.2 / 14.730.3 / 17.7
VCTree [60]30.2 / 6.734.6 / 8.0
+IETrans [66]25.4 / 11.529.3 / 14.0
+HiLo (ours)27.1 / 12.929.8 / 15.2
Transformer [59]30.0 / 7.434.3 / 8.8
+IETrans [66]25.5 / 12.529.6 / 15.0
+HiLo (ours)25.4 / 14.629.2 / 17.6
GPSNet [46]30.3 / 5.935.0 / 7.1
+IETrans [66]25.9 / 14.628.1 / 16.5
+HiLo (ours)25.6 / 15.827.9 / 18.0
+ +SGG. Here we study whether our approach that was developed for the PSG task can also be applied to the SGG task. In Tab. 2, we conduct experiments on the VG-150 dataset. Following IETrans [66], we extend our unbiased HiLo framework (Sec. 3.3) without our PSG-specific baseline (Sec. 3.2) to four state-of-the-art biased SGG methods, namely MOTIF [65], VCTree [60], Transformer [59] and GPSNet [46], using the implementation of [66]. Compared to the unbiased IETrans [66] method, our method improves mean recall without sacrificing recall, demonstrating its effectiveness in enhancing the performance of low-frequency relations. Compared to the biased baselines [65, 60, 59, 46], our method achieves a significantly larger mean recall, while still maintaining an acceptable recall. This indicates that our method can improve the performance of low-frequency relations while also taking into account the performance of high-frequency relations. It also shows that the HiLo framework is a general technique that yields systematic improvements in both the PSG and SGG tasks. + +# 4.5. Ablation Studies + +Consistent with the paper, we use HiLo with a Resnet-50 backbone to perform ablation experiments on the PSG dataset. + +HiLo framework for different baselines. In this section we investigate whether our HiLo framework (Sec. 3.3) yields improvements for other baselines, rather than just the one presented in Sec. 3.2. Tab. 3 shows the results for two baselines, with and without the HiLo framework. We observe that our biased baseline outperforms the previous PSGTR [62] method on all metrics. Furthermore, by applying the HiLo framework, we can substantially improve the performance over both baselines. It is worth mentioning that the HiLo framework improves recall and mean recall simultaneously, whereas other methods typically improve one metric at the cost of the other [59, 63]. + +HiLo relation augmentation. We observe that out of + +Table 2. Comparison between our HiLo framework and other methods on the VG-150 dataset. Similar to [66], we apply IETrans and our own method on top of four leading baselines. + +
BaselineHiLoR/mR@20R/mR@50R/mR@100
HiLo baseline✓34.1 / 23.740.7 / 30.343.0 / 33.1
HiLo baseline-32.6 / 20.938.0 / 27.438.9 / 28.4
PSGTR [62]✓30.1 / 20.236.6 / 23.938.3 / 24.5
PSGTR [62]-28.4 / 16.634.4 / 20.836.3 / 22.1
+ +Table 3. Comparison of different baselines, with and without HiLo framework. Using the HiLo framework, we see significant improvements on both metrics. + +
Relation Aug.Multiple relationsR/mR@50R/mR@100
✓40%40.7 / 30.343.0 / 33.1
-10%40.1 / 28.142.8 / 32.5
+ +Table 4. Ablation study for HiLo relation augmentation. Relation augmentation affects the ratio of subject-object pairs with multiple relations. The larger this ratio, the more relations can be swapped, which leads to better results. + +260,296 labeled triplets in the PSG dataset, only about $10\%$ of subject-object pairs have multiple relations, for which we can apply relation swapping (Sec. 3.3.1). After applying our proposed relation augmentation technique (Sec. 3.3.1), this ratio significantly increases to $40\%$ . Our experimental results in Tab. 4 demonstrate that only applying the HiLo framework on $10\%$ already gives an improvement over the baseline from Tab. 3. As the number of swappable triplets increases due to augmentation, the model's performance is further enhanced, highlighting the potential of our method. + +HiLo prediction alignment. We conduct ablation experiments on the subject-object consistency loss and relation consistency loss (Sec. 3.3.2), which are used to align the predictions from the high and low frequency branches. The results, as presented in Tab. 5, demonstrate that using both losses yields the best performance. It is worth mentioning that we have explored the margin in the relation consistency loss and found that setting the margin to zero leads to a small performance degradation. This finding confirms that there is a partial semantic overlap between swapped relations, indicating that they are not entirely consistent. + +To investigate the impact of relational index exchange (RIE) on the relation consistency loss, we conducted experiments to verify the effect of omitting RIE. In the absence of RIE, we solely compute the consistency loss for relation categories that are not involved in relation swapping and exclude the swapping component from the calculation of the consistency loss. The outcomes of this experiment are presented in Tab. 6, and demonstrate a notable reduction in the mean recall and a decline in the model's performance for relations with relational semantic overlap when RIE is not utilized. + +HiLo inference fusion. We ablate the inference fusion (Sec. 3.3.3) and evaluate the performance of each branch's output separately. In Tab. 7, experimental results suggest that fusion can effectively leverage the uniqueness of the high and low frequency branch predictions to achieve comprehensive improvements on all metrics. + +![](images/7fa600f28fd3137b58e0fa7f7db26250b9afb8bf2fcb8f7b65c7781e73616aab.jpg) +↑ Image + +![](images/3a9f8d2eb981294c28102f3ab364f95e94d73dc3233d2f7fd5d30f65b93a9a82.jpg) +Ground Truth +person walking on sea +surfboard over sea +sky over surf +person attached to surfboard +surfboard attached to person +surfboard on person + +![](images/b3646b8e1b5dbaaeeb361e9b91848002f2fbdce0140cfd6ca683a2f080d46bb7.jpg) +person carrying surfboard +looking at +sky over person +sky over sea + +![](images/c603cee51a1a8b18c820cde2034ae09fa7a704cb872df026ca4d63d2ac96985e.jpg) +Prediction $\rightarrow$ +person holding surfboard + +![](images/6e1a1c82d1b12bb4fcd21ef016055f53c023d0542011c0dee25c22fa2fcba4f8.jpg) +sky over sea +person standing on sea +# +sky over person + +![](images/b7f1aa76cca772df169569b9f3788fe4095be02c539bafa96ca2e7cbbcd0e0ba.jpg) +person on sea +person carrying surfboard +looking at you + +![](images/afd8823e7c5b5591c040c7eb8afaf2e9d9763e319fa5a06197d26292e32f5efe.jpg) +sea in front of person +sea beside person +person playing surfboard +surfboard on sea + +![](images/a04fcf96662337ab4ff31d31471d73864e5cfd71636071d51a96174372a56e2b.jpg) +↑ +Image +Prediction→ + +![](images/ce68394d186232dd3e2c1e3f79ce71c568dc0aa633e68981525fff37f793a0ce.jpg) +Ground Truth +1 +sky over snow +sky over person +person standing on snow +sky over skis +sky over backpack +sky over rock +person climbing snow +person walking on snow +personbesideperson +person on snow +person running on snow +person picking skis +person playing skills +person playing with person attached to skis +person attached to backpack +mountain attached +sky over mountain +# + +![](images/27e21a5518eb23cf7104d3a1cfe7470c4299507d893ece5625e4e1a14f376a3b.jpg) +person carrying backpack +person holding skis +person climbing snow. +sky over snow. + +![](images/b72dddcb15b5e69afbb1e14f07a0b77555faa439c1b6be332ba58fc38efda690.jpg) +person carrying backpack +person holding +skis on : + +![](images/3ba5ceb2c1ff682fe10946d74229fb33427e75719f3bb82142b829b78e8bf237.jpg) +↑ +Image +Prediction $\rightarrow$ + +![](images/bbcb12245522c272c26d46ee07348c4652472801a13605622700bc61e93907c1.jpg) +Ground Truth +1 +chair beside wall +chair in front of +person in front of chair +person looking at +person throwing +wall in front of tree +over chair +person swinging tennis racket +treebesidewall +chair beside person +chair attached to wall +person holding sports ball + +![](images/84a9ce5306300de517c32d31386355b13b59521e84e6b23a884c79dc5296d51e.jpg) +mehohid +person beside barrier +wall beside tree +person playing sports ball + +![](images/09ae112509d1f1e09e86d6acf0847d76b3b96580a4c7206e49724a8b5642d9ac.jpg) + +![](images/7c3d0874c90c90775a6ba793c8315b1e8f848e82cf2f9303f20d2b161763aaa0.jpg) + +![](images/da702fde1726af0638b1d5d7dff8d28d1df8af4307319e948810c447889d64f2.jpg) +tree in front of + +![](images/fd226a19c1fb0ca712398d39c7fb7114f94d2d1d042de71a44bda5b97bde5f41.jpg) +chair leaning on wall +tree over wall +Figure 3. Visualization of panoptic segmentations and the top 20 predicted triplets compared with ground truth. The upper left is the original image, the lower left is the ground truth and on the right are the predictions. The highlighted triplets represent the subject-object pairs with multiple relations, where the blue highlights represent the high frequency relations and the red highlights represent the low frequency relations. The visualization results show that our method can predict both high frequency and low frequency relations. + +![](images/869b41a2d909c1c603bae50fa873afa3ce9122612b09b7cafee5211987245d8c.jpg) +holding +about to +person in front of wall + +
ObjectRelationMarginR/mR@50R/mR@100
✓✓0.540.7 / 30.343.0 / 33.1
-✓0.540.6 / 29.742.8 / 32.8
-✓0.040.5 / 29.542.7 / 32.8
✓--40.4 / 29.042.8 / 32.2
---39.7 / 28.642.4 / 32.0
+ +Table 5. Ablation study for different losses in HiLo prediction alignment. Object refers to the subject-object consistency loss and relation refers to the relation consistency loss. The margin parameter is defined in Eq. 6. + +
Whether to use RIER/mR@20R/mR@50R/mR@100
✓34.1 / 23.740.7 / 30.343.0 / 33.1
-33.5 / 22.340.3 / 29.042.6 / 32.3
+ +We also attempted to average the tensor generated by the two branches and obtain the PSG result through postprocessing. However, we found that this approach leads to a substantial drop in performance, as evident in Tab. 8. This can be attributed to the inconsistent prediction results of the two branches for the same query index. These findings validate that the inference fusion method effectively merges the results from the two branches. Furthermore, our experimental results demonstrate that the query associated with the identical index in two branches does not predict the same subject-object pair. Thus, directly averaging the tensor produced by the two branches results in prediction ambiguity, ultimately leading to a substantial decline in performance. This observation underscores the necessity of conducting triplet query correspondence when performing prediction alignment. In particular, due to the inconsistent query prediction content for the corresponding index + +Table 6. Ablation study for relation consistency loss in HiLo prediction alignment. + +
H-L ResultL-H ResultR/mR@50R/mR@100
✓✓40.7 / 30.343.0 / 33.1
✓-38.8 / 29.939.8 / 30.9
-✓38.5 / 26.539.5 / 27.8
+ +Table 7. Ablation study for HiLo inference fusion. + +
Fusion methodR/mR@20R/mR@50R/mR@100
inference fusion34.1 / 23.740.7 / 30.343.0 / 33.1
average tensor19.6 / 13.123.1 / 15.723.9 / 16.3
+ +Table 8. Ablation study for HiLo inference fusion. Inference fusion refers to the method proposed in the paper to fuse the results of two branches. Average tensor refers to the fusion method that directly averages the tensor output by the two branches. + +
Attention focusR/mR@20R/mR@50R/mR@100
subject-object32.6 / 20.938.0 / 27.438.9 / 28.4
full image30.4 / 19.336.5 / 25.937.1 / 26.8
+ +Table 9. Ablation study for masked relation attention. + +in the two branches, a one-to-one correspondence must be constructed based on the label assigned by each query to achieve prediction alignment. + +Masked relation attention. We investigate the impact of different mask input types for cross-attention on the HiLo baseline (Sec. 3.2) performance. Specifically, we compare two different attention focus regions, namely the subject-object region and the full image. The results are shown in Tab. 9. Focusing on the full image presents a more challenging optimization task for the model since no target region is specified. Consequently, we observe a drop of $1.8\%$ in R@100 and $1.6\%$ in mR@100. This shows that it is crucial to apply masked relation attention to the subject and object. + +
MethodAllRareAllOverlap
PSGTR22.16.222.123.5
HiLo (ours)33.1 (+11.0)20.3 (+14.1)33.1 (+11.0)38.8 (+15.3)
+ +![](images/d374e4df8be64179687aad0660ee28d0f4f8dcb574a20ef3b7d20b31074ce94a.jpg) +Figure 4. Convergence speed analysis of different methods. Our method converges significantly faster than previous methods. + +# 4.6. Analysis + +To further demonstrate the efficacy of our method, we conduct the following analysis on HiLo. + +Long-tail problem and relational semantic overlap. To verify whether the problems of long-tail and relational semantic overlap have been tackled, we conduct experiments shown in Tab. 10. For the long-tail problem, we consider relations appearing less than 500 times (28 relations) in the PSG dataset (relations appear on average 4787 times) as rare relations, and report their mR@100. There is a $14\%$ improvement for rare relations and an $11\%$ improvement for all relations. This suggests that our method addresses the long-tail problem. + +For the relational semantic overlap, we select all test images that have this problem (927 images) and report their mR@100. Our method shows a $15.3\%$ improvement over PSGTR on images with semantic overlap relations, exceeding the overall improvement across all test images. This suggests that our approach addresses the problem of relational semantic overlap. + +Convergence speed and time cost analysis. We evaluate the convergence of our model by assessing its performance on the validation set at various epochs, as illustrated in Fig. 4. Our analysis reveals that our proposed method outperforms prior methods [62] both in terms of final performance and convergence speed. Specifically, PSGTR [62] achieves negligible performance in the initial 12 epochs, requiring 60 epochs to converge, as per the authors. In contrast, our HiLo method achieves better results in just 12 epochs, indicating its superior convergence speed. + +For time cost, we primarily analyze relation augmentation and swapping. 1) Relation augmentation, inspired by + +Table 10. Verification of the improvements in the long-tail problem and relational semantic overlap. All refers to testing on the whole test set. Rare refers to testing only on rare relations. Overlap refers to testing only on data with relation semantic overlap. + +
MethodGFlopsParam. (M)Train Mem. (G)Infer. Time (ms)
PSGTR461.344.226.5140
HiLo (ours)229.458.716.1156
+ +Table 11. Training and inference cost. + +IETrans [66], involves training a baseline model and then using it to predict relation labels so as to augment the original relation labels. For the PSG dataset, following our experimental setup (see Sec. 4.3), this takes 18 hours. Afterwards, the training of our HiLo model requires 18 hours, which makes the whole process 36 hours. In contrast, the PSGTR model training takes 48 hours. Both our method and PSGTR utilize ResNet-50 as the backbone for fair comparison. Our method's rapid convergence (see Sec. 4.5) reduces training time compared to PSGTR, making the additional time cost for relation augmentation tolerable. 2) Relation swapping is a quick operation on relation labels during training and does not significantly contribute to overall time consumption. + +Training and inference cost. Training and inference cost is shown in Tab. 11. Despite using two transformer-based decoders, our method requires less resource. Given the same input sizes (1280, 800) and ResNet-50 as backbone, the resource usage is shown in Tab. 11. PSGTR [62] generates mask features for subject and object separately, while our method reduces this computation by only generating mask features once. Our inference time is marginally higher due to more complex post-processing. Using H-L and L-H data only changes the labels between two branches, not increasing resource use. + +# 5. Conclusion + +In this work we proposed the HiLo framework to tackle the long-tail problem with relational semantic overlap in Panoptic Scene Graph generation. The HiLo framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths by aligning their predictions. We also constructed a HiLo baseline to allow high-quality panoptic segmentation to improve PSG performance. Experimental results demonstrate that our method achieves state-of-the-art performance on the PSG dataset, confirming its effectiveness. In future work, we will investigate how knowledge distillation [28, 23] can be used to fuse the high and low branches in our method, as well as its application to downstream tasks such as visual question answering and image captioning. + +# Acknowledgment + +The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the European Union's Horizon 2020 FET Proactive Program under Agreement 101017857 and Fundamental Research Funds for the Central Universities. + +# References + +[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 173:33-45, 2018. 1 +[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 353-369. Springer, 2022. 1 +[3] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2 +[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6 +[5] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3 +[7] Long Chen, Hanwang Zhang, Jun Xiao, Xiangnan He, Shiliang Pu, and Shih-Fu Chang. Counterfactual critic multiagent training for scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4613-4623, 2019. 6 +[8] Shizhe Chen, Qin Jin, Peng Wang, and Qi Wu. Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9962-9971, 2020. 1 +[9] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2 +[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 4, 6 +[11] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3 +[12] Meng-Jiun Chiou, Henghui Ding, Hanshu Yan, Changhu Wang, Roger Zimmermann, and Jiashi Feng. Recovering the + +unbiased scene graphs from the biased ones. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1581-1590, 2021. 1, 2 +[13] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[14] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 1 +[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1 +[16] Bo Dai, Yuqi Zhang, and Dahua Lin. Detecting visual relationships with deep relational networks. In Proceedings of the IEEE conference on computer vision and Pattern recognition, pages 3076-3086, 2017. 2 +[17] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 266-283. Springer, 2022. 1, 2 +[18] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 1, 2 +[19] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 2 +[20] Mohammed Haroon Dupty, Zhen Zhang, and Wee Sun Lee. Visual relationship detection with low rank non-negative tensor decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10737-10744, 2020. 2 +[21] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the 2018 10th international conference on machine learning and computing, pages 225–229, 2018. 1 +[22] Arushi Goel, Basura Fernando, Frank Keller, and Hakan Bilen. Not all relations are equal: Mining informative labels for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15596-15606, 2022. 2, 4 +[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 9 +[24] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 1969-1978, 2019. 2 +[25] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 2 +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[27] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1 +[28] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 9 +[29] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1 +[30] Zih-Siou Hung, Arun Mallya, and Svetlana Lazebnik. Contextual translation embedding for visual relationship detection and scene graph generation. IEEE transactions on pattern analysis and machine intelligence, 43(11):3820-3832, 2020. 2 +[31] Seong Jae Hwang, Sathya N Ravi, Zirui Tao, Hyunwoo J Kim, Maxwell D Collins, and Vikas Singh. Tensorize, factorize and regularize: Robust visual relationship learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1014-1023, 2018. 2 +[32] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1 +[33] Haeyong Kang and Chang D Yoo. Skew class-balanced reweighting for unbiased scene graph generation. Machine Learning and Knowledge Extraction, 5(1):287-303, 2023. 2 +[34] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15879-15889, 2021. 2 +[35] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 3 +[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6 +[37] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels: + +Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 2 +[38] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19486-19496, 2022. 2 +[39] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 2 +[40] Wei Li, Haiwei Zhang, Qijie Bai, Guoqing Zhao, Ning Jiang, and Xiaojie Yuan. Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2022. 2 +[41] Yikang Li, Wanli Ouyang, Xiaogang Wang, and Xiao'ou Tang. Vip-cnn: Visual phrase guided convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1347-1356, 2017. 2 +[42] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 2 +[43] Wentong Liao, Bodo Rosenhahn, Ling Shuai, and Michael Ying Yang. Natural language guided visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2 +[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 4 +[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[46] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2, 6, 7 +[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6 +[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[49] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In + +Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 1, 2 +[50] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2 +[51] Mengshi Qi, Yunhong Wang, and Annan Li. Online cross-modal scene retrieval by binary representation and semantic graph. In Proceedings of the 25th ACM international conference on Multimedia, pages 744-752, 2017. 1 +[52] Mohammad Amin Sadeghi and Ali Farhadi. Recognition using visual phrases. IEEE, 2011. 2 +[53] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1 +[54] Sahand Sharifzadeh, Sina Moayed Baharlou, Max Berrendorf, Rajat Koner, and Volker Tresp. Improving visual relation detection using depth maps. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 3597-3604. IEEE, 2021. 2 +[55] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1 +[56] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2 +[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 3 +[58] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 4 +[59] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 2, 7 +[60] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6, 7 + +[61] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 2, 6 +[62] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 2, 4, 6, 7, 9 +[63] Jing Yu, Yuan Chai, Yujing Wang, Yue Hu, and Qi Wu. Cogtree: Cognition tree loss for unbiased scene graph generation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1274–1280. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 2, 6, 7 +[64] Xu Chen Ya Zhang Xiao Gu Yue Hu, Siheng Chen. Neural message passing for visual relationship detection. In ICML Workshop on Learning and Reasoning with Graph-Structured Representations, Long Beach, CA, June 2019. 2 +[65] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7 +[66] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 409-424. Springer, 2022. 2, 4, 7, 9 +[67] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and TatSeng Chua. Visual translation embedding network for visual relation detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5532-5540, 2017. 2 +[68] Ji Zhang, Yannis Kalantidis, Marcus Rohrbach, Manohar Paluri, Ahmed Elgammal, and Mohamed Elhoseiny. Large-scale visual relationship understanding. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9185-9194, 2019. 2 +[69] Sipeng Zheng, Shizhe Chen, and Qin Jin. Visual relation detection with multi-level attention. In Proceedings of the 27th ACM International Conference on Multimedia, pages 121-129, 2019. 2 +[70] Guangming Zhu, Liang Zhang, Youliang Jiang, Yixuan Dang, Haoran Hou, Peiyi Shen, Mingtao Feng, Xia Zhao, Qiguang Miao, Syed Afaq Ali Shah, et al. Scene graph generation: A comprehensive survey. arXiv preprint arXiv:2201.00443, 2022. 2 +[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\{\mathrm{detr}\}$ : Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3 +[72] Yaohui Zhu, Shuqiang Jiang, and Xiangyang Li. Visual relationship detection with object spatial distribution. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 379-384. IEEE, 2017. 2 \ No newline at end of file diff --git a/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/images.zip b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b59244eeaf6e4e9d229d35f7274814a6f0af1997 --- /dev/null +++ b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db558c197cba193d22095c6a94b88eefa0bba2c7db1d76318120908c6554d927 +size 570449 diff --git a/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/layout.json b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..40349277382d8d2161643869a65380ff558805b0 --- /dev/null +++ b/2023/HiLo_ Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation/layout.json @@ -0,0 +1,14508 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 103, + 459, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 103, + 459, + 139 + ], + "spans": [ + { + "bbox": [ + 135, + 103, + 459, + 139 + ], + "type": "text", + "content": "HiLo: Exploiting High Low Frequency Relations for Unbiased Panoptic Scene Graph Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 162, + 153, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 162, + 153, + 175 + ], + "spans": [ + { + "bbox": [ + 93, + 162, + 153, + 175 + ], + "type": "text", + "content": "Zijian Zhou" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 178, + 172, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 178, + 172, + 189 + ], + "spans": [ + { + "bbox": [ + 74, + 178, + 172, + 189 + ], + "type": "text", + "content": "Department of Informatics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 191, + 167, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 191, + 167, + 202 + ], + "spans": [ + { + "bbox": [ + 80, + 191, + 167, + 202 + ], + "type": "text", + "content": "King's College London" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 206, + 181, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 206, + 181, + 217 + ], + "spans": [ + { + "bbox": [ + 66, + 206, + 181, + 217 + ], + "type": "text", + "content": "zijian.zhou@kcl.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 260, + 161, + 334, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 161, + 334, + 175 + ], + "spans": [ + { + "bbox": [ + 260, + 161, + 334, + 175 + ], + "type": "text", + "content": "Miaojing Shi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 204, + 178, + 390, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 178, + 390, + 190 + ], + "spans": [ + { + "bbox": [ + 204, + 178, + 390, + 190 + ], + "type": "text", + "content": "College of Electronic and Information Engineering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 264, + 191, + 329, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 191, + 329, + 203 + ], + "spans": [ + { + "bbox": [ + 264, + 191, + 329, + 203 + ], + "type": "text", + "content": "Tongji University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 247, + 206, + 346, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 206, + 346, + 217 + ], + "spans": [ + { + "bbox": [ + 247, + 206, + 346, + 217 + ], + "type": "text", + "content": "mshi@tongji.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 435, + 162, + 506, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 435, + 162, + 506, + 175 + ], + "type": "text", + "content": "Holger Caesar" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 427, + 178, + 515, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 178, + 515, + 189 + ], + "spans": [ + { + "bbox": [ + 427, + 178, + 515, + 189 + ], + "type": "text", + "content": "Intelligent Vehicles Lab" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 413, + 191, + 528, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 191, + 528, + 203 + ], + "spans": [ + { + "bbox": [ + 413, + 191, + 528, + 203 + ], + "type": "text", + "content": "Delft University of Technology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 417, + 206, + 522, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 206, + 522, + 216 + ], + "spans": [ + { + "bbox": [ + 417, + 206, + 522, + 216 + ], + "type": "text", + "content": "h.caesar@tudelft.nl" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 244, + 191, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 191, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 191, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 270, + 290, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 534 + ], + "type": "text", + "content": "Panoptic Scene Graph generation (PSG) is a recently proposed task in image scene understanding that aims to segment the image and extract triplets of subjects, objects and their relations to build a scene graph. This task is particularly challenging for two reasons. First, it suffers from a long-tail problem in its relation categories, making naive biased methods more inclined to high-frequency relations. Existing unbiased methods tackle the long-tail problem by data/loss rebalancing to favor low-frequency relations. Second, a subject-object pair can have two or more semantically overlapping relations. While existing methods favor one over the other, our proposed HiLo framework lets different network branches specialize on low and high frequency relations, enforce their consistency and fuse the results. To the best of our knowledge we are the first to propose an explicitly unbiased PSG method. In extensive experiments we show that our HiLo framework achieves state-of-the-art results on the PSG task. We also apply our method to the Scene Graph Generation task that predicts boxes instead of masks and see improvements over all baseline methods. Code is available at https://github.com/franciszzj/HiLo." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 575, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 671 + ], + "type": "text", + "content": "Scene Graph Generation (SGG) [49] is a crucial task in image scene understanding that extracts triplets in the form of subjects, objects and their relations to build a scene graph. Subjects and objects are represented with bounding boxes. Since this task links vision and text, it holds great potential for a variety of applications, including visual question answering [27], image captioning [21, 8], image retrieval [32, 53, 51] and visual reasoning [1, 55]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 671, + 287, + 696 + ], + "type": "text", + "content": "Recently a novel variant of SGG was proposed, which is Panoptic Scene Graph generation (PSG) [62]. Subjects" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 373, + 245, + 481, + 317 + ], + "blocks": [ + { + "bbox": [ + 373, + 245, + 481, + 317 + ], + "lines": [ + { + "bbox": [ + 373, + 245, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 373, + 245, + 481, + 317 + ], + "type": "image", + "image_path": "0fadbdf098beab5eb95c9713497ea7eb46a1aa1e6e20906838d22af4f7395b5d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 309, + 327, + 385, + 396 + ], + "blocks": [ + { + "bbox": [ + 309, + 327, + 385, + 396 + ], + "lines": [ + { + "bbox": [ + 309, + 327, + 385, + 396 + ], + "spans": [ + { + "bbox": [ + 309, + 327, + 385, + 396 + ], + "type": "image", + "image_path": "26b4006d1727886b770c936e64746194da33e5e442cce1f9422bbff5967c3e67.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 397, + 353, + 407 + ], + "lines": [ + { + "bbox": [ + 342, + 397, + 353, + 407 + ], + "spans": [ + { + "bbox": [ + 342, + 397, + 353, + 407 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 388, + 327, + 463, + 396 + ], + "blocks": [ + { + "bbox": [ + 420, + 318, + 432, + 327 + ], + "lines": [ + { + "bbox": [ + 420, + 318, + 432, + 327 + ], + "spans": [ + { + "bbox": [ + 420, + 318, + 432, + 327 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 327, + 463, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 327, + 463, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 327, + 463, + 396 + ], + "type": "image", + "image_path": "85fcc76f8260fec609d39d1efa6c7be0f9835d1a920f0c86a8b9f4748ae6a805.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 397, + 432, + 407 + ], + "lines": [ + { + "bbox": [ + 421, + 397, + 432, + 407 + ], + "spans": [ + { + "bbox": [ + 421, + 397, + 432, + 407 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 467, + 327, + 544, + 396 + ], + "blocks": [ + { + "bbox": [ + 467, + 327, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 467, + 327, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 467, + 327, + 544, + 396 + ], + "type": "image", + "image_path": "47698a3e5574e48d55fb89438b4df0461f2fe71c2d98da3561b7b81fb835af6e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 397, + 512, + 407 + ], + "lines": [ + { + "bbox": [ + 500, + 397, + 512, + 407 + ], + "spans": [ + { + "bbox": [ + 500, + 397, + 512, + 407 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 418, + 547, + 496 + ], + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 496 + ], + "type": "text", + "content": "Figure 1. An example of the PSG task. We compare the predicted object relations of different approaches. a) An image and its panoptic segmentation. Two persons and a sports ball are shown in different colors. b) A biased method predicts mostly high frequency relations. c) An unbiased method predicts mostly low frequency relations. d) Our method predicts both low and high frequency relations, as well as more relations in total." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 507, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 546, + 555 + ], + "type": "text", + "content": "and objects are represented with panoptic segmentation [35] masks. Unless stated otherwise, in this work we focus on PSG, since it is pixel-level accurate and also covers background classes and their relations with foreground objects." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 556, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 664 + ], + "type": "text", + "content": "The performance of the PSG model is affected by a long-tail problem in its relation categories. For instance, relations such as over, in front of and holding occur tens of thousands of times in the PSG dataset [62], while others like swinging and kissing occur only a few dozen times. This severe class imbalance in the relation categories can lead to model predictions that are more inclined to high-frequency relations, which poses significant challenges to the application of panoptic scene graphs in real-world scenarios." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Previous methods [12, 18, 39, 17] have often treated the long-tail problem of the PSG task as equivalent to the long-tail problem in object-centric tasks such as classification [15, 5, 29] or semantic segmentation [14, 2]. Consequently," + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 144, + 712 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21637" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "these methods have employed re-balancing techniques to address class imbalance, either through re-sampling the data [39] or by using a class-balanced loss [33] that assigns different weights to different relation categories." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 346 + ], + "type": "text", + "content": "In contrast, in the relation-centric PSG task, a subject-object pair can have multiple relations that exhibit relational semantic overlap, such as being partially or fully overlapping. For example, in Fig. 1, there are multiple relations between the boy and the sports ball, such as beside, looking at, playing and chasing. For regular biased models [61, 65, 60, 46], the results are dominated by high-frequency relations (beside, looking at). For specifically unbiased models [66, 62], the results are dominated by low-frequency relations (playing, chasing). However, since the low frequency relations can be more specific (e.g. on and standing on) or only partially overlapping with high frequency relations (e.g. looking at and chasing), it is crucial to include both to fully understand the image. We found that relational semantic overlap occurs in large numbers in the PSG dataset [62] and that current methods do not effectively address it. This is reflected in the increase in the category-averaged mean recall metric of unbiased methods, at the cost of the decrease in global recall (see Sec. 4.4)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 347, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 466 + ], + "type": "text", + "content": "To address the long-tail problem of scene graphs under relational semantic overlap, we introduce the HiLo framework. This framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths with the help of two novel consistency loss functions. We apply our framework on top of a novel baseline. This baseline uses a recent transformer-based approach [10] for panoptic segmentation and adapts triplet queries [62] and masked attention [10] for the PSG task. In summary, we make the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 469, + 286, + 625 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 469, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 469, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 59, + 469, + 286, + 515 + ], + "type": "text", + "content": "- We identify the long-tail problem with relational semantic overlap in the PSG task and propose the HiLo framework to address this problem. The framework is general and can be applied to any PSG method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 519, + 286, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 519, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 59, + 519, + 286, + 564 + ], + "type": "text", + "content": "- We propose a powerful and efficient one-stage end-to-end baseline. This baseline enhances the interaction between mask and relation prediction in the transformer decoder layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 569, + 286, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 569, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 59, + 569, + 286, + 625 + ], + "type": "text", + "content": "- We conduct extensive experiments to demonstrate the effectiveness of our framework and baseline. Our results outperform the state-of-the-art in both recall and mean recall on the PSG dataset and show systematic improvements on the VG dataset." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 639, + 134, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 134, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 134, + 651 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 186, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 186, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 186, + 672 + ], + "type": "text", + "content": "2.1. Scene Graph Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "The Scene Graph Generation (SGG) [49] task plays a crucial role in connecting vision and language, and has received widespread attention in the computer vision commu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "type": "text", + "content": "nity. Many methods have been proposed to improve the performance of SGG, which can be classified into three categories [70]. The first is to introduce multi-modal information, such as appearance [52], space [72], depth [54], and segmentation [34]. The second is to introduce prior information and commonsense knowledge, such as statistical [3, 16, 65, 9] and language prior knowledge [49, 43, 68, 31, 20]. The third category involves designing different model structures, such as message passing [41, 16, 42, 65, 24, 64], attention mechanisms [69, 50], tree structures and visual translation [67, 30]. However, most of these methods are two-stage methods that cannot learn scene graphs end-to-end. In contrast, to improve the learning ability of SGG models, several methods based on transformers have been proposed, including SGTR [38], RelationFormer [56] and RelTR [13]. These are end-to-end trainable in a single stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 284, + 492, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 492, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 492, + 297 + ], + "type": "text", + "content": "2.2. Unbiased Scene Graph Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 302, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 530 + ], + "type": "text", + "content": "Solving the long-tail problem in the SGG task has attracted considerable attention from researchers, and several unbiased methods [59, 63, 12, 18, 25, 39, 19, 22, 37, 40, 17, 66] have been proposed. These methods typically improve the model from the perspective of data re-sampling [39] or a class-balanced loss [33]. BGNN [39] uses a two-layer resampling strategy to provide a more balanced data distribution during training, while CogTree [63] proposal exploits the semantic relation between different predicate classes to design a novel CogTree loss. HML [17] improves the model's ability to solve long-tail problems by designing a staged training process, and IETrans [66] proposes an internal and external transfer method to transfer high frequency relations to low frequency relations and recover missing relations to train the unbiased model. Dong et al. [19] propose to group relations by their frequency and train specialized relation encoders for each group. While these methods are able to mitigate the long-tail problem, they do not address relational semantic overlap." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 539, + 489, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 489, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 489, + 552 + ], + "type": "text", + "content": "2.3. Panoptic Scene Graph Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 700 + ], + "type": "text", + "content": "Contrary to SGG, the PSG task [62] uses panoptic segmentation masks instead of bounding boxes to represent objects, resulting in a more comprehensive scene graph. PSGTR [62], an end-to-end method based on the DETR structure, was proposed to construct a transformer-based PSG model. PSGFormer [62] further improved on PSGTR by introducing Object & Relation Query Learning Blocks and Query Matching Blocks. Their strong performance on most relation classes indicates that their method is implicitly unbiased. In contrast, our approach is the first to explicitly bias a PSG model, creating separate branches for low and high frequency relations, which are then fused together." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Since the scene graph in PSG is built upon the sub" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21638" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 332 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 332 + ], + "type": "image", + "image_path": "d504548dc5565807a602bd6bcf021632dae67a9e444c9a18b35d5e033c982c02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 337, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 547, + 392 + ], + "type": "text", + "content": "Figure 2. An overview of our HiLo framework with HiLo baseline. a) HiLo relation swapping module swaps the multiple relations in the subject-object pair to obtain H-L Data and L-H Data respectively. b) Input data into our HiLo framework with HiLo baseline model, there are two branches, namely H-L decoder and L-H decoder, which learn H-L Data and L-H Data respectively. c) In addition to task losses for PSG, we propose HiLo prediction alignment, which includes subject-object consistency loss and relation consistency loss, so that the parallel branch can be better optimized." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 400, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 534 + ], + "type": "text", + "content": "jects, objects and their relations, strong panoptic segmentation [35] is crucial for PSG. Several DETR-based [6] methods such as Deformable-DETR [71], Segmenter [57], MaskFormer [11] and Mask2Former [10] have recently pushed the envelope in panoptic segmentation. These methods have introduced deformable transformer encoders and decoders, as well as pixel decoders and multi-scale information to improve model performance and speed up convergence. Our proposed baseline builds upon these techniques to further enhance the PSG performance and speed up model convergence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 543, + 103, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 103, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 103, + 555 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 563, + 146, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 146, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 146, + 576 + ], + "type": "text", + "content": "3.1. Problem Setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "The Panoptic Scene Graph generation (PSG) task aims to generate a panoptic scene graph " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " for a given image " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " contains an object set " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and a relational triplet set " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{O}, \\mathcal{T})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "-th object " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{o_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " to represent the object's mask and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " the object's category, i.e. " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "o_i = (m_i, c_i)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "-th triplet in " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_j\\}_{j=1}^M" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " to represent the subject, " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " the object, and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " their relation, i.e. " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "t_j = (s_j, o_j, r_j)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". There are in total " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " object categories and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " relation categories. For the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "-th relation, we denote its frequency in the training set as " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "f_k" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 399, + 394, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 394, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 394, + 411 + ], + "type": "text", + "content": "3.2. HiLo Baseline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 480 + ], + "type": "text", + "content": "High-quality panoptic segmentation is crucial for achieving good PSG performance. We build our method upon the latest advances in DETR-based panoptic segmentation, Mask2Former [10]. Below we present its structure, as well as our proposed modifications for the PSG task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": "Mask2Former. This method comprises three key parts: A backbone (CNN-based or transformer-based) followed by a pixel decoder, a transformer module with a transformer decoder and a task-specific module with different task heads. Specifically, the backbone takes an input image " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": " and generates an image feature " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": ". The pixel decoder then gradually upsamples " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": " to produce multi-scale features " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}} = \\{\\widetilde{F}_i\\}_{i=1}^4" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": ". The transformer decoder takes a set of queries " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": " and multi-scale features " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": " as input and outputs a set of mask features " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": " of the same size with " + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 304, + 482, + 546, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "content": "On top of the transformer decoder, there are two task heads including a linear classifier that predicts the class probability for each mask; and a multi-layer perceptron (MLP) that uses the mask features " + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "content": " to generate the mask embedding " + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "content": ". The mask prediction is obtained by taking the dot product of the mask embedding " + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "content": " with the scale feature of the highest resolution in the multi-scale features " + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}" + }, + { + "bbox": [ + 304, + 604, + 545, + 687 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Triplet queries. The original query in Mask2Former is to predict the object. In order to predict both subject, object" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21639" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "and their relation, we develop the triplet queries " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^t" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " inspired by [62] into our baseline. Each query predicts a triplet that includes subject, object, and relation. Accordingly, our task heads comprise three linear classifiers. Each classifier is responsible for predicting the class probability for subject, object or relation, respectively. Moreover, we devise two MLPs to generate mask embeddings for the subject and object, denoted by " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^s" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^o" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ". We use them to compute the dot product with the scale feature of the highest resolution in multi-scale feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " and obtain the mask prediction for the subject and object, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": "Masked relation attention. [10] proposes a transformer decoder variant with masked attention, which extracts spatial features by adding the predicted mask of the object from the previous decoder layer's mask prediction. It makes the model focus on the object-related area in the feature map. To adapt this scheme to the PSG task, for each relation, we extend the masked attention to take the union of the binary masks of the subject and object as input, which represents the pixels corresponding to the relation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "text", + "content": "Network training. We adopt the same losses as PS-GTR [62], including cross-entropy loss " + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{so\\_cls}" + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "text", + "content": " for object classification of subject-object pairs and a combination " + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{so\\_mask}" + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "text", + "content": " of focal loss [44] and dice loss [58] to jointly supervise mask learning. To supervise the relations, we use the cross-entropy loss " + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rel\\_cls}" + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "text", + "content": ". The baseline loss with " + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "inline_equation", + "content": "(\\lambda_1 = 1, \\lambda_2 = 1, \\lambda_3 = 4)" + }, + { + "bbox": [ + 47, + 312, + 289, + 395 + ], + "type": "text", + "content": " is thus:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b a s e l i n e}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {s o . c l s}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {s o . m a s k}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {r e l . c l s}} \\tag {1}", + "image_path": "1986a7db6fc500e7a7af3f758c31a2f0dd7756804b70636a0cf9f0d5a5aa22ae.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 153, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 153, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 153, + 447 + ], + "type": "text", + "content": "3.3. HiLo Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "content": "The key insight of our HiLo is to build a model that can take into account both high frequency and low frequency relations, and effectively improve the performance on low frequency relations without degrading the performance on high frequency relations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 186, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 186, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 186, + 540 + ], + "type": "text", + "content": "3.3.1 HiLo relation generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": "In the PSG task, multiple relations can be used to describe the connection between a subject and an object from different perspectives, such as spatial relations, actions and prepositions. Since the subject, object and their relative position in the image are fixed, all these relations share the same visual information. This is reflected in the PSG dataset [62] where many subject-object pairs are annotated with multiple relation labels. These relation labels are of different frequencies in the dataset. In this section we introduce the HiLo relation generation module to prepare two sets of training data, biasing towards high and low-frequency relations respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Relation augmentation. Similar to [22], we observe that there are many missing relation annotations in the PSG" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "dataset [62]. To add the missing relation annotations, we design a relation augmentation scheme that is inspired by IETrans [66], which converts high-frequency relations to low-frequency relations and adds a relation to a subject-object pair that has no relation. We adapt it by first training our baseline as a biased model using the original annotated data. For every subject-object pair in the training set, we use this model to predict the relation scores for all predefined relation categories:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 181, + 545, + 312 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 181, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 181, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 181, + 545, + 252 + ], + "type": "text", + "content": "- If this subject-object pair has annotated relation labels, we pick the one with the highest predicted score and use the score as a threshold. For other relations (except for already labeled ones) whose predicted scores are greater than the threshold, we add them as the relation labels of this subject-object pair." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 253, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 545, + 312 + ], + "type": "text", + "content": "- If this subject-object pair has no annotated relation labels, we use the predicted score of the no-relation class as a threshold. For other relations whose predicted scores are greater than this threshold we add them as the relation labels of this subject-object pair." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "type": "text", + "content": "This operation allows us to significantly augment the relation labels for the subject-object pairs in the training data, which is especially relevant for pairs with zero or no annotated relations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "text", + "content": "Relation swapping. We swap the relation labels for each subject-object pair. Specifically, given a subject-object pair " + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "inline_equation", + "content": "(s,o)" + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "text", + "content": " relation labels, we have " + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "text", + "content": " triplets " + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "inline_equation", + "content": "(s,o,r_1),\\ldots ,(s,o,r_K)" + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "text", + "content": ", sorted by their relation frequency in descending order, " + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "inline_equation", + "content": "f_{1} > \\ldots >f_{K}" + }, + { + "bbox": [ + 304, + 363, + 545, + 458 + ], + "type": "text", + "content": ". We denote by H-L and L-H the swapping of high-frequency relations with low-frequency relations and vice versa. This creates two sets of data:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "text", + "content": "H-L Data. Given a triplet " + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "inline_equation", + "content": "(s, o, r_k)" + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "text", + "content": ", we replace its relation label " + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "text", + "content": " with that of the next triplet " + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "inline_equation", + "content": "r_{k+1}" + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "text", + "content": " with lower frequency. We sequentially process all triplets from " + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "inline_equation", + "content": "(s, o, r_1)" + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "inline_equation", + "content": "(s, o, r_{K-1})" + }, + { + "bbox": [ + 304, + 461, + 545, + 509 + ], + "type": "text", + "content": ", keeping the last triplet unchanged." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "L-H Data. Given a triplet " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "(s,o,r_k)" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", we replace its relation label " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " with that of the previous triplet " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "r_{k - 1}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " with higher frequency. We sequentially process all triplets from " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "(s,o,r_K)" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "(s,o,r_2)" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", keeping the first triplet unchanged. Hence, we obtain two new sets of triplets, denoted by " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{H - L}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{L - H}}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". We devise two parallel decoders from the shared encoder of our backbone. They are learned with the H-L and L-H data, respectively. The H-L and L-H decoders favour the predictions for low and high-frequency relations, respectively. Despite their difference, for the corresponding triplet query in the two decoders, their predictions are highly correlated: on one hand, the subject and object predictions should be the same; on the other hand, the distribution of relation predictions should be overlapping. Below we first build the query correspondence in the two decoders and then introduce the HiLo subject-object and relation consistency loss to align the predictions from two the decoders." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21640" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 194, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 194, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 194, + 85 + ], + "type": "text", + "content": "3.3.2 HiLo prediction alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 286, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 286, + 163 + ], + "type": "text", + "content": "Training two different relation distributions simultaneously confuse the model. To make the model differentiate between the two branches, we built a HiLo prediction alignment module, including triplet query correspondence, subject-object consistency loss and relation consistency loss respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 286, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 286, + 329 + ], + "type": "text", + "content": "Triplet query correspondence. In order to construct the subject-object consistency loss and relation consistency loss, we first need to construct the correspondence between the triplet queries in the H-L and L-H decoders. The same query index in the two decoders is not naturally matched. In order to find the query correspondence, we need to rely on the ground truth assignment: we use Hungarian matching to assign the triplet label to the corresponding triplet query, and record the relation label index corresponding to the triplet query. This label index allows us to construct the correspondence between the triplet queries in both decoders. We calculate the consistency loss for the triplet query prediction with the same relation label index in the two decoders." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 330, + 286, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 286, + 401 + ], + "type": "text", + "content": "Subject-object consistency loss. Having the corresponding predictions from the two decoders, both their subjects and objects have the same ground truth and should be equal. We propose a subject-object consistency loss " + }, + { + "bbox": [ + 46, + 330, + 286, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 46, + 330, + 286, + 401 + ], + "type": "text", + "content": " to minimize the mean squared error (MSE) of the corresponding predictions from the two decoders." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 411, + 286, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 411, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 72, + 411, + 286, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o b j} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {m a s k} \\tag {2}", + "image_path": "123d433db2feadd60958c5c594f80b800f3bc115eec4e1a992b0cff645b4c3be.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 426, + 286, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 426, + 286, + 439 + ], + "spans": [ + { + "bbox": [ + 74, + 426, + 286, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = \\left\\| \\operatorname {s o f t m a x} \\left(p _ {c} ^ {\\mathrm {H} - \\mathrm {L}}\\right) - \\operatorname {s o f t m a x} \\left(p _ {c} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\right\\| ^ {2} \\tag {3}", + "image_path": "5d0c9e5c924bf5cf08319afdad6cb247f8c6709d385e30189492710c5ba1d585.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 441, + 286, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 441, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 64, + 441, + 286, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\left\\| \\operatorname {s i g m o i d} \\left(p _ {m} ^ {\\mathrm {H} - \\mathrm {L}}\\right) - \\operatorname {s i g m o i d} \\left(p _ {m} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\right\\| ^ {2} \\tag {4}", + "image_path": "4f1ba821827e0a6893e4c8b61da38b468bd8191034aa5803cd562e1725ee7f77.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 462, + 286, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 286, + 486 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 46, + 462, + 286, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 46, + 462, + 286, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 462, + 286, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 46, + 462, + 286, + 486 + ], + "type": "text", + "content": " represent the MSE losses for class prediction and mask prediction, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "inline_equation", + "content": "p_c^{\\mathrm{H - L}}" + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "inline_equation", + "content": "p_c^{\\mathrm{L - H}}" + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "content": " are the class prediction logits from the H-L and L-H decoders. After a softmax, they are " + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "content": "-dimensional probability vectors. After a sigmoid, " + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "inline_equation", + "content": "p_m^{H - L}" + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "inline_equation", + "content": "p_m^{L - H}" + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "content": " are the mask prediction logits for the H-L and L-H decoders." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 641 + ], + "type": "text", + "content": "Relation consistency loss. Given a pair of subject and object, we have previously swapped the high-low frequency relation labels to create data for the H-L and L-H decoders. For H-L, the prediction of the low-frequency relation logit is of high value; while for L-H, the prediction of the high-frequency relation logit is of high value. For the predictions on the rest logits, they should be similar, since it is the same pair of subject and object for the two decoders. Based on this observation, we introduce the relation consistency loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Specifically, for a pair of subject and object, we use " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "p_r^{\\mathrm{H - L}}" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " to denote the predicted relation logits from the H-L decoder, " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "p_r^{\\mathrm{L - H}}" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": " the predicted relation logits from the L-H decoder. They are " + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "-dimensional probability vectors after softmax. We can map between the distributions of the two vectors by swapping the logit values between the high- and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "low-frequency relation indices, which is named Relational Index Exchange operation. We use " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{RIE}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " to denote this operation. " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{RIE}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " includes a stop gradient operation, which creates copies from original predictions of two branches, enabling value exchanges. For example, for relations " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "r_{k + 1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{RIE}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " exchanges the values between " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "p_{r,k}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "p_{r,k + 1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "p_r" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". We can therefore compute the distance between " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "p_r^{\\mathrm{H - L}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and its mapped counterpart from " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "p_r^{\\mathrm{L - H}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", vice versa:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 173, + 545, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 314, + 173, + 545, + 215 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {D i s t} _ {\\mathrm {H i L o}} = \\left\\| \\text {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {H - L}}\\right) - R I E \\left(\\text {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {L - H}}\\right)\\right) \\right\\| ^ {2} \\\\ + \\| R I E \\left(\\operatorname {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {H} - \\mathrm {L}}\\right)\\right) - \\operatorname {s o f t m a x} \\left(p _ {r} ^ {\\mathrm {L} - \\mathrm {H}}\\right) \\| ^ {2} \\tag {5} \\\\ \\end{array}", + "image_path": "cec8e5d452c320ec95e6a92762bafc34e6a398195a7cfce9352511266f9d405c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 222, + 545, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 246 + ], + "type": "text", + "content": "The relation consistency loss is defined to minimize " + }, + { + "bbox": [ + 304, + 222, + 545, + 246 + ], + "type": "inline_equation", + "content": "\\mathrm{Dist}_{\\mathrm{HiLo}}" + }, + { + "bbox": [ + 304, + 222, + 545, + 246 + ], + "type": "text", + "content": " with a margin of " + }, + { + "bbox": [ + 304, + 222, + 545, + 246 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 222, + 545, + 246 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 249, + 544, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 249, + 544, + 261 + ], + "spans": [ + { + "bbox": [ + 364, + 249, + 544, + 261 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e l} = \\max \\left(\\operatorname {D i s t} _ {\\mathrm {H i L o}} - m, 0\\right), \\tag {6}", + "image_path": "dd0ebba80b36e00543efb77261c4c8cd73f5ca00ad7d124e043b004cd0411f89.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 268, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 268, + 545, + 304 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 268, + 545, + 304 + ], + "type": "text", + "content": " is a small constant. Adding " + }, + { + "bbox": [ + 304, + 268, + 545, + 304 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 268, + 545, + 304 + ], + "type": "text", + "content": " is due to the fact that the high- and low frequency relations might be only partially semantically overlapping." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 304, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 374 + ], + "type": "text", + "content": "Network training. Our subject-object consistency loss and relation consistency loss can seamlessly integrate with the losses of any baseline method to jointly supervise the training of the entire model. Notably, we supervise the output of each transformer decoder layer to ensure effective learning. The final loss " + }, + { + "bbox": [ + 304, + 304, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 304, + 545, + 374 + ], + "type": "text", + "content": " is thus:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 366, + 384, + 544, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 384, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 366, + 384, + 544, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e l i n e}} + \\mathcal {L} _ {\\text {o b j}} + \\mathcal {L} _ {\\text {r e l}} \\tag {7}", + "image_path": "61c8aabcadc2959b135621feff393a8bef68d317186f3a4b316b793d3145689c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 418, + 431, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 431, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 431, + 428 + ], + "type": "text", + "content": "3.3.3 HiLo inference fusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "content": "The H-L and L-H decoders favour low-frequency relation prediction and high-frequency relation prediction, respectively. To combine the strength of both during inference, we introduce the HiLo inference fusion module. Specifically, we denote by " + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{H - L}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{L - H}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "content": " the predicted panoptic scene graphs from H-L and L-H decoders, respectively. There are " + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "content": " triplets in " + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{H - L}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "content": " triplets in " + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{L - H}}" + }, + { + "bbox": [ + 304, + 436, + 545, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 519, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "text", + "content": "- First, we merge the triplets in " + }, + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{H - L}}" + }, + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{L - H}}" + }, + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "text", + "content": " and sort them in the descending order according to their relation scores. We obtain a list of " + }, + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "inline_equation", + "content": "N_{1} + N_{2}" + }, + { + "bbox": [ + 317, + 519, + 544, + 555 + ], + "type": "text", + "content": " triplets." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "text", + "content": "- Second, starting from the first triplet, we de-duplicate the triplet list. For the " + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "text", + "content": "-th triplet " + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "text", + "content": ", we identify its duplicated versions from the " + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "text", + "content": "-th triplet until the end of the list. Remove any follow-up triplet from the list if it 1) has the same subject, object and relation classes to that in " + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "text", + "content": " and 2) has a mask IoU greater than a threshold, e.g. 0.5, between the predicted subject/object and the corresponding subject/object in " + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 317, + 555, + 545, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 651, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 651, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 317, + 651, + 545, + 711 + ], + "type": "text", + "content": "- Third, after dedduplication, for each triplet in the list, we multiply the relation, subject and object scores as an overall score for it. We sort the triplet list according to this score in descending order to obtain the final panoptic scene graph " + }, + { + "bbox": [ + 317, + 651, + 545, + 711 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\mathrm{HiLo}}" + }, + { + "bbox": [ + 317, + 651, + 545, + 711 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21641" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 58, + 492, + 193 + ], + "blocks": [ + { + "bbox": [ + 99, + 58, + 492, + 193 + ], + "lines": [ + { + "bbox": [ + 99, + 58, + 492, + 193 + ], + "spans": [ + { + "bbox": [ + 99, + 58, + 492, + 193 + ], + "type": "table", + "html": "
MethodBackboneScene Graph Detection
R@20mR@20R@50mR@50R@100mR@100
IMP [61]R5016.56.518.27.118.67.2
MOTIF [65]R5020.09.121.79.622.09.7
VCTree [60]R5020.69.722.110.222.510.2
GPSNet [46]R5017.87.019.67.520.17.7
PSGTR [62]R5028.416.634.420.836.322.1
PSGFormer [62]R5018.014.819.617.020.117.6
HiLo (ours)R5034.123.740.730.343.033.1
HiLo (ours)Swin-B38.528.346.235.349.639.1
HiLo (ours)Swin-L40.629.748.737.651.440.9
", + "image_path": "f3d1795f460d33158742d3dbcb18eae607688c65c04929121ff5b239b6feabdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 218 + ], + "type": "text", + "content": "Table 1. Comparison between our HiLo and other methods on the PSG dataset. Our method shows superior performance compared to all previous methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 230, + 128, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 128, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 128, + 243 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 254, + 111, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 111, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 111, + 266 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 288, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 288, + 397 + ], + "type": "text", + "content": "Panoptic Scene Graph Generation (PSG) [62]. This is the first Panoptic Scene Graph generation dataset. It has a total of 48,749 labeled images including 2,177 test images and 46,572 training images. The object categories comprise 80 thing classes and 53 stuff classes, which is the same as the COCO [45] and COCO-Stuff datasets [4]. The relation categories comprise 56 classes, including positional relations, common object-object relations, common actions, human actions, actions in the traffic scene, actions in the sports scene and interactions between backgrounds [62]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": "Visual Genome (VG) [36]. VG is a widely used benchmark dataset for Scene Graph Generation. Following previous work [65, 7], we adopt the widely accepted split, VG-150, which contains 150 object categories and 50 relation categories. The object categories cover a wide range of classes, such as animals, vehicles and household items. The relation categories include both spatial and semantic classes, such as on, in and wearing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 518, + 155, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 155, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 155, + 530 + ], + "type": "text", + "content": "4.2. Tasks and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 541, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 649 + ], + "type": "text", + "content": "Three subtasks have been proposed for the SGG and PSG tasks, which are Predicate Classification, Scene Graph Classification and Scene Graph Detection [61]. We focus on Scene Graph Detection for both datasets, since it is the most comprehensive and addressed by [62]. This subtask requires the model to first localize the objects and then predict the object classes and relations. Note that Scene Graph Detection in PSG includes the detection on stuff classes, while in SGG it does not." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Following previous work [60, 63, 62], we use Recall@K (R@K) and mean Recall@K (mR@K) as our metrics, where the former metric is dominated by high-frequency relations, while the latter assigns equal weight to all relation classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 230, + 437, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 230, + 437, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 437, + 243 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "text", + "content": "In our experiments, we follow the training strategy of PSGTR [62]. We use the AdamW optimizer [48], with a learning rate of " + }, + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "text", + "content": " and weight decay of " + }, + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "text", + "content": ", except for the backbone, which is trained with a learning rate of " + }, + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 304, + 251, + 547, + 430 + ], + "type": "text", + "content": ". For initialization, we use Mask2Former [10] pretrained on COCO [45] to initialize our backbone and pixel decoder. Following Mask2Former [10], we use 100 triplet queries for the H-L and L-H decoders respectively. Additionally, both the H-L and L-H decoders are initialized with Mask2Former's transformer decoder. To ensure consistent comparison with PSGTR, we adopted the same data augmentation settings. Our model is trained for 12 epochs with a step scheduler at epoch 10, taking approximately 18 hours to train on four A100 GPUs with a batch size of 1 for each GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 442, + 489, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 489, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 489, + 455 + ], + "type": "text", + "content": "4.4. Comparison to the state-of-the-art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "content": "PSG. Tab. 1 reports the performance of our method compared to the state-of-the-art on the PSG dataset [62]. We separate the methods into two groups. The first are two-stage methods consisting of a separate segmentor and relation predictor, which are modified for the PSG task in [62]. The second are one-stage end-to-end methods, which are able to simultaneously predict panoptic segmentation and relations. Our method belongs to the second category. For a fair comparison between the different methods, we use the same Resnet-50 [26] backbone. Our method shows superior performance compared to all previous methods. Particularly, it outperforms the previous best-performing method PSGTR [62] by a large margin, i.e. " + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "inline_equation", + "content": "+6\\%" + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "content": " in R@100 and " + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "inline_equation", + "content": "+11\\%" + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "content": " in mR@100. Our model is able to converge within only 12 epochs of training, whereas PSGTR [62] is trained for 60 epochs. We also evaluate our method using pre-trained transformer-based backbones, i.e. Swin-B and Swin-L [47], with the latter being a bigger model. Our results are consistently improved over all metrics due the powerful feature representation ability of the transformer. We also conducted a visual comparison, as shown in Fig. 3." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21642" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 258, + 227 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 258, + 227 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 258, + 227 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 258, + 227 + ], + "type": "table", + "html": "
MethodScene Graph Detection
R/mR@50R/mR@100
MOTIF [65]31.0 / 6.735.1 / 7.7
+IETrans [66]26.4 / 12.430.6 / 14.9
+HiLo (ours)26.2 / 14.730.3 / 17.7
VCTree [60]30.2 / 6.734.6 / 8.0
+IETrans [66]25.4 / 11.529.3 / 14.0
+HiLo (ours)27.1 / 12.929.8 / 15.2
Transformer [59]30.0 / 7.434.3 / 8.8
+IETrans [66]25.5 / 12.529.6 / 15.0
+HiLo (ours)25.4 / 14.629.2 / 17.6
GPSNet [46]30.3 / 5.935.0 / 7.1
+IETrans [66]25.9 / 14.628.1 / 16.5
+HiLo (ours)25.6 / 15.827.9 / 18.0
", + "image_path": "e6b2be7f5804958c7060c5b4f8dae5465d8189635924127a4ff1d6a6f9cd47ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 495 + ], + "type": "text", + "content": "SGG. Here we study whether our approach that was developed for the PSG task can also be applied to the SGG task. In Tab. 2, we conduct experiments on the VG-150 dataset. Following IETrans [66], we extend our unbiased HiLo framework (Sec. 3.3) without our PSG-specific baseline (Sec. 3.2) to four state-of-the-art biased SGG methods, namely MOTIF [65], VCTree [60], Transformer [59] and GPSNet [46], using the implementation of [66]. Compared to the unbiased IETrans [66] method, our method improves mean recall without sacrificing recall, demonstrating its effectiveness in enhancing the performance of low-frequency relations. Compared to the biased baselines [65, 60, 59, 46], our method achieves a significantly larger mean recall, while still maintaining an acceptable recall. This indicates that our method can improve the performance of low-frequency relations while also taking into account the performance of high-frequency relations. It also shows that the HiLo framework is a general technique that yields systematic improvements in both the PSG and SGG tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 503, + 147, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 147, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 147, + 514 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "content": "Consistent with the paper, we use HiLo with a Resnet-50 backbone to perform ablation experiments on the PSG dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 700 + ], + "type": "text", + "content": "HiLo framework for different baselines. In this section we investigate whether our HiLo framework (Sec. 3.3) yields improvements for other baselines, rather than just the one presented in Sec. 3.2. Tab. 3 shows the results for two baselines, with and without the HiLo framework. We observe that our biased baseline outperforms the previous PSGTR [62] method on all metrics. Furthermore, by applying the HiLo framework, we can substantially improve the performance over both baselines. It is worth mentioning that the HiLo framework improves recall and mean recall simultaneously, whereas other methods typically improve one metric at the cost of the other [59, 63]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "HiLo relation augmentation. We observe that out of" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 543, + 125 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 262 + ], + "type": "text", + "content": "Table 2. Comparison between our HiLo framework and other methods on the VG-150 dataset. Similar to [66], we apply IETrans and our own method on top of four leading baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 543, + 125 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 125 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 125 + ], + "type": "table", + "html": "
BaselineHiLoR/mR@20R/mR@50R/mR@100
HiLo baseline✓34.1 / 23.740.7 / 30.343.0 / 33.1
HiLo baseline-32.6 / 20.938.0 / 27.438.9 / 28.4
PSGTR [62]✓30.1 / 20.236.6 / 23.938.3 / 24.5
PSGTR [62]-28.4 / 16.634.4 / 20.836.3 / 22.1
", + "image_path": "f46ecceb452ce1ba47c56fa5a1c05428dd6d8b1dfb206b7d03677afdb9548111.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 167, + 543, + 201 + ], + "blocks": [ + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 163 + ], + "type": "text", + "content": "Table 3. Comparison of different baselines, with and without HiLo framework. Using the HiLo framework, we see significant improvements on both metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 167, + 543, + 201 + ], + "lines": [ + { + "bbox": [ + 309, + 167, + 543, + 201 + ], + "spans": [ + { + "bbox": [ + 309, + 167, + 543, + 201 + ], + "type": "table", + "html": "
Relation Aug.Multiple relationsR/mR@50R/mR@100
✓40%40.7 / 30.343.0 / 33.1
-10%40.1 / 28.142.8 / 32.5
", + "image_path": "3f4a9b1ab653b5c624c1f8333e7c51bd32379c9b1dd7e112fbd6f464b4a684fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 208, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 251 + ], + "type": "text", + "content": "Table 4. Ablation study for HiLo relation augmentation. Relation augmentation affects the ratio of subject-object pairs with multiple relations. The larger this ratio, the more relations can be swapped, which leads to better results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": "260,296 labeled triplets in the PSG dataset, only about " + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": " of subject-object pairs have multiple relations, for which we can apply relation swapping (Sec. 3.3.1). After applying our proposed relation augmentation technique (Sec. 3.3.1), this ratio significantly increases to " + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": ". Our experimental results in Tab. 4 demonstrate that only applying the HiLo framework on " + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 378 + ], + "type": "text", + "content": " already gives an improvement over the baseline from Tab. 3. As the number of swappable triplets increases due to augmentation, the model's performance is further enhanced, highlighting the potential of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 379, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 510 + ], + "type": "text", + "content": "HiLo prediction alignment. We conduct ablation experiments on the subject-object consistency loss and relation consistency loss (Sec. 3.3.2), which are used to align the predictions from the high and low frequency branches. The results, as presented in Tab. 5, demonstrate that using both losses yields the best performance. It is worth mentioning that we have explored the margin in the relation consistency loss and found that setting the margin to zero leads to a small performance degradation. This finding confirms that there is a partial semantic overlap between swapped relations, indicating that they are not entirely consistent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "content": "To investigate the impact of relational index exchange (RIE) on the relation consistency loss, we conducted experiments to verify the effect of omitting RIE. In the absence of RIE, we solely compute the consistency loss for relation categories that are not involved in relation swapping and exclude the swapping component from the calculation of the consistency loss. The outcomes of this experiment are presented in Tab. 6, and demonstrate a notable reduction in the mean recall and a decline in the model's performance for relations with relational semantic overlap when RIE is not utilized." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "HiLo inference fusion. We ablate the inference fusion (Sec. 3.3.3) and evaluate the performance of each branch's output separately. In Tab. 7, experimental results suggest that fusion can effectively leverage the uniqueness of the high and low frequency branch predictions to achieve comprehensive improvements on all metrics." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21643" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 75, + 129, + 128 + ], + "blocks": [ + { + "bbox": [ + 52, + 75, + 129, + 128 + ], + "lines": [ + { + "bbox": [ + 52, + 75, + 129, + 128 + ], + "spans": [ + { + "bbox": [ + 52, + 75, + 129, + 128 + ], + "type": "image", + "image_path": "7fa600f28fd3137b58e0fa7f7db26250b9afb8bf2fcb8f7b65c7781e73616aab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 130, + 100, + 140 + ], + "lines": [ + { + "bbox": [ + 84, + 130, + 100, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 130, + 100, + 140 + ], + "type": "text", + "content": "↑ Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 166, + 129, + 219 + ], + "blocks": [ + { + "bbox": [ + 77, + 150, + 107, + 156 + ], + "lines": [ + { + "bbox": [ + 77, + 150, + 107, + 156 + ], + "spans": [ + { + "bbox": [ + 77, + 150, + 107, + 156 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 166, + 129, + 219 + ], + "lines": [ + { + "bbox": [ + 53, + 166, + 129, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 129, + 219 + ], + "type": "image", + "image_path": "3a9f8d2eb981294c28102f3ab364f95e94d73dc3233d2f7fd5d30f65b93a9a82.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 167, + 185, + 171 + ], + "lines": [ + { + "bbox": [ + 137, + 167, + 185, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 185, + 171 + ], + "type": "text", + "content": "person walking on sea" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 172, + 177, + 177 + ], + "lines": [ + { + "bbox": [ + 137, + 172, + 177, + 177 + ], + "spans": [ + { + "bbox": [ + 137, + 172, + 177, + 177 + ], + "type": "text", + "content": "surfboard over sea" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 178, + 167, + 182 + ], + "lines": [ + { + "bbox": [ + 137, + 178, + 167, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 178, + 167, + 182 + ], + "type": "text", + "content": "sky over surf" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 209, + 198, + 214 + ], + "lines": [ + { + "bbox": [ + 137, + 209, + 198, + 214 + ], + "spans": [ + { + "bbox": [ + 137, + 209, + 198, + 214 + ], + "type": "text", + "content": "person attached to surfboard" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 215, + 198, + 220 + ], + "lines": [ + { + "bbox": [ + 137, + 215, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 137, + 215, + 198, + 220 + ], + "type": "text", + "content": "surfboard attached to person" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 221, + 181, + 227 + ], + "lines": [ + { + "bbox": [ + 137, + 221, + 181, + 227 + ], + "spans": [ + { + "bbox": [ + 137, + 221, + 181, + 227 + ], + "type": "text", + "content": "surfboard on person" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 74, + 228, + 86, + 233 + ], + "blocks": [ + { + "bbox": [ + 56, + 222, + 113, + 227 + ], + "lines": [ + { + "bbox": [ + 56, + 222, + 113, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 113, + 227 + ], + "type": "text", + "content": "person carrying surfboard" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 228, + 86, + 233 + ], + "lines": [ + { + "bbox": [ + 74, + 228, + 86, + 233 + ], + "spans": [ + { + "bbox": [ + 74, + 228, + 86, + 233 + ], + "type": "image", + "image_path": "b3646b8e1b5dbaaeeb361e9b91848002f2fbdce0140cfd6ca683a2f080d46bb7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 234, + 96, + 240 + ], + "lines": [ + { + "bbox": [ + 74, + 234, + 96, + 240 + ], + "spans": [ + { + "bbox": [ + 74, + 234, + 96, + 240 + ], + "type": "text", + "content": "looking at" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 247, + 100, + 252 + ], + "lines": [ + { + "bbox": [ + 56, + 247, + 100, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 100, + 252 + ], + "type": "text", + "content": "sky over person" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 253, + 85, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 253, + 85, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 85, + 259 + ], + "type": "text", + "content": "sky over sea" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 132, + 76, + 209, + 129 + ], + "blocks": [ + { + "bbox": [ + 96, + 144, + 126, + 149 + ], + "lines": [ + { + "bbox": [ + 96, + 144, + 126, + 149 + ], + "spans": [ + { + "bbox": [ + 96, + 144, + 126, + 149 + ], + "type": "text", + "content": "Prediction " + }, + { + "bbox": [ + 96, + 144, + 126, + 149 + ], + "type": "inline_equation", + "content": "\\rightarrow" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 76, + 209, + 129 + ], + "lines": [ + { + "bbox": [ + 132, + 76, + 209, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 76, + 209, + 129 + ], + "type": "image", + "image_path": "c603cee51a1a8b18c820cde2034ae09fa7a704cb872df026ca4d63d2ac96985e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 131, + 191, + 137 + ], + "lines": [ + { + "bbox": [ + 134, + 131, + 191, + 137 + ], + "spans": [ + { + "bbox": [ + 134, + 131, + 191, + 137 + ], + "type": "text", + "content": "person holding surfboard" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 151, + 150, + 165, + 155 + ], + "blocks": [ + { + "bbox": [ + 135, + 138, + 164, + 144 + ], + "lines": [ + { + "bbox": [ + 135, + 138, + 164, + 144 + ], + "spans": [ + { + "bbox": [ + 135, + 138, + 164, + 144 + ], + "type": "text", + "content": "sky over sea" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 144, + 203, + 149 + ], + "lines": [ + { + "bbox": [ + 136, + 144, + 203, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 144, + 203, + 149 + ], + "type": "text", + "content": "person standing on sea" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 150, + 165, + 155 + ], + "lines": [ + { + "bbox": [ + 151, + 150, + 165, + 155 + ], + "spans": [ + { + "bbox": [ + 151, + 150, + 165, + 155 + ], + "type": "image", + "image_path": "6e1a1c82d1b12bb4fcd21ef016055f53c023d0542011c0dee25c22fa2fcba4f8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 156, + 165, + 160 + ], + "lines": [ + { + "bbox": [ + 151, + 156, + 165, + 160 + ], + "spans": [ + { + "bbox": [ + 151, + 156, + 165, + 160 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 161, + 171, + 166 + ], + "lines": [ + { + "bbox": [ + 137, + 161, + 171, + 166 + ], + "spans": [ + { + "bbox": [ + 137, + 161, + 171, + 166 + ], + "type": "text", + "content": "sky over person" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 151, + 196, + 165, + 201 + ], + "blocks": [ + { + "bbox": [ + 137, + 183, + 167, + 189 + ], + "lines": [ + { + "bbox": [ + 137, + 183, + 167, + 189 + ], + "spans": [ + { + "bbox": [ + 137, + 183, + 167, + 189 + ], + "type": "text", + "content": "person on sea" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 190, + 193, + 195 + ], + "lines": [ + { + "bbox": [ + 137, + 190, + 193, + 195 + ], + "spans": [ + { + "bbox": [ + 137, + 190, + 193, + 195 + ], + "type": "text", + "content": "person carrying surfboard" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 196, + 165, + 201 + ], + "lines": [ + { + "bbox": [ + 151, + 196, + 165, + 201 + ], + "spans": [ + { + "bbox": [ + 151, + 196, + 165, + 201 + ], + "type": "image", + "image_path": "b7f1aa76cca772df169569b9f3788fe4095be02c539bafa96ca2e7cbbcd0e0ba.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 201, + 182, + 207 + ], + "lines": [ + { + "bbox": [ + 151, + 201, + 182, + 207 + ], + "spans": [ + { + "bbox": [ + 151, + 201, + 182, + 207 + ], + "type": "text", + "content": "looking at you" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 137, + 240, + 165, + 246 + ], + "blocks": [ + { + "bbox": [ + 137, + 228, + 182, + 233 + ], + "lines": [ + { + "bbox": [ + 137, + 228, + 182, + 233 + ], + "spans": [ + { + "bbox": [ + 137, + 228, + 182, + 233 + ], + "type": "text", + "content": "sea in front of person" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 234, + 175, + 239 + ], + "lines": [ + { + "bbox": [ + 137, + 234, + 175, + 239 + ], + "spans": [ + { + "bbox": [ + 137, + 234, + 175, + 239 + ], + "type": "text", + "content": "sea beside person" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 240, + 165, + 246 + ], + "lines": [ + { + "bbox": [ + 137, + 240, + 165, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 240, + 165, + 246 + ], + "type": "image", + "image_path": "afd8823e7c5b5591c040c7eb8afaf2e9d9763e319fa5a06197d26292e32f5efe.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 247, + 190, + 252 + ], + "lines": [ + { + "bbox": [ + 137, + 247, + 190, + 252 + ], + "spans": [ + { + "bbox": [ + 137, + 247, + 190, + 252 + ], + "type": "text", + "content": "person playing surfboard" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 253, + 173, + 258 + ], + "lines": [ + { + "bbox": [ + 137, + 253, + 173, + 258 + ], + "spans": [ + { + "bbox": [ + 137, + 253, + 173, + 258 + ], + "type": "text", + "content": "surfboard on sea" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 219, + 75, + 295, + 129 + ], + "blocks": [ + { + "bbox": [ + 219, + 75, + 295, + 129 + ], + "lines": [ + { + "bbox": [ + 219, + 75, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 219, + 75, + 295, + 129 + ], + "type": "image", + "image_path": "a04fcf96662337ab4ff31d31471d73864e5cfd71636071d51a96174372a56e2b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 133, + 265, + 138 + ], + "lines": [ + { + "bbox": [ + 250, + 133, + 265, + 138 + ], + "spans": [ + { + "bbox": [ + 250, + 133, + 265, + 138 + ], + "type": "text", + "content": "↑" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 139, + 264, + 144 + ], + "lines": [ + { + "bbox": [ + 251, + 139, + 264, + 144 + ], + "spans": [ + { + "bbox": [ + 251, + 139, + 264, + 144 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 146, + 291, + 152 + ], + "lines": [ + { + "bbox": [ + 242, + 146, + 291, + 152 + ], + "spans": [ + { + "bbox": [ + 242, + 146, + 291, + 152 + ], + "type": "text", + "content": "Prediction→" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 219, + 171, + 296, + 224 + ], + "blocks": [ + { + "bbox": [ + 242, + 153, + 290, + 159 + ], + "lines": [ + { + "bbox": [ + 242, + 153, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 242, + 153, + 290, + 159 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 256, + 160, + 272, + 165 + ], + "lines": [ + { + "bbox": [ + 256, + 160, + 272, + 165 + ], + "spans": [ + { + "bbox": [ + 256, + 160, + 272, + 165 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 171, + 296, + 224 + ], + "lines": [ + { + "bbox": [ + 219, + 171, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 219, + 171, + 296, + 224 + ], + "type": "image", + "image_path": "ce68394d186232dd3e2c1e3f79ce71c568dc0aa633e68981525fff37f793a0ce.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 152, + 335, + 157 + ], + "lines": [ + { + "bbox": [ + 303, + 152, + 335, + 157 + ], + "spans": [ + { + "bbox": [ + 303, + 152, + 335, + 157 + ], + "type": "text", + "content": "sky over snow" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 158, + 337, + 163 + ], + "lines": [ + { + "bbox": [ + 303, + 158, + 337, + 163 + ], + "spans": [ + { + "bbox": [ + 303, + 158, + 337, + 163 + ], + "type": "text", + "content": "sky over person" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 164, + 356, + 170 + ], + "lines": [ + { + "bbox": [ + 303, + 164, + 356, + 170 + ], + "spans": [ + { + "bbox": [ + 303, + 164, + 356, + 170 + ], + "type": "text", + "content": "person standing on snow" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 171, + 332, + 176 + ], + "lines": [ + { + "bbox": [ + 303, + 171, + 332, + 176 + ], + "spans": [ + { + "bbox": [ + 303, + 171, + 332, + 176 + ], + "type": "text", + "content": "sky over skis" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 177, + 343, + 182 + ], + "lines": [ + { + "bbox": [ + 303, + 177, + 343, + 182 + ], + "spans": [ + { + "bbox": [ + 303, + 177, + 343, + 182 + ], + "type": "text", + "content": "sky over backpack" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 183, + 333, + 189 + ], + "lines": [ + { + "bbox": [ + 303, + 183, + 333, + 189 + ], + "spans": [ + { + "bbox": [ + 303, + 183, + 333, + 189 + ], + "type": "text", + "content": "sky over rock" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 190, + 351, + 195 + ], + "lines": [ + { + "bbox": [ + 303, + 190, + 351, + 195 + ], + "spans": [ + { + "bbox": [ + 303, + 190, + 351, + 195 + ], + "type": "text", + "content": "person climbing snow" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 196, + 355, + 201 + ], + "lines": [ + { + "bbox": [ + 303, + 196, + 355, + 201 + ], + "spans": [ + { + "bbox": [ + 303, + 196, + 355, + 201 + ], + "type": "text", + "content": "person walking on snow" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 202, + 354, + 207 + ], + "lines": [ + { + "bbox": [ + 303, + 202, + 354, + 207 + ], + "spans": [ + { + "bbox": [ + 303, + 202, + 354, + 207 + ], + "type": "text", + "content": "personbesideperson" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 208, + 337, + 213 + ], + "lines": [ + { + "bbox": [ + 303, + 208, + 337, + 213 + ], + "spans": [ + { + "bbox": [ + 303, + 208, + 337, + 213 + ], + "type": "text", + "content": "person on snow" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 213, + 354, + 219 + ], + "lines": [ + { + "bbox": [ + 303, + 213, + 354, + 219 + ], + "spans": [ + { + "bbox": [ + 303, + 213, + 354, + 219 + ], + "type": "text", + "content": "person running on snow" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 220, + 354, + 225 + ], + "lines": [ + { + "bbox": [ + 303, + 220, + 354, + 225 + ], + "spans": [ + { + "bbox": [ + 303, + 220, + 354, + 225 + ], + "type": "text", + "content": "person picking skis" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 226, + 344, + 232 + ], + "lines": [ + { + "bbox": [ + 303, + 226, + 344, + 232 + ], + "spans": [ + { + "bbox": [ + 303, + 226, + 344, + 232 + ], + "type": "text", + "content": "person playing skills" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 232, + 353, + 238 + ], + "lines": [ + { + "bbox": [ + 303, + 232, + 353, + 238 + ], + "spans": [ + { + "bbox": [ + 303, + 232, + 353, + 238 + ], + "type": "text", + "content": "person playing with person attached to skis" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 239, + 364, + 244 + ], + "lines": [ + { + "bbox": [ + 303, + 239, + 364, + 244 + ], + "spans": [ + { + "bbox": [ + 303, + 239, + 364, + 244 + ], + "type": "text", + "content": "person attached to backpack" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 245, + 344, + 251 + ], + "lines": [ + { + "bbox": [ + 303, + 245, + 344, + 251 + ], + "spans": [ + { + "bbox": [ + 303, + 245, + 344, + 251 + ], + "type": "text", + "content": "mountain attached" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 251, + 344, + 257 + ], + "lines": [ + { + "bbox": [ + 303, + 251, + 344, + 257 + ], + "spans": [ + { + "bbox": [ + 303, + 251, + 344, + 257 + ], + "type": "text", + "content": "sky over mountain" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 258, + 344, + 262 + ], + "lines": [ + { + "bbox": [ + 303, + 258, + 344, + 262 + ], + "spans": [ + { + "bbox": [ + 303, + 258, + 344, + 262 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 225, + 228, + 264, + 233 + ], + "blocks": [ + { + "bbox": [ + 225, + 228, + 264, + 233 + ], + "lines": [ + { + "bbox": [ + 225, + 228, + 264, + 233 + ], + "spans": [ + { + "bbox": [ + 225, + 228, + 264, + 233 + ], + "type": "image", + "image_path": "27e21a5518eb23cf7104d3a1cfe7470c4299507d893ece5625e4e1a14f376a3b.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 234, + 279, + 239 + ], + "lines": [ + { + "bbox": [ + 223, + 234, + 279, + 239 + ], + "spans": [ + { + "bbox": [ + 223, + 234, + 279, + 239 + ], + "type": "text", + "content": "person carrying backpack" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 240, + 279, + 246 + ], + "lines": [ + { + "bbox": [ + 223, + 240, + 279, + 246 + ], + "spans": [ + { + "bbox": [ + 223, + 240, + 279, + 246 + ], + "type": "text", + "content": "person holding skis" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 247, + 272, + 252 + ], + "lines": [ + { + "bbox": [ + 223, + 247, + 272, + 252 + ], + "spans": [ + { + "bbox": [ + 223, + 247, + 272, + 252 + ], + "type": "text", + "content": "person climbing snow." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 253, + 256, + 258 + ], + "lines": [ + { + "bbox": [ + 223, + 253, + 256, + 258 + ], + "spans": [ + { + "bbox": [ + 223, + 253, + 256, + 258 + ], + "type": "text", + "content": "sky over snow." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 299, + 75, + 375, + 129 + ], + "blocks": [ + { + "bbox": [ + 299, + 75, + 375, + 129 + ], + "lines": [ + { + "bbox": [ + 299, + 75, + 375, + 129 + ], + "spans": [ + { + "bbox": [ + 299, + 75, + 375, + 129 + ], + "type": "image", + "image_path": "b72dddcb15b5e69afbb1e14f07a0b77555faa439c1b6be332ba58fc38efda690.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 133, + 359, + 138 + ], + "lines": [ + { + "bbox": [ + 302, + 133, + 359, + 138 + ], + "spans": [ + { + "bbox": [ + 302, + 133, + 359, + 138 + ], + "type": "text", + "content": "person carrying backpack" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 139, + 335, + 144 + ], + "lines": [ + { + "bbox": [ + 302, + 139, + 335, + 144 + ], + "spans": [ + { + "bbox": [ + 302, + 139, + 335, + 144 + ], + "type": "text", + "content": "person holding" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 146, + 321, + 151 + ], + "lines": [ + { + "bbox": [ + 303, + 146, + 321, + 151 + ], + "spans": [ + { + "bbox": [ + 303, + 146, + 321, + 151 + ], + "type": "text", + "content": "skis on :" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 386, + 76, + 462, + 129 + ], + "blocks": [ + { + "bbox": [ + 386, + 76, + 462, + 129 + ], + "lines": [ + { + "bbox": [ + 386, + 76, + 462, + 129 + ], + "spans": [ + { + "bbox": [ + 386, + 76, + 462, + 129 + ], + "type": "image", + "image_path": "3ba5ceb2c1ff682fe10946d74229fb33427e75719f3bb82142b829b78e8bf237.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 133, + 430, + 138 + ], + "lines": [ + { + "bbox": [ + 420, + 133, + 430, + 138 + ], + "spans": [ + { + "bbox": [ + 420, + 133, + 430, + 138 + ], + "type": "text", + "content": "↑" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 139, + 431, + 145 + ], + "lines": [ + { + "bbox": [ + 417, + 139, + 431, + 145 + ], + "spans": [ + { + "bbox": [ + 417, + 139, + 431, + 145 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 409, + 147, + 459, + 152 + ], + "lines": [ + { + "bbox": [ + 409, + 147, + 459, + 152 + ], + "spans": [ + { + "bbox": [ + 409, + 147, + 459, + 152 + ], + "type": "text", + "content": "Prediction " + }, + { + "bbox": [ + 409, + 147, + 459, + 152 + ], + "type": "inline_equation", + "content": "\\rightarrow" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 386, + 172, + 462, + 224 + ], + "blocks": [ + { + "bbox": [ + 410, + 154, + 441, + 160 + ], + "lines": [ + { + "bbox": [ + 410, + 154, + 441, + 160 + ], + "spans": [ + { + "bbox": [ + 410, + 154, + 441, + 160 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 161, + 425, + 166 + ], + "lines": [ + { + "bbox": [ + 424, + 161, + 425, + 166 + ], + "spans": [ + { + "bbox": [ + 424, + 161, + 425, + 166 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 172, + 462, + 224 + ], + "lines": [ + { + "bbox": [ + 386, + 172, + 462, + 224 + ], + "spans": [ + { + "bbox": [ + 386, + 172, + 462, + 224 + ], + "type": "image", + "image_path": "bbcb12245522c272c26d46ee07348c4652472801a13605622700bc61e93907c1.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 152, + 506, + 157 + ], + "lines": [ + { + "bbox": [ + 469, + 152, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 469, + 152, + 506, + 157 + ], + "type": "text", + "content": "chair beside wall" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 158, + 504, + 163 + ], + "lines": [ + { + "bbox": [ + 469, + 158, + 504, + 163 + ], + "spans": [ + { + "bbox": [ + 469, + 158, + 504, + 163 + ], + "type": "text", + "content": "chair in front of" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 164, + 519, + 170 + ], + "lines": [ + { + "bbox": [ + 469, + 164, + 519, + 170 + ], + "spans": [ + { + "bbox": [ + 469, + 164, + 519, + 170 + ], + "type": "text", + "content": "person in front of chair" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 171, + 506, + 176 + ], + "lines": [ + { + "bbox": [ + 469, + 171, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 469, + 171, + 506, + 176 + ], + "type": "text", + "content": "person looking at" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 177, + 504, + 182 + ], + "lines": [ + { + "bbox": [ + 469, + 177, + 504, + 182 + ], + "spans": [ + { + "bbox": [ + 469, + 177, + 504, + 182 + ], + "type": "text", + "content": "person throwing" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 183, + 513, + 189 + ], + "lines": [ + { + "bbox": [ + 469, + 183, + 513, + 189 + ], + "spans": [ + { + "bbox": [ + 469, + 183, + 513, + 189 + ], + "type": "text", + "content": "wall in front of tree" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 492, + 190, + 515, + 194 + ], + "lines": [ + { + "bbox": [ + 492, + 190, + 515, + 194 + ], + "spans": [ + { + "bbox": [ + 492, + 190, + 515, + 194 + ], + "type": "text", + "content": "over chair" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 195, + 533, + 201 + ], + "lines": [ + { + "bbox": [ + 469, + 195, + 533, + 201 + ], + "spans": [ + { + "bbox": [ + 469, + 195, + 533, + 201 + ], + "type": "text", + "content": "person swinging tennis racket" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 201, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 469, + 201, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 469, + 201, + 504, + 207 + ], + "type": "text", + "content": "treebesidewall" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 208, + 511, + 213 + ], + "lines": [ + { + "bbox": [ + 469, + 208, + 511, + 213 + ], + "spans": [ + { + "bbox": [ + 469, + 208, + 511, + 213 + ], + "type": "text", + "content": "chair beside person" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 213, + 515, + 219 + ], + "lines": [ + { + "bbox": [ + 469, + 213, + 515, + 219 + ], + "spans": [ + { + "bbox": [ + 469, + 213, + 515, + 219 + ], + "type": "text", + "content": "chair attached to wall" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 220, + 525, + 225 + ], + "lines": [ + { + "bbox": [ + 469, + 220, + 525, + 225 + ], + "spans": [ + { + "bbox": [ + 469, + 220, + 525, + 225 + ], + "type": "text", + "content": "person holding sports ball" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_caption" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 389, + 226, + 462, + 232 + ], + "blocks": [ + { + "bbox": [ + 389, + 226, + 462, + 232 + ], + "lines": [ + { + "bbox": [ + 389, + 226, + 462, + 232 + ], + "spans": [ + { + "bbox": [ + 389, + 226, + 462, + 232 + ], + "type": "image", + "image_path": "84a9ce5306300de517c32d31386355b13b59521e84e6b23a884c79dc5296d51e.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 226, + 501, + 231 + ], + "lines": [ + { + "bbox": [ + 469, + 226, + 501, + 231 + ], + "spans": [ + { + "bbox": [ + 469, + 226, + 501, + 231 + ], + "type": "text", + "content": "mehohid" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 232, + 515, + 236 + ], + "lines": [ + { + "bbox": [ + 469, + 232, + 515, + 236 + ], + "spans": [ + { + "bbox": [ + 469, + 232, + 515, + 236 + ], + "type": "text", + "content": "person beside barrier" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 237, + 504, + 242 + ], + "lines": [ + { + "bbox": [ + 469, + 237, + 504, + 242 + ], + "spans": [ + { + "bbox": [ + 469, + 237, + 504, + 242 + ], + "type": "text", + "content": "wall beside tree" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 243, + 524, + 247 + ], + "lines": [ + { + "bbox": [ + 469, + 243, + 524, + 247 + ], + "spans": [ + { + "bbox": [ + 469, + 243, + 524, + 247 + ], + "type": "text", + "content": "person playing sports ball" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_caption" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 389, + 232, + 455, + 237 + ], + "blocks": [ + { + "bbox": [ + 389, + 232, + 455, + 237 + ], + "lines": [ + { + "bbox": [ + 389, + 232, + 455, + 237 + ], + "spans": [ + { + "bbox": [ + 389, + 232, + 455, + 237 + ], + "type": "image", + "image_path": "09ae112509d1f1e09e86d6acf0847d76b3b96580a4c7206e49724a8b5642d9ac.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 406, + 238, + 425, + 243 + ], + "blocks": [ + { + "bbox": [ + 406, + 238, + 425, + 243 + ], + "lines": [ + { + "bbox": [ + 406, + 238, + 425, + 243 + ], + "spans": [ + { + "bbox": [ + 406, + 238, + 425, + 243 + ], + "type": "image", + "image_path": "7c3d0874c90c90775a6ba793c8315b1e8f848e82cf2f9303f20d2b161763aaa0.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 389, + 243, + 455, + 249 + ], + "blocks": [ + { + "bbox": [ + 389, + 243, + 455, + 249 + ], + "lines": [ + { + "bbox": [ + 389, + 243, + 455, + 249 + ], + "spans": [ + { + "bbox": [ + 389, + 243, + 455, + 249 + ], + "type": "image", + "image_path": "da702fde1726af0638b1d5d7dff8d28d1df8af4307319e948810c447889d64f2.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 248, + 501, + 252 + ], + "lines": [ + { + "bbox": [ + 469, + 248, + 501, + 252 + ], + "spans": [ + { + "bbox": [ + 469, + 248, + 501, + 252 + ], + "type": "text", + "content": "tree in front of" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "image_caption" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 389, + 250, + 455, + 255 + ], + "blocks": [ + { + "bbox": [ + 389, + 250, + 455, + 255 + ], + "lines": [ + { + "bbox": [ + 389, + 250, + 455, + 255 + ], + "spans": [ + { + "bbox": [ + 389, + 250, + 455, + 255 + ], + "type": "image", + "image_path": "fd226a19c1fb0ca712398d39c7fb7114f94d2d1d042de71a44bda5b97bde5f41.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 255, + 436, + 260 + ], + "lines": [ + { + "bbox": [ + 389, + 255, + 436, + 260 + ], + "spans": [ + { + "bbox": [ + 389, + 255, + 436, + 260 + ], + "type": "text", + "content": "chair leaning on wall" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 253, + 501, + 258 + ], + "lines": [ + { + "bbox": [ + 469, + 253, + 501, + 258 + ], + "spans": [ + { + "bbox": [ + 469, + 253, + 501, + 258 + ], + "type": "text", + "content": "tree over wall" + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 265, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 547, + 310 + ], + "type": "text", + "content": "Figure 3. Visualization of panoptic segmentations and the top 20 predicted triplets compared with ground truth. The upper left is the original image, the lower left is the ground truth and on the right are the predictions. The highlighted triplets represent the subject-object pairs with multiple relations, where the blue highlights represent the high frequency relations and the red highlights represent the low frequency relations. The visualization results show that our method can predict both high frequency and low frequency relations." + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_caption" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 465, + 76, + 541, + 129 + ], + "blocks": [ + { + "bbox": [ + 465, + 76, + 541, + 129 + ], + "lines": [ + { + "bbox": [ + 465, + 76, + 541, + 129 + ], + "spans": [ + { + "bbox": [ + 465, + 76, + 541, + 129 + ], + "type": "image", + "image_path": "869b41a2d909c1c603bae50fa873afa3ce9122612b09b7cafee5211987245d8c.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 133, + 502, + 138 + ], + "lines": [ + { + "bbox": [ + 485, + 133, + 502, + 138 + ], + "spans": [ + { + "bbox": [ + 485, + 133, + 502, + 138 + ], + "type": "text", + "content": "holding" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 139, + 501, + 144 + ], + "lines": [ + { + "bbox": [ + 485, + 139, + 501, + 144 + ], + "spans": [ + { + "bbox": [ + 485, + 139, + 501, + 144 + ], + "type": "text", + "content": "about to" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 146, + 518, + 151 + ], + "lines": [ + { + "bbox": [ + 469, + 146, + 518, + 151 + ], + "spans": [ + { + "bbox": [ + 469, + 146, + 518, + 151 + ], + "type": "text", + "content": "person in front of wall" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_caption" + } + ], + "index": 79 + }, + { + "type": "table", + "bbox": [ + 50, + 316, + 285, + 387 + ], + "blocks": [ + { + "bbox": [ + 50, + 316, + 285, + 387 + ], + "lines": [ + { + "bbox": [ + 50, + 316, + 285, + 387 + ], + "spans": [ + { + "bbox": [ + 50, + 316, + 285, + 387 + ], + "type": "table", + "html": "
ObjectRelationMarginR/mR@50R/mR@100
✓✓0.540.7 / 30.343.0 / 33.1
-✓0.540.6 / 29.742.8 / 32.8
-✓0.040.5 / 29.542.7 / 32.8
✓--40.4 / 29.042.8 / 32.2
---39.7 / 28.642.4 / 32.0
", + "image_path": "0b2033ee9843e9403dcaddf9f9b0517c08abc8824e0f4ffc3e749950400b1fac.jpg" + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "table_body" + } + ], + "index": 102 + }, + { + "type": "table", + "bbox": [ + 50, + 440, + 285, + 475 + ], + "blocks": [ + { + "bbox": [ + 46, + 393, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 437 + ], + "type": "text", + "content": "Table 5. Ablation study for different losses in HiLo prediction alignment. Object refers to the subject-object consistency loss and relation refers to the relation consistency loss. The margin parameter is defined in Eq. 6." + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 440, + 285, + 475 + ], + "lines": [ + { + "bbox": [ + 50, + 440, + 285, + 475 + ], + "spans": [ + { + "bbox": [ + 50, + 440, + 285, + 475 + ], + "type": "table", + "html": "
Whether to use RIER/mR@20R/mR@50R/mR@100
✓34.1 / 23.740.7 / 30.343.0 / 33.1
-33.5 / 22.340.3 / 29.042.6 / 32.3
", + "image_path": "54d3f5aa593018d6d151c53be9d9465cf28c65b1d8b9be8789cba17f2af78f52.jpg" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "table_body" + } + ], + "index": 104 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "We also attempted to average the tensor generated by the two branches and obtain the PSG result through postprocessing. However, we found that this approach leads to a substantial drop in performance, as evident in Tab. 8. This can be attributed to the inconsistent prediction results of the two branches for the same query index. These findings validate that the inference fusion method effectively merges the results from the two branches. Furthermore, our experimental results demonstrate that the query associated with the identical index in two branches does not predict the same subject-object pair. Thus, directly averaging the tensor produced by the two branches results in prediction ambiguity, ultimately leading to a substantial decline in performance. This observation underscores the necessity of conducting triplet query correspondence when performing prediction alignment. In particular, due to the inconsistent query prediction content for the corresponding index" + } + ] + } + ], + "index": 106 + }, + { + "type": "table", + "bbox": [ + 309, + 316, + 543, + 367 + ], + "blocks": [ + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 503 + ], + "type": "text", + "content": "Table 6. Ablation study for relation consistency loss in HiLo prediction alignment." + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 316, + 543, + 367 + ], + "lines": [ + { + "bbox": [ + 309, + 316, + 543, + 367 + ], + "spans": [ + { + "bbox": [ + 309, + 316, + 543, + 367 + ], + "type": "table", + "html": "
H-L ResultL-H ResultR/mR@50R/mR@100
✓✓40.7 / 30.343.0 / 33.1
✓-38.8 / 29.939.8 / 30.9
-✓38.5 / 26.539.5 / 27.8
", + "image_path": "61bc5568454d8f476932592fb4cd781cd4f9b43fba279802867a9f3ae3be64bc.jpg" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "table_body" + } + ], + "index": 107 + }, + { + "type": "table", + "bbox": [ + 309, + 396, + 543, + 433 + ], + "blocks": [ + { + "bbox": [ + 334, + 373, + 517, + 384 + ], + "lines": [ + { + "bbox": [ + 334, + 373, + 517, + 384 + ], + "spans": [ + { + "bbox": [ + 334, + 373, + 517, + 384 + ], + "type": "text", + "content": "Table 7. Ablation study for HiLo inference fusion." + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 396, + 543, + 433 + ], + "lines": [ + { + "bbox": [ + 309, + 396, + 543, + 433 + ], + "spans": [ + { + "bbox": [ + 309, + 396, + 543, + 433 + ], + "type": "table", + "html": "
Fusion methodR/mR@20R/mR@50R/mR@100
inference fusion34.1 / 23.740.7 / 30.343.0 / 33.1
average tensor19.6 / 13.123.1 / 15.723.9 / 16.3
", + "image_path": "c9c4e69dcc77114d4e305e0a0db0030f2b57ccc8e6c013798bdd472a9a433375.jpg" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "table_body" + } + ], + "index": 109 + }, + { + "type": "table", + "bbox": [ + 309, + 493, + 543, + 531 + ], + "blocks": [ + { + "bbox": [ + 305, + 437, + 546, + 483 + ], + "lines": [ + { + "bbox": [ + 305, + 437, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 546, + 483 + ], + "type": "text", + "content": "Table 8. Ablation study for HiLo inference fusion. Inference fusion refers to the method proposed in the paper to fuse the results of two branches. Average tensor refers to the fusion method that directly averages the tensor output by the two branches." + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 493, + 543, + 531 + ], + "lines": [ + { + "bbox": [ + 309, + 493, + 543, + 531 + ], + "spans": [ + { + "bbox": [ + 309, + 493, + 543, + 531 + ], + "type": "table", + "html": "
Attention focusR/mR@20R/mR@50R/mR@100
subject-object32.6 / 20.938.0 / 27.438.9 / 28.4
full image30.4 / 19.336.5 / 25.937.1 / 26.8
", + "image_path": "c941d116aed8531b044c6bb9595facaba62e23c26f1c3c0b5cff7cee68fde276.jpg" + } + ] + } + ], + "index": 111, + "angle": 0, + "type": "table_body" + } + ], + "index": 111 + }, + { + "bbox": [ + 328, + 536, + 523, + 548 + ], + "lines": [ + { + "bbox": [ + 328, + 536, + 523, + 548 + ], + "spans": [ + { + "bbox": [ + 328, + 536, + 523, + 548 + ], + "type": "text", + "content": "Table 9. Ablation study for masked relation attention." + } + ] + } + ], + "index": 112, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 552, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 552, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 552, + 545, + 588 + ], + "type": "text", + "content": "in the two branches, a one-to-one correspondence must be constructed based on the label assigned by each query to achieve prediction alignment." + } + ] + } + ], + "index": 113 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Masked relation attention. We investigate the impact of different mask input types for cross-attention on the HiLo baseline (Sec. 3.2) performance. Specifically, we compare two different attention focus regions, namely the subject-object region and the full image. The results are shown in Tab. 9. Focusing on the full image presents a more challenging optimization task for the model since no target region is specified. Consequently, we observe a drop of " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " in R@100 and " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " in mR@100. This shows that it is crucial to apply masked relation attention to the subject and object." + } + ] + } + ], + "index": 114 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21644" + } + ] + } + ], + "index": 115 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 285, + 100 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 285, + 100 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 285, + 100 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 285, + 100 + ], + "type": "table", + "html": "
MethodAllRareAllOverlap
PSGTR22.16.222.123.5
HiLo (ours)33.1 (+11.0)20.3 (+14.1)33.1 (+11.0)38.8 (+15.3)
", + "image_path": "d3552fdcf427a91afb71c2ad0ba537c196e0af78455bf9b98cf0e8739a0bf80b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 152, + 285, + 289 + ], + "blocks": [ + { + "bbox": [ + 50, + 152, + 285, + 289 + ], + "lines": [ + { + "bbox": [ + 50, + 152, + 285, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 152, + 285, + 289 + ], + "type": "image", + "image_path": "d374e4df8be64179687aad0660ee28d0f4f8dcb574a20ef3b7d20b31074ce94a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 297, + 286, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 297, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 286, + 319 + ], + "type": "text", + "content": "Figure 4. Convergence speed analysis of different methods. Our method converges significantly faster than previous methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 110, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 110, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 110, + 335 + ], + "type": "text", + "content": "4.6. Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 286, + 365 + ], + "type": "text", + "content": "To further demonstrate the efficacy of our method, we conduct the following analysis on HiLo." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "content": "Long-tail problem and relational semantic overlap. To verify whether the problems of long-tail and relational semantic overlap have been tackled, we conduct experiments shown in Tab. 10. For the long-tail problem, we consider relations appearing less than 500 times (28 relations) in the PSG dataset (relations appear on average 4787 times) as rare relations, and report their mR@100. There is a " + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "content": " improvement for rare relations and an " + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 46, + 365, + 286, + 484 + ], + "type": "text", + "content": " improvement for all relations. This suggests that our method addresses the long-tail problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 568 + ], + "type": "text", + "content": "For the relational semantic overlap, we select all test images that have this problem (927 images) and report their mR@100. Our method shows a " + }, + { + "bbox": [ + 46, + 486, + 286, + 568 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 568 + ], + "type": "text", + "content": " improvement over PSGTR on images with semantic overlap relations, exceeding the overall improvement across all test images. This suggests that our approach addresses the problem of relational semantic overlap." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 689 + ], + "type": "text", + "content": "Convergence speed and time cost analysis. We evaluate the convergence of our model by assessing its performance on the validation set at various epochs, as illustrated in Fig. 4. Our analysis reveals that our proposed method outperforms prior methods [62] both in terms of final performance and convergence speed. Specifically, PSGTR [62] achieves negligible performance in the initial 12 epochs, requiring 60 epochs to converge, as per the authors. In contrast, our HiLo method achieves better results in just 12 epochs, indicating its superior convergence speed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "For time cost, we primarily analyze relation augmentation and swapping. 1) Relation augmentation, inspired by" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 100 + ], + "blocks": [ + { + "bbox": [ + 47, + 106, + 286, + 150 + ], + "lines": [ + { + "bbox": [ + 47, + 106, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 286, + 150 + ], + "type": "text", + "content": "Table 10. Verification of the improvements in the long-tail problem and relational semantic overlap. All refers to testing on the whole test set. Rare refers to testing only on rare relations. Overlap refers to testing only on data with relation semantic overlap." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 544, + 100 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 100 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 100 + ], + "type": "table", + "html": "
MethodGFlopsParam. (M)Train Mem. (G)Infer. Time (ms)
PSGTR461.344.226.5140
HiLo (ours)229.458.716.1156
", + "image_path": "6c5a64a04850bc738007a3fd5e40358837002b7bd5db9145c65b201ff9b6366d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 106, + 494, + 117 + ], + "lines": [ + { + "bbox": [ + 356, + 106, + 494, + 117 + ], + "spans": [ + { + "bbox": [ + 356, + 106, + 494, + 117 + ], + "type": "text", + "content": "Table 11. Training and inference cost." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 119, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 119, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 119, + 545, + 286 + ], + "type": "text", + "content": "IETrans [66], involves training a baseline model and then using it to predict relation labels so as to augment the original relation labels. For the PSG dataset, following our experimental setup (see Sec. 4.3), this takes 18 hours. Afterwards, the training of our HiLo model requires 18 hours, which makes the whole process 36 hours. In contrast, the PSGTR model training takes 48 hours. Both our method and PSGTR utilize ResNet-50 as the backbone for fair comparison. Our method's rapid convergence (see Sec. 4.5) reduces training time compared to PSGTR, making the additional time cost for relation augmentation tolerable. 2) Relation swapping is a quick operation on relation labels during training and does not significantly contribute to overall time consumption." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 287, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 418 + ], + "type": "text", + "content": "Training and inference cost. Training and inference cost is shown in Tab. 11. Despite using two transformer-based decoders, our method requires less resource. Given the same input sizes (1280, 800) and ResNet-50 as backbone, the resource usage is shown in Tab. 11. PSGTR [62] generates mask features for subject and object separately, while our method reduces this computation by only generating mask features once. Our inference time is marginally higher due to more complex post-processing. Using H-L and L-H data only changes the labels between two branches, not increasing resource use." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 423, + 378, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 378, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 378, + 435 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 443, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 611 + ], + "type": "text", + "content": "In this work we proposed the HiLo framework to tackle the long-tail problem with relational semantic overlap in Panoptic Scene Graph generation. The HiLo framework simultaneously learns the high and low frequency relations in different network branches and unifies their strengths by aligning their predictions. We also constructed a HiLo baseline to allow high-quality panoptic segmentation to improve PSG performance. Experimental results demonstrate that our method achieves state-of-the-art performance on the PSG dataset, confirming its effectiveness. In future work, we will investigate how knowledge distillation [28, 23] can be used to fuse the high and low branches in our method, as well as its application to downstream tasks such as visual question answering and image captioning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 616, + 399, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 399, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 399, + 628 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "The authors would like to thank Prof. Tomasz Radzik for helpful discussions. Computing resources provided by King's Computational Research, Engineering and Technology Environment (CREATE). This work was supported by the European Union's Horizon 2020 FET Proactive Program under Agreement 101017857 and Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21645" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 173:33-45, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 201 + ], + "type": "text", + "content": "[2] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, pages 353-369. Springer, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 267 + ], + "type": "text", + "content": "[3] Stephan Baier, Yunpu Ma, and Volker Tresp. Improving visual relationship detection using semantic modeling of scene descriptions. In The Semantic Web-ISWC 2017: 16th International Semantic Web Conference, Vienna, Austria, October 21-25, 2017, Proceedings, Part I 16, pages 53-68. Springer, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "text", + "content": "[5] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 423 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 479 + ], + "type": "text", + "content": "[7] Long Chen, Hanwang Zhang, Jun Xiao, Xiangnan He, Shiliang Pu, and Shih-Fu Chang. Counterfactual critic multiagent training for scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4613-4623, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "text", + "content": "[8] Shizhe Chen, Qin Jin, Peng Wang, and Qi Wu. Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9962-9971, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 288, + 590 + ], + "type": "text", + "content": "[9] Tianshui Chen, Weihao Yu, Riquan Chen, and Liang Lin. Knowledge-embedded routing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6163–6171, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 646 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[11] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[12] Meng-Jiun Chiou, Henghui Ding, Hanshu Yan, Changhu Wang, Roger Zimmermann, and Jiashi Feng. Recovering the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "unbiased scene graphs from the biased ones. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1581-1590, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 150 + ], + "type": "text", + "content": "[13] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[14] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 547, + 297 + ], + "type": "text", + "content": "[16] Bo Dai, Yuqi Zhang, and Dahua Lin. Detecting visual relationships with deep relational networks. In Proceedings of the IEEE conference on computer vision and Pattern recognition, pages 3076-3086, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 299, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 365 + ], + "type": "text", + "content": "[17] Youming Deng, Yansheng Li, Yongjun Zhang, Xiang Xiang, Jian Wang, Jingdong Chen, and Jiayi Ma. Hierarchical memory learning for fine-grained scene graph generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 266-283. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "text", + "content": "[18] Alakh Desai, Tz-Ying Wu, Subarna Tripathi, and Nuno Vasconcelos. Learning of visual relations: The devil is in the tails. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15404-15413, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "text", + "content": "[19] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 488, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 543 + ], + "type": "text", + "content": "[20] Mohammed Haroon Dupty, Zhen Zhang, and Wee Sun Lee. Visual relationship detection with low rank non-negative tensor decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10737-10744, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 544, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 547, + 589 + ], + "type": "text", + "content": "[21] Lizhao Gao, Bo Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the 2018 10th international conference on machine learning and computing, pages 225–229, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 644 + ], + "type": "text", + "content": "[22] Arushi Goel, Basura Fernando, Frank Keller, and Hakan Bilen. Not all relations are equal: Mining informative labels for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15596-15606, 2022. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[23] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 714 + ], + "type": "text", + "content": "[24] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21646" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 1969-1978, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[25] Yuyu Guo, Lianli Gao, Xuanhan Wang, Yuxuan Hu, Xing Xu, Xu Lu, Heng Tao Shen, and Jingkuan Song. From general to specific: Informative scene graph generation via balance adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16383-16392, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "text", + "content": "[27] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 274 + ], + "type": "text", + "content": "[28] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 331 + ], + "type": "text", + "content": "[29] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 386 + ], + "type": "text", + "content": "[30] Zih-Siou Hung, Arun Mallya, and Svetlana Lazebnik. Contextual translation embedding for visual relationship detection and scene graph generation. IEEE transactions on pattern analysis and machine intelligence, 43(11):3820-3832, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 443 + ], + "type": "text", + "content": "[31] Seong Jae Hwang, Sathya N Ravi, Zirui Tao, Hyunwoo J Kim, Maxwell D Collins, and Vikas Singh. Tensorize, factorize and regularize: Robust visual relationship learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1014-1023, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[32] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3668–3678, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[33] Haeyong Kang and Chang D Yoo. Skew class-balanced reweighting for unbiased scene graph generation. Machine Learning and Knowledge Extraction, 5(1):287-303, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[34] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15879-15889, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9404-9413, 2019. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[36] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[37] Lin Li, Long Chen, Yifeng Huang, Zhimeng Zhang, Songyang Zhang, and Jun Xiao. The devil is in the labels:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Noisy label correction for robust scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18869-18878, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 152 + ], + "type": "text", + "content": "[38] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19486-19496, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "text", + "content": "[39] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[40] Wei Li, Haiwei Zhang, Qijie Bai, Guoqing Zhao, Ning Jiang, and Xiaojie Yuan. Ppdl: Predicate probability distribution based loss for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[41] Yikang Li, Wanli Ouyang, Xiaogang Wang, and Xiao'ou Tang. Vip-cnn: Visual phrase guided convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1347-1356, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "text", + "content": "[42] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE international conference on computer vision, pages 1261-1270, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "text", + "content": "[43] Wentong Liao, Bodo Rosenhahn, Ling Shuai, and Michael Ying Yang. Natural language guided visual relationship detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "text", + "content": "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "text", + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 599 + ], + "type": "text", + "content": "[46] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 656 + ], + "type": "text", + "content": "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[48] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[49] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21647" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 852-869. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "text", + "content": "[50] Mengshi Qi, Weijian Li, Zhengyuan Yang, Yunhong Wang, and Jiebo Luo. Attentive relational networks for mapping images to scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3957-3966, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "text", + "content": "[51] Mengshi Qi, Yunhong Wang, and Annan Li. Online cross-modal scene retrieval by binary representation and semantic graph. In Proceedings of the 25th ACM international conference on Multimedia, pages 744-752, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 231 + ], + "type": "text", + "content": "[52] Mohammad Amin Sadeghi and Ali Farhadi. Recognition using visual phrases. IEEE, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 287 + ], + "type": "text", + "content": "[53] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 289, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 342 + ], + "type": "text", + "content": "[54] Sahand Sharifzadeh, Sina Moayed Baharlou, Max Berrendorf, Rajat Koner, and Volker Tresp. Improving visual relation detection using depth maps. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 3597-3604. IEEE, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "type": "text", + "content": "[55] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 390, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 390, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 390, + 287, + 456 + ], + "type": "text", + "content": "[56] Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, et al. Relationformer: A unified framework for image-to-graph generation. In European Conference on Computer Vision, pages 422-439. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 457, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 287, + 501 + ], + "type": "text", + "content": "[57] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 502, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 601 + ], + "type": "text", + "content": "[58] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MICCAI 2017, Quebec City, QC, Canada, September 14, Proceedings 3, pages 240-248. Springer, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[59] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[60] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 2, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[61] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[62] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 1, 2, 4, 6, 7, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 163, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 239 + ], + "type": "text", + "content": "[63] Jing Yu, Yuan Chai, Yujing Wang, Yue Hu, and Qi Wu. Cogtree: Cognition tree loss for unbiased scene graph generation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1274–1280. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "type": "text", + "content": "[64] Xu Chen Ya Zhang Xiao Gu Yue Hu, Siheng Chen. Neural message passing for visual relationship detection. In ICML Workshop on Learning and Reasoning with Graph-Structured Representations, Long Beach, CA, June 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 285, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 338 + ], + "type": "text", + "content": "[65] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 341, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 545, + 407 + ], + "type": "text", + "content": "[66] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 409-424. Springer, 2022. 2, 4, 7, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 407, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 407, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 308, + 407, + 545, + 460 + ], + "type": "text", + "content": "[67] Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, and TatSeng Chua. Visual translation embedding network for visual relation detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5532-5540, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 462, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 462, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 462, + 545, + 518 + ], + "type": "text", + "content": "[68] Ji Zhang, Yannis Kalantidis, Marcus Rohrbach, Manohar Paluri, Ahmed Elgammal, and Mohamed Elhoseiny. Large-scale visual relationship understanding. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9185-9194, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 519, + 545, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 562 + ], + "type": "text", + "content": "[69] Sipeng Zheng, Shizhe Chen, and Qin Jin. Visual relation detection with multi-level attention. In Proceedings of the 27th ACM International Conference on Multimedia, pages 121-129, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 563, + 545, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 563, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 308, + 563, + 545, + 617 + ], + "type": "text", + "content": "[70] Guangming Zhu, Liang Zhang, Youliang Jiang, Yixuan Dang, Haoran Hou, Peiyi Shen, Mingtao Feng, Xia Zhao, Qiguang Miao, Syed Afaq Ali Shah, et al. Scene graph generation: A comprehensive survey. arXiv preprint arXiv:2201.00443, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 619, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 619, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 619, + 545, + 662 + ], + "type": "text", + "content": "[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable " + }, + { + "bbox": [ + 308, + 619, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{detr}\\}" + }, + { + "bbox": [ + 308, + 619, + 545, + 662 + ], + "type": "text", + "content": ": Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 664, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 664, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 664, + 545, + 708 + ], + "type": "text", + "content": "[72] Yaohui Zhu, Shuqiang Jiang, and Xiangyang Li. Visual relationship detection with object spatial distribution. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 379-384. IEEE, 2017. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21648" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_content_list.json b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44f8fe1861d9a37cc0f26d269fffec653df949be --- /dev/null +++ b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_content_list.json @@ -0,0 +1,1613 @@ +[ + { + "type": "text", + "text": "HiTeA: Hierarchical Temporal-Aware Video-Language Pre-training", + "text_level": 1, + "bbox": [ + 142, + 89, + 826, + 112 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qinghao Ye Guohai Xu Ming Yan* Haiyang Xu* Qi Qian Ji Zhang Fei Huang", + "bbox": [ + 264, + 119, + 702, + 157 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DAMO Academy, Alibaba Group yeqinghao.yqh@alibaba-inc.com", + "bbox": [ + 349, + 172, + 620, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 241, + 313, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-language pre-training has advanced the performance of various downstream video-language tasks. However, most previous methods directly inherit or adapt typical image-language pre-training paradigms to video-language pre-training, thus not fully exploiting the unique characteristic of video, i.e., temporal. In this paper, we propose a Hierarchical Temporal-Aware video-language pretraining framework, HiTeA, with two novel pre-training tasks for yielding temporal-aware multi-modal representation with cross-modal fine-grained temporal moment information and temporal contextual relations between video-text multi-modal pairs. First, we propose a cross-modal moment exploration task to explore moments in videos by mining the paired texts, which results in detailed video moment representation. Then, based on the learned detailed moment representations, the inherent temporal contextual relations are captured by aligning video-text pairs as a whole in different time resolutions with multi-modal temporal relation exploration task. Furthermore, we introduce the shuffling test to evaluate the temporal reliance of datasets and video-language pre-training models. We achieve state-of-the-art results on 15 well-established video-language understanding and generation tasks, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with $8.6\\%$ and $11.1\\%$ improvement respectively. HiTeA also demonstrates strong generalization ability when directly transferred to downstream tasks in a zero-shot manner.", + "bbox": [ + 75, + 273, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 712, + 209, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision and language are two primary signals that constitute the real-world perception of humanity. With the success of image-language pre-training [8,22,25,47], video-language pre-training [23,26,27,34] has recently received increasing attention. Large-scale video-language pre-training helps the model to learn effective multi-modal representation, which has shown significant improvement on a variety of video-language downstream tasks, such as videotext retrieval, video question answering and video caption-", + "bbox": [ + 75, + 738, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/483f83779b42714bc182be16f6d694585fb415a0441b17fe29454fad8bd01314.jpg", + "image_caption": [ + "Figure 1: Comparison between existing paradigms and ours for video-language pre-training. (a) Previous methods align video and text within global perspective as the pretext. (b) We introduce HiTeA by varying video in different temporal views and modeling cross-modal temporal information between moments and texts, as well as the temporal contextual relations between multi-modal pairs." + ], + "image_footnote": [], + "bbox": [ + 511, + 241, + 880, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing [5,33,44,48,50,53,59].", + "bbox": [ + 500, + 625, + 687, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the success of image-language pre-training paradigm, various methods [10, 11, 23, 26, 27] have been proposed to adapt it to video-language pre-training. ClipBERT [21] and Singularity [20] directly build on representations from image encoders and aggregate them via score aggregation function and temporal encoder. Furthermore, MILNCE [34] and Frozen [2] switch image encoder to video encoder for spatio-temporal video representation learning and align the video with corresponding text. In addition, some advanced pre-training tasks are designed through modeling entity [23], reconstructing masked patches [10] and predicting frame order [26, 61]. Despite their promising performance on downstream tasks, they treat video within global perspective illustrated in Figure 1(a), thus failing to consider fine-grained temporal information and temporal contextual relations which are essential to video-language pre-training.", + "bbox": [ + 496, + 642, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since untrimmed video contains various temporal details,", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 94, + 886, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15405", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "directly treating the video globally has two main limitations: (1) Less effective in modeling the fine-grained moment information including atomic actions and moments. As illustrated in Figure 1(b), we vary time resolutions and generate two views (long & short) for the input video. As a result, the short-view video clip tends to represent the moment information and the long-view video may express more event-level information. For example, the short-view video clip in Figure 1(b) only describes the moment of \"lick fingers\" rather than \"eating ice cream\". Such fine-grained moment information is hard to be captured by the long-view video under global event perspective; (2) Ignoring the temporal contextual relations implicitly existed in the video. Knowing the event expressed by the text, the moment \"eating ice cream\" can be inferred from the moment \"lick fingers\" shown by short-view video. However, such implicit temporal contextual relations between the moment and the event are rarely explored in previous works.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these problems, we propose a Hierarchical Temporal-Aware video-language pre-training framework, HiTeA, for both multi-modal understanding and generation. Except for the standard pre-training tasks, HiTeA introduces two novel temporal-aware video-language pre-training tasks, named cross-modal moment exploration (CME) and multimodal temporal relation exploration (MTRE), which not only model the fine-grained temporal moment information but also captures temporal contextual information hierarchically, yielding temporal-aware multi-modal representations for both understanding and generation. Specifically, we first generate the long-view and short-view videos with different time resolutions to build hierarchy of the input video. Then, based on the similarities of words and short-view video, we select the most relevant words as positive and leave the rest of the words as hard negatives. The CME pre-training task is applied to align the positive words and short-view video representations in the same embedding space. Moreover, to capture association between moments and the event for temporal contextual modeling, we match different views for the same video. However, directly matching two views visually would be noisy due to the background similarity [39]. To this end, we perform multi-modal pair alignment between videotext pairs via the MTRE pre-training task. More specifically, the short-view video guided by most relevant words and the long-view video guided by text will be aligned, which enables the model to extrapolate the contextual information from the short-view with language signal while enhancing temporal reasoning ability. Empowered by above two novel temporal-aware video-language pre-training tasks, HiTeA is capable of modeling temporal-aware multi-modal information revealed in video-text data including both fine-grained moment information and temporal contextual relations.", + "bbox": [ + 75, + 367, + 472, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In spite of a good performance, recent studies [4, 20] reveal most video-language downstream datasets are biased to", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "wards still objects, scenes, etc., while the temporal dynamics are negligible. To evaluate the temporal performance of the video-language pre-training model and temporal reliance of downstream datasets, we introduce temporal shuffling test for these datasets. This enables a comprehensive evaluation of temporal modeling capability in the video-language pre-training field. Besides, our method achieves significant improvement on the datasets with heavy temporal reliance.", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are the followings:", + "bbox": [ + 517, + 213, + 874, + 227 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel hierarchical temporal-aware video-language pre-training framework with both video-language understanding and generation capabilities.", + "- We introduce temporal-aware pre-training tasks to generate temporal-aware multi-modal representation through modeling fine-grained temporal moment information as well as capturing the temporal contextual relations between moment and event.", + "- Extensive experiments demonstrate the effectiveness and generalization ability of HiTeA, and it achieves state-of-the-art performance on 15 video-language downstream datasets including video-text retrieval, video question answering, and video captioning, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with $8.6\\%$ and $11.1\\%$ improvement respectively." + ], + "bbox": [ + 517, + 238, + 890, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 512, + 640, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Language Pre-training Benefiting from a large number of image/video-text pairs, video-language pretraining (VLP) exhibits superior capabilities on various video-text benchmarks. The method of VLP is constantly evolving. Traditional approaches [26, 30, 42, 62] leverage offline-extracted dense video features for pre-training to circumvent the expensive computation overhead. In contrast, ClipBERT [21] suggests that sparse sampling can enable affordable end-to-end learning and improve performance simultaneously. Recent emerging approaches [2, 11, 14, 21, 23, 27, 52, 57] adopt this strategy and propose new model architectures and pre-training tasks. Frozen [2] trains jointly on image and video datasets via video-text contrastive learning (VTC). ALPRO [23] proposes a new visually-grounded pre-training task combined with VTC, video-text matching (VTM) and masked language modeling (MLM) [9] to learn fine-grained region-entity alignment. LAVENDER [27] formulates all pre-training and downstream tasks as MLM so that a unified architecture can be used for all video-text tasks. Apart from above representative works, frame order modeling (FOM) [26, 61] and masked video modeling (MVM) [10] are designed for VLP. However, the temporal characteristic of video still remains largely unexplored. To this end, we introduce a novel hierar", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15406", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ea68e067c43e3539126ddd5b67b4506b70cb716b5392357a635800d66589dc0.jpg", + "image_caption": [ + "Figure 2: Illustration of the proposed HiTeA. We first generate two different temporal views for the input video, where the long-view is the video itself and the short-view is randomly truncated from the input video. To explore the moment revealed in the short-view, cross-modal moment exploration (CME) selects the candidate words from the input text with $\\mathcal{L}_{\\mathrm{CME}}$ . Then, we perform multi-modal temporal relation exploration (MTRE) for modeling the temporal contextual relations between two video-text pairs with different views by $\\mathcal{L}_{\\mathrm{MTRE}}$ . Note that the multi-modal encoders and the text features are shared." + ], + "image_footnote": [], + "bbox": [ + 147, + 95, + 823, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "chical temporal-aware VLP framework which not only models the fine-grained moment information but also captures contextual relations with different temporal granularities.", + "bbox": [ + 75, + 465, + 470, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Modeling The temporal characteristic acts as a vital role in VLP since it provides the model with the capabilities of reasoning and understanding causality. Previous efforts in this field can be roughly divided into three categories. First, several methods directly transfer image-text models to video-text tasks by simply concatenating video frame [22,24] or building an extra temporal encoder [31,32]. Second, some works [2, 10, 23, 27] switch the image encoder to video encoder for learning spatio-temporal contexts within videos. Third, HERO [26] and MERLOT [61] design FOM task to explicitly recover the correct temporal order of shuffled frames. Nonetheless, ATP [4] and Singularity [20] reveal the existence of a static appearance bias in popular video-language datasets, and they develop single-frame models to achieve strong performance, comparable or even better than above methods with explicit temporal modeling. Therefore, they recommend SSv2 [20] and NExT-QA [48] datasets to test the temporal ability of VLP models. Different from previous approaches, we vary the temporal resolutions and generate two views of video so as to construct the temporal hierarchy, which equips the model with the ability to learn both fine-grained temporal moment information and temporal contextual representation at the same time.", + "bbox": [ + 75, + 551, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 465, + 589, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 493, + 609, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 sketches the overview of the HiTeA. In concrete, our model consists of two unimodal encoders for encoding video and text separately, a multi-modal encoder for video and text interaction, and a text decoder for generation which is omitted here for simplicity and detailed in Appendix.", + "bbox": [ + 496, + 518, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For video representation, previous methods [21, 23, 27] encode the whole input video as a single-view feature, ignoring the rich temporal details contained in the video. Thus, we first treat the video into two views with different time resolutions to build hierarchy of the input video. Specially, the untrimmed video is regarded as a long-view $V^L$ for capturing event information, and a video segment is randomly truncated from the input video as the short-view for capturing moment information denoted as $V^S$ . Then, we use the video encoder to encode an arbitrary view of video $V \\in \\mathbb{R}^{T \\times H \\times W}$ into a sequence of embeddings: $\\mathcal{V} = \\{v_{\\mathrm{cls}}, v_1, \\dots, v_M\\} \\in \\mathbb{R}^{M \\times D}$ , where $M$ is the number of flattened patches, $D$ is hidden size, and $v_{\\mathrm{cls}}$ is the embedding of the visual [CLS] token which provides global representation of the video. For text representation, we use the text encoder to transform the text $T$ into a sequence of embeddings: $\\mathcal{T} = \\{w_{\\mathrm{cls}}, w_1, \\dots, w_N\\} \\in \\mathbb{R}^{N \\times D}$ , where $N$ is length of the text. After that, the multi-modal encoder takes video features $\\mathcal{V}$ and text features $\\mathcal{T}$ as inputs and yields the multi-modal representation $v_{\\mathrm{cls}}$ for the video.", + "bbox": [ + 496, + 598, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15407", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to take full advantage of the different views of the video, we introduce cross-modal moment exploration (CEM) to explore the proper words or phrases from input text to align the short-view video with $\\mathcal{L}_{\\mathrm{CME}}$ for capturing the fine-grained temporal moment information in Section 3.2. Furthermore, to model the temporal contextual relations between the short-view video containing moment information and the long-view video with event information, we propose multi-modal temporal relation exploration (MTRE) to match the multi-modal representation of shortview and long-view videos by $\\mathcal{L}_{\\mathrm{MTRE}}$ in Section 3.3. Lastly, we introduce the overall pre-training objective for training the model in Section 3.4.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cross-Modal Moment Exploration", + "text_level": 1, + "bbox": [ + 76, + 294, + 380, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The video with short temporal range (i.e., short-view of video) with the paired text tends to be accompanied with fine-grained temporal moment information. However, the paired text partially describes the short-view of video bringing noise to the fine-grained moment representation thus degrading the performance. To this end, we propose a novel pre-training task named cross-modal moment exploration (CME), which enables the model to understand fine-grained moment information by leveraging the partially aligned text.", + "bbox": [ + 75, + 316, + 468, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, we first discover the possible positive words for the video in short-view by computing the cosine similarity of the word embedding sequence $\\{w_{1},\\dots ,w_{N}\\}$ from text encoder and the short-view video representation $v_{\\mathrm{cls}}^S$ from video encoder as:", + "bbox": [ + 75, + 454, + 468, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} = \\{\\pi (1), \\dots , \\pi (K) \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 535, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi : \\{1, \\dots, N\\} \\to \\{1, \\dots, N\\}$ is a permutation function for ranking such that $s(w_{\\pi(1)}, v_{\\mathrm{cls}}^S) \\geq \\dots \\geq s(w_{\\pi(N)}, v_{\\mathrm{cls}}^S)$ , and $\\mathcal{K}$ is the set of selected word indices, $K$ is the number of possible selected words, and $s(x, y) = x^T y / \\|x\\|_2 \\|y\\|_2$ represents the cosine similarity between $x$ and $y$ . After obtaining the words for the video in short-view as the positive pair, the cross-modal moment exploration loss $\\mathcal{L}_{\\mathrm{CME}}$ is computed with negative pairs from other words in the input text, which is defined as:", + "bbox": [ + 75, + 558, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C M E}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\frac {1}{| \\mathcal {K} |} \\sum_ {k \\in \\mathcal {K}} \\log \\frac {\\exp \\left(\\left(v _ {\\mathrm {c l s}} ^ {S}\\right) _ {i} ^ {\\top} w _ {i , k} / \\tau\\right)}{\\sum_ {n = 1} ^ {N} \\exp \\left(\\left(v _ {\\mathrm {c l s}} ^ {S}\\right) _ {i} ^ {\\top} w _ {i , n} / \\tau\\right)}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 700, + 488, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is the learnable temperature hyper-parameter that controls the sharpness of the output distribution, and it is initialized as 0.07. As a consequence, the model is able to understand moment information via the proposed cross-modal exploration scheme.", + "bbox": [ + 75, + 763, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Modal Temporal Relation Exploration", + "text_level": 1, + "bbox": [ + 76, + 847, + 455, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the video encoder has demonstrated its effectiveness in learning temporal representation implicitly", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[10, 23, 56], it remains a challenge to discover the inherent temporal contextual relations. As a result, the limited capabilities in temporal modeling deteriorate the downstream task in temporal reasoning. This is in particular a missing point for the existing video-language pre-training paradigm [11, 21, 23, 27], which usually focuses on bridging video and text neglecting the function of text for guiding the video temporal contextual representation learning thus losing the temporal cues.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we introduce multi-modal temporal relation exploration (MTRE), a novel temporal-aware pre-training task that improves models' capacities in capturing temporal context in video with fine-grained text guidance by aligning multi-modal pairs. Specially, the short-view video $V^{S}$ would represent moment information with respect to the whole video. On the contrary, the long-view video $V^{L}$ expresses the event and topical information. To obtain the text-guided video features, we feed videos in different temporal views into the video encoder individually. Then, the text features are extracted and interact with the video features by the multi-modal encoder and yield text-guided video representations $\\boldsymbol{v}_{\\mathrm{cls}}^{L} \\in \\mathbb{R}^{D}$ and $\\boldsymbol{v}_{\\mathrm{cls}}^{S} \\in \\mathbb{R}^{D}$ as follows:", + "bbox": [ + 496, + 227, + 892, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {L} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {L}, v _ {1} ^ {L}, \\dots , v _ {M} ^ {L} \\right\\}, \\left\\{w _ {\\mathrm {c l s}}, w _ {1}, \\dots , w _ {N} \\right\\}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 434, + 890, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {S} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \\dots , v _ {M} ^ {S} \\right\\}, \\left\\{w _ {\\mathrm {c l s}}, w _ {1}, \\dots , w _ {N} \\right\\}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 455, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f(\\mathcal{V},\\mathcal{T})$ represents the multi-modal encoder with video features $\\mathcal{V}$ and text features $\\mathcal{T}$ . However, since the short-view of the video is partially aligned with the text, using the whole text is not reasonable for generating accurate text-guided video feature for short-view. Meanwhile, improper video-text pairs would yield noisy multi-modal representation thus degrading the performance of the model. Therefore, thanks to the positive words mined by cross-modal moment exploration, we can calibrate representation for short-view video by:", + "bbox": [ + 496, + 484, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {S} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \\dots , v _ {M} ^ {S} \\right\\}, \\left\\{w _ {\\mathcal {K} _ {1}}, \\dots , w _ {\\mathcal {K} _ {K}} \\right\\}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 645, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{K}_i\\in \\mathcal{K}$ is the index of the set for selected positive words. Then, we aim to match the representation of produced text-guided video features in different granularities in order to enable the model to predict the past and the future from the short-view of video, which benefits for capturing the general structure of the video. Specifically, we adopt the SimSiam framework [6] for minimizing their negative cosine similarity:", + "bbox": [ + 496, + 676, + 890, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} \\left(p ^ {S}, z ^ {L}\\right) = - \\frac {p ^ {S}}{\\| p ^ {S} \\| _ {2}} \\cdot \\frac {z ^ {L}}{\\| z ^ {L} \\| _ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 808, + 890, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p^S = h(g(\\mathbf{v}_{\\mathrm{cls}}^S))$ and $z^{L} = g(\\mathbf{v}_{\\mathrm{cls}}^{L})$ . The $g$ and $h$ are projection MLP head and prediction MLP head [7,13]. Minimizing $\\mathcal{D}(p^S,z^L)$ is equivalent for minimizing the mean", + "bbox": [ + 496, + 853, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15408", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/12e5a8d044373be85c09e8000995e378578d1532b64159a7de0b4efbd7faa80e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# PT DataMSRVTTDiDeMoLSMDCActivityNet Caption
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
ClipBERT [21]0.2M22.046.859.920.448.060.8---21.349.063.5
Frozen [2]5M31.059.570.531.059.872.415.030.839.8---
ALPRO [23]5M33.960.773.235.967.578.8------
BridgeFormer [11]5M37.664.875.137.062.273.917.935.444.5---
Singularity [20]5M36.865.975.547.475.284.0---43.070.681.3
LAVENDER [27]5M37.863.875.047.474.782.422.243.853.5---
", + "bbox": [ + 119, + 88, + 848, + 191 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4c12622ea141dc372243a42464bec9b5bebe055e8763264f4a98111087cd270e.jpg", + "table_caption": [ + "Models pre-trained on more data" + ], + "table_footnote": [], + "table_body": "
VIOLET [10]183M34.563.073.432.662.874.716.136.641.2---
All-in-one [46]138M37.968.177.132.761.473.5---22.453.767.7
Clip4Clip [31]400M42.171.981.443.470.280.621.641.849.840.572.4-
X-CLIP [32]400M46.173.083.145.274.0-23.343.0-44.374.1-
HiTeA5M44.469.378.951.879.185.327.146.254.545.173.584.2
", + "bbox": [ + 119, + 202, + 844, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Performance comparison on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more pre-training data for fair comparison. # PT Data is the number of video-text pairs for pre-training.", + "bbox": [ + 75, + 273, + 893, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "square error between $p^S$ and $z^L$ , which encourages the videos in different temporal magnitudes to be similar. Following [6, 13], we defined a symmetrized loss as:", + "bbox": [ + 75, + 335, + 470, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M T R E}} = \\frac {1}{2} \\left[ \\mathcal {D} \\left(p ^ {L}, \\mathbf {s g} \\left(z ^ {S}\\right)\\right) + \\mathcal {D} \\left(p ^ {S}, \\mathbf {s g} \\left(z ^ {L}\\right)\\right) \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 392, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathsf{sg}(\\cdot)$ is the stop-gradient operation that prevents the model from collapse during training [6].", + "bbox": [ + 75, + 431, + 468, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Pre-training Objectives", + "text_level": 1, + "bbox": [ + 76, + 470, + 294, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Apart from the two proposed temporal-aware pre-training tasks, we follow proven video-text pre-training approaches [2, 23, 27] to adopt the standard pre-training tasks including video-text contrastive (VTC), video-text matching (VTM), masked language modeling (MLM), and prefix language modeling (PrefixLM) described in the related work. Precisely, VTC and VTM align the video and text from the global perspective, while MLM and PrefixLM contribute to multi-modal understanding and generation capabilities of the model. Details of these objectives are described in the Appendix. We simply combine these as the base training objective $\\mathcal{L}_{base}$ for our model. Therefore, the full pre-training objective is computed as:", + "bbox": [ + 75, + 494, + 470, + 690 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {C M E}} + \\mathcal {L} _ {\\text {M T R E}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 703, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 756, + 253, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training Datasets Following the recent work [2,11,20, 23,27], we pre-train our model on a webly-sourced video dataset WebVid-2M [2] with 2.5M video-text pairs and a image-text dataset Google Conceptual Captions (CC3M) [41] with 3M image-text pairs. Unlike previous methods, we do not pre-train our model on the large-scale video-text datasets like HowTo100M [34] with 136M video-text pairs and YT-Temporal-180M [61] due to the heavy computation.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream Datasets We evaluate our pre-trained model on 18 video-language benchmarks including video-text retrieval, video question answering, and video captioning tasks. Specifically, video question answering (VideoQA) can be categorized as Multiple-Choice (MC) and Open-Ended (OE) settings. The evaluation datasets are briefly summarized in below. Details can be found in the Appendix.", + "bbox": [ + 496, + 335, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Video-Text Retrieval: MSRVTT [53], DiDeMo [1], LSMDC [38], ActivityNet Caption [18], SSv2-Label [20], and SSv2-Template [20];", + "- VideoQA (MC): TGIF-Action, TGIF-Transition [15], MSRVTT-MC [58], LSMDC-MC [44], and NExT-QA [48];", + "- VideoQA (OE): TGIF-Frame [15], MSRVTT-QA, MSVD-QA [50], LSMDC-FiB [33] and ActivityNet-QA [59].", + "- Video Captioning: MSRVTT [53] and MSVD [5]." + ], + "bbox": [ + 500, + 454, + 890, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details Our implementation of HiTeA is based on PyTorch [35]. In detail, we instantiate the video encoder with MViT-Base model [28] pretrained on ImageNet-21K [37]. The text encoder is initialized from first six layers of pre-trained BERT-Base [9], and the multi-modal encoder is initialized with last six layers of pre-trained BERT-Base. We pre-train HiTeA for 10 epochs, using a batch size of 16 on 8 NVIDIA A100 GPUs. We use AdamW [17] optimizer with a weight decay of 0.02 and betas (0.9, 0.98). The learning rate is first warmed up to 5e-5 in the first 1000 iterations, and decays following a cosine schedule. During pre-training stage, following with [10, 14, 23], we sparsely sample 4 frames for short and long view while preserving their order in-between and resize them to $224 \\times 224$ . The duration of short view is restricted as the 1/8 of the whole video duration. $K$ is empirically set to 5. The MLM mask ratio is set to $15\\%$ . Details of fine-tuning stage are described in Appendix.", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15409", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6e61ce32854071d354d2d0356ea2ef693ce0d0eded1ff27a1d8107e5b5a30732.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method#PT DataTGIFMSRVTTLSMDCMSVDActivityNet
ActionTransitionFrameMCQAMCFiBQAQA
ClipBERT [21]0.2M82.887.860.388.237.4----
ALPRO [23]5M----42.1--46.3-
Singularity [20]5M---92.042.7---41.8
LAVENDER [27]5M96.699.172.296.644.286.056.955.4-
Clover [14]5M94.998.071.495.043.983.254.151.9-
", + "bbox": [ + 181, + 88, + 785, + 190 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/34ca937dd55fbb6fa3095f9a11e5751859b3982d83e487498471b2b9f9868630.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Models pre-trained on more data
VIOLET [10]183M92.595.768.991.943.982.853.747.938.9
JustAsk [54]69M----41.5--46.338.9
MERLOT [61]180M94.096.269.590.943.181.752.9-41.4
All-in-one [46]283M95.594.766.392.346.884.4-48.3-
HiTeA5M96.898.872.597.245.485.854.655.645.1
", + "bbox": [ + 181, + 191, + 785, + 276 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7acf1abb5d26f8a77c1f1c5e7f7db875b78e3a167ce16c08752b901e55e13f43.jpg", + "table_caption": [ + "Table 2: Performance comparison on video question answering. Accuracy is reported for evaluation. We gray out methods that use significantly more pre-training data for fair comparison." + ], + "table_footnote": [], + "table_body": "
Method# PT DataMSRVTTMSVD
UniVL [30]180M49.9-
SwinBERT [29]-53.8120.6
MV-GPT [40]53M60.0-
CLIP4Caption [43]400M57.7-
LAVENDER [27]5M58.0142.9
HiTeA5M62.5145.1
", + "bbox": [ + 117, + 321, + 426, + 416 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/17d1b058fa28724f1fe5b7373d8f7830de9620e3331ca20e6d19766c55c097ed.jpg", + "table_caption": [ + "Table 3: Performance comparison on video captioning. CIDEr [45] is reported for evaluation." + ], + "table_footnote": [ + "Table 4: Performance comparison of different SoTA methods with TimeSformer [3]. For text-to-video retrieval, Recall@1/5/10 are reported. For video question answering task, we report the Top-1 accuracy." + ], + "table_body": "
MethodMSRVTT-Ret.LSMDC-Ret.MSRVTT-QAMSVD-QA
Frozen [2]31.0/59.5/70.515.0/30.8/39.8--
ALPRO [23]33.9/60.7/73.2-/-/-42.146.3
BridgeFormer [11]37.6/64.8/75.117.9/35.4/44.5--
Singularity [20]36.8/65.9/75.5-/-/-42.7-
TimeSformer (Lbase)37.6/62.0/72.319.2/38.7/48.041.549.0
HiTeA with TimeSformer39.7/65.0/75.121.8/40.7/49.943.752.4
", + "bbox": [ + 78, + 452, + 468, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to Prior Arts", + "text_level": 1, + "bbox": [ + 76, + 599, + 313, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare HiTeA with numerous state-of-the-art video-language pre-training methods on several downstream datasets under fine-tuning setting.", + "bbox": [ + 75, + 625, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Text-to-Video Retrieval", + "text_level": 1, + "bbox": [ + 76, + 693, + 292, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 summarizes the results on MSRVTT [53], DiDeMo [1], LSMDC [38], and ActivityNet Caption [18] under fine-tuning settings. Our method outperforms all of the existing video-language pre-training model by a large margin under the same data scale. In particular, our method yields $6.6\\%$ lift in terms of R@1 on MSRVTT dataset while only exploiting 5M video-text pairs. Note that we also include the comparison with the recent works that utilize the powerful encoder from CLIP [36], our method still can be comparable with them even surpass them, which shows the validity of the proposed method. Besides, we can notice that our method achieves the best result among all of listed methods", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on LSMDC dataset, which proves that our model can leverage the various moments presented in fruitful movie clips with cross-modal moment exploration.", + "bbox": [ + 496, + 324, + 890, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Video Question Answering", + "text_level": 1, + "bbox": [ + 498, + 388, + 741, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 lists the results of HiTeA and current state-of-the-art approaches on nine VideoQA datasets. It can be noticed that our method achieves the best performance in most of VideoQA datasets even with less pre-training data. Specifically, it achieves absolute improvement $1.1\\%$ on TGIF-FrameQA, $2.2\\%$ on MSRVTT-MC, $1.5\\%$ on MSRVTT-QA, $0.2\\%$ on MSVD-QA, and $3.3\\%$ on ActivityNet-QA. We believe the moments learned by the cross-modal exploration are useful for finding the clue of answers in VideoQA.", + "bbox": [ + 496, + 411, + 893, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 Video Captioning", + "text_level": 1, + "bbox": [ + 498, + 566, + 676, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 compares HiTeA with existing methods on video captioning datasets MSRVTT and MSVD. As shown in the table, although we use less pre-training data than compared approaches, HiTeA still obtains significant improvement compared to those large-scale pre-trained models. On MSRVTT Caption, our method surpasses SoTA method MV-GPT [40] by $2.5\\%$ CIDEr. Note that MV-GPT is pretrained for multi-modal video captioning and it leverages the ASR transcripts from audio as the additional input. By contrast, our method only utilizes video as the input during generation.", + "bbox": [ + 496, + 589, + 890, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Discussion", + "text_level": 1, + "bbox": [ + 500, + 765, + 617, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we discuss the temporal characteristics of our model and the datasets.", + "bbox": [ + 496, + 787, + 893, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization on Plain Backbone. Table 4 delivers the performance of SoTA methods with the plain visual backbone. We instantiate the video encoder with TimeSformer [3] pretrained on ImageNet-21K [37] following [2, 11, 23]. We can observe that HiTeA with TimeSformer significantly", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15410", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5d7f91cf91f805f4b9ce96a440beef3d9a0dba63d89d97eee40f614bf46aa24d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMSRVTT-RetrievalSSv2 Template-RetrievalNExT-QA (Hard)MSVD-QA
R@1R@5R@10AveRR@1R@5R@10AveRAcc@CAcc@TAcc.
\\( \\mathcal{L}_{\\text{base}} \\)40.068.077.161.780.5100.0100.093.544.046.452.7
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{MTRE}} \\)41.669.178.263.083.398.9100.094.146.346.454.8
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{CME}} \\)42.069.379.763.783.999.4100.094.446.348.354.3
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{CME}} + \\mathcal{L}_{\\text{MTRE}} \\)44.469.378.964.285.6100.0100.095.247.848.655.6
", + "bbox": [ + 140, + 88, + 826, + 178 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b29bd20e1e5da51083c8485cb7c48c4d62c7a95f6462e539af84ad39128c17c.jpg", + "table_caption": [ + "Table 5: Evaluation of the proposed methods on four downstream video-language tasks. For text-to-video retrieval, R@1, R@5, R@10, and the average are reported. For video question answering, we report the accuracy." + ], + "table_footnote": [], + "table_body": "
Method# PT DataSSv2-LabelSSv2-Template
R@1R@5R@10R@1R@5R@10
Frozen [2]5M---52.994.899.4
Clip4Clip [31]400M43.171.480.777.096.698.3
Singularity [20]5M44.173.582.277.098.999.4
HiTeA5M55.281.489.185.6100.0100.0
", + "bbox": [ + 78, + 223, + 467, + 294 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9704bd9383311c2196d32947917f65715b249adf72a6662f7bd02071a8024c00.jpg", + "table_caption": [ + "Table 6: Comparison of existing methods on Something-to-Something (SSv2) text-to-video retrieval." + ], + "table_footnote": [], + "table_body": "
Method# PT DataAcc@CAcc@TAcc@DAcc.
Full Set
Human-87.688.690.488.4
HCRN [19]-45.949.353.748.2
HGA [16]-46.350.759.349.7
VGT [49]0.18M53.456.469.556.9
HGA* [16]400M46.852.159.350.4
ATP [4]400M51.350.266.854.3
HiTeA5M62.458.375.663.1
Hard Split
ATP [4]400M38.436.5//
HGA [16]-43.345.3//
HiTeA5M47.848.6//
", + "bbox": [ + 98, + 330, + 444, + 496 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e3b3571c0e93c9d1496b5768a88bea75792b1e1b563c62df77b7db48f125c2b1.jpg", + "table_caption": [ + "Table 7: Comparison of existing methods on NExT-QA [48]. We report accuracy on the Causal (C), Temporal (T), Descriptive (D) splits and overall accuracy on validation set. * stands for using CLIP as the initialization of visual encoder." + ], + "table_footnote": [], + "table_body": "
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT [53]64.263.30.9
DiDeMo [1]72.170.21.9
LSMDC [38]42.641.70.9
ActivityNet Caption [18]67.666.80.8
SSv2 Template [20]95.272.422.8
SSv2 Label [20]76.773.53.2
", + "bbox": [ + 99, + 566, + 444, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8: Dependency on temporal information for text-to-video retrieval datasets with temporal shuffling test. The average recall of Recall@1, Recall@5, and Recall@10 are reported. We evaluate the performance drop when shuffling the input during inference. \"Original\" and \"Shuffled\" denote the original and shuffled input videos, respectively, and \"Gap\" is the difference between the Original and Shuffled metric. The larger \"Gap\" indicates the dataset relies on temporal information, and the model utilizes more temporal information to solve the task.", + "bbox": [ + 75, + 670, + 467, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "outperforms other SoTA methods (e.g., ALPRO, Bridge-Former, and Singularity) with the same pre-training data and the number of frames. In addition, it shows that our proposed temporal-aware pre-training tasks can boost the", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/790f52b371c913cf6f78a8b08a971ba13fd353b958b613999903aee8d2570597.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT-QA [50]45.445.20.2
MSVD-QA [50]55.655.50.1
TGIF-FrameQA [15]72.572.10.4
ActivityNet-QA [59]45.145.00.1
NExT-QA (Hard) [48]47.145.60.5
", + "bbox": [ + 540, + 223, + 848, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9: Dependency on temporal information for video question answering datasets by temporal shuffling test. We report the accuracy for each dataset. For NExT-QA dataset, we evaluate with the hard split of the validation set [4].", + "bbox": [ + 496, + 306, + 890, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance of TimeSformer $(\\mathcal{L}_{\\mathrm{base}})$ with $2.6\\%$ average improvement on MSRVTT-Retrieval and $3.4\\%$ on MSVD-QA. More related discussions are demonstrated in the Appendix.", + "bbox": [ + 496, + 380, + 890, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Loss Terms. We investigate the contribution of individual loss terms and the results are shown in Table 5. It can be observed that the combining both $\\mathcal{L}_{\\mathrm{CME}}$ and $\\mathcal{L}_{\\mathrm{MTRE}}$ improves the performance of text-to-video retrieval and video question answering by at least $1.7\\%$ and $2.9\\%$ in Average Recall and Average accuracy respectively. In addition, we also find that the performance of $\\mathcal{L}_{\\mathrm{CME}}$ surpasses that of $\\mathcal{L}_{\\mathrm{MTRE}}$ on MSRVTT retrieval dataset that largely dominated by the appearance information. This can be explained that the cross-modal moment exploration loss not only select the positive verbs for the video from the text but also choose the acting object for alignment, which can boost the retrieval performance.", + "bbox": [ + 496, + 436, + 890, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Temporal-aware Tasks. Lei et al. [20] reveal that the previous four retrieval datasets are prone to being biased for appearance while rarely relying on temporal information, thus introducing Something-to-Something v2 (SSv2) Template and SSv2 Label retrieval datasets to test models' true temporal modeling capability. In particular, SSv2 Template retrieval task requires a deeper understanding of the moment and temporal relation since no objects information are presented. The performance on these datasets are summarized in Table 6. It can be observed that HiTeA achieves significant improvement with $+8.5\\%$ gains in terms of R@1 on these two temporal-oriented text-to-video retrieval datasets, which demonstrates the effectiveness of our proposed method through exploring fine-grained moment information and modeling temporal relation. In addition, we evaluate our model on NExT-QA [48] dataset that explicitly designed for temporal and causal understanding.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/256a9cf39bd6f337d83a0cedc2418b23b2eb84a26c6ffd8b9e220895a453cbf0.jpg", + "image_caption": [ + "\"Cartoon characters are talking to a pokemon\"", + "Video", + "Frames", + "Baseline", + "Ours", + "Figure 3: Visualizations of learned cross-attention maps from multi-modal encoder. We present samples from MSRVTT [53] and SSv2 Template [20] retrieval dataset. HiTeA attends to the patches related to objects motion by tracking trajectory." + ], + "image_footnote": [], + "bbox": [ + 125, + 104, + 473, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/734fdad53243f68a2c60e2b68744a6772c584a614e769197cc9ff9281ca3d92c.jpg", + "image_caption": [ + "\"Poking [something] so that it spins around\"", + "Video", + "Frames" + ], + "image_footnote": [], + "bbox": [ + 537, + 104, + 885, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e71569ad943b5e14491b9f8b713a210dc182d8412095f3836cdc978df6f21d3.jpg", + "image_caption": [ + "Baseline" + ], + "image_footnote": [], + "bbox": [ + 537, + 142, + 885, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f09a43a20b52cab24ae1cf7b6496a600bb8f5239fbca6ce0d2f16de8ca7f793.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 180, + 885, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As presented in Table 7, our method significantly surpasses its competitive counterparts, even those methods equipped with powerful image-text pre-trained encoders. Quantitatively, HiTeA obtains an absolute improvement $+9\\%$ on the causality split with the help of intrinsic temporal relation. Recently, Buch et al. [4] filter out the trivial question for the dataset, and build the hard split for causality and temporal related questions for evaluate the causality and temporal of the model. As we can see in the table, even for the questions that heavily rely on causality, our model can still achieves a relative gain of $4.1\\%$ on the model with specific design for VideoQA, which indicates that our model do not solely depend on static appearance.", + "bbox": [ + 75, + 284, + 472, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal Reliance of Datasets. Previous methods [11,21, 23] only evaluate the performance of models on the existing datasets to demonstrate the superiority of the methods. However, Buch et al. [4] and Lei et al. [20] reveal that the most of the evaluation are biased towards the static concepts. Here, we investigate the temporal reliance for the evaluated datasets by introducing the temporal shuffling test. Specifically, we compute the performance changes between running inference on the ordered video versus its shuffled version. The large performance drop indicates the dataset has less spatial bias and needs for temporal information. Table 8 and Table 9 conclude the performance gap between ordered and shuffled input video for text-to-video retrieval and VideoQA datasets. For text-to-video retrieval task, SSv2 Template shows the large performance drop after shuffling the input video, which demonstrates that it depends most on the dynamic information thus verifying our assumption. On the contrary, the performance on ActivityNet Caption dataset is barely affected (-0.8 on Mean Recall) since the text almost describes the static objects without relying on temporal information. For video question answering dataset, we observe that the MSVD-QA and ActivityNet-QA are less sensitive to the order of video frames. This is because these two datasets contain more questions requiring frame-region information, such as object categories, scenes, and species. We believe this can be used to evaluate the temporal reliance of the datasets as well as the utilization for temporal cue by", + "bbox": [ + 75, + 493, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/85dc240c675e60c4cf346933c8b1027a96539570da38a5d0f3625563ab8a0119.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# PT DataMSRVTT-QAMSVD-QA
Just Ask [54]69M2.97.5
LAVENDER [27]5M4.511.6
MERLOT Reserve [60]1B5.8-
FrozenBiLM [55]10M6.411.7
HiTeA5M8.618.2
BLIP [24]129M19.235.2
mPLUG [22]400M21.137.2
HiTeA5M21.737.4
", + "bbox": [ + 540, + 281, + 849, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10: Zero-shot evaluation on video question answering. Accuracy is reported. We gray out those methods additionally supervised pre-training on VQA v2 [12] dataset.", + "bbox": [ + 496, + 386, + 890, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models in the future work.", + "bbox": [ + 500, + 439, + 676, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis. To verify that our model can capture the motion information with respected to the given text rather than inferring from the static signal, we present the query text in SSv2 Template dataset which has masked all of the object information, and also visualize the query in MSRVTT dataset. As we can see in the Figure 3, the attention map of atomic action \"talking\" mainly focuses on the mouse of the cartoon characters while the baseline largely focusing on the characters, which indicates that our method can understand the moment better when adopting the temporal-aware pretraining tasks. In another example, the word \"spins\" can reveal the trajectory of the object showing that our method is able to capture the temporal motion presented in the video.", + "bbox": [ + 496, + 460, + 890, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Zero-shot Generalizability", + "text_level": 1, + "bbox": [ + 500, + 665, + 740, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the generalizability of proposed videotext pre-trained model, we perform zero-shot evaluation on video-language downstream tasks. We evaluate the zero-shot performance on VideoQA task in Table 10. Our method attains competitive zero-shot performance on MSRVTT-QA and MSVD-QA datasets even without help of audio signal supervision [60] or additional generated video question pairs [54]. In particular, less pre-training data (i.e. $5\\mathrm{M} < 69\\mathrm{M}$ ) are used while our method can still outperform other SoTA approaches. We also evaluate the zero-shot performance of models supervised on VQA v2 [12]. We can find that our method surpasses the powerful multi-modal SoTA methods (e.g. mPLUG [22]) with only 5M pre-training data showing the better generalization ability of HiTeA.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15412", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/94bda957968573c897032601d98e1bbd5f1d9df4c59f0e315475196ece7c70fe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# PT DataMSRVTTDiDeMoLSMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
Frozen [2]5M18.739.551.621.146.056.29.322.030.1
ALPRO [23]5M24.144.755.423.847.357.9---
BridgeFormer [11]5M26.046.456.425.650.661.112.225.932.2
Singularity [20]5M28.450.259.536.961.669.3---
", + "bbox": [ + 160, + 88, + 803, + 174 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0ac70c2a391ca661f40b0914e3574d25df9e2f5711323514aea0bbb5ad1d58aa.jpg", + "table_caption": [ + "Models pre-trained on more data" + ], + "table_footnote": [], + "table_body": "
VideoCLIP [51]138M10.422.230.016.646.9----
VIOLET [10]183M25.949.559.723.549.859.8---
Clip4Clip [31]400M32.057.066.9---15.128.536.4
HiTeA5M29.954.262.936.160.170.315.531.139.8
", + "bbox": [ + 161, + 186, + 803, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 11: Zero-shot evaluation on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more data for fair comparison.", + "bbox": [ + 75, + 250, + 893, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also perform zero-shot evaluation on text-to-video retrieval task. Table 11 summarizes the performance of our model and compared approaches on text-to-video retrieval. We can observe that our model yields more than $3.4\\%$ lift in R@1 on MSRVTT dataset [53] while exploiting fewer video-text pairs. Besides, our method surpasses all of the compared models in terms of LSMDC dataset showing the superiority of our method's generalizability.", + "bbox": [ + 75, + 297, + 472, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 433, + 194, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we introduce HiTeA, a novel hierarchical temporal-aware video-language pre-training framework with both understanding and generation capabilities. We vary the video with different views and model cross-modal temporal information between moments and texts as well as temporal contextual relations between multi-modal pairs in a hierarchical way. Specifically, a cross-modal moment exploration pre-training task is proposed to explore the fine-grained temporal information between the paired text and video moment by overcoming the partially semantic alignment. Moreover, multi-modal pairs are constructed to learn temporal contextual relations between moments and the event presented by the video with multi-modal temporal relation exploration pre-training task. We also demonstrate that our proposed pre-training tasks can consistently boost performance significantly on downstream tasks regardless of the backbone showing the generalization ability. Even pre-trained on less data, HiTeA still achieves state-of-the-art performance on a wide range of video-language downstream datasets, which clearly shows the superiority of our method.", + "bbox": [ + 75, + 458, + 472, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 776, + 171, + 791 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6, 7", + "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-" + ], + "bbox": [ + 84, + 801, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 3, 5, 6, 7, 9", + "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 6", + "[4] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 2, 3, 7, 8", + "[5] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 1, 5", + "[6] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4, 5", + "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 4", + "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5", + "[10] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 4, 5, 6, 9", + "[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 4, 5, 6, 8, 9" + ], + "bbox": [ + 501, + 297, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15413", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8", + "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4, 5", + "[14] Jingjia Huang, Yinan Li, Jiashi Feng, Xiaoshuai Sun, and Rongrong Ji. Clover: Towards a unified video-language alignment and fusion model. arXiv preprint arXiv:2207.07885, 2022. 2, 5, 6", + "[15] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 5, 7", + "[16] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 7", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[18] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5, 6, 7", + "[19] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for multimodal video question answering. International Journal of Computer Vision, 129(11):3027-3050, 2021. 7", + "[20] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 2, 3, 5, 6, 7, 8, 9", + "[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 4, 5, 6, 8", + "[22] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 1, 3, 8", + "[23] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 3, 4, 5, 6, 8, 9", + "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni-" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fied vision-language understanding and generation. In ICML, 2022. 3, 8", + "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 1", + "[26] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1, 2, 3", + "[27] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. Lavender: Unifying videolanguage understanding as masked language modeling. arXiv preprint arXiv:2206.07160, 2022. 1, 2, 3, 4, 5, 6, 8", + "[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5", + "[29] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 6", + "[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2, 6", + "[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3, 5, 7, 9", + "[32] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022. 3, 5", + "[33] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5", + "[34] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 5", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15414", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6", + "[37] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5, 6", + "[38] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3202-3212, 2015. 5, 6, 7", + "[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022. 2", + "[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 6", + "[41] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 5", + "[42] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2", + "[43] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 6", + "[44] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5", + "[45] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6", + "[46] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022.5.6", + "[47] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 1", + "[48] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9777–9786, 2021. 1, 3, 5, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. arXiv preprint arXiv:2207.05342, 2022. 7", + "[50] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 5, 7", + "[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 9", + "[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. In Proceedings of International Conference of Machine Learning (ICML), 2023. 2", + "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 7, 8, 9", + "[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 6, 8", + "[55] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 8", + "[56] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 4", + "[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2", + "[58] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 5", + "[59] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9127–9134, 2019. 1, 5, 7", + "[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 1, 2, 3, 5, 6", + "[62] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 218 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15416", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_model.json b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3dae340817e743d18599cb153d7ec52023d49e2c --- /dev/null +++ b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_model.json @@ -0,0 +1,2435 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.09, + 0.828, + 0.113 + ], + "angle": 0, + "content": "HiTeA: Hierarchical Temporal-Aware Video-Language Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.121, + 0.704, + 0.158 + ], + "angle": 0, + "content": "Qinghao Ye Guohai Xu Ming Yan* Haiyang Xu* Qi Qian Ji Zhang Fei Huang" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.174, + 0.621, + 0.21 + ], + "angle": 0, + "content": "DAMO Academy, Alibaba Group yeqinghao.yqh@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.242, + 0.314, + 0.258 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Video-language pre-training has advanced the performance of various downstream video-language tasks. However, most previous methods directly inherit or adapt typical image-language pre-training paradigms to video-language pre-training, thus not fully exploiting the unique characteristic of video, i.e., temporal. In this paper, we propose a Hierarchical Temporal-Aware video-language pretraining framework, HiTeA, with two novel pre-training tasks for yielding temporal-aware multi-modal representation with cross-modal fine-grained temporal moment information and temporal contextual relations between video-text multi-modal pairs. First, we propose a cross-modal moment exploration task to explore moments in videos by mining the paired texts, which results in detailed video moment representation. Then, based on the learned detailed moment representations, the inherent temporal contextual relations are captured by aligning video-text pairs as a whole in different time resolutions with multi-modal temporal relation exploration task. Furthermore, we introduce the shuffling test to evaluate the temporal reliance of datasets and video-language pre-training models. We achieve state-of-the-art results on 15 well-established video-language understanding and generation tasks, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with \\(8.6\\%\\) and \\(11.1\\%\\) improvement respectively. HiTeA also demonstrates strong generalization ability when directly transferred to downstream tasks in a zero-shot manner." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.21, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Vision and language are two primary signals that constitute the real-world perception of humanity. With the success of image-language pre-training [8,22,25,47], video-language pre-training [23,26,27,34] has recently received increasing attention. Large-scale video-language pre-training helps the model to learn effective multi-modal representation, which has shown significant improvement on a variety of video-language downstream tasks, such as videotext retrieval, video question answering and video caption-" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.242, + 0.882, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.49, + 0.895, + 0.598 + ], + "angle": 0, + "content": "Figure 1: Comparison between existing paradigms and ours for video-language pre-training. (a) Previous methods align video and text within global perspective as the pretext. (b) We introduce HiTeA by varying video in different temporal views and modeling cross-modal temporal information between moments and texts, as well as the temporal contextual relations between multi-modal pairs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.626, + 0.688, + 0.641 + ], + "angle": 0, + "content": "ing [5,33,44,48,50,53,59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Inspired by the success of image-language pre-training paradigm, various methods [10, 11, 23, 26, 27] have been proposed to adapt it to video-language pre-training. ClipBERT [21] and Singularity [20] directly build on representations from image encoders and aggregate them via score aggregation function and temporal encoder. Furthermore, MILNCE [34] and Frozen [2] switch image encoder to video encoder for spatio-temporal video representation learning and align the video with corresponding text. In addition, some advanced pre-training tasks are designed through modeling entity [23], reconstructing masked patches [10] and predicting frame order [26, 61]. Despite their promising performance on downstream tasks, they treat video within global perspective illustrated in Figure 1(a), thus failing to consider fine-grained temporal information and temporal contextual relations which are essential to video-language pre-training." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Since untrimmed video contains various temporal details," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15405" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "directly treating the video globally has two main limitations: (1) Less effective in modeling the fine-grained moment information including atomic actions and moments. As illustrated in Figure 1(b), we vary time resolutions and generate two views (long & short) for the input video. As a result, the short-view video clip tends to represent the moment information and the long-view video may express more event-level information. For example, the short-view video clip in Figure 1(b) only describes the moment of \"lick fingers\" rather than \"eating ice cream\". Such fine-grained moment information is hard to be captured by the long-view video under global event perspective; (2) Ignoring the temporal contextual relations implicitly existed in the video. Knowing the event expressed by the text, the moment \"eating ice cream\" can be inferred from the moment \"lick fingers\" shown by short-view video. However, such implicit temporal contextual relations between the moment and the event are rarely explored in previous works." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.867 + ], + "angle": 0, + "content": "To address these problems, we propose a Hierarchical Temporal-Aware video-language pre-training framework, HiTeA, for both multi-modal understanding and generation. Except for the standard pre-training tasks, HiTeA introduces two novel temporal-aware video-language pre-training tasks, named cross-modal moment exploration (CME) and multimodal temporal relation exploration (MTRE), which not only model the fine-grained temporal moment information but also captures temporal contextual information hierarchically, yielding temporal-aware multi-modal representations for both understanding and generation. Specifically, we first generate the long-view and short-view videos with different time resolutions to build hierarchy of the input video. Then, based on the similarities of words and short-view video, we select the most relevant words as positive and leave the rest of the words as hard negatives. The CME pre-training task is applied to align the positive words and short-view video representations in the same embedding space. Moreover, to capture association between moments and the event for temporal contextual modeling, we match different views for the same video. However, directly matching two views visually would be noisy due to the background similarity [39]. To this end, we perform multi-modal pair alignment between videotext pairs via the MTRE pre-training task. More specifically, the short-view video guided by most relevant words and the long-view video guided by text will be aligned, which enables the model to extrapolate the contextual information from the short-view with language signal while enhancing temporal reasoning ability. Empowered by above two novel temporal-aware video-language pre-training tasks, HiTeA is capable of modeling temporal-aware multi-modal information revealed in video-text data including both fine-grained moment information and temporal contextual relations." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In spite of a good performance, recent studies [4, 20] reveal most video-language downstream datasets are biased to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "wards still objects, scenes, etc., while the temporal dynamics are negligible. To evaluate the temporal performance of the video-language pre-training model and temporal reliance of downstream datasets, we introduce temporal shuffling test for these datasets. This enables a comprehensive evaluation of temporal modeling capability in the video-language pre-training field. Besides, our method achieves significant improvement on the datasets with heavy temporal reliance." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.214, + 0.875, + 0.228 + ], + "angle": 0, + "content": "In summary, our key contributions are the followings:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.239, + 0.892, + 0.284 + ], + "angle": 0, + "content": "- We propose a novel hierarchical temporal-aware video-language pre-training framework with both video-language understanding and generation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.288, + 0.892, + 0.362 + ], + "angle": 0, + "content": "- We introduce temporal-aware pre-training tasks to generate temporal-aware multi-modal representation through modeling fine-grained temporal moment information as well as capturing the temporal contextual relations between moment and event." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.368, + 0.892, + 0.489 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness and generalization ability of HiTeA, and it achieves state-of-the-art performance on 15 video-language downstream datasets including video-text retrieval, video question answering, and video captioning, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with \\(8.6\\%\\) and \\(11.1\\%\\) improvement respectively." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.239, + 0.892, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.642, + 0.529 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Video-Language Pre-training Benefiting from a large number of image/video-text pairs, video-language pretraining (VLP) exhibits superior capabilities on various video-text benchmarks. The method of VLP is constantly evolving. Traditional approaches [26, 30, 42, 62] leverage offline-extracted dense video features for pre-training to circumvent the expensive computation overhead. In contrast, ClipBERT [21] suggests that sparse sampling can enable affordable end-to-end learning and improve performance simultaneously. Recent emerging approaches [2, 11, 14, 21, 23, 27, 52, 57] adopt this strategy and propose new model architectures and pre-training tasks. Frozen [2] trains jointly on image and video datasets via video-text contrastive learning (VTC). ALPRO [23] proposes a new visually-grounded pre-training task combined with VTC, video-text matching (VTM) and masked language modeling (MLM) [9] to learn fine-grained region-entity alignment. LAVENDER [27] formulates all pre-training and downstream tasks as MLM so that a unified architecture can be used for all video-text tasks. Apart from above representative works, frame order modeling (FOM) [26, 61] and masked video modeling (MVM) [10] are designed for VLP. However, the temporal characteristic of video still remains largely unexplored. To this end, we introduce a novel hierar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.096, + 0.825, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 2: Illustration of the proposed HiTeA. We first generate two different temporal views for the input video, where the long-view is the video itself and the short-view is randomly truncated from the input video. To explore the moment revealed in the short-view, cross-modal moment exploration (CME) selects the candidate words from the input text with \\(\\mathcal{L}_{\\mathrm{CME}}\\). Then, we perform multi-modal temporal relation exploration (MTRE) for modeling the temporal contextual relations between two video-text pairs with different views by \\(\\mathcal{L}_{\\mathrm{MTRE}}\\). Note that the multi-modal encoders and the text features are shared." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.471, + 0.514 + ], + "angle": 0, + "content": "chical temporal-aware VLP framework which not only models the fine-grained moment information but also captures contextual relations with different temporal granularities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Temporal Modeling The temporal characteristic acts as a vital role in VLP since it provides the model with the capabilities of reasoning and understanding causality. Previous efforts in this field can be roughly divided into three categories. First, several methods directly transfer image-text models to video-text tasks by simply concatenating video frame [22,24] or building an extra temporal encoder [31,32]. Second, some works [2, 10, 23, 27] switch the image encoder to video encoder for learning spatio-temporal contexts within videos. Third, HERO [26] and MERLOT [61] design FOM task to explicitly recover the correct temporal order of shuffled frames. Nonetheless, ATP [4] and Singularity [20] reveal the existence of a static appearance bias in popular video-language datasets, and they develop single-frame models to achieve strong performance, comparable or even better than above methods with explicit temporal modeling. Therefore, they recommend SSv2 [20] and NExT-QA [48] datasets to test the temporal ability of VLP models. Different from previous approaches, we vary the temporal resolutions and generate two views of video so as to construct the temporal hierarchy, which equips the model with the ability to learn both fine-grained temporal moment information and temporal contextual representation at the same time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.591, + 0.481 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.494, + 0.611, + 0.509 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Figure 2 sketches the overview of the HiTeA. In concrete, our model consists of two unimodal encoders for encoding video and text separately, a multi-modal encoder for video and text interaction, and a text decoder for generation which is omitted here for simplicity and detailed in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.903 + ], + "angle": 0, + "content": "For video representation, previous methods [21, 23, 27] encode the whole input video as a single-view feature, ignoring the rich temporal details contained in the video. Thus, we first treat the video into two views with different time resolutions to build hierarchy of the input video. Specially, the untrimmed video is regarded as a long-view \\(V^L\\) for capturing event information, and a video segment is randomly truncated from the input video as the short-view for capturing moment information denoted as \\(V^S\\). Then, we use the video encoder to encode an arbitrary view of video \\(V \\in \\mathbb{R}^{T \\times H \\times W}\\) into a sequence of embeddings: \\(\\mathcal{V} = \\{v_{\\mathrm{cls}}, v_1, \\dots, v_M\\} \\in \\mathbb{R}^{M \\times D}\\), where \\(M\\) is the number of flattened patches, \\(D\\) is hidden size, and \\(v_{\\mathrm{cls}}\\) is the embedding of the visual [CLS] token which provides global representation of the video. For text representation, we use the text encoder to transform the text \\(T\\) into a sequence of embeddings: \\(\\mathcal{T} = \\{w_{\\mathrm{cls}}, w_1, \\dots, w_N\\} \\in \\mathbb{R}^{N \\times D}\\), where \\(N\\) is length of the text. After that, the multi-modal encoder takes video features \\(\\mathcal{V}\\) and text features \\(\\mathcal{T}\\) as inputs and yields the multi-modal representation \\(v_{\\mathrm{cls}}\\) for the video." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15407" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "In order to take full advantage of the different views of the video, we introduce cross-modal moment exploration (CEM) to explore the proper words or phrases from input text to align the short-view video with \\(\\mathcal{L}_{\\mathrm{CME}}\\) for capturing the fine-grained temporal moment information in Section 3.2. Furthermore, to model the temporal contextual relations between the short-view video containing moment information and the long-view video with event information, we propose multi-modal temporal relation exploration (MTRE) to match the multi-modal representation of shortview and long-view videos by \\(\\mathcal{L}_{\\mathrm{MTRE}}\\) in Section 3.3. Lastly, we introduce the overall pre-training objective for training the model in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.295, + 0.381, + 0.312 + ], + "angle": 0, + "content": "3.2. Cross-Modal Moment Exploration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.454 + ], + "angle": 0, + "content": "The video with short temporal range (i.e., short-view of video) with the paired text tends to be accompanied with fine-grained temporal moment information. However, the paired text partially describes the short-view of video bringing noise to the fine-grained moment representation thus degrading the performance. To this end, we propose a novel pre-training task named cross-modal moment exploration (CME), which enables the model to understand fine-grained moment information by leveraging the partially aligned text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.529 + ], + "angle": 0, + "content": "Formally, we first discover the possible positive words for the video in short-view by computing the cosine similarity of the word embedding sequence \\(\\{w_{1},\\dots ,w_{N}\\}\\) from text encoder and the short-view video representation \\(v_{\\mathrm{cls}}^S\\) from video encoder as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.536, + 0.469, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} = \\{\\pi (1), \\dots , \\pi (K) \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.695 + ], + "angle": 0, + "content": "where \\(\\pi : \\{1, \\dots, N\\} \\to \\{1, \\dots, N\\}\\) is a permutation function for ranking such that \\(s(w_{\\pi(1)}, v_{\\mathrm{cls}}^S) \\geq \\dots \\geq s(w_{\\pi(N)}, v_{\\mathrm{cls}}^S)\\), and \\(\\mathcal{K}\\) is the set of selected word indices, \\(K\\) is the number of possible selected words, and \\(s(x, y) = x^T y / \\|x\\|_2 \\|y\\|_2\\) represents the cosine similarity between \\(x\\) and \\(y\\). After obtaining the words for the video in short-view as the positive pair, the cross-modal moment exploration loss \\(\\mathcal{L}_{\\mathrm{CME}}\\) is computed with negative pairs from other words in the input text, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.701, + 0.49, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C M E}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\frac {1}{| \\mathcal {K} |} \\sum_ {k \\in \\mathcal {K}} \\log \\frac {\\exp \\left(\\left(v _ {\\mathrm {c l s}} ^ {S}\\right) _ {i} ^ {\\top} w _ {i , k} / \\tau\\right)}{\\sum_ {n = 1} ^ {N} \\exp \\left(\\left(v _ {\\mathrm {c l s}} ^ {S}\\right) _ {i} ^ {\\top} w _ {i , n} / \\tau\\right)}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.84 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the learnable temperature hyper-parameter that controls the sharpness of the output distribution, and it is initialized as 0.07. As a consequence, the model is able to understand moment information via the proposed cross-modal exploration scheme." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.457, + 0.865 + ], + "angle": 0, + "content": "3.3. Multi-Modal Temporal Relation Exploration" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "While the video encoder has demonstrated its effectiveness in learning temporal representation implicitly" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "[10, 23, 56], it remains a challenge to discover the inherent temporal contextual relations. As a result, the limited capabilities in temporal modeling deteriorate the downstream task in temporal reasoning. This is in particular a missing point for the existing video-language pre-training paradigm [11, 21, 23, 27], which usually focuses on bridging video and text neglecting the function of text for guiding the video temporal contextual representation learning thus losing the temporal cues." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.426 + ], + "angle": 0, + "content": "To this end, we introduce multi-modal temporal relation exploration (MTRE), a novel temporal-aware pre-training task that improves models' capacities in capturing temporal context in video with fine-grained text guidance by aligning multi-modal pairs. Specially, the short-view video \\( V^{S} \\) would represent moment information with respect to the whole video. On the contrary, the long-view video \\( V^{L} \\) expresses the event and topical information. To obtain the text-guided video features, we feed videos in different temporal views into the video encoder individually. Then, the text features are extracted and interact with the video features by the multi-modal encoder and yield text-guided video representations \\( \\boldsymbol{v}_{\\mathrm{cls}}^{L} \\in \\mathbb{R}^{D} \\) and \\( \\boldsymbol{v}_{\\mathrm{cls}}^{S} \\in \\mathbb{R}^{D} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.435, + 0.891, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {L} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {L}, v _ {1} ^ {L}, \\dots , v _ {M} ^ {L} \\right\\}, \\left\\{w _ {\\mathrm {c l s}}, w _ {1}, \\dots , w _ {N} \\right\\}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.456, + 0.891, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {S} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \\dots , v _ {M} ^ {S} \\right\\}, \\left\\{w _ {\\mathrm {c l s}}, w _ {1}, \\dots , w _ {N} \\right\\}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.636 + ], + "angle": 0, + "content": "where \\( f(\\mathcal{V},\\mathcal{T}) \\) represents the multi-modal encoder with video features \\( \\mathcal{V} \\) and text features \\( \\mathcal{T} \\). However, since the short-view of the video is partially aligned with the text, using the whole text is not reasonable for generating accurate text-guided video feature for short-view. Meanwhile, improper video-text pairs would yield noisy multi-modal representation thus degrading the performance of the model. Therefore, thanks to the positive words mined by cross-modal moment exploration, we can calibrate representation for short-view video by:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.646, + 0.891, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {S} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \\dots , v _ {M} ^ {S} \\right\\}, \\left\\{w _ {\\mathcal {K} _ {1}}, \\dots , w _ {\\mathcal {K} _ {K}} \\right\\}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.892, + 0.799 + ], + "angle": 0, + "content": "where \\(\\mathcal{K}_i\\in \\mathcal{K}\\) is the index of the set for selected positive words. Then, we aim to match the representation of produced text-guided video features in different granularities in order to enable the model to predict the past and the future from the short-view of video, which benefits for capturing the general structure of the video. Specifically, we adopt the SimSiam framework [6] for minimizing their negative cosine similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.809, + 0.891, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} \\left(p ^ {S}, z ^ {L}\\right) = - \\frac {p ^ {S}}{\\| p ^ {S} \\| _ {2}} \\cdot \\frac {z ^ {L}}{\\| z ^ {L} \\| _ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.854, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(p^S = h(g(\\mathbf{v}_{\\mathrm{cls}}^S))\\) and \\(z^{L} = g(\\mathbf{v}_{\\mathrm{cls}}^{L})\\). The \\(g\\) and \\(h\\) are projection MLP head and prediction MLP head [7,13]. Minimizing \\(\\mathcal{D}(p^S,z^L)\\) is equivalent for minimizing the mean" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15408" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.849, + 0.193 + ], + "angle": 0, + "content": "
Method# PT DataMSRVTTDiDeMoLSMDCActivityNet Caption
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
ClipBERT [21]0.2M22.046.859.920.448.060.8---21.349.063.5
Frozen [2]5M31.059.570.531.059.872.415.030.839.8---
ALPRO [23]5M33.960.773.235.967.578.8------
BridgeFormer [11]5M37.664.875.137.062.273.917.935.444.5---
Singularity [20]5M36.865.975.547.475.284.0---43.070.681.3
LAVENDER [27]5M37.863.875.047.474.782.422.243.853.5---
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.193, + 0.292, + 0.203 + ], + "angle": 0, + "content": "Models pre-trained on more data" + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.203, + 0.846, + 0.269 + ], + "angle": 0, + "content": "
VIOLET [10]183M34.563.073.432.662.874.716.136.641.2---
All-in-one [46]138M37.968.177.132.761.473.5---22.453.767.7
Clip4Clip [31]400M42.171.981.443.470.280.621.641.849.840.572.4-
X-CLIP [32]400M46.173.083.145.274.0-23.343.0-44.374.1-
HiTeA5M44.469.378.951.879.185.327.146.254.545.173.584.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.275, + 0.895, + 0.321 + ], + "angle": 0, + "content": "Table 1: Performance comparison on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more pre-training data for fair comparison. # PT Data is the number of video-text pairs for pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.471, + 0.385 + ], + "angle": 0, + "content": "square error between \\( p^S \\) and \\( z^L \\), which encourages the videos in different temporal magnitudes to be similar. Following [6, 13], we defined a symmetrized loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.393, + 0.47, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M T R E}} = \\frac {1}{2} \\left[ \\mathcal {D} \\left(p ^ {L}, \\mathbf {s g} \\left(z ^ {S}\\right)\\right) + \\mathcal {D} \\left(p ^ {S}, \\mathbf {s g} \\left(z ^ {L}\\right)\\right) \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.464 + ], + "angle": 0, + "content": "where \\(\\mathsf{sg}(\\cdot)\\) is the stop-gradient operation that prevents the model from collapse during training [6]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.471, + 0.295, + 0.488 + ], + "angle": 0, + "content": "3.4. Pre-training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.471, + 0.691 + ], + "angle": 0, + "content": "Apart from the two proposed temporal-aware pre-training tasks, we follow proven video-text pre-training approaches [2, 23, 27] to adopt the standard pre-training tasks including video-text contrastive (VTC), video-text matching (VTM), masked language modeling (MLM), and prefix language modeling (PrefixLM) described in the related work. Precisely, VTC and VTM align the video and text from the global perspective, while MLM and PrefixLM contribute to multi-modal understanding and generation capabilities of the model. Details of these objectives are described in the Appendix. We simply combine these as the base training objective \\(\\mathcal{L}_{base}\\) for our model. Therefore, the full pre-training objective is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.704, + 0.47, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {C M E}} + \\mathcal {L} _ {\\text {M T R E}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.21, + 0.75 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.254, + 0.774 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Pre-training Datasets Following the recent work [2,11,20, 23,27], we pre-train our model on a webly-sourced video dataset WebVid-2M [2] with 2.5M video-text pairs and a image-text dataset Google Conceptual Captions (CC3M) [41] with 3M image-text pairs. Unlike previous methods, we do not pre-train our model on the large-scale video-text datasets like HowTo100M [34] with 136M video-text pairs and YT-Temporal-180M [61] due to the heavy computation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.445 + ], + "angle": 0, + "content": "Downstream Datasets We evaluate our pre-trained model on 18 video-language benchmarks including video-text retrieval, video question answering, and video captioning tasks. Specifically, video question answering (VideoQA) can be categorized as Multiple-Choice (MC) and Open-Ended (OE) settings. The evaluation datasets are briefly summarized in below. Details can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.455, + 0.892, + 0.5 + ], + "angle": 0, + "content": "- Video-Text Retrieval: MSRVTT [53], DiDeMo [1], LSMDC [38], ActivityNet Caption [18], SSv2-Label [20], and SSv2-Template [20];" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.504, + 0.892, + 0.549 + ], + "angle": 0, + "content": "- VideoQA (MC): TGIF-Action, TGIF-Transition [15], MSRVTT-MC [58], LSMDC-MC [44], and NExT-QA [48];" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.553, + 0.892, + 0.598 + ], + "angle": 0, + "content": "- VideoQA (OE): TGIF-Frame [15], MSRVTT-QA, MSVD-QA [50], LSMDC-FiB [33] and ActivityNet-QA [59]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.602, + 0.854, + 0.618 + ], + "angle": 0, + "content": "- Video Captioning: MSRVTT [53] and MSVD [5]." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.455, + 0.892, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Implementation Details Our implementation of HiTeA is based on PyTorch [35]. In detail, we instantiate the video encoder with MViT-Base model [28] pretrained on ImageNet-21K [37]. The text encoder is initialized from first six layers of pre-trained BERT-Base [9], and the multi-modal encoder is initialized with last six layers of pre-trained BERT-Base. We pre-train HiTeA for 10 epochs, using a batch size of 16 on 8 NVIDIA A100 GPUs. We use AdamW [17] optimizer with a weight decay of 0.02 and betas (0.9, 0.98). The learning rate is first warmed up to 5e-5 in the first 1000 iterations, and decays following a cosine schedule. During pre-training stage, following with [10, 14, 23], we sparsely sample 4 frames for short and long view while preserving their order in-between and resize them to \\(224 \\times 224\\). The duration of short view is restricted as the 1/8 of the whole video duration. \\(K\\) is empirically set to 5. The MLM mask ratio is set to \\(15\\%\\). Details of fine-tuning stage are described in Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15409" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.089, + 0.787, + 0.191 + ], + "angle": 0, + "content": "
Method#PT DataTGIFMSRVTTLSMDCMSVDActivityNet
ActionTransitionFrameMCQAMCFiBQAQA
ClipBERT [21]0.2M82.887.860.388.237.4----
ALPRO [23]5M----42.1--46.3-
Singularity [20]5M---92.042.7---41.8
LAVENDER [27]5M96.699.172.296.644.286.056.955.4-
Clover [14]5M94.998.071.495.043.983.254.151.9-
" + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.192, + 0.787, + 0.277 + ], + "angle": 0, + "content": "
Models pre-trained on more data
VIOLET [10]183M92.595.768.991.943.982.853.747.938.9
JustAsk [54]69M----41.5--46.338.9
MERLOT [61]180M94.096.269.590.943.181.752.9-41.4
All-in-one [46]283M95.594.766.392.346.884.4-48.3-
HiTeA5M96.898.872.597.245.485.854.655.645.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.279, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Table 2: Performance comparison on video question answering. Accuracy is reported for evaluation. We gray out methods that use significantly more pre-training data for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.322, + 0.428, + 0.417 + ], + "angle": 0, + "content": "
Method# PT DataMSRVTTMSVD
UniVL [30]180M49.9-
SwinBERT [29]-53.8120.6
MV-GPT [40]53M60.0-
CLIP4Caption [43]400M57.7-
LAVENDER [27]5M58.0142.9
HiTeA5M62.5145.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.422, + 0.47, + 0.451 + ], + "angle": 0, + "content": "Table 3: Performance comparison on video captioning. CIDEr [45] is reported for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.453, + 0.469, + 0.532 + ], + "angle": 0, + "content": "
MethodMSRVTT-Ret.LSMDC-Ret.MSRVTT-QAMSVD-QA
Frozen [2]31.0/59.5/70.515.0/30.8/39.8--
ALPRO [23]33.9/60.7/73.2-/-/-42.146.3
BridgeFormer [11]37.6/64.8/75.117.9/35.4/44.5--
Singularity [20]36.8/65.9/75.5-/-/-42.7-
TimeSformer (Lbase)37.6/62.0/72.319.2/38.7/48.041.549.0
HiTeA with TimeSformer39.7/65.0/75.121.8/40.7/49.943.752.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Table 4: Performance comparison of different SoTA methods with TimeSformer [3]. For text-to-video retrieval, Recall@1/5/10 are reported. For video question answering task, we report the Top-1 accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.314, + 0.618 + ], + "angle": 0, + "content": "4.2. Comparison to Prior Arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.673 + ], + "angle": 0, + "content": "In this section, we compare HiTeA with numerous state-of-the-art video-language pre-training methods on several downstream datasets under fine-tuning setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.294, + 0.71 + ], + "angle": 0, + "content": "4.2.1 Text-to-Video Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Table 1 summarizes the results on MSRVTT [53], DiDeMo [1], LSMDC [38], and ActivityNet Caption [18] under fine-tuning settings. Our method outperforms all of the existing video-language pre-training model by a large margin under the same data scale. In particular, our method yields \\(6.6\\%\\) lift in terms of R@1 on MSRVTT dataset while only exploiting 5M video-text pairs. Note that we also include the comparison with the recent works that utilize the powerful encoder from CLIP [36], our method still can be comparable with them even surpass them, which shows the validity of the proposed method. Besides, we can notice that our method achieves the best result among all of listed methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.372 + ], + "angle": 0, + "content": "on LSMDC dataset, which proves that our model can leverage the various moments presented in fruitful movie clips with cross-modal moment exploration." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.389, + 0.742, + 0.405 + ], + "angle": 0, + "content": "4.2.2 Video Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.549 + ], + "angle": 0, + "content": "Table 2 lists the results of HiTeA and current state-of-the-art approaches on nine VideoQA datasets. It can be noticed that our method achieves the best performance in most of VideoQA datasets even with less pre-training data. Specifically, it achieves absolute improvement \\(1.1\\%\\) on TGIF-FrameQA, \\(2.2\\%\\) on MSRVTT-MC, \\(1.5\\%\\) on MSRVTT-QA, \\(0.2\\%\\) on MSVD-QA, and \\(3.3\\%\\) on ActivityNet-QA. We believe the moments learned by the cross-modal exploration are useful for finding the clue of answers in VideoQA." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.567, + 0.677, + 0.583 + ], + "angle": 0, + "content": "4.2.3 Video Captioning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Table 3 compares HiTeA with existing methods on video captioning datasets MSRVTT and MSVD. As shown in the table, although we use less pre-training data than compared approaches, HiTeA still obtains significant improvement compared to those large-scale pre-trained models. On MSRVTT Caption, our method surpasses SoTA method MV-GPT [40] by \\(2.5\\%\\) CIDEr. Note that MV-GPT is pretrained for multi-modal video captioning and it leverages the ASR transcripts from audio as the additional input. By contrast, our method only utilizes video as the input during generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.766, + 0.618, + 0.781 + ], + "angle": 0, + "content": "4.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.894, + 0.819 + ], + "angle": 0, + "content": "In this section, we discuss the temporal characteristics of our model and the datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Generalization on Plain Backbone. Table 4 delivers the performance of SoTA methods with the plain visual backbone. We instantiate the video encoder with TimeSformer [3] pretrained on ImageNet-21K [37] following [2, 11, 23]. We can observe that HiTeA with TimeSformer significantly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15410" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.141, + 0.089, + 0.828, + 0.179 + ], + "angle": 0, + "content": "
MethodMSRVTT-RetrievalSSv2 Template-RetrievalNExT-QA (Hard)MSVD-QA
R@1R@5R@10AveRR@1R@5R@10AveRAcc@CAcc@TAcc.
\\( \\mathcal{L}_{\\text{base}} \\)40.068.077.161.780.5100.0100.093.544.046.452.7
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{MTRE}} \\)41.669.178.263.083.398.9100.094.146.346.454.8
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{CME}} \\)42.069.379.763.783.999.4100.094.446.348.354.3
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{CME}} + \\mathcal{L}_{\\text{MTRE}} \\)44.469.378.964.285.6100.0100.095.247.848.655.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.18, + 0.892, + 0.211 + ], + "angle": 0, + "content": "Table 5: Evaluation of the proposed methods on four downstream video-language tasks. For text-to-video retrieval, R@1, R@5, R@10, and the average are reported. For video question answering, we report the accuracy." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.224, + 0.468, + 0.295 + ], + "angle": 0, + "content": "
Method# PT DataSSv2-LabelSSv2-Template
R@1R@5R@10R@1R@5R@10
Frozen [2]5M---52.994.899.4
Clip4Clip [31]400M43.171.480.777.096.698.3
Singularity [20]5M44.173.582.277.098.999.4
HiTeA5M55.281.489.185.6100.0100.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.296, + 0.468, + 0.327 + ], + "angle": 0, + "content": "Table 6: Comparison of existing methods on Something-to-Something (SSv2) text-to-video retrieval." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.332, + 0.446, + 0.497 + ], + "angle": 0, + "content": "
Method# PT DataAcc@CAcc@TAcc@DAcc.
Full Set
Human-87.688.690.488.4
HCRN [19]-45.949.353.748.2
HGA [16]-46.350.759.349.7
VGT [49]0.18M53.456.469.556.9
HGA* [16]400M46.852.159.350.4
ATP [4]400M51.350.266.854.3
HiTeA5M62.458.375.663.1
Hard Split
ATP [4]400M38.436.5//
HGA [16]-43.345.3//
HiTeA5M47.848.6//
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.502, + 0.468, + 0.563 + ], + "angle": 0, + "content": "Table 7: Comparison of existing methods on NExT-QA [48]. We report accuracy on the Causal (C), Temporal (T), Descriptive (D) splits and overall accuracy on validation set. * stands for using CLIP as the initialization of visual encoder." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.568, + 0.445, + 0.666 + ], + "angle": 0, + "content": "
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT [53]64.263.30.9
DiDeMo [1]72.170.21.9
LSMDC [38]42.641.70.9
ActivityNet Caption [18]67.666.80.8
SSv2 Template [20]95.272.422.8
SSv2 Label [20]76.773.53.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.468, + 0.822 + ], + "angle": 0, + "content": "Table 8: Dependency on temporal information for text-to-video retrieval datasets with temporal shuffling test. The average recall of Recall@1, Recall@5, and Recall@10 are reported. We evaluate the performance drop when shuffling the input during inference. \"Original\" and \"Shuffled\" denote the original and shuffled input videos, respectively, and \"Gap\" is the difference between the Original and Shuffled metric. The larger \"Gap\" indicates the dataset relies on temporal information, and the model utilizes more temporal information to solve the task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "outperforms other SoTA methods (e.g., ALPRO, Bridge-Former, and Singularity) with the same pre-training data and the number of frames. In addition, it shows that our proposed temporal-aware pre-training tasks can boost the" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.224, + 0.849, + 0.303 + ], + "angle": 0, + "content": "
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT-QA [50]45.445.20.2
MSVD-QA [50]55.655.50.1
TGIF-FrameQA [15]72.572.10.4
ActivityNet-QA [59]45.145.00.1
NExT-QA (Hard) [48]47.145.60.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Table 9: Dependency on temporal information for video question answering datasets by temporal shuffling test. We report the accuracy for each dataset. For NExT-QA dataset, we evaluate with the hard split of the validation set [4]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.428 + ], + "angle": 0, + "content": "performance of TimeSformer \\((\\mathcal{L}_{\\mathrm{base}})\\) with \\(2.6\\%\\) average improvement on MSRVTT-Retrieval and \\(3.4\\%\\) on MSVD-QA. More related discussions are demonstrated in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Impact of Loss Terms. We investigate the contribution of individual loss terms and the results are shown in Table 5. It can be observed that the combining both \\(\\mathcal{L}_{\\mathrm{CME}}\\) and \\(\\mathcal{L}_{\\mathrm{MTRE}}\\) improves the performance of text-to-video retrieval and video question answering by at least \\(1.7\\%\\) and \\(2.9\\%\\) in Average Recall and Average accuracy respectively. In addition, we also find that the performance of \\(\\mathcal{L}_{\\mathrm{CME}}\\) surpasses that of \\(\\mathcal{L}_{\\mathrm{MTRE}}\\) on MSRVTT retrieval dataset that largely dominated by the appearance information. This can be explained that the cross-modal moment exploration loss not only select the positive verbs for the video from the text but also choose the acting object for alignment, which can boost the retrieval performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on Temporal-aware Tasks. Lei et al. [20] reveal that the previous four retrieval datasets are prone to being biased for appearance while rarely relying on temporal information, thus introducing Something-to-Something v2 (SSv2) Template and SSv2 Label retrieval datasets to test models' true temporal modeling capability. In particular, SSv2 Template retrieval task requires a deeper understanding of the moment and temporal relation since no objects information are presented. The performance on these datasets are summarized in Table 6. It can be observed that HiTeA achieves significant improvement with \\(+8.5\\%\\) gains in terms of R@1 on these two temporal-oriented text-to-video retrieval datasets, which demonstrates the effectiveness of our proposed method through exploring fine-grained moment information and modeling temporal relation. In addition, we evaluate our model on NExT-QA [48] dataset that explicitly designed for temporal and causal understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15411" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.092, + 0.406, + 0.103 + ], + "angle": 0, + "content": "\"Cartoon characters are talking to a pokemon\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.113, + 0.117, + 0.122 + ], + "angle": 0, + "content": "Video" + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.123, + 0.116, + 0.131 + ], + "angle": 0, + "content": "Frames" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.158, + 0.119, + 0.166 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.195, + 0.112, + 0.203 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.106, + 0.474, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.092, + 0.824, + 0.103 + ], + "angle": 0, + "content": "\"Poking [something] so that it spins around\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.115, + 0.526, + 0.124 + ], + "angle": 0, + "content": "Video" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.124, + 0.526, + 0.132 + ], + "angle": 0, + "content": "Frames" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.106, + 0.886, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.158, + 0.528, + 0.165 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.143, + 0.886, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.181, + 0.886, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.892, + 0.259 + ], + "angle": 0, + "content": "Figure 3: Visualizations of learned cross-attention maps from multi-modal encoder. We present samples from MSRVTT [53] and SSv2 Template [20] retrieval dataset. HiTeA attends to the patches related to objects motion by tracking trajectory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.473, + 0.482 + ], + "angle": 0, + "content": "As presented in Table 7, our method significantly surpasses its competitive counterparts, even those methods equipped with powerful image-text pre-trained encoders. Quantitatively, HiTeA obtains an absolute improvement \\(+9\\%\\) on the causality split with the help of intrinsic temporal relation. Recently, Buch et al. [4] filter out the trivial question for the dataset, and build the hard split for causality and temporal related questions for evaluate the causality and temporal of the model. As we can see in the table, even for the questions that heavily rely on causality, our model can still achieves a relative gain of \\(4.1\\%\\) on the model with specific design for VideoQA, which indicates that our model do not solely depend on static appearance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Temporal Reliance of Datasets. Previous methods [11,21, 23] only evaluate the performance of models on the existing datasets to demonstrate the superiority of the methods. However, Buch et al. [4] and Lei et al. [20] reveal that the most of the evaluation are biased towards the static concepts. Here, we investigate the temporal reliance for the evaluated datasets by introducing the temporal shuffling test. Specifically, we compute the performance changes between running inference on the ordered video versus its shuffled version. The large performance drop indicates the dataset has less spatial bias and needs for temporal information. Table 8 and Table 9 conclude the performance gap between ordered and shuffled input video for text-to-video retrieval and VideoQA datasets. For text-to-video retrieval task, SSv2 Template shows the large performance drop after shuffling the input video, which demonstrates that it depends most on the dynamic information thus verifying our assumption. On the contrary, the performance on ActivityNet Caption dataset is barely affected (-0.8 on Mean Recall) since the text almost describes the static objects without relying on temporal information. For video question answering dataset, we observe that the MSVD-QA and ActivityNet-QA are less sensitive to the order of video frames. This is because these two datasets contain more questions requiring frame-region information, such as object categories, scenes, and species. We believe this can be used to evaluate the temporal reliance of the datasets as well as the utilization for temporal cue by" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.282, + 0.851, + 0.382 + ], + "angle": 0, + "content": "
Method# PT DataMSRVTT-QAMSVD-QA
Just Ask [54]69M2.97.5
LAVENDER [27]5M4.511.6
MERLOT Reserve [60]1B5.8-
FrozenBiLM [55]10M6.411.7
HiTeA5M8.618.2
BLIP [24]129M19.235.2
mPLUG [22]400M21.137.2
HiTeA5M21.737.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.433 + ], + "angle": 0, + "content": "Table 10: Zero-shot evaluation on video question answering. Accuracy is reported. We gray out those methods additionally supervised pre-training on VQA v2 [12] dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.44, + 0.678, + 0.454 + ], + "angle": 0, + "content": "models in the future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Qualitative Analysis. To verify that our model can capture the motion information with respected to the given text rather than inferring from the static signal, we present the query text in SSv2 Template dataset which has masked all of the object information, and also visualize the query in MSRVTT dataset. As we can see in the Figure 3, the attention map of atomic action \"talking\" mainly focuses on the mouse of the cartoon characters while the baseline largely focusing on the characters, which indicates that our method can understand the moment better when adopting the temporal-aware pretraining tasks. In another example, the word \"spins\" can reveal the trajectory of the object showing that our method is able to capture the temporal motion presented in the video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.741, + 0.683 + ], + "angle": 0, + "content": "4.4. Zero-shot Generalizability" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the generalizability of proposed videotext pre-trained model, we perform zero-shot evaluation on video-language downstream tasks. We evaluate the zero-shot performance on VideoQA task in Table 10. Our method attains competitive zero-shot performance on MSRVTT-QA and MSVD-QA datasets even without help of audio signal supervision [60] or additional generated video question pairs [54]. In particular, less pre-training data (i.e. \\(5\\mathrm{M} < 69\\mathrm{M}\\)) are used while our method can still outperform other SoTA approaches. We also evaluate the zero-shot performance of models supervised on VQA v2 [12]. We can find that our method surpasses the powerful multi-modal SoTA methods (e.g. mPLUG [22]) with only 5M pre-training data showing the better generalization ability of HiTeA." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15412" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.161, + 0.089, + 0.805, + 0.175 + ], + "angle": 0, + "content": "
Method# PT DataMSRVTTDiDeMoLSMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
Frozen [2]5M18.739.551.621.146.056.29.322.030.1
ALPRO [23]5M24.144.755.423.847.357.9---
BridgeFormer [11]5M26.046.456.425.650.661.112.225.932.2
Singularity [20]5M28.450.259.536.961.669.3---
" + }, + { + "type": "table_caption", + "bbox": [ + 0.169, + 0.176, + 0.348, + 0.186 + ], + "angle": 0, + "content": "Models pre-trained on more data" + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.187, + 0.805, + 0.245 + ], + "angle": 0, + "content": "
VideoCLIP [51]138M10.422.230.016.646.9----
VIOLET [10]183M25.949.559.723.549.859.8---
Clip4Clip [31]400M32.057.066.9---15.128.536.4
HiTeA5M29.954.262.936.160.170.315.531.139.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.251, + 0.895, + 0.283 + ], + "angle": 0, + "content": "Table 11: Zero-shot evaluation on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more data for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.473, + 0.421 + ], + "angle": 0, + "content": "We also perform zero-shot evaluation on text-to-video retrieval task. Table 11 summarizes the performance of our model and compared approaches on text-to-video retrieval. We can observe that our model yields more than \\(3.4\\%\\) lift in R@1 on MSRVTT dataset [53] while exploiting fewer video-text pairs. Besides, our method surpasses all of the compared models in terms of LSMDC dataset showing the superiority of our method's generalizability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.196, + 0.45 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.473, + 0.763 + ], + "angle": 0, + "content": "In this work, we introduce HiTeA, a novel hierarchical temporal-aware video-language pre-training framework with both understanding and generation capabilities. We vary the video with different views and model cross-modal temporal information between moments and texts as well as temporal contextual relations between multi-modal pairs in a hierarchical way. Specifically, a cross-modal moment exploration pre-training task is proposed to explore the fine-grained temporal information between the paired text and video moment by overcoming the partially semantic alignment. Moreover, multi-modal pairs are constructed to learn temporal contextual relations between moments and the event presented by the video with multi-modal temporal relation exploration pre-training task. We also demonstrate that our proposed pre-training tasks can consistently boost performance significantly on downstream tasks regardless of the backbone showing the generalization ability. Even pre-trained on less data, HiTeA still achieves state-of-the-art performance on a wide range of video-language downstream datasets, which clearly shows the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.777, + 0.173, + 0.792 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.802, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.299, + 0.892, + 0.34 + ], + "angle": 0, + "content": "end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 3, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.342, + 0.894, + 0.383 + ], + "angle": 0, + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.384, + 0.895, + 0.453 + ], + "angle": 0, + "content": "[4] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.454, + 0.895, + 0.522 + ], + "angle": 0, + "content": "[5] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.524, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[6] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.58, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.636, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.705, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[10] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 4, 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 4, 5, 6, 8, 9" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.299, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15413" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[14] Jingjia Huang, Yinan Li, Jiashi Feng, Xiaoshuai Sun, and Rongrong Ji. Clover: Towards a unified video-language alignment and fusion model. arXiv preprint arXiv:2207.07885, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.374 + ], + "angle": 0, + "content": "[15] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[16] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[18] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[19] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for multimodal video question answering. International Journal of Computer Vision, 129(11):3027-3050, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.631 + ], + "angle": 0, + "content": "[20] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 2, 3, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.471, + 0.715 + ], + "angle": 0, + "content": "[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[22] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[23] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 3, 4, 5, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "fied vision-language understanding and generation. In ICML, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[26] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[27] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. Lavender: Unifying videolanguage understanding as masked language modeling. arXiv preprint arXiv:2206.07160, 2022. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[29] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3, 5, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[32] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[33] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[34] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15414" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[37] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[38] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3202-3212, 2015. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[41] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[42] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[43] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[44] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[45] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[46] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022.5.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[47] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[48] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9777–9786, 2021. 1, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. arXiv preprint arXiv:2207.05342, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[50] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. In Proceedings of International Conference of Machine Learning (ICML), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[55] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[56] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[58] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[59] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9127–9134, 2019. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15415" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[62] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15416" + } + ] +] \ No newline at end of file diff --git a/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_origin.pdf b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc9a96617cafd5a0f17afea6aa95a0fc317f513e --- /dev/null +++ b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/122f003d-e493-4776-85d5-ea92cef4fff9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a28dbb685d05aa2249af19a2d30222e6fb38690014a785ab546eb54e0250070 +size 1573091 diff --git a/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/full.md b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..56adb9b89b8c8baa050113c123ab6663d00a6859 --- /dev/null +++ b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/full.md @@ -0,0 +1,331 @@ +# HiTeA: Hierarchical Temporal-Aware Video-Language Pre-training + +Qinghao Ye Guohai Xu Ming Yan* Haiyang Xu* Qi Qian Ji Zhang Fei Huang + +DAMO Academy, Alibaba Group yeqinghao.yqh@alibaba-inc.com + +# Abstract + +Video-language pre-training has advanced the performance of various downstream video-language tasks. However, most previous methods directly inherit or adapt typical image-language pre-training paradigms to video-language pre-training, thus not fully exploiting the unique characteristic of video, i.e., temporal. In this paper, we propose a Hierarchical Temporal-Aware video-language pretraining framework, HiTeA, with two novel pre-training tasks for yielding temporal-aware multi-modal representation with cross-modal fine-grained temporal moment information and temporal contextual relations between video-text multi-modal pairs. First, we propose a cross-modal moment exploration task to explore moments in videos by mining the paired texts, which results in detailed video moment representation. Then, based on the learned detailed moment representations, the inherent temporal contextual relations are captured by aligning video-text pairs as a whole in different time resolutions with multi-modal temporal relation exploration task. Furthermore, we introduce the shuffling test to evaluate the temporal reliance of datasets and video-language pre-training models. We achieve state-of-the-art results on 15 well-established video-language understanding and generation tasks, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with $8.6\%$ and $11.1\%$ improvement respectively. HiTeA also demonstrates strong generalization ability when directly transferred to downstream tasks in a zero-shot manner. + +# 1. Introduction + +Vision and language are two primary signals that constitute the real-world perception of humanity. With the success of image-language pre-training [8,22,25,47], video-language pre-training [23,26,27,34] has recently received increasing attention. Large-scale video-language pre-training helps the model to learn effective multi-modal representation, which has shown significant improvement on a variety of video-language downstream tasks, such as videotext retrieval, video question answering and video caption- + +![](images/483f83779b42714bc182be16f6d694585fb415a0441b17fe29454fad8bd01314.jpg) +Figure 1: Comparison between existing paradigms and ours for video-language pre-training. (a) Previous methods align video and text within global perspective as the pretext. (b) We introduce HiTeA by varying video in different temporal views and modeling cross-modal temporal information between moments and texts, as well as the temporal contextual relations between multi-modal pairs. + +ing [5,33,44,48,50,53,59]. + +Inspired by the success of image-language pre-training paradigm, various methods [10, 11, 23, 26, 27] have been proposed to adapt it to video-language pre-training. ClipBERT [21] and Singularity [20] directly build on representations from image encoders and aggregate them via score aggregation function and temporal encoder. Furthermore, MILNCE [34] and Frozen [2] switch image encoder to video encoder for spatio-temporal video representation learning and align the video with corresponding text. In addition, some advanced pre-training tasks are designed through modeling entity [23], reconstructing masked patches [10] and predicting frame order [26, 61]. Despite their promising performance on downstream tasks, they treat video within global perspective illustrated in Figure 1(a), thus failing to consider fine-grained temporal information and temporal contextual relations which are essential to video-language pre-training. + +Since untrimmed video contains various temporal details, + +directly treating the video globally has two main limitations: (1) Less effective in modeling the fine-grained moment information including atomic actions and moments. As illustrated in Figure 1(b), we vary time resolutions and generate two views (long & short) for the input video. As a result, the short-view video clip tends to represent the moment information and the long-view video may express more event-level information. For example, the short-view video clip in Figure 1(b) only describes the moment of "lick fingers" rather than "eating ice cream". Such fine-grained moment information is hard to be captured by the long-view video under global event perspective; (2) Ignoring the temporal contextual relations implicitly existed in the video. Knowing the event expressed by the text, the moment "eating ice cream" can be inferred from the moment "lick fingers" shown by short-view video. However, such implicit temporal contextual relations between the moment and the event are rarely explored in previous works. + +To address these problems, we propose a Hierarchical Temporal-Aware video-language pre-training framework, HiTeA, for both multi-modal understanding and generation. Except for the standard pre-training tasks, HiTeA introduces two novel temporal-aware video-language pre-training tasks, named cross-modal moment exploration (CME) and multimodal temporal relation exploration (MTRE), which not only model the fine-grained temporal moment information but also captures temporal contextual information hierarchically, yielding temporal-aware multi-modal representations for both understanding and generation. Specifically, we first generate the long-view and short-view videos with different time resolutions to build hierarchy of the input video. Then, based on the similarities of words and short-view video, we select the most relevant words as positive and leave the rest of the words as hard negatives. The CME pre-training task is applied to align the positive words and short-view video representations in the same embedding space. Moreover, to capture association between moments and the event for temporal contextual modeling, we match different views for the same video. However, directly matching two views visually would be noisy due to the background similarity [39]. To this end, we perform multi-modal pair alignment between videotext pairs via the MTRE pre-training task. More specifically, the short-view video guided by most relevant words and the long-view video guided by text will be aligned, which enables the model to extrapolate the contextual information from the short-view with language signal while enhancing temporal reasoning ability. Empowered by above two novel temporal-aware video-language pre-training tasks, HiTeA is capable of modeling temporal-aware multi-modal information revealed in video-text data including both fine-grained moment information and temporal contextual relations. + +In spite of a good performance, recent studies [4, 20] reveal most video-language downstream datasets are biased to + +wards still objects, scenes, etc., while the temporal dynamics are negligible. To evaluate the temporal performance of the video-language pre-training model and temporal reliance of downstream datasets, we introduce temporal shuffling test for these datasets. This enables a comprehensive evaluation of temporal modeling capability in the video-language pre-training field. Besides, our method achieves significant improvement on the datasets with heavy temporal reliance. + +In summary, our key contributions are the followings: + +- We propose a novel hierarchical temporal-aware video-language pre-training framework with both video-language understanding and generation capabilities. +- We introduce temporal-aware pre-training tasks to generate temporal-aware multi-modal representation through modeling fine-grained temporal moment information as well as capturing the temporal contextual relations between moment and event. +- Extensive experiments demonstrate the effectiveness and generalization ability of HiTeA, and it achieves state-of-the-art performance on 15 video-language downstream datasets including video-text retrieval, video question answering, and video captioning, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with $8.6\%$ and $11.1\%$ improvement respectively. + +# 2. Related Work + +Video-Language Pre-training Benefiting from a large number of image/video-text pairs, video-language pretraining (VLP) exhibits superior capabilities on various video-text benchmarks. The method of VLP is constantly evolving. Traditional approaches [26, 30, 42, 62] leverage offline-extracted dense video features for pre-training to circumvent the expensive computation overhead. In contrast, ClipBERT [21] suggests that sparse sampling can enable affordable end-to-end learning and improve performance simultaneously. Recent emerging approaches [2, 11, 14, 21, 23, 27, 52, 57] adopt this strategy and propose new model architectures and pre-training tasks. Frozen [2] trains jointly on image and video datasets via video-text contrastive learning (VTC). ALPRO [23] proposes a new visually-grounded pre-training task combined with VTC, video-text matching (VTM) and masked language modeling (MLM) [9] to learn fine-grained region-entity alignment. LAVENDER [27] formulates all pre-training and downstream tasks as MLM so that a unified architecture can be used for all video-text tasks. Apart from above representative works, frame order modeling (FOM) [26, 61] and masked video modeling (MVM) [10] are designed for VLP. However, the temporal characteristic of video still remains largely unexplored. To this end, we introduce a novel hierar + +![](images/2ea68e067c43e3539126ddd5b67b4506b70cb716b5392357a635800d66589dc0.jpg) +Figure 2: Illustration of the proposed HiTeA. We first generate two different temporal views for the input video, where the long-view is the video itself and the short-view is randomly truncated from the input video. To explore the moment revealed in the short-view, cross-modal moment exploration (CME) selects the candidate words from the input text with $\mathcal{L}_{\mathrm{CME}}$ . Then, we perform multi-modal temporal relation exploration (MTRE) for modeling the temporal contextual relations between two video-text pairs with different views by $\mathcal{L}_{\mathrm{MTRE}}$ . Note that the multi-modal encoders and the text features are shared. + +chical temporal-aware VLP framework which not only models the fine-grained moment information but also captures contextual relations with different temporal granularities. + +Temporal Modeling The temporal characteristic acts as a vital role in VLP since it provides the model with the capabilities of reasoning and understanding causality. Previous efforts in this field can be roughly divided into three categories. First, several methods directly transfer image-text models to video-text tasks by simply concatenating video frame [22,24] or building an extra temporal encoder [31,32]. Second, some works [2, 10, 23, 27] switch the image encoder to video encoder for learning spatio-temporal contexts within videos. Third, HERO [26] and MERLOT [61] design FOM task to explicitly recover the correct temporal order of shuffled frames. Nonetheless, ATP [4] and Singularity [20] reveal the existence of a static appearance bias in popular video-language datasets, and they develop single-frame models to achieve strong performance, comparable or even better than above methods with explicit temporal modeling. Therefore, they recommend SSv2 [20] and NExT-QA [48] datasets to test the temporal ability of VLP models. Different from previous approaches, we vary the temporal resolutions and generate two views of video so as to construct the temporal hierarchy, which equips the model with the ability to learn both fine-grained temporal moment information and temporal contextual representation at the same time. + +# 3. Method + +# 3.1. Overview + +Figure 2 sketches the overview of the HiTeA. In concrete, our model consists of two unimodal encoders for encoding video and text separately, a multi-modal encoder for video and text interaction, and a text decoder for generation which is omitted here for simplicity and detailed in Appendix. + +For video representation, previous methods [21, 23, 27] encode the whole input video as a single-view feature, ignoring the rich temporal details contained in the video. Thus, we first treat the video into two views with different time resolutions to build hierarchy of the input video. Specially, the untrimmed video is regarded as a long-view $V^L$ for capturing event information, and a video segment is randomly truncated from the input video as the short-view for capturing moment information denoted as $V^S$ . Then, we use the video encoder to encode an arbitrary view of video $V \in \mathbb{R}^{T \times H \times W}$ into a sequence of embeddings: $\mathcal{V} = \{v_{\mathrm{cls}}, v_1, \dots, v_M\} \in \mathbb{R}^{M \times D}$ , where $M$ is the number of flattened patches, $D$ is hidden size, and $v_{\mathrm{cls}}$ is the embedding of the visual [CLS] token which provides global representation of the video. For text representation, we use the text encoder to transform the text $T$ into a sequence of embeddings: $\mathcal{T} = \{w_{\mathrm{cls}}, w_1, \dots, w_N\} \in \mathbb{R}^{N \times D}$ , where $N$ is length of the text. After that, the multi-modal encoder takes video features $\mathcal{V}$ and text features $\mathcal{T}$ as inputs and yields the multi-modal representation $v_{\mathrm{cls}}$ for the video. + +In order to take full advantage of the different views of the video, we introduce cross-modal moment exploration (CEM) to explore the proper words or phrases from input text to align the short-view video with $\mathcal{L}_{\mathrm{CME}}$ for capturing the fine-grained temporal moment information in Section 3.2. Furthermore, to model the temporal contextual relations between the short-view video containing moment information and the long-view video with event information, we propose multi-modal temporal relation exploration (MTRE) to match the multi-modal representation of shortview and long-view videos by $\mathcal{L}_{\mathrm{MTRE}}$ in Section 3.3. Lastly, we introduce the overall pre-training objective for training the model in Section 3.4. + +# 3.2. Cross-Modal Moment Exploration + +The video with short temporal range (i.e., short-view of video) with the paired text tends to be accompanied with fine-grained temporal moment information. However, the paired text partially describes the short-view of video bringing noise to the fine-grained moment representation thus degrading the performance. To this end, we propose a novel pre-training task named cross-modal moment exploration (CME), which enables the model to understand fine-grained moment information by leveraging the partially aligned text. + +Formally, we first discover the possible positive words for the video in short-view by computing the cosine similarity of the word embedding sequence $\{w_{1},\dots ,w_{N}\}$ from text encoder and the short-view video representation $v_{\mathrm{cls}}^S$ from video encoder as: + +$$ +\mathcal {K} = \{\pi (1), \dots , \pi (K) \}, \tag {1} +$$ + +where $\pi : \{1, \dots, N\} \to \{1, \dots, N\}$ is a permutation function for ranking such that $s(w_{\pi(1)}, v_{\mathrm{cls}}^S) \geq \dots \geq s(w_{\pi(N)}, v_{\mathrm{cls}}^S)$ , and $\mathcal{K}$ is the set of selected word indices, $K$ is the number of possible selected words, and $s(x, y) = x^T y / \|x\|_2 \|y\|_2$ represents the cosine similarity between $x$ and $y$ . After obtaining the words for the video in short-view as the positive pair, the cross-modal moment exploration loss $\mathcal{L}_{\mathrm{CME}}$ is computed with negative pairs from other words in the input text, which is defined as: + +$$ +\mathcal {L} _ {\mathrm {C M E}} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \left(\frac {1}{| \mathcal {K} |} \sum_ {k \in \mathcal {K}} \log \frac {\exp \left(\left(v _ {\mathrm {c l s}} ^ {S}\right) _ {i} ^ {\top} w _ {i , k} / \tau\right)}{\sum_ {n = 1} ^ {N} \exp \left(\left(v _ {\mathrm {c l s}} ^ {S}\right) _ {i} ^ {\top} w _ {i , n} / \tau\right)}\right), \tag {2} +$$ + +where $\tau$ is the learnable temperature hyper-parameter that controls the sharpness of the output distribution, and it is initialized as 0.07. As a consequence, the model is able to understand moment information via the proposed cross-modal exploration scheme. + +# 3.3. Multi-Modal Temporal Relation Exploration + +While the video encoder has demonstrated its effectiveness in learning temporal representation implicitly + +[10, 23, 56], it remains a challenge to discover the inherent temporal contextual relations. As a result, the limited capabilities in temporal modeling deteriorate the downstream task in temporal reasoning. This is in particular a missing point for the existing video-language pre-training paradigm [11, 21, 23, 27], which usually focuses on bridging video and text neglecting the function of text for guiding the video temporal contextual representation learning thus losing the temporal cues. + +To this end, we introduce multi-modal temporal relation exploration (MTRE), a novel temporal-aware pre-training task that improves models' capacities in capturing temporal context in video with fine-grained text guidance by aligning multi-modal pairs. Specially, the short-view video $V^{S}$ would represent moment information with respect to the whole video. On the contrary, the long-view video $V^{L}$ expresses the event and topical information. To obtain the text-guided video features, we feed videos in different temporal views into the video encoder individually. Then, the text features are extracted and interact with the video features by the multi-modal encoder and yield text-guided video representations $\boldsymbol{v}_{\mathrm{cls}}^{L} \in \mathbb{R}^{D}$ and $\boldsymbol{v}_{\mathrm{cls}}^{S} \in \mathbb{R}^{D}$ as follows: + +$$ +\boldsymbol {v} _ {\mathrm {c l s}} ^ {L} = f \left(\left\{v _ {\mathrm {c l s}} ^ {L}, v _ {1} ^ {L}, \dots , v _ {M} ^ {L} \right\}, \left\{w _ {\mathrm {c l s}}, w _ {1}, \dots , w _ {N} \right\}\right), \tag {3} +$$ + +$$ +\boldsymbol {v} _ {\mathrm {c l s}} ^ {S} = f \left(\left\{v _ {\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \dots , v _ {M} ^ {S} \right\}, \left\{w _ {\mathrm {c l s}}, w _ {1}, \dots , w _ {N} \right\}\right), \tag {4} +$$ + +where $f(\mathcal{V},\mathcal{T})$ represents the multi-modal encoder with video features $\mathcal{V}$ and text features $\mathcal{T}$ . However, since the short-view of the video is partially aligned with the text, using the whole text is not reasonable for generating accurate text-guided video feature for short-view. Meanwhile, improper video-text pairs would yield noisy multi-modal representation thus degrading the performance of the model. Therefore, thanks to the positive words mined by cross-modal moment exploration, we can calibrate representation for short-view video by: + +$$ +\boldsymbol {v} _ {\mathrm {c l s}} ^ {S} = f \left(\left\{v _ {\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \dots , v _ {M} ^ {S} \right\}, \left\{w _ {\mathcal {K} _ {1}}, \dots , w _ {\mathcal {K} _ {K}} \right\}\right), \tag {5} +$$ + +where $\mathcal{K}_i\in \mathcal{K}$ is the index of the set for selected positive words. Then, we aim to match the representation of produced text-guided video features in different granularities in order to enable the model to predict the past and the future from the short-view of video, which benefits for capturing the general structure of the video. Specifically, we adopt the SimSiam framework [6] for minimizing their negative cosine similarity: + +$$ +\mathcal {D} \left(p ^ {S}, z ^ {L}\right) = - \frac {p ^ {S}}{\| p ^ {S} \| _ {2}} \cdot \frac {z ^ {L}}{\| z ^ {L} \| _ {2}}, \tag {6} +$$ + +where $p^S = h(g(\mathbf{v}_{\mathrm{cls}}^S))$ and $z^{L} = g(\mathbf{v}_{\mathrm{cls}}^{L})$ . The $g$ and $h$ are projection MLP head and prediction MLP head [7,13]. Minimizing $\mathcal{D}(p^S,z^L)$ is equivalent for minimizing the mean + +
Method# PT DataMSRVTTDiDeMoLSMDCActivityNet Caption
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
ClipBERT [21]0.2M22.046.859.920.448.060.8---21.349.063.5
Frozen [2]5M31.059.570.531.059.872.415.030.839.8---
ALPRO [23]5M33.960.773.235.967.578.8------
BridgeFormer [11]5M37.664.875.137.062.273.917.935.444.5---
Singularity [20]5M36.865.975.547.475.284.0---43.070.681.3
LAVENDER [27]5M37.863.875.047.474.782.422.243.853.5---
+ +Models pre-trained on more data + +
VIOLET [10]183M34.563.073.432.662.874.716.136.641.2---
All-in-one [46]138M37.968.177.132.761.473.5---22.453.767.7
Clip4Clip [31]400M42.171.981.443.470.280.621.641.849.840.572.4-
X-CLIP [32]400M46.173.083.145.274.0-23.343.0-44.374.1-
HiTeA5M44.469.378.951.879.185.327.146.254.545.173.584.2
+ +Table 1: Performance comparison on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more pre-training data for fair comparison. # PT Data is the number of video-text pairs for pre-training. + +square error between $p^S$ and $z^L$ , which encourages the videos in different temporal magnitudes to be similar. Following [6, 13], we defined a symmetrized loss as: + +$$ +\mathcal {L} _ {\mathrm {M T R E}} = \frac {1}{2} \left[ \mathcal {D} \left(p ^ {L}, \mathbf {s g} \left(z ^ {S}\right)\right) + \mathcal {D} \left(p ^ {S}, \mathbf {s g} \left(z ^ {L}\right)\right) \right], \tag {7} +$$ + +where $\mathsf{sg}(\cdot)$ is the stop-gradient operation that prevents the model from collapse during training [6]. + +# 3.4. Pre-training Objectives + +Apart from the two proposed temporal-aware pre-training tasks, we follow proven video-text pre-training approaches [2, 23, 27] to adopt the standard pre-training tasks including video-text contrastive (VTC), video-text matching (VTM), masked language modeling (MLM), and prefix language modeling (PrefixLM) described in the related work. Precisely, VTC and VTM align the video and text from the global perspective, while MLM and PrefixLM contribute to multi-modal understanding and generation capabilities of the model. Details of these objectives are described in the Appendix. We simply combine these as the base training objective $\mathcal{L}_{base}$ for our model. Therefore, the full pre-training objective is computed as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {b a s e}} + \mathcal {L} _ {\text {C M E}} + \mathcal {L} _ {\text {M T R E}}. \tag {8} +$$ + +# 4. Experiments + +# 4.1. Experiment Setup + +Pre-training Datasets Following the recent work [2,11,20, 23,27], we pre-train our model on a webly-sourced video dataset WebVid-2M [2] with 2.5M video-text pairs and a image-text dataset Google Conceptual Captions (CC3M) [41] with 3M image-text pairs. Unlike previous methods, we do not pre-train our model on the large-scale video-text datasets like HowTo100M [34] with 136M video-text pairs and YT-Temporal-180M [61] due to the heavy computation. + +Downstream Datasets We evaluate our pre-trained model on 18 video-language benchmarks including video-text retrieval, video question answering, and video captioning tasks. Specifically, video question answering (VideoQA) can be categorized as Multiple-Choice (MC) and Open-Ended (OE) settings. The evaluation datasets are briefly summarized in below. Details can be found in the Appendix. + +- Video-Text Retrieval: MSRVTT [53], DiDeMo [1], LSMDC [38], ActivityNet Caption [18], SSv2-Label [20], and SSv2-Template [20]; +- VideoQA (MC): TGIF-Action, TGIF-Transition [15], MSRVTT-MC [58], LSMDC-MC [44], and NExT-QA [48]; +- VideoQA (OE): TGIF-Frame [15], MSRVTT-QA, MSVD-QA [50], LSMDC-FiB [33] and ActivityNet-QA [59]. +- Video Captioning: MSRVTT [53] and MSVD [5]. + +Implementation Details Our implementation of HiTeA is based on PyTorch [35]. In detail, we instantiate the video encoder with MViT-Base model [28] pretrained on ImageNet-21K [37]. The text encoder is initialized from first six layers of pre-trained BERT-Base [9], and the multi-modal encoder is initialized with last six layers of pre-trained BERT-Base. We pre-train HiTeA for 10 epochs, using a batch size of 16 on 8 NVIDIA A100 GPUs. We use AdamW [17] optimizer with a weight decay of 0.02 and betas (0.9, 0.98). The learning rate is first warmed up to 5e-5 in the first 1000 iterations, and decays following a cosine schedule. During pre-training stage, following with [10, 14, 23], we sparsely sample 4 frames for short and long view while preserving their order in-between and resize them to $224 \times 224$ . The duration of short view is restricted as the 1/8 of the whole video duration. $K$ is empirically set to 5. The MLM mask ratio is set to $15\%$ . Details of fine-tuning stage are described in Appendix. + +
Method#PT DataTGIFMSRVTTLSMDCMSVDActivityNet
ActionTransitionFrameMCQAMCFiBQAQA
ClipBERT [21]0.2M82.887.860.388.237.4----
ALPRO [23]5M----42.1--46.3-
Singularity [20]5M---92.042.7---41.8
LAVENDER [27]5M96.699.172.296.644.286.056.955.4-
Clover [14]5M94.998.071.495.043.983.254.151.9-
+ +
Models pre-trained on more data
VIOLET [10]183M92.595.768.991.943.982.853.747.938.9
JustAsk [54]69M----41.5--46.338.9
MERLOT [61]180M94.096.269.590.943.181.752.9-41.4
All-in-one [46]283M95.594.766.392.346.884.4-48.3-
HiTeA5M96.898.872.597.245.485.854.655.645.1
+ +Table 2: Performance comparison on video question answering. Accuracy is reported for evaluation. We gray out methods that use significantly more pre-training data for fair comparison. + +
Method# PT DataMSRVTTMSVD
UniVL [30]180M49.9-
SwinBERT [29]-53.8120.6
MV-GPT [40]53M60.0-
CLIP4Caption [43]400M57.7-
LAVENDER [27]5M58.0142.9
HiTeA5M62.5145.1
+ +Table 3: Performance comparison on video captioning. CIDEr [45] is reported for evaluation. + +
MethodMSRVTT-Ret.LSMDC-Ret.MSRVTT-QAMSVD-QA
Frozen [2]31.0/59.5/70.515.0/30.8/39.8--
ALPRO [23]33.9/60.7/73.2-/-/-42.146.3
BridgeFormer [11]37.6/64.8/75.117.9/35.4/44.5--
Singularity [20]36.8/65.9/75.5-/-/-42.7-
TimeSformer (Lbase)37.6/62.0/72.319.2/38.7/48.041.549.0
HiTeA with TimeSformer39.7/65.0/75.121.8/40.7/49.943.752.4
+ +Table 4: Performance comparison of different SoTA methods with TimeSformer [3]. For text-to-video retrieval, Recall@1/5/10 are reported. For video question answering task, we report the Top-1 accuracy. + +# 4.2. Comparison to Prior Arts + +In this section, we compare HiTeA with numerous state-of-the-art video-language pre-training methods on several downstream datasets under fine-tuning setting. + +# 4.2.1 Text-to-Video Retrieval + +Table 1 summarizes the results on MSRVTT [53], DiDeMo [1], LSMDC [38], and ActivityNet Caption [18] under fine-tuning settings. Our method outperforms all of the existing video-language pre-training model by a large margin under the same data scale. In particular, our method yields $6.6\%$ lift in terms of R@1 on MSRVTT dataset while only exploiting 5M video-text pairs. Note that we also include the comparison with the recent works that utilize the powerful encoder from CLIP [36], our method still can be comparable with them even surpass them, which shows the validity of the proposed method. Besides, we can notice that our method achieves the best result among all of listed methods + +on LSMDC dataset, which proves that our model can leverage the various moments presented in fruitful movie clips with cross-modal moment exploration. + +# 4.2.2 Video Question Answering + +Table 2 lists the results of HiTeA and current state-of-the-art approaches on nine VideoQA datasets. It can be noticed that our method achieves the best performance in most of VideoQA datasets even with less pre-training data. Specifically, it achieves absolute improvement $1.1\%$ on TGIF-FrameQA, $2.2\%$ on MSRVTT-MC, $1.5\%$ on MSRVTT-QA, $0.2\%$ on MSVD-QA, and $3.3\%$ on ActivityNet-QA. We believe the moments learned by the cross-modal exploration are useful for finding the clue of answers in VideoQA. + +# 4.2.3 Video Captioning + +Table 3 compares HiTeA with existing methods on video captioning datasets MSRVTT and MSVD. As shown in the table, although we use less pre-training data than compared approaches, HiTeA still obtains significant improvement compared to those large-scale pre-trained models. On MSRVTT Caption, our method surpasses SoTA method MV-GPT [40] by $2.5\%$ CIDEr. Note that MV-GPT is pretrained for multi-modal video captioning and it leverages the ASR transcripts from audio as the additional input. By contrast, our method only utilizes video as the input during generation. + +# 4.3. Discussion + +In this section, we discuss the temporal characteristics of our model and the datasets. + +Generalization on Plain Backbone. Table 4 delivers the performance of SoTA methods with the plain visual backbone. We instantiate the video encoder with TimeSformer [3] pretrained on ImageNet-21K [37] following [2, 11, 23]. We can observe that HiTeA with TimeSformer significantly + +
MethodMSRVTT-RetrievalSSv2 Template-RetrievalNExT-QA (Hard)MSVD-QA
R@1R@5R@10AveRR@1R@5R@10AveRAcc@CAcc@TAcc.
\( \mathcal{L}_{\text{base}} \)40.068.077.161.780.5100.0100.093.544.046.452.7
\( \mathcal{L}_{\text{base}} + \mathcal{L}_{\text{MTRE}} \)41.669.178.263.083.398.9100.094.146.346.454.8
\( \mathcal{L}_{\text{base}} + \mathcal{L}_{\text{CME}} \)42.069.379.763.783.999.4100.094.446.348.354.3
\( \mathcal{L}_{\text{base}} + \mathcal{L}_{\text{CME}} + \mathcal{L}_{\text{MTRE}} \)44.469.378.964.285.6100.0100.095.247.848.655.6
+ +Table 5: Evaluation of the proposed methods on four downstream video-language tasks. For text-to-video retrieval, R@1, R@5, R@10, and the average are reported. For video question answering, we report the accuracy. + +
Method# PT DataSSv2-LabelSSv2-Template
R@1R@5R@10R@1R@5R@10
Frozen [2]5M---52.994.899.4
Clip4Clip [31]400M43.171.480.777.096.698.3
Singularity [20]5M44.173.582.277.098.999.4
HiTeA5M55.281.489.185.6100.0100.0
+ +Table 6: Comparison of existing methods on Something-to-Something (SSv2) text-to-video retrieval. + +
Method# PT DataAcc@CAcc@TAcc@DAcc.
Full Set
Human-87.688.690.488.4
HCRN [19]-45.949.353.748.2
HGA [16]-46.350.759.349.7
VGT [49]0.18M53.456.469.556.9
HGA* [16]400M46.852.159.350.4
ATP [4]400M51.350.266.854.3
HiTeA5M62.458.375.663.1
Hard Split
ATP [4]400M38.436.5//
HGA [16]-43.345.3//
HiTeA5M47.848.6//
+ +Table 7: Comparison of existing methods on NExT-QA [48]. We report accuracy on the Causal (C), Temporal (T), Descriptive (D) splits and overall accuracy on validation set. * stands for using CLIP as the initialization of visual encoder. + +
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT [53]64.263.30.9
DiDeMo [1]72.170.21.9
LSMDC [38]42.641.70.9
ActivityNet Caption [18]67.666.80.8
SSv2 Template [20]95.272.422.8
SSv2 Label [20]76.773.53.2
+ +Table 8: Dependency on temporal information for text-to-video retrieval datasets with temporal shuffling test. The average recall of Recall@1, Recall@5, and Recall@10 are reported. We evaluate the performance drop when shuffling the input during inference. "Original" and "Shuffled" denote the original and shuffled input videos, respectively, and "Gap" is the difference between the Original and Shuffled metric. The larger "Gap" indicates the dataset relies on temporal information, and the model utilizes more temporal information to solve the task. + +outperforms other SoTA methods (e.g., ALPRO, Bridge-Former, and Singularity) with the same pre-training data and the number of frames. In addition, it shows that our proposed temporal-aware pre-training tasks can boost the + +
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT-QA [50]45.445.20.2
MSVD-QA [50]55.655.50.1
TGIF-FrameQA [15]72.572.10.4
ActivityNet-QA [59]45.145.00.1
NExT-QA (Hard) [48]47.145.60.5
+ +Table 9: Dependency on temporal information for video question answering datasets by temporal shuffling test. We report the accuracy for each dataset. For NExT-QA dataset, we evaluate with the hard split of the validation set [4]. + +performance of TimeSformer $(\mathcal{L}_{\mathrm{base}})$ with $2.6\%$ average improvement on MSRVTT-Retrieval and $3.4\%$ on MSVD-QA. More related discussions are demonstrated in the Appendix. + +Impact of Loss Terms. We investigate the contribution of individual loss terms and the results are shown in Table 5. It can be observed that the combining both $\mathcal{L}_{\mathrm{CME}}$ and $\mathcal{L}_{\mathrm{MTRE}}$ improves the performance of text-to-video retrieval and video question answering by at least $1.7\%$ and $2.9\%$ in Average Recall and Average accuracy respectively. In addition, we also find that the performance of $\mathcal{L}_{\mathrm{CME}}$ surpasses that of $\mathcal{L}_{\mathrm{MTRE}}$ on MSRVTT retrieval dataset that largely dominated by the appearance information. This can be explained that the cross-modal moment exploration loss not only select the positive verbs for the video from the text but also choose the acting object for alignment, which can boost the retrieval performance. + +Evaluation on Temporal-aware Tasks. Lei et al. [20] reveal that the previous four retrieval datasets are prone to being biased for appearance while rarely relying on temporal information, thus introducing Something-to-Something v2 (SSv2) Template and SSv2 Label retrieval datasets to test models' true temporal modeling capability. In particular, SSv2 Template retrieval task requires a deeper understanding of the moment and temporal relation since no objects information are presented. The performance on these datasets are summarized in Table 6. It can be observed that HiTeA achieves significant improvement with $+8.5\%$ gains in terms of R@1 on these two temporal-oriented text-to-video retrieval datasets, which demonstrates the effectiveness of our proposed method through exploring fine-grained moment information and modeling temporal relation. In addition, we evaluate our model on NExT-QA [48] dataset that explicitly designed for temporal and causal understanding. + +![](images/256a9cf39bd6f337d83a0cedc2418b23b2eb84a26c6ffd8b9e220895a453cbf0.jpg) +"Cartoon characters are talking to a pokemon" +Video +Frames +Baseline +Ours +Figure 3: Visualizations of learned cross-attention maps from multi-modal encoder. We present samples from MSRVTT [53] and SSv2 Template [20] retrieval dataset. HiTeA attends to the patches related to objects motion by tracking trajectory. + +![](images/734fdad53243f68a2c60e2b68744a6772c584a614e769197cc9ff9281ca3d92c.jpg) +"Poking [something] so that it spins around" +Video +Frames + +![](images/9e71569ad943b5e14491b9f8b713a210dc182d8412095f3836cdc978df6f21d3.jpg) +Baseline + +![](images/6f09a43a20b52cab24ae1cf7b6496a600bb8f5239fbca6ce0d2f16de8ca7f793.jpg) + +As presented in Table 7, our method significantly surpasses its competitive counterparts, even those methods equipped with powerful image-text pre-trained encoders. Quantitatively, HiTeA obtains an absolute improvement $+9\%$ on the causality split with the help of intrinsic temporal relation. Recently, Buch et al. [4] filter out the trivial question for the dataset, and build the hard split for causality and temporal related questions for evaluate the causality and temporal of the model. As we can see in the table, even for the questions that heavily rely on causality, our model can still achieves a relative gain of $4.1\%$ on the model with specific design for VideoQA, which indicates that our model do not solely depend on static appearance. + +Temporal Reliance of Datasets. Previous methods [11,21, 23] only evaluate the performance of models on the existing datasets to demonstrate the superiority of the methods. However, Buch et al. [4] and Lei et al. [20] reveal that the most of the evaluation are biased towards the static concepts. Here, we investigate the temporal reliance for the evaluated datasets by introducing the temporal shuffling test. Specifically, we compute the performance changes between running inference on the ordered video versus its shuffled version. The large performance drop indicates the dataset has less spatial bias and needs for temporal information. Table 8 and Table 9 conclude the performance gap between ordered and shuffled input video for text-to-video retrieval and VideoQA datasets. For text-to-video retrieval task, SSv2 Template shows the large performance drop after shuffling the input video, which demonstrates that it depends most on the dynamic information thus verifying our assumption. On the contrary, the performance on ActivityNet Caption dataset is barely affected (-0.8 on Mean Recall) since the text almost describes the static objects without relying on temporal information. For video question answering dataset, we observe that the MSVD-QA and ActivityNet-QA are less sensitive to the order of video frames. This is because these two datasets contain more questions requiring frame-region information, such as object categories, scenes, and species. We believe this can be used to evaluate the temporal reliance of the datasets as well as the utilization for temporal cue by + +
Method# PT DataMSRVTT-QAMSVD-QA
Just Ask [54]69M2.97.5
LAVENDER [27]5M4.511.6
MERLOT Reserve [60]1B5.8-
FrozenBiLM [55]10M6.411.7
HiTeA5M8.618.2
BLIP [24]129M19.235.2
mPLUG [22]400M21.137.2
HiTeA5M21.737.4
+ +Table 10: Zero-shot evaluation on video question answering. Accuracy is reported. We gray out those methods additionally supervised pre-training on VQA v2 [12] dataset. + +models in the future work. + +Qualitative Analysis. To verify that our model can capture the motion information with respected to the given text rather than inferring from the static signal, we present the query text in SSv2 Template dataset which has masked all of the object information, and also visualize the query in MSRVTT dataset. As we can see in the Figure 3, the attention map of atomic action "talking" mainly focuses on the mouse of the cartoon characters while the baseline largely focusing on the characters, which indicates that our method can understand the moment better when adopting the temporal-aware pretraining tasks. In another example, the word "spins" can reveal the trajectory of the object showing that our method is able to capture the temporal motion presented in the video. + +# 4.4. Zero-shot Generalizability + +To demonstrate the generalizability of proposed videotext pre-trained model, we perform zero-shot evaluation on video-language downstream tasks. We evaluate the zero-shot performance on VideoQA task in Table 10. Our method attains competitive zero-shot performance on MSRVTT-QA and MSVD-QA datasets even without help of audio signal supervision [60] or additional generated video question pairs [54]. In particular, less pre-training data (i.e. $5\mathrm{M} < 69\mathrm{M}$ ) are used while our method can still outperform other SoTA approaches. We also evaluate the zero-shot performance of models supervised on VQA v2 [12]. We can find that our method surpasses the powerful multi-modal SoTA methods (e.g. mPLUG [22]) with only 5M pre-training data showing the better generalization ability of HiTeA. + +
Method# PT DataMSRVTTDiDeMoLSMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
Frozen [2]5M18.739.551.621.146.056.29.322.030.1
ALPRO [23]5M24.144.755.423.847.357.9---
BridgeFormer [11]5M26.046.456.425.650.661.112.225.932.2
Singularity [20]5M28.450.259.536.961.669.3---
+ +Models pre-trained on more data + +
VideoCLIP [51]138M10.422.230.016.646.9----
VIOLET [10]183M25.949.559.723.549.859.8---
Clip4Clip [31]400M32.057.066.9---15.128.536.4
HiTeA5M29.954.262.936.160.170.315.531.139.8
+ +Table 11: Zero-shot evaluation on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more data for fair comparison. + +We also perform zero-shot evaluation on text-to-video retrieval task. Table 11 summarizes the performance of our model and compared approaches on text-to-video retrieval. We can observe that our model yields more than $3.4\%$ lift in R@1 on MSRVTT dataset [53] while exploiting fewer video-text pairs. Besides, our method surpasses all of the compared models in terms of LSMDC dataset showing the superiority of our method's generalizability. + +# 5. Conclusion + +In this work, we introduce HiTeA, a novel hierarchical temporal-aware video-language pre-training framework with both understanding and generation capabilities. We vary the video with different views and model cross-modal temporal information between moments and texts as well as temporal contextual relations between multi-modal pairs in a hierarchical way. Specifically, a cross-modal moment exploration pre-training task is proposed to explore the fine-grained temporal information between the paired text and video moment by overcoming the partially semantic alignment. Moreover, multi-modal pairs are constructed to learn temporal contextual relations between moments and the event presented by the video with multi-modal temporal relation exploration pre-training task. We also demonstrate that our proposed pre-training tasks can consistently boost performance significantly on downstream tasks regardless of the backbone showing the generalization ability. Even pre-trained on less data, HiTeA still achieves state-of-the-art performance on a wide range of video-language downstream datasets, which clearly shows the superiority of our method. + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6, 7 +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to- + +end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 3, 5, 6, 7, 9 +[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 6 +[4] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the" video" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 2, 3, 7, 8 +[5] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 1, 5 +[6] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4, 5 +[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 4 +[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5 +[10] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 4, 5, 6, 9 +[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 4, 5, 6, 8, 9 + +[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8 +[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4, 5 +[14] Jingjia Huang, Yinan Li, Jiashi Feng, Xiaoshuai Sun, and Rongrong Ji. Clover: Towards a unified video-language alignment and fusion model. arXiv preprint arXiv:2207.07885, 2022. 2, 5, 6 +[15] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 5, 7 +[16] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 7 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[18] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5, 6, 7 +[19] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for multimodal video question answering. International Journal of Computer Vision, 129(11):3027-3050, 2021. 7 +[20] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 2, 3, 5, 6, 7, 8, 9 +[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 4, 5, 6, 8 +[22] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 1, 3, 8 +[23] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 3, 4, 5, 6, 8, 9 +[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni- + +fied vision-language understanding and generation. In ICML, 2022. 3, 8 +[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 1 +[26] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1, 2, 3 +[27] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. Lavender: Unifying videolanguage understanding as masked language modeling. arXiv preprint arXiv:2206.07160, 2022. 1, 2, 3, 4, 5, 6, 8 +[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5 +[29] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 6 +[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2, 6 +[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3, 5, 7, 9 +[32] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022. 3, 5 +[33] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5 +[34] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 5 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 + +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6 +[37] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5, 6 +[38] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3202-3212, 2015. 5, 6, 7 +[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022. 2 +[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 6 +[41] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 5 +[42] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2 +[43] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 6 +[44] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5 +[45] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6 +[46] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022.5.6 +[47] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 1 +[48] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9777–9786, 2021. 1, 3, 5, 7 + +[49] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. arXiv preprint arXiv:2207.05342, 2022. 7 +[50] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 5, 7 +[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 9 +[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. In Proceedings of International Conference of Machine Learning (ICML), 2023. 2 +[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 7, 8, 9 +[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 6, 8 +[55] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 8 +[56] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 4 +[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 +[58] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 5 +[59] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9127–9134, 2019. 1, 5, 7 +[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 8 + +[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 1, 2, 3, 5, 6 +[62] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2 \ No newline at end of file diff --git a/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/images.zip b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c262e204ade1efa100c0c991a815cd16f5964de0 --- /dev/null +++ b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78619e2833af929f6142ea04ad082c094f3f4bfd9a1ff5d94db3827a48ee2d7a +size 602232 diff --git a/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/layout.json b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cd0d0fcf89d82b6dc074a7fa2e1663869d181204 --- /dev/null +++ b/2023/HiTeA_ Hierarchical Temporal-Aware Video-Language Pre-training/layout.json @@ -0,0 +1,8916 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 71, + 506, + 89 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 71, + 506, + 89 + ], + "spans": [ + { + "bbox": [ + 87, + 71, + 506, + 89 + ], + "type": "text", + "content": "HiTeA: Hierarchical Temporal-Aware Video-Language Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 95, + 430, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 95, + 430, + 125 + ], + "spans": [ + { + "bbox": [ + 162, + 95, + 430, + 125 + ], + "type": "text", + "content": "Qinghao Ye Guohai Xu Ming Yan* Haiyang Xu* Qi Qian Ji Zhang Fei Huang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 137, + 380, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 137, + 380, + 166 + ], + "spans": [ + { + "bbox": [ + 214, + 137, + 380, + 166 + ], + "type": "text", + "content": "DAMO Academy, Alibaba Group yeqinghao.yqh@alibaba-inc.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 191, + 192, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 191, + 192, + 204 + ], + "spans": [ + { + "bbox": [ + 143, + 191, + 192, + 204 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 217, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 290, + 540 + ], + "type": "text", + "content": "Video-language pre-training has advanced the performance of various downstream video-language tasks. However, most previous methods directly inherit or adapt typical image-language pre-training paradigms to video-language pre-training, thus not fully exploiting the unique characteristic of video, i.e., temporal. In this paper, we propose a Hierarchical Temporal-Aware video-language pretraining framework, HiTeA, with two novel pre-training tasks for yielding temporal-aware multi-modal representation with cross-modal fine-grained temporal moment information and temporal contextual relations between video-text multi-modal pairs. First, we propose a cross-modal moment exploration task to explore moments in videos by mining the paired texts, which results in detailed video moment representation. Then, based on the learned detailed moment representations, the inherent temporal contextual relations are captured by aligning video-text pairs as a whole in different time resolutions with multi-modal temporal relation exploration task. Furthermore, we introduce the shuffling test to evaluate the temporal reliance of datasets and video-language pre-training models. We achieve state-of-the-art results on 15 well-established video-language understanding and generation tasks, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with " + }, + { + "bbox": [ + 46, + 217, + 290, + 540 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 46, + 217, + 290, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 290, + 540 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 46, + 217, + 290, + 540 + ], + "type": "text", + "content": " improvement respectively. HiTeA also demonstrates strong generalization ability when directly transferred to downstream tasks in a zero-shot manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "content": "Vision and language are two primary signals that constitute the real-world perception of humanity. With the success of image-language pre-training [8,22,25,47], video-language pre-training [23,26,27,34] has recently received increasing attention. Large-scale video-language pre-training helps the model to learn effective multi-modal representation, which has shown significant improvement on a variety of video-language downstream tasks, such as videotext retrieval, video question answering and video caption-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 191, + 539, + 376 + ], + "blocks": [ + { + "bbox": [ + 313, + 191, + 539, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 191, + 539, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 539, + 376 + ], + "type": "image", + "image_path": "483f83779b42714bc182be16f6d694585fb415a0441b17fe29454fad8bd01314.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 388, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 305, + 388, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 547, + 473 + ], + "type": "text", + "content": "Figure 1: Comparison between existing paradigms and ours for video-language pre-training. (a) Previous methods align video and text within global perspective as the pretext. (b) We introduce HiTeA by varying video in different temporal views and modeling cross-modal temporal information between moments and texts, as well as the temporal contextual relations between multi-modal pairs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 495, + 421, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 421, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 421, + 507 + ], + "type": "text", + "content": "ing [5,33,44,48,50,53,59]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 700 + ], + "type": "text", + "content": "Inspired by the success of image-language pre-training paradigm, various methods [10, 11, 23, 26, 27] have been proposed to adapt it to video-language pre-training. ClipBERT [21] and Singularity [20] directly build on representations from image encoders and aggregate them via score aggregation function and temporal encoder. Furthermore, MILNCE [34] and Frozen [2] switch image encoder to video encoder for spatio-temporal video representation learning and align the video with corresponding text. In addition, some advanced pre-training tasks are designed through modeling entity [23], reconstructing masked patches [10] and predicting frame order [26, 61]. Despite their promising performance on downstream tasks, they treat video within global perspective illustrated in Figure 1(a), thus failing to consider fine-grained temporal information and temporal contextual relations which are essential to video-language pre-training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Since untrimmed video contains various temporal details," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15405" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "directly treating the video globally has two main limitations: (1) Less effective in modeling the fine-grained moment information including atomic actions and moments. As illustrated in Figure 1(b), we vary time resolutions and generate two views (long & short) for the input video. As a result, the short-view video clip tends to represent the moment information and the long-view video may express more event-level information. For example, the short-view video clip in Figure 1(b) only describes the moment of \"lick fingers\" rather than \"eating ice cream\". Such fine-grained moment information is hard to be captured by the long-view video under global event perspective; (2) Ignoring the temporal contextual relations implicitly existed in the video. Knowing the event expressed by the text, the moment \"eating ice cream\" can be inferred from the moment \"lick fingers\" shown by short-view video. However, such implicit temporal contextual relations between the moment and the event are rarely explored in previous works." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 686 + ], + "type": "text", + "content": "To address these problems, we propose a Hierarchical Temporal-Aware video-language pre-training framework, HiTeA, for both multi-modal understanding and generation. Except for the standard pre-training tasks, HiTeA introduces two novel temporal-aware video-language pre-training tasks, named cross-modal moment exploration (CME) and multimodal temporal relation exploration (MTRE), which not only model the fine-grained temporal moment information but also captures temporal contextual information hierarchically, yielding temporal-aware multi-modal representations for both understanding and generation. Specifically, we first generate the long-view and short-view videos with different time resolutions to build hierarchy of the input video. Then, based on the similarities of words and short-view video, we select the most relevant words as positive and leave the rest of the words as hard negatives. The CME pre-training task is applied to align the positive words and short-view video representations in the same embedding space. Moreover, to capture association between moments and the event for temporal contextual modeling, we match different views for the same video. However, directly matching two views visually would be noisy due to the background similarity [39]. To this end, we perform multi-modal pair alignment between videotext pairs via the MTRE pre-training task. More specifically, the short-view video guided by most relevant words and the long-view video guided by text will be aligned, which enables the model to extrapolate the contextual information from the short-view with language signal while enhancing temporal reasoning ability. Empowered by above two novel temporal-aware video-language pre-training tasks, HiTeA is capable of modeling temporal-aware multi-modal information revealed in video-text data including both fine-grained moment information and temporal contextual relations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In spite of a good performance, recent studies [4, 20] reveal most video-language downstream datasets are biased to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "wards still objects, scenes, etc., while the temporal dynamics are negligible. To evaluate the temporal performance of the video-language pre-training model and temporal reliance of downstream datasets, we introduce temporal shuffling test for these datasets. This enables a comprehensive evaluation of temporal modeling capability in the video-language pre-training field. Besides, our method achieves significant improvement on the datasets with heavy temporal reliance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 169, + 535, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 169, + 535, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 169, + 535, + 180 + ], + "type": "text", + "content": "In summary, our key contributions are the followings:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 189, + 545, + 387 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 189, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 545, + 224 + ], + "type": "text", + "content": "- We propose a novel hierarchical temporal-aware video-language pre-training framework with both video-language understanding and generation capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 228, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 545, + 286 + ], + "type": "text", + "content": "- We introduce temporal-aware pre-training tasks to generate temporal-aware multi-modal representation through modeling fine-grained temporal moment information as well as capturing the temporal contextual relations between moment and event." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 291, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 291, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 291, + 545, + 387 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness and generalization ability of HiTeA, and it achieves state-of-the-art performance on 15 video-language downstream datasets including video-text retrieval, video question answering, and video captioning, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with " + }, + { + "bbox": [ + 317, + 291, + 545, + 387 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 317, + 291, + 545, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 291, + 545, + 387 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 317, + 291, + 545, + 387 + ], + "type": "text", + "content": " improvement respectively." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Video-Language Pre-training Benefiting from a large number of image/video-text pairs, video-language pretraining (VLP) exhibits superior capabilities on various video-text benchmarks. The method of VLP is constantly evolving. Traditional approaches [26, 30, 42, 62] leverage offline-extracted dense video features for pre-training to circumvent the expensive computation overhead. In contrast, ClipBERT [21] suggests that sparse sampling can enable affordable end-to-end learning and improve performance simultaneously. Recent emerging approaches [2, 11, 14, 21, 23, 27, 52, 57] adopt this strategy and propose new model architectures and pre-training tasks. Frozen [2] trains jointly on image and video datasets via video-text contrastive learning (VTC). ALPRO [23] proposes a new visually-grounded pre-training task combined with VTC, video-text matching (VTM) and masked language modeling (MLM) [9] to learn fine-grained region-entity alignment. LAVENDER [27] formulates all pre-training and downstream tasks as MLM so that a unified architecture can be used for all video-text tasks. Apart from above representative works, frame order modeling (FOM) [26, 61] and masked video modeling (MVM) [10] are designed for VLP. However, the temporal characteristic of video still remains largely unexplored. To this end, we introduce a novel hierar" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15406" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 76, + 504, + 277 + ], + "blocks": [ + { + "bbox": [ + 90, + 76, + 504, + 277 + ], + "lines": [ + { + "bbox": [ + 90, + 76, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 90, + 76, + 504, + 277 + ], + "type": "image", + "image_path": "2ea68e067c43e3539126ddd5b67b4506b70cb716b5392357a635800d66589dc0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 288, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 547, + 350 + ], + "type": "text", + "content": "Figure 2: Illustration of the proposed HiTeA. We first generate two different temporal views for the input video, where the long-view is the video itself and the short-view is randomly truncated from the input video. To explore the moment revealed in the short-view, cross-modal moment exploration (CME) selects the candidate words from the input text with " + }, + { + "bbox": [ + 46, + 288, + 547, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CME}}" + }, + { + "bbox": [ + 46, + 288, + 547, + 350 + ], + "type": "text", + "content": ". Then, we perform multi-modal temporal relation exploration (MTRE) for modeling the temporal contextual relations between two video-text pairs with different views by " + }, + { + "bbox": [ + 46, + 288, + 547, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MTRE}}" + }, + { + "bbox": [ + 46, + 288, + 547, + 350 + ], + "type": "text", + "content": ". Note that the multi-modal encoders and the text features are shared." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "type": "text", + "content": "chical temporal-aware VLP framework which not only models the fine-grained moment information but also captures contextual relations with different temporal granularities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 715 + ], + "type": "text", + "content": "Temporal Modeling The temporal characteristic acts as a vital role in VLP since it provides the model with the capabilities of reasoning and understanding causality. Previous efforts in this field can be roughly divided into three categories. First, several methods directly transfer image-text models to video-text tasks by simply concatenating video frame [22,24] or building an extra temporal encoder [31,32]. Second, some works [2, 10, 23, 27] switch the image encoder to video encoder for learning spatio-temporal contexts within videos. Third, HERO [26] and MERLOT [61] design FOM task to explicitly recover the correct temporal order of shuffled frames. Nonetheless, ATP [4] and Singularity [20] reveal the existence of a static appearance bias in popular video-language datasets, and they develop single-frame models to achieve strong performance, comparable or even better than above methods with explicit temporal modeling. Therefore, they recommend SSv2 [20] and NExT-QA [48] datasets to test the temporal ability of VLP models. Different from previous approaches, we vary the temporal resolutions and generate two views of video so as to construct the temporal hierarchy, which equips the model with the ability to learn both fine-grained temporal moment information and temporal contextual representation at the same time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 369, + 361, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 361, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 361, + 380 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 391, + 373, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 373, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 373, + 403 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 472 + ], + "type": "text", + "content": "Figure 2 sketches the overview of the HiTeA. In concrete, our model consists of two unimodal encoders for encoding video and text separately, a multi-modal encoder for video and text interaction, and a text decoder for generation which is omitted here for simplicity and detailed in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": "For video representation, previous methods [21, 23, 27] encode the whole input video as a single-view feature, ignoring the rich temporal details contained in the video. Thus, we first treat the video into two views with different time resolutions to build hierarchy of the input video. Specially, the untrimmed video is regarded as a long-view " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "V^L" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " for capturing event information, and a video segment is randomly truncated from the input video as the short-view for capturing moment information denoted as " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "V^S" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": ". Then, we use the video encoder to encode an arbitrary view of video " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times H \\times W}" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " into a sequence of embeddings: " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_{\\mathrm{cls}}, v_1, \\dots, v_M\\} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " is the number of flattened patches, " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " is hidden size, and " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " is the embedding of the visual [CLS] token which provides global representation of the video. For text representation, we use the text encoder to transform the text " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " into a sequence of embeddings: " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{w_{\\mathrm{cls}}, w_1, \\dots, w_N\\} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " is length of the text. After that, the multi-modal encoder takes video features " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " and text features " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " as inputs and yields the multi-modal representation " + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 474, + 546, + 715 + ], + "type": "text", + "content": " for the video." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15407" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "In order to take full advantage of the different views of the video, we introduce cross-modal moment exploration (CEM) to explore the proper words or phrases from input text to align the short-view video with " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CME}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " for capturing the fine-grained temporal moment information in Section 3.2. Furthermore, to model the temporal contextual relations between the short-view video containing moment information and the long-view video with event information, we propose multi-modal temporal relation exploration (MTRE) to match the multi-modal representation of shortview and long-view videos by " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MTRE}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " in Section 3.3. Lastly, we introduce the overall pre-training objective for training the model in Section 3.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 233, + 233, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 233, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 233, + 247 + ], + "type": "text", + "content": "3.2. Cross-Modal Moment Exploration" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 287, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 359 + ], + "type": "text", + "content": "The video with short temporal range (i.e., short-view of video) with the paired text tends to be accompanied with fine-grained temporal moment information. However, the paired text partially describes the short-view of video bringing noise to the fine-grained moment representation thus degrading the performance. To this end, we propose a novel pre-training task named cross-modal moment exploration (CME), which enables the model to understand fine-grained moment information by leveraging the partially aligned text." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 418 + ], + "type": "text", + "content": "Formally, we first discover the possible positive words for the video in short-view by computing the cosine similarity of the word embedding sequence " + }, + { + "bbox": [ + 46, + 360, + 287, + 418 + ], + "type": "inline_equation", + "content": "\\{w_{1},\\dots ,w_{N}\\}" + }, + { + "bbox": [ + 46, + 360, + 287, + 418 + ], + "type": "text", + "content": " from text encoder and the short-view video representation " + }, + { + "bbox": [ + 46, + 360, + 287, + 418 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^S" + }, + { + "bbox": [ + 46, + 360, + 287, + 418 + ], + "type": "text", + "content": " from video encoder as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 424, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 424, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {K} = \\{\\pi (1), \\dots , \\pi (K) \\}, \\tag {1}", + "image_path": "e8ff42e6efb25c02c88ae7b683a5377be21a4015692fdf8827467d5685f9d252.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\pi : \\{1, \\dots, N\\} \\to \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": " is a permutation function for ranking such that " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "s(w_{\\pi(1)}, v_{\\mathrm{cls}}^S) \\geq \\dots \\geq s(w_{\\pi(N)}, v_{\\mathrm{cls}}^S)" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": " is the set of selected word indices, " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": " is the number of possible selected words, and " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "s(x, y) = x^T y / \\|x\\|_2 \\|y\\|_2" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": " represents the cosine similarity between " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": ". After obtaining the words for the video in short-view as the positive pair, the cross-modal moment exploration loss " + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CME}}" + }, + { + "bbox": [ + 46, + 442, + 287, + 550 + ], + "type": "text", + "content": " is computed with negative pairs from other words in the input text, which is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 555, + 299, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 299, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 299, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C M E}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left(\\frac {1}{| \\mathcal {K} |} \\sum_ {k \\in \\mathcal {K}} \\log \\frac {\\exp \\left(\\left(v _ {\\mathrm {c l s}} ^ {S}\\right) _ {i} ^ {\\top} w _ {i , k} / \\tau\\right)}{\\sum_ {n = 1} ^ {N} \\exp \\left(\\left(v _ {\\mathrm {c l s}} ^ {S}\\right) _ {i} ^ {\\top} w _ {i , n} / \\tau\\right)}\\right), \\tag {2}", + "image_path": "3fbfa12250bdedd029ac3632ec0cc1832a560d3261715574f7ee1ae966e0c0e9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": " is the learnable temperature hyper-parameter that controls the sharpness of the output distribution, and it is initialized as 0.07. As a consequence, the model is able to understand moment information via the proposed cross-modal exploration scheme." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 279, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 279, + 685 + ], + "type": "text", + "content": "3.3. Multi-Modal Temporal Relation Exploration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "While the video encoder has demonstrated its effectiveness in learning temporal representation implicitly" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "[10, 23, 56], it remains a challenge to discover the inherent temporal contextual relations. As a result, the limited capabilities in temporal modeling deteriorate the downstream task in temporal reasoning. This is in particular a missing point for the existing video-language pre-training paradigm [11, 21, 23, 27], which usually focuses on bridging video and text neglecting the function of text for guiding the video temporal contextual representation learning thus losing the temporal cues." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "text", + "content": "To this end, we introduce multi-modal temporal relation exploration (MTRE), a novel temporal-aware pre-training task that improves models' capacities in capturing temporal context in video with fine-grained text guidance by aligning multi-modal pairs. Specially, the short-view video " + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "inline_equation", + "content": "V^{S}" + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "text", + "content": " would represent moment information with respect to the whole video. On the contrary, the long-view video " + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "inline_equation", + "content": "V^{L}" + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "text", + "content": " expresses the event and topical information. To obtain the text-guided video features, we feed videos in different temporal views into the video encoder individually. Then, the text features are extracted and interact with the video features by the multi-modal encoder and yield text-guided video representations " + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_{\\mathrm{cls}}^{L} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_{\\mathrm{cls}}^{S} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 180, + 546, + 337 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 344, + 545, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 344, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 315, + 344, + 545, + 359 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {L} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {L}, v _ {1} ^ {L}, \\dots , v _ {M} ^ {L} \\right\\}, \\left\\{w _ {\\mathrm {c l s}}, w _ {1}, \\dots , w _ {N} \\right\\}\\right), \\tag {3}", + "image_path": "3e4068ae54d6382c15042ac0c0430de25190e6e60b70a7e6fc268e224ac3ab1e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 361, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 545, + 376 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {S} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \\dots , v _ {M} ^ {S} \\right\\}, \\left\\{w _ {\\mathrm {c l s}}, w _ {1}, \\dots , w _ {N} \\right\\}\\right), \\tag {4}", + "image_path": "eb9b2e0741d60d87238526588196efc58ef1b15b39f85cae6325507cbb6e2f27.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "inline_equation", + "content": "f(\\mathcal{V},\\mathcal{T})" + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "text", + "content": " represents the multi-modal encoder with video features " + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "text", + "content": " and text features " + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 384, + 545, + 503 + ], + "type": "text", + "content": ". However, since the short-view of the video is partially aligned with the text, using the whole text is not reasonable for generating accurate text-guided video feature for short-view. Meanwhile, improper video-text pairs would yield noisy multi-modal representation thus degrading the performance of the model. Therefore, thanks to the positive words mined by cross-modal moment exploration, we can calibrate representation for short-view video by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 511, + 545, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 511, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 320, + 511, + 545, + 528 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {\\mathrm {c l s}} ^ {S} = f \\left(\\left\\{v _ {\\mathrm {c l s}} ^ {S}, v _ {1} ^ {S}, \\dots , v _ {M} ^ {S} \\right\\}, \\left\\{w _ {\\mathcal {K} _ {1}}, \\dots , w _ {\\mathcal {K} _ {K}} \\right\\}\\right), \\tag {5}", + "image_path": "191dcea37249b21513eda3b5cd125016c7d37f64060d833e7ce84215b30b0416.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 536, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 536, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_i\\in \\mathcal{K}" + }, + { + "bbox": [ + 304, + 536, + 545, + 632 + ], + "type": "text", + "content": " is the index of the set for selected positive words. Then, we aim to match the representation of produced text-guided video features in different granularities in order to enable the model to predict the past and the future from the short-view of video, which benefits for capturing the general structure of the video. Specifically, we adopt the SimSiam framework [6] for minimizing their negative cosine similarity:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 640, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 640, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 359, + 640, + 545, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {D} \\left(p ^ {S}, z ^ {L}\\right) = - \\frac {p ^ {S}}{\\| p ^ {S} \\| _ {2}} \\cdot \\frac {z ^ {L}}{\\| z ^ {L} \\| _ {2}}, \\tag {6}", + "image_path": "9b65e527e72f6f7dac91985dfc3ac0d2974a7febf91a01a850f9c40634b9cee5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "inline_equation", + "content": "p^S = h(g(\\mathbf{v}_{\\mathrm{cls}}^S))" + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "inline_equation", + "content": "z^{L} = g(\\mathbf{v}_{\\mathrm{cls}}^{L})" + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "text", + "content": " are projection MLP head and prediction MLP head [7,13]. Minimizing " + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(p^S,z^L)" + }, + { + "bbox": [ + 304, + 676, + 545, + 714 + ], + "type": "text", + "content": " is equivalent for minimizing the mean" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15408" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 519, + 152 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 519, + 152 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 519, + 152 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 519, + 152 + ], + "type": "table", + "html": "
Method# PT DataMSRVTTDiDeMoLSMDCActivityNet Caption
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
ClipBERT [21]0.2M22.046.859.920.448.060.8---21.349.063.5
Frozen [2]5M31.059.570.531.059.872.415.030.839.8---
ALPRO [23]5M33.960.773.235.967.578.8------
BridgeFormer [11]5M37.664.875.137.062.273.917.935.444.5---
Singularity [20]5M36.865.975.547.475.284.0---43.070.681.3
LAVENDER [27]5M37.863.875.047.474.782.422.243.853.5---
", + "image_path": "12e5a8d044373be85c09e8000995e378578d1532b64159a7de0b4efbd7faa80e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 160, + 517, + 213 + ], + "blocks": [ + { + "bbox": [ + 78, + 152, + 178, + 160 + ], + "lines": [ + { + "bbox": [ + 78, + 152, + 178, + 160 + ], + "spans": [ + { + "bbox": [ + 78, + 152, + 178, + 160 + ], + "type": "text", + "content": "Models pre-trained on more data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 160, + 517, + 213 + ], + "lines": [ + { + "bbox": [ + 73, + 160, + 517, + 213 + ], + "spans": [ + { + "bbox": [ + 73, + 160, + 517, + 213 + ], + "type": "table", + "html": "
VIOLET [10]183M34.563.073.432.662.874.716.136.641.2---
All-in-one [46]138M37.968.177.132.761.473.5---22.453.767.7
Clip4Clip [31]400M42.171.981.443.470.280.621.641.849.840.572.4-
X-CLIP [32]400M46.173.083.145.274.0-23.343.0-44.374.1-
HiTeA5M44.469.378.951.879.185.327.146.254.545.173.584.2
", + "image_path": "4c12622ea141dc372243a42464bec9b5bebe055e8763264f4a98111087cd270e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 217, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 254 + ], + "type": "text", + "content": "Table 1: Performance comparison on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more pre-training data for fair comparison. # PT Data is the number of video-text pairs for pre-training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 266, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 304 + ], + "type": "text", + "content": "square error between " + }, + { + "bbox": [ + 46, + 266, + 288, + 304 + ], + "type": "inline_equation", + "content": "p^S" + }, + { + "bbox": [ + 46, + 266, + 288, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 288, + 304 + ], + "type": "inline_equation", + "content": "z^L" + }, + { + "bbox": [ + 46, + 266, + 288, + 304 + ], + "type": "text", + "content": ", which encourages the videos in different temporal magnitudes to be similar. Following [6, 13], we defined a symmetrized loss as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 311, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 287, + 334 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M T R E}} = \\frac {1}{2} \\left[ \\mathcal {D} \\left(p ^ {L}, \\mathbf {s g} \\left(z ^ {S}\\right)\\right) + \\mathcal {D} \\left(p ^ {S}, \\mathbf {s g} \\left(z ^ {L}\\right)\\right) \\right], \\tag {7}", + "image_path": "b0cac76431787a979ed924389b1da24d730436fc1e104214627d546b7c2395c9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\mathsf{sg}(\\cdot)" + }, + { + "bbox": [ + 46, + 342, + 287, + 367 + ], + "type": "text", + "content": " is the stop-gradient operation that prevents the model from collapse during training [6]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 373, + 180, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 180, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 180, + 386 + ], + "type": "text", + "content": "3.4. Pre-training Objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 392, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 547 + ], + "type": "text", + "content": "Apart from the two proposed temporal-aware pre-training tasks, we follow proven video-text pre-training approaches [2, 23, 27] to adopt the standard pre-training tasks including video-text contrastive (VTC), video-text matching (VTM), masked language modeling (MLM), and prefix language modeling (PrefixLM) described in the related work. Precisely, VTC and VTM align the video and text from the global perspective, while MLM and PrefixLM contribute to multi-modal understanding and generation capabilities of the model. Details of these objectives are described in the Appendix. We simply combine these as the base training objective " + }, + { + "bbox": [ + 46, + 392, + 288, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{base}" + }, + { + "bbox": [ + 46, + 392, + 288, + 547 + ], + "type": "text", + "content": " for our model. Therefore, the full pre-training objective is computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 557, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 557, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 107, + 557, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {b a s e}} + \\mathcal {L} _ {\\text {C M E}} + \\mathcal {L} _ {\\text {M T R E}}. \\tag {8}", + "image_path": "0a5627cb70d463d0d6738e37eef927d8eba0d85842b2fea2397a00f648871191.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 579, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 599, + 155, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 155, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 155, + 613 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Pre-training Datasets Following the recent work [2,11,20, 23,27], we pre-train our model on a webly-sourced video dataset WebVid-2M [2] with 2.5M video-text pairs and a image-text dataset Google Conceptual Captions (CC3M) [41] with 3M image-text pairs. Unlike previous methods, we do not pre-train our model on the large-scale video-text datasets like HowTo100M [34] with 136M video-text pairs and YT-Temporal-180M [61] due to the heavy computation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 266, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 352 + ], + "type": "text", + "content": "Downstream Datasets We evaluate our pre-trained model on 18 video-language benchmarks including video-text retrieval, video question answering, and video captioning tasks. Specifically, video question answering (VideoQA) can be categorized as Multiple-Choice (MC) and Open-Ended (OE) settings. The evaluation datasets are briefly summarized in below. Details can be found in the Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 360, + 545, + 489 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 306, + 360, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 545, + 396 + ], + "type": "text", + "content": "- Video-Text Retrieval: MSRVTT [53], DiDeMo [1], LSMDC [38], ActivityNet Caption [18], SSv2-Label [20], and SSv2-Template [20];" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 399, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 545, + 434 + ], + "type": "text", + "content": "- VideoQA (MC): TGIF-Action, TGIF-Transition [15], MSRVTT-MC [58], LSMDC-MC [44], and NExT-QA [48];" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 437, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 545, + 473 + ], + "type": "text", + "content": "- VideoQA (OE): TGIF-Frame [15], MSRVTT-QA, MSVD-QA [50], LSMDC-FiB [33] and ActivityNet-QA [59]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 476, + 522, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 522, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 522, + 489 + ], + "type": "text", + "content": "- Video Captioning: MSRVTT [53] and MSVD [5]." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "Implementation Details Our implementation of HiTeA is based on PyTorch [35]. In detail, we instantiate the video encoder with MViT-Base model [28] pretrained on ImageNet-21K [37]. The text encoder is initialized from first six layers of pre-trained BERT-Base [9], and the multi-modal encoder is initialized with last six layers of pre-trained BERT-Base. We pre-train HiTeA for 10 epochs, using a batch size of 16 on 8 NVIDIA A100 GPUs. We use AdamW [17] optimizer with a weight decay of 0.02 and betas (0.9, 0.98). The learning rate is first warmed up to 5e-5 in the first 1000 iterations, and decays following a cosine schedule. During pre-training stage, following with [10, 14, 23], we sparsely sample 4 frames for short and long view while preserving their order in-between and resize them to " + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": ". The duration of short view is restricted as the 1/8 of the whole video duration. " + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": " is empirically set to 5. The MLM mask ratio is set to " + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": ". Details of fine-tuning stage are described in Appendix." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15409" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 70, + 481, + 151 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 481, + 151 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 481, + 151 + ], + "type": "table", + "html": "
Method#PT DataTGIFMSRVTTLSMDCMSVDActivityNet
ActionTransitionFrameMCQAMCFiBQAQA
ClipBERT [21]0.2M82.887.860.388.237.4----
ALPRO [23]5M----42.1--46.3-
Singularity [20]5M---92.042.7---41.8
LAVENDER [27]5M96.699.172.296.644.286.056.955.4-
Clover [14]5M94.998.071.495.043.983.254.151.9-
", + "image_path": "6e61ce32854071d354d2d0356ea2ef693ce0d0eded1ff27a1d8107e5b5a30732.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 111, + 152, + 481, + 219 + ], + "blocks": [ + { + "bbox": [ + 111, + 152, + 481, + 219 + ], + "lines": [ + { + "bbox": [ + 111, + 152, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 111, + 152, + 481, + 219 + ], + "type": "table", + "html": "
Models pre-trained on more data
VIOLET [10]183M92.595.768.991.943.982.853.747.938.9
JustAsk [54]69M----41.5--46.338.9
MERLOT [61]180M94.096.269.590.943.181.752.9-41.4
All-in-one [46]283M95.594.766.392.346.884.4-48.3-
HiTeA5M96.898.872.597.245.485.854.655.645.1
", + "image_path": "34ca937dd55fbb6fa3095f9a11e5751859b3982d83e487498471b2b9f9868630.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 255, + 261, + 330 + ], + "blocks": [ + { + "bbox": [ + 46, + 220, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 246 + ], + "type": "text", + "content": "Table 2: Performance comparison on video question answering. Accuracy is reported for evaluation. We gray out methods that use significantly more pre-training data for fair comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 255, + 261, + 330 + ], + "lines": [ + { + "bbox": [ + 72, + 255, + 261, + 330 + ], + "spans": [ + { + "bbox": [ + 72, + 255, + 261, + 330 + ], + "type": "table", + "html": "
Method# PT DataMSRVTTMSVD
UniVL [30]180M49.9-
SwinBERT [29]-53.8120.6
MV-GPT [40]53M60.0-
CLIP4Caption [43]400M57.7-
LAVENDER [27]5M58.0142.9
HiTeA5M62.5145.1
", + "image_path": "7acf1abb5d26f8a77c1f1c5e7f7db875b78e3a167ce16c08752b901e55e13f43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 358, + 287, + 421 + ], + "blocks": [ + { + "bbox": [ + 47, + 334, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 357 + ], + "type": "text", + "content": "Table 3: Performance comparison on video captioning. CIDEr [45] is reported for evaluation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 358, + 287, + 421 + ], + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 421 + ], + "type": "table", + "html": "
MethodMSRVTT-Ret.LSMDC-Ret.MSRVTT-QAMSVD-QA
Frozen [2]31.0/59.5/70.515.0/30.8/39.8--
ALPRO [23]33.9/60.7/73.2-/-/-42.146.3
BridgeFormer [11]37.6/64.8/75.117.9/35.4/44.5--
Singularity [20]36.8/65.9/75.5-/-/-42.7-
TimeSformer (Lbase)37.6/62.0/72.319.2/38.7/48.041.549.0
HiTeA with TimeSformer39.7/65.0/75.121.8/40.7/49.943.752.4
", + "image_path": "17d1b058fa28724f1fe5b7373d8f7830de9620e3331ca20e6d19766c55c097ed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "content": "Table 4: Performance comparison of different SoTA methods with TimeSformer [3]. For text-to-video retrieval, Recall@1/5/10 are reported. For video question answering task, we report the Top-1 accuracy." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 192, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 192, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 192, + 489 + ], + "type": "text", + "content": "4.2. Comparison to Prior Arts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 495, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 533 + ], + "type": "text", + "content": "In this section, we compare HiTeA with numerous state-of-the-art video-language pre-training methods on several downstream datasets under fine-tuning setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 549, + 179, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 179, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 179, + 562 + ], + "type": "text", + "content": "4.2.1 Text-to-Video Retrieval" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Table 1 summarizes the results on MSRVTT [53], DiDeMo [1], LSMDC [38], and ActivityNet Caption [18] under fine-tuning settings. Our method outperforms all of the existing video-language pre-training model by a large margin under the same data scale. In particular, our method yields " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " lift in terms of R@1 on MSRVTT dataset while only exploiting 5M video-text pairs. Note that we also include the comparison with the recent works that utilize the powerful encoder from CLIP [36], our method still can be comparable with them even surpass them, which shows the validity of the proposed method. Besides, we can notice that our method achieves the best result among all of listed methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 294 + ], + "type": "text", + "content": "on LSMDC dataset, which proves that our model can leverage the various moments presented in fruitful movie clips with cross-modal moment exploration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 308, + 454, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 454, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 454, + 320 + ], + "type": "text", + "content": "4.2.2 Video Question Answering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": "Table 2 lists the results of HiTeA and current state-of-the-art approaches on nine VideoQA datasets. It can be noticed that our method achieves the best performance in most of VideoQA datasets even with less pre-training data. Specifically, it achieves absolute improvement " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": " on TGIF-FrameQA, " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": " on MSRVTT-MC, " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": " on MSRVTT-QA, " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": " on MSVD-QA, and " + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 326, + 547, + 434 + ], + "type": "text", + "content": " on ActivityNet-QA. We believe the moments learned by the cross-modal exploration are useful for finding the clue of answers in VideoQA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 449, + 414, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 414, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 414, + 461 + ], + "type": "text", + "content": "4.2.3 Video Captioning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "text", + "content": "Table 3 compares HiTeA with existing methods on video captioning datasets MSRVTT and MSVD. As shown in the table, although we use less pre-training data than compared approaches, HiTeA still obtains significant improvement compared to those large-scale pre-trained models. On MSRVTT Caption, our method surpasses SoTA method MV-GPT [40] by " + }, + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 304, + 467, + 545, + 598 + ], + "type": "text", + "content": " CIDEr. Note that MV-GPT is pretrained for multi-modal video captioning and it leverages the ASR transcripts from audio as the additional input. By contrast, our method only utilizes video as the input during generation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 606, + 378, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 606, + 378, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 606, + 378, + 618 + ], + "type": "text", + "content": "4.3. Discussion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 624, + 547, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 648 + ], + "type": "text", + "content": "In this section, we discuss the temporal characteristics of our model and the datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Generalization on Plain Backbone. Table 4 delivers the performance of SoTA methods with the plain visual backbone. We instantiate the video encoder with TimeSformer [3] pretrained on ImageNet-21K [37] following [2, 11, 23]. We can observe that HiTeA with TimeSformer significantly" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15410" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 70, + 506, + 141 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 506, + 141 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 506, + 141 + ], + "type": "table", + "html": "
MethodMSRVTT-RetrievalSSv2 Template-RetrievalNExT-QA (Hard)MSVD-QA
R@1R@5R@10AveRR@1R@5R@10AveRAcc@CAcc@TAcc.
\\( \\mathcal{L}_{\\text{base}} \\)40.068.077.161.780.5100.0100.093.544.046.452.7
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{MTRE}} \\)41.669.178.263.083.398.9100.094.146.346.454.8
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{CME}} \\)42.069.379.763.783.999.4100.094.446.348.354.3
\\( \\mathcal{L}_{\\text{base}} + \\mathcal{L}_{\\text{CME}} + \\mathcal{L}_{\\text{MTRE}} \\)44.469.378.964.285.6100.0100.095.247.848.655.6
", + "image_path": "5d7f91cf91f805f4b9ce96a440beef3d9a0dba63d89d97eee40f614bf46aa24d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 177, + 286, + 233 + ], + "blocks": [ + { + "bbox": [ + 46, + 142, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 46, + 142, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 545, + 167 + ], + "type": "text", + "content": "Table 5: Evaluation of the proposed methods on four downstream video-language tasks. For text-to-video retrieval, R@1, R@5, R@10, and the average are reported. For video question answering, we report the accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 177, + 286, + 233 + ], + "lines": [ + { + "bbox": [ + 48, + 177, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 286, + 233 + ], + "type": "table", + "html": "
Method# PT DataSSv2-LabelSSv2-Template
R@1R@5R@10R@1R@5R@10
Frozen [2]5M---52.994.899.4
Clip4Clip [31]400M43.171.480.777.096.698.3
Singularity [20]5M44.173.582.277.098.999.4
HiTeA5M55.281.489.185.6100.0100.0
", + "image_path": "7b29bd20e1e5da51083c8485cb7c48c4d62c7a95f6462e539af84ad39128c17c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 60, + 262, + 272, + 393 + ], + "blocks": [ + { + "bbox": [ + 47, + 234, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 234, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 286, + 258 + ], + "type": "text", + "content": "Table 6: Comparison of existing methods on Something-to-Something (SSv2) text-to-video retrieval." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 262, + 272, + 393 + ], + "lines": [ + { + "bbox": [ + 60, + 262, + 272, + 393 + ], + "spans": [ + { + "bbox": [ + 60, + 262, + 272, + 393 + ], + "type": "table", + "html": "
Method# PT DataAcc@CAcc@TAcc@DAcc.
Full Set
Human-87.688.690.488.4
HCRN [19]-45.949.353.748.2
HGA [16]-46.350.759.349.7
VGT [49]0.18M53.456.469.556.9
HGA* [16]400M46.852.159.350.4
ATP [4]400M51.350.266.854.3
HiTeA5M62.458.375.663.1
Hard Split
ATP [4]400M38.436.5//
HGA [16]-43.345.3//
HiTeA5M47.848.6//
", + "image_path": "9704bd9383311c2196d32947917f65715b249adf72a6662f7bd02071a8024c00.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 61, + 449, + 272, + 527 + ], + "blocks": [ + { + "bbox": [ + 46, + 397, + 286, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 397, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 286, + 445 + ], + "type": "text", + "content": "Table 7: Comparison of existing methods on NExT-QA [48]. We report accuracy on the Causal (C), Temporal (T), Descriptive (D) splits and overall accuracy on validation set. * stands for using CLIP as the initialization of visual encoder." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 449, + 272, + 527 + ], + "lines": [ + { + "bbox": [ + 61, + 449, + 272, + 527 + ], + "spans": [ + { + "bbox": [ + 61, + 449, + 272, + 527 + ], + "type": "table", + "html": "
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT [53]64.263.30.9
DiDeMo [1]72.170.21.9
LSMDC [38]42.641.70.9
ActivityNet Caption [18]67.666.80.8
SSv2 Template [20]95.272.422.8
SSv2 Label [20]76.773.53.2
", + "image_path": "e3b3571c0e93c9d1496b5768a88bea75792b1e1b563c62df77b7db48f125c2b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 531, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 286, + 651 + ], + "type": "text", + "content": "Table 8: Dependency on temporal information for text-to-video retrieval datasets with temporal shuffling test. The average recall of Recall@1, Recall@5, and Recall@10 are reported. We evaluate the performance drop when shuffling the input during inference. \"Original\" and \"Shuffled\" denote the original and shuffled input videos, respectively, and \"Gap\" is the difference between the Original and Shuffled metric. The larger \"Gap\" indicates the dataset relies on temporal information, and the model utilizes more temporal information to solve the task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "outperforms other SoTA methods (e.g., ALPRO, Bridge-Former, and Singularity) with the same pre-training data and the number of frames. In addition, it shows that our proposed temporal-aware pre-training tasks can boost the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 331, + 177, + 519, + 239 + ], + "blocks": [ + { + "bbox": [ + 331, + 177, + 519, + 239 + ], + "lines": [ + { + "bbox": [ + 331, + 177, + 519, + 239 + ], + "spans": [ + { + "bbox": [ + 331, + 177, + 519, + 239 + ], + "type": "table", + "html": "
DatasetOriginal ↑Shuffled ↓Gap ↑
MSRVTT-QA [50]45.445.20.2
MSVD-QA [50]55.655.50.1
TGIF-FrameQA [15]72.572.10.4
ActivityNet-QA [59]45.145.00.1
NExT-QA (Hard) [48]47.145.60.5
", + "image_path": "790f52b371c913cf6f78a8b08a971ba13fd353b958b613999903aee8d2570597.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 293 + ], + "type": "text", + "content": "Table 9: Dependency on temporal information for video question answering datasets by temporal shuffling test. We report the accuracy for each dataset. For NExT-QA dataset, we evaluate with the hard split of the validation set [4]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "content": "performance of TimeSformer " + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{base}})" + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "content": " average improvement on MSRVTT-Retrieval and " + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "content": " on MSVD-QA. More related discussions are demonstrated in the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": "Impact of Loss Terms. We investigate the contribution of individual loss terms and the results are shown in Table 5. It can be observed that the combining both " + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CME}}" + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MTRE}}" + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": " improves the performance of text-to-video retrieval and video question answering by at least " + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": " in Average Recall and Average accuracy respectively. In addition, we also find that the performance of " + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CME}}" + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": " surpasses that of " + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MTRE}}" + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": " on MSRVTT retrieval dataset that largely dominated by the appearance information. This can be explained that the cross-modal moment exploration loss not only select the positive verbs for the video from the text but also choose the acting object for alignment, which can boost the retrieval performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on Temporal-aware Tasks. Lei et al. [20] reveal that the previous four retrieval datasets are prone to being biased for appearance while rarely relying on temporal information, thus introducing Something-to-Something v2 (SSv2) Template and SSv2 Label retrieval datasets to test models' true temporal modeling capability. In particular, SSv2 Template retrieval task requires a deeper understanding of the moment and temporal relation since no objects information are presented. The performance on these datasets are summarized in Table 6. It can be observed that HiTeA achieves significant improvement with " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "+8.5\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " gains in terms of R@1 on these two temporal-oriented text-to-video retrieval datasets, which demonstrates the effectiveness of our proposed method through exploring fine-grained moment information and modeling temporal relation. In addition, we evaluate our model on NExT-QA [48] dataset that explicitly designed for temporal and causal understanding." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15411" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 83, + 290, + 171 + ], + "blocks": [ + { + "bbox": [ + 114, + 72, + 248, + 81 + ], + "lines": [ + { + "bbox": [ + 114, + 72, + 248, + 81 + ], + "spans": [ + { + "bbox": [ + 114, + 72, + 248, + 81 + ], + "type": "text", + "content": "\"Cartoon characters are talking to a pokemon\"" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 89, + 71, + 96 + ], + "lines": [ + { + "bbox": [ + 51, + 89, + 71, + 96 + ], + "spans": [ + { + "bbox": [ + 51, + 89, + 71, + 96 + ], + "type": "text", + "content": "Video" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 97, + 70, + 103 + ], + "lines": [ + { + "bbox": [ + 52, + 97, + 70, + 103 + ], + "spans": [ + { + "bbox": [ + 52, + 97, + 70, + 103 + ], + "type": "text", + "content": "Frames" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 125, + 72, + 131 + ], + "lines": [ + { + "bbox": [ + 51, + 125, + 72, + 131 + ], + "spans": [ + { + "bbox": [ + 51, + 125, + 72, + 131 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 154, + 68, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 68, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 68, + 160 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 83, + 290, + 171 + ], + "lines": [ + { + "bbox": [ + 77, + 83, + 290, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 83, + 290, + 171 + ], + "type": "image", + "image_path": "256a9cf39bd6f337d83a0cedc2418b23b2eb84a26c6ffd8b9e220895a453cbf0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 545, + 205 + ], + "type": "text", + "content": "Figure 3: Visualizations of learned cross-attention maps from multi-modal encoder. We present samples from MSRVTT [53] and SSv2 Template [20] retrieval dataset. HiTeA attends to the patches related to objects motion by tracking trajectory." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 329, + 83, + 542, + 110 + ], + "blocks": [ + { + "bbox": [ + 374, + 72, + 504, + 81 + ], + "lines": [ + { + "bbox": [ + 374, + 72, + 504, + 81 + ], + "spans": [ + { + "bbox": [ + 374, + 72, + 504, + 81 + ], + "type": "text", + "content": "\"Poking [something] so that it spins around\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 91, + 321, + 98 + ], + "lines": [ + { + "bbox": [ + 302, + 91, + 321, + 98 + ], + "spans": [ + { + "bbox": [ + 302, + 91, + 321, + 98 + ], + "type": "text", + "content": "Video" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 98, + 321, + 104 + ], + "lines": [ + { + "bbox": [ + 302, + 98, + 321, + 104 + ], + "spans": [ + { + "bbox": [ + 302, + 98, + 321, + 104 + ], + "type": "text", + "content": "Frames" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 83, + 542, + 110 + ], + "lines": [ + { + "bbox": [ + 329, + 83, + 542, + 110 + ], + "spans": [ + { + "bbox": [ + 329, + 83, + 542, + 110 + ], + "type": "image", + "image_path": "734fdad53243f68a2c60e2b68744a6772c584a614e769197cc9ff9281ca3d92c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 113, + 542, + 141 + ], + "blocks": [ + { + "bbox": [ + 302, + 125, + 323, + 130 + ], + "lines": [ + { + "bbox": [ + 302, + 125, + 323, + 130 + ], + "spans": [ + { + "bbox": [ + 302, + 125, + 323, + 130 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 113, + 542, + 141 + ], + "lines": [ + { + "bbox": [ + 329, + 113, + 542, + 141 + ], + "spans": [ + { + "bbox": [ + 329, + 113, + 542, + 141 + ], + "type": "image", + "image_path": "9e71569ad943b5e14491b9f8b713a210dc182d8412095f3836cdc978df6f21d3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 329, + 143, + 542, + 171 + ], + "blocks": [ + { + "bbox": [ + 329, + 143, + 542, + 171 + ], + "lines": [ + { + "bbox": [ + 329, + 143, + 542, + 171 + ], + "spans": [ + { + "bbox": [ + 329, + 143, + 542, + 171 + ], + "type": "image", + "image_path": "6f09a43a20b52cab24ae1cf7b6496a600bb8f5239fbca6ce0d2f16de8ca7f793.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 225, + 289, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 289, + 381 + ], + "type": "text", + "content": "As presented in Table 7, our method significantly surpasses its competitive counterparts, even those methods equipped with powerful image-text pre-trained encoders. Quantitatively, HiTeA obtains an absolute improvement " + }, + { + "bbox": [ + 46, + 225, + 289, + 381 + ], + "type": "inline_equation", + "content": "+9\\%" + }, + { + "bbox": [ + 46, + 225, + 289, + 381 + ], + "type": "text", + "content": " on the causality split with the help of intrinsic temporal relation. Recently, Buch et al. [4] filter out the trivial question for the dataset, and build the hard split for causality and temporal related questions for evaluate the causality and temporal of the model. As we can see in the table, even for the questions that heavily rely on causality, our model can still achieves a relative gain of " + }, + { + "bbox": [ + 46, + 225, + 289, + 381 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 46, + 225, + 289, + 381 + ], + "type": "text", + "content": " on the model with specific design for VideoQA, which indicates that our model do not solely depend on static appearance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": "Temporal Reliance of Datasets. Previous methods [11,21, 23] only evaluate the performance of models on the existing datasets to demonstrate the superiority of the methods. However, Buch et al. [4] and Lei et al. [20] reveal that the most of the evaluation are biased towards the static concepts. Here, we investigate the temporal reliance for the evaluated datasets by introducing the temporal shuffling test. Specifically, we compute the performance changes between running inference on the ordered video versus its shuffled version. The large performance drop indicates the dataset has less spatial bias and needs for temporal information. Table 8 and Table 9 conclude the performance gap between ordered and shuffled input video for text-to-video retrieval and VideoQA datasets. For text-to-video retrieval task, SSv2 Template shows the large performance drop after shuffling the input video, which demonstrates that it depends most on the dynamic information thus verifying our assumption. On the contrary, the performance on ActivityNet Caption dataset is barely affected (-0.8 on Mean Recall) since the text almost describes the static objects without relying on temporal information. For video question answering dataset, we observe that the MSVD-QA and ActivityNet-QA are less sensitive to the order of video frames. This is because these two datasets contain more questions requiring frame-region information, such as object categories, scenes, and species. We believe this can be used to evaluate the temporal reliance of the datasets as well as the utilization for temporal cue by" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 331, + 223, + 520, + 302 + ], + "blocks": [ + { + "bbox": [ + 331, + 223, + 520, + 302 + ], + "lines": [ + { + "bbox": [ + 331, + 223, + 520, + 302 + ], + "spans": [ + { + "bbox": [ + 331, + 223, + 520, + 302 + ], + "type": "table", + "html": "
Method# PT DataMSRVTT-QAMSVD-QA
Just Ask [54]69M2.97.5
LAVENDER [27]5M4.511.6
MERLOT Reserve [60]1B5.8-
FrozenBiLM [55]10M6.411.7
HiTeA5M8.618.2
BLIP [24]129M19.235.2
mPLUG [22]400M21.137.2
HiTeA5M21.737.4
", + "image_path": "85dc240c675e60c4cf346933c8b1027a96539570da38a5d0f3625563ab8a0119.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "type": "text", + "content": "Table 10: Zero-shot evaluation on video question answering. Accuracy is reported. We gray out those methods additionally supervised pre-training on VQA v2 [12] dataset." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 348, + 414, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 414, + 359 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 414, + 359 + ], + "type": "text", + "content": "models in the future work." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 365, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 520 + ], + "type": "text", + "content": "Qualitative Analysis. To verify that our model can capture the motion information with respected to the given text rather than inferring from the static signal, we present the query text in SSv2 Template dataset which has masked all of the object information, and also visualize the query in MSRVTT dataset. As we can see in the Figure 3, the attention map of atomic action \"talking\" mainly focuses on the mouse of the cartoon characters while the baseline largely focusing on the characters, which indicates that our method can understand the moment better when adopting the temporal-aware pretraining tasks. In another example, the word \"spins\" can reveal the trajectory of the object showing that our method is able to capture the temporal motion presented in the video." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 527, + 453, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 453, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 453, + 540 + ], + "type": "text", + "content": "4.4. Zero-shot Generalizability" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "To demonstrate the generalizability of proposed videotext pre-trained model, we perform zero-shot evaluation on video-language downstream tasks. We evaluate the zero-shot performance on VideoQA task in Table 10. Our method attains competitive zero-shot performance on MSRVTT-QA and MSVD-QA datasets even without help of audio signal supervision [60] or additional generated video question pairs [54]. In particular, less pre-training data (i.e. " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "5\\mathrm{M} < 69\\mathrm{M}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": ") are used while our method can still outperform other SoTA approaches. We also evaluate the zero-shot performance of models supervised on VQA v2 [12]. We can find that our method surpasses the powerful multi-modal SoTA methods (e.g. mPLUG [22]) with only 5M pre-training data showing the better generalization ability of HiTeA." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15412" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 98, + 70, + 492, + 138 + ], + "blocks": [ + { + "bbox": [ + 98, + 70, + 492, + 138 + ], + "lines": [ + { + "bbox": [ + 98, + 70, + 492, + 138 + ], + "spans": [ + { + "bbox": [ + 98, + 70, + 492, + 138 + ], + "type": "table", + "html": "
Method# PT DataMSRVTTDiDeMoLSMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
Frozen [2]5M18.739.551.621.146.056.29.322.030.1
ALPRO [23]5M24.144.755.423.847.357.9---
BridgeFormer [11]5M26.046.456.425.650.661.112.225.932.2
Singularity [20]5M28.450.259.536.961.669.3---
", + "image_path": "94bda957968573c897032601d98e1bbd5f1d9df4c59f0e315475196ece7c70fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 99, + 148, + 492, + 194 + ], + "blocks": [ + { + "bbox": [ + 103, + 139, + 212, + 147 + ], + "lines": [ + { + "bbox": [ + 103, + 139, + 212, + 147 + ], + "spans": [ + { + "bbox": [ + 103, + 139, + 212, + 147 + ], + "type": "text", + "content": "Models pre-trained on more data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 148, + 492, + 194 + ], + "lines": [ + { + "bbox": [ + 99, + 148, + 492, + 194 + ], + "spans": [ + { + "bbox": [ + 99, + 148, + 492, + 194 + ], + "type": "table", + "html": "
VideoCLIP [51]138M10.422.230.016.646.9----
VIOLET [10]183M25.949.559.723.549.859.8---
Clip4Clip [31]400M32.057.066.9---15.128.536.4
HiTeA5M29.954.262.936.160.170.315.531.139.8
", + "image_path": "0ac70c2a391ca661f40b0914e3574d25df9e2f5711323514aea0bbb5ad1d58aa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 198, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 224 + ], + "type": "text", + "content": "Table 11: Zero-shot evaluation on text-to-video retrieval. All results are reported on R@1/R@5/R@10. We gray out methods that use significantly more data for fair comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "type": "text", + "content": "We also perform zero-shot evaluation on text-to-video retrieval task. Table 11 summarizes the performance of our model and compared approaches on text-to-video retrieval. We can observe that our model yields more than " + }, + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "type": "text", + "content": " lift in R@1 on MSRVTT dataset [53] while exploiting fewer video-text pairs. Besides, our method surpasses all of the compared models in terms of LSMDC dataset showing the superiority of our method's generalizability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 343, + 119, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 119, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 119, + 356 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 363, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 289, + 604 + ], + "type": "text", + "content": "In this work, we introduce HiTeA, a novel hierarchical temporal-aware video-language pre-training framework with both understanding and generation capabilities. We vary the video with different views and model cross-modal temporal information between moments and texts as well as temporal contextual relations between multi-modal pairs in a hierarchical way. Specifically, a cross-modal moment exploration pre-training task is proposed to explore the fine-grained temporal information between the paired text and video moment by overcoming the partially semantic alignment. Moreover, multi-modal pairs are constructed to learn temporal contextual relations between moments and the event presented by the video with multi-modal temporal relation exploration pre-training task. We also demonstrate that our proposed pre-training tasks can consistently boost performance significantly on downstream tasks regardless of the backbone showing the generalization ability. Even pre-trained on less data, HiTeA still achieves state-of-the-art performance on a wide range of video-language downstream datasets, which clearly shows the superiority of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 615, + 105, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 105, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 105, + 627 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 635, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 52, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 52, + 635, + 287, + 689 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 52, + 691, + 287, + 713 + ], + "type": "text", + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 236, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 326, + 236, + 545, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 236, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 326, + 236, + 545, + 269 + ], + "type": "text", + "content": "end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 3, 5, 6, 7, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 270, + 547, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 270, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 311, + 270, + 547, + 303 + ], + "type": "text", + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 304, + 547, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 304, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 311, + 304, + 547, + 358 + ], + "type": "text", + "content": "[4] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 359, + 547, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 359, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 311, + 359, + 547, + 413 + ], + "type": "text", + "content": "[5] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 190-200, 2011. 1, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 311, + 415, + 545, + 458 + ], + "type": "text", + "content": "[6] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 459, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 459, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 311, + 459, + 545, + 502 + ], + "type": "text", + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 503, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 503, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 311, + 503, + 545, + 556 + ], + "type": "text", + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 558, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 558, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 311, + 558, + 545, + 601 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "type": "text", + "content": "[10] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. arXiv preprint arXiv:2111.12681, 2021. 1, 2, 3, 4, 5, 6, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[11] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16167-16176, 2022. 1, 2, 4, 5, 6, 8, 9" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "15413" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 195 + ], + "type": "text", + "content": "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "text", + "content": "[14] Jingjia Huang, Yinan Li, Jiashi Feng, Xiaoshuai Sun, and Rongrong Ji. Clover: Towards a unified video-language alignment and fusion model. arXiv preprint arXiv:2207.07885, 2022. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 296 + ], + "type": "text", + "content": "[15] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 342 + ], + "type": "text", + "content": "[16] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11109-11116, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 374 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "type": "text", + "content": "[18] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "text", + "content": "[19] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for multimodal video question answering. International Journal of Computer Vision, 129(11):3027-3050, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 499 + ], + "type": "text", + "content": "[20] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428, 2022. 1, 2, 3, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 566 + ], + "type": "text", + "content": "[21] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 622 + ], + "type": "text", + "content": "[22] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 1, 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "text", + "content": "[23] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 1, 2, 3, 4, 5, 6, 8, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for uni-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "fied vision-language understanding and generation. In ICML, 2022. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "text", + "content": "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in Neural Information Processing Systems, 34:9694-9705, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[26] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "type": "text", + "content": "[27] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. Lavender: Unifying videolanguage understanding as masked language modeling. arXiv preprint arXiv:2206.07160, 2022. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 304 + ], + "type": "text", + "content": "[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4804-4814, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 305, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 370 + ], + "type": "text", + "content": "[29] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 426 + ], + "type": "text", + "content": "[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "text", + "content": "[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 3, 5, 7, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "text", + "content": "[32] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 515, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 580 + ], + "type": "text", + "content": "[33] Tegan Maharaj, Nicolas Ballas, Anna Rohrbach, Aaron Courville, and Christopher Pal. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6884–6893, 2017. 1, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 582, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 646 + ], + "type": "text", + "content": "[34] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 1, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15414" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[37] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. arXiv preprint arXiv:2104.10972, 2021. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "type": "text", + "content": "[38] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3202-3212, 2015. 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "text", + "content": "[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Integrating language guidance into vision-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16177-16189, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 326 + ], + "type": "text", + "content": "[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17959-17968, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[41] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of ACL, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[42] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7464-7473, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[43] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[44] Atousa Torabi, Niket Tandon, and Leonid Sigal. Learning language-visual embedding for movie understanding with natural-language. arXiv preprint arXiv:1609.08124, 2016. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 559 + ], + "type": "text", + "content": "[45] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "text", + "content": "[46] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pre-training. arXiv preprint arXiv:2203.07303, 2022.5.6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 658 + ], + "type": "text", + "content": "[47] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[48] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9777–9786, 2021. 1, 3, 5, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. arXiv preprint arXiv:2207.05342, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "text", + "content": "[50] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 1, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 216 + ], + "type": "text", + "content": "[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 292 + ], + "type": "text", + "content": "[52] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video. In Proceedings of International Conference of Machine Learning (ICML), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "type": "text", + "content": "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 1, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[54] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "type": "text", + "content": "[55] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[56] Qinghao Ye, Xiyue Shen, Yuan Gao, Zirui Wang, Qi Bi, Ping Li, and Guang Yang. Temporal cue guided video highlight detection with low-rank audio-visual fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7950–7959, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "text", + "content": "[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[58] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 593, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 647 + ], + "type": "text", + "content": "[59] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9127–9134, 2019. 1, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[60] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15415" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[61] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[62] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8746-8755, 2020. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15416" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_content_list.json b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7340755f465a43aaeee7ef620b593a73031a0907 --- /dev/null +++ b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_content_list.json @@ -0,0 +1,1485 @@ +[ + { + "type": "text", + "text": "HiVLP: Hierarchical Interactive Video-Language Pre-Training", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Shao, Jianzhuang Liu, Renjing Pei, Songcen Xu, Peng Dai, Juwei Lu, Weimian Li, Youliang Yan, Huawei Noah's Ark Lab", + "bbox": [ + 241, + 181, + 725, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shaobin3, liu.jianzhuang, peirenjing, xusongcen, peng.dai, juwei.lu, liweimian, yanyouliang}@huawei.com", + "bbox": [ + 256, + 236, + 710, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-Language Pre-training (VLP) has become one of the most popular research topics in deep learning. However, compared to image-language pre-training, VLP has lagged far behind due to the lack of large amounts of video-text pairs. In this work, we train a VLP model with a hybrid of image-text and video-text pairs, which significantly outperforms pre-training with only the video-text pairs. Besides, existing methods usually model the cross-modal interaction using cross-attention between single-scale visual tokens and textual tokens. These visual features are either of low resolutions lacking fine-grained information, or of high resolutions without high-level semantics. To address the issue, we propose Hierarchical interactive Video-Language Pre-training (HiVLP) that efficiently uses a hierarchical visual feature group for multi-modal cross-attention during pre-training. In the hierarchical framework, low-resolution features are learned with focus on more global high-level semantic information, while high-resolution features carry fine-grained details. As a result, HiVLP has the ability to effectively learn both the global and fine-grained representations to achieve better alignment between video and text inputs. Furthermore, we design a hierarchical multi-scale vision contrastive loss for self-supervised learning to boost the interaction between them. Experimental results show that HiVLP establishes new state-of-the-art results in three downstream tasks, text-video retrieval, video-text retrieval, and video captioning.", + "bbox": [ + 75, + 335, + 473, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the framework of pre-training with large-scale uncurated data and then fine-tuning on some specific downstream tasks has attracted much attention. It firstly emerges in the field of Natural Language Processing (NLP), such as BERT [10], GPT [41] and T5 [42], which are pretrained on a large corpus of web-scraped dataset and then fine-tuned on a wide variety of NLP downstream tasks.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94bbadf5bf352a1c7f4750ededafb07582a2d1ee4e9c22cdc407d4113db61b41.jpg", + "image_caption": [ + "(a) Coarse Visual Feature Injecting" + ], + "image_footnote": [], + "bbox": [ + 501, + 301, + 679, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e16e61e12c61e0b4fdae632a6ed49e853f2da63d3d64235dc9efc3d678c1470.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 301, + 893, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7c1e467c132d8af97264895ab2d4dfd1bbc871f6f1099608efe4cebe40bc622.jpg", + "image_caption": [ + "(b) Fine-Grain Visual Feature Injecting", + "Figure 1. (a) Coarse visual feature injecting to CA blocks. (b) Fine-grained visual feature injecting to CA blocks. (c) Hierarchical visual feature injecting to CA blocks. A CA block consists of a casual/bi-directional self-attention layer, a cross-attention layer, and a feed forward layer. It is with bi-directional self-attention for vision-language understanding and with casual self-attention layer for vision-language generation." + ], + "image_footnote": [], + "bbox": [ + 506, + 419, + 864, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hereafter, it is transferred rapidly to the computer vision area. For examples, CLIP [40], ALIGN [16], Florence [58] and BLIP [20] all use more than 100 million open-domain image-text pairs in Image-Language Pre-training (ILP). However, most of the existing Video-Language Pretraining (VLP) works [47, 22, 51, 64, 30] use either a small-scale dataset (e.g., YouCookII [63] with 14K video-text pairs) or a large-scale dataset with less diversity (e.g., Howto100M [34] sourced from 1.22M videos). To solve this problem, we use a larger-scale dataset with 114M image-text pairs and a dataset with 2.5M video-text pairs to pre-train our model. We show that diversity is more important than the total amount of training pairs, and a small set of image-text pairs can achieve much better performance than using millions of video-text pairs. We believe this is a significant way to enhance VLP models and alleviate the", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13756", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cost of the collection of video-text pairs.", + "bbox": [ + 76, + 90, + 346, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In VLP and ILP, existing works [47, 22, 64, 21, 56] often use cross-attention to model the cross-modal interaction between visual features and text features. However, they usually adopt only the single-scale and low-resolution visual features (i.e., $\\frac{1}{16}$ scale of the input) for cross-attention (CA) blocks, as shown in Figure 1(a). This scheme fails to obtain fine-grained interaction with text features and limits the performance of the pre-training model. For finer-grained interactions, [33] injects the high-resolution visual features (i.e., $\\frac{1}{4}$ scale of the input) to CA blocks as shown in Figure 1(b), but it does not have high-level semantics. To overcome these limitations, we propose Hierarchical Interactive Video-Language Pre-training (HiVLP) that efficiently uses a Hierarchical Visual Feature Group (HVFG) for multi-modal cross-attention. As shown in Figure 1(c), HVFG includes different scales of visual features, where the low-resolution ones with high-level semantics are beneficial for global representation and the high-resolution ones with detailed information are useful for fine-grained interaction. Especially, HVFG is able to achieve much better accuracy because of such a multi-scale.", + "bbox": [ + 75, + 107, + 468, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many works [47, 64, 51] use self-supervised learning to assist the video-language pre-training by reconstructing the masked frame tokens. However, it may introduce noise to interactions between visual and textual features for the masked frame tokens [32]. In this paper, we propose a Multi-level Vision Contrastive (MVC) loss for our HiVLP by applying a global-to-local contrast learning to every scale in HVFG. The MVC loss does not damage the visual tokens and helps the multi-level alignment between visual and textual features.", + "bbox": [ + 75, + 426, + 468, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 578, + 419, + 592 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, our HiVLP is the first work that uses a hierarchical interaction for video-language pre-training. It is able to effectively learn both the global and fine-grained representations for better alignment between visual and textual features.", + "- We design a multi-level vision contrastive (MVC) loss for self-supervised learning that can sufficiently mine multi-level visual information to help video-language pre-training.", + "- We reveal that diversity is more important than the amount of training pairs, and using more diverse image-text pairs benefits a lot for VLP.", + "- Our HiVLP unifies video-language understanding and generation. It achieves state-of-the-art results in text-video retrieval, video-text retrieval, and video captioning." + ], + "bbox": [ + 94, + 611, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-Language Pre-Training (ILP). CLIP [40] is the pioneering work that collects large-scale web data (400M image-text pairs) and achieves competitive zero-shot performance on a variety of downstream tasks [39, 14]. ALIGN [16] is pre-trained with a larger-scale dataset (1.8B) obtaining better performance. FILIP [57] is pre-trained with 300M image-text pairs and designs a cross-modal late interaction mechanism for fine-grained contrastive learning. The key behind their success is that they take the advantages of large-scale datasets which are currently not available in VLP. To deal with this problem, we train our HiVLP model with a hybrid of video-text and image-text pairs.", + "bbox": [ + 496, + 118, + 890, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-Language Pre-Training (VLP). Existing VLP works either use a pre-trained S3D [50] to extract visual features as vision input [30, 52, 51] to speed up the training process, or firstly perform pre-training on video-text pairs and then transfer the model to video-language generation tasks [47, 64, 22]. However, both these two training approaches limit the model performance because they are not trained end-to-end [24]. Our HiVLP jointly trains the model with image-text and video-text pairs end-to-end, and unifies both video-language understanding and generation in one framework. FiT [4] also involves pre-training with both image-text and video-text pairs, but can not do video-language generation.", + "bbox": [ + 496, + 303, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Learning (SSL). To effectively use datasets, many works [47, 64, 51, 49, 55] use self-supervised learning to assist VLP. VideoBERT [47] tokenizes video frames by hierarchical vector quantization, and then performs SSL by predicting the masked visual tokens. ActBERT [64] predicts the action and object words of masked video tokens. VLM [51] masks either all video tokens or all text tokens, and then uses tokens from one modality to recover masked tokens. However, the existing methods may damage the visual tokens, which introduces noise into cross-attention between visual and textual tokens [32]. For better visual representation learning and avoiding damaging visual tokens, we introduce the MVC loss which maximizes the mutual information between multiscale global and local representations, and improves the multi-level alignment between visual and textual features.", + "bbox": [ + 496, + 503, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hierarchical Interaction. As far as we know, there is no related work about hierarchical interaction in VLP. The most related work is VinVL [59] in ILP. VinVL [59] uses an object detector to extract different sizes of objects as visual features to do cross-attention with textual features. The method of VinVL is complicated and the accuracy of its visual features is limited by the object detector. However, HiVLP makes different scales of visual features interact with textual features without the need of an object detector, which is more effective.", + "bbox": [ + 496, + 750, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e229b6afe7cd86bdc79de7941233721cd18b1ef93088ee505a3a320e80ef2360.jpg", + "image_caption": [ + "Figure 2. (a) Overview of HiVLP which consists of a vision encoder, a text encoder, a multi-modal video-text encoder and a multi-modal video-text decoder. Each image (or video clip) is transformed into two augmentations as the inputs to the vision encoder and the momentum vision encoder, respectively. The hierarchical visual feature group (HVFG) has scales $\\left\\{\\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}$ . The multi-level visual contrastive (MVC) loss is for multi-level SSL. Besides, HiVLP also uses the visual features of $\\frac{1}{32}$ scale from the vision encoder and the output of the text encoder to optimize the vision-text contrastive (VTC) loss. Additionally, the multi-modal video-text encoder is trained with a vision-text matching (VTM) loss for the matching of video-/image-text pairs. The multi-modal video-text decoder is trained with a language modeling (LM) loss for video captioning. (b) The light FPN [25] is used for gradually upsampling and fusing the output features of the Swin Transformer. (c) The temporal block consists of a temporal self-attention layer, a SW-MSA layer [29] and a feed forward layer. We only replace the last block of the Swin Transformer with the temporal block for less parameters." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 586, + 468 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/69b14ea2ca6e4e54ff382d171feef75dbe8defcf9214da9d9d49a0ff31f2388f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 90, + 880, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08f4a6bd3543f38c924c253a3858bfd6a7315661fcc64e7bb0f16cc563431223.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 325, + 880, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 611, + 187, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We firstly describe the detailed architecture of HiVLP. Then we present the hierarchical interaction HVFG. Finally, the pre-training objectives are given.", + "bbox": [ + 75, + 637, + 468, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 76, + 695, + 210, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2(a), HiVLP consists of a vision encoder, a momentum vision encoder, a text encoder, a multi-modal video-text encoder, and a multi-modal videotext decoder. The vision encoder and the text encoder extract image and text features respectively, and do not need to interact with each other for a fast approximate nearest neighbor search in inference [33]. The multi-modal videotext encoder re-ranks the top-k video-text pairs by an additional MLP head that predicts whether they are matched or not. The multi-modal video-text decoder is used to perform the video-language generation (e.g., video captioning). The momentum vision encoder has the same archi", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tecture as the vision encoder, and the weights are updated by a momentum-based moving average strategy as in MoCo [15].", + "bbox": [ + 496, + 612, + 890, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main component of the vision encoder is a hierarchical backbone Swin Transformer [29], followed by a light feature pyramid network (FPN) [25]. Besides, we add a temporal self-attention layer in the last block of Swin Transformer to capture the temporal information, as shown in Figure 2(c). In Figure 2(b), the light FPN uses the output features of $\\frac{1}{32}$ , $\\frac{1}{16}$ , and $\\frac{1}{8}$ scales from Swin Transformer as inputs, and then gradually upsamples those features for feature fusion. Note that to reduce computation overhead, we do not use $\\frac{1}{4}$ scale of visual features in HVFG.", + "bbox": [ + 496, + 657, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The text encoder's architecture is the same as BERT [10], which consists of 12 transformer layers. The multi-modal video-text encoder has 12 cross-attention (CA) blocks each with a bi-directional self-attention layer. The multi-modal video-text decoder consists of 12 CA blocks each with a casual self-attention layer.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13758", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a4c078d814f3a0ec3e3c38a2023b63f09a6c0f4cfbb18923540c42a9668571b6.jpg", + "image_caption": [ + "Figure 3. (a) Multi-level global-to-local contrastive loss. (b) Generation of global and local features." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 524, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f67c0f87f724134ec31964016c157eeb897bd9f7c8760fd73335ee0d94686d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 89, + 823, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hierarchical Interaction", + "text_level": 1, + "bbox": [ + 76, + 272, + 302, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing works often inject single-scale visual features to CA blocks to interact with textual features. For example, ALBEF [21] and BLIP [20] use visual features of $\\frac{1}{16}$ scale, which fail to do fine-grained cross-attention. In contrast, [33] adopts features of $\\frac{1}{4}$ scale, but these over detailed features without high-level semantics. In our work, we propose a hierarchical interaction mechanism between visual and textual features via injecting HVFG into the CA blocks as shown in Figure 2(a). HVFG includes visual features of three scales $\\left\\{\\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}$ . The visual features of $\\frac{1}{8}$ scale contain fine-grained information, and those of $\\frac{1}{32}$ scale carry high-level semantics. Therefore, HVFG is able to obtain both global and fine-grained representations to boost the interaction between visual and textual features.", + "bbox": [ + 75, + 296, + 473, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pre-Training Objectives", + "text_level": 1, + "bbox": [ + 76, + 517, + 299, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input of the vision encoder is an image or video clip $X \\in R^{B \\times M \\times 3 \\times H \\times W}$ , which consists of $M$ frames with the resolution $H \\times W$ , where $M = 1$ for images and $B$ is the batch size. The input to the text encoder is a batch of tokenized sequences of words $T \\in N^{B \\times L}$ , where $L$ is the max length of a caption.", + "bbox": [ + 76, + 541, + 468, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our method, an input video clip $X$ is split into non-overlapping patches $\\{x_{i}\\}_{i = 1}^{BMN}, x_{i} \\in R^{3 \\times P \\times P}$ , where $P \\times P$ is input path size and $N = \\frac{H \\times W}{P \\times P}$ . The patches are tokenized by a linear embedding layer which is followed by a linear embedding layer of Swin Transformer.", + "bbox": [ + 76, + 632, + 468, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Text Contrastive (VTC) Loss. VTC aims to pull positive pairs of vision and language representations together and push negative pairs far away. Like ALBEF [21] using a momentum distillation for vision-language pretraining, we introduce two queues to store the most recent $K$ vision-text representations pairs from momentum encoders. Formally, the video-to-text contrastive loss is defined as:", + "bbox": [ + 75, + 708, + 470, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T 2 V} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {- e x p \\left(s \\left(f _ {i} , t _ {i} ^ {+}\\right) / \\rho\\right)}{e x p \\left(s \\left(f _ {i} , t _ {i} ^ {+}\\right) / \\rho\\right) + \\sum_ {j = 1} ^ {K} e x p \\left(s \\left(f _ {i} , t _ {j} ^ {-}\\right) / \\rho\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 849, + 473, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\rho$ is the temperature, $t_i^+$ and $t_j^-$ are the text features of the positive and negative text samples for the $i$ -th vision input $X_i$ respectively, and $f_i$ is the visual feature of $\\frac{1}{32}$ scale from $X_i$ . $s(f_i, t_i^+)$ denotes the cosine similarity between $f_i$ and $t_i^+$ that are matched, and $s(f_i, t_j^-)$ denotes the similarity between $f_i$ and $t_j^-$ that are not matched. Symmetrically, the text-to-video contrastive loss is:", + "bbox": [ + 496, + 272, + 893, + 381 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V 2 T} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {- \\exp \\left(s \\left(t _ {i} , f _ {i} ^ {+}\\right) / \\rho\\right)}{\\exp \\left(s \\left(t _ {i} , f _ {i} ^ {+}\\right) / \\rho\\right) + \\sum_ {j = 1} ^ {K} \\exp \\left(s \\left(t _ {i} , f _ {j} ^ {-}\\right) / \\rho\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 405, + 898, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{i}^{+}$ and $f_{j}^{-}$ are positive and negative visual features of the $i$ -th text feature, respectively.", + "bbox": [ + 496, + 454, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sum of $L_{V2T}$ and $L_{T2V}$ serve as the vision-text contrastive loss:", + "bbox": [ + 498, + 484, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V T C} = \\frac {1}{2} \\left(L _ {V 2 T} + L _ {T 2 V}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 526, + 890, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-level Vision Contrastive (MVC) Loss. As shown in Figure 3(a), MVC aims to pull the global representation of a scale closer to those local patch representations from the same vision input in different views of a scale. The light FPN has three scales $L_{1}$ , $L_{2}$ , and $L_{3}$ (see Figure 2(b) and Figure 3(a)). The global feature $f_{ij}$ of scale $j$ of the $i$ -th visual input is obtained by the average of all the patch features at $L_{j}$ (Figure 3(b)). The local features $f_{ijk}, k = 1, 2, 3, \\dots, s_{j}$ are generated as shown in Figure 3(b) where $f_{ijk}$ is the average feature pooling result of a local window such as $2 \\times 2$ , and $s_{j}$ is the number of the local averaged features at $L_{j}'$ . Finally, the MVC loss is formulated as:", + "bbox": [ + 496, + 556, + 893, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M V C} ^ {i j} = \\frac {\\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j} ^ {+}\\right) / \\rho\\right)}{\\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j} ^ {+}\\right) / \\rho\\right) + \\sum_ {k = 1} ^ {(B - 1) S _ {j}} \\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j k} ^ {-}\\right) / \\rho\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 761, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M V C} = - \\frac {1}{3 B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {3} - l o g L _ {M V C} ^ {i j}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 823, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{f}_{ij}$ and $\\hat{f}_{ijk}$ correspond to $f_{ij}$ and $f_{ijk}$ , respectively, but from the other augmentation of the visual input.", + "bbox": [ + 498, + 868, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13759", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-Text Matching (VTM) Loss. Through a MLP layer, VTM is used to judge whether a video-/image-text pair is matched or not with the cross entropy function. There are $B$ positive pairs from the batch, and $2B$ negative pairs are obtained according to the hard negative mining strategy in [21]. This loss is defined as:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V T M} = - \\frac {1}{3 B} \\sum_ {i = 1} ^ {3 B} \\left[ y _ {i} \\log \\left(p _ {i}\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - p _ {i}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 207, + 468, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{i}$ is the ground-truth, and $p_i$ is the output of the MLP layer of the $i$ -th pair.", + "bbox": [ + 75, + 261, + 468, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Language Modeling (LM) Loss. We inject the hierarchical visual representations into the multi-modal video-text decoder to generate captions. Let $i$ -th input text be $T_{i} = [t_{i}^{1}, t_{i}^{2}, \\dots, t_{i}^{L}]$ . This loss is defined as:", + "bbox": [ + 75, + 292, + 468, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L M} = - \\frac {1}{B L} \\sum_ {i = 1} ^ {B} \\sum_ {l = 1} ^ {L} \\log \\left(p \\left(t _ {i} ^ {l} \\mid t _ {i} ^ {l - 1}, \\dots , t _ {i} ^ {1}, f _ {i 1}, f _ {i 2}, f _ {i 3}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 378, + 467, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(t_i^l | t_i^{l-1}, \\dots, t_i^1, f_{i1}, f_{i2}, f_{i3})$ is the output probability of the multi-modal video-text decoder for the $l$ -th token $t_i^l$ given the previously fed tokens $t_i^{l-1}, \\dots, t_i^1$ and the multilevel visual features $f_{i1}, f_{i2}, f_{i3}$ .", + "bbox": [ + 75, + 431, + 467, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total Loss. Finally, we have the total loss to train HiVLP:", + "bbox": [ + 76, + 494, + 460, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = \\alpha L _ {V T C} + \\beta L _ {M V C} + \\gamma L _ {V T M} + \\delta L _ {L M}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 537, + 468, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ , $\\beta$ , $\\gamma$ , and $\\delta$ are the weight coefficients of the losses.", + "bbox": [ + 75, + 566, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 613, + 209, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the pre-training and downstream datasets, and the implementation details. Then we show the experimental and comparison results, and finally ablate our model.", + "bbox": [ + 75, + 638, + 468, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 710, + 179, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-Training Datasets. Following FiT [4], we involve two popular datasets (CC-3M [46] and WebVid-2M [4]) for pre-training HiVLP. CC-3M consists of 3M image-text pairs and WebVid-2M includes 2.5M video-text pairs, in total 5.5M image-/video-text pairs. Besides, we use LAION100M, D-14M, and WebVid-2M (total 116.5M image-/video-text pairs) as the larger-scale dataset to train our model, resulting in HiVLP*. LAION100M is a subset of LAION [45]. D-14M is a combined dataset of image-text pairs from COCO [26], Visual Genome [17], CC3M [46], CC12M [6], and SBU captions [35].", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downstream Task Datasets. There are four downstream datasets used in our work for video retrieval. (i) MSR-VTT [53] consists of 10K YouTube videos with 200K human-annotated descriptions. Similar to previous methods [4], we use the split of 9K for training and 1K for testing. (ii) MSVD [7] is a smaller dataset with 1970 videos and 78800 sentences, about 40 descriptions per video. Like [4], we use the split of 1200, 100, and 670 videos for training, validation and testing, respectively. (iii) DiDeMo [3] has 10k videos, each of which is described by multiple sentences, resulting in 40K sentences. For a fair comparison, we follow the setting in [4], where all descriptions for one video are concatenated into a single sentence. (iv) LSMDC [43] contains 128K clips. Like [4], we use 7408 clips for validation and 1000 clips for testing. For video captioning, we evaluate our model on MSR-VTT and MSVD. Following the split of [30], on MSR-VTT, we use $6.5K$ training videos and $2.9K$ testing videos; on MSVD, we use $1.2K$ training videos and 670 testing videos.", + "bbox": [ + 496, + 90, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 391, + 718, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model HiVLP is pre-trained on 16 GPUs (32G memory). Our vision encoder is initialized by Swin-B [29] pretrained on ImageNet-21k [9]. The light FPN is randomly initialized. The text encoder, multi-modal video-text encoder, and multi-modal video-text decoder are BERTs initialized from BERTbase [10]. HiVLP is pre-trained for 25 epochs, the first 20 epochs only with image-text pairs with a batch size of 30 and the last 5 epochs with both image-text and video-text pairs. Each video clip consists of 4 frames. We use AdamW optimizer with a weight decay of 0.02, and the learning rate is initialized as $10^{-5}$ and is warmed up to $10^{-4}$ after 3,000 training iterations. We then decrease the learning rate by the cosine decay strategy to $10^{-5}$ . For the hyper-parameters in Equation 8, we set $\\alpha$ , $\\beta$ , and $\\delta$ to 1.0 and $\\gamma$ to 0.001. The input resolution is $224 \\times 224$ with data augmentations used in [21]. And we use 160 GPUs (16G memory) to train HiVLP* with a batch size of 18 per GPU. For speeding up the training of HiVLP*, we sample 1 frame from each clip.", + "bbox": [ + 496, + 415, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the downstream task of text-video retrieval, we sample 4 frames per video for training. Since the multi-modal video-text encoder filters top- $k$ candidates during inference ( $k$ is set to 128). We sample 8 frames per video for testing on MSR-VTT, MSVD, and LSMDC. Because the videos in DiDeMo are longer, we sample 10 frames for testing from it. For the downstream task of video captioning, we evaluate our model on MSR-VTT and MSVD, with 8 random frames per video for training and 16 frames per video for testing. For all downstream tasks, the initial learning rate is set to $5 \\times 10^{-6}$ , and the weight decay, the batch size, and the total number of epochs are set to 0.05, 64, and 10, respectively.", + "bbox": [ + 496, + 704, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/99022c17bc701f250f17b0bfa858711b1ce154a7e6b7c63dbd749786e853b1ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-Training Datasets#PairsR@1↑R@5↑R@10↑MedR↓MR↑
Zero-Shot
SupportSet [38]HowTo100M100M8.723.031.131.020.9
HD-VILA [54]HD-VILA-100M100M14.431.641.617.529.2
FiT [4]CC3M, WebVid-2M5.5M18.739.551.610.036.6
ALPRO [19]CC3M, WebVid-2M5.5M24.144.755.48.041.4
CLIP [40]WIT400M400M30.654.464.34.049.8
Florence [58]FLD-900M900M37.663.872.6-58
BLIP [20]L-115M, D-14M239M43.365.674.72.061.2
HiVLPCC3M, WebVid-2M5.5M26.447.355.77.043.1
HiVLP*L-100M, D-14M, WebVid-2M116.5M43.566.476.42.062.1
Fine-Tuning
ActBERT [64]HowTo100M100M16.342.856.910.038.7
HERO [22]HowTo100M100M16.843.457.7-39.3
NoiseEstimation [2]HowTo100M100M17.441.653.68.037.5
UniVL [30]HowTo100M100M21.249.663.16.044.6
ClipBERT [18]HowTo100M100M22.046.859.96.042.9
AVLnet [44]HowTo100M100M27.155.666.64.049.8
MMT [12]HowTo100M100M26.657.169.64.051.1
SupportSet [38]HowTo100M100M30.158.569.33.052.6
FiT [4]CC3M, WebVid-2M5.5M31.059.570.53.053.7
ALPRO [19]CC3M, WebVid-2M5.5M33.960.773.23.055.9
HD-VILA [54]HD-VILA-100M100M35.065.277.23.059.1
CLIP4Clip [31]WIT400M400M44.571.481.62.065.8
HiVLPCC3M, WebVid-2M5.5M41.165.975.92.061.0
HiVLP*L-100M, D-14M, WebVid-2M116.5M50.976.483.61.070.3
", + "bbox": [ + 133, + 87, + 836, + 458 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c91939e5898ad2903f9c5a6aa8a1b856f256736e5fb14c3bb98759e305457ad2.jpg", + "table_caption": [ + "Table 1. Comparisons with SOTA text-to-video retrieval methods on MSR-VTT. R@K: Recall@K; MedR: Median Rank; MR: Mean Rank; L-115M: LAION115M; L-100M: LAION100M." + ], + "table_footnote": [], + "table_body": "
MethodR@1↑R@5↑R@10↑MR↑
CE [28]19.849.063.844.2
Support Set [38]28.460.072.953.8
FiT [4]33.764.776.357.3
CLIP4Clip [31]46.276.184.668.9
HiVLP39.168.177.261.5
HiVLP*50.278.985.871.6
", + "bbox": [ + 98, + 515, + 450, + 619 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/130209c0f7dd390c1eb665ea6abefd0d19daad8b0d74446663c1d5601fce57a9.jpg", + "table_caption": [ + "Table 2. Text-to-video results on the MSVD dataset." + ], + "table_footnote": [], + "table_body": "
MethodR@1↑R@5↑R@10↑MR↑
ClipBERT [18]20.444.556.740.5
FiT [4]34.665.074.758.1
ALPRO [54]35.967.578.860.7
CLIP4Clip [31]43.470.280.664.7
HiVLP40.568.777.162.1
HiVLP*48.173.882.568.1
", + "bbox": [ + 99, + 656, + 446, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 796, + 452, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare HiVLP with state-of-the-art (SOTA) methods on three popular video-language downstream tasks (text-video retrieval, video-text retrieval, and video captioning).", + "bbox": [ + 76, + 821, + 468, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-Video Retrieval. We use MSR-VTT, MSVD,", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3959813fee75483c7bd345b8d0eeb3be58cd27f8bea84f259f419b566eefdedb.jpg", + "table_caption": [ + "Table 3. Text-to-video results on the DiDeMo dataset." + ], + "table_footnote": [], + "table_body": "
MethodR@1↑R@5↑R@10↑MR↑
MMT [12]12.929.238.827.0
FiT [4]15.030.839.828.5
MDMMT [11]18.838.547.935.1
CLIP4Clip [31]21.641.849.837.7
HiVLP20.438.848.435.9
HiVLP*24.844.154.641.2
", + "bbox": [ + 521, + 516, + 870, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Text-to-video results on the LSMDC dataset.", + "bbox": [ + 535, + 628, + 854, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LSMDC, and DiDeMo datasets to evaluate text-video retrieval. We report zero-shot and fine-tuning results (Table 1) on MSR-VTT. For zero-shot retrieval, pre-trained with the same small amount of data, HiVLP outperforms FiT by a large margin (7.4% on R@1). CLIP, Florence, BLIP, and HiVLP* use large-scale datasets for pre-training, with the numbers of pairs 400M, 900M, 239M, and 116.5M, respectively. BLIP performs best among previous works. However, even with much fewer training pairs (116.5M vs. 239M), our HiVLP* outperforms BLIP on all the metrics except obtaining the same MedR. Note that BLIP uses both filtered and synthetic captions in LAION115M and D-14M, while we only use the filtered captions in LAION100M and D-14M.", + "bbox": [ + 496, + 656, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fine-tuning comparison, HiVLP/HiVLP* is finetuned with the VTC and VTM losses. In Table 1, we see that", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bdd3bf2a915731a5fe9e5dbef156371cf6fb9e67b8cf77ddc0d220cca6e8c676.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMSR-VTTMSVDLDMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP4Clip-meanP+ [31]43.170.581.256.679.784.320.639.447.5
CLIP4Clip-seqTransf+ [31]42.770.980.662.087.392.620.839.048.6
X-Pool+ [13]44.473.384.066.490.094.222.742.651.2
HiVLP*51.575.983.167.289.595.725.645.653.8
", + "bbox": [ + 130, + 88, + 839, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7f9f417cedea22ceb6be247fe2c4f3269b053b6213e3eb6afb014cededbad08c.jpg", + "table_caption": [ + "Table 5. Comparison with SOTA methods on MSR-VTT, MSVD, and LSMDC for video-text retrieval. ${}^{ + }$ indicates using both CLIP's image and text encoders as its encoders." + ], + "table_footnote": [], + "table_body": "
Method#Test FramesMSVDMSR-VTT
B4↑M↑R↑C↑B4↑M↑R↑C↑
SibNet [27]3054.234.871.788.240.927.560.247.5
SAAT [62]2846.533.569.481.039.927.761.251.0
STG-KD [36]1652.236.973.993.040.528.360.947.1
PMI-CAP [8]3254.636.4-95.142.128.7-49.4
ORG-TRL [61]2854.336.473.995.243.628.862.150.9
OpenBook [60]28----33.923.750.252.9
SwinBERT [24]6466.342.480.9149.445.430.664.155.9
HiVLP1667.145.381.8144.547.431.264.662.3
HiVLP*1668.345.182.0151.649.232.465.967.8
", + "bbox": [ + 165, + 228, + 805, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Comparison with SOTA methods on MSVD and MSR-VTT for video captioning.", + "bbox": [ + 214, + 395, + 746, + 411 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e48530dfecfe2719643d9c7563a7bc3b607f04c3cf6d4f5649c2cf5486b304e7.jpg", + "image_caption": [ + "Figure 4. Diversity vs. amount of training pairs." + ], + "image_footnote": [], + "bbox": [ + 96, + 425, + 450, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HiVLP uses the same pre-training datasets as FiT but again outperforms it by a large margin (10.1% on Rank@1). With only about 29% amount of pre-training pairs in CLIP4Clip, HiVLP* exceeds CLIP4Clip on all the metrics significantly. Tables 2, 3, and 4 show the comparison on another three datasets, where HiVLP* outperforms the other methods by a large margin.", + "bbox": [ + 75, + 641, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video-Text Retrieval. In Table 5, we use HiVLP* to compare with the SOTA methods which are built on CLIP. Our method can beat these CLIP-based methods with fewer pretraining pairs (116.5M vs. 400M).", + "bbox": [ + 75, + 748, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Captioning. We also evaluate our model on MSR-VTT and MSVD for video captioning. Four popular metrics BLEU4 [37] (B4), METEOR [5] (M), ROUGE-L [23] (R), and CIDEr [48] (C) are employed. We fine-tune our model with the LM loss for the video-language generation task.", + "bbox": [ + 75, + 809, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 6, HiVLP works best on all the met", + "bbox": [ + 96, + 885, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rics except one. Note that SwinBERT adopts 64 frames for testing.", + "bbox": [ + 498, + 422, + 890, + 453 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 465, + 653, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We randomly choose 1/10 image-text pairs from CC3M, resulting in $300\\mathrm{k}$ image-text pairs to do ablation study. We surprisingly find that our model pre-trained with only $300\\mathrm{k}$ image-text pairs can beat the models SupportSet and HDVILA pre-trained respectively on Howto100M and HDVILA-100M, when they are transferred to the MSR-VTT dataset for text-to-video zero-shot retrieval (45.3 vs. 31.1 vs. 41.6 on R@10).", + "bbox": [ + 496, + 489, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image-Text Pairs to Enhance VLP. As shown in Figure 4, when the number of image-text pairs is increased from 3M to 114M but with the same set of video-text pairs, the Rank@1 on the MSR-VTT dataset raises a lot (from 26.4 to 43.5); when only using much fewer image-text pairs of CC3M than video-text pairs of HD-VILP (0.3M vs. 100M), the Rank@1 performance is much better (17.4 vs. 14.4). It reveals that image-text pairs can bring more diversity than the same amount or more video-text pairs, and more image-text pairs benefit a lot. We believe this is a significant way to enhance the performance of VLP models.", + "bbox": [ + 496, + 612, + 892, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MVC Loss Design. We use the HiVLP without the MVC loss as the baseline model. As shown in Table 7, the visual SSL both G2G and G2L-1 are able to improve the performance of the baseline. Using the G2L-1 contrastive loss is better than the G2G contrastive loss, because G2L-1 can dig out local information while G2G cannot. And using more number of different scales of visual features in the MVC loss (i.e., G2L-2 and G2L-3) can further boost the align", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/292bdb87cbf0f444070b3195f1a7b878f5d40ba1c2ebf4a1ec5581a251dd0f5a.jpg", + "image_caption": [ + "Caption: A jackal is walking around in a field", + "Figure 5. Grad-CAM visualization on the cross-attention maps of the different-scale visual features injected to the CA blocks. We extract the attention maps in the 8th layer of the multi-modal video-text encoder. The word \"jackal\" in the caption is the attention target in the video. (Best viewed on screen.)" + ], + "image_footnote": [], + "bbox": [ + 91, + 104, + 872, + 354 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/672d38c9f19fdbbe81ae2614583aa765cec5e210ff78b3d70c17d414c7ce2c57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMVC lossR@10↑MR↑
Baseline-43.531.1
+G2G(L1,L1′′)44.231.4
+G2L-1(L1,L1′)44.931.8
+G2L-2(L1,L1′), (L2,L2′)45.232.2
+G2L-3(L1,L1′), (L2,L2′), (L3,L3′)45.332.5
", + "bbox": [ + 83, + 420, + 464, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ment between the multi-scale visual features and the textual features.", + "bbox": [ + 75, + 650, + 468, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of HVFG. Table 8 compares our hierarchical vision features with only single-scale ones. HiVLP-1/8 uses large single scales of visual features to achieve the best performance among HiVLP-1/32, HiVLP-1/16, and HiVLP-1/8. However, HiVLP-1/8 does not efficiently use the deeper and small scales of visual features with higher-level semantics. To solve this problem, HiVLP effectively uses both global and fine-grained visual features to interact with textual features, resulting in much better performance than HiVLP-1/8 (32.5 vs. 30.5 on MR).", + "bbox": [ + 75, + 684, + 468, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization. As Figure 5 shown, HiVLP can capture fine-grained information better than HiVLP-1/8. The attention maps of HiVLP-1/16 look random due to its low-resolution feature maps.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4273e4e3e186f500e37c0d0192f1fea96d498ca04a3476e06a4d5114e0e50652.jpg", + "table_caption": [ + "Table 7. Impact of the MVC loss for zero-shot transfer to MSR-VTT. Note that $(L_{j}, L_{j}^{\\prime})$ is the $j$ scale of global-to-local (G2L) contrastive loss, with $L_{j}, L_{j}^{\\prime}, j = 1, 2, 3$ indicated in Figure 3(a). $L_{1}^{\\prime \\prime}$ denotes replacing the dense projection $L_{1}^{\\prime}$ by the global projection as indicated in the upper part of Figure 3(b). And $(L_{1}, L_{1}^{\\prime \\prime})$ denotes the $j = 1$ global-to-global (G2G) contrastive loss. MR: Mean Rank." + ], + "table_footnote": [], + "table_body": "
Modelscales of featuresR@1↑R@10↑MR↑
HiVLP-1/32{1/32, 1/32, 1/32}14.043.430.0
HiVLP-1/16{1/16, 1/16, 1/16}14.343.630.2
HiVLP-1/8{1/8, 1/8, 1/8}14.843.730.5
HiVLP{1/32, 1/16, 1/8}17.445.332.5
", + "bbox": [ + 500, + 421, + 895, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study of different vision features injecting to CA blocks for text-to-video retrieval zero-shot transfer to MSR-VTT.", + "bbox": [ + 498, + 505, + 893, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 559, + 617, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented HiVLP, a novel hierarchical interactive video-language pre-training framework. Different from previous methods that input single-scale visual features to cross-attention blocks, HiVLP injects a hierarchical vision feature group (HVFG) to effectively use both global and fine-grained visual features for interaction with textual features. Additionally, our HiVLP is pre-trained with multilevel self-supervised learning that can further improve the model performance. We also reveal that VLP models benefit a lot from the diversity of image-text pairs. Extensive experimental results of downstream tasks (text-video retrieval, video-text retrieval, and video captioning) on 4 popular benchmark datasets show that HiVLP is able to achieve better performance than previous SOTA methods overall by a large margin.", + "bbox": [ + 496, + 585, + 890, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 829, + 666, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.mindspore.cn/.", + "[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021.", + "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017.", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021.", + "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL, 2005.", + "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021.", + "[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011.", + "[8] Shaoxiang Chen, Wenhao Jiang, Wei Liu, and Yu-Gang Jiang. Learning modality interaction for temporal sentence localization and event captioning in videos. In ECCV, 2020.", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019.", + "[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. Mdmmt: Multidomain multimodal transformer for video retrieval. In CVPR, June 2021.", + "[12] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020.", + "[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022.", + "[14] Peiyan Guan, Renjing Pei, Bin Shao, Jianzhuang Liu, Weimian Li, Jiaxi Gu, Songcen Xu, Youliang Yan, and Edmund Lam. Pidro: Parallel isomeric attention with dynamic routing for text-video retrieval. In ICCV, 2023.", + "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021.", + "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. *IJCV*, 123(1):32–73, 2017." + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021.", + "[19] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In CVPR, 2022.", + "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022.", + "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021.", + "[22] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In EMNLP, 2020.", + "[23] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In ACL, 2004.", + "[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022.", + "[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017.", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014.", + "[27] Sheng Liu, Zhou Ren, and Junsong Yuan. Sibnet: Sibling convolutional encoder for video captioning. In NeurIPS, 2020.", + "[28] Yang Liu, Samuel Albania, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019.", + "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021.", + "[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pretraining model for multimodal understanding and generation. arXiv:2002.06353, 2020.", + "[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv:2104.08860, 2021.", + "[32] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In ACM MM, 2021.", + "[33] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Ef-" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13764", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ficient text-to-visual retrieval with transformers. In CVPR, 2021.", + "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019.", + "[35] Vicente Ordonez, Girish Kulkarni, and Tamara L. Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011.", + "[36] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020.", + "[37] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002.", + "[38] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021.", + "[39] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for videolanguage retrieval. In CVPR, 2023.", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021.", + "[41] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. In ICML, 2018.", + "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020.", + "[43] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015.", + "[44] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv:2006.09199, 2020.", + "[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clipfiltered 400 million image-text pairs. arXiv:2111.02114, 2021.", + "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018.", + "[47] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, 2019." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015.", + "[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022.", + "[50] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018.", + "[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv:2105.09996, 2021.", + "[52] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, 2021.", + "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016.", + "[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022.", + "[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022.", + "[56] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, 2022.", + "[57] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv:2111.07783, 2021.", + "[58] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv:2111.11432, 2021.", + "[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021.", + "[60] Ziqi Zhang, Zhongang Qi, Chunfeng Yuan, Ying Shan, Bing Li, Ying Deng, and Weiming Hu. Open-book video captioning with retrieve-copy-generate network. In CVPR, 2021.", + "[61] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13765", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with teacher-recommended learning for video captioning. In CVPR, 2020.", + "[62] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In CVPR, 2020.", + "[63] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018.", + "[64] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020." + ], + "bbox": [ + 78, + 90, + 467, + 220 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13766", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_model.json b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e75e71c87430e5597020076565185b7596c4bf7 --- /dev/null +++ b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "HiVLP: Hierarchical Interactive Video-Language Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.182, + 0.726, + 0.233 + ], + "angle": 0, + "content": "Bin Shao, Jianzhuang Liu, Renjing Pei, Songcen Xu, Peng Dai, Juwei Lu, Weimian Li, Youliang Yan, Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.237, + 0.711, + 0.27 + ], + "angle": 0, + "content": "{shaobin3, liu.jianzhuang, peirenjing, xusongcen, peng.dai, juwei.lu, liweimian, yanyouliang}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.745 + ], + "angle": 0, + "content": "Video-Language Pre-training (VLP) has become one of the most popular research topics in deep learning. However, compared to image-language pre-training, VLP has lagged far behind due to the lack of large amounts of video-text pairs. In this work, we train a VLP model with a hybrid of image-text and video-text pairs, which significantly outperforms pre-training with only the video-text pairs. Besides, existing methods usually model the cross-modal interaction using cross-attention between single-scale visual tokens and textual tokens. These visual features are either of low resolutions lacking fine-grained information, or of high resolutions without high-level semantics. To address the issue, we propose Hierarchical interactive Video-Language Pre-training (HiVLP) that efficiently uses a hierarchical visual feature group for multi-modal cross-attention during pre-training. In the hierarchical framework, low-resolution features are learned with focus on more global high-level semantic information, while high-resolution features carry fine-grained details. As a result, HiVLP has the ability to effectively learn both the global and fine-grained representations to achieve better alignment between video and text inputs. Furthermore, we design a hierarchical multi-scale vision contrastive loss for self-supervised learning to boost the interaction between them. Experimental results show that HiVLP establishes new state-of-the-art results in three downstream tasks, text-video retrieval, video-text retrieval, and video captioning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.21, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Recently, the framework of pre-training with large-scale uncurated data and then fine-tuning on some specific downstream tasks has attracted much attention. It firstly emerges in the field of Natural Language Processing (NLP), such as BERT [10], GPT [41] and T5 [42], which are pretrained on a large corpus of web-scraped dataset and then fine-tuned on a wide variety of NLP downstream tasks." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.302, + 0.68, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.407, + 0.681, + 0.416 + ], + "angle": 0, + "content": "(a) Coarse Visual Feature Injecting" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.302, + 0.895, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.407, + 0.878, + 0.417 + ], + "angle": 0, + "content": "(b) Fine-Grain Visual Feature Injecting" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.42, + 0.866, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.551, + 0.895, + 0.649 + ], + "angle": 0, + "content": "Figure 1. (a) Coarse visual feature injecting to CA blocks. (b) Fine-grained visual feature injecting to CA blocks. (c) Hierarchical visual feature injecting to CA blocks. A CA block consists of a casual/bi-directional self-attention layer, a cross-attention layer, and a feed forward layer. It is with bi-directional self-attention for vision-language understanding and with casual self-attention layer for vision-language generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Hereafter, it is transferred rapidly to the computer vision area. For examples, CLIP [40], ALIGN [16], Florence [58] and BLIP [20] all use more than 100 million open-domain image-text pairs in Image-Language Pre-training (ILP). However, most of the existing Video-Language Pretraining (VLP) works [47, 22, 51, 64, 30] use either a small-scale dataset (e.g., YouCookII [63] with 14K video-text pairs) or a large-scale dataset with less diversity (e.g., Howto100M [34] sourced from 1.22M videos). To solve this problem, we use a larger-scale dataset with 114M image-text pairs and a dataset with 2.5M video-text pairs to pre-train our model. We show that diversity is more important than the total amount of training pairs, and a small set of image-text pairs can achieve much better performance than using millions of video-text pairs. We believe this is a significant way to enhance VLP models and alleviate the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13756" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.347, + 0.106 + ], + "angle": 0, + "content": "cost of the collection of video-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.47, + 0.424 + ], + "angle": 0, + "content": "In VLP and ILP, existing works [47, 22, 64, 21, 56] often use cross-attention to model the cross-modal interaction between visual features and text features. However, they usually adopt only the single-scale and low-resolution visual features (i.e., \\(\\frac{1}{16}\\) scale of the input) for cross-attention (CA) blocks, as shown in Figure 1(a). This scheme fails to obtain fine-grained interaction with text features and limits the performance of the pre-training model. For finer-grained interactions, [33] injects the high-resolution visual features (i.e., \\(\\frac{1}{4}\\) scale of the input) to CA blocks as shown in Figure 1(b), but it does not have high-level semantics. To overcome these limitations, we propose Hierarchical Interactive Video-Language Pre-training (HiVLP) that efficiently uses a Hierarchical Visual Feature Group (HVFG) for multi-modal cross-attention. As shown in Figure 1(c), HVFG includes different scales of visual features, where the low-resolution ones with high-level semantics are beneficial for global representation and the high-resolution ones with detailed information are useful for fine-grained interaction. Especially, HVFG is able to achieve much better accuracy because of such a multi-scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Many works [47, 64, 51] use self-supervised learning to assist the video-language pre-training by reconstructing the masked frame tokens. However, it may introduce noise to interactions between visual and textual features for the masked frame tokens [32]. In this paper, we propose a Multi-level Vision Contrastive (MVC) loss for our HiVLP by applying a global-to-local contrast learning to every scale in HVFG. The MVC loss does not damage the visual tokens and helps the multi-level alignment between visual and textual features." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.579, + 0.421, + 0.593 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.612, + 0.468, + 0.687 + ], + "angle": 0, + "content": "- To the best of our knowledge, our HiVLP is the first work that uses a hierarchical interaction for video-language pre-training. It is able to effectively learn both the global and fine-grained representations for better alignment between visual and textual features." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.704, + 0.468, + 0.763 + ], + "angle": 0, + "content": "- We design a multi-level vision contrastive (MVC) loss for self-supervised learning that can sufficiently mine multi-level visual information to help video-language pre-training." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.78, + 0.468, + 0.825 + ], + "angle": 0, + "content": "- We reveal that diversity is more important than the amount of training pairs, and using more diverse image-text pairs benefits a lot for VLP." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- Our HiVLP unifies video-language understanding and generation. It achieves state-of-the-art results in text-video retrieval, video-text retrieval, and video captioning." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.612, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.119, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Image-Language Pre-Training (ILP). CLIP [40] is the pioneering work that collects large-scale web data (400M image-text pairs) and achieves competitive zero-shot performance on a variety of downstream tasks [39, 14]. ALIGN [16] is pre-trained with a larger-scale dataset (1.8B) obtaining better performance. FILIP [57] is pre-trained with 300M image-text pairs and designs a cross-modal late interaction mechanism for fine-grained contrastive learning. The key behind their success is that they take the advantages of large-scale datasets which are currently not available in VLP. To deal with this problem, we train our HiVLP model with a hybrid of video-text and image-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Video-Language Pre-Training (VLP). Existing VLP works either use a pre-trained S3D [50] to extract visual features as vision input [30, 52, 51] to speed up the training process, or firstly perform pre-training on video-text pairs and then transfer the model to video-language generation tasks [47, 64, 22]. However, both these two training approaches limit the model performance because they are not trained end-to-end [24]. Our HiVLP jointly trains the model with image-text and video-text pairs end-to-end, and unifies both video-language understanding and generation in one framework. FiT [4] also involves pre-training with both image-text and video-text pairs, but can not do video-language generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Self-Supervised Learning (SSL). To effectively use datasets, many works [47, 64, 51, 49, 55] use self-supervised learning to assist VLP. VideoBERT [47] tokenizes video frames by hierarchical vector quantization, and then performs SSL by predicting the masked visual tokens. ActBERT [64] predicts the action and object words of masked video tokens. VLM [51] masks either all video tokens or all text tokens, and then uses tokens from one modality to recover masked tokens. However, the existing methods may damage the visual tokens, which introduces noise into cross-attention between visual and textual tokens [32]. For better visual representation learning and avoiding damaging visual tokens, we introduce the MVC loss which maximizes the mutual information between multiscale global and local representations, and improves the multi-level alignment between visual and textual features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Hierarchical Interaction. As far as we know, there is no related work about hierarchical interaction in VLP. The most related work is VinVL [59] in ILP. VinVL [59] uses an object detector to extract different sizes of objects as visual features to do cross-attention with textual features. The method of VinVL is complicated and the accuracy of its visual features is limited by the object detector. However, HiVLP makes different scales of visual features interact with textual features without the need of an object detector, which is more effective." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13757" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.587, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.091, + 0.882, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.327, + 0.882, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.475, + 0.892, + 0.601 + ], + "angle": 0, + "content": "Figure 2. (a) Overview of HiVLP which consists of a vision encoder, a text encoder, a multi-modal video-text encoder and a multi-modal video-text decoder. Each image (or video clip) is transformed into two augmentations as the inputs to the vision encoder and the momentum vision encoder, respectively. The hierarchical visual feature group (HVFG) has scales \\(\\left\\{\\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}\\). The multi-level visual contrastive (MVC) loss is for multi-level SSL. Besides, HiVLP also uses the visual features of \\(\\frac{1}{32}\\) scale from the vision encoder and the output of the text encoder to optimize the vision-text contrastive (VTC) loss. Additionally, the multi-modal video-text encoder is trained with a vision-text matching (VTM) loss for the matching of video-/image-text pairs. The multi-modal video-text decoder is trained with a language modeling (LM) loss for video captioning. (b) The light FPN [25] is used for gradually upsampling and fusing the output features of the Swin Transformer. (c) The temporal block consists of a temporal self-attention layer, a SW-MSA layer [29] and a feed forward layer. We only replace the last block of the Swin Transformer with the temporal block for less parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.188, + 0.629 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.684 + ], + "angle": 0, + "content": "We firstly describe the detailed architecture of HiVLP. Then we present the hierarchical interaction HVFG. Finally, the pre-training objectives are given." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.212, + 0.71 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Figure 2(a), HiVLP consists of a vision encoder, a momentum vision encoder, a text encoder, a multi-modal video-text encoder, and a multi-modal videotext decoder. The vision encoder and the text encoder extract image and text features respectively, and do not need to interact with each other for a fast approximate nearest neighbor search in inference [33]. The multi-modal videotext encoder re-ranks the top-k video-text pairs by an additional MLP head that predicts whether they are matched or not. The multi-modal video-text decoder is used to perform the video-language generation (e.g., video captioning). The momentum vision encoder has the same archi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.657 + ], + "angle": 0, + "content": "tecture as the vision encoder, and the weights are updated by a momentum-based moving average strategy as in MoCo [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.81 + ], + "angle": 0, + "content": "The main component of the vision encoder is a hierarchical backbone Swin Transformer [29], followed by a light feature pyramid network (FPN) [25]. Besides, we add a temporal self-attention layer in the last block of Swin Transformer to capture the temporal information, as shown in Figure 2(c). In Figure 2(b), the light FPN uses the output features of \\(\\frac{1}{32}\\), \\(\\frac{1}{16}\\), and \\(\\frac{1}{8}\\) scales from Swin Transformer as inputs, and then gradually upsamples those features for feature fusion. Note that to reduce computation overhead, we do not use \\(\\frac{1}{4}\\) scale of visual features in HVFG." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The text encoder's architecture is the same as BERT [10], which consists of 12 transformer layers. The multi-modal video-text encoder has 12 cross-attention (CA) blocks each with a bi-directional self-attention layer. The multi-modal video-text decoder consists of 12 CA blocks each with a casual self-attention layer." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13758" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.09, + 0.526, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.09, + 0.825, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.247, + 0.78, + 0.261 + ], + "angle": 0, + "content": "Figure 3. (a) Multi-level global-to-local contrastive loss. (b) Generation of global and local features." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.273, + 0.303, + 0.289 + ], + "angle": 0, + "content": "3.2. Hierarchical Interaction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.474, + 0.508 + ], + "angle": 0, + "content": "Existing works often inject single-scale visual features to CA blocks to interact with textual features. For example, ALBEF [21] and BLIP [20] use visual features of \\(\\frac{1}{16}\\) scale, which fail to do fine-grained cross-attention. In contrast, [33] adopts features of \\(\\frac{1}{4}\\) scale, but these over detailed features without high-level semantics. In our work, we propose a hierarchical interaction mechanism between visual and textual features via injecting HVFG into the CA blocks as shown in Figure 2(a). HVFG includes visual features of three scales \\(\\left\\{\\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}\\). The visual features of \\(\\frac{1}{8}\\) scale contain fine-grained information, and those of \\(\\frac{1}{32}\\) scale carry high-level semantics. Therefore, HVFG is able to obtain both global and fine-grained representations to boost the interaction between visual and textual features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.518, + 0.3, + 0.535 + ], + "angle": 0, + "content": "3.3. Pre-Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.542, + 0.47, + 0.633 + ], + "angle": 0, + "content": "The input of the vision encoder is an image or video clip \\( X \\in R^{B \\times M \\times 3 \\times H \\times W} \\), which consists of \\( M \\) frames with the resolution \\( H \\times W \\), where \\( M = 1 \\) for images and \\( B \\) is the batch size. The input to the text encoder is a batch of tokenized sequences of words \\( T \\in N^{B \\times L} \\), where \\( L \\) is the max length of a caption." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.633, + 0.47, + 0.709 + ], + "angle": 0, + "content": "In our method, an input video clip \\( X \\) is split into non-overlapping patches \\( \\{x_{i}\\}_{i = 1}^{BMN}, x_{i} \\in R^{3 \\times P \\times P} \\), where \\( P \\times P \\) is input path size and \\( N = \\frac{H \\times W}{P \\times P} \\). The patches are tokenized by a linear embedding layer which is followed by a linear embedding layer of Swin Transformer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.471, + 0.83 + ], + "angle": 0, + "content": "Vision-Text Contrastive (VTC) Loss. VTC aims to pull positive pairs of vision and language representations together and push negative pairs far away. Like ALBEF [21] using a momentum distillation for vision-language pretraining, we introduce two queues to store the most recent \\( K \\) vision-text representations pairs from momentum encoders. Formally, the video-to-text contrastive loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.851, + 0.475, + 0.902 + ], + "angle": 0, + "content": "\\[\nL _ {T 2 V} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {- e x p \\left(s \\left(f _ {i} , t _ {i} ^ {+}\\right) / \\rho\\right)}{e x p \\left(s \\left(f _ {i} , t _ {i} ^ {+}\\right) / \\rho\\right) + \\sum_ {j = 1} ^ {K} e x p \\left(s \\left(f _ {i} , t _ {j} ^ {-}\\right) / \\rho\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.895, + 0.382 + ], + "angle": 0, + "content": "where \\(\\rho\\) is the temperature, \\(t_i^+\\) and \\(t_j^-\\) are the text features of the positive and negative text samples for the \\(i\\)-th vision input \\(X_i\\) respectively, and \\(f_i\\) is the visual feature of \\(\\frac{1}{32}\\) scale from \\(X_i\\). \\(s(f_i, t_i^+)\\) denotes the cosine similarity between \\(f_i\\) and \\(t_i^+\\) that are matched, and \\(s(f_i, t_j^-)\\) denotes the similarity between \\(f_i\\) and \\(t_j^-\\) that are not matched. Symmetrically, the text-to-video contrastive loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.406, + 0.9, + 0.454 + ], + "angle": 0, + "content": "\\[\nL _ {V 2 T} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {- \\exp \\left(s \\left(t _ {i} , f _ {i} ^ {+}\\right) / \\rho\\right)}{\\exp \\left(s \\left(t _ {i} , f _ {i} ^ {+}\\right) / \\rho\\right) + \\sum_ {j = 1} ^ {K} \\exp \\left(s \\left(t _ {i} , f _ {j} ^ {-}\\right) / \\rho\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.486 + ], + "angle": 0, + "content": "where \\( f_{i}^{+} \\) and \\( f_{j}^{-} \\) are positive and negative visual features of the \\( i \\)-th text feature, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.892, + 0.515 + ], + "angle": 0, + "content": "The sum of \\( L_{V2T} \\) and \\( L_{T2V} \\) serve as the vision-text contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.527, + 0.892, + 0.555 + ], + "angle": 0, + "content": "\\[\nL _ {V T C} = \\frac {1}{2} \\left(L _ {V 2 T} + L _ {T 2 V}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.895, + 0.743 + ], + "angle": 0, + "content": "Multi-level Vision Contrastive (MVC) Loss. As shown in Figure 3(a), MVC aims to pull the global representation of a scale closer to those local patch representations from the same vision input in different views of a scale. The light FPN has three scales \\( L_{1} \\), \\( L_{2} \\), and \\( L_{3} \\) (see Figure 2(b) and Figure 3(a)). The global feature \\( f_{ij} \\) of scale \\( j \\) of the \\( i \\)-th visual input is obtained by the average of all the patch features at \\( L_{j} \\) (Figure 3(b)). The local features \\( f_{ijk}, k = 1, 2, 3, \\dots, s_{j} \\) are generated as shown in Figure 3(b) where \\( f_{ijk} \\) is the average feature pooling result of a local window such as \\( 2 \\times 2 \\), and \\( s_{j} \\) is the number of the local averaged features at \\( L_{j}' \\). Finally, the MVC loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.762, + 0.892, + 0.814 + ], + "angle": 0, + "content": "\\[\nL _ {M V C} ^ {i j} = \\frac {\\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j} ^ {+}\\right) / \\rho\\right)}{\\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j} ^ {+}\\right) / \\rho\\right) + \\sum_ {k = 1} ^ {(B - 1) S _ {j}} \\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j k} ^ {-}\\right) / \\rho\\right)}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.824, + 0.892, + 0.863 + ], + "angle": 0, + "content": "\\[\nL _ {M V C} = - \\frac {1}{3 B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {3} - l o g L _ {M V C} ^ {i j}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.869, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{f}_{ij}\\) and \\(\\hat{f}_{ijk}\\) correspond to \\(f_{ij}\\) and \\(f_{ijk}\\), respectively, but from the other augmentation of the visual input." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13759" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Vision-Text Matching (VTM) Loss. Through a MLP layer, VTM is used to judge whether a video-/image-text pair is matched or not with the cross entropy function. There are \\( B \\) positive pairs from the batch, and \\( 2B \\) negative pairs are obtained according to the hard negative mining strategy in [21]. This loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.208, + 0.47, + 0.261 + ], + "angle": 0, + "content": "\\[\nL _ {V T M} = - \\frac {1}{3 B} \\sum_ {i = 1} ^ {3 B} \\left[ y _ {i} \\log \\left(p _ {i}\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - p _ {i}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.292 + ], + "angle": 0, + "content": "where \\( y_{i} \\) is the ground-truth, and \\( p_i \\) is the output of the MLP layer of the \\( i \\)-th pair." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.47, + 0.355 + ], + "angle": 0, + "content": "Language Modeling (LM) Loss. We inject the hierarchical visual representations into the multi-modal video-text decoder to generate captions. Let \\(i\\)-th input text be \\(T_{i} = [t_{i}^{1}, t_{i}^{2}, \\dots, t_{i}^{L}]\\). This loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.379, + 0.468, + 0.434 + ], + "angle": 0, + "content": "\\[\nL _ {L M} = - \\frac {1}{B L} \\sum_ {i = 1} ^ {B} \\sum_ {l = 1} ^ {L} \\log \\left(p \\left(t _ {i} ^ {l} \\mid t _ {i} ^ {l - 1}, \\dots , t _ {i} ^ {1}, f _ {i 1}, f _ {i 2}, f _ {i 3}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.468, + 0.495 + ], + "angle": 0, + "content": "where \\( p(t_i^l | t_i^{l-1}, \\dots, t_i^1, f_{i1}, f_{i2}, f_{i3}) \\) is the output probability of the multi-modal video-text decoder for the \\( l \\)-th token \\( t_i^l \\) given the previously fed tokens \\( t_i^{l-1}, \\dots, t_i^1 \\) and the multilevel visual features \\( f_{i1}, f_{i2}, f_{i3} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.495, + 0.462, + 0.51 + ], + "angle": 0, + "content": "Total Loss. Finally, we have the total loss to train HiVLP:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.539, + 0.469, + 0.555 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = \\alpha L _ {V T C} + \\beta L _ {M V C} + \\gamma L _ {V T M} + \\delta L _ {L M}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.598 + ], + "angle": 0, + "content": "where \\(\\alpha\\), \\(\\beta\\), \\(\\gamma\\), and \\(\\delta\\) are the weight coefficients of the losses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.614, + 0.21, + 0.631 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.7 + ], + "angle": 0, + "content": "In this section, we first describe the pre-training and downstream datasets, and the implementation details. Then we show the experimental and comparison results, and finally ablate our model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.181, + 0.725 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Pre-Training Datasets. Following FiT [4], we involve two popular datasets (CC-3M [46] and WebVid-2M [4]) for pre-training HiVLP. CC-3M consists of 3M image-text pairs and WebVid-2M includes 2.5M video-text pairs, in total 5.5M image-/video-text pairs. Besides, we use LAION100M, D-14M, and WebVid-2M (total 116.5M image-/video-text pairs) as the larger-scale dataset to train our model, resulting in HiVLP*. LAION100M is a subset of LAION [45]. D-14M is a combined dataset of image-text pairs from COCO [26], Visual Genome [17], CC3M [46], CC12M [6], and SBU captions [35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Downstream Task Datasets. There are four downstream datasets used in our work for video retrieval. (i) MSR-VTT [53] consists of 10K YouTube videos with 200K human-annotated descriptions. Similar to previous methods [4], we use the split of 9K for training and 1K for testing. (ii) MSVD [7] is a smaller dataset with 1970 videos and 78800 sentences, about 40 descriptions per video. Like [4], we use the split of 1200, 100, and 670 videos for training, validation and testing, respectively. (iii) DiDeMo [3] has 10k videos, each of which is described by multiple sentences, resulting in 40K sentences. For a fair comparison, we follow the setting in [4], where all descriptions for one video are concatenated into a single sentence. (iv) LSMDC [43] contains 128K clips. Like [4], we use 7408 clips for validation and 1000 clips for testing. For video captioning, we evaluate our model on MSR-VTT and MSVD. Following the split of [30], on MSR-VTT, we use \\(6.5K\\) training videos and \\(2.9K\\) testing videos; on MSVD, we use \\(1.2K\\) training videos and 670 testing videos." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.392, + 0.719, + 0.408 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Our model HiVLP is pre-trained on 16 GPUs (32G memory). Our vision encoder is initialized by Swin-B [29] pretrained on ImageNet-21k [9]. The light FPN is randomly initialized. The text encoder, multi-modal video-text encoder, and multi-modal video-text decoder are BERTs initialized from BERTbase [10]. HiVLP is pre-trained for 25 epochs, the first 20 epochs only with image-text pairs with a batch size of 30 and the last 5 epochs with both image-text and video-text pairs. Each video clip consists of 4 frames. We use AdamW optimizer with a weight decay of 0.02, and the learning rate is initialized as \\(10^{-5}\\) and is warmed up to \\(10^{-4}\\) after 3,000 training iterations. We then decrease the learning rate by the cosine decay strategy to \\(10^{-5}\\). For the hyper-parameters in Equation 8, we set \\(\\alpha\\), \\(\\beta\\), and \\(\\delta\\) to 1.0 and \\(\\gamma\\) to 0.001. The input resolution is \\(224 \\times 224\\) with data augmentations used in [21]. And we use 160 GPUs (16G memory) to train HiVLP* with a batch size of 18 per GPU. For speeding up the training of HiVLP*, we sample 1 frame from each clip." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.902 + ], + "angle": 0, + "content": "For the downstream task of text-video retrieval, we sample 4 frames per video for training. Since the multi-modal video-text encoder filters top-\\(k\\) candidates during inference (\\(k\\) is set to 128). We sample 8 frames per video for testing on MSR-VTT, MSVD, and LSMDC. Because the videos in DiDeMo are longer, we sample 10 frames for testing from it. For the downstream task of video captioning, we evaluate our model on MSR-VTT and MSVD, with 8 random frames per video for training and 16 frames per video for testing. For all downstream tasks, the initial learning rate is set to \\(5 \\times 10^{-6}\\), and the weight decay, the batch size, and the total number of epochs are set to 0.05, 64, and 10, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13760" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.135, + 0.088, + 0.837, + 0.459 + ], + "angle": 0, + "content": "
MethodPre-Training Datasets#PairsR@1↑R@5↑R@10↑MedR↓MR↑
Zero-Shot
SupportSet [38]HowTo100M100M8.723.031.131.020.9
HD-VILA [54]HD-VILA-100M100M14.431.641.617.529.2
FiT [4]CC3M, WebVid-2M5.5M18.739.551.610.036.6
ALPRO [19]CC3M, WebVid-2M5.5M24.144.755.48.041.4
CLIP [40]WIT400M400M30.654.464.34.049.8
Florence [58]FLD-900M900M37.663.872.6-58
BLIP [20]L-115M, D-14M239M43.365.674.72.061.2
HiVLPCC3M, WebVid-2M5.5M26.447.355.77.043.1
HiVLP*L-100M, D-14M, WebVid-2M116.5M43.566.476.42.062.1
Fine-Tuning
ActBERT [64]HowTo100M100M16.342.856.910.038.7
HERO [22]HowTo100M100M16.843.457.7-39.3
NoiseEstimation [2]HowTo100M100M17.441.653.68.037.5
UniVL [30]HowTo100M100M21.249.663.16.044.6
ClipBERT [18]HowTo100M100M22.046.859.96.042.9
AVLnet [44]HowTo100M100M27.155.666.64.049.8
MMT [12]HowTo100M100M26.657.169.64.051.1
SupportSet [38]HowTo100M100M30.158.569.33.052.6
FiT [4]CC3M, WebVid-2M5.5M31.059.570.53.053.7
ALPRO [19]CC3M, WebVid-2M5.5M33.960.773.23.055.9
HD-VILA [54]HD-VILA-100M100M35.065.277.23.059.1
CLIP4Clip [31]WIT400M400M44.571.481.62.065.8
HiVLPCC3M, WebVid-2M5.5M41.165.975.92.061.0
HiVLP*L-100M, D-14M, WebVid-2M116.5M50.976.483.61.070.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.467, + 0.895, + 0.496 + ], + "angle": 0, + "content": "Table 1. Comparisons with SOTA text-to-video retrieval methods on MSR-VTT. R@K: Recall@K; MedR: Median Rank; MR: Mean Rank; L-115M: LAION115M; L-100M: LAION100M." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.516, + 0.451, + 0.62 + ], + "angle": 0, + "content": "
MethodR@1↑R@5↑R@10↑MR↑
CE [28]19.849.063.844.2
Support Set [38]28.460.072.953.8
FiT [4]33.764.776.357.3
CLIP4Clip [31]46.276.184.668.9
HiVLP39.168.177.261.5
HiVLP*50.278.985.871.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.629, + 0.428, + 0.643 + ], + "angle": 0, + "content": "Table 2. Text-to-video results on the MSVD dataset." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.657, + 0.447, + 0.76 + ], + "angle": 0, + "content": "
MethodR@1↑R@5↑R@10↑MR↑
ClipBERT [18]20.444.556.740.5
FiT [4]34.665.074.758.1
ALPRO [54]35.967.578.860.7
CLIP4Clip [31]43.470.280.664.7
HiVLP40.568.777.162.1
HiVLP*48.173.882.568.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.769, + 0.434, + 0.782 + ], + "angle": 0, + "content": "Table 3. Text-to-video results on the DiDeMo dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.797, + 0.454, + 0.814 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.47, + 0.884 + ], + "angle": 0, + "content": "In this section, we compare HiVLP with state-of-the-art (SOTA) methods on three popular video-language downstream tasks (text-video retrieval, video-text retrieval, and video captioning)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Text-Video Retrieval. We use MSR-VTT, MSVD," + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.517, + 0.872, + 0.62 + ], + "angle": 0, + "content": "
MethodR@1↑R@5↑R@10↑MR↑
MMT [12]12.929.238.827.0
FiT [4]15.030.839.828.5
MDMMT [11]18.838.547.935.1
CLIP4Clip [31]21.641.849.837.7
HiVLP20.438.848.435.9
HiVLP*24.844.154.641.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.629, + 0.855, + 0.643 + ], + "angle": 0, + "content": "Table 4. Text-to-video results on the LSMDC dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.868 + ], + "angle": 0, + "content": "LSMDC, and DiDeMo datasets to evaluate text-video retrieval. We report zero-shot and fine-tuning results (Table 1) on MSR-VTT. For zero-shot retrieval, pre-trained with the same small amount of data, HiVLP outperforms FiT by a large margin (7.4% on R@1). CLIP, Florence, BLIP, and HiVLP* use large-scale datasets for pre-training, with the numbers of pairs 400M, 900M, 239M, and 116.5M, respectively. BLIP performs best among previous works. However, even with much fewer training pairs (116.5M vs. 239M), our HiVLP* outperforms BLIP on all the metrics except obtaining the same MedR. Note that BLIP uses both filtered and synthetic captions in LAION115M and D-14M, while we only use the filtered captions in LAION100M and D-14M." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "For fine-tuning comparison, HiVLP/HiVLP* is finetuned with the VTC and VTM losses. In Table 1, we see that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13761" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.841, + 0.178 + ], + "angle": 0, + "content": "
MethodMSR-VTTMSVDLDMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP4Clip-meanP+ [31]43.170.581.256.679.784.320.639.447.5
CLIP4Clip-seqTransf+ [31]42.770.980.662.087.392.620.839.048.6
X-Pool+ [13]44.473.384.066.490.094.222.742.651.2
HiVLP*51.575.983.167.289.595.725.645.653.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.189, + 0.894, + 0.218 + ], + "angle": 0, + "content": "Table 5. Comparison with SOTA methods on MSR-VTT, MSVD, and LSMDC for video-text retrieval. \\( {}^{ + } \\) indicates using both CLIP's image and text encoders as its encoders." + }, + { + "type": "table", + "bbox": [ + 0.166, + 0.229, + 0.807, + 0.387 + ], + "angle": 0, + "content": "
Method#Test FramesMSVDMSR-VTT
B4↑M↑R↑C↑B4↑M↑R↑C↑
SibNet [27]3054.234.871.788.240.927.560.247.5
SAAT [62]2846.533.569.481.039.927.761.251.0
STG-KD [36]1652.236.973.993.040.528.360.947.1
PMI-CAP [8]3254.636.4-95.142.128.7-49.4
ORG-TRL [61]2854.336.473.995.243.628.862.150.9
OpenBook [60]28----33.923.750.252.9
SwinBERT [24]6466.342.480.9149.445.430.664.155.9
HiVLP1667.145.381.8144.547.431.264.662.3
HiVLP*1668.345.182.0151.649.232.465.967.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.396, + 0.748, + 0.412 + ], + "angle": 0, + "content": "Table 6. Comparison with SOTA methods on MSVD and MSR-VTT for video captioning." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.426, + 0.451, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.614, + 0.416, + 0.629 + ], + "angle": 0, + "content": "Figure 4. Diversity vs. amount of training pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.748 + ], + "angle": 0, + "content": "HiVLP uses the same pre-training datasets as FiT but again outperforms it by a large margin (10.1% on Rank@1). With only about 29% amount of pre-training pairs in CLIP4Clip, HiVLP* exceeds CLIP4Clip on all the metrics significantly. Tables 2, 3, and 4 show the comparison on another three datasets, where HiVLP* outperforms the other methods by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Video-Text Retrieval. In Table 5, we use HiVLP* to compare with the SOTA methods which are built on CLIP. Our method can beat these CLIP-based methods with fewer pretraining pairs (116.5M vs. 400M)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.885 + ], + "angle": 0, + "content": "Video Captioning. We also evaluate our model on MSR-VTT and MSVD for video captioning. Four popular metrics BLEU4 [37] (B4), METEOR [5] (M), ROUGE-L [23] (R), and CIDEr [48] (C) are employed. We fine-tune our model with the LM loss for the video-language generation task." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "As shown in Table 6, HiVLP works best on all the met" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.423, + 0.892, + 0.454 + ], + "angle": 0, + "content": "rics except one. Note that SwinBERT adopts 64 frames for testing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.654, + 0.483 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.611 + ], + "angle": 0, + "content": "We randomly choose 1/10 image-text pairs from CC3M, resulting in \\(300\\mathrm{k}\\) image-text pairs to do ablation study. We surprisingly find that our model pre-trained with only \\(300\\mathrm{k}\\) image-text pairs can beat the models SupportSet and HDVILA pre-trained respectively on Howto100M and HDVILA-100M, when they are transferred to the MSR-VTT dataset for text-to-video zero-shot retrieval (45.3 vs. 31.1 vs. 41.6 on R@10)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Image-Text Pairs to Enhance VLP. As shown in Figure 4, when the number of image-text pairs is increased from 3M to 114M but with the same set of video-text pairs, the Rank@1 on the MSR-VTT dataset raises a lot (from 26.4 to 43.5); when only using much fewer image-text pairs of CC3M than video-text pairs of HD-VILP (0.3M vs. 100M), the Rank@1 performance is much better (17.4 vs. 14.4). It reveals that image-text pairs can bring more diversity than the same amount or more video-text pairs, and more image-text pairs benefit a lot. We believe this is a significant way to enhance the performance of VLP models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.901 + ], + "angle": 0, + "content": "MVC Loss Design. We use the HiVLP without the MVC loss as the baseline model. As shown in Table 7, the visual SSL both G2G and G2L-1 are able to improve the performance of the baseline. Using the G2L-1 contrastive loss is better than the G2G contrastive loss, because G2L-1 can dig out local information while G2G cannot. And using more number of different scales of visual features in the MVC loss (i.e., G2L-2 and G2L-3) can further boost the align" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13762" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.092, + 0.608, + 0.106 + ], + "angle": 0, + "content": "Caption: A jackal is walking around in a field" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.106, + 0.874, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.358, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Figure 5. Grad-CAM visualization on the cross-attention maps of the different-scale visual features injected to the CA blocks. We extract the attention maps in the 8th layer of the multi-modal video-text encoder. The word \"jackal\" in the caption is the attention target in the video. (Best viewed on screen.)" + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.421, + 0.465, + 0.517 + ], + "angle": 0, + "content": "
ModelMVC lossR@10↑MR↑
Baseline-43.531.1
+G2G(L1,L1′′)44.231.4
+G2L-1(L1,L1′)44.931.8
+G2L-2(L1,L1′), (L2,L2′)45.232.2
+G2L-3(L1,L1′), (L2,L2′), (L3,L3′)45.332.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Table 7. Impact of the MVC loss for zero-shot transfer to MSR-VTT. Note that \\((L_{j}, L_{j}^{\\prime})\\) is the \\(j\\) scale of global-to-local (G2L) contrastive loss, with \\(L_{j}, L_{j}^{\\prime}, j = 1, 2, 3\\) indicated in Figure 3(a). \\(L_{1}^{\\prime \\prime}\\) denotes replacing the dense projection \\(L_{1}^{\\prime}\\) by the global projection as indicated in the upper part of Figure 3(b). And \\((L_{1}, L_{1}^{\\prime \\prime})\\) denotes the \\(j = 1\\) global-to-global (G2G) contrastive loss. MR: Mean Rank." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.681 + ], + "angle": 0, + "content": "ment between the multi-scale visual features and the textual features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Impact of HVFG. Table 8 compares our hierarchical vision features with only single-scale ones. HiVLP-1/8 uses large single scales of visual features to achieve the best performance among HiVLP-1/32, HiVLP-1/16, and HiVLP-1/8. However, HiVLP-1/8 does not efficiently use the deeper and small scales of visual features with higher-level semantics. To solve this problem, HiVLP effectively uses both global and fine-grained visual features to interact with textual features, resulting in much better performance than HiVLP-1/8 (32.5 vs. 30.5 on MR)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Visualization. As Figure 5 shown, HiVLP can capture fine-grained information better than HiVLP-1/8. The attention maps of HiVLP-1/16 look random due to its low-resolution feature maps." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.422, + 0.897, + 0.497 + ], + "angle": 0, + "content": "
Modelscales of featuresR@1↑R@10↑MR↑
HiVLP-1/32{1/32, 1/32, 1/32}14.043.430.0
HiVLP-1/16{1/16, 1/16, 1/16}14.343.630.2
HiVLP-1/8{1/8, 1/8, 1/8}14.843.730.5
HiVLP{1/32, 1/16, 1/8}17.445.332.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.506, + 0.895, + 0.548 + ], + "angle": 0, + "content": "Table 8. Ablation study of different vision features injecting to CA blocks for text-to-video retrieval zero-shot transfer to MSR-VTT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.619, + 0.575 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.813 + ], + "angle": 0, + "content": "We have presented HiVLP, a novel hierarchical interactive video-language pre-training framework. Different from previous methods that input single-scale visual features to cross-attention blocks, HiVLP injects a hierarchical vision feature group (HVFG) to effectively use both global and fine-grained visual features for interaction with textual features. Additionally, our HiVLP is pre-trained with multilevel self-supervised learning that can further improve the model performance. We also reveal that VLP models benefit a lot from the diversity of image-text pairs. Extensive experimental results of downstream tasks (text-video retrieval, video-text retrieval, and video captioning) on 4 popular benchmark datasets show that HiVLP is able to achieve better performance than previous SOTA methods overall by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.667, + 0.846 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13763" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.277, + 0.129 + ], + "angle": 0, + "content": "[1] https://www.mindspore.cn/." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.469, + 0.172 + ], + "angle": 0, + "content": "[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.469, + 0.215 + ], + "angle": 0, + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.469, + 0.258 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.39, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[8] Shaoxiang Chen, Wenhao Jiang, Wei Liu, and Yu-Gang Jiang. Learning modality interaction for temporal sentence localization and event captioning in videos. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.468, + 0.472 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. Mdmmt: Multidomain multimodal transformer for video retrieval. In CVPR, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[12] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[14] Peiyan Guan, Renjing Pei, Bin Shao, Jianzhuang Liu, Weimian Li, Jiaxi Gu, Songcen Xu, Youliang Yan, and Edmund Lam. Pidro: Parallel isomeric attention with dynamic routing for text-video retrieval. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. *IJCV*, 123(1):32–73, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[19] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[22] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In EMNLP, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[23] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In ACL, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[27] Sheng Liu, Zhou Ren, and Junsong Yuan. Sibnet: Sibling convolutional encoder for video captioning. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[28] Yang Liu, Samuel Albania, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pretraining model for multimodal understanding and generation. arXiv:2002.06353, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv:2104.08860, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[32] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In ACM MM, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Ef-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13764" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ficient text-to-visual retrieval with transformers. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[35] Vicente Ordonez, Girish Kulkarni, and Tamara L. Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.469, + 0.293 + ], + "angle": 0, + "content": "[36] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.469, + 0.337 + ], + "angle": 0, + "content": "[37] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[38] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.398, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[39] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for videolanguage retrieval. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[41] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. In ICML, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.469, + 0.656 + ], + "angle": 0, + "content": "[43] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[44] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv:2006.09199, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clipfiltered 400 million image-text pairs. arXiv:2111.02114, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[50] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv:2105.09996, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[52] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[56] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[57] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv:2111.07783, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[58] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv:2111.11432, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[60] Ziqi Zhang, Zhongang Qi, Chunfeng Yuan, Ying Shan, Bing Li, Ying Deng, and Weiming Hu. Open-book video captioning with retrieve-copy-generate network. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[61] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13765" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "with teacher-recommended learning for video captioning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[62] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.151, + 0.468, + 0.191 + ], + "angle": 0, + "content": "[63] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.468, + 0.221 + ], + "angle": 0, + "content": "[64] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13766" + } + ] +] \ No newline at end of file diff --git a/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_origin.pdf b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eb6d90c1120a443eb8b03fb9f6d2135652814922 --- /dev/null +++ b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/e5850211-fab6-4860-b782-c2a5d6bad4e4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2245bef972cac7bde20826a2ba7947e6301c071d4c559cf012c90c0e9ab07c5c +size 1610168 diff --git a/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/full.md b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c66a43c1d61ee77ac7603637a095442cf4cac79 --- /dev/null +++ b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/full.md @@ -0,0 +1,309 @@ +# HiVLP: Hierarchical Interactive Video-Language Pre-Training + +Bin Shao, Jianzhuang Liu, Renjing Pei, Songcen Xu, Peng Dai, Juwei Lu, Weimian Li, Youliang Yan, Huawei Noah's Ark Lab + +{shaobin3, liu.jianzhuang, peirenjing, xusongcen, peng.dai, juwei.lu, liweimian, yanyouliang}@huawei.com + +# Abstract + +Video-Language Pre-training (VLP) has become one of the most popular research topics in deep learning. However, compared to image-language pre-training, VLP has lagged far behind due to the lack of large amounts of video-text pairs. In this work, we train a VLP model with a hybrid of image-text and video-text pairs, which significantly outperforms pre-training with only the video-text pairs. Besides, existing methods usually model the cross-modal interaction using cross-attention between single-scale visual tokens and textual tokens. These visual features are either of low resolutions lacking fine-grained information, or of high resolutions without high-level semantics. To address the issue, we propose Hierarchical interactive Video-Language Pre-training (HiVLP) that efficiently uses a hierarchical visual feature group for multi-modal cross-attention during pre-training. In the hierarchical framework, low-resolution features are learned with focus on more global high-level semantic information, while high-resolution features carry fine-grained details. As a result, HiVLP has the ability to effectively learn both the global and fine-grained representations to achieve better alignment between video and text inputs. Furthermore, we design a hierarchical multi-scale vision contrastive loss for self-supervised learning to boost the interaction between them. Experimental results show that HiVLP establishes new state-of-the-art results in three downstream tasks, text-video retrieval, video-text retrieval, and video captioning. + +# 1. Introduction + +Recently, the framework of pre-training with large-scale uncurated data and then fine-tuning on some specific downstream tasks has attracted much attention. It firstly emerges in the field of Natural Language Processing (NLP), such as BERT [10], GPT [41] and T5 [42], which are pretrained on a large corpus of web-scraped dataset and then fine-tuned on a wide variety of NLP downstream tasks. + +![](images/94bbadf5bf352a1c7f4750ededafb07582a2d1ee4e9c22cdc407d4113db61b41.jpg) +(a) Coarse Visual Feature Injecting + +![](images/2e16e61e12c61e0b4fdae632a6ed49e853f2da63d3d64235dc9efc3d678c1470.jpg) + +![](images/e7c1e467c132d8af97264895ab2d4dfd1bbc871f6f1099608efe4cebe40bc622.jpg) +(b) Fine-Grain Visual Feature Injecting +Figure 1. (a) Coarse visual feature injecting to CA blocks. (b) Fine-grained visual feature injecting to CA blocks. (c) Hierarchical visual feature injecting to CA blocks. A CA block consists of a casual/bi-directional self-attention layer, a cross-attention layer, and a feed forward layer. It is with bi-directional self-attention for vision-language understanding and with casual self-attention layer for vision-language generation. + +Hereafter, it is transferred rapidly to the computer vision area. For examples, CLIP [40], ALIGN [16], Florence [58] and BLIP [20] all use more than 100 million open-domain image-text pairs in Image-Language Pre-training (ILP). However, most of the existing Video-Language Pretraining (VLP) works [47, 22, 51, 64, 30] use either a small-scale dataset (e.g., YouCookII [63] with 14K video-text pairs) or a large-scale dataset with less diversity (e.g., Howto100M [34] sourced from 1.22M videos). To solve this problem, we use a larger-scale dataset with 114M image-text pairs and a dataset with 2.5M video-text pairs to pre-train our model. We show that diversity is more important than the total amount of training pairs, and a small set of image-text pairs can achieve much better performance than using millions of video-text pairs. We believe this is a significant way to enhance VLP models and alleviate the + +cost of the collection of video-text pairs. + +In VLP and ILP, existing works [47, 22, 64, 21, 56] often use cross-attention to model the cross-modal interaction between visual features and text features. However, they usually adopt only the single-scale and low-resolution visual features (i.e., $\frac{1}{16}$ scale of the input) for cross-attention (CA) blocks, as shown in Figure 1(a). This scheme fails to obtain fine-grained interaction with text features and limits the performance of the pre-training model. For finer-grained interactions, [33] injects the high-resolution visual features (i.e., $\frac{1}{4}$ scale of the input) to CA blocks as shown in Figure 1(b), but it does not have high-level semantics. To overcome these limitations, we propose Hierarchical Interactive Video-Language Pre-training (HiVLP) that efficiently uses a Hierarchical Visual Feature Group (HVFG) for multi-modal cross-attention. As shown in Figure 1(c), HVFG includes different scales of visual features, where the low-resolution ones with high-level semantics are beneficial for global representation and the high-resolution ones with detailed information are useful for fine-grained interaction. Especially, HVFG is able to achieve much better accuracy because of such a multi-scale. + +Many works [47, 64, 51] use self-supervised learning to assist the video-language pre-training by reconstructing the masked frame tokens. However, it may introduce noise to interactions between visual and textual features for the masked frame tokens [32]. In this paper, we propose a Multi-level Vision Contrastive (MVC) loss for our HiVLP by applying a global-to-local contrast learning to every scale in HVFG. The MVC loss does not damage the visual tokens and helps the multi-level alignment between visual and textual features. + +Our contributions can be summarized as follows: + +- To the best of our knowledge, our HiVLP is the first work that uses a hierarchical interaction for video-language pre-training. It is able to effectively learn both the global and fine-grained representations for better alignment between visual and textual features. +- We design a multi-level vision contrastive (MVC) loss for self-supervised learning that can sufficiently mine multi-level visual information to help video-language pre-training. +- We reveal that diversity is more important than the amount of training pairs, and using more diverse image-text pairs benefits a lot for VLP. +- Our HiVLP unifies video-language understanding and generation. It achieves state-of-the-art results in text-video retrieval, video-text retrieval, and video captioning. + +# 2. Related Work + +Image-Language Pre-Training (ILP). CLIP [40] is the pioneering work that collects large-scale web data (400M image-text pairs) and achieves competitive zero-shot performance on a variety of downstream tasks [39, 14]. ALIGN [16] is pre-trained with a larger-scale dataset (1.8B) obtaining better performance. FILIP [57] is pre-trained with 300M image-text pairs and designs a cross-modal late interaction mechanism for fine-grained contrastive learning. The key behind their success is that they take the advantages of large-scale datasets which are currently not available in VLP. To deal with this problem, we train our HiVLP model with a hybrid of video-text and image-text pairs. + +Video-Language Pre-Training (VLP). Existing VLP works either use a pre-trained S3D [50] to extract visual features as vision input [30, 52, 51] to speed up the training process, or firstly perform pre-training on video-text pairs and then transfer the model to video-language generation tasks [47, 64, 22]. However, both these two training approaches limit the model performance because they are not trained end-to-end [24]. Our HiVLP jointly trains the model with image-text and video-text pairs end-to-end, and unifies both video-language understanding and generation in one framework. FiT [4] also involves pre-training with both image-text and video-text pairs, but can not do video-language generation. + +Self-Supervised Learning (SSL). To effectively use datasets, many works [47, 64, 51, 49, 55] use self-supervised learning to assist VLP. VideoBERT [47] tokenizes video frames by hierarchical vector quantization, and then performs SSL by predicting the masked visual tokens. ActBERT [64] predicts the action and object words of masked video tokens. VLM [51] masks either all video tokens or all text tokens, and then uses tokens from one modality to recover masked tokens. However, the existing methods may damage the visual tokens, which introduces noise into cross-attention between visual and textual tokens [32]. For better visual representation learning and avoiding damaging visual tokens, we introduce the MVC loss which maximizes the mutual information between multiscale global and local representations, and improves the multi-level alignment between visual and textual features. + +Hierarchical Interaction. As far as we know, there is no related work about hierarchical interaction in VLP. The most related work is VinVL [59] in ILP. VinVL [59] uses an object detector to extract different sizes of objects as visual features to do cross-attention with textual features. The method of VinVL is complicated and the accuracy of its visual features is limited by the object detector. However, HiVLP makes different scales of visual features interact with textual features without the need of an object detector, which is more effective. + +![](images/e229b6afe7cd86bdc79de7941233721cd18b1ef93088ee505a3a320e80ef2360.jpg) +Figure 2. (a) Overview of HiVLP which consists of a vision encoder, a text encoder, a multi-modal video-text encoder and a multi-modal video-text decoder. Each image (or video clip) is transformed into two augmentations as the inputs to the vision encoder and the momentum vision encoder, respectively. The hierarchical visual feature group (HVFG) has scales $\left\{\frac{1}{8}, \frac{1}{16}, \frac{1}{32}\right\}$ . The multi-level visual contrastive (MVC) loss is for multi-level SSL. Besides, HiVLP also uses the visual features of $\frac{1}{32}$ scale from the vision encoder and the output of the text encoder to optimize the vision-text contrastive (VTC) loss. Additionally, the multi-modal video-text encoder is trained with a vision-text matching (VTM) loss for the matching of video-/image-text pairs. The multi-modal video-text decoder is trained with a language modeling (LM) loss for video captioning. (b) The light FPN [25] is used for gradually upsampling and fusing the output features of the Swin Transformer. (c) The temporal block consists of a temporal self-attention layer, a SW-MSA layer [29] and a feed forward layer. We only replace the last block of the Swin Transformer with the temporal block for less parameters. + +![](images/69b14ea2ca6e4e54ff382d171feef75dbe8defcf9214da9d9d49a0ff31f2388f.jpg) + +![](images/08f4a6bd3543f38c924c253a3858bfd6a7315661fcc64e7bb0f16cc563431223.jpg) + +# 3. Approach + +We firstly describe the detailed architecture of HiVLP. Then we present the hierarchical interaction HVFG. Finally, the pre-training objectives are given. + +# 3.1. Architecture + +As illustrated in Figure 2(a), HiVLP consists of a vision encoder, a momentum vision encoder, a text encoder, a multi-modal video-text encoder, and a multi-modal videotext decoder. The vision encoder and the text encoder extract image and text features respectively, and do not need to interact with each other for a fast approximate nearest neighbor search in inference [33]. The multi-modal videotext encoder re-ranks the top-k video-text pairs by an additional MLP head that predicts whether they are matched or not. The multi-modal video-text decoder is used to perform the video-language generation (e.g., video captioning). The momentum vision encoder has the same archi + +tecture as the vision encoder, and the weights are updated by a momentum-based moving average strategy as in MoCo [15]. + +The main component of the vision encoder is a hierarchical backbone Swin Transformer [29], followed by a light feature pyramid network (FPN) [25]. Besides, we add a temporal self-attention layer in the last block of Swin Transformer to capture the temporal information, as shown in Figure 2(c). In Figure 2(b), the light FPN uses the output features of $\frac{1}{32}$ , $\frac{1}{16}$ , and $\frac{1}{8}$ scales from Swin Transformer as inputs, and then gradually upsamples those features for feature fusion. Note that to reduce computation overhead, we do not use $\frac{1}{4}$ scale of visual features in HVFG. + +The text encoder's architecture is the same as BERT [10], which consists of 12 transformer layers. The multi-modal video-text encoder has 12 cross-attention (CA) blocks each with a bi-directional self-attention layer. The multi-modal video-text decoder consists of 12 CA blocks each with a casual self-attention layer. + +![](images/a4c078d814f3a0ec3e3c38a2023b63f09a6c0f4cfbb18923540c42a9668571b6.jpg) +Figure 3. (a) Multi-level global-to-local contrastive loss. (b) Generation of global and local features. + +![](images/0f67c0f87f724134ec31964016c157eeb897bd9f7c8760fd73335ee0d94686d2.jpg) + +# 3.2. Hierarchical Interaction + +Existing works often inject single-scale visual features to CA blocks to interact with textual features. For example, ALBEF [21] and BLIP [20] use visual features of $\frac{1}{16}$ scale, which fail to do fine-grained cross-attention. In contrast, [33] adopts features of $\frac{1}{4}$ scale, but these over detailed features without high-level semantics. In our work, we propose a hierarchical interaction mechanism between visual and textual features via injecting HVFG into the CA blocks as shown in Figure 2(a). HVFG includes visual features of three scales $\left\{\frac{1}{8}, \frac{1}{16}, \frac{1}{32}\right\}$ . The visual features of $\frac{1}{8}$ scale contain fine-grained information, and those of $\frac{1}{32}$ scale carry high-level semantics. Therefore, HVFG is able to obtain both global and fine-grained representations to boost the interaction between visual and textual features. + +# 3.3. Pre-Training Objectives + +The input of the vision encoder is an image or video clip $X \in R^{B \times M \times 3 \times H \times W}$ , which consists of $M$ frames with the resolution $H \times W$ , where $M = 1$ for images and $B$ is the batch size. The input to the text encoder is a batch of tokenized sequences of words $T \in N^{B \times L}$ , where $L$ is the max length of a caption. + +In our method, an input video clip $X$ is split into non-overlapping patches $\{x_{i}\}_{i = 1}^{BMN}, x_{i} \in R^{3 \times P \times P}$ , where $P \times P$ is input path size and $N = \frac{H \times W}{P \times P}$ . The patches are tokenized by a linear embedding layer which is followed by a linear embedding layer of Swin Transformer. + +Vision-Text Contrastive (VTC) Loss. VTC aims to pull positive pairs of vision and language representations together and push negative pairs far away. Like ALBEF [21] using a momentum distillation for vision-language pretraining, we introduce two queues to store the most recent $K$ vision-text representations pairs from momentum encoders. Formally, the video-to-text contrastive loss is defined as: + +$$ +L _ {T 2 V} = \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {- e x p \left(s \left(f _ {i} , t _ {i} ^ {+}\right) / \rho\right)}{e x p \left(s \left(f _ {i} , t _ {i} ^ {+}\right) / \rho\right) + \sum_ {j = 1} ^ {K} e x p \left(s \left(f _ {i} , t _ {j} ^ {-}\right) / \rho\right)}, \tag {1} +$$ + +where $\rho$ is the temperature, $t_i^+$ and $t_j^-$ are the text features of the positive and negative text samples for the $i$ -th vision input $X_i$ respectively, and $f_i$ is the visual feature of $\frac{1}{32}$ scale from $X_i$ . $s(f_i, t_i^+)$ denotes the cosine similarity between $f_i$ and $t_i^+$ that are matched, and $s(f_i, t_j^-)$ denotes the similarity between $f_i$ and $t_j^-$ that are not matched. Symmetrically, the text-to-video contrastive loss is: + +$$ +L _ {V 2 T} = \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {- \exp \left(s \left(t _ {i} , f _ {i} ^ {+}\right) / \rho\right)}{\exp \left(s \left(t _ {i} , f _ {i} ^ {+}\right) / \rho\right) + \sum_ {j = 1} ^ {K} \exp \left(s \left(t _ {i} , f _ {j} ^ {-}\right) / \rho\right)}, \tag {2} +$$ + +where $f_{i}^{+}$ and $f_{j}^{-}$ are positive and negative visual features of the $i$ -th text feature, respectively. + +The sum of $L_{V2T}$ and $L_{T2V}$ serve as the vision-text contrastive loss: + +$$ +L _ {V T C} = \frac {1}{2} \left(L _ {V 2 T} + L _ {T 2 V}\right). \tag {3} +$$ + +Multi-level Vision Contrastive (MVC) Loss. As shown in Figure 3(a), MVC aims to pull the global representation of a scale closer to those local patch representations from the same vision input in different views of a scale. The light FPN has three scales $L_{1}$ , $L_{2}$ , and $L_{3}$ (see Figure 2(b) and Figure 3(a)). The global feature $f_{ij}$ of scale $j$ of the $i$ -th visual input is obtained by the average of all the patch features at $L_{j}$ (Figure 3(b)). The local features $f_{ijk}, k = 1, 2, 3, \dots, s_{j}$ are generated as shown in Figure 3(b) where $f_{ijk}$ is the average feature pooling result of a local window such as $2 \times 2$ , and $s_{j}$ is the number of the local averaged features at $L_{j}'$ . Finally, the MVC loss is formulated as: + +$$ +L _ {M V C} ^ {i j} = \frac {\exp \left(s \left(f _ {i j} , \hat {f} _ {i j} ^ {+}\right) / \rho\right)}{\exp \left(s \left(f _ {i j} , \hat {f} _ {i j} ^ {+}\right) / \rho\right) + \sum_ {k = 1} ^ {(B - 1) S _ {j}} \exp \left(s \left(f _ {i j} , \hat {f} _ {i j k} ^ {-}\right) / \rho\right)}, \tag {4} +$$ + +$$ +L _ {M V C} = - \frac {1}{3 B} \sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {3} - l o g L _ {M V C} ^ {i j}, \tag {5} +$$ + +where $\hat{f}_{ij}$ and $\hat{f}_{ijk}$ correspond to $f_{ij}$ and $f_{ijk}$ , respectively, but from the other augmentation of the visual input. + +Vision-Text Matching (VTM) Loss. Through a MLP layer, VTM is used to judge whether a video-/image-text pair is matched or not with the cross entropy function. There are $B$ positive pairs from the batch, and $2B$ negative pairs are obtained according to the hard negative mining strategy in [21]. This loss is defined as: + +$$ +L _ {V T M} = - \frac {1}{3 B} \sum_ {i = 1} ^ {3 B} \left[ y _ {i} \log \left(p _ {i}\right) + \left(1 - y _ {i}\right) \log \left(1 - p _ {i}\right) \right], \tag {6} +$$ + +where $y_{i}$ is the ground-truth, and $p_i$ is the output of the MLP layer of the $i$ -th pair. + +Language Modeling (LM) Loss. We inject the hierarchical visual representations into the multi-modal video-text decoder to generate captions. Let $i$ -th input text be $T_{i} = [t_{i}^{1}, t_{i}^{2}, \dots, t_{i}^{L}]$ . This loss is defined as: + +$$ +L _ {L M} = - \frac {1}{B L} \sum_ {i = 1} ^ {B} \sum_ {l = 1} ^ {L} \log \left(p \left(t _ {i} ^ {l} \mid t _ {i} ^ {l - 1}, \dots , t _ {i} ^ {1}, f _ {i 1}, f _ {i 2}, f _ {i 3}\right)\right), \tag {7} +$$ + +where $p(t_i^l | t_i^{l-1}, \dots, t_i^1, f_{i1}, f_{i2}, f_{i3})$ is the output probability of the multi-modal video-text decoder for the $l$ -th token $t_i^l$ given the previously fed tokens $t_i^{l-1}, \dots, t_i^1$ and the multilevel visual features $f_{i1}, f_{i2}, f_{i3}$ . + +Total Loss. Finally, we have the total loss to train HiVLP: + +$$ +L _ {t o t a l} = \alpha L _ {V T C} + \beta L _ {M V C} + \gamma L _ {V T M} + \delta L _ {L M}, \tag {8} +$$ + +where $\alpha$ , $\beta$ , $\gamma$ , and $\delta$ are the weight coefficients of the losses. + +# 4. Experiments + +In this section, we first describe the pre-training and downstream datasets, and the implementation details. Then we show the experimental and comparison results, and finally ablate our model. + +# 4.1. Datasets + +Pre-Training Datasets. Following FiT [4], we involve two popular datasets (CC-3M [46] and WebVid-2M [4]) for pre-training HiVLP. CC-3M consists of 3M image-text pairs and WebVid-2M includes 2.5M video-text pairs, in total 5.5M image-/video-text pairs. Besides, we use LAION100M, D-14M, and WebVid-2M (total 116.5M image-/video-text pairs) as the larger-scale dataset to train our model, resulting in HiVLP*. LAION100M is a subset of LAION [45]. D-14M is a combined dataset of image-text pairs from COCO [26], Visual Genome [17], CC3M [46], CC12M [6], and SBU captions [35]. + +Downstream Task Datasets. There are four downstream datasets used in our work for video retrieval. (i) MSR-VTT [53] consists of 10K YouTube videos with 200K human-annotated descriptions. Similar to previous methods [4], we use the split of 9K for training and 1K for testing. (ii) MSVD [7] is a smaller dataset with 1970 videos and 78800 sentences, about 40 descriptions per video. Like [4], we use the split of 1200, 100, and 670 videos for training, validation and testing, respectively. (iii) DiDeMo [3] has 10k videos, each of which is described by multiple sentences, resulting in 40K sentences. For a fair comparison, we follow the setting in [4], where all descriptions for one video are concatenated into a single sentence. (iv) LSMDC [43] contains 128K clips. Like [4], we use 7408 clips for validation and 1000 clips for testing. For video captioning, we evaluate our model on MSR-VTT and MSVD. Following the split of [30], on MSR-VTT, we use $6.5K$ training videos and $2.9K$ testing videos; on MSVD, we use $1.2K$ training videos and 670 testing videos. + +# 4.2. Implementation Details + +Our model HiVLP is pre-trained on 16 GPUs (32G memory). Our vision encoder is initialized by Swin-B [29] pretrained on ImageNet-21k [9]. The light FPN is randomly initialized. The text encoder, multi-modal video-text encoder, and multi-modal video-text decoder are BERTs initialized from BERTbase [10]. HiVLP is pre-trained for 25 epochs, the first 20 epochs only with image-text pairs with a batch size of 30 and the last 5 epochs with both image-text and video-text pairs. Each video clip consists of 4 frames. We use AdamW optimizer with a weight decay of 0.02, and the learning rate is initialized as $10^{-5}$ and is warmed up to $10^{-4}$ after 3,000 training iterations. We then decrease the learning rate by the cosine decay strategy to $10^{-5}$ . For the hyper-parameters in Equation 8, we set $\alpha$ , $\beta$ , and $\delta$ to 1.0 and $\gamma$ to 0.001. The input resolution is $224 \times 224$ with data augmentations used in [21]. And we use 160 GPUs (16G memory) to train HiVLP* with a batch size of 18 per GPU. For speeding up the training of HiVLP*, we sample 1 frame from each clip. + +For the downstream task of text-video retrieval, we sample 4 frames per video for training. Since the multi-modal video-text encoder filters top- $k$ candidates during inference ( $k$ is set to 128). We sample 8 frames per video for testing on MSR-VTT, MSVD, and LSMDC. Because the videos in DiDeMo are longer, we sample 10 frames for testing from it. For the downstream task of video captioning, we evaluate our model on MSR-VTT and MSVD, with 8 random frames per video for training and 16 frames per video for testing. For all downstream tasks, the initial learning rate is set to $5 \times 10^{-6}$ , and the weight decay, the batch size, and the total number of epochs are set to 0.05, 64, and 10, respectively. + +
MethodPre-Training Datasets#PairsR@1↑R@5↑R@10↑MedR↓MR↑
Zero-Shot
SupportSet [38]HowTo100M100M8.723.031.131.020.9
HD-VILA [54]HD-VILA-100M100M14.431.641.617.529.2
FiT [4]CC3M, WebVid-2M5.5M18.739.551.610.036.6
ALPRO [19]CC3M, WebVid-2M5.5M24.144.755.48.041.4
CLIP [40]WIT400M400M30.654.464.34.049.8
Florence [58]FLD-900M900M37.663.872.6-58
BLIP [20]L-115M, D-14M239M43.365.674.72.061.2
HiVLPCC3M, WebVid-2M5.5M26.447.355.77.043.1
HiVLP*L-100M, D-14M, WebVid-2M116.5M43.566.476.42.062.1
Fine-Tuning
ActBERT [64]HowTo100M100M16.342.856.910.038.7
HERO [22]HowTo100M100M16.843.457.7-39.3
NoiseEstimation [2]HowTo100M100M17.441.653.68.037.5
UniVL [30]HowTo100M100M21.249.663.16.044.6
ClipBERT [18]HowTo100M100M22.046.859.96.042.9
AVLnet [44]HowTo100M100M27.155.666.64.049.8
MMT [12]HowTo100M100M26.657.169.64.051.1
SupportSet [38]HowTo100M100M30.158.569.33.052.6
FiT [4]CC3M, WebVid-2M5.5M31.059.570.53.053.7
ALPRO [19]CC3M, WebVid-2M5.5M33.960.773.23.055.9
HD-VILA [54]HD-VILA-100M100M35.065.277.23.059.1
CLIP4Clip [31]WIT400M400M44.571.481.62.065.8
HiVLPCC3M, WebVid-2M5.5M41.165.975.92.061.0
HiVLP*L-100M, D-14M, WebVid-2M116.5M50.976.483.61.070.3
+ +Table 1. Comparisons with SOTA text-to-video retrieval methods on MSR-VTT. R@K: Recall@K; MedR: Median Rank; MR: Mean Rank; L-115M: LAION115M; L-100M: LAION100M. + +
MethodR@1↑R@5↑R@10↑MR↑
CE [28]19.849.063.844.2
Support Set [38]28.460.072.953.8
FiT [4]33.764.776.357.3
CLIP4Clip [31]46.276.184.668.9
HiVLP39.168.177.261.5
HiVLP*50.278.985.871.6
+ +Table 2. Text-to-video results on the MSVD dataset. + +
MethodR@1↑R@5↑R@10↑MR↑
ClipBERT [18]20.444.556.740.5
FiT [4]34.665.074.758.1
ALPRO [54]35.967.578.860.7
CLIP4Clip [31]43.470.280.664.7
HiVLP40.568.777.162.1
HiVLP*48.173.882.568.1
+ +# 4.3. Comparisons with State-of-the-Art Methods + +In this section, we compare HiVLP with state-of-the-art (SOTA) methods on three popular video-language downstream tasks (text-video retrieval, video-text retrieval, and video captioning). + +Text-Video Retrieval. We use MSR-VTT, MSVD, + +Table 3. Text-to-video results on the DiDeMo dataset. + +
MethodR@1↑R@5↑R@10↑MR↑
MMT [12]12.929.238.827.0
FiT [4]15.030.839.828.5
MDMMT [11]18.838.547.935.1
CLIP4Clip [31]21.641.849.837.7
HiVLP20.438.848.435.9
HiVLP*24.844.154.641.2
+ +Table 4. Text-to-video results on the LSMDC dataset. + +LSMDC, and DiDeMo datasets to evaluate text-video retrieval. We report zero-shot and fine-tuning results (Table 1) on MSR-VTT. For zero-shot retrieval, pre-trained with the same small amount of data, HiVLP outperforms FiT by a large margin (7.4% on R@1). CLIP, Florence, BLIP, and HiVLP* use large-scale datasets for pre-training, with the numbers of pairs 400M, 900M, 239M, and 116.5M, respectively. BLIP performs best among previous works. However, even with much fewer training pairs (116.5M vs. 239M), our HiVLP* outperforms BLIP on all the metrics except obtaining the same MedR. Note that BLIP uses both filtered and synthetic captions in LAION115M and D-14M, while we only use the filtered captions in LAION100M and D-14M. + +For fine-tuning comparison, HiVLP/HiVLP* is finetuned with the VTC and VTM losses. In Table 1, we see that + +
MethodMSR-VTTMSVDLDMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP4Clip-meanP+ [31]43.170.581.256.679.784.320.639.447.5
CLIP4Clip-seqTransf+ [31]42.770.980.662.087.392.620.839.048.6
X-Pool+ [13]44.473.384.066.490.094.222.742.651.2
HiVLP*51.575.983.167.289.595.725.645.653.8
+ +Table 5. Comparison with SOTA methods on MSR-VTT, MSVD, and LSMDC for video-text retrieval. ${}^{ + }$ indicates using both CLIP's image and text encoders as its encoders. + +
Method#Test FramesMSVDMSR-VTT
B4↑M↑R↑C↑B4↑M↑R↑C↑
SibNet [27]3054.234.871.788.240.927.560.247.5
SAAT [62]2846.533.569.481.039.927.761.251.0
STG-KD [36]1652.236.973.993.040.528.360.947.1
PMI-CAP [8]3254.636.4-95.142.128.7-49.4
ORG-TRL [61]2854.336.473.995.243.628.862.150.9
OpenBook [60]28----33.923.750.252.9
SwinBERT [24]6466.342.480.9149.445.430.664.155.9
HiVLP1667.145.381.8144.547.431.264.662.3
HiVLP*1668.345.182.0151.649.232.465.967.8
+ +Table 6. Comparison with SOTA methods on MSVD and MSR-VTT for video captioning. + +![](images/e48530dfecfe2719643d9c7563a7bc3b607f04c3cf6d4f5649c2cf5486b304e7.jpg) +Figure 4. Diversity vs. amount of training pairs. + +HiVLP uses the same pre-training datasets as FiT but again outperforms it by a large margin (10.1% on Rank@1). With only about 29% amount of pre-training pairs in CLIP4Clip, HiVLP* exceeds CLIP4Clip on all the metrics significantly. Tables 2, 3, and 4 show the comparison on another three datasets, where HiVLP* outperforms the other methods by a large margin. + +Video-Text Retrieval. In Table 5, we use HiVLP* to compare with the SOTA methods which are built on CLIP. Our method can beat these CLIP-based methods with fewer pretraining pairs (116.5M vs. 400M). + +Video Captioning. We also evaluate our model on MSR-VTT and MSVD for video captioning. Four popular metrics BLEU4 [37] (B4), METEOR [5] (M), ROUGE-L [23] (R), and CIDEr [48] (C) are employed. We fine-tune our model with the LM loss for the video-language generation task. + +As shown in Table 6, HiVLP works best on all the met + +rics except one. Note that SwinBERT adopts 64 frames for testing. + +# 4.4. Ablation Study + +We randomly choose 1/10 image-text pairs from CC3M, resulting in $300\mathrm{k}$ image-text pairs to do ablation study. We surprisingly find that our model pre-trained with only $300\mathrm{k}$ image-text pairs can beat the models SupportSet and HDVILA pre-trained respectively on Howto100M and HDVILA-100M, when they are transferred to the MSR-VTT dataset for text-to-video zero-shot retrieval (45.3 vs. 31.1 vs. 41.6 on R@10). + +Image-Text Pairs to Enhance VLP. As shown in Figure 4, when the number of image-text pairs is increased from 3M to 114M but with the same set of video-text pairs, the Rank@1 on the MSR-VTT dataset raises a lot (from 26.4 to 43.5); when only using much fewer image-text pairs of CC3M than video-text pairs of HD-VILP (0.3M vs. 100M), the Rank@1 performance is much better (17.4 vs. 14.4). It reveals that image-text pairs can bring more diversity than the same amount or more video-text pairs, and more image-text pairs benefit a lot. We believe this is a significant way to enhance the performance of VLP models. + +MVC Loss Design. We use the HiVLP without the MVC loss as the baseline model. As shown in Table 7, the visual SSL both G2G and G2L-1 are able to improve the performance of the baseline. Using the G2L-1 contrastive loss is better than the G2G contrastive loss, because G2L-1 can dig out local information while G2G cannot. And using more number of different scales of visual features in the MVC loss (i.e., G2L-2 and G2L-3) can further boost the align + +![](images/292bdb87cbf0f444070b3195f1a7b878f5d40ba1c2ebf4a1ec5581a251dd0f5a.jpg) +Caption: A jackal is walking around in a field +Figure 5. Grad-CAM visualization on the cross-attention maps of the different-scale visual features injected to the CA blocks. We extract the attention maps in the 8th layer of the multi-modal video-text encoder. The word "jackal" in the caption is the attention target in the video. (Best viewed on screen.) + +
ModelMVC lossR@10↑MR↑
Baseline-43.531.1
+G2G(L1,L1′′)44.231.4
+G2L-1(L1,L1′)44.931.8
+G2L-2(L1,L1′), (L2,L2′)45.232.2
+G2L-3(L1,L1′), (L2,L2′), (L3,L3′)45.332.5
+ +ment between the multi-scale visual features and the textual features. + +Impact of HVFG. Table 8 compares our hierarchical vision features with only single-scale ones. HiVLP-1/8 uses large single scales of visual features to achieve the best performance among HiVLP-1/32, HiVLP-1/16, and HiVLP-1/8. However, HiVLP-1/8 does not efficiently use the deeper and small scales of visual features with higher-level semantics. To solve this problem, HiVLP effectively uses both global and fine-grained visual features to interact with textual features, resulting in much better performance than HiVLP-1/8 (32.5 vs. 30.5 on MR). + +Visualization. As Figure 5 shown, HiVLP can capture fine-grained information better than HiVLP-1/8. The attention maps of HiVLP-1/16 look random due to its low-resolution feature maps. + +Table 7. Impact of the MVC loss for zero-shot transfer to MSR-VTT. Note that $(L_{j}, L_{j}^{\prime})$ is the $j$ scale of global-to-local (G2L) contrastive loss, with $L_{j}, L_{j}^{\prime}, j = 1, 2, 3$ indicated in Figure 3(a). $L_{1}^{\prime \prime}$ denotes replacing the dense projection $L_{1}^{\prime}$ by the global projection as indicated in the upper part of Figure 3(b). And $(L_{1}, L_{1}^{\prime \prime})$ denotes the $j = 1$ global-to-global (G2G) contrastive loss. MR: Mean Rank. + +
Modelscales of featuresR@1↑R@10↑MR↑
HiVLP-1/32{1/32, 1/32, 1/32}14.043.430.0
HiVLP-1/16{1/16, 1/16, 1/16}14.343.630.2
HiVLP-1/8{1/8, 1/8, 1/8}14.843.730.5
HiVLP{1/32, 1/16, 1/8}17.445.332.5
+ +Table 8. Ablation study of different vision features injecting to CA blocks for text-to-video retrieval zero-shot transfer to MSR-VTT. + +# 5. Conclusion + +We have presented HiVLP, a novel hierarchical interactive video-language pre-training framework. Different from previous methods that input single-scale visual features to cross-attention blocks, HiVLP injects a hierarchical vision feature group (HVFG) to effectively use both global and fine-grained visual features for interaction with textual features. Additionally, our HiVLP is pre-trained with multilevel self-supervised learning that can further improve the model performance. We also reveal that VLP models benefit a lot from the diversity of image-text pairs. Extensive experimental results of downstream tasks (text-video retrieval, video-text retrieval, and video captioning) on 4 popular benchmark datasets show that HiVLP is able to achieve better performance than previous SOTA methods overall by a large margin. + +# Acknowledgements + +We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research. + +# References + +[1] https://www.mindspore.cn/. +[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021. +[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. +[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL, 2005. +[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. +[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011. +[8] Shaoxiang Chen, Wenhao Jiang, Wei Liu, and Yu-Gang Jiang. Learning modality interaction for temporal sentence localization and event captioning in videos. In ECCV, 2020. +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. +[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. Mdmmt: Multidomain multimodal transformer for video retrieval. In CVPR, June 2021. +[12] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020. +[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022. +[14] Peiyan Guan, Renjing Pei, Bin Shao, Jianzhuang Liu, Weimian Li, Jiaxi Gu, Songcen Xu, Youliang Yan, and Edmund Lam. Pidro: Parallel isomeric attention with dynamic routing for text-video retrieval. In ICCV, 2023. +[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. +[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. *IJCV*, 123(1):32–73, 2017. + +[18] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021. +[19] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In CVPR, 2022. +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. +[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. +[22] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In EMNLP, 2020. +[23] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In ACL, 2004. +[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022. +[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. +[27] Sheng Liu, Zhou Ren, and Junsong Yuan. Sibnet: Sibling convolutional encoder for video captioning. In NeurIPS, 2020. +[28] Yang Liu, Samuel Albania, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019. +[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. +[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pretraining model for multimodal understanding and generation. arXiv:2002.06353, 2020. +[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv:2104.08860, 2021. +[32] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In ACM MM, 2021. +[33] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Ef- + +ficient text-to-visual retrieval with transformers. In CVPR, 2021. +[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019. +[35] Vicente Ordonez, Girish Kulkarni, and Tamara L. Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011. +[36] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020. +[37] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. +[38] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021. +[39] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for videolanguage retrieval. In CVPR, 2023. +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021. +[41] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. In ICML, 2018. +[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. +[43] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. +[44] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv:2006.09199, 2020. +[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clipfiltered 400 million image-text pairs. arXiv:2111.02114, 2021. +[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. +[47] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, 2019. + +[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015. +[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. +[50] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018. +[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv:2105.09996, 2021. +[52] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, 2021. +[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. +[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. +[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. +[56] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, 2022. +[57] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv:2111.07783, 2021. +[58] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv:2111.11432, 2021. +[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. +[60] Ziqi Zhang, Zhongang Qi, Chunfeng Yuan, Ying Shan, Bing Li, Ying Deng, and Weiming Hu. Open-book video captioning with retrieve-copy-generate network. In CVPR, 2021. +[61] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph + +with teacher-recommended learning for video captioning. In CVPR, 2020. +[62] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In CVPR, 2020. +[63] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. +[64] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020. \ No newline at end of file diff --git a/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/images.zip b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab6c6d26fc611a2cb74ff67cf78714fc7a800226 --- /dev/null +++ b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3734756e67f007b488a6c26f7a021887e993b0aa02a701b9351d7fabb0389e5a +size 789139 diff --git a/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/layout.json b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c8c0f181b318876e1f1b2cfab36303009d5278c0 --- /dev/null +++ b/2023/HiVLP_ Hierarchical Interactive Video-Language Pre-Training/layout.json @@ -0,0 +1,8911 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "text", + "content": "HiVLP: Hierarchical Interactive Video-Language Pre-Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 144, + 444, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 144, + 444, + 184 + ], + "spans": [ + { + "bbox": [ + 148, + 144, + 444, + 184 + ], + "type": "text", + "content": "Bin Shao, Jianzhuang Liu, Renjing Pei, Songcen Xu, Peng Dai, Juwei Lu, Weimian Li, Youliang Yan, Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 187, + 435, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 187, + 435, + 213 + ], + "spans": [ + { + "bbox": [ + 157, + 187, + 435, + 213 + ], + "type": "text", + "content": "{shaobin3, liu.jianzhuang, peirenjing, xusongcen, peng.dai, juwei.lu, liweimian, yanyouliang}@huawei.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 590 + ], + "type": "text", + "content": "Video-Language Pre-training (VLP) has become one of the most popular research topics in deep learning. However, compared to image-language pre-training, VLP has lagged far behind due to the lack of large amounts of video-text pairs. In this work, we train a VLP model with a hybrid of image-text and video-text pairs, which significantly outperforms pre-training with only the video-text pairs. Besides, existing methods usually model the cross-modal interaction using cross-attention between single-scale visual tokens and textual tokens. These visual features are either of low resolutions lacking fine-grained information, or of high resolutions without high-level semantics. To address the issue, we propose Hierarchical interactive Video-Language Pre-training (HiVLP) that efficiently uses a hierarchical visual feature group for multi-modal cross-attention during pre-training. In the hierarchical framework, low-resolution features are learned with focus on more global high-level semantic information, while high-resolution features carry fine-grained details. As a result, HiVLP has the ability to effectively learn both the global and fine-grained representations to achieve better alignment between video and text inputs. Furthermore, we design a hierarchical multi-scale vision contrastive loss for self-supervised learning to boost the interaction between them. Experimental results show that HiVLP establishes new state-of-the-art results in three downstream tasks, text-video retrieval, video-text retrieval, and video captioning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Recently, the framework of pre-training with large-scale uncurated data and then fine-tuning on some specific downstream tasks has attracted much attention. It firstly emerges in the field of Natural Language Processing (NLP), such as BERT [10], GPT [41] and T5 [42], which are pretrained on a large corpus of web-scraped dataset and then fine-tuned on a wide variety of NLP downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 239, + 416, + 316 + ], + "blocks": [ + { + "bbox": [ + 307, + 239, + 416, + 316 + ], + "lines": [ + { + "bbox": [ + 307, + 239, + 416, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 416, + 316 + ], + "type": "image", + "image_path": "94bbadf5bf352a1c7f4750ededafb07582a2d1ee4e9c22cdc407d4113db61b41.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 322, + 416, + 329 + ], + "lines": [ + { + "bbox": [ + 343, + 322, + 416, + 329 + ], + "spans": [ + { + "bbox": [ + 343, + 322, + 416, + 329 + ], + "type": "text", + "content": "(a) Coarse Visual Feature Injecting" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 419, + 239, + 547, + 318 + ], + "blocks": [ + { + "bbox": [ + 419, + 239, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 419, + 239, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 419, + 239, + 547, + 318 + ], + "type": "image", + "image_path": "2e16e61e12c61e0b4fdae632a6ed49e853f2da63d3d64235dc9efc3d678c1470.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 332, + 529, + 432 + ], + "blocks": [ + { + "bbox": [ + 453, + 322, + 537, + 330 + ], + "lines": [ + { + "bbox": [ + 453, + 322, + 537, + 330 + ], + "spans": [ + { + "bbox": [ + 453, + 322, + 537, + 330 + ], + "type": "text", + "content": "(b) Fine-Grain Visual Feature Injecting" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 332, + 529, + 432 + ], + "lines": [ + { + "bbox": [ + 310, + 332, + 529, + 432 + ], + "spans": [ + { + "bbox": [ + 310, + 332, + 529, + 432 + ], + "type": "image", + "image_path": "e7c1e467c132d8af97264895ab2d4dfd1bbc871f6f1099608efe4cebe40bc622.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 436, + 547, + 514 + ], + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 514 + ], + "type": "text", + "content": "Figure 1. (a) Coarse visual feature injecting to CA blocks. (b) Fine-grained visual feature injecting to CA blocks. (c) Hierarchical visual feature injecting to CA blocks. A CA block consists of a casual/bi-directional self-attention layer, a cross-attention layer, and a feed forward layer. It is with bi-directional self-attention for vision-language understanding and with casual self-attention layer for vision-language generation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Hereafter, it is transferred rapidly to the computer vision area. For examples, CLIP [40], ALIGN [16], Florence [58] and BLIP [20] all use more than 100 million open-domain image-text pairs in Image-Language Pre-training (ILP). However, most of the existing Video-Language Pretraining (VLP) works [47, 22, 51, 64, 30] use either a small-scale dataset (e.g., YouCookII [63] with 14K video-text pairs) or a large-scale dataset with less diversity (e.g., Howto100M [34] sourced from 1.22M videos). To solve this problem, we use a larger-scale dataset with 114M image-text pairs and a dataset with 2.5M video-text pairs to pre-train our model. We show that diversity is more important than the total amount of training pairs, and a small set of image-text pairs can achieve much better performance than using millions of video-text pairs. We believe this is a significant way to enhance VLP models and alleviate the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13756" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 212, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 212, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 212, + 83 + ], + "type": "text", + "content": "cost of the collection of video-text pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 287, + 335 + ], + "type": "text", + "content": "In VLP and ILP, existing works [47, 22, 64, 21, 56] often use cross-attention to model the cross-modal interaction between visual features and text features. However, they usually adopt only the single-scale and low-resolution visual features (i.e., " + }, + { + "bbox": [ + 46, + 85, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 46, + 85, + 287, + 335 + ], + "type": "text", + "content": " scale of the input) for cross-attention (CA) blocks, as shown in Figure 1(a). This scheme fails to obtain fine-grained interaction with text features and limits the performance of the pre-training model. For finer-grained interactions, [33] injects the high-resolution visual features (i.e., " + }, + { + "bbox": [ + 46, + 85, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 46, + 85, + 287, + 335 + ], + "type": "text", + "content": " scale of the input) to CA blocks as shown in Figure 1(b), but it does not have high-level semantics. To overcome these limitations, we propose Hierarchical Interactive Video-Language Pre-training (HiVLP) that efficiently uses a Hierarchical Visual Feature Group (HVFG) for multi-modal cross-attention. As shown in Figure 1(c), HVFG includes different scales of visual features, where the low-resolution ones with high-level semantics are beneficial for global representation and the high-resolution ones with detailed information are useful for fine-grained interaction. Especially, HVFG is able to achieve much better accuracy because of such a multi-scale." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 338, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 456 + ], + "type": "text", + "content": "Many works [47, 64, 51] use self-supervised learning to assist the video-language pre-training by reconstructing the masked frame tokens. However, it may introduce noise to interactions between visual and textual features for the masked frame tokens [32]. In this paper, we propose a Multi-level Vision Contrastive (MVC) loss for our HiVLP by applying a global-to-local contrast learning to every scale in HVFG. The MVC loss does not damage the visual tokens and helps the multi-level alignment between visual and textual features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 458, + 257, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 458, + 257, + 469 + ], + "spans": [ + { + "bbox": [ + 59, + 458, + 257, + 469 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 484, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 59, + 484, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 484, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 59, + 484, + 286, + 544 + ], + "type": "text", + "content": "- To the best of our knowledge, our HiVLP is the first work that uses a hierarchical interaction for video-language pre-training. It is able to effectively learn both the global and fine-grained representations for better alignment between visual and textual features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 557, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 557, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 58, + 557, + 286, + 604 + ], + "type": "text", + "content": "- We design a multi-level vision contrastive (MVC) loss for self-supervised learning that can sufficiently mine multi-level visual information to help video-language pre-training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 617, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 617, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 58, + 617, + 286, + 653 + ], + "type": "text", + "content": "- We reveal that diversity is more important than the amount of training pairs, and using more diverse image-text pairs benefits a lot for VLP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 286, + 712 + ], + "type": "text", + "content": "- Our HiVLP unifies video-language understanding and generation. It achieves state-of-the-art results in text-video retrieval, video-text retrieval, and video captioning." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 94, + 545, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 94, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 94, + 545, + 237 + ], + "type": "text", + "content": "Image-Language Pre-Training (ILP). CLIP [40] is the pioneering work that collects large-scale web data (400M image-text pairs) and achieves competitive zero-shot performance on a variety of downstream tasks [39, 14]. ALIGN [16] is pre-trained with a larger-scale dataset (1.8B) obtaining better performance. FILIP [57] is pre-trained with 300M image-text pairs and designs a cross-modal late interaction mechanism for fine-grained contrastive learning. The key behind their success is that they take the advantages of large-scale datasets which are currently not available in VLP. To deal with this problem, we train our HiVLP model with a hybrid of video-text and image-text pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 240, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 396 + ], + "type": "text", + "content": "Video-Language Pre-Training (VLP). Existing VLP works either use a pre-trained S3D [50] to extract visual features as vision input [30, 52, 51] to speed up the training process, or firstly perform pre-training on video-text pairs and then transfer the model to video-language generation tasks [47, 64, 22]. However, both these two training approaches limit the model performance because they are not trained end-to-end [24]. Our HiVLP jointly trains the model with image-text and video-text pairs end-to-end, and unifies both video-language understanding and generation in one framework. FiT [4] also involves pre-training with both image-text and video-text pairs, but can not do video-language generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 590 + ], + "type": "text", + "content": "Self-Supervised Learning (SSL). To effectively use datasets, many works [47, 64, 51, 49, 55] use self-supervised learning to assist VLP. VideoBERT [47] tokenizes video frames by hierarchical vector quantization, and then performs SSL by predicting the masked visual tokens. ActBERT [64] predicts the action and object words of masked video tokens. VLM [51] masks either all video tokens or all text tokens, and then uses tokens from one modality to recover masked tokens. However, the existing methods may damage the visual tokens, which introduces noise into cross-attention between visual and textual tokens [32]. For better visual representation learning and avoiding damaging visual tokens, we introduce the MVC loss which maximizes the mutual information between multiscale global and local representations, and improves the multi-level alignment between visual and textual features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": "Hierarchical Interaction. As far as we know, there is no related work about hierarchical interaction in VLP. The most related work is VinVL [59] in ILP. VinVL [59] uses an object detector to extract different sizes of objects as visual features to do cross-attention with textual features. The method of VinVL is complicated and the accuracy of its visual features is limited by the object detector. However, HiVLP makes different scales of visual features interact with textual features without the need of an object detector, which is more effective." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13757" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 359, + 371 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 359, + 371 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 359, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 359, + 371 + ], + "type": "image", + "image_path": "e229b6afe7cd86bdc79de7941233721cd18b1ef93088ee505a3a320e80ef2360.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 545, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 545, + 475 + ], + "type": "text", + "content": "Figure 2. (a) Overview of HiVLP which consists of a vision encoder, a text encoder, a multi-modal video-text encoder and a multi-modal video-text decoder. Each image (or video clip) is transformed into two augmentations as the inputs to the vision encoder and the momentum vision encoder, respectively. The hierarchical visual feature group (HVFG) has scales " + }, + { + "bbox": [ + 46, + 376, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\left\\{\\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}" + }, + { + "bbox": [ + 46, + 376, + 545, + 475 + ], + "type": "text", + "content": ". The multi-level visual contrastive (MVC) loss is for multi-level SSL. Besides, HiVLP also uses the visual features of " + }, + { + "bbox": [ + 46, + 376, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\frac{1}{32}" + }, + { + "bbox": [ + 46, + 376, + 545, + 475 + ], + "type": "text", + "content": " scale from the vision encoder and the output of the text encoder to optimize the vision-text contrastive (VTC) loss. Additionally, the multi-modal video-text encoder is trained with a vision-text matching (VTM) loss for the matching of video-/image-text pairs. The multi-modal video-text decoder is trained with a language modeling (LM) loss for video captioning. (b) The light FPN [25] is used for gradually upsampling and fusing the output features of the Swin Transformer. (c) The temporal block consists of a temporal self-attention layer, a SW-MSA layer [29] and a feed forward layer. We only replace the last block of the Swin Transformer with the temporal block for less parameters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 361, + 72, + 539, + 257 + ], + "blocks": [ + { + "bbox": [ + 361, + 72, + 539, + 257 + ], + "lines": [ + { + "bbox": [ + 361, + 72, + 539, + 257 + ], + "spans": [ + { + "bbox": [ + 361, + 72, + 539, + 257 + ], + "type": "image", + "image_path": "69b14ea2ca6e4e54ff382d171feef75dbe8defcf9214da9d9d49a0ff31f2388f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 359, + 258, + 539, + 371 + ], + "blocks": [ + { + "bbox": [ + 359, + 258, + 539, + 371 + ], + "lines": [ + { + "bbox": [ + 359, + 258, + 539, + 371 + ], + "spans": [ + { + "bbox": [ + 359, + 258, + 539, + 371 + ], + "type": "image", + "image_path": "08f4a6bd3543f38c924c253a3858bfd6a7315661fcc64e7bb0f16cc563431223.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 484, + 115, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 115, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 115, + 498 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 505, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 541 + ], + "type": "text", + "content": "We firstly describe the detailed architecture of HiVLP. Then we present the hierarchical interaction HVFG. Finally, the pre-training objectives are given." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 129, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 129, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 129, + 562 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "As illustrated in Figure 2(a), HiVLP consists of a vision encoder, a momentum vision encoder, a text encoder, a multi-modal video-text encoder, and a multi-modal videotext decoder. The vision encoder and the text encoder extract image and text features respectively, and do not need to interact with each other for a fast approximate nearest neighbor search in inference [33]. The multi-modal videotext encoder re-ranks the top-k video-text pairs by an additional MLP head that predicts whether they are matched or not. The multi-modal video-text decoder is used to perform the video-language generation (e.g., video captioning). The momentum vision encoder has the same archi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 485, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 520 + ], + "type": "text", + "content": "tecture as the vision encoder, and the weights are updated by a momentum-based moving average strategy as in MoCo [15]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": "The main component of the vision encoder is a hierarchical backbone Swin Transformer [29], followed by a light feature pyramid network (FPN) [25]. Besides, we add a temporal self-attention layer in the last block of Swin Transformer to capture the temporal information, as shown in Figure 2(c). In Figure 2(b), the light FPN uses the output features of " + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\frac{1}{32}" + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": " scales from Swin Transformer as inputs, and then gradually upsamples those features for feature fusion. Note that to reduce computation overhead, we do not use " + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": " scale of visual features in HVFG." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "The text encoder's architecture is the same as BERT [10], which consists of 12 transformer layers. The multi-modal video-text encoder has 12 cross-attention (CA) blocks each with a bi-directional self-attention layer. The multi-modal video-text decoder consists of 12 CA blocks each with a casual self-attention layer." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13758" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 321, + 190 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 321, + 190 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 321, + 190 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 321, + 190 + ], + "type": "image", + "image_path": "a4c078d814f3a0ec3e3c38a2023b63f09a6c0f4cfbb18923540c42a9668571b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 195, + 477, + 206 + ], + "lines": [ + { + "bbox": [ + 115, + 195, + 477, + 206 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 477, + 206 + ], + "type": "text", + "content": "Figure 3. (a) Multi-level global-to-local contrastive loss. (b) Generation of global and local features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 71, + 504, + 191 + ], + "blocks": [ + { + "bbox": [ + 325, + 71, + 504, + 191 + ], + "lines": [ + { + "bbox": [ + 325, + 71, + 504, + 191 + ], + "spans": [ + { + "bbox": [ + 325, + 71, + 504, + 191 + ], + "type": "image", + "image_path": "0f67c0f87f724134ec31964016c157eeb897bd9f7c8760fd73335ee0d94686d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 216, + 185, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 185, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 185, + 228 + ], + "type": "text", + "content": "3.2. Hierarchical Interaction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "text", + "content": "Existing works often inject single-scale visual features to CA blocks to interact with textual features. For example, ALBEF [21] and BLIP [20] use visual features of " + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "text", + "content": " scale, which fail to do fine-grained cross-attention. In contrast, [33] adopts features of " + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "text", + "content": " scale, but these over detailed features without high-level semantics. In our work, we propose a hierarchical interaction mechanism between visual and textual features via injecting HVFG into the CA blocks as shown in Figure 2(a). HVFG includes visual features of three scales " + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\left\\{\\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}" + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "text", + "content": ". The visual features of " + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "text", + "content": " scale contain fine-grained information, and those of " + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\frac{1}{32}" + }, + { + "bbox": [ + 46, + 235, + 290, + 402 + ], + "type": "text", + "content": " scale carry high-level semantics. Therefore, HVFG is able to obtain both global and fine-grained representations to boost the interaction between visual and textual features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 410, + 183, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 183, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 183, + 423 + ], + "type": "text", + "content": "3.3. Pre-Training Objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": "The input of the vision encoder is an image or video clip " + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "inline_equation", + "content": "X \\in R^{B \\times M \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": ", which consists of " + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": " frames with the resolution " + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "inline_equation", + "content": "M = 1" + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": " for images and " + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": " is the batch size. The input to the text encoder is a batch of tokenized sequences of words " + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "inline_equation", + "content": "T \\in N^{B \\times L}" + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 429, + 287, + 501 + ], + "type": "text", + "content": " is the max length of a caption." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "text", + "content": "In our method, an input video clip " + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "text", + "content": " is split into non-overlapping patches " + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{BMN}, x_{i} \\in R^{3 \\times P \\times P}" + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "text", + "content": " is input path size and " + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "inline_equation", + "content": "N = \\frac{H \\times W}{P \\times P}" + }, + { + "bbox": [ + 47, + 501, + 287, + 561 + ], + "type": "text", + "content": ". The patches are tokenized by a linear embedding layer which is followed by a linear embedding layer of Swin Transformer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 561, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 288, + 657 + ], + "type": "text", + "content": "Vision-Text Contrastive (VTC) Loss. VTC aims to pull positive pairs of vision and language representations together and push negative pairs far away. Like ALBEF [21] using a momentum distillation for vision-language pretraining, we introduce two queues to store the most recent " + }, + { + "bbox": [ + 46, + 561, + 288, + 657 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 561, + 288, + 657 + ], + "type": "text", + "content": " vision-text representations pairs from momentum encoders. Formally, the video-to-text contrastive loss is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 673, + 290, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 290, + 714 + ], + "type": "interline_equation", + "content": "L _ {T 2 V} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {- e x p \\left(s \\left(f _ {i} , t _ {i} ^ {+}\\right) / \\rho\\right)}{e x p \\left(s \\left(f _ {i} , t _ {i} ^ {+}\\right) / \\rho\\right) + \\sum_ {j = 1} ^ {K} e x p \\left(s \\left(f _ {i} , t _ {j} ^ {-}\\right) / \\rho\\right)}, \\tag {1}", + "image_path": "0df8a16a590fa78e8dadf72feb3c8d272ead82354fe52368610493b31a79ee35.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " is the temperature, " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "t_i^+" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "t_j^-" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " are the text features of the positive and negative text samples for the " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": "-th vision input " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "X_i" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " respectively, and " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " is the visual feature of " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "\\frac{1}{32}" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " scale from " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "X_i" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "s(f_i, t_i^+)" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " denotes the cosine similarity between " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "t_i^+" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " that are matched, and " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "s(f_i, t_j^-)" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " denotes the similarity between " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "inline_equation", + "content": "t_j^-" + }, + { + "bbox": [ + 304, + 216, + 547, + 302 + ], + "type": "text", + "content": " that are not matched. Symmetrically, the text-to-video contrastive loss is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 321, + 550, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 550, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 550, + 359 + ], + "type": "interline_equation", + "content": "L _ {V 2 T} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {- \\exp \\left(s \\left(t _ {i} , f _ {i} ^ {+}\\right) / \\rho\\right)}{\\exp \\left(s \\left(t _ {i} , f _ {i} ^ {+}\\right) / \\rho\\right) + \\sum_ {j = 1} ^ {K} \\exp \\left(s \\left(t _ {i} , f _ {j} ^ {-}\\right) / \\rho\\right)}, \\tag {2}", + "image_path": "84d9bda78e846084eaad38e0d7257d02e230b2b9feb3304987f9d0741e17bcb2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "inline_equation", + "content": "f_{i}^{+}" + }, + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "inline_equation", + "content": "f_{j}^{-}" + }, + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "text", + "content": " are positive and negative visual features of the " + }, + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 360, + 545, + 384 + ], + "type": "text", + "content": "-th text feature, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 384, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 407 + ], + "type": "text", + "content": "The sum of " + }, + { + "bbox": [ + 305, + 384, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_{V2T}" + }, + { + "bbox": [ + 305, + 384, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 384, + 545, + 407 + ], + "type": "inline_equation", + "content": "L_{T2V}" + }, + { + "bbox": [ + 305, + 384, + 545, + 407 + ], + "type": "text", + "content": " serve as the vision-text contrastive loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 417, + 545, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 417, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 370, + 417, + 545, + 439 + ], + "type": "interline_equation", + "content": "L _ {V T C} = \\frac {1}{2} \\left(L _ {V 2 T} + L _ {T 2 V}\\right). \\tag {3}", + "image_path": "8fb9897da5b95ea8e28dc2cfb53b50240ffc018f68cb89822ad6af36f60371b8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": "Multi-level Vision Contrastive (MVC) Loss. As shown in Figure 3(a), MVC aims to pull the global representation of a scale closer to those local patch representations from the same vision input in different views of a scale. The light FPN has three scales " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "L_{3}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": " (see Figure 2(b) and Figure 3(a)). The global feature " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "f_{ij}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": " of scale " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": "-th visual input is obtained by the average of all the patch features at " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "L_{j}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": " (Figure 3(b)). The local features " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "f_{ijk}, k = 1, 2, 3, \\dots, s_{j}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": " are generated as shown in Figure 3(b) where " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "f_{ijk}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": " is the average feature pooling result of a local window such as " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "s_{j}" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": " is the number of the local averaged features at " + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "inline_equation", + "content": "L_{j}'" + }, + { + "bbox": [ + 304, + 441, + 547, + 588 + ], + "type": "text", + "content": ". Finally, the MVC loss is formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 603, + 545, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 545, + 644 + ], + "type": "interline_equation", + "content": "L _ {M V C} ^ {i j} = \\frac {\\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j} ^ {+}\\right) / \\rho\\right)}{\\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j} ^ {+}\\right) / \\rho\\right) + \\sum_ {k = 1} ^ {(B - 1) S _ {j}} \\exp \\left(s \\left(f _ {i j} , \\hat {f} _ {i j k} ^ {-}\\right) / \\rho\\right)}, \\tag {4}", + "image_path": "a6802f682d9464f9a3ec9d1faded6106a1d27d0936bc0e56373a8b05a52facec.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 652, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 652, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 354, + 652, + 545, + 683 + ], + "type": "interline_equation", + "content": "L _ {M V C} = - \\frac {1}{3 B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {3} - l o g L _ {M V C} ^ {i j}, \\tag {5}", + "image_path": "b439fdf4c6165ff60e87768a32213005d120e94cafd1b50dfc8124c810cfa099.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}_{ij}" + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}_{ijk}" + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "text", + "content": " correspond to " + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{ij}" + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{ijk}" + }, + { + "bbox": [ + 305, + 688, + 545, + 714 + ], + "type": "text", + "content": ", respectively, but from the other augmentation of the visual input." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13759" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "Vision-Text Matching (VTM) Loss. Through a MLP layer, VTM is used to judge whether a video-/image-text pair is matched or not with the cross entropy function. There are " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " positive pairs from the batch, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "2B" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " negative pairs are obtained according to the hard negative mining strategy in [21]. This loss is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 164, + 287, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 287, + 206 + ], + "type": "interline_equation", + "content": "L _ {V T M} = - \\frac {1}{3 B} \\sum_ {i = 1} ^ {3 B} \\left[ y _ {i} \\log \\left(p _ {i}\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - p _ {i}\\right) \\right], \\tag {6}", + "image_path": "9c7adffb91f16ed1427d16519cc5e6733b8ac7e1c3d5fc6ebfeee569b724b316.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "text", + "content": " is the ground-truth, and " + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "text", + "content": " is the output of the MLP layer of the " + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 207, + 287, + 231 + ], + "type": "text", + "content": "-th pair." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 281 + ], + "type": "text", + "content": "Language Modeling (LM) Loss. We inject the hierarchical visual representations into the multi-modal video-text decoder to generate captions. Let " + }, + { + "bbox": [ + 46, + 232, + 287, + 281 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 232, + 287, + 281 + ], + "type": "text", + "content": "-th input text be " + }, + { + "bbox": [ + 46, + 232, + 287, + 281 + ], + "type": "inline_equation", + "content": "T_{i} = [t_{i}^{1}, t_{i}^{2}, \\dots, t_{i}^{L}]" + }, + { + "bbox": [ + 46, + 232, + 287, + 281 + ], + "type": "text", + "content": ". This loss is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 300, + 286, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 300, + 286, + 343 + ], + "spans": [ + { + "bbox": [ + 49, + 300, + 286, + 343 + ], + "type": "interline_equation", + "content": "L _ {L M} = - \\frac {1}{B L} \\sum_ {i = 1} ^ {B} \\sum_ {l = 1} ^ {L} \\log \\left(p \\left(t _ {i} ^ {l} \\mid t _ {i} ^ {l - 1}, \\dots , t _ {i} ^ {1}, f _ {i 1}, f _ {i 2}, f _ {i 3}\\right)\\right), \\tag {7}", + "image_path": "302bd86f11c141a2bf1b1a1a7a399954d40a9b562573d4d61677638d2793b0e0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "inline_equation", + "content": "p(t_i^l | t_i^{l-1}, \\dots, t_i^1, f_{i1}, f_{i2}, f_{i3})" + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "text", + "content": " is the output probability of the multi-modal video-text decoder for the " + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "text", + "content": "-th token " + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "inline_equation", + "content": "t_i^l" + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "text", + "content": " given the previously fed tokens " + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "inline_equation", + "content": "t_i^{l-1}, \\dots, t_i^1" + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "text", + "content": " and the multilevel visual features " + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "inline_equation", + "content": "f_{i1}, f_{i2}, f_{i3}" + }, + { + "bbox": [ + 46, + 342, + 286, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 392, + 282, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 282, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 282, + 403 + ], + "type": "text", + "content": "Total Loss. Finally, we have the total loss to train HiVLP:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 426, + 287, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 426, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 58, + 426, + 287, + 439 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = \\alpha L _ {V T C} + \\beta L _ {M V C} + \\gamma L _ {V T M} + \\delta L _ {L M}, \\tag {8}", + "image_path": "ce0209142c268ad460c24f60afa41b2c5b85bfe1df66acd2d4f04605ab7385f1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 449, + 287, + 473 + ], + "type": "text", + "content": " are the weight coefficients of the losses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 486, + 128, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 128, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 128, + 499 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 554 + ], + "type": "text", + "content": "In this section, we first describe the pre-training and downstream datasets, and the implementation details. Then we show the experimental and comparison results, and finally ablate our model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 563, + 110, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 110, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 110, + 574 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Pre-Training Datasets. Following FiT [4], we involve two popular datasets (CC-3M [46] and WebVid-2M [4]) for pre-training HiVLP. CC-3M consists of 3M image-text pairs and WebVid-2M includes 2.5M video-text pairs, in total 5.5M image-/video-text pairs. Besides, we use LAION100M, D-14M, and WebVid-2M (total 116.5M image-/video-text pairs) as the larger-scale dataset to train our model, resulting in HiVLP*. LAION100M is a subset of LAION [45]. D-14M is a combined dataset of image-text pairs from COCO [26], Visual Genome [17], CC3M [46], CC12M [6], and SBU captions [35]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": "Downstream Task Datasets. There are four downstream datasets used in our work for video retrieval. (i) MSR-VTT [53] consists of 10K YouTube videos with 200K human-annotated descriptions. Similar to previous methods [4], we use the split of 9K for training and 1K for testing. (ii) MSVD [7] is a smaller dataset with 1970 videos and 78800 sentences, about 40 descriptions per video. Like [4], we use the split of 1200, 100, and 670 videos for training, validation and testing, respectively. (iii) DiDeMo [3] has 10k videos, each of which is described by multiple sentences, resulting in 40K sentences. For a fair comparison, we follow the setting in [4], where all descriptions for one video are concatenated into a single sentence. (iv) LSMDC [43] contains 128K clips. Like [4], we use 7408 clips for validation and 1000 clips for testing. For video captioning, we evaluate our model on MSR-VTT and MSVD. Following the split of [30], on MSR-VTT, we use " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "6.5K" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " training videos and " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "2.9K" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " testing videos; on MSVD, we use " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "1.2K" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " training videos and 670 testing videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 310, + 440, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 440, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 440, + 323 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": "Our model HiVLP is pre-trained on 16 GPUs (32G memory). Our vision encoder is initialized by Swin-B [29] pretrained on ImageNet-21k [9]. The light FPN is randomly initialized. The text encoder, multi-modal video-text encoder, and multi-modal video-text decoder are BERTs initialized from BERTbase [10]. HiVLP is pre-trained for 25 epochs, the first 20 epochs only with image-text pairs with a batch size of 30 and the last 5 epochs with both image-text and video-text pairs. Each video clip consists of 4 frames. We use AdamW optimizer with a weight decay of 0.02, and the learning rate is initialized as " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": " and is warmed up to " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": " after 3,000 training iterations. We then decrease the learning rate by the cosine decay strategy to " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": ". For the hyper-parameters in Equation 8, we set " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": " to 1.0 and " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": " to 0.001. The input resolution is " + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 329, + 545, + 557 + ], + "type": "text", + "content": " with data augmentations used in [21]. And we use 160 GPUs (16G memory) to train HiVLP* with a batch size of 18 per GPU. For speeding up the training of HiVLP*, we sample 1 frame from each clip." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": "For the downstream task of text-video retrieval, we sample 4 frames per video for training. Since the multi-modal video-text encoder filters top-" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": " candidates during inference (" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": " is set to 128). We sample 8 frames per video for testing on MSR-VTT, MSVD, and LSMDC. Because the videos in DiDeMo are longer, we sample 10 frames for testing from it. For the downstream task of video captioning, we evaluate our model on MSR-VTT and MSVD, with 8 random frames per video for training and 16 frames per video for testing. For all downstream tasks, the initial learning rate is set to " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-6}" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ", and the weight decay, the batch size, and the total number of epochs are set to 0.05, 64, and 10, respectively." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13760" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 69, + 512, + 363 + ], + "blocks": [ + { + "bbox": [ + 82, + 69, + 512, + 363 + ], + "lines": [ + { + "bbox": [ + 82, + 69, + 512, + 363 + ], + "spans": [ + { + "bbox": [ + 82, + 69, + 512, + 363 + ], + "type": "table", + "html": "
MethodPre-Training Datasets#PairsR@1↑R@5↑R@10↑MedR↓MR↑
Zero-Shot
SupportSet [38]HowTo100M100M8.723.031.131.020.9
HD-VILA [54]HD-VILA-100M100M14.431.641.617.529.2
FiT [4]CC3M, WebVid-2M5.5M18.739.551.610.036.6
ALPRO [19]CC3M, WebVid-2M5.5M24.144.755.48.041.4
CLIP [40]WIT400M400M30.654.464.34.049.8
Florence [58]FLD-900M900M37.663.872.6-58
BLIP [20]L-115M, D-14M239M43.365.674.72.061.2
HiVLPCC3M, WebVid-2M5.5M26.447.355.77.043.1
HiVLP*L-100M, D-14M, WebVid-2M116.5M43.566.476.42.062.1
Fine-Tuning
ActBERT [64]HowTo100M100M16.342.856.910.038.7
HERO [22]HowTo100M100M16.843.457.7-39.3
NoiseEstimation [2]HowTo100M100M17.441.653.68.037.5
UniVL [30]HowTo100M100M21.249.663.16.044.6
ClipBERT [18]HowTo100M100M22.046.859.96.042.9
AVLnet [44]HowTo100M100M27.155.666.64.049.8
MMT [12]HowTo100M100M26.657.169.64.051.1
SupportSet [38]HowTo100M100M30.158.569.33.052.6
FiT [4]CC3M, WebVid-2M5.5M31.059.570.53.053.7
ALPRO [19]CC3M, WebVid-2M5.5M33.960.773.23.055.9
HD-VILA [54]HD-VILA-100M100M35.065.277.23.059.1
CLIP4Clip [31]WIT400M400M44.571.481.62.065.8
HiVLPCC3M, WebVid-2M5.5M41.165.975.92.061.0
HiVLP*L-100M, D-14M, WebVid-2M116.5M50.976.483.61.070.3
", + "image_path": "99022c17bc701f250f17b0bfa858711b1ce154a7e6b7c63dbd749786e853b1ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 408, + 276, + 491 + ], + "blocks": [ + { + "bbox": [ + 46, + 369, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 547, + 392 + ], + "type": "text", + "content": "Table 1. Comparisons with SOTA text-to-video retrieval methods on MSR-VTT. R@K: Recall@K; MedR: Median Rank; MR: Mean Rank; L-115M: LAION115M; L-100M: LAION100M." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 408, + 276, + 491 + ], + "lines": [ + { + "bbox": [ + 60, + 408, + 276, + 491 + ], + "spans": [ + { + "bbox": [ + 60, + 408, + 276, + 491 + ], + "type": "table", + "html": "
MethodR@1↑R@5↑R@10↑MR↑
CE [28]19.849.063.844.2
Support Set [38]28.460.072.953.8
FiT [4]33.764.776.357.3
CLIP4Clip [31]46.276.184.668.9
HiVLP39.168.177.261.5
HiVLP*50.278.985.871.6
", + "image_path": "c91939e5898ad2903f9c5a6aa8a1b856f256736e5fb14c3bb98759e305457ad2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 520, + 273, + 601 + ], + "blocks": [ + { + "bbox": [ + 71, + 498, + 261, + 509 + ], + "lines": [ + { + "bbox": [ + 71, + 498, + 261, + 509 + ], + "spans": [ + { + "bbox": [ + 71, + 498, + 261, + 509 + ], + "type": "text", + "content": "Table 2. Text-to-video results on the MSVD dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 520, + 273, + 601 + ], + "lines": [ + { + "bbox": [ + 61, + 520, + 273, + 601 + ], + "spans": [ + { + "bbox": [ + 61, + 520, + 273, + 601 + ], + "type": "table", + "html": "
MethodR@1↑R@5↑R@10↑MR↑
ClipBERT [18]20.444.556.740.5
FiT [4]34.665.074.758.1
ALPRO [54]35.967.578.860.7
CLIP4Clip [31]43.470.280.664.7
HiVLP40.568.777.162.1
HiVLP*48.173.882.568.1
", + "image_path": "130209c0f7dd390c1eb665ea6abefd0d19daad8b0d74446663c1d5601fce57a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 631, + 277, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 277, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 277, + 644 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 651, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 700 + ], + "type": "text", + "content": "In this section, we compare HiVLP with state-of-the-art (SOTA) methods on three popular video-language downstream tasks (text-video retrieval, video-text retrieval, and video captioning)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Text-Video Retrieval. We use MSR-VTT, MSVD," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 409, + 533, + 491 + ], + "blocks": [ + { + "bbox": [ + 69, + 609, + 265, + 619 + ], + "lines": [ + { + "bbox": [ + 69, + 609, + 265, + 619 + ], + "spans": [ + { + "bbox": [ + 69, + 609, + 265, + 619 + ], + "type": "text", + "content": "Table 3. Text-to-video results on the DiDeMo dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 409, + 533, + 491 + ], + "lines": [ + { + "bbox": [ + 319, + 409, + 533, + 491 + ], + "spans": [ + { + "bbox": [ + 319, + 409, + 533, + 491 + ], + "type": "table", + "html": "
MethodR@1↑R@5↑R@10↑MR↑
MMT [12]12.929.238.827.0
FiT [4]15.030.839.828.5
MDMMT [11]18.838.547.935.1
CLIP4Clip [31]21.641.849.837.7
HiVLP20.438.848.435.9
HiVLP*24.844.154.641.2
", + "image_path": "3959813fee75483c7bd345b8d0eeb3be58cd27f8bea84f259f419b566eefdedb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 498, + 523, + 509 + ], + "lines": [ + { + "bbox": [ + 328, + 498, + 523, + 509 + ], + "spans": [ + { + "bbox": [ + 328, + 498, + 523, + 509 + ], + "type": "text", + "content": "Table 4. Text-to-video results on the LSMDC dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 520, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 687 + ], + "type": "text", + "content": "LSMDC, and DiDeMo datasets to evaluate text-video retrieval. We report zero-shot and fine-tuning results (Table 1) on MSR-VTT. For zero-shot retrieval, pre-trained with the same small amount of data, HiVLP outperforms FiT by a large margin (7.4% on R@1). CLIP, Florence, BLIP, and HiVLP* use large-scale datasets for pre-training, with the numbers of pairs 400M, 900M, 239M, and 116.5M, respectively. BLIP performs best among previous works. However, even with much fewer training pairs (116.5M vs. 239M), our HiVLP* outperforms BLIP on all the metrics except obtaining the same MedR. Note that BLIP uses both filtered and synthetic captions in LAION115M and D-14M, while we only use the filtered captions in LAION100M and D-14M." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "For fine-tuning comparison, HiVLP/HiVLP* is finetuned with the VTC and VTM losses. In Table 1, we see that" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13761" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 514, + 140 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 514, + 140 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 514, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 514, + 140 + ], + "type": "table", + "html": "
MethodMSR-VTTMSVDLDMDC
R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP4Clip-meanP+ [31]43.170.581.256.679.784.320.639.447.5
CLIP4Clip-seqTransf+ [31]42.770.980.662.087.392.620.839.048.6
X-Pool+ [13]44.473.384.066.490.094.222.742.651.2
HiVLP*51.575.983.167.289.595.725.645.653.8
", + "image_path": "bdd3bf2a915731a5fe9e5dbef156371cf6fb9e67b8cf77ddc0d220cca6e8c676.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 101, + 181, + 493, + 306 + ], + "blocks": [ + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "type": "text", + "content": "Table 5. Comparison with SOTA methods on MSR-VTT, MSVD, and LSMDC for video-text retrieval. " + }, + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "type": "inline_equation", + "content": "{}^{ + }" + }, + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "type": "text", + "content": " indicates using both CLIP's image and text encoders as its encoders." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 101, + 181, + 493, + 306 + ], + "lines": [ + { + "bbox": [ + 101, + 181, + 493, + 306 + ], + "spans": [ + { + "bbox": [ + 101, + 181, + 493, + 306 + ], + "type": "table", + "html": "
Method#Test FramesMSVDMSR-VTT
B4↑M↑R↑C↑B4↑M↑R↑C↑
SibNet [27]3054.234.871.788.240.927.560.247.5
SAAT [62]2846.533.569.481.039.927.761.251.0
STG-KD [36]1652.236.973.993.040.528.360.947.1
PMI-CAP [8]3254.636.4-95.142.128.7-49.4
ORG-TRL [61]2854.336.473.995.243.628.862.150.9
OpenBook [60]28----33.923.750.252.9
SwinBERT [24]6466.342.480.9149.445.430.664.155.9
HiVLP1667.145.381.8144.547.431.264.662.3
HiVLP*1668.345.182.0151.649.232.465.967.8
", + "image_path": "7f9f417cedea22ceb6be247fe2c4f3269b053b6213e3eb6afb014cededbad08c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 313, + 457, + 326 + ], + "lines": [ + { + "bbox": [ + 131, + 313, + 457, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 457, + 326 + ], + "type": "text", + "content": "Table 6. Comparison with SOTA methods on MSVD and MSR-VTT for video captioning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 337, + 276, + 482 + ], + "blocks": [ + { + "bbox": [ + 59, + 337, + 276, + 482 + ], + "lines": [ + { + "bbox": [ + 59, + 337, + 276, + 482 + ], + "spans": [ + { + "bbox": [ + 59, + 337, + 276, + 482 + ], + "type": "image", + "image_path": "e48530dfecfe2719643d9c7563a7bc3b607f04c3cf6d4f5649c2cf5486b304e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 486, + 254, + 498 + ], + "lines": [ + { + "bbox": [ + 79, + 486, + 254, + 498 + ], + "spans": [ + { + "bbox": [ + 79, + 486, + 254, + 498 + ], + "type": "text", + "content": "Figure 4. Diversity vs. amount of training pairs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 592 + ], + "type": "text", + "content": "HiVLP uses the same pre-training datasets as FiT but again outperforms it by a large margin (10.1% on Rank@1). With only about 29% amount of pre-training pairs in CLIP4Clip, HiVLP* exceeds CLIP4Clip on all the metrics significantly. Tables 2, 3, and 4 show the comparison on another three datasets, where HiVLP* outperforms the other methods by a large margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 593, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 640 + ], + "type": "text", + "content": "Video-Text Retrieval. In Table 5, we use HiVLP* to compare with the SOTA methods which are built on CLIP. Our method can beat these CLIP-based methods with fewer pretraining pairs (116.5M vs. 400M)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "text", + "content": "Video Captioning. We also evaluate our model on MSR-VTT and MSVD for video captioning. Four popular metrics BLEU4 [37] (B4), METEOR [5] (M), ROUGE-L [23] (R), and CIDEr [48] (C) are employed. We fine-tune our model with the LM loss for the video-language generation task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "content": "As shown in Table 6, HiVLP works best on all the met" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 335, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 359 + ], + "type": "text", + "content": "rics except one. Note that SwinBERT adopts 64 frames for testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 369, + 400, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 400, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 400, + 382 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 388, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 483 + ], + "type": "text", + "content": "We randomly choose 1/10 image-text pairs from CC3M, resulting in " + }, + { + "bbox": [ + 304, + 388, + 545, + 483 + ], + "type": "inline_equation", + "content": "300\\mathrm{k}" + }, + { + "bbox": [ + 304, + 388, + 545, + 483 + ], + "type": "text", + "content": " image-text pairs to do ablation study. We surprisingly find that our model pre-trained with only " + }, + { + "bbox": [ + 304, + 388, + 545, + 483 + ], + "type": "inline_equation", + "content": "300\\mathrm{k}" + }, + { + "bbox": [ + 304, + 388, + 545, + 483 + ], + "type": "text", + "content": " image-text pairs can beat the models SupportSet and HDVILA pre-trained respectively on Howto100M and HDVILA-100M, when they are transferred to the MSR-VTT dataset for text-to-video zero-shot retrieval (45.3 vs. 31.1 vs. 41.6 on R@10)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 485, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 616 + ], + "type": "text", + "content": "Image-Text Pairs to Enhance VLP. As shown in Figure 4, when the number of image-text pairs is increased from 3M to 114M but with the same set of video-text pairs, the Rank@1 on the MSR-VTT dataset raises a lot (from 26.4 to 43.5); when only using much fewer image-text pairs of CC3M than video-text pairs of HD-VILP (0.3M vs. 100M), the Rank@1 performance is much better (17.4 vs. 14.4). It reveals that image-text pairs can bring more diversity than the same amount or more video-text pairs, and more image-text pairs benefit a lot. We believe this is a significant way to enhance the performance of VLP models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "MVC Loss Design. We use the HiVLP without the MVC loss as the baseline model. As shown in Table 7, the visual SSL both G2G and G2L-1 are able to improve the performance of the baseline. Using the G2L-1 contrastive loss is better than the G2G contrastive loss, because G2L-1 can dig out local information while G2G cannot. And using more number of different scales of visual features in the MVC loss (i.e., G2L-2 and G2L-3) can further boost the align" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13762" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 83, + 534, + 281 + ], + "blocks": [ + { + "bbox": [ + 193, + 72, + 372, + 83 + ], + "lines": [ + { + "bbox": [ + 193, + 72, + 372, + 83 + ], + "spans": [ + { + "bbox": [ + 193, + 72, + 372, + 83 + ], + "type": "text", + "content": "Caption: A jackal is walking around in a field" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 83, + 534, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 83, + 534, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 83, + 534, + 281 + ], + "type": "image", + "image_path": "292bdb87cbf0f444070b3195f1a7b878f5d40ba1c2ebf4a1ec5581a251dd0f5a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 283, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 547, + 316 + ], + "type": "text", + "content": "Figure 5. Grad-CAM visualization on the cross-attention maps of the different-scale visual features injected to the CA blocks. We extract the attention maps in the 8th layer of the multi-modal video-text encoder. The word \"jackal\" in the caption is the attention target in the video. (Best viewed on screen.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 333, + 284, + 409 + ], + "blocks": [ + { + "bbox": [ + 51, + 333, + 284, + 409 + ], + "lines": [ + { + "bbox": [ + 51, + 333, + 284, + 409 + ], + "spans": [ + { + "bbox": [ + 51, + 333, + 284, + 409 + ], + "type": "table", + "html": "
ModelMVC lossR@10↑MR↑
Baseline-43.531.1
+G2G(L1,L1′′)44.231.4
+G2L-1(L1,L1′)44.931.8
+G2L-2(L1,L1′), (L2,L2′)45.232.2
+G2L-3(L1,L1′), (L2,L2′), (L3,L3′)45.332.5
", + "image_path": "672d38c9f19fdbbe81ae2614583aa765cec5e210ff78b3d70c17d414c7ce2c57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 539 + ], + "type": "text", + "content": "ment between the multi-scale visual features and the textual features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 542, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 662 + ], + "type": "text", + "content": "Impact of HVFG. Table 8 compares our hierarchical vision features with only single-scale ones. HiVLP-1/8 uses large single scales of visual features to achieve the best performance among HiVLP-1/32, HiVLP-1/16, and HiVLP-1/8. However, HiVLP-1/8 does not efficiently use the deeper and small scales of visual features with higher-level semantics. To solve this problem, HiVLP effectively uses both global and fine-grained visual features to interact with textual features, resulting in much better performance than HiVLP-1/8 (32.5 vs. 30.5 on MR)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Visualization. As Figure 5 shown, HiVLP can capture fine-grained information better than HiVLP-1/8. The attention maps of HiVLP-1/16 look random due to its low-resolution feature maps." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 334, + 548, + 393 + ], + "blocks": [ + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": "Table 7. Impact of the MVC loss for zero-shot transfer to MSR-VTT. Note that " + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "inline_equation", + "content": "(L_{j}, L_{j}^{\\prime})" + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": " scale of global-to-local (G2L) contrastive loss, with " + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "inline_equation", + "content": "L_{j}, L_{j}^{\\prime}, j = 1, 2, 3" + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": " indicated in Figure 3(a). " + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "inline_equation", + "content": "L_{1}^{\\prime \\prime}" + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": " denotes replacing the dense projection " + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "inline_equation", + "content": "L_{1}^{\\prime}" + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": " by the global projection as indicated in the upper part of Figure 3(b). And " + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "inline_equation", + "content": "(L_{1}, L_{1}^{\\prime \\prime})" + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "inline_equation", + "content": "j = 1" + }, + { + "bbox": [ + 46, + 416, + 287, + 499 + ], + "type": "text", + "content": " global-to-global (G2G) contrastive loss. MR: Mean Rank." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 334, + 548, + 393 + ], + "lines": [ + { + "bbox": [ + 306, + 334, + 548, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 548, + 393 + ], + "type": "table", + "html": "
Modelscales of featuresR@1↑R@10↑MR↑
HiVLP-1/32{1/32, 1/32, 1/32}14.043.430.0
HiVLP-1/16{1/16, 1/16, 1/16}14.343.630.2
HiVLP-1/8{1/8, 1/8, 1/8}14.843.730.5
HiVLP{1/32, 1/16, 1/8}17.445.332.5
", + "image_path": "4273e4e3e186f500e37c0d0192f1fea96d498ca04a3476e06a4d5114e0e50652.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 400, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 305, + 400, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 547, + 434 + ], + "type": "text", + "content": "Table 8. Ablation study of different vision features injecting to CA blocks for text-to-video retrieval zero-shot transfer to MSR-VTT." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 378, + 455 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 464, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 643 + ], + "type": "text", + "content": "We have presented HiVLP, a novel hierarchical interactive video-language pre-training framework. Different from previous methods that input single-scale visual features to cross-attention blocks, HiVLP injects a hierarchical vision feature group (HVFG) to effectively use both global and fine-grained visual features for interaction with textual features. Additionally, our HiVLP is pre-trained with multilevel self-supervised learning that can further improve the model performance. We also reveal that VLP models benefit a lot from the diversity of image-text pairs. Extensive experimental results of downstream tasks (text-video retrieval, video-text retrieval, and video captioning) on 4 popular benchmark datasets show that HiVLP is able to achieve better performance than previous SOTA methods overall by a large margin." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We gratefully acknowledge the support of MindSpore [1], CANN (Compute Architecture for Neural Networks) and Ascend AI Processor used for this research." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13763" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 169, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 169, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 169, + 102 + ], + "type": "text", + "content": "[1] https://www.mindspore.cn/." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 136 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 136 + ], + "type": "text", + "content": "[2] Elad Amrani, Rami Ben-Ari, Daniel Rotman, and Alex Bronstein. Noise estimation using density estimation for self-supervised multimodal learning. In AAAI, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 170 + ], + "type": "text", + "content": "[3] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 204 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 206, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 287, + 239 + ], + "type": "text", + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL, 2005." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 240, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 287, + 282 + ], + "type": "text", + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 285, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 306 + ], + "type": "text", + "content": "[7] David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, 2011." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 308, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 287, + 341 + ], + "type": "text", + "content": "[8] Shaoxiang Chen, Wenhao Jiang, Wei Liu, and Yu-Gang Jiang. Learning modality interaction for temporal sentence localization and event captioning in videos. In ECCV, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 342, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 286, + 373 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 411, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 442 + ], + "type": "text", + "content": "[11] Maksim Dzabraev, Maksim Kalashnikov, Stepan Komkov, and Aleksandr Petiushko. Mdmmt: Multidomain multimodal transformer for video retrieval. In CVPR, June 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 476 + ], + "type": "text", + "content": "[12] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In ECCV, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "text", + "content": "[13] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In CVPR, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "type": "text", + "content": "[14] Peiyan Guan, Renjing Pei, Bin Shao, Jianzhuang Liu, Weimian Li, Jiaxi Gu, Songcen Xu, Youliang Yan, and Edmund Lam. Pidro: Parallel isomeric attention with dynamic routing for text-video retrieval. In ICCV, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "text", + "content": "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. *IJCV*, 123(1):32–73, 2017." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[19] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven C.H. Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In CVPR, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[21] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "text", + "content": "[22] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In EMNLP, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[23] Chin-Yew Lin and Franz Josef Och. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In ACL, 2004." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 310, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 352 + ], + "type": "text", + "content": "[24] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "text", + "content": "[27] Sheng Liu, Zhou Ren, and Junsong Yuan. Sibnet: Sibling convolutional encoder for video captioning. In NeurIPS, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "text", + "content": "[28] Yang Liu, Samuel Albania, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts. In BMVC, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "type": "text", + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "type": "text", + "content": "[30] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pretraining model for multimodal understanding and generation. arXiv:2002.06353, 2020." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[31] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval. arXiv:2104.08860, 2021." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[32] Jianjie Luo, Yehao Li, Yingwei Pan, Ting Yao, Hongyang Chao, and Tao Mei. Coco-bert: Improving video-language pre-training with contrastive cross-modal matching and denoising. In ACM MM, 2021." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[33] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Ef-" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13764" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "ficient text-to-visual retrieval with transformers. In CVPR, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 151 + ], + "type": "text", + "content": "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "text", + "content": "[35] Vicente Ordonez, Girish Kulkarni, and Tamara L. Berg. Im2text: Describing images using 1 million captioned photographs. In NeurIPS, 2011." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 287, + 232 + ], + "type": "text", + "content": "[36] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 287, + 266 + ], + "type": "text", + "content": "[37] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 269, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 312 + ], + "type": "text", + "content": "[38] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. In ICLR, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 315, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 315, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 287, + 358 + ], + "type": "text", + "content": "[39] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for videolanguage retrieval. In CVPR, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 415 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. ICML, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 417, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 450 + ], + "type": "text", + "content": "[41] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. In ICML, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "text", + "content": "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 497, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 287, + 519 + ], + "type": "text", + "content": "[43] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 575 + ], + "type": "text", + "content": "[44] Andrew Rouditchenko, Angie Boggust, David Harwath, Brian Chen, Dhiraj Joshi, Samuel Thomas, Kartik Audhkhasi, Hilde Kuehne, Rameswar Panda, Rogerio Feris, et al. Avlnet: Learning audio-visual language representations from instructional videos. arXiv:2006.09199, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "text", + "content": "[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clipfiltered 400 million image-text pairs. arXiv:2111.02114, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[47] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, 2019." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 172 + ], + "type": "text", + "content": "[49] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, Sen Xing, Guo Chen, Junting Pan, Jiashuo Yu, Yali Wang, Limin Wang, and Yu Qiao. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[50] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "text", + "content": "[51] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv:2105.09996, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "text", + "content": "[52] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 352 + ], + "type": "text", + "content": "[53] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[54] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[55] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "text", + "content": "[56] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "text", + "content": "[57] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv:2111.07783, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 535, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 610 + ], + "type": "text", + "content": "[58] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv:2111.11432, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "text", + "content": "[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[60] Ziqi Zhang, Zhongang Qi, Chunfeng Yuan, Ying Shan, Bing Li, Ying Deng, and Weiming Hu. Open-book video captioning with retrieve-copy-generate network. In CVPR, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[61] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13765" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 175 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "with teacher-recommended learning for video captioning. In CVPR, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 118 + ], + "type": "text", + "content": "[62] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In CVPR, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 119, + 286, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 119, + 286, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 119, + 286, + 151 + ], + "type": "text", + "content": "[63] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 152, + 286, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 175 + ], + "type": "text", + "content": "[64] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, 2020." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13766" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_content_list.json b/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bbae405088b5fb8a8ca7d0e16cc1fc28ac1014a8 --- /dev/null +++ b/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_content_list.json @@ -0,0 +1,1454 @@ +[ + { + "type": "text", + "text": "Hidden Biases of End-to-End Driving Models", + "text_level": 1, + "bbox": [ + 254, + 130, + 715, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bernhard Jaeger Kashyap Chitta Andreas Geiger University of Tübingen Tübingen AI Center", + "bbox": [ + 271, + 180, + 697, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bernhard.jaeger, kashyap.chitta, a.geiger}@uni-tuebingen.de", + "bbox": [ + 220, + 219, + 750, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end driving systems have recently made rapid progress, in particular on CARLA. Independent of their major contribution, they introduce changes to minor system components. Consequently, the source of improvements is unclear. We identify two biases that recur in nearly all state-of-the-art methods and are critical for the observed progress on CARLA: (1) lateral recovery via a strong inductive bias towards target point following, and (2) longitudinal averaging of multimodal waypoint predictions for slowing down. We investigate the drawbacks of these biases and identify principled alternatives. By incorporating our insights, we develop $TF++$ , a simple end-to-end method that ranks first on the Longest6 and LAV benchmarks, gaining 11 driving score over the best prior work on Longest6.", + "bbox": [ + 75, + 300, + 473, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 527, + 209, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end driving approaches have rapidly improved in performance on the CARLA leaderboard [1], the de-facto standard for fair online evaluation. Driving scores have increased from under 20 [9,22] to over 70 [26,31] in just two years. However, why recent systems work so well is not fully understood, as both methods and training sets differ largely between submissions. Rigorous ablations are expensive due to the large design space of driving systems and the need to simulate large amounts of driving for evaluation.", + "bbox": [ + 75, + 551, + 468, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In particular, recent methods trained with Imitation Learning (IL) have shown strong performance [6, 8, 11, 26, 31]. They are trained using offline datasets, yet they can surprisingly recover from the classic compounding error problem of IL [21, 24], as indicated by their high route completions [8, 11, 26, 31]. While they do not utilize HD maps as input, they are provided with map-based GNSS locations in the center of the lane (spaced $30\\mathrm{m}$ apart on average) called target points (TPs) that describe the route the car should follow. TPs were introduced as an alternative form of conditioning signals to convey driver intent [9, 22]. Prior work [13, 14] used discrete navigation commands or NCs (i.e. follow lane, turn right,...) instead.", + "bbox": [ + 75, + 688, + 468, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we show that TP conditioned models re", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3674815b82a038fb93a973e0c1fdc6dc96c7ee5055a195ecdd133c8db238274.jpg", + "image_caption": [ + "(a) Target point shortcut", + "Figure 1: Hidden biases. (a) When outside their training distribution, current methods extrapolate waypoint predictions to the nearest target point, helping them recover. (b) The future velocity is multi-modal, but current methods commit to a single plan, which leads to interpolation." + ], + "image_footnote": [], + "bbox": [ + 501, + 266, + 691, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/968ffbfeaf7d76cb3925c655a3001bc3703236ddd87551990cc22b85a353e6bf.jpg", + "image_caption": [ + "(b) Waypoint ambiguity" + ], + "image_footnote": [], + "bbox": [ + 697, + 266, + 888, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cover from the compounding error problem because they use geometric information contained in the TP to reset steering errors periodically (at every TP). This makes them implicitly rely on accurate map information, even though they are otherwise HD map free. Steering directly towards a TP is a shortcut [16] that these IL methods learn to exploit. When methods accumulate enough steering error to be out of distribution during deployment, we observe that they steer towards the nearest TP. When the TP is close, this has the effect of driving back to the lane center, where it is in distribution again. This is illustrated in Fig. 1a. However, when the TP is far away, this shortcut can lead to catastrophic steering errors (e.g. cutting a turn). We show examples of this behavior for various SotA architectures in Section 3.1. We demonstrate that the shortcut problem is intrinsically related to the decoder architecture and that a transformer decoder [29] can mitigate it.", + "bbox": [ + 496, + 580, + 892, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another common aspect of the current SotA is that they use waypoints (future positions of an expert driver) as output representations [8, 11, 31]. We point out that this is an ambiguous representation as the future velocity is multi-", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8240", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "modal, yet the model commits to a point estimate. This is illustrated in Fig. 1b. We show that this ambiguity can sometimes be helpful due to the continuous nature of waypoints: the network can continuously interpolate between modes. We propose an alternative that explicitly predicts the uncertainty of the network via target speed classification, and show that interpolating between target speeds weighted by the uncertainty reduces collisions. Our controlled experiments also cover important but sometimes neglected details in the training of end-to-end driving systems, such as augmentation, training schedules, and dataset size. In particular, we revisit the idea of shift and rotation augmentations to aid recovery [2,5]. These were common in early IL methods for CARLA with control outputs [14], but are harder to implement with waypoint outputs and not used by the current SotA. We find that they yield significant improvements.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using these insights, we develop TransFuser++ (TF++, which sets a new SotA on the Longest6 [11] and LAV [8] benchmarks. Two of the ideas we apply, transformer decoder pooling and path-based outputs instead of waypoints, have been used in Interfuser [26]. However, these are presented as minor details, and their impact is not studied in isolation as in our work. TF++ is significantly simpler than Interfuser, yet outperforms it by a large margin on Longest6, as we show in Section 4. We use $\\sim 4\\times$ less data, 1 camera instead of 4, and do not require complex heuristics to extract throttle and brake commands for our path output.", + "bbox": [ + 75, + 333, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions:", + "text_level": 1, + "bbox": [ + 76, + 500, + 184, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that target point conditioned models learn a shortcut that helps them recover from steering errors.", + "- We point out that the waypoint output representation is ambiguous, but its continuous nature helps models collide less by interpolating to slow down.", + "- Using the insights gained from controlled experiments, we propose TransFuser++ which places first on the Longest6 and LAV benchmarks." + ], + "bbox": [ + 94, + 513, + 467, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Code, data, and models are available at https://github.com/ autonomousvision/carla_garage.", + "bbox": [ + 75, + 633, + 468, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 679, + 218, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "IL for autonomous driving dates back over 30 years [20]. It regained traction with the seminal works of [3, 5, 13] and the release of CARLA [15], a 3D simulator used in numerous recent research advances in autonomous driving [9, 10, 14, 17, 19, 25, 28, 33]. Early IL approaches evaluated in CARLA used a discrete navigation command (NC) [7, 9, 14], but their performance is not competitive to modern approaches [8, 11, 26, 31] which predict waypoints conditioned on TPs. In this work, we revisit an idea used in early systems but neglected in modern TP-conditioned methods: geometric shift augmentations [5, 9, 14, 21].", + "bbox": [ + 75, + 704, + 468, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LAV [8] supervises waypoint outputs with additional data by making predictions for other nearby agents in the", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scene. Their waypoints are initially generated with NC conditioning. These are then refined using a GRU [12]. They observe a large (+50) improvement in route completion from this refinement. Our findings suggest that the refinement module improves steering primarily through TP conditioning, which is only provided to the refinement GRU. TCP [31] observes that waypoints are stronger at collision avoidance than directly predicting controls, but sometimes suffer at large turns. They leverage the strength and mitigate the weaknesses of these two representations by creating a situation dependent ensemble. Our study suggests that the waypoints are better at collision avoidance due to an implicit slowdown when the car is uncertain. PlanT [23] investigates planning on CARLA by processing object bounding boxes with a transformer. Its sensorimotor version differs from other SotA models in that it is trained in two stages (not end-to-end). In our setting, we find that end-to-end training is crucial. We show that their observations about dataset scale also hold for end-to-end models.", + "bbox": [ + 496, + 90, + 890, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TransFuser [11,22] is a simple, well-known and widely used baseline for CARLA. We provide an explanation for the large differences in route completion between the target point conditioned TransFuser architecture and its NC conditioned baselines [7,9,14]. Further, we propose modifications to its architecture, output representation and training strategy which lead to significant improvements.", + "bbox": [ + 496, + 377, + 890, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interfuser [26] regresses a path for steering, predicts object density maps, and classifies traffic rule flags. This representation is converted by a forecasting mechanism and hand designed heuristics into control. Like our proposed method, it uses a transformer decoder for pooling features, and disentangles future velocities from the path in the output. While these ideas were already present, they were not studied or discussed as significant. Our work adds to the literature by showing that these design choices are indeed critical to performance and providing explanations as to why. Furthermore, our final system is simpler and significantly outperforms Interfuser on Longest6.", + "bbox": [ + 496, + 483, + 890, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ReasonNet [27], ThinkTwice [18] and CaT [32] are concurrent end-to-end driving approaches that all use TP conditioning and waypoints as output representation.", + "bbox": [ + 496, + 665, + 890, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Hidden Biases of End-to-End Driving", + "text_level": 1, + "bbox": [ + 500, + 723, + 836, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider the task of urban navigation from point A to B [11]. The goal is to complete routes with dense traffic, multiple lanes and complex geometries (e.g. roundabouts) without incurring infractions. Along the way, the agent encounters manually designed pre-crash traffic scenarios. Each route is a list of GNSS coordinates called target points (TPs) which can be up to $50\\mathrm{m}$ apart ( $\\sim 30\\mathrm{m}$ on average).", + "bbox": [ + 496, + 750, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Metrics: We use the CARLA online metrics. Route Completion (RC) is the percentage of the route completed. Infraction Score (IS) is a penalty factor starting at 1.0 that", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8241", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f73ae611e9a505aab52a4e2d51fd1ef542472da56a923283a1cd87bd29913bdd.jpg", + "image_caption": [ + "(a) TransFuser (NC conditioned)" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 346, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/757d0036a0c9d88c7c99088a58efcdfe252f25d64716dbec9decaa89ebc34dba.jpg", + "image_caption": [ + "(b) TransFuser (TP conditioned)" + ], + "image_footnote": [], + "bbox": [ + 349, + 88, + 619, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c9018ef2eb763ec683c5380c78a2d062c24ef8f9cdbcba8cc6ded3fe01c425d6.jpg", + "image_caption": [ + "(c) LAV [8] (TP conditioned)", + "Figure 2: Extrapolation to target point. In unknown situations, TP conditioned methods extrapolate their waypoints towards target points. This periodically resets steering errors and is a form of implicit map based recovery. However, relying on extrapolation is a shortcut that can lead to catastrophic errors in certain situations (see Fig. 3a and Fig. 3b)." + ], + "image_footnote": [], + "bbox": [ + 622, + 88, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "gets reduced multiplicatively for every infraction. Our main metric is the Driving Score (DS) which multiplies RC with the IS. Where insightful, we report infraction per kilometer metrics. For a comprehensive description, refer to [11].", + "bbox": [ + 75, + 431, + 468, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Baseline:** As a baseline, we reproduce TransFuser [11]. This is a simple method representative of current SotA driving systems on CARLA. The main differences in our reproduction are a $360^{\\circ}$ field of view (FOV) LiDAR to enable safe lane changes and a dataset where the expert is driving $2 \\times$ faster to speed up evaluation. A list of other minor changes is provided in the supplementary material.", + "bbox": [ + 75, + 494, + 468, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Benchmark: For all experiments in this section, we train on CARLA towns 01, 03, 04, 06, 07 and 10. We use the 16 validation routes from LAV [8] in the withheld towns 02 and 05 for evaluation. In order to reduce the influence of training and evaluation variance which can be large in CARLA [4, 11], every experiment reports the average of 3 training seeds, evaluated 3 times each. We additionally report the training standard deviation for the main metrics.", + "bbox": [ + 75, + 602, + 468, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. A shortcut for recovery", + "text_level": 1, + "bbox": [ + 76, + 739, + 290, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While current SotA CARLA methods have fundamentally different architectures [8,11,26,31], they are all trained with conditional IL using fixed pre-recorded datasets. The evaluation task involves challenging routes which are $\\sim 1.5\\mathrm{km}$ long, so it would be expected that these methods suffer from compounding errors, a well-known problem for IL [24]. Geometric shift augmentations [5] are a common approach to teach IL methods how to recover from such compounding steering errors [5, 14]. Surprisingly,", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "even though such augmentations are not employed by SotA methods on CARLA, they report high RCs, demonstrating that they are not prone to this expected failure mode.", + "bbox": [ + 496, + 431, + 890, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All the aforementioned methods condition their predictions using the next target point (TP) along the route. To understand the importance of the TP, we train 2 versions of our reproduction of TransFuser: one with the original TP conditioning and another with a discrete (NC) condition. This is a one hot vector that indicates whether the car should follow the lane, turn right at the next intersection, etc. It contains no geometric information about the center of the lanes, but still removes the inherent task ambiguity of inner-city driving [13]. For the NC conditioned model, we also implement shift and rotation augmentations by collecting augmented frames during data collection. We deploy a second camera in the simulator, that changes its position and orientation randomly at every time-step. The output labels are transformed accordingly for training the model (implementation details and examples can be found in the supplementary material). The results are shown in Table 1.", + "bbox": [ + 496, + 479, + 892, + 734 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ceb7e678cd448085bd5fa9be991c882af0ab2182eb183f9c119675c7b9fecf02.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cond.Aug.DS ↑RC ↑Dev ↓
NC-32 ± 856 ± 120.86
NC✓35 ± 354 ± 40.99
TP-39 ± 984 ± 70.00
", + "bbox": [ + 553, + 750, + 839, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Conditioning and Augmentation.", + "bbox": [ + 549, + 832, + 841, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We observe a significant difference in RC of 28 points when switching between TP and NC conditioning. The TP conditioned model has 0 route deviations per km (\"Dev\"),", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8242", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/545be28bd61414f2dd817fdb99088f2423a765d951c7a6f901d0e5d1db571516.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 346, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/531a2b62baf02b0ced74678761cfbd331f200b760a2278217625d3fd21888851.jpg", + "image_caption": [ + "(a) TCP [31] follows a shortcut." + ], + "image_footnote": [], + "bbox": [ + 80, + 152, + 230, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7b9b4a700097ebe9cd4815dcdfe60f590af54af6592135bab170194eed9f5ba1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 88, + 617, + 142 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/882ff92bb508758e6a52f6ea756405da23d92cfe04f62ab4b41dcfc1ee83bc82.jpg", + "image_caption": [ + "(b) TransFuser follows a shortcut." + ], + "image_footnote": [], + "bbox": [ + 349, + 143, + 617, + 351 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6dae0f4b7ce141a1af37a3748af96f3c2d1348f9f011e3237ad2ca27f962ecb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 890, + 142 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/95bcd4da5c5bcfd8bf21868307d38123f417fde75dd2990c98a47239d39596a0.jpg", + "image_caption": [ + "(c) A transformer decoder fixes this." + ], + "image_footnote": [], + "bbox": [ + 622, + 143, + 890, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which indicates that it never takes a wrong turn or drives too far away (more than $30\\mathrm{m}$ ) from the lane. However, this is not the case for NC conditioned models. While we do see a small improvement in DS with augmentation, its RC is still unsatisfactory. This indicates that the geometric information regarding the lane center available with the TP aids recovery in SotA IL methods, prompting further investigation. We inspect the TransFuser and LAV [8] models by constructing situations where the car is forcefully steered out of lane. Fig. 2 visualizes these scenarios with the camera, LiDAR and ground truth HD maps. For TP conditioned models, the predicted waypoints (shown in blue) extrapolate towards the nearest TP (red) even though the situation is out of distribution. This shows that one reason for their strong route following is that the bird's eye view (BEV) TP resets steering errors periodically: methods learn to steer towards nearby TPs because the expert trajectory in the dataset always goes through them. This has the effect that TP conditioned methods periodically drive back to the center of the lane, resetting any accumulated errors. Furthermore, it suggests that SotA IL approaches strongly rely on pre-specified geometric information regarding lane centers for recovery.", + "bbox": [ + 75, + 417, + 472, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We identify this as a form of shortcut learning, which can be useful when the car is close to a target point, but can also lead to catastrophic steering errors when the target point is further away. An example would be directly extrapolating to a target point behind a turn, which leads to cutting the turn. In Fig. 3a and Fig. 3b, we show instances where this happens for TCP [31] and TransFuser [11] in a validation town at nighttime. TransFuser also predicts the BEV segmentation as an auxiliary task, which we overlay", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d3fbcfa66a60c99ada9e90fe765c072327ce6e055601f2432306eb08e192d57.jpg", + "image_caption": [ + "Figure 3: Target point shortcut. When TP conditioned methods extrapolate to spatially distant waypoints, they incur large steering errors. Replacing global average pooling in TransFuser with a cross-attention mechanism mitigates the issue.", + "Figure 4: Pooling. Existing approaches vectorize feature grids either by global average pooling (top, e.g. [8, 11]) or with attention mechanisms (bottom, e.g. [26, 31]). The latter retains spatial information gained via auxiliary tasks." + ], + "image_footnote": [], + "bbox": [ + 501, + 414, + 893, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in the figure (gray: route, yellow: lane marking). For TCP, we render the ground truth map and omit the LiDAR, since it does not use LiDAR and does not predict BEV segmentation. Both methods predict waypoints that are tilted towards the TP instead of following the street. As a result, they drive into the opposing lane. We refer to this as the TP shortcut.", + "bbox": [ + 496, + 622, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Improved pooling and data augmentation", + "text_level": 1, + "bbox": [ + 500, + 724, + 854, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One design choice which differs across SotA architectures is the pooling applied between the encoder and decoder. In Fig. 4, we summarize these. In particular, TransFuser [11] and LAV [8] employ global average pooling (GAP) followed by an MLP. InterFuser [26] pools features via the cross-attention mechanism of a transformer decoder. Finally, TCP [31], which has two decoders, uses GAP for one decoder and attention-based pooling for the other.", + "bbox": [ + 496, + 748, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 4, the networks are encouraged to learn spatially meaningful feature grids through auxiliary convo", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8243", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lutional decoders that predict outputs such as BEV semantic segmentation. However, GAP does not maintain the spatial information in the features. Fig. 3b shows the BEV road segmentation predicted by TransFuser overlaid by its Li-DAR input. Unlike the waypoints, the BEV predictions are quite accurate. This indicates strong features coming from the encoder. Nevertheless, the final conditional waypoint predictions focus on the TP signal in this situation. We hypothesize that the GAP operation makes it difficult for the downstream decoder to utilize the strong BEV features. Note that attention-based pooling, such as the transformer decoder of Interfuser, preserves spatial information through the use of positional encodings.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare the original TransFuser GAP approach with the spatially preserving design of Interfuser in Table 2. For the latter variant, we remove the GAP operation and MLP in TransFuser. We then process the $8 \\times 8$ BEV feature grid as tokens with a transformer decoder. Implementation details can be found in the supplementary material.", + "bbox": [ + 75, + 287, + 472, + 381 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f0c7a04d45982c518fdb0d8fd7ec4b75226d6d98dd0389476ebbec1be2755e4c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PoolingAug.DS ↑RC ↑Stat ↓
GAP + MLP-39 ± 984 ± 71.04
Transformer Decoder-43 ± 693 ± 30.55
Transformer Decoder✓49 ± 890 ± 40.10
", + "bbox": [ + 88, + 393, + 460, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Pooling and augmentation.", + "bbox": [ + 147, + 476, + 398, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After replacing GAP with the transformer decoder, the RC increases by 9 points and collisions per km with static objects (\"Stat\") reduce by a factor of 2. Static objects (such as poles) only occur outside lanes in CARLA, implying that the network manages to stay in the lane far more consistently. Fig. 3c provides qualitative evidence that the transformer decoder can succeed in situations where GAP failed due to the TP shortcut (additional examples in supplementary). In addition, we investigate the inclusion of shift and rotation augmentations designed to aid lateral recovery (as described in Section 3.1). Collisions with the static environment are reduced further by a factor of 5 indicating that augmentation provides strong benefits.", + "bbox": [ + 75, + 500, + 472, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. The ambiguity of waypoints", + "text_level": 1, + "bbox": [ + 76, + 709, + 328, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Waypoints are used in many SotA systems as outputs [8, 11, 31]. They are obtained by recording an expert driver's GNSS locations at fixed time intervals (e.g. 0.5s) and transforming them into a local coordinate frame. The model is then trained to predict future waypoints, typically with an $L_{1}$ regression objective. Waypoints entangle both the path and the future velocities of the vehicle. The velocity after a specific time interval extracted from the waypoints is used by a downstream controller as a target speed.", + "bbox": [ + 75, + 733, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 5, we plot a histogram comparing the target speeds of the expert algorithm (i.e. training dataset) to those", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7496a29c2d4fffcb9f57a32a1cfafa0bfdd5ea9958471d6b60b2c2d80d5686e5.jpg", + "image_caption": [ + "Figure 5: Transfuser interpolates between modes." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 857, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "extracted from TransFuser (best model in Table 2). Interestingly, the distributions are quite different. By design, the expert chooses one of four target speeds: 29, 18, 7, or 0 km/h. These values cover behaviors needed for city driving, corresponding to four situations: regular driving, slowing down in intersections, slowing down near pedestrians, and stopping. In contrast, for TransFuser, the predicted target speeds cover the entire 0-29 km/h range.", + "bbox": [ + 496, + 338, + 890, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the future path that our expert follows is deterministic and unambiguous (center of the lane, lane change at pre-defined locations), future velocities are multi-modal. As both are jointly represented by waypoints, this leads to an ambiguous entangled representation. However, existing methods which utilize this entangled waypoint representation predict only point estimates (i.e., a single set of waypoints) as output, hence only a single mode is modeled. From Fig. 5, we observe that the waypoint based TransFuser model indeed interpolates between modes, a behavior that is expected when modeling multi-modal outputs deterministically. We illustrate this behavior in detail using the examples of (1) approaching an intersection with a green light (Fig. 6a), and (2) a cyclist cutting into the vehicle's path (Fig. 6b). TransFuser slows down in these situations since it is uncertain, e.g., the light may turn red in Fig. 6a. The car stops in time in Fig. 6b because of the decreased speed.", + "bbox": [ + 496, + 460, + 892, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While this averaging is indeed beneficial in some situations, an entangled representation is undesirable as it is less interpretable and does not explicitly expose uncertainty. Moreover, when stopped (all waypoints collapsed to one location), the steering signal is undefined which requires additional heuristics in the controller. To resolve this, we disentangle the future velocities from the path by sampling the expert's position at fixed distances instead of fixed time intervals for training a (deterministic) path predictor. However, when predicting the path instead of time-dependent trajectories, one requires an additional method to determine a target speed for the car. We propose to predict the tar", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8244", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f92f0a9f4dcb7a9453dc9387e8aae22342b37524a874d13ee277f88c6c6cbfb.jpg", + "image_caption": [ + "(a) Slowing down at a green light." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 348, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51c9dde26c5bd6454eab8b33693d7623c562f53a5bdfe26c9c669bc55ac5929d.jpg", + "image_caption": [ + "(b) Slowing down in an intersection." + ], + "image_footnote": [], + "bbox": [ + 349, + 88, + 619, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f295c149d94641124f7afc384c7437a83b1a86a80d044ff403ce9821b08a4157.jpg", + "image_caption": [ + "(c) Disentangling route and target speed." + ], + "image_footnote": [], + "bbox": [ + 620, + 88, + 890, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "get speed by simple classification using an additional MLP head. We incorporate the prediction uncertainties into the final output by using a confidence weighted average of the predicted target speed as input to the controller. Note that this is just one example of how the resulting confidence estimates may be leveraged by a vehicle controller.", + "bbox": [ + 75, + 431, + 470, + 523 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/20eb67207828b7ad0ef9eecf0cecf319175ae342402eb35c911bd3f32d40de32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OutputDS ↑RC ↑Veh ↓Stat ↓
Waypoints49 ± 890 ± 40.700.10
Path + Argmax40 ± 188 ± 21.250.03
Path + Weighted50 ± 388 ± 10.830.02
", + "bbox": [ + 99, + 540, + 446, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to the waypoint representation, naively converting the predicted classes to a target speed via the most likely class (Argmax) increases vehicle collisions (\"Veh\"), as seen in Table 3. Employing the proposed weighting (Weighted) achieves lower collisions. Moreover, the path-based model steers better as indicated by the lower static obstacle collisions (\"Stat\"). The disentangled representation achieves the same driving score as the entangled waypoint representation, providing an alternative with a simpler and more interpretable controller: (1) it eases design since it has identical parameters to the controller in the expert, (2) it has access to an unambiguous path representation even when driving slowly, and (3) it explicitly exposes and makes use of uncertainties with regard to target speeds.", + "bbox": [ + 75, + 641, + 468, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 6c shows a qualitative example. The car is slowing down due to its uncertainty. The lower target speed allows it successfully merge behind the cyclist.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4de65413a2c6ceebd9f2de56d54ee791b05faf6cfcd44c2ea1b5ada5bcc31d27.jpg", + "image_caption": [ + "Figure 6: Waypoints are ambiguous. The model's output representation forces it to predict a single mode for future velocities. There is a possibility that the traffic light might turn red in the future, or that the cyclist either cuts in or yields to the agent. The particular mode (or interpolation thereof) that the model converges to, depends on the training run and dataset.", + "Figure 7: TransFuser++ architecture." + ], + "image_footnote": [], + "bbox": [ + 500, + 430, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Scaling up to TransFuser++", + "text_level": 1, + "bbox": [ + 500, + 710, + 748, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By incorporating these insights, we obtain a significantly improved version of TransFuser [11] that we call TransFuser++ (Fig. 7). We now describe two important implementation changes compared to [11] that involve scaling.", + "bbox": [ + 498, + 736, + 890, + 797 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/819bfe98eca49066af663a4a8902b0f7b968e51fb9def41f4304e5ae9604cdc4.jpg", + "table_caption": [ + "Table 3: Output representation." + ], + "table_footnote": [], + "table_body": "
Two stageFrozenDS ↑RC ↑Veh ↓
--50 ± 388 ± 10.83
✓-54 ± 192 ± 50.79
✓✓44 ± 697 ± 11.21
", + "bbox": [ + 537, + 816, + 856, + 887 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4: Two stage training.", + "bbox": [ + 596, + 898, + 792, + 914 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8245", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training schedule: We observe no benefits from simply doubling the number of training epochs for a single training stage. However, in Table 4, we double the training time with a two-stage approach. First, we pre-train the encoder with only the perception losses (2D depth, 2D and BEV semantic segmentation, and vehicle bounding boxes, as shown in Fig. 7) for the usual number of epochs. We then fine-tune the resulting checkpoint with all losses, after including the GRU decoder and MLP classifier. Initializing the backbone with features that are pre-trained on the auxiliary tasks leads to a 4 DS improvement, consistent with the claims of LAV [8]. We also experiment with freezing the pre-trained backbone and only training the transformer decoder and its heads in the second stage. This leads to a drop of 10 DS, indicating that end-to-end optimization is important.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset scale: Recent work on CARLA [23] shows large improvements by scaling the dataset size. However, [23] considers planning models with privileged inputs on training towns. We investigate the impact of scaling when evaluating on held-out validation towns for end-to-end models. We start with 185k training samples (as in [11]) and scale it up by re-running the training routes 3 times with different traffic (555k frames). Table 5 shows a clear improvement of 6 DS via scaling. This shows that current IL approaches can still benefit from larger datasets. We train for the same amount of epochs, which implies that this improvement comes at the cost of $3 \\times$ longer training.", + "bbox": [ + 75, + 318, + 472, + 501 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1f5084fc042e3f1644c63f03184187a1a040153c153443b964a45c319954c320.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset sizeDS ↑RC ↑Veh ↓
185k54 ± 192 ± 50.79
555k60 ± 698 ± 10.73
", + "bbox": [ + 140, + 513, + 408, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5: Effects of scale.", + "bbox": [ + 187, + 582, + 357, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Comparison to State of the Art", + "text_level": 1, + "bbox": [ + 76, + 617, + 361, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods: This section demonstrates the advantages of $\\mathrm{TF}++$ over several state-of-the-art models, which we list below, starting with the most recent. (1) Interfuser [26] processes multiple cameras and a BEV LiDAR by encoding each of them individually with a CNN. The resulting feature grids are fed into a transformer and decoded into perception outputs and the path to follow. The network does not predict longitudinal controls. Instead, it uses the BEV bounding boxes and a simple motion forecast (extrapolating historical dynamics) and linearly optimizes for the target speed using heuristically chosen objectives. (2) Perception PlanT [23] is a transformer based method trained in two independent stages that uses BEV bounding boxes as an intermediate representation. Its perception is based on TransFuser, and its planner outputs waypoints. (3) TCP [31] is a camera-only model with two output representations: waypoints and controls. During test time, it ensembles the two", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "outputs together with a weighted average that changes based on whether the vehicle is turning. (4) TransFuser [11] fuses perspective cameras with a BEV LiDAR by processing them with individual CNNs and exchanging features using transformers. It uses global average pooling and waypoint outputs, as described in Section 3. (5) LAV [8] processes cameras and LiDAR point clouds into an intermediate BEV representation. It trains a planner on this representation that predicts waypoints for the ego vehicle with a NC conditioned GRU followed by a TP conditioned refinement GRU. During training, the planner is also tasked to predict the trajectory of other surrounding vehicles to increase the number of labels. The authors release two versions of this method, which we call LAV v1 and LAV v2. (6) WOR [7] is an IL method for which the labels are enriched using a reward function to provide dense supervision for all possible actions. Unlike the other baselines, it is conditioned with the NC and does not utilize the TP. WOR is the best NC conditioned baseline that is publicly available.", + "bbox": [ + 496, + 90, + 893, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Benchmarks:** We use two benchmarks to evaluate on seen and unseen towns, with scenarios taken from [11] (type 1,3,4,7,8,9 and 10). Table 6 compares the performance of systems on the Longest6 benchmark [11] which consists of 36 routes in training towns 01 to 06. On Longest6, models trained on any data can be evaluated, so we compare against the author-provided models or directly report numbers from the respective papers. The mean and std of three evaluations is reported. For $\\mathrm{TF}++$ we train 3 models and report the average result. Longest6 has the advantage that it evaluates in dense traffic and has diverse towns and routes. The drawbacks are that it does not penalize stop sign infractions or measure generalization to new towns. Therefore, we additionally compare performance on validation towns using the LAV [8] routes in Table 7. These are 4 routes with 4 weathers (16 combined) in Town 02 and 05, which are withheld during training. We re-train methods with the dataset from the corresponding paper 3 times and evaluate each seed 3 times. Reported results are the mean of all runs and the std between the training seeds. We compare against the reported SotA TCP and our reproduced TransFuser baseline on this benchmark and provide additional baselines in the supplementary.", + "bbox": [ + 496, + 380, + 893, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset: For $\\mathrm{TF}++$ we collect data on the same training routes as in [11] with an improved expert labeling algorithm (described in the supplementary material). We repeat this 3 times on the same routes with different traffic, as in [23]. For validation, we withhold Town 02 and 05 during training, else we train on all towns. In total, we train with 750k frames (550k when withholding Town 02 and 05).", + "bbox": [ + 496, + 731, + 893, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results: We start with the results in training towns (Table 6). The best NC conditioned method, WOR, has significantly lower RC than all TP conditioned systems (22 lower than LAV v1). In particular, the route deviations (Dev) are", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8246", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/86c0e879617727145fd2ffb657f88796ef79160aa86b1c4f6d1fcd7aae2fc4f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Dev ↓TO ↓Block ↓
WOR [7]21 ± 348 ± 40.56 ± 0.030.181.050.371.280.880.080.20
LAV v1 [8]33 ± 170 ± 30.51 ± 0.020.160.830.150.960.060.120.45
Interfuser [26]47 ± 674 ± 10.63 ± 0.070.061.140.110.240.000.520.06
TransFuser [11]47 ± 693 ± 10.50 ± 0.060.032.450.070.160.000.060.10
TCP [31]48 ± 372 ± 30.65 ± 0.040.041.080.230.140.020.180.35
LAV v2 [8]58 ± 183 ± 10.68 ± 0.020.000.690.150.230.080.320.11
Perception PlanT [23]58 ± 588 ± 10.65 ± 0.060.070.970.110.090.000.130.13
TF++ (ours)69 ± 094 ± 20.72 ± 0.010.000.830.010.050.000.070.06
Expert81 ± 390 ± 10.91 ± 0.040.010.210.000.010.000.070.09
", + "bbox": [ + 112, + 88, + 857, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e1574de00cf735e93d3fe286bbf845d8b367ea21f18f5224109304339c41ca3c.jpg", + "table_caption": [ + "Table 6: Performance on training towns (Longest6). Released models, std over 3 evaluations." + ], + "table_footnote": [], + "table_body": "
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Stop ↓Dev ↓TO ↓Block ↓
TransFuser (ours)39 ± 984 ± 70.46 ± 0.060.000.741.040.201.070.000.230.21
TCP [31]58 ± 585 ± 30.67 ± 0.060.000.350.160.011.050.000.190.19
TF++ (ours)70 ± 699 ± 00.70 ± 0.060.010.630.010.040.260.000.050.00
Expert94950.990.000.020.000.020.000.000.000.08
", + "bbox": [ + 96, + 289, + 874, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "10 times higher. Current TP conditioned SotA methods achieve (close to) 0 route deviations with driving scores in the range of 47 DS to 58 DS. TF++ outperforms all baselines on Longest6, showing a $19\\%$ relative improvement over the previous SotA Perception PlanT. TF++ achieves close to expert-level performance on all infractions except for vehicle collisions (Veh). On the validation towns (Table 7) TF++ outperforms our reproduced TransFuser by 31 DS improving all metrics, particularly environmental collisions (Stat). It improves the prior SotA result TCP by $21\\%$ . Runtime: Table 8 compares the runtime of TF++ and TransFuser which have the same sensor setup. TF++ yields a large DS improvement while being only $14\\%$ slower.", + "bbox": [ + 75, + 417, + 468, + 613 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/00b3d66ce55636f08812b19500be937753424e762fd2ba86a4189a0457ca0d13.jpg", + "table_caption": [ + "Table 7: Performance on validation towns (LAV). Reproduced models, std over 3 trainings (for 3 evaluations each)." + ], + "table_footnote": [], + "table_body": "
Method ↑DS ↑Time (ms) ↓
TransFuser (ours)3944
TF++ (ours)7050
", + "bbox": [ + 137, + 626, + 410, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: Runtime. We show the runtime per frame in ms averaged over 300 time steps on a single route on a RTX 3090. TF++ outperforms TransFuser by a wide margin despite using a similar compute budget during inference.", + "bbox": [ + 75, + 695, + 468, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 768, + 194, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we show that recent SotA driving models have exceptional route following abilities because they learn a strong bias towards following nearby TPs. Shortcut learning like this is a general phenomenon in deep neural networks [16] and has been observed in the context of autonomous driving for inputs such as velocity [14] or temporal frames [3, 30]. We add to this literature by observ-", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing shortcut learning with respect to the conditioning signal. While shortcut learning usually has a negative impact on performance, we observe positive (improved recovery) and negative (cutting turns) effects. We show that the negative effects can be mitigated by avoiding global pooling and incorporating data augmentation.", + "bbox": [ + 496, + 417, + 890, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A second commonality in SotA approaches is the use of waypoints as an output representation. We observe that this representation is ambiguous because it predicts a point estimate for multi-modal future velocities. We disambiguate the representation by disentangling future velocities from the deterministic path predictions and classifying target speeds instead. We then weigh target speeds according to their predicted confidence in our controller. Surprisingly, we find that interpolation is helpful for reducing collisions.", + "bbox": [ + 496, + 508, + 890, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose TransFuser++ by improving the popular baseline TransFuser with a series of controlled experiments. TF++ is a simple end-to-end method that sets a new state of the art on the LAV and Longest6 benchmarks.", + "bbox": [ + 496, + 643, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: This study investigates urban driving in CARLA, where all investigated methods drive at relatively low speed $(< 35\\mathrm{km / h})$ . Therefore, problems specific to high-speed driving are not considered. In addition, lanes are free of static obstacles, hence scenarios requiring navigation around them are not included. We point out the strong reliance of current methods on TPs for recovery. This implies a reliance on accurate localization and mapping to obtain these TPs. They are accurately mapped in CARLA, however, this assumption might not hold in real environments.", + "bbox": [ + 496, + 704, + 890, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact: We aim to make progress towards autonomous driving. This technology, if realized, could have massive societal impact, reducing road accidents, trans", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8247", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "portation costs and improving the mobility of elderly people. Potential negative implications include a reduction in jobs for human drivers and possible military applications.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements: Andreas Geiger and Bernhard Jaeger were supported by the ERC Starting Grant LEGO-3D (850533), the BMWi in the project KI Delta Learning (project number 19A19013O) and the DFG EXC number 2064/1 - project number 390727645. Kashyap Chitta was supported by the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. We thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Bernhard Jaeger and Kashyap Chitta as well as Niklas Hanselmann and Katrin Renz for proofreading.", + "bbox": [ + 75, + 143, + 472, + 310 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 323, + 173, + 338 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020.1", + "[2] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 2", + "[3] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. In Proc. Robotics: Science and Systems (RSS), 2019. 2, 8", + "[4] Aseem Behl, Kashyap Chitta, Aditya Prakash, Eshed Ohn-Bar, and Andreas Geiger. Label efficient visual abstractions for autonomous driving. In Proc. IEEE International Conf. on Intelligent Robots and Systems (IROS), 2020. 3", + "[5] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba. End to end learning for self-driving cars. arXiv.org, 1604.07316, 2016. 2, 3", + "[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[7] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2, 7, 8", + "[8] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 7, 8", + "[9] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Proc. Conf. on Robot Learning (CoRL), 2019. 1, 2", + "[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2" + ], + "bbox": [ + 78, + 347, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[12] Kyunghyun Cho, Bart van Merrienboer, Caglar Güçehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proc. of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 2", + "[13] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2018. 1, 2, 3", + "[14] Felipe Codevilla, Eder Santana, Antonio M. López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 8", + "[15] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proc. Conf. on Robot Learning (CoRL), 2017. 2", + "[16] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1, 8", + "[17] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zak Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[18] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[19] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[20] Dean Pomerleau. ALVINN: an autonomous land vehicle in a neural network. In Advances in Neural Information Processing Systems (NIPS), 1988. 2", + "[21] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2", + "[22] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[23] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, Almut Sophia Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8248", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "representations. In Proc. Conf. on Robot Learning (CoRL), 2022. 2, 7, 8", + "[24] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Conference on Artificial Intelligence and Statistics (AISTATS), 2011. 1, 3", + "[25] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In Proc. Conf. on Robot Learning (CoRL), 2018. 2", + "[26] Hao Shao, Letian Wang, RuoBing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Proc. Conf. on Robot Learning (CoRL), 2022. 1, 2, 3, 4, 7, 8", + "[27] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[28] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 1", + "[30] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 8", + "[31] Peng Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 4, 5, 7, 8", + "[32] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[33] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 699 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8249", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_model.json b/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82bc9f150acf5a65748ded47bf702f1f4782a28f --- /dev/null +++ b/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.131, + 0.717, + 0.154 + ], + "angle": 0, + "content": "Hidden Biases of End-to-End Driving Models" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.181, + 0.699, + 0.218 + ], + "angle": 0, + "content": "Bernhard Jaeger Kashyap Chitta Andreas Geiger University of Tübingen Tübingen AI Center" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.22, + 0.75, + 0.235 + ], + "angle": 0, + "content": "{bernhard.jaeger, kashyap.chitta, a.geiger}@uni-tuebingen.de" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.513 + ], + "angle": 0, + "content": "End-to-end driving systems have recently made rapid progress, in particular on CARLA. Independent of their major contribution, they introduce changes to minor system components. Consequently, the source of improvements is unclear. We identify two biases that recur in nearly all state-of-the-art methods and are critical for the observed progress on CARLA: (1) lateral recovery via a strong inductive bias towards target point following, and (2) longitudinal averaging of multimodal waypoint predictions for slowing down. We investigate the drawbacks of these biases and identify principled alternatives. By incorporating our insights, we develop \\( TF++ \\), a simple end-to-end method that ranks first on the Longest6 and LAV benchmarks, gaining 11 driving score over the best prior work on Longest6." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.528, + 0.21, + 0.543 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.689 + ], + "angle": 0, + "content": "End-to-end driving approaches have rapidly improved in performance on the CARLA leaderboard [1], the de-facto standard for fair online evaluation. Driving scores have increased from under 20 [9,22] to over 70 [26,31] in just two years. However, why recent systems work so well is not fully understood, as both methods and training sets differ largely between submissions. Rigorous ablations are expensive due to the large design space of driving systems and the need to simulate large amounts of driving for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.885 + ], + "angle": 0, + "content": "In particular, recent methods trained with Imitation Learning (IL) have shown strong performance [6, 8, 11, 26, 31]. They are trained using offline datasets, yet they can surprisingly recover from the classic compounding error problem of IL [21, 24], as indicated by their high route completions [8, 11, 26, 31]. While they do not utilize HD maps as input, they are provided with map-based GNSS locations in the center of the lane (spaced \\(30\\mathrm{m}\\) apart on average) called target points (TPs) that describe the route the car should follow. TPs were introduced as an alternative form of conditioning signals to convey driver intent [9, 22]. Prior work [13, 14] used discrete navigation commands or NCs (i.e. follow lane, turn right,...) instead." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this paper, we show that TP conditioned models re" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.267, + 0.692, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.475, + 0.669, + 0.489 + ], + "angle": 0, + "content": "(a) Target point shortcut" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.267, + 0.889, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.475, + 0.866, + 0.489 + ], + "angle": 0, + "content": "(b) Waypoint ambiguity" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.494, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Figure 1: Hidden biases. (a) When outside their training distribution, current methods extrapolate waypoint predictions to the nearest target point, helping them recover. (b) The future velocity is multi-modal, but current methods commit to a single plan, which leads to interpolation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.893, + 0.839 + ], + "angle": 0, + "content": "cover from the compounding error problem because they use geometric information contained in the TP to reset steering errors periodically (at every TP). This makes them implicitly rely on accurate map information, even though they are otherwise HD map free. Steering directly towards a TP is a shortcut [16] that these IL methods learn to exploit. When methods accumulate enough steering error to be out of distribution during deployment, we observe that they steer towards the nearest TP. When the TP is close, this has the effect of driving back to the lane center, where it is in distribution again. This is illustrated in Fig. 1a. However, when the TP is far away, this shortcut can lead to catastrophic steering errors (e.g. cutting a turn). We show examples of this behavior for various SotA architectures in Section 3.1. We demonstrate that the shortcut problem is intrinsically related to the decoder architecture and that a transformer decoder [29] can mitigate it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Another common aspect of the current SotA is that they use waypoints (future positions of an expert driver) as output representations [8, 11, 31]. We point out that this is an ambiguous representation as the future velocity is multi-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8240" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "modal, yet the model commits to a point estimate. This is illustrated in Fig. 1b. We show that this ambiguity can sometimes be helpful due to the continuous nature of waypoints: the network can continuously interpolate between modes. We propose an alternative that explicitly predicts the uncertainty of the network via target speed classification, and show that interpolating between target speeds weighted by the uncertainty reduces collisions. Our controlled experiments also cover important but sometimes neglected details in the training of end-to-end driving systems, such as augmentation, training schedules, and dataset size. In particular, we revisit the idea of shift and rotation augmentations to aid recovery [2,5]. These were common in early IL methods for CARLA with control outputs [14], but are harder to implement with waypoint outputs and not used by the current SotA. We find that they yield significant improvements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Using these insights, we develop TransFuser++ (TF++, which sets a new SotA on the Longest6 [11] and LAV [8] benchmarks. Two of the ideas we apply, transformer decoder pooling and path-based outputs instead of waypoints, have been used in Interfuser [26]. However, these are presented as minor details, and their impact is not studied in isolation as in our work. TF++ is significantly simpler than Interfuser, yet outperforms it by a large margin on Longest6, as we show in Section 4. We use \\(\\sim 4\\times\\) less data, 1 camera instead of 4, and do not require complex heuristics to extract throttle and brake commands for our path output." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.5, + 0.185, + 0.513 + ], + "angle": 0, + "content": "Contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.515, + 0.468, + 0.543 + ], + "angle": 0, + "content": "- We show that target point conditioned models learn a shortcut that helps them recover from steering errors." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.545, + 0.468, + 0.589 + ], + "angle": 0, + "content": "- We point out that the waypoint output representation is ambiguous, but its continuous nature helps models collide less by interpolating to slow down." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.59, + 0.468, + 0.634 + ], + "angle": 0, + "content": "- Using the insights gained from controlled experiments, we propose TransFuser++ which places first on the Longest6 and LAV benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.515, + 0.468, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Code, data, and models are available at https://github.com/ autonomousvision/carla_garage." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.68, + 0.219, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.87 + ], + "angle": 0, + "content": "IL for autonomous driving dates back over 30 years [20]. It regained traction with the seminal works of [3, 5, 13] and the release of CARLA [15], a 3D simulator used in numerous recent research advances in autonomous driving [9, 10, 14, 17, 19, 25, 28, 33]. Early IL approaches evaluated in CARLA used a discrete navigation command (NC) [7, 9, 14], but their performance is not competitive to modern approaches [8, 11, 26, 31] which predict waypoints conditioned on TPs. In this work, we revisit an idea used in early systems but neglected in modern TP-conditioned methods: geometric shift augmentations [5, 9, 14, 21]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "LAV [8] supervises waypoint outputs with additional data by making predictions for other nearby agents in the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.378 + ], + "angle": 0, + "content": "scene. Their waypoints are initially generated with NC conditioning. These are then refined using a GRU [12]. They observe a large (+50) improvement in route completion from this refinement. Our findings suggest that the refinement module improves steering primarily through TP conditioning, which is only provided to the refinement GRU. TCP [31] observes that waypoints are stronger at collision avoidance than directly predicting controls, but sometimes suffer at large turns. They leverage the strength and mitigate the weaknesses of these two representations by creating a situation dependent ensemble. Our study suggests that the waypoints are better at collision avoidance due to an implicit slowdown when the car is uncertain. PlanT [23] investigates planning on CARLA by processing object bounding boxes with a transformer. Its sensorimotor version differs from other SotA models in that it is trained in two stages (not end-to-end). In our setting, we find that end-to-end training is crucial. We show that their observations about dataset scale also hold for end-to-end models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.484 + ], + "angle": 0, + "content": "TransFuser [11,22] is a simple, well-known and widely used baseline for CARLA. We provide an explanation for the large differences in route completion between the target point conditioned TransFuser architecture and its NC conditioned baselines [7,9,14]. Further, we propose modifications to its architecture, output representation and training strategy which lead to significant improvements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.665 + ], + "angle": 0, + "content": "Interfuser [26] regresses a path for steering, predicts object density maps, and classifies traffic rule flags. This representation is converted by a forecasting mechanism and hand designed heuristics into control. Like our proposed method, it uses a transformer decoder for pooling features, and disentangles future velocities from the path in the output. While these ideas were already present, they were not studied or discussed as significant. Our work adds to the literature by showing that these design choices are indeed critical to performance and providing explanations as to why. Furthermore, our final system is simpler and significantly outperforms Interfuser on Longest6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.711 + ], + "angle": 0, + "content": "ReasonNet [27], ThinkTwice [18] and CaT [32] are concurrent end-to-end driving approaches that all use TP conditioning and waypoints as output representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.838, + 0.742 + ], + "angle": 0, + "content": "3. Hidden Biases of End-to-End Driving" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.856 + ], + "angle": 0, + "content": "We consider the task of urban navigation from point A to B [11]. The goal is to complete routes with dense traffic, multiple lanes and complex geometries (e.g. roundabouts) without incurring infractions. Along the way, the agent encounters manually designed pre-crash traffic scenarios. Each route is a list of GNSS coordinates called target points (TPs) which can be up to \\(50\\mathrm{m}\\) apart (\\(\\sim 30\\mathrm{m}\\) on average)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Metrics: We use the CARLA online metrics. Route Completion (RC) is the percentage of the route completed. Infraction Score (IS) is a penalty factor starting at 1.0 that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8241" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.348, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.355, + 0.312, + 0.369 + ], + "angle": 0, + "content": "(a) TransFuser (NC conditioned)" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.089, + 0.62, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.355, + 0.582, + 0.369 + ], + "angle": 0, + "content": "(b) TransFuser (TP conditioned)" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.892, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.355, + 0.844, + 0.369 + ], + "angle": 0, + "content": "(c) LAV [8] (TP conditioned)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.894, + 0.418 + ], + "angle": 0, + "content": "Figure 2: Extrapolation to target point. In unknown situations, TP conditioned methods extrapolate their waypoints towards target points. This periodically resets steering errors and is a form of implicit map based recovery. However, relying on extrapolation is a shortcut that can lead to catastrophic errors in certain situations (see Fig. 3a and Fig. 3b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.469, + 0.493 + ], + "angle": 0, + "content": "gets reduced multiplicatively for every infraction. Our main metric is the Driving Score (DS) which multiplies RC with the IS. Where insightful, we report infraction per kilometer metrics. For a comprehensive description, refer to [11]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.469, + 0.602 + ], + "angle": 0, + "content": "**Baseline:** As a baseline, we reproduce TransFuser [11]. This is a simple method representative of current SotA driving systems on CARLA. The main differences in our reproduction are a \\(360^{\\circ}\\) field of view (FOV) LiDAR to enable safe lane changes and a dataset where the expert is driving \\(2 \\times\\) faster to speed up evaluation. A list of other minor changes is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.469, + 0.724 + ], + "angle": 0, + "content": "Benchmark: For all experiments in this section, we train on CARLA towns 01, 03, 04, 06, 07 and 10. We use the 16 validation routes from LAV [8] in the withheld towns 02 and 05 for evaluation. In order to reduce the influence of training and evaluation variance which can be large in CARLA [4, 11], every experiment reports the average of 3 training seeds, evaluated 3 times each. We additionally report the training standard deviation for the main metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.292, + 0.756 + ], + "angle": 0, + "content": "3.1. A shortcut for recovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "While current SotA CARLA methods have fundamentally different architectures [8,11,26,31], they are all trained with conditional IL using fixed pre-recorded datasets. The evaluation task involves challenging routes which are \\(\\sim 1.5\\mathrm{km}\\) long, so it would be expected that these methods suffer from compounding errors, a well-known problem for IL [24]. Geometric shift augmentations [5] are a common approach to teach IL methods how to recover from such compounding steering errors [5, 14]. Surprisingly," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.478 + ], + "angle": 0, + "content": "even though such augmentations are not employed by SotA methods on CARLA, they report high RCs, demonstrating that they are not prone to this expected failure mode." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.893, + 0.736 + ], + "angle": 0, + "content": "All the aforementioned methods condition their predictions using the next target point (TP) along the route. To understand the importance of the TP, we train 2 versions of our reproduction of TransFuser: one with the original TP conditioning and another with a discrete (NC) condition. This is a one hot vector that indicates whether the car should follow the lane, turn right at the next intersection, etc. It contains no geometric information about the center of the lanes, but still removes the inherent task ambiguity of inner-city driving [13]. For the NC conditioned model, we also implement shift and rotation augmentations by collecting augmented frames during data collection. We deploy a second camera in the simulator, that changes its position and orientation randomly at every time-step. The output labels are transformed accordingly for training the model (implementation details and examples can be found in the supplementary material). The results are shown in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.75, + 0.84, + 0.822 + ], + "angle": 0, + "content": "
Cond.Aug.DS ↑RC ↑Dev ↓
NC-32 ± 856 ± 120.86
NC✓35 ± 354 ± 40.99
TP-39 ± 984 ± 70.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.833, + 0.842, + 0.848 + ], + "angle": 0, + "content": "Table 1: Conditioning and Augmentation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We observe a significant difference in RC of 28 points when switching between TP and NC conditioning. The TP conditioned model has 0 route deviations per km (\"Dev\")," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8242" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.348, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.154, + 0.231, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.355, + 0.309, + 0.368 + ], + "angle": 0, + "content": "(a) TCP [31] follows a shortcut." + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.089, + 0.619, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.144, + 0.619, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.355, + 0.585, + 0.368 + ], + "angle": 0, + "content": "(b) TransFuser follows a shortcut." + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.892, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.144, + 0.892, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.355, + 0.863, + 0.368 + ], + "angle": 0, + "content": "(c) A transformer decoder fixes this." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.893, + 0.403 + ], + "angle": 0, + "content": "Figure 3: Target point shortcut. When TP conditioned methods extrapolate to spatially distant waypoints, they incur large steering errors. Replacing global average pooling in TransFuser with a cross-attention mechanism mitigates the issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.473, + 0.751 + ], + "angle": 0, + "content": "which indicates that it never takes a wrong turn or drives too far away (more than \\(30\\mathrm{m}\\)) from the lane. However, this is not the case for NC conditioned models. While we do see a small improvement in DS with augmentation, its RC is still unsatisfactory. This indicates that the geometric information regarding the lane center available with the TP aids recovery in SotA IL methods, prompting further investigation. We inspect the TransFuser and LAV [8] models by constructing situations where the car is forcefully steered out of lane. Fig. 2 visualizes these scenarios with the camera, LiDAR and ground truth HD maps. For TP conditioned models, the predicted waypoints (shown in blue) extrapolate towards the nearest TP (red) even though the situation is out of distribution. This shows that one reason for their strong route following is that the bird's eye view (BEV) TP resets steering errors periodically: methods learn to steer towards nearby TPs because the expert trajectory in the dataset always goes through them. This has the effect that TP conditioned methods periodically drive back to the center of the lane, resetting any accumulated errors. Furthermore, it suggests that SotA IL approaches strongly rely on pre-specified geometric information regarding lane centers for recovery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We identify this as a form of shortcut learning, which can be useful when the car is close to a target point, but can also lead to catastrophic steering errors when the target point is further away. An example would be directly extrapolating to a target point behind a turn, which leads to cutting the turn. In Fig. 3a and Fig. 3b, we show instances where this happens for TCP [31] and TransFuser [11] in a validation town at nighttime. TransFuser also predicts the BEV segmentation as an auxiliary task, which we overlay" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.415, + 0.895, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.545, + 0.893, + 0.605 + ], + "angle": 0, + "content": "Figure 4: Pooling. Existing approaches vectorize feature grids either by global average pooling (top, e.g. [8, 11]) or with attention mechanisms (bottom, e.g. [26, 31]). The latter retains spatial information gained via auxiliary tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.713 + ], + "angle": 0, + "content": "in the figure (gray: route, yellow: lane marking). For TCP, we render the ground truth map and omit the LiDAR, since it does not use LiDAR and does not predict BEV segmentation. Both methods predict waypoints that are tilted towards the TP instead of following the street. As a result, they drive into the opposing lane. We refer to this as the TP shortcut." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.855, + 0.741 + ], + "angle": 0, + "content": "3.2. Improved pooling and data augmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.87 + ], + "angle": 0, + "content": "One design choice which differs across SotA architectures is the pooling applied between the encoder and decoder. In Fig. 4, we summarize these. In particular, TransFuser [11] and LAV [8] employ global average pooling (GAP) followed by an MLP. InterFuser [26] pools features via the cross-attention mechanism of a transformer decoder. Finally, TCP [31], which has two decoders, uses GAP for one decoder and attention-based pooling for the other." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 4, the networks are encouraged to learn spatially meaningful feature grids through auxiliary convo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8243" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "lutional decoders that predict outputs such as BEV semantic segmentation. However, GAP does not maintain the spatial information in the features. Fig. 3b shows the BEV road segmentation predicted by TransFuser overlaid by its Li-DAR input. Unlike the waypoints, the BEV predictions are quite accurate. This indicates strong features coming from the encoder. Nevertheless, the final conditional waypoint predictions focus on the TP signal in this situation. We hypothesize that the GAP operation makes it difficult for the downstream decoder to utilize the strong BEV features. Note that attention-based pooling, such as the transformer decoder of Interfuser, preserves spatial information through the use of positional encodings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.382 + ], + "angle": 0, + "content": "We compare the original TransFuser GAP approach with the spatially preserving design of Interfuser in Table 2. For the latter variant, we remove the GAP operation and MLP in TransFuser. We then process the \\(8 \\times 8\\) BEV feature grid as tokens with a transformer decoder. Implementation details can be found in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.394, + 0.462, + 0.468 + ], + "angle": 0, + "content": "
PoolingAug.DS ↑RC ↑Stat ↓
GAP + MLP-39 ± 984 ± 71.04
Transformer Decoder-43 ± 693 ± 30.55
Transformer Decoder✓49 ± 890 ± 40.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.477, + 0.399, + 0.492 + ], + "angle": 0, + "content": "Table 2: Pooling and augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.698 + ], + "angle": 0, + "content": "After replacing GAP with the transformer decoder, the RC increases by 9 points and collisions per km with static objects (\"Stat\") reduce by a factor of 2. Static objects (such as poles) only occur outside lanes in CARLA, implying that the network manages to stay in the lane far more consistently. Fig. 3c provides qualitative evidence that the transformer decoder can succeed in situations where GAP failed due to the TP shortcut (additional examples in supplementary). In addition, we investigate the inclusion of shift and rotation augmentations designed to aid lateral recovery (as described in Section 3.1). Collisions with the static environment are reduced further by a factor of 5 indicating that augmentation provides strong benefits." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.33, + 0.726 + ], + "angle": 0, + "content": "3.3. The ambiguity of waypoints" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Waypoints are used in many SotA systems as outputs [8, 11, 31]. They are obtained by recording an expert driver's GNSS locations at fixed time intervals (e.g. 0.5s) and transforming them into a local coordinate frame. The model is then trained to predict future waypoints, typically with an \\( L_{1} \\) regression objective. Waypoints entangle both the path and the future velocities of the vehicle. The velocity after a specific time interval extracted from the waypoints is used by a downstream controller as a target speed." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Fig. 5, we plot a histogram comparing the target speeds of the expert algorithm (i.e. training dataset) to those" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.858, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.311, + 0.868, + 0.326 + ], + "angle": 0, + "content": "Figure 5: Transfuser interpolates between modes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.46 + ], + "angle": 0, + "content": "extracted from TransFuser (best model in Table 2). Interestingly, the distributions are quite different. By design, the expert chooses one of four target speeds: 29, 18, 7, or 0 km/h. These values cover behaviors needed for city driving, corresponding to four situations: regular driving, slowing down in intersections, slowing down near pedestrians, and stopping. In contrast, for TransFuser, the predicted target speeds cover the entire 0-29 km/h range." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.893, + 0.718 + ], + "angle": 0, + "content": "While the future path that our expert follows is deterministic and unambiguous (center of the lane, lane change at pre-defined locations), future velocities are multi-modal. As both are jointly represented by waypoints, this leads to an ambiguous entangled representation. However, existing methods which utilize this entangled waypoint representation predict only point estimates (i.e., a single set of waypoints) as output, hence only a single mode is modeled. From Fig. 5, we observe that the waypoint based TransFuser model indeed interpolates between modes, a behavior that is expected when modeling multi-modal outputs deterministically. We illustrate this behavior in detail using the examples of (1) approaching an intersection with a green light (Fig. 6a), and (2) a cyclist cutting into the vehicle's path (Fig. 6b). TransFuser slows down in these situations since it is uncertain, e.g., the light may turn red in Fig. 6a. The car stops in time in Fig. 6b because of the decreased speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "While this averaging is indeed beneficial in some situations, an entangled representation is undesirable as it is less interpretable and does not explicitly expose uncertainty. Moreover, when stopped (all waypoints collapsed to one location), the steering signal is undefined which requires additional heuristics in the controller. To resolve this, we disentangle the future velocities from the path by sampling the expert's position at fixed distances instead of fixed time intervals for training a (deterministic) path predictor. However, when predicting the path instead of time-dependent trajectories, one requires an additional method to determine a target speed for the car. We propose to predict the tar" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8244" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.349, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.355, + 0.315, + 0.369 + ], + "angle": 0, + "content": "(a) Slowing down at a green light." + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.089, + 0.62, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.355, + 0.594, + 0.369 + ], + "angle": 0, + "content": "(b) Slowing down in an intersection." + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.089, + 0.892, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.355, + 0.876, + 0.369 + ], + "angle": 0, + "content": "(c) Disentangling route and target speed." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.894, + 0.418 + ], + "angle": 0, + "content": "Figure 6: Waypoints are ambiguous. The model's output representation forces it to predict a single mode for future velocities. There is a possibility that the traffic light might turn red in the future, or that the cyclist either cuts in or yields to the agent. The particular mode (or interpolation thereof) that the model converges to, depends on the training run and dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.471, + 0.524 + ], + "angle": 0, + "content": "get speed by simple classification using an additional MLP head. We incorporate the prediction uncertainties into the final output by using a confidence weighted average of the predicted target speed as input to the controller. Note that this is just one example of how the resulting confidence estimates may be leveraged by a vehicle controller." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.541, + 0.447, + 0.613 + ], + "angle": 0, + "content": "
OutputDS ↑RC ↑Veh ↓Stat ↓
Waypoints49 ± 890 ± 40.700.10
Path + Argmax40 ± 188 ± 21.250.03
Path + Weighted50 ± 388 ± 10.830.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.624, + 0.384, + 0.639 + ], + "angle": 0, + "content": "Table 3: Output representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Compared to the waypoint representation, naively converting the predicted classes to a target speed via the most likely class (Argmax) increases vehicle collisions (\"Veh\"), as seen in Table 3. Employing the proposed weighting (Weighted) achieves lower collisions. Moreover, the path-based model steers better as indicated by the lower static obstacle collisions (\"Stat\"). The disentangled representation achieves the same driving score as the entangled waypoint representation, providing an alternative with a simpler and more interpretable controller: (1) it eases design since it has identical parameters to the controller in the expert, (2) it has access to an unambiguous path representation even when driving slowly, and (3) it explicitly exposes and makes use of uncertainties with regard to target speeds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Fig. 6c shows a qualitative example. The car is slowing down due to its uncertainty. The lower target speed allows it successfully merge behind the cyclist." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.431, + 0.891, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.667, + 0.826, + 0.683 + ], + "angle": 0, + "content": "Figure 7: TransFuser++ architecture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.749, + 0.728 + ], + "angle": 0, + "content": "3.4. Scaling up to TransFuser++" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.737, + 0.892, + 0.799 + ], + "angle": 0, + "content": "By incorporating these insights, we obtain a significantly improved version of TransFuser [11] that we call TransFuser++ (Fig. 7). We now describe two important implementation changes compared to [11] that involve scaling." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.817, + 0.857, + 0.888 + ], + "angle": 0, + "content": "
Two stageFrozenDS ↑RC ↑Veh ↓
--50 ± 388 ± 10.83
✓-54 ± 192 ± 50.79
✓✓44 ± 697 ± 11.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.899, + 0.794, + 0.915 + ], + "angle": 0, + "content": "Table 4: Two stage training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8245" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "Training schedule: We observe no benefits from simply doubling the number of training epochs for a single training stage. However, in Table 4, we double the training time with a two-stage approach. First, we pre-train the encoder with only the perception losses (2D depth, 2D and BEV semantic segmentation, and vehicle bounding boxes, as shown in Fig. 7) for the usual number of epochs. We then fine-tune the resulting checkpoint with all losses, after including the GRU decoder and MLP classifier. Initializing the backbone with features that are pre-trained on the auxiliary tasks leads to a 4 DS improvement, consistent with the claims of LAV [8]. We also experiment with freezing the pre-trained backbone and only training the transformer decoder and its heads in the second stage. This leads to a drop of 10 DS, indicating that end-to-end optimization is important." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Dataset scale: Recent work on CARLA [23] shows large improvements by scaling the dataset size. However, [23] considers planning models with privileged inputs on training towns. We investigate the impact of scaling when evaluating on held-out validation towns for end-to-end models. We start with 185k training samples (as in [11]) and scale it up by re-running the training routes 3 times with different traffic (555k frames). Table 5 shows a clear improvement of 6 DS via scaling. This shows that current IL approaches can still benefit from larger datasets. We train for the same amount of epochs, which implies that this improvement comes at the cost of \\(3 \\times\\) longer training." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.514, + 0.409, + 0.574 + ], + "angle": 0, + "content": "
Dataset sizeDS ↑RC ↑Veh ↓
185k54 ± 192 ± 50.79
555k60 ± 698 ± 10.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.188, + 0.583, + 0.359, + 0.597 + ], + "angle": 0, + "content": "Table 5: Effects of scale." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.362, + 0.636 + ], + "angle": 0, + "content": "4. Comparison to State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Methods: This section demonstrates the advantages of \\(\\mathrm{TF}++\\) over several state-of-the-art models, which we list below, starting with the most recent. (1) Interfuser [26] processes multiple cameras and a BEV LiDAR by encoding each of them individually with a CNN. The resulting feature grids are fed into a transformer and decoded into perception outputs and the path to follow. The network does not predict longitudinal controls. Instead, it uses the BEV bounding boxes and a simple motion forecast (extrapolating historical dynamics) and linearly optimizes for the target speed using heuristically chosen objectives. (2) Perception PlanT [23] is a transformer based method trained in two independent stages that uses BEV bounding boxes as an intermediate representation. Its perception is based on TransFuser, and its planner outputs waypoints. (3) TCP [31] is a camera-only model with two output representations: waypoints and controls. During test time, it ensembles the two" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.379 + ], + "angle": 0, + "content": "outputs together with a weighted average that changes based on whether the vehicle is turning. (4) TransFuser [11] fuses perspective cameras with a BEV LiDAR by processing them with individual CNNs and exchanging features using transformers. It uses global average pooling and waypoint outputs, as described in Section 3. (5) LAV [8] processes cameras and LiDAR point clouds into an intermediate BEV representation. It trains a planner on this representation that predicts waypoints for the ego vehicle with a NC conditioned GRU followed by a TP conditioned refinement GRU. During training, the planner is also tasked to predict the trajectory of other surrounding vehicles to increase the number of labels. The authors release two versions of this method, which we call LAV v1 and LAV v2. (6) WOR [7] is an IL method for which the labels are enriched using a reward function to provide dense supervision for all possible actions. Unlike the other baselines, it is conditioned with the NC and does not utilize the TP. WOR is the best NC conditioned baseline that is publicly available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.895, + 0.73 + ], + "angle": 0, + "content": "**Benchmarks:** We use two benchmarks to evaluate on seen and unseen towns, with scenarios taken from [11] (type 1,3,4,7,8,9 and 10). Table 6 compares the performance of systems on the Longest6 benchmark [11] which consists of 36 routes in training towns 01 to 06. On Longest6, models trained on any data can be evaluated, so we compare against the author-provided models or directly report numbers from the respective papers. The mean and std of three evaluations is reported. For \\(\\mathrm{TF}++\\) we train 3 models and report the average result. Longest6 has the advantage that it evaluates in dense traffic and has diverse towns and routes. The drawbacks are that it does not penalize stop sign infractions or measure generalization to new towns. Therefore, we additionally compare performance on validation towns using the LAV [8] routes in Table 7. These are 4 routes with 4 weathers (16 combined) in Town 02 and 05, which are withheld during training. We re-train methods with the dataset from the corresponding paper 3 times and evaluate each seed 3 times. Reported results are the mean of all runs and the std between the training seeds. We compare against the reported SotA TCP and our reproduced TransFuser baseline on this benchmark and provide additional baselines in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.894, + 0.838 + ], + "angle": 0, + "content": "Dataset: For \\(\\mathrm{TF}++\\) we collect data on the same training routes as in [11] with an improved expert labeling algorithm (described in the supplementary material). We repeat this 3 times on the same routes with different traffic, as in [23]. For validation, we withhold Town 02 and 05 during training, else we train on all towns. In total, we train with 750k frames (550k when withholding Town 02 and 05)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Results: We start with the results in training towns (Table 6). The best NC conditioned method, WOR, has significantly lower RC than all TP conditioned systems (22 lower than LAV v1). In particular, the route deviations (Dev) are" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8246" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.858, + 0.251 + ], + "angle": 0, + "content": "
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Dev ↓TO ↓Block ↓
WOR [7]21 ± 348 ± 40.56 ± 0.030.181.050.371.280.880.080.20
LAV v1 [8]33 ± 170 ± 30.51 ± 0.020.160.830.150.960.060.120.45
Interfuser [26]47 ± 674 ± 10.63 ± 0.070.061.140.110.240.000.520.06
TransFuser [11]47 ± 693 ± 10.50 ± 0.060.032.450.070.160.000.060.10
TCP [31]48 ± 372 ± 30.65 ± 0.040.041.080.230.140.020.180.35
LAV v2 [8]58 ± 183 ± 10.68 ± 0.020.000.690.150.230.080.320.11
Perception PlanT [23]58 ± 588 ± 10.65 ± 0.060.070.970.110.090.000.130.13
TF++ (ours)69 ± 094 ± 20.72 ± 0.010.000.830.010.050.000.070.06
Expert81 ± 390 ± 10.91 ± 0.040.010.210.000.010.000.070.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.17, + 0.261, + 0.798, + 0.275 + ], + "angle": 0, + "content": "Table 6: Performance on training towns (Longest6). Released models, std over 3 evaluations." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.29, + 0.875, + 0.384 + ], + "angle": 0, + "content": "
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Stop ↓Dev ↓TO ↓Block ↓
TransFuser (ours)39 ± 984 ± 70.46 ± 0.060.000.741.040.201.070.000.230.21
TCP [31]58 ± 585 ± 30.67 ± 0.060.000.350.160.011.050.000.190.19
TF++ (ours)70 ± 699 ± 00.70 ± 0.060.010.630.010.040.260.000.050.00
Expert94950.990.000.020.000.020.000.000.000.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.393, + 0.869, + 0.409 + ], + "angle": 0, + "content": "Table 7: Performance on validation towns (LAV). Reproduced models, std over 3 trainings (for 3 evaluations each)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.614 + ], + "angle": 0, + "content": "10 times higher. Current TP conditioned SotA methods achieve (close to) 0 route deviations with driving scores in the range of 47 DS to 58 DS. TF++ outperforms all baselines on Longest6, showing a \\(19\\%\\) relative improvement over the previous SotA Perception PlanT. TF++ achieves close to expert-level performance on all infractions except for vehicle collisions (Veh). On the validation towns (Table 7) TF++ outperforms our reproduced TransFuser by 31 DS improving all metrics, particularly environmental collisions (Stat). It improves the prior SotA result TCP by \\(21\\%\\). Runtime: Table 8 compares the runtime of TF++ and TransFuser which have the same sensor setup. TF++ yields a large DS improvement while being only \\(14\\%\\) slower." + }, + { + "type": "table", + "bbox": [ + 0.138, + 0.627, + 0.411, + 0.686 + ], + "angle": 0, + "content": "
Method ↑DS ↑Time (ms) ↓
TransFuser (ours)3944
TF++ (ours)7050
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.757 + ], + "angle": 0, + "content": "Table 8: Runtime. We show the runtime per frame in ms averaged over 300 time steps on a single route on a RTX 3090. TF++ outperforms TransFuser by a wide margin despite using a similar compute budget during inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.196, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this work, we show that recent SotA driving models have exceptional route following abilities because they learn a strong bias towards following nearby TPs. Shortcut learning like this is a general phenomenon in deep neural networks [16] and has been observed in the context of autonomous driving for inputs such as velocity [14] or temporal frames [3, 30]. We add to this literature by observ-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.508 + ], + "angle": 0, + "content": "ing shortcut learning with respect to the conditioning signal. While shortcut learning usually has a negative impact on performance, we observe positive (improved recovery) and negative (cutting turns) effects. We show that the negative effects can be mitigated by avoiding global pooling and incorporating data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.644 + ], + "angle": 0, + "content": "A second commonality in SotA approaches is the use of waypoints as an output representation. We observe that this representation is ambiguous because it predicts a point estimate for multi-modal future velocities. We disambiguate the representation by disentangling future velocities from the deterministic path predictions and classifying target speeds instead. We then weigh target speeds according to their predicted confidence in our controller. Surprisingly, we find that interpolation is helpful for reducing collisions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.704 + ], + "angle": 0, + "content": "We propose TransFuser++ by improving the popular baseline TransFuser with a series of controlled experiments. TF++ is a simple end-to-end method that sets a new state of the art on the LAV and Longest6 benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Limitations: This study investigates urban driving in CARLA, where all investigated methods drive at relatively low speed \\((< 35\\mathrm{km / h})\\). Therefore, problems specific to high-speed driving are not considered. In addition, lanes are free of static obstacles, hence scenarios requiring navigation around them are not included. We point out the strong reliance of current methods on TPs for recovery. This implies a reliance on accurate localization and mapping to obtain these TPs. They are accurately mapped in CARLA, however, this assumption might not hold in real environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Broader Impact: We aim to make progress towards autonomous driving. This technology, if realized, could have massive societal impact, reducing road accidents, trans" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8247" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "portation costs and improving the mobility of elderly people. Potential negative implications include a reduction in jobs for human drivers and possible military applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.144, + 0.473, + 0.311 + ], + "angle": 0, + "content": "Acknowledgements: Andreas Geiger and Bernhard Jaeger were supported by the ERC Starting Grant LEGO-3D (850533), the BMWi in the project KI Delta Learning (project number 19A19013O) and the DFG EXC number 2064/1 - project number 390727645. Kashyap Chitta was supported by the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. We thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Bernhard Jaeger and Kashyap Chitta as well as Niklas Hanselmann and Katrin Renz for proofreading." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.324, + 0.174, + 0.339 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.348, + 0.468, + 0.377 + ], + "angle": 0, + "content": "[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.47, + 0.461 + ], + "angle": 0, + "content": "[2] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.463, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[3] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. In Proc. Robotics: Science and Systems (RSS), 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[4] Aseem Behl, Kashyap Chitta, Aditya Prakash, Eshed Ohn-Bar, and Andreas Geiger. Label efficient visual abstractions for autonomous driving. In Proc. IEEE International Conf. on Intelligent Robots and Systems (IROS), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.577, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[5] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba. End to end learning for self-driving cars. arXiv.org, 1604.07316, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.647, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[7] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.76, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[8] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[9] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Proc. Conf. on Robot Learning (CoRL), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[12] Kyunghyun Cho, Bart van Merrienboer, Caglar Güçehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proc. of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[13] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[14] Felipe Codevilla, Eder Santana, Antonio M. López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proc. Conf. on Robot Learning (CoRL), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[16] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[17] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zak Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[18] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[19] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[20] Dean Pomerleau. ALVINN: an autonomous land vehicle in a neural network. In Advances in Neural Information Processing Systems (NIPS), 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[21] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[22] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, Almut Sophia Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8248" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "representations. In Proc. Conf. on Robot Learning (CoRL), 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[24] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Conference on Artificial Intelligence and Statistics (AISTATS), 2011. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[25] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In Proc. Conf. on Robot Learning (CoRL), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[26] Hao Shao, Letian Wang, RuoBing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Proc. Conf. on Robot Learning (CoRL), 2022. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[27] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[28] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[30] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[31] Peng Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[32] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[33] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8249" + } + ] +] \ No newline at end of file diff --git a/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_origin.pdf b/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0b34f91343e507e4443b403ea108ec8726b1b47d --- /dev/null +++ b/2023/Hidden Biases of End-to-End Driving Models/8f4d684a-1555-4674-8f25-206ee2d4af90_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79b97b2a9f6f9d9637d921457c9b189fbe604eb970b08b73455c23472e4c20cb +size 4646203 diff --git a/2023/Hidden Biases of End-to-End Driving Models/full.md b/2023/Hidden Biases of End-to-End Driving Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c542beadc1ea4ac655e93647dc949a8a868083ee --- /dev/null +++ b/2023/Hidden Biases of End-to-End Driving Models/full.md @@ -0,0 +1,268 @@ +# Hidden Biases of End-to-End Driving Models + +Bernhard Jaeger Kashyap Chitta Andreas Geiger University of Tübingen Tübingen AI Center + +{bernhard.jaeger, kashyap.chitta, a.geiger}@uni-tuebingen.de + +# Abstract + +End-to-end driving systems have recently made rapid progress, in particular on CARLA. Independent of their major contribution, they introduce changes to minor system components. Consequently, the source of improvements is unclear. We identify two biases that recur in nearly all state-of-the-art methods and are critical for the observed progress on CARLA: (1) lateral recovery via a strong inductive bias towards target point following, and (2) longitudinal averaging of multimodal waypoint predictions for slowing down. We investigate the drawbacks of these biases and identify principled alternatives. By incorporating our insights, we develop $TF++$ , a simple end-to-end method that ranks first on the Longest6 and LAV benchmarks, gaining 11 driving score over the best prior work on Longest6. + +# 1. Introduction + +End-to-end driving approaches have rapidly improved in performance on the CARLA leaderboard [1], the de-facto standard for fair online evaluation. Driving scores have increased from under 20 [9,22] to over 70 [26,31] in just two years. However, why recent systems work so well is not fully understood, as both methods and training sets differ largely between submissions. Rigorous ablations are expensive due to the large design space of driving systems and the need to simulate large amounts of driving for evaluation. + +In particular, recent methods trained with Imitation Learning (IL) have shown strong performance [6, 8, 11, 26, 31]. They are trained using offline datasets, yet they can surprisingly recover from the classic compounding error problem of IL [21, 24], as indicated by their high route completions [8, 11, 26, 31]. While they do not utilize HD maps as input, they are provided with map-based GNSS locations in the center of the lane (spaced $30\mathrm{m}$ apart on average) called target points (TPs) that describe the route the car should follow. TPs were introduced as an alternative form of conditioning signals to convey driver intent [9, 22]. Prior work [13, 14] used discrete navigation commands or NCs (i.e. follow lane, turn right,...) instead. + +In this paper, we show that TP conditioned models re + +![](images/e3674815b82a038fb93a973e0c1fdc6dc96c7ee5055a195ecdd133c8db238274.jpg) +(a) Target point shortcut +Figure 1: Hidden biases. (a) When outside their training distribution, current methods extrapolate waypoint predictions to the nearest target point, helping them recover. (b) The future velocity is multi-modal, but current methods commit to a single plan, which leads to interpolation. + +![](images/968ffbfeaf7d76cb3925c655a3001bc3703236ddd87551990cc22b85a353e6bf.jpg) +(b) Waypoint ambiguity + +cover from the compounding error problem because they use geometric information contained in the TP to reset steering errors periodically (at every TP). This makes them implicitly rely on accurate map information, even though they are otherwise HD map free. Steering directly towards a TP is a shortcut [16] that these IL methods learn to exploit. When methods accumulate enough steering error to be out of distribution during deployment, we observe that they steer towards the nearest TP. When the TP is close, this has the effect of driving back to the lane center, where it is in distribution again. This is illustrated in Fig. 1a. However, when the TP is far away, this shortcut can lead to catastrophic steering errors (e.g. cutting a turn). We show examples of this behavior for various SotA architectures in Section 3.1. We demonstrate that the shortcut problem is intrinsically related to the decoder architecture and that a transformer decoder [29] can mitigate it. + +Another common aspect of the current SotA is that they use waypoints (future positions of an expert driver) as output representations [8, 11, 31]. We point out that this is an ambiguous representation as the future velocity is multi- + +modal, yet the model commits to a point estimate. This is illustrated in Fig. 1b. We show that this ambiguity can sometimes be helpful due to the continuous nature of waypoints: the network can continuously interpolate between modes. We propose an alternative that explicitly predicts the uncertainty of the network via target speed classification, and show that interpolating between target speeds weighted by the uncertainty reduces collisions. Our controlled experiments also cover important but sometimes neglected details in the training of end-to-end driving systems, such as augmentation, training schedules, and dataset size. In particular, we revisit the idea of shift and rotation augmentations to aid recovery [2,5]. These were common in early IL methods for CARLA with control outputs [14], but are harder to implement with waypoint outputs and not used by the current SotA. We find that they yield significant improvements. + +Using these insights, we develop TransFuser++ (TF++, which sets a new SotA on the Longest6 [11] and LAV [8] benchmarks. Two of the ideas we apply, transformer decoder pooling and path-based outputs instead of waypoints, have been used in Interfuser [26]. However, these are presented as minor details, and their impact is not studied in isolation as in our work. TF++ is significantly simpler than Interfuser, yet outperforms it by a large margin on Longest6, as we show in Section 4. We use $\sim 4\times$ less data, 1 camera instead of 4, and do not require complex heuristics to extract throttle and brake commands for our path output. + +# Contributions: + +- We show that target point conditioned models learn a shortcut that helps them recover from steering errors. +- We point out that the waypoint output representation is ambiguous, but its continuous nature helps models collide less by interpolating to slow down. +- Using the insights gained from controlled experiments, we propose TransFuser++ which places first on the Longest6 and LAV benchmarks. + +Code, data, and models are available at https://github.com/ autonomousvision/carla_garage. + +# 2. Related Work + +IL for autonomous driving dates back over 30 years [20]. It regained traction with the seminal works of [3, 5, 13] and the release of CARLA [15], a 3D simulator used in numerous recent research advances in autonomous driving [9, 10, 14, 17, 19, 25, 28, 33]. Early IL approaches evaluated in CARLA used a discrete navigation command (NC) [7, 9, 14], but their performance is not competitive to modern approaches [8, 11, 26, 31] which predict waypoints conditioned on TPs. In this work, we revisit an idea used in early systems but neglected in modern TP-conditioned methods: geometric shift augmentations [5, 9, 14, 21]. + +LAV [8] supervises waypoint outputs with additional data by making predictions for other nearby agents in the + +scene. Their waypoints are initially generated with NC conditioning. These are then refined using a GRU [12]. They observe a large (+50) improvement in route completion from this refinement. Our findings suggest that the refinement module improves steering primarily through TP conditioning, which is only provided to the refinement GRU. TCP [31] observes that waypoints are stronger at collision avoidance than directly predicting controls, but sometimes suffer at large turns. They leverage the strength and mitigate the weaknesses of these two representations by creating a situation dependent ensemble. Our study suggests that the waypoints are better at collision avoidance due to an implicit slowdown when the car is uncertain. PlanT [23] investigates planning on CARLA by processing object bounding boxes with a transformer. Its sensorimotor version differs from other SotA models in that it is trained in two stages (not end-to-end). In our setting, we find that end-to-end training is crucial. We show that their observations about dataset scale also hold for end-to-end models. + +TransFuser [11,22] is a simple, well-known and widely used baseline for CARLA. We provide an explanation for the large differences in route completion between the target point conditioned TransFuser architecture and its NC conditioned baselines [7,9,14]. Further, we propose modifications to its architecture, output representation and training strategy which lead to significant improvements. + +Interfuser [26] regresses a path for steering, predicts object density maps, and classifies traffic rule flags. This representation is converted by a forecasting mechanism and hand designed heuristics into control. Like our proposed method, it uses a transformer decoder for pooling features, and disentangles future velocities from the path in the output. While these ideas were already present, they were not studied or discussed as significant. Our work adds to the literature by showing that these design choices are indeed critical to performance and providing explanations as to why. Furthermore, our final system is simpler and significantly outperforms Interfuser on Longest6. + +ReasonNet [27], ThinkTwice [18] and CaT [32] are concurrent end-to-end driving approaches that all use TP conditioning and waypoints as output representation. + +# 3. Hidden Biases of End-to-End Driving + +We consider the task of urban navigation from point A to B [11]. The goal is to complete routes with dense traffic, multiple lanes and complex geometries (e.g. roundabouts) without incurring infractions. Along the way, the agent encounters manually designed pre-crash traffic scenarios. Each route is a list of GNSS coordinates called target points (TPs) which can be up to $50\mathrm{m}$ apart ( $\sim 30\mathrm{m}$ on average). + +Metrics: We use the CARLA online metrics. Route Completion (RC) is the percentage of the route completed. Infraction Score (IS) is a penalty factor starting at 1.0 that + +![](images/f73ae611e9a505aab52a4e2d51fd1ef542472da56a923283a1cd87bd29913bdd.jpg) +(a) TransFuser (NC conditioned) + +![](images/757d0036a0c9d88c7c99088a58efcdfe252f25d64716dbec9decaa89ebc34dba.jpg) +(b) TransFuser (TP conditioned) + +![](images/c9018ef2eb763ec683c5380c78a2d062c24ef8f9cdbcba8cc6ded3fe01c425d6.jpg) +(c) LAV [8] (TP conditioned) +Figure 2: Extrapolation to target point. In unknown situations, TP conditioned methods extrapolate their waypoints towards target points. This periodically resets steering errors and is a form of implicit map based recovery. However, relying on extrapolation is a shortcut that can lead to catastrophic errors in certain situations (see Fig. 3a and Fig. 3b). + +gets reduced multiplicatively for every infraction. Our main metric is the Driving Score (DS) which multiplies RC with the IS. Where insightful, we report infraction per kilometer metrics. For a comprehensive description, refer to [11]. + +**Baseline:** As a baseline, we reproduce TransFuser [11]. This is a simple method representative of current SotA driving systems on CARLA. The main differences in our reproduction are a $360^{\circ}$ field of view (FOV) LiDAR to enable safe lane changes and a dataset where the expert is driving $2 \times$ faster to speed up evaluation. A list of other minor changes is provided in the supplementary material. + +Benchmark: For all experiments in this section, we train on CARLA towns 01, 03, 04, 06, 07 and 10. We use the 16 validation routes from LAV [8] in the withheld towns 02 and 05 for evaluation. In order to reduce the influence of training and evaluation variance which can be large in CARLA [4, 11], every experiment reports the average of 3 training seeds, evaluated 3 times each. We additionally report the training standard deviation for the main metrics. + +# 3.1. A shortcut for recovery + +While current SotA CARLA methods have fundamentally different architectures [8,11,26,31], they are all trained with conditional IL using fixed pre-recorded datasets. The evaluation task involves challenging routes which are $\sim 1.5\mathrm{km}$ long, so it would be expected that these methods suffer from compounding errors, a well-known problem for IL [24]. Geometric shift augmentations [5] are a common approach to teach IL methods how to recover from such compounding steering errors [5, 14]. Surprisingly, + +even though such augmentations are not employed by SotA methods on CARLA, they report high RCs, demonstrating that they are not prone to this expected failure mode. + +All the aforementioned methods condition their predictions using the next target point (TP) along the route. To understand the importance of the TP, we train 2 versions of our reproduction of TransFuser: one with the original TP conditioning and another with a discrete (NC) condition. This is a one hot vector that indicates whether the car should follow the lane, turn right at the next intersection, etc. It contains no geometric information about the center of the lanes, but still removes the inherent task ambiguity of inner-city driving [13]. For the NC conditioned model, we also implement shift and rotation augmentations by collecting augmented frames during data collection. We deploy a second camera in the simulator, that changes its position and orientation randomly at every time-step. The output labels are transformed accordingly for training the model (implementation details and examples can be found in the supplementary material). The results are shown in Table 1. + +
Cond.Aug.DS ↑RC ↑Dev ↓
NC-32 ± 856 ± 120.86
NC✓35 ± 354 ± 40.99
TP-39 ± 984 ± 70.00
+ +Table 1: Conditioning and Augmentation. + +We observe a significant difference in RC of 28 points when switching between TP and NC conditioning. The TP conditioned model has 0 route deviations per km ("Dev"), + +![](images/545be28bd61414f2dd817fdb99088f2423a765d951c7a6f901d0e5d1db571516.jpg) + +![](images/531a2b62baf02b0ced74678761cfbd331f200b760a2278217625d3fd21888851.jpg) +(a) TCP [31] follows a shortcut. + +![](images/7b9b4a700097ebe9cd4815dcdfe60f590af54af6592135bab170194eed9f5ba1.jpg) + +![](images/882ff92bb508758e6a52f6ea756405da23d92cfe04f62ab4b41dcfc1ee83bc82.jpg) +(b) TransFuser follows a shortcut. + +![](images/6dae0f4b7ce141a1af37a3748af96f3c2d1348f9f011e3237ad2ca27f962ecb6.jpg) + +![](images/95bcd4da5c5bcfd8bf21868307d38123f417fde75dd2990c98a47239d39596a0.jpg) +(c) A transformer decoder fixes this. + +which indicates that it never takes a wrong turn or drives too far away (more than $30\mathrm{m}$ ) from the lane. However, this is not the case for NC conditioned models. While we do see a small improvement in DS with augmentation, its RC is still unsatisfactory. This indicates that the geometric information regarding the lane center available with the TP aids recovery in SotA IL methods, prompting further investigation. We inspect the TransFuser and LAV [8] models by constructing situations where the car is forcefully steered out of lane. Fig. 2 visualizes these scenarios with the camera, LiDAR and ground truth HD maps. For TP conditioned models, the predicted waypoints (shown in blue) extrapolate towards the nearest TP (red) even though the situation is out of distribution. This shows that one reason for their strong route following is that the bird's eye view (BEV) TP resets steering errors periodically: methods learn to steer towards nearby TPs because the expert trajectory in the dataset always goes through them. This has the effect that TP conditioned methods periodically drive back to the center of the lane, resetting any accumulated errors. Furthermore, it suggests that SotA IL approaches strongly rely on pre-specified geometric information regarding lane centers for recovery. + +We identify this as a form of shortcut learning, which can be useful when the car is close to a target point, but can also lead to catastrophic steering errors when the target point is further away. An example would be directly extrapolating to a target point behind a turn, which leads to cutting the turn. In Fig. 3a and Fig. 3b, we show instances where this happens for TCP [31] and TransFuser [11] in a validation town at nighttime. TransFuser also predicts the BEV segmentation as an auxiliary task, which we overlay + +![](images/0d3fbcfa66a60c99ada9e90fe765c072327ce6e055601f2432306eb08e192d57.jpg) +Figure 3: Target point shortcut. When TP conditioned methods extrapolate to spatially distant waypoints, they incur large steering errors. Replacing global average pooling in TransFuser with a cross-attention mechanism mitigates the issue. +Figure 4: Pooling. Existing approaches vectorize feature grids either by global average pooling (top, e.g. [8, 11]) or with attention mechanisms (bottom, e.g. [26, 31]). The latter retains spatial information gained via auxiliary tasks. + +in the figure (gray: route, yellow: lane marking). For TCP, we render the ground truth map and omit the LiDAR, since it does not use LiDAR and does not predict BEV segmentation. Both methods predict waypoints that are tilted towards the TP instead of following the street. As a result, they drive into the opposing lane. We refer to this as the TP shortcut. + +# 3.2. Improved pooling and data augmentation + +One design choice which differs across SotA architectures is the pooling applied between the encoder and decoder. In Fig. 4, we summarize these. In particular, TransFuser [11] and LAV [8] employ global average pooling (GAP) followed by an MLP. InterFuser [26] pools features via the cross-attention mechanism of a transformer decoder. Finally, TCP [31], which has two decoders, uses GAP for one decoder and attention-based pooling for the other. + +As shown in Fig. 4, the networks are encouraged to learn spatially meaningful feature grids through auxiliary convo + +lutional decoders that predict outputs such as BEV semantic segmentation. However, GAP does not maintain the spatial information in the features. Fig. 3b shows the BEV road segmentation predicted by TransFuser overlaid by its Li-DAR input. Unlike the waypoints, the BEV predictions are quite accurate. This indicates strong features coming from the encoder. Nevertheless, the final conditional waypoint predictions focus on the TP signal in this situation. We hypothesize that the GAP operation makes it difficult for the downstream decoder to utilize the strong BEV features. Note that attention-based pooling, such as the transformer decoder of Interfuser, preserves spatial information through the use of positional encodings. + +We compare the original TransFuser GAP approach with the spatially preserving design of Interfuser in Table 2. For the latter variant, we remove the GAP operation and MLP in TransFuser. We then process the $8 \times 8$ BEV feature grid as tokens with a transformer decoder. Implementation details can be found in the supplementary material. + +
PoolingAug.DS ↑RC ↑Stat ↓
GAP + MLP-39 ± 984 ± 71.04
Transformer Decoder-43 ± 693 ± 30.55
Transformer Decoder✓49 ± 890 ± 40.10
+ +Table 2: Pooling and augmentation. + +After replacing GAP with the transformer decoder, the RC increases by 9 points and collisions per km with static objects ("Stat") reduce by a factor of 2. Static objects (such as poles) only occur outside lanes in CARLA, implying that the network manages to stay in the lane far more consistently. Fig. 3c provides qualitative evidence that the transformer decoder can succeed in situations where GAP failed due to the TP shortcut (additional examples in supplementary). In addition, we investigate the inclusion of shift and rotation augmentations designed to aid lateral recovery (as described in Section 3.1). Collisions with the static environment are reduced further by a factor of 5 indicating that augmentation provides strong benefits. + +# 3.3. The ambiguity of waypoints + +Waypoints are used in many SotA systems as outputs [8, 11, 31]. They are obtained by recording an expert driver's GNSS locations at fixed time intervals (e.g. 0.5s) and transforming them into a local coordinate frame. The model is then trained to predict future waypoints, typically with an $L_{1}$ regression objective. Waypoints entangle both the path and the future velocities of the vehicle. The velocity after a specific time interval extracted from the waypoints is used by a downstream controller as a target speed. + +In Fig. 5, we plot a histogram comparing the target speeds of the expert algorithm (i.e. training dataset) to those + +![](images/7496a29c2d4fffcb9f57a32a1cfafa0bfdd5ea9958471d6b60b2c2d80d5686e5.jpg) +Figure 5: Transfuser interpolates between modes. + +extracted from TransFuser (best model in Table 2). Interestingly, the distributions are quite different. By design, the expert chooses one of four target speeds: 29, 18, 7, or 0 km/h. These values cover behaviors needed for city driving, corresponding to four situations: regular driving, slowing down in intersections, slowing down near pedestrians, and stopping. In contrast, for TransFuser, the predicted target speeds cover the entire 0-29 km/h range. + +While the future path that our expert follows is deterministic and unambiguous (center of the lane, lane change at pre-defined locations), future velocities are multi-modal. As both are jointly represented by waypoints, this leads to an ambiguous entangled representation. However, existing methods which utilize this entangled waypoint representation predict only point estimates (i.e., a single set of waypoints) as output, hence only a single mode is modeled. From Fig. 5, we observe that the waypoint based TransFuser model indeed interpolates between modes, a behavior that is expected when modeling multi-modal outputs deterministically. We illustrate this behavior in detail using the examples of (1) approaching an intersection with a green light (Fig. 6a), and (2) a cyclist cutting into the vehicle's path (Fig. 6b). TransFuser slows down in these situations since it is uncertain, e.g., the light may turn red in Fig. 6a. The car stops in time in Fig. 6b because of the decreased speed. + +While this averaging is indeed beneficial in some situations, an entangled representation is undesirable as it is less interpretable and does not explicitly expose uncertainty. Moreover, when stopped (all waypoints collapsed to one location), the steering signal is undefined which requires additional heuristics in the controller. To resolve this, we disentangle the future velocities from the path by sampling the expert's position at fixed distances instead of fixed time intervals for training a (deterministic) path predictor. However, when predicting the path instead of time-dependent trajectories, one requires an additional method to determine a target speed for the car. We propose to predict the tar + +![](images/6f92f0a9f4dcb7a9453dc9387e8aae22342b37524a874d13ee277f88c6c6cbfb.jpg) +(a) Slowing down at a green light. + +![](images/51c9dde26c5bd6454eab8b33693d7623c562f53a5bdfe26c9c669bc55ac5929d.jpg) +(b) Slowing down in an intersection. + +![](images/f295c149d94641124f7afc384c7437a83b1a86a80d044ff403ce9821b08a4157.jpg) +(c) Disentangling route and target speed. + +get speed by simple classification using an additional MLP head. We incorporate the prediction uncertainties into the final output by using a confidence weighted average of the predicted target speed as input to the controller. Note that this is just one example of how the resulting confidence estimates may be leveraged by a vehicle controller. + +
OutputDS ↑RC ↑Veh ↓Stat ↓
Waypoints49 ± 890 ± 40.700.10
Path + Argmax40 ± 188 ± 21.250.03
Path + Weighted50 ± 388 ± 10.830.02
+ +Compared to the waypoint representation, naively converting the predicted classes to a target speed via the most likely class (Argmax) increases vehicle collisions ("Veh"), as seen in Table 3. Employing the proposed weighting (Weighted) achieves lower collisions. Moreover, the path-based model steers better as indicated by the lower static obstacle collisions ("Stat"). The disentangled representation achieves the same driving score as the entangled waypoint representation, providing an alternative with a simpler and more interpretable controller: (1) it eases design since it has identical parameters to the controller in the expert, (2) it has access to an unambiguous path representation even when driving slowly, and (3) it explicitly exposes and makes use of uncertainties with regard to target speeds. + +Fig. 6c shows a qualitative example. The car is slowing down due to its uncertainty. The lower target speed allows it successfully merge behind the cyclist. + +![](images/4de65413a2c6ceebd9f2de56d54ee791b05faf6cfcd44c2ea1b5ada5bcc31d27.jpg) +Figure 6: Waypoints are ambiguous. The model's output representation forces it to predict a single mode for future velocities. There is a possibility that the traffic light might turn red in the future, or that the cyclist either cuts in or yields to the agent. The particular mode (or interpolation thereof) that the model converges to, depends on the training run and dataset. +Figure 7: TransFuser++ architecture. + +# 3.4. Scaling up to TransFuser++ + +By incorporating these insights, we obtain a significantly improved version of TransFuser [11] that we call TransFuser++ (Fig. 7). We now describe two important implementation changes compared to [11] that involve scaling. + +Table 3: Output representation. + +
Two stageFrozenDS ↑RC ↑Veh ↓
--50 ± 388 ± 10.83
✓-54 ± 192 ± 50.79
✓✓44 ± 697 ± 11.21
+ +Table 4: Two stage training. + +Training schedule: We observe no benefits from simply doubling the number of training epochs for a single training stage. However, in Table 4, we double the training time with a two-stage approach. First, we pre-train the encoder with only the perception losses (2D depth, 2D and BEV semantic segmentation, and vehicle bounding boxes, as shown in Fig. 7) for the usual number of epochs. We then fine-tune the resulting checkpoint with all losses, after including the GRU decoder and MLP classifier. Initializing the backbone with features that are pre-trained on the auxiliary tasks leads to a 4 DS improvement, consistent with the claims of LAV [8]. We also experiment with freezing the pre-trained backbone and only training the transformer decoder and its heads in the second stage. This leads to a drop of 10 DS, indicating that end-to-end optimization is important. + +Dataset scale: Recent work on CARLA [23] shows large improvements by scaling the dataset size. However, [23] considers planning models with privileged inputs on training towns. We investigate the impact of scaling when evaluating on held-out validation towns for end-to-end models. We start with 185k training samples (as in [11]) and scale it up by re-running the training routes 3 times with different traffic (555k frames). Table 5 shows a clear improvement of 6 DS via scaling. This shows that current IL approaches can still benefit from larger datasets. We train for the same amount of epochs, which implies that this improvement comes at the cost of $3 \times$ longer training. + +
Dataset sizeDS ↑RC ↑Veh ↓
185k54 ± 192 ± 50.79
555k60 ± 698 ± 10.73
+ +Table 5: Effects of scale. + +# 4. Comparison to State of the Art + +Methods: This section demonstrates the advantages of $\mathrm{TF}++$ over several state-of-the-art models, which we list below, starting with the most recent. (1) Interfuser [26] processes multiple cameras and a BEV LiDAR by encoding each of them individually with a CNN. The resulting feature grids are fed into a transformer and decoded into perception outputs and the path to follow. The network does not predict longitudinal controls. Instead, it uses the BEV bounding boxes and a simple motion forecast (extrapolating historical dynamics) and linearly optimizes for the target speed using heuristically chosen objectives. (2) Perception PlanT [23] is a transformer based method trained in two independent stages that uses BEV bounding boxes as an intermediate representation. Its perception is based on TransFuser, and its planner outputs waypoints. (3) TCP [31] is a camera-only model with two output representations: waypoints and controls. During test time, it ensembles the two + +outputs together with a weighted average that changes based on whether the vehicle is turning. (4) TransFuser [11] fuses perspective cameras with a BEV LiDAR by processing them with individual CNNs and exchanging features using transformers. It uses global average pooling and waypoint outputs, as described in Section 3. (5) LAV [8] processes cameras and LiDAR point clouds into an intermediate BEV representation. It trains a planner on this representation that predicts waypoints for the ego vehicle with a NC conditioned GRU followed by a TP conditioned refinement GRU. During training, the planner is also tasked to predict the trajectory of other surrounding vehicles to increase the number of labels. The authors release two versions of this method, which we call LAV v1 and LAV v2. (6) WOR [7] is an IL method for which the labels are enriched using a reward function to provide dense supervision for all possible actions. Unlike the other baselines, it is conditioned with the NC and does not utilize the TP. WOR is the best NC conditioned baseline that is publicly available. + +**Benchmarks:** We use two benchmarks to evaluate on seen and unseen towns, with scenarios taken from [11] (type 1,3,4,7,8,9 and 10). Table 6 compares the performance of systems on the Longest6 benchmark [11] which consists of 36 routes in training towns 01 to 06. On Longest6, models trained on any data can be evaluated, so we compare against the author-provided models or directly report numbers from the respective papers. The mean and std of three evaluations is reported. For $\mathrm{TF}++$ we train 3 models and report the average result. Longest6 has the advantage that it evaluates in dense traffic and has diverse towns and routes. The drawbacks are that it does not penalize stop sign infractions or measure generalization to new towns. Therefore, we additionally compare performance on validation towns using the LAV [8] routes in Table 7. These are 4 routes with 4 weathers (16 combined) in Town 02 and 05, which are withheld during training. We re-train methods with the dataset from the corresponding paper 3 times and evaluate each seed 3 times. Reported results are the mean of all runs and the std between the training seeds. We compare against the reported SotA TCP and our reproduced TransFuser baseline on this benchmark and provide additional baselines in the supplementary. + +Dataset: For $\mathrm{TF}++$ we collect data on the same training routes as in [11] with an improved expert labeling algorithm (described in the supplementary material). We repeat this 3 times on the same routes with different traffic, as in [23]. For validation, we withhold Town 02 and 05 during training, else we train on all towns. In total, we train with 750k frames (550k when withholding Town 02 and 05). + +Results: We start with the results in training towns (Table 6). The best NC conditioned method, WOR, has significantly lower RC than all TP conditioned systems (22 lower than LAV v1). In particular, the route deviations (Dev) are + +
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Dev ↓TO ↓Block ↓
WOR [7]21 ± 348 ± 40.56 ± 0.030.181.050.371.280.880.080.20
LAV v1 [8]33 ± 170 ± 30.51 ± 0.020.160.830.150.960.060.120.45
Interfuser [26]47 ± 674 ± 10.63 ± 0.070.061.140.110.240.000.520.06
TransFuser [11]47 ± 693 ± 10.50 ± 0.060.032.450.070.160.000.060.10
TCP [31]48 ± 372 ± 30.65 ± 0.040.041.080.230.140.020.180.35
LAV v2 [8]58 ± 183 ± 10.68 ± 0.020.000.690.150.230.080.320.11
Perception PlanT [23]58 ± 588 ± 10.65 ± 0.060.070.970.110.090.000.130.13
TF++ (ours)69 ± 094 ± 20.72 ± 0.010.000.830.010.050.000.070.06
Expert81 ± 390 ± 10.91 ± 0.040.010.210.000.010.000.070.09
+ +Table 6: Performance on training towns (Longest6). Released models, std over 3 evaluations. + +
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Stop ↓Dev ↓TO ↓Block ↓
TransFuser (ours)39 ± 984 ± 70.46 ± 0.060.000.741.040.201.070.000.230.21
TCP [31]58 ± 585 ± 30.67 ± 0.060.000.350.160.011.050.000.190.19
TF++ (ours)70 ± 699 ± 00.70 ± 0.060.010.630.010.040.260.000.050.00
Expert94950.990.000.020.000.020.000.000.000.08
+ +10 times higher. Current TP conditioned SotA methods achieve (close to) 0 route deviations with driving scores in the range of 47 DS to 58 DS. TF++ outperforms all baselines on Longest6, showing a $19\%$ relative improvement over the previous SotA Perception PlanT. TF++ achieves close to expert-level performance on all infractions except for vehicle collisions (Veh). On the validation towns (Table 7) TF++ outperforms our reproduced TransFuser by 31 DS improving all metrics, particularly environmental collisions (Stat). It improves the prior SotA result TCP by $21\%$ . Runtime: Table 8 compares the runtime of TF++ and TransFuser which have the same sensor setup. TF++ yields a large DS improvement while being only $14\%$ slower. + +Table 7: Performance on validation towns (LAV). Reproduced models, std over 3 trainings (for 3 evaluations each). + +
Method ↑DS ↑Time (ms) ↓
TransFuser (ours)3944
TF++ (ours)7050
+ +Table 8: Runtime. We show the runtime per frame in ms averaged over 300 time steps on a single route on a RTX 3090. TF++ outperforms TransFuser by a wide margin despite using a similar compute budget during inference. + +# 5. Conclusion + +In this work, we show that recent SotA driving models have exceptional route following abilities because they learn a strong bias towards following nearby TPs. Shortcut learning like this is a general phenomenon in deep neural networks [16] and has been observed in the context of autonomous driving for inputs such as velocity [14] or temporal frames [3, 30]. We add to this literature by observ- + +ing shortcut learning with respect to the conditioning signal. While shortcut learning usually has a negative impact on performance, we observe positive (improved recovery) and negative (cutting turns) effects. We show that the negative effects can be mitigated by avoiding global pooling and incorporating data augmentation. + +A second commonality in SotA approaches is the use of waypoints as an output representation. We observe that this representation is ambiguous because it predicts a point estimate for multi-modal future velocities. We disambiguate the representation by disentangling future velocities from the deterministic path predictions and classifying target speeds instead. We then weigh target speeds according to their predicted confidence in our controller. Surprisingly, we find that interpolation is helpful for reducing collisions. + +We propose TransFuser++ by improving the popular baseline TransFuser with a series of controlled experiments. TF++ is a simple end-to-end method that sets a new state of the art on the LAV and Longest6 benchmarks. + +Limitations: This study investigates urban driving in CARLA, where all investigated methods drive at relatively low speed $(< 35\mathrm{km / h})$ . Therefore, problems specific to high-speed driving are not considered. In addition, lanes are free of static obstacles, hence scenarios requiring navigation around them are not included. We point out the strong reliance of current methods on TPs for recovery. This implies a reliance on accurate localization and mapping to obtain these TPs. They are accurately mapped in CARLA, however, this assumption might not hold in real environments. + +Broader Impact: We aim to make progress towards autonomous driving. This technology, if realized, could have massive societal impact, reducing road accidents, trans + +portation costs and improving the mobility of elderly people. Potential negative implications include a reduction in jobs for human drivers and possible military applications. + +Acknowledgements: Andreas Geiger and Bernhard Jaeger were supported by the ERC Starting Grant LEGO-3D (850533), the BMWi in the project KI Delta Learning (project number 19A19013O) and the DFG EXC number 2064/1 - project number 390727645. Kashyap Chitta was supported by the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. We thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Bernhard Jaeger and Kashyap Chitta as well as Niklas Hanselmann and Katrin Renz for proofreading. + +# References + +[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020.1 +[2] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 2 +[3] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. In Proc. Robotics: Science and Systems (RSS), 2019. 2, 8 +[4] Aseem Behl, Kashyap Chitta, Aditya Prakash, Eshed Ohn-Bar, and Andreas Geiger. Label efficient visual abstractions for autonomous driving. In Proc. IEEE International Conf. on Intelligent Robots and Systems (IROS), 2020. 3 +[5] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba. End to end learning for self-driving cars. arXiv.org, 1604.07316, 2016. 2, 3 +[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[7] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2, 7, 8 +[8] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 7, 8 +[9] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Proc. Conf. on Robot Learning (CoRL), 2019. 1, 2 +[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2 + +[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[12] Kyunghyun Cho, Bart van Merrienboer, Caglar Güçehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proc. of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 2 +[13] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2018. 1, 2, 3 +[14] Felipe Codevilla, Eder Santana, Antonio M. López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 8 +[15] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proc. Conf. on Robot Learning (CoRL), 2017. 2 +[16] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1, 8 +[17] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zak Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[18] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[19] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[20] Dean Pomerleau. ALVINN: an autonomous land vehicle in a neural network. In Advances in Neural Information Processing Systems (NIPS), 1988. 2 +[21] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2 +[22] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[23] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, Almut Sophia Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level + +representations. In Proc. Conf. on Robot Learning (CoRL), 2022. 2, 7, 8 +[24] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Conference on Artificial Intelligence and Statistics (AISTATS), 2011. 1, 3 +[25] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In Proc. Conf. on Robot Learning (CoRL), 2018. 2 +[26] Hao Shao, Letian Wang, RuoBing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Proc. Conf. on Robot Learning (CoRL), 2022. 1, 2, 3, 4, 7, 8 +[27] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[28] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 1 +[30] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 8 +[31] Peng Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 4, 5, 7, 8 +[32] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[33] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2 \ No newline at end of file diff --git a/2023/Hidden Biases of End-to-End Driving Models/images.zip b/2023/Hidden Biases of End-to-End Driving Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2f4cfa562cee016f5299333d5b31a482c1cc9994 --- /dev/null +++ b/2023/Hidden Biases of End-to-End Driving Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c562d22e7a8d8f092633df8468e3fc767d638e0412c29b4311d34144a38c0536 +size 504892 diff --git a/2023/Hidden Biases of End-to-End Driving Models/layout.json b/2023/Hidden Biases of End-to-End Driving Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4e7fd61c29cd373a4d38ad19ef212349865fa9d3 --- /dev/null +++ b/2023/Hidden Biases of End-to-End Driving Models/layout.json @@ -0,0 +1,6536 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 103, + 438, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 103, + 438, + 121 + ], + "spans": [ + { + "bbox": [ + 156, + 103, + 438, + 121 + ], + "type": "text", + "content": "Hidden Biases of End-to-End Driving Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 166, + 143, + 427, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 143, + 427, + 172 + ], + "spans": [ + { + "bbox": [ + 166, + 143, + 427, + 172 + ], + "type": "text", + "content": "Bernhard Jaeger Kashyap Chitta Andreas Geiger University of Tübingen Tübingen AI Center" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 174, + 459, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 174, + 459, + 186 + ], + "spans": [ + { + "bbox": [ + 135, + 174, + 459, + 186 + ], + "type": "text", + "content": "{bernhard.jaeger, kashyap.chitta, a.geiger}@uni-tuebingen.de" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "type": "text", + "content": "End-to-end driving systems have recently made rapid progress, in particular on CARLA. Independent of their major contribution, they introduce changes to minor system components. Consequently, the source of improvements is unclear. We identify two biases that recur in nearly all state-of-the-art methods and are critical for the observed progress on CARLA: (1) lateral recovery via a strong inductive bias towards target point following, and (2) longitudinal averaging of multimodal waypoint predictions for slowing down. We investigate the drawbacks of these biases and identify principled alternatives. By incorporating our insights, we develop " + }, + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "type": "inline_equation", + "content": "TF++" + }, + { + "bbox": [ + 46, + 238, + 290, + 406 + ], + "type": "text", + "content": ", a simple end-to-end method that ranks first on the Longest6 and LAV benchmarks, gaining 11 driving score over the best prior work on Longest6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 545 + ], + "type": "text", + "content": "End-to-end driving approaches have rapidly improved in performance on the CARLA leaderboard [1], the de-facto standard for fair online evaluation. Driving scores have increased from under 20 [9,22] to over 70 [26,31] in just two years. However, why recent systems work so well is not fully understood, as both methods and training sets differ largely between submissions. Rigorous ablations are expensive due to the large design space of driving systems and the need to simulate large amounts of driving for evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": "In particular, recent methods trained with Imitation Learning (IL) have shown strong performance [6, 8, 11, 26, 31]. They are trained using offline datasets, yet they can surprisingly recover from the classic compounding error problem of IL [21, 24], as indicated by their high route completions [8, 11, 26, 31]. While they do not utilize HD maps as input, they are provided with map-based GNSS locations in the center of the lane (spaced " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": " apart on average) called target points (TPs) that describe the route the car should follow. TPs were introduced as an alternative form of conditioning signals to convey driver intent [9, 22]. Prior work [13, 14] used discrete navigation commands or NCs (i.e. follow lane, turn right,...) instead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we show that TP conditioned models re" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 211, + 423, + 373 + ], + "blocks": [ + { + "bbox": [ + 307, + 211, + 423, + 373 + ], + "lines": [ + { + "bbox": [ + 307, + 211, + 423, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 423, + 373 + ], + "type": "image", + "image_path": "e3674815b82a038fb93a973e0c1fdc6dc96c7ee5055a195ecdd133c8db238274.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 376, + 409, + 387 + ], + "lines": [ + { + "bbox": [ + 320, + 376, + 409, + 387 + ], + "spans": [ + { + "bbox": [ + 320, + 376, + 409, + 387 + ], + "type": "text", + "content": "(a) Target point shortcut" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 391, + 547, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 391, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 547, + 452 + ], + "type": "text", + "content": "Figure 1: Hidden biases. (a) When outside their training distribution, current methods extrapolate waypoint predictions to the nearest target point, helping them recover. (b) The future velocity is multi-modal, but current methods commit to a single plan, which leads to interpolation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 427, + 211, + 544, + 372 + ], + "blocks": [ + { + "bbox": [ + 427, + 211, + 544, + 372 + ], + "lines": [ + { + "bbox": [ + 427, + 211, + 544, + 372 + ], + "spans": [ + { + "bbox": [ + 427, + 211, + 544, + 372 + ], + "type": "image", + "image_path": "968ffbfeaf7d76cb3925c655a3001bc3703236ddd87551990cc22b85a353e6bf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 376, + 529, + 387 + ], + "lines": [ + { + "bbox": [ + 441, + 376, + 529, + 387 + ], + "spans": [ + { + "bbox": [ + 441, + 376, + 529, + 387 + ], + "type": "text", + "content": "(b) Waypoint ambiguity" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 460, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 664 + ], + "type": "text", + "content": "cover from the compounding error problem because they use geometric information contained in the TP to reset steering errors periodically (at every TP). This makes them implicitly rely on accurate map information, even though they are otherwise HD map free. Steering directly towards a TP is a shortcut [16] that these IL methods learn to exploit. When methods accumulate enough steering error to be out of distribution during deployment, we observe that they steer towards the nearest TP. When the TP is close, this has the effect of driving back to the lane center, where it is in distribution again. This is illustrated in Fig. 1a. However, when the TP is far away, this shortcut can lead to catastrophic steering errors (e.g. cutting a turn). We show examples of this behavior for various SotA architectures in Section 3.1. We demonstrate that the shortcut problem is intrinsically related to the decoder architecture and that a transformer decoder [29] can mitigate it." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Another common aspect of the current SotA is that they use waypoints (future positions of an expert driver) as output representations [8, 11, 31]. We point out that this is an ambiguous representation as the future velocity is multi-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8240" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "modal, yet the model commits to a point estimate. This is illustrated in Fig. 1b. We show that this ambiguity can sometimes be helpful due to the continuous nature of waypoints: the network can continuously interpolate between modes. We propose an alternative that explicitly predicts the uncertainty of the network via target speed classification, and show that interpolating between target speeds weighted by the uncertainty reduces collisions. Our controlled experiments also cover important but sometimes neglected details in the training of end-to-end driving systems, such as augmentation, training schedules, and dataset size. In particular, we revisit the idea of shift and rotation augmentations to aid recovery [2,5]. These were common in early IL methods for CARLA with control outputs [14], but are harder to implement with waypoint outputs and not used by the current SotA. We find that they yield significant improvements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 396 + ], + "type": "text", + "content": "Using these insights, we develop TransFuser++ (TF++, which sets a new SotA on the Longest6 [11] and LAV [8] benchmarks. Two of the ideas we apply, transformer decoder pooling and path-based outputs instead of waypoints, have been used in Interfuser [26]. However, these are presented as minor details, and their impact is not studied in isolation as in our work. TF++ is significantly simpler than Interfuser, yet outperforms it by a large margin on Longest6, as we show in Section 4. We use " + }, + { + "bbox": [ + 46, + 264, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\sim 4\\times" + }, + { + "bbox": [ + 46, + 264, + 289, + 396 + ], + "type": "text", + "content": " less data, 1 camera instead of 4, and do not require complex heuristics to extract throttle and brake commands for our path output." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 396, + 113, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 113, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 113, + 406 + ], + "type": "text", + "content": "Contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 407, + 286, + 502 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 407, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 407, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 58, + 407, + 286, + 430 + ], + "type": "text", + "content": "- We show that target point conditioned models learn a shortcut that helps them recover from steering errors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 431, + 286, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 431, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 58, + 431, + 286, + 466 + ], + "type": "text", + "content": "- We point out that the waypoint output representation is ambiguous, but its continuous nature helps models collide less by interpolating to slow down." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 467, + 286, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 467, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 58, + 467, + 286, + 502 + ], + "type": "text", + "content": "- Using the insights gained from controlled experiments, we propose TransFuser++ which places first on the Longest6 and LAV benchmarks." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 527 + ], + "type": "text", + "content": "Code, data, and models are available at https://github.com/ autonomousvision/carla_garage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 538, + 134, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 134, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 134, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "content": "IL for autonomous driving dates back over 30 years [20]. It regained traction with the seminal works of [3, 5, 13] and the release of CARLA [15], a 3D simulator used in numerous recent research advances in autonomous driving [9, 10, 14, 17, 19, 25, 28, 33]. Early IL approaches evaluated in CARLA used a discrete navigation command (NC) [7, 9, 14], but their performance is not competitive to modern approaches [8, 11, 26, 31] which predict waypoints conditioned on TPs. In this work, we revisit an idea used in early systems but neglected in modern TP-conditioned methods: geometric shift augmentations [5, 9, 14, 21]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "LAV [8] supervises waypoint outputs with additional data by making predictions for other nearby agents in the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "content": "scene. Their waypoints are initially generated with NC conditioning. These are then refined using a GRU [12]. They observe a large (+50) improvement in route completion from this refinement. Our findings suggest that the refinement module improves steering primarily through TP conditioning, which is only provided to the refinement GRU. TCP [31] observes that waypoints are stronger at collision avoidance than directly predicting controls, but sometimes suffer at large turns. They leverage the strength and mitigate the weaknesses of these two representations by creating a situation dependent ensemble. Our study suggests that the waypoints are better at collision avoidance due to an implicit slowdown when the car is uncertain. PlanT [23] investigates planning on CARLA by processing object bounding boxes with a transformer. Its sensorimotor version differs from other SotA models in that it is trained in two stages (not end-to-end). In our setting, we find that end-to-end training is crucial. We show that their observations about dataset scale also hold for end-to-end models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 383 + ], + "type": "text", + "content": "TransFuser [11,22] is a simple, well-known and widely used baseline for CARLA. We provide an explanation for the large differences in route completion between the target point conditioned TransFuser architecture and its NC conditioned baselines [7,9,14]. Further, we propose modifications to its architecture, output representation and training strategy which lead to significant improvements." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 383, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 526 + ], + "type": "text", + "content": "Interfuser [26] regresses a path for steering, predicts object density maps, and classifies traffic rule flags. This representation is converted by a forecasting mechanism and hand designed heuristics into control. Like our proposed method, it uses a transformer decoder for pooling features, and disentangles future velocities from the path in the output. While these ideas were already present, they were not studied or discussed as significant. Our work adds to the literature by showing that these design choices are indeed critical to performance and providing explanations as to why. Furthermore, our final system is simpler and significantly outperforms Interfuser on Longest6." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 527, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 563 + ], + "type": "text", + "content": "ReasonNet [27], ThinkTwice [18] and CaT [32] are concurrent end-to-end driving approaches that all use TP conditioning and waypoints as output representation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 573, + 512, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 512, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 512, + 587 + ], + "type": "text", + "content": "3. Hidden Biases of End-to-End Driving" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": "We consider the task of urban navigation from point A to B [11]. The goal is to complete routes with dense traffic, multiple lanes and complex geometries (e.g. roundabouts) without incurring infractions. Along the way, the agent encounters manually designed pre-crash traffic scenarios. Each route is a list of GNSS coordinates called target points (TPs) which can be up to " + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "50\\mathrm{m}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " apart (" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\sim 30\\mathrm{m}" + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": " on average)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Metrics: We use the CARLA online metrics. Route Completion (RC) is the percentage of the route completed. Infraction Score (IS) is a penalty factor starting at 1.0 that" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8241" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 212, + 277 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 212, + 277 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 212, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 212, + 277 + ], + "type": "image", + "image_path": "f73ae611e9a505aab52a4e2d51fd1ef542472da56a923283a1cd87bd29913bdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 281, + 190, + 292 + ], + "lines": [ + { + "bbox": [ + 71, + 281, + 190, + 292 + ], + "spans": [ + { + "bbox": [ + 71, + 281, + 190, + 292 + ], + "type": "text", + "content": "(a) TransFuser (NC conditioned)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 70, + 379, + 277 + ], + "blocks": [ + { + "bbox": [ + 214, + 70, + 379, + 277 + ], + "lines": [ + { + "bbox": [ + 214, + 70, + 379, + 277 + ], + "spans": [ + { + "bbox": [ + 214, + 70, + 379, + 277 + ], + "type": "image", + "image_path": "757d0036a0c9d88c7c99088a58efcdfe252f25d64716dbec9decaa89ebc34dba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 281, + 356, + 292 + ], + "lines": [ + { + "bbox": [ + 238, + 281, + 356, + 292 + ], + "spans": [ + { + "bbox": [ + 238, + 281, + 356, + 292 + ], + "type": "text", + "content": "(b) TransFuser (TP conditioned)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 545, + 277 + ], + "type": "image", + "image_path": "c9018ef2eb763ec683c5380c78a2d062c24ef8f9cdbcba8cc6ded3fe01c425d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 281, + 516, + 292 + ], + "lines": [ + { + "bbox": [ + 409, + 281, + 516, + 292 + ], + "spans": [ + { + "bbox": [ + 409, + 281, + 516, + 292 + ], + "type": "text", + "content": "(c) LAV [8] (TP conditioned)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 293, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 331 + ], + "type": "text", + "content": "Figure 2: Extrapolation to target point. In unknown situations, TP conditioned methods extrapolate their waypoints towards target points. This periodically resets steering errors and is a form of implicit map based recovery. However, relying on extrapolation is a shortcut that can lead to catastrophic errors in certain situations (see Fig. 3a and Fig. 3b)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "type": "text", + "content": "gets reduced multiplicatively for every infraction. Our main metric is the Driving Score (DS) which multiplies RC with the IS. Where insightful, we report infraction per kilometer metrics. For a comprehensive description, refer to [11]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 392, + 287, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 476 + ], + "type": "text", + "content": "**Baseline:** As a baseline, we reproduce TransFuser [11]. This is a simple method representative of current SotA driving systems on CARLA. The main differences in our reproduction are a " + }, + { + "bbox": [ + 46, + 392, + 287, + 476 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 392, + 287, + 476 + ], + "type": "text", + "content": " field of view (FOV) LiDAR to enable safe lane changes and a dataset where the expert is driving " + }, + { + "bbox": [ + 46, + 392, + 287, + 476 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 392, + 287, + 476 + ], + "type": "text", + "content": " faster to speed up evaluation. A list of other minor changes is provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 477, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 477, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 287, + 573 + ], + "type": "text", + "content": "Benchmark: For all experiments in this section, we train on CARLA towns 01, 03, 04, 06, 07 and 10. We use the 16 validation routes from LAV [8] in the withheld towns 02 and 05 for evaluation. In order to reduce the influence of training and evaluation variance which can be large in CARLA [4, 11], every experiment reports the average of 3 training seeds, evaluated 3 times each. We additionally report the training standard deviation for the main metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 586, + 178, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 178, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 178, + 598 + ], + "type": "text", + "content": "3.1. A shortcut for recovery" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "While current SotA CARLA methods have fundamentally different architectures [8,11,26,31], they are all trained with conditional IL using fixed pre-recorded datasets. The evaluation task involves challenging routes which are " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sim 1.5\\mathrm{km}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " long, so it would be expected that these methods suffer from compounding errors, a well-known problem for IL [24]. Geometric shift augmentations [5] are a common approach to teach IL methods how to recover from such compounding steering errors [5, 14]. Surprisingly," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 378 + ], + "type": "text", + "content": "even though such augmentations are not employed by SotA methods on CARLA, they report high RCs, demonstrating that they are not prone to this expected failure mode." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 380, + 546, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 582 + ], + "type": "text", + "content": "All the aforementioned methods condition their predictions using the next target point (TP) along the route. To understand the importance of the TP, we train 2 versions of our reproduction of TransFuser: one with the original TP conditioning and another with a discrete (NC) condition. This is a one hot vector that indicates whether the car should follow the lane, turn right at the next intersection, etc. It contains no geometric information about the center of the lanes, but still removes the inherent task ambiguity of inner-city driving [13]. For the NC conditioned model, we also implement shift and rotation augmentations by collecting augmented frames during data collection. We deploy a second camera in the simulator, that changes its position and orientation randomly at every time-step. The output labels are transformed accordingly for training the model (implementation details and examples can be found in the supplementary material). The results are shown in Table 1." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 339, + 594, + 514, + 651 + ], + "blocks": [ + { + "bbox": [ + 339, + 594, + 514, + 651 + ], + "lines": [ + { + "bbox": [ + 339, + 594, + 514, + 651 + ], + "spans": [ + { + "bbox": [ + 339, + 594, + 514, + 651 + ], + "type": "table", + "html": "
Cond.Aug.DS ↑RC ↑Dev ↓
NC-32 ± 856 ± 120.86
NC✓35 ± 354 ± 40.99
TP-39 ± 984 ± 70.00
", + "image_path": "ceb7e678cd448085bd5fa9be991c882af0ab2182eb183f9c119675c7b9fecf02.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 336, + 659, + 515, + 671 + ], + "lines": [ + { + "bbox": [ + 336, + 659, + 515, + 671 + ], + "spans": [ + { + "bbox": [ + 336, + 659, + 515, + 671 + ], + "type": "text", + "content": "Table 1: Conditioning and Augmentation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We observe a significant difference in RC of 28 points when switching between TP and NC conditioning. The TP conditioned model has 0 route deviations per km (\"Dev\")," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8242" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 212, + 114 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 212, + 114 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 212, + 114 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 212, + 114 + ], + "type": "image", + "image_path": "545be28bd61414f2dd817fdb99088f2423a765d951c7a6f901d0e5d1db571516.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 121, + 141, + 277 + ], + "blocks": [ + { + "bbox": [ + 49, + 121, + 141, + 277 + ], + "lines": [ + { + "bbox": [ + 49, + 121, + 141, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 121, + 141, + 277 + ], + "type": "image", + "image_path": "531a2b62baf02b0ced74678761cfbd331f200b760a2278217625d3fd21888851.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 281, + 189, + 291 + ], + "lines": [ + { + "bbox": [ + 73, + 281, + 189, + 291 + ], + "spans": [ + { + "bbox": [ + 73, + 281, + 189, + 291 + ], + "type": "text", + "content": "(a) TCP [31] follows a shortcut." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 214, + 70, + 378, + 113 + ], + "blocks": [ + { + "bbox": [ + 214, + 70, + 378, + 113 + ], + "lines": [ + { + "bbox": [ + 214, + 70, + 378, + 113 + ], + "spans": [ + { + "bbox": [ + 214, + 70, + 378, + 113 + ], + "type": "image", + "image_path": "7b9b4a700097ebe9cd4815dcdfe60f590af54af6592135bab170194eed9f5ba1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 214, + 114, + 378, + 278 + ], + "blocks": [ + { + "bbox": [ + 214, + 114, + 378, + 278 + ], + "lines": [ + { + "bbox": [ + 214, + 114, + 378, + 278 + ], + "spans": [ + { + "bbox": [ + 214, + 114, + 378, + 278 + ], + "type": "image", + "image_path": "882ff92bb508758e6a52f6ea756405da23d92cfe04f62ab4b41dcfc1ee83bc82.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 281, + 358, + 291 + ], + "lines": [ + { + "bbox": [ + 235, + 281, + 358, + 291 + ], + "spans": [ + { + "bbox": [ + 235, + 281, + 358, + 291 + ], + "type": "text", + "content": "(b) TransFuser follows a shortcut." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 545, + 113 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 545, + 113 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 545, + 113 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 545, + 113 + ], + "type": "image", + "image_path": "6dae0f4b7ce141a1af37a3748af96f3c2d1348f9f011e3237ad2ca27f962ecb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 381, + 114, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 381, + 114, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 381, + 114, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 381, + 114, + 545, + 277 + ], + "type": "image", + "image_path": "95bcd4da5c5bcfd8bf21868307d38123f417fde75dd2990c98a47239d39596a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 281, + 528, + 291 + ], + "lines": [ + { + "bbox": [ + 397, + 281, + 528, + 291 + ], + "spans": [ + { + "bbox": [ + 397, + 281, + 528, + 291 + ], + "type": "text", + "content": "(c) A transformer decoder fixes this." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 331, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 594 + ], + "type": "text", + "content": "which indicates that it never takes a wrong turn or drives too far away (more than " + }, + { + "bbox": [ + 46, + 331, + 289, + 594 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 46, + 331, + 289, + 594 + ], + "type": "text", + "content": ") from the lane. However, this is not the case for NC conditioned models. While we do see a small improvement in DS with augmentation, its RC is still unsatisfactory. This indicates that the geometric information regarding the lane center available with the TP aids recovery in SotA IL methods, prompting further investigation. We inspect the TransFuser and LAV [8] models by constructing situations where the car is forcefully steered out of lane. Fig. 2 visualizes these scenarios with the camera, LiDAR and ground truth HD maps. For TP conditioned models, the predicted waypoints (shown in blue) extrapolate towards the nearest TP (red) even though the situation is out of distribution. This shows that one reason for their strong route following is that the bird's eye view (BEV) TP resets steering errors periodically: methods learn to steer towards nearby TPs because the expert trajectory in the dataset always goes through them. This has the effect that TP conditioned methods periodically drive back to the center of the lane, resetting any accumulated errors. Furthermore, it suggests that SotA IL approaches strongly rely on pre-specified geometric information regarding lane centers for recovery." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "We identify this as a form of shortcut learning, which can be useful when the car is close to a target point, but can also lead to catastrophic steering errors when the target point is further away. An example would be directly extrapolating to a target point behind a turn, which leads to cutting the turn. In Fig. 3a and Fig. 3b, we show instances where this happens for TCP [31] and TransFuser [11] in a validation town at nighttime. TransFuser also predicts the BEV segmentation as an auxiliary task, which we overlay" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 328, + 547, + 422 + ], + "blocks": [ + { + "bbox": [ + 46, + 293, + 546, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 546, + 319 + ], + "type": "text", + "content": "Figure 3: Target point shortcut. When TP conditioned methods extrapolate to spatially distant waypoints, they incur large steering errors. Replacing global average pooling in TransFuser with a cross-attention mechanism mitigates the issue." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 328, + 547, + 422 + ], + "lines": [ + { + "bbox": [ + 307, + 328, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 547, + 422 + ], + "type": "image", + "image_path": "0d3fbcfa66a60c99ada9e90fe765c072327ce6e055601f2432306eb08e192d57.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 431, + 546, + 479 + ], + "lines": [ + { + "bbox": [ + 304, + 431, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 546, + 479 + ], + "type": "text", + "content": "Figure 4: Pooling. Existing approaches vectorize feature grids either by global average pooling (top, e.g. [8, 11]) or with attention mechanisms (bottom, e.g. [26, 31]). The latter retains spatial information gained via auxiliary tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 493, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 564 + ], + "type": "text", + "content": "in the figure (gray: route, yellow: lane marking). For TCP, we render the ground truth map and omit the LiDAR, since it does not use LiDAR and does not predict BEV segmentation. Both methods predict waypoints that are tilted towards the TP instead of following the street. As a result, they drive into the opposing lane. We refer to this as the TP shortcut." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 574, + 523, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 523, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 523, + 586 + ], + "type": "text", + "content": "3.2. Improved pooling and data augmentation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": "One design choice which differs across SotA architectures is the pooling applied between the encoder and decoder. In Fig. 4, we summarize these. In particular, TransFuser [11] and LAV [8] employ global average pooling (GAP) followed by an MLP. InterFuser [26] pools features via the cross-attention mechanism of a transformer decoder. Finally, TCP [31], which has two decoders, uses GAP for one decoder and attention-based pooling for the other." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig. 4, the networks are encouraged to learn spatially meaningful feature grids through auxiliary convo" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8243" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "lutional decoders that predict outputs such as BEV semantic segmentation. However, GAP does not maintain the spatial information in the features. Fig. 3b shows the BEV road segmentation predicted by TransFuser overlaid by its Li-DAR input. Unlike the waypoints, the BEV predictions are quite accurate. This indicates strong features coming from the encoder. Nevertheless, the final conditional waypoint predictions focus on the TP signal in this situation. We hypothesize that the GAP operation makes it difficult for the downstream decoder to utilize the strong BEV features. Note that attention-based pooling, such as the transformer decoder of Interfuser, preserves spatial information through the use of positional encodings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 302 + ], + "type": "text", + "content": "We compare the original TransFuser GAP approach with the spatially preserving design of Interfuser in Table 2. For the latter variant, we remove the GAP operation and MLP in TransFuser. We then process the " + }, + { + "bbox": [ + 46, + 228, + 289, + 302 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 46, + 228, + 289, + 302 + ], + "type": "text", + "content": " BEV feature grid as tokens with a transformer decoder. Implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 54, + 312, + 282, + 370 + ], + "blocks": [ + { + "bbox": [ + 54, + 312, + 282, + 370 + ], + "lines": [ + { + "bbox": [ + 54, + 312, + 282, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 282, + 370 + ], + "type": "table", + "html": "
PoolingAug.DS ↑RC ↑Stat ↓
GAP + MLP-39 ± 984 ± 71.04
Transformer Decoder-43 ± 693 ± 30.55
Transformer Decoder✓49 ± 890 ± 40.10
", + "image_path": "f0c7a04d45982c518fdb0d8fd7ec4b75226d6d98dd0389476ebbec1be2755e4c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 377, + 244, + 389 + ], + "lines": [ + { + "bbox": [ + 90, + 377, + 244, + 389 + ], + "spans": [ + { + "bbox": [ + 90, + 377, + 244, + 389 + ], + "type": "text", + "content": "Table 2: Pooling and augmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 552 + ], + "type": "text", + "content": "After replacing GAP with the transformer decoder, the RC increases by 9 points and collisions per km with static objects (\"Stat\") reduce by a factor of 2. Static objects (such as poles) only occur outside lanes in CARLA, implying that the network manages to stay in the lane far more consistently. Fig. 3c provides qualitative evidence that the transformer decoder can succeed in situations where GAP failed due to the TP shortcut (additional examples in supplementary). In addition, we investigate the inclusion of shift and rotation augmentations designed to aid lateral recovery (as described in Section 3.1). Collisions with the static environment are reduced further by a factor of 5 indicating that augmentation provides strong benefits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 562, + 201, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 201, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 201, + 574 + ], + "type": "text", + "content": "3.3. The ambiguity of waypoints" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": "Waypoints are used in many SotA systems as outputs [8, 11, 31]. They are obtained by recording an expert driver's GNSS locations at fixed time intervals (e.g. 0.5s) and transforming them into a local coordinate frame. The model is then trained to predict future waypoints, typically with an " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": " regression objective. Waypoints entangle both the path and the future velocities of the vehicle. The velocity after a specific time interval extracted from the waypoints is used by a downstream controller as a target speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In Fig. 5, we plot a histogram comparing the target speeds of the expert algorithm (i.e. training dataset) to those" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 525, + 238 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 525, + 238 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 525, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 525, + 238 + ], + "type": "image", + "image_path": "7496a29c2d4fffcb9f57a32a1cfafa0bfdd5ea9958471d6b60b2c2d80d5686e5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 246, + 531, + 258 + ], + "lines": [ + { + "bbox": [ + 320, + 246, + 531, + 258 + ], + "spans": [ + { + "bbox": [ + 320, + 246, + 531, + 258 + ], + "type": "text", + "content": "Figure 5: Transfuser interpolates between modes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 364 + ], + "type": "text", + "content": "extracted from TransFuser (best model in Table 2). Interestingly, the distributions are quite different. By design, the expert chooses one of four target speeds: 29, 18, 7, or 0 km/h. These values cover behaviors needed for city driving, corresponding to four situations: regular driving, slowing down in intersections, slowing down near pedestrians, and stopping. In contrast, for TransFuser, the predicted target speeds cover the entire 0-29 km/h range." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 365, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 568 + ], + "type": "text", + "content": "While the future path that our expert follows is deterministic and unambiguous (center of the lane, lane change at pre-defined locations), future velocities are multi-modal. As both are jointly represented by waypoints, this leads to an ambiguous entangled representation. However, existing methods which utilize this entangled waypoint representation predict only point estimates (i.e., a single set of waypoints) as output, hence only a single mode is modeled. From Fig. 5, we observe that the waypoint based TransFuser model indeed interpolates between modes, a behavior that is expected when modeling multi-modal outputs deterministically. We illustrate this behavior in detail using the examples of (1) approaching an intersection with a green light (Fig. 6a), and (2) a cyclist cutting into the vehicle's path (Fig. 6b). TransFuser slows down in these situations since it is uncertain, e.g., the light may turn red in Fig. 6a. The car stops in time in Fig. 6b because of the decreased speed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "While this averaging is indeed beneficial in some situations, an entangled representation is undesirable as it is less interpretable and does not explicitly expose uncertainty. Moreover, when stopped (all waypoints collapsed to one location), the steering signal is undefined which requires additional heuristics in the controller. To resolve this, we disentangle the future velocities from the path by sampling the expert's position at fixed distances instead of fixed time intervals for training a (deterministic) path predictor. However, when predicting the path instead of time-dependent trajectories, one requires an additional method to determine a target speed for the car. We propose to predict the tar" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8244" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 213, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 213, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 213, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 213, + 277 + ], + "type": "image", + "image_path": "6f92f0a9f4dcb7a9453dc9387e8aae22342b37524a874d13ee277f88c6c6cbfb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 281, + 192, + 292 + ], + "lines": [ + { + "bbox": [ + 69, + 281, + 192, + 292 + ], + "spans": [ + { + "bbox": [ + 69, + 281, + 192, + 292 + ], + "type": "text", + "content": "(a) Slowing down at a green light." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 70, + 379, + 277 + ], + "blocks": [ + { + "bbox": [ + 214, + 70, + 379, + 277 + ], + "lines": [ + { + "bbox": [ + 214, + 70, + 379, + 277 + ], + "spans": [ + { + "bbox": [ + 214, + 70, + 379, + 277 + ], + "type": "image", + "image_path": "51c9dde26c5bd6454eab8b33693d7623c562f53a5bdfe26c9c669bc55ac5929d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 281, + 363, + 292 + ], + "lines": [ + { + "bbox": [ + 230, + 281, + 363, + 292 + ], + "spans": [ + { + "bbox": [ + 230, + 281, + 363, + 292 + ], + "type": "text", + "content": "(b) Slowing down in an intersection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 380, + 70, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 380, + 70, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 380, + 70, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 380, + 70, + 545, + 277 + ], + "type": "image", + "image_path": "f295c149d94641124f7afc384c7437a83b1a86a80d044ff403ce9821b08a4157.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 281, + 536, + 292 + ], + "lines": [ + { + "bbox": [ + 389, + 281, + 536, + 292 + ], + "spans": [ + { + "bbox": [ + 389, + 281, + 536, + 292 + ], + "type": "text", + "content": "(c) Disentangling route and target speed." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 415 + ], + "type": "text", + "content": "get speed by simple classification using an additional MLP head. We incorporate the prediction uncertainties into the final output by using a confidence weighted average of the predicted target speed as input to the controller. Note that this is just one example of how the resulting confidence estimates may be leveraged by a vehicle controller." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 61, + 428, + 273, + 485 + ], + "blocks": [ + { + "bbox": [ + 61, + 428, + 273, + 485 + ], + "lines": [ + { + "bbox": [ + 61, + 428, + 273, + 485 + ], + "spans": [ + { + "bbox": [ + 61, + 428, + 273, + 485 + ], + "type": "table", + "html": "
OutputDS ↑RC ↑Veh ↓Stat ↓
Waypoints49 ± 890 ± 40.700.10
Path + Argmax40 ± 188 ± 21.250.03
Path + Weighted50 ± 388 ± 10.830.02
", + "image_path": "20eb67207828b7ad0ef9eecf0cecf319175ae342402eb35c911bd3f32d40de32.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 508, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 675 + ], + "type": "text", + "content": "Compared to the waypoint representation, naively converting the predicted classes to a target speed via the most likely class (Argmax) increases vehicle collisions (\"Veh\"), as seen in Table 3. Employing the proposed weighting (Weighted) achieves lower collisions. Moreover, the path-based model steers better as indicated by the lower static obstacle collisions (\"Stat\"). The disentangled representation achieves the same driving score as the entangled waypoint representation, providing an alternative with a simpler and more interpretable controller: (1) it eases design since it has identical parameters to the controller in the expert, (2) it has access to an unambiguous path representation even when driving slowly, and (3) it explicitly exposes and makes use of uncertainties with regard to target speeds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Fig. 6c shows a qualitative example. The car is slowing down due to its uncertainty. The lower target speed allows it successfully merge behind the cyclist." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 341, + 545, + 519 + ], + "blocks": [ + { + "bbox": [ + 46, + 293, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 331 + ], + "type": "text", + "content": "Figure 6: Waypoints are ambiguous. The model's output representation forces it to predict a single mode for future velocities. There is a possibility that the traffic light might turn red in the future, or that the cyclist either cuts in or yields to the agent. The particular mode (or interpolation thereof) that the model converges to, depends on the training run and dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 341, + 545, + 519 + ], + "lines": [ + { + "bbox": [ + 306, + 341, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 545, + 519 + ], + "type": "image", + "image_path": "4de65413a2c6ceebd9f2de56d54ee791b05faf6cfcd44c2ea1b5ada5bcc31d27.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 528, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 345, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 345, + 528, + 505, + 540 + ], + "type": "text", + "content": "Figure 7: TransFuser++ architecture." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 563, + 458, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 458, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 458, + 576 + ], + "type": "text", + "content": "3.4. Scaling up to TransFuser++" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "content": "By incorporating these insights, we obtain a significantly improved version of TransFuser [11] that we call TransFuser++ (Fig. 7). We now describe two important implementation changes compared to [11] that involve scaling." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 329, + 647, + 524, + 703 + ], + "blocks": [ + { + "bbox": [ + 99, + 494, + 235, + 506 + ], + "lines": [ + { + "bbox": [ + 99, + 494, + 235, + 506 + ], + "spans": [ + { + "bbox": [ + 99, + 494, + 235, + 506 + ], + "type": "text", + "content": "Table 3: Output representation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 647, + 524, + 703 + ], + "lines": [ + { + "bbox": [ + 329, + 647, + 524, + 703 + ], + "spans": [ + { + "bbox": [ + 329, + 647, + 524, + 703 + ], + "type": "table", + "html": "
Two stageFrozenDS ↑RC ↑Veh ↓
--50 ± 388 ± 10.83
✓-54 ± 192 ± 50.79
✓✓44 ± 697 ± 11.21
", + "image_path": "819bfe98eca49066af663a4a8902b0f7b968e51fb9def41f4304e5ae9604cdc4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 365, + 712, + 485, + 724 + ], + "lines": [ + { + "bbox": [ + 365, + 712, + 485, + 724 + ], + "spans": [ + { + "bbox": [ + 365, + 712, + 485, + 724 + ], + "type": "text", + "content": "Table 4: Two stage training." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8245" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "Training schedule: We observe no benefits from simply doubling the number of training epochs for a single training stage. However, in Table 4, we double the training time with a two-stage approach. First, we pre-train the encoder with only the perception losses (2D depth, 2D and BEV semantic segmentation, and vehicle bounding boxes, as shown in Fig. 7) for the usual number of epochs. We then fine-tune the resulting checkpoint with all losses, after including the GRU decoder and MLP classifier. Initializing the backbone with features that are pre-trained on the auxiliary tasks leads to a 4 DS improvement, consistent with the claims of LAV [8]. We also experiment with freezing the pre-trained backbone and only training the transformer decoder and its heads in the second stage. This leads to a drop of 10 DS, indicating that end-to-end optimization is important." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 397 + ], + "type": "text", + "content": "Dataset scale: Recent work on CARLA [23] shows large improvements by scaling the dataset size. However, [23] considers planning models with privileged inputs on training towns. We investigate the impact of scaling when evaluating on held-out validation towns for end-to-end models. We start with 185k training samples (as in [11]) and scale it up by re-running the training routes 3 times with different traffic (555k frames). Table 5 shows a clear improvement of 6 DS via scaling. This shows that current IL approaches can still benefit from larger datasets. We train for the same amount of epochs, which implies that this improvement comes at the cost of " + }, + { + "bbox": [ + 46, + 252, + 289, + 397 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 252, + 289, + 397 + ], + "type": "text", + "content": " longer training." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 86, + 407, + 250, + 454 + ], + "blocks": [ + { + "bbox": [ + 86, + 407, + 250, + 454 + ], + "lines": [ + { + "bbox": [ + 86, + 407, + 250, + 454 + ], + "spans": [ + { + "bbox": [ + 86, + 407, + 250, + 454 + ], + "type": "table", + "html": "
Dataset sizeDS ↑RC ↑Veh ↓
185k54 ± 192 ± 50.79
555k60 ± 698 ± 10.73
", + "image_path": "1f5084fc042e3f1644c63f03184187a1a040153c153443b964a45c319954c320.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 461, + 219, + 472 + ], + "lines": [ + { + "bbox": [ + 115, + 461, + 219, + 472 + ], + "spans": [ + { + "bbox": [ + 115, + 461, + 219, + 472 + ], + "type": "text", + "content": "Table 5: Effects of scale." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 489, + 221, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 221, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 221, + 503 + ], + "type": "text", + "content": "4. Comparison to State of the Art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Methods: This section demonstrates the advantages of " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TF}++" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " over several state-of-the-art models, which we list below, starting with the most recent. (1) Interfuser [26] processes multiple cameras and a BEV LiDAR by encoding each of them individually with a CNN. The resulting feature grids are fed into a transformer and decoded into perception outputs and the path to follow. The network does not predict longitudinal controls. Instead, it uses the BEV bounding boxes and a simple motion forecast (extrapolating historical dynamics) and linearly optimizes for the target speed using heuristically chosen objectives. (2) Perception PlanT [23] is a transformer based method trained in two independent stages that uses BEV bounding boxes as an intermediate representation. Its perception is based on TransFuser, and its planner outputs waypoints. (3) TCP [31] is a camera-only model with two output representations: waypoints and controls. During test time, it ensembles the two" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 300 + ], + "type": "text", + "content": "outputs together with a weighted average that changes based on whether the vehicle is turning. (4) TransFuser [11] fuses perspective cameras with a BEV LiDAR by processing them with individual CNNs and exchanging features using transformers. It uses global average pooling and waypoint outputs, as described in Section 3. (5) LAV [8] processes cameras and LiDAR point clouds into an intermediate BEV representation. It trains a planner on this representation that predicts waypoints for the ego vehicle with a NC conditioned GRU followed by a TP conditioned refinement GRU. During training, the planner is also tasked to predict the trajectory of other surrounding vehicles to increase the number of labels. The authors release two versions of this method, which we call LAV v1 and LAV v2. (6) WOR [7] is an IL method for which the labels are enriched using a reward function to provide dense supervision for all possible actions. Unlike the other baselines, it is conditioned with the NC and does not utilize the TP. WOR is the best NC conditioned baseline that is publicly available." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 301, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 547, + 578 + ], + "type": "text", + "content": "**Benchmarks:** We use two benchmarks to evaluate on seen and unseen towns, with scenarios taken from [11] (type 1,3,4,7,8,9 and 10). Table 6 compares the performance of systems on the Longest6 benchmark [11] which consists of 36 routes in training towns 01 to 06. On Longest6, models trained on any data can be evaluated, so we compare against the author-provided models or directly report numbers from the respective papers. The mean and std of three evaluations is reported. For " + }, + { + "bbox": [ + 304, + 301, + 547, + 578 + ], + "type": "inline_equation", + "content": "\\mathrm{TF}++" + }, + { + "bbox": [ + 304, + 301, + 547, + 578 + ], + "type": "text", + "content": " we train 3 models and report the average result. Longest6 has the advantage that it evaluates in dense traffic and has diverse towns and routes. The drawbacks are that it does not penalize stop sign infractions or measure generalization to new towns. Therefore, we additionally compare performance on validation towns using the LAV [8] routes in Table 7. These are 4 routes with 4 weathers (16 combined) in Town 02 and 05, which are withheld during training. We re-train methods with the dataset from the corresponding paper 3 times and evaluate each seed 3 times. Reported results are the mean of all runs and the std between the training seeds. We compare against the reported SotA TCP and our reproduced TransFuser baseline on this benchmark and provide additional baselines in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 579, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 663 + ], + "type": "text", + "content": "Dataset: For " + }, + { + "bbox": [ + 304, + 579, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathrm{TF}++" + }, + { + "bbox": [ + 304, + 579, + 547, + 663 + ], + "type": "text", + "content": " we collect data on the same training routes as in [11] with an improved expert labeling algorithm (described in the supplementary material). We repeat this 3 times on the same routes with different traffic, as in [23]. For validation, we withhold Town 02 and 05 during training, else we train on all towns. In total, we train with 750k frames (550k when withholding Town 02 and 05)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Results: We start with the results in training towns (Table 6). The best NC conditioned method, WOR, has significantly lower RC than all TP conditioned systems (22 lower than LAV v1). In particular, the route deviations (Dev) are" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8246" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 525, + 198 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 525, + 198 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 525, + 198 + ], + "type": "table", + "html": "
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Dev ↓TO ↓Block ↓
WOR [7]21 ± 348 ± 40.56 ± 0.030.181.050.371.280.880.080.20
LAV v1 [8]33 ± 170 ± 30.51 ± 0.020.160.830.150.960.060.120.45
Interfuser [26]47 ± 674 ± 10.63 ± 0.070.061.140.110.240.000.520.06
TransFuser [11]47 ± 693 ± 10.50 ± 0.060.032.450.070.160.000.060.10
TCP [31]48 ± 372 ± 30.65 ± 0.040.041.080.230.140.020.180.35
LAV v2 [8]58 ± 183 ± 10.68 ± 0.020.000.690.150.230.080.320.11
Perception PlanT [23]58 ± 588 ± 10.65 ± 0.060.070.970.110.090.000.130.13
TF++ (ours)69 ± 094 ± 20.72 ± 0.010.000.830.010.050.000.070.06
Expert81 ± 390 ± 10.91 ± 0.040.010.210.000.010.000.070.09
", + "image_path": "86c0e879617727145fd2ffb657f88796ef79160aa86b1c4f6d1fcd7aae2fc4f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 229, + 535, + 304 + ], + "blocks": [ + { + "bbox": [ + 104, + 206, + 488, + 217 + ], + "lines": [ + { + "bbox": [ + 104, + 206, + 488, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 206, + 488, + 217 + ], + "type": "text", + "content": "Table 6: Performance on training towns (Longest6). Released models, std over 3 evaluations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 229, + 535, + 304 + ], + "lines": [ + { + "bbox": [ + 59, + 229, + 535, + 304 + ], + "spans": [ + { + "bbox": [ + 59, + 229, + 535, + 304 + ], + "type": "table", + "html": "
MethodDS ↑RC ↑IS ↑Ped ↓Veh ↓Stat ↓Red ↓Stop ↓Dev ↓TO ↓Block ↓
TransFuser (ours)39 ± 984 ± 70.46 ± 0.060.000.741.040.201.070.000.230.21
TCP [31]58 ± 585 ± 30.67 ± 0.060.000.350.160.011.050.000.190.19
TF++ (ours)70 ± 699 ± 00.70 ± 0.060.010.630.010.040.260.000.050.00
Expert94950.990.000.020.000.020.000.000.000.08
", + "image_path": "e1574de00cf735e93d3fe286bbf845d8b367ea21f18f5224109304339c41ca3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "text", + "content": "10 times higher. Current TP conditioned SotA methods achieve (close to) 0 route deviations with driving scores in the range of 47 DS to 58 DS. TF++ outperforms all baselines on Longest6, showing a " + }, + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "text", + "content": " relative improvement over the previous SotA Perception PlanT. TF++ achieves close to expert-level performance on all infractions except for vehicle collisions (Veh). On the validation towns (Table 7) TF++ outperforms our reproduced TransFuser by 31 DS improving all metrics, particularly environmental collisions (Stat). It improves the prior SotA result TCP by " + }, + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "text", + "content": ". Runtime: Table 8 compares the runtime of TF++ and TransFuser which have the same sensor setup. TF++ yields a large DS improvement while being only " + }, + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 331, + 287, + 486 + ], + "type": "text", + "content": " slower." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 84, + 496, + 251, + 543 + ], + "blocks": [ + { + "bbox": [ + 59, + 311, + 531, + 323 + ], + "lines": [ + { + "bbox": [ + 59, + 311, + 531, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 311, + 531, + 323 + ], + "type": "text", + "content": "Table 7: Performance on validation towns (LAV). Reproduced models, std over 3 trainings (for 3 evaluations each)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 496, + 251, + 543 + ], + "lines": [ + { + "bbox": [ + 84, + 496, + 251, + 543 + ], + "spans": [ + { + "bbox": [ + 84, + 496, + 251, + 543 + ], + "type": "table", + "html": "
Method ↑DS ↑Time (ms) ↓
TransFuser (ours)3944
TF++ (ours)7050
", + "image_path": "00b3d66ce55636f08812b19500be937753424e762fd2ba86a4189a0457ca0d13.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 551, + 287, + 599 + ], + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 599 + ], + "type": "text", + "content": "Table 8: Runtime. We show the runtime per frame in ms averaged over 300 time steps on a single route on a RTX 3090. TF++ outperforms TransFuser by a wide margin despite using a similar compute budget during inference." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 119, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "In this work, we show that recent SotA driving models have exceptional route following abilities because they learn a strong bias towards following nearby TPs. Shortcut learning like this is a general phenomenon in deep neural networks [16] and has been observed in the context of autonomous driving for inputs such as velocity [14] or temporal frames [3, 30]. We add to this literature by observ-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 331, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 402 + ], + "type": "text", + "content": "ing shortcut learning with respect to the conditioning signal. While shortcut learning usually has a negative impact on performance, we observe positive (improved recovery) and negative (cutting turns) effects. We show that the negative effects can be mitigated by avoiding global pooling and incorporating data augmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": "A second commonality in SotA approaches is the use of waypoints as an output representation. We observe that this representation is ambiguous because it predicts a point estimate for multi-modal future velocities. We disambiguate the representation by disentangling future velocities from the deterministic path predictions and classifying target speeds instead. We then weigh target speeds according to their predicted confidence in our controller. Surprisingly, we find that interpolation is helpful for reducing collisions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 557 + ], + "type": "text", + "content": "We propose TransFuser++ by improving the popular baseline TransFuser with a series of controlled experiments. TF++ is a simple end-to-end method that sets a new state of the art on the LAV and Longest6 benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": "Limitations: This study investigates urban driving in CARLA, where all investigated methods drive at relatively low speed " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "(< 35\\mathrm{km / h})" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": ". Therefore, problems specific to high-speed driving are not considered. In addition, lanes are free of static obstacles, hence scenarios requiring navigation around them are not included. We point out the strong reliance of current methods on TPs for recovery. This implies a reliance on accurate localization and mapping to obtain these TPs. They are accurately mapped in CARLA, however, this assumption might not hold in real environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Broader Impact: We aim to make progress towards autonomous driving. This technology, if realized, could have massive societal impact, reducing road accidents, trans" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8247" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "portation costs and improving the mobility of elderly people. Potential negative implications include a reduction in jobs for human drivers and possible military applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 114, + 289, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 114, + 289, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 114, + 289, + 246 + ], + "type": "text", + "content": "Acknowledgements: Andreas Geiger and Bernhard Jaeger were supported by the ERC Starting Grant LEGO-3D (850533), the BMWi in the project KI Delta Learning (project number 19A19013O) and the DFG EXC number 2064/1 - project number 390727645. Kashyap Chitta was supported by the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. We thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Bernhard Jaeger and Kashyap Chitta as well as Niklas Hanselmann and Katrin Renz for proofreading." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 256, + 106, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 106, + 268 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 106, + 268 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 275, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 275, + 286, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 275, + 286, + 298 + ], + "spans": [ + { + "bbox": [ + 53, + 275, + 286, + 298 + ], + "type": "text", + "content": "[1] Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 300, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 287, + 365 + ], + "type": "text", + "content": "[2] Alexander Amini, Tsun-Hsuan Wang, Igor Gilitschenski, Wilko Schwarting, Zhijian Liu, Song Han, Sertac Karaman, and Daniela Rus. Vista 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 366, + 287, + 409 + ], + "type": "text", + "content": "[3] Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. In Proc. Robotics: Science and Systems (RSS), 2019. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 454 + ], + "type": "text", + "content": "[4] Aseem Behl, Kashyap Chitta, Aditya Prakash, Eshed Ohn-Bar, and Andreas Geiger. Label efficient visual abstractions for autonomous driving. In Proc. IEEE International Conf. on Intelligent Robots and Systems (IROS), 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 287, + 510 + ], + "type": "text", + "content": "[5] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba. End to end learning for self-driving cars. arXiv.org, 1604.07316, 2016. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "type": "text", + "content": "[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 600 + ], + "type": "text", + "content": "[7] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 287, + 634 + ], + "type": "text", + "content": "[8] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 668 + ], + "type": "text", + "content": "[9] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Proc. Conf. on Robot Learning (CoRL), 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "text", + "content": "[12] Kyunghyun Cho, Bart van Merrienboer, Caglar Güçehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proc. of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[13] Felipe Codevilla, Matthias Muiller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2018. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "text", + "content": "[14] Felipe Codevilla, Eder Santana, Antonio M. López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 340 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proc. Conf. on Robot Learning (CoRL), 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[16] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2020. 1, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "text", + "content": "[17] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zak Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "type": "text", + "content": "[18] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 500, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 543 + ], + "type": "text", + "content": "[19] Eshed Ohn-Bar, Aditya Prakash, Aseem Behl, Kashyap Chitta, and Andreas Geiger. Learning situational driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 578 + ], + "type": "text", + "content": "[20] Dean Pomerleau. ALVINN: an autonomous land vehicle in a neural network. In Advances in Neural Information Processing Systems (NIPS), 1988. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "text", + "content": "[21] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[22] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[23] Katrin Renz, Kashyap Chitta, Otniel-Bogdan Mercea, Almut Sophia Koepke, Zeynep Akata, and Andreas Geiger. Plant: Explainable planning transformers via object-level" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8248" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 554 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "representations. In Proc. Conf. on Robot Learning (CoRL), 2022. 2, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[24] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Conference on Artificial Intelligence and Statistics (AISTATS), 2011. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[25] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In Proc. Conf. on Robot Learning (CoRL), 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[26] Hao Shao, Letian Wang, RuoBing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Proc. Conf. on Robot Learning (CoRL), 2022. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 272 + ], + "type": "text", + "content": "[27] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 274, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 319 + ], + "type": "text", + "content": "[28] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 373 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NIPS), pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 419 + ], + "type": "text", + "content": "[30] Chuan Wen, Jierui Lin, Trevor Darrell, Dinesh Jayaraman, and Yang Gao. Fighting copycat agents in behavioral cloning from observation histories. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 475 + ], + "type": "text", + "content": "[31] Peng Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 476, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 510 + ], + "type": "text", + "content": "[32] Jimuyang Zhang, Zanming Huang, and Eshed Ohn-Bar. Coaching a teachable student. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "text", + "content": "[33] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8249" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_content_list.json b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b94c98e6c03b815eddc01331a6bcbf3a8d2916c --- /dev/null +++ b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_content_list.json @@ -0,0 +1,1842 @@ +[ + { + "type": "text", + "text": "Hiding Visual Information via Obfuscating Adversarial Perturbations", + "text_level": 1, + "bbox": [ + 133, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhigang Su $^{1,\\ast}$ Dawei Zhou $^{1,\\ast}$ Nannan Wang $^{1,\\dagger}$ Decheng Liu $^{1,\\dagger}$ Zhen Wang $^{2}$ Xinbo Gao $^{3}$ $^{1}$ Xidian University, $^{2}$ Zhejiang Lab \n $^{3}$ Chongqing University of Posts and Telecommunications \n{zgsu, dwzhou}.xidian@gmail.com, {nnwang, dchliu}@xidian.edu.cn wangzhen@zhejianglab.com, gaoxb@cqupt.edu.cn", + "bbox": [ + 200, + 178, + 767, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Growing leakage and misuse of visual information raise security and privacy concerns, which promotes the development of information protection. Existing adversarial perturbations-based methods mainly focus on the de-identification against deep learning models. However, the inherent visual information of the data has not been well protected. In this work, inspired by the Type-I adversarial attack, we propose an Adversarial Visual Information Hiding (AVIH) method to protect the visual privacy of data. Specifically, the method generates obfuscating adversarial perturbations to obscure the visual information of the data. Meanwhile, it maintains the hidden objectives to be correctly predicted by models. In addition, our method does not modify the parameters of the applied model, which makes it flexible for different scenarios. Experimental results on the recognition and classification tasks demonstrate that the proposed method can effectively hide visual information and hardly affect the performances of models. The code is available at https://github.com/suzhigangssz/AVIH.", + "bbox": [ + 75, + 353, + 473, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 684, + 209, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have been widely applied in the computer vision [26, 10, 18]. However, the increasing leakage and misuse of visual information has raised serious concerns [45, 20], especially in fields such as face [27, 28, 51] and medicine [5, 1, 12]. A representative case is the security issue of data stored in the cloud environment [19, 4, 35]. Due to potential vulnerabilities in cyberspace [3], uploaded private images can be easily stolen and used maliciously [38]. Therefore, it is meaningful and urgent to explore effective strategies to protect visual information.", + "bbox": [ + 75, + 710, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A classic strategy is visual information hiding, which", + "bbox": [ + 94, + 862, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eabd8be83dd989dc71e02e2ce7d341c337390bac77350c8c3467a079b24c31f5.jpg", + "image_caption": [ + "Figure 1. Illustration of visual information hiding for face recognition systems in cloud environments. The gallery set is protected and provided to the DNN in the cloud. Protected image has quite different visual information from the original image, but it is still correctly identified. The protected gallery set can be recovered by a key model by the owner." + ], + "image_footnote": [], + "bbox": [ + 506, + 320, + 887, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mainly consists of two types [40]: Homomorphic Encryption (HE)-based methods [2, 46, 32], Perceptual Encryption (PE)-based methods [42, 40, 39, 8, 16] and steganography method [24, 50]. Affected by the nonlinear activation functions in DNNs, HE-based methods are difficult to perform well on advanced DNNs [40]. Although PE-based methods apply to DNNs, existing methods typically require retraining with data in the encrypted domain to guarantee accuracy on encrypted data [42, 40]. This affects the performance of service model on raw data and cause additional resource consumption (especially for large models). The steganography method can hide the visual information of sensitive data into another image, but it does not guarantee that the service model can correctly recognize the protected image.", + "bbox": [ + 496, + 580, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate these negative effects, we expect to hide the visual information without making any modifications to service model. Namely, we hide sensitive visual information only by varying the input image. Previous researches have made some explorations in this regard. The transformation network-based methods [16], which try to protect the original image via a transformation function parameterized by", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. † Corresponding author.", + "bbox": [ + 94, + 886, + 344, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4356", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a neural network, share the same philosophy. However, the method cannot easily recover the original image from the protected image for other purposes. They may suffer from adversarial vulnerability [9, 14] because the introduced neural network may be destroyed by adversarial attacks.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In fact, the negative effects of adversarial attacks can be utilized positively to protect privacy. Some works have exploited adversarial attacks for de-identification [36, 34, 48]. These methods add imperceptible perturbations or nonsuspicious patches generated by adversarial attacks to original images, hindering DNNs to extract effective features and recognize identities, thus protecting the identity privacy in the image. However, in this work, we focus on visual information hiding, which means that the protected image is entirely different from the original image visually but can still be correctly predicted by DNNs (see Figure. 1). Fortunately, we observe a special type of adversarial attack (called Type-I attack) [43] quite different from the type used in previous methods. This type of attack guides DNNs to make consistent predictions on two distinct samples.", + "bbox": [ + 75, + 169, + 468, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the Type-I adversarial attack, in this paper, we propose an Adversarial Visual Information Hiding (AVIH) method. The proposed method hides the visual information in images while preserving their functional (e.g., recognition and classification) features for service model. It can recover original images from protected images for their owners. Specifically, we reduce the visual correlation between the protected and original image while minimizing their distance in the feature space of service model, to generate the protected image. Meanwhile, we exploit a generative model pre-trained in a private training setting as the key model, then optimize the protected image based on it so that the recovered image is similar to the original image. Furthermore, to break through the tough trade-off between the capability of privacy protection and the quality of restored image, we design the variance consistency loss to enhance privacy protection without compromising image recovery (see Section 3.3). Note that the protected image generated by our method can only be accurately recovered by the own key model, other models (even if the model architecture is the same) are difficult to recover well (see Section 4.3).", + "bbox": [ + 75, + 398, + 470, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AVIH can significantly improve the security and flexibility of image storage, which is extremely obvious in the protection of gallery sets for metric learning. Take the cloud-based face recognition system as an example. According to the service face recognition model, the face database manager can generate protected images locally or in the cloud and save the key model locally. The protected image contains no visual information and can be used by the service model to extract features correctly. Moreover, these protected images cannot be recognized by other models. Then, these protected images can be stored in the gallery set in the cloud for normal face recognition tasks. These protected", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images can be recovered using the key model when needed (such as maintaining the dataset or using face images for other tasks, etc.). The process is shown in Figure. 1.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking visual information hiding of gallery set images for cloud-deployed face recognition systems as the basic task, we conduct a comprehensive evaluation of our proposed method in terms of both effectiveness and security. In order to compare with existing methods suitable for information hiding tasks, we extend our method to classification tasks. Experimental results on multiple service models and datasets show that the proposed method is effective. In addition, to prove the effectiveness of our proposed loss, we conduct an ablation study about it to further present the advantages of our proposed method.", + "bbox": [ + 496, + 136, + 890, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 517, + 301, + 772, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Inspired by Type-I adversarial attacks, we propose a visual information hiding method AVIH. To alleviate the difficult trade-off between capability of information hiding and quality of recovered image, we design a variance consistency loss.", + "- Our proposed method has following properties: 1) The visual information in the image is clearly obfuscated. 2) Our method does not require retrain. 3) The protected image can be recovered by the own key model, but external models are difficult to recover.", + "- We validate the effectiveness of the proposed method on the face recognition task and the classification task. In addition, we conduct qualitative and quantitative ablation studies to show efficiency of the proposed loss." + ], + "bbox": [ + 517, + 327, + 890, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 580, + 640, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual Information Hiding", + "text_level": 1, + "bbox": [ + 500, + 604, + 740, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The visual information hiding of images is from the perspective of human vision. It is most directly manifested by the protected images being visually unrecognizable. Since Homomorphic Encryption (HE) only supports additive and multiplicative operations, HE are not suitable for nonlinear computations. Most DNNs contain a large number of nonlinear computations, so HE-based methods are hardly applicable to state-of-the-art DNNs [40]. The steganography methods hide sensitive information in a cover image and require as few changes to the cover image as possible. It does not need to guarantee that the service model can perform normal predictions on hidden information, which is very different from our work. Therefore, these two types of methods are not discussed in this paper. For the methods based on Perceptual Encryption (PE), some works [42, 40, 39, 8] focus on finding an encrypted domain and training the model directly using the encrypted images. However, this has a significant impact on the accuracy of", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4357", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b71863da0bc209636e8331b01482351beab4d50c352386ee0f1ce385d08cd8b4.jpg", + "image_caption": [ + "Figure 2. Overview of adversarial attack-based visual information hiding (AVIH) method. Taking the face recognition as an example, given a service model and a pre-trained key model, we protect the original image $x$ and obtain the protected image $x'$ ." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 890, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the model. To improve the accuracy of the classification model for protected images, Ito et al. [16] trained a transformation network to keep the classifier correctly classified while hiding visual information. However, the weakness of this method is that the protected image cannot be recovered.", + "bbox": [ + 75, + 345, + 468, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the current work used in classification and segmentation tasks, we mainly focus on metric learning tasks represented by face recognition, which are more prone to privacy leakage. To compare with existing methods that can achieve visual information hiding, we also extend the AVIH method to classification tasks. Our proposed method can also compensate for the drawbacks of the above methods. It generates protected images for a specific model that already exists. It takes advantage of the vulnerability in the model itself to provide strong privacy protection to the image. Since our method does not modify the service model, it does not affect the accuracy on the original image.", + "bbox": [ + 75, + 421, + 473, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Adversarial Attack", + "text_level": 1, + "bbox": [ + 76, + 617, + 261, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DNNs are vulnerable to some adversarial examples [9, 14]. There are many adversarial attack methods [9, 33, 6] to find adversarial examples efficiently. Tang et al. [43] divided the adversarial attacks into adversarial attack Type-I and adversarial attack Type-II based on the statistical Type I and Type II error. We take an optimization perspective on adversarial attacks. Then the Type-II attack maximizes the difference in the model output while ensuring a slight difference with the original input samples. Mathematically, it can be formulated as follows:", + "bbox": [ + 75, + 641, + 472, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {x ^ {\\prime}} f (x) - f \\left(x ^ {\\prime}\\right) \\quad s. t. \\quad \\| x ^ {\\prime} - x \\| < \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 804, + 468, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x$ is the original sample, $x'$ is the adversarial sample, and $f(\\cdot)$ is the model which is attacked. The Type-I attack, in contrast to the Type-II attack, looks for an input sample that differs the most from the original input sample", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "but makes the model output as same as possible. Mathematically, it can be formulated as follows:", + "bbox": [ + 498, + 345, + 890, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {x ^ {\\prime}} \\| x ^ {\\prime} - x \\| \\quad s. t. \\quad f (x) = f \\left(x ^ {\\prime}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 387, + 890, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Type-I attack on classification and generative models was implemented in the work of Tang et al. [43]. Then, Sun et al. [41] implemented a Type-I attack against variational autoencoder (VAE) [23]. In this work, we implement Type-I attacks on face recognition and classification tasks. Inspired by these attacks, we propose the AVIH method.", + "bbox": [ + 496, + 420, + 893, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 498, + 523, + 635, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective of our proposed method is to learn a protected image $x'$ which satisfies: 1) $x'$ is completely different from the original image $x$ , 2) for the service model the output $f_s(x')$ is the same as the $f_s(x)$ , and 3) $x'$ can be recovered as the $x$ by the key model. Mathematically,", + "bbox": [ + 496, + 550, + 893, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From $x\\in \\mathcal{X}$ Generate $x^{\\prime} = \\mathcal{A}(x)$", + "bbox": [ + 557, + 637, + 831, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "s.t. $\\| x^{\\prime} - x\\| >\\epsilon$ (3)", + "bbox": [ + 558, + 656, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s} \\left(x ^ {\\prime}\\right) = f _ {s} (x), k e y \\left(x ^ {\\prime}\\right) = x.\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 675, + 805, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of our method is shown in Figure. 2.", + "bbox": [ + 517, + 703, + 846, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image Visual Information Hiding", + "text_level": 1, + "bbox": [ + 500, + 728, + 794, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Exploiting the adversarial vulnerability of DNNs, we can perform Type-I attack on the service model to get images that are visually completely different from the original image but with extremely similar features. Suppose that $d(\\cdot)$ measures the difference between the outputs of the model. For different tasks, it behaves as different functions. Then for a particular service model $f_{s}$ , the difference between the output of the adversarial sample and the original sample is", + "bbox": [ + 496, + 751, + 890, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) = d \\left(f _ {s} (x), f _ {s} \\left(x ^ {\\prime}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4358", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We definite $\\mathcal{L}_d$ as", + "bbox": [ + 76, + 90, + 197, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} \\left(x ^ {\\prime}, x\\right) = \\left\\| x - x ^ {\\prime} \\right\\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 114, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "then the loss of the Type-I attack on the service model is formulated as", + "bbox": [ + 76, + 145, + 468, + 172 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) - \\lambda \\cdot \\mathcal {L} _ {d} \\left(x ^ {\\prime}, x\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 186, + 468, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a positive hyperparameter which balances image level differences and output level differences. Then with a certain number of iterations $K$ , we can optimize $\\mathcal{L}_r$ by the following operations to get the final adversarial sample.", + "bbox": [ + 76, + 214, + 468, + 275 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {k + 1} = \\alpha \\cdot g _ {k} + \\frac {\\nabla \\mathcal {L} \\left(x _ {k} ^ {\\prime} , x\\right)}{\\| \\nabla \\mathcal {L} \\left(x _ {k} ^ {\\prime} , x\\right) \\| _ {2} ^ {2}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 284, + 468, + 321 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {k + 1} ^ {\\prime} = x _ {k} ^ {\\prime} - \\beta \\cdot g _ {k + 1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 334, + 468, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 358, + 122, + 371 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {0} = \\frac {\\nabla \\mathcal {L} \\left(x _ {0} ^ {\\prime} , x\\right)}{\\| \\nabla \\mathcal {L} \\left(x _ {0} ^ {\\prime} , x\\right) \\| _ {2} ^ {2}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 368, + 468, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the first iteration, $x_0'$ can be randomly initialized, or it can be made $x_0' = x$ . The former provides an easier way to get an adversarial example with more significant visual differences. The latter provides a faster way to get an adversarial example that meets the criteria.", + "bbox": [ + 76, + 409, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Recovery", + "text_level": 1, + "bbox": [ + 76, + 494, + 236, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The protected image obtained by Equation. 6 can effectively hide visual information and keep the function of the sensitive information for the service model. However, the protected image is not recoverable. To achieve the goal of Equation. 3, we attach a recovery module that can recover the generated information hidden image. We refer to the generated images with perturbations as protected images. This perturbation has the function of protecting the visual information of the image.", + "bbox": [ + 76, + 518, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To get the protected images, we first train the generative model $G$ , which can generate the same images as the input. Then we perform Type-I attacks on both the service model and the generative model. We define $\\mathcal{L}_g$ as follows:", + "bbox": [ + 76, + 654, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right) = \\| x - G \\left(x ^ {\\prime}\\right) \\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 724, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The loss $\\mathcal{L}_g$ can help to keep the protected images recoverable by the key model we have chosen. Thus, the loss becomes as follows:", + "bbox": [ + 76, + 755, + 468, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) + \\mu \\cdot \\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 811, + 468, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ is a hyperparameter that balances the protection quality and restore quality. With Equation. 7 and Equation. 8, we can optimize $\\mathcal{L}_e$ to obtain the protected image $x^{\\prime}$ , which can satisfy the objective of Equation. 3.", + "bbox": [ + 76, + 840, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Adversarial visual information hiding method \nInput: Service model $f_{s}$ ; key model $G$ ; original image $x$ ; number of iterations $K$ ; gradient $g$ ; number of times the loss has increased $num$ ; maximum number of times the loss increase $maxnum$ .", + "bbox": [ + 500, + 90, + 890, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Protected images $x'$ .", + "bbox": [ + 500, + 167, + 697, + 183 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $g_0 = 0$ ; num = 0;", + "2: Random initialization $x_0'$ ;", + "3: for $k = 0$ to $K$ do", + "4: Compute $\\mathcal{L}_{AVIH}(x_k', x)$ via Equation. 15;" + ], + "bbox": [ + 509, + 186, + 823, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5: $g_{k + 1} = \\alpha \\cdot g_k + \\frac{\\nabla\\mathcal{L}_{AVIH}(x_k',x)}{\\left\\|\\nabla\\mathcal{L}_{AVIH}(x_k',x)\\right\\|_2^2}$ (Equation. 7);", + "bbox": [ + 509, + 246, + 870, + 272 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "6: $x_{k + 1}^{\\prime} = x_{k}^{\\prime} - \\beta \\cdot g_{k + 1}$ (Equation. 8);", + "7: if $\\mathcal{L}_{AVIH}(x_{k - 1}^{\\prime},x) < \\mathcal{L}_{AVIH}(x_{k}^{\\prime},x)$ then", + "8: $num = num + 1$" + ], + "bbox": [ + 509, + 273, + 833, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "9: end if", + "bbox": [ + 509, + 319, + 588, + 330 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: if num > maxnum then", + "11: $\\beta = 0.85\\cdot \\beta$", + "12: $num = 0$", + "13: end if", + "14: end for" + ], + "bbox": [ + 504, + 333, + 717, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Variance Consistency Loss", + "text_level": 1, + "bbox": [ + 500, + 435, + 738, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The protected image obtained by the objective function of Equation. 6 satisfies the requirements of Equation. 3, but its protection quality is not high. The obtained protected images have the problem of the difficult trade-off between protection quality and recovery quality. That is, if we want to obtain an image that is difficult to be cracked successfully, the quality of the image recovered by the key model will be poor. Another problem is that the obtained protected image, although differing greatly from the original in color, has obvious visual information that the original image has, which negatively impacts visual information protection. To solve the above problem, we propose a variance consistency loss. It improves the quality of protection by limiting the differences between each part of the image to make the protected image visually more confusing.", + "bbox": [ + 496, + 459, + 890, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the input image, we divide the image in each channel (R, G, B) into $N$ blocks respectively, i.e., $\\{b_1, b_2, \\ldots, b_N\\}_c$ , where $c \\in \\{R, G, B\\}$ , $b_n \\in \\mathbb{R}^{h \\times w}$ and $h, w$ denote the height and width of the block. The pixels of the blocks are allowed to have overlapping parts between them. Let $p_{i,j}^{b_n} \\in [0, 1]$ denote the normalized pixel value at $(i, j)$ in block $b_n$ . Then, we calculate the sum of each block:", + "bbox": [ + 496, + 685, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n} = \\sum_ {i = 1} ^ {h} \\sum_ {j = 1} ^ {w} p _ {i, j} ^ {b _ {n}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 801, + 890, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use $\\mathcal{S}$ denote the set of sum of the blocks for each channel, i.e., $S_{c} = \\{S_{1}, S_{2}, \\ldots, S_{N}\\}_{c}$ . In our practice, we convolve the image with a convolution kernel of size $h \\times w$ to", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4359", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "obtain blocks. Then, we calculate the variance of $S_{c}$ and obtain $\\sigma_c^2 = \\mathrm{var}(S_c)$ . Finally, we define the variance consistency loss as:", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right) = \\sigma_ {R} ^ {2} + \\sigma_ {G} ^ {2} + \\sigma_ {B} ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 146, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_{R}$ , $\\sigma_{G}$ and $\\sigma_{B}$ denote the variances for R, G, B channels, respectively. By minimizing $\\mathcal{L}_v$ , we can get a protected image with a more uniform distribution of pixel values. Proof-of-concept experiments (see Section 4.4) show that $\\mathcal{L}_v$ can eliminate the visual semantics in the protected images which are similar to the original images, and can help obtain protected images with high quality of protection and recovery. More details on the variance consistency loss can be found in Supplementary Material D.", + "bbox": [ + 75, + 175, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the variance consistency loss, the loss function in Equation. 6 is modified as:", + "bbox": [ + 75, + 311, + 468, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) + \\lambda \\cdot \\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 351, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the Equation. 11 is reformulated as:", + "bbox": [ + 76, + 380, + 344, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A V I H} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) + \\lambda \\cdot \\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right) + \\mu \\cdot \\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 404, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ are hyperparameters. This is our proposed AVIH framework for visual information hiding. The algorithm is summarized in Algorithm 1.", + "bbox": [ + 75, + 431, + 468, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. AVIH Method for Specific Tasks", + "text_level": 1, + "bbox": [ + 76, + 486, + 361, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method can provide effective protection for the stored image and can be applied to a wide range of tasks. In this work, we take face recognition and classification as examples to illustrate the ability of our method.", + "bbox": [ + 75, + 510, + 468, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the face recognition task, we want the features extracted by the service model $f_{s}$ from the protected and original images to be as same as possible. Thus, we modify $\\mathcal{L}_t$ as follows:", + "bbox": [ + 75, + 571, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) = \\left\\| f _ {s} (x) - f _ {s} \\left(x ^ {\\prime}\\right) \\right\\| _ {2} ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 638, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the classification task, compared with directly minimizing the mean square error (MSE) of $f_{s}(x)$ and $f_{s}(x^{\\prime})$ , we find that maximizing $f_{s}^{c}(x^{\\prime})$ where $c$ is the prediction class of the service model on the original image can more effectively reduce the impact of visual information hiding on the accuracy. We thus first obtain the logit output of the original image $f_{s}(x)$ and convert it to one-hot format $\\delta(f_{s}(x))$ . Then, we reify $\\mathcal{L}_t$ as follows:", + "bbox": [ + 75, + 669, + 468, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t}} \\left(x ^ {\\prime}, x\\right) = - \\delta \\left(f _ {s} (x)\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 801, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 829, + 209, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first verify the effectiveness of the AVIH method for face recognition tasks. Then, taking the face recognition system as an example, the security of the", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVIH method and the effectiveness of the variance consistency loss are explored. Finally, we verify the effectiveness of the AVIH method for classification tasks and compare it with other methods.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 162, + 705, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and service models. Our experiments were mainly evaluated based on the Labeled Face in the Wild (LFW) [13] dataset. We choose four face recognition models ArcFace [7], CosFace [44], SphereFace [30] used in [48], and AdaFace [22] to fully evaluate the performance of our method. Among them, the input size of ArcFace and AdaFace [22] is $112 \\times 112$ , and the input size of CosFace and SphereFace is $112 \\times 96$ . Therefore, we first use the MTCNN [49] to align and crop the face images to the corresponding face recognition model's input size.", + "bbox": [ + 496, + 185, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Key models. We choose the pix2pix framework [15] to train our key model. To pre-train the key model, we randomly selected 1,2878 images from the CelebA [31] dataset as the training set. Then we set the input and output of the model to the same image, set the batch size to 1, and train for 4 epochs. Finally, the trained generator is used directly as the key model. Unless otherwise stated, we use only one key model to protect the entire dataset.", + "bbox": [ + 496, + 337, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. To evaluate the effectiveness of our method more realistically and inspired by the evaluation method in MegaFace [21], we modified the evaluation method of LFW. We randomly selected 12 persons from the LFW dataset as the probe set. Each contains more than 12 facial images, comprising 355 images. The other 12878 images, we use as the gallery set. In the testing phase, we take one face image of a person in the probe set and put it into the gallery set. Then use the remaining images of this person as the test set. Next, we use the above-divided dataset to test the accuracy of the face recognition model. In this way, we put each person's image in the probe set into the gallery set in turn to measure the average accuracy. This metric can well demonstrate the impact of our method on face recognition models in practical applications.", + "bbox": [ + 496, + 458, + 890, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the Structural Similarity Index Measure (SSIM) [47] and Learned Perceptual Image Patch Similarity (LPIPS) to quantitatively measure the quality of the protected images. In this work, we use $\\mathrm{SSIM}_e / \\mathrm{LPIPS}_e$ to represent the average of the SSIM/LPIPS values between the protected image $x^{\\prime}$ and the original image $x$ , and use $\\mathrm{SSIM}_d / \\mathrm{LPIPS}_d$ to represent the average of the SSIM/LPIPS values between $x$ and the recovered image $G(x^{\\prime})$ . A higher SSIM indicates that the two images are more similar and a lower LPIPS indicates that the two images are more similar.", + "bbox": [ + 496, + 685, + 890, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Effectiveness of Face Privacy Protection", + "text_level": 1, + "bbox": [ + 500, + 845, + 841, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effectiveness and impact on service model. To the best of our knowledge, our method is the first (PE)-based", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4360", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/634eb4f4eb059f7510b7be3d5e15ae05cca855349c36f72bb1fbf41ed64c5617.jpg", + "image_caption": [ + "Figure 3. Adversarial Visual Information Hiding (AVIH) method protect face images for different service models. We mark below each protected image the cosine similarity between its feature vector and the original image's feature vector." + ], + "image_footnote": [], + "bbox": [ + 84, + 92, + 883, + 320 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b2e995de79acc286b79ff439d2c85b90c91d6991f93107775a0adf62dd04beec.jpg", + "table_caption": [ + "Table 1. Image visual quality metrics and cosine similarity (cosim) between the output features of the service model for original and protected images on LFW." + ], + "table_footnote": [], + "table_body": "
SSIMe↓LPIPSe↑SSIMd↑LPIPSd↓cos-sim ↑
AdaFace0.0280.8860.8930.1440.997
ArcFace0.0300.8880.9030.1280.994
CosFace0.0280.8910.8990.1280.998
SphereFace0.0290.8880.9060.1200.996
", + "bbox": [ + 84, + 440, + 460, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method for face recognition. There is no closely related method yet to compare the performance of protected images. Therefore, we conduct qualitative and quantitative evaluations of protected images. High visual metric scores and a very slight impact on accuracy demonstrate the effectiveness of our method.", + "bbox": [ + 75, + 566, + 468, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the AVIH method to protect the original face image $x$ in the probe set to test its effectiveness on face visual information hiding. Then, we used the obtained protected image $x'$ as the input of the key model $G(\\cdot)$ . Finally get the recovered image $G(x')$ . The results are shown in Figure. 3. From Figure. 3, it is evident that the protected image generated by our method is significantly different from the original image. Since no useful visual information is obtained from the protected image, it has the effect of visual information hiding for the original image. In Table. 1, the average SSIM value between the protected and original images for each service model is less than 0.04, while the cosine similarity between their features is higher than 0.99. Therefore, the protected image generated by AVIH can completely replace the original image while hiding the visual information.", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a3d40375fff7f5230cfe3768e9395701613a2ab9a56ae8eb0d968c9b16807463.jpg", + "table_caption": [ + "Table 2. Accuracy (percentage) of face recognition models for original image and different protected image on $LFW$ . The results of AVIH-ONE and AVIH-ALL are expected to be close to the results of original." + ], + "table_footnote": [], + "table_body": "
ModelsOriginalAVIH-ONE (%)AVIH-ALL (%)
AdaFace98.698.698.6
ArcFace96.596.596.5
CosFace89.489.289.3
SphereFace80.380.080.6
", + "bbox": [ + 501, + 454, + 890, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the impact of AVIH method on the accuracy of the face recognition models, we first generate a protected face image in the probe set before putting it into the gallery set. Then put the protected image into the gallery set instead of the original image. We tested the accuracy of the face recognition models in this way. The result AVIH-ONE is shown in Table. 2. Compared with the original accuracy of the models, it can be concluded that the AVIH method's impact on the models' accuracy is very slight, which can achieve the same accuracy in the ArcFace model.", + "bbox": [ + 496, + 579, + 892, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the actual situation, all the images stored in the gallery set are protected images. So we replaced all the images in the gallery set with protected images and then retested the accuracy of the models, as AVIH-ALL in Table. 2. In this case, our method has a very slight impact on the accuracy of the models and has a beneficial effect on the accuracy of the SphereFace model.", + "bbox": [ + 496, + 732, + 893, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test the average of SSIM values, as shown in Table. 1. Combined with Figure. 3 shows that the recovered image has good quality and can recover most of the visual information of the original image. Since the recovered images", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4361", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/58b660d8f78393e257072cb99766085175badd85065d4437424335db2987cde6.jpg", + "table_caption": [ + "Table 3. The performances of our method when the pre-trained model is based on non-face image (e.g., COCO). The numbers in parentheses represent the difference from the original accuracy of the model. For example, (-0.2) means the accuracy is 0.2 lower than the original." + ], + "table_footnote": [], + "table_body": "
ModelAVIH-ONE (%)SSIMe ↓SSIMd ↑
AdaFace98.6(-0.0)0.0260.891
ArcFace96.5(-0.0)0.0270.909
CosFace89.2(-0.2)0.0310.882
SphereFace79.9(-0.4)0.0310.890
", + "bbox": [ + 84, + 167, + 460, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c10fce1e5365f23a1b107ae16061179db1f80d3a2ee548b70de7d16a529f841.jpg", + "image_caption": [ + "Figure 4. Visual information hiding in real-world scenes. The cosine similarity between recognition features of protected images $(x^{\\prime})$ and original images $(x)$ is marked below $x^{\\prime}$ . The SSIM between recovered images $(G(x^{\\prime}))$ and original images is marked below $G(x^{\\prime})$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 276, + 464, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are generated by the generative model, a better-trained generative model can achieve better recovery quality. We also explore the time spent in protecting images of AVIH in Supplementary Materials A.", + "bbox": [ + 75, + 472, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Randomness of the dataset for training the key model. In this work, we use CelebA, which is also a face dataset, to train the key model. However, in real scenarios, large-scale face images are often difficult to obtain due to the privacy of each individual involved. So we changed CelebA to COCO [29] containing various objects to test the effectiveness of the new key model. We choose the test set of COCO as the train set and train 6 epochs on the key model. The results are shown in Table.3. It can be concluded that the impact of protected images on the model accuracy and the quality of recovered images are slightly different from the key model trained with CelebA. Therefore, we can use different types of datasets to train the key model, not just limited to face images.", + "bbox": [ + 75, + 534, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real world face privacy protection. We randomly selected real-world face photos taken with phones and then protected them to test the performance of the AVIH method for using real-world scenarios. Part of the results is shown in Figure. 4. In realistic scenarios, the cosine similarity of the features between the protected image generated by the AVIH method and the original image is still higher than 0.99. The recovered image can still be recovered well.", + "bbox": [ + 75, + 747, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to noise. We tested the quality of the recovery after adding Gaussian noise $X_{noise} \\sim \\mathrm{N}(0, \\sigma^2)$ to the", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f3b34b519e0bef1d392fe44e8204f47ab8c272eb0ad516239acc6b5c23fc703.jpg", + "table_caption": [ + "Table 4. Robustness of protected images against noise." + ], + "table_footnote": [], + "table_body": "
σ03/2555/2558/25510/255
SSIMd0.930.920.890.840.81
", + "bbox": [ + 511, + 109, + 880, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64f761b6b3634878d1c263ee78d425222d75ba6cb786b176dc8af7fda90eca3d.jpg", + "image_caption": [ + "Figure 5. Partial results of the key model randomness analysis. The left column is the key used for protection, and the top row is the key used for recovery. Every key model is trained with different initialization only. More images of the results are shown in Supplementary Material B." + ], + "image_footnote": [], + "bbox": [ + 517, + 172, + 880, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f82b43cbfa6f0e2b5b23e41f5242fefa9c864736e6ea431aace1fffc2b075be.jpg", + "image_caption": [ + "Figure 6. Crack the protected image using the attack model. The training dataset for the attack model is obtained using key1 protection. An attacker can recover images protected with his own key (key1), but cannot recover images protected with other keys." + ], + "image_footnote": [], + "bbox": [ + 516, + 407, + 885, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "protected image, as shown in Table 4. It can be concluded that our method is robust to small noise.", + "bbox": [ + 498, + 565, + 890, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Security Analysis", + "text_level": 1, + "bbox": [ + 500, + 604, + 669, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key model randomness analysis. To test the randomness of the key model, we use the same strategy as Ding et al. [8]. We first trained 16 key models using the same settings but with different initialization values. Different initialization values may result in significantly different key models. Then, we choose one of them in turn as the private key model to participate in the protection of the image, and use the others as the external key model to try to recover the protected image. From the results shown in Figure. 5, it can be concluded that the protected image obtained by one private key model cannot be recovered by other external key models, demonstrating that our proposed method can greatly improve the security of the face visual information. In addition, it is convenient that we can obtain mutually independent keys by simply changing the initialization without changing factors such as the structure of the key model and the training dataset. We suppose that one reason for this phenomenon is the instability of GAN training, where dif", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4362", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc85188b6de24d89443085d4ac6aa3f9e9a7a9aeafe6acc496b9e4a3939882c6.jpg", + "image_caption": [ + "Figure 7. Results of the ablation study. The SSIM value between the processed image and the original image is shown under each image. The own key represents the key model used for generating the protected image, and the external key represents other key models that are initialized differently during training." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 883, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ferent initialization values lead to different locally optimal solutions. Another reason could be that our method tends to find the boundary points of the input field corresponding to the output error allowed by the key model. However, the instability of GAN training causes this boundary to change significantly when the initialization value is changed.", + "bbox": [ + 75, + 422, + 470, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Responding to a possible attack. For the key model, even if the key model structure and its training settings are leaked, the protection is still difficult to break if the initialization point of the key training model is not leaked. As for the image, if pairs of original and protected images are leaked in large quantities, then an attacker can directly use these image to train a generative model to directly map the protected domain to the original domain. Ito et al. [17] exploited this method to evaluate the robustness of protection. So, we protect all the face images in the gallery set using the same key model, producing pairs of protected and original images. Then we use these paired images as a training set, and train a model using the same structure as the attack model. Using such a model, we attempt to recover the protected images in the probe set. The results are shown in Figure. 6. When the training set images and the protected image to be cracked are protected from the same key model, the attack model can recover the protected image. However, when they are protected with different key models, the trained model cannot crack the protected image. Therefore, we can increase the frequency of changing key models to effectively defend against such attacks.", + "bbox": [ + 75, + 518, + 468, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The protected images obtained by our method also have good statistical properties. The correlation analysis and histogram analysis are shown in Supplementary Material B.", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also verify that the protected image generated by a service face recognition model cannot be recognized by other models with the same function.", + "bbox": [ + 500, + 422, + 890, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 481, + 653, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the variance consistency loss. We compare the variance consistency loss $\\mathcal{L}_v$ with the distance loss $\\mathcal{L}_d$ (in Equation 5) to verify the effectiveness of $\\mathcal{L}_v$ . For each loss, we choose a set of suitable weights $\\lambda$ separately to represent its trade-off between the protection quality and recovery quality. Its results are shown in Figure 7. The protected image generated using $\\mathcal{L}_d$ has a smaller SSIM value compared to the protected image using $\\mathcal{L}_v$ , but a clear outline can be seen. So the visual information of the original image still exists. If the protected image is to be made free of such visual information, there is a significant loss in the quality of the recovered image. In contrast, using $\\mathcal{L}_v$ , it is possible to generate protected images with both no residual visual information and high recovery quality. We also use another key to recover the protected image to further investigate the effect of the two different losses on the protection quality. From Figure 7, it can be concluded that the protected image using $\\mathcal{L}_v$ is more difficult to be recovered by similar key models with different initialization while ensuring the quality of the recovered image. We compare $\\mathcal{L}_v$ with more other losses in Supplementary Material D to further demonstrate the effectiveness of $\\mathcal{L}_v$ .", + "bbox": [ + 496, + 506, + 892, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure. 5, the difference between the results when $\\lambda = 0.001$ and $\\lambda = 0.01$ is very slight for the protected and recovery quality of the protected images. This means that our method requires no carefully adjustment of", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4363", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26afe4253765ac1ba40a06862e3b501cccb7e3a3c08d1debebc91c0b5ab8e6d7.jpg", + "image_caption": [ + "Figure 8. Impact of parameter settings on protection quality and recovery quality. $k$ represents the size of the convolution kernel and $s$ represents the step size. For each pair of images, the former is the protected image and the latter is the recovery image. Below the recovery image is marked the SSIM value between it and the original image." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 233 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7ed96582425f1ae0113e64aa90651a34a0c6682e5995bb7c675928633d86ebfd.jpg", + "image_caption": [ + "Figure 9. Results of AVIH method for Resnet50 as the service model. Each column represents a different sample." + ], + "image_footnote": [], + "bbox": [ + 81, + 339, + 465, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the weight $\\lambda$ for $\\mathcal{L}_v$ , which is very convenient.", + "bbox": [ + 76, + 486, + 382, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Parameters of variance consistency loss. In the variance consistency loss calculation, the process of blocking the image and calculating the internal sum of each block can be regarded as a convolution kernel with all values of 1 to convolve the image. The convolution kernel size represents the size of the block, and the step size of the convolution kernel represents the degree of overlap of the block. We explored the effect of block size and degree of overlap on performance, as shown in Figure 8. When the overlap degree of blocks is certain (especially when it is highly overlapping), the quality of the recovery image decreases as the size of the blocks increases. When the size of blocks is certain, the protection quality decreases as the overlap degree of blocks decreases. In this work, we divide the image into $4 \\times 4$ blocks, with 12 pixels overlapping between adjacent blocks (step size is 1). This setting is applicable to most tasks and there is no need to change it for different samples.", + "bbox": [ + 75, + 503, + 468, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Privacy protection for classification models", + "text_level": 1, + "bbox": [ + 76, + 770, + 442, + 786 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We trained ResNet50 [11], VGG19 [37] using the train set of CIFAR-10 [25] and implemented AVIH method for the images of the test set. The results are shown in Figure. 9. We also compared two existing methods suitable for visual information hiding and the results are shown in Table. 5, where the ITP [16] do not recover protected images. Since the LIE [42] uses protected images to train the model, the", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cd6ceeda66587f6c022f3c3fe8835c0648e2ce354a5bfba927c93938e8aaba6a.jpg", + "table_caption": [ + "Table 5. Impact of privacy protection methods on the accuracy (percentage) of classification models on CIFAR-10." + ], + "table_footnote": [], + "table_body": "
MethodModelOriginal (%)Protect (%)
LIE [42]VGG1910.5987.78
Resnet5011.0091.53
ITP [16]VGG1993.9590.70
Resnet5095.5390.16
AVIH (Our)VGG1993.9593.95
Resnet5095.5395.28
", + "bbox": [ + 506, + 125, + 885, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "prediction accuracy on the original data is low. We show the visual metrics, the results of the two comparative methods, and the results on Imgnet in Supplementary Material C. Our method has minimal impact on the accuracy of the models, and the average SSIM value of the recovered images can reach above 0.9 for both models after our tests.", + "bbox": [ + 496, + 277, + 890, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 398, + 617, + 414 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a visual information hiding method AVIH based on Type-I attack. We evaluate our method by image protection of face recognition systems in the cloud. Experiments show that the AVIH method can protect images while preserving their functionality for the service model. We also propose a variance consistency loss to solve the problem of the difficult trade-off between protection quality and recovery quality. Finally, we use the AVIH method in classification tasks with satisfactory results. Our work provides a new perspective on image visual information protection, which has beneficial implications for adversarial learning communities and private protection communities. The AVIH method requires complete service model information. While this is feasible in most image storage situations, it limits the applicability of the AVIH method to a wider range of image protection scenarios. This is also a problem we will solve in future work.", + "bbox": [ + 496, + 429, + 890, + 685 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 62036007; in part by the Fundamental Research Funds for the Central Universities under Grants QTZX23042 and XJS221502; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01; in part by the Natural Science Basis Research Plan in Shaanxi Province of China under Grant 2022JQ-696; in part by the Fundamental Research Funds for the Central Universities and in part by the Innovation Fund of Xidian University under Grant YJSJ23012.", + "bbox": [ + 496, + 704, + 892, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4364", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Md Zahangir Alom, Chris Yakopcic, Mahmudul Hasan, Tarek M Taha, and Vijayan K Asari. Recurrent residual u-net for medical image segmentation. Journal of Medical Imaging, 6(1):014006-014006, 2019. 1", + "[2] Yoshinori Aono, Takuya Hayashi, Lihua Wang, Shiho Moriai, et al. Privacy-preserving deep learning via additively homomorphic encryption. IEEE Transactions on Information Forensics and Security, 13(5):1333-1345, 2017. 1", + "[3] L Arockiam and S Monikandan. Efficient cloud storage confidentiality to ensure data security. In 2014 International Conference on Computer Communication and Informatics, pages 1-5. IEEE, 2014. 1", + "[4] Deyan Chen and Hong Zhao. Data security and privacy protection issues in cloud computing. In 2012 International Conference on Computer Science and Electronics Engineering, volume 1, pages 647-651. IEEE, 2012. 1", + "[5] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michtitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE Transactions on Medical Imaging, 37(11):2453-2462, 2018. 1", + "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 3", + "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5", + "[8] Yi Ding, Guozheng Wu, Dajiang Chen, Ning Zhang, Linpeng Gong, Mingsheng Cao, and Zhiguang Qin. Deepedn: a deep-learning-based image encryption and decryption network for internet of medical things. IEEE Internet of Things Journal, 8(3):1504-1518, 2020. 1, 2, 7", + "[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2, 3", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9", + "[12] Lei Hu, Da Wei Zhou, Cai Xia Fu, Thomas Benkert, Chun Yu Jiang, Rui Ting Li, Li Ming Wei, and Jun Gong Zhao. Advanced zoomed diffusion-weighted imaging vs. full-field-of-view diffusion-weighted imaging in prostate cancer detection: a radiomic features study. European Radiology, 31:1760-1769, 2021. 1", + "[13] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2, 3", + "[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5", + "[16] Hiroki Ito, Yuma Kinoshita, Maungmaung Aprilpyone, and Hitoshi Kiya. Image to perturbation: An image transformation network for generating visually protected images for privacy-preserving deep neural networks. IEEE Access, 9:64629-64638, 2021. 1, 3, 9", + "[17] Hiroki Ito, Yuma Kinoshita, and Hitoshi Kiya. Image transformation network for privacy-preserving deep neural networks and its security evaluation. In 2020 IEEE 9th Global Conference on Consumer Electronics (GCCE), pages 822-825. IEEE, 2020. 8", + "[18] He Kaiming, Gkioxari Georgia, Dollar Piotr, and Girshick Ross. Mask r-cnn. IEEE Transactions on Pattern Analysis & Machine Intelligence, PP:1-1, 2017. 1", + "[19] Seny Kamara and Kristin Lauter. Cryptographic cloud storage. In International Conference on Financial Cryptography and Data Security, pages 136-149. Springer, 2010. 1", + "[20] Manjit Kaur and Vijay Kumar. A comprehensive review on image encryption techniques. Archives of Computational Methods in Engineering, 27(1):15-43, 2020. 1", + "[21] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 5", + "[22] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18750-18759, 2022. 5", + "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[24] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 1", + "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 9", + "[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1", + "[27] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Towards semi-supervised deep facial expression recognition with an adaptive confidence margin. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4166-4175, 2022. 1", + "[28] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Unconstrained facial expression recognition with no-reference de-elements learning. IEEE Transactions on Affective Computing, 2023. 1" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4365", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7", + "[30] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 5", + "[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5", + "[32] Qian Lou and Lei Jiang. She: A fast and accurate deep neural network for encrypted data. Advances in Neural Information Processing Systems, 32, 2019. 1", + "[33] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 3", + "[34] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 2", + "[35] Siani Pearson and Azzedine Benameur. Privacy, security and trust issues arising from cloud computing. In 2010 IEEE Second International Conference on Cloud Computing Technology and Science, pages 693-702. IEEE, 2010. 1", + "[36] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 acm sigsac conference on computer and communications security, pages 1528-1540, 2016. 2", + "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 9", + "[38] Ashish Singh and Kakali Chatterjee. Cloud security issues and challenges: A survey. Journal of Network and Computer Applications, 79:88-115, 2017. 1", + "[39] Warit Sirichotedumrong and Hitoshi Kiya. A gan-based image transformation scheme for privacy-preserving deep neural networks. In 2020 28th European Signal Processing Conference (EUSIPCO), pages 745-749. IEEE, 2021. 1, 2", + "[40] Warit Sirichotedumrong, Takahiro Maekawa, Yuma Kinoshita, and Hitoshi Kiya. Privacy-preserving deep neural networks with pixel-based image encryption considering data augmentation in the encrypted domain. In 2019 IEEE International Conference on Image Processing (ICIP), pages 674-678. IEEE, 2019. 1, 2", + "[41] Chengjin Sun, Sizhe Chen, Jia Cai, and Xiaolin Huang. Type i attack for generative models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 593-597. IEEE, 2020. 3", + "[42] Masayuki Tanaka. Learnable image encryption. In 2018 IEEE International Conference on Consumer Electronics-Taiwan (ICCE-TW), pages 1-2, 2018. 1, 2, 9" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Sanli Tang, Xiaolin Huang, Mingjian Chen, Chengjin Sun, and Jie Yang. Adversarial attack type i: Cheat classifiers by significant changes. IEEE transactions on pattern analysis and machine intelligence, 43(3):1100-1109, 2019. 2, 3", + "[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 5", + "[45] Xueyu Wang, Jiajun Huang, Siqi Ma, Surya Nepal, and Chang Xu. Deepfake disrupter: The detector of deepfake is my friend. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14920-14929, 2022. 1", + "[46] Yizhi Wang, Jun Lin, and Zhongfeng Wang. An efficient convolution core architecture for privacy-preserving deep learning. In 2018 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5. IEEE, 2018. 1", + "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[48] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 5", + "[49] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 5", + "[50] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 1", + "[51] Dawei Zhou, Nannan Wang, Chunlei Peng, Yi Yu, Xi Yang, and Xinbo Gao. Towards multi-domain face synthesis via domain-invariant representations and multi-level feature parts. IEEE Transactions on Multimedia, 24:3469-3479, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 654 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4366", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_model.json b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1bb50d32fad557fcc4ba590637efbcd87dd4ab26 --- /dev/null +++ b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "Hiding Visual Information via Obfuscating Adversarial Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.179, + 0.768, + 0.288 + ], + "angle": 0, + "content": "Zhigang Su\\(^{1,\\ast}\\) Dawei Zhou\\(^{1,\\ast}\\) Nannan Wang\\(^{1,\\dagger}\\) Decheng Liu\\(^{1,\\dagger}\\) Zhen Wang\\(^{2}\\) Xinbo Gao\\(^{3}\\) \n\\(^{1}\\)Xidian University, \\(^{2}\\)Zhejiang Lab \n\\(^{3}\\)Chongqing University of Posts and Telecommunications \n{zgsu, dwzhou}.xidian@gmail.com, {nnwang, dchliu}@xidian.edu.cn wangzhen@zhejianglab.com, gaoxb@cqupt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.643 + ], + "angle": 0, + "content": "Growing leakage and misuse of visual information raise security and privacy concerns, which promotes the development of information protection. Existing adversarial perturbations-based methods mainly focus on the de-identification against deep learning models. However, the inherent visual information of the data has not been well protected. In this work, inspired by the Type-I adversarial attack, we propose an Adversarial Visual Information Hiding (AVIH) method to protect the visual privacy of data. Specifically, the method generates obfuscating adversarial perturbations to obscure the visual information of the data. Meanwhile, it maintains the hidden objectives to be correctly predicted by models. In addition, our method does not modify the parameters of the applied model, which makes it flexible for different scenarios. Experimental results on the recognition and classification tasks demonstrate that the proposed method can effectively hide visual information and hardly affect the performances of models. The code is available at https://github.com/suzhigangssz/AVIH." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.685, + 0.21, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have been widely applied in the computer vision [26, 10, 18]. However, the increasing leakage and misuse of visual information has raised serious concerns [45, 20], especially in fields such as face [27, 28, 51] and medicine [5, 1, 12]. A representative case is the security issue of data stored in the cloud environment [19, 4, 35]. Due to potential vulnerabilities in cyberspace [3], uploaded private images can be easily stolen and used maliciously [38]. Therefore, it is meaningful and urgent to explore effective strategies to protect visual information." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.863, + 0.47, + 0.878 + ], + "angle": 0, + "content": "A classic strategy is visual information hiding, which" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.321, + 0.888, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.468, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Figure 1. Illustration of visual information hiding for face recognition systems in cloud environments. The gallery set is protected and provided to the DNN in the cloud. Protected image has quite different visual information from the original image, but it is still correctly identified. The protected gallery set can be recovered by a key model by the owner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.893, + 0.793 + ], + "angle": 0, + "content": "mainly consists of two types [40]: Homomorphic Encryption (HE)-based methods [2, 46, 32], Perceptual Encryption (PE)-based methods [42, 40, 39, 8, 16] and steganography method [24, 50]. Affected by the nonlinear activation functions in DNNs, HE-based methods are difficult to perform well on advanced DNNs [40]. Although PE-based methods apply to DNNs, existing methods typically require retraining with data in the encrypted domain to guarantee accuracy on encrypted data [42, 40]. This affects the performance of service model on raw data and cause additional resource consumption (especially for large models). The steganography method can hide the visual information of sensitive data into another image, but it does not guarantee that the service model can correctly recognize the protected image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To alleviate these negative effects, we expect to hide the visual information without making any modifications to service model. Namely, we hide sensitive visual information only by varying the input image. Previous researches have made some explorations in this regard. The transformation network-based methods [16], which try to protect the original image via a transformation function parameterized by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.346, + 0.902 + ], + "angle": 0, + "content": "*Equal contributions. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "a neural network, share the same philosophy. However, the method cannot easily recover the original image from the protected image for other purposes. They may suffer from adversarial vulnerability [9, 14] because the introduced neural network may be destroyed by adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.397 + ], + "angle": 0, + "content": "In fact, the negative effects of adversarial attacks can be utilized positively to protect privacy. Some works have exploited adversarial attacks for de-identification [36, 34, 48]. These methods add imperceptible perturbations or nonsuspicious patches generated by adversarial attacks to original images, hindering DNNs to extract effective features and recognize identities, thus protecting the identity privacy in the image. However, in this work, we focus on visual information hiding, which means that the protected image is entirely different from the original image visually but can still be correctly predicted by DNNs (see Figure. 1). Fortunately, we observe a special type of adversarial attack (called Type-I attack) [43] quite different from the type used in previous methods. This type of attack guides DNNs to make consistent predictions on two distinct samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.717 + ], + "angle": 0, + "content": "Inspired by the Type-I adversarial attack, in this paper, we propose an Adversarial Visual Information Hiding (AVIH) method. The proposed method hides the visual information in images while preserving their functional (e.g., recognition and classification) features for service model. It can recover original images from protected images for their owners. Specifically, we reduce the visual correlation between the protected and original image while minimizing their distance in the feature space of service model, to generate the protected image. Meanwhile, we exploit a generative model pre-trained in a private training setting as the key model, then optimize the protected image based on it so that the recovered image is similar to the original image. Furthermore, to break through the tough trade-off between the capability of privacy protection and the quality of restored image, we design the variance consistency loss to enhance privacy protection without compromising image recovery (see Section 3.3). Note that the protected image generated by our method can only be accurately recovered by the own key model, other models (even if the model architecture is the same) are difficult to recover well (see Section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "AVIH can significantly improve the security and flexibility of image storage, which is extremely obvious in the protection of gallery sets for metric learning. Take the cloud-based face recognition system as an example. According to the service face recognition model, the face database manager can generate protected images locally or in the cloud and save the key model locally. The protected image contains no visual information and can be used by the service model to extract features correctly. Moreover, these protected images cannot be recognized by other models. Then, these protected images can be stored in the gallery set in the cloud for normal face recognition tasks. These protected" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "images can be recovered using the key model when needed (such as maintaining the dataset or using face images for other tasks, etc.). The process is shown in Figure. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Taking visual information hiding of gallery set images for cloud-deployed face recognition systems as the basic task, we conduct a comprehensive evaluation of our proposed method in terms of both effectiveness and security. In order to compare with existing methods suitable for information hiding tasks, we extend our method to classification tasks. Experimental results on multiple service models and datasets show that the proposed method is effective. In addition, to prove the effectiveness of our proposed loss, we conduct an ablation study about it to further present the advantages of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.303, + 0.774, + 0.317 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.328, + 0.892, + 0.404 + ], + "angle": 0, + "content": "- Inspired by Type-I adversarial attacks, we propose a visual information hiding method AVIH. To alleviate the difficult trade-off between capability of information hiding and quality of recovered image, we design a variance consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.413, + 0.892, + 0.488 + ], + "angle": 0, + "content": "- Our proposed method has following properties: 1) The visual information in the image is clearly obfuscated. 2) Our method does not require retrain. 3) The protected image can be recovered by the own key model, but external models are difficult to recover." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.498, + 0.892, + 0.559 + ], + "angle": 0, + "content": "- We validate the effectiveness of the proposed method on the face recognition task and the classification task. In addition, we conduct qualitative and quantitative ablation studies to show efficiency of the proposed loss." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.328, + 0.892, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.642, + 0.597 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.741, + 0.622 + ], + "angle": 0, + "content": "2.1. Visual Information Hiding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The visual information hiding of images is from the perspective of human vision. It is most directly manifested by the protected images being visually unrecognizable. Since Homomorphic Encryption (HE) only supports additive and multiplicative operations, HE are not suitable for nonlinear computations. Most DNNs contain a large number of nonlinear computations, so HE-based methods are hardly applicable to state-of-the-art DNNs [40]. The steganography methods hide sensitive information in a cover image and require as few changes to the cover image as possible. It does not need to guarantee that the service model can perform normal predictions on hidden information, which is very different from our work. Therefore, these two types of methods are not discussed in this paper. For the methods based on Perceptual Encryption (PE), some works [42, 40, 39, 8] focus on finding an encrypted domain and training the model directly using the encrypted images. However, this has a significant impact on the accuracy of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.891, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.895, + 0.325 + ], + "angle": 0, + "content": "Figure 2. Overview of adversarial attack-based visual information hiding (AVIH) method. Taking the face recognition as an example, given a service model and a pre-trained key model, we protect the original image \\( x \\) and obtain the protected image \\( x' \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.47, + 0.422 + ], + "angle": 0, + "content": "the model. To improve the accuracy of the classification model for protected images, Ito et al. [16] trained a transformation network to keep the classifier correctly classified while hiding visual information. However, the weakness of this method is that the protected image cannot be recovered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.474, + 0.606 + ], + "angle": 0, + "content": "Unlike the current work used in classification and segmentation tasks, we mainly focus on metric learning tasks represented by face recognition, which are more prone to privacy leakage. To compare with existing methods that can achieve visual information hiding, we also extend the AVIH method to classification tasks. Our proposed method can also compensate for the drawbacks of the above methods. It generates protected images for a specific model that already exists. It takes advantage of the vulnerability in the model itself to provide strong privacy protection to the image. Since our method does not modify the service model, it does not affect the accuracy on the original image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.262, + 0.633 + ], + "angle": 0, + "content": "2.2. Adversarial Attack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.473, + 0.793 + ], + "angle": 0, + "content": "DNNs are vulnerable to some adversarial examples [9, 14]. There are many adversarial attack methods [9, 33, 6] to find adversarial examples efficiently. Tang et al. [43] divided the adversarial attacks into adversarial attack Type-I and adversarial attack Type-II based on the statistical Type I and Type II error. We take an optimization perspective on adversarial attacks. Then the Type-II attack maximizes the difference in the model output while ensuring a slight difference with the original input samples. Mathematically, it can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.805, + 0.47, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\max _ {x ^ {\\prime}} f (x) - f \\left(x ^ {\\prime}\\right) \\quad s. t. \\quad \\| x ^ {\\prime} - x \\| < \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\( x \\) is the original sample, \\( x' \\) is the adversarial sample, and \\( f(\\cdot) \\) is the model which is attacked. The Type-I attack, in contrast to the Type-II attack, looks for an input sample that differs the most from the original input sample" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.346, + 0.892, + 0.378 + ], + "angle": 0, + "content": "but makes the model output as same as possible. Mathematically, it can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.388, + 0.892, + 0.411 + ], + "angle": 0, + "content": "\\[\n\\max _ {x ^ {\\prime}} \\| x ^ {\\prime} - x \\| \\quad s. t. \\quad f (x) = f \\left(x ^ {\\prime}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.895, + 0.513 + ], + "angle": 0, + "content": "The Type-I attack on classification and generative models was implemented in the work of Tang et al. [43]. Then, Sun et al. [41] implemented a Type-I attack against variational autoencoder (VAE) [23]. In this work, we implement Type-I attacks on face recognition and classification tasks. Inspired by these attacks, we propose the AVIH method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.525, + 0.637, + 0.543 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.894, + 0.628 + ], + "angle": 0, + "content": "The objective of our proposed method is to learn a protected image \\( x' \\) which satisfies: 1) \\( x' \\) is completely different from the original image \\( x \\), 2) for the service model the output \\( f_s(x') \\) is the same as the \\( f_s(x) \\), and 3) \\( x' \\) can be recovered as the \\( x \\) by the key model. Mathematically," + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.638, + 0.833, + 0.655 + ], + "angle": 0, + "content": "From \\(x\\in \\mathcal{X}\\) Generate \\(x^{\\prime} = \\mathcal{A}(x)\\)" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.657, + 0.892, + 0.674 + ], + "angle": 0, + "content": "s.t. \\(\\| x^{\\prime} - x\\| >\\epsilon\\) (3)" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.676, + 0.806, + 0.692 + ], + "angle": 0, + "content": "\\[\nf _ {s} \\left(x ^ {\\prime}\\right) = f _ {s} (x), k e y \\left(x ^ {\\prime}\\right) = x.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.704, + 0.848, + 0.72 + ], + "angle": 0, + "content": "The pipeline of our method is shown in Figure. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.729, + 0.795, + 0.746 + ], + "angle": 0, + "content": "3.1. Image Visual Information Hiding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Exploiting the adversarial vulnerability of DNNs, we can perform Type-I attack on the service model to get images that are visually completely different from the original image but with extremely similar features. Suppose that \\( d(\\cdot) \\) measures the difference between the outputs of the model. For different tasks, it behaves as different functions. Then for a particular service model \\( f_{s} \\), the difference between the output of the adversarial sample and the original sample is" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) = d \\left(f _ {s} (x), f _ {s} \\left(x ^ {\\prime}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4358" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.198, + 0.106 + ], + "angle": 0, + "content": "We definite \\(\\mathcal{L}_d\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.115, + 0.469, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} \\left(x ^ {\\prime}, x\\right) = \\left\\| x - x ^ {\\prime} \\right\\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.146, + 0.47, + 0.174 + ], + "angle": 0, + "content": "then the loss of the Type-I attack on the service model is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.187, + 0.469, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) - \\lambda \\cdot \\mathcal {L} _ {d} \\left(x ^ {\\prime}, x\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.215, + 0.469, + 0.276 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a positive hyperparameter which balances image level differences and output level differences. Then with a certain number of iterations \\(K\\), we can optimize \\(\\mathcal{L}_r\\) by the following operations to get the final adversarial sample." + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.285, + 0.469, + 0.323 + ], + "angle": 0, + "content": "\\[\ng _ {k + 1} = \\alpha \\cdot g _ {k} + \\frac {\\nabla \\mathcal {L} \\left(x _ {k} ^ {\\prime} , x\\right)}{\\| \\nabla \\mathcal {L} \\left(x _ {k} ^ {\\prime} , x\\right) \\| _ {2} ^ {2}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.335, + 0.469, + 0.353 + ], + "angle": 0, + "content": "\\[\nx _ {k + 1} ^ {\\prime} = x _ {k} ^ {\\prime} - \\beta \\cdot g _ {k + 1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.359, + 0.124, + 0.372 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.369, + 0.469, + 0.405 + ], + "angle": 0, + "content": "\\[\ng _ {0} = \\frac {\\nabla \\mathcal {L} \\left(x _ {0} ^ {\\prime} , x\\right)}{\\| \\nabla \\mathcal {L} \\left(x _ {0} ^ {\\prime} , x\\right) \\| _ {2} ^ {2}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.469, + 0.485 + ], + "angle": 0, + "content": "At the first iteration, \\( x_0' \\) can be randomly initialized, or it can be made \\( x_0' = x \\). The former provides an easier way to get an adversarial example with more significant visual differences. The latter provides a faster way to get an adversarial example that meets the criteria." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.495, + 0.238, + 0.511 + ], + "angle": 0, + "content": "3.2. Image Recovery" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.519, + 0.469, + 0.653 + ], + "angle": 0, + "content": "The protected image obtained by Equation. 6 can effectively hide visual information and keep the function of the sensitive information for the service model. However, the protected image is not recoverable. To achieve the goal of Equation. 3, we attach a recovery module that can recover the generated information hidden image. We refer to the generated images with perturbations as protected images. This perturbation has the function of protecting the visual information of the image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.655, + 0.469, + 0.716 + ], + "angle": 0, + "content": "To get the protected images, we first train the generative model \\(G\\), which can generate the same images as the input. Then we perform Type-I attacks on both the service model and the generative model. We define \\(\\mathcal{L}_g\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.726, + 0.469, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right) = \\| x - G \\left(x ^ {\\prime}\\right) \\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.756, + 0.469, + 0.8 + ], + "angle": 0, + "content": "The loss \\(\\mathcal{L}_g\\) can help to keep the protected images recoverable by the key model we have chosen. Thus, the loss becomes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.812, + 0.469, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) + \\mu \\cdot \\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mu\\) is a hyperparameter that balances the protection quality and restore quality. With Equation. 7 and Equation. 8, we can optimize \\(\\mathcal{L}_e\\) to obtain the protected image \\(x^{\\prime}\\), which can satisfy the objective of Equation. 3." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.167 + ], + "angle": 0, + "content": "Algorithm 1 Adversarial visual information hiding method \nInput: Service model \\( f_{s} \\); key model \\( G \\); original image \\( x \\); number of iterations \\( K \\); gradient \\( g \\); number of times the loss has increased \\( num \\); maximum number of times the loss increase \\( maxnum \\)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.169, + 0.698, + 0.184 + ], + "angle": 0, + "content": "Output: Protected images \\( x' \\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.187, + 0.655, + 0.201 + ], + "angle": 0, + "content": "1: \\( g_0 = 0 \\); num = 0;" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.203, + 0.699, + 0.216 + ], + "angle": 0, + "content": "2: Random initialization \\( x_0' \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.218, + 0.653, + 0.23 + ], + "angle": 0, + "content": "3: for \\( k = 0 \\) to \\( K \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.232, + 0.825, + 0.246 + ], + "angle": 0, + "content": "4: Compute \\(\\mathcal{L}_{AVIH}(x_k', x)\\) via Equation. 15;" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.187, + 0.825, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.247, + 0.871, + 0.273 + ], + "angle": 0, + "content": "5: \\(g_{k + 1} = \\alpha \\cdot g_k + \\frac{\\nabla\\mathcal{L}_{AVIH}(x_k',x)}{\\left\\|\\nabla\\mathcal{L}_{AVIH}(x_k',x)\\right\\|_2^2}\\) (Equation. 7);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.274, + 0.79, + 0.288 + ], + "angle": 0, + "content": "6: \\(x_{k + 1}^{\\prime} = x_{k}^{\\prime} - \\beta \\cdot g_{k + 1}\\) (Equation. 8);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.289, + 0.834, + 0.303 + ], + "angle": 0, + "content": "7: if \\(\\mathcal{L}_{AVIH}(x_{k - 1}^{\\prime},x) < \\mathcal{L}_{AVIH}(x_{k}^{\\prime},x)\\) then" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.304, + 0.686, + 0.317 + ], + "angle": 0, + "content": "8: \\( num = num + 1 \\)" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.274, + 0.834, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.32, + 0.589, + 0.331 + ], + "angle": 0, + "content": "9: end if" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.334, + 0.718, + 0.346 + ], + "angle": 0, + "content": "10: if num > maxnum then" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.349, + 0.652, + 0.362 + ], + "angle": 0, + "content": "11: \\(\\beta = 0.85\\cdot \\beta\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.365, + 0.633, + 0.377 + ], + "angle": 0, + "content": "12: \\( num = 0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.379, + 0.59, + 0.391 + ], + "angle": 0, + "content": "13: end if" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.395, + 0.583, + 0.406 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.334, + 0.718, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.74, + 0.452 + ], + "angle": 0, + "content": "3.3. Variance Consistency Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.685 + ], + "angle": 0, + "content": "The protected image obtained by the objective function of Equation. 6 satisfies the requirements of Equation. 3, but its protection quality is not high. The obtained protected images have the problem of the difficult trade-off between protection quality and recovery quality. That is, if we want to obtain an image that is difficult to be cracked successfully, the quality of the image recovered by the key model will be poor. Another problem is that the obtained protected image, although differing greatly from the original in color, has obvious visual information that the original image has, which negatively impacts visual information protection. To solve the above problem, we propose a variance consistency loss. It improves the quality of protection by limiting the differences between each part of the image to make the protected image visually more confusing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.79 + ], + "angle": 0, + "content": "For the input image, we divide the image in each channel (R, G, B) into \\(N\\) blocks respectively, i.e., \\(\\{b_1, b_2, \\ldots, b_N\\}_c\\), where \\(c \\in \\{R, G, B\\}\\), \\(b_n \\in \\mathbb{R}^{h \\times w}\\) and \\(h, w\\) denote the height and width of the block. The pixels of the blocks are allowed to have overlapping parts between them. Let \\(p_{i,j}^{b_n} \\in [0, 1]\\) denote the normalized pixel value at \\((i, j)\\) in block \\(b_n\\). Then, we calculate the sum of each block:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.802, + 0.891, + 0.844 + ], + "angle": 0, + "content": "\\[\nS _ {n} = \\sum_ {i = 1} ^ {h} \\sum_ {j = 1} ^ {w} p _ {i, j} ^ {b _ {n}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use \\(\\mathcal{S}\\) denote the set of sum of the blocks for each channel, i.e., \\(S_{c} = \\{S_{1}, S_{2}, \\ldots, S_{N}\\}_{c}\\). In our practice, we convolve the image with a convolution kernel of size \\(h \\times w\\) to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4359" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "obtain blocks. Then, we calculate the variance of \\(S_{c}\\) and obtain \\(\\sigma_c^2 = \\mathrm{var}(S_c)\\). Finally, we define the variance consistency loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.147, + 0.469, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right) = \\sigma_ {R} ^ {2} + \\sigma_ {G} ^ {2} + \\sigma_ {B} ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.176, + 0.469, + 0.31 + ], + "angle": 0, + "content": "where \\(\\sigma_{R}\\), \\(\\sigma_{G}\\) and \\(\\sigma_{B}\\) denote the variances for R, G, B channels, respectively. By minimizing \\(\\mathcal{L}_v\\), we can get a protected image with a more uniform distribution of pixel values. Proof-of-concept experiments (see Section 4.4) show that \\(\\mathcal{L}_v\\) can eliminate the visual semantics in the protected images which are similar to the original images, and can help obtain protected images with high quality of protection and recovery. More details on the variance consistency loss can be found in Supplementary Material D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.469, + 0.342 + ], + "angle": 0, + "content": "Based on the variance consistency loss, the loss function in Equation. 6 is modified as:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.352, + 0.469, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) + \\lambda \\cdot \\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.381, + 0.345, + 0.395 + ], + "angle": 0, + "content": "and the Equation. 11 is reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.405, + 0.469, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A V I H} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) + \\lambda \\cdot \\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right) + \\mu \\cdot \\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.469, + 0.479 + ], + "angle": 0, + "content": "where \\(\\mu\\) are hyperparameters. This is our proposed AVIH framework for visual information hiding. The algorithm is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.487, + 0.362, + 0.504 + ], + "angle": 0, + "content": "3.4. AVIH Method for Specific Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.469, + 0.571 + ], + "angle": 0, + "content": "Our method can provide effective protection for the stored image and can be applied to a wide range of tasks. In this work, we take face recognition and classification as examples to illustrate the ability of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.469, + 0.631 + ], + "angle": 0, + "content": "For the face recognition task, we want the features extracted by the service model \\( f_{s} \\) from the protected and original images to be as same as possible. Thus, we modify \\( \\mathcal{L}_t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.64, + 0.469, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) = \\left\\| f _ {s} (x) - f _ {s} \\left(x ^ {\\prime}\\right) \\right\\| _ {2} ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.791 + ], + "angle": 0, + "content": "For the classification task, compared with directly minimizing the mean square error (MSE) of \\( f_{s}(x) \\) and \\( f_{s}(x^{\\prime}) \\), we find that maximizing \\( f_{s}^{c}(x^{\\prime}) \\) where \\( c \\) is the prediction class of the service model on the original image can more effectively reduce the impact of visual information hiding on the accuracy. We thus first obtain the logit output of the original image \\( f_{s}(x) \\) and convert it to one-hot format \\( \\delta(f_{s}(x)) \\). Then, we reify \\( \\mathcal{L}_t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.802, + 0.469, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t}} \\left(x ^ {\\prime}, x\\right) = - \\delta \\left(f _ {s} (x)\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right). \\tag {17}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.21, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we first verify the effectiveness of the AVIH method for face recognition tasks. Then, taking the face recognition system as an example, the security of the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "AVIH method and the effectiveness of the variance consistency loss are explored. Finally, we verify the effectiveness of the AVIH method for classification tasks and compare it with other methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.163, + 0.707, + 0.179 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Dataset and service models. Our experiments were mainly evaluated based on the Labeled Face in the Wild (LFW) [13] dataset. We choose four face recognition models ArcFace [7], CosFace [44], SphereFace [30] used in [48], and AdaFace [22] to fully evaluate the performance of our method. Among them, the input size of ArcFace and AdaFace [22] is \\( 112 \\times 112 \\), and the input size of CosFace and SphereFace is \\( 112 \\times 96 \\). Therefore, we first use the MTCNN [49] to align and crop the face images to the corresponding face recognition model's input size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Key models. We choose the pix2pix framework [15] to train our key model. To pre-train the key model, we randomly selected 1,2878 images from the CelebA [31] dataset as the training set. Then we set the input and output of the model to the same image, set the batch size to 1, and train for 4 epochs. Finally, the trained generator is used directly as the key model. Unless otherwise stated, we use only one key model to protect the entire dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Evaluation metrics. To evaluate the effectiveness of our method more realistically and inspired by the evaluation method in MegaFace [21], we modified the evaluation method of LFW. We randomly selected 12 persons from the LFW dataset as the probe set. Each contains more than 12 facial images, comprising 355 images. The other 12878 images, we use as the gallery set. In the testing phase, we take one face image of a person in the probe set and put it into the gallery set. Then use the remaining images of this person as the test set. Next, we use the above-divided dataset to test the accuracy of the face recognition model. In this way, we put each person's image in the probe set into the gallery set in turn to measure the average accuracy. This metric can well demonstrate the impact of our method on face recognition models in practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.836 + ], + "angle": 0, + "content": "We use the Structural Similarity Index Measure (SSIM) [47] and Learned Perceptual Image Patch Similarity (LPIPS) to quantitatively measure the quality of the protected images. In this work, we use \\(\\mathrm{SSIM}_e / \\mathrm{LPIPS}_e\\) to represent the average of the SSIM/LPIPS values between the protected image \\(x^{\\prime}\\) and the original image \\(x\\), and use \\(\\mathrm{SSIM}_d / \\mathrm{LPIPS}_d\\) to represent the average of the SSIM/LPIPS values between \\(x\\) and the recovered image \\(G(x^{\\prime})\\). A higher SSIM indicates that the two images are more similar and a lower LPIPS indicates that the two images are more similar." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.842, + 0.863 + ], + "angle": 0, + "content": "4.2. Effectiveness of Face Privacy Protection" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness and impact on service model. To the best of our knowledge, our method is the first (PE)-based" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4360" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.093, + 0.885, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.34, + 0.893, + 0.369 + ], + "angle": 0, + "content": "Figure 3. Adversarial Visual Information Hiding (AVIH) method protect face images for different service models. We mark below each protected image the cosine similarity between its feature vector and the original image's feature vector." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.389, + 0.471, + 0.432 + ], + "angle": 0, + "content": "Table 1. Image visual quality metrics and cosine similarity (cosim) between the output features of the service model for original and protected images on LFW." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.441, + 0.462, + 0.528 + ], + "angle": 0, + "content": "
SSIMe↓LPIPSe↑SSIMd↑LPIPSd↓cos-sim ↑
AdaFace0.0280.8860.8930.1440.997
ArcFace0.0300.8880.9030.1280.994
CosFace0.0280.8910.8990.1280.998
SphereFace0.0290.8880.9060.1200.996
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.657 + ], + "angle": 0, + "content": "method for face recognition. There is no closely related method yet to compare the performance of protected images. Therefore, we conduct qualitative and quantitative evaluations of protected images. High visual metric scores and a very slight impact on accuracy demonstrate the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We use the AVIH method to protect the original face image \\( x \\) in the probe set to test its effectiveness on face visual information hiding. Then, we used the obtained protected image \\( x' \\) as the input of the key model \\( G(\\cdot) \\). Finally get the recovered image \\( G(x') \\). The results are shown in Figure. 3. From Figure. 3, it is evident that the protected image generated by our method is significantly different from the original image. Since no useful visual information is obtained from the protected image, it has the effect of visual information hiding for the original image. In Table. 1, the average SSIM value between the protected and original images for each service model is less than 0.04, while the cosine similarity between their features is higher than 0.99. Therefore, the protected image generated by AVIH can completely replace the original image while hiding the visual information." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.389, + 0.893, + 0.445 + ], + "angle": 0, + "content": "Table 2. Accuracy (percentage) of face recognition models for original image and different protected image on \\( LFW \\). The results of AVIH-ONE and AVIH-ALL are expected to be close to the results of original." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.455, + 0.892, + 0.543 + ], + "angle": 0, + "content": "
ModelsOriginalAVIH-ONE (%)AVIH-ALL (%)
AdaFace98.698.698.6
ArcFace96.596.596.5
CosFace89.489.289.3
SphereFace80.380.080.6
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.732 + ], + "angle": 0, + "content": "To evaluate the impact of AVIH method on the accuracy of the face recognition models, we first generate a protected face image in the probe set before putting it into the gallery set. Then put the protected image into the gallery set instead of the original image. We tested the accuracy of the face recognition models in this way. The result AVIH-ONE is shown in Table. 2. Compared with the original accuracy of the models, it can be concluded that the AVIH method's impact on the models' accuracy is very slight, which can achieve the same accuracy in the ArcFace model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.894, + 0.839 + ], + "angle": 0, + "content": "Considering the actual situation, all the images stored in the gallery set are protected images. So we replaced all the images in the gallery set with protected images and then retested the accuracy of the models, as AVIH-ALL in Table. 2. In this case, our method has a very slight impact on the accuracy of the models and has a beneficial effect on the accuracy of the SphereFace model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We test the average of SSIM values, as shown in Table. 1. Combined with Figure. 3 shows that the recovered image has good quality and can recover most of the visual information of the original image. Since the recovered images" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4361" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.471, + 0.158 + ], + "angle": 0, + "content": "Table 3. The performances of our method when the pre-trained model is based on non-face image (e.g., COCO). The numbers in parentheses represent the difference from the original accuracy of the model. For example, (-0.2) means the accuracy is 0.2 lower than the original." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.168, + 0.462, + 0.255 + ], + "angle": 0, + "content": "
ModelAVIH-ONE (%)SSIMe ↓SSIMd ↑
AdaFace98.6(-0.0)0.0260.891
ArcFace96.5(-0.0)0.0270.909
CosFace89.2(-0.2)0.0310.882
SphereFace79.9(-0.4)0.0310.890
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.277, + 0.465, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Figure 4. Visual information hiding in real-world scenes. The cosine similarity between recognition features of protected images \\((x^{\\prime})\\) and original images \\((x)\\) is marked below \\(x^{\\prime}\\). The SSIM between recovered images \\((G(x^{\\prime}))\\) and original images is marked below \\(G(x^{\\prime})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.534 + ], + "angle": 0, + "content": "are generated by the generative model, a better-trained generative model can achieve better recovery quality. We also explore the time spent in protecting images of AVIH in Supplementary Materials A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.469, + 0.746 + ], + "angle": 0, + "content": "Randomness of the dataset for training the key model. In this work, we use CelebA, which is also a face dataset, to train the key model. However, in real scenarios, large-scale face images are often difficult to obtain due to the privacy of each individual involved. So we changed CelebA to COCO [29] containing various objects to test the effectiveness of the new key model. We choose the test set of COCO as the train set and train 6 epochs on the key model. The results are shown in Table.3. It can be concluded that the impact of protected images on the model accuracy and the quality of recovered images are slightly different from the key model trained with CelebA. Therefore, we can use different types of datasets to train the key model, not just limited to face images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Real world face privacy protection. We randomly selected real-world face photos taken with phones and then protected them to test the performance of the AVIH method for using real-world scenarios. Part of the results is shown in Figure. 4. In realistic scenarios, the cosine similarity of the features between the protected image generated by the AVIH method and the original image is still higher than 0.99. The recovered image can still be recovered well." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Robustness to noise. We tested the quality of the recovery after adding Gaussian noise \\(X_{noise} \\sim \\mathrm{N}(0, \\sigma^2)\\) to the" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.088, + 0.857, + 0.102 + ], + "angle": 0, + "content": "Table 4. Robustness of protected images against noise." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.11, + 0.882, + 0.154 + ], + "angle": 0, + "content": "
σ03/2555/2558/25510/255
SSIMd0.930.920.890.840.81
" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.174, + 0.882, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.317, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Figure 5. Partial results of the key model randomness analysis. The left column is the key used for protection, and the top row is the key used for recovery. Every key model is trained with different initialization only. More images of the results are shown in Supplementary Material B." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.408, + 0.886, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.494, + 0.891, + 0.551 + ], + "angle": 0, + "content": "Figure 6. Crack the protected image using the attack model. The training dataset for the attack model is obtained using key1 protection. An attacker can recover images protected with his own key (key1), but cannot recover images protected with other keys." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.566, + 0.891, + 0.595 + ], + "angle": 0, + "content": "protected image, as shown in Table 4. It can be concluded that our method is robust to small noise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.671, + 0.623 + ], + "angle": 0, + "content": "4.3. Security Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Key model randomness analysis. To test the randomness of the key model, we use the same strategy as Ding et al. [8]. We first trained 16 key models using the same settings but with different initialization values. Different initialization values may result in significantly different key models. Then, we choose one of them in turn as the private key model to participate in the protection of the image, and use the others as the external key model to try to recover the protected image. From the results shown in Figure. 5, it can be concluded that the protected image obtained by one private key model cannot be recovered by other external key models, demonstrating that our proposed method can greatly improve the security of the face visual information. In addition, it is convenient that we can obtain mutually independent keys by simply changing the initialization without changing factors such as the structure of the key model and the training dataset. We suppose that one reason for this phenomenon is the instability of GAN training, where dif" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4362" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.885, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.357, + 0.893, + 0.4 + ], + "angle": 0, + "content": "Figure 7. Results of the ablation study. The SSIM value between the processed image and the original image is shown under each image. The own key represents the key model used for generating the protected image, and the external key represents other key models that are initialized differently during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.471, + 0.515 + ], + "angle": 0, + "content": "ferent initialization values lead to different locally optimal solutions. Another reason could be that our method tends to find the boundary points of the input field corresponding to the output error allowed by the key model. However, the instability of GAN training causes this boundary to change significantly when the initialization value is changed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Responding to a possible attack. For the key model, even if the key model structure and its training settings are leaked, the protection is still difficult to break if the initialization point of the key training model is not leaked. As for the image, if pairs of original and protected images are leaked in large quantities, then an attacker can directly use these image to train a generative model to directly map the protected domain to the original domain. Ito et al. [17] exploited this method to evaluate the robustness of protection. So, we protect all the face images in the gallery set using the same key model, producing pairs of protected and original images. Then we use these paired images as a training set, and train a model using the same structure as the attack model. Using such a model, we attempt to recover the protected images in the probe set. The results are shown in Figure. 6. When the training set images and the protected image to be cracked are protected from the same key model, the attack model can recover the protected image. However, when they are protected with different key models, the trained model cannot crack the protected image. Therefore, we can increase the frequency of changing key models to effectively defend against such attacks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The protected images obtained by our method also have good statistical properties. The correlation analysis and histogram analysis are shown in Supplementary Material B." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.424, + 0.892, + 0.469 + ], + "angle": 0, + "content": "We also verify that the protected image generated by a service face recognition model cannot be recognized by other models with the same function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.655, + 0.499 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.839 + ], + "angle": 0, + "content": "Effectiveness of the variance consistency loss. We compare the variance consistency loss \\(\\mathcal{L}_v\\) with the distance loss \\(\\mathcal{L}_d\\) (in Equation 5) to verify the effectiveness of \\(\\mathcal{L}_v\\). For each loss, we choose a set of suitable weights \\(\\lambda\\) separately to represent its trade-off between the protection quality and recovery quality. Its results are shown in Figure 7. The protected image generated using \\(\\mathcal{L}_d\\) has a smaller SSIM value compared to the protected image using \\(\\mathcal{L}_v\\), but a clear outline can be seen. So the visual information of the original image still exists. If the protected image is to be made free of such visual information, there is a significant loss in the quality of the recovered image. In contrast, using \\(\\mathcal{L}_v\\), it is possible to generate protected images with both no residual visual information and high recovery quality. We also use another key to recover the protected image to further investigate the effect of the two different losses on the protection quality. From Figure 7, it can be concluded that the protected image using \\(\\mathcal{L}_v\\) is more difficult to be recovered by similar key models with different initialization while ensuring the quality of the recovered image. We compare \\(\\mathcal{L}_v\\) with more other losses in Supplementary Material D to further demonstrate the effectiveness of \\(\\mathcal{L}_v\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure. 5, the difference between the results when \\(\\lambda = 0.001\\) and \\(\\lambda = 0.01\\) is very slight for the protected and recovery quality of the protected images. This means that our method requires no carefully adjustment of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4363" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.327 + ], + "angle": 0, + "content": "Figure 8. Impact of parameter settings on protection quality and recovery quality. \\( k \\) represents the size of the convolution kernel and \\( s \\) represents the step size. For each pair of images, the former is the protected image and the latter is the recovery image. Below the recovery image is marked the SSIM value between it and the original image." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.34, + 0.466, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.433, + 0.468, + 0.462 + ], + "angle": 0, + "content": "Figure 9. Results of AVIH method for Resnet50 as the service model. Each column represents a different sample." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.383, + 0.502 + ], + "angle": 0, + "content": "the weight \\(\\lambda\\) for \\(\\mathcal{L}_v\\), which is very convenient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.469, + 0.76 + ], + "angle": 0, + "content": "Parameters of variance consistency loss. In the variance consistency loss calculation, the process of blocking the image and calculating the internal sum of each block can be regarded as a convolution kernel with all values of 1 to convolve the image. The convolution kernel size represents the size of the block, and the step size of the convolution kernel represents the degree of overlap of the block. We explored the effect of block size and degree of overlap on performance, as shown in Figure 8. When the overlap degree of blocks is certain (especially when it is highly overlapping), the quality of the recovery image decreases as the size of the blocks increases. When the size of blocks is certain, the protection quality decreases as the overlap degree of blocks decreases. In this work, we divide the image into \\(4 \\times 4\\) blocks, with 12 pixels overlapping between adjacent blocks (step size is 1). This setting is applicable to most tasks and there is no need to change it for different samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.443, + 0.787 + ], + "angle": 0, + "content": "4.5. Privacy protection for classification models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We trained ResNet50 [11], VGG19 [37] using the train set of CIFAR-10 [25] and implemented AVIH method for the images of the test set. The results are shown in Figure. 9. We also compared two existing methods suitable for visual information hiding and the results are shown in Table. 5, where the ITP [16] do not recover protected images. Since the LIE [42] uses protected images to train the model, the" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.115 + ], + "angle": 0, + "content": "Table 5. Impact of privacy protection methods on the accuracy (percentage) of classification models on CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.126, + 0.886, + 0.23 + ], + "angle": 0, + "content": "
MethodModelOriginal (%)Protect (%)
LIE [42]VGG1910.5987.78
Resnet5011.0091.53
ITP [16]VGG1993.9590.70
Resnet5095.5390.16
AVIH (Our)VGG1993.9593.95
Resnet5095.5395.28
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.368 + ], + "angle": 0, + "content": "prediction accuracy on the original data is low. We show the visual metrics, the results of the two comparative methods, and the results on Imgnet in Supplementary Material C. Our method has minimal impact on the accuracy of the models, and the average SSIM value of the recovered images can reach above 0.9 for both models after our tests." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.399, + 0.618, + 0.415 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.686 + ], + "angle": 0, + "content": "In this paper, we propose a visual information hiding method AVIH based on Type-I attack. We evaluate our method by image protection of face recognition systems in the cloud. Experiments show that the AVIH method can protect images while preserving their functionality for the service model. We also propose a variance consistency loss to solve the problem of the difficult trade-off between protection quality and recovery quality. Finally, we use the AVIH method in classification tasks with satisfactory results. Our work provides a new perspective on image visual information protection, which has beneficial implications for adversarial learning communities and private protection communities. The AVIH method requires complete service model information. While this is feasible in most image storage situations, it limits the applicability of the AVIH method to a wider range of image protection scenarios. This is also a problem we will solve in future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 62036007; in part by the Fundamental Research Funds for the Central Universities under Grants QTZX23042 and XJS221502; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01; in part by the Natural Science Basis Research Plan in Shaanxi Province of China under Grant 2022JQ-696; in part by the Fundamental Research Funds for the Central Universities and in part by the Innovation Fund of Xidian University under Grant YJSJ23012." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4364" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Md Zahangir Alom, Chris Yakopcic, Mahmudul Hasan, Tarek M Taha, and Vijayan K Asari. Recurrent residual u-net for medical image segmentation. Journal of Medical Imaging, 6(1):014006-014006, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Yoshinori Aono, Takuya Hayashi, Lihua Wang, Shiho Moriai, et al. Privacy-preserving deep learning via additively homomorphic encryption. IEEE Transactions on Information Forensics and Security, 13(5):1333-1345, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] L Arockiam and S Monikandan. Efficient cloud storage confidentiality to ensure data security. In 2014 International Conference on Computer Communication and Informatics, pages 1-5. IEEE, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[4] Deyan Chen and Hong Zhao. Data security and privacy protection issues in cloud computing. In 2012 International Conference on Computer Science and Electronics Engineering, volume 1, pages 647-651. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[5] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michtitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE Transactions on Medical Imaging, 37(11):2453-2462, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.452 + ], + "angle": 0, + "content": "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Yi Ding, Guozheng Wu, Dajiang Chen, Ning Zhang, Linpeng Gong, Mingsheng Cao, and Zhiguang Qin. Deepedn: a deep-learning-based image encryption and decryption network for internet of medical things. IEEE Internet of Things Journal, 8(3):1504-1518, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[12] Lei Hu, Da Wei Zhou, Cai Xia Fu, Thomas Benkert, Chun Yu Jiang, Rui Ting Li, Li Ming Wei, and Jun Gong Zhao. Advanced zoomed diffusion-weighted imaging vs. full-field-of-view diffusion-weighted imaging in prostate cancer detection: a radiomic features study. European Radiology, 31:1760-1769, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[16] Hiroki Ito, Yuma Kinoshita, Maungmaung Aprilpyone, and Hitoshi Kiya. Image to perturbation: An image transformation network for generating visually protected images for privacy-preserving deep neural networks. IEEE Access, 9:64629-64638, 2021. 1, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[17] Hiroki Ito, Yuma Kinoshita, and Hitoshi Kiya. Image transformation network for privacy-preserving deep neural networks and its security evaluation. In 2020 IEEE 9th Global Conference on Consumer Electronics (GCCE), pages 822-825. IEEE, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] He Kaiming, Gkioxari Georgia, Dollar Piotr, and Girshick Ross. Mask r-cnn. IEEE Transactions on Pattern Analysis & Machine Intelligence, PP:1-1, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[19] Seny Kamara and Kristin Lauter. Cryptographic cloud storage. In International Conference on Financial Cryptography and Data Security, pages 136-149. Springer, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[20] Manjit Kaur and Vijay Kumar. A comprehensive review on image encryption techniques. Archives of Computational Methods in Engineering, 27(1):15-43, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[22] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18750-18759, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[24] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[27] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Towards semi-supervised deep facial expression recognition with an adaptive confidence margin. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4166-4175, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Unconstrained facial expression recognition with no-reference de-elements learning. IEEE Transactions on Affective Computing, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4365" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.23 + ], + "angle": 0, + "content": "[30] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[32] Qian Lou and Lei Jiang. She: A fast and accurate deep neural network for encrypted data. Advances in Neural Information Processing Systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[33] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[34] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[35] Siani Pearson and Azzedine Benameur. Privacy, security and trust issues arising from cloud computing. In 2010 IEEE Second International Conference on Cloud Computing Technology and Science, pages 693-702. IEEE, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[36] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 acm sigsac conference on computer and communications security, pages 1528-1540, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.622 + ], + "angle": 0, + "content": "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[38] Ashish Singh and Kakali Chatterjee. Cloud security issues and challenges: A survey. Journal of Network and Computer Applications, 79:88-115, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[39] Warit Sirichotedumrong and Hitoshi Kiya. A gan-based image transformation scheme for privacy-preserving deep neural networks. In 2020 28th European Signal Processing Conference (EUSIPCO), pages 745-749. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[40] Warit Sirichotedumrong, Takahiro Maekawa, Yuma Kinoshita, and Hitoshi Kiya. Privacy-preserving deep neural networks with pixel-based image encryption considering data augmentation in the encrypted domain. In 2019 IEEE International Conference on Image Processing (ICIP), pages 674-678. IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[41] Chengjin Sun, Sizhe Chen, Jia Cai, and Xiaolin Huang. Type i attack for generative models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 593-597. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[42] Masayuki Tanaka. Learnable image encryption. In 2018 IEEE International Conference on Consumer Electronics-Taiwan (ICCE-TW), pages 1-2, 2018. 1, 2, 9" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Sanli Tang, Xiaolin Huang, Mingjian Chen, Chengjin Sun, and Jie Yang. Adversarial attack type i: Cheat classifiers by significant changes. IEEE transactions on pattern analysis and machine intelligence, 43(3):1100-1109, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[45] Xueyu Wang, Jiajun Huang, Siqi Ma, Surya Nepal, and Chang Xu. Deepfake disrupter: The detector of deepfake is my friend. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14920-14929, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[46] Yizhi Wang, Jun Lin, and Zhongfeng Wang. An efficient convolution core architecture for privacy-preserving deep learning. In 2018 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[48] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[49] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[50] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[51] Dawei Zhou, Nannan Wang, Chunlei Peng, Yi Yu, Xi Yang, and Xinbo Gao. Towards multi-domain face synthesis via domain-invariant representations and multi-level feature parts. IEEE Transactions on Multimedia, 24:3469-3479, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4366" + } + ] +] \ No newline at end of file diff --git a/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_origin.pdf b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..78924a76c7357db3f797b6a27f2ffc422ee9d89c --- /dev/null +++ b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/a690bb1f-1210-4744-bd66-f678e3096b46_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cffac5544968ab699e30f2fb9cc76b458de8ffa649411e5020176721872d0cd4 +size 1876397 diff --git a/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/full.md b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..feed344f77f8b1aea0395f656da1ac6445cb24e1 --- /dev/null +++ b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/full.md @@ -0,0 +1,387 @@ +# Hiding Visual Information via Obfuscating Adversarial Perturbations + +Zhigang Su $^{1,\ast}$ Dawei Zhou $^{1,\ast}$ Nannan Wang $^{1,\dagger}$ Decheng Liu $^{1,\dagger}$ Zhen Wang $^{2}$ Xinbo Gao $^{3}$ $^{1}$ Xidian University, $^{2}$ Zhejiang Lab + $^{3}$ Chongqing University of Posts and Telecommunications +{zgsu, dwzhou}.xidian@gmail.com, {nnwang, dchliu}@xidian.edu.cn wangzhen@zhejianglab.com, gaoxb@cqupt.edu.cn + +# Abstract + +Growing leakage and misuse of visual information raise security and privacy concerns, which promotes the development of information protection. Existing adversarial perturbations-based methods mainly focus on the de-identification against deep learning models. However, the inherent visual information of the data has not been well protected. In this work, inspired by the Type-I adversarial attack, we propose an Adversarial Visual Information Hiding (AVIH) method to protect the visual privacy of data. Specifically, the method generates obfuscating adversarial perturbations to obscure the visual information of the data. Meanwhile, it maintains the hidden objectives to be correctly predicted by models. In addition, our method does not modify the parameters of the applied model, which makes it flexible for different scenarios. Experimental results on the recognition and classification tasks demonstrate that the proposed method can effectively hide visual information and hardly affect the performances of models. The code is available at https://github.com/suzhigangssz/AVIH. + +# 1. Introduction + +Deep neural networks (DNNs) have been widely applied in the computer vision [26, 10, 18]. However, the increasing leakage and misuse of visual information has raised serious concerns [45, 20], especially in fields such as face [27, 28, 51] and medicine [5, 1, 12]. A representative case is the security issue of data stored in the cloud environment [19, 4, 35]. Due to potential vulnerabilities in cyberspace [3], uploaded private images can be easily stolen and used maliciously [38]. Therefore, it is meaningful and urgent to explore effective strategies to protect visual information. + +A classic strategy is visual information hiding, which + +![](images/eabd8be83dd989dc71e02e2ce7d341c337390bac77350c8c3467a079b24c31f5.jpg) +Figure 1. Illustration of visual information hiding for face recognition systems in cloud environments. The gallery set is protected and provided to the DNN in the cloud. Protected image has quite different visual information from the original image, but it is still correctly identified. The protected gallery set can be recovered by a key model by the owner. + +mainly consists of two types [40]: Homomorphic Encryption (HE)-based methods [2, 46, 32], Perceptual Encryption (PE)-based methods [42, 40, 39, 8, 16] and steganography method [24, 50]. Affected by the nonlinear activation functions in DNNs, HE-based methods are difficult to perform well on advanced DNNs [40]. Although PE-based methods apply to DNNs, existing methods typically require retraining with data in the encrypted domain to guarantee accuracy on encrypted data [42, 40]. This affects the performance of service model on raw data and cause additional resource consumption (especially for large models). The steganography method can hide the visual information of sensitive data into another image, but it does not guarantee that the service model can correctly recognize the protected image. + +To alleviate these negative effects, we expect to hide the visual information without making any modifications to service model. Namely, we hide sensitive visual information only by varying the input image. Previous researches have made some explorations in this regard. The transformation network-based methods [16], which try to protect the original image via a transformation function parameterized by + +a neural network, share the same philosophy. However, the method cannot easily recover the original image from the protected image for other purposes. They may suffer from adversarial vulnerability [9, 14] because the introduced neural network may be destroyed by adversarial attacks. + +In fact, the negative effects of adversarial attacks can be utilized positively to protect privacy. Some works have exploited adversarial attacks for de-identification [36, 34, 48]. These methods add imperceptible perturbations or nonsuspicious patches generated by adversarial attacks to original images, hindering DNNs to extract effective features and recognize identities, thus protecting the identity privacy in the image. However, in this work, we focus on visual information hiding, which means that the protected image is entirely different from the original image visually but can still be correctly predicted by DNNs (see Figure. 1). Fortunately, we observe a special type of adversarial attack (called Type-I attack) [43] quite different from the type used in previous methods. This type of attack guides DNNs to make consistent predictions on two distinct samples. + +Inspired by the Type-I adversarial attack, in this paper, we propose an Adversarial Visual Information Hiding (AVIH) method. The proposed method hides the visual information in images while preserving their functional (e.g., recognition and classification) features for service model. It can recover original images from protected images for their owners. Specifically, we reduce the visual correlation between the protected and original image while minimizing their distance in the feature space of service model, to generate the protected image. Meanwhile, we exploit a generative model pre-trained in a private training setting as the key model, then optimize the protected image based on it so that the recovered image is similar to the original image. Furthermore, to break through the tough trade-off between the capability of privacy protection and the quality of restored image, we design the variance consistency loss to enhance privacy protection without compromising image recovery (see Section 3.3). Note that the protected image generated by our method can only be accurately recovered by the own key model, other models (even if the model architecture is the same) are difficult to recover well (see Section 4.3). + +AVIH can significantly improve the security and flexibility of image storage, which is extremely obvious in the protection of gallery sets for metric learning. Take the cloud-based face recognition system as an example. According to the service face recognition model, the face database manager can generate protected images locally or in the cloud and save the key model locally. The protected image contains no visual information and can be used by the service model to extract features correctly. Moreover, these protected images cannot be recognized by other models. Then, these protected images can be stored in the gallery set in the cloud for normal face recognition tasks. These protected + +images can be recovered using the key model when needed (such as maintaining the dataset or using face images for other tasks, etc.). The process is shown in Figure. 1. + +Taking visual information hiding of gallery set images for cloud-deployed face recognition systems as the basic task, we conduct a comprehensive evaluation of our proposed method in terms of both effectiveness and security. In order to compare with existing methods suitable for information hiding tasks, we extend our method to classification tasks. Experimental results on multiple service models and datasets show that the proposed method is effective. In addition, to prove the effectiveness of our proposed loss, we conduct an ablation study about it to further present the advantages of our proposed method. + +Our main contributions are as follows: + +- Inspired by Type-I adversarial attacks, we propose a visual information hiding method AVIH. To alleviate the difficult trade-off between capability of information hiding and quality of recovered image, we design a variance consistency loss. +- Our proposed method has following properties: 1) The visual information in the image is clearly obfuscated. 2) Our method does not require retrain. 3) The protected image can be recovered by the own key model, but external models are difficult to recover. +- We validate the effectiveness of the proposed method on the face recognition task and the classification task. In addition, we conduct qualitative and quantitative ablation studies to show efficiency of the proposed loss. + +# 2. Related Work + +# 2.1. Visual Information Hiding + +The visual information hiding of images is from the perspective of human vision. It is most directly manifested by the protected images being visually unrecognizable. Since Homomorphic Encryption (HE) only supports additive and multiplicative operations, HE are not suitable for nonlinear computations. Most DNNs contain a large number of nonlinear computations, so HE-based methods are hardly applicable to state-of-the-art DNNs [40]. The steganography methods hide sensitive information in a cover image and require as few changes to the cover image as possible. It does not need to guarantee that the service model can perform normal predictions on hidden information, which is very different from our work. Therefore, these two types of methods are not discussed in this paper. For the methods based on Perceptual Encryption (PE), some works [42, 40, 39, 8] focus on finding an encrypted domain and training the model directly using the encrypted images. However, this has a significant impact on the accuracy of + +![](images/b71863da0bc209636e8331b01482351beab4d50c352386ee0f1ce385d08cd8b4.jpg) +Figure 2. Overview of adversarial attack-based visual information hiding (AVIH) method. Taking the face recognition as an example, given a service model and a pre-trained key model, we protect the original image $x$ and obtain the protected image $x'$ . + +the model. To improve the accuracy of the classification model for protected images, Ito et al. [16] trained a transformation network to keep the classifier correctly classified while hiding visual information. However, the weakness of this method is that the protected image cannot be recovered. + +Unlike the current work used in classification and segmentation tasks, we mainly focus on metric learning tasks represented by face recognition, which are more prone to privacy leakage. To compare with existing methods that can achieve visual information hiding, we also extend the AVIH method to classification tasks. Our proposed method can also compensate for the drawbacks of the above methods. It generates protected images for a specific model that already exists. It takes advantage of the vulnerability in the model itself to provide strong privacy protection to the image. Since our method does not modify the service model, it does not affect the accuracy on the original image. + +# 2.2. Adversarial Attack + +DNNs are vulnerable to some adversarial examples [9, 14]. There are many adversarial attack methods [9, 33, 6] to find adversarial examples efficiently. Tang et al. [43] divided the adversarial attacks into adversarial attack Type-I and adversarial attack Type-II based on the statistical Type I and Type II error. We take an optimization perspective on adversarial attacks. Then the Type-II attack maximizes the difference in the model output while ensuring a slight difference with the original input samples. Mathematically, it can be formulated as follows: + +$$ +\max _ {x ^ {\prime}} f (x) - f \left(x ^ {\prime}\right) \quad s. t. \quad \| x ^ {\prime} - x \| < \epsilon , \tag {1} +$$ + +where $x$ is the original sample, $x'$ is the adversarial sample, and $f(\cdot)$ is the model which is attacked. The Type-I attack, in contrast to the Type-II attack, looks for an input sample that differs the most from the original input sample + +but makes the model output as same as possible. Mathematically, it can be formulated as follows: + +$$ +\max _ {x ^ {\prime}} \| x ^ {\prime} - x \| \quad s. t. \quad f (x) = f \left(x ^ {\prime}\right). \tag {2} +$$ + +The Type-I attack on classification and generative models was implemented in the work of Tang et al. [43]. Then, Sun et al. [41] implemented a Type-I attack against variational autoencoder (VAE) [23]. In this work, we implement Type-I attacks on face recognition and classification tasks. Inspired by these attacks, we propose the AVIH method. + +# 3. Methodology + +The objective of our proposed method is to learn a protected image $x'$ which satisfies: 1) $x'$ is completely different from the original image $x$ , 2) for the service model the output $f_s(x')$ is the same as the $f_s(x)$ , and 3) $x'$ can be recovered as the $x$ by the key model. Mathematically, + +From $x\in \mathcal{X}$ Generate $x^{\prime} = \mathcal{A}(x)$ + +s.t. $\| x^{\prime} - x\| >\epsilon$ (3) + +$$ +f _ {s} \left(x ^ {\prime}\right) = f _ {s} (x), k e y \left(x ^ {\prime}\right) = x. +$$ + +The pipeline of our method is shown in Figure. 2. + +# 3.1. Image Visual Information Hiding + +Exploiting the adversarial vulnerability of DNNs, we can perform Type-I attack on the service model to get images that are visually completely different from the original image but with extremely similar features. Suppose that $d(\cdot)$ measures the difference between the outputs of the model. For different tasks, it behaves as different functions. Then for a particular service model $f_{s}$ , the difference between the output of the adversarial sample and the original sample is + +$$ +\mathcal {L} _ {t} \left(x ^ {\prime}, x\right) = d \left(f _ {s} (x), f _ {s} \left(x ^ {\prime}\right)\right). \tag {4} +$$ + +We definite $\mathcal{L}_d$ as + +$$ +\mathcal {L} _ {d} \left(x ^ {\prime}, x\right) = \left\| x - x ^ {\prime} \right\| _ {2} ^ {2}, \tag {5} +$$ + +then the loss of the Type-I attack on the service model is formulated as + +$$ +\mathcal {L} _ {r} \left(x ^ {\prime}, x\right) = \mathcal {L} _ {t} \left(x ^ {\prime}, x\right) - \lambda \cdot \mathcal {L} _ {d} \left(x ^ {\prime}, x\right), \tag {6} +$$ + +where $\lambda$ is a positive hyperparameter which balances image level differences and output level differences. Then with a certain number of iterations $K$ , we can optimize $\mathcal{L}_r$ by the following operations to get the final adversarial sample. + +$$ +g _ {k + 1} = \alpha \cdot g _ {k} + \frac {\nabla \mathcal {L} \left(x _ {k} ^ {\prime} , x\right)}{\| \nabla \mathcal {L} \left(x _ {k} ^ {\prime} , x\right) \| _ {2} ^ {2}}, \tag {7} +$$ + +$$ +x _ {k + 1} ^ {\prime} = x _ {k} ^ {\prime} - \beta \cdot g _ {k + 1}, \tag {8} +$$ + +where + +$$ +g _ {0} = \frac {\nabla \mathcal {L} \left(x _ {0} ^ {\prime} , x\right)}{\| \nabla \mathcal {L} \left(x _ {0} ^ {\prime} , x\right) \| _ {2} ^ {2}}. \tag {9} +$$ + +At the first iteration, $x_0'$ can be randomly initialized, or it can be made $x_0' = x$ . The former provides an easier way to get an adversarial example with more significant visual differences. The latter provides a faster way to get an adversarial example that meets the criteria. + +# 3.2. Image Recovery + +The protected image obtained by Equation. 6 can effectively hide visual information and keep the function of the sensitive information for the service model. However, the protected image is not recoverable. To achieve the goal of Equation. 3, we attach a recovery module that can recover the generated information hidden image. We refer to the generated images with perturbations as protected images. This perturbation has the function of protecting the visual information of the image. + +To get the protected images, we first train the generative model $G$ , which can generate the same images as the input. Then we perform Type-I attacks on both the service model and the generative model. We define $\mathcal{L}_g$ as follows: + +$$ +\mathcal {L} _ {g} \left(x ^ {\prime}, x\right) = \| x - G \left(x ^ {\prime}\right) \| _ {2} ^ {2}. \tag {10} +$$ + +The loss $\mathcal{L}_g$ can help to keep the protected images recoverable by the key model we have chosen. Thus, the loss becomes as follows: + +$$ +\mathcal {L} _ {e} \left(x ^ {\prime}, x\right) = \mathcal {L} _ {r} \left(x ^ {\prime}, x\right) + \mu \cdot \mathcal {L} _ {g} \left(x ^ {\prime}, x\right), \tag {11} +$$ + +where $\mu$ is a hyperparameter that balances the protection quality and restore quality. With Equation. 7 and Equation. 8, we can optimize $\mathcal{L}_e$ to obtain the protected image $x^{\prime}$ , which can satisfy the objective of Equation. 3. + +Algorithm 1 Adversarial visual information hiding method +Input: Service model $f_{s}$ ; key model $G$ ; original image $x$ ; number of iterations $K$ ; gradient $g$ ; number of times the loss has increased $num$ ; maximum number of times the loss increase $maxnum$ . + +Output: Protected images $x'$ . + +1: $g_0 = 0$ ; num = 0; +2: Random initialization $x_0'$ ; +3: for $k = 0$ to $K$ do +4: Compute $\mathcal{L}_{AVIH}(x_k', x)$ via Equation. 15; + +5: $g_{k + 1} = \alpha \cdot g_k + \frac{\nabla\mathcal{L}_{AVIH}(x_k',x)}{\left\|\nabla\mathcal{L}_{AVIH}(x_k',x)\right\|_2^2}$ (Equation. 7); + +6: $x_{k + 1}^{\prime} = x_{k}^{\prime} - \beta \cdot g_{k + 1}$ (Equation. 8); +7: if $\mathcal{L}_{AVIH}(x_{k - 1}^{\prime},x) < \mathcal{L}_{AVIH}(x_{k}^{\prime},x)$ then +8: $num = num + 1$ + +9: end if + +10: if num > maxnum then +11: $\beta = 0.85\cdot \beta$ +12: $num = 0$ +13: end if +14: end for + +# 3.3. Variance Consistency Loss + +The protected image obtained by the objective function of Equation. 6 satisfies the requirements of Equation. 3, but its protection quality is not high. The obtained protected images have the problem of the difficult trade-off between protection quality and recovery quality. That is, if we want to obtain an image that is difficult to be cracked successfully, the quality of the image recovered by the key model will be poor. Another problem is that the obtained protected image, although differing greatly from the original in color, has obvious visual information that the original image has, which negatively impacts visual information protection. To solve the above problem, we propose a variance consistency loss. It improves the quality of protection by limiting the differences between each part of the image to make the protected image visually more confusing. + +For the input image, we divide the image in each channel (R, G, B) into $N$ blocks respectively, i.e., $\{b_1, b_2, \ldots, b_N\}_c$ , where $c \in \{R, G, B\}$ , $b_n \in \mathbb{R}^{h \times w}$ and $h, w$ denote the height and width of the block. The pixels of the blocks are allowed to have overlapping parts between them. Let $p_{i,j}^{b_n} \in [0, 1]$ denote the normalized pixel value at $(i, j)$ in block $b_n$ . Then, we calculate the sum of each block: + +$$ +S _ {n} = \sum_ {i = 1} ^ {h} \sum_ {j = 1} ^ {w} p _ {i, j} ^ {b _ {n}}. \tag {12} +$$ + +We use $\mathcal{S}$ denote the set of sum of the blocks for each channel, i.e., $S_{c} = \{S_{1}, S_{2}, \ldots, S_{N}\}_{c}$ . In our practice, we convolve the image with a convolution kernel of size $h \times w$ to + +obtain blocks. Then, we calculate the variance of $S_{c}$ and obtain $\sigma_c^2 = \mathrm{var}(S_c)$ . Finally, we define the variance consistency loss as: + +$$ +\mathcal {L} _ {v} \left(x ^ {\prime}\right) = \sigma_ {R} ^ {2} + \sigma_ {G} ^ {2} + \sigma_ {B} ^ {2}, \tag {13} +$$ + +where $\sigma_{R}$ , $\sigma_{G}$ and $\sigma_{B}$ denote the variances for R, G, B channels, respectively. By minimizing $\mathcal{L}_v$ , we can get a protected image with a more uniform distribution of pixel values. Proof-of-concept experiments (see Section 4.4) show that $\mathcal{L}_v$ can eliminate the visual semantics in the protected images which are similar to the original images, and can help obtain protected images with high quality of protection and recovery. More details on the variance consistency loss can be found in Supplementary Material D. + +Based on the variance consistency loss, the loss function in Equation. 6 is modified as: + +$$ +\mathcal {L} _ {r} \left(x ^ {\prime}, x\right) = \mathcal {L} _ {t} \left(x ^ {\prime}, x\right) + \lambda \cdot \mathcal {L} _ {v} \left(x ^ {\prime}\right), \tag {14} +$$ + +and the Equation. 11 is reformulated as: + +$$ +\mathcal {L} _ {A V I H} \left(x ^ {\prime}, x\right) = \mathcal {L} _ {t} \left(x ^ {\prime}, x\right) + \lambda \cdot \mathcal {L} _ {v} \left(x ^ {\prime}\right) + \mu \cdot \mathcal {L} _ {g} \left(x ^ {\prime}, x\right), \tag {15} +$$ + +where $\mu$ are hyperparameters. This is our proposed AVIH framework for visual information hiding. The algorithm is summarized in Algorithm 1. + +# 3.4. AVIH Method for Specific Tasks + +Our method can provide effective protection for the stored image and can be applied to a wide range of tasks. In this work, we take face recognition and classification as examples to illustrate the ability of our method. + +For the face recognition task, we want the features extracted by the service model $f_{s}$ from the protected and original images to be as same as possible. Thus, we modify $\mathcal{L}_t$ as follows: + +$$ +\mathcal {L} _ {t} \left(x ^ {\prime}, x\right) = \left\| f _ {s} (x) - f _ {s} \left(x ^ {\prime}\right) \right\| _ {2} ^ {2}. \tag {16} +$$ + +For the classification task, compared with directly minimizing the mean square error (MSE) of $f_{s}(x)$ and $f_{s}(x^{\prime})$ , we find that maximizing $f_{s}^{c}(x^{\prime})$ where $c$ is the prediction class of the service model on the original image can more effectively reduce the impact of visual information hiding on the accuracy. We thus first obtain the logit output of the original image $f_{s}(x)$ and convert it to one-hot format $\delta(f_{s}(x))$ . Then, we reify $\mathcal{L}_t$ as follows: + +$$ +\mathcal {L} _ {\mathrm {t}} \left(x ^ {\prime}, x\right) = - \delta \left(f _ {s} (x)\right) \cdot f _ {s} \left(x ^ {\prime}\right). \tag {17} +$$ + +# 4. Experiments + +In this section, we first verify the effectiveness of the AVIH method for face recognition tasks. Then, taking the face recognition system as an example, the security of the + +AVIH method and the effectiveness of the variance consistency loss are explored. Finally, we verify the effectiveness of the AVIH method for classification tasks and compare it with other methods. + +# 4.1. Experimental Settings + +Dataset and service models. Our experiments were mainly evaluated based on the Labeled Face in the Wild (LFW) [13] dataset. We choose four face recognition models ArcFace [7], CosFace [44], SphereFace [30] used in [48], and AdaFace [22] to fully evaluate the performance of our method. Among them, the input size of ArcFace and AdaFace [22] is $112 \times 112$ , and the input size of CosFace and SphereFace is $112 \times 96$ . Therefore, we first use the MTCNN [49] to align and crop the face images to the corresponding face recognition model's input size. + +Key models. We choose the pix2pix framework [15] to train our key model. To pre-train the key model, we randomly selected 1,2878 images from the CelebA [31] dataset as the training set. Then we set the input and output of the model to the same image, set the batch size to 1, and train for 4 epochs. Finally, the trained generator is used directly as the key model. Unless otherwise stated, we use only one key model to protect the entire dataset. + +Evaluation metrics. To evaluate the effectiveness of our method more realistically and inspired by the evaluation method in MegaFace [21], we modified the evaluation method of LFW. We randomly selected 12 persons from the LFW dataset as the probe set. Each contains more than 12 facial images, comprising 355 images. The other 12878 images, we use as the gallery set. In the testing phase, we take one face image of a person in the probe set and put it into the gallery set. Then use the remaining images of this person as the test set. Next, we use the above-divided dataset to test the accuracy of the face recognition model. In this way, we put each person's image in the probe set into the gallery set in turn to measure the average accuracy. This metric can well demonstrate the impact of our method on face recognition models in practical applications. + +We use the Structural Similarity Index Measure (SSIM) [47] and Learned Perceptual Image Patch Similarity (LPIPS) to quantitatively measure the quality of the protected images. In this work, we use $\mathrm{SSIM}_e / \mathrm{LPIPS}_e$ to represent the average of the SSIM/LPIPS values between the protected image $x^{\prime}$ and the original image $x$ , and use $\mathrm{SSIM}_d / \mathrm{LPIPS}_d$ to represent the average of the SSIM/LPIPS values between $x$ and the recovered image $G(x^{\prime})$ . A higher SSIM indicates that the two images are more similar and a lower LPIPS indicates that the two images are more similar. + +# 4.2. Effectiveness of Face Privacy Protection + +Effectiveness and impact on service model. To the best of our knowledge, our method is the first (PE)-based + +![](images/634eb4f4eb059f7510b7be3d5e15ae05cca855349c36f72bb1fbf41ed64c5617.jpg) +Figure 3. Adversarial Visual Information Hiding (AVIH) method protect face images for different service models. We mark below each protected image the cosine similarity between its feature vector and the original image's feature vector. + +Table 1. Image visual quality metrics and cosine similarity (cosim) between the output features of the service model for original and protected images on LFW. + +
SSIMe↓LPIPSe↑SSIMd↑LPIPSd↓cos-sim ↑
AdaFace0.0280.8860.8930.1440.997
ArcFace0.0300.8880.9030.1280.994
CosFace0.0280.8910.8990.1280.998
SphereFace0.0290.8880.9060.1200.996
+ +method for face recognition. There is no closely related method yet to compare the performance of protected images. Therefore, we conduct qualitative and quantitative evaluations of protected images. High visual metric scores and a very slight impact on accuracy demonstrate the effectiveness of our method. + +We use the AVIH method to protect the original face image $x$ in the probe set to test its effectiveness on face visual information hiding. Then, we used the obtained protected image $x'$ as the input of the key model $G(\cdot)$ . Finally get the recovered image $G(x')$ . The results are shown in Figure. 3. From Figure. 3, it is evident that the protected image generated by our method is significantly different from the original image. Since no useful visual information is obtained from the protected image, it has the effect of visual information hiding for the original image. In Table. 1, the average SSIM value between the protected and original images for each service model is less than 0.04, while the cosine similarity between their features is higher than 0.99. Therefore, the protected image generated by AVIH can completely replace the original image while hiding the visual information. + +Table 2. Accuracy (percentage) of face recognition models for original image and different protected image on $LFW$ . The results of AVIH-ONE and AVIH-ALL are expected to be close to the results of original. + +
ModelsOriginalAVIH-ONE (%)AVIH-ALL (%)
AdaFace98.698.698.6
ArcFace96.596.596.5
CosFace89.489.289.3
SphereFace80.380.080.6
+ +To evaluate the impact of AVIH method on the accuracy of the face recognition models, we first generate a protected face image in the probe set before putting it into the gallery set. Then put the protected image into the gallery set instead of the original image. We tested the accuracy of the face recognition models in this way. The result AVIH-ONE is shown in Table. 2. Compared with the original accuracy of the models, it can be concluded that the AVIH method's impact on the models' accuracy is very slight, which can achieve the same accuracy in the ArcFace model. + +Considering the actual situation, all the images stored in the gallery set are protected images. So we replaced all the images in the gallery set with protected images and then retested the accuracy of the models, as AVIH-ALL in Table. 2. In this case, our method has a very slight impact on the accuracy of the models and has a beneficial effect on the accuracy of the SphereFace model. + +We test the average of SSIM values, as shown in Table. 1. Combined with Figure. 3 shows that the recovered image has good quality and can recover most of the visual information of the original image. Since the recovered images + +Table 3. The performances of our method when the pre-trained model is based on non-face image (e.g., COCO). The numbers in parentheses represent the difference from the original accuracy of the model. For example, (-0.2) means the accuracy is 0.2 lower than the original. + +
ModelAVIH-ONE (%)SSIMe ↓SSIMd ↑
AdaFace98.6(-0.0)0.0260.891
ArcFace96.5(-0.0)0.0270.909
CosFace89.2(-0.2)0.0310.882
SphereFace79.9(-0.4)0.0310.890
+ +![](images/2c10fce1e5365f23a1b107ae16061179db1f80d3a2ee548b70de7d16a529f841.jpg) +Figure 4. Visual information hiding in real-world scenes. The cosine similarity between recognition features of protected images $(x^{\prime})$ and original images $(x)$ is marked below $x^{\prime}$ . The SSIM between recovered images $(G(x^{\prime}))$ and original images is marked below $G(x^{\prime})$ . + +are generated by the generative model, a better-trained generative model can achieve better recovery quality. We also explore the time spent in protecting images of AVIH in Supplementary Materials A. + +Randomness of the dataset for training the key model. In this work, we use CelebA, which is also a face dataset, to train the key model. However, in real scenarios, large-scale face images are often difficult to obtain due to the privacy of each individual involved. So we changed CelebA to COCO [29] containing various objects to test the effectiveness of the new key model. We choose the test set of COCO as the train set and train 6 epochs on the key model. The results are shown in Table.3. It can be concluded that the impact of protected images on the model accuracy and the quality of recovered images are slightly different from the key model trained with CelebA. Therefore, we can use different types of datasets to train the key model, not just limited to face images. + +Real world face privacy protection. We randomly selected real-world face photos taken with phones and then protected them to test the performance of the AVIH method for using real-world scenarios. Part of the results is shown in Figure. 4. In realistic scenarios, the cosine similarity of the features between the protected image generated by the AVIH method and the original image is still higher than 0.99. The recovered image can still be recovered well. + +Robustness to noise. We tested the quality of the recovery after adding Gaussian noise $X_{noise} \sim \mathrm{N}(0, \sigma^2)$ to the + +Table 4. Robustness of protected images against noise. + +
σ03/2555/2558/25510/255
SSIMd0.930.920.890.840.81
+ +![](images/64f761b6b3634878d1c263ee78d425222d75ba6cb786b176dc8af7fda90eca3d.jpg) +Figure 5. Partial results of the key model randomness analysis. The left column is the key used for protection, and the top row is the key used for recovery. Every key model is trained with different initialization only. More images of the results are shown in Supplementary Material B. + +![](images/1f82b43cbfa6f0e2b5b23e41f5242fefa9c864736e6ea431aace1fffc2b075be.jpg) +Figure 6. Crack the protected image using the attack model. The training dataset for the attack model is obtained using key1 protection. An attacker can recover images protected with his own key (key1), but cannot recover images protected with other keys. + +protected image, as shown in Table 4. It can be concluded that our method is robust to small noise. + +# 4.3. Security Analysis + +Key model randomness analysis. To test the randomness of the key model, we use the same strategy as Ding et al. [8]. We first trained 16 key models using the same settings but with different initialization values. Different initialization values may result in significantly different key models. Then, we choose one of them in turn as the private key model to participate in the protection of the image, and use the others as the external key model to try to recover the protected image. From the results shown in Figure. 5, it can be concluded that the protected image obtained by one private key model cannot be recovered by other external key models, demonstrating that our proposed method can greatly improve the security of the face visual information. In addition, it is convenient that we can obtain mutually independent keys by simply changing the initialization without changing factors such as the structure of the key model and the training dataset. We suppose that one reason for this phenomenon is the instability of GAN training, where dif + +![](images/fc85188b6de24d89443085d4ac6aa3f9e9a7a9aeafe6acc496b9e4a3939882c6.jpg) +Figure 7. Results of the ablation study. The SSIM value between the processed image and the original image is shown under each image. The own key represents the key model used for generating the protected image, and the external key represents other key models that are initialized differently during training. + +ferent initialization values lead to different locally optimal solutions. Another reason could be that our method tends to find the boundary points of the input field corresponding to the output error allowed by the key model. However, the instability of GAN training causes this boundary to change significantly when the initialization value is changed. + +Responding to a possible attack. For the key model, even if the key model structure and its training settings are leaked, the protection is still difficult to break if the initialization point of the key training model is not leaked. As for the image, if pairs of original and protected images are leaked in large quantities, then an attacker can directly use these image to train a generative model to directly map the protected domain to the original domain. Ito et al. [17] exploited this method to evaluate the robustness of protection. So, we protect all the face images in the gallery set using the same key model, producing pairs of protected and original images. Then we use these paired images as a training set, and train a model using the same structure as the attack model. Using such a model, we attempt to recover the protected images in the probe set. The results are shown in Figure. 6. When the training set images and the protected image to be cracked are protected from the same key model, the attack model can recover the protected image. However, when they are protected with different key models, the trained model cannot crack the protected image. Therefore, we can increase the frequency of changing key models to effectively defend against such attacks. + +The protected images obtained by our method also have good statistical properties. The correlation analysis and histogram analysis are shown in Supplementary Material B. + +We also verify that the protected image generated by a service face recognition model cannot be recognized by other models with the same function. + +# 4.4. Ablation Study + +Effectiveness of the variance consistency loss. We compare the variance consistency loss $\mathcal{L}_v$ with the distance loss $\mathcal{L}_d$ (in Equation 5) to verify the effectiveness of $\mathcal{L}_v$ . For each loss, we choose a set of suitable weights $\lambda$ separately to represent its trade-off between the protection quality and recovery quality. Its results are shown in Figure 7. The protected image generated using $\mathcal{L}_d$ has a smaller SSIM value compared to the protected image using $\mathcal{L}_v$ , but a clear outline can be seen. So the visual information of the original image still exists. If the protected image is to be made free of such visual information, there is a significant loss in the quality of the recovered image. In contrast, using $\mathcal{L}_v$ , it is possible to generate protected images with both no residual visual information and high recovery quality. We also use another key to recover the protected image to further investigate the effect of the two different losses on the protection quality. From Figure 7, it can be concluded that the protected image using $\mathcal{L}_v$ is more difficult to be recovered by similar key models with different initialization while ensuring the quality of the recovered image. We compare $\mathcal{L}_v$ with more other losses in Supplementary Material D to further demonstrate the effectiveness of $\mathcal{L}_v$ . + +As shown in Figure. 5, the difference between the results when $\lambda = 0.001$ and $\lambda = 0.01$ is very slight for the protected and recovery quality of the protected images. This means that our method requires no carefully adjustment of + +![](images/26afe4253765ac1ba40a06862e3b501cccb7e3a3c08d1debebc91c0b5ab8e6d7.jpg) +Figure 8. Impact of parameter settings on protection quality and recovery quality. $k$ represents the size of the convolution kernel and $s$ represents the step size. For each pair of images, the former is the protected image and the latter is the recovery image. Below the recovery image is marked the SSIM value between it and the original image. + +![](images/7ed96582425f1ae0113e64aa90651a34a0c6682e5995bb7c675928633d86ebfd.jpg) +Figure 9. Results of AVIH method for Resnet50 as the service model. Each column represents a different sample. + +the weight $\lambda$ for $\mathcal{L}_v$ , which is very convenient. + +Parameters of variance consistency loss. In the variance consistency loss calculation, the process of blocking the image and calculating the internal sum of each block can be regarded as a convolution kernel with all values of 1 to convolve the image. The convolution kernel size represents the size of the block, and the step size of the convolution kernel represents the degree of overlap of the block. We explored the effect of block size and degree of overlap on performance, as shown in Figure 8. When the overlap degree of blocks is certain (especially when it is highly overlapping), the quality of the recovery image decreases as the size of the blocks increases. When the size of blocks is certain, the protection quality decreases as the overlap degree of blocks decreases. In this work, we divide the image into $4 \times 4$ blocks, with 12 pixels overlapping between adjacent blocks (step size is 1). This setting is applicable to most tasks and there is no need to change it for different samples. + +# 4.5. Privacy protection for classification models + +We trained ResNet50 [11], VGG19 [37] using the train set of CIFAR-10 [25] and implemented AVIH method for the images of the test set. The results are shown in Figure. 9. We also compared two existing methods suitable for visual information hiding and the results are shown in Table. 5, where the ITP [16] do not recover protected images. Since the LIE [42] uses protected images to train the model, the + +Table 5. Impact of privacy protection methods on the accuracy (percentage) of classification models on CIFAR-10. + +
MethodModelOriginal (%)Protect (%)
LIE [42]VGG1910.5987.78
Resnet5011.0091.53
ITP [16]VGG1993.9590.70
Resnet5095.5390.16
AVIH (Our)VGG1993.9593.95
Resnet5095.5395.28
+ +prediction accuracy on the original data is low. We show the visual metrics, the results of the two comparative methods, and the results on Imgnet in Supplementary Material C. Our method has minimal impact on the accuracy of the models, and the average SSIM value of the recovered images can reach above 0.9 for both models after our tests. + +# 5. Conclusion + +In this paper, we propose a visual information hiding method AVIH based on Type-I attack. We evaluate our method by image protection of face recognition systems in the cloud. Experiments show that the AVIH method can protect images while preserving their functionality for the service model. We also propose a variance consistency loss to solve the problem of the difficult trade-off between protection quality and recovery quality. Finally, we use the AVIH method in classification tasks with satisfactory results. Our work provides a new perspective on image visual information protection, which has beneficial implications for adversarial learning communities and private protection communities. The AVIH method requires complete service model information. While this is feasible in most image storage situations, it limits the applicability of the AVIH method to a wider range of image protection scenarios. This is also a problem we will solve in future work. + +Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 62036007; in part by the Fundamental Research Funds for the Central Universities under Grants QTZX23042 and XJS221502; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01; in part by the Natural Science Basis Research Plan in Shaanxi Province of China under Grant 2022JQ-696; in part by the Fundamental Research Funds for the Central Universities and in part by the Innovation Fund of Xidian University under Grant YJSJ23012. + +# References + +[1] Md Zahangir Alom, Chris Yakopcic, Mahmudul Hasan, Tarek M Taha, and Vijayan K Asari. Recurrent residual u-net for medical image segmentation. Journal of Medical Imaging, 6(1):014006-014006, 2019. 1 +[2] Yoshinori Aono, Takuya Hayashi, Lihua Wang, Shiho Moriai, et al. Privacy-preserving deep learning via additively homomorphic encryption. IEEE Transactions on Information Forensics and Security, 13(5):1333-1345, 2017. 1 +[3] L Arockiam and S Monikandan. Efficient cloud storage confidentiality to ensure data security. In 2014 International Conference on Computer Communication and Informatics, pages 1-5. IEEE, 2014. 1 +[4] Deyan Chen and Hong Zhao. Data security and privacy protection issues in cloud computing. In 2012 International Conference on Computer Science and Electronics Engineering, volume 1, pages 647-651. IEEE, 2012. 1 +[5] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michtitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE Transactions on Medical Imaging, 37(11):2453-2462, 2018. 1 +[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 3 +[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5 +[8] Yi Ding, Guozheng Wu, Dajiang Chen, Ning Zhang, Linpeng Gong, Mingsheng Cao, and Zhiguang Qin. Deepedn: a deep-learning-based image encryption and decryption network for internet of medical things. IEEE Internet of Things Journal, 8(3):1504-1518, 2020. 1, 2, 7 +[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2, 3 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9 +[12] Lei Hu, Da Wei Zhou, Cai Xia Fu, Thomas Benkert, Chun Yu Jiang, Rui Ting Li, Li Ming Wei, and Jun Gong Zhao. Advanced zoomed diffusion-weighted imaging vs. full-field-of-view diffusion-weighted imaging in prostate cancer detection: a radiomic features study. European Radiology, 31:1760-1769, 2021. 1 +[13] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5 + +[14] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2, 3 +[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5 +[16] Hiroki Ito, Yuma Kinoshita, Maungmaung Aprilpyone, and Hitoshi Kiya. Image to perturbation: An image transformation network for generating visually protected images for privacy-preserving deep neural networks. IEEE Access, 9:64629-64638, 2021. 1, 3, 9 +[17] Hiroki Ito, Yuma Kinoshita, and Hitoshi Kiya. Image transformation network for privacy-preserving deep neural networks and its security evaluation. In 2020 IEEE 9th Global Conference on Consumer Electronics (GCCE), pages 822-825. IEEE, 2020. 8 +[18] He Kaiming, Gkioxari Georgia, Dollar Piotr, and Girshick Ross. Mask r-cnn. IEEE Transactions on Pattern Analysis & Machine Intelligence, PP:1-1, 2017. 1 +[19] Seny Kamara and Kristin Lauter. Cryptographic cloud storage. In International Conference on Financial Cryptography and Data Security, pages 136-149. Springer, 2010. 1 +[20] Manjit Kaur and Vijay Kumar. A comprehensive review on image encryption techniques. Archives of Computational Methods in Engineering, 27(1):15-43, 2020. 1 +[21] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 5 +[22] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18750-18759, 2022. 5 +[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[24] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 1 +[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 9 +[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1 +[27] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Towards semi-supervised deep facial expression recognition with an adaptive confidence margin. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4166-4175, 2022. 1 +[28] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Unconstrained facial expression recognition with no-reference de-elements learning. IEEE Transactions on Affective Computing, 2023. 1 + +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7 +[30] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 5 +[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5 +[32] Qian Lou and Lei Jiang. She: A fast and accurate deep neural network for encrypted data. Advances in Neural Information Processing Systems, 32, 2019. 1 +[33] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 3 +[34] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 2 +[35] Siani Pearson and Azzedine Benameur. Privacy, security and trust issues arising from cloud computing. In 2010 IEEE Second International Conference on Cloud Computing Technology and Science, pages 693-702. IEEE, 2010. 1 +[36] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 acm sigsac conference on computer and communications security, pages 1528-1540, 2016. 2 +[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 9 +[38] Ashish Singh and Kakali Chatterjee. Cloud security issues and challenges: A survey. Journal of Network and Computer Applications, 79:88-115, 2017. 1 +[39] Warit Sirichotedumrong and Hitoshi Kiya. A gan-based image transformation scheme for privacy-preserving deep neural networks. In 2020 28th European Signal Processing Conference (EUSIPCO), pages 745-749. IEEE, 2021. 1, 2 +[40] Warit Sirichotedumrong, Takahiro Maekawa, Yuma Kinoshita, and Hitoshi Kiya. Privacy-preserving deep neural networks with pixel-based image encryption considering data augmentation in the encrypted domain. In 2019 IEEE International Conference on Image Processing (ICIP), pages 674-678. IEEE, 2019. 1, 2 +[41] Chengjin Sun, Sizhe Chen, Jia Cai, and Xiaolin Huang. Type i attack for generative models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 593-597. IEEE, 2020. 3 +[42] Masayuki Tanaka. Learnable image encryption. In 2018 IEEE International Conference on Consumer Electronics-Taiwan (ICCE-TW), pages 1-2, 2018. 1, 2, 9 + +[43] Sanli Tang, Xiaolin Huang, Mingjian Chen, Chengjin Sun, and Jie Yang. Adversarial attack type i: Cheat classifiers by significant changes. IEEE transactions on pattern analysis and machine intelligence, 43(3):1100-1109, 2019. 2, 3 +[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 5 +[45] Xueyu Wang, Jiajun Huang, Siqi Ma, Surya Nepal, and Chang Xu. Deepfake disrupter: The detector of deepfake is my friend. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14920-14929, 2022. 1 +[46] Yizhi Wang, Jun Lin, and Zhongfeng Wang. An efficient convolution core architecture for privacy-preserving deep learning. In 2018 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5. IEEE, 2018. 1 +[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[48] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 5 +[49] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 5 +[50] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 1 +[51] Dawei Zhou, Nannan Wang, Chunlei Peng, Yi Yu, Xi Yang, and Xinbo Gao. Towards multi-domain face synthesis via domain-invariant representations and multi-level feature parts. IEEE Transactions on Multimedia, 24:3469-3479, 2021. 1 \ No newline at end of file diff --git a/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/images.zip b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bebbd59e227cc15955e58a7b8dfc5ad249d0575d --- /dev/null +++ b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd77c00c9524a29d0a67e3e859051a79e66ea41d9285a6a3f8e16b1ad43e8348 +size 650094 diff --git a/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/layout.json b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0002ffdf81c5cb8a275434881e269beda79c7dcd --- /dev/null +++ b/2023/Hiding Visual Information via Obfuscating Adversarial Perturbations/layout.json @@ -0,0 +1,10158 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "text", + "content": "Hiding Visual Information via Obfuscating Adversarial Perturbations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "spans": [ + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": "Zhigang Su" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": " Dawei Zhou" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": " Nannan Wang" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": " Decheng Liu" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": " Zhen Wang" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": " Xinbo Gao" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": "Xidian University, " + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": "Zhejiang Lab \n" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 141, + 470, + 228 + ], + "type": "text", + "content": "Chongqing University of Posts and Telecommunications \n{zgsu, dwzhou}.xidian@gmail.com, {nnwang, dchliu}@xidian.edu.cn wangzhen@zhejianglab.com, gaoxb@cqupt.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 280, + 290, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 509 + ], + "type": "text", + "content": "Growing leakage and misuse of visual information raise security and privacy concerns, which promotes the development of information protection. Existing adversarial perturbations-based methods mainly focus on the de-identification against deep learning models. However, the inherent visual information of the data has not been well protected. In this work, inspired by the Type-I adversarial attack, we propose an Adversarial Visual Information Hiding (AVIH) method to protect the visual privacy of data. Specifically, the method generates obfuscating adversarial perturbations to obscure the visual information of the data. Meanwhile, it maintains the hidden objectives to be correctly predicted by models. In addition, our method does not modify the parameters of the applied model, which makes it flexible for different scenarios. Experimental results on the recognition and classification tasks demonstrate that the proposed method can effectively hide visual information and hardly affect the performances of models. The code is available at https://github.com/suzhigangssz/AVIH." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 682 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have been widely applied in the computer vision [26, 10, 18]. However, the increasing leakage and misuse of visual information has raised serious concerns [45, 20], especially in fields such as face [27, 28, 51] and medicine [5, 1, 12]. A representative case is the security issue of data stored in the cloud environment [19, 4, 35]. Due to potential vulnerabilities in cyberspace [3], uploaded private images can be easily stolen and used maliciously [38]. Therefore, it is meaningful and urgent to explore effective strategies to protect visual information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 683, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 287, + 695 + ], + "type": "text", + "content": "A classic strategy is visual information hiding, which" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 254, + 543, + 358 + ], + "blocks": [ + { + "bbox": [ + 310, + 254, + 543, + 358 + ], + "lines": [ + { + "bbox": [ + 310, + 254, + 543, + 358 + ], + "spans": [ + { + "bbox": [ + 310, + 254, + 543, + 358 + ], + "type": "image", + "image_path": "eabd8be83dd989dc71e02e2ce7d341c337390bac77350c8c3467a079b24c31f5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 370, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 305, + 370, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 547, + 437 + ], + "type": "text", + "content": "Figure 1. Illustration of visual information hiding for face recognition systems in cloud environments. The gallery set is protected and provided to the DNN in the cloud. Protected image has quite different visual information from the original image, but it is still correctly identified. The protected gallery set can be recovered by a key model by the owner." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 628 + ], + "type": "text", + "content": "mainly consists of two types [40]: Homomorphic Encryption (HE)-based methods [2, 46, 32], Perceptual Encryption (PE)-based methods [42, 40, 39, 8, 16] and steganography method [24, 50]. Affected by the nonlinear activation functions in DNNs, HE-based methods are difficult to perform well on advanced DNNs [40]. Although PE-based methods apply to DNNs, existing methods typically require retraining with data in the encrypted domain to guarantee accuracy on encrypted data [42, 40]. This affects the performance of service model on raw data and cause additional resource consumption (especially for large models). The steganography method can hide the visual information of sensitive data into another image, but it does not guarantee that the service model can correctly recognize the protected image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "To alleviate these negative effects, we expect to hide the visual information without making any modifications to service model. Namely, we hide sensitive visual information only by varying the input image. Previous researches have made some explorations in this regard. The transformation network-based methods [16], which try to protect the original image via a transformation function parameterized by" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 211, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 211, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 211, + 714 + ], + "type": "text", + "content": "*Equal contributions. † Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4356" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "a neural network, share the same philosophy. However, the method cannot easily recover the original image from the protected image for other purposes. They may suffer from adversarial vulnerability [9, 14] because the introduced neural network may be destroyed by adversarial attacks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 314 + ], + "type": "text", + "content": "In fact, the negative effects of adversarial attacks can be utilized positively to protect privacy. Some works have exploited adversarial attacks for de-identification [36, 34, 48]. These methods add imperceptible perturbations or nonsuspicious patches generated by adversarial attacks to original images, hindering DNNs to extract effective features and recognize identities, thus protecting the identity privacy in the image. However, in this work, we focus on visual information hiding, which means that the protected image is entirely different from the original image visually but can still be correctly predicted by DNNs (see Figure. 1). Fortunately, we observe a special type of adversarial attack (called Type-I attack) [43] quite different from the type used in previous methods. This type of attack guides DNNs to make consistent predictions on two distinct samples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 567 + ], + "type": "text", + "content": "Inspired by the Type-I adversarial attack, in this paper, we propose an Adversarial Visual Information Hiding (AVIH) method. The proposed method hides the visual information in images while preserving their functional (e.g., recognition and classification) features for service model. It can recover original images from protected images for their owners. Specifically, we reduce the visual correlation between the protected and original image while minimizing their distance in the feature space of service model, to generate the protected image. Meanwhile, we exploit a generative model pre-trained in a private training setting as the key model, then optimize the protected image based on it so that the recovered image is similar to the original image. Furthermore, to break through the tough trade-off between the capability of privacy protection and the quality of restored image, we design the variance consistency loss to enhance privacy protection without compromising image recovery (see Section 3.3). Note that the protected image generated by our method can only be accurately recovered by the own key model, other models (even if the model architecture is the same) are difficult to recover well (see Section 4.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "AVIH can significantly improve the security and flexibility of image storage, which is extremely obvious in the protection of gallery sets for metric learning. Take the cloud-based face recognition system as an example. According to the service face recognition model, the face database manager can generate protected images locally or in the cloud and save the key model locally. The protected image contains no visual information and can be used by the service model to extract features correctly. Moreover, these protected images cannot be recognized by other models. Then, these protected images can be stored in the gallery set in the cloud for normal face recognition tasks. These protected" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "images can be recovered using the key model when needed (such as maintaining the dataset or using face images for other tasks, etc.). The process is shown in Figure. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 108, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 239 + ], + "type": "text", + "content": "Taking visual information hiding of gallery set images for cloud-deployed face recognition systems as the basic task, we conduct a comprehensive evaluation of our proposed method in terms of both effectiveness and security. In order to compare with existing methods suitable for information hiding tasks, we extend our method to classification tasks. Experimental results on multiple service models and datasets show that the proposed method is effective. In addition, to prove the effectiveness of our proposed loss, we conduct an ablation study about it to further present the advantages of our proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 239, + 473, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 473, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 473, + 251 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 259, + 545, + 442 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 259, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 259, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 259, + 545, + 319 + ], + "type": "text", + "content": "- Inspired by Type-I adversarial attacks, we propose a visual information hiding method AVIH. To alleviate the difficult trade-off between capability of information hiding and quality of recovered image, we design a variance consistency loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 327, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 545, + 386 + ], + "type": "text", + "content": "- Our proposed method has following properties: 1) The visual information in the image is clearly obfuscated. 2) Our method does not require retrain. 3) The protected image can be recovered by the own key model, but external models are difficult to recover." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 394, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 545, + 442 + ], + "type": "text", + "content": "- We validate the effectiveness of the proposed method on the face recognition task and the classification task. In addition, we conduct qualitative and quantitative ablation studies to show efficiency of the proposed loss." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 460, + 392, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 392, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 392, + 472 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 479, + 453, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 453, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 453, + 492 + ], + "type": "text", + "content": "2.1. Visual Information Hiding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "The visual information hiding of images is from the perspective of human vision. It is most directly manifested by the protected images being visually unrecognizable. Since Homomorphic Encryption (HE) only supports additive and multiplicative operations, HE are not suitable for nonlinear computations. Most DNNs contain a large number of nonlinear computations, so HE-based methods are hardly applicable to state-of-the-art DNNs [40]. The steganography methods hide sensitive information in a cover image and require as few changes to the cover image as possible. It does not need to guarantee that the service model can perform normal predictions on hidden information, which is very different from our work. Therefore, these two types of methods are not discussed in this paper. For the methods based on Perceptual Encryption (PE), some works [42, 40, 39, 8] focus on finding an encrypted domain and training the model directly using the encrypted images. However, this has a significant impact on the accuracy of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4357" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 545, + 222 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 545, + 222 + ], + "type": "image", + "image_path": "b71863da0bc209636e8331b01482351beab4d50c352386ee0f1ce385d08cd8b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 257 + ], + "type": "text", + "content": "Figure 2. Overview of adversarial attack-based visual information hiding (AVIH) method. Taking the face recognition as an example, given a service model and a pre-trained key model, we protect the original image " + }, + { + "bbox": [ + 46, + 232, + 547, + 257 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 232, + 547, + 257 + ], + "type": "text", + "content": " and obtain the protected image " + }, + { + "bbox": [ + 46, + 232, + 547, + 257 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 232, + 547, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 334 + ], + "type": "text", + "content": "the model. To improve the accuracy of the classification model for protected images, Ito et al. [16] trained a transformation network to keep the classifier correctly classified while hiding visual information. However, the weakness of this method is that the protected image cannot be recovered." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 290, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 290, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 290, + 479 + ], + "type": "text", + "content": "Unlike the current work used in classification and segmentation tasks, we mainly focus on metric learning tasks represented by face recognition, which are more prone to privacy leakage. To compare with existing methods that can achieve visual information hiding, we also extend the AVIH method to classification tasks. Our proposed method can also compensate for the drawbacks of the above methods. It generates protected images for a specific model that already exists. It takes advantage of the vulnerability in the model itself to provide strong privacy protection to the image. Since our method does not modify the service model, it does not affect the accuracy on the original image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 489, + 160, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 160, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 160, + 501 + ], + "type": "text", + "content": "2.2. Adversarial Attack" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 289, + 628 + ], + "type": "text", + "content": "DNNs are vulnerable to some adversarial examples [9, 14]. There are many adversarial attack methods [9, 33, 6] to find adversarial examples efficiently. Tang et al. [43] divided the adversarial attacks into adversarial attack Type-I and adversarial attack Type-II based on the statistical Type I and Type II error. We take an optimization perspective on adversarial attacks. Then the Type-II attack maximizes the difference in the model output while ensuring a slight difference with the original input samples. Mathematically, it can be formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 637, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 637, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 83, + 637, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\max _ {x ^ {\\prime}} f (x) - f \\left(x ^ {\\prime}\\right) \\quad s. t. \\quad \\| x ^ {\\prime} - x \\| < \\epsilon , \\tag {1}", + "image_path": "e5959432e7c40ec65fcc57b3a98680413afb137796d5ec7668f72c56c9bf48f5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " is the original sample, " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " is the adversarial sample, and " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": " is the model which is attacked. The Type-I attack, in contrast to the Type-II attack, looks for an input sample that differs the most from the original input sample" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 274, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 299 + ], + "type": "text", + "content": "but makes the model output as same as possible. Mathematically, it can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 349, + 307, + 545, + 325 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 307, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 349, + 307, + 545, + 325 + ], + "type": "interline_equation", + "content": "\\max _ {x ^ {\\prime}} \\| x ^ {\\prime} - x \\| \\quad s. t. \\quad f (x) = f \\left(x ^ {\\prime}\\right). \\tag {2}", + "image_path": "1c929f148726923efdd397f5dee76ca256601ec9853ced2594423d03b4387e5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 333, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 547, + 406 + ], + "type": "text", + "content": "The Type-I attack on classification and generative models was implemented in the work of Tang et al. [43]. Then, Sun et al. [41] implemented a Type-I attack against variational autoencoder (VAE) [23]. In this work, we implement Type-I attacks on face recognition and classification tasks. Inspired by these attacks, we propose the AVIH method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 415, + 389, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 389, + 430 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": "The objective of our proposed method is to learn a protected image " + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": " which satisfies: 1) " + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": " is completely different from the original image " + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": ", 2) for the service model the output " + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "inline_equation", + "content": "f_s(x')" + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": " is the same as the " + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "inline_equation", + "content": "f_s(x)" + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": ", and 3) " + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": " can be recovered as the " + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 436, + 547, + 497 + ], + "type": "text", + "content": " by the key model. Mathematically," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 505, + 509, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 505, + 509, + 518 + ], + "spans": [ + { + "bbox": [ + 341, + 505, + 509, + 518 + ], + "type": "text", + "content": "From " + }, + { + "bbox": [ + 341, + 505, + 509, + 518 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{X}" + }, + { + "bbox": [ + 341, + 505, + 509, + 518 + ], + "type": "text", + "content": " Generate " + }, + { + "bbox": [ + 341, + 505, + 509, + 518 + ], + "type": "inline_equation", + "content": "x^{\\prime} = \\mathcal{A}(x)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 520, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 520, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 342, + 520, + 545, + 533 + ], + "type": "text", + "content": "s.t. " + }, + { + "bbox": [ + 342, + 520, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\| x^{\\prime} - x\\| >\\epsilon" + }, + { + "bbox": [ + 342, + 520, + 545, + 533 + ], + "type": "text", + "content": " (3)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 368, + 535, + 493, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 535, + 493, + 548 + ], + "spans": [ + { + "bbox": [ + 368, + 535, + 493, + 548 + ], + "type": "interline_equation", + "content": "f _ {s} \\left(x ^ {\\prime}\\right) = f _ {s} (x), k e y \\left(x ^ {\\prime}\\right) = x.", + "image_path": "0fd1ef9e3011ecd17e9b6fc72251110f4a34ed2eff7e6551307a31c9939956e4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 557, + 518, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 518, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 518, + 570 + ], + "type": "text", + "content": "The pipeline of our method is shown in Figure. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 577, + 486, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 486, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 486, + 590 + ], + "type": "text", + "content": "3.1. Image Visual Information Hiding" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": "Exploiting the adversarial vulnerability of DNNs, we can perform Type-I attack on the service model to get images that are visually completely different from the original image but with extremely similar features. Suppose that " + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": " measures the difference between the outputs of the model. For different tasks, it behaves as different functions. Then for a particular service model " + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 304, + 595, + 545, + 691 + ], + "type": "text", + "content": ", the difference between the output of the adversarial sample and the original sample is" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 365, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) = d \\left(f _ {s} (x), f _ {s} \\left(x ^ {\\prime}\\right)\\right). \\tag {4}", + "image_path": "f265a161ec612bbfd401c362bafe07c750ae2a0fb63920ae041747a7b5db967b.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4358" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 121, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 121, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 121, + 83 + ], + "type": "text", + "content": "We definite " + }, + { + "bbox": [ + 47, + 72, + 121, + 83 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 47, + 72, + 121, + 83 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 91, + 287, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 91, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 118, + 91, + 287, + 107 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} \\left(x ^ {\\prime}, x\\right) = \\left\\| x - x ^ {\\prime} \\right\\| _ {2} ^ {2}, \\tag {5}", + "image_path": "4a5814487a2f3afe1e40da88df20d73b691fe736d92bbca28a72b56f1495f699.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 115, + 287, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 287, + 137 + ], + "type": "text", + "content": "then the loss of the Type-I attack on the service model is formulated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 148, + 287, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 148, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 88, + 148, + 287, + 162 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) - \\lambda \\cdot \\mathcal {L} _ {d} \\left(x ^ {\\prime}, x\\right), \\tag {6}", + "image_path": "6d140fd5ecb9e277576f1c76cf376f9ccd05175df65a438d137d353158d59d8f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "text", + "content": " is a positive hyperparameter which balances image level differences and output level differences. Then with a certain number of iterations " + }, + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "text", + "content": ", we can optimize " + }, + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 47, + 170, + 287, + 218 + ], + "type": "text", + "content": " by the following operations to get the final adversarial sample." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 225, + 287, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 225, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 99, + 225, + 287, + 255 + ], + "type": "interline_equation", + "content": "g _ {k + 1} = \\alpha \\cdot g _ {k} + \\frac {\\nabla \\mathcal {L} \\left(x _ {k} ^ {\\prime} , x\\right)}{\\| \\nabla \\mathcal {L} \\left(x _ {k} ^ {\\prime} , x\\right) \\| _ {2} ^ {2}}, \\tag {7}", + "image_path": "17faa525dd73dd894a9d547f0b849957d4d2e59a80bf407ef8fdd376bb05e899.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 265, + 287, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 265, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 119, + 265, + 287, + 279 + ], + "type": "interline_equation", + "content": "x _ {k + 1} ^ {\\prime} = x _ {k} ^ {\\prime} - \\beta \\cdot g _ {k + 1}, \\tag {8}", + "image_path": "81591a9d7f23daeeff7087208eb233f138038e0a25a3617837a900a3bd03158b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 284, + 75, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 75, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 75, + 294 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 292, + 287, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 292, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 122, + 292, + 287, + 320 + ], + "type": "interline_equation", + "content": "g _ {0} = \\frac {\\nabla \\mathcal {L} \\left(x _ {0} ^ {\\prime} , x\\right)}{\\| \\nabla \\mathcal {L} \\left(x _ {0} ^ {\\prime} , x\\right) \\| _ {2} ^ {2}}. \\tag {9}", + "image_path": "a9f29f30af992c0db7d8882adf2fe49f068611c391b2445943c21a288df29df2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 324, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 287, + 384 + ], + "type": "text", + "content": "At the first iteration, " + }, + { + "bbox": [ + 47, + 324, + 287, + 384 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 47, + 324, + 287, + 384 + ], + "type": "text", + "content": " can be randomly initialized, or it can be made " + }, + { + "bbox": [ + 47, + 324, + 287, + 384 + ], + "type": "inline_equation", + "content": "x_0' = x" + }, + { + "bbox": [ + 47, + 324, + 287, + 384 + ], + "type": "text", + "content": ". The former provides an easier way to get an adversarial example with more significant visual differences. The latter provides a faster way to get an adversarial example that meets the criteria." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 392, + 145, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 145, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 145, + 404 + ], + "type": "text", + "content": "3.2. Image Recovery" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 411, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 517 + ], + "type": "text", + "content": "The protected image obtained by Equation. 6 can effectively hide visual information and keep the function of the sensitive information for the service model. However, the protected image is not recoverable. To achieve the goal of Equation. 3, we attach a recovery module that can recover the generated information hidden image. We refer to the generated images with perturbations as protected images. This perturbation has the function of protecting the visual information of the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 518, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 287, + 567 + ], + "type": "text", + "content": "To get the protected images, we first train the generative model " + }, + { + "bbox": [ + 47, + 518, + 287, + 567 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 518, + 287, + 567 + ], + "type": "text", + "content": ", which can generate the same images as the input. Then we perform Type-I attacks on both the service model and the generative model. We define " + }, + { + "bbox": [ + 47, + 518, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 47, + 518, + 287, + 567 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 110, + 574, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 574, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 110, + 574, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right) = \\| x - G \\left(x ^ {\\prime}\\right) \\| _ {2} ^ {2}. \\tag {10}", + "image_path": "7e29ae3573a4d990a49560195b92d7c376708e886a56147304c7b833e9ecde74.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 598, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 633 + ], + "type": "text", + "content": "The loss " + }, + { + "bbox": [ + 47, + 598, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 47, + 598, + 287, + 633 + ], + "type": "text", + "content": " can help to keep the protected images recoverable by the key model we have chosen. Thus, the loss becomes as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 643, + 287, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 643, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 88, + 643, + 287, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) + \\mu \\cdot \\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right), \\tag {11}", + "image_path": "6511fb44480886ab84c7633c907e4bccf79514aed740ac9ef9b2e0e95a8a44b1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": " is a hyperparameter that balances the protection quality and restore quality. With Equation. 7 and Equation. 8, we can optimize " + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": " to obtain the protected image " + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": ", which can satisfy the objective of Equation. 3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "Algorithm 1 Adversarial visual information hiding method \nInput: Service model " + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "; key model " + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "; original image " + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "; number of iterations " + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "; gradient " + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "; number of times the loss has increased " + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "num" + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "; maximum number of times the loss increase " + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "maxnum" + }, + { + "bbox": [ + 306, + 72, + 545, + 132 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 133, + 427, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 427, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 427, + 145 + ], + "type": "text", + "content": "Output: Protected images " + }, + { + "bbox": [ + 306, + 133, + 427, + 145 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 306, + 133, + 427, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 148, + 504, + 194 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 312, + 148, + 400, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 148, + 400, + 159 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 400, + 159 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 312, + 148, + 400, + 159 + ], + "type": "inline_equation", + "content": "g_0 = 0" + }, + { + "bbox": [ + 312, + 148, + 400, + 159 + ], + "type": "text", + "content": "; num = 0;" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 160, + 427, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 160, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 160, + 427, + 171 + ], + "type": "text", + "content": "2: Random initialization " + }, + { + "bbox": [ + 312, + 160, + 427, + 171 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 312, + 160, + 427, + 171 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 172, + 399, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 172, + 399, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 172, + 399, + 182 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 312, + 172, + 399, + 182 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 312, + 172, + 399, + 182 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 172, + 399, + 182 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 312, + 172, + 399, + 182 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 183, + 504, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 183, + 504, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 183, + 504, + 194 + ], + "type": "text", + "content": "4: Compute " + }, + { + "bbox": [ + 312, + 183, + 504, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AVIH}(x_k', x)" + }, + { + "bbox": [ + 312, + 183, + 504, + 194 + ], + "type": "text", + "content": " via Equation. 15;" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 195, + 533, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 533, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 533, + 216 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 195, + 533, + 216 + ], + "type": "inline_equation", + "content": "g_{k + 1} = \\alpha \\cdot g_k + \\frac{\\nabla\\mathcal{L}_{AVIH}(x_k',x)}{\\left\\|\\nabla\\mathcal{L}_{AVIH}(x_k',x)\\right\\|_2^2}" + }, + { + "bbox": [ + 312, + 195, + 533, + 216 + ], + "type": "text", + "content": " (Equation. 7);" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 217, + 510, + 251 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 217, + 483, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 483, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 483, + 228 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 312, + 217, + 483, + 228 + ], + "type": "inline_equation", + "content": "x_{k + 1}^{\\prime} = x_{k}^{\\prime} - \\beta \\cdot g_{k + 1}" + }, + { + "bbox": [ + 312, + 217, + 483, + 228 + ], + "type": "text", + "content": " (Equation. 8);" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 228, + 510, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 510, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 510, + 239 + ], + "type": "text", + "content": "7: if " + }, + { + "bbox": [ + 312, + 228, + 510, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AVIH}(x_{k - 1}^{\\prime},x) < \\mathcal{L}_{AVIH}(x_{k}^{\\prime},x)" + }, + { + "bbox": [ + 312, + 228, + 510, + 239 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 240, + 419, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 240, + 419, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 240, + 419, + 251 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 312, + 240, + 419, + 251 + ], + "type": "inline_equation", + "content": "num = num + 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 253, + 360, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 253, + 360, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 253, + 360, + 262 + ], + "type": "text", + "content": "9: end if" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 309, + 264, + 439, + 321 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 264, + 439, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 439, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 439, + 274 + ], + "type": "text", + "content": "10: if num > maxnum then" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 309, + 276, + 399, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 276, + 399, + 286 + ], + "spans": [ + { + "bbox": [ + 309, + 276, + 399, + 286 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 309, + 276, + 399, + 286 + ], + "type": "inline_equation", + "content": "\\beta = 0.85\\cdot \\beta" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 310, + 289, + 387, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 289, + 387, + 298 + ], + "spans": [ + { + "bbox": [ + 310, + 289, + 387, + 298 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 310, + 289, + 387, + 298 + ], + "type": "inline_equation", + "content": "num = 0" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 310, + 300, + 361, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 300, + 361, + 309 + ], + "spans": [ + { + "bbox": [ + 310, + 300, + 361, + 309 + ], + "type": "text", + "content": "13: end if" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 310, + 312, + 356, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 312, + 356, + 321 + ], + "spans": [ + { + "bbox": [ + 310, + 312, + 356, + 321 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 345, + 452, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 452, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 452, + 357 + ], + "type": "text", + "content": "3.3. Variance Consistency Loss" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 364, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 542 + ], + "type": "text", + "content": "The protected image obtained by the objective function of Equation. 6 satisfies the requirements of Equation. 3, but its protection quality is not high. The obtained protected images have the problem of the difficult trade-off between protection quality and recovery quality. That is, if we want to obtain an image that is difficult to be cracked successfully, the quality of the image recovered by the key model will be poor. Another problem is that the obtained protected image, although differing greatly from the original in color, has obvious visual information that the original image has, which negatively impacts visual information protection. To solve the above problem, we propose a variance consistency loss. It improves the quality of protection by limiting the differences between each part of the image to make the protected image visually more confusing." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": "For the input image, we divide the image in each channel (R, G, B) into " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": " blocks respectively, i.e., " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\{b_1, b_2, \\ldots, b_N\\}_c" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "c \\in \\{R, G, B\\}" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "b_n \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": " denote the height and width of the block. The pixels of the blocks are allowed to have overlapping parts between them. Let " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "p_{i,j}^{b_n} \\in [0, 1]" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": " denote the normalized pixel value at " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "(i, j)" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": " in block " + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "inline_equation", + "content": "b_n" + }, + { + "bbox": [ + 304, + 543, + 545, + 625 + ], + "type": "text", + "content": ". Then, we calculate the sum of each block:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 387, + 635, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 387, + 635, + 545, + 668 + ], + "type": "interline_equation", + "content": "S _ {n} = \\sum_ {i = 1} ^ {h} \\sum_ {j = 1} ^ {w} p _ {i, j} ^ {b _ {n}}. \\tag {12}", + "image_path": "b4259d0a099ade497f0782527129af142339a600ed126b6bd0f8a0b500f486a5.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " denote the set of sum of the blocks for each channel, i.e., " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{c} = \\{S_{1}, S_{2}, \\ldots, S_{N}\\}_{c}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". In our practice, we convolve the image with a convolution kernel of size " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4359" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "obtain blocks. Then, we calculate the variance of " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\sigma_c^2 = \\mathrm{var}(S_c)" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": ". Finally, we define the variance consistency loss as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 112, + 116, + 287, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 116, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 112, + 116, + 287, + 131 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right) = \\sigma_ {R} ^ {2} + \\sigma_ {G} ^ {2} + \\sigma_ {B} ^ {2}, \\tag {13}", + "image_path": "595500dee8442ac07332ca41ad990262a4790448c4fe86e4b1af090d829bb93e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\sigma_{R}" + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\sigma_{G}" + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\sigma_{B}" + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "text", + "content": " denote the variances for R, G, B channels, respectively. By minimizing " + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "text", + "content": ", we can get a protected image with a more uniform distribution of pixel values. Proof-of-concept experiments (see Section 4.4) show that " + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 46, + 139, + 287, + 245 + ], + "type": "text", + "content": " can eliminate the visual semantics in the protected images which are similar to the original images, and can help obtain protected images with high quality of protection and recovery. More details on the variance consistency loss can be found in Supplementary Material D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 270 + ], + "type": "text", + "content": "Based on the variance consistency loss, the loss function in Equation. 6 is modified as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 278, + 287, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 278, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 93, + 278, + 287, + 293 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) + \\lambda \\cdot \\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right), \\tag {14}", + "image_path": "6025f760c22fd8b8d7636411903d0345b1b6df9b8d63c5701dc5dc0f266f1b21.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 301, + 211, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 211, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 211, + 312 + ], + "type": "text", + "content": "and the Equation. 11 is reformulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 320, + 287, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 320, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 52, + 320, + 287, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A V I H} \\left(x ^ {\\prime}, x\\right) = \\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) + \\lambda \\cdot \\mathcal {L} _ {v} \\left(x ^ {\\prime}\\right) + \\mu \\cdot \\mathcal {L} _ {g} \\left(x ^ {\\prime}, x\\right), \\tag {15}", + "image_path": "6edb7da4ec1328549864c1db9b0dfea9539ff22e6331642ddd6e73e171286b87.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 342, + 287, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 342, + 287, + 379 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 342, + 287, + 379 + ], + "type": "text", + "content": " are hyperparameters. This is our proposed AVIH framework for visual information hiding. The algorithm is summarized in Algorithm 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 385, + 221, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 221, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 221, + 399 + ], + "type": "text", + "content": "3.4. AVIH Method for Specific Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 404, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 452 + ], + "type": "text", + "content": "Our method can provide effective protection for the stored image and can be applied to a wide range of tasks. In this work, we take face recognition and classification as examples to illustrate the ability of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 453, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 499 + ], + "type": "text", + "content": "For the face recognition task, we want the features extracted by the service model " + }, + { + "bbox": [ + 46, + 453, + 287, + 499 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 453, + 287, + 499 + ], + "type": "text", + "content": " from the protected and original images to be as same as possible. Thus, we modify " + }, + { + "bbox": [ + 46, + 453, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 46, + 453, + 287, + 499 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 101, + 506, + 287, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 506, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 101, + 506, + 287, + 523 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} \\left(x ^ {\\prime}, x\\right) = \\left\\| f _ {s} (x) - f _ {s} \\left(x ^ {\\prime}\\right) \\right\\| _ {2} ^ {2}. \\tag {16}", + "image_path": "8fa266a6cb6ee910167c456a0f7a9ffb7a1c980ab7cbed43e66de811161057b0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": "For the classification task, compared with directly minimizing the mean square error (MSE) of " + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "inline_equation", + "content": "f_{s}(x)" + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "inline_equation", + "content": "f_{s}(x^{\\prime})" + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": ", we find that maximizing " + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "inline_equation", + "content": "f_{s}^{c}(x^{\\prime})" + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": " is the prediction class of the service model on the original image can more effectively reduce the impact of visual information hiding on the accuracy. We thus first obtain the logit output of the original image " + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "inline_equation", + "content": "f_{s}(x)" + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": " and convert it to one-hot format " + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "inline_equation", + "content": "\\delta(f_{s}(x))" + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": ". Then, we reify " + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 46, + 530, + 287, + 626 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 101, + 635, + 287, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 635, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 101, + 635, + 287, + 648 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t}} \\left(x ^ {\\prime}, x\\right) = - \\delta \\left(f _ {s} (x)\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right). \\tag {17}", + "image_path": "9edb00da8ae1bb2136074b626ad32d633d8021c0f7c8961d8ff7a88cd991e3bd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 128, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In this section, we first verify the effectiveness of the AVIH method for face recognition tasks. Then, taking the face recognition system as an example, the security of the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "AVIH method and the effectiveness of the variance consistency loss are explored. Finally, we verify the effectiveness of the AVIH method for classification tasks and compare it with other methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 129, + 432, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 432, + 141 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 432, + 141 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 147, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 266 + ], + "type": "text", + "content": "Dataset and service models. Our experiments were mainly evaluated based on the Labeled Face in the Wild (LFW) [13] dataset. We choose four face recognition models ArcFace [7], CosFace [44], SphereFace [30] used in [48], and AdaFace [22] to fully evaluate the performance of our method. Among them, the input size of ArcFace and AdaFace [22] is " + }, + { + "bbox": [ + 304, + 147, + 545, + 266 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 304, + 147, + 545, + 266 + ], + "type": "text", + "content": ", and the input size of CosFace and SphereFace is " + }, + { + "bbox": [ + 304, + 147, + 545, + 266 + ], + "type": "inline_equation", + "content": "112 \\times 96" + }, + { + "bbox": [ + 304, + 147, + 545, + 266 + ], + "type": "text", + "content": ". Therefore, we first use the MTCNN [49] to align and crop the face images to the corresponding face recognition model's input size." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 267, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 361 + ], + "type": "text", + "content": "Key models. We choose the pix2pix framework [15] to train our key model. To pre-train the key model, we randomly selected 1,2878 images from the CelebA [31] dataset as the training set. Then we set the input and output of the model to the same image, set the batch size to 1, and train for 4 epochs. Finally, the trained generator is used directly as the key model. Unless otherwise stated, we use only one key model to protect the entire dataset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 363, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 542 + ], + "type": "text", + "content": "Evaluation metrics. To evaluate the effectiveness of our method more realistically and inspired by the evaluation method in MegaFace [21], we modified the evaluation method of LFW. We randomly selected 12 persons from the LFW dataset as the probe set. Each contains more than 12 facial images, comprising 355 images. The other 12878 images, we use as the gallery set. In the testing phase, we take one face image of a person in the probe set and put it into the gallery set. Then use the remaining images of this person as the test set. Next, we use the above-divided dataset to test the accuracy of the face recognition model. In this way, we put each person's image in the probe set into the gallery set in turn to measure the average accuracy. This metric can well demonstrate the impact of our method on face recognition models in practical applications." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": "We use the Structural Similarity Index Measure (SSIM) [47] and Learned Perceptual Image Patch Similarity (LPIPS) to quantitatively measure the quality of the protected images. In this work, we use " + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{SSIM}_e / \\mathrm{LPIPS}_e" + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": " to represent the average of the SSIM/LPIPS values between the protected image " + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": " and the original image " + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": ", and use " + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{SSIM}_d / \\mathrm{LPIPS}_d" + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": " to represent the average of the SSIM/LPIPS values between " + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": " and the recovered image " + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "inline_equation", + "content": "G(x^{\\prime})" + }, + { + "bbox": [ + 304, + 543, + 545, + 662 + ], + "type": "text", + "content": ". A higher SSIM indicates that the two images are more similar and a lower LPIPS indicates that the two images are more similar." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 670, + 515, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 515, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 515, + 683 + ], + "type": "text", + "content": "4.2. Effectiveness of Face Privacy Protection" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness and impact on service model. To the best of our knowledge, our method is the first (PE)-based" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4360" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 73, + 541, + 254 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 541, + 254 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 541, + 254 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 541, + 254 + ], + "type": "image", + "image_path": "634eb4f4eb059f7510b7be3d5e15ae05cca855349c36f72bb1fbf41ed64c5617.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 269, + 546, + 292 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 546, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 546, + 292 + ], + "type": "text", + "content": "Figure 3. Adversarial Visual Information Hiding (AVIH) method protect face images for different service models. We mark below each protected image the cosine similarity between its feature vector and the original image's feature vector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 349, + 282, + 418 + ], + "blocks": [ + { + "bbox": [ + 46, + 308, + 288, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 342 + ], + "type": "text", + "content": "Table 1. Image visual quality metrics and cosine similarity (cosim) between the output features of the service model for original and protected images on LFW." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 349, + 282, + 418 + ], + "lines": [ + { + "bbox": [ + 52, + 349, + 282, + 418 + ], + "spans": [ + { + "bbox": [ + 52, + 349, + 282, + 418 + ], + "type": "table", + "html": "
SSIMe↓LPIPSe↑SSIMd↑LPIPSd↓cos-sim ↑
AdaFace0.0280.8860.8930.1440.997
ArcFace0.0300.8880.9030.1280.994
CosFace0.0280.8910.8990.1280.998
SphereFace0.0290.8880.9060.1200.996
", + "image_path": "b2e995de79acc286b79ff439d2c85b90c91d6991f93107775a0adf62dd04beec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 520 + ], + "type": "text", + "content": "method for face recognition. There is no closely related method yet to compare the performance of protected images. Therefore, we conduct qualitative and quantitative evaluations of protected images. High visual metric scores and a very slight impact on accuracy demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "We use the AVIH method to protect the original face image " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " in the probe set to test its effectiveness on face visual information hiding. Then, we used the obtained protected image " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " as the input of the key model " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ". Finally get the recovered image " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "G(x')" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ". The results are shown in Figure. 3. From Figure. 3, it is evident that the protected image generated by our method is significantly different from the original image. Since no useful visual information is obtained from the protected image, it has the effect of visual information hiding for the original image. In Table. 1, the average SSIM value between the protected and original images for each service model is less than 0.04, while the cosine similarity between their features is higher than 0.99. Therefore, the protected image generated by AVIH can completely replace the original image while hiding the visual information." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 360, + 545, + 430 + ], + "blocks": [ + { + "bbox": [ + 305, + 308, + 546, + 352 + ], + "lines": [ + { + "bbox": [ + 305, + 308, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 546, + 352 + ], + "type": "text", + "content": "Table 2. Accuracy (percentage) of face recognition models for original image and different protected image on " + }, + { + "bbox": [ + 305, + 308, + 546, + 352 + ], + "type": "inline_equation", + "content": "LFW" + }, + { + "bbox": [ + 305, + 308, + 546, + 352 + ], + "type": "text", + "content": ". The results of AVIH-ONE and AVIH-ALL are expected to be close to the results of original." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 360, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 430 + ], + "type": "table", + "html": "
ModelsOriginalAVIH-ONE (%)AVIH-ALL (%)
AdaFace98.698.698.6
ArcFace96.596.596.5
CosFace89.489.289.3
SphereFace80.380.080.6
", + "image_path": "a3d40375fff7f5230cfe3768e9395701613a2ab9a56ae8eb0d968c9b16807463.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 459, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 579 + ], + "type": "text", + "content": "To evaluate the impact of AVIH method on the accuracy of the face recognition models, we first generate a protected face image in the probe set before putting it into the gallery set. Then put the protected image into the gallery set instead of the original image. We tested the accuracy of the face recognition models in this way. The result AVIH-ONE is shown in Table. 2. Compared with the original accuracy of the models, it can be concluded that the AVIH method's impact on the models' accuracy is very slight, which can achieve the same accuracy in the ArcFace model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 580, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 547, + 664 + ], + "type": "text", + "content": "Considering the actual situation, all the images stored in the gallery set are protected images. So we replaced all the images in the gallery set with protected images and then retested the accuracy of the models, as AVIH-ALL in Table. 2. In this case, our method has a very slight impact on the accuracy of the models and has a beneficial effect on the accuracy of the SphereFace model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "We test the average of SSIM values, as shown in Table. 1. Combined with Figure. 3 shows that the recovered image has good quality and can recover most of the visual information of the original image. Since the recovered images" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4361" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 133, + 282, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 288, + 125 + ], + "type": "text", + "content": "Table 3. The performances of our method when the pre-trained model is based on non-face image (e.g., COCO). The numbers in parentheses represent the difference from the original accuracy of the model. For example, (-0.2) means the accuracy is 0.2 lower than the original." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 133, + 282, + 201 + ], + "lines": [ + { + "bbox": [ + 52, + 133, + 282, + 201 + ], + "spans": [ + { + "bbox": [ + 52, + 133, + 282, + 201 + ], + "type": "table", + "html": "
ModelAVIH-ONE (%)SSIMe ↓SSIMd ↑
AdaFace98.6(-0.0)0.0260.891
ArcFace96.5(-0.0)0.0270.909
CosFace89.2(-0.2)0.0310.882
SphereFace79.9(-0.4)0.0310.890
", + "image_path": "58b660d8f78393e257072cb99766085175badd85065d4437424335db2987cde6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 219, + 284, + 283 + ], + "blocks": [ + { + "bbox": [ + 51, + 219, + 284, + 283 + ], + "lines": [ + { + "bbox": [ + 51, + 219, + 284, + 283 + ], + "spans": [ + { + "bbox": [ + 51, + 219, + 284, + 283 + ], + "type": "image", + "image_path": "2c10fce1e5365f23a1b107ae16061179db1f80d3a2ee548b70de7d16a529f841.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "text", + "content": "Figure 4. Visual information hiding in real-world scenes. The cosine similarity between recognition features of protected images " + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "inline_equation", + "content": "(x^{\\prime})" + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "text", + "content": " and original images " + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "inline_equation", + "content": "(x)" + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "text", + "content": " is marked below " + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "text", + "content": ". The SSIM between recovered images " + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "inline_equation", + "content": "(G(x^{\\prime}))" + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "text", + "content": " and original images is marked below " + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "inline_equation", + "content": "G(x^{\\prime})" + }, + { + "bbox": [ + 46, + 293, + 287, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 422 + ], + "type": "text", + "content": "are generated by the generative model, a better-trained generative model can achieve better recovery quality. We also explore the time spent in protecting images of AVIH in Supplementary Materials A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 423, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 590 + ], + "type": "text", + "content": "Randomness of the dataset for training the key model. In this work, we use CelebA, which is also a face dataset, to train the key model. However, in real scenarios, large-scale face images are often difficult to obtain due to the privacy of each individual involved. So we changed CelebA to COCO [29] containing various objects to test the effectiveness of the new key model. We choose the test set of COCO as the train set and train 6 epochs on the key model. The results are shown in Table.3. It can be concluded that the impact of protected images on the model accuracy and the quality of recovered images are slightly different from the key model trained with CelebA. Therefore, we can use different types of datasets to train the key model, not just limited to face images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": "Real world face privacy protection. We randomly selected real-world face photos taken with phones and then protected them to test the performance of the AVIH method for using real-world scenarios. Part of the results is shown in Figure. 4. In realistic scenarios, the cosine similarity of the features between the protected image generated by the AVIH method and the original image is still higher than 0.99. The recovered image can still be recovered well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Robustness to noise. We tested the quality of the recovery after adding Gaussian noise " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "X_{noise} \\sim \\mathrm{N}(0, \\sigma^2)" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " to the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 87, + 539, + 121 + ], + "blocks": [ + { + "bbox": [ + 326, + 69, + 524, + 80 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 524, + 80 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 524, + 80 + ], + "type": "text", + "content": "Table 4. Robustness of protected images against noise." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 87, + 539, + 121 + ], + "lines": [ + { + "bbox": [ + 313, + 87, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 539, + 121 + ], + "type": "table", + "html": "
σ03/2555/2558/25510/255
SSIMd0.930.920.890.840.81
", + "image_path": "2f3b34b519e0bef1d392fe44e8204f47ab8c272eb0ad516239acc6b5c23fc703.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 137, + 539, + 240 + ], + "blocks": [ + { + "bbox": [ + 317, + 137, + 539, + 240 + ], + "lines": [ + { + "bbox": [ + 317, + 137, + 539, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 137, + 539, + 240 + ], + "type": "image", + "image_path": "64f761b6b3634878d1c263ee78d425222d75ba6cb786b176dc8af7fda90eca3d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 251, + 545, + 306 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 306 + ], + "type": "text", + "content": "Figure 5. Partial results of the key model randomness analysis. The left column is the key used for protection, and the top row is the key used for recovery. Every key model is trained with different initialization only. More images of the results are shown in Supplementary Material B." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 323, + 542, + 379 + ], + "blocks": [ + { + "bbox": [ + 316, + 323, + 542, + 379 + ], + "lines": [ + { + "bbox": [ + 316, + 323, + 542, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 542, + 379 + ], + "type": "image", + "image_path": "1f82b43cbfa6f0e2b5b23e41f5242fefa9c864736e6ea431aace1fffc2b075be.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 391, + 545, + 436 + ], + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 436 + ], + "type": "text", + "content": "Figure 6. Crack the protected image using the attack model. The training dataset for the attack model is obtained using key1 protection. An attacker can recover images protected with his own key (key1), but cannot recover images protected with other keys." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 448, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 545, + 471 + ], + "type": "text", + "content": "protected image, as shown in Table 4. It can be concluded that our method is robust to small noise." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 479, + 410, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 410, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 410, + 493 + ], + "type": "text", + "content": "4.3. Security Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Key model randomness analysis. To test the randomness of the key model, we use the same strategy as Ding et al. [8]. We first trained 16 key models using the same settings but with different initialization values. Different initialization values may result in significantly different key models. Then, we choose one of them in turn as the private key model to participate in the protection of the image, and use the others as the external key model to try to recover the protected image. From the results shown in Figure. 5, it can be concluded that the protected image obtained by one private key model cannot be recovered by other external key models, demonstrating that our proposed method can greatly improve the security of the face visual information. In addition, it is convenient that we can obtain mutually independent keys by simply changing the initialization without changing factors such as the structure of the key model and the training dataset. We suppose that one reason for this phenomenon is the instability of GAN training, where dif" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4362" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 541, + 273 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 273 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 273 + ], + "type": "image", + "image_path": "fc85188b6de24d89443085d4ac6aa3f9e9a7a9aeafe6acc496b9e4a3939882c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 282, + 546, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 546, + 316 + ], + "type": "text", + "content": "Figure 7. Results of the ablation study. The SSIM value between the processed image and the original image is shown under each image. The own key represents the key model used for generating the protected image, and the external key represents other key models that are initialized differently during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 407 + ], + "type": "text", + "content": "ferent initialization values lead to different locally optimal solutions. Another reason could be that our method tends to find the boundary points of the input field corresponding to the output error allowed by the key model. However, the instability of GAN training causes this boundary to change significantly when the initialization value is changed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 674 + ], + "type": "text", + "content": "Responding to a possible attack. For the key model, even if the key model structure and its training settings are leaked, the protection is still difficult to break if the initialization point of the key training model is not leaked. As for the image, if pairs of original and protected images are leaked in large quantities, then an attacker can directly use these image to train a generative model to directly map the protected domain to the original domain. Ito et al. [17] exploited this method to evaluate the robustness of protection. So, we protect all the face images in the gallery set using the same key model, producing pairs of protected and original images. Then we use these paired images as a training set, and train a model using the same structure as the attack model. Using such a model, we attempt to recover the protected images in the probe set. The results are shown in Figure. 6. When the training set images and the protected image to be cracked are protected from the same key model, the attack model can recover the protected image. However, when they are protected with different key models, the trained model cannot crack the protected image. Therefore, we can increase the frequency of changing key models to effectively defend against such attacks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "The protected images obtained by our method also have good statistical properties. The correlation analysis and histogram analysis are shown in Supplementary Material B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 335, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 545, + 371 + ], + "type": "text", + "content": "We also verify that the protected image generated by a service face recognition model cannot be recognized by other models with the same function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 381, + 400, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 400, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 400, + 395 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": "Effectiveness of the variance consistency loss. We compare the variance consistency loss " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": " with the distance loss " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": " (in Equation 5) to verify the effectiveness of " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": ". For each loss, we choose a set of suitable weights " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": " separately to represent its trade-off between the protection quality and recovery quality. Its results are shown in Figure 7. The protected image generated using " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": " has a smaller SSIM value compared to the protected image using " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": ", but a clear outline can be seen. So the visual information of the original image still exists. If the protected image is to be made free of such visual information, there is a significant loss in the quality of the recovered image. In contrast, using " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": ", it is possible to generate protected images with both no residual visual information and high recovery quality. We also use another key to recover the protected image to further investigate the effect of the two different losses on the protection quality. From Figure 7, it can be concluded that the protected image using " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": " is more difficult to be recovered by similar key models with different initialization while ensuring the quality of the recovered image. We compare " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": " with more other losses in Supplementary Material D to further demonstrate the effectiveness of " + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 304, + 401, + 546, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "As shown in Figure. 5, the difference between the results when " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0.001" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " is very slight for the protected and recovery quality of the protected images. This means that our method requires no carefully adjustment of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4363" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 185 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 185 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 185 + ], + "type": "image", + "image_path": "26afe4253765ac1ba40a06862e3b501cccb7e3a3c08d1debebc91c0b5ab8e6d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 258 + ], + "type": "text", + "content": "Figure 8. Impact of parameter settings on protection quality and recovery quality. " + }, + { + "bbox": [ + 46, + 193, + 287, + 258 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 193, + 287, + 258 + ], + "type": "text", + "content": " represents the size of the convolution kernel and " + }, + { + "bbox": [ + 46, + 193, + 287, + 258 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 193, + 287, + 258 + ], + "type": "text", + "content": " represents the step size. For each pair of images, the former is the protected image and the latter is the recovery image. Below the recovery image is marked the SSIM value between it and the original image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 269, + 285, + 331 + ], + "blocks": [ + { + "bbox": [ + 50, + 269, + 285, + 331 + ], + "lines": [ + { + "bbox": [ + 50, + 269, + 285, + 331 + ], + "spans": [ + { + "bbox": [ + 50, + 269, + 285, + 331 + ], + "type": "image", + "image_path": "7ed96582425f1ae0113e64aa90651a34a0c6682e5995bb7c675928633d86ebfd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 286, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 365 + ], + "type": "text", + "content": "Figure 9. Results of AVIH method for Resnet50 as the service model. Each column represents a different sample." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 385, + 234, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 234, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 234, + 397 + ], + "type": "text", + "content": "the weight " + }, + { + "bbox": [ + 47, + 385, + 234, + 397 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 385, + 234, + 397 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 385, + 234, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 47, + 385, + 234, + 397 + ], + "type": "text", + "content": ", which is very convenient." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 601 + ], + "type": "text", + "content": "Parameters of variance consistency loss. In the variance consistency loss calculation, the process of blocking the image and calculating the internal sum of each block can be regarded as a convolution kernel with all values of 1 to convolve the image. The convolution kernel size represents the size of the block, and the step size of the convolution kernel represents the degree of overlap of the block. We explored the effect of block size and degree of overlap on performance, as shown in Figure 8. When the overlap degree of blocks is certain (especially when it is highly overlapping), the quality of the recovery image decreases as the size of the blocks increases. When the size of blocks is certain, the protection quality decreases as the overlap degree of blocks decreases. In this work, we divide the image into " + }, + { + "bbox": [ + 46, + 399, + 287, + 601 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 399, + 287, + 601 + ], + "type": "text", + "content": " blocks, with 12 pixels overlapping between adjacent blocks (step size is 1). This setting is applicable to most tasks and there is no need to change it for different samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 610, + 271, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 271, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 271, + 623 + ], + "type": "text", + "content": "4.5. Privacy protection for classification models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We trained ResNet50 [11], VGG19 [37] using the train set of CIFAR-10 [25] and implemented AVIH method for the images of the test set. The results are shown in Figure. 9. We also compared two existing methods suitable for visual information hiding and the results are shown in Table. 5, where the ITP [16] do not recover protected images. Since the LIE [42] uses protected images to train the model, the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 99, + 542, + 182 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 91 + ], + "type": "text", + "content": "Table 5. Impact of privacy protection methods on the accuracy (percentage) of classification models on CIFAR-10." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 99, + 542, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 99, + 542, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 99, + 542, + 182 + ], + "type": "table", + "html": "
MethodModelOriginal (%)Protect (%)
LIE [42]VGG1910.5987.78
Resnet5011.0091.53
ITP [16]VGG1993.9590.70
Resnet5095.5390.16
AVIH (Our)VGG1993.9593.95
Resnet5095.5395.28
", + "image_path": "cd6ceeda66587f6c022f3c3fe8835c0648e2ce354a5bfba927c93938e8aaba6a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "text", + "content": "prediction accuracy on the original data is low. We show the visual metrics, the results of the two comparative methods, and the results on Imgnet in Supplementary Material C. Our method has minimal impact on the accuracy of the models, and the average SSIM value of the recovered images can reach above 0.9 for both models after our tests." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 316, + 378, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 378, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 378, + 328 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 340, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 543 + ], + "type": "text", + "content": "In this paper, we propose a visual information hiding method AVIH based on Type-I attack. We evaluate our method by image protection of face recognition systems in the cloud. Experiments show that the AVIH method can protect images while preserving their functionality for the service model. We also propose a variance consistency loss to solve the problem of the difficult trade-off between protection quality and recovery quality. Finally, we use the AVIH method in classification tasks with satisfactory results. Our work provides a new perspective on image visual information protection, which has beneficial implications for adversarial learning communities and private protection communities. The AVIH method requires complete service model information. While this is feasible in most image storage situations, it limits the applicability of the AVIH method to a wider range of image protection scenarios. This is also a problem we will solve in future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China under Grants U22A2096 and 62036007; in part by the Fundamental Research Funds for the Central Universities under Grants QTZX23042 and XJS221502; in part by the Technology Innovation Leading Program of Shaanxi under Grant 2022QFY01-15; in part by Open Research Projects of Zhejiang Lab under Grant 2021KG0AB01; in part by the Natural Science Basis Research Plan in Shaanxi Province of China under Grant 2022JQ-696; in part by the Fundamental Research Funds for the Central Universities and in part by the Innovation Fund of Xidian University under Grant YJSJ23012." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4364" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Md Zahangir Alom, Chris Yakopcic, Mahmudul Hasan, Tarek M Taha, and Vijayan K Asari. Recurrent residual u-net for medical image segmentation. Journal of Medical Imaging, 6(1):014006-014006, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "text", + "content": "[2] Yoshinori Aono, Takuya Hayashi, Lihua Wang, Shiho Moriai, et al. Privacy-preserving deep learning via additively homomorphic encryption. IEEE Transactions on Information Forensics and Security, 13(5):1333-1345, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "text", + "content": "[3] L Arockiam and S Monikandan. Efficient cloud storage confidentiality to ensure data security. In 2014 International Conference on Computer Communication and Informatics, pages 1-5. IEEE, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "text", + "content": "[4] Deyan Chen and Hong Zhao. Data security and privacy protection issues in cloud computing. In 2012 International Conference on Computer Science and Electronics Engineering, volume 1, pages 647-651. IEEE, 2012. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 313 + ], + "type": "text", + "content": "[5] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michtitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE Transactions on Medical Imaging, 37(11):2453-2462, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 357 + ], + "type": "text", + "content": "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 412 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 468 + ], + "type": "text", + "content": "[8] Yi Ding, Guozheng Wu, Dajiang Chen, Ning Zhang, Linpeng Gong, Mingsheng Cao, and Zhiguang Qin. Deepedn: a deep-learning-based image encryption and decryption network for internet of medical things. IEEE Internet of Things Journal, 8(3):1504-1518, 2020. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 502 + ], + "type": "text", + "content": "[9] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 547 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 657 + ], + "type": "text", + "content": "[12] Lei Hu, Da Wei Zhou, Cai Xia Fu, Thomas Benkert, Chun Yu Jiang, Rui Ting Li, Li Ming Wei, and Jun Gong Zhao. Advanced zoomed diffusion-weighted imaging vs. full-field-of-view diffusion-weighted imaging in prostate cancer detection: a radiomic features study. European Radiology, 31:1760-1769, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[15] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 226 + ], + "type": "text", + "content": "[16] Hiroki Ito, Yuma Kinoshita, Maungmaung Aprilpyone, and Hitoshi Kiya. Image to perturbation: An image transformation network for generating visually protected images for privacy-preserving deep neural networks. IEEE Access, 9:64629-64638, 2021. 1, 3, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "text", + "content": "[17] Hiroki Ito, Yuma Kinoshita, and Hitoshi Kiya. Image transformation network for privacy-preserving deep neural networks and its security evaluation. In 2020 IEEE 9th Global Conference on Consumer Electronics (GCCE), pages 822-825. IEEE, 2020. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 316 + ], + "type": "text", + "content": "[18] He Kaiming, Gkioxari Georgia, Dollar Piotr, and Girshick Ross. Mask r-cnn. IEEE Transactions on Pattern Analysis & Machine Intelligence, PP:1-1, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 316, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 349 + ], + "type": "text", + "content": "[19] Seny Kamara and Kristin Lauter. Cryptographic cloud storage. In International Conference on Financial Cryptography and Data Security, pages 136-149. Springer, 2010. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "text", + "content": "[20] Manjit Kaur and Vijay Kumar. A comprehensive review on image encryption techniques. Archives of Computational Methods in Engineering, 27(1):15-43, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "text", + "content": "[21] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "type": "text", + "content": "[22] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18750-18759, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 482, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 504 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 548 + ], + "type": "text", + "content": "[24] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "type": "text", + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 9" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "text", + "content": "[26] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "text", + "content": "[27] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Towards semi-supervised deep facial expression recognition with an adaptive confidence margin. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4166-4175, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[28] Hangyu Li, Nannan Wang, Xi Yang, Xiaoyu Wang, and Xinbo Gao. Unconstrained facial expression recognition with no-reference de-elements learning. IEEE Transactions on Affective Computing, 2023. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4365" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 182 + ], + "type": "text", + "content": "[30] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 288, + 226 + ], + "type": "text", + "content": "[31] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "text", + "content": "[32] Qian Lou and Lei Jiang. She: A fast and accurate deep neural network for encrypted data. Advances in Neural Information Processing Systems, 32, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "text", + "content": "[33] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[34] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 1491-1500. IEEE, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "text", + "content": "[35] Siani Pearson and Azzedine Benameur. Privacy, security and trust issues arising from cloud computing. In 2010 IEEE Second International Conference on Cloud Computing Technology and Science, pages 693-702. IEEE, 2010. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[36] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 acm sigsac conference on computer and communications security, pages 1528-1540, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 492 + ], + "type": "text", + "content": "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 288, + 525 + ], + "type": "text", + "content": "[38] Ashish Singh and Kakali Chatterjee. Cloud security issues and challenges: A survey. Journal of Network and Computer Applications, 79:88-115, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 569 + ], + "type": "text", + "content": "[39] Warit Sirichotedumrong and Hitoshi Kiya. A gan-based image transformation scheme for privacy-preserving deep neural networks. In 2020 28th European Signal Processing Conference (EUSIPCO), pages 745-749. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "type": "text", + "content": "[40] Warit Sirichotedumrong, Takahiro Maekawa, Yuma Kinoshita, and Hitoshi Kiya. Privacy-preserving deep neural networks with pixel-based image encryption considering data augmentation in the encrypted domain. In 2019 IEEE International Conference on Image Processing (ICIP), pages 674-678. IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 679 + ], + "type": "text", + "content": "[41] Chengjin Sun, Sizhe Chen, Jia Cai, and Xiaolin Huang. Type i attack for generative models. In 2020 IEEE International Conference on Image Processing (ICIP), pages 593-597. IEEE, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[42] Masayuki Tanaka. Learnable image encryption. In 2018 IEEE International Conference on Consumer Electronics-Taiwan (ICCE-TW), pages 1-2, 2018. 1, 2, 9" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 518 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Sanli Tang, Xiaolin Huang, Mingjian Chen, Chengjin Sun, and Jie Yang. Adversarial attack type i: Cheat classifiers by significant changes. IEEE transactions on pattern analysis and machine intelligence, 43(3):1100-1109, 2019. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[44] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "text", + "content": "[45] Xueyu Wang, Jiajun Huang, Siqi Ma, Surya Nepal, and Chang Xu. Deepfake disrupter: The detector of deepfake is my friend. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14920-14929, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "type": "text", + "content": "[46] Yizhi Wang, Jun Lin, and Zhongfeng Wang. An efficient convolution core architecture for privacy-preserving deep learning. In 2018 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5. IEEE, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 274, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 318 + ], + "type": "text", + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "text", + "content": "[48] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "text", + "content": "[49] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23(10):1499-1503, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 420, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 463 + ], + "type": "text", + "content": "[50] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 518 + ], + "type": "text", + "content": "[51] Dawei Zhou, Nannan Wang, Chunlei Peng, Yi Yu, Xi Yang, and Xinbo Gao. Towards multi-domain face synthesis via domain-invariant representations and multi-level feature parts. IEEE Transactions on Multimedia, 24:3469-3479, 2021. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "4366" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_content_list.json b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a42e1e512f3613f5d56533e9c32981e168df336d --- /dev/null +++ b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_content_list.json @@ -0,0 +1,2425 @@ +[ + { + "type": "text", + "text": "Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection", + "text_level": 1, + "bbox": [ + 143, + 130, + 826, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Feng*", + "bbox": [ + 241, + 202, + 328, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifeng Xu*", + "bbox": [ + 357, + 203, + 449, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guangming Lu†", + "bbox": [ + 478, + 203, + 606, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjie Pei†", + "bbox": [ + 637, + 203, + 730, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Harbin Institute of Technology, Shenzhen", + "bbox": [ + 320, + 222, + 648, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14a709fb49e13fc575435c9f8ae16c57f9a8338dbc3c441f616e33f4835d67d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 246, + 212, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36fb8097904e9b486e760364b2a8dd2169ea7921be8a147c96505e13ceb9c398.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 351, + 212, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94b692096424a9137c9ff1a6050125e20670effcdb55690b63e55c7a56205002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 246, + 348, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccf867ee4f448548b8d8b36a9f98b17b33b178e9c0d1817d26c0ee95edad515b.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 215, + 351, + 348, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8894c44bc6af25eb6213ab01ea5b2c39507db246608c795b5d8d8917621302f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 246, + 483, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fda8e8d2ea298685ef5762a33b012c69589f778e7b35cebec77b82002db8a222.jpg", + "image_caption": [ + "Ours", + "Figure 1: Visual comparison between our method and VCNet [42] (the state-of-the-art method for blind image inpainting) on four tasks: blind image inpainting (graffiti corruption) (up left), blind image inpainting (real-image-patch corruption) (up right), watermark removal (bottom left) and shadow removal (bottom right). The groundtruth mask and the predicted masks by different methods are presented." + ], + "image_footnote": [], + "bbox": [ + 349, + 351, + 483, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad299114892c69d623d37727edc8bcf48a0ea7f94b09a82ffca58f2b48ea36fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 246, + 620, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c396922ecf11dfae74e3046b6f4812f7553a1d22403ad862b4e6b45eec992e72.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 486, + 351, + 619, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/983e989bbd6b5053d53f2fee356fc5bfbc78d344817e1e2f674417bfdf479ac7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 246, + 754, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7358596603db67f7455043ac59e055272d13f06e24c40305884a24d735931bba.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 620, + 351, + 754, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26d73c590f3ac024c68a1917bbd400650446767fbcc06fab4d2b8c7ce99448c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 246, + 890, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41baf0baf8d92866bbee179ea956d4bca23f36571bf18ed8c1abe34d0fab093e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 756, + 351, + 890, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 542, + 313, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Effective image restoration with large-size corruptions, such as blind image inpainting, entails precise detection of corruption region masks which remains extremely challenging due to diverse shapes and patterns of corruptions. In this work, we present a novel method for automatic corruption detection, which allows for blind corruption restoration without known corruption masks. Specifically, we develop a hierarchical contrastive learning framework to detect corrupted regions by capturing the intrinsic semantic distinctions between corrupted and uncorrupted regions. In particular, our model detects the corrupted mask in a coarse-to-fine manner by first predicting a coarse mask by contrastive learning in low-resolution feature space and then refines the uncertain area of the mask by high-resolution contrastive learning. A specialized hierarchical interaction mechanism is designed to facilitate the knowledge propagation of contrastive learning in different scales, boosting the modeling performance substantially. The detected multi-scale corruption masks are then leveraged to guide the corruption restoration. Detecting corrupted regions", + "bbox": [ + 75, + 564, + 472, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by learning the contrastive distinctions rather than the semantic patterns of corruptions, our model has well generalization ability across different corruption patterns. Extensive experiments demonstrate following merits of our model: 1) the superior performance over other methods on both corruption detection and various image restoration tasks including blind inpainting and watermark removal, and 2) strong generalization across different corruption patterns such as graffiti, random noise or other image content. Codes and trained weights are available at https://github.com/xyfJASON/HCL.", + "bbox": [ + 500, + 544, + 892, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 724, + 632, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An essential yet challenging step for image restoration with large-size corruptions, such as blind image inpainting, watermark removal or shadow removal, is to detect the corruption region masks in pixel level precisely. The difficulties lie in the diverse shapes and appearance of corruptions, which can be hardly modeled in a uniform pattern. As a result, blind image inpainting remains a challenging task although image inpainting with known corruption masks [27, 11, 13, 32] has achieved remarkable progress.", + "bbox": [ + 496, + 747, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A straightforward way for corruption detection is to for", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 94, + 875, + 349, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12076", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mulate it as a segmentation task and predict the corruption mask by pixel-wise binary classification. A prominent example following such modeling paradigm is VCNet [42], a state-of-the-art method for blind image inpainting, which regards corruptions as target objects and learns to recognize the semantic patterns of corruptions for detection. While VCNet can successfully detect the corruptions with uniform patterns, it has two potential limitations. First, the corruption may exhibit diverse patterns due to its potentially irregular appearance nature, which VCNet can hardly deal with. Second, it is challenging for VCNet to handle the corruption patterns that are distinct from those appearing in training data, thus it has limited generalization w.r.t. different corruption patterns.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel model for image corruption detection to circumvent the aforementioned limitations of VCNet. Instead of recognizing the semantic patterns of corruptions as VCNet does, we apply metric learning to learn an embedding space in which our model can capture the contrastive semantic distinctions between the corrupted and uncorrupted regions. To this end, we design a hierarchical contrastive learning framework for detecting corrupted regions, which predicts the corruption mask in a coarse-to-fine manner. To be specific, it first predicts a coarse mask by lightweight contrastive learning in a low-resolution feature space. Then it refines the uncertain pixels with low confidence in the predicted mask by high-resolution contrastive learning based on fine-grained features. Note that only a small fraction of predicted masks need to be refined, thus the refining process can be performed quite efficiently. To facilitate the knowledge propagation and consistency of contrastive learning between different scales, we propose a specialized hierarchical interaction mechanism. As a result, the high-resolution contrastive learning can inherit useful knowledge from the low-resolution contrastive learning to achieve precise prediction of corruption masks quite efficiently. The detected multi-scale corruption masks are further used to guide the restoration of the corrupted regions, which also follows a coarse-to-fine generative process.", + "bbox": [ + 75, + 303, + 472, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike VCNet recognizing the semantic patterns of corruptions, our model focuses on capturing the contrastive semantic distinctions between corrupted and uncorrupted regions. Thus our model has well generalization ability across different corruption patterns. As shown in Figure 1. To conclude, we make following contributions.", + "bbox": [ + 75, + 681, + 470, + 773 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a hierarchical contrastive learning framework to detect multi-scale corruption masks by learning the contrastive semantic distinctions between corrupted and uncorrupted regions.", + "- Integrating the proposed hierarchical contrastive learning module, we develop a general-purpose blind image restoration model to perform high-quality image restoration without known corruption masks." + ], + "bbox": [ + 76, + 777, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments validate the effectiveness of our model both quantitatively and qualitatively in three aspects: 1) superior performance on corruption detection (Section 4.2), 2) favorable performance compared to the specialized methods on two challenging image restoration tasks including blind image inpainting and watermark removal (Section 4.3), and 3) strong generalizability across different corruption patterns (Section 4.2 and 4.3.1).", + "bbox": [ + 500, + 90, + 893, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 220, + 640, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Restoration with known Corruption masks. Image restoration with given corruption masks assumes that the corrupted regions are known, and conventional methods [2, 12, 5] leverage uncorrupted content to restore the corrupted regions by pixel diffusion or patch replacement. However, these methods usually generate significant artifacts in the restored image due to the lack of global semantic understanding and generalization ability. With the great success of convolutional neural networks(CNNs) in computer vision, recent methods leverage an encoder-decoder framework to restore corrupted images. A typical way of CNN-based methods [29, 44, 26, 27] employs known uncorrupted regions to infer corrupted content from the outside to the inside. To synthesize reasonable semantics, some methods [31, 35, 20, 30, 16, 28] leverage structure information to improve the quality of semantic structures in the corrupted regions. Promoted by generative models [14, 25], some recent methods [11, 32, 47] attempt to learn generative prior for improving the quality of synthesized images, and generating diverse content for corrupted regions.", + "bbox": [ + 496, + 248, + 893, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image corruption detection. Early methods [3, 33] for image corruption detection assume the corrupted content follows a simple distribution, such as constant values, Gaussian noise, etc. However, previous assumption simplifies the detection of corrupted regions, limiting the application scope of image corruption detection. Recently, Wang et al. [42] relax the definition of image corruption detection by increasing the diversity of corruption types, including watermarking, raindrop, or even random images. They also propose a two-stage framework VCNet for blind image restoration, which first locates corrupted regions in a manner of image segmentation and then fills in content following the typical encoder-decoder based methods for image restoration. However, it sometimes misuses content in corrupted regions while reconstructing the complete image due to the challenge of detecting diverse patterns of corruption.", + "bbox": [ + 496, + 551, + 895, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning. Contrastive learning has been widely used in self-supervised representation learning [10, 17, 4]. Instead of matching an input image to a fixed target, contrastive learning maximizes mutual information in the representation space, keeping the query close to the positive sample and away from the negative sample. Previous works [17, 4, 41] have validated the effectiveness of contrastive", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12077", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2238f8dce237565dd153849e69bff804fa5b767dae51ad0643a14626260294f.jpg", + "image_caption": [ + "Figure 2: Architecture of our method. It leverages the designed Hierarchical Contrastive Learning to detect corruption masks and further synthesizes reasonable content for corrupted regions by the developed Transformer-based Restoration module." + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 831, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learning in high-level vision tasks, due to the inherent suitability for modeling feature contrasts. Recently, some researchers [40, 43] have attempted to introduce contrastive learning into low-level vision tasks. However, pixel-wise contrastive learning on high-resolution images suffers from large computational cost and limited performance. In this paper, we design a hierarchical contrastive learning mechanism to solve above problems and significantly strengthen the performance of image corruption detection.", + "bbox": [ + 75, + 426, + 470, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 565, + 184, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Detecting image corruptions by recognizing the semantic patterns of corruptions is difficult in that the corruptions can potentially exhibit diverse patterns. Our proposed hierarchical contrastive learning framework learns the contrastive semantic distinctions between corrupted and uncorrupted regions, and detects corruption masks in a coarse-to-fine manner. The obtained multi-scale corruption masks are then leveraged to guide the restoration of the corruptions in a coarse-to-fine generative process.", + "bbox": [ + 75, + 585, + 468, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 724, + 186, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem formulation. Given an input image $\\mathbf{I} \\in \\mathbb{R}^{C \\times H \\times W}$ which is corrupted from an intact groundtruth image $\\mathbf{O}_{\\mathrm{gt}} \\in \\mathbb{R}^{C \\times H \\times W}$ , the task of blind image restoration aims to first detect the corruption region mask $\\mathbf{M} \\in \\mathbb{R}^{1 \\times H \\times W}$ and then reconstruct a complete image $\\hat{\\mathbf{O}} \\in \\mathbb{R}^{C \\times H \\times W}$ by synthesizing realistic content $\\hat{\\mathbf{C}}$ for the corrupted regions:", + "bbox": [ + 76, + 741, + 468, + 843 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {I} = \\mathbf {O} _ {\\mathrm {g t}} \\odot \\mathbf {M} + \\mathbf {N} \\odot (1 - \\mathbf {M}), \\\\ \\hat {\\mathbf {O}} = \\mathbf {O} _ {\\mathrm {g t}} \\odot \\mathbf {M} + \\hat {\\mathbf {C}} \\odot (1 - \\mathbf {M}). \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 840, + 468, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Herein, $\\mathbf{M} \\in \\mathbb{R}^{1 \\times H \\times W}$ is a binary map where the values of", + "bbox": [ + 76, + 883, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "corrupted regions are 0 and uncorrupted regions are filled with $1.\\mathbf{N}\\in \\mathbb{R}^{C\\times H\\times W}$ is the noisy content in the corrupted regions. Precise detection of the corruption mask $\\mathbf{M}$ is crucial to the performance of image restoration.", + "bbox": [ + 496, + 426, + 890, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, our model follows the encoder-decoder framework and consists of two core modules: Hierarchical Contrastive Learning module for detecting corruption masks and Transformer-based Restoration module for corruption restoration. The corrupted image is first encoded into multi-scale feature maps by Encoder $\\mathcal{F}_{\\mathrm{enc}}$ of Transformer-based Synthesis module. Then the proposed Hierarchical Contrastive Learning module is employed to perform corruption detection in a coarse-to-fine manner from these feature maps, and predicts multi-scale corruption masks. Finally, the obtained masks are fed into Decoder $\\mathcal{F}_{\\mathrm{Dec}}$ of the Transformer-based Synthesis module for restoration, which is also in a coarse-to-fine generative process. Both Encoder $\\mathcal{F}_{\\mathrm{enc}}$ and Decoder $\\mathcal{F}_{\\mathrm{dec}}$ are mainly composed of basic transformer blocks of MAT [39], while they are equipped with additional down-sampling layers and upsampling layers respectively. We employ a shallow convolutional layer with $5\\times 5$ kernel before Encoder and after Decoder to perform basic feature transformation. Besides, we also employ another Conv-U-Net to refine high-frequency details of output results, leaning upon the local texture refinement capability and efficiency of CNNs.", + "bbox": [ + 496, + 487, + 892, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Corruption Detection by Hierarchical Contrastive Learning", + "text_level": 1, + "bbox": [ + 500, + 821, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design Hierarchical Contrastive Learning framework to guide the learning of multi-scale semantic embedding spaces of Encoder $\\mathcal{F}_{\\mathrm{enc}}$ in such a way that the semantic dis", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12078", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tances between two arbitrary pixels both within uncorrupted or corrupted regions (intra-region distance) should be minimized while the distance between a corrupted pixel and an uncorrupted pixel (inter-region distance) should be maximized. As a result, our model can capture the intrinsic semantic distinctions between corrupted and uncorrupted regions. Then our model performs clustering on all pixels in this learned embedding space to separate them into two clusters, corresponding to the corrupted and uncorrupted regions, and thus achieves the corruption mask.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, Encoder $\\mathcal{F}_{\\mathrm{enc}}$ consists of three encoding stages and produces three scales of encoded feature maps. Accordingly, Our model performs three stages of pixel-level contrastive learning for the corresponding scale of feature maps to guide the learning of its embedding space. We will first describe how our model performs corruption detection in one stage with single-scale contrastive learning. Then we will elaborate on the proposed Hierarchical Interaction Mechanism which enables our model to perform hierarchical contrastive learning quite efficiently for coarse-to-fine multi-scale mask detection.", + "bbox": [ + 75, + 242, + 473, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Single-Scale Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 419, + 372, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised Contrastive Learning. During each stage of contrastive learning, we construct positive pixel pairs by using intra-region pixels, i.e., both pixels are from either the uncorrupted region or the corrupted region. In contrast, each negative training pair consists of two inter-region pixels, one from the corrupted region and the other from the uncorrupted region. To be specific, for a query pixel $q$ from a randomly selected query set $Q$ , we construct the positive set $P$ by randomly selecting pixels from the same region as $q$ and construct the negative set $N$ from the opposite region. Then we apply Circle loss [38] to maximize the cosine similarity of positive pairs while minimizing the similarities of negative pairs. Formally, the contrastive learning in the stage $S_{s}$ is guided by the loss:", + "bbox": [ + 75, + 441, + 472, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L}} ^ {s} = \\sum_ {q \\in Q} \\log \\left[ 1 + \\sum_ {p \\in P} \\exp (- \\mathbf {e} _ {q} \\cdot \\mathbf {e} _ {p} / \\tau) \\cdot \\sum_ {n \\in N} \\exp (\\mathbf {e} _ {q} \\cdot \\mathbf {e} _ {n} / \\tau) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 657, + 468, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is a scale factor. $\\mathbf{e}_q$ is the features for pixel $q$ projected from the corresponding embedding space of Encoder $\\mathcal{F}_{\\mathrm{enc}}^s$ by a projection head in Stage- $s$ ..", + "bbox": [ + 76, + 699, + 470, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {q} = \\mathcal {F} _ {\\operatorname {p r o j}} ^ {s} \\left(\\mathcal {F} _ {\\mathrm {e n c}} ^ {s} \\left(\\mathbf {I} _ {q}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 750, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the projection head $\\mathcal{F}_{\\mathrm{proj}}$ comprises two fully connected layers and a GELU [18] layer in-between for nonlinear transformation.", + "bbox": [ + 76, + 772, + 468, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Corruption Mask Detection. Under the supervision of contrastive learning in Equation 2, the pixels within the same region, either the corrupted or the uncorrupted regions, tend to have similar representations in each scale of embedding space of Encoder $\\mathcal{F}_{\\mathrm{enc}}$ while the pixels from different regions would have dissimilar representations. Thus,", + "bbox": [ + 75, + 816, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3e1f73fb7e08b0078a8017fc0fae87110f948e0df96defb3b4a1ea1fc8b6c507.jpg", + "image_caption": [ + "Figure 3: In the stages $\\mathrm{S}_2$ and $\\mathrm{S}_3$ , both the selected representative samples and uncertain samples in the previous stage are used for contrastive learning in current stage while only the mask labels of uncertain samples are refined during mask detection." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 679, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/439466aac29665d9c911bcad7c57cf98168eab0f6224ef8d9081cc2cf4490c2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 88, + 893, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we can perform clustering to separate them into two clusters. Specifically, our model adopts K-means algorithm for clustering, and the clustering in Stage- $s$ is performed by:", + "bbox": [ + 496, + 382, + 893, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {c} _ {1} ^ {s}, \\mathbf {c} _ {2} ^ {s} \\right], \\left[ \\mathbf {M} _ {1} ^ {s}, \\mathbf {M} _ {2} ^ {s} \\right] = \\mathrm {K - m e a n s} \\left(\\mathcal {F} _ {\\operatorname {p r o j}} ^ {s} \\left(\\mathcal {F} _ {\\mathrm {e n c}} ^ {s} (\\mathbf {I})\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 428, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\mathbf{c}_1^s,\\mathbf{c}_2^s ]$ denotes the embeddings of two produced cluster centers and $[\\mathbf{M}_1^s,\\mathbf{M}_2^s ]$ are the associated binary masks. To identify which cluster corresponds to the corrupted or uncorrupted regions, we train a lightweight binary classifier consisting of two fully connected layers with a ReLU layer, and use it to perform classification on two cluster centers. Thus, the associated mask for the cluster of the corrupted region is the predicted corruption mask.", + "bbox": [ + 496, + 449, + 893, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Hierarchical Interaction Mechanism", + "text_level": 1, + "bbox": [ + 498, + 578, + 808, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model performs Hierarchical Contrastive Learning including three stages to detect the corruption mask in a coarse-to-fine manner. It first performs contrastive learning in the lowest-resolution stage $(\\mathrm{S}_1)$ to predict a coarse mask for corruption. Then it refines the uncertain pixels of the mask with low confidence by higher-resolution contrastive learning in subsequent stages $(\\mathrm{S}_2$ and $\\mathrm{S}_3)$ . We propose Hierarchical Interaction Mechanism to facilitate the interaction and knowledge propagation between different stages of contrastive learning and guarantee the semantic consistency between them. In particular, higher-resolution contrastive learning can inherit useful knowledge from previous stages of contrastive learning, which can improve the learning performance substantially. Quadtree structure are used to build the positional correspondence between feature maps in adjacent stages considering that the resolution of feature maps is always scaled by four times between adjacent stages.", + "bbox": [ + 496, + 598, + 893, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selecting High-Quality Training Samples. During the contrastive learning in $S_1$ , we construct the query set $Q$ as well as the associated positive set $P$ and negative set $N$", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12079", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Equation 2 by randomly selecting pixels from the whole feature map. To improve the efficiency of contrastive learning in the higher stages $(\\mathrm{S}_2$ and $\\mathrm{S}_3)$ , we select a small fraction of pixels from the entire feature maps, which are crucial for contrastive learning, as the high-quality candidate set for constructing $Q$ , $P$ and $N$ .", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be specific, we first measure the prediction confidence of each pixel based on the clustering results of last stage. The prediction confidence of the pixel $q$ in the stage $S_{s}$ is calculated by:", + "bbox": [ + 75, + 181, + 470, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {q} ^ {s} = \\frac {\\exp \\left(- \\left(\\mathbf {e} _ {q} \\cdot \\mathbf {c} _ {y _ {q}} ^ {s} / \\tau\\right)\\right)}{\\sum_ {i = 1} ^ {2} \\exp \\left(- \\left(\\mathbf {e} _ {q} \\cdot \\mathbf {c} _ {i} ^ {s} / \\tau\\right)\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 241, + 468, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{q}$ denotes the cluster index $q$ belongs to and $\\tau$ is the scaling factor. Then we pick out two types of pixels as the high-quality query set for the stage $S_{s+1}$ : 1) the pixels with high confidence which are regarded as representatives of two clusters and are typically close to the cluster centers; 2) the uncertain pixels with low confidence which are mostly near the boundaries between the corrupted and uncorrupted regions. Selecting pixels in such a way for constructing the positive and negative training pairs, our model is able to perform high-resolution contrastive learning in the stages $S_{2}$ and $S_{3}$ quite efficiently and effectively.", + "bbox": [ + 75, + 277, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inter-Stage Semantic Consistency. To guarantee the semantic consistency between different stages of contrastive learning, we reuse the learned features of the previous stage of contrastive learning in current stage. As shown in Figure 3, we concatenate the features in the previous stage with the features in the current stage, which are then fed into the projection head to produce input features for current stage. Accordingly, Equation 3 for high-resolution stages $(S_{2}$ and $S_{3})$ evolves into the following form:", + "bbox": [ + 75, + 444, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} _ {q} = \\mathcal {F} _ {\\text {p r o j}} ^ {s} \\left(\\operatorname {C o n c a t} \\left(\\mathcal {F} _ {\\text {e n c}} ^ {s} \\left(\\mathbf {I} _ {q}\\right), \\mathcal {F} _ {\\text {m a p}} ^ {s} \\left(\\mathcal {F} _ {\\text {e n c}} ^ {s - 1} \\left(\\mathbf {I} _ {q}\\right)\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 580, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the mapping function $\\mathcal{F}_{\\mathrm{map}}^s$ is a linear transformation to regulate the feature dimension of previous stage and balance the effect between current and previous features. Besides the semantic consistency between adjacent stages, another merit of such feature reusing is that our model is able to inherit the learned semantics from previous stage.", + "bbox": [ + 75, + 599, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refining the Mask Prediction of Uncertain Pixels. During the high-resolution contrastive learning in the stages $S_2$ and $S_3$ , we only refine the mask prediction by re-predicting the labels of uncertain pixels with low confidence in the previous stage, i.e., the type-2 samples of selected high-quality query set. The mask labels of other pixels with high confidence are directly inherited from the predictions in the previous stage according to the built quadtree, assuming that the predictions of these pixels are reliable. Thus, such refining process can be performed quite efficiently.", + "bbox": [ + 75, + 691, + 468, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Transformer-based Restoration Module", + "text_level": 1, + "bbox": [ + 76, + 847, + 418, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, the detected corruption masks are fed into Decoder of the Transformer-based Restoration", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "module for corruption restoration, which consists of the designed mask-guided transformer block.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask-guided transformer block. As illustrated in Figure 2, our transformer block contains a multi-head contextual attention (MCA) module, followed by a two-layer MLP with GELU nonlinearity in between. Besides, a layer normalization (LN) [1] layer is applied before each attention and each MLP, and a residual learning connection is employed after each module:", + "bbox": [ + 496, + 119, + 890, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {\\prime s} = \\operatorname {M C A} \\left(\\operatorname {L N} \\left(F _ {i - 1} ^ {s}\\right), \\mathbf {M} ^ {s}\\right) + F _ {i - 1} ^ {s}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 231, + 890, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {s} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(F _ {i} ^ {\\prime s}\\right)\\right) + F _ {i} ^ {\\prime s},\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 250, + 771, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{i}^{s}$ is the output features of the $i$ -th block in the stage $S_{s}$ . To fully exploit the predicted masks by hierarchical contrastive learning, we employ the multi-head contextual attention module proposed in MAT [27], which follows the shifted window manner [34] and is formulated as:", + "bbox": [ + 496, + 276, + 890, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\mathrm {T}} + \\hat {\\mathbf {M}} ^ {s}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 349, + 890, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q},\\mathbf{K},\\mathbf{V}$ are the query, key, and value matrices, respectively. $\\sqrt{d_k}$ is the scaling factor. $\\hat{\\mathbf{M}}$ is defined by", + "bbox": [ + 496, + 382, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {\\hat {M}} _ {i} ^ {s} = \\gamma \\left(\\mathbf {M} _ {i} ^ {s} - 1\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 417, + 890, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is a large positive integer to reduce the values of corrupted pixels.", + "bbox": [ + 496, + 438, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Joint Optimization for Parameter Learning", + "text_level": 1, + "bbox": [ + 500, + 474, + 867, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed Hierarchical Contrastive Learning module and the Restoration module are integrated in a holistic way based on the encoding-decoding framework, forming a general-purpose blind image restoration model, as illustrated in Figure 2. The whole model is optimized jointly. Besides the loss of hierarchical contrastive learning (Equation 2), we also adopt three more supervision signals for corruption restoration, including pixel-wise reconstruction loss, perceptual loss, and adversarial loss.", + "bbox": [ + 496, + 498, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pixel Reconstruction Loss, which pushes the restored image $\\hat{\\mathbf{O}}$ as close as its groundtruth $\\mathbf{O}_{\\mathrm{gt}}$ :", + "bbox": [ + 496, + 633, + 890, + 665 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p i x e l}} = \\left\\| \\hat {\\mathbf {O}} - \\mathbf {O} _ {\\mathrm {g t}} \\right\\| _ {1}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 667, + 890, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perceptual Loss [21], which performs semantic supervision on the restored images in the deep feature space:", + "bbox": [ + 496, + 696, + 888, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p e r c}} = \\sum_ {l = 1} ^ {L} \\frac {1}{C _ {l} H _ {l} W _ {l}} \\left\\| f _ {\\mathrm {v g g}} ^ {l} (\\hat {\\mathbf {O}}) - f _ {\\mathrm {v g g}} ^ {l} \\left(\\mathbf {O} _ {\\mathrm {g t}}\\right) \\right\\| _ {1}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 724, + 890, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Herein $f_{\\mathrm{vgg}}^{l}(\\hat{\\mathbf{O}})$ and $f_{\\mathrm{vgg}}^{l}(\\mathbf{O}_{\\mathrm{gt}})$ are the extracted features from $\\hat{\\mathbf{O}}$ and $\\mathbf{O}_{\\mathrm{gt}}$ respectively from the $l$ -th convolution layer of the pre-trained VGG-19 network [36].", + "bbox": [ + 496, + 762, + 890, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial Loss, which employs WGAN-GP [15] to encourage $\\hat{\\mathbf{O}}$ to be as realistic as its groundtruth $\\mathbf{O}_{\\mathrm{gt}}$ :", + "bbox": [ + 496, + 811, + 890, + 842 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d v}} = - \\mathbb {E} _ {\\hat {\\mathbf {O}} \\sim \\mathbb {P} _ {\\hat {\\mathbf {O}}}} \\left[ D (\\hat {\\mathbf {O}}) \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 843, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D$ is a discriminator. Overall, the whole model is optimized by:", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12080", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/825b180a91c3c7df6991f36072543670b022516ffc282562049e22cbd90147b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 88, + 209, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8727dc59fdae44809fc2cf3f414f39706c1c3a7052ed7c2b2b09a4e9cea79df5.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 119, + 157, + 209, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/682707e653889dfeaec8a7d28f4ded4f2e7993052a6fc86ea1f78b53204828a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 302, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b86f44ed78f713202c481fb0be6f3bbcfe5d400e94695dfc22ed4ee54172ba2.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 210, + 157, + 302, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0e88a4bdcdf442acf0655e2c65c04c925fcd655fee2a13a2016daec75c725e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 88, + 393, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a309e23a2cbe91ba931c5deb44b1b4a79e9f7b1af9ccea058ef6a6e8b64cb214.jpg", + "image_caption": [ + "Ours", + "Figure 4: Visualization of detected masks and reconstructed images on four randomly selected samples from test set." + ], + "image_footnote": [], + "bbox": [ + 303, + 157, + 393, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8bd36a76de3dedc97e4cbacc1e5f17e0d3727a710274ce372eee3779343dd1c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 88, + 483, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5901f0ad0cb9e25095f0378fa34de10fec2ffa6984cf21961cb92553bcb4988.jpg", + "image_caption": [ + "Groundtruth" + ], + "image_footnote": [], + "bbox": [ + 395, + 157, + 483, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b754e9d7dddeb468db2035245c287d63896ceae9285eaf0ea617c877d37032c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 88, + 576, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e4a9c8ade344890d1503f697e7f34dd176c00669102fa031c621fa95801d4d3.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 485, + 157, + 575, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/84106f35af3c94bc3014b5a90ba21772b56c5b732af6aae5ded4cb2adc7b76c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 88, + 666, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e1b3344027f886d9abff9321ed370839605f113d0bc614259de5d4ee8ff38be.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 576, + 157, + 666, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f8a6f712357e38d6915de1f852fcf9422bd984be15549b1ad5315e056bc1e22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 88, + 759, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81649f551577d913c25202a036bb08c24db5dff183cb051c2239625036e6187f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 668, + 157, + 759, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1ce3b6bb84acf4b98d4d16bddc062bc75751e26d82e1c2b6a7e5ba87188e81e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 88, + 849, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/160c77a6319911d8461a1a81e94f17ee6383e61d50ee69c5b53abaf6d7f94620.jpg", + "image_caption": [ + "Groundtruth" + ], + "image_footnote": [], + "bbox": [ + 759, + 157, + 849, + 228 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {p i x e l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}} + \\lambda_ {3} \\sum_ {s = 1} ^ {3} \\mathcal {L} _ {\\mathrm {C L}} ^ {\\mathrm {s}} + \\lambda_ {4} \\mathcal {L} _ {\\text {a d v}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 265, + 468, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3$ , and $\\lambda_4$ are hyper-parameters to balance between different losses.", + "bbox": [ + 76, + 305, + 468, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We integrate the Contrastive Learning module between Encoder and Decoder with consistent coarse-to-fine refining stages for all three modules, which potentially enables synchronized optimization of three modules in each scale.", + "bbox": [ + 76, + 335, + 468, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 407, + 209, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate the performance of our method on image corruption detection, then validates the effectiveness of our method on two challenging image restoration tasks including blind image inpainting and watermark removal. Finally, we conduct extensive ablation study to obtain more insights into our method.", + "bbox": [ + 76, + 433, + 468, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 532, + 284, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the previous work [42] for blind image inpainting, we make two pre-processing operations for data generation: 1) we randomly select natural images from large-scale datasets rather than simple constant values or Gaussian noise as the noisy content for corruptions to increase the difficulty of mask detection in blind image inpainting; 2) we smooth the mask boundaries using alpha blending to avoid distinct boundaries between the corrupted and uncorrupted regions.", + "bbox": [ + 75, + 555, + 468, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Four large-scale datasets are used in our experiments, including FFHQ (faces) [23], CelebA-HQ (faces) [22], ImageNet (objects) [9], and Places (scenes) [48]. For each dataset, images from different datasets are randomly selected as the noise content for corruptions. For instance, the images from CelebA-HQ and ImageNet are used as noise content for FFHQ. We employ the method [29] to generate irregular masks with mask shapes and corruption ratios. We perform experiments on two resolutions of images for comprehensive evaluation: $256 \\times 256$ and $512 \\times 512$ .", + "bbox": [ + 75, + 691, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adam [24] is used as the optimizer by setting $\\beta_{1}$ , $\\beta_{2}$ , initial learning rate and batch size to be 0.9, 0.999, 0.0001, and 4, respectively. More experimental details are provided in the supplementary material.", + "bbox": [ + 75, + 842, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e5077de4cad100baa41cf604936209b62c0832656778f7c31a0c69ec1b87570f.jpg", + "table_caption": [ + "Table 1: Performance comparison for mask detection on three benchmark datasets." + ], + "table_footnote": [], + "table_body": "
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-3030-600-3030-600-3030-60
Acc ↑VCNet0.9480.9430.9720.9780.9740.976
Ours0.9780.9760.9820.9830.9840.985
F1 ↑VCNet0.9670.9480.9820.9800.9840.978
Ours0.9860.9780.9890.9850.9910.987
BCE ↓VCNet0.1260.1370.0730.0550.0640.060
Ours0.0900.0950.0830.0750.0680.064
IoU ↑VCNet0.9310.9000.9660.9620.9690.959
Ours0.9750.9590.9780.9650.9820.974
", + "bbox": [ + 550, + 300, + 839, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Corruption Detection", + "text_level": 1, + "bbox": [ + 500, + 425, + 754, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Accuracy of corruption detection. We first compare the accuracy of corruption detection between our model and VCNet, the state-of-the-art method for corruption detection in blind image inpainting. Four metrics are used for comprehensive evaluation: binary cross entropy (BCE), classification accuracy (Acc), F1 score, and intersection over union (IoU). Table 1 lists the results for different corruption ratios on three datasets in $256 \\times 256$ resolution, which show that our model achieves superior performance of mask detection over VCNet in terms of all metrics except BCE. This is reasonable because VCNet is directly supervised by the BCE loss. Besides, we also make qualitative comparison in Figure 4, in which our model detects more precise corruption masks and restores higher-quality images than VCNet.", + "bbox": [ + 496, + 446, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image inpainting based on detected corruption masks by different methods. As an indirect evaluation of corruption detection, we perform image inpainting over the detected corruption masks from different methods, employing the same inpainting model MAT [27], and then compare the inpainting performance. To have a comprehensive evaluation, we also provide the inpainting performance over the detected masks from a state-of-the-art segmentation model Segmenter [37] and the groundtruth mask. The results on $512 \\times 512$ images in Table 2 show that MAT achieves the highest inpainting performance with the mask detected by our method, which is consistent with the performance comparison of corruption detection.", + "bbox": [ + 496, + 657, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalizability on unseen corruption patterns. To evaluate the generalizability of our model methods across different corruption patterns, we perform training and test over", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12081", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7db2094333dbd694e3187e960e41c373cdd19c63a9a79b4c9f11b863cb57703.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 176, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67f8a8b119b2fa73b73cd72dec6e62d86ec0cf7b63060019ab1edf71c232e922.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 181, + 90, + 269, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14f04c62e0aba61137c0a082b21ff28e4de6e2b4ca4a36c7ae7ac08ce632f3bb.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 272, + 90, + 362, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfbcc05cb16ffb4f5d35e02fb4fb08d9e278ff79d5207152211ed57894715c25.jpg", + "image_caption": [ + "Groundtruth", + "Figure 5: Visual results of blind image inpainting on $512 \\times 512$ resolution images from the Places [48] dataset." + ], + "image_footnote": [], + "bbox": [ + 364, + 90, + 454, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f4c76c69c59fb18b1eaa0aecf826284aa5f5ca66a707d8e6216afff498225627.jpg", + "table_caption": [ + "Table 2: Blind image inpainting based on detected corruption masks by different methods on $512 \\times 512$ images from Places. MAT [27] is finetuned (MAT-F) for fair comparison." + ], + "table_footnote": [], + "table_body": "
PhaseCorruption DetectionImage Inpainting
Corruption DetectionImage InpaintingACC ↑IoU ↑PSNR ↑SSIM ↑
GroundtruthMAT-F--25.340.852
Segmenter [37]MAT-F0.9740.96223.450.830
VCNet [42]MAT-F0.9750.96323.250.829
OursMAT-F0.9800.97024.530.844
VCNetVCNet0.9750.96321.490.764
OursOurs0.9800.97025.280.846
", + "bbox": [ + 101, + 332, + 441, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different corruption patterns. Table 4 presents the results of our model as well as VCNet by training them on Places dataset while performing testing directly on unseen corruption patterns such as random/constant noise and image content from CelebA-HQ [22]. It shows that our model consistently outperforms VCNet in all cases, which reveals the better generalizability of our model over VCNet.", + "bbox": [ + 75, + 426, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Downstream Image Restoration Tasks", + "text_level": 1, + "bbox": [ + 76, + 539, + 403, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Blind Image Inpainting", + "text_level": 1, + "bbox": [ + 76, + 561, + 294, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the experiments of blind image inpainting, we compare our model with VCNet and MPRNet [45], a state-of-the-art approach for image restoration. Table 3 presents the inpainting performance on images of $256 \\times 256$ resolution. Our method achieves the best performance in terms of all metrics and outperforms the other two methods by a large margin. Besides, the qualitative comparison in Figure 4 also demonstrates that our model is able to restore higher-quality images than other two methods, benefiting from the precise corruption detection by the proposed hierarchical contrastive learning mechanism as well as the integrated restoration framework. Note that we also provide the user study in the supplementary material.", + "bbox": [ + 75, + 583, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High-resolution blind image inpainting. We further conduct experiments on images of $512 \\times 512$ resolution in Places [48] dataset for blind image inpainting. As shown in Table 2, our model outperforms VCNet substantially, consistent with the comparison in the case of $256 \\times 256$ resolution. Besides, the qualitative comparison in Figure 5 also validates the superiority of our model. More visual results are presented in the supplementary material.", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a2e0d3278e574a8f0110a1165b8df986c110a917bab6ef44862ca5ede8e3ac2.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 658, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc3fdb375c5dfb549612ba7218098f5f9fa8f7ff0ca88ddc8df3efd8bce30454.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 89, + 727, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16a7579b21d5832bf851d6ff4cf78f4d36c41199721faba9f668a08809bd06c2.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 660, + 142, + 727, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3574c41333a965fce6418e4ee32a65807276fc01b2bdf39b2399dff2f3c988b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 89, + 799, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5e4f7e9b338d8e28c9412e022ca73c3d38f347ca86324dbc3b790d1a217e084.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 728, + 142, + 797, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79b72508848d2cfd23c53d234ed133ccdc14fe160da6c573bdd8fece9de2c16b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 89, + 870, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e7be72b3345aed32b636eaf132987eebbd02f46dbca43d081be3496bf963c8d.jpg", + "image_caption": [ + "Groundtruth" + ], + "image_footnote": [], + "bbox": [ + 800, + 142, + 869, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2b504c647ee84f25366488c465ffe521172168670b0ef9281410445b4bce304.jpg", + "image_caption": [ + "Figure 6: Visual results of bidirectional blind image inpainting on FFHQ [23]. More results are shown in the supplementary material.", + "Input", + "Figure 7: Visual comparison of both corruption detection and blind image inpainting over unseen corruptions. More examples could be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 506, + 252, + 612, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a6aa392cf6a13f7b396fe25824a44fd4538a9f8d66321530f462aa899130b23.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 612, + 252, + 700, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b4ca0f151db8720bea077140a6c69f7f1894b6ac0d0d6f6660b7e8259b0d473.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 702, + 252, + 789, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e1fdc50ed20d77c92d62ed5dac90de8d3ba7217c5c81759be4474b0fdd9c183.jpg", + "image_caption": [ + "Groundtruth" + ], + "image_footnote": [], + "bbox": [ + 792, + 252, + 877, + 386 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/608668cc68c469e7dc4478ee0114659091ccdf0dba5df69c7b2a831b7e1604c0.jpg", + "table_caption": [ + "Table 3: Performance comparison for blind image inpainting in terms of four evaluation metrics on three benchmark datasets." + ], + "table_footnote": [], + "table_body": "
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-2020-4040-600-2020-4040-600-2020-4040-60
PSNR ↑MPRNet33.9326.7522.0633.7626.2621.5033.3725.5921.08
VCNet29.8624.5120.5129.1323.3919.4229.6623.1919.18
Ours34.7927.9823.5433.9126.4021.8034.1126.2021.66
SSIM ↑MPRNet0.9680.8980.7990.9650.8820.7620.9630.8720.743
VCNet0.9330.8490.7360.9240.8180.6780.9420.8330.688
Ours0.9710.9100.8240.9660.8840.7670.9690.8870.746
FID ↓MPRNet3.3712.2029.743.7117.8053.384.1517.6144.73
VCNet6.8014.1728.905.6921.4859.115.9318.6240.21
Ours2.377.2815.482.6912.3541.372.8811.1031.33
LPIPS ↓MPRNet0.0270.0960.2120.0330.1230.2690.0330.1310.283
VCNet0.0440.1030.1880.0450.1280.2460.0460.1380.261
Ours0.0190.0650.1390.0250.0980.2260.0240.0950.221
", + "bbox": [ + 503, + 508, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bidirectional blind image inpainting. When real natural images are used as the noise content for corruption, we can perform bidirectional image restoration: restore either a complete noise image or a complete background image by reversing the corruption mask. It is quite challenging to perform well in inpainting of both directions. The visual results in Figure 6 show that our model performs more robust than VCNet in both corruption detection and image inpainting.", + "bbox": [ + 496, + 672, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization on unseen corruption patterns. Using the same experimental settings as evaluating the generalizability for corruption detection, we also validate the generalizability of our model across unseen corruption patterns on image inpainting. Table 5 shows the image inpainting performance of both our model and VCNet on unseen corruption patterns during training. The experimental results indi", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12082", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e3a4c52c248676b13cb52fdeb2eb0b6abcb72e9b585165b3b9eb3f53e100bd6.jpg", + "table_caption": [ + "Table 4: Generalizability over unseen corruptions for corruption detection. Models are trained on Places." + ], + "table_footnote": [], + "table_body": "
Corruption patternRandom constantCelebA-HQ [22]
Mask ratio (%)0-3030-600-3030-60
Acc ↑VCNet0.9810.9770.9750.974
Ours0.9900.9890.9860.986
F1 ↑VCNet0.9880.9780.9840.976
Ours0.9940.9900.9920.988
BCE ↓VCNet0.0500.0780.0630.068
Ours0.0500.0490.0620.054
IoU ↑VCNet0.9770.9600.9690.954
Ours0.9880.9810.9830.976
", + "bbox": [ + 86, + 142, + 333, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9fd04054a49db703dbcfb43ea502b54e3f2aba3c09539e76e5e90fefef1fb5f9.jpg", + "table_caption": [ + "Table 5: Generalizability over unseen corruptions for blind inpainting. Models are trained on Places." + ], + "table_footnote": [], + "table_body": "
Corruption PatternRandom constantCelebA-HQ [22]
Mask ratio (%)0-2020-4040-600-2020-4040-60
PSNR ↑VCNet30.4824.3319.3029.7723.2118.96
Ours37.6929.0123.0534.6026.5221.82
SSIM ↑VCNet0.9570.8760.7320.9470.8400.685
Ours0.9850.9330.8180.9720.8940.772
FID ↓VCNet6.4116.8442.545.4717.7541.19
Ours1.807.4524.482.6510.5030.42
LPIPS ↓VCNet0.0510.1310.2680.0410.1280.256
Ours0.0130.0630.1820.0210.0890.214
", + "bbox": [ + 354, + 147, + 640, + 260 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a8f10d9405a7b3a5e44eb182704b944f7f0d5feac383e6d15c39b770cdcd6802.jpg", + "table_caption": [ + "Table 6: Ablation study on hierarchical interaction mechanism." + ], + "table_footnote": [], + "table_body": "
Mask ratio (%)0-3030-60
Acc ↑w/o inter-stage consistency0.9690.965
w/o sample selection0.9750.974
Complete model0.9780.976
F1 ↑w/o inter-stage consistency0.9800.969
w/o sample selection0.9840.976
Complete model0.9860.978
BCE ↓w/o inter-stage consistency0.1020.115
w/o sample selection0.0940.101
Complete model0.0900.095
IoU ↑w/o inter-stage consistency0.9620.940
w/o sample selection0.9720.957
Complete model0.9750.959
", + "bbox": [ + 663, + 136, + 885, + 266 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54df3df2d71c4b760caa39a7c7eec07f063f4107a4e885e34831c4521a0fd54d.jpg", + "image_caption": [ + "Figure 8: Hierarchical vs single-stage contrastive learning." + ], + "image_footnote": [], + "bbox": [ + 78, + 273, + 281, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d832645ba19703dcc741d4513b50734e0e36d4d86a0ed32de12fae17216f6925.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 273, + 485, + 386 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3e0b4d47e0d74bf14341ce40a9cbc5c80041b242548d919792daac11953cc569.jpg", + "table_caption": [ + "Table 7: Performance evaluation on watermark removal." + ], + "table_footnote": [], + "table_body": "
MetricsPSNR↑SSIM↑FID↓
SIRF [46]34.630.9780.071
BS2AM [6]34.880.9790.028
DHAN [8]37.670.9860.062
BVMR [19]38.280.9850.018
Split then Refine [7]41.270.9910.011
VCNet [42]32.660.9630.032
Ours41.880.9920.007
", + "bbox": [ + 161, + 431, + 382, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cate that our model performs much better than VCNet. Furthermore, the qualitative comparison in Figure 7 also validates such better generalizability of our model over VCNet on both corruption detection and image inpainting.", + "bbox": [ + 75, + 532, + 468, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Image Watermark Removal", + "text_level": 1, + "bbox": [ + 76, + 597, + 326, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this set of experiments, we further evaluate the performance of our model on LOGO-30K dataset [7] for watermark removal. The comparison in Table 7 shows that our model performs best in terms of all metrics and compares favorably with other specialized methods for watermark removal, which demonstrates the robustness of our model across different image restoration tasks. Moreover, the visual results in Figure 9 also validate the superiority of our model over other methods in watermark removal.", + "bbox": [ + 75, + 614, + 468, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 756, + 230, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of hierarchical contrastive learning. To investigate the effectiveness of the proposed hierarchical contrastive learning framework, we compare its performance with that of single-stage contrastive learning. The results in Figure 8 demonstrate the distinct advantage of the proposed hierarchical contrastive learning framework.", + "bbox": [ + 75, + 779, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of hierarchical interaction mechanism. We further conduct experiments to investigate the effect of hierarchi", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9894bef6b72975635c568964f781b6488187f107bdc193288ebeab30fba341c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 276, + 578, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/987efa1bf8e7ca5ac201355be1dd3649f461cb368d370668e32ef4e179113436.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 276, + 656, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfa12fe0364d403f505267c978999fcb173be067cb7847ad4290cebef70617c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 276, + 733, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4936d45df96aa9456cce80ad225639fc4d0fe4c8ad33c6862f927723abdc0c69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 276, + 813, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f2b7625a05a2b6f362522758840a98ebc05707491a377af433ea5e572ab5754.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 276, + 890, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2f651141f9bbe7cd2261c16a0523f7dafa58897c40d86433018822cec0eb2c3.jpg", + "image_caption": [ + "Input", + "Figure 9: Visual comparison with the state-of-the-art model Split then Refine [7] and VCNet for watermark removal." + ], + "image_footnote": [], + "bbox": [ + 501, + 338, + 578, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe9d70bb59f802e763c8c97388c31c4d5940cea248027d6244af6a6b15231581.jpg", + "image_caption": [ + "VCNet" + ], + "image_footnote": [], + "bbox": [ + 581, + 338, + 656, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b3ed53b4e411fc86705ea930fde9eb86e766cd3e2318957ab0783d18bf59ae2.jpg", + "image_caption": [ + "Split then Refine" + ], + "image_footnote": [], + "bbox": [ + 658, + 338, + 733, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/939c0f0aaedb1a7a7f467e93b9a57e46b396092f152dde16edb62b6986dac88e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 735, + 338, + 813, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2c0b7f9e4134cb596c8a54e1de01b5586c8f1ff7e643b73c5d0165f71aebaed.jpg", + "image_caption": [ + "Groundtruth" + ], + "image_footnote": [], + "bbox": [ + 813, + 338, + 890, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cal interaction mechanism, especially the proposed 'inter-stage semantic consistency' and 'high-quality sample selection strategy'. Table 6 presents the concrete experimental results. While both techniques can boost the performance, 'inter-stage semantic consistency' is more crucial to the performance due to its essential merit: semantic propagation between stages by feature reusing.", + "bbox": [ + 496, + 449, + 890, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 564, + 617, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have designed a novel method, namely hierarchical contrastive learning framework, which can automatically detect corruption masks in pixel level and thus allows for blind image corruption restoration without known corruption masks. Extensive quantitative and qualitative comparisons have demonstrated the superior performance of our method over other methods for various corruption restoration tasks and its well generalization ability across different corruption patterns.", + "bbox": [ + 496, + 584, + 890, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 729, + 679, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the NSFC fund (Grant No. U2013210, 62006060, 62176077), in part by the Shenzhen Key Technical Project under Grant 2022N001, 2020N046, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant (Grant No. 2022A1515010306), in part by Shenzhen Fundamental Research Program (Grant No. JCYJ20210324132210025, JCYJ20220818102415032), and in part by the Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies (Grant No. 2022B1212010005).", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12083", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2", + "[3] Nian Cai, Zhenghang Su, Zhineng Lin, Han Wang, Zhijing Yang, and Bingo Wing-Kuen Ling. Blind inpainting using the fully convolutional neural network. The Visual Computer, 33(2):249-261, 2017. 2", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[5] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9):1200-1212, 2004. 2", + "[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 8", + "[7] Xiaodong Cun and Chi-Man Pun. Split then refine: stacked attention-guided resunets for blind single image visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1184-1192, 2021. 8", + "[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10680-10687, 2020. 8", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7", + "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2", + "[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11368, 2022. 1, 2", + "[12] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, pages 341-346, 2001. 2", + "[13] Xin Feng, Wenjie Pei, Fengjun Li, Fanglin Chen, David Zhang, and Guangming Lu. Generative memory-guided semantic reasoning model for image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2", + "[15] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 5", + "[16] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14134-14143, October 2021. 2", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4", + "[19] Amir Hertz, Sharon Fogel, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Blind visual motif removal from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6858–6867, 2019. 8", + "[20] Yong Shi Jie Yang, Zhiquan Qi. Learning to incorporate structure knowledge for image inpainting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12605-12612, 2020. 2", + "[21] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016. 5", + "[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6, 7, 8", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6, 7", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[26] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 2", + "[27] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10758-10768, 2022. 1, 2, 5, 6, 7", + "[28] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12084", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision, pages 85-100, 2018. 2, 6", + "[30] Hongyu Liu, Bin Jiang, Yibing Song, Wei Huang, and Chao Yang. Rethinking image inpainting via a mutual encoder-decoder with feature equalizations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 725-741. Springer, 2020. 2", + "[31] Hongyu Liu, Bin Jiang, Yi Xiao, and Chao Yang. Coherent semantic attention for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4170-4179, 2019. 2", + "[32] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11347-11357, 2022. 1, 2", + "[33] Yang Liu, Jinshan Pan, and Zhixun Su. Deep blind image inpainting. In International Conference on Intelligent Science and Big Data Engineering, pages 128-141. Springer, 2019. 2", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5", + "[35] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Qureshi, and Mehran Ebrahimi. Edgeconnect: Structure guided image inpainting using edge prediction. In The IEEE International Conference on Computer Vision Workshops, Oct 2019. 2", + "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5", + "[37] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6, 7", + "[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 4", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[40] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 3", + "[41] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2", + "[42] Yi Wang, Ying-Cong Chen, Xin Tao, and Jiaya Jia. Vcnet: A robust approach to blind image inpainting. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2, 6, 7, 8", + "[43] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 3", + "[44] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4471-4480, 2019. 2", + "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 7", + "[46] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 8", + "[47] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1438-1447, 2019. 2", + "[48] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 558 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12085", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_model.json b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_model.json new file mode 100644 index 0000000000000000000000000000000000000000..136ac62694f1e2f2687e75fe6959b509f9441aaa --- /dev/null +++ b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_model.json @@ -0,0 +1,3223 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.828, + 0.177 + ], + "angle": 0, + "content": "Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.203, + 0.33, + 0.221 + ], + "angle": 0, + "content": "Xin Feng*" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.204, + 0.45, + 0.222 + ], + "angle": 0, + "content": "Yifeng Xu*" + }, + { + "type": "text", + "bbox": [ + 0.479, + 0.204, + 0.607, + 0.222 + ], + "angle": 0, + "content": "Guangming Lu†" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.204, + 0.731, + 0.222 + ], + "angle": 0, + "content": "Wenjie Pei†" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.223, + 0.65, + 0.24 + ], + "angle": 0, + "content": "Harbin Institute of Technology, Shenzhen" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.247, + 0.214, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.352, + 0.214, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.458, + 0.165, + 0.471 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.247, + 0.349, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.352, + 0.349, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.458, + 0.304, + 0.47 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.247, + 0.484, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.352, + 0.484, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.458, + 0.434, + 0.47 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.247, + 0.621, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.352, + 0.62, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.458, + 0.571, + 0.47 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.247, + 0.755, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.352, + 0.755, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.458, + 0.706, + 0.47 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.247, + 0.891, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.352, + 0.891, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.458, + 0.831, + 0.469 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.893, + 0.533 + ], + "angle": 0, + "content": "Figure 1: Visual comparison between our method and VCNet [42] (the state-of-the-art method for blind image inpainting) on four tasks: blind image inpainting (graffiti corruption) (up left), blind image inpainting (real-image-patch corruption) (up right), watermark removal (bottom left) and shadow removal (bottom right). The groundtruth mask and the predicted masks by different methods are presented." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.543, + 0.314, + 0.558 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Effective image restoration with large-size corruptions, such as blind image inpainting, entails precise detection of corruption region masks which remains extremely challenging due to diverse shapes and patterns of corruptions. In this work, we present a novel method for automatic corruption detection, which allows for blind corruption restoration without known corruption masks. Specifically, we develop a hierarchical contrastive learning framework to detect corrupted regions by capturing the intrinsic semantic distinctions between corrupted and uncorrupted regions. In particular, our model detects the corrupted mask in a coarse-to-fine manner by first predicting a coarse mask by contrastive learning in low-resolution feature space and then refines the uncertain area of the mask by high-resolution contrastive learning. A specialized hierarchical interaction mechanism is designed to facilitate the knowledge propagation of contrastive learning in different scales, boosting the modeling performance substantially. The detected multi-scale corruption masks are then leveraged to guide the corruption restoration. Detecting corrupted regions" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.545, + 0.893, + 0.711 + ], + "angle": 0, + "content": "by learning the contrastive distinctions rather than the semantic patterns of corruptions, our model has well generalization ability across different corruption patterns. Extensive experiments demonstrate following merits of our model: 1) the superior performance over other methods on both corruption detection and various image restoration tasks including blind inpainting and watermark removal, and 2) strong generalization across different corruption patterns such as graffiti, random noise or other image content. Codes and trained weights are available at https://github.com/xyfJASON/HCL." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.725, + 0.633, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.884 + ], + "angle": 0, + "content": "An essential yet challenging step for image restoration with large-size corruptions, such as blind image inpainting, watermark removal or shadow removal, is to detect the corruption region masks in pixel level precisely. The difficulties lie in the diverse shapes and appearance of corruptions, which can be hardly modeled in a uniform pattern. As a result, blind image inpainting remains a challenging task although image inpainting with known corruption masks [27, 11, 13, 32] has achieved remarkable progress." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A straightforward way for corruption detection is to for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.351, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.351, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12076" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "mulate it as a segmentation task and predict the corruption mask by pixel-wise binary classification. A prominent example following such modeling paradigm is VCNet [42], a state-of-the-art method for blind image inpainting, which regards corruptions as target objects and learns to recognize the semantic patterns of corruptions for detection. While VCNet can successfully detect the corruptions with uniform patterns, it has two potential limitations. First, the corruption may exhibit diverse patterns due to its potentially irregular appearance nature, which VCNet can hardly deal with. Second, it is challenging for VCNet to handle the corruption patterns that are distinct from those appearing in training data, thus it has limited generalization w.r.t. different corruption patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.682 + ], + "angle": 0, + "content": "In this work, we propose a novel model for image corruption detection to circumvent the aforementioned limitations of VCNet. Instead of recognizing the semantic patterns of corruptions as VCNet does, we apply metric learning to learn an embedding space in which our model can capture the contrastive semantic distinctions between the corrupted and uncorrupted regions. To this end, we design a hierarchical contrastive learning framework for detecting corrupted regions, which predicts the corruption mask in a coarse-to-fine manner. To be specific, it first predicts a coarse mask by lightweight contrastive learning in a low-resolution feature space. Then it refines the uncertain pixels with low confidence in the predicted mask by high-resolution contrastive learning based on fine-grained features. Note that only a small fraction of predicted masks need to be refined, thus the refining process can be performed quite efficiently. To facilitate the knowledge propagation and consistency of contrastive learning between different scales, we propose a specialized hierarchical interaction mechanism. As a result, the high-resolution contrastive learning can inherit useful knowledge from the low-resolution contrastive learning to achieve precise prediction of corruption masks quite efficiently. The detected multi-scale corruption masks are further used to guide the restoration of the corrupted regions, which also follows a coarse-to-fine generative process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.472, + 0.775 + ], + "angle": 0, + "content": "Unlike VCNet recognizing the semantic patterns of corruptions, our model focuses on capturing the contrastive semantic distinctions between corrupted and uncorrupted regions. Thus our model has well generalization ability across different corruption patterns. As shown in Figure 1. To conclude, we make following contributions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.778, + 0.47, + 0.839 + ], + "angle": 0, + "content": "- We design a hierarchical contrastive learning framework to detect multi-scale corruption masks by learning the contrastive semantic distinctions between corrupted and uncorrupted regions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- Integrating the proposed hierarchical contrastive learning module, we develop a general-purpose blind image restoration model to perform high-quality image restoration without known corruption masks." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.778, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.214 + ], + "angle": 0, + "content": "- Extensive experiments validate the effectiveness of our model both quantitatively and qualitatively in three aspects: 1) superior performance on corruption detection (Section 4.2), 2) favorable performance compared to the specialized methods on two challenging image restoration tasks including blind image inpainting and watermark removal (Section 4.3), and 3) strong generalizability across different corruption patterns (Section 4.2 and 4.3.1)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.222, + 0.642, + 0.239 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.895, + 0.552 + ], + "angle": 0, + "content": "Image Restoration with known Corruption masks. Image restoration with given corruption masks assumes that the corrupted regions are known, and conventional methods [2, 12, 5] leverage uncorrupted content to restore the corrupted regions by pixel diffusion or patch replacement. However, these methods usually generate significant artifacts in the restored image due to the lack of global semantic understanding and generalization ability. With the great success of convolutional neural networks(CNNs) in computer vision, recent methods leverage an encoder-decoder framework to restore corrupted images. A typical way of CNN-based methods [29, 44, 26, 27] employs known uncorrupted regions to infer corrupted content from the outside to the inside. To synthesize reasonable semantics, some methods [31, 35, 20, 30, 16, 28] leverage structure information to improve the quality of semantic structures in the corrupted regions. Promoted by generative models [14, 25], some recent methods [11, 32, 47] attempt to learn generative prior for improving the quality of synthesized images, and generating diverse content for corrupted regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.896, + 0.794 + ], + "angle": 0, + "content": "Image corruption detection. Early methods [3, 33] for image corruption detection assume the corrupted content follows a simple distribution, such as constant values, Gaussian noise, etc. However, previous assumption simplifies the detection of corrupted regions, limiting the application scope of image corruption detection. Recently, Wang et al. [42] relax the definition of image corruption detection by increasing the diversity of corruption types, including watermarking, raindrop, or even random images. They also propose a two-stage framework VCNet for blind image restoration, which first locates corrupted regions in a manner of image segmentation and then fills in content following the typical encoder-decoder based methods for image restoration. However, it sometimes misuses content in corrupted regions while reconstructing the complete image due to the challenge of detecting diverse patterns of corruption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Contrastive learning. Contrastive learning has been widely used in self-supervised representation learning [10, 17, 4]. Instead of matching an input image to a fixed target, contrastive learning maximizes mutual information in the representation space, keeping the query close to the positive sample and away from the negative sample. Previous works [17, 4, 41] have validated the effectiveness of contrastive" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12077" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.833, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.893, + 0.416 + ], + "angle": 0, + "content": "Figure 2: Architecture of our method. It leverages the designed Hierarchical Contrastive Learning to detect corruption masks and further synthesizes reasonable content for corrupted regions by the developed Transformer-based Restoration module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.471, + 0.563 + ], + "angle": 0, + "content": "learning in high-level vision tasks, due to the inherent suitability for modeling feature contrasts. Recently, some researchers [40, 43] have attempted to introduce contrastive learning into low-level vision tasks. However, pixel-wise contrastive learning on high-resolution images suffers from large computational cost and limited performance. In this paper, we design a hierarchical contrastive learning mechanism to solve above problems and significantly strengthen the performance of image corruption detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.566, + 0.185, + 0.583 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Detecting image corruptions by recognizing the semantic patterns of corruptions is difficult in that the corruptions can potentially exhibit diverse patterns. Our proposed hierarchical contrastive learning framework learns the contrastive semantic distinctions between corrupted and uncorrupted regions, and detects corruption masks in a coarse-to-fine manner. The obtained multi-scale corruption masks are then leveraged to guide the restoration of the corruptions in a coarse-to-fine generative process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.187, + 0.739 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.742, + 0.469, + 0.844 + ], + "angle": 0, + "content": "Problem formulation. Given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{C \\times H \\times W}\\) which is corrupted from an intact groundtruth image \\(\\mathbf{O}_{\\mathrm{gt}} \\in \\mathbb{R}^{C \\times H \\times W}\\), the task of blind image restoration aims to first detect the corruption region mask \\(\\mathbf{M} \\in \\mathbb{R}^{1 \\times H \\times W}\\) and then reconstruct a complete image \\(\\hat{\\mathbf{O}} \\in \\mathbb{R}^{C \\times H \\times W}\\) by synthesizing realistic content \\(\\hat{\\mathbf{C}}\\) for the corrupted regions:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.842, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {I} = \\mathbf {O} _ {\\mathrm {g t}} \\odot \\mathbf {M} + \\mathbf {N} \\odot (1 - \\mathbf {M}), \\\\ \\hat {\\mathbf {O}} = \\mathbf {O} _ {\\mathrm {g t}} \\odot \\mathbf {M} + \\hat {\\mathbf {C}} \\odot (1 - \\mathbf {M}). \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.885, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Herein, \\(\\mathbf{M} \\in \\mathbb{R}^{1 \\times H \\times W}\\) is a binary map where the values of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.488 + ], + "angle": 0, + "content": "corrupted regions are 0 and uncorrupted regions are filled with \\(1.\\mathbf{N}\\in \\mathbb{R}^{C\\times H\\times W}\\) is the noisy content in the corrupted regions. Precise detection of the corruption mask \\(\\mathbf{M}\\) is crucial to the performance of image restoration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.893, + 0.82 + ], + "angle": 0, + "content": "As shown in Figure 2, our model follows the encoder-decoder framework and consists of two core modules: Hierarchical Contrastive Learning module for detecting corruption masks and Transformer-based Restoration module for corruption restoration. The corrupted image is first encoded into multi-scale feature maps by Encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) of Transformer-based Synthesis module. Then the proposed Hierarchical Contrastive Learning module is employed to perform corruption detection in a coarse-to-fine manner from these feature maps, and predicts multi-scale corruption masks. Finally, the obtained masks are fed into Decoder \\(\\mathcal{F}_{\\mathrm{Dec}}\\) of the Transformer-based Synthesis module for restoration, which is also in a coarse-to-fine generative process. Both Encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) and Decoder \\(\\mathcal{F}_{\\mathrm{dec}}\\) are mainly composed of basic transformer blocks of MAT [39], while they are equipped with additional down-sampling layers and upsampling layers respectively. We employ a shallow convolutional layer with \\(5\\times 5\\) kernel before Encoder and after Decoder to perform basic feature transformation. Besides, we also employ another Conv-U-Net to refine high-frequency details of output results, leaning upon the local texture refinement capability and efficiency of CNNs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.892, + 0.854 + ], + "angle": 0, + "content": "3.2. Corruption Detection by Hierarchical Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We design Hierarchical Contrastive Learning framework to guide the learning of multi-scale semantic embedding spaces of Encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) in such a way that the semantic dis" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "tances between two arbitrary pixels both within uncorrupted or corrupted regions (intra-region distance) should be minimized while the distance between a corrupted pixel and an uncorrupted pixel (inter-region distance) should be maximized. As a result, our model can capture the intrinsic semantic distinctions between corrupted and uncorrupted regions. Then our model performs clustering on all pixels in this learned embedding space to separate them into two clusters, corresponding to the corrupted and uncorrupted regions, and thus achieves the corruption mask." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.474, + 0.41 + ], + "angle": 0, + "content": "As shown in Figure 2, Encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) consists of three encoding stages and produces three scales of encoded feature maps. Accordingly, Our model performs three stages of pixel-level contrastive learning for the corresponding scale of feature maps to guide the learning of its embedding space. We will first describe how our model performs corruption detection in one stage with single-scale contrastive learning. Then we will elaborate on the proposed Hierarchical Interaction Mechanism which enables our model to perform hierarchical contrastive learning quite efficiently for coarse-to-fine multi-scale mask detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.42, + 0.373, + 0.435 + ], + "angle": 0, + "content": "3.2.1 Single-Scale Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Supervised Contrastive Learning. During each stage of contrastive learning, we construct positive pixel pairs by using intra-region pixels, i.e., both pixels are from either the uncorrupted region or the corrupted region. In contrast, each negative training pair consists of two inter-region pixels, one from the corrupted region and the other from the uncorrupted region. To be specific, for a query pixel \\( q \\) from a randomly selected query set \\( Q \\), we construct the positive set \\( P \\) by randomly selecting pixels from the same region as \\( q \\) and construct the negative set \\( N \\) from the opposite region. Then we apply Circle loss [38] to maximize the cosine similarity of positive pairs while minimizing the similarities of negative pairs. Formally, the contrastive learning in the stage \\( S_{s} \\) is guided by the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.659, + 0.47, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L}} ^ {s} = \\sum_ {q \\in Q} \\log \\left[ 1 + \\sum_ {p \\in P} \\exp (- \\mathbf {e} _ {q} \\cdot \\mathbf {e} _ {p} / \\tau) \\cdot \\sum_ {n \\in N} \\exp (\\mathbf {e} _ {q} \\cdot \\mathbf {e} _ {n} / \\tau) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.7, + 0.471, + 0.746 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a scale factor. \\(\\mathbf{e}_q\\) is the features for pixel \\(q\\) projected from the corresponding embedding space of Encoder \\(\\mathcal{F}_{\\mathrm{enc}}^s\\) by a projection head in Stage- \\(s\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.75, + 0.47, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {q} = \\mathcal {F} _ {\\operatorname {p r o j}} ^ {s} \\left(\\mathcal {F} _ {\\mathrm {e n c}} ^ {s} \\left(\\mathbf {I} _ {q}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.773, + 0.47, + 0.817 + ], + "angle": 0, + "content": "where the projection head \\(\\mathcal{F}_{\\mathrm{proj}}\\) comprises two fully connected layers and a GELU [18] layer in-between for nonlinear transformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.817, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Corruption Mask Detection. Under the supervision of contrastive learning in Equation 2, the pixels within the same region, either the corrupted or the uncorrupted regions, tend to have similar representations in each scale of embedding space of Encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) while the pixels from different regions would have dissimilar representations. Thus," + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.681, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.089, + 0.895, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.296, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Figure 3: In the stages \\(\\mathrm{S}_2\\) and \\(\\mathrm{S}_3\\), both the selected representative samples and uncertain samples in the previous stage are used for contrastive learning in current stage while only the mask labels of uncertain samples are refined during mask detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.895, + 0.426 + ], + "angle": 0, + "content": "we can perform clustering to separate them into two clusters. Specifically, our model adopts K-means algorithm for clustering, and the clustering in Stage- \\(s\\) is performed by:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.429, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {c} _ {1} ^ {s}, \\mathbf {c} _ {2} ^ {s} \\right], \\left[ \\mathbf {M} _ {1} ^ {s}, \\mathbf {M} _ {2} ^ {s} \\right] = \\mathrm {K - m e a n s} \\left(\\mathcal {F} _ {\\operatorname {p r o j}} ^ {s} \\left(\\mathcal {F} _ {\\mathrm {e n c}} ^ {s} (\\mathbf {I})\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.572 + ], + "angle": 0, + "content": "where \\([\\mathbf{c}_1^s,\\mathbf{c}_2^s ]\\) denotes the embeddings of two produced cluster centers and \\([\\mathbf{M}_1^s,\\mathbf{M}_2^s ]\\) are the associated binary masks. To identify which cluster corresponds to the corrupted or uncorrupted regions, we train a lightweight binary classifier consisting of two fully connected layers with a ReLU layer, and use it to perform classification on two cluster centers. Thus, the associated mask for the cluster of the corrupted region is the predicted corruption mask." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.579, + 0.81, + 0.593 + ], + "angle": 0, + "content": "3.2.2 Hierarchical Interaction Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.856 + ], + "angle": 0, + "content": "Our model performs Hierarchical Contrastive Learning including three stages to detect the corruption mask in a coarse-to-fine manner. It first performs contrastive learning in the lowest-resolution stage \\((\\mathrm{S}_1)\\) to predict a coarse mask for corruption. Then it refines the uncertain pixels of the mask with low confidence by higher-resolution contrastive learning in subsequent stages \\((\\mathrm{S}_2\\) and \\(\\mathrm{S}_3)\\). We propose Hierarchical Interaction Mechanism to facilitate the interaction and knowledge propagation between different stages of contrastive learning and guarantee the semantic consistency between them. In particular, higher-resolution contrastive learning can inherit useful knowledge from previous stages of contrastive learning, which can improve the learning performance substantially. Quadtree structure are used to build the positional correspondence between feature maps in adjacent stages considering that the resolution of feature maps is always scaled by four times between adjacent stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Selecting High-Quality Training Samples. During the contrastive learning in \\( S_1 \\), we construct the query set \\( Q \\) as well as the associated positive set \\( P \\) and negative set \\( N \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12079" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "in Equation 2 by randomly selecting pixels from the whole feature map. To improve the efficiency of contrastive learning in the higher stages \\((\\mathrm{S}_2\\) and \\(\\mathrm{S}_3)\\), we select a small fraction of pixels from the entire feature maps, which are crucial for contrastive learning, as the high-quality candidate set for constructing \\(Q\\), \\(P\\) and \\(N\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.472, + 0.243 + ], + "angle": 0, + "content": "To be specific, we first measure the prediction confidence of each pixel based on the clustering results of last stage. The prediction confidence of the pixel \\( q \\) in the stage \\( S_{s} \\) is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.242, + 0.469, + 0.279 + ], + "angle": 0, + "content": "\\[\nz _ {q} ^ {s} = \\frac {\\exp \\left(- \\left(\\mathbf {e} _ {q} \\cdot \\mathbf {c} _ {y _ {q}} ^ {s} / \\tau\\right)\\right)}{\\sum_ {i = 1} ^ {2} \\exp \\left(- \\left(\\mathbf {e} _ {q} \\cdot \\mathbf {c} _ {i} ^ {s} / \\tau\\right)\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.47, + 0.445 + ], + "angle": 0, + "content": "where \\( y_{q} \\) denotes the cluster index \\( q \\) belongs to and \\( \\tau \\) is the scaling factor. Then we pick out two types of pixels as the high-quality query set for the stage \\( S_{s+1} \\): 1) the pixels with high confidence which are regarded as representatives of two clusters and are typically close to the cluster centers; 2) the uncertain pixels with low confidence which are mostly near the boundaries between the corrupted and uncorrupted regions. Selecting pixels in such a way for constructing the positive and negative training pairs, our model is able to perform high-resolution contrastive learning in the stages \\( S_{2} \\) and \\( S_{3} \\) quite efficiently and effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Inter-Stage Semantic Consistency. To guarantee the semantic consistency between different stages of contrastive learning, we reuse the learned features of the previous stage of contrastive learning in current stage. As shown in Figure 3, we concatenate the features in the previous stage with the features in the current stage, which are then fed into the projection head to produce input features for current stage. Accordingly, Equation 3 for high-resolution stages \\((S_{2}\\) and \\(S_{3})\\) evolves into the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.582, + 0.469, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} _ {q} = \\mathcal {F} _ {\\text {p r o j}} ^ {s} \\left(\\operatorname {C o n c a t} \\left(\\mathcal {F} _ {\\text {e n c}} ^ {s} \\left(\\mathbf {I} _ {q}\\right), \\mathcal {F} _ {\\text {m a p}} ^ {s} \\left(\\mathcal {F} _ {\\text {e n c}} ^ {s - 1} \\left(\\mathbf {I} _ {q}\\right)\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.469, + 0.692 + ], + "angle": 0, + "content": "where the mapping function \\(\\mathcal{F}_{\\mathrm{map}}^s\\) is a linear transformation to regulate the feature dimension of previous stage and balance the effect between current and previous features. Besides the semantic consistency between adjacent stages, another merit of such feature reusing is that our model is able to inherit the learned semantics from previous stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.469, + 0.844 + ], + "angle": 0, + "content": "Refining the Mask Prediction of Uncertain Pixels. During the high-resolution contrastive learning in the stages \\( S_2 \\) and \\( S_3 \\), we only refine the mask prediction by re-predicting the labels of uncertain pixels with low confidence in the previous stage, i.e., the type-2 samples of selected high-quality query set. The mask labels of other pixels with high confidence are directly inherited from the predictions in the previous stage according to the built quadtree, assuming that the predictions of these pixels are reliable. Thus, such refining process can be performed quite efficiently." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.419, + 0.863 + ], + "angle": 0, + "content": "3.3. Transformer-based Restoration Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "As shown in Figure 2, the detected corruption masks are fed into Decoder of the Transformer-based Restoration" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "module for corruption restoration, which consists of the designed mask-guided transformer block." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Mask-guided transformer block. As illustrated in Figure 2, our transformer block contains a multi-head contextual attention (MCA) module, followed by a two-layer MLP with GELU nonlinearity in between. Besides, a layer normalization (LN) [1] layer is applied before each attention and each MLP, and a residual learning connection is employed after each module:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.232, + 0.892, + 0.255 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {\\prime s} = \\operatorname {M C A} \\left(\\operatorname {L N} \\left(F _ {i - 1} ^ {s}\\right), \\mathbf {M} ^ {s}\\right) + F _ {i - 1} ^ {s}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.251, + 0.772, + 0.268 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {s} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(F _ {i} ^ {\\prime s}\\right)\\right) + F _ {i} ^ {\\prime s},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.347 + ], + "angle": 0, + "content": "where \\( F_{i}^{s} \\) is the output features of the \\( i \\)-th block in the stage \\( S_{s} \\). To fully exploit the predicted masks by hierarchical contrastive learning, we employ the multi-head contextual attention module proposed in MAT [27], which follows the shifted window manner [34] and is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.35, + 0.891, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\mathrm {T}} + \\hat {\\mathbf {M}} ^ {s}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.891, + 0.413 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q},\\mathbf{K},\\mathbf{V}\\) are the query, key, and value matrices, respectively. \\(\\sqrt{d_k}\\) is the scaling factor. \\(\\hat{\\mathbf{M}}\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.418, + 0.891, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathbf {\\hat {M}} _ {i} ^ {s} = \\gamma \\left(\\mathbf {M} _ {i} ^ {s} - 1\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.469 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a large positive integer to reduce the values of corrupted pixels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.476, + 0.869, + 0.492 + ], + "angle": 0, + "content": "3.4. Joint Optimization for Parameter Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.634 + ], + "angle": 0, + "content": "The proposed Hierarchical Contrastive Learning module and the Restoration module are integrated in a holistic way based on the encoding-decoding framework, forming a general-purpose blind image restoration model, as illustrated in Figure 2. The whole model is optimized jointly. Besides the loss of hierarchical contrastive learning (Equation 2), we also adopt three more supervision signals for corruption restoration, including pixel-wise reconstruction loss, perceptual loss, and adversarial loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.891, + 0.666 + ], + "angle": 0, + "content": "Pixel Reconstruction Loss, which pushes the restored image \\(\\hat{\\mathbf{O}}\\) as close as its groundtruth \\(\\mathbf{O}_{\\mathrm{gt}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.669, + 0.891, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p i x e l}} = \\left\\| \\hat {\\mathbf {O}} - \\mathbf {O} _ {\\mathrm {g t}} \\right\\| _ {1}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.89, + 0.725 + ], + "angle": 0, + "content": "Perceptual Loss [21], which performs semantic supervision on the restored images in the deep feature space:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.725, + 0.891, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p e r c}} = \\sum_ {l = 1} ^ {L} \\frac {1}{C _ {l} H _ {l} W _ {l}} \\left\\| f _ {\\mathrm {v g g}} ^ {l} (\\hat {\\mathbf {O}}) - f _ {\\mathrm {v g g}} ^ {l} \\left(\\mathbf {O} _ {\\mathrm {g t}}\\right) \\right\\| _ {1}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.891, + 0.812 + ], + "angle": 0, + "content": "Herein \\( f_{\\mathrm{vgg}}^{l}(\\hat{\\mathbf{O}}) \\) and \\( f_{\\mathrm{vgg}}^{l}(\\mathbf{O}_{\\mathrm{gt}}) \\) are the extracted features from \\( \\hat{\\mathbf{O}} \\) and \\( \\mathbf{O}_{\\mathrm{gt}} \\) respectively from the \\( l \\)-th convolution layer of the pre-trained VGG-19 network [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.812, + 0.891, + 0.843 + ], + "angle": 0, + "content": "Adversarial Loss, which employs WGAN-GP [15] to encourage \\(\\hat{\\mathbf{O}}\\) to be as realistic as its groundtruth \\(\\mathbf{O}_{\\mathrm{gt}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.844, + 0.891, + 0.871 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d v}} = - \\mathbb {E} _ {\\hat {\\mathbf {O}} \\sim \\mathbb {P} _ {\\hat {\\mathbf {O}}}} \\left[ D (\\hat {\\mathbf {O}}) \\right], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(D\\) is a discriminator. Overall, the whole model is optimized by:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12080" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.21, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.159, + 0.21, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.23, + 0.178, + 0.24 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.303, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.159, + 0.303, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.229, + 0.274, + 0.239 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.089, + 0.395, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.159, + 0.394, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.23, + 0.361, + 0.24 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.089, + 0.485, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.159, + 0.485, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.23, + 0.468, + 0.24 + ], + "angle": 0, + "content": "Groundtruth" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.089, + 0.577, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.159, + 0.576, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.23, + 0.544, + 0.24 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.089, + 0.668, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.159, + 0.668, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.23, + 0.639, + 0.24 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.089, + 0.76, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.159, + 0.76, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.23, + 0.726, + 0.239 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.089, + 0.851, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.159, + 0.851, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.23, + 0.833, + 0.24 + ], + "angle": 0, + "content": "Groundtruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.242, + 0.865, + 0.257 + ], + "angle": 0, + "content": "Figure 4: Visualization of detected masks and reconstructed images on four randomly selected samples from test set." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.266, + 0.469, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {p i x e l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}} + \\lambda_ {3} \\sum_ {s = 1} ^ {3} \\mathcal {L} _ {\\mathrm {C L}} ^ {\\mathrm {s}} + \\lambda_ {4} \\mathcal {L} _ {\\text {a d v}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.469, + 0.336 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3\\), and \\(\\lambda_4\\) are hyper-parameters to balance between different losses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.469, + 0.398 + ], + "angle": 0, + "content": "We integrate the Contrastive Learning module between Encoder and Decoder with consistent coarse-to-fine refining stages for all three modules, which potentially enables synchronized optimization of three modules in each scale." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.409, + 0.21, + 0.426 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.434, + 0.47, + 0.525 + ], + "angle": 0, + "content": "We first evaluate the performance of our method on image corruption detection, then validates the effectiveness of our method on two challenging image restoration tasks including blind image inpainting and watermark removal. Finally, we conduct extensive ablation study to obtain more insights into our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.533, + 0.285, + 0.55 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.692 + ], + "angle": 0, + "content": "Following the previous work [42] for blind image inpainting, we make two pre-processing operations for data generation: 1) we randomly select natural images from large-scale datasets rather than simple constant values or Gaussian noise as the noisy content for corruptions to increase the difficulty of mask detection in blind image inpainting; 2) we smooth the mask boundaries using alpha blending to avoid distinct boundaries between the corrupted and uncorrupted regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.469, + 0.842 + ], + "angle": 0, + "content": "Four large-scale datasets are used in our experiments, including FFHQ (faces) [23], CelebA-HQ (faces) [22], ImageNet (objects) [9], and Places (scenes) [48]. For each dataset, images from different datasets are randomly selected as the noise content for corruptions. For instance, the images from CelebA-HQ and ImageNet are used as noise content for FFHQ. We employ the method [29] to generate irregular masks with mask shapes and corruption ratios. We perform experiments on two resolutions of images for comprehensive evaluation: \\(256 \\times 256\\) and \\(512 \\times 512\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.843, + 0.469, + 0.904 + ], + "angle": 0, + "content": "Adam [24] is used as the optimizer by setting \\(\\beta_{1}\\), \\(\\beta_{2}\\), initial learning rate and batch size to be 0.9, 0.999, 0.0001, and 4, respectively. More experimental details are provided in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.267, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Table 1: Performance comparison for mask detection on three benchmark datasets." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.301, + 0.84, + 0.419 + ], + "angle": 0, + "content": "
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-3030-600-3030-600-3030-60
Acc ↑VCNet0.9480.9430.9720.9780.9740.976
Ours0.9780.9760.9820.9830.9840.985
F1 ↑VCNet0.9670.9480.9820.9800.9840.978
Ours0.9860.9780.9890.9850.9910.987
BCE ↓VCNet0.1260.1370.0730.0550.0640.060
Ours0.0900.0950.0830.0750.0680.064
IoU ↑VCNet0.9310.9000.9660.9620.9690.959
Ours0.9750.9590.9780.9650.9820.974
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.426, + 0.755, + 0.443 + ], + "angle": 0, + "content": "4.2. Image Corruption Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Accuracy of corruption detection. We first compare the accuracy of corruption detection between our model and VCNet, the state-of-the-art method for corruption detection in blind image inpainting. Four metrics are used for comprehensive evaluation: binary cross entropy (BCE), classification accuracy (Acc), F1 score, and intersection over union (IoU). Table 1 lists the results for different corruption ratios on three datasets in \\(256 \\times 256\\) resolution, which show that our model achieves superior performance of mask detection over VCNet in terms of all metrics except BCE. This is reasonable because VCNet is directly supervised by the BCE loss. Besides, we also make qualitative comparison in Figure 4, in which our model detects more precise corruption masks and restores higher-quality images than VCNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Image inpainting based on detected corruption masks by different methods. As an indirect evaluation of corruption detection, we perform image inpainting over the detected corruption masks from different methods, employing the same inpainting model MAT [27], and then compare the inpainting performance. To have a comprehensive evaluation, we also provide the inpainting performance over the detected masks from a state-of-the-art segmentation model Segmenter [37] and the groundtruth mask. The results on \\(512 \\times 512\\) images in Table 2 show that MAT achieves the highest inpainting performance with the mask detected by our method, which is consistent with the performance comparison of corruption detection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Generalizability on unseen corruption patterns. To evaluate the generalizability of our model methods across different corruption patterns, we perform training and test over" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12081" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.178, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.234, + 0.146, + 0.243 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.091, + 0.271, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.234, + 0.243, + 0.242 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.091, + 0.363, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.234, + 0.331, + 0.242 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.091, + 0.455, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.234, + 0.44, + 0.243 + ], + "angle": 0, + "content": "Groundtruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.243, + 0.468, + 0.273 + ], + "angle": 0, + "content": "Figure 5: Visual results of blind image inpainting on \\(512 \\times 512\\) resolution images from the Places [48] dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.287, + 0.469, + 0.332 + ], + "angle": 0, + "content": "Table 2: Blind image inpainting based on detected corruption masks by different methods on \\(512 \\times 512\\) images from Places. MAT [27] is finetuned (MAT-F) for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.333, + 0.442, + 0.418 + ], + "angle": 0, + "content": "
PhaseCorruption DetectionImage Inpainting
Corruption DetectionImage InpaintingACC ↑IoU ↑PSNR ↑SSIM ↑
GroundtruthMAT-F--25.340.852
Segmenter [37]MAT-F0.9740.96223.450.830
VCNet [42]MAT-F0.9750.96323.250.829
OursMAT-F0.9800.97024.530.844
VCNetVCNet0.9750.96321.490.764
OursOurs0.9800.97025.280.846
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.533 + ], + "angle": 0, + "content": "different corruption patterns. Table 4 presents the results of our model as well as VCNet by training them on Places dataset while performing testing directly on unseen corruption patterns such as random/constant noise and image content from CelebA-HQ [22]. It shows that our model consistently outperforms VCNet in all cases, which reveals the better generalizability of our model over VCNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.405, + 0.556 + ], + "angle": 0, + "content": "4.3. Downstream Image Restoration Tasks" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.563, + 0.295, + 0.579 + ], + "angle": 0, + "content": "4.3.1 Blind Image Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.78 + ], + "angle": 0, + "content": "In the experiments of blind image inpainting, we compare our model with VCNet and MPRNet [45], a state-of-the-art approach for image restoration. Table 3 presents the inpainting performance on images of \\(256 \\times 256\\) resolution. Our method achieves the best performance in terms of all metrics and outperforms the other two methods by a large margin. Besides, the qualitative comparison in Figure 4 also demonstrates that our model is able to restore higher-quality images than other two methods, benefiting from the precise corruption detection by the proposed hierarchical contrastive learning mechanism as well as the integrated restoration framework. Note that we also provide the user study in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.902 + ], + "angle": 0, + "content": "High-resolution blind image inpainting. We further conduct experiments on images of \\(512 \\times 512\\) resolution in Places [48] dataset for blind image inpainting. As shown in Table 2, our model outperforms VCNet substantially, consistent with the comparison in the case of \\(256 \\times 256\\) resolution. Besides, the qualitative comparison in Figure 5 also validates the superiority of our model. More visual results are presented in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.659, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.198, + 0.601, + 0.205 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.09, + 0.728, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.143, + 0.728, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.198, + 0.706, + 0.205 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.09, + 0.8, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.143, + 0.798, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.198, + 0.776, + 0.205 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.09, + 0.871, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.143, + 0.87, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.198, + 0.857, + 0.205 + ], + "angle": 0, + "content": "Groundtruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.207, + 0.891, + 0.251 + ], + "angle": 0, + "content": "Figure 6: Visual results of bidirectional blind image inpainting on FFHQ [23]. More results are shown in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.253, + 0.613, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.389, + 0.581, + 0.397 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.253, + 0.701, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.389, + 0.674, + 0.397 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.253, + 0.79, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.389, + 0.758, + 0.397 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.253, + 0.879, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.389, + 0.862, + 0.397 + ], + "angle": 0, + "content": "Groundtruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.405, + 0.891, + 0.451 + ], + "angle": 0, + "content": "Figure 7: Visual comparison of both corruption detection and blind image inpainting over unseen corruptions. More examples could be found in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.461, + 0.891, + 0.504 + ], + "angle": 0, + "content": "Table 3: Performance comparison for blind image inpainting in terms of four evaluation metrics on three benchmark datasets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.51, + 0.891, + 0.657 + ], + "angle": 0, + "content": "
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-2020-4040-600-2020-4040-600-2020-4040-60
PSNR ↑MPRNet33.9326.7522.0633.7626.2621.5033.3725.5921.08
VCNet29.8624.5120.5129.1323.3919.4229.6623.1919.18
Ours34.7927.9823.5433.9126.4021.8034.1126.2021.66
SSIM ↑MPRNet0.9680.8980.7990.9650.8820.7620.9630.8720.743
VCNet0.9330.8490.7360.9240.8180.6780.9420.8330.688
Ours0.9710.9100.8240.9660.8840.7670.9690.8870.746
FID ↓MPRNet3.3712.2029.743.7117.8053.384.1517.6144.73
VCNet6.8014.1728.905.6921.4859.115.9318.6240.21
Ours2.377.2815.482.6912.3541.372.8811.1031.33
LPIPS ↓MPRNet0.0270.0960.2120.0330.1230.2690.0330.1310.283
VCNet0.0440.1030.1880.0450.1280.2460.0460.1380.261
Ours0.0190.0650.1390.0250.0980.2260.0240.0950.221
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Bidirectional blind image inpainting. When real natural images are used as the noise content for corruption, we can perform bidirectional image restoration: restore either a complete noise image or a complete background image by reversing the corruption mask. It is quite challenging to perform well in inpainting of both directions. The visual results in Figure 6 show that our model performs more robust than VCNet in both corruption detection and image inpainting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Generalization on unseen corruption patterns. Using the same experimental settings as evaluating the generalizability for corruption detection, we also validate the generalizability of our model across unseen corruption patterns on image inpainting. Table 5 shows the image inpainting performance of both our model and VCNet on unseen corruption patterns during training. The experimental results indi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12082" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.098, + 0.335, + 0.141 + ], + "angle": 0, + "content": "Table 4: Generalizability over unseen corruptions for corruption detection. Models are trained on Places." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.143, + 0.334, + 0.259 + ], + "angle": 0, + "content": "
Corruption patternRandom constantCelebA-HQ [22]
Mask ratio (%)0-3030-600-3030-60
Acc ↑VCNet0.9810.9770.9750.974
Ours0.9900.9890.9860.986
F1 ↑VCNet0.9880.9780.9840.976
Ours0.9940.9900.9920.988
BCE ↓VCNet0.0500.0780.0630.068
Ours0.0500.0490.0620.054
IoU ↑VCNet0.9770.9600.9690.954
Ours0.9880.9810.9830.976
" + }, + { + "type": "table_caption", + "bbox": [ + 0.34, + 0.101, + 0.654, + 0.144 + ], + "angle": 0, + "content": "Table 5: Generalizability over unseen corruptions for blind inpainting. Models are trained on Places." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.148, + 0.642, + 0.261 + ], + "angle": 0, + "content": "
Corruption PatternRandom constantCelebA-HQ [22]
Mask ratio (%)0-2020-4040-600-2020-4040-60
PSNR ↑VCNet30.4824.3319.3029.7723.2118.96
Ours37.6929.0123.0534.6026.5221.82
SSIM ↑VCNet0.9570.8760.7320.9470.8400.685
Ours0.9850.9330.8180.9720.8940.772
FID ↓VCNet6.4116.8442.545.4717.7541.19
Ours1.807.4524.482.6510.5030.42
LPIPS ↓VCNet0.0510.1310.2680.0410.1280.256
Ours0.0130.0630.1820.0210.0890.214
" + }, + { + "type": "table_caption", + "bbox": [ + 0.658, + 0.1, + 0.887, + 0.13 + ], + "angle": 0, + "content": "Table 6: Ablation study on hierarchical interaction mechanism." + }, + { + "type": "table", + "bbox": [ + 0.664, + 0.137, + 0.887, + 0.267 + ], + "angle": 0, + "content": "
Mask ratio (%)0-3030-60
Acc ↑w/o inter-stage consistency0.9690.965
w/o sample selection0.9750.974
Complete model0.9780.976
F1 ↑w/o inter-stage consistency0.9800.969
w/o sample selection0.9840.976
Complete model0.9860.978
BCE ↓w/o inter-stage consistency0.1020.115
w/o sample selection0.0940.101
Complete model0.0900.095
IoU ↑w/o inter-stage consistency0.9620.940
w/o sample selection0.9720.957
Complete model0.9750.959
" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.275, + 0.282, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.275, + 0.486, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.387, + 0.466, + 0.403 + ], + "angle": 0, + "content": "Figure 8: Hierarchical vs single-stage contrastive learning." + }, + { + "type": "table_caption", + "bbox": [ + 0.087, + 0.417, + 0.458, + 0.43 + ], + "angle": 0, + "content": "Table 7: Performance evaluation on watermark removal." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.433, + 0.383, + 0.521 + ], + "angle": 0, + "content": "
MetricsPSNR↑SSIM↑FID↓
SIRF [46]34.630.9780.071
BS2AM [6]34.880.9790.028
DHAN [8]37.670.9860.062
BVMR [19]38.280.9850.018
Split then Refine [7]41.270.9910.011
VCNet [42]32.660.9630.032
Ours41.880.9920.007
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.595 + ], + "angle": 0, + "content": "cate that our model performs much better than VCNet. Furthermore, the qualitative comparison in Figure 7 also validates such better generalizability of our model over VCNet on both corruption detection and image inpainting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.598, + 0.327, + 0.613 + ], + "angle": 0, + "content": "4.3.2 Image Watermark Removal" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.752 + ], + "angle": 0, + "content": "In this set of experiments, we further evaluate the performance of our model on LOGO-30K dataset [7] for watermark removal. The comparison in Table 7 shows that our model performs best in terms of all metrics and compares favorably with other specialized methods for watermark removal, which demonstrates the robustness of our model across different image restoration tasks. Moreover, the visual results in Figure 9 also validate the superiority of our model over other methods in watermark removal." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.231, + 0.773 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Effect of hierarchical contrastive learning. To investigate the effectiveness of the proposed hierarchical contrastive learning framework, we compare its performance with that of single-stage contrastive learning. The results in Figure 8 demonstrate the distinct advantage of the proposed hierarchical contrastive learning framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effect of hierarchical interaction mechanism. We further conduct experiments to investigate the effect of hierarchi" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.277, + 0.58, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.277, + 0.657, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.277, + 0.735, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.277, + 0.814, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.277, + 0.892, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.339, + 0.58, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.399, + 0.553, + 0.406 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.339, + 0.658, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.399, + 0.634, + 0.406 + ], + "angle": 0, + "content": "VCNet" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.339, + 0.735, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.399, + 0.73, + 0.406 + ], + "angle": 0, + "content": "Split then Refine" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.339, + 0.814, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.399, + 0.785, + 0.406 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.339, + 0.892, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.399, + 0.876, + 0.406 + ], + "angle": 0, + "content": "Groundtruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.409, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Figure 9: Visual comparison with the state-of-the-art model Split then Refine [7] and VCNet for watermark removal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.556 + ], + "angle": 0, + "content": "cal interaction mechanism, especially the proposed 'inter-stage semantic consistency' and 'high-quality sample selection strategy'. Table 6 presents the concrete experimental results. While both techniques can boost the performance, 'inter-stage semantic consistency' is more crucial to the performance due to its essential merit: semantic propagation between stages by feature reusing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.619, + 0.58 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.722 + ], + "angle": 0, + "content": "In this work, we have designed a novel method, namely hierarchical contrastive learning framework, which can automatically detect corruption masks in pixel level and thus allows for blind image corruption restoration without known corruption masks. Extensive quantitative and qualitative comparisons have demonstrated the superior performance of our method over other methods for various corruption restoration tasks and its well generalization ability across different corruption patterns." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.731, + 0.681, + 0.747 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the NSFC fund (Grant No. U2013210, 62006060, 62176077), in part by the Shenzhen Key Technical Project under Grant 2022N001, 2020N046, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant (Grant No. 2022A1515010306), in part by Shenzhen Fundamental Research Program (Grant No. JCYJ20210324132210025, JCYJ20220818102415032), and in part by the Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies (Grant No. 2022B1212010005)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12083" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.212 + ], + "angle": 0, + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Nian Cai, Zhenghang Su, Zhineng Lin, Han Wang, Zhijing Yang, and Bingo Wing-Kuen Ling. Blind inpainting using the fully convolutional neural network. The Visual Computer, 33(2):249-261, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[5] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9):1200-1212, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Xiaodong Cun and Chi-Man Pun. Split then refine: stacked attention-guided resunets for blind single image visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1184-1192, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.496, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10680-10687, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11368, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[12] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, pages 341-346, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[13] Xin Feng, Wenjie Pei, Fengjun Li, Fanglin Chen, David Zhang, and Guangming Lu. Generative memory-guided semantic reasoning model for image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[15] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14134-14143, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[19] Amir Hertz, Sharon Fogel, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Blind visual motif removal from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6858–6867, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[20] Yong Shi Jie Yang, Zhiquan Qi. Learning to incorporate structure knowledge for image inpainting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12605-12612, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[21] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[26] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[27] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10758-10768, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12084" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision, pages 85-100, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[30] Hongyu Liu, Bin Jiang, Yibing Song, Wei Huang, and Chao Yang. Rethinking image inpainting via a mutual encoder-decoder with feature equalizations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 725-741. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[31] Hongyu Liu, Bin Jiang, Yi Xiao, and Chao Yang. Coherent semantic attention for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4170-4179, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[32] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11347-11357, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.439 + ], + "angle": 0, + "content": "[33] Yang Liu, Jinshan Pan, and Zhixun Su. Deep blind image inpainting. In International Conference on Intelligent Science and Big Data Engineering, pages 128-141. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[35] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Qureshi, and Mehran Ebrahimi. Edgeconnect: Structure guided image inpainting using edge prediction. In The IEEE International Conference on Computer Vision Workshops, Oct 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.471, + 0.664 + ], + "angle": 0, + "content": "[37] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.79 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[40] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[41] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Yi Wang, Ying-Cong Chen, Xin Tao, and Jiaya Jia. Vcnet: A robust approach to blind image inpainting. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[43] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[44] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4471-4480, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[46] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[47] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1438-1447, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[48] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12085" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_origin.pdf b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a68f4da59796cb595c4f2166191349687fdcce2c --- /dev/null +++ b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/2454a4c4-9bf9-4a9e-a255-d46d7a04e834_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83d5afa6cea827b133bcffb827ebcd0ffe02d41126a969779abb016b82c9c7df +size 2222296 diff --git a/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/full.md b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f41071ae11f374a4b06428cea56fd82e7d70dbde --- /dev/null +++ b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/full.md @@ -0,0 +1,482 @@ +# Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection + +Xin Feng* + +Yifeng Xu* + +Guangming Lu† + +Wenjie Pei† + +Harbin Institute of Technology, Shenzhen + +![](images/14a709fb49e13fc575435c9f8ae16c57f9a8338dbc3c441f616e33f4835d67d8.jpg) + +![](images/36fb8097904e9b486e760364b2a8dd2169ea7921be8a147c96505e13ceb9c398.jpg) +Input + +![](images/94b692096424a9137c9ff1a6050125e20670effcdb55690b63e55c7a56205002.jpg) + +![](images/ccf867ee4f448548b8d8b36a9f98b17b33b178e9c0d1817d26c0ee95edad515b.jpg) +VCNet + +![](images/8894c44bc6af25eb6213ab01ea5b2c39507db246608c795b5d8d8917621302f5.jpg) + +![](images/fda8e8d2ea298685ef5762a33b012c69589f778e7b35cebec77b82002db8a222.jpg) +Ours +Figure 1: Visual comparison between our method and VCNet [42] (the state-of-the-art method for blind image inpainting) on four tasks: blind image inpainting (graffiti corruption) (up left), blind image inpainting (real-image-patch corruption) (up right), watermark removal (bottom left) and shadow removal (bottom right). The groundtruth mask and the predicted masks by different methods are presented. + +![](images/ad299114892c69d623d37727edc8bcf48a0ea7f94b09a82ffca58f2b48ea36fc.jpg) + +![](images/c396922ecf11dfae74e3046b6f4812f7553a1d22403ad862b4e6b45eec992e72.jpg) +Input + +![](images/983e989bbd6b5053d53f2fee356fc5bfbc78d344817e1e2f674417bfdf479ac7.jpg) + +![](images/7358596603db67f7455043ac59e055272d13f06e24c40305884a24d735931bba.jpg) +VCNet + +![](images/26d73c590f3ac024c68a1917bbd400650446767fbcc06fab4d2b8c7ce99448c0.jpg) + +![](images/41baf0baf8d92866bbee179ea956d4bca23f36571bf18ed8c1abe34d0fab093e.jpg) +Ours + +# Abstract + +Effective image restoration with large-size corruptions, such as blind image inpainting, entails precise detection of corruption region masks which remains extremely challenging due to diverse shapes and patterns of corruptions. In this work, we present a novel method for automatic corruption detection, which allows for blind corruption restoration without known corruption masks. Specifically, we develop a hierarchical contrastive learning framework to detect corrupted regions by capturing the intrinsic semantic distinctions between corrupted and uncorrupted regions. In particular, our model detects the corrupted mask in a coarse-to-fine manner by first predicting a coarse mask by contrastive learning in low-resolution feature space and then refines the uncertain area of the mask by high-resolution contrastive learning. A specialized hierarchical interaction mechanism is designed to facilitate the knowledge propagation of contrastive learning in different scales, boosting the modeling performance substantially. The detected multi-scale corruption masks are then leveraged to guide the corruption restoration. Detecting corrupted regions + +by learning the contrastive distinctions rather than the semantic patterns of corruptions, our model has well generalization ability across different corruption patterns. Extensive experiments demonstrate following merits of our model: 1) the superior performance over other methods on both corruption detection and various image restoration tasks including blind inpainting and watermark removal, and 2) strong generalization across different corruption patterns such as graffiti, random noise or other image content. Codes and trained weights are available at https://github.com/xyfJASON/HCL. + +# 1. Introduction + +An essential yet challenging step for image restoration with large-size corruptions, such as blind image inpainting, watermark removal or shadow removal, is to detect the corruption region masks in pixel level precisely. The difficulties lie in the diverse shapes and appearance of corruptions, which can be hardly modeled in a uniform pattern. As a result, blind image inpainting remains a challenging task although image inpainting with known corruption masks [27, 11, 13, 32] has achieved remarkable progress. + +A straightforward way for corruption detection is to for + +mulate it as a segmentation task and predict the corruption mask by pixel-wise binary classification. A prominent example following such modeling paradigm is VCNet [42], a state-of-the-art method for blind image inpainting, which regards corruptions as target objects and learns to recognize the semantic patterns of corruptions for detection. While VCNet can successfully detect the corruptions with uniform patterns, it has two potential limitations. First, the corruption may exhibit diverse patterns due to its potentially irregular appearance nature, which VCNet can hardly deal with. Second, it is challenging for VCNet to handle the corruption patterns that are distinct from those appearing in training data, thus it has limited generalization w.r.t. different corruption patterns. + +In this work, we propose a novel model for image corruption detection to circumvent the aforementioned limitations of VCNet. Instead of recognizing the semantic patterns of corruptions as VCNet does, we apply metric learning to learn an embedding space in which our model can capture the contrastive semantic distinctions between the corrupted and uncorrupted regions. To this end, we design a hierarchical contrastive learning framework for detecting corrupted regions, which predicts the corruption mask in a coarse-to-fine manner. To be specific, it first predicts a coarse mask by lightweight contrastive learning in a low-resolution feature space. Then it refines the uncertain pixels with low confidence in the predicted mask by high-resolution contrastive learning based on fine-grained features. Note that only a small fraction of predicted masks need to be refined, thus the refining process can be performed quite efficiently. To facilitate the knowledge propagation and consistency of contrastive learning between different scales, we propose a specialized hierarchical interaction mechanism. As a result, the high-resolution contrastive learning can inherit useful knowledge from the low-resolution contrastive learning to achieve precise prediction of corruption masks quite efficiently. The detected multi-scale corruption masks are further used to guide the restoration of the corrupted regions, which also follows a coarse-to-fine generative process. + +Unlike VCNet recognizing the semantic patterns of corruptions, our model focuses on capturing the contrastive semantic distinctions between corrupted and uncorrupted regions. Thus our model has well generalization ability across different corruption patterns. As shown in Figure 1. To conclude, we make following contributions. + +- We design a hierarchical contrastive learning framework to detect multi-scale corruption masks by learning the contrastive semantic distinctions between corrupted and uncorrupted regions. +- Integrating the proposed hierarchical contrastive learning module, we develop a general-purpose blind image restoration model to perform high-quality image restoration without known corruption masks. + +- Extensive experiments validate the effectiveness of our model both quantitatively and qualitatively in three aspects: 1) superior performance on corruption detection (Section 4.2), 2) favorable performance compared to the specialized methods on two challenging image restoration tasks including blind image inpainting and watermark removal (Section 4.3), and 3) strong generalizability across different corruption patterns (Section 4.2 and 4.3.1). + +# 2. Related Work + +Image Restoration with known Corruption masks. Image restoration with given corruption masks assumes that the corrupted regions are known, and conventional methods [2, 12, 5] leverage uncorrupted content to restore the corrupted regions by pixel diffusion or patch replacement. However, these methods usually generate significant artifacts in the restored image due to the lack of global semantic understanding and generalization ability. With the great success of convolutional neural networks(CNNs) in computer vision, recent methods leverage an encoder-decoder framework to restore corrupted images. A typical way of CNN-based methods [29, 44, 26, 27] employs known uncorrupted regions to infer corrupted content from the outside to the inside. To synthesize reasonable semantics, some methods [31, 35, 20, 30, 16, 28] leverage structure information to improve the quality of semantic structures in the corrupted regions. Promoted by generative models [14, 25], some recent methods [11, 32, 47] attempt to learn generative prior for improving the quality of synthesized images, and generating diverse content for corrupted regions. + +Image corruption detection. Early methods [3, 33] for image corruption detection assume the corrupted content follows a simple distribution, such as constant values, Gaussian noise, etc. However, previous assumption simplifies the detection of corrupted regions, limiting the application scope of image corruption detection. Recently, Wang et al. [42] relax the definition of image corruption detection by increasing the diversity of corruption types, including watermarking, raindrop, or even random images. They also propose a two-stage framework VCNet for blind image restoration, which first locates corrupted regions in a manner of image segmentation and then fills in content following the typical encoder-decoder based methods for image restoration. However, it sometimes misuses content in corrupted regions while reconstructing the complete image due to the challenge of detecting diverse patterns of corruption. + +Contrastive learning. Contrastive learning has been widely used in self-supervised representation learning [10, 17, 4]. Instead of matching an input image to a fixed target, contrastive learning maximizes mutual information in the representation space, keeping the query close to the positive sample and away from the negative sample. Previous works [17, 4, 41] have validated the effectiveness of contrastive + +![](images/b2238f8dce237565dd153849e69bff804fa5b767dae51ad0643a14626260294f.jpg) +Figure 2: Architecture of our method. It leverages the designed Hierarchical Contrastive Learning to detect corruption masks and further synthesizes reasonable content for corrupted regions by the developed Transformer-based Restoration module. + +learning in high-level vision tasks, due to the inherent suitability for modeling feature contrasts. Recently, some researchers [40, 43] have attempted to introduce contrastive learning into low-level vision tasks. However, pixel-wise contrastive learning on high-resolution images suffers from large computational cost and limited performance. In this paper, we design a hierarchical contrastive learning mechanism to solve above problems and significantly strengthen the performance of image corruption detection. + +# 3. Approach + +Detecting image corruptions by recognizing the semantic patterns of corruptions is difficult in that the corruptions can potentially exhibit diverse patterns. Our proposed hierarchical contrastive learning framework learns the contrastive semantic distinctions between corrupted and uncorrupted regions, and detects corruption masks in a coarse-to-fine manner. The obtained multi-scale corruption masks are then leveraged to guide the restoration of the corruptions in a coarse-to-fine generative process. + +# 3.1. Overview + +Problem formulation. Given an input image $\mathbf{I} \in \mathbb{R}^{C \times H \times W}$ which is corrupted from an intact groundtruth image $\mathbf{O}_{\mathrm{gt}} \in \mathbb{R}^{C \times H \times W}$ , the task of blind image restoration aims to first detect the corruption region mask $\mathbf{M} \in \mathbb{R}^{1 \times H \times W}$ and then reconstruct a complete image $\hat{\mathbf{O}} \in \mathbb{R}^{C \times H \times W}$ by synthesizing realistic content $\hat{\mathbf{C}}$ for the corrupted regions: + +$$ +\begin{array}{l} \mathbf {I} = \mathbf {O} _ {\mathrm {g t}} \odot \mathbf {M} + \mathbf {N} \odot (1 - \mathbf {M}), \\ \hat {\mathbf {O}} = \mathbf {O} _ {\mathrm {g t}} \odot \mathbf {M} + \hat {\mathbf {C}} \odot (1 - \mathbf {M}). \tag {1} \\ \end{array} +$$ + +Herein, $\mathbf{M} \in \mathbb{R}^{1 \times H \times W}$ is a binary map where the values of + +corrupted regions are 0 and uncorrupted regions are filled with $1.\mathbf{N}\in \mathbb{R}^{C\times H\times W}$ is the noisy content in the corrupted regions. Precise detection of the corruption mask $\mathbf{M}$ is crucial to the performance of image restoration. + +As shown in Figure 2, our model follows the encoder-decoder framework and consists of two core modules: Hierarchical Contrastive Learning module for detecting corruption masks and Transformer-based Restoration module for corruption restoration. The corrupted image is first encoded into multi-scale feature maps by Encoder $\mathcal{F}_{\mathrm{enc}}$ of Transformer-based Synthesis module. Then the proposed Hierarchical Contrastive Learning module is employed to perform corruption detection in a coarse-to-fine manner from these feature maps, and predicts multi-scale corruption masks. Finally, the obtained masks are fed into Decoder $\mathcal{F}_{\mathrm{Dec}}$ of the Transformer-based Synthesis module for restoration, which is also in a coarse-to-fine generative process. Both Encoder $\mathcal{F}_{\mathrm{enc}}$ and Decoder $\mathcal{F}_{\mathrm{dec}}$ are mainly composed of basic transformer blocks of MAT [39], while they are equipped with additional down-sampling layers and upsampling layers respectively. We employ a shallow convolutional layer with $5\times 5$ kernel before Encoder and after Decoder to perform basic feature transformation. Besides, we also employ another Conv-U-Net to refine high-frequency details of output results, leaning upon the local texture refinement capability and efficiency of CNNs. + +# 3.2. Corruption Detection by Hierarchical Contrastive Learning + +We design Hierarchical Contrastive Learning framework to guide the learning of multi-scale semantic embedding spaces of Encoder $\mathcal{F}_{\mathrm{enc}}$ in such a way that the semantic dis + +tances between two arbitrary pixels both within uncorrupted or corrupted regions (intra-region distance) should be minimized while the distance between a corrupted pixel and an uncorrupted pixel (inter-region distance) should be maximized. As a result, our model can capture the intrinsic semantic distinctions between corrupted and uncorrupted regions. Then our model performs clustering on all pixels in this learned embedding space to separate them into two clusters, corresponding to the corrupted and uncorrupted regions, and thus achieves the corruption mask. + +As shown in Figure 2, Encoder $\mathcal{F}_{\mathrm{enc}}$ consists of three encoding stages and produces three scales of encoded feature maps. Accordingly, Our model performs three stages of pixel-level contrastive learning for the corresponding scale of feature maps to guide the learning of its embedding space. We will first describe how our model performs corruption detection in one stage with single-scale contrastive learning. Then we will elaborate on the proposed Hierarchical Interaction Mechanism which enables our model to perform hierarchical contrastive learning quite efficiently for coarse-to-fine multi-scale mask detection. + +# 3.2.1 Single-Scale Contrastive Learning + +Supervised Contrastive Learning. During each stage of contrastive learning, we construct positive pixel pairs by using intra-region pixels, i.e., both pixels are from either the uncorrupted region or the corrupted region. In contrast, each negative training pair consists of two inter-region pixels, one from the corrupted region and the other from the uncorrupted region. To be specific, for a query pixel $q$ from a randomly selected query set $Q$ , we construct the positive set $P$ by randomly selecting pixels from the same region as $q$ and construct the negative set $N$ from the opposite region. Then we apply Circle loss [38] to maximize the cosine similarity of positive pairs while minimizing the similarities of negative pairs. Formally, the contrastive learning in the stage $S_{s}$ is guided by the loss: + +$$ +\mathcal {L} _ {\mathrm {C L}} ^ {s} = \sum_ {q \in Q} \log \left[ 1 + \sum_ {p \in P} \exp (- \mathbf {e} _ {q} \cdot \mathbf {e} _ {p} / \tau) \cdot \sum_ {n \in N} \exp (\mathbf {e} _ {q} \cdot \mathbf {e} _ {n} / \tau) \right], \tag {2} +$$ + +where $\tau$ is a scale factor. $\mathbf{e}_q$ is the features for pixel $q$ projected from the corresponding embedding space of Encoder $\mathcal{F}_{\mathrm{enc}}^s$ by a projection head in Stage- $s$ .. + +$$ +\mathbf {e} _ {q} = \mathcal {F} _ {\operatorname {p r o j}} ^ {s} \left(\mathcal {F} _ {\mathrm {e n c}} ^ {s} \left(\mathbf {I} _ {q}\right)\right), \tag {3} +$$ + +where the projection head $\mathcal{F}_{\mathrm{proj}}$ comprises two fully connected layers and a GELU [18] layer in-between for nonlinear transformation. + +Corruption Mask Detection. Under the supervision of contrastive learning in Equation 2, the pixels within the same region, either the corrupted or the uncorrupted regions, tend to have similar representations in each scale of embedding space of Encoder $\mathcal{F}_{\mathrm{enc}}$ while the pixels from different regions would have dissimilar representations. Thus, + +![](images/3e1f73fb7e08b0078a8017fc0fae87110f948e0df96defb3b4a1ea1fc8b6c507.jpg) +Figure 3: In the stages $\mathrm{S}_2$ and $\mathrm{S}_3$ , both the selected representative samples and uncertain samples in the previous stage are used for contrastive learning in current stage while only the mask labels of uncertain samples are refined during mask detection. + +![](images/439466aac29665d9c911bcad7c57cf98168eab0f6224ef8d9081cc2cf4490c2c.jpg) + +we can perform clustering to separate them into two clusters. Specifically, our model adopts K-means algorithm for clustering, and the clustering in Stage- $s$ is performed by: + +$$ +\left[ \mathbf {c} _ {1} ^ {s}, \mathbf {c} _ {2} ^ {s} \right], \left[ \mathbf {M} _ {1} ^ {s}, \mathbf {M} _ {2} ^ {s} \right] = \mathrm {K - m e a n s} \left(\mathcal {F} _ {\operatorname {p r o j}} ^ {s} \left(\mathcal {F} _ {\mathrm {e n c}} ^ {s} (\mathbf {I})\right)\right), \tag {4} +$$ + +where $[\mathbf{c}_1^s,\mathbf{c}_2^s ]$ denotes the embeddings of two produced cluster centers and $[\mathbf{M}_1^s,\mathbf{M}_2^s ]$ are the associated binary masks. To identify which cluster corresponds to the corrupted or uncorrupted regions, we train a lightweight binary classifier consisting of two fully connected layers with a ReLU layer, and use it to perform classification on two cluster centers. Thus, the associated mask for the cluster of the corrupted region is the predicted corruption mask. + +# 3.2.2 Hierarchical Interaction Mechanism + +Our model performs Hierarchical Contrastive Learning including three stages to detect the corruption mask in a coarse-to-fine manner. It first performs contrastive learning in the lowest-resolution stage $(\mathrm{S}_1)$ to predict a coarse mask for corruption. Then it refines the uncertain pixels of the mask with low confidence by higher-resolution contrastive learning in subsequent stages $(\mathrm{S}_2$ and $\mathrm{S}_3)$ . We propose Hierarchical Interaction Mechanism to facilitate the interaction and knowledge propagation between different stages of contrastive learning and guarantee the semantic consistency between them. In particular, higher-resolution contrastive learning can inherit useful knowledge from previous stages of contrastive learning, which can improve the learning performance substantially. Quadtree structure are used to build the positional correspondence between feature maps in adjacent stages considering that the resolution of feature maps is always scaled by four times between adjacent stages. + +Selecting High-Quality Training Samples. During the contrastive learning in $S_1$ , we construct the query set $Q$ as well as the associated positive set $P$ and negative set $N$ + +in Equation 2 by randomly selecting pixels from the whole feature map. To improve the efficiency of contrastive learning in the higher stages $(\mathrm{S}_2$ and $\mathrm{S}_3)$ , we select a small fraction of pixels from the entire feature maps, which are crucial for contrastive learning, as the high-quality candidate set for constructing $Q$ , $P$ and $N$ . + +To be specific, we first measure the prediction confidence of each pixel based on the clustering results of last stage. The prediction confidence of the pixel $q$ in the stage $S_{s}$ is calculated by: + +$$ +z _ {q} ^ {s} = \frac {\exp \left(- \left(\mathbf {e} _ {q} \cdot \mathbf {c} _ {y _ {q}} ^ {s} / \tau\right)\right)}{\sum_ {i = 1} ^ {2} \exp \left(- \left(\mathbf {e} _ {q} \cdot \mathbf {c} _ {i} ^ {s} / \tau\right)\right)}, \tag {5} +$$ + +where $y_{q}$ denotes the cluster index $q$ belongs to and $\tau$ is the scaling factor. Then we pick out two types of pixels as the high-quality query set for the stage $S_{s+1}$ : 1) the pixels with high confidence which are regarded as representatives of two clusters and are typically close to the cluster centers; 2) the uncertain pixels with low confidence which are mostly near the boundaries between the corrupted and uncorrupted regions. Selecting pixels in such a way for constructing the positive and negative training pairs, our model is able to perform high-resolution contrastive learning in the stages $S_{2}$ and $S_{3}$ quite efficiently and effectively. + +Inter-Stage Semantic Consistency. To guarantee the semantic consistency between different stages of contrastive learning, we reuse the learned features of the previous stage of contrastive learning in current stage. As shown in Figure 3, we concatenate the features in the previous stage with the features in the current stage, which are then fed into the projection head to produce input features for current stage. Accordingly, Equation 3 for high-resolution stages $(S_{2}$ and $S_{3})$ evolves into the following form: + +$$ +\mathbf {e} _ {q} = \mathcal {F} _ {\text {p r o j}} ^ {s} \left(\operatorname {C o n c a t} \left(\mathcal {F} _ {\text {e n c}} ^ {s} \left(\mathbf {I} _ {q}\right), \mathcal {F} _ {\text {m a p}} ^ {s} \left(\mathcal {F} _ {\text {e n c}} ^ {s - 1} \left(\mathbf {I} _ {q}\right)\right)\right)\right), \tag {6} +$$ + +where the mapping function $\mathcal{F}_{\mathrm{map}}^s$ is a linear transformation to regulate the feature dimension of previous stage and balance the effect between current and previous features. Besides the semantic consistency between adjacent stages, another merit of such feature reusing is that our model is able to inherit the learned semantics from previous stage. + +Refining the Mask Prediction of Uncertain Pixels. During the high-resolution contrastive learning in the stages $S_2$ and $S_3$ , we only refine the mask prediction by re-predicting the labels of uncertain pixels with low confidence in the previous stage, i.e., the type-2 samples of selected high-quality query set. The mask labels of other pixels with high confidence are directly inherited from the predictions in the previous stage according to the built quadtree, assuming that the predictions of these pixels are reliable. Thus, such refining process can be performed quite efficiently. + +# 3.3. Transformer-based Restoration Module + +As shown in Figure 2, the detected corruption masks are fed into Decoder of the Transformer-based Restoration + +module for corruption restoration, which consists of the designed mask-guided transformer block. + +Mask-guided transformer block. As illustrated in Figure 2, our transformer block contains a multi-head contextual attention (MCA) module, followed by a two-layer MLP with GELU nonlinearity in between. Besides, a layer normalization (LN) [1] layer is applied before each attention and each MLP, and a residual learning connection is employed after each module: + +$$ +F _ {i} ^ {\prime s} = \operatorname {M C A} \left(\operatorname {L N} \left(F _ {i - 1} ^ {s}\right), \mathbf {M} ^ {s}\right) + F _ {i - 1} ^ {s}, \tag {7} +$$ + +$$ +F _ {i} ^ {s} = \operatorname {M L P} \left(\operatorname {L N} \left(F _ {i} ^ {\prime s}\right)\right) + F _ {i} ^ {\prime s}, +$$ + +where $F_{i}^{s}$ is the output features of the $i$ -th block in the stage $S_{s}$ . To fully exploit the predicted masks by hierarchical contrastive learning, we employ the multi-head contextual attention module proposed in MAT [27], which follows the shifted window manner [34] and is formulated as: + +$$ +\operatorname {A t t} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {\mathrm {T}} + \hat {\mathbf {M}} ^ {s}}{\sqrt {d _ {k}}}\right) \mathbf {V}, \tag {8} +$$ + +where $\mathbf{Q},\mathbf{K},\mathbf{V}$ are the query, key, and value matrices, respectively. $\sqrt{d_k}$ is the scaling factor. $\hat{\mathbf{M}}$ is defined by + +$$ +\mathbf {\hat {M}} _ {i} ^ {s} = \gamma \left(\mathbf {M} _ {i} ^ {s} - 1\right), \tag {9} +$$ + +where $\gamma$ is a large positive integer to reduce the values of corrupted pixels. + +# 3.4. Joint Optimization for Parameter Learning + +The proposed Hierarchical Contrastive Learning module and the Restoration module are integrated in a holistic way based on the encoding-decoding framework, forming a general-purpose blind image restoration model, as illustrated in Figure 2. The whole model is optimized jointly. Besides the loss of hierarchical contrastive learning (Equation 2), we also adopt three more supervision signals for corruption restoration, including pixel-wise reconstruction loss, perceptual loss, and adversarial loss. + +Pixel Reconstruction Loss, which pushes the restored image $\hat{\mathbf{O}}$ as close as its groundtruth $\mathbf{O}_{\mathrm{gt}}$ : + +$$ +\mathcal {L} _ {\text {p i x e l}} = \left\| \hat {\mathbf {O}} - \mathbf {O} _ {\mathrm {g t}} \right\| _ {1}. \tag {10} +$$ + +Perceptual Loss [21], which performs semantic supervision on the restored images in the deep feature space: + +$$ +\mathcal {L} _ {\text {p e r c}} = \sum_ {l = 1} ^ {L} \frac {1}{C _ {l} H _ {l} W _ {l}} \left\| f _ {\mathrm {v g g}} ^ {l} (\hat {\mathbf {O}}) - f _ {\mathrm {v g g}} ^ {l} \left(\mathbf {O} _ {\mathrm {g t}}\right) \right\| _ {1}. \tag {11} +$$ + +Herein $f_{\mathrm{vgg}}^{l}(\hat{\mathbf{O}})$ and $f_{\mathrm{vgg}}^{l}(\mathbf{O}_{\mathrm{gt}})$ are the extracted features from $\hat{\mathbf{O}}$ and $\mathbf{O}_{\mathrm{gt}}$ respectively from the $l$ -th convolution layer of the pre-trained VGG-19 network [36]. + +Adversarial Loss, which employs WGAN-GP [15] to encourage $\hat{\mathbf{O}}$ to be as realistic as its groundtruth $\mathbf{O}_{\mathrm{gt}}$ : + +$$ +\mathcal {L} _ {\mathrm {a d v}} = - \mathbb {E} _ {\hat {\mathbf {O}} \sim \mathbb {P} _ {\hat {\mathbf {O}}}} \left[ D (\hat {\mathbf {O}}) \right], \tag {12} +$$ + +where $D$ is a discriminator. Overall, the whole model is optimized by: + +![](images/825b180a91c3c7df6991f36072543670b022516ffc282562049e22cbd90147b5.jpg) + +![](images/8727dc59fdae44809fc2cf3f414f39706c1c3a7052ed7c2b2b09a4e9cea79df5.jpg) +Input + +![](images/682707e653889dfeaec8a7d28f4ded4f2e7993052a6fc86ea1f78b53204828a4.jpg) + +![](images/8b86f44ed78f713202c481fb0be6f3bbcfe5d400e94695dfc22ed4ee54172ba2.jpg) +VCNet + +![](images/f0e88a4bdcdf442acf0655e2c65c04c925fcd655fee2a13a2016daec75c725e4.jpg) + +![](images/a309e23a2cbe91ba931c5deb44b1b4a79e9f7b1af9ccea058ef6a6e8b64cb214.jpg) +Ours +Figure 4: Visualization of detected masks and reconstructed images on four randomly selected samples from test set. + +![](images/8bd36a76de3dedc97e4cbacc1e5f17e0d3727a710274ce372eee3779343dd1c4.jpg) + +![](images/b5901f0ad0cb9e25095f0378fa34de10fec2ffa6984cf21961cb92553bcb4988.jpg) +Groundtruth + +![](images/b754e9d7dddeb468db2035245c287d63896ceae9285eaf0ea617c877d37032c8.jpg) + +![](images/5e4a9c8ade344890d1503f697e7f34dd176c00669102fa031c621fa95801d4d3.jpg) +Input + +![](images/84106f35af3c94bc3014b5a90ba21772b56c5b732af6aae5ded4cb2adc7b76c1.jpg) + +![](images/4e1b3344027f886d9abff9321ed370839605f113d0bc614259de5d4ee8ff38be.jpg) +VCNet + +![](images/0f8a6f712357e38d6915de1f852fcf9422bd984be15549b1ad5315e056bc1e22.jpg) + +![](images/81649f551577d913c25202a036bb08c24db5dff183cb051c2239625036e6187f.jpg) +Ours + +![](images/d1ce3b6bb84acf4b98d4d16bddc062bc75751e26d82e1c2b6a7e5ba87188e81e.jpg) + +![](images/160c77a6319911d8461a1a81e94f17ee6383e61d50ee69c5b53abaf6d7f94620.jpg) +Groundtruth + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {\text {p i x e l}} + \lambda_ {2} \mathcal {L} _ {\text {p e r c}} + \lambda_ {3} \sum_ {s = 1} ^ {3} \mathcal {L} _ {\mathrm {C L}} ^ {\mathrm {s}} + \lambda_ {4} \mathcal {L} _ {\text {a d v}}, \tag {13} +$$ + +where $\lambda_1, \lambda_2, \lambda_3$ , and $\lambda_4$ are hyper-parameters to balance between different losses. + +We integrate the Contrastive Learning module between Encoder and Decoder with consistent coarse-to-fine refining stages for all three modules, which potentially enables synchronized optimization of three modules in each scale. + +# 4. Experiments + +We first evaluate the performance of our method on image corruption detection, then validates the effectiveness of our method on two challenging image restoration tasks including blind image inpainting and watermark removal. Finally, we conduct extensive ablation study to obtain more insights into our method. + +# 4.1. Experimental Settings + +Following the previous work [42] for blind image inpainting, we make two pre-processing operations for data generation: 1) we randomly select natural images from large-scale datasets rather than simple constant values or Gaussian noise as the noisy content for corruptions to increase the difficulty of mask detection in blind image inpainting; 2) we smooth the mask boundaries using alpha blending to avoid distinct boundaries between the corrupted and uncorrupted regions. + +Four large-scale datasets are used in our experiments, including FFHQ (faces) [23], CelebA-HQ (faces) [22], ImageNet (objects) [9], and Places (scenes) [48]. For each dataset, images from different datasets are randomly selected as the noise content for corruptions. For instance, the images from CelebA-HQ and ImageNet are used as noise content for FFHQ. We employ the method [29] to generate irregular masks with mask shapes and corruption ratios. We perform experiments on two resolutions of images for comprehensive evaluation: $256 \times 256$ and $512 \times 512$ . + +Adam [24] is used as the optimizer by setting $\beta_{1}$ , $\beta_{2}$ , initial learning rate and batch size to be 0.9, 0.999, 0.0001, and 4, respectively. More experimental details are provided in the supplementary material. + +Table 1: Performance comparison for mask detection on three benchmark datasets. + +
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-3030-600-3030-600-3030-60
Acc ↑VCNet0.9480.9430.9720.9780.9740.976
Ours0.9780.9760.9820.9830.9840.985
F1 ↑VCNet0.9670.9480.9820.9800.9840.978
Ours0.9860.9780.9890.9850.9910.987
BCE ↓VCNet0.1260.1370.0730.0550.0640.060
Ours0.0900.0950.0830.0750.0680.064
IoU ↑VCNet0.9310.9000.9660.9620.9690.959
Ours0.9750.9590.9780.9650.9820.974
+ +# 4.2. Image Corruption Detection + +Accuracy of corruption detection. We first compare the accuracy of corruption detection between our model and VCNet, the state-of-the-art method for corruption detection in blind image inpainting. Four metrics are used for comprehensive evaluation: binary cross entropy (BCE), classification accuracy (Acc), F1 score, and intersection over union (IoU). Table 1 lists the results for different corruption ratios on three datasets in $256 \times 256$ resolution, which show that our model achieves superior performance of mask detection over VCNet in terms of all metrics except BCE. This is reasonable because VCNet is directly supervised by the BCE loss. Besides, we also make qualitative comparison in Figure 4, in which our model detects more precise corruption masks and restores higher-quality images than VCNet. + +Image inpainting based on detected corruption masks by different methods. As an indirect evaluation of corruption detection, we perform image inpainting over the detected corruption masks from different methods, employing the same inpainting model MAT [27], and then compare the inpainting performance. To have a comprehensive evaluation, we also provide the inpainting performance over the detected masks from a state-of-the-art segmentation model Segmenter [37] and the groundtruth mask. The results on $512 \times 512$ images in Table 2 show that MAT achieves the highest inpainting performance with the mask detected by our method, which is consistent with the performance comparison of corruption detection. + +Generalizability on unseen corruption patterns. To evaluate the generalizability of our model methods across different corruption patterns, we perform training and test over + +![](images/c7db2094333dbd694e3187e960e41c373cdd19c63a9a79b4c9f11b863cb57703.jpg) +Input + +![](images/67f8a8b119b2fa73b73cd72dec6e62d86ec0cf7b63060019ab1edf71c232e922.jpg) +VCNet + +![](images/14f04c62e0aba61137c0a082b21ff28e4de6e2b4ca4a36c7ae7ac08ce632f3bb.jpg) +Ours + +![](images/cfbcc05cb16ffb4f5d35e02fb4fb08d9e278ff79d5207152211ed57894715c25.jpg) +Groundtruth +Figure 5: Visual results of blind image inpainting on $512 \times 512$ resolution images from the Places [48] dataset. + +Table 2: Blind image inpainting based on detected corruption masks by different methods on $512 \times 512$ images from Places. MAT [27] is finetuned (MAT-F) for fair comparison. + +
PhaseCorruption DetectionImage Inpainting
Corruption DetectionImage InpaintingACC ↑IoU ↑PSNR ↑SSIM ↑
GroundtruthMAT-F--25.340.852
Segmenter [37]MAT-F0.9740.96223.450.830
VCNet [42]MAT-F0.9750.96323.250.829
OursMAT-F0.9800.97024.530.844
VCNetVCNet0.9750.96321.490.764
OursOurs0.9800.97025.280.846
+ +different corruption patterns. Table 4 presents the results of our model as well as VCNet by training them on Places dataset while performing testing directly on unseen corruption patterns such as random/constant noise and image content from CelebA-HQ [22]. It shows that our model consistently outperforms VCNet in all cases, which reveals the better generalizability of our model over VCNet. + +# 4.3. Downstream Image Restoration Tasks + +# 4.3.1 Blind Image Inpainting + +In the experiments of blind image inpainting, we compare our model with VCNet and MPRNet [45], a state-of-the-art approach for image restoration. Table 3 presents the inpainting performance on images of $256 \times 256$ resolution. Our method achieves the best performance in terms of all metrics and outperforms the other two methods by a large margin. Besides, the qualitative comparison in Figure 4 also demonstrates that our model is able to restore higher-quality images than other two methods, benefiting from the precise corruption detection by the proposed hierarchical contrastive learning mechanism as well as the integrated restoration framework. Note that we also provide the user study in the supplementary material. + +High-resolution blind image inpainting. We further conduct experiments on images of $512 \times 512$ resolution in Places [48] dataset for blind image inpainting. As shown in Table 2, our model outperforms VCNet substantially, consistent with the comparison in the case of $256 \times 256$ resolution. Besides, the qualitative comparison in Figure 5 also validates the superiority of our model. More visual results are presented in the supplementary material. + +![](images/0a2e0d3278e574a8f0110a1165b8df986c110a917bab6ef44862ca5ede8e3ac2.jpg) +Input + +![](images/fc3fdb375c5dfb549612ba7218098f5f9fa8f7ff0ca88ddc8df3efd8bce30454.jpg) + +![](images/16a7579b21d5832bf851d6ff4cf78f4d36c41199721faba9f668a08809bd06c2.jpg) +VCNet + +![](images/3574c41333a965fce6418e4ee32a65807276fc01b2bdf39b2399dff2f3c988b4.jpg) + +![](images/f5e4f7e9b338d8e28c9412e022ca73c3d38f347ca86324dbc3b790d1a217e084.jpg) +Ours + +![](images/79b72508848d2cfd23c53d234ed133ccdc14fe160da6c573bdd8fece9de2c16b.jpg) + +![](images/3e7be72b3345aed32b636eaf132987eebbd02f46dbca43d081be3496bf963c8d.jpg) +Groundtruth + +![](images/c2b504c647ee84f25366488c465ffe521172168670b0ef9281410445b4bce304.jpg) +Figure 6: Visual results of bidirectional blind image inpainting on FFHQ [23]. More results are shown in the supplementary material. +Input +Figure 7: Visual comparison of both corruption detection and blind image inpainting over unseen corruptions. More examples could be found in the supplementary material. + +![](images/4a6aa392cf6a13f7b396fe25824a44fd4538a9f8d66321530f462aa899130b23.jpg) +VCNet + +![](images/8b4ca0f151db8720bea077140a6c69f7f1894b6ac0d0d6f6660b7e8259b0d473.jpg) +Ours + +![](images/0e1fdc50ed20d77c92d62ed5dac90de8d3ba7217c5c81759be4474b0fdd9c183.jpg) +Groundtruth + +Table 3: Performance comparison for blind image inpainting in terms of four evaluation metrics on three benchmark datasets. + +
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-2020-4040-600-2020-4040-600-2020-4040-60
PSNR ↑MPRNet33.9326.7522.0633.7626.2621.5033.3725.5921.08
VCNet29.8624.5120.5129.1323.3919.4229.6623.1919.18
Ours34.7927.9823.5433.9126.4021.8034.1126.2021.66
SSIM ↑MPRNet0.9680.8980.7990.9650.8820.7620.9630.8720.743
VCNet0.9330.8490.7360.9240.8180.6780.9420.8330.688
Ours0.9710.9100.8240.9660.8840.7670.9690.8870.746
FID ↓MPRNet3.3712.2029.743.7117.8053.384.1517.6144.73
VCNet6.8014.1728.905.6921.4859.115.9318.6240.21
Ours2.377.2815.482.6912.3541.372.8811.1031.33
LPIPS ↓MPRNet0.0270.0960.2120.0330.1230.2690.0330.1310.283
VCNet0.0440.1030.1880.0450.1280.2460.0460.1380.261
Ours0.0190.0650.1390.0250.0980.2260.0240.0950.221
+ +Bidirectional blind image inpainting. When real natural images are used as the noise content for corruption, we can perform bidirectional image restoration: restore either a complete noise image or a complete background image by reversing the corruption mask. It is quite challenging to perform well in inpainting of both directions. The visual results in Figure 6 show that our model performs more robust than VCNet in both corruption detection and image inpainting. + +Generalization on unseen corruption patterns. Using the same experimental settings as evaluating the generalizability for corruption detection, we also validate the generalizability of our model across unseen corruption patterns on image inpainting. Table 5 shows the image inpainting performance of both our model and VCNet on unseen corruption patterns during training. The experimental results indi + +Table 4: Generalizability over unseen corruptions for corruption detection. Models are trained on Places. + +
Corruption patternRandom constantCelebA-HQ [22]
Mask ratio (%)0-3030-600-3030-60
Acc ↑VCNet0.9810.9770.9750.974
Ours0.9900.9890.9860.986
F1 ↑VCNet0.9880.9780.9840.976
Ours0.9940.9900.9920.988
BCE ↓VCNet0.0500.0780.0630.068
Ours0.0500.0490.0620.054
IoU ↑VCNet0.9770.9600.9690.954
Ours0.9880.9810.9830.976
+ +Table 5: Generalizability over unseen corruptions for blind inpainting. Models are trained on Places. + +
Corruption PatternRandom constantCelebA-HQ [22]
Mask ratio (%)0-2020-4040-600-2020-4040-60
PSNR ↑VCNet30.4824.3319.3029.7723.2118.96
Ours37.6929.0123.0534.6026.5221.82
SSIM ↑VCNet0.9570.8760.7320.9470.8400.685
Ours0.9850.9330.8180.9720.8940.772
FID ↓VCNet6.4116.8442.545.4717.7541.19
Ours1.807.4524.482.6510.5030.42
LPIPS ↓VCNet0.0510.1310.2680.0410.1280.256
Ours0.0130.0630.1820.0210.0890.214
+ +Table 6: Ablation study on hierarchical interaction mechanism. + +
Mask ratio (%)0-3030-60
Acc ↑w/o inter-stage consistency0.9690.965
w/o sample selection0.9750.974
Complete model0.9780.976
F1 ↑w/o inter-stage consistency0.9800.969
w/o sample selection0.9840.976
Complete model0.9860.978
BCE ↓w/o inter-stage consistency0.1020.115
w/o sample selection0.0940.101
Complete model0.0900.095
IoU ↑w/o inter-stage consistency0.9620.940
w/o sample selection0.9720.957
Complete model0.9750.959
+ +![](images/54df3df2d71c4b760caa39a7c7eec07f063f4107a4e885e34831c4521a0fd54d.jpg) +Figure 8: Hierarchical vs single-stage contrastive learning. + +![](images/d832645ba19703dcc741d4513b50734e0e36d4d86a0ed32de12fae17216f6925.jpg) + +Table 7: Performance evaluation on watermark removal. + +
MetricsPSNR↑SSIM↑FID↓
SIRF [46]34.630.9780.071
BS2AM [6]34.880.9790.028
DHAN [8]37.670.9860.062
BVMR [19]38.280.9850.018
Split then Refine [7]41.270.9910.011
VCNet [42]32.660.9630.032
Ours41.880.9920.007
+ +cate that our model performs much better than VCNet. Furthermore, the qualitative comparison in Figure 7 also validates such better generalizability of our model over VCNet on both corruption detection and image inpainting. + +# 4.3.2 Image Watermark Removal + +In this set of experiments, we further evaluate the performance of our model on LOGO-30K dataset [7] for watermark removal. The comparison in Table 7 shows that our model performs best in terms of all metrics and compares favorably with other specialized methods for watermark removal, which demonstrates the robustness of our model across different image restoration tasks. Moreover, the visual results in Figure 9 also validate the superiority of our model over other methods in watermark removal. + +# 4.4. Ablation Study + +Effect of hierarchical contrastive learning. To investigate the effectiveness of the proposed hierarchical contrastive learning framework, we compare its performance with that of single-stage contrastive learning. The results in Figure 8 demonstrate the distinct advantage of the proposed hierarchical contrastive learning framework. + +Effect of hierarchical interaction mechanism. We further conduct experiments to investigate the effect of hierarchi + +![](images/9894bef6b72975635c568964f781b6488187f107bdc193288ebeab30fba341c0.jpg) + +![](images/987efa1bf8e7ca5ac201355be1dd3649f461cb368d370668e32ef4e179113436.jpg) + +![](images/bfa12fe0364d403f505267c978999fcb173be067cb7847ad4290cebef70617c8.jpg) + +![](images/4936d45df96aa9456cce80ad225639fc4d0fe4c8ad33c6862f927723abdc0c69.jpg) + +![](images/2f2b7625a05a2b6f362522758840a98ebc05707491a377af433ea5e572ab5754.jpg) + +![](images/d2f651141f9bbe7cd2261c16a0523f7dafa58897c40d86433018822cec0eb2c3.jpg) +Input +Figure 9: Visual comparison with the state-of-the-art model Split then Refine [7] and VCNet for watermark removal. + +![](images/fe9d70bb59f802e763c8c97388c31c4d5940cea248027d6244af6a6b15231581.jpg) +VCNet + +![](images/5b3ed53b4e411fc86705ea930fde9eb86e766cd3e2318957ab0783d18bf59ae2.jpg) +Split then Refine + +![](images/939c0f0aaedb1a7a7f467e93b9a57e46b396092f152dde16edb62b6986dac88e.jpg) +Ours + +![](images/e2c0b7f9e4134cb596c8a54e1de01b5586c8f1ff7e643b73c5d0165f71aebaed.jpg) +Groundtruth + +cal interaction mechanism, especially the proposed 'inter-stage semantic consistency' and 'high-quality sample selection strategy'. Table 6 presents the concrete experimental results. While both techniques can boost the performance, 'inter-stage semantic consistency' is more crucial to the performance due to its essential merit: semantic propagation between stages by feature reusing. + +# 5. Conclusion + +In this work, we have designed a novel method, namely hierarchical contrastive learning framework, which can automatically detect corruption masks in pixel level and thus allows for blind image corruption restoration without known corruption masks. Extensive quantitative and qualitative comparisons have demonstrated the superior performance of our method over other methods for various corruption restoration tasks and its well generalization ability across different corruption patterns. + +# 6. Acknowledgement + +This work was supported in part by the NSFC fund (Grant No. U2013210, 62006060, 62176077), in part by the Shenzhen Key Technical Project under Grant 2022N001, 2020N046, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant (Grant No. 2022A1515010306), in part by Shenzhen Fundamental Research Program (Grant No. JCYJ20210324132210025, JCYJ20220818102415032), and in part by the Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies (Grant No. 2022B1212010005). + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2 +[3] Nian Cai, Zhenghang Su, Zhineng Lin, Han Wang, Zhijing Yang, and Bingo Wing-Kuen Ling. Blind inpainting using the fully convolutional neural network. The Visual Computer, 33(2):249-261, 2017. 2 +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[5] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9):1200-1212, 2004. 2 +[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 8 +[7] Xiaodong Cun and Chi-Man Pun. Split then refine: stacked attention-guided resunets for blind single image visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1184-1192, 2021. 8 +[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10680-10687, 2020. 8 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7 +[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2 +[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11368, 2022. 1, 2 +[12] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, pages 341-346, 2001. 2 +[13] Xin Feng, Wenjie Pei, Fengjun Li, Fanglin Chen, David Zhang, and Guangming Lu. Generative memory-guided semantic reasoning model for image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and + +Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2 +[15] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 5 +[16] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14134-14143, October 2021. 2 +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4 +[19] Amir Hertz, Sharon Fogel, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Blind visual motif removal from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6858–6867, 2019. 8 +[20] Yong Shi Jie Yang, Zhiquan Qi. Learning to incorporate structure knowledge for image inpainting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12605-12612, 2020. 2 +[21] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016. 5 +[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6, 7, 8 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6, 7 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[26] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 2 +[27] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10758-10768, 2022. 1, 2, 5, 6, 7 +[28] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 + +[29] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision, pages 85-100, 2018. 2, 6 +[30] Hongyu Liu, Bin Jiang, Yibing Song, Wei Huang, and Chao Yang. Rethinking image inpainting via a mutual encoder-decoder with feature equalizations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 725-741. Springer, 2020. 2 +[31] Hongyu Liu, Bin Jiang, Yi Xiao, and Chao Yang. Coherent semantic attention for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4170-4179, 2019. 2 +[32] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11347-11357, 2022. 1, 2 +[33] Yang Liu, Jinshan Pan, and Zhixun Su. Deep blind image inpainting. In International Conference on Intelligent Science and Big Data Engineering, pages 128-141. Springer, 2019. 2 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5 +[35] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Qureshi, and Mehran Ebrahimi. Edgeconnect: Structure guided image inpainting using edge prediction. In The IEEE International Conference on Computer Vision Workshops, Oct 2019. 2 +[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[37] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6, 7 +[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 4 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[40] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 3 +[41] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised + +visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2 +[42] Yi Wang, Ying-Cong Chen, Xin Tao, and Jiaya Jia. Vcnet: A robust approach to blind image inpainting. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2, 6, 7, 8 +[43] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 3 +[44] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4471-4480, 2019. 2 +[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 7 +[46] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 8 +[47] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1438-1447, 2019. 2 +[48] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 6, 7 \ No newline at end of file diff --git a/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/images.zip b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..970d5c042a470b018648b81789088f0d68231da4 --- /dev/null +++ b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd27845072d07ea166d4196cca25bb16ac86931fd7c9869967b355531702d393 +size 861714 diff --git a/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/layout.json b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b3297bdd7b205b1da73a16469eb52fd932ae18ad --- /dev/null +++ b/2023/Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection/layout.json @@ -0,0 +1,12142 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 506, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 506, + 140 + ], + "type": "text", + "content": "Hierarchical Contrastive Learning for Pattern-Generalizable Image Corruption Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 160, + 201, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 160, + 201, + 175 + ], + "spans": [ + { + "bbox": [ + 148, + 160, + 201, + 175 + ], + "type": "text", + "content": "Xin Feng*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 161, + 275, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 161, + 275, + 175 + ], + "spans": [ + { + "bbox": [ + 219, + 161, + 275, + 175 + ], + "type": "text", + "content": "Yifeng Xu*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 293, + 161, + 371, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 161, + 371, + 175 + ], + "spans": [ + { + "bbox": [ + 293, + 161, + 371, + 175 + ], + "type": "text", + "content": "Guangming Lu†" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 390, + 161, + 447, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 161, + 447, + 175 + ], + "spans": [ + { + "bbox": [ + 390, + 161, + 447, + 175 + ], + "type": "text", + "content": "Wenjie Pei†" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 196, + 176, + 397, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 176, + 397, + 190 + ], + "spans": [ + { + "bbox": [ + 196, + 176, + 397, + 190 + ], + "type": "text", + "content": "Harbin Institute of Technology, Shenzhen" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 195, + 130, + 277 + ], + "blocks": [ + { + "bbox": [ + 50, + 195, + 130, + 277 + ], + "lines": [ + { + "bbox": [ + 50, + 195, + 130, + 277 + ], + "spans": [ + { + "bbox": [ + 50, + 195, + 130, + 277 + ], + "type": "image", + "image_path": "14a709fb49e13fc575435c9f8ae16c57f9a8338dbc3c441f616e33f4835d67d8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 278, + 130, + 361 + ], + "blocks": [ + { + "bbox": [ + 49, + 278, + 130, + 361 + ], + "lines": [ + { + "bbox": [ + 49, + 278, + 130, + 361 + ], + "spans": [ + { + "bbox": [ + 49, + 278, + 130, + 361 + ], + "type": "image", + "image_path": "36fb8097904e9b486e760364b2a8dd2169ea7921be8a147c96505e13ceb9c398.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 362, + 100, + 373 + ], + "lines": [ + { + "bbox": [ + 80, + 362, + 100, + 373 + ], + "spans": [ + { + "bbox": [ + 80, + 362, + 100, + 373 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 132, + 195, + 213, + 277 + ], + "blocks": [ + { + "bbox": [ + 132, + 195, + 213, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 195, + 213, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 213, + 277 + ], + "type": "image", + "image_path": "94b692096424a9137c9ff1a6050125e20670effcdb55690b63e55c7a56205002.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 132, + 278, + 213, + 361 + ], + "blocks": [ + { + "bbox": [ + 132, + 278, + 213, + 361 + ], + "lines": [ + { + "bbox": [ + 132, + 278, + 213, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 213, + 361 + ], + "type": "image", + "image_path": "ccf867ee4f448548b8d8b36a9f98b17b33b178e9c0d1817d26c0ee95edad515b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 362, + 186, + 372 + ], + "lines": [ + { + "bbox": [ + 160, + 362, + 186, + 372 + ], + "spans": [ + { + "bbox": [ + 160, + 362, + 186, + 372 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 214, + 195, + 296, + 277 + ], + "blocks": [ + { + "bbox": [ + 214, + 195, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 214, + 195, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 214, + 195, + 296, + 277 + ], + "type": "image", + "image_path": "8894c44bc6af25eb6213ab01ea5b2c39507db246608c795b5d8d8917621302f5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 214, + 278, + 296, + 361 + ], + "blocks": [ + { + "bbox": [ + 214, + 278, + 296, + 361 + ], + "lines": [ + { + "bbox": [ + 214, + 278, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 214, + 278, + 296, + 361 + ], + "type": "image", + "image_path": "fda8e8d2ea298685ef5762a33b012c69589f778e7b35cebec77b82002db8a222.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 362, + 265, + 372 + ], + "lines": [ + { + "bbox": [ + 246, + 362, + 265, + 372 + ], + "spans": [ + { + "bbox": [ + 246, + 362, + 265, + 372 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 373, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 546, + 422 + ], + "type": "text", + "content": "Figure 1: Visual comparison between our method and VCNet [42] (the state-of-the-art method for blind image inpainting) on four tasks: blind image inpainting (graffiti corruption) (up left), blind image inpainting (real-image-patch corruption) (up right), watermark removal (bottom left) and shadow removal (bottom right). The groundtruth mask and the predicted masks by different methods are presented." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 298, + 195, + 380, + 277 + ], + "blocks": [ + { + "bbox": [ + 298, + 195, + 380, + 277 + ], + "lines": [ + { + "bbox": [ + 298, + 195, + 380, + 277 + ], + "spans": [ + { + "bbox": [ + 298, + 195, + 380, + 277 + ], + "type": "image", + "image_path": "ad299114892c69d623d37727edc8bcf48a0ea7f94b09a82ffca58f2b48ea36fc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 298, + 278, + 379, + 360 + ], + "blocks": [ + { + "bbox": [ + 298, + 278, + 379, + 360 + ], + "lines": [ + { + "bbox": [ + 298, + 278, + 379, + 360 + ], + "spans": [ + { + "bbox": [ + 298, + 278, + 379, + 360 + ], + "type": "image", + "image_path": "c396922ecf11dfae74e3046b6f4812f7553a1d22403ad862b4e6b45eec992e72.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 362, + 349, + 372 + ], + "lines": [ + { + "bbox": [ + 329, + 362, + 349, + 372 + ], + "spans": [ + { + "bbox": [ + 329, + 362, + 349, + 372 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 380, + 195, + 462, + 277 + ], + "blocks": [ + { + "bbox": [ + 380, + 195, + 462, + 277 + ], + "lines": [ + { + "bbox": [ + 380, + 195, + 462, + 277 + ], + "spans": [ + { + "bbox": [ + 380, + 195, + 462, + 277 + ], + "type": "image", + "image_path": "983e989bbd6b5053d53f2fee356fc5bfbc78d344817e1e2f674417bfdf479ac7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 380, + 278, + 462, + 360 + ], + "blocks": [ + { + "bbox": [ + 380, + 278, + 462, + 360 + ], + "lines": [ + { + "bbox": [ + 380, + 278, + 462, + 360 + ], + "spans": [ + { + "bbox": [ + 380, + 278, + 462, + 360 + ], + "type": "image", + "image_path": "7358596603db67f7455043ac59e055272d13f06e24c40305884a24d735931bba.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 362, + 432, + 372 + ], + "lines": [ + { + "bbox": [ + 406, + 362, + 432, + 372 + ], + "spans": [ + { + "bbox": [ + 406, + 362, + 432, + 372 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 463, + 195, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 463, + 195, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 463, + 195, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 463, + 195, + 545, + 277 + ], + "type": "image", + "image_path": "26d73c590f3ac024c68a1917bbd400650446767fbcc06fab4d2b8c7ce99448c0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 463, + 278, + 545, + 360 + ], + "blocks": [ + { + "bbox": [ + 463, + 278, + 545, + 360 + ], + "lines": [ + { + "bbox": [ + 463, + 278, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 463, + 278, + 545, + 360 + ], + "type": "image", + "image_path": "41baf0baf8d92866bbee179ea956d4bca23f36571bf18ed8c1abe34d0fab093e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 362, + 508, + 371 + ], + "lines": [ + { + "bbox": [ + 489, + 362, + 508, + 371 + ], + "spans": [ + { + "bbox": [ + 489, + 362, + 508, + 371 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 143, + 430, + 192, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 430, + 192, + 441 + ], + "spans": [ + { + "bbox": [ + 143, + 430, + 192, + 441 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 447, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 289, + 687 + ], + "type": "text", + "content": "Effective image restoration with large-size corruptions, such as blind image inpainting, entails precise detection of corruption region masks which remains extremely challenging due to diverse shapes and patterns of corruptions. In this work, we present a novel method for automatic corruption detection, which allows for blind corruption restoration without known corruption masks. Specifically, we develop a hierarchical contrastive learning framework to detect corrupted regions by capturing the intrinsic semantic distinctions between corrupted and uncorrupted regions. In particular, our model detects the corrupted mask in a coarse-to-fine manner by first predicting a coarse mask by contrastive learning in low-resolution feature space and then refines the uncertain area of the mask by high-resolution contrastive learning. A specialized hierarchical interaction mechanism is designed to facilitate the knowledge propagation of contrastive learning in different scales, boosting the modeling performance substantially. The detected multi-scale corruption masks are then leveraged to guide the corruption restoration. Detecting corrupted regions" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 431, + 546, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 546, + 563 + ], + "type": "text", + "content": "by learning the contrastive distinctions rather than the semantic patterns of corruptions, our model has well generalization ability across different corruption patterns. Extensive experiments demonstrate following merits of our model: 1) the superior performance over other methods on both corruption detection and various image restoration tasks including blind inpainting and watermark removal, and 2) strong generalization across different corruption patterns such as graffiti, random noise or other image content. Codes and trained weights are available at https://github.com/xyfJASON/HCL." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 574, + 387, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 387, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 387, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 592, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 700 + ], + "type": "text", + "content": "An essential yet challenging step for image restoration with large-size corruptions, such as blind image inpainting, watermark removal or shadow removal, is to detect the corruption region masks in pixel level precisely. The difficulties lie in the diverse shapes and appearance of corruptions, which can be hardly modeled in a uniform pattern. As a result, blind image inpainting remains a challenging task although image inpainting with known corruption masks [27, 11, 13, 32] has achieved remarkable progress." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "A straightforward way for corruption detection is to for" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 214, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 214, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 214, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12076" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "mulate it as a segmentation task and predict the corruption mask by pixel-wise binary classification. A prominent example following such modeling paradigm is VCNet [42], a state-of-the-art method for blind image inpainting, which regards corruptions as target objects and learns to recognize the semantic patterns of corruptions for detection. While VCNet can successfully detect the corruptions with uniform patterns, it has two potential limitations. First, the corruption may exhibit diverse patterns due to its potentially irregular appearance nature, which VCNet can hardly deal with. Second, it is challenging for VCNet to handle the corruption patterns that are distinct from those appearing in training data, thus it has limited generalization w.r.t. different corruption patterns." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 540 + ], + "type": "text", + "content": "In this work, we propose a novel model for image corruption detection to circumvent the aforementioned limitations of VCNet. Instead of recognizing the semantic patterns of corruptions as VCNet does, we apply metric learning to learn an embedding space in which our model can capture the contrastive semantic distinctions between the corrupted and uncorrupted regions. To this end, we design a hierarchical contrastive learning framework for detecting corrupted regions, which predicts the corruption mask in a coarse-to-fine manner. To be specific, it first predicts a coarse mask by lightweight contrastive learning in a low-resolution feature space. Then it refines the uncertain pixels with low confidence in the predicted mask by high-resolution contrastive learning based on fine-grained features. Note that only a small fraction of predicted masks need to be refined, thus the refining process can be performed quite efficiently. To facilitate the knowledge propagation and consistency of contrastive learning between different scales, we propose a specialized hierarchical interaction mechanism. As a result, the high-resolution contrastive learning can inherit useful knowledge from the low-resolution contrastive learning to achieve precise prediction of corruption masks quite efficiently. The detected multi-scale corruption masks are further used to guide the restoration of the corrupted regions, which also follows a coarse-to-fine generative process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 540, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 613 + ], + "type": "text", + "content": "Unlike VCNet recognizing the semantic patterns of corruptions, our model focuses on capturing the contrastive semantic distinctions between corrupted and uncorrupted regions. Thus our model has well generalization ability across different corruption patterns. As shown in Figure 1. To conclude, we make following contributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 616, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 616, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 664 + ], + "type": "text", + "content": "- We design a hierarchical contrastive learning framework to detect multi-scale corruption masks by learning the contrastive semantic distinctions between corrupted and uncorrupted regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": "- Integrating the proposed hierarchical contrastive learning module, we develop a general-purpose blind image restoration model to perform high-quality image restoration without known corruption masks." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 547, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 169 + ], + "type": "text", + "content": "- Extensive experiments validate the effectiveness of our model both quantitatively and qualitatively in three aspects: 1) superior performance on corruption detection (Section 4.2), 2) favorable performance compared to the specialized methods on two challenging image restoration tasks including blind image inpainting and watermark removal (Section 4.3), and 3) strong generalizability across different corruption patterns (Section 4.2 and 4.3.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 175, + 392, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 392, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 392, + 189 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 197, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 547, + 437 + ], + "type": "text", + "content": "Image Restoration with known Corruption masks. Image restoration with given corruption masks assumes that the corrupted regions are known, and conventional methods [2, 12, 5] leverage uncorrupted content to restore the corrupted regions by pixel diffusion or patch replacement. However, these methods usually generate significant artifacts in the restored image due to the lack of global semantic understanding and generalization ability. With the great success of convolutional neural networks(CNNs) in computer vision, recent methods leverage an encoder-decoder framework to restore corrupted images. A typical way of CNN-based methods [29, 44, 26, 27] employs known uncorrupted regions to infer corrupted content from the outside to the inside. To synthesize reasonable semantics, some methods [31, 35, 20, 30, 16, 28] leverage structure information to improve the quality of semantic structures in the corrupted regions. Promoted by generative models [14, 25], some recent methods [11, 32, 47] attempt to learn generative prior for improving the quality of synthesized images, and generating diverse content for corrupted regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 548, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 548, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 548, + 628 + ], + "type": "text", + "content": "Image corruption detection. Early methods [3, 33] for image corruption detection assume the corrupted content follows a simple distribution, such as constant values, Gaussian noise, etc. However, previous assumption simplifies the detection of corrupted regions, limiting the application scope of image corruption detection. Recently, Wang et al. [42] relax the definition of image corruption detection by increasing the diversity of corruption types, including watermarking, raindrop, or even random images. They also propose a two-stage framework VCNet for blind image restoration, which first locates corrupted regions in a manner of image segmentation and then fills in content following the typical encoder-decoder based methods for image restoration. However, it sometimes misuses content in corrupted regions while reconstructing the complete image due to the challenge of detecting diverse patterns of corruption." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Contrastive learning. Contrastive learning has been widely used in self-supervised representation learning [10, 17, 4]. Instead of matching an input image to a fixed target, contrastive learning maximizes mutual information in the representation space, keeping the query close to the positive sample and away from the negative sample. Previous works [17, 4, 41] have validated the effectiveness of contrastive" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12077" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 509, + 302 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 509, + 302 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 509, + 302 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 509, + 302 + ], + "type": "image", + "image_path": "b2238f8dce237565dd153849e69bff804fa5b767dae51ad0643a14626260294f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 546, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 546, + 329 + ], + "type": "text", + "content": "Figure 2: Architecture of our method. It leverages the designed Hierarchical Contrastive Learning to detect corruption masks and further synthesizes reasonable content for corrupted regions by the developed Transformer-based Restoration module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 338, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 445 + ], + "type": "text", + "content": "learning in high-level vision tasks, due to the inherent suitability for modeling feature contrasts. Recently, some researchers [40, 43] have attempted to introduce contrastive learning into low-level vision tasks. However, pixel-wise contrastive learning on high-resolution images suffers from large computational cost and limited performance. In this paper, we design a hierarchical contrastive learning mechanism to solve above problems and significantly strengthen the performance of image corruption detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 448, + 113, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 113, + 461 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 113, + 461 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 464, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 571 + ], + "type": "text", + "content": "Detecting image corruptions by recognizing the semantic patterns of corruptions is difficult in that the corruptions can potentially exhibit diverse patterns. Our proposed hierarchical contrastive learning framework learns the contrastive semantic distinctions between corrupted and uncorrupted regions, and detects corruption masks in a coarse-to-fine manner. The obtained multi-scale corruption masks are then leveraged to guide the restoration of the corruptions in a coarse-to-fine generative process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 114, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 114, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 114, + 585 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "text", + "content": "Problem formulation. Given an input image " + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "text", + "content": " which is corrupted from an intact groundtruth image " + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{gt}} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "text", + "content": ", the task of blind image restoration aims to first detect the corruption region mask " + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "text", + "content": " and then reconstruct a complete image " + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "text", + "content": " by synthesizing realistic content " + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}" + }, + { + "bbox": [ + 47, + 587, + 287, + 668 + ], + "type": "text", + "content": " for the corrupted regions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 666, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 666, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 108, + 666, + 287, + 694 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {I} = \\mathbf {O} _ {\\mathrm {g t}} \\odot \\mathbf {M} + \\mathbf {N} \\odot (1 - \\mathbf {M}), \\\\ \\hat {\\mathbf {O}} = \\mathbf {O} _ {\\mathrm {g t}} \\odot \\mathbf {M} + \\hat {\\mathbf {C}} \\odot (1 - \\mathbf {M}). \\tag {1} \\\\ \\end{array}", + "image_path": "97bd31a7b46bd379081d88cb3f6e5dab629723c8459d4d5e20c1127d8aa06b46.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "text", + "content": "Herein, " + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 47, + 700, + 287, + 713 + ], + "type": "text", + "content": " is a binary map where the values of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "text", + "content": "corrupted regions are 0 and uncorrupted regions are filled with " + }, + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "inline_equation", + "content": "1.\\mathbf{N}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "text", + "content": " is the noisy content in the corrupted regions. Precise detection of the corruption mask " + }, + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 338, + 545, + 386 + ], + "type": "text", + "content": " is crucial to the performance of image restoration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "text", + "content": "As shown in Figure 2, our model follows the encoder-decoder framework and consists of two core modules: Hierarchical Contrastive Learning module for detecting corruption masks and Transformer-based Restoration module for corruption restoration. The corrupted image is first encoded into multi-scale feature maps by Encoder " + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "text", + "content": " of Transformer-based Synthesis module. Then the proposed Hierarchical Contrastive Learning module is employed to perform corruption detection in a coarse-to-fine manner from these feature maps, and predicts multi-scale corruption masks. Finally, the obtained masks are fed into Decoder " + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{Dec}}" + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "text", + "content": " of the Transformer-based Synthesis module for restoration, which is also in a coarse-to-fine generative process. Both Encoder " + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "text", + "content": " and Decoder " + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{dec}}" + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "text", + "content": " are mainly composed of basic transformer blocks of MAT [39], while they are equipped with additional down-sampling layers and upsampling layers respectively. We employ a shallow convolutional layer with " + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "inline_equation", + "content": "5\\times 5" + }, + { + "bbox": [ + 304, + 386, + 546, + 649 + ], + "type": "text", + "content": " kernel before Encoder and after Decoder to perform basic feature transformation. Besides, we also employ another Conv-U-Net to refine high-frequency details of output results, leaning upon the local texture refinement capability and efficiency of CNNs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 651, + 545, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 545, + 676 + ], + "type": "text", + "content": "3.2. Corruption Detection by Hierarchical Contrastive Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We design Hierarchical Contrastive Learning framework to guide the learning of multi-scale semantic embedding spaces of Encoder " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " in such a way that the semantic dis" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12078" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "tances between two arbitrary pixels both within uncorrupted or corrupted regions (intra-region distance) should be minimized while the distance between a corrupted pixel and an uncorrupted pixel (inter-region distance) should be maximized. As a result, our model can capture the intrinsic semantic distinctions between corrupted and uncorrupted regions. Then our model performs clustering on all pixels in this learned embedding space to separate them into two clusters, corresponding to the corrupted and uncorrupted regions, and thus achieves the corruption mask." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "type": "text", + "content": "As shown in Figure 2, Encoder " + }, + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 46, + 192, + 290, + 324 + ], + "type": "text", + "content": " consists of three encoding stages and produces three scales of encoded feature maps. Accordingly, Our model performs three stages of pixel-level contrastive learning for the corresponding scale of feature maps to guide the learning of its embedding space. We will first describe how our model performs corruption detection in one stage with single-scale contrastive learning. Then we will elaborate on the proposed Hierarchical Interaction Mechanism which enables our model to perform hierarchical contrastive learning quite efficiently for coarse-to-fine multi-scale mask detection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 332, + 228, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 228, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 228, + 344 + ], + "type": "text", + "content": "3.2.1 Single-Scale Contrastive Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "content": "Supervised Contrastive Learning. During each stage of contrastive learning, we construct positive pixel pairs by using intra-region pixels, i.e., both pixels are from either the uncorrupted region or the corrupted region. In contrast, each negative training pair consists of two inter-region pixels, one from the corrupted region and the other from the uncorrupted region. To be specific, for a query pixel " + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "content": " from a randomly selected query set " + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "content": ", we construct the positive set " + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "content": " by randomly selecting pixels from the same region as " + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "content": " and construct the negative set " + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "content": " from the opposite region. Then we apply Circle loss [38] to maximize the cosine similarity of positive pairs while minimizing the similarities of negative pairs. Formally, the contrastive learning in the stage " + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "inline_equation", + "content": "S_{s}" + }, + { + "bbox": [ + 46, + 350, + 289, + 518 + ], + "type": "text", + "content": " is guided by the loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 521, + 287, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 287, + 551 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L}} ^ {s} = \\sum_ {q \\in Q} \\log \\left[ 1 + \\sum_ {p \\in P} \\exp (- \\mathbf {e} _ {q} \\cdot \\mathbf {e} _ {p} / \\tau) \\cdot \\sum_ {n \\in N} \\exp (\\mathbf {e} _ {q} \\cdot \\mathbf {e} _ {n} / \\tau) \\right], \\tag {2}", + "image_path": "dc5e37dc943d23c0df75545e98124dfbbaffbe3503dc24861bb9df3060a80508.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "text", + "content": " is a scale factor. " + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_q" + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "text", + "content": " is the features for pixel " + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "text", + "content": " projected from the corresponding embedding space of Encoder " + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}^s" + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "text", + "content": " by a projection head in Stage- " + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 554, + 288, + 590 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 594, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 594, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 122, + 594, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {q} = \\mathcal {F} _ {\\operatorname {p r o j}} ^ {s} \\left(\\mathcal {F} _ {\\mathrm {e n c}} ^ {s} \\left(\\mathbf {I} _ {q}\\right)\\right), \\tag {3}", + "image_path": "1e05be92973cd671f7ea5a33977b781157d27246b5abb6aeef83015af71902b4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 612, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 287, + 647 + ], + "type": "text", + "content": "where the projection head " + }, + { + "bbox": [ + 47, + 612, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{proj}}" + }, + { + "bbox": [ + 47, + 612, + 287, + 647 + ], + "type": "text", + "content": " comprises two fully connected layers and a GELU [18] layer in-between for nonlinear transformation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 647, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 288, + 714 + ], + "type": "text", + "content": "Corruption Mask Detection. Under the supervision of contrastive learning in Equation 2, the pixels within the same region, either the corrupted or the uncorrupted regions, tend to have similar representations in each scale of embedding space of Encoder " + }, + { + "bbox": [ + 46, + 647, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 46, + 647, + 288, + 714 + ], + "type": "text", + "content": " while the pixels from different regions would have dissimilar representations. Thus," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 416, + 223 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 416, + 223 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 416, + 223 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 416, + 223 + ], + "type": "image", + "image_path": "3e1f73fb7e08b0078a8017fc0fae87110f948e0df96defb3b4a1ea1fc8b6c507.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 234, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 304, + 234, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 547, + 293 + ], + "type": "text", + "content": "Figure 3: In the stages " + }, + { + "bbox": [ + 304, + 234, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_2" + }, + { + "bbox": [ + 304, + 234, + 547, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 234, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_3" + }, + { + "bbox": [ + 304, + 234, + 547, + 293 + ], + "type": "text", + "content": ", both the selected representative samples and uncertain samples in the previous stage are used for contrastive learning in current stage while only the mask labels of uncertain samples are refined during mask detection." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 420, + 70, + 547, + 223 + ], + "blocks": [ + { + "bbox": [ + 420, + 70, + 547, + 223 + ], + "lines": [ + { + "bbox": [ + 420, + 70, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 420, + 70, + 547, + 223 + ], + "type": "image", + "image_path": "439466aac29665d9c911bcad7c57cf98168eab0f6224ef8d9081cc2cf4490c2c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 303, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 547, + 337 + ], + "type": "text", + "content": "we can perform clustering to separate them into two clusters. Specifically, our model adopts K-means algorithm for clustering, and the clustering in Stage- " + }, + { + "bbox": [ + 304, + 303, + 547, + 337 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 303, + 547, + 337 + ], + "type": "text", + "content": " is performed by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 339, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 339, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 337, + 339, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {c} _ {1} ^ {s}, \\mathbf {c} _ {2} ^ {s} \\right], \\left[ \\mathbf {M} _ {1} ^ {s}, \\mathbf {M} _ {2} ^ {s} \\right] = \\mathrm {K - m e a n s} \\left(\\mathcal {F} _ {\\operatorname {p r o j}} ^ {s} \\left(\\mathcal {F} _ {\\mathrm {e n c}} ^ {s} (\\mathbf {I})\\right)\\right), \\tag {4}", + "image_path": "0cdd4976d36b5dc193e3e72a2a70261f1e685ed32841ebbf3e9dee4d99462bf2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 356, + 547, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 356, + 547, + 453 + ], + "type": "inline_equation", + "content": "[\\mathbf{c}_1^s,\\mathbf{c}_2^s ]" + }, + { + "bbox": [ + 304, + 356, + 547, + 453 + ], + "type": "text", + "content": " denotes the embeddings of two produced cluster centers and " + }, + { + "bbox": [ + 304, + 356, + 547, + 453 + ], + "type": "inline_equation", + "content": "[\\mathbf{M}_1^s,\\mathbf{M}_2^s ]" + }, + { + "bbox": [ + 304, + 356, + 547, + 453 + ], + "type": "text", + "content": " are the associated binary masks. To identify which cluster corresponds to the corrupted or uncorrupted regions, we train a lightweight binary classifier consisting of two fully connected layers with a ReLU layer, and use it to perform classification on two cluster centers. Thus, the associated mask for the cluster of the corrupted region is the predicted corruption mask." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 458, + 495, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 458, + 495, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 458, + 495, + 469 + ], + "type": "text", + "content": "3.2.2 Hierarchical Interaction Mechanism" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "text", + "content": "Our model performs Hierarchical Contrastive Learning including three stages to detect the corruption mask in a coarse-to-fine manner. It first performs contrastive learning in the lowest-resolution stage " + }, + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_1)" + }, + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "text", + "content": " to predict a coarse mask for corruption. Then it refines the uncertain pixels of the mask with low confidence by higher-resolution contrastive learning in subsequent stages " + }, + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_2" + }, + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_3)" + }, + { + "bbox": [ + 304, + 474, + 547, + 677 + ], + "type": "text", + "content": ". We propose Hierarchical Interaction Mechanism to facilitate the interaction and knowledge propagation between different stages of contrastive learning and guarantee the semantic consistency between them. In particular, higher-resolution contrastive learning can inherit useful knowledge from previous stages of contrastive learning, which can improve the learning performance substantially. Quadtree structure are used to build the positional correspondence between feature maps in adjacent stages considering that the resolution of feature maps is always scaled by four times between adjacent stages." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Selecting High-Quality Training Samples. During the contrastive learning in " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": ", we construct the query set " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " as well as the associated positive set " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": " and negative set " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12079" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "in Equation 2 by randomly selecting pixels from the whole feature map. To improve the efficiency of contrastive learning in the higher stages " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "(\\mathrm{S}_2" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_3)" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ", we select a small fraction of pixels from the entire feature maps, which are crucial for contrastive learning, as the high-quality candidate set for constructing " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 192 + ], + "type": "text", + "content": "To be specific, we first measure the prediction confidence of each pixel based on the clustering results of last stage. The prediction confidence of the pixel " + }, + { + "bbox": [ + 46, + 144, + 288, + 192 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 144, + 288, + 192 + ], + "type": "text", + "content": " in the stage " + }, + { + "bbox": [ + 46, + 144, + 288, + 192 + ], + "type": "inline_equation", + "content": "S_{s}" + }, + { + "bbox": [ + 46, + 144, + 288, + 192 + ], + "type": "text", + "content": " is calculated by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 101, + 191, + 287, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 191, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 101, + 191, + 287, + 220 + ], + "type": "interline_equation", + "content": "z _ {q} ^ {s} = \\frac {\\exp \\left(- \\left(\\mathbf {e} _ {q} \\cdot \\mathbf {c} _ {y _ {q}} ^ {s} / \\tau\\right)\\right)}{\\sum_ {i = 1} ^ {2} \\exp \\left(- \\left(\\mathbf {e} _ {q} \\cdot \\mathbf {c} _ {i} ^ {s} / \\tau\\right)\\right)}, \\tag {5}", + "image_path": "96b8ea3730710bbd075ef1430379b87a70a5d6c09502b29821a81a450791947d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "y_{q}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " denotes the cluster index " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " belongs to and " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " is the scaling factor. Then we pick out two types of pixels as the high-quality query set for the stage " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "S_{s+1}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ": 1) the pixels with high confidence which are regarded as representatives of two clusters and are typically close to the cluster centers; 2) the uncertain pixels with low confidence which are mostly near the boundaries between the corrupted and uncorrupted regions. Selecting pixels in such a way for constructing the positive and negative training pairs, our model is able to perform high-resolution contrastive learning in the stages " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "S_{3}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " quite efficiently and effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": "Inter-Stage Semantic Consistency. To guarantee the semantic consistency between different stages of contrastive learning, we reuse the learned features of the previous stage of contrastive learning in current stage. As shown in Figure 3, we concatenate the features in the previous stage with the features in the current stage, which are then fed into the projection head to produce input features for current stage. Accordingly, Equation 3 for high-resolution stages " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "(S_{2}" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "S_{3})" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": " evolves into the following form:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 460, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 460, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 63, + 460, + 287, + 475 + ], + "type": "interline_equation", + "content": "\\mathbf {e} _ {q} = \\mathcal {F} _ {\\text {p r o j}} ^ {s} \\left(\\operatorname {C o n c a t} \\left(\\mathcal {F} _ {\\text {e n c}} ^ {s} \\left(\\mathbf {I} _ {q}\\right), \\mathcal {F} _ {\\text {m a p}} ^ {s} \\left(\\mathcal {F} _ {\\text {e n c}} ^ {s - 1} \\left(\\mathbf {I} _ {q}\\right)\\right)\\right)\\right), \\tag {6}", + "image_path": "76b5dfab62b4e6d209f5a695f6f15dca82e2d3a41591267e80c6948dc1b3f7c0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "content": "where the mapping function " + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{map}}^s" + }, + { + "bbox": [ + 46, + 475, + 287, + 548 + ], + "type": "text", + "content": " is a linear transformation to regulate the feature dimension of previous stage and balance the effect between current and previous features. Besides the semantic consistency between adjacent stages, another merit of such feature reusing is that our model is able to inherit the learned semantics from previous stage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 548, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 668 + ], + "type": "text", + "content": "Refining the Mask Prediction of Uncertain Pixels. During the high-resolution contrastive learning in the stages " + }, + { + "bbox": [ + 46, + 548, + 287, + 668 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 46, + 548, + 287, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 548, + 287, + 668 + ], + "type": "inline_equation", + "content": "S_3" + }, + { + "bbox": [ + 46, + 548, + 287, + 668 + ], + "type": "text", + "content": ", we only refine the mask prediction by re-predicting the labels of uncertain pixels with low confidence in the previous stage, i.e., the type-2 samples of selected high-quality query set. The mask labels of other pixels with high confidence are directly inherited from the predictions in the previous stage according to the built quadtree, assuming that the predictions of these pixels are reliable. Thus, such refining process can be performed quite efficiently." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 256, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 256, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 256, + 683 + ], + "type": "text", + "content": "3.3. Transformer-based Restoration Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "As shown in Figure 2, the detected corruption masks are fed into Decoder of the Transformer-based Restoration" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "module for corruption restoration, which consists of the designed mask-guided transformer block." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 95, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 179 + ], + "type": "text", + "content": "Mask-guided transformer block. As illustrated in Figure 2, our transformer block contains a multi-head contextual attention (MCA) module, followed by a two-layer MLP with GELU nonlinearity in between. Besides, a layer normalization (LN) [1] layer is applied before each attention and each MLP, and a residual learning connection is employed after each module:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 183, + 545, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 183, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 348, + 183, + 545, + 201 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {\\prime s} = \\operatorname {M C A} \\left(\\operatorname {L N} \\left(F _ {i - 1} ^ {s}\\right), \\mathbf {M} ^ {s}\\right) + F _ {i - 1} ^ {s}, \\tag {7}", + "image_path": "d1a134062bf193a7609bcf5de9b03d6381abd06e7ba111d4610c35ac5f10b3d7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 198, + 472, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 198, + 472, + 212 + ], + "spans": [ + { + "bbox": [ + 351, + 198, + 472, + 212 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {s} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(F _ {i} ^ {\\prime s}\\right)\\right) + F _ {i} ^ {\\prime s},", + "image_path": "44ae3fc562ee58e227c398c10da2b25fee9bc21aeb52184b7a9fe1bc5713b6ec.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "inline_equation", + "content": "F_{i}^{s}" + }, + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "text", + "content": " is the output features of the " + }, + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "text", + "content": "-th block in the stage " + }, + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "inline_equation", + "content": "S_{s}" + }, + { + "bbox": [ + 304, + 219, + 545, + 274 + ], + "type": "text", + "content": ". To fully exploit the predicted masks by hierarchical contrastive learning, we employ the multi-head contextual attention module proposed in MAT [27], which follows the shifted window manner [34] and is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 277, + 545, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 277, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 348, + 277, + 545, + 301 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\mathrm {T}} + \\hat {\\mathbf {M}} ^ {s}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V}, \\tag {8}", + "image_path": "70af2935563469a67b0d1a9866f1224360e67732ea98810ce7cb580c49dee6ca.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{Q},\\mathbf{K},\\mathbf{V}" + }, + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "text", + "content": " are the query, key, and value matrices, respectively. " + }, + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\sqrt{d_k}" + }, + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "text", + "content": " is the scaling factor. " + }, + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 303, + 545, + 327 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 385, + 331, + 545, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 331, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 385, + 331, + 545, + 344 + ], + "type": "interline_equation", + "content": "\\mathbf {\\hat {M}} _ {i} ^ {s} = \\gamma \\left(\\mathbf {M} _ {i} ^ {s} - 1\\right), \\tag {9}", + "image_path": "1fdc5036310dfb6d329e2ac182d7d647a0e93320b18ec268c93852859ba57f4c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 347, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 347, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 347, + 545, + 371 + ], + "type": "text", + "content": " is a large positive integer to reduce the values of corrupted pixels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 376, + 531, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 531, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 531, + 389 + ], + "type": "text", + "content": "3.4. Joint Optimization for Parameter Learning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 395, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 502 + ], + "type": "text", + "content": "The proposed Hierarchical Contrastive Learning module and the Restoration module are integrated in a holistic way based on the encoding-decoding framework, forming a general-purpose blind image restoration model, as illustrated in Figure 2. The whole model is optimized jointly. Besides the loss of hierarchical contrastive learning (Equation 2), we also adopt three more supervision signals for corruption restoration, including pixel-wise reconstruction loss, perceptual loss, and adversarial loss." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": "Pixel Reconstruction Loss, which pushes the restored image " + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}" + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": " as close as its groundtruth " + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 379, + 529, + 545, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 529, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 379, + 529, + 545, + 551 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p i x e l}} = \\left\\| \\hat {\\mathbf {O}} - \\mathbf {O} _ {\\mathrm {g t}} \\right\\| _ {1}. \\tag {10}", + "image_path": "d9a618905a37e8782fc1d48f863f719db590b612c58572a6dbed05b18854a24c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 552, + 544, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 544, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 544, + 574 + ], + "type": "text", + "content": "Perceptual Loss [21], which performs semantic supervision on the restored images in the deep feature space:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 337, + 574, + 545, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 574, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 337, + 574, + 545, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p e r c}} = \\sum_ {l = 1} ^ {L} \\frac {1}{C _ {l} H _ {l} W _ {l}} \\left\\| f _ {\\mathrm {v g g}} ^ {l} (\\hat {\\mathbf {O}}) - f _ {\\mathrm {v g g}} ^ {l} \\left(\\mathbf {O} _ {\\mathrm {g t}}\\right) \\right\\| _ {1}. \\tag {11}", + "image_path": "a4e61d9742f254fc1a1593d76a2b2b8bc3e8101cb1eb96cf2590dfd5639f2e52.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "text", + "content": "Herein " + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{vgg}}^{l}(\\hat{\\mathbf{O}})" + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{vgg}}^{l}(\\mathbf{O}_{\\mathrm{gt}})" + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "text", + "content": " are the extracted features from " + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}" + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "text", + "content": " respectively from the " + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 604, + 545, + 643 + ], + "type": "text", + "content": "-th convolution layer of the pre-trained VGG-19 network [36]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 643, + 545, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 643, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 545, + 667 + ], + "type": "text", + "content": "Adversarial Loss, which employs WGAN-GP [15] to encourage " + }, + { + "bbox": [ + 304, + 643, + 545, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}" + }, + { + "bbox": [ + 304, + 643, + 545, + 667 + ], + "type": "text", + "content": " to be as realistic as its groundtruth " + }, + { + "bbox": [ + 304, + 643, + 545, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 643, + 545, + 667 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 369, + 668, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 668, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 369, + 668, + 545, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d v}} = - \\mathbb {E} _ {\\hat {\\mathbf {O}} \\sim \\mathbb {P} _ {\\hat {\\mathbf {O}}}} \\left[ D (\\hat {\\mathbf {O}}) \\right], \\tag {12}", + "image_path": "8dfa37f09f7de656f31cd15b4cca0a5f69be4f111ecd7da07247d2f632277f26.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " is a discriminator. Overall, the whole model is optimized by:" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12080" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 128, + 125 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 128, + 125 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 128, + 125 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 128, + 125 + ], + "type": "image", + "image_path": "825b180a91c3c7df6991f36072543670b022516ffc282562049e22cbd90147b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 125, + 128, + 180 + ], + "blocks": [ + { + "bbox": [ + 73, + 125, + 128, + 180 + ], + "lines": [ + { + "bbox": [ + 73, + 125, + 128, + 180 + ], + "spans": [ + { + "bbox": [ + 73, + 125, + 128, + 180 + ], + "type": "image", + "image_path": "8727dc59fdae44809fc2cf3f414f39706c1c3a7052ed7c2b2b09a4e9cea79df5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 182, + 108, + 190 + ], + "lines": [ + { + "bbox": [ + 93, + 182, + 108, + 190 + ], + "spans": [ + { + "bbox": [ + 93, + 182, + 108, + 190 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 185, + 125 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 185, + 125 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 185, + 125 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 185, + 125 + ], + "type": "image", + "image_path": "682707e653889dfeaec8a7d28f4ded4f2e7993052a6fc86ea1f78b53204828a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 125, + 185, + 180 + ], + "blocks": [ + { + "bbox": [ + 129, + 125, + 185, + 180 + ], + "lines": [ + { + "bbox": [ + 129, + 125, + 185, + 180 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 185, + 180 + ], + "type": "image", + "image_path": "8b86f44ed78f713202c481fb0be6f3bbcfe5d400e94695dfc22ed4ee54172ba2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 181, + 167, + 189 + ], + "lines": [ + { + "bbox": [ + 146, + 181, + 167, + 189 + ], + "spans": [ + { + "bbox": [ + 146, + 181, + 167, + 189 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 186, + 70, + 241, + 125 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 241, + 125 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 241, + 125 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 241, + 125 + ], + "type": "image", + "image_path": "f0e88a4bdcdf442acf0655e2c65c04c925fcd655fee2a13a2016daec75c725e4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 186, + 125, + 241, + 181 + ], + "blocks": [ + { + "bbox": [ + 186, + 125, + 241, + 181 + ], + "lines": [ + { + "bbox": [ + 186, + 125, + 241, + 181 + ], + "spans": [ + { + "bbox": [ + 186, + 125, + 241, + 181 + ], + "type": "image", + "image_path": "a309e23a2cbe91ba931c5deb44b1b4a79e9f7b1af9ccea058ef6a6e8b64cb214.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 182, + 220, + 190 + ], + "lines": [ + { + "bbox": [ + 205, + 182, + 220, + 190 + ], + "spans": [ + { + "bbox": [ + 205, + 182, + 220, + 190 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 191, + 529, + 203 + ], + "lines": [ + { + "bbox": [ + 60, + 191, + 529, + 203 + ], + "spans": [ + { + "bbox": [ + 60, + 191, + 529, + 203 + ], + "type": "text", + "content": "Figure 4: Visualization of detected masks and reconstructed images on four randomly selected samples from test set." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 242, + 70, + 296, + 125 + ], + "blocks": [ + { + "bbox": [ + 242, + 70, + 296, + 125 + ], + "lines": [ + { + "bbox": [ + 242, + 70, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 242, + 70, + 296, + 125 + ], + "type": "image", + "image_path": "8bd36a76de3dedc97e4cbacc1e5f17e0d3727a710274ce372eee3779343dd1c4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 242, + 125, + 296, + 181 + ], + "blocks": [ + { + "bbox": [ + 242, + 125, + 296, + 181 + ], + "lines": [ + { + "bbox": [ + 242, + 125, + 296, + 181 + ], + "spans": [ + { + "bbox": [ + 242, + 125, + 296, + 181 + ], + "type": "image", + "image_path": "b5901f0ad0cb9e25095f0378fa34de10fec2ffa6984cf21961cb92553bcb4988.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 182, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 251, + 182, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 251, + 182, + 286, + 190 + ], + "type": "text", + "content": "Groundtruth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 297, + 70, + 353, + 125 + ], + "blocks": [ + { + "bbox": [ + 297, + 70, + 353, + 125 + ], + "lines": [ + { + "bbox": [ + 297, + 70, + 353, + 125 + ], + "spans": [ + { + "bbox": [ + 297, + 70, + 353, + 125 + ], + "type": "image", + "image_path": "b754e9d7dddeb468db2035245c287d63896ceae9285eaf0ea617c877d37032c8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 297, + 125, + 352, + 181 + ], + "blocks": [ + { + "bbox": [ + 297, + 125, + 352, + 181 + ], + "lines": [ + { + "bbox": [ + 297, + 125, + 352, + 181 + ], + "spans": [ + { + "bbox": [ + 297, + 125, + 352, + 181 + ], + "type": "image", + "image_path": "5e4a9c8ade344890d1503f697e7f34dd176c00669102fa031c621fa95801d4d3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 182, + 332, + 190 + ], + "lines": [ + { + "bbox": [ + 317, + 182, + 332, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 332, + 190 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 353, + 70, + 408, + 125 + ], + "blocks": [ + { + "bbox": [ + 353, + 70, + 408, + 125 + ], + "lines": [ + { + "bbox": [ + 353, + 70, + 408, + 125 + ], + "spans": [ + { + "bbox": [ + 353, + 70, + 408, + 125 + ], + "type": "image", + "image_path": "84106f35af3c94bc3014b5a90ba21772b56c5b732af6aae5ded4cb2adc7b76c1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 353, + 125, + 408, + 181 + ], + "blocks": [ + { + "bbox": [ + 353, + 125, + 408, + 181 + ], + "lines": [ + { + "bbox": [ + 353, + 125, + 408, + 181 + ], + "spans": [ + { + "bbox": [ + 353, + 125, + 408, + 181 + ], + "type": "image", + "image_path": "4e1b3344027f886d9abff9321ed370839605f113d0bc614259de5d4ee8ff38be.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 182, + 391, + 190 + ], + "lines": [ + { + "bbox": [ + 370, + 182, + 391, + 190 + ], + "spans": [ + { + "bbox": [ + 370, + 182, + 391, + 190 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 409, + 70, + 465, + 125 + ], + "blocks": [ + { + "bbox": [ + 409, + 70, + 465, + 125 + ], + "lines": [ + { + "bbox": [ + 409, + 70, + 465, + 125 + ], + "spans": [ + { + "bbox": [ + 409, + 70, + 465, + 125 + ], + "type": "image", + "image_path": "0f8a6f712357e38d6915de1f852fcf9422bd984be15549b1ad5315e056bc1e22.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 409, + 125, + 465, + 181 + ], + "blocks": [ + { + "bbox": [ + 409, + 125, + 465, + 181 + ], + "lines": [ + { + "bbox": [ + 409, + 125, + 465, + 181 + ], + "spans": [ + { + "bbox": [ + 409, + 125, + 465, + 181 + ], + "type": "image", + "image_path": "81649f551577d913c25202a036bb08c24db5dff183cb051c2239625036e6187f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 182, + 444, + 189 + ], + "lines": [ + { + "bbox": [ + 429, + 182, + 444, + 189 + ], + "spans": [ + { + "bbox": [ + 429, + 182, + 444, + 189 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 520, + 125 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 520, + 125 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 520, + 125 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 520, + 125 + ], + "type": "image", + "image_path": "d1ce3b6bb84acf4b98d4d16bddc062bc75751e26d82e1c2b6a7e5ba87188e81e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 465, + 125, + 520, + 181 + ], + "blocks": [ + { + "bbox": [ + 465, + 125, + 520, + 181 + ], + "lines": [ + { + "bbox": [ + 465, + 125, + 520, + 181 + ], + "spans": [ + { + "bbox": [ + 465, + 125, + 520, + 181 + ], + "type": "image", + "image_path": "160c77a6319911d8461a1a81e94f17ee6383e61d50ee69c5b53abaf6d7f94620.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 182, + 509, + 190 + ], + "lines": [ + { + "bbox": [ + 475, + 182, + 509, + 190 + ], + "spans": [ + { + "bbox": [ + 475, + 182, + 509, + 190 + ], + "type": "text", + "content": "Groundtruth" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 61, + 210, + 287, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 287, + 241 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {p i x e l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}} + \\lambda_ {3} \\sum_ {s = 1} ^ {3} \\mathcal {L} _ {\\mathrm {C L}} ^ {\\mathrm {s}} + \\lambda_ {4} \\mathcal {L} _ {\\text {a d v}}, \\tag {13}", + "image_path": "81c1531ad474af1825abfa914da3bdb742cbca34eba3a9217e0888ce641da4c4.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 242, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 287, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 242, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 47, + 242, + 287, + 266 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 242, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 47, + 242, + 287, + 266 + ], + "type": "text", + "content": " are hyper-parameters to balance between different losses." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 266, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 287, + 315 + ], + "type": "text", + "content": "We integrate the Contrastive Learning module between Encoder and Decoder with consistent coarse-to-fine refining stages for all three modules, which potentially enables synchronized optimization of three modules in each scale." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 323, + 128, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 128, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 128, + 337 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 343, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 287, + 415 + ], + "type": "text", + "content": "We first evaluate the performance of our method on image corruption detection, then validates the effectiveness of our method on two challenging image restoration tasks including blind image inpainting and watermark removal. Finally, we conduct extensive ablation study to obtain more insights into our method." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 422, + 174, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 174, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 174, + 435 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 548 + ], + "type": "text", + "content": "Following the previous work [42] for blind image inpainting, we make two pre-processing operations for data generation: 1) we randomly select natural images from large-scale datasets rather than simple constant values or Gaussian noise as the noisy content for corruptions to increase the difficulty of mask detection in blind image inpainting; 2) we smooth the mask boundaries using alpha blending to avoid distinct boundaries between the corrupted and uncorrupted regions." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": "Four large-scale datasets are used in our experiments, including FFHQ (faces) [23], CelebA-HQ (faces) [22], ImageNet (objects) [9], and Places (scenes) [48]. For each dataset, images from different datasets are randomly selected as the noise content for corruptions. For instance, the images from CelebA-HQ and ImageNet are used as noise content for FFHQ. We employ the method [29] to generate irregular masks with mask shapes and corruption ratios. We perform experiments on two resolutions of images for comprehensive evaluation: " + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "content": "Adam [24] is used as the optimizer by setting " + }, + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 667, + 287, + 715 + ], + "type": "text", + "content": ", initial learning rate and batch size to be 0.9, 0.999, 0.0001, and 4, respectively. More experimental details are provided in the supplementary material." + } + ] + } + ], + "index": 33 + }, + { + "type": "table", + "bbox": [ + 337, + 238, + 514, + 331 + ], + "blocks": [ + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "type": "text", + "content": "Table 1: Performance comparison for mask detection on three benchmark datasets." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 238, + 514, + 331 + ], + "lines": [ + { + "bbox": [ + 337, + 238, + 514, + 331 + ], + "spans": [ + { + "bbox": [ + 337, + 238, + 514, + 331 + ], + "type": "table", + "html": "
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-3030-600-3030-600-3030-60
Acc ↑VCNet0.9480.9430.9720.9780.9740.976
Ours0.9780.9760.9820.9830.9840.985
F1 ↑VCNet0.9670.9480.9820.9800.9840.978
Ours0.9860.9780.9890.9850.9910.987
BCE ↓VCNet0.1260.1370.0730.0550.0640.060
Ours0.0900.0950.0830.0750.0680.064
IoU ↑VCNet0.9310.9000.9660.9620.9690.959
Ours0.9750.9590.9780.9650.9820.974
", + "image_path": "e5077de4cad100baa41cf604936209b62c0832656778f7c31a0c69ec1b87570f.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 337, + 462, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 337, + 462, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 462, + 350 + ], + "type": "text", + "content": "4.2. Image Corruption Detection" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": "Accuracy of corruption detection. We first compare the accuracy of corruption detection between our model and VCNet, the state-of-the-art method for corruption detection in blind image inpainting. Four metrics are used for comprehensive evaluation: binary cross entropy (BCE), classification accuracy (Acc), F1 score, and intersection over union (IoU). Table 1 lists the results for different corruption ratios on three datasets in " + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 354, + 545, + 521 + ], + "type": "text", + "content": " resolution, which show that our model achieves superior performance of mask detection over VCNet in terms of all metrics except BCE. This is reasonable because VCNet is directly supervised by the BCE loss. Besides, we also make qualitative comparison in Figure 4, in which our model detects more precise corruption masks and restores higher-quality images than VCNet." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "content": "Image inpainting based on detected corruption masks by different methods. As an indirect evaluation of corruption detection, we perform image inpainting over the detected corruption masks from different methods, employing the same inpainting model MAT [27], and then compare the inpainting performance. To have a comprehensive evaluation, we also provide the inpainting performance over the detected masks from a state-of-the-art segmentation model Segmenter [37] and the groundtruth mask. The results on " + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 521, + 545, + 677 + ], + "type": "text", + "content": " images in Table 2 show that MAT achieves the highest inpainting performance with the mask detected by our method, which is consistent with the performance comparison of corruption detection." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Generalizability on unseen corruption patterns. To evaluate the generalizability of our model methods across different corruption patterns, we perform training and test over" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12081" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 108, + 183 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 108, + 183 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 108, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 108, + 183 + ], + "type": "image", + "image_path": "c7db2094333dbd694e3187e960e41c373cdd19c63a9a79b4c9f11b863cb57703.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 185, + 89, + 192 + ], + "lines": [ + { + "bbox": [ + 73, + 185, + 89, + 192 + ], + "spans": [ + { + "bbox": [ + 73, + 185, + 89, + 192 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 72, + 165, + 183 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 165, + 183 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 165, + 183 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 165, + 183 + ], + "type": "image", + "image_path": "67f8a8b119b2fa73b73cd72dec6e62d86ec0cf7b63060019ab1edf71c232e922.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 185, + 148, + 191 + ], + "lines": [ + { + "bbox": [ + 128, + 185, + 148, + 191 + ], + "spans": [ + { + "bbox": [ + 128, + 185, + 148, + 191 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 72, + 222, + 182 + ], + "blocks": [ + { + "bbox": [ + 167, + 72, + 222, + 182 + ], + "lines": [ + { + "bbox": [ + 167, + 72, + 222, + 182 + ], + "spans": [ + { + "bbox": [ + 167, + 72, + 222, + 182 + ], + "type": "image", + "image_path": "14f04c62e0aba61137c0a082b21ff28e4de6e2b4ca4a36c7ae7ac08ce632f3bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 185, + 202, + 191 + ], + "lines": [ + { + "bbox": [ + 187, + 185, + 202, + 191 + ], + "spans": [ + { + "bbox": [ + 187, + 185, + 202, + 191 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 223, + 72, + 278, + 182 + ], + "blocks": [ + { + "bbox": [ + 223, + 72, + 278, + 182 + ], + "lines": [ + { + "bbox": [ + 223, + 72, + 278, + 182 + ], + "spans": [ + { + "bbox": [ + 223, + 72, + 278, + 182 + ], + "type": "image", + "image_path": "cfbcc05cb16ffb4f5d35e02fb4fb08d9e278ff79d5207152211ed57894715c25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 185, + 269, + 192 + ], + "lines": [ + { + "bbox": [ + 234, + 185, + 269, + 192 + ], + "spans": [ + { + "bbox": [ + 234, + 185, + 269, + 192 + ], + "type": "text", + "content": "Groundtruth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 192, + 286, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 192, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 286, + 216 + ], + "type": "text", + "content": "Figure 5: Visual results of blind image inpainting on " + }, + { + "bbox": [ + 47, + 192, + 286, + 216 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 192, + 286, + 216 + ], + "type": "text", + "content": " resolution images from the Places [48] dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 62, + 263, + 270, + 331 + ], + "blocks": [ + { + "bbox": [ + 47, + 227, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 227, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 287, + 262 + ], + "type": "text", + "content": "Table 2: Blind image inpainting based on detected corruption masks by different methods on " + }, + { + "bbox": [ + 47, + 227, + 287, + 262 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 227, + 287, + 262 + ], + "type": "text", + "content": " images from Places. MAT [27] is finetuned (MAT-F) for fair comparison." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 263, + 270, + 331 + ], + "lines": [ + { + "bbox": [ + 62, + 263, + 270, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 270, + 331 + ], + "type": "table", + "html": "
PhaseCorruption DetectionImage Inpainting
Corruption DetectionImage InpaintingACC ↑IoU ↑PSNR ↑SSIM ↑
GroundtruthMAT-F--25.340.852
Segmenter [37]MAT-F0.9740.96223.450.830
VCNet [42]MAT-F0.9750.96323.250.829
OursMAT-F0.9800.97024.530.844
VCNetVCNet0.9750.96321.490.764
OursOurs0.9800.97025.280.846
", + "image_path": "f4c76c69c59fb18b1eaa0aecf826284aa5f5ca66a707d8e6216afff498225627.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 422 + ], + "type": "text", + "content": "different corruption patterns. Table 4 presents the results of our model as well as VCNet by training them on Places dataset while performing testing directly on unseen corruption patterns such as random/constant noise and image content from CelebA-HQ [22]. It shows that our model consistently outperforms VCNet in all cases, which reveals the better generalizability of our model over VCNet." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 427, + 247, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 247, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 247, + 440 + ], + "type": "text", + "content": "4.3. Downstream Image Restoration Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 445, + 180, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 180, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 180, + 458 + ], + "type": "text", + "content": "4.3.1 Blind Image Inpainting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "type": "text", + "content": "In the experiments of blind image inpainting, we compare our model with VCNet and MPRNet [45], a state-of-the-art approach for image restoration. Table 3 presents the inpainting performance on images of " + }, + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "type": "text", + "content": " resolution. Our method achieves the best performance in terms of all metrics and outperforms the other two methods by a large margin. Besides, the qualitative comparison in Figure 4 also demonstrates that our model is able to restore higher-quality images than other two methods, benefiting from the precise corruption detection by the proposed hierarchical contrastive learning mechanism as well as the integrated restoration framework. Note that we also provide the user study in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "High-resolution blind image inpainting. We further conduct experiments on images of " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " resolution in Places [48] dataset for blind image inpainting. As shown in Table 2, our model outperforms VCNet substantially, consistent with the comparison in the case of " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " resolution. Besides, the qualitative comparison in Figure 5 also validates the superiority of our model. More visual results are presented in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 403, + 156 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 403, + 156 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 403, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 403, + 156 + ], + "type": "image", + "image_path": "0a2e0d3278e574a8f0110a1165b8df986c110a917bab6ef44862ca5ede8e3ac2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 156, + 367, + 162 + ], + "lines": [ + { + "bbox": [ + 354, + 156, + 367, + 162 + ], + "spans": [ + { + "bbox": [ + 354, + 156, + 367, + 162 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 403, + 71, + 445, + 112 + ], + "blocks": [ + { + "bbox": [ + 403, + 71, + 445, + 112 + ], + "lines": [ + { + "bbox": [ + 403, + 71, + 445, + 112 + ], + "spans": [ + { + "bbox": [ + 403, + 71, + 445, + 112 + ], + "type": "image", + "image_path": "fc3fdb375c5dfb549612ba7218098f5f9fa8f7ff0ca88ddc8df3efd8bce30454.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 404, + 113, + 445, + 156 + ], + "blocks": [ + { + "bbox": [ + 404, + 113, + 445, + 156 + ], + "lines": [ + { + "bbox": [ + 404, + 113, + 445, + 156 + ], + "spans": [ + { + "bbox": [ + 404, + 113, + 445, + 156 + ], + "type": "image", + "image_path": "16a7579b21d5832bf851d6ff4cf78f4d36c41199721faba9f668a08809bd06c2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 156, + 432, + 162 + ], + "lines": [ + { + "bbox": [ + 416, + 156, + 432, + 162 + ], + "spans": [ + { + "bbox": [ + 416, + 156, + 432, + 162 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 446, + 71, + 489, + 112 + ], + "blocks": [ + { + "bbox": [ + 446, + 71, + 489, + 112 + ], + "lines": [ + { + "bbox": [ + 446, + 71, + 489, + 112 + ], + "spans": [ + { + "bbox": [ + 446, + 71, + 489, + 112 + ], + "type": "image", + "image_path": "3574c41333a965fce6418e4ee32a65807276fc01b2bdf39b2399dff2f3c988b4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 446, + 113, + 488, + 156 + ], + "blocks": [ + { + "bbox": [ + 446, + 113, + 488, + 156 + ], + "lines": [ + { + "bbox": [ + 446, + 113, + 488, + 156 + ], + "spans": [ + { + "bbox": [ + 446, + 113, + 488, + 156 + ], + "type": "image", + "image_path": "f5e4f7e9b338d8e28c9412e022ca73c3d38f347ca86324dbc3b790d1a217e084.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 156, + 474, + 162 + ], + "lines": [ + { + "bbox": [ + 462, + 156, + 474, + 162 + ], + "spans": [ + { + "bbox": [ + 462, + 156, + 474, + 162 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 490, + 71, + 533, + 112 + ], + "blocks": [ + { + "bbox": [ + 490, + 71, + 533, + 112 + ], + "lines": [ + { + "bbox": [ + 490, + 71, + 533, + 112 + ], + "spans": [ + { + "bbox": [ + 490, + 71, + 533, + 112 + ], + "type": "image", + "image_path": "79b72508848d2cfd23c53d234ed133ccdc14fe160da6c573bdd8fece9de2c16b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 490, + 113, + 532, + 156 + ], + "blocks": [ + { + "bbox": [ + 490, + 113, + 532, + 156 + ], + "lines": [ + { + "bbox": [ + 490, + 113, + 532, + 156 + ], + "spans": [ + { + "bbox": [ + 490, + 113, + 532, + 156 + ], + "type": "image", + "image_path": "3e7be72b3345aed32b636eaf132987eebbd02f46dbca43d081be3496bf963c8d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 156, + 524, + 162 + ], + "lines": [ + { + "bbox": [ + 498, + 156, + 524, + 162 + ], + "spans": [ + { + "bbox": [ + 498, + 156, + 524, + 162 + ], + "type": "text", + "content": "Groundtruth" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 310, + 200, + 375, + 306 + ], + "blocks": [ + { + "bbox": [ + 306, + 163, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 306, + 163, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 545, + 198 + ], + "type": "text", + "content": "Figure 6: Visual results of bidirectional blind image inpainting on FFHQ [23]. More results are shown in the supplementary material." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 200, + 375, + 306 + ], + "lines": [ + { + "bbox": [ + 310, + 200, + 375, + 306 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 375, + 306 + ], + "type": "image", + "image_path": "c2b504c647ee84f25366488c465ffe521172168670b0ef9281410445b4bce304.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 308, + 355, + 314 + ], + "lines": [ + { + "bbox": [ + 340, + 308, + 355, + 314 + ], + "spans": [ + { + "bbox": [ + 340, + 308, + 355, + 314 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 320, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 306, + 320, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 545, + 357 + ], + "type": "text", + "content": "Figure 7: Visual comparison of both corruption detection and blind image inpainting over unseen corruptions. More examples could be found in the supplementary material." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 375, + 200, + 429, + 306 + ], + "blocks": [ + { + "bbox": [ + 375, + 200, + 429, + 306 + ], + "lines": [ + { + "bbox": [ + 375, + 200, + 429, + 306 + ], + "spans": [ + { + "bbox": [ + 375, + 200, + 429, + 306 + ], + "type": "image", + "image_path": "4a6aa392cf6a13f7b396fe25824a44fd4538a9f8d66321530f462aa899130b23.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 308, + 412, + 314 + ], + "lines": [ + { + "bbox": [ + 393, + 308, + 412, + 314 + ], + "spans": [ + { + "bbox": [ + 393, + 308, + 412, + 314 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 430, + 200, + 483, + 306 + ], + "blocks": [ + { + "bbox": [ + 430, + 200, + 483, + 306 + ], + "lines": [ + { + "bbox": [ + 430, + 200, + 483, + 306 + ], + "spans": [ + { + "bbox": [ + 430, + 200, + 483, + 306 + ], + "type": "image", + "image_path": "8b4ca0f151db8720bea077140a6c69f7f1894b6ac0d0d6f6660b7e8259b0d473.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 308, + 463, + 314 + ], + "lines": [ + { + "bbox": [ + 449, + 308, + 463, + 314 + ], + "spans": [ + { + "bbox": [ + 449, + 308, + 463, + 314 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 485, + 200, + 537, + 306 + ], + "blocks": [ + { + "bbox": [ + 485, + 200, + 537, + 306 + ], + "lines": [ + { + "bbox": [ + 485, + 200, + 537, + 306 + ], + "spans": [ + { + "bbox": [ + 485, + 200, + 537, + 306 + ], + "type": "image", + "image_path": "0e1fdc50ed20d77c92d62ed5dac90de8d3ba7217c5c81759be4474b0fdd9c183.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 494, + 308, + 527, + 314 + ], + "lines": [ + { + "bbox": [ + 494, + 308, + 527, + 314 + ], + "spans": [ + { + "bbox": [ + 494, + 308, + 527, + 314 + ], + "type": "text", + "content": "Groundtruth" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 308, + 403, + 545, + 520 + ], + "blocks": [ + { + "bbox": [ + 306, + 365, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 306, + 365, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 545, + 399 + ], + "type": "text", + "content": "Table 3: Performance comparison for blind image inpainting in terms of four evaluation metrics on three benchmark datasets." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 403, + 545, + 520 + ], + "lines": [ + { + "bbox": [ + 308, + 403, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 403, + 545, + 520 + ], + "type": "table", + "html": "
DatasetFFHQ [23]ImageNet [9]Places [48]
Mask ratio (%)0-2020-4040-600-2020-4040-600-2020-4040-60
PSNR ↑MPRNet33.9326.7522.0633.7626.2621.5033.3725.5921.08
VCNet29.8624.5120.5129.1323.3919.4229.6623.1919.18
Ours34.7927.9823.5433.9126.4021.8034.1126.2021.66
SSIM ↑MPRNet0.9680.8980.7990.9650.8820.7620.9630.8720.743
VCNet0.9330.8490.7360.9240.8180.6780.9420.8330.688
Ours0.9710.9100.8240.9660.8840.7670.9690.8870.746
FID ↓MPRNet3.3712.2029.743.7117.8053.384.1517.6144.73
VCNet6.8014.1728.905.6921.4859.115.9318.6240.21
Ours2.377.2815.482.6912.3541.372.8811.1031.33
LPIPS ↓MPRNet0.0270.0960.2120.0330.1230.2690.0330.1310.283
VCNet0.0440.1030.1880.0450.1280.2460.0460.1380.261
Ours0.0190.0650.1390.0250.0980.2260.0240.0950.221
", + "image_path": "608668cc68c469e7dc4478ee0114659091ccdf0dba5df69c7b2a831b7e1604c0.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "type": "text", + "content": "Bidirectional blind image inpainting. When real natural images are used as the noise content for corruption, we can perform bidirectional image restoration: restore either a complete noise image or a complete background image by reversing the corruption mask. It is quite challenging to perform well in inpainting of both directions. The visual results in Figure 6 show that our model performs more robust than VCNet in both corruption detection and image inpainting." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Generalization on unseen corruption patterns. Using the same experimental settings as evaluating the generalizability for corruption detection, we also validate the generalizability of our model across unseen corruption patterns on image inpainting. Table 5 shows the image inpainting performance of both our model and VCNet on unseen corruption patterns during training. The experimental results indi" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12082" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 113, + 204, + 205 + ], + "blocks": [ + { + "bbox": [ + 52, + 77, + 205, + 111 + ], + "lines": [ + { + "bbox": [ + 52, + 77, + 205, + 111 + ], + "spans": [ + { + "bbox": [ + 52, + 77, + 205, + 111 + ], + "type": "text", + "content": "Table 4: Generalizability over unseen corruptions for corruption detection. Models are trained on Places." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 113, + 204, + 205 + ], + "lines": [ + { + "bbox": [ + 53, + 113, + 204, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 113, + 204, + 205 + ], + "type": "table", + "html": "
Corruption patternRandom constantCelebA-HQ [22]
Mask ratio (%)0-3030-600-3030-60
Acc ↑VCNet0.9810.9770.9750.974
Ours0.9900.9890.9860.986
F1 ↑VCNet0.9880.9780.9840.976
Ours0.9940.9900.9920.988
BCE ↓VCNet0.0500.0780.0630.068
Ours0.0500.0490.0620.054
IoU ↑VCNet0.9770.9600.9690.954
Ours0.9880.9810.9830.976
", + "image_path": "2e3a4c52c248676b13cb52fdeb2eb0b6abcb72e9b585165b3b9eb3f53e100bd6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 217, + 117, + 392, + 206 + ], + "blocks": [ + { + "bbox": [ + 208, + 79, + 400, + 114 + ], + "lines": [ + { + "bbox": [ + 208, + 79, + 400, + 114 + ], + "spans": [ + { + "bbox": [ + 208, + 79, + 400, + 114 + ], + "type": "text", + "content": "Table 5: Generalizability over unseen corruptions for blind inpainting. Models are trained on Places." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 117, + 392, + 206 + ], + "lines": [ + { + "bbox": [ + 217, + 117, + 392, + 206 + ], + "spans": [ + { + "bbox": [ + 217, + 117, + 392, + 206 + ], + "type": "table", + "html": "
Corruption PatternRandom constantCelebA-HQ [22]
Mask ratio (%)0-2020-4040-600-2020-4040-60
PSNR ↑VCNet30.4824.3319.3029.7723.2118.96
Ours37.6929.0123.0534.6026.5221.82
SSIM ↑VCNet0.9570.8760.7320.9470.8400.685
Ours0.9850.9330.8180.9720.8940.772
FID ↓VCNet6.4116.8442.545.4717.7541.19
Ours1.807.4524.482.6510.5030.42
LPIPS ↓VCNet0.0510.1310.2680.0410.1280.256
Ours0.0130.0630.1820.0210.0890.214
", + "image_path": "9fd04054a49db703dbcfb43ea502b54e3f2aba3c09539e76e5e90fefef1fb5f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 406, + 108, + 542, + 211 + ], + "blocks": [ + { + "bbox": [ + 402, + 79, + 542, + 102 + ], + "lines": [ + { + "bbox": [ + 402, + 79, + 542, + 102 + ], + "spans": [ + { + "bbox": [ + 402, + 79, + 542, + 102 + ], + "type": "text", + "content": "Table 6: Ablation study on hierarchical interaction mechanism." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 406, + 108, + 542, + 211 + ], + "lines": [ + { + "bbox": [ + 406, + 108, + 542, + 211 + ], + "spans": [ + { + "bbox": [ + 406, + 108, + 542, + 211 + ], + "type": "table", + "html": "
Mask ratio (%)0-3030-60
Acc ↑w/o inter-stage consistency0.9690.965
w/o sample selection0.9750.974
Complete model0.9780.976
F1 ↑w/o inter-stage consistency0.9800.969
w/o sample selection0.9840.976
Complete model0.9860.978
BCE ↓w/o inter-stage consistency0.1020.115
w/o sample selection0.0940.101
Complete model0.0900.095
IoU ↑w/o inter-stage consistency0.9620.940
w/o sample selection0.9720.957
Complete model0.9750.959
", + "image_path": "a8f10d9405a7b3a5e44eb182704b944f7f0d5feac383e6d15c39b770cdcd6802.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 217, + 172, + 306 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 172, + 306 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 172, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 172, + 306 + ], + "type": "image", + "image_path": "54df3df2d71c4b760caa39a7c7eec07f063f4107a4e885e34831c4521a0fd54d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 306, + 285, + 319 + ], + "lines": [ + { + "bbox": [ + 48, + 306, + 285, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 285, + 319 + ], + "type": "text", + "content": "Figure 8: Hierarchical vs single-stage contrastive learning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 176, + 217, + 297, + 306 + ], + "blocks": [ + { + "bbox": [ + 176, + 217, + 297, + 306 + ], + "lines": [ + { + "bbox": [ + 176, + 217, + 297, + 306 + ], + "spans": [ + { + "bbox": [ + 176, + 217, + 297, + 306 + ], + "type": "image", + "image_path": "d832645ba19703dcc741d4513b50734e0e36d4d86a0ed32de12fae17216f6925.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 99, + 342, + 234, + 412 + ], + "blocks": [ + { + "bbox": [ + 53, + 330, + 280, + 340 + ], + "lines": [ + { + "bbox": [ + 53, + 330, + 280, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 280, + 340 + ], + "type": "text", + "content": "Table 7: Performance evaluation on watermark removal." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 342, + 234, + 412 + ], + "lines": [ + { + "bbox": [ + 99, + 342, + 234, + 412 + ], + "spans": [ + { + "bbox": [ + 99, + 342, + 234, + 412 + ], + "type": "table", + "html": "
MetricsPSNR↑SSIM↑FID↓
SIRF [46]34.630.9780.071
BS2AM [6]34.880.9790.028
DHAN [8]37.670.9860.062
BVMR [19]38.280.9850.018
Split then Refine [7]41.270.9910.011
VCNet [42]32.660.9630.032
Ours41.880.9920.007
", + "image_path": "3e0b4d47e0d74bf14341ce40a9cbc5c80041b242548d919792daac11953cc569.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "content": "cate that our model performs much better than VCNet. Furthermore, the qualitative comparison in Figure 7 also validates such better generalizability of our model over VCNet on both corruption detection and image inpainting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 473, + 200, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 200, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 200, + 485 + ], + "type": "text", + "content": "4.3.2 Image Watermark Removal" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 487, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 595 + ], + "type": "text", + "content": "In this set of experiments, we further evaluate the performance of our model on LOGO-30K dataset [7] for watermark removal. The comparison in Table 7 shows that our model performs best in terms of all metrics and compares favorably with other specialized methods for watermark removal, which demonstrates the robustness of our model across different image restoration tasks. Moreover, the visual results in Figure 9 also validate the superiority of our model over other methods in watermark removal." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 141, + 612 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 689 + ], + "type": "text", + "content": "Effect of hierarchical contrastive learning. To investigate the effectiveness of the proposed hierarchical contrastive learning framework, we compare its performance with that of single-stage contrastive learning. The results in Figure 8 demonstrate the distinct advantage of the proposed hierarchical contrastive learning framework." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Effect of hierarchical interaction mechanism. We further conduct experiments to investigate the effect of hierarchi" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 219, + 354, + 267 + ], + "blocks": [ + { + "bbox": [ + 307, + 219, + 354, + 267 + ], + "lines": [ + { + "bbox": [ + 307, + 219, + 354, + 267 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 354, + 267 + ], + "type": "image", + "image_path": "9894bef6b72975635c568964f781b6488187f107bdc193288ebeab30fba341c0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 356, + 219, + 402, + 267 + ], + "blocks": [ + { + "bbox": [ + 356, + 219, + 402, + 267 + ], + "lines": [ + { + "bbox": [ + 356, + 219, + 402, + 267 + ], + "spans": [ + { + "bbox": [ + 356, + 219, + 402, + 267 + ], + "type": "image", + "image_path": "987efa1bf8e7ca5ac201355be1dd3649f461cb368d370668e32ef4e179113436.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 403, + 219, + 449, + 267 + ], + "blocks": [ + { + "bbox": [ + 403, + 219, + 449, + 267 + ], + "lines": [ + { + "bbox": [ + 403, + 219, + 449, + 267 + ], + "spans": [ + { + "bbox": [ + 403, + 219, + 449, + 267 + ], + "type": "image", + "image_path": "bfa12fe0364d403f505267c978999fcb173be067cb7847ad4290cebef70617c8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 450, + 219, + 498, + 267 + ], + "blocks": [ + { + "bbox": [ + 450, + 219, + 498, + 267 + ], + "lines": [ + { + "bbox": [ + 450, + 219, + 498, + 267 + ], + "spans": [ + { + "bbox": [ + 450, + 219, + 498, + 267 + ], + "type": "image", + "image_path": "4936d45df96aa9456cce80ad225639fc4d0fe4c8ad33c6862f927723abdc0c69.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 498, + 219, + 545, + 267 + ], + "blocks": [ + { + "bbox": [ + 498, + 219, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 498, + 219, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 498, + 219, + 545, + 267 + ], + "type": "image", + "image_path": "2f2b7625a05a2b6f362522758840a98ebc05707491a377af433ea5e572ab5754.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 307, + 268, + 354, + 314 + ], + "blocks": [ + { + "bbox": [ + 307, + 268, + 354, + 314 + ], + "lines": [ + { + "bbox": [ + 307, + 268, + 354, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 354, + 314 + ], + "type": "image", + "image_path": "d2f651141f9bbe7cd2261c16a0523f7dafa58897c40d86433018822cec0eb2c3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 316, + 338, + 321 + ], + "lines": [ + { + "bbox": [ + 325, + 316, + 338, + 321 + ], + "spans": [ + { + "bbox": [ + 325, + 316, + 338, + 321 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 323, + 545, + 347 + ], + "lines": [ + { + "bbox": [ + 306, + 323, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 545, + 347 + ], + "type": "text", + "content": "Figure 9: Visual comparison with the state-of-the-art model Split then Refine [7] and VCNet for watermark removal." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 356, + 268, + 402, + 314 + ], + "blocks": [ + { + "bbox": [ + 356, + 268, + 402, + 314 + ], + "lines": [ + { + "bbox": [ + 356, + 268, + 402, + 314 + ], + "spans": [ + { + "bbox": [ + 356, + 268, + 402, + 314 + ], + "type": "image", + "image_path": "fe9d70bb59f802e763c8c97388c31c4d5940cea248027d6244af6a6b15231581.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 316, + 388, + 321 + ], + "lines": [ + { + "bbox": [ + 369, + 316, + 388, + 321 + ], + "spans": [ + { + "bbox": [ + 369, + 316, + 388, + 321 + ], + "type": "text", + "content": "VCNet" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 403, + 268, + 449, + 314 + ], + "blocks": [ + { + "bbox": [ + 403, + 268, + 449, + 314 + ], + "lines": [ + { + "bbox": [ + 403, + 268, + 449, + 314 + ], + "spans": [ + { + "bbox": [ + 403, + 268, + 449, + 314 + ], + "type": "image", + "image_path": "5b3ed53b4e411fc86705ea930fde9eb86e766cd3e2318957ab0783d18bf59ae2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 316, + 446, + 321 + ], + "lines": [ + { + "bbox": [ + 406, + 316, + 446, + 321 + ], + "spans": [ + { + "bbox": [ + 406, + 316, + 446, + 321 + ], + "type": "text", + "content": "Split then Refine" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 450, + 268, + 498, + 314 + ], + "blocks": [ + { + "bbox": [ + 450, + 268, + 498, + 314 + ], + "lines": [ + { + "bbox": [ + 450, + 268, + 498, + 314 + ], + "spans": [ + { + "bbox": [ + 450, + 268, + 498, + 314 + ], + "type": "image", + "image_path": "939c0f0aaedb1a7a7f467e93b9a57e46b396092f152dde16edb62b6986dac88e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 316, + 480, + 321 + ], + "lines": [ + { + "bbox": [ + 468, + 316, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 468, + 316, + 480, + 321 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 498, + 268, + 545, + 314 + ], + "blocks": [ + { + "bbox": [ + 498, + 268, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 498, + 268, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 498, + 268, + 545, + 314 + ], + "type": "image", + "image_path": "e2c0b7f9e4134cb596c8a54e1de01b5586c8f1ff7e643b73c5d0165f71aebaed.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 316, + 536, + 321 + ], + "lines": [ + { + "bbox": [ + 506, + 316, + 536, + 321 + ], + "spans": [ + { + "bbox": [ + 506, + 316, + 536, + 321 + ], + "type": "text", + "content": "Groundtruth" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 356, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 440 + ], + "type": "text", + "content": "cal interaction mechanism, especially the proposed 'inter-stage semantic consistency' and 'high-quality sample selection strategy'. Table 6 presents the concrete experimental results. While both techniques can boost the performance, 'inter-stage semantic consistency' is more crucial to the performance due to its essential merit: semantic propagation between stages by feature reusing." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "text", + "content": "In this work, we have designed a novel method, namely hierarchical contrastive learning framework, which can automatically detect corruption masks in pixel level and thus allows for blind image corruption restoration without known corruption masks. Extensive quantitative and qualitative comparisons have demonstrated the superior performance of our method over other methods for various corruption restoration tasks and its well generalization ability across different corruption patterns." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 578, + 416, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 416, + 591 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 416, + 591 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the NSFC fund (Grant No. U2013210, 62006060, 62176077), in part by the Shenzhen Key Technical Project under Grant 2022N001, 2020N046, in part by the Guangdong Basic and Applied Basic Research Foundation under Grant (Grant No. 2022A1515010306), in part by Shenzhen Fundamental Research Program (Grant No. JCYJ20210324132210025, JCYJ20220818102415032), and in part by the Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies (Grant No. 2022B1212010005)." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12083" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 288, + 213 + ], + "type": "text", + "content": "[3] Nian Cai, Zhenghang Su, Zhineng Lin, Han Wang, Zhijing Yang, and Bingo Wing-Kuen Ling. Blind inpainting using the fully convolutional neural network. The Visual Computer, 33(2):249-261, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 288, + 257 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "text", + "content": "[5] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9):1200-1212, 2004. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 345 + ], + "type": "text", + "content": "[6] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 390 + ], + "type": "text", + "content": "[7] Xiaodong Cun and Chi-Man Pun. Split then refine: stacked attention-guided resunets for blind single image visible watermark removal. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1184-1192, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "type": "text", + "content": "[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10680-10687, 2020. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[10] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11368, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[12] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, pages 341-346, 2001. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[13] Xin Feng, Wenjie Pei, Fengjun Li, Fanglin Chen, David Zhang, and Guangming Lu. Generative memory-guided semantic reasoning model for image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "text", + "content": "[15] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[16] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14134-14143, October 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 271 + ], + "type": "text", + "content": "[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "text", + "content": "[19] Amir Hertz, Sharon Fogel, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Blind visual motif removal from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6858–6867, 2019. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": "[20] Yong Shi Jie Yang, Zhiquan Qi. Learning to incorporate structure knowledge for image inpainting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12605-12612, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "text", + "content": "[21] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[22] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 535 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 558 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[26] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 657 + ], + "type": "text", + "content": "[27] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10758-10768, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[28] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12084" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision, pages 85-100, 2018. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 194 + ], + "type": "text", + "content": "[30] Hongyu Liu, Bin Jiang, Yibing Song, Wei Huang, and Chao Yang. Rethinking image inpainting via a mutual encoder-decoder with feature equalizations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 725-741. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "text", + "content": "[31] Hongyu Liu, Bin Jiang, Yi Xiao, and Chao Yang. Coherent semantic attention for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4170-4179, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "type": "text", + "content": "[32] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11347-11357, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "type": "text", + "content": "[33] Yang Liu, Jinshan Pan, and Zhixun Su. Deep blind image inpainting. In International Conference on Intelligent Science and Big Data Engineering, pages 128-141. Springer, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[35] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Qureshi, and Mehran Ebrahimi. Edgeconnect: Structure guided image inpainting using edge prediction. In The IEEE International Conference on Computer Vision Workshops, Oct 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 481 + ], + "type": "text", + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 525 + ], + "type": "text", + "content": "[37] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7262-7272, 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "type": "text", + "content": "[38] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 288, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 625 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "text", + "content": "[40] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10581-10590, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[41] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 442 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[42] Yi Wang, Ying-Cong Chen, Xin Tao, and Jiaya Jia. Vcnet: A robust approach to blind image inpainting. In European Conference on Computer Vision, pages 752-768. Springer, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[43] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10551-10560, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 252 + ], + "type": "text", + "content": "[44] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4471-4480, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "text", + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[46] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 353, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 397 + ], + "type": "text", + "content": "[47] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1438-1447, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 398, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 442 + ], + "type": "text", + "content": "[48] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 6, 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12085" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_content_list.json b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d8fa4503604e8eae7158dd8cd327f59b3da1f242 --- /dev/null +++ b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_content_list.json @@ -0,0 +1,2109 @@ +[ + { + "type": "text", + "text": "Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models", + "text_level": 1, + "bbox": [ + 204, + 130, + 767, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huaijin Pi $^{1}$ , Sida Peng $^{1}$ , Minghui Yang $^{2}$ , Xiaowei Zhou $^{1}$ , Hujun Bao $^{1*}$", + "bbox": [ + 197, + 202, + 776, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Ant Group", + "bbox": [ + 222, + 224, + 746, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 277, + 313, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents a novel approach to generating the 3D motion of a human interacting with a target object, with a focus on solving the challenge of synthesizing long-range and diverse motions, which could not be fulfilled by existing auto-regressive models or path planning-based methods. We propose a hierarchical generation framework to solve this challenge. Specifically, our framework first generates a set of milestones and then synthesizes the motion along them. Therefore, the long-range motion generation could be reduced to synthesizing several short motion sequences guided by milestones. The experiments on the NSM, COUCH, and SAMP datasets show that our approach outperforms previous methods by a large margin in both quality and diversity. The source code is available on our project page https://zju3dv.github.io/hghoi.", + "bbox": [ + 75, + 311, + 473, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 571, + 209, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene-aware motion generation [19] aims to synthesize 3D human motion given a 3D scene model to enable virtual humans to naturally wander around scenes and interact with objects, which has a variety of applications in AR/VR, filmmaking, and video games.", + "bbox": [ + 75, + 599, + 468, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike traditional motion generation methods for character control which aim to generate short or repeated motion on the fly guided by a user's control signals [56], we focus on the setting of generating long-term human-object interactions [56, 19, 76] given a starting position of the human and a target object model. This setting brings in new challenges. First, the entire approaching process and the human-object interaction should be coherent, which requires the capability of modeling long-range interaction between the human and the object. Second, in the context of content generation, the generative model should be able to synthesize diverse motions as there are many plausible ways for a real human to approach and interact with the target object.", + "bbox": [ + 75, + 675, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f698d33418202babdfca7f4b06968b8d90235acb490c959682dc07f3ffdd37e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 280, + 885, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b4ada4e7b0ca0ea6c34e2f2fa1ba622b75b166f000e246c077885f84368a241.jpg", + "image_caption": [ + "Figure 1. Genration of human-object interactions. Given an object, our method first predicts a set of milestones, where the rings indicate the positions and the humans with pink clothes represent the local poses. Then the motions are inflated between milestones. This figure shows that our method can generate diverse milestones and motions with the same object. The flow of time is shown with a color code where darker blue denotes the later frame." + ], + "image_footnote": [], + "bbox": [ + 506, + 433, + 885, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods for motion synthesis can be roughly characterized into online generation and offline generation. Most online methods [56, 19, 76] focus on real-time control of characters. Given a target object, they generally use auto-regressive models to recurrently generate future motions by feeding their predictions. Although they have been widely used for interactive scenarios like video games, their motion quality is not satisfactory enough for long-term generation [63]. A plausible cause is the error accumulation in the auto-regressive process, where the errors in previous predictions are fed back as the model input, as discussed in [40, 26, 45, 46, 22]. To improve the motion quality,", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "some recent offline methods [6, 65, 64, 63] employ a multistage framework, which first generates the trajectory and then synthesizes motions. TDNS [63] generates paths by combining the cVAE model [33] and deterministic planning methods like A* [18]. Although this strategy can produce reasonable paths, the path diversity is limited, as demonstrated by our experimental results in Sec. 4.4.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel offline approach for synthesizing long-term and diverse human-object interactions. Our innovation lies in a hierarchical generation strategy, which first predicts a set of milestones and then generates human motions between milestones. Fig. 1 illustrates the basic idea. Specifically, given the starting position and the target object, we design a milestone generation module to synthesize a set of milestones along the motion trajectory, each of which encodes the local pose and indicates the transition point during the human movement. Based on these milestones, we employ a motion generation module to produce the full motion sequence. Thanks to the milestones, we simplify the long-sequence generation into synthesizing several short motion sequences. Furthermore, the local pose at each milestone is generated by a transformer that considers the global dependency [62], leading to temporally consistent results, which further contribute to coherent motions.", + "bbox": [ + 75, + 198, + 470, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to our hierarchical generation framework, we further exploit diffusion models [53, 23, 54] to synthesize human-object interactions. Previous diffusion models for motion synthesis [31, 59] combine transformer [62] and Denoising Diffusion Probabilistic Model (DDPM) [23]. Directly applying them to our setting is prohibitively comput-intensive due to the long motion sequences and may lead to the GPU memory explosion [50]. Because our hierarchical generation framework converts the long-term generation to the synthesis of several short sequences, the required GPU memory is reduced to the same level of short-term motion generation. Therefore, we can efficiently leverage the transformer DDPM to synthesize long-term motion sequences, which improves the generation quality.", + "bbox": [ + 75, + 455, + 468, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our design choices on the NSM [56], COUCH [76], and SAMP [19] datasets with extensive experiments. On these datasets, our hierarchical framework outperforms previous methods significantly in both motion quality and diversity.", + "bbox": [ + 75, + 667, + 468, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 760, + 218, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Motion synthesis", + "text_level": 1, + "bbox": [ + 76, + 785, + 243, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion synthesis is a long-standing problem in computer graphics and vision [68, 51, 42, 9]. With the rapid development of deep learning, recent works have applied neural networks to motion synthesis [14, 40, 26, 72]. Some methods are deterministic [28, 81, 44, 67] while others try to predict stochastic motions by VAE [70, 1, 71, 79, 45] and", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GAN [4, 34]. To further improve the performance, some works apply GCN [67, 10, 66, 35, 39] or transformers [5] to extract the features from the human skeleton. To handle the ambiguity of human motion, some works [26, 55] propose to employ phase signals to guide the motions. Recent works start to consider the scene context [56, 6]. NSM [56] is the first work that aims at synthesizing human motions with object-level interactions with specific action labels. Based on NSM, SAMP [19] applies cVAE to predict diverse motions. These works [56, 19, 76] mainly focus on the interaction with one or two objects while others [6, 65] generate motions with a full scene (including the information about floors and walls) as the input. [6] predicts future motions of the given motion sequence by first predicting the trajectory and then generating motions based on a 2D image of the scene. Similar to this pipeline, [65] applies GAN [16]. Furthermore, [64] proposes a framework that first places pose on a human-provided path and then synthesizes motions in a 3D scene with a full-scene scan. [63] combines $\\mathbf{A}^*$ and cVAE to generate diverse human motions. To further control human motion, [80] employs the gaze, and COUCH [76] explicitly models the hand contacts to guide the prediction. Some methods [7, 21] also enable physically simulated characters to perform scene interaction tasks including sitting [7, 21] and carrying boxes [21]. Some works also focus on grasp [58, 69], manipulation [73] and the interaction with dynamic objects [57].", + "bbox": [ + 496, + 90, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ours vs. others. This work follows the setting of [56, 19] and focuses on object-level interaction. In contrast to [56, 19] which generate motions mainly based on auto-regressive models, we design a hierarchical framework to synthesize motions. Different from [6, 64], our work focuses on much more long-term generation (longer than 10 seconds) while [6] is 2 seconds and [64] is 6 seconds. Instead of planning the path auto-regressively with an extra network to generate diverse trajectories like [63], our method directly predicts a set of milestones to describe the approaching process which is inherently diverse. In addition, most methods [19, 76, 64, 63] rely on cVAE to generate stochastic motions while we exploit DDPM [23] to synthesize trajectories and motions.", + "bbox": [ + 496, + 518, + 890, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion models", + "text_level": 1, + "bbox": [ + 500, + 741, + 666, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [53] are a class of likelihood-based methods that generate samples by gradually removing noise from a signal. Then, [23, 54] develop the diffusion models for high-quality image generation. To control the generated results, [12] proposes classifier guidance for trading off diversity for fidelity. Later, the classifier-free model [24] achieves better results [48] in text-conditioned image generation. In addition, diffusion models have been successfully applied to other domains like the generation of videos", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15062", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e94b5e2623b8eb9c14a45cbda549c0033009a05eed1a24688830e82bd0525162.jpg", + "image_caption": [ + "Figure 2. Overview of our pipeline. Our pipeline consists of three components: First, the goal pose is synthesized given the object. Then, a number of milestones with local poses are predicted based on the goal pose. Finally, the trajectory and the full motion sequences are inflated between the milestones." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 885, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[52, 25] and 3D contents [47].", + "bbox": [ + 76, + 272, + 277, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are some works [74, 59, 31, 38] that apply DDPM to synthesize motions. [3, 8] also explore latent diffusion models [50] for motion generation. However, most works focus on text-conditioned [74, 11], audio-driven [82] and music-driven [60, 82] motion generation while we target human-object interaction. In this work, we apply transformer DDPM [23] to a multi-stage framework, which separately generates trajectories and synthesizes motions.", + "bbox": [ + 75, + 286, + 470, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are some concurrent works [27] that apply DDPM to synthesize motions in a scene. However, [27] is designed for short-term motion generation (around 2 seconds) while we target long-term human motion generation (longer than 10 seconds). Different from existing work [59, 27], we employ DDPM in a multi-stage framework, where trajectories and motions are separately predicted.", + "bbox": [ + 75, + 407, + 468, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 526, + 174, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the object $\\mathbf{I}$ and the starting point $\\mathbf{s}$ , our goal is to synthesize 3D human motions $\\{(\\mathbf{r}_i, \\pmb{\\theta}_i)\\}_{i=1}^N$ with human-object interactions. $\\mathbf{r}_i$ is the root trajectory, and the $\\pmb{\\theta}_i$ indicates local pose at $i$ -th frame.", + "bbox": [ + 75, + 551, + 468, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design a hierarchical motion generation framework as shown in Fig. 2. First, we employ GoalNet[19] to predict an interaction goal on the object. Then, we generate the goal pose (Sec. 3.1) to explicitly model the human-object interaction. Next, our milestone generation module (Sec. 3.2) estimates the length of the milestones, produces the trajectory of the milestones from the starting point to the goal, and places milestone poses. Therefore, the long-range motion generation is decomposed into combinations of short-range motion synthesis. Finally, we design a motion generation module (Sec. 3.3) to synthesize the trajectory between the milestones and infill the motions.", + "bbox": [ + 75, + 612, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Goal pose generation", + "text_level": 1, + "bbox": [ + 76, + 801, + 274, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We call the pose in which a person interacts with an object and remains stationary a goal pose. To synthesize diverse human motion, we first generate a goal pose interacting with the object following [58, 69, 63]. Most methods [78, 75, 20] generate human poses using the cVAE model.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd7b467a6441efe6754b756c098c3fe2730a3f5db2ca96632998109e8c5fe610.jpg", + "image_caption": [ + "Figure 3. Part VQ-VAE. Part VQ-VAE first splits the skeleton into multiple parts and learns the codebooks separately. The composition of different parts is subsequently modeled with the autoregressive prediction model." + ], + "image_footnote": [], + "bbox": [ + 506, + 273, + 885, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "They project the poses to the standard normal distribution in the continuous space [33]. Empirically, we find that cVAE models do not perform well in our setting. To overcome this challenge, we introduce VQ-VAE [61, 49] to model the data distribution which exploits a discrete representation to cluster the data in a finite set of points [37]. We hypothesize that limited data of goal pose from the SAMP dataset [19] can always be clustered by VQ-VAE but may not be enough for learning a continuous latent space for VAE [41]. In addition, based on the observation that different human poses may share similar properties [15, 29] (e.g., humans may sit with different hand positions but the same leg positions), we split joints into $L$ ( $L = 5$ ) different non-overlapping groups like MotionPuzzle [29].", + "bbox": [ + 496, + 556, + 893, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quantization. As shown in Fig. 3, the goal pose $\\theta_{g}$ is split into separate joint groups as $\\theta_{g} = \\{\\theta_{gi}\\}_{i=1}^{L}$ . Then a discrete codebook $\\mathbf{Z}_i$ with a list of vectors compares the output $\\hat{\\mathbf{z}}_i$ from the encoder $E_i$ to find the closest vector in Euclidean distance. The $L$ vectors with minimal distances will be concatenated and fed into a shared decoder $D$ to", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15063", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "reconstruct $\\theta_{g}$ . The loss function is defined as:", + "bbox": [ + 76, + 90, + 387, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\boldsymbol {\\theta} _ {g}, D (\\mathbf {z})\\right) = \\| \\boldsymbol {\\theta} _ {g} - D (\\mathbf {z}) \\| _ {2} ^ {2} + \\sum_ {i = 1} ^ {L} \\| s g \\left[ E _ {i} \\left(\\boldsymbol {\\theta} _ {g i}\\right) \\right] - \\mathbf {z} _ {i} \\| _ {2} ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 118, + 467, + 161 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n+ \\beta \\sum_ {i = 1} ^ {L} \\| s g [ \\mathbf {z} _ {i} ] - E _ {i} (\\boldsymbol {\\theta} _ {g i}) \\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 161, + 434, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 202, + 120, + 215 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\left[ \\mathbf {z} _ {1}, \\dots , \\mathbf {z} _ {L} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 229, + 467, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} = \\underset {\\mathbf {z} _ {i} ^ {\\prime} \\in Z _ {i}} {\\arg \\min } \\| \\mathbf {z} _ {i} ^ {\\prime} - \\hat {\\mathbf {z}} _ {i} \\|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 247, + 467, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {i} = E \\left(\\boldsymbol {\\theta} _ {g i}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 279, + 467, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the term $sg[\\cdot]$ denotes the stop-gradient operator, and $\\| sg[\\mathbf{z}_i] - E_i(\\pmb{\\theta}_{gi})\\|_2^2$ is the commitment loss [49] controlled by a weighting factor $\\beta$ .", + "bbox": [ + 76, + 304, + 467, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation. With $E_{i}$ and $D$ available, we can represent $\\pmb{\\theta}_{g} = \\{\\pmb{\\theta}_{g1},\\dots,\\pmb{\\theta}_{gL}\\}$ by a sequence of the part-based codebook indices. To be more specific, we use $E_{i}$ to extract the feature from $\\pmb{\\theta}_{gi}$ and find the closest vector $\\mathbf{z}_i\\in \\mathbf{Z}_i$ . Then we use $s_i\\in \\{0,\\ldots ,|\\mathbf{Z}_i| - 1\\}$ to indicate the index of $\\mathbf{z}_i$ in $\\mathbf{Z}_i$ . Therefore, $\\pmb{\\theta}_{g}$ can be represented by $\\mathbf{s} = \\{s_1,\\dots,s_L\\}$ .", + "bbox": [ + 76, + 367, + 467, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate a natural goal pose, we convert the problem to predict a sequence of indices that can represent $\\theta_{g}$ . We formulate the inference as a conditional auto-regressive process and employ a transformer [62] to learn to predict the distribution of possible indices [13]. The condition contains the environment around the goal $\\mathbf{O}_g$ , and the action $\\mathbf{a}_g$ . Following NSM [56] and COUCH [76], a cylindrical volume of a pre-defined radius and height is created around the goal. Within this volume, spheres are uniformly sampled and the occupancies corresponding to the object of these spheres are calculated. Then, these occupancies are flattened to form a feature vector denoted as $\\mathbf{O}_g$ . $\\mathbf{a}_g$ is a vector indicating the action type. These variables are fed into the transformer as tokens. Our target is to learn the likelihood of the sequence:", + "bbox": [ + 75, + 459, + 467, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {s} \\mid \\mathbf {O} _ {g}, \\mathbf {a} _ {g}) = \\prod_ {i} p \\left(s _ {i} \\mid s _ {< i}, \\mathbf {O} _ {g}, \\mathbf {a} _ {g}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 681, + 467, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After predicting the indices, we map them back to their corresponding codebook entries to get the quantized features $\\mathbf{z} = [\\mathbf{z}_1,\\dots ,\\mathbf{z}_L]$ , which are fed into the decoder $D$ to generate the goal pose $\\theta_{g}$ .", + "bbox": [ + 76, + 717, + 467, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Milestone generation", + "text_level": 1, + "bbox": [ + 76, + 786, + 272, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the starting pose and the goal pose, we can generate the milestone trajectory and synthesize the local poses at the milestones. Following [74, 59], we build a transformer DDPM [23] and apply it to generate the milestones for better quality. Because the length of motion data is unknown and can be arbitrary (e.g., the human could quickly", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "walk towards the chair and sit down or sit after walking around the chair slowly), we predict the length of milestones, denoted by $N$ . Then we synthesize $N$ milestone points and place the local poses on these points.", + "bbox": [ + 500, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer DDPM. Here we first briefly introduce DDPM [23], which is learned to reverse a diffusion process. Formally, the diffusion model [53] is defined as a latent variable model of the form $p_{\\theta}(\\mathbf{x}_0) \\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T}$ , where $\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)$ is the data and $\\mathbf{x}_1, \\ldots, \\mathbf{x}_T$ are the latents. $p_{\\theta}(\\mathbf{x}_{0:T})$ is formulated as a Markov chain as:", + "bbox": [ + 500, + 172, + 890, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 277, + 890, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 349, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models approximate posterior $q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)$ as a Markov chain that gradually adds Gaussian noise to the data with variance schedules given by $\\beta_t$ :", + "bbox": [ + 500, + 377, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 436, + 888, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 492, + 888, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to adding noises on $\\mathbf{x}_0$ sequentially, DDPM formulates the diffusion process as:", + "bbox": [ + 500, + 523, + 888, + 553 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 566, + 888, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{t} = 1 - \\beta_{t}$ and $\\bar{\\alpha}_{t} = \\prod_{s = 1}^{t}\\alpha_{s}$ . Hence, we can generate $\\mathbf{x}_t$ by sampling a noise $\\epsilon$ as the training data. DDPM employs a neural network to model $p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)$ and the inference is to gradually denoise $\\mathbf{x}_t$ from $t = T$ to $t = 1$ where $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ .", + "bbox": [ + 500, + 597, + 888, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Like existing works [74, 31] that apply DDPM in the motion domain, we employ a transformer decoder [62] as our architecture of DDPM. The transformer takes the noise $\\mathbf{x}_t$ and the condition $\\mathbf{C}$ as input. The condition $\\mathbf{C}$ means variables related to the generation, which will be described in detail in each subsection. The diffusion time-step $t$ is in the sinusoidal position embeddings form [62] and is injected into each block in the transformer. Different from existing works [59, 74] that assume the length of the motion sequences is already given, we insert a parallel branch to estimate the length of milestones by taking the length token $\\mathbf{H}^{tok}$ and the condition $\\mathbf{C}$ as the input, as shown in Fig. 4. The length prediction head is an MLP and predicts a multinomial distribution over discrete length indices $\\{1, 2, \\dots, N_{max}\\}$ like [17], where $N_{max}$ represents the", + "bbox": [ + 500, + 674, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15064", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efd63b3a288c09bc8835f577071820eacdaea1dc20e574f5c86a51b6d80ebdba.jpg", + "image_caption": [ + "Figure 4. Overview of transformer DDPM for milestone generation. The model first takes the length token $\\mathbf{H}^{tok}$ and the condition $\\mathbf{C}$ as the input to predict the data length. Then it constructs the noise sequence $\\mathbf{x}_T^{1:N}$ with length $N$ . In the diffusion process, it is fed with $\\mathbf{C}$ and the sequence $\\mathbf{x}_t^{1:N}$ at time-step $t$ to predict the target $\\hat{\\mathbf{x}}_0^{1:N}$ . For other sub-modules, we remove the length prediction head." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "length of the longest sequences for training. We use cross-entropy loss as the loss function. At inference time, we sample the length $N$ from the estimated distribution. Next, we construct $N$ milestones as the input to the transformer. We predict the denoised data $\\hat{\\mathbf{x}}_0$ based on the condition $\\mathbf{C}$ by the DDPM $f$ . This is formulated as $\\hat{\\mathbf{x}}_0 = f(\\mathbf{x}_t,t,\\mathbf{C})$ , and the training loss is defined as:", + "bbox": [ + 75, + 472, + 472, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\operatorname {E} _ {t \\in [ 1, T ], \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\mathbf {x} _ {0} - \\hat {\\mathbf {x}} _ {0} \\| \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 588, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our setting, we use a two-step strategy like [64, 63], where two transformer DDPMs are applied to first generate the milestone points and then synthesize the local pose at each milestone.", + "bbox": [ + 75, + 632, + 468, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generation of milestone points. The milestone points are conditioned on the object, the information of the starting point, and the goal. We then define the condition as:", + "bbox": [ + 75, + 710, + 468, + 757 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {m}} = \\left\\{\\mathbf {I} _ {s}, \\mathbf {I} _ {g}, \\mathbf {O} _ {s}, \\mathbf {O} _ {g}, \\mathbf {g}, \\mathbf {s} \\right\\}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 767, + 468, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_s$ and $\\mathbf{I}_g$ are the object representation relative to the starting point and the goal. Similar to other methods [56, 19, 76], the object representation is modeled by an $8\\times 8\\times 8$ grid with their positions and occupancies. Following NSM [56] and COUCH [76], we also explicitly model the occupancies around the starting point and the goal with $\\mathbf{O}_s$ and $\\mathbf{O}_g$ (the same form as the occupancy feature described in Sec. 3.1).", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The information of goal $\\mathbf{g}$ is defined as $\\{\\mathbf{r}_g, \\mathbf{a}_g, \\boldsymbol{\\theta}_g\\}$ , where $\\mathbf{r}_g$ means the goal position and orientation in the starting point coordinate system, $\\mathbf{a}_g$ is the target action label at the goal, and $\\boldsymbol{\\theta}_g$ is the goal pose. $\\mathbf{s} = \\{\\mathbf{a}_s, \\boldsymbol{\\theta}_s\\}$ represents the action labels $\\mathbf{a}_s$ and the pose $\\boldsymbol{\\theta}_s$ at the starting point.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The target is to predict the milestone $\\{\\mathbf{m}_1, \\dots, \\mathbf{m}_N\\}$ with length $N$ . Following NSM [56], a bi-directional scheme is employed for the milestone points generation where we predict the roots of milestones in both the starting point coordinate system and the goal coordinate system. The final predicted roots are blended from these two kinds of outputs. The milestone point $\\mathbf{m}_i$ is defined as:", + "bbox": [ + 496, + 167, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i} = \\left\\{\\mathbf {r} _ {i} ^ {b}, \\mathbf {c} _ {i}, \\mathbf {w} _ {i} \\right\\}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 282, + 890, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The representation is similar to previous work [19, 56, 76], where $\\mathbf{r}_i^b$ indicates the root position and forward directions relative to the starting point and the goal, $\\mathbf{c}_i$ is a label vector indicating the contact between the environment and the body, and we use a high-dimensional feature vector $\\mathbf{w}_i$ to encode the character state at the milestone following [56, 19, 76]. The details of these variables are provided in the supplementary material. We use transformer DDPM $f_{\\mathbf{m}}$ to predict the length and synthesize milestone points.", + "bbox": [ + 496, + 310, + 892, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generation of milestone poses. Previous works [64, 63] separately place the poses on the sparse points along the path and infill the motions between, which may lead to unnatural transitions between these poses. On the contrary, we generate the local pose at each milestone with transformer DDPM to build the temporal dependency. With the help of the generated milestone points, we can access the accurate spatial relationship $\\mathbf{I}_i$ between the object and the character and the ego-centric environment occupancies $\\mathbf{O}_i$ [76] at the $i$ -th milestone. The milestone poses also depend on the milestone state including $\\{\\mathbf{c}_i, \\mathbf{w}_i\\}$ at the $i$ -th milestone. Consequently, we define the condition $\\mathbf{C_k}$ as a combination with starting pose $\\theta_s$ , goal pose $\\theta_g$ , and frame-wise condition $\\gamma_i$ at the $i$ -th milestone as:", + "bbox": [ + 496, + 464, + 892, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {k}} = \\left\\{\\boldsymbol {\\theta} _ {s}, \\boldsymbol {\\theta} _ {g}, \\gamma_ {1}, \\dots , \\gamma_ {N} \\right\\}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 686, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad \\gamma_ {i} = \\left\\{\\mathbf {I} _ {i}, \\mathbf {O} _ {i}, \\mathbf {c} _ {i}, \\mathbf {w} _ {i} \\right\\}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 705, + 890, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The local poses at the milestones are predicted by the transformer DDPM $f_{\\mathbf{k}}$ without the length prediction head.", + "bbox": [ + 496, + 732, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Motion generation", + "text_level": 1, + "bbox": [ + 500, + 771, + 679, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of predicting motions frame-by-frame [56, 19, 76], our approach hierarchically synthesizes the full sequence based on the generated milestones. We follow [6, 64, 63] to first generate the trajectory and then synthesize the motions. Specifically, within two consecutive milestones, we first complete the trajectory. Then, the motions are inflated with the guidance of successive milestone", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "poses. The two steps are accomplished using two transformer DDPMs (described in Sec. 3.2), respectively. For each step, we carefully design the condition of DDPM to generate the target output.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory completion. For the trajectory completion between the milestones $\\mathbf{m}_i$ and $\\mathbf{m}_{i + 1}$ , we assume that it is only conditioned on the milestones and the object. Thus we define the condition as:", + "bbox": [ + 75, + 167, + 468, + 229 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {r}} = \\left\\{\\mathbf {I} _ {i}, \\mathbf {I} _ {i + 1}, \\mathbf {O} _ {i}, \\mathbf {O} _ {i + 1}, \\mathbf {m} _ {i}, \\mathbf {m} _ {i + 1}, \\mathbf {t} _ {i} ^ {i + 1} \\right\\}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 236, + 468, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_i$ indicates the object representation relative to the milestone $i$ and $\\mathbf{O}_i$ denotes the ego-centric occupancies (the same form as the occupancy feature described in Sec. 3.1) around the $i$ -th milestone like NSM [56] and COUCH [76]. $\\mathbf{m}_i$ has been shown in Eq. (13). $\\mathbf{t}_i^{i+1}$ represents the position and orientation of the $(i+1)$ -th milestone in the $i$ -th milestone's coordinate system. Between two consecutive milestones, we generate the trajectory with a length of 2 seconds, which is 61 frames. Similar hyperparameters can be found in previous methods [64, 63].", + "bbox": [ + 75, + 262, + 468, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The target output is the trajectory between two consecutive milestones. Similar to the milestone point, we synthesize the trajectory in the $j$ -th frame with a bi-directional scheme. The trajectory is composed of a set of points that have the same representation as the milestone in Eq. (13). We generate the trajectory with a transformer DDPM $f_{\\mathbf{r}}$ similar to the one in Sec. 3.2. Since we assume the trajectory length between two milestones is fixed, the DDPM $f_{\\mathbf{r}}$ does not have the length prediction head.", + "bbox": [ + 75, + 412, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion infilling. To synthesize the long-range motion, we convert a long sequence into several fixed-length short sequences with the help of milestone points and milestone poses. For a sub-sequence between the consecutive milestone poses, our goal is to generate the missing local poses over the trajectory. The generated motion has to satisfy the trajectory and naturally transits from a milestone to the next milestone. Like milestone pose generation, we use the same representation of frame-wise condition in Eq. (15). The condition is defined as:", + "bbox": [ + 75, + 566, + 468, + 717 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {p}} = \\left\\{\\boldsymbol {\\theta} ^ {1}, \\boldsymbol {\\theta} ^ {6 1}, \\gamma^ {1}, \\dots , \\gamma^ {6 1} \\right\\}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 723, + 468, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta^1$ and $\\theta^{61}$ are the local poses of two consecutive milestones. By taking these inputs, we generate smooth motions using another transformer DDPM $f_{\\mathbf{p}}$ without the length prediction head.", + "bbox": [ + 75, + 748, + 468, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 209, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 847, + 290, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train the part VQ-VAE and transformer DDPM models with the Adam optimizer [32]. All the models are trained", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with a fixed learning rate of 0.0001 with batch size 256. The remaining details are in the supplementary material.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets and evaluation metrics", + "text_level": 1, + "bbox": [ + 500, + 132, + 779, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test setting. Our experiments are conducted on the SAMP [19], COUCH [76], and NSM [56] datasets. Provided with a starting point, a starting pose, an object, and an endpoint, the virtual human is asked to approach the object, interact with it, and leave to reach the endpoint. The ablation studies are conducted on the SAMP dataset.", + "bbox": [ + 498, + 156, + 890, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following the previous method [19], we calculate the Fréchet distance (FD) between the generated and ground-truth motions to measure the motion quality. We also conduct user studies and each sequence is evaluated by at least 3 users with scores ranging from 1 to 5. In addition, we calculate the penetration ratio [64, 75, 78] and foot sliding [72, 36] to show the physical plausibility between the 3D object and the synthesized motions. We compute the Average Pairwise Distance (APD) [71, 77] to evaluate the diversity. Specifically, we calculate the APD of synthesized motion, the character's pose during object interactions, and trajectories. Following previous work [56, 19], we calculate PE (positional errors) and RE (rotational errors) to indicate the precision of object interactions. For each test object, we generate multiple sequences. More details are included in the supplementary material.", + "bbox": [ + 498, + 268, + 890, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with other methods", + "text_level": 1, + "bbox": [ + 500, + 522, + 784, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on the SAMP dataset. On the SAMP dataset [19], we compare our method with online methods SAMP [19] and MoE [72]. As our method is offline, we also implement and modify offline methods SLT [64] and TDNS [63] to our setting. For SLT [64], we employ $\\mathrm{A}^*$ [18] to plan a path and select points along the path as subgoals to form the input for SLT. More details about the implementation of [64, 63] are in the supplementary material. Since MoE often fails to finish the action, we do not calculate the penetration ratio for it. As shown in Tab. 1, our approach outperforms other methods in terms of lower FD, higher user study scores, and higher APD. Furthermore, our method achieves much higher trajectory diversity than SAMP [19]. Although TDNS [63] proposes Neural Mapper (NM) which combines $\\mathrm{A}^*$ [18] and cVAE, the diversity of the generated trajectory is inferior to our method, as indicated by $\\mathrm{APD}_T$ .", + "bbox": [ + 498, + 546, + 890, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on the COUCH dataset. As our target is to synthesize diverse motions instead of controlling the characters, we only evaluate the motion quality on the COUCH dataset [76]. Tab. 2 shows that our method outperforms all baselines. Our approach achieves much higher $\\mathrm{APD}_T$ than other methods. We observe that $\\mathrm{APD}_T$ of TDNS [63] is", + "bbox": [ + 498, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15066", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2895bbde53af9e4aeec1e7c7c1127ad7f37b54d192c3b2664d4dc23c8862be14.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFD ↓User study ↑APDM ↑APDP ↑APDT ↑PE↓RE↓Penetration↓Sliding↓
MoE [72]74.332.763.502.6352.46∞∞-1.68
SAMP [19]57.342.863.633.0563.183.444.126.981.02
SLT* [64]68.832.303.082.6640.131.771.604.281.71
TDNS* [63]46.602.903.683.4066.040.450.395.140.94
Ours22.343.624.064.5291.380.390.324.000.50
", + "bbox": [ + 209, + 88, + 767, + 183 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/749b2a24fc7ba13f416a2f57496c67348cb2febbcd40117d5cfbec117ee5443d.jpg", + "table_caption": [ + "Table 1. Quantitative results on the SAMP dataset. SLT* and TDNS* mean we modify and implement them on the SAMP dataset. The subscript “M”, “P” and “T” stand for “Motion”, “Pose” and “Trajectory”. PE and RE represent the positional error and rotation error. Sliding denotes foot sliding. $\\infty$ means the method failed to reach the goal." + ], + "table_footnote": [], + "table_body": "
MethodFD↓User study ↑APDM↑APDP↑APDT↑Penetration↓Sliding↓
NSM [56]118.982.990008.200.59
SAMP [19]160.121.940.890.184.694.940.72
COUCH [76]127.193.051.410.6623.487.430.37
SLT* [64]93.463.011.681.172.863.901.85
TDNS* [63]71.723.292.702.1916.985.121.17
Ours56.354.273.222.3064.963.540.55
", + "bbox": [ + 84, + 258, + 465, + 344 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b541412e76b946b207f101f05679b7483e7bedcb34e0a0ff227c946756e488e.jpg", + "table_caption": [ + "Table 2. Quantitative results on the COUCH dataset. SLT* and TDNS* mean we modify and implement them." + ], + "table_footnote": [], + "table_body": "
MethodFD↓User study↑APDM↑APDP↑APDT↑PE↓RE↓Penetration↓Sliding↓
NSM [56]90.393.950001.720.406.430.69
SAMP [19]68.863.771.210.1120.014.774.847.831.38
Ours57.024.042.620.9362.231.010.284.850.80
", + "bbox": [ + 84, + 402, + 465, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e24db9e6a07fcb9fcc57645ee0972f302e45bbe19bc8d4092e2bb9c0aaed7126.jpg", + "image_caption": [ + "Figure 5. Results in a cluttered scene. Our method can generate motions that avoid obstacles in a cluttered scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 492, + 465, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "higher than the approaches [19, 64] that employ deterministic $\\mathrm{A}^*$ [18], but much lower than our method. Although COUCH [76] exhibits lower foot sliding than our method, it may sometimes get stuck, resulting in lower foot sliding since the character does not move.", + "bbox": [ + 75, + 709, + 468, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on the NSM dataset. On the NSM dataset [56], we compare our approach with SAMP [19] and NSM [56]. Tab. 3 shows that our method outperforms baselines. Compared with the deterministic method NSM, our approach could generate stochastic motion and diverse trajectories.", + "bbox": [ + 75, + 821, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/60d63e97046e651a1b62d78651063507c2e089df3a931c661c1b0225b03fd2eb.jpg", + "table_caption": [ + "Table 3. Quantitative results on the NSM dataset." + ], + "table_footnote": [], + "table_body": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
w/o GP31.073.212.4685.214.100.49
w/o MT27.873.623.01127.945.300.72
w/o MP25.144.013.7885.694.580.43
w/o TC36.772.762.4283.264.340.63
Ours22.344.064.5291.384.000.50
", + "bbox": [ + 509, + 260, + 888, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study of the impact of sub-modules. Although the variant without MT generates more diverse trajectories as shown in $\\mathrm{APD}_T$ , the motion quality is much worse as indicated by the higher FD, penetration, and sliding. w/o: without. GP: goal pose generation. MT: milestone point generation. MP: milestone pose generation. TC: trajectory completion.", + "bbox": [ + 498, + 361, + 890, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results in a cluttered scene. We show our generated results in a cluttered scene in Fig. 5. The percentage of frames with penetration is $3.8\\%$ for our method and $4.9\\%$ for SAMP. More details are in the supplementary material.", + "bbox": [ + 498, + 470, + 890, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. As demonstrated in Fig. 6, our approach achieves better results than baselines [19, 63] on the SAMP dataset. Fig. 7 compares our method with COUCH and TDNS on the COUCH dataset. More qualitative results are in the supplementary material.", + "bbox": [ + 498, + 555, + 890, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 642, + 648, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of each sub-module. To show the effectiveness of our hierarchical design, we evaluate our method against four variants where we remove one sub-module for each variant. Tab. 4 indicates that each component improves performance. Generating a whole trajectory leads to more diverse trajectories but the motion quality is worse and has more foot sliding. This validates the necessity of the separate generation of trajectories and motions.", + "bbox": [ + 496, + 666, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Goal pose generation. We evaluate our goal pose generator with several variants, including cVAE, DDPM, and standard VQ-VAE. For this evaluation, we only replace the goal pose generation module and keep the others the same.", + "bbox": [ + 496, + 810, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 shows that part VQ-VAE generates more diverse poses than continuous latent space models. The comparison", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f9493cac2b81052dc5a587d9c057ebf9b446360c9363ca85d6243807b0f7431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 90, + 356, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ba3fbe2fb50f1044eb3563862c079b92157e47faaadf7473cc907da1152334c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 90, + 612, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8dcac349f5f41eaa22b12a170e752aa93073df06a6aa1efc680f2e89c278a8e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 90, + 867, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc15e7ad015fa69354d59bc3d47de8e65f6232ba24aa41155eb93e183479aea4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 194, + 356, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4dba9b99cd50a9a74f708a7de38f81c37f818b8ffc42918eff0e240b0ab814e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 194, + 612, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af57b0727e3512349c25ebf78da51e9323b9aa52d1e47624881441957ecb2bf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 194, + 867, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8212b3241dd2b7457ded4c158602a2a1bc7f719d8393d20d4beb51b9356eefa4.jpg", + "image_caption": [ + "SAMP" + ], + "image_footnote": [], + "bbox": [ + 101, + 294, + 356, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4dd1961b5870def3098fcee476ff9f9757cab68e23d62861a712b6b58eb06080.jpg", + "image_caption": [ + "TDNS" + ], + "image_footnote": [], + "bbox": [ + 356, + 294, + 612, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61a6ba53db2a0cc9c475b2fb2e92c509b0c4911bbc940644f2c5837ebc30a58a.jpg", + "image_caption": [ + "Ours", + "Figure 6. Qualitative results on the SAMP dataset. We compare our method with the baselines SAMP [19] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that SAMP and TDNS tend to walk backward with more foot sliding. The second row shows that the baselines start to sit and lie down before reaching the object while our method synthesizes more natural results. The third row demonstrates that our framework has the capability to generate a long trajectory and walk naturally along the trajectory while SAMP gets stuck near the object and TDNS synthesizes unnatural motions. Lines on the floor indicate trajectories. The human with pink clothes indicate the start position. Darker color denotes later frames in the sequence." + ], + "image_footnote": [], + "bbox": [ + 612, + 294, + 867, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43cc387d22673748ba60534607c8a53f6cbfb000e1123e50884bca7cd93e8312.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 508, + 359, + 603 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d26e69c4f57a92c329d6c93b3d95018e99321d394730fe2296be3935266467a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 508, + 612, + 603 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abb3e4cab11a0b18ed1742fe33e0c14c3a74bc39e44041eb4e40918ab38b1164.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 508, + 864, + 603 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53e0383f74e904058261f0bf153a081b7d8fb539361621e5f3f85265db8004fd.jpg", + "image_caption": [ + "COUCH" + ], + "image_footnote": [], + "bbox": [ + 107, + 603, + 359, + 698 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/289ff845a0eda095049d9aa78e4f87ea0980709f4cc7b77e8443f280218ea88f.jpg", + "image_caption": [ + "TDNS" + ], + "image_footnote": [], + "bbox": [ + 361, + 603, + 612, + 699 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c898e148c0c93190d136944fdcbeb29117abad7195ef1d95300a5c9c5ff234d6.jpg", + "image_caption": [ + "Ours", + "Figure 7. Qualitative results on the COUCH dataset. We compare our method with the baseline COUCH [76] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that COUCH and TDNS tend to sit when the character is far from the object. The second row shows that COUCH may fail to stand up and TDNS may stand up unnaturally. To better visualize the results of TDNS in the second row, we only keep the frames in which the human stands up and goes to the endpoint." + ], + "image_footnote": [], + "bbox": [ + 612, + 603, + 864, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with standard VQ-VAE shows the necessity of our part design. We also try part VQ-VAE for motion infilling, but the results in Tab. 6 show that its performance is worse.", + "bbox": [ + 75, + 797, + 468, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Milestone generation. To further investigate the impact of milestones, we compare our approach with the variants", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that employ path-planning methods and select points along the path as milestones. For this ablation, we implement $\\mathrm{A}^*$ and NM [63] proposed by TDNS. As demonstrated in Tab. 7, the diversity of generated trajectories of $\\mathrm{A}^*$ [18] is much worse, and the motion quality drops significantly, as indicated by the lower $\\mathrm{APD}_T$ and higher FD. The trajectory", + "bbox": [ + 496, + 797, + 890, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f36f81c4f257b1cbd43493ee2237f269afb7ef3b7a7e1f68e77120c21be8c7aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantsFD ↓APDM ↑APDP ↑APDT ↑Penetration↓Sliding↓
cVAE27.063.363.2790.524.360.47
DDPM34.223.753.3784.764.310.49
VQ-VAE24.773.783.8789.194.270.48
Part VQ-VAE22.344.064.5291.384.000.50
", + "bbox": [ + 86, + 89, + 465, + 161 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/eadc86cd35b81fe387d7f936edd5541252c87804459137e5dbd6907594d53b8d.jpg", + "table_caption": [ + "Table 5. Ablation study of goal pose generation. We implement the goal pose module with different architectures." + ], + "table_footnote": [], + "table_body": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
Part VQ-VAE28.713.853.6283.514.320.89
DDPM22.344.064.5291.384.000.50
", + "bbox": [ + 86, + 212, + 465, + 257 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/fb9bea38f1c9c62796ef15c2e5cf816f726148edc6da0ac185c97b6ac3d4adbc.jpg", + "table_caption": [ + "Table 6. Ablation study of part VQ-VAE for motion infilling. We replace the DDPM as part VQ-VAE to predict motions." + ], + "table_footnote": [], + "table_body": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
A*40.594.074.4242.984.240.88
NM [63]40.544.154.3152.744.150.96
MT22.344.064.5291.384.000.50
", + "bbox": [ + 86, + 303, + 465, + 364 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/bc604761b7ea7d51df11e3f3eec982c34b6a6f4c7dc0be14616180f3e43a1fd9.jpg", + "table_caption": [ + "Table 7. Ablation study of milestone generation. We compare our method with the variant based on the path generated by $\\mathbf{A}^*$ path planning [18]. NM: Neural Mapper [63]. MT: milestone point generation." + ], + "table_footnote": [], + "table_body": "
MethodFD↓APDM↑APDP↑Penetration↓Sliding↓
ConvAE [30]26.503.954.334.820.89
SLT [64]27.953.764.054.130.78
Ours22.344.064.524.000.50
", + "bbox": [ + 86, + 444, + 465, + 511 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "diversity of NM [63] is better than $\\mathbf{A}^*$ , but still worse than our method. The reason why these variants perform poorly might be the low diversity of trajectory that affects the distribution of generated motions for calculating FD.", + "bbox": [ + 75, + 569, + 468, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Motion infilling. To validate our motion infilling module, we compare it with ConvAE [30] and SLT [64]. We only replace the motion infilling module and keep the others the same. The comparison of motion quality and diversity is shown in Tab. 8 and our method outperforms ConvAE [30] and SLT [64] with lower FD.", + "bbox": [ + 75, + 651, + 468, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "DDPM vs. VAE. To show the effectiveness of DDPM, we implement a cVAE variant, where we simply replace our transformer DDPM with transformer cVAE [45, 63]. As shown in Tab. 9, although cVAE models could generate more diverse trajectories, their motion quality is far from satisfactory, indicated by the much higher value of FD.", + "bbox": [ + 75, + 763, + 468, + 854 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison with other diffusion models. Our approach stands out from architectures in MDM [59] and FLAME", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/536701ef207f755236fe451ac93d8c548c408b92e5231b5f62dcf25d76a62518.jpg", + "table_caption": [ + "Table 8. Ablation study of motion infilling module. We compare our method with other motion infilling methods." + ], + "table_footnote": [], + "table_body": "
ArchFD↓R-APDM↑R-APDP↑R-APDT↑Penetration↓Sliding↓
cVAE29.323.653.58109.555.061.12
DDPM22.344.064.5291.384.000.50
", + "bbox": [ + 509, + 89, + 888, + 132 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/720b3d0b137f785e8725acca6e50c1a8cb95d421a0ad67ab8d6b4506cece2d3d.jpg", + "table_caption": [ + "Table 9. Evaluation of the architecture for our generation framework. Although cVAE based variant generates more diverse trajectories, the motion quality drops significantly as indicated by the much higher FD. Arch stands for the architecture type." + ], + "table_footnote": [], + "table_body": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
MDM [59]38.694.013.7772.675.190.70
MDM [59] + C23.974.024.1389.804.870.39
FLAME [31]28.934.024.4665.465.780.89
Ours22.344.064.5291.384.000.50
", + "bbox": [ + 509, + 212, + 888, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 10. Comparison with MDM and FLAME. C denotes the frame-wise conditions.", + "bbox": [ + 500, + 287, + 890, + 314 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[31] by incorporating frame-wise conditions. Tab. 10 demonstrates the significance of the frame-wise conditions.", + "bbox": [ + 500, + 340, + 890, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "More analyses and ablation studies. More detailed analyses and ablation studies of our design choices are provided in the supplementary material.", + "bbox": [ + 498, + 396, + 890, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 455, + 625, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although our method can generate diverse and natural motions, there are still some limitations. Our method is offline and cannot be applied to interactive scenarios. We assume that the objects are static and cannot handle moving objects. The diffusion models require a long inference time. It takes 7.13 seconds on average for a 720-frame sequence on a Titan Xp GPU. The slow speed might be solved by methods that could accelerate diffusion models [43, 2].", + "bbox": [ + 496, + 479, + 890, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 617, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose a novel hierarchical pipeline for motion synthesis of human-object interactions. Our approach first generates the goal pose and then predicts a set of milestones. Next, we synthesize motions with the guide of milestones. Furthermore, we apply DDPM in our hierarchical pipeline. We also show that our framework could generate more diverse and natural human-object interaction motions than other methods.", + "bbox": [ + 496, + 645, + 890, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 782, + 666, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Jintao Lu, Zhi Cen, Zizhang Li, and Kechun Xu for the valuable discussions. This work was partially supported by the Key Research Project of Zhejiang Lab (No. K2022PG1BB01), NSFC (No. 62172364), and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2", + "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 9", + "[3] German Barquero, Sergio Escalera, and Cristina Palmero. BeLFusion: Latent Diffusion for Behavior-Driven Human Motion Prediction. arXiv e-prints, page arXiv:2211.14304, Nov. 2022. 3", + "[4] Emad Barsoum, John R. Kender, and Zicheng Liu. Hpgan: Probabilistic 3d human motion prediction via gan. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1499-149909, 2018. 2", + "[5] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, Ding Liu, Jing Liu, and Nadia Magnenat Thalmann. Learning progressive joint propagation for human motion prediction. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, page 226-242, Berlin, Heidelberg, 2020. Springer-Verlag. 2", + "[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, page 387-404, Berlin, Heidelberg, 2020. Springer-Verlag. 2, 5", + "[7] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 5887-5895, 2021. 2", + "[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your Commands via Motion Diffusion in Latent Space. arXiv e-prints, page arXiv:2212.04048, Dec. 2022. 3", + "[9] Simon Clavet. Motion matching and the road to next-gen animation. In GDC, 2016. 2", + "[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6518-6526, 2020. 2", + "[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A Framework for Denoising-Diffusion-based Motion Synthesis. arXiv eprints, page arXiv:2212.04495, Dec. 2022. 3", + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "volume 34, pages 8780-8794. Curran Associates, Inc., 2021. 2", + "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, June 2021. 4", + "[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 2", + "[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396-1406, 2021. 3", + "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, oct 2020. 2", + "[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, June 2022. 4", + "[18] Peter E. Hart, Nils J. Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. 2, 6, 7, 8, 9", + "[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11374-11384, October 2021. 1, 2, 3, 5, 6, 7, 8", + "[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, June 2021. 3", + "[21] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH '23, New York, NY, USA, 2023. Association for Computing Machinery. 2", + "[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 1", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3, 4", + "[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. arXiv e-prints, page arXiv:2207.12598, July 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. arXiv e-prints, page arXiv:2204.03458, Apr. 2022. 3", + "[26] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4), jul 2017. 1, 2", + "[27] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based Generation, Optimization, and Planning in 3D Scenes. arXiv e-prints, page arXiv:2301.06015, Jan. 2023. 3", + "[28] Ashesh Jain, Amir R. Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5308-5317, 2016. 2", + "[29] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Trans. Graph., 41(3), jun 2022. 3", + "[30] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927, 2020. 9", + "[31] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. FLAME: Freeform Language-based Motion Synthesis & Editing. arXiv e-prints, page arXiv:2209.00349, Sept. 2022. 2, 3, 4, 9", + "[32] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, page arXiv:1412.6980, Dec. 2014. 6", + "[33] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, page arXiv:1312.6114, Dec. 2013. 2, 3", + "[34] Jogendra Nath Kundu, Maharshi Gor, and R. Venkatesh Babu. Bihmp-gan: Bidirectional 3d human motion prediction gan. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'19/IAAI'19/EAAI'19. AAAI Press, 2019. 2", + "[35] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2", + "[36] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), aug 2020. 6", + "[37] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3", + "[38] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained Diffusion Models for Unified Human Motion Synthesis. arXiv e-prints, page arXiv:2212.02837, Dec. 2022. 3", + "[39] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Multi-level motion attention for human motion prediction. International Journal of Computer Vision, 129(9):2513-2535, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2", + "[41] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, June 2022. 3", + "[42] Tomohiko Mukai and Shigeru Kuriyama. Geostatistical motion interpolation. ACM Trans. Graph., 24(3):1062-1070, jul 2005. 2", + "[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8162-8171. PMLR, 18-24 Jul 2021. 9", + "[44] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference (BMVC)*, 2018. 2", + "[45] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 9", + "[46] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1", + "[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. arXiv e-prints, page arXiv:2209.14988, Sept. 2022. 3", + "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, page arXiv:2204.06125, Apr. 2022. 2", + "[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3, 4", + "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 2, 3", + "[51] C. Rose, M.F. Cohen, and B. Bodenheimer. Verbs and adverbs: multidimensional motion interpolation. IEEE Computer Graphics and Applications, 18(5):32-40, 1998. 2", + "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv e-prints, page arXiv:2209.14792, Sept. 2022. 3", + "[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 4", + "[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc. 2", + "[55] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 41(4), jul 2022. 2", + "[56] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6), nov 2019. 1, 2, 4, 5, 6, 7", + "[57] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), aug 2020. 2", + "[58] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 2, 3", + "[59] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 3, 4, 9", + "[60] Jonathan Tseng, Rodrigo Castellon, and C. Karen Liu. EDGE: Editable Dance Generation From Music. arXiv eprints, page arXiv:2211.10658, Nov. 2022. 3", + "[61] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2, 4", + "[63] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20460-20469, June 2022. 1, 2, 3, 5, 6, 7, 8, 9", + "[64] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9401-9411, June 2021. 2, 5, 6, 7, 9", + "[65] J. Wang, S. Yan, B. Dai, and D. Lin. Scene-aware generative network for human motion synthesis. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "(CVPR), pages 12201-12210, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 2", + "[66] Mao Wei, Liu Miaomiao, and Salzemann Mathieu. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2", + "[67] Mao Wei, Liu Miaomiao, Salzemann Mathieu, and Li Hongdong. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2", + "[68] D.J. Wiley and J.K. Hahn. Interpolation synthesis of articulated figure motion. IEEE Computer Graphics and Applications, 17(6):39-45, 1997. 2", + "[69] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3", + "[70] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to generate multimodal human dynamics. In European Conference on Computer Vision, pages 276–293. Springer, 2018. 2", + "[71] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6", + "[72] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 37(4), jul 2018. 2, 6, 7", + "[73] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 40(4), jul 2021. 2", + "[74] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 4", + "[75] S. Zhang, Y. Zhang, Q. Ma, M. J. Black, and S. Tang. Place: Proximity learning of articulation and contact in 3d environments. In 2020 International Conference on 3D Vision (3DV), pages 642-651, Los Alamitos, CA, USA, nov 2020. IEEE Computer Society. 3, 6", + "[76] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 1, 2, 4, 5, 6, 7, 8", + "[77] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 6", + "[78] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3d people in scenes without people. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 6", + "[79] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15072", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 2", + "[81] Yi Zhou, Zimo Li, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent networks for extended complex human motion synthesis. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 2", + "[82] Zixiang Zhou and Baoyuan Wang. UDE: A Unified Driving Engine for Human Motion Generation. arXiv e-prints, page arXiv:2211.16016, Nov. 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 273 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "15073", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_model.json b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1fcdaaf154697012ca6b951fbb2a0f2ee1d1f4dd --- /dev/null +++ b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_model.json @@ -0,0 +1,3053 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.768, + 0.175 + ], + "angle": 0, + "content": "Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.203, + 0.777, + 0.223 + ], + "angle": 0, + "content": "Huaijin Pi\\(^{1}\\), Sida Peng\\(^{1}\\), Minghui Yang\\(^{2}\\), Xiaowei Zhou\\(^{1}\\), Hujun Bao\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.226, + 0.748, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Lab of CAD&CG, Zhejiang University \\(^{2}\\) Ant Group" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.279, + 0.314, + 0.296 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.474, + 0.542 + ], + "angle": 0, + "content": "This paper presents a novel approach to generating the 3D motion of a human interacting with a target object, with a focus on solving the challenge of synthesizing long-range and diverse motions, which could not be fulfilled by existing auto-regressive models or path planning-based methods. We propose a hierarchical generation framework to solve this challenge. Specifically, our framework first generates a set of milestones and then synthesizes the motion along them. Therefore, the long-range motion generation could be reduced to synthesizing several short motion sequences guided by milestones. The experiments on the NSM, COUCH, and SAMP datasets show that our approach outperforms previous methods by a large margin in both quality and diversity. The source code is available on our project page https://zju3dv.github.io/hghoi." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.572, + 0.21, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.675 + ], + "angle": 0, + "content": "Scene-aware motion generation [19] aims to synthesize 3D human motion given a 3D scene model to enable virtual humans to naturally wander around scenes and interact with objects, which has a variety of applications in AR/VR, filmmaking, and video games." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.676, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Unlike traditional motion generation methods for character control which aim to generate short or repeated motion on the fly guided by a user's control signals [56], we focus on the setting of generating long-term human-object interactions [56, 19, 76] given a starting position of the human and a target object model. This setting brings in new challenges. First, the entire approaching process and the human-object interaction should be coherent, which requires the capability of modeling long-range interaction between the human and the object. Second, in the context of content generation, the generative model should be able to synthesize diverse motions as there are many plausible ways for a real human to approach and interact with the target object." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.281, + 0.887, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.434, + 0.887, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.59, + 0.894, + 0.687 + ], + "angle": 0, + "content": "Figure 1. Genration of human-object interactions. Given an object, our method first predicts a set of milestones, where the rings indicate the positions and the humans with pink clothes represent the local poses. Then the motions are inflated between milestones. This figure shows that our method can generate diverse milestones and motions with the same object. The flow of time is shown with a color code where darker blue denotes the later frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Existing methods for motion synthesis can be roughly characterized into online generation and offline generation. Most online methods [56, 19, 76] focus on real-time control of characters. Given a target object, they generally use auto-regressive models to recurrently generate future motions by feeding their predictions. Although they have been widely used for interactive scenarios like video games, their motion quality is not satisfactory enough for long-term generation [63]. A plausible cause is the error accumulation in the auto-regressive process, where the errors in previous predictions are fed back as the model input, as discussed in [40, 26, 45, 46, 22]. To improve the motion quality," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15061" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "some recent offline methods [6, 65, 64, 63] employ a multistage framework, which first generates the trajectory and then synthesizes motions. TDNS [63] generates paths by combining the cVAE model [33] and deterministic planning methods like A* [18]. Although this strategy can produce reasonable paths, the path diversity is limited, as demonstrated by our experimental results in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.455 + ], + "angle": 0, + "content": "In this paper, we propose a novel offline approach for synthesizing long-term and diverse human-object interactions. Our innovation lies in a hierarchical generation strategy, which first predicts a set of milestones and then generates human motions between milestones. Fig. 1 illustrates the basic idea. Specifically, given the starting position and the target object, we design a milestone generation module to synthesize a set of milestones along the motion trajectory, each of which encodes the local pose and indicates the transition point during the human movement. Based on these milestones, we employ a motion generation module to produce the full motion sequence. Thanks to the milestones, we simplify the long-sequence generation into synthesizing several short motion sequences. Furthermore, the local pose at each milestone is generated by a transformer that considers the global dependency [62], leading to temporally consistent results, which further contribute to coherent motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.668 + ], + "angle": 0, + "content": "In addition to our hierarchical generation framework, we further exploit diffusion models [53, 23, 54] to synthesize human-object interactions. Previous diffusion models for motion synthesis [31, 59] combine transformer [62] and Denoising Diffusion Probabilistic Model (DDPM) [23]. Directly applying them to our setting is prohibitively comput-intensive due to the long motion sequences and may lead to the GPU memory explosion [50]. Because our hierarchical generation framework converts the long-term generation to the synthesis of several short sequences, the required GPU memory is reduced to the same level of short-term motion generation. Therefore, we can efficiently leverage the transformer DDPM to synthesize long-term motion sequences, which improves the generation quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.745 + ], + "angle": 0, + "content": "We validate our design choices on the NSM [56], COUCH [76], and SAMP [19] datasets with extensive experiments. On these datasets, our hierarchical framework outperforms previous methods significantly in both motion quality and diversity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.761, + 0.22, + 0.777 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.244, + 0.803 + ], + "angle": 0, + "content": "2.1. Motion synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Motion synthesis is a long-standing problem in computer graphics and vision [68, 51, 42, 9]. With the rapid development of deep learning, recent works have applied neural networks to motion synthesis [14, 40, 26, 72]. Some methods are deterministic [28, 81, 44, 67] while others try to predict stochastic motions by VAE [70, 1, 71, 79, 45] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.501 + ], + "angle": 0, + "content": "GAN [4, 34]. To further improve the performance, some works apply GCN [67, 10, 66, 35, 39] or transformers [5] to extract the features from the human skeleton. To handle the ambiguity of human motion, some works [26, 55] propose to employ phase signals to guide the motions. Recent works start to consider the scene context [56, 6]. NSM [56] is the first work that aims at synthesizing human motions with object-level interactions with specific action labels. Based on NSM, SAMP [19] applies cVAE to predict diverse motions. These works [56, 19, 76] mainly focus on the interaction with one or two objects while others [6, 65] generate motions with a full scene (including the information about floors and walls) as the input. [6] predicts future motions of the given motion sequence by first predicting the trajectory and then generating motions based on a 2D image of the scene. Similar to this pipeline, [65] applies GAN [16]. Furthermore, [64] proposes a framework that first places pose on a human-provided path and then synthesizes motions in a 3D scene with a full-scene scan. [63] combines \\(\\mathbf{A}^*\\) and cVAE to generate diverse human motions. To further control human motion, [80] employs the gaze, and COUCH [76] explicitly models the hand contacts to guide the prediction. Some methods [7, 21] also enable physically simulated characters to perform scene interaction tasks including sitting [7, 21] and carrying boxes [21]. Some works also focus on grasp [58, 69], manipulation [73] and the interaction with dynamic objects [57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Ours vs. others. This work follows the setting of [56, 19] and focuses on object-level interaction. In contrast to [56, 19] which generate motions mainly based on auto-regressive models, we design a hierarchical framework to synthesize motions. Different from [6, 64], our work focuses on much more long-term generation (longer than 10 seconds) while [6] is 2 seconds and [64] is 6 seconds. Instead of planning the path auto-regressively with an extra network to generate diverse trajectories like [63], our method directly predicts a set of milestones to describe the approaching process which is inherently diverse. In addition, most methods [19, 76, 64, 63] rely on cVAE to generate stochastic motions while we exploit DDPM [23] to synthesize trajectories and motions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.667, + 0.756 + ], + "angle": 0, + "content": "2.2. Diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Diffusion models [53] are a class of likelihood-based methods that generate samples by gradually removing noise from a signal. Then, [23, 54] develop the diffusion models for high-quality image generation. To control the generated results, [12] proposes classifier guidance for trading off diversity for fidelity. Later, the classifier-free model [24] achieves better results [48] in text-conditioned image generation. In addition, diffusion models have been successfully applied to other domains like the generation of videos" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15062" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.208, + 0.895, + 0.25 + ], + "angle": 0, + "content": "Figure 2. Overview of our pipeline. Our pipeline consists of three components: First, the goal pose is synthesized given the object. Then, a number of milestones with local poses are predicted based on the goal pose. Finally, the trajectory and the full motion sequences are inflated between the milestones." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.279, + 0.287 + ], + "angle": 0, + "content": "[52, 25] and 3D contents [47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.471, + 0.408 + ], + "angle": 0, + "content": "There are some works [74, 59, 31, 38] that apply DDPM to synthesize motions. [3, 8] also explore latent diffusion models [50] for motion generation. However, most works focus on text-conditioned [74, 11], audio-driven [82] and music-driven [60, 82] motion generation while we target human-object interaction. In this work, we apply transformer DDPM [23] to a multi-stage framework, which separately generates trajectories and synthesizes motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.47, + 0.515 + ], + "angle": 0, + "content": "There are some concurrent works [27] that apply DDPM to synthesize motions in a scene. However, [27] is designed for short-term motion generation (around 2 seconds) while we target long-term human motion generation (longer than 10 seconds). Different from existing work [59, 27], we employ DDPM in a multi-stage framework, where trajectories and motions are separately predicted." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.527, + 0.176, + 0.543 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Given the object \\(\\mathbf{I}\\) and the starting point \\(\\mathbf{s}\\), our goal is to synthesize 3D human motions \\(\\{(\\mathbf{r}_i, \\pmb{\\theta}_i)\\}_{i=1}^N\\) with human-object interactions. \\(\\mathbf{r}_i\\) is the root trajectory, and the \\(\\pmb{\\theta}_i\\) indicates local pose at \\(i\\)-th frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.793 + ], + "angle": 0, + "content": "We design a hierarchical motion generation framework as shown in Fig. 2. First, we employ GoalNet[19] to predict an interaction goal on the object. Then, we generate the goal pose (Sec. 3.1) to explicitly model the human-object interaction. Next, our milestone generation module (Sec. 3.2) estimates the length of the milestones, produces the trajectory of the milestones from the starting point to the goal, and places milestone poses. Therefore, the long-range motion generation is decomposed into combinations of short-range motion synthesis. Finally, we design a motion generation module (Sec. 3.3) to synthesize the trajectory between the milestones and infill the motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.276, + 0.818 + ], + "angle": 0, + "content": "3.1. Goal pose generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We call the pose in which a person interacts with an object and remains stationary a goal pose. To synthesize diverse human motion, we first generate a goal pose interacting with the object following [58, 69, 63]. Most methods [78, 75, 20] generate human poses using the cVAE model." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.275, + 0.887, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.467, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Figure 3. Part VQ-VAE. Part VQ-VAE first splits the skeleton into multiple parts and learns the codebooks separately. The composition of different parts is subsequently modeled with the autoregressive prediction model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.895, + 0.769 + ], + "angle": 0, + "content": "They project the poses to the standard normal distribution in the continuous space [33]. Empirically, we find that cVAE models do not perform well in our setting. To overcome this challenge, we introduce VQ-VAE [61, 49] to model the data distribution which exploits a discrete representation to cluster the data in a finite set of points [37]. We hypothesize that limited data of goal pose from the SAMP dataset [19] can always be clustered by VQ-VAE but may not be enough for learning a continuous latent space for VAE [41]. In addition, based on the observation that different human poses may share similar properties [15, 29] (e.g., humans may sit with different hand positions but the same leg positions), we split joints into \\( L \\) (\\( L = 5 \\)) different non-overlapping groups like MotionPuzzle [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Quantization. As shown in Fig. 3, the goal pose \\(\\theta_{g}\\) is split into separate joint groups as \\(\\theta_{g} = \\{\\theta_{gi}\\}_{i=1}^{L}\\). Then a discrete codebook \\(\\mathbf{Z}_i\\) with a list of vectors compares the output \\(\\hat{\\mathbf{z}}_i\\) from the encoder \\(E_i\\) to find the closest vector in Euclidean distance. The \\(L\\) vectors with minimal distances will be concatenated and fed into a shared decoder \\(D\\) to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15063" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.388, + 0.107 + ], + "angle": 0, + "content": "reconstruct \\(\\theta_{g}\\). The loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.119, + 0.468, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\boldsymbol {\\theta} _ {g}, D (\\mathbf {z})\\right) = \\| \\boldsymbol {\\theta} _ {g} - D (\\mathbf {z}) \\| _ {2} ^ {2} + \\sum_ {i = 1} ^ {L} \\| s g \\left[ E _ {i} \\left(\\boldsymbol {\\theta} _ {g i}\\right) \\right] - \\mathbf {z} _ {i} \\| _ {2} ^ {2} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.162, + 0.435, + 0.193 + ], + "angle": 0, + "content": "\\[\n+ \\beta \\sum_ {i = 1} ^ {L} \\| s g [ \\mathbf {z} _ {i} ] - E _ {i} (\\boldsymbol {\\theta} _ {g i}) \\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.203, + 0.122, + 0.216 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.23, + 0.468, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\left[ \\mathbf {z} _ {1}, \\dots , \\mathbf {z} _ {L} \\right], \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.248, + 0.468, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} = \\underset {\\mathbf {z} _ {i} ^ {\\prime} \\in Z _ {i}} {\\arg \\min } \\| \\mathbf {z} _ {i} ^ {\\prime} - \\hat {\\mathbf {z}} _ {i} \\|, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.28, + 0.468, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {i} = E \\left(\\boldsymbol {\\theta} _ {g i}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.305, + 0.468, + 0.351 + ], + "angle": 0, + "content": "Here, the term \\( sg[\\cdot] \\) denotes the stop-gradient operator, and \\( \\| sg[\\mathbf{z}_i] - E_i(\\pmb{\\theta}_{gi})\\|_2^2 \\) is the commitment loss [49] controlled by a weighting factor \\( \\beta \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.468, + 0.459 + ], + "angle": 0, + "content": "Generation. With \\(E_{i}\\) and \\(D\\) available, we can represent \\(\\pmb{\\theta}_{g} = \\{\\pmb{\\theta}_{g1},\\dots,\\pmb{\\theta}_{gL}\\}\\) by a sequence of the part-based codebook indices. To be more specific, we use \\(E_{i}\\) to extract the feature from \\(\\pmb{\\theta}_{gi}\\) and find the closest vector \\(\\mathbf{z}_i\\in \\mathbf{Z}_i\\). Then we use \\(s_i\\in \\{0,\\ldots ,|\\mathbf{Z}_i| - 1\\}\\) to indicate the index of \\(\\mathbf{z}_i\\) in \\(\\mathbf{Z}_i\\). Therefore, \\(\\pmb{\\theta}_{g}\\) can be represented by \\(\\mathbf{s} = \\{s_1,\\dots,s_L\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.468, + 0.67 + ], + "angle": 0, + "content": "To generate a natural goal pose, we convert the problem to predict a sequence of indices that can represent \\(\\theta_{g}\\). We formulate the inference as a conditional auto-regressive process and employ a transformer [62] to learn to predict the distribution of possible indices [13]. The condition contains the environment around the goal \\(\\mathbf{O}_g\\), and the action \\(\\mathbf{a}_g\\). Following NSM [56] and COUCH [76], a cylindrical volume of a pre-defined radius and height is created around the goal. Within this volume, spheres are uniformly sampled and the occupancies corresponding to the object of these spheres are calculated. Then, these occupancies are flattened to form a feature vector denoted as \\(\\mathbf{O}_g\\). \\(\\mathbf{a}_g\\) is a vector indicating the action type. These variables are fed into the transformer as tokens. Our target is to learn the likelihood of the sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.682, + 0.468, + 0.713 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {s} \\mid \\mathbf {O} _ {g}, \\mathbf {a} _ {g}) = \\prod_ {i} p \\left(s _ {i} \\mid s _ {< i}, \\mathbf {O} _ {g}, \\mathbf {a} _ {g}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.718, + 0.468, + 0.78 + ], + "angle": 0, + "content": "After predicting the indices, we map them back to their corresponding codebook entries to get the quantized features \\(\\mathbf{z} = [\\mathbf{z}_1,\\dots ,\\mathbf{z}_L]\\), which are fed into the decoder \\(D\\) to generate the goal pose \\(\\theta_{g}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.274, + 0.803 + ], + "angle": 0, + "content": "3.2. Milestone generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Based on the starting pose and the goal pose, we can generate the milestone trajectory and synthesize the local poses at the milestones. Following [74, 59], we build a transformer DDPM [23] and apply it to generate the milestones for better quality. Because the length of motion data is unknown and can be arbitrary (e.g., the human could quickly" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "walk towards the chair and sit down or sit after walking around the chair slowly), we predict the length of milestones, denoted by \\( N \\). Then we synthesize \\( N \\) milestone points and place the local poses on these points." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.173, + 0.892, + 0.264 + ], + "angle": 0, + "content": "Transformer DDPM. Here we first briefly introduce DDPM [23], which is learned to reverse a diffusion process. Formally, the diffusion model [53] is defined as a latent variable model of the form \\( p_{\\theta}(\\mathbf{x}_0) \\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T} \\), where \\( \\mathbf{x}_0 \\sim q(\\mathbf{x}_0) \\) is the data and \\( \\mathbf{x}_1, \\ldots, \\mathbf{x}_T \\) are the latents. \\( p_{\\theta}(\\mathbf{x}_{0:T}) \\) is formulated as a Markov chain as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.278, + 0.891, + 0.317 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.35, + 0.891, + 0.365 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.378, + 0.891, + 0.424 + ], + "angle": 0, + "content": "Diffusion models approximate posterior \\( q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) \\) as a Markov chain that gradually adds Gaussian noise to the data with variance schedules given by \\( \\beta_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.437, + 0.89, + 0.476 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.493, + 0.89, + 0.517 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.524, + 0.89, + 0.554 + ], + "angle": 0, + "content": "In contrast to adding noises on \\(\\mathbf{x}_0\\) sequentially, DDPM formulates the diffusion process as:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.568, + 0.89, + 0.583 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.598, + 0.89, + 0.674 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} = 1 - \\beta_{t}\\) and \\(\\bar{\\alpha}_{t} = \\prod_{s = 1}^{t}\\alpha_{s}\\). Hence, we can generate \\(\\mathbf{x}_t\\) by sampling a noise \\(\\epsilon\\) as the training data. DDPM employs a neural network to model \\(p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)\\) and the inference is to gradually denoise \\(\\mathbf{x}_t\\) from \\(t = T\\) to \\(t = 1\\) where \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Like existing works [74, 31] that apply DDPM in the motion domain, we employ a transformer decoder [62] as our architecture of DDPM. The transformer takes the noise \\(\\mathbf{x}_t\\) and the condition \\(\\mathbf{C}\\) as input. The condition \\(\\mathbf{C}\\) means variables related to the generation, which will be described in detail in each subsection. The diffusion time-step \\(t\\) is in the sinusoidal position embeddings form [62] and is injected into each block in the transformer. Different from existing works [59, 74] that assume the length of the motion sequences is already given, we insert a parallel branch to estimate the length of milestones by taking the length token \\(\\mathbf{H}^{tok}\\) and the condition \\(\\mathbf{C}\\) as the input, as shown in Fig. 4. The length prediction head is an MLP and predicts a multinomial distribution over discrete length indices \\(\\{1, 2, \\dots, N_{max}\\}\\) like [17], where \\(N_{max}\\) represents the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15064" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Overview of transformer DDPM for milestone generation. The model first takes the length token \\(\\mathbf{H}^{tok}\\) and the condition \\(\\mathbf{C}\\) as the input to predict the data length. Then it constructs the noise sequence \\(\\mathbf{x}_T^{1:N}\\) with length \\(N\\). In the diffusion process, it is fed with \\(\\mathbf{C}\\) and the sequence \\(\\mathbf{x}_t^{1:N}\\) at time-step \\(t\\) to predict the target \\(\\hat{\\mathbf{x}}_0^{1:N}\\). For other sub-modules, we remove the length prediction head." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.473, + 0.581 + ], + "angle": 0, + "content": "length of the longest sequences for training. We use cross-entropy loss as the loss function. At inference time, we sample the length \\(N\\) from the estimated distribution. Next, we construct \\(N\\) milestones as the input to the transformer. We predict the denoised data \\(\\hat{\\mathbf{x}}_0\\) based on the condition \\(\\mathbf{C}\\) by the DDPM \\(f\\). This is formulated as \\(\\hat{\\mathbf{x}}_0 = f(\\mathbf{x}_t,t,\\mathbf{C})\\), and the training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.589, + 0.47, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\operatorname {E} _ {t \\in [ 1, T ], \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\mathbf {x} _ {0} - \\hat {\\mathbf {x}} _ {0} \\| \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.695 + ], + "angle": 0, + "content": "In our setting, we use a two-step strategy like [64, 63], where two transformer DDPMs are applied to first generate the milestone points and then synthesize the local pose at each milestone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.47, + 0.758 + ], + "angle": 0, + "content": "Generation of milestone points. The milestone points are conditioned on the object, the information of the starting point, and the goal. We then define the condition as:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.768, + 0.47, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {m}} = \\left\\{\\mathbf {I} _ {s}, \\mathbf {I} _ {g}, \\mathbf {O} _ {s}, \\mathbf {O} _ {g}, \\mathbf {g}, \\mathbf {s} \\right\\}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_s\\) and \\(\\mathbf{I}_g\\) are the object representation relative to the starting point and the goal. Similar to other methods [56, 19, 76], the object representation is modeled by an \\(8\\times 8\\times 8\\) grid with their positions and occupancies. Following NSM [56] and COUCH [76], we also explicitly model the occupancies around the starting point and the goal with \\(\\mathbf{O}_s\\) and \\(\\mathbf{O}_g\\) (the same form as the occupancy feature described in Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "The information of goal \\(\\mathbf{g}\\) is defined as \\(\\{\\mathbf{r}_g, \\mathbf{a}_g, \\boldsymbol{\\theta}_g\\}\\), where \\(\\mathbf{r}_g\\) means the goal position and orientation in the starting point coordinate system, \\(\\mathbf{a}_g\\) is the target action label at the goal, and \\(\\boldsymbol{\\theta}_g\\) is the goal pose. \\(\\mathbf{s} = \\{\\mathbf{a}_s, \\boldsymbol{\\theta}_s\\}\\) represents the action labels \\(\\mathbf{a}_s\\) and the pose \\(\\boldsymbol{\\theta}_s\\) at the starting point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.273 + ], + "angle": 0, + "content": "The target is to predict the milestone \\(\\{\\mathbf{m}_1, \\dots, \\mathbf{m}_N\\}\\) with length \\(N\\). Following NSM [56], a bi-directional scheme is employed for the milestone points generation where we predict the roots of milestones in both the starting point coordinate system and the goal coordinate system. The final predicted roots are blended from these two kinds of outputs. The milestone point \\(\\mathbf{m}_i\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.283, + 0.892, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i} = \\left\\{\\mathbf {r} _ {i} ^ {b}, \\mathbf {c} _ {i}, \\mathbf {w} _ {i} \\right\\}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.893, + 0.448 + ], + "angle": 0, + "content": "The representation is similar to previous work [19, 56, 76], where \\(\\mathbf{r}_i^b\\) indicates the root position and forward directions relative to the starting point and the goal, \\(\\mathbf{c}_i\\) is a label vector indicating the contact between the environment and the body, and we use a high-dimensional feature vector \\(\\mathbf{w}_i\\) to encode the character state at the milestone following [56, 19, 76]. The details of these variables are provided in the supplementary material. We use transformer DDPM \\(f_{\\mathbf{m}}\\) to predict the length and synthesize milestone points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.893, + 0.678 + ], + "angle": 0, + "content": "Generation of milestone poses. Previous works [64, 63] separately place the poses on the sparse points along the path and infill the motions between, which may lead to unnatural transitions between these poses. On the contrary, we generate the local pose at each milestone with transformer DDPM to build the temporal dependency. With the help of the generated milestone points, we can access the accurate spatial relationship \\(\\mathbf{I}_i\\) between the object and the character and the ego-centric environment occupancies \\(\\mathbf{O}_i\\) [76] at the \\(i\\)-th milestone. The milestone poses also depend on the milestone state including \\(\\{\\mathbf{c}_i, \\mathbf{w}_i\\}\\) at the \\(i\\)-th milestone. Consequently, we define the condition \\(\\mathbf{C_k}\\) as a combination with starting pose \\(\\theta_s\\), goal pose \\(\\theta_g\\), and frame-wise condition \\(\\gamma_i\\) at the \\(i\\)-th milestone as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.687, + 0.891, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {k}} = \\left\\{\\boldsymbol {\\theta} _ {s}, \\boldsymbol {\\theta} _ {g}, \\gamma_ {1}, \\dots , \\gamma_ {N} \\right\\}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.706, + 0.892, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad \\gamma_ {i} = \\left\\{\\mathbf {I} _ {i}, \\mathbf {O} _ {i}, \\mathbf {c} _ {i}, \\mathbf {w} _ {i} \\right\\}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.765 + ], + "angle": 0, + "content": "The local poses at the milestones are predicted by the transformer DDPM \\( f_{\\mathbf{k}} \\) without the length prediction head." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.681, + 0.789 + ], + "angle": 0, + "content": "3.3. Motion generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Instead of predicting motions frame-by-frame [56, 19, 76], our approach hierarchically synthesizes the full sequence based on the generated milestones. We follow [6, 64, 63] to first generate the trajectory and then synthesize the motions. Specifically, within two consecutive milestones, we first complete the trajectory. Then, the motions are inflated with the guidance of successive milestone" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "poses. The two steps are accomplished using two transformer DDPMs (described in Sec. 3.2), respectively. For each step, we carefully design the condition of DDPM to generate the target output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.47, + 0.23 + ], + "angle": 0, + "content": "Trajectory completion. For the trajectory completion between the milestones \\(\\mathbf{m}_i\\) and \\(\\mathbf{m}_{i + 1}\\), we assume that it is only conditioned on the milestones and the object. Thus we define the condition as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.237, + 0.469, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {r}} = \\left\\{\\mathbf {I} _ {i}, \\mathbf {I} _ {i + 1}, \\mathbf {O} _ {i}, \\mathbf {O} _ {i + 1}, \\mathbf {m} _ {i}, \\mathbf {m} _ {i + 1}, \\mathbf {t} _ {i} ^ {i + 1} \\right\\}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.413 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_i\\) indicates the object representation relative to the milestone \\(i\\) and \\(\\mathbf{O}_i\\) denotes the ego-centric occupancies (the same form as the occupancy feature described in Sec. 3.1) around the \\(i\\)-th milestone like NSM [56] and COUCH [76]. \\(\\mathbf{m}_i\\) has been shown in Eq. (13). \\(\\mathbf{t}_i^{i+1}\\) represents the position and orientation of the \\((i+1)\\)-th milestone in the \\(i\\)-th milestone's coordinate system. Between two consecutive milestones, we generate the trajectory with a length of 2 seconds, which is 61 frames. Similar hyperparameters can be found in previous methods [64, 63]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.55 + ], + "angle": 0, + "content": "The target output is the trajectory between two consecutive milestones. Similar to the milestone point, we synthesize the trajectory in the \\(j\\)-th frame with a bi-directional scheme. The trajectory is composed of a set of points that have the same representation as the milestone in Eq. (13). We generate the trajectory with a transformer DDPM \\(f_{\\mathbf{r}}\\) similar to the one in Sec. 3.2. Since we assume the trajectory length between two milestones is fixed, the DDPM \\(f_{\\mathbf{r}}\\) does not have the length prediction head." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Motion infilling. To synthesize the long-range motion, we convert a long sequence into several fixed-length short sequences with the help of milestone points and milestone poses. For a sub-sequence between the consecutive milestone poses, our goal is to generate the missing local poses over the trajectory. The generated motion has to satisfy the trajectory and naturally transits from a milestone to the next milestone. Like milestone pose generation, we use the same representation of frame-wise condition in Eq. (15). The condition is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.724, + 0.469, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {p}} = \\left\\{\\boldsymbol {\\theta} ^ {1}, \\boldsymbol {\\theta} ^ {6 1}, \\gamma^ {1}, \\dots , \\gamma^ {6 1} \\right\\}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.811 + ], + "angle": 0, + "content": "where \\(\\theta^1\\) and \\(\\theta^{61}\\) are the local poses of two consecutive milestones. By taking these inputs, we generate smooth motions using another transformer DDPM \\(f_{\\mathbf{p}}\\) without the length prediction head." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.21, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.291, + 0.864 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We train the part VQ-VAE and transformer DDPM models with the Adam optimizer [32]. All the models are trained" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "with a fixed learning rate of 0.0001 with batch size 256. The remaining details are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.133, + 0.78, + 0.149 + ], + "angle": 0, + "content": "4.2. Datasets and evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.157, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Test setting. Our experiments are conducted on the SAMP [19], COUCH [76], and NSM [56] datasets. Provided with a starting point, a starting pose, an object, and an endpoint, the virtual human is asked to approach the object, interact with it, and leave to reach the endpoint. The ablation studies are conducted on the SAMP dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.27, + 0.892, + 0.512 + ], + "angle": 0, + "content": "Metrics. Following the previous method [19], we calculate the Fréchet distance (FD) between the generated and ground-truth motions to measure the motion quality. We also conduct user studies and each sequence is evaluated by at least 3 users with scores ranging from 1 to 5. In addition, we calculate the penetration ratio [64, 75, 78] and foot sliding [72, 36] to show the physical plausibility between the 3D object and the synthesized motions. We compute the Average Pairwise Distance (APD) [71, 77] to evaluate the diversity. Specifically, we calculate the APD of synthesized motion, the character's pose during object interactions, and trajectories. Following previous work [56, 19], we calculate PE (positional errors) and RE (rotational errors) to indicate the precision of object interactions. For each test object, we generate multiple sequences. More details are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.785, + 0.539 + ], + "angle": 0, + "content": "4.3. Comparison with other methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.547, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Results on the SAMP dataset. On the SAMP dataset [19], we compare our method with online methods SAMP [19] and MoE [72]. As our method is offline, we also implement and modify offline methods SLT [64] and TDNS [63] to our setting. For SLT [64], we employ \\(\\mathrm{A}^*\\) [18] to plan a path and select points along the path as subgoals to form the input for SLT. More details about the implementation of [64, 63] are in the supplementary material. Since MoE often fails to finish the action, we do not calculate the penetration ratio for it. As shown in Tab. 1, our approach outperforms other methods in terms of lower FD, higher user study scores, and higher APD. Furthermore, our method achieves much higher trajectory diversity than SAMP [19]. Although TDNS [63] proposes Neural Mapper (NM) which combines \\(\\mathrm{A}^*\\) [18] and cVAE, the diversity of the generated trajectory is inferior to our method, as indicated by \\(\\mathrm{APD}_T\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results on the COUCH dataset. As our target is to synthesize diverse motions instead of controlling the characters, we only evaluate the motion quality on the COUCH dataset [76]. Tab. 2 shows that our method outperforms all baselines. Our approach achieves much higher \\(\\mathrm{APD}_T\\) than other methods. We observe that \\(\\mathrm{APD}_T\\) of TDNS [63] is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15066" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.21, + 0.089, + 0.768, + 0.184 + ], + "angle": 0, + "content": "
MethodFD ↓User study ↑APDM ↑APDP ↑APDT ↑PE↓RE↓Penetration↓Sliding↓
MoE [72]74.332.763.502.6352.46∞∞-1.68
SAMP [19]57.342.863.633.0563.183.444.126.981.02
SLT* [64]68.832.303.082.6640.131.771.604.281.71
TDNS* [63]46.602.903.683.4066.040.450.395.140.94
Ours22.343.624.064.5291.380.390.324.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.195, + 0.895, + 0.237 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the SAMP dataset. SLT* and TDNS* mean we modify and implement them on the SAMP dataset. The subscript “M”, “P” and “T” stand for “Motion”, “Pose” and “Trajectory”. PE and RE represent the positional error and rotation error. Sliding denotes foot sliding. \\(\\infty\\) means the method failed to reach the goal." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.26, + 0.466, + 0.345 + ], + "angle": 0, + "content": "
MethodFD↓User study ↑APDM↑APDP↑APDT↑Penetration↓Sliding↓
NSM [56]118.982.990008.200.59
SAMP [19]160.121.940.890.184.694.940.72
COUCH [76]127.193.051.410.6623.487.430.37
SLT* [64]93.463.011.681.172.863.901.85
TDNS* [63]71.723.292.702.1916.985.121.17
Ours56.354.273.222.3064.963.540.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.354, + 0.47, + 0.385 + ], + "angle": 0, + "content": "Table 2. Quantitative results on the COUCH dataset. SLT* and TDNS* mean we modify and implement them." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.403, + 0.466, + 0.449 + ], + "angle": 0, + "content": "
MethodFD↓User study↑APDM↑APDP↑APDT↑PE↓RE↓Penetration↓Sliding↓
NSM [56]90.393.950001.720.406.430.69
SAMP [19]68.863.771.210.1120.014.774.847.831.38
Ours57.024.042.620.9362.231.010.284.850.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.459, + 0.428, + 0.473 + ], + "angle": 0, + "content": "Table 3. Quantitative results on the NSM dataset." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.493, + 0.467, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.652, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Figure 5. Results in a cluttered scene. Our method can generate motions that avoid obstacles in a cluttered scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.785 + ], + "angle": 0, + "content": "higher than the approaches [19, 64] that employ deterministic \\(\\mathrm{A}^*\\) [18], but much lower than our method. Although COUCH [76] exhibits lower foot sliding than our method, it may sometimes get stuck, resulting in lower foot sliding since the character does not move." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Results on the NSM dataset. On the NSM dataset [56], we compare our approach with SAMP [19] and NSM [56]. Tab. 3 shows that our method outperforms baselines. Compared with the deterministic method NSM, our approach could generate stochastic motion and diverse trajectories." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.261, + 0.889, + 0.352 + ], + "angle": 0, + "content": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
w/o GP31.073.212.4685.214.100.49
w/o MT27.873.623.01127.945.300.72
w/o MP25.144.013.7885.694.580.43
w/o TC36.772.762.4283.264.340.63
Ours22.344.064.5291.384.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.362, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Table 4. Ablation study of the impact of sub-modules. Although the variant without MT generates more diverse trajectories as shown in \\(\\mathrm{APD}_T\\), the motion quality is much worse as indicated by the higher FD, penetration, and sliding. w/o: without. GP: goal pose generation. MT: milestone point generation. MP: milestone pose generation. TC: trajectory completion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Results in a cluttered scene. We show our generated results in a cluttered scene in Fig. 5. The percentage of frames with penetration is \\(3.8\\%\\) for our method and \\(4.9\\%\\) for SAMP. More details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.556, + 0.892, + 0.631 + ], + "angle": 0, + "content": "Qualitative results. As demonstrated in Fig. 6, our approach achieves better results than baselines [19, 63] on the SAMP dataset. Fig. 7 compares our method with COUCH and TDNS on the COUCH dataset. More qualitative results are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.65, + 0.659 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Impact of each sub-module. To show the effectiveness of our hierarchical design, we evaluate our method against four variants where we remove one sub-module for each variant. Tab. 4 indicates that each component improves performance. Generating a whole trajectory leads to more diverse trajectories but the motion quality is worse and has more foot sliding. This validates the necessity of the separate generation of trajectories and motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Goal pose generation. We evaluate our goal pose generator with several variants, including cVAE, DDPM, and standard VQ-VAE. For this evaluation, we only replace the goal pose generation module and keep the others the same." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Tab. 5 shows that part VQ-VAE generates more diverse poses than continuous latent space models. The comparison" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.091, + 0.357, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.091, + 0.613, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.091, + 0.869, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.195, + 0.357, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.195, + 0.613, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.195, + 0.869, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.295, + 0.357, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.295, + 0.613, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.295, + 0.869, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.396, + 0.239, + 0.405 + ], + "angle": 0, + "content": "SAMP" + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.396, + 0.5, + 0.405 + ], + "angle": 0, + "content": "TDNS" + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.396, + 0.752, + 0.405 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.409, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on the SAMP dataset. We compare our method with the baselines SAMP [19] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that SAMP and TDNS tend to walk backward with more foot sliding. The second row shows that the baselines start to sit and lie down before reaching the object while our method synthesizes more natural results. The third row demonstrates that our framework has the capability to generate a long trajectory and walk naturally along the trajectory while SAMP gets stuck near the object and TDNS synthesizes unnatural motions. Lines on the floor indicate trajectories. The human with pink clothes indicate the start position. Darker color denotes later frames in the sequence." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.509, + 0.361, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.509, + 0.613, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.509, + 0.865, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.604, + 0.361, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.604, + 0.613, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.604, + 0.865, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.702, + 0.252, + 0.711 + ], + "angle": 0, + "content": "COUCH" + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.702, + 0.501, + 0.711 + ], + "angle": 0, + "content": "TDNS" + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.702, + 0.75, + 0.711 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.718, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on the COUCH dataset. We compare our method with the baseline COUCH [76] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that COUCH and TDNS tend to sit when the character is far from the object. The second row shows that COUCH may fail to stand up and TDNS may stand up unnaturally. To better visualize the results of TDNS in the second row, we only keep the frames in which the human stands up and goes to the endpoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.799, + 0.469, + 0.845 + ], + "angle": 0, + "content": "with standard VQ-VAE shows the necessity of our part design. We also try part VQ-VAE for motion infilling, but the results in Tab. 6 show that its performance is worse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Milestone generation. To further investigate the impact of milestones, we compare our approach with the variants" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.799, + 0.892, + 0.892 + ], + "angle": 0, + "content": "that employ path-planning methods and select points along the path as milestones. For this ablation, we implement \\(\\mathrm{A}^*\\) and NM [63] proposed by TDNS. As demonstrated in Tab. 7, the diversity of generated trajectories of \\(\\mathrm{A}^*\\) [18] is much worse, and the motion quality drops significantly, as indicated by the lower \\(\\mathrm{APD}_T\\) and higher FD. The trajectory" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15068" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.09, + 0.466, + 0.162 + ], + "angle": 0, + "content": "
VariantsFD ↓APDM ↑APDP ↑APDT ↑Penetration↓Sliding↓
cVAE27.063.363.2790.524.360.47
DDPM34.223.753.3784.764.310.49
VQ-VAE24.773.783.8789.194.270.48
Part VQ-VAE22.344.064.5291.384.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.172, + 0.47, + 0.2 + ], + "angle": 0, + "content": "Table 5. Ablation study of goal pose generation. We implement the goal pose module with different architectures." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.213, + 0.466, + 0.258 + ], + "angle": 0, + "content": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
Part VQ-VAE28.713.853.6283.514.320.89
DDPM22.344.064.5291.384.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.262, + 0.468, + 0.29 + ], + "angle": 0, + "content": "Table 6. Ablation study of part VQ-VAE for motion infilling. We replace the DDPM as part VQ-VAE to predict motions." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.304, + 0.466, + 0.366 + ], + "angle": 0, + "content": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
A*40.594.074.4242.984.240.88
NM [63]40.544.154.3152.744.150.96
MT22.344.064.5291.384.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.376, + 0.469, + 0.432 + ], + "angle": 0, + "content": "Table 7. Ablation study of milestone generation. We compare our method with the variant based on the path generated by \\(\\mathbf{A}^*\\) path planning [18]. NM: Neural Mapper [63]. MT: milestone point generation." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.445, + 0.466, + 0.512 + ], + "angle": 0, + "content": "
MethodFD↓APDM↑APDP↑Penetration↓Sliding↓
ConvAE [30]26.503.954.334.820.89
SLT [64]27.953.764.054.130.78
Ours22.344.064.524.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.516, + 0.469, + 0.544 + ], + "angle": 0, + "content": "Table 8. Ablation study of motion infilling module. We compare our method with other motion infilling methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.469, + 0.631 + ], + "angle": 0, + "content": "diversity of NM [63] is better than \\(\\mathbf{A}^*\\), but still worse than our method. The reason why these variants perform poorly might be the low diversity of trajectory that affects the distribution of generated motions for calculating FD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.469, + 0.743 + ], + "angle": 0, + "content": "Motion infilling. To validate our motion infilling module, we compare it with ConvAE [30] and SLT [64]. We only replace the motion infilling module and keep the others the same. The comparison of motion quality and diversity is shown in Tab. 8 and our method outperforms ConvAE [30] and SLT [64] with lower FD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.855 + ], + "angle": 0, + "content": "DDPM vs. VAE. To show the effectiveness of DDPM, we implement a cVAE variant, where we simply replace our transformer DDPM with transformer cVAE [45, 63]. As shown in Tab. 9, although cVAE models could generate more diverse trajectories, their motion quality is far from satisfactory, indicated by the much higher value of FD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Comparison with other diffusion models. Our approach stands out from architectures in MDM [59] and FLAME" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.09, + 0.89, + 0.133 + ], + "angle": 0, + "content": "
ArchFD↓R-APDM↑R-APDP↑R-APDT↑Penetration↓Sliding↓
cVAE29.323.653.58109.555.061.12
DDPM22.344.064.5291.384.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.143, + 0.892, + 0.2 + ], + "angle": 0, + "content": "Table 9. Evaluation of the architecture for our generation framework. Although cVAE based variant generates more diverse trajectories, the motion quality drops significantly as indicated by the much higher FD. Arch stands for the architecture type." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.213, + 0.889, + 0.284 + ], + "angle": 0, + "content": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
MDM [59]38.694.013.7772.675.190.70
MDM [59] + C23.974.024.1389.804.870.39
FLAME [31]28.934.024.4665.465.780.89
Ours22.344.064.5291.384.000.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.315 + ], + "angle": 0, + "content": "Table 10. Comparison with MDM and FLAME. C denotes the frame-wise conditions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.341, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[31] by incorporating frame-wise conditions. Tab. 10 demonstrates the significance of the frame-wise conditions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.892, + 0.443 + ], + "angle": 0, + "content": "More analyses and ablation studies. More detailed analyses and ablation studies of our design choices are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.456, + 0.627, + 0.47 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.601 + ], + "angle": 0, + "content": "Although our method can generate diverse and natural motions, there are still some limitations. Our method is offline and cannot be applied to interactive scenarios. We assume that the objects are static and cannot handle moving objects. The diffusion models require a long inference time. It takes 7.13 seconds on average for a 720-frame sequence on a Titan Xp GPU. The slow speed might be solved by methods that could accelerate diffusion models [43, 2]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.619, + 0.634 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In this work, we propose a novel hierarchical pipeline for motion synthesis of human-object interactions. Our approach first generates the goal pose and then predicts a set of milestones. Next, we synthesize motions with the guide of milestones. Furthermore, we apply DDPM in our hierarchical pipeline. We also show that our framework could generate more diverse and natural human-object interaction motions than other methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.667, + 0.801 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We thank Jintao Lu, Zhi Cen, Zizhang Li, and Kechun Xu for the valuable discussions. This work was partially supported by the Key Research Project of Zhejiang Lab (No. K2022PG1BB01), NSFC (No. 62172364), and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15069" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[3] German Barquero, Sergio Escalera, and Cristina Palmero. BeLFusion: Latent Diffusion for Behavior-Driven Human Motion Prediction. arXiv e-prints, page arXiv:2211.14304, Nov. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[4] Emad Barsoum, John R. Kender, and Zicheng Liu. Hpgan: Probabilistic 3d human motion prediction via gan. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1499-149909, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[5] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, Ding Liu, Jing Liu, and Nadia Magnenat Thalmann. Learning progressive joint propagation for human motion prediction. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, page 226-242, Berlin, Heidelberg, 2020. Springer-Verlag. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.47, + 0.57 + ], + "angle": 0, + "content": "[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, page 387-404, Berlin, Heidelberg, 2020. Springer-Verlag. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.47, + 0.641 + ], + "angle": 0, + "content": "[7] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 5887-5895, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.643, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your Commands via Motion Diffusion in Latent Space. arXiv e-prints, page arXiv:2212.04048, Dec. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.701, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[9] Simon Clavet. Motion matching and the road to next-gen animation. In GDC, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6518-6526, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A Framework for Denoising-Diffusion-based Motion Synthesis. arXiv eprints, page arXiv:2212.04495, Dec. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "volume 34, pages 8780-8794. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, June 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.894, + 0.25 + ], + "angle": 0, + "content": "[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.895, + 0.323 + ], + "angle": 0, + "content": "[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396-1406, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.894, + 0.38 + ], + "angle": 0, + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, oct 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.894, + 0.453 + ], + "angle": 0, + "content": "[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, June 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.894, + 0.511 + ], + "angle": 0, + "content": "[18] Peter E. Hart, Nils J. Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. 2, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.894, + 0.595 + ], + "angle": 0, + "content": "[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11374-11384, October 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[21] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH '23, New York, NY, USA, 2023. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. arXiv e-prints, page arXiv:2207.12598, July 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15070" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. arXiv e-prints, page arXiv:2204.03458, Apr. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.148, + 0.471, + 0.189 + ], + "angle": 0, + "content": "[26] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4), jul 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[27] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based Generation, Optimization, and Planning in 3D Scenes. arXiv e-prints, page arXiv:2301.06015, Jan. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.247, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[28] Ashesh Jain, Amir R. Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5308-5317, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[29] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Trans. Graph., 41(3), jun 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[30] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[31] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. FLAME: Freeform Language-based Motion Synthesis & Editing. arXiv e-prints, page arXiv:2209.00349, Sept. 2022. 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[32] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, page arXiv:1412.6980, Dec. 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.484, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[33] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, page arXiv:1312.6114, Dec. 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[34] Jogendra Nath Kundu, Maharshi Gor, and R. Venkatesh Babu. Bihmp-gan: Bidirectional 3d human motion prediction gan. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'19/IAAI'19/EAAI'19. AAAI Press, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[35] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.707, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[36] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), aug 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.749, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[37] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[38] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained Diffusion Models for Unified Human Motion Synthesis. arXiv e-prints, page arXiv:2212.02837, Dec. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Multi-level motion attention for human motion prediction. International Journal of Computer Vision, 129(9):2513-2535, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[40] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[41] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[42] Tomohiko Mukai and Shigeru Kuriyama. Geostatistical motion interpolation. ACM Trans. Graph., 24(3):1062-1070, jul 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8162-8171. PMLR, 18-24 Jul 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[44] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference (BMVC)*, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[45] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[46] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. arXiv e-prints, page arXiv:2209.14988, Sept. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, page arXiv:2204.06125, Apr. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[51] C. Rose, M.F. Cohen, and B. Bodenheimer. Verbs and adverbs: multidimensional motion interpolation. IEEE Computer Graphics and Applications, 18(5):32-40, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv e-prints, page arXiv:2209.14792, Sept. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15071" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[55] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 41(4), jul 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[56] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6), nov 2019. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[57] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), aug 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.47, + 0.419 + ], + "angle": 0, + "content": "[58] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[59] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.47, + 0.519 + ], + "angle": 0, + "content": "[60] Jonathan Tseng, Rodrigo Castellon, and C. Karen Liu. EDGE: Editable Dance Generation From Music. arXiv eprints, page arXiv:2211.10658, Nov. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[61] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[63] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20460-20469, June 2022. 1, 2, 3, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[64] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9401-9411, June 2021. 2, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[65] J. Wang, S. Yan, B. Dai, and D. Lin. Scene-aware generative network for human motion synthesis. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "(CVPR), pages 12201-12210, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[66] Mao Wei, Liu Miaomiao, and Salzemann Mathieu. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[67] Mao Wei, Liu Miaomiao, Salzemann Mathieu, and Li Hongdong. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[68] D.J. Wiley and J.K. Hahn. Interpolation synthesis of articulated figure motion. IEEE Computer Graphics and Applications, 17(6):39-45, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[69] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[70] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to generate multimodal human dynamics. In European Conference on Computer Vision, pages 276–293. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[71] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[72] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 37(4), jul 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[73] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 40(4), jul 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[74] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[75] S. Zhang, Y. Zhang, Q. Ma, M. J. Black, and S. Tang. Place: Proximity learning of articulation and contact in 3d environments. In 2020 International Conference on 3D Vision (3DV), pages 642-651, Los Alamitos, CA, USA, nov 2020. IEEE Computer Society. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[76] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[77] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[78] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3d people in scenes without people. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[79] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15072" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[80] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[81] Yi Zhou, Zimo Li, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent networks for extended complex human motion synthesis. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[82] Zixiang Zhou and Baoyuan Wang. UDE: A Unified Driving Engine for Human Motion Generation. arXiv e-prints, page arXiv:2211.16016, Nov. 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15073" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_origin.pdf b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..97c4151a2facebd256c71061e8d03e26e8938a33 --- /dev/null +++ b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/c8490beb-3559-4e1e-833f-b399932e108c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:029057ae9dd13d053848142f55c7b8daeff466aa5f57d64a5c05b85d25b693c4 +size 10022533 diff --git a/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/full.md b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..309a9fcafa39d88154a940b95270bad844b6d7f5 --- /dev/null +++ b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/full.md @@ -0,0 +1,448 @@ +# Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models + +Huaijin Pi $^{1}$ , Sida Peng $^{1}$ , Minghui Yang $^{2}$ , Xiaowei Zhou $^{1}$ , Hujun Bao $^{1*}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Ant Group + +# Abstract + +This paper presents a novel approach to generating the 3D motion of a human interacting with a target object, with a focus on solving the challenge of synthesizing long-range and diverse motions, which could not be fulfilled by existing auto-regressive models or path planning-based methods. We propose a hierarchical generation framework to solve this challenge. Specifically, our framework first generates a set of milestones and then synthesizes the motion along them. Therefore, the long-range motion generation could be reduced to synthesizing several short motion sequences guided by milestones. The experiments on the NSM, COUCH, and SAMP datasets show that our approach outperforms previous methods by a large margin in both quality and diversity. The source code is available on our project page https://zju3dv.github.io/hghoi. + +# 1. Introduction + +Scene-aware motion generation [19] aims to synthesize 3D human motion given a 3D scene model to enable virtual humans to naturally wander around scenes and interact with objects, which has a variety of applications in AR/VR, filmmaking, and video games. + +Unlike traditional motion generation methods for character control which aim to generate short or repeated motion on the fly guided by a user's control signals [56], we focus on the setting of generating long-term human-object interactions [56, 19, 76] given a starting position of the human and a target object model. This setting brings in new challenges. First, the entire approaching process and the human-object interaction should be coherent, which requires the capability of modeling long-range interaction between the human and the object. Second, in the context of content generation, the generative model should be able to synthesize diverse motions as there are many plausible ways for a real human to approach and interact with the target object. + +![](images/2f698d33418202babdfca7f4b06968b8d90235acb490c959682dc07f3ffdd37e.jpg) + +![](images/9b4ada4e7b0ca0ea6c34e2f2fa1ba622b75b166f000e246c077885f84368a241.jpg) +Figure 1. Genration of human-object interactions. Given an object, our method first predicts a set of milestones, where the rings indicate the positions and the humans with pink clothes represent the local poses. Then the motions are inflated between milestones. This figure shows that our method can generate diverse milestones and motions with the same object. The flow of time is shown with a color code where darker blue denotes the later frame. + +Existing methods for motion synthesis can be roughly characterized into online generation and offline generation. Most online methods [56, 19, 76] focus on real-time control of characters. Given a target object, they generally use auto-regressive models to recurrently generate future motions by feeding their predictions. Although they have been widely used for interactive scenarios like video games, their motion quality is not satisfactory enough for long-term generation [63]. A plausible cause is the error accumulation in the auto-regressive process, where the errors in previous predictions are fed back as the model input, as discussed in [40, 26, 45, 46, 22]. To improve the motion quality, + +some recent offline methods [6, 65, 64, 63] employ a multistage framework, which first generates the trajectory and then synthesizes motions. TDNS [63] generates paths by combining the cVAE model [33] and deterministic planning methods like A* [18]. Although this strategy can produce reasonable paths, the path diversity is limited, as demonstrated by our experimental results in Sec. 4.4. + +In this paper, we propose a novel offline approach for synthesizing long-term and diverse human-object interactions. Our innovation lies in a hierarchical generation strategy, which first predicts a set of milestones and then generates human motions between milestones. Fig. 1 illustrates the basic idea. Specifically, given the starting position and the target object, we design a milestone generation module to synthesize a set of milestones along the motion trajectory, each of which encodes the local pose and indicates the transition point during the human movement. Based on these milestones, we employ a motion generation module to produce the full motion sequence. Thanks to the milestones, we simplify the long-sequence generation into synthesizing several short motion sequences. Furthermore, the local pose at each milestone is generated by a transformer that considers the global dependency [62], leading to temporally consistent results, which further contribute to coherent motions. + +In addition to our hierarchical generation framework, we further exploit diffusion models [53, 23, 54] to synthesize human-object interactions. Previous diffusion models for motion synthesis [31, 59] combine transformer [62] and Denoising Diffusion Probabilistic Model (DDPM) [23]. Directly applying them to our setting is prohibitively comput-intensive due to the long motion sequences and may lead to the GPU memory explosion [50]. Because our hierarchical generation framework converts the long-term generation to the synthesis of several short sequences, the required GPU memory is reduced to the same level of short-term motion generation. Therefore, we can efficiently leverage the transformer DDPM to synthesize long-term motion sequences, which improves the generation quality. + +We validate our design choices on the NSM [56], COUCH [76], and SAMP [19] datasets with extensive experiments. On these datasets, our hierarchical framework outperforms previous methods significantly in both motion quality and diversity. + +# 2. Related Work + +# 2.1. Motion synthesis + +Motion synthesis is a long-standing problem in computer graphics and vision [68, 51, 42, 9]. With the rapid development of deep learning, recent works have applied neural networks to motion synthesis [14, 40, 26, 72]. Some methods are deterministic [28, 81, 44, 67] while others try to predict stochastic motions by VAE [70, 1, 71, 79, 45] and + +GAN [4, 34]. To further improve the performance, some works apply GCN [67, 10, 66, 35, 39] or transformers [5] to extract the features from the human skeleton. To handle the ambiguity of human motion, some works [26, 55] propose to employ phase signals to guide the motions. Recent works start to consider the scene context [56, 6]. NSM [56] is the first work that aims at synthesizing human motions with object-level interactions with specific action labels. Based on NSM, SAMP [19] applies cVAE to predict diverse motions. These works [56, 19, 76] mainly focus on the interaction with one or two objects while others [6, 65] generate motions with a full scene (including the information about floors and walls) as the input. [6] predicts future motions of the given motion sequence by first predicting the trajectory and then generating motions based on a 2D image of the scene. Similar to this pipeline, [65] applies GAN [16]. Furthermore, [64] proposes a framework that first places pose on a human-provided path and then synthesizes motions in a 3D scene with a full-scene scan. [63] combines $\mathbf{A}^*$ and cVAE to generate diverse human motions. To further control human motion, [80] employs the gaze, and COUCH [76] explicitly models the hand contacts to guide the prediction. Some methods [7, 21] also enable physically simulated characters to perform scene interaction tasks including sitting [7, 21] and carrying boxes [21]. Some works also focus on grasp [58, 69], manipulation [73] and the interaction with dynamic objects [57]. + +Ours vs. others. This work follows the setting of [56, 19] and focuses on object-level interaction. In contrast to [56, 19] which generate motions mainly based on auto-regressive models, we design a hierarchical framework to synthesize motions. Different from [6, 64], our work focuses on much more long-term generation (longer than 10 seconds) while [6] is 2 seconds and [64] is 6 seconds. Instead of planning the path auto-regressively with an extra network to generate diverse trajectories like [63], our method directly predicts a set of milestones to describe the approaching process which is inherently diverse. In addition, most methods [19, 76, 64, 63] rely on cVAE to generate stochastic motions while we exploit DDPM [23] to synthesize trajectories and motions. + +# 2.2. Diffusion models + +Diffusion models [53] are a class of likelihood-based methods that generate samples by gradually removing noise from a signal. Then, [23, 54] develop the diffusion models for high-quality image generation. To control the generated results, [12] proposes classifier guidance for trading off diversity for fidelity. Later, the classifier-free model [24] achieves better results [48] in text-conditioned image generation. In addition, diffusion models have been successfully applied to other domains like the generation of videos + +![](images/e94b5e2623b8eb9c14a45cbda549c0033009a05eed1a24688830e82bd0525162.jpg) +Figure 2. Overview of our pipeline. Our pipeline consists of three components: First, the goal pose is synthesized given the object. Then, a number of milestones with local poses are predicted based on the goal pose. Finally, the trajectory and the full motion sequences are inflated between the milestones. + +[52, 25] and 3D contents [47]. + +There are some works [74, 59, 31, 38] that apply DDPM to synthesize motions. [3, 8] also explore latent diffusion models [50] for motion generation. However, most works focus on text-conditioned [74, 11], audio-driven [82] and music-driven [60, 82] motion generation while we target human-object interaction. In this work, we apply transformer DDPM [23] to a multi-stage framework, which separately generates trajectories and synthesizes motions. + +There are some concurrent works [27] that apply DDPM to synthesize motions in a scene. However, [27] is designed for short-term motion generation (around 2 seconds) while we target long-term human motion generation (longer than 10 seconds). Different from existing work [59, 27], we employ DDPM in a multi-stage framework, where trajectories and motions are separately predicted. + +# 3. Methods + +Given the object $\mathbf{I}$ and the starting point $\mathbf{s}$ , our goal is to synthesize 3D human motions $\{(\mathbf{r}_i, \pmb{\theta}_i)\}_{i=1}^N$ with human-object interactions. $\mathbf{r}_i$ is the root trajectory, and the $\pmb{\theta}_i$ indicates local pose at $i$ -th frame. + +We design a hierarchical motion generation framework as shown in Fig. 2. First, we employ GoalNet[19] to predict an interaction goal on the object. Then, we generate the goal pose (Sec. 3.1) to explicitly model the human-object interaction. Next, our milestone generation module (Sec. 3.2) estimates the length of the milestones, produces the trajectory of the milestones from the starting point to the goal, and places milestone poses. Therefore, the long-range motion generation is decomposed into combinations of short-range motion synthesis. Finally, we design a motion generation module (Sec. 3.3) to synthesize the trajectory between the milestones and infill the motions. + +# 3.1. Goal pose generation + +We call the pose in which a person interacts with an object and remains stationary a goal pose. To synthesize diverse human motion, we first generate a goal pose interacting with the object following [58, 69, 63]. Most methods [78, 75, 20] generate human poses using the cVAE model. + +![](images/dd7b467a6441efe6754b756c098c3fe2730a3f5db2ca96632998109e8c5fe610.jpg) +Figure 3. Part VQ-VAE. Part VQ-VAE first splits the skeleton into multiple parts and learns the codebooks separately. The composition of different parts is subsequently modeled with the autoregressive prediction model. + +They project the poses to the standard normal distribution in the continuous space [33]. Empirically, we find that cVAE models do not perform well in our setting. To overcome this challenge, we introduce VQ-VAE [61, 49] to model the data distribution which exploits a discrete representation to cluster the data in a finite set of points [37]. We hypothesize that limited data of goal pose from the SAMP dataset [19] can always be clustered by VQ-VAE but may not be enough for learning a continuous latent space for VAE [41]. In addition, based on the observation that different human poses may share similar properties [15, 29] (e.g., humans may sit with different hand positions but the same leg positions), we split joints into $L$ ( $L = 5$ ) different non-overlapping groups like MotionPuzzle [29]. + +Quantization. As shown in Fig. 3, the goal pose $\theta_{g}$ is split into separate joint groups as $\theta_{g} = \{\theta_{gi}\}_{i=1}^{L}$ . Then a discrete codebook $\mathbf{Z}_i$ with a list of vectors compares the output $\hat{\mathbf{z}}_i$ from the encoder $E_i$ to find the closest vector in Euclidean distance. The $L$ vectors with minimal distances will be concatenated and fed into a shared decoder $D$ to + +reconstruct $\theta_{g}$ . The loss function is defined as: + +$$ +\mathcal {L} \left(\boldsymbol {\theta} _ {g}, D (\mathbf {z})\right) = \| \boldsymbol {\theta} _ {g} - D (\mathbf {z}) \| _ {2} ^ {2} + \sum_ {i = 1} ^ {L} \| s g \left[ E _ {i} \left(\boldsymbol {\theta} _ {g i}\right) \right] - \mathbf {z} _ {i} \| _ {2} ^ {2} \tag {1} +$$ + +$$ ++ \beta \sum_ {i = 1} ^ {L} \| s g [ \mathbf {z} _ {i} ] - E _ {i} (\boldsymbol {\theta} _ {g i}) \| _ {2} ^ {2}, +$$ + +where + +$$ +\mathbf {z} = \left[ \mathbf {z} _ {1}, \dots , \mathbf {z} _ {L} \right], \tag {2} +$$ + +$$ +\mathbf {z} _ {i} = \underset {\mathbf {z} _ {i} ^ {\prime} \in Z _ {i}} {\arg \min } \| \mathbf {z} _ {i} ^ {\prime} - \hat {\mathbf {z}} _ {i} \|, \tag {3} +$$ + +$$ +\hat {\mathbf {z}} _ {i} = E \left(\boldsymbol {\theta} _ {g i}\right). \tag {4} +$$ + +Here, the term $sg[\cdot]$ denotes the stop-gradient operator, and $\| sg[\mathbf{z}_i] - E_i(\pmb{\theta}_{gi})\|_2^2$ is the commitment loss [49] controlled by a weighting factor $\beta$ . + +Generation. With $E_{i}$ and $D$ available, we can represent $\pmb{\theta}_{g} = \{\pmb{\theta}_{g1},\dots,\pmb{\theta}_{gL}\}$ by a sequence of the part-based codebook indices. To be more specific, we use $E_{i}$ to extract the feature from $\pmb{\theta}_{gi}$ and find the closest vector $\mathbf{z}_i\in \mathbf{Z}_i$ . Then we use $s_i\in \{0,\ldots ,|\mathbf{Z}_i| - 1\}$ to indicate the index of $\mathbf{z}_i$ in $\mathbf{Z}_i$ . Therefore, $\pmb{\theta}_{g}$ can be represented by $\mathbf{s} = \{s_1,\dots,s_L\}$ . + +To generate a natural goal pose, we convert the problem to predict a sequence of indices that can represent $\theta_{g}$ . We formulate the inference as a conditional auto-regressive process and employ a transformer [62] to learn to predict the distribution of possible indices [13]. The condition contains the environment around the goal $\mathbf{O}_g$ , and the action $\mathbf{a}_g$ . Following NSM [56] and COUCH [76], a cylindrical volume of a pre-defined radius and height is created around the goal. Within this volume, spheres are uniformly sampled and the occupancies corresponding to the object of these spheres are calculated. Then, these occupancies are flattened to form a feature vector denoted as $\mathbf{O}_g$ . $\mathbf{a}_g$ is a vector indicating the action type. These variables are fed into the transformer as tokens. Our target is to learn the likelihood of the sequence: + +$$ +p (\mathbf {s} \mid \mathbf {O} _ {g}, \mathbf {a} _ {g}) = \prod_ {i} p \left(s _ {i} \mid s _ {< i}, \mathbf {O} _ {g}, \mathbf {a} _ {g}\right). \tag {5} +$$ + +After predicting the indices, we map them back to their corresponding codebook entries to get the quantized features $\mathbf{z} = [\mathbf{z}_1,\dots ,\mathbf{z}_L]$ , which are fed into the decoder $D$ to generate the goal pose $\theta_{g}$ . + +# 3.2. Milestone generation + +Based on the starting pose and the goal pose, we can generate the milestone trajectory and synthesize the local poses at the milestones. Following [74, 59], we build a transformer DDPM [23] and apply it to generate the milestones for better quality. Because the length of motion data is unknown and can be arbitrary (e.g., the human could quickly + +walk towards the chair and sit down or sit after walking around the chair slowly), we predict the length of milestones, denoted by $N$ . Then we synthesize $N$ milestone points and place the local poses on these points. + +Transformer DDPM. Here we first briefly introduce DDPM [23], which is learned to reverse a diffusion process. Formally, the diffusion model [53] is defined as a latent variable model of the form $p_{\theta}(\mathbf{x}_0) \coloneqq \int p_{\theta}(\mathbf{x}_{0:T}) d\mathbf{x}_{1:T}$ , where $\mathbf{x}_0 \sim q(\mathbf{x}_0)$ is the data and $\mathbf{x}_1, \ldots, \mathbf{x}_T$ are the latents. $p_{\theta}(\mathbf{x}_{0:T})$ is formulated as a Markov chain as: + +$$ +p _ {\theta} \left(\mathbf {x} _ {0: T}\right) := p \left(\mathbf {x} _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right), \tag {6} +$$ + +$$ +p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) := \mathcal {N} \left(\mathbf {x} _ {t - 1}; \boldsymbol {\mu} _ {\theta} \left(\mathbf {x} _ {t}, t\right), \boldsymbol {\Sigma} _ {\theta} \left(\mathbf {x} _ {t}, t\right)\right). \tag {7} +$$ + +Diffusion models approximate posterior $q(\mathbf{x}_{1:T} \mid \mathbf{x}_0)$ as a Markov chain that gradually adds Gaussian noise to the data with variance schedules given by $\beta_t$ : + +$$ +q \left(\mathbf {x} _ {1: T} \mid \mathbf {x} _ {0}\right) := \prod_ {t = 1} ^ {T} q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right), \tag {8} +$$ + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) := \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right). \tag {9} +$$ + +In contrast to adding noises on $\mathbf{x}_0$ sequentially, DDPM formulates the diffusion process as: + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \tag {10} +$$ + +where $\alpha_{t} = 1 - \beta_{t}$ and $\bar{\alpha}_{t} = \prod_{s = 1}^{t}\alpha_{s}$ . Hence, we can generate $\mathbf{x}_t$ by sampling a noise $\epsilon$ as the training data. DDPM employs a neural network to model $p_{\theta}(\mathbf{x}_{t - 1}\mid \mathbf{x}_t)$ and the inference is to gradually denoise $\mathbf{x}_t$ from $t = T$ to $t = 1$ where $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . + +Like existing works [74, 31] that apply DDPM in the motion domain, we employ a transformer decoder [62] as our architecture of DDPM. The transformer takes the noise $\mathbf{x}_t$ and the condition $\mathbf{C}$ as input. The condition $\mathbf{C}$ means variables related to the generation, which will be described in detail in each subsection. The diffusion time-step $t$ is in the sinusoidal position embeddings form [62] and is injected into each block in the transformer. Different from existing works [59, 74] that assume the length of the motion sequences is already given, we insert a parallel branch to estimate the length of milestones by taking the length token $\mathbf{H}^{tok}$ and the condition $\mathbf{C}$ as the input, as shown in Fig. 4. The length prediction head is an MLP and predicts a multinomial distribution over discrete length indices $\{1, 2, \dots, N_{max}\}$ like [17], where $N_{max}$ represents the + +![](images/efd63b3a288c09bc8835f577071820eacdaea1dc20e574f5c86a51b6d80ebdba.jpg) +Figure 4. Overview of transformer DDPM for milestone generation. The model first takes the length token $\mathbf{H}^{tok}$ and the condition $\mathbf{C}$ as the input to predict the data length. Then it constructs the noise sequence $\mathbf{x}_T^{1:N}$ with length $N$ . In the diffusion process, it is fed with $\mathbf{C}$ and the sequence $\mathbf{x}_t^{1:N}$ at time-step $t$ to predict the target $\hat{\mathbf{x}}_0^{1:N}$ . For other sub-modules, we remove the length prediction head. + +length of the longest sequences for training. We use cross-entropy loss as the loss function. At inference time, we sample the length $N$ from the estimated distribution. Next, we construct $N$ milestones as the input to the transformer. We predict the denoised data $\hat{\mathbf{x}}_0$ based on the condition $\mathbf{C}$ by the DDPM $f$ . This is formulated as $\hat{\mathbf{x}}_0 = f(\mathbf{x}_t,t,\mathbf{C})$ , and the training loss is defined as: + +$$ +\mathcal {L} = \operatorname {E} _ {t \in [ 1, T ], \mathbf {x} _ {0} \sim q (\mathbf {x} _ {0}), \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \| \mathbf {x} _ {0} - \hat {\mathbf {x}} _ {0} \| \right]. \tag {11} +$$ + +In our setting, we use a two-step strategy like [64, 63], where two transformer DDPMs are applied to first generate the milestone points and then synthesize the local pose at each milestone. + +Generation of milestone points. The milestone points are conditioned on the object, the information of the starting point, and the goal. We then define the condition as: + +$$ +\mathbf {C} _ {\mathbf {m}} = \left\{\mathbf {I} _ {s}, \mathbf {I} _ {g}, \mathbf {O} _ {s}, \mathbf {O} _ {g}, \mathbf {g}, \mathbf {s} \right\}, \tag {12} +$$ + +where $\mathbf{I}_s$ and $\mathbf{I}_g$ are the object representation relative to the starting point and the goal. Similar to other methods [56, 19, 76], the object representation is modeled by an $8\times 8\times 8$ grid with their positions and occupancies. Following NSM [56] and COUCH [76], we also explicitly model the occupancies around the starting point and the goal with $\mathbf{O}_s$ and $\mathbf{O}_g$ (the same form as the occupancy feature described in Sec. 3.1). + +The information of goal $\mathbf{g}$ is defined as $\{\mathbf{r}_g, \mathbf{a}_g, \boldsymbol{\theta}_g\}$ , where $\mathbf{r}_g$ means the goal position and orientation in the starting point coordinate system, $\mathbf{a}_g$ is the target action label at the goal, and $\boldsymbol{\theta}_g$ is the goal pose. $\mathbf{s} = \{\mathbf{a}_s, \boldsymbol{\theta}_s\}$ represents the action labels $\mathbf{a}_s$ and the pose $\boldsymbol{\theta}_s$ at the starting point. + +The target is to predict the milestone $\{\mathbf{m}_1, \dots, \mathbf{m}_N\}$ with length $N$ . Following NSM [56], a bi-directional scheme is employed for the milestone points generation where we predict the roots of milestones in both the starting point coordinate system and the goal coordinate system. The final predicted roots are blended from these two kinds of outputs. The milestone point $\mathbf{m}_i$ is defined as: + +$$ +\mathbf {m} _ {i} = \left\{\mathbf {r} _ {i} ^ {b}, \mathbf {c} _ {i}, \mathbf {w} _ {i} \right\}. \tag {13} +$$ + +The representation is similar to previous work [19, 56, 76], where $\mathbf{r}_i^b$ indicates the root position and forward directions relative to the starting point and the goal, $\mathbf{c}_i$ is a label vector indicating the contact between the environment and the body, and we use a high-dimensional feature vector $\mathbf{w}_i$ to encode the character state at the milestone following [56, 19, 76]. The details of these variables are provided in the supplementary material. We use transformer DDPM $f_{\mathbf{m}}$ to predict the length and synthesize milestone points. + +Generation of milestone poses. Previous works [64, 63] separately place the poses on the sparse points along the path and infill the motions between, which may lead to unnatural transitions between these poses. On the contrary, we generate the local pose at each milestone with transformer DDPM to build the temporal dependency. With the help of the generated milestone points, we can access the accurate spatial relationship $\mathbf{I}_i$ between the object and the character and the ego-centric environment occupancies $\mathbf{O}_i$ [76] at the $i$ -th milestone. The milestone poses also depend on the milestone state including $\{\mathbf{c}_i, \mathbf{w}_i\}$ at the $i$ -th milestone. Consequently, we define the condition $\mathbf{C_k}$ as a combination with starting pose $\theta_s$ , goal pose $\theta_g$ , and frame-wise condition $\gamma_i$ at the $i$ -th milestone as: + +$$ +\mathbf {C} _ {\mathbf {k}} = \left\{\boldsymbol {\theta} _ {s}, \boldsymbol {\theta} _ {g}, \gamma_ {1}, \dots , \gamma_ {N} \right\}, \tag {14} +$$ + +$$ +\text {w h e r e} \quad \gamma_ {i} = \left\{\mathbf {I} _ {i}, \mathbf {O} _ {i}, \mathbf {c} _ {i}, \mathbf {w} _ {i} \right\}. \tag {15} +$$ + +The local poses at the milestones are predicted by the transformer DDPM $f_{\mathbf{k}}$ without the length prediction head. + +# 3.3. Motion generation + +Instead of predicting motions frame-by-frame [56, 19, 76], our approach hierarchically synthesizes the full sequence based on the generated milestones. We follow [6, 64, 63] to first generate the trajectory and then synthesize the motions. Specifically, within two consecutive milestones, we first complete the trajectory. Then, the motions are inflated with the guidance of successive milestone + +poses. The two steps are accomplished using two transformer DDPMs (described in Sec. 3.2), respectively. For each step, we carefully design the condition of DDPM to generate the target output. + +Trajectory completion. For the trajectory completion between the milestones $\mathbf{m}_i$ and $\mathbf{m}_{i + 1}$ , we assume that it is only conditioned on the milestones and the object. Thus we define the condition as: + +$$ +\mathbf {C} _ {\mathbf {r}} = \left\{\mathbf {I} _ {i}, \mathbf {I} _ {i + 1}, \mathbf {O} _ {i}, \mathbf {O} _ {i + 1}, \mathbf {m} _ {i}, \mathbf {m} _ {i + 1}, \mathbf {t} _ {i} ^ {i + 1} \right\}, \tag {16} +$$ + +where $\mathbf{I}_i$ indicates the object representation relative to the milestone $i$ and $\mathbf{O}_i$ denotes the ego-centric occupancies (the same form as the occupancy feature described in Sec. 3.1) around the $i$ -th milestone like NSM [56] and COUCH [76]. $\mathbf{m}_i$ has been shown in Eq. (13). $\mathbf{t}_i^{i+1}$ represents the position and orientation of the $(i+1)$ -th milestone in the $i$ -th milestone's coordinate system. Between two consecutive milestones, we generate the trajectory with a length of 2 seconds, which is 61 frames. Similar hyperparameters can be found in previous methods [64, 63]. + +The target output is the trajectory between two consecutive milestones. Similar to the milestone point, we synthesize the trajectory in the $j$ -th frame with a bi-directional scheme. The trajectory is composed of a set of points that have the same representation as the milestone in Eq. (13). We generate the trajectory with a transformer DDPM $f_{\mathbf{r}}$ similar to the one in Sec. 3.2. Since we assume the trajectory length between two milestones is fixed, the DDPM $f_{\mathbf{r}}$ does not have the length prediction head. + +Motion infilling. To synthesize the long-range motion, we convert a long sequence into several fixed-length short sequences with the help of milestone points and milestone poses. For a sub-sequence between the consecutive milestone poses, our goal is to generate the missing local poses over the trajectory. The generated motion has to satisfy the trajectory and naturally transits from a milestone to the next milestone. Like milestone pose generation, we use the same representation of frame-wise condition in Eq. (15). The condition is defined as: + +$$ +\mathbf {C} _ {\mathbf {p}} = \left\{\boldsymbol {\theta} ^ {1}, \boldsymbol {\theta} ^ {6 1}, \gamma^ {1}, \dots , \gamma^ {6 1} \right\}, \tag {17} +$$ + +where $\theta^1$ and $\theta^{61}$ are the local poses of two consecutive milestones. By taking these inputs, we generate smooth motions using another transformer DDPM $f_{\mathbf{p}}$ without the length prediction head. + +# 4. Experiments + +# 4.1. Implementation details + +We train the part VQ-VAE and transformer DDPM models with the Adam optimizer [32]. All the models are trained + +with a fixed learning rate of 0.0001 with batch size 256. The remaining details are in the supplementary material. + +# 4.2. Datasets and evaluation metrics + +Test setting. Our experiments are conducted on the SAMP [19], COUCH [76], and NSM [56] datasets. Provided with a starting point, a starting pose, an object, and an endpoint, the virtual human is asked to approach the object, interact with it, and leave to reach the endpoint. The ablation studies are conducted on the SAMP dataset. + +Metrics. Following the previous method [19], we calculate the Fréchet distance (FD) between the generated and ground-truth motions to measure the motion quality. We also conduct user studies and each sequence is evaluated by at least 3 users with scores ranging from 1 to 5. In addition, we calculate the penetration ratio [64, 75, 78] and foot sliding [72, 36] to show the physical plausibility between the 3D object and the synthesized motions. We compute the Average Pairwise Distance (APD) [71, 77] to evaluate the diversity. Specifically, we calculate the APD of synthesized motion, the character's pose during object interactions, and trajectories. Following previous work [56, 19], we calculate PE (positional errors) and RE (rotational errors) to indicate the precision of object interactions. For each test object, we generate multiple sequences. More details are included in the supplementary material. + +# 4.3. Comparison with other methods + +Results on the SAMP dataset. On the SAMP dataset [19], we compare our method with online methods SAMP [19] and MoE [72]. As our method is offline, we also implement and modify offline methods SLT [64] and TDNS [63] to our setting. For SLT [64], we employ $\mathrm{A}^*$ [18] to plan a path and select points along the path as subgoals to form the input for SLT. More details about the implementation of [64, 63] are in the supplementary material. Since MoE often fails to finish the action, we do not calculate the penetration ratio for it. As shown in Tab. 1, our approach outperforms other methods in terms of lower FD, higher user study scores, and higher APD. Furthermore, our method achieves much higher trajectory diversity than SAMP [19]. Although TDNS [63] proposes Neural Mapper (NM) which combines $\mathrm{A}^*$ [18] and cVAE, the diversity of the generated trajectory is inferior to our method, as indicated by $\mathrm{APD}_T$ . + +Results on the COUCH dataset. As our target is to synthesize diverse motions instead of controlling the characters, we only evaluate the motion quality on the COUCH dataset [76]. Tab. 2 shows that our method outperforms all baselines. Our approach achieves much higher $\mathrm{APD}_T$ than other methods. We observe that $\mathrm{APD}_T$ of TDNS [63] is + +
MethodFD ↓User study ↑APDM ↑APDP ↑APDT ↑PE↓RE↓Penetration↓Sliding↓
MoE [72]74.332.763.502.6352.46∞∞-1.68
SAMP [19]57.342.863.633.0563.183.444.126.981.02
SLT* [64]68.832.303.082.6640.131.771.604.281.71
TDNS* [63]46.602.903.683.4066.040.450.395.140.94
Ours22.343.624.064.5291.380.390.324.000.50
+ +Table 1. Quantitative results on the SAMP dataset. SLT* and TDNS* mean we modify and implement them on the SAMP dataset. The subscript “M”, “P” and “T” stand for “Motion”, “Pose” and “Trajectory”. PE and RE represent the positional error and rotation error. Sliding denotes foot sliding. $\infty$ means the method failed to reach the goal. + +
MethodFD↓User study ↑APDM↑APDP↑APDT↑Penetration↓Sliding↓
NSM [56]118.982.990008.200.59
SAMP [19]160.121.940.890.184.694.940.72
COUCH [76]127.193.051.410.6623.487.430.37
SLT* [64]93.463.011.681.172.863.901.85
TDNS* [63]71.723.292.702.1916.985.121.17
Ours56.354.273.222.3064.963.540.55
+ +Table 2. Quantitative results on the COUCH dataset. SLT* and TDNS* mean we modify and implement them. + +
MethodFD↓User study↑APDM↑APDP↑APDT↑PE↓RE↓Penetration↓Sliding↓
NSM [56]90.393.950001.720.406.430.69
SAMP [19]68.863.771.210.1120.014.774.847.831.38
Ours57.024.042.620.9362.231.010.284.850.80
+ +![](images/e24db9e6a07fcb9fcc57645ee0972f302e45bbe19bc8d4092e2bb9c0aaed7126.jpg) +Figure 5. Results in a cluttered scene. Our method can generate motions that avoid obstacles in a cluttered scene. + +higher than the approaches [19, 64] that employ deterministic $\mathrm{A}^*$ [18], but much lower than our method. Although COUCH [76] exhibits lower foot sliding than our method, it may sometimes get stuck, resulting in lower foot sliding since the character does not move. + +Results on the NSM dataset. On the NSM dataset [56], we compare our approach with SAMP [19] and NSM [56]. Tab. 3 shows that our method outperforms baselines. Compared with the deterministic method NSM, our approach could generate stochastic motion and diverse trajectories. + +Table 3. Quantitative results on the NSM dataset. + +
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
w/o GP31.073.212.4685.214.100.49
w/o MT27.873.623.01127.945.300.72
w/o MP25.144.013.7885.694.580.43
w/o TC36.772.762.4283.264.340.63
Ours22.344.064.5291.384.000.50
+ +Table 4. Ablation study of the impact of sub-modules. Although the variant without MT generates more diverse trajectories as shown in $\mathrm{APD}_T$ , the motion quality is much worse as indicated by the higher FD, penetration, and sliding. w/o: without. GP: goal pose generation. MT: milestone point generation. MP: milestone pose generation. TC: trajectory completion. + +Results in a cluttered scene. We show our generated results in a cluttered scene in Fig. 5. The percentage of frames with penetration is $3.8\%$ for our method and $4.9\%$ for SAMP. More details are in the supplementary material. + +Qualitative results. As demonstrated in Fig. 6, our approach achieves better results than baselines [19, 63] on the SAMP dataset. Fig. 7 compares our method with COUCH and TDNS on the COUCH dataset. More qualitative results are in the supplementary material. + +# 4.4. Ablation study + +Impact of each sub-module. To show the effectiveness of our hierarchical design, we evaluate our method against four variants where we remove one sub-module for each variant. Tab. 4 indicates that each component improves performance. Generating a whole trajectory leads to more diverse trajectories but the motion quality is worse and has more foot sliding. This validates the necessity of the separate generation of trajectories and motions. + +Goal pose generation. We evaluate our goal pose generator with several variants, including cVAE, DDPM, and standard VQ-VAE. For this evaluation, we only replace the goal pose generation module and keep the others the same. + +Tab. 5 shows that part VQ-VAE generates more diverse poses than continuous latent space models. The comparison + +![](images/8f9493cac2b81052dc5a587d9c057ebf9b446360c9363ca85d6243807b0f7431.jpg) + +![](images/0ba3fbe2fb50f1044eb3563862c079b92157e47faaadf7473cc907da1152334c.jpg) + +![](images/8dcac349f5f41eaa22b12a170e752aa93073df06a6aa1efc680f2e89c278a8e7.jpg) + +![](images/bc15e7ad015fa69354d59bc3d47de8e65f6232ba24aa41155eb93e183479aea4.jpg) + +![](images/4dba9b99cd50a9a74f708a7de38f81c37f818b8ffc42918eff0e240b0ab814e3.jpg) + +![](images/af57b0727e3512349c25ebf78da51e9323b9aa52d1e47624881441957ecb2bf9.jpg) + +![](images/8212b3241dd2b7457ded4c158602a2a1bc7f719d8393d20d4beb51b9356eefa4.jpg) +SAMP + +![](images/4dd1961b5870def3098fcee476ff9f9757cab68e23d62861a712b6b58eb06080.jpg) +TDNS + +![](images/61a6ba53db2a0cc9c475b2fb2e92c509b0c4911bbc940644f2c5837ebc30a58a.jpg) +Ours +Figure 6. Qualitative results on the SAMP dataset. We compare our method with the baselines SAMP [19] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that SAMP and TDNS tend to walk backward with more foot sliding. The second row shows that the baselines start to sit and lie down before reaching the object while our method synthesizes more natural results. The third row demonstrates that our framework has the capability to generate a long trajectory and walk naturally along the trajectory while SAMP gets stuck near the object and TDNS synthesizes unnatural motions. Lines on the floor indicate trajectories. The human with pink clothes indicate the start position. Darker color denotes later frames in the sequence. + +![](images/43cc387d22673748ba60534607c8a53f6cbfb000e1123e50884bca7cd93e8312.jpg) + +![](images/d26e69c4f57a92c329d6c93b3d95018e99321d394730fe2296be3935266467a3.jpg) + +![](images/abb3e4cab11a0b18ed1742fe33e0c14c3a74bc39e44041eb4e40918ab38b1164.jpg) + +![](images/53e0383f74e904058261f0bf153a081b7d8fb539361621e5f3f85265db8004fd.jpg) +COUCH + +![](images/289ff845a0eda095049d9aa78e4f87ea0980709f4cc7b77e8443f280218ea88f.jpg) +TDNS + +![](images/c898e148c0c93190d136944fdcbeb29117abad7195ef1d95300a5c9c5ff234d6.jpg) +Ours +Figure 7. Qualitative results on the COUCH dataset. We compare our method with the baseline COUCH [76] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that COUCH and TDNS tend to sit when the character is far from the object. The second row shows that COUCH may fail to stand up and TDNS may stand up unnaturally. To better visualize the results of TDNS in the second row, we only keep the frames in which the human stands up and goes to the endpoint. + +with standard VQ-VAE shows the necessity of our part design. We also try part VQ-VAE for motion infilling, but the results in Tab. 6 show that its performance is worse. + +Milestone generation. To further investigate the impact of milestones, we compare our approach with the variants + +that employ path-planning methods and select points along the path as milestones. For this ablation, we implement $\mathrm{A}^*$ and NM [63] proposed by TDNS. As demonstrated in Tab. 7, the diversity of generated trajectories of $\mathrm{A}^*$ [18] is much worse, and the motion quality drops significantly, as indicated by the lower $\mathrm{APD}_T$ and higher FD. The trajectory + +
VariantsFD ↓APDM ↑APDP ↑APDT ↑Penetration↓Sliding↓
cVAE27.063.363.2790.524.360.47
DDPM34.223.753.3784.764.310.49
VQ-VAE24.773.783.8789.194.270.48
Part VQ-VAE22.344.064.5291.384.000.50
+ +Table 5. Ablation study of goal pose generation. We implement the goal pose module with different architectures. + +
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
Part VQ-VAE28.713.853.6283.514.320.89
DDPM22.344.064.5291.384.000.50
+ +Table 6. Ablation study of part VQ-VAE for motion infilling. We replace the DDPM as part VQ-VAE to predict motions. + +
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
A*40.594.074.4242.984.240.88
NM [63]40.544.154.3152.744.150.96
MT22.344.064.5291.384.000.50
+ +Table 7. Ablation study of milestone generation. We compare our method with the variant based on the path generated by $\mathbf{A}^*$ path planning [18]. NM: Neural Mapper [63]. MT: milestone point generation. + +
MethodFD↓APDM↑APDP↑Penetration↓Sliding↓
ConvAE [30]26.503.954.334.820.89
SLT [64]27.953.764.054.130.78
Ours22.344.064.524.000.50
+ +diversity of NM [63] is better than $\mathbf{A}^*$ , but still worse than our method. The reason why these variants perform poorly might be the low diversity of trajectory that affects the distribution of generated motions for calculating FD. + +Motion infilling. To validate our motion infilling module, we compare it with ConvAE [30] and SLT [64]. We only replace the motion infilling module and keep the others the same. The comparison of motion quality and diversity is shown in Tab. 8 and our method outperforms ConvAE [30] and SLT [64] with lower FD. + +DDPM vs. VAE. To show the effectiveness of DDPM, we implement a cVAE variant, where we simply replace our transformer DDPM with transformer cVAE [45, 63]. As shown in Tab. 9, although cVAE models could generate more diverse trajectories, their motion quality is far from satisfactory, indicated by the much higher value of FD. + +Comparison with other diffusion models. Our approach stands out from architectures in MDM [59] and FLAME + +Table 8. Ablation study of motion infilling module. We compare our method with other motion infilling methods. + +
ArchFD↓R-APDM↑R-APDP↑R-APDT↑Penetration↓Sliding↓
cVAE29.323.653.58109.555.061.12
DDPM22.344.064.5291.384.000.50
+ +Table 9. Evaluation of the architecture for our generation framework. Although cVAE based variant generates more diverse trajectories, the motion quality drops significantly as indicated by the much higher FD. Arch stands for the architecture type. + +
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
MDM [59]38.694.013.7772.675.190.70
MDM [59] + C23.974.024.1389.804.870.39
FLAME [31]28.934.024.4665.465.780.89
Ours22.344.064.5291.384.000.50
+ +Table 10. Comparison with MDM and FLAME. C denotes the frame-wise conditions. + +[31] by incorporating frame-wise conditions. Tab. 10 demonstrates the significance of the frame-wise conditions. + +More analyses and ablation studies. More detailed analyses and ablation studies of our design choices are provided in the supplementary material. + +# 4.5. Limitations + +Although our method can generate diverse and natural motions, there are still some limitations. Our method is offline and cannot be applied to interactive scenarios. We assume that the objects are static and cannot handle moving objects. The diffusion models require a long inference time. It takes 7.13 seconds on average for a 720-frame sequence on a Titan Xp GPU. The slow speed might be solved by methods that could accelerate diffusion models [43, 2]. + +# 5. Conclusion + +In this work, we propose a novel hierarchical pipeline for motion synthesis of human-object interactions. Our approach first generates the goal pose and then predicts a set of milestones. Next, we synthesize motions with the guide of milestones. Furthermore, we apply DDPM in our hierarchical pipeline. We also show that our framework could generate more diverse and natural human-object interaction motions than other methods. + +# Acknowledgements + +We thank Jintao Lu, Zhi Cen, Zizhang Li, and Kechun Xu for the valuable discussions. This work was partially supported by the Key Research Project of Zhejiang Lab (No. K2022PG1BB01), NSFC (No. 62172364), and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2 +[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 9 +[3] German Barquero, Sergio Escalera, and Cristina Palmero. BeLFusion: Latent Diffusion for Behavior-Driven Human Motion Prediction. arXiv e-prints, page arXiv:2211.14304, Nov. 2022. 3 +[4] Emad Barsoum, John R. Kender, and Zicheng Liu. Hpgan: Probabilistic 3d human motion prediction via gan. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1499-149909, 2018. 2 +[5] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, Ding Liu, Jing Liu, and Nadia Magnenat Thalmann. Learning progressive joint propagation for human motion prediction. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, page 226-242, Berlin, Heidelberg, 2020. Springer-Verlag. 2 +[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, page 387-404, Berlin, Heidelberg, 2020. Springer-Verlag. 2, 5 +[7] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 5887-5895, 2021. 2 +[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your Commands via Motion Diffusion in Latent Space. arXiv e-prints, page arXiv:2212.04048, Dec. 2022. 3 +[9] Simon Clavet. Motion matching and the road to next-gen animation. In GDC, 2016. 2 +[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6518-6526, 2020. 2 +[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A Framework for Denoising-Diffusion-based Motion Synthesis. arXiv eprints, page arXiv:2212.04495, Dec. 2022. 3 +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, + +volume 34, pages 8780-8794. Curran Associates, Inc., 2021. 2 +[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, June 2021. 4 +[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 2 +[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396-1406, 2021. 3 +[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, oct 2020. 2 +[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, June 2022. 4 +[18] Peter E. Hart, Nils J. Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. 2, 6, 7, 8, 9 +[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11374-11384, October 2021. 1, 2, 3, 5, 6, 7, 8 +[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, June 2021. 3 +[21] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH '23, New York, NY, USA, 2023. Association for Computing Machinery. 2 +[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 1 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3, 4 +[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. arXiv e-prints, page arXiv:2207.12598, July 2022. 2 + +[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. arXiv e-prints, page arXiv:2204.03458, Apr. 2022. 3 +[26] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4), jul 2017. 1, 2 +[27] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based Generation, Optimization, and Planning in 3D Scenes. arXiv e-prints, page arXiv:2301.06015, Jan. 2023. 3 +[28] Ashesh Jain, Amir R. Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5308-5317, 2016. 2 +[29] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Trans. Graph., 41(3), jun 2022. 3 +[30] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927, 2020. 9 +[31] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. FLAME: Freeform Language-based Motion Synthesis & Editing. arXiv e-prints, page arXiv:2209.00349, Sept. 2022. 2, 3, 4, 9 +[32] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, page arXiv:1412.6980, Dec. 2014. 6 +[33] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, page arXiv:1312.6114, Dec. 2013. 2, 3 +[34] Jogendra Nath Kundu, Maharshi Gor, and R. Venkatesh Babu. Bihmp-gan: Bidirectional 3d human motion prediction gan. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'19/IAAI'19/EAAI'19. AAAI Press, 2019. 2 +[35] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2 +[36] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), aug 2020. 6 +[37] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3 +[38] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained Diffusion Models for Unified Human Motion Synthesis. arXiv e-prints, page arXiv:2212.02837, Dec. 2022. 3 +[39] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Multi-level motion attention for human motion prediction. International Journal of Computer Vision, 129(9):2513-2535, 2021. 2 + +[40] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2 +[41] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, June 2022. 3 +[42] Tomohiko Mukai and Shigeru Kuriyama. Geostatistical motion interpolation. ACM Trans. Graph., 24(3):1062-1070, jul 2005. 2 +[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8162-8171. PMLR, 18-24 Jul 2021. 9 +[44] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference (BMVC)*, 2018. 2 +[45] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 9 +[46] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1 +[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. arXiv e-prints, page arXiv:2209.14988, Sept. 2022. 3 +[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, page arXiv:2204.06125, Apr. 2022. 2 +[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3, 4 +[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 2, 3 +[51] C. Rose, M.F. Cohen, and B. Bodenheimer. Verbs and adverbs: multidimensional motion interpolation. IEEE Computer Graphics and Applications, 18(5):32-40, 1998. 2 +[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv e-prints, page arXiv:2209.14792, Sept. 2022. 3 +[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David + +Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 4 +[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc. 2 +[55] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 41(4), jul 2022. 2 +[56] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6), nov 2019. 1, 2, 4, 5, 6, 7 +[57] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), aug 2020. 2 +[58] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 2, 3 +[59] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 3, 4, 9 +[60] Jonathan Tseng, Rodrigo Castellon, and C. Karen Liu. EDGE: Editable Dance Generation From Music. arXiv eprints, page arXiv:2211.10658, Nov. 2022. 3 +[61] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2, 4 +[63] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20460-20469, June 2022. 1, 2, 3, 5, 6, 7, 8, 9 +[64] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9401-9411, June 2021. 2, 5, 6, 7, 9 +[65] J. Wang, S. Yan, B. Dai, and D. Lin. Scene-aware generative network for human motion synthesis. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition + +(CVPR), pages 12201-12210, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 2 +[66] Mao Wei, Liu Miaomiao, and Salzemann Mathieu. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2 +[67] Mao Wei, Liu Miaomiao, Salzemann Mathieu, and Li Hongdong. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2 +[68] D.J. Wiley and J.K. Hahn. Interpolation synthesis of articulated figure motion. IEEE Computer Graphics and Applications, 17(6):39-45, 1997. 2 +[69] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3 +[70] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to generate multimodal human dynamics. In European Conference on Computer Vision, pages 276–293. Springer, 2018. 2 +[71] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6 +[72] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 37(4), jul 2018. 2, 6, 7 +[73] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 40(4), jul 2021. 2 +[74] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 4 +[75] S. Zhang, Y. Zhang, Q. Ma, M. J. Black, and S. Tang. Place: Proximity learning of articulation and contact in 3d environments. In 2020 International Conference on 3D Vision (3DV), pages 642-651, Los Alamitos, CA, USA, nov 2020. IEEE Computer Society. 3, 6 +[76] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 1, 2, 4, 5, 6, 7, 8 +[77] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 6 +[78] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3d people in scenes without people. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 6 +[79] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 2 + +[80] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 2 +[81] Yi Zhou, Zimo Li, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent networks for extended complex human motion synthesis. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 2 +[82] Zixiang Zhou and Baoyuan Wang. UDE: A Unified Driving Engine for Human Motion Generation. arXiv e-prints, page arXiv:2211.16016, Nov. 2022. 3 \ No newline at end of file diff --git a/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/images.zip b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..151da9cfdd94da0137b58dfdd74dc7675d242f6a --- /dev/null +++ b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e744d0b827097a927cb70d988a1c7cffb6388d27dacece7073e224e68f3fc0c +size 658551 diff --git a/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/layout.json b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..01ac3d4a2c8506ddf962fc89559eab8f4eeb616c --- /dev/null +++ b/2023/Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models/layout.json @@ -0,0 +1,12431 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 470, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 470, + 138 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 470, + 138 + ], + "type": "text", + "content": "Hierarchical Generation of Human-Object Interactions with Diffusion Probabilistic Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "spans": [ + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "text", + "content": "Huaijin Pi" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "text", + "content": ", Sida Peng" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "text", + "content": ", Minghui Yang" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "text", + "content": ", Xiaowei Zhou" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "text", + "content": ", Hujun Bao" + }, + { + "bbox": [ + 121, + 160, + 475, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 178, + 457, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 178, + 457, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 178, + 457, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 178, + 457, + 194 + ], + "type": "text", + "content": " State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 136, + 178, + 457, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 178, + 457, + 194 + ], + "type": "text", + "content": " Ant Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 220, + 192, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 220, + 192, + 234 + ], + "spans": [ + { + "bbox": [ + 143, + 220, + 192, + 234 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 247, + 290, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 290, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 290, + 429 + ], + "type": "text", + "content": "This paper presents a novel approach to generating the 3D motion of a human interacting with a target object, with a focus on solving the challenge of synthesizing long-range and diverse motions, which could not be fulfilled by existing auto-regressive models or path planning-based methods. We propose a hierarchical generation framework to solve this challenge. Specifically, our framework first generates a set of milestones and then synthesizes the motion along them. Therefore, the long-range motion generation could be reduced to synthesizing several short motion sequences guided by milestones. The experiments on the NSM, COUCH, and SAMP datasets show that our approach outperforms previous methods by a large margin in both quality and diversity. The source code is available on our project page https://zju3dv.github.io/hghoi." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 453, + 128, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 128, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 128, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 475, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 534 + ], + "type": "text", + "content": "Scene-aware motion generation [19] aims to synthesize 3D human motion given a 3D scene model to enable virtual humans to naturally wander around scenes and interact with objects, which has a variety of applications in AR/VR, filmmaking, and video games." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 535, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 535, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 535, + 288, + 693 + ], + "type": "text", + "content": "Unlike traditional motion generation methods for character control which aim to generate short or repeated motion on the fly guided by a user's control signals [56], we focus on the setting of generating long-term human-object interactions [56, 19, 76] given a starting position of the human and a target object model. This setting brings in new challenges. First, the entire approaching process and the human-object interaction should be coherent, which requires the capability of modeling long-range interaction between the human and the object. Second, in the context of content generation, the generative model should be able to synthesize diverse motions as there are many plausible ways for a real human to approach and interact with the target object." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 222, + 542, + 343 + ], + "blocks": [ + { + "bbox": [ + 310, + 222, + 542, + 343 + ], + "lines": [ + { + "bbox": [ + 310, + 222, + 542, + 343 + ], + "spans": [ + { + "bbox": [ + 310, + 222, + 542, + 343 + ], + "type": "image", + "image_path": "2f698d33418202babdfca7f4b06968b8d90235acb490c959682dc07f3ffdd37e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 343, + 542, + 462 + ], + "blocks": [ + { + "bbox": [ + 310, + 343, + 542, + 462 + ], + "lines": [ + { + "bbox": [ + 310, + 343, + 542, + 462 + ], + "spans": [ + { + "bbox": [ + 310, + 343, + 542, + 462 + ], + "type": "image", + "image_path": "9b4ada4e7b0ca0ea6c34e2f2fa1ba622b75b166f000e246c077885f84368a241.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 467, + 547, + 544 + ], + "lines": [ + { + "bbox": [ + 305, + 467, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 547, + 544 + ], + "type": "text", + "content": "Figure 1. Genration of human-object interactions. Given an object, our method first predicts a set of milestones, where the rings indicate the positions and the humans with pink clothes represent the local poses. Then the motions are inflated between milestones. This figure shows that our method can generate diverse milestones and motions with the same object. The flow of time is shown with a color code where darker blue denotes the later frame." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Existing methods for motion synthesis can be roughly characterized into online generation and offline generation. Most online methods [56, 19, 76] focus on real-time control of characters. Given a target object, they generally use auto-regressive models to recurrently generate future motions by feeding their predictions. Although they have been widely used for interactive scenarios like video games, their motion quality is not satisfactory enough for long-term generation [63]. A plausible cause is the error accumulation in the auto-regressive process, where the errors in previous predictions are fed back as the model input, as discussed in [40, 26, 45, 46, 22]. To improve the motion quality," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15061" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "some recent offline methods [6, 65, 64, 63] employ a multistage framework, which first generates the trajectory and then synthesizes motions. TDNS [63] generates paths by combining the cVAE model [33] and deterministic planning methods like A* [18]. Although this strategy can produce reasonable paths, the path diversity is limited, as demonstrated by our experimental results in Sec. 4.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 360 + ], + "type": "text", + "content": "In this paper, we propose a novel offline approach for synthesizing long-term and diverse human-object interactions. Our innovation lies in a hierarchical generation strategy, which first predicts a set of milestones and then generates human motions between milestones. Fig. 1 illustrates the basic idea. Specifically, given the starting position and the target object, we design a milestone generation module to synthesize a set of milestones along the motion trajectory, each of which encodes the local pose and indicates the transition point during the human movement. Based on these milestones, we employ a motion generation module to produce the full motion sequence. Thanks to the milestones, we simplify the long-sequence generation into synthesizing several short motion sequences. Furthermore, the local pose at each milestone is generated by a transformer that considers the global dependency [62], leading to temporally consistent results, which further contribute to coherent motions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 529 + ], + "type": "text", + "content": "In addition to our hierarchical generation framework, we further exploit diffusion models [53, 23, 54] to synthesize human-object interactions. Previous diffusion models for motion synthesis [31, 59] combine transformer [62] and Denoising Diffusion Probabilistic Model (DDPM) [23]. Directly applying them to our setting is prohibitively comput-intensive due to the long motion sequences and may lead to the GPU memory explosion [50]. Because our hierarchical generation framework converts the long-term generation to the synthesis of several short sequences, the required GPU memory is reduced to the same level of short-term motion generation. Therefore, we can efficiently leverage the transformer DDPM to synthesize long-term motion sequences, which improves the generation quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": "We validate our design choices on the NSM [56], COUCH [76], and SAMP [19] datasets with extensive experiments. On these datasets, our hierarchical framework outperforms previous methods significantly in both motion quality and diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 602, + 134, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 134, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 134, + 615 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 149, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 149, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 149, + 635 + ], + "type": "text", + "content": "2.1. Motion synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "Motion synthesis is a long-standing problem in computer graphics and vision [68, 51, 42, 9]. With the rapid development of deep learning, recent works have applied neural networks to motion synthesis [14, 40, 26, 72]. Some methods are deterministic [28, 81, 44, 67] while others try to predict stochastic motions by VAE [70, 1, 71, 79, 45] and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 396 + ], + "type": "text", + "content": "GAN [4, 34]. To further improve the performance, some works apply GCN [67, 10, 66, 35, 39] or transformers [5] to extract the features from the human skeleton. To handle the ambiguity of human motion, some works [26, 55] propose to employ phase signals to guide the motions. Recent works start to consider the scene context [56, 6]. NSM [56] is the first work that aims at synthesizing human motions with object-level interactions with specific action labels. Based on NSM, SAMP [19] applies cVAE to predict diverse motions. These works [56, 19, 76] mainly focus on the interaction with one or two objects while others [6, 65] generate motions with a full scene (including the information about floors and walls) as the input. [6] predicts future motions of the given motion sequence by first predicting the trajectory and then generating motions based on a 2D image of the scene. Similar to this pipeline, [65] applies GAN [16]. Furthermore, [64] proposes a framework that first places pose on a human-provided path and then synthesizes motions in a 3D scene with a full-scene scan. [63] combines " + }, + { + "bbox": [ + 304, + 72, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^*" + }, + { + "bbox": [ + 304, + 72, + 545, + 396 + ], + "type": "text", + "content": " and cVAE to generate diverse human motions. To further control human motion, [80] employs the gaze, and COUCH [76] explicitly models the hand contacts to guide the prediction. Some methods [7, 21] also enable physically simulated characters to perform scene interaction tasks including sitting [7, 21] and carrying boxes [21]. Some works also focus on grasp [58, 69], manipulation [73] and the interaction with dynamic objects [57]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 579 + ], + "type": "text", + "content": "Ours vs. others. This work follows the setting of [56, 19] and focuses on object-level interaction. In contrast to [56, 19] which generate motions mainly based on auto-regressive models, we design a hierarchical framework to synthesize motions. Different from [6, 64], our work focuses on much more long-term generation (longer than 10 seconds) while [6] is 2 seconds and [64] is 6 seconds. Instead of planning the path auto-regressively with an extra network to generate diverse trajectories like [63], our method directly predicts a set of milestones to describe the approaching process which is inherently diverse. In addition, most methods [19, 76, 64, 63] rely on cVAE to generate stochastic motions while we exploit DDPM [23] to synthesize trajectories and motions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 587, + 408, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 408, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 408, + 598 + ], + "type": "text", + "content": "2.2. Diffusion models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Diffusion models [53] are a class of likelihood-based methods that generate samples by gradually removing noise from a signal. Then, [23, 54] develop the diffusion models for high-quality image generation. To control the generated results, [12] proposes classifier guidance for trading off diversity for fidelity. Later, the classifier-free model [24] achieves better results [48] in text-conditioned image generation. In addition, diffusion models have been successfully applied to other domains like the generation of videos" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15062" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 542, + 162 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 162 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 162 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 162 + ], + "type": "image", + "image_path": "e94b5e2623b8eb9c14a45cbda549c0033009a05eed1a24688830e82bd0525162.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 164, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 164, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 547, + 198 + ], + "type": "text", + "content": "Figure 2. Overview of our pipeline. Our pipeline consists of three components: First, the goal pose is synthesized given the object. Then, a number of milestones with local poses are predicted based on the goal pose. Finally, the trajectory and the full motion sequences are inflated between the milestones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 216, + 170, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 170, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 170, + 227 + ], + "type": "text", + "content": "[52, 25] and 3D contents [47]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 227, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 323 + ], + "type": "text", + "content": "There are some works [74, 59, 31, 38] that apply DDPM to synthesize motions. [3, 8] also explore latent diffusion models [50] for motion generation. However, most works focus on text-conditioned [74, 11], audio-driven [82] and music-driven [60, 82] motion generation while we target human-object interaction. In this work, we apply transformer DDPM [23] to a multi-stage framework, which separately generates trajectories and synthesizes motions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 323, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 407 + ], + "type": "text", + "content": "There are some concurrent works [27] that apply DDPM to synthesize motions in a scene. However, [27] is designed for short-term motion generation (around 2 seconds) while we target long-term human motion generation (longer than 10 seconds). Different from existing work [59, 27], we employ DDPM in a multi-stage framework, where trajectories and motions are separately predicted." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 417, + 107, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 107, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 107, + 430 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": "Given the object " + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": " and the starting point " + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": ", our goal is to synthesize 3D human motions " + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{r}_i, \\pmb{\\theta}_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": " with human-object interactions. " + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": " is the root trajectory, and the " + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_i" + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": " indicates local pose at " + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": "-th frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "content": "We design a hierarchical motion generation framework as shown in Fig. 2. First, we employ GoalNet[19] to predict an interaction goal on the object. Then, we generate the goal pose (Sec. 3.1) to explicitly model the human-object interaction. Next, our milestone generation module (Sec. 3.2) estimates the length of the milestones, produces the trajectory of the milestones from the starting point to the goal, and places milestone poses. Therefore, the long-range motion generation is decomposed into combinations of short-range motion synthesis. Finally, we design a motion generation module (Sec. 3.3) to synthesize the trajectory between the milestones and infill the motions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 168, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 168, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 168, + 647 + ], + "type": "text", + "content": "3.1. Goal pose generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "We call the pose in which a person interacts with an object and remains stationary a goal pose. To synthesize diverse human motion, we first generate a goal pose interacting with the object following [58, 69, 63]. Most methods [78, 75, 20] generate human poses using the cVAE model." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 217, + 542, + 366 + ], + "blocks": [ + { + "bbox": [ + 310, + 217, + 542, + 366 + ], + "lines": [ + { + "bbox": [ + 310, + 217, + 542, + 366 + ], + "spans": [ + { + "bbox": [ + 310, + 217, + 542, + 366 + ], + "type": "image", + "image_path": "dd7b467a6441efe6754b756c098c3fe2730a3f5db2ca96632998109e8c5fe610.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 369, + 545, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 414 + ], + "type": "text", + "content": "Figure 3. Part VQ-VAE. Part VQ-VAE first splits the skeleton into multiple parts and learns the codebooks separately. The composition of different parts is subsequently modeled with the autoregressive prediction model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 609 + ], + "type": "text", + "content": "They project the poses to the standard normal distribution in the continuous space [33]. Empirically, we find that cVAE models do not perform well in our setting. To overcome this challenge, we introduce VQ-VAE [61, 49] to model the data distribution which exploits a discrete representation to cluster the data in a finite set of points [37]. We hypothesize that limited data of goal pose from the SAMP dataset [19] can always be clustered by VQ-VAE but may not be enough for learning a continuous latent space for VAE [41]. In addition, based on the observation that different human poses may share similar properties [15, 29] (e.g., humans may sit with different hand positions but the same leg positions), we split joints into " + }, + { + "bbox": [ + 304, + 441, + 547, + 609 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 441, + 547, + 609 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 441, + 547, + 609 + ], + "type": "inline_equation", + "content": "L = 5" + }, + { + "bbox": [ + 304, + 441, + 547, + 609 + ], + "type": "text", + "content": ") different non-overlapping groups like MotionPuzzle [29]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Quantization. As shown in Fig. 3, the goal pose " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " is split into separate joint groups as " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{g} = \\{\\theta_{gi}\\}_{i=1}^{L}" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": ". Then a discrete codebook " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " with a list of vectors compares the output " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_i" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " from the encoder " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " to find the closest vector in Euclidean distance. The " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " vectors with minimal distances will be concatenated and fed into a shared decoder " + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15063" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 237, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 237, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 237, + 84 + ], + "type": "text", + "content": "reconstruct " + }, + { + "bbox": [ + 47, + 72, + 237, + 84 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 47, + 72, + 237, + 84 + ], + "type": "text", + "content": ". The loss function is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 94, + 286, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 94, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 58, + 94, + 286, + 128 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\boldsymbol {\\theta} _ {g}, D (\\mathbf {z})\\right) = \\| \\boldsymbol {\\theta} _ {g} - D (\\mathbf {z}) \\| _ {2} ^ {2} + \\sum_ {i = 1} ^ {L} \\| s g \\left[ E _ {i} \\left(\\boldsymbol {\\theta} _ {g i}\\right) \\right] - \\mathbf {z} _ {i} \\| _ {2} ^ {2} \\tag {1}", + "image_path": "797c9e4f2c9754dc325088a4998d79b542afd58d1919d4cbc2afe3815eaa1d3b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 161, + 128, + 266, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 128, + 266, + 152 + ], + "spans": [ + { + "bbox": [ + 161, + 128, + 266, + 152 + ], + "type": "interline_equation", + "content": "+ \\beta \\sum_ {i = 1} ^ {L} \\| s g [ \\mathbf {z} _ {i} ] - E _ {i} (\\boldsymbol {\\theta} _ {g i}) \\| _ {2} ^ {2},", + "image_path": "7800cbb2c31cf65017504b9e802ed7a25e461340c5a27a1e68c2aa7bedcf7024.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 160, + 74, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 160, + 74, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 160, + 74, + 171 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 182, + 286, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 182, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 117, + 182, + 286, + 193 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\left[ \\mathbf {z} _ {1}, \\dots , \\mathbf {z} _ {L} \\right], \\tag {2}", + "image_path": "f7d3c8326a5a3b96f9a17a438df445e38687f8893bbc503470cf1b18dbf917b4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 196, + 286, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 196, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 117, + 196, + 286, + 217 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} = \\underset {\\mathbf {z} _ {i} ^ {\\prime} \\in Z _ {i}} {\\arg \\min } \\| \\mathbf {z} _ {i} ^ {\\prime} - \\hat {\\mathbf {z}} _ {i} \\|, \\tag {3}", + "image_path": "e224d5bee56a5f0b0f223c17dcab1838d2d0322d576453986dcffcb425e40a26.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 221, + 286, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 221, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 117, + 221, + 286, + 233 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {i} = E \\left(\\boldsymbol {\\theta} _ {g i}\\right). \\tag {4}", + "image_path": "2a48c6898d5215d2c117f4031c1e053af7ba8b61ddff2217dec199ce9f2bb4c3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "text", + "content": "Here, the term " + }, + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "inline_equation", + "content": "sg[\\cdot]" + }, + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "text", + "content": " denotes the stop-gradient operator, and " + }, + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "inline_equation", + "content": "\\| sg[\\mathbf{z}_i] - E_i(\\pmb{\\theta}_{gi})\\|_2^2" + }, + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "text", + "content": " is the commitment loss [49] controlled by a weighting factor " + }, + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 241, + 286, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": "Generation. With " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " available, we can represent " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{g} = \\{\\pmb{\\theta}_{g1},\\dots,\\pmb{\\theta}_{gL}\\}" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " by a sequence of the part-based codebook indices. To be more specific, we use " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " to extract the feature from " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{gi}" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " and find the closest vector " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\in \\mathbf{Z}_i" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": ". Then we use " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "s_i\\in \\{0,\\ldots ,|\\mathbf{Z}_i| - 1\\}" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " to indicate the index of " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{g}" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": " can be represented by " + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = \\{s_1,\\dots,s_L\\}" + }, + { + "bbox": [ + 47, + 291, + 286, + 363 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "text", + "content": "To generate a natural goal pose, we convert the problem to predict a sequence of indices that can represent " + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "text", + "content": ". We formulate the inference as a conditional auto-regressive process and employ a transformer [62] to learn to predict the distribution of possible indices [13]. The condition contains the environment around the goal " + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_g" + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "text", + "content": ", and the action " + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_g" + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "text", + "content": ". Following NSM [56] and COUCH [76], a cylindrical volume of a pre-defined radius and height is created around the goal. Within this volume, spheres are uniformly sampled and the occupancies corresponding to the object of these spheres are calculated. Then, these occupancies are flattened to form a feature vector denoted as " + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_g" + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_g" + }, + { + "bbox": [ + 46, + 364, + 286, + 530 + ], + "type": "text", + "content": " is a vector indicating the action type. These variables are fed into the transformer as tokens. Our target is to learn the likelihood of the sequence:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 540, + 286, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 540, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 85, + 540, + 286, + 564 + ], + "type": "interline_equation", + "content": "p (\\mathbf {s} \\mid \\mathbf {O} _ {g}, \\mathbf {a} _ {g}) = \\prod_ {i} p \\left(s _ {i} \\mid s _ {< i}, \\mathbf {O} _ {g}, \\mathbf {a} _ {g}\\right). \\tag {5}", + "image_path": "551db2a9cf9f67873858e94df33c7d29c30766741a491fd3032cd2a5a3562bd0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "text", + "content": "After predicting the indices, we map them back to their corresponding codebook entries to get the quantized features " + }, + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = [\\mathbf{z}_1,\\dots ,\\mathbf{z}_L]" + }, + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "text", + "content": ", which are fed into the decoder " + }, + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "text", + "content": " to generate the goal pose " + }, + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 47, + 568, + 286, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 623, + 167, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 167, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 167, + 635 + ], + "type": "text", + "content": "3.2. Milestone generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "Based on the starting pose and the goal pose, we can generate the milestone trajectory and synthesize the local poses at the milestones. Following [74, 59], we build a transformer DDPM [23] and apply it to generate the milestones for better quality. Because the length of motion data is unknown and can be arbitrary (e.g., the human could quickly" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": "walk towards the chair and sit down or sit after walking around the chair slowly), we predict the length of milestones, denoted by " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": ". Then we synthesize " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": " milestone points and place the local poses on these points." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "text", + "content": "Transformer DDPM. Here we first briefly introduce DDPM [23], which is learned to reverse a diffusion process. Formally, the diffusion model [53] is defined as a latent variable model of the form " + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0) \\coloneqq \\int p_{\\theta}(\\mathbf{x}_{0:T}) d\\mathbf{x}_{1:T}" + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "text", + "content": " is the data and " + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1, \\ldots, \\mathbf{x}_T" + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "text", + "content": " are the latents. " + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{0:T})" + }, + { + "bbox": [ + 306, + 137, + 545, + 209 + ], + "type": "text", + "content": " is formulated as a Markov chain as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 220, + 545, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 347, + 220, + 545, + 251 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right), \\tag {6}", + "image_path": "25942e0bd0d4bf8f0bea79101ed99a24fbe382a79c5f91e66b5c6050ad793374.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 277, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 545, + 289 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\boldsymbol {\\mu} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\boldsymbol {\\Sigma} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right). \\tag {7}", + "image_path": "4a0bf3b9afa69f10a4be559383b2a799dcbc0acc09129106a77001a76fe56f57.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 299, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 545, + 335 + ], + "type": "text", + "content": "Diffusion models approximate posterior " + }, + { + "bbox": [ + 306, + 299, + 545, + 335 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 306, + 299, + 545, + 335 + ], + "type": "text", + "content": " as a Markov chain that gradually adds Gaussian noise to the data with variance schedules given by " + }, + { + "bbox": [ + 306, + 299, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 306, + 299, + 545, + 335 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 357, + 346, + 544, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 346, + 544, + 376 + ], + "spans": [ + { + "bbox": [ + 357, + 346, + 544, + 376 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) := \\prod_ {t = 1} ^ {T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right), \\tag {8}", + "image_path": "0dff56a6630ed205cb0e0e4d7d2c2624914a9702b2613e237cf68a42647fb545.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 334, + 390, + 544, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 390, + 544, + 409 + ], + "spans": [ + { + "bbox": [ + 334, + 390, + 544, + 409 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) := \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {9}", + "image_path": "c62bfc015a5f7ee70771b99451894eed9faa13fd34cf17529a2971172b948ef5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 415, + 544, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 544, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 544, + 438 + ], + "type": "text", + "content": "In contrast to adding noises on " + }, + { + "bbox": [ + 306, + 415, + 544, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 306, + 415, + 544, + 438 + ], + "type": "text", + "content": " sequentially, DDPM formulates the diffusion process as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 342, + 449, + 544, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 449, + 544, + 461 + ], + "spans": [ + { + "bbox": [ + 342, + 449, + 544, + 461 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {10}", + "image_path": "bc535fcaabdf9543e406af1ee084d99cc5e444223626a321478e7b4d7c2e8aac.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{s = 1}^{t}\\alpha_{s}" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": ". Hence, we can generate " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": " by sampling a noise " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": " as the training data. DDPM employs a neural network to model " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": " and the inference is to gradually denoise " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 306, + 473, + 544, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": "Like existing works [74, 31] that apply DDPM in the motion domain, we employ a transformer decoder [62] as our architecture of DDPM. The transformer takes the noise " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " and the condition " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " as input. The condition " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " means variables related to the generation, which will be described in detail in each subsection. The diffusion time-step " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " is in the sinusoidal position embeddings form [62] and is injected into each block in the transformer. Different from existing works [59, 74] that assume the length of the motion sequences is already given, we insert a parallel branch to estimate the length of milestones by taking the length token " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{tok}" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " and the condition " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " as the input, as shown in Fig. 4. The length prediction head is an MLP and predicts a multinomial distribution over discrete length indices " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{1, 2, \\dots, N_{max}\\}" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " like [17], where " + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{max}" + }, + { + "bbox": [ + 306, + 534, + 545, + 713 + ], + "type": "text", + "content": " represents the" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15064" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 271 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 271 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 271 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 271 + ], + "type": "image", + "image_path": "efd63b3a288c09bc8835f577071820eacdaea1dc20e574f5c86a51b6d80ebdba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": "Figure 4. Overview of transformer DDPM for milestone generation. The model first takes the length token " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{tok}" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": " and the condition " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": " as the input to predict the data length. Then it constructs the noise sequence " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{1:N}" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": ". In the diffusion process, it is fed with " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": " and the sequence " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{1:N}" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": " at time-step " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": " to predict the target " + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^{1:N}" + }, + { + "bbox": [ + 46, + 279, + 289, + 357 + ], + "type": "text", + "content": ". For other sub-modules, we remove the length prediction head." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "content": "length of the longest sequences for training. We use cross-entropy loss as the loss function. At inference time, we sample the length " + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "content": " from the estimated distribution. Next, we construct " + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "content": " milestones as the input to the transformer. We predict the denoised data " + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "content": " based on the condition " + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "content": " by the DDPM " + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "content": ". This is formulated as " + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = f(\\mathbf{x}_t,t,\\mathbf{C})" + }, + { + "bbox": [ + 46, + 374, + 289, + 460 + ], + "type": "text", + "content": ", and the training loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 466, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\operatorname {E} _ {t \\in [ 1, T ], \\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0}), \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\mathbf {x} _ {0} - \\hat {\\mathbf {x}} _ {0} \\| \\right]. \\tag {11}", + "image_path": "ccab5854f516a949595fa291b60295050050de2a1cb03a651c50cbcfb6ec39c5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 501, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 550 + ], + "type": "text", + "content": "In our setting, we use a two-step strategy like [64, 63], where two transformer DDPMs are applied to first generate the milestone points and then synthesize the local pose at each milestone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "text", + "content": "Generation of milestone points. The milestone points are conditioned on the object, the information of the starting point, and the goal. We then define the condition as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 608, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 608, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {m}} = \\left\\{\\mathbf {I} _ {s}, \\mathbf {I} _ {g}, \\mathbf {O} _ {s}, \\mathbf {O} _ {g}, \\mathbf {g}, \\mathbf {s} \\right\\}, \\tag {12}", + "image_path": "6b8e0ed8e977f90e1680dc1db12588e97067f4a4053a2886456f7204216ef0b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_g" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " are the object representation relative to the starting point and the goal. Similar to other methods [56, 19, 76], the object representation is modeled by an " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "8\\times 8\\times 8" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " grid with their positions and occupancies. Following NSM [56] and COUCH [76], we also explicitly model the occupancies around the starting point and the goal with " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_s" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_g" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " (the same form as the occupancy feature described in Sec. 3.1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "The information of goal " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{r}_g, \\mathbf{a}_g, \\boldsymbol{\\theta}_g\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_g" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " means the goal position and orientation in the starting point coordinate system, " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_g" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the target action label at the goal, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}_g" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the goal pose. " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{s} = \\{\\mathbf{a}_s, \\boldsymbol{\\theta}_s\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " represents the action labels " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_s" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and the pose " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}_s" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " at the starting point." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": "The target is to predict the milestone " + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{m}_1, \\dots, \\mathbf{m}_N\\}" + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": ". Following NSM [56], a bi-directional scheme is employed for the milestone points generation where we predict the roots of milestones in both the starting point coordinate system and the goal coordinate system. The final predicted roots are blended from these two kinds of outputs. The milestone point " + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 385, + 224, + 545, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 224, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 385, + 224, + 545, + 237 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i} = \\left\\{\\mathbf {r} _ {i} ^ {b}, \\mathbf {c} _ {i}, \\mathbf {w} _ {i} \\right\\}. \\tag {13}", + "image_path": "d8b536cc2df3ee53eb012f1e1ade61b386ddd86c12dd8ca5dd84d36cd464ecef.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "text", + "content": "The representation is similar to previous work [19, 56, 76], where " + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i^b" + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "text", + "content": " indicates the root position and forward directions relative to the starting point and the goal, " + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "text", + "content": " is a label vector indicating the contact between the environment and the body, and we use a high-dimensional feature vector " + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "text", + "content": " to encode the character state at the milestone following [56, 19, 76]. The details of these variables are provided in the supplementary material. We use transformer DDPM " + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{m}}" + }, + { + "bbox": [ + 304, + 246, + 546, + 354 + ], + "type": "text", + "content": " to predict the length and synthesize milestone points." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": "Generation of milestone poses. Previous works [64, 63] separately place the poses on the sparse points along the path and infill the motions between, which may lead to unnatural transitions between these poses. On the contrary, we generate the local pose at each milestone with transformer DDPM to build the temporal dependency. With the help of the generated milestone points, we can access the accurate spatial relationship " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": " between the object and the character and the ego-centric environment occupancies " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": " [76] at the " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": "-th milestone. The milestone poses also depend on the milestone state including " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i, \\mathbf{w}_i\\}" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": "-th milestone. Consequently, we define the condition " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{C_k}" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": " as a combination with starting pose " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": ", goal pose " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\theta_g" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": ", and frame-wise condition " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "\\gamma_i" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 368, + 546, + 536 + ], + "type": "text", + "content": "-th milestone as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 544, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 544, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 362, + 544, + 545, + 557 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {k}} = \\left\\{\\boldsymbol {\\theta} _ {s}, \\boldsymbol {\\theta} _ {g}, \\gamma_ {1}, \\dots , \\gamma_ {N} \\right\\}, \\tag {14}", + "image_path": "d28c9d012619959448a94004f950411be5f2dfae58c9682a159968749063f319.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 559, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 559, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 362, + 559, + 545, + 572 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad \\gamma_ {i} = \\left\\{\\mathbf {I} _ {i}, \\mathbf {O} _ {i}, \\mathbf {c} _ {i}, \\mathbf {w} _ {i} \\right\\}. \\tag {15}", + "image_path": "c15705da20be6e1a47fe749ece92a83915d1a405bed55e896ce8676f71f5ac6d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "type": "text", + "content": "The local poses at the milestones are predicted by the transformer DDPM " + }, + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{k}}" + }, + { + "bbox": [ + 304, + 580, + 545, + 605 + ], + "type": "text", + "content": " without the length prediction head." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 611, + 416, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 416, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 416, + 624 + ], + "type": "text", + "content": "3.3. Motion generation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Instead of predicting motions frame-by-frame [56, 19, 76], our approach hierarchically synthesizes the full sequence based on the generated milestones. We follow [6, 64, 63] to first generate the trajectory and then synthesize the motions. Specifically, within two consecutive milestones, we first complete the trajectory. Then, the motions are inflated with the guidance of successive milestone" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15065" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "poses. The two steps are accomplished using two transformer DDPMs (described in Sec. 3.2), respectively. For each step, we carefully design the condition of DDPM to generate the target output." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 182 + ], + "type": "text", + "content": "Trajectory completion. For the trajectory completion between the milestones " + }, + { + "bbox": [ + 46, + 133, + 287, + 182 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 46, + 133, + 287, + 182 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 287, + 182 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{i + 1}" + }, + { + "bbox": [ + 46, + 133, + 287, + 182 + ], + "type": "text", + "content": ", we assume that it is only conditioned on the milestones and the object. Thus we define the condition as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 187, + 287, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 187, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 67, + 187, + 287, + 203 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {r}} = \\left\\{\\mathbf {I} _ {i}, \\mathbf {I} _ {i + 1}, \\mathbf {O} _ {i}, \\mathbf {O} _ {i + 1}, \\mathbf {m} _ {i}, \\mathbf {m} _ {i + 1}, \\mathbf {t} _ {i} ^ {i + 1} \\right\\}, \\tag {16}", + "image_path": "c2221511ae19db0a19251883a0931e06bc780c5459626da088b630362a2fb47e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": " indicates the object representation relative to the milestone " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": " denotes the ego-centric occupancies (the same form as the occupancy feature described in Sec. 3.1) around the " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": "-th milestone like NSM [56] and COUCH [76]. " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": " has been shown in Eq. (13). " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i^{i+1}" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": " represents the position and orientation of the " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "(i+1)" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": "-th milestone in the " + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 208, + 287, + 327 + ], + "type": "text", + "content": "-th milestone's coordinate system. Between two consecutive milestones, we generate the trajectory with a length of 2 seconds, which is 61 frames. Similar hyperparameters can be found in previous methods [64, 63]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "content": "The target output is the trajectory between two consecutive milestones. Similar to the milestone point, we synthesize the trajectory in the " + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "content": "-th frame with a bi-directional scheme. The trajectory is composed of a set of points that have the same representation as the milestone in Eq. (13). We generate the trajectory with a transformer DDPM " + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "content": " similar to the one in Sec. 3.2. Since we assume the trajectory length between two milestones is fixed, the DDPM " + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "content": " does not have the length prediction head." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": "Motion infilling. To synthesize the long-range motion, we convert a long sequence into several fixed-length short sequences with the help of milestone points and milestone poses. For a sub-sequence between the consecutive milestone poses, our goal is to generate the missing local poses over the trajectory. The generated motion has to satisfy the trajectory and naturally transits from a milestone to the next milestone. Like milestone pose generation, we use the same representation of frame-wise condition in Eq. (15). The condition is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 573, + 287, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 573, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 107, + 573, + 287, + 589 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {p}} = \\left\\{\\boldsymbol {\\theta} ^ {1}, \\boldsymbol {\\theta} ^ {6 1}, \\gamma^ {1}, \\dots , \\gamma^ {6 1} \\right\\}, \\tag {17}", + "image_path": "3dc5c0eeebdf9fe8aa945973bcb8a27e2b369da810c8c0f0f469a7177237797c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\theta^1" + }, + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "inline_equation", + "content": "\\theta^{61}" + }, + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "text", + "content": " are the local poses of two consecutive milestones. By taking these inputs, we generate smooth motions using another transformer DDPM " + }, + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{p}}" + }, + { + "bbox": [ + 46, + 593, + 287, + 642 + ], + "type": "text", + "content": " without the length prediction head." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 178, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 178, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 178, + 684 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We train the part VQ-VAE and transformer DDPM models with the Adam optimizer [32]. All the models are trained" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "with a fixed learning rate of 0.0001 with batch size 256. The remaining details are in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 105, + 477, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 105, + 477, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 105, + 477, + 118 + ], + "type": "text", + "content": "4.2. Datasets and evaluation metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 124, + 545, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 124, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 124, + 545, + 196 + ], + "type": "text", + "content": "Test setting. Our experiments are conducted on the SAMP [19], COUCH [76], and NSM [56] datasets. Provided with a starting point, a starting pose, an object, and an endpoint, the virtual human is asked to approach the object, interact with it, and leave to reach the endpoint. The ablation studies are conducted on the SAMP dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 213, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 405 + ], + "type": "text", + "content": "Metrics. Following the previous method [19], we calculate the Fréchet distance (FD) between the generated and ground-truth motions to measure the motion quality. We also conduct user studies and each sequence is evaluated by at least 3 users with scores ranging from 1 to 5. In addition, we calculate the penetration ratio [64, 75, 78] and foot sliding [72, 36] to show the physical plausibility between the 3D object and the synthesized motions. We compute the Average Pairwise Distance (APD) [71, 77] to evaluate the diversity. Specifically, we calculate the APD of synthesized motion, the character's pose during object interactions, and trajectories. Following previous work [56, 19], we calculate PE (positional errors) and RE (rotational errors) to indicate the precision of object interactions. For each test object, we generate multiple sequences. More details are included in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 414, + 480, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 480, + 426 + ], + "type": "text", + "content": "4.3. Comparison with other methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "text", + "content": "Results on the SAMP dataset. On the SAMP dataset [19], we compare our method with online methods SAMP [19] and MoE [72]. As our method is offline, we also implement and modify offline methods SLT [64] and TDNS [63] to our setting. For SLT [64], we employ " + }, + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "text", + "content": " [18] to plan a path and select points along the path as subgoals to form the input for SLT. More details about the implementation of [64, 63] are in the supplementary material. Since MoE often fails to finish the action, we do not calculate the penetration ratio for it. As shown in Tab. 1, our approach outperforms other methods in terms of lower FD, higher user study scores, and higher APD. Furthermore, our method achieves much higher trajectory diversity than SAMP [19]. Although TDNS [63] proposes Neural Mapper (NM) which combines " + }, + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "text", + "content": " [18] and cVAE, the diversity of the generated trajectory is inferior to our method, as indicated by " + }, + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_T" + }, + { + "bbox": [ + 305, + 433, + 545, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": "Results on the COUCH dataset. As our target is to synthesize diverse motions instead of controlling the characters, we only evaluate the motion quality on the COUCH dataset [76]. Tab. 2 shows that our method outperforms all baselines. Our approach achieves much higher " + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_T" + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": " than other methods. We observe that " + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_T" + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": " of TDNS [63] is" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15066" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 128, + 70, + 470, + 145 + ], + "blocks": [ + { + "bbox": [ + 128, + 70, + 470, + 145 + ], + "lines": [ + { + "bbox": [ + 128, + 70, + 470, + 145 + ], + "spans": [ + { + "bbox": [ + 128, + 70, + 470, + 145 + ], + "type": "table", + "html": "
MethodFD ↓User study ↑APDM ↑APDP ↑APDT ↑PE↓RE↓Penetration↓Sliding↓
MoE [72]74.332.763.502.6352.46∞∞-1.68
SAMP [19]57.342.863.633.0563.183.444.126.981.02
SLT* [64]68.832.303.082.6640.131.771.604.281.71
TDNS* [63]46.602.903.683.4066.040.450.395.140.94
Ours22.343.624.064.5291.380.390.324.000.50
", + "image_path": "2895bbde53af9e4aeec1e7c7c1127ad7f37b54d192c3b2664d4dc23c8862be14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 205, + 285, + 273 + ], + "blocks": [ + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "type": "text", + "content": "Table 1. Quantitative results on the SAMP dataset. SLT* and TDNS* mean we modify and implement them on the SAMP dataset. The subscript “M”, “P” and “T” stand for “Motion”, “Pose” and “Trajectory”. PE and RE represent the positional error and rotation error. Sliding denotes foot sliding. " + }, + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 46, + 154, + 547, + 187 + ], + "type": "text", + "content": " means the method failed to reach the goal." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 205, + 285, + 273 + ], + "lines": [ + { + "bbox": [ + 52, + 205, + 285, + 273 + ], + "spans": [ + { + "bbox": [ + 52, + 205, + 285, + 273 + ], + "type": "table", + "html": "
MethodFD↓User study ↑APDM↑APDP↑APDT↑Penetration↓Sliding↓
NSM [56]118.982.990008.200.59
SAMP [19]160.121.940.890.184.694.940.72
COUCH [76]127.193.051.410.6623.487.430.37
SLT* [64]93.463.011.681.172.863.901.85
TDNS* [63]71.723.292.702.1916.985.121.17
Ours56.354.273.222.3064.963.540.55
", + "image_path": "749b2a24fc7ba13f416a2f57496c67348cb2febbcd40117d5cfbec117ee5443d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 319, + 285, + 355 + ], + "blocks": [ + { + "bbox": [ + 47, + 280, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 280, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 287, + 304 + ], + "type": "text", + "content": "Table 2. Quantitative results on the COUCH dataset. SLT* and TDNS* mean we modify and implement them." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 319, + 285, + 355 + ], + "lines": [ + { + "bbox": [ + 52, + 319, + 285, + 355 + ], + "spans": [ + { + "bbox": [ + 52, + 319, + 285, + 355 + ], + "type": "table", + "html": "
MethodFD↓User study↑APDM↑APDP↑APDT↑PE↓RE↓Penetration↓Sliding↓
NSM [56]90.393.950001.720.406.430.69
SAMP [19]68.863.771.210.1120.014.774.847.831.38
Ours57.024.042.620.9362.231.010.284.850.80
", + "image_path": "6b541412e76b946b207f101f05679b7483e7bedcb34e0a0ff227c946756e488e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 390, + 285, + 513 + ], + "blocks": [ + { + "bbox": [ + 50, + 390, + 285, + 513 + ], + "lines": [ + { + "bbox": [ + 50, + 390, + 285, + 513 + ], + "spans": [ + { + "bbox": [ + 50, + 390, + 285, + 513 + ], + "type": "image", + "image_path": "e24db9e6a07fcb9fcc57645ee0972f302e45bbe19bc8d4092e2bb9c0aaed7126.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 516, + 287, + 538 + ], + "lines": [ + { + "bbox": [ + 47, + 516, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 287, + 538 + ], + "type": "text", + "content": "Figure 5. Results in a cluttered scene. Our method can generate motions that avoid obstacles in a cluttered scene." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "type": "text", + "content": "higher than the approaches [19, 64] that employ deterministic " + }, + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 46, + 562, + 287, + 621 + ], + "type": "text", + "content": " [18], but much lower than our method. Although COUCH [76] exhibits lower foot sliding than our method, it may sometimes get stuck, resulting in lower foot sliding since the character does not move." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 651, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 286, + 712 + ], + "type": "text", + "content": "Results on the NSM dataset. On the NSM dataset [56], we compare our approach with SAMP [19] and NSM [56]. Tab. 3 shows that our method outperforms baselines. Compared with the deterministic method NSM, our approach could generate stochastic motion and diverse trajectories." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 312, + 206, + 544, + 278 + ], + "blocks": [ + { + "bbox": [ + 72, + 363, + 261, + 374 + ], + "lines": [ + { + "bbox": [ + 72, + 363, + 261, + 374 + ], + "spans": [ + { + "bbox": [ + 72, + 363, + 261, + 374 + ], + "type": "text", + "content": "Table 3. Quantitative results on the NSM dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 206, + 544, + 278 + ], + "lines": [ + { + "bbox": [ + 312, + 206, + 544, + 278 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 544, + 278 + ], + "type": "table", + "html": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
w/o GP31.073.212.4685.214.100.49
w/o MT27.873.623.01127.945.300.72
w/o MP25.144.013.7885.694.580.43
w/o TC36.772.762.4283.264.340.63
Ours22.344.064.5291.384.000.50
", + "image_path": "60d63e97046e651a1b62d78651063507c2e089df3a931c661c1b0225b03fd2eb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 286, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 305, + 286, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 545, + 353 + ], + "type": "text", + "content": "Table 4. Ablation study of the impact of sub-modules. Although the variant without MT generates more diverse trajectories as shown in " + }, + { + "bbox": [ + 305, + 286, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_T" + }, + { + "bbox": [ + 305, + 286, + 545, + 353 + ], + "type": "text", + "content": ", the motion quality is much worse as indicated by the higher FD, penetration, and sliding. w/o: without. GP: goal pose generation. MT: milestone point generation. MP: milestone pose generation. TC: trajectory completion." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "content": "Results in a cluttered scene. We show our generated results in a cluttered scene in Fig. 5. The percentage of frames with penetration is " + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "content": " for our method and " + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "content": " for SAMP. More details are in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 440, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 440, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 440, + 545, + 499 + ], + "type": "text", + "content": "Qualitative results. As demonstrated in Fig. 6, our approach achieves better results than baselines [19, 63] on the SAMP dataset. Fig. 7 compares our method with COUCH and TDNS on the COUCH dataset. More qualitative results are in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 509, + 397, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 397, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 397, + 521 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 528, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 624 + ], + "type": "text", + "content": "Impact of each sub-module. To show the effectiveness of our hierarchical design, we evaluate our method against four variants where we remove one sub-module for each variant. Tab. 4 indicates that each component improves performance. Generating a whole trajectory leads to more diverse trajectories but the motion quality is worse and has more foot sliding. This validates the necessity of the separate generation of trajectories and motions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "content": "Goal pose generation. We evaluate our goal pose generator with several variants, including cVAE, DDPM, and standard VQ-VAE. For this evaluation, we only replace the goal pose generation module and keep the others the same." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Tab. 5 shows that part VQ-VAE generates more diverse poses than continuous latent space models. The comparison" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15067" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 218, + 153 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 218, + 153 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 218, + 153 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 218, + 153 + ], + "type": "image", + "image_path": "8f9493cac2b81052dc5a587d9c057ebf9b446360c9363ca85d6243807b0f7431.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 218, + 72, + 375, + 152 + ], + "blocks": [ + { + "bbox": [ + 218, + 72, + 375, + 152 + ], + "lines": [ + { + "bbox": [ + 218, + 72, + 375, + 152 + ], + "spans": [ + { + "bbox": [ + 218, + 72, + 375, + 152 + ], + "type": "image", + "image_path": "0ba3fbe2fb50f1044eb3563862c079b92157e47faaadf7473cc907da1152334c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 375, + 72, + 531, + 152 + ], + "blocks": [ + { + "bbox": [ + 375, + 72, + 531, + 152 + ], + "lines": [ + { + "bbox": [ + 375, + 72, + 531, + 152 + ], + "spans": [ + { + "bbox": [ + 375, + 72, + 531, + 152 + ], + "type": "image", + "image_path": "8dcac349f5f41eaa22b12a170e752aa93073df06a6aa1efc680f2e89c278a8e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 154, + 218, + 232 + ], + "blocks": [ + { + "bbox": [ + 62, + 154, + 218, + 232 + ], + "lines": [ + { + "bbox": [ + 62, + 154, + 218, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 154, + 218, + 232 + ], + "type": "image", + "image_path": "bc15e7ad015fa69354d59bc3d47de8e65f6232ba24aa41155eb93e183479aea4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 218, + 154, + 375, + 232 + ], + "blocks": [ + { + "bbox": [ + 218, + 154, + 375, + 232 + ], + "lines": [ + { + "bbox": [ + 218, + 154, + 375, + 232 + ], + "spans": [ + { + "bbox": [ + 218, + 154, + 375, + 232 + ], + "type": "image", + "image_path": "4dba9b99cd50a9a74f708a7de38f81c37f818b8ffc42918eff0e240b0ab814e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 375, + 154, + 531, + 232 + ], + "blocks": [ + { + "bbox": [ + 375, + 154, + 531, + 232 + ], + "lines": [ + { + "bbox": [ + 375, + 154, + 531, + 232 + ], + "spans": [ + { + "bbox": [ + 375, + 154, + 531, + 232 + ], + "type": "image", + "image_path": "af57b0727e3512349c25ebf78da51e9323b9aa52d1e47624881441957ecb2bf9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 233, + 218, + 312 + ], + "blocks": [ + { + "bbox": [ + 62, + 233, + 218, + 312 + ], + "lines": [ + { + "bbox": [ + 62, + 233, + 218, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 233, + 218, + 312 + ], + "type": "image", + "image_path": "8212b3241dd2b7457ded4c158602a2a1bc7f719d8393d20d4beb51b9356eefa4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 313, + 146, + 320 + ], + "lines": [ + { + "bbox": [ + 128, + 313, + 146, + 320 + ], + "spans": [ + { + "bbox": [ + 128, + 313, + 146, + 320 + ], + "type": "text", + "content": "SAMP" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 218, + 233, + 375, + 312 + ], + "blocks": [ + { + "bbox": [ + 218, + 233, + 375, + 312 + ], + "lines": [ + { + "bbox": [ + 218, + 233, + 375, + 312 + ], + "spans": [ + { + "bbox": [ + 218, + 233, + 375, + 312 + ], + "type": "image", + "image_path": "4dd1961b5870def3098fcee476ff9f9757cab68e23d62861a712b6b58eb06080.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 313, + 306, + 320 + ], + "lines": [ + { + "bbox": [ + 288, + 313, + 306, + 320 + ], + "spans": [ + { + "bbox": [ + 288, + 313, + 306, + 320 + ], + "type": "text", + "content": "TDNS" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 375, + 233, + 531, + 312 + ], + "blocks": [ + { + "bbox": [ + 375, + 233, + 531, + 312 + ], + "lines": [ + { + "bbox": [ + 375, + 233, + 531, + 312 + ], + "spans": [ + { + "bbox": [ + 375, + 233, + 531, + 312 + ], + "type": "image", + "image_path": "61a6ba53db2a0cc9c475b2fb2e92c509b0c4911bbc940644f2c5837ebc30a58a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 313, + 460, + 320 + ], + "lines": [ + { + "bbox": [ + 446, + 313, + 460, + 320 + ], + "spans": [ + { + "bbox": [ + 446, + 313, + 460, + 320 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 323, + 545, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 545, + 390 + ], + "type": "text", + "content": "Figure 6. Qualitative results on the SAMP dataset. We compare our method with the baselines SAMP [19] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that SAMP and TDNS tend to walk backward with more foot sliding. The second row shows that the baselines start to sit and lie down before reaching the object while our method synthesizes more natural results. The third row demonstrates that our framework has the capability to generate a long trajectory and walk naturally along the trajectory while SAMP gets stuck near the object and TDNS synthesizes unnatural motions. Lines on the floor indicate trajectories. The human with pink clothes indicate the start position. Darker color denotes later frames in the sequence." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 66, + 403, + 220, + 478 + ], + "blocks": [ + { + "bbox": [ + 66, + 403, + 220, + 478 + ], + "lines": [ + { + "bbox": [ + 66, + 403, + 220, + 478 + ], + "spans": [ + { + "bbox": [ + 66, + 403, + 220, + 478 + ], + "type": "image", + "image_path": "43cc387d22673748ba60534607c8a53f6cbfb000e1123e50884bca7cd93e8312.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 221, + 403, + 375, + 478 + ], + "blocks": [ + { + "bbox": [ + 221, + 403, + 375, + 478 + ], + "lines": [ + { + "bbox": [ + 221, + 403, + 375, + 478 + ], + "spans": [ + { + "bbox": [ + 221, + 403, + 375, + 478 + ], + "type": "image", + "image_path": "d26e69c4f57a92c329d6c93b3d95018e99321d394730fe2296be3935266467a3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 375, + 403, + 529, + 478 + ], + "blocks": [ + { + "bbox": [ + 375, + 403, + 529, + 478 + ], + "lines": [ + { + "bbox": [ + 375, + 403, + 529, + 478 + ], + "spans": [ + { + "bbox": [ + 375, + 403, + 529, + 478 + ], + "type": "image", + "image_path": "abb3e4cab11a0b18ed1742fe33e0c14c3a74bc39e44041eb4e40918ab38b1164.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 66, + 478, + 220, + 553 + ], + "blocks": [ + { + "bbox": [ + 66, + 478, + 220, + 553 + ], + "lines": [ + { + "bbox": [ + 66, + 478, + 220, + 553 + ], + "spans": [ + { + "bbox": [ + 66, + 478, + 220, + 553 + ], + "type": "image", + "image_path": "53e0383f74e904058261f0bf153a081b7d8fb539361621e5f3f85265db8004fd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 555, + 154, + 563 + ], + "lines": [ + { + "bbox": [ + 133, + 555, + 154, + 563 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 154, + 563 + ], + "type": "text", + "content": "COUCH" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 221, + 478, + 375, + 554 + ], + "blocks": [ + { + "bbox": [ + 221, + 478, + 375, + 554 + ], + "lines": [ + { + "bbox": [ + 221, + 478, + 375, + 554 + ], + "spans": [ + { + "bbox": [ + 221, + 478, + 375, + 554 + ], + "type": "image", + "image_path": "289ff845a0eda095049d9aa78e4f87ea0980709f4cc7b77e8443f280218ea88f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 555, + 306, + 563 + ], + "lines": [ + { + "bbox": [ + 290, + 555, + 306, + 563 + ], + "spans": [ + { + "bbox": [ + 290, + 555, + 306, + 563 + ], + "type": "text", + "content": "TDNS" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 375, + 478, + 529, + 554 + ], + "blocks": [ + { + "bbox": [ + 375, + 478, + 529, + 554 + ], + "lines": [ + { + "bbox": [ + 375, + 478, + 529, + 554 + ], + "spans": [ + { + "bbox": [ + 375, + 478, + 529, + 554 + ], + "type": "image", + "image_path": "c898e148c0c93190d136944fdcbeb29117abad7195ef1d95300a5c9c5ff234d6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 555, + 459, + 563 + ], + "lines": [ + { + "bbox": [ + 445, + 555, + 459, + 563 + ], + "spans": [ + { + "bbox": [ + 445, + 555, + 459, + 563 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 568, + 545, + 613 + ], + "lines": [ + { + "bbox": [ + 46, + 568, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 545, + 613 + ], + "type": "text", + "content": "Figure 7. Qualitative results on the COUCH dataset. We compare our method with the baseline COUCH [76] and TDNS [63]. The failure cases are pointed out by the red rectangles. Specifically, the first row indicates that COUCH and TDNS tend to sit when the character is far from the object. The second row shows that COUCH may fail to stand up and TDNS may stand up unnaturally. To better visualize the results of TDNS in the second row, we only keep the frames in which the human stands up and goes to the endpoint." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 632, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 669 + ], + "type": "text", + "content": "with standard VQ-VAE shows the necessity of our part design. We also try part VQ-VAE for motion infilling, but the results in Tab. 6 show that its performance is worse." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Milestone generation. To further investigate the impact of milestones, we compare our approach with the variants" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "text", + "content": "that employ path-planning methods and select points along the path as milestones. For this ablation, we implement " + }, + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "text", + "content": " and NM [63] proposed by TDNS. As demonstrated in Tab. 7, the diversity of generated trajectories of " + }, + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "text", + "content": " [18] is much worse, and the motion quality drops significantly, as indicated by the lower " + }, + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_T" + }, + { + "bbox": [ + 304, + 632, + 545, + 706 + ], + "type": "text", + "content": " and higher FD. The trajectory" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15068" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 71, + 285, + 128 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 285, + 128 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 285, + 128 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 285, + 128 + ], + "type": "table", + "html": "
VariantsFD ↓APDM ↑APDP ↑APDT ↑Penetration↓Sliding↓
cVAE27.063.363.2790.524.360.47
DDPM34.223.753.3784.764.310.49
VQ-VAE24.773.783.8789.194.270.48
Part VQ-VAE22.344.064.5291.384.000.50
", + "image_path": "f36f81c4f257b1cbd43493ee2237f269afb7ef3b7a7e1f68e77120c21be8c7aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 168, + 285, + 204 + ], + "blocks": [ + { + "bbox": [ + 47, + 136, + 287, + 158 + ], + "lines": [ + { + "bbox": [ + 47, + 136, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 287, + 158 + ], + "type": "text", + "content": "Table 5. Ablation study of goal pose generation. We implement the goal pose module with different architectures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 168, + 285, + 204 + ], + "lines": [ + { + "bbox": [ + 53, + 168, + 285, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 285, + 204 + ], + "type": "table", + "html": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
Part VQ-VAE28.713.853.6283.514.320.89
DDPM22.344.064.5291.384.000.50
", + "image_path": "eadc86cd35b81fe387d7f936edd5541252c87804459137e5dbd6907594d53b8d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 53, + 240, + 285, + 289 + ], + "blocks": [ + { + "bbox": [ + 47, + 207, + 286, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 286, + 229 + ], + "type": "text", + "content": "Table 6. Ablation study of part VQ-VAE for motion infilling. We replace the DDPM as part VQ-VAE to predict motions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 240, + 285, + 289 + ], + "lines": [ + { + "bbox": [ + 53, + 240, + 285, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 285, + 289 + ], + "type": "table", + "html": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
A*40.594.074.4242.984.240.88
NM [63]40.544.154.3152.744.150.96
MT22.344.064.5291.384.000.50
", + "image_path": "fb9bea38f1c9c62796ef15c2e5cf816f726148edc6da0ac185c97b6ac3d4adbc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 53, + 352, + 285, + 405 + ], + "blocks": [ + { + "bbox": [ + 46, + 297, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 342 + ], + "type": "text", + "content": "Table 7. Ablation study of milestone generation. We compare our method with the variant based on the path generated by " + }, + { + "bbox": [ + 46, + 297, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^*" + }, + { + "bbox": [ + 46, + 297, + 287, + 342 + ], + "type": "text", + "content": " path planning [18]. NM: Neural Mapper [63]. MT: milestone point generation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 352, + 285, + 405 + ], + "lines": [ + { + "bbox": [ + 53, + 352, + 285, + 405 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 285, + 405 + ], + "type": "table", + "html": "
MethodFD↓APDM↑APDP↑Penetration↓Sliding↓
ConvAE [30]26.503.954.334.820.89
SLT [64]27.953.764.054.130.78
Ours22.344.064.524.000.50
", + "image_path": "bc604761b7ea7d51df11e3f3eec982c34b6a6f4c7dc0be14616180f3e43a1fd9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "type": "text", + "content": "diversity of NM [63] is better than " + }, + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^*" + }, + { + "bbox": [ + 46, + 451, + 287, + 499 + ], + "type": "text", + "content": ", but still worse than our method. The reason why these variants perform poorly might be the low diversity of trajectory that affects the distribution of generated motions for calculating FD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 516, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 588 + ], + "type": "text", + "content": "Motion infilling. To validate our motion infilling module, we compare it with ConvAE [30] and SLT [64]. We only replace the motion infilling module and keep the others the same. The comparison of motion quality and diversity is shown in Tab. 8 and our method outperforms ConvAE [30] and SLT [64] with lower FD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": "DDPM vs. VAE. To show the effectiveness of DDPM, we implement a cVAE variant, where we simply replace our transformer DDPM with transformer cVAE [45, 63]. As shown in Tab. 9, although cVAE models could generate more diverse trajectories, their motion quality is far from satisfactory, indicated by the much higher value of FD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Comparison with other diffusion models. Our approach stands out from architectures in MDM [59] and FLAME" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 312, + 71, + 544, + 105 + ], + "blocks": [ + { + "bbox": [ + 47, + 408, + 287, + 430 + ], + "lines": [ + { + "bbox": [ + 47, + 408, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 287, + 430 + ], + "type": "text", + "content": "Table 8. Ablation study of motion infilling module. We compare our method with other motion infilling methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 71, + 544, + 105 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 544, + 105 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 544, + 105 + ], + "type": "table", + "html": "
ArchFD↓R-APDM↑R-APDP↑R-APDT↑Penetration↓Sliding↓
cVAE29.323.653.58109.555.061.12
DDPM22.344.064.5291.384.000.50
", + "image_path": "536701ef207f755236fe451ac93d8c548c408b92e5231b5f62dcf25d76a62518.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 312, + 168, + 544, + 224 + ], + "blocks": [ + { + "bbox": [ + 305, + 113, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 305, + 113, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 305, + 113, + 545, + 158 + ], + "type": "text", + "content": "Table 9. Evaluation of the architecture for our generation framework. Although cVAE based variant generates more diverse trajectories, the motion quality drops significantly as indicated by the much higher FD. Arch stands for the architecture type." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 168, + 544, + 224 + ], + "lines": [ + { + "bbox": [ + 312, + 168, + 544, + 224 + ], + "spans": [ + { + "bbox": [ + 312, + 168, + 544, + 224 + ], + "type": "table", + "html": "
VariantsFD↓APDM↑APDP↑APDT↑Penetration↓Sliding↓
MDM [59]38.694.013.7772.675.190.70
MDM [59] + C23.974.024.1389.804.870.39
FLAME [31]28.934.024.4665.465.780.89
Ours22.344.064.5291.384.000.50
", + "image_path": "720b3d0b137f785e8725acca6e50c1a8cb95d421a0ad67ab8d6b4506cece2d3d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 228, + 545, + 249 + ], + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 249 + ], + "type": "text", + "content": "Table 10. Comparison with MDM and FLAME. C denotes the frame-wise conditions." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 270, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 293 + ], + "type": "text", + "content": "[31] by incorporating frame-wise conditions. Tab. 10 demonstrates the significance of the frame-wise conditions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "content": "More analyses and ablation studies. More detailed analyses and ablation studies of our design choices are provided in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 361, + 383, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 383, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 383, + 372 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 380, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 475 + ], + "type": "text", + "content": "Although our method can generate diverse and natural motions, there are still some limitations. Our method is offline and cannot be applied to interactive scenarios. We assume that the objects are static and cannot handle moving objects. The diffusion models require a long inference time. It takes 7.13 seconds on average for a 720-frame sequence on a Titan Xp GPU. The slow speed might be solved by methods that could accelerate diffusion models [43, 2]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 511, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 605 + ], + "type": "text", + "content": "In this work, we propose a novel hierarchical pipeline for motion synthesis of human-object interactions. Our approach first generates the goal pose and then predicts a set of milestones. Next, we synthesize motions with the guide of milestones. Furthermore, we apply DDPM in our hierarchical pipeline. We also show that our framework could generate more diverse and natural human-object interaction motions than other methods." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 620, + 408, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 408, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 408, + 634 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "We thank Jintao Lu, Zhi Cen, Zizhang Li, and Kechun Xu for the valuable discussions. This work was partially supported by the Key Research Project of Zhejiang Lab (No. K2022PG1BB01), NSFC (No. 62172364), and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15069" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. In International Conference on Learning Representations, 2022. 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 237 + ], + "type": "text", + "content": "[3] German Barquero, Sergio Escalera, and Cristina Palmero. BeLFusion: Latent Diffusion for Behavior-Driven Human Motion Prediction. arXiv e-prints, page arXiv:2211.14304, Nov. 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 293 + ], + "type": "text", + "content": "[4] Emad Barsoum, John R. Kender, and Zicheng Liu. Hpgan: Probabilistic 3d human motion prediction via gan. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1499-149909, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 384 + ], + "type": "text", + "content": "[5] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, Ding Liu, Jing Liu, and Nadia Magnenat Thalmann. Learning progressive joint propagation for human motion prediction. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, page 226-242, Berlin, Heidelberg, 2020. Springer-Verlag. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 385, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 451 + ], + "type": "text", + "content": "[6] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, page 387-404, Berlin, Heidelberg, 2020. Springer-Verlag. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 507 + ], + "type": "text", + "content": "[7] Yu-Wei Chao, Jimei Yang, Weifeng Chen, and Jia Deng. Learning to sit: Synthesizing human-chair interactions via hierarchical control. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 5887-5895, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 509, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 287, + 553 + ], + "type": "text", + "content": "[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your Commands via Motion Diffusion in Latent Space. arXiv e-prints, page arXiv:2212.04048, Dec. 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 287, + 576 + ], + "type": "text", + "content": "[9] Simon Clavet. Motion matching and the road to next-gen animation. In GDC, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning dynamic relationships for 3d human motion prediction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6518-6526, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[11] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A Framework for Denoising-Diffusion-based Motion Synthesis. arXiv eprints, page arXiv:2212.04495, Dec. 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "volume 34, pages 8780-8794. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 97, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 547, + 151 + ], + "type": "text", + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, June 2021. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 153, + 547, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 547, + 198 + ], + "type": "text", + "content": "[14] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 200, + 547, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 547, + 255 + ], + "type": "text", + "content": "[15] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1396-1406, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 258, + 547, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 547, + 300 + ], + "type": "text", + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, oct 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 303, + 547, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 547, + 358 + ], + "type": "text", + "content": "[17] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, June 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 360, + 547, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 547, + 404 + ], + "type": "text", + "content": "[18] Peter E. Hart, Nils J. Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. 2, 6, 7, 8, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 406, + 547, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 547, + 471 + ], + "type": "text", + "content": "[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11374-11384, October 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 474, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 529 + ], + "type": "text", + "content": "[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J. Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14708-14718, June 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "type": "text", + "content": "[21] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH '23, New York, NY, USA, 2023. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 588, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 620 + ], + "type": "text", + "content": "[22] Chengan He, Jun Saito, James Zachary, Holly Rushmeier, and Yi Zhou. Nemf: Neural motion fields for kinematic animation. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 623, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 545, + 678 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. arXiv e-prints, page arXiv:2207.12598, July 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15070" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. arXiv e-prints, page arXiv:2204.03458, Apr. 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 117, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 117, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 117, + 288, + 149 + ], + "type": "text", + "content": "[26] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Trans. Graph., 36(4), jul 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 194 + ], + "type": "text", + "content": "[27] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based Generation, Optimization, and Planning in 3D Scenes. arXiv e-prints, page arXiv:2301.06015, Jan. 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "text", + "content": "[28] Ashesh Jain, Amir R. Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5308-5317, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 271 + ], + "type": "text", + "content": "[29] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Trans. Graph., 41(3), jun 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 287, + 316 + ], + "type": "text", + "content": "[30] Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pages 918-927, 2020. 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "type": "text", + "content": "[31] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. FLAME: Freeform Language-based Motion Synthesis & Editing. arXiv e-prints, page arXiv:2209.00349, Sept. 2022. 2, 3, 4, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 350, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 380 + ], + "type": "text", + "content": "[32] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv e-prints, page arXiv:1412.6980, Dec. 2014. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 383, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 383, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 383, + 287, + 414 + ], + "type": "text", + "content": "[33] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. arXiv e-prints, page arXiv:1312.6114, Dec. 2013. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 415, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 287, + 502 + ], + "type": "text", + "content": "[34] Jogendra Nath Kundu, Maharshi Gor, and R. Venkatesh Babu. Bihmp-gan: Bidirectional 3d human motion prediction gan. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'19/IAAI'19/EAAI'19. AAAI Press, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 504, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 287, + 557 + ], + "type": "text", + "content": "[35] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 287, + 591 + ], + "type": "text", + "content": "[36] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Trans. Graph., 39(4), aug 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 593, + 287, + 635 + ], + "type": "text", + "content": "[37] Thomas Lucas, Fabien Baradel, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 668 + ], + "type": "text", + "content": "[38] Jianxin Ma, Shuai Bai, and Chang Zhou. Pretrained Diffusion Models for Unified Human Motion Synthesis. arXiv e-prints, page arXiv:2212.02837, Dec. 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[39] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Multi-level motion attention for human motion prediction. International Journal of Computer Vision, 129(9):2513-2535, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[40] Julieta Martinez, Michael J. Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[41] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, June 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "type": "text", + "content": "[42] Tomohiko Mukai and Shigeru Kuriyama. Geostatistical motion interpolation. ACM Trans. Graph., 24(3):1062-1070, jul 2005. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 272 + ], + "type": "text", + "content": "[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8162-8171. PMLR, 18-24 Jul 2021. 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "type": "text", + "content": "[44] Dario Pavllo, David Grangier, and Michael Auli. *Quaternion: A quaternion-based recurrent model for human motion*. In *British Machine Vision Conference (BMVC)*, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[45] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 354, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 396 + ], + "type": "text", + "content": "[46] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "text", + "content": "[47] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. arXiv e-prints, page arXiv:2209.14988, Sept. 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "type": "text", + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv e-prints, page arXiv:2204.06125, Apr. 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "type": "text", + "content": "[51] C. Rose, M.F. Cohen, and B. Bodenheimer. Verbs and adverbs: multidimensional motion interpolation. IEEE Computer Graphics and Applications, 18(5):32-40, 1998. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 614, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 678 + ], + "type": "text", + "content": "[52] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. arXiv e-prints, page arXiv:2209.14792, Sept. 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15071" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[54] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS'20, Red Hook, NY, USA, 2020. Curran Associates Inc. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[55] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Trans. Graph., 41(4), jul 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "text", + "content": "[56] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6), nov 2019. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "text", + "content": "[57] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), aug 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "type": "text", + "content": "[58] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. Goal: Generating 4d whole-body motion for hand-object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "type": "text", + "content": "[59] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 3, 4, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 411 + ], + "type": "text", + "content": "[60] Jonathan Tseng, Rodrigo Castellon, and C. Karen Liu. EDGE: Editable Dance Generation From Music. arXiv eprints, page arXiv:2211.10658, Nov. 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 477 + ], + "type": "text", + "content": "[61] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 555 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 622 + ], + "type": "text", + "content": "[63] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20460-20469, June 2022. 1, 2, 3, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[64] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9401-9411, June 2021. 2, 5, 6, 7, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[65] J. Wang, S. Yan, B. Dai, and D. Lin. Scene-aware generative network for human motion synthesis. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "(CVPR), pages 12201-12210, Los Alamitos, CA, USA, jun 2021. IEEE Computer Society. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 127 + ], + "type": "text", + "content": "[66] Mao Wei, Liu Miaomiao, and Salzemann Mathieu. History repeats itself: Human motion prediction via motion attention. In ECCV, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "text", + "content": "[67] Mao Wei, Liu Miaomiao, Salzemann Mathieu, and Li Hongdong. Learning trajectory dependencies for human motion prediction. In ICCV, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 194 + ], + "type": "text", + "content": "[68] D.J. Wiley and J.K. Hahn. Interpolation synthesis of articulated figure motion. IEEE Computer Graphics and Applications, 17(6):39-45, 1997. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 238 + ], + "type": "text", + "content": "[69] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 303 + ], + "type": "text", + "content": "[70] Xinchen Yan, Akash Rastogi, Ruben Villegas, Kalyan Sunkavalli, Eli Shechtman, Sunil Hadap, Ersin Yumer, and Honglak Lee. Mt-vae: Learning motion transformations to generate multimodal human dynamics. In European Conference on Computer Vision, pages 276–293. Springer, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 305, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 347 + ], + "type": "text", + "content": "[71] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "text", + "content": "[72] He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 37(4), jul 2018. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 384, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 425 + ], + "type": "text", + "content": "[73] He Zhang, Yuting Ye, Takaaki Shiratori, and Taku Komura. Manipnet: Neural manipulation synthesis with a hand-object spatial representation. ACM Trans. Graph., 40(4), jul 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "text", + "content": "[74] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 525 + ], + "type": "text", + "content": "[75] S. Zhang, Y. Zhang, Q. Ma, M. J. Black, and S. Tang. Place: Proximity learning of articulation and contact in 3d environments. In 2020 International Conference on 3D Vision (3DV), pages 642-651, Los Alamitos, CA, USA, nov 2020. IEEE Computer Society. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 526, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 580 + ], + "type": "text", + "content": "[76] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[77] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372-3382, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[78] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J. Black, and Siyu Tang. Generating 3d people in scenes without people. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[79] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20481-20491, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15072" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[80] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "text", + "content": "[81] Yi Zhou, Zimo Li, Shuangjiu Xiao, Chong He, Zeng Huang, and Hao Li. Auto-conditioned recurrent networks for extended complex human motion synthesis. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "text", + "content": "[82] Zixiang Zhou and Baoyuan Wang. UDE: A Unified Driving Engine for Human Motion Generation. arXiv e-prints, page arXiv:2211.16016, Nov. 2022. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15073" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_content_list.json b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e2a8f1a0393d30327285f5902c2a8ca4ce4bf17 --- /dev/null +++ b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_content_list.json @@ -0,0 +1,1870 @@ +[ + { + "type": "text", + "text": "Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation", + "text_level": 1, + "bbox": [ + 109, + 128, + 861, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zongyi Xu $^{1\\dagger}$ Bo Yuan $^{1\\dagger}$ Shanshan Zhao $^{2\\dagger}$ Qianni Zhang $^{3}$ Xinbo Gao $^{1*}$ $^{1}$ Chongqing University of Posts and Telecommunications, China \n $^{2}$ JD Explore Academy, China $^{3}$ Queen Mary University of London, UK \n{xuzy, qaoxb}@cqupt.edu.cn s210201127@stu.cqupt.edu.cn \nsshan.zhao00@gmail.com qianni.zhang@qmul.ac.uk", + "bbox": [ + 179, + 202, + 795, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Impressive performance on point cloud semantic segmentation has been achieved by fully-supervised methods with large amounts of labelled data. As it is labour-intensive to acquire large-scale point cloud data with pointwise labels, many attempts have been made to explore learning 3D point cloud segmentation with limited annotations. Active learning is one of the effective strategies to achieve this purpose but is still under-explored. The most recent methods of this kind measure the uncertainty of each pre-divided region for manual labelling but they suffer from redundant information and require additional efforts for region division. This paper aims at addressing this issue by developing a hierarchical point-based active learning strategy. Specifically, we measure the uncertainty for each point by a hierarchical minimum margin uncertainty module which considers the contextual information at multiple levels. Then, a feature-distance suppression strategy is designed to select important and representative points for manual labelling. Besides, to better exploit the unlabelled data, we build a semi-supervised segmentation framework based on our active strategy. Extensive experiments on the S3DIS and ScanNetV2 datasets demonstrate that the proposed framework achieves $96.5\\%$ and $100\\%$ performance of fully-supervised baseline with only $0.07\\%$ and $0.1\\%$ training data, respectively, outperforming the state-of-the-art weakly-supervised and active learning methods. The code will be available at https://github.com/SmiletoE/HPAL.", + "bbox": [ + 75, + 358, + 472, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud semantic segmentation aims to assign a category label for each 3D point, which can be applied to various scenarios, such as robotics [34], autonomous driving [1], and augmented reality [20]. Recently, deep learning", + "bbox": [ + 75, + 818, + 468, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based methods [38, 52] have achieved impressive performance. These high-performing methods usually rely on large amounts of data with point-wise labels. However, acquiring such dense labels for 3D point clouds is extremely tedious and costly.", + "bbox": [ + 500, + 327, + 890, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To relieve the labour and cost of annotation, many recent methods explore semi-supervised learning to learn 3D segmentation models with limited labelled points that are usually sampled randomly. These methods attempt to employ effective strategies to propagate the label information to the unlabelled points [5, 17, 10, 13]. Although these semi-supervised methods can greatly decrease labelling costs, their performance might be limited due to various factors. Among them, the most important reason is that some of the randomly selected points are in fact redundant while some really important points might be omitted.", + "bbox": [ + 496, + 407, + 892, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As an alternative learning strategy, active learning recently is studied to alleviate such limitations for 3D segmentation. Lin et al. [24] divide the whole point clouds into segments and each segment is utilised as the basic query unit for sample selection. ReDAL [41] proposes to select those informative and diverse sub-scene regions for label acquisition. The entropy, colour discontinuity, and structural complexity are used to measure the information of sub-scene regions. Following it, SSDR-AL [29] groups the original point clouds into superpoints and incrementally selects the most informative regions for annotation. However, these methods usually rely on pre-dividing the point cloud into multiple regions, and the region division strategy might negatively impact the performance. Moreover, as point clouds present strong semantic similarity in local areas, selecting all the points in the local region results in a redundant labelling budget.", + "bbox": [ + 496, + 579, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by the above analysis, this research aims at enhancing the 3D segmentation performance with limited labelled data in the active learning framework. Specifically, in comparison with previous region-based methods, we mea", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. †Equal contribution.", + "bbox": [ + 94, + 886, + 331, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18098", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sure the uncertainty or importance of labelling for each point to avoid the additional efforts and potential bias of region division. However, considering each point individually may introduce extra noise to the training process, leading to unreliable uncertainty values that cannot reflect the real importance of each point. To alleviate this issue, we design a hierarchical minimum margin uncertainty (HMMU) measurement module by considering the local contextual information. The HMMU involves the uncertainty of the point and its neighbourhood. In this way, the acquired score can more effectively represent the importance of each point label. Based on the uncertainty scores acquired by HMMU, we can directly choose the Top- $K$ points for labelling. However, as the point cloud presents semantic similarity in the local region, some redundant uncertain points tend to be distributed densely. To further reduce the labelling cost, we propose a feature-distance suppression module (FDS) to remove those uncertain points with similar features in the neighbourhood. By exploiting the devised HMMU measurement and FDS for redundant point filtering, our hierarchical point-based active learning strategy is able to select more valuable points for labelling and thus reduce labelling costs. Furthermore, inspired by previous semi-supervised methods that exploit the unlabelled points, we utilise a simple teacher-student scheme to enhance the supervision signal by assigning a pseudo label for the unlabelled points.", + "bbox": [ + 76, + 90, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this research are as follows:", + "bbox": [ + 96, + 484, + 455, + 498 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a point-based active learning method for semi-supervised point cloud semantic segmentation, which surpasses current semi-supervised or active learning methods and achieves comparable performance with full supervision methods relying on scarce annotations.", + "- We propose a novel hierarchical minimum margin uncertainty module to measure the uncertainty for each point by progressively perceiving the contextual information at increasing scales.", + "- We propose a feature-distance suppression module to remove redundant points with similar features in the neighbourhood and further minimise the labelling cost." + ], + "bbox": [ + 96, + 503, + 467, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 731, + 218, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 756, + 318, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, a large number of deep learning-based point cloud semantic segmentation methods have been developed, which can be roughly divided into the following categories: 1) 2D projection-based methods [4, 19, 2, 8, 26, 39, 40, 44], which project the 3D point cloud onto 2D images through a variety of viewpoints including multiview, bird's eye view, spherical projection etc. Projection-based methods can make full use of those well-designed 2D", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "convolution networks for 3D scene parsing, but obviously, the projection causes the loss of 3D geometric information and thus limits the performance. 2) voxel-based methods [6, 11, 31, 47, 7, 33]. Point cloud voxelization regularizes point clouds with uneven densities, making it possible to extend regular 2D convolutions to 3D scenes. At an early stage, voxel-based methods usually suffer from the explosion of computation and the loss of information. Fortunately, some sparse convolution methods [11, 31, 7] have been proposed over time to greatly ease the computational costs. 3) point-based methods [14, 27, 28, 35, 23, 42]. These methods directly take the original uneven point cloud as input and use MLPs, 3D point convolution, and other more flexible ways to process the point clouds, which preserve the original 3D information. However, their high performance still relies on a large number of point-level labels.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Weakly-supervised 3D Semantic Segmentation", + "text_level": 1, + "bbox": [ + 500, + 343, + 890, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We collectively refer to all methods training with sparse annotation as weakly supervised methods, which is a well-explored field in point cloud segmentation. These methods are usually based on a small set of randomly labelled data and then exploit more information from the unlabelled data by leveraging techniques such as transfer learning, contrastive learning, consistency regularization, pseudo-label, etc. Based on the transfer learning technologies, pre-training is widely exploited in 2D images, and has recently been gradually applied to point clouds [50, 49, 43, 53, 37, 45]. These methods learn pre-trained knowledge from 3D point clouds or 2D images and then apply it to the target dataset, thus achieving better training results with limited labelled data. Some other works train the model by using consistency regularization [46, 51, 22, 48] and pseudo-label [13, 36, 5, 25, 50, 22]. SQN [13] is one of these works which designs a novel semantic query network to efficiently obtain the pseudo-labels. To further improve weakly supervised training, some approaches also use contrastive learning [22, 25, 12, 18, 43], which usually combine contrastive loss with pre-training, consistency regularization, and pseudo-labels to better exploit the supervision information. Most weakly-supervised methods explore information from the labelled data, so how to select the most informative data to label is a question of crucial importance.", + "bbox": [ + 496, + 367, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Active Learning on 3D Semantic Segmentation", + "text_level": 1, + "bbox": [ + 500, + 755, + 890, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similar to weakly-supervised learning, active learning also aims to train deep models with limited labelled data. The main difference is that weakly-supervised learning mainly focuses on enhancing supervision by exploiting available label information while active learning focuses on the selection of valuable information for labelling.", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lin et al. [24] make the first attempt for active learning on 3D semantic segmentation using segments as the ba", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18099", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ac33a529850ced3fdabce95f9d216f413bb77ad63190df5c58ea36b48a1b59a.jpg", + "image_caption": [ + "Figure 1. The proposed hierarchical point-based active learning framework for semi-supervised point cloud semantic segmentation. (a) A segmentation network using the teacher-student model is trained with the initial labelled and unlabelled training data. The teacher is updated from the student by the exponential moving averaging (EMA) and the student network is obtained as our target segmentation model. (b) Those representative candidate points are measured by the trained model using the proposed hierarchical minimum margin uncertainty module (HMMU). (c) Feature-distance Suppression module (FDS) is then applied to remove those redundant points in the neighbourhood and select the final to-be-annotated points. The selected valuable points are annotated and added to the labelled training set to improve the segmentation model." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 851, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sic query unit, and propose a segment entropy strategy to measure the informativeness of each segment. To improve the efficiency of active learning, superpoints are introduced as the most common basic query units [30, 41]. Considering the limitation of uncertainty, Shi et al. [30] propose to design active acquisition functions from multiple perspectives, including feature diversity, shape diversity, and entropy. ReDAL [41] combines entropy with the geometric structure and colour information and uses feature clustering to prevent selection redundancy. In addition to the entropy-based method, SSDR-AL [29] improves MMU based on super-point by considering the principal class within superpoints. To date, all active learning methods for point cloud semantic segmentation are region based. However, the pre-region division can greatly affect the active learning performance and unnecessary annotation in the semantic-similar areas is prone to occur.", + "bbox": [ + 75, + 425, + 472, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 699, + 187, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 724, + 187, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we describe our proposed hierarchical point-based active learning approach for semi-supervised point cloud semantic segmentation in detail. As shown in Figure 1, the initial labelled and unlabelled training data are utilised to train the point cloud semantic segmentation neural network in a semi-supervised way. A teacher-student model is adopted. With the trained model, the valuable unlabelled points that are able to improve the segmentation performance most can be measured and selected in an active learning manner. The hierarchical minimum margin uncer", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tainty module is proposed to measure the point-wise uncertainty by hierarchically increasing contextual ranges, which enables progressively capturing the local context at different scales. Then, the neighbouring points with similar features are removed by the feature distance suppression module, thereby further reducing the annotation redundancy. The remaining points with high uncertainty scores as well as sparse distributions are regarded as the candidates to be annotated. In this manner, the most valuable and representative unlabelled points are selected to expand the labelled set, with which the semi-supervised point cloud semantic segmentation neural network can be further enhanced. In the following, we describe each component of the proposed framework in detail.", + "bbox": [ + 496, + 425, + 893, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hierarchical Minimum Margin Uncertainty Measurement", + "text_level": 1, + "bbox": [ + 498, + 648, + 890, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the proposed method, we aim to select and label the most valuable points that can bring maximum performance improvements. Intuitively, identifying and using the most informative labelled data for training is the key to obtaining a model with good quality, and consequently accurate segmentation prediction. Merely considering individual points without their surrounding context information cannot reflect the real importance of the point. Thus, we design the hierarchical minimum margin uncertainty measurement module to calculate the uncertainty score for each point by grouping points at multiple scales and progressively perceiving context information in a broader range along the hierarchy for the unlabelled point.", + "bbox": [ + 496, + 688, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 1, the prediction for each point gen", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18100", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "erated with the currently trained model is input into the HMMU module. We first calculate the point-level minimum margin uncertainty score $U_{x}$ for each unlabelled point with Eq. 1:", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU _ {x} = h \\left(x ^ {u}; p _ {1} \\left(x ^ {u}\\right)\\right) - h \\left(x ^ {u}; p _ {2} \\left(x ^ {u}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 165, + 468, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x^{u}$ is the candidate point that is chosen to be labelled; $p_1(x^u)$ and $p_2(x^u)$ are the highest and second-highest prediction of $x^{u}$ under the segmentation predictor $h(\\cdot)$ .", + "bbox": [ + 76, + 196, + 468, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we group a wider range of neighbours through downsampling. The softmax prediction of each downsampled point is obtained by averaging the softmax prediction of its neighbouring points in the original point cloud. Each softmax label of the downsampled point represents the prediction distribution of a local region, which is denoted by $S_{R}$ .", + "bbox": [ + 75, + 243, + 468, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We perform $N$ levels of voxel downsampling. The local context information of the to-be-annotated point $x^{u}$ at the $i_{th}$ level of downsampling, denoted by $S_R^i (x^u)$ , can be represented as the average of predictions of the grouped points in the following:", + "bbox": [ + 75, + 349, + 468, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {R} ^ {i} \\left(x ^ {u}\\right) = \\frac {1}{K} \\sum_ {j = 1} ^ {K} p \\left(x _ {j} ^ {u}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 439, + 468, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{j}^{u}$ is the $j_{th}$ neighbour for $x^{u}$ in the original point cloud; $K$ is the number of neighbours within the predefined voxel radius $v_{r}$ ; and $p(\\cdot)$ is the prediction probability.", + "bbox": [ + 76, + 494, + 468, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each level of downsampling, the voxel-level contextual uncertainty scores $U_{R}^{i}$ for the unlabelled point $x^{u}$ can be obtained with:", + "bbox": [ + 75, + 541, + 468, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU _ {R} ^ {i} = h \\left(x ^ {u}; S _ {R 1} ^ {i} \\left(x ^ {u}\\right)\\right) - h \\left(x ^ {u}; S _ {R 2} ^ {i} \\left(x ^ {u}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 599, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{R1}^{i}(x^{u})$ and $S_{R2}^{i}(x^{u})$ are the highest and the second highest average prediction of the grouped points of $x^{u}$ from the segmentor $h(\\cdot)$ in the $i_{th}$ downsampling.", + "bbox": [ + 76, + 631, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, for each unannotated point $x^u$ , we integrate the point-level and the voxel-level contextual uncertainty scores to get the final uncertainty $v^u$ using Eq. 4:", + "bbox": [ + 75, + 678, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv ^ {u} = U _ {x} + \\sum_ {i = 1} ^ {N} \\omega^ {i} \\times U _ {R} ^ {i}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 737, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\omega^i\\in \\{0.1,0.01,0.001\\}$ is the hyperparameter representing the fusion weight of the contextual uncertainty at the $i_{th}$ level of downsampling and three levels of downsampling is performed in the experiment.", + "bbox": [ + 75, + 792, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With HMMU, the context information is fused to significantly improve performance by merely annotating points rather than regions.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6846c69fb3aba47166d515065c268b83f7dce9c71f1a6cdc34208b84c17e8331.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 694, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab1732e2df70d568dd28cc51d9f005751291274703fbf2bae48f33b69a8dfb4e.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 870, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/530891ab13e28aef388ed5b3b791b1b4c291e0bf98261b155e8229be8963ff26.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 524, + 185, + 694, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b612dccfc37e0870c61f196a023acd436dc0fc37aee3691fb93fdc32fafc9e5.jpg", + "image_caption": [ + "(d)", + "Figure 2. Visualization of selected points. (a) and (c) show the points selected using the Top- $K$ approach (blue dots); (b) and (d) present points chosen by FDS (orange dots). FDS enables a more spread-out point selection." + ], + "image_footnote": [], + "bbox": [ + 697, + 185, + 870, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature-distance Suppression based Point Selection", + "text_level": 1, + "bbox": [ + 500, + 347, + 890, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the point-wise uncertainty score $v^u$ , we can select the Top- $K$ points with the highest uncertainty to label. However, as shown in Figure 2, such a point selection way might make points concentrate in local areas, causing label redundancy. Thus, we propose a feature-distance suppression (FDS) module to ensure the selected points retain a spread-out distribution in the space, thus offering a more effective overall representation.", + "bbox": [ + 498, + 386, + 890, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a distance suppression radius $r$ and a feature similarity threshold $\\tau$ , for a point $x_{i}$ to be selected, we first determine whether there are points within its radius $r$ that have been already selected. A judgment set $D^{i}$ is constructed for the candidate point $x_{i}$ in the current point selection iteration.", + "bbox": [ + 498, + 507, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {i} = D ^ {i} \\cup \\{x _ {j} \\}, i f \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 594, + 890, + 618 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\forall x _ {j} i s s e l e c t e d: d _ {i j} < r,\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 614, + 782, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{j}$ is the neighbour of $x_{i}$ within the radius of $r$ ; $d_{ij}$ is the Euclidean distance between $x_{i}$ and $x_{j}$ ; and $D^{i}$ represents if there are other candidate points in the neighbourhood and $D^{i}$ is initialized to be empty.", + "bbox": [ + 498, + 635, + 890, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If there is no neighbour of $x_{i}$ that is selected as a candidate, i.e. the judgment set $D^{i}$ is empty, $x_{i}$ is regarded as the valuable point that can represent the local area for labelling. Conversely, if the neighbouring point $x_{j}$ is already selected as the uncertainty point, cosine similarity $Sim_{ij}$ between the features of $x_{i}$ and $x_{j}$ is then computed with Eq. 6.", + "bbox": [ + 498, + 696, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} \\left(x _ {i}, x _ {j}\\right) = \\frac {\\mathbf {f} _ {i} \\cdot \\mathbf {f} _ {j}}{\\left| \\left| \\mathbf {f} _ {i} \\right| \\right| \\cdot \\left| \\left| \\mathbf {f} _ {j} \\right| \\right|}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 797, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_i$ and $\\mathbf{f}_j$ are the features of $x_{i}$ and $x_{j}$ extracted with the trained student segmentation network.", + "bbox": [ + 498, + 839, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If there exists an $x_{j}$ in $D^{i}$ for which $Sim_{ij} > \\tau$ , $x_{i}$ is regarded as a redundant point with respect to the labelled", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "data and $x_{i}$ is excluded from labelling.", + "bbox": [ + 76, + 90, + 334, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training objective", + "text_level": 1, + "bbox": [ + 76, + 114, + 253, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train a network with partly annotated data, a usual approach is to apply the cross-entropy loss to the labelled points and ignore the unlabelled ones, which causes inadequate supervision for model learning. Inspired by previous semi-supervised 3D point cloud segmentation methods, in the proposed method we employ a teacher-student framework to exploit unlabelled points and provide additional supervision. Specifically, we construct two segmentation networks using MinkowskiNet [7], denoted as the teacher model $M_{t}$ and student model $M_{s}$ , respectively. A consistency constraint between the teacher model and the student model is exploited to learn knowledge from unlabelled data. Pseudo-labels of unlabelled data are generated using the teacher model. During training, the original sample is input into the teacher network while its augmented counterparts are input into the student. For those labelled points, we apply the standard cross-entropy loss. For the unlabelled ones, we generate pseudo-labels from the output of the teacher and also compute the cross-entropy loss. The losses in both labelled and unlabelled points are then used to optimize the student network by gradient descent. After updating the student model at each step, Exponential Moving Average (EMA) is used to transfer the parameters of the student model to the teacher model:", + "bbox": [ + 75, + 138, + 472, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} ^ {j} = \\alpha \\theta_ {t} ^ {j - 1} + (1 - \\alpha) \\theta_ {s} ^ {j}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 508, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{t}$ and $\\theta_{s}$ are the parameters of the teacher and student networks, respectively, and $j$ denotes the $j$ -th training step. $\\alpha$ is the hyper-parameter to determine the speed of parameter transmission, which is generally close to 1.", + "bbox": [ + 75, + 537, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 609, + 209, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 635, + 179, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed approach on two publicly available benchmark datasets, S3DIS [3] and ScanNetV2 [9], to demonstrate its superior performance.", + "bbox": [ + 75, + 659, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "S3DIS (Stanford Large-Scale 3D Indoor Space Dataset) is a large-scale indoor point cloud dataset captured by Matter-port scanners which mainly dedicate to 3D segmentation tasks. It consists of 6 different large-scale indoor areas with a total of 271 rooms, each room is a separate point cloud sample that includes coordinate, colour, and annotation information for each point. Following the standard evaluation criterion, Area 5 is used as the test set and the rest of the areas are used as the training set.", + "bbox": [ + 75, + 704, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ScanNetV2 is an RGB-D video dataset. It contains 1513 point cloud samples from 707 individual indoor scenes for training and 100 unlabelled samples for testing. For the 3D semantic segmentation task, ScanNetV2 provides up to 40", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "categories of classification labels and selects 20 categories as classification targets to establish the segmentation task. Each point in the dataset contains coordinate, colour, and annotation information. In the experiments, we follow the official data split, dividing the 1513 training data into 1201 training samples and 312 validation samples, and evaluate methods on the test set.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 207, + 718, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segmentation Model. In our experiments, we use the PyTorch implementation of MinkowskiNet in ReDAL [41] as our backbone network with the stochastic gradient descent (SGD) optimizer. The architecture is trained on a single NVIDIA RTX A6000 GPU.", + "bbox": [ + 496, + 231, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Settings. For S3DIS, we set the batch size to 4, and train 60K steps per iteration. The initial learning rate is 0.1, and cosine annealing with a period of 5000 is used as the learning rate decay strategy. For ScanNetV2, we set the batch size to 4, and train 30K steps per iteration. The initial learning rate is 0.1, and the learning rate decay strategy is using polynomial decay with a power of 0.9. In the teacher-student model, the keep rate of EMA and the thresholds for the pseudo-label are set to 0.955 and 0.75, respectively. 5 iterations are performed in the active learning framework. In HMMU, the point cloud is downsampled at each scale with the sampling radius of $10\\mathrm{cm}$ , $50\\mathrm{cm}$ , and $100\\mathrm{cm}$ , respectively. In the FDS module, the radius of the local region is set to $20\\mathrm{cm}$ and the feature similarity threshold is set to 0.8.", + "bbox": [ + 496, + 306, + 892, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with Active Learning Methods", + "text_level": 1, + "bbox": [ + 498, + 544, + 867, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our active learning strategy, we remove the semi-supervised module in the comparisons with active learning methods. As current active learning methods for point cloud semantic segmentation are region-based, we compare our proposed method against them to verify the effectiveness of the point-based active learning method. The compared results are shown in Table 1. Compared with the region-based active learning methods, our point-based selection strategy greatly reduces the labelling cost, with the demand of achieving $90\\%$ performance of the fully-supervised baselines. With only $0.1\\%$ (about 13,000 points) of the downsampled points, our method allows for $90\\%$ of the performance of our fully supervised baseline, i.e. fully-supervised MinkowskiNet. This verifies our theory that the strong semantic similarity in local areas of point clouds results in redundant labelling and ineffective training in region-based active learning.", + "bbox": [ + 496, + 568, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, we compare the performance of our active learning strategy with existing region-based ones and four basic point-based selection strategies including random selection (Random), softmax least confidence (LC), softmax minimum margin uncertainty (MMU) and softmax entropy", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18102", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/afcbbc8cdb50b6641248a5dd8a6d2c752915d3fff7ea7c93e0693bfa193be02f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. The comparison of the percentage of labelled points required to achieve $90\\%$ accuracy of fully-supervised baseline on the S3DIS dataset against current region-based active learning methods. For a fair comparison, the semi-supervised module of our method is removed here which means only labelled data and selected points are used during training." + ], + "table_body": "
MethodsSegmentorsLabelled points
ReDAL [41]SPVCNN13%
ReDAL [41]MinkowskiNet15%
SSDR-AL [29]Randlanet11.7%
OursMinkowskiNet0.1%
", + "bbox": [ + 101, + 88, + 444, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(Entropy). The results are shown in Table 2. It can be seen that the performance of our method is increased by $4.6\\%$ compared with random point selection and surpasses the state-of-the-art region-based active learning method by $1\\%$ with significantly fewer labelled points (0.1% v.s. 11.7%).", + "bbox": [ + 75, + 260, + 467, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e40a29aedde36ad62f66b076ca716754c15395098d5f763322f7d6c174179d06.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. The comparison of mIoU between our active strategy without the semi-supervised module and different active learning methods on the S3DIS dataset." + ], + "table_body": "
TypeMethodsSegmentorslabelled pointsmIoU(%)
Region-basedReDALSPVCNN15%58.0
ReDALMinkowskiNet15%57.3
SSDR-ALRandlanet11.7%58.3
Point-basedRandomMinkowskiNet0.1%54.7
EntropyMinkowskiNet0.1%48.2
LCMinkowskiNet0.1%53.3
MMUMinkowskiNet0.1%55.5
OursMinkowskiNet0.1%59.3
", + "bbox": [ + 81, + 347, + 467, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with Weakly-supervised Methods", + "text_level": 1, + "bbox": [ + 76, + 513, + 465, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare the proposed method with the state-of-the-art weakly-supervised methods. Similar to [22, 48], we present state-of-the-art segmentation methods with different supervision settings. The comparison results on the datasets of S3DIS and ScannetV2 are presented in Table 3 and Table 4, respectively.", + "bbox": [ + 75, + 537, + 467, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "S3DIS. we perform experiments under three different annotation budgets, namely $0.7\\%$ , $0.07\\%$ , and $0.02\\%$ . For the sake of fairness, we follow the SQN [13] to calculate the annotation percentage which is defined as the ratio between the number of labelled data and all training and validation data. Each annotation budget is completed through five iterations. As shown in Table 3, our approach gains consistent improvements in the segmentation performance with labelled points increased from $0.02\\%$ to $0.43\\%$ . It is worth noting that, for the $0.7\\%$ annotation budget, the segmentation performance at the third iteration ( $0.43\\%$ annotations) has already surpassed the fully-supervised MinkowskiNet. This verifies that selecting the most representative points and making full use of the remaining unlabelled data can greatly benefit the performance of point cloud semantic segmentation. Our method also outperforms the state-of-the-art PSD, SQN and HybridCR under $1\\%$ labelled data by $2.2\\%$ , $2.0\\%$ and $0.4\\%$ , respectively. When the number of labelled", + "bbox": [ + 75, + 628, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/483da21a8fb555f6eed08b83abc17ef521275a0c7cd6e9f978c18251052f82ac.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Comparison with existing weakly-supervised methods on S3DIS Area-5. † represents the result of the baseline trained on our own device." + ], + "table_body": "
MethodsLabelled pointsArea 5
PointNet [27]100%41.1
PointCNN [23]100%57.3
KPConv [35]100%67.1
PointTransformer [54]100%70.4
CBL [32]100%69.4
MinkowskiNet†[7]100%64.5
PSD [51]1%63.5
SQN [13]1%63.7
HybridCR [22]1%65.3
Ours0.43%65.7
SQN [13]0.1%61.4
Ours0.07%62.3
PSD [51]0.03%48.2
HybridCR [22]0.03%51.5
MIL [48]0.02%51.4
GaIA [21]0.02%53.7
VIB [36]0.02%52.0
Ours0.02%55.9
", + "bbox": [ + 526, + 88, + 867, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "points is reduced to $0.07\\%$ , our method is also superior to the SQN with $0.1\\%$ labelled points and achieves $96.5\\%$ performance of the fully-supervised baseline (MinkowskiNet). In the case of the least labelling budget, we evaluate our method in the setting of $0.02\\%$ labelled points. It can be seen that our method also achieves the best performance, outperforming the state-of-the-art GaIA, MIL, HybridCR and PSD by $2.2\\%$ , $4.5\\%$ , $4.4\\%$ and $7.7\\%$ .", + "bbox": [ + 496, + 401, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also visualize the segmentation results of our method under different annotation budgets in Figure 3. The results show that with $0.43\\%$ labelled data, our approach can achieve comparable segmentation results with the fully supervised baseline. Benefiting from the points selected by our method, our segmentation results are even more promising in some cases, such as the bookcase and column.", + "bbox": [ + 496, + 522, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanNetV2. As illustrated in Table 4, we use the criteria of percentage and fixed number of labelled points as different annotation budgets to evaluate the methods, including $0.1\\%$ , 200pts, and 20pts. Similar to S3DIS, each annotation budget completes through five iterations, and the third iteration (120pts) is reported in the case of 200pts. We substantially beat SQN's performance by $11.3\\%$ using $0.1\\%$ of annotation. Compared against PSD, HybridCR and GaIA which are trained with $1\\%$ annotations, our method outperforms them with only $0.1\\%$ labels by $13.5\\%$ , $11.4\\%$ and $3\\%$ , respectively. In the case of the fixed number of annotation points, we use 200 points per scene (200pts) and 20 points per scene (20pts) as the annotation budget. In our experiment with 200pts, when the labelling points reach 120pts, our method has achieved comparable performance to the fully supervised baseline. This proves the effectiveness of our active learning strategy on the semi-supervised point cloud semantic segmentation. We also evaluate our", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/519bae80d21186449d66858167662bdc2249649c6d7be3ae3096547a9b6c43e3.jpg", + "image_caption": [], + "image_footnote": [ + "ceiling floor wall column window door chair table bookcase board clutter" + ], + "bbox": [ + 86, + 98, + 217, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/183656d7b8788898e35d22e22cdd9a74fd02c1352f28775cf9ab06dc1b955d38.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 86, + 164, + 217, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5dc9b1d54fa4d239bbf6abbe761d2d1c2fb693e98f1fc9e1761c6e2a9313f9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 98, + 349, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ca96052ef21033b5bf642718808a1e413f4ed71eb88a3e799ecaf296491f55b.jpg", + "image_caption": [ + "(b) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 220, + 165, + 349, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b235e45e92fd9982c177383025f787de549cda90cb8017b04d3d579b411d19f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 98, + 482, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83a618d9ecd74f84398f4cf2fa7c839dd822eed7194447c9ccd9510ee850ab64.jpg", + "image_caption": [ + "(c) Fully-supervised" + ], + "image_footnote": [], + "bbox": [ + 352, + 165, + 482, + 228 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d08a212ffa052a2ec22ce8079b72d6a63718128d7eae9d7ba533d9b01d9d3d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 98, + 614, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ca709480bdfc169dfd7ba13f54f61e2ca3853399073a71b161fa53f65cae22d.jpg", + "image_caption": [ + "(d) Ours $(0.02\\%)$" + ], + "image_footnote": [], + "bbox": [ + 486, + 165, + 614, + 228 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c08f3e6a3257c0bacafeeaa0aa429ffc927389ed668278863a286a1df7252c50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 98, + 750, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/408d87ec47a483138c9fab8dd551491e2101a5ab479db8bd08deb34de316f631.jpg", + "image_caption": [ + "(e) Ours $(0.07\\%)$" + ], + "image_footnote": [], + "bbox": [ + 617, + 165, + 748, + 228 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34b1278c1896af7e5465f113abefce71e5ec62ea716a7f2cd23246a491a1070b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 98, + 882, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b29481131af0468405bccacc07b28559c13c49f5562f63956cb55c2a60c363a5.jpg", + "image_caption": [ + "(f) Ours $(0.43\\%)$" + ], + "image_footnote": [], + "bbox": [ + 751, + 165, + 882, + 228 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/abdfd87f8f294378cac478211b2a023b21df1003270a1f4330b373b1dcad5ba9.jpg", + "image_caption": [ + "Figure 3. Visualization of segmentation results on the test set of S3DIS Area-5. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when $0.43\\%$ of the labelled data is adopted." + ], + "image_footnote": [], + "bbox": [ + 91, + 330, + 222, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53df65a7fb776c9145d1ddc88ee1bb02479172c43bee91222e147391a50628b4.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [ + "wall floor cabinet chair sofa table door bookshelf counter desk window sink curtain bed other furniture" + ], + "bbox": [ + 91, + 396, + 222, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f63f279ee846a0b46ec530267755cccac3adcc1e7793102621ac465942e87fc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 330, + 352, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13aae238025387ca61900aa445135d86cc7f2f58fc7419d01a356cc970bf45c5.jpg", + "image_caption": [ + "(b) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 225, + 395, + 351, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2dd32fc6eda9c971b434907dbba655470dc5bc3723380961c34f7da4c3986a37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 330, + 480, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97e3c0e02c4ebe39d150711c8eece58eeeb3ccdbba83e0ef3e0bedbbec7df215.jpg", + "image_caption": [ + "(c) Fully-supervised" + ], + "image_footnote": [], + "bbox": [ + 352, + 395, + 480, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48ea1038526256f503990dc79d97e53c89d14d850815d8fa7649097c5da86d4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 330, + 609, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85203a59f32ddd4ab34f2aaa47ce30a06d9081210229b3557734f99293668dc4.jpg", + "image_caption": [ + "(d) Ours (20pts)", + "Figure 4. Visualization of segmentation results on the validation set of ScannNetV2. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when 120 labelled points per scene are adopted." + ], + "image_footnote": [], + "bbox": [ + 483, + 395, + 609, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1560259c80f5ef5844010b57f5acff2dd73ee6c6979a897c1eb26235fb929a86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 330, + 741, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/212c0cd5954a32b40ed744be2dfcf8bb64288bb89f839a72f3decbc6710f6939.jpg", + "image_caption": [ + "(e) Ours $(0.1\\%)$" + ], + "image_footnote": [], + "bbox": [ + 616, + 395, + 741, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/973ac929ba247559605112f64640899df48c6c0536824148525b4b1087fdbf5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 330, + 872, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aadae06adf078e22a84d0a8cb68aa24c8ea6eb1b237dd138f5743815c6875625.jpg", + "image_caption": [ + "(f) Ours (120pts)" + ], + "image_footnote": [], + "bbox": [ + 743, + 395, + 872, + 458 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/66ec10cfb4e04ea035f96d28cd8d82e5c43b16244f6f4f1508520b6be78d58f2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. The comparison with existing weakly-supervised methods. ${}^{ \\dagger }$ is the result of the baseline trained on our device." + ], + "table_body": "
MethodsLabelled pointsvaltest
PointNet++ [28]100%-33.9
KPCov [35]100%-68.4
VMNet [16]100%-74.6
BPNet [15]100%-74.9
MinkowskiNet†[7]100%69.368.0
PSD [51]1%-54.7
HybridCR [22]1%56.956.8
GaIA [21]1%-65.2
SQN [13]0.1%-56.9
Ours0.1%69.968.2
SQN [13]200pts-59.8
CSC [12]200pts68.266.5
GaIA [21]200pts-68.5
VIBUS [36]200pts69.669.1
Ours120pts70.269.4
SQN [13]20pts-48.6
CSC [12]20pts55.553.1
MIL [48]20pts57.854.4
VIBUS [36]20pts-58.6
GaIA [21]20pts-63.8
Ours20pts62.262.5
", + "bbox": [ + 101, + 558, + 444, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method on the least annotation setting. In the case of 20pts, our result is much higher than SQN, CSC, MIL,", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and VIBUS, showing that our method effectively selects the most valuable points. However, our result is slightly lower than GaIA, as the performance of our feature extraction backbone is limited given only 20pts (the proportion is less than $0.02\\%$ of ScanNetV2).", + "bbox": [ + 496, + 561, + 890, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 4, we visualize the segmentation results of our method under different annotation budgets on ScanNetV2. It is observed that our method can achieve comparable or even better segmentation results with minor training budgets compared to our fully-supervised baseline.", + "bbox": [ + 496, + 638, + 890, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 724, + 663, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of each module in our method, we conduct the following ablation studies on S3DIS dataset. Following [13], all the ablated networks are trained using $0.1\\%$ labelled points on Areas 1, 2, 3, 4, 6 and tested on Area 5. The number of iterations is set to 5 and $0.02\\%$ data is selected in each iteration. To clearly see the effect of active learning, we present the results as shown in Figure 5. We take random sampling as the basic strategy for active selection where points are randomly selected in each iteration. After we replace the random sam", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18104", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd66a2ae14a1b6265f3ab4f1741216d10c5eb4cc6e8d8f5c861822a7779e4565.jpg", + "image_caption": [ + "Figure 5. The comparison of performance improvements for different combinations of modules." + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 428, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pling with MMU for active sampling, the mIoU is slightly improved by $1.5\\%$ . After utilising HMMU to select points, we acquire an improvement of around $3.0\\%$ compared with MMU in the last iteration. The FDS module is also added to further improve the performance. HMMU+FDS achieves an extra $2.9\\%$ mIoU improvement compared to HMMU. More impressively, merely relying on the active learning modules including HMMU and FDS, our approach already outperforms the random sampling baseline by $8\\%$ and surpasses $90\\%$ performance of full-supervision with only $0.1\\%$ labelling data. The effectiveness of the semi-supervised module is also demonstrated by the experimental results of HMMU+FDS+TS. Obviously, adding the semi-supervised teacher-student module, the segmentation result is further improved by $5\\%$ compared to HMMU + FDS. The proposed method also reaches $95\\%$ of the fully-supervised setting with only $0.1\\%$ labelled data.", + "bbox": [ + 75, + 356, + 468, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 5, we show the effectiveness of different modules by quantitative comparison. We compare different compositions of our module to verify their effectiveness. Shown in the first row, for the base case where unlabelled points are selected by MMU in each iteration, the mIoU is only $55.51\\%$ . After using the HMMU module, the segmentation results are improved to $57.65\\%$ . If we only apply FDS to the MMU selection for each iteration, a slight improvement of $1.17\\%$ is obtained compared with the based case. When the active learning strategy is removed, as shown in the fourth row in Table 5, the semi-supervised segmentation method only achieves the performance of $57.01\\%$ . Point cloud semantic segmentation only using the active learning technique can reach the mIoU of $59.29\\%$ . If we combine semi-supervised and active learning, as shown in the last row, the performance can be improved to $62.26\\%$ .", + "bbox": [ + 75, + 613, + 470, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 6, we show the ablation study of each layer in the HMMU module. PL represents the layer that calculates", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6a8f10f635a846d07a7bed5b9268dc39ad56b8ca38ef1267f2ae436fc02f6a40.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ComponentsmIoU(%)
HMMUFDSTS
base.55.51
✓57.65
✓56.68
✓57.01
✓✓59.29
✓✓✓62.26
", + "bbox": [ + 550, + 87, + 821, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the point-level minimum margin uncertainty score. VL is the layer which calculates the voxel-level contextual uncertainty score. $v_{r}$ and $\\omega$ are the voxel-radius and weight for each VL layer. It is shown that the combination of one layer of point-level uncertainty and three layers of voxel-level uncertainty can achieve the best performance. We also show the ablation study on the hyperparameters of the FDS module in Table 7. We achieve the best segmentation result given that $r$ and $\\tau$ are set to $20\\mathrm{cm}$ and 0.8 respectively. Thus, we adopt this combination in our experiments.", + "bbox": [ + 496, + 222, + 890, + 373 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/78f310caedf015c35e6f4429cabe4cd6a995fddf7f8c93684f7c06e93cc06526.jpg", + "table_caption": [ + "Table 5. Ablation study of different components." + ], + "table_footnote": [], + "table_body": "
LayerHMMUiter 1iter 2iter 3iter 4iter 5
ωvr(cm)
PL--37.8350.3754.3256.1856.68
PL + VL1{0.1}{10}37.8352.7154.8456.5257.03
PL + VL2{0.01}{50}37.8351.4555.8157.2257.43
PL + VL3{0.001}{100}37.8351.0655.6257.3658.03
PL + VL1,2{0.1, 0.01}{10, 50}37.8351.5655.2856.6658.21
PL + VL1,2,3{0.1, 0.01, 0.001}{10, 50, 100}37.8352.0256.0258.4959.29
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.001}{10, 50, 100, 200}37.8351.5656.3257.6157.71
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.0001}{10, 50, 100, 200}37.8351.6956.2258.5359.19
", + "bbox": [ + 503, + 383, + 890, + 468 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47222fd4f8b5afc9dd0d13a9dd6df501212f7f4c524964762ffeb0b353f979b0.jpg", + "table_caption": [ + "Table 6. Ablation study of HMMU layers." + ], + "table_footnote": [ + "Table 7. Ablation study of the hyperparameters in FDS module." + ], + "table_body": "
FDSiteration1iteration2iteration3iteration4iteration5
r(cm)τ
200.837.8352.0256.0258.4959.29
200.937.8351.2754.9056.7958.79
400.837.8351.3555.7157.0957.79
400.937.8351.6555.2657.4559.13
500.9537.8350.8755.7356.4957.82
", + "bbox": [ + 503, + 492, + 893, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 603, + 617, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study active learning for 3D point cloud semantic segmentation. In comparison with previous region-based methods, we propose a hierarchical point-based active learning strategy consisting of two new designs, HMMU and FDS. HMMU serves as an effective way to measure the uncertainty or importance of labelling, while FDS enables us to select the most valuable points by considering the feature similarity and spatial distribution. To better use the unlabelled points during network learning, we also introduce a teacher-student structure to generate pseudo-labels. Extensive experiments are conducted on S3DIS and ScanNet datasets to demonstrate the effectiveness of our method. In the future, a more advanced semi-supervised architecture can be adopted for generating more reliable pseudo-labels to further facilitate active learning based point cloud semantic segmentation tasks.", + "bbox": [ + 496, + 628, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62206033,", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "No.62036007, No.62221005), and the Natural Science Foundation of Chongqing (No. cstc2020jcyj-msxmX0855, No. cstc2021ycjh-bgzxm0339).", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 148, + 173, + 165 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rashid Abbasi, Ali Kashif Bashir, Hasan J Alyamani, Farhan Amin, Jaehyeok Doh, and Jianwen Chen. Lidar point cloud compression, processing and learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 24(1):962-979, 2022.", + "[2] Eren Erdal Aksoy, Saimir Baci, and Selcuk Cavdar. Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In 2020 IEEE intelligent vehicles symposium (IV), pages 926-932. IEEE, 2020.", + "[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016.", + "[4] Alexandre Boulch, Bertrand Le Saux, Nicolas Audebert, et al. Unstructured point cloud semantic labeling using deep segmentation networks. 3dor@eurographics, 3:17-24, 2017.", + "[5] Mingmei Cheng, Le Hui, Jin Xie, and Jian Yang. Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1140-1147, 2021.", + "[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021.", + "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019.", + "[8] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5-7, 2020, Proceedings, Part II 15, pages 207-222. Springer, 2020.", + "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017.", + "[10] Shuang Deng, Qiulei Dong, Bo Liu, and Zhanyi Hu. Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In 2022 International Conference on Robotics and Automation (ICRA), pages 9214-9220. IEEE, 2022.", + "[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the" + ], + "bbox": [ + 78, + 172, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018.", + "[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021.", + "[13] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 600-619. Springer, 2022.", + "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020.", + "[15] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021.", + "[16] Zeyu Hu, Xuyang Bai, Jiaxiang Shang, Runze Zhang, Jiayu Dong, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15488-15498, 2021.", + "[17] Ankang Ji, Yunxiang Zhou, Limao Zhang, Robert LK Tiong, and Xiaolong Xue. Semi-supervised learning-based point cloud network for segmentation of 3d tunnel scenes. Automation in Construction, 146:104668, 2023.", + "[18] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6423-6432, 2021.", + "[19] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020.", + "[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022.", + "[21] Min Seok Lee, Seok Woo Yang, and Sung Won Han. Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 582-591, 2023.", + "[22] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18106", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "contrastive regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14930-14939, 2022.", + "[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018.", + "[24] Y Lin, G Vosselman, Y Cao, and MY Yang. Efficient training of semantic point cloud segmentation via active learning. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2:243-250, 2020.", + "[25] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021.", + "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4213-4220. IEEE, 2019.", + "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017.", + "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017.", + "[29] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022.", + "[30] Xian Shi, Xun Xu, Ke Chen, Lile Cai, Chuan Sheng Foo, and Kui Jia. Label-efficient point cloud semantic segmentation: An active learning approach. arXiv preprint arXiv:2101.06931, 2021.", + "[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII, pages 685–702. Springer, 2020.", + "[32] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8489-8499, 2022.", + "[33] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 2017 international conference on 3D vision (3DV), pages 537-547. IEEE, 2017.", + "[34] Hugues Thomas, Ben Agro, Mona Gridseth, Jian Zhang, and Timothy D Barfoot. Self-supervised learning of lidar segmentation for autonomous indoor navigation. In 2021" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE International Conference on Robotics and Automation (ICRA), pages 14047-14053. IEEE, 2021.", + "[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019.", + "[36] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS Journal of Photogrammetry and Remote Sensing, 194:302-318, 2022.", + "[37] Haiyan Wang, Xuejian Rong, Liang Yang, Jinglun Feng, Jizhong Xiao, and Yingli Tian. Weakly supervised semantic segmentation in 3d graph-structured point clouds of wild scenes. arXiv preprint arXiv:2004.12498, 2020.", + "[38] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10296-10305, 2019.", + "[39] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1887-1893. IEEE, 2018.", + "[40] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019.", + "[41] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15510-15519, 2021.", + "[42] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019.", + "[43] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020.", + "[44] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 1-19. Springer, 2020.", + "[45] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18107", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision, pages 638-656. Springer, 2022.", + "[46] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020.", + "[47] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021.", + "[48] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022.", + "[49] Ping-Chung Yu, Cheng Sun, and Min Sun. Data efficient 3d learner via knowledge transferred from 2d model. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 182-198. Springer, 2022.", + "[50] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3421-3429, 2021.", + "[51] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15520-15528, 2021.", + "[52] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020.", + "[53] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021.", + "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 726 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18108", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_model.json b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_model.json new file mode 100644 index 0000000000000000000000000000000000000000..367000c8ba0ba055d438e67443f2c96e1861c147 --- /dev/null +++ b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_model.json @@ -0,0 +1,2653 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.862, + 0.176 + ], + "angle": 0, + "content": "Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.203, + 0.796, + 0.294 + ], + "angle": 0, + "content": "Zongyi Xu\\(^{1\\dagger}\\) Bo Yuan\\(^{1\\dagger}\\) Shanshan Zhao\\(^{2\\dagger}\\) Qianni Zhang\\(^{3}\\) Xinbo Gao\\(^{1*}\\) \n\\(^{1}\\)Chongqing University of Posts and Telecommunications, China \n\\(^{2}\\)JD Explore Academy, China \\(^{3}\\)Queen Mary University of London, UK \n{xuzy, qaoxb}@cqupt.edu.cn s210201127@stu.cqupt.edu.cn \nsshan.zhao00@gmail.com qianni.zhang@qmul.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.766 + ], + "angle": 0, + "content": "Impressive performance on point cloud semantic segmentation has been achieved by fully-supervised methods with large amounts of labelled data. As it is labour-intensive to acquire large-scale point cloud data with pointwise labels, many attempts have been made to explore learning 3D point cloud segmentation with limited annotations. Active learning is one of the effective strategies to achieve this purpose but is still under-explored. The most recent methods of this kind measure the uncertainty of each pre-divided region for manual labelling but they suffer from redundant information and require additional efforts for region division. This paper aims at addressing this issue by developing a hierarchical point-based active learning strategy. Specifically, we measure the uncertainty for each point by a hierarchical minimum margin uncertainty module which considers the contextual information at multiple levels. Then, a feature-distance suppression strategy is designed to select important and representative points for manual labelling. Besides, to better exploit the unlabelled data, we build a semi-supervised segmentation framework based on our active strategy. Extensive experiments on the S3DIS and ScanNetV2 datasets demonstrate that the proposed framework achieves \\(96.5\\%\\) and \\(100\\%\\) performance of fully-supervised baseline with only \\(0.07\\%\\) and \\(0.1\\%\\) training data, respectively, outperforming the state-of-the-art weakly-supervised and active learning methods. The code will be available at https://github.com/SmiletoE/HPAL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.794, + 0.21, + 0.81 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.819, + 0.47, + 0.881 + ], + "angle": 0, + "content": "Point cloud semantic segmentation aims to assign a category label for each 3D point, which can be applied to various scenarios, such as robotics [34], autonomous driving [1], and augmented reality [20]. Recently, deep learning" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.328, + 0.892, + 0.405 + ], + "angle": 0, + "content": "based methods [38, 52] have achieved impressive performance. These high-performing methods usually rely on large amounts of data with point-wise labels. However, acquiring such dense labels for 3D point clouds is extremely tedious and costly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.893, + 0.576 + ], + "angle": 0, + "content": "To relieve the labour and cost of annotation, many recent methods explore semi-supervised learning to learn 3D segmentation models with limited labelled points that are usually sampled randomly. These methods attempt to employ effective strategies to propagate the label information to the unlabelled points [5, 17, 10, 13]. Although these semi-supervised methods can greatly decrease labelling costs, their performance might be limited due to various factors. Among them, the most important reason is that some of the randomly selected points are in fact redundant while some really important points might be omitted." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.837 + ], + "angle": 0, + "content": "As an alternative learning strategy, active learning recently is studied to alleviate such limitations for 3D segmentation. Lin et al. [24] divide the whole point clouds into segments and each segment is utilised as the basic query unit for sample selection. ReDAL [41] proposes to select those informative and diverse sub-scene regions for label acquisition. The entropy, colour discontinuity, and structural complexity are used to measure the information of sub-scene regions. Following it, SSDR-AL [29] groups the original point clouds into superpoints and incrementally selects the most informative regions for annotation. However, these methods usually rely on pre-dividing the point cloud into multiple regions, and the region division strategy might negatively impact the performance. Moreover, as point clouds present strong semantic similarity in local areas, selecting all the points in the local region results in a redundant labelling budget." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Motivated by the above analysis, this research aims at enhancing the 3D segmentation performance with limited labelled data in the active learning framework. Specifically, in comparison with previous region-based methods, we mea" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.333, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. †Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.485 + ], + "angle": 0, + "content": "sure the uncertainty or importance of labelling for each point to avoid the additional efforts and potential bias of region division. However, considering each point individually may introduce extra noise to the training process, leading to unreliable uncertainty values that cannot reflect the real importance of each point. To alleviate this issue, we design a hierarchical minimum margin uncertainty (HMMU) measurement module by considering the local contextual information. The HMMU involves the uncertainty of the point and its neighbourhood. In this way, the acquired score can more effectively represent the importance of each point label. Based on the uncertainty scores acquired by HMMU, we can directly choose the Top- \\( K \\) points for labelling. However, as the point cloud presents semantic similarity in the local region, some redundant uncertain points tend to be distributed densely. To further reduce the labelling cost, we propose a feature-distance suppression module (FDS) to remove those uncertain points with similar features in the neighbourhood. By exploiting the devised HMMU measurement and FDS for redundant point filtering, our hierarchical point-based active learning strategy is able to select more valuable points for labelling and thus reduce labelling costs. Furthermore, inspired by previous semi-supervised methods that exploit the unlabelled points, we utilise a simple teacher-student scheme to enhance the supervision signal by assigning a pseudo label for the unlabelled points." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.485, + 0.456, + 0.499 + ], + "angle": 0, + "content": "The main contributions of this research are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.505, + 0.468, + 0.594 + ], + "angle": 0, + "content": "- We propose a point-based active learning method for semi-supervised point cloud semantic segmentation, which surpasses current semi-supervised or active learning methods and achieves comparable performance with full supervision methods relying on scarce annotations." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.603, + 0.468, + 0.662 + ], + "angle": 0, + "content": "- We propose a novel hierarchical minimum margin uncertainty module to measure the uncertainty for each point by progressively perceiving the contextual information at increasing scales." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.671, + 0.468, + 0.716 + ], + "angle": 0, + "content": "- We propose a feature-distance suppression module to remove redundant points with similar features in the neighbourhood and further minimise the labelling cost." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.505, + 0.468, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.22, + 0.748 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.32, + 0.773 + ], + "angle": 0, + "content": "2.1. 3D Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In recent years, a large number of deep learning-based point cloud semantic segmentation methods have been developed, which can be roughly divided into the following categories: 1) 2D projection-based methods [4, 19, 2, 8, 26, 39, 40, 44], which project the 3D point cloud onto 2D images through a variety of viewpoints including multiview, bird's eye view, spherical projection etc. Projection-based methods can make full use of those well-designed 2D" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "convolution networks for 3D scene parsing, but obviously, the projection causes the loss of 3D geometric information and thus limits the performance. 2) voxel-based methods [6, 11, 31, 47, 7, 33]. Point cloud voxelization regularizes point clouds with uneven densities, making it possible to extend regular 2D convolutions to 3D scenes. At an early stage, voxel-based methods usually suffer from the explosion of computation and the loss of information. Fortunately, some sparse convolution methods [11, 31, 7] have been proposed over time to greatly ease the computational costs. 3) point-based methods [14, 27, 28, 35, 23, 42]. These methods directly take the original uneven point cloud as input and use MLPs, 3D point convolution, and other more flexible ways to process the point clouds, which preserve the original 3D information. However, their high performance still relies on a large number of point-level labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.344, + 0.892, + 0.361 + ], + "angle": 0, + "content": "2.2. Weakly-supervised 3D Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.745 + ], + "angle": 0, + "content": "We collectively refer to all methods training with sparse annotation as weakly supervised methods, which is a well-explored field in point cloud segmentation. These methods are usually based on a small set of randomly labelled data and then exploit more information from the unlabelled data by leveraging techniques such as transfer learning, contrastive learning, consistency regularization, pseudo-label, etc. Based on the transfer learning technologies, pre-training is widely exploited in 2D images, and has recently been gradually applied to point clouds [50, 49, 43, 53, 37, 45]. These methods learn pre-trained knowledge from 3D point clouds or 2D images and then apply it to the target dataset, thus achieving better training results with limited labelled data. Some other works train the model by using consistency regularization [46, 51, 22, 48] and pseudo-label [13, 36, 5, 25, 50, 22]. SQN [13] is one of these works which designs a novel semantic query network to efficiently obtain the pseudo-labels. To further improve weakly supervised training, some approaches also use contrastive learning [22, 25, 12, 18, 43], which usually combine contrastive loss with pre-training, consistency regularization, and pseudo-labels to better exploit the supervision information. Most weakly-supervised methods explore information from the labelled data, so how to select the most informative data to label is a question of crucial importance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.892, + 0.772 + ], + "angle": 0, + "content": "2.3. Active Learning on 3D Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Similar to weakly-supervised learning, active learning also aims to train deep models with limited labelled data. The main difference is that weakly-supervised learning mainly focuses on enhancing supervision by exploiting available label information while active learning focuses on the selection of valuable information for labelling." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Lin et al. [24] make the first attempt for active learning on 3D semantic segmentation using segments as the ba" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.852, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.895, + 0.414 + ], + "angle": 0, + "content": "Figure 1. The proposed hierarchical point-based active learning framework for semi-supervised point cloud semantic segmentation. (a) A segmentation network using the teacher-student model is trained with the initial labelled and unlabelled training data. The teacher is updated from the student by the exponential moving averaging (EMA) and the student network is obtained as our target segmentation model. (b) Those representative candidate points are measured by the trained model using the proposed hierarchical minimum margin uncertainty module (HMMU). (c) Feature-distance Suppression module (FDS) is then applied to remove those redundant points in the neighbourhood and select the final to-be-annotated points. The selected valuable points are annotated and added to the labelled training set to improve the segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.683 + ], + "angle": 0, + "content": "sic query unit, and propose a segment entropy strategy to measure the informativeness of each segment. To improve the efficiency of active learning, superpoints are introduced as the most common basic query units [30, 41]. Considering the limitation of uncertainty, Shi et al. [30] propose to design active acquisition functions from multiple perspectives, including feature diversity, shape diversity, and entropy. ReDAL [41] combines entropy with the geometric structure and colour information and uses feature clustering to prevent selection redundancy. In addition to the entropy-based method, SSDR-AL [29] improves MMU based on super-point by considering the principal class within superpoints. To date, all active learning methods for point cloud semantic segmentation are region based. However, the pre-region division can greatly affect the active learning performance and unnecessary annotation in the semantic-similar areas is prone to occur." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.188, + 0.718 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.726, + 0.188, + 0.74 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In the following, we describe our proposed hierarchical point-based active learning approach for semi-supervised point cloud semantic segmentation in detail. As shown in Figure 1, the initial labelled and unlabelled training data are utilised to train the point cloud semantic segmentation neural network in a semi-supervised way. A teacher-student model is adopted. With the trained model, the valuable unlabelled points that are able to improve the segmentation performance most can be measured and selected in an active learning manner. The hierarchical minimum margin uncer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.637 + ], + "angle": 0, + "content": "tainty module is proposed to measure the point-wise uncertainty by hierarchically increasing contextual ranges, which enables progressively capturing the local context at different scales. Then, the neighbouring points with similar features are removed by the feature distance suppression module, thereby further reducing the annotation redundancy. The remaining points with high uncertainty scores as well as sparse distributions are regarded as the candidates to be annotated. In this manner, the most valuable and representative unlabelled points are selected to expand the labelled set, with which the semi-supervised point cloud semantic segmentation neural network can be further enhanced. In the following, we describe each component of the proposed framework in detail." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.649, + 0.892, + 0.68 + ], + "angle": 0, + "content": "3.2. Hierarchical Minimum Margin Uncertainty Measurement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.884 + ], + "angle": 0, + "content": "In the proposed method, we aim to select and label the most valuable points that can bring maximum performance improvements. Intuitively, identifying and using the most informative labelled data for training is the key to obtaining a model with good quality, and consequently accurate segmentation prediction. Merely considering individual points without their surrounding context information cannot reflect the real importance of the point. Thus, we design the hierarchical minimum margin uncertainty measurement module to calculate the uncertainty score for each point by grouping points at multiple scales and progressively perceiving context information in a broader range along the hierarchy for the unlabelled point." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 1, the prediction for each point gen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "erated with the currently trained model is input into the HMMU module. We first calculate the point-level minimum margin uncertainty score \\( U_{x} \\) for each unlabelled point with Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.166, + 0.469, + 0.184 + ], + "angle": 0, + "content": "\\[\nU _ {x} = h \\left(x ^ {u}; p _ {1} \\left(x ^ {u}\\right)\\right) - h \\left(x ^ {u}; p _ {2} \\left(x ^ {u}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.197, + 0.469, + 0.243 + ], + "angle": 0, + "content": "where \\( x^{u} \\) is the candidate point that is chosen to be labelled; \\( p_1(x^u) \\) and \\( p_2(x^u) \\) are the highest and second-highest prediction of \\( x^{u} \\) under the segmentation predictor \\( h(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.469, + 0.349 + ], + "angle": 0, + "content": "Then we group a wider range of neighbours through downsampling. The softmax prediction of each downsampled point is obtained by averaging the softmax prediction of its neighbouring points in the original point cloud. Each softmax label of the downsampled point represents the prediction distribution of a local region, which is denoted by \\( S_{R} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.426 + ], + "angle": 0, + "content": "We perform \\(N\\) levels of voxel downsampling. The local context information of the to-be-annotated point \\(x^{u}\\) at the \\(i_{th}\\) level of downsampling, denoted by \\(S_R^i (x^u)\\), can be represented as the average of predictions of the grouped points in the following:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.44, + 0.469, + 0.482 + ], + "angle": 0, + "content": "\\[\nS _ {R} ^ {i} \\left(x ^ {u}\\right) = \\frac {1}{K} \\sum_ {j = 1} ^ {K} p \\left(x _ {j} ^ {u}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.495, + 0.469, + 0.542 + ], + "angle": 0, + "content": "where \\( x_{j}^{u} \\) is the \\( j_{th} \\) neighbour for \\( x^{u} \\) in the original point cloud; \\( K \\) is the number of neighbours within the predefined voxel radius \\( v_{r} \\); and \\( p(\\cdot) \\) is the prediction probability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.469, + 0.587 + ], + "angle": 0, + "content": "For each level of downsampling, the voxel-level contextual uncertainty scores \\( U_{R}^{i} \\) for the unlabelled point \\( x^{u} \\) can be obtained with:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.601, + 0.469, + 0.62 + ], + "angle": 0, + "content": "\\[\nU _ {R} ^ {i} = h \\left(x ^ {u}; S _ {R 1} ^ {i} \\left(x ^ {u}\\right)\\right) - h \\left(x ^ {u}; S _ {R 2} ^ {i} \\left(x ^ {u}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.632, + 0.469, + 0.678 + ], + "angle": 0, + "content": "where \\( S_{R1}^{i}(x^{u}) \\) and \\( S_{R2}^{i}(x^{u}) \\) are the highest and the second highest average prediction of the grouped points of \\( x^{u} \\) from the segmentor \\( h(\\cdot) \\) in the \\( i_{th} \\) downsampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.469, + 0.725 + ], + "angle": 0, + "content": "Finally, for each unannotated point \\( x^u \\), we integrate the point-level and the voxel-level contextual uncertainty scores to get the final uncertainty \\( v^u \\) using Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.738, + 0.469, + 0.78 + ], + "angle": 0, + "content": "\\[\nv ^ {u} = U _ {x} + \\sum_ {i = 1} ^ {N} \\omega^ {i} \\times U _ {R} ^ {i}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Here, \\(\\omega^i\\in \\{0.1,0.01,0.001\\}\\) is the hyperparameter representing the fusion weight of the contextual uncertainty at the \\(i_{th}\\) level of downsampling and three levels of downsampling is performed in the experiment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.469, + 0.901 + ], + "angle": 0, + "content": "With HMMU, the context information is fused to significantly improve performance by merely annotating points rather than regions." + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.089, + 0.696, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.175, + 0.617, + 0.185 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.871, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.175, + 0.792, + 0.185 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.186, + 0.696, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.273, + 0.617, + 0.282 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.186, + 0.871, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.273, + 0.792, + 0.282 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.282, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Visualization of selected points. (a) and (c) show the points selected using the Top- \\(K\\) approach (blue dots); (b) and (d) present points chosen by FDS (orange dots). FDS enables a more spread-out point selection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.348, + 0.891, + 0.378 + ], + "angle": 0, + "content": "3.3. Feature-distance Suppression based Point Selection" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.507 + ], + "angle": 0, + "content": "With the point-wise uncertainty score \\( v^u \\), we can select the Top-\\( K \\) points with the highest uncertainty to label. However, as shown in Figure 2, such a point selection way might make points concentrate in local areas, causing label redundancy. Thus, we propose a feature-distance suppression (FDS) module to ensure the selected points retain a spread-out distribution in the space, thus offering a more effective overall representation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Given a distance suppression radius \\( r \\) and a feature similarity threshold \\( \\tau \\), for a point \\( x_{i} \\) to be selected, we first determine whether there are points within its radius \\( r \\) that have been already selected. A judgment set \\( D^{i} \\) is constructed for the candidate point \\( x_{i} \\) in the current point selection iteration." + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.595, + 0.892, + 0.619 + ], + "angle": 0, + "content": "\\[\nD ^ {i} = D ^ {i} \\cup \\{x _ {j} \\}, i f \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.615, + 0.783, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\forall x _ {j} i s s e l e c t e d: d _ {i j} < r,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.636, + 0.891, + 0.696 + ], + "angle": 0, + "content": "where \\( x_{j} \\) is the neighbour of \\( x_{i} \\) within the radius of \\( r \\); \\( d_{ij} \\) is the Euclidean distance between \\( x_{i} \\) and \\( x_{j} \\); and \\( D^{i} \\) represents if there are other candidate points in the neighbourhood and \\( D^{i} \\) is initialized to be empty." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.697, + 0.892, + 0.788 + ], + "angle": 0, + "content": "If there is no neighbour of \\( x_{i} \\) that is selected as a candidate, i.e. the judgment set \\( D^{i} \\) is empty, \\( x_{i} \\) is regarded as the valuable point that can represent the local area for labelling. Conversely, if the neighbouring point \\( x_{j} \\) is already selected as the uncertainty point, cosine similarity \\( Sim_{ij} \\) between the features of \\( x_{i} \\) and \\( x_{j} \\) is then computed with Eq. 6." + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.798, + 0.891, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} \\left(x _ {i}, x _ {j}\\right) = \\frac {\\mathbf {f} _ {i} \\cdot \\mathbf {f} _ {j}}{\\left| \\left| \\mathbf {f} _ {i} \\right| \\right| \\cdot \\left| \\left| \\mathbf {f} _ {j} \\right| \\right|}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.891, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_i\\) and \\(\\mathbf{f}_j\\) are the features of \\(x_{i}\\) and \\(x_{j}\\) extracted with the trained student segmentation network." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "If there exists an \\( x_{j} \\) in \\( D^{i} \\) for which \\( Sim_{ij} > \\tau \\), \\( x_{i} \\) is regarded as a redundant point with respect to the labelled" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.335, + 0.108 + ], + "angle": 0, + "content": "data and \\( x_{i} \\) is excluded from labelling." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.254, + 0.132 + ], + "angle": 0, + "content": "3.4. Training objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.473, + 0.501 + ], + "angle": 0, + "content": "To train a network with partly annotated data, a usual approach is to apply the cross-entropy loss to the labelled points and ignore the unlabelled ones, which causes inadequate supervision for model learning. Inspired by previous semi-supervised 3D point cloud segmentation methods, in the proposed method we employ a teacher-student framework to exploit unlabelled points and provide additional supervision. Specifically, we construct two segmentation networks using MinkowskiNet [7], denoted as the teacher model \\(M_{t}\\) and student model \\(M_{s}\\), respectively. A consistency constraint between the teacher model and the student model is exploited to learn knowledge from unlabelled data. Pseudo-labels of unlabelled data are generated using the teacher model. During training, the original sample is input into the teacher network while its augmented counterparts are input into the student. For those labelled points, we apply the standard cross-entropy loss. For the unlabelled ones, we generate pseudo-labels from the output of the teacher and also compute the cross-entropy loss. The losses in both labelled and unlabelled points are then used to optimize the student network by gradient descent. After updating the student model at each step, Exponential Moving Average (EMA) is used to transfer the parameters of the student model to the teacher model:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.51, + 0.47, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} ^ {j} = \\alpha \\theta_ {t} ^ {j - 1} + (1 - \\alpha) \\theta_ {s} ^ {j}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.6 + ], + "angle": 0, + "content": "where \\(\\theta_{t}\\) and \\(\\theta_{s}\\) are the parameters of the teacher and student networks, respectively, and \\(j\\) denotes the \\(j\\)-th training step. \\(\\alpha\\) is the hyper-parameter to determine the speed of parameter transmission, which is generally close to 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.611, + 0.21, + 0.629 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.636, + 0.181, + 0.651 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.704 + ], + "angle": 0, + "content": "We evaluate the proposed approach on two publicly available benchmark datasets, S3DIS [3] and ScanNetV2 [9], to demonstrate its superior performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.84 + ], + "angle": 0, + "content": "S3DIS (Stanford Large-Scale 3D Indoor Space Dataset) is a large-scale indoor point cloud dataset captured by Matter-port scanners which mainly dedicate to 3D segmentation tasks. It consists of 6 different large-scale indoor areas with a total of 271 rooms, each room is a separate point cloud sample that includes coordinate, colour, and annotation information for each point. Following the standard evaluation criterion, Area 5 is used as the test set and the rest of the areas are used as the training set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "ScanNetV2 is an RGB-D video dataset. It contains 1513 point cloud samples from 707 individual indoor scenes for training and 100 unlabelled samples for testing. For the 3D semantic segmentation task, ScanNetV2 provides up to 40" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "categories of classification labels and selects 20 categories as classification targets to establish the segmentation task. Each point in the dataset contains coordinate, colour, and annotation information. In the experiments, we follow the official data split, dividing the 1513 training data into 1201 training samples and 312 validation samples, and evaluate methods on the test set." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.208, + 0.719, + 0.224 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Segmentation Model. In our experiments, we use the PyTorch implementation of MinkowskiNet in ReDAL [41] as our backbone network with the stochastic gradient descent (SGD) optimizer. The architecture is trained on a single NVIDIA RTX A6000 GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Training Settings. For S3DIS, we set the batch size to 4, and train 60K steps per iteration. The initial learning rate is 0.1, and cosine annealing with a period of 5000 is used as the learning rate decay strategy. For ScanNetV2, we set the batch size to 4, and train 30K steps per iteration. The initial learning rate is 0.1, and the learning rate decay strategy is using polynomial decay with a power of 0.9. In the teacher-student model, the keep rate of EMA and the thresholds for the pseudo-label are set to 0.955 and 0.75, respectively. 5 iterations are performed in the active learning framework. In HMMU, the point cloud is downsampled at each scale with the sampling radius of \\(10\\mathrm{cm}\\), \\(50\\mathrm{cm}\\), and \\(100\\mathrm{cm}\\), respectively. In the FDS module, the radius of the local region is set to \\(20\\mathrm{cm}\\) and the feature similarity threshold is set to 0.8." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.869, + 0.561 + ], + "angle": 0, + "content": "4.3. Comparison with Active Learning Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.825 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our active learning strategy, we remove the semi-supervised module in the comparisons with active learning methods. As current active learning methods for point cloud semantic segmentation are region-based, we compare our proposed method against them to verify the effectiveness of the point-based active learning method. The compared results are shown in Table 1. Compared with the region-based active learning methods, our point-based selection strategy greatly reduces the labelling cost, with the demand of achieving \\(90\\%\\) performance of the fully-supervised baselines. With only \\(0.1\\%\\) (about 13,000 points) of the downsampled points, our method allows for \\(90\\%\\) of the performance of our fully supervised baseline, i.e. fully-supervised MinkowskiNet. This verifies our theory that the strong semantic similarity in local areas of point clouds results in redundant labelling and ineffective training in region-based active learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Furthermore, we compare the performance of our active learning strategy with existing region-based ones and four basic point-based selection strategies including random selection (Random), softmax least confidence (LC), softmax minimum margin uncertainty (MMU) and softmax entropy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.445, + 0.164 + ], + "angle": 0, + "content": "
MethodsSegmentorsLabelled points
ReDAL [41]SPVCNN13%
ReDAL [41]MinkowskiNet15%
SSDR-AL [29]Randlanet11.7%
OursMinkowskiNet0.1%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.166, + 0.47, + 0.25 + ], + "angle": 0, + "content": "Table 1. The comparison of the percentage of labelled points required to achieve \\(90\\%\\) accuracy of fully-supervised baseline on the S3DIS dataset against current region-based active learning methods. For a fair comparison, the semi-supervised module of our method is removed here which means only labelled data and selected points are used during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.468, + 0.337 + ], + "angle": 0, + "content": "(Entropy). The results are shown in Table 2. It can be seen that the performance of our method is increased by \\(4.6\\%\\) compared with random point selection and surpasses the state-of-the-art region-based active learning method by \\(1\\%\\) with significantly fewer labelled points (0.1% v.s. 11.7%)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.348, + 0.468, + 0.454 + ], + "angle": 0, + "content": "
TypeMethodsSegmentorslabelled pointsmIoU(%)
Region-basedReDALSPVCNN15%58.0
ReDALMinkowskiNet15%57.3
SSDR-ALRandlanet11.7%58.3
Point-basedRandomMinkowskiNet0.1%54.7
EntropyMinkowskiNet0.1%48.2
LCMinkowskiNet0.1%53.3
MMUMinkowskiNet0.1%55.5
OursMinkowskiNet0.1%59.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.495 + ], + "angle": 0, + "content": "Table 2. The comparison of mIoU between our active strategy without the semi-supervised module and different active learning methods on the S3DIS dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.467, + 0.531 + ], + "angle": 0, + "content": "4.4. Comparison with Weakly-supervised Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.468, + 0.629 + ], + "angle": 0, + "content": "We also compare the proposed method with the state-of-the-art weakly-supervised methods. Similar to [22, 48], we present state-of-the-art segmentation methods with different supervision settings. The comparison results on the datasets of S3DIS and ScannetV2 are presented in Table 3 and Table 4, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.468, + 0.901 + ], + "angle": 0, + "content": "S3DIS. we perform experiments under three different annotation budgets, namely \\(0.7\\%\\), \\(0.07\\%\\), and \\(0.02\\%\\). For the sake of fairness, we follow the SQN [13] to calculate the annotation percentage which is defined as the ratio between the number of labelled data and all training and validation data. Each annotation budget is completed through five iterations. As shown in Table 3, our approach gains consistent improvements in the segmentation performance with labelled points increased from \\(0.02\\%\\) to \\(0.43\\%\\). It is worth noting that, for the \\(0.7\\%\\) annotation budget, the segmentation performance at the third iteration (\\(0.43\\%\\) annotations) has already surpassed the fully-supervised MinkowskiNet. This verifies that selecting the most representative points and making full use of the remaining unlabelled data can greatly benefit the performance of point cloud semantic segmentation. Our method also outperforms the state-of-the-art PSD, SQN and HybridCR under \\(1\\%\\) labelled data by \\(2.2\\%\\), \\(2.0\\%\\) and \\(0.4\\%\\), respectively. When the number of labelled" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.868, + 0.346 + ], + "angle": 0, + "content": "
MethodsLabelled pointsArea 5
PointNet [27]100%41.1
PointCNN [23]100%57.3
KPConv [35]100%67.1
PointTransformer [54]100%70.4
CBL [32]100%69.4
MinkowskiNet†[7]100%64.5
PSD [51]1%63.5
SQN [13]1%63.7
HybridCR [22]1%65.3
Ours0.43%65.7
SQN [13]0.1%61.4
Ours0.07%62.3
PSD [51]0.03%48.2
HybridCR [22]0.03%51.5
MIL [48]0.02%51.4
GaIA [21]0.02%53.7
VIB [36]0.02%52.0
Ours0.02%55.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.347, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Table 3. Comparison with existing weakly-supervised methods on S3DIS Area-5. † represents the result of the baseline trained on our own device." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.522 + ], + "angle": 0, + "content": "points is reduced to \\(0.07\\%\\), our method is also superior to the SQN with \\(0.1\\%\\) labelled points and achieves \\(96.5\\%\\) performance of the fully-supervised baseline (MinkowskiNet). In the case of the least labelling budget, we evaluate our method in the setting of \\(0.02\\%\\) labelled points. It can be seen that our method also achieves the best performance, outperforming the state-of-the-art GaIA, MIL, HybridCR and PSD by \\(2.2\\%\\), \\(4.5\\%\\), \\(4.4\\%\\) and \\(7.7\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.628 + ], + "angle": 0, + "content": "We also visualize the segmentation results of our method under different annotation budgets in Figure 3. The results show that with \\(0.43\\%\\) labelled data, our approach can achieve comparable segmentation results with the fully supervised baseline. Benefiting from the points selected by our method, our segmentation results are even more promising in some cases, such as the bookcase and column." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ScanNetV2. As illustrated in Table 4, we use the criteria of percentage and fixed number of labelled points as different annotation budgets to evaluate the methods, including \\(0.1\\%\\), 200pts, and 20pts. Similar to S3DIS, each annotation budget completes through five iterations, and the third iteration (120pts) is reported in the case of 200pts. We substantially beat SQN's performance by \\(11.3\\%\\) using \\(0.1\\%\\) of annotation. Compared against PSD, HybridCR and GaIA which are trained with \\(1\\%\\) annotations, our method outperforms them with only \\(0.1\\%\\) labels by \\(13.5\\%\\), \\(11.4\\%\\) and \\(3\\%\\), respectively. In the case of the fixed number of annotation points, we use 200 points per scene (200pts) and 20 points per scene (20pts) as the annotation budget. In our experiment with 200pts, when the labelling points reach 120pts, our method has achieved comparable performance to the fully supervised baseline. This proves the effectiveness of our active learning strategy on the semi-supervised point cloud semantic segmentation. We also evaluate our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.099, + 0.218, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.165, + 0.218, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.242, + 0.18, + 0.255 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.099, + 0.351, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.166, + 0.351, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.242, + 0.336, + 0.254 + ], + "angle": 0, + "content": "(b) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.099, + 0.483, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.166, + 0.483, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.242, + 0.482, + 0.255 + ], + "angle": 0, + "content": "(c) Fully-supervised" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.099, + 0.615, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.166, + 0.615, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.242, + 0.6, + 0.255 + ], + "angle": 0, + "content": "(d) Ours \\((0.02\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.099, + 0.75, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.166, + 0.749, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.242, + 0.733, + 0.255 + ], + "angle": 0, + "content": "(e) Ours \\((0.07\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.099, + 0.883, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.166, + 0.883, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.242, + 0.866, + 0.255 + ], + "angle": 0, + "content": "(f) Ours \\((0.43\\%)\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.198, + 0.263, + 0.782, + 0.276 + ], + "angle": 0, + "content": "ceiling floor wall column window door chair table bookcase board clutter" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.893, + 0.316 + ], + "angle": 0, + "content": "Figure 3. Visualization of segmentation results on the test set of S3DIS Area-5. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when \\(0.43\\%\\) of the labelled data is adopted." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.331, + 0.223, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.397, + 0.223, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.471, + 0.184, + 0.484 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.332, + 0.353, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.396, + 0.352, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.471, + 0.337, + 0.483 + ], + "angle": 0, + "content": "(b) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.331, + 0.481, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.396, + 0.481, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.471, + 0.48, + 0.484 + ], + "angle": 0, + "content": "(c) Fully-supervised" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.331, + 0.611, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.396, + 0.611, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.471, + 0.593, + 0.484 + ], + "angle": 0, + "content": "(d) Ours (20pts)" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.331, + 0.742, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.396, + 0.742, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.471, + 0.723, + 0.483 + ], + "angle": 0, + "content": "(e) Ours \\((0.1\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.331, + 0.874, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.396, + 0.873, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.471, + 0.857, + 0.483 + ], + "angle": 0, + "content": "(f) Ours (120pts)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.091, + 0.491, + 0.884, + 0.506 + ], + "angle": 0, + "content": "wall floor cabinet chair sofa table door bookshelf counter desk window sink curtain bed other furniture" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.522, + 0.893, + 0.551 + ], + "angle": 0, + "content": "Figure 4. Visualization of segmentation results on the validation set of ScannNetV2. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when 120 labelled points per scene are adopted." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.559, + 0.446, + 0.839 + ], + "angle": 0, + "content": "
MethodsLabelled pointsvaltest
PointNet++ [28]100%-33.9
KPCov [35]100%-68.4
VMNet [16]100%-74.6
BPNet [15]100%-74.9
MinkowskiNet†[7]100%69.368.0
PSD [51]1%-54.7
HybridCR [22]1%56.956.8
GaIA [21]1%-65.2
SQN [13]0.1%-56.9
Ours0.1%69.968.2
SQN [13]200pts-59.8
CSC [12]200pts68.266.5
GaIA [21]200pts-68.5
VIBUS [36]200pts69.669.1
Ours120pts70.269.4
SQN [13]20pts-48.6
CSC [12]20pts55.553.1
MIL [48]20pts57.854.4
VIBUS [36]20pts-58.6
GaIA [21]20pts-63.8
Ours20pts62.262.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.839, + 0.468, + 0.865 + ], + "angle": 0, + "content": "Table 4. The comparison with existing weakly-supervised methods. \\( {}^{ \\dagger } \\) is the result of the baseline trained on our device." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "method on the least annotation setting. In the case of 20pts, our result is much higher than SQN, CSC, MIL," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.636 + ], + "angle": 0, + "content": "and VIBUS, showing that our method effectively selects the most valuable points. However, our result is slightly lower than GaIA, as the performance of our feature extraction backbone is limited given only 20pts (the proportion is less than \\(0.02\\%\\) of ScanNetV2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.714 + ], + "angle": 0, + "content": "As shown in Fig. 4, we visualize the segmentation results of our method under different annotation budgets on ScanNetV2. It is observed that our method can achieve comparable or even better segmentation results with minor training budgets compared to our fully-supervised baseline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.665, + 0.74 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of each module in our method, we conduct the following ablation studies on S3DIS dataset. Following [13], all the ablated networks are trained using \\(0.1\\%\\) labelled points on Areas 1, 2, 3, 4, 6 and tested on Area 5. The number of iterations is set to 5 and \\(0.02\\%\\) data is selected in each iteration. To clearly see the effect of active learning, we present the results as shown in Figure 5. We take random sampling as the basic strategy for active selection where points are randomly selected in each iteration. After we replace the random sam" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18104" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.429, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.348 + ], + "angle": 0, + "content": "Figure 5. The comparison of performance improvements for different combinations of modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.614 + ], + "angle": 0, + "content": "pling with MMU for active sampling, the mIoU is slightly improved by \\(1.5\\%\\). After utilising HMMU to select points, we acquire an improvement of around \\(3.0\\%\\) compared with MMU in the last iteration. The FDS module is also added to further improve the performance. HMMU+FDS achieves an extra \\(2.9\\%\\) mIoU improvement compared to HMMU. More impressively, merely relying on the active learning modules including HMMU and FDS, our approach already outperforms the random sampling baseline by \\(8\\%\\) and surpasses \\(90\\%\\) performance of full-supervision with only \\(0.1\\%\\) labelling data. The effectiveness of the semi-supervised module is also demonstrated by the experimental results of HMMU+FDS+TS. Obviously, adding the semi-supervised teacher-student module, the segmentation result is further improved by \\(5\\%\\) compared to HMMU + FDS. The proposed method also reaches \\(95\\%\\) of the fully-supervised setting with only \\(0.1\\%\\) labelled data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.856 + ], + "angle": 0, + "content": "In Table 5, we show the effectiveness of different modules by quantitative comparison. We compare different compositions of our module to verify their effectiveness. Shown in the first row, for the base case where unlabelled points are selected by MMU in each iteration, the mIoU is only \\(55.51\\%\\). After using the HMMU module, the segmentation results are improved to \\(57.65\\%\\). If we only apply FDS to the MMU selection for each iteration, a slight improvement of \\(1.17\\%\\) is obtained compared with the based case. When the active learning strategy is removed, as shown in the fourth row in Table 5, the semi-supervised segmentation method only achieves the performance of \\(57.01\\%\\). Point cloud semantic segmentation only using the active learning technique can reach the mIoU of \\(59.29\\%\\). If we combine semi-supervised and active learning, as shown in the last row, the performance can be improved to \\(62.26\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Table 6, we show the ablation study of each layer in the HMMU module. PL represents the layer that calculates" + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.088, + 0.822, + 0.196 + ], + "angle": 0, + "content": "
ComponentsmIoU(%)
HMMUFDSTS
base.55.51
✓57.65
✓56.68
✓57.01
✓✓59.29
✓✓✓62.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.552, + 0.196, + 0.84, + 0.208 + ], + "angle": 0, + "content": "Table 5. Ablation study of different components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.892, + 0.374 + ], + "angle": 0, + "content": "the point-level minimum margin uncertainty score. VL is the layer which calculates the voxel-level contextual uncertainty score. \\( v_{r} \\) and \\( \\omega \\) are the voxel-radius and weight for each VL layer. It is shown that the combination of one layer of point-level uncertainty and three layers of voxel-level uncertainty can achieve the best performance. We also show the ablation study on the hyperparameters of the FDS module in Table 7. We achieve the best segmentation result given that \\( r \\) and \\( \\tau \\) are set to \\( 20\\mathrm{cm} \\) and 0.8 respectively. Thus, we adopt this combination in our experiments." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.469 + ], + "angle": 0, + "content": "
LayerHMMUiter 1iter 2iter 3iter 4iter 5
ωvr(cm)
PL--37.8350.3754.3256.1856.68
PL + VL1{0.1}{10}37.8352.7154.8456.5257.03
PL + VL2{0.01}{50}37.8351.4555.8157.2257.43
PL + VL3{0.001}{100}37.8351.0655.6257.3658.03
PL + VL1,2{0.1, 0.01}{10, 50}37.8351.5655.2856.6658.21
PL + VL1,2,3{0.1, 0.01, 0.001}{10, 50, 100}37.8352.0256.0258.4959.29
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.001}{10, 50, 100, 200}37.8351.5656.3257.6157.71
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.0001}{10, 50, 100, 200}37.8351.6956.2258.5359.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.573, + 0.469, + 0.82, + 0.48 + ], + "angle": 0, + "content": "Table 6. Ablation study of HMMU layers." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.493, + 0.894, + 0.578 + ], + "angle": 0, + "content": "
FDSiteration1iteration2iteration3iteration4iteration5
r(cm)τ
200.837.8352.0256.0258.4959.29
200.937.8351.2754.9056.7958.79
400.837.8351.3555.7157.0957.79
400.937.8351.6555.2657.4559.13
500.9537.8350.8755.7356.4957.82
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.502, + 0.578, + 0.878, + 0.589 + ], + "angle": 0, + "content": "Table 7. Ablation study of the hyperparameters in FDS module." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.619, + 0.62 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In this paper, we study active learning for 3D point cloud semantic segmentation. In comparison with previous region-based methods, we propose a hierarchical point-based active learning strategy consisting of two new designs, HMMU and FDS. HMMU serves as an effective way to measure the uncertainty or importance of labelling, while FDS enables us to select the most valuable points by considering the feature similarity and spatial distribution. To better use the unlabelled points during network learning, we also introduce a teacher-student structure to generate pseudo-labels. Extensive experiments are conducted on S3DIS and ScanNet datasets to demonstrate the effectiveness of our method. In the future, a more advanced semi-supervised architecture can be adopted for generating more reliable pseudo-labels to further facilitate active learning based point cloud semantic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62206033," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "No.62036007, No.62221005), and the Natural Science Foundation of Chongqing (No. cstc2020jcyj-msxmX0855, No. cstc2021ycjh-bgzxm0339)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.149, + 0.174, + 0.166 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.47, + 0.243 + ], + "angle": 0, + "content": "[1] Rashid Abbasi, Ali Kashif Bashir, Hasan J Alyamani, Farhan Amin, Jaehyeok Doh, and Jianwen Chen. Lidar point cloud compression, processing and learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 24(1):962-979, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[2] Eren Erdal Aksoy, Saimir Baci, and Selcuk Cavdar. Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In 2020 IEEE intelligent vehicles symposium (IV), pages 926-932. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.472, + 0.369 + ], + "angle": 0, + "content": "[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.37, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[4] Alexandre Boulch, Bertrand Le Saux, Nicolas Audebert, et al. Unstructured point cloud semantic labeling using deep segmentation networks. 3dor@eurographics, 3:17-24, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.47, + 0.482 + ], + "angle": 0, + "content": "[5] Mingmei Cheng, Le Hui, Jin Xie, and Jian Yang. Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1140-1147, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.566, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.636, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[8] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5-7, 2020, Proceedings, Part II 15, pages 207-222. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[10] Shuang Deng, Qiulei Dong, Bo Liu, and Zhanyi Hu. Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In 2022 International Conference on Robotics and Automation (ICRA), pages 9214-9220. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.174, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[13] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 600-619. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[15] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[16] Zeyu Hu, Xuyang Bai, Jiaxiang Shang, Runze Zhang, Jiayu Dong, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15488-15498, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[17] Ankang Ji, Yunxiang Zhou, Limao Zhang, Robert LK Tiong, and Xiaolong Xue. Semi-supervised learning-based point cloud network for segmentation of 3d tunnel scenes. Automation in Construction, 146:104668, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[18] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6423-6432, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[19] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.894, + 0.789 + ], + "angle": 0, + "content": "[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[21] Min Seok Lee, Seok Woo Yang, and Sung Won Han. Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 582-591, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[22] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18106" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "contrastive regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14930-14939, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[24] Y Lin, G Vosselman, Y Cao, and MY Yang. Efficient training of semantic point cloud segmentation via active learning. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2:243-250, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[25] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4213-4220. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.471, + 0.518 + ], + "angle": 0, + "content": "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[29] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[30] Xian Shi, Xun Xu, Ke Chen, Lile Cai, Chuan Sheng Foo, and Kui Jia. Label-efficient point cloud semantic segmentation: An active learning approach. arXiv preprint arXiv:2101.06931, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII, pages 685–702. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[32] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8489-8499, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[33] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 2017 international conference on 3D vision (3DV), pages 537-547. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[34] Hugues Thomas, Ben Agro, Mona Gridseth, Jian Zhang, and Timothy D Barfoot. Self-supervised learning of lidar segmentation for autonomous indoor navigation. In 2021" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE International Conference on Robotics and Automation (ICRA), pages 14047-14053. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[36] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS Journal of Photogrammetry and Remote Sensing, 194:302-318, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[37] Haiyan Wang, Xuejian Rong, Liang Yang, Jinglun Feng, Jizhong Xiao, and Yingli Tian. Weakly supervised semantic segmentation in 3d graph-structured point clouds of wild scenes. arXiv preprint arXiv:2004.12498, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[38] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10296-10305, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.379, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[39] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1887-1893. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.45, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[40] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[41] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15510-15519, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[42] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[43] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[44] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 1-19. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[45] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18107" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision, pages 638-656. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[46] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[47] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[48] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[49] Ping-Chung Yu, Cheng Sun, and Min Sun. Data efficient 3d learner via knowledge transferred from 2d model. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 182-198. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[50] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3421-3429, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[51] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15520-15528, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[52] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[53] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18108" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_origin.pdf b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ce85ebcaaeaa264585c2df0444ba1c22a303309c --- /dev/null +++ b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/67399329-b319-4f94-bd3d-1cdd42128783_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:892093dd0cda3f685e303fa545a309cac0c2fd2b77d1c6251d639e0953866bdb +size 5640723 diff --git a/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/full.md b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..15395265551b79c93bd74e36490c3d7c60e9a4b9 --- /dev/null +++ b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/full.md @@ -0,0 +1,378 @@ +# Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation + +Zongyi Xu $^{1\dagger}$ Bo Yuan $^{1\dagger}$ Shanshan Zhao $^{2\dagger}$ Qianni Zhang $^{3}$ Xinbo Gao $^{1*}$ $^{1}$ Chongqing University of Posts and Telecommunications, China + $^{2}$ JD Explore Academy, China $^{3}$ Queen Mary University of London, UK +{xuzy, qaoxb}@cqupt.edu.cn s210201127@stu.cqupt.edu.cn +sshan.zhao00@gmail.com qianni.zhang@qmul.ac.uk + +# Abstract + +Impressive performance on point cloud semantic segmentation has been achieved by fully-supervised methods with large amounts of labelled data. As it is labour-intensive to acquire large-scale point cloud data with pointwise labels, many attempts have been made to explore learning 3D point cloud segmentation with limited annotations. Active learning is one of the effective strategies to achieve this purpose but is still under-explored. The most recent methods of this kind measure the uncertainty of each pre-divided region for manual labelling but they suffer from redundant information and require additional efforts for region division. This paper aims at addressing this issue by developing a hierarchical point-based active learning strategy. Specifically, we measure the uncertainty for each point by a hierarchical minimum margin uncertainty module which considers the contextual information at multiple levels. Then, a feature-distance suppression strategy is designed to select important and representative points for manual labelling. Besides, to better exploit the unlabelled data, we build a semi-supervised segmentation framework based on our active strategy. Extensive experiments on the S3DIS and ScanNetV2 datasets demonstrate that the proposed framework achieves $96.5\%$ and $100\%$ performance of fully-supervised baseline with only $0.07\%$ and $0.1\%$ training data, respectively, outperforming the state-of-the-art weakly-supervised and active learning methods. The code will be available at https://github.com/SmiletoE/HPAL. + +# 1. Introduction + +Point cloud semantic segmentation aims to assign a category label for each 3D point, which can be applied to various scenarios, such as robotics [34], autonomous driving [1], and augmented reality [20]. Recently, deep learning + +based methods [38, 52] have achieved impressive performance. These high-performing methods usually rely on large amounts of data with point-wise labels. However, acquiring such dense labels for 3D point clouds is extremely tedious and costly. + +To relieve the labour and cost of annotation, many recent methods explore semi-supervised learning to learn 3D segmentation models with limited labelled points that are usually sampled randomly. These methods attempt to employ effective strategies to propagate the label information to the unlabelled points [5, 17, 10, 13]. Although these semi-supervised methods can greatly decrease labelling costs, their performance might be limited due to various factors. Among them, the most important reason is that some of the randomly selected points are in fact redundant while some really important points might be omitted. + +As an alternative learning strategy, active learning recently is studied to alleviate such limitations for 3D segmentation. Lin et al. [24] divide the whole point clouds into segments and each segment is utilised as the basic query unit for sample selection. ReDAL [41] proposes to select those informative and diverse sub-scene regions for label acquisition. The entropy, colour discontinuity, and structural complexity are used to measure the information of sub-scene regions. Following it, SSDR-AL [29] groups the original point clouds into superpoints and incrementally selects the most informative regions for annotation. However, these methods usually rely on pre-dividing the point cloud into multiple regions, and the region division strategy might negatively impact the performance. Moreover, as point clouds present strong semantic similarity in local areas, selecting all the points in the local region results in a redundant labelling budget. + +Motivated by the above analysis, this research aims at enhancing the 3D segmentation performance with limited labelled data in the active learning framework. Specifically, in comparison with previous region-based methods, we mea + +sure the uncertainty or importance of labelling for each point to avoid the additional efforts and potential bias of region division. However, considering each point individually may introduce extra noise to the training process, leading to unreliable uncertainty values that cannot reflect the real importance of each point. To alleviate this issue, we design a hierarchical minimum margin uncertainty (HMMU) measurement module by considering the local contextual information. The HMMU involves the uncertainty of the point and its neighbourhood. In this way, the acquired score can more effectively represent the importance of each point label. Based on the uncertainty scores acquired by HMMU, we can directly choose the Top- $K$ points for labelling. However, as the point cloud presents semantic similarity in the local region, some redundant uncertain points tend to be distributed densely. To further reduce the labelling cost, we propose a feature-distance suppression module (FDS) to remove those uncertain points with similar features in the neighbourhood. By exploiting the devised HMMU measurement and FDS for redundant point filtering, our hierarchical point-based active learning strategy is able to select more valuable points for labelling and thus reduce labelling costs. Furthermore, inspired by previous semi-supervised methods that exploit the unlabelled points, we utilise a simple teacher-student scheme to enhance the supervision signal by assigning a pseudo label for the unlabelled points. + +The main contributions of this research are as follows: + +- We propose a point-based active learning method for semi-supervised point cloud semantic segmentation, which surpasses current semi-supervised or active learning methods and achieves comparable performance with full supervision methods relying on scarce annotations. +- We propose a novel hierarchical minimum margin uncertainty module to measure the uncertainty for each point by progressively perceiving the contextual information at increasing scales. +- We propose a feature-distance suppression module to remove redundant points with similar features in the neighbourhood and further minimise the labelling cost. + +# 2. Related Work + +# 2.1. 3D Semantic Segmentation + +In recent years, a large number of deep learning-based point cloud semantic segmentation methods have been developed, which can be roughly divided into the following categories: 1) 2D projection-based methods [4, 19, 2, 8, 26, 39, 40, 44], which project the 3D point cloud onto 2D images through a variety of viewpoints including multiview, bird's eye view, spherical projection etc. Projection-based methods can make full use of those well-designed 2D + +convolution networks for 3D scene parsing, but obviously, the projection causes the loss of 3D geometric information and thus limits the performance. 2) voxel-based methods [6, 11, 31, 47, 7, 33]. Point cloud voxelization regularizes point clouds with uneven densities, making it possible to extend regular 2D convolutions to 3D scenes. At an early stage, voxel-based methods usually suffer from the explosion of computation and the loss of information. Fortunately, some sparse convolution methods [11, 31, 7] have been proposed over time to greatly ease the computational costs. 3) point-based methods [14, 27, 28, 35, 23, 42]. These methods directly take the original uneven point cloud as input and use MLPs, 3D point convolution, and other more flexible ways to process the point clouds, which preserve the original 3D information. However, their high performance still relies on a large number of point-level labels. + +# 2.2. Weakly-supervised 3D Semantic Segmentation + +We collectively refer to all methods training with sparse annotation as weakly supervised methods, which is a well-explored field in point cloud segmentation. These methods are usually based on a small set of randomly labelled data and then exploit more information from the unlabelled data by leveraging techniques such as transfer learning, contrastive learning, consistency regularization, pseudo-label, etc. Based on the transfer learning technologies, pre-training is widely exploited in 2D images, and has recently been gradually applied to point clouds [50, 49, 43, 53, 37, 45]. These methods learn pre-trained knowledge from 3D point clouds or 2D images and then apply it to the target dataset, thus achieving better training results with limited labelled data. Some other works train the model by using consistency regularization [46, 51, 22, 48] and pseudo-label [13, 36, 5, 25, 50, 22]. SQN [13] is one of these works which designs a novel semantic query network to efficiently obtain the pseudo-labels. To further improve weakly supervised training, some approaches also use contrastive learning [22, 25, 12, 18, 43], which usually combine contrastive loss with pre-training, consistency regularization, and pseudo-labels to better exploit the supervision information. Most weakly-supervised methods explore information from the labelled data, so how to select the most informative data to label is a question of crucial importance. + +# 2.3. Active Learning on 3D Semantic Segmentation + +Similar to weakly-supervised learning, active learning also aims to train deep models with limited labelled data. The main difference is that weakly-supervised learning mainly focuses on enhancing supervision by exploiting available label information while active learning focuses on the selection of valuable information for labelling. + +Lin et al. [24] make the first attempt for active learning on 3D semantic segmentation using segments as the ba + +![](images/0ac33a529850ced3fdabce95f9d216f413bb77ad63190df5c58ea36b48a1b59a.jpg) +Figure 1. The proposed hierarchical point-based active learning framework for semi-supervised point cloud semantic segmentation. (a) A segmentation network using the teacher-student model is trained with the initial labelled and unlabelled training data. The teacher is updated from the student by the exponential moving averaging (EMA) and the student network is obtained as our target segmentation model. (b) Those representative candidate points are measured by the trained model using the proposed hierarchical minimum margin uncertainty module (HMMU). (c) Feature-distance Suppression module (FDS) is then applied to remove those redundant points in the neighbourhood and select the final to-be-annotated points. The selected valuable points are annotated and added to the labelled training set to improve the segmentation model. + +sic query unit, and propose a segment entropy strategy to measure the informativeness of each segment. To improve the efficiency of active learning, superpoints are introduced as the most common basic query units [30, 41]. Considering the limitation of uncertainty, Shi et al. [30] propose to design active acquisition functions from multiple perspectives, including feature diversity, shape diversity, and entropy. ReDAL [41] combines entropy with the geometric structure and colour information and uses feature clustering to prevent selection redundancy. In addition to the entropy-based method, SSDR-AL [29] improves MMU based on super-point by considering the principal class within superpoints. To date, all active learning methods for point cloud semantic segmentation are region based. However, the pre-region division can greatly affect the active learning performance and unnecessary annotation in the semantic-similar areas is prone to occur. + +# 3. Approach + +# 3.1. Overview + +In the following, we describe our proposed hierarchical point-based active learning approach for semi-supervised point cloud semantic segmentation in detail. As shown in Figure 1, the initial labelled and unlabelled training data are utilised to train the point cloud semantic segmentation neural network in a semi-supervised way. A teacher-student model is adopted. With the trained model, the valuable unlabelled points that are able to improve the segmentation performance most can be measured and selected in an active learning manner. The hierarchical minimum margin uncer + +tainty module is proposed to measure the point-wise uncertainty by hierarchically increasing contextual ranges, which enables progressively capturing the local context at different scales. Then, the neighbouring points with similar features are removed by the feature distance suppression module, thereby further reducing the annotation redundancy. The remaining points with high uncertainty scores as well as sparse distributions are regarded as the candidates to be annotated. In this manner, the most valuable and representative unlabelled points are selected to expand the labelled set, with which the semi-supervised point cloud semantic segmentation neural network can be further enhanced. In the following, we describe each component of the proposed framework in detail. + +# 3.2. Hierarchical Minimum Margin Uncertainty Measurement + +In the proposed method, we aim to select and label the most valuable points that can bring maximum performance improvements. Intuitively, identifying and using the most informative labelled data for training is the key to obtaining a model with good quality, and consequently accurate segmentation prediction. Merely considering individual points without their surrounding context information cannot reflect the real importance of the point. Thus, we design the hierarchical minimum margin uncertainty measurement module to calculate the uncertainty score for each point by grouping points at multiple scales and progressively perceiving context information in a broader range along the hierarchy for the unlabelled point. + +As shown in Figure 1, the prediction for each point gen + +erated with the currently trained model is input into the HMMU module. We first calculate the point-level minimum margin uncertainty score $U_{x}$ for each unlabelled point with Eq. 1: + +$$ +U _ {x} = h \left(x ^ {u}; p _ {1} \left(x ^ {u}\right)\right) - h \left(x ^ {u}; p _ {2} \left(x ^ {u}\right)\right), \tag {1} +$$ + +where $x^{u}$ is the candidate point that is chosen to be labelled; $p_1(x^u)$ and $p_2(x^u)$ are the highest and second-highest prediction of $x^{u}$ under the segmentation predictor $h(\cdot)$ . + +Then we group a wider range of neighbours through downsampling. The softmax prediction of each downsampled point is obtained by averaging the softmax prediction of its neighbouring points in the original point cloud. Each softmax label of the downsampled point represents the prediction distribution of a local region, which is denoted by $S_{R}$ . + +We perform $N$ levels of voxel downsampling. The local context information of the to-be-annotated point $x^{u}$ at the $i_{th}$ level of downsampling, denoted by $S_R^i (x^u)$ , can be represented as the average of predictions of the grouped points in the following: + +$$ +S _ {R} ^ {i} \left(x ^ {u}\right) = \frac {1}{K} \sum_ {j = 1} ^ {K} p \left(x _ {j} ^ {u}\right), \tag {2} +$$ + +where $x_{j}^{u}$ is the $j_{th}$ neighbour for $x^{u}$ in the original point cloud; $K$ is the number of neighbours within the predefined voxel radius $v_{r}$ ; and $p(\cdot)$ is the prediction probability. + +For each level of downsampling, the voxel-level contextual uncertainty scores $U_{R}^{i}$ for the unlabelled point $x^{u}$ can be obtained with: + +$$ +U _ {R} ^ {i} = h \left(x ^ {u}; S _ {R 1} ^ {i} \left(x ^ {u}\right)\right) - h \left(x ^ {u}; S _ {R 2} ^ {i} \left(x ^ {u}\right)\right), \tag {3} +$$ + +where $S_{R1}^{i}(x^{u})$ and $S_{R2}^{i}(x^{u})$ are the highest and the second highest average prediction of the grouped points of $x^{u}$ from the segmentor $h(\cdot)$ in the $i_{th}$ downsampling. + +Finally, for each unannotated point $x^u$ , we integrate the point-level and the voxel-level contextual uncertainty scores to get the final uncertainty $v^u$ using Eq. 4: + +$$ +v ^ {u} = U _ {x} + \sum_ {i = 1} ^ {N} \omega^ {i} \times U _ {R} ^ {i}. \tag {4} +$$ + +Here, $\omega^i\in \{0.1,0.01,0.001\}$ is the hyperparameter representing the fusion weight of the contextual uncertainty at the $i_{th}$ level of downsampling and three levels of downsampling is performed in the experiment. + +With HMMU, the context information is fused to significantly improve performance by merely annotating points rather than regions. + +![](images/6846c69fb3aba47166d515065c268b83f7dce9c71f1a6cdc34208b84c17e8331.jpg) +(a) + +![](images/ab1732e2df70d568dd28cc51d9f005751291274703fbf2bae48f33b69a8dfb4e.jpg) +(b) + +![](images/530891ab13e28aef388ed5b3b791b1b4c291e0bf98261b155e8229be8963ff26.jpg) +(c) + +![](images/1b612dccfc37e0870c61f196a023acd436dc0fc37aee3691fb93fdc32fafc9e5.jpg) +(d) +Figure 2. Visualization of selected points. (a) and (c) show the points selected using the Top- $K$ approach (blue dots); (b) and (d) present points chosen by FDS (orange dots). FDS enables a more spread-out point selection. + +# 3.3. Feature-distance Suppression based Point Selection + +With the point-wise uncertainty score $v^u$ , we can select the Top- $K$ points with the highest uncertainty to label. However, as shown in Figure 2, such a point selection way might make points concentrate in local areas, causing label redundancy. Thus, we propose a feature-distance suppression (FDS) module to ensure the selected points retain a spread-out distribution in the space, thus offering a more effective overall representation. + +Given a distance suppression radius $r$ and a feature similarity threshold $\tau$ , for a point $x_{i}$ to be selected, we first determine whether there are points within its radius $r$ that have been already selected. A judgment set $D^{i}$ is constructed for the candidate point $x_{i}$ in the current point selection iteration. + +$$ +D ^ {i} = D ^ {i} \cup \{x _ {j} \}, i f \tag {5} +$$ + +$$ +\forall x _ {j} i s s e l e c t e d: d _ {i j} < r, +$$ + +where $x_{j}$ is the neighbour of $x_{i}$ within the radius of $r$ ; $d_{ij}$ is the Euclidean distance between $x_{i}$ and $x_{j}$ ; and $D^{i}$ represents if there are other candidate points in the neighbourhood and $D^{i}$ is initialized to be empty. + +If there is no neighbour of $x_{i}$ that is selected as a candidate, i.e. the judgment set $D^{i}$ is empty, $x_{i}$ is regarded as the valuable point that can represent the local area for labelling. Conversely, if the neighbouring point $x_{j}$ is already selected as the uncertainty point, cosine similarity $Sim_{ij}$ between the features of $x_{i}$ and $x_{j}$ is then computed with Eq. 6. + +$$ +\operatorname {S i m} \left(x _ {i}, x _ {j}\right) = \frac {\mathbf {f} _ {i} \cdot \mathbf {f} _ {j}}{\left| \left| \mathbf {f} _ {i} \right| \right| \cdot \left| \left| \mathbf {f} _ {j} \right| \right|}, \tag {6} +$$ + +where $\mathbf{f}_i$ and $\mathbf{f}_j$ are the features of $x_{i}$ and $x_{j}$ extracted with the trained student segmentation network. + +If there exists an $x_{j}$ in $D^{i}$ for which $Sim_{ij} > \tau$ , $x_{i}$ is regarded as a redundant point with respect to the labelled + +data and $x_{i}$ is excluded from labelling. + +# 3.4. Training objective + +To train a network with partly annotated data, a usual approach is to apply the cross-entropy loss to the labelled points and ignore the unlabelled ones, which causes inadequate supervision for model learning. Inspired by previous semi-supervised 3D point cloud segmentation methods, in the proposed method we employ a teacher-student framework to exploit unlabelled points and provide additional supervision. Specifically, we construct two segmentation networks using MinkowskiNet [7], denoted as the teacher model $M_{t}$ and student model $M_{s}$ , respectively. A consistency constraint between the teacher model and the student model is exploited to learn knowledge from unlabelled data. Pseudo-labels of unlabelled data are generated using the teacher model. During training, the original sample is input into the teacher network while its augmented counterparts are input into the student. For those labelled points, we apply the standard cross-entropy loss. For the unlabelled ones, we generate pseudo-labels from the output of the teacher and also compute the cross-entropy loss. The losses in both labelled and unlabelled points are then used to optimize the student network by gradient descent. After updating the student model at each step, Exponential Moving Average (EMA) is used to transfer the parameters of the student model to the teacher model: + +$$ +\theta_ {t} ^ {j} = \alpha \theta_ {t} ^ {j - 1} + (1 - \alpha) \theta_ {s} ^ {j}, \tag {7} +$$ + +where $\theta_{t}$ and $\theta_{s}$ are the parameters of the teacher and student networks, respectively, and $j$ denotes the $j$ -th training step. $\alpha$ is the hyper-parameter to determine the speed of parameter transmission, which is generally close to 1. + +# 4. Experiments + +# 4.1. Datasets + +We evaluate the proposed approach on two publicly available benchmark datasets, S3DIS [3] and ScanNetV2 [9], to demonstrate its superior performance. + +S3DIS (Stanford Large-Scale 3D Indoor Space Dataset) is a large-scale indoor point cloud dataset captured by Matter-port scanners which mainly dedicate to 3D segmentation tasks. It consists of 6 different large-scale indoor areas with a total of 271 rooms, each room is a separate point cloud sample that includes coordinate, colour, and annotation information for each point. Following the standard evaluation criterion, Area 5 is used as the test set and the rest of the areas are used as the training set. + +ScanNetV2 is an RGB-D video dataset. It contains 1513 point cloud samples from 707 individual indoor scenes for training and 100 unlabelled samples for testing. For the 3D semantic segmentation task, ScanNetV2 provides up to 40 + +categories of classification labels and selects 20 categories as classification targets to establish the segmentation task. Each point in the dataset contains coordinate, colour, and annotation information. In the experiments, we follow the official data split, dividing the 1513 training data into 1201 training samples and 312 validation samples, and evaluate methods on the test set. + +# 4.2. Implementation Details + +Segmentation Model. In our experiments, we use the PyTorch implementation of MinkowskiNet in ReDAL [41] as our backbone network with the stochastic gradient descent (SGD) optimizer. The architecture is trained on a single NVIDIA RTX A6000 GPU. + +Training Settings. For S3DIS, we set the batch size to 4, and train 60K steps per iteration. The initial learning rate is 0.1, and cosine annealing with a period of 5000 is used as the learning rate decay strategy. For ScanNetV2, we set the batch size to 4, and train 30K steps per iteration. The initial learning rate is 0.1, and the learning rate decay strategy is using polynomial decay with a power of 0.9. In the teacher-student model, the keep rate of EMA and the thresholds for the pseudo-label are set to 0.955 and 0.75, respectively. 5 iterations are performed in the active learning framework. In HMMU, the point cloud is downsampled at each scale with the sampling radius of $10\mathrm{cm}$ , $50\mathrm{cm}$ , and $100\mathrm{cm}$ , respectively. In the FDS module, the radius of the local region is set to $20\mathrm{cm}$ and the feature similarity threshold is set to 0.8. + +# 4.3. Comparison with Active Learning Methods + +To demonstrate the effectiveness of our active learning strategy, we remove the semi-supervised module in the comparisons with active learning methods. As current active learning methods for point cloud semantic segmentation are region-based, we compare our proposed method against them to verify the effectiveness of the point-based active learning method. The compared results are shown in Table 1. Compared with the region-based active learning methods, our point-based selection strategy greatly reduces the labelling cost, with the demand of achieving $90\%$ performance of the fully-supervised baselines. With only $0.1\%$ (about 13,000 points) of the downsampled points, our method allows for $90\%$ of the performance of our fully supervised baseline, i.e. fully-supervised MinkowskiNet. This verifies our theory that the strong semantic similarity in local areas of point clouds results in redundant labelling and ineffective training in region-based active learning. + +Furthermore, we compare the performance of our active learning strategy with existing region-based ones and four basic point-based selection strategies including random selection (Random), softmax least confidence (LC), softmax minimum margin uncertainty (MMU) and softmax entropy + +
MethodsSegmentorsLabelled points
ReDAL [41]SPVCNN13%
ReDAL [41]MinkowskiNet15%
SSDR-AL [29]Randlanet11.7%
OursMinkowskiNet0.1%
+ +Table 1. The comparison of the percentage of labelled points required to achieve $90\%$ accuracy of fully-supervised baseline on the S3DIS dataset against current region-based active learning methods. For a fair comparison, the semi-supervised module of our method is removed here which means only labelled data and selected points are used during training. + +(Entropy). The results are shown in Table 2. It can be seen that the performance of our method is increased by $4.6\%$ compared with random point selection and surpasses the state-of-the-art region-based active learning method by $1\%$ with significantly fewer labelled points (0.1% v.s. 11.7%). + +
TypeMethodsSegmentorslabelled pointsmIoU(%)
Region-basedReDALSPVCNN15%58.0
ReDALMinkowskiNet15%57.3
SSDR-ALRandlanet11.7%58.3
Point-basedRandomMinkowskiNet0.1%54.7
EntropyMinkowskiNet0.1%48.2
LCMinkowskiNet0.1%53.3
MMUMinkowskiNet0.1%55.5
OursMinkowskiNet0.1%59.3
+ +Table 2. The comparison of mIoU between our active strategy without the semi-supervised module and different active learning methods on the S3DIS dataset. + +# 4.4. Comparison with Weakly-supervised Methods + +We also compare the proposed method with the state-of-the-art weakly-supervised methods. Similar to [22, 48], we present state-of-the-art segmentation methods with different supervision settings. The comparison results on the datasets of S3DIS and ScannetV2 are presented in Table 3 and Table 4, respectively. + +S3DIS. we perform experiments under three different annotation budgets, namely $0.7\%$ , $0.07\%$ , and $0.02\%$ . For the sake of fairness, we follow the SQN [13] to calculate the annotation percentage which is defined as the ratio between the number of labelled data and all training and validation data. Each annotation budget is completed through five iterations. As shown in Table 3, our approach gains consistent improvements in the segmentation performance with labelled points increased from $0.02\%$ to $0.43\%$ . It is worth noting that, for the $0.7\%$ annotation budget, the segmentation performance at the third iteration ( $0.43\%$ annotations) has already surpassed the fully-supervised MinkowskiNet. This verifies that selecting the most representative points and making full use of the remaining unlabelled data can greatly benefit the performance of point cloud semantic segmentation. Our method also outperforms the state-of-the-art PSD, SQN and HybridCR under $1\%$ labelled data by $2.2\%$ , $2.0\%$ and $0.4\%$ , respectively. When the number of labelled + +
MethodsLabelled pointsArea 5
PointNet [27]100%41.1
PointCNN [23]100%57.3
KPConv [35]100%67.1
PointTransformer [54]100%70.4
CBL [32]100%69.4
MinkowskiNet†[7]100%64.5
PSD [51]1%63.5
SQN [13]1%63.7
HybridCR [22]1%65.3
Ours0.43%65.7
SQN [13]0.1%61.4
Ours0.07%62.3
PSD [51]0.03%48.2
HybridCR [22]0.03%51.5
MIL [48]0.02%51.4
GaIA [21]0.02%53.7
VIB [36]0.02%52.0
Ours0.02%55.9
+ +Table 3. Comparison with existing weakly-supervised methods on S3DIS Area-5. † represents the result of the baseline trained on our own device. + +points is reduced to $0.07\%$ , our method is also superior to the SQN with $0.1\%$ labelled points and achieves $96.5\%$ performance of the fully-supervised baseline (MinkowskiNet). In the case of the least labelling budget, we evaluate our method in the setting of $0.02\%$ labelled points. It can be seen that our method also achieves the best performance, outperforming the state-of-the-art GaIA, MIL, HybridCR and PSD by $2.2\%$ , $4.5\%$ , $4.4\%$ and $7.7\%$ . + +We also visualize the segmentation results of our method under different annotation budgets in Figure 3. The results show that with $0.43\%$ labelled data, our approach can achieve comparable segmentation results with the fully supervised baseline. Benefiting from the points selected by our method, our segmentation results are even more promising in some cases, such as the bookcase and column. + +ScanNetV2. As illustrated in Table 4, we use the criteria of percentage and fixed number of labelled points as different annotation budgets to evaluate the methods, including $0.1\%$ , 200pts, and 20pts. Similar to S3DIS, each annotation budget completes through five iterations, and the third iteration (120pts) is reported in the case of 200pts. We substantially beat SQN's performance by $11.3\%$ using $0.1\%$ of annotation. Compared against PSD, HybridCR and GaIA which are trained with $1\%$ annotations, our method outperforms them with only $0.1\%$ labels by $13.5\%$ , $11.4\%$ and $3\%$ , respectively. In the case of the fixed number of annotation points, we use 200 points per scene (200pts) and 20 points per scene (20pts) as the annotation budget. In our experiment with 200pts, when the labelling points reach 120pts, our method has achieved comparable performance to the fully supervised baseline. This proves the effectiveness of our active learning strategy on the semi-supervised point cloud semantic segmentation. We also evaluate our + +![](images/519bae80d21186449d66858167662bdc2249649c6d7be3ae3096547a9b6c43e3.jpg) +ceiling floor wall column window door chair table bookcase board clutter + +![](images/183656d7b8788898e35d22e22cdd9a74fd02c1352f28775cf9ab06dc1b955d38.jpg) +(a) Input + +![](images/b5dc9b1d54fa4d239bbf6abbe761d2d1c2fb693e98f1fc9e1761c6e2a9313f9a.jpg) + +![](images/7ca96052ef21033b5bf642718808a1e413f4ed71eb88a3e799ecaf296491f55b.jpg) +(b) Ground Truth + +![](images/8b235e45e92fd9982c177383025f787de549cda90cb8017b04d3d579b411d19f.jpg) + +![](images/83a618d9ecd74f84398f4cf2fa7c839dd822eed7194447c9ccd9510ee850ab64.jpg) +(c) Fully-supervised + +![](images/4d08a212ffa052a2ec22ce8079b72d6a63718128d7eae9d7ba533d9b01d9d3d8.jpg) + +![](images/9ca709480bdfc169dfd7ba13f54f61e2ca3853399073a71b161fa53f65cae22d.jpg) +(d) Ours $(0.02\%)$ + +![](images/c08f3e6a3257c0bacafeeaa0aa429ffc927389ed668278863a286a1df7252c50.jpg) + +![](images/408d87ec47a483138c9fab8dd551491e2101a5ab479db8bd08deb34de316f631.jpg) +(e) Ours $(0.07\%)$ + +![](images/34b1278c1896af7e5465f113abefce71e5ec62ea716a7f2cd23246a491a1070b.jpg) + +![](images/b29481131af0468405bccacc07b28559c13c49f5562f63956cb55c2a60c363a5.jpg) +(f) Ours $(0.43\%)$ + +![](images/abdfd87f8f294378cac478211b2a023b21df1003270a1f4330b373b1dcad5ba9.jpg) +Figure 3. Visualization of segmentation results on the test set of S3DIS Area-5. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when $0.43\%$ of the labelled data is adopted. + +(a) Input +![](images/53df65a7fb776c9145d1ddc88ee1bb02479172c43bee91222e147391a50628b4.jpg) +wall floor cabinet chair sofa table door bookshelf counter desk window sink curtain bed other furniture + +![](images/f63f279ee846a0b46ec530267755cccac3adcc1e7793102621ac465942e87fc2.jpg) + +![](images/13aae238025387ca61900aa445135d86cc7f2f58fc7419d01a356cc970bf45c5.jpg) +(b) Ground Truth + +![](images/2dd32fc6eda9c971b434907dbba655470dc5bc3723380961c34f7da4c3986a37.jpg) + +![](images/97e3c0e02c4ebe39d150711c8eece58eeeb3ccdbba83e0ef3e0bedbbec7df215.jpg) +(c) Fully-supervised + +![](images/48ea1038526256f503990dc79d97e53c89d14d850815d8fa7649097c5da86d4e.jpg) + +![](images/85203a59f32ddd4ab34f2aaa47ce30a06d9081210229b3557734f99293668dc4.jpg) +(d) Ours (20pts) +Figure 4. Visualization of segmentation results on the validation set of ScannNetV2. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when 120 labelled points per scene are adopted. + +![](images/1560259c80f5ef5844010b57f5acff2dd73ee6c6979a897c1eb26235fb929a86.jpg) + +![](images/212c0cd5954a32b40ed744be2dfcf8bb64288bb89f839a72f3decbc6710f6939.jpg) +(e) Ours $(0.1\%)$ + +![](images/973ac929ba247559605112f64640899df48c6c0536824148525b4b1087fdbf5a.jpg) + +![](images/aadae06adf078e22a84d0a8cb68aa24c8ea6eb1b237dd138f5743815c6875625.jpg) +(f) Ours (120pts) + +
MethodsLabelled pointsvaltest
PointNet++ [28]100%-33.9
KPCov [35]100%-68.4
VMNet [16]100%-74.6
BPNet [15]100%-74.9
MinkowskiNet†[7]100%69.368.0
PSD [51]1%-54.7
HybridCR [22]1%56.956.8
GaIA [21]1%-65.2
SQN [13]0.1%-56.9
Ours0.1%69.968.2
SQN [13]200pts-59.8
CSC [12]200pts68.266.5
GaIA [21]200pts-68.5
VIBUS [36]200pts69.669.1
Ours120pts70.269.4
SQN [13]20pts-48.6
CSC [12]20pts55.553.1
MIL [48]20pts57.854.4
VIBUS [36]20pts-58.6
GaIA [21]20pts-63.8
Ours20pts62.262.5
+ +Table 4. The comparison with existing weakly-supervised methods. ${}^{ \dagger }$ is the result of the baseline trained on our device. + +method on the least annotation setting. In the case of 20pts, our result is much higher than SQN, CSC, MIL, + +and VIBUS, showing that our method effectively selects the most valuable points. However, our result is slightly lower than GaIA, as the performance of our feature extraction backbone is limited given only 20pts (the proportion is less than $0.02\%$ of ScanNetV2). + +As shown in Fig. 4, we visualize the segmentation results of our method under different annotation budgets on ScanNetV2. It is observed that our method can achieve comparable or even better segmentation results with minor training budgets compared to our fully-supervised baseline. + +# 4.5. Ablation Studies + +To demonstrate the effectiveness of each module in our method, we conduct the following ablation studies on S3DIS dataset. Following [13], all the ablated networks are trained using $0.1\%$ labelled points on Areas 1, 2, 3, 4, 6 and tested on Area 5. The number of iterations is set to 5 and $0.02\%$ data is selected in each iteration. To clearly see the effect of active learning, we present the results as shown in Figure 5. We take random sampling as the basic strategy for active selection where points are randomly selected in each iteration. After we replace the random sam + +![](images/fd66a2ae14a1b6265f3ab4f1741216d10c5eb4cc6e8d8f5c861822a7779e4565.jpg) +Figure 5. The comparison of performance improvements for different combinations of modules. + +pling with MMU for active sampling, the mIoU is slightly improved by $1.5\%$ . After utilising HMMU to select points, we acquire an improvement of around $3.0\%$ compared with MMU in the last iteration. The FDS module is also added to further improve the performance. HMMU+FDS achieves an extra $2.9\%$ mIoU improvement compared to HMMU. More impressively, merely relying on the active learning modules including HMMU and FDS, our approach already outperforms the random sampling baseline by $8\%$ and surpasses $90\%$ performance of full-supervision with only $0.1\%$ labelling data. The effectiveness of the semi-supervised module is also demonstrated by the experimental results of HMMU+FDS+TS. Obviously, adding the semi-supervised teacher-student module, the segmentation result is further improved by $5\%$ compared to HMMU + FDS. The proposed method also reaches $95\%$ of the fully-supervised setting with only $0.1\%$ labelled data. + +In Table 5, we show the effectiveness of different modules by quantitative comparison. We compare different compositions of our module to verify their effectiveness. Shown in the first row, for the base case where unlabelled points are selected by MMU in each iteration, the mIoU is only $55.51\%$ . After using the HMMU module, the segmentation results are improved to $57.65\%$ . If we only apply FDS to the MMU selection for each iteration, a slight improvement of $1.17\%$ is obtained compared with the based case. When the active learning strategy is removed, as shown in the fourth row in Table 5, the semi-supervised segmentation method only achieves the performance of $57.01\%$ . Point cloud semantic segmentation only using the active learning technique can reach the mIoU of $59.29\%$ . If we combine semi-supervised and active learning, as shown in the last row, the performance can be improved to $62.26\%$ . + +In Table 6, we show the ablation study of each layer in the HMMU module. PL represents the layer that calculates + +
ComponentsmIoU(%)
HMMUFDSTS
base.55.51
✓57.65
✓56.68
✓57.01
✓✓59.29
✓✓✓62.26
+ +the point-level minimum margin uncertainty score. VL is the layer which calculates the voxel-level contextual uncertainty score. $v_{r}$ and $\omega$ are the voxel-radius and weight for each VL layer. It is shown that the combination of one layer of point-level uncertainty and three layers of voxel-level uncertainty can achieve the best performance. We also show the ablation study on the hyperparameters of the FDS module in Table 7. We achieve the best segmentation result given that $r$ and $\tau$ are set to $20\mathrm{cm}$ and 0.8 respectively. Thus, we adopt this combination in our experiments. + +Table 5. Ablation study of different components. + +
LayerHMMUiter 1iter 2iter 3iter 4iter 5
ωvr(cm)
PL--37.8350.3754.3256.1856.68
PL + VL1{0.1}{10}37.8352.7154.8456.5257.03
PL + VL2{0.01}{50}37.8351.4555.8157.2257.43
PL + VL3{0.001}{100}37.8351.0655.6257.3658.03
PL + VL1,2{0.1, 0.01}{10, 50}37.8351.5655.2856.6658.21
PL + VL1,2,3{0.1, 0.01, 0.001}{10, 50, 100}37.8352.0256.0258.4959.29
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.001}{10, 50, 100, 200}37.8351.5656.3257.6157.71
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.0001}{10, 50, 100, 200}37.8351.6956.2258.5359.19
+ +Table 6. Ablation study of HMMU layers. + +
FDSiteration1iteration2iteration3iteration4iteration5
r(cm)τ
200.837.8352.0256.0258.4959.29
200.937.8351.2754.9056.7958.79
400.837.8351.3555.7157.0957.79
400.937.8351.6555.2657.4559.13
500.9537.8350.8755.7356.4957.82
+ +Table 7. Ablation study of the hyperparameters in FDS module. + +# 5. Conclusion + +In this paper, we study active learning for 3D point cloud semantic segmentation. In comparison with previous region-based methods, we propose a hierarchical point-based active learning strategy consisting of two new designs, HMMU and FDS. HMMU serves as an effective way to measure the uncertainty or importance of labelling, while FDS enables us to select the most valuable points by considering the feature similarity and spatial distribution. To better use the unlabelled points during network learning, we also introduce a teacher-student structure to generate pseudo-labels. Extensive experiments are conducted on S3DIS and ScanNet datasets to demonstrate the effectiveness of our method. In the future, a more advanced semi-supervised architecture can be adopted for generating more reliable pseudo-labels to further facilitate active learning based point cloud semantic segmentation tasks. + +Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62206033, + +No.62036007, No.62221005), and the Natural Science Foundation of Chongqing (No. cstc2020jcyj-msxmX0855, No. cstc2021ycjh-bgzxm0339). + +# References + +[1] Rashid Abbasi, Ali Kashif Bashir, Hasan J Alyamani, Farhan Amin, Jaehyeok Doh, and Jianwen Chen. Lidar point cloud compression, processing and learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 24(1):962-979, 2022. +[2] Eren Erdal Aksoy, Saimir Baci, and Selcuk Cavdar. Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In 2020 IEEE intelligent vehicles symposium (IV), pages 926-932. IEEE, 2020. +[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. +[4] Alexandre Boulch, Bertrand Le Saux, Nicolas Audebert, et al. Unstructured point cloud semantic labeling using deep segmentation networks. 3dor@eurographics, 3:17-24, 2017. +[5] Mingmei Cheng, Le Hui, Jin Xie, and Jian Yang. Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1140-1147, 2021. +[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. +[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. +[8] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5-7, 2020, Proceedings, Part II 15, pages 207-222. Springer, 2020. +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. +[10] Shuang Deng, Qiulei Dong, Bo Liu, and Zhanyi Hu. Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In 2022 International Conference on Robotics and Automation (ICRA), pages 9214-9220. IEEE, 2022. +[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the + +IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. +[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021. +[13] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 600-619. Springer, 2022. +[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020. +[15] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021. +[16] Zeyu Hu, Xuyang Bai, Jiaxiang Shang, Runze Zhang, Jiayu Dong, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15488-15498, 2021. +[17] Ankang Ji, Yunxiang Zhou, Limao Zhang, Robert LK Tiong, and Xiaolong Xue. Semi-supervised learning-based point cloud network for segmentation of 3d tunnel scenes. Automation in Construction, 146:104668, 2023. +[18] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6423-6432, 2021. +[19] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020. +[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. +[21] Min Seok Lee, Seok Woo Yang, and Sung Won Han. Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 582-591, 2023. +[22] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid + +contrastive regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14930-14939, 2022. +[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. +[24] Y Lin, G Vosselman, Y Cao, and MY Yang. Efficient training of semantic point cloud segmentation via active learning. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2:243-250, 2020. +[25] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021. +[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4213-4220. IEEE, 2019. +[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. +[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. +[29] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022. +[30] Xian Shi, Xun Xu, Ke Chen, Lile Cai, Chuan Sheng Foo, and Kui Jia. Label-efficient point cloud semantic segmentation: An active learning approach. arXiv preprint arXiv:2101.06931, 2021. +[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII, pages 685–702. Springer, 2020. +[32] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8489-8499, 2022. +[33] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 2017 international conference on 3D vision (3DV), pages 537-547. IEEE, 2017. +[34] Hugues Thomas, Ben Agro, Mona Gridseth, Jian Zhang, and Timothy D Barfoot. Self-supervised learning of lidar segmentation for autonomous indoor navigation. In 2021 + +IEEE International Conference on Robotics and Automation (ICRA), pages 14047-14053. IEEE, 2021. +[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. +[36] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS Journal of Photogrammetry and Remote Sensing, 194:302-318, 2022. +[37] Haiyan Wang, Xuejian Rong, Liang Yang, Jinglun Feng, Jizhong Xiao, and Yingli Tian. Weakly supervised semantic segmentation in 3d graph-structured point clouds of wild scenes. arXiv preprint arXiv:2004.12498, 2020. +[38] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10296-10305, 2019. +[39] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1887-1893. IEEE, 2018. +[40] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019. +[41] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15510-15519, 2021. +[42] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019. +[43] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. +[44] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 1-19. Springer, 2020. +[45] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In + +European Conference on Computer Vision, pages 638-656. Springer, 2022. +[46] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020. +[47] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021. +[48] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022. +[49] Ping-Chung Yu, Cheng Sun, and Min Sun. Data efficient 3d learner via knowledge transferred from 2d model. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 182-198. Springer, 2022. +[50] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3421-3429, 2021. +[51] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15520-15528, 2021. +[52] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020. +[53] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. +[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. \ No newline at end of file diff --git a/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/images.zip b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5ae3049ae0390f32bd2df93103cf777b64f1508 --- /dev/null +++ b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cddc3e8f2c03e9dc7d23a4d95dad6c3fa878af9d9f511c40c896b4ffc65ab4a9 +size 623093 diff --git a/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/layout.json b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..da7073831642bfc633444efb3f83ac87749bbfbc --- /dev/null +++ b/2023/Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation/layout.json @@ -0,0 +1,11269 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 527, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 527, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 527, + 139 + ], + "type": "text", + "content": "Hierarchical Point-based Active Learning for Semi-supervised Point Cloud Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "spans": [ + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": "Zongyi Xu" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": " Bo Yuan" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": " Shanshan Zhao" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": " Qianni Zhang" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": " Xinbo Gao" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": "Chongqing University of Posts and Telecommunications, China \n" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": "JD Explore Academy, China " + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 160, + 487, + 232 + ], + "type": "text", + "content": "Queen Mary University of London, UK \n{xuzy, qaoxb}@cqupt.edu.cn s210201127@stu.cqupt.edu.cn \nsshan.zhao00@gmail.com qianni.zhang@qmul.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "text", + "content": "Impressive performance on point cloud semantic segmentation has been achieved by fully-supervised methods with large amounts of labelled data. As it is labour-intensive to acquire large-scale point cloud data with pointwise labels, many attempts have been made to explore learning 3D point cloud segmentation with limited annotations. Active learning is one of the effective strategies to achieve this purpose but is still under-explored. The most recent methods of this kind measure the uncertainty of each pre-divided region for manual labelling but they suffer from redundant information and require additional efforts for region division. This paper aims at addressing this issue by developing a hierarchical point-based active learning strategy. Specifically, we measure the uncertainty for each point by a hierarchical minimum margin uncertainty module which considers the contextual information at multiple levels. Then, a feature-distance suppression strategy is designed to select important and representative points for manual labelling. Besides, to better exploit the unlabelled data, we build a semi-supervised segmentation framework based on our active strategy. Extensive experiments on the S3DIS and ScanNetV2 datasets demonstrate that the proposed framework achieves " + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "inline_equation", + "content": "96.5\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "text", + "content": " performance of fully-supervised baseline with only " + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 284, + 289, + 606 + ], + "type": "text", + "content": " training data, respectively, outperforming the state-of-the-art weakly-supervised and active learning methods. The code will be available at https://github.com/SmiletoE/HPAL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 648, + 287, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 648, + 287, + 697 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 287, + 697 + ], + "type": "text", + "content": "Point cloud semantic segmentation aims to assign a category label for each 3D point, which can be applied to various scenarios, such as robotics [34], autonomous driving [1], and augmented reality [20]. Recently, deep learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 259, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 545, + 320 + ], + "type": "text", + "content": "based methods [38, 52] have achieved impressive performance. These high-performing methods usually rely on large amounts of data with point-wise labels. However, acquiring such dense labels for 3D point clouds is extremely tedious and costly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 323, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 456 + ], + "type": "text", + "content": "To relieve the labour and cost of annotation, many recent methods explore semi-supervised learning to learn 3D segmentation models with limited labelled points that are usually sampled randomly. These methods attempt to employ effective strategies to propagate the label information to the unlabelled points [5, 17, 10, 13]. Although these semi-supervised methods can greatly decrease labelling costs, their performance might be limited due to various factors. Among them, the most important reason is that some of the randomly selected points are in fact redundant while some really important points might be omitted." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 459, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 662 + ], + "type": "text", + "content": "As an alternative learning strategy, active learning recently is studied to alleviate such limitations for 3D segmentation. Lin et al. [24] divide the whole point clouds into segments and each segment is utilised as the basic query unit for sample selection. ReDAL [41] proposes to select those informative and diverse sub-scene regions for label acquisition. The entropy, colour discontinuity, and structural complexity are used to measure the information of sub-scene regions. Following it, SSDR-AL [29] groups the original point clouds into superpoints and incrementally selects the most informative regions for annotation. However, these methods usually rely on pre-dividing the point cloud into multiple regions, and the region division strategy might negatively impact the performance. Moreover, as point clouds present strong semantic similarity in local areas, selecting all the points in the local region results in a redundant labelling budget." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Motivated by the above analysis, this research aims at enhancing the 3D segmentation performance with limited labelled data in the active learning framework. Specifically, in comparison with previous region-based methods, we mea" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "text", + "content": "*Corresponding author. †Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18098" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": "sure the uncertainty or importance of labelling for each point to avoid the additional efforts and potential bias of region division. However, considering each point individually may introduce extra noise to the training process, leading to unreliable uncertainty values that cannot reflect the real importance of each point. To alleviate this issue, we design a hierarchical minimum margin uncertainty (HMMU) measurement module by considering the local contextual information. The HMMU involves the uncertainty of the point and its neighbourhood. In this way, the acquired score can more effectively represent the importance of each point label. Based on the uncertainty scores acquired by HMMU, we can directly choose the Top- " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": " points for labelling. However, as the point cloud presents semantic similarity in the local region, some redundant uncertain points tend to be distributed densely. To further reduce the labelling cost, we propose a feature-distance suppression module (FDS) to remove those uncertain points with similar features in the neighbourhood. By exploiting the devised HMMU measurement and FDS for redundant point filtering, our hierarchical point-based active learning strategy is able to select more valuable points for labelling and thus reduce labelling costs. Furthermore, inspired by previous semi-supervised methods that exploit the unlabelled points, we utilise a simple teacher-student scheme to enhance the supervision signal by assigning a pseudo label for the unlabelled points." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 384, + 279, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 384, + 279, + 395 + ], + "spans": [ + { + "bbox": [ + 59, + 384, + 279, + 395 + ], + "type": "text", + "content": "The main contributions of this research are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 399, + 286, + 567 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 399, + 286, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 399, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 59, + 399, + 286, + 470 + ], + "type": "text", + "content": "- We propose a point-based active learning method for semi-supervised point cloud semantic segmentation, which surpasses current semi-supervised or active learning methods and achieves comparable performance with full supervision methods relying on scarce annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 477, + 286, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 477, + 286, + 524 + ], + "spans": [ + { + "bbox": [ + 59, + 477, + 286, + 524 + ], + "type": "text", + "content": "- We propose a novel hierarchical minimum margin uncertainty module to measure the uncertainty for each point by progressively perceiving the contextual information at increasing scales." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 531, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 531, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 59, + 531, + 286, + 567 + ], + "type": "text", + "content": "- We propose a feature-distance suppression module to remove redundant points with similar features in the neighbourhood and further minimise the labelling cost." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 579, + 134, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 134, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 134, + 592 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 195, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 195, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 195, + 612 + ], + "type": "text", + "content": "2.1. 3D Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "In recent years, a large number of deep learning-based point cloud semantic segmentation methods have been developed, which can be roughly divided into the following categories: 1) 2D projection-based methods [4, 19, 2, 8, 26, 39, 40, 44], which project the 3D point cloud onto 2D images through a variety of viewpoints including multiview, bird's eye view, spherical projection etc. Projection-based methods can make full use of those well-designed 2D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "convolution networks for 3D scene parsing, but obviously, the projection causes the loss of 3D geometric information and thus limits the performance. 2) voxel-based methods [6, 11, 31, 47, 7, 33]. Point cloud voxelization regularizes point clouds with uneven densities, making it possible to extend regular 2D convolutions to 3D scenes. At an early stage, voxel-based methods usually suffer from the explosion of computation and the loss of information. Fortunately, some sparse convolution methods [11, 31, 7] have been proposed over time to greatly ease the computational costs. 3) point-based methods [14, 27, 28, 35, 23, 42]. These methods directly take the original uneven point cloud as input and use MLPs, 3D point convolution, and other more flexible ways to process the point clouds, which preserve the original 3D information. However, their high performance still relies on a large number of point-level labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 272, + 545, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 545, + 285 + ], + "type": "text", + "content": "2.2. Weakly-supervised 3D Semantic Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 291, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 590 + ], + "type": "text", + "content": "We collectively refer to all methods training with sparse annotation as weakly supervised methods, which is a well-explored field in point cloud segmentation. These methods are usually based on a small set of randomly labelled data and then exploit more information from the unlabelled data by leveraging techniques such as transfer learning, contrastive learning, consistency regularization, pseudo-label, etc. Based on the transfer learning technologies, pre-training is widely exploited in 2D images, and has recently been gradually applied to point clouds [50, 49, 43, 53, 37, 45]. These methods learn pre-trained knowledge from 3D point clouds or 2D images and then apply it to the target dataset, thus achieving better training results with limited labelled data. Some other works train the model by using consistency regularization [46, 51, 22, 48] and pseudo-label [13, 36, 5, 25, 50, 22]. SQN [13] is one of these works which designs a novel semantic query network to efficiently obtain the pseudo-labels. To further improve weakly supervised training, some approaches also use contrastive learning [22, 25, 12, 18, 43], which usually combine contrastive loss with pre-training, consistency regularization, and pseudo-labels to better exploit the supervision information. Most weakly-supervised methods explore information from the labelled data, so how to select the most informative data to label is a question of crucial importance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 598, + 545, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 545, + 611 + ], + "type": "text", + "content": "2.3. Active Learning on 3D Semantic Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "Similar to weakly-supervised learning, active learning also aims to train deep models with limited labelled data. The main difference is that weakly-supervised learning mainly focuses on enhancing supervision by exploiting available label information while active learning focuses on the selection of valuable information for labelling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Lin et al. [24] make the first attempt for active learning on 3D semantic segmentation using segments as the ba" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18099" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 521, + 246 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 521, + 246 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 521, + 246 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 521, + 246 + ], + "type": "image", + "image_path": "0ac33a529850ced3fdabce95f9d216f413bb77ad63190df5c58ea36b48a1b59a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 547, + 327 + ], + "type": "text", + "content": "Figure 1. The proposed hierarchical point-based active learning framework for semi-supervised point cloud semantic segmentation. (a) A segmentation network using the teacher-student model is trained with the initial labelled and unlabelled training data. The teacher is updated from the student by the exponential moving averaging (EMA) and the student network is obtained as our target segmentation model. (b) Those representative candidate points are measured by the trained model using the proposed hierarchical minimum margin uncertainty module (HMMU). (c) Feature-distance Suppression module (FDS) is then applied to remove those redundant points in the neighbourhood and select the final to-be-annotated points. The selected valuable points are annotated and added to the labelled training set to improve the segmentation model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 337, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 540 + ], + "type": "text", + "content": "sic query unit, and propose a segment entropy strategy to measure the informativeness of each segment. To improve the efficiency of active learning, superpoints are introduced as the most common basic query units [30, 41]. Considering the limitation of uncertainty, Shi et al. [30] propose to design active acquisition functions from multiple perspectives, including feature diversity, shape diversity, and entropy. ReDAL [41] combines entropy with the geometric structure and colour information and uses feature clustering to prevent selection redundancy. In addition to the entropy-based method, SSDR-AL [29] improves MMU based on super-point by considering the principal class within superpoints. To date, all active learning methods for point cloud semantic segmentation are region based. However, the pre-region division can greatly affect the active learning performance and unnecessary annotation in the semantic-similar areas is prone to occur." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 554, + 115, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 115, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 115, + 568 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 574, + 115, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 115, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 115, + 586 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "In the following, we describe our proposed hierarchical point-based active learning approach for semi-supervised point cloud semantic segmentation in detail. As shown in Figure 1, the initial labelled and unlabelled training data are utilised to train the point cloud semantic segmentation neural network in a semi-supervised way. A teacher-student model is adopted. With the trained model, the valuable unlabelled points that are able to improve the segmentation performance most can be measured and selected in an active learning manner. The hierarchical minimum margin uncer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 337, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 504 + ], + "type": "text", + "content": "tainty module is proposed to measure the point-wise uncertainty by hierarchically increasing contextual ranges, which enables progressively capturing the local context at different scales. Then, the neighbouring points with similar features are removed by the feature distance suppression module, thereby further reducing the annotation redundancy. The remaining points with high uncertainty scores as well as sparse distributions are regarded as the candidates to be annotated. In this manner, the most valuable and representative unlabelled points are selected to expand the labelled set, with which the semi-supervised point cloud semantic segmentation neural network can be further enhanced. In the following, we describe each component of the proposed framework in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 514, + 545, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 545, + 538 + ], + "type": "text", + "content": "3.2. Hierarchical Minimum Margin Uncertainty Measurement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 700 + ], + "type": "text", + "content": "In the proposed method, we aim to select and label the most valuable points that can bring maximum performance improvements. Intuitively, identifying and using the most informative labelled data for training is the key to obtaining a model with good quality, and consequently accurate segmentation prediction. Merely considering individual points without their surrounding context information cannot reflect the real importance of the point. Thus, we design the hierarchical minimum margin uncertainty measurement module to calculate the uncertainty score for each point by grouping points at multiple scales and progressively perceiving context information in a broader range along the hierarchy for the unlabelled point." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "As shown in Figure 1, the prediction for each point gen" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18100" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "erated with the currently trained model is input into the HMMU module. We first calculate the point-level minimum margin uncertainty score " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "U_{x}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " for each unlabelled point with Eq. 1:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 131, + 287, + 145 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 131, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 83, + 131, + 287, + 145 + ], + "type": "interline_equation", + "content": "U _ {x} = h \\left(x ^ {u}; p _ {1} \\left(x ^ {u}\\right)\\right) - h \\left(x ^ {u}; p _ {2} \\left(x ^ {u}\\right)\\right), \\tag {1}", + "image_path": "9f103477d531eab3a2de843483a85319ae88ded8667df45fca6652c6f7c46418.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "text", + "content": " is the candidate point that is chosen to be labelled; " + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "inline_equation", + "content": "p_1(x^u)" + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "inline_equation", + "content": "p_2(x^u)" + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "text", + "content": " are the highest and second-highest prediction of " + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "text", + "content": " under the segmentation predictor " + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 47, + 156, + 287, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 193, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 276 + ], + "type": "text", + "content": "Then we group a wider range of neighbours through downsampling. The softmax prediction of each downsampled point is obtained by averaging the softmax prediction of its neighbouring points in the original point cloud. Each softmax label of the downsampled point represents the prediction distribution of a local region, which is denoted by " + }, + { + "bbox": [ + 46, + 193, + 287, + 276 + ], + "type": "inline_equation", + "content": "S_{R}" + }, + { + "bbox": [ + 46, + 193, + 287, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "content": "We perform " + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "content": " levels of voxel downsampling. The local context information of the to-be-annotated point " + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "content": " level of downsampling, denoted by " + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "inline_equation", + "content": "S_R^i (x^u)" + }, + { + "bbox": [ + 46, + 277, + 287, + 337 + ], + "type": "text", + "content": ", can be represented as the average of predictions of the grouped points in the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 348, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 348, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 116, + 348, + 287, + 381 + ], + "type": "interline_equation", + "content": "S _ {R} ^ {i} \\left(x ^ {u}\\right) = \\frac {1}{K} \\sum_ {j = 1} ^ {K} p \\left(x _ {j} ^ {u}\\right), \\tag {2}", + "image_path": "4b1afbf8943f180d90ce07fe89f5e136e46f890075f11082caad501691da187b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "inline_equation", + "content": "x_{j}^{u}" + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "inline_equation", + "content": "j_{th}" + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "content": " neighbour for " + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "content": " in the original point cloud; " + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "content": " is the number of neighbours within the predefined voxel radius " + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "content": "; and " + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 47, + 392, + 287, + 429 + ], + "type": "text", + "content": " is the prediction probability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 429, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 464 + ], + "type": "text", + "content": "For each level of downsampling, the voxel-level contextual uncertainty scores " + }, + { + "bbox": [ + 46, + 429, + 287, + 464 + ], + "type": "inline_equation", + "content": "U_{R}^{i}" + }, + { + "bbox": [ + 46, + 429, + 287, + 464 + ], + "type": "text", + "content": " for the unlabelled point " + }, + { + "bbox": [ + 46, + 429, + 287, + 464 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 46, + 429, + 287, + 464 + ], + "type": "text", + "content": " can be obtained with:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 475, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 475, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 75, + 475, + 287, + 491 + ], + "type": "interline_equation", + "content": "U _ {R} ^ {i} = h \\left(x ^ {u}; S _ {R 1} ^ {i} \\left(x ^ {u}\\right)\\right) - h \\left(x ^ {u}; S _ {R 2} ^ {i} \\left(x ^ {u}\\right)\\right), \\tag {3}", + "image_path": "8c732dc01273043b4a88549476cb9e9b9dc236ec7bf06a9decd6a5b7bba985a5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "inline_equation", + "content": "S_{R1}^{i}(x^{u})" + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "inline_equation", + "content": "S_{R2}^{i}(x^{u})" + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "text", + "content": " are the highest and the second highest average prediction of the grouped points of " + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "text", + "content": " from the segmentor " + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 47, + 500, + 287, + 536 + ], + "type": "text", + "content": " downsampling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 537, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 574 + ], + "type": "text", + "content": "Finally, for each unannotated point " + }, + { + "bbox": [ + 46, + 537, + 287, + 574 + ], + "type": "inline_equation", + "content": "x^u" + }, + { + "bbox": [ + 46, + 537, + 287, + 574 + ], + "type": "text", + "content": ", we integrate the point-level and the voxel-level contextual uncertainty scores to get the final uncertainty " + }, + { + "bbox": [ + 46, + 537, + 287, + 574 + ], + "type": "inline_equation", + "content": "v^u" + }, + { + "bbox": [ + 46, + 537, + 287, + 574 + ], + "type": "text", + "content": " using Eq. 4:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 584, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 584, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 110, + 584, + 287, + 617 + ], + "type": "interline_equation", + "content": "v ^ {u} = U _ {x} + \\sum_ {i = 1} ^ {N} \\omega^ {i} \\times U _ {R} ^ {i}. \\tag {4}", + "image_path": "19e64c8f4be03243f9b0866e592526503af7a08b64becc30804a5a565db31730.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\omega^i\\in \\{0.1,0.01,0.001\\}" + }, + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "text", + "content": " is the hyperparameter representing the fusion weight of the contextual uncertainty at the " + }, + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 46, + 628, + 287, + 676 + ], + "type": "text", + "content": " level of downsampling and three levels of downsampling is performed in the experiment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "With HMMU, the context information is fused to significantly improve performance by merely annotating points rather than regions." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 425, + 137 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 425, + 137 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 425, + 137 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 425, + 137 + ], + "type": "image", + "image_path": "6846c69fb3aba47166d515065c268b83f7dce9c71f1a6cdc34208b84c17e8331.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 138, + 377, + 146 + ], + "lines": [ + { + "bbox": [ + 368, + 138, + 377, + 146 + ], + "spans": [ + { + "bbox": [ + 368, + 138, + 377, + 146 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 533, + 137 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 533, + 137 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 533, + 137 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 533, + 137 + ], + "type": "image", + "image_path": "ab1732e2df70d568dd28cc51d9f005751291274703fbf2bae48f33b69a8dfb4e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 138, + 484, + 146 + ], + "lines": [ + { + "bbox": [ + 474, + 138, + 484, + 146 + ], + "spans": [ + { + "bbox": [ + 474, + 138, + 484, + 146 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 321, + 147, + 425, + 213 + ], + "blocks": [ + { + "bbox": [ + 321, + 147, + 425, + 213 + ], + "lines": [ + { + "bbox": [ + 321, + 147, + 425, + 213 + ], + "spans": [ + { + "bbox": [ + 321, + 147, + 425, + 213 + ], + "type": "image", + "image_path": "530891ab13e28aef388ed5b3b791b1b4c291e0bf98261b155e8229be8963ff26.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 216, + 377, + 223 + ], + "lines": [ + { + "bbox": [ + 368, + 216, + 377, + 223 + ], + "spans": [ + { + "bbox": [ + 368, + 216, + 377, + 223 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 427, + 147, + 533, + 213 + ], + "blocks": [ + { + "bbox": [ + 427, + 147, + 533, + 213 + ], + "lines": [ + { + "bbox": [ + 427, + 147, + 533, + 213 + ], + "spans": [ + { + "bbox": [ + 427, + 147, + 533, + 213 + ], + "type": "image", + "image_path": "1b612dccfc37e0870c61f196a023acd436dc0fc37aee3691fb93fdc32fafc9e5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 216, + 484, + 223 + ], + "lines": [ + { + "bbox": [ + 475, + 216, + 484, + 223 + ], + "spans": [ + { + "bbox": [ + 475, + 216, + 484, + 223 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 223, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 305, + 223, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 545, + 266 + ], + "type": "text", + "content": "Figure 2. Visualization of selected points. (a) and (c) show the points selected using the Top- " + }, + { + "bbox": [ + 305, + 223, + 545, + 266 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 223, + 545, + 266 + ], + "type": "text", + "content": " approach (blue dots); (b) and (d) present points chosen by FDS (orange dots). FDS enables a more spread-out point selection." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 275, + 545, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 275, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 306, + 275, + 545, + 299 + ], + "type": "text", + "content": "3.3. Feature-distance Suppression based Point Selection" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 306, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 401 + ], + "type": "text", + "content": "With the point-wise uncertainty score " + }, + { + "bbox": [ + 305, + 306, + 545, + 401 + ], + "type": "inline_equation", + "content": "v^u" + }, + { + "bbox": [ + 305, + 306, + 545, + 401 + ], + "type": "text", + "content": ", we can select the Top-" + }, + { + "bbox": [ + 305, + 306, + 545, + 401 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 306, + 545, + 401 + ], + "type": "text", + "content": " points with the highest uncertainty to label. However, as shown in Figure 2, such a point selection way might make points concentrate in local areas, causing label redundancy. Thus, we propose a feature-distance suppression (FDS) module to ensure the selected points retain a spread-out distribution in the space, thus offering a more effective overall representation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "content": "Given a distance suppression radius " + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "content": " and a feature similarity threshold " + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "content": ", for a point " + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "content": " to be selected, we first determine whether there are points within its radius " + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "content": " that have been already selected. A judgment set " + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "inline_equation", + "content": "D^{i}" + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "content": " is constructed for the candidate point " + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 402, + 545, + 471 + ], + "type": "text", + "content": " in the current point selection iteration." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 376, + 471, + 545, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 471, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 376, + 471, + 545, + 490 + ], + "type": "interline_equation", + "content": "D ^ {i} = D ^ {i} \\cup \\{x _ {j} \\}, i f \\tag {5}", + "image_path": "8db04057cfb2cda9e18a9fdd5406f848fbf83dfa5c0cba4366c348e213f1c2f9.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 372, + 487, + 479, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 487, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 372, + 487, + 479, + 499 + ], + "type": "interline_equation", + "content": "\\forall x _ {j} i s s e l e c t e d: d _ {i j} < r,", + "image_path": "c28930f6143c30489fbe862f3d967007a3e6fa25e9f1955fd293ddf5241bd089.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": " is the neighbour of " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": " within the radius of " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": " is the Euclidean distance between " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": "; and " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "D^{i}" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": " represents if there are other candidate points in the neighbourhood and " + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "inline_equation", + "content": "D^{i}" + }, + { + "bbox": [ + 305, + 503, + 545, + 551 + ], + "type": "text", + "content": " is initialized to be empty." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": "If there is no neighbour of " + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": " that is selected as a candidate, i.e. the judgment set " + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "D^{i}" + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": " is empty, " + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": " is regarded as the valuable point that can represent the local area for labelling. Conversely, if the neighbouring point " + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": " is already selected as the uncertainty point, cosine similarity " + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "Sim_{ij}" + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": " between the features of " + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 305, + 552, + 545, + 624 + ], + "type": "text", + "content": " is then computed with Eq. 6." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 363, + 632, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 632, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 363, + 632, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} \\left(x _ {i}, x _ {j}\\right) = \\frac {\\mathbf {f} _ {i} \\cdot \\mathbf {f} _ {j}}{\\left| \\left| \\mathbf {f} _ {i} \\right| \\right| \\cdot \\left| \\left| \\mathbf {f} _ {j} \\right| \\right|}, \\tag {6}", + "image_path": "ab20ae889d02938e0d6d96ecba3bfa00a668ce2d7b92f6a5393736ab2a18a6e1.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j" + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "text", + "content": " are the features of " + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 305, + 665, + 545, + 689 + ], + "type": "text", + "content": " extracted with the trained student segmentation network." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "If there exists an " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "D^{i}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " for which " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "Sim_{ij} > \\tau" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " is regarded as a redundant point with respect to the labelled" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18101" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 205, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 205, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 205, + 85 + ], + "type": "text", + "content": "data and " + }, + { + "bbox": [ + 47, + 72, + 205, + 85 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 72, + 205, + 85 + ], + "type": "text", + "content": " is excluded from labelling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 155, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 155, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 155, + 104 + ], + "type": "text", + "content": "3.4. Training objective" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 110, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 289, + 396 + ], + "type": "text", + "content": "To train a network with partly annotated data, a usual approach is to apply the cross-entropy loss to the labelled points and ignore the unlabelled ones, which causes inadequate supervision for model learning. Inspired by previous semi-supervised 3D point cloud segmentation methods, in the proposed method we employ a teacher-student framework to exploit unlabelled points and provide additional supervision. Specifically, we construct two segmentation networks using MinkowskiNet [7], denoted as the teacher model " + }, + { + "bbox": [ + 46, + 110, + 289, + 396 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 46, + 110, + 289, + 396 + ], + "type": "text", + "content": " and student model " + }, + { + "bbox": [ + 46, + 110, + 289, + 396 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 46, + 110, + 289, + 396 + ], + "type": "text", + "content": ", respectively. A consistency constraint between the teacher model and the student model is exploited to learn knowledge from unlabelled data. Pseudo-labels of unlabelled data are generated using the teacher model. During training, the original sample is input into the teacher network while its augmented counterparts are input into the student. For those labelled points, we apply the standard cross-entropy loss. For the unlabelled ones, we generate pseudo-labels from the output of the teacher and also compute the cross-entropy loss. The losses in both labelled and unlabelled points are then used to optimize the student network by gradient descent. After updating the student model at each step, Exponential Moving Average (EMA) is used to transfer the parameters of the student model to the teacher model:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 403, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 403, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 113, + 403, + 287, + 420 + ], + "type": "interline_equation", + "content": "\\theta_ {t} ^ {j} = \\alpha \\theta_ {t} ^ {j - 1} + (1 - \\alpha) \\theta_ {s} ^ {j}, \\tag {7}", + "image_path": "e9ff679cdc80b1ff0e9adb6de094ef0a023a88b0f5ffb01ec6fda544cb7e3359.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": " are the parameters of the teacher and student networks, respectively, and " + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": "-th training step. " + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": " is the hyper-parameter to determine the speed of parameter transmission, which is generally close to 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 483, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 128, + 498 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 503, + 110, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 110, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 110, + 515 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "type": "text", + "content": "We evaluate the proposed approach on two publicly available benchmark datasets, S3DIS [3] and ScanNetV2 [9], to demonstrate its superior performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 665 + ], + "type": "text", + "content": "S3DIS (Stanford Large-Scale 3D Indoor Space Dataset) is a large-scale indoor point cloud dataset captured by Matter-port scanners which mainly dedicate to 3D segmentation tasks. It consists of 6 different large-scale indoor areas with a total of 271 rooms, each room is a separate point cloud sample that includes coordinate, colour, and annotation information for each point. Following the standard evaluation criterion, Area 5 is used as the test set and the rest of the areas are used as the training set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "ScanNetV2 is an RGB-D video dataset. It contains 1513 point cloud samples from 707 individual indoor scenes for training and 100 unlabelled samples for testing. For the 3D semantic segmentation task, ScanNetV2 provides up to 40" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "categories of classification labels and selects 20 categories as classification targets to establish the segmentation task. Each point in the dataset contains coordinate, colour, and annotation information. In the experiments, we follow the official data split, dividing the 1513 training data into 1201 training samples and 312 validation samples, and evaluate methods on the test set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 164, + 440, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 440, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 440, + 177 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 183, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 243 + ], + "type": "text", + "content": "Segmentation Model. In our experiments, we use the PyTorch implementation of MinkowskiNet in ReDAL [41] as our backbone network with the stochastic gradient descent (SGD) optimizer. The architecture is trained on a single NVIDIA RTX A6000 GPU." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "text", + "content": "Training Settings. For S3DIS, we set the batch size to 4, and train 60K steps per iteration. The initial learning rate is 0.1, and cosine annealing with a period of 5000 is used as the learning rate decay strategy. For ScanNetV2, we set the batch size to 4, and train 30K steps per iteration. The initial learning rate is 0.1, and the learning rate decay strategy is using polynomial decay with a power of 0.9. In the teacher-student model, the keep rate of EMA and the thresholds for the pseudo-label are set to 0.955 and 0.75, respectively. 5 iterations are performed in the active learning framework. In HMMU, the point cloud is downsampled at each scale with the sampling radius of " + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "inline_equation", + "content": "50\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "inline_equation", + "content": "100\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "text", + "content": ", respectively. In the FDS module, the radius of the local region is set to " + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "inline_equation", + "content": "20\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 243, + 546, + 422 + ], + "type": "text", + "content": " and the feature similarity threshold is set to 0.8." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 431, + 531, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 531, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 531, + 444 + ], + "type": "text", + "content": "4.3. Comparison with Active Learning Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our active learning strategy, we remove the semi-supervised module in the comparisons with active learning methods. As current active learning methods for point cloud semantic segmentation are region-based, we compare our proposed method against them to verify the effectiveness of the point-based active learning method. The compared results are shown in Table 1. Compared with the region-based active learning methods, our point-based selection strategy greatly reduces the labelling cost, with the demand of achieving " + }, + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "text", + "content": " performance of the fully-supervised baselines. With only " + }, + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "text", + "content": " (about 13,000 points) of the downsampled points, our method allows for " + }, + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 653 + ], + "type": "text", + "content": " of the performance of our fully supervised baseline, i.e. fully-supervised MinkowskiNet. This verifies our theory that the strong semantic similarity in local areas of point clouds results in redundant labelling and ineffective training in region-based active learning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "Furthermore, we compare the performance of our active learning strategy with existing region-based ones and four basic point-based selection strategies including random selection (Random), softmax least confidence (LC), softmax minimum margin uncertainty (MMU) and softmax entropy" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18102" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 272, + 129 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 272, + 129 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 272, + 129 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 272, + 129 + ], + "type": "table", + "html": "
MethodsSegmentorsLabelled points
ReDAL [41]SPVCNN13%
ReDAL [41]MinkowskiNet15%
SSDR-AL [29]Randlanet11.7%
OursMinkowskiNet0.1%
", + "image_path": "afcbbc8cdb50b6641248a5dd8a6d2c752915d3fff7ea7c93e0693bfa193be02f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 131, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 131, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 131, + 287, + 198 + ], + "type": "text", + "content": "Table 1. The comparison of the percentage of labelled points required to achieve " + }, + { + "bbox": [ + 46, + 131, + 287, + 198 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 131, + 287, + 198 + ], + "type": "text", + "content": " accuracy of fully-supervised baseline on the S3DIS dataset against current region-based active learning methods. For a fair comparison, the semi-supervised module of our method is removed here which means only labelled data and selected points are used during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "text", + "content": "(Entropy). The results are shown in Table 2. It can be seen that the performance of our method is increased by " + }, + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "text", + "content": " compared with random point selection and surpasses the state-of-the-art region-based active learning method by " + }, + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "text", + "content": " with significantly fewer labelled points (0.1% v.s. 11.7%)." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 275, + 286, + 359 + ], + "blocks": [ + { + "bbox": [ + 50, + 275, + 286, + 359 + ], + "lines": [ + { + "bbox": [ + 50, + 275, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 50, + 275, + 286, + 359 + ], + "type": "table", + "html": "
TypeMethodsSegmentorslabelled pointsmIoU(%)
Region-basedReDALSPVCNN15%58.0
ReDALMinkowskiNet15%57.3
SSDR-ALRandlanet11.7%58.3
Point-basedRandomMinkowskiNet0.1%54.7
EntropyMinkowskiNet0.1%48.2
LCMinkowskiNet0.1%53.3
MMUMinkowskiNet0.1%55.5
OursMinkowskiNet0.1%59.3
", + "image_path": "e40a29aedde36ad62f66b076ca716754c15395098d5f763322f7d6c174179d06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 360, + 286, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 392 + ], + "type": "text", + "content": "Table 2. The comparison of mIoU between our active strategy without the semi-supervised module and different active learning methods on the S3DIS dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 285, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 285, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 285, + 420 + ], + "type": "text", + "content": "4.4. Comparison with Weakly-supervised Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 498 + ], + "type": "text", + "content": "We also compare the proposed method with the state-of-the-art weakly-supervised methods. Similar to [22, 48], we present state-of-the-art segmentation methods with different supervision settings. The comparison results on the datasets of S3DIS and ScannetV2 are presented in Table 3 and Table 4, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": "S3DIS. we perform experiments under three different annotation budgets, namely " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": ". For the sake of fairness, we follow the SQN [13] to calculate the annotation percentage which is defined as the ratio between the number of labelled data and all training and validation data. Each annotation budget is completed through five iterations. As shown in Table 3, our approach gains consistent improvements in the segmentation performance with labelled points increased from " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.43\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": ". It is worth noting that, for the " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " annotation budget, the segmentation performance at the third iteration (" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.43\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " annotations) has already surpassed the fully-supervised MinkowskiNet. This verifies that selecting the most representative points and making full use of the remaining unlabelled data can greatly benefit the performance of point cloud semantic segmentation. Our method also outperforms the state-of-the-art PSD, SQN and HybridCR under " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " labelled data by " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": ", respectively. When the number of labelled" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 531, + 274 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 531, + 274 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 531, + 274 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 531, + 274 + ], + "type": "table", + "html": "
MethodsLabelled pointsArea 5
PointNet [27]100%41.1
PointCNN [23]100%57.3
KPConv [35]100%67.1
PointTransformer [54]100%70.4
CBL [32]100%69.4
MinkowskiNet†[7]100%64.5
PSD [51]1%63.5
SQN [13]1%63.7
HybridCR [22]1%65.3
Ours0.43%65.7
SQN [13]0.1%61.4
Ours0.07%62.3
PSD [51]0.03%48.2
HybridCR [22]0.03%51.5
MIL [48]0.02%51.4
GaIA [21]0.02%53.7
VIB [36]0.02%52.0
Ours0.02%55.9
", + "image_path": "483da21a8fb555f6eed08b83abc17ef521275a0c7cd6e9f978c18251052f82ac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 274, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 545, + 305 + ], + "type": "text", + "content": "Table 3. Comparison with existing weakly-supervised methods on S3DIS Area-5. † represents the result of the baseline trained on our own device." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": "points is reduced to " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": ", our method is also superior to the SQN with " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": " labelled points and achieves " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "96.5\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": " performance of the fully-supervised baseline (MinkowskiNet). In the case of the least labelling budget, we evaluate our method in the setting of " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": " labelled points. It can be seen that our method also achieves the best performance, outperforming the state-of-the-art GaIA, MIL, HybridCR and PSD by " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "type": "text", + "content": "We also visualize the segmentation results of our method under different annotation budgets in Figure 3. The results show that with " + }, + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "type": "inline_equation", + "content": "0.43\\%" + }, + { + "bbox": [ + 304, + 414, + 545, + 497 + ], + "type": "text", + "content": " labelled data, our approach can achieve comparable segmentation results with the fully supervised baseline. Benefiting from the points selected by our method, our segmentation results are even more promising in some cases, such as the bookcase and column." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "ScanNetV2. As illustrated in Table 4, we use the criteria of percentage and fixed number of labelled points as different annotation budgets to evaluate the methods, including " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", 200pts, and 20pts. Similar to S3DIS, each annotation budget completes through five iterations, and the third iteration (120pts) is reported in the case of 200pts. We substantially beat SQN's performance by " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "11.3\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " of annotation. Compared against PSD, HybridCR and GaIA which are trained with " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " annotations, our method outperforms them with only " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " labels by " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "13.5\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ", respectively. In the case of the fixed number of annotation points, we use 200 points per scene (200pts) and 20 points per scene (20pts) as the annotation budget. In our experiment with 200pts, when the labelling points reach 120pts, our method has achieved comparable performance to the fully supervised baseline. This proves the effectiveness of our active learning strategy on the semi-supervised point cloud semantic segmentation. We also evaluate our" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18103" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 78, + 133, + 129 + ], + "blocks": [ + { + "bbox": [ + 53, + 78, + 133, + 129 + ], + "lines": [ + { + "bbox": [ + 53, + 78, + 133, + 129 + ], + "spans": [ + { + "bbox": [ + 53, + 78, + 133, + 129 + ], + "type": "image", + "image_path": "519bae80d21186449d66858167662bdc2249649c6d7be3ae3096547a9b6c43e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 208, + 478, + 218 + ], + "lines": [ + { + "bbox": [ + 121, + 208, + 478, + 218 + ], + "spans": [ + { + "bbox": [ + 121, + 208, + 478, + 218 + ], + "type": "text", + "content": "ceiling floor wall column window door chair table bookcase board clutter" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 130, + 133, + 180 + ], + "blocks": [ + { + "bbox": [ + 53, + 130, + 133, + 180 + ], + "lines": [ + { + "bbox": [ + 53, + 130, + 133, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 133, + 180 + ], + "type": "image", + "image_path": "183656d7b8788898e35d22e22cdd9a74fd02c1352f28775cf9ab06dc1b955d38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 191, + 110, + 201 + ], + "lines": [ + { + "bbox": [ + 78, + 191, + 110, + 201 + ], + "spans": [ + { + "bbox": [ + 78, + 191, + 110, + 201 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 135, + 78, + 214, + 130 + ], + "blocks": [ + { + "bbox": [ + 135, + 78, + 214, + 130 + ], + "lines": [ + { + "bbox": [ + 135, + 78, + 214, + 130 + ], + "spans": [ + { + "bbox": [ + 135, + 78, + 214, + 130 + ], + "type": "image", + "image_path": "b5dc9b1d54fa4d239bbf6abbe761d2d1c2fb693e98f1fc9e1761c6e2a9313f9a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 131, + 214, + 180 + ], + "blocks": [ + { + "bbox": [ + 135, + 131, + 214, + 180 + ], + "lines": [ + { + "bbox": [ + 135, + 131, + 214, + 180 + ], + "spans": [ + { + "bbox": [ + 135, + 131, + 214, + 180 + ], + "type": "image", + "image_path": "7ca96052ef21033b5bf642718808a1e413f4ed71eb88a3e799ecaf296491f55b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 191, + 205, + 201 + ], + "lines": [ + { + "bbox": [ + 144, + 191, + 205, + 201 + ], + "spans": [ + { + "bbox": [ + 144, + 191, + 205, + 201 + ], + "type": "text", + "content": "(b) Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 78, + 295, + 130 + ], + "blocks": [ + { + "bbox": [ + 216, + 78, + 295, + 130 + ], + "lines": [ + { + "bbox": [ + 216, + 78, + 295, + 130 + ], + "spans": [ + { + "bbox": [ + 216, + 78, + 295, + 130 + ], + "type": "image", + "image_path": "8b235e45e92fd9982c177383025f787de549cda90cb8017b04d3d579b411d19f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 216, + 131, + 295, + 181 + ], + "blocks": [ + { + "bbox": [ + 216, + 131, + 295, + 181 + ], + "lines": [ + { + "bbox": [ + 216, + 131, + 295, + 181 + ], + "spans": [ + { + "bbox": [ + 216, + 131, + 295, + 181 + ], + "type": "image", + "image_path": "83a618d9ecd74f84398f4cf2fa7c839dd822eed7194447c9ccd9510ee850ab64.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 191, + 294, + 201 + ], + "lines": [ + { + "bbox": [ + 224, + 191, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 224, + 191, + 294, + 201 + ], + "type": "text", + "content": "(c) Fully-supervised" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 296, + 78, + 376, + 130 + ], + "blocks": [ + { + "bbox": [ + 296, + 78, + 376, + 130 + ], + "lines": [ + { + "bbox": [ + 296, + 78, + 376, + 130 + ], + "spans": [ + { + "bbox": [ + 296, + 78, + 376, + 130 + ], + "type": "image", + "image_path": "4d08a212ffa052a2ec22ce8079b72d6a63718128d7eae9d7ba533d9b01d9d3d8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 298, + 131, + 376, + 181 + ], + "blocks": [ + { + "bbox": [ + 298, + 131, + 376, + 181 + ], + "lines": [ + { + "bbox": [ + 298, + 131, + 376, + 181 + ], + "spans": [ + { + "bbox": [ + 298, + 131, + 376, + 181 + ], + "type": "image", + "image_path": "9ca709480bdfc169dfd7ba13f54f61e2ca3853399073a71b161fa53f65cae22d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 191, + 367, + 201 + ], + "lines": [ + { + "bbox": [ + 307, + 191, + 367, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 367, + 201 + ], + "type": "text", + "content": "(d) Ours " + }, + { + "bbox": [ + 307, + 191, + 367, + 201 + ], + "type": "inline_equation", + "content": "(0.02\\%)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 378, + 78, + 459, + 130 + ], + "blocks": [ + { + "bbox": [ + 378, + 78, + 459, + 130 + ], + "lines": [ + { + "bbox": [ + 378, + 78, + 459, + 130 + ], + "spans": [ + { + "bbox": [ + 378, + 78, + 459, + 130 + ], + "type": "image", + "image_path": "c08f3e6a3257c0bacafeeaa0aa429ffc927389ed668278863a286a1df7252c50.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 378, + 131, + 458, + 181 + ], + "blocks": [ + { + "bbox": [ + 378, + 131, + 458, + 181 + ], + "lines": [ + { + "bbox": [ + 378, + 131, + 458, + 181 + ], + "spans": [ + { + "bbox": [ + 378, + 131, + 458, + 181 + ], + "type": "image", + "image_path": "408d87ec47a483138c9fab8dd551491e2101a5ab479db8bd08deb34de316f631.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 191, + 448, + 201 + ], + "lines": [ + { + "bbox": [ + 389, + 191, + 448, + 201 + ], + "spans": [ + { + "bbox": [ + 389, + 191, + 448, + 201 + ], + "type": "text", + "content": "(e) Ours " + }, + { + "bbox": [ + 389, + 191, + 448, + 201 + ], + "type": "inline_equation", + "content": "(0.07\\%)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 460, + 78, + 540, + 130 + ], + "blocks": [ + { + "bbox": [ + 460, + 78, + 540, + 130 + ], + "lines": [ + { + "bbox": [ + 460, + 78, + 540, + 130 + ], + "spans": [ + { + "bbox": [ + 460, + 78, + 540, + 130 + ], + "type": "image", + "image_path": "34b1278c1896af7e5465f113abefce71e5ec62ea716a7f2cd23246a491a1070b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 460, + 131, + 540, + 181 + ], + "blocks": [ + { + "bbox": [ + 460, + 131, + 540, + 181 + ], + "lines": [ + { + "bbox": [ + 460, + 131, + 540, + 181 + ], + "spans": [ + { + "bbox": [ + 460, + 131, + 540, + 181 + ], + "type": "image", + "image_path": "b29481131af0468405bccacc07b28559c13c49f5562f63956cb55c2a60c363a5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 191, + 529, + 201 + ], + "lines": [ + { + "bbox": [ + 471, + 191, + 529, + 201 + ], + "spans": [ + { + "bbox": [ + 471, + 191, + 529, + 201 + ], + "type": "text", + "content": "(f) Ours " + }, + { + "bbox": [ + 471, + 191, + 529, + 201 + ], + "type": "inline_equation", + "content": "(0.43\\%)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 56, + 262, + 136, + 313 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 546, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 546, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 546, + 250 + ], + "type": "text", + "content": "Figure 3. Visualization of segmentation results on the test set of S3DIS Area-5. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when " + }, + { + "bbox": [ + 46, + 228, + 546, + 250 + ], + "type": "inline_equation", + "content": "0.43\\%" + }, + { + "bbox": [ + 46, + 228, + 546, + 250 + ], + "type": "text", + "content": " of the labelled data is adopted." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 262, + 136, + 313 + ], + "lines": [ + { + "bbox": [ + 56, + 262, + 136, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 136, + 313 + ], + "type": "image", + "image_path": "abdfd87f8f294378cac478211b2a023b21df1003270a1f4330b373b1dcad5ba9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 56, + 314, + 136, + 363 + ], + "blocks": [ + { + "bbox": [ + 56, + 314, + 136, + 363 + ], + "lines": [ + { + "bbox": [ + 56, + 314, + 136, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 136, + 363 + ], + "type": "image", + "image_path": "53df65a7fb776c9145d1ddc88ee1bb02479172c43bee91222e147391a50628b4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 373, + 112, + 383 + ], + "lines": [ + { + "bbox": [ + 81, + 373, + 112, + 383 + ], + "spans": [ + { + "bbox": [ + 81, + 373, + 112, + 383 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 388, + 541, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 541, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 541, + 400 + ], + "type": "text", + "content": "wall floor cabinet chair sofa table door bookshelf counter desk window sink curtain bed other furniture" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 137, + 262, + 216, + 312 + ], + "blocks": [ + { + "bbox": [ + 137, + 262, + 216, + 312 + ], + "lines": [ + { + "bbox": [ + 137, + 262, + 216, + 312 + ], + "spans": [ + { + "bbox": [ + 137, + 262, + 216, + 312 + ], + "type": "image", + "image_path": "f63f279ee846a0b46ec530267755cccac3adcc1e7793102621ac465942e87fc2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 138, + 313, + 215, + 364 + ], + "blocks": [ + { + "bbox": [ + 138, + 313, + 215, + 364 + ], + "lines": [ + { + "bbox": [ + 138, + 313, + 215, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 215, + 364 + ], + "type": "image", + "image_path": "13aae238025387ca61900aa445135d86cc7f2f58fc7419d01a356cc970bf45c5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 373, + 206, + 382 + ], + "lines": [ + { + "bbox": [ + 146, + 373, + 206, + 382 + ], + "spans": [ + { + "bbox": [ + 146, + 373, + 206, + 382 + ], + "type": "text", + "content": "(b) Ground Truth" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 216, + 262, + 294, + 312 + ], + "blocks": [ + { + "bbox": [ + 216, + 262, + 294, + 312 + ], + "lines": [ + { + "bbox": [ + 216, + 262, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 216, + 262, + 294, + 312 + ], + "type": "image", + "image_path": "2dd32fc6eda9c971b434907dbba655470dc5bc3723380961c34f7da4c3986a37.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 216, + 313, + 294, + 363 + ], + "blocks": [ + { + "bbox": [ + 216, + 313, + 294, + 363 + ], + "lines": [ + { + "bbox": [ + 216, + 313, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 216, + 313, + 294, + 363 + ], + "type": "image", + "image_path": "97e3c0e02c4ebe39d150711c8eece58eeeb3ccdbba83e0ef3e0bedbbec7df215.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 373, + 293, + 383 + ], + "lines": [ + { + "bbox": [ + 223, + 373, + 293, + 383 + ], + "spans": [ + { + "bbox": [ + 223, + 373, + 293, + 383 + ], + "type": "text", + "content": "(c) Fully-supervised" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 295, + 262, + 373, + 312 + ], + "blocks": [ + { + "bbox": [ + 295, + 262, + 373, + 312 + ], + "lines": [ + { + "bbox": [ + 295, + 262, + 373, + 312 + ], + "spans": [ + { + "bbox": [ + 295, + 262, + 373, + 312 + ], + "type": "image", + "image_path": "48ea1038526256f503990dc79d97e53c89d14d850815d8fa7649097c5da86d4e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 296, + 313, + 373, + 363 + ], + "blocks": [ + { + "bbox": [ + 296, + 313, + 373, + 363 + ], + "lines": [ + { + "bbox": [ + 296, + 313, + 373, + 363 + ], + "spans": [ + { + "bbox": [ + 296, + 313, + 373, + 363 + ], + "type": "image", + "image_path": "85203a59f32ddd4ab34f2aaa47ce30a06d9081210229b3557734f99293668dc4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 373, + 362, + 383 + ], + "lines": [ + { + "bbox": [ + 307, + 373, + 362, + 383 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 362, + 383 + ], + "type": "text", + "content": "(d) Ours (20pts)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 413, + 546, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 546, + 436 + ], + "type": "text", + "content": "Figure 4. Visualization of segmentation results on the validation set of ScannNetV2. Our method achieves comparable or even better results than our fully-supervised baseline (MinkowskiNet) when 120 labelled points per scene are adopted." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 376, + 262, + 454, + 312 + ], + "blocks": [ + { + "bbox": [ + 376, + 262, + 454, + 312 + ], + "lines": [ + { + "bbox": [ + 376, + 262, + 454, + 312 + ], + "spans": [ + { + "bbox": [ + 376, + 262, + 454, + 312 + ], + "type": "image", + "image_path": "1560259c80f5ef5844010b57f5acff2dd73ee6c6979a897c1eb26235fb929a86.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 377, + 313, + 454, + 363 + ], + "blocks": [ + { + "bbox": [ + 377, + 313, + 454, + 363 + ], + "lines": [ + { + "bbox": [ + 377, + 313, + 454, + 363 + ], + "spans": [ + { + "bbox": [ + 377, + 313, + 454, + 363 + ], + "type": "image", + "image_path": "212c0cd5954a32b40ed744be2dfcf8bb64288bb89f839a72f3decbc6710f6939.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 373, + 442, + 382 + ], + "lines": [ + { + "bbox": [ + 388, + 373, + 442, + 382 + ], + "spans": [ + { + "bbox": [ + 388, + 373, + 442, + 382 + ], + "type": "text", + "content": "(e) Ours " + }, + { + "bbox": [ + 388, + 373, + 442, + 382 + ], + "type": "inline_equation", + "content": "(0.1\\%)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 455, + 262, + 534, + 312 + ], + "blocks": [ + { + "bbox": [ + 455, + 262, + 534, + 312 + ], + "lines": [ + { + "bbox": [ + 455, + 262, + 534, + 312 + ], + "spans": [ + { + "bbox": [ + 455, + 262, + 534, + 312 + ], + "type": "image", + "image_path": "973ac929ba247559605112f64640899df48c6c0536824148525b4b1087fdbf5a.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 455, + 313, + 534, + 363 + ], + "blocks": [ + { + "bbox": [ + 455, + 313, + 534, + 363 + ], + "lines": [ + { + "bbox": [ + 455, + 313, + 534, + 363 + ], + "spans": [ + { + "bbox": [ + 455, + 313, + 534, + 363 + ], + "type": "image", + "image_path": "aadae06adf078e22a84d0a8cb68aa24c8ea6eb1b237dd138f5743815c6875625.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 373, + 524, + 382 + ], + "lines": [ + { + "bbox": [ + 465, + 373, + 524, + 382 + ], + "spans": [ + { + "bbox": [ + 465, + 373, + 524, + 382 + ], + "type": "text", + "content": "(f) Ours (120pts)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 62, + 442, + 272, + 664 + ], + "blocks": [ + { + "bbox": [ + 62, + 442, + 272, + 664 + ], + "lines": [ + { + "bbox": [ + 62, + 442, + 272, + 664 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 272, + 664 + ], + "type": "table", + "html": "
MethodsLabelled pointsvaltest
PointNet++ [28]100%-33.9
KPCov [35]100%-68.4
VMNet [16]100%-74.6
BPNet [15]100%-74.9
MinkowskiNet†[7]100%69.368.0
PSD [51]1%-54.7
HybridCR [22]1%56.956.8
GaIA [21]1%-65.2
SQN [13]0.1%-56.9
Ours0.1%69.968.2
SQN [13]200pts-59.8
CSC [12]200pts68.266.5
GaIA [21]200pts-68.5
VIBUS [36]200pts69.669.1
Ours120pts70.269.4
SQN [13]20pts-48.6
CSC [12]20pts55.553.1
MIL [48]20pts57.854.4
VIBUS [36]20pts-58.6
GaIA [21]20pts-63.8
Ours20pts62.262.5
", + "image_path": "66ec10cfb4e04ea035f96d28cd8d82e5c43b16244f6f4f1508520b6be78d58f2.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 664, + 286, + 685 + ], + "lines": [ + { + "bbox": [ + 46, + 664, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 286, + 685 + ], + "type": "text", + "content": "Table 4. The comparison with existing weakly-supervised methods. " + }, + { + "bbox": [ + 46, + 664, + 286, + 685 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 46, + 664, + 286, + 685 + ], + "type": "text", + "content": " is the result of the baseline trained on our device." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "method on the least annotation setting. In the case of 20pts, our result is much higher than SQN, CSC, MIL," + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 445, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 503 + ], + "type": "text", + "content": "and VIBUS, showing that our method effectively selects the most valuable points. However, our result is slightly lower than GaIA, as the performance of our feature extraction backbone is limited given only 20pts (the proportion is less than " + }, + { + "bbox": [ + 304, + 445, + 545, + 503 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 304, + 445, + 545, + 503 + ], + "type": "text", + "content": " of ScanNetV2)." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 506, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 565 + ], + "type": "text", + "content": "As shown in Fig. 4, we visualize the segmentation results of our method under different annotation budgets on ScanNetV2. It is observed that our method can achieve comparable or even better segmentation results with minor training budgets compared to our fully-supervised baseline." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 574, + 406, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 406, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 406, + 586 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "To demonstrate the effectiveness of each module in our method, we conduct the following ablation studies on S3DIS dataset. Following [13], all the ablated networks are trained using " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " labelled points on Areas 1, 2, 3, 4, 6 and tested on Area 5. The number of iterations is set to 5 and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " data is selected in each iteration. To clearly see the effect of active learning, we present the results as shown in Figure 5. We take random sampling as the basic strategy for active selection where points are randomly selected in each iteration. After we replace the random sam" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18104" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 262, + 247 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 262, + 247 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 262, + 247 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 262, + 247 + ], + "type": "image", + "image_path": "fd66a2ae14a1b6265f3ab4f1741216d10c5eb4cc6e8d8f5c861822a7779e4565.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 287, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 275 + ], + "type": "text", + "content": "Figure 5. The comparison of performance improvements for different combinations of modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": "pling with MMU for active sampling, the mIoU is slightly improved by " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": ". After utilising HMMU to select points, we acquire an improvement of around " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " compared with MMU in the last iteration. The FDS module is also added to further improve the performance. HMMU+FDS achieves an extra " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " mIoU improvement compared to HMMU. More impressively, merely relying on the active learning modules including HMMU and FDS, our approach already outperforms the random sampling baseline by " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " and surpasses " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " performance of full-supervision with only " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " labelling data. The effectiveness of the semi-supervised module is also demonstrated by the experimental results of HMMU+FDS+TS. Obviously, adding the semi-supervised teacher-student module, the segmentation result is further improved by " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " compared to HMMU + FDS. The proposed method also reaches " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " of the fully-supervised setting with only " + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 282, + 287, + 486 + ], + "type": "text", + "content": " labelled data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "content": "In Table 5, we show the effectiveness of different modules by quantitative comparison. We compare different compositions of our module to verify their effectiveness. Shown in the first row, for the base case where unlabelled points are selected by MMU in each iteration, the mIoU is only " + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "inline_equation", + "content": "55.51\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "content": ". After using the HMMU module, the segmentation results are improved to " + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "inline_equation", + "content": "57.65\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "content": ". If we only apply FDS to the MMU selection for each iteration, a slight improvement of " + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "inline_equation", + "content": "1.17\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "content": " is obtained compared with the based case. When the active learning strategy is removed, as shown in the fourth row in Table 5, the semi-supervised segmentation method only achieves the performance of " + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "inline_equation", + "content": "57.01\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "content": ". Point cloud semantic segmentation only using the active learning technique can reach the mIoU of " + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "inline_equation", + "content": "59.29\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "content": ". If we combine semi-supervised and active learning, as shown in the last row, the performance can be improved to " + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "inline_equation", + "content": "62.26\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In Table 6, we show the ablation study of each layer in the HMMU module. PL represents the layer that calculates" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 337, + 69, + 503, + 155 + ], + "blocks": [ + { + "bbox": [ + 337, + 69, + 503, + 155 + ], + "lines": [ + { + "bbox": [ + 337, + 69, + 503, + 155 + ], + "spans": [ + { + "bbox": [ + 337, + 69, + 503, + 155 + ], + "type": "table", + "html": "
ComponentsmIoU(%)
HMMUFDSTS
base.55.51
✓57.65
✓56.68
✓57.01
✓✓59.29
✓✓✓62.26
", + "image_path": "6a8f10f635a846d07a7bed5b9268dc39ad56b8ca38ef1267f2ae436fc02f6a40.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": "the point-level minimum margin uncertainty score. VL is the layer which calculates the voxel-level contextual uncertainty score. " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " are the voxel-radius and weight for each VL layer. It is shown that the combination of one layer of point-level uncertainty and three layers of voxel-level uncertainty can achieve the best performance. We also show the ablation study on the hyperparameters of the FDS module in Table 7. We achieve the best segmentation result given that " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "20\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " and 0.8 respectively. Thus, we adopt this combination in our experiments." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 304, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 337, + 155, + 514, + 164 + ], + "lines": [ + { + "bbox": [ + 337, + 155, + 514, + 164 + ], + "spans": [ + { + "bbox": [ + 337, + 155, + 514, + 164 + ], + "type": "text", + "content": "Table 5. Ablation study of different components." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 304, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 371 + ], + "type": "table", + "html": "
LayerHMMUiter 1iter 2iter 3iter 4iter 5
ωvr(cm)
PL--37.8350.3754.3256.1856.68
PL + VL1{0.1}{10}37.8352.7154.8456.5257.03
PL + VL2{0.01}{50}37.8351.4555.8157.2257.43
PL + VL3{0.001}{100}37.8351.0655.6257.3658.03
PL + VL1,2{0.1, 0.01}{10, 50}37.8351.5655.2856.6658.21
PL + VL1,2,3{0.1, 0.01, 0.001}{10, 50, 100}37.8352.0256.0258.4959.29
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.001}{10, 50, 100, 200}37.8351.5656.3257.6157.71
PL + VL1,2,3,4{0.1, 0.01, 0.001, 0.0001}{10, 50, 100, 200}37.8351.6956.2258.5359.19
", + "image_path": "78f310caedf015c35e6f4429cabe4cd6a995fddf7f8c93684f7c06e93cc06526.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 390, + 547, + 457 + ], + "blocks": [ + { + "bbox": [ + 350, + 371, + 501, + 380 + ], + "lines": [ + { + "bbox": [ + 350, + 371, + 501, + 380 + ], + "spans": [ + { + "bbox": [ + 350, + 371, + 501, + 380 + ], + "type": "text", + "content": "Table 6. Ablation study of HMMU layers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 390, + 547, + 457 + ], + "lines": [ + { + "bbox": [ + 308, + 390, + 547, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 390, + 547, + 457 + ], + "type": "table", + "html": "
FDSiteration1iteration2iteration3iteration4iteration5
r(cm)τ
200.837.8352.0256.0258.4959.29
200.937.8351.2754.9056.7958.79
400.837.8351.3555.7157.0957.79
400.937.8351.6555.2657.4559.13
500.9537.8350.8755.7356.4957.82
", + "image_path": "47222fd4f8b5afc9dd0d13a9dd6df501212f7f4c524964762ffeb0b353f979b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 307, + 457, + 537, + 466 + ], + "lines": [ + { + "bbox": [ + 307, + 457, + 537, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 457, + 537, + 466 + ], + "type": "text", + "content": "Table 7. Ablation study of the hyperparameters in FDS module." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 478, + 378, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 378, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 378, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 689 + ], + "type": "text", + "content": "In this paper, we study active learning for 3D point cloud semantic segmentation. In comparison with previous region-based methods, we propose a hierarchical point-based active learning strategy consisting of two new designs, HMMU and FDS. HMMU serves as an effective way to measure the uncertainty or importance of labelling, while FDS enables us to select the most valuable points by considering the feature similarity and spatial distribution. To better use the unlabelled points during network learning, we also introduce a teacher-student structure to generate pseudo-labels. Extensive experiments are conducted on S3DIS and ScanNet datasets to demonstrate the effectiveness of our method. In the future, a more advanced semi-supervised architecture can be adopted for generating more reliable pseudo-labels to further facilitate active learning based point cloud semantic segmentation tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the National Natural Science Foundation of China (No. 62206033," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18105" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "No.62036007, No.62221005), and the Natural Science Foundation of Chongqing (No. cstc2020jcyj-msxmX0855, No. cstc2021ycjh-bgzxm0339)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 106, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 106, + 131 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 106, + 131 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 137, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 192 + ], + "type": "text", + "content": "[1] Rashid Abbasi, Ali Kashif Bashir, Hasan J Alyamani, Farhan Amin, Jaehyeok Doh, and Jianwen Chen. Lidar point cloud compression, processing and learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 24(1):962-979, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[2] Eren Erdal Aksoy, Saimir Baci, and Selcuk Cavdar. Salsanet: Fast road and vehicle segmentation in lidar point clouds for autonomous driving. In 2020 IEEE intelligent vehicles symposium (IV), pages 926-932. IEEE, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 288, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 288, + 292 + ], + "type": "text", + "content": "[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 293, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 287, + 335 + ], + "type": "text", + "content": "[4] Alexandre Boulch, Bertrand Le Saux, Nicolas Audebert, et al. Unstructured point cloud semantic labeling using deep segmentation networks. 3dor@eurographics, 3:17-24, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 337, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 287, + 381 + ], + "type": "text", + "content": "[5] Mingmei Cheng, Le Hui, Jin Xie, and Jian Yang. Sspc-net: Semi-supervised semantic 3d point cloud segmentation network. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 1140-1147, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 382, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 287, + 446 + ], + "type": "text", + "content": "[6] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 448, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 287, + 502 + ], + "type": "text", + "content": "[7] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 503, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 287, + 569 + ], + "type": "text", + "content": "[8] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5-7, 2020, Proceedings, Part II 15, pages 207-222. Springer, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 287, + 624 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 678 + ], + "type": "text", + "content": "[10] Shuang Deng, Qiulei Dong, Bo Liu, and Zhanyi Hu. Superpoint-guided semi-supervised semantic segmentation of 3d point clouds. In 2022 International Conference on Robotics and Automation (ICRA), pages 9214-9220. IEEE, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587-15597, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 217 + ], + "type": "text", + "content": "[13] Qingyong Hu, Bo Yang, Guangchi Fang, Yulan Guo, Ales Leonardis, Niki Trigoni, and Andrew Markham. Sqn: Weakly-supervised semantic segmentation of large-scale 3d point clouds. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 600-619. Springer, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "type": "text", + "content": "[14] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11108-11117, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "text", + "content": "[15] Wenbo Hu, Hengshuang Zhao, Li Jiang, Jiaya Jia, and Tien-Tsin Wong. Bidirectional projection network for cross dimension scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14373-14382, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "text", + "content": "[16] Zeyu Hu, Xuyang Bai, Jiaxiang Shang, Runze Zhang, Jiayu Dong, Xin Wang, Guangyuan Sun, Hongbo Fu, and Chiew-Lan Tai. Vmnet: Voxel-mesh network for geodesic-aware 3d semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15488-15498, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[17] Ankang Ji, Yunxiang Zhou, Limao Zhang, Robert LK Tiong, and Xiaolong Xue. Semi-supervised learning-based point cloud network for segmentation of 3d tunnel scenes. Automation in Construction, 146:104668, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "text", + "content": "[18] Li Jiang, Shaoshuai Shi, Zhuotao Tian, Xin Lai, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Guided point contrastive learning for semi-supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6423-6432, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "type": "text", + "content": "[19] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 518-535. Springer, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 570, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 547, + 624 + ], + "type": "text", + "content": "[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "text", + "content": "[21] Min Seok Lee, Seok Woo Yang, and Sung Won Han. Gaia: Graphical information gain based attention network for weakly supervised point cloud semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 582-591, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[22] Mengtian Li, Yuan Xie, Yunhang Shen, Bo Ke, Ruizhi Qiao, Bo Ren, Shaohui Lin, and Lizhuang Ma. Hybridcr: Weakly-supervised 3d point cloud semantic segmentation via hybrid" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18106" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "contrastive regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14930-14939, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[24] Y Lin, G Vosselman, Y Cao, and MY Yang. Efficient training of semantic point cloud segmentation via active learning. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2:243-250, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 252 + ], + "type": "text", + "content": "[25] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1736, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "text", + "content": "[26] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 4213-4220. IEEE, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 363 + ], + "type": "text", + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 410 + ], + "type": "text", + "content": "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 465 + ], + "type": "text", + "content": "[29] Feifei Shao, Yawei Luo, Ping Liu, Jie Chen, Yi Yang, Yulei Lu, and Jun Xiao. Active learning for point cloud semantic segmentation via spatial-structural diversity reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2575–2585, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 510 + ], + "type": "text", + "content": "[30] Xian Shi, Xun Xu, Ke Chen, Lile Cai, Chuan Sheng Foo, and Kui Jia. Label-efficient point cloud semantic segmentation: An active learning approach. arXiv preprint arXiv:2101.06931, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 578 + ], + "type": "text", + "content": "[31] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII, pages 685–702. Springer, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[32] Liyao Tang, Yibing Zhan, Zhe Chen, Baosheng Yu, and Dacheng Tao. Contrastive boundary learning for point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8489-8499, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[33] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 2017 international conference on 3D vision (3DV), pages 537-547. IEEE, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[34] Hugues Thomas, Ben Agro, Mona Gridseth, Jian Zhang, and Timothy D Barfoot. Self-supervised learning of lidar segmentation for autonomous indoor navigation. In 2021" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE International Conference on Robotics and Automation (ICRA), pages 14047-14053. IEEE, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 97, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 547, + 152 + ], + "type": "text", + "content": "[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "type": "text", + "content": "[36] Beiwen Tian, Liyi Luo, Hao Zhao, and Guyue Zhou. Vibus: Data-efficient 3d scene parsing with viewpoint bottleneck and uncertainty-spectrum modeling. ISPRS Journal of Photogrammetry and Remote Sensing, 194:302-318, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 242 + ], + "type": "text", + "content": "[37] Haiyan Wang, Xuejian Rong, Liang Yang, Jinglun Feng, Jizhong Xiao, and Yingli Tian. Weakly supervised semantic segmentation in 3d graph-structured point clouds of wild scenes. arXiv preprint arXiv:2004.12498, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 297 + ], + "type": "text", + "content": "[38] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10296-10305, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 300, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 300, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 545, + 354 + ], + "type": "text", + "content": "[39] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1887-1893. IEEE, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 356, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 356, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 356, + 545, + 421 + ], + "type": "text", + "content": "[40] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 International Conference on Robotics and Automation (ICRA), pages 4376-4382. IEEE, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 488 + ], + "type": "text", + "content": "[41] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15510-15519, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "text", + "content": "[42] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 9621-9630, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 535, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 601 + ], + "type": "text", + "content": "[43] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 602, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 668 + ], + "type": "text", + "content": "[44] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeeze-segv3: Spatially-adaptive convolution for efficient pointcloud segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 1-19. Springer, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[45] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18107" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 575 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision, pages 638-656. Springer, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[46] Xun Xu and Gim Hee Lee. Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13706-13715, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[47] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3101-3109, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "text", + "content": "[48] Cheng-Kun Yang, Ji-Jia Wu, Kai-Syun Chen, Yung-Yu Chuang, and Yen-Yu Lin. An mil-derived transformer for weakly supervised point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11830-11839, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 307 + ], + "type": "text", + "content": "[49] Ping-Chung Yu, Cheng Sun, and Min Sun. Data efficient 3d learner via knowledge transferred from 2d model. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 182-198. Springer, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 361 + ], + "type": "text", + "content": "[50] Yachao Zhang, Zonghao Li, Yuan Xie, Yanyun Qu, Cuihua Li, and Tao Mei. Weakly supervised semantic segmentation for large-scale point cloud. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3421-3429, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 418 + ], + "type": "text", + "content": "[51] Yachao Zhang, Yanyun Qu, Yuan Xie, Zonghao Li, Shanshan Zheng, and Cuihua Li. Perturbed self-distillation: Weakly supervised large-scale point cloud semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15520-15528, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 485 + ], + "type": "text", + "content": "[52] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9610, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "text", + "content": "[53] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "text", + "content": "[54] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18108" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_content_list.json b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d781118cf7b8313bb902deaf6ceb57fcee7a120b --- /dev/null +++ b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Hierarchical Prior Mining for Non-local Multi-View Stereo", + "text_level": 1, + "bbox": [ + 187, + 130, + 782, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chunlin Ren1 Qingshan Xu2 Shikun Zhang1 Jiaqi Yang1*", + "bbox": [ + 178, + 179, + 767, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Northwestern Polytechnical University 2 Nanyang Technological University", + "bbox": [ + 179, + 199, + 789, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{renchunlin, zhangshikun}@mail.nwpu.edu.cn; qingshan.xu@ntu.edu.sg;", + "bbox": [ + 187, + 219, + 779, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jqyang@nwpu.edu.cn", + "bbox": [ + 405, + 237, + 565, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a fundamental problem in computer vision, multiview stereo (MVS) aims at recovering the 3D geometry of a target from a set of 2D images. Recent advances in MVS have shown that it is important to perceive non-local structured information for recovering geometry in low-textured areas. In this work, we propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS). The key characteristics are the following techniques that exploit non-local information to assist MVS: 1) A Non-local Extensible Sampling Pattern (NESP), which is able to adaptively change the size of sampled areas without becoming snared in locally optimal solutions. 2) A new approach to leverage non-local reliable points and construct a planar prior model based on K-Nearest Neighbor (KNN), to obtain potential hypotheses for the regions where prior construction is challenging. 3) A Hierarchical Prior Mining (HPM) framework, which is used to mine extensive non-local prior information at different scales to assist 3D model recovery, this strategy can achieve a considerable balance between the reconstruction of details and low-textured areas. Experimental results on the ETH3D and Tanks & Temples have verified the superior performance and strong generalization capability of our method. Our code will be available at https://github.com/CLinuxx/HPM-MVS.", + "bbox": [ + 75, + 318, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 710, + 209, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-View Stereo (MVS) is one of the key problems in the field of 3D computer vision, which contributes greatly to virtual reality [5], 3D object recognition [14], and autonomous driving [23]. The objective of MVS method is to reconstruct the 3D geometry of a target from a series of images along with camera parameters. Over the past few years, the emergence of numerous datasets [15, 26, 45] has tremendously contributed to the continuous advancement of MVS methods [8, 27, 9]. However, influenced by low", + "bbox": [ + 75, + 738, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6dfe36af85d2e859cf7240f782c9b071901f876893459ad415cabc92d7a20d11.jpg", + "image_caption": [ + "Figure 1: $F_{1}$ score $(\\uparrow)$ comparisons with SOTA traditional [25, 38, 36] and learning-based [30, 29, 32] MVS methods on ETH3D [26] and Tanks & Temples [17]." + ], + "image_footnote": [], + "bbox": [ + 501, + 284, + 696, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5e7581498d2fe79dff5922272199a7392435be2a31ed3bedecc612028357034.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 284, + 887, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "textured areas, illumination variation and other factors [1], MVS is still a challenging problem.", + "bbox": [ + 498, + 521, + 890, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To perform an accurate and robust 3D reconstruction, two different types of methods have been investigated, including learning-based MVS [43, 44, 37, 22] and traditional MVS [9, 38, 36]. Learning-based MVS implements deep networks to extract high-level features and make predictions. However, it needs a large amount of training and its generalization ability still needs further improvement [16]. Traditional MVS can also be broadly classified into two categories: plane-sweeping-based MVS [7, 3, 10] and PatchMatch MVS [9, 25, 38]. Although plane-sweeping-based MVS produces good results for sufficiently textured and unoccluded surfaces, it performs poorly for scenes composed of large planar surfaces, while PatchMatch MVS has successfully overcome this restriction.", + "bbox": [ + 496, + 551, + 892, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "PatchMatch MVS usually consists of four steps, including random initialization, hypothesis propagation, multiview matching cost evaluation and refinement [36]. In essence, it is a process of sampling and verification. The hypothesis propagation samples appropriate hypotheses from neighboring pixels to construct solution space while the multi-view matching cost evaluation defines a criterion to verify the reliability of the sampled hypotheses. Therefore, these two steps play an important role in PatchMatch", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3611", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MVS. Following this popular four-step pipeline, Patch-Match MVS produces remarkable outcomes.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Even so, geometry recovery still suffers a lot from low-textured areas because the previous PatchMatch MVS methods place too much emphasis on local information. Therefore, we exploit non-local information to achieve high-quality reconstruction. We propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS) method, which defines a novel hypothesis propagation pattern and a novel way for constructing planar prior models to assist multi-view matching cost evaluation.", + "bbox": [ + 75, + 125, + 467, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In terms of hypothesis propagation, sequential propagation [2, 46, 25] and diffusion-like propagation [9, 38] are all popular strategies. The previous one only considers the nearest pixels for updating. In contrast, the diffusion-like propagation can update simultaneously based on the neighbor hypotheses around the central point. To further leverage structured region information, ACMH [38] proposes an adaptive checkerboard sampling scheme. Although these previous approaches have greatly improved the efficiency and performance of hypothesis propagation, they both process in the local neighborhood; thus misestimates can only be updated when several better hypotheses in the local area are sampled. Based on the above observations, we first propose the basic method with a Non-local Extensible Sampling Pattern (NESP). Intuitively, a non-local operation removes sampling points around the center point. It allows distant correct pixels to have a greater possibility to contribute to the update of the current pixel. Moreover, the extensible architecture maintains the variable sampling size and can efficiently select suitable candidate hypotheses.", + "bbox": [ + 75, + 263, + 467, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In terms of the multi-view matching cost evaluation, since the photometric consistency usually causes ambiguities during the depth optimization of low-textured areas, many previous methods [35, 11, 40] introduce planar prior models to help the evaluation. They assume low-textured areas are usually on smooth homogeneous surfaces, and construct a planar prior model at the current scale to assist depth estimation. To establish a more robust and more comprehensive planar prior, we employ K-Nearest Neighbor (KNN) to search non-local reliable points and obtain potential hypotheses for the marginal regions where prior construction is difficult. Inspired by the coarse-to-fine strategy, we further design an HPM framework. In a coarse-to-fine manner, the prior knowledge at the early stages is built upon non-local credible sparse correspondences at low-resolution scales, which expands the receptive field and leads to an effective depth estimate for smooth homogeneous surfaces. Subsequently, the prior knowledge at the later stages is constructed at the higher-resolution scales by using previously rectified hypotheses, which restores the depth information of details in the image. This way, the depth information can be successfully recovered by our HPM architecture.", + "bbox": [ + 75, + 568, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on different competitive datasets verify the effectiveness of our method (Fig. 1). The results demonstrate that HPM-MVS has excellent performance, and also holds strong generalization capability to more complex scenes. In a nutshell, our contributions are three-fold as follows:", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present an NESP module to adaptively determine the number of potential sampling points while avoiding becoming snared in locally optimal solutions.", + "- To construct a more robust and more comprehensive planar prior model, we propose an approach based on KNN to search non-local neighbor reliable hypotheses and construct a planar prior model for marginal areas where prior construction is challenging.", + "- We design an HPM framework to explore prior knowledge at multiple scales, which efficiently balances out the reconstruction of details and low-textured areas." + ], + "bbox": [ + 517, + 188, + 890, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 383, + 640, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Traditional Multi-View Stereo", + "text_level": 1, + "bbox": [ + 500, + 409, + 766, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PatchMatch Multi-View Stereo. The PatchMatch method was initially proposed by Barnes et al. [4], and its primary goal is to quickly identify the approximate nearest neighbor among patches of two images. The extension of PatchMatch idea to MVS was proposed by Shen [28]. In order to enhance efficiency, several methods [46, 2, 34] employ the sequential propagation scheme. However, this propagation scheme is still time-consuming. To this end, Galliani et al. [9] proposed Gipuma, a massively parallel multi-view extension of PatchMatch MVS which leverages a red-black checkerboard pattern to perform a diffusion-like propagation scheme. This makes better use of the parallelism of GPUs to improve efficiency. Nonetheless, the major drawback is that each pixel's sampling points are preset, which in turn causes regional information to be ignored. Further, Xu and Tao [38] designed a method with Adaptive Checkerboard sampling and Multi-Hypothesis joint view selection (ACMH) to address this shortcoming. In allusion to depth estimation for low-textured areas, they further combined ACMH with multi-scale geometric consistency guidance to present ACMM. In addition, MARMVS [41] selects the optimal scale for each pixel to alleviate the ambiguity caused by regions with raw texture. Although the aforementioned approaches significantly enhance the performance of PatchMatch MVS method, how to recover the depth information of low-textured regions is always a challenging problem.", + "bbox": [ + 496, + 431, + 890, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Planar Prior Assistance. In order to deal with this longstanding ill-posed problem, the planar prior is considered. Romanoni et al. [24] proposed TAPA-MVS, they separated the image into two different scales of superpixels and utilized these as regional structures. The authors fitted a plane", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d2d72280a756345706a9503d76bd3c741e8f8459af11f69f511be28227fe8a04.jpg", + "image_caption": [ + "Figure 2: An overview of HPM-MVS. Starting with the input images, we first apply the basic MVS with NESP (Fig. 3) to obtain the initial hypotheses. Then, we downsample them to the coarsest scale and generate the prior model. Third, we upsample the model to the current scale and leverage it to assist hypothesis prediction. After operating on different scales, we further use geometric consistency (GC) to optimize the results. Finally, the hypotheses will be fused to a point cloud." + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 810, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "within the superpixels by using trustworthy 3D points and the RANSAC method, then leveraged the prior hypotheses to optimize the low-textured regions. ACMP [40] constructs a planar prior depth map based on triangulations, and subsequently mosaics the prior hypotheses into the MVS process via a probabilistic graphical model. Recently, Xu et al. [36] designed a multi-scale geometric consistency guided and planar prior assisted Multi-View Stereo (ACMMP) by combining ACMM [38] and ACMP [40]. It fully exploits depth information at various scales and leverages prior assistance to efficiently target low-textured areas. Previous approaches only consider the prior information from the current image scale, we nonetheless believe that prior hypotheses should not be restricted to just one scale. To this aim, we embrace an HPM framework to discover more meaningful and comprehensive non-local prior knowledge.", + "bbox": [ + 75, + 412, + 472, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Learning-based Multi-View Stereo", + "text_level": 1, + "bbox": [ + 76, + 665, + 380, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep learning has been widely used in 3D vision and has proven to be critical for MVS. MVSNet [43] introduces an end-to-end deep learning network that builds the 3D cost volumes from 2D image features. R-MVSNet [44] regularizes the cost volume along the depth direction with the convolutional GRU to make reconstruction feasible. To further decrease memory and boost performance, CIDER [39] builds a lightweight cost volume using an average group-wise correlation similarity metric. In recent years, the efficiency of learning-based MVS gains more attention. CVP-MVSNet [42], CasMVSNet [12] and UCS-Net [6] build a cascade cost volume by integrating the coarse-to-fine strategy. Furthermore, PatchMatchNet [30] embeds PatchMatch MVS into a deep learning framework to achieve high", + "bbox": [ + 75, + 689, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quality reconstruction with low memory consumption.", + "bbox": [ + 498, + 412, + 859, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although the existing deep learning methods have achieved great success in MVS, they usually require a large quantity of training datasets, which may not be practical in real-world applications. In addition, the generalization ability still remains an issue for learning-based methods [16]. As such, we still focus on traditional methods for MVS.", + "bbox": [ + 496, + 429, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 535, + 635, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first briefly review the classical Patch-Match MVS. Then, we detail our proposed HPM-MVS. An overview of the method is illustrated in Fig. 2 (the pseudocode is shown in the supplementary).", + "bbox": [ + 496, + 561, + 892, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Review of PatchMatch MVS", + "text_level": 1, + "bbox": [ + 500, + 633, + 756, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PatchMatch MVS [9, 25, 38] is a method for rapidly finding correspondences between image patches. To compute depth maps for input images, each image is selected in turn as a reference image and its corresponding source images are determined based on the Structure-from-Motion (SfM) results. In general, the process consists of four steps [36], i.e., random initialization, hypothesis propagation, multiview cost evaluation and refinement. The last three steps are iterated until convergence, the details are as follows:", + "bbox": [ + 496, + 657, + 893, + 792 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Random initialization means randomly generating a plane hypothesis for each pixel in the reference image based on the results of SfM. A key observation behind is that a non-trivial fraction of a large field of random offset assignments is likely to be a good guess.", + "2) Hypothesis propagation samples the hypotheses from neighboring pixels because of the high spatial coherence of" + ], + "bbox": [ + 498, + 794, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af597405f6e5d7fc1736ad130a548b317cab60a7421af727d104ef3001b915ef.jpg", + "image_caption": [ + "Figure 3: Non-local Extensible Sampling Pattern. The circles represent the pixels in the red-black checkerboard. The red areas represent the sampling regions; the lighter the color, the less adaptive extension is performed. The solid yellow circles show the sampled points." + ], + "image_footnote": [], + "bbox": [ + 145, + 90, + 406, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "neighbors in images. Moreover, a proper hypothesis can be propagated to a relatively large region to ensure the smoothness constraint of nearby pixels.", + "bbox": [ + 75, + 349, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3) Multi-view matching cost evaluation aims at robustly integrating matching cost from multiple views to select the best plane hypotheses.", + "4) Refinement generates two additional hypotheses to enrich the diversity of the solution space. The one with the lowest cost will be selected as the refined estimation." + ], + "bbox": [ + 75, + 395, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we concentrate on hypothesis propagation and multi-view matching cost evaluation, which are essential components of PatchMatch MVS methods, to capture non-local information for recovering better geometry in low-textured areas.", + "bbox": [ + 75, + 486, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Non-local Extensible Sampling Pattern", + "text_level": 1, + "bbox": [ + 76, + 573, + 410, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hypothesis propagation is a prerequisite for multi-view matching cost evaluation and refinement. Hence, choosing a reasonable sampling pattern will lead to accurate results. The pixels within a relatively large area can be represented by one of the pixels. So considering local information repeatedly is redundant. Our NESP is inspired by these two key insights to sample more robust hypotheses. Following the PatchMatch MVS pipeline introduced above, we design our basic MVS with NESP.", + "bbox": [ + 75, + 597, + 468, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Random Initialization. First, we randomly generate a hypothesis $\\pmb{\\theta}_{x} = [d_{x},\\pmb{n}_{x}]$ for each pixel $x$ , where $d_{x}$ represents the depth information and $\\pmb{n}_{x}$ is a normal vector. Second, a matching cost is calculated from each source image via homography [13]. Finally, the initial multi-view matching cost is obtained by averaging top- $k$ best costs.", + "bbox": [ + 75, + 734, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hypothesis Propagation. The diffusion-like propagation scheme can tremendously enhance efficiency while maintaining high-quality hypothesis propagation. Following [38], we first partition all pixels in the reference image into a red-black checkerboard pattern which uses hypothe", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ses within eight regions as candidates to update. Second, we divide eight sampling areas into four long strip areas and four polyline areas; each polyline area starts with 8 samples, compared to each long strip area's initial 5 samples. Then, we consider two strategies to construct our NESP (Fig. 3).", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Extensible Strategy. To achieve the purpose of extension, we first choose the best hypothesis in each area according to the multi-view matching costs from random initialization or the previous iteration (the initial costs in random initialization are used for the first propagation, while the other iterations use the costs computed in the previous iteration) and define the candidate hypothesis set as $\\Theta = \\{\\pmb{\\theta}_i | i = 1 \\dots 8\\}$ . Second, we calculate their matching costs with respect to different source images and embed them into a cost matrix,", + "bbox": [ + 517, + 180, + 890, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} = \\left[ \\begin{array}{c c c c} m _ {1, 1} & m _ {1, 2} & \\dots & m _ {1, N - 1} \\\\ m _ {2, 1} & m _ {2, 2} & \\dots & m _ {2, N - 1} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ m _ {8, 1} & m _ {8, 2} & \\dots & m _ {8, N - 1} \\end{array} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 342, + 890, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the number of input images, $m_{i,j}$ represents the matching cost of the sampled point in the $i$ -th region scored by the $j$ -th view. Based on this matrix, a voting scheme is implemented in each line to determine whether this region needs to be extended. Two thresholds are defined here: 1) A good matching cost threshold is,", + "bbox": [ + 529, + 420, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tau \\left(t _ {i t e r}, t _ {e x t}\\right) = \\tau_ {g o o d} \\cdot e ^ {- \\frac {t _ {i t e r} {} ^ {2} \\cdot \\left(N _ {e x t} - t _ {e x t}\\right)}{\\alpha}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 537, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_{iter}$ and $t_{ext}$ represent the $t_{iter}$ -th iteration of hypothesis propagation and the $t_{ext}$ -th regional expansion, respectively; $\\tau_{good}$ is the initial good matching cost threshold, $\\alpha$ is a constant and $N_{ext}$ represents the maximum number of extensions. 2) A bad matching cost threshold is defined as $\\tau_{bad}$ . For a specific region, there should exist at least $n_{good}$ matching costs smaller than $\\tau(t_{iter}, t_{ext})$ , and at most $n_{bad}$ matching costs meeting the condition: $m_{i,j} > \\tau_{bad}$ . When the above conditions are satisfied, the expansion of the current sampling area will be stopped. Otherwise, this area should extend by itself while each extension doubles the number of sampling points in the area. Notably, the set $\\Theta$ and the matrix $M$ are renewed dynamically in each expansion using the rule described above.", + "bbox": [ + 529, + 571, + 890, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Non-local Strategy. There are two key insights behind the non-local strategy: 1) The surrounding pixels of the central point share the same structured region information, and their hypotheses can be contained by Eq. 4 when in the process of refinement. 2) The misestimates in low-textured areas always appear as small", + "bbox": [ + 517, + 809, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "isolated speckles in the depth map; this phenomenon indicates that these regions have fallen into local optimal solutions. Therefore, we use the non-local strategy (Fig. 3) which abandons the sampling points within the radius $R$ of the center pixels. The motivation is to concentrate more on non-local areas instead of focusing on the local surrounding information.", + "bbox": [ + 107, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the end, we combine these two strategies to form NESP, which helps to collect more reasonable candidate hypotheses from non-local neighboring pixels.", + "bbox": [ + 76, + 208, + 468, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-View Matching Cost Evaluation. In order to determine the best hypothesis from the above candidate hypothesis set, we follow the previous studies [46, 25, 38] to define the multi-view matching cost via photo consistency in our basic MVS with NESP,", + "bbox": [ + 76, + 253, + 468, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc _ {p h o t o} \\left(\\boldsymbol {\\theta} _ {i}\\right) = \\frac {\\sum_ {j} w _ {j} \\cdot m _ {i , j}}{\\sum_ {j} w _ {j}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 337, + 468, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{j}$ is the weight of the $j$ -th view, which is computed by the view selection strategy in [38]. The hypothesis with the least matching cost will be chosen as the current best estimate for the center pixel.", + "bbox": [ + 76, + 383, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refinement. We generate two new hypotheses in refinement. One hypothesis is obtained by perturbing the current hypothesis $[d_x^p, n_x^p]$ , while the other is randomly generated $[d_x^r, n_x^r]$ . Subsequently, the two new hypotheses are randomly arranged and combined with the current hypothesis to form an ensemble of seven hypotheses,", + "bbox": [ + 76, + 445, + 470, + 536 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\begin{array}{c} \\left\\{\\left[ d _ {x}, \\boldsymbol {n} _ {x} \\right], \\left[ d _ {x}, \\boldsymbol {n} _ {x} ^ {p} \\right], \\left[ d _ {x}, \\boldsymbol {n} _ {x} ^ {r} \\right], \\left[ d _ {x} ^ {r}, \\boldsymbol {n} _ {x} \\right], \\right. \\\\ \\left. \\left[ d _ {x} ^ {r}, \\boldsymbol {n} _ {x} ^ {r} \\right], \\left[ d _ {x} ^ {p}, \\boldsymbol {n} _ {x} \\right], \\left[ d _ {x} ^ {p}, \\boldsymbol {n} _ {x} ^ {p} \\right] \\right\\}. \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 545, + 468, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final estimation for pixel $x$ will be the hypothesis with the minimum cost.", + "bbox": [ + 76, + 587, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Planar Prior Construction", + "text_level": 1, + "bbox": [ + 76, + 627, + 316, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The local image patches of low-textured areas are highly similar, and the photometric consistency, as a metric function of MVS, always leads to incorrect estimates due to its defects. Inspired by [40], we construct a better planar prior model to assist multi-view matching cost evaluation. After executing the basic MVS, a hypothesis with its final cost for each pixel can be obtained. A lower cost represents a more precise estimate, therefore we collect the credible correspondences into a set $I_{\\text{cred}}$ ,", + "bbox": [ + 76, + 650, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {\\text {c r e d}} = \\left\\{\\boldsymbol {\\theta} _ {x} \\mid c _ {\\text {p h o t o}} \\left(\\boldsymbol {\\theta} _ {x}\\right) < \\tau_ {\\text {c r e d}} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 797, + 468, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{\\text{red}}$ represents the credible threshold. Then, the Delaunay triangulation is applied to generate triangular surfaces of various sizes. Although these triangles cover a large area, there are still some marginal areas in the image that remain unaffected. This suggests that the local information", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/97591c5205d532d941d8e59b1a360e296e2ac032cd52f8f7d3e34474504bdc10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 630, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5bf72a045e5878075f616bdf268b6b454a70c23998a9b35af8f7ce102a22b8bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 88, + 759, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d08ee48d1d5c634b66e5167c9e9a435548f96f511d504e37ac86107f3af95751.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 88, + 890, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/47f40daefacd977002d3e1c9ece2ae137d60bdfe91cccf2fea9f19404ccab315.jpg", + "image_caption": [ + "(a)", + "(d)", + "Figure 4: Planar prior models at different scales and resulting depth maps. (a) The color image (the low-textured area is in the white box while pixels in the red box are details); (b) the planar prior model constructed at the low-resolution scale; (c) the planar prior model constructed at the high-resolution scale (the dark blue areas show that there are anomalies and have already been filtered out); (d) the depth map obtained without prior models; (e) the depth map based on the prior model constructed at the low-resolution scale; (f) the depth map based on prior models at different scales." + ], + "image_footnote": [], + "bbox": [ + 501, + 175, + 630, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/240c2f46207ca9ec3919e69a5778c1727aeba93dbde47bd7150349b1a64dee78.jpg", + "image_caption": [ + "(b)", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 632, + 175, + 761, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/23800d8bb04a57c8a3fbc6fe7648634d0242af18cc6640efd0e5c60f3825bee6.jpg", + "image_caption": [ + "(c)", + "(f)" + ], + "image_footnote": [], + "bbox": [ + 761, + 175, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cannot satisfy the requirements of developing a planar prior model in these parts of the regions. To address this issue, we propose a method based on KNN to construct the planar prior model via non-local credible hypotheses. First, we set up a KD-tree [47] for the reliable points in $I_{\\text{cred}}$ . Second, the top- $K$ nearest non-local neighbors are searched for pixels in the regions where prior construction is challenging. In addition, we use Heron's formula to avoid the case of co-linearity of proximity points. Then, each pixel will have three reliable neighboring points. We project them into the coordinate of the reference camera and set up a matrix $A$ to get the plane parameters $z_{opt}$ ,", + "bbox": [ + 498, + 453, + 892, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {\\text {o p t}} = \\underset {\\boldsymbol {z} ^ {*}: | | \\boldsymbol {z} ^ {*} | | = 1} {\\arg \\min } \\left| \\left| \\boldsymbol {A} \\cdot \\boldsymbol {z} ^ {*} \\right| \\right|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 647, + 890, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z^{*}$ is a unit-length variable. To further enhance the validity of the prior model, the ultimate prior knowledge for each image is filtered by its depth range. At last, we consider Planar Prior Assistance [40] which can better leverage the planar prior model to assist multi-view matching cost evaluation in low-textured areas.", + "bbox": [ + 498, + 685, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Hierarchical Prior Mining", + "text_level": 1, + "bbox": [ + 500, + 786, + 738, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The multi-scale structure is employed in some Patch-Match MVS methods to infer the depth information of low-textured areas. Previous approaches [21, 38, 33] usually construct image pyramids to fuse depth estimation at different scales. In addition, several methods [40, 24] focus on constructing planar priors models at a single scale to help", + "bbox": [ + 498, + 810, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3615", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the depth estimation. In our work, we demonstrate that constructing planar prior models in a multi-scale way will further facilitate depth estimation.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Fig. 4(b), when the planar prior model is built in low-textured regions, the performance at the low-resolution scale is obviously more advanced than the high-resolution scale. Because the information is condensed after downsampling, resulting in a larger receptive field for each pixel, and a wider range of non-local information can be excavated than the original scale. However, only using the prior from the coarsest scale may lead ambiguity in detail areas. We find that the aforementioned problem of prior construction is expertly managed at the high-resolution scale (Fig. 4(c)).", + "bbox": [ + 75, + 136, + 468, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the depth perception advantages of planar prior models at different scales are complementary. Inspired by this key observation, we design an HPM framework to construct prior models at different scales, which assists the depth estimation of low-textured regions without sacrificing details. The depth map can be continually optimized via the HPM framework (Fig. 4(d)-(f)).", + "bbox": [ + 75, + 303, + 468, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Excavating hierarchical prior is essentially an iterative coarse-to-fine optimization process (Fig. 2). After employing the basic MVS method with NESP, we first downsample the obtained hypotheses (i.e. depth and normal) to the coarsest scale and build a planar prior model. Second, with the joint bilateral upsampler [18], we propagate the planar prior to the original scale. Third, the planar prior model is embedded into the PatchMatch MVS method with Planar Prior Assistance [36] to approximate the depth in low-textured areas better. Finally, we downsample the newly generated hypotheses to the medium scale again, and repeat the same operation until iterating up to the original scale. For images with high resolutions, more medium scales can be considered. To further optimize the results, we apply geometric consistency as [25, 38] do. This way, the final depth estimation can achieve a better trade-off between details and low-textured areas.", + "bbox": [ + 75, + 409, + 468, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, to save computing resources and improve time efficiency, we propose a fast version of the HPM framework. This version only uses the planar prior model constructed at the coarsest scale to assist MVS. The rationale behind this is that the prior model obtained at the coarsest scale can effectively target the low-textured region and exactly compensate for the deficiencies of the basic MVS method in that region. Both HPM and its fast version will be experimentally analyzed.", + "bbox": [ + 75, + 665, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 815, + 209, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on two challenging MVS datasets, ETH3D benchmark [26] and Tanks & Temples datasets [17], from two perspectives, i.e., benchmark performance and analysis experiments.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ad3fb7f3152340efeab16fac4d190b8b81261c3dbe7a503146c0adbe5308aebe.jpg", + "table_caption": [ + "Table 1: Point cloud evaluation on ETH3D benchmark [26]. We show accuracy / completeness / $F_{1}$ score (in %) at different thresholds (2cm and 10cm). The best results are marked in bold while the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
Method2cm10cm
Train.i) Traditional
COLMAP [25]91.85 / 55.13 / 67.6698.75 / 79.47 / 87.61
PCF-MVS [19]84.11 / 75.73 / 79.4295.98 / 90.42 / 92.98
ACMM [38]90.67 / 70.42 / 78.8698.12 / 86.40 / 91.70
ACMP [40]90.12 / 72.15 / 79.7997.97 / 87.15 / 92.03
ACMMP [36]91.03 / 77.27 / 83.4297.96 / 93.19 / 95.46
ii) Learning
PatchMatchNet [30]64.81 / 65.43 / 64.2189.98 / 83.28 / 85.70
IterMVS [29]73.62 / 61.87 / 66.3694.48 / 78.85 / 85.25
MVSTER [32]68.08 / 76.92 / 72.0691.97 / 91.91 / 91.73
HPM-MVSfast91.17 / 73.20 / 80.8698.23 / 92.41 / 94.97
HPM-MVS90.66 / 79.50 / 84.5897.97 / 95.59 / 96.22
Testi) Traditional
COLMAP [25]91.97 / 62.98 / 73.0198.25 / 84.54 / 90.40
PCF-MVS [19]82.15 / 79.29 / 80.3892.12 / 91.26 / 91.56
ACMM [38]90.65 / 74.34 / 80.7898.05 / 88.77 / 92.96
ACMP [40]90.45 / 75.58 / 81.5197.47 / 88.71 / 92.62
ACMMP [36]91.91 / 82.10 / 85.8998.05 / 94.67 / 96.27
ii) Learning
PatchMatchNet [30]79.71 / 77.46 / 73.1291.98 / 92.05 / 91.91
IterMVS [29]76.91 / 72.65 / 74.2995.42 / 85.81 / 90.15
MVSTER [32]77.09 / 82.47 / 79.0194.21 / 92.71 / 92.30
HPM-MVSfast92.50 / 80.25 / 85.3598.32 / 94.89 / 96.51
HPM-MVS92.13 / 83.25 / 87.1198.11 / 95.41 / 96.69
", + "bbox": [ + 504, + 160, + 890, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 468, + 689, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ETH3D high-resolution multi-view benchmark [26] consists of training datasets and test datasets, which contain indoor and outdoor images at a resolution of $6048 \\times 4032^1$ . Tanks & Temples datasets are divided into Intermediate and Advanced datasets based on the complexity of the scene. However, these datasets do not provide official camera poses. In that case, we use COLMAP [25] to estimate camera parameters.", + "bbox": [ + 496, + 491, + 890, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All of our experiments are executed on a computer with an Intel i5-12600 CPU and an RTX 3070 GPU. In the basic MVS with NESP, $\\{k, N_{ext}, \\tau_{good}, \\tau_{bad}, \\alpha, n_{good}, n_{bad}, R\\} = \\{4, 3, 0.8, 1.2, 90, 1, 2, 4\\}$ . In the process of planar prior construction, $\\{\\tau_{cred}, K\\} = \\{0.1, 6\\}$ .", + "bbox": [ + 498, + 613, + 890, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmark Performance", + "text_level": 1, + "bbox": [ + 500, + 700, + 728, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on ETH3D Benchmark. Both traditional methods [25, 19, 38, 40, 36] and learning-based [30, 29, 32] methods are considered for comparison. Table 1 reports the accuracy, completeness and $F_{1}$ score of the point clouds.", + "bbox": [ + 496, + 723, + 890, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in the table, HPM-MVS can achieve outstanding performance in terms of completeness and $F_{1}$ score among all the methods, and HPM-MVS fast can outperform in accuracy while guaranteeing completeness and $F_{1}$ score. Notably, HPM-MVS obtains the highest completeness than other methods on both training and test datasets which con", + "bbox": [ + 496, + 785, + 890, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We downsample the undistorted images to $3200 \\times 2130$ as in [25].", + "bbox": [ + 517, + 886, + 880, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3616", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05e14e7e6da81555aca19be8d88dc6879a00bf56e246349bf0a70e056eb423f2.jpg", + "image_caption": [ + "(a) Images" + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 217, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/827904327740f1d3cb869735f7653944c122222ce6f93bcc841dce356c195e06.jpg", + "image_caption": [ + "(b) COLMAP [25]" + ], + "image_footnote": [], + "bbox": [ + 218, + 88, + 351, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f284a1f851b7c1e1eaf3162640edb1693b018d84557c13888a81bb67324a3bc.jpg", + "image_caption": [ + "(c)ACMMP [36]", + "Figure 5: Qualitative point cloud comparisons between different MVS methods on several high-resolution multi-view test scans (old computer and terrace 2) of ETH3D benchmark [26]. Some challenging areas are shown in red boxes." + ], + "image_footnote": [], + "bbox": [ + 352, + 88, + 483, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/536571a31e4ed0743ce38e864d00358c58a2e09cc7986685400e3a7adf340ac4.jpg", + "image_caption": [ + "(d) IterMVS [29]" + ], + "image_footnote": [], + "bbox": [ + 483, + 88, + 617, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1141416b38400e85332de8c61528263f9dfb5e682023c35a0cf27e1cdd9870b7.jpg", + "image_caption": [ + "(e) HPM-MVSfast" + ], + "image_footnote": [], + "bbox": [ + 617, + 88, + 751, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6b69daf01831a12ea420011182e78fc6801a427d9e035ee0b49d4ef58370f79.jpg", + "image_caption": [ + "(f) HPM-MVS" + ], + "image_footnote": [], + "bbox": [ + 753, + 88, + 885, + 205 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8e14d7c419b577249cf7627fd835531f5267c2e1c5454de8702628dbfb3e33e.jpg", + "table_caption": [ + "Table 2: $F_{1}$ score (in %) comparisons of different methods on Tanks & Temples datasets [17] at given evaluation threshold." + ], + "table_footnote": [], + "table_body": "
MethodIntermediateAdvanced
TraditionalCOLMAP [25]42.1427.24
PCF-MVS [19]55.8834.59
ACMM [38]57.2734.02
ACMP [40]58.4137.44
ACMMP [36]59.3837.84
LearningR-MVSNet [44]48.4024.91
PatchMatch-RL [20]51.8131.78
PatchMatchNet [30]53.1532.31
IterMVS [29]56.2233.24
PVSNet [37]56.8833.46
Effi-MVS [31]56.8834.39
MVSTER [32]60.9237.53
OursHPM-MVSfast61.5939.65
HPM-MVS61.3940.80
", + "bbox": [ + 89, + 348, + 455, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tain large low-textured areas, because HPM-MVS perceives non-local structured information for recovering geometry. The visual comparisons are shown in Fig. 5. It can be clearly observed that our methods produce higher quality point clouds than other competitors, especially in challenging low-textured areas, e.g., red boxes in Fig. 5.", + "bbox": [ + 75, + 568, + 468, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Tanks & Temples Datasets. To further illustrate the robustness of our methods, we test our methods on Tanks & Temples datasets without any fine-tuning. The quantitative results of state-of-the-art methods on both Intermediate and Advanced sets are reported in Table 2.", + "bbox": [ + 75, + 657, + 468, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our methods achieve outstanding performance among all the methods. Even compared with learning-based methods, our methods still achieve better performance without any data training. Remarkably, HPM-MVS can greatly improve the performance of Advanced datasets which contain complex details and more low-textured areas difficult to recover. Our method ranks $\\mathbf{1}^{st}$ in Advanced datasets over all the existing traditional MVS methods, and it outperforms the previous best record by $2.96\\%$ . This result well confirms that our method can achieve a good balance for MVS reconstruction in details and low-textured areas. Fig. 6", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "presents some visualization examples of point cloud recall comparisons between different methods, which verifies our proposed methods can produce more complete 3D geometries than the others. These obvious advantages suggest that HPM-MVS and HPM-MVSfast hold powerful generalization ability and can adapt to different application scenarios without any data training.", + "bbox": [ + 496, + 292, + 892, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis Experiments", + "text_level": 1, + "bbox": [ + 500, + 411, + 702, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. In this section, we conduct an ablation study on the multi-view high-resolution training datasets of ETH3D benchmark to validate the performance of different parts in our proposed methods. We divide them into five sections, which are Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework. The baseline method is $\\mathrm{ACMH}^2$ [38]. Table 3 summarizes the effectiveness of each individual part in our proposed methods.", + "bbox": [ + 496, + 434, + 892, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the basic MVS method with NESP, we remove the no-local part and the extensible part, respectively. The results show that these two proposals can both enhance the completeness and $F_{1}$ score. In particular, the improvement of the non-local strategy is more obvious than the extensible strategy. In some circumstances, especially in low-textured areas, several severe errors always cause the hypothesis propagation into local optimal solutions. Hence, the non-local operation can effectively prevent this situation from getting worse.", + "bbox": [ + 496, + 587, + 890, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on NESP, the planar prior model constructed at the current scale generates a potential hypothesis for each pixel in the image. As for two versions of the HPM framework, one can make several observations from the results: 1) HPM and $\\mathrm{HPM}_{\\mathrm{fast}}$ achieve outstanding performance at relatively high thresholds (e.g., $5cm$ and $10cm$ ), because they have a strong focus on sizable low-textured areas. 2) We notice that $\\mathrm{HPM}_{\\mathrm{fast}}$ will suffer some loss of completeness at low thresholds (e.g., $1cm$ ), because it obtains the prior model", + "bbox": [ + 496, + 738, + 890, + 875 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2Note that, we combine ACMH with geometric consistency here.", + "bbox": [ + 514, + 886, + 862, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b44c67490e7b44509f7e927195a22f77ad6cc1581a28f3692eee820db27d3e86.jpg", + "image_caption": [ + "(a) ACMM [38]" + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 209, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc2707246729b29fe26ccd694e56534f185981625b44a2ac759ed1f643ecfc8a.jpg", + "image_caption": [ + "(b)ACMMP [36]" + ], + "image_footnote": [], + "bbox": [ + 210, + 95, + 339, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4c729c1eb5b77f2923a7857434e9b78f985dcba6f4b9d673eb1bb830da6053d.jpg", + "image_caption": [ + "(c) IterMVS [29]", + "Figure 6: Qualitative point cloud recall comparisons between MVS methods on Tanks & Temples datasets [17] (Horse and Temple). The darker the pixel color is, the greater the error will be. $\\tau$ is the recommended threshold for each dataset." + ], + "image_footnote": [], + "bbox": [ + 341, + 94, + 473, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c447d4aac95d29e091c5b3b45699a21fb33bbc9c56471fefffc27d20bc3beb49.jpg", + "image_caption": [ + "(d) MVSTER [32]" + ], + "image_footnote": [], + "bbox": [ + 473, + 94, + 604, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a90f1108c6e7b1cb5ba66e589188dfb2834ac9d1c1d3e3f54af2dfe429a08e3a.jpg", + "image_caption": [ + "(e) HPM-MVSfast" + ], + "image_footnote": [], + "bbox": [ + 604, + 94, + 736, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77bfd1b20116410f1aa48c24e8dc813953d4115d6ca12a78d1d86b8b835d368a.jpg", + "image_caption": [ + "(f) HPM-MVS" + ], + "image_footnote": [], + "bbox": [ + 738, + 90, + 893, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d2e12fe0ca338398df04b596307a02f3c1aab9d25c405849cf5fe5a8c5728bd.jpg", + "table_caption": [ + "Table 3: Comparative results of our methods under different settings on the training datasets of ETH3D benchmark [26]. N, E, PA, $\\mathrm{HPM}_{\\mathrm{fast}}$ and HPM mean Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSettingsETH3D train.
NEPAHPMfastHPM1cm2cm5cm10cm
ACMH [38]82.41 / 50.63 / 61.7191.09 / 63.04 / 73.6096.61 / 74.73 / 83.5098.25 / 82.20 / 88.57
NSP✓82.52 / 53.46 / 63.8490.86 / 65.59 / 75.2696.63 / 76.15 / 84.5298.46 / 82.40 / 89.19
ESP✓82.66 / 51.40 / 62.2290.93 / 63.83 / 74.0596.64 / 75.37 / 84.0498.47 / 82.09 / 89.03
NESP✓✓82.47 / 53.98 / 64.2290.78 / 66.02 / 75.5996.60 / 76.63 / 84.8798.47 / 82.73 / 89.44
NESP+PA✓✓✓82.88 / 62.37 / 71.3290.51 / 74.96 / 81.6696.48 / 84.29 / 88.9097.89 / 88.57 / 92.73
HPMfast✓✓83.32 / 53.53 / 64.6691.16 / 70.48 / 79.0496.39 / 84.67 / 89.8497.89 / 88.57 / 92.73
HPM✓✓83.01 / 62.07 / 70.6790.65 / 77.03 / 83.0696.06 / 88.42 / 91.9797.91 / 93.62 / 95.66
HPM-MVSfast✓✓✓✓82.95 / 57.36 / 67.3991.17 / 73.20 / 80.8696.51 / 86.22 / 90.8998.23 / 92.14 / 94.97
HPM-MVS✓✓✓✓82.93 / 65.16 / 72.7390.66 / 79.50 / 84.5896.13 / 89.98 / 92.8897.97 / 94.59 / 96.22
", + "bbox": [ + 81, + 339, + 890, + 483 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/356f8e53a9e737db617e0740426760f135ba7190d1d36360cbfaf271df5a2528.jpg", + "table_caption": [ + "Table 4: Generalization performance of NESP on ETH3D high-resolution training datasets [26]." + ], + "table_footnote": [], + "table_body": "
MethodETH3D train.
1cm2cm5cm10cm20cm
ACMM [38]67.5878.8687.6891.7094.41
ACMP [40]68.7279.7988.3292.0394.43
ACMMP [36]71.5783.4292.0395.4697.37
ACMM+NESP70.7081.0188.8092.2694.55
3.12↑2.15↑1.12↑0.56↑0.14↑
ACMP+NESP70.8781.4589.4392.7294.78
2.15↑1.66↑1.11↑0.69↑0.35↑
ACMMP+NESP74.5485.3393.2596.4597.99
2.97↑1.91↑1.22↑0.99↑0.62↑
", + "bbox": [ + 101, + 547, + 444, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "at the lowest resolution scale which may lead to ambiguities in details. 3) The method with NESP as the basis for subsequent processes can provide a set of reliable points for planar prior model construction.", + "bbox": [ + 75, + 718, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization Performance of NESP. In addition, we perform a more extensive comparison on the multi-view high-resolution training datasets of ETH3D benchmark to show the generalization ability of NESP. Several kinds of state-of-the-art diffusion-like propagation methods are integrated with NESP for evaluation. The considered methods are ACMM [38], ACMP [40] and ACMMP [36]. Note that ACMMC is a very recent state-of-the-art geometric-only", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method. Each method is tested under different thresholds, and results are shown in Table 4.", + "bbox": [ + 498, + 507, + 888, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impressively, NESP dramatically improves the $F_{1}$ score for all tested methods at relatively low thresholds, e.g., 1cm and 2cm. The improvement of NESP working with multi-scale strategy is more obvious, because the non-local and extensible strategy works even better at low-resolution scale. Notably, the combination of NESP and ACMMP outperforms all the competitors and achieves state-of-the-art $F_{1}$ score of $85.33\\% / 96.45\\%$ at 2cm/10cm thresholds. Based on the results, the following conclusions can be drawn: 1) NESP is a generalized module that can be adapted to other MVS methods and boost their performance. 2) Combined with the multi-scale strategy, NESP can achieve better performance.", + "bbox": [ + 496, + 539, + 892, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More analysis experiments and point cloud visualizations can be found in the supplementary.", + "bbox": [ + 498, + 736, + 890, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 782, + 617, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented an HPM-MVS method which can efficiently perceive non-local structured information to achieve high-quality reconstruction. Based on the non-local and extensible strategy, we first propose our NESP module. Focusing on the planar prior construction in marginal regions, we employ KNN to search non-local credible points", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and obtain potential hypotheses. Further, we design an HPM framework which explores planar prior at different scales to assist MVS. Through extensive experiments, we have demonstrated the rationality of each key component of our method and its superiority against existing methods. In the future, we plan to combine our method with a probabilistic graphical model to further improve its ability to handle the reconstruction of details and low-textured regions.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (No. 62002295), and NWPU international cooperation and exchange promotion projects (No. 02100-23GH0501).", + "bbox": [ + 76, + 218, + 472, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 292, + 174, + 309 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 1", + "[2] Christian Bailer, Manuel Finckh, and Hendrik Lensch. Scale robust multi view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 398-411. Springer, 2012. 2", + "[3] Caroline Baillard and Andrew Zisserman. A plane-sweep strategy for the 3d reconstruction of buildings from multiple images. International Archives of Photogrammetry and Remote Sensing, 33(B2; PART 2):56-62, 2000. 1", + "[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Transactions on Graphics, 28(3):24, 2009. 2", + "[5] Fabio Bruno, Stefano Bruno, Giovanna De Sensi, Maria-Laura Luchi, Stefania Mancuso, and Maurizio Muzzupappa. From 3d reconstruction to virtual reality: A complete methodology for digital archaeological exhibition. Journal of Cultural Heritage, 11(1):42-49, 2010. 1", + "[6] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2020. 3", + "[7] Robert T Collins. A space-sweep approach to true multi-image matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 358-363. IEEE, 1996. 1", + "[8] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2009. 1", + "[9] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 1, 2, 3", + "[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, Qingxiong Yang, and Marc Pollefeys. Real-time plane-sweeping stereo with multiple sweeping directions. In Pro-" + ], + "bbox": [ + 78, + 316, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1", + "[11] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1418-1425. IEEE, 2010. 2", + "[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 3", + "[13] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 4", + "[14] Mohsen Hejrati and Deva Ramanan. Analysis by synthesis: 3d object recognition by object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2456, 2014. 1", + "[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 1", + "[16] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016. 1, 3", + "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 1, 6, 7, 8", + "[18] Johannes Kopf, Michael F Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Transactions on Graphics, 26(3):96-es, 2007. 6", + "[19] Andreas Kuhn, Shan Lin, and Oliver Erdler. Plane completion and filtering for multi-view stereo reconstruction. In Proceedings of the German Conference on Pattern Recognition, pages 18-32. Springer, 2019. 6, 7", + "[20] Jae Yong Lee, Joseph DeGol, Chuhang Zou, and Derek Hoiem. Patchmatch-rl: Deep mvs with pixelwise depth, normal, and visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6158–6167, 2021. 7", + "[21] Jie Liao, Yanping Fu, Qingan Yan, and Chunxia Xiao. Pyramid multi-view stereo with local consistency. In Computer Graphics Forum, volume 38, pages 335-346. Wiley Online Library, 2019. 5", + "[22] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1", + "[23] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE International Conference on Computer Vision, pages 6851–6860, 2019. 1" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3619", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Andrea Romanoni and Matteo Matteucci. Tapa-mvs: Textureless-aware patchmatch multi-view stereo. In Proceedings of the IEEE International Conference on Computer Vision, pages 10413-10422, 2019. 2, 5", + "[25] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 3, 5, 6, 7", + "[26] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 1, 6, 7, 8", + "[27] Qi Shan, Riley Adams, Brian Curless, Yasutaka Furukawa, and Steven M Seitz. The visual turing test for scene reconstruction. In International Conference on 3D Vision-3DV, pages 25–32. IEEE, 2013. 1", + "[28] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. IEEE Transactions on Image Processing, 22(5):1901-1914, 2013. 2", + "[29] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8606-8615, 2022. 1, 6, 7, 8", + "[30] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14194-14203, 2021. 1, 3, 6, 7", + "[31] Shaoqian Wang, Bo Li, and Yuchao Dai. Efficient multiview stereo by iterative dynamic cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8655-8664, 2022. 7", + "[32] Xiaofeng Wang, Zheng Zhu, Guan Huang, Fangbo Qin, Yun Ye, Yijia He, Xu Chi, and Xingang Wang. Mvster: Epipolar transformer for efficient multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 573-591. Springer, 2022. 1, 6, 7, 8", + "[33] Yuesong Wang, Tao Guan, Zhuo Chen, Yawei Luo, Keyang Luo, and Lili Ju. Mesh-guided multi-view stereo with pyramid architecture. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2039-2048, 2020. 5", + "[34] Jian Wei, Benjamin Resch, and Hendrik PA Lensch. Multiview depth map estimation with cross-view consistency. In Proceedings of the British Machine Vision Conference, 2014. 2", + "[35] Oliver Woodford, Philip Torr, Ian Reid, and Andrew Fitzgibbon. Global stereo reconstruction under second-order smoothness priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(12):2115-2128, 2009. 2", + "[36] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 1, 7", + "[38] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 1, 2, 3, 4, 5, 6, 7, 8", + "[39] Qingshan Xu and Wenbing Tao. Learning inverse depth regression for multi-view stereo with correlation cost volume. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12508-12515, 2020. 3", + "[40] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 2, 3, 5, 6, 7, 8", + "[41] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5981-5990, 2020. 2", + "[42] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3", + "[43] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 767-783, 2018. 1, 3", + "[44] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvvsnet for high-resolution multiview stereo depth inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 1, 3, 7", + "[45] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A largescale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 1", + "[46] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2, 5", + "[47] Kun Zhou, Qiming Hou, Rui Wang, and Baining Guo. Real-time kd-tree construction on graphics hardware. ACM Transactions on Graphics, 27(5):1-11, 2008. 5" + ], + "bbox": [ + 501, + 92, + 890, + 768 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3620", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_model.json b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0fba554d8f855202604a7bd9133d05491b2394d7 --- /dev/null +++ b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.131, + 0.784, + 0.152 + ], + "angle": 0, + "content": "Hierarchical Prior Mining for Non-local Multi-View Stereo" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.18, + 0.768, + 0.2 + ], + "angle": 0, + "content": "Chunlin Ren1 Qingshan Xu2 Shikun Zhang1 Jiaqi Yang1*" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.2, + 0.79, + 0.217 + ], + "angle": 0, + "content": "1 Northwestern Polytechnical University 2 Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.22, + 0.78, + 0.234 + ], + "angle": 0, + "content": "{renchunlin, zhangshikun}@mail.nwpu.edu.cn; qingshan.xu@ntu.edu.sg;" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.238, + 0.566, + 0.251 + ], + "angle": 0, + "content": "jqyang@nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.683 + ], + "angle": 0, + "content": "As a fundamental problem in computer vision, multiview stereo (MVS) aims at recovering the 3D geometry of a target from a set of 2D images. Recent advances in MVS have shown that it is important to perceive non-local structured information for recovering geometry in low-textured areas. In this work, we propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS). The key characteristics are the following techniques that exploit non-local information to assist MVS: 1) A Non-local Extensible Sampling Pattern (NESP), which is able to adaptively change the size of sampled areas without becoming snared in locally optimal solutions. 2) A new approach to leverage non-local reliable points and construct a planar prior model based on K-Nearest Neighbor (KNN), to obtain potential hypotheses for the regions where prior construction is challenging. 3) A Hierarchical Prior Mining (HPM) framework, which is used to mine extensive non-local prior information at different scales to assist 3D model recovery, this strategy can achieve a considerable balance between the reconstruction of details and low-textured areas. Experimental results on the ETH3D and Tanks & Temples have verified the superior performance and strong generalization capability of our method. Our code will be available at https://github.com/CLinuxx/HPM-MVS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.21, + 0.728 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Multi-View Stereo (MVS) is one of the key problems in the field of 3D computer vision, which contributes greatly to virtual reality [5], 3D object recognition [14], and autonomous driving [23]. The objective of MVS method is to reconstruct the 3D geometry of a target from a series of images along with camera parameters. Over the past few years, the emergence of numerous datasets [15, 26, 45] has tremendously contributed to the continuous advancement of MVS methods [8, 27, 9]. However, influenced by low" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.285, + 0.697, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.285, + 0.888, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.448, + 0.893, + 0.495 + ], + "angle": 0, + "content": "Figure 1: \\(F_{1}\\) score \\((\\uparrow)\\) comparisons with SOTA traditional [25, 38, 36] and learning-based [30, 29, 32] MVS methods on ETH3D [26] and Tanks & Temples [17]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.892, + 0.553 + ], + "angle": 0, + "content": "textured areas, illumination variation and other factors [1], MVS is still a challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.764 + ], + "angle": 0, + "content": "To perform an accurate and robust 3D reconstruction, two different types of methods have been investigated, including learning-based MVS [43, 44, 37, 22] and traditional MVS [9, 38, 36]. Learning-based MVS implements deep networks to extract high-level features and make predictions. However, it needs a large amount of training and its generalization ability still needs further improvement [16]. Traditional MVS can also be broadly classified into two categories: plane-sweeping-based MVS [7, 3, 10] and PatchMatch MVS [9, 25, 38]. Although plane-sweeping-based MVS produces good results for sufficiently textured and unoccluded surfaces, it performs poorly for scenes composed of large planar surfaces, while PatchMatch MVS has successfully overcome this restriction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "PatchMatch MVS usually consists of four steps, including random initialization, hypothesis propagation, multiview matching cost evaluation and refinement [36]. In essence, it is a process of sampling and verification. The hypothesis propagation samples appropriate hypotheses from neighboring pixels to construct solution space while the multi-view matching cost evaluation defines a criterion to verify the reliability of the sampled hypotheses. Therefore, these two steps play an important role in PatchMatch" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3611" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "MVS. Following this popular four-step pipeline, Patch-Match MVS produces remarkable outcomes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.125, + 0.468, + 0.261 + ], + "angle": 0, + "content": "Even so, geometry recovery still suffers a lot from low-textured areas because the previous PatchMatch MVS methods place too much emphasis on local information. Therefore, we exploit non-local information to achieve high-quality reconstruction. We propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS) method, which defines a novel hypothesis propagation pattern and a novel way for constructing planar prior models to assist multi-view matching cost evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.468, + 0.566 + ], + "angle": 0, + "content": "In terms of hypothesis propagation, sequential propagation [2, 46, 25] and diffusion-like propagation [9, 38] are all popular strategies. The previous one only considers the nearest pixels for updating. In contrast, the diffusion-like propagation can update simultaneously based on the neighbor hypotheses around the central point. To further leverage structured region information, ACMH [38] proposes an adaptive checkerboard sampling scheme. Although these previous approaches have greatly improved the efficiency and performance of hypothesis propagation, they both process in the local neighborhood; thus misestimates can only be updated when several better hypotheses in the local area are sampled. Based on the above observations, we first propose the basic method with a Non-local Extensible Sampling Pattern (NESP). Intuitively, a non-local operation removes sampling points around the center point. It allows distant correct pixels to have a greater possibility to contribute to the update of the current pixel. Moreover, the extensible architecture maintains the variable sampling size and can efficiently select suitable candidate hypotheses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In terms of the multi-view matching cost evaluation, since the photometric consistency usually causes ambiguities during the depth optimization of low-textured areas, many previous methods [35, 11, 40] introduce planar prior models to help the evaluation. They assume low-textured areas are usually on smooth homogeneous surfaces, and construct a planar prior model at the current scale to assist depth estimation. To establish a more robust and more comprehensive planar prior, we employ K-Nearest Neighbor (KNN) to search non-local reliable points and obtain potential hypotheses for the marginal regions where prior construction is difficult. Inspired by the coarse-to-fine strategy, we further design an HPM framework. In a coarse-to-fine manner, the prior knowledge at the early stages is built upon non-local credible sparse correspondences at low-resolution scales, which expands the receptive field and leads to an effective depth estimate for smooth homogeneous surfaces. Subsequently, the prior knowledge at the later stages is constructed at the higher-resolution scales by using previously rectified hypotheses, which restores the depth information of details in the image. This way, the depth information can be successfully recovered by our HPM architecture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "Extensive experiments on different competitive datasets verify the effectiveness of our method (Fig. 1). The results demonstrate that HPM-MVS has excellent performance, and also holds strong generalization capability to more complex scenes. In a nutshell, our contributions are three-fold as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.189, + 0.891, + 0.235 + ], + "angle": 0, + "content": "- We present an NESP module to adaptively determine the number of potential sampling points while avoiding becoming snared in locally optimal solutions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.244, + 0.892, + 0.319 + ], + "angle": 0, + "content": "- To construct a more robust and more comprehensive planar prior model, we propose an approach based on KNN to search non-local neighbor reliable hypotheses and construct a planar prior model for marginal areas where prior construction is challenging." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.328, + 0.892, + 0.372 + ], + "angle": 0, + "content": "- We design an HPM framework to explore prior knowledge at multiple scales, which efficiently balances out the reconstruction of details and low-textured areas." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.189, + 0.892, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.385, + 0.642, + 0.401 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.767, + 0.424 + ], + "angle": 0, + "content": "2.1. Traditional Multi-View Stereo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.825 + ], + "angle": 0, + "content": "PatchMatch Multi-View Stereo. The PatchMatch method was initially proposed by Barnes et al. [4], and its primary goal is to quickly identify the approximate nearest neighbor among patches of two images. The extension of PatchMatch idea to MVS was proposed by Shen [28]. In order to enhance efficiency, several methods [46, 2, 34] employ the sequential propagation scheme. However, this propagation scheme is still time-consuming. To this end, Galliani et al. [9] proposed Gipuma, a massively parallel multi-view extension of PatchMatch MVS which leverages a red-black checkerboard pattern to perform a diffusion-like propagation scheme. This makes better use of the parallelism of GPUs to improve efficiency. Nonetheless, the major drawback is that each pixel's sampling points are preset, which in turn causes regional information to be ignored. Further, Xu and Tao [38] designed a method with Adaptive Checkerboard sampling and Multi-Hypothesis joint view selection (ACMH) to address this shortcoming. In allusion to depth estimation for low-textured areas, they further combined ACMH with multi-scale geometric consistency guidance to present ACMM. In addition, MARMVS [41] selects the optimal scale for each pixel to alleviate the ambiguity caused by regions with raw texture. Although the aforementioned approaches significantly enhance the performance of PatchMatch MVS method, how to recover the depth information of low-textured regions is always a challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Planar Prior Assistance. In order to deal with this longstanding ill-posed problem, the planar prior is considered. Romanoni et al. [24] proposed TAPA-MVS, they separated the image into two different scales of superpixels and utilized these as regional structures. The authors fitted a plane" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3612" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.089, + 0.812, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.896, + 0.39 + ], + "angle": 0, + "content": "Figure 2: An overview of HPM-MVS. Starting with the input images, we first apply the basic MVS with NESP (Fig. 3) to obtain the initial hypotheses. Then, we downsample them to the coarsest scale and generate the prior model. Third, we upsample the model to the current scale and leverage it to assist hypothesis prediction. After operating on different scales, we further use geometric consistency (GC) to optimize the results. Finally, the hypotheses will be fused to a point cloud." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.473, + 0.657 + ], + "angle": 0, + "content": "within the superpixels by using trustworthy 3D points and the RANSAC method, then leveraged the prior hypotheses to optimize the low-textured regions. ACMP [40] constructs a planar prior depth map based on triangulations, and subsequently mosaics the prior hypotheses into the MVS process via a probabilistic graphical model. Recently, Xu et al. [36] designed a multi-scale geometric consistency guided and planar prior assisted Multi-View Stereo (ACMMP) by combining ACMM [38] and ACMP [40]. It fully exploits depth information at various scales and leverages prior assistance to efficiently target low-textured areas. Previous approaches only consider the prior information from the current image scale, we nonetheless believe that prior hypotheses should not be restricted to just one scale. To this aim, we embrace an HPM framework to discover more meaningful and comprehensive non-local prior knowledge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.381, + 0.683 + ], + "angle": 0, + "content": "2.2. Learning-based Multi-View Stereo" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Deep learning has been widely used in 3D vision and has proven to be critical for MVS. MVSNet [43] introduces an end-to-end deep learning network that builds the 3D cost volumes from 2D image features. R-MVSNet [44] regularizes the cost volume along the depth direction with the convolutional GRU to make reconstruction feasible. To further decrease memory and boost performance, CIDER [39] builds a lightweight cost volume using an average group-wise correlation similarity metric. In recent years, the efficiency of learning-based MVS gains more attention. CVP-MVSNet [42], CasMVSNet [12] and UCS-Net [6] build a cascade cost volume by integrating the coarse-to-fine strategy. Furthermore, PatchMatchNet [30] embeds PatchMatch MVS into a deep learning framework to achieve high" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.86, + 0.43 + ], + "angle": 0, + "content": "quality reconstruction with low memory consumption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Although the existing deep learning methods have achieved great success in MVS, they usually require a large quantity of training datasets, which may not be practical in real-world applications. In addition, the generalization ability still remains an issue for learning-based methods [16]. As such, we still focus on traditional methods for MVS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.637, + 0.554 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.893, + 0.624 + ], + "angle": 0, + "content": "In this section, we first briefly review the classical Patch-Match MVS. Then, we detail our proposed HPM-MVS. An overview of the method is illustrated in Fig. 2 (the pseudocode is shown in the supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.757, + 0.65 + ], + "angle": 0, + "content": "3.1. Review of PatchMatch MVS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.794 + ], + "angle": 0, + "content": "PatchMatch MVS [9, 25, 38] is a method for rapidly finding correspondences between image patches. To compute depth maps for input images, each image is selected in turn as a reference image and its corresponding source images are determined based on the Structure-from-Motion (SfM) results. In general, the process consists of four steps [36], i.e., random initialization, hypothesis propagation, multiview cost evaluation and refinement. The last three steps are iterated until convergence, the details are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.894, + 0.87 + ], + "angle": 0, + "content": "1) Random initialization means randomly generating a plane hypothesis for each pixel in the reference image based on the results of SfM. A key observation behind is that a non-trivial fraction of a large field of random offset assignments is likely to be a good guess." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "2) Hypothesis propagation samples the hypotheses from neighboring pixels because of the high spatial coherence of" + }, + { + "type": "list", + "bbox": [ + 0.499, + 0.795, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3613" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.146, + 0.091, + 0.408, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.473, + 0.322 + ], + "angle": 0, + "content": "Figure 3: Non-local Extensible Sampling Pattern. The circles represent the pixels in the red-black checkerboard. The red areas represent the sampling regions; the lighter the color, the less adaptive extension is performed. The solid yellow circles show the sampled points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.395 + ], + "angle": 0, + "content": "neighbors in images. Moreover, a proper hypothesis can be propagated to a relatively large region to ensure the smoothness constraint of nearby pixels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.47, + 0.44 + ], + "angle": 0, + "content": "3) Multi-view matching cost evaluation aims at robustly integrating matching cost from multiple views to select the best plane hypotheses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.469, + 0.485 + ], + "angle": 0, + "content": "4) Refinement generates two additional hypotheses to enrich the diversity of the solution space. The one with the lowest cost will be selected as the refined estimation." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.396, + 0.47, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.562 + ], + "angle": 0, + "content": "In this work, we concentrate on hypothesis propagation and multi-view matching cost evaluation, which are essential components of PatchMatch MVS methods, to capture non-local information for recovering better geometry in low-textured areas." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.411, + 0.59 + ], + "angle": 0, + "content": "3.2. Non-local Extensible Sampling Pattern" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.733 + ], + "angle": 0, + "content": "Hypothesis propagation is a prerequisite for multi-view matching cost evaluation and refinement. Hence, choosing a reasonable sampling pattern will lead to accurate results. The pixels within a relatively large area can be represented by one of the pixels. So considering local information repeatedly is redundant. Our NESP is inspired by these two key insights to sample more robust hypotheses. Following the PatchMatch MVS pipeline introduced above, we design our basic MVS with NESP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Random Initialization. First, we randomly generate a hypothesis \\(\\pmb{\\theta}_{x} = [d_{x},\\pmb{n}_{x}]\\) for each pixel \\(x\\), where \\(d_{x}\\) represents the depth information and \\(\\pmb{n}_{x}\\) is a normal vector. Second, a matching cost is calculated from each source image via homography [13]. Finally, the initial multi-view matching cost is obtained by averaging top-\\(k\\) best costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Hypothesis Propagation. The diffusion-like propagation scheme can tremendously enhance efficiency while maintaining high-quality hypothesis propagation. Following [38], we first partition all pixels in the reference image into a red-black checkerboard pattern which uses hypothe" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "ses within eight regions as candidates to update. Second, we divide eight sampling areas into four long strip areas and four polyline areas; each polyline area starts with 8 samples, compared to each long strip area's initial 5 samples. Then, we consider two strategies to construct our NESP (Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.181, + 0.892, + 0.332 + ], + "angle": 0, + "content": "- Extensible Strategy. To achieve the purpose of extension, we first choose the best hypothesis in each area according to the multi-view matching costs from random initialization or the previous iteration (the initial costs in random initialization are used for the first propagation, while the other iterations use the costs computed in the previous iteration) and define the candidate hypothesis set as \\(\\Theta = \\{\\pmb{\\theta}_i | i = 1 \\dots 8\\}\\). Second, we calculate their matching costs with respect to different source images and embed them into a cost matrix," + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.343, + 0.892, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} = \\left[ \\begin{array}{c c c c} m _ {1, 1} & m _ {1, 2} & \\dots & m _ {1, N - 1} \\\\ m _ {2, 1} & m _ {2, 2} & \\dots & m _ {2, N - 1} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ m _ {8, 1} & m _ {8, 2} & \\dots & m _ {8, N - 1} \\end{array} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.421, + 0.892, + 0.526 + ], + "angle": 0, + "content": "where \\(N\\) is the number of input images, \\(m_{i,j}\\) represents the matching cost of the sampled point in the \\(i\\)-th region scored by the \\(j\\)-th view. Based on this matrix, a voting scheme is implemented in each line to determine whether this region needs to be extended. Two thresholds are defined here: 1) A good matching cost threshold is," + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.538, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\tau \\left(t _ {i t e r}, t _ {e x t}\\right) = \\tau_ {g o o d} \\cdot e ^ {- \\frac {t _ {i t e r} {} ^ {2} \\cdot \\left(N _ {e x t} - t _ {e x t}\\right)}{\\alpha}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.573, + 0.892, + 0.799 + ], + "angle": 0, + "content": "where \\( t_{iter} \\) and \\( t_{ext} \\) represent the \\( t_{iter} \\)-th iteration of hypothesis propagation and the \\( t_{ext} \\)-th regional expansion, respectively; \\( \\tau_{good} \\) is the initial good matching cost threshold, \\( \\alpha \\) is a constant and \\( N_{ext} \\) represents the maximum number of extensions. 2) A bad matching cost threshold is defined as \\( \\tau_{bad} \\). For a specific region, there should exist at least \\( n_{good} \\) matching costs smaller than \\( \\tau(t_{iter}, t_{ext}) \\), and at most \\( n_{bad} \\) matching costs meeting the condition: \\( m_{i,j} > \\tau_{bad} \\). When the above conditions are satisfied, the expansion of the current sampling area will be stopped. Otherwise, this area should extend by itself while each extension doubles the number of sampling points in the area. Notably, the set \\( \\Theta \\) and the matrix \\( M \\) are renewed dynamically in each expansion using the rule described above." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Non-local Strategy. There are two key insights behind the non-local strategy: 1) The surrounding pixels of the central point share the same structured region information, and their hypotheses can be contained by Eq. 4 when in the process of refinement. 2) The misestimates in low-textured areas always appear as small" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3614" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "isolated speckles in the depth map; this phenomenon indicates that these regions have fallen into local optimal solutions. Therefore, we use the non-local strategy (Fig. 3) which abandons the sampling points within the radius \\( R \\) of the center pixels. The motivation is to concentrate more on non-local areas instead of focusing on the local surrounding information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.209, + 0.47, + 0.255 + ], + "angle": 0, + "content": "In the end, we combine these two strategies to form NESP, which helps to collect more reasonable candidate hypotheses from non-local neighboring pixels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.255, + 0.469, + 0.329 + ], + "angle": 0, + "content": "Multi-View Matching Cost Evaluation. In order to determine the best hypothesis from the above candidate hypothesis set, we follow the previous studies [46, 25, 38] to define the multi-view matching cost via photo consistency in our basic MVS with NESP," + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.338, + 0.47, + 0.377 + ], + "angle": 0, + "content": "\\[\nc _ {p h o t o} \\left(\\boldsymbol {\\theta} _ {i}\\right) = \\frac {\\sum_ {j} w _ {j} \\cdot m _ {i , j}}{\\sum_ {j} w _ {j}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.47, + 0.446 + ], + "angle": 0, + "content": "where \\( w_{j} \\) is the weight of the \\( j \\)-th view, which is computed by the view selection strategy in [38]. The hypothesis with the least matching cost will be chosen as the current best estimate for the center pixel." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.446, + 0.471, + 0.537 + ], + "angle": 0, + "content": "Refinement. We generate two new hypotheses in refinement. One hypothesis is obtained by perturbing the current hypothesis \\([d_x^p, n_x^p]\\), while the other is randomly generated \\([d_x^r, n_x^r]\\). Subsequently, the two new hypotheses are randomly arranged and combined with the current hypothesis to form an ensemble of seven hypotheses," + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.546, + 0.469, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\left. \\begin{array}{c} \\left\\{\\left[ d _ {x}, \\boldsymbol {n} _ {x} \\right], \\left[ d _ {x}, \\boldsymbol {n} _ {x} ^ {p} \\right], \\left[ d _ {x}, \\boldsymbol {n} _ {x} ^ {r} \\right], \\left[ d _ {x} ^ {r}, \\boldsymbol {n} _ {x} \\right], \\right. \\\\ \\left. \\left[ d _ {x} ^ {r}, \\boldsymbol {n} _ {x} ^ {r} \\right], \\left[ d _ {x} ^ {p}, \\boldsymbol {n} _ {x} \\right], \\left[ d _ {x} ^ {p}, \\boldsymbol {n} _ {x} ^ {p} \\right] \\right\\}. \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.588, + 0.47, + 0.618 + ], + "angle": 0, + "content": "The final estimation for pixel \\( x \\) will be the hypothesis with the minimum cost." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.628, + 0.317, + 0.643 + ], + "angle": 0, + "content": "3.3. Planar Prior Construction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.651, + 0.47, + 0.788 + ], + "angle": 0, + "content": "The local image patches of low-textured areas are highly similar, and the photometric consistency, as a metric function of MVS, always leads to incorrect estimates due to its defects. Inspired by [40], we construct a better planar prior model to assist multi-view matching cost evaluation. After executing the basic MVS, a hypothesis with its final cost for each pixel can be obtained. A lower cost represents a more precise estimate, therefore we collect the credible correspondences into a set \\( I_{\\text{cred}} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.798, + 0.469, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {\\text {c r e d}} = \\left\\{\\boldsymbol {\\theta} _ {x} \\mid c _ {\\text {p h o t o}} \\left(\\boldsymbol {\\theta} _ {x}\\right) < \\tau_ {\\text {c r e d}} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\tau_{\\text{red}}\\) represents the credible threshold. Then, the Delaunay triangulation is applied to generate triangular surfaces of various sizes. Although these triangles cover a large area, there are still some marginal areas in the image that remain unaffected. This suggests that the local information" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.632, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.163, + 0.576, + 0.174 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.089, + 0.761, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.163, + 0.706, + 0.174 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.089, + 0.891, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.163, + 0.836, + 0.174 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.176, + 0.632, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.25, + 0.577, + 0.262 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.176, + 0.762, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.251, + 0.706, + 0.262 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.176, + 0.891, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.25, + 0.835, + 0.262 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.275, + 0.893, + 0.427 + ], + "angle": 0, + "content": "Figure 4: Planar prior models at different scales and resulting depth maps. (a) The color image (the low-textured area is in the white box while pixels in the red box are details); (b) the planar prior model constructed at the low-resolution scale; (c) the planar prior model constructed at the high-resolution scale (the dark blue areas show that there are anomalies and have already been filtered out); (d) the depth map obtained without prior models; (e) the depth map based on the prior model constructed at the low-resolution scale; (f) the depth map based on prior models at different scales." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.454, + 0.893, + 0.636 + ], + "angle": 0, + "content": "cannot satisfy the requirements of developing a planar prior model in these parts of the regions. To address this issue, we propose a method based on KNN to construct the planar prior model via non-local credible hypotheses. First, we set up a KD-tree [47] for the reliable points in \\( I_{\\text{cred}} \\). Second, the top-\\( K \\) nearest non-local neighbors are searched for pixels in the regions where prior construction is challenging. In addition, we use Heron's formula to avoid the case of co-linearity of proximity points. Then, each pixel will have three reliable neighboring points. We project them into the coordinate of the reference camera and set up a matrix \\( A \\) to get the plane parameters \\( z_{opt} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.648, + 0.891, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {\\text {o p t}} = \\underset {\\boldsymbol {z} ^ {*}: | | \\boldsymbol {z} ^ {*} | | = 1} {\\arg \\min } \\left| \\left| \\boldsymbol {A} \\cdot \\boldsymbol {z} ^ {*} \\right| \\right|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.686, + 0.892, + 0.777 + ], + "angle": 0, + "content": "where \\( z^{*} \\) is a unit-length variable. To further enhance the validity of the prior model, the ultimate prior knowledge for each image is filtered by its depth range. At last, we consider Planar Prior Assistance [40] which can better leverage the planar prior model to assist multi-view matching cost evaluation in low-textured areas." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.74, + 0.803 + ], + "angle": 0, + "content": "3.4. Hierarchical Prior Mining" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The multi-scale structure is employed in some Patch-Match MVS methods to infer the depth information of low-textured areas. Previous approaches [21, 38, 33] usually construct image pyramids to fuse depth estimation at different scales. In addition, several methods [40, 24] focus on constructing planar priors models at a single scale to help" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3615" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "the depth estimation. In our work, we demonstrate that constructing planar prior models in a multi-scale way will further facilitate depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.303 + ], + "angle": 0, + "content": "As depicted in Fig. 4(b), when the planar prior model is built in low-textured regions, the performance at the low-resolution scale is obviously more advanced than the high-resolution scale. Because the information is condensed after downsampling, resulting in a larger receptive field for each pixel, and a wider range of non-local information can be excavated than the original scale. However, only using the prior from the coarsest scale may lead ambiguity in detail areas. We find that the aforementioned problem of prior construction is expertly managed at the high-resolution scale (Fig. 4(c))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.409 + ], + "angle": 0, + "content": "Therefore, the depth perception advantages of planar prior models at different scales are complementary. Inspired by this key observation, we design an HPM framework to construct prior models at different scales, which assists the depth estimation of low-textured regions without sacrificing details. The depth map can be continually optimized via the HPM framework (Fig. 4(d)-(f))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Excavating hierarchical prior is essentially an iterative coarse-to-fine optimization process (Fig. 2). After employing the basic MVS method with NESP, we first downsample the obtained hypotheses (i.e. depth and normal) to the coarsest scale and build a planar prior model. Second, with the joint bilateral upsampler [18], we propagate the planar prior to the original scale. Third, the planar prior model is embedded into the PatchMatch MVS method with Planar Prior Assistance [36] to approximate the depth in low-textured areas better. Finally, we downsample the newly generated hypotheses to the medium scale again, and repeat the same operation until iterating up to the original scale. For images with high resolutions, more medium scales can be considered. To further optimize the results, we apply geometric consistency as [25, 38] do. This way, the final depth estimation can achieve a better trade-off between details and low-textured areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Moreover, to save computing resources and improve time efficiency, we propose a fast version of the HPM framework. This version only uses the planar prior model constructed at the coarsest scale to assist MVS. The rationale behind this is that the prior model obtained at the coarsest scale can effectively target the low-textured region and exactly compensate for the deficiencies of the basic MVS method in that region. Both HPM and its fast version will be experimentally analyzed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.21, + 0.833 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We evaluate our method on two challenging MVS datasets, ETH3D benchmark [26] and Tanks & Temples datasets [17], from two perspectives, i.e., benchmark performance and analysis experiments." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.15 + ], + "angle": 0, + "content": "Table 1: Point cloud evaluation on ETH3D benchmark [26]. We show accuracy / completeness / \\(F_{1}\\) score (in %) at different thresholds (2cm and 10cm). The best results are marked in bold while the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.161, + 0.892, + 0.445 + ], + "angle": 0, + "content": "
Method2cm10cm
Train.i) Traditional
COLMAP [25]91.85 / 55.13 / 67.6698.75 / 79.47 / 87.61
PCF-MVS [19]84.11 / 75.73 / 79.4295.98 / 90.42 / 92.98
ACMM [38]90.67 / 70.42 / 78.8698.12 / 86.40 / 91.70
ACMP [40]90.12 / 72.15 / 79.7997.97 / 87.15 / 92.03
ACMMP [36]91.03 / 77.27 / 83.4297.96 / 93.19 / 95.46
ii) Learning
PatchMatchNet [30]64.81 / 65.43 / 64.2189.98 / 83.28 / 85.70
IterMVS [29]73.62 / 61.87 / 66.3694.48 / 78.85 / 85.25
MVSTER [32]68.08 / 76.92 / 72.0691.97 / 91.91 / 91.73
HPM-MVSfast91.17 / 73.20 / 80.8698.23 / 92.41 / 94.97
HPM-MVS90.66 / 79.50 / 84.5897.97 / 95.59 / 96.22
Testi) Traditional
COLMAP [25]91.97 / 62.98 / 73.0198.25 / 84.54 / 90.40
PCF-MVS [19]82.15 / 79.29 / 80.3892.12 / 91.26 / 91.56
ACMM [38]90.65 / 74.34 / 80.7898.05 / 88.77 / 92.96
ACMP [40]90.45 / 75.58 / 81.5197.47 / 88.71 / 92.62
ACMMP [36]91.91 / 82.10 / 85.8998.05 / 94.67 / 96.27
ii) Learning
PatchMatchNet [30]79.71 / 77.46 / 73.1291.98 / 92.05 / 91.91
IterMVS [29]76.91 / 72.65 / 74.2995.42 / 85.81 / 90.15
MVSTER [32]77.09 / 82.47 / 79.0194.21 / 92.71 / 92.30
HPM-MVSfast92.50 / 80.25 / 85.3598.32 / 94.89 / 96.51
HPM-MVS92.13 / 83.25 / 87.1198.11 / 95.41 / 96.69
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.691, + 0.485 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.613 + ], + "angle": 0, + "content": "ETH3D high-resolution multi-view benchmark [26] consists of training datasets and test datasets, which contain indoor and outdoor images at a resolution of \\(6048 \\times 4032^1\\). Tanks & Temples datasets are divided into Intermediate and Advanced datasets based on the complexity of the scene. However, these datasets do not provide official camera poses. In that case, we use COLMAP [25] to estimate camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.69 + ], + "angle": 0, + "content": "All of our experiments are executed on a computer with an Intel i5-12600 CPU and an RTX 3070 GPU. In the basic MVS with NESP, \\(\\{k, N_{ext}, \\tau_{good}, \\tau_{bad}, \\alpha, n_{good}, n_{bad}, R\\} = \\{4, 3, 0.8, 1.2, 90, 1, 2, 4\\}\\). In the process of planar prior construction, \\(\\{\\tau_{cred}, K\\} = \\{0.1, 6\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.701, + 0.73, + 0.715 + ], + "angle": 0, + "content": "4.2. Benchmark Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Evaluation on ETH3D Benchmark. Both traditional methods [25, 19, 38, 40, 36] and learning-based [30, 29, 32] methods are considered for comparison. Table 1 reports the accuracy, completeness and \\(F_{1}\\) score of the point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.892, + 0.876 + ], + "angle": 0, + "content": "As shown in the table, HPM-MVS can achieve outstanding performance in terms of completeness and \\( F_{1} \\) score among all the methods, and HPM-MVS fast can outperform in accuracy while guaranteeing completeness and \\( F_{1} \\) score. Notably, HPM-MVS obtains the highest completeness than other methods on both training and test datasets which con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.881, + 0.901 + ], + "angle": 0, + "content": "1We downsample the undistorted images to \\(3200 \\times 2130\\) as in [25]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.088, + 0.218, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.211, + 0.185, + 0.225 + ], + "angle": 0, + "content": "(a) Images" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.089, + 0.352, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.211, + 0.342, + 0.225 + ], + "angle": 0, + "content": "(b) COLMAP [25]" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.089, + 0.485, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.211, + 0.471, + 0.225 + ], + "angle": 0, + "content": "(c)ACMMP [36]" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.089, + 0.618, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.211, + 0.603, + 0.225 + ], + "angle": 0, + "content": "(d) IterMVS [29]" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.089, + 0.753, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.211, + 0.741, + 0.225 + ], + "angle": 0, + "content": "(e) HPM-MVSfast" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.089, + 0.887, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.211, + 0.864, + 0.224 + ], + "angle": 0, + "content": "(f) HPM-MVS" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.894, + 0.268 + ], + "angle": 0, + "content": "Figure 5: Qualitative point cloud comparisons between different MVS methods on several high-resolution multi-view test scans (old computer and terrace 2) of ETH3D benchmark [26]. Some challenging areas are shown in red boxes." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.337 + ], + "angle": 0, + "content": "Table 2: \\(F_{1}\\) score (in %) comparisons of different methods on Tanks & Temples datasets [17] at given evaluation threshold." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.349, + 0.456, + 0.543 + ], + "angle": 0, + "content": "
MethodIntermediateAdvanced
TraditionalCOLMAP [25]42.1427.24
PCF-MVS [19]55.8834.59
ACMM [38]57.2734.02
ACMP [40]58.4137.44
ACMMP [36]59.3837.84
LearningR-MVSNet [44]48.4024.91
PatchMatch-RL [20]51.8131.78
PatchMatchNet [30]53.1532.31
IterMVS [29]56.2233.24
PVSNet [37]56.8833.46
Effi-MVS [31]56.8834.39
MVSTER [32]60.9237.53
OursHPM-MVSfast61.5939.65
HPM-MVS61.3940.80
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.659 + ], + "angle": 0, + "content": "tain large low-textured areas, because HPM-MVS perceives non-local structured information for recovering geometry. The visual comparisons are shown in Fig. 5. It can be clearly observed that our methods produce higher quality point clouds than other competitors, especially in challenging low-textured areas, e.g., red boxes in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.735 + ], + "angle": 0, + "content": "Evaluation on Tanks & Temples Datasets. To further illustrate the robustness of our methods, we test our methods on Tanks & Temples datasets without any fine-tuning. The quantitative results of state-of-the-art methods on both Intermediate and Advanced sets are reported in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our methods achieve outstanding performance among all the methods. Even compared with learning-based methods, our methods still achieve better performance without any data training. Remarkably, HPM-MVS can greatly improve the performance of Advanced datasets which contain complex details and more low-textured areas difficult to recover. Our method ranks \\( \\mathbf{1}^{st} \\) in Advanced datasets over all the existing traditional MVS methods, and it outperforms the previous best record by \\( 2.96\\% \\). This result well confirms that our method can achieve a good balance for MVS reconstruction in details and low-textured areas. Fig. 6" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.893, + 0.402 + ], + "angle": 0, + "content": "presents some visualization examples of point cloud recall comparisons between different methods, which verifies our proposed methods can produce more complete 3D geometries than the others. These obvious advantages suggest that HPM-MVS and HPM-MVSfast hold powerful generalization ability and can adapt to different application scenarios without any data training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.412, + 0.704, + 0.429 + ], + "angle": 0, + "content": "4.3. Analysis Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Ablation Study. In this section, we conduct an ablation study on the multi-view high-resolution training datasets of ETH3D benchmark to validate the performance of different parts in our proposed methods. We divide them into five sections, which are Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework. The baseline method is \\(\\mathrm{ACMH}^2\\) [38]. Table 3 summarizes the effectiveness of each individual part in our proposed methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.739 + ], + "angle": 0, + "content": "For the basic MVS method with NESP, we remove the no-local part and the extensible part, respectively. The results show that these two proposals can both enhance the completeness and \\( F_{1} \\) score. In particular, the improvement of the non-local strategy is more obvious than the extensible strategy. In some circumstances, especially in low-textured areas, several severe errors always cause the hypothesis propagation into local optimal solutions. Hence, the non-local operation can effectively prevent this situation from getting worse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Based on NESP, the planar prior model constructed at the current scale generates a potential hypothesis for each pixel in the image. As for two versions of the HPM framework, one can make several observations from the results: 1) HPM and \\(\\mathrm{HPM}_{\\mathrm{fast}}\\) achieve outstanding performance at relatively high thresholds (e.g., \\(5cm\\) and \\(10cm\\)), because they have a strong focus on sizable low-textured areas. 2) We notice that \\(\\mathrm{HPM}_{\\mathrm{fast}}\\) will suffer some loss of completeness at low thresholds (e.g., \\(1cm\\)), because it obtains the prior model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.863, + 0.9 + ], + "angle": 0, + "content": "2Note that, we combine ACMH with geometric consistency here." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.21, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.211, + 0.195, + 0.226 + ], + "angle": 0, + "content": "(a) ACMM [38]" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.096, + 0.34, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.211, + 0.331, + 0.226 + ], + "angle": 0, + "content": "(b)ACMMP [36]" + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.095, + 0.475, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.212, + 0.461, + 0.226 + ], + "angle": 0, + "content": "(c) IterMVS [29]" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.095, + 0.605, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.212, + 0.597, + 0.226 + ], + "angle": 0, + "content": "(d) MVSTER [32]" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.095, + 0.738, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.212, + 0.728, + 0.226 + ], + "angle": 0, + "content": "(e) HPM-MVSfast" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.092, + 0.894, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.212, + 0.85, + 0.225 + ], + "angle": 0, + "content": "(f) HPM-MVS" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.894, + 0.269 + ], + "angle": 0, + "content": "Figure 6: Qualitative point cloud recall comparisons between MVS methods on Tanks & Temples datasets [17] (Horse and Temple). The darker the pixel color is, the greater the error will be. \\(\\tau\\) is the recommended threshold for each dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.283, + 0.893, + 0.33 + ], + "angle": 0, + "content": "Table 3: Comparative results of our methods under different settings on the training datasets of ETH3D benchmark [26]. N, E, PA, \\(\\mathrm{HPM}_{\\mathrm{fast}}\\) and HPM mean Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.34, + 0.891, + 0.484 + ], + "angle": 0, + "content": "
MethodSettingsETH3D train.
NEPAHPMfastHPM1cm2cm5cm10cm
ACMH [38]82.41 / 50.63 / 61.7191.09 / 63.04 / 73.6096.61 / 74.73 / 83.5098.25 / 82.20 / 88.57
NSP✓82.52 / 53.46 / 63.8490.86 / 65.59 / 75.2696.63 / 76.15 / 84.5298.46 / 82.40 / 89.19
ESP✓82.66 / 51.40 / 62.2290.93 / 63.83 / 74.0596.64 / 75.37 / 84.0498.47 / 82.09 / 89.03
NESP✓✓82.47 / 53.98 / 64.2290.78 / 66.02 / 75.5996.60 / 76.63 / 84.8798.47 / 82.73 / 89.44
NESP+PA✓✓✓82.88 / 62.37 / 71.3290.51 / 74.96 / 81.6696.48 / 84.29 / 88.9097.89 / 88.57 / 92.73
HPMfast✓✓83.32 / 53.53 / 64.6691.16 / 70.48 / 79.0496.39 / 84.67 / 89.8497.89 / 88.57 / 92.73
HPM✓✓83.01 / 62.07 / 70.6790.65 / 77.03 / 83.0696.06 / 88.42 / 91.9797.91 / 93.62 / 95.66
HPM-MVSfast✓✓✓✓82.95 / 57.36 / 67.3991.17 / 73.20 / 80.8696.51 / 86.22 / 90.8998.23 / 92.14 / 94.97
HPM-MVS✓✓✓✓82.93 / 65.16 / 72.7390.66 / 79.50 / 84.5896.13 / 89.98 / 92.8897.97 / 94.59 / 96.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Table 4: Generalization performance of NESP on ETH3D high-resolution training datasets [26]." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.548, + 0.446, + 0.692 + ], + "angle": 0, + "content": "
MethodETH3D train.
1cm2cm5cm10cm20cm
ACMM [38]67.5878.8687.6891.7094.41
ACMP [40]68.7279.7988.3292.0394.43
ACMMP [36]71.5783.4292.0395.4697.37
ACMM+NESP70.7081.0188.8092.2694.55
3.12↑2.15↑1.12↑0.56↑0.14↑
ACMP+NESP70.8781.4589.4392.7294.78
2.15↑1.66↑1.11↑0.69↑0.35↑
ACMMP+NESP74.5485.3393.2596.4597.99
2.97↑1.91↑1.22↑0.99↑0.62↑
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.779 + ], + "angle": 0, + "content": "at the lowest resolution scale which may lead to ambiguities in details. 3) The method with NESP as the basis for subsequent processes can provide a set of reliable points for planar prior model construction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Generalization Performance of NESP. In addition, we perform a more extensive comparison on the multi-view high-resolution training datasets of ETH3D benchmark to show the generalization ability of NESP. Several kinds of state-of-the-art diffusion-like propagation methods are integrated with NESP for evaluation. The considered methods are ACMM [38], ACMP [40] and ACMMP [36]. Note that ACMMC is a very recent state-of-the-art geometric-only" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.89, + 0.538 + ], + "angle": 0, + "content": "method. Each method is tested under different thresholds, and results are shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.893, + 0.736 + ], + "angle": 0, + "content": "Impressively, NESP dramatically improves the \\( F_{1} \\) score for all tested methods at relatively low thresholds, e.g., 1cm and 2cm. The improvement of NESP working with multi-scale strategy is more obvious, because the non-local and extensible strategy works even better at low-resolution scale. Notably, the combination of NESP and ACMMP outperforms all the competitors and achieves state-of-the-art \\( F_{1} \\) score of \\( 85.33\\% / 96.45\\% \\) at 2cm/10cm thresholds. Based on the results, the following conclusions can be drawn: 1) NESP is a generalized module that can be adapted to other MVS methods and boost their performance. 2) Combined with the multi-scale strategy, NESP can achieve better performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.737, + 0.892, + 0.768 + ], + "angle": 0, + "content": "More analysis experiments and point cloud visualizations can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.619, + 0.8 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this paper, we presented an HPM-MVS method which can efficiently perceive non-local structured information to achieve high-quality reconstruction. Based on the non-local and extensible strategy, we first propose our NESP module. Focusing on the planar prior construction in marginal regions, we employ KNN to search non-local credible points" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3618" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "and obtain potential hypotheses. Further, we design an HPM framework which explores planar prior at different scales to assist MVS. Through extensive experiments, we have demonstrated the rationality of each key component of our method and its superiority against existing methods. In the future, we plan to combine our method with a probabilistic graphical model to further improve its ability to handle the reconstruction of details and low-textured regions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.219, + 0.473, + 0.281 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (No. 62002295), and NWPU international cooperation and exchange promotion projects (No. 02100-23GH0501)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.294, + 0.175, + 0.31 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.376, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[2] Christian Bailer, Manuel Finckh, and Hendrik Lensch. Scale robust multi view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 398-411. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.434, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[3] Caroline Baillard and Andrew Zisserman. A plane-sweep strategy for the 3d reconstruction of buildings from multiple images. International Archives of Photogrammetry and Remote Sensing, 33(B2; PART 2):56-62, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Transactions on Graphics, 28(3):24, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[5] Fabio Bruno, Stefano Bruno, Giovanna De Sensi, Maria-Laura Luchi, Stefania Mancuso, and Maurizio Muzzupappa. From 3d reconstruction to virtual reality: A complete methodology for digital archaeological exhibition. Journal of Cultural Heritage, 11(1):42-49, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[6] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[7] Robert T Collins. A space-sweep approach to true multi-image matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 358-363. IEEE, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.746, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[8] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[9] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, Qingxiong Yang, and Marc Pollefeys. Real-time plane-sweeping stereo with multiple sweeping directions. In Pro-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.318, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[11] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1418-1425. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[13] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[14] Mohsen Hejrati and Deva Ramanan. Analysis by synthesis: 3d object recognition by object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2456, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[16] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[18] Johannes Kopf, Michael F Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Transactions on Graphics, 26(3):96-es, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[19] Andreas Kuhn, Shan Lin, and Oliver Erdler. Plane completion and filtering for multi-view stereo reconstruction. In Proceedings of the German Conference on Pattern Recognition, pages 18-32. Springer, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[20] Jae Yong Lee, Joseph DeGol, Chuhang Zou, and Derek Hoiem. Patchmatch-rl: Deep mvs with pixelwise depth, normal, and visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6158–6167, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[21] Jie Liao, Yanping Fu, Qingan Yan, and Chunxia Xiao. Pyramid multi-view stereo with local consistency. In Computer Graphics Forum, volume 38, pages 335-346. Wiley Online Library, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[22] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE International Conference on Computer Vision, pages 6851–6860, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3619" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[24] Andrea Romanoni and Matteo Matteucci. Tapa-mvs: Textureless-aware patchmatch multi-view stereo. In Proceedings of the IEEE International Conference on Computer Vision, pages 10413-10422, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[25] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.471, + 0.301 + ], + "angle": 0, + "content": "[26] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[27] Qi Shan, Riley Adams, Brian Curless, Yasutaka Furukawa, and Steven M Seitz. The visual turing test for scene reconstruction. In International Conference on 3D Vision-3DV, pages 25–32. IEEE, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[28] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. IEEE Transactions on Image Processing, 22(5):1901-1914, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[29] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8606-8615, 2022. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[30] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14194-14203, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[31] Shaoqian Wang, Bo Li, and Yuchao Dai. Efficient multiview stereo by iterative dynamic cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8655-8664, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[32] Xiaofeng Wang, Zheng Zhu, Guan Huang, Fangbo Qin, Yun Ye, Yijia He, Xu Chi, and Xingang Wang. Mvster: Epipolar transformer for efficient multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 573-591. Springer, 2022. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[33] Yuesong Wang, Tao Guan, Zhuo Chen, Yawei Luo, Keyang Luo, and Lili Ju. Mesh-guided multi-view stereo with pyramid architecture. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2039-2048, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[34] Jian Wei, Benjamin Resch, and Hendrik PA Lensch. Multiview depth map estimation with cross-view consistency. In Proceedings of the British Machine Vision Conference, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[35] Oliver Woodford, Philip Torr, Ian Reid, and Andrew Fitzgibbon. Global stereo reconstruction under second-order smoothness priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(12):2115-2128, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[36] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[37] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[38] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[39] Qingshan Xu and Wenbing Tao. Learning inverse depth regression for multi-view stereo with correlation cost volume. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12508-12515, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[40] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[41] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5981-5990, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[42] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[43] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 767-783, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[44] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvvsnet for high-resolution multiview stereo depth inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[45] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A largescale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[46] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[47] Kun Zhou, Qiming Hou, Rui Wang, and Baining Guo. Real-time kd-tree construction on graphics hardware. ACM Transactions on Graphics, 27(5):1-11, 2008. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3620" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_origin.pdf b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..739b5340a6ea8565d0b8ad0ef1384b55fcde2c12 --- /dev/null +++ b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/32540140-b677-4721-9ff6-75d372d0916d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f0eeda30b75165889e805dd20cb776a42be46c2278cbc859fa9255038844a40 +size 10393008 diff --git a/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/full.md b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19ef60110e7fc57610ac103e3431981982c3e87e --- /dev/null +++ b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/full.md @@ -0,0 +1,339 @@ +# Hierarchical Prior Mining for Non-local Multi-View Stereo + +Chunlin Ren1 Qingshan Xu2 Shikun Zhang1 Jiaqi Yang1* + +1 Northwestern Polytechnical University 2 Nanyang Technological University + +{renchunlin, zhangshikun}@mail.nwpu.edu.cn; qingshan.xu@ntu.edu.sg; + +jqyang@nwpu.edu.cn + +# Abstract + +As a fundamental problem in computer vision, multiview stereo (MVS) aims at recovering the 3D geometry of a target from a set of 2D images. Recent advances in MVS have shown that it is important to perceive non-local structured information for recovering geometry in low-textured areas. In this work, we propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS). The key characteristics are the following techniques that exploit non-local information to assist MVS: 1) A Non-local Extensible Sampling Pattern (NESP), which is able to adaptively change the size of sampled areas without becoming snared in locally optimal solutions. 2) A new approach to leverage non-local reliable points and construct a planar prior model based on K-Nearest Neighbor (KNN), to obtain potential hypotheses for the regions where prior construction is challenging. 3) A Hierarchical Prior Mining (HPM) framework, which is used to mine extensive non-local prior information at different scales to assist 3D model recovery, this strategy can achieve a considerable balance between the reconstruction of details and low-textured areas. Experimental results on the ETH3D and Tanks & Temples have verified the superior performance and strong generalization capability of our method. Our code will be available at https://github.com/CLinuxx/HPM-MVS. + +# 1. Introduction + +Multi-View Stereo (MVS) is one of the key problems in the field of 3D computer vision, which contributes greatly to virtual reality [5], 3D object recognition [14], and autonomous driving [23]. The objective of MVS method is to reconstruct the 3D geometry of a target from a series of images along with camera parameters. Over the past few years, the emergence of numerous datasets [15, 26, 45] has tremendously contributed to the continuous advancement of MVS methods [8, 27, 9]. However, influenced by low + +![](images/6dfe36af85d2e859cf7240f782c9b071901f876893459ad415cabc92d7a20d11.jpg) +Figure 1: $F_{1}$ score $(\uparrow)$ comparisons with SOTA traditional [25, 38, 36] and learning-based [30, 29, 32] MVS methods on ETH3D [26] and Tanks & Temples [17]. + +![](images/a5e7581498d2fe79dff5922272199a7392435be2a31ed3bedecc612028357034.jpg) + +textured areas, illumination variation and other factors [1], MVS is still a challenging problem. + +To perform an accurate and robust 3D reconstruction, two different types of methods have been investigated, including learning-based MVS [43, 44, 37, 22] and traditional MVS [9, 38, 36]. Learning-based MVS implements deep networks to extract high-level features and make predictions. However, it needs a large amount of training and its generalization ability still needs further improvement [16]. Traditional MVS can also be broadly classified into two categories: plane-sweeping-based MVS [7, 3, 10] and PatchMatch MVS [9, 25, 38]. Although plane-sweeping-based MVS produces good results for sufficiently textured and unoccluded surfaces, it performs poorly for scenes composed of large planar surfaces, while PatchMatch MVS has successfully overcome this restriction. + +PatchMatch MVS usually consists of four steps, including random initialization, hypothesis propagation, multiview matching cost evaluation and refinement [36]. In essence, it is a process of sampling and verification. The hypothesis propagation samples appropriate hypotheses from neighboring pixels to construct solution space while the multi-view matching cost evaluation defines a criterion to verify the reliability of the sampled hypotheses. Therefore, these two steps play an important role in PatchMatch + +MVS. Following this popular four-step pipeline, Patch-Match MVS produces remarkable outcomes. + +Even so, geometry recovery still suffers a lot from low-textured areas because the previous PatchMatch MVS methods place too much emphasis on local information. Therefore, we exploit non-local information to achieve high-quality reconstruction. We propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS) method, which defines a novel hypothesis propagation pattern and a novel way for constructing planar prior models to assist multi-view matching cost evaluation. + +In terms of hypothesis propagation, sequential propagation [2, 46, 25] and diffusion-like propagation [9, 38] are all popular strategies. The previous one only considers the nearest pixels for updating. In contrast, the diffusion-like propagation can update simultaneously based on the neighbor hypotheses around the central point. To further leverage structured region information, ACMH [38] proposes an adaptive checkerboard sampling scheme. Although these previous approaches have greatly improved the efficiency and performance of hypothesis propagation, they both process in the local neighborhood; thus misestimates can only be updated when several better hypotheses in the local area are sampled. Based on the above observations, we first propose the basic method with a Non-local Extensible Sampling Pattern (NESP). Intuitively, a non-local operation removes sampling points around the center point. It allows distant correct pixels to have a greater possibility to contribute to the update of the current pixel. Moreover, the extensible architecture maintains the variable sampling size and can efficiently select suitable candidate hypotheses. + +In terms of the multi-view matching cost evaluation, since the photometric consistency usually causes ambiguities during the depth optimization of low-textured areas, many previous methods [35, 11, 40] introduce planar prior models to help the evaluation. They assume low-textured areas are usually on smooth homogeneous surfaces, and construct a planar prior model at the current scale to assist depth estimation. To establish a more robust and more comprehensive planar prior, we employ K-Nearest Neighbor (KNN) to search non-local reliable points and obtain potential hypotheses for the marginal regions where prior construction is difficult. Inspired by the coarse-to-fine strategy, we further design an HPM framework. In a coarse-to-fine manner, the prior knowledge at the early stages is built upon non-local credible sparse correspondences at low-resolution scales, which expands the receptive field and leads to an effective depth estimate for smooth homogeneous surfaces. Subsequently, the prior knowledge at the later stages is constructed at the higher-resolution scales by using previously rectified hypotheses, which restores the depth information of details in the image. This way, the depth information can be successfully recovered by our HPM architecture. + +Extensive experiments on different competitive datasets verify the effectiveness of our method (Fig. 1). The results demonstrate that HPM-MVS has excellent performance, and also holds strong generalization capability to more complex scenes. In a nutshell, our contributions are three-fold as follows: + +- We present an NESP module to adaptively determine the number of potential sampling points while avoiding becoming snared in locally optimal solutions. +- To construct a more robust and more comprehensive planar prior model, we propose an approach based on KNN to search non-local neighbor reliable hypotheses and construct a planar prior model for marginal areas where prior construction is challenging. +- We design an HPM framework to explore prior knowledge at multiple scales, which efficiently balances out the reconstruction of details and low-textured areas. + +# 2. Related Work + +# 2.1. Traditional Multi-View Stereo + +PatchMatch Multi-View Stereo. The PatchMatch method was initially proposed by Barnes et al. [4], and its primary goal is to quickly identify the approximate nearest neighbor among patches of two images. The extension of PatchMatch idea to MVS was proposed by Shen [28]. In order to enhance efficiency, several methods [46, 2, 34] employ the sequential propagation scheme. However, this propagation scheme is still time-consuming. To this end, Galliani et al. [9] proposed Gipuma, a massively parallel multi-view extension of PatchMatch MVS which leverages a red-black checkerboard pattern to perform a diffusion-like propagation scheme. This makes better use of the parallelism of GPUs to improve efficiency. Nonetheless, the major drawback is that each pixel's sampling points are preset, which in turn causes regional information to be ignored. Further, Xu and Tao [38] designed a method with Adaptive Checkerboard sampling and Multi-Hypothesis joint view selection (ACMH) to address this shortcoming. In allusion to depth estimation for low-textured areas, they further combined ACMH with multi-scale geometric consistency guidance to present ACMM. In addition, MARMVS [41] selects the optimal scale for each pixel to alleviate the ambiguity caused by regions with raw texture. Although the aforementioned approaches significantly enhance the performance of PatchMatch MVS method, how to recover the depth information of low-textured regions is always a challenging problem. + +Planar Prior Assistance. In order to deal with this longstanding ill-posed problem, the planar prior is considered. Romanoni et al. [24] proposed TAPA-MVS, they separated the image into two different scales of superpixels and utilized these as regional structures. The authors fitted a plane + +![](images/d2d72280a756345706a9503d76bd3c741e8f8459af11f69f511be28227fe8a04.jpg) +Figure 2: An overview of HPM-MVS. Starting with the input images, we first apply the basic MVS with NESP (Fig. 3) to obtain the initial hypotheses. Then, we downsample them to the coarsest scale and generate the prior model. Third, we upsample the model to the current scale and leverage it to assist hypothesis prediction. After operating on different scales, we further use geometric consistency (GC) to optimize the results. Finally, the hypotheses will be fused to a point cloud. + +within the superpixels by using trustworthy 3D points and the RANSAC method, then leveraged the prior hypotheses to optimize the low-textured regions. ACMP [40] constructs a planar prior depth map based on triangulations, and subsequently mosaics the prior hypotheses into the MVS process via a probabilistic graphical model. Recently, Xu et al. [36] designed a multi-scale geometric consistency guided and planar prior assisted Multi-View Stereo (ACMMP) by combining ACMM [38] and ACMP [40]. It fully exploits depth information at various scales and leverages prior assistance to efficiently target low-textured areas. Previous approaches only consider the prior information from the current image scale, we nonetheless believe that prior hypotheses should not be restricted to just one scale. To this aim, we embrace an HPM framework to discover more meaningful and comprehensive non-local prior knowledge. + +# 2.2. Learning-based Multi-View Stereo + +Deep learning has been widely used in 3D vision and has proven to be critical for MVS. MVSNet [43] introduces an end-to-end deep learning network that builds the 3D cost volumes from 2D image features. R-MVSNet [44] regularizes the cost volume along the depth direction with the convolutional GRU to make reconstruction feasible. To further decrease memory and boost performance, CIDER [39] builds a lightweight cost volume using an average group-wise correlation similarity metric. In recent years, the efficiency of learning-based MVS gains more attention. CVP-MVSNet [42], CasMVSNet [12] and UCS-Net [6] build a cascade cost volume by integrating the coarse-to-fine strategy. Furthermore, PatchMatchNet [30] embeds PatchMatch MVS into a deep learning framework to achieve high + +quality reconstruction with low memory consumption. + +Although the existing deep learning methods have achieved great success in MVS, they usually require a large quantity of training datasets, which may not be practical in real-world applications. In addition, the generalization ability still remains an issue for learning-based methods [16]. As such, we still focus on traditional methods for MVS. + +# 3. Methodology + +In this section, we first briefly review the classical Patch-Match MVS. Then, we detail our proposed HPM-MVS. An overview of the method is illustrated in Fig. 2 (the pseudocode is shown in the supplementary). + +# 3.1. Review of PatchMatch MVS + +PatchMatch MVS [9, 25, 38] is a method for rapidly finding correspondences between image patches. To compute depth maps for input images, each image is selected in turn as a reference image and its corresponding source images are determined based on the Structure-from-Motion (SfM) results. In general, the process consists of four steps [36], i.e., random initialization, hypothesis propagation, multiview cost evaluation and refinement. The last three steps are iterated until convergence, the details are as follows: + +1) Random initialization means randomly generating a plane hypothesis for each pixel in the reference image based on the results of SfM. A key observation behind is that a non-trivial fraction of a large field of random offset assignments is likely to be a good guess. +2) Hypothesis propagation samples the hypotheses from neighboring pixels because of the high spatial coherence of + +![](images/af597405f6e5d7fc1736ad130a548b317cab60a7421af727d104ef3001b915ef.jpg) +Figure 3: Non-local Extensible Sampling Pattern. The circles represent the pixels in the red-black checkerboard. The red areas represent the sampling regions; the lighter the color, the less adaptive extension is performed. The solid yellow circles show the sampled points. + +neighbors in images. Moreover, a proper hypothesis can be propagated to a relatively large region to ensure the smoothness constraint of nearby pixels. + +3) Multi-view matching cost evaluation aims at robustly integrating matching cost from multiple views to select the best plane hypotheses. +4) Refinement generates two additional hypotheses to enrich the diversity of the solution space. The one with the lowest cost will be selected as the refined estimation. + +In this work, we concentrate on hypothesis propagation and multi-view matching cost evaluation, which are essential components of PatchMatch MVS methods, to capture non-local information for recovering better geometry in low-textured areas. + +# 3.2. Non-local Extensible Sampling Pattern + +Hypothesis propagation is a prerequisite for multi-view matching cost evaluation and refinement. Hence, choosing a reasonable sampling pattern will lead to accurate results. The pixels within a relatively large area can be represented by one of the pixels. So considering local information repeatedly is redundant. Our NESP is inspired by these two key insights to sample more robust hypotheses. Following the PatchMatch MVS pipeline introduced above, we design our basic MVS with NESP. + +Random Initialization. First, we randomly generate a hypothesis $\pmb{\theta}_{x} = [d_{x},\pmb{n}_{x}]$ for each pixel $x$ , where $d_{x}$ represents the depth information and $\pmb{n}_{x}$ is a normal vector. Second, a matching cost is calculated from each source image via homography [13]. Finally, the initial multi-view matching cost is obtained by averaging top- $k$ best costs. + +Hypothesis Propagation. The diffusion-like propagation scheme can tremendously enhance efficiency while maintaining high-quality hypothesis propagation. Following [38], we first partition all pixels in the reference image into a red-black checkerboard pattern which uses hypothe + +ses within eight regions as candidates to update. Second, we divide eight sampling areas into four long strip areas and four polyline areas; each polyline area starts with 8 samples, compared to each long strip area's initial 5 samples. Then, we consider two strategies to construct our NESP (Fig. 3). + +- Extensible Strategy. To achieve the purpose of extension, we first choose the best hypothesis in each area according to the multi-view matching costs from random initialization or the previous iteration (the initial costs in random initialization are used for the first propagation, while the other iterations use the costs computed in the previous iteration) and define the candidate hypothesis set as $\Theta = \{\pmb{\theta}_i | i = 1 \dots 8\}$ . Second, we calculate their matching costs with respect to different source images and embed them into a cost matrix, + +$$ +\boldsymbol {M} = \left[ \begin{array}{c c c c} m _ {1, 1} & m _ {1, 2} & \dots & m _ {1, N - 1} \\ m _ {2, 1} & m _ {2, 2} & \dots & m _ {2, N - 1} \\ \vdots & \vdots & \ddots & \vdots \\ m _ {8, 1} & m _ {8, 2} & \dots & m _ {8, N - 1} \end{array} \right], \tag {1} +$$ + +where $N$ is the number of input images, $m_{i,j}$ represents the matching cost of the sampled point in the $i$ -th region scored by the $j$ -th view. Based on this matrix, a voting scheme is implemented in each line to determine whether this region needs to be extended. Two thresholds are defined here: 1) A good matching cost threshold is, + +$$ +\tau \left(t _ {i t e r}, t _ {e x t}\right) = \tau_ {g o o d} \cdot e ^ {- \frac {t _ {i t e r} {} ^ {2} \cdot \left(N _ {e x t} - t _ {e x t}\right)}{\alpha}}, \tag {2} +$$ + +where $t_{iter}$ and $t_{ext}$ represent the $t_{iter}$ -th iteration of hypothesis propagation and the $t_{ext}$ -th regional expansion, respectively; $\tau_{good}$ is the initial good matching cost threshold, $\alpha$ is a constant and $N_{ext}$ represents the maximum number of extensions. 2) A bad matching cost threshold is defined as $\tau_{bad}$ . For a specific region, there should exist at least $n_{good}$ matching costs smaller than $\tau(t_{iter}, t_{ext})$ , and at most $n_{bad}$ matching costs meeting the condition: $m_{i,j} > \tau_{bad}$ . When the above conditions are satisfied, the expansion of the current sampling area will be stopped. Otherwise, this area should extend by itself while each extension doubles the number of sampling points in the area. Notably, the set $\Theta$ and the matrix $M$ are renewed dynamically in each expansion using the rule described above. + +- Non-local Strategy. There are two key insights behind the non-local strategy: 1) The surrounding pixels of the central point share the same structured region information, and their hypotheses can be contained by Eq. 4 when in the process of refinement. 2) The misestimates in low-textured areas always appear as small + +isolated speckles in the depth map; this phenomenon indicates that these regions have fallen into local optimal solutions. Therefore, we use the non-local strategy (Fig. 3) which abandons the sampling points within the radius $R$ of the center pixels. The motivation is to concentrate more on non-local areas instead of focusing on the local surrounding information. + +In the end, we combine these two strategies to form NESP, which helps to collect more reasonable candidate hypotheses from non-local neighboring pixels. + +Multi-View Matching Cost Evaluation. In order to determine the best hypothesis from the above candidate hypothesis set, we follow the previous studies [46, 25, 38] to define the multi-view matching cost via photo consistency in our basic MVS with NESP, + +$$ +c _ {p h o t o} \left(\boldsymbol {\theta} _ {i}\right) = \frac {\sum_ {j} w _ {j} \cdot m _ {i , j}}{\sum_ {j} w _ {j}}, \tag {3} +$$ + +where $w_{j}$ is the weight of the $j$ -th view, which is computed by the view selection strategy in [38]. The hypothesis with the least matching cost will be chosen as the current best estimate for the center pixel. + +Refinement. We generate two new hypotheses in refinement. One hypothesis is obtained by perturbing the current hypothesis $[d_x^p, n_x^p]$ , while the other is randomly generated $[d_x^r, n_x^r]$ . Subsequently, the two new hypotheses are randomly arranged and combined with the current hypothesis to form an ensemble of seven hypotheses, + +$$ +\left. \begin{array}{c} \left\{\left[ d _ {x}, \boldsymbol {n} _ {x} \right], \left[ d _ {x}, \boldsymbol {n} _ {x} ^ {p} \right], \left[ d _ {x}, \boldsymbol {n} _ {x} ^ {r} \right], \left[ d _ {x} ^ {r}, \boldsymbol {n} _ {x} \right], \right. \\ \left. \left[ d _ {x} ^ {r}, \boldsymbol {n} _ {x} ^ {r} \right], \left[ d _ {x} ^ {p}, \boldsymbol {n} _ {x} \right], \left[ d _ {x} ^ {p}, \boldsymbol {n} _ {x} ^ {p} \right] \right\}. \end{array} \right. \tag {4} +$$ + +The final estimation for pixel $x$ will be the hypothesis with the minimum cost. + +# 3.3. Planar Prior Construction + +The local image patches of low-textured areas are highly similar, and the photometric consistency, as a metric function of MVS, always leads to incorrect estimates due to its defects. Inspired by [40], we construct a better planar prior model to assist multi-view matching cost evaluation. After executing the basic MVS, a hypothesis with its final cost for each pixel can be obtained. A lower cost represents a more precise estimate, therefore we collect the credible correspondences into a set $I_{\text{cred}}$ , + +$$ +\boldsymbol {I} _ {\text {c r e d}} = \left\{\boldsymbol {\theta} _ {x} \mid c _ {\text {p h o t o}} \left(\boldsymbol {\theta} _ {x}\right) < \tau_ {\text {c r e d}} \right\}, \tag {5} +$$ + +where $\tau_{\text{red}}$ represents the credible threshold. Then, the Delaunay triangulation is applied to generate triangular surfaces of various sizes. Although these triangles cover a large area, there are still some marginal areas in the image that remain unaffected. This suggests that the local information + +![](images/97591c5205d532d941d8e59b1a360e296e2ac032cd52f8f7d3e34474504bdc10.jpg) + +![](images/5bf72a045e5878075f616bdf268b6b454a70c23998a9b35af8f7ce102a22b8bf.jpg) + +![](images/d08ee48d1d5c634b66e5167c9e9a435548f96f511d504e37ac86107f3af95751.jpg) + +![](images/47f40daefacd977002d3e1c9ece2ae137d60bdfe91cccf2fea9f19404ccab315.jpg) +(a) +(d) +Figure 4: Planar prior models at different scales and resulting depth maps. (a) The color image (the low-textured area is in the white box while pixels in the red box are details); (b) the planar prior model constructed at the low-resolution scale; (c) the planar prior model constructed at the high-resolution scale (the dark blue areas show that there are anomalies and have already been filtered out); (d) the depth map obtained without prior models; (e) the depth map based on the prior model constructed at the low-resolution scale; (f) the depth map based on prior models at different scales. + +![](images/240c2f46207ca9ec3919e69a5778c1727aeba93dbde47bd7150349b1a64dee78.jpg) +(b) +(e) + +![](images/23800d8bb04a57c8a3fbc6fe7648634d0242af18cc6640efd0e5c60f3825bee6.jpg) +(c) +(f) + +cannot satisfy the requirements of developing a planar prior model in these parts of the regions. To address this issue, we propose a method based on KNN to construct the planar prior model via non-local credible hypotheses. First, we set up a KD-tree [47] for the reliable points in $I_{\text{cred}}$ . Second, the top- $K$ nearest non-local neighbors are searched for pixels in the regions where prior construction is challenging. In addition, we use Heron's formula to avoid the case of co-linearity of proximity points. Then, each pixel will have three reliable neighboring points. We project them into the coordinate of the reference camera and set up a matrix $A$ to get the plane parameters $z_{opt}$ , + +$$ +\boldsymbol {z} _ {\text {o p t}} = \underset {\boldsymbol {z} ^ {*}: | | \boldsymbol {z} ^ {*} | | = 1} {\arg \min } \left| \left| \boldsymbol {A} \cdot \boldsymbol {z} ^ {*} \right| \right|, \tag {6} +$$ + +where $z^{*}$ is a unit-length variable. To further enhance the validity of the prior model, the ultimate prior knowledge for each image is filtered by its depth range. At last, we consider Planar Prior Assistance [40] which can better leverage the planar prior model to assist multi-view matching cost evaluation in low-textured areas. + +# 3.4. Hierarchical Prior Mining + +The multi-scale structure is employed in some Patch-Match MVS methods to infer the depth information of low-textured areas. Previous approaches [21, 38, 33] usually construct image pyramids to fuse depth estimation at different scales. In addition, several methods [40, 24] focus on constructing planar priors models at a single scale to help + +the depth estimation. In our work, we demonstrate that constructing planar prior models in a multi-scale way will further facilitate depth estimation. + +As depicted in Fig. 4(b), when the planar prior model is built in low-textured regions, the performance at the low-resolution scale is obviously more advanced than the high-resolution scale. Because the information is condensed after downsampling, resulting in a larger receptive field for each pixel, and a wider range of non-local information can be excavated than the original scale. However, only using the prior from the coarsest scale may lead ambiguity in detail areas. We find that the aforementioned problem of prior construction is expertly managed at the high-resolution scale (Fig. 4(c)). + +Therefore, the depth perception advantages of planar prior models at different scales are complementary. Inspired by this key observation, we design an HPM framework to construct prior models at different scales, which assists the depth estimation of low-textured regions without sacrificing details. The depth map can be continually optimized via the HPM framework (Fig. 4(d)-(f)). + +Excavating hierarchical prior is essentially an iterative coarse-to-fine optimization process (Fig. 2). After employing the basic MVS method with NESP, we first downsample the obtained hypotheses (i.e. depth and normal) to the coarsest scale and build a planar prior model. Second, with the joint bilateral upsampler [18], we propagate the planar prior to the original scale. Third, the planar prior model is embedded into the PatchMatch MVS method with Planar Prior Assistance [36] to approximate the depth in low-textured areas better. Finally, we downsample the newly generated hypotheses to the medium scale again, and repeat the same operation until iterating up to the original scale. For images with high resolutions, more medium scales can be considered. To further optimize the results, we apply geometric consistency as [25, 38] do. This way, the final depth estimation can achieve a better trade-off between details and low-textured areas. + +Moreover, to save computing resources and improve time efficiency, we propose a fast version of the HPM framework. This version only uses the planar prior model constructed at the coarsest scale to assist MVS. The rationale behind this is that the prior model obtained at the coarsest scale can effectively target the low-textured region and exactly compensate for the deficiencies of the basic MVS method in that region. Both HPM and its fast version will be experimentally analyzed. + +# 4. Experiments + +We evaluate our method on two challenging MVS datasets, ETH3D benchmark [26] and Tanks & Temples datasets [17], from two perspectives, i.e., benchmark performance and analysis experiments. + +Table 1: Point cloud evaluation on ETH3D benchmark [26]. We show accuracy / completeness / $F_{1}$ score (in %) at different thresholds (2cm and 10cm). The best results are marked in bold while the second-best results are underlined. + +
Method2cm10cm
Train.i) Traditional
COLMAP [25]91.85 / 55.13 / 67.6698.75 / 79.47 / 87.61
PCF-MVS [19]84.11 / 75.73 / 79.4295.98 / 90.42 / 92.98
ACMM [38]90.67 / 70.42 / 78.8698.12 / 86.40 / 91.70
ACMP [40]90.12 / 72.15 / 79.7997.97 / 87.15 / 92.03
ACMMP [36]91.03 / 77.27 / 83.4297.96 / 93.19 / 95.46
ii) Learning
PatchMatchNet [30]64.81 / 65.43 / 64.2189.98 / 83.28 / 85.70
IterMVS [29]73.62 / 61.87 / 66.3694.48 / 78.85 / 85.25
MVSTER [32]68.08 / 76.92 / 72.0691.97 / 91.91 / 91.73
HPM-MVSfast91.17 / 73.20 / 80.8698.23 / 92.41 / 94.97
HPM-MVS90.66 / 79.50 / 84.5897.97 / 95.59 / 96.22
Testi) Traditional
COLMAP [25]91.97 / 62.98 / 73.0198.25 / 84.54 / 90.40
PCF-MVS [19]82.15 / 79.29 / 80.3892.12 / 91.26 / 91.56
ACMM [38]90.65 / 74.34 / 80.7898.05 / 88.77 / 92.96
ACMP [40]90.45 / 75.58 / 81.5197.47 / 88.71 / 92.62
ACMMP [36]91.91 / 82.10 / 85.8998.05 / 94.67 / 96.27
ii) Learning
PatchMatchNet [30]79.71 / 77.46 / 73.1291.98 / 92.05 / 91.91
IterMVS [29]76.91 / 72.65 / 74.2995.42 / 85.81 / 90.15
MVSTER [32]77.09 / 82.47 / 79.0194.21 / 92.71 / 92.30
HPM-MVSfast92.50 / 80.25 / 85.3598.32 / 94.89 / 96.51
HPM-MVS92.13 / 83.25 / 87.1198.11 / 95.41 / 96.69
+ +# 4.1. Experimental Setup + +ETH3D high-resolution multi-view benchmark [26] consists of training datasets and test datasets, which contain indoor and outdoor images at a resolution of $6048 \times 4032^1$ . Tanks & Temples datasets are divided into Intermediate and Advanced datasets based on the complexity of the scene. However, these datasets do not provide official camera poses. In that case, we use COLMAP [25] to estimate camera parameters. + +All of our experiments are executed on a computer with an Intel i5-12600 CPU and an RTX 3070 GPU. In the basic MVS with NESP, $\{k, N_{ext}, \tau_{good}, \tau_{bad}, \alpha, n_{good}, n_{bad}, R\} = \{4, 3, 0.8, 1.2, 90, 1, 2, 4\}$ . In the process of planar prior construction, $\{\tau_{cred}, K\} = \{0.1, 6\}$ . + +# 4.2. Benchmark Performance + +Evaluation on ETH3D Benchmark. Both traditional methods [25, 19, 38, 40, 36] and learning-based [30, 29, 32] methods are considered for comparison. Table 1 reports the accuracy, completeness and $F_{1}$ score of the point clouds. + +As shown in the table, HPM-MVS can achieve outstanding performance in terms of completeness and $F_{1}$ score among all the methods, and HPM-MVS fast can outperform in accuracy while guaranteeing completeness and $F_{1}$ score. Notably, HPM-MVS obtains the highest completeness than other methods on both training and test datasets which con + +![](images/05e14e7e6da81555aca19be8d88dc6879a00bf56e246349bf0a70e056eb423f2.jpg) +(a) Images + +![](images/827904327740f1d3cb869735f7653944c122222ce6f93bcc841dce356c195e06.jpg) +(b) COLMAP [25] + +![](images/7f284a1f851b7c1e1eaf3162640edb1693b018d84557c13888a81bb67324a3bc.jpg) +(c)ACMMP [36] +Figure 5: Qualitative point cloud comparisons between different MVS methods on several high-resolution multi-view test scans (old computer and terrace 2) of ETH3D benchmark [26]. Some challenging areas are shown in red boxes. + +![](images/536571a31e4ed0743ce38e864d00358c58a2e09cc7986685400e3a7adf340ac4.jpg) +(d) IterMVS [29] + +![](images/1141416b38400e85332de8c61528263f9dfb5e682023c35a0cf27e1cdd9870b7.jpg) +(e) HPM-MVSfast + +![](images/e6b69daf01831a12ea420011182e78fc6801a427d9e035ee0b49d4ef58370f79.jpg) +(f) HPM-MVS + +Table 2: $F_{1}$ score (in %) comparisons of different methods on Tanks & Temples datasets [17] at given evaluation threshold. + +
MethodIntermediateAdvanced
TraditionalCOLMAP [25]42.1427.24
PCF-MVS [19]55.8834.59
ACMM [38]57.2734.02
ACMP [40]58.4137.44
ACMMP [36]59.3837.84
LearningR-MVSNet [44]48.4024.91
PatchMatch-RL [20]51.8131.78
PatchMatchNet [30]53.1532.31
IterMVS [29]56.2233.24
PVSNet [37]56.8833.46
Effi-MVS [31]56.8834.39
MVSTER [32]60.9237.53
OursHPM-MVSfast61.5939.65
HPM-MVS61.3940.80
+ +tain large low-textured areas, because HPM-MVS perceives non-local structured information for recovering geometry. The visual comparisons are shown in Fig. 5. It can be clearly observed that our methods produce higher quality point clouds than other competitors, especially in challenging low-textured areas, e.g., red boxes in Fig. 5. + +Evaluation on Tanks & Temples Datasets. To further illustrate the robustness of our methods, we test our methods on Tanks & Temples datasets without any fine-tuning. The quantitative results of state-of-the-art methods on both Intermediate and Advanced sets are reported in Table 2. + +Our methods achieve outstanding performance among all the methods. Even compared with learning-based methods, our methods still achieve better performance without any data training. Remarkably, HPM-MVS can greatly improve the performance of Advanced datasets which contain complex details and more low-textured areas difficult to recover. Our method ranks $\mathbf{1}^{st}$ in Advanced datasets over all the existing traditional MVS methods, and it outperforms the previous best record by $2.96\%$ . This result well confirms that our method can achieve a good balance for MVS reconstruction in details and low-textured areas. Fig. 6 + +presents some visualization examples of point cloud recall comparisons between different methods, which verifies our proposed methods can produce more complete 3D geometries than the others. These obvious advantages suggest that HPM-MVS and HPM-MVSfast hold powerful generalization ability and can adapt to different application scenarios without any data training. + +# 4.3. Analysis Experiments + +Ablation Study. In this section, we conduct an ablation study on the multi-view high-resolution training datasets of ETH3D benchmark to validate the performance of different parts in our proposed methods. We divide them into five sections, which are Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework. The baseline method is $\mathrm{ACMH}^2$ [38]. Table 3 summarizes the effectiveness of each individual part in our proposed methods. + +For the basic MVS method with NESP, we remove the no-local part and the extensible part, respectively. The results show that these two proposals can both enhance the completeness and $F_{1}$ score. In particular, the improvement of the non-local strategy is more obvious than the extensible strategy. In some circumstances, especially in low-textured areas, several severe errors always cause the hypothesis propagation into local optimal solutions. Hence, the non-local operation can effectively prevent this situation from getting worse. + +Based on NESP, the planar prior model constructed at the current scale generates a potential hypothesis for each pixel in the image. As for two versions of the HPM framework, one can make several observations from the results: 1) HPM and $\mathrm{HPM}_{\mathrm{fast}}$ achieve outstanding performance at relatively high thresholds (e.g., $5cm$ and $10cm$ ), because they have a strong focus on sizable low-textured areas. 2) We notice that $\mathrm{HPM}_{\mathrm{fast}}$ will suffer some loss of completeness at low thresholds (e.g., $1cm$ ), because it obtains the prior model + +![](images/b44c67490e7b44509f7e927195a22f77ad6cc1581a28f3692eee820db27d3e86.jpg) +(a) ACMM [38] + +![](images/cc2707246729b29fe26ccd694e56534f185981625b44a2ac759ed1f643ecfc8a.jpg) +(b)ACMMP [36] + +![](images/d4c729c1eb5b77f2923a7857434e9b78f985dcba6f4b9d673eb1bb830da6053d.jpg) +(c) IterMVS [29] +Figure 6: Qualitative point cloud recall comparisons between MVS methods on Tanks & Temples datasets [17] (Horse and Temple). The darker the pixel color is, the greater the error will be. $\tau$ is the recommended threshold for each dataset. + +![](images/c447d4aac95d29e091c5b3b45699a21fb33bbc9c56471fefffc27d20bc3beb49.jpg) +(d) MVSTER [32] + +![](images/a90f1108c6e7b1cb5ba66e589188dfb2834ac9d1c1d3e3f54af2dfe429a08e3a.jpg) +(e) HPM-MVSfast + +![](images/77bfd1b20116410f1aa48c24e8dc813953d4115d6ca12a78d1d86b8b835d368a.jpg) +(f) HPM-MVS + +Table 3: Comparative results of our methods under different settings on the training datasets of ETH3D benchmark [26]. N, E, PA, $\mathrm{HPM}_{\mathrm{fast}}$ and HPM mean Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework, respectively. + +
MethodSettingsETH3D train.
NEPAHPMfastHPM1cm2cm5cm10cm
ACMH [38]82.41 / 50.63 / 61.7191.09 / 63.04 / 73.6096.61 / 74.73 / 83.5098.25 / 82.20 / 88.57
NSP✓82.52 / 53.46 / 63.8490.86 / 65.59 / 75.2696.63 / 76.15 / 84.5298.46 / 82.40 / 89.19
ESP✓82.66 / 51.40 / 62.2290.93 / 63.83 / 74.0596.64 / 75.37 / 84.0498.47 / 82.09 / 89.03
NESP✓✓82.47 / 53.98 / 64.2290.78 / 66.02 / 75.5996.60 / 76.63 / 84.8798.47 / 82.73 / 89.44
NESP+PA✓✓✓82.88 / 62.37 / 71.3290.51 / 74.96 / 81.6696.48 / 84.29 / 88.9097.89 / 88.57 / 92.73
HPMfast✓✓83.32 / 53.53 / 64.6691.16 / 70.48 / 79.0496.39 / 84.67 / 89.8497.89 / 88.57 / 92.73
HPM✓✓83.01 / 62.07 / 70.6790.65 / 77.03 / 83.0696.06 / 88.42 / 91.9797.91 / 93.62 / 95.66
HPM-MVSfast✓✓✓✓82.95 / 57.36 / 67.3991.17 / 73.20 / 80.8696.51 / 86.22 / 90.8998.23 / 92.14 / 94.97
HPM-MVS✓✓✓✓82.93 / 65.16 / 72.7390.66 / 79.50 / 84.5896.13 / 89.98 / 92.8897.97 / 94.59 / 96.22
+ +Table 4: Generalization performance of NESP on ETH3D high-resolution training datasets [26]. + +
MethodETH3D train.
1cm2cm5cm10cm20cm
ACMM [38]67.5878.8687.6891.7094.41
ACMP [40]68.7279.7988.3292.0394.43
ACMMP [36]71.5783.4292.0395.4697.37
ACMM+NESP70.7081.0188.8092.2694.55
3.12↑2.15↑1.12↑0.56↑0.14↑
ACMP+NESP70.8781.4589.4392.7294.78
2.15↑1.66↑1.11↑0.69↑0.35↑
ACMMP+NESP74.5485.3393.2596.4597.99
2.97↑1.91↑1.22↑0.99↑0.62↑
+ +at the lowest resolution scale which may lead to ambiguities in details. 3) The method with NESP as the basis for subsequent processes can provide a set of reliable points for planar prior model construction. + +Generalization Performance of NESP. In addition, we perform a more extensive comparison on the multi-view high-resolution training datasets of ETH3D benchmark to show the generalization ability of NESP. Several kinds of state-of-the-art diffusion-like propagation methods are integrated with NESP for evaluation. The considered methods are ACMM [38], ACMP [40] and ACMMP [36]. Note that ACMMC is a very recent state-of-the-art geometric-only + +method. Each method is tested under different thresholds, and results are shown in Table 4. + +Impressively, NESP dramatically improves the $F_{1}$ score for all tested methods at relatively low thresholds, e.g., 1cm and 2cm. The improvement of NESP working with multi-scale strategy is more obvious, because the non-local and extensible strategy works even better at low-resolution scale. Notably, the combination of NESP and ACMMP outperforms all the competitors and achieves state-of-the-art $F_{1}$ score of $85.33\% / 96.45\%$ at 2cm/10cm thresholds. Based on the results, the following conclusions can be drawn: 1) NESP is a generalized module that can be adapted to other MVS methods and boost their performance. 2) Combined with the multi-scale strategy, NESP can achieve better performance. + +More analysis experiments and point cloud visualizations can be found in the supplementary. + +# 5. Conclusion + +In this paper, we presented an HPM-MVS method which can efficiently perceive non-local structured information to achieve high-quality reconstruction. Based on the non-local and extensible strategy, we first propose our NESP module. Focusing on the planar prior construction in marginal regions, we employ KNN to search non-local credible points + +and obtain potential hypotheses. Further, we design an HPM framework which explores planar prior at different scales to assist MVS. Through extensive experiments, we have demonstrated the rationality of each key component of our method and its superiority against existing methods. In the future, we plan to combine our method with a probabilistic graphical model to further improve its ability to handle the reconstruction of details and low-textured regions. + +Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (No. 62002295), and NWPU international cooperation and exchange promotion projects (No. 02100-23GH0501). + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 1 +[2] Christian Bailer, Manuel Finckh, and Hendrik Lensch. Scale robust multi view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 398-411. Springer, 2012. 2 +[3] Caroline Baillard and Andrew Zisserman. A plane-sweep strategy for the 3d reconstruction of buildings from multiple images. International Archives of Photogrammetry and Remote Sensing, 33(B2; PART 2):56-62, 2000. 1 +[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Transactions on Graphics, 28(3):24, 2009. 2 +[5] Fabio Bruno, Stefano Bruno, Giovanna De Sensi, Maria-Laura Luchi, Stefania Mancuso, and Maurizio Muzzupappa. From 3d reconstruction to virtual reality: A complete methodology for digital archaeological exhibition. Journal of Cultural Heritage, 11(1):42-49, 2010. 1 +[6] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2020. 3 +[7] Robert T Collins. A space-sweep approach to true multi-image matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 358-363. IEEE, 1996. 1 +[8] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2009. 1 +[9] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 1, 2, 3 +[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, Qingxiong Yang, and Marc Pollefeys. Real-time plane-sweeping stereo with multiple sweeping directions. In Pro- + +ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1 +[11] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1418-1425. IEEE, 2010. 2 +[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 3 +[13] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 4 +[14] Mohsen Hejrati and Deva Ramanan. Analysis by synthesis: 3d object recognition by object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2456, 2014. 1 +[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 1 +[16] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016. 1, 3 +[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 1, 6, 7, 8 +[18] Johannes Kopf, Michael F Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Transactions on Graphics, 26(3):96-es, 2007. 6 +[19] Andreas Kuhn, Shan Lin, and Oliver Erdler. Plane completion and filtering for multi-view stereo reconstruction. In Proceedings of the German Conference on Pattern Recognition, pages 18-32. Springer, 2019. 6, 7 +[20] Jae Yong Lee, Joseph DeGol, Chuhang Zou, and Derek Hoiem. Patchmatch-rl: Deep mvs with pixelwise depth, normal, and visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6158–6167, 2021. 7 +[21] Jie Liao, Yanping Fu, Qingan Yan, and Chunxia Xiao. Pyramid multi-view stereo with local consistency. In Computer Graphics Forum, volume 38, pages 335-346. Wiley Online Library, 2019. 5 +[22] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1 +[23] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE International Conference on Computer Vision, pages 6851–6860, 2019. 1 + +[24] Andrea Romanoni and Matteo Matteucci. Tapa-mvs: Textureless-aware patchmatch multi-view stereo. In Proceedings of the IEEE International Conference on Computer Vision, pages 10413-10422, 2019. 2, 5 +[25] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 3, 5, 6, 7 +[26] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 1, 6, 7, 8 +[27] Qi Shan, Riley Adams, Brian Curless, Yasutaka Furukawa, and Steven M Seitz. The visual turing test for scene reconstruction. In International Conference on 3D Vision-3DV, pages 25–32. IEEE, 2013. 1 +[28] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. IEEE Transactions on Image Processing, 22(5):1901-1914, 2013. 2 +[29] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8606-8615, 2022. 1, 6, 7, 8 +[30] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14194-14203, 2021. 1, 3, 6, 7 +[31] Shaoqian Wang, Bo Li, and Yuchao Dai. Efficient multiview stereo by iterative dynamic cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8655-8664, 2022. 7 +[32] Xiaofeng Wang, Zheng Zhu, Guan Huang, Fangbo Qin, Yun Ye, Yijia He, Xu Chi, and Xingang Wang. Mvster: Epipolar transformer for efficient multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 573-591. Springer, 2022. 1, 6, 7, 8 +[33] Yuesong Wang, Tao Guan, Zhuo Chen, Yawei Luo, Keyang Luo, and Lili Ju. Mesh-guided multi-view stereo with pyramid architecture. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2039-2048, 2020. 5 +[34] Jian Wei, Benjamin Resch, and Hendrik PA Lensch. Multiview depth map estimation with cross-view consistency. In Proceedings of the British Machine Vision Conference, 2014. 2 +[35] Oliver Woodford, Philip Torr, Ian Reid, and Andrew Fitzgibbon. Global stereo reconstruction under second-order smoothness priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(12):2115-2128, 2009. 2 +[36] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3, 6, 7, 8 + +[37] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 1, 7 +[38] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 1, 2, 3, 4, 5, 6, 7, 8 +[39] Qingshan Xu and Wenbing Tao. Learning inverse depth regression for multi-view stereo with correlation cost volume. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12508-12515, 2020. 3 +[40] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 2, 3, 5, 6, 7, 8 +[41] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5981-5990, 2020. 2 +[42] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3 +[43] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 767-783, 2018. 1, 3 +[44] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvvsnet for high-resolution multiview stereo depth inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 1, 3, 7 +[45] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A largescale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 1 +[46] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2, 5 +[47] Kun Zhou, Qiming Hou, Rui Wang, and Baining Guo. Real-time kd-tree construction on graphics hardware. ACM Transactions on Graphics, 27(5):1-11, 2008. 5 \ No newline at end of file diff --git a/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/images.zip b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..90892f47a7eb9041642d398a9d8c45474a9f2634 --- /dev/null +++ b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee0d9c6ae5233f164b69102edb08f37462b35e5104d516218d0fcdd9d83e5b95 +size 644119 diff --git a/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/layout.json b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e062f0b9f73fed105409c186a1b31fe517c0cba --- /dev/null +++ b/2023/Hierarchical Prior Mining for Non-local Multi-View Stereo/layout.json @@ -0,0 +1,8881 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 479, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 479, + 120 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 479, + 120 + ], + "type": "text", + "content": "Hierarchical Prior Mining for Non-local Multi-View Stereo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 142, + 470, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 142, + 470, + 158 + ], + "spans": [ + { + "bbox": [ + 109, + 142, + 470, + 158 + ], + "type": "text", + "content": "Chunlin Ren1 Qingshan Xu2 Shikun Zhang1 Jiaqi Yang1*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 158, + 483, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 158, + 483, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 158, + 483, + 171 + ], + "type": "text", + "content": "1 Northwestern Polytechnical University 2 Nanyang Technological University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 174, + 477, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 174, + 477, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 174, + 477, + 185 + ], + "type": "text", + "content": "{renchunlin, zhangshikun}@mail.nwpu.edu.cn; qingshan.xu@ntu.edu.sg;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 248, + 188, + 346, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 188, + 346, + 198 + ], + "spans": [ + { + "bbox": [ + 248, + 188, + 346, + 198 + ], + "type": "text", + "content": "jqyang@nwpu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 252, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 540 + ], + "type": "text", + "content": "As a fundamental problem in computer vision, multiview stereo (MVS) aims at recovering the 3D geometry of a target from a set of 2D images. Recent advances in MVS have shown that it is important to perceive non-local structured information for recovering geometry in low-textured areas. In this work, we propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS). The key characteristics are the following techniques that exploit non-local information to assist MVS: 1) A Non-local Extensible Sampling Pattern (NESP), which is able to adaptively change the size of sampled areas without becoming snared in locally optimal solutions. 2) A new approach to leverage non-local reliable points and construct a planar prior model based on K-Nearest Neighbor (KNN), to obtain potential hypotheses for the regions where prior construction is challenging. 3) A Hierarchical Prior Mining (HPM) framework, which is used to mine extensive non-local prior information at different scales to assist 3D model recovery, this strategy can achieve a considerable balance between the reconstruction of details and low-textured areas. Experimental results on the ETH3D and Tanks & Temples have verified the superior performance and strong generalization capability of our method. Our code will be available at https://github.com/CLinuxx/HPM-MVS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 563, + 128, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 128, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 128, + 576 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "content": "Multi-View Stereo (MVS) is one of the key problems in the field of 3D computer vision, which contributes greatly to virtual reality [5], 3D object recognition [14], and autonomous driving [23]. The objective of MVS method is to reconstruct the 3D geometry of a target from a series of images along with camera parameters. Over the past few years, the emergence of numerous datasets [15, 26, 45] has tremendously contributed to the continuous advancement of MVS methods [8, 27, 9]. However, influenced by low" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 225, + 426, + 346 + ], + "blocks": [ + { + "bbox": [ + 307, + 225, + 426, + 346 + ], + "lines": [ + { + "bbox": [ + 307, + 225, + 426, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 225, + 426, + 346 + ], + "type": "image", + "image_path": "6dfe36af85d2e859cf7240f782c9b071901f876893459ad415cabc92d7a20d11.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 354, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 354, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 546, + 392 + ], + "type": "text", + "content": "Figure 1: " + }, + { + "bbox": [ + 305, + 354, + 546, + 392 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 305, + 354, + 546, + 392 + ], + "type": "text", + "content": " score " + }, + { + "bbox": [ + 305, + 354, + 546, + 392 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 305, + 354, + 546, + 392 + ], + "type": "text", + "content": " comparisons with SOTA traditional [25, 38, 36] and learning-based [30, 29, 32] MVS methods on ETH3D [26] and Tanks & Temples [17]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 225, + 543, + 346 + ], + "blocks": [ + { + "bbox": [ + 426, + 225, + 543, + 346 + ], + "lines": [ + { + "bbox": [ + 426, + 225, + 543, + 346 + ], + "spans": [ + { + "bbox": [ + 426, + 225, + 543, + 346 + ], + "type": "image", + "image_path": "a5e7581498d2fe79dff5922272199a7392435be2a31ed3bedecc612028357034.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 437 + ], + "type": "text", + "content": "textured areas, illumination variation and other factors [1], MVS is still a challenging problem." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 437, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 605 + ], + "type": "text", + "content": "To perform an accurate and robust 3D reconstruction, two different types of methods have been investigated, including learning-based MVS [43, 44, 37, 22] and traditional MVS [9, 38, 36]. Learning-based MVS implements deep networks to extract high-level features and make predictions. However, it needs a large amount of training and its generalization ability still needs further improvement [16]. Traditional MVS can also be broadly classified into two categories: plane-sweeping-based MVS [7, 3, 10] and PatchMatch MVS [9, 25, 38]. Although plane-sweeping-based MVS produces good results for sufficiently textured and unoccluded surfaces, it performs poorly for scenes composed of large planar surfaces, while PatchMatch MVS has successfully overcome this restriction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "PatchMatch MVS usually consists of four steps, including random initialization, hypothesis propagation, multiview matching cost evaluation and refinement [36]. In essence, it is a process of sampling and verification. The hypothesis propagation samples appropriate hypotheses from neighboring pixels to construct solution space while the multi-view matching cost evaluation defines a criterion to verify the reliability of the sampled hypotheses. Therefore, these two steps play an important role in PatchMatch" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3611" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "MVS. Following this popular four-step pipeline, Patch-Match MVS produces remarkable outcomes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 286, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 286, + 206 + ], + "type": "text", + "content": "Even so, geometry recovery still suffers a lot from low-textured areas because the previous PatchMatch MVS methods place too much emphasis on local information. Therefore, we exploit non-local information to achieve high-quality reconstruction. We propose a Hierarchical Prior Mining for Non-local Multi-View Stereo (HPM-MVS) method, which defines a novel hypothesis propagation pattern and a novel way for constructing planar prior models to assist multi-view matching cost evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 286, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 286, + 448 + ], + "type": "text", + "content": "In terms of hypothesis propagation, sequential propagation [2, 46, 25] and diffusion-like propagation [9, 38] are all popular strategies. The previous one only considers the nearest pixels for updating. In contrast, the diffusion-like propagation can update simultaneously based on the neighbor hypotheses around the central point. To further leverage structured region information, ACMH [38] proposes an adaptive checkerboard sampling scheme. Although these previous approaches have greatly improved the efficiency and performance of hypothesis propagation, they both process in the local neighborhood; thus misestimates can only be updated when several better hypotheses in the local area are sampled. Based on the above observations, we first propose the basic method with a Non-local Extensible Sampling Pattern (NESP). Intuitively, a non-local operation removes sampling points around the center point. It allows distant correct pixels to have a greater possibility to contribute to the update of the current pixel. Moreover, the extensible architecture maintains the variable sampling size and can efficiently select suitable candidate hypotheses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 450, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 713 + ], + "type": "text", + "content": "In terms of the multi-view matching cost evaluation, since the photometric consistency usually causes ambiguities during the depth optimization of low-textured areas, many previous methods [35, 11, 40] introduce planar prior models to help the evaluation. They assume low-textured areas are usually on smooth homogeneous surfaces, and construct a planar prior model at the current scale to assist depth estimation. To establish a more robust and more comprehensive planar prior, we employ K-Nearest Neighbor (KNN) to search non-local reliable points and obtain potential hypotheses for the marginal regions where prior construction is difficult. Inspired by the coarse-to-fine strategy, we further design an HPM framework. In a coarse-to-fine manner, the prior knowledge at the early stages is built upon non-local credible sparse correspondences at low-resolution scales, which expands the receptive field and leads to an effective depth estimate for smooth homogeneous surfaces. Subsequently, the prior knowledge at the later stages is constructed at the higher-resolution scales by using previously rectified hypotheses, which restores the depth information of details in the image. This way, the depth information can be successfully recovered by our HPM architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "Extensive experiments on different competitive datasets verify the effectiveness of our method (Fig. 1). The results demonstrate that HPM-MVS has excellent performance, and also holds strong generalization capability to more complex scenes. In a nutshell, our contributions are three-fold as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 149, + 545, + 294 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 149, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 149, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 149, + 545, + 186 + ], + "type": "text", + "content": "- We present an NESP module to adaptively determine the number of potential sampling points while avoiding becoming snared in locally optimal solutions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 193, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 193, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 545, + 252 + ], + "type": "text", + "content": "- To construct a more robust and more comprehensive planar prior model, we propose an approach based on KNN to search non-local neighbor reliable hypotheses and construct a planar prior model for marginal areas where prior construction is challenging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 259, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 259, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 259, + 545, + 294 + ], + "type": "text", + "content": "- We design an HPM framework to explore prior knowledge at multiple scales, which efficiently balances out the reconstruction of details and low-textured areas." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 304, + 392, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 392, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 392, + 317 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 324, + 469, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 469, + 335 + ], + "type": "text", + "content": "2.1. Traditional Multi-View Stereo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 653 + ], + "type": "text", + "content": "PatchMatch Multi-View Stereo. The PatchMatch method was initially proposed by Barnes et al. [4], and its primary goal is to quickly identify the approximate nearest neighbor among patches of two images. The extension of PatchMatch idea to MVS was proposed by Shen [28]. In order to enhance efficiency, several methods [46, 2, 34] employ the sequential propagation scheme. However, this propagation scheme is still time-consuming. To this end, Galliani et al. [9] proposed Gipuma, a massively parallel multi-view extension of PatchMatch MVS which leverages a red-black checkerboard pattern to perform a diffusion-like propagation scheme. This makes better use of the parallelism of GPUs to improve efficiency. Nonetheless, the major drawback is that each pixel's sampling points are preset, which in turn causes regional information to be ignored. Further, Xu and Tao [38] designed a method with Adaptive Checkerboard sampling and Multi-Hypothesis joint view selection (ACMH) to address this shortcoming. In allusion to depth estimation for low-textured areas, they further combined ACMH with multi-scale geometric consistency guidance to present ACMM. In addition, MARMVS [41] selects the optimal scale for each pixel to alleviate the ambiguity caused by regions with raw texture. Although the aforementioned approaches significantly enhance the performance of PatchMatch MVS method, how to recover the depth information of low-textured regions is always a challenging problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Planar Prior Assistance. In order to deal with this longstanding ill-posed problem, the planar prior is considered. Romanoni et al. [24] proposed TAPA-MVS, they separated the image into two different scales of superpixels and utilized these as regional structures. The authors fitted a plane" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3612" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 496, + 251 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 496, + 251 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 496, + 251 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 496, + 251 + ], + "type": "image", + "image_path": "d2d72280a756345706a9503d76bd3c741e8f8459af11f69f511be28227fe8a04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 548, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 548, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 548, + 308 + ], + "type": "text", + "content": "Figure 2: An overview of HPM-MVS. Starting with the input images, we first apply the basic MVS with NESP (Fig. 3) to obtain the initial hypotheses. Then, we downsample them to the coarsest scale and generate the prior model. Third, we upsample the model to the current scale and leverage it to assist hypothesis prediction. After operating on different scales, we further use geometric consistency (GC) to optimize the results. Finally, the hypotheses will be fused to a point cloud." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 327, + 289, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 520 + ], + "type": "text", + "content": "within the superpixels by using trustworthy 3D points and the RANSAC method, then leveraged the prior hypotheses to optimize the low-textured regions. ACMP [40] constructs a planar prior depth map based on triangulations, and subsequently mosaics the prior hypotheses into the MVS process via a probabilistic graphical model. Recently, Xu et al. [36] designed a multi-scale geometric consistency guided and planar prior assisted Multi-View Stereo (ACMMP) by combining ACMM [38] and ACMP [40]. It fully exploits depth information at various scales and leverages prior assistance to efficiently target low-textured areas. Previous approaches only consider the prior information from the current image scale, we nonetheless believe that prior hypotheses should not be restricted to just one scale. To this aim, we embrace an HPM framework to discover more meaningful and comprehensive non-local prior knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 527, + 233, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 233, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 233, + 540 + ], + "type": "text", + "content": "2.2. Learning-based Multi-View Stereo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 546, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 715 + ], + "type": "text", + "content": "Deep learning has been widely used in 3D vision and has proven to be critical for MVS. MVSNet [43] introduces an end-to-end deep learning network that builds the 3D cost volumes from 2D image features. R-MVSNet [44] regularizes the cost volume along the depth direction with the convolutional GRU to make reconstruction feasible. To further decrease memory and boost performance, CIDER [39] builds a lightweight cost volume using an average group-wise correlation similarity metric. In recent years, the efficiency of learning-based MVS gains more attention. CVP-MVSNet [42], CasMVSNet [12] and UCS-Net [6] build a cascade cost volume by integrating the coarse-to-fine strategy. Furthermore, PatchMatchNet [30] embeds PatchMatch MVS into a deep learning framework to achieve high" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 327, + 526, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 526, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 526, + 340 + ], + "type": "text", + "content": "quality reconstruction with low memory consumption." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 411 + ], + "type": "text", + "content": "Although the existing deep learning methods have achieved great success in MVS, they usually require a large quantity of training datasets, which may not be practical in real-world applications. In addition, the generalization ability still remains an issue for learning-based methods [16]. As such, we still focus on traditional methods for MVS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 424, + 389, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 389, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 389, + 438 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 445, + 546, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 494 + ], + "type": "text", + "content": "In this section, we first briefly review the classical Patch-Match MVS. Then, we detail our proposed HPM-MVS. An overview of the method is illustrated in Fig. 2 (the pseudocode is shown in the supplementary)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 502, + 463, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 463, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 463, + 514 + ], + "type": "text", + "content": "3.1. Review of PatchMatch MVS" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "text", + "content": "PatchMatch MVS [9, 25, 38] is a method for rapidly finding correspondences between image patches. To compute depth maps for input images, each image is selected in turn as a reference image and its corresponding source images are determined based on the Structure-from-Motion (SfM) results. In general, the process consists of four steps [36], i.e., random initialization, hypothesis propagation, multiview cost evaluation and refinement. The last three steps are iterated until convergence, the details are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 629, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 305, + 629, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 547, + 689 + ], + "type": "text", + "content": "1) Random initialization means randomly generating a plane hypothesis for each pixel in the reference image based on the results of SfM. A key observation behind is that a non-trivial fraction of a large field of random offset assignments is likely to be a good guess." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "2) Hypothesis propagation samples the hypotheses from neighboring pixels because of the high spatial coherence of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3613" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 72, + 249, + 183 + ], + "blocks": [ + { + "bbox": [ + 89, + 72, + 249, + 183 + ], + "lines": [ + { + "bbox": [ + 89, + 72, + 249, + 183 + ], + "spans": [ + { + "bbox": [ + 89, + 72, + 249, + 183 + ], + "type": "image", + "image_path": "af597405f6e5d7fc1736ad130a548b317cab60a7421af727d104ef3001b915ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 289, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 255 + ], + "type": "text", + "content": "Figure 3: Non-local Extensible Sampling Pattern. The circles represent the pixels in the red-black checkerboard. The red areas represent the sampling regions; the lighter the color, the less adaptive extension is performed. The solid yellow circles show the sampled points." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 312 + ], + "type": "text", + "content": "neighbors in images. Moreover, a proper hypothesis can be propagated to a relatively large region to ensure the smoothness constraint of nearby pixels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 313, + 287, + 384 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 348 + ], + "type": "text", + "content": "3) Multi-view matching cost evaluation aims at robustly integrating matching cost from multiple views to select the best plane hypotheses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 350, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 384 + ], + "type": "text", + "content": "4) Refinement generates two additional hypotheses to enrich the diversity of the solution space. The one with the lowest cost will be selected as the refined estimation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 385, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 445 + ], + "type": "text", + "content": "In this work, we concentrate on hypothesis propagation and multi-view matching cost evaluation, which are essential components of PatchMatch MVS methods, to capture non-local information for recovering better geometry in low-textured areas." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 454, + 251, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 251, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 251, + 467 + ], + "type": "text", + "content": "3.2. Non-local Extensible Sampling Pattern" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 473, + 287, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 580 + ], + "type": "text", + "content": "Hypothesis propagation is a prerequisite for multi-view matching cost evaluation and refinement. Hence, choosing a reasonable sampling pattern will lead to accurate results. The pixels within a relatively large area can be represented by one of the pixels. So considering local information repeatedly is redundant. Our NESP is inspired by these two key insights to sample more robust hypotheses. Following the PatchMatch MVS pipeline introduced above, we design our basic MVS with NESP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": "Random Initialization. First, we randomly generate a hypothesis " + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{x} = [d_{x},\\pmb{n}_{x}]" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": " for each pixel " + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "inline_equation", + "content": "d_{x}" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": " represents the depth information and " + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{n}_{x}" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": " is a normal vector. Second, a matching cost is calculated from each source image via homography [13]. Finally, the initial multi-view matching cost is obtained by averaging top-" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": " best costs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Hypothesis Propagation. The diffusion-like propagation scheme can tremendously enhance efficiency while maintaining high-quality hypothesis propagation. Following [38], we first partition all pixels in the reference image into a red-black checkerboard pattern which uses hypothe" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "ses within eight regions as candidates to update. Second, we divide eight sampling areas into four long strip areas and four polyline areas; each polyline area starts with 8 samples, compared to each long strip area's initial 5 samples. Then, we consider two strategies to construct our NESP (Fig. 3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 143, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 545, + 262 + ], + "type": "text", + "content": "- Extensible Strategy. To achieve the purpose of extension, we first choose the best hypothesis in each area according to the multi-view matching costs from random initialization or the previous iteration (the initial costs in random initialization are used for the first propagation, while the other iterations use the costs computed in the previous iteration) and define the candidate hypothesis set as " + }, + { + "bbox": [ + 317, + 143, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\pmb{\\theta}_i | i = 1 \\dots 8\\}" + }, + { + "bbox": [ + 317, + 143, + 545, + 262 + ], + "type": "text", + "content": ". Second, we calculate their matching costs with respect to different source images and embed them into a cost matrix," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 271, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 271, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 349, + 271, + 545, + 328 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} = \\left[ \\begin{array}{c c c c} m _ {1, 1} & m _ {1, 2} & \\dots & m _ {1, N - 1} \\\\ m _ {2, 1} & m _ {2, 2} & \\dots & m _ {2, N - 1} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ m _ {8, 1} & m _ {8, 2} & \\dots & m _ {8, N - 1} \\end{array} \\right], \\tag {1}", + "image_path": "a9f567b920cb6a2f355d4514ce92d95ee7069b171e64d32fe70708ec06a711f9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "text", + "content": " is the number of input images, " + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "inline_equation", + "content": "m_{i,j}" + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "text", + "content": " represents the matching cost of the sampled point in the " + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "text", + "content": "-th region scored by the " + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 324, + 333, + 545, + 416 + ], + "type": "text", + "content": "-th view. Based on this matrix, a voting scheme is implemented in each line to determine whether this region needs to be extended. Two thresholds are defined here: 1) A good matching cost threshold is," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 342, + 426, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 426, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 342, + 426, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\tau \\left(t _ {i t e r}, t _ {e x t}\\right) = \\tau_ {g o o d} \\cdot e ^ {- \\frac {t _ {i t e r} {} ^ {2} \\cdot \\left(N _ {e x t} - t _ {e x t}\\right)}{\\alpha}}, \\tag {2}", + "image_path": "6364e910215ae8fb594e7b4dd2ff5a3915800ded4725b1754d9933796b1f02e3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "t_{iter}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "t_{ext}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "t_{iter}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": "-th iteration of hypothesis propagation and the " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "t_{ext}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": "-th regional expansion, respectively; " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\tau_{good}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " is the initial good matching cost threshold, " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " is a constant and " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "N_{ext}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " represents the maximum number of extensions. 2) A bad matching cost threshold is defined as " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\tau_{bad}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": ". For a specific region, there should exist at least " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "n_{good}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " matching costs smaller than " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\tau(t_{iter}, t_{ext})" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": ", and at most " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "n_{bad}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " matching costs meeting the condition: " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "m_{i,j} > \\tau_{bad}" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": ". When the above conditions are satisfied, the expansion of the current sampling area will be stopped. Otherwise, this area should extend by itself while each extension doubles the number of sampling points in the area. Notably, the set " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " and the matrix " + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 324, + 453, + 545, + 632 + ], + "type": "text", + "content": " are renewed dynamically in each expansion using the rule described above." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 641, + 545, + 713 + ], + "type": "text", + "content": "- Non-local Strategy. There are two key insights behind the non-local strategy: 1) The surrounding pixels of the central point share the same structured region information, and their hypotheses can be contained by Eq. 4 when in the process of refinement. 2) The misestimates in low-textured areas always appear as small" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3614" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 289, + 156 + ], + "type": "text", + "content": "isolated speckles in the depth map; this phenomenon indicates that these regions have fallen into local optimal solutions. Therefore, we use the non-local strategy (Fig. 3) which abandons the sampling points within the radius " + }, + { + "bbox": [ + 66, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 66, + 72, + 289, + 156 + ], + "type": "text", + "content": " of the center pixels. The motivation is to concentrate more on non-local areas instead of focusing on the local surrounding information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 165, + 287, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 287, + 201 + ], + "type": "text", + "content": "In the end, we combine these two strategies to form NESP, which helps to collect more reasonable candidate hypotheses from non-local neighboring pixels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 201, + 287, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 287, + 260 + ], + "type": "text", + "content": "Multi-View Matching Cost Evaluation. In order to determine the best hypothesis from the above candidate hypothesis set, we follow the previous studies [46, 25, 38] to define the multi-view matching cost via photo consistency in our basic MVS with NESP," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 267, + 287, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 267, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 108, + 267, + 287, + 298 + ], + "type": "interline_equation", + "content": "c _ {p h o t o} \\left(\\boldsymbol {\\theta} _ {i}\\right) = \\frac {\\sum_ {j} w _ {j} \\cdot m _ {i , j}}{\\sum_ {j} w _ {j}}, \\tag {3}", + "image_path": "50b6414aa1039f69503400ac98c718ae8afcf66b8e9b96b130c25b0281917312.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 304, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 287, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 304, + 287, + 353 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 47, + 304, + 287, + 353 + ], + "type": "text", + "content": " is the weight of the " + }, + { + "bbox": [ + 47, + 304, + 287, + 353 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 304, + 287, + 353 + ], + "type": "text", + "content": "-th view, which is computed by the view selection strategy in [38]. The hypothesis with the least matching cost will be chosen as the current best estimate for the center pixel." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 353, + 288, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 288, + 425 + ], + "type": "text", + "content": "Refinement. We generate two new hypotheses in refinement. One hypothesis is obtained by perturbing the current hypothesis " + }, + { + "bbox": [ + 47, + 353, + 288, + 425 + ], + "type": "inline_equation", + "content": "[d_x^p, n_x^p]" + }, + { + "bbox": [ + 47, + 353, + 288, + 425 + ], + "type": "text", + "content": ", while the other is randomly generated " + }, + { + "bbox": [ + 47, + 353, + 288, + 425 + ], + "type": "inline_equation", + "content": "[d_x^r, n_x^r]" + }, + { + "bbox": [ + 47, + 353, + 288, + 425 + ], + "type": "text", + "content": ". Subsequently, the two new hypotheses are randomly arranged and combined with the current hypothesis to form an ensemble of seven hypotheses," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 432, + 287, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 432, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 90, + 432, + 287, + 458 + ], + "type": "interline_equation", + "content": "\\left. \\begin{array}{c} \\left\\{\\left[ d _ {x}, \\boldsymbol {n} _ {x} \\right], \\left[ d _ {x}, \\boldsymbol {n} _ {x} ^ {p} \\right], \\left[ d _ {x}, \\boldsymbol {n} _ {x} ^ {r} \\right], \\left[ d _ {x} ^ {r}, \\boldsymbol {n} _ {x} \\right], \\right. \\\\ \\left. \\left[ d _ {x} ^ {r}, \\boldsymbol {n} _ {x} ^ {r} \\right], \\left[ d _ {x} ^ {p}, \\boldsymbol {n} _ {x} \\right], \\left[ d _ {x} ^ {p}, \\boldsymbol {n} _ {x} ^ {p} \\right] \\right\\}. \\end{array} \\right. \\tag {4}", + "image_path": "728eefab0e3672bb96956cd905ab3d30f780ccb1cc83c995e0d092927ca80530.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 465, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 287, + 489 + ], + "type": "text", + "content": "The final estimation for pixel " + }, + { + "bbox": [ + 47, + 465, + 287, + 489 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 465, + 287, + 489 + ], + "type": "text", + "content": " will be the hypothesis with the minimum cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 497, + 194, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 194, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 194, + 509 + ], + "type": "text", + "content": "3.3. Planar Prior Construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "type": "text", + "content": "The local image patches of low-textured areas are highly similar, and the photometric consistency, as a metric function of MVS, always leads to incorrect estimates due to its defects. Inspired by [40], we construct a better planar prior model to assist multi-view matching cost evaluation. After executing the basic MVS, a hypothesis with its final cost for each pixel can be obtained. A lower cost represents a more precise estimate, therefore we collect the credible correspondences into a set " + }, + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "type": "inline_equation", + "content": "I_{\\text{cred}}" + }, + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 632, + 287, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 632, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 97, + 632, + 287, + 645 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {\\text {c r e d}} = \\left\\{\\boldsymbol {\\theta} _ {x} \\mid c _ {\\text {p h o t o}} \\left(\\boldsymbol {\\theta} _ {x}\\right) < \\tau_ {\\text {c r e d}} \\right\\}, \\tag {5}", + "image_path": "7c355613c9b86df9d9953d91f3608fb9f5206affae6ce6fea1803526838dad0b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\text{red}}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " represents the credible threshold. Then, the Delaunay triangulation is applied to generate triangular surfaces of various sizes. Although these triangles cover a large area, there are still some marginal areas in the image that remain unaffected. This suggests that the local information" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 386, + 123 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 386, + 123 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 386, + 123 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 386, + 123 + ], + "type": "image", + "image_path": "97591c5205d532d941d8e59b1a360e296e2ac032cd52f8f7d3e34474504bdc10.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 387, + 70, + 465, + 123 + ], + "blocks": [ + { + "bbox": [ + 387, + 70, + 465, + 123 + ], + "lines": [ + { + "bbox": [ + 387, + 70, + 465, + 123 + ], + "spans": [ + { + "bbox": [ + 387, + 70, + 465, + 123 + ], + "type": "image", + "image_path": "5bf72a045e5878075f616bdf268b6b454a70c23998a9b35af8f7ce102a22b8bf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 466, + 70, + 545, + 123 + ], + "blocks": [ + { + "bbox": [ + 466, + 70, + 545, + 123 + ], + "lines": [ + { + "bbox": [ + 466, + 70, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 466, + 70, + 545, + 123 + ], + "type": "image", + "image_path": "d08ee48d1d5c634b66e5167c9e9a435548f96f511d504e37ac86107f3af95751.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 139, + 386, + 192 + ], + "blocks": [ + { + "bbox": [ + 342, + 129, + 352, + 137 + ], + "lines": [ + { + "bbox": [ + 342, + 129, + 352, + 137 + ], + "spans": [ + { + "bbox": [ + 342, + 129, + 352, + 137 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 139, + 386, + 192 + ], + "lines": [ + { + "bbox": [ + 307, + 139, + 386, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 386, + 192 + ], + "type": "image", + "image_path": "47f40daefacd977002d3e1c9ece2ae137d60bdfe91cccf2fea9f19404ccab315.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 198, + 353, + 207 + ], + "lines": [ + { + "bbox": [ + 342, + 198, + 353, + 207 + ], + "spans": [ + { + "bbox": [ + 342, + 198, + 353, + 207 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 217, + 546, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 546, + 338 + ], + "type": "text", + "content": "Figure 4: Planar prior models at different scales and resulting depth maps. (a) The color image (the low-textured area is in the white box while pixels in the red box are details); (b) the planar prior model constructed at the low-resolution scale; (c) the planar prior model constructed at the high-resolution scale (the dark blue areas show that there are anomalies and have already been filtered out); (d) the depth map obtained without prior models; (e) the depth map based on the prior model constructed at the low-resolution scale; (f) the depth map based on prior models at different scales." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 387, + 139, + 466, + 192 + ], + "blocks": [ + { + "bbox": [ + 421, + 129, + 432, + 137 + ], + "lines": [ + { + "bbox": [ + 421, + 129, + 432, + 137 + ], + "spans": [ + { + "bbox": [ + 421, + 129, + 432, + 137 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 387, + 139, + 466, + 192 + ], + "lines": [ + { + "bbox": [ + 387, + 139, + 466, + 192 + ], + "spans": [ + { + "bbox": [ + 387, + 139, + 466, + 192 + ], + "type": "image", + "image_path": "240c2f46207ca9ec3919e69a5778c1727aeba93dbde47bd7150349b1a64dee78.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 198, + 432, + 207 + ], + "lines": [ + { + "bbox": [ + 421, + 198, + 432, + 207 + ], + "spans": [ + { + "bbox": [ + 421, + 198, + 432, + 207 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 139, + 545, + 192 + ], + "blocks": [ + { + "bbox": [ + 500, + 129, + 511, + 137 + ], + "lines": [ + { + "bbox": [ + 500, + 129, + 511, + 137 + ], + "spans": [ + { + "bbox": [ + 500, + 129, + 511, + 137 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 139, + 545, + 192 + ], + "lines": [ + { + "bbox": [ + 466, + 139, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 466, + 139, + 545, + 192 + ], + "type": "image", + "image_path": "23800d8bb04a57c8a3fbc6fe7648634d0242af18cc6640efd0e5c60f3825bee6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 198, + 511, + 207 + ], + "lines": [ + { + "bbox": [ + 500, + 198, + 511, + 207 + ], + "spans": [ + { + "bbox": [ + 500, + 198, + 511, + 207 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "text", + "content": "cannot satisfy the requirements of developing a planar prior model in these parts of the regions. To address this issue, we propose a method based on KNN to construct the planar prior model via non-local credible hypotheses. First, we set up a KD-tree [47] for the reliable points in " + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "inline_equation", + "content": "I_{\\text{cred}}" + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "text", + "content": ". Second, the top-" + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "text", + "content": " nearest non-local neighbors are searched for pixels in the regions where prior construction is challenging. In addition, we use Heron's formula to avoid the case of co-linearity of proximity points. Then, each pixel will have three reliable neighboring points. We project them into the coordinate of the reference camera and set up a matrix " + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "text", + "content": " to get the plane parameters " + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "inline_equation", + "content": "z_{opt}" + }, + { + "bbox": [ + 305, + 359, + 546, + 503 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 368, + 513, + 545, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 513, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 368, + 513, + 545, + 534 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {\\text {o p t}} = \\underset {\\boldsymbol {z} ^ {*}: | | \\boldsymbol {z} ^ {*} | | = 1} {\\arg \\min } \\left| \\left| \\boldsymbol {A} \\cdot \\boldsymbol {z} ^ {*} \\right| \\right|, \\tag {6}", + "image_path": "ab244dfe82232e959c7949d69e8fc1ebb6424fb83da513a0b6a9d8720f87a1fa.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 543, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 543, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 305, + 543, + 545, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 543, + 545, + 615 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 305, + 543, + 545, + 615 + ], + "type": "text", + "content": " is a unit-length variable. To further enhance the validity of the prior model, the ultimate prior knowledge for each image is filtered by its depth range. At last, we consider Planar Prior Assistance [40] which can better leverage the planar prior model to assist multi-view matching cost evaluation in low-textured areas." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 623, + 452, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 452, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 452, + 635 + ], + "type": "text", + "content": "3.4. Hierarchical Prior Mining" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 714 + ], + "type": "text", + "content": "The multi-scale structure is employed in some Patch-Match MVS methods to infer the depth information of low-textured areas. Previous approaches [21, 38, 33] usually construct image pyramids to fuse depth estimation at different scales. In addition, several methods [40, 24] focus on constructing planar priors models at a single scale to help" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3615" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "the depth estimation. In our work, we demonstrate that constructing planar prior models in a multi-scale way will further facilitate depth estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 239 + ], + "type": "text", + "content": "As depicted in Fig. 4(b), when the planar prior model is built in low-textured regions, the performance at the low-resolution scale is obviously more advanced than the high-resolution scale. Because the information is condensed after downsampling, resulting in a larger receptive field for each pixel, and a wider range of non-local information can be excavated than the original scale. However, only using the prior from the coarsest scale may lead ambiguity in detail areas. We find that the aforementioned problem of prior construction is expertly managed at the high-resolution scale (Fig. 4(c))." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 323 + ], + "type": "text", + "content": "Therefore, the depth perception advantages of planar prior models at different scales are complementary. Inspired by this key observation, we design an HPM framework to construct prior models at different scales, which assists the depth estimation of low-textured regions without sacrificing details. The depth map can be continually optimized via the HPM framework (Fig. 4(d)-(f))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 526 + ], + "type": "text", + "content": "Excavating hierarchical prior is essentially an iterative coarse-to-fine optimization process (Fig. 2). After employing the basic MVS method with NESP, we first downsample the obtained hypotheses (i.e. depth and normal) to the coarsest scale and build a planar prior model. Second, with the joint bilateral upsampler [18], we propagate the planar prior to the original scale. Third, the planar prior model is embedded into the PatchMatch MVS method with Planar Prior Assistance [36] to approximate the depth in low-textured areas better. Finally, we downsample the newly generated hypotheses to the medium scale again, and repeat the same operation until iterating up to the original scale. For images with high resolutions, more medium scales can be considered. To further optimize the results, we apply geometric consistency as [25, 38] do. This way, the final depth estimation can achieve a better trade-off between details and low-textured areas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 527, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 635 + ], + "type": "text", + "content": "Moreover, to save computing resources and improve time efficiency, we propose a fast version of the HPM framework. This version only uses the planar prior model constructed at the coarsest scale to assist MVS. The rationale behind this is that the prior model obtained at the coarsest scale can effectively target the low-textured region and exactly compensate for the deficiencies of the basic MVS method in that region. Both HPM and its fast version will be experimentally analyzed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 646, + 128, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 128, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 128, + 659 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We evaluate our method on two challenging MVS datasets, ETH3D benchmark [26] and Tanks & Temples datasets [17], from two perspectives, i.e., benchmark performance and analysis experiments." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 127, + 545, + 352 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "type": "text", + "content": "Table 1: Point cloud evaluation on ETH3D benchmark [26]. We show accuracy / completeness / " + }, + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 305, + 71, + 545, + 118 + ], + "type": "text", + "content": " score (in %) at different thresholds (2cm and 10cm). The best results are marked in bold while the second-best results are underlined." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 127, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 309, + 127, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 309, + 127, + 545, + 352 + ], + "type": "table", + "html": "
Method2cm10cm
Train.i) Traditional
COLMAP [25]91.85 / 55.13 / 67.6698.75 / 79.47 / 87.61
PCF-MVS [19]84.11 / 75.73 / 79.4295.98 / 90.42 / 92.98
ACMM [38]90.67 / 70.42 / 78.8698.12 / 86.40 / 91.70
ACMP [40]90.12 / 72.15 / 79.7997.97 / 87.15 / 92.03
ACMMP [36]91.03 / 77.27 / 83.4297.96 / 93.19 / 95.46
ii) Learning
PatchMatchNet [30]64.81 / 65.43 / 64.2189.98 / 83.28 / 85.70
IterMVS [29]73.62 / 61.87 / 66.3694.48 / 78.85 / 85.25
MVSTER [32]68.08 / 76.92 / 72.0691.97 / 91.91 / 91.73
HPM-MVSfast91.17 / 73.20 / 80.8698.23 / 92.41 / 94.97
HPM-MVS90.66 / 79.50 / 84.5897.97 / 95.59 / 96.22
Testi) Traditional
COLMAP [25]91.97 / 62.98 / 73.0198.25 / 84.54 / 90.40
PCF-MVS [19]82.15 / 79.29 / 80.3892.12 / 91.26 / 91.56
ACMM [38]90.65 / 74.34 / 80.7898.05 / 88.77 / 92.96
ACMP [40]90.45 / 75.58 / 81.5197.47 / 88.71 / 92.62
ACMMP [36]91.91 / 82.10 / 85.8998.05 / 94.67 / 96.27
ii) Learning
PatchMatchNet [30]79.71 / 77.46 / 73.1291.98 / 92.05 / 91.91
IterMVS [29]76.91 / 72.65 / 74.2995.42 / 85.81 / 90.15
MVSTER [32]77.09 / 82.47 / 79.0194.21 / 92.71 / 92.30
HPM-MVSfast92.50 / 80.25 / 85.3598.32 / 94.89 / 96.51
HPM-MVS92.13 / 83.25 / 87.1198.11 / 95.41 / 96.69
", + "image_path": "ad3fb7f3152340efeab16fac4d190b8b81261c3dbe7a503146c0adbe5308aebe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 371, + 422, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 422, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 422, + 384 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "content": "ETH3D high-resolution multi-view benchmark [26] consists of training datasets and test datasets, which contain indoor and outdoor images at a resolution of " + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "inline_equation", + "content": "6048 \\times 4032^1" + }, + { + "bbox": [ + 304, + 389, + 545, + 485 + ], + "type": "text", + "content": ". Tanks & Temples datasets are divided into Intermediate and Advanced datasets based on the complexity of the scene. However, these datasets do not provide official camera poses. In that case, we use COLMAP [25] to estimate camera parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 486, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 546 + ], + "type": "text", + "content": "All of our experiments are executed on a computer with an Intel i5-12600 CPU and an RTX 3070 GPU. In the basic MVS with NESP, " + }, + { + "bbox": [ + 305, + 486, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\{k, N_{ext}, \\tau_{good}, \\tau_{bad}, \\alpha, n_{good}, n_{bad}, R\\} = \\{4, 3, 0.8, 1.2, 90, 1, 2, 4\\}" + }, + { + "bbox": [ + 305, + 486, + 545, + 546 + ], + "type": "text", + "content": ". In the process of planar prior construction, " + }, + { + "bbox": [ + 305, + 486, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\{\\tau_{cred}, K\\} = \\{0.1, 6\\}" + }, + { + "bbox": [ + 305, + 486, + 545, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 555, + 446, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 446, + 566 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 446, + 566 + ], + "type": "text", + "content": "4.2. Benchmark Performance" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "content": "Evaluation on ETH3D Benchmark. Both traditional methods [25, 19, 38, 40, 36] and learning-based [30, 29, 32] methods are considered for comparison. Table 1 reports the accuracy, completeness and " + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 573, + 545, + 621 + ], + "type": "text", + "content": " score of the point clouds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 622, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 545, + 693 + ], + "type": "text", + "content": "As shown in the table, HPM-MVS can achieve outstanding performance in terms of completeness and " + }, + { + "bbox": [ + 304, + 622, + 545, + 693 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 622, + 545, + 693 + ], + "type": "text", + "content": " score among all the methods, and HPM-MVS fast can outperform in accuracy while guaranteeing completeness and " + }, + { + "bbox": [ + 304, + 622, + 545, + 693 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 622, + 545, + 693 + ], + "type": "text", + "content": " score. Notably, HPM-MVS obtains the highest completeness than other methods on both training and test datasets which con" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 539, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 539, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 539, + 713 + ], + "type": "text", + "content": "1We downsample the undistorted images to " + }, + { + "bbox": [ + 317, + 702, + 539, + 713 + ], + "type": "inline_equation", + "content": "3200 \\times 2130" + }, + { + "bbox": [ + 317, + 702, + 539, + 713 + ], + "type": "text", + "content": " as in [25]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3616" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 133, + 163 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 133, + 163 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 133, + 163 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 133, + 163 + ], + "type": "image", + "image_path": "05e14e7e6da81555aca19be8d88dc6879a00bf56e246349bf0a70e056eb423f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 167, + 113, + 178 + ], + "lines": [ + { + "bbox": [ + 72, + 167, + 113, + 178 + ], + "spans": [ + { + "bbox": [ + 72, + 167, + 113, + 178 + ], + "type": "text", + "content": "(a) Images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 134, + 70, + 215, + 163 + ], + "blocks": [ + { + "bbox": [ + 134, + 70, + 215, + 163 + ], + "lines": [ + { + "bbox": [ + 134, + 70, + 215, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 215, + 163 + ], + "type": "image", + "image_path": "827904327740f1d3cb869735f7653944c122222ce6f93bcc841dce356c195e06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 167, + 209, + 178 + ], + "lines": [ + { + "bbox": [ + 140, + 167, + 209, + 178 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 209, + 178 + ], + "type": "text", + "content": "(b) COLMAP [25]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 70, + 296, + 163 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 296, + 163 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 296, + 163 + ], + "type": "image", + "image_path": "7f284a1f851b7c1e1eaf3162640edb1693b018d84557c13888a81bb67324a3bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 167, + 288, + 178 + ], + "lines": [ + { + "bbox": [ + 224, + 167, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 224, + 167, + 288, + 178 + ], + "type": "text", + "content": "(c)ACMMP [36]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 187, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 547, + 212 + ], + "type": "text", + "content": "Figure 5: Qualitative point cloud comparisons between different MVS methods on several high-resolution multi-view test scans (old computer and terrace 2) of ETH3D benchmark [26]. Some challenging areas are shown in red boxes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 378, + 163 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 378, + 163 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 378, + 163 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 378, + 163 + ], + "type": "image", + "image_path": "536571a31e4ed0743ce38e864d00358c58a2e09cc7986685400e3a7adf340ac4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 167, + 369, + 178 + ], + "lines": [ + { + "bbox": [ + 306, + 167, + 369, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 369, + 178 + ], + "type": "text", + "content": "(d) IterMVS [29]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 460, + 163 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 460, + 163 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 460, + 163 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 460, + 163 + ], + "type": "image", + "image_path": "1141416b38400e85332de8c61528263f9dfb5e682023c35a0cf27e1cdd9870b7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 167, + 453, + 178 + ], + "lines": [ + { + "bbox": [ + 386, + 167, + 453, + 178 + ], + "spans": [ + { + "bbox": [ + 386, + 167, + 453, + 178 + ], + "type": "text", + "content": "(e) HPM-MVSfast" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 461, + 70, + 542, + 163 + ], + "blocks": [ + { + "bbox": [ + 461, + 70, + 542, + 163 + ], + "lines": [ + { + "bbox": [ + 461, + 70, + 542, + 163 + ], + "spans": [ + { + "bbox": [ + 461, + 70, + 542, + 163 + ], + "type": "image", + "image_path": "e6b69daf01831a12ea420011182e78fc6801a427d9e035ee0b49d4ef58370f79.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 167, + 528, + 177 + ], + "lines": [ + { + "bbox": [ + 474, + 167, + 528, + 177 + ], + "spans": [ + { + "bbox": [ + 474, + 167, + 528, + 177 + ], + "type": "text", + "content": "(f) HPM-MVS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 55, + 276, + 279, + 430 + ], + "blocks": [ + { + "bbox": [ + 46, + 231, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 266 + ], + "type": "text", + "content": "Table 2: " + }, + { + "bbox": [ + 46, + 231, + 288, + 266 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 46, + 231, + 288, + 266 + ], + "type": "text", + "content": " score (in %) comparisons of different methods on Tanks & Temples datasets [17] at given evaluation threshold." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 276, + 279, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 279, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 279, + 430 + ], + "type": "table", + "html": "
MethodIntermediateAdvanced
TraditionalCOLMAP [25]42.1427.24
PCF-MVS [19]55.8834.59
ACMM [38]57.2734.02
ACMP [40]58.4137.44
ACMMP [36]59.3837.84
LearningR-MVSNet [44]48.4024.91
PatchMatch-RL [20]51.8131.78
PatchMatchNet [30]53.1532.31
IterMVS [29]56.2233.24
PVSNet [37]56.8833.46
Effi-MVS [31]56.8834.39
MVSTER [32]60.9237.53
OursHPM-MVSfast61.5939.65
HPM-MVS61.3940.80
", + "image_path": "d8e14d7c419b577249cf7627fd835531f5267c2e1c5454de8702628dbfb3e33e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 521 + ], + "type": "text", + "content": "tain large low-textured areas, because HPM-MVS perceives non-local structured information for recovering geometry. The visual comparisons are shown in Fig. 5. It can be clearly observed that our methods produce higher quality point clouds than other competitors, especially in challenging low-textured areas, e.g., red boxes in Fig. 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 582 + ], + "type": "text", + "content": "Evaluation on Tanks & Temples Datasets. To further illustrate the robustness of our methods, we test our methods on Tanks & Temples datasets without any fine-tuning. The quantitative results of state-of-the-art methods on both Intermediate and Advanced sets are reported in Table 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "Our methods achieve outstanding performance among all the methods. Even compared with learning-based methods, our methods still achieve better performance without any data training. Remarkably, HPM-MVS can greatly improve the performance of Advanced datasets which contain complex details and more low-textured areas difficult to recover. Our method ranks " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{1}^{st}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " in Advanced datasets over all the existing traditional MVS methods, and it outperforms the previous best record by " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "2.96\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ". This result well confirms that our method can achieve a good balance for MVS reconstruction in details and low-textured areas. Fig. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 232, + 546, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 546, + 318 + ], + "type": "text", + "content": "presents some visualization examples of point cloud recall comparisons between different methods, which verifies our proposed methods can produce more complete 3D geometries than the others. These obvious advantages suggest that HPM-MVS and HPM-MVSfast hold powerful generalization ability and can adapt to different application scenarios without any data training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 326, + 430, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 326, + 430, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 326, + 430, + 339 + ], + "type": "text", + "content": "4.3. Analysis Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 344, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 546, + 464 + ], + "type": "text", + "content": "Ablation Study. In this section, we conduct an ablation study on the multi-view high-resolution training datasets of ETH3D benchmark to validate the performance of different parts in our proposed methods. We divide them into five sections, which are Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework. The baseline method is " + }, + { + "bbox": [ + 304, + 344, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\mathrm{ACMH}^2" + }, + { + "bbox": [ + 304, + 344, + 546, + 464 + ], + "type": "text", + "content": " [38]. Table 3 summarizes the effectiveness of each individual part in our proposed methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 465, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 585 + ], + "type": "text", + "content": "For the basic MVS method with NESP, we remove the no-local part and the extensible part, respectively. The results show that these two proposals can both enhance the completeness and " + }, + { + "bbox": [ + 304, + 465, + 545, + 585 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 465, + 545, + 585 + ], + "type": "text", + "content": " score. In particular, the improvement of the non-local strategy is more obvious than the extensible strategy. In some circumstances, especially in low-textured areas, several severe errors always cause the hypothesis propagation into local optimal solutions. Hence, the non-local operation can effectively prevent this situation from getting worse." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "text", + "content": "Based on NESP, the planar prior model constructed at the current scale generates a potential hypothesis for each pixel in the image. As for two versions of the HPM framework, one can make several observations from the results: 1) HPM and " + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathrm{HPM}_{\\mathrm{fast}}" + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "text", + "content": " achieve outstanding performance at relatively high thresholds (e.g., " + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "inline_equation", + "content": "5cm" + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "inline_equation", + "content": "10cm" + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "text", + "content": "), because they have a strong focus on sizable low-textured areas. 2) We notice that " + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathrm{HPM}_{\\mathrm{fast}}" + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "text", + "content": " will suffer some loss of completeness at low thresholds (e.g., " + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "inline_equation", + "content": "1cm" + }, + { + "bbox": [ + 304, + 585, + 545, + 693 + ], + "type": "text", + "content": "), because it obtains the prior model" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 528, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 528, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 528, + 712 + ], + "type": "text", + "content": "2Note that, we combine ACMH with geometric consistency here." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3617" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 128, + 163 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 128, + 163 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 128, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 128, + 163 + ], + "type": "image", + "image_path": "b44c67490e7b44509f7e927195a22f77ad6cc1581a28f3692eee820db27d3e86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 167, + 119, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 167, + 119, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 167, + 119, + 178 + ], + "type": "text", + "content": "(a) ACMM [38]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 76, + 208, + 163 + ], + "blocks": [ + { + "bbox": [ + 129, + 76, + 208, + 163 + ], + "lines": [ + { + "bbox": [ + 129, + 76, + 208, + 163 + ], + "spans": [ + { + "bbox": [ + 129, + 76, + 208, + 163 + ], + "type": "image", + "image_path": "cc2707246729b29fe26ccd694e56534f185981625b44a2ac759ed1f643ecfc8a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 167, + 202, + 178 + ], + "lines": [ + { + "bbox": [ + 137, + 167, + 202, + 178 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 202, + 178 + ], + "type": "text", + "content": "(b)ACMMP [36]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 75, + 290, + 163 + ], + "blocks": [ + { + "bbox": [ + 209, + 75, + 290, + 163 + ], + "lines": [ + { + "bbox": [ + 209, + 75, + 290, + 163 + ], + "spans": [ + { + "bbox": [ + 209, + 75, + 290, + 163 + ], + "type": "image", + "image_path": "d4c729c1eb5b77f2923a7857434e9b78f985dcba6f4b9d673eb1bb830da6053d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 167, + 282, + 178 + ], + "lines": [ + { + "bbox": [ + 219, + 167, + 282, + 178 + ], + "spans": [ + { + "bbox": [ + 219, + 167, + 282, + 178 + ], + "type": "text", + "content": "(c) IterMVS [29]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 188, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 213 + ], + "type": "text", + "content": "Figure 6: Qualitative point cloud recall comparisons between MVS methods on Tanks & Temples datasets [17] (Horse and Temple). The darker the pixel color is, the greater the error will be. " + }, + { + "bbox": [ + 46, + 188, + 547, + 213 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 188, + 547, + 213 + ], + "type": "text", + "content": " is the recommended threshold for each dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 290, + 75, + 370, + 163 + ], + "blocks": [ + { + "bbox": [ + 290, + 75, + 370, + 163 + ], + "lines": [ + { + "bbox": [ + 290, + 75, + 370, + 163 + ], + "spans": [ + { + "bbox": [ + 290, + 75, + 370, + 163 + ], + "type": "image", + "image_path": "c447d4aac95d29e091c5b3b45699a21fb33bbc9c56471fefffc27d20bc3beb49.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 167, + 365, + 178 + ], + "lines": [ + { + "bbox": [ + 297, + 167, + 365, + 178 + ], + "spans": [ + { + "bbox": [ + 297, + 167, + 365, + 178 + ], + "type": "text", + "content": "(d) MVSTER [32]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 370, + 75, + 451, + 163 + ], + "blocks": [ + { + "bbox": [ + 370, + 75, + 451, + 163 + ], + "lines": [ + { + "bbox": [ + 370, + 75, + 451, + 163 + ], + "spans": [ + { + "bbox": [ + 370, + 75, + 451, + 163 + ], + "type": "image", + "image_path": "a90f1108c6e7b1cb5ba66e589188dfb2834ac9d1c1d3e3f54af2dfe429a08e3a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 167, + 445, + 178 + ], + "lines": [ + { + "bbox": [ + 378, + 167, + 445, + 178 + ], + "spans": [ + { + "bbox": [ + 378, + 167, + 445, + 178 + ], + "type": "text", + "content": "(e) HPM-MVSfast" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 452, + 72, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 452, + 72, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 452, + 72, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 452, + 72, + 547, + 163 + ], + "type": "image", + "image_path": "77bfd1b20116410f1aa48c24e8dc813953d4115d6ca12a78d1d86b8b835d368a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 167, + 520, + 178 + ], + "lines": [ + { + "bbox": [ + 465, + 167, + 520, + 178 + ], + "spans": [ + { + "bbox": [ + 465, + 167, + 520, + 178 + ], + "type": "text", + "content": "(f) HPM-MVS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 50, + 269, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 46, + 224, + 546, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 546, + 261 + ], + "type": "text", + "content": "Table 3: Comparative results of our methods under different settings on the training datasets of ETH3D benchmark [26]. N, E, PA, " + }, + { + "bbox": [ + 46, + 224, + 546, + 261 + ], + "type": "inline_equation", + "content": "\\mathrm{HPM}_{\\mathrm{fast}}" + }, + { + "bbox": [ + 46, + 224, + 546, + 261 + ], + "type": "text", + "content": " and HPM mean Non-local Sampling Pattern, Extensible Sampling Pattern, planar prior model assistance, Hierarchical Prior Mining framework fast version and Hierarchical Prior Mining framework, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 269, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 50, + 269, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 50, + 269, + 545, + 383 + ], + "type": "table", + "html": "
MethodSettingsETH3D train.
NEPAHPMfastHPM1cm2cm5cm10cm
ACMH [38]82.41 / 50.63 / 61.7191.09 / 63.04 / 73.6096.61 / 74.73 / 83.5098.25 / 82.20 / 88.57
NSP✓82.52 / 53.46 / 63.8490.86 / 65.59 / 75.2696.63 / 76.15 / 84.5298.46 / 82.40 / 89.19
ESP✓82.66 / 51.40 / 62.2290.93 / 63.83 / 74.0596.64 / 75.37 / 84.0498.47 / 82.09 / 89.03
NESP✓✓82.47 / 53.98 / 64.2290.78 / 66.02 / 75.5996.60 / 76.63 / 84.8798.47 / 82.73 / 89.44
NESP+PA✓✓✓82.88 / 62.37 / 71.3290.51 / 74.96 / 81.6696.48 / 84.29 / 88.9097.89 / 88.57 / 92.73
HPMfast✓✓83.32 / 53.53 / 64.6691.16 / 70.48 / 79.0496.39 / 84.67 / 89.8497.89 / 88.57 / 92.73
HPM✓✓83.01 / 62.07 / 70.6790.65 / 77.03 / 83.0696.06 / 88.42 / 91.9797.91 / 93.62 / 95.66
HPM-MVSfast✓✓✓✓82.95 / 57.36 / 67.3991.17 / 73.20 / 80.8696.51 / 86.22 / 90.8998.23 / 92.14 / 94.97
HPM-MVS✓✓✓✓82.93 / 65.16 / 72.7390.66 / 79.50 / 84.5896.13 / 89.98 / 92.8897.97 / 94.59 / 96.22
", + "image_path": "5d2e12fe0ca338398df04b596307a02f3c1aab9d25c405849cf5fe5a8c5728bd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 62, + 434, + 272, + 548 + ], + "blocks": [ + { + "bbox": [ + 46, + 400, + 287, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 426 + ], + "type": "text", + "content": "Table 4: Generalization performance of NESP on ETH3D high-resolution training datasets [26]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 434, + 272, + 548 + ], + "lines": [ + { + "bbox": [ + 62, + 434, + 272, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 272, + 548 + ], + "type": "table", + "html": "
MethodETH3D train.
1cm2cm5cm10cm20cm
ACMM [38]67.5878.8687.6891.7094.41
ACMP [40]68.7279.7988.3292.0394.43
ACMMP [36]71.5783.4292.0395.4697.37
ACMM+NESP70.7081.0188.8092.2694.55
3.12↑2.15↑1.12↑0.56↑0.14↑
ACMP+NESP70.8781.4589.4392.7294.78
2.15↑1.66↑1.11↑0.69↑0.35↑
ACMMP+NESP74.5485.3393.2596.4597.99
2.97↑1.91↑1.22↑0.99↑0.62↑
", + "image_path": "356f8e53a9e737db617e0740426760f135ba7190d1d36360cbfaf271df5a2528.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 616 + ], + "type": "text", + "content": "at the lowest resolution scale which may lead to ambiguities in details. 3) The method with NESP as the basis for subsequent processes can provide a set of reliable points for planar prior model construction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Generalization Performance of NESP. In addition, we perform a more extensive comparison on the multi-view high-resolution training datasets of ETH3D benchmark to show the generalization ability of NESP. Several kinds of state-of-the-art diffusion-like propagation methods are integrated with NESP for evaluation. The considered methods are ACMM [38], ACMP [40] and ACMMP [36]. Note that ACMMC is a very recent state-of-the-art geometric-only" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 402, + 544, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 544, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 544, + 426 + ], + "type": "text", + "content": "method. Each method is tested under different thresholds, and results are shown in Table 4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "text", + "content": "Impressively, NESP dramatically improves the " + }, + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "text", + "content": " score for all tested methods at relatively low thresholds, e.g., 1cm and 2cm. The improvement of NESP working with multi-scale strategy is more obvious, because the non-local and extensible strategy works even better at low-resolution scale. Notably, the combination of NESP and ACMMP outperforms all the competitors and achieves state-of-the-art " + }, + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "text", + "content": " score of " + }, + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "inline_equation", + "content": "85.33\\% / 96.45\\%" + }, + { + "bbox": [ + 304, + 427, + 546, + 582 + ], + "type": "text", + "content": " at 2cm/10cm thresholds. Based on the results, the following conclusions can be drawn: 1) NESP is a generalized module that can be adapted to other MVS methods and boost their performance. 2) Combined with the multi-scale strategy, NESP can achieve better performance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 583, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 583, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 305, + 583, + 545, + 608 + ], + "type": "text", + "content": "More analysis experiments and point cloud visualizations can be found in the supplementary." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 620, + 378, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 378, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 378, + 633 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "In this paper, we presented an HPM-MVS method which can efficiently perceive non-local structured information to achieve high-quality reconstruction. Based on the non-local and extensible strategy, we first propose our NESP module. Focusing on the planar prior construction in marginal regions, we employ KNN to search non-local credible points" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3618" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "and obtain potential hypotheses. Further, we design an HPM framework which explores planar prior at different scales to assist MVS. Through extensive experiments, we have demonstrated the rationality of each key component of our method and its superiority against existing methods. In the future, we plan to combine our method with a probabilistic graphical model to further improve its ability to handle the reconstruction of details and low-textured regions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 173, + 289, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 289, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 289, + 222 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (No. 62002295), and NWPU international cooperation and exchange promotion projects (No. 02100-23GH0501)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 232, + 107, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 107, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 107, + 245 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 251, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 251, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 297 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 297, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 342 + ], + "type": "text", + "content": "[2] Christian Bailer, Manuel Finckh, and Hendrik Lensch. Scale robust multi view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 398-411. Springer, 2012. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 343, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 343, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 343, + 288, + 387 + ], + "type": "text", + "content": "[3] Caroline Baillard and Andrew Zisserman. A plane-sweep strategy for the 3d reconstruction of buildings from multiple images. International Archives of Photogrammetry and Remote Sensing, 33(B2; PART 2):56-62, 2000. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 431 + ], + "type": "text", + "content": "[4] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Transactions on Graphics, 28(3):24, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 433, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 288, + 487 + ], + "type": "text", + "content": "[5] Fabio Bruno, Stefano Bruno, Giovanna De Sensi, Maria-Laura Luchi, Stefania Mancuso, and Maurizio Muzzupappa. From 3d reconstruction to virtual reality: A complete methodology for digital archaeological exhibition. Journal of Cultural Heritage, 11(1):42-49, 2010. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 288, + 544 + ], + "type": "text", + "content": "[6] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 545, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 588 + ], + "type": "text", + "content": "[7] Robert T Collins. A space-sweep approach to true multi-image matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 358-363. IEEE, 1996. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 590, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 288, + 633 + ], + "type": "text", + "content": "[8] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2009. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 679 + ], + "type": "text", + "content": "[9] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[10] David Gallup, Jan-Michael Frahm, Philippos Mordohai, Qingxiong Yang, and Marc Pollefeys. Real-time plane-sweeping stereo with multiple sweeping directions. In Pro-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2007. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[11] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1418-1425. IEEE, 2010. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 205 + ], + "type": "text", + "content": "[12] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "text", + "content": "[13] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[14] Mohsen Hejrati and Deva Ramanan. Analysis by synthesis: 3d object recognition by object reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2449-2456, 2014. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "text", + "content": "[16] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "type": "text", + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 1, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[18] Johannes Kopf, Michael F Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Transactions on Graphics, 26(3):96-es, 2007. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[19] Andreas Kuhn, Shan Lin, and Oliver Erdler. Plane completion and filtering for multi-view stereo reconstruction. In Proceedings of the German Conference on Pattern Recognition, pages 18-32. Springer, 2019. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[20] Jae Yong Lee, Joseph DeGol, Chuhang Zou, and Derek Hoiem. Patchmatch-rl: Deep mvs with pixelwise depth, normal, and visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6158–6167, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[21] Jie Liao, Yanping Fu, Qingan Yan, and Chunxia Xiao. Pyramid multi-view stereo with local consistency. In Computer Graphics Forum, volume 38, pages 335-346. Wiley Online Library, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[22] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8665-8675, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[23] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In Proceedings of the IEEE International Conference on Computer Vision, pages 6851–6860, 2019. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3619" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[24] Andrea Romanoni and Matteo Matteucci. Tapa-mvs: Textureless-aware patchmatch multi-view stereo. In Proceedings of the IEEE International Conference on Computer Vision, pages 10413-10422, 2019. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "text", + "content": "[25] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 238 + ], + "type": "text", + "content": "[26] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 1, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "text", + "content": "[27] Qi Shan, Riley Adams, Brian Curless, Yasutaka Furukawa, and Steven M Seitz. The visual turing test for scene reconstruction. In International Conference on 3D Vision-3DV, pages 25–32. IEEE, 2013. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 284, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 288, + 316 + ], + "type": "text", + "content": "[28] Shuhan Shen. Accurate multiple view 3d reconstruction using patch-based stereo for large-scale scenes. IEEE Transactions on Image Processing, 22(5):1901-1914, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 370 + ], + "type": "text", + "content": "[29] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8606-8615, 2022. 1, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 288, + 426 + ], + "type": "text", + "content": "[30] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14194-14203, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[31] Shaoqian Wang, Bo Li, and Yuchao Dai. Efficient multiview stereo by iterative dynamic cost volume. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8655-8664, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[32] Xiaofeng Wang, Zheng Zhu, Guan Huang, Fangbo Qin, Yun Ye, Yijia He, Xu Chi, and Xingang Wang. Mvster: Epipolar transformer for efficient multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 573-591. Springer, 2022. 1, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 580 + ], + "type": "text", + "content": "[33] Yuesong Wang, Tao Guan, Zhuo Chen, Yawei Luo, Keyang Luo, and Lili Ju. Mesh-guided multi-view stereo with pyramid architecture. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2039-2048, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 623 + ], + "type": "text", + "content": "[34] Jian Wei, Benjamin Resch, and Hendrik PA Lensch. Multiview depth map estimation with cross-view consistency. In Proceedings of the British Machine Vision Conference, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "type": "text", + "content": "[35] Oliver Woodford, Philip Torr, Ian Reid, and Andrew Fitzgibbon. Global stereo reconstruction under second-order smoothness priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(12):2115-2128, 2009. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[36] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 609 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[37] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 1, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[38] Qingshan Xu and Wenbing Tao. Multi-scale geometric consistency guided multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5483-5492, 2019. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[39] Qingshan Xu and Wenbing Tao. Learning inverse depth regression for multi-view stereo with correlation cost volume. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12508-12515, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[40] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 263, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 319 + ], + "type": "text", + "content": "[41] Zhenyu Xu, Yiguang Liu, Xuelei Shi, Ying Wang, and Yunan Zheng. Marmvs: Matching ambiguity reduced multiple view stereo for efficient large scale scene reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5981-5990, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 363 + ], + "type": "text", + "content": "[42] Jiayu Yang, Wei Mao, Jose M Alvarez, and Miaomiao Liu. Cost volume pyramid based depth inference for multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4877-4886, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[43] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the IEEE European Conference on Computer Vision, pages 767-783, 2018. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 410, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 463 + ], + "type": "text", + "content": "[44] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvvsnet for high-resolution multiview stereo depth inference. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 1, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 520 + ], + "type": "text", + "content": "[45] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A largescale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 521, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 574 + ], + "type": "text", + "content": "[46] Enliang Zheng, Enrique Dunn, Vladimir Jojic, and Jan-Michael Frahm. Patchmatch based joint view selection and depthmap estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1517, 2014. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "type": "text", + "content": "[47] Kun Zhou, Qiming Hou, Rui Wang, and Baining Guo. Real-time kd-tree construction on graphics hardware. ACM Transactions on Graphics, 27(5):1-11, 2008. 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3620" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_content_list.json b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac072cc84e2e271ab1326822e83af4d633e01423 --- /dev/null +++ b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_content_list.json @@ -0,0 +1,1502 @@ +[ + { + "type": "text", + "text": "Hierarchical Spatio-Temporal Representation Learning for Gait Recognition", + "text_level": 1, + "bbox": [ + 99, + 161, + 870, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Wang $^{1,2}$ , Bo Liu $^{1,2*}$ , Fangfang Liang $^{1,2}$ , Bincheng Wang $^{1,2}$", + "bbox": [ + 240, + 206, + 724, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 College of Information Science and Technology, Hebei Agricultural University, China \n2 Hebei Key Laboratory of Agricultural Big Data, China", + "bbox": [ + 141, + 224, + 828, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{20212060107, 20212060108}@pgs.hebau.edu.cn, {boliu, liangfangfang}@hebau.edu.cn", + "bbox": [ + 132, + 258, + 833, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gait recognition is a biometric technique that identifies individuals by their unique walking styles, which is suitable for unconstrained environments and has a wide range of applications. While current methods focus on exploiting body part-based representations, they often neglect the hierarchical dependencies between local motion patterns. In this paper, we propose a hierarchical spatio-temporal representation learning (HSTL) framework for extracting gait features from coarse to fine. Our framework starts with a hierarchical clustering analysis to recover multilevel body structures from the whole body to local details. Next, an adaptive region-based motion extractor (ARME) is designed to learn region-independent motion features. The proposed HSTL then stacks multiple ARMEs in a top-down manner, with each ARME corresponding to a specific partition level of the hierarchy. An adaptive spatiotemporal pooling (ASTP) module is used to capture gait features at different levels of detail to perform hierarchical feature mapping. Finally, a frame-level temporal aggregation (FTA) module is employed to reduce redundant information in gait sequences through multi-scale temporal downsampling. Extensive experiments on CASIA-B, OUMVLP, GREW, and Gait3D datasets demonstrate that our method outperforms the state-of-the-art while maintaining a reasonable balance between model accuracy and complexity. Code is available at: https://github.com/gudaochangsheng/HSTL.", + "bbox": [ + 75, + 334, + 473, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 735, + 206, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike other biometric technologies such as fingerprint, iris, and face, human gait can be captured at a distance without subject cooperation [34]. By evaluating individual-specific walking patterns, gait recognition has been applied in a variety of fields, including criminal investigations [31, 29], sports science [17, 6], and smart transportation", + "bbox": [ + 75, + 758, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5c08c72dcfa0c9537bddd60dc16b987cfe8a966154eafa67b2997ed616ebac1.jpg", + "image_caption": [ + "Figure 1: The motivation for our approach. Left: the motion distributions of body parts. The same color indicates the same spatial location across multiple gait sequences in the CASIA-B dataset. Right: an example hierarchy of body-part sequences where $T$ denotes the temporal dimension of the sequence." + ], + "image_footnote": [], + "bbox": [ + 502, + 302, + 894, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[47]. However, the recognition can be challenging due to large variations in viewpoint [28, 18], occlusion [33, 43], and wearing [50, 48].", + "bbox": [ + 497, + 530, + 892, + 572 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these issues, various approaches have been proposed for extracting gait features from silhouette sequences [4, 25, 16, 15, 55], 3D human structures [1, 22, 42, 59, 20], or gait templates [10, 35, 51]. Silhouette-based gait recognition methods have gained increasing attention due to the ease of obtaining silhouettes from raw videos while preserving essential temporal information. The alignment of the input silhouette makes it possible for some methods to extract local body features by horizontally slicing the silhouette image [56] or intermediate-layer features [8, 27]. This partitioning strategy, first introduced in person re-identification (ReID) [38], has been proven to be effective for gait recognition [4, 8, 12, 3].", + "bbox": [ + 497, + 574, + 892, + 754 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the main limitation of the above part-based approaches is that they do not consider the hierarchical nature of local body movements [2]. For instance, within a gait cycle, the feet and lower body have distinct motion characteristics. Therefore, it is important to treat these body regions separately and investigate their part-whole relationships. Our motivation stems from the examination of body-part-specific motion clues. Specifically, each raw gait se", + "bbox": [ + 497, + 755, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 95, + 8, + 169, + 46 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 231, + 7, + 815, + 48 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 95, + 853, + 223, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19639", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quence in the CASIA-B [52] dataset is uniformly divided into eight part sequences along the body axis, so that each division roughly match a particular body part. The distributions of all body parts are shown in the left part of Fig. 1. Observably, some parts, e.g., the head and feet, are easily separated owing to their large changes in walking kinematics. Whereas other parts, such as the thighs and calves, overlap due to the strong motion correlations between them. Further, to identify the relational structure among the part sequences, a hierarchical clustering analysis [7] is performed. The results are shown in the right part of Fig. 1, indicating that the semantic body regions can be captured in the higher clustering levels without precise localization of the body parts.", + "bbox": [ + 75, + 125, + 472, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the above findings, we propose a novel hierarchical spatio-temporal representation learning (HSTL) framework for gait representation. The HSTL framework consists of multiple adaptive region-based motion extractor (ARME) modules, which are stacked to learn hierarchical motion patterns implied in a gait sequence (as shown in Fig.1). In the ARME module, to account for inter-regional differences, non-shared 3D convolutions are used in correspondence with individual body regions. There regions are pre-identified by a hierarchical clustering process performed on fixed horizontal partitions, allowing each body region to cover one or more body parts. Consequently, the deeper the ARME is, the more local features it tends to extract. Moreover, an adaptive spatio-temporal pooling (ASTP) module is proposed, which couples with an ARME module on the corresponding level to obtain hierarchical gait embeddings.", + "bbox": [ + 75, + 319, + 472, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, changes in gait speed or sampling frequency may result in several redundant frames in a gait sequence. Although several temporal fusion strategies have been proposed, they lose spatial information [8, 15] or lack adaptability [27, 25]. To address this issue, we propose a frame-level temporal aggregation strategy (FTA). FTA fuses temporal features at multiple time steps, preserving significant motion information while compressing the sequence length. The main contributions of this paper are summarized as follows.", + "bbox": [ + 75, + 555, + 472, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL takes into account the dependencies of body regions in gait motions, ensuring simplicity and scalability of the architectural design.", + "2. We introduce an adaptive region-based motion extractor (ARME) module to learn region-independent spatio-temporal representation for gait sequences, an adaptive spatio-temporal pooling (ASTP) module to" + ], + "bbox": [ + 87, + 706, + 470, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perform hierarchical feature mapping, and a frame-level temporal aggregation (FTA) strategy to compress a gait sequence by removing redundant frames.", + "bbox": [ + 531, + 125, + 892, + 168 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Extensive experiments on the widely used gait datasets CASIA-B [52], including OUMVLP [39], GREW [60] and Gait3D [59], demonstrate that our method achieves state-of-the-art performance while offering a suitable trade-off between model accuracy and complexity.", + "bbox": [ + 512, + 178, + 894, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 276, + 640, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Gait Recognition", + "text_level": 1, + "bbox": [ + 499, + 300, + 667, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning-based gait recognition methods can be broadly categorized into two categories: model-based and appearance-based. Model-based approaches extract structure and motion information from gait videos with the aid of pose estimation [1, 22, 42, 20, 41, 59]. Although these methods are robust to changes in viewpoint and appearance, they are sensitive to the accuracy of the pose parameters, making them incapable of handling low-resolution data. On the other hand, appearance-based approaches learn the feature representation from raw videos [57, 37, 24], or binary silhouette sequences [10, 35, 51, 4, 12, 8, 13, 27, 16, 14, 3, 5], which offer greater flexibility than model-based approaches. Our proposed method belongs to the family of appearance-based methods and uses silhouette sequences as inputs.", + "bbox": [ + 497, + 321, + 894, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Hierarchical Model", + "text_level": 1, + "bbox": [ + 499, + 539, + 687, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hierarchical feature representation has been successfully applied to a wide range of vision tasks. Here, we provide a brief review of the hierarchical object ReID approaches related to gait recognition.", + "bbox": [ + 497, + 561, + 890, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In person ReID, some approaches [30, 54, 45, 40, 53] hierarchically learn local descriptions and aggregate appearance features at different levels. For example, Matsukawa et al. [30] described an image patch via hierarchical Gaussian distribution. Zhang et al. [53] proposed a framework to learn coarse-grained and fine-grained features according to body structure. To solve the occlusion problem, Tan et al. [40] devised a hierarchical mask generator to learn from both occluded and holistic joint images. In vehicle ReID, some approaches [46, 36, 19] extract features from vehicle images in a hierarchical manner. For instance, Wei et al. [46] proposed an RNN-based module for extracting latent cues from the model level to the vehicle level. Shyam et al. [36] developed an attention-based hierarchical feature extractor. In addition, Li et al. [19] proposed a global structural embedding module for investigating hierarchical relationships between vehicle characteristics by incorporating attribute and state information.", + "bbox": [ + 497, + 617, + 894, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Each part is pooled into a vector for visualization using t-SNE [44].", + "bbox": [ + 95, + 853, + 460, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19640", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/01ddaa05a50ac8566503c24c507354c4f738ba5c94c89ee856cb3c177c889603.jpg", + "image_caption": [ + "Figure 2: The framework of HFSL. It mainly consists of three modules: ARME (adaptive region-based motion extractor), ASTP (adaptive spatio-temporal pooling), and FTA (frame-level temporal aggregation). During pre-processing, a hierarchy of walking is obtained to guide the architectural design of HFSL. The framework uses multiple ARMEs to extract gait features from the entire body to individual regions. The ASTP module performs hierarchical feature mapping for the output of each level of ARME. The FTA module compresses local clips of each gait sequence to reduce the number of redundant frames. $T$ , $H$ and $W$ denote dimensions of the feature maps. © represents the concatenation operation." + ], + "image_footnote": [], + "bbox": [ + 77, + 123, + 890, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For gait recognition, fusing features of multiple granularities can improve performance [56, 8, 27, 25, 16, 3]. In particular, CSTL [15] proposed a temporal modeling network that integrates multi-scale temporal features adaptively. By combining part-level and sequence-level features, GaitPart [8] obtained a part-independent spatio-temporal expression. Additionally, GaitGL [27] considered both full body-based and part-based information to achieve discriminative feature learning. Instead of equally dividing the feature maps, in 3D Local [16], a localization operation was developed to find 3D volumes of body parts in a sequence. However, most existing gait recognition methods do not sufficiently exploit the hierarchical dependencies among body parts during walking. In this paper, the proposed HSTL performs a coarse-to-fine hierarchical strategy that integrates multi-level motion patterns from gait sequences.", + "bbox": [ + 75, + 520, + 470, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Temporal Model", + "text_level": 1, + "bbox": [ + 77, + 747, + 240, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal cues play a crucial role in gait recognition due to the periodic changes in body shape. Previous methods treat a gait sequence as an unordered set, either compressing it into a single gait template during preprocessing [35, 21] or learning order-independent gait representations from silhouette sets [4, 12, 13]. These methods assume that different subjects share similar global gait patterns, making or", + "bbox": [ + 75, + 769, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dering inputs unnecessary for gait assessment. However, ignoring the temporal nature of the gait sequence can result in missing discriminative local motion information. Recently, some approaches have achieved significant performance gains by explicitly modeling temporal information using LSTM [56], 1D convolution [15], and 3D convolution [26, 16]. Nevertheless, these spatio-temporal operators also significantly increase computational costs. Although some methods have been proposed to reduce video length by aggregating local clips [27, 25, 3], they lack adaptability to variations in pace. The main difference between our approach and others [23, 11] is that we employ multi-scale temporal pooling at the frame level while considering variations in motion across body regions, leading to a more adaptable reduction of the gait sequence length.", + "bbox": [ + 497, + 520, + 892, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 499, + 775, + 672, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the detailed description of HSTL, including the adaptive region-based motion extractor (ARME), the adaptive spatio-temporal pooling (ASTP), and the frame-level temporal aggregation (FTA).", + "bbox": [ + 497, + 811, + 892, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19641", + "bbox": [ + 465, + 887, + 502, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Pipeline", + "text_level": 1, + "bbox": [ + 77, + 125, + 272, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of our HSTL is presented in Fig. 2. Given a gait dataset $\\mathcal{D} = \\{S_i\\}_{i=1}^N$ with $N$ gait sequences, where each sequence $S_i \\in \\mathbb{R}^{C \\times T \\times H \\times W}$ is represented as a 4D tensor with $C$ channels, $T$ frames, and $H \\times W$ pixels. During the preprocessing stage, each gait sequence $S_i$ is divided horizontally and uniformly into $k$ part sequences, indexed from 1 to $k$ . Then, a hierarchical clustering algorithm [7] is applied to these part sequences to obtain a generic hierarchy of gait motions, which is denoted as $\\mathcal{P} = \\{\\mathcal{P}^{(l)}\\}_{l=1}^L$ . Here, $L$ is the number of levels in the hierarchy and $\\mathcal{P}^{(l)}$ is the set of partitions at level $l$ . The partitions at level $l$ are defined as $\\mathcal{P}^{(l)} = \\{P_1^{(l)}, P_2^{(l)}, \\ldots, P_{K_l}^{(l)}\\}$ , where $P_j^{(l)}$ is the $j$ -th subset of part indices and $K_l$ is the number of groups at level $l$ . For instance, the top level $\\mathcal{P}^{(1)} = \\{\\{1, 2, \\ldots, k-1, k\\}\\}$ means all the $k$ parts can be considered as a whole for the gait analysis. This hierarchy provides a structured property of the gait motion patterns and can be utilized to guide gait feature extraction. To achieve this, our proposed HSTL employs three modules: ARME for extracting independent multi-granularity motion features, ASTP for generating vectorized gait embeddings, and FTA for reducing redundant information at the frame level. The HSTL stacks these three modules according to the division in $\\mathcal{P}$ , and the main branch of the HSTL for the input sequence $S_{in}$ can be formalized as:", + "bbox": [ + 75, + 146, + 472, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY ^ {M} = \\Gamma^ {(L)} \\circ \\Psi^ {(L - 1)} \\circ \\dots \\circ \\Omega^ {(2)} \\circ \\Psi^ {(2)} \\circ \\Psi^ {(1)} (S _ {i n}), (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 500, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Psi^{(l)},\\Gamma^{(l)}$ , and $\\Omega^{(l)}$ represent the ARME, ASTP and FTA modules at the $l$ -th level of $\\mathcal{P}$ , respectively. Since FTA uses inter-frame compression to reduce redundant information, it is employed only once at the $l_{\\Omega}$ -th level in $\\mathcal{P}$ (e.g., $l_{\\Omega} = 2$ in Eq.(1)) to prevent excessive loss of information.", + "bbox": [ + 75, + 523, + 468, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the hierarchical gait embeddings, the output $Y^{(l)}$ of each $\\Psi^{(l)}$ at levels $l \\in \\{1,2,\\dots ,L - 1\\}$ and the output of $\\Omega^{(l_{\\Omega})}$ at level $l_{\\Omega}$ , denoted as $Y_{\\Omega}^{(l_{\\Omega})}$ , are fed into the corresponding $\\Gamma^{(l)}$ . The resulting outputs from these $L$ auxiliary branches are concatenated with the output of the main branch defined in Eq. (1), forming the final result, denoted as $Y$ , which is given by:", + "bbox": [ + 75, + 593, + 470, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Y = \\left[ Y ^ {M}, \\Gamma^ {(L - 1)} \\left(Y ^ {(L - 1)}\\right), \\dots , \\right. \\\\ \\left. \\Gamma^ {\\left(l _ {\\Omega}\\right)} \\left(Y _ {\\Omega} ^ {\\left(l _ {\\Omega}\\right)}\\right), \\Gamma^ {(2)} \\left(Y ^ {(2)}\\right), \\Gamma^ {(1)} \\left(Y ^ {(1)}\\right) \\right], \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 92, + 699, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[,]$ denotes the concatenation operation.", + "bbox": [ + 75, + 769, + 384, + 783 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, $Y$ undergoes feature mapping through separate fully connected layers. The model is then trained using a combination of triplet loss $\\mathcal{L}_{tri}$ and cross-entropy loss $\\mathcal{L}_{ce}$ , which is a commonly adopted practice in gait recognition [12, 27, 15, 16, 5]. Further details regarding the relevant modules are described in the following subsections.", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adaptive Region-based Motion Extractor", + "text_level": 1, + "bbox": [ + 499, + 125, + 853, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The adaptive region-based motion extractor (ARME) aims to extract independent spatio-temporal patterns that are associated with different human body parts in a gait sequence. Unlike existing methods that uniformly slice gait images or sequences along the height axis [56, 8, 27, 25], ARME considers the inherent hierarchical relationships among different part sequences that are consistent with walking patterns. This allows ARME to effectively capture the unique walking kinematics of each part.", + "bbox": [ + 497, + 147, + 892, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the hierarchical relation $\\mathcal{P}$ introduced in Section 3.1, ARME first divides the input sequence $X$ into $K_{l}$ regions based on the partition of the $l$ -th level $\\mathcal{P}^{(l)}$ , resulting in the set of regions $\\{X_{j}^{(l)}\\}_{j=1}^{K_{l}}$ , where $X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}$ . $H_{j}^{(l)} = \\frac{\\left|P_{j}^{(l)}\\right|}{k} H$ is the height of the $j$ -th region of the $l$ -th level. Then the $l$ -th level of ARME, $\\Psi^{(l)}$ , can be defined as follows:", + "bbox": [ + 497, + 271, + 892, + 384 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {\\Psi} ^ {(l)} = \\Psi^ {(l)} \\left(X ^ {(l)}\\right) = \\left[ f _ {1} \\left(X _ {1} ^ {(l)}\\right), f _ {2} \\left(X _ {2} ^ {(l)}\\right), \\dots , f _ {K _ {l}} \\left(X _ {K _ {l}} ^ {(l)}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 499, + 391, + 895, + 427 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{j}(\\cdot)$ represents the independent 3D convolution operation applied to the $j$ -th region. The output feature map $Y_{\\Psi}^{(l)} \\in \\mathbb{R}^{C(l) \\times T \\times H \\times W}$ of level $l$ has $C^{(l)}$ channels. This module only modifies the number of channels, preserving the spatial and temporal resolutions of the input feature map.", + "bbox": [ + 497, + 428, + 892, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Spatio-Temporal Pooling", + "text_level": 1, + "bbox": [ + 497, + 523, + 803, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is a common procedure in gait recognition to obtain a compact and fixed-length feature representation by performing horizontal and uniform slicing of feature maps and strip-based pooling [4, 9, 26, 27, 15]. However, a nonuniform division of the feature maps is better at capturing the gait motion characteristics. Thus, the adaptive spatiotemporal pooling (ASTP) is devised to construct hierarchical feature mapping (as shown in Fig. 2). Similar to the ARME module described in Section 3.2, the hierarchy $\\mathcal{P}$ enables us to obtain the $j$ -th region of the $l$ -th level, denoted as $X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}$ . The corresponding ASTP, denoted as $\\Gamma^{(l)}$ , can be expressed as follow:", + "bbox": [ + 497, + 546, + 892, + 718 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {\\Gamma , j} ^ {(l)} = \\Gamma_ {j} ^ {(l)} \\left(X _ {j} ^ {(l)}\\right) = \\operatorname {G e M} _ {j} \\circ \\operatorname {F C} \\circ \\operatorname {M a x} \\left(X _ {j} ^ {(l)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 729, + 890, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Max}(\\cdot)$ represents a max pooling operation along the temporal dimension, $\\mathrm{FC}(\\cdot)$ represents a fully connected layer, $\\mathrm{GeM}_j(\\cdot)$ represents a generalized mean pooling (GeM) operation [32] for $j$ -th region, and $\\Gamma_j^{(l)}$ : $\\mathbb{R}^{C\\times T\\times H_j^{(l)}\\times W}\\mapsto \\mathbb{R}^{C\\times 1\\times H_j^{(l)}\\times W}\\mapsto \\mathbb{R}^{C^{(l)}\\times 1\\times H_j^{(l)}\\times W}\\mapsto$ $\\mathbb{R}^{C^{(l)}\\times 1\\times 1\\times 1}$ , in other words, $Y_{\\Gamma ,j}^{(l)}\\in \\mathbb{R}^{C^{(l)}\\times 1\\times 1\\times 1}$ . Therefore, by concatenating outputs of $K_{l}$ regions, we can", + "bbox": [ + 497, + 761, + 892, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19642", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea613ae139d910f745d276d8f510e9148a04742ff49bca542a91df61ab271baf.jpg", + "image_caption": [ + "Figure 3: The detailed structure of the frame-level temporal aggregation (FTA). For simplicity, we omit the channel dimension $C$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 122, + 472, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "obtain $Y_{\\Gamma}^{(l)} = \\left[Y_{\\Gamma,1}^{(l)},Y_{\\Gamma,2}^{(l)},\\dots ,Y_{\\Gamma,K_l}^{(l)}\\right]$ , where $Y_{\\Gamma}^{(l)} \\in \\mathbb{R}^{C^{(l)} \\times 1 \\times K_l \\times 1}$ is the output of ASTP at level $l$ .", + "bbox": [ + 75, + 365, + 468, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Frame-level Temporal Aggregation", + "text_level": 1, + "bbox": [ + 77, + 410, + 381, + 427 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A gait sequence may contain several redundant frames due to factors such as the acquisition frame rate and pace frequency. To reduce computational costs, some methods compress a gait sequence by aggregating its local clips [27, 25]. In the proposed frame-level temporal aggregation (FTA) strategy, we consider both the gait structure and the multiscale temporal information. Given the $j$ -th gait region at the $l$ -th level, $X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}$ , we first fuse the features of the two temporal scales using the following formula:", + "bbox": [ + 75, + 433, + 468, + 572 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {U} _ {j} ^ {(l)} = U _ {j, 1} ^ {(l)} + U _ {j, 2} ^ {(l)} \\\\ = \\operatorname {M a x} _ {3 \\times 1 \\times 1} ^ {3 \\times 1 \\times 1} \\left(X _ {j} ^ {(l)}\\right) + \\operatorname {M a x} _ {5 \\times 1 \\times 1} ^ {3 \\times 1 \\times 1} \\left(X _ {j} ^ {(l)}\\right), \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 579, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Max}_{3\\times 1\\times 1}^{3\\times 1\\times 1}(\\cdot)$ and $\\mathrm{Max}_{5\\times 1\\times 1}^{3\\times 1\\times 1}(\\cdot)$ denote max pooling operations with kernel sizes of $3\\times 1\\times 1$ and $5\\times 1\\times 1$ respectively, both with stride of $3\\times 1\\times 1$ . $\\hat{U}_j^{(l)}, U_{j,1}^{(l)}$ and $U_{j,2}^{(l)}$ have the same size of $(C,\\frac{T}{3},H_j^{(l)},W)$ . The output of Eq.(5), $\\hat{U}_j^{(l)}$ , is the element-wise summation of the aggregation results of the two scales, $U_{j,1}^{(l)}$ and $U_{j,2}^{(l)}$ , which reduces the temporal dimension of the input from $T$ to $\\frac{T}{3}$ .", + "bbox": [ + 75, + 631, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the FTA model produces frame-level weights, which can be expressed as:", + "bbox": [ + 75, + 747, + 468, + 774 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Z _ {j, 1} ^ {(l)} = \\operatorname {F C} _ {j, 1} ^ {(l)} \\left(\\operatorname {G A P} \\left(\\hat {U} _ {j} ^ {(l)}\\right)\\right), \\tag {6} \\\\ Z _ {j, 2} ^ {(l)} = \\operatorname {F C} _ {j, 2} ^ {(l)} \\left(\\operatorname {G A P} \\left(\\hat {U} _ {j} ^ {(l)}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 781, + 468, + 831 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{GAP}(\\cdot)$ represents the global mean pooling along spatial dimension. $\\mathrm{FC}_{j,1}(\\cdot)$ and $\\mathrm{FC}_{j,2}(\\cdot)$ are two independ-", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dent fully connected layers that generate the frame selection weighting tensors, $Z_{j,1}^{(l)}$ and $Z_{j,2}^{(l)} \\in \\mathbb{R}^{C \\times \\frac{T}{3} \\times 1 \\times 1}$ , for $U_{j,1}^{(l)}$ and $U_{j,2}^{(l)}$ , respectively. The weights are further normalized across the two scales, which can be written as follows:", + "bbox": [ + 497, + 125, + 892, + 187 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {j, s, c, t} ^ {(l)} = \\frac {e ^ {Z _ {j , s , c , t} ^ {(l)}}}{e ^ {Z _ {j , 1 , c , t} ^ {(l)}} + e ^ {Z _ {j , 2 , c , t} ^ {(l)}}} \\quad s \\in \\{1, 2 \\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 195, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}_{j,s,c,t}^{(l)}\\in \\mathbb{R}^{1\\times 1\\times 1\\times 1}$ is the weight value of the $c$ -th channel of the $t$ -th frame. Combining Eq. (5) and Eq. (7), the $j$ -th output region feature $Y_{\\Omega ,j}^{(l)}\\in \\mathbb{R}^{C^{(l)}\\times \\frac{T}{3}\\times H_j^{(l)}\\times W}$ for the $l$ -th level of FTA can be obtained as follows:", + "bbox": [ + 497, + 243, + 892, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY _ {\\Omega , j} ^ {(l)} = \\mathcal {W} _ {j, 1} ^ {(l)} \\odot U _ {j, 1} ^ {(l)} + \\mathcal {W} _ {j, 2} ^ {(l)} \\odot U _ {j, 2} ^ {(l)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 312, + 890, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}_{j,1}^{(l)},\\mathcal{W}_{j,2}^{(l)}\\in \\mathbb{R}^{C\\times \\frac{T}{3}\\times 1\\times 1}$ are two weight tensors calculated using Eq. (5), and $\\odot$ represents element-wise multiplication operation. The FTA module outputs $Y_{\\Omega}^{(l)}\\in$ $\\mathbb{R}^{C\\times \\frac{T}{3}\\times H\\times W}$ by concatenating the $K_{l}$ gait regions of level $l$ where $Y_{\\Omega}^{(l)} = \\left[Y_{\\Omega ,1}^{(l)},Y_{\\Omega ,2}^{(l)},\\ldots ,Y_{\\Omega ,K_l}^{(l)}\\right]$", + "bbox": [ + 497, + 342, + 892, + 427 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 499, + 435, + 633, + 451 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 499, + 458, + 800, + 471 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CASIA-B. The CASIA-B [52] dataset is a widely used benchmark for gait recognition. It contains video sequences of 124 subjects with 11 different views and three walking conditions (normal walking (NM), walking with a bag (BG), and walking with a coat (CL)). Our study follows the protocol outlined in previous works [4, 8, 27, 25, 15, 16]. The first 74 subjects are used for training, and the remaining 50 subjects are used for testing. During testing, the first four sequences under NM (NM#01-04) are regarded as the gallery set, and the rest (NM#05-06, BG#01-02, CL#01-02) are regarded as the probe set.", + "bbox": [ + 497, + 479, + 890, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "OUMVLP. The OUMVLP [39] is one of the largest gait datasets, containing silhouette sequences of 10,307 subjects. Each subject has a single normal walking condition (NM) with 14 views. According to the protocol provided by the dataset, the first 5,153 subjects are used for training while the remaining 5,154 subjects are used for testing. During the testing phase, the sequences of NM#01 are assigned to the gallery set, and the sequences of NM#02 are considered as the probe set.", + "bbox": [ + 497, + 631, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GREW. GREW [60] is the first large-scale dataset for gait recognition in the wild, consisting of 128,671 sequences from 26,345 individuals captured by 882 cameras. It includes four data types: four data types: silhouettes, optical flow, 2D pose, and 3D pose. The dataset is divided into a training set with 20,000 subjects and 102,887 sequences, and a testing set with 6,000 subjects and 24,000 sequences. In the testing phase, each subject has two sequences for the", + "bbox": [ + 497, + 755, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19643", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d48851b29735f913159e9d0a776d016e16e222032da6cda922a3c06d1c9bbc8c.jpg", + "table_caption": [ + "Table 1: Rank-1 accuracy (%) on CASIA-B under all views and different conditions, excluding identical-view cases. Std denotes the performance sample standard deviation across 11 views." + ], + "table_footnote": [], + "table_body": "
Gallery NM #1-40° - 180°MeanStd
Probe0°18°36°54°72°90°108°126°144°162°180°
NM #5-6GaitSet [4]90.897.999.496.993.691.795.097.898.996.885.895.03.5
GaitPart [8]94.198.699.398.594.092.395.998.499.297.890.496.23.1
3D Local [16]96.099.099.598.997.194.296.399.098.898.595.297.51.8
CSTL [15]97.299.099.298.196.295.597.798.799.298.996.597.81.3
GaitGL [27]96.098.399.097.996.995.497.098.999.398.894.097.41.7
LagrangeGait [3]95.798.199.198.396.495.297.599.099.398.994.997.51.6
MetaGait [5]97.399.299.599.197.295.597.699.199.399.196.798.11.3
Ours97.698.099.698.297.496.597.999.399.498.497.098.11.0
BG #1-2GaitSet [4]83.891.291.888.883.381.084.190.092.294.479.087.24.9
GaitPart [8]89.194.896.795.188.384.989.093.596.193.885.891.54.2
3D Local [16]92.995.997.896.293.087.892.796.397.998.088.594.33.5
CSTL [15]91.796.597.095.490.988.091.595.897.095.590.393.63.0
GaitGL [27]92.696.696.895.593.589.392.296.598.296.991.594.52.8
LagrangeGait [3]94.296.296.895.894.389.591.796.898.097.090.994.62.7
MetaGait [5]92.996.797.196.494.790.492.997.298.598.192.395.22.6
Ours95.096.597.396.695.393.394.696.898.697.792.995.91.7
CL #1-2GaitSet [4]61.475.480.777.372.170.171.573.573.568.450.070.48.0
GaitPart [8]70.785.586.983.377.172.576.982.283.880.266.578.76.6
3D Local [16]78.290.292.087.183.076.883.186.686.884.170.983.76.2
CSTL [15]78.189.491.686.682.179.981.886.388.786.675.384.24.9
GaitGL [27]76.690.090.387.184.579.084.187.087.384.469.583.66.3
LagrangeGait [3]77.490.693.290.284.780.385.287.789.386.671.085.16.3
MetaGait [5]80.091.893.087.886.582.985.290.090.889.378.486.94.6
Ours82.494.295.091.788.283.388.092.393.191.078.588.95.1
", + "bbox": [ + 80, + 163, + 892, + 509 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gallery set and two for the probe set. The GREW dataset also includes a distractor set, which contains 233,857 unlabeled sequences.", + "bbox": [ + 75, + 533, + 468, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gait3D. The Gait3D dataset [59] is a newly proposed dataset for gait recognition in uncontrolled indoor environments, particularly in large supermarkets. It contains 25,309 sequences of 4,000 subjects extracted from 39 cameras, with 18,940 sequences from 3,000 subjects for training and 6,369 sequences from 1,000 subjects for testing. The dataset mainly includes four data types: silhouettes, 2D pose, 3D pose, and 3D mesh. During testing, one sequence per subject is used as the probe set, while the remaining sequences are used as the gallery set. To evaluate the model's performance, we use accuracy as well as mean average precision (mAP) and mean inverse negative penalty (mINP) [49], which consider multiple instances and hard sample recall.", + "bbox": [ + 75, + 574, + 468, + 769 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 77, + 775, + 294, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. 1) In our implementation, the margin $m$ of the triplet loss is set to 0.2, and the parameter $p$ of the GeM function used in the ASTP module is initialized to 6.5. In the hierarchy of gait motion, the number of partitions $k$ at the bottom level is set to 8. 2) The batch size is set to (8,8)", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: The detailed architecture of the proposed HSTL on CASIA-B. The first column denotes the levels of the gait hierarchy and $K_{l}$ is the number of groups at level $l$ . $C_{in}$ and $C_{out}$ represent the input channel and output channel of each layer respectively. The body parts are indexed in spatial order from top to bottom, numbered 1 to 8.", + "bbox": [ + 497, + 532, + 890, + 615 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4ad98b02a4fa046e472213cee361469b0e943d8a696d872552eb5e859c07ed7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LevelBlockLayer\\( {C}_{in} \\)\\( {C}_{out} \\)Kernel\\( {K}_{l} \\)Parts Grouping
1ARMEConv3d132(3,3,3)1\\( \\{ \\{ 1,2,3,4,5,6,7,8\\} \\} \\)
ASTP
2ARMEConv3d3232(3,3,3)2\\( \\{ \\{ 1,2,3,4,5\\} ,\\{ 6,7,8\\} \\} \\)
Conv3d3264(3,3,3)
ASTP
2FTAMaxPool6464(3,1,1)2\\( \\{ \\{ 1,2,3,4,5\\} ,\\{ 6,7,8\\} \\} \\)
(5,1,1)
ASTP
3ARMEConv3d64128(3,3,3)4\\( \\{ \\{ 1\\} ,\\{ 2,3,4,5\\} ,\\{ 6,7\\} ,\\{ 8\\} \\} \\)
Conv3d128128(3,3,3)
ASTP
4ASTP8\\( \\{ \\{ 1\\} ,\\{ 2\\} ,\\{ 3\\} ,\\{ 4\\} , \\) \\( \\{ 5\\} ,\\{ 6\\} ,\\{ 7\\} ,\\{ 8\\} \\} \\)
", + "bbox": [ + 502, + 625, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for CASIA-B, (32,8) for OUMVLP, (32,4) for GREW, and (32,4) for Gait3D. 3) We use gait silhouettes as the input", + "bbox": [ + 499, + 838, + 892, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19644", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c2b3c965b11b5d1fda20c6444107874709f56e54377223a8f1560bd413df11c9.jpg", + "table_caption": [ + "Table 3: Rank-1 accuracy (%) on OUMVLP under all views, excluding identical-view cases. Std denotes the performance sample standard deviation across 14 views." + ], + "table_footnote": [], + "table_body": "
MethodProbe ViewMeanStd
0°15°30°45°60°75°90°180°195°210°225°240°255°270°
GaitSet [4]79.387.990.090.188.088.787.781.886.589.089.287.287.686.287.14.0
GaitPart [8]82.688.990.891.089.789.989.585.288.190.090.189.089.188.288.72.3
GLN [12]83.890.091.091.290.390.089.485.389.190.590.689.689.388.589.22.1
CSTL [15]87.191.091.591.890.690.890.689.490.290.590.789.890.089.490.21.1
GaitGL [27]84.990.291.191.591.190.890.388.588.690.390.489.689.588.889.71.7
3D Local [16]86.191.292.692.992.291.391.186.990.892.292.391.391.190.290.92.0
LagrangeGait [3]85.990.691.391.591.291.090.688.989.290.590.689.989.889.290.01.4
MetaGait [5]88.292.393.093.593.192.792.689.391.292.092.692.391.991.191.91.4
Ours91.492.992.793.092.992.592.592.792.392.192.392.291.891.892.40.5
", + "bbox": [ + 80, + 162, + 892, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "modality. During the training phase, we sample 30 frames following the strategy proposed in [8], while during testing, all frames are fed into the model. Moreover, we align each frame following the strategy presented in [39], and the input image size is cropped to $64 \\times 44$ for all datasets. 4) For CASIA-B, the optimizer used is Adam with a weight decay of $5e-4$ . The model is trained for 100K iterations with an initial learning rate (LR) of $1e-5$ , and the LR is multiplied by 0.1 at 70K iterations. For OUMVLP, GREW, and Gait3D, the model is trained for 250K, 250K, and 210K iterations, respectively, with an initial LR of 0.1. The LR is multiplied by 0.1 at 150K and 200K iterations. The optimizer used for these datasets is SGD with a weight decay of $5e-4$ .", + "bbox": [ + 75, + 326, + 468, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architecture details. Table 2 presents the detailed architecture of the model used for the CASIA-B dataset. To handle datasets with more subjects, such as OUMVLP, GREW, and Gait3D, we incorporate the label smoothing operation into the cross-entropy loss function, and deepen the network by adding an extra ARME module to the third level of the hierarchy. The output channels for the four ARMEs are set to 64, 64, 128 and 256, respectively. Additionally, we include a layer of spatial downsampling after the first ARME to improve the training efficiency.", + "bbox": [ + 75, + 519, + 468, + 659 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/70616d4e9ab9b0717aa965c7a424f88f1c2fe3479ba9762367f4a4a5cfad8757.jpg", + "table_caption": [ + "Table 4: Rank-1 accuracy (\\%), Rank-5 accuracy (\\%), Rank-10 accuracy (\\%), Rank-20 accuracy (\\%) on GREW." + ], + "table_footnote": [], + "table_body": "
MethodsRank-1Rank-5Rank-10Rank-20
GaitSet [4]46.2863.5870.2676.82
GaitPart [8]44.0160.6867.2573.47
GaitGL [27]47.2863.5669.3274.18
MTSGait [58]55.3271.2876.8581.55
Ours62.7276.5781.3285.24
", + "bbox": [ + 82, + 706, + 468, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 77, + 817, + 445, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on CASIA-B. Table 1 compares the performance of the proposed HSTL with seven state-of-the-art", + "bbox": [ + 77, + 838, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b958a642034e427d47f8326f8a8b5804d5a10ec370dca672ce9d1a144f606adf.jpg", + "table_caption": [ + "Table 5: Rank-1 accuracy $(\\%)$ , Rank-5 accuracy $(\\%)$ , mAP $(\\%)$ and mINP on Gait3D." + ], + "table_footnote": [], + "table_body": "
MethodsRank-1Rank-5mAPmINP
GaitSet [4]36.7058.3030.0117.30
GaitPart [8]28.2047.6021.5812.36
GLN [12]31.4052.9024.7413.58
GaitGL [27]29.7048.5022.2913.26
CSTL [15]11.7019.205.592.59
SMPLGait [59]46.3064.5037.1622.23
MTSGait [58]48.7067.1037.6321.92
Ours61.3076.3055.4834.77
", + "bbox": [ + 504, + 362, + 892, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods on the CASIA-B dataset. It can be seen that the proposed method obtains the best results in all three walking conditions while maintaining considerable stability across different views. The experimental results reveal that 1) the mean accuracy of the proposed method for the BG and CL walking conditions is $95.9\\%$ and $88.9\\%$ , respectively, which are higher by $0.7\\%$ and $2.0\\%$ than the second-best method (MetaGait [5]), demonstrating the advantage of our method in cross-view gait recognition. 2) The decrease in accuracy from NM to CL is $9.2\\%$ for our method, compared to $12.4\\%$ for 3D Local [16]. This indicates that the ARME module is more adaptable to various walking conditions. In addition, our method achieves an accuracy of $88.9\\%$ in the challenging CL condition, which is $5.2\\%$ higher than 3D Local. This may be due to the complex clothing conditions that affect the part localization accuracy in 3D Local. 3) Both CSTL [15] and our method extracts multi-scale temporal features but in different ways. CSTL first extracts spatial information and then fuses three scales of motion features. In contrast, we propose an FTA module to aggregate spatiotemporal information from multiple body regions. Therefore, the average rank-1 accuracy of our method is higher or equivalent to that of CSTL in all views. This suggests that FTA can be more adaptive to spatio-temporal changes.", + "bbox": [ + 497, + 534, + 892, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19645", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b3df54efb68d4f7d1d4dd6f5e5575c510b7ee677da840905a1554140abb5d6a.jpg", + "image_caption": [ + "(a) Average rank-1 accuracy" + ], + "image_footnote": [], + "bbox": [ + 80, + 122, + 322, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fcfaaf69f6c0eaa2d1ceeeff340112c9d6f6103cc441745270a83b708dd7a20.jpg", + "image_caption": [ + "(b) Age distribution", + "Figure 4: Cross-age comparison results for the OUMVLP dataset. (a) Average rank-1 accuracy in adults vs children. (b) Distribution of two age groups." + ], + "image_footnote": [], + "bbox": [ + 327, + 169, + 467, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on OUMVLP. To verify the model's generalizability, we conduct experiments on the large-scale OUMVLP dataset. As shown in Table 3, our approach achieves competitive results in most views. In particular, the proposed method outperforms the second-best method (MetaGait) by an average of $2.5\\%$ under three extreme views, i.e., $0^{\\circ}$ , $180^{\\circ}$ , and $270^{\\circ}$ , resulting in the best mean performance and cross-view stability. In addition, the OUMVLP dataset also provides annotations for the ages of the subjects. To evaluate the impact of age differences on recognition performance, we divided all subjects into two groups: adults (18-87 years old) and children (2-17 years old), as shown in Fig. 4(b). Fig. 4(a) presents the recognition accuracy based on age. It can be observed that, as adult sequences make up about $70\\%$ of the total sequences, all the compared methods show a bias toward the recognition results of adults. However, compared to other methods, our model effectively improves the accuracy of gait recognition for children, demonstrating the effectiveness of our hierarchical gait representation across ages.", + "bbox": [ + 75, + 325, + 468, + 600 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on GREW and Gait3D. Gait3D and GREW are two recently introduced datasets that contain challenging conditions, such as misalignment of the human body and partial occlusion. Tables 4 and 5 show the results of the comparison between the proposed method and the state-of-the-art methods. Our method shows superior performance in all metrics, indicating its ability to effectively model gait characteristics in realistic scenarios.", + "bbox": [ + 75, + 600, + 468, + 711 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 77, + 720, + 230, + 735 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of hierarchical feature extraction. To verify the effectiveness of our hierarchical gait partitioning, we conduct experiment with various grouping strategies. As shown in Table 2, this comparison only considers the different settings of the first three layers since a uniform partition is used at layer 4. Specifically, 1-1-1 indicates that the first three levels have the same number of groupings, and they are divided uniformly. $1 - 2 - 4^{*}$ refers to the nonuniform division used in the proposed model. As shown in", + "bbox": [ + 75, + 741, + 468, + 865 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f157c36fffc6bc2e3a2fb3c459dafda257bdcb183813c7e78d9c9725f005e33f.jpg", + "image_caption": [ + "Figure 5: Ablation study on the effectiveness of hierarchical feature extraction on the CASIA-B dataset (best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 519, + 122, + 873, + 281 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2e005c7e03450ec5dd125490067d690fe893c82e73cb938094dcc2f211e63023.jpg", + "table_caption": [ + "Table 6: Ablation study on the effectiveness of ARME, ASTP, and FTA modules in terms of average rank-1 accuracy on the CASIA-B dataset." + ], + "table_footnote": [], + "table_body": "
SettingARMEASTPFTANMBGCLMean
35
a97.094.584.291.9
b✓97.895.385.993.0
c✓✓97.895.487.093.4
d✓✓✓97.795.387.593.5
e✓✓✓97.795.287.293.4
f✓✓✓97.895.487.593.6
g✓✓✓✓98.195.988.994.3
", + "bbox": [ + 504, + 396, + 890, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5, the hierarchical feature extraction setting, e.g., 1-2-4, outperforms the other non-hierarchical approaches. A further mean performance improvement of $1.0\\%$ is achieved when the motion relationships between different body regions, i.e., 1-2-4*, are considered.", + "bbox": [ + 497, + 532, + 890, + 600 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of ARME, ASTP, and FTA. The results of the ablation experiments for ARME, ASTP, and FTA are shown in Tab. 6 The results indicate that the ARME module significantly contributes to the improvement of the recognition accuracy, with an average improvement of $1.1\\%$ compared to the baseline model (non-grouping version of ARME). The mean accuracy is further improved by $1.3\\%$ when the ASTP module is integrated and multi-scale fusion is utilized in FTA. These results demonstrate the effectiveness and complementarity of the three modules in the proposed gait recognition framework.", + "bbox": [ + 497, + 600, + 890, + 754 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Trade-off between accuracy and efficiency", + "text_level": 1, + "bbox": [ + 499, + 762, + 858, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, we evaluate the relationship between accuracy and efficiency for each compared method. As shown in Fig. 6, the 3D convolution-based approaches, such as GaitGL [27], 3D Local [16], LagrangeGait [3] and MetaGait [5], outperform the 2D convolution-based methods, like GaitSet [4] and GaitPart [8], in terms of accuracy, but", + "bbox": [ + 497, + 783, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19646", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e37d6a628cb52e34fa5fd7fa799c99e534ef52eb2e2b158c8103251db87897ae.jpg", + "image_caption": [ + "Figure 6: The trade-off between accuracy and FLOPs of our method and other comparison methods on the OUMVLP dataset." + ], + "image_footnote": [], + "bbox": [ + 78, + 123, + 472, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "at the cost of a significant increase in FLOPs (floating point operations). Our approach has a better trade-off between accuracy and efficiency. The main reason is that the proposed hierarchical learning architecture can extract multilevel motion features while reducing the number of 3D convolutions. More experimental results and ablation analysis are provided in the supplementary material.", + "bbox": [ + 75, + 385, + 468, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 77, + 496, + 194, + 509 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper presents a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL stacks multiple adaptive region-based motion extractors (ARMEs) and learns walking patterns in a coarse-to-fine manner. An adaptive spatio-temporal pooling (ASTP) module is proposed to perform hierarchical feature mapping for the output of each level of ARME. Additionally, a frame-level temporal aggregation module (FTA) is designed to compress local clips by fusing temporal information. The effectiveness of the proposed HSTL framework is demonstrated through extensive experiments conducted on four public datasets (CASIA-B, OUMVLP, GREW, and Gait3D).", + "bbox": [ + 75, + 519, + 472, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 77, + 712, + 243, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 61972132, 62106065) and the Research Project for Self-cultivating Talents of Hebei Agricultural University (Grant No. PY201810).", + "bbox": [ + 75, + 735, + 468, + 805 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 817, + 174, + 831 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[1] Weizhi An, Shiqi Yu, Yasushi Makihara, Xinhui Wu, Chi Xu, Yang Yu, Rijun Liao, and Yasushi Yagi. Performance evalua", + "bbox": [ + 84, + 839, + 468, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion of model-based gait on multi-view very large population database with pose sequences. IEEE TBIOM, 2(4):421-430, 2020. 1, 2", + "[2] Pia Bideau, Aruni RoyChowdhury, Rakesh R Menon, and Erik Learned-Miller. The best of both worlds: Combining cnns and geometric constraints for hierarchical motion segmentation. In CVPR, pages 508-517, 2018. 1", + "[3] Tianrui Chai, Annan Li, Shaoxiong Zhang, Zilong Li, and Yunhong Wang. Lagrange motion analysis and view embeddings for improved gait recognition. In CVPR, pages 20249-20258, 2022. 1, 2, 3, 6, 7, 8", + "[4] Hanqing Chao, Yiwei He, Junping Zhang, and Jianfeng Feng. Gaitset: Regarding gait as a set for cross-view gait recognition. In AAAI, pages 8126-8133, 2019. 1, 2, 3, 4, 5, 6, 7, 8", + "[5] Huanzhang Dou, Pengyi Zhang, Wei Su, Yunlong Yu, and Xi Li. Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In ECCV, pages 357-374. Springer, 2022. 2, 4, 6, 7, 8", + "[6] Jessica Maria Echterhoff, Juan Haladjian, and Bernd Brügge. Gait and jump classification in modern equestrian sports. In ISWC, pages 88-91, 2018. 1", + "[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 2, 4", + "[8] Chao Fan, Yunjie Peng, Chunshui Cao, Xu Liu, Saihui Hou, Jiannan Chi, Yongzhen Huang, Qing Li, and Zhiqiang He. Gaitpart: Temporal part-based model for gait recognition. In CVPR, pages 14225-14233, 2020. 1, 2, 3, 4, 5, 6, 7, 8", + "[9] Yang Fu, Yunchao Wei, Yuqian Zhou, Honghui Shi, Gao Huang, Xinchao Wang, Zhiqiang Yao, and Thomas Huang. Horizontal pyramid matching for person re-identification. In AAAI, pages 8295–8302, 2019. 4", + "[10] Jinguang Han and Bir Bhanu. Individual recognition using gait energy image. IEEE TPAMI, 28(2):316-322, 2005. 1, 2", + "[11] Ruibing Hou, Hong Chang, Bingpeng Ma, Rui Huang, and Shiguang Shan. Bicnet-tks: Learning efficient spatial-temporal representation for video person re-identification. In CVPR, pages 2014–2023, 2021. 3", + "[12] Saihui Hou, Chunshui Cao, Xu Liu, and Yongzhen Huang. Gait lateral network: Learning discriminative and compact representations for gait recognition. In ECCV, pages 382-398, 2020. 1, 2, 3, 4, 7", + "[13] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Set residual network for silhouette-based gait recognition. IEEE TBIOM, 3(3):384-393, 2021. 2, 3", + "[14] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Gait quality aware network: Toward the interpretability of silhouette-based gait recognition. IEEE TNNLS, 2022. 2", + "[15] Xiaohu Huang, Duowang Zhu, Hao Wang, Xinggang Wang, Bo Yang, Botao He, Wenyu Liu, and Bin Feng. Context-sensitive temporal feature learning for gait recognition. In ICCV, pages 12909-12918, 2021. 1, 2, 3, 4, 5, 6, 7", + "[16] Zhen Huang, Dixiu Xue, Xu Shen, Xinmei Tian, Houqiang Li, Jianqiang Huang, and Xian-Sheng Hua. 3d local convolutional neural networks for gait recognition. In ICCV, pages 14920-14929, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 500, + 125, + 894, + 865 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19647", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Munish Kumar, Navdeep Singh, Ravinder Kumar, Shubham Goel, and Krishan Kumar. Gait recognition based on vision systems: A systematic survey. Journal of Visual Communication and Image Representation, 75:103052, 2021. 1", + "[18] Worapan Kusakunniran. Review of gait recognition approaches and their challenges on view changes. IET Biometrics, 9(6):238-250, 2020. 1", + "[19] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE TIP, 31:5949-5962, 2022. 2", + "[20] Xiang Li, Yasushi Makihara, Chi Xu, and Yasushi Yagi. End-to-end model-based gait recognition using synchronized multi-view pose constraint. In ICCVW, pages 4106-4115, 2021. 1, 2", + "[21] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, and Mingwu Ren. Gait recognition via semi-supervised disentangled representation learning to identity and covariate features. In CVPR, pages 13309-13319, 2020. 3", + "[22] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, Shiqi Yu, and Mingwu Ren. End-to-end model-based gait recognition. In ACCV, pages 3-20, 2020. 1, 2", + "[23] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In CVPR, pages 510-519, 2019. 3", + "[24] Junhao Liang, Chao Fan, Saihui Hou, Chuanfu Shen, Yongzhen Huang, and Shiqi Yu. Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In ECCV, pages 375-390. Springer, 2022. 2", + "[25] Beibei Lin, Yu Liu, and Shunli Zhang. Gaitmask: Mask-based model for gait recognition. In BMVC, page 363, 2021. 1, 2, 3, 4, 5", + "[26] Beibei Lin, Shunli Zhang, and Feng Bao. Gait recognition with multiple-temporal-scale 3d convolutional neural network. In ACM MM, pages 3054-3062, 2020. 3, 4", + "[27] Beibei Lin, Shunli Zhang, and Xin Yu. Gait recognition via effective global-local feature representation and local temporal aggregation. In ICCV, pages 14648-14656, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[28] Nini Liu and Yap-Peng Tan. View invariant gait recognition. In ICASSP, pages 1410-1413, 2010. 1", + "[29] Yasushi Makihara, Mark S Nixon, and Yasushi Yagi. Gait recognition: Databases, representations, and applications. Computer Vision: A Reference Guide, pages 1-13, 2020. 1", + "[30] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical gaussian descriptor for person re-identification. In CVPR, pages 1363–1372, 2016. 2", + "[31] Daigo Muramatsu, Yasushi Makihara, Haruyuki Iwama, Takuya Tanoue, and Yasushi Yagi. Gait verification system for supporting criminal investigation. In IAPR, pages 747-748, 2013. 1", + "[32] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 41(7):1655-1668, 2018. 4", + "[33] Imad Rida, Noor Almaadeed, and Somaya Almaadeed. Robust gait recognition: a comprehensive survey. IET Biometrics, 8(1):14-28, 2019. 1" + ], + "bbox": [ + 78, + 125, + 468, + 865 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Alireza Sepas-Moghaddam and Ali Etemad. Deep gait recognition: A survey. IEEE TPAMI, 2022. 1", + "[35] Kohei Shiraga, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Geinet: View-invariant gait recognition using a convolutional neural network. In ICB, pages 1-8, 2016. 1, 2, 3", + "[36] Pranjay Shyam, Kuk-Jin Yoon, and Kyung-Soo Kim. Adversarily-trained hierarchical feature extractor for vehicle re-identification. In ICRA, pages 13400–13407, 2021. 2", + "[37] Chunfeng Song, Yongzhen Huang, Yan Huang, Ning Jia, and Liang Wang. Gaitnet: An end-to-end network for gait based human identification. PR, 96:106988, 2019. 2", + "[38] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 1", + "[39] Noriko Takemura, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ TCVA, 10:1-14, 2018. 2, 5, 7", + "[40] Lei Tan, Pingyang Dai, Rongrong Ji, and Yongjian Wu. Dynamic prototype mask for occluded person re-identification. In ACM MM, pages 531–540, 2022. 2", + "[41] Torben Teepe, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Towards a deeper understanding of skeleton-based gait recognition. In CVPRW, pages 1569-1577, 2022. 2", + "[42] Torben Teepe, Ali Khan, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In ICIP, pages 2314-2318, 2021. 1, 2", + "[43] Md Uddin, Daigo Muramatsu, Noriko Takemura, Md Ahad, Atiqur Rahman, Yasushi Yagi, et al. Spatio-temporal silhouette sequence reconstruction for gait recognition against occlusion. IPSJ TCVA, 11(1):1-18, 2019. 1", + "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2", + "[45] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. IEEE TCSVT, 2021. 2", + "[46] Xiu-Shen Wei, Chen-Lin Zhang, Lingqiao Liu, Chunhua Shen, and Jianxin Wu. Coarse-to-fine: A rnn-based hierarchical attention model for vehicle re-identification. In ACCV, pages 575–591, 2018. 2", + "[47] Jiachen Yang, Jianxiong Zhou, Dayong Fan, and Haibin Lv. Design of intelligent recognition system based on gait recognition technology in smart transportation. Multimedia Tools and Applications, 75(24):17501-17514, 2016. 1", + "[48] Lingxiang Yao, Worapan Kusakunniran, Qiang Wu, Jingsong Xu, and Jian Zhang. Collaborative feature learning for gait recognition under cloth changes. IEEE TCSVT, 2021. 1", + "[49] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6):2872-2893, 2021. 6" + ], + "bbox": [ + 502, + 127, + 890, + 865 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19648", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] TzeWei Yeoh, Hernan E Aguirre, and Kiyoshi Tanaka. Clothing-invariant gait recognition using convolutional neural network. In ISPACS, pages 1-5, 2016. 1", + "[51] Shiqi Yu, Haifeng Chen, Qing Wang, Linlin Shen, and Yongzhen Huang. Invariant feature extraction for gait recognition using only one uniform model. Neurocomputing, 239:81-93, 2017. 1, 2", + "[52] Shiqi Yu, Daoliang Tan, and Tieniu Tan. A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In ICPR, pages 441-444, 2006. 2, 5", + "[53] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 2", + "[54] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 17:516-530, 2022. 2", + "[55] Shaoxiong Zhang, Yunhong Wang, and Annan Li. Crossview gait recognition with deep universal linear embeddings. In CVPR, pages 9095-9104, 2021. 1", + "[56] Yuqi Zhang, Yongzhen Huang, Shiqi Yu, and Liang Wang. Cross-view gait recognition by discriminative feature learning. IEEE TIP, 29:1001-1015, 2019. 1, 3, 4", + "[57] Ziyuan Zhang, Luan Tran, Xi Yin, Yousef Atoum, Xiaoming Liu, Jian Wan, and Nanxin Wang. Gait recognition via disentangled representation learning. In CVPR, pages 4710-4719, 2019. 2", + "[58] Jinkai Zheng, Xinchen Liu, Xiaoyan Gu, Yaoqi Sun, Chuang Gan, Jiyong Zhang, Wu Liu, and Chenggang Yan. Gait recognition in the wild with multi-hop temporal switch. In ACM MM, pages 6136-6145, 2022. 7", + "[59] Jinkai Zheng, Xinchen Liu, Wu Liu, Lingxiao He, Chenggang Yan, and Tao Mei. Gait recognition in the wild with dense 3d representations and a benchmark. In CVPR, pages 20228-20237, 2022. 1, 2, 6, 7", + "[60] Zheng Zhu, Xianda Guo, Tian Yang, Junjie Huang, Jiankang Deng, Guan Huang, Dalong Du, Jiwen Lu, and Jie Zhou. Gait recognition in the wild: A benchmark. In ICCV, pages 14789-14799, 2021. 2, 5" + ], + "bbox": [ + 78, + 125, + 468, + 656 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19649", + "bbox": [ + 467, + 887, + 504, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_model.json b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7a0686026505ae76ac10cf45812e2b996ba0cedc --- /dev/null +++ b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.096, + 0.009, + 0.171, + 0.047 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.232, + 0.008, + 0.816, + 0.049 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.162, + 0.871, + 0.182 + ], + "angle": 0, + "content": "Hierarchical Spatio-Temporal Representation Learning for Gait Recognition" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.207, + 0.726, + 0.225 + ], + "angle": 0, + "content": "Lei Wang\\(^{1,2}\\), Bo Liu\\(^{1,2*}\\), Fangfang Liang\\(^{1,2}\\), Bincheng Wang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.225, + 0.829, + 0.257 + ], + "angle": 0, + "content": "1 College of Information Science and Technology, Hebei Agricultural University, China \n2 Hebei Key Laboratory of Agricultural Big Data, China" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.259, + 0.835, + 0.273 + ], + "angle": 0, + "content": "{20212060107, 20212060108}@pgs.hebau.edu.cn, {boliu, liangfangfang}@hebau.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.319 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.335, + 0.474, + 0.709 + ], + "angle": 0, + "content": "Gait recognition is a biometric technique that identifies individuals by their unique walking styles, which is suitable for unconstrained environments and has a wide range of applications. While current methods focus on exploiting body part-based representations, they often neglect the hierarchical dependencies between local motion patterns. In this paper, we propose a hierarchical spatio-temporal representation learning (HSTL) framework for extracting gait features from coarse to fine. Our framework starts with a hierarchical clustering analysis to recover multilevel body structures from the whole body to local details. Next, an adaptive region-based motion extractor (ARME) is designed to learn region-independent motion features. The proposed HSTL then stacks multiple ARMEs in a top-down manner, with each ARME corresponding to a specific partition level of the hierarchy. An adaptive spatiotemporal pooling (ASTP) module is used to capture gait features at different levels of detail to perform hierarchical feature mapping. Finally, a frame-level temporal aggregation (FTA) module is employed to reduce redundant information in gait sequences through multi-scale temporal downsampling. Extensive experiments on CASIA-B, OUMVLP, GREW, and Gait3D datasets demonstrate that our method outperforms the state-of-the-art while maintaining a reasonable balance between model accuracy and complexity. Code is available at: https://github.com/gudaochangsheng/HSTL." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.736, + 0.208, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.759, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Unlike other biometric technologies such as fingerprint, iris, and face, human gait can be captured at a distance without subject cooperation [34]. By evaluating individual-specific walking patterns, gait recognition has been applied in a variety of fields, including criminal investigations [31, 29], sports science [17, 6], and smart transportation" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.304, + 0.895, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.42, + 0.893, + 0.504 + ], + "angle": 0, + "content": "Figure 1: The motivation for our approach. Left: the motion distributions of body parts. The same color indicates the same spatial location across multiple gait sequences in the CASIA-B dataset. Right: an example hierarchy of body-part sequences where \\( T \\) denotes the temporal dimension of the sequence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.531, + 0.893, + 0.573 + ], + "angle": 0, + "content": "[47]. However, the recognition can be challenging due to large variations in viewpoint [28, 18], occlusion [33, 43], and wearing [50, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.893, + 0.755 + ], + "angle": 0, + "content": "To address these issues, various approaches have been proposed for extracting gait features from silhouette sequences [4, 25, 16, 15, 55], 3D human structures [1, 22, 42, 59, 20], or gait templates [10, 35, 51]. Silhouette-based gait recognition methods have gained increasing attention due to the ease of obtaining silhouettes from raw videos while preserving essential temporal information. The alignment of the input silhouette makes it possible for some methods to extract local body features by horizontally slicing the silhouette image [56] or intermediate-layer features [8, 27]. This partitioning strategy, first introduced in person re-identification (ReID) [38], has been proven to be effective for gait recognition [4, 8, 12, 3]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.894, + 0.868 + ], + "angle": 0, + "content": "However, the main limitation of the above part-based approaches is that they do not consider the hierarchical nature of local body movements [2]. For instance, within a gait cycle, the feet and lower body have distinct motion characteristics. Therefore, it is important to treat these body regions separately and investigate their part-whole relationships. Our motivation stems from the examination of body-part-specific motion clues. Specifically, each raw gait se" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.854, + 0.225, + 0.867 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19639" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.319 + ], + "angle": 0, + "content": "quence in the CASIA-B [52] dataset is uniformly divided into eight part sequences along the body axis, so that each division roughly match a particular body part. The distributions of all body parts are shown in the left part of Fig. 1. Observably, some parts, e.g., the head and feet, are easily separated owing to their large changes in walking kinematics. Whereas other parts, such as the thighs and calves, overlap due to the strong motion correlations between them. Further, to identify the relational structure among the part sequences, a hierarchical clustering analysis [7] is performed. The results are shown in the right part of Fig. 1, indicating that the semantic body regions can be captured in the higher clustering levels without precise localization of the body parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.473, + 0.555 + ], + "angle": 0, + "content": "Following the above findings, we propose a novel hierarchical spatio-temporal representation learning (HSTL) framework for gait representation. The HSTL framework consists of multiple adaptive region-based motion extractor (ARME) modules, which are stacked to learn hierarchical motion patterns implied in a gait sequence (as shown in Fig.1). In the ARME module, to account for inter-regional differences, non-shared 3D convolutions are used in correspondence with individual body regions. There regions are pre-identified by a hierarchical clustering process performed on fixed horizontal partitions, allowing each body region to cover one or more body parts. Consequently, the deeper the ARME is, the more local features it tends to extract. Moreover, an adaptive spatio-temporal pooling (ASTP) module is proposed, which couples with an ARME module on the corresponding level to obtain hierarchical gait embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.473, + 0.695 + ], + "angle": 0, + "content": "In addition, changes in gait speed or sampling frequency may result in several redundant frames in a gait sequence. Although several temporal fusion strategies have been proposed, they lose spatial information [8, 15] or lack adaptability [27, 25]. To address this issue, we propose a frame-level temporal aggregation strategy (FTA). FTA fuses temporal features at multiple time steps, preserving significant motion information while compressing the sequence length. The main contributions of this paper are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.707, + 0.472, + 0.778 + ], + "angle": 0, + "content": "1. We propose a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL takes into account the dependencies of body regions in gait motions, ensuring simplicity and scalability of the architectural design." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.788, + 0.472, + 0.845 + ], + "angle": 0, + "content": "2. We introduce an adaptive region-based motion extractor (ARME) module to learn region-independent spatio-temporal representation for gait sequences, an adaptive spatio-temporal pooling (ASTP) module to" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.707, + 0.472, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.127, + 0.894, + 0.169 + ], + "angle": 0, + "content": "perform hierarchical feature mapping, and a frame-level temporal aggregation (FTA) strategy to compress a gait sequence by removing redundant frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.179, + 0.895, + 0.264 + ], + "angle": 0, + "content": "3. Extensive experiments on the widely used gait datasets CASIA-B [52], including OUMVLP [39], GREW [60] and Gait3D [59], demonstrate that our method achieves state-of-the-art performance while offering a suitable trade-off between model accuracy and complexity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.277, + 0.642, + 0.292 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.668, + 0.315 + ], + "angle": 0, + "content": "2.1. Gait Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.895, + 0.53 + ], + "angle": 0, + "content": "Deep learning-based gait recognition methods can be broadly categorized into two categories: model-based and appearance-based. Model-based approaches extract structure and motion information from gait videos with the aid of pose estimation [1, 22, 42, 20, 41, 59]. Although these methods are robust to changes in viewpoint and appearance, they are sensitive to the accuracy of the pose parameters, making them incapable of handling low-resolution data. On the other hand, appearance-based approaches learn the feature representation from raw videos [57, 37, 24], or binary silhouette sequences [10, 35, 51, 4, 12, 8, 13, 27, 16, 14, 3, 5], which offer greater flexibility than model-based approaches. Our proposed method belongs to the family of appearance-based methods and uses silhouette sequences as inputs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.688, + 0.555 + ], + "angle": 0, + "content": "2.2. Hierarchical Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Hierarchical feature representation has been successfully applied to a wide range of vision tasks. Here, we provide a brief review of the hierarchical object ReID approaches related to gait recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.895, + 0.867 + ], + "angle": 0, + "content": "In person ReID, some approaches [30, 54, 45, 40, 53] hierarchically learn local descriptions and aggregate appearance features at different levels. For example, Matsukawa et al. [30] described an image patch via hierarchical Gaussian distribution. Zhang et al. [53] proposed a framework to learn coarse-grained and fine-grained features according to body structure. To solve the occlusion problem, Tan et al. [40] devised a hierarchical mask generator to learn from both occluded and holistic joint images. In vehicle ReID, some approaches [46, 36, 19] extract features from vehicle images in a hierarchical manner. For instance, Wei et al. [46] proposed an RNN-based module for extracting latent cues from the model level to the vehicle level. Shyam et al. [36] developed an attention-based hierarchical feature extractor. In addition, Li et al. [19] proposed a global structural embedding module for investigating hierarchical relationships between vehicle characteristics by incorporating attribute and state information." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.854, + 0.461, + 0.867 + ], + "angle": 0, + "content": "Each part is pooled into a vector for visualization using t-SNE [44]." + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.124, + 0.891, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Figure 2: The framework of HFSL. It mainly consists of three modules: ARME (adaptive region-based motion extractor), ASTP (adaptive spatio-temporal pooling), and FTA (frame-level temporal aggregation). During pre-processing, a hierarchy of walking is obtained to guide the architectural design of HFSL. The framework uses multiple ARMEs to extract gait features from the entire body to individual regions. The ASTP module performs hierarchical feature mapping for the output of each level of ARME. The FTA module compresses local clips of each gait sequence to reduce the number of redundant frames. \\( T \\), \\( H \\) and \\( W \\) denote dimensions of the feature maps. © represents the concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.471, + 0.743 + ], + "angle": 0, + "content": "For gait recognition, fusing features of multiple granularities can improve performance [56, 8, 27, 25, 16, 3]. In particular, CSTL [15] proposed a temporal modeling network that integrates multi-scale temporal features adaptively. By combining part-level and sequence-level features, GaitPart [8] obtained a part-independent spatio-temporal expression. Additionally, GaitGL [27] considered both full body-based and part-based information to achieve discriminative feature learning. Instead of equally dividing the feature maps, in 3D Local [16], a localization operation was developed to find 3D volumes of body parts in a sequence. However, most existing gait recognition methods do not sufficiently exploit the hierarchical dependencies among body parts during walking. In this paper, the proposed HSTL performs a coarse-to-fine hierarchical strategy that integrates multi-level motion patterns from gait sequences." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.748, + 0.242, + 0.763 + ], + "angle": 0, + "content": "2.3. Temporal Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Temporal cues play a crucial role in gait recognition due to the periodic changes in body shape. Previous methods treat a gait sequence as an unordered set, either compressing it into a single gait template during preprocessing [35, 21] or learning order-independent gait representations from silhouette sets [4, 12, 13]. These methods assume that different subjects share similar global gait patterns, making or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.893, + 0.73 + ], + "angle": 0, + "content": "dering inputs unnecessary for gait assessment. However, ignoring the temporal nature of the gait sequence can result in missing discriminative local motion information. Recently, some approaches have achieved significant performance gains by explicitly modeling temporal information using LSTM [56], 1D convolution [15], and 3D convolution [26, 16]. Nevertheless, these spatio-temporal operators also significantly increase computational costs. Although some methods have been proposed to reduce video length by aggregating local clips [27, 25, 3], they lack adaptability to variations in pace. The main difference between our approach and others [23, 11] is that we employ multi-scale temporal pooling at the frame level while considering variations in motion across body regions, leading to a more adaptable reduction of the gait sequence length." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.776, + 0.673, + 0.792 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.812, + 0.893, + 0.868 + ], + "angle": 0, + "content": "In this section, we present the detailed description of HSTL, including the adaptive region-based motion extractor (ARME), the adaptive spatio-temporal pooling (ASTP), and the frame-level temporal aggregation (FTA)." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.504, + 0.898 + ], + "angle": 0, + "content": "19641" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.126, + 0.273, + 0.14 + ], + "angle": 0, + "content": "3.1. Framework Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.147, + 0.473, + 0.493 + ], + "angle": 0, + "content": "The overview of our HSTL is presented in Fig. 2. Given a gait dataset \\(\\mathcal{D} = \\{S_i\\}_{i=1}^N\\) with \\(N\\) gait sequences, where each sequence \\(S_i \\in \\mathbb{R}^{C \\times T \\times H \\times W}\\) is represented as a 4D tensor with \\(C\\) channels, \\(T\\) frames, and \\(H \\times W\\) pixels. During the preprocessing stage, each gait sequence \\(S_i\\) is divided horizontally and uniformly into \\(k\\) part sequences, indexed from 1 to \\(k\\). Then, a hierarchical clustering algorithm [7] is applied to these part sequences to obtain a generic hierarchy of gait motions, which is denoted as \\(\\mathcal{P} = \\{\\mathcal{P}^{(l)}\\}_{l=1}^L\\). Here, \\(L\\) is the number of levels in the hierarchy and \\(\\mathcal{P}^{(l)}\\) is the set of partitions at level \\(l\\). The partitions at level \\(l\\) are defined as \\(\\mathcal{P}^{(l)} = \\{P_1^{(l)}, P_2^{(l)}, \\ldots, P_{K_l}^{(l)}\\}\\), where \\(P_j^{(l)}\\) is the \\(j\\)-th subset of part indices and \\(K_l\\) is the number of groups at level \\(l\\). For instance, the top level \\(\\mathcal{P}^{(1)} = \\{\\{1, 2, \\ldots, k-1, k\\}\\}\\) means all the \\(k\\) parts can be considered as a whole for the gait analysis. This hierarchy provides a structured property of the gait motion patterns and can be utilized to guide gait feature extraction. To achieve this, our proposed HSTL employs three modules: ARME for extracting independent multi-granularity motion features, ASTP for generating vectorized gait embeddings, and FTA for reducing redundant information at the frame level. The HSTL stacks these three modules according to the division in \\(\\mathcal{P}\\), and the main branch of the HSTL for the input sequence \\(S_{in}\\) can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.5, + 0.47, + 0.518 + ], + "angle": 0, + "content": "\\[\nY ^ {M} = \\Gamma^ {(L)} \\circ \\Psi^ {(L - 1)} \\circ \\dots \\circ \\Omega^ {(2)} \\circ \\Psi^ {(2)} \\circ \\Psi^ {(1)} (S _ {i n}), (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.47, + 0.594 + ], + "angle": 0, + "content": "where \\(\\Psi^{(l)},\\Gamma^{(l)}\\) , and \\(\\Omega^{(l)}\\) represent the ARME, ASTP and FTA modules at the \\(l\\) -th level of \\(\\mathcal{P}\\) , respectively. Since FTA uses inter-frame compression to reduce redundant information, it is employed only once at the \\(l_{\\Omega}\\) -th level in \\(\\mathcal{P}\\) (e.g., \\(l_{\\Omega} = 2\\) in Eq.(1)) to prevent excessive loss of information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.472, + 0.695 + ], + "angle": 0, + "content": "To obtain the hierarchical gait embeddings, the output \\( Y^{(l)} \\) of each \\( \\Psi^{(l)} \\) at levels \\( l \\in \\{1,2,\\dots ,L - 1\\} \\) and the output of \\( \\Omega^{(l_{\\Omega})} \\) at level \\( l_{\\Omega} \\), denoted as \\( Y_{\\Omega}^{(l_{\\Omega})} \\), are fed into the corresponding \\( \\Gamma^{(l)} \\). The resulting outputs from these \\( L \\) auxiliary branches are concatenated with the output of the main branch defined in Eq. (1), forming the final result, denoted as \\( Y \\), which is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.7, + 0.47, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Y = \\left[ Y ^ {M}, \\Gamma^ {(L - 1)} \\left(Y ^ {(L - 1)}\\right), \\dots , \\right. \\\\ \\left. \\Gamma^ {\\left(l _ {\\Omega}\\right)} \\left(Y _ {\\Omega} ^ {\\left(l _ {\\Omega}\\right)}\\right), \\Gamma^ {(2)} \\left(Y ^ {(2)}\\right), \\Gamma^ {(1)} \\left(Y ^ {(1)}\\right) \\right], \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.77, + 0.386, + 0.784 + ], + "angle": 0, + "content": "where \\([,]\\) denotes the concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Finally, \\( Y \\) undergoes feature mapping through separate fully connected layers. The model is then trained using a combination of triplet loss \\( \\mathcal{L}_{tri} \\) and cross-entropy loss \\( \\mathcal{L}_{ce} \\), which is a commonly adopted practice in gait recognition [12, 27, 15, 16, 5]. Further details regarding the relevant modules are described in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.126, + 0.855, + 0.14 + ], + "angle": 0, + "content": "3.2. Adaptive Region-based Motion Extractor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.148, + 0.894, + 0.272 + ], + "angle": 0, + "content": "The adaptive region-based motion extractor (ARME) aims to extract independent spatio-temporal patterns that are associated with different human body parts in a gait sequence. Unlike existing methods that uniformly slice gait images or sequences along the height axis [56, 8, 27, 25], ARME considers the inherent hierarchical relationships among different part sequences that are consistent with walking patterns. This allows ARME to effectively capture the unique walking kinematics of each part." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.894, + 0.385 + ], + "angle": 0, + "content": "Given the hierarchical relation \\(\\mathcal{P}\\) introduced in Section 3.1, ARME first divides the input sequence \\(X\\) into \\(K_{l}\\) regions based on the partition of the \\(l\\)-th level \\(\\mathcal{P}^{(l)}\\), resulting in the set of regions \\(\\{X_{j}^{(l)}\\}_{j=1}^{K_{l}}\\), where \\(X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}\\). \\(H_{j}^{(l)} = \\frac{\\left|P_{j}^{(l)}\\right|}{k} H\\) is the height of the \\(j\\)-th region of the \\(l\\)-th level. Then the \\(l\\)-th level of ARME, \\(\\Psi^{(l)}\\), can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.393, + 0.896, + 0.428 + ], + "angle": 0, + "content": "\\[\nY _ {\\Psi} ^ {(l)} = \\Psi^ {(l)} \\left(X ^ {(l)}\\right) = \\left[ f _ {1} \\left(X _ {1} ^ {(l)}\\right), f _ {2} \\left(X _ {2} ^ {(l)}\\right), \\dots , f _ {K _ {l}} \\left(X _ {K _ {l}} ^ {(l)}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.894, + 0.514 + ], + "angle": 0, + "content": "where \\(f_{j}(\\cdot)\\) represents the independent 3D convolution operation applied to the \\(j\\)-th region. The output feature map \\(Y_{\\Psi}^{(l)} \\in \\mathbb{R}^{C(l) \\times T \\times H \\times W}\\) of level \\(l\\) has \\(C^{(l)}\\) channels. This module only modifies the number of channels, preserving the spatial and temporal resolutions of the input feature map." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.524, + 0.804, + 0.54 + ], + "angle": 0, + "content": "3.3. Adaptive Spatio-Temporal Pooling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.894, + 0.719 + ], + "angle": 0, + "content": "It is a common procedure in gait recognition to obtain a compact and fixed-length feature representation by performing horizontal and uniform slicing of feature maps and strip-based pooling [4, 9, 26, 27, 15]. However, a nonuniform division of the feature maps is better at capturing the gait motion characteristics. Thus, the adaptive spatiotemporal pooling (ASTP) is devised to construct hierarchical feature mapping (as shown in Fig. 2). Similar to the ARME module described in Section 3.2, the hierarchy \\(\\mathcal{P}\\) enables us to obtain the \\(j\\)-th region of the \\(l\\)-th level, denoted as \\(X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}\\). The corresponding ASTP, denoted as \\(\\Gamma^{(l)}\\), can be expressed as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.73, + 0.892, + 0.752 + ], + "angle": 0, + "content": "\\[\nY _ {\\Gamma , j} ^ {(l)} = \\Gamma_ {j} ^ {(l)} \\left(X _ {j} ^ {(l)}\\right) = \\operatorname {G e M} _ {j} \\circ \\operatorname {F C} \\circ \\operatorname {M a x} \\left(X _ {j} ^ {(l)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.894, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathrm{Max}(\\cdot)\\) represents a max pooling operation along the temporal dimension, \\(\\mathrm{FC}(\\cdot)\\) represents a fully connected layer, \\(\\mathrm{GeM}_j(\\cdot)\\) represents a generalized mean pooling (GeM) operation [32] for \\(j\\)-th region, and \\(\\Gamma_j^{(l)}\\) : \\(\\mathbb{R}^{C\\times T\\times H_j^{(l)}\\times W}\\mapsto \\mathbb{R}^{C\\times 1\\times H_j^{(l)}\\times W}\\mapsto \\mathbb{R}^{C^{(l)}\\times 1\\times H_j^{(l)}\\times W}\\mapsto\\) \\(\\mathbb{R}^{C^{(l)}\\times 1\\times 1\\times 1}\\), in other words, \\(Y_{\\Gamma ,j}^{(l)}\\in \\mathbb{R}^{C^{(l)}\\times 1\\times 1\\times 1}\\). Therefore, by concatenating outputs of \\(K_{l}\\) regions, we can" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19642" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.123, + 0.473, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.304, + 0.471, + 0.346 + ], + "angle": 0, + "content": "Figure 3: The detailed structure of the frame-level temporal aggregation (FTA). For simplicity, we omit the channel dimension \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.47, + 0.406 + ], + "angle": 0, + "content": "obtain \\( Y_{\\Gamma}^{(l)} = \\left[Y_{\\Gamma,1}^{(l)},Y_{\\Gamma,2}^{(l)},\\dots ,Y_{\\Gamma,K_l}^{(l)}\\right] \\), where \\( Y_{\\Gamma}^{(l)} \\in \\mathbb{R}^{C^{(l)} \\times 1 \\times K_l \\times 1} \\) is the output of ASTP at level \\( l \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.412, + 0.383, + 0.428 + ], + "angle": 0, + "content": "3.4. Frame-level Temporal Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.434, + 0.47, + 0.573 + ], + "angle": 0, + "content": "A gait sequence may contain several redundant frames due to factors such as the acquisition frame rate and pace frequency. To reduce computational costs, some methods compress a gait sequence by aggregating its local clips [27, 25]. In the proposed frame-level temporal aggregation (FTA) strategy, we consider both the gait structure and the multiscale temporal information. Given the \\(j\\)-th gait region at the \\(l\\)-th level, \\(X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}\\), we first fuse the features of the two temporal scales using the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.58, + 0.469, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {U} _ {j} ^ {(l)} = U _ {j, 1} ^ {(l)} + U _ {j, 2} ^ {(l)} \\\\ = \\operatorname {M a x} _ {3 \\times 1 \\times 1} ^ {3 \\times 1 \\times 1} \\left(X _ {j} ^ {(l)}\\right) + \\operatorname {M a x} _ {5 \\times 1 \\times 1} ^ {3 \\times 1 \\times 1} \\left(X _ {j} ^ {(l)}\\right), \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.633, + 0.47, + 0.748 + ], + "angle": 0, + "content": "where \\(\\mathrm{Max}_{3\\times 1\\times 1}^{3\\times 1\\times 1}(\\cdot)\\) and \\(\\mathrm{Max}_{5\\times 1\\times 1}^{3\\times 1\\times 1}(\\cdot)\\) denote max pooling operations with kernel sizes of \\(3\\times 1\\times 1\\) and \\(5\\times 1\\times 1\\) respectively, both with stride of \\(3\\times 1\\times 1\\). \\(\\hat{U}_j^{(l)}, U_{j,1}^{(l)}\\) and \\(U_{j,2}^{(l)}\\) have the same size of \\((C,\\frac{T}{3},H_j^{(l)},W)\\). The output of Eq.(5), \\(\\hat{U}_j^{(l)}\\), is the element-wise summation of the aggregation results of the two scales, \\(U_{j,1}^{(l)}\\) and \\(U_{j,2}^{(l)}\\), which reduces the temporal dimension of the input from \\(T\\) to \\(\\frac{T}{3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.748, + 0.469, + 0.775 + ], + "angle": 0, + "content": "Then, the FTA model produces frame-level weights, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.782, + 0.47, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Z _ {j, 1} ^ {(l)} = \\operatorname {F C} _ {j, 1} ^ {(l)} \\left(\\operatorname {G A P} \\left(\\hat {U} _ {j} ^ {(l)}\\right)\\right), \\tag {6} \\\\ Z _ {j, 2} ^ {(l)} = \\operatorname {F C} _ {j, 2} ^ {(l)} \\left(\\operatorname {G A P} \\left(\\hat {U} _ {j} ^ {(l)}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathrm{GAP}(\\cdot)\\) represents the global mean pooling along spatial dimension. \\(\\mathrm{FC}_{j,1}(\\cdot)\\) and \\(\\mathrm{FC}_{j,2}(\\cdot)\\) are two independ-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.894, + 0.188 + ], + "angle": 0, + "content": "dent fully connected layers that generate the frame selection weighting tensors, \\( Z_{j,1}^{(l)} \\) and \\( Z_{j,2}^{(l)} \\in \\mathbb{R}^{C \\times \\frac{T}{3} \\times 1 \\times 1} \\), for \\( U_{j,1}^{(l)} \\) and \\( U_{j,2}^{(l)} \\), respectively. The weights are further normalized across the two scales, which can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.196, + 0.892, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {j, s, c, t} ^ {(l)} = \\frac {e ^ {Z _ {j , s , c , t} ^ {(l)}}}{e ^ {Z _ {j , 1 , c , t} ^ {(l)}} + e ^ {Z _ {j , 2 , c , t} ^ {(l)}}} \\quad s \\in \\{1, 2 \\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.244, + 0.894, + 0.306 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}_{j,s,c,t}^{(l)}\\in \\mathbb{R}^{1\\times 1\\times 1\\times 1}\\) is the weight value of the \\(c\\) -th channel of the \\(t\\) -th frame. Combining Eq. (5) and Eq. (7), the \\(j\\) -th output region feature \\(Y_{\\Omega ,j}^{(l)}\\in \\mathbb{R}^{C^{(l)}\\times \\frac{T}{3}\\times H_j^{(l)}\\times W}\\) for the \\(l\\) -th level of FTA can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.313, + 0.892, + 0.334 + ], + "angle": 0, + "content": "\\[\nY _ {\\Omega , j} ^ {(l)} = \\mathcal {W} _ {j, 1} ^ {(l)} \\odot U _ {j, 1} ^ {(l)} + \\mathcal {W} _ {j, 2} ^ {(l)} \\odot U _ {j, 2} ^ {(l)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.343, + 0.893, + 0.428 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}_{j,1}^{(l)},\\mathcal{W}_{j,2}^{(l)}\\in \\mathbb{R}^{C\\times \\frac{T}{3}\\times 1\\times 1}\\) are two weight tensors calculated using Eq. (5), and \\(\\odot\\) represents element-wise multiplication operation. The FTA module outputs \\(Y_{\\Omega}^{(l)}\\in\\) \\(\\mathbb{R}^{C\\times \\frac{T}{3}\\times H\\times W}\\) by concatenating the \\(K_{l}\\) gait regions of level \\(l\\) where \\(Y_{\\Omega}^{(l)} = \\left[Y_{\\Omega ,1}^{(l)},Y_{\\Omega ,2}^{(l)},\\ldots ,Y_{\\Omega ,K_l}^{(l)}\\right]\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.634, + 0.452 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.459, + 0.8, + 0.472 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.48, + 0.892, + 0.632 + ], + "angle": 0, + "content": "CASIA-B. The CASIA-B [52] dataset is a widely used benchmark for gait recognition. It contains video sequences of 124 subjects with 11 different views and three walking conditions (normal walking (NM), walking with a bag (BG), and walking with a coat (CL)). Our study follows the protocol outlined in previous works [4, 8, 27, 25, 15, 16]. The first 74 subjects are used for training, and the remaining 50 subjects are used for testing. During testing, the first four sequences under NM (NM#01-04) are regarded as the gallery set, and the rest (NM#05-06, BG#01-02, CL#01-02) are regarded as the probe set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.756 + ], + "angle": 0, + "content": "OUMVLP. The OUMVLP [39] is one of the largest gait datasets, containing silhouette sequences of 10,307 subjects. Each subject has a single normal walking condition (NM) with 14 views. According to the protocol provided by the dataset, the first 5,153 subjects are used for training while the remaining 5,154 subjects are used for testing. During the testing phase, the sequences of NM#01 are assigned to the gallery set, and the sequences of NM#02 are considered as the probe set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.867 + ], + "angle": 0, + "content": "GREW. GREW [60] is the first large-scale dataset for gait recognition in the wild, consisting of 128,671 sequences from 26,345 individuals captured by 882 cameras. It includes four data types: four data types: silhouettes, optical flow, 2D pose, and 3D pose. The dataset is divided into a training set with 20,000 subjects and 102,887 sequences, and a testing set with 6,000 subjects and 24,000 sequences. In the testing phase, each subject has two sequences for the" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19643" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.124, + 0.894, + 0.154 + ], + "angle": 0, + "content": "Table 1: Rank-1 accuracy (%) on CASIA-B under all views and different conditions, excluding identical-view cases. Std denotes the performance sample standard deviation across 11 views." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.164, + 0.893, + 0.51 + ], + "angle": 0, + "content": "
Gallery NM #1-40° - 180°MeanStd
Probe0°18°36°54°72°90°108°126°144°162°180°
NM #5-6GaitSet [4]90.897.999.496.993.691.795.097.898.996.885.895.03.5
GaitPart [8]94.198.699.398.594.092.395.998.499.297.890.496.23.1
3D Local [16]96.099.099.598.997.194.296.399.098.898.595.297.51.8
CSTL [15]97.299.099.298.196.295.597.798.799.298.996.597.81.3
GaitGL [27]96.098.399.097.996.995.497.098.999.398.894.097.41.7
LagrangeGait [3]95.798.199.198.396.495.297.599.099.398.994.997.51.6
MetaGait [5]97.399.299.599.197.295.597.699.199.399.196.798.11.3
Ours97.698.099.698.297.496.597.999.399.498.497.098.11.0
BG #1-2GaitSet [4]83.891.291.888.883.381.084.190.092.294.479.087.24.9
GaitPart [8]89.194.896.795.188.384.989.093.596.193.885.891.54.2
3D Local [16]92.995.997.896.293.087.892.796.397.998.088.594.33.5
CSTL [15]91.796.597.095.490.988.091.595.897.095.590.393.63.0
GaitGL [27]92.696.696.895.593.589.392.296.598.296.991.594.52.8
LagrangeGait [3]94.296.296.895.894.389.591.796.898.097.090.994.62.7
MetaGait [5]92.996.797.196.494.790.492.997.298.598.192.395.22.6
Ours95.096.597.396.695.393.394.696.898.697.792.995.91.7
CL #1-2GaitSet [4]61.475.480.777.372.170.171.573.573.568.450.070.48.0
GaitPart [8]70.785.586.983.377.172.576.982.283.880.266.578.76.6
3D Local [16]78.290.292.087.183.076.883.186.686.884.170.983.76.2
CSTL [15]78.189.491.686.682.179.981.886.388.786.675.384.24.9
GaitGL [27]76.690.090.387.184.579.084.187.087.384.469.583.66.3
LagrangeGait [3]77.490.693.290.284.780.385.287.789.386.671.085.16.3
MetaGait [5]80.091.893.087.886.582.985.290.090.889.378.486.94.6
Ours82.494.295.091.788.283.388.092.393.191.078.588.95.1
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.534, + 0.47, + 0.576 + ], + "angle": 0, + "content": "gallery set and two for the probe set. The GREW dataset also includes a distractor set, which contains 233,857 unlabeled sequences." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.47, + 0.77 + ], + "angle": 0, + "content": "Gait3D. The Gait3D dataset [59] is a newly proposed dataset for gait recognition in uncontrolled indoor environments, particularly in large supermarkets. It contains 25,309 sequences of 4,000 subjects extracted from 39 cameras, with 18,940 sequences from 3,000 subjects for training and 6,369 sequences from 1,000 subjects for testing. The dataset mainly includes four data types: silhouettes, 2D pose, 3D pose, and 3D mesh. During testing, one sequence per subject is used as the probe set, while the remaining sequences are used as the gallery set. To evaluate the model's performance, we use accuracy as well as mean average precision (mAP) and mean inverse negative penalty (mINP) [49], which consider multiple instances and hard sample recall." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.776, + 0.295, + 0.791 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.798, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Training details. 1) In our implementation, the margin \\( m \\) of the triplet loss is set to 0.2, and the parameter \\( p \\) of the GeM function used in the ASTP module is initialized to 6.5. In the hierarchy of gait motion, the number of partitions \\( k \\) at the bottom level is set to 8. 2) The batch size is set to (8,8)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.616 + ], + "angle": 0, + "content": "Table 2: The detailed architecture of the proposed HSTL on CASIA-B. The first column denotes the levels of the gait hierarchy and \\( K_{l} \\) is the number of groups at level \\( l \\). \\( C_{in} \\) and \\( C_{out} \\) represent the input channel and output channel of each layer respectively. The body parts are indexed in spatial order from top to bottom, numbered 1 to 8." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.627, + 0.892, + 0.807 + ], + "angle": 0, + "content": "
LevelBlockLayer\\( {C}_{in} \\)\\( {C}_{out} \\)Kernel\\( {K}_{l} \\)Parts Grouping
1ARMEConv3d132(3,3,3)1\\( \\{ \\{ 1,2,3,4,5,6,7,8\\} \\} \\)
ASTP
2ARMEConv3d3232(3,3,3)2\\( \\{ \\{ 1,2,3,4,5\\} ,\\{ 6,7,8\\} \\} \\)
Conv3d3264(3,3,3)
ASTP
2FTAMaxPool6464(3,1,1)2\\( \\{ \\{ 1,2,3,4,5\\} ,\\{ 6,7,8\\} \\} \\)
(5,1,1)
ASTP
3ARMEConv3d64128(3,3,3)4\\( \\{ \\{ 1\\} ,\\{ 2,3,4,5\\} ,\\{ 6,7\\} ,\\{ 8\\} \\} \\)
Conv3d128128(3,3,3)
ASTP
4ASTP8\\( \\{ \\{ 1\\} ,\\{ 2\\} ,\\{ 3\\} ,\\{ 4\\} , \\) \\( \\{ 5\\} ,\\{ 6\\} ,\\{ 7\\} ,\\{ 8\\} \\} \\)
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.893, + 0.868 + ], + "angle": 0, + "content": "for CASIA-B, (32,8) for OUMVLP, (32,4) for GREW, and (32,4) for Gait3D. 3) We use gait silhouettes as the input" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19644" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.124, + 0.894, + 0.154 + ], + "angle": 0, + "content": "Table 3: Rank-1 accuracy (%) on OUMVLP under all views, excluding identical-view cases. Std denotes the performance sample standard deviation across 14 views." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.163, + 0.894, + 0.306 + ], + "angle": 0, + "content": "
MethodProbe ViewMeanStd
0°15°30°45°60°75°90°180°195°210°225°240°255°270°
GaitSet [4]79.387.990.090.188.088.787.781.886.589.089.287.287.686.287.14.0
GaitPart [8]82.688.990.891.089.789.989.585.288.190.090.189.089.188.288.72.3
GLN [12]83.890.091.091.290.390.089.485.389.190.590.689.689.388.589.22.1
CSTL [15]87.191.091.591.890.690.890.689.490.290.590.789.890.089.490.21.1
GaitGL [27]84.990.291.191.591.190.890.388.588.690.390.489.689.588.889.71.7
3D Local [16]86.191.292.692.992.291.391.186.990.892.292.391.391.190.290.92.0
LagrangeGait [3]85.990.691.391.591.291.090.688.989.290.590.689.989.889.290.01.4
MetaGait [5]88.292.393.093.593.192.792.689.391.292.092.692.391.991.191.91.4
Ours91.492.992.793.092.992.592.592.792.392.192.392.291.891.892.40.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.519 + ], + "angle": 0, + "content": "modality. During the training phase, we sample 30 frames following the strategy proposed in [8], while during testing, all frames are fed into the model. Moreover, we align each frame following the strategy presented in [39], and the input image size is cropped to \\(64 \\times 44\\) for all datasets. 4) For CASIA-B, the optimizer used is Adam with a weight decay of \\(5e-4\\). The model is trained for 100K iterations with an initial learning rate (LR) of \\(1e-5\\), and the LR is multiplied by 0.1 at 70K iterations. For OUMVLP, GREW, and Gait3D, the model is trained for 250K, 250K, and 210K iterations, respectively, with an initial LR of 0.1. The LR is multiplied by 0.1 at 150K and 200K iterations. The optimizer used for these datasets is SGD with a weight decay of \\(5e-4\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Architecture details. Table 2 presents the detailed architecture of the model used for the CASIA-B dataset. To handle datasets with more subjects, such as OUMVLP, GREW, and Gait3D, we incorporate the label smoothing operation into the cross-entropy loss function, and deepen the network by adding an extra ARME module to the third level of the hierarchy. The output channels for the four ARMEs are set to 64, 64, 128 and 256, respectively. Additionally, we include a layer of spatial downsampling after the first ARME to improve the training efficiency." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.669, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Table 4: Rank-1 accuracy (\\%), Rank-5 accuracy (\\%), Rank-10 accuracy (\\%), Rank-20 accuracy (\\%) on GREW." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.707, + 0.47, + 0.798 + ], + "angle": 0, + "content": "
MethodsRank-1Rank-5Rank-10Rank-20
GaitSet [4]46.2863.5870.2676.82
GaitPart [8]44.0160.6867.2573.47
GaitGL [27]47.2863.5669.3274.18
MTSGait [58]55.3271.2876.8581.55
Ours62.7276.5781.3285.24
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.818, + 0.446, + 0.833 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.839, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Evaluation on CASIA-B. Table 1 compares the performance of the proposed HSTL with seven state-of-the-art" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.326, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Table 5: Rank-1 accuracy \\((\\%)\\), Rank-5 accuracy \\((\\%)\\), mAP \\((\\%)\\) and mINP on Gait3D." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.363, + 0.893, + 0.503 + ], + "angle": 0, + "content": "
MethodsRank-1Rank-5mAPmINP
GaitSet [4]36.7058.3030.0117.30
GaitPart [8]28.2047.6021.5812.36
GLN [12]31.4052.9024.7413.58
GaitGL [27]29.7048.5022.2913.26
CSTL [15]11.7019.205.592.59
SMPLGait [59]46.3064.5037.1622.23
MTSGait [58]48.7067.1037.6321.92
Ours61.3076.3055.4834.77
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.868 + ], + "angle": 0, + "content": "methods on the CASIA-B dataset. It can be seen that the proposed method obtains the best results in all three walking conditions while maintaining considerable stability across different views. The experimental results reveal that 1) the mean accuracy of the proposed method for the BG and CL walking conditions is \\(95.9\\%\\) and \\(88.9\\%\\), respectively, which are higher by \\(0.7\\%\\) and \\(2.0\\%\\) than the second-best method (MetaGait [5]), demonstrating the advantage of our method in cross-view gait recognition. 2) The decrease in accuracy from NM to CL is \\(9.2\\%\\) for our method, compared to \\(12.4\\%\\) for 3D Local [16]. This indicates that the ARME module is more adaptable to various walking conditions. In addition, our method achieves an accuracy of \\(88.9\\%\\) in the challenging CL condition, which is \\(5.2\\%\\) higher than 3D Local. This may be due to the complex clothing conditions that affect the part localization accuracy in 3D Local. 3) Both CSTL [15] and our method extracts multi-scale temporal features but in different ways. CSTL first extracts spatial information and then fuses three scales of motion features. In contrast, we propose an FTA module to aggregate spatiotemporal information from multiple body regions. Therefore, the average rank-1 accuracy of our method is higher or equivalent to that of CSTL in all views. This suggests that FTA can be more adaptive to spatio-temporal changes." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19645" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.123, + 0.323, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.238, + 0.277, + 0.249 + ], + "angle": 0, + "content": "(a) Average rank-1 accuracy" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.17, + 0.468, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.238, + 0.449, + 0.248 + ], + "angle": 0, + "content": "(b) Age distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.26, + 0.47, + 0.302 + ], + "angle": 0, + "content": "Figure 4: Cross-age comparison results for the OUMVLP dataset. (a) Average rank-1 accuracy in adults vs children. (b) Distribution of two age groups." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.47, + 0.602 + ], + "angle": 0, + "content": "Evaluation on OUMVLP. To verify the model's generalizability, we conduct experiments on the large-scale OUMVLP dataset. As shown in Table 3, our approach achieves competitive results in most views. In particular, the proposed method outperforms the second-best method (MetaGait) by an average of \\(2.5\\%\\) under three extreme views, i.e., \\(0^{\\circ}\\), \\(180^{\\circ}\\), and \\(270^{\\circ}\\), resulting in the best mean performance and cross-view stability. In addition, the OUMVLP dataset also provides annotations for the ages of the subjects. To evaluate the impact of age differences on recognition performance, we divided all subjects into two groups: adults (18-87 years old) and children (2-17 years old), as shown in Fig. 4(b). Fig. 4(a) presents the recognition accuracy based on age. It can be observed that, as adult sequences make up about \\(70\\%\\) of the total sequences, all the compared methods show a bias toward the recognition results of adults. However, compared to other methods, our model effectively improves the accuracy of gait recognition for children, demonstrating the effectiveness of our hierarchical gait representation across ages." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.602, + 0.469, + 0.712 + ], + "angle": 0, + "content": "Evaluation on GREW and Gait3D. Gait3D and GREW are two recently introduced datasets that contain challenging conditions, such as misalignment of the human body and partial occlusion. Tables 4 and 5 show the results of the comparison between the proposed method and the state-of-the-art methods. Our method shows superior performance in all metrics, indicating its ability to effectively model gait characteristics in realistic scenarios." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.721, + 0.231, + 0.736 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.742, + 0.469, + 0.866 + ], + "angle": 0, + "content": "Effectiveness of hierarchical feature extraction. To verify the effectiveness of our hierarchical gait partitioning, we conduct experiment with various grouping strategies. As shown in Table 2, this comparison only considers the different settings of the first three layers since a uniform partition is used at layer 4. Specifically, 1-1-1 indicates that the first three levels have the same number of groupings, and they are divided uniformly. \\(1 - 2 - 4^{*}\\) refers to the nonuniform division used in the proposed model. As shown in" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.123, + 0.875, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.29, + 0.892, + 0.331 + ], + "angle": 0, + "content": "Figure 5: Ablation study on the effectiveness of hierarchical feature extraction on the CASIA-B dataset (best viewed in color)." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.345, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Table 6: Ablation study on the effectiveness of ARME, ASTP, and FTA modules in terms of average rank-1 accuracy on the CASIA-B dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.397, + 0.892, + 0.509 + ], + "angle": 0, + "content": "
SettingARMEASTPFTANMBGCLMean
35
a97.094.584.291.9
b✓97.895.385.993.0
c✓✓97.895.487.093.4
d✓✓✓97.795.387.593.5
e✓✓✓97.795.287.293.4
f✓✓✓97.895.487.593.6
g✓✓✓✓98.195.988.994.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.601 + ], + "angle": 0, + "content": "Fig. 5, the hierarchical feature extraction setting, e.g., 1-2-4, outperforms the other non-hierarchical approaches. A further mean performance improvement of \\(1.0\\%\\) is achieved when the motion relationships between different body regions, i.e., 1-2-4*, are considered." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.602, + 0.892, + 0.755 + ], + "angle": 0, + "content": "Effectiveness of ARME, ASTP, and FTA. The results of the ablation experiments for ARME, ASTP, and FTA are shown in Tab. 6 The results indicate that the ARME module significantly contributes to the improvement of the recognition accuracy, with an average improvement of \\(1.1\\%\\) compared to the baseline model (non-grouping version of ARME). The mean accuracy is further improved by \\(1.3\\%\\) when the ASTP module is integrated and multi-scale fusion is utilized in FTA. These results demonstrate the effectiveness and complementarity of the three modules in the proposed gait recognition framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.859, + 0.778 + ], + "angle": 0, + "content": "4.5. Trade-off between accuracy and efficiency" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.784, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In this subsection, we evaluate the relationship between accuracy and efficiency for each compared method. As shown in Fig. 6, the 3D convolution-based approaches, such as GaitGL [27], 3D Local [16], LagrangeGait [3] and MetaGait [5], outperform the 2D convolution-based methods, like GaitSet [4] and GaitPart [8], in terms of accuracy, but" + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.898 + ], + "angle": 0, + "content": "19646" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.124, + 0.473, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.32, + 0.472, + 0.362 + ], + "angle": 0, + "content": "Figure 6: The trade-off between accuracy and FLOPs of our method and other comparison methods on the OUMVLP dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.387, + 0.47, + 0.485 + ], + "angle": 0, + "content": "at the cost of a significant increase in FLOPs (floating point operations). Our approach has a better trade-off between accuracy and efficiency. The main reason is that the proposed hierarchical learning architecture can extract multilevel motion features while reducing the number of 3D convolutions. More experimental results and ablation analysis are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.497, + 0.196, + 0.51 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.473, + 0.7 + ], + "angle": 0, + "content": "This paper presents a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL stacks multiple adaptive region-based motion extractors (ARMEs) and learns walking patterns in a coarse-to-fine manner. An adaptive spatio-temporal pooling (ASTP) module is proposed to perform hierarchical feature mapping for the output of each level of ARME. Additionally, a frame-level temporal aggregation module (FTA) is designed to compress local clips by fusing temporal information. The effectiveness of the proposed HSTL framework is demonstrated through extensive experiments conducted on four public datasets (CASIA-B, OUMVLP, GREW, and Gait3D)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.245, + 0.728 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.736, + 0.47, + 0.806 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 61972132, 62106065) and the Research Project for Self-cultivating Talents of Hebei Agricultural University (Grant No. PY201810)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.818, + 0.175, + 0.832 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.84, + 0.47, + 0.867 + ], + "angle": 0, + "content": "[1] Weizhi An, Shiqi Yu, Yasushi Makihara, Xinhui Wu, Chi Xu, Yang Yu, Rijun Liao, and Yasushi Yagi. Performance evalua" + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.893, + 0.165 + ], + "angle": 0, + "content": "tion of model-based gait on multi-view very large population database with pose sequences. IEEE TBIOM, 2(4):421-430, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.166, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[2] Pia Bideau, Aruni RoyChowdhury, Rakesh R Menon, and Erik Learned-Miller. The best of both worlds: Combining cnns and geometric constraints for hierarchical motion segmentation. In CVPR, pages 508-517, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.218, + 0.895, + 0.267 + ], + "angle": 0, + "content": "[3] Tianrui Chai, Annan Li, Shaoxiong Zhang, Zilong Li, and Yunhong Wang. Lagrange motion analysis and view embeddings for improved gait recognition. In CVPR, pages 20249-20258, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.268, + 0.895, + 0.317 + ], + "angle": 0, + "content": "[4] Hanqing Chao, Yiwei He, Junping Zhang, and Jianfeng Feng. Gaitset: Regarding gait as a set for cross-view gait recognition. In AAAI, pages 8126-8133, 2019. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.318, + 0.895, + 0.37 + ], + "angle": 0, + "content": "[5] Huanzhang Dou, Pengyi Zhang, Wei Su, Yunlong Yu, and Xi Li. Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In ECCV, pages 357-374. Springer, 2022. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.37, + 0.895, + 0.408 + ], + "angle": 0, + "content": "[6] Jessica Maria Echterhoff, Juan Haladjian, and Bernd Brügge. Gait and jump classification in modern equestrian sports. In ISWC, pages 88-91, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.409, + 0.894, + 0.458 + ], + "angle": 0, + "content": "[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.459, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[8] Chao Fan, Yunjie Peng, Chunshui Cao, Xu Liu, Saihui Hou, Jiannan Chi, Yongzhen Huang, Qing Li, and Zhiqiang He. Gaitpart: Temporal part-based model for gait recognition. In CVPR, pages 14225-14233, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.51, + 0.893, + 0.56 + ], + "angle": 0, + "content": "[9] Yang Fu, Yunchao Wei, Yuqian Zhou, Honghui Shi, Gao Huang, Xinchao Wang, Zhiqiang Yao, and Thomas Huang. Horizontal pyramid matching for person re-identification. In AAAI, pages 8295–8302, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.561, + 0.893, + 0.586 + ], + "angle": 0, + "content": "[10] Jinguang Han and Bir Bhanu. Individual recognition using gait energy image. IEEE TPAMI, 28(2):316-322, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.893, + 0.637 + ], + "angle": 0, + "content": "[11] Ruibing Hou, Hong Chang, Bingpeng Ma, Rui Huang, and Shiguang Shan. Bicnet-tks: Learning efficient spatial-temporal representation for video person re-identification. In CVPR, pages 2014–2023, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[12] Saihui Hou, Chunshui Cao, Xu Liu, and Yongzhen Huang. Gait lateral network: Learning discriminative and compact representations for gait recognition. In ECCV, pages 382-398, 2020. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.893, + 0.726 + ], + "angle": 0, + "content": "[13] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Set residual network for silhouette-based gait recognition. IEEE TBIOM, 3(3):384-393, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.727, + 0.894, + 0.765 + ], + "angle": 0, + "content": "[14] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Gait quality aware network: Toward the interpretability of silhouette-based gait recognition. IEEE TNNLS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.766, + 0.893, + 0.816 + ], + "angle": 0, + "content": "[15] Xiaohu Huang, Duowang Zhu, Hao Wang, Xinggang Wang, Bo Yang, Botao He, Wenyu Liu, and Bin Feng. Context-sensitive temporal feature learning for gait recognition. In ICCV, pages 12909-12918, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.866 + ], + "angle": 0, + "content": "[16] Zhen Huang, Dixiu Xue, Xu Shen, Xinmei Tian, Houqiang Li, Jianqiang Huang, and Xian-Sheng Hua. 3d local convolutional neural networks for gait recognition. In ICCV, pages 14920-14929, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.127, + 0.895, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "19647" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[17] Munish Kumar, Navdeep Singh, Ravinder Kumar, Shubham Goel, and Krishan Kumar. Gait recognition based on vision systems: A systematic survey. Journal of Visual Communication and Image Representation, 75:103052, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.18, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[18] Worapan Kusakunniran. Review of gait recognition approaches and their challenges on view changes. IET Biometrics, 9(6):238-250, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.47, + 0.269 + ], + "angle": 0, + "content": "[19] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE TIP, 31:5949-5962, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.272, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[20] Xiang Li, Yasushi Makihara, Chi Xu, and Yasushi Yagi. End-to-end model-based gait recognition using synchronized multi-view pose constraint. In ICCVW, pages 4106-4115, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.325, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[21] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, and Mingwu Ren. Gait recognition via semi-supervised disentangled representation learning to identity and covariate features. In CVPR, pages 13309-13319, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[22] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, Shiqi Yu, and Mingwu Ren. End-to-end model-based gait recognition. In ACCV, pages 3-20, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.417, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[23] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In CVPR, pages 510-519, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.444, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[24] Junhao Liang, Chao Fan, Saihui Hou, Chuanfu Shen, Yongzhen Huang, and Shiqi Yu. Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In ECCV, pages 375-390. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.497, + 0.469, + 0.534 + ], + "angle": 0, + "content": "[25] Beibei Lin, Yu Liu, and Shunli Zhang. Gaitmask: Mask-based model for gait recognition. In BMVC, page 363, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.537, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[26] Beibei Lin, Shunli Zhang, and Feng Bao. Gait recognition with multiple-temporal-scale 3d convolutional neural network. In ACM MM, pages 3054-3062, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.577, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[27] Beibei Lin, Shunli Zhang, and Xin Yu. Gait recognition via effective global-local feature representation and local temporal aggregation. In ICCV, pages 14648-14656, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.629, + 0.469, + 0.654 + ], + "angle": 0, + "content": "[28] Nini Liu and Yap-Peng Tan. View invariant gait recognition. In ICASSP, pages 1410-1413, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.656, + 0.469, + 0.694 + ], + "angle": 0, + "content": "[29] Yasushi Makihara, Mark S Nixon, and Yasushi Yagi. Gait recognition: Databases, representations, and applications. Computer Vision: A Reference Guide, pages 1-13, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.696, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[30] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical gaussian descriptor for person re-identification. In CVPR, pages 1363–1372, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.736, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[31] Daigo Muramatsu, Yasushi Makihara, Haruyuki Iwama, Takuya Tanoue, and Yasushi Yagi. Gait verification system for supporting criminal investigation. In IAPR, pages 747-748, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.826 + ], + "angle": 0, + "content": "[32] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 41(7):1655-1668, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.828, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[33] Imad Rida, Noor Almaadeed, and Somaya Almaadeed. Robust gait recognition: a comprehensive survey. IET Biometrics, 8(1):14-28, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.153 + ], + "angle": 0, + "content": "[34] Alireza Sepas-Moghaddam and Ali Etemad. Deep gait recognition: A survey. IEEE TPAMI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[35] Kohei Shiraga, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Geinet: View-invariant gait recognition using a convolutional neural network. In ICB, pages 1-8, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[36] Pranjay Shyam, Kuk-Jin Yoon, and Kyung-Soo Kim. Adversarily-trained hierarchical feature extractor for vehicle re-identification. In ICRA, pages 13400–13407, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.284 + ], + "angle": 0, + "content": "[37] Chunfeng Song, Yongzhen Huang, Yan Huang, Ning Jia, and Liang Wang. Gaitnet: An end-to-end network for gait based human identification. PR, 96:106988, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.287, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[38] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[39] Noriko Takemura, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ TCVA, 10:1-14, 2018. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[40] Lei Tan, Pingyang Dai, Rongrong Ji, and Yongjian Wu. Dynamic prototype mask for occluded person re-identification. In ACM MM, pages 531–540, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[41] Torben Teepe, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Towards a deeper understanding of skeleton-based gait recognition. In CVPRW, pages 1569-1577, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.892, + 0.535 + ], + "angle": 0, + "content": "[42] Torben Teepe, Ali Khan, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In ICIP, pages 2314-2318, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.538, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[43] Md Uddin, Daigo Muramatsu, Noriko Takemura, Md Ahad, Atiqur Rahman, Yasushi Yagi, et al. Spatio-temporal silhouette sequence reconstruction for gait recognition against occlusion. IPSJ TCVA, 11(1):1-18, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[45] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. IEEE TCSVT, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.67, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[46] Xiu-Shen Wei, Chen-Lin Zhang, Lingqiao Liu, Chunhua Shen, and Jianxin Wu. Coarse-to-fine: A rnn-based hierarchical attention model for vehicle re-identification. In ACCV, pages 575–591, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.723, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[47] Jiachen Yang, Jianxiong Zhou, Dayong Fan, and Haibin Lv. Design of intelligent recognition system based on gait recognition technology in smart transportation. Multimedia Tools and Applications, 75(24):17501-17514, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[48] Lingxiang Yao, Worapan Kusakunniran, Qiang Wu, Jingsong Xu, and Jian Zhang. Collaborative feature learning for gait recognition under cloth changes. IEEE TCSVT, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.866 + ], + "angle": 0, + "content": "[49] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6):2872-2893, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19648" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.166 + ], + "angle": 0, + "content": "[50] TzeWei Yeoh, Hernan E Aguirre, and Kiyoshi Tanaka. Clothing-invariant gait recognition using convolutional neural network. In ISPACS, pages 1-5, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[51] Shiqi Yu, Haifeng Chen, Qing Wang, Linlin Shen, and Yongzhen Huang. Invariant feature extraction for gait recognition using only one uniform model. Neurocomputing, 239:81-93, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[52] Shiqi Yu, Daoliang Tan, and Tieniu Tan. A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In ICPR, pages 441-444, 2006. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.27, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[53] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[54] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 17:516-530, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.374, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[55] Shaoxiong Zhang, Yunhong Wang, and Annan Li. Crossview gait recognition with deep universal linear embeddings. In CVPR, pages 9095-9104, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[56] Yuqi Zhang, Yongzhen Huang, Shiqi Yu, and Liang Wang. Cross-view gait recognition by discriminative feature learning. IEEE TIP, 29:1001-1015, 2019. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.452, + 0.47, + 0.501 + ], + "angle": 0, + "content": "[57] Ziyuan Zhang, Luan Tran, Xi Yin, Yousef Atoum, Xiaoming Liu, Jian Wan, and Nanxin Wang. Gait recognition via disentangled representation learning. In CVPR, pages 4710-4719, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.47, + 0.553 + ], + "angle": 0, + "content": "[58] Jinkai Zheng, Xinchen Liu, Xiaoyan Gu, Yaoqi Sun, Chuang Gan, Jiyong Zhang, Wu Liu, and Chenggang Yan. Gait recognition in the wild with multi-hop temporal switch. In ACM MM, pages 6136-6145, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[59] Jinkai Zheng, Xinchen Liu, Wu Liu, Lingxiao He, Chenggang Yan, and Tao Mei. Gait recognition in the wild with dense 3d representations and a benchmark. In CVPR, pages 20228-20237, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[60] Zheng Zhu, Xianda Guo, Tian Yang, Junjie Huang, Jiankang Deng, Guan Huang, Dalong Du, Jiwen Lu, and Jie Zhou. Gait recognition in the wild: A benchmark. In ICCV, pages 14789-14799, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.468, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "19649" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_origin.pdf b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8331e27fcf82d1dcae5282565c9f8c31a91d970c --- /dev/null +++ b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/6240fae9-3133-48a9-86fe-481f88ab16fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2639d14283cfb236ef1c36bf0e10f86af85c7b1dba44d21b2657faf57dcabb61 +size 2999509 diff --git a/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/full.md b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b66702bf1b68a6a556bdc39d5e9837c15c67135 --- /dev/null +++ b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/full.md @@ -0,0 +1,305 @@ +# Hierarchical Spatio-Temporal Representation Learning for Gait Recognition + +Lei Wang $^{1,2}$ , Bo Liu $^{1,2*}$ , Fangfang Liang $^{1,2}$ , Bincheng Wang $^{1,2}$ + +1 College of Information Science and Technology, Hebei Agricultural University, China +2 Hebei Key Laboratory of Agricultural Big Data, China + +{20212060107, 20212060108}@pgs.hebau.edu.cn, {boliu, liangfangfang}@hebau.edu.cn + +# Abstract + +Gait recognition is a biometric technique that identifies individuals by their unique walking styles, which is suitable for unconstrained environments and has a wide range of applications. While current methods focus on exploiting body part-based representations, they often neglect the hierarchical dependencies between local motion patterns. In this paper, we propose a hierarchical spatio-temporal representation learning (HSTL) framework for extracting gait features from coarse to fine. Our framework starts with a hierarchical clustering analysis to recover multilevel body structures from the whole body to local details. Next, an adaptive region-based motion extractor (ARME) is designed to learn region-independent motion features. The proposed HSTL then stacks multiple ARMEs in a top-down manner, with each ARME corresponding to a specific partition level of the hierarchy. An adaptive spatiotemporal pooling (ASTP) module is used to capture gait features at different levels of detail to perform hierarchical feature mapping. Finally, a frame-level temporal aggregation (FTA) module is employed to reduce redundant information in gait sequences through multi-scale temporal downsampling. Extensive experiments on CASIA-B, OUMVLP, GREW, and Gait3D datasets demonstrate that our method outperforms the state-of-the-art while maintaining a reasonable balance between model accuracy and complexity. Code is available at: https://github.com/gudaochangsheng/HSTL. + +# 1. Introduction + +Unlike other biometric technologies such as fingerprint, iris, and face, human gait can be captured at a distance without subject cooperation [34]. By evaluating individual-specific walking patterns, gait recognition has been applied in a variety of fields, including criminal investigations [31, 29], sports science [17, 6], and smart transportation + +![](images/a5c08c72dcfa0c9537bddd60dc16b987cfe8a966154eafa67b2997ed616ebac1.jpg) +Figure 1: The motivation for our approach. Left: the motion distributions of body parts. The same color indicates the same spatial location across multiple gait sequences in the CASIA-B dataset. Right: an example hierarchy of body-part sequences where $T$ denotes the temporal dimension of the sequence. + +[47]. However, the recognition can be challenging due to large variations in viewpoint [28, 18], occlusion [33, 43], and wearing [50, 48]. + +To address these issues, various approaches have been proposed for extracting gait features from silhouette sequences [4, 25, 16, 15, 55], 3D human structures [1, 22, 42, 59, 20], or gait templates [10, 35, 51]. Silhouette-based gait recognition methods have gained increasing attention due to the ease of obtaining silhouettes from raw videos while preserving essential temporal information. The alignment of the input silhouette makes it possible for some methods to extract local body features by horizontally slicing the silhouette image [56] or intermediate-layer features [8, 27]. This partitioning strategy, first introduced in person re-identification (ReID) [38], has been proven to be effective for gait recognition [4, 8, 12, 3]. + +However, the main limitation of the above part-based approaches is that they do not consider the hierarchical nature of local body movements [2]. For instance, within a gait cycle, the feet and lower body have distinct motion characteristics. Therefore, it is important to treat these body regions separately and investigate their part-whole relationships. Our motivation stems from the examination of body-part-specific motion clues. Specifically, each raw gait se + +quence in the CASIA-B [52] dataset is uniformly divided into eight part sequences along the body axis, so that each division roughly match a particular body part. The distributions of all body parts are shown in the left part of Fig. 1. Observably, some parts, e.g., the head and feet, are easily separated owing to their large changes in walking kinematics. Whereas other parts, such as the thighs and calves, overlap due to the strong motion correlations between them. Further, to identify the relational structure among the part sequences, a hierarchical clustering analysis [7] is performed. The results are shown in the right part of Fig. 1, indicating that the semantic body regions can be captured in the higher clustering levels without precise localization of the body parts. + +Following the above findings, we propose a novel hierarchical spatio-temporal representation learning (HSTL) framework for gait representation. The HSTL framework consists of multiple adaptive region-based motion extractor (ARME) modules, which are stacked to learn hierarchical motion patterns implied in a gait sequence (as shown in Fig.1). In the ARME module, to account for inter-regional differences, non-shared 3D convolutions are used in correspondence with individual body regions. There regions are pre-identified by a hierarchical clustering process performed on fixed horizontal partitions, allowing each body region to cover one or more body parts. Consequently, the deeper the ARME is, the more local features it tends to extract. Moreover, an adaptive spatio-temporal pooling (ASTP) module is proposed, which couples with an ARME module on the corresponding level to obtain hierarchical gait embeddings. + +In addition, changes in gait speed or sampling frequency may result in several redundant frames in a gait sequence. Although several temporal fusion strategies have been proposed, they lose spatial information [8, 15] or lack adaptability [27, 25]. To address this issue, we propose a frame-level temporal aggregation strategy (FTA). FTA fuses temporal features at multiple time steps, preserving significant motion information while compressing the sequence length. The main contributions of this paper are summarized as follows. + +1. We propose a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL takes into account the dependencies of body regions in gait motions, ensuring simplicity and scalability of the architectural design. +2. We introduce an adaptive region-based motion extractor (ARME) module to learn region-independent spatio-temporal representation for gait sequences, an adaptive spatio-temporal pooling (ASTP) module to + +perform hierarchical feature mapping, and a frame-level temporal aggregation (FTA) strategy to compress a gait sequence by removing redundant frames. + +3. Extensive experiments on the widely used gait datasets CASIA-B [52], including OUMVLP [39], GREW [60] and Gait3D [59], demonstrate that our method achieves state-of-the-art performance while offering a suitable trade-off between model accuracy and complexity. + +# 2. Related Work + +# 2.1. Gait Recognition + +Deep learning-based gait recognition methods can be broadly categorized into two categories: model-based and appearance-based. Model-based approaches extract structure and motion information from gait videos with the aid of pose estimation [1, 22, 42, 20, 41, 59]. Although these methods are robust to changes in viewpoint and appearance, they are sensitive to the accuracy of the pose parameters, making them incapable of handling low-resolution data. On the other hand, appearance-based approaches learn the feature representation from raw videos [57, 37, 24], or binary silhouette sequences [10, 35, 51, 4, 12, 8, 13, 27, 16, 14, 3, 5], which offer greater flexibility than model-based approaches. Our proposed method belongs to the family of appearance-based methods and uses silhouette sequences as inputs. + +# 2.2. Hierarchical Model + +Hierarchical feature representation has been successfully applied to a wide range of vision tasks. Here, we provide a brief review of the hierarchical object ReID approaches related to gait recognition. + +In person ReID, some approaches [30, 54, 45, 40, 53] hierarchically learn local descriptions and aggregate appearance features at different levels. For example, Matsukawa et al. [30] described an image patch via hierarchical Gaussian distribution. Zhang et al. [53] proposed a framework to learn coarse-grained and fine-grained features according to body structure. To solve the occlusion problem, Tan et al. [40] devised a hierarchical mask generator to learn from both occluded and holistic joint images. In vehicle ReID, some approaches [46, 36, 19] extract features from vehicle images in a hierarchical manner. For instance, Wei et al. [46] proposed an RNN-based module for extracting latent cues from the model level to the vehicle level. Shyam et al. [36] developed an attention-based hierarchical feature extractor. In addition, Li et al. [19] proposed a global structural embedding module for investigating hierarchical relationships between vehicle characteristics by incorporating attribute and state information. + +![](images/01ddaa05a50ac8566503c24c507354c4f738ba5c94c89ee856cb3c177c889603.jpg) +Figure 2: The framework of HFSL. It mainly consists of three modules: ARME (adaptive region-based motion extractor), ASTP (adaptive spatio-temporal pooling), and FTA (frame-level temporal aggregation). During pre-processing, a hierarchy of walking is obtained to guide the architectural design of HFSL. The framework uses multiple ARMEs to extract gait features from the entire body to individual regions. The ASTP module performs hierarchical feature mapping for the output of each level of ARME. The FTA module compresses local clips of each gait sequence to reduce the number of redundant frames. $T$ , $H$ and $W$ denote dimensions of the feature maps. © represents the concatenation operation. + +For gait recognition, fusing features of multiple granularities can improve performance [56, 8, 27, 25, 16, 3]. In particular, CSTL [15] proposed a temporal modeling network that integrates multi-scale temporal features adaptively. By combining part-level and sequence-level features, GaitPart [8] obtained a part-independent spatio-temporal expression. Additionally, GaitGL [27] considered both full body-based and part-based information to achieve discriminative feature learning. Instead of equally dividing the feature maps, in 3D Local [16], a localization operation was developed to find 3D volumes of body parts in a sequence. However, most existing gait recognition methods do not sufficiently exploit the hierarchical dependencies among body parts during walking. In this paper, the proposed HSTL performs a coarse-to-fine hierarchical strategy that integrates multi-level motion patterns from gait sequences. + +# 2.3. Temporal Model + +Temporal cues play a crucial role in gait recognition due to the periodic changes in body shape. Previous methods treat a gait sequence as an unordered set, either compressing it into a single gait template during preprocessing [35, 21] or learning order-independent gait representations from silhouette sets [4, 12, 13]. These methods assume that different subjects share similar global gait patterns, making or + +dering inputs unnecessary for gait assessment. However, ignoring the temporal nature of the gait sequence can result in missing discriminative local motion information. Recently, some approaches have achieved significant performance gains by explicitly modeling temporal information using LSTM [56], 1D convolution [15], and 3D convolution [26, 16]. Nevertheless, these spatio-temporal operators also significantly increase computational costs. Although some methods have been proposed to reduce video length by aggregating local clips [27, 25, 3], they lack adaptability to variations in pace. The main difference between our approach and others [23, 11] is that we employ multi-scale temporal pooling at the frame level while considering variations in motion across body regions, leading to a more adaptable reduction of the gait sequence length. + +# 3. Proposed Method + +In this section, we present the detailed description of HSTL, including the adaptive region-based motion extractor (ARME), the adaptive spatio-temporal pooling (ASTP), and the frame-level temporal aggregation (FTA). + +# 3.1. Framework Pipeline + +The overview of our HSTL is presented in Fig. 2. Given a gait dataset $\mathcal{D} = \{S_i\}_{i=1}^N$ with $N$ gait sequences, where each sequence $S_i \in \mathbb{R}^{C \times T \times H \times W}$ is represented as a 4D tensor with $C$ channels, $T$ frames, and $H \times W$ pixels. During the preprocessing stage, each gait sequence $S_i$ is divided horizontally and uniformly into $k$ part sequences, indexed from 1 to $k$ . Then, a hierarchical clustering algorithm [7] is applied to these part sequences to obtain a generic hierarchy of gait motions, which is denoted as $\mathcal{P} = \{\mathcal{P}^{(l)}\}_{l=1}^L$ . Here, $L$ is the number of levels in the hierarchy and $\mathcal{P}^{(l)}$ is the set of partitions at level $l$ . The partitions at level $l$ are defined as $\mathcal{P}^{(l)} = \{P_1^{(l)}, P_2^{(l)}, \ldots, P_{K_l}^{(l)}\}$ , where $P_j^{(l)}$ is the $j$ -th subset of part indices and $K_l$ is the number of groups at level $l$ . For instance, the top level $\mathcal{P}^{(1)} = \{\{1, 2, \ldots, k-1, k\}\}$ means all the $k$ parts can be considered as a whole for the gait analysis. This hierarchy provides a structured property of the gait motion patterns and can be utilized to guide gait feature extraction. To achieve this, our proposed HSTL employs three modules: ARME for extracting independent multi-granularity motion features, ASTP for generating vectorized gait embeddings, and FTA for reducing redundant information at the frame level. The HSTL stacks these three modules according to the division in $\mathcal{P}$ , and the main branch of the HSTL for the input sequence $S_{in}$ can be formalized as: + +$$ +Y ^ {M} = \Gamma^ {(L)} \circ \Psi^ {(L - 1)} \circ \dots \circ \Omega^ {(2)} \circ \Psi^ {(2)} \circ \Psi^ {(1)} (S _ {i n}), (1) +$$ + +where $\Psi^{(l)},\Gamma^{(l)}$ , and $\Omega^{(l)}$ represent the ARME, ASTP and FTA modules at the $l$ -th level of $\mathcal{P}$ , respectively. Since FTA uses inter-frame compression to reduce redundant information, it is employed only once at the $l_{\Omega}$ -th level in $\mathcal{P}$ (e.g., $l_{\Omega} = 2$ in Eq.(1)) to prevent excessive loss of information. + +To obtain the hierarchical gait embeddings, the output $Y^{(l)}$ of each $\Psi^{(l)}$ at levels $l \in \{1,2,\dots ,L - 1\}$ and the output of $\Omega^{(l_{\Omega})}$ at level $l_{\Omega}$ , denoted as $Y_{\Omega}^{(l_{\Omega})}$ , are fed into the corresponding $\Gamma^{(l)}$ . The resulting outputs from these $L$ auxiliary branches are concatenated with the output of the main branch defined in Eq. (1), forming the final result, denoted as $Y$ , which is given by: + +$$ +\begin{array}{l} Y = \left[ Y ^ {M}, \Gamma^ {(L - 1)} \left(Y ^ {(L - 1)}\right), \dots , \right. \\ \left. \Gamma^ {\left(l _ {\Omega}\right)} \left(Y _ {\Omega} ^ {\left(l _ {\Omega}\right)}\right), \Gamma^ {(2)} \left(Y ^ {(2)}\right), \Gamma^ {(1)} \left(Y ^ {(1)}\right) \right], \tag {2} \\ \end{array} +$$ + +where $[,]$ denotes the concatenation operation. + +Finally, $Y$ undergoes feature mapping through separate fully connected layers. The model is then trained using a combination of triplet loss $\mathcal{L}_{tri}$ and cross-entropy loss $\mathcal{L}_{ce}$ , which is a commonly adopted practice in gait recognition [12, 27, 15, 16, 5]. Further details regarding the relevant modules are described in the following subsections. + +# 3.2. Adaptive Region-based Motion Extractor + +The adaptive region-based motion extractor (ARME) aims to extract independent spatio-temporal patterns that are associated with different human body parts in a gait sequence. Unlike existing methods that uniformly slice gait images or sequences along the height axis [56, 8, 27, 25], ARME considers the inherent hierarchical relationships among different part sequences that are consistent with walking patterns. This allows ARME to effectively capture the unique walking kinematics of each part. + +Given the hierarchical relation $\mathcal{P}$ introduced in Section 3.1, ARME first divides the input sequence $X$ into $K_{l}$ regions based on the partition of the $l$ -th level $\mathcal{P}^{(l)}$ , resulting in the set of regions $\{X_{j}^{(l)}\}_{j=1}^{K_{l}}$ , where $X_{j}^{(l)} \in \mathbb{R}^{C \times T \times H_{j}^{(l)} \times W}$ . $H_{j}^{(l)} = \frac{\left|P_{j}^{(l)}\right|}{k} H$ is the height of the $j$ -th region of the $l$ -th level. Then the $l$ -th level of ARME, $\Psi^{(l)}$ , can be defined as follows: + +$$ +Y _ {\Psi} ^ {(l)} = \Psi^ {(l)} \left(X ^ {(l)}\right) = \left[ f _ {1} \left(X _ {1} ^ {(l)}\right), f _ {2} \left(X _ {2} ^ {(l)}\right), \dots , f _ {K _ {l}} \left(X _ {K _ {l}} ^ {(l)}\right) \right], \tag {3} +$$ + +where $f_{j}(\cdot)$ represents the independent 3D convolution operation applied to the $j$ -th region. The output feature map $Y_{\Psi}^{(l)} \in \mathbb{R}^{C(l) \times T \times H \times W}$ of level $l$ has $C^{(l)}$ channels. This module only modifies the number of channels, preserving the spatial and temporal resolutions of the input feature map. + +# 3.3. Adaptive Spatio-Temporal Pooling + +It is a common procedure in gait recognition to obtain a compact and fixed-length feature representation by performing horizontal and uniform slicing of feature maps and strip-based pooling [4, 9, 26, 27, 15]. However, a nonuniform division of the feature maps is better at capturing the gait motion characteristics. Thus, the adaptive spatiotemporal pooling (ASTP) is devised to construct hierarchical feature mapping (as shown in Fig. 2). Similar to the ARME module described in Section 3.2, the hierarchy $\mathcal{P}$ enables us to obtain the $j$ -th region of the $l$ -th level, denoted as $X_{j}^{(l)} \in \mathbb{R}^{C \times T \times H_{j}^{(l)} \times W}$ . The corresponding ASTP, denoted as $\Gamma^{(l)}$ , can be expressed as follow: + +$$ +Y _ {\Gamma , j} ^ {(l)} = \Gamma_ {j} ^ {(l)} \left(X _ {j} ^ {(l)}\right) = \operatorname {G e M} _ {j} \circ \operatorname {F C} \circ \operatorname {M a x} \left(X _ {j} ^ {(l)}\right), \tag {4} +$$ + +where $\mathrm{Max}(\cdot)$ represents a max pooling operation along the temporal dimension, $\mathrm{FC}(\cdot)$ represents a fully connected layer, $\mathrm{GeM}_j(\cdot)$ represents a generalized mean pooling (GeM) operation [32] for $j$ -th region, and $\Gamma_j^{(l)}$ : $\mathbb{R}^{C\times T\times H_j^{(l)}\times W}\mapsto \mathbb{R}^{C\times 1\times H_j^{(l)}\times W}\mapsto \mathbb{R}^{C^{(l)}\times 1\times H_j^{(l)}\times W}\mapsto$ $\mathbb{R}^{C^{(l)}\times 1\times 1\times 1}$ , in other words, $Y_{\Gamma ,j}^{(l)}\in \mathbb{R}^{C^{(l)}\times 1\times 1\times 1}$ . Therefore, by concatenating outputs of $K_{l}$ regions, we can + +![](images/ea613ae139d910f745d276d8f510e9148a04742ff49bca542a91df61ab271baf.jpg) +Figure 3: The detailed structure of the frame-level temporal aggregation (FTA). For simplicity, we omit the channel dimension $C$ . + +obtain $Y_{\Gamma}^{(l)} = \left[Y_{\Gamma,1}^{(l)},Y_{\Gamma,2}^{(l)},\dots ,Y_{\Gamma,K_l}^{(l)}\right]$ , where $Y_{\Gamma}^{(l)} \in \mathbb{R}^{C^{(l)} \times 1 \times K_l \times 1}$ is the output of ASTP at level $l$ . + +# 3.4. Frame-level Temporal Aggregation + +A gait sequence may contain several redundant frames due to factors such as the acquisition frame rate and pace frequency. To reduce computational costs, some methods compress a gait sequence by aggregating its local clips [27, 25]. In the proposed frame-level temporal aggregation (FTA) strategy, we consider both the gait structure and the multiscale temporal information. Given the $j$ -th gait region at the $l$ -th level, $X_{j}^{(l)} \in \mathbb{R}^{C \times T \times H_{j}^{(l)} \times W}$ , we first fuse the features of the two temporal scales using the following formula: + +$$ +\begin{array}{l} \hat {U} _ {j} ^ {(l)} = U _ {j, 1} ^ {(l)} + U _ {j, 2} ^ {(l)} \\ = \operatorname {M a x} _ {3 \times 1 \times 1} ^ {3 \times 1 \times 1} \left(X _ {j} ^ {(l)}\right) + \operatorname {M a x} _ {5 \times 1 \times 1} ^ {3 \times 1 \times 1} \left(X _ {j} ^ {(l)}\right), \tag {5} \\ \end{array} +$$ + +where $\mathrm{Max}_{3\times 1\times 1}^{3\times 1\times 1}(\cdot)$ and $\mathrm{Max}_{5\times 1\times 1}^{3\times 1\times 1}(\cdot)$ denote max pooling operations with kernel sizes of $3\times 1\times 1$ and $5\times 1\times 1$ respectively, both with stride of $3\times 1\times 1$ . $\hat{U}_j^{(l)}, U_{j,1}^{(l)}$ and $U_{j,2}^{(l)}$ have the same size of $(C,\frac{T}{3},H_j^{(l)},W)$ . The output of Eq.(5), $\hat{U}_j^{(l)}$ , is the element-wise summation of the aggregation results of the two scales, $U_{j,1}^{(l)}$ and $U_{j,2}^{(l)}$ , which reduces the temporal dimension of the input from $T$ to $\frac{T}{3}$ . + +Then, the FTA model produces frame-level weights, which can be expressed as: + +$$ +\begin{array}{l} Z _ {j, 1} ^ {(l)} = \operatorname {F C} _ {j, 1} ^ {(l)} \left(\operatorname {G A P} \left(\hat {U} _ {j} ^ {(l)}\right)\right), \tag {6} \\ Z _ {j, 2} ^ {(l)} = \operatorname {F C} _ {j, 2} ^ {(l)} \left(\operatorname {G A P} \left(\hat {U} _ {j} ^ {(l)}\right)\right), \\ \end{array} +$$ + +where $\mathrm{GAP}(\cdot)$ represents the global mean pooling along spatial dimension. $\mathrm{FC}_{j,1}(\cdot)$ and $\mathrm{FC}_{j,2}(\cdot)$ are two independ- + +dent fully connected layers that generate the frame selection weighting tensors, $Z_{j,1}^{(l)}$ and $Z_{j,2}^{(l)} \in \mathbb{R}^{C \times \frac{T}{3} \times 1 \times 1}$ , for $U_{j,1}^{(l)}$ and $U_{j,2}^{(l)}$ , respectively. The weights are further normalized across the two scales, which can be written as follows: + +$$ +\mathcal {W} _ {j, s, c, t} ^ {(l)} = \frac {e ^ {Z _ {j , s , c , t} ^ {(l)}}}{e ^ {Z _ {j , 1 , c , t} ^ {(l)}} + e ^ {Z _ {j , 2 , c , t} ^ {(l)}}} \quad s \in \{1, 2 \}, \tag {7} +$$ + +where $\mathcal{W}_{j,s,c,t}^{(l)}\in \mathbb{R}^{1\times 1\times 1\times 1}$ is the weight value of the $c$ -th channel of the $t$ -th frame. Combining Eq. (5) and Eq. (7), the $j$ -th output region feature $Y_{\Omega ,j}^{(l)}\in \mathbb{R}^{C^{(l)}\times \frac{T}{3}\times H_j^{(l)}\times W}$ for the $l$ -th level of FTA can be obtained as follows: + +$$ +Y _ {\Omega , j} ^ {(l)} = \mathcal {W} _ {j, 1} ^ {(l)} \odot U _ {j, 1} ^ {(l)} + \mathcal {W} _ {j, 2} ^ {(l)} \odot U _ {j, 2} ^ {(l)}, \tag {8} +$$ + +where $\mathcal{W}_{j,1}^{(l)},\mathcal{W}_{j,2}^{(l)}\in \mathbb{R}^{C\times \frac{T}{3}\times 1\times 1}$ are two weight tensors calculated using Eq. (5), and $\odot$ represents element-wise multiplication operation. The FTA module outputs $Y_{\Omega}^{(l)}\in$ $\mathbb{R}^{C\times \frac{T}{3}\times H\times W}$ by concatenating the $K_{l}$ gait regions of level $l$ where $Y_{\Omega}^{(l)} = \left[Y_{\Omega ,1}^{(l)},Y_{\Omega ,2}^{(l)},\ldots ,Y_{\Omega ,K_l}^{(l)}\right]$ + +# 4. Experiments + +# 4.1. Datasets and Evaluation Protocols + +CASIA-B. The CASIA-B [52] dataset is a widely used benchmark for gait recognition. It contains video sequences of 124 subjects with 11 different views and three walking conditions (normal walking (NM), walking with a bag (BG), and walking with a coat (CL)). Our study follows the protocol outlined in previous works [4, 8, 27, 25, 15, 16]. The first 74 subjects are used for training, and the remaining 50 subjects are used for testing. During testing, the first four sequences under NM (NM#01-04) are regarded as the gallery set, and the rest (NM#05-06, BG#01-02, CL#01-02) are regarded as the probe set. + +OUMVLP. The OUMVLP [39] is one of the largest gait datasets, containing silhouette sequences of 10,307 subjects. Each subject has a single normal walking condition (NM) with 14 views. According to the protocol provided by the dataset, the first 5,153 subjects are used for training while the remaining 5,154 subjects are used for testing. During the testing phase, the sequences of NM#01 are assigned to the gallery set, and the sequences of NM#02 are considered as the probe set. + +GREW. GREW [60] is the first large-scale dataset for gait recognition in the wild, consisting of 128,671 sequences from 26,345 individuals captured by 882 cameras. It includes four data types: four data types: silhouettes, optical flow, 2D pose, and 3D pose. The dataset is divided into a training set with 20,000 subjects and 102,887 sequences, and a testing set with 6,000 subjects and 24,000 sequences. In the testing phase, each subject has two sequences for the + +Table 1: Rank-1 accuracy (%) on CASIA-B under all views and different conditions, excluding identical-view cases. Std denotes the performance sample standard deviation across 11 views. + +
Gallery NM #1-40° - 180°MeanStd
Probe0°18°36°54°72°90°108°126°144°162°180°
NM #5-6GaitSet [4]90.897.999.496.993.691.795.097.898.996.885.895.03.5
GaitPart [8]94.198.699.398.594.092.395.998.499.297.890.496.23.1
3D Local [16]96.099.099.598.997.194.296.399.098.898.595.297.51.8
CSTL [15]97.299.099.298.196.295.597.798.799.298.996.597.81.3
GaitGL [27]96.098.399.097.996.995.497.098.999.398.894.097.41.7
LagrangeGait [3]95.798.199.198.396.495.297.599.099.398.994.997.51.6
MetaGait [5]97.399.299.599.197.295.597.699.199.399.196.798.11.3
Ours97.698.099.698.297.496.597.999.399.498.497.098.11.0
BG #1-2GaitSet [4]83.891.291.888.883.381.084.190.092.294.479.087.24.9
GaitPart [8]89.194.896.795.188.384.989.093.596.193.885.891.54.2
3D Local [16]92.995.997.896.293.087.892.796.397.998.088.594.33.5
CSTL [15]91.796.597.095.490.988.091.595.897.095.590.393.63.0
GaitGL [27]92.696.696.895.593.589.392.296.598.296.991.594.52.8
LagrangeGait [3]94.296.296.895.894.389.591.796.898.097.090.994.62.7
MetaGait [5]92.996.797.196.494.790.492.997.298.598.192.395.22.6
Ours95.096.597.396.695.393.394.696.898.697.792.995.91.7
CL #1-2GaitSet [4]61.475.480.777.372.170.171.573.573.568.450.070.48.0
GaitPart [8]70.785.586.983.377.172.576.982.283.880.266.578.76.6
3D Local [16]78.290.292.087.183.076.883.186.686.884.170.983.76.2
CSTL [15]78.189.491.686.682.179.981.886.388.786.675.384.24.9
GaitGL [27]76.690.090.387.184.579.084.187.087.384.469.583.66.3
LagrangeGait [3]77.490.693.290.284.780.385.287.789.386.671.085.16.3
MetaGait [5]80.091.893.087.886.582.985.290.090.889.378.486.94.6
Ours82.494.295.091.788.283.388.092.393.191.078.588.95.1
+ +gallery set and two for the probe set. The GREW dataset also includes a distractor set, which contains 233,857 unlabeled sequences. + +Gait3D. The Gait3D dataset [59] is a newly proposed dataset for gait recognition in uncontrolled indoor environments, particularly in large supermarkets. It contains 25,309 sequences of 4,000 subjects extracted from 39 cameras, with 18,940 sequences from 3,000 subjects for training and 6,369 sequences from 1,000 subjects for testing. The dataset mainly includes four data types: silhouettes, 2D pose, 3D pose, and 3D mesh. During testing, one sequence per subject is used as the probe set, while the remaining sequences are used as the gallery set. To evaluate the model's performance, we use accuracy as well as mean average precision (mAP) and mean inverse negative penalty (mINP) [49], which consider multiple instances and hard sample recall. + +# 4.2. Implementation Details + +Training details. 1) In our implementation, the margin $m$ of the triplet loss is set to 0.2, and the parameter $p$ of the GeM function used in the ASTP module is initialized to 6.5. In the hierarchy of gait motion, the number of partitions $k$ at the bottom level is set to 8. 2) The batch size is set to (8,8) + +Table 2: The detailed architecture of the proposed HSTL on CASIA-B. The first column denotes the levels of the gait hierarchy and $K_{l}$ is the number of groups at level $l$ . $C_{in}$ and $C_{out}$ represent the input channel and output channel of each layer respectively. The body parts are indexed in spatial order from top to bottom, numbered 1 to 8. + +
LevelBlockLayer\( {C}_{in} \)\( {C}_{out} \)Kernel\( {K}_{l} \)Parts Grouping
1ARMEConv3d132(3,3,3)1\( \{ \{ 1,2,3,4,5,6,7,8\} \} \)
ASTP
2ARMEConv3d3232(3,3,3)2\( \{ \{ 1,2,3,4,5\} ,\{ 6,7,8\} \} \)
Conv3d3264(3,3,3)
ASTP
2FTAMaxPool6464(3,1,1)2\( \{ \{ 1,2,3,4,5\} ,\{ 6,7,8\} \} \)
(5,1,1)
ASTP
3ARMEConv3d64128(3,3,3)4\( \{ \{ 1\} ,\{ 2,3,4,5\} ,\{ 6,7\} ,\{ 8\} \} \)
Conv3d128128(3,3,3)
ASTP
4ASTP8\( \{ \{ 1\} ,\{ 2\} ,\{ 3\} ,\{ 4\} , \) \( \{ 5\} ,\{ 6\} ,\{ 7\} ,\{ 8\} \} \)
+ +for CASIA-B, (32,8) for OUMVLP, (32,4) for GREW, and (32,4) for Gait3D. 3) We use gait silhouettes as the input + +Table 3: Rank-1 accuracy (%) on OUMVLP under all views, excluding identical-view cases. Std denotes the performance sample standard deviation across 14 views. + +
MethodProbe ViewMeanStd
0°15°30°45°60°75°90°180°195°210°225°240°255°270°
GaitSet [4]79.387.990.090.188.088.787.781.886.589.089.287.287.686.287.14.0
GaitPart [8]82.688.990.891.089.789.989.585.288.190.090.189.089.188.288.72.3
GLN [12]83.890.091.091.290.390.089.485.389.190.590.689.689.388.589.22.1
CSTL [15]87.191.091.591.890.690.890.689.490.290.590.789.890.089.490.21.1
GaitGL [27]84.990.291.191.591.190.890.388.588.690.390.489.689.588.889.71.7
3D Local [16]86.191.292.692.992.291.391.186.990.892.292.391.391.190.290.92.0
LagrangeGait [3]85.990.691.391.591.291.090.688.989.290.590.689.989.889.290.01.4
MetaGait [5]88.292.393.093.593.192.792.689.391.292.092.692.391.991.191.91.4
Ours91.492.992.793.092.992.592.592.792.392.192.392.291.891.892.40.5
+ +modality. During the training phase, we sample 30 frames following the strategy proposed in [8], while during testing, all frames are fed into the model. Moreover, we align each frame following the strategy presented in [39], and the input image size is cropped to $64 \times 44$ for all datasets. 4) For CASIA-B, the optimizer used is Adam with a weight decay of $5e-4$ . The model is trained for 100K iterations with an initial learning rate (LR) of $1e-5$ , and the LR is multiplied by 0.1 at 70K iterations. For OUMVLP, GREW, and Gait3D, the model is trained for 250K, 250K, and 210K iterations, respectively, with an initial LR of 0.1. The LR is multiplied by 0.1 at 150K and 200K iterations. The optimizer used for these datasets is SGD with a weight decay of $5e-4$ . + +Architecture details. Table 2 presents the detailed architecture of the model used for the CASIA-B dataset. To handle datasets with more subjects, such as OUMVLP, GREW, and Gait3D, we incorporate the label smoothing operation into the cross-entropy loss function, and deepen the network by adding an extra ARME module to the third level of the hierarchy. The output channels for the four ARMEs are set to 64, 64, 128 and 256, respectively. Additionally, we include a layer of spatial downsampling after the first ARME to improve the training efficiency. + +Table 4: Rank-1 accuracy (\%), Rank-5 accuracy (\%), Rank-10 accuracy (\%), Rank-20 accuracy (\%) on GREW. + +
MethodsRank-1Rank-5Rank-10Rank-20
GaitSet [4]46.2863.5870.2676.82
GaitPart [8]44.0160.6867.2573.47
GaitGL [27]47.2863.5669.3274.18
MTSGait [58]55.3271.2876.8581.55
Ours62.7276.5781.3285.24
+ +# 4.3. Comparison with State-of-the-Art Methods + +Evaluation on CASIA-B. Table 1 compares the performance of the proposed HSTL with seven state-of-the-art + +Table 5: Rank-1 accuracy $(\%)$ , Rank-5 accuracy $(\%)$ , mAP $(\%)$ and mINP on Gait3D. + +
MethodsRank-1Rank-5mAPmINP
GaitSet [4]36.7058.3030.0117.30
GaitPart [8]28.2047.6021.5812.36
GLN [12]31.4052.9024.7413.58
GaitGL [27]29.7048.5022.2913.26
CSTL [15]11.7019.205.592.59
SMPLGait [59]46.3064.5037.1622.23
MTSGait [58]48.7067.1037.6321.92
Ours61.3076.3055.4834.77
+ +methods on the CASIA-B dataset. It can be seen that the proposed method obtains the best results in all three walking conditions while maintaining considerable stability across different views. The experimental results reveal that 1) the mean accuracy of the proposed method for the BG and CL walking conditions is $95.9\%$ and $88.9\%$ , respectively, which are higher by $0.7\%$ and $2.0\%$ than the second-best method (MetaGait [5]), demonstrating the advantage of our method in cross-view gait recognition. 2) The decrease in accuracy from NM to CL is $9.2\%$ for our method, compared to $12.4\%$ for 3D Local [16]. This indicates that the ARME module is more adaptable to various walking conditions. In addition, our method achieves an accuracy of $88.9\%$ in the challenging CL condition, which is $5.2\%$ higher than 3D Local. This may be due to the complex clothing conditions that affect the part localization accuracy in 3D Local. 3) Both CSTL [15] and our method extracts multi-scale temporal features but in different ways. CSTL first extracts spatial information and then fuses three scales of motion features. In contrast, we propose an FTA module to aggregate spatiotemporal information from multiple body regions. Therefore, the average rank-1 accuracy of our method is higher or equivalent to that of CSTL in all views. This suggests that FTA can be more adaptive to spatio-temporal changes. + +![](images/1b3df54efb68d4f7d1d4dd6f5e5575c510b7ee677da840905a1554140abb5d6a.jpg) +(a) Average rank-1 accuracy + +![](images/8fcfaaf69f6c0eaa2d1ceeeff340112c9d6f6103cc441745270a83b708dd7a20.jpg) +(b) Age distribution +Figure 4: Cross-age comparison results for the OUMVLP dataset. (a) Average rank-1 accuracy in adults vs children. (b) Distribution of two age groups. + +Evaluation on OUMVLP. To verify the model's generalizability, we conduct experiments on the large-scale OUMVLP dataset. As shown in Table 3, our approach achieves competitive results in most views. In particular, the proposed method outperforms the second-best method (MetaGait) by an average of $2.5\%$ under three extreme views, i.e., $0^{\circ}$ , $180^{\circ}$ , and $270^{\circ}$ , resulting in the best mean performance and cross-view stability. In addition, the OUMVLP dataset also provides annotations for the ages of the subjects. To evaluate the impact of age differences on recognition performance, we divided all subjects into two groups: adults (18-87 years old) and children (2-17 years old), as shown in Fig. 4(b). Fig. 4(a) presents the recognition accuracy based on age. It can be observed that, as adult sequences make up about $70\%$ of the total sequences, all the compared methods show a bias toward the recognition results of adults. However, compared to other methods, our model effectively improves the accuracy of gait recognition for children, demonstrating the effectiveness of our hierarchical gait representation across ages. + +Evaluation on GREW and Gait3D. Gait3D and GREW are two recently introduced datasets that contain challenging conditions, such as misalignment of the human body and partial occlusion. Tables 4 and 5 show the results of the comparison between the proposed method and the state-of-the-art methods. Our method shows superior performance in all metrics, indicating its ability to effectively model gait characteristics in realistic scenarios. + +# 4.4. Ablation Study + +Effectiveness of hierarchical feature extraction. To verify the effectiveness of our hierarchical gait partitioning, we conduct experiment with various grouping strategies. As shown in Table 2, this comparison only considers the different settings of the first three layers since a uniform partition is used at layer 4. Specifically, 1-1-1 indicates that the first three levels have the same number of groupings, and they are divided uniformly. $1 - 2 - 4^{*}$ refers to the nonuniform division used in the proposed model. As shown in + +![](images/f157c36fffc6bc2e3a2fb3c459dafda257bdcb183813c7e78d9c9725f005e33f.jpg) +Figure 5: Ablation study on the effectiveness of hierarchical feature extraction on the CASIA-B dataset (best viewed in color). + +Table 6: Ablation study on the effectiveness of ARME, ASTP, and FTA modules in terms of average rank-1 accuracy on the CASIA-B dataset. + +
SettingARMEASTPFTANMBGCLMean
35
a97.094.584.291.9
b✓97.895.385.993.0
c✓✓97.895.487.093.4
d✓✓✓97.795.387.593.5
e✓✓✓97.795.287.293.4
f✓✓✓97.895.487.593.6
g✓✓✓✓98.195.988.994.3
+ +Fig. 5, the hierarchical feature extraction setting, e.g., 1-2-4, outperforms the other non-hierarchical approaches. A further mean performance improvement of $1.0\%$ is achieved when the motion relationships between different body regions, i.e., 1-2-4*, are considered. + +Effectiveness of ARME, ASTP, and FTA. The results of the ablation experiments for ARME, ASTP, and FTA are shown in Tab. 6 The results indicate that the ARME module significantly contributes to the improvement of the recognition accuracy, with an average improvement of $1.1\%$ compared to the baseline model (non-grouping version of ARME). The mean accuracy is further improved by $1.3\%$ when the ASTP module is integrated and multi-scale fusion is utilized in FTA. These results demonstrate the effectiveness and complementarity of the three modules in the proposed gait recognition framework. + +# 4.5. Trade-off between accuracy and efficiency + +In this subsection, we evaluate the relationship between accuracy and efficiency for each compared method. As shown in Fig. 6, the 3D convolution-based approaches, such as GaitGL [27], 3D Local [16], LagrangeGait [3] and MetaGait [5], outperform the 2D convolution-based methods, like GaitSet [4] and GaitPart [8], in terms of accuracy, but + +![](images/e37d6a628cb52e34fa5fd7fa799c99e534ef52eb2e2b158c8103251db87897ae.jpg) +Figure 6: The trade-off between accuracy and FLOPs of our method and other comparison methods on the OUMVLP dataset. + +at the cost of a significant increase in FLOPs (floating point operations). Our approach has a better trade-off between accuracy and efficiency. The main reason is that the proposed hierarchical learning architecture can extract multilevel motion features while reducing the number of 3D convolutions. More experimental results and ablation analysis are provided in the supplementary material. + +# 5. Conclusion + +This paper presents a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL stacks multiple adaptive region-based motion extractors (ARMEs) and learns walking patterns in a coarse-to-fine manner. An adaptive spatio-temporal pooling (ASTP) module is proposed to perform hierarchical feature mapping for the output of each level of ARME. Additionally, a frame-level temporal aggregation module (FTA) is designed to compress local clips by fusing temporal information. The effectiveness of the proposed HSTL framework is demonstrated through extensive experiments conducted on four public datasets (CASIA-B, OUMVLP, GREW, and Gait3D). + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China (Grant Nos. 61972132, 62106065) and the Research Project for Self-cultivating Talents of Hebei Agricultural University (Grant No. PY201810). + +# References + +[1] Weizhi An, Shiqi Yu, Yasushi Makihara, Xinhui Wu, Chi Xu, Yang Yu, Rijun Liao, and Yasushi Yagi. Performance evalua + +tion of model-based gait on multi-view very large population database with pose sequences. IEEE TBIOM, 2(4):421-430, 2020. 1, 2 +[2] Pia Bideau, Aruni RoyChowdhury, Rakesh R Menon, and Erik Learned-Miller. The best of both worlds: Combining cnns and geometric constraints for hierarchical motion segmentation. In CVPR, pages 508-517, 2018. 1 +[3] Tianrui Chai, Annan Li, Shaoxiong Zhang, Zilong Li, and Yunhong Wang. Lagrange motion analysis and view embeddings for improved gait recognition. In CVPR, pages 20249-20258, 2022. 1, 2, 3, 6, 7, 8 +[4] Hanqing Chao, Yiwei He, Junping Zhang, and Jianfeng Feng. Gaitset: Regarding gait as a set for cross-view gait recognition. In AAAI, pages 8126-8133, 2019. 1, 2, 3, 4, 5, 6, 7, 8 +[5] Huanzhang Dou, Pengyi Zhang, Wei Su, Yunlong Yu, and Xi Li. Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In ECCV, pages 357-374. Springer, 2022. 2, 4, 6, 7, 8 +[6] Jessica Maria Echterhoff, Juan Haladjian, and Bernd Brügge. Gait and jump classification in modern equestrian sports. In ISWC, pages 88-91, 2018. 1 +[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 2, 4 +[8] Chao Fan, Yunjie Peng, Chunshui Cao, Xu Liu, Saihui Hou, Jiannan Chi, Yongzhen Huang, Qing Li, and Zhiqiang He. Gaitpart: Temporal part-based model for gait recognition. In CVPR, pages 14225-14233, 2020. 1, 2, 3, 4, 5, 6, 7, 8 +[9] Yang Fu, Yunchao Wei, Yuqian Zhou, Honghui Shi, Gao Huang, Xinchao Wang, Zhiqiang Yao, and Thomas Huang. Horizontal pyramid matching for person re-identification. In AAAI, pages 8295–8302, 2019. 4 +[10] Jinguang Han and Bir Bhanu. Individual recognition using gait energy image. IEEE TPAMI, 28(2):316-322, 2005. 1, 2 +[11] Ruibing Hou, Hong Chang, Bingpeng Ma, Rui Huang, and Shiguang Shan. Bicnet-tks: Learning efficient spatial-temporal representation for video person re-identification. In CVPR, pages 2014–2023, 2021. 3 +[12] Saihui Hou, Chunshui Cao, Xu Liu, and Yongzhen Huang. Gait lateral network: Learning discriminative and compact representations for gait recognition. In ECCV, pages 382-398, 2020. 1, 2, 3, 4, 7 +[13] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Set residual network for silhouette-based gait recognition. IEEE TBIOM, 3(3):384-393, 2021. 2, 3 +[14] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Gait quality aware network: Toward the interpretability of silhouette-based gait recognition. IEEE TNNLS, 2022. 2 +[15] Xiaohu Huang, Duowang Zhu, Hao Wang, Xinggang Wang, Bo Yang, Botao He, Wenyu Liu, and Bin Feng. Context-sensitive temporal feature learning for gait recognition. In ICCV, pages 12909-12918, 2021. 1, 2, 3, 4, 5, 6, 7 +[16] Zhen Huang, Dixiu Xue, Xu Shen, Xinmei Tian, Houqiang Li, Jianqiang Huang, and Xian-Sheng Hua. 3d local convolutional neural networks for gait recognition. In ICCV, pages 14920-14929, 2021. 1, 2, 3, 4, 5, 6, 7, 8 + +[17] Munish Kumar, Navdeep Singh, Ravinder Kumar, Shubham Goel, and Krishan Kumar. Gait recognition based on vision systems: A systematic survey. Journal of Visual Communication and Image Representation, 75:103052, 2021. 1 +[18] Worapan Kusakunniran. Review of gait recognition approaches and their challenges on view changes. IET Biometrics, 9(6):238-250, 2020. 1 +[19] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE TIP, 31:5949-5962, 2022. 2 +[20] Xiang Li, Yasushi Makihara, Chi Xu, and Yasushi Yagi. End-to-end model-based gait recognition using synchronized multi-view pose constraint. In ICCVW, pages 4106-4115, 2021. 1, 2 +[21] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, and Mingwu Ren. Gait recognition via semi-supervised disentangled representation learning to identity and covariate features. In CVPR, pages 13309-13319, 2020. 3 +[22] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, Shiqi Yu, and Mingwu Ren. End-to-end model-based gait recognition. In ACCV, pages 3-20, 2020. 1, 2 +[23] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In CVPR, pages 510-519, 2019. 3 +[24] Junhao Liang, Chao Fan, Saihui Hou, Chuanfu Shen, Yongzhen Huang, and Shiqi Yu. Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In ECCV, pages 375-390. Springer, 2022. 2 +[25] Beibei Lin, Yu Liu, and Shunli Zhang. Gaitmask: Mask-based model for gait recognition. In BMVC, page 363, 2021. 1, 2, 3, 4, 5 +[26] Beibei Lin, Shunli Zhang, and Feng Bao. Gait recognition with multiple-temporal-scale 3d convolutional neural network. In ACM MM, pages 3054-3062, 2020. 3, 4 +[27] Beibei Lin, Shunli Zhang, and Xin Yu. Gait recognition via effective global-local feature representation and local temporal aggregation. In ICCV, pages 14648-14656, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[28] Nini Liu and Yap-Peng Tan. View invariant gait recognition. In ICASSP, pages 1410-1413, 2010. 1 +[29] Yasushi Makihara, Mark S Nixon, and Yasushi Yagi. Gait recognition: Databases, representations, and applications. Computer Vision: A Reference Guide, pages 1-13, 2020. 1 +[30] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical gaussian descriptor for person re-identification. In CVPR, pages 1363–1372, 2016. 2 +[31] Daigo Muramatsu, Yasushi Makihara, Haruyuki Iwama, Takuya Tanoue, and Yasushi Yagi. Gait verification system for supporting criminal investigation. In IAPR, pages 747-748, 2013. 1 +[32] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 41(7):1655-1668, 2018. 4 +[33] Imad Rida, Noor Almaadeed, and Somaya Almaadeed. Robust gait recognition: a comprehensive survey. IET Biometrics, 8(1):14-28, 2019. 1 + +[34] Alireza Sepas-Moghaddam and Ali Etemad. Deep gait recognition: A survey. IEEE TPAMI, 2022. 1 +[35] Kohei Shiraga, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Geinet: View-invariant gait recognition using a convolutional neural network. In ICB, pages 1-8, 2016. 1, 2, 3 +[36] Pranjay Shyam, Kuk-Jin Yoon, and Kyung-Soo Kim. Adversarily-trained hierarchical feature extractor for vehicle re-identification. In ICRA, pages 13400–13407, 2021. 2 +[37] Chunfeng Song, Yongzhen Huang, Yan Huang, Ning Jia, and Liang Wang. Gaitnet: An end-to-end network for gait based human identification. PR, 96:106988, 2019. 2 +[38] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 1 +[39] Noriko Takemura, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ TCVA, 10:1-14, 2018. 2, 5, 7 +[40] Lei Tan, Pingyang Dai, Rongrong Ji, and Yongjian Wu. Dynamic prototype mask for occluded person re-identification. In ACM MM, pages 531–540, 2022. 2 +[41] Torben Teepe, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Towards a deeper understanding of skeleton-based gait recognition. In CVPRW, pages 1569-1577, 2022. 2 +[42] Torben Teepe, Ali Khan, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In ICIP, pages 2314-2318, 2021. 1, 2 +[43] Md Uddin, Daigo Muramatsu, Noriko Takemura, Md Ahad, Atiqur Rahman, Yasushi Yagi, et al. Spatio-temporal silhouette sequence reconstruction for gait recognition against occlusion. IPSJ TCVA, 11(1):1-18, 2019. 1 +[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2 +[45] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. IEEE TCSVT, 2021. 2 +[46] Xiu-Shen Wei, Chen-Lin Zhang, Lingqiao Liu, Chunhua Shen, and Jianxin Wu. Coarse-to-fine: A rnn-based hierarchical attention model for vehicle re-identification. In ACCV, pages 575–591, 2018. 2 +[47] Jiachen Yang, Jianxiong Zhou, Dayong Fan, and Haibin Lv. Design of intelligent recognition system based on gait recognition technology in smart transportation. Multimedia Tools and Applications, 75(24):17501-17514, 2016. 1 +[48] Lingxiang Yao, Worapan Kusakunniran, Qiang Wu, Jingsong Xu, and Jian Zhang. Collaborative feature learning for gait recognition under cloth changes. IEEE TCSVT, 2021. 1 +[49] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6):2872-2893, 2021. 6 + +[50] TzeWei Yeoh, Hernan E Aguirre, and Kiyoshi Tanaka. Clothing-invariant gait recognition using convolutional neural network. In ISPACS, pages 1-5, 2016. 1 +[51] Shiqi Yu, Haifeng Chen, Qing Wang, Linlin Shen, and Yongzhen Huang. Invariant feature extraction for gait recognition using only one uniform model. Neurocomputing, 239:81-93, 2017. 1, 2 +[52] Shiqi Yu, Daoliang Tan, and Tieniu Tan. A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In ICPR, pages 441-444, 2006. 2, 5 +[53] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 2 +[54] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 17:516-530, 2022. 2 +[55] Shaoxiong Zhang, Yunhong Wang, and Annan Li. Crossview gait recognition with deep universal linear embeddings. In CVPR, pages 9095-9104, 2021. 1 +[56] Yuqi Zhang, Yongzhen Huang, Shiqi Yu, and Liang Wang. Cross-view gait recognition by discriminative feature learning. IEEE TIP, 29:1001-1015, 2019. 1, 3, 4 +[57] Ziyuan Zhang, Luan Tran, Xi Yin, Yousef Atoum, Xiaoming Liu, Jian Wan, and Nanxin Wang. Gait recognition via disentangled representation learning. In CVPR, pages 4710-4719, 2019. 2 +[58] Jinkai Zheng, Xinchen Liu, Xiaoyan Gu, Yaoqi Sun, Chuang Gan, Jiyong Zhang, Wu Liu, and Chenggang Yan. Gait recognition in the wild with multi-hop temporal switch. In ACM MM, pages 6136-6145, 2022. 7 +[59] Jinkai Zheng, Xinchen Liu, Wu Liu, Lingxiao He, Chenggang Yan, and Tao Mei. Gait recognition in the wild with dense 3d representations and a benchmark. In CVPR, pages 20228-20237, 2022. 1, 2, 6, 7 +[60] Zheng Zhu, Xianda Guo, Tian Yang, Junjie Huang, Jiankang Deng, Guan Huang, Dalong Du, Jiwen Lu, and Jie Zhou. Gait recognition in the wild: A benchmark. In ICCV, pages 14789-14799, 2021. 2, 5 \ No newline at end of file diff --git a/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/images.zip b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d9dbf1f7f006fd043b91b29bf3d2c32f1e06bc41 --- /dev/null +++ b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aceeab06bdf069abc8ed2d5685bc97075612e6361b2fcbe247a586bd68f10bcb +size 743794 diff --git a/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/layout.json b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..24421c7732feaf36125d415843e1531a301ca7e1 --- /dev/null +++ b/2023/Hierarchical Spatio-Temporal Representation Learning for Gait Recognition/layout.json @@ -0,0 +1,9723 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 136, + 518, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 136, + 518, + 153 + ], + "spans": [ + { + "bbox": [ + 59, + 136, + 518, + 153 + ], + "type": "text", + "content": "Hierarchical Spatio-Temporal Representation Learning for Gait Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "spans": [ + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "text", + "content": "Lei Wang" + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "text", + "content": ", Bo Liu" + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "text", + "content": ", Fangfang Liang" + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "text", + "content": ", Bincheng Wang" + }, + { + "bbox": [ + 143, + 174, + 431, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 189, + 493, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 189, + 493, + 216 + ], + "spans": [ + { + "bbox": [ + 84, + 189, + 493, + 216 + ], + "type": "text", + "content": "1 College of Information Science and Technology, Hebei Agricultural University, China \n2 Hebei Key Laboratory of Agricultural Big Data, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 218, + 496, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 218, + 496, + 229 + ], + "spans": [ + { + "bbox": [ + 79, + 218, + 496, + 229 + ], + "type": "text", + "content": "{20212060107, 20212060108}@pgs.hebau.edu.cn, {boliu, liangfangfang}@hebau.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 256, + 186, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 256, + 186, + 268 + ], + "spans": [ + { + "bbox": [ + 139, + 256, + 186, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 282, + 282, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 282, + 282, + 596 + ], + "spans": [ + { + "bbox": [ + 45, + 282, + 282, + 596 + ], + "type": "text", + "content": "Gait recognition is a biometric technique that identifies individuals by their unique walking styles, which is suitable for unconstrained environments and has a wide range of applications. While current methods focus on exploiting body part-based representations, they often neglect the hierarchical dependencies between local motion patterns. In this paper, we propose a hierarchical spatio-temporal representation learning (HSTL) framework for extracting gait features from coarse to fine. Our framework starts with a hierarchical clustering analysis to recover multilevel body structures from the whole body to local details. Next, an adaptive region-based motion extractor (ARME) is designed to learn region-independent motion features. The proposed HSTL then stacks multiple ARMEs in a top-down manner, with each ARME corresponding to a specific partition level of the hierarchy. An adaptive spatiotemporal pooling (ASTP) module is used to capture gait features at different levels of detail to perform hierarchical feature mapping. Finally, a frame-level temporal aggregation (FTA) module is employed to reduce redundant information in gait sequences through multi-scale temporal downsampling. Extensive experiments on CASIA-B, OUMVLP, GREW, and Gait3D datasets demonstrate that our method outperforms the state-of-the-art while maintaining a reasonable balance between model accuracy and complexity. Code is available at: https://github.com/gudaochangsheng/HSTL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 619, + 123, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 123, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 123, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 639, + 279, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 639, + 279, + 710 + ], + "spans": [ + { + "bbox": [ + 45, + 639, + 279, + 710 + ], + "type": "text", + "content": "Unlike other biometric technologies such as fingerprint, iris, and face, human gait can be captured at a distance without subject cooperation [34]. By evaluating individual-specific walking patterns, gait recognition has been applied in a variety of fields, including criminal investigations [31, 29], sports science [17, 6], and smart transportation" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 299, + 255, + 532, + 346 + ], + "blocks": [ + { + "bbox": [ + 299, + 255, + 532, + 346 + ], + "lines": [ + { + "bbox": [ + 299, + 255, + 532, + 346 + ], + "spans": [ + { + "bbox": [ + 299, + 255, + 532, + 346 + ], + "type": "image", + "image_path": "a5c08c72dcfa0c9537bddd60dc16b987cfe8a966154eafa67b2997ed616ebac1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 353, + 531, + 424 + ], + "lines": [ + { + "bbox": [ + 296, + 353, + 531, + 424 + ], + "spans": [ + { + "bbox": [ + 296, + 353, + 531, + 424 + ], + "type": "text", + "content": "Figure 1: The motivation for our approach. Left: the motion distributions of body parts. The same color indicates the same spatial location across multiple gait sequences in the CASIA-B dataset. Right: an example hierarchy of body-part sequences where " + }, + { + "bbox": [ + 296, + 353, + 531, + 424 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 353, + 531, + 424 + ], + "type": "text", + "content": " denotes the temporal dimension of the sequence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 447, + 531, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 447, + 531, + 482 + ], + "spans": [ + { + "bbox": [ + 296, + 447, + 531, + 482 + ], + "type": "text", + "content": "[47]. However, the recognition can be challenging due to large variations in viewpoint [28, 18], occlusion [33, 43], and wearing [50, 48]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 484, + 531, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 484, + 531, + 635 + ], + "spans": [ + { + "bbox": [ + 296, + 484, + 531, + 635 + ], + "type": "text", + "content": "To address these issues, various approaches have been proposed for extracting gait features from silhouette sequences [4, 25, 16, 15, 55], 3D human structures [1, 22, 42, 59, 20], or gait templates [10, 35, 51]. Silhouette-based gait recognition methods have gained increasing attention due to the ease of obtaining silhouettes from raw videos while preserving essential temporal information. The alignment of the input silhouette makes it possible for some methods to extract local body features by horizontally slicing the silhouette image [56] or intermediate-layer features [8, 27]. This partitioning strategy, first introduced in person re-identification (ReID) [38], has been proven to be effective for gait recognition [4, 8, 12, 3]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "However, the main limitation of the above part-based approaches is that they do not consider the hierarchical nature of local body movements [2]. For instance, within a gait cycle, the feet and lower body have distinct motion characteristics. Therefore, it is important to treat these body regions separately and investigate their part-whole relationships. Our motivation stems from the examination of body-part-specific motion clues. Specifically, each raw gait se" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 7, + 101, + 39 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 7, + 101, + 39 + ], + "spans": [ + { + "bbox": [ + 57, + 7, + 101, + 39 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 6, + 485, + 41 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 6, + 485, + 41 + ], + "spans": [ + { + "bbox": [ + 138, + 6, + 485, + 41 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19639" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 268 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 268 + ], + "type": "text", + "content": "quence in the CASIA-B [52] dataset is uniformly divided into eight part sequences along the body axis, so that each division roughly match a particular body part. The distributions of all body parts are shown in the left part of Fig. 1. Observably, some parts, e.g., the head and feet, are easily separated owing to their large changes in walking kinematics. Whereas other parts, such as the thighs and calves, overlap due to the strong motion correlations between them. Further, to identify the relational structure among the part sequences, a hierarchical clustering analysis [7] is performed. The results are shown in the right part of Fig. 1, indicating that the semantic body regions can be captured in the higher clustering levels without precise localization of the body parts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 269, + 281, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 269, + 281, + 467 + ], + "spans": [ + { + "bbox": [ + 45, + 269, + 281, + 467 + ], + "type": "text", + "content": "Following the above findings, we propose a novel hierarchical spatio-temporal representation learning (HSTL) framework for gait representation. The HSTL framework consists of multiple adaptive region-based motion extractor (ARME) modules, which are stacked to learn hierarchical motion patterns implied in a gait sequence (as shown in Fig.1). In the ARME module, to account for inter-regional differences, non-shared 3D convolutions are used in correspondence with individual body regions. There regions are pre-identified by a hierarchical clustering process performed on fixed horizontal partitions, allowing each body region to cover one or more body parts. Consequently, the deeper the ARME is, the more local features it tends to extract. Moreover, an adaptive spatio-temporal pooling (ASTP) module is proposed, which couples with an ARME module on the corresponding level to obtain hierarchical gait embeddings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 468, + 281, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 468, + 281, + 585 + ], + "spans": [ + { + "bbox": [ + 45, + 468, + 281, + 585 + ], + "type": "text", + "content": "In addition, changes in gait speed or sampling frequency may result in several redundant frames in a gait sequence. Although several temporal fusion strategies have been proposed, they lose spatial information [8, 15] or lack adaptability [27, 25]. To address this issue, we propose a frame-level temporal aggregation strategy (FTA). FTA fuses temporal features at multiple time steps, preserving significant motion information while compressing the sequence length. The main contributions of this paper are summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 595, + 280, + 711 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 54, + 595, + 280, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 595, + 280, + 655 + ], + "spans": [ + { + "bbox": [ + 54, + 595, + 280, + 655 + ], + "type": "text", + "content": "1. We propose a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL takes into account the dependencies of body regions in gait motions, ensuring simplicity and scalability of the architectural design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 663, + 280, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 663, + 280, + 711 + ], + "spans": [ + { + "bbox": [ + 52, + 663, + 280, + 711 + ], + "type": "text", + "content": "2. We introduce an adaptive region-based motion extractor (ARME) module to learn region-independent spatio-temporal representation for gait sequences, an adaptive spatio-temporal pooling (ASTP) module to" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 106, + 531, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 531, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 531, + 142 + ], + "type": "text", + "content": "perform hierarchical feature mapping, and a frame-level temporal aggregation (FTA) strategy to compress a gait sequence by removing redundant frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 150, + 532, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 150, + 532, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 150, + 532, + 222 + ], + "type": "text", + "content": "3. Extensive experiments on the widely used gait datasets CASIA-B [52], including OUMVLP [39], GREW [60] and Gait3D [59], demonstrate that our method achieves state-of-the-art performance while offering a suitable trade-off between model accuracy and complexity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 233, + 381, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 233, + 381, + 245 + ], + "spans": [ + { + "bbox": [ + 297, + 233, + 381, + 245 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 253, + 397, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 253, + 397, + 265 + ], + "spans": [ + { + "bbox": [ + 297, + 253, + 397, + 265 + ], + "type": "text", + "content": "2.1. Gait Recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 271, + 532, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 271, + 532, + 446 + ], + "spans": [ + { + "bbox": [ + 296, + 271, + 532, + 446 + ], + "type": "text", + "content": "Deep learning-based gait recognition methods can be broadly categorized into two categories: model-based and appearance-based. Model-based approaches extract structure and motion information from gait videos with the aid of pose estimation [1, 22, 42, 20, 41, 59]. Although these methods are robust to changes in viewpoint and appearance, they are sensitive to the accuracy of the pose parameters, making them incapable of handling low-resolution data. On the other hand, appearance-based approaches learn the feature representation from raw videos [57, 37, 24], or binary silhouette sequences [10, 35, 51, 4, 12, 8, 13, 27, 16, 14, 3, 5], which offer greater flexibility than model-based approaches. Our proposed method belongs to the family of appearance-based methods and uses silhouette sequences as inputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 454, + 409, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 454, + 409, + 467 + ], + "spans": [ + { + "bbox": [ + 297, + 454, + 409, + 467 + ], + "type": "text", + "content": "2.2. Hierarchical Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 473, + 530, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 473, + 530, + 519 + ], + "spans": [ + { + "bbox": [ + 296, + 473, + 530, + 519 + ], + "type": "text", + "content": "Hierarchical feature representation has been successfully applied to a wide range of vision tasks. Here, we provide a brief review of the hierarchical object ReID approaches related to gait recognition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 520, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 520, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 520, + 532, + 730 + ], + "type": "text", + "content": "In person ReID, some approaches [30, 54, 45, 40, 53] hierarchically learn local descriptions and aggregate appearance features at different levels. For example, Matsukawa et al. [30] described an image patch via hierarchical Gaussian distribution. Zhang et al. [53] proposed a framework to learn coarse-grained and fine-grained features according to body structure. To solve the occlusion problem, Tan et al. [40] devised a hierarchical mask generator to learn from both occluded and holistic joint images. In vehicle ReID, some approaches [46, 36, 19] extract features from vehicle images in a hierarchical manner. For instance, Wei et al. [46] proposed an RNN-based module for extracting latent cues from the model level to the vehicle level. Shyam et al. [36] developed an attention-based hierarchical feature extractor. In addition, Li et al. [19] proposed a global structural embedding module for investigating hierarchical relationships between vehicle characteristics by incorporating attribute and state information." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 719, + 274, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 274, + 730 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 274, + 730 + ], + "type": "text", + "content": "Each part is pooled into a vector for visualization using t-SNE [44]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19640" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 46, + 104, + 530, + 339 + ], + "blocks": [ + { + "bbox": [ + 46, + 104, + 530, + 339 + ], + "lines": [ + { + "bbox": [ + 46, + 104, + 530, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 104, + 530, + 339 + ], + "type": "image", + "image_path": "01ddaa05a50ac8566503c24c507354c4f738ba5c94c89ee856cb3c177c889603.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "lines": [ + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "spans": [ + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "type": "text", + "content": "Figure 2: The framework of HFSL. It mainly consists of three modules: ARME (adaptive region-based motion extractor), ASTP (adaptive spatio-temporal pooling), and FTA (frame-level temporal aggregation). During pre-processing, a hierarchy of walking is obtained to guide the architectural design of HFSL. The framework uses multiple ARMEs to extract gait features from the entire body to individual regions. The ASTP module performs hierarchical feature mapping for the output of each level of ARME. The FTA module compresses local clips of each gait sequence to reduce the number of redundant frames. " + }, + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 45, + 347, + 531, + 417 + ], + "type": "text", + "content": " denote dimensions of the feature maps. © represents the concatenation operation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 438, + 280, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 438, + 280, + 625 + ], + "spans": [ + { + "bbox": [ + 45, + 438, + 280, + 625 + ], + "type": "text", + "content": "For gait recognition, fusing features of multiple granularities can improve performance [56, 8, 27, 25, 16, 3]. In particular, CSTL [15] proposed a temporal modeling network that integrates multi-scale temporal features adaptively. By combining part-level and sequence-level features, GaitPart [8] obtained a part-independent spatio-temporal expression. Additionally, GaitGL [27] considered both full body-based and part-based information to achieve discriminative feature learning. Instead of equally dividing the feature maps, in 3D Local [16], a localization operation was developed to find 3D volumes of body parts in a sequence. However, most existing gait recognition methods do not sufficiently exploit the hierarchical dependencies among body parts during walking. In this paper, the proposed HSTL performs a coarse-to-fine hierarchical strategy that integrates multi-level motion patterns from gait sequences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 143, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 143, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 143, + 642 + ], + "type": "text", + "content": "2.3. Temporal Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 279, + 730 + ], + "type": "text", + "content": "Temporal cues play a crucial role in gait recognition due to the periodic changes in body shape. Previous methods treat a gait sequence as an unordered set, either compressing it into a single gait template during preprocessing [35, 21] or learning order-independent gait representations from silhouette sets [4, 12, 13]. These methods assume that different subjects share similar global gait patterns, making or" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 438, + 531, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 438, + 531, + 614 + ], + "spans": [ + { + "bbox": [ + 296, + 438, + 531, + 614 + ], + "type": "text", + "content": "dering inputs unnecessary for gait assessment. However, ignoring the temporal nature of the gait sequence can result in missing discriminative local motion information. Recently, some approaches have achieved significant performance gains by explicitly modeling temporal information using LSTM [56], 1D convolution [15], and 3D convolution [26, 16]. Nevertheless, these spatio-temporal operators also significantly increase computational costs. Although some methods have been proposed to reduce video length by aggregating local clips [27, 25, 3], they lack adaptability to variations in pace. The main difference between our approach and others [23, 11] is that we employ multi-scale temporal pooling at the frame level while considering variations in motion across body regions, leading to a more adaptable reduction of the gait sequence length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 653, + 400, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 653, + 400, + 666 + ], + "spans": [ + { + "bbox": [ + 297, + 653, + 400, + 666 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 683, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 683, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 683, + 531, + 730 + ], + "type": "text", + "content": "In this section, we present the detailed description of HSTL, including the adaptive region-based motion extractor (ARME), the adaptive spatio-temporal pooling (ASTP), and the frame-level temporal aggregation (FTA)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "text", + "content": "19641" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 106, + 162, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 106, + 162, + 117 + ], + "spans": [ + { + "bbox": [ + 46, + 106, + 162, + 117 + ], + "type": "text", + "content": "3.1. Framework Pipeline" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "spans": [ + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": "The overview of our HSTL is presented in Fig. 2. Given a gait dataset " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{S_i\\}_{i=1}^N" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " gait sequences, where each sequence " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "S_i \\in \\mathbb{R}^{C \\times T \\times H \\times W}" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " is represented as a 4D tensor with " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " channels, " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " frames, and " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " pixels. During the preprocessing stage, each gait sequence " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " is divided horizontally and uniformly into " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " part sequences, indexed from 1 to " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": ". Then, a hierarchical clustering algorithm [7] is applied to these part sequences to obtain a generic hierarchy of gait motions, which is denoted as " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathcal{P}^{(l)}\\}_{l=1}^L" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " is the number of levels in the hierarchy and " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(l)}" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " is the set of partitions at level " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": ". The partitions at level " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " are defined as " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(l)} = \\{P_1^{(l)}, P_2^{(l)}, \\ldots, P_{K_l}^{(l)}\\}" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "P_j^{(l)}" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": "-th subset of part indices and " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "K_l" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " is the number of groups at level " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": ". For instance, the top level " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(1)} = \\{\\{1, 2, \\ldots, k-1, k\\}\\}" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " means all the " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " parts can be considered as a whole for the gait analysis. This hierarchy provides a structured property of the gait motion patterns and can be utilized to guide gait feature extraction. To achieve this, our proposed HSTL employs three modules: ARME for extracting independent multi-granularity motion features, ASTP for generating vectorized gait embeddings, and FTA for reducing redundant information at the frame level. The HSTL stacks these three modules according to the division in " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": ", and the main branch of the HSTL for the input sequence " + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "inline_equation", + "content": "S_{in}" + }, + { + "bbox": [ + 45, + 123, + 281, + 415 + ], + "type": "text", + "content": " can be formalized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 421, + 279, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 421, + 279, + 436 + ], + "spans": [ + { + "bbox": [ + 50, + 421, + 279, + 436 + ], + "type": "interline_equation", + "content": "Y ^ {M} = \\Gamma^ {(L)} \\circ \\Psi^ {(L - 1)} \\circ \\dots \\circ \\Omega^ {(2)} \\circ \\Psi^ {(2)} \\circ \\Psi^ {(1)} (S _ {i n}), (1)", + "image_path": "cffe0fb6499dfb04e8d347d14b0a5d67516b18c8646f689a4b50ad48efd7c991.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "spans": [ + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "inline_equation", + "content": "\\Psi^{(l)},\\Gamma^{(l)}" + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "inline_equation", + "content": "\\Omega^{(l)}" + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": " represent the ARME, ASTP and FTA modules at the " + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": " -th level of " + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": " , respectively. Since FTA uses inter-frame compression to reduce redundant information, it is employed only once at the " + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "inline_equation", + "content": "l_{\\Omega}" + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": " -th level in " + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "inline_equation", + "content": "l_{\\Omega} = 2" + }, + { + "bbox": [ + 45, + 441, + 279, + 500 + ], + "type": "text", + "content": " in Eq.(1)) to prevent excessive loss of information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "spans": [ + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": "To obtain the hierarchical gait embeddings, the output " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "Y^{(l)}" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": " of each " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "\\Psi^{(l)}" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": " at levels " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "l \\in \\{1,2,\\dots ,L - 1\\}" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": " and the output of " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "\\Omega^{(l_{\\Omega})}" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "l_{\\Omega}" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "Y_{\\Omega}^{(l_{\\Omega})}" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": ", are fed into the corresponding " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "\\Gamma^{(l)}" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": ". The resulting outputs from these " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": " auxiliary branches are concatenated with the output of the main branch defined in Eq. (1), forming the final result, denoted as " + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 45, + 500, + 280, + 585 + ], + "type": "text", + "content": ", which is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 589, + 279, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 279, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 279, + 642 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Y = \\left[ Y ^ {M}, \\Gamma^ {(L - 1)} \\left(Y ^ {(L - 1)}\\right), \\dots , \\right. \\\\ \\left. \\Gamma^ {\\left(l _ {\\Omega}\\right)} \\left(Y _ {\\Omega} ^ {\\left(l _ {\\Omega}\\right)}\\right), \\Gamma^ {(2)} \\left(Y ^ {(2)}\\right), \\Gamma^ {(1)} \\left(Y ^ {(1)}\\right) \\right], \\tag {2} \\\\ \\end{array}", + "image_path": "943dcd5ceb8208079b317ea4049f25538443aa66ea8d99cec576012f768d20c7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 648, + 229, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 229, + 660 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 229, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 648, + 229, + 660 + ], + "type": "inline_equation", + "content": "[,]" + }, + { + "bbox": [ + 45, + 648, + 229, + 660 + ], + "type": "text", + "content": " denotes the concatenation operation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " undergoes feature mapping through separate fully connected layers. The model is then trained using a combination of triplet loss " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tri}" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " and cross-entropy loss " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": ", which is a commonly adopted practice in gait recognition [12, 27, 15, 16, 5]. Further details regarding the relevant modules are described in the following subsections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 106, + 508, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 106, + 508, + 117 + ], + "spans": [ + { + "bbox": [ + 297, + 106, + 508, + 117 + ], + "type": "text", + "content": "3.2. Adaptive Region-based Motion Extractor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 124, + 531, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 124, + 531, + 229 + ], + "spans": [ + { + "bbox": [ + 296, + 124, + 531, + 229 + ], + "type": "text", + "content": "The adaptive region-based motion extractor (ARME) aims to extract independent spatio-temporal patterns that are associated with different human body parts in a gait sequence. Unlike existing methods that uniformly slice gait images or sequences along the height axis [56, 8, 27, 25], ARME considers the inherent hierarchical relationships among different part sequences that are consistent with walking patterns. This allows ARME to effectively capture the unique walking kinematics of each part." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "spans": [ + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": "Given the hierarchical relation " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": " introduced in Section 3.1, ARME first divides the input sequence " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "K_{l}" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": " regions based on the partition of the " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": "-th level " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(l)}" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": ", resulting in the set of regions " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "\\{X_{j}^{(l)}\\}_{j=1}^{K_{l}}" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "H_{j}^{(l)} = \\frac{\\left|P_{j}^{(l)}\\right|}{k} H" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": " is the height of the " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": "-th region of the " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": "-th level. Then the " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": "-th level of ARME, " + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "inline_equation", + "content": "\\Psi^{(l)}" + }, + { + "bbox": [ + 296, + 229, + 531, + 324 + ], + "type": "text", + "content": ", can be defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 330, + 533, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 330, + 533, + 360 + ], + "spans": [ + { + "bbox": [ + 297, + 330, + 533, + 360 + ], + "type": "interline_equation", + "content": "Y _ {\\Psi} ^ {(l)} = \\Psi^ {(l)} \\left(X ^ {(l)}\\right) = \\left[ f _ {1} \\left(X _ {1} ^ {(l)}\\right), f _ {2} \\left(X _ {2} ^ {(l)}\\right), \\dots , f _ {K _ {l}} \\left(X _ {K _ {l}} ^ {(l)}\\right) \\right], \\tag {3}", + "image_path": "350ce7da5f669bd8e77b935348fa8d4f59df56883ed66b63041c4dc2484106f1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "spans": [ + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "inline_equation", + "content": "f_{j}(\\cdot)" + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "text", + "content": " represents the independent 3D convolution operation applied to the " + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "text", + "content": "-th region. The output feature map " + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "inline_equation", + "content": "Y_{\\Psi}^{(l)} \\in \\mathbb{R}^{C(l) \\times T \\times H \\times W}" + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "text", + "content": " of level " + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "inline_equation", + "content": "C^{(l)}" + }, + { + "bbox": [ + 296, + 361, + 531, + 432 + ], + "type": "text", + "content": " channels. This module only modifies the number of channels, preserving the spatial and temporal resolutions of the input feature map." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 441, + 478, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 441, + 478, + 454 + ], + "spans": [ + { + "bbox": [ + 296, + 441, + 478, + 454 + ], + "type": "text", + "content": "3.3. Adaptive Spatio-Temporal Pooling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "spans": [ + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "text", + "content": "It is a common procedure in gait recognition to obtain a compact and fixed-length feature representation by performing horizontal and uniform slicing of feature maps and strip-based pooling [4, 9, 26, 27, 15]. However, a nonuniform division of the feature maps is better at capturing the gait motion characteristics. Thus, the adaptive spatiotemporal pooling (ASTP) is devised to construct hierarchical feature mapping (as shown in Fig. 2). Similar to the ARME module described in Section 3.2, the hierarchy " + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "text", + "content": " enables us to obtain the " + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "text", + "content": "-th region of the " + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "text", + "content": "-th level, denoted as " + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "inline_equation", + "content": "X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}" + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "text", + "content": ". The corresponding ASTP, denoted as " + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "inline_equation", + "content": "\\Gamma^{(l)}" + }, + { + "bbox": [ + 296, + 460, + 531, + 605 + ], + "type": "text", + "content": ", can be expressed as follow:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 614, + 530, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 614, + 530, + 633 + ], + "spans": [ + { + "bbox": [ + 320, + 614, + 530, + 633 + ], + "type": "interline_equation", + "content": "Y _ {\\Gamma , j} ^ {(l)} = \\Gamma_ {j} ^ {(l)} \\left(X _ {j} ^ {(l)}\\right) = \\operatorname {G e M} _ {j} \\circ \\operatorname {F C} \\circ \\operatorname {M a x} \\left(X _ {j} ^ {(l)}\\right), \\tag {4}", + "image_path": "f39fe886823325c0c0a68781213da582239a81469b8f5523f1a1dce432595773.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{Max}(\\cdot)" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": " represents a max pooling operation along the temporal dimension, " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}(\\cdot)" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": " represents a fully connected layer, " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{GeM}_j(\\cdot)" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": " represents a generalized mean pooling (GeM) operation [32] for " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": "-th region, and " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\Gamma_j^{(l)}" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": " : " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times T\\times H_j^{(l)}\\times W}\\mapsto \\mathbb{R}^{C\\times 1\\times H_j^{(l)}\\times W}\\mapsto \\mathbb{R}^{C^{(l)}\\times 1\\times H_j^{(l)}\\times W}\\mapsto" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C^{(l)}\\times 1\\times 1\\times 1}" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": ", in other words, " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "Y_{\\Gamma ,j}^{(l)}\\in \\mathbb{R}^{C^{(l)}\\times 1\\times 1\\times 1}" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": ". Therefore, by concatenating outputs of " + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "inline_equation", + "content": "K_{l}" + }, + { + "bbox": [ + 296, + 641, + 531, + 730 + ], + "type": "text", + "content": " regions, we can" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19642" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 103, + 281, + 250 + ], + "blocks": [ + { + "bbox": [ + 47, + 103, + 281, + 250 + ], + "lines": [ + { + "bbox": [ + 47, + 103, + 281, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 281, + 250 + ], + "type": "image", + "image_path": "ea613ae139d910f745d276d8f510e9148a04742ff49bca542a91df61ab271baf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 255, + 280, + 291 + ], + "lines": [ + { + "bbox": [ + 45, + 255, + 280, + 291 + ], + "spans": [ + { + "bbox": [ + 45, + 255, + 280, + 291 + ], + "type": "text", + "content": "Figure 3: The detailed structure of the frame-level temporal aggregation (FTA). For simplicity, we omit the channel dimension " + }, + { + "bbox": [ + 45, + 255, + 280, + 291 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 255, + 280, + 291 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "spans": [ + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "text", + "content": "obtain " + }, + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "inline_equation", + "content": "Y_{\\Gamma}^{(l)} = \\left[Y_{\\Gamma,1}^{(l)},Y_{\\Gamma,2}^{(l)},\\dots ,Y_{\\Gamma,K_l}^{(l)}\\right]" + }, + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "inline_equation", + "content": "Y_{\\Gamma}^{(l)} \\in \\mathbb{R}^{C^{(l)} \\times 1 \\times K_l \\times 1}" + }, + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "text", + "content": " is the output of ASTP at level " + }, + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 308, + 279, + 341 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 227, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 227, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 227, + 360 + ], + "type": "text", + "content": "3.4. Frame-level Temporal Aggregation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "spans": [ + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "text", + "content": "A gait sequence may contain several redundant frames due to factors such as the acquisition frame rate and pace frequency. To reduce computational costs, some methods compress a gait sequence by aggregating its local clips [27, 25]. In the proposed frame-level temporal aggregation (FTA) strategy, we consider both the gait structure and the multiscale temporal information. Given the " + }, + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "text", + "content": "-th gait region at the " + }, + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "text", + "content": "-th level, " + }, + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "inline_equation", + "content": "X_{j}^{(l)} \\in \\mathbb{R}^{C \\times T \\times H_{j}^{(l)} \\times W}" + }, + { + "bbox": [ + 45, + 365, + 279, + 482 + ], + "type": "text", + "content": ", we first fuse the features of the two temporal scales using the following formula:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 488, + 279, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 488, + 279, + 527 + ], + "spans": [ + { + "bbox": [ + 59, + 488, + 279, + 527 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {U} _ {j} ^ {(l)} = U _ {j, 1} ^ {(l)} + U _ {j, 2} ^ {(l)} \\\\ = \\operatorname {M a x} _ {3 \\times 1 \\times 1} ^ {3 \\times 1 \\times 1} \\left(X _ {j} ^ {(l)}\\right) + \\operatorname {M a x} _ {5 \\times 1 \\times 1} ^ {3 \\times 1 \\times 1} \\left(X _ {j} ^ {(l)}\\right), \\tag {5} \\\\ \\end{array}", + "image_path": "380fca9842abd4e5fd8d34c5f2b0ab7d136c199de167dcf7fdf025183a703884.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "spans": [ + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{Max}_{3\\times 1\\times 1}^{3\\times 1\\times 1}(\\cdot)" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{Max}_{5\\times 1\\times 1}^{3\\times 1\\times 1}(\\cdot)" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " denote max pooling operations with kernel sizes of " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "3\\times 1\\times 1" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "5\\times 1\\times 1" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " respectively, both with stride of " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "3\\times 1\\times 1" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "\\hat{U}_j^{(l)}, U_{j,1}^{(l)}" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "U_{j,2}^{(l)}" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " have the same size of " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "(C,\\frac{T}{3},H_j^{(l)},W)" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": ". The output of Eq.(5), " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "\\hat{U}_j^{(l)}" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": ", is the element-wise summation of the aggregation results of the two scales, " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "U_{j,1}^{(l)}" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "U_{j,2}^{(l)}" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": ", which reduces the temporal dimension of the input from " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "inline_equation", + "content": "\\frac{T}{3}" + }, + { + "bbox": [ + 45, + 532, + 279, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 629, + 279, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 629, + 279, + 652 + ], + "spans": [ + { + "bbox": [ + 45, + 629, + 279, + 652 + ], + "type": "text", + "content": "Then, the FTA model produces frame-level weights, which can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 101, + 658, + 279, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 658, + 279, + 700 + ], + "spans": [ + { + "bbox": [ + 101, + 658, + 279, + 700 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Z _ {j, 1} ^ {(l)} = \\operatorname {F C} _ {j, 1} ^ {(l)} \\left(\\operatorname {G A P} \\left(\\hat {U} _ {j} ^ {(l)}\\right)\\right), \\tag {6} \\\\ Z _ {j, 2} ^ {(l)} = \\operatorname {F C} _ {j, 2} ^ {(l)} \\left(\\operatorname {G A P} \\left(\\hat {U} _ {j} ^ {(l)}\\right)\\right), \\\\ \\end{array}", + "image_path": "6cd6b40e46934153504ce7a52db06e09f96f30c5c136dce8edfcd0fd5f2a647a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{GAP}(\\cdot)" + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": " represents the global mean pooling along spatial dimension. " + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}_{j,1}(\\cdot)" + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}_{j,2}(\\cdot)" + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": " are two independ-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "text", + "content": "dent fully connected layers that generate the frame selection weighting tensors, " + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "inline_equation", + "content": "Z_{j,1}^{(l)}" + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "inline_equation", + "content": "Z_{j,2}^{(l)} \\in \\mathbb{R}^{C \\times \\frac{T}{3} \\times 1 \\times 1}" + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "inline_equation", + "content": "U_{j,1}^{(l)}" + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "inline_equation", + "content": "U_{j,2}^{(l)}" + }, + { + "bbox": [ + 296, + 106, + 531, + 158 + ], + "type": "text", + "content": ", respectively. The weights are further normalized across the two scales, which can be written as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 327, + 165, + 530, + 197 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 165, + 530, + 197 + ], + "spans": [ + { + "bbox": [ + 327, + 165, + 530, + 197 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {j, s, c, t} ^ {(l)} = \\frac {e ^ {Z _ {j , s , c , t} ^ {(l)}}}{e ^ {Z _ {j , 1 , c , t} ^ {(l)}} + e ^ {Z _ {j , 2 , c , t} ^ {(l)}}} \\quad s \\in \\{1, 2 \\}, \\tag {7}", + "image_path": "3e133599f7244a84afe44782aa94e494656d390cf739fb8ded537be7b9078818.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "spans": [ + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{j,s,c,t}^{(l)}\\in \\mathbb{R}^{1\\times 1\\times 1\\times 1}" + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "content": " is the weight value of the " + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "content": " -th channel of the " + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "content": " -th frame. Combining Eq. (5) and Eq. (7), the " + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "content": " -th output region feature " + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "inline_equation", + "content": "Y_{\\Omega ,j}^{(l)}\\in \\mathbb{R}^{C^{(l)}\\times \\frac{T}{3}\\times H_j^{(l)}\\times W}" + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 205, + 531, + 257 + ], + "type": "text", + "content": " -th level of FTA can be obtained as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 263, + 530, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 263, + 530, + 281 + ], + "spans": [ + { + "bbox": [ + 341, + 263, + 530, + 281 + ], + "type": "interline_equation", + "content": "Y _ {\\Omega , j} ^ {(l)} = \\mathcal {W} _ {j, 1} ^ {(l)} \\odot U _ {j, 1} ^ {(l)} + \\mathcal {W} _ {j, 2} ^ {(l)} \\odot U _ {j, 2} ^ {(l)}, \\tag {8}", + "image_path": "13cad37b763653ff3b896582aed5389a25d0f26fb984c7d670a327d135f02083.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "spans": [ + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{j,1}^{(l)},\\mathcal{W}_{j,2}^{(l)}\\in \\mathbb{R}^{C\\times \\frac{T}{3}\\times 1\\times 1}" + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "text", + "content": " are two weight tensors calculated using Eq. (5), and " + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "text", + "content": " represents element-wise multiplication operation. The FTA module outputs " + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "inline_equation", + "content": "Y_{\\Omega}^{(l)}\\in" + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times \\frac{T}{3}\\times H\\times W}" + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "text", + "content": " by concatenating the " + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "inline_equation", + "content": "K_{l}" + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "text", + "content": " gait regions of level " + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 296, + 288, + 531, + 360 + ], + "type": "inline_equation", + "content": "Y_{\\Omega}^{(l)} = \\left[Y_{\\Omega ,1}^{(l)},Y_{\\Omega ,2}^{(l)},\\ldots ,Y_{\\Omega ,K_l}^{(l)}\\right]" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 367, + 377, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 367, + 377, + 380 + ], + "spans": [ + { + "bbox": [ + 297, + 367, + 377, + 380 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 386, + 476, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 386, + 476, + 397 + ], + "spans": [ + { + "bbox": [ + 297, + 386, + 476, + 397 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 404, + 530, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 404, + 530, + 532 + ], + "spans": [ + { + "bbox": [ + 296, + 404, + 530, + 532 + ], + "type": "text", + "content": "CASIA-B. The CASIA-B [52] dataset is a widely used benchmark for gait recognition. It contains video sequences of 124 subjects with 11 different views and three walking conditions (normal walking (NM), walking with a bag (BG), and walking with a coat (CL)). Our study follows the protocol outlined in previous works [4, 8, 27, 25, 15, 16]. The first 74 subjects are used for training, and the remaining 50 subjects are used for testing. During testing, the first four sequences under NM (NM#01-04) are regarded as the gallery set, and the rest (NM#05-06, BG#01-02, CL#01-02) are regarded as the probe set." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 532, + 530, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 532, + 530, + 636 + ], + "spans": [ + { + "bbox": [ + 296, + 532, + 530, + 636 + ], + "type": "text", + "content": "OUMVLP. The OUMVLP [39] is one of the largest gait datasets, containing silhouette sequences of 10,307 subjects. Each subject has a single normal walking condition (NM) with 14 views. According to the protocol provided by the dataset, the first 5,153 subjects are used for training while the remaining 5,154 subjects are used for testing. During the testing phase, the sequences of NM#01 are assigned to the gallery set, and the sequences of NM#02 are considered as the probe set." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "content": "GREW. GREW [60] is the first large-scale dataset for gait recognition in the wild, consisting of 128,671 sequences from 26,345 individuals captured by 882 cameras. It includes four data types: four data types: silhouettes, optical flow, 2D pose, and 3D pose. The dataset is divided into a training set with 20,000 subjects and 102,887 sequences, and a testing set with 6,000 subjects and 24,000 sequences. In the testing phase, each subject has two sequences for the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19643" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 138, + 531, + 429 + ], + "blocks": [ + { + "bbox": [ + 45, + 104, + 531, + 129 + ], + "lines": [ + { + "bbox": [ + 45, + 104, + 531, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 104, + 531, + 129 + ], + "type": "text", + "content": "Table 1: Rank-1 accuracy (%) on CASIA-B under all views and different conditions, excluding identical-view cases. Std denotes the performance sample standard deviation across 11 views." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 138, + 531, + 429 + ], + "lines": [ + { + "bbox": [ + 48, + 138, + 531, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 531, + 429 + ], + "type": "table", + "html": "
Gallery NM #1-40° - 180°MeanStd
Probe0°18°36°54°72°90°108°126°144°162°180°
NM #5-6GaitSet [4]90.897.999.496.993.691.795.097.898.996.885.895.03.5
GaitPart [8]94.198.699.398.594.092.395.998.499.297.890.496.23.1
3D Local [16]96.099.099.598.997.194.296.399.098.898.595.297.51.8
CSTL [15]97.299.099.298.196.295.597.798.799.298.996.597.81.3
GaitGL [27]96.098.399.097.996.995.497.098.999.398.894.097.41.7
LagrangeGait [3]95.798.199.198.396.495.297.599.099.398.994.997.51.6
MetaGait [5]97.399.299.599.197.295.597.699.199.399.196.798.11.3
Ours97.698.099.698.297.496.597.999.399.498.497.098.11.0
BG #1-2GaitSet [4]83.891.291.888.883.381.084.190.092.294.479.087.24.9
GaitPart [8]89.194.896.795.188.384.989.093.596.193.885.891.54.2
3D Local [16]92.995.997.896.293.087.892.796.397.998.088.594.33.5
CSTL [15]91.796.597.095.490.988.091.595.897.095.590.393.63.0
GaitGL [27]92.696.696.895.593.589.392.296.598.296.991.594.52.8
LagrangeGait [3]94.296.296.895.894.389.591.796.898.097.090.994.62.7
MetaGait [5]92.996.797.196.494.790.492.997.298.598.192.395.22.6
Ours95.096.597.396.695.393.394.696.898.697.792.995.91.7
CL #1-2GaitSet [4]61.475.480.777.372.170.171.573.573.568.450.070.48.0
GaitPart [8]70.785.586.983.377.172.576.982.283.880.266.578.76.6
3D Local [16]78.290.292.087.183.076.883.186.686.884.170.983.76.2
CSTL [15]78.189.491.686.682.179.981.886.388.786.675.384.24.9
GaitGL [27]76.690.090.387.184.579.084.187.087.384.469.583.66.3
LagrangeGait [3]77.490.693.290.284.780.385.287.789.386.671.085.16.3
MetaGait [5]80.091.893.087.886.582.985.290.090.889.378.486.94.6
Ours82.494.295.091.788.283.388.092.393.191.078.588.95.1
", + "image_path": "d48851b29735f913159e9d0a776d016e16e222032da6cda922a3c06d1c9bbc8c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 449, + 279, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 449, + 279, + 484 + ], + "spans": [ + { + "bbox": [ + 45, + 449, + 279, + 484 + ], + "type": "text", + "content": "gallery set and two for the probe set. The GREW dataset also includes a distractor set, which contains 233,857 unlabeled sequences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 484, + 279, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 484, + 279, + 648 + ], + "spans": [ + { + "bbox": [ + 45, + 484, + 279, + 648 + ], + "type": "text", + "content": "Gait3D. The Gait3D dataset [59] is a newly proposed dataset for gait recognition in uncontrolled indoor environments, particularly in large supermarkets. It contains 25,309 sequences of 4,000 subjects extracted from 39 cameras, with 18,940 sequences from 3,000 subjects for training and 6,369 sequences from 1,000 subjects for testing. The dataset mainly includes four data types: silhouettes, 2D pose, 3D pose, and 3D mesh. During testing, one sequence per subject is used as the probe set, while the remaining sequences are used as the gallery set. To evaluate the model's performance, we use accuracy as well as mean average precision (mAP) and mean inverse negative penalty (mINP) [49], which consider multiple instances and hard sample recall." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 175, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 175, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 175, + 666 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "Training details. 1) In our implementation, the margin " + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": " of the triplet loss is set to 0.2, and the parameter " + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": " of the GeM function used in the ASTP module is initialized to 6.5. In the hierarchy of gait motion, the number of partitions " + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": " at the bottom level is set to 8. 2) The batch size is set to (8,8)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "spans": [ + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "text", + "content": "Table 2: The detailed architecture of the proposed HSTL on CASIA-B. The first column denotes the levels of the gait hierarchy and " + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "inline_equation", + "content": "K_{l}" + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "text", + "content": " is the number of groups at level " + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "inline_equation", + "content": "C_{out}" + }, + { + "bbox": [ + 296, + 448, + 530, + 518 + ], + "type": "text", + "content": " represent the input channel and output channel of each layer respectively. The body parts are indexed in spatial order from top to bottom, numbered 1 to 8." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 299, + 527, + 530, + 679 + ], + "blocks": [ + { + "bbox": [ + 299, + 527, + 530, + 679 + ], + "lines": [ + { + "bbox": [ + 299, + 527, + 530, + 679 + ], + "spans": [ + { + "bbox": [ + 299, + 527, + 530, + 679 + ], + "type": "table", + "html": "
LevelBlockLayer\\( {C}_{in} \\)\\( {C}_{out} \\)Kernel\\( {K}_{l} \\)Parts Grouping
1ARMEConv3d132(3,3,3)1\\( \\{ \\{ 1,2,3,4,5,6,7,8\\} \\} \\)
ASTP
2ARMEConv3d3232(3,3,3)2\\( \\{ \\{ 1,2,3,4,5\\} ,\\{ 6,7,8\\} \\} \\)
Conv3d3264(3,3,3)
ASTP
2FTAMaxPool6464(3,1,1)2\\( \\{ \\{ 1,2,3,4,5\\} ,\\{ 6,7,8\\} \\} \\)
(5,1,1)
ASTP
3ARMEConv3d64128(3,3,3)4\\( \\{ \\{ 1\\} ,\\{ 2,3,4,5\\} ,\\{ 6,7\\} ,\\{ 8\\} \\} \\)
Conv3d128128(3,3,3)
ASTP
4ASTP8\\( \\{ \\{ 1\\} ,\\{ 2\\} ,\\{ 3\\} ,\\{ 4\\} , \\) \\( \\{ 5\\} ,\\{ 6\\} ,\\{ 7\\} ,\\{ 8\\} \\} \\)
", + "image_path": "4ad98b02a4fa046e472213cee361469b0e943d8a696d872552eb5e859c07ed7b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 706, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 531, + 730 + ], + "type": "text", + "content": "for CASIA-B, (32,8) for OUMVLP, (32,4) for GREW, and (32,4) for Gait3D. 3) We use gait silhouettes as the input" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19644" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 137, + 531, + 257 + ], + "blocks": [ + { + "bbox": [ + 45, + 104, + 531, + 129 + ], + "lines": [ + { + "bbox": [ + 45, + 104, + 531, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 104, + 531, + 129 + ], + "type": "text", + "content": "Table 3: Rank-1 accuracy (%) on OUMVLP under all views, excluding identical-view cases. Std denotes the performance sample standard deviation across 14 views." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 137, + 531, + 257 + ], + "lines": [ + { + "bbox": [ + 48, + 137, + 531, + 257 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 531, + 257 + ], + "type": "table", + "html": "
MethodProbe ViewMeanStd
0°15°30°45°60°75°90°180°195°210°225°240°255°270°
GaitSet [4]79.387.990.090.188.088.787.781.886.589.089.287.287.686.287.14.0
GaitPart [8]82.688.990.891.089.789.989.585.288.190.090.189.089.188.288.72.3
GLN [12]83.890.091.091.290.390.089.485.389.190.590.689.689.388.589.22.1
CSTL [15]87.191.091.591.890.690.890.689.490.290.590.789.890.089.490.21.1
GaitGL [27]84.990.291.191.591.190.890.388.588.690.390.489.689.588.889.71.7
3D Local [16]86.191.292.692.992.291.391.186.990.892.292.391.391.190.290.92.0
LagrangeGait [3]85.990.691.391.591.291.090.688.989.290.590.689.989.889.290.01.4
MetaGait [5]88.292.393.093.593.192.792.689.391.292.092.692.391.991.191.91.4
Ours91.492.992.793.092.992.592.592.792.392.192.392.291.891.892.40.5
", + "image_path": "c2b3c965b11b5d1fda20c6444107874709f56e54377223a8f1560bd413df11c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "spans": [ + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "text", + "content": "modality. During the training phase, we sample 30 frames following the strategy proposed in [8], while during testing, all frames are fed into the model. Moreover, we align each frame following the strategy presented in [39], and the input image size is cropped to " + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "inline_equation", + "content": "64 \\times 44" + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "text", + "content": " for all datasets. 4) For CASIA-B, the optimizer used is Adam with a weight decay of " + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "inline_equation", + "content": "5e-4" + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "text", + "content": ". The model is trained for 100K iterations with an initial learning rate (LR) of " + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "inline_equation", + "content": "1e-5" + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "text", + "content": ", and the LR is multiplied by 0.1 at 70K iterations. For OUMVLP, GREW, and Gait3D, the model is trained for 250K, 250K, and 210K iterations, respectively, with an initial LR of 0.1. The LR is multiplied by 0.1 at 150K and 200K iterations. The optimizer used for these datasets is SGD with a weight decay of " + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "inline_equation", + "content": "5e-4" + }, + { + "bbox": [ + 45, + 275, + 279, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 437, + 279, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 437, + 279, + 555 + ], + "spans": [ + { + "bbox": [ + 45, + 437, + 279, + 555 + ], + "type": "text", + "content": "Architecture details. Table 2 presents the detailed architecture of the model used for the CASIA-B dataset. To handle datasets with more subjects, such as OUMVLP, GREW, and Gait3D, we incorporate the label smoothing operation into the cross-entropy loss function, and deepen the network by adding an extra ARME module to the third level of the hierarchy. The output channels for the four ARMEs are set to 64, 64, 128 and 256, respectively. Additionally, we include a layer of spatial downsampling after the first ARME to improve the training efficiency." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 49, + 595, + 279, + 671 + ], + "blocks": [ + { + "bbox": [ + 46, + 563, + 279, + 587 + ], + "lines": [ + { + "bbox": [ + 46, + 563, + 279, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 279, + 587 + ], + "type": "text", + "content": "Table 4: Rank-1 accuracy (\\%), Rank-5 accuracy (\\%), Rank-10 accuracy (\\%), Rank-20 accuracy (\\%) on GREW." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 595, + 279, + 671 + ], + "lines": [ + { + "bbox": [ + 49, + 595, + 279, + 671 + ], + "spans": [ + { + "bbox": [ + 49, + 595, + 279, + 671 + ], + "type": "table", + "html": "
MethodsRank-1Rank-5Rank-10Rank-20
GaitSet [4]46.2863.5870.2676.82
GaitPart [8]44.0160.6867.2573.47
GaitGL [27]47.2863.5669.3274.18
MTSGait [58]55.3271.2876.8581.55
Ours62.7276.5781.3285.24
", + "image_path": "70616d4e9ab9b0717aa965c7a424f88f1c2fe3479ba9762367f4a4a5cfad8757.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 688, + 265, + 701 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 688, + 265, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 688, + 265, + 701 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 706, + 279, + 730 + ], + "type": "text", + "content": "Evaluation on CASIA-B. Table 1 compares the performance of the proposed HSTL with seven state-of-the-art" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 300, + 305, + 531, + 423 + ], + "blocks": [ + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "lines": [ + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "spans": [ + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "type": "text", + "content": "Table 5: Rank-1 accuracy " + }, + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "type": "text", + "content": ", Rank-5 accuracy " + }, + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "type": "text", + "content": ", mAP " + }, + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 297, + 274, + 530, + 298 + ], + "type": "text", + "content": " and mINP on Gait3D." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 300, + 305, + 531, + 423 + ], + "lines": [ + { + "bbox": [ + 300, + 305, + 531, + 423 + ], + "spans": [ + { + "bbox": [ + 300, + 305, + 531, + 423 + ], + "type": "table", + "html": "
MethodsRank-1Rank-5mAPmINP
GaitSet [4]36.7058.3030.0117.30
GaitPart [8]28.2047.6021.5812.36
GLN [12]31.4052.9024.7413.58
GaitGL [27]29.7048.5022.2913.26
CSTL [15]11.7019.205.592.59
SMPLGait [59]46.3064.5037.1622.23
MTSGait [58]48.7067.1037.6321.92
Ours61.3076.3055.4834.77
", + "image_path": "b958a642034e427d47f8326f8a8b5804d5a10ec370dca672ce9d1a144f606adf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": "methods on the CASIA-B dataset. It can be seen that the proposed method obtains the best results in all three walking conditions while maintaining considerable stability across different views. The experimental results reveal that 1) the mean accuracy of the proposed method for the BG and CL walking conditions is " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "95.9\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "88.9\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": ", respectively, which are higher by " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": " than the second-best method (MetaGait [5]), demonstrating the advantage of our method in cross-view gait recognition. 2) The decrease in accuracy from NM to CL is " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": " for our method, compared to " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": " for 3D Local [16]. This indicates that the ARME module is more adaptable to various walking conditions. In addition, our method achieves an accuracy of " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "88.9\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": " in the challenging CL condition, which is " + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 296, + 450, + 531, + 730 + ], + "type": "text", + "content": " higher than 3D Local. This may be due to the complex clothing conditions that affect the part localization accuracy in 3D Local. 3) Both CSTL [15] and our method extracts multi-scale temporal features but in different ways. CSTL first extracts spatial information and then fuses three scales of motion features. In contrast, we propose an FTA module to aggregate spatiotemporal information from multiple body regions. Therefore, the average rank-1 accuracy of our method is higher or equivalent to that of CSTL in all views. This suggests that FTA can be more adaptive to spatio-temporal changes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19645" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 103, + 192, + 197 + ], + "blocks": [ + { + "bbox": [ + 48, + 103, + 192, + 197 + ], + "lines": [ + { + "bbox": [ + 48, + 103, + 192, + 197 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 192, + 197 + ], + "type": "image", + "image_path": "1b3df54efb68d4f7d1d4dd6f5e5575c510b7ee677da840905a1554140abb5d6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 200, + 164, + 209 + ], + "lines": [ + { + "bbox": [ + 76, + 200, + 164, + 209 + ], + "spans": [ + { + "bbox": [ + 76, + 200, + 164, + 209 + ], + "type": "text", + "content": "(a) Average rank-1 accuracy" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 143, + 278, + 197 + ], + "blocks": [ + { + "bbox": [ + 195, + 143, + 278, + 197 + ], + "lines": [ + { + "bbox": [ + 195, + 143, + 278, + 197 + ], + "spans": [ + { + "bbox": [ + 195, + 143, + 278, + 197 + ], + "type": "image", + "image_path": "8fcfaaf69f6c0eaa2d1ceeeff340112c9d6f6103cc441745270a83b708dd7a20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 200, + 267, + 208 + ], + "lines": [ + { + "bbox": [ + 204, + 200, + 267, + 208 + ], + "spans": [ + { + "bbox": [ + 204, + 200, + 267, + 208 + ], + "type": "text", + "content": "(b) Age distribution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 218, + 279, + 254 + ], + "lines": [ + { + "bbox": [ + 45, + 218, + 279, + 254 + ], + "spans": [ + { + "bbox": [ + 45, + 218, + 279, + 254 + ], + "type": "text", + "content": "Figure 4: Cross-age comparison results for the OUMVLP dataset. (a) Average rank-1 accuracy in adults vs children. (b) Distribution of two age groups." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "text", + "content": "Evaluation on OUMVLP. To verify the model's generalizability, we conduct experiments on the large-scale OUMVLP dataset. As shown in Table 3, our approach achieves competitive results in most views. In particular, the proposed method outperforms the second-best method (MetaGait) by an average of " + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "text", + "content": " under three extreme views, i.e., " + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "text", + "content": ", resulting in the best mean performance and cross-view stability. In addition, the OUMVLP dataset also provides annotations for the ages of the subjects. To evaluate the impact of age differences on recognition performance, we divided all subjects into two groups: adults (18-87 years old) and children (2-17 years old), as shown in Fig. 4(b). Fig. 4(a) presents the recognition accuracy based on age. It can be observed that, as adult sequences make up about " + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 45, + 274, + 279, + 506 + ], + "type": "text", + "content": " of the total sequences, all the compared methods show a bias toward the recognition results of adults. However, compared to other methods, our model effectively improves the accuracy of gait recognition for children, demonstrating the effectiveness of our hierarchical gait representation across ages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "content": "Evaluation on GREW and Gait3D. Gait3D and GREW are two recently introduced datasets that contain challenging conditions, such as misalignment of the human body and partial occlusion. Tables 4 and 5 show the results of the comparison between the proposed method and the state-of-the-art methods. Our method shows superior performance in all metrics, indicating its ability to effectively model gait characteristics in realistic scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 607, + 137, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 137, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 137, + 619 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "text", + "content": "Effectiveness of hierarchical feature extraction. To verify the effectiveness of our hierarchical gait partitioning, we conduct experiment with various grouping strategies. As shown in Table 2, this comparison only considers the different settings of the first three layers since a uniform partition is used at layer 4. Specifically, 1-1-1 indicates that the first three levels have the same number of groupings, and they are divided uniformly. " + }, + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "inline_equation", + "content": "1 - 2 - 4^{*}" + }, + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "text", + "content": " refers to the nonuniform division used in the proposed model. As shown in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 103, + 520, + 237 + ], + "blocks": [ + { + "bbox": [ + 309, + 103, + 520, + 237 + ], + "lines": [ + { + "bbox": [ + 309, + 103, + 520, + 237 + ], + "spans": [ + { + "bbox": [ + 309, + 103, + 520, + 237 + ], + "type": "image", + "image_path": "f157c36fffc6bc2e3a2fb3c459dafda257bdcb183813c7e78d9c9725f005e33f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 244, + 530, + 278 + ], + "lines": [ + { + "bbox": [ + 297, + 244, + 530, + 278 + ], + "spans": [ + { + "bbox": [ + 297, + 244, + 530, + 278 + ], + "type": "text", + "content": "Figure 5: Ablation study on the effectiveness of hierarchical feature extraction on the CASIA-B dataset (best viewed in color)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 300, + 334, + 530, + 428 + ], + "blocks": [ + { + "bbox": [ + 297, + 290, + 530, + 325 + ], + "lines": [ + { + "bbox": [ + 297, + 290, + 530, + 325 + ], + "spans": [ + { + "bbox": [ + 297, + 290, + 530, + 325 + ], + "type": "text", + "content": "Table 6: Ablation study on the effectiveness of ARME, ASTP, and FTA modules in terms of average rank-1 accuracy on the CASIA-B dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 300, + 334, + 530, + 428 + ], + "lines": [ + { + "bbox": [ + 300, + 334, + 530, + 428 + ], + "spans": [ + { + "bbox": [ + 300, + 334, + 530, + 428 + ], + "type": "table", + "html": "
SettingARMEASTPFTANMBGCLMean
35
a97.094.584.291.9
b✓97.895.385.993.0
c✓✓97.895.487.093.4
d✓✓✓97.795.387.593.5
e✓✓✓97.795.287.293.4
f✓✓✓97.895.487.593.6
g✓✓✓✓98.195.988.994.3
", + "image_path": "2e005c7e03450ec5dd125490067d690fe893c82e73cb938094dcc2f211e63023.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 448, + 530, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 448, + 530, + 506 + ], + "spans": [ + { + "bbox": [ + 296, + 448, + 530, + 506 + ], + "type": "text", + "content": "Fig. 5, the hierarchical feature extraction setting, e.g., 1-2-4, outperforms the other non-hierarchical approaches. A further mean performance improvement of " + }, + { + "bbox": [ + 296, + 448, + 530, + 506 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 296, + 448, + 530, + 506 + ], + "type": "text", + "content": " is achieved when the motion relationships between different body regions, i.e., 1-2-4*, are considered." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 506, + 530, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 506, + 530, + 635 + ], + "spans": [ + { + "bbox": [ + 296, + 506, + 530, + 635 + ], + "type": "text", + "content": "Effectiveness of ARME, ASTP, and FTA. The results of the ablation experiments for ARME, ASTP, and FTA are shown in Tab. 6 The results indicate that the ARME module significantly contributes to the improvement of the recognition accuracy, with an average improvement of " + }, + { + "bbox": [ + 296, + 506, + 530, + 635 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 296, + 506, + 530, + 635 + ], + "type": "text", + "content": " compared to the baseline model (non-grouping version of ARME). The mean accuracy is further improved by " + }, + { + "bbox": [ + 296, + 506, + 530, + 635 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 296, + 506, + 530, + 635 + ], + "type": "text", + "content": " when the ASTP module is integrated and multi-scale fusion is utilized in FTA. These results demonstrate the effectiveness and complementarity of the three modules in the proposed gait recognition framework." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 642, + 511, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 642, + 511, + 655 + ], + "spans": [ + { + "bbox": [ + 297, + 642, + 511, + 655 + ], + "type": "text", + "content": "4.5. Trade-off between accuracy and efficiency" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "content": "In this subsection, we evaluate the relationship between accuracy and efficiency for each compared method. As shown in Fig. 6, the 3D convolution-based approaches, such as GaitGL [27], 3D Local [16], LagrangeGait [3] and MetaGait [5], outperform the 2D convolution-based methods, like GaitSet [4] and GaitPart [8], in terms of accuracy, but" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19646" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 104, + 281, + 262 + ], + "blocks": [ + { + "bbox": [ + 47, + 104, + 281, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 104, + 281, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 104, + 281, + 262 + ], + "type": "image", + "image_path": "e37d6a628cb52e34fa5fd7fa799c99e534ef52eb2e2b158c8103251db87897ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 269, + 280, + 304 + ], + "lines": [ + { + "bbox": [ + 45, + 269, + 280, + 304 + ], + "spans": [ + { + "bbox": [ + 45, + 269, + 280, + 304 + ], + "type": "text", + "content": "Figure 6: The trade-off between accuracy and FLOPs of our method and other comparison methods on the OUMVLP dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 325, + 279, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 325, + 279, + 408 + ], + "spans": [ + { + "bbox": [ + 45, + 325, + 279, + 408 + ], + "type": "text", + "content": "at the cost of a significant increase in FLOPs (floating point operations). Our approach has a better trade-off between accuracy and efficiency. The main reason is that the proposed hierarchical learning architecture can extract multilevel motion features while reducing the number of 3D convolutions. More experimental results and ablation analysis are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 116, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 116, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 116, + 429 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 437, + 281, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 437, + 281, + 589 + ], + "spans": [ + { + "bbox": [ + 45, + 437, + 281, + 589 + ], + "type": "text", + "content": "This paper presents a hierarchical spatio-temporal representation learning (HSTL) framework for gait recognition. HSTL stacks multiple adaptive region-based motion extractors (ARMEs) and learns walking patterns in a coarse-to-fine manner. An adaptive spatio-temporal pooling (ASTP) module is proposed to perform hierarchical feature mapping for the output of each level of ARME. Additionally, a frame-level temporal aggregation module (FTA) is designed to compress local clips by fusing temporal information. The effectiveness of the proposed HSTL framework is demonstrated through extensive experiments conducted on four public datasets (CASIA-B, OUMVLP, GREW, and Gait3D)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 600, + 145, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 145, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 145, + 612 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 619, + 279, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 619, + 279, + 678 + ], + "spans": [ + { + "bbox": [ + 45, + 619, + 279, + 678 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 61972132, 62106065) and the Research Project for Self-cultivating Talents of Hebei Agricultural University (Grant No. PY201810)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 688, + 104, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 688, + 104, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 688, + 104, + 700 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 50, + 707, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 707, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 50, + 707, + 279, + 730 + ], + "type": "text", + "content": "[1] Weizhi An, Shiqi Yu, Yasushi Makihara, Xinhui Wu, Chi Xu, Yang Yu, Rijun Liao, and Yasushi Yagi. Performance evalua" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 298, + 106, + 532, + 729 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 317, + 106, + 531, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 531, + 138 + ], + "type": "text", + "content": "tion of model-based gait on multi-view very large population database with pose sequences. IEEE TBIOM, 2(4):421-430, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 139, + 531, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 139, + 531, + 182 + ], + "spans": [ + { + "bbox": [ + 302, + 139, + 531, + 182 + ], + "type": "text", + "content": "[2] Pia Bideau, Aruni RoyChowdhury, Rakesh R Menon, and Erik Learned-Miller. The best of both worlds: Combining cnns and geometric constraints for hierarchical motion segmentation. In CVPR, pages 508-517, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 183, + 532, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 183, + 532, + 224 + ], + "spans": [ + { + "bbox": [ + 302, + 183, + 532, + 224 + ], + "type": "text", + "content": "[3] Tianrui Chai, Annan Li, Shaoxiong Zhang, Zilong Li, and Yunhong Wang. Lagrange motion analysis and view embeddings for improved gait recognition. In CVPR, pages 20249-20258, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 225, + 532, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 225, + 532, + 266 + ], + "spans": [ + { + "bbox": [ + 302, + 225, + 532, + 266 + ], + "type": "text", + "content": "[4] Hanqing Chao, Yiwei He, Junping Zhang, and Jianfeng Feng. Gaitset: Regarding gait as a set for cross-view gait recognition. In AAAI, pages 8126-8133, 2019. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 267, + 532, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 267, + 532, + 311 + ], + "spans": [ + { + "bbox": [ + 302, + 267, + 532, + 311 + ], + "type": "text", + "content": "[5] Huanzhang Dou, Pengyi Zhang, Wei Su, Yunlong Yu, and Xi Li. Metagait: Learning to learn an omni sample adaptive representation for gait recognition. In ECCV, pages 357-374. Springer, 2022. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 311, + 532, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 311, + 532, + 343 + ], + "spans": [ + { + "bbox": [ + 302, + 311, + 532, + 343 + ], + "type": "text", + "content": "[6] Jessica Maria Echterhoff, Juan Haladjian, and Bernd Brügge. Gait and jump classification in modern equestrian sports. In ISWC, pages 88-91, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 344, + 531, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 344, + 531, + 385 + ], + "spans": [ + { + "bbox": [ + 302, + 344, + 531, + 385 + ], + "type": "text", + "content": "[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD, pages 226-231, 1996. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 386, + 531, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 386, + 531, + 428 + ], + "spans": [ + { + "bbox": [ + 302, + 386, + 531, + 428 + ], + "type": "text", + "content": "[8] Chao Fan, Yunjie Peng, Chunshui Cao, Xu Liu, Saihui Hou, Jiannan Chi, Yongzhen Huang, Qing Li, and Zhiqiang He. Gaitpart: Temporal part-based model for gait recognition. In CVPR, pages 14225-14233, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 429, + 531, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 429, + 531, + 471 + ], + "spans": [ + { + "bbox": [ + 302, + 429, + 531, + 471 + ], + "type": "text", + "content": "[9] Yang Fu, Yunchao Wei, Yuqian Zhou, Honghui Shi, Gao Huang, Xinchao Wang, Zhiqiang Yao, and Thomas Huang. Horizontal pyramid matching for person re-identification. In AAAI, pages 8295–8302, 2019. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 298, + 472, + 531, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 472, + 531, + 493 + ], + "spans": [ + { + "bbox": [ + 298, + 472, + 531, + 493 + ], + "type": "text", + "content": "[10] Jinguang Han and Bir Bhanu. Individual recognition using gait energy image. IEEE TPAMI, 28(2):316-322, 2005. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 494, + 531, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 494, + 531, + 536 + ], + "spans": [ + { + "bbox": [ + 299, + 494, + 531, + 536 + ], + "type": "text", + "content": "[11] Ruibing Hou, Hong Chang, Bingpeng Ma, Rui Huang, and Shiguang Shan. Bicnet-tks: Learning efficient spatial-temporal representation for video person re-identification. In CVPR, pages 2014–2023, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 537, + 531, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 537, + 531, + 578 + ], + "spans": [ + { + "bbox": [ + 299, + 537, + 531, + 578 + ], + "type": "text", + "content": "[12] Saihui Hou, Chunshui Cao, Xu Liu, and Yongzhen Huang. Gait lateral network: Learning discriminative and compact representations for gait recognition. In ECCV, pages 382-398, 2020. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 580, + 531, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 580, + 531, + 611 + ], + "spans": [ + { + "bbox": [ + 299, + 580, + 531, + 611 + ], + "type": "text", + "content": "[13] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Set residual network for silhouette-based gait recognition. IEEE TBIOM, 3(3):384-393, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 612, + 531, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 612, + 531, + 644 + ], + "spans": [ + { + "bbox": [ + 299, + 612, + 531, + 644 + ], + "type": "text", + "content": "[14] Saihui Hou, Xu Liu, Chunshui Cao, and Yongzhen Huang. Gait quality aware network: Toward the interpretability of silhouette-based gait recognition. IEEE TNNLS, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 644, + 531, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 644, + 531, + 687 + ], + "spans": [ + { + "bbox": [ + 299, + 644, + 531, + 687 + ], + "type": "text", + "content": "[15] Xiaohu Huang, Duowang Zhu, Hao Wang, Xinggang Wang, Bo Yang, Botao He, Wenyu Liu, and Bin Feng. Context-sensitive temporal feature learning for gait recognition. In ICCV, pages 12909-12918, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 687, + 531, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 531, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 531, + 729 + ], + "type": "text", + "content": "[16] Zhen Huang, Dixiu Xue, Xu Shen, Xinmei Tian, Houqiang Li, Jianqiang Huang, and Xian-Sheng Hua. 3d local convolutional neural networks for gait recognition. In ICCV, pages 14920-14929, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "19647" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 149 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 149 + ], + "type": "text", + "content": "[17] Munish Kumar, Navdeep Singh, Ravinder Kumar, Shubham Goel, and Krishan Kumar. Gait recognition based on vision systems: A systematic survey. Journal of Visual Communication and Image Representation, 75:103052, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 151, + 279, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 279, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 279, + 183 + ], + "type": "text", + "content": "[18] Worapan Kusakunniran. Review of gait recognition approaches and their challenges on view changes. IET Biometrics, 9(6):238-250, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 279, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 279, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 279, + 226 + ], + "type": "text", + "content": "[19] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE TIP, 31:5949-5962, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 279, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 279, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 279, + 271 + ], + "type": "text", + "content": "[20] Xiang Li, Yasushi Makihara, Chi Xu, and Yasushi Yagi. End-to-end model-based gait recognition using synchronized multi-view pose constraint. In ICCVW, pages 4106-4115, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 273, + 279, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 273, + 279, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 273, + 279, + 315 + ], + "type": "text", + "content": "[21] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, and Mingwu Ren. Gait recognition via semi-supervised disentangled representation learning to identity and covariate features. In CVPR, pages 13309-13319, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 318, + 279, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 318, + 279, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 318, + 279, + 349 + ], + "type": "text", + "content": "[22] Xiang Li, Yasushi Makihara, Chi Xu, Yasushi Yagi, Shiqi Yu, and Mingwu Ren. End-to-end model-based gait recognition. In ACCV, pages 3-20, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 351, + 279, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 351, + 279, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 351, + 279, + 372 + ], + "type": "text", + "content": "[23] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In CVPR, pages 510-519, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 373, + 279, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 279, + 416 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 279, + 416 + ], + "type": "text", + "content": "[24] Junhao Liang, Chao Fan, Saihui Hou, Chuanfu Shen, Yongzhen Huang, and Shiqi Yu. Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In ECCV, pages 375-390. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 418, + 279, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 279, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 279, + 449 + ], + "type": "text", + "content": "[25] Beibei Lin, Yu Liu, and Shunli Zhang. Gaitmask: Mask-based model for gait recognition. In BMVC, page 363, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 452, + 279, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 279, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 279, + 483 + ], + "type": "text", + "content": "[26] Beibei Lin, Shunli Zhang, and Feng Bao. Gait recognition with multiple-temporal-scale 3d convolutional neural network. In ACM MM, pages 3054-3062, 2020. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 485, + 279, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 279, + 527 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 279, + 527 + ], + "type": "text", + "content": "[27] Beibei Lin, Shunli Zhang, and Xin Yu. Gait recognition via effective global-local feature representation and local temporal aggregation. In ICCV, pages 14648-14656, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 529, + 279, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 279, + 550 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 279, + 550 + ], + "type": "text", + "content": "[28] Nini Liu and Yap-Peng Tan. View invariant gait recognition. In ICASSP, pages 1410-1413, 2010. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 552, + 279, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 279, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 279, + 584 + ], + "type": "text", + "content": "[29] Yasushi Makihara, Mark S Nixon, and Yasushi Yagi. Gait recognition: Databases, representations, and applications. Computer Vision: A Reference Guide, pages 1-13, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 586, + 279, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 279, + 618 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 279, + 618 + ], + "type": "text", + "content": "[30] Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, and Yoichi Sato. Hierarchical gaussian descriptor for person re-identification. In CVPR, pages 1363–1372, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 619, + 279, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 619, + 279, + 660 + ], + "spans": [ + { + "bbox": [ + 48, + 619, + 279, + 660 + ], + "type": "text", + "content": "[31] Daigo Muramatsu, Yasushi Makihara, Haruyuki Iwama, Takuya Tanoue, and Yasushi Yagi. Gait verification system for supporting criminal investigation. In IAPR, pages 747-748, 2013. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 663, + 279, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 663, + 279, + 695 + ], + "spans": [ + { + "bbox": [ + 48, + 663, + 279, + 695 + ], + "type": "text", + "content": "[32] Filip Radenović, Giorgos Tolias, and Ondřej Chum. Finetuning cnn image retrieval with no human annotation. IEEE TPAMI, 41(7):1655-1668, 2018. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 697, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 697, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 48, + 697, + 279, + 729 + ], + "type": "text", + "content": "[33] Imad Rida, Noor Almaadeed, and Somaya Almaadeed. Robust gait recognition: a comprehensive survey. IET Biometrics, 8(1):14-28, 2019. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 729 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 128 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 128 + ], + "type": "text", + "content": "[34] Alireza Sepas-Moghaddam and Ali Etemad. Deep gait recognition: A survey. IEEE TPAMI, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 130, + 530, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 130, + 530, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 130, + 530, + 171 + ], + "type": "text", + "content": "[35] Kohei Shiraga, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Geinet: View-invariant gait recognition using a convolutional neural network. In ICB, pages 1-8, 2016. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 174, + 530, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 174, + 530, + 206 + ], + "spans": [ + { + "bbox": [ + 299, + 174, + 530, + 206 + ], + "type": "text", + "content": "[36] Pranjay Shyam, Kuk-Jin Yoon, and Kyung-Soo Kim. Adversarily-trained hierarchical feature extractor for vehicle re-identification. In ICRA, pages 13400–13407, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 207, + 530, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 207, + 530, + 239 + ], + "spans": [ + { + "bbox": [ + 299, + 207, + 530, + 239 + ], + "type": "text", + "content": "[37] Chunfeng Song, Yongzhen Huang, Yan Huang, Ning Jia, and Liang Wang. Gaitnet: An end-to-end network for gait based human identification. PR, 96:106988, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 241, + 530, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 241, + 530, + 283 + ], + "spans": [ + { + "bbox": [ + 299, + 241, + 530, + 283 + ], + "type": "text", + "content": "[38] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In ECCV, pages 480-496, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 286, + 530, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 286, + 530, + 328 + ], + "spans": [ + { + "bbox": [ + 299, + 286, + 530, + 328 + ], + "type": "text", + "content": "[39] Noriko Takemura, Yasushi Makihara, Daigo Muramatsu, Tomio Echigo, and Yasushi Yagi. Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ TCVA, 10:1-14, 2018. 2, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 330, + 530, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 330, + 530, + 362 + ], + "spans": [ + { + "bbox": [ + 299, + 330, + 530, + 362 + ], + "type": "text", + "content": "[40] Lei Tan, Pingyang Dai, Rongrong Ji, and Yongjian Wu. Dynamic prototype mask for occluded person re-identification. In ACM MM, pages 531–540, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 364, + 530, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 364, + 530, + 405 + ], + "spans": [ + { + "bbox": [ + 299, + 364, + 530, + 405 + ], + "type": "text", + "content": "[41] Torben Teepe, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Towards a deeper understanding of skeleton-based gait recognition. In CVPRW, pages 1569-1577, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 408, + 530, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 408, + 530, + 450 + ], + "spans": [ + { + "bbox": [ + 299, + 408, + 530, + 450 + ], + "type": "text", + "content": "[42] Torben Teepe, Ali Khan, Johannes Gilg, Fabian Herzog, Stefan Hörmann, and Gerhard Rigoll. Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In ICIP, pages 2314-2318, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 452, + 530, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 452, + 530, + 495 + ], + "spans": [ + { + "bbox": [ + 299, + 452, + 530, + 495 + ], + "type": "text", + "content": "[43] Md Uddin, Daigo Muramatsu, Noriko Takemura, Md Ahad, Atiqur Rahman, Yasushi Yagi, et al. Spatio-temporal silhouette sequence reconstruction for gait recognition against occlusion. IPSJ TCVA, 11(1):1-18, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 497, + 530, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 497, + 530, + 518 + ], + "spans": [ + { + "bbox": [ + 299, + 497, + 530, + 518 + ], + "type": "text", + "content": "[44] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 299, + 520, + 530, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 520, + 530, + 561 + ], + "spans": [ + { + "bbox": [ + 299, + 520, + 530, + 561 + ], + "type": "text", + "content": "[45] Zhikang Wang, Lihuo He, Xiaoguang Tu, Jian Zhao, Xinbo Gao, Shengmei Shen, and Jiashi Feng. Robust video-based person re-identification by hierarchical mining. IEEE TCSVT, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 299, + 564, + 530, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 564, + 530, + 606 + ], + "spans": [ + { + "bbox": [ + 299, + 564, + 530, + 606 + ], + "type": "text", + "content": "[46] Xiu-Shen Wei, Chen-Lin Zhang, Lingqiao Liu, Chunhua Shen, and Jianxin Wu. Coarse-to-fine: A rnn-based hierarchical attention model for vehicle re-identification. In ACCV, pages 575–591, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 299, + 608, + 530, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 608, + 530, + 650 + ], + "spans": [ + { + "bbox": [ + 299, + 608, + 530, + 650 + ], + "type": "text", + "content": "[47] Jiachen Yang, Jianxiong Zhou, Dayong Fan, and Haibin Lv. Design of intelligent recognition system based on gait recognition technology in smart transportation. Multimedia Tools and Applications, 75(24):17501-17514, 2016. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 299, + 653, + 530, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 653, + 530, + 685 + ], + "spans": [ + { + "bbox": [ + 299, + 653, + 530, + 685 + ], + "type": "text", + "content": "[48] Lingxiang Yao, Worapan Kusakunniran, Qiang Wu, Jingsong Xu, and Jian Zhang. Collaborative feature learning for gait recognition under cloth changes. IEEE TCSVT, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 299, + 687, + 530, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 530, + 729 + ], + "type": "text", + "content": "[49] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE TPAMI, 44(6):2872-2893, 2021. 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "19648" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 553 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 139 + ], + "type": "text", + "content": "[50] TzeWei Yeoh, Hernan E Aguirre, and Kiyoshi Tanaka. Clothing-invariant gait recognition using convolutional neural network. In ISPACS, pages 1-5, 2016. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 279, + 182 + ], + "type": "text", + "content": "[51] Shiqi Yu, Haifeng Chen, Qing Wang, Linlin Shen, and Yongzhen Huang. Invariant feature extraction for gait recognition using only one uniform model. Neurocomputing, 239:81-93, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 183, + 279, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 279, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 279, + 225 + ], + "type": "text", + "content": "[52] Shiqi Yu, Daoliang Tan, and Tieniu Tan. A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In ICPR, pages 441-444, 2006. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 227, + 279, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 279, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 279, + 269 + ], + "type": "text", + "content": "[53] Anguo Zhang, Yueming Gao, Yuzhen Niu, Wenxi Liu, and Yongcheng Zhou. Coarse-to-fine person re-identification with auxiliary-domain classification and second-order information bottleneck. In CVPR, pages 598–607, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 271, + 279, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 279, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 279, + 313 + ], + "type": "text", + "content": "[54] Mingyang Zhang, Yang Xiao, Fu Xiong, Shuai Li, Zhiguo Cao, Zhiwen Fang, and Joey Tianyi Zhou. Person re-identification with hierarchical discriminative spatial aggregation. IEEE TIFS, 17:516-530, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 314, + 279, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 279, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 279, + 346 + ], + "type": "text", + "content": "[55] Shaoxiong Zhang, Yunhong Wang, and Annan Li. Crossview gait recognition with deep universal linear embeddings. In CVPR, pages 9095-9104, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 347, + 279, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 279, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 279, + 379 + ], + "type": "text", + "content": "[56] Yuqi Zhang, Yongzhen Huang, Shiqi Yu, and Liang Wang. Cross-view gait recognition by discriminative feature learning. IEEE TIP, 29:1001-1015, 2019. 1, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 380, + 279, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 279, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 279, + 421 + ], + "type": "text", + "content": "[57] Ziyuan Zhang, Luan Tran, Xi Yin, Yousef Atoum, Xiaoming Liu, Jian Wan, and Nanxin Wang. Gait recognition via disentangled representation learning. In CVPR, pages 4710-4719, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 423, + 279, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 279, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 279, + 465 + ], + "type": "text", + "content": "[58] Jinkai Zheng, Xinchen Liu, Xiaoyan Gu, Yaoqi Sun, Chuang Gan, Jiyong Zhang, Wu Liu, and Chenggang Yan. Gait recognition in the wild with multi-hop temporal switch. In ACM MM, pages 6136-6145, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 467, + 279, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 279, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 279, + 509 + ], + "type": "text", + "content": "[59] Jinkai Zheng, Xinchen Liu, Wu Liu, Lingxiao He, Chenggang Yan, and Tao Mei. Gait recognition in the wild with dense 3d representations and a benchmark. In CVPR, pages 20228-20237, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 511, + 279, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 279, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 279, + 553 + ], + "type": "text", + "content": "[60] Zheng Zhu, Xianda Guo, Tian Yang, Junjie Huang, Jiankang Deng, Guan Huang, Dalong Du, Jiwen Lu, and Jie Zhou. Gait recognition in the wild: A benchmark. In ICCV, pages 14789-14799, 2021. 2, 5" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "text", + "content": "19649" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_content_list.json b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..49b1d6f0b051cb702d818f16997a247ca1146b92 --- /dev/null +++ b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_content_list.json @@ -0,0 +1,1560 @@ +[ + { + "type": "text", + "text": "Hierarchical Visual Categories Modeling: A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection", + "text_level": 1, + "bbox": [ + 85, + 161, + 884, + 201 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinglun Li $^{1}$ Xinyu Zhou $^{2}$ Pinxue Guo $^{1}$ Yixuan Sun $^{1}$ Yiwen Huang $^{2}$", + "bbox": [ + 179, + 226, + 789, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weifeng Ge²† Wenqiang Zhang¹,²†", + "bbox": [ + 344, + 245, + 631, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Academy for Engineering and Technology, Fudan University", + "bbox": [ + 240, + 261, + 727, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computer Science, Fudan University", + "bbox": [ + 284, + 277, + 663, + 293 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jingli960423@gmail.com,zhouxinyu20@fudan.edu.cn", + "bbox": [ + 268, + 296, + 697, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pxguo21，21210860014，21210240056}@m.fudan.edu.cn,", + "bbox": [ + 260, + 312, + 705, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "weifeng.ge.ic@gmail.com, wqzhang@fudan.edu.cn", + "bbox": [ + 285, + 328, + 682, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 372, + 312, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting out-of-distribution inputs for visual recognition models has become critical in safe deep learning. This paper proposes a novel hierarchical visual category modeling scheme to separate out-of-distribution data from in-distribution data through joint representation learning and statistical modeling. We learn a mixture of Gaussian models for each in-distribution category. There are many Gaussian mixture models to model different visual categories. With these Gaussian models, we design an in-distribution score function by aggregating multiple Mahalanobis-based metrics. We don't use any auxiliary outlier data as training samples, which may hurt the generalization ability of out-of-distribution detection algorithms. We split the ImageNet-1k dataset into ten folds randomly. We use one fold as the in-distribution dataset and the others as out-of-distribution datasets to evaluate the proposed method. We also conduct experiments on seven popular benchmarks, including CIFAR, iNaturalist, SUN, Places, Textures, ImageNet-O, and OpenImage-O. Extensive experiments indicate that the proposed method outperforms state-of-the-art algorithms clearly. Meanwhile, we find that our visual representation has a competitive performance when compared with features learned by classical methods. These results demonstrate that the proposed method hasn't weakened the discriminative ability of visual recognition models and keeps high efficiency in detecting out-of-distribution samples.", + "bbox": [ + 75, + 403, + 472, + 764 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 790, + 208, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern deep neural networks have shown strong generalization ability when training and test data are from the", + "bbox": [ + 75, + 814, + 468, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "same distribution [44, 18, 51, 10, 33]. However, encountering unexpected scenarios is inevitable in real-world applications. Thus assuring that training and test data share the same distribution becomes problematic. In applications like autonomous driving [3, 4] and medical image analysis [55, 14, 43], it is critical for models to identify inputs beyond their recognition capacity – known as out-of-distribution (OOD) detection. OOD detection algorithms can enable the system to warn humans promptly in many safety-related scenarios. Moreover, it has become an important research topic in the research community of safe artificial intelligence [21, 29, 23, 32, 56].", + "bbox": [ + 499, + 374, + 890, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many popular OOD detection methods aim to build probability models to describe training distributions [29, 56, 46, 24, 23]. With these probability models, they built a score function that can calculate in-distribution scores for test samples. These in-distribution scores reflect whether these samples fall into a given distribution. Then the test sample can be evaluated by the score function to decide whether it is an OOD sample or not. Thus modeling features of in-distribution data become extremely important. Previous works [29, 16, 5, 12, 38] build a distribution over the whole training data. Since training images may come from various visual categories, the decision boundary between In-Distribution (InD) and OOD data becomes extremely complex. To solve this problem, subsequent studies [24, 7, 15, 56] decomposed the whole dataset into several subgroups to simplify the decision boundary. Although representative algorithms like MOS [24], have gotten impressive performance in identifying OOD samples, they failed to detect near OOD samples. Because when different visual categories are grouped together, the OOD decision boundary will become even more uncertain.", + "bbox": [ + 497, + 543, + 892, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A typical framework for out-of-distribution detection involves two key steps: 1) learning a compact feature rep", + "bbox": [ + 499, + 838, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 253, + 48, + 789, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^\\dagger$ indicates corresponding authors.", + "bbox": [ + 95, + 852, + 273, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23425", + "bbox": [ + 463, + 887, + 505, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representation that can fit probability models easily; 2) modeling features of in-distribution data in complex distributions accurately. The above two problems are mutually connected because more compact features will make modeling the data distribution easier, and stronger probability modeling techniques will exploit fewer restrictions on representation learning. However, achieving the above goals is difficult, even when there are a lot of breakthroughs in deep learning. Because if training samples from the same category are too close in the feature space, it will usually lead to overfitting. Meanwhile, in-distribution samples may come from different visual categories that have large variations in appearance and semantic information, which makes modeling complex training distributions become challenging even for excellent statisticians.", + "bbox": [ + 75, + 125, + 472, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new out-of-distribution detection framework, called hierarchical visual category modeling, to solve the above two issues simultaneously. We hold an assumption that given a training set that contains multiple visual categories, we can learn a probability model for every category independently. The out-of-distribution detection problem can be solved easily by aggregating probability models of known categories. Our motivation is that decomposing the whole dataset into subsets and modeling each category independently can avoid finding common characteristics shared by different categories. However, modeling an individual visual category is still challenging since classical supervised learning won't lead to compact feature representation. Thus, for each input sample, we need to force its feature representation to match the corresponding statistical model. That means we need to jointly conduct density estimation and representation learning. If we can jointly learn visual representations and optimize statistical models end-to-end, we can get good feature representations falling into distributions of the corresponding visual categories. Besides, we exploit knowledge distillation as done in [6] to learn robust feature representation. In this way, we can describe the complex training distribution with multiple Gaussian mixture models while not impairing the generalization ability of visual features.", + "bbox": [ + 75, + 337, + 472, + 682 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In practice, to learn visual concepts that are in complex distributions, we build a Gaussian mixture model (GMM [39]) for each visual category. Given input samples, we extract their deep features and project these features into a high-dimensional attribute space. Different from classical Gaussian mixture models that send the same input into $K$ different Gaussian models, we divide the attribute space into multiple groups and build a Gaussian model in each group independently. This strategy can give every attribute group a clear learning target and lead to better convergence. Experimental results indicate that this strategy works quite well. After the visual representation learning and statistical model parameters optimization, we can directly aggre", + "bbox": [ + 75, + 687, + 472, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gate these statistical models to judge whether a test sample comes from the training distribution or not. To evaluate our OOD detector, we split ImageNet into ten folds randomly and select one of these splits as the training set and all other splits as the OOD dataset to conduct extensive tests. Experiments indicate that the proposed method has a strong ability to identify OOD samples. We also evaluate our method on seven popular OOD benchmarks. Experimental results demonstrate that the proposed method not only can identify OOD samples efficiently but also improves the discriminative ability of learned visual representations.", + "bbox": [ + 497, + 125, + 894, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper are summarized as follows:", + "bbox": [ + 497, + 277, + 894, + 305 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new out-of-distribution detection scheme, called hierarchical visual category modeling to conduct joint representation learning and density estimation. It provides a new perspective for out-of-distribution detection to learn visual representation and probability models end-to-end.", + "- We exploit multiple Gaussian mixture models to model visual concepts in complex distributions. Visual attributes are divided into subgroups and modeled by different Gaussian components, which makes parameter learning much more efficient.", + "- We conducted comprehensive experiments and ablation studies on popular benchmarks to investigate the effectiveness of the proposed method. Experiments demonstrated that our out-of-distribution detection models achieve better performance clearly when compared with previous methods." + ], + "bbox": [ + 517, + 315, + 890, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 593, + 640, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Out-Of-Distribution Detection. Out-of-distribution detection aims to distinguish out-of-distribution samples from in-distribution data. Numerous methods have been proposed for OOD detection. Maximum softmax probability (MSP) [21] has been recognized as a strong baseline by using the maximum score across all classes as an OOD score. ODIN [30] improves MSP by perturbing the inputs and adjusting the logits via rescaling. Gaussian discriminant analysis has been employed in [29, 56] to detect OOD samples. ReAct [45] uses rectified activation to reduce model overconfidence in OOD data. Shama et al. in [42] utilized Gram matrices to measure feature correlations for OOD detection. Bibas et al. in [2] proposed pNML regret to detect OOD samples with a single-layer neural network. The Generalized-ODIN approach [23] decomposes the confidence of class probability using a dividend/divisor structure to incorporate prior knowledge. Another promising line of work focuses on designing new learning objectives to train", + "bbox": [ + 497, + 617, + 894, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23426", + "bbox": [ + 463, + 887, + 505, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/760d42cd0d1840d52c8fc98770bd29c917eefb93770bdbbe6ff06e498505f1a5.jpg", + "image_caption": [ + "Figure 1. Illustration of the training pipeline of hierarchical visual category modeling, written as HVCM. In HVCM, we jointly learn the visual representation and parameters of probabilistic models. We get two different views of an input image and send the outputs into a knowledge distillation framework (DINO [6]). Image representations are projected into a high-dimensional attribute space. Then these attributes are divided into different groups and pass SoftMax functions to get attribute distributions. We match attributes in each group with stored attribute centers of the target visual category. The whole model is trained in an end-to-end manner." + ], + "image_footnote": [], + "bbox": [ + 85, + 124, + 884, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "deep models from scratch. Leave-out Classifiers [52] introduces a margin loss to ensure that InD and OOD samples are separated in the feature space. Lee et al. in [28] proposed a novel confidence loss to give lower confidence for OOD samples. There are also some methods [58, 41, 35] designed to conduct OOD detection based on generative models. Unlike these methods, our proposed method jointly conducts representation learning and density estimation to improve the OOD detection performance and keep the learned feature with strong discriminative ability.", + "bbox": [ + 75, + 454, + 472, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Density Estimation in Deep Learning. Density estimation tries to describe the probabilistic density distribution of observed data accurately and has been investigated in deep learning for a long time [47, 11, 31]. In [1], Chong et al. employed variational autoencoders for anomaly detection, where the reconstruction probability was used to compute the anomaly score of each sample. Papamakarios et al. [36] proposed a novel method for density estimation based on masked autoregressive flow. Zhou et al. [62] extended variational autoencoders by incorporating a mixture of Gaussians to model the latent space distribution, allowing for more flexible and expressive representations. Yang et al. [13] combined Flow-based generative models with generative adversarial networks for density estimation and sample generation. Zhao et al. [11] used discrete latent variables to conduct density estimation, which has been applied to many research topics in natural language processing and image processing. In this paper, we build probabilistic models that can be jointly learned with deep networks. We wish", + "bbox": [ + 75, + 603, + 472, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "latent representations of the visual categories could easily be modeled by Gaussian mixture models even when complex distributions exist. Besides, we do not use post-hoc methods to conduct feature learning and density estimation separately. We aim to use the probabilistic model to guide the feature-learning process of deep networks.", + "bbox": [ + 497, + 454, + 890, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 499, + 555, + 591, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework", + "text_level": 1, + "bbox": [ + 499, + 580, + 626, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State-of-the-art methods [21, 30, 29, 24, 46, 32] typically assume all InD data follow the same distribution. However, such an assumption will lead to the decision boundary of an OOD detector being doped with some uncertain space. In this paper, we avoid modeling the whole InD dataset and only focus on independently modeling each object category's distribution. We propose hierarchical visual categories modeling to achieve this goal while maintaining high classification accuracy. In hierarchical visual categories modeling, we first project image features into a high-dimensional attribute space (usually 8192 dimensions). These attributes can be grouped into multiple sub-visual concepts as components of an image category which can be easily modeled by Gaussian distributions. Then combinations of multiple sub-visual concepts (abbreviated as sub-concepts) can be grouped to describe a more complex visual concept of an in-distribution category. Since we define visual categories based on sub-visual concepts, simply modeling distributions of these sub-concepts and de", + "bbox": [ + 497, + 603, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23427", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scribing visual categories with these sub-concepts hierarchically can model complicated training distributions.", + "bbox": [ + 75, + 125, + 467, + 153 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given a visual recognition model $f$ , it maps an input image $\\mathbf{x}$ with label $y$ to a high-dimensional feature vector $z \\in \\mathbb{R}^q$ . As described above, we project $z$ into an attribute space $S \\in \\mathbb{R}^d$ with a higher dimension and get an attribute description $\\mathbf{a}$ . Attributes in $\\mathbf{a}$ can be grouped into multiple attribute subgroups $\\{a_i\\}_{i=1}^G$ where $a_i \\in S_i \\subset \\mathbb{R}^{d/G}$ and $i \\in \\{1,2,\\dots,G\\}$ . For images from a visual category $c$ , we assume that attributes in their $i$ -th attribute group follow a simple Gaussian distribution $\\mathcal{N}(\\mu_i^c, \\Sigma_i^c)$ , where $\\mu_i^c$ and $\\Sigma_i^c$ are the mean and variance respectively. Since attributes are divided into $G$ groups, we will have $G$ different Gaussian distributions for each visual category. The benefits of dividing the whole attribute space into several subspaces come from two folds: (1) modeling attribute distributions in these groups becomes easier; (2) assemble of distributions in these attribute groups can describe complex distributions, which will lead to more accurate decision boundaries. To combine all attributes to describe every visual category, our OOD detector learns weights $\\{w_i^c\\}_{i=1}^G$ of all attribute groups through exponential moving averages. Therefore, for each in-distribution class $c$ , we can model the corresponding visual concept in a probability perspective with a mixture of Gaussian models:", + "bbox": [ + 75, + 154, + 472, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {x}; c\\right) = \\sum_ {i = 1} ^ {G} w _ {i} ^ {c} \\mathcal {N} \\left(\\boldsymbol {a} _ {i}; \\boldsymbol {\\mu} _ {i} ^ {c}, \\boldsymbol {\\Sigma} _ {i} ^ {c}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 159, + 480, + 468, + 519 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}^{c}\\in \\mathbb{R},\\pmb{\\mu}_{i}^{c}\\in \\mathbb{R}^{d / G},\\pmb{\\Sigma}_{i}^{a}\\in \\mathbb{R}^{d / G\\times d / G}$ and $\\mathcal{N}(\\cdot)$ means a normal distribution. We build a Gaussian mixture model for each class and get $C$ different Gaussian mixture models. Then, our proposed HVCM focuses on training deep neural networks to learn image features that follow the above distributions and parameters of these probability models jointly.", + "bbox": [ + 75, + 524, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the above Gaussian probability models, given a test sample $\\pmb{x}^{\\prime}$ , we define the score function $g(\\pmb{x}^{\\prime};\\pmb{w}^{c},\\pmb{\\mu}^{c},\\pmb{\\Sigma}^{c})$ to measure whether it belongs to the $c$ -th visual category using the learned probability density function. Here, $\\pmb{w}^{c} = \\{w_{i}^{c}\\}_{i = 1}^{G}$ , $\\pmb{\\mu}^{c} = \\{\\pmb{\\mu}_{i}^{c}\\}_{i = 1}^{G}$ and $\\pmb{\\Sigma}^{c} = \\{\\pmb{\\Sigma}_{i}^{c}\\}_{i = 1}^{G}$ . This score function can be used as a reliable metric to detect OOD samples:", + "bbox": [ + 75, + 623, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh \\left(\\boldsymbol {x} ^ {\\prime}\\right) = \\left\\{ \\begin{array}{l l} \\operatorname {I n D}, & \\text {i f} \\min _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) \\geq \\gamma , \\\\ \\operatorname {O O D}, & \\text {i f} \\min _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) < \\gamma , \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 95, + 729, + 468, + 774 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ are thresholds to be determined in subsequent sections. Eq. (2) indicates that we use the minimal InD score among all $C$ categories to make the final decision.", + "bbox": [ + 75, + 783, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our framework is illustrated in Figure 1, where there are two steps: (1) jointly learn deep features that fit our probability models and parameters of these models; (2) calculate", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the InD score based on a set of Gaussian mixture models as the metric for out-of-distribution detection. In the following subsections, we give more details.", + "bbox": [ + 497, + 125, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Joint Visual Representation Learning and Parameter Optimization of Probability Models", + "text_level": 1, + "bbox": [ + 497, + 175, + 890, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn visual representations that follow Gaussian mixture models and keep their discriminative ability at the same time, we exploit the knowledge distillation framework DINO [6] to perform the joint learning. As shown in Figure 1, for an image $\\pmb{x}$ , we sample ten different views of $\\pmb{x}$ and send them into teacher and student branches simultaneously to perform self-distillation. During knowledge distillation, we project ResNet50 [19] features in 2048 dimensions into an attribute space with dimension $d(= 8192)$ . Apart from the learning objective produced by self-distillation, we force the attributes of each class to follow a class-specific Gaussian mixture model. We divide the attributes $\\pmb{a} \\in \\mathbb{R}^d$ of $\\pmb{x}$ into $G$ groups to learn the Gaussian mixture model parameters. However, in practice, it's too hard to directly learn the mean and variance of Gaussian models. We follow He et. al [20] to learn attribute centers $\\{\\pmb{\\mu}_i^c\\}_{i = 1}^G$ of the $c$ -th category (the image label $y$ is $c$ ). A linear projection layer is exploited to predict the weights $\\{w_i^c (\\pmb {x})\\}_{i = 1}^G$ of $\\pmb{x}$ on all $G$ attribute groups. Our learning objective can be written as follows:", + "bbox": [ + 497, + 209, + 892, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {K D} + \\alpha \\sum_ {i = 1} ^ {G} \\mathrm {K L} \\left(\\boldsymbol {a} _ {i} \\mid \\mid \\boldsymbol {\\mu} _ {i} ^ {c}\\right) + \\beta \\sum_ {i = 1} ^ {T} w _ {i} ^ {c} (\\boldsymbol {x}) \\mathrm {K L} \\left(\\boldsymbol {\\mu} _ {i} ^ {c} \\mid \\mid \\boldsymbol {a} _ {i}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 510, + 473, + 890, + 519 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{KD}$ stands for the cross entropy loss in self-distillation, KL stands for the Kullback-Leibler(KL) divergence, and $\\alpha$ and $\\beta$ are hyperparameters. The reason why we exploit two KL divergences is that experiments indicate that the term $\\mathrm{KL}(\\pmb{a}_i||\\pmb{\\mu}_i^c)$ will favor learning attribute centers and the term $\\mathrm{KL}(\\pmb{\\mu}_i^c||\\pmb{a}_i)$ is more suitable to learn image attribute descriptions and group weights. Note that softmax operations normalize attributes and their learnable centers in each group before calculating learning objectives.", + "bbox": [ + 497, + 527, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the learning objective in Eq. (3), student network parameters $\\theta_{s}$ , and weights $\\{w_i^c\\}_{i = 1}^G$ and attribute centers $\\{\\pmb {\\mu}_i^c\\}_{i = 1}^G$ of all groups are learnt at the same time. They are updated using the following equations:", + "bbox": [ + 497, + 652, + 890, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {s} ^ {t + 1} = \\theta_ {s} ^ {t} - \\gamma_ {1} \\frac {\\partial \\mathcal {L}}{\\partial \\theta_ {s} ^ {t}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 718, + 890, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {i} ^ {c, t + 1} = \\boldsymbol {\\mu} _ {i} ^ {c, t} - \\gamma_ {2} \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {\\mu} _ {i} ^ {c , t}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 749, + 890, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {i} ^ {c, t + 1} = (1 - \\gamma_ {3}) \\boldsymbol {w} _ {i} ^ {c, t} + \\gamma_ {3} w _ {i} ^ {c} (\\boldsymbol {x}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 574, + 782, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{1},\\gamma_{2}$ and $\\gamma_{3}$ are the learning rates. $\\gamma_{1},\\gamma_{2}$ control the gradient update speed, while $\\gamma_{3}$ controls the updating speed of the group weights. Following He et. al [20], $\\theta_s,\\{\\pmb {w}^c\\}_{c = 1}^C$ and $\\{\\pmb {\\mu}^c\\}_{c = 1}^C$ are initialized with Gaussian noises. We use", + "bbox": [ + 497, + 809, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23428", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86bd137e38afbec0ea0d638fb82abf313231d1da57021098cda98e3b8d685939.jpg", + "image_caption": [ + "Figure 2. Illustration of attribute group visualization and Mahalanobis distance distribution. Both attribute groups are visualized by t-SNE [49]. The colors encode different in-distribution data(ImageNet), and out-of-distribution features(SUN) marked as gray points. Models are trained on ResNet-50 [18] using DINO(a) and HVCM(b)." + ], + "image_footnote": [], + "bbox": [ + 80, + 123, + 467, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adam optimizer [25] with momentum to update both $\\theta_{s}$ and $\\{\\pmb{\\mu}^c\\}_{c=1}^C$ . While the attribute weights are learned through exponential moving averages [17].", + "bbox": [ + 75, + 440, + 468, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Out-of-Distribution Detection based on Probability Models in Hierarchy", + "text_level": 1, + "bbox": [ + 75, + 492, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the training of our hierarchical probability model converges, we will get visual attributes that follow learned Gaussian distributions hierarchically for in-distribution samples. Meanwhile, we also get weights and centers of all attribute groups for each class. However, the mean attributes are updated frequently during the training and are thus not suitable to be used as group centers. So we need to recalculate the attribute centers for each visual category. Given an in-distribution visual category $c$ , we estimate the mean vector and covariation matrix of the $i$ -th attribute group:", + "bbox": [ + 75, + 529, + 468, + 682 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {i} ^ {c} = \\frac {1}{N _ {c}} \\sum_ {m = 1} ^ {N _ {c}} \\boldsymbol {a} _ {i} ^ {m} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 694, + 468, + 730 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {i} ^ {c} = \\frac {1}{N _ {c} - 1} \\sum_ {m = 1} ^ {N _ {c}} \\left(\\boldsymbol {a} _ {i} ^ {m} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) \\left(\\boldsymbol {a} _ {i} ^ {m} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) ^ {\\top}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 733, + 468, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{c}$ notes the number of samples in the $c$ -th class, and $\\pmb{a}_i^m$ is the sub-attribute vector of the $m$ -th sample. With these weights, means, and covariances, we can describe each visual category in the hierarchy accurately. We try to use the probability density function in Eq. (1) as the in-distribution function. However, we will encounter the", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "problem of numerical overflow when calculating the determinants of covariance matrices as [48]. Instead, we compute the Mahalanobis distance between the $i$ -th sub-visual attributes $a_{i}^{\\prime}$ of a test sample $\\pmb{x}^{\\prime}$ and the corresponding attribute center $\\mu_{i}^{c}$ to measure the likelihood of these attributes belongs to the target category:", + "bbox": [ + 497, + 125, + 890, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} ^ {c} (\\boldsymbol {x}) = - \\left(\\boldsymbol {a} _ {i} ^ {\\prime} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) ^ {\\top} \\left(\\boldsymbol {\\Sigma} _ {i} ^ {c}\\right) ^ {- 1} \\left(\\boldsymbol {a} _ {i} ^ {\\prime} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 217, + 890, + 235 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above equation gives the in-distribution measure for one attribute group. While, for every category, we have multiple attribute groups and need to assemble related in-distribution measures to get the class-level in-distribution score. Since we have gotten the weights of attribute groups for each category, we can easily assemble them and get the class-level score function:", + "bbox": [ + 497, + 244, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) = \\sum_ {i = 1} ^ {G} w _ {i} ^ {c} M _ {i} ^ {c} \\left(\\boldsymbol {x} ^ {\\prime}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 349, + 890, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this score function, we can easily get the in-distribution score of a test sample on each visual category. Since there are $C$ categories in the whole in-distribution dataset, we get the maximal in-distribution score across different visual categories as the in-distribution score on the whole dataset:", + "bbox": [ + 497, + 396, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(\\boldsymbol {x} ^ {\\prime}\\right) = \\max _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 489, + 890, + 509 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A high in-distribution score $g(\\boldsymbol{x}^{\\prime})$ indicates that the semantic attributes of a test sample lie very close to one or multiple in-distribution visual categories (as shown in Figure 2). On the contrary, if a sample does not belong to the previously modeled categories, it will get a low in-distribution score. We follow Eq. (2) to set thresholds to judge whether a sample is an out-of-distribution sample. In the experiment section, we discuss how to set these thresholds.", + "bbox": [ + 497, + 519, + 890, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 499, + 642, + 633, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 499, + 665, + 690, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In-distribution Datasets. We use ImageNet-1K [40] and CIFAR10 [27] as our in-distribution datasets. ImageNet-1K is a large-scale visual recognition dataset containing 1000 object categories and 1281167 images. We split it into 10 folds randomly and ensured each fold contains 100 object categories. Since our computation resources are limited, we randomly select one fold as the in-distribution dataset. There other nine folds are used as OOD datasets as other popular benchmarks to test the performance of the proposed method in detecting near OOD samples. For CIFAR10 [27], there are 60000 color images in 10 classes. We conduct OOD algorithm evaluation as previous methods [30, 9, 46, 32, 45, 54].", + "bbox": [ + 497, + 686, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23429", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/28aa2845c2fe9ccd44992735eaa867eec281065b3c312e2e7391c7ab1a8e8b26.jpg", + "table_caption": [ + "Table 1. OOD detection performance comparison of HVCM and existing methods. All comparison methods rely on ResNet-50 trained with cross-entropy loss. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages." + ], + "table_footnote": [], + "table_body": "
MethodOOD DatasetsAverageInD Acc
iNaturalistSUNPlacesTexturesFPR95↓AUROC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]68.1287.4858.5989.7659.5389.1872.6682.7164.7387.28
ODIN [30]55.7885.9260.4783.8360.9488.0664.0681.2860.3184.77
Maha [29]97.0055.1298.8051.9397.2050.5220.0094.9978.2563.14
Energy [32]58.9186.4527.0393.4438.7591.5156.2587.2645.2489.6785.74
GODIN* [23]72.0079.8660.0984.5864.9482.3039.5089.2859.1384.01
MOS* [24]52.9491.4167.7886.8271.3184.3873.6580.5666.4285.79
ReAct [45]58.4882.6078.1869.1186.3359.8450.5387.0368.3874.65
HVCM(Ours)21.5692.1917.2094.4419.9893.6229.2290.6821.9992.7388.28
", + "bbox": [ + 77, + 163, + 892, + 312 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/995fb7f8c378028a5bda00a00a4145fd1ea9489a121af12937fd915d0aef56d2.jpg", + "table_caption": [ + "Table 2. Evaluation on more challenging detection tasks. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages." + ], + "table_footnote": [], + "table_body": "
MethodImageNet-OOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]72.5083.3389.1157.99
ODIN [30]73.4471.0364.0679.88
Maha [29]54.4079.3059.2077.62
Energy* [32]66.5681.0662.3484.61
GODIN* [23]71.5579.8973.5777.27
MOS* [24]87.4064.8766.7881.42
ReAct [45]87.0564.1584.1764.30
HVCM(Ours)42.8686.7228.5890.26
", + "bbox": [ + 78, + 385, + 470, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Out-of-distribution Dataset. On ImageNet, we follow Huang et al. [24] to test our methods and use Texture [8], iNaturalist [50], Places365 [61], and SUN [57]) as OOD test sets. To further explore the limitation of our approach, we evaluate our method on another two OOD datasets, OpenImage-O [26] and ImageNet-O [22]. For CIFAR10, as in [46, 32, 45, 54], we selected eight widely-used datasets, including Texture [8], SVHN [34], Places365 [61], iSUN [59], LSUN-Crop [60], LSUN-Resize [60], ImageNet-Resize [40], and ImageNet-Fix [40] as our test sets. Besides, to test the ability of HVCM to identify near OOD datasets, we use the remaining nine ImageNet subsets as the OOD test sets. Note that since our evaluation on ImageNet differs from previous methods [24, 45], we implement these algorithms with open source provided by authors and follow standard experimental settings.", + "bbox": [ + 75, + 530, + 468, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We employ the commonly used metrics in OOD detection [24, 45] to evaluate our approach, including AUROC, FPR95 and InD Acc. AUROC stands for the area under the receiver operating characteristic curve, FPR95 is short for TPR@FPR95 and represents the false positive rate when the true positive rate is $95\\%$ , and InD Acc is the classification accuracy of in-distribution data.", + "bbox": [ + 75, + 754, + 468, + 850 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We utilize ResNet-50 [19] as the feature", + "bbox": [ + 75, + 851, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "backbone for ImageNet and the dimension of the attribute space is set to 8192. The training is finished in 300 epochs. On CIFAR10, we use ResNet-18 [19] as our feature backbone, and the dimension of the attribute space is set to 1024. The training on CIFAR10 is finished in 200 epochs. The number of attribute groups is set to 32, and $\\alpha$ , $\\beta$ , $\\gamma_1$ , $\\gamma_2$ , and $\\gamma_3$ are set to 1, 0.1, 1, 1 and $1 \\times 10^{-4}$ , respectively. Updating $\\mu$ too fast can lead to the oscillation of group centers, negatively influencing the precision of probabilistic modeling. So we utilize a smaller $\\beta$ than $\\alpha$ to update $\\mu$ . All the hyperparameters are tuned according to the experimental results. We employ SGD with a momentum of 0.9, an initial learning rate of 0.1, and a batch size of 128. The learning rate is reduced by a factor of 10 at $50\\%$ and $75\\%$ of the total training epochs. We train all backbones from scratch using random initialization. All experiments are performed using PyTorch [37] with default parameters on four NVIDIA GeForce RTX 3090.", + "bbox": [ + 497, + 334, + 890, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Algorithms", + "text_level": 1, + "bbox": [ + 497, + 593, + 887, + 610 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Standard evaluation on ImageNet. We compare our HVCM with seven popular OOD detection methods, including MSP [21], ODIN [30], GODIN [23], Maha [29], Energy [32], MOS [24], and ReAct [45]. For datasets that describe objects or scenes, such as SUN, Places, and iNaturalist, HVCM achieves better AUROC and FPR95 metrics. When we summarize the results of all four datasets, HVCM achieves $21.99\\%$ on FPR95 and $92.73\\%$ on AUROC, which outperforms the previous best method Energy [32] by $23.25\\%$ and $3.06\\%$ . This is a significant improvement, which demonstrates that end-to-end training is very important to get good results. When compared to Maha [29], our proposed method exhibits inferior performance in terms of both FPR95 and AUROC. This observation suggests that our method is less effective in describing textures. We attribute this limitation to the fact that textures often encompass numerous repeated patterns, which differ from the characteristics required for general object", + "bbox": [ + 497, + 617, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23430", + "bbox": [ + 463, + 887, + 505, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b59f54d80b5cb03f7b48621c7cf5931310a1c8f0f8bc94af148ffc2c64a58d17.jpg", + "table_caption": [ + "Table 3. Comparison of OOD detection performance of HVCM and existing methods on CIFAR10 dataset. All compared methods use ResNet-18 trained with cross-entropy loss except our proposed method, which uses HVCMLoss. The performance is evaluated based on AUROC (A) and FPR95 (F). $\\uparrow$ indicates larger values are better and $\\downarrow$ indicates the opposite. Bold numbers indicate superior results. All values are expressed in percentages." + ], + "table_footnote": [], + "table_body": "
MethodOOD DatasetAverage
TextureSVHNPlaces365iSUNLSUN(C)LSUN(R)ImageNet(R)ImageNet(F)
F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑
CELossMSP56.4790.2058.4090.5651.8691.9853.8491.6446.2292.7449.1092.4859.6589.3559.9089.4054.4391.04
ODIN40.3791.9827.8293.2831.8094.2317.5996.6025.2395.3114.7497.2127.1894.6545.2690.6228.7594.24
Gram10.8197.732.5899.3930.1593.371.1199.7613.9197.030.5299.861.6198.3866.9482.0215.9595.94
Gram+pNML7.1898.501.6399.6023.2195.130.8399.809.4298.000.4299.881.2498.7657.9085.5312.7396.90
HVCMLossHVCM(ours)1.8899.311.3299.470.9599.540.6599.710.7799.660.5199.961.8099.204.7798.191.5899.38
", + "bbox": [ + 77, + 176, + 892, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da919b464e0ae709993ce4e367f77c534caab7e229e8b77e6a2242ad410f9db8.jpg", + "image_caption": [ + "Figure 3. HVCM performance comparison as the increasing distances between InD and OOD data." + ], + "image_footnote": [], + "bbox": [ + 80, + 296, + 470, + 432 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "recognition. It is important to note that our method only utilizes features from the last layer of the network, while Maha [29] leverages features from both intermediate and deep network layers. However, although our results are worse than Maha [29], we are better than all the other methods. This still indicates that the proposed method is very robust in identifying different types of outliers. Furthermore, we build a cosine classifier with the learned attribute centers for image classification. We got $88.28\\%$ accuracy which is $2.57\\%$ higher than our supervised learning baseline. This is solid evidence to show that our proposed method can model the InD data accurately and ensure the learned features keep high discriminative ability simultaneously.", + "bbox": [ + 75, + 489, + 468, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Standard evaluation on CIFAR10. We performed a more conventional OOD detection task on CIFAR10. This experiment aims twofold: firstly, to demonstrate that our method is not dependent on self-distillation (DINO [6]), and secondly, to validate the effectiveness and robustness of our HVCM. Table 3 compares our method with several classic and top-performing algorithms. All comparison methods use ResNet-18 as the main backbone network and are trained with cross-entropy loss, while our method uses only the loss function in Eq. (3). As shown in Table 3, the proposed method outperforms the previous best methods Gram [42] and pNML [2] on both average FPR95 and AUROC obviously. These results demonstrate the proposed method can perform well even on small datasets. Mean", + "bbox": [ + 75, + 672, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bbb82db78981acddb32227a133d65fb4eed3827d9365e8954d2825463b51a5b2.jpg", + "table_caption": [ + "Table 4. The performance of HVCM with varying numbers of group centers $G$ . Results are averaged across four standard OOD datasets, consistent with the main results." + ], + "table_footnote": [], + "table_body": "
Concepts NumberFPR95↓AUROC↑
G=824.3791.73
G=1623.0792.24
G=3221.9992.73
", + "bbox": [ + 519, + 337, + 873, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while, there is no self-distillation exploited, which indicates our joint representation learning and statistical modeling is independent of self-supervised learning algorithms [6].", + "bbox": [ + 497, + 439, + 890, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on more challenging OOD datasets. To overcome the limitations of current OOD benchmarks [53] and evaluate the robustness of our approach against adversarial attacks, we conducted experiments on two challenging datasets, namely OpenImage-O [26] and ImageNet-O [22]. As shown in Table 2, HVCM achieves the highest AUROC and lowest FPR95 among all methods on the OpenImage-O dataset. Although ImageNet-O contains adversarial examples and is more challenging, HVCM still outperforms other methods on this dataset.", + "bbox": [ + 497, + 492, + 892, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on near-to-far OOD datasets. To investigate the ability of the proposed method to detect near OOD samples, we construct 9 OOD test sets with the remaining ImageNet images. We rank the semantic distance between the remaining 900 visual categories with the 100 categories in the InD dataset. We use the average cosine distance as the measure and construct 9 different OOD test sets. Details are introduced in the supplementary material. For convenience, we denote these datasets from OOD 1 to OOD 9. The experimental results are depicted in Figure 3, and several samples are displayed in Figure 4. We can find that the proposed method achieves good AUROC even when the test set is very close to the InD dataset. When the test sets become farther, FPR95 decreases quickly, which indicates the proposed method is very sensitive to the semantic distances of OOD datasets.", + "bbox": [ + 497, + 644, + 894, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23431", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1c79ce06faf27ee74bfd7b4a2c70170003acac168ce8aaa49399a37655b8327.jpg", + "image_caption": [ + "Figure 4. Each dataset exhibits displayed two categories of images. The leftmost samples belong to the InD dataset, while the categories on the right correspond to nine OOD datasets arranged in ascending order of distance. It can be observed that the gap between the OOD samples and InD samples gradually widens as the distance increases." + ], + "image_footnote": [], + "bbox": [ + 82, + 124, + 166, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79651569599fe1d7aca39573a1699880ee4a1289dc51241faf73ac364295315a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 124, + 248, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5dc7a8b3738a0cf3d7598a81fa2fb85c8d552a1283b0872119f7a1b187d4a756.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 124, + 332, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f4ce035735d759670a962608b80b58d238afd3785c623f59ae38cb04e8679ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 124, + 413, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47ed9d314f95959e1c6602f53c134914d98af91d49c9408e538ad2e1099b0d7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 124, + 489, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5be6a798eff03983b97e80e9ac01c488c711fb6728032ad59a80b7d363ddc396.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 124, + 568, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/011fe1e45a40ba52df17c45226c7da0cd740d1dddc56cb96df60d1f96e1dfd21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 569, + 124, + 645, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5936316381cac47bd3ded9236f78ee5a778445f1e04d1e45e8a33e166ba3461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 124, + 727, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3afd3b53e5037dde691369b79ebd6a707edb32d3d29bf2ca7b69a24f49946f40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 729, + 124, + 803, + 249 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bd8ebf552e027c3b5072d428aecb91dd41495a42176e999b66b29157a0cefcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 124, + 887, + 249 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b0ec4c4ef195ef426b1be0307d06987fe82cb755714e27db1512e9640f2a9eb3.jpg", + "table_caption": [ + "Table 5. A set of ablation results about HVCM. The top row investigates the effect of using MSE, KL divergence, or JS divergence on the model performance; the bottle row compares the performance of different OOD detection methods. Results are averaged across four standard OOD datasets following the main results." + ], + "table_footnote": [], + "table_body": "
Strategy AblationFPR95↓AUROC↑
Learning ObjectiveL223.5992.26
KL22.9992.58
JS21.9992.73
InD DistanceCosine66.3784.06
Linear36.7786.59
Maha21.9992.73
", + "bbox": [ + 97, + 377, + 450, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 75, + 492, + 228, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of attribute groups. We varied the number of attribute groups from 8 to 32 to analyze the components in Gaussian mixture models. In Table 4, we find a positive correlation between the number of attribute groups and model performance, with the best performance achieved when $G$ is set to 32. We also try to set $G$ to bigger numbers and list the results in the supplementary material. However, more attribute groups will lead to bigger correlation matrices when computing the InD score. Thus, we finally set $G$ to 32 to balance performance and inference speed.", + "bbox": [ + 75, + 516, + 468, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of learning objectives. The first row of Table 5 investigates how the choice of the learning objective influences the performance of HVCM. We test three learning objectives, including the L2 loss, JS divergence loss, and KL divergence loss. The results show that compared with the L2 loss and KL divergence loss, the JS divergence loss achieves the lowest FPR95, demonstrating its superiority for statistical modeling. We attribute this to the fact that the group centers and features need to learn from each other, and the JS divergence loss is symmetrical in enclosing them.", + "bbox": [ + 75, + 656, + 468, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different InD distance metrics. In Table 5, we also compared our Maha metric with two different InD distance metrics. The cosine distance metric directly measures the distance by calculating the cosine similarity between the input feature and the mean of Gaussian distribution models. The", + "bbox": [ + 75, + 796, + 468, + 865 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "linear distance metric is used to calculate the distance with trainable linear layers. The results show that our Maha metric is an effective metric compared with its counterparts. We attribute this to the Maha distance space can better fit the training distribution in realistic scenes.", + "bbox": [ + 497, + 313, + 890, + 382 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/285be144585e931334d9d0e34d834848d001ef2bcd9292cae5e44e8cff5caf4a.jpg", + "image_caption": [ + "Figure 5. The performance of HVCM is evaluated as the number of InD classes increases across four OOD datasets." + ], + "image_footnote": [], + "bbox": [ + 512, + 403, + 878, + 538 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Increasing numbers of InD Categories in ImageNet. We test how the OOD detection performance varies with the increasing of object categories in the in-distribution dataset. Following Wang et al. [53], we test on four popular benchmarks and set $C$ to $\\{50, 100, 200, 300\\}$ respectively. As depicted in Figure 5, the performance of HVCM fluctuates on different datasets with the increasing of the InD object categories, which suggests that the number of categories has little impact on our approach. These results validate our assumption that we only need to model in-distribution image categories, and the out-of-distribution samples can be detected easily then.", + "bbox": [ + 497, + 585, + 890, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different thresholds for OOD detection. Figure 6 illustrates the accuracy of OOD detection for our method across various datasets. On most datasets, our method exhibits the same trend for accuracy variation with thresholds. This indicates that our approach has strong generalization and ideal performance to domains with significant differences. We explain the performance of Imagenet-O as its task difficulty with adversarial samples.", + "bbox": [ + 497, + 755, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23432", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/baf1b3963df841cae5c7bf8398f96ca7a512ee5e051b2e91981f86624548fa52.jpg", + "image_caption": [ + "Figure 6. The accuracy of HVAC varies with different thresholds across all OOD datasets." + ], + "image_footnote": [], + "bbox": [ + 99, + 123, + 440, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 77, + 384, + 194, + 399 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduce a hierarchical visual category modeling scheme for out-of-distribution detection, which combines visual representation learning and parameter optimization of probability models. It provides a novel perspective for OOD detection by conducting representation learning and density estimation end-to-end. By modeling visual categories with mixtures of Gaussian models, we describe visual categories in very complex distribution and don't rely on outlier training data to perform OOD detection. Experiments demonstrate that the proposed method outperforms state-of-the-art algorithms clearly and does not hinder the discriminative ability of deep features.", + "bbox": [ + 75, + 413, + 468, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations. However, our method needs to map deep features into high-dimensional attribute spaces and build plentiful Gaussian mixture models. These Gaussian mixture models bring a lot of computational costs and make the inference of the OOD detector become inefficient. Thus, simplifying the probability models and accelerating the inference process will be the future direction.", + "bbox": [ + 75, + 585, + 468, + 682 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 75, + 712, + 247, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (No.62072112), Scientific and Technological innovation action plan of Shanghai Science and Technology Committee (No.22511102202), National Key R&D Program of China (2020AAA0108301), National Natural Science Foundation of China under Grant (No. 62106051), Shanghai Pujiang Program (No. 21PJ1400600), and National Key R&D Program of China (2022YFC3601405).", + "bbox": [ + 75, + 741, + 468, + 865 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 124, + 596, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinwon An and Sungzoon Cho. Variational autoencoder based anomaly detection using reconstruction probability. Special lecture on IE, 2(1):1-18, 2015.", + "[2] Koby Bibas, Meir Feder, and Tal Hassner. Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems, 34:1179-1191, 2021.", + "[3] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019.", + "[4] Daniel Bogdoll, Jasmin Breitenstein, Florian Heidecker, Maarten Bieshaar, Bernhard Sick, Tim Fingscheidt, and Marius Zöllner. Description of corner cases in automated driving: Goals and challenges. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1023-1028, 2021.", + "[5] Senqi Cao and Zhongfei Zhang. Deep hybrid models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2022.", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021.", + "[7] Xingyu Chen, Xuguang Lan, Fuchun Sun, and Nanning Zheng. A boundary based out-of-distribution classifier for generalized zero-shot learning. In European Conference on Computer Vision, pages 572-588. Springer, 2020.", + "[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014.", + "[9] Roy De Maesschalck, Delphine Jouan-Rimbaud, and Désiré L Massart. The mahalanobis distance. Chemometrics and intelligent laboratory systems, 50(1):1-18, 2000.", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "[11] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016.", + "[12] Zhen Fang, Jie Lu, Anjin Liu, Feng Liu, and Guangquan Zhang. Learning bounds for open-set learning. In International Conference on Machine Learning, pages 3122-3132. PMLR, 2021.", + "[13] Aditya Grover, Manik Dhar, and Stefano Ermon. Flow-gan: Combining maximum likelihood and adversarial learning in generative models. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018.", + "[14] Xiaoyuan Guo, Judy Wawira Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Cvad: A generic" + ], + "bbox": [ + 500, + 147, + 894, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23433", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "medical anomaly detector based on cascade vae. arXiv preprint arXiv:2110.15811, 2021.", + "[15] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9591-9600, 2022.", + "[16] Matan Haroush, Tzviel Frostig, Ruth Heller, and Daniel Soudry. A statistical framework for efficient out of distribution detection in deep neural networks. In International Conference on Learning Representations, 2021.", + "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022.", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016.", + "[20] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022.", + "[21] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. Proceedings of International Conference on Learning Representations, 2017.", + "[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021.", + "[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.", + "[24] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021.", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.", + "[26] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, et al. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2(3):18, 2017.", + "[27] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009." + ], + "bbox": [ + 78, + 125, + 468, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In International Conference on Learning Representations, 2018.", + "[29] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018.", + "[30] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018.", + "[31] Qiao Liu, Jiaze Xu, Rui Jiang, and Wing Hung Wong. Density estimation using deep generative neural networks. Proceedings of the National Academy of Sciences, 118(15):e2101344118, 2021.", + "[32] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020.", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021.", + "[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011.", + "[35] Philipp Oberdiek, Gernot A Fink, and Matthias Rottmann. Uqgan: A unified model for uncertainty quantification of deep classifiers trained via conditional gans. arXiv preprint arXiv:2201.13279, 2022.", + "[36] George Papamakarios, Theo Pavlakou, and Iain Murray. Masked autoregressive flow for density estimation. Advances in neural information processing systems, 30, 2017.", + "[37] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019.", + "[38] Chen Qiu, Aodong Li, Marius Kloft, Maja Rudolph, and Stephan Mandt. Latent outlier exposure for anomaly detection with contaminated data. arXiv preprint arXiv:2202.08088, 2022.", + "[39] Carl Rasmussen. The infinite gaussian mixture model. Advances in neural information processing systems, 12, 1999.", + "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015.", + "[41] Murat Sensoy, Lance Kaplan, Federico Cerutti, and Maryam Saleki. Uncertainty-aware deep classifiers using generative" + ], + "bbox": [ + 502, + 127, + 890, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23434", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5620-5627, 2020.", + "[42] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with in-distribution examples and gram matrices. arXiv e-prints, pages arXiv-1912, 2019.", + "[43] Nina Shvetsova, Bart Bakker, Irina Fedulova, Heinrich Schulz, and Dmitry V. Dylov. Anomaly detection in medical imaging with deep perceptual autoencoders. IEEE Access, 9:118571-118583, 2021.", + "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[45] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In Advances in Neural Information Processing Systems, 2021.", + "[46] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. ICML, 2022.", + "[47] Esteban G Tabak and Cristina V Turner. A family of nonparametric density estimation algorithms. Communications on Pure and Applied Mathematics, 66(2):145-164, 2013.", + "[48] Shengjin Tang, Chuanqiang Yu, Xiaoyan Sun, Hongdong Fan, and Xiaosheng Si. A note on parameters estimation for nonlinear wiener processes with measurement errors. IEEE Access, 7:176756-176766, 2019.", + "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008.", + "[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018.", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[52] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), pages 550-564, 2018.", + "[53] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[54] Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. Mitigating neural network overconfidence with logit normalization. In International Conference on Machine Learning, pages 23631-23644. PMLR, 2022.", + "[55] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y. Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Nicholas Petrick and Kensaku Mori, editors, Medical Imaging 2018: Computer-Aided Diagnosis, volume 10575, pages 375 - 380. International Society for Optics and Photonics, SPIE, 2018." + ], + "bbox": [ + 80, + 125, + 468, + 865 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Jim Winkens, Rudy Bunel, Abhijit Guha Roy, Robert Stanforth, Vivek Natarajan, Joseph R. Ledsam, Patricia MacWilliams, Pushmeet Kohli, Alan Karthikesalingam, Simon Kohl, Taylan Cemgil, S. M. Ali Eslami, and Olaf Ronneberger. Contrastive training for improved out-of-distribution detection. arXiv preprint arXiv:2007.05566, 2020.", + "[57] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010.", + "[58] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. Advances in neural information processing systems, 33:20685-20696, 2020.", + "[59] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015.", + "[60] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015.", + "[61] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018.", + "[62] Cangqi Zhou, Hao Ban, Jing Zhang, Qianmu Li, and Yinghua Zhang. Gaussian mixture variational autoencoder for semi-supervised topic modeling. IEEE Access, 8:106843-106854, 2020." + ], + "bbox": [ + 502, + 127, + 890, + 536 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23435", + "bbox": [ + 463, + 887, + 504, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_model.json b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..08d74298e8c24420cb6aa1887146d17aaf42c20d --- /dev/null +++ b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.254, + 0.049, + 0.79, + 0.088 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.162, + 0.885, + 0.202 + ], + "angle": 0, + "content": "Hierarchical Visual Categories Modeling: A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.228, + 0.79, + 0.245 + ], + "angle": 0, + "content": "Jinglun Li\\(^{1}\\) Xinyu Zhou\\(^{2}\\) Pinxue Guo\\(^{1}\\) Yixuan Sun\\(^{1}\\) Yiwen Huang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.246, + 0.632, + 0.262 + ], + "angle": 0, + "content": "Weifeng Ge²† Wenqiang Zhang¹,²†" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.262, + 0.729, + 0.278 + ], + "angle": 0, + "content": "\\(^{1}\\)Academy for Engineering and Technology, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.278, + 0.664, + 0.294 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Computer Science, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.297, + 0.699, + 0.31 + ], + "angle": 0, + "content": "jingli960423@gmail.com,zhouxinyu20@fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.313, + 0.706, + 0.326 + ], + "angle": 0, + "content": "{pxguo21，21210860014，21210240056}@m.fudan.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.33, + 0.684, + 0.341 + ], + "angle": 0, + "content": "weifeng.ge.ic@gmail.com, wqzhang@fudan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.374, + 0.313, + 0.388 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.473, + 0.765 + ], + "angle": 0, + "content": "Detecting out-of-distribution inputs for visual recognition models has become critical in safe deep learning. This paper proposes a novel hierarchical visual category modeling scheme to separate out-of-distribution data from in-distribution data through joint representation learning and statistical modeling. We learn a mixture of Gaussian models for each in-distribution category. There are many Gaussian mixture models to model different visual categories. With these Gaussian models, we design an in-distribution score function by aggregating multiple Mahalanobis-based metrics. We don't use any auxiliary outlier data as training samples, which may hurt the generalization ability of out-of-distribution detection algorithms. We split the ImageNet-1k dataset into ten folds randomly. We use one fold as the in-distribution dataset and the others as out-of-distribution datasets to evaluate the proposed method. We also conduct experiments on seven popular benchmarks, including CIFAR, iNaturalist, SUN, Places, Textures, ImageNet-O, and OpenImage-O. Extensive experiments indicate that the proposed method outperforms state-of-the-art algorithms clearly. Meanwhile, we find that our visual representation has a competitive performance when compared with features learned by classical methods. These results demonstrate that the proposed method hasn't weakened the discriminative ability of visual recognition models and keeps high efficiency in detecting out-of-distribution samples." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.791, + 0.21, + 0.806 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.815, + 0.47, + 0.843 + ], + "angle": 0, + "content": "Modern deep neural networks have shown strong generalization ability when training and test data are from the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.375, + 0.892, + 0.542 + ], + "angle": 0, + "content": "same distribution [44, 18, 51, 10, 33]. However, encountering unexpected scenarios is inevitable in real-world applications. Thus assuring that training and test data share the same distribution becomes problematic. In applications like autonomous driving [3, 4] and medical image analysis [55, 14, 43], it is critical for models to identify inputs beyond their recognition capacity – known as out-of-distribution (OOD) detection. OOD detection algorithms can enable the system to warn humans promptly in many safety-related scenarios. Moreover, it has become an important research topic in the research community of safe artificial intelligence [21, 29, 23, 32, 56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.893, + 0.836 + ], + "angle": 0, + "content": "Many popular OOD detection methods aim to build probability models to describe training distributions [29, 56, 46, 24, 23]. With these probability models, they built a score function that can calculate in-distribution scores for test samples. These in-distribution scores reflect whether these samples fall into a given distribution. Then the test sample can be evaluated by the score function to decide whether it is an OOD sample or not. Thus modeling features of in-distribution data become extremely important. Previous works [29, 16, 5, 12, 38] build a distribution over the whole training data. Since training images may come from various visual categories, the decision boundary between In-Distribution (InD) and OOD data becomes extremely complex. To solve this problem, subsequent studies [24, 7, 15, 56] decomposed the whole dataset into several subgroups to simplify the decision boundary. Although representative algorithms like MOS [24], have gotten impressive performance in identifying OOD samples, they failed to detect near OOD samples. Because when different visual categories are grouped together, the OOD decision boundary will become even more uncertain." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.868 + ], + "angle": 0, + "content": "A typical framework for out-of-distribution detection involves two key steps: 1) learning a compact feature rep" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.853, + 0.275, + 0.867 + ], + "angle": 0, + "content": "\\(^\\dagger\\)indicates corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "23425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.334 + ], + "angle": 0, + "content": "representation that can fit probability models easily; 2) modeling features of in-distribution data in complex distributions accurately. The above two problems are mutually connected because more compact features will make modeling the data distribution easier, and stronger probability modeling techniques will exploit fewer restrictions on representation learning. However, achieving the above goals is difficult, even when there are a lot of breakthroughs in deep learning. Because if training samples from the same category are too close in the feature space, it will usually lead to overfitting. Meanwhile, in-distribution samples may come from different visual categories that have large variations in appearance and semantic information, which makes modeling complex training distributions become challenging even for excellent statisticians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.684 + ], + "angle": 0, + "content": "In this paper, we propose a new out-of-distribution detection framework, called hierarchical visual category modeling, to solve the above two issues simultaneously. We hold an assumption that given a training set that contains multiple visual categories, we can learn a probability model for every category independently. The out-of-distribution detection problem can be solved easily by aggregating probability models of known categories. Our motivation is that decomposing the whole dataset into subsets and modeling each category independently can avoid finding common characteristics shared by different categories. However, modeling an individual visual category is still challenging since classical supervised learning won't lead to compact feature representation. Thus, for each input sample, we need to force its feature representation to match the corresponding statistical model. That means we need to jointly conduct density estimation and representation learning. If we can jointly learn visual representations and optimize statistical models end-to-end, we can get good feature representations falling into distributions of the corresponding visual categories. Besides, we exploit knowledge distillation as done in [6] to learn robust feature representation. In this way, we can describe the complex training distribution with multiple Gaussian mixture models while not impairing the generalization ability of visual features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.473, + 0.868 + ], + "angle": 0, + "content": "In practice, to learn visual concepts that are in complex distributions, we build a Gaussian mixture model (GMM [39]) for each visual category. Given input samples, we extract their deep features and project these features into a high-dimensional attribute space. Different from classical Gaussian mixture models that send the same input into \\(K\\) different Gaussian models, we divide the attribute space into multiple groups and build a Gaussian model in each group independently. This strategy can give every attribute group a clear learning target and lead to better convergence. Experimental results indicate that this strategy works quite well. After the visual representation learning and statistical model parameters optimization, we can directly aggre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.895, + 0.278 + ], + "angle": 0, + "content": "gate these statistical models to judge whether a test sample comes from the training distribution or not. To evaluate our OOD detector, we split ImageNet into ten folds randomly and select one of these splits as the training set and all other splits as the OOD dataset to conduct extensive tests. Experiments indicate that the proposed method has a strong ability to identify OOD samples. We also evaluate our method on seven popular OOD benchmarks. Experimental results demonstrate that the proposed method not only can identify OOD samples efficiently but also improves the discriminative ability of learned visual representations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.279, + 0.895, + 0.306 + ], + "angle": 0, + "content": "The contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.316, + 0.892, + 0.4 + ], + "angle": 0, + "content": "- We introduce a new out-of-distribution detection scheme, called hierarchical visual category modeling to conduct joint representation learning and density estimation. It provides a new perspective for out-of-distribution detection to learn visual representation and probability models end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.41, + 0.892, + 0.479 + ], + "angle": 0, + "content": "- We exploit multiple Gaussian mixture models to model visual concepts in complex distributions. Visual attributes are divided into subgroups and modeled by different Gaussian components, which makes parameter learning much more efficient." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.489, + 0.892, + 0.572 + ], + "angle": 0, + "content": "- We conducted comprehensive experiments and ablation studies on popular benchmarks to investigate the effectiveness of the proposed method. Experiments demonstrated that our out-of-distribution detection models achieve better performance clearly when compared with previous methods." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.316, + 0.892, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.642, + 0.609 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Out-Of-Distribution Detection. Out-of-distribution detection aims to distinguish out-of-distribution samples from in-distribution data. Numerous methods have been proposed for OOD detection. Maximum softmax probability (MSP) [21] has been recognized as a strong baseline by using the maximum score across all classes as an OOD score. ODIN [30] improves MSP by perturbing the inputs and adjusting the logits via rescaling. Gaussian discriminant analysis has been employed in [29, 56] to detect OOD samples. ReAct [45] uses rectified activation to reduce model overconfidence in OOD data. Shama et al. in [42] utilized Gram matrices to measure feature correlations for OOD detection. Bibas et al. in [2] proposed pNML regret to detect OOD samples with a single-layer neural network. The Generalized-ODIN approach [23] decomposes the confidence of class probability using a dividend/divisor structure to incorporate prior knowledge. Another promising line of work focuses on designing new learning objectives to train" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "23426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.125, + 0.885, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Figure 1. Illustration of the training pipeline of hierarchical visual category modeling, written as HVCM. In HVCM, we jointly learn the visual representation and parameters of probabilistic models. We get two different views of an input image and send the outputs into a knowledge distillation framework (DINO [6]). Image representations are projected into a high-dimensional attribute space. Then these attributes are divided into different groups and pass SoftMax functions to get attribute distributions. We match attributes in each group with stored attribute centers of the target visual category. The whole model is trained in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.473, + 0.595 + ], + "angle": 0, + "content": "deep models from scratch. Leave-out Classifiers [52] introduces a margin loss to ensure that InD and OOD samples are separated in the feature space. Lee et al. in [28] proposed a novel confidence loss to give lower confidence for OOD samples. There are also some methods [58, 41, 35] designed to conduct OOD detection based on generative models. Unlike these methods, our proposed method jointly conducts representation learning and density estimation to improve the OOD detection performance and keep the learned feature with strong discriminative ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.473, + 0.868 + ], + "angle": 0, + "content": "Density Estimation in Deep Learning. Density estimation tries to describe the probabilistic density distribution of observed data accurately and has been investigated in deep learning for a long time [47, 11, 31]. In [1], Chong et al. employed variational autoencoders for anomaly detection, where the reconstruction probability was used to compute the anomaly score of each sample. Papamakarios et al. [36] proposed a novel method for density estimation based on masked autoregressive flow. Zhou et al. [62] extended variational autoencoders by incorporating a mixture of Gaussians to model the latent space distribution, allowing for more flexible and expressive representations. Yang et al. [13] combined Flow-based generative models with generative adversarial networks for density estimation and sample generation. Zhao et al. [11] used discrete latent variables to conduct density estimation, which has been applied to many research topics in natural language processing and image processing. In this paper, we build probabilistic models that can be jointly learned with deep networks. We wish" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.541 + ], + "angle": 0, + "content": "latent representations of the visual categories could easily be modeled by Gaussian mixture models even when complex distributions exist. Besides, we do not use post-hoc methods to conduct feature learning and density estimation separately. We aim to use the probabilistic model to guide the feature-learning process of deep networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.593, + 0.571 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.628, + 0.595 + ], + "angle": 0, + "content": "3.1. Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.868 + ], + "angle": 0, + "content": "State-of-the-art methods [21, 30, 29, 24, 46, 32] typically assume all InD data follow the same distribution. However, such an assumption will lead to the decision boundary of an OOD detector being doped with some uncertain space. In this paper, we avoid modeling the whole InD dataset and only focus on independently modeling each object category's distribution. We propose hierarchical visual categories modeling to achieve this goal while maintaining high classification accuracy. In hierarchical visual categories modeling, we first project image features into a high-dimensional attribute space (usually 8192 dimensions). These attributes can be grouped into multiple sub-visual concepts as components of an image category which can be easily modeled by Gaussian distributions. Then combinations of multiple sub-visual concepts (abbreviated as sub-concepts) can be grouped to describe a more complex visual concept of an in-distribution category. Since we define visual categories based on sub-visual concepts, simply modeling distributions of these sub-concepts and de" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "23427" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.468, + 0.154 + ], + "angle": 0, + "content": "scribing visual categories with these sub-concepts hierarchically can model complicated training distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.473, + 0.474 + ], + "angle": 0, + "content": "Formally, given a visual recognition model \\( f \\), it maps an input image \\( \\mathbf{x} \\) with label \\( y \\) to a high-dimensional feature vector \\( z \\in \\mathbb{R}^q \\). As described above, we project \\( z \\) into an attribute space \\( S \\in \\mathbb{R}^d \\) with a higher dimension and get an attribute description \\( \\mathbf{a} \\). Attributes in \\( \\mathbf{a} \\) can be grouped into multiple attribute subgroups \\( \\{a_i\\}_{i=1}^G \\) where \\( a_i \\in S_i \\subset \\mathbb{R}^{d/G} \\) and \\( i \\in \\{1,2,\\dots,G\\} \\). For images from a visual category \\( c \\), we assume that attributes in their \\( i \\)-th attribute group follow a simple Gaussian distribution \\( \\mathcal{N}(\\mu_i^c, \\Sigma_i^c) \\), where \\( \\mu_i^c \\) and \\( \\Sigma_i^c \\) are the mean and variance respectively. Since attributes are divided into \\( G \\) groups, we will have \\( G \\) different Gaussian distributions for each visual category. The benefits of dividing the whole attribute space into several subspaces come from two folds: (1) modeling attribute distributions in these groups becomes easier; (2) assemble of distributions in these attribute groups can describe complex distributions, which will lead to more accurate decision boundaries. To combine all attributes to describe every visual category, our OOD detector learns weights \\( \\{w_i^c\\}_{i=1}^G \\) of all attribute groups through exponential moving averages. Therefore, for each in-distribution class \\( c \\), we can model the corresponding visual concept in a probability perspective with a mixture of Gaussian models:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.482, + 0.47, + 0.52 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {x}; c\\right) = \\sum_ {i = 1} ^ {G} w _ {i} ^ {c} \\mathcal {N} \\left(\\boldsymbol {a} _ {i}; \\boldsymbol {\\mu} _ {i} ^ {c}, \\boldsymbol {\\Sigma} _ {i} ^ {c}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.623 + ], + "angle": 0, + "content": "where \\(w_{i}^{c}\\in \\mathbb{R},\\pmb{\\mu}_{i}^{c}\\in \\mathbb{R}^{d / G},\\pmb{\\Sigma}_{i}^{a}\\in \\mathbb{R}^{d / G\\times d / G}\\) and \\(\\mathcal{N}(\\cdot)\\) means a normal distribution. We build a Gaussian mixture model for each class and get \\(C\\) different Gaussian mixture models. Then, our proposed HVCM focuses on training deep neural networks to learn image features that follow the above distributions and parameters of these probability models jointly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.624, + 0.47, + 0.721 + ], + "angle": 0, + "content": "With the above Gaussian probability models, given a test sample \\(\\pmb{x}^{\\prime}\\), we define the score function \\(g(\\pmb{x}^{\\prime};\\pmb{w}^{c},\\pmb{\\mu}^{c},\\pmb{\\Sigma}^{c})\\) to measure whether it belongs to the \\(c\\)-th visual category using the learned probability density function. Here, \\(\\pmb{w}^{c} = \\{w_{i}^{c}\\}_{i = 1}^{G}\\), \\(\\pmb{\\mu}^{c} = \\{\\pmb{\\mu}_{i}^{c}\\}_{i = 1}^{G}\\) and \\(\\pmb{\\Sigma}^{c} = \\{\\pmb{\\Sigma}_{i}^{c}\\}_{i = 1}^{G}\\). This score function can be used as a reliable metric to detect OOD samples:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.73, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\nh \\left(\\boldsymbol {x} ^ {\\prime}\\right) = \\left\\{ \\begin{array}{l l} \\operatorname {I n D}, & \\text {i f} \\min _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) \\geq \\gamma , \\\\ \\operatorname {O O D}, & \\text {i f} \\min _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) < \\gamma , \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.784, + 0.469, + 0.825 + ], + "angle": 0, + "content": "where \\(\\gamma\\) are thresholds to be determined in subsequent sections. Eq. (2) indicates that we use the minimal InD score among all \\(C\\) categories to make the final decision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Our framework is illustrated in Figure 1, where there are two steps: (1) jointly learn deep features that fit our probability models and parameters of these models; (2) calculate" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.892, + 0.168 + ], + "angle": 0, + "content": "the InD score based on a set of Gaussian mixture models as the metric for out-of-distribution detection. In the following subsections, we give more details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.176, + 0.892, + 0.204 + ], + "angle": 0, + "content": "3.2. Joint Visual Representation Learning and Parameter Optimization of Probability Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.21, + 0.893, + 0.466 + ], + "angle": 0, + "content": "To learn visual representations that follow Gaussian mixture models and keep their discriminative ability at the same time, we exploit the knowledge distillation framework DINO [6] to perform the joint learning. As shown in Figure 1, for an image \\( \\pmb{x} \\), we sample ten different views of \\( \\pmb{x} \\) and send them into teacher and student branches simultaneously to perform self-distillation. During knowledge distillation, we project ResNet50 [19] features in 2048 dimensions into an attribute space with dimension \\( d(= 8192) \\). Apart from the learning objective produced by self-distillation, we force the attributes of each class to follow a class-specific Gaussian mixture model. We divide the attributes \\( \\pmb{a} \\in \\mathbb{R}^d \\) of \\( \\pmb{x} \\) into \\( G \\) groups to learn the Gaussian mixture model parameters. However, in practice, it's too hard to directly learn the mean and variance of Gaussian models. We follow He et. al [20] to learn attribute centers \\( \\{\\pmb{\\mu}_i^c\\}_{i = 1}^G \\) of the \\( c \\)-th category (the image label \\( y \\) is \\( c \\)). A linear projection layer is exploited to predict the weights \\( \\{w_i^c (\\pmb {x})\\}_{i = 1}^G \\) of \\( \\pmb{x} \\) on all \\( G \\) attribute groups. Our learning objective can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.474, + 0.891, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {K D} + \\alpha \\sum_ {i = 1} ^ {G} \\mathrm {K L} \\left(\\boldsymbol {a} _ {i} \\mid \\mid \\boldsymbol {\\mu} _ {i} ^ {c}\\right) + \\beta \\sum_ {i = 1} ^ {T} w _ {i} ^ {c} (\\boldsymbol {x}) \\mathrm {K L} \\left(\\boldsymbol {\\mu} _ {i} ^ {c} \\mid \\mid \\boldsymbol {a} _ {i}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.653 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{KD}\\) stands for the cross entropy loss in self-distillation, KL stands for the Kullback-Leibler(KL) divergence, and \\(\\alpha\\) and \\(\\beta\\) are hyperparameters. The reason why we exploit two KL divergences is that experiments indicate that the term \\(\\mathrm{KL}(\\pmb{a}_i||\\pmb{\\mu}_i^c)\\) will favor learning attribute centers and the term \\(\\mathrm{KL}(\\pmb{\\mu}_i^c||\\pmb{a}_i)\\) is more suitable to learn image attribute descriptions and group weights. Note that softmax operations normalize attributes and their learnable centers in each group before calculating learning objectives." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.653, + 0.892, + 0.71 + ], + "angle": 0, + "content": "With the learning objective in Eq. (3), student network parameters \\(\\theta_{s}\\), and weights \\(\\{w_i^c\\}_{i = 1}^G\\) and attribute centers \\(\\{\\pmb {\\mu}_i^c\\}_{i = 1}^G\\) of all groups are learnt at the same time. They are updated using the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.719, + 0.891, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\theta_ {s} ^ {t + 1} = \\theta_ {s} ^ {t} - \\gamma_ {1} \\frac {\\partial \\mathcal {L}}{\\partial \\theta_ {s} ^ {t}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.75, + 0.891, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {i} ^ {c, t + 1} = \\boldsymbol {\\mu} _ {i} ^ {c, t} - \\gamma_ {2} \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {\\mu} _ {i} ^ {c , t}}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.783, + 0.891, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {i} ^ {c, t + 1} = (1 - \\gamma_ {3}) \\boldsymbol {w} _ {i} ^ {c, t} + \\gamma_ {3} w _ {i} ^ {c} (\\boldsymbol {x}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.868 + ], + "angle": 0, + "content": "where \\(\\gamma_{1},\\gamma_{2}\\) and \\(\\gamma_{3}\\) are the learning rates. \\(\\gamma_{1},\\gamma_{2}\\) control the gradient update speed, while \\(\\gamma_{3}\\) controls the updating speed of the group weights. Following He et. al [20], \\(\\theta_s,\\{\\pmb {w}^c\\}_{c = 1}^C\\) and \\(\\{\\pmb {\\mu}^c\\}_{c = 1}^C\\) are initialized with Gaussian noises. We use" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "23428" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.124, + 0.468, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.471, + 0.417 + ], + "angle": 0, + "content": "Figure 2. Illustration of attribute group visualization and Mahalanobis distance distribution. Both attribute groups are visualized by t-SNE [49]. The colors encode different in-distribution data(ImageNet), and out-of-distribution features(SUN) marked as gray points. Models are trained on ResNet-50 [18] using DINO(a) and HVCM(b)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.441, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Adam optimizer [25] with momentum to update both \\(\\theta_{s}\\) and \\(\\{\\pmb{\\mu}^c\\}_{c=1}^C\\). While the attribute weights are learned through exponential moving averages [17]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.469, + 0.524 + ], + "angle": 0, + "content": "3.3. Out-of-Distribution Detection based on Probability Models in Hierarchy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.683 + ], + "angle": 0, + "content": "When the training of our hierarchical probability model converges, we will get visual attributes that follow learned Gaussian distributions hierarchically for in-distribution samples. Meanwhile, we also get weights and centers of all attribute groups for each class. However, the mean attributes are updated frequently during the training and are thus not suitable to be used as group centers. So we need to recalculate the attribute centers for each visual category. Given an in-distribution visual category \\( c \\), we estimate the mean vector and covariation matrix of the \\( i \\)-th attribute group:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.695, + 0.47, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {i} ^ {c} = \\frac {1}{N _ {c}} \\sum_ {m = 1} ^ {N _ {c}} \\boldsymbol {a} _ {i} ^ {m} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.735, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {i} ^ {c} = \\frac {1}{N _ {c} - 1} \\sum_ {m = 1} ^ {N _ {c}} \\left(\\boldsymbol {a} _ {i} ^ {m} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) \\left(\\boldsymbol {a} _ {i} ^ {m} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) ^ {\\top}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.47, + 0.867 + ], + "angle": 0, + "content": "where \\(N_{c}\\) notes the number of samples in the \\(c\\)-th class, and \\(\\pmb{a}_i^m\\) is the sub-attribute vector of the \\(m\\)-th sample. With these weights, means, and covariances, we can describe each visual category in the hierarchy accurately. We try to use the probability density function in Eq. (1) as the in-distribution function. However, we will encounter the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.21 + ], + "angle": 0, + "content": "problem of numerical overflow when calculating the determinants of covariance matrices as [48]. Instead, we compute the Mahalanobis distance between the \\(i\\)-th sub-visual attributes \\(a_{i}^{\\prime}\\) of a test sample \\(\\pmb{x}^{\\prime}\\) and the corresponding attribute center \\(\\mu_{i}^{c}\\) to measure the likelihood of these attributes belongs to the target category:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.218, + 0.892, + 0.236 + ], + "angle": 0, + "content": "\\[\nM _ {i} ^ {c} (\\boldsymbol {x}) = - \\left(\\boldsymbol {a} _ {i} ^ {\\prime} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) ^ {\\top} \\left(\\boldsymbol {\\Sigma} _ {i} ^ {c}\\right) ^ {- 1} \\left(\\boldsymbol {a} _ {i} ^ {\\prime} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.342 + ], + "angle": 0, + "content": "The above equation gives the in-distribution measure for one attribute group. While, for every category, we have multiple attribute groups and need to assemble related in-distribution measures to get the class-level in-distribution score. Since we have gotten the weights of attribute groups for each category, we can easily assemble them and get the class-level score function:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.35, + 0.892, + 0.387 + ], + "angle": 0, + "content": "\\[\ng \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) = \\sum_ {i = 1} ^ {G} w _ {i} ^ {c} M _ {i} ^ {c} \\left(\\boldsymbol {x} ^ {\\prime}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.48 + ], + "angle": 0, + "content": "With this score function, we can easily get the in-distribution score of a test sample on each visual category. Since there are \\( C \\) categories in the whole in-distribution dataset, we get the maximal in-distribution score across different visual categories as the in-distribution score on the whole dataset:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.49, + 0.891, + 0.51 + ], + "angle": 0, + "content": "\\[\ng \\left(\\boldsymbol {x} ^ {\\prime}\\right) = \\max _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.631 + ], + "angle": 0, + "content": "A high in-distribution score \\( g(\\boldsymbol{x}^{\\prime}) \\) indicates that the semantic attributes of a test sample lie very close to one or multiple in-distribution visual categories (as shown in Figure 2). On the contrary, if a sample does not belong to the previously modeled categories, it will get a low in-distribution score. We follow Eq. (2) to set thresholds to judge whether a sample is an out-of-distribution sample. In the experiment section, we discuss how to set these thresholds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.634, + 0.659 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.691, + 0.681 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In-distribution Datasets. We use ImageNet-1K [40] and CIFAR10 [27] as our in-distribution datasets. ImageNet-1K is a large-scale visual recognition dataset containing 1000 object categories and 1281167 images. We split it into 10 folds randomly and ensured each fold contains 100 object categories. Since our computation resources are limited, we randomly select one fold as the in-distribution dataset. There other nine folds are used as OOD datasets as other popular benchmarks to test the performance of the proposed method in detecting near OOD samples. For CIFAR10 [27], there are 60000 color images in 10 classes. We conduct OOD algorithm evaluation as previous methods [30, 9, 46, 32, 45, 54]." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "23429" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.122, + 0.894, + 0.163 + ], + "angle": 0, + "content": "Table 1. OOD detection performance comparison of HVCM and existing methods. All comparison methods rely on ResNet-50 trained with cross-entropy loss. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.165, + 0.894, + 0.313 + ], + "angle": 0, + "content": "
MethodOOD DatasetsAverageInD Acc
iNaturalistSUNPlacesTexturesFPR95↓AUROC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]68.1287.4858.5989.7659.5389.1872.6682.7164.7387.28
ODIN [30]55.7885.9260.4783.8360.9488.0664.0681.2860.3184.77
Maha [29]97.0055.1298.8051.9397.2050.5220.0094.9978.2563.14
Energy [32]58.9186.4527.0393.4438.7591.5156.2587.2645.2489.6785.74
GODIN* [23]72.0079.8660.0984.5864.9482.3039.5089.2859.1384.01
MOS* [24]52.9491.4167.7886.8271.3184.3873.6580.5666.4285.79
ReAct [45]58.4882.6078.1869.1186.3359.8450.5387.0368.3874.65
HVCM(Ours)21.5692.1917.2094.4419.9893.6229.2290.6821.9992.7388.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.331, + 0.472, + 0.385 + ], + "angle": 0, + "content": "Table 2. Evaluation on more challenging detection tasks. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.386, + 0.472, + 0.508 + ], + "angle": 0, + "content": "
MethodImageNet-OOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]72.5083.3389.1157.99
ODIN [30]73.4471.0364.0679.88
Maha [29]54.4079.3059.2077.62
Energy* [32]66.5681.0662.3484.61
GODIN* [23]71.5579.8973.5777.27
MOS* [24]87.4064.8766.7881.42
ReAct [45]87.0564.1584.1764.30
HVCM(Ours)42.8686.7228.5890.26
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.753 + ], + "angle": 0, + "content": "Out-of-distribution Dataset. On ImageNet, we follow Huang et al. [24] to test our methods and use Texture [8], iNaturalist [50], Places365 [61], and SUN [57]) as OOD test sets. To further explore the limitation of our approach, we evaluate our method on another two OOD datasets, OpenImage-O [26] and ImageNet-O [22]. For CIFAR10, as in [46, 32, 45, 54], we selected eight widely-used datasets, including Texture [8], SVHN [34], Places365 [61], iSUN [59], LSUN-Crop [60], LSUN-Resize [60], ImageNet-Resize [40], and ImageNet-Fix [40] as our test sets. Besides, to test the ability of HVCM to identify near OOD datasets, we use the remaining nine ImageNet subsets as the OOD test sets. Note that since our evaluation on ImageNet differs from previous methods [24, 45], we implement these algorithms with open source provided by authors and follow standard experimental settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.47, + 0.851 + ], + "angle": 0, + "content": "Evaluation Metrics. We employ the commonly used metrics in OOD detection [24, 45] to evaluate our approach, including AUROC, FPR95 and InD Acc. AUROC stands for the area under the receiver operating characteristic curve, FPR95 is short for TPR@FPR95 and represents the false positive rate when the true positive rate is \\(95\\%\\), and InD Acc is the classification accuracy of in-distribution data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.852, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Training Details. We utilize ResNet-50 [19] as the feature" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.584 + ], + "angle": 0, + "content": "backbone for ImageNet and the dimension of the attribute space is set to 8192. The training is finished in 300 epochs. On CIFAR10, we use ResNet-18 [19] as our feature backbone, and the dimension of the attribute space is set to 1024. The training on CIFAR10 is finished in 200 epochs. The number of attribute groups is set to 32, and \\(\\alpha\\), \\(\\beta\\), \\(\\gamma_1\\), \\(\\gamma_2\\), and \\(\\gamma_3\\) are set to 1, 0.1, 1, 1 and \\(1 \\times 10^{-4}\\), respectively. Updating \\(\\mu\\) too fast can lead to the oscillation of group centers, negatively influencing the precision of probabilistic modeling. So we utilize a smaller \\(\\beta\\) than \\(\\alpha\\) to update \\(\\mu\\). All the hyperparameters are tuned according to the experimental results. We employ SGD with a momentum of 0.9, an initial learning rate of 0.1, and a batch size of 128. The learning rate is reduced by a factor of 10 at \\(50\\%\\) and \\(75\\%\\) of the total training epochs. We train all backbones from scratch using random initialization. All experiments are performed using PyTorch [37] with default parameters on four NVIDIA GeForce RTX 3090." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.595, + 0.888, + 0.611 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Algorithms" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Standard evaluation on ImageNet. We compare our HVCM with seven popular OOD detection methods, including MSP [21], ODIN [30], GODIN [23], Maha [29], Energy [32], MOS [24], and ReAct [45]. For datasets that describe objects or scenes, such as SUN, Places, and iNaturalist, HVCM achieves better AUROC and FPR95 metrics. When we summarize the results of all four datasets, HVCM achieves \\(21.99\\%\\) on FPR95 and \\(92.73\\%\\) on AUROC, which outperforms the previous best method Energy [32] by \\(23.25\\%\\) and \\(3.06\\%\\). This is a significant improvement, which demonstrates that end-to-end training is very important to get good results. When compared to Maha [29], our proposed method exhibits inferior performance in terms of both FPR95 and AUROC. This observation suggests that our method is less effective in describing textures. We attribute this limitation to the fact that textures often encompass numerous repeated patterns, which differ from the characteristics required for general object" + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "23430" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.122, + 0.894, + 0.175 + ], + "angle": 0, + "content": "Table 3. Comparison of OOD detection performance of HVCM and existing methods on CIFAR10 dataset. All compared methods use ResNet-18 trained with cross-entropy loss except our proposed method, which uses HVCMLoss. The performance is evaluated based on AUROC (A) and FPR95 (F). \\(\\uparrow\\) indicates larger values are better and \\(\\downarrow\\) indicates the opposite. Bold numbers indicate superior results. All values are expressed in percentages." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.177, + 0.894, + 0.278 + ], + "angle": 0, + "content": "
MethodOOD DatasetAverage
TextureSVHNPlaces365iSUNLSUN(C)LSUN(R)ImageNet(R)ImageNet(F)
F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑
CELossMSP56.4790.2058.4090.5651.8691.9853.8491.6446.2292.7449.1092.4859.6589.3559.9089.4054.4391.04
ODIN40.3791.9827.8293.2831.8094.2317.5996.6025.2395.3114.7497.2127.1894.6545.2690.6228.7594.24
Gram10.8197.732.5899.3930.1593.371.1199.7613.9197.030.5299.861.6198.3866.9482.0215.9595.94
Gram+pNML7.1898.501.6399.6023.2195.130.8399.809.4298.000.4299.881.2498.7657.9085.5312.7396.90
HVCMLossHVCM(ours)1.8899.311.3299.470.9599.540.6599.710.7799.660.5199.961.8099.204.7798.191.5899.38
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.298, + 0.472, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.435, + 0.471, + 0.46 + ], + "angle": 0, + "content": "Figure 3. HVCM performance comparison as the increasing distances between InD and OOD data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.67 + ], + "angle": 0, + "content": "recognition. It is important to note that our method only utilizes features from the last layer of the network, while Maha [29] leverages features from both intermediate and deep network layers. However, although our results are worse than Maha [29], we are better than all the other methods. This still indicates that the proposed method is very robust in identifying different types of outliers. Furthermore, we build a cosine classifier with the learned attribute centers for image classification. We got \\(88.28\\%\\) accuracy which is \\(2.57\\%\\) higher than our supervised learning baseline. This is solid evidence to show that our proposed method can model the InD data accurately and ensure the learned features keep high discriminative ability simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Standard evaluation on CIFAR10. We performed a more conventional OOD detection task on CIFAR10. This experiment aims twofold: firstly, to demonstrate that our method is not dependent on self-distillation (DINO [6]), and secondly, to validate the effectiveness and robustness of our HVCM. Table 3 compares our method with several classic and top-performing algorithms. All comparison methods use ResNet-18 as the main backbone network and are trained with cross-entropy loss, while our method uses only the loss function in Eq. (3). As shown in Table 3, the proposed method outperforms the previous best methods Gram [42] and pNML [2] on both average FPR95 and AUROC obviously. These results demonstrate the proposed method can perform well even on small datasets. Mean" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.296, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Table 4. The performance of HVCM with varying numbers of group centers \\( G \\). Results are averaged across four standard OOD datasets, consistent with the main results." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.338, + 0.875, + 0.392 + ], + "angle": 0, + "content": "
Concepts NumberFPR95↓AUROC↑
G=824.3791.73
G=1623.0792.24
G=3221.9992.73
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.483 + ], + "angle": 0, + "content": "while, there is no self-distillation exploited, which indicates our joint representation learning and statistical modeling is independent of self-supervised learning algorithms [6]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.893, + 0.633 + ], + "angle": 0, + "content": "Results on more challenging OOD datasets. To overcome the limitations of current OOD benchmarks [53] and evaluate the robustness of our approach against adversarial attacks, we conducted experiments on two challenging datasets, namely OpenImage-O [26] and ImageNet-O [22]. As shown in Table 2, HVCM achieves the highest AUROC and lowest FPR95 among all methods on the OpenImage-O dataset. Although ImageNet-O contains adversarial examples and is more challenging, HVCM still outperforms other methods on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.895, + 0.867 + ], + "angle": 0, + "content": "Results on near-to-far OOD datasets. To investigate the ability of the proposed method to detect near OOD samples, we construct 9 OOD test sets with the remaining ImageNet images. We rank the semantic distance between the remaining 900 visual categories with the 100 categories in the InD dataset. We use the average cosine distance as the measure and construct 9 different OOD test sets. Details are introduced in the supplementary material. For convenience, we denote these datasets from OOD 1 to OOD 9. The experimental results are depicted in Figure 3, and several samples are displayed in Figure 4. We can find that the proposed method achieves good AUROC even when the test set is very close to the InD dataset. When the test sets become farther, FPR95 decreases quickly, which indicates the proposed method is very sensitive to the semantic distances of OOD datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "23431" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.125, + 0.167, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.125, + 0.25, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.125, + 0.334, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.125, + 0.414, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.125, + 0.49, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.125, + 0.569, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.125, + 0.646, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.125, + 0.728, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.125, + 0.805, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.125, + 0.888, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.253, + 0.893, + 0.292 + ], + "angle": 0, + "content": "Figure 4. Each dataset exhibits displayed two categories of images. The leftmost samples belong to the InD dataset, while the categories on the right correspond to nine OOD datasets arranged in ascending order of distance. It can be observed that the gap between the OOD samples and InD samples gradually widens as the distance increases." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.374 + ], + "angle": 0, + "content": "Table 5. A set of ablation results about HVCM. The top row investigates the effect of using MSE, KL divergence, or JS divergence on the model performance; the bottle row compares the performance of different OOD detection methods. Results are averaged across four standard OOD datasets following the main results." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.378, + 0.451, + 0.466 + ], + "angle": 0, + "content": "
Strategy AblationFPR95↓AUROC↑
Learning ObjectiveL223.5992.26
KL22.9992.58
JS21.9992.73
InD DistanceCosine66.3784.06
Linear36.7786.59
Maha21.9992.73
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.493, + 0.23, + 0.509 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Number of attribute groups. We varied the number of attribute groups from 8 to 32 to analyze the components in Gaussian mixture models. In Table 4, we find a positive correlation between the number of attribute groups and model performance, with the best performance achieved when \\( G \\) is set to 32. We also try to set \\( G \\) to bigger numbers and list the results in the supplementary material. However, more attribute groups will lead to bigger correlation matrices when computing the InD score. Thus, we finally set \\( G \\) to 32 to balance performance and inference speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Choice of learning objectives. The first row of Table 5 investigates how the choice of the learning objective influences the performance of HVCM. We test three learning objectives, including the L2 loss, JS divergence loss, and KL divergence loss. The results show that compared with the L2 loss and KL divergence loss, the JS divergence loss achieves the lowest FPR95, demonstrating its superiority for statistical modeling. We attribute this to the fact that the group centers and features need to learn from each other, and the JS divergence loss is symmetrical in enclosing them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.469, + 0.866 + ], + "angle": 0, + "content": "Different InD distance metrics. In Table 5, we also compared our Maha metric with two different InD distance metrics. The cosine distance metric directly measures the distance by calculating the cosine similarity between the input feature and the mean of Gaussian distribution models. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.383 + ], + "angle": 0, + "content": "linear distance metric is used to calculate the distance with trainable linear layers. The results show that our Maha metric is an effective metric compared with its counterparts. We attribute this to the Maha distance space can better fit the training distribution in realistic scenes." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.404, + 0.88, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.541, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Figure 5. The performance of HVCM is evaluated as the number of InD classes increases across four OOD datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.752 + ], + "angle": 0, + "content": "Increasing numbers of InD Categories in ImageNet. We test how the OOD detection performance varies with the increasing of object categories in the in-distribution dataset. Following Wang et al. [53], we test on four popular benchmarks and set \\( C \\) to \\( \\{50, 100, 200, 300\\} \\) respectively. As depicted in Figure 5, the performance of HVCM fluctuates on different datasets with the increasing of the InD object categories, which suggests that the number of categories has little impact on our approach. These results validate our assumption that we only need to model in-distribution image categories, and the out-of-distribution samples can be detected easily then." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Different thresholds for OOD detection. Figure 6 illustrates the accuracy of OOD detection for our method across various datasets. On most datasets, our method exhibits the same trend for accuracy variation with thresholds. This indicates that our approach has strong generalization and ideal performance to domains with significant differences. We explain the performance of Imagenet-O as its task difficulty with adversarial samples." + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "23432" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.124, + 0.442, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.328, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Figure 6. The accuracy of HVAC varies with different thresholds across all OOD datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.385, + 0.196, + 0.4 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.581 + ], + "angle": 0, + "content": "In this paper, we introduce a hierarchical visual category modeling scheme for out-of-distribution detection, which combines visual representation learning and parameter optimization of probability models. It provides a novel perspective for OOD detection by conducting representation learning and density estimation end-to-end. By modeling visual categories with mixtures of Gaussian models, we describe visual categories in very complex distribution and don't rely on outlier training data to perform OOD detection. Experiments demonstrate that the proposed method outperforms state-of-the-art algorithms clearly and does not hinder the discriminative ability of deep features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.683 + ], + "angle": 0, + "content": "Limitations. However, our method needs to map deep features into high-dimensional attribute spaces and build plentiful Gaussian mixture models. These Gaussian mixture models bring a lot of computational costs and make the inference of the OOD detector become inefficient. Thus, simplifying the probability models and accelerating the inference process will be the future direction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.248, + 0.73 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.866 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (No.62072112), Scientific and Technological innovation action plan of Shanghai Science and Technology Committee (No.22511102202), National Key R&D Program of China (2020AAA0108301), National Natural Science Foundation of China under Grant (No. 62106051), Shanghai Pujiang Program (No. 21PJ1400600), and National Key R&D Program of China (2022YFC3601405)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.125, + 0.597, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.148, + 0.892, + 0.186 + ], + "angle": 0, + "content": "[1] Jinwon An and Sungzoon Cho. Variational autoencoder based anomaly detection using reconstruction probability. Special lecture on IE, 2(1):1-18, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.188, + 0.894, + 0.239 + ], + "angle": 0, + "content": "[2] Koby Bibas, Meir Feder, and Tal Hassner. Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems, 34:1179-1191, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.241, + 0.895, + 0.304 + ], + "angle": 0, + "content": "[3] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.305, + 0.894, + 0.382 + ], + "angle": 0, + "content": "[4] Daniel Bogdoll, Jasmin Breitenstein, Florian Heidecker, Maarten Bieshaar, Bernhard Sick, Tim Fingscheidt, and Marius Zöllner. Description of corner cases in automated driving: Goals and challenges. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1023-1028, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.383, + 0.895, + 0.434 + ], + "angle": 0, + "content": "[5] Senqi Cao and Zhongfei Zhang. Deep hybrid models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.436, + 0.895, + 0.499 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.5, + 0.894, + 0.551 + ], + "angle": 0, + "content": "[7] Xingyu Chen, Xuguang Lan, Fuchun Sun, and Nanning Zheng. A boundary based out-of-distribution classifier for generalized zero-shot learning. In European Conference on Computer Vision, pages 572-588. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.553, + 0.895, + 0.604 + ], + "angle": 0, + "content": "[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.605, + 0.895, + 0.643 + ], + "angle": 0, + "content": "[9] Roy De Maesschalck, Delphine Jouan-Rimbaud, and Désiré L Massart. The mahalanobis distance. Chemometrics and intelligent laboratory systems, 50(1):1-18, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.644, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.696, + 0.892, + 0.735 + ], + "angle": 0, + "content": "[11] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.736, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[12] Zhen Fang, Jie Lu, Anjin Liu, Feng Liu, and Guangquan Zhang. Learning bounds for open-set learning. In International Conference on Machine Learning, pages 3122-3132. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.788, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[13] Aditya Grover, Manik Dhar, and Stefano Ermon. Flow-gan: Combining maximum likelihood and adversarial learning in generative models. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.841, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[14] Xiaoyuan Guo, Judy Wawira Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Cvad: A generic" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.148, + 0.895, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "23433" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.127, + 0.469, + 0.153 + ], + "angle": 0, + "content": "medical anomaly detector based on cascade vae. arXiv preprint arXiv:2110.15811, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.154, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[15] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9591-9600, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.217, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[16] Matan Haroush, Tzviel Frostig, Ruth Heller, and Daniel Soudry. A statistical framework for efficient out of distribution detection in deep neural networks. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.331, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[20] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[21] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. Proceedings of International Conference on Learning Representations, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.598 + ], + "angle": 0, + "content": "[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[24] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.47, + 0.751 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.752, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[26] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, et al. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2(3):18, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.47, + 0.867 + ], + "angle": 0, + "content": "[27] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[28] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.254 + ], + "angle": 0, + "content": "[29] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.255, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[30] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[31] Qiao Liu, Jiaze Xu, Rui Jiang, and Wing Hung Wong. Density estimation using deep generative neural networks. Proceedings of the National Academy of Sciences, 118(15):e2101344118, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[32] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[35] Philipp Oberdiek, Gernot A Fink, and Matthias Rottmann. Uqgan: A unified model for uncertainty quantification of deep classifiers trained via conditional gans. arXiv preprint arXiv:2201.13279, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[36] George Papamakarios, Theo Pavlakou, and Iain Murray. Masked autoregressive flow for density estimation. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[37] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.751 + ], + "angle": 0, + "content": "[38] Chen Qiu, Aodong Li, Marius Kloft, Maja Rudolph, and Stephan Mandt. Latent outlier exposure for anomaly detection with contaminated data. arXiv preprint arXiv:2202.08088, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.752, + 0.892, + 0.777 + ], + "angle": 0, + "content": "[39] Carl Rasmussen. The infinite gaussian mixture model. Advances in neural information processing systems, 12, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.778, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.842, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[41] Murat Sensoy, Lance Kaplan, Federico Cerutti, and Maryam Saleki. Uncertainty-aware deep classifiers using generative" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "23434" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.127, + 0.469, + 0.153 + ], + "angle": 0, + "content": "models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5620-5627, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.155, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[42] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with in-distribution examples and gram matrices. arXiv e-prints, pages arXiv-1912, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.194, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[43] Nina Shvetsova, Bart Bakker, Irina Fedulova, Heinrich Schulz, and Dmitry V. Dylov. Anomaly detection in medical imaging with deep perceptual autoencoders. IEEE Access, 9:118571-118583, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.246, + 0.469, + 0.283 + ], + "angle": 0, + "content": "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.285, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[45] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.324, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[46] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[47] Esteban G Tabak and Cristina V Turner. A family of nonparametric density estimation algorithms. Communications on Pure and Applied Mathematics, 66(2):145-164, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.403, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[48] Shengjin Tang, Chuanqiang Yu, Xiaoyan Sun, Hongdong Fan, and Xiaosheng Si. A note on parameters estimation for nonlinear wiener processes with measurement errors. IEEE Access, 7:176756-176766, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.454, + 0.469, + 0.491 + ], + "angle": 0, + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.494, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.571, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[52] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), pages 550-564, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.687, + 0.469, + 0.736 + ], + "angle": 0, + "content": "[53] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.739, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[54] Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. Mitigating neural network overconfidence with logit normalization. In International Conference on Machine Learning, pages 23631-23644. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.791, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[55] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y. Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Nicholas Petrick and Kensaku Mori, editors, Medical Imaging 2018: Computer-Aided Diagnosis, volume 10575, pages 375 - 380. International Society for Optics and Photonics, SPIE, 2018." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.127, + 0.469, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[56] Jim Winkens, Rudy Bunel, Abhijit Guha Roy, Robert Stanforth, Vivek Natarajan, Joseph R. Ledsam, Patricia MacWilliams, Pushmeet Kohli, Alan Karthikesalingam, Simon Kohl, Taylan Cemgil, S. M. Ali Eslami, and Olaf Ronneberger. Contrastive training for improved out-of-distribution detection. arXiv preprint arXiv:2007.05566, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[57] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[58] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. Advances in neural information processing systems, 33:20685-20696, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[59] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[60] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.438, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[61] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[62] Cangqi Zhou, Hao Ban, Jing Zhang, Qianmu Li, and Yinghua Zhang. Gaussian mixture variational autoencoder for semi-supervised topic modeling. IEEE Access, 8:106843-106854, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.465, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "23435" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_origin.pdf b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9084fb6a568c0b536f84bb78e7e297a4ef1e3d99 --- /dev/null +++ b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/bd3288f5-edbd-4923-8ad4-4c190bb5dbf1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e39ef0dea5df57a7310db455d6c81fccd8810c851b4e3c541e23f6573abf970d +size 1020284 diff --git a/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/full.md b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..981decd72df254999e4b1d4fef6bfd9f78fd9e03 --- /dev/null +++ b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/full.md @@ -0,0 +1,322 @@ +# Hierarchical Visual Categories Modeling: A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection + +Jinglun Li $^{1}$ Xinyu Zhou $^{2}$ Pinxue Guo $^{1}$ Yixuan Sun $^{1}$ Yiwen Huang $^{2}$ + +Weifeng Ge²† Wenqiang Zhang¹,²† + +$^{1}$ Academy for Engineering and Technology, Fudan University + +$^{2}$ School of Computer Science, Fudan University + +jingli960423@gmail.com,zhouxinyu20@fudan.edu.cn + +{pxguo21，21210860014，21210240056}@m.fudan.edu.cn, + +weifeng.ge.ic@gmail.com, wqzhang@fudan.edu.cn + +# Abstract + +Detecting out-of-distribution inputs for visual recognition models has become critical in safe deep learning. This paper proposes a novel hierarchical visual category modeling scheme to separate out-of-distribution data from in-distribution data through joint representation learning and statistical modeling. We learn a mixture of Gaussian models for each in-distribution category. There are many Gaussian mixture models to model different visual categories. With these Gaussian models, we design an in-distribution score function by aggregating multiple Mahalanobis-based metrics. We don't use any auxiliary outlier data as training samples, which may hurt the generalization ability of out-of-distribution detection algorithms. We split the ImageNet-1k dataset into ten folds randomly. We use one fold as the in-distribution dataset and the others as out-of-distribution datasets to evaluate the proposed method. We also conduct experiments on seven popular benchmarks, including CIFAR, iNaturalist, SUN, Places, Textures, ImageNet-O, and OpenImage-O. Extensive experiments indicate that the proposed method outperforms state-of-the-art algorithms clearly. Meanwhile, we find that our visual representation has a competitive performance when compared with features learned by classical methods. These results demonstrate that the proposed method hasn't weakened the discriminative ability of visual recognition models and keeps high efficiency in detecting out-of-distribution samples. + +# 1. Introduction + +Modern deep neural networks have shown strong generalization ability when training and test data are from the + +same distribution [44, 18, 51, 10, 33]. However, encountering unexpected scenarios is inevitable in real-world applications. Thus assuring that training and test data share the same distribution becomes problematic. In applications like autonomous driving [3, 4] and medical image analysis [55, 14, 43], it is critical for models to identify inputs beyond their recognition capacity – known as out-of-distribution (OOD) detection. OOD detection algorithms can enable the system to warn humans promptly in many safety-related scenarios. Moreover, it has become an important research topic in the research community of safe artificial intelligence [21, 29, 23, 32, 56]. + +Many popular OOD detection methods aim to build probability models to describe training distributions [29, 56, 46, 24, 23]. With these probability models, they built a score function that can calculate in-distribution scores for test samples. These in-distribution scores reflect whether these samples fall into a given distribution. Then the test sample can be evaluated by the score function to decide whether it is an OOD sample or not. Thus modeling features of in-distribution data become extremely important. Previous works [29, 16, 5, 12, 38] build a distribution over the whole training data. Since training images may come from various visual categories, the decision boundary between In-Distribution (InD) and OOD data becomes extremely complex. To solve this problem, subsequent studies [24, 7, 15, 56] decomposed the whole dataset into several subgroups to simplify the decision boundary. Although representative algorithms like MOS [24], have gotten impressive performance in identifying OOD samples, they failed to detect near OOD samples. Because when different visual categories are grouped together, the OOD decision boundary will become even more uncertain. + +A typical framework for out-of-distribution detection involves two key steps: 1) learning a compact feature rep + +representation that can fit probability models easily; 2) modeling features of in-distribution data in complex distributions accurately. The above two problems are mutually connected because more compact features will make modeling the data distribution easier, and stronger probability modeling techniques will exploit fewer restrictions on representation learning. However, achieving the above goals is difficult, even when there are a lot of breakthroughs in deep learning. Because if training samples from the same category are too close in the feature space, it will usually lead to overfitting. Meanwhile, in-distribution samples may come from different visual categories that have large variations in appearance and semantic information, which makes modeling complex training distributions become challenging even for excellent statisticians. + +In this paper, we propose a new out-of-distribution detection framework, called hierarchical visual category modeling, to solve the above two issues simultaneously. We hold an assumption that given a training set that contains multiple visual categories, we can learn a probability model for every category independently. The out-of-distribution detection problem can be solved easily by aggregating probability models of known categories. Our motivation is that decomposing the whole dataset into subsets and modeling each category independently can avoid finding common characteristics shared by different categories. However, modeling an individual visual category is still challenging since classical supervised learning won't lead to compact feature representation. Thus, for each input sample, we need to force its feature representation to match the corresponding statistical model. That means we need to jointly conduct density estimation and representation learning. If we can jointly learn visual representations and optimize statistical models end-to-end, we can get good feature representations falling into distributions of the corresponding visual categories. Besides, we exploit knowledge distillation as done in [6] to learn robust feature representation. In this way, we can describe the complex training distribution with multiple Gaussian mixture models while not impairing the generalization ability of visual features. + +In practice, to learn visual concepts that are in complex distributions, we build a Gaussian mixture model (GMM [39]) for each visual category. Given input samples, we extract their deep features and project these features into a high-dimensional attribute space. Different from classical Gaussian mixture models that send the same input into $K$ different Gaussian models, we divide the attribute space into multiple groups and build a Gaussian model in each group independently. This strategy can give every attribute group a clear learning target and lead to better convergence. Experimental results indicate that this strategy works quite well. After the visual representation learning and statistical model parameters optimization, we can directly aggre + +gate these statistical models to judge whether a test sample comes from the training distribution or not. To evaluate our OOD detector, we split ImageNet into ten folds randomly and select one of these splits as the training set and all other splits as the OOD dataset to conduct extensive tests. Experiments indicate that the proposed method has a strong ability to identify OOD samples. We also evaluate our method on seven popular OOD benchmarks. Experimental results demonstrate that the proposed method not only can identify OOD samples efficiently but also improves the discriminative ability of learned visual representations. + +The contributions of this paper are summarized as follows: + +- We introduce a new out-of-distribution detection scheme, called hierarchical visual category modeling to conduct joint representation learning and density estimation. It provides a new perspective for out-of-distribution detection to learn visual representation and probability models end-to-end. +- We exploit multiple Gaussian mixture models to model visual concepts in complex distributions. Visual attributes are divided into subgroups and modeled by different Gaussian components, which makes parameter learning much more efficient. +- We conducted comprehensive experiments and ablation studies on popular benchmarks to investigate the effectiveness of the proposed method. Experiments demonstrated that our out-of-distribution detection models achieve better performance clearly when compared with previous methods. + +# 2. Related Work + +Out-Of-Distribution Detection. Out-of-distribution detection aims to distinguish out-of-distribution samples from in-distribution data. Numerous methods have been proposed for OOD detection. Maximum softmax probability (MSP) [21] has been recognized as a strong baseline by using the maximum score across all classes as an OOD score. ODIN [30] improves MSP by perturbing the inputs and adjusting the logits via rescaling. Gaussian discriminant analysis has been employed in [29, 56] to detect OOD samples. ReAct [45] uses rectified activation to reduce model overconfidence in OOD data. Shama et al. in [42] utilized Gram matrices to measure feature correlations for OOD detection. Bibas et al. in [2] proposed pNML regret to detect OOD samples with a single-layer neural network. The Generalized-ODIN approach [23] decomposes the confidence of class probability using a dividend/divisor structure to incorporate prior knowledge. Another promising line of work focuses on designing new learning objectives to train + +![](images/760d42cd0d1840d52c8fc98770bd29c917eefb93770bdbbe6ff06e498505f1a5.jpg) +Figure 1. Illustration of the training pipeline of hierarchical visual category modeling, written as HVCM. In HVCM, we jointly learn the visual representation and parameters of probabilistic models. We get two different views of an input image and send the outputs into a knowledge distillation framework (DINO [6]). Image representations are projected into a high-dimensional attribute space. Then these attributes are divided into different groups and pass SoftMax functions to get attribute distributions. We match attributes in each group with stored attribute centers of the target visual category. The whole model is trained in an end-to-end manner. + +deep models from scratch. Leave-out Classifiers [52] introduces a margin loss to ensure that InD and OOD samples are separated in the feature space. Lee et al. in [28] proposed a novel confidence loss to give lower confidence for OOD samples. There are also some methods [58, 41, 35] designed to conduct OOD detection based on generative models. Unlike these methods, our proposed method jointly conducts representation learning and density estimation to improve the OOD detection performance and keep the learned feature with strong discriminative ability. + +Density Estimation in Deep Learning. Density estimation tries to describe the probabilistic density distribution of observed data accurately and has been investigated in deep learning for a long time [47, 11, 31]. In [1], Chong et al. employed variational autoencoders for anomaly detection, where the reconstruction probability was used to compute the anomaly score of each sample. Papamakarios et al. [36] proposed a novel method for density estimation based on masked autoregressive flow. Zhou et al. [62] extended variational autoencoders by incorporating a mixture of Gaussians to model the latent space distribution, allowing for more flexible and expressive representations. Yang et al. [13] combined Flow-based generative models with generative adversarial networks for density estimation and sample generation. Zhao et al. [11] used discrete latent variables to conduct density estimation, which has been applied to many research topics in natural language processing and image processing. In this paper, we build probabilistic models that can be jointly learned with deep networks. We wish + +latent representations of the visual categories could easily be modeled by Gaussian mixture models even when complex distributions exist. Besides, we do not use post-hoc methods to conduct feature learning and density estimation separately. We aim to use the probabilistic model to guide the feature-learning process of deep networks. + +# 3. Method + +# 3.1. Framework + +State-of-the-art methods [21, 30, 29, 24, 46, 32] typically assume all InD data follow the same distribution. However, such an assumption will lead to the decision boundary of an OOD detector being doped with some uncertain space. In this paper, we avoid modeling the whole InD dataset and only focus on independently modeling each object category's distribution. We propose hierarchical visual categories modeling to achieve this goal while maintaining high classification accuracy. In hierarchical visual categories modeling, we first project image features into a high-dimensional attribute space (usually 8192 dimensions). These attributes can be grouped into multiple sub-visual concepts as components of an image category which can be easily modeled by Gaussian distributions. Then combinations of multiple sub-visual concepts (abbreviated as sub-concepts) can be grouped to describe a more complex visual concept of an in-distribution category. Since we define visual categories based on sub-visual concepts, simply modeling distributions of these sub-concepts and de + +scribing visual categories with these sub-concepts hierarchically can model complicated training distributions. + +Formally, given a visual recognition model $f$ , it maps an input image $\mathbf{x}$ with label $y$ to a high-dimensional feature vector $z \in \mathbb{R}^q$ . As described above, we project $z$ into an attribute space $S \in \mathbb{R}^d$ with a higher dimension and get an attribute description $\mathbf{a}$ . Attributes in $\mathbf{a}$ can be grouped into multiple attribute subgroups $\{a_i\}_{i=1}^G$ where $a_i \in S_i \subset \mathbb{R}^{d/G}$ and $i \in \{1,2,\dots,G\}$ . For images from a visual category $c$ , we assume that attributes in their $i$ -th attribute group follow a simple Gaussian distribution $\mathcal{N}(\mu_i^c, \Sigma_i^c)$ , where $\mu_i^c$ and $\Sigma_i^c$ are the mean and variance respectively. Since attributes are divided into $G$ groups, we will have $G$ different Gaussian distributions for each visual category. The benefits of dividing the whole attribute space into several subspaces come from two folds: (1) modeling attribute distributions in these groups becomes easier; (2) assemble of distributions in these attribute groups can describe complex distributions, which will lead to more accurate decision boundaries. To combine all attributes to describe every visual category, our OOD detector learns weights $\{w_i^c\}_{i=1}^G$ of all attribute groups through exponential moving averages. Therefore, for each in-distribution class $c$ , we can model the corresponding visual concept in a probability perspective with a mixture of Gaussian models: + +$$ +p \left(\boldsymbol {x}; c\right) = \sum_ {i = 1} ^ {G} w _ {i} ^ {c} \mathcal {N} \left(\boldsymbol {a} _ {i}; \boldsymbol {\mu} _ {i} ^ {c}, \boldsymbol {\Sigma} _ {i} ^ {c}\right), \tag {1} +$$ + +where $w_{i}^{c}\in \mathbb{R},\pmb{\mu}_{i}^{c}\in \mathbb{R}^{d / G},\pmb{\Sigma}_{i}^{a}\in \mathbb{R}^{d / G\times d / G}$ and $\mathcal{N}(\cdot)$ means a normal distribution. We build a Gaussian mixture model for each class and get $C$ different Gaussian mixture models. Then, our proposed HVCM focuses on training deep neural networks to learn image features that follow the above distributions and parameters of these probability models jointly. + +With the above Gaussian probability models, given a test sample $\pmb{x}^{\prime}$ , we define the score function $g(\pmb{x}^{\prime};\pmb{w}^{c},\pmb{\mu}^{c},\pmb{\Sigma}^{c})$ to measure whether it belongs to the $c$ -th visual category using the learned probability density function. Here, $\pmb{w}^{c} = \{w_{i}^{c}\}_{i = 1}^{G}$ , $\pmb{\mu}^{c} = \{\pmb{\mu}_{i}^{c}\}_{i = 1}^{G}$ and $\pmb{\Sigma}^{c} = \{\pmb{\Sigma}_{i}^{c}\}_{i = 1}^{G}$ . This score function can be used as a reliable metric to detect OOD samples: + +$$ +h \left(\boldsymbol {x} ^ {\prime}\right) = \left\{ \begin{array}{l l} \operatorname {I n D}, & \text {i f} \min _ {c} g \left(\boldsymbol {x} ^ {\prime}; \boldsymbol {w} ^ {c}, \boldsymbol {\mu} ^ {c}, \boldsymbol {\Sigma} ^ {c}\right) \geq \gamma , \\ \operatorname {O O D}, & \text {i f} \min _ {c} g \left(\boldsymbol {x} ^ {\prime}; \boldsymbol {w} ^ {c}, \boldsymbol {\mu} ^ {c}, \boldsymbol {\Sigma} ^ {c}\right) < \gamma , \end{array} \right. \tag {2} +$$ + +where $\gamma$ are thresholds to be determined in subsequent sections. Eq. (2) indicates that we use the minimal InD score among all $C$ categories to make the final decision. + +Our framework is illustrated in Figure 1, where there are two steps: (1) jointly learn deep features that fit our probability models and parameters of these models; (2) calculate + +the InD score based on a set of Gaussian mixture models as the metric for out-of-distribution detection. In the following subsections, we give more details. + +# 3.2. Joint Visual Representation Learning and Parameter Optimization of Probability Models + +To learn visual representations that follow Gaussian mixture models and keep their discriminative ability at the same time, we exploit the knowledge distillation framework DINO [6] to perform the joint learning. As shown in Figure 1, for an image $\pmb{x}$ , we sample ten different views of $\pmb{x}$ and send them into teacher and student branches simultaneously to perform self-distillation. During knowledge distillation, we project ResNet50 [19] features in 2048 dimensions into an attribute space with dimension $d(= 8192)$ . Apart from the learning objective produced by self-distillation, we force the attributes of each class to follow a class-specific Gaussian mixture model. We divide the attributes $\pmb{a} \in \mathbb{R}^d$ of $\pmb{x}$ into $G$ groups to learn the Gaussian mixture model parameters. However, in practice, it's too hard to directly learn the mean and variance of Gaussian models. We follow He et. al [20] to learn attribute centers $\{\pmb{\mu}_i^c\}_{i = 1}^G$ of the $c$ -th category (the image label $y$ is $c$ ). A linear projection layer is exploited to predict the weights $\{w_i^c (\pmb {x})\}_{i = 1}^G$ of $\pmb{x}$ on all $G$ attribute groups. Our learning objective can be written as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {K D} + \alpha \sum_ {i = 1} ^ {G} \mathrm {K L} \left(\boldsymbol {a} _ {i} \mid \mid \boldsymbol {\mu} _ {i} ^ {c}\right) + \beta \sum_ {i = 1} ^ {T} w _ {i} ^ {c} (\boldsymbol {x}) \mathrm {K L} \left(\boldsymbol {\mu} _ {i} ^ {c} \mid \mid \boldsymbol {a} _ {i}\right) \tag {3} +$$ + +where $\mathcal{L}_{KD}$ stands for the cross entropy loss in self-distillation, KL stands for the Kullback-Leibler(KL) divergence, and $\alpha$ and $\beta$ are hyperparameters. The reason why we exploit two KL divergences is that experiments indicate that the term $\mathrm{KL}(\pmb{a}_i||\pmb{\mu}_i^c)$ will favor learning attribute centers and the term $\mathrm{KL}(\pmb{\mu}_i^c||\pmb{a}_i)$ is more suitable to learn image attribute descriptions and group weights. Note that softmax operations normalize attributes and their learnable centers in each group before calculating learning objectives. + +With the learning objective in Eq. (3), student network parameters $\theta_{s}$ , and weights $\{w_i^c\}_{i = 1}^G$ and attribute centers $\{\pmb {\mu}_i^c\}_{i = 1}^G$ of all groups are learnt at the same time. They are updated using the following equations: + +$$ +\theta_ {s} ^ {t + 1} = \theta_ {s} ^ {t} - \gamma_ {1} \frac {\partial \mathcal {L}}{\partial \theta_ {s} ^ {t}}, \tag {4} +$$ + +$$ +\boldsymbol {\mu} _ {i} ^ {c, t + 1} = \boldsymbol {\mu} _ {i} ^ {c, t} - \gamma_ {2} \frac {\partial \mathcal {L}}{\partial \boldsymbol {\mu} _ {i} ^ {c , t}}, \tag {5} +$$ + +$$ +\boldsymbol {w} _ {i} ^ {c, t + 1} = (1 - \gamma_ {3}) \boldsymbol {w} _ {i} ^ {c, t} + \gamma_ {3} w _ {i} ^ {c} (\boldsymbol {x}), \tag {6} +$$ + +where $\gamma_{1},\gamma_{2}$ and $\gamma_{3}$ are the learning rates. $\gamma_{1},\gamma_{2}$ control the gradient update speed, while $\gamma_{3}$ controls the updating speed of the group weights. Following He et. al [20], $\theta_s,\{\pmb {w}^c\}_{c = 1}^C$ and $\{\pmb {\mu}^c\}_{c = 1}^C$ are initialized with Gaussian noises. We use + +![](images/86bd137e38afbec0ea0d638fb82abf313231d1da57021098cda98e3b8d685939.jpg) +Figure 2. Illustration of attribute group visualization and Mahalanobis distance distribution. Both attribute groups are visualized by t-SNE [49]. The colors encode different in-distribution data(ImageNet), and out-of-distribution features(SUN) marked as gray points. Models are trained on ResNet-50 [18] using DINO(a) and HVCM(b). + +Adam optimizer [25] with momentum to update both $\theta_{s}$ and $\{\pmb{\mu}^c\}_{c=1}^C$ . While the attribute weights are learned through exponential moving averages [17]. + +# 3.3. Out-of-Distribution Detection based on Probability Models in Hierarchy + +When the training of our hierarchical probability model converges, we will get visual attributes that follow learned Gaussian distributions hierarchically for in-distribution samples. Meanwhile, we also get weights and centers of all attribute groups for each class. However, the mean attributes are updated frequently during the training and are thus not suitable to be used as group centers. So we need to recalculate the attribute centers for each visual category. Given an in-distribution visual category $c$ , we estimate the mean vector and covariation matrix of the $i$ -th attribute group: + +$$ +\boldsymbol {\mu} _ {i} ^ {c} = \frac {1}{N _ {c}} \sum_ {m = 1} ^ {N _ {c}} \boldsymbol {a} _ {i} ^ {m} \tag {7} +$$ + +$$ +\boldsymbol {\Sigma} _ {i} ^ {c} = \frac {1}{N _ {c} - 1} \sum_ {m = 1} ^ {N _ {c}} \left(\boldsymbol {a} _ {i} ^ {m} - \boldsymbol {\mu} _ {i} ^ {c}\right) \left(\boldsymbol {a} _ {i} ^ {m} - \boldsymbol {\mu} _ {i} ^ {c}\right) ^ {\top}, \tag {8} +$$ + +where $N_{c}$ notes the number of samples in the $c$ -th class, and $\pmb{a}_i^m$ is the sub-attribute vector of the $m$ -th sample. With these weights, means, and covariances, we can describe each visual category in the hierarchy accurately. We try to use the probability density function in Eq. (1) as the in-distribution function. However, we will encounter the + +problem of numerical overflow when calculating the determinants of covariance matrices as [48]. Instead, we compute the Mahalanobis distance between the $i$ -th sub-visual attributes $a_{i}^{\prime}$ of a test sample $\pmb{x}^{\prime}$ and the corresponding attribute center $\mu_{i}^{c}$ to measure the likelihood of these attributes belongs to the target category: + +$$ +M _ {i} ^ {c} (\boldsymbol {x}) = - \left(\boldsymbol {a} _ {i} ^ {\prime} - \boldsymbol {\mu} _ {i} ^ {c}\right) ^ {\top} \left(\boldsymbol {\Sigma} _ {i} ^ {c}\right) ^ {- 1} \left(\boldsymbol {a} _ {i} ^ {\prime} - \boldsymbol {\mu} _ {i} ^ {c}\right). \tag {9} +$$ + +The above equation gives the in-distribution measure for one attribute group. While, for every category, we have multiple attribute groups and need to assemble related in-distribution measures to get the class-level in-distribution score. Since we have gotten the weights of attribute groups for each category, we can easily assemble them and get the class-level score function: + +$$ +g \left(\boldsymbol {x} ^ {\prime}; \boldsymbol {w} ^ {c}, \boldsymbol {\mu} ^ {c}, \boldsymbol {\Sigma} ^ {c}\right) = \sum_ {i = 1} ^ {G} w _ {i} ^ {c} M _ {i} ^ {c} \left(\boldsymbol {x} ^ {\prime}\right). \tag {10} +$$ + +With this score function, we can easily get the in-distribution score of a test sample on each visual category. Since there are $C$ categories in the whole in-distribution dataset, we get the maximal in-distribution score across different visual categories as the in-distribution score on the whole dataset: + +$$ +g \left(\boldsymbol {x} ^ {\prime}\right) = \max _ {c} g \left(\boldsymbol {x} ^ {\prime}; \boldsymbol {w} ^ {c}, \boldsymbol {\mu} ^ {c}, \boldsymbol {\Sigma} ^ {c}\right). \tag {11} +$$ + +A high in-distribution score $g(\boldsymbol{x}^{\prime})$ indicates that the semantic attributes of a test sample lie very close to one or multiple in-distribution visual categories (as shown in Figure 2). On the contrary, if a sample does not belong to the previously modeled categories, it will get a low in-distribution score. We follow Eq. (2) to set thresholds to judge whether a sample is an out-of-distribution sample. In the experiment section, we discuss how to set these thresholds. + +# 4. Experiments + +# 4.1. Experimental Setup + +In-distribution Datasets. We use ImageNet-1K [40] and CIFAR10 [27] as our in-distribution datasets. ImageNet-1K is a large-scale visual recognition dataset containing 1000 object categories and 1281167 images. We split it into 10 folds randomly and ensured each fold contains 100 object categories. Since our computation resources are limited, we randomly select one fold as the in-distribution dataset. There other nine folds are used as OOD datasets as other popular benchmarks to test the performance of the proposed method in detecting near OOD samples. For CIFAR10 [27], there are 60000 color images in 10 classes. We conduct OOD algorithm evaluation as previous methods [30, 9, 46, 32, 45, 54]. + +Table 1. OOD detection performance comparison of HVCM and existing methods. All comparison methods rely on ResNet-50 trained with cross-entropy loss. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages. + +
MethodOOD DatasetsAverageInD Acc
iNaturalistSUNPlacesTexturesFPR95↓AUROC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]68.1287.4858.5989.7659.5389.1872.6682.7164.7387.28
ODIN [30]55.7885.9260.4783.8360.9488.0664.0681.2860.3184.77
Maha [29]97.0055.1298.8051.9397.2050.5220.0094.9978.2563.14
Energy [32]58.9186.4527.0393.4438.7591.5156.2587.2645.2489.6785.74
GODIN* [23]72.0079.8660.0984.5864.9482.3039.5089.2859.1384.01
MOS* [24]52.9491.4167.7886.8271.3184.3873.6580.5666.4285.79
ReAct [45]58.4882.6078.1869.1186.3359.8450.5387.0368.3874.65
HVCM(Ours)21.5692.1917.2094.4419.9893.6229.2290.6821.9992.7388.28
+ +Table 2. Evaluation on more challenging detection tasks. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages. + +
MethodImageNet-OOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]72.5083.3389.1157.99
ODIN [30]73.4471.0364.0679.88
Maha [29]54.4079.3059.2077.62
Energy* [32]66.5681.0662.3484.61
GODIN* [23]71.5579.8973.5777.27
MOS* [24]87.4064.8766.7881.42
ReAct [45]87.0564.1584.1764.30
HVCM(Ours)42.8686.7228.5890.26
+ +Out-of-distribution Dataset. On ImageNet, we follow Huang et al. [24] to test our methods and use Texture [8], iNaturalist [50], Places365 [61], and SUN [57]) as OOD test sets. To further explore the limitation of our approach, we evaluate our method on another two OOD datasets, OpenImage-O [26] and ImageNet-O [22]. For CIFAR10, as in [46, 32, 45, 54], we selected eight widely-used datasets, including Texture [8], SVHN [34], Places365 [61], iSUN [59], LSUN-Crop [60], LSUN-Resize [60], ImageNet-Resize [40], and ImageNet-Fix [40] as our test sets. Besides, to test the ability of HVCM to identify near OOD datasets, we use the remaining nine ImageNet subsets as the OOD test sets. Note that since our evaluation on ImageNet differs from previous methods [24, 45], we implement these algorithms with open source provided by authors and follow standard experimental settings. + +Evaluation Metrics. We employ the commonly used metrics in OOD detection [24, 45] to evaluate our approach, including AUROC, FPR95 and InD Acc. AUROC stands for the area under the receiver operating characteristic curve, FPR95 is short for TPR@FPR95 and represents the false positive rate when the true positive rate is $95\%$ , and InD Acc is the classification accuracy of in-distribution data. + +Training Details. We utilize ResNet-50 [19] as the feature + +backbone for ImageNet and the dimension of the attribute space is set to 8192. The training is finished in 300 epochs. On CIFAR10, we use ResNet-18 [19] as our feature backbone, and the dimension of the attribute space is set to 1024. The training on CIFAR10 is finished in 200 epochs. The number of attribute groups is set to 32, and $\alpha$ , $\beta$ , $\gamma_1$ , $\gamma_2$ , and $\gamma_3$ are set to 1, 0.1, 1, 1 and $1 \times 10^{-4}$ , respectively. Updating $\mu$ too fast can lead to the oscillation of group centers, negatively influencing the precision of probabilistic modeling. So we utilize a smaller $\beta$ than $\alpha$ to update $\mu$ . All the hyperparameters are tuned according to the experimental results. We employ SGD with a momentum of 0.9, an initial learning rate of 0.1, and a batch size of 128. The learning rate is reduced by a factor of 10 at $50\%$ and $75\%$ of the total training epochs. We train all backbones from scratch using random initialization. All experiments are performed using PyTorch [37] with default parameters on four NVIDIA GeForce RTX 3090. + +# 4.2. Comparison with State-of-the-Art Algorithms + +Standard evaluation on ImageNet. We compare our HVCM with seven popular OOD detection methods, including MSP [21], ODIN [30], GODIN [23], Maha [29], Energy [32], MOS [24], and ReAct [45]. For datasets that describe objects or scenes, such as SUN, Places, and iNaturalist, HVCM achieves better AUROC and FPR95 metrics. When we summarize the results of all four datasets, HVCM achieves $21.99\%$ on FPR95 and $92.73\%$ on AUROC, which outperforms the previous best method Energy [32] by $23.25\%$ and $3.06\%$ . This is a significant improvement, which demonstrates that end-to-end training is very important to get good results. When compared to Maha [29], our proposed method exhibits inferior performance in terms of both FPR95 and AUROC. This observation suggests that our method is less effective in describing textures. We attribute this limitation to the fact that textures often encompass numerous repeated patterns, which differ from the characteristics required for general object + +Table 3. Comparison of OOD detection performance of HVCM and existing methods on CIFAR10 dataset. All compared methods use ResNet-18 trained with cross-entropy loss except our proposed method, which uses HVCMLoss. The performance is evaluated based on AUROC (A) and FPR95 (F). $\uparrow$ indicates larger values are better and $\downarrow$ indicates the opposite. Bold numbers indicate superior results. All values are expressed in percentages. + +
MethodOOD DatasetAverage
TextureSVHNPlaces365iSUNLSUN(C)LSUN(R)ImageNet(R)ImageNet(F)
F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑
CELossMSP56.4790.2058.4090.5651.8691.9853.8491.6446.2292.7449.1092.4859.6589.3559.9089.4054.4391.04
ODIN40.3791.9827.8293.2831.8094.2317.5996.6025.2395.3114.7497.2127.1894.6545.2690.6228.7594.24
Gram10.8197.732.5899.3930.1593.371.1199.7613.9197.030.5299.861.6198.3866.9482.0215.9595.94
Gram+pNML7.1898.501.6399.6023.2195.130.8399.809.4298.000.4299.881.2498.7657.9085.5312.7396.90
HVCMLossHVCM(ours)1.8899.311.3299.470.9599.540.6599.710.7799.660.5199.961.8099.204.7798.191.5899.38
+ +![](images/da919b464e0ae709993ce4e367f77c534caab7e229e8b77e6a2242ad410f9db8.jpg) +Figure 3. HVCM performance comparison as the increasing distances between InD and OOD data. + +recognition. It is important to note that our method only utilizes features from the last layer of the network, while Maha [29] leverages features from both intermediate and deep network layers. However, although our results are worse than Maha [29], we are better than all the other methods. This still indicates that the proposed method is very robust in identifying different types of outliers. Furthermore, we build a cosine classifier with the learned attribute centers for image classification. We got $88.28\%$ accuracy which is $2.57\%$ higher than our supervised learning baseline. This is solid evidence to show that our proposed method can model the InD data accurately and ensure the learned features keep high discriminative ability simultaneously. + +Standard evaluation on CIFAR10. We performed a more conventional OOD detection task on CIFAR10. This experiment aims twofold: firstly, to demonstrate that our method is not dependent on self-distillation (DINO [6]), and secondly, to validate the effectiveness and robustness of our HVCM. Table 3 compares our method with several classic and top-performing algorithms. All comparison methods use ResNet-18 as the main backbone network and are trained with cross-entropy loss, while our method uses only the loss function in Eq. (3). As shown in Table 3, the proposed method outperforms the previous best methods Gram [42] and pNML [2] on both average FPR95 and AUROC obviously. These results demonstrate the proposed method can perform well even on small datasets. Mean + +Table 4. The performance of HVCM with varying numbers of group centers $G$ . Results are averaged across four standard OOD datasets, consistent with the main results. + +
Concepts NumberFPR95↓AUROC↑
G=824.3791.73
G=1623.0792.24
G=3221.9992.73
+ +while, there is no self-distillation exploited, which indicates our joint representation learning and statistical modeling is independent of self-supervised learning algorithms [6]. + +Results on more challenging OOD datasets. To overcome the limitations of current OOD benchmarks [53] and evaluate the robustness of our approach against adversarial attacks, we conducted experiments on two challenging datasets, namely OpenImage-O [26] and ImageNet-O [22]. As shown in Table 2, HVCM achieves the highest AUROC and lowest FPR95 among all methods on the OpenImage-O dataset. Although ImageNet-O contains adversarial examples and is more challenging, HVCM still outperforms other methods on this dataset. + +Results on near-to-far OOD datasets. To investigate the ability of the proposed method to detect near OOD samples, we construct 9 OOD test sets with the remaining ImageNet images. We rank the semantic distance between the remaining 900 visual categories with the 100 categories in the InD dataset. We use the average cosine distance as the measure and construct 9 different OOD test sets. Details are introduced in the supplementary material. For convenience, we denote these datasets from OOD 1 to OOD 9. The experimental results are depicted in Figure 3, and several samples are displayed in Figure 4. We can find that the proposed method achieves good AUROC even when the test set is very close to the InD dataset. When the test sets become farther, FPR95 decreases quickly, which indicates the proposed method is very sensitive to the semantic distances of OOD datasets. + +![](images/d1c79ce06faf27ee74bfd7b4a2c70170003acac168ce8aaa49399a37655b8327.jpg) +Figure 4. Each dataset exhibits displayed two categories of images. The leftmost samples belong to the InD dataset, while the categories on the right correspond to nine OOD datasets arranged in ascending order of distance. It can be observed that the gap between the OOD samples and InD samples gradually widens as the distance increases. + +![](images/79651569599fe1d7aca39573a1699880ee4a1289dc51241faf73ac364295315a.jpg) + +![](images/5dc7a8b3738a0cf3d7598a81fa2fb85c8d552a1283b0872119f7a1b187d4a756.jpg) + +![](images/6f4ce035735d759670a962608b80b58d238afd3785c623f59ae38cb04e8679ba.jpg) + +![](images/47ed9d314f95959e1c6602f53c134914d98af91d49c9408e538ad2e1099b0d7b.jpg) + +![](images/5be6a798eff03983b97e80e9ac01c488c711fb6728032ad59a80b7d363ddc396.jpg) + +![](images/011fe1e45a40ba52df17c45226c7da0cd740d1dddc56cb96df60d1f96e1dfd21.jpg) + +![](images/e5936316381cac47bd3ded9236f78ee5a778445f1e04d1e45e8a33e166ba3461.jpg) + +![](images/3afd3b53e5037dde691369b79ebd6a707edb32d3d29bf2ca7b69a24f49946f40.jpg) + +![](images/7bd8ebf552e027c3b5072d428aecb91dd41495a42176e999b66b29157a0cefcb.jpg) + +Table 5. A set of ablation results about HVCM. The top row investigates the effect of using MSE, KL divergence, or JS divergence on the model performance; the bottle row compares the performance of different OOD detection methods. Results are averaged across four standard OOD datasets following the main results. + +
Strategy AblationFPR95↓AUROC↑
Learning ObjectiveL223.5992.26
KL22.9992.58
JS21.9992.73
InD DistanceCosine66.3784.06
Linear36.7786.59
Maha21.9992.73
+ +# 4.3. Ablation Study + +Number of attribute groups. We varied the number of attribute groups from 8 to 32 to analyze the components in Gaussian mixture models. In Table 4, we find a positive correlation between the number of attribute groups and model performance, with the best performance achieved when $G$ is set to 32. We also try to set $G$ to bigger numbers and list the results in the supplementary material. However, more attribute groups will lead to bigger correlation matrices when computing the InD score. Thus, we finally set $G$ to 32 to balance performance and inference speed. + +Choice of learning objectives. The first row of Table 5 investigates how the choice of the learning objective influences the performance of HVCM. We test three learning objectives, including the L2 loss, JS divergence loss, and KL divergence loss. The results show that compared with the L2 loss and KL divergence loss, the JS divergence loss achieves the lowest FPR95, demonstrating its superiority for statistical modeling. We attribute this to the fact that the group centers and features need to learn from each other, and the JS divergence loss is symmetrical in enclosing them. + +Different InD distance metrics. In Table 5, we also compared our Maha metric with two different InD distance metrics. The cosine distance metric directly measures the distance by calculating the cosine similarity between the input feature and the mean of Gaussian distribution models. The + +linear distance metric is used to calculate the distance with trainable linear layers. The results show that our Maha metric is an effective metric compared with its counterparts. We attribute this to the Maha distance space can better fit the training distribution in realistic scenes. + +![](images/285be144585e931334d9d0e34d834848d001ef2bcd9292cae5e44e8cff5caf4a.jpg) +Figure 5. The performance of HVCM is evaluated as the number of InD classes increases across four OOD datasets. + +Increasing numbers of InD Categories in ImageNet. We test how the OOD detection performance varies with the increasing of object categories in the in-distribution dataset. Following Wang et al. [53], we test on four popular benchmarks and set $C$ to $\{50, 100, 200, 300\}$ respectively. As depicted in Figure 5, the performance of HVCM fluctuates on different datasets with the increasing of the InD object categories, which suggests that the number of categories has little impact on our approach. These results validate our assumption that we only need to model in-distribution image categories, and the out-of-distribution samples can be detected easily then. + +Different thresholds for OOD detection. Figure 6 illustrates the accuracy of OOD detection for our method across various datasets. On most datasets, our method exhibits the same trend for accuracy variation with thresholds. This indicates that our approach has strong generalization and ideal performance to domains with significant differences. We explain the performance of Imagenet-O as its task difficulty with adversarial samples. + +![](images/baf1b3963df841cae5c7bf8398f96ca7a512ee5e051b2e91981f86624548fa52.jpg) +Figure 6. The accuracy of HVAC varies with different thresholds across all OOD datasets. + +# 5. Conclusion + +In this paper, we introduce a hierarchical visual category modeling scheme for out-of-distribution detection, which combines visual representation learning and parameter optimization of probability models. It provides a novel perspective for OOD detection by conducting representation learning and density estimation end-to-end. By modeling visual categories with mixtures of Gaussian models, we describe visual categories in very complex distribution and don't rely on outlier training data to perform OOD detection. Experiments demonstrate that the proposed method outperforms state-of-the-art algorithms clearly and does not hinder the discriminative ability of deep features. + +Limitations. However, our method needs to map deep features into high-dimensional attribute spaces and build plentiful Gaussian mixture models. These Gaussian mixture models bring a lot of computational costs and make the inference of the OOD detector become inefficient. Thus, simplifying the probability models and accelerating the inference process will be the future direction. + +# 6. Acknowledgment + +This work was supported by National Natural Science Foundation of China (No.62072112), Scientific and Technological innovation action plan of Shanghai Science and Technology Committee (No.22511102202), National Key R&D Program of China (2020AAA0108301), National Natural Science Foundation of China under Grant (No. 62106051), Shanghai Pujiang Program (No. 21PJ1400600), and National Key R&D Program of China (2022YFC3601405). + +# References + +[1] Jinwon An and Sungzoon Cho. Variational autoencoder based anomaly detection using reconstruction probability. Special lecture on IE, 2(1):1-18, 2015. +[2] Koby Bibas, Meir Feder, and Tal Hassner. Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems, 34:1179-1191, 2021. +[3] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. +[4] Daniel Bogdoll, Jasmin Breitenstein, Florian Heidecker, Maarten Bieshaar, Bernhard Sick, Tim Fingscheidt, and Marius Zöllner. Description of corner cases in automated driving: Goals and challenges. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1023-1028, 2021. +[5] Senqi Cao and Zhongfei Zhang. Deep hybrid models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2022. +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. +[7] Xingyu Chen, Xuguang Lan, Fuchun Sun, and Nanning Zheng. A boundary based out-of-distribution classifier for generalized zero-shot learning. In European Conference on Computer Vision, pages 572-588. Springer, 2020. +[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. +[9] Roy De Maesschalck, Delphine Jouan-Rimbaud, and Désiré L Massart. The mahalanobis distance. Chemometrics and intelligent laboratory systems, 50(1):1-18, 2000. +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. +[11] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016. +[12] Zhen Fang, Jie Lu, Anjin Liu, Feng Liu, and Guangquan Zhang. Learning bounds for open-set learning. In International Conference on Machine Learning, pages 3122-3132. PMLR, 2021. +[13] Aditya Grover, Manik Dhar, and Stefano Ermon. Flow-gan: Combining maximum likelihood and adversarial learning in generative models. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. +[14] Xiaoyuan Guo, Judy Wawira Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Cvad: A generic + +medical anomaly detector based on cascade vae. arXiv preprint arXiv:2110.15811, 2021. +[15] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9591-9600, 2022. +[16] Matan Haroush, Tzviel Frostig, Ruth Heller, and Daniel Soudry. A statistical framework for efficient out of distribution detection in deep neural networks. In International Conference on Learning Representations, 2021. +[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. +[20] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022. +[21] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. Proceedings of International Conference on Learning Representations, 2017. +[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. +[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. +[24] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[26] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, et al. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2(3):18, 2017. +[27] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009. + +[28] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In International Conference on Learning Representations, 2018. +[29] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018. +[30] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. +[31] Qiao Liu, Jiaze Xu, Rui Jiang, and Wing Hung Wong. Density estimation using deep generative neural networks. Proceedings of the National Academy of Sciences, 118(15):e2101344118, 2021. +[32] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020. +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. +[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. +[35] Philipp Oberdiek, Gernot A Fink, and Matthias Rottmann. Uqgan: A unified model for uncertainty quantification of deep classifiers trained via conditional gans. arXiv preprint arXiv:2201.13279, 2022. +[36] George Papamakarios, Theo Pavlakou, and Iain Murray. Masked autoregressive flow for density estimation. Advances in neural information processing systems, 30, 2017. +[37] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. +[38] Chen Qiu, Aodong Li, Marius Kloft, Maja Rudolph, and Stephan Mandt. Latent outlier exposure for anomaly detection with contaminated data. arXiv preprint arXiv:2202.08088, 2022. +[39] Carl Rasmussen. The infinite gaussian mixture model. Advances in neural information processing systems, 12, 1999. +[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. +[41] Murat Sensoy, Lance Kaplan, Federico Cerutti, and Maryam Saleki. Uncertainty-aware deep classifiers using generative + +models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5620-5627, 2020. +[42] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with in-distribution examples and gram matrices. arXiv e-prints, pages arXiv-1912, 2019. +[43] Nina Shvetsova, Bart Bakker, Irina Fedulova, Heinrich Schulz, and Dmitry V. Dylov. Anomaly detection in medical imaging with deep perceptual autoencoders. IEEE Access, 9:118571-118583, 2021. +[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[45] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In Advances in Neural Information Processing Systems, 2021. +[46] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. ICML, 2022. +[47] Esteban G Tabak and Cristina V Turner. A family of nonparametric density estimation algorithms. Communications on Pure and Applied Mathematics, 66(2):145-164, 2013. +[48] Shengjin Tang, Chuanqiang Yu, Xiaoyan Sun, Hongdong Fan, and Xiaosheng Si. A note on parameters estimation for nonlinear wiener processes with measurement errors. IEEE Access, 7:176756-176766, 2019. +[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. +[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[52] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), pages 550-564, 2018. +[53] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[54] Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. Mitigating neural network overconfidence with logit normalization. In International Conference on Machine Learning, pages 23631-23644. PMLR, 2022. +[55] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y. Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Nicholas Petrick and Kensaku Mori, editors, Medical Imaging 2018: Computer-Aided Diagnosis, volume 10575, pages 375 - 380. International Society for Optics and Photonics, SPIE, 2018. + +[56] Jim Winkens, Rudy Bunel, Abhijit Guha Roy, Robert Stanforth, Vivek Natarajan, Joseph R. Ledsam, Patricia MacWilliams, Pushmeet Kohli, Alan Karthikesalingam, Simon Kohl, Taylan Cemgil, S. M. Ali Eslami, and Olaf Ronneberger. Contrastive training for improved out-of-distribution detection. arXiv preprint arXiv:2007.05566, 2020. +[57] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. +[58] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. Advances in neural information processing systems, 33:20685-20696, 2020. +[59] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015. +[60] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. +[61] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. +[62] Cangqi Zhou, Hao Ban, Jing Zhang, Qianmu Li, and Yinghua Zhang. Gaussian mixture variational autoencoder for semi-supervised topic modeling. IEEE Access, 8:106843-106854, 2020. \ No newline at end of file diff --git a/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/images.zip b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eae867f988a5e4759874906b5fbae57f5254fe84 --- /dev/null +++ b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a088f012c262172faff8f3042e6d2d39b9ae3c93c91d8522e0db8e5814925841 +size 551880 diff --git a/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/layout.json b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fd248b488bf41f10ab21cba54ccc26d0484d432 --- /dev/null +++ b/2023/Hierarchical Visual Categories Modeling_ A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection/layout.json @@ -0,0 +1,9121 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 136, + 526, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 136, + 526, + 170 + ], + "spans": [ + { + "bbox": [ + 51, + 136, + 526, + 170 + ], + "type": "text", + "content": "Hierarchical Visual Categories Modeling: A Joint Representation Learning and Density Estimation Framework for Out-of-Distribution Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "spans": [ + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "text", + "content": "Jinglun Li" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "text", + "content": " Xinyu Zhou" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "text", + "content": " Pinxue Guo" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "text", + "content": " Yixuan Sun" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "text", + "content": " Yiwen Huang" + }, + { + "bbox": [ + 107, + 191, + 470, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 207, + 376, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 207, + 376, + 220 + ], + "spans": [ + { + "bbox": [ + 205, + 207, + 376, + 220 + ], + "type": "text", + "content": "Weifeng Ge²† Wenqiang Zhang¹,²†" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 220, + 433, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 220, + 433, + 234 + ], + "spans": [ + { + "bbox": [ + 143, + 220, + 433, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 220, + 433, + 234 + ], + "type": "text", + "content": "Academy for Engineering and Technology, Fudan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 169, + 234, + 395, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 234, + 395, + 247 + ], + "spans": [ + { + "bbox": [ + 169, + 234, + 395, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 169, + 234, + 395, + 247 + ], + "type": "text", + "content": "School of Computer Science, Fudan University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 250, + 415, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 250, + 415, + 261 + ], + "spans": [ + { + "bbox": [ + 160, + 250, + 415, + 261 + ], + "type": "text", + "content": "jingli960423@gmail.com,zhouxinyu20@fudan.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 155, + 263, + 420, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 263, + 420, + 274 + ], + "spans": [ + { + "bbox": [ + 155, + 263, + 420, + 274 + ], + "type": "text", + "content": "{pxguo21，21210860014，21210240056}@m.fudan.edu.cn," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 170, + 277, + 406, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 277, + 406, + 287 + ], + "spans": [ + { + "bbox": [ + 170, + 277, + 406, + 287 + ], + "type": "text", + "content": "weifeng.ge.ic@gmail.com, wqzhang@fudan.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "spans": [ + { + "bbox": [ + 139, + 314, + 186, + 326 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 340, + 281, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 340, + 281, + 644 + ], + "spans": [ + { + "bbox": [ + 45, + 340, + 281, + 644 + ], + "type": "text", + "content": "Detecting out-of-distribution inputs for visual recognition models has become critical in safe deep learning. This paper proposes a novel hierarchical visual category modeling scheme to separate out-of-distribution data from in-distribution data through joint representation learning and statistical modeling. We learn a mixture of Gaussian models for each in-distribution category. There are many Gaussian mixture models to model different visual categories. With these Gaussian models, we design an in-distribution score function by aggregating multiple Mahalanobis-based metrics. We don't use any auxiliary outlier data as training samples, which may hurt the generalization ability of out-of-distribution detection algorithms. We split the ImageNet-1k dataset into ten folds randomly. We use one fold as the in-distribution dataset and the others as out-of-distribution datasets to evaluate the proposed method. We also conduct experiments on seven popular benchmarks, including CIFAR, iNaturalist, SUN, Places, Textures, ImageNet-O, and OpenImage-O. Extensive experiments indicate that the proposed method outperforms state-of-the-art algorithms clearly. Meanwhile, we find that our visual representation has a competitive performance when compared with features learned by classical methods. These results demonstrate that the proposed method hasn't weakened the discriminative ability of visual recognition models and keeps high efficiency in detecting out-of-distribution samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 666, + 124, + 678 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 124, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 124, + 678 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 686, + 279, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 686, + 279, + 709 + ], + "spans": [ + { + "bbox": [ + 45, + 686, + 279, + 709 + ], + "type": "text", + "content": "Modern deep neural networks have shown strong generalization ability when training and test data are from the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 315, + 530, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 315, + 530, + 456 + ], + "spans": [ + { + "bbox": [ + 297, + 315, + 530, + 456 + ], + "type": "text", + "content": "same distribution [44, 18, 51, 10, 33]. However, encountering unexpected scenarios is inevitable in real-world applications. Thus assuring that training and test data share the same distribution becomes problematic. In applications like autonomous driving [3, 4] and medical image analysis [55, 14, 43], it is critical for models to identify inputs beyond their recognition capacity – known as out-of-distribution (OOD) detection. OOD detection algorithms can enable the system to warn humans promptly in many safety-related scenarios. Moreover, it has become an important research topic in the research community of safe artificial intelligence [21, 29, 23, 32, 56]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 458, + 531, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 458, + 531, + 703 + ], + "spans": [ + { + "bbox": [ + 296, + 458, + 531, + 703 + ], + "type": "text", + "content": "Many popular OOD detection methods aim to build probability models to describe training distributions [29, 56, 46, 24, 23]. With these probability models, they built a score function that can calculate in-distribution scores for test samples. These in-distribution scores reflect whether these samples fall into a given distribution. Then the test sample can be evaluated by the score function to decide whether it is an OOD sample or not. Thus modeling features of in-distribution data become extremely important. Previous works [29, 16, 5, 12, 38] build a distribution over the whole training data. Since training images may come from various visual categories, the decision boundary between In-Distribution (InD) and OOD data becomes extremely complex. To solve this problem, subsequent studies [24, 7, 15, 56] decomposed the whole dataset into several subgroups to simplify the decision boundary. Although representative algorithms like MOS [24], have gotten impressive performance in identifying OOD samples, they failed to detect near OOD samples. Because when different visual categories are grouped together, the OOD decision boundary will become even more uncertain." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": "A typical framework for out-of-distribution detection involves two key steps: 1) learning a compact feature rep" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 41, + 470, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 41, + 470, + 74 + ], + "spans": [ + { + "bbox": [ + 151, + 41, + 470, + 74 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 718, + 163, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 718, + 163, + 730 + ], + "spans": [ + { + "bbox": [ + 57, + 718, + 163, + 730 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 57, + 718, + 163, + 730 + ], + "type": "text", + "content": "indicates corresponding authors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "type": "text", + "content": "23425" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 281 + ], + "type": "text", + "content": "representation that can fit probability models easily; 2) modeling features of in-distribution data in complex distributions accurately. The above two problems are mutually connected because more compact features will make modeling the data distribution easier, and stronger probability modeling techniques will exploit fewer restrictions on representation learning. However, achieving the above goals is difficult, even when there are a lot of breakthroughs in deep learning. Because if training samples from the same category are too close in the feature space, it will usually lead to overfitting. Meanwhile, in-distribution samples may come from different visual categories that have large variations in appearance and semantic information, which makes modeling complex training distributions become challenging even for excellent statisticians." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 284, + 281, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 284, + 281, + 575 + ], + "spans": [ + { + "bbox": [ + 45, + 284, + 281, + 575 + ], + "type": "text", + "content": "In this paper, we propose a new out-of-distribution detection framework, called hierarchical visual category modeling, to solve the above two issues simultaneously. We hold an assumption that given a training set that contains multiple visual categories, we can learn a probability model for every category independently. The out-of-distribution detection problem can be solved easily by aggregating probability models of known categories. Our motivation is that decomposing the whole dataset into subsets and modeling each category independently can avoid finding common characteristics shared by different categories. However, modeling an individual visual category is still challenging since classical supervised learning won't lead to compact feature representation. Thus, for each input sample, we need to force its feature representation to match the corresponding statistical model. That means we need to jointly conduct density estimation and representation learning. If we can jointly learn visual representations and optimize statistical models end-to-end, we can get good feature representations falling into distributions of the corresponding visual categories. Besides, we exploit knowledge distillation as done in [6] to learn robust feature representation. In this way, we can describe the complex training distribution with multiple Gaussian mixture models while not impairing the generalization ability of visual features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 579, + 281, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 579, + 281, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 579, + 281, + 730 + ], + "type": "text", + "content": "In practice, to learn visual concepts that are in complex distributions, we build a Gaussian mixture model (GMM [39]) for each visual category. Given input samples, we extract their deep features and project these features into a high-dimensional attribute space. Different from classical Gaussian mixture models that send the same input into " + }, + { + "bbox": [ + 45, + 579, + 281, + 730 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 579, + 281, + 730 + ], + "type": "text", + "content": " different Gaussian models, we divide the attribute space into multiple groups and build a Gaussian model in each group independently. This strategy can give every attribute group a clear learning target and lead to better convergence. Experimental results indicate that this strategy works quite well. After the visual representation learning and statistical model parameters optimization, we can directly aggre" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 296, + 106, + 532, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 532, + 234 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 532, + 234 + ], + "type": "text", + "content": "gate these statistical models to judge whether a test sample comes from the training distribution or not. To evaluate our OOD detector, we split ImageNet into ten folds randomly and select one of these splits as the training set and all other splits as the OOD dataset to conduct extensive tests. Experiments indicate that the proposed method has a strong ability to identify OOD samples. We also evaluate our method on seven popular OOD benchmarks. Experimental results demonstrate that the proposed method not only can identify OOD samples efficiently but also improves the discriminative ability of learned visual representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 234, + 532, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 234, + 532, + 257 + ], + "spans": [ + { + "bbox": [ + 296, + 234, + 532, + 257 + ], + "type": "text", + "content": "The contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 266, + 530, + 481 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 308, + 266, + 530, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 530, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 530, + 336 + ], + "type": "text", + "content": "- We introduce a new out-of-distribution detection scheme, called hierarchical visual category modeling to conduct joint representation learning and density estimation. It provides a new perspective for out-of-distribution detection to learn visual representation and probability models end-to-end." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 345, + 530, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 345, + 530, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 345, + 530, + 403 + ], + "type": "text", + "content": "- We exploit multiple Gaussian mixture models to model visual concepts in complex distributions. Visual attributes are divided into subgroups and modeled by different Gaussian components, which makes parameter learning much more efficient." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 411, + 530, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 530, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 530, + 481 + ], + "type": "text", + "content": "- We conducted comprehensive experiments and ablation studies on popular benchmarks to investigate the effectiveness of the proposed method. Experiments demonstrated that our out-of-distribution detection models achieve better performance clearly when compared with previous methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 297, + 500, + 381, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 500, + 381, + 512 + ], + "spans": [ + { + "bbox": [ + 297, + 500, + 381, + 512 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 520, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 520, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 520, + 532, + 730 + ], + "type": "text", + "content": "Out-Of-Distribution Detection. Out-of-distribution detection aims to distinguish out-of-distribution samples from in-distribution data. Numerous methods have been proposed for OOD detection. Maximum softmax probability (MSP) [21] has been recognized as a strong baseline by using the maximum score across all classes as an OOD score. ODIN [30] improves MSP by perturbing the inputs and adjusting the logits via rescaling. Gaussian discriminant analysis has been employed in [29, 56] to detect OOD samples. ReAct [45] uses rectified activation to reduce model overconfidence in OOD data. Shama et al. in [42] utilized Gram matrices to measure feature correlations for OOD detection. Bibas et al. in [2] proposed pNML regret to detect OOD samples with a single-layer neural network. The Generalized-ODIN approach [23] decomposes the confidence of class probability using a dividend/divisor structure to incorporate prior knowledge. Another promising line of work focuses on designing new learning objectives to train" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "type": "text", + "content": "23426" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 105, + 526, + 309 + ], + "blocks": [ + { + "bbox": [ + 51, + 105, + 526, + 309 + ], + "lines": [ + { + "bbox": [ + 51, + 105, + 526, + 309 + ], + "spans": [ + { + "bbox": [ + 51, + 105, + 526, + 309 + ], + "type": "image", + "image_path": "760d42cd0d1840d52c8fc98770bd29c917eefb93770bdbbe6ff06e498505f1a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 311, + 531, + 366 + ], + "lines": [ + { + "bbox": [ + 45, + 311, + 531, + 366 + ], + "spans": [ + { + "bbox": [ + 45, + 311, + 531, + 366 + ], + "type": "text", + "content": "Figure 1. Illustration of the training pipeline of hierarchical visual category modeling, written as HVCM. In HVCM, we jointly learn the visual representation and parameters of probabilistic models. We get two different views of an input image and send the outputs into a knowledge distillation framework (DINO [6]). Image representations are projected into a high-dimensional attribute space. Then these attributes are divided into different groups and pass SoftMax functions to get attribute distributions. We match attributes in each group with stored attribute centers of the target visual category. The whole model is trained in an end-to-end manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 383, + 281, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 383, + 281, + 500 + ], + "spans": [ + { + "bbox": [ + 45, + 383, + 281, + 500 + ], + "type": "text", + "content": "deep models from scratch. Leave-out Classifiers [52] introduces a margin loss to ensure that InD and OOD samples are separated in the feature space. Lee et al. in [28] proposed a novel confidence loss to give lower confidence for OOD samples. There are also some methods [58, 41, 35] designed to conduct OOD detection based on generative models. Unlike these methods, our proposed method jointly conducts representation learning and density estimation to improve the OOD detection performance and keep the learned feature with strong discriminative ability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 508, + 281, + 730 + ], + "type": "text", + "content": "Density Estimation in Deep Learning. Density estimation tries to describe the probabilistic density distribution of observed data accurately and has been investigated in deep learning for a long time [47, 11, 31]. In [1], Chong et al. employed variational autoencoders for anomaly detection, where the reconstruction probability was used to compute the anomaly score of each sample. Papamakarios et al. [36] proposed a novel method for density estimation based on masked autoregressive flow. Zhou et al. [62] extended variational autoencoders by incorporating a mixture of Gaussians to model the latent space distribution, allowing for more flexible and expressive representations. Yang et al. [13] combined Flow-based generative models with generative adversarial networks for density estimation and sample generation. Zhao et al. [11] used discrete latent variables to conduct density estimation, which has been applied to many research topics in natural language processing and image processing. In this paper, we build probabilistic models that can be jointly learned with deep networks. We wish" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 383, + 530, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 383, + 530, + 455 + ], + "spans": [ + { + "bbox": [ + 296, + 383, + 530, + 455 + ], + "type": "text", + "content": "latent representations of the visual categories could easily be modeled by Gaussian mixture models even when complex distributions exist. Besides, we do not use post-hoc methods to conduct feature learning and density estimation separately. We aim to use the probabilistic model to guide the feature-learning process of deep networks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 468, + 352, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 468, + 352, + 480 + ], + "spans": [ + { + "bbox": [ + 297, + 468, + 352, + 480 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 489, + 373, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 489, + 373, + 500 + ], + "spans": [ + { + "bbox": [ + 297, + 489, + 373, + 500 + ], + "type": "text", + "content": "3.1. Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 508, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 508, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 508, + 530, + 730 + ], + "type": "text", + "content": "State-of-the-art methods [21, 30, 29, 24, 46, 32] typically assume all InD data follow the same distribution. However, such an assumption will lead to the decision boundary of an OOD detector being doped with some uncertain space. In this paper, we avoid modeling the whole InD dataset and only focus on independently modeling each object category's distribution. We propose hierarchical visual categories modeling to achieve this goal while maintaining high classification accuracy. In hierarchical visual categories modeling, we first project image features into a high-dimensional attribute space (usually 8192 dimensions). These attributes can be grouped into multiple sub-visual concepts as components of an image category which can be easily modeled by Gaussian distributions. Then combinations of multiple sub-visual concepts (abbreviated as sub-concepts) can be grouped to describe a more complex visual concept of an in-distribution category. Since we define visual categories based on sub-visual concepts, simply modeling distributions of these sub-concepts and de" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23427" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 278, + 129 + ], + "type": "text", + "content": "scribing visual categories with these sub-concepts hierarchically can model complicated training distributions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "spans": [ + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": "Formally, given a visual recognition model " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": ", it maps an input image " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " to a high-dimensional feature vector " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^q" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": ". As described above, we project " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " into an attribute space " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " with a higher dimension and get an attribute description " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": ". Attributes in " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " can be grouped into multiple attribute subgroups " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\{a_i\\}_{i=1}^G" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "a_i \\in S_i \\subset \\mathbb{R}^{d/G}" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,G\\}" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": ". For images from a visual category " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": ", we assume that attributes in their " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": "-th attribute group follow a simple Gaussian distribution " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_i^c, \\Sigma_i^c)" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\mu_i^c" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\Sigma_i^c" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " are the mean and variance respectively. Since attributes are divided into " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " groups, we will have " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " different Gaussian distributions for each visual category. The benefits of dividing the whole attribute space into several subspaces come from two folds: (1) modeling attribute distributions in these groups becomes easier; (2) assemble of distributions in these attribute groups can describe complex distributions, which will lead to more accurate decision boundaries. To combine all attributes to describe every visual category, our OOD detector learns weights " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "\\{w_i^c\\}_{i=1}^G" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": " of all attribute groups through exponential moving averages. Therefore, for each in-distribution class " + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 45, + 130, + 281, + 399 + ], + "type": "text", + "content": ", we can model the corresponding visual concept in a probability perspective with a mixture of Gaussian models:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 95, + 405, + 279, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 405, + 279, + 437 + ], + "spans": [ + { + "bbox": [ + 95, + 405, + 279, + 437 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {x}; c\\right) = \\sum_ {i = 1} ^ {G} w _ {i} ^ {c} \\mathcal {N} \\left(\\boldsymbol {a} _ {i}; \\boldsymbol {\\mu} _ {i} ^ {c}, \\boldsymbol {\\Sigma} _ {i} ^ {c}\\right), \\tag {1}", + "image_path": "0173ee623bba8010d90548356b6321783792cf7901d167bda0a1a30c8f21927e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "spans": [ + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "inline_equation", + "content": "w_{i}^{c}\\in \\mathbb{R},\\pmb{\\mu}_{i}^{c}\\in \\mathbb{R}^{d / G},\\pmb{\\Sigma}_{i}^{a}\\in \\mathbb{R}^{d / G\\times d / G}" + }, + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\cdot)" + }, + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "text", + "content": " means a normal distribution. We build a Gaussian mixture model for each class and get " + }, + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 442, + 279, + 524 + ], + "type": "text", + "content": " different Gaussian mixture models. Then, our proposed HVCM focuses on training deep neural networks to learn image features that follow the above distributions and parameters of these probability models jointly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "spans": [ + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "content": "With the above Gaussian probability models, given a test sample " + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "content": ", we define the score function " + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "inline_equation", + "content": "g(\\pmb{x}^{\\prime};\\pmb{w}^{c},\\pmb{\\mu}^{c},\\pmb{\\Sigma}^{c})" + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "content": " to measure whether it belongs to the " + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "content": "-th visual category using the learned probability density function. Here, " + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{c} = \\{w_{i}^{c}\\}_{i = 1}^{G}" + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{c} = \\{\\pmb{\\mu}_{i}^{c}\\}_{i = 1}^{G}" + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{c} = \\{\\pmb{\\Sigma}_{i}^{c}\\}_{i = 1}^{G}" + }, + { + "bbox": [ + 45, + 525, + 279, + 607 + ], + "type": "text", + "content": ". This score function can be used as a reliable metric to detect OOD samples:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 614, + 279, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 279, + 652 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 279, + 652 + ], + "type": "interline_equation", + "content": "h \\left(\\boldsymbol {x} ^ {\\prime}\\right) = \\left\\{ \\begin{array}{l l} \\operatorname {I n D}, & \\text {i f} \\min _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) \\geq \\gamma , \\\\ \\operatorname {O O D}, & \\text {i f} \\min _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) < \\gamma , \\end{array} \\right. \\tag {2}", + "image_path": "0343508c58ddf83edffc68217c9d1b688b45f1d89cbd0dd7054d46ef9373d8c2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 660, + 279, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 694 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 694 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 660, + 279, + 694 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 45, + 660, + 279, + 694 + ], + "type": "text", + "content": " are thresholds to be determined in subsequent sections. Eq. (2) indicates that we use the minimal InD score among all " + }, + { + "bbox": [ + 45, + 660, + 279, + 694 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 45, + 660, + 279, + 694 + ], + "type": "text", + "content": " categories to make the final decision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "Our framework is illustrated in Figure 1, where there are two steps: (1) jointly learn deep features that fit our probability models and parameters of these models; (2) calculate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 106, + 530, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 141 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 141 + ], + "type": "text", + "content": "the InD score based on a set of Gaussian mixture models as the metric for out-of-distribution detection. In the following subsections, we give more details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 148, + 530, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 148, + 530, + 171 + ], + "spans": [ + { + "bbox": [ + 296, + 148, + 530, + 171 + ], + "type": "text", + "content": "3.2. Joint Visual Representation Learning and Parameter Optimization of Probability Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "spans": [ + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": "To learn visual representations that follow Gaussian mixture models and keep their discriminative ability at the same time, we exploit the knowledge distillation framework DINO [6] to perform the joint learning. As shown in Figure 1, for an image " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": ", we sample ten different views of " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " and send them into teacher and student branches simultaneously to perform self-distillation. During knowledge distillation, we project ResNet50 [19] features in 2048 dimensions into an attribute space with dimension " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "d(= 8192)" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": ". Apart from the learning objective produced by self-distillation, we force the attributes of each class to follow a class-specific Gaussian mixture model. We divide the attributes " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{a} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " groups to learn the Gaussian mixture model parameters. However, in practice, it's too hard to directly learn the mean and variance of Gaussian models. We follow He et. al [20] to learn attribute centers " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}_i^c\\}_{i = 1}^G" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": "-th category (the image label " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": "). A linear projection layer is exploited to predict the weights " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "\\{w_i^c (\\pmb {x})\\}_{i = 1}^G" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " on all " + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 296, + 176, + 531, + 392 + ], + "type": "text", + "content": " attribute groups. Our learning objective can be written as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 530, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 530, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 530, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {K D} + \\alpha \\sum_ {i = 1} ^ {G} \\mathrm {K L} \\left(\\boldsymbol {a} _ {i} \\mid \\mid \\boldsymbol {\\mu} _ {i} ^ {c}\\right) + \\beta \\sum_ {i = 1} ^ {T} w _ {i} ^ {c} (\\boldsymbol {x}) \\mathrm {K L} \\left(\\boldsymbol {\\mu} _ {i} ^ {c} \\mid \\mid \\boldsymbol {a} _ {i}\\right) \\tag {3}", + "image_path": "48238342d8708e78c6650da2841a02b04b1ebe1b1acdcb839b3c0f96856f6603.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "spans": [ + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KD}" + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "text", + "content": " stands for the cross entropy loss in self-distillation, KL stands for the Kullback-Leibler(KL) divergence, and " + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "text", + "content": " are hyperparameters. The reason why we exploit two KL divergences is that experiments indicate that the term " + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\pmb{a}_i||\\pmb{\\mu}_i^c)" + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "text", + "content": " will favor learning attribute centers and the term " + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\pmb{\\mu}_i^c||\\pmb{a}_i)" + }, + { + "bbox": [ + 296, + 444, + 530, + 549 + ], + "type": "text", + "content": " is more suitable to learn image attribute descriptions and group weights. Note that softmax operations normalize attributes and their learnable centers in each group before calculating learning objectives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "spans": [ + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "text", + "content": "With the learning objective in Eq. (3), student network parameters " + }, + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "text", + "content": ", and weights " + }, + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "inline_equation", + "content": "\\{w_i^c\\}_{i = 1}^G" + }, + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "text", + "content": " and attribute centers " + }, + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\mu}_i^c\\}_{i = 1}^G" + }, + { + "bbox": [ + 296, + 549, + 530, + 597 + ], + "type": "text", + "content": " of all groups are learnt at the same time. They are updated using the following equations:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 605, + 530, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 605, + 530, + 629 + ], + "spans": [ + { + "bbox": [ + 351, + 605, + 530, + 629 + ], + "type": "interline_equation", + "content": "\\theta_ {s} ^ {t + 1} = \\theta_ {s} ^ {t} - \\gamma_ {1} \\frac {\\partial \\mathcal {L}}{\\partial \\theta_ {s} ^ {t}}, \\tag {4}", + "image_path": "2d97c42448c75fe3a78930849a97396a313d91b0040cb385016b3f20abe28d0f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 343, + 631, + 530, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 631, + 530, + 657 + ], + "spans": [ + { + "bbox": [ + 343, + 631, + 530, + 657 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {i} ^ {c, t + 1} = \\boldsymbol {\\mu} _ {i} ^ {c, t} - \\gamma_ {2} \\frac {\\partial \\mathcal {L}}{\\partial \\boldsymbol {\\mu} _ {i} ^ {c , t}}, \\tag {5}", + "image_path": "13f9e3dd041aae12e1e4dec326fe498abc98ff044d4035349f2e05d2d7d8084a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 659, + 530, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 659, + 530, + 673 + ], + "spans": [ + { + "bbox": [ + 342, + 659, + 530, + 673 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {i} ^ {c, t + 1} = (1 - \\gamma_ {3}) \\boldsymbol {w} _ {i} ^ {c, t} + \\gamma_ {3} w _ {i} ^ {c} (\\boldsymbol {x}), \\tag {6}", + "image_path": "b2364252368c614250ea958989685a78d45ff91373c77e235908e38b7f9fbd88.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2}" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\gamma_{3}" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " are the learning rates. " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2}" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " control the gradient update speed, while " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\gamma_{3}" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " controls the updating speed of the group weights. Following He et. al [20], " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\theta_s,\\{\\pmb {w}^c\\}_{c = 1}^C" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\mu}^c\\}_{c = 1}^C" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " are initialized with Gaussian noises. We use" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23428" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 104, + 278, + 284 + ], + "blocks": [ + { + "bbox": [ + 48, + 104, + 278, + 284 + ], + "lines": [ + { + "bbox": [ + 48, + 104, + 278, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 104, + 278, + 284 + ], + "type": "image", + "image_path": "86bd137e38afbec0ea0d638fb82abf313231d1da57021098cda98e3b8d685939.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 287, + 280, + 351 + ], + "lines": [ + { + "bbox": [ + 45, + 287, + 280, + 351 + ], + "spans": [ + { + "bbox": [ + 45, + 287, + 280, + 351 + ], + "type": "text", + "content": "Figure 2. Illustration of attribute group visualization and Mahalanobis distance distribution. Both attribute groups are visualized by t-SNE [49]. The colors encode different in-distribution data(ImageNet), and out-of-distribution features(SUN) marked as gray points. Models are trained on ResNet-50 [18] using DINO(a) and HVCM(b)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 371, + 279, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 371, + 279, + 407 + ], + "spans": [ + { + "bbox": [ + 45, + 371, + 279, + 407 + ], + "type": "text", + "content": "Adam optimizer [25] with momentum to update both " + }, + { + "bbox": [ + 45, + 371, + 279, + 407 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 45, + 371, + 279, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 371, + 279, + 407 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}^c\\}_{c=1}^C" + }, + { + "bbox": [ + 45, + 371, + 279, + 407 + ], + "type": "text", + "content": ". While the attribute weights are learned through exponential moving averages [17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 415, + 279, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 415, + 279, + 441 + ], + "spans": [ + { + "bbox": [ + 45, + 415, + 279, + 441 + ], + "type": "text", + "content": "3.3. Out-of-Distribution Detection based on Probability Models in Hierarchy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 446, + 279, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 446, + 279, + 575 + ], + "spans": [ + { + "bbox": [ + 45, + 446, + 279, + 575 + ], + "type": "text", + "content": "When the training of our hierarchical probability model converges, we will get visual attributes that follow learned Gaussian distributions hierarchically for in-distribution samples. Meanwhile, we also get weights and centers of all attribute groups for each class. However, the mean attributes are updated frequently during the training and are thus not suitable to be used as group centers. So we need to recalculate the attribute centers for each visual category. Given an in-distribution visual category " + }, + { + "bbox": [ + 45, + 446, + 279, + 575 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 45, + 446, + 279, + 575 + ], + "type": "text", + "content": ", we estimate the mean vector and covariation matrix of the " + }, + { + "bbox": [ + 45, + 446, + 279, + 575 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 45, + 446, + 279, + 575 + ], + "type": "text", + "content": "-th attribute group:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 585, + 279, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 585, + 279, + 615 + ], + "spans": [ + { + "bbox": [ + 76, + 585, + 279, + 615 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {i} ^ {c} = \\frac {1}{N _ {c}} \\sum_ {m = 1} ^ {N _ {c}} \\boldsymbol {a} _ {i} ^ {m} \\tag {7}", + "image_path": "43b5f2aece5c6019038d7e37af396b8e227fdecc3fb1a6f24c10e4f9b5a8ad12.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 618, + 279, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 618, + 279, + 650 + ], + "spans": [ + { + "bbox": [ + 76, + 618, + 279, + 650 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {i} ^ {c} = \\frac {1}{N _ {c} - 1} \\sum_ {m = 1} ^ {N _ {c}} \\left(\\boldsymbol {a} _ {i} ^ {m} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) \\left(\\boldsymbol {a} _ {i} ^ {m} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) ^ {\\top}, \\tag {8}", + "image_path": "02b6a40f30b31c6c5b3b450b68bc97b352210a640f6869601ddf148ba5f8673a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " notes the number of samples in the " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "-th class, and " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\pmb{a}_i^m" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " is the sub-attribute vector of the " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "-th sample. With these weights, means, and covariances, we can describe each visual category in the hierarchy accurately. We try to use the probability density function in Eq. (1) as the in-distribution function. However, we will encounter the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "text", + "content": "problem of numerical overflow when calculating the determinants of covariance matrices as [48]. Instead, we compute the Mahalanobis distance between the " + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "text", + "content": "-th sub-visual attributes " + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "inline_equation", + "content": "a_{i}^{\\prime}" + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "text", + "content": " of a test sample " + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "text", + "content": " and the corresponding attribute center " + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "inline_equation", + "content": "\\mu_{i}^{c}" + }, + { + "bbox": [ + 296, + 106, + 530, + 176 + ], + "type": "text", + "content": " to measure the likelihood of these attributes belongs to the target category:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 327, + 183, + 530, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 183, + 530, + 198 + ], + "spans": [ + { + "bbox": [ + 327, + 183, + 530, + 198 + ], + "type": "interline_equation", + "content": "M _ {i} ^ {c} (\\boldsymbol {x}) = - \\left(\\boldsymbol {a} _ {i} ^ {\\prime} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right) ^ {\\top} \\left(\\boldsymbol {\\Sigma} _ {i} ^ {c}\\right) ^ {- 1} \\left(\\boldsymbol {a} _ {i} ^ {\\prime} - \\boldsymbol {\\mu} _ {i} ^ {c}\\right). \\tag {9}", + "image_path": "496d87d6facb2c7e9334347cbc89e1aafbe7c9da4f37e0a24833020ce1b1989c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 206, + 530, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 206, + 530, + 287 + ], + "spans": [ + { + "bbox": [ + 296, + 206, + 530, + 287 + ], + "type": "text", + "content": "The above equation gives the in-distribution measure for one attribute group. While, for every category, we have multiple attribute groups and need to assemble related in-distribution measures to get the class-level in-distribution score. Since we have gotten the weights of attribute groups for each category, we can easily assemble them and get the class-level score function:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 294, + 530, + 325 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 294, + 530, + 325 + ], + "spans": [ + { + "bbox": [ + 340, + 294, + 530, + 325 + ], + "type": "interline_equation", + "content": "g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right) = \\sum_ {i = 1} ^ {G} w _ {i} ^ {c} M _ {i} ^ {c} \\left(\\boldsymbol {x} ^ {\\prime}\\right). \\tag {10}", + "image_path": "fa453d164aa91ef4a18de8addccc71ca850ab95d580028e33f2631def371583e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 334, + 530, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 334, + 530, + 404 + ], + "spans": [ + { + "bbox": [ + 296, + 334, + 530, + 404 + ], + "type": "text", + "content": "With this score function, we can easily get the in-distribution score of a test sample on each visual category. Since there are " + }, + { + "bbox": [ + 296, + 334, + 530, + 404 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 296, + 334, + 530, + 404 + ], + "type": "text", + "content": " categories in the whole in-distribution dataset, we get the maximal in-distribution score across different visual categories as the in-distribution score on the whole dataset:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 412, + 530, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 412, + 530, + 429 + ], + "spans": [ + { + "bbox": [ + 348, + 412, + 530, + 429 + ], + "type": "interline_equation", + "content": "g \\left(\\boldsymbol {x} ^ {\\prime}\\right) = \\max _ {c} g \\left(\\boldsymbol {x} ^ {\\prime}; \\boldsymbol {w} ^ {c}, \\boldsymbol {\\mu} ^ {c}, \\boldsymbol {\\Sigma} ^ {c}\\right). \\tag {11}", + "image_path": "fbe5f8167294c49fd75679e35c5c5f976a7ff78bd0c5ab3b98c29d50835db4f8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 437, + 530, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 437, + 530, + 531 + ], + "spans": [ + { + "bbox": [ + 296, + 437, + 530, + 531 + ], + "type": "text", + "content": "A high in-distribution score " + }, + { + "bbox": [ + 296, + 437, + 530, + 531 + ], + "type": "inline_equation", + "content": "g(\\boldsymbol{x}^{\\prime})" + }, + { + "bbox": [ + 296, + 437, + 530, + 531 + ], + "type": "text", + "content": " indicates that the semantic attributes of a test sample lie very close to one or multiple in-distribution visual categories (as shown in Figure 2). On the contrary, if a sample does not belong to the previously modeled categories, it will get a low in-distribution score. We follow Eq. (2) to set thresholds to judge whether a sample is an out-of-distribution sample. In the experiment section, we discuss how to set these thresholds." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 541, + 377, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 541, + 377, + 554 + ], + "spans": [ + { + "bbox": [ + 297, + 541, + 377, + 554 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 560, + 411, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 560, + 411, + 573 + ], + "spans": [ + { + "bbox": [ + 297, + 560, + 411, + 573 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 578, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 578, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 578, + 530, + 730 + ], + "type": "text", + "content": "In-distribution Datasets. We use ImageNet-1K [40] and CIFAR10 [27] as our in-distribution datasets. ImageNet-1K is a large-scale visual recognition dataset containing 1000 object categories and 1281167 images. We split it into 10 folds randomly and ensured each fold contains 100 object categories. Since our computation resources are limited, we randomly select one fold as the in-distribution dataset. There other nine folds are used as OOD datasets as other popular benchmarks to test the performance of the proposed method in detecting near OOD samples. For CIFAR10 [27], there are 60000 color images in 10 classes. We conduct OOD algorithm evaluation as previous methods [30, 9, 46, 32, 45, 54]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23429" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 46, + 138, + 531, + 263 + ], + "blocks": [ + { + "bbox": [ + 45, + 102, + 531, + 137 + ], + "lines": [ + { + "bbox": [ + 45, + 102, + 531, + 137 + ], + "spans": [ + { + "bbox": [ + 45, + 102, + 531, + 137 + ], + "type": "text", + "content": "Table 1. OOD detection performance comparison of HVCM and existing methods. All comparison methods rely on ResNet-50 trained with cross-entropy loss. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 46, + 138, + 531, + 263 + ], + "lines": [ + { + "bbox": [ + 46, + 138, + 531, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 531, + 263 + ], + "type": "table", + "html": "
MethodOOD DatasetsAverageInD Acc
iNaturalistSUNPlacesTexturesFPR95↓AUROC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]68.1287.4858.5989.7659.5389.1872.6682.7164.7387.28
ODIN [30]55.7885.9260.4783.8360.9488.0664.0681.2860.3184.77
Maha [29]97.0055.1298.8051.9397.2050.5220.0094.9978.2563.14
Energy [32]58.9186.4527.0393.4438.7591.5156.2587.2645.2489.6785.74
GODIN* [23]72.0079.8660.0984.5864.9482.3039.5089.2859.1384.01
MOS* [24]52.9491.4167.7886.8271.3184.3873.6580.5666.4285.79
ReAct [45]58.4882.6078.1869.1186.3359.8450.5387.0368.3874.65
HVCM(Ours)21.5692.1917.2094.4419.9893.6229.2290.6821.9992.7388.28
", + "image_path": "28aa2845c2fe9ccd44992735eaa867eec281065b3c312e2e7391c7ab1a8e8b26.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 325, + 280, + 427 + ], + "blocks": [ + { + "bbox": [ + 45, + 278, + 280, + 324 + ], + "lines": [ + { + "bbox": [ + 45, + 278, + 280, + 324 + ], + "spans": [ + { + "bbox": [ + 45, + 278, + 280, + 324 + ], + "type": "text", + "content": "Table 2. Evaluation on more challenging detection tasks. * indicates that the method is fine-tuned on InD data. ↑ indicates larger values are better, and ↓ is the opposite. Bold numbers are superior results. All values are percentages." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 325, + 280, + 427 + ], + "lines": [ + { + "bbox": [ + 47, + 325, + 280, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 280, + 427 + ], + "type": "table", + "html": "
MethodImageNet-OOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑
MSP [21]72.5083.3389.1157.99
ODIN [30]73.4471.0364.0679.88
Maha [29]54.4079.3059.2077.62
Energy* [32]66.5681.0662.3484.61
GODIN* [23]71.5579.8973.5777.27
MOS* [24]87.4064.8766.7881.42
ReAct [45]87.0564.1584.1764.30
HVCM(Ours)42.8686.7228.5890.26
", + "image_path": "995fb7f8c378028a5bda00a00a4145fd1ea9489a121af12937fd915d0aef56d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 447, + 279, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 447, + 279, + 634 + ], + "spans": [ + { + "bbox": [ + 45, + 447, + 279, + 634 + ], + "type": "text", + "content": "Out-of-distribution Dataset. On ImageNet, we follow Huang et al. [24] to test our methods and use Texture [8], iNaturalist [50], Places365 [61], and SUN [57]) as OOD test sets. To further explore the limitation of our approach, we evaluate our method on another two OOD datasets, OpenImage-O [26] and ImageNet-O [22]. For CIFAR10, as in [46, 32, 45, 54], we selected eight widely-used datasets, including Texture [8], SVHN [34], Places365 [61], iSUN [59], LSUN-Crop [60], LSUN-Resize [60], ImageNet-Resize [40], and ImageNet-Fix [40] as our test sets. Besides, to test the ability of HVCM to identify near OOD datasets, we use the remaining nine ImageNet subsets as the OOD test sets. Note that since our evaluation on ImageNet differs from previous methods [24, 45], we implement these algorithms with open source provided by authors and follow standard experimental settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 635, + 279, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 635, + 279, + 716 + ], + "spans": [ + { + "bbox": [ + 45, + 635, + 279, + 716 + ], + "type": "text", + "content": "Evaluation Metrics. We employ the commonly used metrics in OOD detection [24, 45] to evaluate our approach, including AUROC, FPR95 and InD Acc. AUROC stands for the area under the receiver operating characteristic curve, FPR95 is short for TPR@FPR95 and represents the false positive rate when the true positive rate is " + }, + { + "bbox": [ + 45, + 635, + 279, + 716 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 45, + 635, + 279, + 716 + ], + "type": "text", + "content": ", and InD Acc is the classification accuracy of in-distribution data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 717, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 717, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 717, + 279, + 730 + ], + "type": "text", + "content": "Training Details. We utilize ResNet-50 [19] as the feature" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "spans": [ + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": "backbone for ImageNet and the dimension of the attribute space is set to 8192. The training is finished in 300 epochs. On CIFAR10, we use ResNet-18 [19] as our feature backbone, and the dimension of the attribute space is set to 1024. The training on CIFAR10 is finished in 200 epochs. The number of attribute groups is set to 32, and " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\gamma_1" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\gamma_2" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\gamma_3" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": " are set to 1, 0.1, 1, 1 and " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": ", respectively. Updating " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": " too fast can lead to the oscillation of group centers, negatively influencing the precision of probabilistic modeling. So we utilize a smaller " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": " to update " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": ". All the hyperparameters are tuned according to the experimental results. We employ SGD with a momentum of 0.9, an initial learning rate of 0.1, and a batch size of 128. The learning rate is reduced by a factor of 10 at " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 296, + 282, + 530, + 491 + ], + "type": "text", + "content": " of the total training epochs. We train all backbones from scratch using random initialization. All experiments are performed using PyTorch [37] with default parameters on four NVIDIA GeForce RTX 3090." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 500, + 528, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 500, + 528, + 514 + ], + "spans": [ + { + "bbox": [ + 296, + 500, + 528, + 514 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Algorithms" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "content": "Standard evaluation on ImageNet. We compare our HVCM with seven popular OOD detection methods, including MSP [21], ODIN [30], GODIN [23], Maha [29], Energy [32], MOS [24], and ReAct [45]. For datasets that describe objects or scenes, such as SUN, Places, and iNaturalist, HVCM achieves better AUROC and FPR95 metrics. When we summarize the results of all four datasets, HVCM achieves " + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "inline_equation", + "content": "21.99\\%" + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "content": " on FPR95 and " + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "inline_equation", + "content": "92.73\\%" + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "content": " on AUROC, which outperforms the previous best method Energy [32] by " + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "inline_equation", + "content": "23.25\\%" + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "inline_equation", + "content": "3.06\\%" + }, + { + "bbox": [ + 296, + 520, + 530, + 730 + ], + "type": "text", + "content": ". This is a significant improvement, which demonstrates that end-to-end training is very important to get good results. When compared to Maha [29], our proposed method exhibits inferior performance in terms of both FPR95 and AUROC. This observation suggests that our method is less effective in describing textures. We attribute this limitation to the fact that textures often encompass numerous repeated patterns, which differ from the characteristics required for general object" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 301, + 756 + ], + "type": "text", + "content": "23430" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 46, + 149, + 531, + 234 + ], + "blocks": [ + { + "bbox": [ + 45, + 102, + 531, + 147 + ], + "lines": [ + { + "bbox": [ + 45, + 102, + 531, + 147 + ], + "spans": [ + { + "bbox": [ + 45, + 102, + 531, + 147 + ], + "type": "text", + "content": "Table 3. Comparison of OOD detection performance of HVCM and existing methods on CIFAR10 dataset. All compared methods use ResNet-18 trained with cross-entropy loss except our proposed method, which uses HVCMLoss. The performance is evaluated based on AUROC (A) and FPR95 (F). " + }, + { + "bbox": [ + 45, + 102, + 531, + 147 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 45, + 102, + 531, + 147 + ], + "type": "text", + "content": " indicates larger values are better and " + }, + { + "bbox": [ + 45, + 102, + 531, + 147 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 45, + 102, + 531, + 147 + ], + "type": "text", + "content": " indicates the opposite. Bold numbers indicate superior results. All values are expressed in percentages." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 46, + 149, + 531, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 531, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 531, + 234 + ], + "type": "table", + "html": "
MethodOOD DatasetAverage
TextureSVHNPlaces365iSUNLSUN(C)LSUN(R)ImageNet(R)ImageNet(F)
F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑F↓A↑
CELossMSP56.4790.2058.4090.5651.8691.9853.8491.6446.2292.7449.1092.4859.6589.3559.9089.4054.4391.04
ODIN40.3791.9827.8293.2831.8094.2317.5996.6025.2395.3114.7497.2127.1894.6545.2690.6228.7594.24
Gram10.8197.732.5899.3930.1593.371.1199.7613.9197.030.5299.861.6198.3866.9482.0215.9595.94
Gram+pNML7.1898.501.6399.6023.2195.130.8399.809.4298.000.4299.881.2498.7657.9085.5312.7396.90
HVCMLossHVCM(ours)1.8899.311.3299.470.9599.540.6599.710.7799.660.5199.961.8099.204.7798.191.5899.38
", + "image_path": "b59f54d80b5cb03f7b48621c7cf5931310a1c8f0f8bc94af148ffc2c64a58d17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 250, + 280, + 364 + ], + "blocks": [ + { + "bbox": [ + 48, + 250, + 280, + 364 + ], + "lines": [ + { + "bbox": [ + 48, + 250, + 280, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 280, + 364 + ], + "type": "image", + "image_path": "da919b464e0ae709993ce4e367f77c534caab7e229e8b77e6a2242ad410f9db8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 366, + 280, + 387 + ], + "lines": [ + { + "bbox": [ + 45, + 366, + 280, + 387 + ], + "spans": [ + { + "bbox": [ + 45, + 366, + 280, + 387 + ], + "type": "text", + "content": "Figure 3. HVCM performance comparison as the increasing distances between InD and OOD data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 412, + 279, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 412, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 45, + 412, + 279, + 564 + ], + "type": "text", + "content": "recognition. It is important to note that our method only utilizes features from the last layer of the network, while Maha [29] leverages features from both intermediate and deep network layers. However, although our results are worse than Maha [29], we are better than all the other methods. This still indicates that the proposed method is very robust in identifying different types of outliers. Furthermore, we build a cosine classifier with the learned attribute centers for image classification. We got " + }, + { + "bbox": [ + 45, + 412, + 279, + 564 + ], + "type": "inline_equation", + "content": "88.28\\%" + }, + { + "bbox": [ + 45, + 412, + 279, + 564 + ], + "type": "text", + "content": " accuracy which is " + }, + { + "bbox": [ + 45, + 412, + 279, + 564 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 45, + 412, + 279, + 564 + ], + "type": "text", + "content": " higher than our supervised learning baseline. This is solid evidence to show that our proposed method can model the InD data accurately and ensure the learned features keep high discriminative ability simultaneously." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 566, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 566, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 566, + 279, + 730 + ], + "type": "text", + "content": "Standard evaluation on CIFAR10. We performed a more conventional OOD detection task on CIFAR10. This experiment aims twofold: firstly, to demonstrate that our method is not dependent on self-distillation (DINO [6]), and secondly, to validate the effectiveness and robustness of our HVCM. Table 3 compares our method with several classic and top-performing algorithms. All comparison methods use ResNet-18 as the main backbone network and are trained with cross-entropy loss, while our method uses only the loss function in Eq. (3). As shown in Table 3, the proposed method outperforms the previous best methods Gram [42] and pNML [2] on both average FPR95 and AUROC obviously. These results demonstrate the proposed method can perform well even on small datasets. Mean" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 284, + 520, + 330 + ], + "blocks": [ + { + "bbox": [ + 296, + 249, + 531, + 281 + ], + "lines": [ + { + "bbox": [ + 296, + 249, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 296, + 249, + 531, + 281 + ], + "type": "text", + "content": "Table 4. The performance of HVCM with varying numbers of group centers " + }, + { + "bbox": [ + 296, + 249, + 531, + 281 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 296, + 249, + 531, + 281 + ], + "type": "text", + "content": ". Results are averaged across four standard OOD datasets, consistent with the main results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 284, + 520, + 330 + ], + "lines": [ + { + "bbox": [ + 309, + 284, + 520, + 330 + ], + "spans": [ + { + "bbox": [ + 309, + 284, + 520, + 330 + ], + "type": "table", + "html": "
Concepts NumberFPR95↓AUROC↑
G=824.3791.73
G=1623.0792.24
G=3221.9992.73
", + "image_path": "bbb82db78981acddb32227a133d65fb4eed3827d9365e8954d2825463b51a5b2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 370, + 530, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 370, + 530, + 406 + ], + "spans": [ + { + "bbox": [ + 296, + 370, + 530, + 406 + ], + "type": "text", + "content": "while, there is no self-distillation exploited, which indicates our joint representation learning and statistical modeling is independent of self-supervised learning algorithms [6]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 415, + 531, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 415, + 531, + 532 + ], + "spans": [ + { + "bbox": [ + 296, + 415, + 531, + 532 + ], + "type": "text", + "content": "Results on more challenging OOD datasets. To overcome the limitations of current OOD benchmarks [53] and evaluate the robustness of our approach against adversarial attacks, we conducted experiments on two challenging datasets, namely OpenImage-O [26] and ImageNet-O [22]. As shown in Table 2, HVCM achieves the highest AUROC and lowest FPR95 among all methods on the OpenImage-O dataset. Although ImageNet-O contains adversarial examples and is more challenging, HVCM still outperforms other methods on this dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 543, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 543, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 543, + 532, + 730 + ], + "type": "text", + "content": "Results on near-to-far OOD datasets. To investigate the ability of the proposed method to detect near OOD samples, we construct 9 OOD test sets with the remaining ImageNet images. We rank the semantic distance between the remaining 900 visual categories with the 100 categories in the InD dataset. We use the average cosine distance as the measure and construct 9 different OOD test sets. Details are introduced in the supplementary material. For convenience, we denote these datasets from OOD 1 to OOD 9. The experimental results are depicted in Figure 3, and several samples are displayed in Figure 4. We can find that the proposed method achieves good AUROC even when the test set is very close to the InD dataset. When the test sets become farther, FPR95 decreases quickly, which indicates the proposed method is very sensitive to the semantic distances of OOD datasets." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23431" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 105, + 99, + 210 + ], + "blocks": [ + { + "bbox": [ + 49, + 105, + 99, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 105, + 99, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 105, + 99, + 210 + ], + "type": "image", + "image_path": "d1c79ce06faf27ee74bfd7b4a2c70170003acac168ce8aaa49399a37655b8327.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 213, + 531, + 245 + ], + "lines": [ + { + "bbox": [ + 45, + 213, + 531, + 245 + ], + "spans": [ + { + "bbox": [ + 45, + 213, + 531, + 245 + ], + "type": "text", + "content": "Figure 4. Each dataset exhibits displayed two categories of images. The leftmost samples belong to the InD dataset, while the categories on the right correspond to nine OOD datasets arranged in ascending order of distance. It can be observed that the gap between the OOD samples and InD samples gradually widens as the distance increases." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 102, + 105, + 148, + 210 + ], + "blocks": [ + { + "bbox": [ + 102, + 105, + 148, + 210 + ], + "lines": [ + { + "bbox": [ + 102, + 105, + 148, + 210 + ], + "spans": [ + { + "bbox": [ + 102, + 105, + 148, + 210 + ], + "type": "image", + "image_path": "79651569599fe1d7aca39573a1699880ee4a1289dc51241faf73ac364295315a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 151, + 105, + 198, + 210 + ], + "blocks": [ + { + "bbox": [ + 151, + 105, + 198, + 210 + ], + "lines": [ + { + "bbox": [ + 151, + 105, + 198, + 210 + ], + "spans": [ + { + "bbox": [ + 151, + 105, + 198, + 210 + ], + "type": "image", + "image_path": "5dc7a8b3738a0cf3d7598a81fa2fb85c8d552a1283b0872119f7a1b187d4a756.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 199, + 105, + 246, + 210 + ], + "blocks": [ + { + "bbox": [ + 199, + 105, + 246, + 210 + ], + "lines": [ + { + "bbox": [ + 199, + 105, + 246, + 210 + ], + "spans": [ + { + "bbox": [ + 199, + 105, + 246, + 210 + ], + "type": "image", + "image_path": "6f4ce035735d759670a962608b80b58d238afd3785c623f59ae38cb04e8679ba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 246, + 105, + 291, + 210 + ], + "blocks": [ + { + "bbox": [ + 246, + 105, + 291, + 210 + ], + "lines": [ + { + "bbox": [ + 246, + 105, + 291, + 210 + ], + "spans": [ + { + "bbox": [ + 246, + 105, + 291, + 210 + ], + "type": "image", + "image_path": "47ed9d314f95959e1c6602f53c134914d98af91d49c9408e538ad2e1099b0d7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 292, + 105, + 338, + 210 + ], + "blocks": [ + { + "bbox": [ + 292, + 105, + 338, + 210 + ], + "lines": [ + { + "bbox": [ + 292, + 105, + 338, + 210 + ], + "spans": [ + { + "bbox": [ + 292, + 105, + 338, + 210 + ], + "type": "image", + "image_path": "5be6a798eff03983b97e80e9ac01c488c711fb6728032ad59a80b7d363ddc396.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 339, + 105, + 384, + 210 + ], + "blocks": [ + { + "bbox": [ + 339, + 105, + 384, + 210 + ], + "lines": [ + { + "bbox": [ + 339, + 105, + 384, + 210 + ], + "spans": [ + { + "bbox": [ + 339, + 105, + 384, + 210 + ], + "type": "image", + "image_path": "011fe1e45a40ba52df17c45226c7da0cd740d1dddc56cb96df60d1f96e1dfd21.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 384, + 105, + 433, + 210 + ], + "blocks": [ + { + "bbox": [ + 384, + 105, + 433, + 210 + ], + "lines": [ + { + "bbox": [ + 384, + 105, + 433, + 210 + ], + "spans": [ + { + "bbox": [ + 384, + 105, + 433, + 210 + ], + "type": "image", + "image_path": "e5936316381cac47bd3ded9236f78ee5a778445f1e04d1e45e8a33e166ba3461.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 105, + 478, + 210 + ], + "blocks": [ + { + "bbox": [ + 434, + 105, + 478, + 210 + ], + "lines": [ + { + "bbox": [ + 434, + 105, + 478, + 210 + ], + "spans": [ + { + "bbox": [ + 434, + 105, + 478, + 210 + ], + "type": "image", + "image_path": "3afd3b53e5037dde691369b79ebd6a707edb32d3d29bf2ca7b69a24f49946f40.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 479, + 105, + 528, + 210 + ], + "blocks": [ + { + "bbox": [ + 479, + 105, + 528, + 210 + ], + "lines": [ + { + "bbox": [ + 479, + 105, + 528, + 210 + ], + "spans": [ + { + "bbox": [ + 479, + 105, + 528, + 210 + ], + "type": "image", + "image_path": "7bd8ebf552e027c3b5072d428aecb91dd41495a42176e999b66b29157a0cefcb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 58, + 318, + 268, + 392 + ], + "blocks": [ + { + "bbox": [ + 45, + 261, + 279, + 314 + ], + "lines": [ + { + "bbox": [ + 45, + 261, + 279, + 314 + ], + "spans": [ + { + "bbox": [ + 45, + 261, + 279, + 314 + ], + "type": "text", + "content": "Table 5. A set of ablation results about HVCM. The top row investigates the effect of using MSE, KL divergence, or JS divergence on the model performance; the bottle row compares the performance of different OOD detection methods. Results are averaged across four standard OOD datasets following the main results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 318, + 268, + 392 + ], + "lines": [ + { + "bbox": [ + 58, + 318, + 268, + 392 + ], + "spans": [ + { + "bbox": [ + 58, + 318, + 268, + 392 + ], + "type": "table", + "html": "
Strategy AblationFPR95↓AUROC↑
Learning ObjectiveL223.5992.26
KL22.9992.58
JS21.9992.73
InD DistanceCosine66.3784.06
Linear36.7786.59
Maha21.9992.73
", + "image_path": "b0ec4c4ef195ef426b1be0307d06987fe82cb755714e27db1512e9640f2a9eb3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 415, + 136, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 415, + 136, + 428 + ], + "spans": [ + { + "bbox": [ + 45, + 415, + 136, + 428 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "spans": [ + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "text", + "content": "Number of attribute groups. We varied the number of attribute groups from 8 to 32 to analyze the components in Gaussian mixture models. In Table 4, we find a positive correlation between the number of attribute groups and model performance, with the best performance achieved when " + }, + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "text", + "content": " is set to 32. We also try to set " + }, + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "text", + "content": " to bigger numbers and list the results in the supplementary material. However, more attribute groups will lead to bigger correlation matrices when computing the InD score. Thus, we finally set " + }, + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 435, + 279, + 551 + ], + "type": "text", + "content": " to 32 to balance performance and inference speed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 553, + 279, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 553, + 279, + 670 + ], + "spans": [ + { + "bbox": [ + 45, + 553, + 279, + 670 + ], + "type": "text", + "content": "Choice of learning objectives. The first row of Table 5 investigates how the choice of the learning objective influences the performance of HVCM. We test three learning objectives, including the L2 loss, JS divergence loss, and KL divergence loss. The results show that compared with the L2 loss and KL divergence loss, the JS divergence loss achieves the lowest FPR95, demonstrating its superiority for statistical modeling. We attribute this to the fact that the group centers and features need to learn from each other, and the JS divergence loss is symmetrical in enclosing them." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 671, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 729 + ], + "type": "text", + "content": "Different InD distance metrics. In Table 5, we also compared our Maha metric with two different InD distance metrics. The cosine distance metric directly measures the distance by calculating the cosine similarity between the input feature and the mean of Gaussian distribution models. The" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 264, + 530, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 264, + 530, + 322 + ], + "spans": [ + { + "bbox": [ + 296, + 264, + 530, + 322 + ], + "type": "text", + "content": "linear distance metric is used to calculate the distance with trainable linear layers. The results show that our Maha metric is an effective metric compared with its counterparts. We attribute this to the Maha distance space can better fit the training distribution in realistic scenes." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 305, + 340, + 523, + 453 + ], + "blocks": [ + { + "bbox": [ + 305, + 340, + 523, + 453 + ], + "lines": [ + { + "bbox": [ + 305, + 340, + 523, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 523, + 453 + ], + "type": "image", + "image_path": "285be144585e931334d9d0e34d834848d001ef2bcd9292cae5e44e8cff5caf4a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 455, + 530, + 475 + ], + "lines": [ + { + "bbox": [ + 296, + 455, + 530, + 475 + ], + "spans": [ + { + "bbox": [ + 296, + 455, + 530, + 475 + ], + "type": "text", + "content": "Figure 5. The performance of HVCM is evaluated as the number of InD classes increases across four OOD datasets." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 493, + 530, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 493, + 530, + 633 + ], + "spans": [ + { + "bbox": [ + 296, + 493, + 530, + 633 + ], + "type": "text", + "content": "Increasing numbers of InD Categories in ImageNet. We test how the OOD detection performance varies with the increasing of object categories in the in-distribution dataset. Following Wang et al. [53], we test on four popular benchmarks and set " + }, + { + "bbox": [ + 296, + 493, + 530, + 633 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 296, + 493, + 530, + 633 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 493, + 530, + 633 + ], + "type": "inline_equation", + "content": "\\{50, 100, 200, 300\\}" + }, + { + "bbox": [ + 296, + 493, + 530, + 633 + ], + "type": "text", + "content": " respectively. As depicted in Figure 5, the performance of HVCM fluctuates on different datasets with the increasing of the InD object categories, which suggests that the number of categories has little impact on our approach. These results validate our assumption that we only need to model in-distribution image categories, and the out-of-distribution samples can be detected easily then." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "content": "Different thresholds for OOD detection. Figure 6 illustrates the accuracy of OOD detection for our method across various datasets. On most datasets, our method exhibits the same trend for accuracy variation with thresholds. This indicates that our approach has strong generalization and ideal performance to domains with significant differences. We explain the performance of Imagenet-O as its task difficulty with adversarial samples." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23432" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 104, + 262, + 274 + ], + "blocks": [ + { + "bbox": [ + 59, + 104, + 262, + 274 + ], + "lines": [ + { + "bbox": [ + 59, + 104, + 262, + 274 + ], + "spans": [ + { + "bbox": [ + 59, + 104, + 262, + 274 + ], + "type": "image", + "image_path": "baf1b3963df841cae5c7bf8398f96ca7a512ee5e051b2e91981f86624548fa52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 276, + 279, + 298 + ], + "lines": [ + { + "bbox": [ + 45, + 276, + 279, + 298 + ], + "spans": [ + { + "bbox": [ + 45, + 276, + 279, + 298 + ], + "type": "text", + "content": "Figure 6. The accuracy of HVAC varies with different thresholds across all OOD datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 324, + 116, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 116, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 116, + 336 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 348, + 279, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 348, + 279, + 489 + ], + "spans": [ + { + "bbox": [ + 45, + 348, + 279, + 489 + ], + "type": "text", + "content": "In this paper, we introduce a hierarchical visual category modeling scheme for out-of-distribution detection, which combines visual representation learning and parameter optimization of probability models. It provides a novel perspective for OOD detection by conducting representation learning and density estimation end-to-end. By modeling visual categories with mixtures of Gaussian models, we describe visual categories in very complex distribution and don't rely on outlier training data to perform OOD detection. Experiments demonstrate that the proposed method outperforms state-of-the-art algorithms clearly and does not hinder the discriminative ability of deep features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 493, + 279, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 493, + 279, + 575 + ], + "spans": [ + { + "bbox": [ + 45, + 493, + 279, + 575 + ], + "type": "text", + "content": "Limitations. However, our method needs to map deep features into high-dimensional attribute spaces and build plentiful Gaussian mixture models. These Gaussian mixture models bring a lot of computational costs and make the inference of the OOD detector become inefficient. Thus, simplifying the probability models and accelerating the inference process will be the future direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 600, + 147, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 600, + 147, + 614 + ], + "spans": [ + { + "bbox": [ + 45, + 600, + 147, + 614 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (No.62072112), Scientific and Technological innovation action plan of Shanghai Science and Technology Committee (No.22511102202), National Key R&D Program of China (2020AAA0108301), National Natural Science Foundation of China under Grant (No. 62106051), Shanghai Pujiang Program (No. 21PJ1400600), and National Key R&D Program of China (2022YFC3601405)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 298, + 105, + 355, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 105, + 355, + 117 + ], + "spans": [ + { + "bbox": [ + 298, + 105, + 355, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 298, + 124, + 532, + 730 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 302, + 124, + 530, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 124, + 530, + 156 + ], + "spans": [ + { + "bbox": [ + 302, + 124, + 530, + 156 + ], + "type": "text", + "content": "[1] Jinwon An and Sungzoon Cho. Variational autoencoder based anomaly detection using reconstruction probability. Special lecture on IE, 2(1):1-18, 2015." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 158, + 531, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 158, + 531, + 201 + ], + "spans": [ + { + "bbox": [ + 302, + 158, + 531, + 201 + ], + "type": "text", + "content": "[2] Koby Bibas, Meir Feder, and Tal Hassner. Single layer predictive normalized maximum likelihood for out-of-distribution detection. Advances in Neural Information Processing Systems, 34:1179-1191, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 202, + 532, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 202, + 532, + 255 + ], + "spans": [ + { + "bbox": [ + 302, + 202, + 532, + 255 + ], + "type": "text", + "content": "[3] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 256, + 531, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 256, + 531, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 256, + 531, + 321 + ], + "type": "text", + "content": "[4] Daniel Bogdoll, Jasmin Breitenstein, Florian Heidecker, Maarten Bieshaar, Bernhard Sick, Tim Fingscheidt, and Marius Zöllner. Description of corner cases in automated driving: Goals and challenges. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1023-1028, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 322, + 532, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 322, + 532, + 365 + ], + "spans": [ + { + "bbox": [ + 302, + 322, + 532, + 365 + ], + "type": "text", + "content": "[5] Senqi Cao and Zhongfei Zhang. Deep hybrid models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 367, + 532, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 367, + 532, + 420 + ], + "spans": [ + { + "bbox": [ + 302, + 367, + 532, + 420 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 421, + 531, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 421, + 531, + 463 + ], + "spans": [ + { + "bbox": [ + 302, + 421, + 531, + 463 + ], + "type": "text", + "content": "[7] Xingyu Chen, Xuguang Lan, Fuchun Sun, and Nanning Zheng. A boundary based out-of-distribution classifier for generalized zero-shot learning. In European Conference on Computer Vision, pages 572-588. Springer, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 465, + 532, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 465, + 532, + 508 + ], + "spans": [ + { + "bbox": [ + 302, + 465, + 532, + 508 + ], + "type": "text", + "content": "[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 509, + 532, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 509, + 532, + 541 + ], + "spans": [ + { + "bbox": [ + 302, + 509, + 532, + 541 + ], + "type": "text", + "content": "[9] Roy De Maesschalck, Delphine Jouan-Rimbaud, and Désiré L Massart. The mahalanobis distance. Chemometrics and intelligent laboratory systems, 50(1):1-18, 2000." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 298, + 542, + 530, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 542, + 530, + 585 + ], + "spans": [ + { + "bbox": [ + 298, + 542, + 530, + 585 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 298, + 586, + 530, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 586, + 530, + 618 + ], + "spans": [ + { + "bbox": [ + 298, + 586, + 530, + 618 + ], + "type": "text", + "content": "[11] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. arXiv preprint arXiv:1605.08803, 2016." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 298, + 619, + 530, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 619, + 530, + 662 + ], + "spans": [ + { + "bbox": [ + 298, + 619, + 530, + 662 + ], + "type": "text", + "content": "[12] Zhen Fang, Jie Lu, Anjin Liu, Feng Liu, and Guangquan Zhang. Learning bounds for open-set learning. In International Conference on Machine Learning, pages 3122-3132. PMLR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 298, + 663, + 530, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 663, + 530, + 707 + ], + "spans": [ + { + "bbox": [ + 298, + 663, + 530, + 707 + ], + "type": "text", + "content": "[13] Aditya Grover, Manik Dhar, and Stefano Ermon. Flow-gan: Combining maximum likelihood and adversarial learning in generative models. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 298, + 708, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 708, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 298, + 708, + 530, + 730 + ], + "type": "text", + "content": "[14] Xiaoyuan Guo, Judy Wawira Gichoya, Saptarshi Purkayastha, and Imon Banerjee. Cvad: A generic" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23433" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 64, + 106, + 279, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 106, + 279, + 128 + ], + "spans": [ + { + "bbox": [ + 64, + 106, + 279, + 128 + ], + "type": "text", + "content": "medical anomaly detector based on cascade vae. arXiv preprint arXiv:2110.15811, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 279, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 279, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 279, + 182 + ], + "type": "text", + "content": "[15] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9591-9600, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 182, + 279, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 279, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 279, + 225 + ], + "type": "text", + "content": "[16] Matan Haroush, Tzviel Frostig, Ruth Heller, and Daniel Soudry. A statistical framework for efficient out of distribution detection in deep neural networks. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 225, + 279, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 279, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 279, + 277 + ], + "type": "text", + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 278, + 279, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 279, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 279, + 321 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 321, + 279, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 321, + 279, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 279, + 364 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 365, + 279, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 279, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 279, + 428 + ], + "type": "text", + "content": "[20] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 429, + 279, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 279, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 279, + 471 + ], + "type": "text", + "content": "[21] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. Proceedings of International Conference on Learning Representations, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 472, + 279, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 279, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 279, + 503 + ], + "type": "text", + "content": "[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 504, + 279, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 279, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 279, + 557 + ], + "type": "text", + "content": "[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 558, + 279, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 279, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 279, + 600 + ], + "type": "text", + "content": "[24] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 601, + 279, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 279, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 279, + 632 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 633, + 279, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 279, + 697 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 279, + 697 + ], + "type": "text", + "content": "[26] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, et al. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2(3):18, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 698, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 698, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 698, + 279, + 730 + ], + "type": "text", + "content": "[27] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto, 2009." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 730 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 149 + ], + "type": "text", + "content": "[28] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 150, + 530, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 150, + 530, + 213 + ], + "spans": [ + { + "bbox": [ + 299, + 150, + 530, + 213 + ], + "type": "text", + "content": "[29] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 214, + 530, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 214, + 530, + 256 + ], + "spans": [ + { + "bbox": [ + 299, + 214, + 530, + 256 + ], + "type": "text", + "content": "[30] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 257, + 530, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 257, + 530, + 299 + ], + "spans": [ + { + "bbox": [ + 299, + 257, + 530, + 299 + ], + "type": "text", + "content": "[31] Qiao Liu, Jiaze Xu, Rui Jiang, and Wing Hung Wong. Density estimation using deep generative neural networks. Proceedings of the National Academy of Sciences, 118(15):e2101344118, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 300, + 530, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 300, + 530, + 353 + ], + "spans": [ + { + "bbox": [ + 299, + 300, + 530, + 353 + ], + "type": "text", + "content": "[32] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21464-21475. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 354, + 530, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 354, + 530, + 406 + ], + "spans": [ + { + "bbox": [ + 299, + 354, + 530, + 406 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 407, + 530, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 407, + 530, + 460 + ], + "spans": [ + { + "bbox": [ + 299, + 407, + 530, + 460 + ], + "type": "text", + "content": "[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 461, + 530, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 461, + 530, + 503 + ], + "spans": [ + { + "bbox": [ + 299, + 461, + 530, + 503 + ], + "type": "text", + "content": "[35] Philipp Oberdiek, Gernot A Fink, and Matthias Rottmann. Uqgan: A unified model for uncertainty quantification of deep classifiers trained via conditional gans. arXiv preprint arXiv:2201.13279, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 504, + 530, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 504, + 530, + 536 + ], + "spans": [ + { + "bbox": [ + 299, + 504, + 530, + 536 + ], + "type": "text", + "content": "[36] George Papamakarios, Theo Pavlakou, and Iain Murray. Masked autoregressive flow for density estimation. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 537, + 530, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 537, + 530, + 589 + ], + "spans": [ + { + "bbox": [ + 299, + 537, + 530, + 589 + ], + "type": "text", + "content": "[37] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 590, + 530, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 590, + 530, + 632 + ], + "spans": [ + { + "bbox": [ + 299, + 590, + 530, + 632 + ], + "type": "text", + "content": "[38] Chen Qiu, Aodong Li, Marius Kloft, Maja Rudolph, and Stephan Mandt. Latent outlier exposure for anomaly detection with contaminated data. arXiv preprint arXiv:2202.08088, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 633, + 530, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 633, + 530, + 654 + ], + "spans": [ + { + "bbox": [ + 299, + 633, + 530, + 654 + ], + "type": "text", + "content": "[39] Carl Rasmussen. The infinite gaussian mixture model. Advances in neural information processing systems, 12, 1999." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 655, + 530, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 655, + 530, + 708 + ], + "spans": [ + { + "bbox": [ + 299, + 655, + 530, + 708 + ], + "type": "text", + "content": "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "type": "text", + "content": "[41] Murat Sensoy, Lance Kaplan, Federico Cerutti, and Maryam Saleki. Uncertainty-aware deep classifiers using generative" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23434" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 106, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 65, + 106, + 279, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 106, + 279, + 128 + ], + "spans": [ + { + "bbox": [ + 65, + 106, + 279, + 128 + ], + "type": "text", + "content": "models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5620-5627, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 279, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 279, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 279, + 161 + ], + "type": "text", + "content": "[42] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with in-distribution examples and gram matrices. arXiv e-prints, pages arXiv-1912, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 279, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 279, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 279, + 204 + ], + "type": "text", + "content": "[43] Nina Shvetsova, Bart Bakker, Irina Fedulova, Heinrich Schulz, and Dmitry V. Dylov. Anomaly detection in medical imaging with deep perceptual autoencoders. IEEE Access, 9:118571-118583, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 207, + 279, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 279, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 279, + 238 + ], + "type": "text", + "content": "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 279, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 279, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 279, + 271 + ], + "type": "text", + "content": "[45] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 279, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 279, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 279, + 303 + ], + "type": "text", + "content": "[46] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. ICML, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 279, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 279, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 279, + 337 + ], + "type": "text", + "content": "[47] Esteban G Tabak and Cristina V Turner. A family of nonparametric density estimation algorithms. Communications on Pure and Applied Mathematics, 66(2):145-164, 2013." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 339, + 279, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 339, + 279, + 379 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 279, + 379 + ], + "type": "text", + "content": "[48] Shengjin Tang, Chuanqiang Yu, Xiaoyan Sun, Hongdong Fan, and Xiaosheng Si. A note on parameters estimation for nonlinear wiener processes with measurement errors. IEEE Access, 7:176756-176766, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 279, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 279, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 279, + 413 + ], + "type": "text", + "content": "[49] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 279, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 279, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 279, + 478 + ], + "type": "text", + "content": "[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 279, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 279, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 279, + 522 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 524, + 279, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 279, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 279, + 576 + ], + "type": "text", + "content": "[52] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), pages 550-564, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 279, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 279, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 279, + 619 + ], + "type": "text", + "content": "[53] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 622, + 279, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 279, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 279, + 664 + ], + "type": "text", + "content": "[54] Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. Mitigating neural network overconfidence with logit normalization. In International Conference on Machine Learning, pages 23631-23644. PMLR, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 666, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 666, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 48, + 666, + 279, + 729 + ], + "type": "text", + "content": "[55] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y. Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Nicholas Petrick and Kensaku Mori, editors, Medical Imaging 2018: Computer-Aided Diagnosis, volume 10575, pages 375 - 380. International Society for Optics and Photonics, SPIE, 2018." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 452 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 181 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 181 + ], + "type": "text", + "content": "[56] Jim Winkens, Rudy Bunel, Abhijit Guha Roy, Robert Stanforth, Vivek Natarajan, Joseph R. Ledsam, Patricia MacWilliams, Pushmeet Kohli, Alan Karthikesalingam, Simon Kohl, Taylan Cemgil, S. M. Ali Eslami, and Olaf Ronneberger. Contrastive training for improved out-of-distribution detection. arXiv preprint arXiv:2007.05566, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 183, + 530, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 183, + 530, + 235 + ], + "spans": [ + { + "bbox": [ + 299, + 183, + 530, + 235 + ], + "type": "text", + "content": "[57] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 237, + 530, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 237, + 530, + 278 + ], + "spans": [ + { + "bbox": [ + 299, + 237, + 530, + 278 + ], + "type": "text", + "content": "[58] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. Advances in neural information processing systems, 33:20685-20696, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 281, + 530, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 281, + 530, + 322 + ], + "spans": [ + { + "bbox": [ + 299, + 281, + 530, + 322 + ], + "type": "text", + "content": "[59] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 324, + 530, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 324, + 530, + 367 + ], + "spans": [ + { + "bbox": [ + 299, + 324, + 530, + 367 + ], + "type": "text", + "content": "[60] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 368, + 530, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 368, + 530, + 410 + ], + "spans": [ + { + "bbox": [ + 299, + 368, + 530, + 410 + ], + "type": "text", + "content": "[61] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 411, + 530, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 411, + 530, + 452 + ], + "spans": [ + { + "bbox": [ + 299, + 411, + 530, + 452 + ], + "type": "text", + "content": "[62] Cangqi Zhou, Hao Ban, Jing Zhang, Qianmu Li, and Yinghua Zhang. Gaussian mixture variational autoencoder for semi-supervised topic modeling. IEEE Access, 8:106843-106854, 2020." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 276, + 747, + 300, + 756 + ], + "type": "text", + "content": "23435" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_content_list.json b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1933722d13299ca1122d4e242876e79b47131e54 --- /dev/null +++ b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_content_list.json @@ -0,0 +1,1992 @@ +[ + { + "type": "text", + "text": "Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning", + "text_level": 1, + "bbox": [ + 96, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanjae Kim", + "bbox": [ + 168, + 179, + 274, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiyoung Lee²", + "bbox": [ + 310, + 180, + 419, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seongheon Park", + "bbox": [ + 454, + 180, + 591, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kwanghoon Sohn $^{1,3,*}$", + "bbox": [ + 629, + 180, + 800, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Yonsei University", + "bbox": [ + 98, + 204, + 250, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ NAVER AI Lab", + "bbox": [ + 289, + 204, + 424, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Korea Institute", + "bbox": [ + 465, + 204, + 593, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of Science and Technology (KIST)", + "bbox": [ + 599, + 205, + 870, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{incohjk,sam121796,khsohn}@yonsei.ac.kr", + "bbox": [ + 196, + 229, + 542, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lee.j@navercorp.com", + "bbox": [ + 601, + 229, + 772, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 294, + 312, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compositional zero-shot learning (CZSL) aims to recognize unseen compositions with prior knowledge of known primitives (attribute and object). Previous works for CZSL often suffer from grasping the contextuality between attribute and object, as well as the discriminability of visual features, and the long-tailed distribution of real-world compositional data. We propose a simple and scalable framework called Composition Transformer (CoT) to address these issues. CoT employs object and attribute experts in distinctive manners to generate representative embeddings, using the visual network hierarchically. The object expert extracts representative object embeddings from the final layer in a bottom-up manner, while the attribute expert makes attribute embeddings in a top-down manner with a proposed object-guided attention module that models contextuality explicitly. To remedy biased prediction caused by imbalanced data distribution, we develop a simple minority attribute augmentation (MAA) that synthesizes virtual samples by mixing two images and oversampling minority attribute classes. Our method achieves SoTA performance on several benchmarks, including MIT-States, C-GQA, and VAW-CZSL. We also demonstrate the effectiveness of CoT in improving visual discrimination and addressing the model bias from the imbalanced data distribution. The code is available at https://github.com/HanjaeKim98/CoT.", + "bbox": [ + 75, + 327, + 473, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans perceive entities as hierarchies of parts; for example, we recognize 'Cute Cat' by composing the meaning of 'Cute' and 'Cat'. People can even perceive new concepts by composing the primitive meanings they already", + "bbox": [ + 75, + 765, + 468, + 825 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8cfc7ef69ae570faf24074f64d6c08a8f23aaebbb84dbebffd51e2968bc66485.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 280, + 705, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d75d6cbdc356857837d3947d57f3b52a8cff261348df50fa120d3b8160139a50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 294, + 888, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74bbbd01f18c8ba45fc28da780044b8c0f47d3cd8554f3b496409ee554e597d5.jpg", + "image_caption": [ + "(a)OADis" + ], + "image_footnote": [], + "bbox": [ + 503, + 393, + 705, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab0c3570a37ec99c73300541daa30a0c4021a884fb4d4755db8bdd6fd3f0c9eb.jpg", + "image_caption": [ + "(b) CoT (Ours)", + "Figure 1: Visual feature distribution of previous state-of-the-art approach (OADis) [53] and CoT (Ours). The features are gathered with respect to attribute 'Blue -' (Top) and object ' - Wall' (Bottom). Notably, the previous method suffers from degraded visual discrimination such as ('Blue, Jeans' vs. 'Blue, Shirt'), or all attributes composed into object 'Wall'." + ], + "image_footnote": [], + "bbox": [ + 707, + 407, + 888, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "knew. Such compositionality is a fundamental ability of humans for all cognition. For this reason, compositional zero-shot learning (CZSL), recognizing a novel composition of known components (i.e., object and attribute), has been regarded as a crucial problem in the research community.", + "bbox": [ + 496, + 612, + 893, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A naive approach for CZSL is to combine attribute and object scores estimated from individually trained classifiers [39, 21]. However, it is hard for these independent predictions to consider the interactions that result from a combination of different primitives, which is contextuality (e.g. different meaning of 'Old' in 'Old Car' and 'Old Cat') in the composition [39]. Previous approaches [42, 76, 36, 41, 53, 80] have solved the problem by modeling the compositional label (text) embedding for each class. They leverage external knowledge bases like Glove [37, 47] to extract attribute and object semantic vectors, and concatenate them with a few layers into the composition label embedding. The embeddings are then aligned with visual features in the shared embedding space, where the recognition of unseen image", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 839, + 222, + 851 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was supported by the Institute of Information Communications Technology Planning Evaluation (IITP) grant funded by the Korean government (MSIT) (No. 2021-0-02068, Artificial Intelligence Innovation Hub)", + "bbox": [ + 76, + 851, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5675", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "query becomes the nearest neighbor search problem [36]. Nevertheless, the extent to which contextuality is taken into account in the visual domain has still been restricted.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works [52, 53] have also pointed out the importance of discriminability in visual features, which is related to generalization performance towards unseen composition recognition [62, 38]. A popular solution is disentangle-ment [52, 53, 27, 80], in which independent layers are allocated to extract intermediate visual representations of attribute and object. However, as shown in Fig. 1, it may be challenging to extract their unique characteristics to understand the heterogeneous interaction. We carefully argue that this problem arises from how the visual backbone (e.g. ResNet18 [16]) is used. This is because attribute and object features, which require completely different characteristics [30, 22], are based on the same visual representation from the deep layer of the backbone.", + "bbox": [ + 75, + 138, + 468, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another challenge in CZSL is the long-tailed distribution of real-world compositional data [2, 63, 3]. Few attribute classes are dominantly associated with objects, which may cause a hubness problem [10, 12] among visual features. The visual features from the head (frequent) composition become a hub, which aggravates the visual features to be indistinguishable in the embedding space [12], and induces a biased prediction towards dominant composition [61, 81].", + "bbox": [ + 75, + 352, + 468, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, based on the discussions above, we propose Composition Transformer (CoT) to enlarge the visual discrimination for robust CZSL. Motivated by bottom-up and top-down attention [31, 32, 28], the CoT presents object and attribute experts, each forming its feature in different layers of the visual network. Specifically, the object expert generates a representative object embedding from the last layer (i.e., bottom-up pathway) that is most robust to identify object category with high-level semantics [72, 44]. Then we explicitly model contextuality through an object-guided attention module that explores intermediate layers of the backbone network and builds attribute-related features associated with the object (i.e., top-down pathway). Based on this module, the attribute expert generates a distinctive attribute embedding in conjunction with the object embedding. By utilizing all the features of each layer exhibiting different characteristics, our method comprehensively leverages a visual network to diversify the components' features in the shared embedding space.", + "bbox": [ + 75, + 474, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we further develop a simple minority attribute augmentation (MAA) methodology tailored for CZSL to address the biased prediction caused by imbalanced data distribution [61]. Unlike GAN-based augmentations [56, 27] that lead to overwhelming computation during training, our method simplifies the synthesis process of the virtual sample by blending two images while oversampling minority attribute classes [7, 45]. Thanks to the label smoothing effect of the balanced data distribution by augmenta", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion [79, 59], the CoT is well generalized with minimal computational costs, resolving the bias problem induced by the majority classes.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized in three-folds:", + "bbox": [ + 517, + 137, + 844, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To enhance visual discrimination and contextuality, we propose a Composition Transformer (CoT). In this framework, object and attribute experts hierarchically estimate primitive embeddings by fully utilizing intermediate outputs of the visual network.", + "- We introduce a simple yet robust MAA that alleviates dominant prediction on head compositions with majority attributes.", + "- The remarkable experimental results demonstrate that our CoT and MAA are harmonized to improve the performance on several CZSL benchmarks, showing state-of-the-art performance." + ], + "bbox": [ + 517, + 155, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 361, + 640, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a set of objects and their associated attributes, compositional zero-shot learning (CZSL) aims to recognize unseen compositions using primitive knowledge. Contrary to conventional methods [34] that recognize the object and attribute independently, recent approaches [39, 42, 29] have regarded contextuality as a crucial standpoint for understanding the interpretation of the attribute depending on the object class. For example, AttrOps [42] represented the attribute as a linear transformation of the object's state, and SymNets [29] regularized visual features to satisfy the group axioms with auxiliary losses. Some works [41, 75] have leveraged graph networks to model the global dependencies among primitives and their compositions.", + "bbox": [ + 496, + 386, + 890, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, those methods have often suffered in distinguishing visual composition features. Meta-learning [13] has been utilized for CZSL frameworks [49, 67] to enhance the discriminative power of visual features but induce high training costs. Other works [52, 27, 53] have disentangled representations of objects and attributes for modeling visual composition. Those methods have shown promising generalization performance than conventional methods. However, the aforementioned approaches have often failed to discriminate the minority classes in object and attribute, which is caused by severe model bias toward seen compositions [61, 81]. They have used the mostly-shared visual backbone for attribute and object representations, which needs more deep concern to be extracted differently according to its characteristics. In specific, the attribute feature is changed in a wide variety of aspects as the property of the object changes, and the object feature must be perceived as semantically the same despite changing the attribute.", + "bbox": [ + 496, + 583, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is also closely related to visual attribute prediction [46, 26], where the goal is to recognize visual attributes describing states or properties within an object.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5676", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6570432d57a19d81c1db8062ac8362c6c3d3a36e8686e094024126723cb18130.jpg", + "image_caption": [ + "Figure 2: The overall architecture of composition transformer. Object and attribute experts produce object- and attribute-specific features at different layers. An object-guided attention module is attached to each attribute experts to highlight attribute-specific features. Object-attribute embeddings are composed, and then projected to a joint embedding space with label embedding for optimization." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 888, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the fine-grained local information (e.g. color, shape, and style) is required to the task, several approaches [55, 60, 20, 58, 48] utilize low-level features in backbone network, which contain textual details and objects' contour information [74, 69]. Meanwhile, with a similar motivation, some methods for CZSL have integrated low-level features with a multi-scale fusion approach to acquire robust visual features [70, 75]. Although a simple integrated feature can recover spatial details slightly, it is hard to localize the RoI of the attribute that varies with the object for compositional recognition; for example, a puddle in an object 'Street' for a composition class 'Wet Street'.", + "bbox": [ + 75, + 286, + 472, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, attributes in generic objects are long-tailed distributed in nature [55, 48]. The imbalance issue occurs as a biased recognition problem, because deep networks tend to be over-confident in head composition classes [54, 65]. This phenomenon is also known as a hubness problem [10, 12], where few visual features from head composition classes become indistinguishable from many tail composition samples. One major solution for the data imbalance problem is the resampling [15, 57, 14, 35] by flattening the long-tailed dataset by over-sampling of minority classes or under-sampling of majority classes. However, directly applying such sampling strategies in CZSL could intensify the overfitting problem on over-sampled classes, and discard valuable samples for learning the alignment between visual and semantic (linguistic) features in the joint embedding space [81]. Another line of work is generation methods [68, 66] that hallucinate a new sample from rare classes but they require complicated training strategies for data generation. In this work, we are inspired by recent data augmentation techniques [77, 7, 45, 79, 8], and combine both resampling and generation approaches to tailor for CZSL, which facilitates the regularization of robust embedding space even in the minority classes via augmented virtual samples.", + "bbox": [ + 75, + 468, + 472, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Composition Transformer (CoT)", + "text_level": 1, + "bbox": [ + 76, + 844, + 380, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In compositional zero-shot learning (CZSL), given an object set $\\mathcal{O}$ and an attribute set $\\mathcal{A}$ , object-attribute", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pairs comprise a composition set $\\mathcal{V} = \\mathcal{A}\\times \\mathcal{O}$ . The network is trained with an observable (seen) composition set $\\mathcal{D}_{\\mathrm{tr}} = \\{\\mathcal{X}_{\\mathrm{s}},\\mathcal{Y}_{\\mathrm{s}}\\}$ , where $\\mathcal{X}_{\\mathrm{s}}$ is an image set corresponding to a composition label set $\\mathcal{V}_{\\mathrm{s}}$ . Following the generalized zero-shot setting [49, 5, 73], a test set $\\mathcal{D}_{\\mathrm{te}} = \\{\\mathcal{X}_{\\mathrm{te}},\\mathcal{Y}_{\\mathrm{te}}\\}$ consists of samples drawn from both seen and unseen composition classes, i.e. $\\mathcal{V}_{\\mathrm{te}} = \\mathcal{V}_{\\mathrm{s}}\\cup \\mathcal{V}_{\\mathrm{u}}$ and $\\mathcal{V}_{\\mathrm{s}}\\cap \\mathcal{V}_{\\mathrm{u}} = \\emptyset$ . For this task, a powerful visual network is required to extract distinct features of objects and attributes respectively, in order to recognize both seen and unseen compositions well.", + "bbox": [ + 496, + 286, + 892, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose a novel composition transformer (CoT) that includes object and attribute expert modules as illustrated in Fig. 2. Our CoT is basically built on the vision transformer (i.e. ViT [11]) and fully utilizes intermediate blocks to capture the attribute and object representations. These representations are projected into the joint embedding space to calculate the similarity between given compositional labels and images. We also introduce a minority attribute sampling method for data augmentation such that the biased recognition with head composition class is alleviated by label-smoothing regularization effect [40]. Next, we describe in detail the configurations of CoT.", + "bbox": [ + 496, + 438, + 893, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Object expert", + "text_level": 1, + "bbox": [ + 500, + 628, + 640, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The CoT is composed of a sequence of transformer blocks including multi-head self-attention and MLP layers. Generally, as the block becomes deeper, class discrimination capability is increased while being invariant from the attribute deformation [78, 43]. Therefore, we select the final output from the last block to an input of the object expert in a bottom-up manner [31, 1].", + "bbox": [ + 496, + 651, + 893, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, let us denote $\\{\\mathbf{z}^n\\}_{n=1}^N$ as the output features of the blocks in CoT. The feature of $n$ -th transformer blocks, $\\mathbf{z}^n$ , consists of a class token (i.e., [CLS]) $\\mathbf{z}_c^n \\in \\mathbb{R}^{1 \\times D}$ and a set of patch tokens $\\mathbf{z}_p^n \\in \\mathbb{R}^{HW \\times D}$ . The object expert $E_o$ encodes the last feature to present an object embedding:", + "bbox": [ + 496, + 756, + 893, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {o} = E _ {o} \\left(\\mathbf {z} _ {c} ^ {N}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 842, + 892, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{o}$ consists of a fully-connected (FC) layer. To optimize the parameters in the object expert, we compute a", + "bbox": [ + 496, + 869, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5677", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cross-entropy loss with an object classifier whose weights are initialized from object word-embeddings as follows:", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o b j}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {o} , w \\left(o _ {i}\\right)\\right) / \\tau_ {o} \\right\\}}{\\sum_ {o _ {k} \\in \\mathcal {O}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {o} , w \\left(o _ {k}\\right)\\right) / \\tau_ {o} \\right\\}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 131, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature parameter, $w(o)$ denotes the word-embedding [47] corresponding to the object $o\\in \\mathcal{O}$ .", + "bbox": [ + 76, + 176, + 468, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Attribute expert", + "text_level": 1, + "bbox": [ + 76, + 217, + 238, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Capturing both attribute and object in the same deeply condensed feature is nontrivial as they require pretty different visual representations [30, 22]. The straightforward solution is to design two-stream networks without any parameter sharing. However, such individual training could not capture the contextuality [39]. Hence, we build an attribute expert which utilizes semantics estimated from the object expert as a condition, and extracts attribute-related features on intermediate layers of visual backbone in a top-down fashion [31, 1]. We devise an object-guided attention module that leverages convolution kernel-based attention [71, 6, 24], which excels at capturing fine-grained local information required for attribute recognition [18].", + "bbox": [ + 75, + 239, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first reshape $\\mathbf{z}_p^n$ to $\\mathbf{Z}_p^n$ with output resolution of $H\\times W\\times D$ . Following [51], we tile $\\mathbf{p}_o$ to match the spatial dimension with $\\mathbf{Z}_p^n$ and concatenate them to obtain an object-contextualized feature $\\{\\mathbf{U}^n = [\\mathbf{p}_o,\\mathbf{Z}_p^n ]\\} \\in \\mathbb{R}^{H\\times W\\times 2D}$ . As shown in Fig. 2, we generate a channel attention map $\\mathbf{A}_c^n\\in \\mathbb{R}^{1\\times 1\\times D}$ from the averaged feature across the spatial axis:", + "bbox": [ + 76, + 436, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {c} ^ {n} = \\sigma \\left(\\operatorname {c o n v} _ {1 \\times 1} \\left(\\operatorname {a v g p o o l} _ {\\text {s p a t i a l}} \\left(\\mathbf {U} ^ {n}\\right)\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 555, + 468, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes a sigmoid function. We also generate a spatial attention map $\\mathbf{A}_s^n\\in \\mathbb{R}^{H\\times W\\times 1}$ from the averaged feature across the channel axis:", + "bbox": [ + 76, + 584, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {s} ^ {n} = \\sigma \\left(\\operatorname {c o n v} _ {3 \\times 3} \\left(\\operatorname {a v g p o o l} _ {\\text {c h a n n e l}} \\left(\\mathbf {U} ^ {n}\\right)\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 641, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we aggregate two attention maps into a unified attention map $\\mathbf{A}^n\\in \\mathbb{R}^{H\\times W\\times D}$ as follows:", + "bbox": [ + 76, + 669, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} ^ {n} = \\mathbf {A} _ {s} ^ {n} \\times \\mathbf {A} _ {c} ^ {n}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 710, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output of our object-guided attention module is defined as attribute features with estimated attention maps, followed by the residual connection:", + "bbox": [ + 76, + 739, + 468, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {Z}} _ {p} ^ {n} = \\mathbf {Z} _ {p} ^ {n} + \\mathbf {A} ^ {n} \\odot \\mathbf {Z} _ {p} ^ {n}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 795, + 468, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ is an operation for element-wise multiplication.", + "bbox": [ + 76, + 825, + 450, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $\\tilde{\\mathbf{Z}}_p^n$ from each intermediate layer, we ensemble the features from different blocks by taking advantage of the hierarchical architecture [33, 43]. It is required for recognizing heterogeneous types of attributes such as colors or", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "shapes [48]. Especially, we choose low-, middle-, and high-level features from the backbone, denoted as $\\{\\tilde{\\mathbf{Z}}_p^l,\\tilde{\\mathbf{Z}}_p^m,\\tilde{\\mathbf{Z}}_p^h\\}$ , and ensemble those triplets with global average pooling (GAP) followed by concatenation. The attribute expert $E_{a}$ then generates an attribute embedding with the triplets as", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {a} = E _ {a} \\left(\\left[ \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {l}\\right), \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {m}\\right), \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {h}\\right) \\right]\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 174, + 890, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{a}$ consists of a FC layer. Similar to the object expert, we optimize the attribute expert with the cross-entropy loss:", + "bbox": [ + 498, + 199, + 890, + 231 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a t t}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {a} , w \\left(a _ {i}\\right)\\right) / \\tau_ {a} \\right\\}}{\\sum_ {a _ {k} \\in \\mathcal {A}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {a} , w \\left(a _ {k}\\right)\\right) / \\tau_ {a} \\right\\}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 250, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Mapping visual-semantic space", + "text_level": 1, + "bbox": [ + 500, + 292, + 777, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual compositional embedding $\\mathbf{p}_x$ of the input image $x$ is calculated from object and attribute experts, formulated as", + "bbox": [ + 498, + 315, + 890, + 358 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {x} = E _ {c} \\left(\\left[ \\mathbf {p} _ {o}, \\mathbf {p} _ {a} \\right]\\right)). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 359, + 890, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The embedding function $E_{c}$ includes an FC layer, projecting the visual features into the joint embedding space that aligns visual and semantic representations. Following previous works [42, 36], we estimate semantic label embedding $\\mathbf{p}_y$ of each seen composition label $y \\in \\mathcal{V}_s$ by projecting the concatenated word vectors into the joint space:", + "bbox": [ + 498, + 381, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {y} = g ([ w (o), w (a) ]), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 479, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g(\\cdot)$ is a label embedding network that consists of 3 FC layers and ReLU activation function. During training, we minimize the cosine distance of visual and semantic embeddings with cross-entropy loss as follows:", + "bbox": [ + 498, + 505, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o m p}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {x} , \\mathbf {s} _ {y}\\right) / \\tau_ {c} \\right\\}}{\\sum_ {y _ {k} \\in \\mathcal {Y} _ {s}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {x} , \\mathbf {s} _ {y _ {k}}\\right)\\right) / \\tau_ {c} \\}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 574, + 890, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau_{c}$ is a temperature parameter for the composition loss $\\mathcal{L}_{\\mathrm{comp}}$ . Finally, the total objectives of CoT is then formulated as follows:", + "bbox": [ + 498, + 616, + 890, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o m p}} + \\alpha \\mathcal {L} _ {\\text {a t t}} + \\beta \\mathcal {L} _ {\\text {o b j}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 670, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are balance weights.", + "bbox": [ + 500, + 695, + 736, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, we measure the cosine similarity between a visual embeddings and all label embeddings from $\\mathcal{V}_{\\mathrm{te}}$ and regard it as a feasibility score of the image and composition labels [36]. Following [80], we also use the classification scores from object and attribute experts as a form of cosine similarity. Therefore, a final feasibility score of label $y = (a,o)$ is derived by adding the above three scores, formulated as:", + "bbox": [ + 498, + 710, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc (y) = \\cos \\left(\\mathbf {p} _ {x}, \\mathbf {s} _ {y}\\right) + \\cos \\left(\\mathbf {p} _ {o}, w (o)\\right) + \\cos \\left(\\mathbf {p} _ {a}, w (a)\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 839, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we predict the label with the highest score as the final composition label.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5678", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c7e939d1597c896f8580d538c27c5397b248b3f0c572ad155a5902f77c49ca9e.jpg", + "image_caption": [ + "Figure 3: Attribute class distributions on VAW-CZSL. Every attributes in each plot are composed into a fixed object; (a) Grass and (b) Player." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 269, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bd24f994b76243dcb3e9cc538ec455ae52d1e59cdf7f07d99b15bdde94160da6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 90, + 460, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Minority attribute augmentation", + "text_level": 1, + "bbox": [ + 76, + 297, + 364, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With a carefully designed architecture, our CoT can generate distinctive visual features for CZSL. Nevertheless, the imbalanced composition samples in the training dataset restrict the performance since attributes often co-occur with certain objects in the real world [30]. For example in Fig. 3, 'Green Grass' is naturally more common than 'White Grass'. Such long-tailed data makes the deep models be over-confident in the composition with head attributes [81], and thus hurts the generalization on unseen compositions.", + "bbox": [ + 75, + 321, + 467, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we present a minority attribute augmentation (MAA) for training, which diversifies the minority composition. First, two compositional embeddings from COt are selected as inputs to the augmentation, where $\\mathbf{p}_{x_A}$ and $\\mathbf{p}_{x_B}$ share the same object class. Especially, to balance the major and minor attributes, we assign higher sampling weights on minority attributes with the weighted sampler [14]. Formally, let $\\zeta_{o_i}$ denote the number of attribute samples in the $i$ -th attribute class composed on an object class $o$ . The sampling weight of the attribute-object pair $\\kappa(a_i,o)$ is formulated by an inverse attribute class frequency according to:", + "bbox": [ + 75, + 458, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa \\left(a _ {i}, o\\right) = \\frac {1 / \\zeta_ {o _ {i}}}{\\sum_ {i} 1 / \\zeta_ {o _ {i}}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 635, + 468, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by Mixup [79, 64], we generate a virtual sample $\\mathbf{p}_{x_M}$ with alpha compositing and its attribute semantic embedding $w(a_{M})$ by linear combination:", + "bbox": [ + 76, + 674, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {x _ {M}} = \\lambda \\mathbf {p} _ {x _ {A}} + (1 - \\lambda) \\mathbf {p} _ {x _ {B}},\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 733, + 364, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(a _ {M}\\right) = \\lambda w \\left(a _ {A}\\right) + (1 - \\lambda) w \\left(a _ {B}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 751, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mixing factor $\\lambda$ is sampled from a uniform distribution.", + "bbox": [ + 76, + 779, + 467, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We notice that there are two significant differences from Mixup: (1) Compared to Mixup sampling arbitrary images, MAA constructs a composition pair fastening an object with the minority aware sampling to address the imbalanced data issue in CZSL. (2) Mixup mixes one-hot vector labels [79, 77] for classification, while MAA generates a virtual label in word vector space, to jointly regularize vi", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sual and linguistic (label) embedding functions against the majority attribute classes with Eq. (12).", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Extension to other backbones", + "text_level": 1, + "bbox": [ + 500, + 131, + 763, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our idea of CoT is generic and scalable, so we can simply extend to other backbones such as ResNet-18 (R18) [16] by selecting specific layers from the backbone composed of multiple layers or blocks. For a fair comparison to prior arts [41, 53, 23], we study the impact of our contribution on the ResNet backbone based on the convolutional operation. In specific, the object expert generates the object embedding from the last fifth convolution block. The classification token is replaced by a global average pooled feature. Similar to using ViT [11], we concatenate the intermediate layers into a feature vector for the attribute expert. The remained settings such as the architecture of $E_{\\mathrm{comp}}$ and $g(\\cdot)$ are the same as ViT.", + "bbox": [ + 496, + 154, + 893, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 354, + 633, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 500, + 378, + 687, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. MIT-States [19] dataset is a common dataset in CZSL that most previous methods adopted, containing 115 states, 245 objects, and 1962 compositions. C-GQA [41] and VAW-CZSL [53] are two brand-new and large-scale benchmarks collected from in-the-wild images with more general compositions, such as 'Stained Wall' and 'Hazy Mountain'. Detailed data statistics including long-tailed distribution can be found in Appendix A.", + "bbox": [ + 496, + 401, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We follow common CZSL setting [49], testing performance on both seen and unseen pairs [5]. Under the setting, we report area under the curve (AUC) (\\%) between seen and unseen accuracies at different bias term in validation and test splits. We also compute the best seen and unseen accuracies in the curve, and calculate the best harmonic mean to measure the trade-off between the two values. Lastly, we follow [41] and report the attribute and object accuracies on unseen labels to measure the discrimination power of each representation.", + "bbox": [ + 496, + 522, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our CoT with 9 recent CZSL approaches: AoP [42], LE+ [39], TMN [49], Symnet [29], CompCos [36], CGE [41], SCEN [27], OADis [53], CAPE [23]. All baselines were built on pretrained ResNet-18 (R18) [16], with pre-trained word embedding such as GloVe [47], FastText [4], and Word2Vec [37]. To show the contribution of our proposed elements, we carry on the most recent methods including CGE and OADis with a transformer-based backbone (ViT-B) [11]. We also clarify that CGE and OADis utilize the knowledge of unseen composition labels at training for building a relation graph and hallucinating virtual samples respectively, while other methods do not assess the unseen composition label sets. The results of other baselines were obtained from [41, 27, 23, 53] or re-implemented from their official code.", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5679", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8763f875621c4bb45e9e747809b95a083277926e9c53c14f6e7f824a82c47a16.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Quantitative comparison with prior arts. Following [41, 53], we measure AUC values on validation (V) and test (T) datasets, at top-1 (@1) on C-GQA and top-3 (@3) on VAW-CZSL, respectively. In addition, we report harmonic mean (HM), seen (S), unseen (US), attribute (A), and object (O) accuracies. * denotes the models accessing unseen composition labels at training. The best result and the second best result are boldfaced and underlined, respectively." + ], + "table_body": "
MethodsBackboneMIT-StatesC-GQAVAW-CZSL
V@1T@1HMSUV@1T@1HMSUV@3T@3HMSU
With frozen R18
AoP [42]R182.52.010.716.618.40.90.32.911.83.91.41.49.116.411.7
LE+ [39]R183.52.311.516.221.21.20.65.316.15.01.51.69.816.213.2
TMN [49]R183.32.611.822.717.12.21.17.721.66.32.22.311.919.915.4
Symnet [29]R184.53.413.824.820.03.31.89.825.29.22.32.312.219.115.8
CompCos [36]R186.94.816.926.924.53.52.612.128.111.83.13.214.223.918.0
CGE* [41]R187.25.318.128.925.03.62.511.927.511.72.72.913.023.416.8
SCEN [27]R187.25.318.429.925.24.02.912.428.912.1-----
OADis* [53]R187.65.918.931.125.6-----3.53.615.224.918.7
CAPE [23]R18-5.819.130.526.2-4.216.332.915.6-----
CoTR187.76.219.630.826.84.94.516.633.116.63.63.815.724.619.1
With frozen ViT-B
CGE* [41]ViT-B8.77.321.333.528.64.53.815.631.314.43.73.915.726.319.7
OADis* [53]ViT-B9.07.521.934.229.35.94.616.232.515.34.14.316.626.920.8
CoTViT-B9.27.823.234.831.56.75.117.534.018.84.44.717.726.922.2
With finetuning ViT-B
CGE* [41]ViT-B11.49.724.839.731.67.35.418.538.017.15.96.220.130.125.7
OADis* [53]ViT-B11.810.125.239.232.18.17.020.138.319.86.26.520.431.326.1
CoTViT-B12.110.525.839.533.08.77.422.139.222.76.67.221.732.928.2
", + "bbox": [ + 81, + 88, + 895, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 474, + 290, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We use ViT-B [11] and ResNet18 [16] pre-trained on ImageNet [9] as a visual backbone. In ViT-B backbone, Each patch size is $16 \\times 16$ , and the number of patch tokens in each layer is $14 \\times 14$ , having 768 channels per token. The label embedding network $g$ is composed of 3 FC layers, two dropout layers with ReLU activation functions, where hidden dimensions are 900. We use one FC layer for $E_{o}$ and $E_{a}$ to produce 300-dimensional prototypical vectors that are matched with GloVe [47] embedding vectors from composition labels. Therefore, both $E_{c}$ and $g(\\cdot)$ produce 300-dimensional embedding vectors, with $l_{2}$ normalization. In the attribute expert, we use the outputs of the 3rd, 6th and 9th blocks (denoted $\\mathbf{z}^{l}, \\mathbf{z}^{m}, \\mathbf{z}^{h}$ respectively) for multi-level feature fusion. For ResNet18, we use the 2nd, 3rd and 4th blocks as $\\mathbf{z}^{l}, \\mathbf{z}^{m}, \\mathbf{z}^{h}$ , resulting in a 438-dimensional feature vector.", + "bbox": [ + 75, + 502, + 470, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training setup. We train CoT with Adam optimizer [25]. The input images are augmented with random crop and horizontal flip, and finally resized into $224 \\times 224$ . In ViT, learning rate is initialized to 1e-4 for all three datasets, and decayed by a factor of 0.1 for 10 epochs. In R18, the initial learning rate is 5e-5 with the same decay parameter of the ViT setting. Note that we freeze the GloVe [47] word embedding $w$ for fair comparisons. For all datasets, we use the same temperature parameters [36, 53, 27] of each cross-entropy loss, $\\tau_{c}$ , $\\tau_{o}$ , and $\\tau_{a}$ as 0.05, 0.01 and 0.01 respec", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e5a931173816bbd0bbdf14d9db7c463d6c762920195841d22defbfeda577c3f8.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Comparison of unseen attribute (Attr.) and object (Obj.) accuracies with the latest CZSL methods [41, 53]. All results are reported on fine-tuned ViT backbone." + ], + "table_body": "
MethodsMIT-StatesC-GQAVAW-CZSL
Attr.Obj.Attr.Obj.Attr.Obj.
CGE [41]35.744.417.534.624.551.6
OADis [53]35.245.214.742.022.353.8
CoT37.346.019.840.227.354.2
", + "bbox": [ + 514, + 472, + 877, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tively. We use the loss balance weights $\\alpha$ and $\\beta$ as (0.5, 0.5) on C-GQA, and (0.4, 0.6) on VAW-CZSL and MIT-States. We note that to prevent overfitting of COT we apply MAA from 15 and 30 epochs to ViT and R18, respectively.", + "bbox": [ + 498, + 638, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main evaluation", + "text_level": 1, + "bbox": [ + 500, + 710, + 661, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares generalization performance between our methods and the baselines with ResNet18 [16] (R18) and ViT-B [11] backbone setting. To further compare visual discrimination, we also compare unseen attribute and object accuracies of our methods and the latest CZSL methods [41, 53] in Table 2. In the following, we analyze the results from the three datasets.", + "bbox": [ + 496, + 733, + 890, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MIT-States. From Table 1, the CoT achieves the best AUC values (V@1 and H@1) and harmonic mean (HM) with both R18 and ViT backbone. These results sufficiently demonstrate our experts' applicability on the CNN-based", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5680", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a59829124f7d95109ce6246478b586fdf86ebe5ce84113faa9412d865358aca0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CoTMAAAUCHM
5.8118.1
✓6.0519.0
✓6.9320.8
✓✓7.2021.7
", + "bbox": [ + 181, + 98, + 343, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2c1c66eba70725811a7cdf5bdb5455fe7f5c762d23575bf26dcbd980259b8dc9.jpg", + "table_caption": [ + "(a) Component analysis." + ], + "table_footnote": [], + "table_body": "
Loss componentAUCHM
Lcomp5.9418.9
Lcomp+Lobj6.2219.6
Lcomp+Lattr6.9121.4
Lcomp+Lattr+Lobj7.2021.7
", + "bbox": [ + 387, + 98, + 573, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f3a6626fc535f0416c7d051dd2d283dda02ba8acefb6b8724817de1fcba2e13.jpg", + "table_caption": [ + "(b) Impact of loss functions." + ], + "table_footnote": [], + "table_body": "
AugmentationAUCHM
Manifold Mixup [64]5.9318.8
CutMix [77]6.9520.7
Mixup [79]7.0421.0
MAA7.2021.7
", + "bbox": [ + 614, + 98, + 810, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e1c1210f43a59659bd5099fd82035510378a1be359ba4f90438069af4a63675d.jpg", + "table_caption": [ + "(c) Impact of data augmentation." + ], + "table_footnote": [], + "table_body": "
GuidanceAUCHMSUAO
w/o guidance6.720.530.725.822.553.4
Attribute guidance6.921.031.526.425.753.9
Object guidance7.221.732.928.227.354.2
", + "bbox": [ + 140, + 205, + 452, + 273 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d046198ae3328d3c1040b054f3eac870f243543573d13eca28adadf362b1e513.jpg", + "table_caption": [ + "(d) Impact of object guidance." + ], + "table_footnote": [], + "table_body": "
EnsemblezlzmzhAUCHMSUAO
Low2346.620.730.328.725.852.3
Mid5676.921.231.929.725.255.9
High910116.821.433.426.624.154.5
Mixture3697.221.732.928.227.354.2
", + "bbox": [ + 470, + 200, + 843, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(e) Different configurations of intermediate blocks in attribute experts.", + "bbox": [ + 467, + 282, + 839, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Ablation experiments on VAW-CZSL dataset. Note that all results are reported on CoT (ViT-B), with a fine-tuning setting.", + "bbox": [ + 94, + 303, + 870, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "backbone. With a finetuned ViT-B, the CoT obtains the best validation AUC of $12.1\\%$ and test AUC of $10.5\\%$ , surpassing previous CZSL methods with a large margin. Our method also improves attribute (Attr.) and object (Obj.) accuracies as in Table 2, showing the effectiveness of CoT for visual discrimination.", + "bbox": [ + 75, + 343, + 468, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "C-GQA. Similar trends can be observed in the C-GQA dataset. In the frozen R18 setting, the CoT achieves the best test AUC (T@1) of $4.5\\%$ , comparable with previous state-of-the-art CAPE. Although the T@1 scores from CGE and OADis are also increased upon replacing the backbone from R18 to ViT-B, our CoT with ViT-B outperforms OADis with larger margins in all metrics, as compared to the performance with R18. It is noticeable that our proposed augmentation technique contributes the unbiased prediction, improving unseen accuracies into $18.8\\%$ with a minority-aware sampling. The performance regarding AUC is improved with end-to-end training with backbone, $2.0\\%$ in V@1 and $2.3\\%$ in T@1. Of note, OADis and CGE access unseen composition labels at training, so it is unfair to directly compare with other methods including CoT. Nevertheless, more surprisingly, our CoT shows comparable and superior performance to OADis. Finally, in Table 2, our model also improves the attribute accuracy by achieving $19.8\\%$ , having comparable $40.2\\%$ object accuracy compared to OADis and CGE.", + "bbox": [ + 75, + 440, + 468, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VAW-CZSL. We also achieve SoTA performance on VAW-CZSL, the most challenging benchmark having almost 10K composition labels for evaluation. Even if we do not use additional unseen composition information like OADis, our method achieves the highest T@1 of $7.2\\%$ with the best unseen label accuracy of $28.2\\%$ . We also obtain both the highest attribute and object score (Table 2), especially with a significant improvement in attribute prediction by $2.8\\%$ compared to CGE. It proves that the careful modeling of the visual network in CoT is essential to substantially improve", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the generalization capability. In addition, the results suggest that finetuning the visual backbone to suit composition recognition improves the performance in all benchmarks.", + "bbox": [ + 496, + 343, + 890, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 402, + 648, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We ablate our CoT and MAA with different design choices on Table 3. We use the VAW-CZSL dataset and fine-tuned ViT-B backbone as our main ablation setting.", + "bbox": [ + 496, + 428, + 890, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component analysis. We report ablation results of CoT and MAA in Table 3a. In 'without CoT' case, we use CompCos [36] baseline. CoT boosts the AUC and HM significantly about $1.1\\%$ and $2.3\\%$ compared to the baseline. MAA offers better results to both CoT and baseline, showing scalability to other methods. Notably, employing MAA on CoT gives higher gains of $0.3\\%$ AUC and $2.3\\%$ HM compared to the baseline, demonstrating that the two components create a synergy effect.", + "bbox": [ + 496, + 474, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of loss functions. Table 3b shows the impact of each loss function. Compared to the case using $\\mathcal{L}_{\\mathrm{comp}}$ only, $\\mathcal{L}_{\\mathrm{obj}}$ and $\\mathcal{L}_{\\mathrm{attr}}$ bring a gain on both AUC and HM. Using both losses to $\\mathcal{L}_{\\mathrm{comp}}$ improves AUC and HM significantly by $1.26\\%$ and $2.8\\%$ respectively, showing the importance of object and attribute losses to learn each expert.", + "bbox": [ + 496, + 612, + 890, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different data augmentations. In Table 3c, we compare MAA with various data mixing methodologies including Manifold Mixup [64], CutMix [77] and Mixup [79]. To apply Manifold Mixup in CoT, we select 3,6,9 and the last layer as a mixup layer, and mix these intermediate features. MAA achieves the highest score on both AUC and HM, outperforming other augmentation methods including standard Mixup. This demonstrates the effectiveness of MAA which balances the data distribution with a composition pair. Interestingly, we found that Manifold Mixup performs at worst, even degrading pure CoT performance (AUC: 6.93, HM: 20.4) in Table 3a. We speculate that the mixed representation at intermediate layers can lead to significant underfit", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5681", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f42adfde4d5eb773bcde54cab35a89d2ebb499cb29e0c8ccd614af226591ee50.jpg", + "image_caption": [ + "Figure 4: Distribution of k-occurrence counts with $k = 20$ of C-GQA and VAW-CZSL test set. All results are reported with ViT-B backbone. x-axis stands for composition class ids which are ordered by decreasing count numbers. For better visualization, we show the top $10\\%$ head composition classes where the hubness is predominant. Circles at start points indicate the most frequent hubs. The result on MIT-States is in the Appendix C.1." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ting of the attention module.", + "bbox": [ + 75, + 351, + 266, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of object guidance. As shown in Table 3d, the object guidance significantly boosts the performance. We found that reversing the guidance (attribute guidance) decreases the performance of AUC by $0.3\\%$ and HM by $0.7\\%$ . The results validate the usage of object guidance for contextualized attribute prediction.", + "bbox": [ + 75, + 366, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design choice of feature ensemble. Table 3e shows the ablations with different configurations of intermediate blocks used in the attribute expert. We evaluate four types of block feature ensembles: low, mid, high, and mixture. Among them, the mixture type achieves the best performance on AUC, harmonic mean, and attribute accuracy. We therefore choose this mixture ensemble as a default setting. Another finding is low-level feature is more adequate for attribute and unseen composition recognition, while high-level feature is robust to object and seen composition accuracy.", + "bbox": [ + 75, + 458, + 468, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 633, + 192, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mitigating hubness problem. We illustrate the hubness effect of a baseline [36] and our method on C-GQA and VAW-CZSL datasets in Fig. 4. Concretely, we count $k$ -occurrences $(N_{k})$ , the number of times that a sample appears in the $k$ nearest neighbors of other points [50] for all class ids, and calculate the skewness of $N_{k}$ distribution for hubness measurement. We observe that CoT significantly alleviates the hubness problem of baseline across datasets, enlarging the visual discrimination in the embedding space. MAA further reduces the skewness of the distribution by mitigating biased prediction on head composition classes.", + "bbox": [ + 75, + 657, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object-guided attention visualization. Qualitatively, we visualize attention maps from the object-guided attention module in Fig. 5. We observe that the attention module could capture the contextualized meaning of attributes in composition, e.g., wet skin of horse vs. puddle in street. It", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f276bc4dd14e73b5ec5444822ca5e18bbb2920da2515767590b10ed3438525e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 888, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6a7ea8ed80d52633e8d036bb4ca5cb40b471006edd68918b7a3fe6da80422f2.jpg", + "image_caption": [ + "Figure 5: Visualization of object-guided attention on VAW-CZSL dataset. Each row represents the attended regions in image pairs, which are related to different attributes contextualized with a specific object. More results in the Appendix C.5." + ], + "image_footnote": [], + "bbox": [ + 501, + 181, + 887, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ba0c5c88fa9fe2ea9383b762b07160bc76ad1a2070cca0500779df7cc311e2f.jpg", + "image_caption": [ + "Graite Ground" + ], + "image_footnote": [], + "bbox": [ + 506, + 368, + 888, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cabdee7b0e0a882dd56a5548893f2db4e16a43e51e71986cc57a43bfe6f40bff.jpg", + "image_caption": [ + "(d)", + "(g)", + "Figure 6: Ground-truth composition labels and top-3 prediction results on MIT-States, C-GQA and VAW-CZSL datasets. More results in the Appendix C.6." + ], + "image_footnote": [], + "bbox": [ + 506, + 491, + 635, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f38e241c7dbebadba09033625aaf18fbe5a4d186c0a6bb4a5c8193c2a68fcea.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 645, + 491, + 759, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/439a78825ff7be7f6787cad581a1dc46ecaa9cf97ac9e76f871459c5c423b0a7.jpg", + "image_caption": [ + "(i)" + ], + "image_footnote": [], + "bbox": [ + 771, + 491, + 883, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "validates that the attention module assists CoT to extract distinctive attribute embedding for composition by localizing specific RoI.", + "bbox": [ + 498, + 648, + 890, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 6 illustrates the qualitative results of CoT. Even though some predictions do not correspond to annotated image labels, they explain the composition in another view; e.g., 'Furry Dog' and 'Shaggy Dog' in Fig. 6 (i). Moreover, in-the-wild images usually have multiple objects with their own attributes; e.g., 'White Wall' and 'Reading Man' in Fig. 6 (h). To recognize such compositions in the real world, it is necessary to handle the multi-label nature of composition, while distinguishing each attribute-object interaction [17] in an image. It encourages us to construct a new CZSL benchmark along with a novel evaluation metric handling both multi-label prediction and multiple compositions in the image. We leave it as a future work.", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5682", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 89, + 194, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose Composition Transformer (CoT) with simple minority attribute augmentation (MAA), a contextual, discriminative, and unbiased CZSL framework. The proposed object and attribute experts hierarchically utilize the entire powerful visual backbone to generate a composition embedding. The simple but effective MAA balances the long-tail distributed composition labels. Extensive studies with comprehensive analysis demonstrate the effectiveness of each component, and show that CoT surpasses existing methods across several benchmarks.", + "bbox": [ + 75, + 111, + 468, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This research was supported by the Yonsei Signature Research Cluster Program of 2022 (2022-22-0002) and the KIST Institutional Program (Project No.2E31051-21-203).", + "bbox": [ + 76, + 268, + 470, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 344, + 174, + 359 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 3, 4", + "[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 2", + "[3] Yuval Atzmon, Felix Kreuk, Uri Shalit, and Gal Chechik. A causal view of compositional zero-shot recognition. In NeurIPS, 2020. 2", + "[4] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. TACL, 2017. 5", + "[5] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In ECCV, 2016.3,5", + "[6] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020. 4", + "[7] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In ECCV, 2020. 2, 3", + "[8] Yu-Ying Chou, Hsuan-Tien Lin, and Tyng-Luh Liu. Adaptive and generative zero-shot learning. In International conference on learning representations, 2020. 3", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6", + "[10] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2, 3", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 369, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3, 5, 6", + "[12] Nanyi Fei, Yizhao Gao, Zhiwu Lu, and Tao Xiang. Z-score normalization, hubness, and few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 142-151, 2021. 2, 3", + "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2", + "[14] Yonatan Geifman and Ran El-Yaniv. Deep active learning over the long tail. arXiv preprint arXiv:1711.00941, 2017. 3,5", + "[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE TKDE, 2009. 3", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 5, 6", + "[17] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In ECCV, 2020. 8", + "[18] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In CVPR, 2020. 4", + "[19] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 5", + "[20] Mahdi M Kalayeh and Mubarak Shah. On symbiosis of attribute prediction and semantic segmentation. IEEE TPAMI, 2019. 3", + "[21] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Revisiting visual product for compositional zero-shot learning. In NeurIPSW, 2021. 1", + "[22] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In CVPR, 2022. 2, 4", + "[23] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Alain Pagani, Didier Stricker, and Muhammad Zeshan Afzal. Learning attention propagation for compositional zero-shot learning. arXiv preprint arXiv:2210.11557, 2022. 5, 6", + "[24] Hanjae Kim, Sunghun Joung, Ig-Jae Kim, and Kwanghoon Sohn. Prototype-guided saliency feature learning for person search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4865-4874, 2021. 4", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6", + "[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2", + "[27] Xiangyu Li, Xu Yang, Kun Wei, Cheng Deng, and Muli Yang. Siamese contrastive embedding network for compositional zero-shot learning. In CVPR, 2022. 2, 5, 6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5683", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2", + "[29] Yong-Lu Li, Yue Xu, Xiaohan Mao, and Cewu Lu. Symmetry and group in attribute-object compositions. In CVPR, 2020. 2, 5, 6", + "[30] Kongming Liang, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Unifying visual attribute learning with object recognition in a multiplicative framework. IEEE TPAMI, 2018. 2, 4, 5", + "[31] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2, 3, 4", + "[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 2", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4", + "[34] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, 2016. 2", + "[35] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 3", + "[36] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8", + "[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 1, 5", + "[38] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In CVPR, pages 12664-12673, 2020. 2", + "[39] Ishan Misra, Abhinav Gupta, and Martial Hebert. From red wine to red tomato: Composition with context. In CVPR, 2017. 1, 2, 4, 5, 6", + "[40] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 3", + "[41] Muhammad Ferjad Naeem, Yongqin Xian, Federico Tombari, and Zeynep Akata. Learning graph embeddings for compositional zero-shot learning. In CVPR, 2021. 1, 2, 5, 6", + "[42] Tushar Nagarajan and Kristen Grauman. Attributes as operators: factorizing unseen attribute-object compositions. In ECCV, 2018. 1, 2, 4, 5, 6", + "[43] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, 2021. 3, 4", + "[44] Zizheng Pan, Bohan Zhuang, Jing Liu, Haoyu He, and Jianfei Cai. Scalable vision transformers with hierarchical pool" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. In Proceedings of the IEEE/cvf international conference on computer vision, pages 377-386, 2021. 2", + "[45] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In CVPR, 2022. 2, 3", + "[46] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 2", + "[47] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 1, 4, 5, 6", + "[48] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 3, 4", + "[49] Senthil Purushwalkam, Maximilian Nickel, Abhinav Gupta, and Marc'Aurelio Ranzato. Task-driven modular networks for zero-shot compositional learning. In ICCV, 2019. 2, 3, 5, 6", + "[50] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 8", + "[51] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4", + "[52] Frank Ruis, Gertjan Burghouts, and Doina Bucur. Independent prototype propagation for zero-shot compositionality. In NeurIPS, 2021. 2", + "[53] Nirat Saini, Khoi Pham, and Abhinav Shrivastava. Disentangling visual embeddings for attributes and objects. In CVPR, 2022. 1, 2, 5, 6", + "[54] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In WACV, 2021. 3", + "[55] Nikolaos Sarafianos, Xiang Xu, and Ioannis A Kakadiaris. Deep imbalanced attribute classification using visual attention aggregation. In ECCV, 2018. 3", + "[56] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, 2019. 2", + "[57] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In ECCV, 2016. 3", + "[58] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In CVPR, 2021. 3", + "[59] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 2", + "[60] Chufeng Tang, Lu Sheng, Zhaoxiang Zhang, and Xiaolin Hu. Improving pedestrian attribute recognition with weakly-supervised multi-scale attribute-specific localization. In ICCV, 2019. 3", + "[61] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5684", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bad momentum causal effect. Advances in Neural Information Processing Systems, 33:1513-1524, 2020. 2", + "[62] Bin Tong, Chao Wang, Martin Klinkigt, Yoshiyuki Kobayashi, and Yuuichi Nonaka. Hierarchical disentanglement of discriminative latent features for zero-shot learning. In CVPR, pages 11467-11476, 2019. 2", + "[63] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 2", + "[64] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 5, 7", + "[65] Byron C Wallace and Issa J Dahabreh. Improving class probability estimates for imbalanced data. KAIS, 2014. 3", + "[66] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 3", + "[67] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, 2019. 2", + "[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018. 3", + "[69] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In CVPR, 2021. 3", + "[70] Kun Wei, Muli Yang, Hao Wang, Cheng Deng, and Xianglong Liu. Adversarial fine-grained composition learning for unseen attribute-object recognition. In ICCV, 2019. 3", + "[71] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 4", + "[72] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3907-3916, 2019. 2", + "[73] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 2018. 3", + "[74] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature based activation map for weakly supervised object localization. In ICCV, 2021. 3", + "[75] Ziwei Xu, Guangzhi Wang, Yongkang Wong, and Mohan S. Kankanhalli. Relation-aware compositional zero-shot learning for attribute-object pair recognition. IEEE TMM, 2021. 2, 3", + "[76] Muli Yang, Cheng Deng, Junchi Yan, Xianglong Liu, and Dacheng Tao. Learning unseen concepts via hierarchical decomposition and composition. In CVPR, 2020. 1" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 3, 5, 7", + "[78] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 3", + "[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 3, 5, 7", + "[80] Tian Zhang, Kongming Liang, Ruoyi Du, Xian Sun, Zhanyu Ma, and Jun Guo. Learning invariant visual representations for compositional zero-shot learning. arXiv preprint arXiv:2206.00415, 2022. 1, 2, 4", + "[81] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5685", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_model.json b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05619c436b839bbac24d5c172fb42bf7a39d3925 --- /dev/null +++ b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_model.json @@ -0,0 +1,2972 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.18, + 0.276, + 0.2 + ], + "angle": 0, + "content": "Hanjae Kim" + }, + { + "type": "text", + "bbox": [ + 0.311, + 0.181, + 0.42, + 0.199 + ], + "angle": 0, + "content": "Jiyoung Lee²" + }, + { + "type": "text", + "bbox": [ + 0.455, + 0.181, + 0.593, + 0.199 + ], + "angle": 0, + "content": "Seongheon Park" + }, + { + "type": "text", + "bbox": [ + 0.63, + 0.181, + 0.802, + 0.199 + ], + "angle": 0, + "content": "Kwanghoon Sohn\\(^{1,3,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.205, + 0.25, + 0.223 + ], + "angle": 0, + "content": "\\(^{1}\\)Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.205, + 0.426, + 0.222 + ], + "angle": 0, + "content": "\\(^{2}\\)NAVER AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.466, + 0.205, + 0.594, + 0.223 + ], + "angle": 0, + "content": "\\(^{3}\\)Korea Institute" + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.206, + 0.872, + 0.224 + ], + "angle": 0, + "content": "of Science and Technology (KIST)" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.23, + 0.543, + 0.245 + ], + "angle": 0, + "content": "{incohjk,sam121796,khsohn}@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.602, + 0.231, + 0.774, + 0.244 + ], + "angle": 0, + "content": "lee.j@navercorp.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.295, + 0.313, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.474, + 0.719 + ], + "angle": 0, + "content": "Compositional zero-shot learning (CZSL) aims to recognize unseen compositions with prior knowledge of known primitives (attribute and object). Previous works for CZSL often suffer from grasping the contextuality between attribute and object, as well as the discriminability of visual features, and the long-tailed distribution of real-world compositional data. We propose a simple and scalable framework called Composition Transformer (CoT) to address these issues. CoT employs object and attribute experts in distinctive manners to generate representative embeddings, using the visual network hierarchically. The object expert extracts representative object embeddings from the final layer in a bottom-up manner, while the attribute expert makes attribute embeddings in a top-down manner with a proposed object-guided attention module that models contextuality explicitly. To remedy biased prediction caused by imbalanced data distribution, we develop a simple minority attribute augmentation (MAA) that synthesizes virtual samples by mixing two images and oversampling minority attribute classes. Our method achieves SoTA performance on several benchmarks, including MIT-States, C-GQA, and VAW-CZSL. We also demonstrate the effectiveness of CoT in improving visual discrimination and addressing the model bias from the imbalanced data distribution. The code is available at https://github.com/HanjaeKim98/CoT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.47, + 0.827 + ], + "angle": 0, + "content": "Humans perceive entities as hierarchies of parts; for example, we recognize 'Cute Cat' by composing the meaning of 'Cute' and 'Cat'. People can even perceive new concepts by composing the primitive meanings they already" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.281, + 0.707, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.295, + 0.889, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.394, + 0.706, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.408, + 0.889, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.497, + 0.643, + 0.509 + ], + "angle": 0, + "content": "(a)OADis" + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.497, + 0.843, + 0.509 + ], + "angle": 0, + "content": "(b) CoT (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.598 + ], + "angle": 0, + "content": "Figure 1: Visual feature distribution of previous state-of-the-art approach (OADis) [53] and CoT (Ours). The features are gathered with respect to attribute 'Blue -' (Top) and object ' - Wall' (Bottom). Notably, the previous method suffers from degraded visual discrimination such as ('Blue, Jeans' vs. 'Blue, Shirt'), or all attributes composed into object 'Wall'." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.689 + ], + "angle": 0, + "content": "knew. Such compositionality is a fundamental ability of humans for all cognition. For this reason, compositional zero-shot learning (CZSL), recognizing a novel composition of known components (i.e., object and attribute), has been regarded as a crucial problem in the research community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "A naive approach for CZSL is to combine attribute and object scores estimated from individually trained classifiers [39, 21]. However, it is hard for these independent predictions to consider the interactions that result from a combination of different primitives, which is contextuality (e.g. different meaning of 'Old' in 'Old Car' and 'Old Cat') in the composition [39]. Previous approaches [42, 76, 36, 41, 53, 80] have solved the problem by modeling the compositional label (text) embedding for each class. They leverage external knowledge bases like Glove [37, 47] to extract attribute and object semantic vectors, and concatenate them with a few layers into the composition label embedding. The embeddings are then aligned with visual features in the shared embedding space, where the recognition of unseen image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.84, + 0.223, + 0.852 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.852, + 0.47, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the Institute of Information Communications Technology Planning Evaluation (IITP) grant funded by the Korean government (MSIT) (No. 2021-0-02068, Artificial Intelligence Innovation Hub)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "query becomes the nearest neighbor search problem [36]. Nevertheless, the extent to which contextuality is taken into account in the visual domain has still been restricted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.35 + ], + "angle": 0, + "content": "Prior works [52, 53] have also pointed out the importance of discriminability in visual features, which is related to generalization performance towards unseen composition recognition [62, 38]. A popular solution is disentangle-ment [52, 53, 27, 80], in which independent layers are allocated to extract intermediate visual representations of attribute and object. However, as shown in Fig. 1, it may be challenging to extract their unique characteristics to understand the heterogeneous interaction. We carefully argue that this problem arises from how the visual backbone (e.g. ResNet18 [16]) is used. This is because attribute and object features, which require completely different characteristics [30, 22], are based on the same visual representation from the deep layer of the backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Another challenge in CZSL is the long-tailed distribution of real-world compositional data [2, 63, 3]. Few attribute classes are dominantly associated with objects, which may cause a hubness problem [10, 12] among visual features. The visual features from the head (frequent) composition become a hub, which aggravates the visual features to be indistinguishable in the embedding space [12], and induces a biased prediction towards dominant composition [61, 81]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.763 + ], + "angle": 0, + "content": "In this paper, based on the discussions above, we propose Composition Transformer (CoT) to enlarge the visual discrimination for robust CZSL. Motivated by bottom-up and top-down attention [31, 32, 28], the CoT presents object and attribute experts, each forming its feature in different layers of the visual network. Specifically, the object expert generates a representative object embedding from the last layer (i.e., bottom-up pathway) that is most robust to identify object category with high-level semantics [72, 44]. Then we explicitly model contextuality through an object-guided attention module that explores intermediate layers of the backbone network and builds attribute-related features associated with the object (i.e., top-down pathway). Based on this module, the attribute expert generates a distinctive attribute embedding in conjunction with the object embedding. By utilizing all the features of each layer exhibiting different characteristics, our method comprehensively leverages a visual network to diversify the components' features in the shared embedding space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Finally, we further develop a simple minority attribute augmentation (MAA) methodology tailored for CZSL to address the biased prediction caused by imbalanced data distribution [61]. Unlike GAN-based augmentations [56, 27] that lead to overwhelming computation during training, our method simplifies the synthesis process of the virtual sample by blending two images while oversampling minority attribute classes [7, 45]. Thanks to the label smoothing effect of the balanced data distribution by augmenta" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "tion [79, 59], the CoT is well generalized with minimal computational costs, resolving the bias problem induced by the majority classes." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.138, + 0.845, + 0.152 + ], + "angle": 0, + "content": "Our contributions are summarized in three-folds:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.156, + 0.892, + 0.232 + ], + "angle": 0, + "content": "- To enhance visual discrimination and contextuality, we propose a Composition Transformer (CoT). In this framework, object and attribute experts hierarchically estimate primitive embeddings by fully utilizing intermediate outputs of the visual network." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.237, + 0.892, + 0.282 + ], + "angle": 0, + "content": "- We introduce a simple yet robust MAA that alleviates dominant prediction on head compositions with majority attributes." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.287, + 0.892, + 0.348 + ], + "angle": 0, + "content": "- The remarkable experimental results demonstrate that our CoT and MAA are harmonized to improve the performance on several CZSL benchmarks, showing state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.156, + 0.892, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.642, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Given a set of objects and their associated attributes, compositional zero-shot learning (CZSL) aims to recognize unseen compositions using primitive knowledge. Contrary to conventional methods [34] that recognize the object and attribute independently, recent approaches [39, 42, 29] have regarded contextuality as a crucial standpoint for understanding the interpretation of the attribute depending on the object class. For example, AttrOps [42] represented the attribute as a linear transformation of the object's state, and SymNets [29] regularized visual features to satisfy the group axioms with auxiliary losses. Some works [41, 75] have leveraged graph networks to model the global dependencies among primitives and their compositions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.854 + ], + "angle": 0, + "content": "However, those methods have often suffered in distinguishing visual composition features. Meta-learning [13] has been utilized for CZSL frameworks [49, 67] to enhance the discriminative power of visual features but induce high training costs. Other works [52, 27, 53] have disentangled representations of objects and attributes for modeling visual composition. Those methods have shown promising generalization performance than conventional methods. However, the aforementioned approaches have often failed to discriminate the minority classes in object and attribute, which is caused by severe model bias toward seen compositions [61, 81]. They have used the mostly-shared visual backbone for attribute and object representations, which needs more deep concern to be extracted differently according to its characteristics. In specific, the attribute feature is changed in a wide variety of aspects as the property of the object changes, and the object feature must be perceived as semantically the same despite changing the attribute." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method is also closely related to visual attribute prediction [46, 26], where the goal is to recognize visual attributes describing states or properties within an object." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.889, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.893, + 0.275 + ], + "angle": 0, + "content": "Figure 2: The overall architecture of composition transformer. Object and attribute experts produce object- and attribute-specific features at different layers. An object-guided attention module is attached to each attribute experts to highlight attribute-specific features. Object-attribute embeddings are composed, and then projected to a joint embedding space with label embedding for optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.473, + 0.468 + ], + "angle": 0, + "content": "Since the fine-grained local information (e.g. color, shape, and style) is required to the task, several approaches [55, 60, 20, 58, 48] utilize low-level features in backbone network, which contain textual details and objects' contour information [74, 69]. Meanwhile, with a similar motivation, some methods for CZSL have integrated low-level features with a multi-scale fusion approach to acquire robust visual features [70, 75]. Although a simple integrated feature can recover spatial details slightly, it is hard to localize the RoI of the attribute that varies with the object for compositional recognition; for example, a puddle in an object 'Street' for a composition class 'Wet Street'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.473, + 0.833 + ], + "angle": 0, + "content": "Moreover, attributes in generic objects are long-tailed distributed in nature [55, 48]. The imbalance issue occurs as a biased recognition problem, because deep networks tend to be over-confident in head composition classes [54, 65]. This phenomenon is also known as a hubness problem [10, 12], where few visual features from head composition classes become indistinguishable from many tail composition samples. One major solution for the data imbalance problem is the resampling [15, 57, 14, 35] by flattening the long-tailed dataset by over-sampling of minority classes or under-sampling of majority classes. However, directly applying such sampling strategies in CZSL could intensify the overfitting problem on over-sampled classes, and discard valuable samples for learning the alignment between visual and semantic (linguistic) features in the joint embedding space [81]. Another line of work is generation methods [68, 66] that hallucinate a new sample from rare classes but they require complicated training strategies for data generation. In this work, we are inspired by recent data augmentation techniques [77, 7, 45, 79, 8], and combine both resampling and generation approaches to tailor for CZSL, which facilitates the regularization of robust embedding space even in the minority classes via augmented virtual samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.382, + 0.863 + ], + "angle": 0, + "content": "3. Composition Transformer (CoT)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In compositional zero-shot learning (CZSL), given an object set \\(\\mathcal{O}\\) and an attribute set \\(\\mathcal{A}\\), object-attribute" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.893, + 0.438 + ], + "angle": 0, + "content": "pairs comprise a composition set \\(\\mathcal{V} = \\mathcal{A}\\times \\mathcal{O}\\). The network is trained with an observable (seen) composition set \\(\\mathcal{D}_{\\mathrm{tr}} = \\{\\mathcal{X}_{\\mathrm{s}},\\mathcal{Y}_{\\mathrm{s}}\\}\\), where \\(\\mathcal{X}_{\\mathrm{s}}\\) is an image set corresponding to a composition label set \\(\\mathcal{V}_{\\mathrm{s}}\\). Following the generalized zero-shot setting [49, 5, 73], a test set \\(\\mathcal{D}_{\\mathrm{te}} = \\{\\mathcal{X}_{\\mathrm{te}},\\mathcal{Y}_{\\mathrm{te}}\\}\\) consists of samples drawn from both seen and unseen composition classes, i.e. \\(\\mathcal{V}_{\\mathrm{te}} = \\mathcal{V}_{\\mathrm{s}}\\cup \\mathcal{V}_{\\mathrm{u}}\\) and \\(\\mathcal{V}_{\\mathrm{s}}\\cap \\mathcal{V}_{\\mathrm{u}} = \\emptyset\\). For this task, a powerful visual network is required to extract distinct features of objects and attributes respectively, in order to recognize both seen and unseen compositions well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.62 + ], + "angle": 0, + "content": "To this end, we propose a novel composition transformer (CoT) that includes object and attribute expert modules as illustrated in Fig. 2. Our CoT is basically built on the vision transformer (i.e. ViT [11]) and fully utilizes intermediate blocks to capture the attribute and object representations. These representations are projected into the joint embedding space to calculate the similarity between given compositional labels and images. We also introduce a minority attribute sampling method for data augmentation such that the biased recognition with head composition class is alleviated by label-smoothing regularization effect [40]. Next, we describe in detail the configurations of CoT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.642, + 0.645 + ], + "angle": 0, + "content": "3.1. Object expert" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.894, + 0.756 + ], + "angle": 0, + "content": "The CoT is composed of a sequence of transformer blocks including multi-head self-attention and MLP layers. Generally, as the block becomes deeper, class discrimination capability is increased while being invariant from the attribute deformation [78, 43]. Therefore, we select the final output from the last block to an input of the object expert in a bottom-up manner [31, 1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.757, + 0.894, + 0.834 + ], + "angle": 0, + "content": "Formally, let us denote \\(\\{\\mathbf{z}^n\\}_{n=1}^N\\) as the output features of the blocks in CoT. The feature of \\(n\\)-th transformer blocks, \\(\\mathbf{z}^n\\), consists of a class token (i.e., [CLS]) \\(\\mathbf{z}_c^n \\in \\mathbb{R}^{1 \\times D}\\) and a set of patch tokens \\(\\mathbf{z}_p^n \\in \\mathbb{R}^{HW \\times D}\\). The object expert \\(E_o\\) encodes the last feature to present an object embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.843, + 0.893, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {o} = E _ {o} \\left(\\mathbf {z} _ {c} ^ {N}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(E_{o}\\) consists of a fully-connected (FC) layer. To optimize the parameters in the object expert, we compute a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5677" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "cross-entropy loss with an object classifier whose weights are initialized from object word-embeddings as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.132, + 0.469, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o b j}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {o} , w \\left(o _ {i}\\right)\\right) / \\tau_ {o} \\right\\}}{\\sum_ {o _ {k} \\in \\mathcal {O}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {o} , w \\left(o _ {k}\\right)\\right) / \\tau_ {o} \\right\\}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.178, + 0.469, + 0.209 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature parameter, \\(w(o)\\) denotes the word-embedding [47] corresponding to the object \\(o\\in \\mathcal{O}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.218, + 0.24, + 0.234 + ], + "angle": 0, + "content": "3.2. Attribute expert" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.241, + 0.469, + 0.437 + ], + "angle": 0, + "content": "Capturing both attribute and object in the same deeply condensed feature is nontrivial as they require pretty different visual representations [30, 22]. The straightforward solution is to design two-stream networks without any parameter sharing. However, such individual training could not capture the contextuality [39]. Hence, we build an attribute expert which utilizes semantics estimated from the object expert as a condition, and extracts attribute-related features on intermediate layers of visual backbone in a top-down fashion [31, 1]. We devise an object-guided attention module that leverages convolution kernel-based attention [71, 6, 24], which excels at capturing fine-grained local information required for attribute recognition [18]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.438, + 0.469, + 0.545 + ], + "angle": 0, + "content": "Specifically, we first reshape \\(\\mathbf{z}_p^n\\) to \\(\\mathbf{Z}_p^n\\) with output resolution of \\(H\\times W\\times D\\). Following [51], we tile \\(\\mathbf{p}_o\\) to match the spatial dimension with \\(\\mathbf{Z}_p^n\\) and concatenate them to obtain an object-contextualized feature \\(\\{\\mathbf{U}^n = [\\mathbf{p}_o,\\mathbf{Z}_p^n ]\\} \\in \\mathbb{R}^{H\\times W\\times 2D}\\). As shown in Fig. 2, we generate a channel attention map \\(\\mathbf{A}_c^n\\in \\mathbb{R}^{1\\times 1\\times D}\\) from the averaged feature across the spatial axis:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.556, + 0.469, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {c} ^ {n} = \\sigma \\left(\\operatorname {c o n v} _ {1 \\times 1} \\left(\\operatorname {a v g p o o l} _ {\\text {s p a t i a l}} \\left(\\mathbf {U} ^ {n}\\right)\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.585, + 0.469, + 0.629 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes a sigmoid function. We also generate a spatial attention map \\(\\mathbf{A}_s^n\\in \\mathbb{R}^{H\\times W\\times 1}\\) from the averaged feature across the channel axis:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.642, + 0.469, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {s} ^ {n} = \\sigma \\left(\\operatorname {c o n v} _ {3 \\times 3} \\left(\\operatorname {a v g p o o l} _ {\\text {c h a n n e l}} \\left(\\mathbf {U} ^ {n}\\right)\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.67, + 0.469, + 0.7 + ], + "angle": 0, + "content": "Finally, we aggregate two attention maps into a unified attention map \\(\\mathbf{A}^n\\in \\mathbb{R}^{H\\times W\\times D}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.712, + 0.469, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} ^ {n} = \\mathbf {A} _ {s} ^ {n} \\times \\mathbf {A} _ {c} ^ {n}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.74, + 0.469, + 0.785 + ], + "angle": 0, + "content": "The output of our object-guided attention module is defined as attribute features with estimated attention maps, followed by the residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.796, + 0.469, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {Z}} _ {p} ^ {n} = \\mathbf {Z} _ {p} ^ {n} + \\mathbf {A} ^ {n} \\odot \\mathbf {Z} _ {p} ^ {n}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.452, + 0.84 + ], + "angle": 0, + "content": "where \\(\\odot\\) is an operation for element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given \\(\\tilde{\\mathbf{Z}}_p^n\\) from each intermediate layer, we ensemble the features from different blocks by taking advantage of the hierarchical architecture [33, 43]. It is required for recognizing heterogeneous types of attributes such as colors or" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "shapes [48]. Especially, we choose low-, middle-, and high-level features from the backbone, denoted as \\(\\{\\tilde{\\mathbf{Z}}_p^l,\\tilde{\\mathbf{Z}}_p^m,\\tilde{\\mathbf{Z}}_p^h\\}\\), and ensemble those triplets with global average pooling (GAP) followed by concatenation. The attribute expert \\(E_{a}\\) then generates an attribute embedding with the triplets as" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.175, + 0.892, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {a} = E _ {a} \\left(\\left[ \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {l}\\right), \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {m}\\right), \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {h}\\right) \\right]\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.2, + 0.892, + 0.232 + ], + "angle": 0, + "content": "where \\(E_{a}\\) consists of a FC layer. Similar to the object expert, we optimize the attribute expert with the cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.251, + 0.892, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a t t}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {a} , w \\left(a _ {i}\\right)\\right) / \\tau_ {a} \\right\\}}{\\sum_ {a _ {k} \\in \\mathcal {A}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {a} , w \\left(a _ {k}\\right)\\right) / \\tau_ {a} \\right\\}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.293, + 0.779, + 0.31 + ], + "angle": 0, + "content": "3.3. Mapping visual-semantic space" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.892, + 0.359 + ], + "angle": 0, + "content": "The visual compositional embedding \\(\\mathbf{p}_x\\) of the input image \\(x\\) is calculated from object and attribute experts, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.361, + 0.892, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {x} = E _ {c} \\left(\\left[ \\mathbf {p} _ {o}, \\mathbf {p} _ {a} \\right]\\right)). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.473 + ], + "angle": 0, + "content": "The embedding function \\( E_{c} \\) includes an FC layer, projecting the visual features into the joint embedding space that aligns visual and semantic representations. Following previous works [42, 36], we estimate semantic label embedding \\( \\mathbf{p}_y \\) of each seen composition label \\( y \\in \\mathcal{V}_s \\) by projecting the concatenated word vectors into the joint space:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.481, + 0.892, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {y} = g ([ w (o), w (a) ]), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.506, + 0.892, + 0.567 + ], + "angle": 0, + "content": "where \\( g(\\cdot) \\) is a label embedding network that consists of 3 FC layers and ReLU activation function. During training, we minimize the cosine distance of visual and semantic embeddings with cross-entropy loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.575, + 0.892, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o m p}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {x} , \\mathbf {s} _ {y}\\right) / \\tau_ {c} \\right\\}}{\\sum_ {y _ {k} \\in \\mathcal {Y} _ {s}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {x} , \\mathbf {s} _ {y _ {k}}\\right)\\right) / \\tau_ {c} \\}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.617, + 0.892, + 0.66 + ], + "angle": 0, + "content": "where \\(\\tau_{c}\\) is a temperature parameter for the composition loss \\(\\mathcal{L}_{\\mathrm{comp}}\\). Finally, the total objectives of CoT is then formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.671, + 0.892, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o m p}} + \\alpha \\mathcal {L} _ {\\text {a t t}} + \\beta \\mathcal {L} _ {\\text {o b j}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.696, + 0.738, + 0.71 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are balance weights." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.711, + 0.892, + 0.83 + ], + "angle": 0, + "content": "During inference, we measure the cosine similarity between a visual embeddings and all label embeddings from \\(\\mathcal{V}_{\\mathrm{te}}\\) and regard it as a feasibility score of the image and composition labels [36]. Following [80], we also use the classification scores from object and attribute experts as a form of cosine similarity. Therefore, a final feasibility score of label \\(y = (a,o)\\) is derived by adding the above three scores, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.84, + 0.892, + 0.871 + ], + "angle": 0, + "content": "\\[\nc (y) = \\cos \\left(\\mathbf {p} _ {x}, \\mathbf {s} _ {y}\\right) + \\cos \\left(\\mathbf {p} _ {o}, w (o)\\right) + \\cos \\left(\\mathbf {p} _ {a}, w (a)\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where we predict the label with the highest score as the final composition label." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.27, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.091, + 0.462, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 3: Attribute class distributions on VAW-CZSL. Every attributes in each plot are composed into a fixed object; (a) Grass and (b) Player." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.299, + 0.366, + 0.316 + ], + "angle": 0, + "content": "3.4. Minority attribute augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.468, + 0.459 + ], + "angle": 0, + "content": "With a carefully designed architecture, our CoT can generate distinctive visual features for CZSL. Nevertheless, the imbalanced composition samples in the training dataset restrict the performance since attributes often co-occur with certain objects in the real world [30]. For example in Fig. 3, 'Green Grass' is naturally more common than 'White Grass'. Such long-tailed data makes the deep models be over-confident in the composition with head attributes [81], and thus hurts the generalization on unseen compositions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.626 + ], + "angle": 0, + "content": "To this end, we present a minority attribute augmentation (MAA) for training, which diversifies the minority composition. First, two compositional embeddings from COt are selected as inputs to the augmentation, where \\(\\mathbf{p}_{x_A}\\) and \\(\\mathbf{p}_{x_B}\\) share the same object class. Especially, to balance the major and minor attributes, we assign higher sampling weights on minority attributes with the weighted sampler [14]. Formally, let \\(\\zeta_{o_i}\\) denote the number of attribute samples in the \\(i\\)-th attribute class composed on an object class \\(o\\). The sampling weight of the attribute-object pair \\(\\kappa(a_i,o)\\) is formulated by an inverse attribute class frequency according to:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.636, + 0.47, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\kappa \\left(a _ {i}, o\\right) = \\frac {1 / \\zeta_ {o _ {i}}}{\\sum_ {i} 1 / \\zeta_ {o _ {i}}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.47, + 0.721 + ], + "angle": 0, + "content": "Inspired by Mixup [79, 64], we generate a virtual sample \\(\\mathbf{p}_{x_M}\\) with alpha compositing and its attribute semantic embedding \\(w(a_{M})\\) by linear combination:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.734, + 0.365, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {x _ {M}} = \\lambda \\mathbf {p} _ {x _ {A}} + (1 - \\lambda) \\mathbf {p} _ {x _ {B}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.752, + 0.469, + 0.769 + ], + "angle": 0, + "content": "\\[\nw \\left(a _ {M}\\right) = \\lambda w \\left(a _ {A}\\right) + (1 - \\lambda) w \\left(a _ {B}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.468, + 0.795 + ], + "angle": 0, + "content": "The mixing factor \\(\\lambda\\) is sampled from a uniform distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We notice that there are two significant differences from Mixup: (1) Compared to Mixup sampling arbitrary images, MAA constructs a composition pair fastening an object with the minority aware sampling to address the imbalanced data issue in CZSL. (2) Mixup mixes one-hot vector labels [79, 77] for classification, while MAA generates a virtual label in word vector space, to jointly regularize vi" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "sual and linguistic (label) embedding functions against the majority attribute classes with Eq. (12)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.132, + 0.764, + 0.147 + ], + "angle": 0, + "content": "3.5. Extension to other backbones" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.894, + 0.351 + ], + "angle": 0, + "content": "Our idea of CoT is generic and scalable, so we can simply extend to other backbones such as ResNet-18 (R18) [16] by selecting specific layers from the backbone composed of multiple layers or blocks. For a fair comparison to prior arts [41, 53, 23], we study the impact of our contribution on the ResNet backbone based on the convolutional operation. In specific, the object expert generates the object embedding from the last fifth convolution block. The classification token is replaced by a global average pooled feature. Similar to using ViT [11], we concatenate the intermediate layers into a feature vector for the attribute expert. The remained settings such as the architecture of \\( E_{\\mathrm{comp}} \\) and \\( g(\\cdot) \\) are the same as ViT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.355, + 0.634, + 0.371 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.379, + 0.688, + 0.396 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Datasets. MIT-States [19] dataset is a common dataset in CZSL that most previous methods adopted, containing 115 states, 245 objects, and 1962 compositions. C-GQA [41] and VAW-CZSL [53] are two brand-new and large-scale benchmarks collected from in-the-wild images with more general compositions, such as 'Stained Wall' and 'Hazy Mountain'. Detailed data statistics including long-tailed distribution can be found in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Metrics. We follow common CZSL setting [49], testing performance on both seen and unseen pairs [5]. Under the setting, we report area under the curve (AUC) (\\%) between seen and unseen accuracies at different bias term in validation and test splits. We also compute the best seen and unseen accuracies in the curve, and calculate the best harmonic mean to measure the trade-off between the two values. Lastly, we follow [41] and report the attribute and object accuracies on unseen labels to measure the discrimination power of each representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our CoT with 9 recent CZSL approaches: AoP [42], LE+ [39], TMN [49], Symnet [29], CompCos [36], CGE [41], SCEN [27], OADis [53], CAPE [23]. All baselines were built on pretrained ResNet-18 (R18) [16], with pre-trained word embedding such as GloVe [47], FastText [4], and Word2Vec [37]. To show the contribution of our proposed elements, we carry on the most recent methods including CGE and OADis with a transformer-based backbone (ViT-B) [11]. We also clarify that CGE and OADis utilize the knowledge of unseen composition labels at training for building a relation graph and hallucinating virtual samples respectively, while other methods do not assess the unseen composition label sets. The results of other baselines were obtained from [41, 27, 23, 53] or re-implemented from their official code." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5679" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.897, + 0.389 + ], + "angle": 0, + "content": "
MethodsBackboneMIT-StatesC-GQAVAW-CZSL
V@1T@1HMSUV@1T@1HMSUV@3T@3HMSU
With frozen R18
AoP [42]R182.52.010.716.618.40.90.32.911.83.91.41.49.116.411.7
LE+ [39]R183.52.311.516.221.21.20.65.316.15.01.51.69.816.213.2
TMN [49]R183.32.611.822.717.12.21.17.721.66.32.22.311.919.915.4
Symnet [29]R184.53.413.824.820.03.31.89.825.29.22.32.312.219.115.8
CompCos [36]R186.94.816.926.924.53.52.612.128.111.83.13.214.223.918.0
CGE* [41]R187.25.318.128.925.03.62.511.927.511.72.72.913.023.416.8
SCEN [27]R187.25.318.429.925.24.02.912.428.912.1-----
OADis* [53]R187.65.918.931.125.6-----3.53.615.224.918.7
CAPE [23]R18-5.819.130.526.2-4.216.332.915.6-----
CoTR187.76.219.630.826.84.94.516.633.116.63.63.815.724.619.1
With frozen ViT-B
CGE* [41]ViT-B8.77.321.333.528.64.53.815.631.314.43.73.915.726.319.7
OADis* [53]ViT-B9.07.521.934.229.35.94.616.232.515.34.14.316.626.920.8
CoTViT-B9.27.823.234.831.56.75.117.534.018.84.44.717.726.922.2
With finetuning ViT-B
CGE* [41]ViT-B11.49.724.839.731.67.35.418.538.017.15.96.220.130.125.7
OADis* [53]ViT-B11.810.125.239.232.18.17.020.138.319.86.26.520.431.326.1
CoTViT-B12.110.525.839.533.08.77.422.139.222.76.67.221.732.928.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.393, + 0.895, + 0.45 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison with prior arts. Following [41, 53], we measure AUC values on validation (V) and test (T) datasets, at top-1 (@1) on C-GQA and top-3 (@3) on VAW-CZSL, respectively. In addition, we report harmonic mean (HM), seen (S), unseen (US), attribute (A), and object (O) accuracies. * denotes the models accessing unseen composition labels at training. The best result and the second best result are boldfaced and underlined, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.475, + 0.292, + 0.491 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.471, + 0.744 + ], + "angle": 0, + "content": "Models. We use ViT-B [11] and ResNet18 [16] pre-trained on ImageNet [9] as a visual backbone. In ViT-B backbone, Each patch size is \\( 16 \\times 16 \\), and the number of patch tokens in each layer is \\( 14 \\times 14 \\), having 768 channels per token. The label embedding network \\( g \\) is composed of 3 FC layers, two dropout layers with ReLU activation functions, where hidden dimensions are 900. We use one FC layer for \\( E_{o} \\) and \\( E_{a} \\) to produce 300-dimensional prototypical vectors that are matched with GloVe [47] embedding vectors from composition labels. Therefore, both \\( E_{c} \\) and \\( g(\\cdot) \\) produce 300-dimensional embedding vectors, with \\( l_{2} \\) normalization. In the attribute expert, we use the outputs of the 3rd, 6th and 9th blocks (denoted \\( \\mathbf{z}^{l}, \\mathbf{z}^{m}, \\mathbf{z}^{h} \\) respectively) for multi-level feature fusion. For ResNet18, we use the 2nd, 3rd and 4th blocks as \\( \\mathbf{z}^{l}, \\mathbf{z}^{m}, \\mathbf{z}^{h} \\), resulting in a 438-dimensional feature vector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Training setup. We train CoT with Adam optimizer [25]. The input images are augmented with random crop and horizontal flip, and finally resized into \\(224 \\times 224\\). In ViT, learning rate is initialized to 1e-4 for all three datasets, and decayed by a factor of 0.1 for 10 epochs. In R18, the initial learning rate is 5e-5 with the same decay parameter of the ViT setting. Note that we freeze the GloVe [47] word embedding \\(w\\) for fair comparisons. For all datasets, we use the same temperature parameters [36, 53, 27] of each cross-entropy loss, \\(\\tau_{c}\\), \\(\\tau_{o}\\), and \\(\\tau_{a}\\) as 0.05, 0.01 and 0.01 respec" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.473, + 0.879, + 0.564 + ], + "angle": 0, + "content": "
MethodsMIT-StatesC-GQAVAW-CZSL
Attr.Obj.Attr.Obj.Attr.Obj.
CGE [41]35.744.417.534.624.551.6
OADis [53]35.245.214.742.022.353.8
CoT37.346.019.840.227.354.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.57, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Table 2: Comparison of unseen attribute (Attr.) and object (Obj.) accuracies with the latest CZSL methods [41, 53]. All results are reported on fine-tuned ViT backbone." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.639, + 0.892, + 0.701 + ], + "angle": 0, + "content": "tively. We use the loss balance weights \\(\\alpha\\) and \\(\\beta\\) as (0.5, 0.5) on C-GQA, and (0.4, 0.6) on VAW-CZSL and MIT-States. We note that to prevent overfitting of COT we apply MAA from 15 and 30 epochs to ViT and R18, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.663, + 0.725 + ], + "angle": 0, + "content": "4.3. Main evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Table 1 compares generalization performance between our methods and the baselines with ResNet18 [16] (R18) and ViT-B [11] backbone setting. To further compare visual discrimination, we also compare unseen attribute and object accuracies of our methods and the latest CZSL methods [41, 53] in Table 2. In the following, we analyze the results from the three datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "MIT-States. From Table 1, the CoT achieves the best AUC values (V@1 and H@1) and harmonic mean (HM) with both R18 and ViT backbone. These results sufficiently demonstrate our experts' applicability on the CNN-based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5680" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.099, + 0.344, + 0.178 + ], + "angle": 0, + "content": "
CoTMAAAUCHM
5.8118.1
✓6.0519.0
✓6.9320.8
✓✓7.2021.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.182, + 0.326, + 0.194 + ], + "angle": 0, + "content": "(a) Component analysis." + }, + { + "type": "table", + "bbox": [ + 0.388, + 0.099, + 0.574, + 0.178 + ], + "angle": 0, + "content": "
Loss componentAUCHM
Lcomp5.9418.9
Lcomp+Lobj6.2219.6
Lcomp+Lattr6.9121.4
Lcomp+Lattr+Lobj7.2021.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.402, + 0.182, + 0.555, + 0.194 + ], + "angle": 0, + "content": "(b) Impact of loss functions." + }, + { + "type": "table", + "bbox": [ + 0.615, + 0.099, + 0.811, + 0.178 + ], + "angle": 0, + "content": "
AugmentationAUCHM
Manifold Mixup [64]5.9318.8
CutMix [77]6.9520.7
Mixup [79]7.0421.0
MAA7.2021.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.621, + 0.182, + 0.799, + 0.194 + ], + "angle": 0, + "content": "(c) Impact of data augmentation." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.206, + 0.453, + 0.274 + ], + "angle": 0, + "content": "
GuidanceAUCHMSUAO
w/o guidance6.720.530.725.822.553.4
Attribute guidance6.921.031.526.425.753.9
Object guidance7.221.732.928.227.354.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.277, + 0.379, + 0.29 + ], + "angle": 0, + "content": "(d) Impact of object guidance." + }, + { + "type": "table", + "bbox": [ + 0.472, + 0.202, + 0.844, + 0.281 + ], + "angle": 0, + "content": "
EnsemblezlzmzhAUCHMSUAO
Low2346.620.730.328.725.852.3
Mid5676.921.231.929.725.255.9
High910116.821.433.426.624.154.5
Mixture3697.221.732.928.227.354.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.468, + 0.284, + 0.84, + 0.297 + ], + "angle": 0, + "content": "(e) Different configurations of intermediate blocks in attribute experts." + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.304, + 0.872, + 0.318 + ], + "angle": 0, + "content": "Table 3: Ablation experiments on VAW-CZSL dataset. Note that all results are reported on CoT (ViT-B), with a fine-tuning setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.47, + 0.435 + ], + "angle": 0, + "content": "backbone. With a finetuned ViT-B, the CoT obtains the best validation AUC of \\(12.1\\%\\) and test AUC of \\(10.5\\%\\), surpassing previous CZSL methods with a large margin. Our method also improves attribute (Attr.) and object (Obj.) accuracies as in Table 2, showing the effectiveness of CoT for visual discrimination." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.744 + ], + "angle": 0, + "content": "C-GQA. Similar trends can be observed in the C-GQA dataset. In the frozen R18 setting, the CoT achieves the best test AUC (T@1) of \\(4.5\\%\\), comparable with previous state-of-the-art CAPE. Although the T@1 scores from CGE and OADis are also increased upon replacing the backbone from R18 to ViT-B, our CoT with ViT-B outperforms OADis with larger margins in all metrics, as compared to the performance with R18. It is noticeable that our proposed augmentation technique contributes the unbiased prediction, improving unseen accuracies into \\(18.8\\%\\) with a minority-aware sampling. The performance regarding AUC is improved with end-to-end training with backbone, \\(2.0\\%\\) in V@1 and \\(2.3\\%\\) in T@1. Of note, OADis and CGE access unseen composition labels at training, so it is unfair to directly compare with other methods including CoT. Nevertheless, more surprisingly, our CoT shows comparable and superior performance to OADis. Finally, in Table 2, our model also improves the attribute accuracy by achieving \\(19.8\\%\\), having comparable \\(40.2\\%\\) object accuracy compared to OADis and CGE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "VAW-CZSL. We also achieve SoTA performance on VAW-CZSL, the most challenging benchmark having almost 10K composition labels for evaluation. Even if we do not use additional unseen composition information like OADis, our method achieves the highest T@1 of \\(7.2\\%\\) with the best unseen label accuracy of \\(28.2\\%\\). We also obtain both the highest attribute and object score (Table 2), especially with a significant improvement in attribute prediction by \\(2.8\\%\\) compared to CGE. It proves that the careful modeling of the visual network in CoT is essential to substantially improve" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.391 + ], + "angle": 0, + "content": "the generalization capability. In addition, the results suggest that finetuning the visual backbone to suit composition recognition improves the performance in all benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.404, + 0.65, + 0.42 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.474 + ], + "angle": 0, + "content": "We ablate our CoT and MAA with different design choices on Table 3. We use the VAW-CZSL dataset and fine-tuned ViT-B backbone as our main ablation setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Component analysis. We report ablation results of CoT and MAA in Table 3a. In 'without CoT' case, we use CompCos [36] baseline. CoT boosts the AUC and HM significantly about \\(1.1\\%\\) and \\(2.3\\%\\) compared to the baseline. MAA offers better results to both CoT and baseline, showing scalability to other methods. Notably, employing MAA on CoT gives higher gains of \\(0.3\\%\\) AUC and \\(2.3\\%\\) HM compared to the baseline, demonstrating that the two components create a synergy effect." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Impact of loss functions. Table 3b shows the impact of each loss function. Compared to the case using \\(\\mathcal{L}_{\\mathrm{comp}}\\) only, \\(\\mathcal{L}_{\\mathrm{obj}}\\) and \\(\\mathcal{L}_{\\mathrm{attr}}\\) bring a gain on both AUC and HM. Using both losses to \\(\\mathcal{L}_{\\mathrm{comp}}\\) improves AUC and HM significantly by \\(1.26\\%\\) and \\(2.8\\%\\) respectively, showing the importance of object and attribute losses to learn each expert." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Different data augmentations. In Table 3c, we compare MAA with various data mixing methodologies including Manifold Mixup [64], CutMix [77] and Mixup [79]. To apply Manifold Mixup in CoT, we select 3,6,9 and the last layer as a mixup layer, and mix these intermediate features. MAA achieves the highest score on both AUC and HM, outperforming other augmentation methods including standard Mixup. This demonstrates the effectiveness of MAA which balances the data distribution with a composition pair. Interestingly, we found that Manifold Mixup performs at worst, even degrading pure CoT performance (AUC: 6.93, HM: 20.4) in Table 3a. We speculate that the mixed representation at intermediate layers can lead to significant underfit" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5681" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.227, + 0.47, + 0.324 + ], + "angle": 0, + "content": "Figure 4: Distribution of k-occurrence counts with \\(k = 20\\) of C-GQA and VAW-CZSL test set. All results are reported with ViT-B backbone. x-axis stands for composition class ids which are ordered by decreasing count numbers. For better visualization, we show the top \\(10\\%\\) head composition classes where the hubness is predominant. Circles at start points indicate the most frequent hubs. The result on MIT-States is in the Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.267, + 0.366 + ], + "angle": 0, + "content": "ting of the attention module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.469, + 0.457 + ], + "angle": 0, + "content": "Impact of object guidance. As shown in Table 3d, the object guidance significantly boosts the performance. We found that reversing the guidance (attribute guidance) decreases the performance of AUC by \\(0.3\\%\\) and HM by \\(0.7\\%\\). The results validate the usage of object guidance for contextualized attribute prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Design choice of feature ensemble. Table 3e shows the ablations with different configurations of intermediate blocks used in the attribute expert. We evaluate four types of block feature ensembles: low, mid, high, and mixture. Among them, the mixture type achieves the best performance on AUC, harmonic mean, and attribute accuracy. We therefore choose this mixture ensemble as a default setting. Another finding is low-level feature is more adequate for attribute and unseen composition recognition, while high-level feature is robust to object and seen composition accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.194, + 0.649 + ], + "angle": 0, + "content": "4.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Mitigating hubness problem. We illustrate the hubness effect of a baseline [36] and our method on C-GQA and VAW-CZSL datasets in Fig. 4. Concretely, we count \\( k \\)-occurrences \\( (N_{k}) \\), the number of times that a sample appears in the \\( k \\) nearest neighbors of other points [50] for all class ids, and calculate the skewness of \\( N_{k} \\) distribution for hubness measurement. We observe that CoT significantly alleviates the hubness problem of baseline across datasets, enlarging the visual discrimination in the embedding space. MAA further reduces the skewness of the distribution by mitigating biased prediction on head composition classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Object-guided attention visualization. Qualitatively, we visualize attention maps from the object-guided attention module in Fig. 5. We observe that the attention module could capture the contextualized meaning of attributes in composition, e.g., wet skin of horse vs. puddle in street. It" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.183, + 0.888, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Figure 5: Visualization of object-guided attention on VAW-CZSL dataset. Each row represents the attended regions in image pairs, which are related to different attributes contextualized with a specific object. More results in the Appendix C.5." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.369, + 0.89, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.47, + 0.641, + 0.48 + ], + "angle": 0, + "content": "Graite Ground" + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.482, + 0.56, + 0.491 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.492, + 0.637, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.544, + 0.56, + 0.554 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.492, + 0.76, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.544, + 0.69, + 0.553 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.492, + 0.884, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.544, + 0.814, + 0.553 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.562, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Figure 6: Ground-truth composition labels and top-3 prediction results on MIT-States, C-GQA and VAW-CZSL datasets. More results in the Appendix C.6." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.65, + 0.892, + 0.695 + ], + "angle": 0, + "content": "validates that the attention module assists CoT to extract distinctive attribute embedding for composition by localizing specific RoI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Qualitative results. Fig. 6 illustrates the qualitative results of CoT. Even though some predictions do not correspond to annotated image labels, they explain the composition in another view; e.g., 'Furry Dog' and 'Shaggy Dog' in Fig. 6 (i). Moreover, in-the-wild images usually have multiple objects with their own attributes; e.g., 'White Wall' and 'Reading Man' in Fig. 6 (h). To recognize such compositions in the real world, it is necessary to handle the multi-label nature of composition, while distinguishing each attribute-object interaction [17] in an image. It encourages us to construct a new CZSL benchmark along with a novel evaluation metric handling both multi-label prediction and multiple compositions in the image. We leave it as a future work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5682" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.196, + 0.106 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.112, + 0.47, + 0.263 + ], + "angle": 0, + "content": "In this paper, we propose Composition Transformer (CoT) with simple minority attribute augmentation (MAA), a contextual, discriminative, and unbiased CZSL framework. The proposed object and attribute experts hierarchically utilize the entire powerful visual backbone to generate a composition embedding. The simple but effective MAA balances the long-tail distributed composition labels. Extensive studies with comprehensive analysis demonstrate the effectiveness of each component, and show that CoT surpasses existing methods across several benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.471, + 0.329 + ], + "angle": 0, + "content": "Acknowledgement. This research was supported by the Yonsei Signature Research Cluster Program of 2022 (2022-22-0002) and the KIST Institutional Program (Project No.2E31051-21-203)." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.345, + 0.175, + 0.36 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.37, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.485, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[3] Yuval Atzmon, Felix Kreuk, Uri Shalit, and Gal Chechik. A causal view of compositional zero-shot recognition. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.528, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[4] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. TACL, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.572, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[5] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In ECCV, 2016.3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.629, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[6] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.7, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[7] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.744, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[8] Yu-Ying Chou, Hsuan-Tien Lin, and Tyng-Luh Liu. Adaptive and generative zero-shot learning. In International conference on learning representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.787, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[10] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.37, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[12] Nanyi Fei, Yizhao Gao, Zhiwu Lu, and Tao Xiang. Z-score normalization, hubness, and few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 142-151, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[14] Yonatan Geifman and Ran El-Yaniv. Deep active learning over the long tail. arXiv preprint arXiv:1711.00941, 2017. 3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE TKDE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[17] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In ECCV, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[18] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[20] Mahdi M Kalayeh and Mubarak Shah. On symbiosis of attribute prediction and semantic segmentation. IEEE TPAMI, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[21] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Revisiting visual product for compositional zero-shot learning. In NeurIPSW, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[22] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[23] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Alain Pagani, Didier Stricker, and Muhammad Zeshan Afzal. Learning attention propagation for compositional zero-shot learning. arXiv preprint arXiv:2210.11557, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[24] Hanjae Kim, Sunghun Joung, Ig-Jae Kim, and Kwanghoon Sohn. Prototype-guided saliency feature learning for person search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4865-4874, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.791, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Xiangyu Li, Xu Yang, Kun Wei, Cheng Deng, and Muli Yang. Siamese contrastive embedding network for compositional zero-shot learning. In CVPR, 2022. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5683" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[29] Yong-Lu Li, Yue Xu, Xiaohan Mao, and Cewu Lu. Symmetry and group in attribute-object compositions. In CVPR, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[30] Kongming Liang, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Unifying visual attribute learning with object recognition in a multiplicative framework. IEEE TPAMI, 2018. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[34] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[35] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[36] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[38] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In CVPR, pages 12664-12673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[39] Ishan Misra, Abhinav Gupta, and Martial Hebert. From red wine to red tomato: Composition with context. In CVPR, 2017. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[40] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[41] Muhammad Ferjad Naeem, Yongqin Xian, Federico Tombari, and Zeynep Akata. Learning graph embeddings for compositional zero-shot learning. In CVPR, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[42] Tushar Nagarajan and Kristen Grauman. Attributes as operators: factorizing unseen attribute-object compositions. In ECCV, 2018. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[43] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[44] Zizheng Pan, Bohan Zhuang, Jing Liu, Haoyu He, and Jianfei Cai. Scalable vision transformers with hierarchical pool" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ing. In Proceedings of the IEEE/cvf international conference on computer vision, pages 377-386, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[45] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[46] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[47] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[48] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[49] Senthil Purushwalkam, Maximilian Nickel, Abhinav Gupta, and Marc'Aurelio Ranzato. Task-driven modular networks for zero-shot compositional learning. In ICCV, 2019. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[50] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[51] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[52] Frank Ruis, Gertjan Burghouts, and Doina Bucur. Independent prototype propagation for zero-shot compositionality. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[53] Nirat Saini, Khoi Pham, and Abhinav Shrivastava. Disentangling visual embeddings for attributes and objects. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[54] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In WACV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[55] Nikolaos Sarafianos, Xiang Xu, and Ioannis A Kakadiaris. Deep imbalanced attribute classification using visual attention aggregation. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[56] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[57] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[58] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[59] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[60] Chufeng Tang, Lu Sheng, Zhaoxiang Zhang, and Xiaolin Hu. Improving pedestrian attribute recognition with weakly-supervised multi-scale attribute-specific localization. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[61] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5684" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "bad momentum causal effect. Advances in Neural Information Processing Systems, 33:1513-1524, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.176 + ], + "angle": 0, + "content": "[62] Bin Tong, Chao Wang, Martin Klinkigt, Yoshiyuki Kobayashi, and Yuuichi Nonaka. Hierarchical disentanglement of discriminative latent features for zero-shot learning. In CVPR, pages 11467-11476, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[63] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[64] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[65] Byron C Wallace and Issa J Dahabreh. Improving class probability estimates for imbalanced data. KAIS, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[66] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[67] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.468, + 0.503 + ], + "angle": 0, + "content": "[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.468, + 0.545 + ], + "angle": 0, + "content": "[69] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.468, + 0.588 + ], + "angle": 0, + "content": "[70] Kun Wei, Muli Yang, Hao Wang, Cheng Deng, and Xianglong Liu. Adversarial fine-grained composition learning for unseen attribute-object recognition. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.468, + 0.63 + ], + "angle": 0, + "content": "[71] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.468, + 0.688 + ], + "angle": 0, + "content": "[72] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3907-3916, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[73] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.468, + 0.8 + ], + "angle": 0, + "content": "[74] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature based activation map for weakly supervised object localization. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[75] Ziwei Xu, Guangzhi Wang, Yongkang Wong, and Mohan S. Kankanhalli. Relation-aware compositional zero-shot learning for attribute-object pair recognition. IEEE TMM, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[76] Muli Yang, Cheng Deng, Junchi Yan, Xianglong Liu, and Dacheng Tao. Learning unseen concepts via hierarchical decomposition and composition. In CVPR, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[78] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[80] Tian Zhang, Kongming Liang, Ruoyi Du, Xian Sun, Zhanyu Ma, and Jun Guo. Learning invariant visual representations for compositional zero-shot learning. arXiv preprint arXiv:2206.00415, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[81] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5685" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_origin.pdf b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..042f8aae5fc5598d827dd81c7b320efdc70312fc --- /dev/null +++ b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/d892901c-ab65-4154-a39e-e041473b0ce9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ad8366e5c73cdd805ac4414caf2c176478b06e8463c859438539501599ad214 +size 1019081 diff --git a/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/full.md b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65ef685091512f71b5584344d5146cfc534195cb --- /dev/null +++ b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/full.md @@ -0,0 +1,428 @@ +# Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning + +Hanjae Kim + +Jiyoung Lee² + +Seongheon Park + +Kwanghoon Sohn $^{1,3,*}$ + +$^{1}$ Yonsei University + +$^{2}$ NAVER AI Lab + +$^{3}$ Korea Institute + +of Science and Technology (KIST) + +{incohjk,sam121796,khsohn}@yonsei.ac.kr + +lee.j@navercorp.com + +# Abstract + +Compositional zero-shot learning (CZSL) aims to recognize unseen compositions with prior knowledge of known primitives (attribute and object). Previous works for CZSL often suffer from grasping the contextuality between attribute and object, as well as the discriminability of visual features, and the long-tailed distribution of real-world compositional data. We propose a simple and scalable framework called Composition Transformer (CoT) to address these issues. CoT employs object and attribute experts in distinctive manners to generate representative embeddings, using the visual network hierarchically. The object expert extracts representative object embeddings from the final layer in a bottom-up manner, while the attribute expert makes attribute embeddings in a top-down manner with a proposed object-guided attention module that models contextuality explicitly. To remedy biased prediction caused by imbalanced data distribution, we develop a simple minority attribute augmentation (MAA) that synthesizes virtual samples by mixing two images and oversampling minority attribute classes. Our method achieves SoTA performance on several benchmarks, including MIT-States, C-GQA, and VAW-CZSL. We also demonstrate the effectiveness of CoT in improving visual discrimination and addressing the model bias from the imbalanced data distribution. The code is available at https://github.com/HanjaeKim98/CoT. + +# 1. Introduction + +Humans perceive entities as hierarchies of parts; for example, we recognize 'Cute Cat' by composing the meaning of 'Cute' and 'Cat'. People can even perceive new concepts by composing the primitive meanings they already + +![](images/8cfc7ef69ae570faf24074f64d6c08a8f23aaebbb84dbebffd51e2968bc66485.jpg) + +![](images/d75d6cbdc356857837d3947d57f3b52a8cff261348df50fa120d3b8160139a50.jpg) + +![](images/74bbbd01f18c8ba45fc28da780044b8c0f47d3cd8554f3b496409ee554e597d5.jpg) +(a)OADis + +![](images/ab0c3570a37ec99c73300541daa30a0c4021a884fb4d4755db8bdd6fd3f0c9eb.jpg) +(b) CoT (Ours) +Figure 1: Visual feature distribution of previous state-of-the-art approach (OADis) [53] and CoT (Ours). The features are gathered with respect to attribute 'Blue -' (Top) and object ' - Wall' (Bottom). Notably, the previous method suffers from degraded visual discrimination such as ('Blue, Jeans' vs. 'Blue, Shirt'), or all attributes composed into object 'Wall'. + +knew. Such compositionality is a fundamental ability of humans for all cognition. For this reason, compositional zero-shot learning (CZSL), recognizing a novel composition of known components (i.e., object and attribute), has been regarded as a crucial problem in the research community. + +A naive approach for CZSL is to combine attribute and object scores estimated from individually trained classifiers [39, 21]. However, it is hard for these independent predictions to consider the interactions that result from a combination of different primitives, which is contextuality (e.g. different meaning of 'Old' in 'Old Car' and 'Old Cat') in the composition [39]. Previous approaches [42, 76, 36, 41, 53, 80] have solved the problem by modeling the compositional label (text) embedding for each class. They leverage external knowledge bases like Glove [37, 47] to extract attribute and object semantic vectors, and concatenate them with a few layers into the composition label embedding. The embeddings are then aligned with visual features in the shared embedding space, where the recognition of unseen image + +query becomes the nearest neighbor search problem [36]. Nevertheless, the extent to which contextuality is taken into account in the visual domain has still been restricted. + +Prior works [52, 53] have also pointed out the importance of discriminability in visual features, which is related to generalization performance towards unseen composition recognition [62, 38]. A popular solution is disentangle-ment [52, 53, 27, 80], in which independent layers are allocated to extract intermediate visual representations of attribute and object. However, as shown in Fig. 1, it may be challenging to extract their unique characteristics to understand the heterogeneous interaction. We carefully argue that this problem arises from how the visual backbone (e.g. ResNet18 [16]) is used. This is because attribute and object features, which require completely different characteristics [30, 22], are based on the same visual representation from the deep layer of the backbone. + +Another challenge in CZSL is the long-tailed distribution of real-world compositional data [2, 63, 3]. Few attribute classes are dominantly associated with objects, which may cause a hubness problem [10, 12] among visual features. The visual features from the head (frequent) composition become a hub, which aggravates the visual features to be indistinguishable in the embedding space [12], and induces a biased prediction towards dominant composition [61, 81]. + +In this paper, based on the discussions above, we propose Composition Transformer (CoT) to enlarge the visual discrimination for robust CZSL. Motivated by bottom-up and top-down attention [31, 32, 28], the CoT presents object and attribute experts, each forming its feature in different layers of the visual network. Specifically, the object expert generates a representative object embedding from the last layer (i.e., bottom-up pathway) that is most robust to identify object category with high-level semantics [72, 44]. Then we explicitly model contextuality through an object-guided attention module that explores intermediate layers of the backbone network and builds attribute-related features associated with the object (i.e., top-down pathway). Based on this module, the attribute expert generates a distinctive attribute embedding in conjunction with the object embedding. By utilizing all the features of each layer exhibiting different characteristics, our method comprehensively leverages a visual network to diversify the components' features in the shared embedding space. + +Finally, we further develop a simple minority attribute augmentation (MAA) methodology tailored for CZSL to address the biased prediction caused by imbalanced data distribution [61]. Unlike GAN-based augmentations [56, 27] that lead to overwhelming computation during training, our method simplifies the synthesis process of the virtual sample by blending two images while oversampling minority attribute classes [7, 45]. Thanks to the label smoothing effect of the balanced data distribution by augmenta + +tion [79, 59], the CoT is well generalized with minimal computational costs, resolving the bias problem induced by the majority classes. + +Our contributions are summarized in three-folds: + +- To enhance visual discrimination and contextuality, we propose a Composition Transformer (CoT). In this framework, object and attribute experts hierarchically estimate primitive embeddings by fully utilizing intermediate outputs of the visual network. +- We introduce a simple yet robust MAA that alleviates dominant prediction on head compositions with majority attributes. +- The remarkable experimental results demonstrate that our CoT and MAA are harmonized to improve the performance on several CZSL benchmarks, showing state-of-the-art performance. + +# 2. Related Work + +Given a set of objects and their associated attributes, compositional zero-shot learning (CZSL) aims to recognize unseen compositions using primitive knowledge. Contrary to conventional methods [34] that recognize the object and attribute independently, recent approaches [39, 42, 29] have regarded contextuality as a crucial standpoint for understanding the interpretation of the attribute depending on the object class. For example, AttrOps [42] represented the attribute as a linear transformation of the object's state, and SymNets [29] regularized visual features to satisfy the group axioms with auxiliary losses. Some works [41, 75] have leveraged graph networks to model the global dependencies among primitives and their compositions. + +However, those methods have often suffered in distinguishing visual composition features. Meta-learning [13] has been utilized for CZSL frameworks [49, 67] to enhance the discriminative power of visual features but induce high training costs. Other works [52, 27, 53] have disentangled representations of objects and attributes for modeling visual composition. Those methods have shown promising generalization performance than conventional methods. However, the aforementioned approaches have often failed to discriminate the minority classes in object and attribute, which is caused by severe model bias toward seen compositions [61, 81]. They have used the mostly-shared visual backbone for attribute and object representations, which needs more deep concern to be extracted differently according to its characteristics. In specific, the attribute feature is changed in a wide variety of aspects as the property of the object changes, and the object feature must be perceived as semantically the same despite changing the attribute. + +Our method is also closely related to visual attribute prediction [46, 26], where the goal is to recognize visual attributes describing states or properties within an object. + +![](images/6570432d57a19d81c1db8062ac8362c6c3d3a36e8686e094024126723cb18130.jpg) +Figure 2: The overall architecture of composition transformer. Object and attribute experts produce object- and attribute-specific features at different layers. An object-guided attention module is attached to each attribute experts to highlight attribute-specific features. Object-attribute embeddings are composed, and then projected to a joint embedding space with label embedding for optimization. + +Since the fine-grained local information (e.g. color, shape, and style) is required to the task, several approaches [55, 60, 20, 58, 48] utilize low-level features in backbone network, which contain textual details and objects' contour information [74, 69]. Meanwhile, with a similar motivation, some methods for CZSL have integrated low-level features with a multi-scale fusion approach to acquire robust visual features [70, 75]. Although a simple integrated feature can recover spatial details slightly, it is hard to localize the RoI of the attribute that varies with the object for compositional recognition; for example, a puddle in an object 'Street' for a composition class 'Wet Street'. + +Moreover, attributes in generic objects are long-tailed distributed in nature [55, 48]. The imbalance issue occurs as a biased recognition problem, because deep networks tend to be over-confident in head composition classes [54, 65]. This phenomenon is also known as a hubness problem [10, 12], where few visual features from head composition classes become indistinguishable from many tail composition samples. One major solution for the data imbalance problem is the resampling [15, 57, 14, 35] by flattening the long-tailed dataset by over-sampling of minority classes or under-sampling of majority classes. However, directly applying such sampling strategies in CZSL could intensify the overfitting problem on over-sampled classes, and discard valuable samples for learning the alignment between visual and semantic (linguistic) features in the joint embedding space [81]. Another line of work is generation methods [68, 66] that hallucinate a new sample from rare classes but they require complicated training strategies for data generation. In this work, we are inspired by recent data augmentation techniques [77, 7, 45, 79, 8], and combine both resampling and generation approaches to tailor for CZSL, which facilitates the regularization of robust embedding space even in the minority classes via augmented virtual samples. + +# 3. Composition Transformer (CoT) + +In compositional zero-shot learning (CZSL), given an object set $\mathcal{O}$ and an attribute set $\mathcal{A}$ , object-attribute + +pairs comprise a composition set $\mathcal{V} = \mathcal{A}\times \mathcal{O}$ . The network is trained with an observable (seen) composition set $\mathcal{D}_{\mathrm{tr}} = \{\mathcal{X}_{\mathrm{s}},\mathcal{Y}_{\mathrm{s}}\}$ , where $\mathcal{X}_{\mathrm{s}}$ is an image set corresponding to a composition label set $\mathcal{V}_{\mathrm{s}}$ . Following the generalized zero-shot setting [49, 5, 73], a test set $\mathcal{D}_{\mathrm{te}} = \{\mathcal{X}_{\mathrm{te}},\mathcal{Y}_{\mathrm{te}}\}$ consists of samples drawn from both seen and unseen composition classes, i.e. $\mathcal{V}_{\mathrm{te}} = \mathcal{V}_{\mathrm{s}}\cup \mathcal{V}_{\mathrm{u}}$ and $\mathcal{V}_{\mathrm{s}}\cap \mathcal{V}_{\mathrm{u}} = \emptyset$ . For this task, a powerful visual network is required to extract distinct features of objects and attributes respectively, in order to recognize both seen and unseen compositions well. + +To this end, we propose a novel composition transformer (CoT) that includes object and attribute expert modules as illustrated in Fig. 2. Our CoT is basically built on the vision transformer (i.e. ViT [11]) and fully utilizes intermediate blocks to capture the attribute and object representations. These representations are projected into the joint embedding space to calculate the similarity between given compositional labels and images. We also introduce a minority attribute sampling method for data augmentation such that the biased recognition with head composition class is alleviated by label-smoothing regularization effect [40]. Next, we describe in detail the configurations of CoT. + +# 3.1. Object expert + +The CoT is composed of a sequence of transformer blocks including multi-head self-attention and MLP layers. Generally, as the block becomes deeper, class discrimination capability is increased while being invariant from the attribute deformation [78, 43]. Therefore, we select the final output from the last block to an input of the object expert in a bottom-up manner [31, 1]. + +Formally, let us denote $\{\mathbf{z}^n\}_{n=1}^N$ as the output features of the blocks in CoT. The feature of $n$ -th transformer blocks, $\mathbf{z}^n$ , consists of a class token (i.e., [CLS]) $\mathbf{z}_c^n \in \mathbb{R}^{1 \times D}$ and a set of patch tokens $\mathbf{z}_p^n \in \mathbb{R}^{HW \times D}$ . The object expert $E_o$ encodes the last feature to present an object embedding: + +$$ +\mathbf {p} _ {o} = E _ {o} \left(\mathbf {z} _ {c} ^ {N}\right), \tag {1} +$$ + +where $E_{o}$ consists of a fully-connected (FC) layer. To optimize the parameters in the object expert, we compute a + +cross-entropy loss with an object classifier whose weights are initialized from object word-embeddings as follows: + +$$ +\mathcal {L} _ {\mathrm {o b j}} = - \log \frac {\exp \left\{\cos \left(\mathbf {p} _ {o} , w \left(o _ {i}\right)\right) / \tau_ {o} \right\}}{\sum_ {o _ {k} \in \mathcal {O}} \exp \left\{\cos \left(\mathbf {p} _ {o} , w \left(o _ {k}\right)\right) / \tau_ {o} \right\}}, \tag {2} +$$ + +where $\tau$ is a temperature parameter, $w(o)$ denotes the word-embedding [47] corresponding to the object $o\in \mathcal{O}$ . + +# 3.2. Attribute expert + +Capturing both attribute and object in the same deeply condensed feature is nontrivial as they require pretty different visual representations [30, 22]. The straightforward solution is to design two-stream networks without any parameter sharing. However, such individual training could not capture the contextuality [39]. Hence, we build an attribute expert which utilizes semantics estimated from the object expert as a condition, and extracts attribute-related features on intermediate layers of visual backbone in a top-down fashion [31, 1]. We devise an object-guided attention module that leverages convolution kernel-based attention [71, 6, 24], which excels at capturing fine-grained local information required for attribute recognition [18]. + +Specifically, we first reshape $\mathbf{z}_p^n$ to $\mathbf{Z}_p^n$ with output resolution of $H\times W\times D$ . Following [51], we tile $\mathbf{p}_o$ to match the spatial dimension with $\mathbf{Z}_p^n$ and concatenate them to obtain an object-contextualized feature $\{\mathbf{U}^n = [\mathbf{p}_o,\mathbf{Z}_p^n ]\} \in \mathbb{R}^{H\times W\times 2D}$ . As shown in Fig. 2, we generate a channel attention map $\mathbf{A}_c^n\in \mathbb{R}^{1\times 1\times D}$ from the averaged feature across the spatial axis: + +$$ +\mathbf {A} _ {c} ^ {n} = \sigma \left(\operatorname {c o n v} _ {1 \times 1} \left(\operatorname {a v g p o o l} _ {\text {s p a t i a l}} \left(\mathbf {U} ^ {n}\right)\right)\right), \tag {3} +$$ + +where $\sigma$ denotes a sigmoid function. We also generate a spatial attention map $\mathbf{A}_s^n\in \mathbb{R}^{H\times W\times 1}$ from the averaged feature across the channel axis: + +$$ +\mathbf {A} _ {s} ^ {n} = \sigma \left(\operatorname {c o n v} _ {3 \times 3} \left(\operatorname {a v g p o o l} _ {\text {c h a n n e l}} \left(\mathbf {U} ^ {n}\right)\right)\right). \tag {4} +$$ + +Finally, we aggregate two attention maps into a unified attention map $\mathbf{A}^n\in \mathbb{R}^{H\times W\times D}$ as follows: + +$$ +\mathbf {A} ^ {n} = \mathbf {A} _ {s} ^ {n} \times \mathbf {A} _ {c} ^ {n}. \tag {5} +$$ + +The output of our object-guided attention module is defined as attribute features with estimated attention maps, followed by the residual connection: + +$$ +\tilde {\mathbf {Z}} _ {p} ^ {n} = \mathbf {Z} _ {p} ^ {n} + \mathbf {A} ^ {n} \odot \mathbf {Z} _ {p} ^ {n}, \tag {6} +$$ + +where $\odot$ is an operation for element-wise multiplication. + +Given $\tilde{\mathbf{Z}}_p^n$ from each intermediate layer, we ensemble the features from different blocks by taking advantage of the hierarchical architecture [33, 43]. It is required for recognizing heterogeneous types of attributes such as colors or + +shapes [48]. Especially, we choose low-, middle-, and high-level features from the backbone, denoted as $\{\tilde{\mathbf{Z}}_p^l,\tilde{\mathbf{Z}}_p^m,\tilde{\mathbf{Z}}_p^h\}$ , and ensemble those triplets with global average pooling (GAP) followed by concatenation. The attribute expert $E_{a}$ then generates an attribute embedding with the triplets as + +$$ +\mathbf {p} _ {a} = E _ {a} \left(\left[ \operatorname {G A P} \left(\tilde {\mathbf {Z}} _ {p} ^ {l}\right), \operatorname {G A P} \left(\tilde {\mathbf {Z}} _ {p} ^ {m}\right), \operatorname {G A P} \left(\tilde {\mathbf {Z}} _ {p} ^ {h}\right) \right]\right), \tag {7} +$$ + +where $E_{a}$ consists of a FC layer. Similar to the object expert, we optimize the attribute expert with the cross-entropy loss: + +$$ +\mathcal {L} _ {\mathrm {a t t}} = - \log \frac {\exp \left\{\cos \left(\mathbf {p} _ {a} , w \left(a _ {i}\right)\right) / \tau_ {a} \right\}}{\sum_ {a _ {k} \in \mathcal {A}} \exp \left\{\cos \left(\mathbf {p} _ {a} , w \left(a _ {k}\right)\right) / \tau_ {a} \right\}}. \tag {8} +$$ + +# 3.3. Mapping visual-semantic space + +The visual compositional embedding $\mathbf{p}_x$ of the input image $x$ is calculated from object and attribute experts, formulated as + +$$ +\mathbf {p} _ {x} = E _ {c} \left(\left[ \mathbf {p} _ {o}, \mathbf {p} _ {a} \right]\right)). \tag {9} +$$ + +The embedding function $E_{c}$ includes an FC layer, projecting the visual features into the joint embedding space that aligns visual and semantic representations. Following previous works [42, 36], we estimate semantic label embedding $\mathbf{p}_y$ of each seen composition label $y \in \mathcal{V}_s$ by projecting the concatenated word vectors into the joint space: + +$$ +\mathbf {s} _ {y} = g ([ w (o), w (a) ]), \tag {10} +$$ + +where $g(\cdot)$ is a label embedding network that consists of 3 FC layers and ReLU activation function. During training, we minimize the cosine distance of visual and semantic embeddings with cross-entropy loss as follows: + +$$ +\mathcal {L} _ {\text {c o m p}} = - \log \frac {\exp \left\{\cos \left(\mathbf {p} _ {x} , \mathbf {s} _ {y}\right) / \tau_ {c} \right\}}{\sum_ {y _ {k} \in \mathcal {Y} _ {s}} \exp \left\{\cos \left(\mathbf {p} _ {x} , \mathbf {s} _ {y _ {k}}\right)\right) / \tau_ {c} \}, \tag {11} +$$ + +where $\tau_{c}$ is a temperature parameter for the composition loss $\mathcal{L}_{\mathrm{comp}}$ . Finally, the total objectives of CoT is then formulated as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c o m p}} + \alpha \mathcal {L} _ {\text {a t t}} + \beta \mathcal {L} _ {\text {o b j}}, \tag {12} +$$ + +where $\alpha$ and $\beta$ are balance weights. + +During inference, we measure the cosine similarity between a visual embeddings and all label embeddings from $\mathcal{V}_{\mathrm{te}}$ and regard it as a feasibility score of the image and composition labels [36]. Following [80], we also use the classification scores from object and attribute experts as a form of cosine similarity. Therefore, a final feasibility score of label $y = (a,o)$ is derived by adding the above three scores, formulated as: + +$$ +c (y) = \cos \left(\mathbf {p} _ {x}, \mathbf {s} _ {y}\right) + \cos \left(\mathbf {p} _ {o}, w (o)\right) + \cos \left(\mathbf {p} _ {a}, w (a)\right), \tag {13} +$$ + +where we predict the label with the highest score as the final composition label. + +![](images/c7e939d1597c896f8580d538c27c5397b248b3f0c572ad155a5902f77c49ca9e.jpg) +Figure 3: Attribute class distributions on VAW-CZSL. Every attributes in each plot are composed into a fixed object; (a) Grass and (b) Player. + +![](images/bd24f994b76243dcb3e9cc538ec455ae52d1e59cdf7f07d99b15bdde94160da6.jpg) + +# 3.4. Minority attribute augmentation + +With a carefully designed architecture, our CoT can generate distinctive visual features for CZSL. Nevertheless, the imbalanced composition samples in the training dataset restrict the performance since attributes often co-occur with certain objects in the real world [30]. For example in Fig. 3, 'Green Grass' is naturally more common than 'White Grass'. Such long-tailed data makes the deep models be over-confident in the composition with head attributes [81], and thus hurts the generalization on unseen compositions. + +To this end, we present a minority attribute augmentation (MAA) for training, which diversifies the minority composition. First, two compositional embeddings from COt are selected as inputs to the augmentation, where $\mathbf{p}_{x_A}$ and $\mathbf{p}_{x_B}$ share the same object class. Especially, to balance the major and minor attributes, we assign higher sampling weights on minority attributes with the weighted sampler [14]. Formally, let $\zeta_{o_i}$ denote the number of attribute samples in the $i$ -th attribute class composed on an object class $o$ . The sampling weight of the attribute-object pair $\kappa(a_i,o)$ is formulated by an inverse attribute class frequency according to: + +$$ +\kappa \left(a _ {i}, o\right) = \frac {1 / \zeta_ {o _ {i}}}{\sum_ {i} 1 / \zeta_ {o _ {i}}}. \tag {14} +$$ + +Inspired by Mixup [79, 64], we generate a virtual sample $\mathbf{p}_{x_M}$ with alpha compositing and its attribute semantic embedding $w(a_{M})$ by linear combination: + +$$ +\mathbf {p} _ {x _ {M}} = \lambda \mathbf {p} _ {x _ {A}} + (1 - \lambda) \mathbf {p} _ {x _ {B}}, +$$ + +$$ +w \left(a _ {M}\right) = \lambda w \left(a _ {A}\right) + (1 - \lambda) w \left(a _ {B}\right). \tag {15} +$$ + +The mixing factor $\lambda$ is sampled from a uniform distribution. + +We notice that there are two significant differences from Mixup: (1) Compared to Mixup sampling arbitrary images, MAA constructs a composition pair fastening an object with the minority aware sampling to address the imbalanced data issue in CZSL. (2) Mixup mixes one-hot vector labels [79, 77] for classification, while MAA generates a virtual label in word vector space, to jointly regularize vi + +sual and linguistic (label) embedding functions against the majority attribute classes with Eq. (12). + +# 3.5. Extension to other backbones + +Our idea of CoT is generic and scalable, so we can simply extend to other backbones such as ResNet-18 (R18) [16] by selecting specific layers from the backbone composed of multiple layers or blocks. For a fair comparison to prior arts [41, 53, 23], we study the impact of our contribution on the ResNet backbone based on the convolutional operation. In specific, the object expert generates the object embedding from the last fifth convolution block. The classification token is replaced by a global average pooled feature. Similar to using ViT [11], we concatenate the intermediate layers into a feature vector for the attribute expert. The remained settings such as the architecture of $E_{\mathrm{comp}}$ and $g(\cdot)$ are the same as ViT. + +# 4. Experiments + +# 4.1. Experimental setup + +Datasets. MIT-States [19] dataset is a common dataset in CZSL that most previous methods adopted, containing 115 states, 245 objects, and 1962 compositions. C-GQA [41] and VAW-CZSL [53] are two brand-new and large-scale benchmarks collected from in-the-wild images with more general compositions, such as 'Stained Wall' and 'Hazy Mountain'. Detailed data statistics including long-tailed distribution can be found in Appendix A. + +Metrics. We follow common CZSL setting [49], testing performance on both seen and unseen pairs [5]. Under the setting, we report area under the curve (AUC) (\%) between seen and unseen accuracies at different bias term in validation and test splits. We also compute the best seen and unseen accuracies in the curve, and calculate the best harmonic mean to measure the trade-off between the two values. Lastly, we follow [41] and report the attribute and object accuracies on unseen labels to measure the discrimination power of each representation. + +Baselines. We compare our CoT with 9 recent CZSL approaches: AoP [42], LE+ [39], TMN [49], Symnet [29], CompCos [36], CGE [41], SCEN [27], OADis [53], CAPE [23]. All baselines were built on pretrained ResNet-18 (R18) [16], with pre-trained word embedding such as GloVe [47], FastText [4], and Word2Vec [37]. To show the contribution of our proposed elements, we carry on the most recent methods including CGE and OADis with a transformer-based backbone (ViT-B) [11]. We also clarify that CGE and OADis utilize the knowledge of unseen composition labels at training for building a relation graph and hallucinating virtual samples respectively, while other methods do not assess the unseen composition label sets. The results of other baselines were obtained from [41, 27, 23, 53] or re-implemented from their official code. + +
MethodsBackboneMIT-StatesC-GQAVAW-CZSL
V@1T@1HMSUV@1T@1HMSUV@3T@3HMSU
With frozen R18
AoP [42]R182.52.010.716.618.40.90.32.911.83.91.41.49.116.411.7
LE+ [39]R183.52.311.516.221.21.20.65.316.15.01.51.69.816.213.2
TMN [49]R183.32.611.822.717.12.21.17.721.66.32.22.311.919.915.4
Symnet [29]R184.53.413.824.820.03.31.89.825.29.22.32.312.219.115.8
CompCos [36]R186.94.816.926.924.53.52.612.128.111.83.13.214.223.918.0
CGE* [41]R187.25.318.128.925.03.62.511.927.511.72.72.913.023.416.8
SCEN [27]R187.25.318.429.925.24.02.912.428.912.1-----
OADis* [53]R187.65.918.931.125.6-----3.53.615.224.918.7
CAPE [23]R18-5.819.130.526.2-4.216.332.915.6-----
CoTR187.76.219.630.826.84.94.516.633.116.63.63.815.724.619.1
With frozen ViT-B
CGE* [41]ViT-B8.77.321.333.528.64.53.815.631.314.43.73.915.726.319.7
OADis* [53]ViT-B9.07.521.934.229.35.94.616.232.515.34.14.316.626.920.8
CoTViT-B9.27.823.234.831.56.75.117.534.018.84.44.717.726.922.2
With finetuning ViT-B
CGE* [41]ViT-B11.49.724.839.731.67.35.418.538.017.15.96.220.130.125.7
OADis* [53]ViT-B11.810.125.239.232.18.17.020.138.319.86.26.520.431.326.1
CoTViT-B12.110.525.839.533.08.77.422.139.222.76.67.221.732.928.2
+ +Table 1: Quantitative comparison with prior arts. Following [41, 53], we measure AUC values on validation (V) and test (T) datasets, at top-1 (@1) on C-GQA and top-3 (@3) on VAW-CZSL, respectively. In addition, we report harmonic mean (HM), seen (S), unseen (US), attribute (A), and object (O) accuracies. * denotes the models accessing unseen composition labels at training. The best result and the second best result are boldfaced and underlined, respectively. + +# 4.2. Implementation details + +Models. We use ViT-B [11] and ResNet18 [16] pre-trained on ImageNet [9] as a visual backbone. In ViT-B backbone, Each patch size is $16 \times 16$ , and the number of patch tokens in each layer is $14 \times 14$ , having 768 channels per token. The label embedding network $g$ is composed of 3 FC layers, two dropout layers with ReLU activation functions, where hidden dimensions are 900. We use one FC layer for $E_{o}$ and $E_{a}$ to produce 300-dimensional prototypical vectors that are matched with GloVe [47] embedding vectors from composition labels. Therefore, both $E_{c}$ and $g(\cdot)$ produce 300-dimensional embedding vectors, with $l_{2}$ normalization. In the attribute expert, we use the outputs of the 3rd, 6th and 9th blocks (denoted $\mathbf{z}^{l}, \mathbf{z}^{m}, \mathbf{z}^{h}$ respectively) for multi-level feature fusion. For ResNet18, we use the 2nd, 3rd and 4th blocks as $\mathbf{z}^{l}, \mathbf{z}^{m}, \mathbf{z}^{h}$ , resulting in a 438-dimensional feature vector. + +Training setup. We train CoT with Adam optimizer [25]. The input images are augmented with random crop and horizontal flip, and finally resized into $224 \times 224$ . In ViT, learning rate is initialized to 1e-4 for all three datasets, and decayed by a factor of 0.1 for 10 epochs. In R18, the initial learning rate is 5e-5 with the same decay parameter of the ViT setting. Note that we freeze the GloVe [47] word embedding $w$ for fair comparisons. For all datasets, we use the same temperature parameters [36, 53, 27] of each cross-entropy loss, $\tau_{c}$ , $\tau_{o}$ , and $\tau_{a}$ as 0.05, 0.01 and 0.01 respec + +
MethodsMIT-StatesC-GQAVAW-CZSL
Attr.Obj.Attr.Obj.Attr.Obj.
CGE [41]35.744.417.534.624.551.6
OADis [53]35.245.214.742.022.353.8
CoT37.346.019.840.227.354.2
+ +Table 2: Comparison of unseen attribute (Attr.) and object (Obj.) accuracies with the latest CZSL methods [41, 53]. All results are reported on fine-tuned ViT backbone. + +tively. We use the loss balance weights $\alpha$ and $\beta$ as (0.5, 0.5) on C-GQA, and (0.4, 0.6) on VAW-CZSL and MIT-States. We note that to prevent overfitting of COT we apply MAA from 15 and 30 epochs to ViT and R18, respectively. + +# 4.3. Main evaluation + +Table 1 compares generalization performance between our methods and the baselines with ResNet18 [16] (R18) and ViT-B [11] backbone setting. To further compare visual discrimination, we also compare unseen attribute and object accuracies of our methods and the latest CZSL methods [41, 53] in Table 2. In the following, we analyze the results from the three datasets. + +MIT-States. From Table 1, the CoT achieves the best AUC values (V@1 and H@1) and harmonic mean (HM) with both R18 and ViT backbone. These results sufficiently demonstrate our experts' applicability on the CNN-based + +
CoTMAAAUCHM
5.8118.1
✓6.0519.0
✓6.9320.8
✓✓7.2021.7
+ +(a) Component analysis. + +
Loss componentAUCHM
Lcomp5.9418.9
Lcomp+Lobj6.2219.6
Lcomp+Lattr6.9121.4
Lcomp+Lattr+Lobj7.2021.7
+ +(b) Impact of loss functions. + +
AugmentationAUCHM
Manifold Mixup [64]5.9318.8
CutMix [77]6.9520.7
Mixup [79]7.0421.0
MAA7.2021.7
+ +(c) Impact of data augmentation. + +
GuidanceAUCHMSUAO
w/o guidance6.720.530.725.822.553.4
Attribute guidance6.921.031.526.425.753.9
Object guidance7.221.732.928.227.354.2
+ +(d) Impact of object guidance. + +
EnsemblezlzmzhAUCHMSUAO
Low2346.620.730.328.725.852.3
Mid5676.921.231.929.725.255.9
High910116.821.433.426.624.154.5
Mixture3697.221.732.928.227.354.2
+ +(e) Different configurations of intermediate blocks in attribute experts. + +Table 3: Ablation experiments on VAW-CZSL dataset. Note that all results are reported on CoT (ViT-B), with a fine-tuning setting. + +backbone. With a finetuned ViT-B, the CoT obtains the best validation AUC of $12.1\%$ and test AUC of $10.5\%$ , surpassing previous CZSL methods with a large margin. Our method also improves attribute (Attr.) and object (Obj.) accuracies as in Table 2, showing the effectiveness of CoT for visual discrimination. + +C-GQA. Similar trends can be observed in the C-GQA dataset. In the frozen R18 setting, the CoT achieves the best test AUC (T@1) of $4.5\%$ , comparable with previous state-of-the-art CAPE. Although the T@1 scores from CGE and OADis are also increased upon replacing the backbone from R18 to ViT-B, our CoT with ViT-B outperforms OADis with larger margins in all metrics, as compared to the performance with R18. It is noticeable that our proposed augmentation technique contributes the unbiased prediction, improving unseen accuracies into $18.8\%$ with a minority-aware sampling. The performance regarding AUC is improved with end-to-end training with backbone, $2.0\%$ in V@1 and $2.3\%$ in T@1. Of note, OADis and CGE access unseen composition labels at training, so it is unfair to directly compare with other methods including CoT. Nevertheless, more surprisingly, our CoT shows comparable and superior performance to OADis. Finally, in Table 2, our model also improves the attribute accuracy by achieving $19.8\%$ , having comparable $40.2\%$ object accuracy compared to OADis and CGE. + +VAW-CZSL. We also achieve SoTA performance on VAW-CZSL, the most challenging benchmark having almost 10K composition labels for evaluation. Even if we do not use additional unseen composition information like OADis, our method achieves the highest T@1 of $7.2\%$ with the best unseen label accuracy of $28.2\%$ . We also obtain both the highest attribute and object score (Table 2), especially with a significant improvement in attribute prediction by $2.8\%$ compared to CGE. It proves that the careful modeling of the visual network in CoT is essential to substantially improve + +the generalization capability. In addition, the results suggest that finetuning the visual backbone to suit composition recognition improves the performance in all benchmarks. + +# 4.4. Ablation study + +We ablate our CoT and MAA with different design choices on Table 3. We use the VAW-CZSL dataset and fine-tuned ViT-B backbone as our main ablation setting. + +Component analysis. We report ablation results of CoT and MAA in Table 3a. In 'without CoT' case, we use CompCos [36] baseline. CoT boosts the AUC and HM significantly about $1.1\%$ and $2.3\%$ compared to the baseline. MAA offers better results to both CoT and baseline, showing scalability to other methods. Notably, employing MAA on CoT gives higher gains of $0.3\%$ AUC and $2.3\%$ HM compared to the baseline, demonstrating that the two components create a synergy effect. + +Impact of loss functions. Table 3b shows the impact of each loss function. Compared to the case using $\mathcal{L}_{\mathrm{comp}}$ only, $\mathcal{L}_{\mathrm{obj}}$ and $\mathcal{L}_{\mathrm{attr}}$ bring a gain on both AUC and HM. Using both losses to $\mathcal{L}_{\mathrm{comp}}$ improves AUC and HM significantly by $1.26\%$ and $2.8\%$ respectively, showing the importance of object and attribute losses to learn each expert. + +Different data augmentations. In Table 3c, we compare MAA with various data mixing methodologies including Manifold Mixup [64], CutMix [77] and Mixup [79]. To apply Manifold Mixup in CoT, we select 3,6,9 and the last layer as a mixup layer, and mix these intermediate features. MAA achieves the highest score on both AUC and HM, outperforming other augmentation methods including standard Mixup. This demonstrates the effectiveness of MAA which balances the data distribution with a composition pair. Interestingly, we found that Manifold Mixup performs at worst, even degrading pure CoT performance (AUC: 6.93, HM: 20.4) in Table 3a. We speculate that the mixed representation at intermediate layers can lead to significant underfit + +![](images/f42adfde4d5eb773bcde54cab35a89d2ebb499cb29e0c8ccd614af226591ee50.jpg) +Figure 4: Distribution of k-occurrence counts with $k = 20$ of C-GQA and VAW-CZSL test set. All results are reported with ViT-B backbone. x-axis stands for composition class ids which are ordered by decreasing count numbers. For better visualization, we show the top $10\%$ head composition classes where the hubness is predominant. Circles at start points indicate the most frequent hubs. The result on MIT-States is in the Appendix C.1. + +ting of the attention module. + +Impact of object guidance. As shown in Table 3d, the object guidance significantly boosts the performance. We found that reversing the guidance (attribute guidance) decreases the performance of AUC by $0.3\%$ and HM by $0.7\%$ . The results validate the usage of object guidance for contextualized attribute prediction. + +Design choice of feature ensemble. Table 3e shows the ablations with different configurations of intermediate blocks used in the attribute expert. We evaluate four types of block feature ensembles: low, mid, high, and mixture. Among them, the mixture type achieves the best performance on AUC, harmonic mean, and attribute accuracy. We therefore choose this mixture ensemble as a default setting. Another finding is low-level feature is more adequate for attribute and unseen composition recognition, while high-level feature is robust to object and seen composition accuracy. + +# 4.5. Discussion + +Mitigating hubness problem. We illustrate the hubness effect of a baseline [36] and our method on C-GQA and VAW-CZSL datasets in Fig. 4. Concretely, we count $k$ -occurrences $(N_{k})$ , the number of times that a sample appears in the $k$ nearest neighbors of other points [50] for all class ids, and calculate the skewness of $N_{k}$ distribution for hubness measurement. We observe that CoT significantly alleviates the hubness problem of baseline across datasets, enlarging the visual discrimination in the embedding space. MAA further reduces the skewness of the distribution by mitigating biased prediction on head composition classes. + +Object-guided attention visualization. Qualitatively, we visualize attention maps from the object-guided attention module in Fig. 5. We observe that the attention module could capture the contextualized meaning of attributes in composition, e.g., wet skin of horse vs. puddle in street. It + +![](images/7f276bc4dd14e73b5ec5444822ca5e18bbb2920da2515767590b10ed3438525e.jpg) + +![](images/d6a7ea8ed80d52633e8d036bb4ca5cb40b471006edd68918b7a3fe6da80422f2.jpg) +Figure 5: Visualization of object-guided attention on VAW-CZSL dataset. Each row represents the attended regions in image pairs, which are related to different attributes contextualized with a specific object. More results in the Appendix C.5. + +![](images/3ba0c5c88fa9fe2ea9383b762b07160bc76ad1a2070cca0500779df7cc311e2f.jpg) +Graite Ground + +![](images/cabdee7b0e0a882dd56a5548893f2db4e16a43e51e71986cc57a43bfe6f40bff.jpg) +(d) +(g) +Figure 6: Ground-truth composition labels and top-3 prediction results on MIT-States, C-GQA and VAW-CZSL datasets. More results in the Appendix C.6. + +![](images/7f38e241c7dbebadba09033625aaf18fbe5a4d186c0a6bb4a5c8193c2a68fcea.jpg) +(h) + +![](images/439a78825ff7be7f6787cad581a1dc46ecaa9cf97ac9e76f871459c5c423b0a7.jpg) +(i) + +validates that the attention module assists CoT to extract distinctive attribute embedding for composition by localizing specific RoI. + +Qualitative results. Fig. 6 illustrates the qualitative results of CoT. Even though some predictions do not correspond to annotated image labels, they explain the composition in another view; e.g., 'Furry Dog' and 'Shaggy Dog' in Fig. 6 (i). Moreover, in-the-wild images usually have multiple objects with their own attributes; e.g., 'White Wall' and 'Reading Man' in Fig. 6 (h). To recognize such compositions in the real world, it is necessary to handle the multi-label nature of composition, while distinguishing each attribute-object interaction [17] in an image. It encourages us to construct a new CZSL benchmark along with a novel evaluation metric handling both multi-label prediction and multiple compositions in the image. We leave it as a future work. + +# 5. Conclusion + +In this paper, we propose Composition Transformer (CoT) with simple minority attribute augmentation (MAA), a contextual, discriminative, and unbiased CZSL framework. The proposed object and attribute experts hierarchically utilize the entire powerful visual backbone to generate a composition embedding. The simple but effective MAA balances the long-tail distributed composition labels. Extensive studies with comprehensive analysis demonstrate the effectiveness of each component, and show that CoT surpasses existing methods across several benchmarks. + +Acknowledgement. This research was supported by the Yonsei Signature Research Cluster Program of 2022 (2022-22-0002) and the KIST Institutional Program (Project No.2E31051-21-203). + +# References + +[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 3, 4 +[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 2 +[3] Yuval Atzmon, Felix Kreuk, Uri Shalit, and Gal Chechik. A causal view of compositional zero-shot recognition. In NeurIPS, 2020. 2 +[4] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. TACL, 2017. 5 +[5] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In ECCV, 2016.3,5 +[6] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020. 4 +[7] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In ECCV, 2020. 2, 3 +[8] Yu-Ying Chou, Hsuan-Tien Lin, and Tyng-Luh Liu. Adaptive and generative zero-shot learning. In International conference on learning representations, 2020. 3 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6 +[10] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2, 3 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3, 5, 6 +[12] Nanyi Fei, Yizhao Gao, Zhiwu Lu, and Tao Xiang. Z-score normalization, hubness, and few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 142-151, 2021. 2, 3 +[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2 +[14] Yonatan Geifman and Ran El-Yaniv. Deep active learning over the long tail. arXiv preprint arXiv:1711.00941, 2017. 3,5 +[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE TKDE, 2009. 3 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 5, 6 +[17] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In ECCV, 2020. 8 +[18] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In CVPR, 2020. 4 +[19] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 5 +[20] Mahdi M Kalayeh and Mubarak Shah. On symbiosis of attribute prediction and semantic segmentation. IEEE TPAMI, 2019. 3 +[21] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Revisiting visual product for compositional zero-shot learning. In NeurIPSW, 2021. 1 +[22] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In CVPR, 2022. 2, 4 +[23] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Alain Pagani, Didier Stricker, and Muhammad Zeshan Afzal. Learning attention propagation for compositional zero-shot learning. arXiv preprint arXiv:2210.11557, 2022. 5, 6 +[24] Hanjae Kim, Sunghun Joung, Ig-Jae Kim, and Kwanghoon Sohn. Prototype-guided saliency feature learning for person search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4865-4874, 2021. 4 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2 +[27] Xiangyu Li, Xu Yang, Kun Wei, Cheng Deng, and Muli Yang. Siamese contrastive embedding network for compositional zero-shot learning. In CVPR, 2022. 2, 5, 6 + +[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2 +[29] Yong-Lu Li, Yue Xu, Xiaohan Mao, and Cewu Lu. Symmetry and group in attribute-object compositions. In CVPR, 2020. 2, 5, 6 +[30] Kongming Liang, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Unifying visual attribute learning with object recognition in a multiplicative framework. IEEE TPAMI, 2018. 2, 4, 5 +[31] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2, 3, 4 +[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 2 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4 +[34] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, 2016. 2 +[35] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 3 +[36] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8 +[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 1, 5 +[38] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In CVPR, pages 12664-12673, 2020. 2 +[39] Ishan Misra, Abhinav Gupta, and Martial Hebert. From red wine to red tomato: Composition with context. In CVPR, 2017. 1, 2, 4, 5, 6 +[40] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 3 +[41] Muhammad Ferjad Naeem, Yongqin Xian, Federico Tombari, and Zeynep Akata. Learning graph embeddings for compositional zero-shot learning. In CVPR, 2021. 1, 2, 5, 6 +[42] Tushar Nagarajan and Kristen Grauman. Attributes as operators: factorizing unseen attribute-object compositions. In ECCV, 2018. 1, 2, 4, 5, 6 +[43] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, 2021. 3, 4 +[44] Zizheng Pan, Bohan Zhuang, Jing Liu, Haoyu He, and Jianfei Cai. Scalable vision transformers with hierarchical pool + +ing. In Proceedings of the IEEE/cvf international conference on computer vision, pages 377-386, 2021. 2 +[45] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In CVPR, 2022. 2, 3 +[46] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 2 +[47] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 1, 4, 5, 6 +[48] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 3, 4 +[49] Senthil Purushwalkam, Maximilian Nickel, Abhinav Gupta, and Marc'Aurelio Ranzato. Task-driven modular networks for zero-shot compositional learning. In ICCV, 2019. 2, 3, 5, 6 +[50] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 8 +[51] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4 +[52] Frank Ruis, Gertjan Burghouts, and Doina Bucur. Independent prototype propagation for zero-shot compositionality. In NeurIPS, 2021. 2 +[53] Nirat Saini, Khoi Pham, and Abhinav Shrivastava. Disentangling visual embeddings for attributes and objects. In CVPR, 2022. 1, 2, 5, 6 +[54] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In WACV, 2021. 3 +[55] Nikolaos Sarafianos, Xiang Xu, and Ioannis A Kakadiaris. Deep imbalanced attribute classification using visual attention aggregation. In ECCV, 2018. 3 +[56] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, 2019. 2 +[57] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In ECCV, 2016. 3 +[58] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In CVPR, 2021. 3 +[59] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 2 +[60] Chufeng Tang, Lu Sheng, Zhaoxiang Zhang, and Xiaolin Hu. Improving pedestrian attribute recognition with weakly-supervised multi-scale attribute-specific localization. In ICCV, 2019. 3 +[61] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the + +bad momentum causal effect. Advances in Neural Information Processing Systems, 33:1513-1524, 2020. 2 +[62] Bin Tong, Chao Wang, Martin Klinkigt, Yoshiyuki Kobayashi, and Yuuichi Nonaka. Hierarchical disentanglement of discriminative latent features for zero-shot learning. In CVPR, pages 11467-11476, 2019. 2 +[63] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 2 +[64] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 5, 7 +[65] Byron C Wallace and Issa J Dahabreh. Improving class probability estimates for imbalanced data. KAIS, 2014. 3 +[66] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 3 +[67] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, 2019. 2 +[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018. 3 +[69] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In CVPR, 2021. 3 +[70] Kun Wei, Muli Yang, Hao Wang, Cheng Deng, and Xianglong Liu. Adversarial fine-grained composition learning for unseen attribute-object recognition. In ICCV, 2019. 3 +[71] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 4 +[72] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3907-3916, 2019. 2 +[73] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 2018. 3 +[74] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature based activation map for weakly supervised object localization. In ICCV, 2021. 3 +[75] Ziwei Xu, Guangzhi Wang, Yongkang Wong, and Mohan S. Kankanhalli. Relation-aware compositional zero-shot learning for attribute-object pair recognition. IEEE TMM, 2021. 2, 3 +[76] Muli Yang, Cheng Deng, Junchi Yan, Xianglong Liu, and Dacheng Tao. Learning unseen concepts via hierarchical decomposition and composition. In CVPR, 2020. 1 + +[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 3, 5, 7 +[78] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 3 +[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 3, 5, 7 +[80] Tian Zhang, Kongming Liang, Ruoyi Du, Xian Sun, Zhanyu Ma, and Jun Guo. Learning invariant visual representations for compositional zero-shot learning. arXiv preprint arXiv:2206.00415, 2022. 1, 2, 4 +[81] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 3, 5 \ No newline at end of file diff --git a/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/images.zip b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..10e3f847a0e5e1b6b69d1ea62444daa9eff6157f --- /dev/null +++ b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1b0788f773987f6176c31bcc70453e0af89bf49b26c26ba8999190a71033017 +size 517610 diff --git a/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/layout.json b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5d28fa446873d92fb0bbf5f9f29a4f9c894d40bf --- /dev/null +++ b/2023/Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning/layout.json @@ -0,0 +1,11180 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "text", + "content": "Hierarchical Visual Primitive Experts for Compositional Zero-Shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 142, + 168, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 142, + 168, + 158 + ], + "spans": [ + { + "bbox": [ + 103, + 142, + 168, + 158 + ], + "type": "text", + "content": "Hanjae Kim" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 143, + 257, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 143, + 257, + 157 + ], + "spans": [ + { + "bbox": [ + 190, + 143, + 257, + 157 + ], + "type": "text", + "content": "Jiyoung Lee²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 278, + 143, + 362, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 143, + 362, + 157 + ], + "spans": [ + { + "bbox": [ + 278, + 143, + 362, + 157 + ], + "type": "text", + "content": "Seongheon Park" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 385, + 143, + 490, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 143, + 490, + 157 + ], + "spans": [ + { + "bbox": [ + 385, + 143, + 490, + 157 + ], + "type": "text", + "content": "Kwanghoon Sohn" + }, + { + "bbox": [ + 385, + 143, + 490, + 157 + ], + "type": "inline_equation", + "content": "^{1,3,*}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 162, + 153, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 162, + 153, + 176 + ], + "spans": [ + { + "bbox": [ + 60, + 162, + 153, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 162, + 153, + 176 + ], + "type": "text", + "content": "Yonsei University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 177, + 162, + 260, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 162, + 260, + 175 + ], + "spans": [ + { + "bbox": [ + 177, + 162, + 260, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 177, + 162, + 260, + 175 + ], + "type": "text", + "content": "NAVER AI Lab" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 285, + 162, + 363, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 162, + 363, + 176 + ], + "spans": [ + { + "bbox": [ + 285, + 162, + 363, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 285, + 162, + 363, + 176 + ], + "type": "text", + "content": "Korea Institute" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 163, + 533, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 163, + 533, + 177 + ], + "spans": [ + { + "bbox": [ + 367, + 163, + 533, + 177 + ], + "type": "text", + "content": "of Science and Technology (KIST)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 120, + 182, + 332, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 182, + 332, + 194 + ], + "spans": [ + { + "bbox": [ + 120, + 182, + 332, + 194 + ], + "type": "text", + "content": "{incohjk,sam121796,khsohn}@yonsei.ac.kr" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 368, + 182, + 473, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 182, + 473, + 193 + ], + "spans": [ + { + "bbox": [ + 368, + 182, + 473, + 193 + ], + "type": "text", + "content": "lee.j@navercorp.com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 233, + 191, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 233, + 191, + 245 + ], + "spans": [ + { + "bbox": [ + 143, + 233, + 191, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 259, + 290, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 290, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 290, + 569 + ], + "type": "text", + "content": "Compositional zero-shot learning (CZSL) aims to recognize unseen compositions with prior knowledge of known primitives (attribute and object). Previous works for CZSL often suffer from grasping the contextuality between attribute and object, as well as the discriminability of visual features, and the long-tailed distribution of real-world compositional data. We propose a simple and scalable framework called Composition Transformer (CoT) to address these issues. CoT employs object and attribute experts in distinctive manners to generate representative embeddings, using the visual network hierarchically. The object expert extracts representative object embeddings from the final layer in a bottom-up manner, while the attribute expert makes attribute embeddings in a top-down manner with a proposed object-guided attention module that models contextuality explicitly. To remedy biased prediction caused by imbalanced data distribution, we develop a simple minority attribute augmentation (MAA) that synthesizes virtual samples by mixing two images and oversampling minority attribute classes. Our method achieves SoTA performance on several benchmarks, including MIT-States, C-GQA, and VAW-CZSL. We also demonstrate the effectiveness of CoT in improving visual discrimination and addressing the model bias from the imbalanced data distribution. The code is available at https://github.com/HanjaeKim98/CoT." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "text", + "content": "Humans perceive entities as hierarchies of parts; for example, we recognize 'Cute Cat' by composing the meaning of 'Cute' and 'Cat'. People can even perceive new concepts by composing the primitive meanings they already" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 222, + 432, + 304 + ], + "blocks": [ + { + "bbox": [ + 309, + 222, + 432, + 304 + ], + "lines": [ + { + "bbox": [ + 309, + 222, + 432, + 304 + ], + "spans": [ + { + "bbox": [ + 309, + 222, + 432, + 304 + ], + "type": "image", + "image_path": "8cfc7ef69ae570faf24074f64d6c08a8f23aaebbb84dbebffd51e2968bc66485.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 433, + 233, + 544, + 304 + ], + "blocks": [ + { + "bbox": [ + 433, + 233, + 544, + 304 + ], + "lines": [ + { + "bbox": [ + 433, + 233, + 544, + 304 + ], + "spans": [ + { + "bbox": [ + 433, + 233, + 544, + 304 + ], + "type": "image", + "image_path": "d75d6cbdc356857837d3947d57f3b52a8cff261348df50fa120d3b8160139a50.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 308, + 312, + 432, + 392 + ], + "blocks": [ + { + "bbox": [ + 308, + 312, + 432, + 392 + ], + "lines": [ + { + "bbox": [ + 308, + 312, + 432, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 312, + 432, + 392 + ], + "type": "image", + "image_path": "74bbbd01f18c8ba45fc28da780044b8c0f47d3cd8554f3b496409ee554e597d5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 393, + 393, + 403 + ], + "lines": [ + { + "bbox": [ + 356, + 393, + 393, + 403 + ], + "spans": [ + { + "bbox": [ + 356, + 393, + 393, + 403 + ], + "type": "text", + "content": "(a)OADis" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 433, + 323, + 544, + 392 + ], + "blocks": [ + { + "bbox": [ + 433, + 323, + 544, + 392 + ], + "lines": [ + { + "bbox": [ + 433, + 323, + 544, + 392 + ], + "spans": [ + { + "bbox": [ + 433, + 323, + 544, + 392 + ], + "type": "image", + "image_path": "ab0c3570a37ec99c73300541daa30a0c4021a884fb4d4755db8bdd6fd3f0c9eb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 393, + 515, + 403 + ], + "lines": [ + { + "bbox": [ + 462, + 393, + 515, + 403 + ], + "spans": [ + { + "bbox": [ + 462, + 393, + 515, + 403 + ], + "type": "text", + "content": "(b) CoT (Ours)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 406, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 473 + ], + "type": "text", + "content": "Figure 1: Visual feature distribution of previous state-of-the-art approach (OADis) [53] and CoT (Ours). The features are gathered with respect to attribute 'Blue -' (Top) and object ' - Wall' (Bottom). Notably, the previous method suffers from degraded visual discrimination such as ('Blue, Jeans' vs. 'Blue, Shirt'), or all attributes composed into object 'Wall'." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "content": "knew. Such compositionality is a fundamental ability of humans for all cognition. For this reason, compositional zero-shot learning (CZSL), recognizing a novel composition of known components (i.e., object and attribute), has been regarded as a crucial problem in the research community." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "A naive approach for CZSL is to combine attribute and object scores estimated from individually trained classifiers [39, 21]. However, it is hard for these independent predictions to consider the interactions that result from a combination of different primitives, which is contextuality (e.g. different meaning of 'Old' in 'Old Car' and 'Old Cat') in the composition [39]. Previous approaches [42, 76, 36, 41, 53, 80] have solved the problem by modeling the compositional label (text) embedding for each class. They leverage external knowledge bases like Glove [37, 47] to extract attribute and object semantic vectors, and concatenate them with a few layers into the composition label embedding. The embeddings are then aligned with visual features in the shared embedding space, where the recognition of unseen image" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 665, + 136, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 136, + 674 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 136, + 674 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 287, + 712 + ], + "type": "text", + "content": "This work was supported by the Institute of Information Communications Technology Planning Evaluation (IITP) grant funded by the Korean government (MSIT) (No. 2021-0-02068, Artificial Intelligence Innovation Hub)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5675" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "query becomes the nearest neighbor search problem [36]. Nevertheless, the extent to which contextuality is taken into account in the visual domain has still been restricted." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 277 + ], + "type": "text", + "content": "Prior works [52, 53] have also pointed out the importance of discriminability in visual features, which is related to generalization performance towards unseen composition recognition [62, 38]. A popular solution is disentangle-ment [52, 53, 27, 80], in which independent layers are allocated to extract intermediate visual representations of attribute and object. However, as shown in Fig. 1, it may be challenging to extract their unique characteristics to understand the heterogeneous interaction. We carefully argue that this problem arises from how the visual backbone (e.g. ResNet18 [16]) is used. This is because attribute and object features, which require completely different characteristics [30, 22], are based on the same visual representation from the deep layer of the backbone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 376 + ], + "type": "text", + "content": "Another challenge in CZSL is the long-tailed distribution of real-world compositional data [2, 63, 3]. Few attribute classes are dominantly associated with objects, which may cause a hubness problem [10, 12] among visual features. The visual features from the head (frequent) composition become a hub, which aggravates the visual features to be indistinguishable in the embedding space [12], and induces a biased prediction towards dominant composition [61, 81]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 604 + ], + "type": "text", + "content": "In this paper, based on the discussions above, we propose Composition Transformer (CoT) to enlarge the visual discrimination for robust CZSL. Motivated by bottom-up and top-down attention [31, 32, 28], the CoT presents object and attribute experts, each forming its feature in different layers of the visual network. Specifically, the object expert generates a representative object embedding from the last layer (i.e., bottom-up pathway) that is most robust to identify object category with high-level semantics [72, 44]. Then we explicitly model contextuality through an object-guided attention module that explores intermediate layers of the backbone network and builds attribute-related features associated with the object (i.e., top-down pathway). Based on this module, the attribute expert generates a distinctive attribute embedding in conjunction with the object embedding. By utilizing all the features of each layer exhibiting different characteristics, our method comprehensively leverages a visual network to diversify the components' features in the shared embedding space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Finally, we further develop a simple minority attribute augmentation (MAA) methodology tailored for CZSL to address the biased prediction caused by imbalanced data distribution [61]. Unlike GAN-based augmentations [56, 27] that lead to overwhelming computation during training, our method simplifies the synthesis process of the virtual sample by blending two images while oversampling minority attribute classes [7, 45]. Thanks to the label smoothing effect of the balanced data distribution by augmenta" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "tion [79, 59], the CoT is well generalized with minimal computational costs, resolving the bias problem induced by the majority classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 109, + 517, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 109, + 517, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 109, + 517, + 120 + ], + "type": "text", + "content": "Our contributions are summarized in three-folds:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 123, + 545, + 275 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 123, + 545, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 123, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 123, + 545, + 183 + ], + "type": "text", + "content": "- To enhance visual discrimination and contextuality, we propose a Composition Transformer (CoT). In this framework, object and attribute experts hierarchically estimate primitive embeddings by fully utilizing intermediate outputs of the visual network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 187, + 545, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 187, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 545, + 223 + ], + "type": "text", + "content": "- We introduce a simple yet robust MAA that alleviates dominant prediction on head compositions with majority attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 227, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 227, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 227, + 545, + 275 + ], + "type": "text", + "content": "- The remarkable experimental results demonstrate that our CoT and MAA are harmonized to improve the performance on several CZSL benchmarks, showing state-of-the-art performance." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 286, + 392, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 392, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 392, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 306, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 461 + ], + "type": "text", + "content": "Given a set of objects and their associated attributes, compositional zero-shot learning (CZSL) aims to recognize unseen compositions using primitive knowledge. Contrary to conventional methods [34] that recognize the object and attribute independently, recent approaches [39, 42, 29] have regarded contextuality as a crucial standpoint for understanding the interpretation of the attribute depending on the object class. For example, AttrOps [42] represented the attribute as a linear transformation of the object's state, and SymNets [29] regularized visual features to satisfy the group axioms with auxiliary losses. Some works [41, 75] have leveraged graph networks to model the global dependencies among primitives and their compositions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 676 + ], + "type": "text", + "content": "However, those methods have often suffered in distinguishing visual composition features. Meta-learning [13] has been utilized for CZSL frameworks [49, 67] to enhance the discriminative power of visual features but induce high training costs. Other works [52, 27, 53] have disentangled representations of objects and attributes for modeling visual composition. Those methods have shown promising generalization performance than conventional methods. However, the aforementioned approaches have often failed to discriminate the minority classes in object and attribute, which is caused by severe model bias toward seen compositions [61, 81]. They have used the mostly-shared visual backbone for attribute and object representations, which needs more deep concern to be extracted differently according to its characteristics. In specific, the attribute feature is changed in a wide variety of aspects as the property of the object changes, and the object feature must be perceived as semantically the same despite changing the attribute." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Our method is also closely related to visual attribute prediction [46, 26], where the goal is to recognize visual attributes describing states or properties within an object." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5676" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 544, + 177 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 544, + 177 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 544, + 177 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 544, + 177 + ], + "type": "image", + "image_path": "6570432d57a19d81c1db8062ac8362c6c3d3a36e8686e094024126723cb18130.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 546, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 546, + 217 + ], + "type": "text", + "content": "Figure 2: The overall architecture of composition transformer. Object and attribute experts produce object- and attribute-specific features at different layers. An object-guided attention module is attached to each attribute experts to highlight attribute-specific features. Object-attribute embeddings are composed, and then projected to a joint embedding space with label embedding for optimization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 227, + 289, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 289, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 289, + 370 + ], + "type": "text", + "content": "Since the fine-grained local information (e.g. color, shape, and style) is required to the task, several approaches [55, 60, 20, 58, 48] utilize low-level features in backbone network, which contain textual details and objects' contour information [74, 69]. Meanwhile, with a similar motivation, some methods for CZSL have integrated low-level features with a multi-scale fusion approach to acquire robust visual features [70, 75]. Although a simple integrated feature can recover spatial details slightly, it is hard to localize the RoI of the attribute that varies with the object for compositional recognition; for example, a puddle in an object 'Street' for a composition class 'Wet Street'." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 371, + 289, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 289, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 289, + 659 + ], + "type": "text", + "content": "Moreover, attributes in generic objects are long-tailed distributed in nature [55, 48]. The imbalance issue occurs as a biased recognition problem, because deep networks tend to be over-confident in head composition classes [54, 65]. This phenomenon is also known as a hubness problem [10, 12], where few visual features from head composition classes become indistinguishable from many tail composition samples. One major solution for the data imbalance problem is the resampling [15, 57, 14, 35] by flattening the long-tailed dataset by over-sampling of minority classes or under-sampling of majority classes. However, directly applying such sampling strategies in CZSL could intensify the overfitting problem on over-sampled classes, and discard valuable samples for learning the alignment between visual and semantic (linguistic) features in the joint embedding space [81]. Another line of work is generation methods [68, 66] that hallucinate a new sample from rare classes but they require complicated training strategies for data generation. In this work, we are inspired by recent data augmentation techniques [77, 7, 45, 79, 8], and combine both resampling and generation approaches to tailor for CZSL, which facilitates the regularization of robust embedding space even in the minority classes via augmented virtual samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 669, + 233, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 233, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 233, + 683 + ], + "type": "text", + "content": "3. Composition Transformer (CoT)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In compositional zero-shot learning (CZSL), given an object set " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " and an attribute set " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": ", object-attribute" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": "pairs comprise a composition set " + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\mathcal{A}\\times \\mathcal{O}" + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": ". The network is trained with an observable (seen) composition set " + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{tr}} = \\{\\mathcal{X}_{\\mathrm{s}},\\mathcal{Y}_{\\mathrm{s}}\\}" + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": " is an image set corresponding to a composition label set " + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": ". Following the generalized zero-shot setting [49, 5, 73], a test set " + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{te}} = \\{\\mathcal{X}_{\\mathrm{te}},\\mathcal{Y}_{\\mathrm{te}}\\}" + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": " consists of samples drawn from both seen and unseen composition classes, i.e. " + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{te}} = \\mathcal{V}_{\\mathrm{s}}\\cup \\mathcal{V}_{\\mathrm{u}}" + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{s}}\\cap \\mathcal{V}_{\\mathrm{u}} = \\emptyset" + }, + { + "bbox": [ + 304, + 227, + 546, + 346 + ], + "type": "text", + "content": ". For this task, a powerful visual network is required to extract distinct features of objects and attributes respectively, in order to recognize both seen and unseen compositions well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 347, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 491 + ], + "type": "text", + "content": "To this end, we propose a novel composition transformer (CoT) that includes object and attribute expert modules as illustrated in Fig. 2. Our CoT is basically built on the vision transformer (i.e. ViT [11]) and fully utilizes intermediate blocks to capture the attribute and object representations. These representations are projected into the joint embedding space to calculate the similarity between given compositional labels and images. We also introduce a minority attribute sampling method for data augmentation such that the biased recognition with head composition class is alleviated by label-smoothing regularization effect [40]. Next, we describe in detail the configurations of CoT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 498, + 392, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 392, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 392, + 510 + ], + "type": "text", + "content": "3.1. Object expert" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 516, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 598 + ], + "type": "text", + "content": "The CoT is composed of a sequence of transformer blocks including multi-head self-attention and MLP layers. Generally, as the block becomes deeper, class discrimination capability is increased while being invariant from the attribute deformation [78, 43]. Therefore, we select the final output from the last block to an input of the object expert in a bottom-up manner [31, 1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": "Formally, let us denote " + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}^n\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": " as the output features of the blocks in CoT. The feature of " + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": "-th transformer blocks, " + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^n" + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": ", consists of a class token (i.e., [CLS]) " + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_c^n \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": " and a set of patch tokens " + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_p^n \\in \\mathbb{R}^{HW \\times D}" + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": ". The object expert " + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "inline_equation", + "content": "E_o" + }, + { + "bbox": [ + 304, + 599, + 547, + 660 + ], + "type": "text", + "content": " encodes the last feature to present an object embedding:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 394, + 667, + 546, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 667, + 546, + 681 + ], + "spans": [ + { + "bbox": [ + 394, + 667, + 546, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {o} = E _ {o} \\left(\\mathbf {z} _ {c} ^ {N}\\right), \\tag {1}", + "image_path": "c3b9d03be45d008f5e7a67069bdc4f9ca076d819402b7eadbc9b97a458c07c8c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "E_{o}" + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": " consists of a fully-connected (FC) layer. To optimize the parameters in the object expert, we compute a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5677" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "cross-entropy loss with an object classifier whose weights are initialized from object word-embeddings as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 104, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 104, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 64, + 104, + 287, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o b j}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {o} , w \\left(o _ {i}\\right)\\right) / \\tau_ {o} \\right\\}}{\\sum_ {o _ {k} \\in \\mathcal {O}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {o} , w \\left(o _ {k}\\right)\\right) / \\tau_ {o} \\right\\}}, \\tag {2}", + "image_path": "b846a847820ac09d2dc703d7cadb2f8897e9541499cd9371fcbda2cf4c5712db.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "text", + "content": " is a temperature parameter, " + }, + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "inline_equation", + "content": "w(o)" + }, + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "text", + "content": " denotes the word-embedding [47] corresponding to the object " + }, + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "inline_equation", + "content": "o\\in \\mathcal{O}" + }, + { + "bbox": [ + 47, + 140, + 287, + 165 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 172, + 146, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 172, + 146, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 146, + 185 + ], + "type": "text", + "content": "3.2. Attribute expert" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 190, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 346 + ], + "type": "text", + "content": "Capturing both attribute and object in the same deeply condensed feature is nontrivial as they require pretty different visual representations [30, 22]. The straightforward solution is to design two-stream networks without any parameter sharing. However, such individual training could not capture the contextuality [39]. Hence, we build an attribute expert which utilizes semantics estimated from the object expert as a condition, and extracts attribute-related features on intermediate layers of visual backbone in a top-down fashion [31, 1]. We devise an object-guided attention module that leverages convolution kernel-based attention [71, 6, 24], which excels at capturing fine-grained local information required for attribute recognition [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": "Specifically, we first reshape " + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_p^n" + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_p^n" + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": " with output resolution of " + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "inline_equation", + "content": "H\\times W\\times D" + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": ". Following [51], we tile " + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_o" + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": " to match the spatial dimension with " + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_p^n" + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": " and concatenate them to obtain an object-contextualized feature " + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{U}^n = [\\mathbf{p}_o,\\mathbf{Z}_p^n ]\\} \\in \\mathbb{R}^{H\\times W\\times 2D}" + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": ". As shown in Fig. 2, we generate a channel attention map " + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_c^n\\in \\mathbb{R}^{1\\times 1\\times D}" + }, + { + "bbox": [ + 47, + 346, + 287, + 431 + ], + "type": "text", + "content": " from the averaged feature across the spatial axis:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 440, + 287, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 440, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 79, + 440, + 287, + 454 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {c} ^ {n} = \\sigma \\left(\\operatorname {c o n v} _ {1 \\times 1} \\left(\\operatorname {a v g p o o l} _ {\\text {s p a t i a l}} \\left(\\mathbf {U} ^ {n}\\right)\\right)\\right), \\tag {3}", + "image_path": "e423c654ca6f004c6b84f70cbc99327610bcad47af8e87e379c021ed36c3c6b4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 463, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 463, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 463, + 287, + 498 + ], + "type": "text", + "content": " denotes a sigmoid function. We also generate a spatial attention map " + }, + { + "bbox": [ + 47, + 463, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s^n\\in \\mathbb{R}^{H\\times W\\times 1}" + }, + { + "bbox": [ + 47, + 463, + 287, + 498 + ], + "type": "text", + "content": " from the averaged feature across the channel axis:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 508, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 508, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 77, + 508, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {s} ^ {n} = \\sigma \\left(\\operatorname {c o n v} _ {3 \\times 3} \\left(\\operatorname {a v g p o o l} _ {\\text {c h a n n e l}} \\left(\\mathbf {U} ^ {n}\\right)\\right)\\right). \\tag {4}", + "image_path": "9905cd79d3116133e9d4c6460f3689e145db25b41b15d0834ca3c5adcff527c0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "text", + "content": "Finally, we aggregate two attention maps into a unified attention map " + }, + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^n\\in \\mathbb{R}^{H\\times W\\times D}" + }, + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 563, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {A} ^ {n} = \\mathbf {A} _ {s} ^ {n} \\times \\mathbf {A} _ {c} ^ {n}. \\tag {5}", + "image_path": "dc28d09cfb921fc1eac0d95c6559dddb630371f0857db68dea8fbb61969d365c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 586, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 621 + ], + "type": "text", + "content": "The output of our object-guided attention module is defined as attribute features with estimated attention maps, followed by the residual connection:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 119, + 630, + 287, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 630, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 119, + 630, + 287, + 646 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {Z}} _ {p} ^ {n} = \\mathbf {Z} _ {p} ^ {n} + \\mathbf {A} ^ {n} \\odot \\mathbf {Z} _ {p} ^ {n}, \\tag {6}", + "image_path": "5e8f38860ff928af2f4d1bace249d4c6d2a0d8d7f6b5cbaa9eeaa89af904a645.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 654, + 276, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 276, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 276, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 276, + 665 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 654, + 276, + 665 + ], + "type": "text", + "content": " is an operation for element-wise multiplication." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Z}}_p^n" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": " from each intermediate layer, we ensemble the features from different blocks by taking advantage of the hierarchical architecture [33, 43]. It is required for recognizing heterogeneous types of attributes such as colors or" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "shapes [48]. Especially, we choose low-, middle-, and high-level features from the backbone, denoted as " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{Z}}_p^l,\\tilde{\\mathbf{Z}}_p^m,\\tilde{\\mathbf{Z}}_p^h\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ", and ensemble those triplets with global average pooling (GAP) followed by concatenation. The attribute expert " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " then generates an attribute embedding with the triplets as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 138, + 545, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 138, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 335, + 138, + 545, + 153 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {a} = E _ {a} \\left(\\left[ \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {l}\\right), \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {m}\\right), \\operatorname {G A P} \\left(\\tilde {\\mathbf {Z}} _ {p} ^ {h}\\right) \\right]\\right), \\tag {7}", + "image_path": "865612f58de37af5109d24c1ae0c0c282ce3348abf499ef34b976db417c43a65.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 158, + 545, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 158, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 545, + 183 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 158, + 545, + 183 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 305, + 158, + 545, + 183 + ], + "type": "text", + "content": " consists of a FC layer. Similar to the object expert, we optimize the attribute expert with the cross-entropy loss:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 322, + 198, + 545, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 198, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 322, + 198, + 545, + 227 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a t t}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {a} , w \\left(a _ {i}\\right)\\right) / \\tau_ {a} \\right\\}}{\\sum_ {a _ {k} \\in \\mathcal {A}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {a} , w \\left(a _ {k}\\right)\\right) / \\tau_ {a} \\right\\}}. \\tag {8}", + "image_path": "7d630a68df69f21644ff68fe689f9ed75c6d2c5c939eed6d92aa359a3a10bdad.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 232, + 476, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 476, + 245 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 476, + 245 + ], + "type": "text", + "content": "3.3. Mapping visual-semantic space" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 250, + 545, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 284 + ], + "type": "text", + "content": "The visual compositional embedding " + }, + { + "bbox": [ + 305, + 250, + 545, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_x" + }, + { + "bbox": [ + 305, + 250, + 545, + 284 + ], + "type": "text", + "content": " of the input image " + }, + { + "bbox": [ + 305, + 250, + 545, + 284 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 250, + 545, + 284 + ], + "type": "text", + "content": " is calculated from object and attribute experts, formulated as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 382, + 285, + 545, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 285, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 382, + 285, + 545, + 299 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {x} = E _ {c} \\left(\\left[ \\mathbf {p} _ {o}, \\mathbf {p} _ {a} \\right]\\right)). \\tag {9}", + "image_path": "52d7bdfed54204516093da79137db83d5e8d7bdb000b7476630bb1bca0b06390.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "content": "The embedding function " + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "content": " includes an FC layer, projecting the visual features into the joint embedding space that aligns visual and semantic representations. Following previous works [42, 36], we estimate semantic label embedding " + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_y" + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "content": " of each seen composition label " + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{V}_s" + }, + { + "bbox": [ + 305, + 302, + 545, + 374 + ], + "type": "text", + "content": " by projecting the concatenated word vectors into the joint space:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 378, + 380, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 380, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 378, + 380, + 545, + 396 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {y} = g ([ w (o), w (a) ]), \\tag {10}", + "image_path": "52ae812cbc4055c880d0b56ea5e59465614581e9d9daacaf9d8d0a83d340183e.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 400, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 545, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 400, + 545, + 449 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 305, + 400, + 545, + 449 + ], + "type": "text", + "content": " is a label embedding network that consists of 3 FC layers and ReLU activation function. During training, we minimize the cosine distance of visual and semantic embeddings with cross-entropy loss as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 455, + 545, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 455, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 318, + 455, + 545, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o m p}} = - \\log \\frac {\\exp \\left\\{\\cos \\left(\\mathbf {p} _ {x} , \\mathbf {s} _ {y}\\right) / \\tau_ {c} \\right\\}}{\\sum_ {y _ {k} \\in \\mathcal {Y} _ {s}} \\exp \\left\\{\\cos \\left(\\mathbf {p} _ {x} , \\mathbf {s} _ {y _ {k}}\\right)\\right) / \\tau_ {c} \\}, \\tag {11}", + "image_path": "5e22d08c4e7643cd2bb140554cbcf31ec3e850254cfe2b6312933260d9df22bf.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 488, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 488, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 545, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 488, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 305, + 488, + 545, + 522 + ], + "type": "text", + "content": " is a temperature parameter for the composition loss " + }, + { + "bbox": [ + 305, + 488, + 545, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{comp}}" + }, + { + "bbox": [ + 305, + 488, + 545, + 522 + ], + "type": "text", + "content": ". Finally, the total objectives of CoT is then formulated as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 361, + 531, + 545, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 531, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 361, + 531, + 545, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o m p}} + \\alpha \\mathcal {L} _ {\\text {a t t}} + \\beta \\mathcal {L} _ {\\text {o b j}}, \\tag {12}", + "image_path": "7b9ad009cf256f8cc729dc5b1a3d48ba7fa93fb8ba6ca2fc04716faf98f493df.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 551, + 451, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 451, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 451, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 551, + 451, + 562 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 551, + 451, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 551, + 451, + 562 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 551, + 451, + 562 + ], + "type": "text", + "content": " are balance weights." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 563, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 657 + ], + "type": "text", + "content": "During inference, we measure the cosine similarity between a visual embeddings and all label embeddings from " + }, + { + "bbox": [ + 305, + 563, + 545, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{te}}" + }, + { + "bbox": [ + 305, + 563, + 545, + 657 + ], + "type": "text", + "content": " and regard it as a feasibility score of the image and composition labels [36]. Following [80], we also use the classification scores from object and attribute experts as a form of cosine similarity. Therefore, a final feasibility score of label " + }, + { + "bbox": [ + 305, + 563, + 545, + 657 + ], + "type": "inline_equation", + "content": "y = (a,o)" + }, + { + "bbox": [ + 305, + 563, + 545, + 657 + ], + "type": "text", + "content": " is derived by adding the above three scores, formulated as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 665, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 665, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 665, + 545, + 689 + ], + "type": "interline_equation", + "content": "c (y) = \\cos \\left(\\mathbf {p} _ {x}, \\mathbf {s} _ {y}\\right) + \\cos \\left(\\mathbf {p} _ {o}, w (o)\\right) + \\cos \\left(\\mathbf {p} _ {a}, w (a)\\right), \\tag {13}", + "image_path": "80b312d63cdd70ee359f03ef4d90bf253e41c4ecf38c845f7752d33120999d68.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where we predict the label with the highest score as the final composition label." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5678" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 165, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 165, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 165, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 165, + 178 + ], + "type": "image", + "image_path": "c7e939d1597c896f8580d538c27c5397b248b3f0c572ad155a5902f77c49ca9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "text", + "content": "Figure 3: Attribute class distributions on VAW-CZSL. Every attributes in each plot are composed into a fixed object; (a) Grass and (b) Player." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 72, + 282, + 178 + ], + "blocks": [ + { + "bbox": [ + 168, + 72, + 282, + 178 + ], + "lines": [ + { + "bbox": [ + 168, + 72, + 282, + 178 + ], + "spans": [ + { + "bbox": [ + 168, + 72, + 282, + 178 + ], + "type": "image", + "image_path": "bd24f994b76243dcb3e9cc538ec455ae52d1e59cdf7f07d99b15bdde94160da6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 236, + 223, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 223, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 223, + 250 + ], + "type": "text", + "content": "3.4. Minority attribute augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 255, + 286, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 363 + ], + "type": "text", + "content": "With a carefully designed architecture, our CoT can generate distinctive visual features for CZSL. Nevertheless, the imbalanced composition samples in the training dataset restrict the performance since attributes often co-occur with certain objects in the real world [30]. For example in Fig. 3, 'Green Grass' is naturally more common than 'White Grass'. Such long-tailed data makes the deep models be over-confident in the composition with head attributes [81], and thus hurts the generalization on unseen compositions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": "To this end, we present a minority attribute augmentation (MAA) for training, which diversifies the minority composition. First, two compositional embeddings from COt are selected as inputs to the augmentation, where " + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{x_A}" + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{x_B}" + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": " share the same object class. Especially, to balance the major and minor attributes, we assign higher sampling weights on minority attributes with the weighted sampler [14]. Formally, let " + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\zeta_{o_i}" + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": " denote the number of attribute samples in the " + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": "-th attribute class composed on an object class " + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": ". The sampling weight of the attribute-object pair " + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\kappa(a_i,o)" + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": " is formulated by an inverse attribute class frequency according to:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 503, + 287, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 503, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 121, + 503, + 287, + 531 + ], + "type": "interline_equation", + "content": "\\kappa \\left(a _ {i}, o\\right) = \\frac {1 / \\zeta_ {o _ {i}}}{\\sum_ {i} 1 / \\zeta_ {o _ {i}}}. \\tag {14}", + "image_path": "3f7c973ae008e006ea68b8cd182ea8dbac4fa7bed357f04f01991dfcce319cad.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 534, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 571 + ], + "type": "text", + "content": "Inspired by Mixup [79, 64], we generate a virtual sample " + }, + { + "bbox": [ + 47, + 534, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{x_M}" + }, + { + "bbox": [ + 47, + 534, + 287, + 571 + ], + "type": "text", + "content": " with alpha compositing and its attribute semantic embedding " + }, + { + "bbox": [ + 47, + 534, + 287, + 571 + ], + "type": "inline_equation", + "content": "w(a_{M})" + }, + { + "bbox": [ + 47, + 534, + 287, + 571 + ], + "type": "text", + "content": " by linear combination:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 101, + 581, + 223, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 581, + 223, + 594 + ], + "spans": [ + { + "bbox": [ + 101, + 581, + 223, + 594 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {x _ {M}} = \\lambda \\mathbf {p} _ {x _ {A}} + (1 - \\lambda) \\mathbf {p} _ {x _ {B}},", + "image_path": "78e3194804b87c1802ff033c8dfc53e7bfb0c91bafc992b6bdfa4d1c1e6ccc83.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 595, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 595, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 91, + 595, + 287, + 609 + ], + "type": "interline_equation", + "content": "w \\left(a _ {M}\\right) = \\lambda w \\left(a _ {A}\\right) + (1 - \\lambda) w \\left(a _ {B}\\right). \\tag {15}", + "image_path": "38683d7305f99ec0f3a93b2007a9420b5051591296c64aff06286aa32dd24adc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 617, + 286, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 286, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 286, + 629 + ], + "type": "text", + "content": "The mixing factor " + }, + { + "bbox": [ + 47, + 617, + 286, + 629 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 617, + 286, + 629 + ], + "type": "text", + "content": " is sampled from a uniform distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "We notice that there are two significant differences from Mixup: (1) Compared to Mixup sampling arbitrary images, MAA constructs a composition pair fastening an object with the minority aware sampling to address the imbalanced data issue in CZSL. (2) Mixup mixes one-hot vector labels [79, 77] for classification, while MAA generates a virtual label in word vector space, to jointly regularize vi" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "sual and linguistic (label) embedding functions against the majority attribute classes with Eq. (12)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 104, + 467, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 104, + 467, + 116 + ], + "spans": [ + { + "bbox": [ + 306, + 104, + 467, + 116 + ], + "type": "text", + "content": "3.5. Extension to other backbones" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 122, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 547, + 277 + ], + "type": "text", + "content": "Our idea of CoT is generic and scalable, so we can simply extend to other backbones such as ResNet-18 (R18) [16] by selecting specific layers from the backbone composed of multiple layers or blocks. For a fair comparison to prior arts [41, 53, 23], we study the impact of our contribution on the ResNet backbone based on the convolutional operation. In specific, the object expert generates the object embedding from the last fifth convolution block. The classification token is replaced by a global average pooled feature. Similar to using ViT [11], we concatenate the intermediate layers into a feature vector for the attribute expert. The remained settings such as the architecture of " + }, + { + "bbox": [ + 304, + 122, + 547, + 277 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{comp}}" + }, + { + "bbox": [ + 304, + 122, + 547, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 122, + 547, + 277 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 122, + 547, + 277 + ], + "type": "text", + "content": " are the same as ViT." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 281, + 388, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 388, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 388, + 293 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 300, + 421, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 421, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 421, + 313 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 413 + ], + "type": "text", + "content": "Datasets. MIT-States [19] dataset is a common dataset in CZSL that most previous methods adopted, containing 115 states, 245 objects, and 1962 compositions. C-GQA [41] and VAW-CZSL [53] are two brand-new and large-scale benchmarks collected from in-the-wild images with more general compositions, such as 'Stained Wall' and 'Hazy Mountain'. Detailed data statistics including long-tailed distribution can be found in Appendix A." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 414, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 533 + ], + "type": "text", + "content": "Metrics. We follow common CZSL setting [49], testing performance on both seen and unseen pairs [5]. Under the setting, we report area under the curve (AUC) (\\%) between seen and unseen accuracies at different bias term in validation and test splits. We also compute the best seen and unseen accuracies in the curve, and calculate the best harmonic mean to measure the trade-off between the two values. Lastly, we follow [41] and report the attribute and object accuracies on unseen labels to measure the discrimination power of each representation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Baselines. We compare our CoT with 9 recent CZSL approaches: AoP [42], LE+ [39], TMN [49], Symnet [29], CompCos [36], CGE [41], SCEN [27], OADis [53], CAPE [23]. All baselines were built on pretrained ResNet-18 (R18) [16], with pre-trained word embedding such as GloVe [47], FastText [4], and Word2Vec [37]. To show the contribution of our proposed elements, we carry on the most recent methods including CGE and OADis with a transformer-based backbone (ViT-B) [11]. We also clarify that CGE and OADis utilize the knowledge of unseen composition labels at training for building a relation graph and hallucinating virtual samples respectively, while other methods do not assess the unseen composition label sets. The results of other baselines were obtained from [41, 27, 23, 53] or re-implemented from their official code." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5679" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 548, + 308 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 548, + 308 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 548, + 308 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 548, + 308 + ], + "type": "table", + "html": "
MethodsBackboneMIT-StatesC-GQAVAW-CZSL
V@1T@1HMSUV@1T@1HMSUV@3T@3HMSU
With frozen R18
AoP [42]R182.52.010.716.618.40.90.32.911.83.91.41.49.116.411.7
LE+ [39]R183.52.311.516.221.21.20.65.316.15.01.51.69.816.213.2
TMN [49]R183.32.611.822.717.12.21.17.721.66.32.22.311.919.915.4
Symnet [29]R184.53.413.824.820.03.31.89.825.29.22.32.312.219.115.8
CompCos [36]R186.94.816.926.924.53.52.612.128.111.83.13.214.223.918.0
CGE* [41]R187.25.318.128.925.03.62.511.927.511.72.72.913.023.416.8
SCEN [27]R187.25.318.429.925.24.02.912.428.912.1-----
OADis* [53]R187.65.918.931.125.6-----3.53.615.224.918.7
CAPE [23]R18-5.819.130.526.2-4.216.332.915.6-----
CoTR187.76.219.630.826.84.94.516.633.116.63.63.815.724.619.1
With frozen ViT-B
CGE* [41]ViT-B8.77.321.333.528.64.53.815.631.314.43.73.915.726.319.7
OADis* [53]ViT-B9.07.521.934.229.35.94.616.232.515.34.14.316.626.920.8
CoTViT-B9.27.823.234.831.56.75.117.534.018.84.44.717.726.922.2
With finetuning ViT-B
CGE* [41]ViT-B11.49.724.839.731.67.35.418.538.017.15.96.220.130.125.7
OADis* [53]ViT-B11.810.125.239.232.18.17.020.138.319.86.26.520.431.326.1
CoTViT-B12.110.525.839.533.08.77.422.139.222.76.67.221.732.928.2
", + "image_path": "8763f875621c4bb45e9e747809b95a083277926e9c53c14f6e7f824a82c47a16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 311, + 547, + 356 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 547, + 356 + ], + "type": "text", + "content": "Table 1: Quantitative comparison with prior arts. Following [41, 53], we measure AUC values on validation (V) and test (T) datasets, at top-1 (@1) on C-GQA and top-3 (@3) on VAW-CZSL, respectively. In addition, we report harmonic mean (HM), seen (S), unseen (US), attribute (A), and object (O) accuracies. * denotes the models accessing unseen composition labels at training. The best result and the second best result are boldfaced and underlined, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 376, + 178, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 178, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 178, + 388 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": "Models. We use ViT-B [11] and ResNet18 [16] pre-trained on ImageNet [9] as a visual backbone. In ViT-B backbone, Each patch size is " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": ", and the number of patch tokens in each layer is " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": ", having 768 channels per token. The label embedding network " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": " is composed of 3 FC layers, two dropout layers with ReLU activation functions, where hidden dimensions are 900. We use one FC layer for " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "E_{o}" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": " to produce 300-dimensional prototypical vectors that are matched with GloVe [47] embedding vectors from composition labels. Therefore, both " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": " produce 300-dimensional embedding vectors, with " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": " normalization. In the attribute expert, we use the outputs of the 3rd, 6th and 9th blocks (denoted " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{l}, \\mathbf{z}^{m}, \\mathbf{z}^{h}" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": " respectively) for multi-level feature fusion. For ResNet18, we use the 2nd, 3rd and 4th blocks as " + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{l}, \\mathbf{z}^{m}, \\mathbf{z}^{h}" + }, + { + "bbox": [ + 46, + 398, + 288, + 589 + ], + "type": "text", + "content": ", resulting in a 438-dimensional feature vector." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Training setup. We train CoT with Adam optimizer [25]. The input images are augmented with random crop and horizontal flip, and finally resized into " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". In ViT, learning rate is initialized to 1e-4 for all three datasets, and decayed by a factor of 0.1 for 10 epochs. In R18, the initial learning rate is 5e-5 with the same decay parameter of the ViT setting. Note that we freeze the GloVe [47] word embedding " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " for fair comparisons. For all datasets, we use the same temperature parameters [36, 53, 27] of each cross-entropy loss, " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{o}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{a}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " as 0.05, 0.01 and 0.01 respec" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 374, + 537, + 446 + ], + "blocks": [ + { + "bbox": [ + 315, + 374, + 537, + 446 + ], + "lines": [ + { + "bbox": [ + 315, + 374, + 537, + 446 + ], + "spans": [ + { + "bbox": [ + 315, + 374, + 537, + 446 + ], + "type": "table", + "html": "
MethodsMIT-StatesC-GQAVAW-CZSL
Attr.Obj.Attr.Obj.Attr.Obj.
CGE [41]35.744.417.534.624.551.6
OADis [53]35.245.214.742.022.353.8
CoT37.346.019.840.227.354.2
", + "image_path": "e5a931173816bbd0bbdf14d9db7c463d6c762920195841d22defbfeda577c3f8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 451, + 545, + 484 + ], + "lines": [ + { + "bbox": [ + 305, + 451, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 545, + 484 + ], + "type": "text", + "content": "Table 2: Comparison of unseen attribute (Attr.) and object (Obj.) accuracies with the latest CZSL methods [41, 53]. All results are reported on fine-tuned ViT backbone." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 506, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 555 + ], + "type": "text", + "content": "tively. We use the loss balance weights " + }, + { + "bbox": [ + 305, + 506, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 506, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 506, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 506, + 545, + 555 + ], + "type": "text", + "content": " as (0.5, 0.5) on C-GQA, and (0.4, 0.6) on VAW-CZSL and MIT-States. We note that to prevent overfitting of COT we apply MAA from 15 and 30 epochs to ViT and R18, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 563, + 405, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 405, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 405, + 574 + ], + "type": "text", + "content": "4.3. Main evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 581, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 663 + ], + "type": "text", + "content": "Table 1 compares generalization performance between our methods and the baselines with ResNet18 [16] (R18) and ViT-B [11] backbone setting. To further compare visual discrimination, we also compare unseen attribute and object accuracies of our methods and the latest CZSL methods [41, 53] in Table 2. In the following, we analyze the results from the three datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "MIT-States. From Table 1, the CoT achieves the best AUC values (V@1 and H@1) and harmonic mean (HM) with both R18 and ViT backbone. These results sufficiently demonstrate our experts' applicability on the CNN-based" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5680" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 78, + 210, + 140 + ], + "blocks": [ + { + "bbox": [ + 111, + 78, + 210, + 140 + ], + "lines": [ + { + "bbox": [ + 111, + 78, + 210, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 78, + 210, + 140 + ], + "type": "table", + "html": "
CoTMAAAUCHM
5.8118.1
✓6.0519.0
✓6.9320.8
✓✓7.2021.7
", + "image_path": "a59829124f7d95109ce6246478b586fdf86ebe5ce84113faa9412d865358aca0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 237, + 78, + 351, + 140 + ], + "blocks": [ + { + "bbox": [ + 119, + 144, + 199, + 153 + ], + "lines": [ + { + "bbox": [ + 119, + 144, + 199, + 153 + ], + "spans": [ + { + "bbox": [ + 119, + 144, + 199, + 153 + ], + "type": "text", + "content": "(a) Component analysis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 237, + 78, + 351, + 140 + ], + "lines": [ + { + "bbox": [ + 237, + 78, + 351, + 140 + ], + "spans": [ + { + "bbox": [ + 237, + 78, + 351, + 140 + ], + "type": "table", + "html": "
Loss componentAUCHM
Lcomp5.9418.9
Lcomp+Lobj6.2219.6
Lcomp+Lattr6.9121.4
Lcomp+Lattr+Lobj7.2021.7
", + "image_path": "2c1c66eba70725811a7cdf5bdb5455fe7f5c762d23575bf26dcbd980259b8dc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 376, + 78, + 496, + 140 + ], + "blocks": [ + { + "bbox": [ + 246, + 144, + 339, + 153 + ], + "lines": [ + { + "bbox": [ + 246, + 144, + 339, + 153 + ], + "spans": [ + { + "bbox": [ + 246, + 144, + 339, + 153 + ], + "type": "text", + "content": "(b) Impact of loss functions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 78, + 496, + 140 + ], + "lines": [ + { + "bbox": [ + 376, + 78, + 496, + 140 + ], + "spans": [ + { + "bbox": [ + 376, + 78, + 496, + 140 + ], + "type": "table", + "html": "
AugmentationAUCHM
Manifold Mixup [64]5.9318.8
CutMix [77]6.9520.7
Mixup [79]7.0421.0
MAA7.2021.7
", + "image_path": "9f3a6626fc535f0416c7d051dd2d283dda02ba8acefb6b8724817de1fcba2e13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 86, + 163, + 277, + 217 + ], + "blocks": [ + { + "bbox": [ + 380, + 144, + 488, + 153 + ], + "lines": [ + { + "bbox": [ + 380, + 144, + 488, + 153 + ], + "spans": [ + { + "bbox": [ + 380, + 144, + 488, + 153 + ], + "type": "text", + "content": "(c) Impact of data augmentation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 163, + 277, + 217 + ], + "lines": [ + { + "bbox": [ + 86, + 163, + 277, + 217 + ], + "spans": [ + { + "bbox": [ + 86, + 163, + 277, + 217 + ], + "type": "table", + "html": "
GuidanceAUCHMSUAO
w/o guidance6.720.530.725.822.553.4
Attribute guidance6.921.031.526.425.753.9
Object guidance7.221.732.928.227.354.2
", + "image_path": "e1c1210f43a59659bd5099fd82035510378a1be359ba4f90438069af4a63675d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 288, + 159, + 516, + 222 + ], + "blocks": [ + { + "bbox": [ + 131, + 219, + 231, + 229 + ], + "lines": [ + { + "bbox": [ + 131, + 219, + 231, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 219, + 231, + 229 + ], + "type": "text", + "content": "(d) Impact of object guidance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 288, + 159, + 516, + 222 + ], + "lines": [ + { + "bbox": [ + 288, + 159, + 516, + 222 + ], + "spans": [ + { + "bbox": [ + 288, + 159, + 516, + 222 + ], + "type": "table", + "html": "
EnsemblezlzmzhAUCHMSUAO
Low2346.620.730.328.725.852.3
Mid5676.921.231.929.725.255.9
High910116.821.433.426.624.154.5
Mixture3697.221.732.928.227.354.2
", + "image_path": "d046198ae3328d3c1040b054f3eac870f243543573d13eca28adadf362b1e513.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 286, + 224, + 514, + 235 + ], + "lines": [ + { + "bbox": [ + 286, + 224, + 514, + 235 + ], + "spans": [ + { + "bbox": [ + 286, + 224, + 514, + 235 + ], + "type": "text", + "content": "(e) Different configurations of intermediate blocks in attribute experts." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 58, + 240, + 533, + 251 + ], + "lines": [ + { + "bbox": [ + 58, + 240, + 533, + 251 + ], + "spans": [ + { + "bbox": [ + 58, + 240, + 533, + 251 + ], + "type": "text", + "content": "Table 3: Ablation experiments on VAW-CZSL dataset. Note that all results are reported on CoT (ViT-B), with a fine-tuning setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 272, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 344 + ], + "type": "text", + "content": "backbone. With a finetuned ViT-B, the CoT obtains the best validation AUC of " + }, + { + "bbox": [ + 46, + 272, + 287, + 344 + ], + "type": "inline_equation", + "content": "12.1\\%" + }, + { + "bbox": [ + 46, + 272, + 287, + 344 + ], + "type": "text", + "content": " and test AUC of " + }, + { + "bbox": [ + 46, + 272, + 287, + 344 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 46, + 272, + 287, + 344 + ], + "type": "text", + "content": ", surpassing previous CZSL methods with a large margin. Our method also improves attribute (Attr.) and object (Obj.) accuracies as in Table 2, showing the effectiveness of CoT for visual discrimination." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "content": "C-GQA. Similar trends can be observed in the C-GQA dataset. In the frozen R18 setting, the CoT achieves the best test AUC (T@1) of " + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "content": ", comparable with previous state-of-the-art CAPE. Although the T@1 scores from CGE and OADis are also increased upon replacing the backbone from R18 to ViT-B, our CoT with ViT-B outperforms OADis with larger margins in all metrics, as compared to the performance with R18. It is noticeable that our proposed augmentation technique contributes the unbiased prediction, improving unseen accuracies into " + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "inline_equation", + "content": "18.8\\%" + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "content": " with a minority-aware sampling. The performance regarding AUC is improved with end-to-end training with backbone, " + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "content": " in V@1 and " + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "content": " in T@1. Of note, OADis and CGE access unseen composition labels at training, so it is unfair to directly compare with other methods including CoT. Nevertheless, more surprisingly, our CoT shows comparable and superior performance to OADis. Finally, in Table 2, our model also improves the attribute accuracy by achieving " + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "content": ", having comparable " + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "inline_equation", + "content": "40.2\\%" + }, + { + "bbox": [ + 46, + 349, + 287, + 589 + ], + "type": "text", + "content": " object accuracy compared to OADis and CGE." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "VAW-CZSL. We also achieve SoTA performance on VAW-CZSL, the most challenging benchmark having almost 10K composition labels for evaluation. Even if we do not use additional unseen composition information like OADis, our method achieves the highest T@1 of " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " with the best unseen label accuracy of " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "28.2\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". We also obtain both the highest attribute and object score (Table 2), especially with a significant improvement in attribute prediction by " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " compared to CGE. It proves that the careful modeling of the visual network in CoT is essential to substantially improve" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 272, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 309 + ], + "type": "text", + "content": "the generalization capability. In addition, the results suggest that finetuning the visual backbone to suit composition recognition improves the performance in all benchmarks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 319, + 397, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 397, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 397, + 332 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 339, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 375 + ], + "type": "text", + "content": "We ablate our CoT and MAA with different design choices on Table 3. We use the VAW-CZSL dataset and fine-tuned ViT-B backbone as our main ablation setting." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": "Component analysis. We report ablation results of CoT and MAA in Table 3a. In 'without CoT' case, we use CompCos [36] baseline. CoT boosts the AUC and HM significantly about " + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": " compared to the baseline. MAA offers better results to both CoT and baseline, showing scalability to other methods. Notably, employing MAA on CoT gives higher gains of " + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": " AUC and " + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": " HM compared to the baseline, demonstrating that the two components create a synergy effect." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "content": "Impact of loss functions. Table 3b shows the impact of each loss function. Compared to the case using " + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{comp}}" + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "content": " only, " + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{attr}}" + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "content": " bring a gain on both AUC and HM. Using both losses to " + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{comp}}" + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "content": " improves AUC and HM significantly by " + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "inline_equation", + "content": "1.26\\%" + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 485, + 545, + 557 + ], + "type": "text", + "content": " respectively, showing the importance of object and attribute losses to learn each expert." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Different data augmentations. In Table 3c, we compare MAA with various data mixing methodologies including Manifold Mixup [64], CutMix [77] and Mixup [79]. To apply Manifold Mixup in CoT, we select 3,6,9 and the last layer as a mixup layer, and mix these intermediate features. MAA achieves the highest score on both AUC and HM, outperforming other augmentation methods including standard Mixup. This demonstrates the effectiveness of MAA which balances the data distribution with a composition pair. Interestingly, we found that Manifold Mixup performs at worst, even degrading pure CoT performance (AUC: 6.93, HM: 20.4) in Table 3a. We speculate that the mixed representation at intermediate layers can lead to significant underfit" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5681" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 171 + ], + "type": "image", + "image_path": "f42adfde4d5eb773bcde54cab35a89d2ebb499cb29e0c8ccd614af226591ee50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 179, + 287, + 256 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 287, + 256 + ], + "type": "text", + "content": "Figure 4: Distribution of k-occurrence counts with " + }, + { + "bbox": [ + 46, + 179, + 287, + 256 + ], + "type": "inline_equation", + "content": "k = 20" + }, + { + "bbox": [ + 46, + 179, + 287, + 256 + ], + "type": "text", + "content": " of C-GQA and VAW-CZSL test set. All results are reported with ViT-B backbone. x-axis stands for composition class ids which are ordered by decreasing count numbers. For better visualization, we show the top " + }, + { + "bbox": [ + 46, + 179, + 287, + 256 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 179, + 287, + 256 + ], + "type": "text", + "content": " head composition classes where the hubness is predominant. Circles at start points indicate the most frequent hubs. The result on MIT-States is in the Appendix C.1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 278, + 163, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 163, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 163, + 289 + ], + "type": "text", + "content": "ting of the attention module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "text", + "content": "Impact of object guidance. As shown in Table 3d, the object guidance significantly boosts the performance. We found that reversing the guidance (attribute guidance) decreases the performance of AUC by " + }, + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "text", + "content": " and HM by " + }, + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "text", + "content": ". The results validate the usage of object guidance for contextualized attribute prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": "Design choice of feature ensemble. Table 3e shows the ablations with different configurations of intermediate blocks used in the attribute expert. We evaluate four types of block feature ensembles: low, mid, high, and mixture. Among them, the mixture type achieves the best performance on AUC, harmonic mean, and attribute accuracy. We therefore choose this mixture ensemble as a default setting. Another finding is low-level feature is more adequate for attribute and unseen composition recognition, while high-level feature is robust to object and seen composition accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 502, + 118, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 118, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 118, + 514 + ], + "type": "text", + "content": "4.5. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "Mitigating hubness problem. We illustrate the hubness effect of a baseline [36] and our method on C-GQA and VAW-CZSL datasets in Fig. 4. Concretely, we count " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "-occurrences " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "(N_{k})" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": ", the number of times that a sample appears in the " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " nearest neighbors of other points [50] for all class ids, and calculate the skewness of " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " distribution for hubness measurement. We observe that CoT significantly alleviates the hubness problem of baseline across datasets, enlarging the visual discrimination in the embedding space. MAA further reduces the skewness of the distribution by mitigating biased prediction on head composition classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Object-guided attention visualization. Qualitatively, we visualize attention maps from the object-guided attention module in Fig. 5. We observe that the attention module could capture the contextualized meaning of attributes in composition, e.g., wet skin of horse vs. puddle in street. It" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 544, + 140 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 140 + ], + "type": "image", + "image_path": "7f276bc4dd14e73b5ec5444822ca5e18bbb2920da2515767590b10ed3438525e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 144, + 543, + 215 + ], + "blocks": [ + { + "bbox": [ + 307, + 144, + 543, + 215 + ], + "lines": [ + { + "bbox": [ + 307, + 144, + 543, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 144, + 543, + 215 + ], + "type": "image", + "image_path": "d6a7ea8ed80d52633e8d036bb4ca5cb40b471006edd68918b7a3fe6da80422f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 218, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 262 + ], + "type": "text", + "content": "Figure 5: Visualization of object-guided attention on VAW-CZSL dataset. Each row represents the attended regions in image pairs, which are related to different attributes contextualized with a specific object. More results in the Appendix C.5." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 292, + 544, + 371 + ], + "blocks": [ + { + "bbox": [ + 310, + 292, + 544, + 371 + ], + "lines": [ + { + "bbox": [ + 310, + 292, + 544, + 371 + ], + "spans": [ + { + "bbox": [ + 310, + 292, + 544, + 371 + ], + "type": "image", + "image_path": "3ba0c5c88fa9fe2ea9383b762b07160bc76ad1a2070cca0500779df7cc311e2f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 372, + 392, + 380 + ], + "lines": [ + { + "bbox": [ + 317, + 372, + 392, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 392, + 380 + ], + "type": "text", + "content": "Graite Ground" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 389, + 389, + 430 + ], + "blocks": [ + { + "bbox": [ + 334, + 381, + 342, + 388 + ], + "lines": [ + { + "bbox": [ + 334, + 381, + 342, + 388 + ], + "spans": [ + { + "bbox": [ + 334, + 381, + 342, + 388 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 389, + 389, + 430 + ], + "lines": [ + { + "bbox": [ + 310, + 389, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 310, + 389, + 389, + 430 + ], + "type": "image", + "image_path": "cabdee7b0e0a882dd56a5548893f2db4e16a43e51e71986cc57a43bfe6f40bff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 430, + 342, + 438 + ], + "lines": [ + { + "bbox": [ + 334, + 430, + 342, + 438 + ], + "spans": [ + { + "bbox": [ + 334, + 430, + 342, + 438 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 445, + 545, + 478 + ], + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 478 + ], + "type": "text", + "content": "Figure 6: Ground-truth composition labels and top-3 prediction results on MIT-States, C-GQA and VAW-CZSL datasets. More results in the Appendix C.6." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 395, + 389, + 465, + 430 + ], + "blocks": [ + { + "bbox": [ + 395, + 389, + 465, + 430 + ], + "lines": [ + { + "bbox": [ + 395, + 389, + 465, + 430 + ], + "spans": [ + { + "bbox": [ + 395, + 389, + 465, + 430 + ], + "type": "image", + "image_path": "7f38e241c7dbebadba09033625aaf18fbe5a4d186c0a6bb4a5c8193c2a68fcea.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 430, + 422, + 437 + ], + "lines": [ + { + "bbox": [ + 413, + 430, + 422, + 437 + ], + "spans": [ + { + "bbox": [ + 413, + 430, + 422, + 437 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 472, + 389, + 541, + 430 + ], + "blocks": [ + { + "bbox": [ + 472, + 389, + 541, + 430 + ], + "lines": [ + { + "bbox": [ + 472, + 389, + 541, + 430 + ], + "spans": [ + { + "bbox": [ + 472, + 389, + 541, + 430 + ], + "type": "image", + "image_path": "439a78825ff7be7f6787cad581a1dc46ecaa9cf97ac9e76f871459c5c423b0a7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 430, + 498, + 437 + ], + "lines": [ + { + "bbox": [ + 489, + 430, + 498, + 437 + ], + "spans": [ + { + "bbox": [ + 489, + 430, + 498, + 437 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 514, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 545, + 550 + ], + "type": "text", + "content": "validates that the attention module assists CoT to extract distinctive attribute embedding for composition by localizing specific RoI." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "Qualitative results. Fig. 6 illustrates the qualitative results of CoT. Even though some predictions do not correspond to annotated image labels, they explain the composition in another view; e.g., 'Furry Dog' and 'Shaggy Dog' in Fig. 6 (i). Moreover, in-the-wild images usually have multiple objects with their own attributes; e.g., 'White Wall' and 'Reading Man' in Fig. 6 (h). To recognize such compositions in the real world, it is necessary to handle the multi-label nature of composition, while distinguishing each attribute-object interaction [17] in an image. It encourages us to construct a new CZSL benchmark along with a novel evaluation metric handling both multi-label prediction and multiple compositions in the image. We leave it as a future work." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5682" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 88, + 287, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 88, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 88, + 287, + 208 + ], + "type": "text", + "content": "In this paper, we propose Composition Transformer (CoT) with simple minority attribute augmentation (MAA), a contextual, discriminative, and unbiased CZSL framework. The proposed object and attribute experts hierarchically utilize the entire powerful visual backbone to generate a composition embedding. The simple but effective MAA balances the long-tail distributed composition labels. Extensive studies with comprehensive analysis demonstrate the effectiveness of each component, and show that CoT surpasses existing methods across several benchmarks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 213, + 288, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 288, + 260 + ], + "type": "text", + "content": "Acknowledgement. This research was supported by the Yonsei Signature Research Cluster Program of 2022 (2022-22-0002) and the KIST Institutional Program (Project No.2E31051-21-203)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 273, + 107, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 273, + 107, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 273, + 107, + 285 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 293, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 287, + 357 + ], + "type": "text", + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 360, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 287, + 381 + ], + "type": "text", + "content": "[2] Arthur Asuncion and David Newman. Uci machine learning repository, 2007. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 415 + ], + "type": "text", + "content": "[3] Yuval Atzmon, Felix Kreuk, Uri Shalit, and Gal Chechik. A causal view of compositional zero-shot recognition. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 418, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 287, + 449 + ], + "type": "text", + "content": "[4] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. TACL, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 495 + ], + "type": "text", + "content": "[5] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In ECCV, 2016.3,5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 551 + ], + "type": "text", + "content": "[6] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001-3011, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 554, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 554, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 554, + 287, + 586 + ], + "type": "text", + "content": "[7] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 589, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 589, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 53, + 589, + 287, + 620 + ], + "type": "text", + "content": "[8] Yu-Ying Chou, Hsuan-Tien Lin, and Tyng-Luh Liu. Adaptive and generative zero-shot learning. In International conference on learning representations, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 623, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 623, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 53, + 623, + 287, + 654 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[10] Georgiana Dinu, Angeliki Lazaridou, and Marco Baroni. Improving zero-shot learning by mitigating the hubness problem. arXiv preprint arXiv:1412.6568, 2014. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "text", + "content": "[12] Nanyi Fei, Yizhao Gao, Zhiwu Lu, and Tao Xiang. Z-score normalization, hubness, and few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 142-151, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 193 + ], + "type": "text", + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "text", + "content": "[14] Yonatan Geifman and Ran El-Yaniv. Deep active learning over the long tail. arXiv preprint arXiv:1711.00941, 2017. 3,5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 248 + ], + "type": "text", + "content": "[15] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE TKDE, 2009. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 281 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 283, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 315 + ], + "type": "text", + "content": "[17] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In ECCV, 2020. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "text", + "content": "[18] Dat Huynh and Ehsan Elhamifar. Fine-grained generalized zero-shot learning via dense attribute-based attention. In CVPR, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "text", + "content": "[19] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "text", + "content": "[20] Mahdi M Kalayeh and Mubarak Shah. On symbiosis of attribute prediction and semantic segmentation. IEEE TPAMI, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "text", + "content": "[21] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Revisiting visual product for compositional zero-shot learning. In NeurIPSW, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[22] Shyamgopal Karthik, Massimiliano Mancini, and Zeynep Akata. Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 494, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 546 + ], + "type": "text", + "content": "[23] Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool, Alain Pagani, Didier Stricker, and Muhammad Zeshan Afzal. Learning attention propagation for compositional zero-shot learning. arXiv preprint arXiv:2210.11557, 2022. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[24] Hanjae Kim, Sunghun Joung, Ig-Jae Kim, and Kwanghoon Sohn. Prototype-guided saliency feature learning for person search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4865-4874, 2021. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 626, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 626, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 626, + 545, + 679 + ], + "type": "text", + "content": "[26] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[27] Xiangyu Li, Xu Yang, Kun Wei, Cheng Deng, and Muli Yang. Siamese contrastive embedding network for compositional zero-shot learning. In CVPR, 2022. 2, 5, 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5683" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[29] Yong-Lu Li, Yue Xu, Xiaohan Mao, and Cewu Lu. Symmetry and group in attribute-object compositions. In CVPR, 2020. 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[30] Kongming Liang, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Unifying visual attribute learning with object recognition in a multiplicative framework. IEEE TPAMI, 2018. 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "text", + "content": "[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8759-8768, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "text", + "content": "[34] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In ECCV, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[35] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "text", + "content": "[36] Massimiliano Mancini, Muhammad Ferjad Naeem, Yongqin Xian, and Zeynep Akata. Open world compositional zero-shot learning. In CVPR, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 433, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 464 + ], + "type": "text", + "content": "[37] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "text", + "content": "[38] Shaobo Min, Hantao Yao, Hongtao Xie, Chaoqun Wang, Zheng-Jun Zha, and Yongdong Zhang. Domain-aware visual bias eliminating for generalized zero-shot learning. In CVPR, pages 12664-12673, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "text", + "content": "[39] Ishan Misra, Abhinav Gupta, and Martial Hebert. From red wine to red tomato: Composition with context. In CVPR, 2017. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 567 + ], + "type": "text", + "content": "[40] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[41] Muhammad Ferjad Naeem, Yongqin Xian, Federico Tombari, and Zeynep Akata. Learning graph embeddings for compositional zero-shot learning. In CVPR, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[42] Tushar Nagarajan and Kristen Grauman. Attributes as operators: factorizing unseen attribute-object compositions. In ECCV, 2018. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[43] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, 2021. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[44] Zizheng Pan, Bohan Zhuang, Jing Liu, Haoyu He, and Jianfei Cai. Scalable vision transformers with hierarchical pool" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ing. In Proceedings of the IEEE/cvf international conference on computer vision, pages 377-386, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[45] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[46] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[47] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 1, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[48] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[49] Senthil Purushwalkam, Maximilian Nickel, Abhinav Gupta, and Marc'Aurelio Ranzato. Task-driven modular networks for zero-shot compositional learning. In ICCV, 2019. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[50] Milos Radovanovic, Alexandros Nanopoulos, and Mirjana Ivanovic. Hubs in space: Popular nearest neighbors in high-dimensional data. Journal of Machine Learning Research, 11(sept):2487-2531, 2010. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 354 + ], + "type": "text", + "content": "[51] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[52] Frank Ruis, Gertjan Burghouts, and Doina Bucur. Independent prototype propagation for zero-shot compositionality. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "type": "text", + "content": "[53] Nirat Saini, Khoi Pham, and Abhinav Shrivastava. Disentangling visual embeddings for attributes and objects. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[54] Dvir Samuel, Yuval Atzmon, and Gal Chechik. From generalized zero-shot learning to long-tail with class descriptors. In WACV, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[55] Nikolaos Sarafianos, Xiang Xu, and Ioannis A Kakadiaris. Deep imbalanced attribute classification using visual attention aggregation. In ECCV, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "text", + "content": "[56] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized zero-and few-shot learning via aligned variational autoencoders. In CVPR, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "text", + "content": "[57] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In ECCV, 2016. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 611 + ], + "type": "text", + "content": "[58] Ying Shu, Yan Yan, Si Chen, Jing-Hao Xue, Chunhua Shen, and Hanzi Wang. Learning spatial-semantic relationship for facial attribute recognition with limited labeled data. In CVPR, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "text", + "content": "[59] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[60] Chufeng Tang, Lu Sheng, Zhaoxiang Zhang, and Xiaolin Hu. Improving pedestrian attribute recognition with weakly-supervised multi-scale attribute-specific localization. In ICCV, 2019. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[61] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5684" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "bad momentum causal effect. Advances in Neural Information Processing Systems, 33:1513-1524, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 139 + ], + "type": "text", + "content": "[62] Bin Tong, Chao Wang, Martin Klinkigt, Yoshiyuki Kobayashi, and Yuuichi Nonaka. Hierarchical disentanglement of discriminative latent features for zero-shot learning. In CVPR, pages 11467-11476, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 184 + ], + "type": "text", + "content": "[63] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 239 + ], + "type": "text", + "content": "[64] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 286, + 262 + ], + "type": "text", + "content": "[65] Byron C Wallace and Issa J Dahabreh. Improving class probability estimates for imbalanced data. KAIS, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 286, + 319 + ], + "type": "text", + "content": "[66] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3784-3793, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 286, + 352 + ], + "type": "text", + "content": "[67] Xin Wang, Fisher Yu, Ruth Wang, Trevor Darrell, and Joseph E Gonzalez. Tafe-net: Task-aware feature embeddings for low shot learning. In CVPR, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 286, + 398 + ], + "type": "text", + "content": "[68] Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from imaginary data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7278-7286, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 286, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 431 + ], + "type": "text", + "content": "[69] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In CVPR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 286, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 286, + 465 + ], + "type": "text", + "content": "[70] Kun Wei, Muli Yang, Hao Wang, Cheng Deng, and Xianglong Liu. Adversarial fine-grained composition learning for unseen attribute-object recognition. In ICCV, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 286, + 498 + ], + "type": "text", + "content": "[71] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 286, + 544 + ], + "type": "text", + "content": "[72] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3907-3916, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 546, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 286, + 588 + ], + "type": "text", + "content": "[73] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 286, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 633 + ], + "type": "text", + "content": "[74] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature based activation map for weakly supervised object localization. In ICCV, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 677 + ], + "type": "text", + "content": "[75] Ziwei Xu, Guangzhi Wang, Yongkang Wong, and Mohan S. Kankanhalli. Relation-aware compositional zero-shot learning for attribute-object pair recognition. IEEE TMM, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[76] Muli Yang, Cheng Deng, Junchi Yan, Xianglong Liu, and Dacheng Tao. Learning unseen concepts via hierarchical decomposition and composition. In CVPR, 2020. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 251 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 3, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "text", + "content": "[78] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 3, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[80] Tian Zhang, Kongming Liang, Ruoyi Du, Xian Sun, Zhanyu Ma, and Jun Guo. Learning invariant visual representations for compositional zero-shot learning. arXiv preprint arXiv:2206.00415, 2022. 1, 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[81] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5685" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_content_list.json b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6915e29208e919b72eda9281649b46a070b37b48 --- /dev/null +++ b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_content_list.json @@ -0,0 +1,1483 @@ +[ + { + "type": "text", + "text": "Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jungho Lee $^{1}$ , Minhyeok Lee $^{1}$ , Dogyoon Lee $^{1}$ , Sangyoun Lee $^{1,2}$", + "bbox": [ + 225, + 202, + 743, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electrical and Electronic Engineering, Yonsei University $^{2}$ AIonFlow Research", + "bbox": [ + 217, + 238, + 751, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{2015142131,hydragon516,nemotio,sylee}@yonsei.ac.kr", + "bbox": [ + 240, + 277, + 727, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 327, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graph convolutional networks (GCNs) are the most commonly used methods for skeleton-based action recognition and have achieved remarkable performance. Generating adjacency matrices with semantically meaningful edges is particularly important for this task, but extracting such edges is challenging problem. To solve this, we propose a hierarchically decomposed graph convolutional network (HD-GCN) architecture with a novel hierarchically decomposed graph (HD-Graph). The proposed HD-GCN effectively decomposes every joint node into several sets to extract major structurally adjacent and distant edges, and uses them to construct an HD-Graph containing those edges in the same semantic spaces of a human skeleton. In addition, we introduce an attention-guided hierarchy aggregation (A-HA) module to highlight the dominant hierarchical edge sets of the HD-Graph. Furthermore, we apply a new six-way ensemble method, which uses only joint and bone stream without any motion stream. The proposed model is evaluated and achieves state-of-the-art performance on four large, popular datasets. Finally, we demonstrate the effectiveness of our model with various comparative experiments. Code is available at https://github.com/Jho-Yonsei/HD-GCN.", + "bbox": [ + 76, + 359, + 473, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 738, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human action recognition (HAR) is a task that categorizes action classes by receiving video data as input. HAR is used in many applications, such as human-computer interaction and virtual reality. Recently, several RGB-based and skeleton-based HAR methods have been proposed with the development of deep learning technology. However, RGB-based methods [31, 29] cannot robustly recognize human actions because they are strongly influenced by environmental noises such as background color, brightness of light,", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/561382c75aab24245f23319b495f169ee1faddbc9fa580c31210991f4432bc30.jpg", + "image_caption": [ + "Figure 1. The framework of HD-GCN. The input skeleton is applied with various edge sets through a hierarchically decomposed graph (HD-Graph). The red lines are the edges included in the corresponding hierarchy edge set. The network highlights the major edge sets through the attention map. The darker the color of red line, the more highlighted the edge set, and dotted lines denote unconnected edges." + ], + "image_footnote": [], + "bbox": [ + 521, + 329, + 879, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and clothing. Therefore, methods using skeleton modality [35, 24, 36, 26, 5, 4, 18, 2, 15] have attracted attention because they are not affected by these noises. These methods recognize action by receiving 2D or 3D coordinates of major human joints as time-series inputs.", + "bbox": [ + 496, + 612, + 893, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches [24, 18, 4, 2] have adopted graph convolutional networks (GCNs) to apply human-skeleton graphs to convolutional layers. However, existing GCN-based methods [35, 24, 25, 4, 2] have the following limitations. (1) With the widely used handcrafted graph, the relationships between distant joint nodes are not identified since they use only the relationships of PC edges in the human skeleton. Although the graph with PC edges has a semantic significance, the graph with only PC edges suffers from long-range dependency problem as they are heuristically fixed. However, for humans to recognize actions, relationships between structurally distant joints as well as between adjacent joints are strongly correlated. Although several methods [24, 2] have attempted to solve such limita", + "bbox": [ + 496, + 688, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10444", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion by training attention-guided learnable graphs, they still use [35]'s handcrafted graph with their learnable graphs. Moreover, as the element values of [35]'s graph are more dominant than those of the learnable graphs, they do not adequately highlight the relationships between distant nodes. (2) Some recent methods [35, 24, 4, 18] risk falling into suboptimality by simply aggregating the edge features and ignoring the contribution of each edge, thus incompletely recognizing which edges are significant for each skeleton sample. For example, in the case of a 'squat down' action, the interactions between the legs and arms should be highlighted.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these limitations, we propose a hierarchically decomposed graph convolutional network (HD-GCN) with a hierarchically decomposed graph (HD-Graph) and attention-guided hierarchy aggregation (A-HA) module. In addition, we present a six-way ensemble method to effectively utilize our HD-Graph. The framework of our proposed methods is shown in Fig. 1 for 'squat down' action.", + "bbox": [ + 75, + 275, + 472, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The HD-GCN incorporates GCNs with our HD-Graph, which identifies the relationships between distant joint nodes in the same semantic spaces (e.g. right and left hands, right and left feet). The same semantic spaces are formed by moving out step by step from the Center of Mass (CoM) node of the graph. For example, if belly is a CoM node, the first semantic space includes the belly node, the next space includes the chest and hip nodes, and the subsequent space includes the left and right shoulder and the left and right hip nodes. The nodes in the same semantic space are defined as hierarchy node set. To detect the relationships between distant joint nodes, network should have large receptive field. The proposed HD-Graph contains both meaningful adjacent and distant joint nodes by connecting all the nodes in neighboring hierarchy node sets and identifies the connectivity between those nodes for large receptive field. We adopt rooted tree-like structure to effectively represent every edges. We apply a spatial edge convolution (S-EdgeConv) layer to consider semantically close edges which cannot be captured by the HD-Graph for each sample. To create the S-EdgeConv layer, we borrow the structure of [33], which is widely used in 3D point clouds.", + "bbox": [ + 75, + 383, + 472, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To consider the contribution of each edge set, the process of selecting the dominant hierarchical information should depend on the action data sample to give proper attention to the most dominant edge sets. For example, in order to recognize the \"clapping\" action, a hierarchy edge set that includes both hands must be emphasized. To tackle this issue, we propose an attention-guided hierarchy aggregation (A-HA) module, which consists of two submodules: representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv). A scaling bias problem occurs if we use the spatial average pooling module without any node extraction process because each node has a different", + "bbox": [ + 75, + 719, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "number of adjacent nodes. To prevent this, we apply RSAP, which includes a representative node extraction process that triggers features after the pooling layer to represent each node. To effectively manage hierarchical features obtained by RSAP, we apply a hierarchical edge convolution (H-EdgeConv) layer. The H-EdgeConv treats each hierarchical feature as a graph node and identifies which hierarchical features should be highlighted via the Euclidean distance in feature space. With the RSAP and the H-EdgeConv, our model successfully determines which hierarchy edge sets and joints should be emphasized among the input features.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing ensemble method uses four-stream data composed of the joint, bone, joint motion, and bone motion streams, which are the original skeletal coordinates, spatial differential between joint coordinates, and temporal differential of joint, and temporal differential of the bone, respectively. Most existing ensemble methods [25, 2] use additional motion data, but models that solely utilize motion data exhibit relatively inferior performance. To address this problem, we present a new method, a six-way ensemble. We apply this ensemble method by setting three HDGraphs with joint and bone stream data. Each graph has different CoM nodes to extract features of different semantic spaces (see Appendix).", + "bbox": [ + 496, + 258, + 890, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on four benchmark action recognition datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Kinetics-Skeleton [11], and Northwestern-UCLA [30].", + "bbox": [ + 496, + 454, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 517, + 515, + 857, + 529 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a hierarchically decomposed graph (HDGraph) to thoroughly identify the significant distant edges between the same hierarchy node sets.", + "- We propose an attention-guided hierarchy aggregation (A-HA) module to highlight the key edge sets with representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv).", + "- We use a new six-way ensemble method for skeleton-based action recognition with HD-Graphs that have different center of mass (CoM), which outperforms regular ensemble without any motion data.", + "- Our HD-GCN outperforms the state-of-the-arts on four benchmarks for skeleton-based action recognition." + ], + "bbox": [ + 516, + 541, + 890, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 776, + 640, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Action Recognition with GCNs", + "text_level": 1, + "bbox": [ + 500, + 801, + 772, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In skeleton-based action recognition, human skeletal data are represented by a graph with joint nodes. Most recent approaches use GCN-based methods [35, 24, 5, 18, 2] with [35]'s graph structure, which identifies physical connections in the human skeleton. Those GCN-based meth", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10445", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6431cfc3343be6978763e7228e41897481c24b8d0b638e94f751810a862b6d27.jpg", + "image_caption": [ + "Figure 2. (a) Structure of HD-Graph with physically connected (PC) edges. The human skeleton graph is decomposed into a rooted tree, where PC edges are included in hierarchy sets. (b) Structure of HD-Graph with fully connected (FC) edges. Edges between all nodes in the same semantic space are obtained by connecting all the nodes in adjacent hierarchy edge sets. Blue and red lines stand for PC and FC edges, respectively." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 467, + 354 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f59d474ca5410eb86c120ce0161f0c1cd7bf7ed1832e26fb3834ac67d5a50a81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 874, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ods perform remarkably better than methods using handcrafted features [7, 8, 13, 20, 21]. They extract the spatial features representing the relationships between physically connected edges among human skeleton, and they outperform other methods by using them to construct the major relationships between joint nodes in the human skeleton. In particular, [25] and [2] propose adaptive attention-based graph structures to learn the sample-wise topological features. However, they might fall into suboptimality because they do not consider the physical prior of the human skeletal structure and allow too much flexibility in network training. To address this issue, we introduce a novel HD-Graph, referencing the known tendencies of human perception", + "bbox": [ + 75, + 436, + 472, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Attention Modules for Action Recognition", + "text_level": 1, + "bbox": [ + 76, + 670, + 434, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attention mechanism is an essential element for constructing a deep neural network. Using recent attention modules [10, 34], networks emphasize important information along a specific dimension. For example, Hu et al. [10] applies channel-wise attention, and Woo et al. [34] applies both channel-wise and spatial-wise attention. These techniques are divided into two categories for GCNs: (1) attention-based graph construction [24, 2] which is a method of forming topologies using a non-local block [32] or customized correlation matrices, and (2) spatial-wise, temporal-wise, channel-wise attention, which are commonly used attentions in [25, 27], and several other networks.", + "bbox": [ + 75, + 704, + 472, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 435, + 633, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Sec. 3.2, we detail the HD-Graph convolution to solve the problems of the conventional human-skeleton graph [35], which includes only PC edges. We also explain the A-HA module in Sec. 3.3 to highlight dominant hierarchical features. In Sec. 3.4, we replace the widely used four-stream ensemble method [25, 2, 4] with a six-way ensemble without motion data streams. Finally, we introduce the HD-GCN, which uses these proposed methods.", + "bbox": [ + 496, + 460, + 890, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 590, + 638, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notations. The spatio-temporal graph for human skeleton is represented by $\\mathcal{G}(\\mathcal{V},\\mathcal{E})$ , where $\\mathcal{V}$ and $\\mathcal{E}$ denote the joint and edge groups, respectively. Physically connected edges and fully-connected edges used in Sec. 3.2 are denoted as PC-edges and FC-edges, respectively.", + "bbox": [ + 496, + 614, + 890, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Graph Convolutional Networks. 3D time-series skeletal data are represented by $\\mathbf{X} \\in \\mathbb{R}^{3 \\times T \\times V}$ , where $V$ and $T$ are the number of joint nodes and the temporal window size, respectively. GCN's operation with input feature map $\\mathbf{F}_{in} \\in \\mathbb{R}^{C \\times T \\times V}$ is as follows:", + "bbox": [ + 496, + 709, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {o u t} = \\sum_ {s \\in S} \\widehat {\\mathbf {A}} _ {s} \\mathbf {F} _ {i n} \\Theta_ {s}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 795, + 890, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S = \\{s_{id}, s_{cf}, s_{cp}\\}$ denotes graph subsets, and $s_{id}, s_{cf},$ and $s_{cp}$ indicate identity, centrifugal, and centripetal joint subsets, respectively. $\\Theta_s$ denotes the pointwise convolution operation. The normalized adjacency matrix $\\widehat{\\mathbf{A}}$ is", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10446", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "initialized as $\\Lambda^{-\\frac{1}{2}}\\mathbf{A}\\Lambda^{-\\frac{1}{2}}\\in \\mathbb{R}^{N_S\\times V\\times V}$ , where $\\Lambda$ is a diagonal matrix for normalization and $N_{S} = 3$ .", + "bbox": [ + 75, + 88, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hierarchically Decomposed Graph", + "text_level": 1, + "bbox": [ + 76, + 132, + 382, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most recent methods have adopted the handcrafted graph proposed by Yan et al. [35], but the HD-Graph is derived through a newly presented method. Fig. 2 shows the framework of the HD-Graph.", + "bbox": [ + 75, + 156, + 468, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decomposition into a Rooted Tree. The first step is to decompose the graph with PC edges and construct a rooted tree. To decompose a given skeleton into the tree, we need to determine a CoM node, which allows nodes in the same hierarchy edge set to exist in the same semantic space. For example, nodes in the same semantic space, such as elbow and knee joints, or hands and feet, must exist in a hierarchy node set. After choosing the CoM node, the graph is converted into a rooted tree, which includes the hierarchical information of the graph, and defines the directed adjacency matrix $\\vec{\\mathbf{A}}_{\\mathrm{HD}} \\in \\mathbb{R}^{N_L \\times V \\times V}$ with $N_L$ hierarchy layers for $N_H$ hierarchy edge sets:", + "bbox": [ + 75, + 238, + 470, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overrightarrow {\\mathbf {A}} _ {\\mathrm {H D}} = \\left[ \\mathcal {E} \\left(H _ {1} \\rightarrow H _ {2}\\right), \\dots , \\mathcal {E} \\left(H _ {N _ {H} - 1} \\rightarrow H _ {N _ {H}}\\right)\\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 431, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H_{k}$ denotes the $k$ -th hierarchy node set and $\\mathcal{E}(H_k\\to$ $H_{k + 1})$ denotes a set of edges from $H_{k}$ to $H_{k + 1}$ . $N_{L}$ and $N_{H}$ are the number of hierarchy layers and hierarchy edge sets, respectively, and $N_{L} = N_{H} - 1$ . However, $\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{HD}}$ includes only the directed centrifugal edges. For consistency with existing methods, all the reverse-directed edges from the leaf nodes of the rooted tree in Fig. 2 to the CoM node must be reflected in the adjacency matrices to cover the centripetal edges. In addition, to get the features of the nodes themselves, the identity edges for each hierarchy node set must be considered. Thus, the adjacency matrices $\\stackrel {\\leftrightarrow}{\\mathbf{A}}_{\\mathrm{HD}}\\in \\mathbb{R}^{N_L\\times N_S\\times V\\times V}$ are defined as follows:", + "bbox": [ + 75, + 465, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overleftrightarrow {\\mathbf {A}} _ {\\mathrm {H D}} = \\left[ \\begin{array}{l} \\mathcal {E} _ {1}, \\mathcal {E} _ {2}, \\dots , \\mathcal {E} _ {N _ {L}} \\end{array} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 661, + 468, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {k} = \\mathcal {E} (\\underbrace {H _ {k} \\cup H _ {k + 1}} _ {s _ {i d}}, \\underbrace {H _ {k} \\rightarrow H _ {k + 1}} _ {s _ {c p}}, \\underbrace {H _ {k + 1} \\rightarrow H _ {k}} _ {s _ {c f}}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 723, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_k$ denotes the concatenation of the three edge subsets of $S = \\{s_{id}, s_{cp}, s_{cf}\\}$ and $s_{id}, s_{cp}, s_{cf}$ indicate the identity, centripetal, and centrifugal edge subsets, respectively. Through this construction policy, we create a skeletal graph with bidirectional and identity edges.", + "bbox": [ + 75, + 771, + 468, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fully Connected Inter-Hierarchy Edges. Decomposed graph $\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{HD}}$ has a different number of edge sets from the", + "bbox": [ + 75, + 868, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "conventional graph, but the edges are all the same. To identify the relationships between major distant joint nodes, especially those in the same semantic space, we connect all nodes between neighboring hierarchy node sets. In addition, since [35]'s graph contains the connectivity of only PC edges, not distant relationships, the receptive field is very small with this sparse graph. Applying our fully connected (FC) edges to the rooted tree, the graph becomes denser and makes the receptive field larger than before with more meaningful distant connectivity as shown in Fig. 2 (b). Then, the adjacency matrices are normalized with degree matrices for training stability and we leave all elements of the matrices as learnable parameters for training adaptability.", + "bbox": [ + 496, + 90, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HD-Graph Convolution. Our HD-Graph convolution includes four parallel branch operations: three graph convolution through HD-Graph and an additional EdgeConv [33] operation. To reduce the computational complexity, a linear transformation is applied to all four operations. For three of these operations, our method performs a subset-wise GCN operation in the same way as [35, 24] for each hierarchy edge set with three edge subsets. However, rather than summing the output values for each subset as in Eq. (1), we concatenate these output values to the channel dimension:", + "bbox": [ + 496, + 324, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} = \\left\\| _ {s \\in S} \\left\\{\\overleftrightarrow {\\mathbf {A}} _ {\\mathrm {H D}; s} ^ {(k)} \\Phi \\left(\\mathbf {F} _ {i n}\\right) \\Theta_ {s} ^ {(k)} \\right\\}, \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 486, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{\\mathrm{HD}}^{(k)}$ denotes the output feature map of the HD-Graph convolution and function $\\Phi$ denotes a linear transformation with parameter $\\mathbf{W} \\in \\mathbb{R}^{C' \\times C}$ . Note that $\\|\\cdot\\|$ is a concatenation operation.", + "bbox": [ + 496, + 536, + 890, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although our HD-Graph defines more meaningful node relationships than conventional graph, it may still not be able to extract sample-wise key relationships that reflect the similarities between all nodes in the feature space. To improve this limitation, we adopt EdgeConv [33] as the remaining operation, which is used for extracting graphical features through local neighborhood graphs in the feature space. With spatial EdgeConv (S-EdgeConv), our network extracts sample-wise node connectivity, which the HD-Graph does not capture. For our method, S-EdgeConv initially takes the average pooling as the temporal dimension for computational efficiency. Local graphs with local edges are then formed via k-nearest neighbor (k-NN) based on the Euclidean distance, and the local edges as well as identity edges based on the graphs are aggregated via trainable parameters $\\mathbf{W}_{edge} \\in \\mathbb{R}^{C' \\times 2C'}$ . For the deep neural network, physically close edges are reflected to the initial shallow layers, but as they become deeper, the relationship between semantically similar edges in the feature space are identified and learned. Our whole GCN process is shown in", + "bbox": [ + 496, + 599, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10447", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce2a61ad8f85e7fba75a75e2369c18ca05de9d336ba454eb005d37323ad5ec0b.jpg", + "image_caption": [ + "HD-Graph Conv. on $k$ th Hierarchy Set" + ], + "image_footnote": [], + "bbox": [ + 83, + 137, + 272, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/72adcd076ced693715d0de3a9a72bfbe3bd645a076bc60c736f8acf1d7bf9d91.jpg", + "image_caption": [ + "HD - Graph Conv. Block with A-HA" + ], + "image_footnote": [], + "bbox": [ + 282, + 133, + 455, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/921068da55285c260351013e722232f18c7728ce26637a2c2d5ca9b28f3de247.jpg", + "image_caption": [ + "Figure 3. HD-Graph convolution operation block with an A-HA module. The left side shows an operation for one hierarchy edge set, and the right side shows an operation block that concatenates the results for $N_{L}$ edge sets and applies A-HA. The lower part of the figure is EdgeConv, where the EdgeConv subscript indicates the feature space to extract graphical features. The $\\times$ and $\\cdot$ operations denote matrix and element-wise multiplication.", + "Fig. 3 and computed as follows:" + ], + "image_footnote": [], + "bbox": [ + 86, + 373, + 454, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {H D}} \\leftarrow \\sum_ {k = 1} ^ {N _ {L}} \\left[ \\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} \\left| \\right| z _ {\\mathcal {V}} ^ {(k)} \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} \\Phi (\\mathbf {F} _ {i n})\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 590, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{\\mathcal{V}}$ denotes the S-EdgeConv operation.", + "bbox": [ + 76, + 638, + 375, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All four branch outputs are concatenated to the channel dimension, with all four computed in the same way for $N_{L}$ hierarchy edge sets. Due to the inherent characteristics of skeletal data, the number of joint nodes included in each dataset is different, and, consequently, the number of hierarchy sets is different. Therefore, we adopt an addition policy for $N_{L}$ hierarchy-wise outputs and a concatenation policy for $N_{S}$ subset-wise outputs. In this way, the dimensionality is maintained, and the common hierarchy-wise aggregation policy is followed for every skeletal dataset by adding all the outputs for different numbers of hierarchical sets.", + "bbox": [ + 75, + 654, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Attention-Guided Hierarchy Aggregation", + "text_level": 1, + "bbox": [ + 76, + 830, + 429, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The HD-Graph convolution uses an aggregation policy of adding all the hierarchy-wise outputs. However, because each data sample has relationship between specific major", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "edges, we propose an attention-guided hierarchy aggregation (A-HA) module, which applies a weighted-sum policy to the hierarchy dimension with proper attention to the hierarchy-wise outputs. The framework of the HD-Graph convolution with the A-HA module is shown in Fig. 3.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Representative Spatial Average Pooling. Our A-HA module is applied to feature map $\\mathbf{F}_{\\mathrm{HD}} \\in \\mathbb{R}^{C \\times N_L \\times T \\times V}$ after the HD-Graph convolution. The first step is to extract the temporal frame with the highest score on $\\mathbf{F}_{\\mathrm{HD}}$ . RSAP, $\\Psi$ , is then applied, which is preceded by the extraction of representative nodes in each hierarchy layer. If spatial average pooling is applied without this extraction process, scaling bias occurs because the number of edges connected to each node is different. Therefore, representative node extraction is essential to obtain an appropriate score for attention without any bias. After the extraction, spatial average pooling is applied to hierarchy-wise outputs. Our pooling function $\\Psi$ is as follows:", + "bbox": [ + 496, + 185, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)}\\right) = \\frac {1}{N _ {k} + N _ {k + 1}} \\sum_ {v \\in H _ {k} \\cup H _ {k + 1}} \\max _ {t} \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} (v)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 386, + 890, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{k}$ denotes the number of vertices in the $H_{k}$ set.", + "bbox": [ + 500, + 429, + 864, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical Edge Convolution. After the RSAP layer, $N_{L}$ hierarchy-wise features in attention feature map $\\mathbf{M}$ have not yet shared their information with each other. We treat all $N_{L}$ features as nodes on a graph to learn and reflect similarities in the hierarchical feature space. To apply this process, representative features of these nodes are fed into EdgeConv [33], and the similarities of those nodes are learned based on the Euclidean distance. We also include the self-loop shown in the bottom section of Fig. 3 so that the node's own features can be reflected. Our attention map $\\mathbf{M}$ operates as follows:", + "bbox": [ + 496, + 462, + 890, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\sigma \\left(z _ {L} \\left(\\left. \\right| \\left. \\right| _ {k \\in L} \\left\\{\\Psi \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)}\\right) \\right\\}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 637, + 890, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{L}$ and $\\sigma$ denote H-EdgeConv and the sigmoid function, respectively.", + "bbox": [ + 496, + 689, + 888, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The attention map $\\mathbf{M}$ obtained is multiplied by the HD-Graph convolution output feature map $\\mathbf{F}_{\\mathrm{HD}}$ , and the output feature map $\\mathbf{F}_{out}$ is obtained through a weighted sum to the hierarchy axis as shown in Fig. 3. Similar to S-EdgeConv in Sec. 3.2, The H-EdgeConv method incorporates the concept of hierarchical edge sets in a physically proximal manner in earlier layers, while in the deeper layers, it emphasizes the presence of semantically similar edge sets. This approach highlights different hierarchical edge sets for each sample and enables the model to learn meaningful representations that capture both physical and semantic properties of the input data.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10448", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76c5616145acdd07134aec4bf6699640056bc71e696898ea50ed5c3115a3366e.jpg", + "image_caption": [ + "Figure 4. The architecture of HD-GCN. HD-GCN receives the skeleton sequence as input and obtains the class label through nine GCN blocks, and an FC layer, and the softmax function." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 467, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Six-Way Ensemble", + "text_level": 1, + "bbox": [ + 76, + 308, + 258, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shi et al. [24, 25] have applied a four-stream ensemble method using streams for joints, bones, joint motion, and bone motion. However, as the performances of motion streams are relatively poorer than the performances of joint and bone streams, we adopt an ensemble method with the joint and bone streams without any motion streams. We use three different HD-Graphs, and each graph is used for training with joint and bone streams. The three HD-Graphs have different CoM nodes, which are chest, belly, hip nodes, respectively. In other words, we train joint and bone streams with HD-Graph with the CoM node of chest, and we train the same when the CoM node is belly or hip node. As models with the three different graphs should be trained in different aspects, each of the graphs is composed of different edge sets. For example, if the CoM node is belly, both thigh edges and both upper arm edges are included in the same edge set, whereas when the CoM node is chest, both thigh edges and both forearm edges are included in the same edge set. The details of our six-way ensemble are specified on our Appendix.", + "bbox": [ + 75, + 333, + 472, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 648, + 282, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 4, we adopt [24] as our baseline network architecture with a total of nine stacked GCN blocks. The numbers of output channels for the blocks are 64, 64, 64, 128, 128, 128, 256, 256, and 256. Each block contains a residual connection [9] and is divided into a spatial module, in which the GCN operation proceeds, and a temporal module, which includes the temporal convolutions. Our method uses the temporal module of [2], whose baseline module is [18, 28]. This module consists of four branch operations. Two are dilated temporal convolutions with kernel size five and dilation one and two, respectively. The remaining branch operations are pointwise convolution and max pooling with kernel size three. Our spatial module consists of an HD-Graph convolution operation and an A-HA module, as introduced in Sec. 3.2 and Sec. 3.3. After passing", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "through all GCN layers with attention to the hierarchy-wise features, the network compresses the feature map through the global average pooling layer and classifies the action sample through the softmax function.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 166, + 633, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 191, + 808, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NTU-RGB+D 60. NTU-RGB+D 60 [22] is a large dataset used in skeletal action recognition. It contains 56,880 skeleton action samples, performed by 40 different participants and classified into 60 classes. The authors of this dataset recommend two benchmarks. (1) Cross-Subject (X-Sub): 20 of the 40 subjects' actions are used for training, and the remaining 20 are for validation. (2) Cross-View (X-View): Two of the three camera-views are used for training, and the other one is used for validation.", + "bbox": [ + 496, + 214, + 890, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NTU-RGB+D 120. NTU-RGB+D 120 [16] is a dataset in which 57,367 new action samples are added to the NTU-RGB+D 60 dataset. It contains a total of 114,480 skeleton action samples over 120 classes, performed by 106 different subjects. The authors of this dataset recommend two benchmarks: (1) Cross-Subject (X-Sub): 53 of the 106 subjects' actions are used for training, and the remaining 53 are used for validation. (2) Cross-Setup (X-Set): Of the 32 setups, data with even setup IDs are used for training, and the remaining data with odd IDs are used for validation.", + "bbox": [ + 496, + 369, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Kinetics-Skeleton. The Kinetics-Skeleton dataset is derived from the Kinetics 400 video dataset [11], utilizing the OpenPose pose estimation [1] to extract 240,436 training and 19,796 testing skeleton sequences across 400 classes. The dataset restricts the number of skeletons per time step to two and eliminates skeletons with lower confidence scores, ensuring high-quality sequences for human action recognition and pose estimation research.", + "bbox": [ + 496, + 542, + 890, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Northwestern-UCLA. The Northwestern-UCLA skeleton dataset [30] contains 1494 video clips over 10 classes. Each action is captured through three Kinect cameras with different camera views and is performed by 10 subjects. We adopt the same protocol as NW-UCLA: Two of the three camera-views are used for training, and the other one is used for validation.", + "bbox": [ + 496, + 675, + 890, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Settings. In our experiments, we adopt [24] as the backbone. The SGD optimizer is employed with a Nesterov momentum of 0.9 and a weight decay of 0.0004. The number of learning epochs is set to 90, with a warm-up strategy [9] applied to the first five epochs for more stable learning. We set the learning rate to decay with cosine annealing [19], with a maximum", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10449", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b8ada711316ec3e0ab452edc2763a14e1ed2dcf43ebdc56756aafb452c90fceb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPublicationMotion StreamNTU-RGB+D 60NTU-RGB+D 120Kinetics-SkeletonNorthwestern
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)Top-1 (%)Top-5 (%)UCLA (%)
ST-GCN [35]AAAI 2018✘81.588.370.773.230.752.8-
2s-AGCN [24]CVPR 2019✘88.595.182.584.236.158.7-
SGN [36]CVPR 2020✓89.094.579.281.5--92.5
AGC-LSTM [26]CVPR 2019✘89.295.0----93.3
DGNN [23]CVPR 2019✓89.996.1--36.959.6-
Shift-GCN [5]CVPR 2020✓90.796.585.987.6--94.6
DC-GCN+ADG [4]ECCV 2020✓90.896.686.588.1--95.3
DDGCN [14]ECCV 2020✓91.197.1--38.160.8-
MS-G3D [18]CVPR 2020✘91.596.286.988.438.060.9-
MST-GCN [3]AAAI 2021✓91.596.687.588.838.160.8-
CTR-GCN [2]ICCV 2021✓92.496.888.990.6--96.5
EfficientGCN-B4 [27]TPAMI 2022✓91.795.788.389.1---
STF [12]AAAI 2022✘92.596.988.989.939.9--
InfoGCN (4-ensemble) [6]CVPR 2022✓92.796.989.490.7--96.6
InfoGCN (6-ensemble) [6]CVPR 2022✓93.097.189.891.2--97.0
HD-GCN (2-ensemble)✘92.496.689.190.638.961.796.6
HD-GCN (4-ensemble)✘93.097.089.891.240.363.096.9
HD-GCN (6-ensemble)✘93.497.290.191.640.963.597.2
", + "bbox": [ + 81, + 90, + 890, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learning rate of 0.1 and a minimum learning rate of 0.0001. For the NTU-RGB+D datasets, we set the batch size to 64 and use the data preprocessing method from [36]. For Kinetics-Skeleton, the batch size is set to 128. In addition, to overcome the absence of belly and hip nodes in the Kinetics-Skeleton, we define the center of both hip joints as CoM hip node, and the center of chest and the hip node as CoM belly node, resulting in a total of 20 nodes. For the Northwestern-UCLA dataset, we set the batch size to 16 and use the data preprocessing method from [5]. All our experiments are conducted on a single RTX 3090 GPU.", + "bbox": [ + 75, + 386, + 468, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Arts Methods", + "text_level": 1, + "bbox": [ + 76, + 570, + 450, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Most recent state-of-the-art networks [25, 5, 4, 2] adopt a four-way ensemble method, but we adopt the six-way ensemble method described in Sec. 3.4.", + "bbox": [ + 76, + 595, + 468, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare ours with state-of-the-art networks on three datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Northwestern-UCLA [30], and Kinetics-Skeleton [11]. Comparisons for each dataset are shown in Tab. 1. The recognition performance of our HD-GCN has exceeded the state-of-the-arts on every dataset without any motion streams, as shown in Tab. 1. With our proposed ensemble method, HD-GCN outperforms the state-of-the-art and shows comparable performance to the 6-way ensemble state-of-the-art using only 4-way ensemble method.", + "bbox": [ + 75, + 645, + 470, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 814, + 230, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we demonstrate the effectiveness of the proposed HD-GCN. Performance is specified as the cross-subject and cross-setup classification accuracy on the NTU-RGB+D 120 [16] joint stream data.", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a54ef3dfa5e56387f1f188a12bac37070bb94294407936e72b966eda62dc94b8.jpg", + "table_caption": [ + "Table 1. Comparisons of the top-1 accuracy (%) against state-of-the-art methods on the NTU-RGB+D 60, 120, Northwestern-UCLA, and Kinetics-Skeleton datasets. The orange and yellow cells respectively indicate the highest and second-highest value." + ], + "table_footnote": [], + "table_body": "
Graph typeEdgesS-EdgeConvX-Sub (%)X-Set (%)
ConventionalPC✘83.585.4
HD-Graph
APC✘84.3 (↑ 0.8)86.1 (↑ 0.7)
BPC✓84.6 (↑ 1.1)86.3 (↑ 0.9)
CFC✘84.9 (↑ 1.4)86.5 (↑ 1.1)
DFC✓85.1 (↑ 1.6)86.7 (↑ 1.3)
", + "bbox": [ + 500, + 383, + 890, + 479 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2bdbc49ba7dd6e1f236e3b7613cd6ed00e37188a036bb6d249b3f0218fbc7a3d.jpg", + "table_caption": [ + "Table 2. Comparison of the conventional graph and four types of HD-Graph." + ], + "table_footnote": [], + "table_body": "
MethodH-EdgeConvX-Sub (%)X-Set (%)
Baseline✘83.585.4
HD-GCN
w/o A-HA✘85.1 (↑ 1.6)86.7 (↑ 1.3)
w/ SAP✘85.2 (↑ 1.7)86.7 (↑ 1.3)
w/ SAP✓85.4 (↑ 1.9)87.0 (↑ 1.6)
w/ RSAP✘85.5 (↑ 2.0)87.0 (↑ 1.6)
w/ RSAP✓85.7 (↑ 2.2)87.3 (↑ 1.9)
", + "bbox": [ + 501, + 512, + 890, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of various types of attention modules. SAP denotes the spatial average pooling.", + "bbox": [ + 500, + 622, + 890, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hierarchically Decomposed Graph. To proceed with the ablation study for HD-Graph, we set Yan et al. [35]'s graph as the conventional graph. Here, we use the temporal convolution module of [18], as mentioned in Sec. 3.5, to compare the performance of networks fairly with various graphs. The experimental results are shown in 2.", + "bbox": [ + 496, + 657, + 890, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We set the edges of the HD-Graph in different ways to show a gradual performance increase according to the type of graph. There are four main versions of HD-Graph, the first of which is graph A containing only the PC edges. Unlike the conventional graph with one edge set including three fixed subsets, HD-Graph has a flexible number of edge sets, each divided by hierarchy layers with three subsets. Graph B is an extension of A, with the additional operation S-EdgeConv. Graph C contains FC edges for $N_{H}$ hierarchy node sets, and graph D is similar to C but includes", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10450", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb9b86c77d6e50d56a0b1c929499240e687be1cbcdf22485b7ade65a62477c99.jpg", + "image_caption": [ + "Figure 5. Hierarchy-wise attention scores by A-HA for \"Running\" and \"Kicking\" class. Each score indicates the value of the attention map M of the A-HA module." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 464, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "S-EdgeConv. The HD-Graph with only PC edges performs better than the conventional graph by a large margin, even though they share the same edges. This proves that it is meaningful to divide the joint nodes by hierarchy edge sets. In addition, the HD-Graph with FC edges and S-EdgeConv performs better for every datasets.", + "bbox": [ + 75, + 400, + 468, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attention-Guided Hierarchy Aggregation. To prove the effectiveness of the A-HA module, we use a method to change or remove specific parts of our attention module, with the results shown in Tab. 3. Spatial average pooling (SAP) simply averages along the spatial axis without the representative node extraction process, which performs worse than our RSAP. The poorer performance is due to two factors: (1) scaling bias occurs because the number of nodes in each hierarchy node set is different, and (2) attention through SAP does not represent the corresponding hierarchy node set because it brings the average of the feature vectors of all nodes, not a specific node set. Furthermore, it performs better with H-EdgeConv, which recognizes each hierarchy edge set as a graph node. This proves that because the major edge sets are different for each data sample, it is important to find and highlight edge sets with high similarity based on the Euclidean distance through H-EdgeConv.", + "bbox": [ + 75, + 507, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of the attention score $\\mathbf{M}$ of our A-HA module are shown in Fig. 5. These results show that our module scores edge sets 4 and 5 higher for the \"running\" class, which includes knees and feet, elbows and hands. For the \"Kicking\" class, A-HA gives the highest score to edge set 3, which includes shoulders and hips, followed by edge set 4 and 5. It is reasonable for human visual recognition that the dynamically moving edge set 4, 5 are more important than the stationary and barely moving edge set 3 when running", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/48a55ca5f7de1576bdb2e84253913f033ee16b2578a8fb0bc0fa3383b8dfc0c0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
DC-GCN [4]84.0*86.1*2.743.45
MS-G3D [18]84.9*86.8*5.223.22
CTR-GCN [2]84.986.5*1.971.46
InfoGCN [6]85.186.31.681.57
HD-GCN85.787.31.601.68
", + "bbox": [ + 500, + 88, + 890, + 175 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f899662ea4e68f01c34adb652188266ebd819833fd32176bf1cb7596d29885a2.jpg", + "table_caption": [ + "Table 4. Comparison of complexity of the single-stream state-of-the-arts. $\\star$ : results obtained by the released codes." + ], + "table_footnote": [], + "table_body": "
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
CTR-GCN † [2]88.990.67.885.84
InfoGCN † [6]89.490.76.726.28
InfoGCN ‡ [6]89.891.210.089.42
HD-GCN †89.891.26.406.72
HD-GCN ‡90.191.69.6010.08
", + "bbox": [ + 501, + 209, + 890, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison of complexity of the multi-stream state-of-the-arts. †: 4-ensemble, ‡: 6-ensemble", + "bbox": [ + 498, + 299, + 890, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rather than when kicking something.", + "bbox": [ + 500, + 332, + 743, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Six-Way Ensemble. We use the ensemble method to which three graphs with different CoM nodes are applied, excluding motion streams. Tab. 1 shows that the HD-GCN with 4-way ensemble outperforms the state-of-the-art [6] 4-way methods with motion data and shows comparable performance with the state-of-the-art 6-way method. In addition, when the 6-way ensemble with three different graphs is applied to HD-GCN, it outperforms the state-of-the-art methods. This proves that the features extracted with different CoM nodes are learned in different learning aspects.", + "bbox": [ + 496, + 369, + 890, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Comparison of Complexity with Other Models", + "text_level": 1, + "bbox": [ + 498, + 532, + 890, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our model has multiple branch layers for multiple edge sets, it does not cause high complexity because it precedes channel reduction layers. Comparisons of computational complexity with other models are shown in Tab. 4, where the window size is fixed to 64. Our model shows the best performance on NTU-RGB+D 120 joint stream by a large margin even though the computational complexity of our model is the lowest. For multi-stream ensemble, our 4-stream HD-GCN shows almost similar performance to 6-stream InfoGCN [6] while having 3.68G fewer FLOPs and 2.70M fewer parameters as shown in Tab. 5.", + "bbox": [ + 496, + 556, + 890, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 738, + 625, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel hierarchically decomposed graph convolutional network (HD-GCN) for skeleton-based action recognition. We also propose a new framework (HD-Graph) that replaces the existing framework, decomposes all the joint nodes by hierarchy edge sets and considers the connectivity between major distant nodes, which is difficult to identify naturally. We also present an effective attention module (A-HA) composed of representative spatial average pooling (RSAP) layer and hierarchical", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10451", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "edge convolution (H-EdgeConv), which applies hierarchywise attention for the HD-Graph. In addition, our HD-GCN learns graph-wise features with different patterns through a six-way ensemble method. We derive an effective feature extractor by combining these three methods and empirically verify its effectiveness. Our approach outperforms current state-of-the-art methods on four benchmark datasets.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by Alon-Flow Research and the KIST Institutional Program (Project No.2E32283-23-064).", + "bbox": [ + 76, + 218, + 470, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 277, + 174, + 294 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 6", + "[2] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 1, 2, 3, 6, 7, 8", + "[3] Zhan Chen, Sicheng Li, Bing Yang, Qinghan Li, and Hong Liu. Multi-scale spatial temporal graph convolutional network for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1113-1122, 2021. 7", + "[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 7, 8", + "[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1, 2, 7", + "[6] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogen: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 7, 8", + "[7] David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alán Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems, 28, 2015. 3", + "[8] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 3", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + ], + "bbox": [ + 86, + 304, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3", + "[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 7", + "[12] Lipeng Ke, Kuan-Chuan Peng, and Siwei Lyu. Towards to-at spatio-temporal focus for skeleton-based action recognition. arXiv preprint arXiv:2202.02314, 2022. 7", + "[13] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning, pages 2688-2697. PMLR, 2018. 3", + "[14] Matthew Korban and Xin Li. Ddgcn: A dynamic directed graph convolutional network for action recognition. In European Conference on Computer Vision, pages 761-776. Springer, 2020. 7", + "[15] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, and Sangyoun Lee. Leveraging spatio-temporal dependency for skeleton-based action recognition. arXiv preprint arXiv:2212.04761, 2022. 1", + "[16] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 2, 6, 7", + "[17] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In European conference on computer vision, pages 816-833. Springer, 2016.", + "[18] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 143-152, 2020. 1, 2, 6, 7, 8", + "[19] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[20] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and Michael M Bronstein. Geometric deep learning on graphs and manifolds using mixture model cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5115-5124, 2017. 3", + "[21] Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. Learning convolutional neural networks for graphs. In International conference on machine learning, pages 2014-2023. PMLR, 2016. 3", + "[22] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu $\\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016. 2, 6, 7", + "[23] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10452", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 7", + "[24] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1, 2, 3, 4, 6, 7", + "[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. IEEE Transactions on Image Processing, 29:9532-9545, 2020. 1, 2, 3, 6, 7", + "[26] Chenyang Si, Wentao Chen, Wei Wang, Liang Wang, and Tieniu Tan. An attention enhanced graph convolutional LSTM network for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1227-1236, 2019. 1, 7", + "[27] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 7", + "[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6", + "[29] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 1", + "[30] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning and recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2649-2656, 2014. 2, 6, 7", + "[31] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1", + "[32] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 3", + "[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 4, 5", + "[34] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3", + "[35] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1, 2, 3, 4, 7", + "[36] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "networks for efficient skeleton-based human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1112-1121, 2020. 1, 7", + "bbox": [ + 532, + 90, + 890, + 146 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10453", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_model.json b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a3495a0b98d8d57ad649cdd09ab683ea916dc128 --- /dev/null +++ b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.804, + 0.177 + ], + "angle": 0, + "content": "Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.203, + 0.744, + 0.223 + ], + "angle": 0, + "content": "Jungho Lee\\(^{1}\\), Minhyeok Lee\\(^{1}\\), Dogyoon Lee\\(^{1}\\), Sangyoun Lee\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.239, + 0.753, + 0.274 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electrical and Electronic Engineering, Yonsei University \\(^{2}\\)AIonFlow Research" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.278, + 0.728, + 0.293 + ], + "angle": 0, + "content": "{2015142131,hydragon516,nemotio,sylee}@yonsei.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.328, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.474, + 0.709 + ], + "angle": 0, + "content": "Graph convolutional networks (GCNs) are the most commonly used methods for skeleton-based action recognition and have achieved remarkable performance. Generating adjacency matrices with semantically meaningful edges is particularly important for this task, but extracting such edges is challenging problem. To solve this, we propose a hierarchically decomposed graph convolutional network (HD-GCN) architecture with a novel hierarchically decomposed graph (HD-Graph). The proposed HD-GCN effectively decomposes every joint node into several sets to extract major structurally adjacent and distant edges, and uses them to construct an HD-Graph containing those edges in the same semantic spaces of a human skeleton. In addition, we introduce an attention-guided hierarchy aggregation (A-HA) module to highlight the dominant hierarchical edge sets of the HD-Graph. Furthermore, we apply a new six-way ensemble method, which uses only joint and bone stream without any motion stream. The proposed model is evaluated and achieves state-of-the-art performance on four large, popular datasets. Finally, we demonstrate the effectiveness of our model with various comparative experiments. Code is available at https://github.com/Jho-Yonsei/HD-GCN." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Human action recognition (HAR) is a task that categorizes action classes by receiving video data as input. HAR is used in many applications, such as human-computer interaction and virtual reality. Recently, several RGB-based and skeleton-based HAR methods have been proposed with the development of deep learning technology. However, RGB-based methods [31, 29] cannot robustly recognize human actions because they are strongly influenced by environmental noises such as background color, brightness of light," + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.33, + 0.88, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.507, + 0.895, + 0.605 + ], + "angle": 0, + "content": "Figure 1. The framework of HD-GCN. The input skeleton is applied with various edge sets through a hierarchically decomposed graph (HD-Graph). The red lines are the edges included in the corresponding hierarchy edge set. The network highlights the major edge sets through the attention map. The darker the color of red line, the more highlighted the edge set, and dotted lines denote unconnected edges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.689 + ], + "angle": 0, + "content": "and clothing. Therefore, methods using skeleton modality [35, 24, 36, 26, 5, 4, 18, 2, 15] have attracted attention because they are not affected by these noises. These methods recognize action by receiving 2D or 3D coordinates of major human joints as time-series inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Recent approaches [24, 18, 4, 2] have adopted graph convolutional networks (GCNs) to apply human-skeleton graphs to convolutional layers. However, existing GCN-based methods [35, 24, 25, 4, 2] have the following limitations. (1) With the widely used handcrafted graph, the relationships between distant joint nodes are not identified since they use only the relationships of PC edges in the human skeleton. Although the graph with PC edges has a semantic significance, the graph with only PC edges suffers from long-range dependency problem as they are heuristically fixed. However, for humans to recognize actions, relationships between structurally distant joints as well as between adjacent joints are strongly correlated. Although several methods [24, 2] have attempted to solve such limita" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10444" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "tion by training attention-guided learnable graphs, they still use [35]'s handcrafted graph with their learnable graphs. Moreover, as the element values of [35]'s graph are more dominant than those of the learnable graphs, they do not adequately highlight the relationships between distant nodes. (2) Some recent methods [35, 24, 4, 18] risk falling into suboptimality by simply aggregating the edge features and ignoring the contribution of each edge, thus incompletely recognizing which edges are significant for each skeleton sample. For example, in the case of a 'squat down' action, the interactions between the legs and arms should be highlighted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.382 + ], + "angle": 0, + "content": "Motivated by these limitations, we propose a hierarchically decomposed graph convolutional network (HD-GCN) with a hierarchically decomposed graph (HD-Graph) and attention-guided hierarchy aggregation (A-HA) module. In addition, we present a six-way ensemble method to effectively utilize our HD-Graph. The framework of our proposed methods is shown in Fig. 1 for 'squat down' action." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.473, + 0.718 + ], + "angle": 0, + "content": "The HD-GCN incorporates GCNs with our HD-Graph, which identifies the relationships between distant joint nodes in the same semantic spaces (e.g. right and left hands, right and left feet). The same semantic spaces are formed by moving out step by step from the Center of Mass (CoM) node of the graph. For example, if belly is a CoM node, the first semantic space includes the belly node, the next space includes the chest and hip nodes, and the subsequent space includes the left and right shoulder and the left and right hip nodes. The nodes in the same semantic space are defined as hierarchy node set. To detect the relationships between distant joint nodes, network should have large receptive field. The proposed HD-Graph contains both meaningful adjacent and distant joint nodes by connecting all the nodes in neighboring hierarchy node sets and identifies the connectivity between those nodes for large receptive field. We adopt rooted tree-like structure to effectively represent every edges. We apply a spatial edge convolution (S-EdgeConv) layer to consider semantically close edges which cannot be captured by the HD-Graph for each sample. To create the S-EdgeConv layer, we borrow the structure of [33], which is widely used in 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To consider the contribution of each edge set, the process of selecting the dominant hierarchical information should depend on the action data sample to give proper attention to the most dominant edge sets. For example, in order to recognize the \"clapping\" action, a hierarchy edge set that includes both hands must be emphasized. To tackle this issue, we propose an attention-guided hierarchy aggregation (A-HA) module, which consists of two submodules: representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv). A scaling bias problem occurs if we use the spatial average pooling module without any node extraction process because each node has a different" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "number of adjacent nodes. To prevent this, we apply RSAP, which includes a representative node extraction process that triggers features after the pooling layer to represent each node. To effectively manage hierarchical features obtained by RSAP, we apply a hierarchical edge convolution (H-EdgeConv) layer. The H-EdgeConv treats each hierarchical feature as a graph node and identifies which hierarchical features should be highlighted via the Euclidean distance in feature space. With the RSAP and the H-EdgeConv, our model successfully determines which hierarchy edge sets and joints should be emphasized among the input features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.454 + ], + "angle": 0, + "content": "The existing ensemble method uses four-stream data composed of the joint, bone, joint motion, and bone motion streams, which are the original skeletal coordinates, spatial differential between joint coordinates, and temporal differential of joint, and temporal differential of the bone, respectively. Most existing ensemble methods [25, 2] use additional motion data, but models that solely utilize motion data exhibit relatively inferior performance. To address this problem, we present a new method, a six-way ensemble. We apply this ensemble method by setting three HDGraphs with joint and bone stream data. Each graph has different CoM nodes to extract features of different semantic spaces (see Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.515 + ], + "angle": 0, + "content": "We conduct extensive experiments on four benchmark action recognition datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Kinetics-Skeleton [11], and Northwestern-UCLA [30]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.516, + 0.858, + 0.53 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.542, + 0.892, + 0.587 + ], + "angle": 0, + "content": "- We propose a hierarchically decomposed graph (HDGraph) to thoroughly identify the significant distant edges between the same hierarchy node sets." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.595, + 0.892, + 0.656 + ], + "angle": 0, + "content": "- We propose an attention-guided hierarchy aggregation (A-HA) module to highlight the key edge sets with representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv)." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.664, + 0.892, + 0.724 + ], + "angle": 0, + "content": "- We use a new six-way ensemble method for skeleton-based action recognition with HD-Graphs that have different center of mass (CoM), which outperforms regular ensemble without any motion data." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.732, + 0.892, + 0.762 + ], + "angle": 0, + "content": "- Our HD-GCN outperforms the state-of-the-arts on four benchmarks for skeleton-based action recognition." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.542, + 0.892, + 0.762 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.777, + 0.642, + 0.792 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.774, + 0.818 + ], + "angle": 0, + "content": "2.1. Action Recognition with GCNs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In skeleton-based action recognition, human skeletal data are represented by a graph with joint nodes. Most recent approaches use GCN-based methods [35, 24, 5, 18, 2] with [35]'s graph structure, which identifies physical connections in the human skeleton. Those GCN-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10445" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.468, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.875, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.893, + 0.42 + ], + "angle": 0, + "content": "Figure 2. (a) Structure of HD-Graph with physically connected (PC) edges. The human skeleton graph is decomposed into a rooted tree, where PC edges are included in hierarchy sets. (b) Structure of HD-Graph with fully connected (FC) edges. Edges between all nodes in the same semantic space are obtained by connecting all the nodes in adjacent hierarchy edge sets. Blue and red lines stand for PC and FC edges, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.473, + 0.634 + ], + "angle": 0, + "content": "ods perform remarkably better than methods using handcrafted features [7, 8, 13, 20, 21]. They extract the spatial features representing the relationships between physically connected edges among human skeleton, and they outperform other methods by using them to construct the major relationships between joint nodes in the human skeleton. In particular, [25] and [2] propose adaptive attention-based graph structures to learn the sample-wise topological features. However, they might fall into suboptimality because they do not consider the physical prior of the human skeletal structure and allow too much flexibility in network training. To address this issue, we introduce a novel HD-Graph, referencing the known tendencies of human perception" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.436, + 0.687 + ], + "angle": 0, + "content": "2.2. Attention Modules for Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.9 + ], + "angle": 0, + "content": "The attention mechanism is an essential element for constructing a deep neural network. Using recent attention modules [10, 34], networks emphasize important information along a specific dimension. For example, Hu et al. [10] applies channel-wise attention, and Woo et al. [34] applies both channel-wise and spatial-wise attention. These techniques are divided into two categories for GCNs: (1) attention-based graph construction [24, 2] which is a method of forming topologies using a non-local block [32] or customized correlation matrices, and (2) spatial-wise, temporal-wise, channel-wise attention, which are commonly used attentions in [25, 27], and several other networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.635, + 0.453 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.583 + ], + "angle": 0, + "content": "In Sec. 3.2, we detail the HD-Graph convolution to solve the problems of the conventional human-skeleton graph [35], which includes only PC edges. We also explain the A-HA module in Sec. 3.3 to highlight dominant hierarchical features. In Sec. 3.4, we replace the widely used four-stream ensemble method [25, 2, 4] with a six-way ensemble without motion data streams. Finally, we introduce the HD-GCN, which uses these proposed methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.592, + 0.64, + 0.607 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Notations. The spatio-temporal graph for human skeleton is represented by \\(\\mathcal{G}(\\mathcal{V},\\mathcal{E})\\), where \\(\\mathcal{V}\\) and \\(\\mathcal{E}\\) denote the joint and edge groups, respectively. Physically connected edges and fully-connected edges used in Sec. 3.2 are denoted as PC-edges and FC-edges, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Graph Convolutional Networks. 3D time-series skeletal data are represented by \\(\\mathbf{X} \\in \\mathbb{R}^{3 \\times T \\times V}\\), where \\(V\\) and \\(T\\) are the number of joint nodes and the temporal window size, respectively. GCN's operation with input feature map \\(\\mathbf{F}_{in} \\in \\mathbb{R}^{C \\times T \\times V}\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.796, + 0.891, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {o u t} = \\sum_ {s \\in S} \\widehat {\\mathbf {A}} _ {s} \\mathbf {F} _ {i n} \\Theta_ {s}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( S = \\{s_{id}, s_{cf}, s_{cp}\\} \\) denotes graph subsets, and \\( s_{id}, s_{cf}, \\) and \\( s_{cp} \\) indicate identity, centrifugal, and centripetal joint subsets, respectively. \\( \\Theta_s \\) denotes the pointwise convolution operation. The normalized adjacency matrix \\( \\widehat{\\mathbf{A}} \\) is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10446" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.089, + 0.47, + 0.122 + ], + "angle": 0, + "content": "initialized as \\(\\Lambda^{-\\frac{1}{2}}\\mathbf{A}\\Lambda^{-\\frac{1}{2}}\\in \\mathbb{R}^{N_S\\times V\\times V}\\), where \\(\\Lambda\\) is a diagonal matrix for normalization and \\(N_{S} = 3\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.133, + 0.383, + 0.15 + ], + "angle": 0, + "content": "3.2. Hierarchically Decomposed Graph" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.157, + 0.47, + 0.218 + ], + "angle": 0, + "content": "Most recent methods have adopted the handcrafted graph proposed by Yan et al. [35], but the HD-Graph is derived through a newly presented method. Fig. 2 shows the framework of the HD-Graph." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.471, + 0.422 + ], + "angle": 0, + "content": "Decomposition into a Rooted Tree. The first step is to decompose the graph with PC edges and construct a rooted tree. To decompose a given skeleton into the tree, we need to determine a CoM node, which allows nodes in the same hierarchy edge set to exist in the same semantic space. For example, nodes in the same semantic space, such as elbow and knee joints, or hands and feet, must exist in a hierarchy node set. After choosing the CoM node, the graph is converted into a rooted tree, which includes the hierarchical information of the graph, and defines the directed adjacency matrix \\(\\vec{\\mathbf{A}}_{\\mathrm{HD}} \\in \\mathbb{R}^{N_L \\times V \\times V}\\) with \\(N_L\\) hierarchy layers for \\(N_H\\) hierarchy edge sets:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.433, + 0.47, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\overrightarrow {\\mathbf {A}} _ {\\mathrm {H D}} = \\left[ \\mathcal {E} \\left(H _ {1} \\rightarrow H _ {2}\\right), \\dots , \\mathcal {E} \\left(H _ {N _ {H} - 1} \\rightarrow H _ {N _ {H}}\\right)\\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.653 + ], + "angle": 0, + "content": "where \\(H_{k}\\) denotes the \\(k\\)-th hierarchy node set and \\(\\mathcal{E}(H_k\\to\\) \\(H_{k + 1})\\) denotes a set of edges from \\(H_{k}\\) to \\(H_{k + 1}\\). \\(N_{L}\\) and \\(N_{H}\\) are the number of hierarchy layers and hierarchy edge sets, respectively, and \\(N_{L} = N_{H} - 1\\). However, \\(\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{HD}}\\) includes only the directed centrifugal edges. For consistency with existing methods, all the reverse-directed edges from the leaf nodes of the rooted tree in Fig. 2 to the CoM node must be reflected in the adjacency matrices to cover the centripetal edges. In addition, to get the features of the nodes themselves, the identity edges for each hierarchy node set must be considered. Thus, the adjacency matrices \\(\\stackrel {\\leftrightarrow}{\\mathbf{A}}_{\\mathrm{HD}}\\in \\mathbb{R}^{N_L\\times N_S\\times V\\times V}\\) are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.662, + 0.47, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\overleftrightarrow {\\mathbf {A}} _ {\\mathrm {H D}} = \\left[ \\begin{array}{l} \\mathcal {E} _ {1}, \\mathcal {E} _ {2}, \\dots , \\mathcal {E} _ {N _ {L}} \\end{array} \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.724, + 0.47, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {k} = \\mathcal {E} (\\underbrace {H _ {k} \\cup H _ {k + 1}} _ {s _ {i d}}, \\underbrace {H _ {k} \\rightarrow H _ {k + 1}} _ {s _ {c p}}, \\underbrace {H _ {k + 1} \\rightarrow H _ {k}} _ {s _ {c f}}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.469, + 0.849 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_k\\) denotes the concatenation of the three edge subsets of \\(S = \\{s_{id}, s_{cp}, s_{cf}\\}\\) and \\(s_{id}, s_{cp}, s_{cf}\\) indicate the identity, centripetal, and centrifugal edge subsets, respectively. Through this construction policy, we create a skeletal graph with bidirectional and identity edges." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.869, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Fully Connected Inter-Hierarchy Edges. Decomposed graph \\(\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{HD}}\\) has a different number of edge sets from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.304 + ], + "angle": 0, + "content": "conventional graph, but the edges are all the same. To identify the relationships between major distant joint nodes, especially those in the same semantic space, we connect all nodes between neighboring hierarchy node sets. In addition, since [35]'s graph contains the connectivity of only PC edges, not distant relationships, the receptive field is very small with this sparse graph. Applying our fully connected (FC) edges to the rooted tree, the graph becomes denser and makes the receptive field larger than before with more meaningful distant connectivity as shown in Fig. 2 (b). Then, the adjacency matrices are normalized with degree matrices for training stability and we leave all elements of the matrices as learnable parameters for training adaptability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.476 + ], + "angle": 0, + "content": "HD-Graph Convolution. Our HD-Graph convolution includes four parallel branch operations: three graph convolution through HD-Graph and an additional EdgeConv [33] operation. To reduce the computational complexity, a linear transformation is applied to all four operations. For three of these operations, our method performs a subset-wise GCN operation in the same way as [35, 24] for each hierarchy edge set with three edge subsets. However, rather than summing the output values for each subset as in Eq. (1), we concatenate these output values to the channel dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.487, + 0.892, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} = \\left\\| _ {s \\in S} \\left\\{\\overleftrightarrow {\\mathbf {A}} _ {\\mathrm {H D}; s} ^ {(k)} \\Phi \\left(\\mathbf {F} _ {i n}\\right) \\Theta_ {s} ^ {(k)} \\right\\}, \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.598 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{\\mathrm{HD}}^{(k)}\\) denotes the output feature map of the HD-Graph convolution and function \\(\\Phi\\) denotes a linear transformation with parameter \\(\\mathbf{W} \\in \\mathbb{R}^{C' \\times C}\\). Note that \\(\\|\\cdot\\|\\) is a concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although our HD-Graph defines more meaningful node relationships than conventional graph, it may still not be able to extract sample-wise key relationships that reflect the similarities between all nodes in the feature space. To improve this limitation, we adopt EdgeConv [33] as the remaining operation, which is used for extracting graphical features through local neighborhood graphs in the feature space. With spatial EdgeConv (S-EdgeConv), our network extracts sample-wise node connectivity, which the HD-Graph does not capture. For our method, S-EdgeConv initially takes the average pooling as the temporal dimension for computational efficiency. Local graphs with local edges are then formed via k-nearest neighbor (k-NN) based on the Euclidean distance, and the local edges as well as identity edges based on the graphs are aggregated via trainable parameters \\(\\mathbf{W}_{edge} \\in \\mathbb{R}^{C' \\times 2C'}\\). For the deep neural network, physically close edges are reflected to the initial shallow layers, but as they become deeper, the relationship between semantically similar edges in the feature space are identified and learned. Our whole GCN process is shown in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10447" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.102, + 0.235, + 0.127 + ], + "angle": 0, + "content": "HD-Graph Conv. on \\(k\\) th Hierarchy Set" + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.102, + 0.44, + 0.125 + ], + "angle": 0, + "content": "HD - Graph Conv. Block with A-HA" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.138, + 0.274, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.135, + 0.457, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.374, + 0.455, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Figure 3. HD-Graph convolution operation block with an A-HA module. The left side shows an operation for one hierarchy edge set, and the right side shows an operation block that concatenates the results for \\( N_{L} \\) edge sets and applies A-HA. The lower part of the figure is EdgeConv, where the EdgeConv subscript indicates the feature space to extract graphical features. The \\( \\times \\) and \\( \\cdot \\) operations denote matrix and element-wise multiplication." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.565, + 0.291, + 0.58 + ], + "angle": 0, + "content": "Fig. 3 and computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.591, + 0.469, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {H D}} \\leftarrow \\sum_ {k = 1} ^ {N _ {L}} \\left[ \\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} \\left| \\right| z _ {\\mathcal {V}} ^ {(k)} \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} \\Phi (\\mathbf {F} _ {i n})\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.377, + 0.654 + ], + "angle": 0, + "content": "where \\( z_{\\mathcal{V}} \\) denotes the S-EdgeConv operation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.469, + 0.82 + ], + "angle": 0, + "content": "All four branch outputs are concatenated to the channel dimension, with all four computed in the same way for \\( N_{L} \\) hierarchy edge sets. Due to the inherent characteristics of skeletal data, the number of joint nodes included in each dataset is different, and, consequently, the number of hierarchy sets is different. Therefore, we adopt an addition policy for \\( N_{L} \\) hierarchy-wise outputs and a concatenation policy for \\( N_{S} \\) subset-wise outputs. In this way, the dimensionality is maintained, and the common hierarchy-wise aggregation policy is followed for every skeletal dataset by adding all the outputs for different numbers of hierarchical sets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.431, + 0.848 + ], + "angle": 0, + "content": "3.3. Attention-Guided Hierarchy Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The HD-Graph convolution uses an aggregation policy of adding all the hierarchy-wise outputs. However, because each data sample has relationship between specific major" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "edges, we propose an attention-guided hierarchy aggregation (A-HA) module, which applies a weighted-sum policy to the hierarchy dimension with proper attention to the hierarchy-wise outputs. The framework of the HD-Graph convolution with the A-HA module is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Representative Spatial Average Pooling. Our A-HA module is applied to feature map \\(\\mathbf{F}_{\\mathrm{HD}} \\in \\mathbb{R}^{C \\times N_L \\times T \\times V}\\) after the HD-Graph convolution. The first step is to extract the temporal frame with the highest score on \\(\\mathbf{F}_{\\mathrm{HD}}\\). RSAP, \\(\\Psi\\), is then applied, which is preceded by the extraction of representative nodes in each hierarchy layer. If spatial average pooling is applied without this extraction process, scaling bias occurs because the number of edges connected to each node is different. Therefore, representative node extraction is essential to obtain an appropriate score for attention without any bias. After the extraction, spatial average pooling is applied to hierarchy-wise outputs. Our pooling function \\(\\Psi\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.387, + 0.892, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\Psi \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)}\\right) = \\frac {1}{N _ {k} + N _ {k + 1}} \\sum_ {v \\in H _ {k} \\cup H _ {k + 1}} \\max _ {t} \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} (v)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.43, + 0.866, + 0.444 + ], + "angle": 0, + "content": "where \\(N_{k}\\) denotes the number of vertices in the \\(H_{k}\\) set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Hierarchical Edge Convolution. After the RSAP layer, \\( N_{L} \\) hierarchy-wise features in attention feature map \\( \\mathbf{M} \\) have not yet shared their information with each other. We treat all \\( N_{L} \\) features as nodes on a graph to learn and reflect similarities in the hierarchical feature space. To apply this process, representative features of these nodes are fed into EdgeConv [33], and the similarities of those nodes are learned based on the Euclidean distance. We also include the self-loop shown in the bottom section of Fig. 3 so that the node's own features can be reflected. Our attention map \\( \\mathbf{M} \\) operates as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.638, + 0.892, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\sigma \\left(z _ {L} \\left(\\left. \\right| \\left. \\right| _ {k \\in L} \\left\\{\\Psi \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)}\\right) \\right\\}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.89, + 0.719 + ], + "angle": 0, + "content": "where \\(z_{L}\\) and \\(\\sigma\\) denote H-EdgeConv and the sigmoid function, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The attention map \\(\\mathbf{M}\\) obtained is multiplied by the HD-Graph convolution output feature map \\(\\mathbf{F}_{\\mathrm{HD}}\\), and the output feature map \\(\\mathbf{F}_{out}\\) is obtained through a weighted sum to the hierarchy axis as shown in Fig. 3. Similar to S-EdgeConv in Sec. 3.2, The H-EdgeConv method incorporates the concept of hierarchical edge sets in a physically proximal manner in earlier layers, while in the deeper layers, it emphasizes the presence of semantically similar edge sets. This approach highlights different hierarchical edge sets for each sample and enables the model to learn meaningful representations that capture both physical and semantic properties of the input data." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10448" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.468, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.472, + 0.3 + ], + "angle": 0, + "content": "Figure 4. The architecture of HD-GCN. HD-GCN receives the skeleton sequence as input and obtains the class label through nine GCN blocks, and an FC layer, and the softmax function." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.309, + 0.259, + 0.325 + ], + "angle": 0, + "content": "3.4. Six-Way Ensemble" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.637 + ], + "angle": 0, + "content": "Shi et al. [24, 25] have applied a four-stream ensemble method using streams for joints, bones, joint motion, and bone motion. However, as the performances of motion streams are relatively poorer than the performances of joint and bone streams, we adopt an ensemble method with the joint and bone streams without any motion streams. We use three different HD-Graphs, and each graph is used for training with joint and bone streams. The three HD-Graphs have different CoM nodes, which are chest, belly, hip nodes, respectively. In other words, we train joint and bone streams with HD-Graph with the CoM node of chest, and we train the same when the CoM node is belly or hip node. As models with the three different graphs should be trained in different aspects, each of the graphs is composed of different edge sets. For example, if the CoM node is belly, both thigh edges and both upper arm edges are included in the same edge set, whereas when the CoM node is chest, both thigh edges and both forearm edges are included in the same edge set. The details of our six-way ensemble are specified on our Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.283, + 0.665 + ], + "angle": 0, + "content": "3.5. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 4, we adopt [24] as our baseline network architecture with a total of nine stacked GCN blocks. The numbers of output channels for the blocks are 64, 64, 64, 128, 128, 128, 256, 256, and 256. Each block contains a residual connection [9] and is divided into a spatial module, in which the GCN operation proceeds, and a temporal module, which includes the temporal convolutions. Our method uses the temporal module of [2], whose baseline module is [18, 28]. This module consists of four branch operations. Two are dilated temporal convolutions with kernel size five and dilation one and two, respectively. The remaining branch operations are pointwise convolution and max pooling with kernel size three. Our spatial module consists of an HD-Graph convolution operation and an A-HA module, as introduced in Sec. 3.2 and Sec. 3.3. After passing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "through all GCN layers with attention to the hierarchy-wise features, the network compresses the feature map through the global average pooling layer and classifies the action sample through the softmax function." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.167, + 0.634, + 0.184 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.192, + 0.81, + 0.209 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.351 + ], + "angle": 0, + "content": "NTU-RGB+D 60. NTU-RGB+D 60 [22] is a large dataset used in skeletal action recognition. It contains 56,880 skeleton action samples, performed by 40 different participants and classified into 60 classes. The authors of this dataset recommend two benchmarks. (1) Cross-Subject (X-Sub): 20 of the 40 subjects' actions are used for training, and the remaining 20 are for validation. (2) Cross-View (X-View): Two of the three camera-views are used for training, and the other one is used for validation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.522 + ], + "angle": 0, + "content": "NTU-RGB+D 120. NTU-RGB+D 120 [16] is a dataset in which 57,367 new action samples are added to the NTU-RGB+D 60 dataset. It contains a total of 114,480 skeleton action samples over 120 classes, performed by 106 different subjects. The authors of this dataset recommend two benchmarks: (1) Cross-Subject (X-Sub): 53 of the 106 subjects' actions are used for training, and the remaining 53 are used for validation. (2) Cross-Setup (X-Set): Of the 32 setups, data with even setup IDs are used for training, and the remaining data with odd IDs are used for validation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.663 + ], + "angle": 0, + "content": "Kinetics-Skeleton. The Kinetics-Skeleton dataset is derived from the Kinetics 400 video dataset [11], utilizing the OpenPose pose estimation [1] to extract 240,436 training and 19,796 testing skeleton sequences across 400 classes. The dataset restricts the number of skeletons per time step to two and eliminates skeletons with lower confidence scores, ensuring high-quality sequences for human action recognition and pose estimation research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.892, + 0.782 + ], + "angle": 0, + "content": "Northwestern-UCLA. The Northwestern-UCLA skeleton dataset [30] contains 1494 video clips over 10 classes. Each action is captured through three Kinect cameras with different camera views and is performed by 10 subjects. We adopt the same protocol as NW-UCLA: Two of the three camera-views are used for training, and the other one is used for validation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Experimental Settings. In our experiments, we adopt [24] as the backbone. The SGD optimizer is employed with a Nesterov momentum of 0.9 and a weight decay of 0.0004. The number of learning epochs is set to 90, with a warm-up strategy [9] applied to the first five epochs for more stable learning. We set the learning rate to decay with cosine annealing [19], with a maximum" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10449" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.091, + 0.891, + 0.334 + ], + "angle": 0, + "content": "
MethodsPublicationMotion StreamNTU-RGB+D 60NTU-RGB+D 120Kinetics-SkeletonNorthwestern
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)Top-1 (%)Top-5 (%)UCLA (%)
ST-GCN [35]AAAI 2018✘81.588.370.773.230.752.8-
2s-AGCN [24]CVPR 2019✘88.595.182.584.236.158.7-
SGN [36]CVPR 2020✓89.094.579.281.5--92.5
AGC-LSTM [26]CVPR 2019✘89.295.0----93.3
DGNN [23]CVPR 2019✓89.996.1--36.959.6-
Shift-GCN [5]CVPR 2020✓90.796.585.987.6--94.6
DC-GCN+ADG [4]ECCV 2020✓90.896.686.588.1--95.3
DDGCN [14]ECCV 2020✓91.197.1--38.160.8-
MS-G3D [18]CVPR 2020✘91.596.286.988.438.060.9-
MST-GCN [3]AAAI 2021✓91.596.687.588.838.160.8-
CTR-GCN [2]ICCV 2021✓92.496.888.990.6--96.5
EfficientGCN-B4 [27]TPAMI 2022✓91.795.788.389.1---
STF [12]AAAI 2022✘92.596.988.989.939.9--
InfoGCN (4-ensemble) [6]CVPR 2022✓92.796.989.490.7--96.6
InfoGCN (6-ensemble) [6]CVPR 2022✓93.097.189.891.2--97.0
HD-GCN (2-ensemble)✘92.496.689.190.638.961.796.6
HD-GCN (4-ensemble)✘93.097.089.891.240.363.096.9
HD-GCN (6-ensemble)✘93.497.290.191.640.963.597.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.337, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Table 1. Comparisons of the top-1 accuracy (%) against state-of-the-art methods on the NTU-RGB+D 60, 120, Northwestern-UCLA, and Kinetics-Skeleton datasets. The orange and yellow cells respectively indicate the highest and second-highest value." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.552 + ], + "angle": 0, + "content": "learning rate of 0.1 and a minimum learning rate of 0.0001. For the NTU-RGB+D datasets, we set the batch size to 64 and use the data preprocessing method from [36]. For Kinetics-Skeleton, the batch size is set to 128. In addition, to overcome the absence of belly and hip nodes in the Kinetics-Skeleton, we define the center of both hip joints as CoM hip node, and the center of chest and the hip node as CoM belly node, resulting in a total of 20 nodes. For the Northwestern-UCLA dataset, we set the batch size to 16 and use the data preprocessing method from [5]. All our experiments are conducted on a single RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.571, + 0.452, + 0.587 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Arts Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.597, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Most recent state-of-the-art networks [25, 5, 4, 2] adopt a four-way ensemble method, but we adopt the six-way ensemble method described in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.471, + 0.797 + ], + "angle": 0, + "content": "We compare ours with state-of-the-art networks on three datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Northwestern-UCLA [30], and Kinetics-Skeleton [11]. Comparisons for each dataset are shown in Tab. 1. The recognition performance of our HD-GCN has exceeded the state-of-the-arts on every dataset without any motion streams, as shown in Tab. 1. With our proposed ensemble method, HD-GCN outperforms the state-of-the-art and shows comparable performance to the 6-way ensemble state-of-the-art using only 4-way ensemble method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.231, + 0.831 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this section, we demonstrate the effectiveness of the proposed HD-GCN. Performance is specified as the cross-subject and cross-setup classification accuracy on the NTU-RGB+D 120 [16] joint stream data." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.384, + 0.892, + 0.48 + ], + "angle": 0, + "content": "
Graph typeEdgesS-EdgeConvX-Sub (%)X-Set (%)
ConventionalPC✘83.585.4
HD-Graph
APC✘84.3 (↑ 0.8)86.1 (↑ 0.7)
BPC✓84.6 (↑ 1.1)86.3 (↑ 0.9)
CFC✘84.9 (↑ 1.4)86.5 (↑ 1.1)
DFC✓85.1 (↑ 1.6)86.7 (↑ 1.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.481, + 0.892, + 0.509 + ], + "angle": 0, + "content": "Table 2. Comparison of the conventional graph and four types of HD-Graph." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.513, + 0.892, + 0.619 + ], + "angle": 0, + "content": "
MethodH-EdgeConvX-Sub (%)X-Set (%)
Baseline✘83.585.4
HD-GCN
w/o A-HA✘85.1 (↑ 1.6)86.7 (↑ 1.3)
w/ SAP✘85.2 (↑ 1.7)86.7 (↑ 1.3)
w/ SAP✓85.4 (↑ 1.9)87.0 (↑ 1.6)
w/ RSAP✘85.5 (↑ 2.0)87.0 (↑ 1.6)
w/ RSAP✓85.7 (↑ 2.2)87.3 (↑ 1.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.623, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Table 3. Comparison of various types of attention modules. SAP denotes the spatial average pooling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Hierarchically Decomposed Graph. To proceed with the ablation study for HD-Graph, we set Yan et al. [35]'s graph as the conventional graph. Here, we use the temporal convolution module of [18], as mentioned in Sec. 3.5, to compare the performance of networks fairly with various graphs. The experimental results are shown in 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We set the edges of the HD-Graph in different ways to show a gradual performance increase according to the type of graph. There are four main versions of HD-Graph, the first of which is graph A containing only the PC edges. Unlike the conventional graph with one edge set including three fixed subsets, HD-Graph has a flexible number of edge sets, each divided by hierarchy layers with three subsets. Graph B is an extension of A, with the additional operation S-EdgeConv. Graph C contains FC edges for \\( N_{H} \\) hierarchy node sets, and graph D is similar to C but includes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10450" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.465, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Figure 5. Hierarchy-wise attention scores by A-HA for \"Running\" and \"Kicking\" class. Each score indicates the value of the attention map M of the A-HA module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.469, + 0.492 + ], + "angle": 0, + "content": "S-EdgeConv. The HD-Graph with only PC edges performs better than the conventional graph by a large margin, even though they share the same edges. This proves that it is meaningful to divide the joint nodes by hierarchy edge sets. In addition, the HD-Graph with FC edges and S-EdgeConv performs better for every datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Attention-Guided Hierarchy Aggregation. To prove the effectiveness of the A-HA module, we use a method to change or remove specific parts of our attention module, with the results shown in Tab. 3. Spatial average pooling (SAP) simply averages along the spatial axis without the representative node extraction process, which performs worse than our RSAP. The poorer performance is due to two factors: (1) scaling bias occurs because the number of nodes in each hierarchy node set is different, and (2) attention through SAP does not represent the corresponding hierarchy node set because it brings the average of the feature vectors of all nodes, not a specific node set. Furthermore, it performs better with H-EdgeConv, which recognizes each hierarchy edge set as a graph node. This proves that because the major edge sets are different for each data sample, it is important to find and highlight edge sets with high similarity based on the Euclidean distance through H-EdgeConv." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The results of the attention score \\( \\mathbf{M} \\) of our A-HA module are shown in Fig. 5. These results show that our module scores edge sets 4 and 5 higher for the \"running\" class, which includes knees and feet, elbows and hands. For the \"Kicking\" class, A-HA gives the highest score to edge set 3, which includes shoulders and hips, followed by edge set 4 and 5. It is reasonable for human visual recognition that the dynamically moving edge set 4, 5 are more important than the stationary and barely moving edge set 3 when running" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.892, + 0.176 + ], + "angle": 0, + "content": "
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
DC-GCN [4]84.0*86.1*2.743.45
MS-G3D [18]84.9*86.8*5.223.22
CTR-GCN [2]84.986.5*1.971.46
InfoGCN [6]85.186.31.681.57
HD-GCN85.787.31.601.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.179, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 4. Comparison of complexity of the single-stream state-of-the-arts. \\(\\star\\): results obtained by the released codes." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.21, + 0.892, + 0.296 + ], + "angle": 0, + "content": "
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
CTR-GCN † [2]88.990.67.885.84
InfoGCN † [6]89.490.76.726.28
InfoGCN ‡ [6]89.891.210.089.42
HD-GCN †89.891.26.406.72
HD-GCN ‡90.191.69.6010.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Table 5. Comparison of complexity of the multi-stream state-of-the-arts. †: 4-ensemble, ‡: 6-ensemble" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.333, + 0.744, + 0.348 + ], + "angle": 0, + "content": "rather than when kicking something." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Six-Way Ensemble. We use the ensemble method to which three graphs with different CoM nodes are applied, excluding motion streams. Tab. 1 shows that the HD-GCN with 4-way ensemble outperforms the state-of-the-art [6] 4-way methods with motion data and shows comparable performance with the state-of-the-art 6-way method. In addition, when the 6-way ensemble with three different graphs is applied to HD-GCN, it outperforms the state-of-the-art methods. This proves that the features extracted with different CoM nodes are learned in different learning aspects." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.55 + ], + "angle": 0, + "content": "4.4. Comparison of Complexity with Other Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Although our model has multiple branch layers for multiple edge sets, it does not cause high complexity because it precedes channel reduction layers. Comparisons of computational complexity with other models are shown in Tab. 4, where the window size is fixed to 64. Our model shows the best performance on NTU-RGB+D 120 joint stream by a large margin even though the computational complexity of our model is the lowest. For multi-stream ensemble, our 4-stream HD-GCN shows almost similar performance to 6-stream InfoGCN [6] while having 3.68G fewer FLOPs and 2.70M fewer parameters as shown in Tab. 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.627, + 0.755 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a novel hierarchically decomposed graph convolutional network (HD-GCN) for skeleton-based action recognition. We also propose a new framework (HD-Graph) that replaces the existing framework, decomposes all the joint nodes by hierarchy edge sets and considers the connectivity between major distant nodes, which is difficult to identify naturally. We also present an effective attention module (A-HA) composed of representative spatial average pooling (RSAP) layer and hierarchical" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10451" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "edge convolution (H-EdgeConv), which applies hierarchywise attention for the HD-Graph. In addition, our HD-GCN learns graph-wise features with different patterns through a six-way ensemble method. We derive an effective feature extractor by combining these three methods and empirically verify its effectiveness. Our approach outperforms current state-of-the-art methods on four benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.219, + 0.472, + 0.264 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by Alon-Flow Research and the KIST Institutional Program (Project No.2E32283-23-064)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.279, + 0.175, + 0.295 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.362, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[2] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.434, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[3] Zhan Chen, Sicheng Li, Bing Yang, Qinghan Li, and Hong Liu. Multi-scale spatial temporal graph convolutional network for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1113-1122, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.575, + 0.472, + 0.645 + ], + "angle": 0, + "content": "[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.647, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[6] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogen: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.731, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[7] David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alán Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[8] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.305, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[12] Lipeng Ke, Kuan-Chuan Peng, and Siwei Lyu. Towards to-at spatio-temporal focus for skeleton-based action recognition. arXiv preprint arXiv:2202.02314, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[13] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning, pages 2688-2697. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[14] Matthew Korban and Xin Li. Ddgcn: A dynamic directed graph convolutional network for action recognition. In European Conference on Computer Vision, pages 761-776. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[15] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, and Sangyoun Lee. Leveraging spatio-temporal dependency for skeleton-based action recognition. arXiv preprint arXiv:2212.04761, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[16] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[17] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In European conference on computer vision, pages 816-833. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[18] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 143-152, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[20] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and Michael M Bronstein. Geometric deep learning on graphs and manifolds using mixture model cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5115-5124, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[21] Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. Learning convolutional neural networks for graphs. In International conference on machine learning, pages 2014-2023. PMLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[22] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu \\(\\mathrm{rgb + }\\) d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10452" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[24] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. IEEE Transactions on Image Processing, 29:9532-9545, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[26] Chenyang Si, Wentao Chen, Wei Wang, Liang Wang, and Tieniu Tan. An attention enhanced graph convolutional LSTM network for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1227-1236, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[27] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[29] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[30] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning and recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2649-2656, 2014. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[31] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[32] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[34] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[35] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "networks for efficient skeleton-based human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1112-1121, 2020. 1, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10453" + } + ] +] \ No newline at end of file diff --git a/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_origin.pdf b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5113412604cc610e4de36ddab8316de2ce4d202b --- /dev/null +++ b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/4da439a0-ed06-4c50-a7bd-03aa64888b02_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae6ca60f618c90863281bacdfe964f6727105393ade055d01ca5bf91a6dbfa20 +size 576648 diff --git a/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/full.md b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b209f26b62fe9de44924fbb00369aee065e0f1d --- /dev/null +++ b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/full.md @@ -0,0 +1,288 @@ +# Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition + +Jungho Lee $^{1}$ , Minhyeok Lee $^{1}$ , Dogyoon Lee $^{1}$ , Sangyoun Lee $^{1,2}$ + +$^{1}$ School of Electrical and Electronic Engineering, Yonsei University $^{2}$ AIonFlow Research + +{2015142131,hydragon516,nemotio,sylee}@yonsei.ac.kr + +# Abstract + +Graph convolutional networks (GCNs) are the most commonly used methods for skeleton-based action recognition and have achieved remarkable performance. Generating adjacency matrices with semantically meaningful edges is particularly important for this task, but extracting such edges is challenging problem. To solve this, we propose a hierarchically decomposed graph convolutional network (HD-GCN) architecture with a novel hierarchically decomposed graph (HD-Graph). The proposed HD-GCN effectively decomposes every joint node into several sets to extract major structurally adjacent and distant edges, and uses them to construct an HD-Graph containing those edges in the same semantic spaces of a human skeleton. In addition, we introduce an attention-guided hierarchy aggregation (A-HA) module to highlight the dominant hierarchical edge sets of the HD-Graph. Furthermore, we apply a new six-way ensemble method, which uses only joint and bone stream without any motion stream. The proposed model is evaluated and achieves state-of-the-art performance on four large, popular datasets. Finally, we demonstrate the effectiveness of our model with various comparative experiments. Code is available at https://github.com/Jho-Yonsei/HD-GCN. + +# 1. Introduction + +Human action recognition (HAR) is a task that categorizes action classes by receiving video data as input. HAR is used in many applications, such as human-computer interaction and virtual reality. Recently, several RGB-based and skeleton-based HAR methods have been proposed with the development of deep learning technology. However, RGB-based methods [31, 29] cannot robustly recognize human actions because they are strongly influenced by environmental noises such as background color, brightness of light, + +![](images/561382c75aab24245f23319b495f169ee1faddbc9fa580c31210991f4432bc30.jpg) +Figure 1. The framework of HD-GCN. The input skeleton is applied with various edge sets through a hierarchically decomposed graph (HD-Graph). The red lines are the edges included in the corresponding hierarchy edge set. The network highlights the major edge sets through the attention map. The darker the color of red line, the more highlighted the edge set, and dotted lines denote unconnected edges. + +and clothing. Therefore, methods using skeleton modality [35, 24, 36, 26, 5, 4, 18, 2, 15] have attracted attention because they are not affected by these noises. These methods recognize action by receiving 2D or 3D coordinates of major human joints as time-series inputs. + +Recent approaches [24, 18, 4, 2] have adopted graph convolutional networks (GCNs) to apply human-skeleton graphs to convolutional layers. However, existing GCN-based methods [35, 24, 25, 4, 2] have the following limitations. (1) With the widely used handcrafted graph, the relationships between distant joint nodes are not identified since they use only the relationships of PC edges in the human skeleton. Although the graph with PC edges has a semantic significance, the graph with only PC edges suffers from long-range dependency problem as they are heuristically fixed. However, for humans to recognize actions, relationships between structurally distant joints as well as between adjacent joints are strongly correlated. Although several methods [24, 2] have attempted to solve such limita + +tion by training attention-guided learnable graphs, they still use [35]'s handcrafted graph with their learnable graphs. Moreover, as the element values of [35]'s graph are more dominant than those of the learnable graphs, they do not adequately highlight the relationships between distant nodes. (2) Some recent methods [35, 24, 4, 18] risk falling into suboptimality by simply aggregating the edge features and ignoring the contribution of each edge, thus incompletely recognizing which edges are significant for each skeleton sample. For example, in the case of a 'squat down' action, the interactions between the legs and arms should be highlighted. + +Motivated by these limitations, we propose a hierarchically decomposed graph convolutional network (HD-GCN) with a hierarchically decomposed graph (HD-Graph) and attention-guided hierarchy aggregation (A-HA) module. In addition, we present a six-way ensemble method to effectively utilize our HD-Graph. The framework of our proposed methods is shown in Fig. 1 for 'squat down' action. + +The HD-GCN incorporates GCNs with our HD-Graph, which identifies the relationships between distant joint nodes in the same semantic spaces (e.g. right and left hands, right and left feet). The same semantic spaces are formed by moving out step by step from the Center of Mass (CoM) node of the graph. For example, if belly is a CoM node, the first semantic space includes the belly node, the next space includes the chest and hip nodes, and the subsequent space includes the left and right shoulder and the left and right hip nodes. The nodes in the same semantic space are defined as hierarchy node set. To detect the relationships between distant joint nodes, network should have large receptive field. The proposed HD-Graph contains both meaningful adjacent and distant joint nodes by connecting all the nodes in neighboring hierarchy node sets and identifies the connectivity between those nodes for large receptive field. We adopt rooted tree-like structure to effectively represent every edges. We apply a spatial edge convolution (S-EdgeConv) layer to consider semantically close edges which cannot be captured by the HD-Graph for each sample. To create the S-EdgeConv layer, we borrow the structure of [33], which is widely used in 3D point clouds. + +To consider the contribution of each edge set, the process of selecting the dominant hierarchical information should depend on the action data sample to give proper attention to the most dominant edge sets. For example, in order to recognize the "clapping" action, a hierarchy edge set that includes both hands must be emphasized. To tackle this issue, we propose an attention-guided hierarchy aggregation (A-HA) module, which consists of two submodules: representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv). A scaling bias problem occurs if we use the spatial average pooling module without any node extraction process because each node has a different + +number of adjacent nodes. To prevent this, we apply RSAP, which includes a representative node extraction process that triggers features after the pooling layer to represent each node. To effectively manage hierarchical features obtained by RSAP, we apply a hierarchical edge convolution (H-EdgeConv) layer. The H-EdgeConv treats each hierarchical feature as a graph node and identifies which hierarchical features should be highlighted via the Euclidean distance in feature space. With the RSAP and the H-EdgeConv, our model successfully determines which hierarchy edge sets and joints should be emphasized among the input features. + +The existing ensemble method uses four-stream data composed of the joint, bone, joint motion, and bone motion streams, which are the original skeletal coordinates, spatial differential between joint coordinates, and temporal differential of joint, and temporal differential of the bone, respectively. Most existing ensemble methods [25, 2] use additional motion data, but models that solely utilize motion data exhibit relatively inferior performance. To address this problem, we present a new method, a six-way ensemble. We apply this ensemble method by setting three HDGraphs with joint and bone stream data. Each graph has different CoM nodes to extract features of different semantic spaces (see Appendix). + +We conduct extensive experiments on four benchmark action recognition datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Kinetics-Skeleton [11], and Northwestern-UCLA [30]. + +Our main contributions are summarized as follows: + +- We propose a hierarchically decomposed graph (HDGraph) to thoroughly identify the significant distant edges between the same hierarchy node sets. +- We propose an attention-guided hierarchy aggregation (A-HA) module to highlight the key edge sets with representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv). +- We use a new six-way ensemble method for skeleton-based action recognition with HD-Graphs that have different center of mass (CoM), which outperforms regular ensemble without any motion data. +- Our HD-GCN outperforms the state-of-the-arts on four benchmarks for skeleton-based action recognition. + +# 2. Related Work + +# 2.1. Action Recognition with GCNs + +In skeleton-based action recognition, human skeletal data are represented by a graph with joint nodes. Most recent approaches use GCN-based methods [35, 24, 5, 18, 2] with [35]'s graph structure, which identifies physical connections in the human skeleton. Those GCN-based meth + +![](images/6431cfc3343be6978763e7228e41897481c24b8d0b638e94f751810a862b6d27.jpg) +Figure 2. (a) Structure of HD-Graph with physically connected (PC) edges. The human skeleton graph is decomposed into a rooted tree, where PC edges are included in hierarchy sets. (b) Structure of HD-Graph with fully connected (FC) edges. Edges between all nodes in the same semantic space are obtained by connecting all the nodes in adjacent hierarchy edge sets. Blue and red lines stand for PC and FC edges, respectively. + +![](images/f59d474ca5410eb86c120ce0161f0c1cd7bf7ed1832e26fb3834ac67d5a50a81.jpg) + +ods perform remarkably better than methods using handcrafted features [7, 8, 13, 20, 21]. They extract the spatial features representing the relationships between physically connected edges among human skeleton, and they outperform other methods by using them to construct the major relationships between joint nodes in the human skeleton. In particular, [25] and [2] propose adaptive attention-based graph structures to learn the sample-wise topological features. However, they might fall into suboptimality because they do not consider the physical prior of the human skeletal structure and allow too much flexibility in network training. To address this issue, we introduce a novel HD-Graph, referencing the known tendencies of human perception + +# 2.2. Attention Modules for Action Recognition + +The attention mechanism is an essential element for constructing a deep neural network. Using recent attention modules [10, 34], networks emphasize important information along a specific dimension. For example, Hu et al. [10] applies channel-wise attention, and Woo et al. [34] applies both channel-wise and spatial-wise attention. These techniques are divided into two categories for GCNs: (1) attention-based graph construction [24, 2] which is a method of forming topologies using a non-local block [32] or customized correlation matrices, and (2) spatial-wise, temporal-wise, channel-wise attention, which are commonly used attentions in [25, 27], and several other networks. + +# 3. Methodology + +In Sec. 3.2, we detail the HD-Graph convolution to solve the problems of the conventional human-skeleton graph [35], which includes only PC edges. We also explain the A-HA module in Sec. 3.3 to highlight dominant hierarchical features. In Sec. 3.4, we replace the widely used four-stream ensemble method [25, 2, 4] with a six-way ensemble without motion data streams. Finally, we introduce the HD-GCN, which uses these proposed methods. + +# 3.1. Preliminaries + +Notations. The spatio-temporal graph for human skeleton is represented by $\mathcal{G}(\mathcal{V},\mathcal{E})$ , where $\mathcal{V}$ and $\mathcal{E}$ denote the joint and edge groups, respectively. Physically connected edges and fully-connected edges used in Sec. 3.2 are denoted as PC-edges and FC-edges, respectively. + +Graph Convolutional Networks. 3D time-series skeletal data are represented by $\mathbf{X} \in \mathbb{R}^{3 \times T \times V}$ , where $V$ and $T$ are the number of joint nodes and the temporal window size, respectively. GCN's operation with input feature map $\mathbf{F}_{in} \in \mathbb{R}^{C \times T \times V}$ is as follows: + +$$ +\mathbf {F} _ {o u t} = \sum_ {s \in S} \widehat {\mathbf {A}} _ {s} \mathbf {F} _ {i n} \Theta_ {s}, \tag {1} +$$ + +where $S = \{s_{id}, s_{cf}, s_{cp}\}$ denotes graph subsets, and $s_{id}, s_{cf},$ and $s_{cp}$ indicate identity, centrifugal, and centripetal joint subsets, respectively. $\Theta_s$ denotes the pointwise convolution operation. The normalized adjacency matrix $\widehat{\mathbf{A}}$ is + +initialized as $\Lambda^{-\frac{1}{2}}\mathbf{A}\Lambda^{-\frac{1}{2}}\in \mathbb{R}^{N_S\times V\times V}$ , where $\Lambda$ is a diagonal matrix for normalization and $N_{S} = 3$ . + +# 3.2. Hierarchically Decomposed Graph + +Most recent methods have adopted the handcrafted graph proposed by Yan et al. [35], but the HD-Graph is derived through a newly presented method. Fig. 2 shows the framework of the HD-Graph. + +Decomposition into a Rooted Tree. The first step is to decompose the graph with PC edges and construct a rooted tree. To decompose a given skeleton into the tree, we need to determine a CoM node, which allows nodes in the same hierarchy edge set to exist in the same semantic space. For example, nodes in the same semantic space, such as elbow and knee joints, or hands and feet, must exist in a hierarchy node set. After choosing the CoM node, the graph is converted into a rooted tree, which includes the hierarchical information of the graph, and defines the directed adjacency matrix $\vec{\mathbf{A}}_{\mathrm{HD}} \in \mathbb{R}^{N_L \times V \times V}$ with $N_L$ hierarchy layers for $N_H$ hierarchy edge sets: + +$$ +\overrightarrow {\mathbf {A}} _ {\mathrm {H D}} = \left[ \mathcal {E} \left(H _ {1} \rightarrow H _ {2}\right), \dots , \mathcal {E} \left(H _ {N _ {H} - 1} \rightarrow H _ {N _ {H}}\right)\right], \tag {2} +$$ + +where $H_{k}$ denotes the $k$ -th hierarchy node set and $\mathcal{E}(H_k\to$ $H_{k + 1})$ denotes a set of edges from $H_{k}$ to $H_{k + 1}$ . $N_{L}$ and $N_{H}$ are the number of hierarchy layers and hierarchy edge sets, respectively, and $N_{L} = N_{H} - 1$ . However, $\overleftrightarrow{\mathbf{A}}_{\mathrm{HD}}$ includes only the directed centrifugal edges. For consistency with existing methods, all the reverse-directed edges from the leaf nodes of the rooted tree in Fig. 2 to the CoM node must be reflected in the adjacency matrices to cover the centripetal edges. In addition, to get the features of the nodes themselves, the identity edges for each hierarchy node set must be considered. Thus, the adjacency matrices $\stackrel {\leftrightarrow}{\mathbf{A}}_{\mathrm{HD}}\in \mathbb{R}^{N_L\times N_S\times V\times V}$ are defined as follows: + +$$ +\overleftrightarrow {\mathbf {A}} _ {\mathrm {H D}} = \left[ \begin{array}{l} \mathcal {E} _ {1}, \mathcal {E} _ {2}, \dots , \mathcal {E} _ {N _ {L}} \end{array} \right], \tag {3} +$$ + +$$ +\mathcal {E} _ {k} = \mathcal {E} (\underbrace {H _ {k} \cup H _ {k + 1}} _ {s _ {i d}}, \underbrace {H _ {k} \rightarrow H _ {k + 1}} _ {s _ {c p}}, \underbrace {H _ {k + 1} \rightarrow H _ {k}} _ {s _ {c f}}), \tag {4} +$$ + +where $\mathcal{E}_k$ denotes the concatenation of the three edge subsets of $S = \{s_{id}, s_{cp}, s_{cf}\}$ and $s_{id}, s_{cp}, s_{cf}$ indicate the identity, centripetal, and centrifugal edge subsets, respectively. Through this construction policy, we create a skeletal graph with bidirectional and identity edges. + +Fully Connected Inter-Hierarchy Edges. Decomposed graph $\overleftrightarrow{\mathbf{A}}_{\mathrm{HD}}$ has a different number of edge sets from the + +conventional graph, but the edges are all the same. To identify the relationships between major distant joint nodes, especially those in the same semantic space, we connect all nodes between neighboring hierarchy node sets. In addition, since [35]'s graph contains the connectivity of only PC edges, not distant relationships, the receptive field is very small with this sparse graph. Applying our fully connected (FC) edges to the rooted tree, the graph becomes denser and makes the receptive field larger than before with more meaningful distant connectivity as shown in Fig. 2 (b). Then, the adjacency matrices are normalized with degree matrices for training stability and we leave all elements of the matrices as learnable parameters for training adaptability. + +HD-Graph Convolution. Our HD-Graph convolution includes four parallel branch operations: three graph convolution through HD-Graph and an additional EdgeConv [33] operation. To reduce the computational complexity, a linear transformation is applied to all four operations. For three of these operations, our method performs a subset-wise GCN operation in the same way as [35, 24] for each hierarchy edge set with three edge subsets. However, rather than summing the output values for each subset as in Eq. (1), we concatenate these output values to the channel dimension: + +$$ +\mathbf {F} _ {\mathrm {H D}} ^ {(k)} = \left\| _ {s \in S} \left\{\overleftrightarrow {\mathbf {A}} _ {\mathrm {H D}; s} ^ {(k)} \Phi \left(\mathbf {F} _ {i n}\right) \Theta_ {s} ^ {(k)} \right\}, \right. \tag {5} +$$ + +where $\mathbf{F}_{\mathrm{HD}}^{(k)}$ denotes the output feature map of the HD-Graph convolution and function $\Phi$ denotes a linear transformation with parameter $\mathbf{W} \in \mathbb{R}^{C' \times C}$ . Note that $\|\cdot\|$ is a concatenation operation. + +Although our HD-Graph defines more meaningful node relationships than conventional graph, it may still not be able to extract sample-wise key relationships that reflect the similarities between all nodes in the feature space. To improve this limitation, we adopt EdgeConv [33] as the remaining operation, which is used for extracting graphical features through local neighborhood graphs in the feature space. With spatial EdgeConv (S-EdgeConv), our network extracts sample-wise node connectivity, which the HD-Graph does not capture. For our method, S-EdgeConv initially takes the average pooling as the temporal dimension for computational efficiency. Local graphs with local edges are then formed via k-nearest neighbor (k-NN) based on the Euclidean distance, and the local edges as well as identity edges based on the graphs are aggregated via trainable parameters $\mathbf{W}_{edge} \in \mathbb{R}^{C' \times 2C'}$ . For the deep neural network, physically close edges are reflected to the initial shallow layers, but as they become deeper, the relationship between semantically similar edges in the feature space are identified and learned. Our whole GCN process is shown in + +![](images/ce2a61ad8f85e7fba75a75e2369c18ca05de9d336ba454eb005d37323ad5ec0b.jpg) +HD-Graph Conv. on $k$ th Hierarchy Set + +![](images/72adcd076ced693715d0de3a9a72bfbe3bd645a076bc60c736f8acf1d7bf9d91.jpg) +HD - Graph Conv. Block with A-HA + +![](images/921068da55285c260351013e722232f18c7728ce26637a2c2d5ca9b28f3de247.jpg) +Figure 3. HD-Graph convolution operation block with an A-HA module. The left side shows an operation for one hierarchy edge set, and the right side shows an operation block that concatenates the results for $N_{L}$ edge sets and applies A-HA. The lower part of the figure is EdgeConv, where the EdgeConv subscript indicates the feature space to extract graphical features. The $\times$ and $\cdot$ operations denote matrix and element-wise multiplication. +Fig. 3 and computed as follows: + +$$ +\mathbf {F} _ {\mathrm {H D}} \leftarrow \sum_ {k = 1} ^ {N _ {L}} \left[ \mathbf {F} _ {\mathrm {H D}} ^ {(k)} \left| \right| z _ {\mathcal {V}} ^ {(k)} \left(\frac {1}{T} \sum_ {t = 1} ^ {T} \Phi (\mathbf {F} _ {i n})\right) \right], \tag {6} +$$ + +where $z_{\mathcal{V}}$ denotes the S-EdgeConv operation. + +All four branch outputs are concatenated to the channel dimension, with all four computed in the same way for $N_{L}$ hierarchy edge sets. Due to the inherent characteristics of skeletal data, the number of joint nodes included in each dataset is different, and, consequently, the number of hierarchy sets is different. Therefore, we adopt an addition policy for $N_{L}$ hierarchy-wise outputs and a concatenation policy for $N_{S}$ subset-wise outputs. In this way, the dimensionality is maintained, and the common hierarchy-wise aggregation policy is followed for every skeletal dataset by adding all the outputs for different numbers of hierarchical sets. + +# 3.3. Attention-Guided Hierarchy Aggregation + +The HD-Graph convolution uses an aggregation policy of adding all the hierarchy-wise outputs. However, because each data sample has relationship between specific major + +edges, we propose an attention-guided hierarchy aggregation (A-HA) module, which applies a weighted-sum policy to the hierarchy dimension with proper attention to the hierarchy-wise outputs. The framework of the HD-Graph convolution with the A-HA module is shown in Fig. 3. + +Representative Spatial Average Pooling. Our A-HA module is applied to feature map $\mathbf{F}_{\mathrm{HD}} \in \mathbb{R}^{C \times N_L \times T \times V}$ after the HD-Graph convolution. The first step is to extract the temporal frame with the highest score on $\mathbf{F}_{\mathrm{HD}}$ . RSAP, $\Psi$ , is then applied, which is preceded by the extraction of representative nodes in each hierarchy layer. If spatial average pooling is applied without this extraction process, scaling bias occurs because the number of edges connected to each node is different. Therefore, representative node extraction is essential to obtain an appropriate score for attention without any bias. After the extraction, spatial average pooling is applied to hierarchy-wise outputs. Our pooling function $\Psi$ is as follows: + +$$ +\Psi \left(\mathbf {F} _ {\mathrm {H D}} ^ {(k)}\right) = \frac {1}{N _ {k} + N _ {k + 1}} \sum_ {v \in H _ {k} \cup H _ {k + 1}} \max _ {t} \left(\mathbf {F} _ {\mathrm {H D}} ^ {(k)} (v)\right), \tag {7} +$$ + +where $N_{k}$ denotes the number of vertices in the $H_{k}$ set. + +Hierarchical Edge Convolution. After the RSAP layer, $N_{L}$ hierarchy-wise features in attention feature map $\mathbf{M}$ have not yet shared their information with each other. We treat all $N_{L}$ features as nodes on a graph to learn and reflect similarities in the hierarchical feature space. To apply this process, representative features of these nodes are fed into EdgeConv [33], and the similarities of those nodes are learned based on the Euclidean distance. We also include the self-loop shown in the bottom section of Fig. 3 so that the node's own features can be reflected. Our attention map $\mathbf{M}$ operates as follows: + +$$ +\mathbf {M} = \sigma \left(z _ {L} \left(\left. \right| \left. \right| _ {k \in L} \left\{\Psi \left(\mathbf {F} _ {\mathrm {H D}} ^ {(k)}\right) \right\}\right)\right), \tag {8} +$$ + +where $z_{L}$ and $\sigma$ denote H-EdgeConv and the sigmoid function, respectively. + +The attention map $\mathbf{M}$ obtained is multiplied by the HD-Graph convolution output feature map $\mathbf{F}_{\mathrm{HD}}$ , and the output feature map $\mathbf{F}_{out}$ is obtained through a weighted sum to the hierarchy axis as shown in Fig. 3. Similar to S-EdgeConv in Sec. 3.2, The H-EdgeConv method incorporates the concept of hierarchical edge sets in a physically proximal manner in earlier layers, while in the deeper layers, it emphasizes the presence of semantically similar edge sets. This approach highlights different hierarchical edge sets for each sample and enables the model to learn meaningful representations that capture both physical and semantic properties of the input data. + +![](images/76c5616145acdd07134aec4bf6699640056bc71e696898ea50ed5c3115a3366e.jpg) +Figure 4. The architecture of HD-GCN. HD-GCN receives the skeleton sequence as input and obtains the class label through nine GCN blocks, and an FC layer, and the softmax function. + +# 3.4. Six-Way Ensemble + +Shi et al. [24, 25] have applied a four-stream ensemble method using streams for joints, bones, joint motion, and bone motion. However, as the performances of motion streams are relatively poorer than the performances of joint and bone streams, we adopt an ensemble method with the joint and bone streams without any motion streams. We use three different HD-Graphs, and each graph is used for training with joint and bone streams. The three HD-Graphs have different CoM nodes, which are chest, belly, hip nodes, respectively. In other words, we train joint and bone streams with HD-Graph with the CoM node of chest, and we train the same when the CoM node is belly or hip node. As models with the three different graphs should be trained in different aspects, each of the graphs is composed of different edge sets. For example, if the CoM node is belly, both thigh edges and both upper arm edges are included in the same edge set, whereas when the CoM node is chest, both thigh edges and both forearm edges are included in the same edge set. The details of our six-way ensemble are specified on our Appendix. + +# 3.5. Network Architecture + +As shown in Fig. 4, we adopt [24] as our baseline network architecture with a total of nine stacked GCN blocks. The numbers of output channels for the blocks are 64, 64, 64, 128, 128, 128, 256, 256, and 256. Each block contains a residual connection [9] and is divided into a spatial module, in which the GCN operation proceeds, and a temporal module, which includes the temporal convolutions. Our method uses the temporal module of [2], whose baseline module is [18, 28]. This module consists of four branch operations. Two are dilated temporal convolutions with kernel size five and dilation one and two, respectively. The remaining branch operations are pointwise convolution and max pooling with kernel size three. Our spatial module consists of an HD-Graph convolution operation and an A-HA module, as introduced in Sec. 3.2 and Sec. 3.3. After passing + +through all GCN layers with attention to the hierarchy-wise features, the network compresses the feature map through the global average pooling layer and classifies the action sample through the softmax function. + +# 4. Experiments + +# 4.1. Datasets and Experimental Settings + +NTU-RGB+D 60. NTU-RGB+D 60 [22] is a large dataset used in skeletal action recognition. It contains 56,880 skeleton action samples, performed by 40 different participants and classified into 60 classes. The authors of this dataset recommend two benchmarks. (1) Cross-Subject (X-Sub): 20 of the 40 subjects' actions are used for training, and the remaining 20 are for validation. (2) Cross-View (X-View): Two of the three camera-views are used for training, and the other one is used for validation. + +NTU-RGB+D 120. NTU-RGB+D 120 [16] is a dataset in which 57,367 new action samples are added to the NTU-RGB+D 60 dataset. It contains a total of 114,480 skeleton action samples over 120 classes, performed by 106 different subjects. The authors of this dataset recommend two benchmarks: (1) Cross-Subject (X-Sub): 53 of the 106 subjects' actions are used for training, and the remaining 53 are used for validation. (2) Cross-Setup (X-Set): Of the 32 setups, data with even setup IDs are used for training, and the remaining data with odd IDs are used for validation. + +Kinetics-Skeleton. The Kinetics-Skeleton dataset is derived from the Kinetics 400 video dataset [11], utilizing the OpenPose pose estimation [1] to extract 240,436 training and 19,796 testing skeleton sequences across 400 classes. The dataset restricts the number of skeletons per time step to two and eliminates skeletons with lower confidence scores, ensuring high-quality sequences for human action recognition and pose estimation research. + +Northwestern-UCLA. The Northwestern-UCLA skeleton dataset [30] contains 1494 video clips over 10 classes. Each action is captured through three Kinect cameras with different camera views and is performed by 10 subjects. We adopt the same protocol as NW-UCLA: Two of the three camera-views are used for training, and the other one is used for validation. + +Experimental Settings. In our experiments, we adopt [24] as the backbone. The SGD optimizer is employed with a Nesterov momentum of 0.9 and a weight decay of 0.0004. The number of learning epochs is set to 90, with a warm-up strategy [9] applied to the first five epochs for more stable learning. We set the learning rate to decay with cosine annealing [19], with a maximum + +
MethodsPublicationMotion StreamNTU-RGB+D 60NTU-RGB+D 120Kinetics-SkeletonNorthwestern
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)Top-1 (%)Top-5 (%)UCLA (%)
ST-GCN [35]AAAI 2018✘81.588.370.773.230.752.8-
2s-AGCN [24]CVPR 2019✘88.595.182.584.236.158.7-
SGN [36]CVPR 2020✓89.094.579.281.5--92.5
AGC-LSTM [26]CVPR 2019✘89.295.0----93.3
DGNN [23]CVPR 2019✓89.996.1--36.959.6-
Shift-GCN [5]CVPR 2020✓90.796.585.987.6--94.6
DC-GCN+ADG [4]ECCV 2020✓90.896.686.588.1--95.3
DDGCN [14]ECCV 2020✓91.197.1--38.160.8-
MS-G3D [18]CVPR 2020✘91.596.286.988.438.060.9-
MST-GCN [3]AAAI 2021✓91.596.687.588.838.160.8-
CTR-GCN [2]ICCV 2021✓92.496.888.990.6--96.5
EfficientGCN-B4 [27]TPAMI 2022✓91.795.788.389.1---
STF [12]AAAI 2022✘92.596.988.989.939.9--
InfoGCN (4-ensemble) [6]CVPR 2022✓92.796.989.490.7--96.6
InfoGCN (6-ensemble) [6]CVPR 2022✓93.097.189.891.2--97.0
HD-GCN (2-ensemble)✘92.496.689.190.638.961.796.6
HD-GCN (4-ensemble)✘93.097.089.891.240.363.096.9
HD-GCN (6-ensemble)✘93.497.290.191.640.963.597.2
+ +learning rate of 0.1 and a minimum learning rate of 0.0001. For the NTU-RGB+D datasets, we set the batch size to 64 and use the data preprocessing method from [36]. For Kinetics-Skeleton, the batch size is set to 128. In addition, to overcome the absence of belly and hip nodes in the Kinetics-Skeleton, we define the center of both hip joints as CoM hip node, and the center of chest and the hip node as CoM belly node, resulting in a total of 20 nodes. For the Northwestern-UCLA dataset, we set the batch size to 16 and use the data preprocessing method from [5]. All our experiments are conducted on a single RTX 3090 GPU. + +# 4.2. Comparison with State-of-the-Arts Methods + +Most recent state-of-the-art networks [25, 5, 4, 2] adopt a four-way ensemble method, but we adopt the six-way ensemble method described in Sec. 3.4. + +We compare ours with state-of-the-art networks on three datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Northwestern-UCLA [30], and Kinetics-Skeleton [11]. Comparisons for each dataset are shown in Tab. 1. The recognition performance of our HD-GCN has exceeded the state-of-the-arts on every dataset without any motion streams, as shown in Tab. 1. With our proposed ensemble method, HD-GCN outperforms the state-of-the-art and shows comparable performance to the 6-way ensemble state-of-the-art using only 4-way ensemble method. + +# 4.3. Ablation Study + +In this section, we demonstrate the effectiveness of the proposed HD-GCN. Performance is specified as the cross-subject and cross-setup classification accuracy on the NTU-RGB+D 120 [16] joint stream data. + +Table 1. Comparisons of the top-1 accuracy (%) against state-of-the-art methods on the NTU-RGB+D 60, 120, Northwestern-UCLA, and Kinetics-Skeleton datasets. The orange and yellow cells respectively indicate the highest and second-highest value. + +
Graph typeEdgesS-EdgeConvX-Sub (%)X-Set (%)
ConventionalPC✘83.585.4
HD-Graph
APC✘84.3 (↑ 0.8)86.1 (↑ 0.7)
BPC✓84.6 (↑ 1.1)86.3 (↑ 0.9)
CFC✘84.9 (↑ 1.4)86.5 (↑ 1.1)
DFC✓85.1 (↑ 1.6)86.7 (↑ 1.3)
+ +Table 2. Comparison of the conventional graph and four types of HD-Graph. + +
MethodH-EdgeConvX-Sub (%)X-Set (%)
Baseline✘83.585.4
HD-GCN
w/o A-HA✘85.1 (↑ 1.6)86.7 (↑ 1.3)
w/ SAP✘85.2 (↑ 1.7)86.7 (↑ 1.3)
w/ SAP✓85.4 (↑ 1.9)87.0 (↑ 1.6)
w/ RSAP✘85.5 (↑ 2.0)87.0 (↑ 1.6)
w/ RSAP✓85.7 (↑ 2.2)87.3 (↑ 1.9)
+ +Table 3. Comparison of various types of attention modules. SAP denotes the spatial average pooling. + +Hierarchically Decomposed Graph. To proceed with the ablation study for HD-Graph, we set Yan et al. [35]'s graph as the conventional graph. Here, we use the temporal convolution module of [18], as mentioned in Sec. 3.5, to compare the performance of networks fairly with various graphs. The experimental results are shown in 2. + +We set the edges of the HD-Graph in different ways to show a gradual performance increase according to the type of graph. There are four main versions of HD-Graph, the first of which is graph A containing only the PC edges. Unlike the conventional graph with one edge set including three fixed subsets, HD-Graph has a flexible number of edge sets, each divided by hierarchy layers with three subsets. Graph B is an extension of A, with the additional operation S-EdgeConv. Graph C contains FC edges for $N_{H}$ hierarchy node sets, and graph D is similar to C but includes + +![](images/fb9b86c77d6e50d56a0b1c929499240e687be1cbcdf22485b7ade65a62477c99.jpg) +Figure 5. Hierarchy-wise attention scores by A-HA for "Running" and "Kicking" class. Each score indicates the value of the attention map M of the A-HA module. + +S-EdgeConv. The HD-Graph with only PC edges performs better than the conventional graph by a large margin, even though they share the same edges. This proves that it is meaningful to divide the joint nodes by hierarchy edge sets. In addition, the HD-Graph with FC edges and S-EdgeConv performs better for every datasets. + +Attention-Guided Hierarchy Aggregation. To prove the effectiveness of the A-HA module, we use a method to change or remove specific parts of our attention module, with the results shown in Tab. 3. Spatial average pooling (SAP) simply averages along the spatial axis without the representative node extraction process, which performs worse than our RSAP. The poorer performance is due to two factors: (1) scaling bias occurs because the number of nodes in each hierarchy node set is different, and (2) attention through SAP does not represent the corresponding hierarchy node set because it brings the average of the feature vectors of all nodes, not a specific node set. Furthermore, it performs better with H-EdgeConv, which recognizes each hierarchy edge set as a graph node. This proves that because the major edge sets are different for each data sample, it is important to find and highlight edge sets with high similarity based on the Euclidean distance through H-EdgeConv. + +The results of the attention score $\mathbf{M}$ of our A-HA module are shown in Fig. 5. These results show that our module scores edge sets 4 and 5 higher for the "running" class, which includes knees and feet, elbows and hands. For the "Kicking" class, A-HA gives the highest score to edge set 3, which includes shoulders and hips, followed by edge set 4 and 5. It is reasonable for human visual recognition that the dynamically moving edge set 4, 5 are more important than the stationary and barely moving edge set 3 when running + +
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
DC-GCN [4]84.0*86.1*2.743.45
MS-G3D [18]84.9*86.8*5.223.22
CTR-GCN [2]84.986.5*1.971.46
InfoGCN [6]85.186.31.681.57
HD-GCN85.787.31.601.68
+ +Table 4. Comparison of complexity of the single-stream state-of-the-arts. $\star$ : results obtained by the released codes. + +
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
CTR-GCN † [2]88.990.67.885.84
InfoGCN † [6]89.490.76.726.28
InfoGCN ‡ [6]89.891.210.089.42
HD-GCN †89.891.26.406.72
HD-GCN ‡90.191.69.6010.08
+ +Table 5. Comparison of complexity of the multi-stream state-of-the-arts. †: 4-ensemble, ‡: 6-ensemble + +rather than when kicking something. + +Six-Way Ensemble. We use the ensemble method to which three graphs with different CoM nodes are applied, excluding motion streams. Tab. 1 shows that the HD-GCN with 4-way ensemble outperforms the state-of-the-art [6] 4-way methods with motion data and shows comparable performance with the state-of-the-art 6-way method. In addition, when the 6-way ensemble with three different graphs is applied to HD-GCN, it outperforms the state-of-the-art methods. This proves that the features extracted with different CoM nodes are learned in different learning aspects. + +# 4.4. Comparison of Complexity with Other Models + +Although our model has multiple branch layers for multiple edge sets, it does not cause high complexity because it precedes channel reduction layers. Comparisons of computational complexity with other models are shown in Tab. 4, where the window size is fixed to 64. Our model shows the best performance on NTU-RGB+D 120 joint stream by a large margin even though the computational complexity of our model is the lowest. For multi-stream ensemble, our 4-stream HD-GCN shows almost similar performance to 6-stream InfoGCN [6] while having 3.68G fewer FLOPs and 2.70M fewer parameters as shown in Tab. 5. + +# 5. Conclusions + +In this work, we propose a novel hierarchically decomposed graph convolutional network (HD-GCN) for skeleton-based action recognition. We also propose a new framework (HD-Graph) that replaces the existing framework, decomposes all the joint nodes by hierarchy edge sets and considers the connectivity between major distant nodes, which is difficult to identify naturally. We also present an effective attention module (A-HA) composed of representative spatial average pooling (RSAP) layer and hierarchical + +edge convolution (H-EdgeConv), which applies hierarchywise attention for the HD-Graph. In addition, our HD-GCN learns graph-wise features with different patterns through a six-way ensemble method. We derive an effective feature extractor by combining these three methods and empirically verify its effectiveness. Our approach outperforms current state-of-the-art methods on four benchmark datasets. + +Acknowledgements. This work was supported by Alon-Flow Research and the KIST Institutional Program (Project No.2E32283-23-064). + +# References + +[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 6 +[2] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 1, 2, 3, 6, 7, 8 +[3] Zhan Chen, Sicheng Li, Bing Yang, Qinghan Li, and Hong Liu. Multi-scale spatial temporal graph convolutional network for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1113-1122, 2021. 7 +[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 7, 8 +[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1, 2, 7 +[6] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogen: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 7, 8 +[7] David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alán Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems, 28, 2015. 3 +[8] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 3 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 + +[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3 +[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 7 +[12] Lipeng Ke, Kuan-Chuan Peng, and Siwei Lyu. Towards to-at spatio-temporal focus for skeleton-based action recognition. arXiv preprint arXiv:2202.02314, 2022. 7 +[13] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning, pages 2688-2697. PMLR, 2018. 3 +[14] Matthew Korban and Xin Li. Ddgcn: A dynamic directed graph convolutional network for action recognition. In European Conference on Computer Vision, pages 761-776. Springer, 2020. 7 +[15] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, and Sangyoun Lee. Leveraging spatio-temporal dependency for skeleton-based action recognition. arXiv preprint arXiv:2212.04761, 2022. 1 +[16] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 2, 6, 7 +[17] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In European conference on computer vision, pages 816-833. Springer, 2016. +[18] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 143-152, 2020. 1, 2, 6, 7, 8 +[19] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[20] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and Michael M Bronstein. Geometric deep learning on graphs and manifolds using mixture model cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5115-5124, 2017. 3 +[21] Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. Learning convolutional neural networks for graphs. In International conference on machine learning, pages 2014-2023. PMLR, 2016. 3 +[22] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu $\mathrm{rgb + }$ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016. 2, 6, 7 +[23] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 7 +[24] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1, 2, 3, 4, 6, 7 +[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. IEEE Transactions on Image Processing, 29:9532-9545, 2020. 1, 2, 3, 6, 7 +[26] Chenyang Si, Wentao Chen, Wei Wang, Liang Wang, and Tieniu Tan. An attention enhanced graph convolutional LSTM network for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1227-1236, 2019. 1, 7 +[27] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 7 +[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6 +[29] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 1 +[30] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning and recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2649-2656, 2014. 2, 6, 7 +[31] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1 +[32] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 3 +[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 4, 5 +[34] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3 +[35] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1, 2, 3, 4, 7 +[36] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural + +networks for efficient skeleton-based human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1112-1121, 2020. 1, 7 \ No newline at end of file diff --git a/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/images.zip b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eff0785e26e2f54d663a37fa1f495d5d63617b01 --- /dev/null +++ b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4497b3df0be90b66cb01b6f762fe6af756b7ed92a7d508bfa7c1be5462039c44 +size 421612 diff --git a/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/layout.json b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7277ccf4789394f100a9ac7a7f475932932965c1 --- /dev/null +++ b/2023/Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition/layout.json @@ -0,0 +1,7374 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 140 + ], + "type": "text", + "content": "Hierarchically Decomposed Graph Convolutional Networks for Skeleton-Based Action Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "text", + "content": "Jungho Lee" + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "text", + "content": ", Minhyeok Lee" + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "text", + "content": ", Dogyoon Lee" + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "text", + "content": ", Sangyoun Lee" + }, + { + "bbox": [ + 138, + 160, + 455, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 189, + 460, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 189, + 460, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 189, + 460, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 189, + 460, + 217 + ], + "type": "text", + "content": "School of Electrical and Electronic Engineering, Yonsei University " + }, + { + "bbox": [ + 133, + 189, + 460, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 189, + 460, + 217 + ], + "type": "text", + "content": "AIonFlow Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 220, + 445, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 220, + 445, + 232 + ], + "spans": [ + { + "bbox": [ + 147, + 220, + 445, + 232 + ], + "type": "text", + "content": "{2015142131,hydragon516,nemotio,sylee}@yonsei.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 259, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 259, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 259, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 285, + 290, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 290, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 290, + 561 + ], + "type": "text", + "content": "Graph convolutional networks (GCNs) are the most commonly used methods for skeleton-based action recognition and have achieved remarkable performance. Generating adjacency matrices with semantically meaningful edges is particularly important for this task, but extracting such edges is challenging problem. To solve this, we propose a hierarchically decomposed graph convolutional network (HD-GCN) architecture with a novel hierarchically decomposed graph (HD-Graph). The proposed HD-GCN effectively decomposes every joint node into several sets to extract major structurally adjacent and distant edges, and uses them to construct an HD-Graph containing those edges in the same semantic spaces of a human skeleton. In addition, we introduce an attention-guided hierarchy aggregation (A-HA) module to highlight the dominant hierarchical edge sets of the HD-Graph. Furthermore, we apply a new six-way ensemble method, which uses only joint and bone stream without any motion stream. The proposed model is evaluated and achieves state-of-the-art performance on four large, popular datasets. Finally, we demonstrate the effectiveness of our model with various comparative experiments. Code is available at https://github.com/Jho-Yonsei/HD-GCN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Human action recognition (HAR) is a task that categorizes action classes by receiving video data as input. HAR is used in many applications, such as human-computer interaction and virtual reality. Recently, several RGB-based and skeleton-based HAR methods have been proposed with the development of deep learning technology. However, RGB-based methods [31, 29] cannot robustly recognize human actions because they are strongly influenced by environmental noises such as background color, brightness of light," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 261, + 538, + 396 + ], + "blocks": [ + { + "bbox": [ + 319, + 261, + 538, + 396 + ], + "lines": [ + { + "bbox": [ + 319, + 261, + 538, + 396 + ], + "spans": [ + { + "bbox": [ + 319, + 261, + 538, + 396 + ], + "type": "image", + "image_path": "561382c75aab24245f23319b495f169ee1faddbc9fa580c31210991f4432bc30.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 401, + 547, + 479 + ], + "lines": [ + { + "bbox": [ + 304, + 401, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 547, + 479 + ], + "type": "text", + "content": "Figure 1. The framework of HD-GCN. The input skeleton is applied with various edge sets through a hierarchically decomposed graph (HD-Graph). The red lines are the edges included in the corresponding hierarchy edge set. The network highlights the major edge sets through the attention map. The darker the color of red line, the more highlighted the edge set, and dotted lines denote unconnected edges." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "content": "and clothing. Therefore, methods using skeleton modality [35, 24, 36, 26, 5, 4, 18, 2, 15] have attracted attention because they are not affected by these noises. These methods recognize action by receiving 2D or 3D coordinates of major human joints as time-series inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": "Recent approaches [24, 18, 4, 2] have adopted graph convolutional networks (GCNs) to apply human-skeleton graphs to convolutional layers. However, existing GCN-based methods [35, 24, 25, 4, 2] have the following limitations. (1) With the widely used handcrafted graph, the relationships between distant joint nodes are not identified since they use only the relationships of PC edges in the human skeleton. Although the graph with PC edges has a semantic significance, the graph with only PC edges suffers from long-range dependency problem as they are heuristically fixed. However, for humans to recognize actions, relationships between structurally distant joints as well as between adjacent joints are strongly correlated. Although several methods [24, 2] have attempted to solve such limita" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10444" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "tion by training attention-guided learnable graphs, they still use [35]'s handcrafted graph with their learnable graphs. Moreover, as the element values of [35]'s graph are more dominant than those of the learnable graphs, they do not adequately highlight the relationships between distant nodes. (2) Some recent methods [35, 24, 4, 18] risk falling into suboptimality by simply aggregating the edge features and ignoring the contribution of each edge, thus incompletely recognizing which edges are significant for each skeleton sample. For example, in the case of a 'squat down' action, the interactions between the legs and arms should be highlighted." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 302 + ], + "type": "text", + "content": "Motivated by these limitations, we propose a hierarchically decomposed graph convolutional network (HD-GCN) with a hierarchically decomposed graph (HD-Graph) and attention-guided hierarchy aggregation (A-HA) module. In addition, we present a six-way ensemble method to effectively utilize our HD-Graph. The framework of our proposed methods is shown in Fig. 1 for 'squat down' action." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 568 + ], + "type": "text", + "content": "The HD-GCN incorporates GCNs with our HD-Graph, which identifies the relationships between distant joint nodes in the same semantic spaces (e.g. right and left hands, right and left feet). The same semantic spaces are formed by moving out step by step from the Center of Mass (CoM) node of the graph. For example, if belly is a CoM node, the first semantic space includes the belly node, the next space includes the chest and hip nodes, and the subsequent space includes the left and right shoulder and the left and right hip nodes. The nodes in the same semantic space are defined as hierarchy node set. To detect the relationships between distant joint nodes, network should have large receptive field. The proposed HD-Graph contains both meaningful adjacent and distant joint nodes by connecting all the nodes in neighboring hierarchy node sets and identifies the connectivity between those nodes for large receptive field. We adopt rooted tree-like structure to effectively represent every edges. We apply a spatial edge convolution (S-EdgeConv) layer to consider semantically close edges which cannot be captured by the HD-Graph for each sample. To create the S-EdgeConv layer, we borrow the structure of [33], which is widely used in 3D point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "content": "To consider the contribution of each edge set, the process of selecting the dominant hierarchical information should depend on the action data sample to give proper attention to the most dominant edge sets. For example, in order to recognize the \"clapping\" action, a hierarchy edge set that includes both hands must be emphasized. To tackle this issue, we propose an attention-guided hierarchy aggregation (A-HA) module, which consists of two submodules: representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv). A scaling bias problem occurs if we use the spatial average pooling module without any node extraction process because each node has a different" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "number of adjacent nodes. To prevent this, we apply RSAP, which includes a representative node extraction process that triggers features after the pooling layer to represent each node. To effectively manage hierarchical features obtained by RSAP, we apply a hierarchical edge convolution (H-EdgeConv) layer. The H-EdgeConv treats each hierarchical feature as a graph node and identifies which hierarchical features should be highlighted via the Euclidean distance in feature space. With the RSAP and the H-EdgeConv, our model successfully determines which hierarchy edge sets and joints should be emphasized among the input features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 205, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 359 + ], + "type": "text", + "content": "The existing ensemble method uses four-stream data composed of the joint, bone, joint motion, and bone motion streams, which are the original skeletal coordinates, spatial differential between joint coordinates, and temporal differential of joint, and temporal differential of the bone, respectively. Most existing ensemble methods [25, 2] use additional motion data, but models that solely utilize motion data exhibit relatively inferior performance. To address this problem, we present a new method, a six-way ensemble. We apply this ensemble method by setting three HDGraphs with joint and bone stream data. Each graph has different CoM nodes to extract features of different semantic spaces (see Appendix)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 360, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 407 + ], + "type": "text", + "content": "We conduct extensive experiments on four benchmark action recognition datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Kinetics-Skeleton [11], and Northwestern-UCLA [30]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 408, + 525, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 408, + 525, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 408, + 525, + 419 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 429, + 545, + 603 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 316, + 429, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 545, + 464 + ], + "type": "text", + "content": "- We propose a hierarchically decomposed graph (HDGraph) to thoroughly identify the significant distant edges between the same hierarchy node sets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 471, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 545, + 519 + ], + "type": "text", + "content": "- We propose an attention-guided hierarchy aggregation (A-HA) module to highlight the key edge sets with representative spatial average pooling (RSAP) and hierarchical edge convolution (H-EdgeConv)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 525, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 545, + 573 + ], + "type": "text", + "content": "- We use a new six-way ensemble method for skeleton-based action recognition with HD-Graphs that have different center of mass (CoM), which outperforms regular ensemble without any motion data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 579, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 545, + 603 + ], + "type": "text", + "content": "- Our HD-GCN outperforms the state-of-the-arts on four benchmarks for skeleton-based action recognition." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 615, + 392, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 615, + 392, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 615, + 392, + 627 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 635, + 473, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 473, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 473, + 647 + ], + "type": "text", + "content": "2.1. Action Recognition with GCNs" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "In skeleton-based action recognition, human skeletal data are represented by a graph with joint nodes. Most recent approaches use GCN-based methods [35, 24, 5, 18, 2] with [35]'s graph structure, which identifies physical connections in the human skeleton. Those GCN-based meth" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10445" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 286, + 281 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 286, + 281 + ], + "type": "image", + "image_path": "6431cfc3343be6978763e7228e41897481c24b8d0b638e94f751810a862b6d27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 287, + 546, + 332 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 546, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 546, + 332 + ], + "type": "text", + "content": "Figure 2. (a) Structure of HD-Graph with physically connected (PC) edges. The human skeleton graph is decomposed into a rooted tree, where PC edges are included in hierarchy sets. (b) Structure of HD-Graph with fully connected (FC) edges. Edges between all nodes in the same semantic space are obtained by connecting all the nodes in adjacent hierarchy edge sets. Blue and red lines stand for PC and FC edges, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 535, + 281 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 535, + 281 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 535, + 281 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 535, + 281 + ], + "type": "image", + "image_path": "f59d474ca5410eb86c120ce0161f0c1cd7bf7ed1832e26fb3834ac67d5a50a81.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 346, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 289, + 502 + ], + "type": "text", + "content": "ods perform remarkably better than methods using handcrafted features [7, 8, 13, 20, 21]. They extract the spatial features representing the relationships between physically connected edges among human skeleton, and they outperform other methods by using them to construct the major relationships between joint nodes in the human skeleton. In particular, [25] and [2] propose adaptive attention-based graph structures to learn the sample-wise topological features. However, they might fall into suboptimality because they do not consider the physical prior of the human skeletal structure and allow too much flexibility in network training. To address this issue, we introduce a novel HD-Graph, referencing the known tendencies of human perception" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 531, + 266, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 266, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 266, + 544 + ], + "type": "text", + "content": "2.2. Attention Modules for Action Recognition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 712 + ], + "type": "text", + "content": "The attention mechanism is an essential element for constructing a deep neural network. Using recent attention modules [10, 34], networks emphasize important information along a specific dimension. For example, Hu et al. [10] applies channel-wise attention, and Woo et al. [34] applies both channel-wise and spatial-wise attention. These techniques are divided into two categories for GCNs: (1) attention-based graph construction [24, 2] which is a method of forming topologies using a non-local block [32] or customized correlation matrices, and (2) spatial-wise, temporal-wise, channel-wise attention, which are commonly used attentions in [25, 27], and several other networks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 345, + 388, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 388, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 388, + 358 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 365, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 461 + ], + "type": "text", + "content": "In Sec. 3.2, we detail the HD-Graph convolution to solve the problems of the conventional human-skeleton graph [35], which includes only PC edges. We also explain the A-HA module in Sec. 3.3 to highlight dominant hierarchical features. In Sec. 3.4, we replace the widely used four-stream ensemble method [25, 2, 4] with a six-way ensemble without motion data streams. Finally, we introduce the HD-GCN, which uses these proposed methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 468, + 391, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 391, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 391, + 480 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": "Notations. The spatio-temporal graph for human skeleton is represented by " + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": " denote the joint and edge groups, respectively. Physically connected edges and fully-connected edges used in Sec. 3.2 are denoted as PC-edges and FC-edges, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "text", + "content": "Graph Convolutional Networks. 3D time-series skeletal data are represented by " + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{3 \\times T \\times V}" + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "text", + "content": " are the number of joint nodes and the temporal window size, respectively. GCN's operation with input feature map " + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{in} \\in \\mathbb{R}^{C \\times T \\times V}" + }, + { + "bbox": [ + 304, + 562, + 545, + 622 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 630, + 545, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 630, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 377, + 630, + 545, + 656 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {o u t} = \\sum_ {s \\in S} \\widehat {\\mathbf {A}} _ {s} \\mathbf {F} _ {i n} \\Theta_ {s}, \\tag {1}", + "image_path": "5627e760d916369e629688ae67d6c713a698b8b27ebc53ac601cd9fcd213954c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "S = \\{s_{id}, s_{cf}, s_{cp}\\}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " denotes graph subsets, and " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "s_{id}, s_{cf}," + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "s_{cp}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " indicate identity, centrifugal, and centripetal joint subsets, respectively. " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\Theta_s" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " denotes the pointwise convolution operation. The normalized adjacency matrix " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{A}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10446" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "text", + "content": "initialized as " + }, + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\Lambda^{-\\frac{1}{2}}\\mathbf{A}\\Lambda^{-\\frac{1}{2}}\\in \\mathbb{R}^{N_S\\times V\\times V}" + }, + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "text", + "content": " is a diagonal matrix for normalization and " + }, + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "inline_equation", + "content": "N_{S} = 3" + }, + { + "bbox": [ + 46, + 70, + 287, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 105, + 234, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 234, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 234, + 118 + ], + "type": "text", + "content": "3.2. Hierarchically Decomposed Graph" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 124, + 287, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 124, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 124, + 287, + 172 + ], + "type": "text", + "content": "Most recent methods have adopted the handcrafted graph proposed by Yan et al. [35], but the HD-Graph is derived through a newly presented method. Fig. 2 shows the framework of the HD-Graph." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "text", + "content": "Decomposition into a Rooted Tree. The first step is to decompose the graph with PC edges and construct a rooted tree. To decompose a given skeleton into the tree, we need to determine a CoM node, which allows nodes in the same hierarchy edge set to exist in the same semantic space. For example, nodes in the same semantic space, such as elbow and knee joints, or hands and feet, must exist in a hierarchy node set. After choosing the CoM node, the graph is converted into a rooted tree, which includes the hierarchical information of the graph, and defines the directed adjacency matrix " + }, + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{A}}_{\\mathrm{HD}} \\in \\mathbb{R}^{N_L \\times V \\times V}" + }, + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "inline_equation", + "content": "N_L" + }, + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "text", + "content": " hierarchy layers for " + }, + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "inline_equation", + "content": "N_H" + }, + { + "bbox": [ + 46, + 189, + 288, + 334 + ], + "type": "text", + "content": " hierarchy edge sets:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 342, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\overrightarrow {\\mathbf {A}} _ {\\mathrm {H D}} = \\left[ \\mathcal {E} \\left(H _ {1} \\rightarrow H _ {2}\\right), \\dots , \\mathcal {E} \\left(H _ {N _ {H} - 1} \\rightarrow H _ {N _ {H}}\\right)\\right], \\tag {2}", + "image_path": "f04913cb28b0ba87d45ec6407897652dc13932203be558bf8f4683abdbe7b22d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "H_{k}" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": "-th hierarchy node set and " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(H_k\\to" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "H_{k + 1})" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": " denotes a set of edges from " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "H_{k}" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "H_{k + 1}" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "N_{H}" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": " are the number of hierarchy layers and hierarchy edge sets, respectively, and " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "N_{L} = N_{H} - 1" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{HD}}" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": " includes only the directed centrifugal edges. For consistency with existing methods, all the reverse-directed edges from the leaf nodes of the rooted tree in Fig. 2 to the CoM node must be reflected in the adjacency matrices to cover the centripetal edges. In addition, to get the features of the nodes themselves, the identity edges for each hierarchy node set must be considered. Thus, the adjacency matrices " + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\stackrel {\\leftrightarrow}{\\mathbf{A}}_{\\mathrm{HD}}\\in \\mathbb{R}^{N_L\\times N_S\\times V\\times V}" + }, + { + "bbox": [ + 46, + 369, + 287, + 517 + ], + "type": "text", + "content": " are defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 524, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 524, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 102, + 524, + 287, + 544 + ], + "type": "interline_equation", + "content": "\\overleftrightarrow {\\mathbf {A}} _ {\\mathrm {H D}} = \\left[ \\begin{array}{l} \\mathcal {E} _ {1}, \\mathcal {E} _ {2}, \\dots , \\mathcal {E} _ {N _ {L}} \\end{array} \\right], \\tag {3}", + "image_path": "5b97b2df975ab1ebfc8d0549382ee9eb9fe7cf14466b284c0903c211e9283367.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 573, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 573, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 573, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {k} = \\mathcal {E} (\\underbrace {H _ {k} \\cup H _ {k + 1}} _ {s _ {i d}}, \\underbrace {H _ {k} \\rightarrow H _ {k + 1}} _ {s _ {c p}}, \\underbrace {H _ {k + 1} \\rightarrow H _ {k}} _ {s _ {c f}}), \\tag {4}", + "image_path": "f1bf823fbaa022598ef5b7cdcea6e3f1a9dbd6d04a527634f1486af494c9517c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_k" + }, + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "text", + "content": " denotes the concatenation of the three edge subsets of " + }, + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "inline_equation", + "content": "S = \\{s_{id}, s_{cp}, s_{cf}\\}" + }, + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "inline_equation", + "content": "s_{id}, s_{cp}, s_{cf}" + }, + { + "bbox": [ + 46, + 611, + 287, + 672 + ], + "type": "text", + "content": " indicate the identity, centripetal, and centrifugal edge subsets, respectively. Through this construction policy, we create a skeletal graph with bidirectional and identity edges." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 688, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 688, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 688, + 287, + 714 + ], + "type": "text", + "content": "Fully Connected Inter-Hierarchy Edges. Decomposed graph " + }, + { + "bbox": [ + 46, + 688, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{HD}}" + }, + { + "bbox": [ + 46, + 688, + 287, + 714 + ], + "type": "text", + "content": " has a different number of edge sets from the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "conventional graph, but the edges are all the same. To identify the relationships between major distant joint nodes, especially those in the same semantic space, we connect all nodes between neighboring hierarchy node sets. In addition, since [35]'s graph contains the connectivity of only PC edges, not distant relationships, the receptive field is very small with this sparse graph. Applying our fully connected (FC) edges to the rooted tree, the graph becomes denser and makes the receptive field larger than before with more meaningful distant connectivity as shown in Fig. 2 (b). Then, the adjacency matrices are normalized with degree matrices for training stability and we leave all elements of the matrices as learnable parameters for training adaptability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 376 + ], + "type": "text", + "content": "HD-Graph Convolution. Our HD-Graph convolution includes four parallel branch operations: three graph convolution through HD-Graph and an additional EdgeConv [33] operation. To reduce the computational complexity, a linear transformation is applied to all four operations. For three of these operations, our method performs a subset-wise GCN operation in the same way as [35, 24] for each hierarchy edge set with three edge subsets. However, rather than summing the output values for each subset as in Eq. (1), we concatenate these output values to the channel dimension:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 385, + 545, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 385, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 351, + 385, + 545, + 412 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} = \\left\\| _ {s \\in S} \\left\\{\\overleftrightarrow {\\mathbf {A}} _ {\\mathrm {H D}; s} ^ {(k)} \\Phi \\left(\\mathbf {F} _ {i n}\\right) \\Theta_ {s} ^ {(k)} \\right\\}, \\right. \\tag {5}", + "image_path": "01df08a0709d31b61bd91c9c59132e890ded518ce15b269c5c222f94144c6813.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{HD}}^{(k)}" + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "text", + "content": " denotes the output feature map of the HD-Graph convolution and function " + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "text", + "content": " denotes a linear transformation with parameter " + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{C' \\times C}" + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 304, + 425, + 545, + 473 + ], + "type": "text", + "content": " is a concatenation operation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": "Although our HD-Graph defines more meaningful node relationships than conventional graph, it may still not be able to extract sample-wise key relationships that reflect the similarities between all nodes in the feature space. To improve this limitation, we adopt EdgeConv [33] as the remaining operation, which is used for extracting graphical features through local neighborhood graphs in the feature space. With spatial EdgeConv (S-EdgeConv), our network extracts sample-wise node connectivity, which the HD-Graph does not capture. For our method, S-EdgeConv initially takes the average pooling as the temporal dimension for computational efficiency. Local graphs with local edges are then formed via k-nearest neighbor (k-NN) based on the Euclidean distance, and the local edges as well as identity edges based on the graphs are aggregated via trainable parameters " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{edge} \\in \\mathbb{R}^{C' \\times 2C'}" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": ". For the deep neural network, physically close edges are reflected to the initial shallow layers, but as they become deeper, the relationship between semantically similar edges in the feature space are identified and learned. Our whole GCN process is shown in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10447" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 109, + 167, + 286 + ], + "blocks": [ + { + "bbox": [ + 75, + 80, + 143, + 100 + ], + "lines": [ + { + "bbox": [ + 75, + 80, + 143, + 100 + ], + "spans": [ + { + "bbox": [ + 75, + 80, + 143, + 100 + ], + "type": "text", + "content": "HD-Graph Conv. on " + }, + { + "bbox": [ + 75, + 80, + 143, + 100 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 75, + 80, + 143, + 100 + ], + "type": "text", + "content": " th Hierarchy Set" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 109, + 167, + 286 + ], + "lines": [ + { + "bbox": [ + 51, + 109, + 167, + 286 + ], + "spans": [ + { + "bbox": [ + 51, + 109, + 167, + 286 + ], + "type": "image", + "image_path": "ce2a61ad8f85e7fba75a75e2369c18ca05de9d336ba454eb005d37323ad5ec0b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 106, + 279, + 285 + ], + "blocks": [ + { + "bbox": [ + 184, + 80, + 269, + 99 + ], + "lines": [ + { + "bbox": [ + 184, + 80, + 269, + 99 + ], + "spans": [ + { + "bbox": [ + 184, + 80, + 269, + 99 + ], + "type": "text", + "content": "HD - Graph Conv. Block with A-HA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 106, + 279, + 285 + ], + "lines": [ + { + "bbox": [ + 173, + 106, + 279, + 285 + ], + "spans": [ + { + "bbox": [ + 173, + 106, + 279, + 285 + ], + "type": "image", + "image_path": "72adcd076ced693715d0de3a9a72bfbe3bd645a076bc60c736f8acf1d7bf9d91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 296, + 278, + 355 + ], + "blocks": [ + { + "bbox": [ + 53, + 296, + 278, + 355 + ], + "lines": [ + { + "bbox": [ + 53, + 296, + 278, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 278, + 355 + ], + "type": "image", + "image_path": "921068da55285c260351013e722232f18c7728ce26637a2c2d5ca9b28f3de247.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "type": "text", + "content": "Figure 3. HD-Graph convolution operation block with an A-HA module. The left side shows an operation for one hierarchy edge set, and the right side shows an operation block that concatenates the results for " + }, + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "type": "text", + "content": " edge sets and applies A-HA. The lower part of the figure is EdgeConv, where the EdgeConv subscript indicates the feature space to extract graphical features. The " + }, + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 363, + 287, + 440 + ], + "type": "text", + "content": " operations denote matrix and element-wise multiplication." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 447, + 178, + 459 + ], + "lines": [ + { + "bbox": [ + 47, + 447, + 178, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 178, + 459 + ], + "type": "text", + "content": "Fig. 3 and computed as follows:" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 468, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 468, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 78, + 468, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {H D}} \\leftarrow \\sum_ {k = 1} ^ {N _ {L}} \\left[ \\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} \\left| \\right| z _ {\\mathcal {V}} ^ {(k)} \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} \\Phi (\\mathbf {F} _ {i n})\\right) \\right], \\tag {6}", + "image_path": "24edaa482670383a1c1e85322e1f1175ef1443d472426f612a4c504c1f89e876.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 506, + 230, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 230, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 230, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 506, + 230, + 517 + ], + "type": "inline_equation", + "content": "z_{\\mathcal{V}}" + }, + { + "bbox": [ + 47, + 506, + 230, + 517 + ], + "type": "text", + "content": " denotes the S-EdgeConv operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "text", + "content": "All four branch outputs are concatenated to the channel dimension, with all four computed in the same way for " + }, + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "text", + "content": " hierarchy edge sets. Due to the inherent characteristics of skeletal data, the number of joint nodes included in each dataset is different, and, consequently, the number of hierarchy sets is different. Therefore, we adopt an addition policy for " + }, + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "text", + "content": " hierarchy-wise outputs and a concatenation policy for " + }, + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "inline_equation", + "content": "N_{S}" + }, + { + "bbox": [ + 46, + 518, + 287, + 649 + ], + "type": "text", + "content": " subset-wise outputs. In this way, the dimensionality is maintained, and the common hierarchy-wise aggregation policy is followed for every skeletal dataset by adding all the outputs for different numbers of hierarchical sets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 263, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 263, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 263, + 671 + ], + "type": "text", + "content": "3.3. Attention-Guided Hierarchy Aggregation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "The HD-Graph convolution uses an aggregation policy of adding all the hierarchy-wise outputs. However, because each data sample has relationship between specific major" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "edges, we propose an attention-guided hierarchy aggregation (A-HA) module, which applies a weighted-sum policy to the hierarchy dimension with proper attention to the hierarchy-wise outputs. The framework of the HD-Graph convolution with the A-HA module is shown in Fig. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "text", + "content": "Representative Spatial Average Pooling. Our A-HA module is applied to feature map " + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{HD}} \\in \\mathbb{R}^{C \\times N_L \\times T \\times V}" + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "text", + "content": " after the HD-Graph convolution. The first step is to extract the temporal frame with the highest score on " + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{HD}}" + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "text", + "content": ". RSAP, " + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "text", + "content": ", is then applied, which is preceded by the extraction of representative nodes in each hierarchy layer. If spatial average pooling is applied without this extraction process, scaling bias occurs because the number of edges connected to each node is different. Therefore, representative node extraction is essential to obtain an appropriate score for attention without any bias. After the extraction, spatial average pooling is applied to hierarchy-wise outputs. Our pooling function " + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 147, + 545, + 300 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 306, + 545, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 306, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 318, + 306, + 545, + 333 + ], + "type": "interline_equation", + "content": "\\Psi \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)}\\right) = \\frac {1}{N _ {k} + N _ {k + 1}} \\sum_ {v \\in H _ {k} \\cup H _ {k + 1}} \\max _ {t} \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)} (v)\\right), \\tag {7}", + "image_path": "55182423f29d4fdea0a462ebb44e437d6157dab4c19558fbd192e2823b0a56f2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 340, + 529, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 529, + 351 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 529, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 340, + 529, + 351 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 306, + 340, + 529, + 351 + ], + "type": "text", + "content": " denotes the number of vertices in the " + }, + { + "bbox": [ + 306, + 340, + 529, + 351 + ], + "type": "inline_equation", + "content": "H_{k}" + }, + { + "bbox": [ + 306, + 340, + 529, + 351 + ], + "type": "text", + "content": " set." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "content": "Hierarchical Edge Convolution. After the RSAP layer, " + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "content": " hierarchy-wise features in attention feature map " + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "content": " have not yet shared their information with each other. We treat all " + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "content": " features as nodes on a graph to learn and reflect similarities in the hierarchical feature space. To apply this process, representative features of these nodes are fed into EdgeConv [33], and the similarities of those nodes are learned based on the Euclidean distance. We also include the self-loop shown in the bottom section of Fig. 3 so that the node's own features can be reflected. Our attention map " + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "content": " operates as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 505, + 545, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 505, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 348, + 505, + 545, + 538 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\sigma \\left(z _ {L} \\left(\\left. \\right| \\left. \\right| _ {k \\in L} \\left\\{\\Psi \\left(\\mathbf {F} _ {\\mathrm {H D}} ^ {(k)}\\right) \\right\\}\\right)\\right), \\tag {8}", + "image_path": "0c7ac0198c4b6832d045c3cc74b35e789b4ac916301b03c40600241a126bfd44.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 544, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 544, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 544, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 546, + 544, + 569 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 304, + 546, + 544, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 544, + 569 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 546, + 544, + 569 + ], + "type": "text", + "content": " denote H-EdgeConv and the sigmoid function, respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "The attention map " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " obtained is multiplied by the HD-Graph convolution output feature map " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{HD}}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", and the output feature map " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{out}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " is obtained through a weighted sum to the hierarchy axis as shown in Fig. 3. Similar to S-EdgeConv in Sec. 3.2, The H-EdgeConv method incorporates the concept of hierarchical edge sets in a physically proximal manner in earlier layers, while in the deeper layers, it emphasizes the presence of semantically similar edge sets. This approach highlights different hierarchical edge sets for each sample and enables the model to learn meaningful representations that capture both physical and semantic properties of the input data." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10448" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 286, + 201 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 286, + 201 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 286, + 201 + ], + "type": "image", + "image_path": "76c5616145acdd07134aec4bf6699640056bc71e696898ea50ed5c3115a3366e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 288, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 237 + ], + "type": "text", + "content": "Figure 4. The architecture of HD-GCN. HD-GCN receives the skeleton sequence as input and obtains the class label through nine GCN blocks, and an FC layer, and the softmax function." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 244, + 158, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 244, + 158, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 244, + 158, + 257 + ], + "type": "text", + "content": "3.4. Six-Way Ensemble" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 264, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 504 + ], + "type": "text", + "content": "Shi et al. [24, 25] have applied a four-stream ensemble method using streams for joints, bones, joint motion, and bone motion. However, as the performances of motion streams are relatively poorer than the performances of joint and bone streams, we adopt an ensemble method with the joint and bone streams without any motion streams. We use three different HD-Graphs, and each graph is used for training with joint and bone streams. The three HD-Graphs have different CoM nodes, which are chest, belly, hip nodes, respectively. In other words, we train joint and bone streams with HD-Graph with the CoM node of chest, and we train the same when the CoM node is belly or hip node. As models with the three different graphs should be trained in different aspects, each of the graphs is composed of different edge sets. For example, if the CoM node is belly, both thigh edges and both upper arm edges are included in the same edge set, whereas when the CoM node is chest, both thigh edges and both forearm edges are included in the same edge set. The details of our six-way ensemble are specified on our Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 173, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 173, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 173, + 526 + ], + "type": "text", + "content": "3.5. Network Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "As shown in Fig. 4, we adopt [24] as our baseline network architecture with a total of nine stacked GCN blocks. The numbers of output channels for the blocks are 64, 64, 64, 128, 128, 128, 256, 256, and 256. Each block contains a residual connection [9] and is divided into a spatial module, in which the GCN operation proceeds, and a temporal module, which includes the temporal convolutions. Our method uses the temporal module of [2], whose baseline module is [18, 28]. This module consists of four branch operations. Two are dilated temporal convolutions with kernel size five and dilation one and two, respectively. The remaining branch operations are pointwise convolution and max pooling with kernel size three. Our spatial module consists of an HD-Graph convolution operation and an A-HA module, as introduced in Sec. 3.2 and Sec. 3.3. After passing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "through all GCN layers with attention to the hierarchy-wise features, the network compresses the feature map through the global average pooling layer and classifies the action sample through the softmax function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 132, + 388, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 132, + 388, + 145 + ], + "spans": [ + { + "bbox": [ + 305, + 132, + 388, + 145 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 152, + 495, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 495, + 165 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 495, + 165 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 170, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 277 + ], + "type": "text", + "content": "NTU-RGB+D 60. NTU-RGB+D 60 [22] is a large dataset used in skeletal action recognition. It contains 56,880 skeleton action samples, performed by 40 different participants and classified into 60 classes. The authors of this dataset recommend two benchmarks. (1) Cross-Subject (X-Sub): 20 of the 40 subjects' actions are used for training, and the remaining 20 are for validation. (2) Cross-View (X-View): Two of the three camera-views are used for training, and the other one is used for validation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "type": "text", + "content": "NTU-RGB+D 120. NTU-RGB+D 120 [16] is a dataset in which 57,367 new action samples are added to the NTU-RGB+D 60 dataset. It contains a total of 114,480 skeleton action samples over 120 classes, performed by 106 different subjects. The authors of this dataset recommend two benchmarks: (1) Cross-Subject (X-Sub): 53 of the 106 subjects' actions are used for training, and the remaining 53 are used for validation. (2) Cross-Setup (X-Set): Of the 32 setups, data with even setup IDs are used for training, and the remaining data with odd IDs are used for validation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": "Kinetics-Skeleton. The Kinetics-Skeleton dataset is derived from the Kinetics 400 video dataset [11], utilizing the OpenPose pose estimation [1] to extract 240,436 training and 19,796 testing skeleton sequences across 400 classes. The dataset restricts the number of skeletons per time step to two and eliminates skeletons with lower confidence scores, ensuring high-quality sequences for human action recognition and pose estimation research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 535, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 545, + 619 + ], + "type": "text", + "content": "Northwestern-UCLA. The Northwestern-UCLA skeleton dataset [30] contains 1494 video clips over 10 classes. Each action is captured through three Kinect cameras with different camera views and is performed by 10 subjects. We adopt the same protocol as NW-UCLA: Two of the three camera-views are used for training, and the other one is used for validation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Experimental Settings. In our experiments, we adopt [24] as the backbone. The SGD optimizer is employed with a Nesterov momentum of 0.9 and a weight decay of 0.0004. The number of learning epochs is set to 90, with a warm-up strategy [9] applied to the first five epochs for more stable learning. We set the learning rate to decay with cosine annealing [19], with a maximum" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10449" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 72, + 545, + 264 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 264 + ], + "type": "table", + "html": "
MethodsPublicationMotion StreamNTU-RGB+D 60NTU-RGB+D 120Kinetics-SkeletonNorthwestern
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)Top-1 (%)Top-5 (%)UCLA (%)
ST-GCN [35]AAAI 2018✘81.588.370.773.230.752.8-
2s-AGCN [24]CVPR 2019✘88.595.182.584.236.158.7-
SGN [36]CVPR 2020✓89.094.579.281.5--92.5
AGC-LSTM [26]CVPR 2019✘89.295.0----93.3
DGNN [23]CVPR 2019✓89.996.1--36.959.6-
Shift-GCN [5]CVPR 2020✓90.796.585.987.6--94.6
DC-GCN+ADG [4]ECCV 2020✓90.896.686.588.1--95.3
DDGCN [14]ECCV 2020✓91.197.1--38.160.8-
MS-G3D [18]CVPR 2020✘91.596.286.988.438.060.9-
MST-GCN [3]AAAI 2021✓91.596.687.588.838.160.8-
CTR-GCN [2]ICCV 2021✓92.496.888.990.6--96.5
EfficientGCN-B4 [27]TPAMI 2022✓91.795.788.389.1---
STF [12]AAAI 2022✘92.596.988.989.939.9--
InfoGCN (4-ensemble) [6]CVPR 2022✓92.796.989.490.7--96.6
InfoGCN (6-ensemble) [6]CVPR 2022✓93.097.189.891.2--97.0
HD-GCN (2-ensemble)✘92.496.689.190.638.961.796.6
HD-GCN (4-ensemble)✘93.097.089.891.240.363.096.9
HD-GCN (6-ensemble)✘93.497.290.191.640.963.597.2
", + "image_path": "b8ada711316ec3e0ab452edc2763a14e1ed2dcf43ebdc56756aafb452c90fceb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 306, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 437 + ], + "type": "text", + "content": "learning rate of 0.1 and a minimum learning rate of 0.0001. For the NTU-RGB+D datasets, we set the batch size to 64 and use the data preprocessing method from [36]. For Kinetics-Skeleton, the batch size is set to 128. In addition, to overcome the absence of belly and hip nodes in the Kinetics-Skeleton, we define the center of both hip joints as CoM hip node, and the center of chest and the hip node as CoM belly node, resulting in a total of 20 nodes. For the Northwestern-UCLA dataset, we set the batch size to 16 and use the data preprocessing method from [5]. All our experiments are conducted on a single RTX 3090 GPU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 452, + 276, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 276, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 276, + 464 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Arts Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": "Most recent state-of-the-art networks [25, 5, 4, 2] adopt a four-way ensemble method, but we adopt the six-way ensemble method described in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 511, + 288, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 288, + 631 + ], + "type": "text", + "content": "We compare ours with state-of-the-art networks on three datasets: NTU-RGB+D 60 [22], NTU-RGB+D 120 [16], Northwestern-UCLA [30], and Kinetics-Skeleton [11]. Comparisons for each dataset are shown in Tab. 1. The recognition performance of our HD-GCN has exceeded the state-of-the-arts on every dataset without any motion streams, as shown in Tab. 1. With our proposed ensemble method, HD-GCN outperforms the state-of-the-art and shows comparable performance to the 6-way ensemble state-of-the-art using only 4-way ensemble method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "In this section, we demonstrate the effectiveness of the proposed HD-GCN. Performance is specified as the cross-subject and cross-setup classification accuracy on the NTU-RGB+D 120 [16] joint stream data." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 304, + 545, + 380 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 546, + 289 + ], + "type": "text", + "content": "Table 1. Comparisons of the top-1 accuracy (%) against state-of-the-art methods on the NTU-RGB+D 60, 120, Northwestern-UCLA, and Kinetics-Skeleton datasets. The orange and yellow cells respectively indicate the highest and second-highest value." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 304, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 304, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 545, + 380 + ], + "type": "table", + "html": "
Graph typeEdgesS-EdgeConvX-Sub (%)X-Set (%)
ConventionalPC✘83.585.4
HD-Graph
APC✘84.3 (↑ 0.8)86.1 (↑ 0.7)
BPC✓84.6 (↑ 1.1)86.3 (↑ 0.9)
CFC✘84.9 (↑ 1.4)86.5 (↑ 1.1)
DFC✓85.1 (↑ 1.6)86.7 (↑ 1.3)
", + "image_path": "a54ef3dfa5e56387f1f188a12bac37070bb94294407936e72b966eda62dc94b8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 406, + 545, + 490 + ], + "blocks": [ + { + "bbox": [ + 306, + 380, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 545, + 403 + ], + "type": "text", + "content": "Table 2. Comparison of the conventional graph and four types of HD-Graph." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 406, + 545, + 490 + ], + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 490 + ], + "type": "table", + "html": "
MethodH-EdgeConvX-Sub (%)X-Set (%)
Baseline✘83.585.4
HD-GCN
w/o A-HA✘85.1 (↑ 1.6)86.7 (↑ 1.3)
w/ SAP✘85.2 (↑ 1.7)86.7 (↑ 1.3)
w/ SAP✓85.4 (↑ 1.9)87.0 (↑ 1.6)
w/ RSAP✘85.5 (↑ 2.0)87.0 (↑ 1.6)
w/ RSAP✓85.7 (↑ 2.2)87.3 (↑ 1.9)
", + "image_path": "2bdbc49ba7dd6e1f236e3b7613cd6ed00e37188a036bb6d249b3f0218fbc7a3d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 493, + 545, + 516 + ], + "lines": [ + { + "bbox": [ + 306, + 493, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 493, + 545, + 516 + ], + "type": "text", + "content": "Table 3. Comparison of various types of attention modules. SAP denotes the spatial average pooling." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "content": "Hierarchically Decomposed Graph. To proceed with the ablation study for HD-Graph, we set Yan et al. [35]'s graph as the conventional graph. Here, we use the temporal convolution module of [18], as mentioned in Sec. 3.5, to compare the performance of networks fairly with various graphs. The experimental results are shown in 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "We set the edges of the HD-Graph in different ways to show a gradual performance increase according to the type of graph. There are four main versions of HD-Graph, the first of which is graph A containing only the PC edges. Unlike the conventional graph with one edge set including three fixed subsets, HD-Graph has a flexible number of edge sets, each divided by hierarchy layers with three subsets. Graph B is an extension of A, with the additional operation S-EdgeConv. Graph C contains FC edges for " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "N_{H}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " hierarchy node sets, and graph D is similar to C but includes" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10450" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 284, + 270 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 284, + 270 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 284, + 270 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 284, + 270 + ], + "type": "image", + "image_path": "fb9b86c77d6e50d56a0b1c929499240e687be1cbcdf22485b7ade65a62477c99.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 287, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 310 + ], + "type": "text", + "content": "Figure 5. Hierarchy-wise attention scores by A-HA for \"Running\" and \"Kicking\" class. Each score indicates the value of the attention map M of the A-HA module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 389 + ], + "type": "text", + "content": "S-EdgeConv. The HD-Graph with only PC edges performs better than the conventional graph by a large margin, even though they share the same edges. This proves that it is meaningful to divide the joint nodes by hierarchy edge sets. In addition, the HD-Graph with FC edges and S-EdgeConv performs better for every datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 402, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 605 + ], + "type": "text", + "content": "Attention-Guided Hierarchy Aggregation. To prove the effectiveness of the A-HA module, we use a method to change or remove specific parts of our attention module, with the results shown in Tab. 3. Spatial average pooling (SAP) simply averages along the spatial axis without the representative node extraction process, which performs worse than our RSAP. The poorer performance is due to two factors: (1) scaling bias occurs because the number of nodes in each hierarchy node set is different, and (2) attention through SAP does not represent the corresponding hierarchy node set because it brings the average of the feature vectors of all nodes, not a specific node set. Furthermore, it performs better with H-EdgeConv, which recognizes each hierarchy edge set as a graph node. This proves that because the major edge sets are different for each data sample, it is important to find and highlight edge sets with high similarity based on the Euclidean distance through H-EdgeConv." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "The results of the attention score " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " of our A-HA module are shown in Fig. 5. These results show that our module scores edge sets 4 and 5 higher for the \"running\" class, which includes knees and feet, elbows and hands. For the \"Kicking\" class, A-HA gives the highest score to edge set 3, which includes shoulders and hips, followed by edge set 4 and 5. It is reasonable for human visual recognition that the dynamically moving edge set 4, 5 are more important than the stationary and barely moving edge set 3 when running" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 545, + 139 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 139 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 139 + ], + "type": "table", + "html": "
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
DC-GCN [4]84.0*86.1*2.743.45
MS-G3D [18]84.9*86.8*5.223.22
CTR-GCN [2]84.986.5*1.971.46
InfoGCN [6]85.186.31.681.57
HD-GCN85.787.31.601.68
", + "image_path": "48a55ca5f7de1576bdb2e84253913f033ee16b2578a8fb0bc0fa3383b8dfc0c0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 166, + 545, + 234 + ], + "blocks": [ + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "type": "text", + "content": "Table 4. Comparison of complexity of the single-stream state-of-the-arts. " + }, + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "type": "text", + "content": ": results obtained by the released codes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 166, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 307, + 166, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 307, + 166, + 545, + 234 + ], + "type": "table", + "html": "
X-Sub (%)X-Set (%)GFLOPs#Param. (M)
CTR-GCN † [2]88.990.67.885.84
InfoGCN † [6]89.490.76.726.28
InfoGCN ‡ [6]89.891.210.089.42
HD-GCN †89.891.26.406.72
HD-GCN ‡90.191.69.6010.08
", + "image_path": "f899662ea4e68f01c34adb652188266ebd819833fd32176bf1cb7596d29885a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 237, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 305, + 237, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 545, + 259 + ], + "type": "text", + "content": "Table 5. Comparison of complexity of the multi-stream state-of-the-arts. †: 4-ensemble, ‡: 6-ensemble" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 263, + 455, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 455, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 455, + 275 + ], + "type": "text", + "content": "rather than when kicking something." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "type": "text", + "content": "Six-Way Ensemble. We use the ensemble method to which three graphs with different CoM nodes are applied, excluding motion streams. Tab. 1 shows that the HD-GCN with 4-way ensemble outperforms the state-of-the-art [6] 4-way methods with motion data and shows comparable performance with the state-of-the-art 6-way method. In addition, when the 6-way ensemble with three different graphs is applied to HD-GCN, it outperforms the state-of-the-art methods. This proves that the features extracted with different CoM nodes are learned in different learning aspects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 422, + 545, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 435 + ], + "type": "text", + "content": "4.4. Comparison of Complexity with Other Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 441, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 573 + ], + "type": "text", + "content": "Although our model has multiple branch layers for multiple edge sets, it does not cause high complexity because it precedes channel reduction layers. Comparisons of computational complexity with other models are shown in Tab. 4, where the window size is fixed to 64. Our model shows the best performance on NTU-RGB+D 120 joint stream by a large margin even though the computational complexity of our model is the lowest. For multi-stream ensemble, our 4-stream HD-GCN shows almost similar performance to 6-stream InfoGCN [6] while having 3.68G fewer FLOPs and 2.70M fewer parameters as shown in Tab. 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 585, + 383, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 383, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 383, + 597 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose a novel hierarchically decomposed graph convolutional network (HD-GCN) for skeleton-based action recognition. We also propose a new framework (HD-Graph) that replaces the existing framework, decomposes all the joint nodes by hierarchy edge sets and considers the connectivity between major distant nodes, which is difficult to identify naturally. We also present an effective attention module (A-HA) composed of representative spatial average pooling (RSAP) layer and hierarchical" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10451" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "edge convolution (H-EdgeConv), which applies hierarchywise attention for the HD-Graph. In addition, our HD-GCN learns graph-wise features with different patterns through a six-way ensemble method. We derive an effective feature extractor by combining these three methods and empirically verify its effectiveness. Our approach outperforms current state-of-the-art methods on four benchmark datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 173, + 288, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 288, + 209 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by Alon-Flow Research and the KIST Institutional Program (Project No.2E32283-23-064)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 220, + 107, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 107, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 107, + 233 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 241, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 287, + 285 + ], + "type": "text", + "content": "[1] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 286, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 288, + 342 + ], + "type": "text", + "content": "[2] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359-13368, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 343, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 343, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 343, + 287, + 397 + ], + "type": "text", + "content": "[3] Zhan Chen, Sicheng Li, Bing Yang, Qinghan Li, and Hong Liu. Multi-scale spatial temporal graph convolutional network for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1113-1122, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "type": "text", + "content": "[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 455, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 455, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 455, + 288, + 510 + ], + "type": "text", + "content": "[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 512, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 577 + ], + "type": "text", + "content": "[6] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogen: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "type": "text", + "content": "[7] David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alán Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 667 + ], + "type": "text", + "content": "[8] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "text", + "content": "[11] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[12] Lipeng Ke, Kuan-Chuan Peng, and Siwei Lyu. Towards to-at spatio-temporal focus for skeleton-based action recognition. arXiv preprint arXiv:2202.02314, 2022. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "text", + "content": "[13] Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems. In International Conference on Machine Learning, pages 2688-2697. PMLR, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 282 + ], + "type": "text", + "content": "[14] Matthew Korban and Xin Li. Ddgcn: A dynamic directed graph convolutional network for action recognition. In European Conference on Computer Vision, pages 761-776. Springer, 2020. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 284, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 326 + ], + "type": "text", + "content": "[15] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, and Sangyoun Lee. Leveraging spatio-temporal dependency for skeleton-based action recognition. arXiv preprint arXiv:2212.04761, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[16] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[17] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In European conference on computer vision, pages 816-833. Springer, 2016." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "text", + "content": "[18] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 143-152, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 513 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 580 + ], + "type": "text", + "content": "[20] Federico Monti, Davide Boscaini, Jonathan Masci, Emanuele Rodola, Jan Svoboda, and Michael M Bronstein. Geometric deep learning on graphs and manifolds using mixture model cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5115-5124, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[21] Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. Learning convolutional neural networks for graphs. In International conference on machine learning, pages 2014-2023. PMLR, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "text", + "content": "[22] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu " + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathrm{rgb + }" + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "text", + "content": " d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1010-1019, 2016. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[23] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10452" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[24] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "text", + "content": "[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. IEEE Transactions on Image Processing, 29:9532-9545, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 249 + ], + "type": "text", + "content": "[26] Chenyang Si, Wentao Chen, Wei Wang, Liang Wang, and Tieniu Tan. An attention enhanced graph convolutional LSTM network for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1227-1236, 2019. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "text", + "content": "[27] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 358 + ], + "type": "text", + "content": "[28] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[29] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "text", + "content": "[30] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning and recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2649-2656, 2014. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[31] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[32] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[33] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1-12, 2019. 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "text", + "content": "[34] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[35] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[36] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "type": "text", + "content": "networks for efficient skeleton-based human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1112-1121, 2020. 1, 7" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10453" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_content_list.json b/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4ecfa995048aeb7efc39c27316073b7ce32cbc56 --- /dev/null +++ b/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_content_list.json @@ -0,0 +1,1588 @@ +[ + { + "type": "text", + "text": "High Quality Entity Segmentation", + "text_level": 1, + "bbox": [ + 310, + 132, + 656, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lu Qi $^{1*}$ , Jason Kuen $^{2*}$ , Tiancheng Shen $^{4*}$ , Jiumiang Gu $^{2}$ , Wenbo Li $^{4}$ , Weidong Guo $^{3\\dagger}$ , Jiaya Jia $^{4}$ , Zhe Lin $^{2}$ , Ming-Hsuan Yang $^{1}$ , $^{1}$ The University of California, Merced $^{2}$ Adobe Research $^{3}$ QQ Brower Lab, Tencent, $^{4}$ The Chinese University of Hong Kong", + "bbox": [ + 178, + 181, + 776, + 262 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abff55552ee80cb243784d84900bcdac555cd2beafbed9552b936cd0a4ce0506.jpg", + "image_caption": [ + "Figure 1: Segmentation results on in-the-wild test images: (1) Mask2Former [7] trained for panoptic segmentation [20] on COCO dataset [31]; (2) Mask2Former trained for entity segmentation [41] on COCO dataset [31]; (3) our CropFormer trained on the proposed EntitySeg Dataset. For a fair comparison, all three models use Swin-Large backbone [33] and are trained until full convergence. Our approach provides far more desirable and useful results for many real-world applications." + ], + "image_footnote": [], + "bbox": [ + 88, + 330, + 883, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 235, + 670, + 313, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense image segmentation tasks (e.g., semantic, panoptic) are useful for image editing, but existing methods can hardly generalize well in an in-the-wild setting where there are unrestricted image domains, classes, and image resolution & quality variations. Motivated by these observations, we construct a new entity segmentation dataset, with a strong focus on high-quality dense segmentation in the wild. The dataset contains images spanning diverse image domains and entities, along with plentiful high-resolution images and high-quality mask annotations for training and testing. Given the high-quality and -resolution nature of the dataset, we propose CropFormer which is designed", + "bbox": [ + 76, + 695, + 472, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to tackle the intractability of instance-level segmentation on high-resolution images. It improves mask prediction by fusing high-res image crops that provides more fine-grained image details and the full image. CropFormer is the first query-based Transformer architecture that can effectively fuse mask predictions from multiple image views, by learning queries that effectively associate the same entities across the full image and its crop. With CropFormer, we achieve a significant AP gain of 1.9 on the challenging entity segmentation task. Furthermore, CropFormer consistently improves the accuracy of traditional segmentation tasks and datasets. The dataset and code are released at http://luqi.info-entityv2.github.io/.", + "bbox": [ + 496, + 672, + 890, + 868 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † indicates corresponding author.", + "bbox": [ + 96, + 886, + 382, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 90, + 209, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Entity segmentation [41] is an emerging segmentation task that focuses on open-world and class-agnostic dense image segmentation, designed to have superior generalization capabilities for unseen-category segmentation. As such, it has great potential in image editing applications, where users can interact with visual entities that exist \"in the wild\". Recent works have leveraged entity segmentation for image manipulation [51, 54] and image retouching [34], where users usually specify in-the-wild text queries (e.g., Oculus, platypus) for local editing. In comparison to traditional closed-set category-based segmentation tasks such as panoptic/instance/semantic segmentation, entity segmentation has the advantage of being all-inclusive and unrestricted, making it easier to find masks that semantically match well with user-specified queries.", + "bbox": [ + 76, + 116, + 472, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous work on entity segmentation [41] has shown promising results, yet it only employed existing panoptic segmentation datasets to verify the feasibility and benefits of entity segmentation. However, existing panoptic segmentation datasets (e.g., COCO [31], Cityscapes [9], ADE20K [65]) suffer from the same issues of closed-set categories that entity segmentation [41] aims to avoid. These datasets are collected and annotated for their original panoptic segmentation task requirements. In view of the lack of a proper entity segmentation dataset, it is difficult to further study and explore on this important task. To address this issue, we present EntitySeg dataset, a new large-scale and high-quality dataset for entity segmentation.", + "bbox": [ + 76, + 343, + 472, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Apart from in-the-wild generalization, we design EntitySeg to prioritize high-quality mask annotations and high-resolution images, which are essential for user-driven image editing in the current era of high-resolution imagery. Image editing workflows usually require pixel-precise segmentation masks to ensure a smooth user experience. However, existing panoptic segmentation datasets generally contain low-resolution images (e.g., COCO[31] has zero images that are at least 1K) and coarsely-annotated masks, making it difficult for models trained on such datasets to produce accurate mask predictions for higher resolution images. This is why we focus on high-resolution images and high-quality masks in the EntitySeg dataset, which clearly distinguishes it from existing datasets.", + "bbox": [ + 76, + 537, + 472, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The EntitySeg dataset poses new challenges to existing segmentation methods due to its high-resolution images and high-quality mask annotations. Standard practices of resizing images and mask annotations to 800px used in existing methods [43, 15, 7], result in a heavy loss of fine-grained details. On the other hand, training or evaluating existing segmentation models directly on our high-resolution/quality dataset is intractable due to excessive computation and memory overheads. A naive workaround is to divide the high-resolution image into smaller patches and apply seg", + "bbox": [ + 76, + 750, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mentation model to them separately, but it is unable to exploit global context to resolve ambiguity. Moreover, it is unclear how to combine mask predictions from multiple patches using state-of-the-art Transformer-based segmentation methods which rely on set-prediction [7, 6, 50, 61], as shown in Fig 4 and our supplementary material.", + "bbox": [ + 496, + 93, + 890, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce CropFormer, a novel approach for tackling the problem of high-resolution entity segmentation. CropFormer learns to dynamically generate tokens to associate the full image with its high-resolution crops and then fuses them to generate the final mask prediction. During training, one of the predefined four corner crops is randomly selected and fused with the full image. For inference, CropFormer is applied to all four corner crops and the full image to achieve complete fusion. The strength of CropFormer lies in its ability to dynamically generate queries to link the same entities in the full image and its higher-resolution crops, allowing us to benefit from both global image context and high-quality image/annotation details to obtain improved results.", + "bbox": [ + 496, + 184, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are as follows:", + "bbox": [ + 517, + 396, + 854, + 410 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a large-scale, high-quality entity segmentation dataset, comprising of images collected from various domains with diverse resolutions and high-quality masks. Our dataset allows for the evaluation of the generalization and robustness of segmentation models in a unified and open-world manner, due to its domain-diverse and high-resolution setting.", + "- We benchmark existing segmentation models on our newly collected dataset. We find that the traditional training setting cannot accommodate our dataset, which comprises over $70\\%$ images with very high resolution.", + "- We propose a novel model, CropFormer, with an innovative association module and batch-level decoder to enable the fusion of the full image and its higher-resolution crops.", + "- Extensive experiments and thorough analysis demonstrate the efficacy of the proposed model and dataset on various segmentation tasks and settings. In Fig. 1, we compare the results from our best CropFormer (trained on EntitySeg) with the best Mask2Formers [7] trained on COCO-Panoptic [20] and COCO-Entity [41]." + ], + "bbox": [ + 517, + 412, + 888, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 758, + 640, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Segmentation Dataset. Numerous datasets have been proposed for semantic, instance, and panoptic segmentation, e.g., Microsoft COCO [31], ADE20K [65], KITTI [12], LVIS [14], PASCAL-VOC [11], OpenImages [23], Cityscapes [9], and so on. In addition, there are some datasets specially designed for specific scenarios, such as amodal segmentation (COCO-Amodal [66], KINS [39]), human segmentation (CelebAMask-HQ [25],", + "bbox": [ + 496, + 777, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4048", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c9e86e1de4e05392c10c8d13360e02ad841f4e13b68c90ee0e5fcc7f46f97de.jpg", + "image_caption": [ + "COCO (height: 427px, width: 640px)" + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 500, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8dc526205554609af5c87d1a00aac115bc66a9dfc8b0a9333b88a52d955f3978.jpg", + "image_caption": [ + "ADE20K (height: 512px, width: 683px)" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 866, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ae77d0b6ad3698bfbceb1084fda15d79b5bcd5e798c715c50d8e3942424a7960.jpg", + "image_caption": [ + "Unsplash (height: 4220px, width: 3176px)" + ], + "image_footnote": [], + "bbox": [ + 101, + 184, + 401, + 328 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d0c275b075b8c8857a13a25a089c0735af4e9974840c38205ef33243b160da1d.jpg", + "image_caption": [ + "LAION-400M (height: 3637px, width: 2909px)", + "Figure 2: High-quality mask annotations for low- and high-resolution images collected from existing datasets such as COCO [31], ADE20K [65] and Cityscapes [9] as well as from the internet are presented. For the images collected from the aforementioned datasets, a visual comparison between the original and our annotations is provided in the middle and rightmost sub-figures, where the unannotated regions are shaded in black. It is worth noting that the RGB and mask images shown here have been downsampled, resulting in some quality degradation compared to the original datasets." + ], + "image_footnote": [], + "bbox": [ + 403, + 185, + 702, + 330 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5184b01d2a1999eae0caab85da41c9397422fe810de2c8ab1214e4ce0b6fea3a.jpg", + "image_caption": [ + "Clipart (340, 339)" + ], + "image_footnote": [], + "bbox": [ + 704, + 185, + 866, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LIP [13], MHP [63]) and domain adaptation (Synscapes [57], GTA5 [44]). Despite the significant contributions from these datasets, there is still a need to fulfill the requirements of real-world applications with high-quality images of large diversity. For example, a segmentation model should be robust to high-resolution images from different domains and able to segment unseen objects in the openworld setting. Most relevant to our work is the ADE20K dataset [65], which has large-scale open-vocabulary categories. However, the collected images are of low resolution and from narrowed domains. In contrast, our dataset collects images from multiple domains, including indoor, outdoor, street scenes, cartoon, and remote images, with over $80\\%$ of the resolution of the images falling within the high-resolution range of 2000px to 8000px. Differently from the ADE20K [64] and LVIS [14] dataset with a predefined category list, our annotation process first conducts class-agnostic mask annotation on each entity and then labels the category information in an open-vocabulary manner.", + "bbox": [ + 76, + 436, + 472, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Parsing Methods. Scene parsing methods mainly rely on convolution-based dense prediction or transformer-based query prediction. Explicit localization information, such as bounding box proposals or pixel position, is often used for pixel-level mask prediction in these approaches. Examples include FCN [35], DeepLab [5], PSPNet [62], Mask R-CNN [15], PANet [32], SOLO [53], CondInst [48], PanopticFPN [19], and PanopticFCN [28]. Inspired by the success of transformer-based detection tasks such as DETR [4] [29, 30, 10, 40], a number of methods, such as Max-Deeplab [50] and Mask2Former [7], have been pro", + "bbox": [ + 78, + 733, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "posed to directly predict segmentation masks without relying on bounding boxes. The key to these methods is to use learnable queries trained by a set-prediction loss to model the implicit location of the potential instance area. However, query-based set-prediction methods cannot effectively perform multi-view fusion, as the query-to-object assignment can drastically change with the change of image view. This is because the set prediction loss typically adopted by such methods does not enforce any assignment constraints across multiple image views.", + "bbox": [ + 496, + 436, + 890, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. EntitySeg Dataset", + "text_level": 1, + "bbox": [ + 500, + 599, + 671, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The EntitySeg dataset contains 33,227 images with high-quality mask annotations, as illustrated in Fig. 2. Compared with existing datasets, it has three distinctive features. Firstly, $71.25\\%$ and $86.23\\%$ of the images are of high quality, with at least $2000\\mathrm{px} \\times 2000\\mathrm{px}$ and $1000\\mathrm{px} \\times 1000\\mathrm{px}$ , respectively, staying in line with the current digital imaging trends. Secondly, our dataset is open-world and not restricted to predefined classes. We consider each semantically-coherent region in the images as an entity, even if it is blurred or hard to recognize semantically. Thirdly, the mask annotation along the boundaries is more precise than those in other datasets. In the following, we will provide an in-depth description of our EntitySeg dataset, as well as an extensive analysis.", + "bbox": [ + 496, + 625, + 890, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image Collection", + "text_level": 1, + "bbox": [ + 500, + 838, + 666, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the collection criterion of COCO [31], we collected most non-iconic images with plentiful contextual information and non-canonical viewpoints. To en", + "bbox": [ + 496, + 853, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4049", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c073d643f54a4396a8035f1509c7b87c6fadc96b1c183a21fd434442315a4a4.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 336, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3f66fd9c6b5940b2013e2191548c1b87a29630c0f8f892aa0b95c0fadc546d0.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 339, + 92, + 580, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/514c4f458271ae9adc8bea39c7abc348a45279afd0f362876ab476853c4eb968.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 581, + 93, + 862, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sure substantial domain diversity, we have sourced the images from various public datasets and the Internet, with permission for academic research. Specifically, part of the images are from COCO [31], ADE20K [65], Pascal VOC [11], Cityscapes [9], Mapillary Vistas [38], ImageNet [22], Open Images [23], DIV2K [3], Flickr2K [55], Clipart [17], Comic [17], DOTA [59], Synscapes [57] and GTA5 [44]. Moreover, we have crawled high-resolution images from Pixabay [1], Unsplash [2], and LAION-400M [46]. As shown in Fig. 3(c), most of the images are from high-resolution sources like Pixabay, Unsplash, and LAION-400M. Moreover, Fig. 3(a) reveals that our dataset has more high-resolution images, with resolutions normally distributed within the range of $0\\mathrm{px}$ to $15,000\\mathrm{px}$ , in comparison to ADE20K and COCO with images all under $1000\\mathrm{px}$ .", + "bbox": [ + 76, + 313, + 473, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Annotation", + "text_level": 1, + "bbox": [ + 78, + 542, + 253, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For our dataset, the annotation process consists of the following steps. Firstly, entities in the image are annotated with class-agnostic pixel-level masks, ensuring that no overlaps exist between masks. Subsequently, each mask is annotated with a class label in an open vocabulary manner. Two special considerations should be taken into account here: 1) we aonotate the entity as 'unknown' or 'blurred' if it cannot be easily named or is severely blurred; 2) the entity is annotated as a 'supercategory_other' if we only know it is a supercategory but unable to determine its fine-grained class. Our annotation process is the opposite of those of the popular segmentation datasets like COCO [31] and ADE20K [64], which first predefine their class sets before annotating masks based on the predefined classes.", + "bbox": [ + 76, + 561, + 468, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2 indicates that our dataset has a more considerable proportion of covered image areas and more masks on average compared to COCO- and ADE20K-Panoptic. This is attributed to our novel annotation process that enables us to take all semantically-coherent regions into account and assign class labels accordingly. Besides, our annotation procedure is more similar to the human visual system. As", + "bbox": [ + 76, + 772, + 470, + 878 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/335fa5049313402eea53b520bbfaf7215329d038a2ddc769011a7638ccffb2a4.jpg", + "image_caption": [ + "Figure 3: Sub-figures (a) and (b) present the distributions of image resolutions and average number of entities among ADE20K, COCO, and EntitySeg, respectively. Moreover, sub-figure (c) displays the distribution of image sources from which the EntitySeg images were collected.", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 519, + 309, + 653, + 352 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a6fba9d53e44bbb944da2b26f2fca4878fc5e44672c760ac7fa89fc7368e879.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 730, + 310, + 862, + 352 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bf6b6fd1ebf66a923b65d8dfefd2c3b06be3eae779f62e70a4ed6d0776b5c28d.jpg", + "table_caption": [ + "Table 1: Annotation consistency of class-agnostic localization and class-aware categories among annotators." + ], + "table_footnote": [], + "table_body": "
DatasetImageRes (avg)↑EntityNum (avg)↑EntityNum (max)↑Entity Complexity↓Entity Simplicity↓Valid Area↑
COCO522.511.2950.7580.5810.891
ADE20K461.313.62550.8020.6060.914
EntitySeg2700.718.12360.7190.5380.999
", + "bbox": [ + 501, + 398, + 901, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: The statistical comparisons between COCO [31], ADE20K [65] and EntitySeg. 'ImageRes (avg)', 'EntityNum (avg)', 'Valid Area', 'Entity Complexity', and 'Entity Simplicity' refer to the average value of image resolution size, entity numbers per image, valid area ratio per image, average entity complexity and simplicity respectively. 'EntityNum (max)' means the maximum per-image number of entities within each dataset.", + "bbox": [ + 498, + 450, + 890, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "evidenced in [36], the human vision system is intrinsically class-agnostic and can recognize entities without comprehending their usage and purpose.", + "bbox": [ + 498, + 580, + 888, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotation consistency is crucial to any human-labeled dataset since it implies whether the annotation task is well-defined and proper. To study this, we randomly selected 500 images from EntitySeg (1.5% of the entire dataset), and asked another two annotators to annotate them after four months since the first round. Table 1(a) shows that the class-agnostic mask mAP in the first step of the annotation process compared to those of the other two annotators. We use one as ground truth and another one as prediction results. In Table 1(b), we also evaluate the category consistency by accuracy (ACC) under the same mask annotations. These two tables demonstrate that our EntitySeg has high annotation consistency in the mask and category labeling stages.", + "bbox": [ + 496, + 627, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Statistics", + "text_level": 1, + "bbox": [ + 500, + 833, + 666, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our EntitySeg dataset, we annotate all images at the entity level, regardless of whether they belong to thing or stuff in previous datasets. As shown in Fig. 1, we show", + "bbox": [ + 498, + 853, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We removed not-safe-for-work images via strict CLIP filtering.", + "bbox": [ + 96, + 883, + 434, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4050", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "some low- and high-resolution photos and their mask annotations to highlight our high-quality pixel-level annotation quality. For example, one can zoom in on the Unsplash image to see how high-quality the railing's mask is.", + "bbox": [ + 76, + 93, + 468, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 shows the quantitative comparison among our dataset, COCO-Panoptic [31], and ADE20K-Panoptic [65]. In the EntitySeg dataset, each image has 18.1 entities on average, which is more than 11.2 and 13.6 entities in COCO and ADE20K. For a fair comparison, we treat each uncountable stuff region as one segment although it is disconnected. The detailed comparison among those three datasets on the distribution of entity number is shown in Fig. 3(b). Furthermore, the shapes of entities in our dataset are more complex than those COCO [31] and ADE20K [65] as represented by the columns 'Entity Complexity' and 'Entity Simplicity' of Table 2 where an entity with a large convexity and simplicity value means it is a simple shape (and both metrics achieve their maximum value of 1.0 for a circle [66]). More details about how complexity and simplicity are calculated can be found in the supplementary material.", + "bbox": [ + 76, + 154, + 470, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Class-Aware Annotation. We select a subset of 11,580 images from the dataset for class labeling, for investigating the diversity of semantic classes in our dataset. In our class annotation process, the class is labeled in an open-vocabulary and free-form manner, then the class set is continually expanded. After the class annotation process, we organize and merge the annotated classes according to Word-Tree [45]. This results in 535 things and 109 stuff classes for those 11,580 images that we use to construct the class-aware EntityClass dataset. Please refer to our supplementary file for more analysis on EntityClass dataset.", + "bbox": [ + 76, + 398, + 468, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. CropFormer", + "text_level": 1, + "bbox": [ + 78, + 579, + 210, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose CropFormer which is the first Transformer-based set-prediction segmentation method that is capable of fusing the mask predictions from multiple image views. Inspired by existing methods [8, 7, 50, 61], we design CropFormer to learn $N$ number of $K$ -dimensional queries $\\mathbf{Q} \\in \\mathbb{R}^{N \\times K}$ to generate mask embeddings $\\mathbf{E} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}$ . $\\mathbf{E}$ then acts as convolution filters that are applied to pixel-level mask features $\\mathbf{P}_2 \\in \\mathbb{R}^{T \\times H \\times W}$ to generate $N$ segmentation masks $\\mathbf{U}^{\\mathbf{m}} \\in \\mathbb{R}^{N \\times T \\times H \\times W}$ , where $T, H$ and $W$ correspond to the image view, height, and width dimensions. However, such a rudimentary design can only work well with single-view inputs and cannot effectively fuse the results from multiple views. Here, we introduce a novel association module and batch-level decoder to achieve the goal of CropFormer, which is to exploit global context from full image and fine-grained local details from crops for high-quality segmentation. CropFormer is illustrated in Fig. 4.", + "bbox": [ + 76, + 604, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Crop Dataloger. In CropFormer framework, the crop dataloger is designed to generate a batch of images that si", + "bbox": [ + 76, + 868, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "multaneously include the full images and their corresponding crops. There are two steps in our dataloader: crop and resize. First, we augment the full image $I^o$ with a crop that is part of the full image. The crop is randomly extracted from one of the fixed image corners: upper-left, upper-right, bottom-left, bottom-right. The crop size is controlled by a fixed ratio hyperparameter $\\delta \\in \\mathbb{R}$ relative to the full image size. Then, we resize both the full image and crop to the same size. In this way, we construct an input tensor $\\mathbf{I} \\in \\mathbb{R}^{1 \\times 2 \\times H' \\times W' \\times 3}$ with 2 elements (full image & 1 crop) along the view dimension. Compared to the full image, our corner crops preserve more image details (e.g., local information) that are useful for high-quality segmentation. Specifically, given an dataloader $\\mathcal{D}$ , we define the data preprocessing as:", + "bbox": [ + 496, + 93, + 890, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = \\left\\{I ^ {o}, I ^ {c} \\right\\}, \\quad I ^ {o} \\sim \\mathcal {D} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 330, + 888, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I^o$ is sampled from $\\mathcal{D}$ , and $I^c$ is the cropped image controlled by fixed ratio $\\delta$ . Unlike the random cropping in [58, 16], we adopt a fixed cropping solution to keep the training and inference consistent. Such a rigid cropping strategy provides better inference efficiency since, rather than bunches of random crops, evaluating a fixed number of crops during inference is sufficient enough.", + "bbox": [ + 496, + 361, + 890, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Encoder and Decoder. Based on learnable image-level queries $\\mathbf{Q}_i\\in \\mathbb{R}^{N\\times K}$ , we use image encoder $\\Theta$ and image-level decoder $\\Phi_{i}$ to generate image-level embeddings $\\mathbf{E_i}\\in \\mathbb{R}^{N\\times 2\\times 1\\times 1\\times K}$ for the full input image and its crop. Given the input tensor (I) and queries $(\\mathbf{Q}_i)$ , we define $\\mathbf{E_i}$ as follows:", + "bbox": [ + 496, + 474, + 888, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {\\mathbf {i}} = \\Phi_ {i} \\left(\\mathbf {Q} _ {\\mathbf {i}}, \\Theta (\\mathbf {I})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 566, + 888, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_i(\\cdot)$ is a Transformer-based image-level decoder. $\\mathbf{E_i}$ is used for image-level entityness prediction and pixellevel prediction using the low-level image feature $\\mathbf{P_2}$ (derived from the image encoder). The prediction process can be formulated as:", + "bbox": [ + 496, + 592, + 888, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {\\mathbf {i}} ^ {\\mathrm {e}}, \\mathbf {U} _ {\\mathbf {i}} ^ {\\mathrm {m}} = \\operatorname {P r e d H e a d} _ {\\mathbf {i}} \\left(\\mathbf {E} _ {\\mathbf {i}}, \\mathbf {P} _ {\\mathbf {2}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 680, + 888, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{U}_{\\mathrm{i}}^{\\mathrm{e}}$ and $\\mathbf{U}_{\\mathrm{i}}^{\\mathrm{m}}$ denote the entityness prediction and pixel-level mask outputs. Here, we employ $i$ subscript to differentiate image-level embeddings and mask outputs from those of the association module.", + "bbox": [ + 496, + 709, + 888, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Association Module. Following [6], our association module aims to generate batch queries $\\mathbf{Q}_{\\mathrm{b}}$ that are fully shared by the full image and its crop to represent the same entities consistently. However, instead of treating $\\mathbf{Q}_{\\mathrm{b}}$ as learnable network parameters, considering that $\\mathbf{E}_{\\mathrm{i}}$ already contains strong segmentation features, we generate $\\mathbf{Q}_{\\mathrm{b}}$ directly from $\\mathbf{E}_{\\mathrm{i}} = \\{\\mathbf{E}_{I^o},\\mathbf{E}_{I^c}\\}$ . Particularly, we use a Transformer architecture with cross-attention ( $f_{\\mathrm{XAtt}}$ ) and self", + "bbox": [ + 496, + 777, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4051", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea2028887fe9c83e2cdf29b95dc6a8252d85cfa04898818ff8d0ec1fdcfb0b7b.jpg", + "image_caption": [ + "Figure 4: Framework of the proposed CropFormer. The red box indicates cropped region randomly sampled from four fixed image corners. In image-level prediction, the same entity across different image views may be assigned to different queries. Our association module and batch decoder can effectively associate the same entity across different views with a single query." + ], + "image_footnote": [], + "bbox": [ + 83, + 106, + 885, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "attention $(f_{\\mathrm{SAtt}})$ to obtain $\\mathbf{Q}_{\\mathbf{b}}$ ..", + "bbox": [ + 78, + 337, + 277, + 353 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\mathbf {b}} = \\operatorname {F F N} \\left(f _ {\\mathrm {S A t t}} \\left(f _ {\\mathrm {X A t t}} \\left(\\underbrace {f _ {\\mathrm {q}} \\left(\\mathbf {E} _ {I ^ {o}}\\right)} _ {\\text {q u e r y}}, \\underbrace {f _ {\\mathrm {k}} \\left(\\mathbf {E} _ {\\mathbf {i}}\\right)} _ {\\text {k e y}}, \\underbrace {f _ {\\mathrm {v}} \\left(\\mathbf {E} _ {\\mathbf {i}}\\right)} _ {\\text {v a l u e}}\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 359, + 468, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where FFN is a feed-forward network, $f_{\\{\\mathbf{q}, \\mathbf{k}, \\mathbf{v}\\}}(\\cdot)$ are linear transformations. Considering the importance of full image for entity segmentation, inspired by [49, 26] we take the image-level embeddings of full image $\\mathbf{E}_{I^o}$ as query, while treating all image-level embeddings $\\mathbf{E}_i$ as key and value.", + "bbox": [ + 78, + 402, + 468, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Batch-level Decoder. Given $\\mathbf{Q}_{\\mathbf{b}} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}$ , we obtain batch embeddings $\\mathbf{E}_{\\mathbf{b}} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}$ :", + "bbox": [ + 78, + 479, + 468, + 512 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {\\mathbf {b}} = \\Phi_ {b} (\\mathbf {Q} _ {\\mathbf {b}}, \\Theta (\\mathbf {I})), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 518, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Phi_b(\\cdot)$ denotes the batch-level decoder. Thus, we broadcast $\\mathbf{E}_{\\mathbf{b}}$ to the shape of $N\\times 2\\times 1\\times 1\\times K$ to share it between the full image and its crop. Finally, the broadcasted $\\mathbf{E}_{\\mathbf{b}}$ and low-level batch features $\\mathbf{P_2}$ are used for batch-level mask classification and pixel-wise prediction:", + "bbox": [ + 78, + 542, + 468, + 617 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {\\mathrm {b}} ^ {\\mathrm {e}}, \\mathbf {U} _ {\\mathrm {b}} ^ {\\mathrm {m}} = \\operatorname {P r e d H e a d} _ {\\mathrm {b}} \\left(\\mathbf {E} _ {\\mathrm {b}}, \\mathbf {P} _ {2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 625, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the image-level decoder, the queries are separate for each image/crop. Whereas, for the batch-level decoder, the queries are shared between the image and its crop. We illustrate the differences of the image- and batch-level decoders in Fig. 5, where we regard two mage views as an entirety.", + "bbox": [ + 78, + 648, + 468, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. During training, we employ two separate losses $\\mathcal{L}_i$ and $\\mathcal{L}_b$ for image- and batch-level predictions, with respect to ground truth $\\mathbf{G}$ . The main difference between the two losses lies in whether the same entities in full image and crop are tied to the same queries or not.", + "bbox": [ + 78, + 727, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall training loss for our model is defined as:", + "bbox": [ + 98, + 801, + 442, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\sum_ {\\mathbf {k} \\in \\{\\mathbf {i}, \\mathbf {b} \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {e}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {e}}\\right) + \\sum_ {\\mathbf {k} \\in \\{\\mathbf {i}, \\mathbf {b} \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {b c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {m}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {m}}\\right) + \\\\ \\sum_ {\\mathbf {k} \\in \\{i, b \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {d i c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {m}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {m}}\\right), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 823, + 468, + 895 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22fe4a5893d2414a865955c60b45a292367c6d822fb2e86860b5fe4745ab2811.jpg", + "image_caption": [ + "Figure 5: The illustration of image- and batch-level decoder and association module." + ], + "image_footnote": [], + "bbox": [ + 517, + 339, + 625, + 450 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25ddc2919141a5656a1bbc3e7648bb011049428a03e41ca3796e5ab54312478d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 339, + 750, + 450 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d293bd3770dd7d3a51207b77f08a3d5ffab0105e1d67e57ccbeea78403fef5d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 339, + 867, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{ce}}$ and $\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{ce}}$ denote binary cross-entropy loss for image- and batch-level entityness prediction. Similarly, $\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{bce}}$ , $\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{bce}}$ , $\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{dice}}$ and $\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{dice}}$ denote the binary cross-entropy and dice loss for image- and batch-level mask prediction.", + "bbox": [ + 500, + 497, + 888, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. In CropFormer, the same entities between the full image and its crop are represented by the same queries by virtue of the association module. For the final segmentation output, we fuse the per-pixel mask predictions obtained from the full image and every crop of the 4 corners by using an average operation. The confidence score of each entity comes from the batch-level entityness prediction score.", + "bbox": [ + 498, + 569, + 888, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 691, + 632, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct our main experiments on the entity segmentation with proposed EntitySeg Dataset in this work. First, we benchmark our dataset in several segmentation tasks, including entity, instance, semantic and panoptic segmentation. Then, we ablate our proposed CropFormer to demonstrate its effectiveness. Note that, in Sec. 4, we omit the conventional data batch dimension to improve clarity, but in practice, we train CropFormer with a batch of multiple dataset images and their crops, following existing methods.", + "bbox": [ + 498, + 718, + 888, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We leave out some details like the training schedule and some ablation studies due to paper length limitations (see supplementary for details).", + "bbox": [ + 498, + 854, + 888, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4052", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dc1812bcf7fc6e7bcc73e3e250217083ae523484aa06c92fe617a8fd0abb0fe1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMetricsMethodBackbonePretrainPerformance
SEMmIoUDeeplabv3R-50ImageNet27.9
Mask2FormerR-50ImageNet37.8
R-50COCO43.3
Swin-TCOCO45.0
Swin-LCOCO50.5
INSAPMask-RCNNR-50ImageNet5.0
Mask2FormerR-50ImageNet13.0
R-50COCO20.3
Swin-TCOCO22.7
Swin-LCOCO30.3
PANPQPanopticFPNR-50ImageNet3.6
Mask2FormerR-50ImageNet5.5
R-50COCO9.6
Swin-TCOCO9.8
Swin-LCOCO13.4
ENTAPeMask-RCNNR-50ImageNet24.9
Mask2FormerR-50ImageNet26.0
R-50COCO35.2
Swin-TCOCO42.9
Swin-LCOCO46.2
", + "bbox": [ + 81, + 89, + 475, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5840bc1499f4d4d472686a69955b18abc2e5825690386c4216bf23b5b3b0479a.jpg", + "table_caption": [ + "Table 3: Benchmark on several segmentation tasks on the EntitySeg Dataset. The 'SEM', 'INS', 'PAN', and 'ENT' indicate the task of semantic, instance, panoptic and entity segmentation. We note that we leave out all details and more experiments in the supplementary file due to paper length limitations." + ], + "table_footnote": [], + "table_body": "
Transfer DatasetTypeSEM (mIoU)INS (AP)PAN (PQ)ENT (APe)
ADE20KImage57.2(+1.0)36.0(+1.1)49.0(+0.9)40.6(+1.5)
YVIS2019 [60]Video-61.3(+0.9)--
VIPSeg [37]Video--37.0(+1.1)-
UVO [52]Video---29.5(+2.2)
", + "bbox": [ + 80, + 421, + 468, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Dataset Benchmarks", + "text_level": 1, + "bbox": [ + 78, + 560, + 274, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark on EntitySeg Dataset. In Table 3, we provide results on semantic, instance, panoptic, and entity segmentation on EntitySeg dataset with some baseline methods. Similar to existing segmentation datasets like COCO and ADE20K, the use of COCO pre-trained weight and a stronger backbone can benefit each segmentation performance of EntitySeg Dataset. Compared to COCO and ADE20K, EntitySeg dataset is more challenging as evidenced by the smaller performance numbers (e.g., Swin-L Mask2Former's PQ on COCO-PAN is $57.8\\%$ ).", + "bbox": [ + 76, + 580, + 468, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transferring EntitySeg to other general segmentation tasks. Table 4 shows the performance comparison of transferring models pretrained on either COCO or EntitySeg dataset to other image/video-level segmentation tasks. Owing to its emphasis on in-the-wild generalization, EntitySeg provides greater improvements than COCO dataset.", + "bbox": [ + 76, + 739, + 468, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefit of EntitySeg for high-quality image segmentation. In Table 5, we transfer Mask2Former pretrained on EntitySeg dataset for finetuning on Dichotomous Image Segmentation [42]. Compared to state-of-the-art ISNet [42]", + "bbox": [ + 76, + 838, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4afcaefe5b3ab1bee75a015442d86b867c3337fdfa1cdefb88f1f8edfcf0aa35.jpg", + "table_caption": [ + "Table 4: The relative performance gains of EntitySeg pretraining over COCO-Panoptic pretraining when fine-tuned on various image/video-level segmentation datasets." + ], + "table_footnote": [], + "table_body": "
MethodFβmx↑Fβw↑M↓Sα↑Eφm↑HCEγ↓
IS-Net [42].791.717.074.813.8561116
Ours.872.789.035.896.913969
", + "bbox": [ + 524, + 90, + 867, + 125 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7932118177984624e80060f10899fc4c36f5014152d9a86622f16bb1c19ec921.jpg", + "table_caption": [ + "Table 5: Evaluation on Dichotomous Image Segmentation (DIS5K dataset [42].) 'Ours' indicates Mask2Former pretrained on EntitySeg and then finetuned on DIS5K." + ], + "table_footnote": [], + "table_body": "
Train DatasetTaskLVIS [14]OCID [47]CAMO [24]FSS [27]Avg. AR
COCOPAN24.3839.9443.8064.9643.27
COCOENT27.7545.8859.7075.6052.23
SAM-H [21]Prompt49.2555.1775.3373.6063.34
EntitySegENT33.5676.6373.2086.2667.41
", + "bbox": [ + 504, + 196, + 885, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6: In-the-wild segmentation performance (AR@100 with IoU threshold 0.5). We directly evaluate models trained on either EntitySeg or COCO on large-vocabulary segmentation LVIS [14], object clutter indoor segmentation for robot grab [47], camouflaged instance segmentation CAMO [24], few-shot segmentation FSS [27]. All the models are based on Mask2Former with Swin-Large backbone.", + "bbox": [ + 496, + 261, + 890, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that is carefully designed for the task, Mask2Former pretrained on EntitySeg outperforms ISNet by a large margin, despite not having any task-specific design. This is by virtue of the high-quality mask annotations in EntitySeg dataset.", + "bbox": [ + 496, + 377, + 890, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation 'in the wild' Table 6 compares in-the-wild segmentation performance between Mask2Formers trained on either EntitySeg or COCO. Here, we directly use the trained models without finetuning on the evaluated datasets. The EntitySeg outperforms COCO-Panoptic [20] and COCO-Entity [41] by large margins.", + "bbox": [ + 496, + 446, + 890, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. A Comprehensive Analysis of CropFormer", + "text_level": 1, + "bbox": [ + 500, + 546, + 862, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first ablate Cropformer on the entity segmentation task and then report results on other segmentation datasets.", + "bbox": [ + 496, + 566, + 888, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7 shows the benefits of CropFormer on segmentation performance and computation cost compared to the Mask2Former [7], its multi-scale extension, as well as related methods [6, 16] that supports multi-view fusion. The first four rows are our baseline Mask2Former with the single-scale inference (800, 1040, or 2700 for shortest side) and (1333, 1732 or 4500 for longest side) for the full images, where $1040 = 800 \\times \\delta$ and $1732 = 1333 \\times \\delta$ with $\\delta = 0.7$ by default. For the results in the fifth row, we use test-time Hungarian matching to associate the same entities obtained with multi-scale images. We find minimal improvement with such an inference strategy. In the sixth and seventh rows, directly using video-level Mask2Former or VITA, the state-of-the-art video instance segmentation framework, merely brings marginal performance gains. Whereas, using the crop output from CropFormer's batch decoder achieves a significant $\\mathrm{AP^e}$ gain as indicated by the last row. By combining the full image and four crop outputs from CropFormer (final row), we obtain even stronger $1.5\\mathrm{AP^e}$ gains compared to the baseline (first row). Compared to the", + "bbox": [ + 496, + 598, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4053", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9eaab0af6855c10b9a0171556f66cea9780c0aaaa73dfab6c2144eda4521ed81.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInput PreprocessingGPU Type\\(AP^e\\)GPU Memory (Train)GPU Memory (Inference)Train Time (GPU day)Inference Time (ms)
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.55.8G3.0G2.4637
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (2700, 4500)\\)A100-40GOOMOOM-OOM-
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (2700, 4500)\\)A100-80G40.958.8G10.4G60.83312
V-Mask2Former [6]\\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.710.4G6.7G6.81503
VITA [16]\\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.911.9G6.8G10.21512
CropFormer\\(\\xrightarrow{\\text{crop}(\\delta)} (800, 1333)\\)A100-40G41.011.4G6.9G9.81503
", + "bbox": [ + 109, + 90, + 859, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/149b176edaaea1262a1752a11e11b338a55167c8b0f334f79a772657ed93636a.jpg", + "table_caption": [ + "Table 7: Comparison with high-resolution methods on EntitySeg with multi-scale training. $(\\mathrm{x},\\mathrm{y})$ indicates number of pixels along the shortest (x) and longest (y) sides of image. Row 1, 2, & 3 are single-scale Mask2Formers trained and evaluated at various single-scale resolutions. (V-Mask2Former & VITA) mask predictions from full image and four crops are fused with either video-level Mask2Former or VITA." + ], + "table_footnote": [], + "table_body": "
Iterations (Train)Mask2FormerCropFormerDecoderAPeAP50AP75
1×39.541.0 (+1.5)I-O39.356.739.8
2×40.242.1 (+1.9)B-O39.156.639.7
3×40.942.8 (+1.9)B-C40.257.540.8
4×40.942.7 (+1.8)B-OC41.058.441.9
(a) metric: APe(b)
", + "bbox": [ + 80, + 261, + 478, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask2Former baselines with a naive increase in image resolution, CropFormer is far more efficient computationally.", + "bbox": [ + 78, + 422, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Longer Training Schedules Table 8(a) indicates that CropFormer consistently improves over the baseline Mask2Former when we adopt $2 \\times /3 \\times$ training schedules. A similar trend is observed when using Swin-L [33] backbone, which is provided in our supplementary file.", + "bbox": [ + 76, + 465, + 468, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoder Performance Table 8(b) shows the impact of using different decoder combinations in CropFormer for mask prediction during inference. Using just the crop outputs from CropFormer's batch decoder (B-C) outperforms using just the full image view (I-O, B-O). Combining both the full image and crops (B-OC) gives the best result.", + "bbox": [ + 76, + 554, + 468, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CropDataloger Table 9(a) and (b) show the ablation study on the usage of crop ratio and crop type. As indicated by Table 9(a), CropFormer obtains the best performance with a crop ratio $\\delta$ of 0.7. Smaller crop ratios perform worse and it might be a problem caused by queries that are not robust against drastic image transformation.", + "bbox": [ + 76, + 657, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Association Module Table 9(c) shows the ablation study on the structure of the association module. There is a slight performance difference between using self-attention or a feedforward module following cross-attention.", + "bbox": [ + 76, + 760, + 468, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to other high-quality segmentation methods. Table 10(a) shows that our CropFormer significantly outperforms recent high-quality instance-level segmentation methods like Mask Transfiner [18] and PatchDCT [56] on both traditional and boundary APs.", + "bbox": [ + 76, + 830, + 468, + 906 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3ede0a76b697c1c7725fb1354d8b9c3099b6068534bb688929b39e6e008421e8.jpg", + "table_caption": [ + "Table 8: Ablation study on: (a) training schedules; (b) CropFormer's decoder combinations. In sub-table(b), the 'Decoder' column indicates whether we use the predictions of the full image ('O'), four cropped patches ('C'), or both of them ('OC') from the 'Image' or 'Batch' decoder." + ], + "table_footnote": [], + "table_body": "
δAPeTrainTestAPeXAttSAttFFNAPe
0.538.5RandomFixed (4)39.7✓○○40.7
0.640.2Fixed (4)Fixed (4)41.0✓○✓40.8
0.741.0Fixed (4)Fixed (8)41.3✓✓○40.8
0.840.9Fixed (8)Fixed (8)41.0✓✓✓41.0
(a)(b)(c)
", + "bbox": [ + 506, + 261, + 877, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e69df0e84f71413b854bf369f50dcba49d7b44f1504764b09d147a24cd96809f.jpg", + "table_caption": [ + "Table 9: Ablation study on the usage of crop ratio $\\delta$ , crop type and association module in CropFormer. In sub-table (b), 'Random' indicates random crops and 'Fixed (4/8)' indicates 4 or 8 fixed corner crops. In sub-table(c), $\\checkmark$ and $\\circ$ means whether we use the module or not." + ], + "table_footnote": [], + "table_body": "
MethodAPeAPbTaskMask2FormerCropFormer
Mask Transfiner [18]33.725.0COCO-INS50.151.9
PatchDCT [56]35.427.4COCO-PAN57.859.1
Mask2Former39.531.2ADE20K-INS34.936.4
Ours41.033.7ADE20K-PAN48.149.7
(a)(b) metrics: AP for INS; PQ for PAN
", + "bbox": [ + 506, + 416, + 901, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10: (a) Comparison with high-quality segmentation methods on EntitySeg task/dataset using $\\mathrm{AP}^e$ and boundary AP $(\\mathrm{AP}^b)$ . (b) evaluation on instance (INS) and panoptic (PAN) segmentation tasks on COCO and ADE20K datasets.", + "bbox": [ + 498, + 483, + 890, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Improvement on class-aware segmentation tasks Table 10(b) shows the class-aware task performance of CropFromer with Swin-L [33] backbone and $3\\times$ training schedule. CropFormer can still achieve competitive gains on the instance and panoptic segmentation tasks on COCO and ADE20K datasets. This suggests that CropFormer may provide other benefits that are beyond high-resolution segmentation (e.g., prediction robustness).", + "bbox": [ + 498, + 556, + 890, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 680, + 617, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes an EntitySeg dataset and a CropFormer framework for high-quality entity segmentation, with a strong focus on in-the-wild generalization and high-quality dense segmentation. The EntitySeg dataset contains about 33K images from various domains and resolutions, accompanied by high-quality mask annotations for training and testing. A novel CropFormer framework is also proposed to overcome the computation challenges introduced by EntitySeg dataset's high-quality and high-resolution characteristics. Specifically, it enables fusing multi-view predictions from a low-res full image and its high-res crops by learning batch-level queries. We hope this work will serve as a catalyst for future research on open-world and high-quality image segmentation tasks.", + "bbox": [ + 496, + 704, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4054", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 80, + 87, + 174, + 102 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pixabay. https://pixabay.com. Accessed: 2022-04-17. 4", + "[2] Unsplash. https://unsplash.com. Accessed: 2022-04-17. 4", + "[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPR Workshops, 2017. 4", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI, 2017. 3", + "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2, 5, 7, 8", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 3, 5, 7", + "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 5", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4", + "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 3", + "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 2010. 2, 4", + "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2", + "[13] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In CVPR, 2017. 3", + "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 2, 3, 7", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2, 3", + "[16] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurlPS, 2022. 5, 7, 8", + "[17] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, 2018. 4" + ], + "bbox": [ + 81, + 112, + 470, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 8", + "[19] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 3", + "[20] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019. 1, 2, 7", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In arXiv, 2023. 7", + "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 4", + "[23] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2, 4", + "[24] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in the-wild: Dataset, method, and benchmark suite. TIP, 2021. 7", + "[25] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2", + "[26] Peizhao Li, Jiuming Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. Selfdoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 6", + "[27] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 7", + "[28] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. TPAMI, 2022. 3", + "[29] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3", + "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 3", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 2, 3, 4, 5", + "[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018. 3", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 8" + ], + "bbox": [ + 501, + 94, + 888, + 897 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4055", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Zerun Liu, Fan Zhang, Jingxuan He, Jin Wang, Zhangye Wang, and Lechao Cheng. Text-guided mask-free local image retouching. arXiv preprint arXiv:2212.07603, 2022. 2", + "[35] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3", + "[36] David Marr. Vision: A computational investigation into the human representation and processing of visual information. San Francisco: W.H. Freeman. Print., 1982. 4", + "[37] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 7", + "[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 4", + "[39] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2", + "[40] Lu Qi, Jason Kuen, Jiumiang Gu, Zhe Lin, Yi Wang, Yukang Chen, Yanwei Li, and Jiaya Jia. Multi-scale aligned distillation for low-resolution detection. In CVPR, 2021. 3", + "[41] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. TPAMI, 2022. 1, 2, 7", + "[42] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, et al. Highly accurate dichotomous image segmentation. In ECCV, 2022. 7", + "[43] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2", + "[44] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3, 4", + "[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 5", + "[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshop, 2021. 4", + "[47] Markus Suchi, Timothy Patten, David Fischinger, and Markus Vincze. Easylabel: A semi-automatic pixel-wise object annotation tool for creating robotic RGB-D datasets. In ICRA, 2019. 7", + "[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 3", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6", + "[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021. 2, 3, 5" + ], + "bbox": [ + 81, + 94, + 468, + 896 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Chunjing Xu, and Yanwei Fu. Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10707-10717, 2022. 2", + "[52] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. arXiv preprint arXiv:2104.04691, 2021.7", + "[53] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, 2020. 3", + "[54] Yikai Wang, Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Wei Zhang, and Yanwei Fu. Entity-level text-guided image manipulation. arXiv preprint arXiv:2302.11383, 2023. 2", + "[55] Yingqian Wang, Longguang Wang, Jungang Yang, Wei An, and Yulan Guo. Flickr1024: A large-scale dataset for stereo image super-resolution. In ICCV Workshops, 2019. 4", + "[56] Qinrou Wen, Jirui Yang, Xue Yang, and Kewei Liang. Patchdct: Patch refinement for high quality instance segmentation. In ICLR, 2023. 8", + "[57] Magnus Wrenninger and Jonas Unger. Synscapes: A photorealistic synthetic dataset for street scene parsing. arXiv preprint arXiv:1810.08705, 2018. 3, 4", + "[58] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 5", + "[59] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, 2018. 4", + "[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 7", + "[61] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In ECCV, 2022. 2, 5", + "[62] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 3", + "[63] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACMMM, 2018. 3", + "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 3, 4", + "[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 3, 4, 5", + "[66] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2, 5" + ], + "bbox": [ + 501, + 94, + 890, + 868 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4056", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_model.json b/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e142f6dda66bd31c197b96a6682fac8024452a3b --- /dev/null +++ b/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.312, + 0.133, + 0.658, + 0.155 + ], + "angle": 0, + "content": "High Quality Entity Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.182, + 0.777, + 0.263 + ], + "angle": 0, + "content": "Lu Qi\\(^{1*}\\), Jason Kuen\\(^{2*}\\), Tiancheng Shen\\(^{4*}\\), Jiumiang Gu\\(^{2}\\), Wenbo Li\\(^{4}\\), Weidong Guo\\(^{3\\dagger}\\), Jiaya Jia\\(^{4}\\), Zhe Lin\\(^{2}\\), Ming-Hsuan Yang\\(^{1}\\), \\(^{1}\\)The University of California, Merced \\(^{2}\\)Adobe Research \\(^{3}\\)QQ Brower Lab, Tencent, \\(^{4}\\)The Chinese University of Hong Kong" + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.332, + 0.885, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.6, + 0.893, + 0.66 + ], + "angle": 0, + "content": "Figure 1: Segmentation results on in-the-wild test images: (1) Mask2Former [7] trained for panoptic segmentation [20] on COCO dataset [31]; (2) Mask2Former trained for entity segmentation [41] on COCO dataset [31]; (3) our CropFormer trained on the proposed EntitySeg Dataset. For a fair comparison, all three models use Swin-Large backbone [33] and are trained until full convergence. Our approach provides far more desirable and useful results for many real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.671, + 0.314, + 0.687 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.696, + 0.473, + 0.877 + ], + "angle": 0, + "content": "Dense image segmentation tasks (e.g., semantic, panoptic) are useful for image editing, but existing methods can hardly generalize well in an in-the-wild setting where there are unrestricted image domains, classes, and image resolution & quality variations. Motivated by these observations, we construct a new entity segmentation dataset, with a strong focus on high-quality dense segmentation in the wild. The dataset contains images spanning diverse image domains and entities, along with plentiful high-resolution images and high-quality mask annotations for training and testing. Given the high-quality and -resolution nature of the dataset, we propose CropFormer which is designed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.869 + ], + "angle": 0, + "content": "to tackle the intractability of instance-level segmentation on high-resolution images. It improves mask prediction by fusing high-res image crops that provides more fine-grained image details and the full image. CropFormer is the first query-based Transformer architecture that can effectively fuse mask predictions from multiple image views, by learning queries that effectively associate the same entities across the full image and its crop. With CropFormer, we achieve a significant AP gain of 1.9 on the challenging entity segmentation task. Furthermore, CropFormer consistently improves the accuracy of traditional segmentation tasks and datasets. The dataset and code are released at http://luqi.info-entityv2.github.io/." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.887, + 0.383, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution. † indicates corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4047" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.092, + 0.21, + 0.108 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.117, + 0.473, + 0.344 + ], + "angle": 0, + "content": "Entity segmentation [41] is an emerging segmentation task that focuses on open-world and class-agnostic dense image segmentation, designed to have superior generalization capabilities for unseen-category segmentation. As such, it has great potential in image editing applications, where users can interact with visual entities that exist \"in the wild\". Recent works have leveraged entity segmentation for image manipulation [51, 54] and image retouching [34], where users usually specify in-the-wild text queries (e.g., Oculus, platypus) for local editing. In comparison to traditional closed-set category-based segmentation tasks such as panoptic/instance/semantic segmentation, entity segmentation has the advantage of being all-inclusive and unrestricted, making it easier to find masks that semantically match well with user-specified queries." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.344, + 0.473, + 0.539 + ], + "angle": 0, + "content": "Previous work on entity segmentation [41] has shown promising results, yet it only employed existing panoptic segmentation datasets to verify the feasibility and benefits of entity segmentation. However, existing panoptic segmentation datasets (e.g., COCO [31], Cityscapes [9], ADE20K [65]) suffer from the same issues of closed-set categories that entity segmentation [41] aims to avoid. These datasets are collected and annotated for their original panoptic segmentation task requirements. In view of the lack of a proper entity segmentation dataset, it is difficult to further study and explore on this important task. To address this issue, we present EntitySeg dataset, a new large-scale and high-quality dataset for entity segmentation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.539, + 0.473, + 0.749 + ], + "angle": 0, + "content": "Apart from in-the-wild generalization, we design EntitySeg to prioritize high-quality mask annotations and high-resolution images, which are essential for user-driven image editing in the current era of high-resolution imagery. Image editing workflows usually require pixel-precise segmentation masks to ensure a smooth user experience. However, existing panoptic segmentation datasets generally contain low-resolution images (e.g., COCO[31] has zero images that are at least 1K) and coarsely-annotated masks, making it difficult for models trained on such datasets to produce accurate mask predictions for higher resolution images. This is why we focus on high-resolution images and high-quality masks in the EntitySeg dataset, which clearly distinguishes it from existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.75, + 0.473, + 0.901 + ], + "angle": 0, + "content": "The EntitySeg dataset poses new challenges to existing segmentation methods due to its high-resolution images and high-quality mask annotations. Standard practices of resizing images and mask annotations to 800px used in existing methods [43, 15, 7], result in a heavy loss of fine-grained details. On the other hand, training or evaluating existing segmentation models directly on our high-resolution/quality dataset is intractable due to excessive computation and memory overheads. A naive workaround is to divide the high-resolution image into smaller patches and apply seg" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.094, + 0.892, + 0.184 + ], + "angle": 0, + "content": "mentation model to them separately, but it is unable to exploit global context to resolve ambiguity. Moreover, it is unclear how to combine mask predictions from multiple patches using state-of-the-art Transformer-based segmentation methods which rely on set-prediction [7, 6, 50, 61], as shown in Fig 4 and our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.892, + 0.395 + ], + "angle": 0, + "content": "To this end, we introduce CropFormer, a novel approach for tackling the problem of high-resolution entity segmentation. CropFormer learns to dynamically generate tokens to associate the full image with its high-resolution crops and then fuses them to generate the final mask prediction. During training, one of the predefined four corner crops is randomly selected and fused with the full image. For inference, CropFormer is applied to all four corner crops and the full image to achieve complete fusion. The strength of CropFormer lies in its ability to dynamically generate queries to link the same entities in the full image and its higher-resolution crops, allowing us to benefit from both global image context and high-quality image/annotation details to obtain improved results." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.397, + 0.856, + 0.411 + ], + "angle": 0, + "content": "The main contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.413, + 0.89, + 0.517 + ], + "angle": 0, + "content": "- We present a large-scale, high-quality entity segmentation dataset, comprising of images collected from various domains with diverse resolutions and high-quality masks. Our dataset allows for the evaluation of the generalization and robustness of segmentation models in a unified and open-world manner, due to its domain-diverse and high-resolution setting." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.518, + 0.89, + 0.591 + ], + "angle": 0, + "content": "- We benchmark existing segmentation models on our newly collected dataset. We find that the traditional training setting cannot accommodate our dataset, which comprises over \\(70\\%\\) images with very high resolution." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.593, + 0.89, + 0.652 + ], + "angle": 0, + "content": "- We propose a novel model, CropFormer, with an innovative association module and batch-level decoder to enable the fusion of the full image and its higher-resolution crops." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.653, + 0.89, + 0.743 + ], + "angle": 0, + "content": "- Extensive experiments and thorough analysis demonstrate the efficacy of the proposed model and dataset on various segmentation tasks and settings. In Fig. 1, we compare the results from our best CropFormer (trained on EntitySeg) with the best Mask2Formers [7] trained on COCO-Panoptic [20] and COCO-Entity [41]." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.413, + 0.89, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.642, + 0.775 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image Segmentation Dataset. Numerous datasets have been proposed for semantic, instance, and panoptic segmentation, e.g., Microsoft COCO [31], ADE20K [65], KITTI [12], LVIS [14], PASCAL-VOC [11], OpenImages [23], Cityscapes [9], and so on. In addition, there are some datasets specially designed for specific scenarios, such as amodal segmentation (COCO-Amodal [66], KINS [39]), human segmentation (CelebAMask-HQ [25]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4048" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.501, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.172, + 0.403, + 0.184 + ], + "angle": 0, + "content": "COCO (height: 427px, width: 640px)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.867, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.172, + 0.785, + 0.184 + ], + "angle": 0, + "content": "ADE20K (height: 512px, width: 683px)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.185, + 0.402, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.331, + 0.358, + 0.343 + ], + "angle": 0, + "content": "Unsplash (height: 4220px, width: 3176px)" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.186, + 0.703, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.186, + 0.867, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.241, + 0.834, + 0.253 + ], + "angle": 0, + "content": "Clipart (340, 339)" + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.332, + 0.673, + 0.345 + ], + "angle": 0, + "content": "LAION-400M (height: 3637px, width: 2909px)" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.354, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Figure 2: High-quality mask annotations for low- and high-resolution images collected from existing datasets such as COCO [31], ADE20K [65] and Cityscapes [9] as well as from the internet are presented. For the images collected from the aforementioned datasets, a visual comparison between the original and our annotations is provided in the middle and rightmost sub-figures, where the unannotated regions are shaded in black. It is worth noting that the RGB and mask images shown here have been downsampled, resulting in some quality degradation compared to the original datasets." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.438, + 0.473, + 0.725 + ], + "angle": 0, + "content": "LIP [13], MHP [63]) and domain adaptation (Synscapes [57], GTA5 [44]). Despite the significant contributions from these datasets, there is still a need to fulfill the requirements of real-world applications with high-quality images of large diversity. For example, a segmentation model should be robust to high-resolution images from different domains and able to segment unseen objects in the openworld setting. Most relevant to our work is the ADE20K dataset [65], which has large-scale open-vocabulary categories. However, the collected images are of low resolution and from narrowed domains. In contrast, our dataset collects images from multiple domains, including indoor, outdoor, street scenes, cartoon, and remote images, with over \\(80\\%\\) of the resolution of the images falling within the high-resolution range of 2000px to 8000px. Differently from the ADE20K [64] and LVIS [14] dataset with a predefined category list, our annotation process first conducts class-agnostic mask annotation on each entity and then labels the category information in an open-vocabulary manner." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.734, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Scene Parsing Methods. Scene parsing methods mainly rely on convolution-based dense prediction or transformer-based query prediction. Explicit localization information, such as bounding box proposals or pixel position, is often used for pixel-level mask prediction in these approaches. Examples include FCN [35], DeepLab [5], PSPNet [62], Mask R-CNN [15], PANet [32], SOLO [53], CondInst [48], PanopticFPN [19], and PanopticFCN [28]. Inspired by the success of transformer-based detection tasks such as DETR [4] [29, 30, 10, 40], a number of methods, such as Max-Deeplab [50] and Mask2Former [7], have been pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.59 + ], + "angle": 0, + "content": "posed to directly predict segmentation masks without relying on bounding boxes. The key to these methods is to use learnable queries trained by a set-prediction loss to model the implicit location of the potential instance area. However, query-based set-prediction methods cannot effectively perform multi-view fusion, as the query-to-object assignment can drastically change with the change of image view. This is because the set prediction loss typically adopted by such methods does not enforce any assignment constraints across multiple image views." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.601, + 0.673, + 0.618 + ], + "angle": 0, + "content": "3. EntitySeg Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.837 + ], + "angle": 0, + "content": "The EntitySeg dataset contains 33,227 images with high-quality mask annotations, as illustrated in Fig. 2. Compared with existing datasets, it has three distinctive features. Firstly, \\(71.25\\%\\) and \\(86.23\\%\\) of the images are of high quality, with at least \\(2000\\mathrm{px} \\times 2000\\mathrm{px}\\) and \\(1000\\mathrm{px} \\times 1000\\mathrm{px}\\), respectively, staying in line with the current digital imaging trends. Secondly, our dataset is open-world and not restricted to predefined classes. We consider each semantically-coherent region in the images as an entity, even if it is blurred or hard to recognize semantically. Thirdly, the mask annotation along the boundaries is more precise than those in other datasets. In the following, we will provide an in-depth description of our EntitySeg dataset, as well as an extensive analysis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.839, + 0.667, + 0.853 + ], + "angle": 0, + "content": "3.1. Image Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.854, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Inspired by the collection criterion of COCO [31], we collected most non-iconic images with plentiful contextual information and non-canonical viewpoints. To en" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4049" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.093, + 0.338, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.245, + 0.225, + 0.256 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.093, + 0.581, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.245, + 0.473, + 0.256 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.094, + 0.863, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.245, + 0.758, + 0.256 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.26, + 0.892, + 0.306 + ], + "angle": 0, + "content": "Figure 3: Sub-figures (a) and (b) present the distributions of image resolutions and average number of entities among ADE20K, COCO, and EntitySeg, respectively. Moreover, sub-figure (c) displays the distribution of image sources from which the EntitySeg images were collected." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.474, + 0.541 + ], + "angle": 0, + "content": "sure substantial domain diversity, we have sourced the images from various public datasets and the Internet, with permission for academic research. Specifically, part of the images are from COCO [31], ADE20K [65], Pascal VOC [11], Cityscapes [9], Mapillary Vistas [38], ImageNet [22], Open Images [23], DIV2K [3], Flickr2K [55], Clipart [17], Comic [17], DOTA [59], Synscapes [57] and GTA5 [44]. Moreover, we have crawled high-resolution images from Pixabay [1], Unsplash [2], and LAION-400M \\[46\\]. As shown in Fig. 3(c), most of the images are from high-resolution sources like Pixabay, Unsplash, and LAION-400M. Moreover, Fig. 3(a) reveals that our dataset has more high-resolution images, with resolutions normally distributed within the range of \\(0\\mathrm{px}\\) to \\(15,000\\mathrm{px}\\), in comparison to ADE20K and COCO with images all under \\(1000\\mathrm{px}\\)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.544, + 0.254, + 0.559 + ], + "angle": 0, + "content": "3.2. Image Annotation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.563, + 0.47, + 0.773 + ], + "angle": 0, + "content": "For our dataset, the annotation process consists of the following steps. Firstly, entities in the image are annotated with class-agnostic pixel-level masks, ensuring that no overlaps exist between masks. Subsequently, each mask is annotated with a class label in an open vocabulary manner. Two special considerations should be taken into account here: 1) we aonotate the entity as 'unknown' or 'blurred' if it cannot be easily named or is severely blurred; 2) the entity is annotated as a 'supercategory_other' if we only know it is a supercategory but unable to determine its fine-grained class. Our annotation process is the opposite of those of the popular segmentation datasets like COCO [31] and ADE20K [64], which first predefine their class sets before annotating masks based on the predefined classes." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.773, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Table 2 indicates that our dataset has a more considerable proportion of covered image areas and more masks on average compared to COCO- and ADE20K-Panoptic. This is attributed to our novel annotation process that enables us to take all semantically-coherent regions into account and assign class labels accordingly. Besides, our annotation procedure is more similar to the human visual system. As" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.31, + 0.655, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.354, + 0.596, + 0.363 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.311, + 0.864, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.354, + 0.806, + 0.363 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.369, + 0.89, + 0.398 + ], + "angle": 0, + "content": "Table 1: Annotation consistency of class-agnostic localization and class-aware categories among annotators." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.4, + 0.902, + 0.441 + ], + "angle": 0, + "content": "
DatasetImageRes (avg)↑EntityNum (avg)↑EntityNum (max)↑Entity Complexity↓Entity Simplicity↓Valid Area↑
COCO522.511.2950.7580.5810.891
ADE20K461.313.62550.8020.6060.914
EntitySeg2700.718.12360.7190.5380.999
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.451, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Table 2: The statistical comparisons between COCO [31], ADE20K [65] and EntitySeg. 'ImageRes (avg)', 'EntityNum (avg)', 'Valid Area', 'Entity Complexity', and 'Entity Simplicity' refer to the average value of image resolution size, entity numbers per image, valid area ratio per image, average entity complexity and simplicity respectively. 'EntityNum (max)' means the maximum per-image number of entities within each dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.581, + 0.89, + 0.627 + ], + "angle": 0, + "content": "evidenced in [36], the human vision system is intrinsically class-agnostic and can recognize entities without comprehending their usage and purpose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Annotation consistency is crucial to any human-labeled dataset since it implies whether the annotation task is well-defined and proper. To study this, we randomly selected 500 images from EntitySeg (1.5% of the entire dataset), and asked another two annotators to annotate them after four months since the first round. Table 1(a) shows that the class-agnostic mask mAP in the first step of the annotation process compared to those of the other two annotators. We use one as ground truth and another one as prediction results. In Table 1(b), we also evaluate the category consistency by accuracy (ACC) under the same mask annotations. These two tables demonstrate that our EntitySeg has high annotation consistency in the mask and category labeling stages." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.834, + 0.667, + 0.849 + ], + "angle": 0, + "content": "3.3. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.854, + 0.891, + 0.9 + ], + "angle": 0, + "content": "In our EntitySeg dataset, we annotate all images at the entity level, regardless of whether they belong to thing or stuff in previous datasets. As shown in Fig. 1, we show" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.885, + 0.435, + 0.899 + ], + "angle": 0, + "content": "1We removed not-safe-for-work images via strict CLIP filtering." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4050" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.094, + 0.47, + 0.154 + ], + "angle": 0, + "content": "some low- and high-resolution photos and their mask annotations to highlight our high-quality pixel-level annotation quality. For example, one can zoom in on the Unsplash image to see how high-quality the railing's mask is." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.155, + 0.471, + 0.395 + ], + "angle": 0, + "content": "Table 2 shows the quantitative comparison among our dataset, COCO-Panoptic [31], and ADE20K-Panoptic [65]. In the EntitySeg dataset, each image has 18.1 entities on average, which is more than 11.2 and 13.6 entities in COCO and ADE20K. For a fair comparison, we treat each uncountable stuff region as one segment although it is disconnected. The detailed comparison among those three datasets on the distribution of entity number is shown in Fig. 3(b). Furthermore, the shapes of entities in our dataset are more complex than those COCO [31] and ADE20K [65] as represented by the columns 'Entity Complexity' and 'Entity Simplicity' of Table 2 where an entity with a large convexity and simplicity value means it is a simple shape (and both metrics achieve their maximum value of 1.0 for a circle [66]). More details about how complexity and simplicity are calculated can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.399, + 0.47, + 0.565 + ], + "angle": 0, + "content": "Class-Aware Annotation. We select a subset of 11,580 images from the dataset for class labeling, for investigating the diversity of semantic classes in our dataset. In our class annotation process, the class is labeled in an open-vocabulary and free-form manner, then the class set is continually expanded. After the class annotation process, we organize and merge the annotated classes according to Word-Tree [45]. This results in 535 things and 109 stuff classes for those 11,580 images that we use to construct the class-aware EntityClass dataset. Please refer to our supplementary file for more analysis on EntityClass dataset." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.58, + 0.212, + 0.597 + ], + "angle": 0, + "content": "4. CropFormer" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.47, + 0.863 + ], + "angle": 0, + "content": "We propose CropFormer which is the first Transformer-based set-prediction segmentation method that is capable of fusing the mask predictions from multiple image views. Inspired by existing methods [8, 7, 50, 61], we design CropFormer to learn \\(N\\) number of \\(K\\)-dimensional queries \\(\\mathbf{Q} \\in \\mathbb{R}^{N \\times K}\\) to generate mask embeddings \\(\\mathbf{E} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}\\). \\(\\mathbf{E}\\) then acts as convolution filters that are applied to pixel-level mask features \\(\\mathbf{P}_2 \\in \\mathbb{R}^{T \\times H \\times W}\\) to generate \\(N\\) segmentation masks \\(\\mathbf{U}^{\\mathbf{m}} \\in \\mathbb{R}^{N \\times T \\times H \\times W}\\), where \\(T, H\\) and \\(W\\) correspond to the image view, height, and width dimensions. However, such a rudimentary design can only work well with single-view inputs and cannot effectively fuse the results from multiple views. Here, we introduce a novel association module and batch-level decoder to achieve the goal of CropFormer, which is to exploit global context from full image and fine-grained local details from crops for high-quality segmentation. CropFormer is illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.869, + 0.47, + 0.899 + ], + "angle": 0, + "content": "Crop Dataloger. In CropFormer framework, the crop dataloger is designed to generate a batch of images that si" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.094, + 0.892, + 0.32 + ], + "angle": 0, + "content": "multaneously include the full images and their corresponding crops. There are two steps in our dataloader: crop and resize. First, we augment the full image \\( I^o \\) with a crop that is part of the full image. The crop is randomly extracted from one of the fixed image corners: upper-left, upper-right, bottom-left, bottom-right. The crop size is controlled by a fixed ratio hyperparameter \\( \\delta \\in \\mathbb{R} \\) relative to the full image size. Then, we resize both the full image and crop to the same size. In this way, we construct an input tensor \\( \\mathbf{I} \\in \\mathbb{R}^{1 \\times 2 \\times H' \\times W' \\times 3} \\) with 2 elements (full image & 1 crop) along the view dimension. Compared to the full image, our corner crops preserve more image details (e.g., local information) that are useful for high-quality segmentation. Specifically, given an dataloader \\( \\mathcal{D} \\), we define the data preprocessing as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.332, + 0.89, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = \\left\\{I ^ {o}, I ^ {c} \\right\\}, \\quad I ^ {o} \\sim \\mathcal {D} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.891, + 0.468 + ], + "angle": 0, + "content": "where \\(I^o\\) is sampled from \\(\\mathcal{D}\\), and \\(I^c\\) is the cropped image controlled by fixed ratio \\(\\delta\\). Unlike the random cropping in [58, 16], we adopt a fixed cropping solution to keep the training and inference consistent. Such a rigid cropping strategy provides better inference efficiency since, rather than bunches of random crops, evaluating a fixed number of crops during inference is sufficient enough." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.89, + 0.565 + ], + "angle": 0, + "content": "Image Encoder and Decoder. Based on learnable image-level queries \\(\\mathbf{Q}_i\\in \\mathbb{R}^{N\\times K}\\), we use image encoder \\(\\Theta\\) and image-level decoder \\(\\Phi_{i}\\) to generate image-level embeddings \\(\\mathbf{E_i}\\in \\mathbb{R}^{N\\times 2\\times 1\\times 1\\times K}\\) for the full input image and its crop. Given the input tensor (I) and queries \\((\\mathbf{Q}_i)\\), we define \\(\\mathbf{E_i}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.567, + 0.89, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {\\mathbf {i}} = \\Phi_ {i} \\left(\\mathbf {Q} _ {\\mathbf {i}}, \\Theta (\\mathbf {I})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.89, + 0.667 + ], + "angle": 0, + "content": "where \\(\\Phi_i(\\cdot)\\) is a Transformer-based image-level decoder. \\(\\mathbf{E_i}\\) is used for image-level entityness prediction and pixellevel prediction using the low-level image feature \\(\\mathbf{P_2}\\) (derived from the image encoder). The prediction process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.681, + 0.89, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {\\mathbf {i}} ^ {\\mathrm {e}}, \\mathbf {U} _ {\\mathbf {i}} ^ {\\mathrm {m}} = \\operatorname {P r e d H e a d} _ {\\mathbf {i}} \\left(\\mathbf {E} _ {\\mathbf {i}}, \\mathbf {P} _ {\\mathbf {2}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.89, + 0.77 + ], + "angle": 0, + "content": "where \\(\\mathbf{U}_{\\mathrm{i}}^{\\mathrm{e}}\\) and \\(\\mathbf{U}_{\\mathrm{i}}^{\\mathrm{m}}\\) denote the entityness prediction and pixel-level mask outputs. Here, we employ \\(i\\) subscript to differentiate image-level embeddings and mask outputs from those of the association module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Association Module. Following [6], our association module aims to generate batch queries \\(\\mathbf{Q}_{\\mathrm{b}}\\) that are fully shared by the full image and its crop to represent the same entities consistently. However, instead of treating \\(\\mathbf{Q}_{\\mathrm{b}}\\) as learnable network parameters, considering that \\(\\mathbf{E}_{\\mathrm{i}}\\) already contains strong segmentation features, we generate \\(\\mathbf{Q}_{\\mathrm{b}}\\) directly from \\(\\mathbf{E}_{\\mathrm{i}} = \\{\\mathbf{E}_{I^o},\\mathbf{E}_{I^c}\\}\\). Particularly, we use a Transformer architecture with cross-attention (\\(f_{\\mathrm{XAtt}}\\)) and self" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.107, + 0.887, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.284, + 0.893, + 0.331 + ], + "angle": 0, + "content": "Figure 4: Framework of the proposed CropFormer. The red box indicates cropped region randomly sampled from four fixed image corners. In image-level prediction, the same entity across different image views may be assigned to different queries. Our association module and batch decoder can effectively associate the same entity across different views with a single query." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.338, + 0.279, + 0.354 + ], + "angle": 0, + "content": "attention \\((f_{\\mathrm{SAtt}})\\) to obtain \\(\\mathbf{Q}_{\\mathbf{b}}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.36, + 0.47, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\mathbf {b}} = \\operatorname {F F N} \\left(f _ {\\mathrm {S A t t}} \\left(f _ {\\mathrm {X A t t}} \\left(\\underbrace {f _ {\\mathrm {q}} \\left(\\mathbf {E} _ {I ^ {o}}\\right)} _ {\\text {q u e r y}}, \\underbrace {f _ {\\mathrm {k}} \\left(\\mathbf {E} _ {\\mathbf {i}}\\right)} _ {\\text {k e y}}, \\underbrace {f _ {\\mathrm {v}} \\left(\\mathbf {E} _ {\\mathbf {i}}\\right)} _ {\\text {v a l u e}}\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.403, + 0.47, + 0.478 + ], + "angle": 0, + "content": "where FFN is a feed-forward network, \\( f_{\\{\\mathbf{q}, \\mathbf{k}, \\mathbf{v}\\}}(\\cdot) \\) are linear transformations. Considering the importance of full image for entity segmentation, inspired by [49, 26] we take the image-level embeddings of full image \\( \\mathbf{E}_{I^o} \\) as query, while treating all image-level embeddings \\( \\mathbf{E}_i \\) as key and value." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.481, + 0.47, + 0.513 + ], + "angle": 0, + "content": "Batch-level Decoder. Given \\(\\mathbf{Q}_{\\mathbf{b}} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}\\), we obtain batch embeddings \\(\\mathbf{E}_{\\mathbf{b}} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.52, + 0.47, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {\\mathbf {b}} = \\Phi_ {b} (\\mathbf {Q} _ {\\mathbf {b}}, \\Theta (\\mathbf {I})), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.543, + 0.47, + 0.618 + ], + "angle": 0, + "content": "where \\(\\Phi_b(\\cdot)\\) denotes the batch-level decoder. Thus, we broadcast \\(\\mathbf{E}_{\\mathbf{b}}\\) to the shape of \\(N\\times 2\\times 1\\times 1\\times K\\) to share it between the full image and its crop. Finally, the broadcasted \\(\\mathbf{E}_{\\mathbf{b}}\\) and low-level batch features \\(\\mathbf{P_2}\\) are used for batch-level mask classification and pixel-wise prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.626, + 0.47, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {\\mathrm {b}} ^ {\\mathrm {e}}, \\mathbf {U} _ {\\mathrm {b}} ^ {\\mathrm {m}} = \\operatorname {P r e d H e a d} _ {\\mathrm {b}} \\left(\\mathbf {E} _ {\\mathrm {b}}, \\mathbf {P} _ {2}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.65, + 0.47, + 0.725 + ], + "angle": 0, + "content": "For the image-level decoder, the queries are separate for each image/crop. Whereas, for the batch-level decoder, the queries are shared between the image and its crop. We illustrate the differences of the image- and batch-level decoders in Fig. 5, where we regard two mage views as an entirety." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.728, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Training. During training, we employ two separate losses \\(\\mathcal{L}_i\\) and \\(\\mathcal{L}_b\\) for image- and batch-level predictions, with respect to ground truth \\(\\mathbf{G}\\). The main difference between the two losses lies in whether the same entities in full image and crop are tied to the same queries or not." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.803, + 0.443, + 0.818 + ], + "angle": 0, + "content": "The overall training loss for our model is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.824, + 0.469, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\sum_ {\\mathbf {k} \\in \\{\\mathbf {i}, \\mathbf {b} \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {e}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {e}}\\right) + \\sum_ {\\mathbf {k} \\in \\{\\mathbf {i}, \\mathbf {b} \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {b c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {m}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {m}}\\right) + \\\\ \\sum_ {\\mathbf {k} \\in \\{i, b \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {d i c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {m}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {m}}\\right), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.34, + 0.627, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.34, + 0.75, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.34, + 0.868, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.457, + 0.89, + 0.486 + ], + "angle": 0, + "content": "Figure 5: The illustration of image- and batch-level decoder and association module." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.498, + 0.89, + 0.56 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{ce}}\\) and \\(\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{ce}}\\) denote binary cross-entropy loss for image- and batch-level entityness prediction. Similarly, \\(\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{bce}}\\), \\(\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{bce}}\\), \\(\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{dice}}\\) and \\(\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{dice}}\\) denote the binary cross-entropy and dice loss for image- and batch-level mask prediction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.57, + 0.89, + 0.676 + ], + "angle": 0, + "content": "Inference. In CropFormer, the same entities between the full image and its crop are represented by the same queries by virtue of the association module. For the final segmentation output, we fuse the per-pixel mask predictions obtained from the full image and every crop of the 4 corners by using an average operation. The confidence score of each entity comes from the batch-level entityness prediction score." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.633, + 0.71 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.719, + 0.89, + 0.854 + ], + "angle": 0, + "content": "We conduct our main experiments on the entity segmentation with proposed EntitySeg Dataset in this work. First, we benchmark our dataset in several segmentation tasks, including entity, instance, semantic and panoptic segmentation. Then, we ablate our proposed CropFormer to demonstrate its effectiveness. Note that, in Sec. 4, we omit the conventional data batch dimension to improve clarity, but in practice, we train CropFormer with a batch of multiple dataset images and their crops, following existing methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Training Details. We leave out some details like the training schedule and some ablation studies due to paper length limitations (see supplementary for details)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4052" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.477, + 0.307 + ], + "angle": 0, + "content": "
TaskMetricsMethodBackbonePretrainPerformance
SEMmIoUDeeplabv3R-50ImageNet27.9
Mask2FormerR-50ImageNet37.8
R-50COCO43.3
Swin-TCOCO45.0
Swin-LCOCO50.5
INSAPMask-RCNNR-50ImageNet5.0
Mask2FormerR-50ImageNet13.0
R-50COCO20.3
Swin-TCOCO22.7
Swin-LCOCO30.3
PANPQPanopticFPNR-50ImageNet3.6
Mask2FormerR-50ImageNet5.5
R-50COCO9.6
Swin-TCOCO9.8
Swin-LCOCO13.4
ENTAPeMask-RCNNR-50ImageNet24.9
Mask2FormerR-50ImageNet26.0
R-50COCO35.2
Swin-TCOCO42.9
Swin-LCOCO46.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.317, + 0.471, + 0.407 + ], + "angle": 0, + "content": "Table 3: Benchmark on several segmentation tasks on the EntitySeg Dataset. The 'SEM', 'INS', 'PAN', and 'ENT' indicate the task of semantic, instance, panoptic and entity segmentation. We note that we leave out all details and more experiments in the supplementary file due to paper length limitations." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.422, + 0.47, + 0.478 + ], + "angle": 0, + "content": "
Transfer DatasetTypeSEM (mIoU)INS (AP)PAN (PQ)ENT (APe)
ADE20KImage57.2(+1.0)36.0(+1.1)49.0(+0.9)40.6(+1.5)
YVIS2019 [60]Video-61.3(+0.9)--
VIPSeg [37]Video--37.0(+1.1)-
UVO [52]Video---29.5(+2.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.488, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Table 4: The relative performance gains of EntitySeg pretraining over COCO-Panoptic pretraining when fine-tuned on various image/video-level segmentation datasets." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.561, + 0.275, + 0.575 + ], + "angle": 0, + "content": "5.1. Dataset Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.581, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Benchmark on EntitySeg Dataset. In Table 3, we provide results on semantic, instance, panoptic, and entity segmentation on EntitySeg dataset with some baseline methods. Similar to existing segmentation datasets like COCO and ADE20K, the use of COCO pre-trained weight and a stronger backbone can benefit each segmentation performance of EntitySeg Dataset. Compared to COCO and ADE20K, EntitySeg dataset is more challenging as evidenced by the smaller performance numbers (e.g., Swin-L Mask2Former's PQ on COCO-PAN is \\(57.8\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.74, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Transferring EntitySeg to other general segmentation tasks. Table 4 shows the performance comparison of transferring models pretrained on either COCO or EntitySeg dataset to other image/video-level segmentation tasks. Owing to its emphasis on in-the-wild generalization, EntitySeg provides greater improvements than COCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.839, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Benefit of EntitySeg for high-quality image segmentation. In Table 5, we transfer Mask2Former pretrained on EntitySeg dataset for finetuning on Dichotomous Image Segmentation [42]. Compared to state-of-the-art ISNet [42]" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.091, + 0.868, + 0.126 + ], + "angle": 0, + "content": "
MethodFβmx↑Fβw↑M↓Sα↑Eφm↑HCEγ↓
IS-Net [42].791.717.074.813.8561116
Ours.872.789.035.896.913969
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.136, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Table 5: Evaluation on Dichotomous Image Segmentation (DIS5K dataset [42].) 'Ours' indicates Mask2Former pretrained on EntitySeg and then finetuned on DIS5K." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.197, + 0.887, + 0.253 + ], + "angle": 0, + "content": "
Train DatasetTaskLVIS [14]OCID [47]CAMO [24]FSS [27]Avg. AR
COCOPAN24.3839.9443.8064.9643.27
COCOENT27.7545.8859.7075.6052.23
SAM-H [21]Prompt49.2555.1775.3373.6063.34
EntitySegENT33.5676.6373.2086.2667.41
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Table 6: In-the-wild segmentation performance (AR@100 with IoU threshold 0.5). We directly evaluate models trained on either EntitySeg or COCO on large-vocabulary segmentation LVIS [14], object clutter indoor segmentation for robot grab [47], camouflaged instance segmentation CAMO [24], few-shot segmentation FSS [27]. All the models are based on Mask2Former with Swin-Large backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.44 + ], + "angle": 0, + "content": "that is carefully designed for the task, Mask2Former pretrained on EntitySeg outperforms ISNet by a large margin, despite not having any task-specific design. This is by virtue of the high-quality mask annotations in EntitySeg dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.539 + ], + "angle": 0, + "content": "Segmentation 'in the wild' Table 6 compares in-the-wild segmentation performance between Mask2Formers trained on either EntitySeg or COCO. Here, we directly use the trained models without finetuning on the evaluated datasets. The EntitySeg outperforms COCO-Panoptic [20] and COCO-Entity [41] by large margins." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.547, + 0.864, + 0.563 + ], + "angle": 0, + "content": "5.2. A Comprehensive Analysis of CropFormer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.89, + 0.598 + ], + "angle": 0, + "content": "We first ablate Cropformer on the entity segmentation task and then report results on other segmentation datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Table 7 shows the benefits of CropFormer on segmentation performance and computation cost compared to the Mask2Former [7], its multi-scale extension, as well as related methods [6, 16] that supports multi-view fusion. The first four rows are our baseline Mask2Former with the single-scale inference (800, 1040, or 2700 for shortest side) and (1333, 1732 or 4500 for longest side) for the full images, where \\(1040 = 800 \\times \\delta\\) and \\(1732 = 1333 \\times \\delta\\) with \\(\\delta = 0.7\\) by default. For the results in the fifth row, we use test-time Hungarian matching to associate the same entities obtained with multi-scale images. We find minimal improvement with such an inference strategy. In the sixth and seventh rows, directly using video-level Mask2Former or VITA, the state-of-the-art video instance segmentation framework, merely brings marginal performance gains. Whereas, using the crop output from CropFormer's batch decoder achieves a significant \\(\\mathrm{AP^e}\\) gain as indicated by the last row. By combining the full image and four crop outputs from CropFormer (final row), we obtain even stronger \\(1.5\\mathrm{AP^e}\\) gains compared to the baseline (first row). Compared to the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4053" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.091, + 0.861, + 0.187 + ], + "angle": 0, + "content": "
MethodInput PreprocessingGPU Type\\(AP^e\\)GPU Memory (Train)GPU Memory (Inference)Train Time (GPU day)Inference Time (ms)
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.55.8G3.0G2.4637
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (2700, 4500)\\)A100-40GOOMOOM-OOM-
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (2700, 4500)\\)A100-80G40.958.8G10.4G60.83312
V-Mask2Former [6]\\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.710.4G6.7G6.81503
VITA [16]\\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.911.9G6.8G10.21512
CropFormer\\(\\xrightarrow{\\text{crop}(\\delta)} (800, 1333)\\)A100-40G41.011.4G6.9G9.81503
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.197, + 0.892, + 0.256 + ], + "angle": 0, + "content": "Table 7: Comparison with high-resolution methods on EntitySeg with multi-scale training. \\((\\mathrm{x},\\mathrm{y})\\) indicates number of pixels along the shortest (x) and longest (y) sides of image. Row 1, 2, & 3 are single-scale Mask2Formers trained and evaluated at various single-scale resolutions. (V-Mask2Former & VITA) mask predictions from full image and four crops are fused with either video-level Mask2Former or VITA." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.262, + 0.479, + 0.327 + ], + "angle": 0, + "content": "
Iterations (Train)Mask2FormerCropFormerDecoderAPeAP50AP75
1×39.541.0 (+1.5)I-O39.356.739.8
2×40.242.1 (+1.9)B-O39.156.639.7
3×40.942.8 (+1.9)B-C40.257.540.8
4×40.942.7 (+1.8)B-OC41.058.441.9
(a) metric: APe(b)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.337, + 0.47, + 0.412 + ], + "angle": 0, + "content": "Table 8: Ablation study on: (a) training schedules; (b) CropFormer's decoder combinations. In sub-table(b), the 'Decoder' column indicates whether we use the predictions of the full image ('O'), four cropped patches ('C'), or both of them ('OC') from the 'Image' or 'Batch' decoder." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.424, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Mask2Former baselines with a naive increase in image resolution, CropFormer is far more efficient computationally." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.467, + 0.47, + 0.543 + ], + "angle": 0, + "content": "Longer Training Schedules Table 8(a) indicates that CropFormer consistently improves over the baseline Mask2Former when we adopt \\(2 \\times /3 \\times\\) training schedules. A similar trend is observed when using Swin-L [33] backbone, which is provided in our supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.555, + 0.47, + 0.646 + ], + "angle": 0, + "content": "Decoder Performance Table 8(b) shows the impact of using different decoder combinations in CropFormer for mask prediction during inference. Using just the crop outputs from CropFormer's batch decoder (B-C) outperforms using just the full image view (I-O, B-O). Combining both the full image and crops (B-OC) gives the best result." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.658, + 0.47, + 0.749 + ], + "angle": 0, + "content": "CropDataloger Table 9(a) and (b) show the ablation study on the usage of crop ratio and crop type. As indicated by Table 9(a), CropFormer obtains the best performance with a crop ratio \\(\\delta\\) of 0.7. Smaller crop ratios perform worse and it might be a problem caused by queries that are not robust against drastic image transformation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.761, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Association Module Table 9(c) shows the ablation study on the structure of the association module. There is a slight performance difference between using self-attention or a feedforward module following cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.832, + 0.47, + 0.907 + ], + "angle": 0, + "content": "Comparison to other high-quality segmentation methods. Table 10(a) shows that our CropFormer significantly outperforms recent high-quality instance-level segmentation methods like Mask Transfiner [18] and PatchDCT [56] on both traditional and boundary APs." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.262, + 0.878, + 0.327 + ], + "angle": 0, + "content": "
δAPeTrainTestAPeXAttSAttFFNAPe
0.538.5RandomFixed (4)39.7✓○○40.7
0.640.2Fixed (4)Fixed (4)41.0✓○✓40.8
0.741.0Fixed (4)Fixed (8)41.3✓✓○40.8
0.840.9Fixed (8)Fixed (8)41.0✓✓✓41.0
(a)(b)(c)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.411 + ], + "angle": 0, + "content": "Table 9: Ablation study on the usage of crop ratio \\( \\delta \\), crop type and association module in CropFormer. In sub-table (b), 'Random' indicates random crops and 'Fixed (4/8)' indicates 4 or 8 fixed corner crops. In sub-table(c), \\( \\checkmark \\) and \\( \\circ \\) means whether we use the module or not." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.417, + 0.902, + 0.482 + ], + "angle": 0, + "content": "
MethodAPeAPbTaskMask2FormerCropFormer
Mask Transfiner [18]33.725.0COCO-INS50.151.9
PatchDCT [56]35.427.4COCO-PAN57.859.1
Mask2Former39.531.2ADE20K-INS34.936.4
Ours41.033.7ADE20K-PAN48.149.7
(a)(b) metrics: AP for INS; PQ for PAN
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.484, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Table 10: (a) Comparison with high-quality segmentation methods on EntitySeg task/dataset using \\(\\mathrm{AP}^e\\) and boundary AP \\((\\mathrm{AP}^b)\\). (b) evaluation on instance (INS) and panoptic (PAN) segmentation tasks on COCO and ADE20K datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Improvement on class-aware segmentation tasks Table 10(b) shows the class-aware task performance of CropFromer with Swin-L [33] backbone and \\(3\\times\\) training schedule. CropFormer can still achieve competitive gains on the instance and panoptic segmentation tasks on COCO and ADE20K datasets. This suggests that CropFormer may provide other benefits that are beyond high-resolution segmentation (e.g., prediction robustness)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.619, + 0.697 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This paper proposes an EntitySeg dataset and a CropFormer framework for high-quality entity segmentation, with a strong focus on in-the-wild generalization and high-quality dense segmentation. The EntitySeg dataset contains about 33K images from various domains and resolutions, accompanied by high-quality mask annotations for training and testing. A novel CropFormer framework is also proposed to overcome the computation challenges introduced by EntitySeg dataset's high-quality and high-resolution characteristics. Specifically, it enables fusing multi-view predictions from a low-res full image and its high-res crops by learning batch-level queries. We hope this work will serve as a catalyst for future research on open-world and high-quality image segmentation tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4054" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.088, + 0.176, + 0.103 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.113, + 0.47, + 0.14 + ], + "angle": 0, + "content": "[1] Pixabay. https://pixabay.com. Accessed: 2022-04-17. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.143, + 0.472, + 0.169 + ], + "angle": 0, + "content": "[2] Unsplash. https://unsplash.com. Accessed: 2022-04-17. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.172, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPR Workshops, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.215, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.259, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.315, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.372, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.429, + 0.471, + 0.469 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.472, + 0.471, + 0.54 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.543, + 0.47, + 0.583 + ], + "angle": 0, + "content": "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.586, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 2010. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.629, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.672, + 0.47, + 0.726 + ], + "angle": 0, + "content": "[13] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.729, + 0.47, + 0.769 + ], + "angle": 0, + "content": "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.772, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.802, + 0.47, + 0.842 + ], + "angle": 0, + "content": "[16] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurlPS, 2022. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.844, + 0.47, + 0.898 + ], + "angle": 0, + "content": "[17] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.113, + 0.472, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.095, + 0.89, + 0.136 + ], + "angle": 0, + "content": "[18] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.89, + 0.176 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.89, + 0.218 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.89, + 0.275 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In arXiv, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.89, + 0.316 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.89, + 0.385 + ], + "angle": 0, + "content": "[23] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.89, + 0.453 + ], + "angle": 0, + "content": "[24] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in the-wild: Dataset, method, and benchmark suite. TIP, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.89, + 0.496 + ], + "angle": 0, + "content": "[25] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.89, + 0.552 + ], + "angle": 0, + "content": "[26] Peizhao Li, Jiuming Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. Selfdoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.89, + 0.593 + ], + "angle": 0, + "content": "[27] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.89, + 0.649 + ], + "angle": 0, + "content": "[28] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.89, + 0.69 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.89, + 0.732 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.89, + 0.788 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.89, + 0.828 + ], + "angle": 0, + "content": "[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.89, + 0.898 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.095, + 0.89, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4055" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.095, + 0.47, + 0.136 + ], + "angle": 0, + "content": "[34] Zerun Liu, Fan Zhang, Jingxuan He, Jin Wang, Zhangye Wang, and Lechao Cheng. Text-guided mask-free local image retouching. arXiv preprint arXiv:2212.07603, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.138, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[35] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.182, + 0.469, + 0.222 + ], + "angle": 0, + "content": "[36] David Marr. Vision: A computational investigation into the human representation and processing of visual information. San Francisco: W.H. Freeman. Print., 1982. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.225, + 0.469, + 0.265 + ], + "angle": 0, + "content": "[37] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.268, + 0.469, + 0.309 + ], + "angle": 0, + "content": "[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.312, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[39] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.355, + 0.469, + 0.396 + ], + "angle": 0, + "content": "[40] Lu Qi, Jason Kuen, Jiumiang Gu, Zhe Lin, Yi Wang, Yukang Chen, Yanwei Li, and Jiaya Jia. Multi-scale aligned distillation for low-resolution detection. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.399, + 0.469, + 0.439 + ], + "angle": 0, + "content": "[41] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. TPAMI, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.442, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[42] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, et al. Highly accurate dichotomous image segmentation. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.485, + 0.469, + 0.526 + ], + "angle": 0, + "content": "[43] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.529, + 0.469, + 0.57 + ], + "angle": 0, + "content": "[44] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.572, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.643, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshop, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.714, + 0.469, + 0.768 + ], + "angle": 0, + "content": "[47] Markus Suchi, Timothy Patten, David Fischinger, and Markus Vincze. Easylabel: A semi-automatic pixel-wise object annotation tool for creating robotic RGB-D datasets. In ICRA, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.771, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.801, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.844, + 0.469, + 0.897 + ], + "angle": 0, + "content": "[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.095, + 0.47, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.095, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[51] Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Chunjing Xu, and Yanwei Fu. Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10707-10717, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[52] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. arXiv preprint arXiv:2104.04691, 2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[53] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[54] Yikai Wang, Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Wei Zhang, and Yanwei Fu. Entity-level text-guided image manipulation. arXiv preprint arXiv:2302.11383, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[55] Yingqian Wang, Longguang Wang, Jungang Yang, Wei An, and Yulan Guo. Flickr1024: A large-scale dataset for stereo image super-resolution. In ICCV Workshops, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[56] Qinrou Wen, Jirui Yang, Xue Yang, and Kewei Liang. Patchdct: Patch refinement for high quality instance segmentation. In ICLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[57] Magnus Wrenninger and Jonas Unger. Synscapes: A photorealistic synthetic dataset for street scene parsing. arXiv preprint arXiv:1810.08705, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[58] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[59] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[61] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In ECCV, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[62] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[63] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACMMM, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.869 + ], + "angle": 0, + "content": "[66] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.095, + 0.892, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4056" + } + ] +] \ No newline at end of file diff --git a/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_origin.pdf b/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a985eb1fe7c667b88c48132dda06b6a72dd4fdab --- /dev/null +++ b/2023/High Quality Entity Segmentation/8c45756a-01dd-4842-b194-1be6a90fce5f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6657d505752f113348ba0313d038786f375258919fe98af245169da9384c363f +size 4858383 diff --git a/2023/High Quality Entity Segmentation/full.md b/2023/High Quality Entity Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8dc5653538d71d59dad355320c85dcf1ab0da10b --- /dev/null +++ b/2023/High Quality Entity Segmentation/full.md @@ -0,0 +1,331 @@ +# High Quality Entity Segmentation + +Lu Qi $^{1*}$ , Jason Kuen $^{2*}$ , Tiancheng Shen $^{4*}$ , Jiumiang Gu $^{2}$ , Wenbo Li $^{4}$ , Weidong Guo $^{3\dagger}$ , Jiaya Jia $^{4}$ , Zhe Lin $^{2}$ , Ming-Hsuan Yang $^{1}$ , $^{1}$ The University of California, Merced $^{2}$ Adobe Research $^{3}$ QQ Brower Lab, Tencent, $^{4}$ The Chinese University of Hong Kong + +![](images/abff55552ee80cb243784d84900bcdac555cd2beafbed9552b936cd0a4ce0506.jpg) +Figure 1: Segmentation results on in-the-wild test images: (1) Mask2Former [7] trained for panoptic segmentation [20] on COCO dataset [31]; (2) Mask2Former trained for entity segmentation [41] on COCO dataset [31]; (3) our CropFormer trained on the proposed EntitySeg Dataset. For a fair comparison, all three models use Swin-Large backbone [33] and are trained until full convergence. Our approach provides far more desirable and useful results for many real-world applications. + +# Abstract + +Dense image segmentation tasks (e.g., semantic, panoptic) are useful for image editing, but existing methods can hardly generalize well in an in-the-wild setting where there are unrestricted image domains, classes, and image resolution & quality variations. Motivated by these observations, we construct a new entity segmentation dataset, with a strong focus on high-quality dense segmentation in the wild. The dataset contains images spanning diverse image domains and entities, along with plentiful high-resolution images and high-quality mask annotations for training and testing. Given the high-quality and -resolution nature of the dataset, we propose CropFormer which is designed + +to tackle the intractability of instance-level segmentation on high-resolution images. It improves mask prediction by fusing high-res image crops that provides more fine-grained image details and the full image. CropFormer is the first query-based Transformer architecture that can effectively fuse mask predictions from multiple image views, by learning queries that effectively associate the same entities across the full image and its crop. With CropFormer, we achieve a significant AP gain of 1.9 on the challenging entity segmentation task. Furthermore, CropFormer consistently improves the accuracy of traditional segmentation tasks and datasets. The dataset and code are released at http://luqi.info-entityv2.github.io/. + +# 1. Introduction + +Entity segmentation [41] is an emerging segmentation task that focuses on open-world and class-agnostic dense image segmentation, designed to have superior generalization capabilities for unseen-category segmentation. As such, it has great potential in image editing applications, where users can interact with visual entities that exist "in the wild". Recent works have leveraged entity segmentation for image manipulation [51, 54] and image retouching [34], where users usually specify in-the-wild text queries (e.g., Oculus, platypus) for local editing. In comparison to traditional closed-set category-based segmentation tasks such as panoptic/instance/semantic segmentation, entity segmentation has the advantage of being all-inclusive and unrestricted, making it easier to find masks that semantically match well with user-specified queries. + +Previous work on entity segmentation [41] has shown promising results, yet it only employed existing panoptic segmentation datasets to verify the feasibility and benefits of entity segmentation. However, existing panoptic segmentation datasets (e.g., COCO [31], Cityscapes [9], ADE20K [65]) suffer from the same issues of closed-set categories that entity segmentation [41] aims to avoid. These datasets are collected and annotated for their original panoptic segmentation task requirements. In view of the lack of a proper entity segmentation dataset, it is difficult to further study and explore on this important task. To address this issue, we present EntitySeg dataset, a new large-scale and high-quality dataset for entity segmentation. + +Apart from in-the-wild generalization, we design EntitySeg to prioritize high-quality mask annotations and high-resolution images, which are essential for user-driven image editing in the current era of high-resolution imagery. Image editing workflows usually require pixel-precise segmentation masks to ensure a smooth user experience. However, existing panoptic segmentation datasets generally contain low-resolution images (e.g., COCO[31] has zero images that are at least 1K) and coarsely-annotated masks, making it difficult for models trained on such datasets to produce accurate mask predictions for higher resolution images. This is why we focus on high-resolution images and high-quality masks in the EntitySeg dataset, which clearly distinguishes it from existing datasets. + +The EntitySeg dataset poses new challenges to existing segmentation methods due to its high-resolution images and high-quality mask annotations. Standard practices of resizing images and mask annotations to 800px used in existing methods [43, 15, 7], result in a heavy loss of fine-grained details. On the other hand, training or evaluating existing segmentation models directly on our high-resolution/quality dataset is intractable due to excessive computation and memory overheads. A naive workaround is to divide the high-resolution image into smaller patches and apply seg + +mentation model to them separately, but it is unable to exploit global context to resolve ambiguity. Moreover, it is unclear how to combine mask predictions from multiple patches using state-of-the-art Transformer-based segmentation methods which rely on set-prediction [7, 6, 50, 61], as shown in Fig 4 and our supplementary material. + +To this end, we introduce CropFormer, a novel approach for tackling the problem of high-resolution entity segmentation. CropFormer learns to dynamically generate tokens to associate the full image with its high-resolution crops and then fuses them to generate the final mask prediction. During training, one of the predefined four corner crops is randomly selected and fused with the full image. For inference, CropFormer is applied to all four corner crops and the full image to achieve complete fusion. The strength of CropFormer lies in its ability to dynamically generate queries to link the same entities in the full image and its higher-resolution crops, allowing us to benefit from both global image context and high-quality image/annotation details to obtain improved results. + +The main contributions of this work are as follows: + +- We present a large-scale, high-quality entity segmentation dataset, comprising of images collected from various domains with diverse resolutions and high-quality masks. Our dataset allows for the evaluation of the generalization and robustness of segmentation models in a unified and open-world manner, due to its domain-diverse and high-resolution setting. +- We benchmark existing segmentation models on our newly collected dataset. We find that the traditional training setting cannot accommodate our dataset, which comprises over $70\%$ images with very high resolution. +- We propose a novel model, CropFormer, with an innovative association module and batch-level decoder to enable the fusion of the full image and its higher-resolution crops. +- Extensive experiments and thorough analysis demonstrate the efficacy of the proposed model and dataset on various segmentation tasks and settings. In Fig. 1, we compare the results from our best CropFormer (trained on EntitySeg) with the best Mask2Formers [7] trained on COCO-Panoptic [20] and COCO-Entity [41]. + +# 2. Related Work + +Image Segmentation Dataset. Numerous datasets have been proposed for semantic, instance, and panoptic segmentation, e.g., Microsoft COCO [31], ADE20K [65], KITTI [12], LVIS [14], PASCAL-VOC [11], OpenImages [23], Cityscapes [9], and so on. In addition, there are some datasets specially designed for specific scenarios, such as amodal segmentation (COCO-Amodal [66], KINS [39]), human segmentation (CelebAMask-HQ [25], + +![](images/8c9e86e1de4e05392c10c8d13360e02ad841f4e13b68c90ee0e5fcc7f46f97de.jpg) +COCO (height: 427px, width: 640px) + +![](images/8dc526205554609af5c87d1a00aac115bc66a9dfc8b0a9333b88a52d955f3978.jpg) +ADE20K (height: 512px, width: 683px) + +![](images/ae77d0b6ad3698bfbceb1084fda15d79b5bcd5e798c715c50d8e3942424a7960.jpg) +Unsplash (height: 4220px, width: 3176px) + +![](images/d0c275b075b8c8857a13a25a089c0735af4e9974840c38205ef33243b160da1d.jpg) +LAION-400M (height: 3637px, width: 2909px) +Figure 2: High-quality mask annotations for low- and high-resolution images collected from existing datasets such as COCO [31], ADE20K [65] and Cityscapes [9] as well as from the internet are presented. For the images collected from the aforementioned datasets, a visual comparison between the original and our annotations is provided in the middle and rightmost sub-figures, where the unannotated regions are shaded in black. It is worth noting that the RGB and mask images shown here have been downsampled, resulting in some quality degradation compared to the original datasets. + +![](images/5184b01d2a1999eae0caab85da41c9397422fe810de2c8ab1214e4ce0b6fea3a.jpg) +Clipart (340, 339) + +LIP [13], MHP [63]) and domain adaptation (Synscapes [57], GTA5 [44]). Despite the significant contributions from these datasets, there is still a need to fulfill the requirements of real-world applications with high-quality images of large diversity. For example, a segmentation model should be robust to high-resolution images from different domains and able to segment unseen objects in the openworld setting. Most relevant to our work is the ADE20K dataset [65], which has large-scale open-vocabulary categories. However, the collected images are of low resolution and from narrowed domains. In contrast, our dataset collects images from multiple domains, including indoor, outdoor, street scenes, cartoon, and remote images, with over $80\%$ of the resolution of the images falling within the high-resolution range of 2000px to 8000px. Differently from the ADE20K [64] and LVIS [14] dataset with a predefined category list, our annotation process first conducts class-agnostic mask annotation on each entity and then labels the category information in an open-vocabulary manner. + +Scene Parsing Methods. Scene parsing methods mainly rely on convolution-based dense prediction or transformer-based query prediction. Explicit localization information, such as bounding box proposals or pixel position, is often used for pixel-level mask prediction in these approaches. Examples include FCN [35], DeepLab [5], PSPNet [62], Mask R-CNN [15], PANet [32], SOLO [53], CondInst [48], PanopticFPN [19], and PanopticFCN [28]. Inspired by the success of transformer-based detection tasks such as DETR [4] [29, 30, 10, 40], a number of methods, such as Max-Deeplab [50] and Mask2Former [7], have been pro + +posed to directly predict segmentation masks without relying on bounding boxes. The key to these methods is to use learnable queries trained by a set-prediction loss to model the implicit location of the potential instance area. However, query-based set-prediction methods cannot effectively perform multi-view fusion, as the query-to-object assignment can drastically change with the change of image view. This is because the set prediction loss typically adopted by such methods does not enforce any assignment constraints across multiple image views. + +# 3. EntitySeg Dataset + +The EntitySeg dataset contains 33,227 images with high-quality mask annotations, as illustrated in Fig. 2. Compared with existing datasets, it has three distinctive features. Firstly, $71.25\%$ and $86.23\%$ of the images are of high quality, with at least $2000\mathrm{px} \times 2000\mathrm{px}$ and $1000\mathrm{px} \times 1000\mathrm{px}$ , respectively, staying in line with the current digital imaging trends. Secondly, our dataset is open-world and not restricted to predefined classes. We consider each semantically-coherent region in the images as an entity, even if it is blurred or hard to recognize semantically. Thirdly, the mask annotation along the boundaries is more precise than those in other datasets. In the following, we will provide an in-depth description of our EntitySeg dataset, as well as an extensive analysis. + +# 3.1. Image Collection + +Inspired by the collection criterion of COCO [31], we collected most non-iconic images with plentiful contextual information and non-canonical viewpoints. To en + +![](images/7c073d643f54a4396a8035f1509c7b87c6fadc96b1c183a21fd434442315a4a4.jpg) +(a) + +![](images/b3f66fd9c6b5940b2013e2191548c1b87a29630c0f8f892aa0b95c0fadc546d0.jpg) +(b) + +![](images/514c4f458271ae9adc8bea39c7abc348a45279afd0f362876ab476853c4eb968.jpg) +(c) + +sure substantial domain diversity, we have sourced the images from various public datasets and the Internet, with permission for academic research. Specifically, part of the images are from COCO [31], ADE20K [65], Pascal VOC [11], Cityscapes [9], Mapillary Vistas [38], ImageNet [22], Open Images [23], DIV2K [3], Flickr2K [55], Clipart [17], Comic [17], DOTA [59], Synscapes [57] and GTA5 [44]. Moreover, we have crawled high-resolution images from Pixabay [1], Unsplash [2], and LAION-400M [46]. As shown in Fig. 3(c), most of the images are from high-resolution sources like Pixabay, Unsplash, and LAION-400M. Moreover, Fig. 3(a) reveals that our dataset has more high-resolution images, with resolutions normally distributed within the range of $0\mathrm{px}$ to $15,000\mathrm{px}$ , in comparison to ADE20K and COCO with images all under $1000\mathrm{px}$ . + +# 3.2. Image Annotation + +For our dataset, the annotation process consists of the following steps. Firstly, entities in the image are annotated with class-agnostic pixel-level masks, ensuring that no overlaps exist between masks. Subsequently, each mask is annotated with a class label in an open vocabulary manner. Two special considerations should be taken into account here: 1) we aonotate the entity as 'unknown' or 'blurred' if it cannot be easily named or is severely blurred; 2) the entity is annotated as a 'supercategory_other' if we only know it is a supercategory but unable to determine its fine-grained class. Our annotation process is the opposite of those of the popular segmentation datasets like COCO [31] and ADE20K [64], which first predefine their class sets before annotating masks based on the predefined classes. + +Table 2 indicates that our dataset has a more considerable proportion of covered image areas and more masks on average compared to COCO- and ADE20K-Panoptic. This is attributed to our novel annotation process that enables us to take all semantically-coherent regions into account and assign class labels accordingly. Besides, our annotation procedure is more similar to the human visual system. As + +![](images/335fa5049313402eea53b520bbfaf7215329d038a2ddc769011a7638ccffb2a4.jpg) +Figure 3: Sub-figures (a) and (b) present the distributions of image resolutions and average number of entities among ADE20K, COCO, and EntitySeg, respectively. Moreover, sub-figure (c) displays the distribution of image sources from which the EntitySeg images were collected. +(a) + +![](images/3a6fba9d53e44bbb944da2b26f2fca4878fc5e44672c760ac7fa89fc7368e879.jpg) +(b) + +Table 1: Annotation consistency of class-agnostic localization and class-aware categories among annotators. + +
DatasetImageRes (avg)↑EntityNum (avg)↑EntityNum (max)↑Entity Complexity↓Entity Simplicity↓Valid Area↑
COCO522.511.2950.7580.5810.891
ADE20K461.313.62550.8020.6060.914
EntitySeg2700.718.12360.7190.5380.999
+ +Table 2: The statistical comparisons between COCO [31], ADE20K [65] and EntitySeg. 'ImageRes (avg)', 'EntityNum (avg)', 'Valid Area', 'Entity Complexity', and 'Entity Simplicity' refer to the average value of image resolution size, entity numbers per image, valid area ratio per image, average entity complexity and simplicity respectively. 'EntityNum (max)' means the maximum per-image number of entities within each dataset. + +evidenced in [36], the human vision system is intrinsically class-agnostic and can recognize entities without comprehending their usage and purpose. + +Annotation consistency is crucial to any human-labeled dataset since it implies whether the annotation task is well-defined and proper. To study this, we randomly selected 500 images from EntitySeg (1.5% of the entire dataset), and asked another two annotators to annotate them after four months since the first round. Table 1(a) shows that the class-agnostic mask mAP in the first step of the annotation process compared to those of the other two annotators. We use one as ground truth and another one as prediction results. In Table 1(b), we also evaluate the category consistency by accuracy (ACC) under the same mask annotations. These two tables demonstrate that our EntitySeg has high annotation consistency in the mask and category labeling stages. + +# 3.3. Dataset Statistics + +In our EntitySeg dataset, we annotate all images at the entity level, regardless of whether they belong to thing or stuff in previous datasets. As shown in Fig. 1, we show + +some low- and high-resolution photos and their mask annotations to highlight our high-quality pixel-level annotation quality. For example, one can zoom in on the Unsplash image to see how high-quality the railing's mask is. + +Table 2 shows the quantitative comparison among our dataset, COCO-Panoptic [31], and ADE20K-Panoptic [65]. In the EntitySeg dataset, each image has 18.1 entities on average, which is more than 11.2 and 13.6 entities in COCO and ADE20K. For a fair comparison, we treat each uncountable stuff region as one segment although it is disconnected. The detailed comparison among those three datasets on the distribution of entity number is shown in Fig. 3(b). Furthermore, the shapes of entities in our dataset are more complex than those COCO [31] and ADE20K [65] as represented by the columns 'Entity Complexity' and 'Entity Simplicity' of Table 2 where an entity with a large convexity and simplicity value means it is a simple shape (and both metrics achieve their maximum value of 1.0 for a circle [66]). More details about how complexity and simplicity are calculated can be found in the supplementary material. + +Class-Aware Annotation. We select a subset of 11,580 images from the dataset for class labeling, for investigating the diversity of semantic classes in our dataset. In our class annotation process, the class is labeled in an open-vocabulary and free-form manner, then the class set is continually expanded. After the class annotation process, we organize and merge the annotated classes according to Word-Tree [45]. This results in 535 things and 109 stuff classes for those 11,580 images that we use to construct the class-aware EntityClass dataset. Please refer to our supplementary file for more analysis on EntityClass dataset. + +# 4. CropFormer + +We propose CropFormer which is the first Transformer-based set-prediction segmentation method that is capable of fusing the mask predictions from multiple image views. Inspired by existing methods [8, 7, 50, 61], we design CropFormer to learn $N$ number of $K$ -dimensional queries $\mathbf{Q} \in \mathbb{R}^{N \times K}$ to generate mask embeddings $\mathbf{E} \in \mathbb{R}^{N \times 1 \times 1 \times 1 \times K}$ . $\mathbf{E}$ then acts as convolution filters that are applied to pixel-level mask features $\mathbf{P}_2 \in \mathbb{R}^{T \times H \times W}$ to generate $N$ segmentation masks $\mathbf{U}^{\mathbf{m}} \in \mathbb{R}^{N \times T \times H \times W}$ , where $T, H$ and $W$ correspond to the image view, height, and width dimensions. However, such a rudimentary design can only work well with single-view inputs and cannot effectively fuse the results from multiple views. Here, we introduce a novel association module and batch-level decoder to achieve the goal of CropFormer, which is to exploit global context from full image and fine-grained local details from crops for high-quality segmentation. CropFormer is illustrated in Fig. 4. + +Crop Dataloger. In CropFormer framework, the crop dataloger is designed to generate a batch of images that si + +multaneously include the full images and their corresponding crops. There are two steps in our dataloader: crop and resize. First, we augment the full image $I^o$ with a crop that is part of the full image. The crop is randomly extracted from one of the fixed image corners: upper-left, upper-right, bottom-left, bottom-right. The crop size is controlled by a fixed ratio hyperparameter $\delta \in \mathbb{R}$ relative to the full image size. Then, we resize both the full image and crop to the same size. In this way, we construct an input tensor $\mathbf{I} \in \mathbb{R}^{1 \times 2 \times H' \times W' \times 3}$ with 2 elements (full image & 1 crop) along the view dimension. Compared to the full image, our corner crops preserve more image details (e.g., local information) that are useful for high-quality segmentation. Specifically, given an dataloader $\mathcal{D}$ , we define the data preprocessing as: + +$$ +\mathbf {I} = \left\{I ^ {o}, I ^ {c} \right\}, \quad I ^ {o} \sim \mathcal {D} \tag {1} +$$ + +where $I^o$ is sampled from $\mathcal{D}$ , and $I^c$ is the cropped image controlled by fixed ratio $\delta$ . Unlike the random cropping in [58, 16], we adopt a fixed cropping solution to keep the training and inference consistent. Such a rigid cropping strategy provides better inference efficiency since, rather than bunches of random crops, evaluating a fixed number of crops during inference is sufficient enough. + +Image Encoder and Decoder. Based on learnable image-level queries $\mathbf{Q}_i\in \mathbb{R}^{N\times K}$ , we use image encoder $\Theta$ and image-level decoder $\Phi_{i}$ to generate image-level embeddings $\mathbf{E_i}\in \mathbb{R}^{N\times 2\times 1\times 1\times K}$ for the full input image and its crop. Given the input tensor (I) and queries $(\mathbf{Q}_i)$ , we define $\mathbf{E_i}$ as follows: + +$$ +\mathbf {E} _ {\mathbf {i}} = \Phi_ {i} \left(\mathbf {Q} _ {\mathbf {i}}, \Theta (\mathbf {I})\right), \tag {2} +$$ + +where $\Phi_i(\cdot)$ is a Transformer-based image-level decoder. $\mathbf{E_i}$ is used for image-level entityness prediction and pixellevel prediction using the low-level image feature $\mathbf{P_2}$ (derived from the image encoder). The prediction process can be formulated as: + +$$ +\mathbf {U} _ {\mathbf {i}} ^ {\mathrm {e}}, \mathbf {U} _ {\mathbf {i}} ^ {\mathrm {m}} = \operatorname {P r e d H e a d} _ {\mathbf {i}} \left(\mathbf {E} _ {\mathbf {i}}, \mathbf {P} _ {\mathbf {2}}\right), \tag {3} +$$ + +where $\mathbf{U}_{\mathrm{i}}^{\mathrm{e}}$ and $\mathbf{U}_{\mathrm{i}}^{\mathrm{m}}$ denote the entityness prediction and pixel-level mask outputs. Here, we employ $i$ subscript to differentiate image-level embeddings and mask outputs from those of the association module. + +Association Module. Following [6], our association module aims to generate batch queries $\mathbf{Q}_{\mathrm{b}}$ that are fully shared by the full image and its crop to represent the same entities consistently. However, instead of treating $\mathbf{Q}_{\mathrm{b}}$ as learnable network parameters, considering that $\mathbf{E}_{\mathrm{i}}$ already contains strong segmentation features, we generate $\mathbf{Q}_{\mathrm{b}}$ directly from $\mathbf{E}_{\mathrm{i}} = \{\mathbf{E}_{I^o},\mathbf{E}_{I^c}\}$ . Particularly, we use a Transformer architecture with cross-attention ( $f_{\mathrm{XAtt}}$ ) and self + +![](images/ea2028887fe9c83e2cdf29b95dc6a8252d85cfa04898818ff8d0ec1fdcfb0b7b.jpg) +Figure 4: Framework of the proposed CropFormer. The red box indicates cropped region randomly sampled from four fixed image corners. In image-level prediction, the same entity across different image views may be assigned to different queries. Our association module and batch decoder can effectively associate the same entity across different views with a single query. + +attention $(f_{\mathrm{SAtt}})$ to obtain $\mathbf{Q}_{\mathbf{b}}$ .. + +$$ +\mathbf {Q} _ {\mathbf {b}} = \operatorname {F F N} \left(f _ {\mathrm {S A t t}} \left(f _ {\mathrm {X A t t}} \left(\underbrace {f _ {\mathrm {q}} \left(\mathbf {E} _ {I ^ {o}}\right)} _ {\text {q u e r y}}, \underbrace {f _ {\mathrm {k}} \left(\mathbf {E} _ {\mathbf {i}}\right)} _ {\text {k e y}}, \underbrace {f _ {\mathrm {v}} \left(\mathbf {E} _ {\mathbf {i}}\right)} _ {\text {v a l u e}}\right)\right)\right), \tag {4} +$$ + +where FFN is a feed-forward network, $f_{\{\mathbf{q}, \mathbf{k}, \mathbf{v}\}}(\cdot)$ are linear transformations. Considering the importance of full image for entity segmentation, inspired by [49, 26] we take the image-level embeddings of full image $\mathbf{E}_{I^o}$ as query, while treating all image-level embeddings $\mathbf{E}_i$ as key and value. + +Batch-level Decoder. Given $\mathbf{Q}_{\mathbf{b}} \in \mathbb{R}^{N \times 1 \times 1 \times 1 \times K}$ , we obtain batch embeddings $\mathbf{E}_{\mathbf{b}} \in \mathbb{R}^{N \times 1 \times 1 \times 1 \times K}$ : + +$$ +\mathbf {E} _ {\mathbf {b}} = \Phi_ {b} (\mathbf {Q} _ {\mathbf {b}}, \Theta (\mathbf {I})), \tag {5} +$$ + +where $\Phi_b(\cdot)$ denotes the batch-level decoder. Thus, we broadcast $\mathbf{E}_{\mathbf{b}}$ to the shape of $N\times 2\times 1\times 1\times K$ to share it between the full image and its crop. Finally, the broadcasted $\mathbf{E}_{\mathbf{b}}$ and low-level batch features $\mathbf{P_2}$ are used for batch-level mask classification and pixel-wise prediction: + +$$ +\mathbf {U} _ {\mathrm {b}} ^ {\mathrm {e}}, \mathbf {U} _ {\mathrm {b}} ^ {\mathrm {m}} = \operatorname {P r e d H e a d} _ {\mathrm {b}} \left(\mathbf {E} _ {\mathrm {b}}, \mathbf {P} _ {2}\right), \tag {6} +$$ + +For the image-level decoder, the queries are separate for each image/crop. Whereas, for the batch-level decoder, the queries are shared between the image and its crop. We illustrate the differences of the image- and batch-level decoders in Fig. 5, where we regard two mage views as an entirety. + +Training. During training, we employ two separate losses $\mathcal{L}_i$ and $\mathcal{L}_b$ for image- and batch-level predictions, with respect to ground truth $\mathbf{G}$ . The main difference between the two losses lies in whether the same entities in full image and crop are tied to the same queries or not. + +The overall training loss for our model is defined as: + +$$ +\begin{array}{l} \mathcal {L} = \sum_ {\mathbf {k} \in \{\mathbf {i}, \mathbf {b} \}} \mathcal {L} _ {\mathbf {k}} ^ {\mathrm {c e}} \left(\mathbf {U} _ {\mathbf {k}} ^ {\mathrm {e}}, \mathbf {G} _ {\mathbf {k}} ^ {\mathrm {e}}\right) + \sum_ {\mathbf {k} \in \{\mathbf {i}, \mathbf {b} \}} \mathcal {L} _ {\mathbf {k}} ^ {\mathrm {b c e}} \left(\mathbf {U} _ {\mathbf {k}} ^ {\mathrm {m}}, \mathbf {G} _ {\mathbf {k}} ^ {\mathrm {m}}\right) + \\ \sum_ {\mathbf {k} \in \{i, b \}} \mathcal {L} _ {\mathbf {k}} ^ {\mathrm {d i c e}} \left(\mathbf {U} _ {\mathbf {k}} ^ {\mathrm {m}}, \mathbf {G} _ {\mathbf {k}} ^ {\mathrm {m}}\right), \tag {7} \\ \end{array} +$$ + +![](images/22fe4a5893d2414a865955c60b45a292367c6d822fb2e86860b5fe4745ab2811.jpg) +Figure 5: The illustration of image- and batch-level decoder and association module. + +![](images/25ddc2919141a5656a1bbc3e7648bb011049428a03e41ca3796e5ab54312478d.jpg) + +![](images/d293bd3770dd7d3a51207b77f08a3d5ffab0105e1d67e57ccbeea78403fef5d1.jpg) + +where $\mathcal{L}_{\mathrm{i}}^{\mathrm{ce}}$ and $\mathcal{L}_{\mathrm{b}}^{\mathrm{ce}}$ denote binary cross-entropy loss for image- and batch-level entityness prediction. Similarly, $\mathcal{L}_{\mathrm{i}}^{\mathrm{bce}}$ , $\mathcal{L}_{\mathrm{b}}^{\mathrm{bce}}$ , $\mathcal{L}_{\mathrm{i}}^{\mathrm{dice}}$ and $\mathcal{L}_{\mathrm{b}}^{\mathrm{dice}}$ denote the binary cross-entropy and dice loss for image- and batch-level mask prediction. + +Inference. In CropFormer, the same entities between the full image and its crop are represented by the same queries by virtue of the association module. For the final segmentation output, we fuse the per-pixel mask predictions obtained from the full image and every crop of the 4 corners by using an average operation. The confidence score of each entity comes from the batch-level entityness prediction score. + +# 5. Experiments + +We conduct our main experiments on the entity segmentation with proposed EntitySeg Dataset in this work. First, we benchmark our dataset in several segmentation tasks, including entity, instance, semantic and panoptic segmentation. Then, we ablate our proposed CropFormer to demonstrate its effectiveness. Note that, in Sec. 4, we omit the conventional data batch dimension to improve clarity, but in practice, we train CropFormer with a batch of multiple dataset images and their crops, following existing methods. + +Training Details. We leave out some details like the training schedule and some ablation studies due to paper length limitations (see supplementary for details). + +
TaskMetricsMethodBackbonePretrainPerformance
SEMmIoUDeeplabv3R-50ImageNet27.9
Mask2FormerR-50ImageNet37.8
R-50COCO43.3
Swin-TCOCO45.0
Swin-LCOCO50.5
INSAPMask-RCNNR-50ImageNet5.0
Mask2FormerR-50ImageNet13.0
R-50COCO20.3
Swin-TCOCO22.7
Swin-LCOCO30.3
PANPQPanopticFPNR-50ImageNet3.6
Mask2FormerR-50ImageNet5.5
R-50COCO9.6
Swin-TCOCO9.8
Swin-LCOCO13.4
ENTAPeMask-RCNNR-50ImageNet24.9
Mask2FormerR-50ImageNet26.0
R-50COCO35.2
Swin-TCOCO42.9
Swin-LCOCO46.2
+ +Table 3: Benchmark on several segmentation tasks on the EntitySeg Dataset. The 'SEM', 'INS', 'PAN', and 'ENT' indicate the task of semantic, instance, panoptic and entity segmentation. We note that we leave out all details and more experiments in the supplementary file due to paper length limitations. + +
Transfer DatasetTypeSEM (mIoU)INS (AP)PAN (PQ)ENT (APe)
ADE20KImage57.2(+1.0)36.0(+1.1)49.0(+0.9)40.6(+1.5)
YVIS2019 [60]Video-61.3(+0.9)--
VIPSeg [37]Video--37.0(+1.1)-
UVO [52]Video---29.5(+2.2)
+ +# 5.1. Dataset Benchmarks + +Benchmark on EntitySeg Dataset. In Table 3, we provide results on semantic, instance, panoptic, and entity segmentation on EntitySeg dataset with some baseline methods. Similar to existing segmentation datasets like COCO and ADE20K, the use of COCO pre-trained weight and a stronger backbone can benefit each segmentation performance of EntitySeg Dataset. Compared to COCO and ADE20K, EntitySeg dataset is more challenging as evidenced by the smaller performance numbers (e.g., Swin-L Mask2Former's PQ on COCO-PAN is $57.8\%$ ). + +Transferring EntitySeg to other general segmentation tasks. Table 4 shows the performance comparison of transferring models pretrained on either COCO or EntitySeg dataset to other image/video-level segmentation tasks. Owing to its emphasis on in-the-wild generalization, EntitySeg provides greater improvements than COCO dataset. + +Benefit of EntitySeg for high-quality image segmentation. In Table 5, we transfer Mask2Former pretrained on EntitySeg dataset for finetuning on Dichotomous Image Segmentation [42]. Compared to state-of-the-art ISNet [42] + +Table 4: The relative performance gains of EntitySeg pretraining over COCO-Panoptic pretraining when fine-tuned on various image/video-level segmentation datasets. + +
MethodFβmx↑Fβw↑M↓Sα↑Eφm↑HCEγ↓
IS-Net [42].791.717.074.813.8561116
Ours.872.789.035.896.913969
+ +Table 5: Evaluation on Dichotomous Image Segmentation (DIS5K dataset [42].) 'Ours' indicates Mask2Former pretrained on EntitySeg and then finetuned on DIS5K. + +
Train DatasetTaskLVIS [14]OCID [47]CAMO [24]FSS [27]Avg. AR
COCOPAN24.3839.9443.8064.9643.27
COCOENT27.7545.8859.7075.6052.23
SAM-H [21]Prompt49.2555.1775.3373.6063.34
EntitySegENT33.5676.6373.2086.2667.41
+ +Table 6: In-the-wild segmentation performance (AR@100 with IoU threshold 0.5). We directly evaluate models trained on either EntitySeg or COCO on large-vocabulary segmentation LVIS [14], object clutter indoor segmentation for robot grab [47], camouflaged instance segmentation CAMO [24], few-shot segmentation FSS [27]. All the models are based on Mask2Former with Swin-Large backbone. + +that is carefully designed for the task, Mask2Former pretrained on EntitySeg outperforms ISNet by a large margin, despite not having any task-specific design. This is by virtue of the high-quality mask annotations in EntitySeg dataset. + +Segmentation 'in the wild' Table 6 compares in-the-wild segmentation performance between Mask2Formers trained on either EntitySeg or COCO. Here, we directly use the trained models without finetuning on the evaluated datasets. The EntitySeg outperforms COCO-Panoptic [20] and COCO-Entity [41] by large margins. + +# 5.2. A Comprehensive Analysis of CropFormer + +We first ablate Cropformer on the entity segmentation task and then report results on other segmentation datasets. + +Table 7 shows the benefits of CropFormer on segmentation performance and computation cost compared to the Mask2Former [7], its multi-scale extension, as well as related methods [6, 16] that supports multi-view fusion. The first four rows are our baseline Mask2Former with the single-scale inference (800, 1040, or 2700 for shortest side) and (1333, 1732 or 4500 for longest side) for the full images, where $1040 = 800 \times \delta$ and $1732 = 1333 \times \delta$ with $\delta = 0.7$ by default. For the results in the fifth row, we use test-time Hungarian matching to associate the same entities obtained with multi-scale images. We find minimal improvement with such an inference strategy. In the sixth and seventh rows, directly using video-level Mask2Former or VITA, the state-of-the-art video instance segmentation framework, merely brings marginal performance gains. Whereas, using the crop output from CropFormer's batch decoder achieves a significant $\mathrm{AP^e}$ gain as indicated by the last row. By combining the full image and four crop outputs from CropFormer (final row), we obtain even stronger $1.5\mathrm{AP^e}$ gains compared to the baseline (first row). Compared to the + +
MethodInput PreprocessingGPU Type\(AP^e\)GPU Memory (Train)GPU Memory (Inference)Train Time (GPU day)Inference Time (ms)
Mask2Formerhigh-res image \(\xrightarrow{\text{resize}} (800, 1333)\)A100-40G39.55.8G3.0G2.4637
Mask2Formerhigh-res image \(\xrightarrow{\text{resize}} (2700, 4500)\)A100-40GOOMOOM-OOM-
Mask2Formerhigh-res image \(\xrightarrow{\text{resize}} (2700, 4500)\)A100-80G40.958.8G10.4G60.83312
V-Mask2Former [6]\(\xrightarrow{\text{resize}} (800, 1333)\)A100-40G39.710.4G6.7G6.81503
VITA [16]\(\xrightarrow{\text{resize}} (800, 1333)\)A100-40G39.911.9G6.8G10.21512
CropFormer\(\xrightarrow{\text{crop}(\delta)} (800, 1333)\)A100-40G41.011.4G6.9G9.81503
+ +Table 7: Comparison with high-resolution methods on EntitySeg with multi-scale training. $(\mathrm{x},\mathrm{y})$ indicates number of pixels along the shortest (x) and longest (y) sides of image. Row 1, 2, & 3 are single-scale Mask2Formers trained and evaluated at various single-scale resolutions. (V-Mask2Former & VITA) mask predictions from full image and four crops are fused with either video-level Mask2Former or VITA. + +
Iterations (Train)Mask2FormerCropFormerDecoderAPeAP50AP75
1×39.541.0 (+1.5)I-O39.356.739.8
2×40.242.1 (+1.9)B-O39.156.639.7
3×40.942.8 (+1.9)B-C40.257.540.8
4×40.942.7 (+1.8)B-OC41.058.441.9
(a) metric: APe(b)
+ +Mask2Former baselines with a naive increase in image resolution, CropFormer is far more efficient computationally. + +Longer Training Schedules Table 8(a) indicates that CropFormer consistently improves over the baseline Mask2Former when we adopt $2 \times /3 \times$ training schedules. A similar trend is observed when using Swin-L [33] backbone, which is provided in our supplementary file. + +Decoder Performance Table 8(b) shows the impact of using different decoder combinations in CropFormer for mask prediction during inference. Using just the crop outputs from CropFormer's batch decoder (B-C) outperforms using just the full image view (I-O, B-O). Combining both the full image and crops (B-OC) gives the best result. + +CropDataloger Table 9(a) and (b) show the ablation study on the usage of crop ratio and crop type. As indicated by Table 9(a), CropFormer obtains the best performance with a crop ratio $\delta$ of 0.7. Smaller crop ratios perform worse and it might be a problem caused by queries that are not robust against drastic image transformation. + +Association Module Table 9(c) shows the ablation study on the structure of the association module. There is a slight performance difference between using self-attention or a feedforward module following cross-attention. + +Comparison to other high-quality segmentation methods. Table 10(a) shows that our CropFormer significantly outperforms recent high-quality instance-level segmentation methods like Mask Transfiner [18] and PatchDCT [56] on both traditional and boundary APs. + +Table 8: Ablation study on: (a) training schedules; (b) CropFormer's decoder combinations. In sub-table(b), the 'Decoder' column indicates whether we use the predictions of the full image ('O'), four cropped patches ('C'), or both of them ('OC') from the 'Image' or 'Batch' decoder. + +
δAPeTrainTestAPeXAttSAttFFNAPe
0.538.5RandomFixed (4)39.7✓○○40.7
0.640.2Fixed (4)Fixed (4)41.0✓○✓40.8
0.741.0Fixed (4)Fixed (8)41.3✓✓○40.8
0.840.9Fixed (8)Fixed (8)41.0✓✓✓41.0
(a)(b)(c)
+ +Table 9: Ablation study on the usage of crop ratio $\delta$ , crop type and association module in CropFormer. In sub-table (b), 'Random' indicates random crops and 'Fixed (4/8)' indicates 4 or 8 fixed corner crops. In sub-table(c), $\checkmark$ and $\circ$ means whether we use the module or not. + +
MethodAPeAPbTaskMask2FormerCropFormer
Mask Transfiner [18]33.725.0COCO-INS50.151.9
PatchDCT [56]35.427.4COCO-PAN57.859.1
Mask2Former39.531.2ADE20K-INS34.936.4
Ours41.033.7ADE20K-PAN48.149.7
(a)(b) metrics: AP for INS; PQ for PAN
+ +Table 10: (a) Comparison with high-quality segmentation methods on EntitySeg task/dataset using $\mathrm{AP}^e$ and boundary AP $(\mathrm{AP}^b)$ . (b) evaluation on instance (INS) and panoptic (PAN) segmentation tasks on COCO and ADE20K datasets. + +Improvement on class-aware segmentation tasks Table 10(b) shows the class-aware task performance of CropFromer with Swin-L [33] backbone and $3\times$ training schedule. CropFormer can still achieve competitive gains on the instance and panoptic segmentation tasks on COCO and ADE20K datasets. This suggests that CropFormer may provide other benefits that are beyond high-resolution segmentation (e.g., prediction robustness). + +# 6. Conclusion + +This paper proposes an EntitySeg dataset and a CropFormer framework for high-quality entity segmentation, with a strong focus on in-the-wild generalization and high-quality dense segmentation. The EntitySeg dataset contains about 33K images from various domains and resolutions, accompanied by high-quality mask annotations for training and testing. A novel CropFormer framework is also proposed to overcome the computation challenges introduced by EntitySeg dataset's high-quality and high-resolution characteristics. Specifically, it enables fusing multi-view predictions from a low-res full image and its high-res crops by learning batch-level queries. We hope this work will serve as a catalyst for future research on open-world and high-quality image segmentation tasks. + +# References + +[1] Pixabay. https://pixabay.com. Accessed: 2022-04-17. 4 +[2] Unsplash. https://unsplash.com. Accessed: 2022-04-17. 4 +[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPR Workshops, 2017. 4 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI, 2017. 3 +[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2, 5, 7, 8 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 3, 5, 7 +[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 5 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4 +[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 3 +[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 2010. 2, 4 +[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2 +[13] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In CVPR, 2017. 3 +[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 2, 3, 7 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2, 3 +[16] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurlPS, 2022. 5, 7, 8 +[17] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, 2018. 4 + +[18] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 8 +[19] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 3 +[20] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019. 1, 2, 7 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In arXiv, 2023. 7 +[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 4 +[23] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2, 4 +[24] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in the-wild: Dataset, method, and benchmark suite. TIP, 2021. 7 +[25] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2 +[26] Peizhao Li, Jiuming Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. Selfdoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 6 +[27] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 7 +[28] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. TPAMI, 2022. 3 +[29] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3 +[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 3 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 2, 3, 4, 5 +[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018. 3 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 8 + +[34] Zerun Liu, Fan Zhang, Jingxuan He, Jin Wang, Zhangye Wang, and Lechao Cheng. Text-guided mask-free local image retouching. arXiv preprint arXiv:2212.07603, 2022. 2 +[35] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3 +[36] David Marr. Vision: A computational investigation into the human representation and processing of visual information. San Francisco: W.H. Freeman. Print., 1982. 4 +[37] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 7 +[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 4 +[39] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2 +[40] Lu Qi, Jason Kuen, Jiumiang Gu, Zhe Lin, Yi Wang, Yukang Chen, Yanwei Li, and Jiaya Jia. Multi-scale aligned distillation for low-resolution detection. In CVPR, 2021. 3 +[41] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. TPAMI, 2022. 1, 2, 7 +[42] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, et al. Highly accurate dichotomous image segmentation. In ECCV, 2022. 7 +[43] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2 +[44] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3, 4 +[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 5 +[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshop, 2021. 4 +[47] Markus Suchi, Timothy Patten, David Fischinger, and Markus Vincze. Easylabel: A semi-automatic pixel-wise object annotation tool for creating robotic RGB-D datasets. In ICRA, 2019. 7 +[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 3 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6 +[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021. 2, 3, 5 + +[51] Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Chunjing Xu, and Yanwei Fu. Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10707-10717, 2022. 2 +[52] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. arXiv preprint arXiv:2104.04691, 2021.7 +[53] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, 2020. 3 +[54] Yikai Wang, Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Wei Zhang, and Yanwei Fu. Entity-level text-guided image manipulation. arXiv preprint arXiv:2302.11383, 2023. 2 +[55] Yingqian Wang, Longguang Wang, Jungang Yang, Wei An, and Yulan Guo. Flickr1024: A large-scale dataset for stereo image super-resolution. In ICCV Workshops, 2019. 4 +[56] Qinrou Wen, Jirui Yang, Xue Yang, and Kewei Liang. Patchdct: Patch refinement for high quality instance segmentation. In ICLR, 2023. 8 +[57] Magnus Wrenninger and Jonas Unger. Synscapes: A photorealistic synthetic dataset for street scene parsing. arXiv preprint arXiv:1810.08705, 2018. 3, 4 +[58] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 5 +[59] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, 2018. 4 +[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 7 +[61] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In ECCV, 2022. 2, 5 +[62] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 3 +[63] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACMMM, 2018. 3 +[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 3, 4 +[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 3, 4, 5 +[66] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2, 5 \ No newline at end of file diff --git a/2023/High Quality Entity Segmentation/images.zip b/2023/High Quality Entity Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..586402af1748987ecbbe89b56e80f50affc48fdf --- /dev/null +++ b/2023/High Quality Entity Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a07bdbdcd34b2345333ece8f603b70d42311dca60b815eabddda868ee9ce75 +size 603299 diff --git a/2023/High Quality Entity Segmentation/layout.json b/2023/High Quality Entity Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e6cbbc3f315d1baa30274b1034aa795bedd9fc59 --- /dev/null +++ b/2023/High Quality Entity Segmentation/layout.json @@ -0,0 +1,9399 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 190, + 105, + 402, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 105, + 402, + 122 + ], + "spans": [ + { + "bbox": [ + 190, + 105, + 402, + 122 + ], + "type": "text", + "content": "High Quality Entity Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "spans": [ + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": "Lu Qi" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Jason Kuen" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Tiancheng Shen" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Jiumiang Gu" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Wenbo Li" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Weidong Guo" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Jiaya Jia" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Zhe Lin" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", Ming-Hsuan Yang" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": "The University of California, Merced " + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": "Adobe Research " + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": "QQ Brower Lab, Tencent, " + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 109, + 144, + 475, + 208 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 54, + 262, + 541, + 475 + ], + "blocks": [ + { + "bbox": [ + 54, + 262, + 541, + 475 + ], + "lines": [ + { + "bbox": [ + 54, + 262, + 541, + 475 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 541, + 475 + ], + "type": "image", + "image_path": "abff55552ee80cb243784d84900bcdac555cd2beafbed9552b936cd0a4ce0506.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 475, + 546, + 522 + ], + "lines": [ + { + "bbox": [ + 47, + 475, + 546, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 546, + 522 + ], + "type": "text", + "content": "Figure 1: Segmentation results on in-the-wild test images: (1) Mask2Former [7] trained for panoptic segmentation [20] on COCO dataset [31]; (2) Mask2Former trained for entity segmentation [41] on COCO dataset [31]; (3) our CropFormer trained on the proposed EntitySeg Dataset. For a fair comparison, all three models use Swin-Large backbone [33] and are trained until full convergence. Our approach provides far more desirable and useful results for many real-world applications." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 144, + 531, + 192, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 531, + 192, + 544 + ], + "spans": [ + { + "bbox": [ + 144, + 531, + 192, + 544 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 551, + 289, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 289, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 289, + 694 + ], + "type": "text", + "content": "Dense image segmentation tasks (e.g., semantic, panoptic) are useful for image editing, but existing methods can hardly generalize well in an in-the-wild setting where there are unrestricted image domains, classes, and image resolution & quality variations. Motivated by these observations, we construct a new entity segmentation dataset, with a strong focus on high-quality dense segmentation in the wild. The dataset contains images spanning diverse image domains and entities, along with plentiful high-resolution images and high-quality mask annotations for training and testing. Given the high-quality and -resolution nature of the dataset, we propose CropFormer which is designed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "type": "text", + "content": "to tackle the intractability of instance-level segmentation on high-resolution images. It improves mask prediction by fusing high-res image crops that provides more fine-grained image details and the full image. CropFormer is the first query-based Transformer architecture that can effectively fuse mask predictions from multiple image views, by learning queries that effectively associate the same entities across the full image and its crop. With CropFormer, we achieve a significant AP gain of 1.9 on the challenging entity segmentation task. Furthermore, CropFormer consistently improves the accuracy of traditional segmentation tasks and datasets. The dataset and code are released at http://luqi.info-entityv2.github.io/." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 702, + 234, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 234, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 234, + 712 + ], + "type": "text", + "content": "*Equal contribution. † indicates corresponding author." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4047" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 128, + 85 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 92, + 289, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 92, + 289, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 289, + 272 + ], + "type": "text", + "content": "Entity segmentation [41] is an emerging segmentation task that focuses on open-world and class-agnostic dense image segmentation, designed to have superior generalization capabilities for unseen-category segmentation. As such, it has great potential in image editing applications, where users can interact with visual entities that exist \"in the wild\". Recent works have leveraged entity segmentation for image manipulation [51, 54] and image retouching [34], where users usually specify in-the-wild text queries (e.g., Oculus, platypus) for local editing. In comparison to traditional closed-set category-based segmentation tasks such as panoptic/instance/semantic segmentation, entity segmentation has the advantage of being all-inclusive and unrestricted, making it easier to find masks that semantically match well with user-specified queries." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 272, + 289, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 289, + 426 + ], + "type": "text", + "content": "Previous work on entity segmentation [41] has shown promising results, yet it only employed existing panoptic segmentation datasets to verify the feasibility and benefits of entity segmentation. However, existing panoptic segmentation datasets (e.g., COCO [31], Cityscapes [9], ADE20K [65]) suffer from the same issues of closed-set categories that entity segmentation [41] aims to avoid. These datasets are collected and annotated for their original panoptic segmentation task requirements. In view of the lack of a proper entity segmentation dataset, it is difficult to further study and explore on this important task. To address this issue, we present EntitySeg dataset, a new large-scale and high-quality dataset for entity segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 426, + 289, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 289, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 289, + 593 + ], + "type": "text", + "content": "Apart from in-the-wild generalization, we design EntitySeg to prioritize high-quality mask annotations and high-resolution images, which are essential for user-driven image editing in the current era of high-resolution imagery. Image editing workflows usually require pixel-precise segmentation masks to ensure a smooth user experience. However, existing panoptic segmentation datasets generally contain low-resolution images (e.g., COCO[31] has zero images that are at least 1K) and coarsely-annotated masks, making it difficult for models trained on such datasets to produce accurate mask predictions for higher resolution images. This is why we focus on high-resolution images and high-quality masks in the EntitySeg dataset, which clearly distinguishes it from existing datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 594, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 289, + 713 + ], + "type": "text", + "content": "The EntitySeg dataset poses new challenges to existing segmentation methods due to its high-resolution images and high-quality mask annotations. Standard practices of resizing images and mask annotations to 800px used in existing methods [43, 15, 7], result in a heavy loss of fine-grained details. On the other hand, training or evaluating existing segmentation models directly on our high-resolution/quality dataset is intractable due to excessive computation and memory overheads. A naive workaround is to divide the high-resolution image into smaller patches and apply seg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 74, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 74, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 74, + 545, + 145 + ], + "type": "text", + "content": "mentation model to them separately, but it is unable to exploit global context to resolve ambiguity. Moreover, it is unclear how to combine mask predictions from multiple patches using state-of-the-art Transformer-based segmentation methods which rely on set-prediction [7, 6, 50, 61], as shown in Fig 4 and our supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 146, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 545, + 312 + ], + "type": "text", + "content": "To this end, we introduce CropFormer, a novel approach for tackling the problem of high-resolution entity segmentation. CropFormer learns to dynamically generate tokens to associate the full image with its high-resolution crops and then fuses them to generate the final mask prediction. During training, one of the predefined four corner crops is randomly selected and fused with the full image. For inference, CropFormer is applied to all four corner crops and the full image to achieve complete fusion. The strength of CropFormer lies in its ability to dynamically generate queries to link the same entities in the full image and its higher-resolution crops, allowing us to benefit from both global image context and high-quality image/annotation details to obtain improved results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 314, + 523, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 314, + 523, + 325 + ], + "spans": [ + { + "bbox": [ + 317, + 314, + 523, + 325 + ], + "type": "text", + "content": "The main contributions of this work are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 327, + 544, + 588 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 327, + 544, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 544, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 544, + 409 + ], + "type": "text", + "content": "- We present a large-scale, high-quality entity segmentation dataset, comprising of images collected from various domains with diverse resolutions and high-quality masks. Our dataset allows for the evaluation of the generalization and robustness of segmentation models in a unified and open-world manner, due to its domain-diverse and high-resolution setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 410, + 544, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 410, + 544, + 468 + ], + "spans": [ + { + "bbox": [ + 317, + 410, + 544, + 468 + ], + "type": "text", + "content": "- We benchmark existing segmentation models on our newly collected dataset. We find that the traditional training setting cannot accommodate our dataset, which comprises over " + }, + { + "bbox": [ + 317, + 410, + 544, + 468 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 317, + 410, + 544, + 468 + ], + "type": "text", + "content": " images with very high resolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 469, + 544, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 469, + 544, + 516 + ], + "spans": [ + { + "bbox": [ + 317, + 469, + 544, + 516 + ], + "type": "text", + "content": "- We propose a novel model, CropFormer, with an innovative association module and batch-level decoder to enable the fusion of the full image and its higher-resolution crops." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 517, + 544, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 517, + 544, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 517, + 544, + 588 + ], + "type": "text", + "content": "- Extensive experiments and thorough analysis demonstrate the efficacy of the proposed model and dataset on various segmentation tasks and settings. In Fig. 1, we compare the results from our best CropFormer (trained on EntitySeg) with the best Mask2Formers [7] trained on COCO-Panoptic [20] and COCO-Entity [41]." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 601, + 392, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 392, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 392, + 613 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 713 + ], + "type": "text", + "content": "Image Segmentation Dataset. Numerous datasets have been proposed for semantic, instance, and panoptic segmentation, e.g., Microsoft COCO [31], ADE20K [65], KITTI [12], LVIS [14], PASCAL-VOC [11], OpenImages [23], Cityscapes [9], and so on. In addition, there are some datasets specially designed for specific scenarios, such as amodal segmentation (COCO-Amodal [66], KINS [39]), human segmentation (CelebAMask-HQ [25]," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4048" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 306, + 135 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 306, + 135 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 306, + 135 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 306, + 135 + ], + "type": "image", + "image_path": "8c9e86e1de4e05392c10c8d13360e02ad841f4e13b68c90ee0e5fcc7f46f97de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 136, + 246, + 145 + ], + "lines": [ + { + "bbox": [ + 130, + 136, + 246, + 145 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 246, + 145 + ], + "type": "text", + "content": "COCO (height: 427px, width: 640px)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 530, + 135 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 530, + 135 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 530, + 135 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 530, + 135 + ], + "type": "image", + "image_path": "8dc526205554609af5c87d1a00aac115bc66a9dfc8b0a9333b88a52d955f3978.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 136, + 480, + 145 + ], + "lines": [ + { + "bbox": [ + 356, + 136, + 480, + 145 + ], + "spans": [ + { + "bbox": [ + 356, + 136, + 480, + 145 + ], + "type": "text", + "content": "ADE20K (height: 512px, width: 683px)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 146, + 246, + 260 + ], + "blocks": [ + { + "bbox": [ + 62, + 146, + 246, + 260 + ], + "lines": [ + { + "bbox": [ + 62, + 146, + 246, + 260 + ], + "spans": [ + { + "bbox": [ + 62, + 146, + 246, + 260 + ], + "type": "image", + "image_path": "ae77d0b6ad3698bfbceb1084fda15d79b5bcd5e798c715c50d8e3942424a7960.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 262, + 219, + 271 + ], + "lines": [ + { + "bbox": [ + 88, + 262, + 219, + 271 + ], + "spans": [ + { + "bbox": [ + 88, + 262, + 219, + 271 + ], + "type": "text", + "content": "Unsplash (height: 4220px, width: 3176px)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 247, + 147, + 430, + 262 + ], + "blocks": [ + { + "bbox": [ + 247, + 147, + 430, + 262 + ], + "lines": [ + { + "bbox": [ + 247, + 147, + 430, + 262 + ], + "spans": [ + { + "bbox": [ + 247, + 147, + 430, + 262 + ], + "type": "image", + "image_path": "d0c275b075b8c8857a13a25a089c0735af4e9974840c38205ef33243b160da1d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 262, + 411, + 273 + ], + "lines": [ + { + "bbox": [ + 264, + 262, + 411, + 273 + ], + "spans": [ + { + "bbox": [ + 264, + 262, + 411, + 273 + ], + "type": "text", + "content": "LAION-400M (height: 3637px, width: 2909px)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 280, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 280, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 545, + 340 + ], + "type": "text", + "content": "Figure 2: High-quality mask annotations for low- and high-resolution images collected from existing datasets such as COCO [31], ADE20K [65] and Cityscapes [9] as well as from the internet are presented. For the images collected from the aforementioned datasets, a visual comparison between the original and our annotations is provided in the middle and rightmost sub-figures, where the unannotated regions are shaded in black. It is worth noting that the RGB and mask images shown here have been downsampled, resulting in some quality degradation compared to the original datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 431, + 147, + 530, + 190 + ], + "blocks": [ + { + "bbox": [ + 431, + 147, + 530, + 190 + ], + "lines": [ + { + "bbox": [ + 431, + 147, + 530, + 190 + ], + "spans": [ + { + "bbox": [ + 431, + 147, + 530, + 190 + ], + "type": "image", + "image_path": "5184b01d2a1999eae0caab85da41c9397422fe810de2c8ab1214e4ce0b6fea3a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 190, + 510, + 200 + ], + "lines": [ + { + "bbox": [ + 454, + 190, + 510, + 200 + ], + "spans": [ + { + "bbox": [ + 454, + 190, + 510, + 200 + ], + "type": "text", + "content": "Clipart (340, 339)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 346, + 289, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 289, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 289, + 574 + ], + "type": "text", + "content": "LIP [13], MHP [63]) and domain adaptation (Synscapes [57], GTA5 [44]). Despite the significant contributions from these datasets, there is still a need to fulfill the requirements of real-world applications with high-quality images of large diversity. For example, a segmentation model should be robust to high-resolution images from different domains and able to segment unseen objects in the openworld setting. Most relevant to our work is the ADE20K dataset [65], which has large-scale open-vocabulary categories. However, the collected images are of low resolution and from narrowed domains. In contrast, our dataset collects images from multiple domains, including indoor, outdoor, street scenes, cartoon, and remote images, with over " + }, + { + "bbox": [ + 47, + 346, + 289, + 574 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 47, + 346, + 289, + 574 + ], + "type": "text", + "content": " of the resolution of the images falling within the high-resolution range of 2000px to 8000px. Differently from the ADE20K [64] and LVIS [14] dataset with a predefined category list, our annotation process first conducts class-agnostic mask annotation on each entity and then labels the category information in an open-vocabulary manner." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 712 + ], + "type": "text", + "content": "Scene Parsing Methods. Scene parsing methods mainly rely on convolution-based dense prediction or transformer-based query prediction. Explicit localization information, such as bounding box proposals or pixel position, is often used for pixel-level mask prediction in these approaches. Examples include FCN [35], DeepLab [5], PSPNet [62], Mask R-CNN [15], PANet [32], SOLO [53], CondInst [48], PanopticFPN [19], and PanopticFCN [28]. Inspired by the success of transformer-based detection tasks such as DETR [4] [29, 30, 10, 40], a number of methods, such as Max-Deeplab [50] and Mask2Former [7], have been pro" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 346, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 467 + ], + "type": "text", + "content": "posed to directly predict segmentation masks without relying on bounding boxes. The key to these methods is to use learnable queries trained by a set-prediction loss to model the implicit location of the potential instance area. However, query-based set-prediction methods cannot effectively perform multi-view fusion, as the query-to-object assignment can drastically change with the change of image view. This is because the set prediction loss typically adopted by such methods does not enforce any assignment constraints across multiple image views." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 475, + 411, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 411, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 411, + 489 + ], + "type": "text", + "content": "3. EntitySeg Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": "The EntitySeg dataset contains 33,227 images with high-quality mask annotations, as illustrated in Fig. 2. Compared with existing datasets, it has three distinctive features. Firstly, " + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "inline_equation", + "content": "71.25\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "inline_equation", + "content": "86.23\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": " of the images are of high quality, with at least " + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "inline_equation", + "content": "2000\\mathrm{px} \\times 2000\\mathrm{px}" + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "inline_equation", + "content": "1000\\mathrm{px} \\times 1000\\mathrm{px}" + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": ", respectively, staying in line with the current digital imaging trends. Secondly, our dataset is open-world and not restricted to predefined classes. We consider each semantically-coherent region in the images as an entity, even if it is blurred or hard to recognize semantically. Thirdly, the mask annotation along the boundaries is more precise than those in other datasets. In the following, we will provide an in-depth description of our EntitySeg dataset, as well as an extensive analysis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 664, + 408, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 408, + 675 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 408, + 675 + ], + "type": "text", + "content": "3.1. Image Collection" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 676, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 676, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 676, + 545, + 712 + ], + "type": "text", + "content": "Inspired by the collection criterion of COCO [31], we collected most non-iconic images with plentiful contextual information and non-canonical viewpoints. To en" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4049" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 206, + 187 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 206, + 187 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 206, + 187 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 206, + 187 + ], + "type": "image", + "image_path": "7c073d643f54a4396a8035f1509c7b87c6fadc96b1c183a21fd434442315a4a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 194, + 137, + 202 + ], + "lines": [ + { + "bbox": [ + 126, + 194, + 137, + 202 + ], + "spans": [ + { + "bbox": [ + 126, + 194, + 137, + 202 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 208, + 73, + 355, + 190 + ], + "blocks": [ + { + "bbox": [ + 208, + 73, + 355, + 190 + ], + "lines": [ + { + "bbox": [ + 208, + 73, + 355, + 190 + ], + "spans": [ + { + "bbox": [ + 208, + 73, + 355, + 190 + ], + "type": "image", + "image_path": "b3f66fd9c6b5940b2013e2191548c1b87a29630c0f8f892aa0b95c0fadc546d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 194, + 289, + 202 + ], + "lines": [ + { + "bbox": [ + 278, + 194, + 289, + 202 + ], + "spans": [ + { + "bbox": [ + 278, + 194, + 289, + 202 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 356, + 74, + 528, + 190 + ], + "blocks": [ + { + "bbox": [ + 356, + 74, + 528, + 190 + ], + "lines": [ + { + "bbox": [ + 356, + 74, + 528, + 190 + ], + "spans": [ + { + "bbox": [ + 356, + 74, + 528, + 190 + ], + "type": "image", + "image_path": "514c4f458271ae9adc8bea39c7abc348a45279afd0f362876ab476853c4eb968.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 194, + 463, + 202 + ], + "lines": [ + { + "bbox": [ + 453, + 194, + 463, + 202 + ], + "spans": [ + { + "bbox": [ + 453, + 194, + 463, + 202 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "text", + "content": "sure substantial domain diversity, we have sourced the images from various public datasets and the Internet, with permission for academic research. Specifically, part of the images are from COCO [31], ADE20K [65], Pascal VOC [11], Cityscapes [9], Mapillary Vistas [38], ImageNet [22], Open Images [23], DIV2K [3], Flickr2K [55], Clipart [17], Comic [17], DOTA [59], Synscapes [57] and GTA5 [44]. Moreover, we have crawled high-resolution images from Pixabay [1], Unsplash [2], and LAION-400M [46]. As shown in Fig. 3(c), most of the images are from high-resolution sources like Pixabay, Unsplash, and LAION-400M. Moreover, Fig. 3(a) reveals that our dataset has more high-resolution images, with resolutions normally distributed within the range of " + }, + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "inline_equation", + "content": "0\\mathrm{px}" + }, + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "inline_equation", + "content": "15,000\\mathrm{px}" + }, + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "text", + "content": ", in comparison to ADE20K and COCO with images all under " + }, + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "inline_equation", + "content": "1000\\mathrm{px}" + }, + { + "bbox": [ + 47, + 248, + 290, + 428 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 430, + 155, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 155, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 155, + 442 + ], + "type": "text", + "content": "3.2. Image Annotation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 445, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 287, + 612 + ], + "type": "text", + "content": "For our dataset, the annotation process consists of the following steps. Firstly, entities in the image are annotated with class-agnostic pixel-level masks, ensuring that no overlaps exist between masks. Subsequently, each mask is annotated with a class label in an open vocabulary manner. Two special considerations should be taken into account here: 1) we aonotate the entity as 'unknown' or 'blurred' if it cannot be easily named or is severely blurred; 2) the entity is annotated as a 'supercategory_other' if we only know it is a supercategory but unable to determine its fine-grained class. Our annotation process is the opposite of those of the popular segmentation datasets like COCO [31] and ADE20K [64], which first predefine their class sets before annotating masks based on the predefined classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 612, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 288, + 696 + ], + "type": "text", + "content": "Table 2 indicates that our dataset has a more considerable proportion of covered image areas and more masks on average compared to COCO- and ADE20K-Panoptic. This is attributed to our novel annotation process that enables us to take all semantically-coherent regions into account and assign class labels accordingly. Besides, our annotation procedure is more similar to the human visual system. As" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 245, + 400, + 279 + ], + "blocks": [ + { + "bbox": [ + 47, + 205, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 545, + 242 + ], + "type": "text", + "content": "Figure 3: Sub-figures (a) and (b) present the distributions of image resolutions and average number of entities among ADE20K, COCO, and EntitySeg, respectively. Moreover, sub-figure (c) displays the distribution of image sources from which the EntitySeg images were collected." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 245, + 400, + 279 + ], + "lines": [ + { + "bbox": [ + 318, + 245, + 400, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 245, + 400, + 279 + ], + "type": "image", + "image_path": "335fa5049313402eea53b520bbfaf7215329d038a2ddc769011a7638ccffb2a4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 280, + 364, + 287 + ], + "lines": [ + { + "bbox": [ + 355, + 280, + 364, + 287 + ], + "spans": [ + { + "bbox": [ + 355, + 280, + 364, + 287 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 447, + 246, + 528, + 279 + ], + "blocks": [ + { + "bbox": [ + 447, + 246, + 528, + 279 + ], + "lines": [ + { + "bbox": [ + 447, + 246, + 528, + 279 + ], + "spans": [ + { + "bbox": [ + 447, + 246, + 528, + 279 + ], + "type": "image", + "image_path": "3a6fba9d53e44bbb944da2b26f2fca4878fc5e44672c760ac7fa89fc7368e879.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 280, + 493, + 287 + ], + "lines": [ + { + "bbox": [ + 484, + 280, + 493, + 287 + ], + "spans": [ + { + "bbox": [ + 484, + 280, + 493, + 287 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 307, + 316, + 552, + 349 + ], + "blocks": [ + { + "bbox": [ + 306, + 292, + 544, + 315 + ], + "lines": [ + { + "bbox": [ + 306, + 292, + 544, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 544, + 315 + ], + "type": "text", + "content": "Table 1: Annotation consistency of class-agnostic localization and class-aware categories among annotators." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 316, + 552, + 349 + ], + "lines": [ + { + "bbox": [ + 307, + 316, + 552, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 552, + 349 + ], + "type": "table", + "html": "
DatasetImageRes (avg)↑EntityNum (avg)↑EntityNum (max)↑Entity Complexity↓Entity Simplicity↓Valid Area↑
COCO522.511.2950.7580.5810.891
ADE20K461.313.62550.8020.6060.914
EntitySeg2700.718.12360.7190.5380.999
", + "image_path": "bf6b6fd1ebf66a923b65d8dfefd2c3b06be3eae779f62e70a4ed6d0776b5c28d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 357, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 545, + 452 + ], + "type": "text", + "content": "Table 2: The statistical comparisons between COCO [31], ADE20K [65] and EntitySeg. 'ImageRes (avg)', 'EntityNum (avg)', 'Valid Area', 'Entity Complexity', and 'Entity Simplicity' refer to the average value of image resolution size, entity numbers per image, valid area ratio per image, average entity complexity and simplicity respectively. 'EntityNum (max)' means the maximum per-image number of entities within each dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 460, + 544, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 544, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 544, + 496 + ], + "type": "text", + "content": "evidenced in [36], the human vision system is intrinsically class-agnostic and can recognize entities without comprehending their usage and purpose." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 497, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 652 + ], + "type": "text", + "content": "Annotation consistency is crucial to any human-labeled dataset since it implies whether the annotation task is well-defined and proper. To study this, we randomly selected 500 images from EntitySeg (1.5% of the entire dataset), and asked another two annotators to annotate them after four months since the first round. Table 1(a) shows that the class-agnostic mask mAP in the first step of the annotation process compared to those of the other two annotators. We use one as ground truth and another one as prediction results. In Table 1(b), we also evaluate the category consistency by accuracy (ACC) under the same mask annotations. These two tables demonstrate that our EntitySeg has high annotation consistency in the mask and category labeling stages." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 660, + 408, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 660, + 408, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 660, + 408, + 672 + ], + "type": "text", + "content": "3.3. Dataset Statistics" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 676, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 676, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 676, + 545, + 712 + ], + "type": "text", + "content": "In our EntitySeg dataset, we annotate all images at the entity level, regardless of whether they belong to thing or stuff in previous datasets. As shown in Fig. 1, we show" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 59, + 700, + 266, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 700, + 266, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 266, + 712 + ], + "type": "text", + "content": "1We removed not-safe-for-work images via strict CLIP filtering." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4050" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 74, + 287, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 74, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 74, + 287, + 121 + ], + "type": "text", + "content": "some low- and high-resolution photos and their mask annotations to highlight our high-quality pixel-level annotation quality. For example, one can zoom in on the Unsplash image to see how high-quality the railing's mask is." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 122, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 288, + 312 + ], + "type": "text", + "content": "Table 2 shows the quantitative comparison among our dataset, COCO-Panoptic [31], and ADE20K-Panoptic [65]. In the EntitySeg dataset, each image has 18.1 entities on average, which is more than 11.2 and 13.6 entities in COCO and ADE20K. For a fair comparison, we treat each uncountable stuff region as one segment although it is disconnected. The detailed comparison among those three datasets on the distribution of entity number is shown in Fig. 3(b). Furthermore, the shapes of entities in our dataset are more complex than those COCO [31] and ADE20K [65] as represented by the columns 'Entity Complexity' and 'Entity Simplicity' of Table 2 where an entity with a large convexity and simplicity value means it is a simple shape (and both metrics achieve their maximum value of 1.0 for a circle [66]). More details about how complexity and simplicity are calculated can be found in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 316, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 287, + 447 + ], + "type": "text", + "content": "Class-Aware Annotation. We select a subset of 11,580 images from the dataset for class labeling, for investigating the diversity of semantic classes in our dataset. In our class annotation process, the class is labeled in an open-vocabulary and free-form manner, then the class set is continually expanded. After the class annotation process, we organize and merge the annotated classes according to Word-Tree [45]. This results in 535 things and 109 stuff classes for those 11,580 images that we use to construct the class-aware EntityClass dataset. Please refer to our supplementary file for more analysis on EntityClass dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 459, + 129, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 129, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 129, + 472 + ], + "type": "text", + "content": "4. CropFormer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": "We propose CropFormer which is the first Transformer-based set-prediction segmentation method that is capable of fusing the mask predictions from multiple image views. Inspired by existing methods [8, 7, 50, 61], we design CropFormer to learn " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": " number of " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": "-dimensional queries " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{N \\times K}" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": " to generate mask embeddings " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": " then acts as convolution filters that are applied to pixel-level mask features " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_2 \\in \\mathbb{R}^{T \\times H \\times W}" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": " segmentation masks " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{U}^{\\mathbf{m}} \\in \\mathbb{R}^{N \\times T \\times H \\times W}" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "T, H" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 479, + 287, + 683 + ], + "type": "text", + "content": " correspond to the image view, height, and width dimensions. However, such a rudimentary design can only work well with single-view inputs and cannot effectively fuse the results from multiple views. Here, we introduce a novel association module and batch-level decoder to achieve the goal of CropFormer, which is to exploit global context from full image and fine-grained local details from crops for high-quality segmentation. CropFormer is illustrated in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 688, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 688, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 688, + 287, + 712 + ], + "type": "text", + "content": "Crop Dataloger. In CropFormer framework, the crop dataloger is designed to generate a batch of images that si" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "text", + "content": "multaneously include the full images and their corresponding crops. There are two steps in our dataloader: crop and resize. First, we augment the full image " + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "inline_equation", + "content": "I^o" + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "text", + "content": " with a crop that is part of the full image. The crop is randomly extracted from one of the fixed image corners: upper-left, upper-right, bottom-left, bottom-right. The crop size is controlled by a fixed ratio hyperparameter " + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "text", + "content": " relative to the full image size. Then, we resize both the full image and crop to the same size. In this way, we construct an input tensor " + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{1 \\times 2 \\times H' \\times W' \\times 3}" + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "text", + "content": " with 2 elements (full image & 1 crop) along the view dimension. Compared to the full image, our corner crops preserve more image details (e.g., local information) that are useful for high-quality segmentation. Specifically, given an dataloader " + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 74, + 545, + 253 + ], + "type": "text", + "content": ", we define the data preprocessing as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 377, + 262, + 544, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 262, + 544, + 277 + ], + "spans": [ + { + "bbox": [ + 377, + 262, + 544, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = \\left\\{I ^ {o}, I ^ {c} \\right\\}, \\quad I ^ {o} \\sim \\mathcal {D} \\tag {1}", + "image_path": "183f4da7843816fae56068d56f1afc940784015c58fbc4bd1d092b0fb1966be3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "inline_equation", + "content": "I^o" + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "inline_equation", + "content": "I^c" + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "content": " is the cropped image controlled by fixed ratio " + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "content": ". Unlike the random cropping in [58, 16], we adopt a fixed cropping solution to keep the training and inference consistent. Such a rigid cropping strategy provides better inference efficiency since, rather than bunches of random crops, evaluating a fixed number of crops during inference is sufficient enough." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "content": "Image Encoder and Decoder. Based on learnable image-level queries " + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i\\in \\mathbb{R}^{N\\times K}" + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "content": ", we use image encoder " + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "content": " and image-level decoder " + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "inline_equation", + "content": "\\Phi_{i}" + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "content": " to generate image-level embeddings " + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{E_i}\\in \\mathbb{R}^{N\\times 2\\times 1\\times 1\\times K}" + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "content": " for the full input image and its crop. Given the input tensor (I) and queries " + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_i)" + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{E_i}" + }, + { + "bbox": [ + 304, + 376, + 544, + 447 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 383, + 449, + 544, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 449, + 544, + 462 + ], + "spans": [ + { + "bbox": [ + 383, + 449, + 544, + 462 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {\\mathbf {i}} = \\Phi_ {i} \\left(\\mathbf {Q} _ {\\mathbf {i}}, \\Theta (\\mathbf {I})\\right), \\tag {2}", + "image_path": "f612abeae2c944b881bb1f9a07e613a81640eda8aff97d1a87718a8865579895.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "inline_equation", + "content": "\\Phi_i(\\cdot)" + }, + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "text", + "content": " is a Transformer-based image-level decoder. " + }, + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{E_i}" + }, + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "text", + "content": " is used for image-level entityness prediction and pixellevel prediction using the low-level image feature " + }, + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{P_2}" + }, + { + "bbox": [ + 304, + 469, + 544, + 528 + ], + "type": "text", + "content": " (derived from the image encoder). The prediction process can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 539, + 544, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 539, + 544, + 553 + ], + "spans": [ + { + "bbox": [ + 359, + 539, + 544, + 553 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {\\mathbf {i}} ^ {\\mathrm {e}}, \\mathbf {U} _ {\\mathbf {i}} ^ {\\mathrm {m}} = \\operatorname {P r e d H e a d} _ {\\mathbf {i}} \\left(\\mathbf {E} _ {\\mathbf {i}}, \\mathbf {P} _ {\\mathbf {2}}\\right), \\tag {3}", + "image_path": "6712eeba96b27372bbc1a3109eeedb1d9834502fccec1bce01fba21edeb69034.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathrm{i}}^{\\mathrm{e}}" + }, + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathrm{i}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "text", + "content": " denote the entityness prediction and pixel-level mask outputs. Here, we employ " + }, + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 562, + 544, + 609 + ], + "type": "text", + "content": " subscript to differentiate image-level embeddings and mask outputs from those of the association module." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "content": "Association Module. Following [6], our association module aims to generate batch queries " + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "content": " that are fully shared by the full image and its crop to represent the same entities consistently. However, instead of treating " + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "content": " as learnable network parameters, considering that " + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "content": " already contains strong segmentation features, we generate " + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "content": " directly from " + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{i}} = \\{\\mathbf{E}_{I^o},\\mathbf{E}_{I^c}\\}" + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "content": ". Particularly, we use a Transformer architecture with cross-attention (" + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{XAtt}}" + }, + { + "bbox": [ + 304, + 616, + 545, + 712 + ], + "type": "text", + "content": ") and self" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4051" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 84, + 542, + 222 + ], + "blocks": [ + { + "bbox": [ + 51, + 84, + 542, + 222 + ], + "lines": [ + { + "bbox": [ + 51, + 84, + 542, + 222 + ], + "spans": [ + { + "bbox": [ + 51, + 84, + 542, + 222 + ], + "type": "image", + "image_path": "ea2028887fe9c83e2cdf29b95dc6a8252d85cfa04898818ff8d0ec1fdcfb0b7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 224, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 546, + 262 + ], + "type": "text", + "content": "Figure 4: Framework of the proposed CropFormer. The red box indicates cropped region randomly sampled from four fixed image corners. In image-level prediction, the same entity across different image views may be assigned to different queries. Our association module and batch decoder can effectively associate the same entity across different views with a single query." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 267, + 170, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 170, + 280 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 170, + 280 + ], + "type": "text", + "content": "attention " + }, + { + "bbox": [ + 48, + 267, + 170, + 280 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{SAtt}})" + }, + { + "bbox": [ + 48, + 267, + 170, + 280 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 48, + 267, + 170, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{b}}" + }, + { + "bbox": [ + 48, + 267, + 170, + 280 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 285, + 287, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 285, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 287, + 314 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\mathbf {b}} = \\operatorname {F F N} \\left(f _ {\\mathrm {S A t t}} \\left(f _ {\\mathrm {X A t t}} \\left(\\underbrace {f _ {\\mathrm {q}} \\left(\\mathbf {E} _ {I ^ {o}}\\right)} _ {\\text {q u e r y}}, \\underbrace {f _ {\\mathrm {k}} \\left(\\mathbf {E} _ {\\mathbf {i}}\\right)} _ {\\text {k e y}}, \\underbrace {f _ {\\mathrm {v}} \\left(\\mathbf {E} _ {\\mathbf {i}}\\right)} _ {\\text {v a l u e}}\\right)\\right)\\right), \\tag {4}", + "image_path": "0ba107f89564a696ce44e5275b8f6d30b457c1e04038e064144072886e4ddd37.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "text", + "content": "where FFN is a feed-forward network, " + }, + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "inline_equation", + "content": "f_{\\{\\mathbf{q}, \\mathbf{k}, \\mathbf{v}\\}}(\\cdot)" + }, + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "text", + "content": " are linear transformations. Considering the importance of full image for entity segmentation, inspired by [49, 26] we take the image-level embeddings of full image " + }, + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{I^o}" + }, + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "text", + "content": " as query, while treating all image-level embeddings " + }, + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_i" + }, + { + "bbox": [ + 48, + 319, + 287, + 378 + ], + "type": "text", + "content": " as key and value." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 380, + 287, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 380, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 380, + 287, + 406 + ], + "type": "text", + "content": "Batch-level Decoder. Given " + }, + { + "bbox": [ + 48, + 380, + 287, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{b}} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}" + }, + { + "bbox": [ + 48, + 380, + 287, + 406 + ], + "type": "text", + "content": ", we obtain batch embeddings " + }, + { + "bbox": [ + 48, + 380, + 287, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{b}} \\in \\mathbb{R}^{N \\times 1 \\times 1 \\times 1 \\times K}" + }, + { + "bbox": [ + 48, + 380, + 287, + 406 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 123, + 411, + 287, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 411, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 123, + 411, + 287, + 425 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {\\mathbf {b}} = \\Phi_ {b} (\\mathbf {Q} _ {\\mathbf {b}}, \\Theta (\\mathbf {I})), \\tag {5}", + "image_path": "923633b04e9348dee7bfce086f6f18a4ed5483e4da0ac375ac61f5d30245bce6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\Phi_b(\\cdot)" + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "text", + "content": " denotes the batch-level decoder. Thus, we broadcast " + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{b}}" + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "text", + "content": " to the shape of " + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "inline_equation", + "content": "N\\times 2\\times 1\\times 1\\times K" + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "text", + "content": " to share it between the full image and its crop. Finally, the broadcasted " + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{b}}" + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "text", + "content": " and low-level batch features " + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{P_2}" + }, + { + "bbox": [ + 48, + 430, + 287, + 489 + ], + "type": "text", + "content": " are used for batch-level mask classification and pixel-wise prediction:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 495, + 287, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 495, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 99, + 495, + 287, + 509 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {\\mathrm {b}} ^ {\\mathrm {e}}, \\mathbf {U} _ {\\mathrm {b}} ^ {\\mathrm {m}} = \\operatorname {P r e d H e a d} _ {\\mathrm {b}} \\left(\\mathbf {E} _ {\\mathrm {b}}, \\mathbf {P} _ {2}\\right), \\tag {6}", + "image_path": "a7c55ce0e855048326d0a5f9c1c6f144ecfec3b247152b11e8cbdc8fcd7b0fa3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 574 + ], + "type": "text", + "content": "For the image-level decoder, the queries are separate for each image/crop. Whereas, for the batch-level decoder, the queries are shared between the image and its crop. We illustrate the differences of the image- and batch-level decoders in Fig. 5, where we regard two mage views as an entirety." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "text", + "content": "Training. During training, we employ two separate losses " + }, + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i" + }, + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "text", + "content": " for image- and batch-level predictions, with respect to ground truth " + }, + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 48, + 576, + 287, + 635 + ], + "type": "text", + "content": ". The main difference between the two losses lies in whether the same entities in full image and crop are tied to the same queries or not." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 60, + 635, + 271, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 635, + 271, + 647 + ], + "spans": [ + { + "bbox": [ + 60, + 635, + 271, + 647 + ], + "type": "text", + "content": "The overall training loss for our model is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 652, + 287, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 652, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 58, + 652, + 287, + 709 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\sum_ {\\mathbf {k} \\in \\{\\mathbf {i}, \\mathbf {b} \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {e}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {e}}\\right) + \\sum_ {\\mathbf {k} \\in \\{\\mathbf {i}, \\mathbf {b} \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {b c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {m}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {m}}\\right) + \\\\ \\sum_ {\\mathbf {k} \\in \\{i, b \\}} \\mathcal {L} _ {\\mathbf {k}} ^ {\\mathrm {d i c e}} \\left(\\mathbf {U} _ {\\mathbf {k}} ^ {\\mathrm {m}}, \\mathbf {G} _ {\\mathbf {k}} ^ {\\mathrm {m}}\\right), \\tag {7} \\\\ \\end{array}", + "image_path": "d75e7df3de59d1133312a527b48fcd9a0704699ac97d82d3eb7cb86751f1e231.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 269, + 383, + 357 + ], + "blocks": [ + { + "bbox": [ + 317, + 269, + 383, + 357 + ], + "lines": [ + { + "bbox": [ + 317, + 269, + 383, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 383, + 357 + ], + "type": "image", + "image_path": "22fe4a5893d2414a865955c60b45a292367c6d822fb2e86860b5fe4745ab2811.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 361, + 544, + 384 + ], + "lines": [ + { + "bbox": [ + 306, + 361, + 544, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 544, + 384 + ], + "type": "text", + "content": "Figure 5: The illustration of image- and batch-level decoder and association module." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 402, + 269, + 459, + 357 + ], + "blocks": [ + { + "bbox": [ + 402, + 269, + 459, + 357 + ], + "lines": [ + { + "bbox": [ + 402, + 269, + 459, + 357 + ], + "spans": [ + { + "bbox": [ + 402, + 269, + 459, + 357 + ], + "type": "image", + "image_path": "25ddc2919141a5656a1bbc3e7648bb011049428a03e41ca3796e5ab54312478d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 478, + 269, + 531, + 356 + ], + "blocks": [ + { + "bbox": [ + 478, + 269, + 531, + 356 + ], + "lines": [ + { + "bbox": [ + 478, + 269, + 531, + 356 + ], + "spans": [ + { + "bbox": [ + 478, + 269, + 531, + 356 + ], + "type": "image", + "image_path": "d293bd3770dd7d3a51207b77f08a3d5ffab0105e1d67e57ccbeea78403fef5d1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{ce}}" + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{ce}}" + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "content": " denote binary cross-entropy loss for image- and batch-level entityness prediction. Similarly, " + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{bce}}" + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{bce}}" + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{i}}^{\\mathrm{dice}}" + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{b}}^{\\mathrm{dice}}" + }, + { + "bbox": [ + 306, + 394, + 544, + 443 + ], + "type": "text", + "content": " denote the binary cross-entropy and dice loss for image- and batch-level mask prediction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 451, + 544, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 451, + 544, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 544, + 535 + ], + "type": "text", + "content": "Inference. In CropFormer, the same entities between the full image and its crop are represented by the same queries by virtue of the association module. For the final segmentation output, we fuse the per-pixel mask predictions obtained from the full image and every crop of the 4 corners by using an average operation. The confidence score of each entity comes from the batch-level entityness prediction score." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 548, + 387, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 387, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 387, + 562 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 569, + 544, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 569, + 544, + 676 + ], + "spans": [ + { + "bbox": [ + 305, + 569, + 544, + 676 + ], + "type": "text", + "content": "We conduct our main experiments on the entity segmentation with proposed EntitySeg Dataset in this work. First, we benchmark our dataset in several segmentation tasks, including entity, instance, semantic and panoptic segmentation. Then, we ablate our proposed CropFormer to demonstrate its effectiveness. Note that, in Sec. 4, we omit the conventional data batch dimension to improve clarity, but in practice, we train CropFormer with a batch of multiple dataset images and their crops, following existing methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 677, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 544, + 712 + ], + "type": "text", + "content": "Training Details. We leave out some details like the training schedule and some ablation studies due to paper length limitations (see supplementary for details)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4052" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 291, + 243 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 291, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 291, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 291, + 243 + ], + "type": "table", + "html": "
TaskMetricsMethodBackbonePretrainPerformance
SEMmIoUDeeplabv3R-50ImageNet27.9
Mask2FormerR-50ImageNet37.8
R-50COCO43.3
Swin-TCOCO45.0
Swin-LCOCO50.5
INSAPMask-RCNNR-50ImageNet5.0
Mask2FormerR-50ImageNet13.0
R-50COCO20.3
Swin-TCOCO22.7
Swin-LCOCO30.3
PANPQPanopticFPNR-50ImageNet3.6
Mask2FormerR-50ImageNet5.5
R-50COCO9.6
Swin-TCOCO9.8
Swin-LCOCO13.4
ENTAPeMask-RCNNR-50ImageNet24.9
Mask2FormerR-50ImageNet26.0
R-50COCO35.2
Swin-TCOCO42.9
Swin-LCOCO46.2
", + "image_path": "dc1812bcf7fc6e7bcc73e3e250217083ae523484aa06c92fe617a8fd0abb0fe1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 334, + 287, + 378 + ], + "blocks": [ + { + "bbox": [ + 47, + 251, + 288, + 322 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 288, + 322 + ], + "type": "text", + "content": "Table 3: Benchmark on several segmentation tasks on the EntitySeg Dataset. The 'SEM', 'INS', 'PAN', and 'ENT' indicate the task of semantic, instance, panoptic and entity segmentation. We note that we leave out all details and more experiments in the supplementary file due to paper length limitations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 334, + 287, + 378 + ], + "lines": [ + { + "bbox": [ + 49, + 334, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 49, + 334, + 287, + 378 + ], + "type": "table", + "html": "
Transfer DatasetTypeSEM (mIoU)INS (AP)PAN (PQ)ENT (APe)
ADE20KImage57.2(+1.0)36.0(+1.1)49.0(+0.9)40.6(+1.5)
YVIS2019 [60]Video-61.3(+0.9)--
VIPSeg [37]Video--37.0(+1.1)-
UVO [52]Video---29.5(+2.2)
", + "image_path": "5840bc1499f4d4d472686a69955b18abc2e5825690386c4216bf23b5b3b0479a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 444, + 168, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 168, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 168, + 455 + ], + "type": "text", + "content": "5.1. Dataset Benchmarks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 460, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 579 + ], + "type": "text", + "content": "Benchmark on EntitySeg Dataset. In Table 3, we provide results on semantic, instance, panoptic, and entity segmentation on EntitySeg dataset with some baseline methods. Similar to existing segmentation datasets like COCO and ADE20K, the use of COCO pre-trained weight and a stronger backbone can benefit each segmentation performance of EntitySeg Dataset. Compared to COCO and ADE20K, EntitySeg dataset is more challenging as evidenced by the smaller performance numbers (e.g., Swin-L Mask2Former's PQ on COCO-PAN is " + }, + { + "bbox": [ + 47, + 460, + 287, + 579 + ], + "type": "inline_equation", + "content": "57.8\\%" + }, + { + "bbox": [ + 47, + 460, + 287, + 579 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 658 + ], + "type": "text", + "content": "Transferring EntitySeg to other general segmentation tasks. Table 4 shows the performance comparison of transferring models pretrained on either COCO or EntitySeg dataset to other image/video-level segmentation tasks. Owing to its emphasis on in-the-wild generalization, EntitySeg provides greater improvements than COCO dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 664, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 288, + 712 + ], + "type": "text", + "content": "Benefit of EntitySeg for high-quality image segmentation. In Table 5, we transfer Mask2Former pretrained on EntitySeg dataset for finetuning on Dichotomous Image Segmentation [42]. Compared to state-of-the-art ISNet [42]" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 72, + 531, + 99 + ], + "blocks": [ + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 422 + ], + "type": "text", + "content": "Table 4: The relative performance gains of EntitySeg pretraining over COCO-Panoptic pretraining when fine-tuned on various image/video-level segmentation datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 72, + 531, + 99 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 531, + 99 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 531, + 99 + ], + "type": "table", + "html": "
MethodFβmx↑Fβw↑M↓Sα↑Eφm↑HCEγ↓
IS-Net [42].791.717.074.813.8561116
Ours.872.789.035.896.913969
", + "image_path": "4afcaefe5b3ab1bee75a015442d86b867c3337fdfa1cdefb88f1f8edfcf0aa35.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 156, + 542, + 200 + ], + "blocks": [ + { + "bbox": [ + 305, + 107, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 305, + 107, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 107, + 545, + 144 + ], + "type": "text", + "content": "Table 5: Evaluation on Dichotomous Image Segmentation (DIS5K dataset [42].) 'Ours' indicates Mask2Former pretrained on EntitySeg and then finetuned on DIS5K." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 156, + 542, + 200 + ], + "lines": [ + { + "bbox": [ + 309, + 156, + 542, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 542, + 200 + ], + "type": "table", + "html": "
Train DatasetTaskLVIS [14]OCID [47]CAMO [24]FSS [27]Avg. AR
COCOPAN24.3839.9443.8064.9643.27
COCOENT27.7545.8859.7075.6052.23
SAM-H [21]Prompt49.2555.1775.3373.6063.34
EntitySegENT33.5676.6373.2086.2667.41
", + "image_path": "7932118177984624e80060f10899fc4c36f5014152d9a86622f16bb1c19ec921.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 207, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 291 + ], + "type": "text", + "content": "Table 6: In-the-wild segmentation performance (AR@100 with IoU threshold 0.5). We directly evaluate models trained on either EntitySeg or COCO on large-vocabulary segmentation LVIS [14], object clutter indoor segmentation for robot grab [47], camouflaged instance segmentation CAMO [24], few-shot segmentation FSS [27]. All the models are based on Mask2Former with Swin-Large backbone." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 299, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 348 + ], + "type": "text", + "content": "that is carefully designed for the task, Mask2Former pretrained on EntitySeg outperforms ISNet by a large margin, despite not having any task-specific design. This is by virtue of the high-quality mask annotations in EntitySeg dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "type": "text", + "content": "Segmentation 'in the wild' Table 6 compares in-the-wild segmentation performance between Mask2Formers trained on either EntitySeg or COCO. Here, we directly use the trained models without finetuning on the evaluated datasets. The EntitySeg outperforms COCO-Panoptic [20] and COCO-Entity [41] by large margins." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 433, + 528, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 528, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 528, + 445 + ], + "type": "text", + "content": "5.2. A Comprehensive Analysis of CropFormer" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 544, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 544, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 544, + 473 + ], + "type": "text", + "content": "We first ablate Cropformer on the entity segmentation task and then report results on other segmentation datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "content": "Table 7 shows the benefits of CropFormer on segmentation performance and computation cost compared to the Mask2Former [7], its multi-scale extension, as well as related methods [6, 16] that supports multi-view fusion. The first four rows are our baseline Mask2Former with the single-scale inference (800, 1040, or 2700 for shortest side) and (1333, 1732 or 4500 for longest side) for the full images, where " + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "inline_equation", + "content": "1040 = 800 \\times \\delta" + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "inline_equation", + "content": "1732 = 1333 \\times \\delta" + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\delta = 0.7" + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "content": " by default. For the results in the fifth row, we use test-time Hungarian matching to associate the same entities obtained with multi-scale images. We find minimal improvement with such an inference strategy. In the sixth and seventh rows, directly using video-level Mask2Former or VITA, the state-of-the-art video instance segmentation framework, merely brings marginal performance gains. Whereas, using the crop output from CropFormer's batch decoder achieves a significant " + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^e}" + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "content": " gain as indicated by the last row. By combining the full image and four crop outputs from CropFormer (final row), we obtain even stronger " + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{AP^e}" + }, + { + "bbox": [ + 304, + 474, + 545, + 712 + ], + "type": "text", + "content": " gains compared to the baseline (first row). Compared to the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4053" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 72, + 526, + 148 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 526, + 148 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 526, + 148 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 526, + 148 + ], + "type": "table", + "html": "
MethodInput PreprocessingGPU Type\\(AP^e\\)GPU Memory (Train)GPU Memory (Inference)Train Time (GPU day)Inference Time (ms)
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.55.8G3.0G2.4637
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (2700, 4500)\\)A100-40GOOMOOM-OOM-
Mask2Formerhigh-res image \\(\\xrightarrow{\\text{resize}} (2700, 4500)\\)A100-80G40.958.8G10.4G60.83312
V-Mask2Former [6]\\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.710.4G6.7G6.81503
VITA [16]\\(\\xrightarrow{\\text{resize}} (800, 1333)\\)A100-40G39.911.9G6.8G10.21512
CropFormer\\(\\xrightarrow{\\text{crop}(\\delta)} (800, 1333)\\)A100-40G41.011.4G6.9G9.81503
", + "image_path": "9eaab0af6855c10b9a0171556f66cea9780c0aaaa73dfab6c2144eda4521ed81.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 207, + 293, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 156, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 545, + 202 + ], + "type": "text", + "content": "Table 7: Comparison with high-resolution methods on EntitySeg with multi-scale training. " + }, + { + "bbox": [ + 47, + 156, + 545, + 202 + ], + "type": "inline_equation", + "content": "(\\mathrm{x},\\mathrm{y})" + }, + { + "bbox": [ + 47, + 156, + 545, + 202 + ], + "type": "text", + "content": " indicates number of pixels along the shortest (x) and longest (y) sides of image. Row 1, 2, & 3 are single-scale Mask2Formers trained and evaluated at various single-scale resolutions. (V-Mask2Former & VITA) mask predictions from full image and four crops are fused with either video-level Mask2Former or VITA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 207, + 293, + 258 + ], + "lines": [ + { + "bbox": [ + 49, + 207, + 293, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 207, + 293, + 258 + ], + "type": "table", + "html": "
Iterations (Train)Mask2FormerCropFormerDecoderAPeAP50AP75
1×39.541.0 (+1.5)I-O39.356.739.8
2×40.242.1 (+1.9)B-O39.156.639.7
3×40.942.8 (+1.9)B-C40.257.540.8
4×40.942.7 (+1.8)B-OC41.058.441.9
(a) metric: APe(b)
", + "image_path": "149b176edaaea1262a1752a11e11b338a55167c8b0f334f79a772657ed93636a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 335, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 287, + 361 + ], + "type": "text", + "content": "Mask2Former baselines with a naive increase in image resolution, CropFormer is far more efficient computationally." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 430 + ], + "type": "text", + "content": "Longer Training Schedules Table 8(a) indicates that CropFormer consistently improves over the baseline Mask2Former when we adopt " + }, + { + "bbox": [ + 47, + 369, + 287, + 430 + ], + "type": "inline_equation", + "content": "2 \\times /3 \\times" + }, + { + "bbox": [ + 47, + 369, + 287, + 430 + ], + "type": "text", + "content": " training schedules. A similar trend is observed when using Swin-L [33] backbone, which is provided in our supplementary file." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 439, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 287, + 511 + ], + "type": "text", + "content": "Decoder Performance Table 8(b) shows the impact of using different decoder combinations in CropFormer for mask prediction during inference. Using just the crop outputs from CropFormer's batch decoder (B-C) outperforms using just the full image view (I-O, B-O). Combining both the full image and crops (B-OC) gives the best result." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": "CropDataloger Table 9(a) and (b) show the ablation study on the usage of crop ratio and crop type. As indicated by Table 9(a), CropFormer obtains the best performance with a crop ratio " + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 521, + 287, + 593 + ], + "type": "text", + "content": " of 0.7. Smaller crop ratios perform worse and it might be a problem caused by queries that are not robust against drastic image transformation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 602, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 287, + 651 + ], + "type": "text", + "content": "Association Module Table 9(c) shows the ablation study on the structure of the association module. There is a slight performance difference between using self-attention or a feedforward module following cross-attention." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 287, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 718 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 718 + ], + "type": "text", + "content": "Comparison to other high-quality segmentation methods. Table 10(a) shows that our CropFormer significantly outperforms recent high-quality instance-level segmentation methods like Mask Transfiner [18] and PatchDCT [56] on both traditional and boundary APs." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 207, + 537, + 258 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 287, + 326 + ], + "type": "text", + "content": "Table 8: Ablation study on: (a) training schedules; (b) CropFormer's decoder combinations. In sub-table(b), the 'Decoder' column indicates whether we use the predictions of the full image ('O'), four cropped patches ('C'), or both of them ('OC') from the 'Image' or 'Batch' decoder." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 207, + 537, + 258 + ], + "lines": [ + { + "bbox": [ + 310, + 207, + 537, + 258 + ], + "spans": [ + { + "bbox": [ + 310, + 207, + 537, + 258 + ], + "type": "table", + "html": "
δAPeTrainTestAPeXAttSAttFFNAPe
0.538.5RandomFixed (4)39.7✓○○40.7
0.640.2Fixed (4)Fixed (4)41.0✓○✓40.8
0.741.0Fixed (4)Fixed (8)41.3✓✓○40.8
0.840.9Fixed (8)Fixed (8)41.0✓✓✓41.0
(a)(b)(c)
", + "image_path": "3ede0a76b697c1c7725fb1354d8b9c3099b6068534bb688929b39e6e008421e8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 330, + 552, + 381 + ], + "blocks": [ + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "type": "text", + "content": "Table 9: Ablation study on the usage of crop ratio " + }, + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "type": "text", + "content": ", crop type and association module in CropFormer. In sub-table (b), 'Random' indicates random crops and 'Fixed (4/8)' indicates 4 or 8 fixed corner crops. In sub-table(c), " + }, + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 305, + 266, + 545, + 325 + ], + "type": "text", + "content": " means whether we use the module or not." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 330, + 552, + 381 + ], + "lines": [ + { + "bbox": [ + 310, + 330, + 552, + 381 + ], + "spans": [ + { + "bbox": [ + 310, + 330, + 552, + 381 + ], + "type": "table", + "html": "
MethodAPeAPbTaskMask2FormerCropFormer
Mask Transfiner [18]33.725.0COCO-INS50.151.9
PatchDCT [56]35.427.4COCO-PAN57.859.1
Mask2Former39.531.2ADE20K-INS34.936.4
Ours41.033.7ADE20K-PAN48.149.7
(a)(b) metrics: AP for INS; PQ for PAN
", + "image_path": "e69df0e84f71413b854bf369f50dcba49d7b44f1504764b09d147a24cd96809f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "type": "text", + "content": "Table 10: (a) Comparison with high-quality segmentation methods on EntitySeg task/dataset using " + }, + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^e" + }, + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "type": "text", + "content": " and boundary AP " + }, + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^b)" + }, + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "type": "text", + "content": ". (b) evaluation on instance (INS) and panoptic (PAN) segmentation tasks on COCO and ADE20K datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 441, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 536 + ], + "type": "text", + "content": "Improvement on class-aware segmentation tasks Table 10(b) shows the class-aware task performance of CropFromer with Swin-L [33] backbone and " + }, + { + "bbox": [ + 305, + 441, + 545, + 536 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 305, + 441, + 545, + 536 + ], + "type": "text", + "content": " training schedule. CropFormer can still achieve competitive gains on the instance and panoptic segmentation tasks on COCO and ADE20K datasets. This suggests that CropFormer may provide other benefits that are beyond high-resolution segmentation (e.g., prediction robustness)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 539, + 378, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 378, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 378, + 552 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "type": "text", + "content": "This paper proposes an EntitySeg dataset and a CropFormer framework for high-quality entity segmentation, with a strong focus on in-the-wild generalization and high-quality dense segmentation. The EntitySeg dataset contains about 33K images from various domains and resolutions, accompanied by high-quality mask annotations for training and testing. A novel CropFormer framework is also proposed to overcome the computation challenges introduced by EntitySeg dataset's high-quality and high-resolution characteristics. Specifically, it enables fusing multi-view predictions from a low-res full image and its high-res crops by learning batch-level queries. We hope this work will serve as a catalyst for future research on open-world and high-quality image segmentation tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4054" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 69, + 107, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 69, + 107, + 81 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 107, + 81 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 89, + 288, + 711 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 89, + 287, + 110 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 89, + 287, + 110 + ], + "spans": [ + { + "bbox": [ + 53, + 89, + 287, + 110 + ], + "type": "text", + "content": "[1] Pixabay. https://pixabay.com. Accessed: 2022-04-17. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 113, + 288, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 113, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 113, + 288, + 133 + ], + "type": "text", + "content": "[2] Unsplash. https://unsplash.com. Accessed: 2022-04-17. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 136, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 136, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 54, + 136, + 288, + 168 + ], + "type": "text", + "content": "[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPR Workshops, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 170, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 288, + 203 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 205, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 288, + 247 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. PAMI, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 249, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 288, + 293 + ], + "type": "text", + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 294, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 288, + 337 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 339, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 288, + 371 + ], + "type": "text", + "content": "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 373, + 288, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 288, + 427 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 430, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 430, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 50, + 430, + 287, + 461 + ], + "type": "text", + "content": "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 464, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 464, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 50, + 464, + 287, + 496 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 2010. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 498, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 498, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 50, + 498, + 287, + 529 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 532, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 532, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 50, + 532, + 287, + 574 + ], + "type": "text", + "content": "[13] Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In CVPR, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 577, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 577, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 50, + 577, + 287, + 609 + ], + "type": "text", + "content": "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 2, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 611, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 611, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 50, + 611, + 287, + 632 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 50, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 50, + 635, + 287, + 666 + ], + "type": "text", + "content": "[16] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurlPS, 2022. 5, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 50, + 668, + 287, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 668, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 50, + 668, + 287, + 711 + ], + "type": "text", + "content": "[17] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, 2018. 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 75, + 544, + 711 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 75, + 544, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 75, + 544, + 107 + ], + "spans": [ + { + "bbox": [ + 307, + 75, + 544, + 107 + ], + "type": "text", + "content": "[18] Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask transfiner for high-quality instance segmentation. In CVPR, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 108, + 544, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 544, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 544, + 139 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 544, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 544, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 544, + 172 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 544, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 544, + 217 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In arXiv, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 219, + 544, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 544, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 544, + 250 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2012. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 251, + 544, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 544, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 544, + 304 + ], + "type": "text", + "content": "[23] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 306, + 544, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 306, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 544, + 358 + ], + "type": "text", + "content": "[24] Trung-Nghia Le, Yubo Cao, Tan-Cong Nguyen, Minh-Quan Le, Khanh-Duy Nguyen, Thanh-Toan Do, Minh-Triet Tran, and Tam V Nguyen. Camouflaged instance segmentation in the-wild: Dataset, method, and benchmark suite. TIP, 2021. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 361, + 544, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 544, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 544, + 392 + ], + "type": "text", + "content": "[25] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 394, + 544, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 544, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 544, + 437 + ], + "type": "text", + "content": "[26] Peizhao Li, Jiuming Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Rajiv Jain, Varun Manjunatha, and Hongfu Liu. Selfdoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 438, + 544, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 544, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 544, + 469 + ], + "type": "text", + "content": "[27] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 472, + 544, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 544, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 544, + 514 + ], + "type": "text", + "content": "[28] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. TPAMI, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 515, + 544, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 544, + 546 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 544, + 546 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 548, + 544, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 544, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 544, + 579 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 581, + 544, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 544, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 544, + 624 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 624, + 544, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 544, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 544, + 655 + ], + "type": "text", + "content": "[32] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 658, + 544, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 544, + 711 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 544, + 711 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 8" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4055" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 50, + 75, + 287, + 710 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 50, + 75, + 287, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 75, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 50, + 75, + 287, + 107 + ], + "type": "text", + "content": "[34] Zerun Liu, Fan Zhang, Jingxuan He, Jin Wang, Zhangye Wang, and Lechao Cheng. Text-guided mask-free local image retouching. arXiv preprint arXiv:2212.07603, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 109, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 109, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 109, + 287, + 140 + ], + "type": "text", + "content": "[35] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 144, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 144, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 50, + 144, + 287, + 175 + ], + "type": "text", + "content": "[36] David Marr. Vision: A computational investigation into the human representation and processing of visual information. San Francisco: W.H. Freeman. Print., 1982. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 178, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 178, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 50, + 178, + 287, + 209 + ], + "type": "text", + "content": "[37] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 212, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 212, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 50, + 212, + 287, + 244 + ], + "type": "text", + "content": "[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 247, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 247, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 50, + 247, + 287, + 278 + ], + "type": "text", + "content": "[39] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 281, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 281, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 50, + 281, + 287, + 313 + ], + "type": "text", + "content": "[40] Lu Qi, Jason Kuen, Jiumiang Gu, Zhe Lin, Yi Wang, Yukang Chen, Yanwei Li, and Jiaya Jia. Multi-scale aligned distillation for low-resolution detection. In CVPR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 50, + 316, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 316, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 50, + 316, + 287, + 347 + ], + "type": "text", + "content": "[41] Lu Qi, Jason Kuen, Yi Wang, Jiumiang Gu, Hengshuang Zhao, Zhe Lin, Philip Torr, and Jiaya Jia. Open-world entity segmentation. TPAMI, 2022. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 50, + 350, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 350, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 50, + 350, + 287, + 381 + ], + "type": "text", + "content": "[42] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, et al. Highly accurate dichotomous image segmentation. In ECCV, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 50, + 384, + 287, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 384, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 50, + 384, + 287, + 416 + ], + "type": "text", + "content": "[43] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 418, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 418, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 50, + 418, + 287, + 451 + ], + "type": "text", + "content": "[44] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 453, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 453, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 50, + 453, + 287, + 507 + ], + "type": "text", + "content": "[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 509, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 509, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 50, + 509, + 287, + 563 + ], + "type": "text", + "content": "[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshop, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 565, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 565, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 50, + 565, + 287, + 608 + ], + "type": "text", + "content": "[47] Markus Suchi, Timothy Patten, David Fischinger, and Markus Vincze. Easylabel: A semi-automatic pixel-wise object annotation tool for creating robotic RGB-D datasets. In ICRA, 2019. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 610, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 610, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 50, + 610, + 287, + 632 + ], + "type": "text", + "content": "[48] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 634, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 634, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 50, + 634, + 287, + 666 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 50, + 668, + 287, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 668, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 50, + 668, + 287, + 710 + ], + "type": "text", + "content": "[50] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 75, + 545, + 688 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 75, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 75, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 75, + 545, + 139 + ], + "type": "text", + "content": "[51] Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Chunjing Xu, and Yanwei Fu. Manitrans: Entity-level text-guided image manipulation via token-wise semantic alignment and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10707-10717, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[52] Weiyao Wang, Matt Feiszli, Heng Wang, and Du Tran. Unidentified video objects: A benchmark for dense, open-world segmentation. arXiv preprint arXiv:2104.04691, 2021.7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[53] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "type": "text", + "content": "[54] Yikai Wang, Jianan Wang, Guansong Lu, Hang Xu, Zhenguo Li, Wei Zhang, and Yanwei Fu. Entity-level text-guided image manipulation. arXiv preprint arXiv:2302.11383, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[55] Yingqian Wang, Longguang Wang, Jungang Yang, Wei An, and Yulan Guo. Flickr1024: A large-scale dataset for stereo image super-resolution. In ICCV Workshops, 2019. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "text", + "content": "[56] Qinrou Wen, Jirui Yang, Xue Yang, and Kewei Liang. Patchdct: Patch refinement for high quality instance segmentation. In ICLR, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 332, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 364 + ], + "type": "text", + "content": "[57] Magnus Wrenninger and Jonas Unger. Synscapes: A photorealistic synthetic dataset for street scene parsing. arXiv preprint arXiv:1810.08705, 2018. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "type": "text", + "content": "[58] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[59] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, 2018. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 444, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 464 + ], + "type": "text", + "content": "[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "text", + "content": "[61] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In ECCV, 2022. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 500, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 532 + ], + "type": "text", + "content": "[62] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "text", + "content": "[63] Jian Zhao, Jianshu Li, Yu Cheng, Terence Sim, Shuicheng Yan, and Jiashi Feng. Understanding humans in crowded scenes: Deep nested adversarial learning and a new benchmark for multi-human parsing. In ACMMM, 2018. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "text", + "content": "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 3, 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "text", + "content": "[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 3, 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 657, + 545, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 688 + ], + "type": "text", + "content": "[66] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dollár. Semantic amodal segmentation. In CVPR, 2017. 2, 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4056" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_content_list.json b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0b2400c6bd81d4a32d9e28e6021171e21f5d9bce --- /dev/null +++ b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_content_list.json @@ -0,0 +1,1758 @@ +[ + { + "type": "text", + "text": "High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net", + "text_level": 1, + "bbox": [ + 102, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zinuo Li* Xuhang Chen* Chi-Man Pun† Xiaodong Cun† University of Macau", + "bbox": [ + 215, + 202, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shadows often occur when we capture the document with casual equipment, which influences the visual quality and readability of the digital copies. Different from the algorithms for natural shadow removal, the algorithms in document shadow removal need to preserve the details of fonts and figures in high-resolution input. Previous works ignore this problem and remove the shadows via approximate attention and small datasets, which might not work in real-world situations. We handle high-resolution document shadow removal directly via a larger-scale real-world dataset and a carefully-designed frequency-aware network. As for the dataset, we acquire over 7k couples of high-resolution $(2462 \\times 3699)$ images of real-world documents pairs with various samples under different lighting circumstances, which is 10 times larger than existing datasets. As for the design of the network, we decouple the high-resolution images in the frequency domain, where the low-frequency details and high-frequency boundaries can be effectively learned via the carefully designed network structure. Powered by our network and dataset, the proposed method shows a clearly better performance than previous methods in terms of visual quality and numerical results. The code, models, and dataset are available at https://github.com/CXH-Research/DocShadow-SD7K.", + "bbox": [ + 75, + 306, + 473, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When occluders obstruct the light source, shadows are cast on the paper. These shadows reduce the readability and hinder subsequent text-related intelligent tasks such as document enhancement, optical character recognition (OCR), key information extraction, and semantic entity labeling [17, 36, 28], etc.", + "bbox": [ + 75, + 742, + 468, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Removing the cast shadow of the digital copy has a long need in the real world and also there are a lot of meth-", + "bbox": [ + 76, + 834, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28248561b5a96f35328a5d259765c94aa8659814812532f61cc9123f682b37dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 272, + 627, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a58314bacef103a319dbf819aa6b55f9edbe16120dffaedc308bf8265f676658.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 501, + 415, + 627, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ada9ef044eab22ffb70a56b5816b42e5e29c695afe9c69b1b4a4b124a433333e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 287, + 756, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3313398616b56ebe2fc1975c8b7236339da7c8d8d03cc98d76d941f1b50fabae.jpg", + "image_caption": [ + "(b) Shah et al. [27]" + ], + "image_footnote": [], + "bbox": [ + 633, + 415, + 759, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2cae5e479813c2403bb3f00ba8334ec4981e4ad64ef4cffc5a54a239dea3aea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 272, + 890, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01b4daa90f855adf0da9244c45fd9be6d7dc0e6165bd89e5376bc12dc662201c.jpg", + "image_caption": [ + "(c) Fu et al. [8]" + ], + "image_footnote": [], + "bbox": [ + 764, + 415, + 890, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8bdf7babb2ed771fb7b544ddca8ce680fbe6abd578c9a64fcc490e23d07d089.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 503, + 627, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8da33ff966afc1227b2d19f6063a88ec014c8a50b65b9d4b317e44359efbf6d0.jpg", + "image_caption": [ + "(d) Liu et al. [25]" + ], + "image_footnote": [], + "bbox": [ + 501, + 647, + 627, + 712 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87f49fc9a1a0ca8a5a4b9e49553b42eeb5972a8a95d8703585afadb91836e09d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 503, + 759, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5df81b1c025cbda53448a1652da314adbe0511c77a6726d39d6f011e0019c28.jpg", + "image_caption": [ + "(e) Hu et al. [15]" + ], + "image_footnote": [], + "bbox": [ + 633, + 647, + 759, + 712 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ad08578990237c1cc1107e91198ec1632793025d7a85ce750802b97dbf9ae98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 503, + 890, + 647 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ab42bdea443cf76f2a254b30a1996da4c1767e6c950f569492ac1401c862200.jpg", + "image_caption": [ + "(f) Ours", + "Figure 1. Visual results of input document shadow image (a), classic method (b), supervised method (c), weakly-supervised method (d), unsupervised method (e) and ours (f). Our model removes the shadow while preserving the original document's content and aspect ratio." + ], + "image_footnote": [], + "bbox": [ + 764, + 647, + 890, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ods, including the traditional methods that applied physics-based illumination models [16, 30, 32]. In reality, these approaches are significantly constrained due to the fact that the assumptions utilized in illumination models [27] are often too rigid to account for real-world shadow. Thus, they may", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 96, + 887, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12449", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "incorrectly compute the exposure and reflectance as shown in Fig. 1.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent methods apply deep learning for this task since it has achieved remarkable progress in natural shadow removal [6] and other computer vision tasks [12, 13, 35]. However, several obstacles make it challenging to transfer previous natural shadow removal methods to the document domain. On the one hand, in contrast to natural shadow removal, documents are presented by high-resolution images in the real world, which needs to preserve fine-grained information such as fonts and figures. However, the majority of document shadow removal methods [24, 4] are only designed for relatively low-resolution images and is hard to handle high-resolution situations directly since they are relying on the low-resolution approximation of attention for guidance. On the other hand, there are no large-scale and high-resolution datasets designed specifically for document shadow removal at this time. Consequently, given limited image pairs for training, existing methods usually result in different artifacts (as shown in Fig. 1 (c)). Although weakly-supervised [18] and unsupervised [15, 25] shadow removal methods alleviate the requirements for data, they hold a strong assumption of the statistical similarity between shadowed and unshadowed domains. As illustrated in Fig. 1 (d) (e), once the training domain is very different from the test, these methods create hallucination contents and suffer from unstable results.", + "bbox": [ + 75, + 125, + 472, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above problems, we first offer a large-scale real-world shadow dataset, Shadowed Document 7K (SD7K), which contains more than 7k real-world and high-resolution shadow and shadow-free document image pairs with also the manually annotated shadow masks for related research, e.g., shadow detection [6]. Compared to existing document shadow removal datasets, SD7K includes various document types, e.g., manga, papers, and figures. Also, different from previous work which only considers one specific light environment, we use three distinct light sources to increase the diversity of the training samples. Besides the dataset, we also propose a robust network for high-resolution image processing specifically, named Frequency-aware Shadow Erasing Net (FSENet). In detail, inspired by the laplacian pyramid [2, 23], we divide the document image into two frequency portions, each with its own domain-specific properties. The low-frequency component is primarily responsible for illuminations or colors while high-frequency components are more related to content details. As for the low-frequency deshading module, we involve the transformer [7] based network to model the global illuminations alternation and a high-frequency restoration module based on cascaded aggregations for adaptive pixel enhancement. For high-frequency modules, we involve several dilated convolution-based texture recovery modules to restore the refined details similar to [6]. With the aforementioned", + "bbox": [ + 75, + 508, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "techniques, our model achieves the state-of-the-art performance of previous methods on two common small benchmarks and the proposed SD7K.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the main contribution as follows:", + "bbox": [ + 517, + 137, + 839, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide SD7K, a large-scale real-world dataset consisting of high-resolution shadow and the associated shadow-free images under various illumination conditions.", + "- We propose FSENet, a frequency-aware network with a carefully designed network structure to handle high-resolution document shadows.", + "- Both qualitative and quantitative results on all available public datasets indicate that the proposed FSENet performs favorably against state-of-the-art methods." + ], + "bbox": [ + 517, + 165, + 890, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 359, + 640, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Document Shadow Dataset. The collection of genuine shadow images and their corresponding shadow-free counterparts is a challenging task due to the varying lighting conditions. As a result, existing datasets are limited in size. The Bako dataset [1] comprises outdated, inaccessible images in tiny quantities. Similarly, Jung's dataset [16] is limited in size and favors light shadows. Kligler's dataset [17] emphasizes colorful symbols, which do not convey a significant amount of information. Among these datasets, the RDSRD dataset [24], while still modest in size, has a better quality in comparison. Different from the existing datasets, we propose a high-resolution and large-scale dataset SD7K that addresses the aforementioned problem, which can facilitate the research in document shadow removal.", + "bbox": [ + 496, + 386, + 890, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Natural Shadow Removal. Methods for natural shadow removal might be informative and applicable to document images. On the one hand, GAN-based methods such as STCGAN [31], use two stacked conditional GANs to jointly detect and eliminate shadows. Mask-ShadowGAN [15] extend the paired GAN to unpaired setting via re-formulated cycle-consistency constraints. To fully utilize the dataset, DHAN [6] synthesize pseudo shadow images and learned border artifacts-free images via a dual hierarchical aggregation network. SG-ShadowNet [29] studies the importance of normalization and learns the style representation of the non-shadow region to harmonize the shadow and shadow-free parts. On the other hand, CNN-based methods such as $\\mathrm{SP + M}$ Net [19] and $\\mathrm{SP + M + I}$ Net [20] utilize the physical linear transformation model to enhance the shadow region and image decomposition to reconstruct the shadow-free image. AEF-Net [8] predicts the exposure parameters of the shadowed regions and fused a series of overexposed shadow images with the original shadow image to restore the shadow region. ShadowFormer [10] is a transformer", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12450", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/55fd7bde2bee0219654fd79164bf136d5e2a20f37ad6f6b3339e8e514aa67ca3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTriplets #Entities #Mean ResolutionDocument TypeWith MaskLight Temperature Control
Bako [1]81N/AN/ATextNoNo
Kligler [17]300252424 × 2013TextNoNo
Jung [16]8787602 × 860TextNoNo
RDSRD [24]540N/AN/ATextYesN/A
SD7K7620350+2462 × 3699Text and pictureYesDifferent lamps & sunlight
", + "bbox": [ + 81, + 89, + 890, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of SD7K with other real-world datasets. The proposed dataset covers a large number of document shadow images and it is the only dataset that satisfies all important data features about document shadow currently.", + "bbox": [ + 75, + 195, + 892, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based network to exploit the context correlation between shadow and non-shadow regions.", + "bbox": [ + 75, + 250, + 468, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Document Image Shadow Removal. Traditional document shadow removal methods usually demand prior knowledge regarding shadow and document context. This includes estimating the background color [30, 32] and modeling shading and lighting effects [16, 27]. Deep learning-based methods are also designed to remove the shadows of documents. Currently, BEDSR-Net [24] stands as the state-of-the-art model in document shadow removal. It predicts global background color and encodes shadow position as an attention map, allowing for the effective removal of shadows. However, as discussed in our introduction, it needs to predict the approximate attention maps from the feature map, which is not suitable for the high resolution inputs.", + "bbox": [ + 75, + 280, + 470, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SD7K Dataset", + "text_level": 1, + "bbox": [ + 76, + 493, + 220, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Tab. 1 and Fig. 3 (a), there are only a limited number of datasets [16, 24, 1, 17] available for document shadow removal, which is typically small and comprises only hundreds of real-world samples. To train a network on such-scale datasets, synthesizing training samples via computer graphic engines is required [24]. However, the rendered samples still show different distributions from the real samples, making it difficult in real-world applications.", + "bbox": [ + 75, + 521, + 468, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome this limitation, we introduce a massive real-world dataset named SD7K for document shadow removal in terms of the occluder entities, the numbers, the resolution, the light source, etc. Below, we give the details of how to collect and process the dataset.", + "bbox": [ + 75, + 643, + 468, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we use a Canon EOS M6 camera to capture printed document images from real-world scenarios. To capture paired shadow and shadow-free images, we make all the light close and leave one light source. All the settings of the camera are also fixed, including focal length, aperture, exposure, and ISO, to avoid the influence of the camera algorithm. Then, we use a camera stand to hold the camera and make it shoot on the paper vertically as shown in Fig. 2 (a). To capture the paired shadow and shadow-free image without involving the changes caused by the manual presses, as in Fig. 2 (b), we use a remote control to capture the photo between the shadowed and shadow-free paper by removing", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92ade6c40849ca1920b29d8c785b2725bd4ca06b8850112fa8e87a8c1751b336.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 246, + 627, + 392 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ba9ec9be2845cbbd327b3554db80b7b1bfcac9937ec554fe8614a442c33bb9c4.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 632, + 246, + 759, + 392 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0780e631cfb7cd38c83c6109a0d5418dbf1aebf3261d2d13fe91a352a83757af.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 763, + 246, + 890, + 392 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3971a0ab029c7bea5f66208b8a97e4f9b59c8aa6086467d9e60a5779c1df6d87.jpg", + "image_caption": [ + "(a) Quantitative comparison", + "Figure 3. Data distribution of SD7K and quantitative comparison across all document shadow datasets." + ], + "image_footnote": [], + "bbox": [ + 504, + 477, + 696, + 580 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3851803b58f5cbb5b8ff55f23ae8252460da19811130a35483792220ad985e44.jpg", + "image_caption": [ + "Figure 2. (a) our data acquisition setup for constructing the dataset. (b) a remote phone control shutter where the user only needs to click a button to capture a pair of shadow-free/shadow images. (c) some showcases of our occluders.", + "(b) Light distribution" + ], + "image_footnote": [], + "bbox": [ + 700, + 476, + 888, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the occluder. We also take the pairs as short as possible to avoid the light difference between captures.", + "bbox": [ + 498, + 643, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further enhance diversity, we alternated the positions of the document and occluder after capturing a few batches of documents. Overall, we use over 30 types of occluders along with more than 350 documents to contribute to the dataset, some of the occluders are shown in Fig. 2 (c). These occluders have the shape of both regular and irregular forms, which provides adequate coverage for various situations, we calculate the shadow area percentage of the whole dataset, which is $41.38\\% \\pm 13.58\\%$ , indicating that it is sufficiently divergent for our shadow removal task.", + "bbox": [ + 496, + 674, + 892, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also consider the fluctuations in sunlight intensity and random placement of shadows, which will be significantly influenced by the light variance. To simulate the results in different light sources, we introduce three types of light sources in our dataset for collection, including reg", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12451", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/743ae3ba46e21e912060a27d8559cad02da70f55ec3ef8cf72cef89640311f1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 205, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7df2cb277b10a39bcbfae849c6eb66b057f287807b34ff6085eb1c29ab4ee82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 336, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5b8298c76ad7ea611cbc46abc3b2407a59330f2cbb769867eca96e29938d215e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 88, + 468, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a6902a37fa99561e7933b2991cabfcbd781acae515ef2540f19fa28de99871d3.jpg", + "image_caption": [ + "Cool light" + ], + "image_footnote": [], + "bbox": [ + 80, + 234, + 205, + 381 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d54e187e74aa06fe4b74897d7c48018bf5e9fe8a33e2545d8e8e97630f776ac.jpg", + "image_caption": [ + "Warm light" + ], + "image_footnote": [], + "bbox": [ + 210, + 236, + 336, + 381 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/334c8217b4404318abfb486c6f029e1369960eb2f68e018ebd3258fa6b717bdf.jpg", + "image_caption": [ + "Sunlight", + "Figure 4. Example shadow and shadow-free images from SD7K." + ], + "image_footnote": [], + "bbox": [ + 341, + 234, + 468, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cular sunlight, a stationary 80W LED yellow light having a color temperature of $3000\\mathrm{K}$ , and a stationary 80W LED white light having a color temperature of $6000\\mathrm{K}$ . The light distribution of the whole dataset is shown in Fig. 3 (b).", + "bbox": [ + 75, + 433, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For data cleaning purposes, we first remove images with noticeable motion distortion or lack of focus. Since some photo still contains outliers and the context is duplicated, we then review the photographs in groups and remove them. Finally, we crop the areas outside of the margins of the document, resulting in a total of 7620 samples, where some of the samples are shown in Fig. 4. We also use a threshold-based method to extract the mask of the training pair since the shadow mask will also benefit related tasks.", + "bbox": [ + 75, + 494, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Frequency-Aware Shadow Erasing Net", + "text_level": 1, + "bbox": [ + 76, + 647, + 426, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since real-world document shadow removal needs huge GPU memory requirements, our network initially divides the input image into different frequency components using the Laplacian Pyramid (LP) similar to [23] to reduce the computational budgets. As shown in Fig. 5, we propose a powerful frequency-aware shadow removal network named FSENet by designing different modules to handle the high-frequency and low-frequency computation individually. After training, the multi-frequency components are mixed to produce the final results. In the rest of this section, we first give the details of the Laplacian pyramid in Sec. 4.1. Then, in Sec. 4.2 and Sec. 4.3, we give the details of the low-frequency deshading module and high-frequency restoration module to precisely modify the color details and the high-frequency details individually. Finally, we give the", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "loss functions in Sec. 4.4.", + "bbox": [ + 500, + 90, + 671, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Laplacian Pyramid and Method Overview", + "text_level": 1, + "bbox": [ + 500, + 114, + 859, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Laplacian Pyramid (LP) [2] plays an important role in image processing. By decomposing and restoring images in a lossless manner, LP provides a rapid processing speed and is well-suited for dealing with high-resolution images. Inspired by [23] for image translation, we implement LP for document shadow removal and train the network on multifrequency.", + "bbox": [ + 496, + 138, + 890, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In detail, given an image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , the $D$ layers LP contains several high-frequency components $[L_0, \\dots, L_{D-1}]$ and a low-frequency component $L_D \\in \\mathbb{R}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times 3}$ , where $L_D$ is $D$ times downsample of $I$ . Each high-frequency component can be described as $L_i = I_i - \\mathrm{PyrUp}(\\mathrm{PyrDown}(I_i))$ , where $I_i$ is the $i$ -th down-sampled image and $\\mathrm{PyrUp}$ and $\\mathrm{PyrDown}$ stand for upsampling and downsampling, respectively. As illustrated in Fig. 5, the $L_1$ and $L_2$ include the high-frequency component edges and textures, while the low-frequency component $L_3 = I_3$ majorly contains the color information. For reversible reconstruction, the final $\\hat{I}$ can be obtained by reversed update the equation of $\\hat{I}_{D-1} = \\mathrm{PyrUp}(\\hat{I}_D) + \\hat{L}_{D-1}$ from $D$ to $0$ , where $\\hat{I}_D$ is the learned low-frequency features and $\\hat{L}_{D-1}$ is the learned high-frequency features from the proposed network. We give detailed instructions for each network below.", + "bbox": [ + 496, + 244, + 892, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Low-Frequency Deshading Module", + "text_level": 1, + "bbox": [ + 498, + 498, + 805, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After getting the low-frequency components, we use several transformer-based blocks to reduce the color and illumination distortions of the shadow images. Compared to CNNs, transformers are superior at modeling non-local self-similarity and long-range dependencies. As shown in Fig. 5, we implement a transformer-based module called Dimension-Aware Transformer (DAT) inspired by [33] with a lightweight UNet-like sub-network [3] for low-frequency feature modeling.", + "bbox": [ + 496, + 521, + 890, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given a low-frequency component $L_{3} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 3}$ of input image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , we first feed $L_{3}$ into three sequential DAT blocks to extract rich features. Each DAT block is built via a sequential height-attention and width-attention to learn the feature in different aspects individually and a convolution-based layer for local feature merging, where we denote the intermediate features of each DAT block as $F1, F2, F3 \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}$ . Please refer to the original [33] for further details of the height-attention and width-attention.", + "bbox": [ + 496, + 657, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then feed these features $F1, F2, F3$ into a Tri-layer Attention Alignment (TAA) block for feature remixing. Given concatenated features $F_{\\mathrm{in}} \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4} \\times C} = \\text{Concat}([F1, F2, F3])$ , where $N = 3$ in this case, $F_{\\mathrm{in}}$ is first reshaped to $\\hat{F}_{\\mathrm{in}} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 3C}$ . Then, we send $\\hat{F}_{\\mathrm{in}}$ to a $1 \\times 1$ convolution layer to aggregate pixel-level cross", + "bbox": [ + 496, + 808, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12452", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/809596f7670c898143aa330172b16331ccdde3b95a768080cdcafdb18eb9ff71.jpg", + "image_caption": [ + "Figure 5. The network structure of our proposed FSENet. Following [23], given a high-resolution image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , we first use Laplacian Pyramid (e.g. Depth $D = 2$ in this case) to decompose the images to multiple frequency components. The Black arrows: The low-frequency part $L_{3} \\in \\mathbb{R}_{\\frac{H}{2D}}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times C}$ are refined to $L_{3}' \\in \\mathbb{R}_{\\frac{H}{2D}}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times C}$ utilizing DAT, TAA and DFE block. Cyan arrows: For the high-frequency part $L_{2} \\in \\mathbb{R}_{\\frac{H}{2D-1}}^{\\frac{H}{2D-1} \\times \\frac{W}{2D-1} \\times C}$ , a contour $C_{L_{2}} \\in \\mathbb{R}_{\\frac{H}{2D-1}}^{\\frac{H}{2D-1} \\times \\frac{W}{2D-1} \\times 1}$ is learned to bridge the low-frequency and high-frequency features. Red arrows: For the remaining components with higher frequency, the learned contour is successively upsampled and refined using the proposed SPP and TRM." + ], + "image_footnote": [], + "bbox": [ + 91, + 93, + 879, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "channel context, followed by three $3 \\times 3$ depth-wise convolution layers to generate query $\\mathbf{Q}$ , key $\\mathbf{K}$ , and value $\\mathbf{V}$ , individually. The $\\mathbf{Q}, \\mathbf{K}$ are then reshaped to $\\hat{\\mathbf{Q}} \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\frac{W}{4} C}$ and $\\hat{\\mathbf{K}} \\in \\mathbb{R}^{\\frac{H}{4} \\frac{W}{4} C \\times N}$ respectively to calculate an attention matrix $\\mathbf{A} \\in \\mathbb{R}^{\\mathbf{N} \\times \\mathbf{N}}$ . Next, we reshape the $\\mathbf{V}$ to $\\hat{\\mathbf{V}} \\in \\mathbb{R}^{\\frac{H}{4} \\frac{W}{4} C \\times N}$ to multiply $\\hat{\\mathbf{V}}$ and $\\mathbf{A}$ , and the input feature $\\hat{F}_{\\mathrm{in}}$ is added to generate the output feature $F4$ , which can be described as:", + "bbox": [ + 75, + 455, + 472, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF 4 = \\operatorname {C o n v} _ {1 \\times 1} L A (\\hat {\\mathbf {Q}}, \\hat {\\mathbf {K}}, \\hat {\\mathbf {V}}) + \\hat {F} _ {\\mathrm {i n}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 582, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Conv_{1\\times 1}$ represents a convolution layer with the kernel of $1\\times 1$ . LA indicates Layer Attention [33], which can be written as:", + "bbox": [ + 75, + 606, + 468, + 651 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL A (\\hat {\\mathbf {Q}}, \\hat {\\mathbf {K}}, \\hat {\\mathbf {V}}) = \\hat {\\mathbf {V}} \\operatorname {s o f t m a x} (\\hat {\\mathbf {Q}} \\hat {\\mathbf {K}} / \\alpha), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 656, + 468, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the temperature of the attention. Please refer to the supplementary material for more details of the TAA.", + "bbox": [ + 75, + 681, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting $F4$ , as shown in Fig. 5, We also use a UNet-like [3, 35] sub-network built by Deeper Feature extraction (DFE) blocks to learn multi-scale features. Each DFE block begins with a LayerNorm to stabilize the training procedure, followed by two convolutions and a deformable convolution. Then, it utilizes SimpleGate [3] to increase the diversity of the features by dividing the features along the channel dimension into two parts and multiplying them together. We also utilize Simplified Channel Attention (SCA) similar to [3] to capture global information in a computationally efficient manner, which can be denoted as:", + "bbox": [ + 75, + 710, + 470, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS C A (\\mathbf {X}) = \\mathbf {X} * \\mathbf {W} \\operatorname {p o o l} (\\mathbf {X}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}$ represents a fully-connected layer, pool represents the global average pooling procedure that combines spatial data into channels, and $*$ represents a channel-wise multiplication. The skip-connections [13] are connected before the LayerNorm to the output similar to [3]. We also give a detailed figure of the DFE block in the supplementary materials.", + "bbox": [ + 496, + 455, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the refined feature $\\hat{F}^4$ , it then passes through three transformer blocks and a TAA Block to produce the enhanced features for image reconstruction using DAT and TAA blocks similar to the feature extractions of $F1, F2, F3$ . Finally, a convolution layer with the kernel size $3 \\times 3$ is performed on the enhanced features to produce the low-frequency component $L_3' \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}$ .", + "bbox": [ + 496, + 561, + 892, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. High-Frequency Restoration Module", + "text_level": 1, + "bbox": [ + 498, + 680, + 821, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting the low-frequency features, we also design novel components to recover the high-frequency details of the learned features. In detail, we first upsample $L_{3}$ and its corresponding result $L_{3}^{\\prime}$ to match the size of $L_{2}$ , then we learn a contour on the concatenation of $[up(L_3), up(L_3'), L_2]$ via several residual blocks [13], where $up(\\cdot)$ means bilinear upsample operation, as illustrated in Fig. 5. This is a straightforward method for optimizing global correction compared to normal RGB images because the bands of the high-frequency feature map are very sparse, allowing us to reduce the volume of calculations as discussed in [23]. This contour will be expanded progressively into a collection of contours to match", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12453", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/811c2e3d6a36dc1b4e14644b5d54f44f92fdb06e0c42fe4e6acc49d9cfb9d430.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 87, + 236, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74164336ac43927d45013c270e30b5f3f13d72b3823f2c0d4c9a5399a3b35a6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 88, + 400, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd8c0bafec9d9c4762f2c68ec2b00102f6c001c5ac12f67fc54020df493447e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 88, + 563, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51233e8dd291e6af8272dfd95225800d5c7e629b65b83acaad949dbfaaada90b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 88, + 728, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/052222441240f32fe624d20415976a38a0d77f4e07765e852893031c63a9864f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 88, + 892, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d95750c311867ea181e1d9c44d6a3814ec267b64402fc6674d657d74dc9ee0ff.jpg", + "image_caption": [ + "Figure 6. Qualitative results of the methods comparison in high resolution, the first row represents the results on the Kligler dataset and the second row depicts the results on the SD7K dataset." + ], + "image_footnote": [], + "bbox": [ + 76, + 287, + 236, + 526 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bba7869dc5a5072f88b969330c0b4fb6118831728d8c9dc73372cefb2261debf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 290, + 398, + 527 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/958643130421a86ba976679e75915cfc439838dfd7af486159017b0bd5fc53de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 290, + 563, + 527 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a4ceb403dec194759a11a0e52a790c814b46fae5fdd82fb1e12cf7487b705ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 290, + 728, + 527 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f57eb4477adf4279f8d6fb43a845c9754ed2ec17d3f2a8eca2e7e9449c6faabe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 290, + 892, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the remaining high-frequency components. As illustrated in Fig. 5, given the high-frequency component $L_{2}$ and the contour $C_{L_2}$ , the output $L_{2}^{\\prime}$ can be written as $L_2^\\prime = L_2\\otimes C_{L_2}$ where $\\otimes$ represents the pixel-wise multiplication.", + "bbox": [ + 75, + 585, + 468, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After multiplication, the Texture Recovery Module (TRM) is then designed to improve elaboration, inspired by the successful natural shadow removal methods [6, 37], which contain a series of the dilated convolutions and attentive aggregation nodes to recover the detailed features.", + "bbox": [ + 75, + 652, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In detail, to increase the reception fields on the high-resolution map, we incorporate dilated convolution [38]. Dilated convolution inserts spaces between the kernel values, effectively increasing the receptive field of the filter without losing resolution or coverage over the input. The amount of dilation is controlled by a dilation rate parameter. For a dilation rate of 1, the layer behaves like a regular convolution. For rates larger than 1, the filter is effectively widened to have a larger field of view. This allows the network to incorporate a wider context and capture features at different scales in a computationally effi", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cient manner compared to using very large filters. As for Attentive Aggregation Node, to re-weight the significance of feature maps, each aggregation node employs a squeeze-and-excitation block [14] to reweight the feature. A $3 \\times 3$ convolution layer is then employed to compress the features and match the original channels. Finally, a spatial pooling pyramid (SPP) [12] is introduced at the conclusion of the TRM to facilitate the remixing of multi-context features. After getting the learned high-frequency of each layer, we use the method introduced in Sec. 4.1 to recover the image and calculate the supervision.", + "bbox": [ + 496, + 585, + 893, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Loss functions", + "text_level": 1, + "bbox": [ + 500, + 768, + 648, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training, we utilize the loss functions of SmoothL [9] $\\mathcal{L}_{l1}^{smooth}$ and SSIM [34] $\\mathcal{L}_{SSIM}$ between the target image and the reconstructed image from LP, both two losses are helpful to preserve the details of the image, which can be represented as: $\\mathcal{L}_{total} = \\mathcal{L}_{l1}^{smooth} + \\lambda * \\mathcal{L}_{SSIM}$ , where we set the weight of $\\mathcal{L}_{SSIM} \\lambda$ to 0.4 empirically.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12454", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4a1277b95bf2f12107e3baa78d3db619343a560f756892882c3ac1ce7fa92d6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSD7KParam (M)
512 × 5121024 × 10242462 × 3699 (Full Size)
PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓
Input15.950.8944.09N/A15.950.9044.09N/A15.940.9144.14N/AN/A
Wang et al. [32]15.310.8247.88>1015.310.8547.88>3015.290.8647.95>60N/A
Wang et al. [30]13.320.6867.48>1013.320.7167.48>3013.260.7468.07>60N/A
Shah et al. [27]9.890.7186.35>109.900.7686.35>309.880.7986.46>60N/A
Jung et al. [16]19.860.9226.76>519.860.9326.76>1019.820.9226.86>30N/A
MaskGAN [15]24.820.8715.430.9624.820.8515.432.0724.670.8615.727.5928.29
DHAN [6]25.610.8514.270.2125.60*0.83*14.27*N/A25.42*0.85*14.60*N/A27.82
AEFNet [8]24.180.9516.830.1123.00*0.90*19.14*N/A22.94*0.90*19.28*N/A391.10
BEDSR-Net [24]21.500.9030.520.8519.81*0.85*33.03*N/A19.74*0.86*33.18*N/A29.44
ShadowFormer [10]23.710.9017.540.1522.69*0.84*19.65*N/A22.64*0.87*19.79*N/A11.35
BMNet [39]24.860.8015.590.1824.84*0.79*15.67*N/A24.70*0.83*15.96*N/A2.11
Ours28.690.979.980.2428.680.979.980.7528.670.9610.007.9329.34
", + "bbox": [ + 76, + 88, + 893, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of the ablation study on three datasets in different resolutions. The best result is highlighted in red and bold. The asterisk (*) indicates that the result is obtained from a low-resolution result $(512\\times 512)$ and then improved using a super-resolution technique [21] to address an out-of-memory error.", + "bbox": [ + 75, + 276, + 892, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 342, + 209, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 367, + 294, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the performance of different network structures with different resolutions on the proposed SD7K dataset, where the batch size is set to 4. We also give more experiments on the low-resolution of $512 \\times 512$ of Jung [16], Kligler [17], and SD7K for fair comparison following previous methods, we give more details in the supplementary materials. To further utilize the dataset, we also incorporate a simple shadow synthesis pipeline to further increase the diversity of the training process. We investigate several potential options, including alpha composition [26] and Perlin noise generation [11] for comparison, and we empirically choose alpha composition to synthesize shadows through the shadow-free image and the mask on the fly via random alpha values in the range between [0.2, 0.7].", + "bbox": [ + 75, + 392, + 468, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for the evaluation metrics, we employ three widely-used metrics: Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity (SSIM). While PSNR and SSIM are commonly used in image restoration evaluation [3, 5, 22], RMSE is the most important and frequently used metric specifically for shadow removal [6, 10].", + "bbox": [ + 75, + 604, + 468, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 723, + 452, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As presented in Tab. 2, we compared FSENet with both traditional and learning-based models of various resolutions using publicly available implementations of these methods, where the proposed method outperforms all other methodologies in most metrics.", + "bbox": [ + 75, + 748, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, our analysis reveals that traditional methods, such as [16, 27], rely on physics-based assumptions that are not always correct, resulting in unstable performance across different datasets as shown in Fig. 6. Deep learning models, on the other hand, are highly dependent on", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the dataset magnitude, and their performance is relatively consistent across different scenarios.", + "bbox": [ + 498, + 342, + 890, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also find that some deep learning models that are effective in low-resolution shadow removal do not scale well to high-resolution documents. They either compromise the aspect ratio of the document or lead to out-of-memory errors. While traditional methods can still remove shadows, they are time-consuming, requiring pixel-by-pixel calculations of illumination difference [27] and background estimation [30]. At present, only Mask-ShadowGAN [15] and the proposed FSENet can effectively remove document shadows in high-resolution scenarios. Despite having more parameters and requiring more time to produce results, FSENet outperforms other methods quantitatively.", + "bbox": [ + 496, + 375, + 892, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 571, + 663, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As presented in Tab. 3 and Fig. 7, we conduct our ablation studies at the original high-resolution documents, examining the effectiveness of several components in our algorithm. These components include the depth of the Laplacian Pyramid, the efficiency of TRM, DFE, and DAT modules, and the effectiveness of our shadow synthesis pipeline.", + "bbox": [ + 496, + 595, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first evaluate the depth of the proposed Laplacian Pyramid. From Tab. 3, with the depth of the Laplacian Pyramid increases (Our w/ Depth=3), the magnitude of the low-frequency components decreases, affecting the low-frequency deshadowing and the reconstructed outcome. Consequently, the model's performance progressively declines as the depth increases. We then evaluate the performance of the proposed DFE and DAT modules, which are also responsible for low-frequency deshading, primarily influencing the pixel-wise variation of the final result. Although each module individually contributes to enhancing shadow elimination performance, their combined network produces the best results. As for the high-frequency modules, the TRM module handles high-frequency components", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12455", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8bf1ee1001d2333a56685fcae959ccd998592f9e8ce7d03160b259847ea7b4c.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 277, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2b71b2473a74fc47d21e7090e0faa4a405a304f178d602bc752a83e750e9bd7.jpg", + "image_caption": [ + "Ours w/ Depth=3" + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 482, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d49ad78ca0f402fd9156691721451865fa442f6f13852be1813a0f1ae517f36.jpg", + "image_caption": [ + "Ours w/o TRM" + ], + "image_footnote": [], + "bbox": [ + 488, + 90, + 687, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd754a23ac84ec75a9c029bc4c2474eb33d8b3dfd8d7fc03d4f06965a73de2dd.jpg", + "image_caption": [ + "Ours w/o DFE" + ], + "image_footnote": [], + "bbox": [ + 694, + 90, + 892, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/697bd36004825364ebaf1d0bbe529a15062849b4ca8ff9e707d54cd83eec9b19.jpg", + "image_caption": [ + "Ours w/o Synthesis" + ], + "image_footnote": [], + "bbox": [ + 78, + 219, + 277, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/846cfcdd1283cd3bf5d968fec9cf8c7f789139c23a13832db5edd4e858f24189.jpg", + "image_caption": [ + "Ours w/o DAT" + ], + "image_footnote": [], + "bbox": [ + 282, + 220, + 482, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b6b38b3df1b37d87a291cc928919d6bbdfe3c85f3dd3ffd75223e59c3059fe0.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 488, + 220, + 687, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70e3eb1319516a0897be357754bcef8fa5beab42bc78c4764d06a2ffabcea459.jpg", + "image_caption": [ + "Target", + "Figure 7. Qualitative results of the ablation studies on different modules of the proposed FSENet." + ], + "image_footnote": [], + "bbox": [ + 694, + 219, + 892, + 328 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/51fd5fe35ad792eb499cc1e72e16ab0581922887b076327d49b0402de38a0d6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablation StudyJung (Original)Kliger (Original)SD7K (Original)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
Input12.930.7961.3413.250.8156.7515.940.9144.14
Ours w/ Depth=320.970.8323.1926.330.9113.6620.830.9324.02
Ours w/o TRM19.500.8127.6027.100.9212.7818.440.8635.92
Ours w/o DFE21.900.8421.0225.160.9214.8919.700.8829.07
Ours w/o DAT22.680.8519.2822.700.8820.1919.030.8732.94
Ours w/o Synthesis22.210.8521.3324.790.9215.5422.940.9321.93
Ours Full23.600.8517.5628.980.9310.6028.670.9610.00
", + "bbox": [ + 96, + 383, + 870, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative results of the ablation study on three datasets in high-resolution. The best result is highlighted in red and bold.", + "bbox": [ + 76, + 534, + 854, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "such as contours and edges, thus contributing significantly to the overall efficacy of our algorithm. Finally, we also evaluate the importance of the proposed shadow synthesis pipeline, where w/o shadow synthesis reduces the performance a lot, especially on the two small datasets.", + "bbox": [ + 75, + 573, + 468, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Limitations", + "text_level": 1, + "bbox": [ + 76, + 657, + 202, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our model is capable of removing shadows in high-resolution images, the current network is not able to run on real-time or on edge devices, such as mobile phones, due to the larger parameters and heavy GPU computations. In the future, we will explore different approaches to further reduce the required parameters, making it usable on smart devices.", + "bbox": [ + 75, + 680, + 468, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 800, + 194, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we tackle the problem of removing shadows from high-resolution documents. To achieve this goal, we consider the contribution of the dataset and the network. As for the dataset, we gather a high-resolution dataset SD7K, which is comprised of over 7000 triplets of real-", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "world document images with distinct characteristics under various illumination conditions. This dataset is an order of magnitude greater than previous datasets and allows us to validate our approach in a more diverse and comprehensive manner. As for the network structure, we propose FSENet, a frequency-aware method to remove shadows in higher resolution via frequency decomposition. To separate the high-frequency contours from the low-resolution content, we employ a Laplacian Pyramid and a series of meticulously designed attention blocks, enabling each component to be learned separately. Powered by our dataset and network, our method achieves a new state-of-the-art for document shadow removal. We believe both the network structure and the dataset will benefit the community for research on this topic.", + "bbox": [ + 496, + 573, + 892, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by the University of Macau under Grant MYRG2022-00190-FST, in part by the Science and Technology Development Fund, Macau SAR, under Grant 0034/2019/AMJ, Grant 0087/2020/A2 and Grant 0049/2021/A.", + "bbox": [ + 496, + 825, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12456", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Steve Bako, Soheil Darabi, Eli Shechtman, Jue Wang, Kalyan Sunkavalli, and Pradeep Sen. Removing shadows from images of documents. In ACCV, pages 173-183. Springer, 2016. 2, 3", + "[2] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 2, 4", + "[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33. Springer, 2022. 4, 5, 7", + "[4] Xuhang Chen, Xiaodong Cun, Chi-Man Pun, and Shuqiang Wang. Shadocnet: Learning spatial-aware tokens in transformer for document shadow removal. In IEEE ICASSP, pages 1-5. IEEE, 2023. 2", + "[5] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 7", + "[6] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2, 6, 7", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2", + "[8] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In CVPR, pages 10571-10580, 2021. 1, 2, 7", + "[9] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 6", + "[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 2, 7", + "[11] John C Hart. Perlin noise pixel shaders. In Proceedings of the ACM SIGGRAPH/EUROGRAPHICS workshop on Graphics hardware, pages 87-94, 2001. 7", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2, 6", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 5", + "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 6", + "[15] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2, 6, 7", + "[16] Seungjun Jung, Muhammad Abul Hasan, and Changick Kim. Water-filling: An efficient algorithm for digitized document shadow removal. In Asian Conference on Computer Vision, pages 398-414. Springer, 2018. 1, 2, 3, 6, 7" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Netanel Kligler, Sagi Katz, and Ayellet Tal. Document enhancement using visibility detection. In CVPR, pages 2374-2382, 2018. 1, 2, 3, 7", + "[18] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, August 2020. 2", + "[19] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, pages 264-281. Springer, 2020. 2", + "[20] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 2", + "[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017. 7", + "[22] Jingtang Liang, Xiaodong Cun, Chi-Man Pun, and Jue Wang. Spatial-separated curve rendering network for efficient and high-resolution image harmonization. In European Conference on Computer Vision, pages 334-349. Springer, 2022. 7", + "[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In CVPR, pages 9392-9400, 2021. 2, 4, 5", + "[24] Yun-Hsuan Lin, Wen-Chin Chen, and Yung-Yu Chuang. Bedsr-net: A deep shadow removal network from a single document image. In CVPR, pages 12905-12914, 2020. 2, 3, 7", + "[25] Zhihao Liu, Hui Yin, Yang Mi, Mengyang Pu, and Song Wang. Shadow removal by a lightness-guided network with training on unpaired data. IEEE Transactions on Image Processing, 30:1853-1865, 2021. 1, 2", + "[26] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In SIGGRAPH, pages 231–242, 1998. 7", + "[27] Vatsal Shah and Vineet Gandhi. An iterative approach for shadow removal in document images. IEEE ICASSP, pages 1892-1896, 2018. 1, 3, 7", + "[28] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1", + "[29] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV, pages 361-378. Springer, 2022. 2", + "[30] Bingshu Wang and CL Philip Chen. An effective background estimation method for shadows removal of document images. In IEEE International Conference on Image Processing (ICIP), pages 3611-3615. IEEE, 2019. 1, 3, 7", + "[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12457", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Jian-Ren Wang and Yung-Yu Chuang. Shadow removal of text document images by estimating local and global background colors. IEEE ICASSP, pages 1534-1538, 2020. 1, 3, 7", + "[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. AAAI, 2023. 4, 5", + "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17683-17693, 2022. 2, 5", + "[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In SIGKDD, pages 1192-1200, 2020. 1", + "[37] Yimin Xu, Mingbao Lin, Hong Yang, Ke Li, Yunhang Shen, Fei Chao, and Rongrong Ji. Shadow-aware dynamic convolution for shadow removal. arXiv preprint arXiv:2205.04908, 2022. 6", + "[38] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. *ICLR*, 2015. 6", + "[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, June 2022. 7" + ], + "bbox": [ + 78, + 90, + 468, + 501 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12458", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_model.json b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..571b1e07273f78991a0d38e846d230f23938ed8c --- /dev/null +++ b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.203, + 0.761, + 0.24 + ], + "angle": 0, + "content": "Zinuo Li* Xuhang Chen* Chi-Man Pun† Xiaodong Cun† University of Macau" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.474, + 0.685 + ], + "angle": 0, + "content": "Shadows often occur when we capture the document with casual equipment, which influences the visual quality and readability of the digital copies. Different from the algorithms for natural shadow removal, the algorithms in document shadow removal need to preserve the details of fonts and figures in high-resolution input. Previous works ignore this problem and remove the shadows via approximate attention and small datasets, which might not work in real-world situations. We handle high-resolution document shadow removal directly via a larger-scale real-world dataset and a carefully-designed frequency-aware network. As for the dataset, we acquire over 7k couples of high-resolution \\((2462 \\times 3699)\\) images of real-world documents pairs with various samples under different lighting circumstances, which is 10 times larger than existing datasets. As for the design of the network, we decouple the high-resolution images in the frequency domain, where the low-frequency details and high-frequency boundaries can be effectively learned via the carefully designed network structure. Powered by our network and dataset, the proposed method shows a clearly better performance than previous methods in terms of visual quality and numerical results. The code, models, and dataset are available at https://github.com/CXH-Research/DocShadow-SD7K." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.21, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.834 + ], + "angle": 0, + "content": "When occluders obstruct the light source, shadows are cast on the paper. These shadows reduce the readability and hinder subsequent text-related intelligent tasks such as document enhancement, optical character recognition (OCR), key information extraction, and semantic entity labeling [17, 36, 28], etc." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.835, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Removing the cast shadow of the digital copy has a long need in the real world and also there are a lot of meth-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.273, + 0.629, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.416, + 0.629, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.484, + 0.596, + 0.499 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.289, + 0.758, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.416, + 0.761, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.484, + 0.76, + 0.499 + ], + "angle": 0, + "content": "(b) Shah et al. [27]" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.273, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.416, + 0.892, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.484, + 0.879, + 0.499 + ], + "angle": 0, + "content": "(c) Fu et al. [8]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.505, + 0.628, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.648, + 0.629, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.716, + 0.623, + 0.731 + ], + "angle": 0, + "content": "(d) Liu et al. [25]" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.505, + 0.76, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.648, + 0.761, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.716, + 0.753, + 0.731 + ], + "angle": 0, + "content": "(e) Hu et al. [15]" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.505, + 0.892, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.648, + 0.892, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.716, + 0.856, + 0.731 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.737, + 0.894, + 0.807 + ], + "angle": 0, + "content": "Figure 1. Visual results of input document shadow image (a), classic method (b), supervised method (c), weakly-supervised method (d), unsupervised method (e) and ours (f). Our model removes the shadow while preserving the original document's content and aspect ratio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "ods, including the traditional methods that applied physics-based illumination models [16, 30, 32]. In reality, these approaches are significantly constrained due to the fact that the assumptions utilized in illumination models [27] are often too rigid to account for real-world shadow. Thus, they may" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.225, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.225, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12449" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "incorrectly compute the exposure and reflectance as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.504 + ], + "angle": 0, + "content": "Recent methods apply deep learning for this task since it has achieved remarkable progress in natural shadow removal [6] and other computer vision tasks [12, 13, 35]. However, several obstacles make it challenging to transfer previous natural shadow removal methods to the document domain. On the one hand, in contrast to natural shadow removal, documents are presented by high-resolution images in the real world, which needs to preserve fine-grained information such as fonts and figures. However, the majority of document shadow removal methods [24, 4] are only designed for relatively low-resolution images and is hard to handle high-resolution situations directly since they are relying on the low-resolution approximation of attention for guidance. On the other hand, there are no large-scale and high-resolution datasets designed specifically for document shadow removal at this time. Consequently, given limited image pairs for training, existing methods usually result in different artifacts (as shown in Fig. 1 (c)). Although weakly-supervised [18] and unsupervised [15, 25] shadow removal methods alleviate the requirements for data, they hold a strong assumption of the statistical similarity between shadowed and unshadowed domains. As illustrated in Fig. 1 (d) (e), once the training domain is very different from the test, these methods create hallucination contents and suffer from unstable results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To solve the above problems, we first offer a large-scale real-world shadow dataset, Shadowed Document 7K (SD7K), which contains more than 7k real-world and high-resolution shadow and shadow-free document image pairs with also the manually annotated shadow masks for related research, e.g., shadow detection [6]. Compared to existing document shadow removal datasets, SD7K includes various document types, e.g., manga, papers, and figures. Also, different from previous work which only considers one specific light environment, we use three distinct light sources to increase the diversity of the training samples. Besides the dataset, we also propose a robust network for high-resolution image processing specifically, named Frequency-aware Shadow Erasing Net (FSENet). In detail, inspired by the laplacian pyramid [2, 23], we divide the document image into two frequency portions, each with its own domain-specific properties. The low-frequency component is primarily responsible for illuminations or colors while high-frequency components are more related to content details. As for the low-frequency deshading module, we involve the transformer [7] based network to model the global illuminations alternation and a high-frequency restoration module based on cascaded aggregations for adaptive pixel enhancement. For high-frequency modules, we involve several dilated convolution-based texture recovery modules to restore the refined details similar to [6]. With the aforementioned" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "techniques, our model achieves the state-of-the-art performance of previous methods on two common small benchmarks and the proposed SD7K." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.138, + 0.841, + 0.152 + ], + "angle": 0, + "content": "We summarize the main contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.166, + 0.892, + 0.225 + ], + "angle": 0, + "content": "- We provide SD7K, a large-scale real-world dataset consisting of high-resolution shadow and the associated shadow-free images under various illumination conditions." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.24, + 0.892, + 0.284 + ], + "angle": 0, + "content": "- We propose FSENet, a frequency-aware network with a carefully designed network structure to handle high-resolution document shadows." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.299, + 0.892, + 0.344 + ], + "angle": 0, + "content": "- Both qualitative and quantitative results on all available public datasets indicate that the proposed FSENet performs favorably against state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.166, + 0.892, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.642, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Document Shadow Dataset. The collection of genuine shadow images and their corresponding shadow-free counterparts is a challenging task due to the varying lighting conditions. As a result, existing datasets are limited in size. The Bako dataset [1] comprises outdated, inaccessible images in tiny quantities. Similarly, Jung's dataset [16] is limited in size and favors light shadows. Kligler's dataset [17] emphasizes colorful symbols, which do not convey a significant amount of information. Among these datasets, the RDSRD dataset [24], while still modest in size, has a better quality in comparison. Different from the existing datasets, we propose a high-resolution and large-scale dataset SD7K that addresses the aforementioned problem, which can facilitate the research in document shadow removal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Natural Shadow Removal. Methods for natural shadow removal might be informative and applicable to document images. On the one hand, GAN-based methods such as STCGAN [31], use two stacked conditional GANs to jointly detect and eliminate shadows. Mask-ShadowGAN [15] extend the paired GAN to unpaired setting via re-formulated cycle-consistency constraints. To fully utilize the dataset, DHAN [6] synthesize pseudo shadow images and learned border artifacts-free images via a dual hierarchical aggregation network. SG-ShadowNet [29] studies the importance of normalization and learns the style representation of the non-shadow region to harmonize the shadow and shadow-free parts. On the other hand, CNN-based methods such as \\(\\mathrm{SP + M}\\) Net [19] and \\(\\mathrm{SP + M + I}\\) Net [20] utilize the physical linear transformation model to enhance the shadow region and image decomposition to reconstruct the shadow-free image. AEF-Net [8] predicts the exposure parameters of the shadowed regions and fused a series of overexposed shadow images with the original shadow image to restore the shadow region. ShadowFormer [10] is a transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12450" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.189 + ], + "angle": 0, + "content": "
DatasetTriplets #Entities #Mean ResolutionDocument TypeWith MaskLight Temperature Control
Bako [1]81N/AN/ATextNoNo
Kligler [17]300252424 × 2013TextNoNo
Jung [16]8787602 × 860TextNoNo
RDSRD [24]540N/AN/ATextYesN/A
SD7K7620350+2462 × 3699Text and pictureYesDifferent lamps & sunlight
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.196, + 0.893, + 0.226 + ], + "angle": 0, + "content": "Table 1. Comparison of SD7K with other real-world datasets. The proposed dataset covers a large number of document shadow images and it is the only dataset that satisfies all important data features about document shadow currently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.279 + ], + "angle": 0, + "content": "based network to exploit the context correlation between shadow and non-shadow regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.471, + 0.478 + ], + "angle": 0, + "content": "Document Image Shadow Removal. Traditional document shadow removal methods usually demand prior knowledge regarding shadow and document context. This includes estimating the background color [30, 32] and modeling shading and lighting effects [16, 27]. Deep learning-based methods are also designed to remove the shadows of documents. Currently, BEDSR-Net [24] stands as the state-of-the-art model in document shadow removal. It predicts global background color and encodes shadow position as an attention map, allowing for the effective removal of shadows. However, as discussed in our introduction, it needs to predict the approximate attention maps from the feature map, which is not suitable for the high resolution inputs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.221, + 0.51 + ], + "angle": 0, + "content": "3. SD7K Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.643 + ], + "angle": 0, + "content": "As shown in Tab. 1 and Fig. 3 (a), there are only a limited number of datasets [16, 24, 1, 17] available for document shadow removal, which is typically small and comprises only hundreds of real-world samples. To train a network on such-scale datasets, synthesizing training samples via computer graphic engines is required [24]. However, the rendered samples still show different distributions from the real samples, making it difficult in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.718 + ], + "angle": 0, + "content": "To overcome this limitation, we introduce a massive real-world dataset named SD7K for document shadow removal in terms of the occluder entities, the numbers, the resolution, the light source, etc. Below, we give the details of how to collect and process the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "First, we use a Canon EOS M6 camera to capture printed document images from real-world scenarios. To capture paired shadow and shadow-free images, we make all the light close and leave one light source. All the settings of the camera are also fixed, including focal length, aperture, exposure, and ISO, to avoid the influence of the camera algorithm. Then, we use a camera stand to hold the camera and make it shoot on the paper vertically as shown in Fig. 2 (a). To capture the paired shadow and shadow-free image without involving the changes caused by the manual presses, as in Fig. 2 (b), we use a remote control to capture the photo between the shadowed and shadow-free paper by removing" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.247, + 0.629, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.397, + 0.576, + 0.41 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.247, + 0.761, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.397, + 0.708, + 0.41 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.247, + 0.892, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.397, + 0.838, + 0.41 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.416, + 0.893, + 0.471 + ], + "angle": 0, + "content": "Figure 2. (a) our data acquisition setup for constructing the dataset. (b) a remote phone control shutter where the user only needs to click a button to capture a pair of shadow-free/shadow images. (c) some showcases of our occluders." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.478, + 0.697, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.585, + 0.692, + 0.599 + ], + "angle": 0, + "content": "(a) Quantitative comparison" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.477, + 0.889, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.584, + 0.865, + 0.599 + ], + "angle": 0, + "content": "(b) Light distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.606, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Figure 3. Data distribution of SD7K and quantitative comparison across all document shadow datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.644, + 0.892, + 0.674 + ], + "angle": 0, + "content": "the occluder. We also take the pairs as short as possible to avoid the light difference between captures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.825 + ], + "angle": 0, + "content": "To further enhance diversity, we alternated the positions of the document and occluder after capturing a few batches of documents. Overall, we use over 30 types of occluders along with more than 350 documents to contribute to the dataset, some of the occluders are shown in Fig. 2 (c). These occluders have the shape of both regular and irregular forms, which provides adequate coverage for various situations, we calculate the shadow area percentage of the whole dataset, which is \\(41.38\\% \\pm 13.58\\%\\), indicating that it is sufficiently divergent for our shadow removal task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We also consider the fluctuations in sunlight intensity and random placement of shadows, which will be significantly influenced by the light variance. To simulate the results in different light sources, we introduce three types of light sources in our dataset for collection, including reg" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12451" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.207, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.089, + 0.338, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.089, + 0.47, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.236, + 0.207, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.385, + 0.177, + 0.4 + ], + "angle": 0, + "content": "Cool light" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.237, + 0.338, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.385, + 0.312, + 0.4 + ], + "angle": 0, + "content": "Warm light" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.236, + 0.47, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.385, + 0.434, + 0.4 + ], + "angle": 0, + "content": "Sunlight" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.406, + 0.461, + 0.421 + ], + "angle": 0, + "content": "Figure 4. Example shadow and shadow-free images from SD7K." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.469, + 0.495 + ], + "angle": 0, + "content": "cular sunlight, a stationary 80W LED yellow light having a color temperature of \\(3000\\mathrm{K}\\), and a stationary 80W LED white light having a color temperature of \\(6000\\mathrm{K}\\). The light distribution of the whole dataset is shown in Fig. 3 (b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.469, + 0.631 + ], + "angle": 0, + "content": "For data cleaning purposes, we first remove images with noticeable motion distortion or lack of focus. Since some photo still contains outliers and the context is duplicated, we then review the photographs in groups and remove them. Finally, we crop the areas outside of the margins of the document, resulting in a total of 7620 samples, where some of the samples are shown in Fig. 4. We also use a threshold-based method to extract the mask of the training pair since the shadow mask will also benefit related tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.427, + 0.666 + ], + "angle": 0, + "content": "4. Frequency-Aware Shadow Erasing Net" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Since real-world document shadow removal needs huge GPU memory requirements, our network initially divides the input image into different frequency components using the Laplacian Pyramid (LP) similar to [23] to reduce the computational budgets. As shown in Fig. 5, we propose a powerful frequency-aware shadow removal network named FSENet by designing different modules to handle the high-frequency and low-frequency computation individually. After training, the multi-frequency components are mixed to produce the final results. In the rest of this section, we first give the details of the Laplacian pyramid in Sec. 4.1. Then, in Sec. 4.2 and Sec. 4.3, we give the details of the low-frequency deshading module and high-frequency restoration module to precisely modify the color details and the high-frequency details individually. Finally, we give the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.673, + 0.106 + ], + "angle": 0, + "content": "loss functions in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.116, + 0.86, + 0.133 + ], + "angle": 0, + "content": "4.1. Laplacian Pyramid and Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Laplacian Pyramid (LP) [2] plays an important role in image processing. By decomposing and restoring images in a lossless manner, LP provides a rapid processing speed and is well-suited for dealing with high-resolution images. Inspired by [23] for image translation, we implement LP for document shadow removal and train the network on multifrequency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.893, + 0.49 + ], + "angle": 0, + "content": "In detail, given an image \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\), the \\(D\\) layers LP contains several high-frequency components \\([L_0, \\dots, L_{D-1}]\\) and a low-frequency component \\(L_D \\in \\mathbb{R}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times 3}\\), where \\(L_D\\) is \\(D\\) times downsample of \\(I\\). Each high-frequency component can be described as \\(L_i = I_i - \\mathrm{PyrUp}(\\mathrm{PyrDown}(I_i))\\), where \\(I_i\\) is the \\(i\\)-th down-sampled image and \\(\\mathrm{PyrUp}\\) and \\(\\mathrm{PyrDown}\\) stand for upsampling and downsampling, respectively. As illustrated in Fig. 5, the \\(L_1\\) and \\(L_2\\) include the high-frequency component edges and textures, while the low-frequency component \\(L_3 = I_3\\) majorly contains the color information. For reversible reconstruction, the final \\(\\hat{I}\\) can be obtained by reversed update the equation of \\(\\hat{I}_{D-1} = \\mathrm{PyrUp}(\\hat{I}_D) + \\hat{L}_{D-1}\\) from \\(D\\) to \\(0\\), where \\(\\hat{I}_D\\) is the learned low-frequency features and \\(\\hat{L}_{D-1}\\) is the learned high-frequency features from the proposed network. We give detailed instructions for each network below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.499, + 0.807, + 0.515 + ], + "angle": 0, + "content": "4.2. Low-Frequency Deshading Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.657 + ], + "angle": 0, + "content": "After getting the low-frequency components, we use several transformer-based blocks to reduce the color and illumination distortions of the shadow images. Compared to CNNs, transformers are superior at modeling non-local self-similarity and long-range dependencies. As shown in Fig. 5, we implement a transformer-based module called Dimension-Aware Transformer (DAT) inspired by [33] with a lightweight UNet-like sub-network [3] for low-frequency feature modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Formally, given a low-frequency component \\(L_{3} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 3}\\) of input image \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\), we first feed \\(L_{3}\\) into three sequential DAT blocks to extract rich features. Each DAT block is built via a sequential height-attention and width-attention to learn the feature in different aspects individually and a convolution-based layer for local feature merging, where we denote the intermediate features of each DAT block as \\(F1, F2, F3 \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}\\). Please refer to the original [33] for further details of the height-attention and width-attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.809, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We then feed these features \\( F1, F2, F3 \\) into a Tri-layer Attention Alignment (TAA) block for feature remixing. Given concatenated features \\( F_{\\mathrm{in}} \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4} \\times C} = \\text{Concat}([F1, F2, F3]) \\), where \\( N = 3 \\) in this case, \\( F_{\\mathrm{in}} \\) is first reshaped to \\( \\hat{F}_{\\mathrm{in}} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 3C} \\). Then, we send \\( \\hat{F}_{\\mathrm{in}} \\) to a \\( 1 \\times 1 \\) convolution layer to aggregate pixel-level cross" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12452" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.094, + 0.88, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.896, + 0.433 + ], + "angle": 0, + "content": "Figure 5. The network structure of our proposed FSENet. Following [23], given a high-resolution image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), we first use Laplacian Pyramid (e.g. Depth \\( D = 2 \\) in this case) to decompose the images to multiple frequency components. The Black arrows: The low-frequency part \\( L_{3} \\in \\mathbb{R}_{\\frac{H}{2D}}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times C} \\) are refined to \\( L_{3}' \\in \\mathbb{R}_{\\frac{H}{2D}}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times C} \\) utilizing DAT, TAA and DFE block. Cyan arrows: For the high-frequency part \\( L_{2} \\in \\mathbb{R}_{\\frac{H}{2D-1}}^{\\frac{H}{2D-1} \\times \\frac{W}{2D-1} \\times C} \\), a contour \\( C_{L_{2}} \\in \\mathbb{R}_{\\frac{H}{2D-1}}^{\\frac{H}{2D-1} \\times \\frac{W}{2D-1} \\times 1} \\) is learned to bridge the low-frequency and high-frequency features. Red arrows: For the remaining components with higher frequency, the learned contour is successively upsampled and refined using the proposed SPP and TRM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.473, + 0.576 + ], + "angle": 0, + "content": "channel context, followed by three \\(3 \\times 3\\) depth-wise convolution layers to generate query \\(\\mathbf{Q}\\), key \\(\\mathbf{K}\\), and value \\(\\mathbf{V}\\), individually. The \\(\\mathbf{Q}, \\mathbf{K}\\) are then reshaped to \\(\\hat{\\mathbf{Q}} \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\frac{W}{4} C}\\) and \\(\\hat{\\mathbf{K}} \\in \\mathbb{R}^{\\frac{H}{4} \\frac{W}{4} C \\times N}\\) respectively to calculate an attention matrix \\(\\mathbf{A} \\in \\mathbb{R}^{\\mathbf{N} \\times \\mathbf{N}}\\). Next, we reshape the \\(\\mathbf{V}\\) to \\(\\hat{\\mathbf{V}} \\in \\mathbb{R}^{\\frac{H}{4} \\frac{W}{4} C \\times N}\\) to multiply \\(\\hat{\\mathbf{V}}\\) and \\(\\mathbf{A}\\), and the input feature \\(\\hat{F}_{\\mathrm{in}}\\) is added to generate the output feature \\(F4\\), which can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.583, + 0.47, + 0.601 + ], + "angle": 0, + "content": "\\[\nF 4 = \\operatorname {C o n v} _ {1 \\times 1} L A (\\hat {\\mathbf {Q}}, \\hat {\\mathbf {K}}, \\hat {\\mathbf {V}}) + \\hat {F} _ {\\mathrm {i n}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.652 + ], + "angle": 0, + "content": "where \\( Conv_{1\\times 1} \\) represents a convolution layer with the kernel of \\( 1\\times 1 \\). LA indicates Layer Attention [33], which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.657, + 0.47, + 0.677 + ], + "angle": 0, + "content": "\\[\nL A (\\hat {\\mathbf {Q}}, \\hat {\\mathbf {K}}, \\hat {\\mathbf {V}}) = \\hat {\\mathbf {V}} \\operatorname {s o f t m a x} (\\hat {\\mathbf {Q}} \\hat {\\mathbf {K}} / \\alpha), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.712 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the temperature of the attention. Please refer to the supplementary material for more details of the TAA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.471, + 0.879 + ], + "angle": 0, + "content": "After getting \\( F4 \\), as shown in Fig. 5, We also use a UNet-like [3, 35] sub-network built by Deeper Feature extraction (DFE) blocks to learn multi-scale features. Each DFE block begins with a LayerNorm to stabilize the training procedure, followed by two convolutions and a deformable convolution. Then, it utilizes SimpleGate [3] to increase the diversity of the features by dividing the features along the channel dimension into two parts and multiplying them together. We also utilize Simplified Channel Attention (SCA) similar to [3] to capture global information in a computationally efficient manner, which can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.885, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nS C A (\\mathbf {X}) = \\mathbf {X} * \\mathbf {W} \\operatorname {p o o l} (\\mathbf {X}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.561 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}\\) represents a fully-connected layer, pool represents the global average pooling procedure that combines spatial data into channels, and \\(*\\) represents a channel-wise multiplication. The skip-connections [13] are connected before the LayerNorm to the output similar to [3]. We also give a detailed figure of the DFE block in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.893, + 0.67 + ], + "angle": 0, + "content": "For the refined feature \\(\\hat{F}^4\\), it then passes through three transformer blocks and a TAA Block to produce the enhanced features for image reconstruction using DAT and TAA blocks similar to the feature extractions of \\(F1, F2, F3\\). Finally, a convolution layer with the kernel size \\(3 \\times 3\\) is performed on the enhanced features to produce the low-frequency component \\(L_3' \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.681, + 0.822, + 0.698 + ], + "angle": 0, + "content": "4.3. High-Frequency Restoration Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "After getting the low-frequency features, we also design novel components to recover the high-frequency details of the learned features. In detail, we first upsample \\( L_{3} \\) and its corresponding result \\( L_{3}^{\\prime} \\) to match the size of \\( L_{2} \\), then we learn a contour on the concatenation of \\( [up(L_3), up(L_3'), L_2] \\) via several residual blocks [13], where \\( up(\\cdot) \\) means bilinear upsample operation, as illustrated in Fig. 5. This is a straightforward method for optimizing global correction compared to normal RGB images because the bands of the high-frequency feature map are very sparse, allowing us to reduce the volume of calculations as discussed in [23]. This contour will be expanded progressively into a collection of contours to match" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12453" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.237, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.089, + 0.401, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.089, + 0.565, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.089, + 0.729, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.089, + 0.893, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.289, + 0.237, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.291, + 0.4, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.291, + 0.565, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.291, + 0.729, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.291, + 0.893, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.534, + 0.893, + 0.562 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of the methods comparison in high resolution, the first row represents the results on the Kligler dataset and the second row depicts the results on the SD7K dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.649 + ], + "angle": 0, + "content": "the remaining high-frequency components. As illustrated in Fig. 5, given the high-frequency component \\( L_{2} \\) and the contour \\( C_{L_2} \\), the output \\( L_{2}^{\\prime} \\) can be written as \\( L_2^\\prime = L_2\\otimes C_{L_2} \\) where \\( \\otimes \\) represents the pixel-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.73 + ], + "angle": 0, + "content": "After multiplication, the Texture Recovery Module (TRM) is then designed to improve elaboration, inspired by the successful natural shadow removal methods [6, 37], which contain a series of the dilated convolutions and attentive aggregation nodes to recover the detailed features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In detail, to increase the reception fields on the high-resolution map, we incorporate dilated convolution [38]. Dilated convolution inserts spaces between the kernel values, effectively increasing the receptive field of the filter without losing resolution or coverage over the input. The amount of dilation is controlled by a dilation rate parameter. For a dilation rate of 1, the layer behaves like a regular convolution. For rates larger than 1, the filter is effectively widened to have a larger field of view. This allows the network to incorporate a wider context and capture features at different scales in a computationally effi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.895, + 0.753 + ], + "angle": 0, + "content": "cient manner compared to using very large filters. As for Attentive Aggregation Node, to re-weight the significance of feature maps, each aggregation node employs a squeeze-and-excitation block [14] to reweight the feature. A \\(3 \\times 3\\) convolution layer is then employed to compress the features and match the original channels. Finally, a spatial pooling pyramid (SPP) [12] is introduced at the conclusion of the TRM to facilitate the remixing of multi-context features. After getting the learned high-frequency of each layer, we use the method introduced in Sec. 4.1 to recover the image and calculate the supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.649, + 0.785 + ], + "angle": 0, + "content": "4.4. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For training, we utilize the loss functions of SmoothL [9] \\(\\mathcal{L}_{l1}^{smooth}\\) and SSIM [34] \\(\\mathcal{L}_{SSIM}\\) between the target image and the reconstructed image from LP, both two losses are helpful to preserve the details of the image, which can be represented as: \\(\\mathcal{L}_{total} = \\mathcal{L}_{l1}^{smooth} + \\lambda * \\mathcal{L}_{SSIM}\\), where we set the weight of \\(\\mathcal{L}_{SSIM} \\lambda\\) to 0.4 empirically." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12454" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.267 + ], + "angle": 0, + "content": "
MethodSD7KParam (M)
512 × 5121024 × 10242462 × 3699 (Full Size)
PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓
Input15.950.8944.09N/A15.950.9044.09N/A15.940.9144.14N/AN/A
Wang et al. [32]15.310.8247.88>1015.310.8547.88>3015.290.8647.95>60N/A
Wang et al. [30]13.320.6867.48>1013.320.7167.48>3013.260.7468.07>60N/A
Shah et al. [27]9.890.7186.35>109.900.7686.35>309.880.7986.46>60N/A
Jung et al. [16]19.860.9226.76>519.860.9326.76>1019.820.9226.86>30N/A
MaskGAN [15]24.820.8715.430.9624.820.8515.432.0724.670.8615.727.5928.29
DHAN [6]25.610.8514.270.2125.60*0.83*14.27*N/A25.42*0.85*14.60*N/A27.82
AEFNet [8]24.180.9516.830.1123.00*0.90*19.14*N/A22.94*0.90*19.28*N/A391.10
BEDSR-Net [24]21.500.9030.520.8519.81*0.85*33.03*N/A19.74*0.86*33.18*N/A29.44
ShadowFormer [10]23.710.9017.540.1522.69*0.84*19.65*N/A22.64*0.87*19.79*N/A11.35
BMNet [39]24.860.8015.590.1824.84*0.79*15.67*N/A24.70*0.83*15.96*N/A2.11
Ours28.690.979.980.2428.680.979.980.7528.670.9610.007.9329.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.277, + 0.893, + 0.32 + ], + "angle": 0, + "content": "Table 2. Quantitative results of the ablation study on three datasets in different resolutions. The best result is highlighted in red and bold. The asterisk (*) indicates that the result is obtained from a low-resolution result \\((512\\times 512)\\) and then improved using a super-resolution technique [21] to address an out-of-memory error." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.343, + 0.21, + 0.359 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.295, + 0.384 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.603 + ], + "angle": 0, + "content": "We evaluate the performance of different network structures with different resolutions on the proposed SD7K dataset, where the batch size is set to 4. We also give more experiments on the low-resolution of \\(512 \\times 512\\) of Jung [16], Kligler [17], and SD7K for fair comparison following previous methods, we give more details in the supplementary materials. To further utilize the dataset, we also incorporate a simple shadow synthesis pipeline to further increase the diversity of the training process. We investigate several potential options, including alpha composition [26] and Perlin noise generation [11] for comparison, and we empirically choose alpha composition to synthesize shadows through the shadow-free image and the mask on the fly via random alpha values in the range between [0.2, 0.7]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.711 + ], + "angle": 0, + "content": "As for the evaluation metrics, we employ three widely-used metrics: Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity (SSIM). While PSNR and SSIM are commonly used in image restoration evaluation [3, 5, 22], RMSE is the most important and frequently used metric specifically for shadow removal [6, 10]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.453, + 0.74 + ], + "angle": 0, + "content": "5.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.824 + ], + "angle": 0, + "content": "As presented in Tab. 2, we compared FSENet with both traditional and learning-based models of various resolutions using publicly available implementations of these methods, where the proposed method outperforms all other methodologies in most metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Furthermore, our analysis reveals that traditional methods, such as [16, 27], rely on physics-based assumptions that are not always correct, resulting in unstable performance across different datasets as shown in Fig. 6. Deep learning models, on the other hand, are highly dependent on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.373 + ], + "angle": 0, + "content": "the dataset magnitude, and their performance is relatively consistent across different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.893, + 0.557 + ], + "angle": 0, + "content": "We also find that some deep learning models that are effective in low-resolution shadow removal do not scale well to high-resolution documents. They either compromise the aspect ratio of the document or lead to out-of-memory errors. While traditional methods can still remove shadows, they are time-consuming, requiring pixel-by-pixel calculations of illumination difference [27] and background estimation [30]. At present, only Mask-ShadowGAN [15] and the proposed FSENet can effectively remove document shadows in high-resolution scenarios. Despite having more parameters and requiring more time to produce results, FSENet outperforms other methods quantitatively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.665, + 0.587 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.688 + ], + "angle": 0, + "content": "As presented in Tab. 3 and Fig. 7, we conduct our ablation studies at the original high-resolution documents, examining the effectiveness of several components in our algorithm. These components include the depth of the Laplacian Pyramid, the efficiency of TRM, DFE, and DAT modules, and the effectiveness of our shadow synthesis pipeline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We first evaluate the depth of the proposed Laplacian Pyramid. From Tab. 3, with the depth of the Laplacian Pyramid increases (Our w/ Depth=3), the magnitude of the low-frequency components decreases, affecting the low-frequency deshadowing and the reconstructed outcome. Consequently, the model's performance progressively declines as the depth increases. We then evaluate the performance of the proposed DFE and DAT modules, which are also responsible for low-frequency deshading, primarily influencing the pixel-wise variation of the final result. Although each module individually contributes to enhancing shadow elimination performance, their combined network produces the best results. As for the high-frequency modules, the TRM module handles high-frequency components" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.088, + 0.278, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.201, + 0.197, + 0.215 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.483, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.2, + 0.441, + 0.215 + ], + "angle": 0, + "content": "Ours w/ Depth=3" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.688, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.203, + 0.64, + 0.216 + ], + "angle": 0, + "content": "Ours w/o TRM" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.091, + 0.893, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.203, + 0.842, + 0.215 + ], + "angle": 0, + "content": "Ours w/o DFE" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.22, + 0.278, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.332, + 0.243, + 0.347 + ], + "angle": 0, + "content": "Ours w/o Synthesis" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.222, + 0.483, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.335, + 0.433, + 0.347 + ], + "angle": 0, + "content": "Ours w/o DAT" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.222, + 0.688, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.335, + 0.606, + 0.347 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.22, + 0.893, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.333, + 0.815, + 0.347 + ], + "angle": 0, + "content": "Target" + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.357, + 0.772, + 0.372 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of the ablation studies on different modules of the proposed FSENet." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.384, + 0.871, + 0.53 + ], + "angle": 0, + "content": "
Ablation StudyJung (Original)Kliger (Original)SD7K (Original)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
Input12.930.7961.3413.250.8156.7515.940.9144.14
Ours w/ Depth=320.970.8323.1926.330.9113.6620.830.9324.02
Ours w/o TRM19.500.8127.6027.100.9212.7818.440.8635.92
Ours w/o DFE21.900.8421.0225.160.9214.8919.700.8829.07
Ours w/o DAT22.680.8519.2822.700.8820.1919.030.8732.94
Ours w/o Synthesis22.210.8521.3324.790.9215.5422.940.9321.93
Ours Full23.600.8517.5628.980.9310.6028.670.9610.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.535, + 0.855, + 0.549 + ], + "angle": 0, + "content": "Table 3. Quantitative results of the ablation study on three datasets in high-resolution. The best result is highlighted in red and bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.65 + ], + "angle": 0, + "content": "such as contours and edges, thus contributing significantly to the overall efficacy of our algorithm. Finally, we also evaluate the importance of the proposed shadow synthesis pipeline, where w/o shadow synthesis reduces the performance a lot, especially on the two small datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.658, + 0.204, + 0.673 + ], + "angle": 0, + "content": "5.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Although our model is capable of removing shadows in high-resolution images, the current network is not able to run on real-time or on edge devices, such as mobile phones, due to the larger parameters and heavy GPU computations. In the future, we will explore different approaches to further reduce the required parameters, making it usable on smart devices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.196, + 0.816 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this paper, we tackle the problem of removing shadows from high-resolution documents. To achieve this goal, we consider the contribution of the dataset and the network. As for the dataset, we gather a high-resolution dataset SD7K, which is comprised of over 7000 triplets of real-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.893, + 0.801 + ], + "angle": 0, + "content": "world document images with distinct characteristics under various illumination conditions. This dataset is an order of magnitude greater than previous datasets and allows us to validate our approach in a more diverse and comprehensive manner. As for the network structure, we propose FSENet, a frequency-aware method to remove shadows in higher resolution via frequency decomposition. To separate the high-frequency contours from the low-resolution content, we employ a Laplacian Pyramid and a series of meticulously designed attention blocks, enabling each component to be learned separately. Powered by our dataset and network, our method achieves a new state-of-the-art for document shadow removal. We believe both the network structure and the dataset will benefit the community for research on this topic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by the University of Macau under Grant MYRG2022-00190-FST, in part by the Science and Technology Development Fund, Macau SAR, under Grant 0034/2019/AMJ, Grant 0087/2020/A2 and Grant 0049/2021/A." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12456" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Steve Bako, Soheil Darabi, Eli Shechtman, Jue Wang, Kalyan Sunkavalli, and Pradeep Sen. Removing shadows from images of documents. In ACCV, pages 173-183. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33. Springer, 2022. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[4] Xuhang Chen, Xiaodong Cun, Chi-Man Pun, and Shuqiang Wang. Shadocnet: Learning spatial-aware tokens in transformer for document shadow removal. In IEEE ICASSP, pages 1-5. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[5] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[6] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In CVPR, pages 10571-10580, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[9] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.592, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[11] John C Hart. Perlin noise pixel shaders. In Proceedings of the ACM SIGGRAPH/EUROGRAPHICS workshop on Graphics hardware, pages 87-94, 2001. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[15] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] Seungjun Jung, Muhammad Abul Hasan, and Changick Kim. Water-filling: An efficient algorithm for digitized document shadow removal. In Asian Conference on Computer Vision, pages 398-414. Springer, 2018. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Netanel Kligler, Sagi Katz, and Ayellet Tal. Document enhancement using visibility detection. In CVPR, pages 2374-2382, 2018. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[18] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, August 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[19] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, pages 264-281. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.21, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[20] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[22] Jingtang Liang, Xiaodong Cun, Chi-Man Pun, and Jue Wang. Spatial-separated curve rendering network for efficient and high-resolution image harmonization. In European Conference on Computer Vision, pages 334-349. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In CVPR, pages 9392-9400, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[24] Yun-Hsuan Lin, Wen-Chin Chen, and Yung-Yu Chuang. Bedsr-net: A deep shadow removal network from a single document image. In CVPR, pages 12905-12914, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[25] Zhihao Liu, Hui Yin, Yang Mi, Mengyang Pu, and Song Wang. Shadow removal by a lightness-guided network with training on unpaired data. IEEE Transactions on Image Processing, 30:1853-1865, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[26] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In SIGGRAPH, pages 231–242, 1998. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[27] Vatsal Shah and Vineet Gandhi. An iterative approach for shadow removal in document images. IEEE ICASSP, pages 1892-1896, 2018. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[28] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[29] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV, pages 361-378. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[30] Bingshu Wang and CL Philip Chen. An effective background estimation method for shadows removal of document images. In IEEE International Conference on Image Processing (ICIP), pages 3611-3615. IEEE, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12457" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[32] Jian-Ren Wang and Yung-Yu Chuang. Shadow removal of text document images by estimating local and global background colors. IEEE ICASSP, pages 1534-1538, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. AAAI, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17683-17693, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In SIGKDD, pages 1192-1200, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[37] Yimin Xu, Mingbao Lin, Hong Yang, Ke Li, Yunhang Shen, Fei Chao, and Rongrong Ji. Shadow-aware dynamic convolution for shadow removal. arXiv preprint arXiv:2205.04908, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[38] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. *ICLR*, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, June 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12458" + } + ] +] \ No newline at end of file diff --git a/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_origin.pdf b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c136e8b37ff5b2643e85f54dcb94cf3ec9bb211e --- /dev/null +++ b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/f8df0628-9a27-4d45-a2b6-2c8c71bf57b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdc8d5233055d88a0667cb9fdaabf3d31e9bf77ef01a230cf737e6f2557420cd +size 1467774 diff --git a/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/full.md b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/full.md new file mode 100644 index 0000000000000000000000000000000000000000..96be35577edeb6f5b95de5c419803d83ac5c04d9 --- /dev/null +++ b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/full.md @@ -0,0 +1,324 @@ +# High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net + +Zinuo Li* Xuhang Chen* Chi-Man Pun† Xiaodong Cun† University of Macau + +# Abstract + +Shadows often occur when we capture the document with casual equipment, which influences the visual quality and readability of the digital copies. Different from the algorithms for natural shadow removal, the algorithms in document shadow removal need to preserve the details of fonts and figures in high-resolution input. Previous works ignore this problem and remove the shadows via approximate attention and small datasets, which might not work in real-world situations. We handle high-resolution document shadow removal directly via a larger-scale real-world dataset and a carefully-designed frequency-aware network. As for the dataset, we acquire over 7k couples of high-resolution $(2462 \times 3699)$ images of real-world documents pairs with various samples under different lighting circumstances, which is 10 times larger than existing datasets. As for the design of the network, we decouple the high-resolution images in the frequency domain, where the low-frequency details and high-frequency boundaries can be effectively learned via the carefully designed network structure. Powered by our network and dataset, the proposed method shows a clearly better performance than previous methods in terms of visual quality and numerical results. The code, models, and dataset are available at https://github.com/CXH-Research/DocShadow-SD7K. + +# 1. Introduction + +When occluders obstruct the light source, shadows are cast on the paper. These shadows reduce the readability and hinder subsequent text-related intelligent tasks such as document enhancement, optical character recognition (OCR), key information extraction, and semantic entity labeling [17, 36, 28], etc. + +Removing the cast shadow of the digital copy has a long need in the real world and also there are a lot of meth- + +![](images/28248561b5a96f35328a5d259765c94aa8659814812532f61cc9123f682b37dc.jpg) + +![](images/a58314bacef103a319dbf819aa6b55f9edbe16120dffaedc308bf8265f676658.jpg) +(a) Input + +![](images/ada9ef044eab22ffb70a56b5816b42e5e29c695afe9c69b1b4a4b124a433333e.jpg) + +![](images/3313398616b56ebe2fc1975c8b7236339da7c8d8d03cc98d76d941f1b50fabae.jpg) +(b) Shah et al. [27] + +![](images/f2cae5e479813c2403bb3f00ba8334ec4981e4ad64ef4cffc5a54a239dea3aea.jpg) + +![](images/01b4daa90f855adf0da9244c45fd9be6d7dc0e6165bd89e5376bc12dc662201c.jpg) +(c) Fu et al. [8] + +![](images/d8bdf7babb2ed771fb7b544ddca8ce680fbe6abd578c9a64fcc490e23d07d089.jpg) + +![](images/8da33ff966afc1227b2d19f6063a88ec014c8a50b65b9d4b317e44359efbf6d0.jpg) +(d) Liu et al. [25] + +![](images/87f49fc9a1a0ca8a5a4b9e49553b42eeb5972a8a95d8703585afadb91836e09d.jpg) + +![](images/c5df81b1c025cbda53448a1652da314adbe0511c77a6726d39d6f011e0019c28.jpg) +(e) Hu et al. [15] + +![](images/5ad08578990237c1cc1107e91198ec1632793025d7a85ce750802b97dbf9ae98.jpg) + +![](images/3ab42bdea443cf76f2a254b30a1996da4c1767e6c950f569492ac1401c862200.jpg) +(f) Ours +Figure 1. Visual results of input document shadow image (a), classic method (b), supervised method (c), weakly-supervised method (d), unsupervised method (e) and ours (f). Our model removes the shadow while preserving the original document's content and aspect ratio. + +ods, including the traditional methods that applied physics-based illumination models [16, 30, 32]. In reality, these approaches are significantly constrained due to the fact that the assumptions utilized in illumination models [27] are often too rigid to account for real-world shadow. Thus, they may + +incorrectly compute the exposure and reflectance as shown in Fig. 1. + +Recent methods apply deep learning for this task since it has achieved remarkable progress in natural shadow removal [6] and other computer vision tasks [12, 13, 35]. However, several obstacles make it challenging to transfer previous natural shadow removal methods to the document domain. On the one hand, in contrast to natural shadow removal, documents are presented by high-resolution images in the real world, which needs to preserve fine-grained information such as fonts and figures. However, the majority of document shadow removal methods [24, 4] are only designed for relatively low-resolution images and is hard to handle high-resolution situations directly since they are relying on the low-resolution approximation of attention for guidance. On the other hand, there are no large-scale and high-resolution datasets designed specifically for document shadow removal at this time. Consequently, given limited image pairs for training, existing methods usually result in different artifacts (as shown in Fig. 1 (c)). Although weakly-supervised [18] and unsupervised [15, 25] shadow removal methods alleviate the requirements for data, they hold a strong assumption of the statistical similarity between shadowed and unshadowed domains. As illustrated in Fig. 1 (d) (e), once the training domain is very different from the test, these methods create hallucination contents and suffer from unstable results. + +To solve the above problems, we first offer a large-scale real-world shadow dataset, Shadowed Document 7K (SD7K), which contains more than 7k real-world and high-resolution shadow and shadow-free document image pairs with also the manually annotated shadow masks for related research, e.g., shadow detection [6]. Compared to existing document shadow removal datasets, SD7K includes various document types, e.g., manga, papers, and figures. Also, different from previous work which only considers one specific light environment, we use three distinct light sources to increase the diversity of the training samples. Besides the dataset, we also propose a robust network for high-resolution image processing specifically, named Frequency-aware Shadow Erasing Net (FSENet). In detail, inspired by the laplacian pyramid [2, 23], we divide the document image into two frequency portions, each with its own domain-specific properties. The low-frequency component is primarily responsible for illuminations or colors while high-frequency components are more related to content details. As for the low-frequency deshading module, we involve the transformer [7] based network to model the global illuminations alternation and a high-frequency restoration module based on cascaded aggregations for adaptive pixel enhancement. For high-frequency modules, we involve several dilated convolution-based texture recovery modules to restore the refined details similar to [6]. With the aforementioned + +techniques, our model achieves the state-of-the-art performance of previous methods on two common small benchmarks and the proposed SD7K. + +We summarize the main contribution as follows: + +- We provide SD7K, a large-scale real-world dataset consisting of high-resolution shadow and the associated shadow-free images under various illumination conditions. +- We propose FSENet, a frequency-aware network with a carefully designed network structure to handle high-resolution document shadows. +- Both qualitative and quantitative results on all available public datasets indicate that the proposed FSENet performs favorably against state-of-the-art methods. + +# 2. Related Work + +Document Shadow Dataset. The collection of genuine shadow images and their corresponding shadow-free counterparts is a challenging task due to the varying lighting conditions. As a result, existing datasets are limited in size. The Bako dataset [1] comprises outdated, inaccessible images in tiny quantities. Similarly, Jung's dataset [16] is limited in size and favors light shadows. Kligler's dataset [17] emphasizes colorful symbols, which do not convey a significant amount of information. Among these datasets, the RDSRD dataset [24], while still modest in size, has a better quality in comparison. Different from the existing datasets, we propose a high-resolution and large-scale dataset SD7K that addresses the aforementioned problem, which can facilitate the research in document shadow removal. + +Natural Shadow Removal. Methods for natural shadow removal might be informative and applicable to document images. On the one hand, GAN-based methods such as STCGAN [31], use two stacked conditional GANs to jointly detect and eliminate shadows. Mask-ShadowGAN [15] extend the paired GAN to unpaired setting via re-formulated cycle-consistency constraints. To fully utilize the dataset, DHAN [6] synthesize pseudo shadow images and learned border artifacts-free images via a dual hierarchical aggregation network. SG-ShadowNet [29] studies the importance of normalization and learns the style representation of the non-shadow region to harmonize the shadow and shadow-free parts. On the other hand, CNN-based methods such as $\mathrm{SP + M}$ Net [19] and $\mathrm{SP + M + I}$ Net [20] utilize the physical linear transformation model to enhance the shadow region and image decomposition to reconstruct the shadow-free image. AEF-Net [8] predicts the exposure parameters of the shadowed regions and fused a series of overexposed shadow images with the original shadow image to restore the shadow region. ShadowFormer [10] is a transformer + +
DatasetTriplets #Entities #Mean ResolutionDocument TypeWith MaskLight Temperature Control
Bako [1]81N/AN/ATextNoNo
Kligler [17]300252424 × 2013TextNoNo
Jung [16]8787602 × 860TextNoNo
RDSRD [24]540N/AN/ATextYesN/A
SD7K7620350+2462 × 3699Text and pictureYesDifferent lamps & sunlight
+ +Table 1. Comparison of SD7K with other real-world datasets. The proposed dataset covers a large number of document shadow images and it is the only dataset that satisfies all important data features about document shadow currently. + +based network to exploit the context correlation between shadow and non-shadow regions. + +Document Image Shadow Removal. Traditional document shadow removal methods usually demand prior knowledge regarding shadow and document context. This includes estimating the background color [30, 32] and modeling shading and lighting effects [16, 27]. Deep learning-based methods are also designed to remove the shadows of documents. Currently, BEDSR-Net [24] stands as the state-of-the-art model in document shadow removal. It predicts global background color and encodes shadow position as an attention map, allowing for the effective removal of shadows. However, as discussed in our introduction, it needs to predict the approximate attention maps from the feature map, which is not suitable for the high resolution inputs. + +# 3. SD7K Dataset + +As shown in Tab. 1 and Fig. 3 (a), there are only a limited number of datasets [16, 24, 1, 17] available for document shadow removal, which is typically small and comprises only hundreds of real-world samples. To train a network on such-scale datasets, synthesizing training samples via computer graphic engines is required [24]. However, the rendered samples still show different distributions from the real samples, making it difficult in real-world applications. + +To overcome this limitation, we introduce a massive real-world dataset named SD7K for document shadow removal in terms of the occluder entities, the numbers, the resolution, the light source, etc. Below, we give the details of how to collect and process the dataset. + +First, we use a Canon EOS M6 camera to capture printed document images from real-world scenarios. To capture paired shadow and shadow-free images, we make all the light close and leave one light source. All the settings of the camera are also fixed, including focal length, aperture, exposure, and ISO, to avoid the influence of the camera algorithm. Then, we use a camera stand to hold the camera and make it shoot on the paper vertically as shown in Fig. 2 (a). To capture the paired shadow and shadow-free image without involving the changes caused by the manual presses, as in Fig. 2 (b), we use a remote control to capture the photo between the shadowed and shadow-free paper by removing + +![](images/92ade6c40849ca1920b29d8c785b2725bd4ca06b8850112fa8e87a8c1751b336.jpg) +(a) + +![](images/ba9ec9be2845cbbd327b3554db80b7b1bfcac9937ec554fe8614a442c33bb9c4.jpg) +(b) + +![](images/0780e631cfb7cd38c83c6109a0d5418dbf1aebf3261d2d13fe91a352a83757af.jpg) +(c) + +![](images/3971a0ab029c7bea5f66208b8a97e4f9b59c8aa6086467d9e60a5779c1df6d87.jpg) +(a) Quantitative comparison +Figure 3. Data distribution of SD7K and quantitative comparison across all document shadow datasets. + +![](images/3851803b58f5cbb5b8ff55f23ae8252460da19811130a35483792220ad985e44.jpg) +Figure 2. (a) our data acquisition setup for constructing the dataset. (b) a remote phone control shutter where the user only needs to click a button to capture a pair of shadow-free/shadow images. (c) some showcases of our occluders. +(b) Light distribution + +the occluder. We also take the pairs as short as possible to avoid the light difference between captures. + +To further enhance diversity, we alternated the positions of the document and occluder after capturing a few batches of documents. Overall, we use over 30 types of occluders along with more than 350 documents to contribute to the dataset, some of the occluders are shown in Fig. 2 (c). These occluders have the shape of both regular and irregular forms, which provides adequate coverage for various situations, we calculate the shadow area percentage of the whole dataset, which is $41.38\% \pm 13.58\%$ , indicating that it is sufficiently divergent for our shadow removal task. + +We also consider the fluctuations in sunlight intensity and random placement of shadows, which will be significantly influenced by the light variance. To simulate the results in different light sources, we introduce three types of light sources in our dataset for collection, including reg + +![](images/743ae3ba46e21e912060a27d8559cad02da70f55ec3ef8cf72cef89640311f1f.jpg) + +![](images/f7df2cb277b10a39bcbfae849c6eb66b057f287807b34ff6085eb1c29ab4ee82.jpg) + +![](images/5b8298c76ad7ea611cbc46abc3b2407a59330f2cbb769867eca96e29938d215e.jpg) + +![](images/a6902a37fa99561e7933b2991cabfcbd781acae515ef2540f19fa28de99871d3.jpg) +Cool light + +![](images/0d54e187e74aa06fe4b74897d7c48018bf5e9fe8a33e2545d8e8e97630f776ac.jpg) +Warm light + +![](images/334c8217b4404318abfb486c6f029e1369960eb2f68e018ebd3258fa6b717bdf.jpg) +Sunlight +Figure 4. Example shadow and shadow-free images from SD7K. + +cular sunlight, a stationary 80W LED yellow light having a color temperature of $3000\mathrm{K}$ , and a stationary 80W LED white light having a color temperature of $6000\mathrm{K}$ . The light distribution of the whole dataset is shown in Fig. 3 (b). + +For data cleaning purposes, we first remove images with noticeable motion distortion or lack of focus. Since some photo still contains outliers and the context is duplicated, we then review the photographs in groups and remove them. Finally, we crop the areas outside of the margins of the document, resulting in a total of 7620 samples, where some of the samples are shown in Fig. 4. We also use a threshold-based method to extract the mask of the training pair since the shadow mask will also benefit related tasks. + +# 4. Frequency-Aware Shadow Erasing Net + +Since real-world document shadow removal needs huge GPU memory requirements, our network initially divides the input image into different frequency components using the Laplacian Pyramid (LP) similar to [23] to reduce the computational budgets. As shown in Fig. 5, we propose a powerful frequency-aware shadow removal network named FSENet by designing different modules to handle the high-frequency and low-frequency computation individually. After training, the multi-frequency components are mixed to produce the final results. In the rest of this section, we first give the details of the Laplacian pyramid in Sec. 4.1. Then, in Sec. 4.2 and Sec. 4.3, we give the details of the low-frequency deshading module and high-frequency restoration module to precisely modify the color details and the high-frequency details individually. Finally, we give the + +loss functions in Sec. 4.4. + +# 4.1. Laplacian Pyramid and Method Overview + +Laplacian Pyramid (LP) [2] plays an important role in image processing. By decomposing and restoring images in a lossless manner, LP provides a rapid processing speed and is well-suited for dealing with high-resolution images. Inspired by [23] for image translation, we implement LP for document shadow removal and train the network on multifrequency. + +In detail, given an image $I \in \mathbb{R}^{H \times W \times 3}$ , the $D$ layers LP contains several high-frequency components $[L_0, \dots, L_{D-1}]$ and a low-frequency component $L_D \in \mathbb{R}^{\frac{H}{2D} \times \frac{W}{2D} \times 3}$ , where $L_D$ is $D$ times downsample of $I$ . Each high-frequency component can be described as $L_i = I_i - \mathrm{PyrUp}(\mathrm{PyrDown}(I_i))$ , where $I_i$ is the $i$ -th down-sampled image and $\mathrm{PyrUp}$ and $\mathrm{PyrDown}$ stand for upsampling and downsampling, respectively. As illustrated in Fig. 5, the $L_1$ and $L_2$ include the high-frequency component edges and textures, while the low-frequency component $L_3 = I_3$ majorly contains the color information. For reversible reconstruction, the final $\hat{I}$ can be obtained by reversed update the equation of $\hat{I}_{D-1} = \mathrm{PyrUp}(\hat{I}_D) + \hat{L}_{D-1}$ from $D$ to $0$ , where $\hat{I}_D$ is the learned low-frequency features and $\hat{L}_{D-1}$ is the learned high-frequency features from the proposed network. We give detailed instructions for each network below. + +# 4.2. Low-Frequency Deshading Module + +After getting the low-frequency components, we use several transformer-based blocks to reduce the color and illumination distortions of the shadow images. Compared to CNNs, transformers are superior at modeling non-local self-similarity and long-range dependencies. As shown in Fig. 5, we implement a transformer-based module called Dimension-Aware Transformer (DAT) inspired by [33] with a lightweight UNet-like sub-network [3] for low-frequency feature modeling. + +Formally, given a low-frequency component $L_{3} \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times 3}$ of input image $I \in \mathbb{R}^{H \times W \times 3}$ , we first feed $L_{3}$ into three sequential DAT blocks to extract rich features. Each DAT block is built via a sequential height-attention and width-attention to learn the feature in different aspects individually and a convolution-based layer for local feature merging, where we denote the intermediate features of each DAT block as $F1, F2, F3 \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times C}$ . Please refer to the original [33] for further details of the height-attention and width-attention. + +We then feed these features $F1, F2, F3$ into a Tri-layer Attention Alignment (TAA) block for feature remixing. Given concatenated features $F_{\mathrm{in}} \in \mathbb{R}^{N \times \frac{H}{4} \times \frac{W}{4} \times C} = \text{Concat}([F1, F2, F3])$ , where $N = 3$ in this case, $F_{\mathrm{in}}$ is first reshaped to $\hat{F}_{\mathrm{in}} \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times 3C}$ . Then, we send $\hat{F}_{\mathrm{in}}$ to a $1 \times 1$ convolution layer to aggregate pixel-level cross + +![](images/809596f7670c898143aa330172b16331ccdde3b95a768080cdcafdb18eb9ff71.jpg) +Figure 5. The network structure of our proposed FSENet. Following [23], given a high-resolution image $I \in \mathbb{R}^{H \times W \times 3}$ , we first use Laplacian Pyramid (e.g. Depth $D = 2$ in this case) to decompose the images to multiple frequency components. The Black arrows: The low-frequency part $L_{3} \in \mathbb{R}_{\frac{H}{2D}}^{\frac{H}{2D} \times \frac{W}{2D} \times C}$ are refined to $L_{3}' \in \mathbb{R}_{\frac{H}{2D}}^{\frac{H}{2D} \times \frac{W}{2D} \times C}$ utilizing DAT, TAA and DFE block. Cyan arrows: For the high-frequency part $L_{2} \in \mathbb{R}_{\frac{H}{2D-1}}^{\frac{H}{2D-1} \times \frac{W}{2D-1} \times C}$ , a contour $C_{L_{2}} \in \mathbb{R}_{\frac{H}{2D-1}}^{\frac{H}{2D-1} \times \frac{W}{2D-1} \times 1}$ is learned to bridge the low-frequency and high-frequency features. Red arrows: For the remaining components with higher frequency, the learned contour is successively upsampled and refined using the proposed SPP and TRM. + +channel context, followed by three $3 \times 3$ depth-wise convolution layers to generate query $\mathbf{Q}$ , key $\mathbf{K}$ , and value $\mathbf{V}$ , individually. The $\mathbf{Q}, \mathbf{K}$ are then reshaped to $\hat{\mathbf{Q}} \in \mathbb{R}^{N \times \frac{H}{4} \frac{W}{4} C}$ and $\hat{\mathbf{K}} \in \mathbb{R}^{\frac{H}{4} \frac{W}{4} C \times N}$ respectively to calculate an attention matrix $\mathbf{A} \in \mathbb{R}^{\mathbf{N} \times \mathbf{N}}$ . Next, we reshape the $\mathbf{V}$ to $\hat{\mathbf{V}} \in \mathbb{R}^{\frac{H}{4} \frac{W}{4} C \times N}$ to multiply $\hat{\mathbf{V}}$ and $\mathbf{A}$ , and the input feature $\hat{F}_{\mathrm{in}}$ is added to generate the output feature $F4$ , which can be described as: + +$$ +F 4 = \operatorname {C o n v} _ {1 \times 1} L A (\hat {\mathbf {Q}}, \hat {\mathbf {K}}, \hat {\mathbf {V}}) + \hat {F} _ {\mathrm {i n}}, \tag {1} +$$ + +where $Conv_{1\times 1}$ represents a convolution layer with the kernel of $1\times 1$ . LA indicates Layer Attention [33], which can be written as: + +$$ +L A (\hat {\mathbf {Q}}, \hat {\mathbf {K}}, \hat {\mathbf {V}}) = \hat {\mathbf {V}} \operatorname {s o f t m a x} (\hat {\mathbf {Q}} \hat {\mathbf {K}} / \alpha), \tag {2} +$$ + +where $\alpha$ is the temperature of the attention. Please refer to the supplementary material for more details of the TAA. + +After getting $F4$ , as shown in Fig. 5, We also use a UNet-like [3, 35] sub-network built by Deeper Feature extraction (DFE) blocks to learn multi-scale features. Each DFE block begins with a LayerNorm to stabilize the training procedure, followed by two convolutions and a deformable convolution. Then, it utilizes SimpleGate [3] to increase the diversity of the features by dividing the features along the channel dimension into two parts and multiplying them together. We also utilize Simplified Channel Attention (SCA) similar to [3] to capture global information in a computationally efficient manner, which can be denoted as: + +$$ +S C A (\mathbf {X}) = \mathbf {X} * \mathbf {W} \operatorname {p o o l} (\mathbf {X}), \tag {3} +$$ + +where $\mathbf{W}$ represents a fully-connected layer, pool represents the global average pooling procedure that combines spatial data into channels, and $*$ represents a channel-wise multiplication. The skip-connections [13] are connected before the LayerNorm to the output similar to [3]. We also give a detailed figure of the DFE block in the supplementary materials. + +For the refined feature $\hat{F}^4$ , it then passes through three transformer blocks and a TAA Block to produce the enhanced features for image reconstruction using DAT and TAA blocks similar to the feature extractions of $F1, F2, F3$ . Finally, a convolution layer with the kernel size $3 \times 3$ is performed on the enhanced features to produce the low-frequency component $L_3' \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times C}$ . + +# 4.3. High-Frequency Restoration Module + +After getting the low-frequency features, we also design novel components to recover the high-frequency details of the learned features. In detail, we first upsample $L_{3}$ and its corresponding result $L_{3}^{\prime}$ to match the size of $L_{2}$ , then we learn a contour on the concatenation of $[up(L_3), up(L_3'), L_2]$ via several residual blocks [13], where $up(\cdot)$ means bilinear upsample operation, as illustrated in Fig. 5. This is a straightforward method for optimizing global correction compared to normal RGB images because the bands of the high-frequency feature map are very sparse, allowing us to reduce the volume of calculations as discussed in [23]. This contour will be expanded progressively into a collection of contours to match + +![](images/811c2e3d6a36dc1b4e14644b5d54f44f92fdb06e0c42fe4e6acc49d9cfb9d430.jpg) + +![](images/74164336ac43927d45013c270e30b5f3f13d72b3823f2c0d4c9a5399a3b35a6b.jpg) + +![](images/fd8c0bafec9d9c4762f2c68ec2b00102f6c001c5ac12f67fc54020df493447e8.jpg) + +![](images/51233e8dd291e6af8272dfd95225800d5c7e629b65b83acaad949dbfaaada90b.jpg) + +![](images/052222441240f32fe624d20415976a38a0d77f4e07765e852893031c63a9864f.jpg) + +![](images/d95750c311867ea181e1d9c44d6a3814ec267b64402fc6674d657d74dc9ee0ff.jpg) +Figure 6. Qualitative results of the methods comparison in high resolution, the first row represents the results on the Kligler dataset and the second row depicts the results on the SD7K dataset. + +![](images/bba7869dc5a5072f88b969330c0b4fb6118831728d8c9dc73372cefb2261debf.jpg) + +![](images/958643130421a86ba976679e75915cfc439838dfd7af486159017b0bd5fc53de.jpg) + +![](images/1a4ceb403dec194759a11a0e52a790c814b46fae5fdd82fb1e12cf7487b705ff.jpg) + +![](images/f57eb4477adf4279f8d6fb43a845c9754ed2ec17d3f2a8eca2e7e9449c6faabe.jpg) + +the remaining high-frequency components. As illustrated in Fig. 5, given the high-frequency component $L_{2}$ and the contour $C_{L_2}$ , the output $L_{2}^{\prime}$ can be written as $L_2^\prime = L_2\otimes C_{L_2}$ where $\otimes$ represents the pixel-wise multiplication. + +After multiplication, the Texture Recovery Module (TRM) is then designed to improve elaboration, inspired by the successful natural shadow removal methods [6, 37], which contain a series of the dilated convolutions and attentive aggregation nodes to recover the detailed features. + +In detail, to increase the reception fields on the high-resolution map, we incorporate dilated convolution [38]. Dilated convolution inserts spaces between the kernel values, effectively increasing the receptive field of the filter without losing resolution or coverage over the input. The amount of dilation is controlled by a dilation rate parameter. For a dilation rate of 1, the layer behaves like a regular convolution. For rates larger than 1, the filter is effectively widened to have a larger field of view. This allows the network to incorporate a wider context and capture features at different scales in a computationally effi + +cient manner compared to using very large filters. As for Attentive Aggregation Node, to re-weight the significance of feature maps, each aggregation node employs a squeeze-and-excitation block [14] to reweight the feature. A $3 \times 3$ convolution layer is then employed to compress the features and match the original channels. Finally, a spatial pooling pyramid (SPP) [12] is introduced at the conclusion of the TRM to facilitate the remixing of multi-context features. After getting the learned high-frequency of each layer, we use the method introduced in Sec. 4.1 to recover the image and calculate the supervision. + +# 4.4. Loss functions + +For training, we utilize the loss functions of SmoothL [9] $\mathcal{L}_{l1}^{smooth}$ and SSIM [34] $\mathcal{L}_{SSIM}$ between the target image and the reconstructed image from LP, both two losses are helpful to preserve the details of the image, which can be represented as: $\mathcal{L}_{total} = \mathcal{L}_{l1}^{smooth} + \lambda * \mathcal{L}_{SSIM}$ , where we set the weight of $\mathcal{L}_{SSIM} \lambda$ to 0.4 empirically. + +
MethodSD7KParam (M)
512 × 5121024 × 10242462 × 3699 (Full Size)
PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓
Input15.950.8944.09N/A15.950.9044.09N/A15.940.9144.14N/AN/A
Wang et al. [32]15.310.8247.88>1015.310.8547.88>3015.290.8647.95>60N/A
Wang et al. [30]13.320.6867.48>1013.320.7167.48>3013.260.7468.07>60N/A
Shah et al. [27]9.890.7186.35>109.900.7686.35>309.880.7986.46>60N/A
Jung et al. [16]19.860.9226.76>519.860.9326.76>1019.820.9226.86>30N/A
MaskGAN [15]24.820.8715.430.9624.820.8515.432.0724.670.8615.727.5928.29
DHAN [6]25.610.8514.270.2125.60*0.83*14.27*N/A25.42*0.85*14.60*N/A27.82
AEFNet [8]24.180.9516.830.1123.00*0.90*19.14*N/A22.94*0.90*19.28*N/A391.10
BEDSR-Net [24]21.500.9030.520.8519.81*0.85*33.03*N/A19.74*0.86*33.18*N/A29.44
ShadowFormer [10]23.710.9017.540.1522.69*0.84*19.65*N/A22.64*0.87*19.79*N/A11.35
BMNet [39]24.860.8015.590.1824.84*0.79*15.67*N/A24.70*0.83*15.96*N/A2.11
Ours28.690.979.980.2428.680.979.980.7528.670.9610.007.9329.34
+ +Table 2. Quantitative results of the ablation study on three datasets in different resolutions. The best result is highlighted in red and bold. The asterisk (*) indicates that the result is obtained from a low-resolution result $(512\times 512)$ and then improved using a super-resolution technique [21] to address an out-of-memory error. + +# 5. Experiments + +# 5.1. Implementation Details + +We evaluate the performance of different network structures with different resolutions on the proposed SD7K dataset, where the batch size is set to 4. We also give more experiments on the low-resolution of $512 \times 512$ of Jung [16], Kligler [17], and SD7K for fair comparison following previous methods, we give more details in the supplementary materials. To further utilize the dataset, we also incorporate a simple shadow synthesis pipeline to further increase the diversity of the training process. We investigate several potential options, including alpha composition [26] and Perlin noise generation [11] for comparison, and we empirically choose alpha composition to synthesize shadows through the shadow-free image and the mask on the fly via random alpha values in the range between [0.2, 0.7]. + +As for the evaluation metrics, we employ three widely-used metrics: Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity (SSIM). While PSNR and SSIM are commonly used in image restoration evaluation [3, 5, 22], RMSE is the most important and frequently used metric specifically for shadow removal [6, 10]. + +# 5.2. Comparisons with State-of-the-Art Methods + +As presented in Tab. 2, we compared FSENet with both traditional and learning-based models of various resolutions using publicly available implementations of these methods, where the proposed method outperforms all other methodologies in most metrics. + +Furthermore, our analysis reveals that traditional methods, such as [16, 27], rely on physics-based assumptions that are not always correct, resulting in unstable performance across different datasets as shown in Fig. 6. Deep learning models, on the other hand, are highly dependent on + +the dataset magnitude, and their performance is relatively consistent across different scenarios. + +We also find that some deep learning models that are effective in low-resolution shadow removal do not scale well to high-resolution documents. They either compromise the aspect ratio of the document or lead to out-of-memory errors. While traditional methods can still remove shadows, they are time-consuming, requiring pixel-by-pixel calculations of illumination difference [27] and background estimation [30]. At present, only Mask-ShadowGAN [15] and the proposed FSENet can effectively remove document shadows in high-resolution scenarios. Despite having more parameters and requiring more time to produce results, FSENet outperforms other methods quantitatively. + +# 5.3. Ablation Studies + +As presented in Tab. 3 and Fig. 7, we conduct our ablation studies at the original high-resolution documents, examining the effectiveness of several components in our algorithm. These components include the depth of the Laplacian Pyramid, the efficiency of TRM, DFE, and DAT modules, and the effectiveness of our shadow synthesis pipeline. + +We first evaluate the depth of the proposed Laplacian Pyramid. From Tab. 3, with the depth of the Laplacian Pyramid increases (Our w/ Depth=3), the magnitude of the low-frequency components decreases, affecting the low-frequency deshadowing and the reconstructed outcome. Consequently, the model's performance progressively declines as the depth increases. We then evaluate the performance of the proposed DFE and DAT modules, which are also responsible for low-frequency deshading, primarily influencing the pixel-wise variation of the final result. Although each module individually contributes to enhancing shadow elimination performance, their combined network produces the best results. As for the high-frequency modules, the TRM module handles high-frequency components + +![](images/c8bf1ee1001d2333a56685fcae959ccd998592f9e8ce7d03160b259847ea7b4c.jpg) +Input + +![](images/b2b71b2473a74fc47d21e7090e0faa4a405a304f178d602bc752a83e750e9bd7.jpg) +Ours w/ Depth=3 + +![](images/5d49ad78ca0f402fd9156691721451865fa442f6f13852be1813a0f1ae517f36.jpg) +Ours w/o TRM + +![](images/cd754a23ac84ec75a9c029bc4c2474eb33d8b3dfd8d7fc03d4f06965a73de2dd.jpg) +Ours w/o DFE + +![](images/697bd36004825364ebaf1d0bbe529a15062849b4ca8ff9e707d54cd83eec9b19.jpg) +Ours w/o Synthesis + +![](images/846cfcdd1283cd3bf5d968fec9cf8c7f789139c23a13832db5edd4e858f24189.jpg) +Ours w/o DAT + +![](images/3b6b38b3df1b37d87a291cc928919d6bbdfe3c85f3dd3ffd75223e59c3059fe0.jpg) +Ours + +![](images/70e3eb1319516a0897be357754bcef8fa5beab42bc78c4764d06a2ffabcea459.jpg) +Target +Figure 7. Qualitative results of the ablation studies on different modules of the proposed FSENet. + +
Ablation StudyJung (Original)Kliger (Original)SD7K (Original)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
Input12.930.7961.3413.250.8156.7515.940.9144.14
Ours w/ Depth=320.970.8323.1926.330.9113.6620.830.9324.02
Ours w/o TRM19.500.8127.6027.100.9212.7818.440.8635.92
Ours w/o DFE21.900.8421.0225.160.9214.8919.700.8829.07
Ours w/o DAT22.680.8519.2822.700.8820.1919.030.8732.94
Ours w/o Synthesis22.210.8521.3324.790.9215.5422.940.9321.93
Ours Full23.600.8517.5628.980.9310.6028.670.9610.00
+ +Table 3. Quantitative results of the ablation study on three datasets in high-resolution. The best result is highlighted in red and bold. + +such as contours and edges, thus contributing significantly to the overall efficacy of our algorithm. Finally, we also evaluate the importance of the proposed shadow synthesis pipeline, where w/o shadow synthesis reduces the performance a lot, especially on the two small datasets. + +# 5.4. Limitations + +Although our model is capable of removing shadows in high-resolution images, the current network is not able to run on real-time or on edge devices, such as mobile phones, due to the larger parameters and heavy GPU computations. In the future, we will explore different approaches to further reduce the required parameters, making it usable on smart devices. + +# 6. Conclusion + +In this paper, we tackle the problem of removing shadows from high-resolution documents. To achieve this goal, we consider the contribution of the dataset and the network. As for the dataset, we gather a high-resolution dataset SD7K, which is comprised of over 7000 triplets of real- + +world document images with distinct characteristics under various illumination conditions. This dataset is an order of magnitude greater than previous datasets and allows us to validate our approach in a more diverse and comprehensive manner. As for the network structure, we propose FSENet, a frequency-aware method to remove shadows in higher resolution via frequency decomposition. To separate the high-frequency contours from the low-resolution content, we employ a Laplacian Pyramid and a series of meticulously designed attention blocks, enabling each component to be learned separately. Powered by our dataset and network, our method achieves a new state-of-the-art for document shadow removal. We believe both the network structure and the dataset will benefit the community for research on this topic. + +Acknowledgements. This work is supported in part by the University of Macau under Grant MYRG2022-00190-FST, in part by the Science and Technology Development Fund, Macau SAR, under Grant 0034/2019/AMJ, Grant 0087/2020/A2 and Grant 0049/2021/A. + +# References + +[1] Steve Bako, Soheil Darabi, Eli Shechtman, Jue Wang, Kalyan Sunkavalli, and Pradeep Sen. Removing shadows from images of documents. In ACCV, pages 173-183. Springer, 2016. 2, 3 +[2] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 2, 4 +[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33. Springer, 2022. 4, 5, 7 +[4] Xuhang Chen, Xiaodong Cun, Chi-Man Pun, and Shuqiang Wang. Shadocnet: Learning spatial-aware tokens in transformer for document shadow removal. In IEEE ICASSP, pages 1-5. IEEE, 2023. 2 +[5] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 7 +[6] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2, 6, 7 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2 +[8] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In CVPR, pages 10571-10580, 2021. 1, 2, 7 +[9] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 6 +[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 2, 7 +[11] John C Hart. Perlin noise pixel shaders. In Proceedings of the ACM SIGGRAPH/EUROGRAPHICS workshop on Graphics hardware, pages 87-94, 2001. 7 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2, 6 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 5 +[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 6 +[15] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2, 6, 7 +[16] Seungjun Jung, Muhammad Abul Hasan, and Changick Kim. Water-filling: An efficient algorithm for digitized document shadow removal. In Asian Conference on Computer Vision, pages 398-414. Springer, 2018. 1, 2, 3, 6, 7 + +[17] Netanel Kligler, Sagi Katz, and Ayellet Tal. Document enhancement using visibility detection. In CVPR, pages 2374-2382, 2018. 1, 2, 3, 7 +[18] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, August 2020. 2 +[19] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, pages 264-281. Springer, 2020. 2 +[20] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 2 +[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017. 7 +[22] Jingtang Liang, Xiaodong Cun, Chi-Man Pun, and Jue Wang. Spatial-separated curve rendering network for efficient and high-resolution image harmonization. In European Conference on Computer Vision, pages 334-349. Springer, 2022. 7 +[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In CVPR, pages 9392-9400, 2021. 2, 4, 5 +[24] Yun-Hsuan Lin, Wen-Chin Chen, and Yung-Yu Chuang. Bedsr-net: A deep shadow removal network from a single document image. In CVPR, pages 12905-12914, 2020. 2, 3, 7 +[25] Zhihao Liu, Hui Yin, Yang Mi, Mengyang Pu, and Song Wang. Shadow removal by a lightness-guided network with training on unpaired data. IEEE Transactions on Image Processing, 30:1853-1865, 2021. 1, 2 +[26] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In SIGGRAPH, pages 231–242, 1998. 7 +[27] Vatsal Shah and Vineet Gandhi. An iterative approach for shadow removal in document images. IEEE ICASSP, pages 1892-1896, 2018. 1, 3, 7 +[28] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1 +[29] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV, pages 361-378. Springer, 2022. 2 +[30] Bingshu Wang and CL Philip Chen. An effective background estimation method for shadows removal of document images. In IEEE International Conference on Image Processing (ICIP), pages 3611-3615. IEEE, 2019. 1, 3, 7 +[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2 + +[32] Jian-Ren Wang and Yung-Yu Chuang. Shadow removal of text document images by estimating local and global background colors. IEEE ICASSP, pages 1534-1538, 2020. 1, 3, 7 +[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. AAAI, 2023. 4, 5 +[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17683-17693, 2022. 2, 5 +[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In SIGKDD, pages 1192-1200, 2020. 1 +[37] Yimin Xu, Mingbao Lin, Hong Yang, Ke Li, Yunhang Shen, Fei Chao, and Rongrong Ji. Shadow-aware dynamic convolution for shadow removal. arXiv preprint arXiv:2205.04908, 2022. 6 +[38] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. *ICLR*, 2015. 6 +[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, June 2022. 7 \ No newline at end of file diff --git a/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/images.zip b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14dda206a0511617238c5844c6c8c5de61e144e1 --- /dev/null +++ b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c806e349b835cd7631638a4a9faf02a2330296899813dd736626acbb77d9ba52 +size 911457 diff --git a/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/layout.json b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9778baa5f36c7677cbc1905d83a9b7af475d9757 --- /dev/null +++ b/2023/High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net/layout.json @@ -0,0 +1,9249 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "type": "text", + "content": "High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 465, + 190 + ], + "type": "text", + "content": "Zinuo Li* Xuhang Chen* Chi-Man Pun† Xiaodong Cun† University of Macau" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 243, + 290, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 290, + 542 + ], + "type": "text", + "content": "Shadows often occur when we capture the document with casual equipment, which influences the visual quality and readability of the digital copies. Different from the algorithms for natural shadow removal, the algorithms in document shadow removal need to preserve the details of fonts and figures in high-resolution input. Previous works ignore this problem and remove the shadows via approximate attention and small datasets, which might not work in real-world situations. We handle high-resolution document shadow removal directly via a larger-scale real-world dataset and a carefully-designed frequency-aware network. As for the dataset, we acquire over 7k couples of high-resolution " + }, + { + "bbox": [ + 46, + 243, + 290, + 542 + ], + "type": "inline_equation", + "content": "(2462 \\times 3699)" + }, + { + "bbox": [ + 46, + 243, + 290, + 542 + ], + "type": "text", + "content": " images of real-world documents pairs with various samples under different lighting circumstances, which is 10 times larger than existing datasets. As for the design of the network, we decouple the high-resolution images in the frequency domain, where the low-frequency details and high-frequency boundaries can be effectively learned via the carefully designed network structure. Powered by our network and dataset, the proposed method shows a clearly better performance than previous methods in terms of visual quality and numerical results. The code, models, and dataset are available at https://github.com/CXH-Research/DocShadow-SD7K." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 588, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 660 + ], + "type": "text", + "content": "When occluders obstruct the light source, shadows are cast on the paper. These shadows reduce the readability and hinder subsequent text-related intelligent tasks such as document enhancement, optical character recognition (OCR), key information extraction, and semantic entity labeling [17, 36, 28], etc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 661, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 287, + 684 + ], + "type": "text", + "content": "Removing the cast shadow of the digital copy has a long need in the real world and also there are a lot of meth-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 216, + 384, + 329 + ], + "blocks": [ + { + "bbox": [ + 307, + 216, + 384, + 329 + ], + "lines": [ + { + "bbox": [ + 307, + 216, + 384, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 384, + 329 + ], + "type": "image", + "image_path": "28248561b5a96f35328a5d259765c94aa8659814812532f61cc9123f682b37dc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 329, + 384, + 380 + ], + "blocks": [ + { + "bbox": [ + 307, + 329, + 384, + 380 + ], + "lines": [ + { + "bbox": [ + 307, + 329, + 384, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 384, + 380 + ], + "type": "image", + "image_path": "a58314bacef103a319dbf819aa6b55f9edbe16120dffaedc308bf8265f676658.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 383, + 364, + 395 + ], + "lines": [ + { + "bbox": [ + 328, + 383, + 364, + 395 + ], + "spans": [ + { + "bbox": [ + 328, + 383, + 364, + 395 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 228, + 463, + 327 + ], + "blocks": [ + { + "bbox": [ + 389, + 228, + 463, + 327 + ], + "lines": [ + { + "bbox": [ + 389, + 228, + 463, + 327 + ], + "spans": [ + { + "bbox": [ + 389, + 228, + 463, + 327 + ], + "type": "image", + "image_path": "ada9ef044eab22ffb70a56b5816b42e5e29c695afe9c69b1b4a4b124a433333e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 388, + 329, + 465, + 380 + ], + "blocks": [ + { + "bbox": [ + 388, + 329, + 465, + 380 + ], + "lines": [ + { + "bbox": [ + 388, + 329, + 465, + 380 + ], + "spans": [ + { + "bbox": [ + 388, + 329, + 465, + 380 + ], + "type": "image", + "image_path": "3313398616b56ebe2fc1975c8b7236339da7c8d8d03cc98d76d941f1b50fabae.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 383, + 465, + 395 + ], + "lines": [ + { + "bbox": [ + 388, + 383, + 465, + 395 + ], + "spans": [ + { + "bbox": [ + 388, + 383, + 465, + 395 + ], + "type": "text", + "content": "(b) Shah et al. [27]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 468, + 216, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 468, + 216, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 468, + 216, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 468, + 216, + 545, + 329 + ], + "type": "image", + "image_path": "f2cae5e479813c2403bb3f00ba8334ec4981e4ad64ef4cffc5a54a239dea3aea.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 468, + 329, + 545, + 380 + ], + "blocks": [ + { + "bbox": [ + 468, + 329, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 468, + 329, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 468, + 329, + 545, + 380 + ], + "type": "image", + "image_path": "01b4daa90f855adf0da9244c45fd9be6d7dc0e6165bd89e5376bc12dc662201c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 383, + 537, + 395 + ], + "lines": [ + { + "bbox": [ + 475, + 383, + 537, + 395 + ], + "spans": [ + { + "bbox": [ + 475, + 383, + 537, + 395 + ], + "type": "text", + "content": "(c) Fu et al. [8]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 399, + 384, + 513 + ], + "blocks": [ + { + "bbox": [ + 307, + 399, + 384, + 513 + ], + "lines": [ + { + "bbox": [ + 307, + 399, + 384, + 513 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 384, + 513 + ], + "type": "image", + "image_path": "d8bdf7babb2ed771fb7b544ddca8ce680fbe6abd578c9a64fcc490e23d07d089.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 513, + 384, + 564 + ], + "blocks": [ + { + "bbox": [ + 307, + 513, + 384, + 564 + ], + "lines": [ + { + "bbox": [ + 307, + 513, + 384, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 384, + 564 + ], + "type": "image", + "image_path": "8da33ff966afc1227b2d19f6063a88ec014c8a50b65b9d4b317e44359efbf6d0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 567, + 381, + 578 + ], + "lines": [ + { + "bbox": [ + 310, + 567, + 381, + 578 + ], + "spans": [ + { + "bbox": [ + 310, + 567, + 381, + 578 + ], + "type": "text", + "content": "(d) Liu et al. [25]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 388, + 399, + 465, + 513 + ], + "blocks": [ + { + "bbox": [ + 388, + 399, + 465, + 513 + ], + "lines": [ + { + "bbox": [ + 388, + 399, + 465, + 513 + ], + "spans": [ + { + "bbox": [ + 388, + 399, + 465, + 513 + ], + "type": "image", + "image_path": "87f49fc9a1a0ca8a5a4b9e49553b42eeb5972a8a95d8703585afadb91836e09d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 388, + 513, + 465, + 564 + ], + "blocks": [ + { + "bbox": [ + 388, + 513, + 465, + 564 + ], + "lines": [ + { + "bbox": [ + 388, + 513, + 465, + 564 + ], + "spans": [ + { + "bbox": [ + 388, + 513, + 465, + 564 + ], + "type": "image", + "image_path": "c5df81b1c025cbda53448a1652da314adbe0511c77a6726d39d6f011e0019c28.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 567, + 460, + 578 + ], + "lines": [ + { + "bbox": [ + 392, + 567, + 460, + 578 + ], + "spans": [ + { + "bbox": [ + 392, + 567, + 460, + 578 + ], + "type": "text", + "content": "(e) Hu et al. [15]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 468, + 399, + 545, + 513 + ], + "blocks": [ + { + "bbox": [ + 468, + 399, + 545, + 513 + ], + "lines": [ + { + "bbox": [ + 468, + 399, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 468, + 399, + 545, + 513 + ], + "type": "image", + "image_path": "5ad08578990237c1cc1107e91198ec1632793025d7a85ce750802b97dbf9ae98.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 468, + 513, + 545, + 564 + ], + "blocks": [ + { + "bbox": [ + 468, + 513, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 468, + 513, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 468, + 513, + 545, + 564 + ], + "type": "image", + "image_path": "3ab42bdea443cf76f2a254b30a1996da4c1767e6c950f569492ac1401c862200.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 567, + 523, + 578 + ], + "lines": [ + { + "bbox": [ + 490, + 567, + 523, + 578 + ], + "spans": [ + { + "bbox": [ + 490, + 567, + 523, + 578 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 583, + 547, + 639 + ], + "lines": [ + { + "bbox": [ + 306, + 583, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 547, + 639 + ], + "type": "text", + "content": "Figure 1. Visual results of input document shadow image (a), classic method (b), supervised method (c), weakly-supervised method (d), unsupervised method (e) and ours (f). Our model removes the shadow while preserving the original document's content and aspect ratio." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "ods, including the traditional methods that applied physics-based illumination models [16, 30, 32]. In reality, these approaches are significantly constrained due to the fact that the assumptions utilized in illumination models [27] are often too rigid to account for real-world shadow. Thus, they may" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contributions" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12449" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "incorrectly compute the exposure and reflectance as shown in Fig. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 289, + 399 + ], + "type": "text", + "content": "Recent methods apply deep learning for this task since it has achieved remarkable progress in natural shadow removal [6] and other computer vision tasks [12, 13, 35]. However, several obstacles make it challenging to transfer previous natural shadow removal methods to the document domain. On the one hand, in contrast to natural shadow removal, documents are presented by high-resolution images in the real world, which needs to preserve fine-grained information such as fonts and figures. However, the majority of document shadow removal methods [24, 4] are only designed for relatively low-resolution images and is hard to handle high-resolution situations directly since they are relying on the low-resolution approximation of attention for guidance. On the other hand, there are no large-scale and high-resolution datasets designed specifically for document shadow removal at this time. Consequently, given limited image pairs for training, existing methods usually result in different artifacts (as shown in Fig. 1 (c)). Although weakly-supervised [18] and unsupervised [15, 25] shadow removal methods alleviate the requirements for data, they hold a strong assumption of the statistical similarity between shadowed and unshadowed domains. As illustrated in Fig. 1 (d) (e), once the training domain is very different from the test, these methods create hallucination contents and suffer from unstable results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 403, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 714 + ], + "type": "text", + "content": "To solve the above problems, we first offer a large-scale real-world shadow dataset, Shadowed Document 7K (SD7K), which contains more than 7k real-world and high-resolution shadow and shadow-free document image pairs with also the manually annotated shadow masks for related research, e.g., shadow detection [6]. Compared to existing document shadow removal datasets, SD7K includes various document types, e.g., manga, papers, and figures. Also, different from previous work which only considers one specific light environment, we use three distinct light sources to increase the diversity of the training samples. Besides the dataset, we also propose a robust network for high-resolution image processing specifically, named Frequency-aware Shadow Erasing Net (FSENet). In detail, inspired by the laplacian pyramid [2, 23], we divide the document image into two frequency portions, each with its own domain-specific properties. The low-frequency component is primarily responsible for illuminations or colors while high-frequency components are more related to content details. As for the low-frequency deshading module, we involve the transformer [7] based network to model the global illuminations alternation and a high-frequency restoration module based on cascaded aggregations for adaptive pixel enhancement. For high-frequency modules, we involve several dilated convolution-based texture recovery modules to restore the refined details similar to [6]. With the aforementioned" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "techniques, our model achieves the state-of-the-art performance of previous methods on two common small benchmarks and the proposed SD7K." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 109, + 514, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 109, + 514, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 109, + 514, + 120 + ], + "type": "text", + "content": "We summarize the main contribution as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 131, + 545, + 272 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 131, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 131, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 317, + 131, + 545, + 178 + ], + "type": "text", + "content": "- We provide SD7K, a large-scale real-world dataset consisting of high-resolution shadow and the associated shadow-free images under various illumination conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 190, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 190, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 190, + 545, + 224 + ], + "type": "text", + "content": "- We propose FSENet, a frequency-aware network with a carefully designed network structure to handle high-resolution document shadows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 236, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 236, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 545, + 272 + ], + "type": "text", + "content": "- Both qualitative and quantitative results on all available public datasets indicate that the proposed FSENet performs favorably against state-of-the-art methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 285, + 392, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 392, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 392, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 306, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 473 + ], + "type": "text", + "content": "Document Shadow Dataset. The collection of genuine shadow images and their corresponding shadow-free counterparts is a challenging task due to the varying lighting conditions. As a result, existing datasets are limited in size. The Bako dataset [1] comprises outdated, inaccessible images in tiny quantities. Similarly, Jung's dataset [16] is limited in size and favors light shadows. Kligler's dataset [17] emphasizes colorful symbols, which do not convey a significant amount of information. Among these datasets, the RDSRD dataset [24], while still modest in size, has a better quality in comparison. Different from the existing datasets, we propose a high-resolution and large-scale dataset SD7K that addresses the aforementioned problem, which can facilitate the research in document shadow removal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "Natural Shadow Removal. Methods for natural shadow removal might be informative and applicable to document images. On the one hand, GAN-based methods such as STCGAN [31], use two stacked conditional GANs to jointly detect and eliminate shadows. Mask-ShadowGAN [15] extend the paired GAN to unpaired setting via re-formulated cycle-consistency constraints. To fully utilize the dataset, DHAN [6] synthesize pseudo shadow images and learned border artifacts-free images via a dual hierarchical aggregation network. SG-ShadowNet [29] studies the importance of normalization and learns the style representation of the non-shadow region to harmonize the shadow and shadow-free parts. On the other hand, CNN-based methods such as " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SP + M}" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": " Net [19] and " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SP + M + I}" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": " Net [20] utilize the physical linear transformation model to enhance the shadow region and image decomposition to reconstruct the shadow-free image. AEF-Net [8] predicts the exposure parameters of the shadowed regions and fused a series of overexposed shadow images with the original shadow image to restore the shadow region. ShadowFormer [10] is a transformer" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12450" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 149 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 149 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 149 + ], + "type": "table", + "html": "
DatasetTriplets #Entities #Mean ResolutionDocument TypeWith MaskLight Temperature Control
Bako [1]81N/AN/ATextNoNo
Kligler [17]300252424 × 2013TextNoNo
Jung [16]8787602 × 860TextNoNo
RDSRD [24]540N/AN/ATextYesN/A
SD7K7620350+2462 × 3699Text and pictureYesDifferent lamps & sunlight
", + "image_path": "55fd7bde2bee0219654fd79164bf136d5e2a20f37ad6f6b3339e8e514aa67ca3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 155, + 546, + 178 + ], + "lines": [ + { + "bbox": [ + 46, + 155, + 546, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 546, + 178 + ], + "type": "text", + "content": "Table 1. Comparison of SD7K with other real-world datasets. The proposed dataset covers a large number of document shadow images and it is the only dataset that satisfies all important data features about document shadow currently." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 198, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 220 + ], + "type": "text", + "content": "based network to exploit the context correlation between shadow and non-shadow regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 222, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 288, + 378 + ], + "type": "text", + "content": "Document Image Shadow Removal. Traditional document shadow removal methods usually demand prior knowledge regarding shadow and document context. This includes estimating the background color [30, 32] and modeling shading and lighting effects [16, 27]. Deep learning-based methods are also designed to remove the shadows of documents. Currently, BEDSR-Net [24] stands as the state-of-the-art model in document shadow removal. It predicts global background color and encodes shadow position as an attention map, allowing for the effective removal of shadows. However, as discussed in our introduction, it needs to predict the approximate attention maps from the feature map, which is not suitable for the high resolution inputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 391, + 135, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 135, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 135, + 403 + ], + "type": "text", + "content": "3. SD7K Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 509 + ], + "type": "text", + "content": "As shown in Tab. 1 and Fig. 3 (a), there are only a limited number of datasets [16, 24, 1, 17] available for document shadow removal, which is typically small and comprises only hundreds of real-world samples. To train a network on such-scale datasets, synthesizing training samples via computer graphic engines is required [24]. However, the rendered samples still show different distributions from the real samples, making it difficult in real-world applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 568 + ], + "type": "text", + "content": "To overcome this limitation, we introduce a massive real-world dataset named SD7K for document shadow removal in terms of the occluder entities, the numbers, the resolution, the light source, etc. Below, we give the details of how to collect and process the dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "First, we use a Canon EOS M6 camera to capture printed document images from real-world scenarios. To capture paired shadow and shadow-free images, we make all the light close and leave one light source. All the settings of the camera are also fixed, including focal length, aperture, exposure, and ISO, to avoid the influence of the camera algorithm. Then, we use a camera stand to hold the camera and make it shoot on the paper vertically as shown in Fig. 2 (a). To capture the paired shadow and shadow-free image without involving the changes caused by the manual presses, as in Fig. 2 (b), we use a remote control to capture the photo between the shadowed and shadow-free paper by removing" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 195, + 384, + 311 + ], + "blocks": [ + { + "bbox": [ + 307, + 195, + 384, + 311 + ], + "lines": [ + { + "bbox": [ + 307, + 195, + 384, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 384, + 311 + ], + "type": "image", + "image_path": "92ade6c40849ca1920b29d8c785b2725bd4ca06b8850112fa8e87a8c1751b336.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 314, + 352, + 324 + ], + "lines": [ + { + "bbox": [ + 340, + 314, + 352, + 324 + ], + "spans": [ + { + "bbox": [ + 340, + 314, + 352, + 324 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 387, + 195, + 465, + 311 + ], + "blocks": [ + { + "bbox": [ + 387, + 195, + 465, + 311 + ], + "lines": [ + { + "bbox": [ + 387, + 195, + 465, + 311 + ], + "spans": [ + { + "bbox": [ + 387, + 195, + 465, + 311 + ], + "type": "image", + "image_path": "ba9ec9be2845cbbd327b3554db80b7b1bfcac9937ec554fe8614a442c33bb9c4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 314, + 433, + 324 + ], + "lines": [ + { + "bbox": [ + 419, + 314, + 433, + 324 + ], + "spans": [ + { + "bbox": [ + 419, + 314, + 433, + 324 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 467, + 195, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 467, + 195, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 467, + 195, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 467, + 195, + 545, + 311 + ], + "type": "image", + "image_path": "0780e631cfb7cd38c83c6109a0d5418dbf1aebf3261d2d13fe91a352a83757af.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 314, + 512, + 324 + ], + "lines": [ + { + "bbox": [ + 501, + 314, + 512, + 324 + ], + "spans": [ + { + "bbox": [ + 501, + 314, + 512, + 324 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 378, + 426, + 460 + ], + "blocks": [ + { + "bbox": [ + 309, + 378, + 426, + 460 + ], + "lines": [ + { + "bbox": [ + 309, + 378, + 426, + 460 + ], + "spans": [ + { + "bbox": [ + 309, + 378, + 426, + 460 + ], + "type": "image", + "image_path": "3971a0ab029c7bea5f66208b8a97e4f9b59c8aa6086467d9e60a5779c1df6d87.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 463, + 423, + 474 + ], + "lines": [ + { + "bbox": [ + 310, + 463, + 423, + 474 + ], + "spans": [ + { + "bbox": [ + 310, + 463, + 423, + 474 + ], + "type": "text", + "content": "(a) Quantitative comparison" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 479, + 545, + 501 + ], + "lines": [ + { + "bbox": [ + 306, + 479, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 545, + 501 + ], + "type": "text", + "content": "Figure 3. Data distribution of SD7K and quantitative comparison across all document shadow datasets." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 429, + 377, + 544, + 452 + ], + "blocks": [ + { + "bbox": [ + 305, + 329, + 546, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 329, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 546, + 373 + ], + "type": "text", + "content": "Figure 2. (a) our data acquisition setup for constructing the dataset. (b) a remote phone control shutter where the user only needs to click a button to capture a pair of shadow-free/shadow images. (c) some showcases of our occluders." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 377, + 544, + 452 + ], + "lines": [ + { + "bbox": [ + 429, + 377, + 544, + 452 + ], + "spans": [ + { + "bbox": [ + 429, + 377, + 544, + 452 + ], + "type": "image", + "image_path": "3851803b58f5cbb5b8ff55f23ae8252460da19811130a35483792220ad985e44.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 462, + 529, + 474 + ], + "lines": [ + { + "bbox": [ + 443, + 462, + 529, + 474 + ], + "spans": [ + { + "bbox": [ + 443, + 462, + 529, + 474 + ], + "type": "text", + "content": "(b) Light distribution" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 510, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 545, + 533 + ], + "type": "text", + "content": "the occluder. We also take the pairs as short as possible to avoid the light difference between captures." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 534, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 653 + ], + "type": "text", + "content": "To further enhance diversity, we alternated the positions of the document and occluder after capturing a few batches of documents. Overall, we use over 30 types of occluders along with more than 350 documents to contribute to the dataset, some of the occluders are shown in Fig. 2 (c). These occluders have the shape of both regular and irregular forms, which provides adequate coverage for various situations, we calculate the shadow area percentage of the whole dataset, which is " + }, + { + "bbox": [ + 304, + 534, + 546, + 653 + ], + "type": "inline_equation", + "content": "41.38\\% \\pm 13.58\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 653 + ], + "type": "text", + "content": ", indicating that it is sufficiently divergent for our shadow removal task." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "We also consider the fluctuations in sunlight intensity and random placement of shadows, which will be significantly influenced by the light variance. To simulate the results in different light sources, we introduce three types of light sources in our dataset for collection, including reg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12451" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 126, + 186 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 126, + 186 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 126, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 126, + 186 + ], + "type": "image", + "image_path": "743ae3ba46e21e912060a27d8559cad02da70f55ec3ef8cf72cef89640311f1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 206, + 186 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 206, + 186 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 206, + 186 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 206, + 186 + ], + "type": "image", + "image_path": "f7df2cb277b10a39bcbfae849c6eb66b057f287807b34ff6085eb1c29ab4ee82.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 209, + 70, + 287, + 186 + ], + "blocks": [ + { + "bbox": [ + 209, + 70, + 287, + 186 + ], + "lines": [ + { + "bbox": [ + 209, + 70, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 209, + 70, + 287, + 186 + ], + "type": "image", + "image_path": "5b8298c76ad7ea611cbc46abc3b2407a59330f2cbb769867eca96e29938d215e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 186, + 126, + 302 + ], + "blocks": [ + { + "bbox": [ + 49, + 186, + 126, + 302 + ], + "lines": [ + { + "bbox": [ + 49, + 186, + 126, + 302 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 126, + 302 + ], + "type": "image", + "image_path": "a6902a37fa99561e7933b2991cabfcbd781acae515ef2540f19fa28de99871d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 304, + 108, + 316 + ], + "lines": [ + { + "bbox": [ + 66, + 304, + 108, + 316 + ], + "spans": [ + { + "bbox": [ + 66, + 304, + 108, + 316 + ], + "type": "text", + "content": "Cool light" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 187, + 206, + 302 + ], + "blocks": [ + { + "bbox": [ + 129, + 187, + 206, + 302 + ], + "lines": [ + { + "bbox": [ + 129, + 187, + 206, + 302 + ], + "spans": [ + { + "bbox": [ + 129, + 187, + 206, + 302 + ], + "type": "image", + "image_path": "0d54e187e74aa06fe4b74897d7c48018bf5e9fe8a33e2545d8e8e97630f776ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 304, + 190, + 316 + ], + "lines": [ + { + "bbox": [ + 145, + 304, + 190, + 316 + ], + "spans": [ + { + "bbox": [ + 145, + 304, + 190, + 316 + ], + "type": "text", + "content": "Warm light" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 209, + 186, + 287, + 302 + ], + "blocks": [ + { + "bbox": [ + 209, + 186, + 287, + 302 + ], + "lines": [ + { + "bbox": [ + 209, + 186, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 209, + 186, + 287, + 302 + ], + "type": "image", + "image_path": "334c8217b4404318abfb486c6f029e1369960eb2f68e018ebd3258fa6b717bdf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 304, + 265, + 316 + ], + "lines": [ + { + "bbox": [ + 230, + 304, + 265, + 316 + ], + "spans": [ + { + "bbox": [ + 230, + 304, + 265, + 316 + ], + "type": "text", + "content": "Sunlight" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 321, + 282, + 333 + ], + "lines": [ + { + "bbox": [ + 47, + 321, + 282, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 282, + 333 + ], + "type": "text", + "content": "Figure 4. Example shadow and shadow-free images from SD7K." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 343, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 392 + ], + "type": "text", + "content": "cular sunlight, a stationary 80W LED yellow light having a color temperature of " + }, + { + "bbox": [ + 46, + 343, + 287, + 392 + ], + "type": "inline_equation", + "content": "3000\\mathrm{K}" + }, + { + "bbox": [ + 46, + 343, + 287, + 392 + ], + "type": "text", + "content": ", and a stationary 80W LED white light having a color temperature of " + }, + { + "bbox": [ + 46, + 343, + 287, + 392 + ], + "type": "inline_equation", + "content": "6000\\mathrm{K}" + }, + { + "bbox": [ + 46, + 343, + 287, + 392 + ], + "type": "text", + "content": ". The light distribution of the whole dataset is shown in Fig. 3 (b)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 499 + ], + "type": "text", + "content": "For data cleaning purposes, we first remove images with noticeable motion distortion or lack of focus. Since some photo still contains outliers and the context is duplicated, we then review the photographs in groups and remove them. Finally, we crop the areas outside of the margins of the document, resulting in a total of 7620 samples, where some of the samples are shown in Fig. 4. We also use a threshold-based method to extract the mask of the training pair since the shadow mask will also benefit related tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 513, + 261, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 261, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 261, + 527 + ], + "type": "text", + "content": "4. Frequency-Aware Shadow Erasing Net" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "Since real-world document shadow removal needs huge GPU memory requirements, our network initially divides the input image into different frequency components using the Laplacian Pyramid (LP) similar to [23] to reduce the computational budgets. As shown in Fig. 5, we propose a powerful frequency-aware shadow removal network named FSENet by designing different modules to handle the high-frequency and low-frequency computation individually. After training, the multi-frequency components are mixed to produce the final results. In the rest of this section, we first give the details of the Laplacian pyramid in Sec. 4.1. Then, in Sec. 4.2 and Sec. 4.3, we give the details of the low-frequency deshading module and high-frequency restoration module to precisely modify the color details and the high-frequency details individually. Finally, we give the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 411, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 411, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 411, + 83 + ], + "type": "text", + "content": "loss functions in Sec. 4.4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 91, + 526, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 526, + 105 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 526, + 105 + ], + "type": "text", + "content": "4.1. Laplacian Pyramid and Method Overview" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 110, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 193 + ], + "type": "text", + "content": "Laplacian Pyramid (LP) [2] plays an important role in image processing. By decomposing and restoring images in a lossless manner, LP provides a rapid processing speed and is well-suited for dealing with high-resolution images. Inspired by [23] for image translation, we implement LP for document shadow removal and train the network on multifrequency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": "In detail, given an image " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " layers LP contains several high-frequency components " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "[L_0, \\dots, L_{D-1}]" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " and a low-frequency component " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "L_D \\in \\mathbb{R}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times 3}" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "L_D" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " times downsample of " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": ". Each high-frequency component can be described as " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "L_i = I_i - \\mathrm{PyrUp}(\\mathrm{PyrDown}(I_i))" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": "-th down-sampled image and " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "\\mathrm{PyrUp}" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "\\mathrm{PyrDown}" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " stand for upsampling and downsampling, respectively. As illustrated in Fig. 5, the " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " include the high-frequency component edges and textures, while the low-frequency component " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "L_3 = I_3" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " majorly contains the color information. For reversible reconstruction, the final " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " can be obtained by reversed update the equation of " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "\\hat{I}_{D-1} = \\mathrm{PyrUp}(\\hat{I}_D) + \\hat{L}_{D-1}" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "\\hat{I}_D" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " is the learned low-frequency features and " + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "inline_equation", + "content": "\\hat{L}_{D-1}" + }, + { + "bbox": [ + 304, + 194, + 546, + 388 + ], + "type": "text", + "content": " is the learned high-frequency features from the proposed network. We give detailed instructions for each network below." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 395, + 493, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 493, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 493, + 407 + ], + "type": "text", + "content": "4.2. Low-Frequency Deshading Module" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "content": "After getting the low-frequency components, we use several transformer-based blocks to reduce the color and illumination distortions of the shadow images. Compared to CNNs, transformers are superior at modeling non-local self-similarity and long-range dependencies. As shown in Fig. 5, we implement a transformer-based module called Dimension-Aware Transformer (DAT) inspired by [33] with a lightweight UNet-like sub-network [3] for low-frequency feature modeling." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "text", + "content": "Formally, given a low-frequency component " + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "inline_equation", + "content": "L_{3} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 3}" + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "text", + "content": " of input image " + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "text", + "content": ", we first feed " + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "inline_equation", + "content": "L_{3}" + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "text", + "content": " into three sequential DAT blocks to extract rich features. Each DAT block is built via a sequential height-attention and width-attention to learn the feature in different aspects individually and a convolution-based layer for local feature merging, where we denote the intermediate features of each DAT block as " + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "inline_equation", + "content": "F1, F2, F3 \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}" + }, + { + "bbox": [ + 304, + 521, + 545, + 639 + ], + "type": "text", + "content": ". Please refer to the original [33] for further details of the height-attention and width-attention." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": "We then feed these features " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "F1, F2, F3" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " into a Tri-layer Attention Alignment (TAA) block for feature remixing. Given concatenated features " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{in}} \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\times \\frac{W}{4} \\times C} = \\text{Concat}([F1, F2, F3])" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " in this case, " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " is first reshaped to " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{in}} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 3C}" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": ". Then, we send " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 640, + 545, + 713 + ], + "type": "text", + "content": " convolution layer to aggregate pixel-level cross" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12452" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 74, + 538, + 262 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 538, + 262 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 538, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 538, + 262 + ], + "type": "image", + "image_path": "809596f7670c898143aa330172b16331ccdde3b95a768080cdcafdb18eb9ff71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "text", + "content": "Figure 5. The network structure of our proposed FSENet. Following [23], given a high-resolution image " + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "text", + "content": ", we first use Laplacian Pyramid (e.g. Depth " + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "inline_equation", + "content": "D = 2" + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "text", + "content": " in this case) to decompose the images to multiple frequency components. The Black arrows: The low-frequency part " + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "inline_equation", + "content": "L_{3} \\in \\mathbb{R}_{\\frac{H}{2D}}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times C}" + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "text", + "content": " are refined to " + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "inline_equation", + "content": "L_{3}' \\in \\mathbb{R}_{\\frac{H}{2D}}^{\\frac{H}{2D} \\times \\frac{W}{2D} \\times C}" + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "text", + "content": " utilizing DAT, TAA and DFE block. Cyan arrows: For the high-frequency part " + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "inline_equation", + "content": "L_{2} \\in \\mathbb{R}_{\\frac{H}{2D-1}}^{\\frac{H}{2D-1} \\times \\frac{W}{2D-1} \\times C}" + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "text", + "content": ", a contour " + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "inline_equation", + "content": "C_{L_{2}} \\in \\mathbb{R}_{\\frac{H}{2D-1}}^{\\frac{H}{2D-1} \\times \\frac{W}{2D-1} \\times 1}" + }, + { + "bbox": [ + 46, + 270, + 548, + 342 + ], + "type": "text", + "content": " is learned to bridge the low-frequency and high-frequency features. Red arrows: For the remaining components with higher frequency, the learned contour is successively upsampled and refined using the proposed SPP and TRM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": "channel context, followed by three " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " depth-wise convolution layers to generate query " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": ", individually. The " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}, \\mathbf{K}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " are then reshaped to " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Q}} \\in \\mathbb{R}^{N \\times \\frac{H}{4} \\frac{W}{4} C}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{K}} \\in \\mathbb{R}^{\\frac{H}{4} \\frac{W}{4} C \\times N}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " respectively to calculate an attention matrix " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{\\mathbf{N} \\times \\mathbf{N}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": ". Next, we reshape the " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}} \\in \\mathbb{R}^{\\frac{H}{4} \\frac{W}{4} C \\times N}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " to multiply " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": ", and the input feature " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": " is added to generate the output feature " + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 46, + 361, + 289, + 456 + ], + "type": "text", + "content": ", which can be described as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 461, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 461, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 91, + 461, + 287, + 475 + ], + "type": "interline_equation", + "content": "F 4 = \\operatorname {C o n v} _ {1 \\times 1} L A (\\hat {\\mathbf {Q}}, \\hat {\\mathbf {K}}, \\hat {\\mathbf {V}}) + \\hat {F} _ {\\mathrm {i n}}, \\tag {1}", + "image_path": "ca5425c330800b1497cdcccb64dd2c09e6cd25a3d4da76870d33cd2fa653e5a1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 480, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 480, + 287, + 516 + ], + "type": "inline_equation", + "content": "Conv_{1\\times 1}" + }, + { + "bbox": [ + 46, + 480, + 287, + 516 + ], + "type": "text", + "content": " represents a convolution layer with the kernel of " + }, + { + "bbox": [ + 46, + 480, + 287, + 516 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 480, + 287, + 516 + ], + "type": "text", + "content": ". LA indicates Layer Attention [33], which can be written as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 520, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 520, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 88, + 520, + 287, + 536 + ], + "type": "interline_equation", + "content": "L A (\\hat {\\mathbf {Q}}, \\hat {\\mathbf {K}}, \\hat {\\mathbf {V}}) = \\hat {\\mathbf {V}} \\operatorname {s o f t m a x} (\\hat {\\mathbf {Q}} \\hat {\\mathbf {K}} / \\alpha), \\tag {2}", + "image_path": "563356a99c18a5fa9e3cb56bcef15cda5877d3902cbcd3d2d10b8e10a496138e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 540, + 287, + 563 + ], + "type": "text", + "content": " is the temperature of the attention. Please refer to the supplementary material for more details of the TAA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "content": "After getting " + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "inline_equation", + "content": "F4" + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "content": ", as shown in Fig. 5, We also use a UNet-like [3, 35] sub-network built by Deeper Feature extraction (DFE) blocks to learn multi-scale features. Each DFE block begins with a LayerNorm to stabilize the training procedure, followed by two convolutions and a deformable convolution. Then, it utilizes SimpleGate [3] to increase the diversity of the features by dividing the features along the channel dimension into two parts and multiplying them together. We also utilize Simplified Channel Attention (SCA) similar to [3] to capture global information in a computationally efficient manner, which can be denoted as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "S C A (\\mathbf {X}) = \\mathbf {X} * \\mathbf {W} \\operatorname {p o o l} (\\mathbf {X}), \\tag {3}", + "image_path": "76845f0fcea567009b74be18f14cf8cef183c3e3e87d7e309fc5f704dd03b0ff.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "text", + "content": " represents a fully-connected layer, pool represents the global average pooling procedure that combines spatial data into channels, and " + }, + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 304, + 361, + 545, + 444 + ], + "type": "text", + "content": " represents a channel-wise multiplication. The skip-connections [13] are connected before the LayerNorm to the output similar to [3]. We also give a detailed figure of the DFE block in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "text", + "content": "For the refined feature " + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "inline_equation", + "content": "\\hat{F}^4" + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "text", + "content": ", it then passes through three transformer blocks and a TAA Block to produce the enhanced features for image reconstruction using DAT and TAA blocks similar to the feature extractions of " + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "inline_equation", + "content": "F1, F2, F3" + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "text", + "content": ". Finally, a convolution layer with the kernel size " + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "text", + "content": " is performed on the enhanced features to produce the low-frequency component " + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "inline_equation", + "content": "L_3' \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}" + }, + { + "bbox": [ + 304, + 445, + 546, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 539, + 503, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 503, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 503, + 552 + ], + "type": "text", + "content": "4.3. High-Frequency Restoration Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "After getting the low-frequency features, we also design novel components to recover the high-frequency details of the learned features. In detail, we first upsample " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "L_{3}" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " and its corresponding result " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "L_{3}^{\\prime}" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " to match the size of " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": ", then we learn a contour on the concatenation of " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "[up(L_3), up(L_3'), L_2]" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " via several residual blocks [13], where " + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "inline_equation", + "content": "up(\\cdot)" + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": " means bilinear upsample operation, as illustrated in Fig. 5. This is a straightforward method for optimizing global correction compared to normal RGB images because the bands of the high-frequency feature map are very sparse, allowing us to reduce the volume of calculations as discussed in [23]. This contour will be expanded progressively into a collection of contours to match" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12453" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 145, + 227 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 145, + 227 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 145, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 145, + 227 + ], + "type": "image", + "image_path": "811c2e3d6a36dc1b4e14644b5d54f44f92fdb06e0c42fe4e6acc49d9cfb9d430.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 148, + 70, + 245, + 227 + ], + "blocks": [ + { + "bbox": [ + 148, + 70, + 245, + 227 + ], + "lines": [ + { + "bbox": [ + 148, + 70, + 245, + 227 + ], + "spans": [ + { + "bbox": [ + 148, + 70, + 245, + 227 + ], + "type": "image", + "image_path": "74164336ac43927d45013c270e30b5f3f13d72b3823f2c0d4c9a5399a3b35a6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 248, + 70, + 345, + 227 + ], + "blocks": [ + { + "bbox": [ + 248, + 70, + 345, + 227 + ], + "lines": [ + { + "bbox": [ + 248, + 70, + 345, + 227 + ], + "spans": [ + { + "bbox": [ + 248, + 70, + 345, + 227 + ], + "type": "image", + "image_path": "fd8c0bafec9d9c4762f2c68ec2b00102f6c001c5ac12f67fc54020df493447e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 349, + 70, + 446, + 227 + ], + "blocks": [ + { + "bbox": [ + 349, + 70, + 446, + 227 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 446, + 227 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 446, + 227 + ], + "type": "image", + "image_path": "51233e8dd291e6af8272dfd95225800d5c7e629b65b83acaad949dbfaaada90b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 449, + 70, + 546, + 227 + ], + "blocks": [ + { + "bbox": [ + 449, + 70, + 546, + 227 + ], + "lines": [ + { + "bbox": [ + 449, + 70, + 546, + 227 + ], + "spans": [ + { + "bbox": [ + 449, + 70, + 546, + 227 + ], + "type": "image", + "image_path": "052222441240f32fe624d20415976a38a0d77f4e07765e852893031c63a9864f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 47, + 228, + 145, + 417 + ], + "blocks": [ + { + "bbox": [ + 47, + 228, + 145, + 417 + ], + "lines": [ + { + "bbox": [ + 47, + 228, + 145, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 145, + 417 + ], + "type": "image", + "image_path": "d95750c311867ea181e1d9c44d6a3814ec267b64402fc6674d657d74dc9ee0ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 422, + 546, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 546, + 445 + ], + "type": "text", + "content": "Figure 6. Qualitative results of the methods comparison in high resolution, the first row represents the results on the Kligler dataset and the second row depicts the results on the SD7K dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 147, + 230, + 244, + 418 + ], + "blocks": [ + { + "bbox": [ + 147, + 230, + 244, + 418 + ], + "lines": [ + { + "bbox": [ + 147, + 230, + 244, + 418 + ], + "spans": [ + { + "bbox": [ + 147, + 230, + 244, + 418 + ], + "type": "image", + "image_path": "bba7869dc5a5072f88b969330c0b4fb6118831728d8c9dc73372cefb2261debf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 247, + 230, + 345, + 418 + ], + "blocks": [ + { + "bbox": [ + 247, + 230, + 345, + 418 + ], + "lines": [ + { + "bbox": [ + 247, + 230, + 345, + 418 + ], + "spans": [ + { + "bbox": [ + 247, + 230, + 345, + 418 + ], + "type": "image", + "image_path": "958643130421a86ba976679e75915cfc439838dfd7af486159017b0bd5fc53de.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 349, + 230, + 446, + 418 + ], + "blocks": [ + { + "bbox": [ + 349, + 230, + 446, + 418 + ], + "lines": [ + { + "bbox": [ + 349, + 230, + 446, + 418 + ], + "spans": [ + { + "bbox": [ + 349, + 230, + 446, + 418 + ], + "type": "image", + "image_path": "1a4ceb403dec194759a11a0e52a790c814b46fae5fdd82fb1e12cf7487b705ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 449, + 230, + 546, + 418 + ], + "blocks": [ + { + "bbox": [ + 449, + 230, + 546, + 418 + ], + "lines": [ + { + "bbox": [ + 449, + 230, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 449, + 230, + 546, + 418 + ], + "type": "image", + "image_path": "f57eb4477adf4279f8d6fb43a845c9754ed2ec17d3f2a8eca2e7e9449c6faabe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": "the remaining high-frequency components. As illustrated in Fig. 5, given the high-frequency component " + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": " and the contour " + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "inline_equation", + "content": "C_{L_2}" + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "inline_equation", + "content": "L_{2}^{\\prime}" + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": " can be written as " + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "inline_equation", + "content": "L_2^\\prime = L_2\\otimes C_{L_2}" + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 46, + 464, + 287, + 514 + ], + "type": "text", + "content": " represents the pixel-wise multiplication." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "text", + "content": "After multiplication, the Texture Recovery Module (TRM) is then designed to improve elaboration, inspired by the successful natural shadow removal methods [6, 37], which contain a series of the dilated convolutions and attentive aggregation nodes to recover the detailed features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "In detail, to increase the reception fields on the high-resolution map, we incorporate dilated convolution [38]. Dilated convolution inserts spaces between the kernel values, effectively increasing the receptive field of the filter without losing resolution or coverage over the input. The amount of dilation is controlled by a dilation rate parameter. For a dilation rate of 1, the layer behaves like a regular convolution. For rates larger than 1, the filter is effectively widened to have a larger field of view. This allows the network to incorporate a wider context and capture features at different scales in a computationally effi" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 464, + 547, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 596 + ], + "type": "text", + "content": "cient manner compared to using very large filters. As for Attentive Aggregation Node, to re-weight the significance of feature maps, each aggregation node employs a squeeze-and-excitation block [14] to reweight the feature. A " + }, + { + "bbox": [ + 304, + 464, + 547, + 596 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 464, + 547, + 596 + ], + "type": "text", + "content": " convolution layer is then employed to compress the features and match the original channels. Finally, a spatial pooling pyramid (SPP) [12] is introduced at the conclusion of the TRM to facilitate the remixing of multi-context features. After getting the learned high-frequency of each layer, we use the method introduced in Sec. 4.1 to recover the image and calculate the supervision." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 609, + 397, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 397, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 397, + 621 + ], + "type": "text", + "content": "4.4. Loss functions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "For training, we utilize the loss functions of SmoothL [9] " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l1}^{smooth}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and SSIM [34] " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SSIM}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " between the target image and the reconstructed image from LP, both two losses are helpful to preserve the details of the image, which can be represented as: " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{total} = \\mathcal{L}_{l1}^{smooth} + \\lambda * \\mathcal{L}_{SSIM}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", where we set the weight of " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SSIM} \\lambda" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " to 0.4 empirically." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12454" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 547, + 211 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 211 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 211 + ], + "type": "table", + "html": "
MethodSD7KParam (M)
512 × 5121024 × 10242462 × 3699 (Full Size)
PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓PSNR↑SSIM↑RMSE↓Time(s) ↓
Input15.950.8944.09N/A15.950.9044.09N/A15.940.9144.14N/AN/A
Wang et al. [32]15.310.8247.88>1015.310.8547.88>3015.290.8647.95>60N/A
Wang et al. [30]13.320.6867.48>1013.320.7167.48>3013.260.7468.07>60N/A
Shah et al. [27]9.890.7186.35>109.900.7686.35>309.880.7986.46>60N/A
Jung et al. [16]19.860.9226.76>519.860.9326.76>1019.820.9226.86>30N/A
MaskGAN [15]24.820.8715.430.9624.820.8515.432.0724.670.8615.727.5928.29
DHAN [6]25.610.8514.270.2125.60*0.83*14.27*N/A25.42*0.85*14.60*N/A27.82
AEFNet [8]24.180.9516.830.1123.00*0.90*19.14*N/A22.94*0.90*19.28*N/A391.10
BEDSR-Net [24]21.500.9030.520.8519.81*0.85*33.03*N/A19.74*0.86*33.18*N/A29.44
ShadowFormer [10]23.710.9017.540.1522.69*0.84*19.65*N/A22.64*0.87*19.79*N/A11.35
BMNet [39]24.860.8015.590.1824.84*0.79*15.67*N/A24.70*0.83*15.96*N/A2.11
Ours28.690.979.980.2428.680.979.980.7528.670.9610.007.9329.34
", + "image_path": "4a1277b95bf2f12107e3baa78d3db619343a560f756892882c3ac1ce7fa92d6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 219, + 546, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 546, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 546, + 253 + ], + "type": "text", + "content": "Table 2. Quantitative results of the ablation study on three datasets in different resolutions. The best result is highlighted in red and bold. The asterisk (*) indicates that the result is obtained from a low-resolution result " + }, + { + "bbox": [ + 46, + 219, + 546, + 253 + ], + "type": "inline_equation", + "content": "(512\\times 512)" + }, + { + "bbox": [ + 46, + 219, + 546, + 253 + ], + "type": "text", + "content": " and then improved using a super-resolution technique [21] to address an out-of-memory error." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 271, + 128, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 128, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 128, + 284 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 291, + 180, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 180, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 180, + 304 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 311, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 477 + ], + "type": "text", + "content": "We evaluate the performance of different network structures with different resolutions on the proposed SD7K dataset, where the batch size is set to 4. We also give more experiments on the low-resolution of " + }, + { + "bbox": [ + 46, + 311, + 287, + 477 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 311, + 287, + 477 + ], + "type": "text", + "content": " of Jung [16], Kligler [17], and SD7K for fair comparison following previous methods, we give more details in the supplementary materials. To further utilize the dataset, we also incorporate a simple shadow synthesis pipeline to further increase the diversity of the training process. We investigate several potential options, including alpha composition [26] and Perlin noise generation [11] for comparison, and we empirically choose alpha composition to synthesize shadows through the shadow-free image and the mask on the fly via random alpha values in the range between [0.2, 0.7]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 563 + ], + "type": "text", + "content": "As for the evaluation metrics, we employ three widely-used metrics: Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity (SSIM). While PSNR and SSIM are commonly used in image restoration evaluation [3, 5, 22], RMSE is the most important and frequently used metric specifically for shadow removal [6, 10]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 573, + 277, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 277, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 277, + 586 + ], + "type": "text", + "content": "5.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 652 + ], + "type": "text", + "content": "As presented in Tab. 2, we compared FSENet with both traditional and learning-based models of various resolutions using publicly available implementations of these methods, where the proposed method outperforms all other methodologies in most metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Furthermore, our analysis reveals that traditional methods, such as [16, 27], rely on physics-based assumptions that are not always correct, resulting in unstable performance across different datasets as shown in Fig. 6. Deep learning models, on the other hand, are highly dependent on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 271, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 295 + ], + "type": "text", + "content": "the dataset magnitude, and their performance is relatively consistent across different scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 297, + 546, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 546, + 441 + ], + "type": "text", + "content": "We also find that some deep learning models that are effective in low-resolution shadow removal do not scale well to high-resolution documents. They either compromise the aspect ratio of the document or lead to out-of-memory errors. While traditional methods can still remove shadows, they are time-consuming, requiring pixel-by-pixel calculations of illumination difference [27] and background estimation [30]. At present, only Mask-ShadowGAN [15] and the proposed FSENet can effectively remove document shadows in high-resolution scenarios. Despite having more parameters and requiring more time to produce results, FSENet outperforms other methods quantitatively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 453, + 406, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 406, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 406, + 464 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 544 + ], + "type": "text", + "content": "As presented in Tab. 3 and Fig. 7, we conduct our ablation studies at the original high-resolution documents, examining the effectiveness of several components in our algorithm. These components include the depth of the Laplacian Pyramid, the efficiency of TRM, DFE, and DAT modules, and the effectiveness of our shadow synthesis pipeline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "We first evaluate the depth of the proposed Laplacian Pyramid. From Tab. 3, with the depth of the Laplacian Pyramid increases (Our w/ Depth=3), the magnitude of the low-frequency components decreases, affecting the low-frequency deshadowing and the reconstructed outcome. Consequently, the model's performance progressively declines as the depth increases. We then evaluate the performance of the proposed DFE and DAT modules, which are also responsible for low-frequency deshading, primarily influencing the pixel-wise variation of the final result. Although each module individually contributes to enhancing shadow elimination performance, their combined network produces the best results. As for the high-frequency modules, the TRM module handles high-frequency components" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12455" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 170, + 156 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 170, + 156 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 170, + 156 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 170, + 156 + ], + "type": "image", + "image_path": "c8bf1ee1001d2333a56685fcae959ccd998592f9e8ce7d03160b259847ea7b4c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 159, + 120, + 170 + ], + "lines": [ + { + "bbox": [ + 97, + 159, + 120, + 170 + ], + "spans": [ + { + "bbox": [ + 97, + 159, + 120, + 170 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 295, + 156 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 295, + 156 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 295, + 156 + ], + "type": "image", + "image_path": "b2b71b2473a74fc47d21e7090e0faa4a405a304f178d602bc752a83e750e9bd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 158, + 269, + 170 + ], + "lines": [ + { + "bbox": [ + 198, + 158, + 269, + 170 + ], + "spans": [ + { + "bbox": [ + 198, + 158, + 269, + 170 + ], + "type": "text", + "content": "Ours w/ Depth=3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 421, + 158 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 421, + 158 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 421, + 158 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 421, + 158 + ], + "type": "image", + "image_path": "5d49ad78ca0f402fd9156691721451865fa442f6f13852be1813a0f1ae517f36.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 160, + 391, + 171 + ], + "lines": [ + { + "bbox": [ + 328, + 160, + 391, + 171 + ], + "spans": [ + { + "bbox": [ + 328, + 160, + 391, + 171 + ], + "type": "text", + "content": "Ours w/o TRM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 546, + 158 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 546, + 158 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 546, + 158 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 546, + 158 + ], + "type": "image", + "image_path": "cd754a23ac84ec75a9c029bc4c2474eb33d8b3dfd8d7fc03d4f06965a73de2dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 160, + 515, + 170 + ], + "lines": [ + { + "bbox": [ + 455, + 160, + 515, + 170 + ], + "spans": [ + { + "bbox": [ + 455, + 160, + 515, + 170 + ], + "type": "text", + "content": "Ours w/o DFE" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 174, + 170, + 260 + ], + "blocks": [ + { + "bbox": [ + 48, + 174, + 170, + 260 + ], + "lines": [ + { + "bbox": [ + 48, + 174, + 170, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 170, + 260 + ], + "type": "image", + "image_path": "697bd36004825364ebaf1d0bbe529a15062849b4ca8ff9e707d54cd83eec9b19.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 262, + 148, + 274 + ], + "lines": [ + { + "bbox": [ + 69, + 262, + 148, + 274 + ], + "spans": [ + { + "bbox": [ + 69, + 262, + 148, + 274 + ], + "type": "text", + "content": "Ours w/o Synthesis" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 173, + 175, + 295, + 262 + ], + "blocks": [ + { + "bbox": [ + 173, + 175, + 295, + 262 + ], + "lines": [ + { + "bbox": [ + 173, + 175, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 173, + 175, + 295, + 262 + ], + "type": "image", + "image_path": "846cfcdd1283cd3bf5d968fec9cf8c7f789139c23a13832db5edd4e858f24189.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 265, + 264, + 274 + ], + "lines": [ + { + "bbox": [ + 204, + 265, + 264, + 274 + ], + "spans": [ + { + "bbox": [ + 204, + 265, + 264, + 274 + ], + "type": "text", + "content": "Ours w/o DAT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 299, + 175, + 421, + 262 + ], + "blocks": [ + { + "bbox": [ + 299, + 175, + 421, + 262 + ], + "lines": [ + { + "bbox": [ + 299, + 175, + 421, + 262 + ], + "spans": [ + { + "bbox": [ + 299, + 175, + 421, + 262 + ], + "type": "image", + "image_path": "3b6b38b3df1b37d87a291cc928919d6bbdfe3c85f3dd3ffd75223e59c3059fe0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 265, + 370, + 274 + ], + "lines": [ + { + "bbox": [ + 349, + 265, + 370, + 274 + ], + "spans": [ + { + "bbox": [ + 349, + 265, + 370, + 274 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 174, + 546, + 260 + ], + "blocks": [ + { + "bbox": [ + 425, + 174, + 546, + 260 + ], + "lines": [ + { + "bbox": [ + 425, + 174, + 546, + 260 + ], + "spans": [ + { + "bbox": [ + 425, + 174, + 546, + 260 + ], + "type": "image", + "image_path": "70e3eb1319516a0897be357754bcef8fa5beab42bc78c4764d06a2ffabcea459.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 263, + 498, + 274 + ], + "lines": [ + { + "bbox": [ + 471, + 263, + 498, + 274 + ], + "spans": [ + { + "bbox": [ + 471, + 263, + 498, + 274 + ], + "type": "text", + "content": "Target" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 282, + 472, + 294 + ], + "lines": [ + { + "bbox": [ + 119, + 282, + 472, + 294 + ], + "spans": [ + { + "bbox": [ + 119, + 282, + 472, + 294 + ], + "type": "text", + "content": "Figure 7. Qualitative results of the ablation studies on different modules of the proposed FSENet." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 59, + 304, + 533, + 419 + ], + "blocks": [ + { + "bbox": [ + 59, + 304, + 533, + 419 + ], + "lines": [ + { + "bbox": [ + 59, + 304, + 533, + 419 + ], + "spans": [ + { + "bbox": [ + 59, + 304, + 533, + 419 + ], + "type": "table", + "html": "
Ablation StudyJung (Original)Kliger (Original)SD7K (Original)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
Input12.930.7961.3413.250.8156.7515.940.9144.14
Ours w/ Depth=320.970.8323.1926.330.9113.6620.830.9324.02
Ours w/o TRM19.500.8127.6027.100.9212.7818.440.8635.92
Ours w/o DFE21.900.8421.0225.160.9214.8919.700.8829.07
Ours w/o DAT22.680.8519.2822.700.8820.1919.030.8732.94
Ours w/o Synthesis22.210.8521.3324.790.9215.5422.940.9321.93
Ours Full23.600.8517.5628.980.9310.6028.670.9610.00
", + "image_path": "51fd5fe35ad792eb499cc1e72e16ab0581922887b076327d49b0402de38a0d6b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 423, + 523, + 434 + ], + "lines": [ + { + "bbox": [ + 47, + 423, + 523, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 523, + 434 + ], + "type": "text", + "content": "Table 3. Quantitative results of the ablation study on three datasets in high-resolution. The best result is highlighted in red and bold." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 454, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 514 + ], + "type": "text", + "content": "such as contours and edges, thus contributing significantly to the overall efficacy of our algorithm. Finally, we also evaluate the importance of the proposed shadow synthesis pipeline, where w/o shadow synthesis reduces the performance a lot, especially on the two small datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 521, + 124, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 124, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 124, + 533 + ], + "type": "text", + "content": "5.4. Limitations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 622 + ], + "type": "text", + "content": "Although our model is capable of removing shadows in high-resolution images, the current network is not able to run on real-time or on edge devices, such as mobile phones, due to the larger parameters and heavy GPU computations. In the future, we will explore different approaches to further reduce the required parameters, making it usable on smart devices." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we tackle the problem of removing shadows from high-resolution documents. To achieve this goal, we consider the contribution of the dataset and the network. As for the dataset, we gather a high-resolution dataset SD7K, which is comprised of over 7000 triplets of real-" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 454, + 546, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 546, + 634 + ], + "type": "text", + "content": "world document images with distinct characteristics under various illumination conditions. This dataset is an order of magnitude greater than previous datasets and allows us to validate our approach in a more diverse and comprehensive manner. As for the network structure, we propose FSENet, a frequency-aware method to remove shadows in higher resolution via frequency decomposition. To separate the high-frequency contours from the low-resolution content, we employ a Laplacian Pyramid and a series of meticulously designed attention blocks, enabling each component to be learned separately. Powered by our dataset and network, our method achieves a new state-of-the-art for document shadow removal. We believe both the network structure and the dataset will benefit the community for research on this topic." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by the University of Macau under Grant MYRG2022-00190-FST, in part by the Science and Technology Development Fund, Macau SAR, under Grant 0034/2019/AMJ, Grant 0087/2020/A2 and Grant 0049/2021/A." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12456" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Steve Bako, Soheil Darabi, Eli Shechtman, Jue Wang, Kalyan Sunkavalli, and Pradeep Sen. Removing shadows from images of documents. In ACCV, pages 173-183. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 169 + ], + "type": "text", + "content": "[2] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, pages 671-679. Elsevier, 1987. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 288, + 203 + ], + "type": "text", + "content": "[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, pages 17-33. Springer, 2022. 4, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 247 + ], + "type": "text", + "content": "[4] Xuhang Chen, Xiaodong Cun, Chi-Man Pun, and Shuqiang Wang. Shadocnet: Learning spatial-aware tokens in transformer for document shadow removal. In IEEE ICASSP, pages 1-5. IEEE, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "text", + "content": "[5] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[6] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In AAAI, pages 10680-10687, 2020. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 401 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[8] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In CVPR, pages 10571-10580, 2021. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 467 + ], + "type": "text", + "content": "[9] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 501 + ], + "type": "text", + "content": "[10] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps image shadow removal. In AAAI, 2023. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "text", + "content": "[11] John C Hart. Perlin noise pixel shaders. In Proceedings of the ACM SIGGRAPH/EUROGRAPHICS workshop on Graphics hardware, pages 87-94, 2001. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 614, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 635 + ], + "type": "text", + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[15] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[16] Seungjun Jung, Muhammad Abul Hasan, and Changick Kim. Water-filling: An efficient algorithm for digitized document shadow removal. In Asian Conference on Computer Vision, pages 398-414. Springer, 2018. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Netanel Kligler, Sagi Katz, and Ayellet Tal. Document enhancement using visibility detection. In CVPR, pages 2374-2382, 2018. 1, 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "text", + "content": "[18] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, August 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "type": "text", + "content": "[19] Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. In ECCV, pages 264-281. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 166, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 166, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 166, + 545, + 209 + ], + "type": "text", + "content": "[20] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 212, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 267 + ], + "type": "text", + "content": "[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 269, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 323 + ], + "type": "text", + "content": "[22] Jingtang Liang, Xiaodong Cun, Chi-Man Pun, and Jue Wang. Spatial-separated curve rendering network for efficient and high-resolution image harmonization. In European Conference on Computer Vision, pages 334-349. Springer, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "text", + "content": "[23] Jie Liang, Hui Zeng, and Lei Zhang. High-resolution photorealistic image translation in real-time: A laplacian pyramid translation network. In CVPR, pages 9392-9400, 2021. 2, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 371, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 414 + ], + "type": "text", + "content": "[24] Yun-Hsuan Lin, Wen-Chin Chen, and Yung-Yu Chuang. Bedsr-net: A deep shadow removal network from a single document image. In CVPR, pages 12905-12914, 2020. 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 416, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 460 + ], + "type": "text", + "content": "[25] Zhihao Liu, Hui Yin, Yang Mi, Mengyang Pu, and Song Wang. Shadow removal by a lightness-guided network with training on unpaired data. IEEE Transactions on Image Processing, 30:1853-1865, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 545, + 495 + ], + "type": "text", + "content": "[26] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In SIGGRAPH, pages 231–242, 1998. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "type": "text", + "content": "[27] Vatsal Shah and Vineet Gandhi. An iterative approach for shadow removal in document images. IEEE ICASSP, pages 1892-1896, 2018. 1, 3, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "text", + "content": "[28] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 588, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 622 + ], + "type": "text", + "content": "[29] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV, pages 361-378. Springer, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[30] Bingshu Wang and CL Philip Chen. An effective background estimation method for shadows removal of document images. In IEEE International Conference on Image Processing (ICIP), pages 3611-3615. IEEE, 2019. 1, 3, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12457" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 397 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[32] Jian-Ren Wang and Yung-Yu Chuang. Shadow removal of text document images by estimating local and global background colors. IEEE ICASSP, pages 1534-1538, 2020. 1, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[33] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. AAAI, 2023. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "text", + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 252 + ], + "type": "text", + "content": "[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17683-17693, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 286, + 296 + ], + "type": "text", + "content": "[36] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In SIGKDD, pages 1192-1200, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 287, + 341 + ], + "type": "text", + "content": "[37] Yimin Xu, Mingbao Lin, Hong Yang, Ke Li, Yunhang Shen, Fei Chao, and Rongrong Ji. Shadow-aware dynamic convolution for shadow removal. arXiv preprint arXiv:2205.04908, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 342, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 286, + 364 + ], + "type": "text", + "content": "[38] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. *ICLR*, 2015. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 286, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 397 + ], + "type": "text", + "content": "[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5627-5636, June 2022. 7" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12458" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_content_list.json b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6ce799818350032a0cbe65d7816bd8ffca587a4 --- /dev/null +++ b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_content_list.json @@ -0,0 +1,1514 @@ +[ + { + "type": "text", + "text": "Holistic Geometric Feature Learning for Structured Reconstruction", + "text_level": 1, + "bbox": [ + 143, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziqiong Lu* Linxi Huan* Qiyuan Ma Xianwei Zheng† \nThe State Key Lab. LIESMARS, Wuhan University \n{zql2018, whu.hlx, qiyuanma, zhengxw}@whu.edu.cn", + "bbox": [ + 209, + 179, + 756, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 282, + 313, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The inference of topological principles is a key problem in structured reconstruction. We observe that wrongly predicted topological relationships are often incurred by the lack of holistic geometry clues in low-level features. Inspired by the fact that massive signals can be compactly described with frequency analysis, we experimentally explore the efficiency and tendency of learning structure geometry in the frequency domain. Accordingly, we propose a frequency-domain feature learning strategy (F-Learn) to fuse scattered geometric fragments holistically for topology-intact structure reasoning. Benefiting from the parsimonious design, the F-Learn strategy can be easily deployed into a deep reconstructor with a lightweight model modification. Experiments demonstrate that the F-Learn strategy can effectively introduce structure awareness into geometric primitive detection and topology inference, bringing significant performance improvement to final structured reconstruction. Code and pre-trained models are available at https://github.com/Geo-Tell/F-Learn.", + "bbox": [ + 75, + 315, + 472, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The structured reconstruction models the shape grammar/topology that depicts procedural shape generation [24], which can facilitate various downstream vision tasks, such as feature matching [19, 44], 3D modeling [23, 32] and 3D scene understanding [47, 42]. The recovery of a shape topology is generally realized via two steps: (i) geometric primitive detection and (ii) graph inference based on extracted geometric primitives (e.g., corners and edges). Following this primitive-to-structure principle, previous researchers usually resort to heatmaps of corners and edges for subsequent structure reasoning [18, 47, 42, 2]. With low-level parsing tasks promoted by deep learning techniques, many works have greatly improved the performance of high-level structure recovery in recent years [42, 24, 2].", + "bbox": [ + 75, + 655, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Researchers commonly dedicate to graph inference with geometric features generated by proven hierarchical backbones (e.g., ResNet [10]) for structured reconstruction. However, incorrect topology inference remains a significant problem due to the lack of holistic geometric clues in the low-level features. Specifically, the geometric fragments extracted by shallow layers can hardly be fused into structurally informative features, leading to wrongly reasoned topological principles (as shown in Fig. 1 (a)). Different from the prior arts that focus on graph inference, this paper turns to study an efficient strategy for holistically learning structure-related features with low-level geometric fragments.", + "bbox": [ + 500, + 268, + 892, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-level geometric features fundamentally support high-level feature extraction and structure recovery. Specifically, the low-level shallow-layer feature maps provide information for precise geometry localization, which is essential for high-level primitive detection and topology inference. Restricted by the limited receptive field, it is hard to holistically capture structure geometries in the shallow layers. We consider that the low efficiency of processing low-level geometric fragments in the space domain is the root cause.", + "bbox": [ + 496, + 470, + 892, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the fact that low-level features can be compactly encoded in the frequency domain, we are interested in achieving efficient geometry feature learning with frequency analysis. In the frequency map of a given feature map, every value encodes the global information of the feature map regarding the corresponding frequency. Therefore, performing convolution operations in the frequency domain actually means directly combining the spatial information in a holistic manner.", + "bbox": [ + 496, + 625, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To be concrete, given a set of low-level feature maps $\\{f_n(x,y)\\}$ and the corresponding frequency map $\\{F_{n}(u,v)\\}$ , the information encoded at $(u_{i},v_{j})$ describes a certain changing pattern of signals in the space domain. When applying a convolution operation to a frequency position $(u_{i},v_{j})$ , the feature components of related frequencies in every map of $\\{f_n(x,y)\\}$ will be merged across the channel dimension in the space domain. As the low-level geometric information generally belongs to high-frequency sig", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding author.", + "bbox": [ + 99, + 875, + 227, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21807", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/121d539bc33c61a8cabbe7c46eba0dbcf3d40d5d1cab69d0e675aec630564733.jpg", + "image_caption": [ + "Figure 1. Comparison of learning low-level features in the space and frequency domains. (a) Structure reconstruction with low-level geometric features learned in the space domain. The inefficient fusion of low-level geometric fragments loses the structure clues of the bottom left roof region and consequently results in missing topological relations in the final reconstructed structure graph. (b) Structure reconstruction with low-level geometric features learned with the proposed frequency-domain feature learning strategy (F-Learn). With the geometric features compactly processed in the frequency domain, our F-Learn strategy effectively achieves holistic geometry fusion for inferring the right topological principle for structure reconstruction. $E_{p}$ and $E_{r}$ denote the scores of precision and recall for detected edges, while $R_{p}$ and $R_{r}$ refer to those of reconstructed regions." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 890, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nals, a frequency-domain convolution directly realizes the integration of geometric primitives scatted in $\\{f_n(x,y)\\}$ . The frequency-domain convolution thereby works more efficiently in holistic geometry learning than the spacedomain counterpart, which combines local geometric clues without a global view.", + "bbox": [ + 75, + 449, + 468, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the discussion above, we propose a frequency-domain feature learning strategy (F-Learn) to efficiently extract holistic geometry features for guiding the inference of structure topology. Working in the frequency domain, the F-Learn strategy holistically fuses the separated geometric primitives into structurally informative features. The proposed F-Learn strategy can be readily applied to a primitive-to-structure framework for structured reconstruction.", + "bbox": [ + 75, + 541, + 470, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this paper are:", + "bbox": [ + 96, + 662, + 447, + 679 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Exploration with a simple geometry recovery task sheds light on the difference in learning tendency between frequency- and space-domain convolutions. The results also validate the high efficiency of frequency-domain convolution in learning holistic geometry.", + "- A parsimonious frequency-domain feature learning strategy (F-Learn) is proposed to generate structurally informative geometry features for structured reconstruction.", + "- Experiments on vectorizing world buildings demonstrate that our F-Learn strategy greatly improves the performance of structured reconstruction." + ], + "bbox": [ + 94, + 691, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 446, + 640, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Geometric Primitive Detection", + "text_level": 1, + "bbox": [ + 500, + 481, + 769, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The detection of geometric primitives is a long-standing vision task that has been extensively explored with handcrafted descriptors in the early stage [26, 9]. Recently, with the learning ability of hierarchical features, deep learning models have significantly promoted the completeness and precision of detected geometric primitives. A popular way for low-level geometry extraction is to perform heatmap regression or pixel-wise binary classification. The heatmap regression technique is widely adopted in corner detection. For instance, methods like CornerNet [17] and CenterNet [6] learn corner heatmaps for the downstream task of object detection. Binary classification is often used to localize edge pixels. Representative edge detectors like HED [34] and RCF [21] focus on fusing multi-level features for edge pixel classification, while some works pay attention to learning crisp edges or semantic contours [40, 5, 11].", + "bbox": [ + 496, + 513, + 890, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to the pixel-wise inference of corners and edges, the detection of line primitives is more structural as a line segment should be defined by two endpoints [7, 16]. Since the task of wireframe parsing was introduced by [14], increasing interest has been witnessed in inferring line segment candidates with learned junctions [45, 38, 20]. With the advance in low- and mid-level primitive detection, researchers have started to make efforts on structured reconstruction in recent years.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21808", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Structured Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 90, + 318, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The structured reconstruction requires reasoning the overall topology of a given shape. Generally, structured reconstruction is a large research field that contains various tasks ranging from 3D object CAD modeling [32], layout estimation [18, 47], to roof extraction [43]. For instance-level structure reconstruction, many methods resort to generative models [32] or detecting key points with fixed topologies (e.g., skeletons of human bodies and hands) [33, 46].", + "bbox": [ + 75, + 113, + 468, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As for scene-level structural modeling, the reconstruction targets usually belong to special semantic-related structures, such as planar building roofs and indoor floorplans. For these tasks, the derived structures have to match semantic regions with robustness to other irrelevant structure information. For example, room layout estimation has to ensure that connected lines can form wall regions without extra lines from windows and doors. Related approaches can be roughly categorized into two kinds: (i) the local primitive based and (ii) the global information based. Pioneering local primitive based works derive structural representation from an image by post-processing heatmaps of corners and edges [18, 47].", + "bbox": [ + 75, + 250, + 468, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With respect to the global information based methods, it is a popular choice to form the final structure graph with structure-related regions segmented from the input image [41, 28]. Different from the segmentation based methods, recent studies consider a holistic graph inference with geometric primitives as nodes for high-level planer extraction. For example, Conv-MPN [42] adopts a variant of the graph neural network to pass messages across the whole graph, and HEAT [2] learns the topological pattern of edges with an attention transformer. Beyond exploring the high-level information in structure recovery, we are interested in improving the efficiency of exploiting low-level geometric features, which serve as the foundation of accurate structure topology inference.", + "bbox": [ + 75, + 446, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Frequency Analysis in Deep Learning", + "text_level": 1, + "bbox": [ + 76, + 665, + 403, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frequency analysis is a technique that provides a compressed representation of signals in the frequency domain. Not only commonly used in classical digital signal processing, but frequency analysis also remains powerful in the era of deep learning. [37] and [1] investigate the training and generalization of neural networks via frequency analysis. [35] analyzes the spectral bias of deep models regarding multiple vision parsing tasks. [8] applies frequency analysis to generated images for improving image synthesis quality, while [15] introduces a focal frequency loss that forces generative models to learn hard frequencies. [25] designs a frequency channel attention mechanism to compress channelwise information with scalars. [3] and [31] utilize frequency analysis in model compression. More applications of the", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frequency-domain representation can also be found in domain adaptation [12, 13] and position embedding [22, 29].", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although frequency analysis has been combined with deep learning techniques in recent years, it is generally used as a mature tool in prior arts without a deep investigation of the reasons why frequency analysis is useful. To this end, we experimentally studied the behavior tendency of frequency-domain convolutions in learning holistic geometries. Based on the investigation results, we designed the F-Learn strategy to efficiently learn holistic geometry clues in low-level feature maps for structured reconstruction.", + "bbox": [ + 496, + 121, + 893, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 268, + 589, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometric features extracted at the early convolutional stage fundamentally support the structure reconstruction with precise geometry localization clues. Therefore, we are interested in addressing a key problem for learning reliable low-level geometric features: the low efficiency of fusing geometric fragments in shallow convolution layers. Considering that geometric information is generally high-frequency signals, we are motivated to study the problem mentioned before with frequency analysis for structure reconstruction. In the following of this section, we first present the preliminaries of frequency analysis. Next, we validate the high efficiency of learning in the frequency domain for holistic geometry recovery. Finally, we introduce our frequency-domain feature learning strategy (F-Learn) and the application of the F-Learn in a given structure reconstruction model.", + "bbox": [ + 496, + 294, + 892, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries of Frequency Analysis", + "text_level": 1, + "bbox": [ + 500, + 544, + 815, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The discrete Fourier transform (DFT) is necessary for analyzing image data in the frequency domain. With $I(x,y)$ denoting the color signal at the spatial position $(x,y)$ of an $M\\times N$ image, the DFT converts the information of $I$ to a frequency map $F$ in the frequency domain by Eq. (1).", + "bbox": [ + 498, + 568, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF (u, v) = \\sum_ {x = 0} ^ {M - 1} \\sum_ {y = 0} ^ {N - 1} I (x, y) e ^ {- j 2 \\pi (u x / M + v y / N)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 651, + 890, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Eq. (1), $F(u, v)$ is a value that encodes a two-dimensional signal changing that belongs to the frequency determined by $u$ and $v$ . The frequency map excels in holistically representing signals that show a similar frequency pattern. Therefore, we assume that the DFT is a powerful tool for efficiently learning structure-related geometry features. To validate this hypothesis, we conducted experiments with a geometry recovery task, and analyze the results quantitatively and qualitatively as below.", + "bbox": [ + 496, + 702, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Learning Efficiency Analysis", + "text_level": 1, + "bbox": [ + 500, + 847, + 756, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we study the learning efficiency of fusing geometric fragments in the space and frequency domains", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21809", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with a simple geometry recovery task. We first decomposed a geometric structure composed of a circle and a square into several geometric fragments as illustrated in Fig. 2 (a), and left overlapping between neighboring fragments to avoid trial solutions.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the purpose to recover the binary image of the original structure with the geometric fragments, we studied the learning efficiency of frequency-domain and space-domain convolution models. We constructed a frequency-domain learning model (F-Learn) and several space-domain convolution models. The baseline was set as a space-domain convolution module (BConv) built with two $1 \\times 1$ convolution layers and one $3 \\times 3$ layer in between. The F-Learn model consists of three key components: (1) a DFT; (2) two parallel convolution modules same with the baseline for separately processing the real and imaginary parts; and (3) an inverse DFT (IDFT) followed by magnitude computation of complex values. For a fair comparison, we also combined a pair of baseline modules into two kinds of space-domain learning models with parallel and cascade structures, namely Conv-Casv1 and Conv-Parv1. Extra $3 \\times 3$ convolutions were also added into the cascade and parallel space-domain models to simulate the function of DFT and IDFT. The modified models were named Conv-Casv2 and Conv-Parv2 for simplicity. For all models, the channel number of the intermediate feature maps was 64, and a $1 \\times 1$ convolution was used as the final binary classifier. Detailed model settings are depicted in Fig. 2(b).", + "bbox": [ + 75, + 167, + 470, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The experiments were conducted on $128 \\times 128$ binary images of the geometric structure and fragments. The training of all models was completed with an Adam optimizer. The learning rate and training period were set to 0.1 and 50 epochs, respectively. The binary cross entropy loss (BCELoss) was chosen as the training guide. To alleviate the influence of randomness, we generated 100 seeds to set 100 trials, and all models were trained with the same seed in every trial. The performance of all models was evaluated with an F1-score measurement at a threshold of 0.5. Quantitative and qualitative results are shown in Fig. 3 and Fig. 4.", + "bbox": [ + 75, + 517, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 draws the averaged F1-score curves of 100 trials. Compared with the space-domain counterparts, the frequency-domain model (F-Learn) shows superiority in recovering the complete structure in a more efficient manner.", + "bbox": [ + 75, + 700, + 468, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visualized results in Fig. 4 also show that the frequency-domain learning model exceeds the basic spatial convolution model with faster convergence speed and better structure recovery quality.", + "bbox": [ + 75, + 762, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 4 reveals that the F-Learn model easily rebuilds the original structure, while the space-domain methods encounter difficulties in efficiently reconstructing the circle. We also find the difference in learning tendency between the F-Learn and space-domain models. The F-Learn model pri", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dd3c7f7e769a96ad8d12190a426b4d23b5dadc0d71229d0a00c72ff1a291c2b2.jpg", + "image_caption": [ + "(a) Geometry Recovery Task", + "(b) The Model Settings" + ], + "image_footnote": [], + "bbox": [ + 506, + 116, + 883, + 188 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e0c7bc7a7d8c30fbf478c5db138fb44e7b5ca7b9d8c61e6bb5ba0467091fb25.jpg", + "image_caption": [ + "Figure 2. The settings of the simple geometry recovery task. (a) The task of recovering the holistic structure with geometric fragments. (b) The detailed settings of compared models." + ], + "image_footnote": [], + "bbox": [ + 513, + 232, + 883, + 424 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eeabac41c684490d02950e54d9c8ca26e382dac3578a1b5d6ed917ca292fc84b.jpg", + "image_caption": [ + "Figure 3. The averaged F1-score curves of 100 trials for the F-Learn model and the compared space-domain methods." + ], + "image_footnote": [], + "bbox": [ + 504, + 484, + 885, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "oritizes holistic integration before further enlarging the gap between the foreground and background. In contrast, the space-domain models first generate high responses to local areas that have strong signal intensity in the input. Hence, the space-domain models prefer overlapped areas without considering structural continuity. In practice, the holistic learning pattern of F-Learn will benefit the topology inference in two ways: (i) holistic geometry clues are ready for learning high-level features at the early stage, which alleviates the burden on learning structure-related low-level features with far supervision signal; (ii) the holistic geometry learning excels in preserving the connection in-", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21810", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c8c685bca19215ac7a484ae97450d8ba69eea8563ed213f38cb9939ef79f28e.jpg", + "image_caption": [ + "Figure 4. The visualized results of F-Learn, Bconv and Cov-Parv1 at different epochs." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formation in the high-resolution low-level feature maps, which are often reused to provide accurate structure localization for topology inference.", + "bbox": [ + 75, + 345, + 467, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We attribute the advantages of the frequency domain based model to its ability to holistically process signals that belong to similar frequencies. Based on the findings above, we designed a frequency-domain feature learning strategy to generate holistic geometric clues in the low-level feature maps for structured reconstruction.", + "bbox": [ + 75, + 392, + 467, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Frequency-domain Learning Strategy", + "text_level": 1, + "bbox": [ + 76, + 491, + 401, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed frequency-domain learning strategy (Flearn) inherits the architecture shown in Fig. 2. As the FLearn strategy will face more complex data and larger batch size in practice than in Sec. 3.2, we replace each convolution layer between DFT and IDFT with a combination (CBR) of a convolution layer, a batch normalization operation, and a relu activation function. We refer to all operations between DFT and IDFT collectively as a CBR group (CBRG). Given a set of geometric feature maps $\\{f_i|i = 0,\\dots ,N\\}$ , the FLearn strategy is formulated as", + "bbox": [ + 75, + 515, + 467, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(F _ {i} ^ {\\text {r e}}, F _ {i} ^ {\\text {i m}}\\right) = \\operatorname {D F T} \\left(f _ {i}\\right);\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 678, + 344, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\hat {F} _ {i} ^ {\\mathrm {r e}} \\}, \\{\\hat {F} _ {i} ^ {\\mathrm {i m}} \\} = \\mathrm {C B R G} (\\{F _ {i} ^ {\\mathrm {r e}} \\}), \\mathrm {C B R G} (\\{F _ {i} ^ {\\mathrm {i m}} \\}); \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 698, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\hat {f} _ {i} \\} = | \\mathrm {I D F T} (\\{\\hat {F} _ {i} ^ {\\mathrm {r e}} \\}, \\{\\hat {F} _ {i} ^ {\\mathrm {i m}} \\}) |.\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 720, + 388, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The F-learn strategy first converts every $f_{i}$ into a pair of maps that record the real and imaginary parts of the DFT, i.e., $F_{i}^{\\mathrm{re}}$ and $F_{i}^{\\mathrm{im}}$ . The maps of the real part $\\{F_{i}^{\\mathrm{re}}\\}$ are subsequently processed by a CBRG that contains two $1 \\times 1$ layers with a $3 \\times 3$ one in between, and the same operation is also applied to $\\{F_{i}^{\\mathrm{im}}\\}$ . In the frequency domain, the $1 \\times 1$ convolution directly fuses the information of the same frequency across the channel dimension, while the $3 \\times 3$ counterpart enhances features with signals that have similar frequencies. The high-frequency geometric fragments", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in $\\{f_i\\}$ are easily combined and enhanced in such a holistic way. The structurally informative space-domain features $\\{\\hat{f}_i\\}$ can be finally obtained through an IDFT along with magnitude computation $|\\cdot|$ . With the parsimonious design, the F-learn strategy is readily inserted into a hierarchical convolutional backbone for holistic geometry feature learning.", + "bbox": [ + 498, + 90, + 890, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For feature propagation, the $\\{f_i\\}$ and $\\{\\hat{f}_i\\}$ are fused into $\\{\\tilde{f}_i\\}$ via", + "bbox": [ + 500, + 196, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\tilde {f} _ {i} \\right\\} = \\mathrm {C} _ {1 \\times 1} \\mathrm {B R} \\left(\\operatorname {C o n c a t} \\left(\\left\\{f _ {i} \\right\\}, \\left\\{\\hat {f} _ {i} \\right\\}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 234, + 890, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Concat}$ and $\\mathbf{C}_{1\\times 1}\\mathbf{BR}$ denote the concatenation operation and a $1\\times 1$ convolution based CBR, respectively.", + "bbox": [ + 498, + 262, + 890, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The enhanced geometric features $\\{\\tilde{f}_i\\}$ are then fed into the next convolution stage for higher-level feature learning and used to offer precise geometric localization for later structure inference.", + "bbox": [ + 498, + 292, + 890, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The F-Learn strategy can be easily inserted into a convolution base backbone. With the commonly-used ResNet50 backbone [10] as an example, the F-Learn strategy can be directly deployed after the first convolution to holistically learn geometric features. In experiments, we implemented our F-Learn strategy in a state-of-the-art approach, i.e., the holistic edge attention transformer (HEAT) [2], for structured reconstruction.", + "bbox": [ + 498, + 352, + 890, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 484, + 630, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 500, + 510, + 691, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Dataset", + "text_level": 1, + "bbox": [ + 500, + 534, + 606, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We tested our method on a dataset introduced by [24] for vectorizing world buildings. This dataset is built on the SpaceNet dataset [30] with 2001 images annotated with roof planar graphs. Each image contains a building instance, and the image is processed into a size of $256 \\times 256$ . In line with the prior arts [2], the training/validation/testing split is set to 1601/50/350.", + "bbox": [ + 498, + 556, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 681, + 717, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluated the F-learn strategy with the HEAT model for structured reconstruction. The HEAT model was built on a ResNet50 backbone pretrained on ImageNet [4]. We performed our experiments with Pytorch [27] in Python3.7 and used a workstation with one NVIDIA RTX 3090 GPU. We adopted the same settings of loss functions as the HEAT model. The training of the model was completed with an AdamW optimizer, and the training period lasts 800 epochs. The learning rate was initialized as $2e - 4$ and multiplied by a factor of 0.1 for the last $25\\%$ epoch. In line with previous works [2, 24], we used precision, recall, and F1 score to evaluate the quality of structure reconstruction in terms of corner detection, edge inference, and region recovery.", + "bbox": [ + 498, + 703, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21811", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4448dbec1fa095cef36ab1002c0f6eca899b2315d7d148296f8136a9d078ae75.jpg", + "table_caption": [ + "Table 1. Quantitative comparison between the F-Learn strategy and other methods in terms of corners detection, edge extraction, and region reconstruction. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + ], + "table_footnote": [], + "table_body": "
Evaluation Type → \nMethodFully-neuralJointCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
IP [24]----74.5--53.1--55.7
Exp-Cls [43]--92.275.983.275.460.467.174.954.763.5
ConvMPN [42]✓-78.079.778.857.059.758.152.456.554.4
HAWP [39]✓✓90.981.285.776.668.172.174.155.463.4
LETR [36]✓✓87.874.880.859.758.659.168.348.756.8
HEAT [2]✓✓91.783.087.180.672.376.276.465.670.6
HEAT(retrain) [2]✓✓91.683.087.180.372.476.175.565.370.0
F-Learn (Ours)✓✓93.284.488.683.675.079.179.568.173.4
", + "bbox": [ + 109, + 133, + 854, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3 Competing Methods", + "text_level": 1, + "bbox": [ + 76, + 339, + 274, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the F-Learn strategy with six methods: HEAT [2], ConvMPN [42], IP [24], Exp-cls [43], HAWP [39] and LETR [36].", + "bbox": [ + 75, + 364, + 468, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HEAT is an attention-based method that takes a 2D raster image as an input and reconstructs a planar graph in an end-to-end manner. HEAT works via three steps: 1) extracting hierarchical features by ResNet50; 2) detecting corners with multi-scale features enriched by a deformable attention module; and 3) employing two weight-sharing Transformer decoders to classify edges and reason structures with detected corners. Our F-Learn strategy is adopted in the first step for learning structure-related features.", + "bbox": [ + 75, + 411, + 470, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conv-MPN uses a graph neural network to infer edges with a pre-trained corner detector. IP firstly detects geometric primitives and then reconstructs a planar graph through several post-processing steps. Exp-cls is based on the geometric primitives produced by other methods (e.g., IP and Conv-MPN) and reconstructs a planar graph through an explore-and-classify framework. LCNN and HAWP are methods specifically proposed for wireframe parsing. LETR is a transformer-based method that directly generates lines without post-processing and heuristic guidance.", + "bbox": [ + 75, + 547, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison and Analysis", + "text_level": 1, + "bbox": [ + 76, + 712, + 312, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Quantitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 737, + 282, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative results displayed in Tab. 1 show that our F-Learn strategy advances roof structure reconstruction with state-of-the-art performance. The F-Learn strategy edges out other compared methods by significant performance gains regarding corner detection, edge inference, and region reconstruction under all metrics.", + "bbox": [ + 75, + 763, + 468, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With respect to the detection of corners and edges, the F-Learn strategy outperforms the second-best by at least $1.0\\%$ and $3.0\\%$ in the precision measurement, while $1.4\\%$ and", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$2.6\\%$ under the recall metric. The improvement in extracting low-level geometric primitives indicates the outstanding ability of the F-Learn strategy for providing holistic and accurate geometric clues. The F-Learn strategy also brings at least a $2.8\\%$ gain in F1 score in terms of region reconstruction, which reveals the assistance of the F-Learn strategy in inferring correct topological relations. Besides, it is worth noticing that the F-Learn strategy significantly outperforms the original HEAT model under all metrics, but only brings a little computation increase compared to HEAT. Tab. 2 presents a detailed comparison of computation requirements.", + "bbox": [ + 496, + 338, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/711ab14b8c40dcbc93953df1575009e9d0f95288e49d78a5e59831ad4cf09a0b.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of the change of computing efficiency brought by F-Learn. Param. and FLOPs are the abbreviations of the parameter quantity and floating point operations, used to measure the complexity of an algorithm/model. Time per Image represents the time it takes the model to infer an image and is used to evaluate the calculation speed of the model." + ], + "table_footnote": [], + "table_body": "
Param.FLOPsTime per Image
w/o F-Learn68.747M670.919G0.10s
w/ F-Learn68.847M684.039G0.11s
", + "bbox": [ + 501, + 625, + 880, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, the computation increase of F-Learn strategy in terms of parameters, FLOPs, and inference time per image can be clearly seen, where the amount is little. Especially, the parameter increase brought by F-Learn strategy is only $0.145\\%$ , which is almost negligible. The results further demonstrates that the F-Learn strategy is simple yet effective.", + "bbox": [ + 496, + 700, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 500, + 829, + 694, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For perceptual comparison, we also present the visualized reconstruction results in Fig. 5. Compared to other methods, our F-Learn is capable of detecting low-level geometric", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21812", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/079dc50752f9b4a8e44be9b0762e6f2bf5577cb93537c68ca7246206f1013004.jpg", + "image_caption": [ + "Figure 5. Qualitative results on outdoor roof structure reconstruction. Other methods build wrong topological relations due to redundant or missing geometric primitives. Our F-Learn strategy derives the correct topological principles with corners and edges more fitted to the holistic structures." + ], + "image_footnote": [], + "bbox": [ + 80, + 97, + 883, + 829 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "primitives that are consistent with the holistic structure.", + "bbox": [ + 76, + 90, + 442, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When reconstructing the roof shown in the first row, ConvMPN and Exp-cls tend to detect redundant corners due to the shadow inference and therefore wrongly separate the roof regions. The original HEAT outputs a non-closed roof structure because key edges are missing. In contrast, the corners and edges learned with the F-Learn strategy are better fitted into the overall structure. This is because the F-Learn strategy can offer low-level features that are rich in holistic geometric clues for topology-preserved roof structure reconstruction. The results in the following rows additionally validate that the F-Learn strategy can effectively extract geometric primitives that are easily ignored by other methods, especially the original HEAT model. By comparing the inferred planar graphs in the fifth row, one can also find that the F-Learn improves the HEAT model with robustness to irrelevant structure information.", + "bbox": [ + 75, + 107, + 470, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 359, + 241, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1 Comparison with Space-domain Methods", + "text_level": 1, + "bbox": [ + 76, + 383, + 419, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to the exploration with a simple geometry fusion task in Sec. 3.2, we further study the F-Learn strategy with the space-domain learning strategies in the real scene for roof structure reconstruction. Similar to Sec. 3.2, we construct a space-domain learning baseline (BConv) with a basic module composed of two $1 \\times 1$ convolution layers and one $3 \\times 3$ layer in between. Furthermore, we design two more space-domain strategies with the basic modules arranged in parallel and cascade, denoted as Conv-Parv. and Conv-Casv., respectively. Same with the F-Learn, additional operations of batch normalization and ReLU activation are added after each convolution layer used in every space-domain strategy.", + "bbox": [ + 75, + 409, + 468, + 604 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/469e7380f27f2bd7fc9eda8adbf5971846bf45ec19effb16677996ac6e830573.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons with space-domain learning methods. The F-Learn strategy is compared with the space-domain counterparts under the F1-score measurement in terms of corner detection, edge inference, and region reconstruction. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + ], + "table_footnote": [], + "table_body": "
MethodCornerEdgeRegion
HEAT87.176.170.0
HEAT-BConv87.476.270.7
HEAT-Conv-Cas.88.177.170.2
HEAT-Conv-Par.87.776.869.5
F-Learn88.679.173.4
", + "bbox": [ + 106, + 709, + 434, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the F1-score measurement, Tab. 3 presents the performance of the F-learn and space-domain learning strategies in detecting geometric primitives. Although some space-domain methods are comparable to the proposed F", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learn strategy in corner extraction, it can be seen that the F-Learn strategy gains over these methods in detecting edges and regions by at least $2\\%$ and $2.7\\%$ , respectively. Because edges and regions are primitives closely related to roof topology, this phenomenon demonstrates the holistic geometry learned by the F-Learn is beneficial for the topology inference for roof reconstruction.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 F-Learn Strategy at Different Layers", + "text_level": 1, + "bbox": [ + 498, + 224, + 812, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the effect of applying the F-Learn strategy to feature maps of different levels with the HEAT model and the hierarchical ResNet50 backbone. The F-learn strategy is deployed to process feature maps generated at the first convolution layer, the first residual learning stage, and the second one. The sizes of the corresponding feature maps are $128 \\times 128$ , $64 \\times 64$ , and $32 \\times 32$ , respectively. The quantitative results are presented in Tab. 4.", + "bbox": [ + 496, + 250, + 890, + 372 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a51f31154f28b9f9a0d520ed00072512d331846508cd8ed71b2fa47af85abbc.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of applying the F-Learn strategy to features at different levels. L0: the first convolution layer; L1: the first residual learning stage; L2: the second residual learning stage. Size: the feature map size. The F1-score metric is used for evaluating the performance of corner detection, edge inference, and region recovery. Higher scores mean better results. The bold values represent the best performance. (unit:%)" + ], + "table_footnote": [], + "table_body": "
MethodSizeCornerEdgeRegion
HEAT-87.176.170.0
F-Learn-L0128 × 12888.679.173.4
F-Learn-L164 × 6487.776.470.9
F-Learn-L232 × 3286.775.369.0
", + "bbox": [ + 504, + 492, + 879, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 implies that the higher the resolution is, the better the F-Learn strategy works. It is because high-resolution low-level feature maps contain more high-frequency geometric clues, while the low-resolution high-level ones are richer in abstract semantic information but poorer in geometric details. Therefore, the high-resolution feature maps are more suitable than the low-resolution counterparts for the F-Learn strategy to learn structurally informative geometries in roof reconstruction.", + "bbox": [ + 496, + 604, + 890, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3 F-Learn Strategy with Various Backbones", + "text_level": 1, + "bbox": [ + 498, + 768, + 846, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of the F-learn strategy, we evaluate the performance improvement brought by the F-Learn strategy with different backbones. The quantitative results are presented in Tab. 5. As shown by Tab. 5, with the deployment of the F-learn strategy, the performance of HEAT is consistently improved in terms of corners detection, edge extraction, and region reconstruction.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21814", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/003692ed79ba1062e10fcc25d1de26f89bd631d70f17203b7692f9f24c870666.jpg", + "table_caption": [ + "Table 5. Quantitative comparison of F-Learn strategy with HEAT built on different backbones, including ResNet18, ResNet34, and ResNet50. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + ], + "table_footnote": [], + "table_body": "
BackboneCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
w/o F-LearnResNet1891.681.986.579.970.174.776.361.668.2
w/ F-LearnResNet1891.782.286.780.870.475.277.162.969.3
w/o F-LearnResNet3490.582.686.478.371.174.573.463.468.0
w/ F-LearnResNet3492.183.287.480.271.875.876.264.169.6
w/o F-LearnResNet5091.683.087.180.372.476.175.565.370.0
w/ F-LearnResNet5093.284.488.683.675.079.179.568.173.4
", + "bbox": [ + 166, + 136, + 795, + 294 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3.4 F-Learn Strategy in Topology Inference", + "text_level": 1, + "bbox": [ + 76, + 318, + 408, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we further explore the effectiveness of the F-Learn strategy in topology inference. Edges and regions are primitives that highly relate to the inference of topological relationships, and these primitives are learned with corners and low-level features in the HEAT model. Therefore, we directly use the ground-truth corner map to focus on topology inference. We compare our F-Learn strategy with the space-domain methods used in Sec. 4.3.1.", + "bbox": [ + 75, + 340, + 470, + 464 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/fdbac89246cc86d99e6b63f4e6a23208fe05702dba899ff502d583d2b601042f.jpg", + "table_caption": [ + "Table 6. Quantitative results of the F-Learn and space-domain strategies with respect to topology inference. The HEAT method with corner ground truth is set as the baseline. With corner annotations, all methods are studied with a focus on topology inference. The higher scores mean better results, and the best performance is bolded. (unit: $\\%$ )" + ], + "table_footnote": [], + "table_body": "
Evaluation Type → MethodEdgeRegion
PrecRecallF1PrecRecallF1
HEAT93.584.188.590.273.881.2
HEAT-BConv94.483.988.889.471.779.6
HEAT-Conv-Cas94.885.389.891.373.781.6
HEAT-Conv-Par93.884.889.189.773.981.0
F-Learn95.187.591.189.577.683.1
", + "bbox": [ + 81, + 563, + 467, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 6 presents the numeric results, and it can be seen that our F-Learn strategy brings the highest gains in most metrics for edge inference and region reconstruction. This phenomenon demonstrates that the F-Learn strategy effectively supports topology inference with an outstanding ability for holistic geometry learning.", + "bbox": [ + 75, + 696, + 468, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 800, + 204, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we present a frequency-domain feature learning strategy (F-Learn) to tackle the issue of wrong topology recovery caused by the lack of holistic clues in low-level features. Experiments with a geometry recovery task convincingly validate the efficiency of the F-Learn", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "strategy in learning holistic geometry. In terms of the real scene, the F-Learn strategy achieves significant performance improvement of topological principle inference for roof structure reconstruction. The ablation studies verify that the F-Learn strategy outperforms the space-domain learning counterparts in capturing holistic geometric features regarding complex real scenes. We believe that it is promising to further explore the holistic learning ability brought by frequency analysis in more vision tasks.", + "bbox": [ + 496, + 318, + 893, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 467, + 679, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by NSFC projects under Grant 42071370, the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084.", + "bbox": [ + 496, + 492, + 893, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 595, + 596, + 609 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ronen Basri, Meirav Galun, Amnon Geifman, David Jacobs, Yoni Kasten, and Shira Kritchman. Frequency bias in neural networks for input of non-uniform density. In International Conference on Machine Learning, pages 685-694. PMLR, 2020. 3", + "[2] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In Proceedings of CVPR, pages 3866-3875, 2022. 1, 3, 5, 6", + "[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Q Weinberger, and Yixin Chen. Compressing convolutional neural networks in the frequency domain. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1475-1484, 2016. 3", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5", + "[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In ECCV, pages 570-586, 2018. 2" + ], + "bbox": [ + 506, + 619, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21815", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[6] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 2", + "[7] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. Communications of the ACM, 15(1):11-15, 1972. 2", + "[8] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3", + "[9] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the CVPR, pages 770-778, 2016. 1, 5", + "[11] Linxi Huan, Nan Xue, Xianwei Zheng, Wei He, Jianya Gong, and Gui-Song Xia. Unmixing convolutional features for crisp edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6602-6609, 2021. 2", + "[12] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 3", + "[13] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Rda: Robust domain adaptation via fourier adversarial attacking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8988-8999, 2021. 3", + "[14] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 2", + "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3", + "[16] Varsha Kamat-Sadekar and Subramaniam Ganesan. Complete description of multiple line segments using the hough transform. Image and Vision Computing, 16(9-10):597-613, 1998. 2", + "[17] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2", + "[18] Chen-Yu Lee, Vijay Badrinarayanan, Tomasz Malisiewicz, and Andrew Rabinovich. Roomnet: End-to-end room layout estimation. In Proceedings of the ICCV, pages 4865-4874, 2017. 1, 3", + "[19] Hongjie Li, Xianwei Zheng, Mingyue Dong, Gui-Song Xia, and Hanjiang Xiong. Locally nonlinear affine verification for multisensor image matching. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 1", + "[20] Yancong Lin, Silvia L Pintea, and Jan C van Gemert. Deep though-transform line priors. In Computer Vision-ECCV" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 323-340. Springer, 2020. 2", + "[21] Yun Liu, Ming-Ming Cheng, Xiaowei Hu, Jia-Wang Bian, Le Zhang, Xiang Bai, and Jinhui Tang. Richer convolutional features for edge detection. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1939-1946, 2019. 2", + "[22] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3", + "[23] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229. PMLR, 2020. 1", + "[24] Nelson Nauata and Yasutaka Furukawa. Vectorizing world buildings: Planar graph reconstruction by primitive detection and relationship inference. In ECCV 2020, pages 711-726. Springer, 2020. 1, 5, 6", + "[25] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3", + "[26] Lawrence G Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2", + "[27] Benoit Steiner, Zachary Devito, Soumith Chintala, Sam Gross, Adam Paszke, Francisco Massa, Adam Lerer, Gregory Chanan, Zeming Lin, Edward Yang, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5", + "[28] Sinisa Stekovic, Mahdi Rad, Friedrich Fraundorfer, and Vincent Lepetit. Montefloor: Extending mcts for reconstructing accurate large-scale floor plans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16034-16043, 2021. 3", + "[29] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3", + "[30] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 5", + "[31] Yunhe Wang, Chang Xu, Chao Xu, and Dacheng Tao. Packing convolutional neural networks in the frequency domain. IEEE transactions on pattern analysis and machine intelligence, 41(10):2495-2510, 2018. 3", + "[32] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the ICCV, pages 6772–6782, 2021. 1, 3", + "[33] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the ECCV, pages 466-481, 2018. 3", + "[34] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. International Journal of Computer Vision, 125(1):3-18, 2017. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21816", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1740-1749, 2020. 3", + "[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 6", + "[37] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3", + "[38] Nan Xue, Song Bai, Fudong Wang, Gui-Song Xia, Tianfu Wu, and Liangpei Zhang. Learning attraction field representation for robust line segment detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1595-1603, 2019. 2", + "[39] Nan Xue, Tianfu Wu, Song Bai, Fudong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2788-2797, 2020. 6", + "[40] Jimei Yang, Brian Price, Scott Cohen, Honglak Lee, and Ming-Hsuan Yang. Object contour detection with a fully convolutional encoder-decoder network. In CVPR, pages 193-202, 2016. 2", + "[41] Zhiliang Zeng, Xianzhi Li, Ying Kin Yu, and Chi-Wing Fu. Deep floor plan recognition using a multi-task network with room-boundary-guided attention. In Proceedings of the ICCV, pages 9096-9104, 2019. 3", + "[42] Fuyang Zhang, Nelson Nauata, and Yasutaka Furukawa. Conv-mpn: Convolutional message passing neural network for structured outdoor architecture reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2798-2807, 2020. 1, 3, 6", + "[43] Fuyang Zhang, Xiang Xu, Nelson Nauata, and Yasutaka Furukawa. Structured outdoor architecture reconstruction by exploration and classification. In Proceedings of ICCV, pages 12427-12435, 2021. 3, 6", + "[44] Xianwei Zheng, Zhuang Yuan, Zhen Dong, Mingyue Dong, Jianya Gong, and Hanjiang Xiong. Smoothly varying projective transformation for line segment matching. ISPRS Journal of Photogrammetry and Remote Sensing, 183:129-146, 2022. 1", + "[45] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 962-971, 2019. 2", + "[46] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the ICCV, pages 4903-4911, 2017. 3", + "[47] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the CVPR, pages 2051-2059, 2018. 1, 3" + ], + "bbox": [ + 78, + 90, + 468, + 853 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21817", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_model.json b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b596e083c3122354d48616a4e2d70129d39ecd7b --- /dev/null +++ b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_model.json @@ -0,0 +1,2092 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.144, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "Holistic Geometric Feature Learning for Structured Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.18, + 0.758, + 0.234 + ], + "angle": 0, + "content": "Ziqiong Lu* Linxi Huan* Qiyuan Ma Xianwei Zheng† \nThe State Key Lab. LIESMARS, Wuhan University \n{zql2018, whu.hlx, qiyuanma, zhengxw}@whu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.284, + 0.314, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.473, + 0.604 + ], + "angle": 0, + "content": "The inference of topological principles is a key problem in structured reconstruction. We observe that wrongly predicted topological relationships are often incurred by the lack of holistic geometry clues in low-level features. Inspired by the fact that massive signals can be compactly described with frequency analysis, we experimentally explore the efficiency and tendency of learning structure geometry in the frequency domain. Accordingly, we propose a frequency-domain feature learning strategy (F-Learn) to fuse scattered geometric fragments holistically for topology-intact structure reasoning. Benefiting from the parsimonious design, the F-Learn strategy can be easily deployed into a deep reconstructor with a lightweight model modification. Experiments demonstrate that the F-Learn strategy can effectively introduce structure awareness into geometric primitive detection and topology inference, bringing significant performance improvement to final structured reconstruction. Code and pre-trained models are available at https://github.com/Geo-Tell/F-Learn." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.21, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.868 + ], + "angle": 0, + "content": "The structured reconstruction models the shape grammar/topology that depicts procedural shape generation [24], which can facilitate various downstream vision tasks, such as feature matching [19, 44], 3D modeling [23, 32] and 3D scene understanding [47, 42]. The recovery of a shape topology is generally realized via two steps: (i) geometric primitive detection and (ii) graph inference based on extracted geometric primitives (e.g., corners and edges). Following this primitive-to-structure principle, previous researchers usually resort to heatmaps of corners and edges for subsequent structure reasoning [18, 47, 42, 2]. With low-level parsing tasks promoted by deep learning techniques, many works have greatly improved the performance of high-level structure recovery in recent years [42, 24, 2]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.893, + 0.466 + ], + "angle": 0, + "content": "Researchers commonly dedicate to graph inference with geometric features generated by proven hierarchical backbones (e.g., ResNet [10]) for structured reconstruction. However, incorrect topology inference remains a significant problem due to the lack of holistic geometric clues in the low-level features. Specifically, the geometric fragments extracted by shallow layers can hardly be fused into structurally informative features, leading to wrongly reasoned topological principles (as shown in Fig. 1 (a)). Different from the prior arts that focus on graph inference, this paper turns to study an efficient strategy for holistically learning structure-related features with low-level geometric fragments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.621 + ], + "angle": 0, + "content": "Low-level geometric features fundamentally support high-level feature extraction and structure recovery. Specifically, the low-level shallow-layer feature maps provide information for precise geometry localization, which is essential for high-level primitive detection and topology inference. Restricted by the limited receptive field, it is hard to holistically capture structure geometries in the shallow layers. We consider that the low efficiency of processing low-level geometric fragments in the space domain is the root cause." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Inspired by the fact that low-level features can be compactly encoded in the frequency domain, we are interested in achieving efficient geometry feature learning with frequency analysis. In the frequency map of a given feature map, every value encodes the global information of the feature map regarding the corresponding frequency. Therefore, performing convolution operations in the frequency domain actually means directly combining the spatial information in a holistic manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To be concrete, given a set of low-level feature maps \\(\\{f_n(x,y)\\}\\) and the corresponding frequency map \\(\\{F_{n}(u,v)\\}\\), the information encoded at \\((u_{i},v_{j})\\) describes a certain changing pattern of signals in the space domain. When applying a convolution operation to a frequency position \\((u_{i},v_{j})\\), the feature components of related frequencies in every map of \\(\\{f_n(x,y)\\}\\) will be merged across the channel dimension in the space domain. As the low-level geometric information generally belongs to high-frequency sig" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.875, + 0.228, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.891, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.895, + 0.425 + ], + "angle": 0, + "content": "Figure 1. Comparison of learning low-level features in the space and frequency domains. (a) Structure reconstruction with low-level geometric features learned in the space domain. The inefficient fusion of low-level geometric fragments loses the structure clues of the bottom left roof region and consequently results in missing topological relations in the final reconstructed structure graph. (b) Structure reconstruction with low-level geometric features learned with the proposed frequency-domain feature learning strategy (F-Learn). With the geometric features compactly processed in the frequency domain, our F-Learn strategy effectively achieves holistic geometry fusion for inferring the right topological principle for structure reconstruction. \\( E_{p} \\) and \\( E_{r} \\) denote the scores of precision and recall for detected edges, while \\( R_{p} \\) and \\( R_{r} \\) refer to those of reconstructed regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.47, + 0.54 + ], + "angle": 0, + "content": "nals, a frequency-domain convolution directly realizes the integration of geometric primitives scatted in \\(\\{f_n(x,y)\\}\\). The frequency-domain convolution thereby works more efficiently in holistic geometry learning than the spacedomain counterpart, which combines local geometric clues without a global view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.471, + 0.662 + ], + "angle": 0, + "content": "Based on the discussion above, we propose a frequency-domain feature learning strategy (F-Learn) to efficiently extract holistic geometry features for guiding the inference of structure topology. Working in the frequency domain, the F-Learn strategy holistically fuses the separated geometric primitives into structurally informative features. The proposed F-Learn strategy can be readily applied to a primitive-to-structure framework for structured reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.664, + 0.449, + 0.68 + ], + "angle": 0, + "content": "In summary, the main contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.692, + 0.47, + 0.768 + ], + "angle": 0, + "content": "- Exploration with a simple geometry recovery task sheds light on the difference in learning tendency between frequency- and space-domain convolutions. The results also validate the high efficiency of frequency-domain convolution in learning holistic geometry." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.782, + 0.471, + 0.841 + ], + "angle": 0, + "content": "- A parsimonious frequency-domain feature learning strategy (F-Learn) is proposed to generate structurally informative geometry features for structured reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "- Experiments on vectorizing world buildings demonstrate that our F-Learn strategy greatly improves the performance of structured reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.692, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.448, + 0.642, + 0.464 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.771, + 0.497 + ], + "angle": 0, + "content": "2.1. Geometric Primitive Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.756 + ], + "angle": 0, + "content": "The detection of geometric primitives is a long-standing vision task that has been extensively explored with handcrafted descriptors in the early stage [26, 9]. Recently, with the learning ability of hierarchical features, deep learning models have significantly promoted the completeness and precision of detected geometric primitives. A popular way for low-level geometry extraction is to perform heatmap regression or pixel-wise binary classification. The heatmap regression technique is widely adopted in corner detection. For instance, methods like CornerNet [17] and CenterNet [6] learn corner heatmaps for the downstream task of object detection. Binary classification is often used to localize edge pixels. Representative edge detectors like HED [34] and RCF [21] focus on fusing multi-level features for edge pixel classification, while some works pay attention to learning crisp edges or semantic contours [40, 5, 11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Compared to the pixel-wise inference of corners and edges, the detection of line primitives is more structural as a line segment should be defined by two endpoints [7, 16]. Since the task of wireframe parsing was introduced by [14], increasing interest has been witnessed in inferring line segment candidates with learned junctions [45, 38, 20]. With the advance in low- and mid-level primitive detection, researchers have started to make efforts on structured reconstruction in recent years." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21808" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.319, + 0.106 + ], + "angle": 0, + "content": "2.2. Structured Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.249 + ], + "angle": 0, + "content": "The structured reconstruction requires reasoning the overall topology of a given shape. Generally, structured reconstruction is a large research field that contains various tasks ranging from 3D object CAD modeling [32], layout estimation [18, 47], to roof extraction [43]. For instance-level structure reconstruction, many methods resort to generative models [32] or detecting key points with fixed topologies (e.g., skeletons of human bodies and hands) [33, 46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.469, + 0.446 + ], + "angle": 0, + "content": "As for scene-level structural modeling, the reconstruction targets usually belong to special semantic-related structures, such as planar building roofs and indoor floorplans. For these tasks, the derived structures have to match semantic regions with robustness to other irrelevant structure information. For example, room layout estimation has to ensure that connected lines can form wall regions without extra lines from windows and doors. Related approaches can be roughly categorized into two kinds: (i) the local primitive based and (ii) the global information based. Pioneering local primitive based works derive structural representation from an image by post-processing heatmaps of corners and edges [18, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.658 + ], + "angle": 0, + "content": "With respect to the global information based methods, it is a popular choice to form the final structure graph with structure-related regions segmented from the input image [41, 28]. Different from the segmentation based methods, recent studies consider a holistic graph inference with geometric primitives as nodes for high-level planer extraction. For example, Conv-MPN [42] adopts a variant of the graph neural network to pass messages across the whole graph, and HEAT [2] learns the topological pattern of edges with an attention transformer. Beyond exploring the high-level information in structure recovery, we are interested in improving the efficiency of exploiting low-level geometric features, which serve as the foundation of accurate structure topology inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.405, + 0.684 + ], + "angle": 0, + "content": "2.3. Frequency Analysis in Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Frequency analysis is a technique that provides a compressed representation of signals in the frequency domain. Not only commonly used in classical digital signal processing, but frequency analysis also remains powerful in the era of deep learning. [37] and [1] investigate the training and generalization of neural networks via frequency analysis. [35] analyzes the spectral bias of deep models regarding multiple vision parsing tasks. [8] applies frequency analysis to generated images for improving image synthesis quality, while [15] introduces a focal frequency loss that forces generative models to learn hard frequencies. [25] designs a frequency channel attention mechanism to compress channelwise information with scalars. [3] and [31] utilize frequency analysis in model compression. More applications of the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "frequency-domain representation can also be found in domain adaptation [12, 13] and position embedding [22, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.258 + ], + "angle": 0, + "content": "Although frequency analysis has been combined with deep learning techniques in recent years, it is generally used as a mature tool in prior arts without a deep investigation of the reasons why frequency analysis is useful. To this end, we experimentally studied the behavior tendency of frequency-domain convolutions in learning holistic geometries. Based on the investigation results, we designed the F-Learn strategy to efficiently learn holistic geometry clues in low-level feature maps for structured reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.27, + 0.591, + 0.285 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.536 + ], + "angle": 0, + "content": "Geometric features extracted at the early convolutional stage fundamentally support the structure reconstruction with precise geometry localization clues. Therefore, we are interested in addressing a key problem for learning reliable low-level geometric features: the low efficiency of fusing geometric fragments in shallow convolution layers. Considering that geometric information is generally high-frequency signals, we are motivated to study the problem mentioned before with frequency analysis for structure reconstruction. In the following of this section, we first present the preliminaries of frequency analysis. Next, we validate the high efficiency of learning in the frequency domain for holistic geometry recovery. Finally, we introduce our frequency-domain feature learning strategy (F-Learn) and the application of the F-Learn in a given structure reconstruction model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.816, + 0.562 + ], + "angle": 0, + "content": "3.1. Preliminaries of Frequency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.569, + 0.892, + 0.645 + ], + "angle": 0, + "content": "The discrete Fourier transform (DFT) is necessary for analyzing image data in the frequency domain. With \\( I(x,y) \\) denoting the color signal at the spatial position \\( (x,y) \\) of an \\( M\\times N \\) image, the DFT converts the information of \\( I \\) to a frequency map \\( F \\) in the frequency domain by Eq. (1)." + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.652, + 0.892, + 0.695 + ], + "angle": 0, + "content": "\\[\nF (u, v) = \\sum_ {x = 0} ^ {M - 1} \\sum_ {y = 0} ^ {N - 1} I (x, y) e ^ {- j 2 \\pi (u x / M + v y / N)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In Eq. (1), \\( F(u, v) \\) is a value that encodes a two-dimensional signal changing that belongs to the frequency determined by \\( u \\) and \\( v \\). The frequency map excels in holistically representing signals that show a similar frequency pattern. Therefore, we assume that the DFT is a powerful tool for efficiently learning structure-related geometry features. To validate this hypothesis, we conducted experiments with a geometry recovery task, and analyze the results quantitatively and qualitatively as below." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.757, + 0.864 + ], + "angle": 0, + "content": "3.2. Learning Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we study the learning efficiency of fusing geometric fragments in the space and frequency domains" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "with a simple geometry recovery task. We first decomposed a geometric structure composed of a circle and a square into several geometric fragments as illustrated in Fig. 2 (a), and left overlapping between neighboring fragments to avoid trial solutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.471, + 0.516 + ], + "angle": 0, + "content": "With the purpose to recover the binary image of the original structure with the geometric fragments, we studied the learning efficiency of frequency-domain and space-domain convolution models. We constructed a frequency-domain learning model (F-Learn) and several space-domain convolution models. The baseline was set as a space-domain convolution module (BConv) built with two \\(1 \\times 1\\) convolution layers and one \\(3 \\times 3\\) layer in between. The F-Learn model consists of three key components: (1) a DFT; (2) two parallel convolution modules same with the baseline for separately processing the real and imaginary parts; and (3) an inverse DFT (IDFT) followed by magnitude computation of complex values. For a fair comparison, we also combined a pair of baseline modules into two kinds of space-domain learning models with parallel and cascade structures, namely Conv-Casv1 and Conv-Parv1. Extra \\(3 \\times 3\\) convolutions were also added into the cascade and parallel space-domain models to simulate the function of DFT and IDFT. The modified models were named Conv-Casv2 and Conv-Parv2 for simplicity. For all models, the channel number of the intermediate feature maps was 64, and a \\(1 \\times 1\\) convolution was used as the final binary classifier. Detailed model settings are depicted in Fig. 2(b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.699 + ], + "angle": 0, + "content": "The experiments were conducted on \\(128 \\times 128\\) binary images of the geometric structure and fragments. The training of all models was completed with an Adam optimizer. The learning rate and training period were set to 0.1 and 50 epochs, respectively. The binary cross entropy loss (BCELoss) was chosen as the training guide. To alleviate the influence of randomness, we generated 100 seeds to set 100 trials, and all models were trained with the same seed in every trial. The performance of all models was evaluated with an F1-score measurement at a threshold of 0.5. Quantitative and qualitative results are shown in Fig. 3 and Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Fig. 3 draws the averaged F1-score curves of 100 trials. Compared with the space-domain counterparts, the frequency-domain model (F-Learn) shows superiority in recovering the complete structure in a more efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.825 + ], + "angle": 0, + "content": "The visualized results in Fig. 4 also show that the frequency-domain learning model exceeds the basic spatial convolution model with faster convergence speed and better structure recovery quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 reveals that the F-Learn model easily rebuilds the original structure, while the space-domain methods encounter difficulties in efficiently reconstructing the circle. We also find the difference in learning tendency between the F-Learn and space-domain models. The F-Learn model pri" + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.093, + 0.803, + 0.108 + ], + "angle": 0, + "content": "(a) Geometry Recovery Task" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.117, + 0.884, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.203, + 0.781, + 0.219 + ], + "angle": 0, + "content": "(b) The Model Settings" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.233, + 0.884, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Figure 2. The settings of the simple geometry recovery task. (a) The task of recovering the holistic structure with geometric fragments. (b) The detailed settings of compared models." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.485, + 0.886, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.665, + 0.892, + 0.694 + ], + "angle": 0, + "content": "Figure 3. The averaged F1-score curves of 100 trials for the F-Learn model and the compared space-domain methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "oritizes holistic integration before further enlarging the gap between the foreground and background. In contrast, the space-domain models first generate high responses to local areas that have strong signal intensity in the input. Hence, the space-domain models prefer overlapped areas without considering structural continuity. In practice, the holistic learning pattern of F-Learn will benefit the topology inference in two ways: (i) holistic geometry clues are ready for learning high-level features at the early stage, which alleviates the burden on learning structure-related low-level features with far supervision signal; (ii) the holistic geometry learning excels in preserving the connection in-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.294, + 0.468, + 0.322 + ], + "angle": 0, + "content": "Figure 4. The visualized results of F-Learn, Bconv and Cov-Parv1 at different epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.468, + 0.392 + ], + "angle": 0, + "content": "formation in the high-resolution low-level feature maps, which are often reused to provide accurate structure localization for topology inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.468, + 0.482 + ], + "angle": 0, + "content": "We attribute the advantages of the frequency domain based model to its ability to holistically process signals that belong to similar frequencies. Based on the findings above, we designed a frequency-domain feature learning strategy to generate holistic geometric clues in the low-level feature maps for structured reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.492, + 0.403, + 0.509 + ], + "angle": 0, + "content": "3.3. Frequency-domain Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.468, + 0.667 + ], + "angle": 0, + "content": "The proposed frequency-domain learning strategy (Flearn) inherits the architecture shown in Fig. 2. As the FLearn strategy will face more complex data and larger batch size in practice than in Sec. 3.2, we replace each convolution layer between DFT and IDFT with a combination (CBR) of a convolution layer, a batch normalization operation, and a relu activation function. We refer to all operations between DFT and IDFT collectively as a CBR group (CBRG). Given a set of geometric feature maps \\(\\{f_i|i = 0,\\dots ,N\\}\\), the FLearn strategy is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.679, + 0.345, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\left(F _ {i} ^ {\\text {r e}}, F _ {i} ^ {\\text {i m}}\\right) = \\operatorname {D F T} \\left(f _ {i}\\right);\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.699, + 0.469, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\{\\hat {F} _ {i} ^ {\\mathrm {r e}} \\}, \\{\\hat {F} _ {i} ^ {\\mathrm {i m}} \\} = \\mathrm {C B R G} (\\{F _ {i} ^ {\\mathrm {r e}} \\}), \\mathrm {C B R G} (\\{F _ {i} ^ {\\mathrm {i m}} \\}); \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.721, + 0.389, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\{\\hat {f} _ {i} \\} = | \\mathrm {I D F T} (\\{\\hat {F} _ {i} ^ {\\mathrm {r e}} \\}, \\{\\hat {F} _ {i} ^ {\\mathrm {i m}} \\}) |.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The F-learn strategy first converts every \\( f_{i} \\) into a pair of maps that record the real and imaginary parts of the DFT, i.e., \\( F_{i}^{\\mathrm{re}} \\) and \\( F_{i}^{\\mathrm{im}} \\). The maps of the real part \\( \\{F_{i}^{\\mathrm{re}}\\} \\) are subsequently processed by a CBRG that contains two \\( 1 \\times 1 \\) layers with a \\( 3 \\times 3 \\) one in between, and the same operation is also applied to \\( \\{F_{i}^{\\mathrm{im}}\\} \\). In the frequency domain, the \\( 1 \\times 1 \\) convolution directly fuses the information of the same frequency across the channel dimension, while the \\( 3 \\times 3 \\) counterpart enhances features with signals that have similar frequencies. The high-frequency geometric fragments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.196 + ], + "angle": 0, + "content": "in \\(\\{f_i\\}\\) are easily combined and enhanced in such a holistic way. The structurally informative space-domain features \\(\\{\\hat{f}_i\\}\\) can be finally obtained through an IDFT along with magnitude computation \\(|\\cdot|\\). With the parsimonious design, the F-learn strategy is readily inserted into a hierarchical convolutional backbone for holistic geometry feature learning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.197, + 0.892, + 0.228 + ], + "angle": 0, + "content": "For feature propagation, the \\(\\{f_i\\}\\) and \\(\\{\\hat{f}_i\\}\\) are fused into \\(\\{\\tilde{f}_i\\}\\) via" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.236, + 0.892, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\tilde {f} _ {i} \\right\\} = \\mathrm {C} _ {1 \\times 1} \\mathrm {B R} \\left(\\operatorname {C o n c a t} \\left(\\left\\{f _ {i} \\right\\}, \\left\\{\\hat {f} _ {i} \\right\\}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.263, + 0.891, + 0.292 + ], + "angle": 0, + "content": "where \\(\\mathrm{Concat}\\) and \\(\\mathbf{C}_{1\\times 1}\\mathbf{BR}\\) denote the concatenation operation and a \\(1\\times 1\\) convolution based CBR, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.293, + 0.892, + 0.352 + ], + "angle": 0, + "content": "The enhanced geometric features \\(\\{\\tilde{f}_i\\}\\) are then fed into the next convolution stage for higher-level feature learning and used to offer precise geometric localization for later structure inference." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.353, + 0.892, + 0.473 + ], + "angle": 0, + "content": "The F-Learn strategy can be easily inserted into a convolution base backbone. With the commonly-used ResNet50 backbone [10] as an example, the F-Learn strategy can be directly deployed after the first convolution to holistically learn geometric features. In experiments, we implemented our F-Learn strategy in a state-of-the-art approach, i.e., the holistic edge attention transformer (HEAT) [2], for structured reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.486, + 0.632, + 0.503 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.511, + 0.692, + 0.527 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.607, + 0.548 + ], + "angle": 0, + "content": "4.1.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.663 + ], + "angle": 0, + "content": "We tested our method on a dataset introduced by [24] for vectorizing world buildings. This dataset is built on the SpaceNet dataset [30] with 2001 images annotated with roof planar graphs. Each image contains a building instance, and the image is processed into a size of \\(256 \\times 256\\). In line with the prior arts [2], the training/validation/testing split is set to 1601/50/350." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.718, + 0.696 + ], + "angle": 0, + "content": "4.1.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluated the F-learn strategy with the HEAT model for structured reconstruction. The HEAT model was built on a ResNet50 backbone pretrained on ImageNet [4]. We performed our experiments with Pytorch [27] in Python3.7 and used a workstation with one NVIDIA RTX 3090 GPU. We adopted the same settings of loss functions as the HEAT model. The training of the model was completed with an AdamW optimizer, and the training period lasts 800 epochs. The learning rate was initialized as \\(2e - 4\\) and multiplied by a factor of 0.1 for the last \\(25\\%\\) epoch. In line with previous works [2, 24], we used precision, recall, and F1 score to evaluate the quality of structure reconstruction in terms of corner detection, edge inference, and region recovery." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21811" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.088, + 0.894, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between the F-Learn strategy and other methods in terms of corners detection, edge extraction, and region reconstruction. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.135, + 0.856, + 0.316 + ], + "angle": 0, + "content": "
Evaluation Type → \nMethodFully-neuralJointCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
IP [24]----74.5--53.1--55.7
Exp-Cls [43]--92.275.983.275.460.467.174.954.763.5
ConvMPN [42]✓-78.079.778.857.059.758.152.456.554.4
HAWP [39]✓✓90.981.285.776.668.172.174.155.463.4
LETR [36]✓✓87.874.880.859.758.659.168.348.756.8
HEAT [2]✓✓91.783.087.180.672.376.276.465.670.6
HEAT(retrain) [2]✓✓91.683.087.180.372.476.175.565.370.0
F-Learn (Ours)✓✓93.284.488.683.675.079.179.568.173.4
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.275, + 0.357 + ], + "angle": 0, + "content": "4.1.3 Competing Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.412 + ], + "angle": 0, + "content": "We compare the F-Learn strategy with six methods: HEAT [2], ConvMPN [42], IP [24], Exp-cls [43], HAWP [39] and LETR [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.471, + 0.548 + ], + "angle": 0, + "content": "HEAT is an attention-based method that takes a 2D raster image as an input and reconstructs a planar graph in an end-to-end manner. HEAT works via three steps: 1) extracting hierarchical features by ResNet50; 2) detecting corners with multi-scale features enriched by a deformable attention module; and 3) employing two weight-sharing Transformer decoders to classify edges and reason structures with detected corners. Our F-Learn strategy is adopted in the first step for learning structure-related features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.702 + ], + "angle": 0, + "content": "Conv-MPN uses a graph neural network to infer edges with a pre-trained corner detector. IP firstly detects geometric primitives and then reconstructs a planar graph through several post-processing steps. Exp-cls is based on the geometric primitives produced by other methods (e.g., IP and Conv-MPN) and reconstructs a planar graph through an explore-and-classify framework. LCNN and HAWP are methods specifically proposed for wireframe parsing. LETR is a transformer-based method that directly generates lines without post-processing and heuristic guidance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.313, + 0.73 + ], + "angle": 0, + "content": "4.2. Comparison and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.738, + 0.283, + 0.754 + ], + "angle": 0, + "content": "4.2.1 Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.854 + ], + "angle": 0, + "content": "The quantitative results displayed in Tab. 1 show that our F-Learn strategy advances roof structure reconstruction with state-of-the-art performance. The F-Learn strategy edges out other compared methods by significant performance gains regarding corner detection, edge inference, and region reconstruction under all metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "With respect to the detection of corners and edges, the F-Learn strategy outperforms the second-best by at least \\(1.0\\%\\) and \\(3.0\\%\\) in the precision measurement, while \\(1.4\\%\\) and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.522 + ], + "angle": 0, + "content": "\\(2.6\\%\\) under the recall metric. The improvement in extracting low-level geometric primitives indicates the outstanding ability of the F-Learn strategy for providing holistic and accurate geometric clues. The F-Learn strategy also brings at least a \\(2.8\\%\\) gain in F1 score in terms of region reconstruction, which reveals the assistance of the F-Learn strategy in inferring correct topological relations. Besides, it is worth noticing that the F-Learn strategy significantly outperforms the original HEAT model under all metrics, but only brings a little computation increase compared to HEAT. Tab. 2 presents a detailed comparison of computation requirements." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.533, + 0.894, + 0.618 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of the change of computing efficiency brought by F-Learn. Param. and FLOPs are the abbreviations of the parameter quantity and floating point operations, used to measure the complexity of an algorithm/model. Time per Image represents the time it takes the model to infer an image and is used to evaluate the calculation speed of the model." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.625, + 0.882, + 0.687 + ], + "angle": 0, + "content": "
Param.FLOPsTime per Image
w/o F-Learn68.747M670.919G0.10s
w/ F-Learn68.847M684.039G0.11s
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In Tab. 2, the computation increase of F-Learn strategy in terms of parameters, FLOPs, and inference time per image can be clearly seen, where the amount is little. Especially, the parameter increase brought by F-Learn strategy is only \\(0.145\\%\\), which is almost negligible. The results further demonstrates that the F-Learn strategy is simple yet effective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.696, + 0.846 + ], + "angle": 0, + "content": "4.2.2 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For perceptual comparison, we also present the visualized reconstruction results in Fig. 5. Compared to other methods, our F-Learn is capable of detecting low-level geometric" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.098, + 0.885, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.838, + 0.893, + 0.879 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on outdoor roof structure reconstruction. Other methods build wrong topological relations due to redundant or missing geometric primitives. Our F-Learn strategy derives the correct topological principles with corners and edges more fitted to the holistic structures." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21813" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.444, + 0.106 + ], + "angle": 0, + "content": "primitives that are consistent with the holistic structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.471, + 0.349 + ], + "angle": 0, + "content": "When reconstructing the roof shown in the first row, ConvMPN and Exp-cls tend to detect redundant corners due to the shadow inference and therefore wrongly separate the roof regions. The original HEAT outputs a non-closed roof structure because key edges are missing. In contrast, the corners and edges learned with the F-Learn strategy are better fitted into the overall structure. This is because the F-Learn strategy can offer low-level features that are rich in holistic geometric clues for topology-preserved roof structure reconstruction. The results in the following rows additionally validate that the F-Learn strategy can effectively extract geometric primitives that are easily ignored by other methods, especially the original HEAT model. By comparing the inferred planar graphs in the fifth row, one can also find that the F-Learn improves the HEAT model with robustness to irrelevant structure information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.361, + 0.242, + 0.375 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.385, + 0.421, + 0.4 + ], + "angle": 0, + "content": "4.3.1 Comparison with Space-domain Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.606 + ], + "angle": 0, + "content": "In addition to the exploration with a simple geometry fusion task in Sec. 3.2, we further study the F-Learn strategy with the space-domain learning strategies in the real scene for roof structure reconstruction. Similar to Sec. 3.2, we construct a space-domain learning baseline (BConv) with a basic module composed of two \\(1 \\times 1\\) convolution layers and one \\(3 \\times 3\\) layer in between. Furthermore, we design two more space-domain strategies with the basic modules arranged in parallel and cascade, denoted as Conv-Parv. and Conv-Casv., respectively. Same with the F-Learn, additional operations of batch normalization and ReLU activation are added after each convolution layer used in every space-domain strategy." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.619, + 0.471, + 0.703 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons with space-domain learning methods. The F-Learn strategy is compared with the space-domain counterparts under the F1-score measurement in terms of corner detection, edge inference, and region reconstruction. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.71, + 0.435, + 0.826 + ], + "angle": 0, + "content": "
MethodCornerEdgeRegion
HEAT87.176.170.0
HEAT-BConv87.476.270.7
HEAT-Conv-Cas.88.177.170.2
HEAT-Conv-Par.87.776.869.5
F-Learn88.679.173.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "With the F1-score measurement, Tab. 3 presents the performance of the F-learn and space-domain learning strategies in detecting geometric primitives. Although some space-domain methods are comparable to the proposed F" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Learn strategy in corner extraction, it can be seen that the F-Learn strategy gains over these methods in detecting edges and regions by at least \\(2\\%\\) and \\(2.7\\%\\), respectively. Because edges and regions are primitives closely related to roof topology, this phenomenon demonstrates the holistic geometry learned by the F-Learn is beneficial for the topology inference for roof reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.225, + 0.813, + 0.24 + ], + "angle": 0, + "content": "4.3.2 F-Learn Strategy at Different Layers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.373 + ], + "angle": 0, + "content": "We investigate the effect of applying the F-Learn strategy to feature maps of different levels with the HEAT model and the hierarchical ResNet50 backbone. The F-learn strategy is deployed to process feature maps generated at the first convolution layer, the first residual learning stage, and the second one. The sizes of the corresponding feature maps are \\(128 \\times 128\\), \\(64 \\times 64\\), and \\(32 \\times 32\\), respectively. The quantitative results are presented in Tab. 4." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of applying the F-Learn strategy to features at different levels. L0: the first convolution layer; L1: the first residual learning stage; L2: the second residual learning stage. Size: the feature map size. The F1-score metric is used for evaluating the performance of corner detection, edge inference, and region recovery. Higher scores mean better results. The bold values represent the best performance. (unit:%)" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.493, + 0.88, + 0.587 + ], + "angle": 0, + "content": "
MethodSizeCornerEdgeRegion
HEAT-87.176.170.0
F-Learn-L0128 × 12888.679.173.4
F-Learn-L164 × 6487.776.470.9
F-Learn-L232 × 3286.775.369.0
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.741 + ], + "angle": 0, + "content": "Tab. 4 implies that the higher the resolution is, the better the F-Learn strategy works. It is because high-resolution low-level feature maps contain more high-frequency geometric clues, while the low-resolution high-level ones are richer in abstract semantic information but poorer in geometric details. Therefore, the high-resolution feature maps are more suitable than the low-resolution counterparts for the F-Learn strategy to learn structurally informative geometries in roof reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.769, + 0.848, + 0.784 + ], + "angle": 0, + "content": "4.3.3 F-Learn Strategy with Various Backbones" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of the F-learn strategy, we evaluate the performance improvement brought by the F-Learn strategy with different backbones. The quantitative results are presented in Tab. 5. As shown by Tab. 5, with the deployment of the F-learn strategy, the performance of HEAT is consistently improved in terms of corners detection, edge extraction, and region reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21814" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.087, + 0.895, + 0.13 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of F-Learn strategy with HEAT built on different backbones, including ResNet18, ResNet34, and ResNet50. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.137, + 0.797, + 0.295 + ], + "angle": 0, + "content": "
BackboneCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
w/o F-LearnResNet1891.681.986.579.970.174.776.361.668.2
w/ F-LearnResNet1891.782.286.780.870.475.277.162.969.3
w/o F-LearnResNet3490.582.686.478.371.174.573.463.468.0
w/ F-LearnResNet3492.183.287.480.271.875.876.264.169.6
w/o F-LearnResNet5091.683.087.180.372.476.175.565.370.0
w/ F-LearnResNet5093.284.488.683.675.079.179.568.173.4
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.319, + 0.409, + 0.334 + ], + "angle": 0, + "content": "4.3.4 F-Learn Strategy in Topology Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.471, + 0.465 + ], + "angle": 0, + "content": "In this section, we further explore the effectiveness of the F-Learn strategy in topology inference. Edges and regions are primitives that highly relate to the inference of topological relationships, and these primitives are learned with corners and low-level features in the HEAT model. Therefore, we directly use the ground-truth corner map to focus on topology inference. We compare our F-Learn strategy with the space-domain methods used in Sec. 4.3.1." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.474, + 0.471, + 0.558 + ], + "angle": 0, + "content": "Table 6. Quantitative results of the F-Learn and space-domain strategies with respect to topology inference. The HEAT method with corner ground truth is set as the baseline. With corner annotations, all methods are studied with a focus on topology inference. The higher scores mean better results, and the best performance is bolded. (unit: \\(\\%\\) )" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.564, + 0.468, + 0.682 + ], + "angle": 0, + "content": "
Evaluation Type → MethodEdgeRegion
PrecRecallF1PrecRecallF1
HEAT93.584.188.590.273.881.2
HEAT-BConv94.483.988.889.471.779.6
HEAT-Conv-Cas94.885.389.891.373.781.6
HEAT-Conv-Par93.884.889.189.773.981.0
F-Learn95.187.591.189.577.683.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Tab. 6 presents the numeric results, and it can be seen that our F-Learn strategy brings the highest gains in most metrics for edge inference and region reconstruction. This phenomenon demonstrates that the F-Learn strategy effectively supports topology inference with an outstanding ability for holistic geometry learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.205, + 0.816 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present a frequency-domain feature learning strategy (F-Learn) to tackle the issue of wrong topology recovery caused by the lack of holistic clues in low-level features. Experiments with a geometry recovery task convincingly validate the efficiency of the F-Learn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.457 + ], + "angle": 0, + "content": "strategy in learning holistic geometry. In terms of the real scene, the F-Learn strategy achieves significant performance improvement of topological principle inference for roof structure reconstruction. The ablation studies verify that the F-Learn strategy outperforms the space-domain learning counterparts in capturing holistic geometric features regarding complex real scenes. We believe that it is promising to further explore the holistic learning ability brought by frequency analysis in more vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.681, + 0.485 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.583 + ], + "angle": 0, + "content": "This research is supported by NSFC projects under Grant 42071370, the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.596, + 0.597, + 0.611 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.621, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[1] Ronen Basri, Meirav Galun, Amnon Geifman, David Jacobs, Yoni Kasten, and Shira Kritchman. Frequency bias in neural networks for input of non-uniform density. In International Conference on Machine Learning, pages 685-694. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.692, + 0.894, + 0.746 + ], + "angle": 0, + "content": "[2] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In Proceedings of CVPR, pages 3866-3875, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.748, + 0.895, + 0.816 + ], + "angle": 0, + "content": "[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Q Weinberger, and Yixin Chen. Compressing convolutional neural networks in the frequency domain. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1475-1484, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.818, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.86, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In ECCV, pages 570-586, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.621, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[6] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[7] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. Communications of the ACM, 15(1):11-15, 1972. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.206, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[8] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[9] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the CVPR, pages 770-778, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[11] Linxi Huan, Nan Xue, Xianwei Zheng, Wei He, Jianya Gong, and Gui-Song Xia. Unmixing convolutional features for crisp edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6602-6609, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[12] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[13] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Rda: Robust domain adaptation via fourier adversarial attacking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8988-8999, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[14] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[16] Varsha Kamat-Sadekar and Subramaniam Ganesan. Complete description of multiple line segments using the hough transform. Image and Vision Computing, 16(9-10):597-613, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[17] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[18] Chen-Yu Lee, Vijay Badrinarayanan, Tomasz Malisiewicz, and Andrew Rabinovich. Roomnet: End-to-end room layout estimation. In Proceedings of the ICCV, pages 4865-4874, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[19] Hongjie Li, Xianwei Zheng, Mingyue Dong, Gui-Song Xia, and Hanjiang Xiong. Locally nonlinear affine verification for multisensor image matching. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[20] Yancong Lin, Silvia L Pintea, and Jan C van Gemert. Deep though-transform line priors. In Computer Vision-ECCV" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 323-340. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[21] Yun Liu, Ming-Ming Cheng, Xiaowei Hu, Jia-Wang Bian, Le Zhang, Xiang Bai, and Jinhui Tang. Richer convolutional features for edge detection. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1939-1946, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[22] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[23] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[24] Nelson Nauata and Yasutaka Furukawa. Vectorizing world buildings: Planar graph reconstruction by primitive detection and relationship inference. In ECCV 2020, pages 711-726. Springer, 2020. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[25] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.894, + 0.447 + ], + "angle": 0, + "content": "[26] Lawrence G Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[27] Benoit Steiner, Zachary Devito, Soumith Chintala, Sam Gross, Adam Paszke, Francisco Massa, Adam Lerer, Gregory Chanan, Zeming Lin, Edward Yang, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[28] Sinisa Stekovic, Mahdi Rad, Friedrich Fraundorfer, and Vincent Lepetit. Montefloor: Extending mcts for reconstructing accurate large-scale floor plans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16034-16043, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[29] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[30] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[31] Yunhe Wang, Chang Xu, Chao Xu, and Dacheng Tao. Packing convolutional neural networks in the frequency domain. IEEE transactions on pattern analysis and machine intelligence, 41(10):2495-2510, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[32] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the ICCV, pages 6772–6782, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[33] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the ECCV, pages 466-481, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. International Journal of Computer Vision, 125(1):3-18, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21816" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[35] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1740-1749, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[37] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[38] Nan Xue, Song Bai, Fudong Wang, Gui-Song Xia, Tianfu Wu, and Liangpei Zhang. Learning attraction field representation for robust line segment detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1595-1603, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[39] Nan Xue, Tianfu Wu, Song Bai, Fudong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2788-2797, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.404, + 0.469, + 0.458 + ], + "angle": 0, + "content": "[40] Jimei Yang, Brian Price, Scott Cohen, Honglak Lee, and Ming-Hsuan Yang. Object contour detection with a fully convolutional encoder-decoder network. In CVPR, pages 193-202, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.46, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[41] Zhiliang Zeng, Xianzhi Li, Ying Kin Yu, and Chi-Wing Fu. Deep floor plan recognition using a multi-task network with room-boundary-guided attention. In Proceedings of the ICCV, pages 9096-9104, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.517, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[42] Fuyang Zhang, Nelson Nauata, and Yasutaka Furukawa. Conv-mpn: Convolutional message passing neural network for structured outdoor architecture reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2798-2807, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.588, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[43] Fuyang Zhang, Xiang Xu, Nelson Nauata, and Yasutaka Furukawa. Structured outdoor architecture reconstruction by exploration and classification. In Proceedings of ICCV, pages 12427-12435, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.644, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[44] Xianwei Zheng, Zhuang Yuan, Zhen Dong, Mingyue Dong, Jianya Gong, and Hanjiang Xiong. Smoothly varying projective transformation for line segment matching. ISPRS Journal of Photogrammetry and Remote Sensing, 183:129-146, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.715, + 0.469, + 0.756 + ], + "angle": 0, + "content": "[45] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 962-971, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[46] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the ICCV, pages 4903-4911, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.801, + 0.469, + 0.854 + ], + "angle": 0, + "content": "[47] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the CVPR, pages 2051-2059, 2018. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21817" + } + ] +] \ No newline at end of file diff --git a/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_origin.pdf b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..08928135425aa67b9cc7f3cb0eae13c066602640 --- /dev/null +++ b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/6569e137-9378-406b-9cde-ca409ae1924f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42c6147f229546432704a942bf52dc0a5a340d892a4bb07a16c956c2017699c4 +size 5058385 diff --git a/2023/Holistic Geometric Feature Learning for Structured Reconstruction/full.md b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a7bae5c179d1e2508720b4ca489009d8fcdd3e60 --- /dev/null +++ b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/full.md @@ -0,0 +1,297 @@ +# Holistic Geometric Feature Learning for Structured Reconstruction + +Ziqiong Lu* Linxi Huan* Qiyuan Ma Xianwei Zheng† +The State Key Lab. LIESMARS, Wuhan University +{zql2018, whu.hlx, qiyuanma, zhengxw}@whu.edu.cn + +# Abstract + +The inference of topological principles is a key problem in structured reconstruction. We observe that wrongly predicted topological relationships are often incurred by the lack of holistic geometry clues in low-level features. Inspired by the fact that massive signals can be compactly described with frequency analysis, we experimentally explore the efficiency and tendency of learning structure geometry in the frequency domain. Accordingly, we propose a frequency-domain feature learning strategy (F-Learn) to fuse scattered geometric fragments holistically for topology-intact structure reasoning. Benefiting from the parsimonious design, the F-Learn strategy can be easily deployed into a deep reconstructor with a lightweight model modification. Experiments demonstrate that the F-Learn strategy can effectively introduce structure awareness into geometric primitive detection and topology inference, bringing significant performance improvement to final structured reconstruction. Code and pre-trained models are available at https://github.com/Geo-Tell/F-Learn. + +# 1. Introduction + +The structured reconstruction models the shape grammar/topology that depicts procedural shape generation [24], which can facilitate various downstream vision tasks, such as feature matching [19, 44], 3D modeling [23, 32] and 3D scene understanding [47, 42]. The recovery of a shape topology is generally realized via two steps: (i) geometric primitive detection and (ii) graph inference based on extracted geometric primitives (e.g., corners and edges). Following this primitive-to-structure principle, previous researchers usually resort to heatmaps of corners and edges for subsequent structure reasoning [18, 47, 42, 2]. With low-level parsing tasks promoted by deep learning techniques, many works have greatly improved the performance of high-level structure recovery in recent years [42, 24, 2]. + +Researchers commonly dedicate to graph inference with geometric features generated by proven hierarchical backbones (e.g., ResNet [10]) for structured reconstruction. However, incorrect topology inference remains a significant problem due to the lack of holistic geometric clues in the low-level features. Specifically, the geometric fragments extracted by shallow layers can hardly be fused into structurally informative features, leading to wrongly reasoned topological principles (as shown in Fig. 1 (a)). Different from the prior arts that focus on graph inference, this paper turns to study an efficient strategy for holistically learning structure-related features with low-level geometric fragments. + +Low-level geometric features fundamentally support high-level feature extraction and structure recovery. Specifically, the low-level shallow-layer feature maps provide information for precise geometry localization, which is essential for high-level primitive detection and topology inference. Restricted by the limited receptive field, it is hard to holistically capture structure geometries in the shallow layers. We consider that the low efficiency of processing low-level geometric fragments in the space domain is the root cause. + +Inspired by the fact that low-level features can be compactly encoded in the frequency domain, we are interested in achieving efficient geometry feature learning with frequency analysis. In the frequency map of a given feature map, every value encodes the global information of the feature map regarding the corresponding frequency. Therefore, performing convolution operations in the frequency domain actually means directly combining the spatial information in a holistic manner. + +To be concrete, given a set of low-level feature maps $\{f_n(x,y)\}$ and the corresponding frequency map $\{F_{n}(u,v)\}$ , the information encoded at $(u_{i},v_{j})$ describes a certain changing pattern of signals in the space domain. When applying a convolution operation to a frequency position $(u_{i},v_{j})$ , the feature components of related frequencies in every map of $\{f_n(x,y)\}$ will be merged across the channel dimension in the space domain. As the low-level geometric information generally belongs to high-frequency sig + +![](images/121d539bc33c61a8cabbe7c46eba0dbcf3d40d5d1cab69d0e675aec630564733.jpg) +Figure 1. Comparison of learning low-level features in the space and frequency domains. (a) Structure reconstruction with low-level geometric features learned in the space domain. The inefficient fusion of low-level geometric fragments loses the structure clues of the bottom left roof region and consequently results in missing topological relations in the final reconstructed structure graph. (b) Structure reconstruction with low-level geometric features learned with the proposed frequency-domain feature learning strategy (F-Learn). With the geometric features compactly processed in the frequency domain, our F-Learn strategy effectively achieves holistic geometry fusion for inferring the right topological principle for structure reconstruction. $E_{p}$ and $E_{r}$ denote the scores of precision and recall for detected edges, while $R_{p}$ and $R_{r}$ refer to those of reconstructed regions. + +nals, a frequency-domain convolution directly realizes the integration of geometric primitives scatted in $\{f_n(x,y)\}$ . The frequency-domain convolution thereby works more efficiently in holistic geometry learning than the spacedomain counterpart, which combines local geometric clues without a global view. + +Based on the discussion above, we propose a frequency-domain feature learning strategy (F-Learn) to efficiently extract holistic geometry features for guiding the inference of structure topology. Working in the frequency domain, the F-Learn strategy holistically fuses the separated geometric primitives into structurally informative features. The proposed F-Learn strategy can be readily applied to a primitive-to-structure framework for structured reconstruction. + +In summary, the main contributions of this paper are: + +- Exploration with a simple geometry recovery task sheds light on the difference in learning tendency between frequency- and space-domain convolutions. The results also validate the high efficiency of frequency-domain convolution in learning holistic geometry. +- A parsimonious frequency-domain feature learning strategy (F-Learn) is proposed to generate structurally informative geometry features for structured reconstruction. +- Experiments on vectorizing world buildings demonstrate that our F-Learn strategy greatly improves the performance of structured reconstruction. + +# 2. Related Work + +# 2.1. Geometric Primitive Detection + +The detection of geometric primitives is a long-standing vision task that has been extensively explored with handcrafted descriptors in the early stage [26, 9]. Recently, with the learning ability of hierarchical features, deep learning models have significantly promoted the completeness and precision of detected geometric primitives. A popular way for low-level geometry extraction is to perform heatmap regression or pixel-wise binary classification. The heatmap regression technique is widely adopted in corner detection. For instance, methods like CornerNet [17] and CenterNet [6] learn corner heatmaps for the downstream task of object detection. Binary classification is often used to localize edge pixels. Representative edge detectors like HED [34] and RCF [21] focus on fusing multi-level features for edge pixel classification, while some works pay attention to learning crisp edges or semantic contours [40, 5, 11]. + +Compared to the pixel-wise inference of corners and edges, the detection of line primitives is more structural as a line segment should be defined by two endpoints [7, 16]. Since the task of wireframe parsing was introduced by [14], increasing interest has been witnessed in inferring line segment candidates with learned junctions [45, 38, 20]. With the advance in low- and mid-level primitive detection, researchers have started to make efforts on structured reconstruction in recent years. + +# 2.2. Structured Reconstruction + +The structured reconstruction requires reasoning the overall topology of a given shape. Generally, structured reconstruction is a large research field that contains various tasks ranging from 3D object CAD modeling [32], layout estimation [18, 47], to roof extraction [43]. For instance-level structure reconstruction, many methods resort to generative models [32] or detecting key points with fixed topologies (e.g., skeletons of human bodies and hands) [33, 46]. + +As for scene-level structural modeling, the reconstruction targets usually belong to special semantic-related structures, such as planar building roofs and indoor floorplans. For these tasks, the derived structures have to match semantic regions with robustness to other irrelevant structure information. For example, room layout estimation has to ensure that connected lines can form wall regions without extra lines from windows and doors. Related approaches can be roughly categorized into two kinds: (i) the local primitive based and (ii) the global information based. Pioneering local primitive based works derive structural representation from an image by post-processing heatmaps of corners and edges [18, 47]. + +With respect to the global information based methods, it is a popular choice to form the final structure graph with structure-related regions segmented from the input image [41, 28]. Different from the segmentation based methods, recent studies consider a holistic graph inference with geometric primitives as nodes for high-level planer extraction. For example, Conv-MPN [42] adopts a variant of the graph neural network to pass messages across the whole graph, and HEAT [2] learns the topological pattern of edges with an attention transformer. Beyond exploring the high-level information in structure recovery, we are interested in improving the efficiency of exploiting low-level geometric features, which serve as the foundation of accurate structure topology inference. + +# 2.3. Frequency Analysis in Deep Learning + +Frequency analysis is a technique that provides a compressed representation of signals in the frequency domain. Not only commonly used in classical digital signal processing, but frequency analysis also remains powerful in the era of deep learning. [37] and [1] investigate the training and generalization of neural networks via frequency analysis. [35] analyzes the spectral bias of deep models regarding multiple vision parsing tasks. [8] applies frequency analysis to generated images for improving image synthesis quality, while [15] introduces a focal frequency loss that forces generative models to learn hard frequencies. [25] designs a frequency channel attention mechanism to compress channelwise information with scalars. [3] and [31] utilize frequency analysis in model compression. More applications of the + +frequency-domain representation can also be found in domain adaptation [12, 13] and position embedding [22, 29]. + +Although frequency analysis has been combined with deep learning techniques in recent years, it is generally used as a mature tool in prior arts without a deep investigation of the reasons why frequency analysis is useful. To this end, we experimentally studied the behavior tendency of frequency-domain convolutions in learning holistic geometries. Based on the investigation results, we designed the F-Learn strategy to efficiently learn holistic geometry clues in low-level feature maps for structured reconstruction. + +# 3. Method + +Geometric features extracted at the early convolutional stage fundamentally support the structure reconstruction with precise geometry localization clues. Therefore, we are interested in addressing a key problem for learning reliable low-level geometric features: the low efficiency of fusing geometric fragments in shallow convolution layers. Considering that geometric information is generally high-frequency signals, we are motivated to study the problem mentioned before with frequency analysis for structure reconstruction. In the following of this section, we first present the preliminaries of frequency analysis. Next, we validate the high efficiency of learning in the frequency domain for holistic geometry recovery. Finally, we introduce our frequency-domain feature learning strategy (F-Learn) and the application of the F-Learn in a given structure reconstruction model. + +# 3.1. Preliminaries of Frequency Analysis + +The discrete Fourier transform (DFT) is necessary for analyzing image data in the frequency domain. With $I(x,y)$ denoting the color signal at the spatial position $(x,y)$ of an $M\times N$ image, the DFT converts the information of $I$ to a frequency map $F$ in the frequency domain by Eq. (1). + +$$ +F (u, v) = \sum_ {x = 0} ^ {M - 1} \sum_ {y = 0} ^ {N - 1} I (x, y) e ^ {- j 2 \pi (u x / M + v y / N)}. \tag {1} +$$ + +In Eq. (1), $F(u, v)$ is a value that encodes a two-dimensional signal changing that belongs to the frequency determined by $u$ and $v$ . The frequency map excels in holistically representing signals that show a similar frequency pattern. Therefore, we assume that the DFT is a powerful tool for efficiently learning structure-related geometry features. To validate this hypothesis, we conducted experiments with a geometry recovery task, and analyze the results quantitatively and qualitatively as below. + +# 3.2. Learning Efficiency Analysis + +In this section, we study the learning efficiency of fusing geometric fragments in the space and frequency domains + +with a simple geometry recovery task. We first decomposed a geometric structure composed of a circle and a square into several geometric fragments as illustrated in Fig. 2 (a), and left overlapping between neighboring fragments to avoid trial solutions. + +With the purpose to recover the binary image of the original structure with the geometric fragments, we studied the learning efficiency of frequency-domain and space-domain convolution models. We constructed a frequency-domain learning model (F-Learn) and several space-domain convolution models. The baseline was set as a space-domain convolution module (BConv) built with two $1 \times 1$ convolution layers and one $3 \times 3$ layer in between. The F-Learn model consists of three key components: (1) a DFT; (2) two parallel convolution modules same with the baseline for separately processing the real and imaginary parts; and (3) an inverse DFT (IDFT) followed by magnitude computation of complex values. For a fair comparison, we also combined a pair of baseline modules into two kinds of space-domain learning models with parallel and cascade structures, namely Conv-Casv1 and Conv-Parv1. Extra $3 \times 3$ convolutions were also added into the cascade and parallel space-domain models to simulate the function of DFT and IDFT. The modified models were named Conv-Casv2 and Conv-Parv2 for simplicity. For all models, the channel number of the intermediate feature maps was 64, and a $1 \times 1$ convolution was used as the final binary classifier. Detailed model settings are depicted in Fig. 2(b). + +The experiments were conducted on $128 \times 128$ binary images of the geometric structure and fragments. The training of all models was completed with an Adam optimizer. The learning rate and training period were set to 0.1 and 50 epochs, respectively. The binary cross entropy loss (BCELoss) was chosen as the training guide. To alleviate the influence of randomness, we generated 100 seeds to set 100 trials, and all models were trained with the same seed in every trial. The performance of all models was evaluated with an F1-score measurement at a threshold of 0.5. Quantitative and qualitative results are shown in Fig. 3 and Fig. 4. + +Fig. 3 draws the averaged F1-score curves of 100 trials. Compared with the space-domain counterparts, the frequency-domain model (F-Learn) shows superiority in recovering the complete structure in a more efficient manner. + +The visualized results in Fig. 4 also show that the frequency-domain learning model exceeds the basic spatial convolution model with faster convergence speed and better structure recovery quality. + +Fig. 4 reveals that the F-Learn model easily rebuilds the original structure, while the space-domain methods encounter difficulties in efficiently reconstructing the circle. We also find the difference in learning tendency between the F-Learn and space-domain models. The F-Learn model pri + +![](images/dd3c7f7e769a96ad8d12190a426b4d23b5dadc0d71229d0a00c72ff1a291c2b2.jpg) +(a) Geometry Recovery Task +(b) The Model Settings + +![](images/1e0c7bc7a7d8c30fbf478c5db138fb44e7b5ca7b9d8c61e6bb5ba0467091fb25.jpg) +Figure 2. The settings of the simple geometry recovery task. (a) The task of recovering the holistic structure with geometric fragments. (b) The detailed settings of compared models. + +![](images/eeabac41c684490d02950e54d9c8ca26e382dac3578a1b5d6ed917ca292fc84b.jpg) +Figure 3. The averaged F1-score curves of 100 trials for the F-Learn model and the compared space-domain methods. + +oritizes holistic integration before further enlarging the gap between the foreground and background. In contrast, the space-domain models first generate high responses to local areas that have strong signal intensity in the input. Hence, the space-domain models prefer overlapped areas without considering structural continuity. In practice, the holistic learning pattern of F-Learn will benefit the topology inference in two ways: (i) holistic geometry clues are ready for learning high-level features at the early stage, which alleviates the burden on learning structure-related low-level features with far supervision signal; (ii) the holistic geometry learning excels in preserving the connection in- + +![](images/6c8c685bca19215ac7a484ae97450d8ba69eea8563ed213f38cb9939ef79f28e.jpg) +Figure 4. The visualized results of F-Learn, Bconv and Cov-Parv1 at different epochs. + +formation in the high-resolution low-level feature maps, which are often reused to provide accurate structure localization for topology inference. + +We attribute the advantages of the frequency domain based model to its ability to holistically process signals that belong to similar frequencies. Based on the findings above, we designed a frequency-domain feature learning strategy to generate holistic geometric clues in the low-level feature maps for structured reconstruction. + +# 3.3. Frequency-domain Learning Strategy + +The proposed frequency-domain learning strategy (Flearn) inherits the architecture shown in Fig. 2. As the FLearn strategy will face more complex data and larger batch size in practice than in Sec. 3.2, we replace each convolution layer between DFT and IDFT with a combination (CBR) of a convolution layer, a batch normalization operation, and a relu activation function. We refer to all operations between DFT and IDFT collectively as a CBR group (CBRG). Given a set of geometric feature maps $\{f_i|i = 0,\dots ,N\}$ , the FLearn strategy is formulated as + +$$ +\left(F _ {i} ^ {\text {r e}}, F _ {i} ^ {\text {i m}}\right) = \operatorname {D F T} \left(f _ {i}\right); +$$ + +$$ +\{\hat {F} _ {i} ^ {\mathrm {r e}} \}, \{\hat {F} _ {i} ^ {\mathrm {i m}} \} = \mathrm {C B R G} (\{F _ {i} ^ {\mathrm {r e}} \}), \mathrm {C B R G} (\{F _ {i} ^ {\mathrm {i m}} \}); \tag {2} +$$ + +$$ +\{\hat {f} _ {i} \} = | \mathrm {I D F T} (\{\hat {F} _ {i} ^ {\mathrm {r e}} \}, \{\hat {F} _ {i} ^ {\mathrm {i m}} \}) |. +$$ + +The F-learn strategy first converts every $f_{i}$ into a pair of maps that record the real and imaginary parts of the DFT, i.e., $F_{i}^{\mathrm{re}}$ and $F_{i}^{\mathrm{im}}$ . The maps of the real part $\{F_{i}^{\mathrm{re}}\}$ are subsequently processed by a CBRG that contains two $1 \times 1$ layers with a $3 \times 3$ one in between, and the same operation is also applied to $\{F_{i}^{\mathrm{im}}\}$ . In the frequency domain, the $1 \times 1$ convolution directly fuses the information of the same frequency across the channel dimension, while the $3 \times 3$ counterpart enhances features with signals that have similar frequencies. The high-frequency geometric fragments + +in $\{f_i\}$ are easily combined and enhanced in such a holistic way. The structurally informative space-domain features $\{\hat{f}_i\}$ can be finally obtained through an IDFT along with magnitude computation $|\cdot|$ . With the parsimonious design, the F-learn strategy is readily inserted into a hierarchical convolutional backbone for holistic geometry feature learning. + +For feature propagation, the $\{f_i\}$ and $\{\hat{f}_i\}$ are fused into $\{\tilde{f}_i\}$ via + +$$ +\left\{\tilde {f} _ {i} \right\} = \mathrm {C} _ {1 \times 1} \mathrm {B R} \left(\operatorname {C o n c a t} \left(\left\{f _ {i} \right\}, \left\{\hat {f} _ {i} \right\}\right)\right), \tag {3} +$$ + +where $\mathrm{Concat}$ and $\mathbf{C}_{1\times 1}\mathbf{BR}$ denote the concatenation operation and a $1\times 1$ convolution based CBR, respectively. + +The enhanced geometric features $\{\tilde{f}_i\}$ are then fed into the next convolution stage for higher-level feature learning and used to offer precise geometric localization for later structure inference. + +The F-Learn strategy can be easily inserted into a convolution base backbone. With the commonly-used ResNet50 backbone [10] as an example, the F-Learn strategy can be directly deployed after the first convolution to holistically learn geometric features. In experiments, we implemented our F-Learn strategy in a state-of-the-art approach, i.e., the holistic edge attention transformer (HEAT) [2], for structured reconstruction. + +# 4. Experiments + +# 4.1. Experiment Settings + +# 4.1.1 Dataset + +We tested our method on a dataset introduced by [24] for vectorizing world buildings. This dataset is built on the SpaceNet dataset [30] with 2001 images annotated with roof planar graphs. Each image contains a building instance, and the image is processed into a size of $256 \times 256$ . In line with the prior arts [2], the training/validation/testing split is set to 1601/50/350. + +# 4.1.2 Implementation Details + +We evaluated the F-learn strategy with the HEAT model for structured reconstruction. The HEAT model was built on a ResNet50 backbone pretrained on ImageNet [4]. We performed our experiments with Pytorch [27] in Python3.7 and used a workstation with one NVIDIA RTX 3090 GPU. We adopted the same settings of loss functions as the HEAT model. The training of the model was completed with an AdamW optimizer, and the training period lasts 800 epochs. The learning rate was initialized as $2e - 4$ and multiplied by a factor of 0.1 for the last $25\%$ epoch. In line with previous works [2, 24], we used precision, recall, and F1 score to evaluate the quality of structure reconstruction in terms of corner detection, edge inference, and region recovery. + +Table 1. Quantitative comparison between the F-Learn strategy and other methods in terms of corners detection, edge extraction, and region reconstruction. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%) + +
Evaluation Type → +MethodFully-neuralJointCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
IP [24]----74.5--53.1--55.7
Exp-Cls [43]--92.275.983.275.460.467.174.954.763.5
ConvMPN [42]✓-78.079.778.857.059.758.152.456.554.4
HAWP [39]✓✓90.981.285.776.668.172.174.155.463.4
LETR [36]✓✓87.874.880.859.758.659.168.348.756.8
HEAT [2]✓✓91.783.087.180.672.376.276.465.670.6
HEAT(retrain) [2]✓✓91.683.087.180.372.476.175.565.370.0
F-Learn (Ours)✓✓93.284.488.683.675.079.179.568.173.4
+ +# 4.1.3 Competing Methods + +We compare the F-Learn strategy with six methods: HEAT [2], ConvMPN [42], IP [24], Exp-cls [43], HAWP [39] and LETR [36]. + +HEAT is an attention-based method that takes a 2D raster image as an input and reconstructs a planar graph in an end-to-end manner. HEAT works via three steps: 1) extracting hierarchical features by ResNet50; 2) detecting corners with multi-scale features enriched by a deformable attention module; and 3) employing two weight-sharing Transformer decoders to classify edges and reason structures with detected corners. Our F-Learn strategy is adopted in the first step for learning structure-related features. + +Conv-MPN uses a graph neural network to infer edges with a pre-trained corner detector. IP firstly detects geometric primitives and then reconstructs a planar graph through several post-processing steps. Exp-cls is based on the geometric primitives produced by other methods (e.g., IP and Conv-MPN) and reconstructs a planar graph through an explore-and-classify framework. LCNN and HAWP are methods specifically proposed for wireframe parsing. LETR is a transformer-based method that directly generates lines without post-processing and heuristic guidance. + +# 4.2. Comparison and Analysis + +# 4.2.1 Quantitative Analysis + +The quantitative results displayed in Tab. 1 show that our F-Learn strategy advances roof structure reconstruction with state-of-the-art performance. The F-Learn strategy edges out other compared methods by significant performance gains regarding corner detection, edge inference, and region reconstruction under all metrics. + +With respect to the detection of corners and edges, the F-Learn strategy outperforms the second-best by at least $1.0\%$ and $3.0\%$ in the precision measurement, while $1.4\%$ and + +$2.6\%$ under the recall metric. The improvement in extracting low-level geometric primitives indicates the outstanding ability of the F-Learn strategy for providing holistic and accurate geometric clues. The F-Learn strategy also brings at least a $2.8\%$ gain in F1 score in terms of region reconstruction, which reveals the assistance of the F-Learn strategy in inferring correct topological relations. Besides, it is worth noticing that the F-Learn strategy significantly outperforms the original HEAT model under all metrics, but only brings a little computation increase compared to HEAT. Tab. 2 presents a detailed comparison of computation requirements. + +Table 2. Quantitative evaluation of the change of computing efficiency brought by F-Learn. Param. and FLOPs are the abbreviations of the parameter quantity and floating point operations, used to measure the complexity of an algorithm/model. Time per Image represents the time it takes the model to infer an image and is used to evaluate the calculation speed of the model. + +
Param.FLOPsTime per Image
w/o F-Learn68.747M670.919G0.10s
w/ F-Learn68.847M684.039G0.11s
+ +In Tab. 2, the computation increase of F-Learn strategy in terms of parameters, FLOPs, and inference time per image can be clearly seen, where the amount is little. Especially, the parameter increase brought by F-Learn strategy is only $0.145\%$ , which is almost negligible. The results further demonstrates that the F-Learn strategy is simple yet effective. + +# 4.2.2 Qualitative Analysis + +For perceptual comparison, we also present the visualized reconstruction results in Fig. 5. Compared to other methods, our F-Learn is capable of detecting low-level geometric + +![](images/079dc50752f9b4a8e44be9b0762e6f2bf5577cb93537c68ca7246206f1013004.jpg) +Figure 5. Qualitative results on outdoor roof structure reconstruction. Other methods build wrong topological relations due to redundant or missing geometric primitives. Our F-Learn strategy derives the correct topological principles with corners and edges more fitted to the holistic structures. + +primitives that are consistent with the holistic structure. + +When reconstructing the roof shown in the first row, ConvMPN and Exp-cls tend to detect redundant corners due to the shadow inference and therefore wrongly separate the roof regions. The original HEAT outputs a non-closed roof structure because key edges are missing. In contrast, the corners and edges learned with the F-Learn strategy are better fitted into the overall structure. This is because the F-Learn strategy can offer low-level features that are rich in holistic geometric clues for topology-preserved roof structure reconstruction. The results in the following rows additionally validate that the F-Learn strategy can effectively extract geometric primitives that are easily ignored by other methods, especially the original HEAT model. By comparing the inferred planar graphs in the fifth row, one can also find that the F-Learn improves the HEAT model with robustness to irrelevant structure information. + +# 4.3. Ablation Studies + +# 4.3.1 Comparison with Space-domain Methods + +In addition to the exploration with a simple geometry fusion task in Sec. 3.2, we further study the F-Learn strategy with the space-domain learning strategies in the real scene for roof structure reconstruction. Similar to Sec. 3.2, we construct a space-domain learning baseline (BConv) with a basic module composed of two $1 \times 1$ convolution layers and one $3 \times 3$ layer in between. Furthermore, we design two more space-domain strategies with the basic modules arranged in parallel and cascade, denoted as Conv-Parv. and Conv-Casv., respectively. Same with the F-Learn, additional operations of batch normalization and ReLU activation are added after each convolution layer used in every space-domain strategy. + +Table 3. Quantitative comparisons with space-domain learning methods. The F-Learn strategy is compared with the space-domain counterparts under the F1-score measurement in terms of corner detection, edge inference, and region reconstruction. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%) + +
MethodCornerEdgeRegion
HEAT87.176.170.0
HEAT-BConv87.476.270.7
HEAT-Conv-Cas.88.177.170.2
HEAT-Conv-Par.87.776.869.5
F-Learn88.679.173.4
+ +With the F1-score measurement, Tab. 3 presents the performance of the F-learn and space-domain learning strategies in detecting geometric primitives. Although some space-domain methods are comparable to the proposed F + +Learn strategy in corner extraction, it can be seen that the F-Learn strategy gains over these methods in detecting edges and regions by at least $2\%$ and $2.7\%$ , respectively. Because edges and regions are primitives closely related to roof topology, this phenomenon demonstrates the holistic geometry learned by the F-Learn is beneficial for the topology inference for roof reconstruction. + +# 4.3.2 F-Learn Strategy at Different Layers + +We investigate the effect of applying the F-Learn strategy to feature maps of different levels with the HEAT model and the hierarchical ResNet50 backbone. The F-learn strategy is deployed to process feature maps generated at the first convolution layer, the first residual learning stage, and the second one. The sizes of the corresponding feature maps are $128 \times 128$ , $64 \times 64$ , and $32 \times 32$ , respectively. The quantitative results are presented in Tab. 4. + +Table 4. Quantitative comparison of applying the F-Learn strategy to features at different levels. L0: the first convolution layer; L1: the first residual learning stage; L2: the second residual learning stage. Size: the feature map size. The F1-score metric is used for evaluating the performance of corner detection, edge inference, and region recovery. Higher scores mean better results. The bold values represent the best performance. (unit:%) + +
MethodSizeCornerEdgeRegion
HEAT-87.176.170.0
F-Learn-L0128 × 12888.679.173.4
F-Learn-L164 × 6487.776.470.9
F-Learn-L232 × 3286.775.369.0
+ +Tab. 4 implies that the higher the resolution is, the better the F-Learn strategy works. It is because high-resolution low-level feature maps contain more high-frequency geometric clues, while the low-resolution high-level ones are richer in abstract semantic information but poorer in geometric details. Therefore, the high-resolution feature maps are more suitable than the low-resolution counterparts for the F-Learn strategy to learn structurally informative geometries in roof reconstruction. + +# 4.3.3 F-Learn Strategy with Various Backbones + +To further demonstrate the effectiveness of the F-learn strategy, we evaluate the performance improvement brought by the F-Learn strategy with different backbones. The quantitative results are presented in Tab. 5. As shown by Tab. 5, with the deployment of the F-learn strategy, the performance of HEAT is consistently improved in terms of corners detection, edge extraction, and region reconstruction. + +Table 5. Quantitative comparison of F-Learn strategy with HEAT built on different backbones, including ResNet18, ResNet34, and ResNet50. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%) + +
BackboneCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
w/o F-LearnResNet1891.681.986.579.970.174.776.361.668.2
w/ F-LearnResNet1891.782.286.780.870.475.277.162.969.3
w/o F-LearnResNet3490.582.686.478.371.174.573.463.468.0
w/ F-LearnResNet3492.183.287.480.271.875.876.264.169.6
w/o F-LearnResNet5091.683.087.180.372.476.175.565.370.0
w/ F-LearnResNet5093.284.488.683.675.079.179.568.173.4
+ +# 4.3.4 F-Learn Strategy in Topology Inference + +In this section, we further explore the effectiveness of the F-Learn strategy in topology inference. Edges and regions are primitives that highly relate to the inference of topological relationships, and these primitives are learned with corners and low-level features in the HEAT model. Therefore, we directly use the ground-truth corner map to focus on topology inference. We compare our F-Learn strategy with the space-domain methods used in Sec. 4.3.1. + +Table 6. Quantitative results of the F-Learn and space-domain strategies with respect to topology inference. The HEAT method with corner ground truth is set as the baseline. With corner annotations, all methods are studied with a focus on topology inference. The higher scores mean better results, and the best performance is bolded. (unit: $\%$ ) + +
Evaluation Type → MethodEdgeRegion
PrecRecallF1PrecRecallF1
HEAT93.584.188.590.273.881.2
HEAT-BConv94.483.988.889.471.779.6
HEAT-Conv-Cas94.885.389.891.373.781.6
HEAT-Conv-Par93.884.889.189.773.981.0
F-Learn95.187.591.189.577.683.1
+ +Tab. 6 presents the numeric results, and it can be seen that our F-Learn strategy brings the highest gains in most metrics for edge inference and region reconstruction. This phenomenon demonstrates that the F-Learn strategy effectively supports topology inference with an outstanding ability for holistic geometry learning. + +# 5. Conclusions + +In this paper, we present a frequency-domain feature learning strategy (F-Learn) to tackle the issue of wrong topology recovery caused by the lack of holistic clues in low-level features. Experiments with a geometry recovery task convincingly validate the efficiency of the F-Learn + +strategy in learning holistic geometry. In terms of the real scene, the F-Learn strategy achieves significant performance improvement of topological principle inference for roof structure reconstruction. The ablation studies verify that the F-Learn strategy outperforms the space-domain learning counterparts in capturing holistic geometric features regarding complex real scenes. We believe that it is promising to further explore the holistic learning ability brought by frequency analysis in more vision tasks. + +# 6. Acknowledgement + +This research is supported by NSFC projects under Grant 42071370, the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084. + +# References + +[1] Ronen Basri, Meirav Galun, Amnon Geifman, David Jacobs, Yoni Kasten, and Shira Kritchman. Frequency bias in neural networks for input of non-uniform density. In International Conference on Machine Learning, pages 685-694. PMLR, 2020. 3 +[2] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In Proceedings of CVPR, pages 3866-3875, 2022. 1, 3, 5, 6 +[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Q Weinberger, and Yixin Chen. Compressing convolutional neural networks in the frequency domain. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1475-1484, 2016. 3 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5 +[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In ECCV, pages 570-586, 2018. 2 + +[6] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 2 +[7] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. Communications of the ACM, 15(1):11-15, 1972. 2 +[8] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3 +[9] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the CVPR, pages 770-778, 2016. 1, 5 +[11] Linxi Huan, Nan Xue, Xianwei Zheng, Wei He, Jianya Gong, and Gui-Song Xia. Unmixing convolutional features for crisp edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6602-6609, 2021. 2 +[12] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 3 +[13] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Rda: Robust domain adaptation via fourier adversarial attacking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8988-8999, 2021. 3 +[14] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 2 +[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3 +[16] Varsha Kamat-Sadekar and Subramaniam Ganesan. Complete description of multiple line segments using the hough transform. Image and Vision Computing, 16(9-10):597-613, 1998. 2 +[17] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2 +[18] Chen-Yu Lee, Vijay Badrinarayanan, Tomasz Malisiewicz, and Andrew Rabinovich. Roomnet: End-to-end room layout estimation. In Proceedings of the ICCV, pages 4865-4874, 2017. 1, 3 +[19] Hongjie Li, Xianwei Zheng, Mingyue Dong, Gui-Song Xia, and Hanjiang Xiong. Locally nonlinear affine verification for multisensor image matching. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 1 +[20] Yancong Lin, Silvia L Pintea, and Jan C van Gemert. Deep though-transform line priors. In Computer Vision-ECCV + +2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 323-340. Springer, 2020. 2 +[21] Yun Liu, Ming-Ming Cheng, Xiaowei Hu, Jia-Wang Bian, Le Zhang, Xiang Bai, and Jinhui Tang. Richer convolutional features for edge detection. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1939-1946, 2019. 2 +[22] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3 +[23] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229. PMLR, 2020. 1 +[24] Nelson Nauata and Yasutaka Furukawa. Vectorizing world buildings: Planar graph reconstruction by primitive detection and relationship inference. In ECCV 2020, pages 711-726. Springer, 2020. 1, 5, 6 +[25] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3 +[26] Lawrence G Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2 +[27] Benoit Steiner, Zachary Devito, Soumith Chintala, Sam Gross, Adam Paszke, Francisco Massa, Adam Lerer, Gregory Chanan, Zeming Lin, Edward Yang, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5 +[28] Sinisa Stekovic, Mahdi Rad, Friedrich Fraundorfer, and Vincent Lepetit. Montefloor: Extending mcts for reconstructing accurate large-scale floor plans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16034-16043, 2021. 3 +[29] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3 +[30] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 5 +[31] Yunhe Wang, Chang Xu, Chao Xu, and Dacheng Tao. Packing convolutional neural networks in the frequency domain. IEEE transactions on pattern analysis and machine intelligence, 41(10):2495-2510, 2018. 3 +[32] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the ICCV, pages 6772–6782, 2021. 1, 3 +[33] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the ECCV, pages 466-481, 2018. 3 +[34] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. International Journal of Computer Vision, 125(1):3-18, 2017. 2 + +[35] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1740-1749, 2020. 3 +[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 6 +[37] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3 +[38] Nan Xue, Song Bai, Fudong Wang, Gui-Song Xia, Tianfu Wu, and Liangpei Zhang. Learning attraction field representation for robust line segment detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1595-1603, 2019. 2 +[39] Nan Xue, Tianfu Wu, Song Bai, Fudong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2788-2797, 2020. 6 +[40] Jimei Yang, Brian Price, Scott Cohen, Honglak Lee, and Ming-Hsuan Yang. Object contour detection with a fully convolutional encoder-decoder network. In CVPR, pages 193-202, 2016. 2 +[41] Zhiliang Zeng, Xianzhi Li, Ying Kin Yu, and Chi-Wing Fu. Deep floor plan recognition using a multi-task network with room-boundary-guided attention. In Proceedings of the ICCV, pages 9096-9104, 2019. 3 +[42] Fuyang Zhang, Nelson Nauata, and Yasutaka Furukawa. Conv-mpn: Convolutional message passing neural network for structured outdoor architecture reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2798-2807, 2020. 1, 3, 6 +[43] Fuyang Zhang, Xiang Xu, Nelson Nauata, and Yasutaka Furukawa. Structured outdoor architecture reconstruction by exploration and classification. In Proceedings of ICCV, pages 12427-12435, 2021. 3, 6 +[44] Xianwei Zheng, Zhuang Yuan, Zhen Dong, Mingyue Dong, Jianya Gong, and Hanjiang Xiong. Smoothly varying projective transformation for line segment matching. ISPRS Journal of Photogrammetry and Remote Sensing, 183:129-146, 2022. 1 +[45] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 962-971, 2019. 2 +[46] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the ICCV, pages 4903-4911, 2017. 3 +[47] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the CVPR, pages 2051-2059, 2018. 1, 3 \ No newline at end of file diff --git a/2023/Holistic Geometric Feature Learning for Structured Reconstruction/images.zip b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..87576981dd4566f2431a68589aadde1d8256404c --- /dev/null +++ b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca058570aac65a679f82d1e9e2965a4af9f264ef0ac77af4913272c9988644ea +size 820773 diff --git a/2023/Holistic Geometric Feature Learning for Structured Reconstruction/layout.json b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d96021fbc9db678017540bb39f8935c0f34d07f3 --- /dev/null +++ b/2023/Holistic Geometric Feature Learning for Structured Reconstruction/layout.json @@ -0,0 +1,7581 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 506, + 121 + ], + "type": "text", + "content": "Holistic Geometric Feature Learning for Structured Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 142, + 463, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 142, + 463, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 142, + 463, + 185 + ], + "type": "text", + "content": "Ziqiong Lu* Linxi Huan* Qiyuan Ma Xianwei Zheng† \nThe State Key Lab. LIESMARS, Wuhan University \n{zql2018, whu.hlx, qiyuanma, zhengxw}@whu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 250, + 289, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 289, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 289, + 478 + ], + "type": "text", + "content": "The inference of topological principles is a key problem in structured reconstruction. We observe that wrongly predicted topological relationships are often incurred by the lack of holistic geometry clues in low-level features. Inspired by the fact that massive signals can be compactly described with frequency analysis, we experimentally explore the efficiency and tendency of learning structure geometry in the frequency domain. Accordingly, we propose a frequency-domain feature learning strategy (F-Learn) to fuse scattered geometric fragments holistically for topology-intact structure reasoning. Benefiting from the parsimonious design, the F-Learn strategy can be easily deployed into a deep reconstructor with a lightweight model modification. Experiments demonstrate that the F-Learn strategy can effectively introduce structure awareness into geometric primitive detection and topology inference, bringing significant performance improvement to final structured reconstruction. Code and pre-trained models are available at https://github.com/Geo-Tell/F-Learn." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 519, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 687 + ], + "type": "text", + "content": "The structured reconstruction models the shape grammar/topology that depicts procedural shape generation [24], which can facilitate various downstream vision tasks, such as feature matching [19, 44], 3D modeling [23, 32] and 3D scene understanding [47, 42]. The recovery of a shape topology is generally realized via two steps: (i) geometric primitive detection and (ii) graph inference based on extracted geometric primitives (e.g., corners and edges). Following this primitive-to-structure principle, previous researchers usually resort to heatmaps of corners and edges for subsequent structure reasoning [18, 47, 42, 2]. With low-level parsing tasks promoted by deep learning techniques, many works have greatly improved the performance of high-level structure recovery in recent years [42, 24, 2]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 213, + 546, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 546, + 369 + ], + "type": "text", + "content": "Researchers commonly dedicate to graph inference with geometric features generated by proven hierarchical backbones (e.g., ResNet [10]) for structured reconstruction. However, incorrect topology inference remains a significant problem due to the lack of holistic geometric clues in the low-level features. Specifically, the geometric fragments extracted by shallow layers can hardly be fused into structurally informative features, leading to wrongly reasoned topological principles (as shown in Fig. 1 (a)). Different from the prior arts that focus on graph inference, this paper turns to study an efficient strategy for holistically learning structure-related features with low-level geometric fragments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 373, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 491 + ], + "type": "text", + "content": "Low-level geometric features fundamentally support high-level feature extraction and structure recovery. Specifically, the low-level shallow-layer feature maps provide information for precise geometry localization, which is essential for high-level primitive detection and topology inference. Restricted by the limited receptive field, it is hard to holistically capture structure geometries in the shallow layers. We consider that the low efficiency of processing low-level geometric fragments in the space domain is the root cause." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 495, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 602 + ], + "type": "text", + "content": "Inspired by the fact that low-level features can be compactly encoded in the frequency domain, we are interested in achieving efficient geometry feature learning with frequency analysis. In the frequency map of a given feature map, every value encodes the global information of the feature map regarding the corresponding frequency. Therefore, performing convolution operations in the frequency domain actually means directly combining the spatial information in a holistic manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "To be concrete, given a set of low-level feature maps " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{f_n(x,y)\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " and the corresponding frequency map " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{F_{n}(u,v)\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", the information encoded at " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(u_{i},v_{j})" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " describes a certain changing pattern of signals in the space domain. When applying a convolution operation to a frequency position " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(u_{i},v_{j})" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", the feature components of related frequencies in every map of " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{f_n(x,y)\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " will be merged across the channel dimension in the space domain. As the low-level geometric information generally belongs to high-frequency sig" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 139, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 139, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 139, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21807" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 545, + 257 + ], + "type": "image", + "image_path": "121d539bc33c61a8cabbe7c46eba0dbcf3d40d5d1cab69d0e675aec630564733.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "text", + "content": "Figure 1. Comparison of learning low-level features in the space and frequency domains. (a) Structure reconstruction with low-level geometric features learned in the space domain. The inefficient fusion of low-level geometric fragments loses the structure clues of the bottom left roof region and consequently results in missing topological relations in the final reconstructed structure graph. (b) Structure reconstruction with low-level geometric features learned with the proposed frequency-domain feature learning strategy (F-Learn). With the geometric features compactly processed in the frequency domain, our F-Learn strategy effectively achieves holistic geometry fusion for inferring the right topological principle for structure reconstruction. " + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "inline_equation", + "content": "E_{r}" + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "text", + "content": " denote the scores of precision and recall for detected edges, while " + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "inline_equation", + "content": "R_{p}" + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "inline_equation", + "content": "R_{r}" + }, + { + "bbox": [ + 46, + 259, + 547, + 336 + ], + "type": "text", + "content": " refer to those of reconstructed regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 356, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 427 + ], + "type": "text", + "content": "nals, a frequency-domain convolution directly realizes the integration of geometric primitives scatted in " + }, + { + "bbox": [ + 46, + 356, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\{f_n(x,y)\\}" + }, + { + "bbox": [ + 46, + 356, + 287, + 427 + ], + "type": "text", + "content": ". The frequency-domain convolution thereby works more efficiently in holistic geometry learning than the spacedomain counterpart, which combines local geometric clues without a global view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 429, + 288, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 288, + 524 + ], + "type": "text", + "content": "Based on the discussion above, we propose a frequency-domain feature learning strategy (F-Learn) to efficiently extract holistic geometry features for guiding the inference of structure topology. Working in the frequency domain, the F-Learn strategy holistically fuses the separated geometric primitives into structurally informative features. The proposed F-Learn strategy can be readily applied to a primitive-to-structure framework for structured reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 525, + 274, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 525, + 274, + 538 + ], + "spans": [ + { + "bbox": [ + 59, + 525, + 274, + 538 + ], + "type": "text", + "content": "In summary, the main contributions of this paper are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 548, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 548, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 548, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 58, + 548, + 287, + 608 + ], + "type": "text", + "content": "- Exploration with a simple geometry recovery task sheds light on the difference in learning tendency between frequency- and space-domain convolutions. The results also validate the high efficiency of frequency-domain convolution in learning holistic geometry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 619, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 619, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 58, + 619, + 288, + 666 + ], + "type": "text", + "content": "- A parsimonious frequency-domain feature learning strategy (F-Learn) is proposed to generate structurally informative geometry features for structured reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 288, + 713 + ], + "type": "text", + "content": "- Experiments on vectorizing world buildings demonstrate that our F-Learn strategy greatly improves the performance of structured reconstruction." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 354, + 392, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 392, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 392, + 367 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 381, + 471, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 471, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 471, + 393 + ], + "type": "text", + "content": "2.1. Geometric Primitive Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 407, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 598 + ], + "type": "text", + "content": "The detection of geometric primitives is a long-standing vision task that has been extensively explored with handcrafted descriptors in the early stage [26, 9]. Recently, with the learning ability of hierarchical features, deep learning models have significantly promoted the completeness and precision of detected geometric primitives. A popular way for low-level geometry extraction is to perform heatmap regression or pixel-wise binary classification. The heatmap regression technique is widely adopted in corner detection. For instance, methods like CornerNet [17] and CenterNet [6] learn corner heatmaps for the downstream task of object detection. Binary classification is often used to localize edge pixels. Representative edge detectors like HED [34] and RCF [21] focus on fusing multi-level features for edge pixel classification, while some works pay attention to learning crisp edges or semantic contours [40, 5, 11]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Compared to the pixel-wise inference of corners and edges, the detection of line primitives is more structural as a line segment should be defined by two endpoints [7, 16]. Since the task of wireframe parsing was introduced by [14], increasing interest has been witnessed in inferring line segment candidates with learned junctions [45, 38, 20]. With the advance in low- and mid-level primitive detection, researchers have started to make efforts on structured reconstruction in recent years." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21808" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 195, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 195, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 195, + 83 + ], + "type": "text", + "content": "2.2. Structured Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 197 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 197 + ], + "type": "text", + "content": "The structured reconstruction requires reasoning the overall topology of a given shape. Generally, structured reconstruction is a large research field that contains various tasks ranging from 3D object CAD modeling [32], layout estimation [18, 47], to roof extraction [43]. For instance-level structure reconstruction, many methods resort to generative models [32] or detecting key points with fixed topologies (e.g., skeletons of human bodies and hands) [33, 46]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 353 + ], + "type": "text", + "content": "As for scene-level structural modeling, the reconstruction targets usually belong to special semantic-related structures, such as planar building roofs and indoor floorplans. For these tasks, the derived structures have to match semantic regions with robustness to other irrelevant structure information. For example, room layout estimation has to ensure that connected lines can form wall regions without extra lines from windows and doors. Related approaches can be roughly categorized into two kinds: (i) the local primitive based and (ii) the global information based. Pioneering local primitive based works derive structural representation from an image by post-processing heatmaps of corners and edges [18, 47]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 521 + ], + "type": "text", + "content": "With respect to the global information based methods, it is a popular choice to form the final structure graph with structure-related regions segmented from the input image [41, 28]. Different from the segmentation based methods, recent studies consider a holistic graph inference with geometric primitives as nodes for high-level planer extraction. For example, Conv-MPN [42] adopts a variant of the graph neural network to pass messages across the whole graph, and HEAT [2] learns the topological pattern of edges with an attention transformer. Beyond exploring the high-level information in structure recovery, we are interested in improving the efficiency of exploiting low-level geometric features, which serve as the foundation of accurate structure topology inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 527, + 247, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 247, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 247, + 541 + ], + "type": "text", + "content": "2.3. Frequency Analysis in Deep Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "Frequency analysis is a technique that provides a compressed representation of signals in the frequency domain. Not only commonly used in classical digital signal processing, but frequency analysis also remains powerful in the era of deep learning. [37] and [1] investigate the training and generalization of neural networks via frequency analysis. [35] analyzes the spectral bias of deep models regarding multiple vision parsing tasks. [8] applies frequency analysis to generated images for improving image synthesis quality, while [15] introduces a focal frequency loss that forces generative models to learn hard frequencies. [25] designs a frequency channel attention mechanism to compress channelwise information with scalars. [3] and [31] utilize frequency analysis in model compression. More applications of the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "frequency-domain representation can also be found in domain adaptation [12, 13] and position embedding [22, 29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 96, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 204 + ], + "type": "text", + "content": "Although frequency analysis has been combined with deep learning techniques in recent years, it is generally used as a mature tool in prior arts without a deep investigation of the reasons why frequency analysis is useful. To this end, we experimentally studied the behavior tendency of frequency-domain convolutions in learning holistic geometries. Based on the investigation results, we designed the F-Learn strategy to efficiently learn holistic geometry clues in low-level feature maps for structured reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 213, + 361, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 361, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 361, + 225 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 233, + 546, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 424 + ], + "type": "text", + "content": "Geometric features extracted at the early convolutional stage fundamentally support the structure reconstruction with precise geometry localization clues. Therefore, we are interested in addressing a key problem for learning reliable low-level geometric features: the low efficiency of fusing geometric fragments in shallow convolution layers. Considering that geometric information is generally high-frequency signals, we are motivated to study the problem mentioned before with frequency analysis for structure reconstruction. In the following of this section, we first present the preliminaries of frequency analysis. Next, we validate the high efficiency of learning in the frequency domain for holistic geometry recovery. Finally, we introduce our frequency-domain feature learning strategy (F-Learn) and the application of the F-Learn in a given structure reconstruction model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 431, + 499, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 499, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 499, + 445 + ], + "type": "text", + "content": "3.1. Preliminaries of Frequency Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": "The discrete Fourier transform (DFT) is necessary for analyzing image data in the frequency domain. With " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "I(x,y)" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": " denoting the color signal at the spatial position " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": " of an " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "M\\times N" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": " image, the DFT converts the information of " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": " to a frequency map " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": " in the frequency domain by Eq. (1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 516, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 516, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 322, + 516, + 545, + 550 + ], + "type": "interline_equation", + "content": "F (u, v) = \\sum_ {x = 0} ^ {M - 1} \\sum_ {y = 0} ^ {N - 1} I (x, y) e ^ {- j 2 \\pi (u x / M + v y / N)}. \\tag {1}", + "image_path": "0af8f511f2591f4cc2023a148e5b2821cc10cc88c26f04c825078c6e4310906d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": "In Eq. (1), " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "F(u, v)" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " is a value that encodes a two-dimensional signal changing that belongs to the frequency determined by " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 556, + 545, + 665 + ], + "type": "text", + "content": ". The frequency map excels in holistically representing signals that show a similar frequency pattern. Therefore, we assume that the DFT is a powerful tool for efficiently learning structure-related geometry features. To validate this hypothesis, we conducted experiments with a geometry recovery task, and analyze the results quantitatively and qualitatively as below." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 463, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 463, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 463, + 684 + ], + "type": "text", + "content": "3.2. Learning Efficiency Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In this section, we study the learning efficiency of fusing geometric fragments in the space and frequency domains" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21809" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "with a simple geometry recovery task. We first decomposed a geometric structure composed of a circle and a square into several geometric fragments as illustrated in Fig. 2 (a), and left overlapping between neighboring fragments to avoid trial solutions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "text", + "content": "With the purpose to recover the binary image of the original structure with the geometric fragments, we studied the learning efficiency of frequency-domain and space-domain convolution models. We constructed a frequency-domain learning model (F-Learn) and several space-domain convolution models. The baseline was set as a space-domain convolution module (BConv) built with two " + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "text", + "content": " convolution layers and one " + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "text", + "content": " layer in between. The F-Learn model consists of three key components: (1) a DFT; (2) two parallel convolution modules same with the baseline for separately processing the real and imaginary parts; and (3) an inverse DFT (IDFT) followed by magnitude computation of complex values. For a fair comparison, we also combined a pair of baseline modules into two kinds of space-domain learning models with parallel and cascade structures, namely Conv-Casv1 and Conv-Parv1. Extra " + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "text", + "content": " convolutions were also added into the cascade and parallel space-domain models to simulate the function of DFT and IDFT. The modified models were named Conv-Casv2 and Conv-Parv2 for simplicity. For all models, the channel number of the intermediate feature maps was 64, and a " + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 133, + 288, + 408 + ], + "type": "text", + "content": " convolution was used as the final binary classifier. Detailed model settings are depicted in Fig. 2(b)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "type": "text", + "content": "The experiments were conducted on " + }, + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 410, + 287, + 553 + ], + "type": "text", + "content": " binary images of the geometric structure and fragments. The training of all models was completed with an Adam optimizer. The learning rate and training period were set to 0.1 and 50 epochs, respectively. The binary cross entropy loss (BCELoss) was chosen as the training guide. To alleviate the influence of randomness, we generated 100 seeds to set 100 trials, and all models were trained with the same seed in every trial. The performance of all models was evaluated with an F1-score measurement at a threshold of 0.5. Quantitative and qualitative results are shown in Fig. 3 and Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 603 + ], + "type": "text", + "content": "Fig. 3 draws the averaged F1-score curves of 100 trials. Compared with the space-domain counterparts, the frequency-domain model (F-Learn) shows superiority in recovering the complete structure in a more efficient manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 604, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 653 + ], + "type": "text", + "content": "The visualized results in Fig. 4 also show that the frequency-domain learning model exceeds the basic spatial convolution model with faster convergence speed and better structure recovery quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Fig. 4 reveals that the F-Learn model easily rebuilds the original structure, while the space-domain methods encounter difficulties in efficiently reconstructing the circle. We also find the difference in learning tendency between the F-Learn and space-domain models. The F-Learn model pri" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 92, + 541, + 149 + ], + "blocks": [ + { + "bbox": [ + 362, + 73, + 491, + 85 + ], + "lines": [ + { + "bbox": [ + 362, + 73, + 491, + 85 + ], + "spans": [ + { + "bbox": [ + 362, + 73, + 491, + 85 + ], + "type": "text", + "content": "(a) Geometry Recovery Task" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 92, + 541, + 149 + ], + "lines": [ + { + "bbox": [ + 310, + 92, + 541, + 149 + ], + "spans": [ + { + "bbox": [ + 310, + 92, + 541, + 149 + ], + "type": "image", + "image_path": "dd3c7f7e769a96ad8d12190a426b4d23b5dadc0d71229d0a00c72ff1a291c2b2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 160, + 477, + 173 + ], + "lines": [ + { + "bbox": [ + 375, + 160, + 477, + 173 + ], + "spans": [ + { + "bbox": [ + 375, + 160, + 477, + 173 + ], + "type": "text", + "content": "(b) The Model Settings" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 314, + 184, + 541, + 336 + ], + "blocks": [ + { + "bbox": [ + 314, + 184, + 541, + 336 + ], + "lines": [ + { + "bbox": [ + 314, + 184, + 541, + 336 + ], + "spans": [ + { + "bbox": [ + 314, + 184, + 541, + 336 + ], + "type": "image", + "image_path": "1e0c7bc7a7d8c30fbf478c5db138fb44e7b5ca7b9d8c61e6bb5ba0467091fb25.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 372 + ], + "type": "text", + "content": "Figure 2. The settings of the simple geometry recovery task. (a) The task of recovering the holistic structure with geometric fragments. (b) The detailed settings of compared models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 384, + 542, + 525 + ], + "blocks": [ + { + "bbox": [ + 309, + 384, + 542, + 525 + ], + "lines": [ + { + "bbox": [ + 309, + 384, + 542, + 525 + ], + "spans": [ + { + "bbox": [ + 309, + 384, + 542, + 525 + ], + "type": "image", + "image_path": "eeabac41c684490d02950e54d9c8ca26e382dac3578a1b5d6ed917ca292fc84b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 526, + 545, + 549 + ], + "lines": [ + { + "bbox": [ + 305, + 526, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 545, + 549 + ], + "type": "text", + "content": "Figure 3. The averaged F1-score curves of 100 trials for the F-Learn model and the compared space-domain methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "oritizes holistic integration before further enlarging the gap between the foreground and background. In contrast, the space-domain models first generate high responses to local areas that have strong signal intensity in the input. Hence, the space-domain models prefer overlapped areas without considering structural continuity. In practice, the holistic learning pattern of F-Learn will benefit the topology inference in two ways: (i) holistic geometry clues are ready for learning high-level features at the early stage, which alleviates the burden on learning structure-related low-level features with far supervision signal; (ii) the holistic geometry learning excels in preserving the connection in-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21810" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 230 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 230 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 230 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 230 + ], + "type": "image", + "image_path": "6c8c685bca19215ac7a484ae97450d8ba69eea8563ed213f38cb9939ef79f28e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 232, + 286, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 232, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 286, + 255 + ], + "type": "text", + "content": "Figure 4. The visualized results of F-Learn, Bconv and Cov-Parv1 at different epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 286, + 310 + ], + "type": "text", + "content": "formation in the high-resolution low-level feature maps, which are often reused to provide accurate structure localization for topology inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 286, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 286, + 381 + ], + "type": "text", + "content": "We attribute the advantages of the frequency domain based model to its ability to holistically process signals that belong to similar frequencies. Based on the findings above, we designed a frequency-domain feature learning strategy to generate holistic geometric clues in the low-level feature maps for structured reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 389, + 246, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 246, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 246, + 403 + ], + "type": "text", + "content": "3.3. Frequency-domain Learning Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 408, + 286, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 286, + 528 + ], + "type": "text", + "content": "The proposed frequency-domain learning strategy (Flearn) inherits the architecture shown in Fig. 2. As the FLearn strategy will face more complex data and larger batch size in practice than in Sec. 3.2, we replace each convolution layer between DFT and IDFT with a combination (CBR) of a convolution layer, a batch normalization operation, and a relu activation function. We refer to all operations between DFT and IDFT collectively as a CBR group (CBRG). Given a set of geometric feature maps " + }, + { + "bbox": [ + 46, + 408, + 286, + 528 + ], + "type": "inline_equation", + "content": "\\{f_i|i = 0,\\dots ,N\\}" + }, + { + "bbox": [ + 46, + 408, + 286, + 528 + ], + "type": "text", + "content": ", the FLearn strategy is formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 537, + 211, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 537, + 211, + 551 + ], + "spans": [ + { + "bbox": [ + 118, + 537, + 211, + 551 + ], + "type": "interline_equation", + "content": "\\left(F _ {i} ^ {\\text {r e}}, F _ {i} ^ {\\text {i m}}\\right) = \\operatorname {D F T} \\left(f _ {i}\\right);", + "image_path": "c317bc38adda9ae1838ddba51ab659856c1cf66f9e6ef733f20c0ddf516809d6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 553, + 287, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 553, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 66, + 553, + 287, + 568 + ], + "type": "interline_equation", + "content": "\\{\\hat {F} _ {i} ^ {\\mathrm {r e}} \\}, \\{\\hat {F} _ {i} ^ {\\mathrm {i m}} \\} = \\mathrm {C B R G} (\\{F _ {i} ^ {\\mathrm {r e}} \\}), \\mathrm {C B R G} (\\{F _ {i} ^ {\\mathrm {i m}} \\}); \\tag {2}", + "image_path": "b9e14e3c65f6b15d14132074b71c6eed0549f90cab3204f12dfccdad33b614a7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 571, + 238, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 571, + 238, + 585 + ], + "spans": [ + { + "bbox": [ + 115, + 571, + 238, + 585 + ], + "type": "interline_equation", + "content": "\\{\\hat {f} _ {i} \\} = | \\mathrm {I D F T} (\\{\\hat {F} _ {i} ^ {\\mathrm {r e}} \\}, \\{\\hat {F} _ {i} ^ {\\mathrm {i m}} \\}) |.", + "image_path": "d0dd1710dad1e1f9a158cab7eac694e1a17c1d9d6525050657b8640240df2a15.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "The F-learn strategy first converts every " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " into a pair of maps that record the real and imaginary parts of the DFT, i.e., " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathrm{re}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathrm{im}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ". The maps of the real part " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{F_{i}^{\\mathrm{re}}\\}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " are subsequently processed by a CBRG that contains two " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " layers with a " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " one in between, and the same operation is also applied to " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{F_{i}^{\\mathrm{im}}\\}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ". In the frequency domain, the " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " convolution directly fuses the information of the same frequency across the channel dimension, while the " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " counterpart enhances features with signals that have similar frequencies. The high-frequency geometric fragments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "text", + "content": "in " + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "inline_equation", + "content": "\\{f_i\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "text", + "content": " are easily combined and enhanced in such a holistic way. The structurally informative space-domain features " + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_i\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "text", + "content": " can be finally obtained through an IDFT along with magnitude computation " + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "text", + "content": ". With the parsimonious design, the F-learn strategy is readily inserted into a hierarchical convolutional backbone for holistic geometry feature learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "text", + "content": "For feature propagation, the " + }, + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\{f_i\\}" + }, + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_i\\}" + }, + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "text", + "content": " are fused into " + }, + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\{\\tilde{f}_i\\}" + }, + { + "bbox": [ + 306, + 156, + 545, + 180 + ], + "type": "text", + "content": " via" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 186, + 545, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 186, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 348, + 186, + 545, + 201 + ], + "type": "interline_equation", + "content": "\\left\\{\\tilde {f} _ {i} \\right\\} = \\mathrm {C} _ {1 \\times 1} \\mathrm {B R} \\left(\\operatorname {C o n c a t} \\left(\\left\\{f _ {i} \\right\\}, \\left\\{\\hat {f} _ {i} \\right\\}\\right)\\right), \\tag {3}", + "image_path": "dadcbda7399463a6f06d2f46d8b1a4e549800778ae26731f2ca7417bb62b6d22.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}" + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{1\\times 1}\\mathbf{BR}" + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "text", + "content": " denote the concatenation operation and a " + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 305, + 208, + 545, + 231 + ], + "type": "text", + "content": " convolution based CBR, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 232, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 278 + ], + "type": "text", + "content": "The enhanced geometric features " + }, + { + "bbox": [ + 305, + 232, + 545, + 278 + ], + "type": "inline_equation", + "content": "\\{\\tilde{f}_i\\}" + }, + { + "bbox": [ + 305, + 232, + 545, + 278 + ], + "type": "text", + "content": " are then fed into the next convolution stage for higher-level feature learning and used to offer precise geometric localization for later structure inference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 279, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 279, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 279, + 545, + 374 + ], + "type": "text", + "content": "The F-Learn strategy can be easily inserted into a convolution base backbone. With the commonly-used ResNet50 backbone [10] as an example, the F-Learn strategy can be directly deployed after the first convolution to holistically learn geometric features. In experiments, we implemented our F-Learn strategy in a state-of-the-art approach, i.e., the holistic edge attention transformer (HEAT) [2], for structured reconstruction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 384, + 386, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 386, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 386, + 398 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 404, + 423, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 404, + 423, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 404, + 423, + 417 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 423, + 371, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 371, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 371, + 434 + ], + "type": "text", + "content": "4.1.1 Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 441, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 525 + ], + "type": "text", + "content": "We tested our method on a dataset introduced by [24] for vectorizing world buildings. This dataset is built on the SpaceNet dataset [30] with 2001 images annotated with roof planar graphs. Each image contains a building instance, and the image is processed into a size of " + }, + { + "bbox": [ + 305, + 441, + 545, + 525 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 305, + 441, + 545, + 525 + ], + "type": "text", + "content": ". In line with the prior arts [2], the training/validation/testing split is set to 1601/50/350." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 540, + 439, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 439, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 439, + 551 + ], + "type": "text", + "content": "4.1.2 Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "content": "We evaluated the F-learn strategy with the HEAT model for structured reconstruction. The HEAT model was built on a ResNet50 backbone pretrained on ImageNet [4]. We performed our experiments with Pytorch [27] in Python3.7 and used a workstation with one NVIDIA RTX 3090 GPU. We adopted the same settings of loss functions as the HEAT model. The training of the model was completed with an AdamW optimizer, and the training period lasts 800 epochs. The learning rate was initialized as " + }, + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "inline_equation", + "content": "2e - 4" + }, + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "content": " and multiplied by a factor of 0.1 for the last " + }, + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "content": " epoch. In line with previous works [2, 24], we used precision, recall, and F1 score to evaluate the quality of structure reconstruction in terms of corner detection, edge inference, and region recovery." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21811" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 106, + 523, + 250 + ], + "blocks": [ + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 69, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between the F-Learn strategy and other methods in terms of corners detection, edge extraction, and region reconstruction. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 106, + 523, + 250 + ], + "lines": [ + { + "bbox": [ + 67, + 106, + 523, + 250 + ], + "spans": [ + { + "bbox": [ + 67, + 106, + 523, + 250 + ], + "type": "table", + "html": "
Evaluation Type → \nMethodFully-neuralJointCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
IP [24]----74.5--53.1--55.7
Exp-Cls [43]--92.275.983.275.460.467.174.954.763.5
ConvMPN [42]✓-78.079.778.857.059.758.152.456.554.4
HAWP [39]✓✓90.981.285.776.668.172.174.155.463.4
LETR [36]✓✓87.874.880.859.758.659.168.348.756.8
HEAT [2]✓✓91.783.087.180.672.376.276.465.670.6
HEAT(retrain) [2]✓✓91.683.087.180.372.476.175.565.370.0
F-Learn (Ours)✓✓93.284.488.683.675.079.179.568.173.4
", + "image_path": "4448dbec1fa095cef36ab1002c0f6eca899b2315d7d148296f8136a9d078ae75.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 269, + 168, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 168, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 168, + 282 + ], + "type": "text", + "content": "4.1.3 Competing Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 326 + ], + "type": "text", + "content": "We compare the F-Learn strategy with six methods: HEAT [2], ConvMPN [42], IP [24], Exp-cls [43], HAWP [39] and LETR [36]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 326, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 434 + ], + "type": "text", + "content": "HEAT is an attention-based method that takes a 2D raster image as an input and reconstructs a planar graph in an end-to-end manner. HEAT works via three steps: 1) extracting hierarchical features by ResNet50; 2) detecting corners with multi-scale features enriched by a deformable attention module; and 3) employing two weight-sharing Transformer decoders to classify edges and reason structures with detected corners. Our F-Learn strategy is adopted in the first step for learning structure-related features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 555 + ], + "type": "text", + "content": "Conv-MPN uses a graph neural network to infer edges with a pre-trained corner detector. IP firstly detects geometric primitives and then reconstructs a planar graph through several post-processing steps. Exp-cls is based on the geometric primitives produced by other methods (e.g., IP and Conv-MPN) and reconstructs a planar graph through an explore-and-classify framework. LCNN and HAWP are methods specifically proposed for wireframe parsing. LETR is a transformer-based method that directly generates lines without post-processing and heuristic guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 564, + 191, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 191, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 191, + 578 + ], + "type": "text", + "content": "4.2. Comparison and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 584, + 173, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 173, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 173, + 597 + ], + "type": "text", + "content": "4.2.1 Quantitative Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": "The quantitative results displayed in Tab. 1 show that our F-Learn strategy advances roof structure reconstruction with state-of-the-art performance. The F-Learn strategy edges out other compared methods by significant performance gains regarding corner detection, edge inference, and region reconstruction under all metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "With respect to the detection of corners and edges, the F-Learn strategy outperforms the second-best by at least " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " in the precision measurement, while " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 268, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 413 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 268, + 545, + 413 + ], + "type": "text", + "content": " under the recall metric. The improvement in extracting low-level geometric primitives indicates the outstanding ability of the F-Learn strategy for providing holistic and accurate geometric clues. The F-Learn strategy also brings at least a " + }, + { + "bbox": [ + 304, + 268, + 545, + 413 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 268, + 545, + 413 + ], + "type": "text", + "content": " gain in F1 score in terms of region reconstruction, which reveals the assistance of the F-Learn strategy in inferring correct topological relations. Besides, it is worth noticing that the F-Learn strategy significantly outperforms the original HEAT model under all metrics, but only brings a little computation increase compared to HEAT. Tab. 2 presents a detailed comparison of computation requirements." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 495, + 539, + 544 + ], + "blocks": [ + { + "bbox": [ + 304, + 422, + 547, + 489 + ], + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 489 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of the change of computing efficiency brought by F-Learn. Param. and FLOPs are the abbreviations of the parameter quantity and floating point operations, used to measure the complexity of an algorithm/model. Time per Image represents the time it takes the model to infer an image and is used to evaluate the calculation speed of the model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 495, + 539, + 544 + ], + "lines": [ + { + "bbox": [ + 307, + 495, + 539, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 495, + 539, + 544 + ], + "type": "table", + "html": "
Param.FLOPsTime per Image
w/o F-Learn68.747M670.919G0.10s
w/ F-Learn68.847M684.039G0.11s
", + "image_path": "711ab14b8c40dcbc93953df1575009e9d0f95288e49d78a5e59831ad4cf09a0b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": "In Tab. 2, the computation increase of F-Learn strategy in terms of parameters, FLOPs, and inference time per image can be clearly seen, where the amount is little. Especially, the parameter increase brought by F-Learn strategy is only " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "0.145\\%" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": ", which is almost negligible. The results further demonstrates that the F-Learn strategy is simple yet effective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 657, + 425, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 425, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 425, + 670 + ], + "type": "text", + "content": "4.2.2 Qualitative Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "For perceptual comparison, we also present the visualized reconstruction results in Fig. 5. Compared to other methods, our F-Learn is capable of detecting low-level geometric" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21812" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 77, + 541, + 657 + ], + "blocks": [ + { + "bbox": [ + 49, + 77, + 541, + 657 + ], + "lines": [ + { + "bbox": [ + 49, + 77, + 541, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 77, + 541, + 657 + ], + "type": "image", + "image_path": "079dc50752f9b4a8e44be9b0762e6f2bf5577cb93537c68ca7246206f1013004.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 663, + 546, + 696 + ], + "lines": [ + { + "bbox": [ + 46, + 663, + 546, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 546, + 696 + ], + "type": "text", + "content": "Figure 5. Qualitative results on outdoor roof structure reconstruction. Other methods build wrong topological relations due to redundant or missing geometric primitives. Our F-Learn strategy derives the correct topological principles with corners and edges more fitted to the holistic structures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21813" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 271, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 271, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 271, + 83 + ], + "type": "text", + "content": "primitives that are consistent with the holistic structure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 288, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 288, + 276 + ], + "type": "text", + "content": "When reconstructing the roof shown in the first row, ConvMPN and Exp-cls tend to detect redundant corners due to the shadow inference and therefore wrongly separate the roof regions. The original HEAT outputs a non-closed roof structure because key edges are missing. In contrast, the corners and edges learned with the F-Learn strategy are better fitted into the overall structure. This is because the F-Learn strategy can offer low-level features that are rich in holistic geometric clues for topology-preserved roof structure reconstruction. The results in the following rows additionally validate that the F-Learn strategy can effectively extract geometric primitives that are easily ignored by other methods, especially the original HEAT model. By comparing the inferred planar graphs in the fifth row, one can also find that the F-Learn improves the HEAT model with robustness to irrelevant structure information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 285, + 148, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 148, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 148, + 297 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 304, + 257, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 257, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 257, + 316 + ], + "type": "text", + "content": "4.3.1 Comparison with Space-domain Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "text", + "content": "In addition to the exploration with a simple geometry fusion task in Sec. 3.2, we further study the F-Learn strategy with the space-domain learning strategies in the real scene for roof structure reconstruction. Similar to Sec. 3.2, we construct a space-domain learning baseline (BConv) with a basic module composed of two " + }, + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "text", + "content": " convolution layers and one " + }, + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 324, + 287, + 479 + ], + "type": "text", + "content": " layer in between. Furthermore, we design two more space-domain strategies with the basic modules arranged in parallel and cascade, denoted as Conv-Parv. and Conv-Casv., respectively. Same with the F-Learn, additional operations of batch normalization and ReLU activation are added after each convolution layer used in every space-domain strategy." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 65, + 562, + 266, + 654 + ], + "blocks": [ + { + "bbox": [ + 46, + 490, + 288, + 556 + ], + "lines": [ + { + "bbox": [ + 46, + 490, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 288, + 556 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons with space-domain learning methods. The F-Learn strategy is compared with the space-domain counterparts under the F1-score measurement in terms of corner detection, edge inference, and region reconstruction. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 562, + 266, + 654 + ], + "lines": [ + { + "bbox": [ + 65, + 562, + 266, + 654 + ], + "spans": [ + { + "bbox": [ + 65, + 562, + 266, + 654 + ], + "type": "table", + "html": "
MethodCornerEdgeRegion
HEAT87.176.170.0
HEAT-BConv87.476.270.7
HEAT-Conv-Cas.88.177.170.2
HEAT-Conv-Par.87.776.869.5
F-Learn88.679.173.4
", + "image_path": "469e7380f27f2bd7fc9eda8adbf5971846bf45ec19effb16677996ac6e830573.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "With the F1-score measurement, Tab. 3 presents the performance of the F-learn and space-domain learning strategies in detecting geometric primitives. Although some space-domain methods are comparable to the proposed F" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "Learn strategy in corner extraction, it can be seen that the F-Learn strategy gains over these methods in detecting edges and regions by at least " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", respectively. Because edges and regions are primitives closely related to roof topology, this phenomenon demonstrates the holistic geometry learned by the F-Learn is beneficial for the topology inference for roof reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 178, + 497, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 178, + 497, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 497, + 190 + ], + "type": "text", + "content": "4.3.2 F-Learn Strategy at Different Layers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "text", + "content": "We investigate the effect of applying the F-Learn strategy to feature maps of different levels with the HEAT model and the hierarchical ResNet50 backbone. The F-learn strategy is deployed to process feature maps generated at the first convolution layer, the first residual learning stage, and the second one. The sizes of the corresponding feature maps are " + }, + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 198, + 545, + 295 + ], + "type": "text", + "content": ", respectively. The quantitative results are presented in Tab. 4." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 390, + 538, + 464 + ], + "blocks": [ + { + "bbox": [ + 304, + 306, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 384 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of applying the F-Learn strategy to features at different levels. L0: the first convolution layer; L1: the first residual learning stage; L2: the second residual learning stage. Size: the feature map size. The F1-score metric is used for evaluating the performance of corner detection, edge inference, and region recovery. Higher scores mean better results. The bold values represent the best performance. (unit:%)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 390, + 538, + 464 + ], + "lines": [ + { + "bbox": [ + 309, + 390, + 538, + 464 + ], + "spans": [ + { + "bbox": [ + 309, + 390, + 538, + 464 + ], + "type": "table", + "html": "
MethodSizeCornerEdgeRegion
HEAT-87.176.170.0
F-Learn-L0128 × 12888.679.173.4
F-Learn-L164 × 6487.776.470.9
F-Learn-L232 × 3286.775.369.0
", + "image_path": "1a51f31154f28b9f9a0d520ed00072512d331846508cd8ed71b2fa47af85abbc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 479, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 586 + ], + "type": "text", + "content": "Tab. 4 implies that the higher the resolution is, the better the F-Learn strategy works. It is because high-resolution low-level feature maps contain more high-frequency geometric clues, while the low-resolution high-level ones are richer in abstract semantic information but poorer in geometric details. Therefore, the high-resolution feature maps are more suitable than the low-resolution counterparts for the F-Learn strategy to learn structurally informative geometries in roof reconstruction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 609, + 518, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 518, + 620 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 518, + 620 + ], + "type": "text", + "content": "4.3.3 F-Learn Strategy with Various Backbones" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of the F-learn strategy, we evaluate the performance improvement brought by the F-Learn strategy with different backbones. The quantitative results are presented in Tab. 5. As shown by Tab. 5, with the deployment of the F-learn strategy, the performance of HEAT is consistently improved in terms of corners detection, edge extraction, and region reconstruction." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21814" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 108, + 487, + 233 + ], + "blocks": [ + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "lines": [ + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of F-Learn strategy with HEAT built on different backbones, including ResNet18, ResNet34, and ResNet50. Prec and F1 are the abbreviations of the precision and f1-score metrics. The higher the scores are, the better the performance is. The best results are marked bold. (unit:%)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 108, + 487, + 233 + ], + "lines": [ + { + "bbox": [ + 102, + 108, + 487, + 233 + ], + "spans": [ + { + "bbox": [ + 102, + 108, + 487, + 233 + ], + "type": "table", + "html": "
BackboneCornerEdgeRegion
PrecRecallF1PrecRecallF1PrecRecallF1
w/o F-LearnResNet1891.681.986.579.970.174.776.361.668.2
w/ F-LearnResNet1891.782.286.780.870.475.277.162.969.3
w/o F-LearnResNet3490.582.686.478.371.174.573.463.468.0
w/ F-LearnResNet3492.183.287.480.271.875.876.264.169.6
w/o F-LearnResNet5091.683.087.180.372.476.175.565.370.0
w/ F-LearnResNet5093.284.488.683.675.079.179.568.173.4
", + "image_path": "003692ed79ba1062e10fcc25d1de26f89bd631d70f17203b7692f9f24c870666.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 252, + 250, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 250, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 250, + 264 + ], + "type": "text", + "content": "4.3.4 F-Learn Strategy in Topology Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 288, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 288, + 368 + ], + "type": "text", + "content": "In this section, we further explore the effectiveness of the F-Learn strategy in topology inference. Edges and regions are primitives that highly relate to the inference of topological relationships, and these primitives are learned with corners and low-level features in the HEAT model. Therefore, we directly use the ground-truth corner map to focus on topology inference. We compare our F-Learn strategy with the space-domain methods used in Sec. 4.3.1." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 446, + 286, + 540 + ], + "blocks": [ + { + "bbox": [ + 46, + 375, + 288, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 288, + 441 + ], + "type": "text", + "content": "Table 6. Quantitative results of the F-Learn and space-domain strategies with respect to topology inference. The HEAT method with corner ground truth is set as the baseline. With corner annotations, all methods are studied with a focus on topology inference. The higher scores mean better results, and the best performance is bolded. (unit: " + }, + { + "bbox": [ + 46, + 375, + 288, + 441 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 46, + 375, + 288, + 441 + ], + "type": "text", + "content": " )" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 446, + 286, + 540 + ], + "lines": [ + { + "bbox": [ + 50, + 446, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 50, + 446, + 286, + 540 + ], + "type": "table", + "html": "
Evaluation Type → MethodEdgeRegion
PrecRecallF1PrecRecallF1
HEAT93.584.188.590.273.881.2
HEAT-BConv94.483.988.889.471.779.6
HEAT-Conv-Cas94.885.389.891.373.781.6
HEAT-Conv-Par93.884.889.189.773.981.0
F-Learn95.187.591.189.577.683.1
", + "image_path": "fdbac89246cc86d99e6b63f4e6a23208fe05702dba899ff502d583d2b601042f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": "Tab. 6 presents the numeric results, and it can be seen that our F-Learn strategy brings the highest gains in most metrics for edge inference and region reconstruction. This phenomenon demonstrates that the F-Learn strategy effectively supports topology inference with an outstanding ability for holistic geometry learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 125, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 125, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 125, + 646 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "In this paper, we present a frequency-domain feature learning strategy (F-Learn) to tackle the issue of wrong topology recovery caused by the lack of holistic clues in low-level features. Experiments with a geometry recovery task convincingly validate the efficiency of the F-Learn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "content": "strategy in learning holistic geometry. In terms of the real scene, the F-Learn strategy achieves significant performance improvement of topological principle inference for roof structure reconstruction. The ablation studies verify that the F-Learn strategy outperforms the space-domain learning counterparts in capturing holistic geometric features regarding complex real scenes. We believe that it is promising to further explore the holistic learning ability brought by frequency analysis in more vision tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 370, + 416, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 416, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 416, + 384 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 461 + ], + "type": "text", + "content": "This research is supported by NSFC projects under Grant 42071370, the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001, and the Open fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources under Grant KF202106084." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 472, + 365, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 365, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 365, + 483 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 491, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 310, + 491, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 491, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 310, + 491, + 545, + 545 + ], + "type": "text", + "content": "[1] Ronen Basri, Meirav Galun, Amnon Geifman, David Jacobs, Yoni Kasten, and Shira Kritchman. Frequency bias in neural networks for input of non-uniform density. In International Conference on Machine Learning, pages 685-694. PMLR, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 548, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 548, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 311, + 548, + 547, + 590 + ], + "type": "text", + "content": "[2] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In Proceedings of CVPR, pages 3866-3875, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 592, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 592, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 311, + 592, + 547, + 646 + ], + "type": "text", + "content": "[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Q Weinberger, and Yixin Chen. Compressing convolutional neural networks in the frequency domain. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1475-1484, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 647, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 647, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 311, + 647, + 547, + 680 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 681, + 547, + 714 + ], + "type": "text", + "content": "[5] Ruoxi Deng, Chunhua Shen, Shengjun Liu, Huibing Wang, and Xinru Liu. Learning to predict crisp boundaries. In ECCV, pages 570-586, 2018. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21815" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "text", + "content": "[6] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 287, + 161 + ], + "type": "text", + "content": "[7] Richard O Duda and Peter E Hart. Use of the hough transformation to detect lines and curves in pictures. Communications of the ACM, 15(1):11-15, 1972. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 206 + ], + "type": "text", + "content": "[8] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 208, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 287, + 239 + ], + "type": "text", + "content": "[9] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the CVPR, pages 770-778, 2016. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[11] Linxi Huan, Nan Xue, Xianwei Zheng, Wei He, Jianya Gong, and Gui-Song Xia. Unmixing convolutional features for crisp edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6602-6609, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 374 + ], + "type": "text", + "content": "[12] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 421 + ], + "type": "text", + "content": "[13] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Rda: Robust domain adaptation via fourier adversarial attacking. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8988-8999, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 476 + ], + "type": "text", + "content": "[14] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 521 + ], + "type": "text", + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[16] Varsha Kamat-Sadekar and Subramaniam Ganesan. Complete description of multiple line segments using the hough transform. Image and Vision Computing, 16(9-10):597-613, 1998. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 601 + ], + "type": "text", + "content": "[17] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[18] Chen-Yu Lee, Vijay Badrinarayanan, Tomasz Malisiewicz, and Andrew Rabinovich. Roomnet: End-to-end room layout estimation. In Proceedings of the ICCV, pages 4865-4874, 2017. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[19] Hongjie Li, Xianwei Zheng, Mingyue Dong, Gui-Song Xia, and Hanjiang Xiong. Locally nonlinear affine verification for multisensor image matching. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[20] Yancong Lin, Silvia L Pintea, and Jan C van Gemert. Deep though-transform line priors. In Computer Vision-ECCV" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 323-340. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[21] Yun Liu, Ming-Ming Cheng, Xiaowei Hu, Jia-Wang Bian, Le Zhang, Xiang Bai, and Jinhui Tang. Richer convolutional features for edge detection. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1939-1946, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[22] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "text", + "content": "[23] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229. PMLR, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[24] Nelson Nauata and Yasutaka Furukawa. Vectorizing world buildings: Planar graph reconstruction by primitive detection and relationship inference. In ECCV 2020, pages 711-726. Springer, 2020. 1, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 547, + 319 + ], + "type": "text", + "content": "[25] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 783-792, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 321, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 547, + 354 + ], + "type": "text", + "content": "[26] Lawrence G Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 355, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 410 + ], + "type": "text", + "content": "[27] Benoit Steiner, Zachary Devito, Soumith Chintala, Sam Gross, Adam Paszke, Francisco Massa, Adam Lerer, Gregory Chanan, Zeming Lin, Edward Yang, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "text", + "content": "[28] Sinisa Stekovic, Mahdi Rad, Friedrich Fraundorfer, and Vincent Lepetit. Montefloor: Extending mcts for reconstructing accurate large-scale floor plans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16034-16043, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 467, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 532 + ], + "type": "text", + "content": "[29] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 566 + ], + "type": "text", + "content": "[30] Adam Van Etten, Dave Lindenbaum, and Todd M Bacastow. Spacenet: A remote sensing dataset and challenge series. arXiv preprint arXiv:1807.01232, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 567, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 611 + ], + "type": "text", + "content": "[31] Yunhe Wang, Chang Xu, Chao Xu, and Dacheng Tao. Packing convolutional neural networks in the frequency domain. IEEE transactions on pattern analysis and machine intelligence, 41(10):2495-2510, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 612, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 645 + ], + "type": "text", + "content": "[32] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the ICCV, pages 6772–6782, 2021. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[33] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the ECCV, pages 466-481, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[34] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. International Journal of Computer Vision, 125(1):3-18, 2017. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21816" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 676 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[35] Kai Xu, Minghai Qin, Fei Sun, Yuhao Wang, Yen-Kuang Chen, and Fengbo Ren. Learning in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1740-1749, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[36] Yifan Xu, Weijian Xu, David Cheung, and Zhuowen Tu. Line segment detection using transformers without edges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4257-4266, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "text", + "content": "[37] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 287, + 262 + ], + "type": "text", + "content": "[38] Nan Xue, Song Bai, Fudong Wang, Gui-Song Xia, Tianfu Wu, and Liangpei Zhang. Learning attraction field representation for robust line segment detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1595-1603, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 287, + 319 + ], + "type": "text", + "content": "[39] Nan Xue, Tianfu Wu, Song Bai, Fudong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2788-2797, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 319, + 287, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 287, + 362 + ], + "type": "text", + "content": "[40] Jimei Yang, Brian Price, Scott Cohen, Honglak Lee, and Ming-Hsuan Yang. Object contour detection with a fully convolutional encoder-decoder network. In CVPR, pages 193-202, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 364, + 287, + 407 + ], + "type": "text", + "content": "[41] Zhiliang Zeng, Xianzhi Li, Ying Kin Yu, and Chi-Wing Fu. Deep floor plan recognition using a multi-task network with room-boundary-guided attention. In Proceedings of the ICCV, pages 9096-9104, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 409, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 409, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 409, + 287, + 464 + ], + "type": "text", + "content": "[42] Fuyang Zhang, Nelson Nauata, and Yasutaka Furukawa. Conv-mpn: Convolutional message passing neural network for structured outdoor architecture reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2798-2807, 2020. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 465, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 465, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 49, + 465, + 287, + 508 + ], + "type": "text", + "content": "[43] Fuyang Zhang, Xiang Xu, Nelson Nauata, and Yasutaka Furukawa. Structured outdoor architecture reconstruction by exploration and classification. In Proceedings of ICCV, pages 12427-12435, 2021. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 510, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 510, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 49, + 510, + 287, + 563 + ], + "type": "text", + "content": "[44] Xianwei Zheng, Zhuang Yuan, Zhen Dong, Mingyue Dong, Jianya Gong, and Hanjiang Xiong. Smoothly varying projective transformation for line segment matching. ISPRS Journal of Photogrammetry and Remote Sensing, 183:129-146, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 566, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 566, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 49, + 566, + 287, + 598 + ], + "type": "text", + "content": "[45] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 962-971, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 600, + 287, + 632 + ], + "type": "text", + "content": "[46] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the ICCV, pages 4903-4911, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 634, + 287, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 634, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 49, + 634, + 287, + 676 + ], + "type": "text", + "content": "[47] Chuhang Zou, Alex Colburn, Qi Shan, and Derek Hoiem. Layoutnet: Reconstructing the 3d room layout from a single rgb image. In Proceedings of the CVPR, pages 2051-2059, 2018. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21817" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_content_list.json b/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..938fd109afc6b418530d71bcad40e5e801ce83c8 --- /dev/null +++ b/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_content_list.json @@ -0,0 +1,1405 @@ +[ + { + "type": "text", + "text": "Holistic Label Correction for Noisy Multi-Label Classification", + "text_level": 1, + "bbox": [ + 171, + 130, + 797, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaobo Xia $^{1}$ , Jiankang Deng $^{2}$ , Wei Bao $^{1}$ , xuan Du $^{3}$ , Bo Han $^{4}$ , Shiguang Shan $^{5,6}$ , Tongliang Liu $^{1*}$", + "bbox": [ + 285, + 179, + 717, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Sydney 2Imperial College London", + "bbox": [ + 263, + 215, + 705, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ JD Explore Academy $^{4}$ Hong Kong Baptist University", + "bbox": [ + 261, + 233, + 709, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Chinese Academy of Sciences $^{6}$ University of Chinese Academy of Sciences", + "bbox": [ + 173, + 251, + 797, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-label classification aims to learn classification models from instances associated with multiple labels. It is pivotal to learn and utilize the label dependence among multiple labels in multi-label classification. As a result of today's big and complex data, noisy labels are inevitable, making it looming to target multi-label classification with noisy labels. Although the importance of label dependence has been shown in multi-label classification with clean labels, it is challenging and hard to bring label dependence to the problem of multi-label classification with noisy labels. The issues are, that we do not understand why label dependence is helpful in the problem, and how to learn and utilize label dependence only using training data with noisy multiple labels. In this paper, we bring label dependence to tackle the problem of multi-label classification with noisy labels. Specifically, we first provide a high-level understanding of why label dependence helps distinguish the examples with clean/noisy multiple labels. Benefiting from the memorization effect in handling noisy labels, a novel algorithm is then proposed to learn the label dependence by only employing training data with noisy multiple labels, and utilize the learned dependence to help correct noisy multiple labels to clean ones. We prove that the use of label dependence could bring a higher success rate for recovering correct multiple labels. Empirical evaluations justify our claims and demonstrate the superiority of our algorithm.", + "bbox": [ + 76, + 335, + 472, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-label classification assigns a set of multiple labels for each instance [71]. As a practical learning paradigm, multi-label classification has been widely applied in various domains, ranging from computer vision [7] and natural language processing [41], to recommendation systems [69] and bioinformatics [8]. Consensually, compared", + "bbox": [ + 75, + 784, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with multi-class classification [20, 23, 24], where each instance is assigned with a single label, multi-label classification is more challenging [35]. Plenty of advanced methods are proposed in recent years for multi-label classification [77, 45, 14, 74, 37, 9, 19, 64].", + "bbox": [ + 500, + 305, + 890, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The great majority of the methods assume that training data are annotated precisely. However, noisy labels are inevitable in multi-label classification [36], especially for classification with big and complex data. They may be resulted by unintentional mistakes of manual and automatic annotators [51, 75, 13], or intentional corruptions on clean labels [50, 44]. Noisy labels severely impair the generalization of learned models, over-parameterized deep models in particular [26, 61, 58, 59, 55, 56]. A straightforward way to address the problem of multi-label classification with noisy labels is to treat each label in isolation and convert the multi-label problem into a number of binary classification problems. Afterward, the methods in multi-class classification with noisy labels [16, 47] are applied to train independent binary classifiers, which capture instance-label dependence robustly to strengthen classification. This way is a remedy to handle noisy labels, but ignores the label dependence among multiple labels. It is essential to learn and utilize the label dependence in multi-label classification [70, 18, 11, 31].", + "bbox": [ + 496, + 383, + 892, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior works [65, 6, 52] illustrate the successes of considering the label dependence among multiple labels in multilabel classification with clean labels. In different ways, e.g., helping learn inter-dependent classifiers [7], the label dependence can be used to boost the learning of the instance-label dependence, which improves final classification. Inspired by the successes, it is concerned that label dependence could be exploited to handle the problem of multilabel classification with noisy labels. However, there are few attempts before for this important problem. At least three questions make the solution remain mysterious. First, in intuition, we need to understand why label dependence is helpful for the problem. Second, in technique, we need to know how to learn and utilize the label dependence in the", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (tongliang.liu@sydney.edu.au).", + "bbox": [ + 94, + 886, + 387, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1483", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4b3ab45ab5805c96c8957631002d585b9fbac6a93b1b94f640454d10a95c141.jpg", + "image_caption": [ + "Figure 1. The illustration of why the label dependence among multiple labels helps distinguish the examples with noisy/clean multiple labels. The arrow presents the label dependence between a label pair. For the labels “a” and “b”, “a → b” means that, when “a” appears, “b” will also occur with high probability. The example comes from a web search. The set of clean multiple labels is {Sea, Human, Motorboat}, where the label dependence is strong with both “Motorboat → Sea” and “Motorboat → Human”. However, due to label corruption, Motorboat is flipped to be Motorcycle, which causes “Motorcycle → Sea”. Therefore, the label dependence among noisy multiple labels is weaker than the label dependence among corresponding clean ones." + ], + "image_footnote": [], + "bbox": [ + 137, + 90, + 834, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "problem. As we only have training data with noisy labels, both the accurate catch and application of the label dependence are challenging. Third, in verification, we need to know what improvements the label dependence can bring.", + "bbox": [ + 75, + 323, + 468, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we answer the three questions one by one. The first answer is illustrated in Figure 1. That is, compared with noisy multiple labels, the label dependence among clean multiple labels is stronger with high probability. Therefore, such dependence could help distinguish the examples with noisy/clean multiple labels for our problem. The second answer is given by the proposed holistic correction for multi-label classification with noisy labels (aka HLC). Specifically, HLC inherits the memorization effect in handling noisy labels [1, 25, 53]: the deep model would firstly memorize the training examples with clean labels, leading to reliable model predictions in early training. In HLC, the label dependence is learned by a dynamic graph [65], and then applied to correcting noisy multiple labels. In more detail, the holistic score in HLC is proposed to measure the instance-label and label dependencies in an example. The stronger instance-label and label dependencies make a larger holistic score. We compare the ratio between the holistic scores of the example with noisy multiple labels and its variant with predicted multiple labels, with an easily determined threshold. The noisy multiple labels are corrected or changeless based on the comparison result. Benefiting from the memorization effect, both dependence learning and multi-label correction are useful. Besides, they fulfill a positive cycle [3]. Namely, better dependence learning results in a better multi-label correction, and better multi-label correction makes better dependence learning, leading to final enhanced classification.", + "bbox": [ + 75, + 385, + 470, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The third answer is given by both theoretical analyses and empirical evaluations. Theoretically, we show that the additional use of label dependence brings a higher probability to handle noisy multiple labels successfully than the sole use of instance-label dependence under some conditions. Empirically, we demonstrate the power of label de", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pendence through experiments and show that, in most situations, HLC outperforms comparison methods with large margins.", + "bbox": [ + 496, + 323, + 890, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper are summarized as follows. (1) We focus on a realistic problem of multi-label classification with noisy labels. The challenges of using label dependency to address the problem are carefully analyzed, which benefits future research on the problem. (2) We propose an effective method to handle noisy labels in multi-label classification. The method measures simultaneous instance-label and label dependencies in an example for follow-up label correction. (3) Theoretical analysis is provided to explain the success rate of the proposed method. Besides, we confirm that the use of label dependence is indeed powerful under some conditions. (4) Extensive empirical results on multiple benchmarks demonstrate the superiority of our method. Detailed ablation studies and discussions are also provided. Codes are attached in the supplementary material.", + "bbox": [ + 496, + 369, + 892, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 627, + 637, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem setup. Let $\\mathcal{X} \\in \\mathbb{R}^d$ denote the input space and $\\mathcal{Y} \\in \\{l_1, \\dots, l_q\\}$ denote the label space with $q$ class labels. An example with multiple labels is denoted as $(x, y)$ , where $x \\in \\mathcal{X}$ is the feature vector of an instance, and $y \\subseteq \\mathcal{Y}$ is its set of associated labels. Denote the size of the label set $y$ as $|y|$ . For the feature vector $x$ , its label set $y$ may be corrupted and is flipped into $\\bar{y} \\subseteq \\mathcal{Y}$ . We utilize a class-dependent noise transition matrix $\\pmb{T}$ [43, 46, 34, 60] to characterize the label flip process. Formally, for any $i \\neq j$ , $T_{ij} = \\mathbb{P}(l_j \\in \\bar{y} \\wedge l_i \\notin \\bar{y} | l_j \\notin y \\wedge l_i \\in y)$ represents the probability of the $i$ -th class label to be flipped into the $j$ -th class label. Consider a noisy multi-label dataset comprising several examples $(x, \\bar{y})$ . The aim is to learn a classification model robustly by only using the noisy dataset. Given an instance in testing, with the learned model, we can predict its relevant label set precisely.", + "bbox": [ + 496, + 657, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1484", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that some works employ another problem setting that the total number of multiple labels can be changed after label flipping, which is referred to as multi-label classification with missing or redundant labels. For the former setting, it is not accurate to consider it as a classification with noisy labels, since all annotated labels are correct [67, 57]. For the latter setting, it is normally called partial multi-label learning [62], which is different from the problem setting of this paper, as detailed in Appendix C.4. Our setting, i.e., the total number of labels is preserved after label flipping, is realistic. In many practical situations, it is easy to determine the number of objects in an image, in particular with object detection techniques. In contrast, it can be harder to annotate the objects perfectly, resulting in noisy labels. In addition, in Section 4.4, we will show that our problem setting well fits the realistic situation. That is to say, the proposed method can achieve superior performance on a realistic noisy multi-label dataset.", + "bbox": [ + 76, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preparation technology. As discussed, we need both the instance-label dependence and the label dependence among multiple labels. Given an example $(\\pmb{x},\\pmb{y})$ , for the instance-label dependence, it can be learned with the conditional probability of $l_{i}\\in \\pmb{y}$ given $\\pmb{x}$ according to model's probability outputs. For the label dependence among multiple labels, it is often estimated by counting the occurrence of label pairs in training data [7].", + "bbox": [ + 76, + 367, + 470, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the graph convolutional network (GCN) is used in multi-label classification and achieves great successes [7, 65, 6]. The advantage of the GCN-based methods is that they can capture the instance-label and label dependencies simultaneously during training. In this paper, we inherit the advantage of the GCN-based methods and build HLC based on ADDGCN [65]. ADDGCN designs a semantic attention module (SAM) to estimate the content-aware class-label representations for each class from the extracted feature map. The representations are fed into a GCN module (GCNM) for final classification. We provide the technical details of ADDGCN [65] in Appendix C.1. Before delving into the next section, readers only need to remember that the instance-label and label dependencies can be learned during training. Note that we also review prior works on multi-class classification with noisy labels and multi-label classification with clean/noisy labels in Appendix C.2 and Appendix C.3.", + "bbox": [ + 76, + 488, + 472, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 772, + 250, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Holistic Judgment in Multi-Label Classification", + "text_level": 1, + "bbox": [ + 76, + 796, + 468, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Holistic score. We begin with an example with clean multiple labels. Given an example $(\\pmb{x},\\pmb{y})$ , we can measure the instance-label dependence $S^f$ , and the label dependence $S^l$ . Denote the variable of clean multiple labels by $\\pmb{Y}$ . Mathematically, we define two dependen", + "bbox": [ + 76, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Holistic Correction." + ], + "code_body": "Input: $(\\pmb {x},\\bar{\\pmb{y}})$ $h$ ,and $\\hat{\\delta}$ \nOutput: $\\bar{y}_{new}$ \n1: $\\pmb {y}^{*} = h(\\pmb {x})$ . \n2: $\\kappa (h,\\pmb {x},\\bar{\\pmb{y}}) = \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}(\\pmb {x}) / \\hat{\\bar{S}}_{\\pmb{y}^{*}}(\\pmb {x})$ \n3: if $\\kappa (h,\\pmb {x},\\bar{\\pmb{y}})\\leq \\hat{\\delta}$ then $\\bar{y}_{new} = y^{*}$ \n4: else $\\bar{y}_{new} = \\bar{y}$", + "bbox": [ + 506, + 111, + 777, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cies as $S_{\\pmb{z}}^{f}(\\pmb{x}) \\coloneqq \\sum_{\\{\\pmb{Y} = \\pmb{z}, l_{i} \\in \\pmb{z}\\}} \\mathbb{P}(l_{i}|\\pmb{x})$ and $S_{\\pmb{z}}^{l}(\\pmb{x}) \\coloneqq \\sum_{\\{\\pmb{Y} = \\pmb{z}, l_{i}, l_{j} \\in \\pmb{z}\\}} \\frac{1}{2} [\\mathbb{P}(l_{j}|l_{i}, \\pmb{x}) + \\mathbb{P}(l_{i}|l_{j}, \\pmb{x})]$ , where $\\pmb{z}$ is the value of the random variable $\\pmb{Y}$ . The holistic score of the example $(\\pmb{x}, \\pmb{y})$ considers two dependencies at the same time. Formally, we denote the holistic score of $(\\pmb{x}, \\pmb{y})$ as $S_{\\pmb{y}}(\\pmb{x})$ and define it as", + "bbox": [ + 496, + 239, + 893, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\boldsymbol {y}} (\\boldsymbol {x}) := S _ {\\boldsymbol {y}} ^ {f} (\\boldsymbol {x}) + S _ {\\boldsymbol {y}} ^ {l} (\\boldsymbol {x}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 342, + 890, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Afterward, denote the variable of noisy multiple labels by $\\bar{Y}$ . For the example with noisy multiple labels, i.e., $(\\pmb{x},\\bar{\\pmb{y}})$ , the instance-label dependence and label dependence are measure by $\\bar{S}_z^f (\\pmb {x})\\coloneqq \\sum_{\\{\\bar{Y} = \\pmb {z},l_i\\in \\pmb {z}\\}}\\mathbb{P}(l_i|\\pmb {x})$ and $\\bar{S}_z^l (\\pmb {x})\\coloneqq$ $\\sum_{\\{\\bar{Y} = z,l_i,l_j\\in \\pmb {z}\\}}\\frac{1}{2}\\left[\\mathbb{P}(l_j|l_i,\\pmb {x}) + \\mathbb{P}(l_i|l_j,\\pmb {x})\\right]$ . Accordingly, the holistic score of the example $(\\pmb {x},\\bar{\\pmb{y}})$ is denoted by $\\bar{S}_{\\bar{\\pmb{y}}}(\\pmb {x})$ , which is defined as $\\bar{S}_{\\bar{\\pmb{y}}}(\\pmb {x})\\coloneqq \\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x}) + \\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})$ . Note that, during training, we cannot access $\\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})$ and $\\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})$ . Instead, the estimated posterior probabilities are used. We denote the estimations of $\\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})$ and $\\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})$ as $\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})$ and $\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})$ . The estimation of the holistic score is $\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}(\\pmb {x}) = \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x}) + \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})$ . With preparation technology discussed in Section 2 and Appendix C.1, $\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})$ and $\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})$ can be obtained.", + "bbox": [ + 496, + 369, + 893, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Holistic correction. For the example $(\\pmb{x},\\bar{\\pmb{y}})$ , we feed it into the deep network $h$ included in ADDGCN [65]. The memorization effect in handling noisy labels [25, 33] shows that the deep network would first memorize the training data with clean labels and then the training data with noisy labels. Therefore, early in training, the outputs of the deep network are relatively reliable and can be used for label correction. For $(\\pmb{x},\\bar{\\pmb{y}})$ , we denote its set of predicted multiple labels as $\\pmb{y}^*$ . Here, the set of predicted labels is obtained with the top $|\\bar{\\pmb{y}}|$ predictions based on the model's probability outputs.", + "bbox": [ + 496, + 613, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recall that the holistic score of an example holistically measures the instance-label dependence and label dependence among multiple labels simultaneously. From both human and machine cognition, if an example is annotated accurately, both dependencies should be strong [70, 18, 68, 28, 7] with high probability. Namely, the holistic score is large. We propose to check the ratio between the holistic score on $(\\pmb{x},\\bar{\\pmb{y}})$ and holistic score on $(\\pmb{x},\\pmb{y}^{*})$ . Specifically,", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1485", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we check", + "bbox": [ + 76, + 92, + 143, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa (h, \\boldsymbol {x}, \\bar {\\boldsymbol {y}}) = \\hat {\\tilde {S}} _ {\\bar {\\boldsymbol {y}}} (\\boldsymbol {x}) / \\hat {\\tilde {S}} _ {\\boldsymbol {y} ^ {*}} (\\boldsymbol {x}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 102, + 468, + 123 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compare this ratio with a predetermined threshold $\\hat{\\delta}$ . The value of $\\hat{\\delta}$ is given in the next subsection. If $\\kappa(h, \\boldsymbol{x}, \\bar{\\boldsymbol{y}}) \\leq \\hat{\\delta}$ , we flip the labels $\\bar{\\boldsymbol{y}}_{\\text{new}} = \\boldsymbol{y}^*$ . Otherwise, the labels remain unchanged with $\\bar{\\boldsymbol{y}}_{\\text{new}} = \\bar{\\boldsymbol{y}}$ . The detailed algorithm of holistic correction for multi-label classification with noisy labels (aka HLC) is provided in Algorithm 1. After holistic correction for noisy labels, we use $(\\boldsymbol{x}, \\bar{\\boldsymbol{y}}_{\\text{new}})$ to train the deep network $h$ based on ADDGCN [65].", + "bbox": [ + 75, + 128, + 468, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Theoretical Insights", + "text_level": 1, + "bbox": [ + 76, + 258, + 266, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We extend the Tsybakov condition [75, 2, 15, 49] from multi-class classification to multi-label classification. Specifically, denote by $\\pmb{a}_{\\pmb{x}}$ the label set predicted based on $S^{f}(\\pmb{x})$ with $\\pmb{a}_{\\pmb{x}} := h^{*}(\\pmb{x}) = \\arg \\max_{\\pmb{z}} S_{\\pmb{z}}^{f}(\\pmb{x})$ . Besides, denote by $\\pmb{b}_{\\pmb{x}}$ the second best prediction with $\\pmb{b}_{\\pmb{x}} := \\arg \\max_{\\pmb{z} \\neq \\pmb{a}_{\\pmb{x}}} S_{\\pmb{z}}^{f}(\\pmb{x})$ . The maximum length of a label set is denoted as $m$ ( $m \\ll q$ ). In this paper, we call the predicted label set by the Bayes optimal classifier for an instance as the correct label set. We present the Tsybakov condition on instance-label (abbreviated as ins.-label here) dependence and holistic Tsybakov condition as follows.", + "bbox": [ + 75, + 282, + 468, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 (Tsybakov condition on ins.-label dependence) $\\exists C_1, \\lambda_1 > 0$ and $\\exists t_0 \\in (0, m]$ , such that for all $t \\leq t_0$ , we have", + "bbox": [ + 75, + 462, + 486, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left[ S _ {\\boldsymbol {a} _ {\\boldsymbol {x}}} ^ {f} (\\boldsymbol {x}) - S _ {\\boldsymbol {b} _ {\\boldsymbol {x}}} ^ {f} (\\boldsymbol {x}) \\leq t \\right] \\leq C _ {1} t ^ {\\lambda_ {1}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 505, + 468, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 2 (Holistic Tsybakov condition) $\\exists C_2, \\lambda_2 > 0$ and $\\exists t_0 \\in (0, m]$ , such that for all $t \\leq t_0$ , we have", + "bbox": [ + 75, + 535, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left[ S _ {\\boldsymbol {a} _ {\\boldsymbol {x}}} (\\boldsymbol {x}) - S _ {\\boldsymbol {b} _ {\\boldsymbol {x}}} (\\boldsymbol {x}) \\leq t \\right] \\leq C _ {2} t ^ {\\lambda_ {2}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 577, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 1 Definition 1 stipulates that the uncertainty of $S^f$ is bounded. The margin region that is close to the decision boundary has a bounded volume. Definition 2 shares the similar idea and bound the uncertainty of $S$ .", + "bbox": [ + 75, + 606, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 Suppose $S(\\pmb{x})$ fulfills the holis-tic Tsybakov condition for constants $C_2, \\lambda_2 > 0$ , and $t_0 \\in (0, m]$ . We define $\\epsilon := \\max_{\\pmb{x}, \\pmb{z}} \\left[ |\\hat{\\bar{S}}_z^f(\\pmb{x}) - \\bar{S}_z^f(\\pmb{x})|, |\\hat{\\bar{S}}_z^l(\\pmb{x}) - \\bar{S}_z^l(\\pmb{x})|, |\\bar{S}_z^l(\\pmb{x}) - S_z^l(\\pmb{x})| \\right]$ and $\\tau := \\min_i T_{ii}$ . We analyze two cases:", + "bbox": [ + 75, + 680, + 496, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) If $\\bar{\\pmb{y}}$ is corrected by $\\kappa(h, \\pmb{x}, \\bar{\\pmb{y}})$ with $\\hat{\\delta}$ , let $\\delta_1 = \\min \\left[\\frac{\\tau S_{\\pmb{b}_\\pmb{x}}(\\pmb{x}) + \\sum_{l_j \\in \\bar{\\pmb{y}}} \\sum_{i \\neq j} T_{ij} \\mathbb{P}(l_i | \\pmb{x})}{\\hat{S}_{\\pmb{y}^*}(\\pmb{x})}\\right]$ and $\\rho_1 := |\\hat{\\delta} - \\delta_1|$ . Assume that $\\epsilon \\leq \\frac{t_0 \\tau - \\rho_1 m}{3}$ . Then, $\\mathbb{P}[\\bar{\\pmb{y}}_{new} = h^*(\\pmb{x}), \\bar{\\pmb{y}}$ is flipped] is at least $1 - C_2 [O(\\max(\\epsilon, \\rho_1))]^{\\lambda_2} - \\mathbb{P}[\\pmb{a}_x \\neq \\{\\pmb{y}^*, \\bar{\\pmb{y}}\\}]$ .", + "bbox": [ + 75, + 761, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2) If $\\bar{\\pmb{y}}$ is not corrected by $\\kappa(h, \\pmb{x}, \\bar{\\pmb{y}})$ with $\\hat{\\delta}$ , let $\\delta_2 = \\max \\left[\\frac{\\hat{\\tilde{S}}_{\\bar{\\pmb{y}}}(\\pmb{x})}{\\tau S_{b_{\\pmb{x}}}(\\pmb{x}) + \\sum_{l_j \\in \\pmb{y}^*} \\sum_{i \\neq j} T_{ij} \\mathbb{P}(l_i | \\pmb{x})}\\right]$ and $\\rho_2 := |\\hat{\\delta} - \\delta_2|$ .", + "bbox": [ + 75, + 856, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assume that $\\epsilon \\leq \\frac{t_0\\delta_2^2\\tau - \\rho_2m - \\rho_2^2m}{3\\delta_2^2}$ . Then, $\\mathbb{P}[\\bar{\\pmb{y}}_{new} = h^{*}(\\pmb{x}),\\bar{\\pmb{y}}$ is accepted] is at least $1 - C_2[O(\\max (\\epsilon ,\\rho_2))]^{\\lambda_2} - \\mathbb{P}[\\pmb {a}_x\\neq \\{\\pmb {y}^*,\\bar{\\pmb{y}}\\} ]$ .", + "bbox": [ + 498, + 88, + 890, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Theorem 1 is provided in Appendix B.1. Theorem 1 extends the theoretical results of [75] to multilabel classification with noisy labels. It claims that, even though with noisy multiple labels, the holistic correction has a guaranteed success rate to make proper corrections. Besides, if we can reasonably approximate the optimal $\\delta$ with $\\hat{\\delta}$ , our algorithm flips noisy multiple labels to correct ones with a good chance. Below, as a corollary of Theorem 1, we show that, there are certain circumstances, the use of holistic scores has a better chance to make corrections satisfactorily, than the sole use of instance-label dependence.", + "bbox": [ + 496, + 154, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Corollary 1 Suppose that $S(\\pmb{x})$ fulfills the holistic Tsybakov condition. Denote the set threshold $\\hat{\\delta}$ and optimal threshold $\\delta$ . We define $\\rho := \\max |\\hat{\\delta} - \\delta|$ . We have that, $\\exists \\epsilon$ and $\\rho$ , if $C_2[O(\\max(\\epsilon, \\rho))]^{\\lambda_2} < C_1[O(\\max(\\epsilon, \\rho))]^{\\lambda_1}$ , holistic correction brings higher probability to handle noisy labels successfully than instance-label dependence.", + "bbox": [ + 496, + 348, + 890, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Corollary 1 is provided in Appendix B.2. Corollary 1 claims that there exist cases where holistic scores better combat noisy labels. Note that, from a theoretical view, we do not state that holistic scores can work better in all circumstances of multi-label classification. Nevertheless, with the determination of the threshold $\\hat{\\delta}$ , holistic scores can perform better in the experiments of this paper, which demonstrates the help of label dependence to handle noisy multiple labels.", + "bbox": [ + 496, + 452, + 890, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 601, + 633, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 626, + 689, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We verify the effectiveness of the proposed method on the synthetic noisy versions of three datasets, i.e., Pascal-VOC 2007 [12], Pascal-VOC 2012 [12], and MS-COCO [32]. Pascal-VOC 2007 contains 5,011 images in train and validation sets, while Pascal-VOC 2012 consists of 11,540 images in train and validation sets. The images come from 20 common object categories. For Pascal-VOC 2007 and Pascal-VOC 2012, we train methods using the noisy training and validation sets, and evaluate them on the test set of Pascal-VOC 2007 that has 4,952 images [14]. MS-COCO contains 82,081 training images and 40,137 validation images from 80 common object categories. As did in [74, 5, 65, 77], we evaluate the performance of methods using validation images.", + "bbox": [ + 496, + 654, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Noisy-label generation. The class-dependent noise transition matrix $T$ [42, 21, 46, 72] is used to corrupt the three", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1486", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/5b8c76122664480be695258f8ea9f2a7e080fd8268bb636019098ee8ad07ac15.jpg", + "table_caption": [ + "Table 1. Comparisons with advanced methods on noisy Pascal-VOC 2007. The mean and standard deviation of results (%) are presented." + ], + "table_footnote": [], + "table_body": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE64.50±1.2058.65±2.1648.19±0.2371.77±1.1560.94±4.2548.72±2.13
CSRA66.99±0.4859.62±0.6146.97±0.4872.45±0.6963.58±1.4852.72±1.52
ADDGCN63.89±0.9455.75±1.9844.14±1.3771.02±0.9561.05±0.0650.18±2.70
APL66.79±1.1958.86±1.5347.64±1.8172.61±0.9961.99±0.7849.10±0.15
CDR67.35±1.7060.05±1.0649.12±0.5972.66±0.7964.58±0.6050.51±2.49
JOINT67.43±0.7363.37±0.9253.27±4.7070.28±1.8568.70±2.8858.57±2.75
WSIC65.43±0.5559.53±0.7348.34±0.4772.57±1.0361.88±2.5750.15±0.86
CCMN69.97±1.3662.58±1.4753.20±1.2870.68±1.0860.94±3.1248.62±1.26
HLC†72.07±0.6770.20±0.4668.00±0.8974.83±0.6469.86±1.6160.09±1.73
OF1 ↑BCE63.52±0.4856.70±2.4548.10±1.4368.28±0.6958.30±2.8251.18±3.10
CSRA65.40±0.4759.39±0.8148.32±1.5069.72±0.5061.89±0.4351.56±2.28
ADDGCN62.63±0.1855.50±1.8744.38±2.9268.95±0.6459.64±0.5653.12±0.62
APL64.85±1.4656.51±1.7047.54±2.4068.89±0.8958.04±0.9752.27±2.20
CDR65.31±0.9957.93±1.0548.86±1.7169.53±0.6559.89±1.0751.68±3.83
JOINT69.72±0.8867.93±0.7761.62±1.4071.24±1.0364.20±0.8860.30±1.24
WSIC63.45±0.9757.96±1.2548.38±2.4169.88±1.2257.97±2.1951.99±1.65
CCMN69.66±1.5560.43±1.3153.84±0.6967.12±0.6159.55±1.4553.46±1.04
HLC†71.03±0.3369.08±1.0068.62±0.4872.09±0.7465.76±2.3960.71±1.37
CF1 ↑BCE58.91±1.3453.21±2.0443.66±0.5365.93±0.8157.03±3.4347.21±1.89
CSRA62.31±0.5055.67±0.6143.11±0.7667.39±0.8059.66±1.0451.13±1.12
ADDGCN60.41±1.0453.72±1.3842.42±0.5966.05±0.9757.81±0.5848.89±2.64
APL60.23±1.5352.85±2.1842.38±1.6766.59±0.7158.33±0.4947.67±1.83
CDR61.37±1.4754.17±0.8643.60±0.8267.11±0.6359.91±0.3948.40±1.98
JOINT63.13±0.3860.22±1.6848.17±5.0166.03±1.2562.05±2.9854.03±3.17
WSIC59.54±1.1054.22±0.5343.82±0.6266.97±1.0058.04±1.7048.19±0.96
CCMN65.19±1.1058.55±1.3149.85±1.0665.47±0.9358.05±2.2448.46±0.80
HLC†68.87±0.1066.62±0.8164.82±0.4869.95±1.1965.13±1.0457.54±1.84
", + "bbox": [ + 150, + 99, + 816, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets. Here, for any $i \\neq j$ , $T_{ij} = \\mathbb{P}(l_j \\in \\bar{\\mathbf{y}} \\wedge l_i \\notin \\bar{\\mathbf{y}} | l_j \\notin \\mathbf{y} \\wedge l_i \\in \\mathbf{y})$ represents the probability of the $i$ -th class label to be flipped into the $j$ -th class label. We consider both symmetric (abbreviated as Sym.) and pairflip (abbreviated as Pair.) noise settings [17]. The details of the transition matrix are provided in Appendix D.2. For symmetric noise, the noise rate is set to $30\\%$ , $40\\%$ , and $50\\%$ . For pairflip noise, the noise rate is set to $20\\%$ , $30\\%$ , and $40\\%$ .", + "bbox": [ + 75, + 470, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We exploit three types of baselines in total. Specifically, Type-I baselines contain the methods that are designed for multi-label classification with clean labels. Type-II baselines consider the methods for multi-class classification with noisy labels. Type-III baselines consider the methods that focus on multi-label classification with noisy labels. It should be noted that, there are relatively few methods belonging to this type [36]. More advanced methods belonging to Type-III baselines need to be investigated [36], which is also our focus in this paper. In more detail, Type-I baselines include CSRA [77] and ADDGCN [65]. Type-II baselines include APL [40], CDR [58], and JOINT [48]. Type-III baselines include WSIC [22] and CCMN [63]. As a simple baseline, we compare our method with the standard deep network that directly trains on noisy datasets (abbreviated as BCE). We detail all baselines in Appendix D.1.", + "bbox": [ + 75, + 603, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network & Optimizer. We use a ResNet-50 network [20] pretrained on ImageNet as the backbone for all methods. We train the models for 30 epochs in total. We utilize", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adam [27] for the network optimization. The batch size is set to 128 for all the datasets. The learning rate is fixed to $5 \\times 10^{-5}$ . The images in Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO resize to $224 \\times 224$ . Note that, to make experiments more comprehensive, we also employ different experimental settings, e.g., different networks and different image sizes. The details are provided in Section 4.3.", + "bbox": [ + 496, + 470, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Measurement. As did in multi-label classification [77, 7], evaluation metrics include the mean average precision (mAP) [71], the average F1-measure (OF1), and the average per-class F1-measure (CF1). For fair comparison, we implement all methods with default parameters by PyTorch, and conduct all experiments on NVIDIA GTX3090 GPUs. All experiments are repeated three times with different random seeds. Following the works in learning with noisy labels [17, 54, 29, 30], the mean and standard deviation of results in the last epoch are reported. In addition, for different evaluation metrics, we report the mean and standard deviation of best results. Supplementary results are shown in Appendix E. Afterwards, the best mean results are highlighted in red. The second best mean results are also highlighted in blue.", + "bbox": [ + 496, + 597, + 892, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with the State-of-the-Arts", + "text_level": 1, + "bbox": [ + 500, + 832, + 831, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO are shown in Table 1, Table 2, and Table 3 respectively. In summary, HLC consistently works", + "bbox": [ + 500, + 854, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1487", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/861655df74c45240a28c1723f94f1251d3b60b26ee722a3de574f3b5ea616970.jpg", + "table_caption": [ + "Table 2. Comparisons with advanced methods on noisy Pascal-VOC 2012. The mean and standard deviation of results (%) are presented." + ], + "table_footnote": [], + "table_body": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE66.74±0.8056.07±0.5045.15±1.5670.91±1.1357.61±1.1449.85±0.36
CSRA66.35±0.5056.20±1.3545.54±1.1471.29±0.8360.71±1.1847.63±1.56
ADDGCN63.34±0.9654.54±0.8644.88±1.7170.41±0.5457.96±0.6847.66±1.08
APL67.07±1.0456.79±1.8643.51±1.9371.32±1.6059.59±1.2748.14±1.16
CDR66.13±1.4956.85±0.4844.84±1.1171.55±1.8760.13±1.8949.44±1.81
JOINT65.19±2.1758.40±2.8745.13±1.6968.93±2.5461.64±1.7853.64±1.61
WSIC65.96±0.7956.34±0.4144.80±0.5470.40±1.1159.40±1.8748.95±1.34
CCMN69.15±0.6661.00±1.0150.71±0.2669.08±1.7859.72±2.3246.67±2.78
HLC†72.14±0.6670.11±0.2768.69±1.0474.51±0.6769.90±0.4364.20±1.26
OF1 ↑BCE64.99±1.1056.92±2.0845.49±2.2368.48±2.2860.21±1.3554.05±1.95
CSRA64.08±0.3756.25±2.5748.67±3.1469.06±0.6559.75±1.7052.89±0.95
ADDGCN63.53±1.4154.28±0.8647.56±2.6747.62±2.3957.90±1.7852.33±0.56
APL64.70±1.1758.05±1.6845.74±1.5570.68±1.0360.22±1.5851.38±1.55
CDR64.06±1.3857.31±1.2146.51±0.9570.45±1.4460.57±1.2452.26±2.42
JOINT67.35±1.8664.57±2.3954.37±3.3370.81±1.4064.40±1.7656.27±1.29
WSIC62.74±2.1057.13±0.7345.52±1.2869.72±1.1959.11±2.0452.49±1.38
CCMN65.77±0.2359.91±0.9351.45±0.9467.93±1.7359.26±0.5148.61±4.71
HLC†71.14±0.6069.50±0.4067.80±0.3372.13±0.2667.59±0.9664.28±0.81
CF1 ↑BCE62.47±0.4453.26±0.4143.43±1.6766.03±1.6955.90±0.7049.29±0.64
CSRA62.08±0.7053.23±1.2743.23±1.2566.02±0.7457.71±1.0247.46±1.68
ADDGCN59.67±1.1452.61±0.5244.33±1.9965.22±0.8655.32±0.7647.30±1.12
APL62.99±1.0753.69±1.8041.72±1.4266.44±1.4057.52±0.9148.14±1.02
CDR62.18±1.0453.61±0.4542.83±0.8766.29±2.1257.23±1.4349.03±1.50
JOINT60.57±2.8254.39±3.7240.48±7.7066.30±2.3359.72±2.1255.06±0.36
WSIC61.70±0.9253.10±0.7442.72±0.5465.34±1.4857.21±1.6248.51±1.12
CCMN64.46±0.6257.45±0.9948.27±0.6867.48±1.4456.93±1.6947.01±1.82
HLC†69.54±0.5667.35±0.4865.72±1.4870.07±0.4165.68±0.9460.57±1.27
", + "bbox": [ + 150, + 101, + 816, + 446 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/697140414eba5919b482e3829a30c42e9c953af112581ccf822b39fb5b8b2458.jpg", + "table_caption": [ + "Table 3. Comparisons with advanced methods on noisy MS-COCO. The mean and standard deviation of results (%) are presented." + ], + "table_footnote": [], + "table_body": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE53.23±0.1547.33±0.7940.25±0.2656.58±0.2249.16±0.0441.57±0.64
CSRA53.89±0.4047.64±0.8639.58±0.1958.27±0.2350.95±0.0743.07±0.64
ADDGCN51.08±0.9544.75±1.1538.66±1.3056.94±0.6150.28±0.8141.45±0.19
APL54.34±0.3248.61±0.7243.55±1.4357.73±0.2050.87±0.3441.77±0.50
CDR54.01±0.0449.01±0.2643.94±1.2557.03±0.2850.99±0.7742.71±0.09
JOINT53.93±0.4148.01±1.0445.27±0.6857.30±0.3351.94±0.2042.74±0.55
WSIC52.99±0.5346.84±0.8639.76±0.6456.66±0.3149.46±0.2542.52±0.62
CCMN51.73±0.1850.36±0.7145.32±0.8958.13±0.4451.17±0.2942.12±0.76
HLC†54.87±0.6851.09±0.5348.15±0.5058.55±0.0953.41±0.1345.91±0.39
OF1 ↑BCE51.34±1.7044.36±0.8234.85±1.2459.16±0.9552.44±0.8142.94±1.13
CSRA52.03±1.8641.63±1.4133.47±3.1859.17±0.1450.27±0.8841.75±1.36
ADDGCN55.67±1.4847.79±0.4035.95±3.7360.96±0.6555.05±1.7847.47±0.77
APL51.07±1.3243.93±2.7033.90±4.0060.04±1.1650.64±2.8644.34±1.99
CDR53.43±1.1645.10±0.8334.91±0.9059.34±0.6152.72±0.6344.17±0.61
JOINT54.56±0.0649.00±1.6637.78±0.9358.20±0.4053.21±0.1746.55±0.61
WSIC50.91±0.5242.93±0.8535.47±1.5258.89±1.1351.63±1.5743.99±1.47
CCMN52.71±1.0443.24±1.1934.62±1.3858.61±1.1852.18±0.7645.92±0.59
HLC†59.92±0.6557.84±0.3855.47±0.9562.28±0.0658.56±0.3751.09±0.60
CF1 ↑BCE45.92±0.2338.96±1.6131.34±0.2752.54±0.5845.54±0.6339.79±0.99
CSRA44.97±1.8837.49±1.7328.96±1.1652.18±0.4444.96±0.4336.88±0.21
ADDGCN46.77±1.8039.35±1.8330.57±1.5754.18±0.2347.55±0.1839.44±0.33
APL42.91±0.5438.38±0.7728.17±2.5052.87±1.0746.27±1.2737.76±1.02
CDR46.62±0.4239.47±0.5429.59±2.5252.51±0.6945.75±0.8139.15±0.53
JOINT49.51±0.8142.38±1.2124.24±0.6154.39±0.1749.90±0.8538.34±0.55
WSIC45.30±1.0939.15±1.6231.42±0.9452.04±0.2845.76±0.7039.44±1.11
CCMN44.20±1.1935.18±1.0127.90±1.2553.23±0.5846.88±0.9240.55±0.89
HLC†51.94±0.6349.24±0.3046.69±0.6655.44±0.1350.91±0.4843.35±0.82
", + "bbox": [ + 150, + 469, + 816, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "best across all noise settings. In many cases, the best results achieved by HLC outperform the second best results by a large margin, especially when the noise level is high. Below, we further discuss the results based on the comparisons", + "bbox": [ + 75, + 834, + 468, + 895 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with three different types of baselines.", + "bbox": [ + 500, + 834, + 751, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared with Type-I baselines. We first notice that Type-I baselines are fragile to noisy labels in multi-label classification. Without considering the side-effect of noisy", + "bbox": [ + 498, + 854, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1488", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "labels, in many cases, they perform worse than BCE, which clearly illustrates the necessity for attention to handling noisy labels. Second, we compare HLC with ADDGCN. Without the proposed correction method for combating noisy labels, HLC will reduce to ADDGCN. As shown in the reported results, HLC performs much better than ADDGCN. To be specific, on noisy Pascal-VOC 2007, for Sym. $40\\%$ , HLC brings about $+15\\%$ performance improvement w.r.t. three evaluation metrics over ADDGCN. For Sym. $50\\%$ , the performance improvement is increased to more than $+20\\%$ . Also, for Pair. $30\\%$ and Pair. $40\\%$ , HLC enhances ADDGCN with about $+10\\%$ improvement. On noisy Pascal-VOC 2012 and MS-COCO, the performance improvement is also very clear.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared with Type-II baselines. On noisy Pascal-VOC 2007, with Sym. noise, we can see that HLC outperforms APL, CDR, and JOINT clearly, especially for Sym. $50\\%$ . Additionally, with Pair. noise, although the improvement is less than the cases with Sym. noise, HLC still performs best. On noisy Pascal-VOC 2012, for both Sym. and Pair. noise, the improvement is significant. Lastly, for noisy MS-COCO, HLC works better than all Type-II baselines with varying enhancement.", + "bbox": [ + 75, + 306, + 470, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that, compared with APL and CDR, JOINT seems to be a stronger baseline. Benefiting from label correction, after a few training epochs, JOINT less overfits to wrong labels, following better performance. Nevertheless, the proposed label-correction paradigm is argued to be more advanced. As shown in all results, HLC surpasses JOINT, which verifies the effectiveness of our method.", + "bbox": [ + 75, + 444, + 470, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared with Type-III baselines. On noisy Pascal-VOC 2007 and noisy Pascal-VOC 2012, HLC outperforms WSIC and CCMN distinctly. For example, with Sym. $50\\%$ noise, more than $+10\\%$ performance promotion is brought by our method. On noisy MS-COCO, although WSIC and CCMN are sometimes competitive w.r.t. mAP, they are inferior w.r.t. both OF1 and CF1.", + "bbox": [ + 75, + 556, + 470, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. More Analyses and Justifications", + "text_level": 1, + "bbox": [ + 76, + 674, + 366, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we conduct performance analysis in more detail. The experiments are conducted with Sym. $50\\%$ noise, which is more challenging than the experiments in low-noise-rate cases.", + "bbox": [ + 75, + 698, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Role of label dependence. We study the effect of removing the consideration of label dependence to provide insights into what makes HLC successful. The experiments are conducted on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to $224 \\times 224$ . Recall that HLC considers instance-label and label dependences simultaneously. When we remove the consideration of the label dependence in HLC, the correspond", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a08a55ece4d032c52d71ab6150c92b4f8a475f8d6b46af72ba8b2d8aff598e9c.jpg", + "table_caption": [ + "Table 4. Ablation study results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The mean and standard deviation of results are presented. The best result in each case is in bold." + ], + "table_footnote": [], + "table_body": "
DatasetNoisy Pascal-VOC 2007
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.06±0.4167.23±1.9263.42±0.58
HLC68.00±0.8968.62±0.4864.82±0.48
DatasetNoisy Pascal-VOC 2012
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.88±0.7566.30±1.2864.33±1.67
HLC68.69±1.0467.80±0.3365.72±1.48
DatasetNoisy MS-COCO
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.46.21±0.3652.90±0.9244.51±1.29
HLC48.15±0.5055.47±0.9546.69±0.66
", + "bbox": [ + 527, + 128, + 861, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3dc4cc0c589ddd7719ba50a4ffd2c9521624dae137206fe861a7090f5299cc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 303, + 823, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/514fc574be132d5f5c5d4bca61e6f2621079a405bfa991e66ed9f4d718c2b0e6.jpg", + "image_caption": [ + "Figure 2. Ablation study results with different values of the set threshold $\\hat{\\delta}$ . The experiments are conducted on noisy Pascal-VOC 2007 (Top) and noisy Pascal-VOC 2012 (Bottom)." + ], + "image_footnote": [], + "bbox": [ + 563, + 416, + 823, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing method is named as HLC w/o l. here. For both HLC w/o l. and HLC, the value of the threshold $\\hat{\\delta}$ is searched in the range $\\{0.25, 0.30, 0.35, 0.40, 0.45\\}$ . We use the $10\\%$ noisy training data as a validation set for the threshold determination and performance report. The results are shown in Table 4. As can be seen, HLC outperforms HLC w/o l.. The results justify our claims that the label dependence could help combat the noisy labels in multi-label classification, which demonstrate the effectiveness of the proposed holistic correction.", + "bbox": [ + 496, + 585, + 893, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of the threshold $\\hat{\\delta}$ . We analyze the influence of different values of $\\hat{\\delta}$ . The experiments are conducted on noisy Pascal-VOC 2007 and Pascal-VOC 2012. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to $224 \\times 224$ . The value of the threshold $\\hat{\\delta}$ is chosen in $\\{0.25, 0.30, 0.35, 0.40, 0.45\\}$ . Figure 2 shows that HLC is robust to the determination of the threshold $\\hat{\\delta}$ in the certain range, which facilitates the practical application of our method.", + "bbox": [ + 496, + 742, + 893, + 880 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluations with different networks. We use pretrained", + "bbox": [ + 500, + 885, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1489", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/15b07f682159307d4b6ee3ff4467ee549233a1c227b64a6bf4517f2e7e6997d2.jpg", + "table_caption": [ + "Table 5. Comparisons with advanced methods on noisy MS COCO with different networks. The mean and standard deviation of results $(\\%)$ are presented." + ], + "table_footnote": [], + "table_body": "
MetricsMethodsResNet-34ResNet-101
mAP ↑BCE42.63±0.7438.17±0.41
CSRA41.35±0.1837.24±1.20
ADDGCN40.15±0.9836.13±0.69
APL44.82±0.7040.90±1.51
CDR45.43±0.6541.00±0.38
JOINT44.81±0.7739.96±1.30
WSIC41.86±0.6237.49±0.68
CCMN45.31±0.4746.01±1.01
HLC†46.05±0.8146.24±2.13
OF1 ↑BCE37.65±2.4638.65±2.50
CSRA35.05±0.7834.28±2.40
ADDGCN35.11±1.2637.18±0.50
APL34.31±1.7337.89±1.30
CDR36.67±3.4339.88±1.15
JOINT39.66±1.1341.36±0.88
WSIC35.08±1.7438.44±0.57
CCMN32.86±1.5037.03±1.48
HLC†44.11±0.8047.79±4.19
CF1 ↑BCE28.95±2.0934.11±1.13
CSRA27.40±0.4431.21±1.57
ADDGCN26.11±0.8630.41±0.72
APL26.64±2.3931.97±1.95
CDR27.13±2.3935.17±1.06
JOINT30.77±1.6337.63±0.81
WSIC27.62±0.6533.83±0.61
CCMN24.75±0.4826.65±0.26
HLC†34.79±1.4140.88±2.42
", + "bbox": [ + 117, + 128, + 424, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ResNet-50 before. To show that our method is robust to the choice of network structures, we use different networks in experiments. Specifically, we employ pretrained ResNet-34 [20] and pretrained ResNet-101 [20] respectively. The noisy MS-COCO is considered. The image size is $224 \\times 224$ . The results on mAP are reported in Table 5. As can be seen, with different networks, HLC still works well.", + "bbox": [ + 75, + 500, + 468, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluations with different image sizes. We resize the image size to $224 \\times 224$ before. To test the performance of advanced methods with different image sizes, we further consider $112 \\times 112$ , $384 \\times 384$ , and $448 \\times 448$ image sizes. Pretrained ResNet-50 is used. The results are reported in Table 6. For mAP, we can see that HLC is competitive compared with CCMN and CSRA. For OF1 and CF1, HLC works better than all baselines with a clear margin.", + "bbox": [ + 75, + 609, + 470, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Experiments on the Real-world Dataset", + "text_level": 1, + "bbox": [ + 76, + 741, + 415, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate that our problem setting can be adapted to the real world and our method can well handle practical scenes, we employ the real-world dataset NUS-WIDE [10] that originally contained 269,648 images from Flicker, which have been manually annotated with 81 visual concepts. Since some urls for download have been deleted, we employ the dataset version in [45]. A standard 70-30 train-test split is used. The backbone is chosen as ResNet-101. As the computation cost of training on NUS-WIDE is rela", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c39cf49eecf790512ab2e512c1187d0e9bab806e265c764fd711603f7cbe1e71.jpg", + "table_caption": [ + "Table 6. Comparisons with advanced methods on noisy MS COCO. The mean and standard deviation of results $(\\%)$ are presented. Difference image sizes are considered here." + ], + "table_footnote": [], + "table_body": "
MetricsImage sizes112 × 112384 × 384448 × 448
mAP ↑BCE32.22±0.6939.40±1.3635.24±1.73
CSRA29.55±0.1643.55±0.7044.56±0.75
ADDGCN32.34±0.4638.72±1.6434.87±1.89
APL34.41±0.4843.65±0.2841.44±1.21
CDR34.75±0.3943.26±0.7239.97±1.40
JOINT32.89±0.1642.95±0.8840.17±1.26
WSIC31.98±0.2339.57±1.0236.08±0.23
CCMN36.17±0.4144.39±0.3944.03±0.17
HLC†35.98±1.0545.12±0.1344.23±1.20
OF1 ↑BCE26.70±0.8834.71±2.7626.72±3.35
CSRA20.14±1.2836.41±0.7138.54±0.78
ADDGCN26.36±2.2942.83±2.0540.73±1.04
APL24.02±1.2234.68±1.4630.73±2.64
CDR26.50±1.2331.31±1.7631.15±3.46
JOINT34.11±0.9538.67±1.2538.11±0.69
WSIC24.61±1.1034.09±2.9430.69±0.97
CCMN23.89±1.4936.16±2.1225.03±2.48
HLC†39.05±2.6846.55±3.7745.14±2.34
CF1 ↑BCE19.61±0.6130.77±2.2824.94±3.31
CSRA13.34±1.2932.66±0.9832.68±0.22
ADDGCN18.67±1.4735.63±1.3733.89±2.41
APL17.21±0.7829.24±0.0925.01±1.31
CDR18.04±1.0729.62±1.1926.28±1.23
JOINT20.76±0.7534.90±1.8833.75±1.31
WSIC19.07±0.5731.63±2.4928.50±1.36
CCMN16.75±1.2130.09±2.3226.68±1.57
HLC†29.70±2.0740.34±2.1437.48±2.36
", + "bbox": [ + 501, + 128, + 893, + 472 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c805c867923ebf9002a5329848b886e12165cb85d529a4ff3bc2f0ed68d619a.jpg", + "table_caption": [ + "Table 7. Comparison of our method to known state-of-the-art models on the NUS-WIDE dataset. Metrics are in %." + ], + "table_footnote": [], + "table_body": "
MethodmAP ↑OF1 ↑CF1 ↑
S-CLs [39]60.173.758.7
MS-CMA [66]61.473.860.5
SRN [76]62.073.458.5
ICME [7]62.874.160.7
ASL [45]63.974.662.7
HLC†63.174.662.9
HLC+ASL†64.575.163.4
", + "bbox": [ + 560, + 506, + 826, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tively large, we run experiments one time. Here we compare our method with S-CLs [39], MS-CMA [66], SRN [76], ICME [7], and ASL [45]. For convenient comparison, we refer to the results of their original papers. Note that to further improve the performance on NUS-WIDE, we utilize ASL to replace the loss function of our method. We name the new method \"HLC+ASL\". Results are provided in Table 7, which demonstrate the effectiveness of our method on the real-world dataset.", + "bbox": [ + 496, + 614, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 768, + 617, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on the realistic problem of multi-label classification with noisy labels. We learn and utilize the label dependence among multiple labels to handle this problem. With the help of label dependence, a novel algorithm named HLC is proposed to correct noisy multiple labels to clean ones. We demonstrate the effectiveness of our algorithm both theoretically and empirically. For future", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1490", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work, we are interested in adapting HLC to other domains such as natural language processing and recommendation systems. We are also interested in promoting our algorithm to tackle instance-dependent label noise [73, 4, 78, 38] in multi-label classification.", + "bbox": [ + 76, + 90, + 472, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 180, + 243, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bo Han was supported by NSFC Young Scientists Fund No. 62006202 and Guangdong Basic and Applied Basic Research Foundation No. 2022A1515011652. Tongliang Liu is partially supported by ARC projects: FT220100318, DP220102121, LP220100527, LP220200949, and IC190100031.", + "bbox": [ + 76, + 205, + 470, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 309, + 174, + 325 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, pages 233-242, 2017. 2", + "[2] Dara Bahri, Heinrich Jiang, and Maya Gupta. Deep k-nn for noisy labels. In ICML, pages 540–550, 2020. 4", + "[3] Yingbin Bai and Tongliang Liu. Me-momentum: Extracting hard confident examples from noisily labeled data. In ICCV, pages 9312-9321, 2021. 2", + "[4] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. arXiv preprint arXiv:2001.03772, 2020. 9", + "[5] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In ICCV, pages 522-531, 2019. 4", + "[6] Zhaomin Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Learning graph convolutional networks for multi-label recognition and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3", + "[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME, pages 622-627, 2019. 1, 3, 5, 8", + "[8] Xiang Cheng, Shu-Guang Zhao, Xuan Xiao, and Kuo-Chen Chou. iatc-misf: a multi-label classifier for predicting the classes of anatomical therapeutic chemicals. Bioinformatics, 33(3):341-346, 2017. 1", + "[9] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. arXiv preprint arXiv:2109.04997, 2021. 1", + "[10] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In ACM International Conference on Image and Video Retrieval, pages 1-9,2009.8" + ], + "bbox": [ + 78, + 335, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Zijun Cui, Yong Zhang, and Qiang Ji. Label error correction and generation through label relationships. In AAAI, pages 3693-3700, 2020. 1", + "[12] Mark Everingham, Andrew Zisserman, Christopher KI Williams, Luc Van Gool, Moray Allan, Christopher M Bishop, Olivier Chapelle, Navneet Dalal, Thomas Deselaers, Gyuri Dorkó, et al. The pascal visual object classes challenge 2007 (voc2007) results. 2008. 4", + "[13] Lei Feng, Senlin Shu, Zhuoyi Lin, Fengmao Lv, Li Li, and Bo An. Can cross entropy loss be robust to label noise? In IJCAI, pages 2206-2212, 2021. 1", + "[14] Bin-Bin Gao and Hong-Yu Zhou. Learning to discover multi-class attentional regions for multi-label image recognition. IEEE Transactions on Image Processing, 30:5920-5932, 2021. 1, 4", + "[15] Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy labels. arXiv preprint arXiv:1607.07526, 2016. 4", + "[16] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor Tsang, and Masashi Sugiyama. *Sigua: Forgetting may make learning with noisy labels more robust.* In ICML, pages 4006–4016, 2020. 1", + "[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages 8527-8537, 2018. 5", + "[18] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3", + "[19] Jun-Yi Hang and Min-Ling Zhang. Dual perspective of label-specific feature learning for multi-label classification. In ICML, pages 8375-8386, 2022. 1", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 8", + "[21] Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In NeurIPS, 2018. 4", + "[22] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly supervised image classification through noise regularization. In CVPR, pages 11517-11525, 2019. 5", + "[23] Zhuo Huang, Xiaobo Xia, Li Shen, Bo Han, Mingming Gong, Chen Gong, and Tongliang Liu. Harnessing out-of-distribution examples via augmenting content and style. In ICLR, 2023. 1", + "[24] Zhuo Huang, Miaoxi Zhu, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, Bo Du, and Tongliang Liu. Robust generalization against photon-limited corruptions via worst-case sharpness minimization. In CVPR, pages 16175-16185, 2023. 1", + "[25] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, pages 2309-2318, 2018. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1491", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. Nlnl: Negative learning for noisy labels. In ICCV, pages 101-110, 2019. 1", + "[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[28] Cheng Li, Bingyu Wang, Virgil Pavlu, and Javed Aslam. Conditional bernoulli mixtures for multi-label classification. In ICML, pages 2482-2491, 2016. 3", + "[29] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. 5", + "[30] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In ICCV, pages 9485-9494, 2021. 5", + "[31] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, 2022. 1", + "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 4", + "[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. In NeurIPS, 2020. 3", + "[34] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2016. 2", + "[35] Weiwei Liu, Ivor W Tsang, and Klaus-Robert Müller. An easy-to-hard learning paradigm for multiple classes and multiple labels. Journal of Machine Learning Research, 18, 2017. 1", + "[36] Weiwei Liu, Haobo Wang, Xiaobo Shen, and Ivor Tsang. The emerging trends of multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 5", + "[37] Weiwei Liu, Donna Xu, Ivor W Tsang, and Wenjie Zhang. Metric learning for multi-output tasks. Transactions on Pattern Analysis and Machine Intelligence, 41(2):408-422, 2018. 1", + "[38] Yang Liu. Identifiability of label noise transition matrix. arXiv preprint arXiv:2202.02016, 2022. 9", + "[39] Yongcheng Liu, Lu Sheng, Jing Shao, Junjie Yan, Shiming Xiang, and Chunhong Pan. Multi-label image classification via knowledge distillation from weakly-supervised detection. In ACMMM, pages 700-708, 2018. 8", + "[40] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. In ICML, pages 6543-6553, 2020. 5", + "[41] Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. Modeling fine-grained entity types with box embeddings. In ACL, 2021. 1", + "[42] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4", + "[43] Cosmin Octavian Pene, Amirmasoud Ghiassi, Taraneh Younesian, Robert Birke, and Lydia Y Chen. Multi-label gold asymmetric loss correction with single-label regulators. arXiv preprint arXiv:2108.02032, 2021. 2", + "[44] Geoff Pleiss, Tianyi Zhang, Ethan R Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. In NeurIPS, 2020. 1", + "[45] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In ICCV, 2021. 1, 8", + "[46] Jun Shu, Qian Zhao, Zengben Xu, and Deyu Meng. Meta transition adaptation for robust deep learning with noisy labels. arXiv preprint arXiv:2006.05697, 2020. 2, 4", + "[47] Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1", + "[48] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, 2018. 5", + "[49] Alexander B Tsybakov. Optimal aggregation of classifiers in statistical learning. The Annals of Statistics, 32(1):135-166, 2004. 4", + "[50] Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks. In NeurIPS, pages 5596-5605, 2017. 1", + "[51] Andreas Veit, Neil Alldrin, Gal Chechik, Ivan Krasin, Abhinav Gupta, and Serge Belongie. Learning from noisy large-scale datasets with minimal supervision. In CVPR, 2017. 1", + "[52] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, pages 2285-2294, 2016. 1", + "[53] Xinshao Wang, Yang Hua, Elyor Kodirov, David A Clifton, and Neil M Robertson. Proselflc: Progressive self label correction for training robust deep neural networks. In CVPR, pages 752-761, 2021. 2", + "[54] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pages 8688-8696, 2018. 5", + "[55] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, pages 13726–13735, 2020. 1", + "[56] Hongxin Wei, Lue Tao, Renchunzi Xie, and Bo An. Open-set label noise can improve robustness against inherent label noise. In NeurIPS, pages 7978-7992, 2021. 1", + "[57] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed graph. In ICCV, pages 4157-4165, 2015. 3", + "[58] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hinding the memorization of noisy labels. In ICLR, 2021. 1, 5" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1492", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In ICLR, 2022. 1", + "[60] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 2", + "[61] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835-6846, 2019. 1", + "[62] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning. In AAAI, 2018. 3", + "[63] Ming-Kun Xie and Sheng-Jun Huang. Ccmn: A general framework for learning with class-conditional multi-label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 5", + "[64] Ming-Kun Xie, Jia-Hao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. In NeurIPS, 2022. 1", + "[65] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, pages 649-665, 2020. 1, 2, 3, 4, 5", + "[66] Renchun You, Zhiyao Guo, Lei Cui, Xiang Long, Yingze Bao, and Shilei Wen. Cross-modality attention with semantic graph embedding for multi-label classification. In AAAI, pages 12709-12716, 2020. 8", + "[67] Hsiang-Fu Yu, Prateek Jain, Purushottam Kar, and Inderjit Dhillon. Large-scale multi-label learning with missing labels. In ICML, pages 593-601, 2014. 3", + "[68] Ze-Bang Yu and Min-Ling Zhang. Multi-label classification with label-specific feature generation: A wrapped approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3", + "[69] Jiong Zhang, Wei-cheng Chang, Hsiang-fu Yu, and Inderjit Dhillon. Fast multi-resolution transformer fine-tuning for extreme multi-label text classification. In NeurIPS, 2021. 1", + "[70] Min-Ling Zhang and Kun Zhang. Multi-label learning by exploiting label dependency. In KDD, pages 999-1008, 2010. 1, 3", + "[71] Min-Ling Zhang and Zhi-Hua Zhou. A review on multi-label learning algorithms. IEEE Transactions on Knowledge and Data Engineering, 26(8):1819-1837, 2013. 1, 5", + "[72] Yivan Zhang, Gang Niu, and Masashi Sugiyama. Learning noise transition matrix from only noisy labels via total variation regularization. In ICML, 2021. 4", + "[73] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In ICLR, 2021. 9", + "[74] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In ICCV, pages 163-172, 2021. 1, 4", + "[75] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "correction of noisy labels. In ICML, pages 11447-11457, 2020. 1, 4", + "[76] Feng Zhu, Hongsheng Li, Wanli Ouyang, Nenghai Yu, and Xiaogang Wang. Learning spatial regularization with image-level supervisions for multi-label image classification. In CVPR, pages 5513-5522, 2017. 8", + "[77] Ke Zhu and Jianxin Wu. Residual attention: A simple but effective method for multi-label recognition. In ICCV, pages 184-193, 2021. 1, 4, 5", + "[78] Zhaowei Zhu, Tongliang Liu, and Yang Liu. A second-order approach to learning with instance-dependent label noise. In CVPR, pages 10113-10123, 2021. 9" + ], + "bbox": [ + 501, + 92, + 890, + 262 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1493", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_model.json b/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b94aa9b9f90e626cd9c30168a058f82bc043b421 --- /dev/null +++ b/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.173, + 0.131, + 0.799, + 0.152 + ], + "angle": 0, + "content": "Holistic Label Correction for Noisy Multi-Label Classification" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.18, + 0.718, + 0.216 + ], + "angle": 0, + "content": "Xiaobo Xia\\(^{1}\\), Jiankang Deng\\(^{2}\\), Wei Bao\\(^{1}\\), xuan Du\\(^{3}\\), Bo Han\\(^{4}\\), Shiguang Shan\\(^{5,6}\\), Tongliang Liu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.217, + 0.707, + 0.234 + ], + "angle": 0, + "content": "1The University of Sydney 2Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.234, + 0.71, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\)JD Explore Academy \\(^{4}\\)Hong Kong Baptist University" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.252, + 0.798, + 0.27 + ], + "angle": 0, + "content": "\\(^{5}\\)Chinese Academy of Sciences \\(^{6}\\)University of Chinese Academy of Sciences" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.473, + 0.73 + ], + "angle": 0, + "content": "Multi-label classification aims to learn classification models from instances associated with multiple labels. It is pivotal to learn and utilize the label dependence among multiple labels in multi-label classification. As a result of today's big and complex data, noisy labels are inevitable, making it looming to target multi-label classification with noisy labels. Although the importance of label dependence has been shown in multi-label classification with clean labels, it is challenging and hard to bring label dependence to the problem of multi-label classification with noisy labels. The issues are, that we do not understand why label dependence is helpful in the problem, and how to learn and utilize label dependence only using training data with noisy multiple labels. In this paper, we bring label dependence to tackle the problem of multi-label classification with noisy labels. Specifically, we first provide a high-level understanding of why label dependence helps distinguish the examples with clean/noisy multiple labels. Benefiting from the memorization effect in handling noisy labels, a novel algorithm is then proposed to learn the label dependence by only employing training data with noisy multiple labels, and utilize the learned dependence to help correct noisy multiple labels to clean ones. We prove that the use of label dependence could bring a higher success rate for recovering correct multiple labels. Empirical evaluations justify our claims and demonstrate the superiority of our algorithm." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Multi-label classification assigns a set of multiple labels for each instance [71]. As a practical learning paradigm, multi-label classification has been widely applied in various domains, ranging from computer vision [7] and natural language processing [41], to recommendation systems [69] and bioinformatics [8]. Consensually, compared" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.892, + 0.382 + ], + "angle": 0, + "content": "with multi-class classification [20, 23, 24], where each instance is assigned with a single label, multi-label classification is more challenging [35]. Plenty of advanced methods are proposed in recent years for multi-label classification [77, 45, 14, 74, 37, 9, 19, 64]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.893, + 0.686 + ], + "angle": 0, + "content": "The great majority of the methods assume that training data are annotated precisely. However, noisy labels are inevitable in multi-label classification [36], especially for classification with big and complex data. They may be resulted by unintentional mistakes of manual and automatic annotators [51, 75, 13], or intentional corruptions on clean labels [50, 44]. Noisy labels severely impair the generalization of learned models, over-parameterized deep models in particular [26, 61, 58, 59, 55, 56]. A straightforward way to address the problem of multi-label classification with noisy labels is to treat each label in isolation and convert the multi-label problem into a number of binary classification problems. Afterward, the methods in multi-class classification with noisy labels [16, 47] are applied to train independent binary classifiers, which capture instance-label dependence robustly to strengthen classification. This way is a remedy to handle noisy labels, but ignores the label dependence among multiple labels. It is essential to learn and utilize the label dependence in multi-label classification [70, 18, 11, 31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Prior works [65, 6, 52] illustrate the successes of considering the label dependence among multiple labels in multilabel classification with clean labels. In different ways, e.g., helping learn inter-dependent classifiers [7], the label dependence can be used to boost the learning of the instance-label dependence, which improves final classification. Inspired by the successes, it is concerned that label dependence could be exploited to handle the problem of multilabel classification with noisy labels. However, there are few attempts before for this important problem. At least three questions make the solution remain mysterious. First, in intuition, we need to understand why label dependence is helpful for the problem. Second, in technique, we need to know how to learn and utilize the label dependence in the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.388, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (tongliang.liu@sydney.edu.au)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.092, + 0.835, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 1. The illustration of why the label dependence among multiple labels helps distinguish the examples with noisy/clean multiple labels. The arrow presents the label dependence between a label pair. For the labels “a” and “b”, “a → b” means that, when “a” appears, “b” will also occur with high probability. The example comes from a web search. The set of clean multiple labels is {Sea, Human, Motorboat}, where the label dependence is strong with both “Motorboat → Sea” and “Motorboat → Human”. However, due to label corruption, Motorboat is flipped to be Motorcycle, which causes “Motorcycle → Sea”. Therefore, the label dependence among noisy multiple labels is weaker than the label dependence among corresponding clean ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.385 + ], + "angle": 0, + "content": "problem. As we only have training data with noisy labels, both the accurate catch and application of the label dependence are challenging. Third, in verification, we need to know what improvements the label dependence can bring." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.471, + 0.809 + ], + "angle": 0, + "content": "In this paper, we answer the three questions one by one. The first answer is illustrated in Figure 1. That is, compared with noisy multiple labels, the label dependence among clean multiple labels is stronger with high probability. Therefore, such dependence could help distinguish the examples with noisy/clean multiple labels for our problem. The second answer is given by the proposed holistic correction for multi-label classification with noisy labels (aka HLC). Specifically, HLC inherits the memorization effect in handling noisy labels [1, 25, 53]: the deep model would firstly memorize the training examples with clean labels, leading to reliable model predictions in early training. In HLC, the label dependence is learned by a dynamic graph [65], and then applied to correcting noisy multiple labels. In more detail, the holistic score in HLC is proposed to measure the instance-label and label dependencies in an example. The stronger instance-label and label dependencies make a larger holistic score. We compare the ratio between the holistic scores of the example with noisy multiple labels and its variant with predicted multiple labels, with an easily determined threshold. The noisy multiple labels are corrected or changeless based on the comparison result. Benefiting from the memorization effect, both dependence learning and multi-label correction are useful. Besides, they fulfill a positive cycle [3]. Namely, better dependence learning results in a better multi-label correction, and better multi-label correction makes better dependence learning, leading to final enhanced classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The third answer is given by both theoretical analyses and empirical evaluations. Theoretically, we show that the additional use of label dependence brings a higher probability to handle noisy multiple labels successfully than the sole use of instance-label dependence under some conditions. Empirically, we demonstrate the power of label de" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.369 + ], + "angle": 0, + "content": "pendence through experiments and show that, in most situations, HLC outperforms comparison methods with large margins." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.893, + 0.612 + ], + "angle": 0, + "content": "The contributions of this paper are summarized as follows. (1) We focus on a realistic problem of multi-label classification with noisy labels. The challenges of using label dependency to address the problem are carefully analyzed, which benefits future research on the problem. (2) We propose an effective method to handle noisy labels in multi-label classification. The method measures simultaneous instance-label and label dependencies in an example for follow-up label correction. (3) Theoretical analysis is provided to explain the success rate of the proposed method. Besides, we confirm that the use of label dependence is indeed powerful under some conditions. (4) Extensive empirical results on multiple benchmarks demonstrate the superiority of our method. Detailed ablation studies and discussions are also provided. Codes are attached in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.638, + 0.644 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Problem setup. Let \\(\\mathcal{X} \\in \\mathbb{R}^d\\) denote the input space and \\(\\mathcal{Y} \\in \\{l_1, \\dots, l_q\\}\\) denote the label space with \\(q\\) class labels. An example with multiple labels is denoted as \\((x, y)\\), where \\(x \\in \\mathcal{X}\\) is the feature vector of an instance, and \\(y \\subseteq \\mathcal{Y}\\) is its set of associated labels. Denote the size of the label set \\(y\\) as \\(|y|\\). For the feature vector \\(x\\), its label set \\(y\\) may be corrupted and is flipped into \\(\\bar{y} \\subseteq \\mathcal{Y}\\). We utilize a class-dependent noise transition matrix \\(\\pmb{T}\\) [43, 46, 34, 60] to characterize the label flip process. Formally, for any \\(i \\neq j\\), \\(T_{ij} = \\mathbb{P}(l_j \\in \\bar{y} \\wedge l_i \\notin \\bar{y} | l_j \\notin y \\wedge l_i \\in y)\\) represents the probability of the \\(i\\)-th class label to be flipped into the \\(j\\)-th class label. Consider a noisy multi-label dataset comprising several examples \\((x, \\bar{y})\\). The aim is to learn a classification model robustly by only using the noisy dataset. Given an instance in testing, with the learned model, we can predict its relevant label set precisely." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1484" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Note that some works employ another problem setting that the total number of multiple labels can be changed after label flipping, which is referred to as multi-label classification with missing or redundant labels. For the former setting, it is not accurate to consider it as a classification with noisy labels, since all annotated labels are correct [67, 57]. For the latter setting, it is normally called partial multi-label learning [62], which is different from the problem setting of this paper, as detailed in Appendix C.4. Our setting, i.e., the total number of labels is preserved after label flipping, is realistic. In many practical situations, it is easy to determine the number of objects in an image, in particular with object detection techniques. In contrast, it can be harder to annotate the objects perfectly, resulting in noisy labels. In addition, in Section 4.4, we will show that our problem setting well fits the realistic situation. That is to say, the proposed method can achieve superior performance on a realistic noisy multi-label dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.472, + 0.489 + ], + "angle": 0, + "content": "Preparation technology. As discussed, we need both the instance-label dependence and the label dependence among multiple labels. Given an example \\((\\pmb{x},\\pmb{y})\\), for the instance-label dependence, it can be learned with the conditional probability of \\(l_{i}\\in \\pmb{y}\\) given \\(\\pmb{x}\\) according to model's probability outputs. For the label dependence among multiple labels, it is often estimated by counting the occurrence of label pairs in training data [7]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.473, + 0.761 + ], + "angle": 0, + "content": "Recently, the graph convolutional network (GCN) is used in multi-label classification and achieves great successes [7, 65, 6]. The advantage of the GCN-based methods is that they can capture the instance-label and label dependencies simultaneously during training. In this paper, we inherit the advantage of the GCN-based methods and build HLC based on ADDGCN [65]. ADDGCN designs a semantic attention module (SAM) to estimate the content-aware class-label representations for each class from the extracted feature map. The representations are fed into a GCN module (GCNM) for final classification. We provide the technical details of ADDGCN [65] in Appendix C.1. Before delving into the next section, readers only need to remember that the instance-label and label dependencies can be learned during training. Note that we also review prior works on multi-class classification with noisy labels and multi-label classification with clean/noisy labels in Appendix C.2 and Appendix C.3." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.773, + 0.251, + 0.79 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.797, + 0.47, + 0.814 + ], + "angle": 0, + "content": "3.1. Holistic Judgment in Multi-Label Classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Holistic score. We begin with an example with clean multiple labels. Given an example \\((\\pmb{x},\\pmb{y})\\), we can measure the instance-label dependence \\(S^f\\), and the label dependence \\(S^l\\). Denote the variable of clean multiple labels by \\(\\pmb{Y}\\). Mathematically, we define two dependen" + }, + { + "type": "code_caption", + "bbox": [ + 0.508, + 0.094, + 0.74, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Holistic Correction." + }, + { + "type": "algorithm", + "bbox": [ + 0.508, + 0.112, + 0.778, + 0.212 + ], + "angle": 0, + "content": "Input: \\((\\pmb {x},\\bar{\\pmb{y}})\\) \\(h\\) ,and \\(\\hat{\\delta}\\) \nOutput: \\(\\bar{y}_{new}\\) \n1: \\(\\pmb {y}^{*} = h(\\pmb {x})\\) . \n2: \\(\\kappa (h,\\pmb {x},\\bar{\\pmb{y}}) = \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}(\\pmb {x}) / \\hat{\\bar{S}}_{\\pmb{y}^{*}}(\\pmb {x})\\) \n3: if \\(\\kappa (h,\\pmb {x},\\bar{\\pmb{y}})\\leq \\hat{\\delta}\\) then \\(\\bar{y}_{new} = y^{*}\\) \n4: else \\(\\bar{y}_{new} = \\bar{y}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.895, + 0.336 + ], + "angle": 0, + "content": "cies as \\( S_{\\pmb{z}}^{f}(\\pmb{x}) \\coloneqq \\sum_{\\{\\pmb{Y} = \\pmb{z}, l_{i} \\in \\pmb{z}\\}} \\mathbb{P}(l_{i}|\\pmb{x}) \\) and \\( S_{\\pmb{z}}^{l}(\\pmb{x}) \\coloneqq \\sum_{\\{\\pmb{Y} = \\pmb{z}, l_{i}, l_{j} \\in \\pmb{z}\\}} \\frac{1}{2} [\\mathbb{P}(l_{j}|l_{i}, \\pmb{x}) + \\mathbb{P}(l_{i}|l_{j}, \\pmb{x})] \\), where \\( \\pmb{z} \\) is the value of the random variable \\( \\pmb{Y} \\). The holistic score of the example \\( (\\pmb{x}, \\pmb{y}) \\) considers two dependencies at the same time. Formally, we denote the holistic score of \\( (\\pmb{x}, \\pmb{y}) \\) as \\( S_{\\pmb{y}}(\\pmb{x}) \\) and define it as" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.343, + 0.892, + 0.363 + ], + "angle": 0, + "content": "\\[\nS _ {\\boldsymbol {y}} (\\boldsymbol {x}) := S _ {\\boldsymbol {y}} ^ {f} (\\boldsymbol {x}) + S _ {\\boldsymbol {y}} ^ {l} (\\boldsymbol {x}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.895, + 0.609 + ], + "angle": 0, + "content": "Afterward, denote the variable of noisy multiple labels by \\(\\bar{Y}\\). For the example with noisy multiple labels, i.e., \\((\\pmb{x},\\bar{\\pmb{y}})\\), the instance-label dependence and label dependence are measure by \\(\\bar{S}_z^f (\\pmb {x})\\coloneqq \\sum_{\\{\\bar{Y} = \\pmb {z},l_i\\in \\pmb {z}\\}}\\mathbb{P}(l_i|\\pmb {x})\\) and \\(\\bar{S}_z^l (\\pmb {x})\\coloneqq\\) \\(\\sum_{\\{\\bar{Y} = z,l_i,l_j\\in \\pmb {z}\\}}\\frac{1}{2}\\left[\\mathbb{P}(l_j|l_i,\\pmb {x}) + \\mathbb{P}(l_i|l_j,\\pmb {x})\\right]\\). Accordingly, the holistic score of the example \\((\\pmb {x},\\bar{\\pmb{y}})\\) is denoted by \\(\\bar{S}_{\\bar{\\pmb{y}}}(\\pmb {x})\\), which is defined as \\(\\bar{S}_{\\bar{\\pmb{y}}}(\\pmb {x})\\coloneqq \\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x}) + \\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})\\). Note that, during training, we cannot access \\(\\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})\\) and \\(\\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})\\). Instead, the estimated posterior probabilities are used. We denote the estimations of \\(\\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})\\) and \\(\\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})\\) as \\(\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})\\) and \\(\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})\\). The estimation of the holistic score is \\(\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}(\\pmb {x}) = \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x}) + \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})\\). With preparation technology discussed in Section 2 and Appendix C.1, \\(\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})\\) and \\(\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})\\) can be obtained." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Holistic correction. For the example \\((\\pmb{x},\\bar{\\pmb{y}})\\), we feed it into the deep network \\(h\\) included in ADDGCN [65]. The memorization effect in handling noisy labels [25, 33] shows that the deep network would first memorize the training data with clean labels and then the training data with noisy labels. Therefore, early in training, the outputs of the deep network are relatively reliable and can be used for label correction. For \\((\\pmb{x},\\bar{\\pmb{y}})\\), we denote its set of predicted multiple labels as \\(\\pmb{y}^*\\). Here, the set of predicted labels is obtained with the top \\(|\\bar{\\pmb{y}}|\\) predictions based on the model's probability outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recall that the holistic score of an example holistically measures the instance-label dependence and label dependence among multiple labels simultaneously. From both human and machine cognition, if an example is annotated accurately, both dependencies should be strong [70, 18, 68, 28, 7] with high probability. Namely, the holistic score is large. We propose to check the ratio between the holistic score on \\((\\pmb{x},\\bar{\\pmb{y}})\\) and holistic score on \\((\\pmb{x},\\pmb{y}^{*})\\). Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1485" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.144, + 0.105 + ], + "angle": 0, + "content": "we check" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.103, + 0.469, + 0.124 + ], + "angle": 0, + "content": "\\[\n\\kappa (h, \\boldsymbol {x}, \\bar {\\boldsymbol {y}}) = \\hat {\\tilde {S}} _ {\\bar {\\boldsymbol {y}}} (\\boldsymbol {x}) / \\hat {\\tilde {S}} _ {\\boldsymbol {y} ^ {*}} (\\boldsymbol {x}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.13, + 0.47, + 0.251 + ], + "angle": 0, + "content": "We compare this ratio with a predetermined threshold \\(\\hat{\\delta}\\). The value of \\(\\hat{\\delta}\\) is given in the next subsection. If \\(\\kappa(h, \\boldsymbol{x}, \\bar{\\boldsymbol{y}}) \\leq \\hat{\\delta}\\), we flip the labels \\(\\bar{\\boldsymbol{y}}_{\\text{new}} = \\boldsymbol{y}^*\\). Otherwise, the labels remain unchanged with \\(\\bar{\\boldsymbol{y}}_{\\text{new}} = \\bar{\\boldsymbol{y}}\\). The detailed algorithm of holistic correction for multi-label classification with noisy labels (aka HLC) is provided in Algorithm 1. After holistic correction for noisy labels, we use \\((\\boldsymbol{x}, \\bar{\\boldsymbol{y}}_{\\text{new}})\\) to train the deep network \\(h\\) based on ADDGCN [65]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.26, + 0.267, + 0.276 + ], + "angle": 0, + "content": "3.2. Theoretical Insights" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.47, + 0.449 + ], + "angle": 0, + "content": "We extend the Tsybakov condition [75, 2, 15, 49] from multi-class classification to multi-label classification. Specifically, denote by \\( \\pmb{a}_{\\pmb{x}} \\) the label set predicted based on \\( S^{f}(\\pmb{x}) \\) with \\( \\pmb{a}_{\\pmb{x}} := h^{*}(\\pmb{x}) = \\arg \\max_{\\pmb{z}} S_{\\pmb{z}}^{f}(\\pmb{x}) \\). Besides, denote by \\( \\pmb{b}_{\\pmb{x}} \\) the second best prediction with \\( \\pmb{b}_{\\pmb{x}} := \\arg \\max_{\\pmb{z} \\neq \\pmb{a}_{\\pmb{x}}} S_{\\pmb{z}}^{f}(\\pmb{x}) \\). The maximum length of a label set is denoted as \\( m \\) (\\( m \\ll q \\)). In this paper, we call the predicted label set by the Bayes optimal classifier for an instance as the correct label set. We present the Tsybakov condition on instance-label (abbreviated as ins.-label here) dependence and holistic Tsybakov condition as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.488, + 0.506 + ], + "angle": 0, + "content": "Definition 1 (Tsybakov condition on ins.-label dependence) \\(\\exists C_1, \\lambda_1 > 0\\) and \\(\\exists t_0 \\in (0, m]\\), such that for all \\(t \\leq t_0\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.506, + 0.469, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left[ S _ {\\boldsymbol {a} _ {\\boldsymbol {x}}} ^ {f} (\\boldsymbol {x}) - S _ {\\boldsymbol {b} _ {\\boldsymbol {x}}} ^ {f} (\\boldsymbol {x}) \\leq t \\right] \\leq C _ {1} t ^ {\\lambda_ {1}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.469, + 0.567 + ], + "angle": 0, + "content": "Definition 2 (Holistic Tsybakov condition) \\(\\exists C_2, \\lambda_2 > 0\\) and \\(\\exists t_0 \\in (0, m]\\), such that for all \\(t \\leq t_0\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.578, + 0.469, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left[ S _ {\\boldsymbol {a} _ {\\boldsymbol {x}}} (\\boldsymbol {x}) - S _ {\\boldsymbol {b} _ {\\boldsymbol {x}}} (\\boldsymbol {x}) \\leq t \\right] \\leq C _ {2} t ^ {\\lambda_ {2}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.469, + 0.668 + ], + "angle": 0, + "content": "Remark 1 Definition 1 stipulates that the uncertainty of \\( S^f \\) is bounded. The margin region that is close to the decision boundary has a bounded volume. Definition 2 shares the similar idea and bound the uncertainty of \\( S \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.498, + 0.762 + ], + "angle": 0, + "content": "Theorem 1 Suppose \\( S(\\pmb{x}) \\) fulfills the holis-tic Tsybakov condition for constants \\( C_2, \\lambda_2 > 0 \\), and \\( t_0 \\in (0, m] \\). We define \\( \\epsilon := \\max_{\\pmb{x}, \\pmb{z}} \\left[ |\\hat{\\bar{S}}_z^f(\\pmb{x}) - \\bar{S}_z^f(\\pmb{x})|, |\\hat{\\bar{S}}_z^l(\\pmb{x}) - \\bar{S}_z^l(\\pmb{x})|, |\\bar{S}_z^l(\\pmb{x}) - S_z^l(\\pmb{x})| \\right] \\) and \\( \\tau := \\min_i T_{ii} \\). We analyze two cases:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.856 + ], + "angle": 0, + "content": "(1) If \\(\\bar{\\pmb{y}}\\) is corrected by \\(\\kappa(h, \\pmb{x}, \\bar{\\pmb{y}})\\) with \\(\\hat{\\delta}\\), let \\(\\delta_1 = \\min \\left[\\frac{\\tau S_{\\pmb{b}_\\pmb{x}}(\\pmb{x}) + \\sum_{l_j \\in \\bar{\\pmb{y}}} \\sum_{i \\neq j} T_{ij} \\mathbb{P}(l_i | \\pmb{x})}{\\hat{S}_{\\pmb{y}^*}(\\pmb{x})}\\right]\\) and \\(\\rho_1 := |\\hat{\\delta} - \\delta_1|\\). Assume that \\(\\epsilon \\leq \\frac{t_0 \\tau - \\rho_1 m}{3}\\). Then, \\(\\mathbb{P}[\\bar{\\pmb{y}}_{new} = h^*(\\pmb{x}), \\bar{\\pmb{y}}\\) is flipped] is at least \\(1 - C_2 [O(\\max(\\epsilon, \\rho_1))]^{\\lambda_2} - \\mathbb{P}[\\pmb{a}_x \\neq \\{\\pmb{y}^*, \\bar{\\pmb{y}}\\}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.857, + 0.469, + 0.905 + ], + "angle": 0, + "content": "(2) If \\(\\bar{\\pmb{y}}\\) is not corrected by \\(\\kappa(h, \\pmb{x}, \\bar{\\pmb{y}})\\) with \\(\\hat{\\delta}\\), let \\(\\delta_2 = \\max \\left[\\frac{\\hat{\\tilde{S}}_{\\bar{\\pmb{y}}}(\\pmb{x})}{\\tau S_{b_{\\pmb{x}}}(\\pmb{x}) + \\sum_{l_j \\in \\pmb{y}^*} \\sum_{i \\neq j} T_{ij} \\mathbb{P}(l_i | \\pmb{x})}\\right]\\) and \\(\\rho_2 := |\\hat{\\delta} - \\delta_2|\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Assume that \\(\\epsilon \\leq \\frac{t_0\\delta_2^2\\tau - \\rho_2m - \\rho_2^2m}{3\\delta_2^2}\\). Then, \\(\\mathbb{P}[\\bar{\\pmb{y}}_{new} = h^{*}(\\pmb{x}),\\bar{\\pmb{y}}\\) is accepted] is at least \\(1 - C_2[O(\\max (\\epsilon ,\\rho_2))]^{\\lambda_2} - \\mathbb{P}[\\pmb {a}_x\\neq \\{\\pmb {y}^*,\\bar{\\pmb{y}}\\} ]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.892, + 0.336 + ], + "angle": 0, + "content": "The proof of Theorem 1 is provided in Appendix B.1. Theorem 1 extends the theoretical results of [75] to multilabel classification with noisy labels. It claims that, even though with noisy multiple labels, the holistic correction has a guaranteed success rate to make proper corrections. Besides, if we can reasonably approximate the optimal \\(\\delta\\) with \\(\\hat{\\delta}\\), our algorithm flips noisy multiple labels to correct ones with a good chance. Below, as a corollary of Theorem 1, we show that, there are certain circumstances, the use of holistic scores has a better chance to make corrections satisfactorily, than the sole use of instance-label dependence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.441 + ], + "angle": 0, + "content": "Corollary 1 Suppose that \\( S(\\pmb{x}) \\) fulfills the holistic Tsybakov condition. Denote the set threshold \\( \\hat{\\delta} \\) and optimal threshold \\( \\delta \\). We define \\( \\rho := \\max |\\hat{\\delta} - \\delta| \\). We have that, \\( \\exists \\epsilon \\) and \\( \\rho \\), if \\( C_2[O(\\max(\\epsilon, \\rho))]^{\\lambda_2} < C_1[O(\\max(\\epsilon, \\rho))]^{\\lambda_1} \\), holistic correction brings higher probability to handle noisy labels successfully than instance-label dependence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.589 + ], + "angle": 0, + "content": "The proof of Corollary 1 is provided in Appendix B.2. Corollary 1 claims that there exist cases where holistic scores better combat noisy labels. Note that, from a theoretical view, we do not state that holistic scores can work better in all circumstances of multi-label classification. Nevertheless, with the determination of the threshold \\(\\hat{\\delta}\\), holistic scores can perform better in the experiments of this paper, which demonstrates the help of label dependence to handle noisy multiple labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.634, + 0.619 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.691, + 0.644 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Datasets. We verify the effectiveness of the proposed method on the synthetic noisy versions of three datasets, i.e., Pascal-VOC 2007 [12], Pascal-VOC 2012 [12], and MS-COCO [32]. Pascal-VOC 2007 contains 5,011 images in train and validation sets, while Pascal-VOC 2012 consists of 11,540 images in train and validation sets. The images come from 20 common object categories. For Pascal-VOC 2007 and Pascal-VOC 2012, we train methods using the noisy training and validation sets, and evaluate them on the test set of Pascal-VOC 2007 that has 4,952 images [14]. MS-COCO contains 82,081 training images and 40,137 validation images from 80 common object categories. As did in [74, 5, 65, 77], we evaluate the performance of methods using validation images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Noisy-label generation. The class-dependent noise transition matrix \\( T \\) [42, 21, 46, 72] is used to corrupt the three" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1486" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.088, + 0.885, + 0.101 + ], + "angle": 0, + "content": "Table 1. Comparisons with advanced methods on noisy Pascal-VOC 2007. The mean and standard deviation of results (%) are presented." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.101, + 0.818, + 0.447 + ], + "angle": 0, + "content": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE64.50±1.2058.65±2.1648.19±0.2371.77±1.1560.94±4.2548.72±2.13
CSRA66.99±0.4859.62±0.6146.97±0.4872.45±0.6963.58±1.4852.72±1.52
ADDGCN63.89±0.9455.75±1.9844.14±1.3771.02±0.9561.05±0.0650.18±2.70
APL66.79±1.1958.86±1.5347.64±1.8172.61±0.9961.99±0.7849.10±0.15
CDR67.35±1.7060.05±1.0649.12±0.5972.66±0.7964.58±0.6050.51±2.49
JOINT67.43±0.7363.37±0.9253.27±4.7070.28±1.8568.70±2.8858.57±2.75
WSIC65.43±0.5559.53±0.7348.34±0.4772.57±1.0361.88±2.5750.15±0.86
CCMN69.97±1.3662.58±1.4753.20±1.2870.68±1.0860.94±3.1248.62±1.26
HLC†72.07±0.6770.20±0.4668.00±0.8974.83±0.6469.86±1.6160.09±1.73
OF1 ↑BCE63.52±0.4856.70±2.4548.10±1.4368.28±0.6958.30±2.8251.18±3.10
CSRA65.40±0.4759.39±0.8148.32±1.5069.72±0.5061.89±0.4351.56±2.28
ADDGCN62.63±0.1855.50±1.8744.38±2.9268.95±0.6459.64±0.5653.12±0.62
APL64.85±1.4656.51±1.7047.54±2.4068.89±0.8958.04±0.9752.27±2.20
CDR65.31±0.9957.93±1.0548.86±1.7169.53±0.6559.89±1.0751.68±3.83
JOINT69.72±0.8867.93±0.7761.62±1.4071.24±1.0364.20±0.8860.30±1.24
WSIC63.45±0.9757.96±1.2548.38±2.4169.88±1.2257.97±2.1951.99±1.65
CCMN69.66±1.5560.43±1.3153.84±0.6967.12±0.6159.55±1.4553.46±1.04
HLC†71.03±0.3369.08±1.0068.62±0.4872.09±0.7465.76±2.3960.71±1.37
CF1 ↑BCE58.91±1.3453.21±2.0443.66±0.5365.93±0.8157.03±3.4347.21±1.89
CSRA62.31±0.5055.67±0.6143.11±0.7667.39±0.8059.66±1.0451.13±1.12
ADDGCN60.41±1.0453.72±1.3842.42±0.5966.05±0.9757.81±0.5848.89±2.64
APL60.23±1.5352.85±2.1842.38±1.6766.59±0.7158.33±0.4947.67±1.83
CDR61.37±1.4754.17±0.8643.60±0.8267.11±0.6359.91±0.3948.40±1.98
JOINT63.13±0.3860.22±1.6848.17±5.0166.03±1.2562.05±2.9854.03±3.17
WSIC59.54±1.1054.22±0.5343.82±0.6266.97±1.0058.04±1.7048.19±0.96
CCMN65.19±1.1058.55±1.3149.85±1.0665.47±0.9358.05±2.2448.46±0.80
HLC†68.87±0.1066.62±0.8164.82±0.4869.95±1.1965.13±1.0457.54±1.84
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.593 + ], + "angle": 0, + "content": "datasets. Here, for any \\(i \\neq j\\), \\(T_{ij} = \\mathbb{P}(l_j \\in \\bar{\\mathbf{y}} \\wedge l_i \\notin \\bar{\\mathbf{y}} | l_j \\notin \\mathbf{y} \\wedge l_i \\in \\mathbf{y})\\) represents the probability of the \\(i\\)-th class label to be flipped into the \\(j\\)-th class label. We consider both symmetric (abbreviated as Sym.) and pairflip (abbreviated as Pair.) noise settings [17]. The details of the transition matrix are provided in Appendix D.2. For symmetric noise, the noise rate is set to \\(30\\%\\), \\(40\\%\\), and \\(50\\%\\). For pairflip noise, the noise rate is set to \\(20\\%\\), \\(30\\%\\), and \\(40\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Baselines. We exploit three types of baselines in total. Specifically, Type-I baselines contain the methods that are designed for multi-label classification with clean labels. Type-II baselines consider the methods for multi-class classification with noisy labels. Type-III baselines consider the methods that focus on multi-label classification with noisy labels. It should be noted that, there are relatively few methods belonging to this type [36]. More advanced methods belonging to Type-III baselines need to be investigated [36], which is also our focus in this paper. In more detail, Type-I baselines include CSRA [77] and ADDGCN [65]. Type-II baselines include APL [40], CDR [58], and JOINT [48]. Type-III baselines include WSIC [22] and CCMN [63]. As a simple baseline, we compare our method with the standard deep network that directly trains on noisy datasets (abbreviated as BCE). We detail all baselines in Appendix D.1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Network & Optimizer. We use a ResNet-50 network [20] pretrained on ImageNet as the backbone for all methods. We train the models for 30 epochs in total. We utilize" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Adam [27] for the network optimization. The batch size is set to 128 for all the datasets. The learning rate is fixed to \\(5 \\times 10^{-5}\\). The images in Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO resize to \\(224 \\times 224\\). Note that, to make experiments more comprehensive, we also employ different experimental settings, e.g., different networks and different image sizes. The details are provided in Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.825 + ], + "angle": 0, + "content": "Measurement. As did in multi-label classification [77, 7], evaluation metrics include the mean average precision (mAP) [71], the average F1-measure (OF1), and the average per-class F1-measure (CF1). For fair comparison, we implement all methods with default parameters by PyTorch, and conduct all experiments on NVIDIA GTX3090 GPUs. All experiments are repeated three times with different random seeds. Following the works in learning with noisy labels [17, 54, 29, 30], the mean and standard deviation of results in the last epoch are reported. In addition, for different evaluation metrics, we report the mean and standard deviation of best results. Supplementary results are shown in Appendix E. Afterwards, the best mean results are highlighted in red. The second best mean results are also highlighted in blue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.833, + 0.849 + ], + "angle": 0, + "content": "4.2. Comparison with the State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO are shown in Table 1, Table 2, and Table 3 respectively. In summary, HLC consistently works" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1487" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.088, + 0.885, + 0.101 + ], + "angle": 0, + "content": "Table 2. Comparisons with advanced methods on noisy Pascal-VOC 2012. The mean and standard deviation of results (%) are presented." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.102, + 0.818, + 0.447 + ], + "angle": 0, + "content": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE66.74±0.8056.07±0.5045.15±1.5670.91±1.1357.61±1.1449.85±0.36
CSRA66.35±0.5056.20±1.3545.54±1.1471.29±0.8360.71±1.1847.63±1.56
ADDGCN63.34±0.9654.54±0.8644.88±1.7170.41±0.5457.96±0.6847.66±1.08
APL67.07±1.0456.79±1.8643.51±1.9371.32±1.6059.59±1.2748.14±1.16
CDR66.13±1.4956.85±0.4844.84±1.1171.55±1.8760.13±1.8949.44±1.81
JOINT65.19±2.1758.40±2.8745.13±1.6968.93±2.5461.64±1.7853.64±1.61
WSIC65.96±0.7956.34±0.4144.80±0.5470.40±1.1159.40±1.8748.95±1.34
CCMN69.15±0.6661.00±1.0150.71±0.2669.08±1.7859.72±2.3246.67±2.78
HLC†72.14±0.6670.11±0.2768.69±1.0474.51±0.6769.90±0.4364.20±1.26
OF1 ↑BCE64.99±1.1056.92±2.0845.49±2.2368.48±2.2860.21±1.3554.05±1.95
CSRA64.08±0.3756.25±2.5748.67±3.1469.06±0.6559.75±1.7052.89±0.95
ADDGCN63.53±1.4154.28±0.8647.56±2.6747.62±2.3957.90±1.7852.33±0.56
APL64.70±1.1758.05±1.6845.74±1.5570.68±1.0360.22±1.5851.38±1.55
CDR64.06±1.3857.31±1.2146.51±0.9570.45±1.4460.57±1.2452.26±2.42
JOINT67.35±1.8664.57±2.3954.37±3.3370.81±1.4064.40±1.7656.27±1.29
WSIC62.74±2.1057.13±0.7345.52±1.2869.72±1.1959.11±2.0452.49±1.38
CCMN65.77±0.2359.91±0.9351.45±0.9467.93±1.7359.26±0.5148.61±4.71
HLC†71.14±0.6069.50±0.4067.80±0.3372.13±0.2667.59±0.9664.28±0.81
CF1 ↑BCE62.47±0.4453.26±0.4143.43±1.6766.03±1.6955.90±0.7049.29±0.64
CSRA62.08±0.7053.23±1.2743.23±1.2566.02±0.7457.71±1.0247.46±1.68
ADDGCN59.67±1.1452.61±0.5244.33±1.9965.22±0.8655.32±0.7647.30±1.12
APL62.99±1.0753.69±1.8041.72±1.4266.44±1.4057.52±0.9148.14±1.02
CDR62.18±1.0453.61±0.4542.83±0.8766.29±2.1257.23±1.4349.03±1.50
JOINT60.57±2.8254.39±3.7240.48±7.7066.30±2.3359.72±2.1255.06±0.36
WSIC61.70±0.9253.10±0.7442.72±0.5465.34±1.4857.21±1.6248.51±1.12
CCMN64.46±0.6257.45±0.9948.27±0.6867.48±1.4456.93±1.6947.01±1.82
HLC†69.54±0.5667.35±0.4865.72±1.4870.07±0.4165.68±0.9460.57±1.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.456, + 0.868, + 0.469 + ], + "angle": 0, + "content": "Table 3. Comparisons with advanced methods on noisy MS-COCO. The mean and standard deviation of results (%) are presented." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.47, + 0.818, + 0.815 + ], + "angle": 0, + "content": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE53.23±0.1547.33±0.7940.25±0.2656.58±0.2249.16±0.0441.57±0.64
CSRA53.89±0.4047.64±0.8639.58±0.1958.27±0.2350.95±0.0743.07±0.64
ADDGCN51.08±0.9544.75±1.1538.66±1.3056.94±0.6150.28±0.8141.45±0.19
APL54.34±0.3248.61±0.7243.55±1.4357.73±0.2050.87±0.3441.77±0.50
CDR54.01±0.0449.01±0.2643.94±1.2557.03±0.2850.99±0.7742.71±0.09
JOINT53.93±0.4148.01±1.0445.27±0.6857.30±0.3351.94±0.2042.74±0.55
WSIC52.99±0.5346.84±0.8639.76±0.6456.66±0.3149.46±0.2542.52±0.62
CCMN51.73±0.1850.36±0.7145.32±0.8958.13±0.4451.17±0.2942.12±0.76
HLC†54.87±0.6851.09±0.5348.15±0.5058.55±0.0953.41±0.1345.91±0.39
OF1 ↑BCE51.34±1.7044.36±0.8234.85±1.2459.16±0.9552.44±0.8142.94±1.13
CSRA52.03±1.8641.63±1.4133.47±3.1859.17±0.1450.27±0.8841.75±1.36
ADDGCN55.67±1.4847.79±0.4035.95±3.7360.96±0.6555.05±1.7847.47±0.77
APL51.07±1.3243.93±2.7033.90±4.0060.04±1.1650.64±2.8644.34±1.99
CDR53.43±1.1645.10±0.8334.91±0.9059.34±0.6152.72±0.6344.17±0.61
JOINT54.56±0.0649.00±1.6637.78±0.9358.20±0.4053.21±0.1746.55±0.61
WSIC50.91±0.5242.93±0.8535.47±1.5258.89±1.1351.63±1.5743.99±1.47
CCMN52.71±1.0443.24±1.1934.62±1.3858.61±1.1852.18±0.7645.92±0.59
HLC†59.92±0.6557.84±0.3855.47±0.9562.28±0.0658.56±0.3751.09±0.60
CF1 ↑BCE45.92±0.2338.96±1.6131.34±0.2752.54±0.5845.54±0.6339.79±0.99
CSRA44.97±1.8837.49±1.7328.96±1.1652.18±0.4444.96±0.4336.88±0.21
ADDGCN46.77±1.8039.35±1.8330.57±1.5754.18±0.2347.55±0.1839.44±0.33
APL42.91±0.5438.38±0.7728.17±2.5052.87±1.0746.27±1.2737.76±1.02
CDR46.62±0.4239.47±0.5429.59±2.5252.51±0.6945.75±0.8139.15±0.53
JOINT49.51±0.8142.38±1.2124.24±0.6154.39±0.1749.90±0.8538.34±0.55
WSIC45.30±1.0939.15±1.6231.42±0.9452.04±0.2845.76±0.7039.44±1.11
CCMN44.20±1.1935.18±1.0127.90±1.2553.23±0.5846.88±0.9240.55±0.89
HLC†51.94±0.6349.24±0.3046.69±0.6655.44±0.1350.91±0.4843.35±0.82
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.835, + 0.47, + 0.896 + ], + "angle": 0, + "content": "best across all noise settings. In many cases, the best results achieved by HLC outperform the second best results by a large margin, especially when the noise level is high. Below, we further discuss the results based on the comparisons" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.835, + 0.753, + 0.85 + ], + "angle": 0, + "content": "with three different types of baselines." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Compared with Type-I baselines. We first notice that Type-I baselines are fragile to noisy labels in multi-label classification. Without considering the side-effect of noisy" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1488" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.304 + ], + "angle": 0, + "content": "labels, in many cases, they perform worse than BCE, which clearly illustrates the necessity for attention to handling noisy labels. Second, we compare HLC with ADDGCN. Without the proposed correction method for combating noisy labels, HLC will reduce to ADDGCN. As shown in the reported results, HLC performs much better than ADDGCN. To be specific, on noisy Pascal-VOC 2007, for Sym. \\(40\\%\\), HLC brings about \\(+15\\%\\) performance improvement w.r.t. three evaluation metrics over ADDGCN. For Sym. \\(50\\%\\), the performance improvement is increased to more than \\(+20\\%\\). Also, for Pair. \\(30\\%\\) and Pair. \\(40\\%\\), HLC enhances ADDGCN with about \\(+10\\%\\) improvement. On noisy Pascal-VOC 2012 and MS-COCO, the performance improvement is also very clear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.472, + 0.445 + ], + "angle": 0, + "content": "Compared with Type-II baselines. On noisy Pascal-VOC 2007, with Sym. noise, we can see that HLC outperforms APL, CDR, and JOINT clearly, especially for Sym. \\(50\\%\\). Additionally, with Pair. noise, although the improvement is less than the cases with Sym. noise, HLC still performs best. On noisy Pascal-VOC 2012, for both Sym. and Pair. noise, the improvement is significant. Lastly, for noisy MS-COCO, HLC works better than all Type-II baselines with varying enhancement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.472, + 0.552 + ], + "angle": 0, + "content": "Note that, compared with APL and CDR, JOINT seems to be a stronger baseline. Benefiting from label correction, after a few training epochs, JOINT less overfits to wrong labels, following better performance. Nevertheless, the proposed label-correction paradigm is argued to be more advanced. As shown in all results, HLC surpasses JOINT, which verifies the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.472, + 0.664 + ], + "angle": 0, + "content": "Compared with Type-III baselines. On noisy Pascal-VOC 2007 and noisy Pascal-VOC 2012, HLC outperforms WSIC and CCMN distinctly. For example, with Sym. \\(50\\%\\) noise, more than \\(+10\\%\\) performance promotion is brought by our method. On noisy MS-COCO, although WSIC and CCMN are sometimes competitive w.r.t. mAP, they are inferior w.r.t. both OF1 and CF1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.675, + 0.367, + 0.691 + ], + "angle": 0, + "content": "4.3. More Analyses and Justifications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.759 + ], + "angle": 0, + "content": "In this subsection, we conduct performance analysis in more detail. The experiments are conducted with Sym. \\(50\\%\\) noise, which is more challenging than the experiments in low-noise-rate cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Role of label dependence. We study the effect of removing the consideration of label dependence to provide insights into what makes HLC successful. The experiments are conducted on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to \\(224 \\times 224\\). Recall that HLC considers instance-label and label dependences simultaneously. When we remove the consideration of the label dependence in HLC, the correspond" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.894, + 0.129 + ], + "angle": 0, + "content": "Table 4. Ablation study results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The mean and standard deviation of results are presented. The best result in each case is in bold." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.129, + 0.862, + 0.285 + ], + "angle": 0, + "content": "
DatasetNoisy Pascal-VOC 2007
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.06±0.4167.23±1.9263.42±0.58
HLC68.00±0.8968.62±0.4864.82±0.48
DatasetNoisy Pascal-VOC 2012
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.88±0.7566.30±1.2864.33±1.67
HLC68.69±1.0467.80±0.3365.72±1.48
DatasetNoisy MS-COCO
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.46.21±0.3652.90±0.9244.51±1.29
HLC48.15±0.5055.47±0.9546.69±0.66
" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.304, + 0.825, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.417, + 0.824, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.529, + 0.894, + 0.571 + ], + "angle": 0, + "content": "Figure 2. Ablation study results with different values of the set threshold \\(\\hat{\\delta}\\). The experiments are conducted on noisy Pascal-VOC 2007 (Top) and noisy Pascal-VOC 2012 (Bottom)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.895, + 0.738 + ], + "angle": 0, + "content": "ing method is named as HLC w/o l. here. For both HLC w/o l. and HLC, the value of the threshold \\(\\hat{\\delta}\\) is searched in the range \\(\\{0.25, 0.30, 0.35, 0.40, 0.45\\}\\). We use the \\(10\\%\\) noisy training data as a validation set for the threshold determination and performance report. The results are shown in Table 4. As can be seen, HLC outperforms HLC w/o l.. The results justify our claims that the label dependence could help combat the noisy labels in multi-label classification, which demonstrate the effectiveness of the proposed holistic correction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.895, + 0.881 + ], + "angle": 0, + "content": "Analysis of the threshold \\(\\hat{\\delta}\\). We analyze the influence of different values of \\(\\hat{\\delta}\\). The experiments are conducted on noisy Pascal-VOC 2007 and Pascal-VOC 2012. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to \\(224 \\times 224\\). The value of the threshold \\(\\hat{\\delta}\\) is chosen in \\(\\{0.25, 0.30, 0.35, 0.40, 0.45\\}\\). Figure 2 shows that HLC is robust to the determination of the threshold \\(\\hat{\\delta}\\) in the certain range, which facilitates the practical application of our method." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Evaluations with different networks. We use pretrained" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1489" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.129 + ], + "angle": 0, + "content": "Table 5. Comparisons with advanced methods on noisy MS COCO with different networks. The mean and standard deviation of results \\((\\%)\\) are presented." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.129, + 0.426, + 0.475 + ], + "angle": 0, + "content": "
MetricsMethodsResNet-34ResNet-101
mAP ↑BCE42.63±0.7438.17±0.41
CSRA41.35±0.1837.24±1.20
ADDGCN40.15±0.9836.13±0.69
APL44.82±0.7040.90±1.51
CDR45.43±0.6541.00±0.38
JOINT44.81±0.7739.96±1.30
WSIC41.86±0.6237.49±0.68
CCMN45.31±0.4746.01±1.01
HLC†46.05±0.8146.24±2.13
OF1 ↑BCE37.65±2.4638.65±2.50
CSRA35.05±0.7834.28±2.40
ADDGCN35.11±1.2637.18±0.50
APL34.31±1.7337.89±1.30
CDR36.67±3.4339.88±1.15
JOINT39.66±1.1341.36±0.88
WSIC35.08±1.7438.44±0.57
CCMN32.86±1.5037.03±1.48
HLC†44.11±0.8047.79±4.19
CF1 ↑BCE28.95±2.0934.11±1.13
CSRA27.40±0.4431.21±1.57
ADDGCN26.11±0.8630.41±0.72
APL26.64±2.3931.97±1.95
CDR27.13±2.3935.17±1.06
JOINT30.77±1.6337.63±0.81
WSIC27.62±0.6533.83±0.61
CCMN24.75±0.4826.65±0.26
HLC†34.79±1.4140.88±2.42
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.606 + ], + "angle": 0, + "content": "ResNet-50 before. To show that our method is robust to the choice of network structures, we use different networks in experiments. Specifically, we employ pretrained ResNet-34 [20] and pretrained ResNet-101 [20] respectively. The noisy MS-COCO is considered. The image size is \\(224 \\times 224\\). The results on mAP are reported in Table 5. As can be seen, with different networks, HLC still works well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.732 + ], + "angle": 0, + "content": "Evaluations with different image sizes. We resize the image size to \\(224 \\times 224\\) before. To test the performance of advanced methods with different image sizes, we further consider \\(112 \\times 112\\), \\(384 \\times 384\\), and \\(448 \\times 448\\) image sizes. Pretrained ResNet-50 is used. The results are reported in Table 6. For mAP, we can see that HLC is competitive compared with CCMN and CSRA. For OF1 and CF1, HLC works better than all baselines with a clear margin." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.416, + 0.758 + ], + "angle": 0, + "content": "4.4. Experiments on the Real-world Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To demonstrate that our problem setting can be adapted to the real world and our method can well handle practical scenes, we employ the real-world dataset NUS-WIDE [10] that originally contained 269,648 images from Flicker, which have been manually annotated with 81 visual concepts. Since some urls for download have been deleted, we employ the dataset version in [45]. A standard 70-30 train-test split is used. The backbone is chosen as ResNet-101. As the computation cost of training on NUS-WIDE is rela" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.129 + ], + "angle": 0, + "content": "Table 6. Comparisons with advanced methods on noisy MS COCO. The mean and standard deviation of results \\((\\%)\\) are presented. Difference image sizes are considered here." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.129, + 0.895, + 0.473 + ], + "angle": 0, + "content": "
MetricsImage sizes112 × 112384 × 384448 × 448
mAP ↑BCE32.22±0.6939.40±1.3635.24±1.73
CSRA29.55±0.1643.55±0.7044.56±0.75
ADDGCN32.34±0.4638.72±1.6434.87±1.89
APL34.41±0.4843.65±0.2841.44±1.21
CDR34.75±0.3943.26±0.7239.97±1.40
JOINT32.89±0.1642.95±0.8840.17±1.26
WSIC31.98±0.2339.57±1.0236.08±0.23
CCMN36.17±0.4144.39±0.3944.03±0.17
HLC†35.98±1.0545.12±0.1344.23±1.20
OF1 ↑BCE26.70±0.8834.71±2.7626.72±3.35
CSRA20.14±1.2836.41±0.7138.54±0.78
ADDGCN26.36±2.2942.83±2.0540.73±1.04
APL24.02±1.2234.68±1.4630.73±2.64
CDR26.50±1.2331.31±1.7631.15±3.46
JOINT34.11±0.9538.67±1.2538.11±0.69
WSIC24.61±1.1034.09±2.9430.69±0.97
CCMN23.89±1.4936.16±2.1225.03±2.48
HLC†39.05±2.6846.55±3.7745.14±2.34
CF1 ↑BCE19.61±0.6130.77±2.2824.94±3.31
CSRA13.34±1.2932.66±0.9832.68±0.22
ADDGCN18.67±1.4735.63±1.3733.89±2.41
APL17.21±0.7829.24±0.0925.01±1.31
CDR18.04±1.0729.62±1.1926.28±1.23
JOINT20.76±0.7534.90±1.8833.75±1.31
WSIC19.07±0.5731.63±2.4928.50±1.36
CCMN16.75±1.2130.09±2.3226.68±1.57
HLC†29.70±2.0740.34±2.1437.48±2.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.481, + 0.891, + 0.507 + ], + "angle": 0, + "content": "Table 7. Comparison of our method to known state-of-the-art models on the NUS-WIDE dataset. Metrics are in %." + }, + { + "type": "table", + "bbox": [ + 0.562, + 0.507, + 0.828, + 0.606 + ], + "angle": 0, + "content": "
MethodmAP ↑OF1 ↑CF1 ↑
S-CLs [39]60.173.758.7
MS-CMA [66]61.473.860.5
SRN [76]62.073.458.5
ICME [7]62.874.160.7
ASL [45]63.974.662.7
HLC†63.174.662.9
HLC+ASL†64.575.163.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.751 + ], + "angle": 0, + "content": "tively large, we run experiments one time. Here we compare our method with S-CLs [39], MS-CMA [66], SRN [76], ICME [7], and ASL [45]. For convenient comparison, we refer to the results of their original papers. Note that to further improve the performance on NUS-WIDE, we utilize ASL to replace the loss function of our method. We name the new method \"HLC+ASL\". Results are provided in Table 7, which demonstrate the effectiveness of our method on the real-world dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.619, + 0.784 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this paper, we focus on the realistic problem of multi-label classification with noisy labels. We learn and utilize the label dependence among multiple labels to handle this problem. With the help of label dependence, a novel algorithm named HLC is proposed to correct noisy multiple labels to clean ones. We demonstrate the effectiveness of our algorithm both theoretically and empirically. For future" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1490" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "work, we are interested in adapting HLC to other domains such as natural language processing and recommendation systems. We are also interested in promoting our algorithm to tackle instance-dependent label noise [73, 4, 78, 38] in multi-label classification." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.181, + 0.245, + 0.198 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.472, + 0.298 + ], + "angle": 0, + "content": "Bo Han was supported by NSFC Young Scientists Fund No. 62006202 and Guangdong Basic and Applied Basic Research Foundation No. 2022A1515011652. Tongliang Liu is partially supported by ARC projects: FT220100318, DP220102121, LP220100527, LP220200949, and IC190100031." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.31, + 0.175, + 0.326 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.336, + 0.47, + 0.405 + ], + "angle": 0, + "content": "[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, pages 233-242, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.407, + 0.472, + 0.434 + ], + "angle": 0, + "content": "[2] Dara Bahri, Heinrich Jiang, and Maya Gupta. Deep k-nn for noisy labels. In ICML, pages 540–550, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.436, + 0.47, + 0.476 + ], + "angle": 0, + "content": "[3] Yingbin Bai and Tongliang Liu. Me-momentum: Extracting hard confident examples from noisily labeled data. In ICCV, pages 9312-9321, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.479, + 0.47, + 0.533 + ], + "angle": 0, + "content": "[4] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. arXiv preprint arXiv:2001.03772, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.535, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[5] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In ICCV, pages 522-531, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.592, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[6] Zhaomin Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Learning graph convolutional networks for multi-label recognition and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.649, + 0.468, + 0.703 + ], + "angle": 0, + "content": "[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME, pages 622-627, 2019. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.706, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[8] Xiang Cheng, Shu-Guang Zhao, Xuan Xiao, and Kuo-Chen Chou. iatc-misf: a multi-label classifier for predicting the classes of anatomical therapeutic chemicals. Bioinformatics, 33(3):341-346, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.762, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[9] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. arXiv preprint arXiv:2109.04997, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[10] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In ACM International Conference on Image and Video Retrieval, pages 1-9,2009.8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.336, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[11] Zijun Cui, Yong Zhang, and Qiang Ji. Label error correction and generation through label relationships. In AAAI, pages 3693-3700, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[12] Mark Everingham, Andrew Zisserman, Christopher KI Williams, Luc Van Gool, Moray Allan, Christopher M Bishop, Olivier Chapelle, Navneet Dalal, Thomas Deselaers, Gyuri Dorkó, et al. The pascal visual object classes challenge 2007 (voc2007) results. 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.209, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[13] Lei Feng, Senlin Shu, Zhuoyi Lin, Fengmao Lv, Li Li, and Bo An. Can cross entropy loss be robust to label noise? In IJCAI, pages 2206-2212, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.253, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[14] Bin-Bin Gao and Hong-Yu Zhou. Learning to discover multi-class attentional regions for multi-label image recognition. IEEE Transactions on Image Processing, 30:5920-5932, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[15] Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy labels. arXiv preprint arXiv:1607.07526, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.354, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[16] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor Tsang, and Masashi Sugiyama. *Sigua: Forgetting may make learning with noisy labels more robust.* In ICML, pages 4006–4016, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.478 + ], + "angle": 0, + "content": "[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages 8527-8537, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[18] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[19] Jun-Yi Hang and Min-Ling Zhang. Dual perspective of label-specific feature learning for multi-label classification. In ICML, pages 8375-8386, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.585, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.629, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In NeurIPS, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[22] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly supervised image classification through noise regularization. In CVPR, pages 11517-11525, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[23] Zhuo Huang, Xiaobo Xia, Li Shen, Bo Han, Mingming Gong, Chen Gong, and Tongliang Liu. Harnessing out-of-distribution examples via augmenting content and style. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[24] Zhuo Huang, Miaoxi Zhu, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, Bo Du, and Tongliang Liu. Robust generalization against photon-limited corruptions via worst-case sharpness minimization. In CVPR, pages 16175-16185, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, pages 2309-2318, 2018. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1491" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[26] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. Nlnl: Negative learning for noisy labels. In ICCV, pages 101-110, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[28] Cheng Li, Bingyu Wang, Virgil Pavlu, and Javed Aslam. Conditional bernoulli mixtures for multi-label classification. In ICML, pages 2482-2491, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[29] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[30] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In ICCV, pages 9485-9494, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[31] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[34] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[35] Weiwei Liu, Ivor W Tsang, and Klaus-Robert Müller. An easy-to-hard learning paradigm for multiple classes and multiple labels. Journal of Machine Learning Research, 18, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[36] Weiwei Liu, Haobo Wang, Xiaobo Shen, and Ivor Tsang. The emerging trends of multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[37] Weiwei Liu, Donna Xu, Ivor W Tsang, and Wenjie Zhang. Metric learning for multi-output tasks. Transactions on Pattern Analysis and Machine Intelligence, 41(2):408-422, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[38] Yang Liu. Identifiability of label noise transition matrix. arXiv preprint arXiv:2202.02016, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[39] Yongcheng Liu, Lu Sheng, Jing Shao, Junjie Yan, Shiming Xiang, and Chunhong Pan. Multi-label image classification via knowledge distillation from weakly-supervised detection. In ACMMM, pages 700-708, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.775, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[40] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. In ICML, pages 6543-6553, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[41] Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. Modeling fine-grained entity types with box embeddings. In ACL, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[43] Cosmin Octavian Pene, Amirmasoud Ghiassi, Taraneh Younesian, Robert Birke, and Lydia Y Chen. Multi-label gold asymmetric loss correction with single-label regulators. arXiv preprint arXiv:2108.02032, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[44] Geoff Pleiss, Tianyi Zhang, Ethan R Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[45] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In ICCV, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[46] Jun Shu, Qian Zhao, Zengben Xu, and Deyu Meng. Meta transition adaptation for robust deep learning with noisy labels. arXiv preprint arXiv:2006.05697, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[47] Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[48] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[49] Alexander B Tsybakov. Optimal aggregation of classifiers in statistical learning. The Annals of Statistics, 32(1):135-166, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[50] Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks. In NeurIPS, pages 5596-5605, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[51] Andreas Veit, Neil Alldrin, Gal Chechik, Ivan Krasin, Abhinav Gupta, and Serge Belongie. Learning from noisy large-scale datasets with minimal supervision. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[52] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, pages 2285-2294, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[53] Xinshao Wang, Yang Hua, Elyor Kodirov, David A Clifton, and Neil M Robertson. Proselflc: Progressive self label correction for training robust deep neural networks. In CVPR, pages 752-761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[54] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pages 8688-8696, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[55] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, pages 13726–13735, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[56] Hongxin Wei, Lue Tao, Renchunzi Xie, and Bo An. Open-set label noise can improve robustness against inherent label noise. In NeurIPS, pages 7978-7992, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[57] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed graph. In ICCV, pages 4157-4165, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[58] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hinding the memorization of noisy labels. In ICLR, 2021. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1492" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[60] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[61] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835-6846, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[62] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning. In AAAI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[63] Ming-Kun Xie and Sheng-Jun Huang. Ccmn: A general framework for learning with class-conditional multi-label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[64] Ming-Kun Xie, Jia-Hao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[65] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, pages 649-665, 2020. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[66] Renchun You, Zhiyao Guo, Lei Cui, Xiang Long, Yingze Bao, and Shilei Wen. Cross-modality attention with semantic graph embedding for multi-label classification. In AAAI, pages 12709-12716, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[67] Hsiang-Fu Yu, Prateek Jain, Purushottam Kar, and Inderjit Dhillon. Large-scale multi-label learning with missing labels. In ICML, pages 593-601, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[68] Ze-Bang Yu and Min-Ling Zhang. Multi-label classification with label-specific feature generation: A wrapped approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[69] Jiong Zhang, Wei-cheng Chang, Hsiang-fu Yu, and Inderjit Dhillon. Fast multi-resolution transformer fine-tuning for extreme multi-label text classification. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[70] Min-Ling Zhang and Kun Zhang. Multi-label learning by exploiting label dependency. In KDD, pages 999-1008, 2010. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[71] Min-Ling Zhang and Zhi-Hua Zhou. A review on multi-label learning algorithms. IEEE Transactions on Knowledge and Data Engineering, 26(8):1819-1837, 2013. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[72] Yivan Zhang, Gang Niu, and Masashi Sugiyama. Learning noise transition matrix from only noisy labels via total variation regularization. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[73] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In ICLR, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[74] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In ICCV, pages 163-172, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[75] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "correction of noisy labels. In ICML, pages 11447-11457, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[76] Feng Zhu, Hongsheng Li, Wanli Ouyang, Nenghai Yu, and Xiaogang Wang. Learning spatial regularization with image-level supervisions for multi-label image classification. In CVPR, pages 5513-5522, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[77] Ke Zhu and Jianxin Wu. Residual attention: A simple but effective method for multi-label recognition. In ICCV, pages 184-193, 2021. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[78] Zhaowei Zhu, Tongliang Liu, and Yang Liu. A second-order approach to learning with instance-dependent label noise. In CVPR, pages 10113-10123, 2021. 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1493" + } + ] +] \ No newline at end of file diff --git a/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_origin.pdf b/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b9ed593304bef4a92174dc8684b9a3ba0a2eb97 --- /dev/null +++ b/2023/Holistic Label Correction for Noisy Multi-Label Classification/c24a07fb-5f1b-44b9-bb06-0d6a49b47146_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f33902570b511de473e01eed36c01c7e40f4b63714ea41de46ca38328d1be909 +size 1072263 diff --git a/2023/Holistic Label Correction for Noisy Multi-Label Classification/full.md b/2023/Holistic Label Correction for Noisy Multi-Label Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..515d2b285eeb92ae0a7b5170147d3282c025322b --- /dev/null +++ b/2023/Holistic Label Correction for Noisy Multi-Label Classification/full.md @@ -0,0 +1,301 @@ +# Holistic Label Correction for Noisy Multi-Label Classification + +Xiaobo Xia $^{1}$ , Jiankang Deng $^{2}$ , Wei Bao $^{1}$ , xuan Du $^{3}$ , Bo Han $^{4}$ , Shiguang Shan $^{5,6}$ , Tongliang Liu $^{1*}$ + +1The University of Sydney 2Imperial College London + +$^{3}$ JD Explore Academy $^{4}$ Hong Kong Baptist University + +$^{5}$ Chinese Academy of Sciences $^{6}$ University of Chinese Academy of Sciences + +# Abstract + +Multi-label classification aims to learn classification models from instances associated with multiple labels. It is pivotal to learn and utilize the label dependence among multiple labels in multi-label classification. As a result of today's big and complex data, noisy labels are inevitable, making it looming to target multi-label classification with noisy labels. Although the importance of label dependence has been shown in multi-label classification with clean labels, it is challenging and hard to bring label dependence to the problem of multi-label classification with noisy labels. The issues are, that we do not understand why label dependence is helpful in the problem, and how to learn and utilize label dependence only using training data with noisy multiple labels. In this paper, we bring label dependence to tackle the problem of multi-label classification with noisy labels. Specifically, we first provide a high-level understanding of why label dependence helps distinguish the examples with clean/noisy multiple labels. Benefiting from the memorization effect in handling noisy labels, a novel algorithm is then proposed to learn the label dependence by only employing training data with noisy multiple labels, and utilize the learned dependence to help correct noisy multiple labels to clean ones. We prove that the use of label dependence could bring a higher success rate for recovering correct multiple labels. Empirical evaluations justify our claims and demonstrate the superiority of our algorithm. + +# 1. Introduction + +Multi-label classification assigns a set of multiple labels for each instance [71]. As a practical learning paradigm, multi-label classification has been widely applied in various domains, ranging from computer vision [7] and natural language processing [41], to recommendation systems [69] and bioinformatics [8]. Consensually, compared + +with multi-class classification [20, 23, 24], where each instance is assigned with a single label, multi-label classification is more challenging [35]. Plenty of advanced methods are proposed in recent years for multi-label classification [77, 45, 14, 74, 37, 9, 19, 64]. + +The great majority of the methods assume that training data are annotated precisely. However, noisy labels are inevitable in multi-label classification [36], especially for classification with big and complex data. They may be resulted by unintentional mistakes of manual and automatic annotators [51, 75, 13], or intentional corruptions on clean labels [50, 44]. Noisy labels severely impair the generalization of learned models, over-parameterized deep models in particular [26, 61, 58, 59, 55, 56]. A straightforward way to address the problem of multi-label classification with noisy labels is to treat each label in isolation and convert the multi-label problem into a number of binary classification problems. Afterward, the methods in multi-class classification with noisy labels [16, 47] are applied to train independent binary classifiers, which capture instance-label dependence robustly to strengthen classification. This way is a remedy to handle noisy labels, but ignores the label dependence among multiple labels. It is essential to learn and utilize the label dependence in multi-label classification [70, 18, 11, 31]. + +Prior works [65, 6, 52] illustrate the successes of considering the label dependence among multiple labels in multilabel classification with clean labels. In different ways, e.g., helping learn inter-dependent classifiers [7], the label dependence can be used to boost the learning of the instance-label dependence, which improves final classification. Inspired by the successes, it is concerned that label dependence could be exploited to handle the problem of multilabel classification with noisy labels. However, there are few attempts before for this important problem. At least three questions make the solution remain mysterious. First, in intuition, we need to understand why label dependence is helpful for the problem. Second, in technique, we need to know how to learn and utilize the label dependence in the + +![](images/b4b3ab45ab5805c96c8957631002d585b9fbac6a93b1b94f640454d10a95c141.jpg) +Figure 1. The illustration of why the label dependence among multiple labels helps distinguish the examples with noisy/clean multiple labels. The arrow presents the label dependence between a label pair. For the labels “a” and “b”, “a → b” means that, when “a” appears, “b” will also occur with high probability. The example comes from a web search. The set of clean multiple labels is {Sea, Human, Motorboat}, where the label dependence is strong with both “Motorboat → Sea” and “Motorboat → Human”. However, due to label corruption, Motorboat is flipped to be Motorcycle, which causes “Motorcycle → Sea”. Therefore, the label dependence among noisy multiple labels is weaker than the label dependence among corresponding clean ones. + +problem. As we only have training data with noisy labels, both the accurate catch and application of the label dependence are challenging. Third, in verification, we need to know what improvements the label dependence can bring. + +In this paper, we answer the three questions one by one. The first answer is illustrated in Figure 1. That is, compared with noisy multiple labels, the label dependence among clean multiple labels is stronger with high probability. Therefore, such dependence could help distinguish the examples with noisy/clean multiple labels for our problem. The second answer is given by the proposed holistic correction for multi-label classification with noisy labels (aka HLC). Specifically, HLC inherits the memorization effect in handling noisy labels [1, 25, 53]: the deep model would firstly memorize the training examples with clean labels, leading to reliable model predictions in early training. In HLC, the label dependence is learned by a dynamic graph [65], and then applied to correcting noisy multiple labels. In more detail, the holistic score in HLC is proposed to measure the instance-label and label dependencies in an example. The stronger instance-label and label dependencies make a larger holistic score. We compare the ratio between the holistic scores of the example with noisy multiple labels and its variant with predicted multiple labels, with an easily determined threshold. The noisy multiple labels are corrected or changeless based on the comparison result. Benefiting from the memorization effect, both dependence learning and multi-label correction are useful. Besides, they fulfill a positive cycle [3]. Namely, better dependence learning results in a better multi-label correction, and better multi-label correction makes better dependence learning, leading to final enhanced classification. + +The third answer is given by both theoretical analyses and empirical evaluations. Theoretically, we show that the additional use of label dependence brings a higher probability to handle noisy multiple labels successfully than the sole use of instance-label dependence under some conditions. Empirically, we demonstrate the power of label de + +pendence through experiments and show that, in most situations, HLC outperforms comparison methods with large margins. + +The contributions of this paper are summarized as follows. (1) We focus on a realistic problem of multi-label classification with noisy labels. The challenges of using label dependency to address the problem are carefully analyzed, which benefits future research on the problem. (2) We propose an effective method to handle noisy labels in multi-label classification. The method measures simultaneous instance-label and label dependencies in an example for follow-up label correction. (3) Theoretical analysis is provided to explain the success rate of the proposed method. Besides, we confirm that the use of label dependence is indeed powerful under some conditions. (4) Extensive empirical results on multiple benchmarks demonstrate the superiority of our method. Detailed ablation studies and discussions are also provided. Codes are attached in the supplementary material. + +# 2. Preliminaries + +Problem setup. Let $\mathcal{X} \in \mathbb{R}^d$ denote the input space and $\mathcal{Y} \in \{l_1, \dots, l_q\}$ denote the label space with $q$ class labels. An example with multiple labels is denoted as $(x, y)$ , where $x \in \mathcal{X}$ is the feature vector of an instance, and $y \subseteq \mathcal{Y}$ is its set of associated labels. Denote the size of the label set $y$ as $|y|$ . For the feature vector $x$ , its label set $y$ may be corrupted and is flipped into $\bar{y} \subseteq \mathcal{Y}$ . We utilize a class-dependent noise transition matrix $\pmb{T}$ [43, 46, 34, 60] to characterize the label flip process. Formally, for any $i \neq j$ , $T_{ij} = \mathbb{P}(l_j \in \bar{y} \wedge l_i \notin \bar{y} | l_j \notin y \wedge l_i \in y)$ represents the probability of the $i$ -th class label to be flipped into the $j$ -th class label. Consider a noisy multi-label dataset comprising several examples $(x, \bar{y})$ . The aim is to learn a classification model robustly by only using the noisy dataset. Given an instance in testing, with the learned model, we can predict its relevant label set precisely. + +Note that some works employ another problem setting that the total number of multiple labels can be changed after label flipping, which is referred to as multi-label classification with missing or redundant labels. For the former setting, it is not accurate to consider it as a classification with noisy labels, since all annotated labels are correct [67, 57]. For the latter setting, it is normally called partial multi-label learning [62], which is different from the problem setting of this paper, as detailed in Appendix C.4. Our setting, i.e., the total number of labels is preserved after label flipping, is realistic. In many practical situations, it is easy to determine the number of objects in an image, in particular with object detection techniques. In contrast, it can be harder to annotate the objects perfectly, resulting in noisy labels. In addition, in Section 4.4, we will show that our problem setting well fits the realistic situation. That is to say, the proposed method can achieve superior performance on a realistic noisy multi-label dataset. + +Preparation technology. As discussed, we need both the instance-label dependence and the label dependence among multiple labels. Given an example $(\pmb{x},\pmb{y})$ , for the instance-label dependence, it can be learned with the conditional probability of $l_{i}\in \pmb{y}$ given $\pmb{x}$ according to model's probability outputs. For the label dependence among multiple labels, it is often estimated by counting the occurrence of label pairs in training data [7]. + +Recently, the graph convolutional network (GCN) is used in multi-label classification and achieves great successes [7, 65, 6]. The advantage of the GCN-based methods is that they can capture the instance-label and label dependencies simultaneously during training. In this paper, we inherit the advantage of the GCN-based methods and build HLC based on ADDGCN [65]. ADDGCN designs a semantic attention module (SAM) to estimate the content-aware class-label representations for each class from the extracted feature map. The representations are fed into a GCN module (GCNM) for final classification. We provide the technical details of ADDGCN [65] in Appendix C.1. Before delving into the next section, readers only need to remember that the instance-label and label dependencies can be learned during training. Note that we also review prior works on multi-class classification with noisy labels and multi-label classification with clean/noisy labels in Appendix C.2 and Appendix C.3. + +# 3. Proposed Method + +# 3.1. Holistic Judgment in Multi-Label Classification + +Holistic score. We begin with an example with clean multiple labels. Given an example $(\pmb{x},\pmb{y})$ , we can measure the instance-label dependence $S^f$ , and the label dependence $S^l$ . Denote the variable of clean multiple labels by $\pmb{Y}$ . Mathematically, we define two dependen + +Algorithm 1: Holistic Correction. +Input: $(\pmb {x},\bar{\pmb{y}})$ $h$ ,and $\hat{\delta}$ +Output: $\bar{y}_{new}$ +1: $\pmb {y}^{*} = h(\pmb {x})$ . +2: $\kappa (h,\pmb {x},\bar{\pmb{y}}) = \hat{\bar{S}}_{\bar{\pmb{y}}}(\pmb {x}) / \hat{\bar{S}}_{\pmb{y}^{*}}(\pmb {x})$ +3: if $\kappa (h,\pmb {x},\bar{\pmb{y}})\leq \hat{\delta}$ then $\bar{y}_{new} = y^{*}$ +4: else $\bar{y}_{new} = \bar{y}$ + +cies as $S_{\pmb{z}}^{f}(\pmb{x}) \coloneqq \sum_{\{\pmb{Y} = \pmb{z}, l_{i} \in \pmb{z}\}} \mathbb{P}(l_{i}|\pmb{x})$ and $S_{\pmb{z}}^{l}(\pmb{x}) \coloneqq \sum_{\{\pmb{Y} = \pmb{z}, l_{i}, l_{j} \in \pmb{z}\}} \frac{1}{2} [\mathbb{P}(l_{j}|l_{i}, \pmb{x}) + \mathbb{P}(l_{i}|l_{j}, \pmb{x})]$ , where $\pmb{z}$ is the value of the random variable $\pmb{Y}$ . The holistic score of the example $(\pmb{x}, \pmb{y})$ considers two dependencies at the same time. Formally, we denote the holistic score of $(\pmb{x}, \pmb{y})$ as $S_{\pmb{y}}(\pmb{x})$ and define it as + +$$ +S _ {\boldsymbol {y}} (\boldsymbol {x}) := S _ {\boldsymbol {y}} ^ {f} (\boldsymbol {x}) + S _ {\boldsymbol {y}} ^ {l} (\boldsymbol {x}). \tag {1} +$$ + +Afterward, denote the variable of noisy multiple labels by $\bar{Y}$ . For the example with noisy multiple labels, i.e., $(\pmb{x},\bar{\pmb{y}})$ , the instance-label dependence and label dependence are measure by $\bar{S}_z^f (\pmb {x})\coloneqq \sum_{\{\bar{Y} = \pmb {z},l_i\in \pmb {z}\}}\mathbb{P}(l_i|\pmb {x})$ and $\bar{S}_z^l (\pmb {x})\coloneqq$ $\sum_{\{\bar{Y} = z,l_i,l_j\in \pmb {z}\}}\frac{1}{2}\left[\mathbb{P}(l_j|l_i,\pmb {x}) + \mathbb{P}(l_i|l_j,\pmb {x})\right]$ . Accordingly, the holistic score of the example $(\pmb {x},\bar{\pmb{y}})$ is denoted by $\bar{S}_{\bar{\pmb{y}}}(\pmb {x})$ , which is defined as $\bar{S}_{\bar{\pmb{y}}}(\pmb {x})\coloneqq \bar{S}_{\bar{\pmb{y}}}^{f}(\pmb {x}) + \bar{S}_{\bar{\pmb{y}}}^{l}(\pmb {x})$ . Note that, during training, we cannot access $\bar{S}_{\bar{\pmb{y}}}^{f}(\pmb {x})$ and $\bar{S}_{\bar{\pmb{y}}}^{l}(\pmb {x})$ . Instead, the estimated posterior probabilities are used. We denote the estimations of $\bar{S}_{\bar{\pmb{y}}}^{f}(\pmb {x})$ and $\bar{S}_{\bar{\pmb{y}}}^{l}(\pmb {x})$ as $\hat{\bar{S}}_{\bar{\pmb{y}}}^{f}(\pmb {x})$ and $\hat{\bar{S}}_{\bar{\pmb{y}}}^{l}(\pmb {x})$ . The estimation of the holistic score is $\hat{\bar{S}}_{\bar{\pmb{y}}}(\pmb {x}) = \hat{\bar{S}}_{\bar{\pmb{y}}}^{f}(\pmb {x}) + \hat{\bar{S}}_{\bar{\pmb{y}}}^{l}(\pmb {x})$ . With preparation technology discussed in Section 2 and Appendix C.1, $\hat{\bar{S}}_{\bar{\pmb{y}}}^{f}(\pmb {x})$ and $\hat{\bar{S}}_{\bar{\pmb{y}}}^{l}(\pmb {x})$ can be obtained. + +Holistic correction. For the example $(\pmb{x},\bar{\pmb{y}})$ , we feed it into the deep network $h$ included in ADDGCN [65]. The memorization effect in handling noisy labels [25, 33] shows that the deep network would first memorize the training data with clean labels and then the training data with noisy labels. Therefore, early in training, the outputs of the deep network are relatively reliable and can be used for label correction. For $(\pmb{x},\bar{\pmb{y}})$ , we denote its set of predicted multiple labels as $\pmb{y}^*$ . Here, the set of predicted labels is obtained with the top $|\bar{\pmb{y}}|$ predictions based on the model's probability outputs. + +Recall that the holistic score of an example holistically measures the instance-label dependence and label dependence among multiple labels simultaneously. From both human and machine cognition, if an example is annotated accurately, both dependencies should be strong [70, 18, 68, 28, 7] with high probability. Namely, the holistic score is large. We propose to check the ratio between the holistic score on $(\pmb{x},\bar{\pmb{y}})$ and holistic score on $(\pmb{x},\pmb{y}^{*})$ . Specifically, + +we check + +$$ +\kappa (h, \boldsymbol {x}, \bar {\boldsymbol {y}}) = \hat {\tilde {S}} _ {\bar {\boldsymbol {y}}} (\boldsymbol {x}) / \hat {\tilde {S}} _ {\boldsymbol {y} ^ {*}} (\boldsymbol {x}). \tag {2} +$$ + +We compare this ratio with a predetermined threshold $\hat{\delta}$ . The value of $\hat{\delta}$ is given in the next subsection. If $\kappa(h, \boldsymbol{x}, \bar{\boldsymbol{y}}) \leq \hat{\delta}$ , we flip the labels $\bar{\boldsymbol{y}}_{\text{new}} = \boldsymbol{y}^*$ . Otherwise, the labels remain unchanged with $\bar{\boldsymbol{y}}_{\text{new}} = \bar{\boldsymbol{y}}$ . The detailed algorithm of holistic correction for multi-label classification with noisy labels (aka HLC) is provided in Algorithm 1. After holistic correction for noisy labels, we use $(\boldsymbol{x}, \bar{\boldsymbol{y}}_{\text{new}})$ to train the deep network $h$ based on ADDGCN [65]. + +# 3.2. Theoretical Insights + +We extend the Tsybakov condition [75, 2, 15, 49] from multi-class classification to multi-label classification. Specifically, denote by $\pmb{a}_{\pmb{x}}$ the label set predicted based on $S^{f}(\pmb{x})$ with $\pmb{a}_{\pmb{x}} := h^{*}(\pmb{x}) = \arg \max_{\pmb{z}} S_{\pmb{z}}^{f}(\pmb{x})$ . Besides, denote by $\pmb{b}_{\pmb{x}}$ the second best prediction with $\pmb{b}_{\pmb{x}} := \arg \max_{\pmb{z} \neq \pmb{a}_{\pmb{x}}} S_{\pmb{z}}^{f}(\pmb{x})$ . The maximum length of a label set is denoted as $m$ ( $m \ll q$ ). In this paper, we call the predicted label set by the Bayes optimal classifier for an instance as the correct label set. We present the Tsybakov condition on instance-label (abbreviated as ins.-label here) dependence and holistic Tsybakov condition as follows. + +Definition 1 (Tsybakov condition on ins.-label dependence) $\exists C_1, \lambda_1 > 0$ and $\exists t_0 \in (0, m]$ , such that for all $t \leq t_0$ , we have + +$$ +\mathbb {P} \left[ S _ {\boldsymbol {a} _ {\boldsymbol {x}}} ^ {f} (\boldsymbol {x}) - S _ {\boldsymbol {b} _ {\boldsymbol {x}}} ^ {f} (\boldsymbol {x}) \leq t \right] \leq C _ {1} t ^ {\lambda_ {1}}. \tag {3} +$$ + +Definition 2 (Holistic Tsybakov condition) $\exists C_2, \lambda_2 > 0$ and $\exists t_0 \in (0, m]$ , such that for all $t \leq t_0$ , we have + +$$ +\mathbb {P} \left[ S _ {\boldsymbol {a} _ {\boldsymbol {x}}} (\boldsymbol {x}) - S _ {\boldsymbol {b} _ {\boldsymbol {x}}} (\boldsymbol {x}) \leq t \right] \leq C _ {2} t ^ {\lambda_ {2}}. \tag {4} +$$ + +Remark 1 Definition 1 stipulates that the uncertainty of $S^f$ is bounded. The margin region that is close to the decision boundary has a bounded volume. Definition 2 shares the similar idea and bound the uncertainty of $S$ . + +Theorem 1 Suppose $S(\pmb{x})$ fulfills the holis-tic Tsybakov condition for constants $C_2, \lambda_2 > 0$ , and $t_0 \in (0, m]$ . We define $\epsilon := \max_{\pmb{x}, \pmb{z}} \left[ |\hat{\bar{S}}_z^f(\pmb{x}) - \bar{S}_z^f(\pmb{x})|, |\hat{\bar{S}}_z^l(\pmb{x}) - \bar{S}_z^l(\pmb{x})|, |\bar{S}_z^l(\pmb{x}) - S_z^l(\pmb{x})| \right]$ and $\tau := \min_i T_{ii}$ . We analyze two cases: + +(1) If $\bar{\pmb{y}}$ is corrected by $\kappa(h, \pmb{x}, \bar{\pmb{y}})$ with $\hat{\delta}$ , let $\delta_1 = \min \left[\frac{\tau S_{\pmb{b}_\pmb{x}}(\pmb{x}) + \sum_{l_j \in \bar{\pmb{y}}} \sum_{i \neq j} T_{ij} \mathbb{P}(l_i | \pmb{x})}{\hat{S}_{\pmb{y}^*}(\pmb{x})}\right]$ and $\rho_1 := |\hat{\delta} - \delta_1|$ . Assume that $\epsilon \leq \frac{t_0 \tau - \rho_1 m}{3}$ . Then, $\mathbb{P}[\bar{\pmb{y}}_{new} = h^*(\pmb{x}), \bar{\pmb{y}}$ is flipped] is at least $1 - C_2 [O(\max(\epsilon, \rho_1))]^{\lambda_2} - \mathbb{P}[\pmb{a}_x \neq \{\pmb{y}^*, \bar{\pmb{y}}\}]$ . + +(2) If $\bar{\pmb{y}}$ is not corrected by $\kappa(h, \pmb{x}, \bar{\pmb{y}})$ with $\hat{\delta}$ , let $\delta_2 = \max \left[\frac{\hat{\tilde{S}}_{\bar{\pmb{y}}}(\pmb{x})}{\tau S_{b_{\pmb{x}}}(\pmb{x}) + \sum_{l_j \in \pmb{y}^*} \sum_{i \neq j} T_{ij} \mathbb{P}(l_i | \pmb{x})}\right]$ and $\rho_2 := |\hat{\delta} - \delta_2|$ . + +Assume that $\epsilon \leq \frac{t_0\delta_2^2\tau - \rho_2m - \rho_2^2m}{3\delta_2^2}$ . Then, $\mathbb{P}[\bar{\pmb{y}}_{new} = h^{*}(\pmb{x}),\bar{\pmb{y}}$ is accepted] is at least $1 - C_2[O(\max (\epsilon ,\rho_2))]^{\lambda_2} - \mathbb{P}[\pmb {a}_x\neq \{\pmb {y}^*,\bar{\pmb{y}}\} ]$ . + +The proof of Theorem 1 is provided in Appendix B.1. Theorem 1 extends the theoretical results of [75] to multilabel classification with noisy labels. It claims that, even though with noisy multiple labels, the holistic correction has a guaranteed success rate to make proper corrections. Besides, if we can reasonably approximate the optimal $\delta$ with $\hat{\delta}$ , our algorithm flips noisy multiple labels to correct ones with a good chance. Below, as a corollary of Theorem 1, we show that, there are certain circumstances, the use of holistic scores has a better chance to make corrections satisfactorily, than the sole use of instance-label dependence. + +Corollary 1 Suppose that $S(\pmb{x})$ fulfills the holistic Tsybakov condition. Denote the set threshold $\hat{\delta}$ and optimal threshold $\delta$ . We define $\rho := \max |\hat{\delta} - \delta|$ . We have that, $\exists \epsilon$ and $\rho$ , if $C_2[O(\max(\epsilon, \rho))]^{\lambda_2} < C_1[O(\max(\epsilon, \rho))]^{\lambda_1}$ , holistic correction brings higher probability to handle noisy labels successfully than instance-label dependence. + +The proof of Corollary 1 is provided in Appendix B.2. Corollary 1 claims that there exist cases where holistic scores better combat noisy labels. Note that, from a theoretical view, we do not state that holistic scores can work better in all circumstances of multi-label classification. Nevertheless, with the determination of the threshold $\hat{\delta}$ , holistic scores can perform better in the experiments of this paper, which demonstrates the help of label dependence to handle noisy multiple labels. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We verify the effectiveness of the proposed method on the synthetic noisy versions of three datasets, i.e., Pascal-VOC 2007 [12], Pascal-VOC 2012 [12], and MS-COCO [32]. Pascal-VOC 2007 contains 5,011 images in train and validation sets, while Pascal-VOC 2012 consists of 11,540 images in train and validation sets. The images come from 20 common object categories. For Pascal-VOC 2007 and Pascal-VOC 2012, we train methods using the noisy training and validation sets, and evaluate them on the test set of Pascal-VOC 2007 that has 4,952 images [14]. MS-COCO contains 82,081 training images and 40,137 validation images from 80 common object categories. As did in [74, 5, 65, 77], we evaluate the performance of methods using validation images. + +Noisy-label generation. The class-dependent noise transition matrix $T$ [42, 21, 46, 72] is used to corrupt the three + +Table 1. Comparisons with advanced methods on noisy Pascal-VOC 2007. The mean and standard deviation of results (%) are presented. + +
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE64.50±1.2058.65±2.1648.19±0.2371.77±1.1560.94±4.2548.72±2.13
CSRA66.99±0.4859.62±0.6146.97±0.4872.45±0.6963.58±1.4852.72±1.52
ADDGCN63.89±0.9455.75±1.9844.14±1.3771.02±0.9561.05±0.0650.18±2.70
APL66.79±1.1958.86±1.5347.64±1.8172.61±0.9961.99±0.7849.10±0.15
CDR67.35±1.7060.05±1.0649.12±0.5972.66±0.7964.58±0.6050.51±2.49
JOINT67.43±0.7363.37±0.9253.27±4.7070.28±1.8568.70±2.8858.57±2.75
WSIC65.43±0.5559.53±0.7348.34±0.4772.57±1.0361.88±2.5750.15±0.86
CCMN69.97±1.3662.58±1.4753.20±1.2870.68±1.0860.94±3.1248.62±1.26
HLC†72.07±0.6770.20±0.4668.00±0.8974.83±0.6469.86±1.6160.09±1.73
OF1 ↑BCE63.52±0.4856.70±2.4548.10±1.4368.28±0.6958.30±2.8251.18±3.10
CSRA65.40±0.4759.39±0.8148.32±1.5069.72±0.5061.89±0.4351.56±2.28
ADDGCN62.63±0.1855.50±1.8744.38±2.9268.95±0.6459.64±0.5653.12±0.62
APL64.85±1.4656.51±1.7047.54±2.4068.89±0.8958.04±0.9752.27±2.20
CDR65.31±0.9957.93±1.0548.86±1.7169.53±0.6559.89±1.0751.68±3.83
JOINT69.72±0.8867.93±0.7761.62±1.4071.24±1.0364.20±0.8860.30±1.24
WSIC63.45±0.9757.96±1.2548.38±2.4169.88±1.2257.97±2.1951.99±1.65
CCMN69.66±1.5560.43±1.3153.84±0.6967.12±0.6159.55±1.4553.46±1.04
HLC†71.03±0.3369.08±1.0068.62±0.4872.09±0.7465.76±2.3960.71±1.37
CF1 ↑BCE58.91±1.3453.21±2.0443.66±0.5365.93±0.8157.03±3.4347.21±1.89
CSRA62.31±0.5055.67±0.6143.11±0.7667.39±0.8059.66±1.0451.13±1.12
ADDGCN60.41±1.0453.72±1.3842.42±0.5966.05±0.9757.81±0.5848.89±2.64
APL60.23±1.5352.85±2.1842.38±1.6766.59±0.7158.33±0.4947.67±1.83
CDR61.37±1.4754.17±0.8643.60±0.8267.11±0.6359.91±0.3948.40±1.98
JOINT63.13±0.3860.22±1.6848.17±5.0166.03±1.2562.05±2.9854.03±3.17
WSIC59.54±1.1054.22±0.5343.82±0.6266.97±1.0058.04±1.7048.19±0.96
CCMN65.19±1.1058.55±1.3149.85±1.0665.47±0.9358.05±2.2448.46±0.80
HLC†68.87±0.1066.62±0.8164.82±0.4869.95±1.1965.13±1.0457.54±1.84
+ +datasets. Here, for any $i \neq j$ , $T_{ij} = \mathbb{P}(l_j \in \bar{\mathbf{y}} \wedge l_i \notin \bar{\mathbf{y}} | l_j \notin \mathbf{y} \wedge l_i \in \mathbf{y})$ represents the probability of the $i$ -th class label to be flipped into the $j$ -th class label. We consider both symmetric (abbreviated as Sym.) and pairflip (abbreviated as Pair.) noise settings [17]. The details of the transition matrix are provided in Appendix D.2. For symmetric noise, the noise rate is set to $30\%$ , $40\%$ , and $50\%$ . For pairflip noise, the noise rate is set to $20\%$ , $30\%$ , and $40\%$ . + +Baselines. We exploit three types of baselines in total. Specifically, Type-I baselines contain the methods that are designed for multi-label classification with clean labels. Type-II baselines consider the methods for multi-class classification with noisy labels. Type-III baselines consider the methods that focus on multi-label classification with noisy labels. It should be noted that, there are relatively few methods belonging to this type [36]. More advanced methods belonging to Type-III baselines need to be investigated [36], which is also our focus in this paper. In more detail, Type-I baselines include CSRA [77] and ADDGCN [65]. Type-II baselines include APL [40], CDR [58], and JOINT [48]. Type-III baselines include WSIC [22] and CCMN [63]. As a simple baseline, we compare our method with the standard deep network that directly trains on noisy datasets (abbreviated as BCE). We detail all baselines in Appendix D.1. + +Network & Optimizer. We use a ResNet-50 network [20] pretrained on ImageNet as the backbone for all methods. We train the models for 30 epochs in total. We utilize + +Adam [27] for the network optimization. The batch size is set to 128 for all the datasets. The learning rate is fixed to $5 \times 10^{-5}$ . The images in Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO resize to $224 \times 224$ . Note that, to make experiments more comprehensive, we also employ different experimental settings, e.g., different networks and different image sizes. The details are provided in Section 4.3. + +Measurement. As did in multi-label classification [77, 7], evaluation metrics include the mean average precision (mAP) [71], the average F1-measure (OF1), and the average per-class F1-measure (CF1). For fair comparison, we implement all methods with default parameters by PyTorch, and conduct all experiments on NVIDIA GTX3090 GPUs. All experiments are repeated three times with different random seeds. Following the works in learning with noisy labels [17, 54, 29, 30], the mean and standard deviation of results in the last epoch are reported. In addition, for different evaluation metrics, we report the mean and standard deviation of best results. Supplementary results are shown in Appendix E. Afterwards, the best mean results are highlighted in red. The second best mean results are also highlighted in blue. + +# 4.2. Comparison with the State-of-the-Arts + +The results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO are shown in Table 1, Table 2, and Table 3 respectively. In summary, HLC consistently works + +Table 2. Comparisons with advanced methods on noisy Pascal-VOC 2012. The mean and standard deviation of results (%) are presented. + +
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE66.74±0.8056.07±0.5045.15±1.5670.91±1.1357.61±1.1449.85±0.36
CSRA66.35±0.5056.20±1.3545.54±1.1471.29±0.8360.71±1.1847.63±1.56
ADDGCN63.34±0.9654.54±0.8644.88±1.7170.41±0.5457.96±0.6847.66±1.08
APL67.07±1.0456.79±1.8643.51±1.9371.32±1.6059.59±1.2748.14±1.16
CDR66.13±1.4956.85±0.4844.84±1.1171.55±1.8760.13±1.8949.44±1.81
JOINT65.19±2.1758.40±2.8745.13±1.6968.93±2.5461.64±1.7853.64±1.61
WSIC65.96±0.7956.34±0.4144.80±0.5470.40±1.1159.40±1.8748.95±1.34
CCMN69.15±0.6661.00±1.0150.71±0.2669.08±1.7859.72±2.3246.67±2.78
HLC†72.14±0.6670.11±0.2768.69±1.0474.51±0.6769.90±0.4364.20±1.26
OF1 ↑BCE64.99±1.1056.92±2.0845.49±2.2368.48±2.2860.21±1.3554.05±1.95
CSRA64.08±0.3756.25±2.5748.67±3.1469.06±0.6559.75±1.7052.89±0.95
ADDGCN63.53±1.4154.28±0.8647.56±2.6747.62±2.3957.90±1.7852.33±0.56
APL64.70±1.1758.05±1.6845.74±1.5570.68±1.0360.22±1.5851.38±1.55
CDR64.06±1.3857.31±1.2146.51±0.9570.45±1.4460.57±1.2452.26±2.42
JOINT67.35±1.8664.57±2.3954.37±3.3370.81±1.4064.40±1.7656.27±1.29
WSIC62.74±2.1057.13±0.7345.52±1.2869.72±1.1959.11±2.0452.49±1.38
CCMN65.77±0.2359.91±0.9351.45±0.9467.93±1.7359.26±0.5148.61±4.71
HLC†71.14±0.6069.50±0.4067.80±0.3372.13±0.2667.59±0.9664.28±0.81
CF1 ↑BCE62.47±0.4453.26±0.4143.43±1.6766.03±1.6955.90±0.7049.29±0.64
CSRA62.08±0.7053.23±1.2743.23±1.2566.02±0.7457.71±1.0247.46±1.68
ADDGCN59.67±1.1452.61±0.5244.33±1.9965.22±0.8655.32±0.7647.30±1.12
APL62.99±1.0753.69±1.8041.72±1.4266.44±1.4057.52±0.9148.14±1.02
CDR62.18±1.0453.61±0.4542.83±0.8766.29±2.1257.23±1.4349.03±1.50
JOINT60.57±2.8254.39±3.7240.48±7.7066.30±2.3359.72±2.1255.06±0.36
WSIC61.70±0.9253.10±0.7442.72±0.5465.34±1.4857.21±1.6248.51±1.12
CCMN64.46±0.6257.45±0.9948.27±0.6867.48±1.4456.93±1.6947.01±1.82
HLC†69.54±0.5667.35±0.4865.72±1.4870.07±0.4165.68±0.9460.57±1.27
+ +Table 3. Comparisons with advanced methods on noisy MS-COCO. The mean and standard deviation of results (%) are presented. + +
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE53.23±0.1547.33±0.7940.25±0.2656.58±0.2249.16±0.0441.57±0.64
CSRA53.89±0.4047.64±0.8639.58±0.1958.27±0.2350.95±0.0743.07±0.64
ADDGCN51.08±0.9544.75±1.1538.66±1.3056.94±0.6150.28±0.8141.45±0.19
APL54.34±0.3248.61±0.7243.55±1.4357.73±0.2050.87±0.3441.77±0.50
CDR54.01±0.0449.01±0.2643.94±1.2557.03±0.2850.99±0.7742.71±0.09
JOINT53.93±0.4148.01±1.0445.27±0.6857.30±0.3351.94±0.2042.74±0.55
WSIC52.99±0.5346.84±0.8639.76±0.6456.66±0.3149.46±0.2542.52±0.62
CCMN51.73±0.1850.36±0.7145.32±0.8958.13±0.4451.17±0.2942.12±0.76
HLC†54.87±0.6851.09±0.5348.15±0.5058.55±0.0953.41±0.1345.91±0.39
OF1 ↑BCE51.34±1.7044.36±0.8234.85±1.2459.16±0.9552.44±0.8142.94±1.13
CSRA52.03±1.8641.63±1.4133.47±3.1859.17±0.1450.27±0.8841.75±1.36
ADDGCN55.67±1.4847.79±0.4035.95±3.7360.96±0.6555.05±1.7847.47±0.77
APL51.07±1.3243.93±2.7033.90±4.0060.04±1.1650.64±2.8644.34±1.99
CDR53.43±1.1645.10±0.8334.91±0.9059.34±0.6152.72±0.6344.17±0.61
JOINT54.56±0.0649.00±1.6637.78±0.9358.20±0.4053.21±0.1746.55±0.61
WSIC50.91±0.5242.93±0.8535.47±1.5258.89±1.1351.63±1.5743.99±1.47
CCMN52.71±1.0443.24±1.1934.62±1.3858.61±1.1852.18±0.7645.92±0.59
HLC†59.92±0.6557.84±0.3855.47±0.9562.28±0.0658.56±0.3751.09±0.60
CF1 ↑BCE45.92±0.2338.96±1.6131.34±0.2752.54±0.5845.54±0.6339.79±0.99
CSRA44.97±1.8837.49±1.7328.96±1.1652.18±0.4444.96±0.4336.88±0.21
ADDGCN46.77±1.8039.35±1.8330.57±1.5754.18±0.2347.55±0.1839.44±0.33
APL42.91±0.5438.38±0.7728.17±2.5052.87±1.0746.27±1.2737.76±1.02
CDR46.62±0.4239.47±0.5429.59±2.5252.51±0.6945.75±0.8139.15±0.53
JOINT49.51±0.8142.38±1.2124.24±0.6154.39±0.1749.90±0.8538.34±0.55
WSIC45.30±1.0939.15±1.6231.42±0.9452.04±0.2845.76±0.7039.44±1.11
CCMN44.20±1.1935.18±1.0127.90±1.2553.23±0.5846.88±0.9240.55±0.89
HLC†51.94±0.6349.24±0.3046.69±0.6655.44±0.1350.91±0.4843.35±0.82
+ +best across all noise settings. In many cases, the best results achieved by HLC outperform the second best results by a large margin, especially when the noise level is high. Below, we further discuss the results based on the comparisons + +with three different types of baselines. + +Compared with Type-I baselines. We first notice that Type-I baselines are fragile to noisy labels in multi-label classification. Without considering the side-effect of noisy + +labels, in many cases, they perform worse than BCE, which clearly illustrates the necessity for attention to handling noisy labels. Second, we compare HLC with ADDGCN. Without the proposed correction method for combating noisy labels, HLC will reduce to ADDGCN. As shown in the reported results, HLC performs much better than ADDGCN. To be specific, on noisy Pascal-VOC 2007, for Sym. $40\%$ , HLC brings about $+15\%$ performance improvement w.r.t. three evaluation metrics over ADDGCN. For Sym. $50\%$ , the performance improvement is increased to more than $+20\%$ . Also, for Pair. $30\%$ and Pair. $40\%$ , HLC enhances ADDGCN with about $+10\%$ improvement. On noisy Pascal-VOC 2012 and MS-COCO, the performance improvement is also very clear. + +Compared with Type-II baselines. On noisy Pascal-VOC 2007, with Sym. noise, we can see that HLC outperforms APL, CDR, and JOINT clearly, especially for Sym. $50\%$ . Additionally, with Pair. noise, although the improvement is less than the cases with Sym. noise, HLC still performs best. On noisy Pascal-VOC 2012, for both Sym. and Pair. noise, the improvement is significant. Lastly, for noisy MS-COCO, HLC works better than all Type-II baselines with varying enhancement. + +Note that, compared with APL and CDR, JOINT seems to be a stronger baseline. Benefiting from label correction, after a few training epochs, JOINT less overfits to wrong labels, following better performance. Nevertheless, the proposed label-correction paradigm is argued to be more advanced. As shown in all results, HLC surpasses JOINT, which verifies the effectiveness of our method. + +Compared with Type-III baselines. On noisy Pascal-VOC 2007 and noisy Pascal-VOC 2012, HLC outperforms WSIC and CCMN distinctly. For example, with Sym. $50\%$ noise, more than $+10\%$ performance promotion is brought by our method. On noisy MS-COCO, although WSIC and CCMN are sometimes competitive w.r.t. mAP, they are inferior w.r.t. both OF1 and CF1. + +# 4.3. More Analyses and Justifications + +In this subsection, we conduct performance analysis in more detail. The experiments are conducted with Sym. $50\%$ noise, which is more challenging than the experiments in low-noise-rate cases. + +Role of label dependence. We study the effect of removing the consideration of label dependence to provide insights into what makes HLC successful. The experiments are conducted on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to $224 \times 224$ . Recall that HLC considers instance-label and label dependences simultaneously. When we remove the consideration of the label dependence in HLC, the correspond + +Table 4. Ablation study results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The mean and standard deviation of results are presented. The best result in each case is in bold. + +
DatasetNoisy Pascal-VOC 2007
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.06±0.4167.23±1.9263.42±0.58
HLC68.00±0.8968.62±0.4864.82±0.48
DatasetNoisy Pascal-VOC 2012
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.88±0.7566.30±1.2864.33±1.67
HLC68.69±1.0467.80±0.3365.72±1.48
DatasetNoisy MS-COCO
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.46.21±0.3652.90±0.9244.51±1.29
HLC48.15±0.5055.47±0.9546.69±0.66
+ +![](images/a3dc4cc0c589ddd7719ba50a4ffd2c9521624dae137206fe861a7090f5299cc2.jpg) + +![](images/514fc574be132d5f5c5d4bca61e6f2621079a405bfa991e66ed9f4d718c2b0e6.jpg) +Figure 2. Ablation study results with different values of the set threshold $\hat{\delta}$ . The experiments are conducted on noisy Pascal-VOC 2007 (Top) and noisy Pascal-VOC 2012 (Bottom). + +ing method is named as HLC w/o l. here. For both HLC w/o l. and HLC, the value of the threshold $\hat{\delta}$ is searched in the range $\{0.25, 0.30, 0.35, 0.40, 0.45\}$ . We use the $10\%$ noisy training data as a validation set for the threshold determination and performance report. The results are shown in Table 4. As can be seen, HLC outperforms HLC w/o l.. The results justify our claims that the label dependence could help combat the noisy labels in multi-label classification, which demonstrate the effectiveness of the proposed holistic correction. + +Analysis of the threshold $\hat{\delta}$ . We analyze the influence of different values of $\hat{\delta}$ . The experiments are conducted on noisy Pascal-VOC 2007 and Pascal-VOC 2012. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to $224 \times 224$ . The value of the threshold $\hat{\delta}$ is chosen in $\{0.25, 0.30, 0.35, 0.40, 0.45\}$ . Figure 2 shows that HLC is robust to the determination of the threshold $\hat{\delta}$ in the certain range, which facilitates the practical application of our method. + +Evaluations with different networks. We use pretrained + +Table 5. Comparisons with advanced methods on noisy MS COCO with different networks. The mean and standard deviation of results $(\%)$ are presented. + +
MetricsMethodsResNet-34ResNet-101
mAP ↑BCE42.63±0.7438.17±0.41
CSRA41.35±0.1837.24±1.20
ADDGCN40.15±0.9836.13±0.69
APL44.82±0.7040.90±1.51
CDR45.43±0.6541.00±0.38
JOINT44.81±0.7739.96±1.30
WSIC41.86±0.6237.49±0.68
CCMN45.31±0.4746.01±1.01
HLC†46.05±0.8146.24±2.13
OF1 ↑BCE37.65±2.4638.65±2.50
CSRA35.05±0.7834.28±2.40
ADDGCN35.11±1.2637.18±0.50
APL34.31±1.7337.89±1.30
CDR36.67±3.4339.88±1.15
JOINT39.66±1.1341.36±0.88
WSIC35.08±1.7438.44±0.57
CCMN32.86±1.5037.03±1.48
HLC†44.11±0.8047.79±4.19
CF1 ↑BCE28.95±2.0934.11±1.13
CSRA27.40±0.4431.21±1.57
ADDGCN26.11±0.8630.41±0.72
APL26.64±2.3931.97±1.95
CDR27.13±2.3935.17±1.06
JOINT30.77±1.6337.63±0.81
WSIC27.62±0.6533.83±0.61
CCMN24.75±0.4826.65±0.26
HLC†34.79±1.4140.88±2.42
+ +ResNet-50 before. To show that our method is robust to the choice of network structures, we use different networks in experiments. Specifically, we employ pretrained ResNet-34 [20] and pretrained ResNet-101 [20] respectively. The noisy MS-COCO is considered. The image size is $224 \times 224$ . The results on mAP are reported in Table 5. As can be seen, with different networks, HLC still works well. + +Evaluations with different image sizes. We resize the image size to $224 \times 224$ before. To test the performance of advanced methods with different image sizes, we further consider $112 \times 112$ , $384 \times 384$ , and $448 \times 448$ image sizes. Pretrained ResNet-50 is used. The results are reported in Table 6. For mAP, we can see that HLC is competitive compared with CCMN and CSRA. For OF1 and CF1, HLC works better than all baselines with a clear margin. + +# 4.4. Experiments on the Real-world Dataset + +To demonstrate that our problem setting can be adapted to the real world and our method can well handle practical scenes, we employ the real-world dataset NUS-WIDE [10] that originally contained 269,648 images from Flicker, which have been manually annotated with 81 visual concepts. Since some urls for download have been deleted, we employ the dataset version in [45]. A standard 70-30 train-test split is used. The backbone is chosen as ResNet-101. As the computation cost of training on NUS-WIDE is rela + +Table 6. Comparisons with advanced methods on noisy MS COCO. The mean and standard deviation of results $(\%)$ are presented. Difference image sizes are considered here. + +
MetricsImage sizes112 × 112384 × 384448 × 448
mAP ↑BCE32.22±0.6939.40±1.3635.24±1.73
CSRA29.55±0.1643.55±0.7044.56±0.75
ADDGCN32.34±0.4638.72±1.6434.87±1.89
APL34.41±0.4843.65±0.2841.44±1.21
CDR34.75±0.3943.26±0.7239.97±1.40
JOINT32.89±0.1642.95±0.8840.17±1.26
WSIC31.98±0.2339.57±1.0236.08±0.23
CCMN36.17±0.4144.39±0.3944.03±0.17
HLC†35.98±1.0545.12±0.1344.23±1.20
OF1 ↑BCE26.70±0.8834.71±2.7626.72±3.35
CSRA20.14±1.2836.41±0.7138.54±0.78
ADDGCN26.36±2.2942.83±2.0540.73±1.04
APL24.02±1.2234.68±1.4630.73±2.64
CDR26.50±1.2331.31±1.7631.15±3.46
JOINT34.11±0.9538.67±1.2538.11±0.69
WSIC24.61±1.1034.09±2.9430.69±0.97
CCMN23.89±1.4936.16±2.1225.03±2.48
HLC†39.05±2.6846.55±3.7745.14±2.34
CF1 ↑BCE19.61±0.6130.77±2.2824.94±3.31
CSRA13.34±1.2932.66±0.9832.68±0.22
ADDGCN18.67±1.4735.63±1.3733.89±2.41
APL17.21±0.7829.24±0.0925.01±1.31
CDR18.04±1.0729.62±1.1926.28±1.23
JOINT20.76±0.7534.90±1.8833.75±1.31
WSIC19.07±0.5731.63±2.4928.50±1.36
CCMN16.75±1.2130.09±2.3226.68±1.57
HLC†29.70±2.0740.34±2.1437.48±2.36
+ +Table 7. Comparison of our method to known state-of-the-art models on the NUS-WIDE dataset. Metrics are in %. + +
MethodmAP ↑OF1 ↑CF1 ↑
S-CLs [39]60.173.758.7
MS-CMA [66]61.473.860.5
SRN [76]62.073.458.5
ICME [7]62.874.160.7
ASL [45]63.974.662.7
HLC†63.174.662.9
HLC+ASL†64.575.163.4
+ +tively large, we run experiments one time. Here we compare our method with S-CLs [39], MS-CMA [66], SRN [76], ICME [7], and ASL [45]. For convenient comparison, we refer to the results of their original papers. Note that to further improve the performance on NUS-WIDE, we utilize ASL to replace the loss function of our method. We name the new method "HLC+ASL". Results are provided in Table 7, which demonstrate the effectiveness of our method on the real-world dataset. + +# 5. Conclusion + +In this paper, we focus on the realistic problem of multi-label classification with noisy labels. We learn and utilize the label dependence among multiple labels to handle this problem. With the help of label dependence, a novel algorithm named HLC is proposed to correct noisy multiple labels to clean ones. We demonstrate the effectiveness of our algorithm both theoretically and empirically. For future + +work, we are interested in adapting HLC to other domains such as natural language processing and recommendation systems. We are also interested in promoting our algorithm to tackle instance-dependent label noise [73, 4, 78, 38] in multi-label classification. + +# Acknowledgements + +Bo Han was supported by NSFC Young Scientists Fund No. 62006202 and Guangdong Basic and Applied Basic Research Foundation No. 2022A1515011652. Tongliang Liu is partially supported by ARC projects: FT220100318, DP220102121, LP220100527, LP220200949, and IC190100031. + +# References + +[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, pages 233-242, 2017. 2 +[2] Dara Bahri, Heinrich Jiang, and Maya Gupta. Deep k-nn for noisy labels. In ICML, pages 540–550, 2020. 4 +[3] Yingbin Bai and Tongliang Liu. Me-momentum: Extracting hard confident examples from noisily labeled data. In ICCV, pages 9312-9321, 2021. 2 +[4] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. arXiv preprint arXiv:2001.03772, 2020. 9 +[5] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In ICCV, pages 522-531, 2019. 4 +[6] Zhaomin Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Learning graph convolutional networks for multi-label recognition and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3 +[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME, pages 622-627, 2019. 1, 3, 5, 8 +[8] Xiang Cheng, Shu-Guang Zhao, Xuan Xiao, and Kuo-Chen Chou. iatc-misf: a multi-label classifier for predicting the classes of anatomical therapeutic chemicals. Bioinformatics, 33(3):341-346, 2017. 1 +[9] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. arXiv preprint arXiv:2109.04997, 2021. 1 +[10] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In ACM International Conference on Image and Video Retrieval, pages 1-9,2009.8 + +[11] Zijun Cui, Yong Zhang, and Qiang Ji. Label error correction and generation through label relationships. In AAAI, pages 3693-3700, 2020. 1 +[12] Mark Everingham, Andrew Zisserman, Christopher KI Williams, Luc Van Gool, Moray Allan, Christopher M Bishop, Olivier Chapelle, Navneet Dalal, Thomas Deselaers, Gyuri Dorkó, et al. The pascal visual object classes challenge 2007 (voc2007) results. 2008. 4 +[13] Lei Feng, Senlin Shu, Zhuoyi Lin, Fengmao Lv, Li Li, and Bo An. Can cross entropy loss be robust to label noise? In IJCAI, pages 2206-2212, 2021. 1 +[14] Bin-Bin Gao and Hong-Yu Zhou. Learning to discover multi-class attentional regions for multi-label image recognition. IEEE Transactions on Image Processing, 30:5920-5932, 2021. 1, 4 +[15] Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy labels. arXiv preprint arXiv:1607.07526, 2016. 4 +[16] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor Tsang, and Masashi Sugiyama. *Sigua: Forgetting may make learning with noisy labels more robust.* In ICML, pages 4006–4016, 2020. 1 +[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages 8527-8537, 2018. 5 +[18] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3 +[19] Jun-Yi Hang and Min-Ling Zhang. Dual perspective of label-specific feature learning for multi-label classification. In ICML, pages 8375-8386, 2022. 1 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 8 +[21] Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In NeurIPS, 2018. 4 +[22] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly supervised image classification through noise regularization. In CVPR, pages 11517-11525, 2019. 5 +[23] Zhuo Huang, Xiaobo Xia, Li Shen, Bo Han, Mingming Gong, Chen Gong, and Tongliang Liu. Harnessing out-of-distribution examples via augmenting content and style. In ICLR, 2023. 1 +[24] Zhuo Huang, Miaoxi Zhu, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, Bo Du, and Tongliang Liu. Robust generalization against photon-limited corruptions via worst-case sharpness minimization. In CVPR, pages 16175-16185, 2023. 1 +[25] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, pages 2309-2318, 2018. 2, 3 + +[26] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. Nlnl: Negative learning for noisy labels. In ICCV, pages 101-110, 2019. 1 +[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[28] Cheng Li, Bingyu Wang, Virgil Pavlu, and Javed Aslam. Conditional bernoulli mixtures for multi-label classification. In ICML, pages 2482-2491, 2016. 3 +[29] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. 5 +[30] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In ICCV, pages 9485-9494, 2021. 5 +[31] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, 2022. 1 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 4 +[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. In NeurIPS, 2020. 3 +[34] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2016. 2 +[35] Weiwei Liu, Ivor W Tsang, and Klaus-Robert Müller. An easy-to-hard learning paradigm for multiple classes and multiple labels. Journal of Machine Learning Research, 18, 2017. 1 +[36] Weiwei Liu, Haobo Wang, Xiaobo Shen, and Ivor Tsang. The emerging trends of multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 5 +[37] Weiwei Liu, Donna Xu, Ivor W Tsang, and Wenjie Zhang. Metric learning for multi-output tasks. Transactions on Pattern Analysis and Machine Intelligence, 41(2):408-422, 2018. 1 +[38] Yang Liu. Identifiability of label noise transition matrix. arXiv preprint arXiv:2202.02016, 2022. 9 +[39] Yongcheng Liu, Lu Sheng, Jing Shao, Junjie Yan, Shiming Xiang, and Chunhong Pan. Multi-label image classification via knowledge distillation from weakly-supervised detection. In ACMMM, pages 700-708, 2018. 8 +[40] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. In ICML, pages 6543-6553, 2020. 5 +[41] Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. Modeling fine-grained entity types with box embeddings. In ACL, 2021. 1 +[42] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks + +robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4 +[43] Cosmin Octavian Pene, Amirmasoud Ghiassi, Taraneh Younesian, Robert Birke, and Lydia Y Chen. Multi-label gold asymmetric loss correction with single-label regulators. arXiv preprint arXiv:2108.02032, 2021. 2 +[44] Geoff Pleiss, Tianyi Zhang, Ethan R Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. In NeurIPS, 2020. 1 +[45] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In ICCV, 2021. 1, 8 +[46] Jun Shu, Qian Zhao, Zengben Xu, and Deyu Meng. Meta transition adaptation for robust deep learning with noisy labels. arXiv preprint arXiv:2006.05697, 2020. 2, 4 +[47] Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1 +[48] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, 2018. 5 +[49] Alexander B Tsybakov. Optimal aggregation of classifiers in statistical learning. The Annals of Statistics, 32(1):135-166, 2004. 4 +[50] Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks. In NeurIPS, pages 5596-5605, 2017. 1 +[51] Andreas Veit, Neil Alldrin, Gal Chechik, Ivan Krasin, Abhinav Gupta, and Serge Belongie. Learning from noisy large-scale datasets with minimal supervision. In CVPR, 2017. 1 +[52] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, pages 2285-2294, 2016. 1 +[53] Xinshao Wang, Yang Hua, Elyor Kodirov, David A Clifton, and Neil M Robertson. Proselflc: Progressive self label correction for training robust deep neural networks. In CVPR, pages 752-761, 2021. 2 +[54] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pages 8688-8696, 2018. 5 +[55] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, pages 13726–13735, 2020. 1 +[56] Hongxin Wei, Lue Tao, Renchunzi Xie, and Bo An. Open-set label noise can improve robustness against inherent label noise. In NeurIPS, pages 7978-7992, 2021. 1 +[57] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed graph. In ICCV, pages 4157-4165, 2015. 3 +[58] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hinding the memorization of noisy labels. In ICLR, 2021. 1, 5 + +[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In ICLR, 2022. 1 +[60] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 2 +[61] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835-6846, 2019. 1 +[62] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning. In AAAI, 2018. 3 +[63] Ming-Kun Xie and Sheng-Jun Huang. Ccmn: A general framework for learning with class-conditional multi-label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 5 +[64] Ming-Kun Xie, Jia-Hao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. In NeurIPS, 2022. 1 +[65] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, pages 649-665, 2020. 1, 2, 3, 4, 5 +[66] Renchun You, Zhiyao Guo, Lei Cui, Xiang Long, Yingze Bao, and Shilei Wen. Cross-modality attention with semantic graph embedding for multi-label classification. In AAAI, pages 12709-12716, 2020. 8 +[67] Hsiang-Fu Yu, Prateek Jain, Purushottam Kar, and Inderjit Dhillon. Large-scale multi-label learning with missing labels. In ICML, pages 593-601, 2014. 3 +[68] Ze-Bang Yu and Min-Ling Zhang. Multi-label classification with label-specific feature generation: A wrapped approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3 +[69] Jiong Zhang, Wei-cheng Chang, Hsiang-fu Yu, and Inderjit Dhillon. Fast multi-resolution transformer fine-tuning for extreme multi-label text classification. In NeurIPS, 2021. 1 +[70] Min-Ling Zhang and Kun Zhang. Multi-label learning by exploiting label dependency. In KDD, pages 999-1008, 2010. 1, 3 +[71] Min-Ling Zhang and Zhi-Hua Zhou. A review on multi-label learning algorithms. IEEE Transactions on Knowledge and Data Engineering, 26(8):1819-1837, 2013. 1, 5 +[72] Yivan Zhang, Gang Niu, and Masashi Sugiyama. Learning noise transition matrix from only noisy labels via total variation regularization. In ICML, 2021. 4 +[73] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In ICLR, 2021. 9 +[74] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In ICCV, pages 163-172, 2021. 1, 4 +[75] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded + +correction of noisy labels. In ICML, pages 11447-11457, 2020. 1, 4 +[76] Feng Zhu, Hongsheng Li, Wanli Ouyang, Nenghai Yu, and Xiaogang Wang. Learning spatial regularization with image-level supervisions for multi-label image classification. In CVPR, pages 5513-5522, 2017. 8 +[77] Ke Zhu and Jianxin Wu. Residual attention: A simple but effective method for multi-label recognition. In ICCV, pages 184-193, 2021. 1, 4, 5 +[78] Zhaowei Zhu, Tongliang Liu, and Yang Liu. A second-order approach to learning with instance-dependent label noise. In CVPR, pages 10113-10123, 2021. 9 \ No newline at end of file diff --git a/2023/Holistic Label Correction for Noisy Multi-Label Classification/images.zip b/2023/Holistic Label Correction for Noisy Multi-Label Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ff521cbe4b7a9e3d28f68535fbbb5332a4e5eb28 --- /dev/null +++ b/2023/Holistic Label Correction for Noisy Multi-Label Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:241b000114427f056de244a0102e51a87ffc52f565c3d71b4649cbcc64a6253e +size 1000546 diff --git a/2023/Holistic Label Correction for Noisy Multi-Label Classification/layout.json b/2023/Holistic Label Correction for Noisy Multi-Label Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b23b0102d4ee55f0139c2103c60c74329913c7bd --- /dev/null +++ b/2023/Holistic Label Correction for Noisy Multi-Label Classification/layout.json @@ -0,0 +1,10196 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 488, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 488, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 488, + 120 + ], + "type": "text", + "content": "Holistic Label Correction for Noisy Multi-Label Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "spans": [ + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "content": "Xiaobo Xia" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "content": ", Jiankang Deng" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "content": ", Wei Bao" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "content": ", xuan Du" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "content": ", Bo Han" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "content": ", Shiguang Shan" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{5,6}" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "text", + "content": ", Tongliang Liu" + }, + { + "bbox": [ + 175, + 142, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 171, + 432, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 171, + 432, + 185 + ], + "spans": [ + { + "bbox": [ + 161, + 171, + 432, + 185 + ], + "type": "text", + "content": "1The University of Sydney 2Imperial College London" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 185, + 434, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 185, + 434, + 199 + ], + "spans": [ + { + "bbox": [ + 160, + 185, + 434, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 160, + 185, + 434, + 199 + ], + "type": "text", + "content": "JD Explore Academy " + }, + { + "bbox": [ + 160, + 185, + 434, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 160, + 185, + 434, + 199 + ], + "type": "text", + "content": "Hong Kong Baptist University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 199, + 488, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 199, + 488, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 488, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 106, + 199, + 488, + 213 + ], + "type": "text", + "content": "Chinese Academy of Sciences " + }, + { + "bbox": [ + 106, + 199, + 488, + 213 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 106, + 199, + 488, + 213 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 266, + 289, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 289, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 289, + 578 + ], + "type": "text", + "content": "Multi-label classification aims to learn classification models from instances associated with multiple labels. It is pivotal to learn and utilize the label dependence among multiple labels in multi-label classification. As a result of today's big and complex data, noisy labels are inevitable, making it looming to target multi-label classification with noisy labels. Although the importance of label dependence has been shown in multi-label classification with clean labels, it is challenging and hard to bring label dependence to the problem of multi-label classification with noisy labels. The issues are, that we do not understand why label dependence is helpful in the problem, and how to learn and utilize label dependence only using training data with noisy multiple labels. In this paper, we bring label dependence to tackle the problem of multi-label classification with noisy labels. Specifically, we first provide a high-level understanding of why label dependence helps distinguish the examples with clean/noisy multiple labels. Benefiting from the memorization effect in handling noisy labels, a novel algorithm is then proposed to learn the label dependence by only employing training data with noisy multiple labels, and utilize the learned dependence to help correct noisy multiple labels to clean ones. We prove that the use of label dependence could bring a higher success rate for recovering correct multiple labels. Empirical evaluations justify our claims and demonstrate the superiority of our algorithm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "type": "text", + "content": "Multi-label classification assigns a set of multiple labels for each instance [71]. As a practical learning paradigm, multi-label classification has been widely applied in various domains, ranging from computer vision [7] and natural language processing [41], to recommendation systems [69] and bioinformatics [8]. Consensually, compared" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 242, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 545, + 302 + ], + "type": "text", + "content": "with multi-class classification [20, 23, 24], where each instance is assigned with a single label, multi-label classification is more challenging [35]. Plenty of advanced methods are proposed in recent years for multi-label classification [77, 45, 14, 74, 37, 9, 19, 64]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 304, + 546, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 543 + ], + "type": "text", + "content": "The great majority of the methods assume that training data are annotated precisely. However, noisy labels are inevitable in multi-label classification [36], especially for classification with big and complex data. They may be resulted by unintentional mistakes of manual and automatic annotators [51, 75, 13], or intentional corruptions on clean labels [50, 44]. Noisy labels severely impair the generalization of learned models, over-parameterized deep models in particular [26, 61, 58, 59, 55, 56]. A straightforward way to address the problem of multi-label classification with noisy labels is to treat each label in isolation and convert the multi-label problem into a number of binary classification problems. Afterward, the methods in multi-class classification with noisy labels [16, 47] are applied to train independent binary classifiers, which capture instance-label dependence robustly to strengthen classification. This way is a remedy to handle noisy labels, but ignores the label dependence among multiple labels. It is essential to learn and utilize the label dependence in multi-label classification [70, 18, 11, 31]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "Prior works [65, 6, 52] illustrate the successes of considering the label dependence among multiple labels in multilabel classification with clean labels. In different ways, e.g., helping learn inter-dependent classifiers [7], the label dependence can be used to boost the learning of the instance-label dependence, which improves final classification. Inspired by the successes, it is concerned that label dependence could be exploited to handle the problem of multilabel classification with noisy labels. However, there are few attempts before for this important problem. At least three questions make the solution remain mysterious. First, in intuition, we need to understand why label dependence is helpful for the problem. Second, in technique, we need to know how to learn and utilize the label dependence in the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "type": "text", + "content": "*Corresponding author (tongliang.liu@sydney.edu.au)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1483" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 72, + 511, + 167 + ], + "blocks": [ + { + "bbox": [ + 84, + 72, + 511, + 167 + ], + "lines": [ + { + "bbox": [ + 84, + 72, + 511, + 167 + ], + "spans": [ + { + "bbox": [ + 84, + 72, + 511, + 167 + ], + "type": "image", + "image_path": "b4b3ab45ab5805c96c8957631002d585b9fbac6a93b1b94f640454d10a95c141.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 547, + 237 + ], + "type": "text", + "content": "Figure 1. The illustration of why the label dependence among multiple labels helps distinguish the examples with noisy/clean multiple labels. The arrow presents the label dependence between a label pair. For the labels “a” and “b”, “a → b” means that, when “a” appears, “b” will also occur with high probability. The example comes from a web search. The set of clean multiple labels is {Sea, Human, Motorboat}, where the label dependence is strong with both “Motorboat → Sea” and “Motorboat → Human”. However, due to label corruption, Motorboat is flipped to be Motorcycle, which causes “Motorcycle → Sea”. Therefore, the label dependence among noisy multiple labels is weaker than the label dependence among corresponding clean ones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 256, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 304 + ], + "type": "text", + "content": "problem. As we only have training data with noisy labels, both the accurate catch and application of the label dependence are challenging. Third, in verification, we need to know what improvements the label dependence can bring." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 288, + 640 + ], + "type": "text", + "content": "In this paper, we answer the three questions one by one. The first answer is illustrated in Figure 1. That is, compared with noisy multiple labels, the label dependence among clean multiple labels is stronger with high probability. Therefore, such dependence could help distinguish the examples with noisy/clean multiple labels for our problem. The second answer is given by the proposed holistic correction for multi-label classification with noisy labels (aka HLC). Specifically, HLC inherits the memorization effect in handling noisy labels [1, 25, 53]: the deep model would firstly memorize the training examples with clean labels, leading to reliable model predictions in early training. In HLC, the label dependence is learned by a dynamic graph [65], and then applied to correcting noisy multiple labels. In more detail, the holistic score in HLC is proposed to measure the instance-label and label dependencies in an example. The stronger instance-label and label dependencies make a larger holistic score. We compare the ratio between the holistic scores of the example with noisy multiple labels and its variant with predicted multiple labels, with an easily determined threshold. The noisy multiple labels are corrected or changeless based on the comparison result. Benefiting from the memorization effect, both dependence learning and multi-label correction are useful. Besides, they fulfill a positive cycle [3]. Namely, better dependence learning results in a better multi-label correction, and better multi-label correction makes better dependence learning, leading to final enhanced classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "The third answer is given by both theoretical analyses and empirical evaluations. Theoretically, we show that the additional use of label dependence brings a higher probability to handle noisy multiple labels successfully than the sole use of instance-label dependence under some conditions. Empirically, we demonstrate the power of label de" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 256, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 292 + ], + "type": "text", + "content": "pendence through experiments and show that, in most situations, HLC outperforms comparison methods with large margins." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 293, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 484 + ], + "type": "text", + "content": "The contributions of this paper are summarized as follows. (1) We focus on a realistic problem of multi-label classification with noisy labels. The challenges of using label dependency to address the problem are carefully analyzed, which benefits future research on the problem. (2) We propose an effective method to handle noisy labels in multi-label classification. The method measures simultaneous instance-label and label dependencies in an example for follow-up label correction. (3) Theoretical analysis is provided to explain the success rate of the proposed method. Besides, we confirm that the use of label dependence is indeed powerful under some conditions. (4) Extensive empirical results on multiple benchmarks demonstrate the superiority of our method. Detailed ablation studies and discussions are also provided. Codes are attached in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 497, + 390, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 390, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 390, + 510 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": "Problem setup. Let " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " denote the input space and " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} \\in \\{l_1, \\dots, l_q\\}" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " denote the label space with " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " class labels. An example with multiple labels is denoted as " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " is the feature vector of an instance, and " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "y \\subseteq \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " is its set of associated labels. Denote the size of the label set " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "|y|" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ". For the feature vector " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ", its label set " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " may be corrupted and is flipped into " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\bar{y} \\subseteq \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ". We utilize a class-dependent noise transition matrix " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " [43, 46, 34, 60] to characterize the label flip process. Formally, for any " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "T_{ij} = \\mathbb{P}(l_j \\in \\bar{y} \\wedge l_i \\notin \\bar{y} | l_j \\notin y \\wedge l_i \\in y)" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": " represents the probability of the " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": "-th class label to be flipped into the " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": "-th class label. Consider a noisy multi-label dataset comprising several examples " + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "inline_equation", + "content": "(x, \\bar{y})" + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": ". The aim is to learn a classification model robustly by only using the noisy dataset. Given an instance in testing, with the learned model, we can predict its relevant label set precisely." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1484" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 288 + ], + "type": "text", + "content": "Note that some works employ another problem setting that the total number of multiple labels can be changed after label flipping, which is referred to as multi-label classification with missing or redundant labels. For the former setting, it is not accurate to consider it as a classification with noisy labels, since all annotated labels are correct [67, 57]. For the latter setting, it is normally called partial multi-label learning [62], which is different from the problem setting of this paper, as detailed in Appendix C.4. Our setting, i.e., the total number of labels is preserved after label flipping, is realistic. In many practical situations, it is easy to determine the number of objects in an image, in particular with object detection techniques. In contrast, it can be harder to annotate the objects perfectly, resulting in noisy labels. In addition, in Section 4.4, we will show that our problem setting well fits the realistic situation. That is to say, the proposed method can achieve superior performance on a realistic noisy multi-label dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "text", + "content": "Preparation technology. As discussed, we need both the instance-label dependence and the label dependence among multiple labels. Given an example " + }, + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "text", + "content": ", for the instance-label dependence, it can be learned with the conditional probability of " + }, + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "inline_equation", + "content": "l_{i}\\in \\pmb{y}" + }, + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 291, + 288, + 387 + ], + "type": "text", + "content": " according to model's probability outputs. For the label dependence among multiple labels, it is often estimated by counting the occurrence of label pairs in training data [7]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 387, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 289, + 602 + ], + "type": "text", + "content": "Recently, the graph convolutional network (GCN) is used in multi-label classification and achieves great successes [7, 65, 6]. The advantage of the GCN-based methods is that they can capture the instance-label and label dependencies simultaneously during training. In this paper, we inherit the advantage of the GCN-based methods and build HLC based on ADDGCN [65]. ADDGCN designs a semantic attention module (SAM) to estimate the content-aware class-label representations for each class from the extracted feature map. The representations are fed into a GCN module (GCNM) for final classification. We provide the technical details of ADDGCN [65] in Appendix C.1. Before delving into the next section, readers only need to remember that the instance-label and label dependencies can be learned during training. Note that we also review prior works on multi-class classification with noisy labels and multi-label classification with clean/noisy labels in Appendix C.2 and Appendix C.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 612, + 153, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 153, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 153, + 625 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 631, + 287, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 287, + 644 + ], + "type": "text", + "content": "3.1. Holistic Judgment in Multi-Label Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": "Holistic score. We begin with an example with clean multiple labels. Given an example " + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": ", we can measure the instance-label dependence " + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "S^f" + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": ", and the label dependence " + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "S^l" + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": ". Denote the variable of clean multiple labels by " + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{Y}" + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": ". Mathematically, we define two dependen" + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 310, + 88, + 476, + 167 + ], + "blocks": [ + { + "bbox": [ + 310, + 74, + 452, + 87 + ], + "lines": [ + { + "bbox": [ + 310, + 74, + 452, + 87 + ], + "spans": [ + { + "bbox": [ + 310, + 74, + 452, + 87 + ], + "type": "text", + "content": "Algorithm 1: Holistic Correction." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "lines": [ + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "spans": [ + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "(\\pmb {x},\\bar{\\pmb{y}})" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": " \nOutput: " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "\\bar{y}_{new}" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "\\pmb {y}^{*} = h(\\pmb {x})" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": " . \n2: " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "\\kappa (h,\\pmb {x},\\bar{\\pmb{y}}) = \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}(\\pmb {x}) / \\hat{\\bar{S}}_{\\pmb{y}^{*}}(\\pmb {x})" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": " \n3: if " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "\\kappa (h,\\pmb {x},\\bar{\\pmb{y}})\\leq \\hat{\\delta}" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "\\bar{y}_{new} = y^{*}" + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "text", + "content": " \n4: else " + }, + { + "bbox": [ + 310, + 88, + 476, + 167 + ], + "type": "inline_equation", + "content": "\\bar{y}_{new} = \\bar{y}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": "cies as " + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "inline_equation", + "content": "S_{\\pmb{z}}^{f}(\\pmb{x}) \\coloneqq \\sum_{\\{\\pmb{Y} = \\pmb{z}, l_{i} \\in \\pmb{z}\\}} \\mathbb{P}(l_{i}|\\pmb{x})" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "inline_equation", + "content": "S_{\\pmb{z}}^{l}(\\pmb{x}) \\coloneqq \\sum_{\\{\\pmb{Y} = \\pmb{z}, l_{i}, l_{j} \\in \\pmb{z}\\}} \\frac{1}{2} [\\mathbb{P}(l_{j}|l_{i}, \\pmb{x}) + \\mathbb{P}(l_{i}|l_{j}, \\pmb{x})]" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": " is the value of the random variable " + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{Y}" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": ". The holistic score of the example " + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "inline_equation", + "content": "(\\pmb{x}, \\pmb{y})" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": " considers two dependencies at the same time. Formally, we denote the holistic score of " + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "inline_equation", + "content": "(\\pmb{x}, \\pmb{y})" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "inline_equation", + "content": "S_{\\pmb{y}}(\\pmb{x})" + }, + { + "bbox": [ + 304, + 190, + 547, + 266 + ], + "type": "text", + "content": " and define it as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 271, + 545, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 271, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 369, + 271, + 545, + 287 + ], + "type": "interline_equation", + "content": "S _ {\\boldsymbol {y}} (\\boldsymbol {x}) := S _ {\\boldsymbol {y}} ^ {f} (\\boldsymbol {x}) + S _ {\\boldsymbol {y}} ^ {l} (\\boldsymbol {x}). \\tag {1}", + "image_path": "c67b2a878ce4b038c946053080544fa98aed231b234ba5d348228f561f0f67dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": "Afterward, denote the variable of noisy multiple labels by " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{Y}" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ". For the example with noisy multiple labels, i.e., " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\bar{\\pmb{y}})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ", the instance-label dependence and label dependence are measure by " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_z^f (\\pmb {x})\\coloneqq \\sum_{\\{\\bar{Y} = \\pmb {z},l_i\\in \\pmb {z}\\}}\\mathbb{P}(l_i|\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_z^l (\\pmb {x})\\coloneqq" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\sum_{\\{\\bar{Y} = z,l_i,l_j\\in \\pmb {z}\\}}\\frac{1}{2}\\left[\\mathbb{P}(l_j|l_i,\\pmb {x}) + \\mathbb{P}(l_i|l_j,\\pmb {x})\\right]" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ". Accordingly, the holistic score of the example " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "(\\pmb {x},\\bar{\\pmb{y}})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\bar{\\pmb{y}}}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ", which is defined as " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\bar{\\pmb{y}}}(\\pmb {x})\\coloneqq \\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x}) + \\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ". Note that, during training, we cannot access " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ". Instead, the estimated posterior probabilities are used. We denote the estimations of " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\bar{S}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ". The estimation of the holistic score is " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}(\\pmb {x}) = \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x}) + \\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": ". With preparation technology discussed in Section 2 and Appendix C.1, " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{f}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "inline_equation", + "content": "\\hat{\\bar{S}}_{\\bar{\\pmb{y}}}^{l}(\\pmb {x})" + }, + { + "bbox": [ + 304, + 293, + 547, + 482 + ], + "type": "text", + "content": " can be obtained." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "content": "Holistic correction. For the example " + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\bar{\\pmb{y}})" + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "content": ", we feed it into the deep network " + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "content": " included in ADDGCN [65]. The memorization effect in handling noisy labels [25, 33] shows that the deep network would first memorize the training data with clean labels and then the training data with noisy labels. Therefore, early in training, the outputs of the deep network are relatively reliable and can be used for label correction. For " + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\bar{\\pmb{y}})" + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "content": ", we denote its set of predicted multiple labels as " + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{y}^*" + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "content": ". Here, the set of predicted labels is obtained with the top " + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "inline_equation", + "content": "|\\bar{\\pmb{y}}|" + }, + { + "bbox": [ + 304, + 486, + 547, + 617 + ], + "type": "text", + "content": " predictions based on the model's probability outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Recall that the holistic score of an example holistically measures the instance-label dependence and label dependence among multiple labels simultaneously. From both human and machine cognition, if an example is annotated accurately, both dependencies should be strong [70, 18, 68, 28, 7] with high probability. Namely, the holistic score is large. We propose to check the ratio between the holistic score on " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\bar{\\pmb{y}})" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and holistic score on " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y}^{*})" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1485" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 88, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 88, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 88, + 83 + ], + "type": "text", + "content": "we check" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 81, + 287, + 98 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 81, + 287, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 287, + 98 + ], + "type": "interline_equation", + "content": "\\kappa (h, \\boldsymbol {x}, \\bar {\\boldsymbol {y}}) = \\hat {\\tilde {S}} _ {\\bar {\\boldsymbol {y}}} (\\boldsymbol {x}) / \\hat {\\tilde {S}} _ {\\boldsymbol {y} ^ {*}} (\\boldsymbol {x}). \\tag {2}", + "image_path": "bc86216b285acaa5c7aaa9be5b83cc8ecd73a740efc3c00f1c75ee7957c2938b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": "We compare this ratio with a predetermined threshold " + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": ". The value of " + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": " is given in the next subsection. If " + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\kappa(h, \\boldsymbol{x}, \\bar{\\boldsymbol{y}}) \\leq \\hat{\\delta}" + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": ", we flip the labels " + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\bar{\\boldsymbol{y}}_{\\text{new}} = \\boldsymbol{y}^*" + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": ". Otherwise, the labels remain unchanged with " + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\bar{\\boldsymbol{y}}_{\\text{new}} = \\bar{\\boldsymbol{y}}" + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": ". The detailed algorithm of holistic correction for multi-label classification with noisy labels (aka HLC) is provided in Algorithm 1. After holistic correction for noisy labels, we use " + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{x}, \\bar{\\boldsymbol{y}}_{\\text{new}})" + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": " to train the deep network " + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 102, + 287, + 198 + ], + "type": "text", + "content": " based on ADDGCN [65]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 205, + 163, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 163, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 163, + 218 + ], + "type": "text", + "content": "3.2. Theoretical Insights" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": "We extend the Tsybakov condition [75, 2, 15, 49] from multi-class classification to multi-label classification. Specifically, denote by " + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{\\pmb{x}}" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": " the label set predicted based on " + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "inline_equation", + "content": "S^{f}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{\\pmb{x}} := h^{*}(\\pmb{x}) = \\arg \\max_{\\pmb{z}} S_{\\pmb{z}}^{f}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": ". Besides, denote by " + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\pmb{b}_{\\pmb{x}}" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": " the second best prediction with " + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "inline_equation", + "content": "\\pmb{b}_{\\pmb{x}} := \\arg \\max_{\\pmb{z} \\neq \\pmb{a}_{\\pmb{x}}} S_{\\pmb{z}}^{f}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": ". The maximum length of a label set is denoted as " + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "inline_equation", + "content": "m \\ll q" + }, + { + "bbox": [ + 46, + 224, + 287, + 355 + ], + "type": "text", + "content": "). In this paper, we call the predicted label set by the Bayes optimal classifier for an instance as the correct label set. We present the Tsybakov condition on instance-label (abbreviated as ins.-label here) dependence and holistic Tsybakov condition as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "text", + "content": "Definition 1 (Tsybakov condition on ins.-label dependence) " + }, + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "inline_equation", + "content": "\\exists C_1, \\lambda_1 > 0" + }, + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "inline_equation", + "content": "\\exists t_0 \\in (0, m]" + }, + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "text", + "content": ", such that for all " + }, + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "inline_equation", + "content": "t \\leq t_0" + }, + { + "bbox": [ + 46, + 366, + 298, + 400 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 400, + 287, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 400, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 93, + 400, + 287, + 417 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left[ S _ {\\boldsymbol {a} _ {\\boldsymbol {x}}} ^ {f} (\\boldsymbol {x}) - S _ {\\boldsymbol {b} _ {\\boldsymbol {x}}} ^ {f} (\\boldsymbol {x}) \\leq t \\right] \\leq C _ {1} t ^ {\\lambda_ {1}}. \\tag {3}", + "image_path": "b7668a4d0f18ea7690ac103b4ad423393d7f37b0e29919d343131f85d0749c0f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "text", + "content": "Definition 2 (Holistic Tsybakov condition) " + }, + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\exists C_2, \\lambda_2 > 0" + }, + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\exists t_0 \\in (0, m]" + }, + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "text", + "content": ", such that for all " + }, + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "inline_equation", + "content": "t \\leq t_0" + }, + { + "bbox": [ + 46, + 424, + 287, + 449 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 457, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 457, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 93, + 457, + 287, + 472 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left[ S _ {\\boldsymbol {a} _ {\\boldsymbol {x}}} (\\boldsymbol {x}) - S _ {\\boldsymbol {b} _ {\\boldsymbol {x}}} (\\boldsymbol {x}) \\leq t \\right] \\leq C _ {2} t ^ {\\lambda_ {2}}. \\tag {4}", + "image_path": "852d92948811385a71e1ae4105652ec41b108a46cb2ea6855250945ead458cc3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": "Remark 1 Definition 1 stipulates that the uncertainty of " + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "inline_equation", + "content": "S^f" + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": " is bounded. The margin region that is close to the decision boundary has a bounded volume. Definition 2 shares the similar idea and bound the uncertainty of " + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "text", + "content": "Theorem 1 Suppose " + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "inline_equation", + "content": "S(\\pmb{x})" + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "text", + "content": " fulfills the holis-tic Tsybakov condition for constants " + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "inline_equation", + "content": "C_2, \\lambda_2 > 0" + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "inline_equation", + "content": "t_0 \\in (0, m]" + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "inline_equation", + "content": "\\epsilon := \\max_{\\pmb{x}, \\pmb{z}} \\left[ |\\hat{\\bar{S}}_z^f(\\pmb{x}) - \\bar{S}_z^f(\\pmb{x})|, |\\hat{\\bar{S}}_z^l(\\pmb{x}) - \\bar{S}_z^l(\\pmb{x})|, |\\bar{S}_z^l(\\pmb{x}) - S_z^l(\\pmb{x})| \\right]" + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "inline_equation", + "content": "\\tau := \\min_i T_{ii}" + }, + { + "bbox": [ + 46, + 539, + 304, + 603 + ], + "type": "text", + "content": ". We analyze two cases:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": "(1) If " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{y}}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " is corrected by " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\kappa(h, \\pmb{x}, \\bar{\\pmb{y}})" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\delta_1 = \\min \\left[\\frac{\\tau S_{\\pmb{b}_\\pmb{x}}(\\pmb{x}) + \\sum_{l_j \\in \\bar{\\pmb{y}}} \\sum_{i \\neq j} T_{ij} \\mathbb{P}(l_i | \\pmb{x})}{\\hat{S}_{\\pmb{y}^*}(\\pmb{x})}\\right]" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\rho_1 := |\\hat{\\delta} - \\delta_1|" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": ". Assume that " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon \\leq \\frac{t_0 \\tau - \\rho_1 m}{3}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{P}[\\bar{\\pmb{y}}_{new} = h^*(\\pmb{x}), \\bar{\\pmb{y}}" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": " is flipped] is at least " + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "inline_equation", + "content": "1 - C_2 [O(\\max(\\epsilon, \\rho_1))]^{\\lambda_2} - \\mathbb{P}[\\pmb{a}_x \\neq \\{\\pmb{y}^*, \\bar{\\pmb{y}}\\}]" + }, + { + "bbox": [ + 46, + 603, + 287, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "text", + "content": "(2) If " + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{y}}" + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "text", + "content": " is not corrected by " + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "inline_equation", + "content": "\\kappa(h, \\pmb{x}, \\bar{\\pmb{y}})" + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "inline_equation", + "content": "\\delta_2 = \\max \\left[\\frac{\\hat{\\tilde{S}}_{\\bar{\\pmb{y}}}(\\pmb{x})}{\\tau S_{b_{\\pmb{x}}}(\\pmb{x}) + \\sum_{l_j \\in \\pmb{y}^*} \\sum_{i \\neq j} T_{ij} \\mathbb{P}(l_i | \\pmb{x})}\\right]" + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "inline_equation", + "content": "\\rho_2 := |\\hat{\\delta} - \\delta_2|" + }, + { + "bbox": [ + 46, + 678, + 287, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": "Assume that " + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\epsilon \\leq \\frac{t_0\\delta_2^2\\tau - \\rho_2m - \\rho_2^2m}{3\\delta_2^2}" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "\\mathbb{P}[\\bar{\\pmb{y}}_{new} = h^{*}(\\pmb{x}),\\bar{\\pmb{y}}" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": " is accepted] is at least " + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "inline_equation", + "content": "1 - C_2[O(\\max (\\epsilon ,\\rho_2))]^{\\lambda_2} - \\mathbb{P}[\\pmb {a}_x\\neq \\{\\pmb {y}^*,\\bar{\\pmb{y}}\\} ]" + }, + { + "bbox": [ + 305, + 70, + 545, + 114 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 122, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 545, + 266 + ], + "type": "text", + "content": "The proof of Theorem 1 is provided in Appendix B.1. Theorem 1 extends the theoretical results of [75] to multilabel classification with noisy labels. It claims that, even though with noisy multiple labels, the holistic correction has a guaranteed success rate to make proper corrections. Besides, if we can reasonably approximate the optimal " + }, + { + "bbox": [ + 304, + 122, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 122, + 545, + 266 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 122, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 122, + 545, + 266 + ], + "type": "text", + "content": ", our algorithm flips noisy multiple labels to correct ones with a good chance. Below, as a corollary of Theorem 1, we show that, there are certain circumstances, the use of holistic scores has a better chance to make corrections satisfactorily, than the sole use of instance-label dependence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": "Corollary 1 Suppose that " + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "inline_equation", + "content": "S(\\pmb{x})" + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": " fulfills the holistic Tsybakov condition. Denote the set threshold " + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": " and optimal threshold " + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "inline_equation", + "content": "\\rho := \\max |\\hat{\\delta} - \\delta|" + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": ". We have that, " + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "inline_equation", + "content": "\\exists \\epsilon" + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "inline_equation", + "content": "C_2[O(\\max(\\epsilon, \\rho))]^{\\lambda_2} < C_1[O(\\max(\\epsilon, \\rho))]^{\\lambda_1}" + }, + { + "bbox": [ + 304, + 276, + 545, + 349 + ], + "type": "text", + "content": ", holistic correction brings higher probability to handle noisy labels successfully than instance-label dependence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 358, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 466 + ], + "type": "text", + "content": "The proof of Corollary 1 is provided in Appendix B.2. Corollary 1 claims that there exist cases where holistic scores better combat noisy labels. Note that, from a theoretical view, we do not state that holistic scores can work better in all circumstances of multi-label classification. Nevertheless, with the determination of the threshold " + }, + { + "bbox": [ + 304, + 358, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 358, + 545, + 466 + ], + "type": "text", + "content": ", holistic scores can perform better in the experiments of this paper, which demonstrates the help of label dependence to handle noisy multiple labels." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 476, + 388, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 388, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 388, + 490 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 496, + 422, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 422, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 422, + 510 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 518, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 686 + ], + "type": "text", + "content": "Datasets. We verify the effectiveness of the proposed method on the synthetic noisy versions of three datasets, i.e., Pascal-VOC 2007 [12], Pascal-VOC 2012 [12], and MS-COCO [32]. Pascal-VOC 2007 contains 5,011 images in train and validation sets, while Pascal-VOC 2012 consists of 11,540 images in train and validation sets. The images come from 20 common object categories. For Pascal-VOC 2007 and Pascal-VOC 2012, we train methods using the noisy training and validation sets, and evaluate them on the test set of Pascal-VOC 2007 that has 4,952 images [14]. MS-COCO contains 82,081 training images and 40,137 validation images from 80 common object categories. As did in [74, 5, 65, 77], we evaluate the performance of methods using validation images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Noisy-label generation. The class-dependent noise transition matrix " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " [42, 21, 46, 72] is used to corrupt the three" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1486" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 79, + 500, + 354 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 541, + 79 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 541, + 79 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 541, + 79 + ], + "type": "text", + "content": "Table 1. Comparisons with advanced methods on noisy Pascal-VOC 2007. The mean and standard deviation of results (%) are presented." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 79, + 500, + 354 + ], + "lines": [ + { + "bbox": [ + 92, + 79, + 500, + 354 + ], + "spans": [ + { + "bbox": [ + 92, + 79, + 500, + 354 + ], + "type": "table", + "html": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE64.50±1.2058.65±2.1648.19±0.2371.77±1.1560.94±4.2548.72±2.13
CSRA66.99±0.4859.62±0.6146.97±0.4872.45±0.6963.58±1.4852.72±1.52
ADDGCN63.89±0.9455.75±1.9844.14±1.3771.02±0.9561.05±0.0650.18±2.70
APL66.79±1.1958.86±1.5347.64±1.8172.61±0.9961.99±0.7849.10±0.15
CDR67.35±1.7060.05±1.0649.12±0.5972.66±0.7964.58±0.6050.51±2.49
JOINT67.43±0.7363.37±0.9253.27±4.7070.28±1.8568.70±2.8858.57±2.75
WSIC65.43±0.5559.53±0.7348.34±0.4772.57±1.0361.88±2.5750.15±0.86
CCMN69.97±1.3662.58±1.4753.20±1.2870.68±1.0860.94±3.1248.62±1.26
HLC†72.07±0.6770.20±0.4668.00±0.8974.83±0.6469.86±1.6160.09±1.73
OF1 ↑BCE63.52±0.4856.70±2.4548.10±1.4368.28±0.6958.30±2.8251.18±3.10
CSRA65.40±0.4759.39±0.8148.32±1.5069.72±0.5061.89±0.4351.56±2.28
ADDGCN62.63±0.1855.50±1.8744.38±2.9268.95±0.6459.64±0.5653.12±0.62
APL64.85±1.4656.51±1.7047.54±2.4068.89±0.8958.04±0.9752.27±2.20
CDR65.31±0.9957.93±1.0548.86±1.7169.53±0.6559.89±1.0751.68±3.83
JOINT69.72±0.8867.93±0.7761.62±1.4071.24±1.0364.20±0.8860.30±1.24
WSIC63.45±0.9757.96±1.2548.38±2.4169.88±1.2257.97±2.1951.99±1.65
CCMN69.66±1.5560.43±1.3153.84±0.6967.12±0.6159.55±1.4553.46±1.04
HLC†71.03±0.3369.08±1.0068.62±0.4872.09±0.7465.76±2.3960.71±1.37
CF1 ↑BCE58.91±1.3453.21±2.0443.66±0.5365.93±0.8157.03±3.4347.21±1.89
CSRA62.31±0.5055.67±0.6143.11±0.7667.39±0.8059.66±1.0451.13±1.12
ADDGCN60.41±1.0453.72±1.3842.42±0.5966.05±0.9757.81±0.5848.89±2.64
APL60.23±1.5352.85±2.1842.38±1.6766.59±0.7158.33±0.4947.67±1.83
CDR61.37±1.4754.17±0.8643.60±0.8267.11±0.6359.91±0.3948.40±1.98
JOINT63.13±0.3860.22±1.6848.17±5.0166.03±1.2562.05±2.9854.03±3.17
WSIC59.54±1.1054.22±0.5343.82±0.6266.97±1.0058.04±1.7048.19±0.96
CCMN65.19±1.1058.55±1.3149.85±1.0665.47±0.9358.05±2.2448.46±0.80
HLC†68.87±0.1066.62±0.8164.82±0.4869.95±1.1965.13±1.0457.54±1.84
", + "image_path": "5b8c76122664480be695258f8ea9f2a7e080fd8268bb636019098ee8ad07ac15.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": "datasets. Here, for any " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "T_{ij} = \\mathbb{P}(l_j \\in \\bar{\\mathbf{y}} \\wedge l_i \\notin \\bar{\\mathbf{y}} | l_j \\notin \\mathbf{y} \\wedge l_i \\in \\mathbf{y})" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": " represents the probability of the " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": "-th class label to be flipped into the " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": "-th class label. We consider both symmetric (abbreviated as Sym.) and pairflip (abbreviated as Pair.) noise settings [17]. The details of the transition matrix are provided in Appendix D.2. For symmetric noise, the noise rate is set to " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": ". For pairflip noise, the noise rate is set to " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 373, + 287, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 478, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 670 + ], + "type": "text", + "content": "Baselines. We exploit three types of baselines in total. Specifically, Type-I baselines contain the methods that are designed for multi-label classification with clean labels. Type-II baselines consider the methods for multi-class classification with noisy labels. Type-III baselines consider the methods that focus on multi-label classification with noisy labels. It should be noted that, there are relatively few methods belonging to this type [36]. More advanced methods belonging to Type-III baselines need to be investigated [36], which is also our focus in this paper. In more detail, Type-I baselines include CSRA [77] and ADDGCN [65]. Type-II baselines include APL [40], CDR [58], and JOINT [48]. Type-III baselines include WSIC [22] and CCMN [63]. As a simple baseline, we compare our method with the standard deep network that directly trains on noisy datasets (abbreviated as BCE). We detail all baselines in Appendix D.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Network & Optimizer. We use a ResNet-50 network [20] pretrained on ImageNet as the backbone for all methods. We train the models for 30 epochs in total. We utilize" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 373, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 469 + ], + "type": "text", + "content": "Adam [27] for the network optimization. The batch size is set to 128 for all the datasets. The learning rate is fixed to " + }, + { + "bbox": [ + 304, + 373, + 545, + 469 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 373, + 545, + 469 + ], + "type": "text", + "content": ". The images in Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO resize to " + }, + { + "bbox": [ + 304, + 373, + 545, + 469 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 373, + 545, + 469 + ], + "type": "text", + "content": ". Note that, to make experiments more comprehensive, we also employ different experimental settings, e.g., different networks and different image sizes. The details are provided in Section 4.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 473, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 653 + ], + "type": "text", + "content": "Measurement. As did in multi-label classification [77, 7], evaluation metrics include the mean average precision (mAP) [71], the average F1-measure (OF1), and the average per-class F1-measure (CF1). For fair comparison, we implement all methods with default parameters by PyTorch, and conduct all experiments on NVIDIA GTX3090 GPUs. All experiments are repeated three times with different random seeds. Following the works in learning with noisy labels [17, 54, 29, 30], the mean and standard deviation of results in the last epoch are reported. In addition, for different evaluation metrics, we report the mean and standard deviation of best results. Supplementary results are shown in Appendix E. Afterwards, the best mean results are highlighted in red. The second best mean results are also highlighted in blue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 659, + 509, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 509, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 509, + 672 + ], + "type": "text", + "content": "4.2. Comparison with the State-of-the-Arts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 546, + 713 + ], + "type": "text", + "content": "The results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO are shown in Table 1, Table 2, and Table 3 respectively. In summary, HLC consistently works" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1487" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 80, + 500, + 354 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 541, + 79 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 541, + 79 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 541, + 79 + ], + "type": "text", + "content": "Table 2. Comparisons with advanced methods on noisy Pascal-VOC 2012. The mean and standard deviation of results (%) are presented." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 80, + 500, + 354 + ], + "lines": [ + { + "bbox": [ + 92, + 80, + 500, + 354 + ], + "spans": [ + { + "bbox": [ + 92, + 80, + 500, + 354 + ], + "type": "table", + "html": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE66.74±0.8056.07±0.5045.15±1.5670.91±1.1357.61±1.1449.85±0.36
CSRA66.35±0.5056.20±1.3545.54±1.1471.29±0.8360.71±1.1847.63±1.56
ADDGCN63.34±0.9654.54±0.8644.88±1.7170.41±0.5457.96±0.6847.66±1.08
APL67.07±1.0456.79±1.8643.51±1.9371.32±1.6059.59±1.2748.14±1.16
CDR66.13±1.4956.85±0.4844.84±1.1171.55±1.8760.13±1.8949.44±1.81
JOINT65.19±2.1758.40±2.8745.13±1.6968.93±2.5461.64±1.7853.64±1.61
WSIC65.96±0.7956.34±0.4144.80±0.5470.40±1.1159.40±1.8748.95±1.34
CCMN69.15±0.6661.00±1.0150.71±0.2669.08±1.7859.72±2.3246.67±2.78
HLC†72.14±0.6670.11±0.2768.69±1.0474.51±0.6769.90±0.4364.20±1.26
OF1 ↑BCE64.99±1.1056.92±2.0845.49±2.2368.48±2.2860.21±1.3554.05±1.95
CSRA64.08±0.3756.25±2.5748.67±3.1469.06±0.6559.75±1.7052.89±0.95
ADDGCN63.53±1.4154.28±0.8647.56±2.6747.62±2.3957.90±1.7852.33±0.56
APL64.70±1.1758.05±1.6845.74±1.5570.68±1.0360.22±1.5851.38±1.55
CDR64.06±1.3857.31±1.2146.51±0.9570.45±1.4460.57±1.2452.26±2.42
JOINT67.35±1.8664.57±2.3954.37±3.3370.81±1.4064.40±1.7656.27±1.29
WSIC62.74±2.1057.13±0.7345.52±1.2869.72±1.1959.11±2.0452.49±1.38
CCMN65.77±0.2359.91±0.9351.45±0.9467.93±1.7359.26±0.5148.61±4.71
HLC†71.14±0.6069.50±0.4067.80±0.3372.13±0.2667.59±0.9664.28±0.81
CF1 ↑BCE62.47±0.4453.26±0.4143.43±1.6766.03±1.6955.90±0.7049.29±0.64
CSRA62.08±0.7053.23±1.2743.23±1.2566.02±0.7457.71±1.0247.46±1.68
ADDGCN59.67±1.1452.61±0.5244.33±1.9965.22±0.8655.32±0.7647.30±1.12
APL62.99±1.0753.69±1.8041.72±1.4266.44±1.4057.52±0.9148.14±1.02
CDR62.18±1.0453.61±0.4542.83±0.8766.29±2.1257.23±1.4349.03±1.50
JOINT60.57±2.8254.39±3.7240.48±7.7066.30±2.3359.72±2.1255.06±0.36
WSIC61.70±0.9253.10±0.7442.72±0.5465.34±1.4857.21±1.6248.51±1.12
CCMN64.46±0.6257.45±0.9948.27±0.6867.48±1.4456.93±1.6947.01±1.82
HLC†69.54±0.5667.35±0.4865.72±1.4870.07±0.4165.68±0.9460.57±1.27
", + "image_path": "861655df74c45240a28c1723f94f1251d3b60b26ee722a3de574f3b5ea616970.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 92, + 372, + 500, + 645 + ], + "blocks": [ + { + "bbox": [ + 61, + 361, + 531, + 371 + ], + "lines": [ + { + "bbox": [ + 61, + 361, + 531, + 371 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 531, + 371 + ], + "type": "text", + "content": "Table 3. Comparisons with advanced methods on noisy MS-COCO. The mean and standard deviation of results (%) are presented." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 372, + 500, + 645 + ], + "lines": [ + { + "bbox": [ + 92, + 372, + 500, + 645 + ], + "spans": [ + { + "bbox": [ + 92, + 372, + 500, + 645 + ], + "type": "table", + "html": "
MetricsMethods / NoiseSym. 30%Sym. 40%Sym. 50%Pair. 20%Pair. 30%Pair. 40%
mAP ↑BCE53.23±0.1547.33±0.7940.25±0.2656.58±0.2249.16±0.0441.57±0.64
CSRA53.89±0.4047.64±0.8639.58±0.1958.27±0.2350.95±0.0743.07±0.64
ADDGCN51.08±0.9544.75±1.1538.66±1.3056.94±0.6150.28±0.8141.45±0.19
APL54.34±0.3248.61±0.7243.55±1.4357.73±0.2050.87±0.3441.77±0.50
CDR54.01±0.0449.01±0.2643.94±1.2557.03±0.2850.99±0.7742.71±0.09
JOINT53.93±0.4148.01±1.0445.27±0.6857.30±0.3351.94±0.2042.74±0.55
WSIC52.99±0.5346.84±0.8639.76±0.6456.66±0.3149.46±0.2542.52±0.62
CCMN51.73±0.1850.36±0.7145.32±0.8958.13±0.4451.17±0.2942.12±0.76
HLC†54.87±0.6851.09±0.5348.15±0.5058.55±0.0953.41±0.1345.91±0.39
OF1 ↑BCE51.34±1.7044.36±0.8234.85±1.2459.16±0.9552.44±0.8142.94±1.13
CSRA52.03±1.8641.63±1.4133.47±3.1859.17±0.1450.27±0.8841.75±1.36
ADDGCN55.67±1.4847.79±0.4035.95±3.7360.96±0.6555.05±1.7847.47±0.77
APL51.07±1.3243.93±2.7033.90±4.0060.04±1.1650.64±2.8644.34±1.99
CDR53.43±1.1645.10±0.8334.91±0.9059.34±0.6152.72±0.6344.17±0.61
JOINT54.56±0.0649.00±1.6637.78±0.9358.20±0.4053.21±0.1746.55±0.61
WSIC50.91±0.5242.93±0.8535.47±1.5258.89±1.1351.63±1.5743.99±1.47
CCMN52.71±1.0443.24±1.1934.62±1.3858.61±1.1852.18±0.7645.92±0.59
HLC†59.92±0.6557.84±0.3855.47±0.9562.28±0.0658.56±0.3751.09±0.60
CF1 ↑BCE45.92±0.2338.96±1.6131.34±0.2752.54±0.5845.54±0.6339.79±0.99
CSRA44.97±1.8837.49±1.7328.96±1.1652.18±0.4444.96±0.4336.88±0.21
ADDGCN46.77±1.8039.35±1.8330.57±1.5754.18±0.2347.55±0.1839.44±0.33
APL42.91±0.5438.38±0.7728.17±2.5052.87±1.0746.27±1.2737.76±1.02
CDR46.62±0.4239.47±0.5429.59±2.5252.51±0.6945.75±0.8139.15±0.53
JOINT49.51±0.8142.38±1.2124.24±0.6154.39±0.1749.90±0.8538.34±0.55
WSIC45.30±1.0939.15±1.6231.42±0.9452.04±0.2845.76±0.7039.44±1.11
CCMN44.20±1.1935.18±1.0127.90±1.2553.23±0.5846.88±0.9240.55±0.89
HLC†51.94±0.6349.24±0.3046.69±0.6655.44±0.1350.91±0.4843.35±0.82
", + "image_path": "697140414eba5919b482e3829a30c42e9c953af112581ccf822b39fb5b8b2458.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 661, + 287, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 661, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 661, + 287, + 709 + ], + "type": "text", + "content": "best across all noise settings. In many cases, the best results achieved by HLC outperform the second best results by a large margin, especially when the noise level is high. Below, we further discuss the results based on the comparisons" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 661, + 460, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 661, + 460, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 661, + 460, + 673 + ], + "type": "text", + "content": "with three different types of baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "Compared with Type-I baselines. We first notice that Type-I baselines are fragile to noisy labels in multi-label classification. Without considering the side-effect of noisy" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1488" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "labels, in many cases, they perform worse than BCE, which clearly illustrates the necessity for attention to handling noisy labels. Second, we compare HLC with ADDGCN. Without the proposed correction method for combating noisy labels, HLC will reduce to ADDGCN. As shown in the reported results, HLC performs much better than ADDGCN. To be specific, on noisy Pascal-VOC 2007, for Sym. " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ", HLC brings about " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "+15\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " performance improvement w.r.t. three evaluation metrics over ADDGCN. For Sym. " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ", the performance improvement is increased to more than " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "+20\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ". Also, for Pair. " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " and Pair. " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ", HLC enhances ADDGCN with about " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "+10\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " improvement. On noisy Pascal-VOC 2012 and MS-COCO, the performance improvement is also very clear." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 352 + ], + "type": "text", + "content": "Compared with Type-II baselines. On noisy Pascal-VOC 2007, with Sym. noise, we can see that HLC outperforms APL, CDR, and JOINT clearly, especially for Sym. " + }, + { + "bbox": [ + 46, + 243, + 288, + 352 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 243, + 288, + 352 + ], + "type": "text", + "content": ". Additionally, with Pair. noise, although the improvement is less than the cases with Sym. noise, HLC still performs best. On noisy Pascal-VOC 2012, for both Sym. and Pair. noise, the improvement is significant. Lastly, for noisy MS-COCO, HLC works better than all Type-II baselines with varying enhancement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 288, + 437 + ], + "type": "text", + "content": "Note that, compared with APL and CDR, JOINT seems to be a stronger baseline. Benefiting from label correction, after a few training epochs, JOINT less overfits to wrong labels, following better performance. Nevertheless, the proposed label-correction paradigm is argued to be more advanced. As shown in all results, HLC surpasses JOINT, which verifies the effectiveness of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 441, + 288, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 525 + ], + "type": "text", + "content": "Compared with Type-III baselines. On noisy Pascal-VOC 2007 and noisy Pascal-VOC 2012, HLC outperforms WSIC and CCMN distinctly. For example, with Sym. " + }, + { + "bbox": [ + 46, + 441, + 288, + 525 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 441, + 288, + 525 + ], + "type": "text", + "content": " noise, more than " + }, + { + "bbox": [ + 46, + 441, + 288, + 525 + ], + "type": "inline_equation", + "content": "+10\\%" + }, + { + "bbox": [ + 46, + 441, + 288, + 525 + ], + "type": "text", + "content": " performance promotion is brought by our method. On noisy MS-COCO, although WSIC and CCMN are sometimes competitive w.r.t. mAP, they are inferior w.r.t. both OF1 and CF1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 534, + 224, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 224, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 224, + 547 + ], + "type": "text", + "content": "4.3. More Analyses and Justifications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "type": "text", + "content": "In this subsection, we conduct performance analysis in more detail. The experiments are conducted with Sym. " + }, + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 553, + 287, + 601 + ], + "type": "text", + "content": " noise, which is more challenging than the experiments in low-noise-rate cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Role of label dependence. We study the effect of removing the consideration of label dependence to provide insights into what makes HLC successful. The experiments are conducted on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". Recall that HLC considers instance-label and label dependences simultaneously. When we remove the consideration of the label dependence in HLC, the correspond" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 102, + 527, + 225 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 547, + 102 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 547, + 102 + ], + "type": "text", + "content": "Table 4. Ablation study results on noisy Pascal-VOC 2007, Pascal-VOC 2012, and MS-COCO. The mean and standard deviation of results are presented. The best result in each case is in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 102, + 527, + 225 + ], + "lines": [ + { + "bbox": [ + 323, + 102, + 527, + 225 + ], + "spans": [ + { + "bbox": [ + 323, + 102, + 527, + 225 + ], + "type": "table", + "html": "
DatasetNoisy Pascal-VOC 2007
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.06±0.4167.23±1.9263.42±0.58
HLC68.00±0.8968.62±0.4864.82±0.48
DatasetNoisy Pascal-VOC 2012
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.67.88±0.7566.30±1.2864.33±1.67
HLC68.69±1.0467.80±0.3365.72±1.48
DatasetNoisy MS-COCO
MethodsmAP ↑OF1 ↑CF1 ↑
HLC w/o l.46.21±0.3652.90±0.9244.51±1.29
HLC48.15±0.5055.47±0.9546.69±0.66
", + "image_path": "a08a55ece4d032c52d71ab6150c92b4f8a475f8d6b46af72ba8b2d8aff598e9c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 345, + 240, + 504, + 329 + ], + "blocks": [ + { + "bbox": [ + 345, + 240, + 504, + 329 + ], + "lines": [ + { + "bbox": [ + 345, + 240, + 504, + 329 + ], + "spans": [ + { + "bbox": [ + 345, + 240, + 504, + 329 + ], + "type": "image", + "image_path": "a3dc4cc0c589ddd7719ba50a4ffd2c9521624dae137206fe861a7090f5299cc2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 330, + 504, + 418 + ], + "blocks": [ + { + "bbox": [ + 345, + 330, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 345, + 330, + 504, + 418 + ], + "spans": [ + { + "bbox": [ + 345, + 330, + 504, + 418 + ], + "type": "image", + "image_path": "514fc574be132d5f5c5d4bca61e6f2621079a405bfa991e66ed9f4d718c2b0e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 418, + 547, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 547, + 452 + ], + "type": "text", + "content": "Figure 2. Ablation study results with different values of the set threshold " + }, + { + "bbox": [ + 305, + 418, + 547, + 452 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 305, + 418, + 547, + 452 + ], + "type": "text", + "content": ". The experiments are conducted on noisy Pascal-VOC 2007 (Top) and noisy Pascal-VOC 2012 (Bottom)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "text", + "content": "ing method is named as HLC w/o l. here. For both HLC w/o l. and HLC, the value of the threshold " + }, + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "text", + "content": " is searched in the range " + }, + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\{0.25, 0.30, 0.35, 0.40, 0.45\\}" + }, + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "text", + "content": ". We use the " + }, + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 464, + 547, + 584 + ], + "type": "text", + "content": " noisy training data as a validation set for the threshold determination and performance report. The results are shown in Table 4. As can be seen, HLC outperforms HLC w/o l.. The results justify our claims that the label dependence could help combat the noisy labels in multi-label classification, which demonstrate the effectiveness of the proposed holistic correction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "content": "Analysis of the threshold " + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "content": ". We analyze the influence of different values of " + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "content": ". The experiments are conducted on noisy Pascal-VOC 2007 and Pascal-VOC 2012. The ResNet-50 network pretrained on ImageNet is used as the backbone. The image size is set to " + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "content": ". The value of the threshold " + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "content": " is chosen in " + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "inline_equation", + "content": "\\{0.25, 0.30, 0.35, 0.40, 0.45\\}" + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "content": ". Figure 2 shows that HLC is robust to the determination of the threshold " + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "inline_equation", + "content": "\\hat{\\delta}" + }, + { + "bbox": [ + 304, + 588, + 547, + 697 + ], + "type": "text", + "content": " in the certain range, which facilitates the practical application of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "content": "Evaluations with different networks. We use pretrained" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1489" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 102, + 260, + 376 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 102 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 102 + ], + "type": "text", + "content": "Table 5. Comparisons with advanced methods on noisy MS COCO with different networks. The mean and standard deviation of results " + }, + { + "bbox": [ + 47, + 69, + 287, + 102 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 69, + 287, + 102 + ], + "type": "text", + "content": " are presented." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 102, + 260, + 376 + ], + "lines": [ + { + "bbox": [ + 72, + 102, + 260, + 376 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 260, + 376 + ], + "type": "table", + "html": "
MetricsMethodsResNet-34ResNet-101
mAP ↑BCE42.63±0.7438.17±0.41
CSRA41.35±0.1837.24±1.20
ADDGCN40.15±0.9836.13±0.69
APL44.82±0.7040.90±1.51
CDR45.43±0.6541.00±0.38
JOINT44.81±0.7739.96±1.30
WSIC41.86±0.6237.49±0.68
CCMN45.31±0.4746.01±1.01
HLC†46.05±0.8146.24±2.13
OF1 ↑BCE37.65±2.4638.65±2.50
CSRA35.05±0.7834.28±2.40
ADDGCN35.11±1.2637.18±0.50
APL34.31±1.7337.89±1.30
CDR36.67±3.4339.88±1.15
JOINT39.66±1.1341.36±0.88
WSIC35.08±1.7438.44±0.57
CCMN32.86±1.5037.03±1.48
HLC†44.11±0.8047.79±4.19
CF1 ↑BCE28.95±2.0934.11±1.13
CSRA27.40±0.4431.21±1.57
ADDGCN26.11±0.8630.41±0.72
APL26.64±2.3931.97±1.95
CDR27.13±2.3935.17±1.06
JOINT30.77±1.6337.63±0.81
WSIC27.62±0.6533.83±0.61
CCMN24.75±0.4826.65±0.26
HLC†34.79±1.4140.88±2.42
", + "image_path": "15b07f682159307d4b6ee3ff4467ee549233a1c227b64a6bf4517f2e7e6997d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": "ResNet-50 before. To show that our method is robust to the choice of network structures, we use different networks in experiments. Specifically, we employ pretrained ResNet-34 [20] and pretrained ResNet-101 [20] respectively. The noisy MS-COCO is considered. The image size is " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": ". The results on mAP are reported in Table 5. As can be seen, with different networks, HLC still works well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "content": "Evaluations with different image sizes. We resize the image size to " + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "content": " before. To test the performance of advanced methods with different image sizes, we further consider " + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 46, + 483, + 288, + 579 + ], + "type": "text", + "content": " image sizes. Pretrained ResNet-50 is used. The results are reported in Table 6. For mAP, we can see that HLC is competitive compared with CCMN and CSRA. For OF1 and CF1, HLC works better than all baselines with a clear margin." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 587, + 254, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 254, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 254, + 600 + ], + "type": "text", + "content": "4.4. Experiments on the Real-world Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "To demonstrate that our problem setting can be adapted to the real world and our method can well handle practical scenes, we employ the real-world dataset NUS-WIDE [10] that originally contained 269,648 images from Flicker, which have been manually annotated with 81 visual concepts. Since some urls for download have been deleted, we employ the dataset version in [45]. A standard 70-30 train-test split is used. The backbone is chosen as ResNet-101. As the computation cost of training on NUS-WIDE is rela" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 102, + 547, + 374 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 102 + ], + "type": "text", + "content": "Table 6. Comparisons with advanced methods on noisy MS COCO. The mean and standard deviation of results " + }, + { + "bbox": [ + 305, + 69, + 545, + 102 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 69, + 545, + 102 + ], + "type": "text", + "content": " are presented. Difference image sizes are considered here." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 102, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 307, + 102, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 102, + 547, + 374 + ], + "type": "table", + "html": "
MetricsImage sizes112 × 112384 × 384448 × 448
mAP ↑BCE32.22±0.6939.40±1.3635.24±1.73
CSRA29.55±0.1643.55±0.7044.56±0.75
ADDGCN32.34±0.4638.72±1.6434.87±1.89
APL34.41±0.4843.65±0.2841.44±1.21
CDR34.75±0.3943.26±0.7239.97±1.40
JOINT32.89±0.1642.95±0.8840.17±1.26
WSIC31.98±0.2339.57±1.0236.08±0.23
CCMN36.17±0.4144.39±0.3944.03±0.17
HLC†35.98±1.0545.12±0.1344.23±1.20
OF1 ↑BCE26.70±0.8834.71±2.7626.72±3.35
CSRA20.14±1.2836.41±0.7138.54±0.78
ADDGCN26.36±2.2942.83±2.0540.73±1.04
APL24.02±1.2234.68±1.4630.73±2.64
CDR26.50±1.2331.31±1.7631.15±3.46
JOINT34.11±0.9538.67±1.2538.11±0.69
WSIC24.61±1.1034.09±2.9430.69±0.97
CCMN23.89±1.4936.16±2.1225.03±2.48
HLC†39.05±2.6846.55±3.7745.14±2.34
CF1 ↑BCE19.61±0.6130.77±2.2824.94±3.31
CSRA13.34±1.2932.66±0.9832.68±0.22
ADDGCN18.67±1.4735.63±1.3733.89±2.41
APL17.21±0.7829.24±0.0925.01±1.31
CDR18.04±1.0729.62±1.1926.28±1.23
JOINT20.76±0.7534.90±1.8833.75±1.31
WSIC19.07±0.5731.63±2.4928.50±1.36
CCMN16.75±1.2130.09±2.3226.68±1.57
HLC†29.70±2.0740.34±2.1437.48±2.36
", + "image_path": "c39cf49eecf790512ab2e512c1187d0e9bab806e265c764fd711603f7cbe1e71.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 343, + 401, + 506, + 479 + ], + "blocks": [ + { + "bbox": [ + 305, + 380, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 401 + ], + "type": "text", + "content": "Table 7. Comparison of our method to known state-of-the-art models on the NUS-WIDE dataset. Metrics are in %." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 401, + 506, + 479 + ], + "lines": [ + { + "bbox": [ + 343, + 401, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 343, + 401, + 506, + 479 + ], + "type": "table", + "html": "
MethodmAP ↑OF1 ↑CF1 ↑
S-CLs [39]60.173.758.7
MS-CMA [66]61.473.860.5
SRN [76]62.073.458.5
ICME [7]62.874.160.7
ASL [45]63.974.662.7
HLC†63.174.662.9
HLC+ASL†64.575.163.4
", + "image_path": "0c805c867923ebf9002a5329848b886e12165cb85d529a4ff3bc2f0ed68d619a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 487, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 594 + ], + "type": "text", + "content": "tively large, we run experiments one time. Here we compare our method with S-CLs [39], MS-CMA [66], SRN [76], ICME [7], and ASL [45]. For convenient comparison, we refer to the results of their original papers. Note that to further improve the performance on NUS-WIDE, we utilize ASL to replace the loss function of our method. We name the new method \"HLC+ASL\". Results are provided in Table 7, which demonstrate the effectiveness of our method on the real-world dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 609, + 378, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 378, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 378, + 620 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "In this paper, we focus on the realistic problem of multi-label classification with noisy labels. We learn and utilize the label dependence among multiple labels to handle this problem. With the help of label dependence, a novel algorithm named HLC is proposed to correct noisy multiple labels to clean ones. We demonstrate the effectiveness of our algorithm both theoretically and empirically. For future" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1490" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": "work, we are interested in adapting HLC to other domains such as natural language processing and recommendation systems. We are also interested in promoting our algorithm to tackle instance-dependent label noise [73, 4, 78, 38] in multi-label classification." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 149, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 149, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 149, + 156 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 288, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 288, + 236 + ], + "type": "text", + "content": "Bo Han was supported by NSFC Young Scientists Fund No. 62006202 and Guangdong Basic and Applied Basic Research Foundation No. 2022A1515011652. Tongliang Liu is partially supported by ARC projects: FT220100318, DP220102121, LP220100527, LP220200949, and IC190100031." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 245, + 107, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 107, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 107, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 266, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 266, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 287, + 320 + ], + "type": "text", + "content": "[1] Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, pages 233-242, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 322, + 288, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 322, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 53, + 322, + 288, + 343 + ], + "type": "text", + "content": "[2] Dara Bahri, Heinrich Jiang, and Maya Gupta. Deep k-nn for noisy labels. In ICML, pages 540–550, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 345, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 345, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 345, + 287, + 376 + ], + "type": "text", + "content": "[3] Yingbin Bai and Tongliang Liu. Me-momentum: Extracting hard confident examples from noisily labeled data. In ICCV, pages 9312-9321, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 379, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 379, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 379, + 287, + 422 + ], + "type": "text", + "content": "[4] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. arXiv preprint arXiv:2001.03772, 2020. 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 423, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 287, + 465 + ], + "type": "text", + "content": "[5] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In ICCV, pages 522-531, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 512 + ], + "type": "text", + "content": "[6] Zhaomin Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Learning graph convolutional networks for multi-label recognition and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 514, + 286, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 286, + 556 + ], + "type": "text", + "content": "[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME, pages 622-627, 2019. 1, 3, 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 559, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 286, + 601 + ], + "type": "text", + "content": "[8] Xiang Cheng, Shu-Guang Zhao, Xuan Xiao, and Kuo-Chen Chou. iatc-misf: a multi-label classifier for predicting the classes of anatomical therapeutic chemicals. Bioinformatics, 33(3):341-346, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 603, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 286, + 657 + ], + "type": "text", + "content": "[9] Tejas Chheda, Purujit Goyal, Trang Tran, Dhruvesh Patel, Michael Boratko, Shib Sankar Dasgupta, and Andrew McCallum. Box embeddings: An open-source library for representation learning using geometric structures. arXiv preprint arXiv:2109.04997, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "text", + "content": "[10] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In ACM International Conference on Image and Video Retrieval, pages 1-9,2009.8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[11] Zijun Cui, Yong Zhang, and Qiang Ji. Label error correction and generation through label relationships. In AAAI, pages 3693-3700, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[12] Mark Everingham, Andrew Zisserman, Christopher KI Williams, Luc Van Gool, Moray Allan, Christopher M Bishop, Olivier Chapelle, Navneet Dalal, Thomas Deselaers, Gyuri Dorkó, et al. The pascal visual object classes challenge 2007 (voc2007) results. 2008. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 165, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 165, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 545, + 198 + ], + "type": "text", + "content": "[13] Lei Feng, Senlin Shu, Zhuoyi Lin, Fengmao Lv, Li Li, and Bo An. Can cross entropy loss be robust to label noise? In IJCAI, pages 2206-2212, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 200, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 200, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 308, + 200, + 545, + 242 + ], + "type": "text", + "content": "[14] Bin-Bin Gao and Hong-Yu Zhou. Learning to discover multi-class attentional regions for multi-label image recognition. IEEE Transactions on Image Processing, 30:5920-5932, 2021. 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 277 + ], + "type": "text", + "content": "[15] Wei Gao, Bin-Bin Yang, and Zhi-Hua Zhou. On the resistance of nearest neighbor to random noisy labels. arXiv preprint arXiv:1607.07526, 2016. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 280, + 545, + 323 + ], + "type": "text", + "content": "[16] Bo Han, Gang Niu, Xingrui Yu, Quanming Yao, Miao Xu, Ivor Tsang, and Masashi Sugiyama. *Sigua: Forgetting may make learning with noisy labels more robust.* In ICML, pages 4006–4016, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 326, + 545, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 378 + ], + "type": "text", + "content": "[17] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, pages 8527-8537, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 426 + ], + "type": "text", + "content": "[18] Jun-Yi Hang and Min-Ling Zhang. Collaborative learning of label semantics and deep label-specific features for multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 428, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 460 + ], + "type": "text", + "content": "[19] Jun-Yi Hang and Min-Ling Zhang. Dual perspective of label-specific feature learning for multi-label classification. In ICML, pages 8375-8386, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 463, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 463, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 463, + 545, + 495 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 498, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 530 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Mantas Mazeika, Duncan Wilson, and Kevin Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In NeurIPS, 2018. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 533, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 565 + ], + "type": "text", + "content": "[22] Mengying Hu, Hu Han, Shiguang Shan, and Xilin Chen. Weakly supervised image classification through noise regularization. In CVPR, pages 11517-11525, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "text", + "content": "[23] Zhuo Huang, Xiaobo Xia, Li Shen, Bo Han, Mingming Gong, Chen Gong, and Tongliang Liu. Harnessing out-of-distribution examples via augmenting content and style. In ICLR, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 666 + ], + "type": "text", + "content": "[24] Zhuo Huang, Miaoxi Zhu, Xiaobo Xia, Li Shen, Jun Yu, Chen Gong, Bo Han, Bo Du, and Tongliang Liu. Robust generalization against photon-limited corruptions via worst-case sharpness minimization. In CVPR, pages 16175-16185, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. MentorNet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, pages 2309-2318, 2018. 2, 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1491" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[26] Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. Nlnl: Negative learning for noisy labels. In ICCV, pages 101-110, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 139 + ], + "type": "text", + "content": "[27] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 174 + ], + "type": "text", + "content": "[28] Cheng Li, Bingyu Wang, Virgil Pavlu, and Javed Aslam. Conditional bernoulli mixtures for multi-label classification. In ICML, pages 2482-2491, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 207 + ], + "type": "text", + "content": "[29] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In ICLR, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 241 + ], + "type": "text", + "content": "[30] Junnan Li, Caiming Xiong, and Steven CH Hoi. Learning from noisy data with robust representation learning. In ICCV, pages 9485-9494, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 287, + 285 + ], + "type": "text", + "content": "[31] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 287, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 287, + 330 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 332, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 287, + 364 + ], + "type": "text", + "content": "[33] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 408 + ], + "type": "text", + "content": "[34] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "text", + "content": "[35] Weiwei Liu, Ivor W Tsang, and Klaus-Robert Müller. An easy-to-hard learning paradigm for multiple classes and multiple labels. Journal of Machine Learning Research, 18, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 456, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 497 + ], + "type": "text", + "content": "[36] Weiwei Liu, Haobo Wang, Xiaobo Shen, and Ivor Tsang. The emerging trends of multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 501, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 542 + ], + "type": "text", + "content": "[37] Weiwei Liu, Donna Xu, Ivor W Tsang, and Wenjie Zhang. Metric learning for multi-output tasks. Transactions on Pattern Analysis and Machine Intelligence, 41(2):408-422, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 545, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 287, + 567 + ], + "type": "text", + "content": "[38] Yang Liu. Identifiability of label noise transition matrix. arXiv preprint arXiv:2202.02016, 2022. 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 611 + ], + "type": "text", + "content": "[39] Yongcheng Liu, Lu Sheng, Jing Shao, Junjie Yan, Shiming Xiang, and Chunhong Pan. Multi-label image classification via knowledge distillation from weakly-supervised detection. In ACMMM, pages 700-708, 2018. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 613, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 655 + ], + "type": "text", + "content": "[40] Xingjun Ma, Hanxun Huang, Yisen Wang, Simone Romano, Sarah Erfani, and James Bailey. Normalized loss functions for deep learning with noisy labels. In ICML, pages 6543-6553, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "type": "text", + "content": "[41] Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. Modeling fine-grained entity types with box embeddings. In ACL, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[42] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[43] Cosmin Octavian Pene, Amirmasoud Ghiassi, Taraneh Younesian, Robert Birke, and Lydia Y Chen. Multi-label gold asymmetric loss correction with single-label regulators. arXiv preprint arXiv:2108.02032, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "text", + "content": "[44] Geoff Pleiss, Tianyi Zhang, Ethan R Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "text", + "content": "[45] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In ICCV, 2021. 1, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "text", + "content": "[46] Jun Shu, Qian Zhao, Zengben Xu, and Deyu Meng. Meta transition adaptation for robust deep learning with noisy labels. arXiv preprint arXiv:2006.05697, 2020. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[47] Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. Learning from noisy labels with deep neural networks: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "text", + "content": "[48] Daiki Tanaka, Daiki Ikami, Toshihiko Yamasaki, and Kiyoharu Aizawa. Joint optimization framework for learning with noisy labels. In CVPR, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "text", + "content": "[49] Alexander B Tsybakov. Optimal aggregation of classifiers in statistical learning. The Annals of Statistics, 32(1):135-166, 2004. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "text", + "content": "[50] Arash Vahdat. Toward robustness against label noise in training deep discriminative neural networks. In NeurIPS, pages 5596-5605, 2017. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "text", + "content": "[51] Andreas Veit, Neil Alldrin, Gal Chechik, Ivan Krasin, Abhinav Gupta, and Serge Belongie. Learning from noisy large-scale datasets with minimal supervision. In CVPR, 2017. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[52] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, pages 2285-2294, 2016. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "text", + "content": "[53] Xinshao Wang, Yang Hua, Elyor Kodirov, David A Clifton, and Neil M Robertson. Proselflc: Progressive self label correction for training robust deep neural networks. In CVPR, pages 752-761, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 516, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 516, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 545, + 557 + ], + "type": "text", + "content": "[54] Yisen Wang, Weiyang Liu, Xingjun Ma, James Bailey, Hongyuan Zha, Le Song, and Shu-Tao Xia. Iterative learning with open-set noisy labels. In CVPR, pages 8688-8696, 2018. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "text", + "content": "[55] Hongxin Wei, Lei Feng, Xiangyu Chen, and Bo An. Combating noisy labels by agreement: A joint training method with co-regularization. In CVPR, pages 13726–13735, 2020. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "text", + "content": "[56] Hongxin Wei, Lue Tao, Renchunzi Xie, and Bo An. Open-set label noise can improve robustness against inherent label noise. In NeurIPS, pages 7978-7992, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "text", + "content": "[57] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed graph. In ICCV, pages 4157-4165, 2015. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[58] Xiaobo Xia, Tongliang Liu, Bo Han, Chen Gong, Nannan Wang, Zongyuan Ge, and Yi Chang. Robust early-learning: Hinding the memorization of noisy labels. In ICLR, 2021. 1, 5" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1492" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In ICLR, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[60] Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. Part-dependent label noise: Towards instance-dependent label noise. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[61] Xiaobo Xia, Tongliang Liu, Nannan Wang, Bo Han, Chen Gong, Gang Niu, and Masashi Sugiyama. Are anchor points really indispensable in label-noise learning? In NeurIPS, pages 6835-6846, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 229 + ], + "type": "text", + "content": "[62] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning. In AAAI, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "text", + "content": "[63] Ming-Kun Xie and Sheng-Jun Huang. Ccmn: A general framework for learning with class-conditional multi-label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "text", + "content": "[64] Ming-Kun Xie, Jia-Hao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "text", + "content": "[65] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, pages 649-665, 2020. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[66] Renchun You, Zhiyao Guo, Lei Cui, Xiang Long, Yingze Bao, and Shilei Wen. Cross-modality attention with semantic graph embedding for multi-label classification. In AAAI, pages 12709-12716, 2020. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 431 + ], + "type": "text", + "content": "[67] Hsiang-Fu Yu, Prateek Jain, Purushottam Kar, and Inderjit Dhillon. Large-scale multi-label learning with missing labels. In ICML, pages 593-601, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 475 + ], + "type": "text", + "content": "[68] Ze-Bang Yu and Min-Ling Zhang. Multi-label classification with label-specific feature generation: A wrapped approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "text", + "content": "[69] Jiong Zhang, Wei-cheng Chang, Hsiang-fu Yu, and Inderjit Dhillon. Fast multi-resolution transformer fine-tuning for extreme multi-label text classification. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "text", + "content": "[70] Min-Ling Zhang and Kun Zhang. Multi-label learning by exploiting label dependency. In KDD, pages 999-1008, 2010. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 578 + ], + "type": "text", + "content": "[71] Min-Ling Zhang and Zhi-Hua Zhou. A review on multi-label learning algorithms. IEEE Transactions on Knowledge and Data Engineering, 26(8):1819-1837, 2013. 1, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "text", + "content": "[72] Yivan Zhang, Gang Niu, and Masashi Sugiyama. Learning noise transition matrix from only noisy labels via total variation regularization. In ICML, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[73] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In ICLR, 2021. 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[74] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In ICCV, pages 163-172, 2021. 1, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[75] Songzhu Zheng, Pengxiang Wu, Aman Goswami, Mayank Goswami, Dimitris Metaxas, and Chao Chen. Error-bounded" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 208 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "correction of noisy labels. In ICML, pages 11447-11457, 2020. 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[76] Feng Zhu, Hongsheng Li, Wanli Ouyang, Nenghai Yu, and Xiaogang Wang. Learning spatial regularization with image-level supervisions for multi-label image classification. In CVPR, pages 5513-5522, 2017. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[77] Ke Zhu and Jianxin Wu. Residual attention: A simple but effective method for multi-label recognition. In ICCV, pages 184-193, 2021. 1, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "text", + "content": "[78] Zhaowei Zhu, Tongliang Liu, and Yang Liu. A second-order approach to learning with instance-dependent label noise. In CVPR, pages 10113-10123, 2021. 9" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1493" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_content_list.json b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9721285e0f5e3bd01600949ee5e389ca660c8be4 --- /dev/null +++ b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_content_list.json @@ -0,0 +1,1468 @@ +[ + { + "type": "text", + "text": "HollowNeRF: Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation", + "text_level": 1, + "bbox": [ + 250, + 130, + 722, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiufeng Xie, Riccardo Gherardi, Zhihong Pan, Stephen Huang", + "bbox": [ + 236, + 203, + 733, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Oppo Mobile Telecommunications Corp.", + "bbox": [ + 321, + 222, + 645, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2479 E Bayshore Rd, Palo Alto, CA, 94303, USA", + "bbox": [ + 287, + 239, + 681, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xxie28@gmail.com", + "bbox": [ + 411, + 260, + 557, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural radiance fields (NeRF) have garnered significant attention, with recent works such as Instant-NGP accelerating NeRF training and evaluation through a combination of hashgrid-based positional encoding and neural networks. However, effectively leveraging the spatial sparsity of 3D scenes remains a challenge. To cull away unnecessary regions of the feature grid, existing solutions rely on prior knowledge of object shape or periodically estimate object shape during training by repeated model evaluations, which are costly and wasteful. To address this issue, we propose HollowNeRF, a novel compression solution for hashgrid-based NeRF which automatically sparsifies the feature grid during the training phase. Instead of directly compressing dense features, HollowNeRF trains a coarse 3D saliency mask that guides efficient feature pruning, and employs an alternating direction method of multipliers (ADMM) pruner to sparsify the 3D saliency mask during training. By exploiting the sparsity in the 3D scene to redistribute hash collisions, HollowNeRF improves rendering quality while using a fraction of the parameters of comparable state-of-the-art solutions, leading to a better cost-accuracy trade-off. Our method delivers comparable rendering quality to Instant-NGP, while utilizing just $31\\%$ of the parameters. In addition, our solution can achieve a PSNR accuracy gain of up to 1dB using only $56\\%$ of the parameters.", + "bbox": [ + 76, + 342, + 473, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF [1, 15]) have gained widespread recognition across academia and industry due to their remarkable capability to generate photorealistic novel views of 3D scenes from a collection of 2D images. Inspired by the volumetric representation, NeRF models the scene as a continuous 5D plenoptic function, enabling the creation of high-fidelity renderings with accurate lighting and shading effects. This technique has found versatile ap", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff438e54c8660f1313d917272bc057f83e8b0dbac6b448ddd4515490aab21663.jpg", + "image_caption": [ + "Figure 1: Cross-section views of rendered 3D scenes. The left column shows the hollow interior of HollowNeRF's rendering, while the middle column shows the solid interior of Instant-NGP's rendering. The right column displays HollowNeRF's rendering without slicing." + ], + "image_footnote": [], + "bbox": [ + 504, + 306, + 890, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plications in fields such as computer graphics, virtual and augmented reality, as well as robotics.", + "bbox": [ + 496, + 628, + 892, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training and evaluating NeRF models can be computationally expensive, and many recent works [20, 16, 4, 22, 18, 23] have focused on improving the efficiency of NeRF. Instant-NGP [16] is an established solution with state-of-the-art training speed. It employs a lightweight hashgrid for input encoding and a small multi-layer perceptron (MLP) to disambiguate hash collisions. However, the lightweight hashgrid unavoidably causes severe collisions at fine resolutions. These collisions are evenly scattered across the occupied voxels, resulting in a suboptimal accuracy. Another line of research focuses on accelerating NeRF rendering by only sampling near the surface of interest, but has to rely on prior knowledge of surface geometries (either from conventional algorithms such as shape-from-silhouette [13] or an MLP predicting the depth distribution along each camera ray [18]). However, a coarse surface estimation degrades", + "bbox": [ + 496, + 659, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3480", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NeRF rendering quality, while a precise surface estimation adds too much complexity, defeating the purpose of acceleration.", + "bbox": [ + 75, + 90, + 468, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose HollowNeRF, a novel NeRF compression method using trainable hash collision mitigation to improve rendering accuracy while consuming less parameters than existing NeRF methods. Built on a hash-based pipeline from Instant-NGP, HollowNeRF prioritizes the important features (of visible voxels) and prunes unnecessary features (of invisible voxels), leading to a redistribution of hash collision probability across the 3D volume. When two voxels direct to the same hash bucket, Instant-NGP shapes the shared feature in this bucket as a mixture of the desired features, which harms the accuracy. In contrast, HollowNeRF steers the shared feature to fit the more important voxel, and prunes the feature of the less important voxel toward 0, reducing interference to features sharing the same bucket. Specifically, when reading the feature of a certain voxel, HollowNeRF further scales the feature by a trainable saliency weight whose value captures the voxel's visibility. To reduce the cost, we divide the 3D space into coarse grid regions and assign a saliency weight to capture each region's visibility, forming a trainable 3D saliency grid. Unlike existing methods [13, 18] that require prior knowledge of the surface geometries, HollowNeRF learns to prioritize the important features by training the saliency grid, which converges to a \"hollow\" saliency distribution across the 3D volume. Figure 1 showcases this \"hollow\" rendering result.", + "bbox": [ + 75, + 136, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed design consists of three main components: a trainable 3D saliency grid to guide the compression of dense features ( $\\S 3.1$ ); a soft zero-skipping gate that enhances the MLP in the NeRF model to ensure a feature compressed to 0 translates to a zero density in the 3D space ( $\\S 3.2$ ); a pruner to further push unnecessary features to exact 0 instead of a small non-zero value by alternating direction method of multipliers or ADMM [3] ( $\\S 3.3$ ). Our experiments demonstrate that HollowNeRF achieves better accuracy (PSNR and LPIPS) than state-of-the-art methods while using significantly fewer parameters.", + "bbox": [ + 75, + 513, + 472, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key contributions of this work are:", + "bbox": [ + 96, + 680, + 356, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel NeRF compression solution, HollowNeRF, that learns to prioritize features defining the visible surface and prune invisible internal features without prior knowledge of surface geometries.", + "- We use an ADMM-based optimization framework to prune unnecessary features during NeRF training and enhance the MLP with a soft zero-skipping gate to ensure that pruned features correctly map to zero density.", + "- We evaluate the performance of HollowNeRF on popular NeRF datasets and our solution demonstrates a significantly superior balance between cost and accuracy than state-of-the-art solutions." + ], + "bbox": [ + 94, + 702, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and related work", + "text_level": 1, + "bbox": [ + 500, + 90, + 774, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Differentiable rendering [15, 7, 5, 22, 17] has emerged as a prominent alternative framework for novel view synthesis, alongside the conventional 3D rendering approaches. Following the seminal NeRF paper [15], the research community has matured the approach and enables it to handle arbitrarily large scenes [24, 29, 27, 2], complex reflections [8, 34, 21, 28, 14], and small training datasets with a limited number of views [10, 6, 19]. Recent works also substantially improve the efficiency of NeRF, reducing the run-time complexity at training [16, 22] and inference [16, 12, 18, 30, 32, 31, 9].", + "bbox": [ + 496, + 114, + 890, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comparatively fewer works have tackled the problem of maximizing visual fidelity for a given space complexity budget [16, 26], which is the focus of this paper. The original NeRF technique [15] encodes the entire 3D scene in the weights of an MLP that predicts opacity $(\\sigma)$ and color $(r,g,b)$ given a 3D position $(x,y,z)$ and direction $(\\theta ,\\phi)$ . While spatial complexity (438K parameters) is not the primary bottleneck, the substantial limitation of NeRF lies in the remarkably slow training and inference speeds. This is primarily attributed to the deep MLP with 8 hidden layers, each having a width of 256 neurons. It has since been shown [22] that differentiable rendering does not necessarily require neural networks. In [22], a scene is encoded as a neuron-free sparse voxel grid of opacities and RGB spherical harmonics. Without the MLP, training and inference run fast, at the cost of a significantly high spatial complexity.", + "bbox": [ + 496, + 281, + 892, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the continuum between the two extremes discussed above, there are techniques that harness the strengths of both ends and fare better in the balance between training speed and spatial complexity. Instant-NGP [16] is a representative example, which encodes features using a multi-resolution hashgrid and then feed the features into a lightweight MLP with only 2 hidden layers to decode the color and density; both the features in the hashgrid and the MLP are trained in conjunction. Our work extends the hash-based pipeline from Instant-NGP by achieving higher quality (PSNR & LPIPS) with less number of parameters. We achieve this goal by reclaiming the resources spent on empty, invisible, or internal regions. Instead of using an auxiliary MLP to predict space occupancy like DONeRF [18], we learn it through a trainable lightweight volumetric saliency grid.", + "bbox": [ + 496, + 523, + 892, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some existing literature has applied model compression techniques to NeRF: CC-NeRF [26] uses tensor decomposition to obtain a low-rank approximation of the learned network. Similarly, TensoRF [4] represents the volume as a 4D tensor and factorizes it into low-rank components. Unlike these compression works, HollowNeRF vets information based on its actual impact on the rendering accuracy, rather than raw entropy. The relative performance of these approaches are investigated in §4.4.", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3481", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ca2d7e5c5b0f5261f90d2957e8bf5d044998ade43110992e4a123fdd358d1c9.jpg", + "image_caption": [ + "Figure 2: Overview of the HollowNeRF workflow." + ], + "image_footnote": [], + "bbox": [ + 130, + 90, + 831, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. HollowNeRF design", + "text_level": 1, + "bbox": [ + 76, + 335, + 269, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents the system design of HollowNeRF, as outlined in Figure 2: (i) Given an input coordinate $\\boldsymbol{x} = (x,y,z)$ , we fetch the corresponding feature $\\boldsymbol{f}$ from the multi-level hashgrid, following the Instant-NGP method [16]. (ii) We predict the saliency weight $p$ of position $\\boldsymbol{x}$ by processing the information from a trainable 3D saliency grid $\\mathcal{G}$ through trilinear interpolation and a sigmoid function (§3.1), and use $p$ to scale the feature $\\boldsymbol{f}$ by $\\boldsymbol{v} = p\\boldsymbol{f}$ . Figure 3a showcases a trained saliency grid $\\mathcal{G}$ , demonstrating that training $\\mathcal{G}$ can mitigate hash collisions by suppressing the unnecessary features in empty or invisible regions. (iii) The weighted feature $\\boldsymbol{v}$ is fed into an MLP, which decodes $\\boldsymbol{v}$ to obtain the density $\\sigma$ and color at position $\\boldsymbol{x}$ . To ensure that $\\boldsymbol{v} = \\mathbf{0}$ maps to $\\sigma = 0$ , we introduce a zero-skipping gate (§3.2). (iv) The density and color outputs of the MLP are used for volume rendering, and the resulting image is compared to the ground truth to obtain a loss function $L$ . During training, an ADMM pruner (§3.3) sparsifies the saliency value distribution across the 3D space by enforcing a sparsity constraint when optimizing the loss $L$ . Figure 3a shows that a large portion of the unnecessary features are not entirely eliminated by training the saliency grid, with saliency weights diminished but not reaching 0. Therefore, we introduce the ADMM pruner to explicitly prune saliency weights of the empty or invisible regions to 0, as shown in Figure 3b.", + "bbox": [ + 75, + 363, + 473, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Trainable 3D saliency grid", + "text_level": 1, + "bbox": [ + 76, + 770, + 316, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In typical 3D scenes without large transparent objects, most regions of the space do not contribute to the final rendering as they are either empty or invisible from any view angle. To leverage this sparsity, methods such as Instant-NGP use a coarse-grained binary mask to capture space occupancy, which is periodically updated during training by evaluating the NeRF model at each voxel and checking if the density is", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/54dd887c1bdb73629434dc99e8034bb76592f0a2f4b4d4af68209341634de09a.jpg", + "image_caption": [ + "(a) Without ADMM pruner." + ], + "image_footnote": [], + "bbox": [ + 509, + 335, + 578, + 458 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/252d8b4c9f4d1fba89113877adbbe843346d1f634954c62beb788dd92b424405.jpg", + "image_caption": [ + "(b) With ADMM pruner." + ], + "image_footnote": [], + "bbox": [ + 589, + 337, + 689, + 459 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca227756aa46323dccd8640169d8ee11dc735ca070c41aaec0b287913d16a594.jpg", + "image_caption": [ + "Figure 3: Slices of a $64 \\times 64 \\times 64$ saliency grid trained on the \"Chair\" scene from the NeRF synthetic dataset[15]." + ], + "image_footnote": [], + "bbox": [ + 692, + 338, + 769, + 458 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/31f42bb1348737bb4fdbda2a83573ccdc7c685d8eb4279f1099b658e0b4869d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 338, + 883, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "above a threshold. This occupancy mask guides sampling in ray marching by skipping unoccupied regions. However, it still keeps unnecessary features in internal occluded regions that have non-zero densities but make no contributions to the visible surfaces from any view angle, as shown in Figure 1.", + "bbox": [ + 496, + 551, + 892, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this insight, we propose to reclaim the capacity spent on internal regions by learning the sparsity pattern through training, effectively making objects hollow. The benefits of this approach are twofold: it reduces space complexity by requiring fewer features, and improves accuracy by reducing hash collisions. However, precisely locating the invisible regions of objects can be challenging without any prior knowledge of the object's shape.", + "bbox": [ + 496, + 643, + 893, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To addresses this challenge, our approach utilizes the hashgrid-based positional encoding from Instant-NGP [16] and extends it by introducing a trainable saliency grid $\\mathcal{G}$ with a resolution of $T\\times T\\times T$ . Specifically, HollowNeRF trains a 3D tensor data structure $\\mathcal{G}$ alongside the NeRF model (feature hashgrid $\\mathcal{H}$ and MLP weights $\\mathcal{W}$ ), which gradually learns which regions to prioritize, such as a nontransparent object surface. We split the space coarsely as $T\\times T\\times T$ uniform grids and the 3D tensor $\\mathcal{G}$ stores train", + "bbox": [ + 496, + 763, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3482", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "able values representing the saliency of features near each grid voxel. Instead of letting multiple dense features in the same coarse grid share one value, we interpolate the values associated to the 8 voxels surrounding the input coordinate $\\pmb{x}$ to get a smoothed saliency distribution $\\mathcal{P}$ over the 3D space. The trainable saliency grid represents in itself an opportunity to compress the NeRF model using differentiable model compression techniques, as discussed later in §3.3.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We initialize $\\mathcal{G}$ as an all-ones tensor, where $T$ is typically much smaller than the resolution of the dense feature grid (In §4.2, we further investigate how different $T$ values affect the performance). Given a 3D coordinate $\\pmb{x} = (x,y,z)$ , we fetch its corresponding feature vector $\\pmb{f}$ from the multi-level hashgrid [16]. We then locate the 8 voxels $(\\pmb{x}_1,\\pmb{x}_2,\\dots,\\pmb{x}_8)$ surrounding $\\pmb{x}$ in the 3D saliency grid $\\mathcal{G}$ , and fetch its corresponding saliency value $g_i$ . Specifically, the coordinate $\\pmb{x}_i = (x_i,y_i,z_i)$ are quantized to the values $(\\hat{x}_i,\\hat{y}_i,\\hat{z}_i)$ between 0 and $T - 1$ , which are then used to index the saliency grid $\\mathcal{G}$ . This gives us the saliency value $g_i = \\mathcal{G}(\\hat{x}_i,\\hat{y}_i,\\hat{z}_i)$ . After obtaining $g_i$ for each of the 8 saliency grid voxels surrounding $\\pmb{x}$ , we calculate the saliency weight $p$ at coordinate $\\pmb{x}$ by trilinear interpolation followed by a sigmoid operator, so that $0 < p < 1$ . The final weighted feature vector $\\pmb{v}$ is obtained by multiplying the saliency weight $p$ with the feature vector $\\pmb{f}$ fetched from the hashgrid, i.e., $\\pmb{v} = p\\pmb{f}$ .", + "bbox": [ + 75, + 210, + 470, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For mathematical analysis, we define the saliency distribution $\\mathcal{P}$ as the collection of saliency weights $p$ at every possible coordinate $\\boldsymbol{x}$ , such that $p = \\mathcal{P}(\\boldsymbol{x})$ . Our proposed method, HollowNeRF, learns the saliency distribution $\\mathcal{P}$ by training the saliency grid $\\mathcal{G}$ using gradient descent. It is worth noting that we inject $\\mathcal{P}$ (and thus $\\mathcal{G}$ ) to the training pipeline by feeding the weighted feature $\\boldsymbol{v}$ instead of the raw feature $\\boldsymbol{f}$ to the MLP for decoding; the rationale for this choice is that since the sigmoid function in the above workflow projects the saliency weights to the range $(0,1)$ , we can view the 3D saliency distribution as a probability distribution $\\mathcal{P}(\\boldsymbol{x})$ indicating the likelihood $p$ that a feature is non-zero at a given 3D position $\\boldsymbol{x}$ . There are only two possibilities for the feature at a given position: either the position is visible with probability $p = \\mathcal{P}(\\boldsymbol{x})$ , and we retrieve the feature vector $\\boldsymbol{f}$ from the multi-level hashgrid as in [16]; otherwise the position is empty or occluded with probability $1 - p$ , and the corresponding feature vector is an all-zero vector $\\mathbf{0}$ . Taking both cases into account, we can calculate the expectation $\\boldsymbol{v}$ of the feature as:", + "bbox": [ + 75, + 468, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} = \\mathcal {P} (\\boldsymbol {x}) \\boldsymbol {f} + (1 - \\mathcal {P} (\\boldsymbol {x})) \\boldsymbol {0} = p \\boldsymbol {f}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 782, + 392, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus the weighted feature $\\pmb{v}$ fed to MLP decoder has a physical meaning, namely the expectation of the feature at $\\pmb{x}$ being either $\\pmb{f}$ fetched from feature hashgrid or 0. Then training the saliency grid is equivalent to learning the probability distribution $\\mathcal{P}(\\pmb{x})$ . This probability distribution represents the likelihood that a given voxel $\\pmb{x}$ contains a non-zero fea", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture. A feature with a lower saliency weight $p$ is more likely to be 0 and will cause less interference to other features assigned to the same bucket in the feature hashgrid. When the saliency weight converges to 0, that feature is pruned with a probability of 1 and becomes a zero vector 0.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training the saliency distribution can also be viewed as hash collision mitigation. Instant-NGP employs hash encoding and distributes hash collisions evenly across voxels with non-zero density. By learning to shuffle hash collisions across the space, HollowNeRF guarantees less interference for more important features. When the feature of a voxel is scaled by a higher saliency weight, it tends to accumulate more gradients during backward propagation, and the information at this voxel will dominate the trained feature value inside the hash bucket shared with other voxels. On the other hand, a feature scaled by zero saliency weight has a zero gradient, hence causes no harm to colliding features.", + "bbox": [ + 496, + 167, + 892, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Soft zero-skipping gate", + "text_level": 1, + "bbox": [ + 500, + 356, + 714, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While HollowNeRF mitigates hash collisions by pruning unnecessary features to zero, the MLP used to decode these features into densities can break the sparsity in the 3D space: the MLP can be viewed as a function $\\sigma = \\mathcal{M}(\\boldsymbol{v})$ mapping the feature vector $\\boldsymbol{v}$ to the density $\\sigma$ , and typically $\\mathcal{M}(\\mathbf{0}) \\neq 0$ . Even though the saliency grid and ADMM pruner ensure most voxels direct to $\\mathbf{0}$ features in the hash-grid, a typical MLP translates such a sparse feature domain to a non-sparse 3D space with non-zero densities scattered everywhere, which is evidently inaccurate. To enforce a 0 output density when the input feature is $\\mathbf{0}$ , we introduce a zero-skipping gate $g(\\boldsymbol{v})$ to the MLP decoder:", + "bbox": [ + 496, + 378, + 890, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {M}} (\\boldsymbol {v}) = \\hat {g} (\\boldsymbol {v}) \\cdot \\mathcal {M} (\\boldsymbol {v})\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 571, + 799, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad \\hat {g} (\\boldsymbol {v}) := \\tanh (\\alpha \\| \\boldsymbol {v} \\| _ {2}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 590, + 890, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The gate $\\hat{g}(\\pmb{v})$ is a differentiable approximation of the \"hard\" gate in Eq. (2) to enforce the desired property $\\hat{\\mathcal{M}}(\\mathbf{0}) = 0$ , and we use a constant $\\alpha$ to control the steepness of the 0-to-1 transition.", + "bbox": [ + 496, + 617, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (\\boldsymbol {v}) := \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\| \\boldsymbol {v} \\| _ {2} = 0 \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 685, + 890, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The non-differentiable \"hard\" gate in Eq. (2) may cause abrupt density value changes during training. Also, the gradient $\\frac{\\partial g}{\\partial p}$ of $g(\\boldsymbol{v}) = g(p\\boldsymbol{f})$ is 0 with $p \\neq 0$ , this 0 gradient prevents the ADMM pruner from suppressing a small saliency value $p$ to 0, resulting in insufficient feature pruning and little improvement in hash collision. Our experiments in §4.3 confirm that adding the \"hard\" gate $g(\\boldsymbol{v})$ provides little gain over adding no gate.", + "bbox": [ + 496, + 734, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, we propose a \"soft\" zero-skipping gate $\\hat{g}(\\pmb{v})$ in Eq. (1) that smoothly reduces the density output to zero for input features being pruned. As an approximation of $g(\\pmb{v})$ ,", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3483", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the value of $\\hat{g}(\\boldsymbol{v})$ should smoothly transition from 0 to 1 for input $\\boldsymbol{v}$ with small magnitudes $\\|\\boldsymbol{v}\\|_2$ , while otherwise staying close to 1 to avoid perturbing necessary information. The tanh function is a suitable choice. In our implementation we adjust the $\\alpha$ value with a schedule, to gradually \"harden\" the soft gate as the training progresses. Specifically, for epochs below 1000, we set $\\alpha$ to $10^4$ to ensure fast convergence. Afterward we increase $\\alpha$ to $10^5$ to minimize the perturbation to the fine-tuning process.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. ADMM pruner", + "text_level": 1, + "bbox": [ + 76, + 244, + 230, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training the 3D saliency grid improves the accuracy by redistributing hash collisions, which, However, cannot fully eliminate unnecessary features, as shown in Figure 3a. To address this, we introduce an ADMM pruning that enforces sparsity in the saliency grid $\\mathcal{G}$ during training.", + "bbox": [ + 75, + 272, + 468, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first motivate the use of ADMM pruner rigorously. Assuming a $K \\times K \\times K$ volume where each voxel $\\pmb{x}$ corresponds to a feature $\\pmb{f}$ , and the ground truth feature $\\pmb{f}_1$ at voxel $\\pmb{x}_1 = (x_1, y_1, z_1)$ and $\\pmb{f}_2$ at a different voxel $\\pmb{x}_2 = (x_2, y_2, z_2)$ are typically different for a 3D scene. Instant-NGP encodes this $K \\times K \\times K$ feature grid into a small hashgrid $\\mathcal{H}$ whose size $<< (K \\times K \\times K)$ . When $\\pmb{f}_1$ and $\\pmb{f}_2$ collide with each other by directing to the same hash bucket, their values queried from the hashgrid become identical $\\pmb{f}_1^* = \\pmb{f}_2^* = \\pmb{f}_H$ , where $\\pmb{f}_H$ is the trained feature stored in the shared hash bucket. Since the ground truth features $\\pmb{f}_1 \\neq \\pmb{f}_2$ while the queried features $\\pmb{f}_1^* = \\pmb{f}_2^*$ , the queried values differ from the ground truth, causing compression artifacts. To solve this problem, HollowNeRF allows the colliding features to be different by introducing the scalar saliency weight $p$ to scale the queried features: $\\pmb{f}_1^* = p(\\pmb{x}_1)\\pmb{f}_H$ and $\\pmb{f}_2^* = p(\\pmb{x}_2)\\pmb{f}_H$ . However, during training, the share feature $\\pmb{f}_H$ may not find a value that satisfies both equations, as they can be written as $\\pmb{f}_1^* = \\frac{p(\\pmb{x}_1)}{p(\\pmb{x}_2)}\\pmb{f}_2^*$ which is not always true especially when $\\pmb{f}_1$ and $\\pmb{f}_2$ are nonzero. Then the queried features $\\pmb{f}_1^*$ and $\\pmb{f}_2^*$ may still cause compression artifacts (although less than instant-NGP). The ADMM pruner addresses this challenge by pruning the ground truth $\\pmb{f}_1$ or $\\pmb{f}_2$ to 0 when $\\pmb{x}_1$ or $\\pmb{x}_2$ is empty/invisible. For example, if voxel $\\pmb{x}_1$ is invisible from any view angle $(\\pmb{f}_1 = 0)$ , there exists a single $\\pmb{f}_H$ value to satisfy both $\\pmb{f}_1 = p(\\pmb{x}_1)\\pmb{f}_H$ and $\\pmb{f}_2 = p(\\pmb{x}_2)\\pmb{f}_H$ , that is $\\pmb{f}_H = \\pmb{f}_2$ , and there is no compression artifacts. The optimizer can converge to this optimal $\\pmb{f}_H$ value through gradient descent by learning that $p(\\pmb{x}_1)$ is 0. In summary, making the feature grid sparse reduces the compression artifacts when encoding features into a hashgrid.", + "bbox": [ + 75, + 349, + 472, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In what follows, we describe two methods for achieving sparsity: a basic L1-regularization approach and our ADMM pruner. To sparsify the saliency grid $\\mathcal{G}$ , one simple approach is to add an L1 regularization term to the original", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSE loss function $L$ typically used in NeRF training:", + "bbox": [ + 500, + 90, + 856, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {W}, \\mathcal {H}, \\mathcal {G}} \\left(L + \\lambda \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 118, + 890, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a constant weight to control the amount of sparsity and compression, and $\\mathcal{H}$ and $\\mathcal{W}$ denote respectively the embeddings in the multi-level hashgrid and the weights in the MLP. We use the sigmoid operator $S_{\\mathrm{sig}}(\\cdot)$ here because the forward pipeline (Figure 2) uses the saliency value after the sigmoid operator to weight the feature vector, and our ultimate goal is a sparse distribution of saliency value to prune unnecessary features. Note that $S_{\\mathrm{sig}}(\\mathcal{G})$ here denotes applying sigmoid operator on each element $g$ of the 3D tensor $\\mathcal{G}$ , and the result is a 3D tensor with all non-negative elements. Therefore, the absolute operator in the L1 regularization term can be removed:", + "bbox": [ + 496, + 154, + 893, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\| S _ {\\operatorname {s i g}} (\\mathcal {G}) \\| _ {1} = \\sum_ {g \\in \\mathcal {G}} | S _ {\\operatorname {s i g}} (g) | = \\sum_ {g \\in \\mathcal {G}} S _ {\\operatorname {s i g}} (g)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 345, + 836, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Removing the absolute operator from the regularization term makes it differentiable and compatible with the training pipeline, eliminating the need to use more complex optimization algorithms such as proximal gradient descent to handle non-differentiable regularization terms.", + "bbox": [ + 496, + 390, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the usage of L1 regularization is already a marked improved over the baseline (see §4.3), we find it is often difficult to select a single $\\lambda$ hyperparameter to control sparsity and compression. During training when the 3D saliency grid $\\mathcal{G}$ has already become sparse enough, the regularization term $\\lambda \\| S_{\\mathrm{sig}}(\\mathcal{G})\\|_1$ may still provide gradients that further prune $\\mathcal{G}$ . If $\\lambda$ is set too high or if the algorithm is let to run for too long, visible features may be pruned resulting in performance degradation.", + "bbox": [ + 496, + 465, + 890, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use an ADMM pruner to overcome the limitations of simple L1 regularization. By introducing a configurable sparsity constraint $C$ such that $\\| S_{\\mathrm{sig}}(\\mathcal{G})\\|_1 < C$ , feature pruning becomes the constrained optimization problem:", + "bbox": [ + 496, + 602, + 890, + 664 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\mathcal {M}, \\mathcal {H}, \\mathcal {G}} L \\left(\\mathcal {M}, \\mathcal {H}, \\mathcal {G}\\right) \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1} < C \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 691, + 815, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We transform this constrained optimization problem into an equivalent unconstrained minimax problem in Problem (5) by using the augmented Lagrangian method which employs a trainable dual variable $\\gamma$ , also known as a Lagrange multiplier. In essence, $\\gamma$ serves as a trainable alternative to the fixed hyperparameter $\\lambda$ in Problem (3).", + "bbox": [ + 496, + 747, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\mathcal {W}, \\mathcal {H}, \\mathcal {G}} \\max _ {\\gamma \\geq 0} L \\left(\\mathcal {W}, \\mathcal {H}, \\mathcal {G}\\right) + \\frac {\\rho_ {\\gamma}}{2} \\left[ \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1} - C \\right] _ {+} ^ {2} \\\\ + \\gamma \\left(\\left\\| S _ {\\text {s i g}} (\\mathcal {G}) \\right\\| _ {1} - C\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 849, + 890, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3484", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/90012d9cda07e88f118596a16af90f890752d6776442d5996673628c8f720d8d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHashgrid size214215216217218219
Param #
Instant-NGP0.50M0.94M1.77M3.34M6.22M11.49M
HollowNeRFT=640.76M1.21M2.04M3.60M6.48M11.75M
T=961.39M1.83M2.66M4.22M7.10M12.37M
T=1282.60M3.04M3.87M5.43M8.32M13.58M
", + "bbox": [ + 207, + 88, + 764, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: The total number of model parameters, depending on the hashgrid size and saliency grid resolution $T$ . This number counts the parameters in the hashgrid, MLP, and saliency grid. (both HollowNeRF and Instant-NGP employ a multi-resolution hashgrid with 16 levels. We use the maximum entries per level to denote the hashgrid size, following the notation in the Instant-NGP paper [16].)", + "bbox": [ + 75, + 217, + 893, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\rho_{\\gamma}$ denotes the learning rate of dual variable $\\gamma$ and $[\\cdot ]_{+}$ clamps the input to non-negative values.", + "bbox": [ + 76, + 305, + 470, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At each iteration $t$ , the optimizer first updates the coarse saliency grid $\\mathcal{G}$ , the feature hashgrid $\\mathcal{H}$ , and the MLP weights $\\mathcal{W}$ to minimize the augmented loss function in Problem (5) using gradient descent, where the dual variable $\\gamma$ is treated as a constant. $\\gamma$ is then updated using gradient ascent with $\\gamma^{(t + 1)} = \\left[\\gamma^{(t)} + \\rho_{\\gamma}\\left(\\| S_{\\mathrm{sig}}(\\mathcal{G})\\| _1 - C\\right)\\right]_+$", + "bbox": [ + 75, + 335, + 470, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 443, + 194, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first investigate the effect of different hyper-parameters, including hashgrid sizes, saliency grid sizes, and sparsity constraints, on the performance (§4.2). We then conduct an ablation study to validate each of HollowNeRF's design components (§4.3). Finally, we evaluate the performance of HollowNeRF on various 3D scenes and compare it with the state-of-the-art works (§4.4).", + "bbox": [ + 75, + 469, + 472, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 588, + 266, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement HollowNeRF on top of torch-ngp [25, 26], a PyTorch CUDA extension implementation of instant-NGP [16]. To ensure a fair comparison, we evaluated both HollowNeRF and instant-NGP using the same torch-ngp implementation. Both methods use a 16-level hashgrid with a feature dimension of 2, where the coarsest resolution of the feature grid is 16, and the finest resolution is 1024. Both methods employ a MLP decoder that has only two hidden layers with a width of 64 neurons, while HollowNeRF augments the MLP with a zero-skipping gate as described in §3.2.", + "bbox": [ + 75, + 612, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the peak signal-to-noise ratio (PSNR) and the Learned Perceptual Image Patch Similarity (LPIPS) [33] with the default AlexNet backend as the performance metrics to assess the rendering quality. Each experiments was conducted on a single NVIDIA Tesla A100 GPU for both training and evaluation. The machine used was equipped with an AMD EPYC 7742 64-Core CPU, 1TB of memory, and 8 GPU cards in total.", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fec9c0c29f1d0be494664c7383dcc15824fbd1ac9f6f9e7d0d7b93b1b8fef20d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 306, + 877, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1315eb5158ec21a3c15de2ad3ecb11f5c8b4f4362c69cb00ea0cf85fa043166.jpg", + "image_caption": [ + "(a) PSNR $\\uparrow$ (higher is better).", + "(b) LPIPS $\\downarrow$ (lower is better).", + "Figure 4: Performance vs. saliency grid size $T$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 503, + 888, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unless otherwise specified, we train the model for 300,000 steps using the Adam optimizer with an initial learning rate of $1 \\times 10^{-2}$ . Further details regarding the appropriate selection of HollowNeRF hyper-parameters, such as $T$ and $C$ , are discussed in §4.2.", + "bbox": [ + 496, + 758, + 893, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Investigating the cost-accuracy tradeoff", + "text_level": 1, + "bbox": [ + 498, + 845, + 841, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first investigate the cost-accuracy tradeoff of HollowNeRF to determine the appropriate hyper-parameter config", + "bbox": [ + 498, + 869, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3485", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9dbaf5eb65d4f1502ae0024b024709fbd374d2e26d68348c924e71c2aa26a471.jpg", + "image_caption": [ + "(a) PSNR $\\uparrow$ (higher is better)." + ], + "image_footnote": [], + "bbox": [ + 89, + 92, + 465, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efa6804fc207a5453459d80664d9ac47fb1f60fe32017b609be1af4d890a4651.jpg", + "image_caption": [ + "(b) LPIPS $\\downarrow$ (lower is better)." + ], + "image_footnote": [], + "bbox": [ + 91, + 292, + 460, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "uration for larger scale experiments in §4.4. We conducted experiments on the chair scene from the NeRF synthetic dataset, varying the hashgrid size from $2^{14}$ to $2^{19}$ and testing the saliency grid resolution of 64, 96, and 128 for each hashgrid size. Table 1 summarizes the cost of different configurations, including the number of parameters in the hashgrid, MLP, and saliency grid. We use a sparsity constraint of $C = 0.04$ in this experiment.", + "bbox": [ + 75, + 549, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Figure 4 indicate that increasing the saliency grid resolution $T$ from 64 to 96 or 128 does not improve accuracy much but results in a higher parameter count, which harms the cost-accuracy tradeoff. On the other hand, lowering $T$ below 64 causes unstable convergence. Consequently, we use a grid resolution of $T = 64$ for the subsequent experiments. Figure 4 further shows that HollowNeRF outperforms the baseline over a wide range of model sizes, achieving higher PSNR and lower LPIPS values for comparable parameter counts. The best overall model, HollowNeRF with $T = 64$ , uses 6.48M parameters to attain a 34.85dB PSNR, as compared to Instant-NGP's 33.86dB using 11.49M. In other words, it achieves about 1dB higher PSNR than Instant-NGP, while utilizing only 56% of the parameters employed by Instant-NGP.", + "bbox": [ + 75, + 672, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac5e9e6d256221fc1eec60364db84c60553091efd7f38c5c44a1d72d53f3f9b0.jpg", + "image_caption": [ + "(a) PSNR $\\uparrow$ (higher is better)." + ], + "image_footnote": [], + "bbox": [ + 514, + 93, + 883, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f841c4e3717ef361fabb2406887c4c495123d69e3d776b763d6b00e2328204b0.jpg", + "image_caption": [ + "Figure 5: Performance vs. sparsity constraint $C$ .", + "(b) LPIPS $\\downarrow$ (lower is better).", + "Figure 6: Ablation study for the design components." + ], + "image_footnote": [], + "bbox": [ + 511, + 292, + 887, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also repeated the training and evaluation for each hashgrid size five times for HollowNeRF with $T = 64$ and instant-NGP, and plotted the standard deviations across the five tests as error bars in Figure 4. The results demonstrate that our approach is more stable than Instant-NGP in terms of convergence, with lower performance variation.", + "bbox": [ + 496, + 559, + 890, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the impact of the choice of sparsity constraint $C$ on performance, we conducted additional experiments following the same setup as previous tests, while varying the $C$ values, and report the results in Figure 5. First, from Figure 5a, we observe that the $C = 0.04$ configuration yields the highest PSNR, in other words, HollowNeRF achieves the highest accuracy when $4\\%$ of the spatial voxels are non-zero. $C = 0.02$ exhibits slightly diminished accuracy, potentially due to excessive pruning in the 3D space, which compromises crucial information for scene reconstruction. Conversely, $C = 0.08$ shows the lowest accuracy across all tested HollowNeRF configurations, suggesting that keeping $8\\%$ non-zero voxels causes excessive hash collisions when packing them into the hashgrid, and the 3D scene to capture could be sparser. We have similar observation for LPIPS, as depicted in Figure 5b.", + "bbox": [ + 496, + 657, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3486", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b398dc24f851b638ba4004a1fc9aa6645146ea1857ac019ae983e8a3e005e2c1.jpg", + "image_caption": [ + "Figure 7: Qualitative comparison with instant-NGP over scenes from the NeRF synthetic dataset." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 493, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6a24caaa983c0d5743c48e5622edb063d25b6cfb296aa9a1ec8541d200ed7ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 88, + 888, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9202d24c5e0f9f79f255c311d39ff25a9fa89ae4c57c33d8d9caec8df9d9adf.jpg", + "image_caption": [ + "Figure 8: Comparison of PSNR(↑) and LPIPS (↓) performance between HollowNeRF and Instant-NGP on various scenes." + ], + "image_footnote": [], + "bbox": [ + 94, + 539, + 880, + 851 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3487", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8d0aca4088252db84726f814d567a6c69f22ce6b9963f4b6ef9d9116c773c86.jpg", + "image_caption": [ + "Figure 9: Comparison of HollowNeRF and recent methods in terms of the trade-off between PSNR and model size (\"HollowNeRF-H19\" means using a hashgrid size of $2^{19}$ )." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 459, + 289 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 376, + 227, + 391 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this ablation study, we analyze the impact of each component in our solution by comparing the convergence speed and resulting performance of the configurations with and without each design component. We use a saliency grid resolution of 64 and a hashgrid size of $2^{18}$ (total 6.48M parameters) for all HollowNeRF configurations, while the baseline is instant-NGP with a hashgrid size of $2^{19}$ (total 11.49M parameters). To investigate how each component affects the convergence speed, we evaluate the performance on the test sets every 300 epochs during the total 3000 epochs.", + "bbox": [ + 75, + 398, + 468, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results presented in Figure 6 show that each design component presence gracefully increases the overall performance, without any catastrophic failure. Each of the HollowNeRF ablated configurations still outperforms instant-NGP, while employing less parameters.", + "bbox": [ + 75, + 565, + 468, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Comparisons on standardized datasets", + "text_level": 1, + "bbox": [ + 76, + 650, + 410, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, we evaluate the performance of HollowNeRF on 3D scenes from the NeRF synthetic dataset [15] and the more complicated Tanks and Temples dataset [11]. Based on the cost-accuracy tradeoff investigated in Sec. 4.2, we set HollowNeRF's hyper-parameters to a sparsity constraint of $C = 0.04$ and a saliency grid resolution of $T = 64$ .", + "bbox": [ + 75, + 672, + 468, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first provide qualitative example views comparing HollowNeRF with Instant-NGP. The distinct advantages of HollowNeRF are evident through the visuals showcased in Figure 7. For example, from the lego scene in column 1, we observe that HollowNeRF with 6.48M parameters achieves notably superior rendering accuracy over Instant-NGP with $77\\%$ more parameters (11.49M). The advantage is particularly pronounced within the bulldozer track region. Furthermore, HollowNeRF with 11.75M parameters exhibits", + "bbox": [ + 75, + 765, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the highest accuracy, capturing fine details such as the tiny indentation on the edge of the black rod located at the left border of the view.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Then we show the PSNR and LPIPS performance of HollowNeRF and Instant-NGP on various 3D scenes (we sample 10 different scenes due to the page limit) in Figure 8. For each scene, we test three hashgrid sizes $(2^{17}, 2^{18}, 2^{19})$ for both methods. We can observe that HollowNeRF consistently outperforms Instant-NGP on all the 10 scenes tested, achieving better PSNR and LPIPS performance with fewer parameters when using the same hashgrid size. For seven out of ten tested scenes, the HollowNeRF with the smallest hashgrid size $(2^{17}$ , for a total parameter count including MLP of $3.34\\mathrm{M}$ ) achieves higher PSNR than the best Instant-NGP model (hashgrid size of $2^{19}$ , corresponding to $11.49\\mathrm{M}$ parameters including the MLP).", + "bbox": [ + 496, + 138, + 892, + 335 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, we compare the average PSNR across the NeRF synthetic dataset [16] and the corresponding model size for HollowNeRF and several recent works. Figure 9 shows that HollowNeRF achieves the best cost-accuracy tradeoff among all tested methods. In particular, HollowNeRF with a hashgrid size of $2^{17}$ achieves an average PSNR of 32.53dB with only a mode size of 14.0MB.", + "bbox": [ + 496, + 335, + 893, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 462, + 612, + 478 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The compression gains of HollowNeRF require the 3D scene being sparse, i.e. mostly filled with either empty or occluded space, which holds for typical 3D scenes. HollowNeRF can still operate when this assumption does not hold, such as for scenes containing smoke, fire, clouds, but its performance will regress to the baseline without pruning.", + "bbox": [ + 496, + 488, + 890, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The current version of HollowNeRF, like its predecessor Instant-NGP, faces challenges in modeling objects with reflective surfaces. In future works, we plan to extend the concept of HollowNeRF to the frameworks [8, 34, 21, 28, 14] with better support for reflective surfaces.", + "bbox": [ + 496, + 580, + 890, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 676, + 627, + 693 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we present HollowNeRF, a novel hashgrid-based NeRF technique that achieves superior rendering quality than state-of-the-art solutions like instant-NGP while using only a fraction of the parameters. HollowNeRF mitigates hash collisions by a simple feature pruning mechanism. Unlike existing methods that rely on explicit surface geometries for feature pruning, HollowNeRF learns a 3D saliency grid to guide feature compression during training, and employs an ADMM pruner to enforce a sparse feature domain. Our experiments demonstrate that HollowNeRF achieves a better balance between cost and accuracy than state-of-the-art solutions, making a solid step towards lightweight and ubiquitous NeRF applications.", + "bbox": [ + 496, + 703, + 892, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, October 2021. 1", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2", + "[3] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 2", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 9", + "[5] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6124-6132, June 2022. 2", + "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, June 2022. 2", + "[7] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 2", + "[8] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 2, 9", + "[9] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2", + "[10] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021. 2, 9", + "[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9", + "[12] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2", + "[13] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 1, 2", + "[14] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, June 2022. 2, 9", + "[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 9", + "[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 4, 6, 9", + "[17] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, June 2022. 2", + "[18] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 40(4), 2021. 1, 2", + "[19] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, June 2022. 2", + "[20] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In International Conference on Computer Vision (ICCV), 2021. 1", + "[21] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2, 9", + "[22] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 9", + "[23] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 9", + "[24] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, June 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3489", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp.6", + "[26] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6, 9", + "[27] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12922–12931, June 2022. 2", + "[28] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, June 2022. 2, 9", + "[29] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 2", + "[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, June 2022. 2", + "[31] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 9", + "[32] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2", + "[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[34] Chengxuan Zhu, Renjie Wan, and Boxin Shi. Neural transmitted radiance fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 9" + ], + "bbox": [ + 78, + 90, + 470, + 753 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3490", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_model.json b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_model.json new file mode 100644 index 0000000000000000000000000000000000000000..43714bb9ce1b48bfd12c0c4ab72b3ab4c849b403 --- /dev/null +++ b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_model.json @@ -0,0 +1,1960 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.131, + 0.723, + 0.176 + ], + "angle": 0, + "content": "HollowNeRF: Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.204, + 0.734, + 0.222 + ], + "angle": 0, + "content": "Xiufeng Xie, Riccardo Gherardi, Zhihong Pan, Stephen Huang" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.223, + 0.646, + 0.24 + ], + "angle": 0, + "content": "Oppo Mobile Telecommunications Corp." + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.24, + 0.682, + 0.257 + ], + "angle": 0, + "content": "2479 E Bayshore Rd, Palo Alto, CA, 94303, USA" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.261, + 0.558, + 0.274 + ], + "angle": 0, + "content": "xxie28@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.343, + 0.474, + 0.721 + ], + "angle": 0, + "content": "Neural radiance fields (NeRF) have garnered significant attention, with recent works such as Instant-NGP accelerating NeRF training and evaluation through a combination of hashgrid-based positional encoding and neural networks. However, effectively leveraging the spatial sparsity of 3D scenes remains a challenge. To cull away unnecessary regions of the feature grid, existing solutions rely on prior knowledge of object shape or periodically estimate object shape during training by repeated model evaluations, which are costly and wasteful. To address this issue, we propose HollowNeRF, a novel compression solution for hashgrid-based NeRF which automatically sparsifies the feature grid during the training phase. Instead of directly compressing dense features, HollowNeRF trains a coarse 3D saliency mask that guides efficient feature pruning, and employs an alternating direction method of multipliers (ADMM) pruner to sparsify the 3D saliency mask during training. By exploiting the sparsity in the 3D scene to redistribute hash collisions, HollowNeRF improves rendering quality while using a fraction of the parameters of comparable state-of-the-art solutions, leading to a better cost-accuracy trade-off. Our method delivers comparable rendering quality to Instant-NGP, while utilizing just \\(31\\%\\) of the parameters. In addition, our solution can achieve a PSNR accuracy gain of up to 1dB using only \\(56\\%\\) of the parameters." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.754, + 0.21, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF [1, 15]) have gained widespread recognition across academia and industry due to their remarkable capability to generate photorealistic novel views of 3D scenes from a collection of 2D images. Inspired by the volumetric representation, NeRF models the scene as a continuous 5D plenoptic function, enabling the creation of high-fidelity renderings with accurate lighting and shading effects. This technique has found versatile ap" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.308, + 0.891, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.526, + 0.894, + 0.603 + ], + "angle": 0, + "content": "Figure 1: Cross-section views of rendered 3D scenes. The left column shows the hollow interior of HollowNeRF's rendering, while the middle column shows the solid interior of Instant-NGP's rendering. The right column displays HollowNeRF's rendering without slicing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.659 + ], + "angle": 0, + "content": "plications in fields such as computer graphics, virtual and augmented reality, as well as robotics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Training and evaluating NeRF models can be computationally expensive, and many recent works [20, 16, 4, 22, 18, 23] have focused on improving the efficiency of NeRF. Instant-NGP [16] is an established solution with state-of-the-art training speed. It employs a lightweight hashgrid for input encoding and a small multi-layer perceptron (MLP) to disambiguate hash collisions. However, the lightweight hashgrid unavoidably causes severe collisions at fine resolutions. These collisions are evenly scattered across the occupied voxels, resulting in a suboptimal accuracy. Another line of research focuses on accelerating NeRF rendering by only sampling near the surface of interest, but has to rely on prior knowledge of surface geometries (either from conventional algorithms such as shape-from-silhouette [13] or an MLP predicting the depth distribution along each camera ray [18]). However, a coarse surface estimation degrades" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3480" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "NeRF rendering quality, while a precise surface estimation adds too much complexity, defeating the purpose of acceleration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.473, + 0.514 + ], + "angle": 0, + "content": "We propose HollowNeRF, a novel NeRF compression method using trainable hash collision mitigation to improve rendering accuracy while consuming less parameters than existing NeRF methods. Built on a hash-based pipeline from Instant-NGP, HollowNeRF prioritizes the important features (of visible voxels) and prunes unnecessary features (of invisible voxels), leading to a redistribution of hash collision probability across the 3D volume. When two voxels direct to the same hash bucket, Instant-NGP shapes the shared feature in this bucket as a mixture of the desired features, which harms the accuracy. In contrast, HollowNeRF steers the shared feature to fit the more important voxel, and prunes the feature of the less important voxel toward 0, reducing interference to features sharing the same bucket. Specifically, when reading the feature of a certain voxel, HollowNeRF further scales the feature by a trainable saliency weight whose value captures the voxel's visibility. To reduce the cost, we divide the 3D space into coarse grid regions and assign a saliency weight to capture each region's visibility, forming a trainable 3D saliency grid. Unlike existing methods [13, 18] that require prior knowledge of the surface geometries, HollowNeRF learns to prioritize the important features by training the saliency grid, which converges to a \"hollow\" saliency distribution across the 3D volume. Figure 1 showcases this \"hollow\" rendering result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.473, + 0.68 + ], + "angle": 0, + "content": "The proposed design consists of three main components: a trainable 3D saliency grid to guide the compression of dense features (\\(\\S 3.1\\)); a soft zero-skipping gate that enhances the MLP in the NeRF model to ensure a feature compressed to 0 translates to a zero density in the 3D space (\\(\\S 3.2\\)); a pruner to further push unnecessary features to exact 0 instead of a small non-zero value by alternating direction method of multipliers or ADMM [3] (\\(\\S 3.3\\)). Our experiments demonstrate that HollowNeRF achieves better accuracy (PSNR and LPIPS) than state-of-the-art methods while using significantly fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.681, + 0.357, + 0.695 + ], + "angle": 0, + "content": "The key contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.703, + 0.469, + 0.763 + ], + "angle": 0, + "content": "- We propose a novel NeRF compression solution, HollowNeRF, that learns to prioritize features defining the visible surface and prune invisible internal features without prior knowledge of surface geometries." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.772, + 0.469, + 0.832 + ], + "angle": 0, + "content": "- We use an ADMM-based optimization framework to prune unnecessary features during NeRF training and enhance the MLP with a soft zero-skipping gate to ensure that pruned features correctly map to zero density." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "- We evaluate the performance of HollowNeRF on popular NeRF datasets and our solution demonstrates a significantly superior balance between cost and accuracy than state-of-the-art solutions." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.703, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.776, + 0.108 + ], + "angle": 0, + "content": "2. Background and related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Differentiable rendering [15, 7, 5, 22, 17] has emerged as a prominent alternative framework for novel view synthesis, alongside the conventional 3D rendering approaches. Following the seminal NeRF paper [15], the research community has matured the approach and enables it to handle arbitrarily large scenes [24, 29, 27, 2], complex reflections [8, 34, 21, 28, 14], and small training datasets with a limited number of views [10, 6, 19]. Recent works also substantially improve the efficiency of NeRF, reducing the run-time complexity at training [16, 22] and inference [16, 12, 18, 30, 32, 31, 9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.893, + 0.523 + ], + "angle": 0, + "content": "Comparatively fewer works have tackled the problem of maximizing visual fidelity for a given space complexity budget [16, 26], which is the focus of this paper. The original NeRF technique [15] encodes the entire 3D scene in the weights of an MLP that predicts opacity \\((\\sigma)\\) and color \\((r,g,b)\\) given a 3D position \\((x,y,z)\\) and direction \\((\\theta ,\\phi)\\). While spatial complexity (438K parameters) is not the primary bottleneck, the substantial limitation of NeRF lies in the remarkably slow training and inference speeds. This is primarily attributed to the deep MLP with 8 hidden layers, each having a width of 256 neurons. It has since been shown [22] that differentiable rendering does not necessarily require neural networks. In [22], a scene is encoded as a neuron-free sparse voxel grid of opacities and RGB spherical harmonics. Without the MLP, training and inference run fast, at the cost of a significantly high spatial complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.893, + 0.765 + ], + "angle": 0, + "content": "In the continuum between the two extremes discussed above, there are techniques that harness the strengths of both ends and fare better in the balance between training speed and spatial complexity. Instant-NGP [16] is a representative example, which encodes features using a multi-resolution hashgrid and then feed the features into a lightweight MLP with only 2 hidden layers to decode the color and density; both the features in the hashgrid and the MLP are trained in conjunction. Our work extends the hash-based pipeline from Instant-NGP by achieving higher quality (PSNR & LPIPS) with less number of parameters. We achieve this goal by reclaiming the resources spent on empty, invisible, or internal regions. Instead of using an auxiliary MLP to predict space occupancy like DONeRF [18], we learn it through a trainable lightweight volumetric saliency grid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Some existing literature has applied model compression techniques to NeRF: CC-NeRF [26] uses tensor decomposition to obtain a low-rank approximation of the learned network. Similarly, TensoRF [4] represents the volume as a 4D tensor and factorizes it into low-rank components. Unlike these compression works, HollowNeRF vets information based on its actual impact on the rendering accuracy, rather than raw entropy. The relative performance of these approaches are investigated in §4.4." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "3481" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.092, + 0.832, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.297, + 0.655, + 0.312 + ], + "angle": 0, + "content": "Figure 2: Overview of the HollowNeRF workflow." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.337, + 0.271, + 0.356 + ], + "angle": 0, + "content": "3. HollowNeRF design" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.474, + 0.759 + ], + "angle": 0, + "content": "This section presents the system design of HollowNeRF, as outlined in Figure 2: (i) Given an input coordinate \\( \\boldsymbol{x} = (x,y,z) \\), we fetch the corresponding feature \\( \\boldsymbol{f} \\) from the multi-level hashgrid, following the Instant-NGP method [16]. (ii) We predict the saliency weight \\( p \\) of position \\( \\boldsymbol{x} \\) by processing the information from a trainable 3D saliency grid \\( \\mathcal{G} \\) through trilinear interpolation and a sigmoid function (§3.1), and use \\( p \\) to scale the feature \\( \\boldsymbol{f} \\) by \\( \\boldsymbol{v} = p\\boldsymbol{f} \\). Figure 3a showcases a trained saliency grid \\( \\mathcal{G} \\), demonstrating that training \\( \\mathcal{G} \\) can mitigate hash collisions by suppressing the unnecessary features in empty or invisible regions. (iii) The weighted feature \\( \\boldsymbol{v} \\) is fed into an MLP, which decodes \\( \\boldsymbol{v} \\) to obtain the density \\( \\sigma \\) and color at position \\( \\boldsymbol{x} \\). To ensure that \\( \\boldsymbol{v} = \\mathbf{0} \\) maps to \\( \\sigma = 0 \\), we introduce a zero-skipping gate (§3.2). (iv) The density and color outputs of the MLP are used for volume rendering, and the resulting image is compared to the ground truth to obtain a loss function \\( L \\). During training, an ADMM pruner (§3.3) sparsifies the saliency value distribution across the 3D space by enforcing a sparsity constraint when optimizing the loss \\( L \\). Figure 3a shows that a large portion of the unnecessary features are not entirely eliminated by training the saliency grid, with saliency weights diminished but not reaching 0. Therefore, we introduce the ADMM pruner to explicitly prune saliency weights of the empty or invisible regions to 0, as shown in Figure 3b." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.318, + 0.788 + ], + "angle": 0, + "content": "3.1. Trainable 3D saliency grid" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In typical 3D scenes without large transparent objects, most regions of the space do not contribute to the final rendering as they are either empty or invisible from any view angle. To leverage this sparsity, methods such as Instant-NGP use a coarse-grained binary mask to capture space occupancy, which is periodically updated during training by evaluating the NeRF model at each voxel and checking if the density is" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.337, + 0.58, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.469, + 0.686, + 0.484 + ], + "angle": 0, + "content": "(a) Without ADMM pruner." + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.338, + 0.69, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.339, + 0.77, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.469, + 0.866, + 0.484 + ], + "angle": 0, + "content": "(b) With ADMM pruner." + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.339, + 0.885, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.495, + 0.893, + 0.527 + ], + "angle": 0, + "content": "Figure 3: Slices of a \\(64 \\times 64 \\times 64\\) saliency grid trained on the \"Chair\" scene from the NeRF synthetic dataset[15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.643 + ], + "angle": 0, + "content": "above a threshold. This occupancy mask guides sampling in ray marching by skipping unoccupied regions. However, it still keeps unnecessary features in internal occluded regions that have non-zero densities but make no contributions to the visible surfaces from any view angle, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.765 + ], + "angle": 0, + "content": "With this insight, we propose to reclaim the capacity spent on internal regions by learning the sparsity pattern through training, effectively making objects hollow. The benefits of this approach are twofold: it reduces space complexity by requiring fewer features, and improves accuracy by reducing hash collisions. However, precisely locating the invisible regions of objects can be challenging without any prior knowledge of the object's shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.902 + ], + "angle": 0, + "content": "To addresses this challenge, our approach utilizes the hashgrid-based positional encoding from Instant-NGP [16] and extends it by introducing a trainable saliency grid \\(\\mathcal{G}\\) with a resolution of \\(T\\times T\\times T\\). Specifically, HollowNeRF trains a 3D tensor data structure \\(\\mathcal{G}\\) alongside the NeRF model (feature hashgrid \\(\\mathcal{H}\\) and MLP weights \\(\\mathcal{W}\\)), which gradually learns which regions to prioritize, such as a nontransparent object surface. We split the space coarsely as \\(T\\times T\\times T\\) uniform grids and the 3D tensor \\(\\mathcal{G}\\) stores train" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3482" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "able values representing the saliency of features near each grid voxel. Instead of letting multiple dense features in the same coarse grid share one value, we interpolate the values associated to the 8 voxels surrounding the input coordinate \\( \\pmb{x} \\) to get a smoothed saliency distribution \\( \\mathcal{P} \\) over the 3D space. The trainable saliency grid represents in itself an opportunity to compress the NeRF model using differentiable model compression techniques, as discussed later in §3.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.469 + ], + "angle": 0, + "content": "We initialize \\(\\mathcal{G}\\) as an all-ones tensor, where \\(T\\) is typically much smaller than the resolution of the dense feature grid (In §4.2, we further investigate how different \\(T\\) values affect the performance). Given a 3D coordinate \\(\\pmb{x} = (x,y,z)\\), we fetch its corresponding feature vector \\(\\pmb{f}\\) from the multi-level hashgrid [16]. We then locate the 8 voxels \\((\\pmb{x}_1,\\pmb{x}_2,\\dots,\\pmb{x}_8)\\) surrounding \\(\\pmb{x}\\) in the 3D saliency grid \\(\\mathcal{G}\\), and fetch its corresponding saliency value \\(g_i\\). Specifically, the coordinate \\(\\pmb{x}_i = (x_i,y_i,z_i)\\) are quantized to the values \\((\\hat{x}_i,\\hat{y}_i,\\hat{z}_i)\\) between 0 and \\(T - 1\\), which are then used to index the saliency grid \\(\\mathcal{G}\\). This gives us the saliency value \\(g_i = \\mathcal{G}(\\hat{x}_i,\\hat{y}_i,\\hat{z}_i)\\). After obtaining \\(g_i\\) for each of the 8 saliency grid voxels surrounding \\(\\pmb{x}\\), we calculate the saliency weight \\(p\\) at coordinate \\(\\pmb{x}\\) by trilinear interpolation followed by a sigmoid operator, so that \\(0 < p < 1\\). The final weighted feature vector \\(\\pmb{v}\\) is obtained by multiplying the saliency weight \\(p\\) with the feature vector \\(\\pmb{f}\\) fetched from the hashgrid, i.e., \\(\\pmb{v} = p\\pmb{f}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.771 + ], + "angle": 0, + "content": "For mathematical analysis, we define the saliency distribution \\(\\mathcal{P}\\) as the collection of saliency weights \\(p\\) at every possible coordinate \\(\\boldsymbol{x}\\), such that \\(p = \\mathcal{P}(\\boldsymbol{x})\\). Our proposed method, HollowNeRF, learns the saliency distribution \\(\\mathcal{P}\\) by training the saliency grid \\(\\mathcal{G}\\) using gradient descent. It is worth noting that we inject \\(\\mathcal{P}\\) (and thus \\(\\mathcal{G}\\)) to the training pipeline by feeding the weighted feature \\(\\boldsymbol{v}\\) instead of the raw feature \\(\\boldsymbol{f}\\) to the MLP for decoding; the rationale for this choice is that since the sigmoid function in the above workflow projects the saliency weights to the range \\((0,1)\\), we can view the 3D saliency distribution as a probability distribution \\(\\mathcal{P}(\\boldsymbol{x})\\) indicating the likelihood \\(p\\) that a feature is non-zero at a given 3D position \\(\\boldsymbol{x}\\). There are only two possibilities for the feature at a given position: either the position is visible with probability \\(p = \\mathcal{P}(\\boldsymbol{x})\\), and we retrieve the feature vector \\(\\boldsymbol{f}\\) from the multi-level hashgrid as in [16]; otherwise the position is empty or occluded with probability \\(1 - p\\), and the corresponding feature vector is an all-zero vector \\(\\mathbf{0}\\). Taking both cases into account, we can calculate the expectation \\(\\boldsymbol{v}\\) of the feature as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.784, + 0.393, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} = \\mathcal {P} (\\boldsymbol {x}) \\boldsymbol {f} + (1 - \\mathcal {P} (\\boldsymbol {x})) \\boldsymbol {0} = p \\boldsymbol {f}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Thus the weighted feature \\( \\pmb{v} \\) fed to MLP decoder has a physical meaning, namely the expectation of the feature at \\( \\pmb{x} \\) being either \\( \\pmb{f} \\) fetched from feature hashgrid or 0. Then training the saliency grid is equivalent to learning the probability distribution \\( \\mathcal{P}(\\pmb{x}) \\). This probability distribution represents the likelihood that a given voxel \\( \\pmb{x} \\) contains a non-zero fea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "ture. A feature with a lower saliency weight \\( p \\) is more likely to be 0 and will cause less interference to other features assigned to the same bucket in the feature hashgrid. When the saliency weight converges to 0, that feature is pruned with a probability of 1 and becomes a zero vector 0." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.349 + ], + "angle": 0, + "content": "Training the saliency distribution can also be viewed as hash collision mitigation. Instant-NGP employs hash encoding and distributes hash collisions evenly across voxels with non-zero density. By learning to shuffle hash collisions across the space, HollowNeRF guarantees less interference for more important features. When the feature of a voxel is scaled by a higher saliency weight, it tends to accumulate more gradients during backward propagation, and the information at this voxel will dominate the trained feature value inside the hash bucket shared with other voxels. On the other hand, a feature scaled by zero saliency weight has a zero gradient, hence causes no harm to colliding features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.715, + 0.374 + ], + "angle": 0, + "content": "3.2. Soft zero-skipping gate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.562 + ], + "angle": 0, + "content": "While HollowNeRF mitigates hash collisions by pruning unnecessary features to zero, the MLP used to decode these features into densities can break the sparsity in the 3D space: the MLP can be viewed as a function \\(\\sigma = \\mathcal{M}(\\boldsymbol{v})\\) mapping the feature vector \\(\\boldsymbol{v}\\) to the density \\(\\sigma\\), and typically \\(\\mathcal{M}(\\mathbf{0}) \\neq 0\\). Even though the saliency grid and ADMM pruner ensure most voxels direct to \\(\\mathbf{0}\\) features in the hash-grid, a typical MLP translates such a sparse feature domain to a non-sparse 3D space with non-zero densities scattered everywhere, which is evidently inaccurate. To enforce a 0 output density when the input feature is \\(\\mathbf{0}\\), we introduce a zero-skipping gate \\(g(\\boldsymbol{v})\\) to the MLP decoder:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.573, + 0.8, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {M}} (\\boldsymbol {v}) = \\hat {g} (\\boldsymbol {v}) \\cdot \\mathcal {M} (\\boldsymbol {v})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.592, + 0.891, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad \\hat {g} (\\boldsymbol {v}) := \\tanh (\\alpha \\| \\boldsymbol {v} \\| _ {2}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.678 + ], + "angle": 0, + "content": "The gate \\(\\hat{g}(\\pmb{v})\\) is a differentiable approximation of the \"hard\" gate in Eq. (2) to enforce the desired property \\(\\hat{\\mathcal{M}}(\\mathbf{0}) = 0\\), and we use a constant \\(\\alpha\\) to control the steepness of the 0-to-1 transition." + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.686, + 0.892, + 0.726 + ], + "angle": 0, + "content": "\\[\ng (\\boldsymbol {v}) := \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\| \\boldsymbol {v} \\| _ {2} = 0 \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.856 + ], + "angle": 0, + "content": "The non-differentiable \"hard\" gate in Eq. (2) may cause abrupt density value changes during training. Also, the gradient \\(\\frac{\\partial g}{\\partial p}\\) of \\(g(\\boldsymbol{v}) = g(p\\boldsymbol{f})\\) is 0 with \\(p \\neq 0\\), this 0 gradient prevents the ADMM pruner from suppressing a small saliency value \\(p\\) to 0, resulting in insufficient feature pruning and little improvement in hash collision. Our experiments in §4.3 confirm that adding the \"hard\" gate \\(g(\\boldsymbol{v})\\) provides little gain over adding no gate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Thus, we propose a \"soft\" zero-skipping gate \\(\\hat{g}(\\pmb{v})\\) in Eq. (1) that smoothly reduces the density output to zero for input features being pruned. As an approximation of \\(g(\\pmb{v})\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3483" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "the value of \\(\\hat{g}(\\boldsymbol{v})\\) should smoothly transition from 0 to 1 for input \\(\\boldsymbol{v}\\) with small magnitudes \\(\\|\\boldsymbol{v}\\|_2\\), while otherwise staying close to 1 to avoid perturbing necessary information. The tanh function is a suitable choice. In our implementation we adjust the \\(\\alpha\\) value with a schedule, to gradually \"harden\" the soft gate as the training progresses. Specifically, for epochs below 1000, we set \\(\\alpha\\) to \\(10^4\\) to ensure fast convergence. Afterward we increase \\(\\alpha\\) to \\(10^5\\) to minimize the perturbation to the fine-tuning process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.246, + 0.232, + 0.263 + ], + "angle": 0, + "content": "3.3. ADMM pruner" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Training the 3D saliency grid improves the accuracy by redistributing hash collisions, which, However, cannot fully eliminate unnecessary features, as shown in Figure 3a. To address this, we introduce an ADMM pruning that enforces sparsity in the saliency grid \\(\\mathcal{G}\\) during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.838 + ], + "angle": 0, + "content": "We first motivate the use of ADMM pruner rigorously. Assuming a \\(K \\times K \\times K\\) volume where each voxel \\(\\pmb{x}\\) corresponds to a feature \\(\\pmb{f}\\), and the ground truth feature \\(\\pmb{f}_1\\) at voxel \\(\\pmb{x}_1 = (x_1, y_1, z_1)\\) and \\(\\pmb{f}_2\\) at a different voxel \\(\\pmb{x}_2 = (x_2, y_2, z_2)\\) are typically different for a 3D scene. Instant-NGP encodes this \\(K \\times K \\times K\\) feature grid into a small hashgrid \\(\\mathcal{H}\\) whose size \\(<< (K \\times K \\times K)\\). When \\(\\pmb{f}_1\\) and \\(\\pmb{f}_2\\) collide with each other by directing to the same hash bucket, their values queried from the hashgrid become identical \\(\\pmb{f}_1^* = \\pmb{f}_2^* = \\pmb{f}_H\\), where \\(\\pmb{f}_H\\) is the trained feature stored in the shared hash bucket. Since the ground truth features \\(\\pmb{f}_1 \\neq \\pmb{f}_2\\) while the queried features \\(\\pmb{f}_1^* = \\pmb{f}_2^*\\), the queried values differ from the ground truth, causing compression artifacts. To solve this problem, HollowNeRF allows the colliding features to be different by introducing the scalar saliency weight \\(p\\) to scale the queried features: \\(\\pmb{f}_1^* = p(\\pmb{x}_1)\\pmb{f}_H\\) and \\(\\pmb{f}_2^* = p(\\pmb{x}_2)\\pmb{f}_H\\). However, during training, the share feature \\(\\pmb{f}_H\\) may not find a value that satisfies both equations, as they can be written as \\(\\pmb{f}_1^* = \\frac{p(\\pmb{x}_1)}{p(\\pmb{x}_2)}\\pmb{f}_2^*\\) which is not always true especially when \\(\\pmb{f}_1\\) and \\(\\pmb{f}_2\\) are nonzero. Then the queried features \\(\\pmb{f}_1^*\\) and \\(\\pmb{f}_2^*\\) may still cause compression artifacts (although less than instant-NGP). The ADMM pruner addresses this challenge by pruning the ground truth \\(\\pmb{f}_1\\) or \\(\\pmb{f}_2\\) to 0 when \\(\\pmb{x}_1\\) or \\(\\pmb{x}_2\\) is empty/invisible. For example, if voxel \\(\\pmb{x}_1\\) is invisible from any view angle \\((\\pmb{f}_1 = 0)\\), there exists a single \\(\\pmb{f}_H\\) value to satisfy both \\(\\pmb{f}_1 = p(\\pmb{x}_1)\\pmb{f}_H\\) and \\(\\pmb{f}_2 = p(\\pmb{x}_2)\\pmb{f}_H\\), that is \\(\\pmb{f}_H = \\pmb{f}_2\\), and there is no compression artifacts. The optimizer can converge to this optimal \\(\\pmb{f}_H\\) value through gradient descent by learning that \\(p(\\pmb{x}_1)\\) is 0. In summary, making the feature grid sparse reduces the compression artifacts when encoding features into a hashgrid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In what follows, we describe two methods for achieving sparsity: a basic L1-regularization approach and our ADMM pruner. To sparsify the saliency grid \\(\\mathcal{G}\\), one simple approach is to add an L1 regularization term to the original" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.857, + 0.108 + ], + "angle": 0, + "content": "MSE loss function \\(L\\) typically used in NeRF training:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.119, + 0.892, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {W}, \\mathcal {H}, \\mathcal {G}} \\left(L + \\lambda \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.895, + 0.336 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a constant weight to control the amount of sparsity and compression, and \\(\\mathcal{H}\\) and \\(\\mathcal{W}\\) denote respectively the embeddings in the multi-level hashgrid and the weights in the MLP. We use the sigmoid operator \\(S_{\\mathrm{sig}}(\\cdot)\\) here because the forward pipeline (Figure 2) uses the saliency value after the sigmoid operator to weight the feature vector, and our ultimate goal is a sparse distribution of saliency value to prune unnecessary features. Note that \\(S_{\\mathrm{sig}}(\\mathcal{G})\\) here denotes applying sigmoid operator on each element \\(g\\) of the 3D tensor \\(\\mathcal{G}\\), and the result is a 3D tensor with all non-negative elements. Therefore, the absolute operator in the L1 regularization term can be removed:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.346, + 0.837, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\| S _ {\\operatorname {s i g}} (\\mathcal {G}) \\| _ {1} = \\sum_ {g \\in \\mathcal {G}} | S _ {\\operatorname {s i g}} (g) | = \\sum_ {g \\in \\mathcal {G}} S _ {\\operatorname {s i g}} (g)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Removing the absolute operator from the regularization term makes it differentiable and compatible with the training pipeline, eliminating the need to use more complex optimization algorithms such as proximal gradient descent to handle non-differentiable regularization terms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Although the usage of L1 regularization is already a marked improved over the baseline (see §4.3), we find it is often difficult to select a single \\(\\lambda\\) hyperparameter to control sparsity and compression. During training when the 3D saliency grid \\(\\mathcal{G}\\) has already become sparse enough, the regularization term \\(\\lambda \\| S_{\\mathrm{sig}}(\\mathcal{G})\\|_1\\) may still provide gradients that further prune \\(\\mathcal{G}\\). If \\(\\lambda\\) is set too high or if the algorithm is let to run for too long, visible features may be pruned resulting in performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.665 + ], + "angle": 0, + "content": "We use an ADMM pruner to overcome the limitations of simple L1 regularization. By introducing a configurable sparsity constraint \\( C \\) such that \\( \\| S_{\\mathrm{sig}}(\\mathcal{G})\\|_1 < C \\), feature pruning becomes the constrained optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.693, + 0.816, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\mathcal {M}, \\mathcal {H}, \\mathcal {G}} L \\left(\\mathcal {M}, \\mathcal {H}, \\mathcal {G}\\right) \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1} < C \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We transform this constrained optimization problem into an equivalent unconstrained minimax problem in Problem (5) by using the augmented Lagrangian method which employs a trainable dual variable \\(\\gamma\\), also known as a Lagrange multiplier. In essence, \\(\\gamma\\) serves as a trainable alternative to the fixed hyperparameter \\(\\lambda\\) in Problem (3)." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.85, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\mathcal {W}, \\mathcal {H}, \\mathcal {G}} \\max _ {\\gamma \\geq 0} L \\left(\\mathcal {W}, \\mathcal {H}, \\mathcal {G}\\right) + \\frac {\\rho_ {\\gamma}}{2} \\left[ \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1} - C \\right] _ {+} ^ {2} \\\\ + \\gamma \\left(\\left\\| S _ {\\text {s i g}} (\\mathcal {G}) \\right\\| _ {1} - C\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3484" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.209, + 0.089, + 0.765, + 0.208 + ], + "angle": 0, + "content": "
MethodHashgrid size214215216217218219
Param #
Instant-NGP0.50M0.94M1.77M3.34M6.22M11.49M
HollowNeRFT=640.76M1.21M2.04M3.60M6.48M11.75M
T=961.39M1.83M2.66M4.22M7.10M12.37M
T=1282.60M3.04M3.87M5.43M8.32M13.58M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.218, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Table 1: The total number of model parameters, depending on the hashgrid size and saliency grid resolution \\(T\\). This number counts the parameters in the hashgrid, MLP, and saliency grid. (both HollowNeRF and Instant-NGP employ a multi-resolution hashgrid with 16 levels. We use the maximum entries per level to denote the hashgrid size, following the notation in the Instant-NGP paper [16].)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.471, + 0.336 + ], + "angle": 0, + "content": "where \\(\\rho_{\\gamma}\\) denotes the learning rate of dual variable \\(\\gamma\\) and \\([\\cdot ]_{+}\\) clamps the input to non-negative values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.471, + 0.432 + ], + "angle": 0, + "content": "At each iteration \\(t\\), the optimizer first updates the coarse saliency grid \\(\\mathcal{G}\\), the feature hashgrid \\(\\mathcal{H}\\), and the MLP weights \\(\\mathcal{W}\\) to minimize the augmented loss function in Problem (5) using gradient descent, where the dual variable \\(\\gamma\\) is treated as a constant. \\(\\gamma\\) is then updated using gradient ascent with \\(\\gamma^{(t + 1)} = \\left[\\gamma^{(t)} + \\rho_{\\gamma}\\left(\\| S_{\\mathrm{sig}}(\\mathcal{G})\\| _1 - C\\right)\\right]_+\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.444, + 0.196, + 0.46 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.473, + 0.579 + ], + "angle": 0, + "content": "In this section, we first investigate the effect of different hyper-parameters, including hashgrid sizes, saliency grid sizes, and sparsity constraints, on the performance (§4.2). We then conduct an ablation study to validate each of HollowNeRF's design components (§4.3). Finally, we evaluate the performance of HollowNeRF on various 3D scenes and compare it with the state-of-the-art works (§4.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.267, + 0.606 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.779 + ], + "angle": 0, + "content": "We implement HollowNeRF on top of torch-ngp [25, 26], a PyTorch CUDA extension implementation of instant-NGP [16]. To ensure a fair comparison, we evaluated both HollowNeRF and instant-NGP using the same torch-ngp implementation. Both methods use a 16-level hashgrid with a feature dimension of 2, where the coarsest resolution of the feature grid is 16, and the finest resolution is 1024. Both methods employ a MLP decoder that has only two hidden layers with a width of 64 neurons, while HollowNeRF augments the MLP with a zero-skipping gate as described in §3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We use the peak signal-to-noise ratio (PSNR) and the Learned Perceptual Image Patch Similarity (LPIPS) [33] with the default AlexNet backend as the performance metrics to assess the rendering quality. Each experiments was conducted on a single NVIDIA Tesla A100 GPU for both training and evaluation. The machine used was equipped with an AMD EPYC 7742 64-Core CPU, 1TB of memory, and 8 GPU cards in total." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.308, + 0.878, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.489, + 0.785, + 0.502 + ], + "angle": 0, + "content": "(a) PSNR \\(\\uparrow\\) (higher is better)." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.504, + 0.889, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.688, + 0.783, + 0.702 + ], + "angle": 0, + "content": "(b) LPIPS \\(\\downarrow\\) (lower is better)." + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.715, + 0.854, + 0.731 + ], + "angle": 0, + "content": "Figure 4: Performance vs. saliency grid size \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.894, + 0.835 + ], + "angle": 0, + "content": "Unless otherwise specified, we train the model for 300,000 steps using the Adam optimizer with an initial learning rate of \\(1 \\times 10^{-2}\\). Further details regarding the appropriate selection of HollowNeRF hyper-parameters, such as \\(T\\) and \\(C\\), are discussed in §4.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.847, + 0.842, + 0.864 + ], + "angle": 0, + "content": "4.2. Investigating the cost-accuracy tradeoff" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "We first investigate the cost-accuracy tradeoff of HollowNeRF to determine the appropriate hyper-parameter config" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3485" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.093, + 0.466, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.274, + 0.362, + 0.289 + ], + "angle": 0, + "content": "(a) PSNR \\(\\uparrow\\) (higher is better)." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.293, + 0.462, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.474, + 0.361, + 0.488 + ], + "angle": 0, + "content": "(b) LPIPS \\(\\downarrow\\) (lower is better)." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.5, + 0.434, + 0.516 + ], + "angle": 0, + "content": "Figure 5: Performance vs. sparsity constraint \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.672 + ], + "angle": 0, + "content": "uration for larger scale experiments in §4.4. We conducted experiments on the chair scene from the NeRF synthetic dataset, varying the hashgrid size from \\(2^{14}\\) to \\(2^{19}\\) and testing the saliency grid resolution of 64, 96, and 128 for each hashgrid size. Table 1 summarizes the cost of different configurations, including the number of parameters in the hashgrid, MLP, and saliency grid. We use a sparsity constraint of \\(C = 0.04\\) in this experiment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The results in Figure 4 indicate that increasing the saliency grid resolution \\( T \\) from 64 to 96 or 128 does not improve accuracy much but results in a higher parameter count, which harms the cost-accuracy tradeoff. On the other hand, lowering \\( T \\) below 64 causes unstable convergence. Consequently, we use a grid resolution of \\( T = 64 \\) for the subsequent experiments. Figure 4 further shows that HollowNeRF outperforms the baseline over a wide range of model sizes, achieving higher PSNR and lower LPIPS values for comparable parameter counts. The best overall model, HollowNeRF with \\( T = 64 \\), uses 6.48M parameters to attain a 34.85dB PSNR, as compared to Instant-NGP's 33.86dB using 11.49M. In other words, it achieves about 1dB higher PSNR than Instant-NGP, while utilizing only 56% of the parameters employed by Instant-NGP." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.094, + 0.885, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.274, + 0.786, + 0.289 + ], + "angle": 0, + "content": "(a) PSNR \\(\\uparrow\\) (higher is better)." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.293, + 0.888, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.474, + 0.784, + 0.488 + ], + "angle": 0, + "content": "(b) LPIPS \\(\\downarrow\\) (lower is better)." + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.5, + 0.868, + 0.516 + ], + "angle": 0, + "content": "Figure 6: Ablation study for the design components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.651 + ], + "angle": 0, + "content": "We also repeated the training and evaluation for each hashgrid size five times for HollowNeRF with \\( T = 64 \\) and instant-NGP, and plotted the standard deviations across the five tests as error bars in Figure 4. The results demonstrate that our approach is more stable than Instant-NGP in terms of convergence, with lower performance variation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.902 + ], + "angle": 0, + "content": "To investigate the impact of the choice of sparsity constraint \\( C \\) on performance, we conducted additional experiments following the same setup as previous tests, while varying the \\( C \\) values, and report the results in Figure 5. First, from Figure 5a, we observe that the \\( C = 0.04 \\) configuration yields the highest PSNR, in other words, HollowNeRF achieves the highest accuracy when \\( 4\\% \\) of the spatial voxels are non-zero. \\( C = 0.02 \\) exhibits slightly diminished accuracy, potentially due to excessive pruning in the 3D space, which compromises crucial information for scene reconstruction. Conversely, \\( C = 0.08 \\) shows the lowest accuracy across all tested HollowNeRF configurations, suggesting that keeping \\( 8\\% \\) non-zero voxels causes excessive hash collisions when packing them into the hashgrid, and the 3D scene to capture could be sparser. We have similar observation for LPIPS, as depicted in Figure 5b." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3486" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.495, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.089, + 0.889, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.513, + 0.804, + 0.53 + ], + "angle": 0, + "content": "Figure 7: Qualitative comparison with instant-NGP over scenes from the NeRF synthetic dataset." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.54, + 0.881, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.86, + 0.885, + 0.876 + ], + "angle": 0, + "content": "Figure 8: Comparison of PSNR(↑) and LPIPS (↓) performance between HollowNeRF and Instant-NGP on various scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3487" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.46, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Figure 9: Comparison of HollowNeRF and recent methods in terms of the trade-off between PSNR and model size (\"HollowNeRF-H19\" means using a hashgrid size of \\(2^{19}\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.377, + 0.228, + 0.392 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.566 + ], + "angle": 0, + "content": "In this ablation study, we analyze the impact of each component in our solution by comparing the convergence speed and resulting performance of the configurations with and without each design component. We use a saliency grid resolution of 64 and a hashgrid size of \\(2^{18}\\) (total 6.48M parameters) for all HollowNeRF configurations, while the baseline is instant-NGP with a hashgrid size of \\(2^{19}\\) (total 11.49M parameters). To investigate how each component affects the convergence speed, we evaluate the performance on the test sets every 300 epochs during the total 3000 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.47, + 0.643 + ], + "angle": 0, + "content": "The results presented in Figure 6 show that each design component presence gracefully increases the overall performance, without any catastrophic failure. Each of the HollowNeRF ablated configurations still outperforms instant-NGP, while employing less parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.411, + 0.667 + ], + "angle": 0, + "content": "4.4. Comparisons on standardized datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Finally, we evaluate the performance of HollowNeRF on 3D scenes from the NeRF synthetic dataset [15] and the more complicated Tanks and Temples dataset [11]. Based on the cost-accuracy tradeoff investigated in Sec. 4.2, we set HollowNeRF's hyper-parameters to a sparsity constraint of \\( C = 0.04 \\) and a saliency grid resolution of \\( T = 64 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We first provide qualitative example views comparing HollowNeRF with Instant-NGP. The distinct advantages of HollowNeRF are evident through the visuals showcased in Figure 7. For example, from the lego scene in column 1, we observe that HollowNeRF with 6.48M parameters achieves notably superior rendering accuracy over Instant-NGP with \\(77\\%\\) more parameters (11.49M). The advantage is particularly pronounced within the bulldozer track region. Furthermore, HollowNeRF with 11.75M parameters exhibits" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "the highest accuracy, capturing fine details such as the tiny indentation on the edge of the black rod located at the left border of the view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.893, + 0.336 + ], + "angle": 0, + "content": "Then we show the PSNR and LPIPS performance of HollowNeRF and Instant-NGP on various 3D scenes (we sample 10 different scenes due to the page limit) in Figure 8. For each scene, we test three hashgrid sizes \\((2^{17}, 2^{18}, 2^{19})\\) for both methods. We can observe that HollowNeRF consistently outperforms Instant-NGP on all the 10 scenes tested, achieving better PSNR and LPIPS performance with fewer parameters when using the same hashgrid size. For seven out of ten tested scenes, the HollowNeRF with the smallest hashgrid size \\((2^{17}\\), for a total parameter count including MLP of \\(3.34\\mathrm{M}\\)) achieves higher PSNR than the best Instant-NGP model (hashgrid size of \\(2^{19}\\), corresponding to \\(11.49\\mathrm{M}\\) parameters including the MLP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.894, + 0.443 + ], + "angle": 0, + "content": "Finally, we compare the average PSNR across the NeRF synthetic dataset [16] and the corresponding model size for HollowNeRF and several recent works. Figure 9 shows that HollowNeRF achieves the best cost-accuracy tradeoff among all tested methods. In particular, HollowNeRF with a hashgrid size of \\(2^{17}\\) achieves an average PSNR of 32.53dB with only a mode size of 14.0MB." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.614, + 0.479 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.581 + ], + "angle": 0, + "content": "The compression gains of HollowNeRF require the 3D scene being sparse, i.e. mostly filled with either empty or occluded space, which holds for typical 3D scenes. HollowNeRF can still operate when this assumption does not hold, such as for scenes containing smoke, fire, clouds, but its performance will regress to the baseline without pruning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.658 + ], + "angle": 0, + "content": "The current version of HollowNeRF, like its predecessor Instant-NGP, faces challenges in modeling objects with reflective surfaces. In future works, we plan to extend the concept of HollowNeRF to the frameworks [8, 34, 21, 28, 14] with better support for reflective surfaces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.628, + 0.694 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present HollowNeRF, a novel hashgrid-based NeRF technique that achieves superior rendering quality than state-of-the-art solutions like instant-NGP while using only a fraction of the parameters. HollowNeRF mitigates hash collisions by a simple feature pruning mechanism. Unlike existing methods that rely on explicit surface geometries for feature pruning, HollowNeRF learns a 3D saliency grid to guide feature compression during training, and employs an ADMM pruner to enforce a sparse feature domain. Our experiments demonstrate that HollowNeRF achieves a better balance between cost and accuracy than state-of-the-art solutions, making a solid step towards lightweight and ubiquitous NeRF applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3488" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, October 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[3] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.353 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[5] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6124-6132, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[7] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.619 + ], + "angle": 0, + "content": "[8] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[9] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[10] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[12] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[14] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, June 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[17] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[18] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 40(4), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[19] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.598, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[20] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.656, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[21] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[22] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[23] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, June 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3489" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[25] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[26] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[27] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12922–12931, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[28] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, June 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.361, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[29] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.432, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.502, + 0.469, + 0.57 + ], + "angle": 0, + "content": "[31] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.572, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[32] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.657, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.7, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[34] Chengxuan Zhu, Renjie Wan, and Boxin Shi. Neural transmitted radiance fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 9" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3490" + } + ] +] \ No newline at end of file diff --git a/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_origin.pdf b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c0c5f8800f50a7bdb07c84b6ae821e77eebe795 --- /dev/null +++ b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/1b4a3dc1-627b-4861-a9e0-292ebe468c19_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f20488bc30532212ef0e397c42eb521eeee2ef73ac47f11a3409cef3dffd97fc +size 2701287 diff --git a/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/full.md b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e5d21394d3280799041aedba6fcce8cbabf63070 --- /dev/null +++ b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/full.md @@ -0,0 +1,282 @@ +# HollowNeRF: Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation + +Xiufeng Xie, Riccardo Gherardi, Zhihong Pan, Stephen Huang + +Oppo Mobile Telecommunications Corp. + +2479 E Bayshore Rd, Palo Alto, CA, 94303, USA + +xxie28@gmail.com + +# Abstract + +Neural radiance fields (NeRF) have garnered significant attention, with recent works such as Instant-NGP accelerating NeRF training and evaluation through a combination of hashgrid-based positional encoding and neural networks. However, effectively leveraging the spatial sparsity of 3D scenes remains a challenge. To cull away unnecessary regions of the feature grid, existing solutions rely on prior knowledge of object shape or periodically estimate object shape during training by repeated model evaluations, which are costly and wasteful. To address this issue, we propose HollowNeRF, a novel compression solution for hashgrid-based NeRF which automatically sparsifies the feature grid during the training phase. Instead of directly compressing dense features, HollowNeRF trains a coarse 3D saliency mask that guides efficient feature pruning, and employs an alternating direction method of multipliers (ADMM) pruner to sparsify the 3D saliency mask during training. By exploiting the sparsity in the 3D scene to redistribute hash collisions, HollowNeRF improves rendering quality while using a fraction of the parameters of comparable state-of-the-art solutions, leading to a better cost-accuracy trade-off. Our method delivers comparable rendering quality to Instant-NGP, while utilizing just $31\%$ of the parameters. In addition, our solution can achieve a PSNR accuracy gain of up to 1dB using only $56\%$ of the parameters. + +# 1. Introduction + +Neural Radiance Fields (NeRF [1, 15]) have gained widespread recognition across academia and industry due to their remarkable capability to generate photorealistic novel views of 3D scenes from a collection of 2D images. Inspired by the volumetric representation, NeRF models the scene as a continuous 5D plenoptic function, enabling the creation of high-fidelity renderings with accurate lighting and shading effects. This technique has found versatile ap + +![](images/ff438e54c8660f1313d917272bc057f83e8b0dbac6b448ddd4515490aab21663.jpg) +Figure 1: Cross-section views of rendered 3D scenes. The left column shows the hollow interior of HollowNeRF's rendering, while the middle column shows the solid interior of Instant-NGP's rendering. The right column displays HollowNeRF's rendering without slicing. + +plications in fields such as computer graphics, virtual and augmented reality, as well as robotics. + +Training and evaluating NeRF models can be computationally expensive, and many recent works [20, 16, 4, 22, 18, 23] have focused on improving the efficiency of NeRF. Instant-NGP [16] is an established solution with state-of-the-art training speed. It employs a lightweight hashgrid for input encoding and a small multi-layer perceptron (MLP) to disambiguate hash collisions. However, the lightweight hashgrid unavoidably causes severe collisions at fine resolutions. These collisions are evenly scattered across the occupied voxels, resulting in a suboptimal accuracy. Another line of research focuses on accelerating NeRF rendering by only sampling near the surface of interest, but has to rely on prior knowledge of surface geometries (either from conventional algorithms such as shape-from-silhouette [13] or an MLP predicting the depth distribution along each camera ray [18]). However, a coarse surface estimation degrades + +NeRF rendering quality, while a precise surface estimation adds too much complexity, defeating the purpose of acceleration. + +We propose HollowNeRF, a novel NeRF compression method using trainable hash collision mitigation to improve rendering accuracy while consuming less parameters than existing NeRF methods. Built on a hash-based pipeline from Instant-NGP, HollowNeRF prioritizes the important features (of visible voxels) and prunes unnecessary features (of invisible voxels), leading to a redistribution of hash collision probability across the 3D volume. When two voxels direct to the same hash bucket, Instant-NGP shapes the shared feature in this bucket as a mixture of the desired features, which harms the accuracy. In contrast, HollowNeRF steers the shared feature to fit the more important voxel, and prunes the feature of the less important voxel toward 0, reducing interference to features sharing the same bucket. Specifically, when reading the feature of a certain voxel, HollowNeRF further scales the feature by a trainable saliency weight whose value captures the voxel's visibility. To reduce the cost, we divide the 3D space into coarse grid regions and assign a saliency weight to capture each region's visibility, forming a trainable 3D saliency grid. Unlike existing methods [13, 18] that require prior knowledge of the surface geometries, HollowNeRF learns to prioritize the important features by training the saliency grid, which converges to a "hollow" saliency distribution across the 3D volume. Figure 1 showcases this "hollow" rendering result. + +The proposed design consists of three main components: a trainable 3D saliency grid to guide the compression of dense features ( $\S 3.1$ ); a soft zero-skipping gate that enhances the MLP in the NeRF model to ensure a feature compressed to 0 translates to a zero density in the 3D space ( $\S 3.2$ ); a pruner to further push unnecessary features to exact 0 instead of a small non-zero value by alternating direction method of multipliers or ADMM [3] ( $\S 3.3$ ). Our experiments demonstrate that HollowNeRF achieves better accuracy (PSNR and LPIPS) than state-of-the-art methods while using significantly fewer parameters. + +The key contributions of this work are: + +- We propose a novel NeRF compression solution, HollowNeRF, that learns to prioritize features defining the visible surface and prune invisible internal features without prior knowledge of surface geometries. +- We use an ADMM-based optimization framework to prune unnecessary features during NeRF training and enhance the MLP with a soft zero-skipping gate to ensure that pruned features correctly map to zero density. +- We evaluate the performance of HollowNeRF on popular NeRF datasets and our solution demonstrates a significantly superior balance between cost and accuracy than state-of-the-art solutions. + +# 2. Background and related work + +Differentiable rendering [15, 7, 5, 22, 17] has emerged as a prominent alternative framework for novel view synthesis, alongside the conventional 3D rendering approaches. Following the seminal NeRF paper [15], the research community has matured the approach and enables it to handle arbitrarily large scenes [24, 29, 27, 2], complex reflections [8, 34, 21, 28, 14], and small training datasets with a limited number of views [10, 6, 19]. Recent works also substantially improve the efficiency of NeRF, reducing the run-time complexity at training [16, 22] and inference [16, 12, 18, 30, 32, 31, 9]. + +Comparatively fewer works have tackled the problem of maximizing visual fidelity for a given space complexity budget [16, 26], which is the focus of this paper. The original NeRF technique [15] encodes the entire 3D scene in the weights of an MLP that predicts opacity $(\sigma)$ and color $(r,g,b)$ given a 3D position $(x,y,z)$ and direction $(\theta ,\phi)$ . While spatial complexity (438K parameters) is not the primary bottleneck, the substantial limitation of NeRF lies in the remarkably slow training and inference speeds. This is primarily attributed to the deep MLP with 8 hidden layers, each having a width of 256 neurons. It has since been shown [22] that differentiable rendering does not necessarily require neural networks. In [22], a scene is encoded as a neuron-free sparse voxel grid of opacities and RGB spherical harmonics. Without the MLP, training and inference run fast, at the cost of a significantly high spatial complexity. + +In the continuum between the two extremes discussed above, there are techniques that harness the strengths of both ends and fare better in the balance between training speed and spatial complexity. Instant-NGP [16] is a representative example, which encodes features using a multi-resolution hashgrid and then feed the features into a lightweight MLP with only 2 hidden layers to decode the color and density; both the features in the hashgrid and the MLP are trained in conjunction. Our work extends the hash-based pipeline from Instant-NGP by achieving higher quality (PSNR & LPIPS) with less number of parameters. We achieve this goal by reclaiming the resources spent on empty, invisible, or internal regions. Instead of using an auxiliary MLP to predict space occupancy like DONeRF [18], we learn it through a trainable lightweight volumetric saliency grid. + +Some existing literature has applied model compression techniques to NeRF: CC-NeRF [26] uses tensor decomposition to obtain a low-rank approximation of the learned network. Similarly, TensoRF [4] represents the volume as a 4D tensor and factorizes it into low-rank components. Unlike these compression works, HollowNeRF vets information based on its actual impact on the rendering accuracy, rather than raw entropy. The relative performance of these approaches are investigated in §4.4. + +![](images/3ca2d7e5c5b0f5261f90d2957e8bf5d044998ade43110992e4a123fdd358d1c9.jpg) +Figure 2: Overview of the HollowNeRF workflow. + +# 3. HollowNeRF design + +This section presents the system design of HollowNeRF, as outlined in Figure 2: (i) Given an input coordinate $\boldsymbol{x} = (x,y,z)$ , we fetch the corresponding feature $\boldsymbol{f}$ from the multi-level hashgrid, following the Instant-NGP method [16]. (ii) We predict the saliency weight $p$ of position $\boldsymbol{x}$ by processing the information from a trainable 3D saliency grid $\mathcal{G}$ through trilinear interpolation and a sigmoid function (§3.1), and use $p$ to scale the feature $\boldsymbol{f}$ by $\boldsymbol{v} = p\boldsymbol{f}$ . Figure 3a showcases a trained saliency grid $\mathcal{G}$ , demonstrating that training $\mathcal{G}$ can mitigate hash collisions by suppressing the unnecessary features in empty or invisible regions. (iii) The weighted feature $\boldsymbol{v}$ is fed into an MLP, which decodes $\boldsymbol{v}$ to obtain the density $\sigma$ and color at position $\boldsymbol{x}$ . To ensure that $\boldsymbol{v} = \mathbf{0}$ maps to $\sigma = 0$ , we introduce a zero-skipping gate (§3.2). (iv) The density and color outputs of the MLP are used for volume rendering, and the resulting image is compared to the ground truth to obtain a loss function $L$ . During training, an ADMM pruner (§3.3) sparsifies the saliency value distribution across the 3D space by enforcing a sparsity constraint when optimizing the loss $L$ . Figure 3a shows that a large portion of the unnecessary features are not entirely eliminated by training the saliency grid, with saliency weights diminished but not reaching 0. Therefore, we introduce the ADMM pruner to explicitly prune saliency weights of the empty or invisible regions to 0, as shown in Figure 3b. + +# 3.1. Trainable 3D saliency grid + +In typical 3D scenes without large transparent objects, most regions of the space do not contribute to the final rendering as they are either empty or invisible from any view angle. To leverage this sparsity, methods such as Instant-NGP use a coarse-grained binary mask to capture space occupancy, which is periodically updated during training by evaluating the NeRF model at each voxel and checking if the density is + +![](images/54dd887c1bdb73629434dc99e8034bb76592f0a2f4b4d4af68209341634de09a.jpg) +(a) Without ADMM pruner. + +![](images/252d8b4c9f4d1fba89113877adbbe843346d1f634954c62beb788dd92b424405.jpg) +(b) With ADMM pruner. + +![](images/ca227756aa46323dccd8640169d8ee11dc735ca070c41aaec0b287913d16a594.jpg) +Figure 3: Slices of a $64 \times 64 \times 64$ saliency grid trained on the "Chair" scene from the NeRF synthetic dataset[15]. + +![](images/31f42bb1348737bb4fdbda2a83573ccdc7c685d8eb4279f1099b658e0b4869d7.jpg) + +above a threshold. This occupancy mask guides sampling in ray marching by skipping unoccupied regions. However, it still keeps unnecessary features in internal occluded regions that have non-zero densities but make no contributions to the visible surfaces from any view angle, as shown in Figure 1. + +With this insight, we propose to reclaim the capacity spent on internal regions by learning the sparsity pattern through training, effectively making objects hollow. The benefits of this approach are twofold: it reduces space complexity by requiring fewer features, and improves accuracy by reducing hash collisions. However, precisely locating the invisible regions of objects can be challenging without any prior knowledge of the object's shape. + +To addresses this challenge, our approach utilizes the hashgrid-based positional encoding from Instant-NGP [16] and extends it by introducing a trainable saliency grid $\mathcal{G}$ with a resolution of $T\times T\times T$ . Specifically, HollowNeRF trains a 3D tensor data structure $\mathcal{G}$ alongside the NeRF model (feature hashgrid $\mathcal{H}$ and MLP weights $\mathcal{W}$ ), which gradually learns which regions to prioritize, such as a nontransparent object surface. We split the space coarsely as $T\times T\times T$ uniform grids and the 3D tensor $\mathcal{G}$ stores train + +able values representing the saliency of features near each grid voxel. Instead of letting multiple dense features in the same coarse grid share one value, we interpolate the values associated to the 8 voxels surrounding the input coordinate $\pmb{x}$ to get a smoothed saliency distribution $\mathcal{P}$ over the 3D space. The trainable saliency grid represents in itself an opportunity to compress the NeRF model using differentiable model compression techniques, as discussed later in §3.3. + +We initialize $\mathcal{G}$ as an all-ones tensor, where $T$ is typically much smaller than the resolution of the dense feature grid (In §4.2, we further investigate how different $T$ values affect the performance). Given a 3D coordinate $\pmb{x} = (x,y,z)$ , we fetch its corresponding feature vector $\pmb{f}$ from the multi-level hashgrid [16]. We then locate the 8 voxels $(\pmb{x}_1,\pmb{x}_2,\dots,\pmb{x}_8)$ surrounding $\pmb{x}$ in the 3D saliency grid $\mathcal{G}$ , and fetch its corresponding saliency value $g_i$ . Specifically, the coordinate $\pmb{x}_i = (x_i,y_i,z_i)$ are quantized to the values $(\hat{x}_i,\hat{y}_i,\hat{z}_i)$ between 0 and $T - 1$ , which are then used to index the saliency grid $\mathcal{G}$ . This gives us the saliency value $g_i = \mathcal{G}(\hat{x}_i,\hat{y}_i,\hat{z}_i)$ . After obtaining $g_i$ for each of the 8 saliency grid voxels surrounding $\pmb{x}$ , we calculate the saliency weight $p$ at coordinate $\pmb{x}$ by trilinear interpolation followed by a sigmoid operator, so that $0 < p < 1$ . The final weighted feature vector $\pmb{v}$ is obtained by multiplying the saliency weight $p$ with the feature vector $\pmb{f}$ fetched from the hashgrid, i.e., $\pmb{v} = p\pmb{f}$ . + +For mathematical analysis, we define the saliency distribution $\mathcal{P}$ as the collection of saliency weights $p$ at every possible coordinate $\boldsymbol{x}$ , such that $p = \mathcal{P}(\boldsymbol{x})$ . Our proposed method, HollowNeRF, learns the saliency distribution $\mathcal{P}$ by training the saliency grid $\mathcal{G}$ using gradient descent. It is worth noting that we inject $\mathcal{P}$ (and thus $\mathcal{G}$ ) to the training pipeline by feeding the weighted feature $\boldsymbol{v}$ instead of the raw feature $\boldsymbol{f}$ to the MLP for decoding; the rationale for this choice is that since the sigmoid function in the above workflow projects the saliency weights to the range $(0,1)$ , we can view the 3D saliency distribution as a probability distribution $\mathcal{P}(\boldsymbol{x})$ indicating the likelihood $p$ that a feature is non-zero at a given 3D position $\boldsymbol{x}$ . There are only two possibilities for the feature at a given position: either the position is visible with probability $p = \mathcal{P}(\boldsymbol{x})$ , and we retrieve the feature vector $\boldsymbol{f}$ from the multi-level hashgrid as in [16]; otherwise the position is empty or occluded with probability $1 - p$ , and the corresponding feature vector is an all-zero vector $\mathbf{0}$ . Taking both cases into account, we can calculate the expectation $\boldsymbol{v}$ of the feature as: + +$$ +\boldsymbol {v} = \mathcal {P} (\boldsymbol {x}) \boldsymbol {f} + (1 - \mathcal {P} (\boldsymbol {x})) \boldsymbol {0} = p \boldsymbol {f} +$$ + +Thus the weighted feature $\pmb{v}$ fed to MLP decoder has a physical meaning, namely the expectation of the feature at $\pmb{x}$ being either $\pmb{f}$ fetched from feature hashgrid or 0. Then training the saliency grid is equivalent to learning the probability distribution $\mathcal{P}(\pmb{x})$ . This probability distribution represents the likelihood that a given voxel $\pmb{x}$ contains a non-zero fea + +ture. A feature with a lower saliency weight $p$ is more likely to be 0 and will cause less interference to other features assigned to the same bucket in the feature hashgrid. When the saliency weight converges to 0, that feature is pruned with a probability of 1 and becomes a zero vector 0. + +Training the saliency distribution can also be viewed as hash collision mitigation. Instant-NGP employs hash encoding and distributes hash collisions evenly across voxels with non-zero density. By learning to shuffle hash collisions across the space, HollowNeRF guarantees less interference for more important features. When the feature of a voxel is scaled by a higher saliency weight, it tends to accumulate more gradients during backward propagation, and the information at this voxel will dominate the trained feature value inside the hash bucket shared with other voxels. On the other hand, a feature scaled by zero saliency weight has a zero gradient, hence causes no harm to colliding features. + +# 3.2. Soft zero-skipping gate + +While HollowNeRF mitigates hash collisions by pruning unnecessary features to zero, the MLP used to decode these features into densities can break the sparsity in the 3D space: the MLP can be viewed as a function $\sigma = \mathcal{M}(\boldsymbol{v})$ mapping the feature vector $\boldsymbol{v}$ to the density $\sigma$ , and typically $\mathcal{M}(\mathbf{0}) \neq 0$ . Even though the saliency grid and ADMM pruner ensure most voxels direct to $\mathbf{0}$ features in the hash-grid, a typical MLP translates such a sparse feature domain to a non-sparse 3D space with non-zero densities scattered everywhere, which is evidently inaccurate. To enforce a 0 output density when the input feature is $\mathbf{0}$ , we introduce a zero-skipping gate $g(\boldsymbol{v})$ to the MLP decoder: + +$$ +\tilde {\mathcal {M}} (\boldsymbol {v}) = \hat {g} (\boldsymbol {v}) \cdot \mathcal {M} (\boldsymbol {v}) +$$ + +$$ +\text {w h e r e} \quad \hat {g} (\boldsymbol {v}) := \tanh (\alpha \| \boldsymbol {v} \| _ {2}) \tag {1} +$$ + +The gate $\hat{g}(\pmb{v})$ is a differentiable approximation of the "hard" gate in Eq. (2) to enforce the desired property $\hat{\mathcal{M}}(\mathbf{0}) = 0$ , and we use a constant $\alpha$ to control the steepness of the 0-to-1 transition. + +$$ +g (\boldsymbol {v}) := \left\{ \begin{array}{l l} 0, & \text {i f} \| \boldsymbol {v} \| _ {2} = 0 \\ 1, & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +The non-differentiable "hard" gate in Eq. (2) may cause abrupt density value changes during training. Also, the gradient $\frac{\partial g}{\partial p}$ of $g(\boldsymbol{v}) = g(p\boldsymbol{f})$ is 0 with $p \neq 0$ , this 0 gradient prevents the ADMM pruner from suppressing a small saliency value $p$ to 0, resulting in insufficient feature pruning and little improvement in hash collision. Our experiments in §4.3 confirm that adding the "hard" gate $g(\boldsymbol{v})$ provides little gain over adding no gate. + +Thus, we propose a "soft" zero-skipping gate $\hat{g}(\pmb{v})$ in Eq. (1) that smoothly reduces the density output to zero for input features being pruned. As an approximation of $g(\pmb{v})$ , + +the value of $\hat{g}(\boldsymbol{v})$ should smoothly transition from 0 to 1 for input $\boldsymbol{v}$ with small magnitudes $\|\boldsymbol{v}\|_2$ , while otherwise staying close to 1 to avoid perturbing necessary information. The tanh function is a suitable choice. In our implementation we adjust the $\alpha$ value with a schedule, to gradually "harden" the soft gate as the training progresses. Specifically, for epochs below 1000, we set $\alpha$ to $10^4$ to ensure fast convergence. Afterward we increase $\alpha$ to $10^5$ to minimize the perturbation to the fine-tuning process. + +# 3.3. ADMM pruner + +Training the 3D saliency grid improves the accuracy by redistributing hash collisions, which, However, cannot fully eliminate unnecessary features, as shown in Figure 3a. To address this, we introduce an ADMM pruning that enforces sparsity in the saliency grid $\mathcal{G}$ during training. + +We first motivate the use of ADMM pruner rigorously. Assuming a $K \times K \times K$ volume where each voxel $\pmb{x}$ corresponds to a feature $\pmb{f}$ , and the ground truth feature $\pmb{f}_1$ at voxel $\pmb{x}_1 = (x_1, y_1, z_1)$ and $\pmb{f}_2$ at a different voxel $\pmb{x}_2 = (x_2, y_2, z_2)$ are typically different for a 3D scene. Instant-NGP encodes this $K \times K \times K$ feature grid into a small hashgrid $\mathcal{H}$ whose size $<< (K \times K \times K)$ . When $\pmb{f}_1$ and $\pmb{f}_2$ collide with each other by directing to the same hash bucket, their values queried from the hashgrid become identical $\pmb{f}_1^* = \pmb{f}_2^* = \pmb{f}_H$ , where $\pmb{f}_H$ is the trained feature stored in the shared hash bucket. Since the ground truth features $\pmb{f}_1 \neq \pmb{f}_2$ while the queried features $\pmb{f}_1^* = \pmb{f}_2^*$ , the queried values differ from the ground truth, causing compression artifacts. To solve this problem, HollowNeRF allows the colliding features to be different by introducing the scalar saliency weight $p$ to scale the queried features: $\pmb{f}_1^* = p(\pmb{x}_1)\pmb{f}_H$ and $\pmb{f}_2^* = p(\pmb{x}_2)\pmb{f}_H$ . However, during training, the share feature $\pmb{f}_H$ may not find a value that satisfies both equations, as they can be written as $\pmb{f}_1^* = \frac{p(\pmb{x}_1)}{p(\pmb{x}_2)}\pmb{f}_2^*$ which is not always true especially when $\pmb{f}_1$ and $\pmb{f}_2$ are nonzero. Then the queried features $\pmb{f}_1^*$ and $\pmb{f}_2^*$ may still cause compression artifacts (although less than instant-NGP). The ADMM pruner addresses this challenge by pruning the ground truth $\pmb{f}_1$ or $\pmb{f}_2$ to 0 when $\pmb{x}_1$ or $\pmb{x}_2$ is empty/invisible. For example, if voxel $\pmb{x}_1$ is invisible from any view angle $(\pmb{f}_1 = 0)$ , there exists a single $\pmb{f}_H$ value to satisfy both $\pmb{f}_1 = p(\pmb{x}_1)\pmb{f}_H$ and $\pmb{f}_2 = p(\pmb{x}_2)\pmb{f}_H$ , that is $\pmb{f}_H = \pmb{f}_2$ , and there is no compression artifacts. The optimizer can converge to this optimal $\pmb{f}_H$ value through gradient descent by learning that $p(\pmb{x}_1)$ is 0. In summary, making the feature grid sparse reduces the compression artifacts when encoding features into a hashgrid. + +In what follows, we describe two methods for achieving sparsity: a basic L1-regularization approach and our ADMM pruner. To sparsify the saliency grid $\mathcal{G}$ , one simple approach is to add an L1 regularization term to the original + +MSE loss function $L$ typically used in NeRF training: + +$$ +\min _ {\mathcal {W}, \mathcal {H}, \mathcal {G}} \left(L + \lambda \| S _ {\text {s i g}} (\mathcal {G}) \| _ {1}\right) \tag {3} +$$ + +where $\lambda$ is a constant weight to control the amount of sparsity and compression, and $\mathcal{H}$ and $\mathcal{W}$ denote respectively the embeddings in the multi-level hashgrid and the weights in the MLP. We use the sigmoid operator $S_{\mathrm{sig}}(\cdot)$ here because the forward pipeline (Figure 2) uses the saliency value after the sigmoid operator to weight the feature vector, and our ultimate goal is a sparse distribution of saliency value to prune unnecessary features. Note that $S_{\mathrm{sig}}(\mathcal{G})$ here denotes applying sigmoid operator on each element $g$ of the 3D tensor $\mathcal{G}$ , and the result is a 3D tensor with all non-negative elements. Therefore, the absolute operator in the L1 regularization term can be removed: + +$$ +\| S _ {\operatorname {s i g}} (\mathcal {G}) \| _ {1} = \sum_ {g \in \mathcal {G}} | S _ {\operatorname {s i g}} (g) | = \sum_ {g \in \mathcal {G}} S _ {\operatorname {s i g}} (g) +$$ + +Removing the absolute operator from the regularization term makes it differentiable and compatible with the training pipeline, eliminating the need to use more complex optimization algorithms such as proximal gradient descent to handle non-differentiable regularization terms. + +Although the usage of L1 regularization is already a marked improved over the baseline (see §4.3), we find it is often difficult to select a single $\lambda$ hyperparameter to control sparsity and compression. During training when the 3D saliency grid $\mathcal{G}$ has already become sparse enough, the regularization term $\lambda \| S_{\mathrm{sig}}(\mathcal{G})\|_1$ may still provide gradients that further prune $\mathcal{G}$ . If $\lambda$ is set too high or if the algorithm is let to run for too long, visible features may be pruned resulting in performance degradation. + +We use an ADMM pruner to overcome the limitations of simple L1 regularization. By introducing a configurable sparsity constraint $C$ such that $\| S_{\mathrm{sig}}(\mathcal{G})\|_1 < C$ , feature pruning becomes the constrained optimization problem: + +$$ +\begin{array}{l} \min _ {\mathcal {M}, \mathcal {H}, \mathcal {G}} L \left(\mathcal {M}, \mathcal {H}, \mathcal {G}\right) \\ \begin{array}{l} \text {s . t .} \quad \| S _ {\text {s i g}} (\mathcal {G}) \| _ {1} < C \end{array} \\ \end{array} +$$ + +We transform this constrained optimization problem into an equivalent unconstrained minimax problem in Problem (5) by using the augmented Lagrangian method which employs a trainable dual variable $\gamma$ , also known as a Lagrange multiplier. In essence, $\gamma$ serves as a trainable alternative to the fixed hyperparameter $\lambda$ in Problem (3). + +$$ +\begin{array}{l} \min _ {\mathcal {W}, \mathcal {H}, \mathcal {G}} \max _ {\gamma \geq 0} L \left(\mathcal {W}, \mathcal {H}, \mathcal {G}\right) + \frac {\rho_ {\gamma}}{2} \left[ \| S _ {\text {s i g}} (\mathcal {G}) \| _ {1} - C \right] _ {+} ^ {2} \\ + \gamma \left(\left\| S _ {\text {s i g}} (\mathcal {G}) \right\| _ {1} - C\right) \tag {5} \\ \end{array} +$$ + +
MethodHashgrid size214215216217218219
Param #
Instant-NGP0.50M0.94M1.77M3.34M6.22M11.49M
HollowNeRFT=640.76M1.21M2.04M3.60M6.48M11.75M
T=961.39M1.83M2.66M4.22M7.10M12.37M
T=1282.60M3.04M3.87M5.43M8.32M13.58M
+ +Table 1: The total number of model parameters, depending on the hashgrid size and saliency grid resolution $T$ . This number counts the parameters in the hashgrid, MLP, and saliency grid. (both HollowNeRF and Instant-NGP employ a multi-resolution hashgrid with 16 levels. We use the maximum entries per level to denote the hashgrid size, following the notation in the Instant-NGP paper [16].) + +where $\rho_{\gamma}$ denotes the learning rate of dual variable $\gamma$ and $[\cdot ]_{+}$ clamps the input to non-negative values. + +At each iteration $t$ , the optimizer first updates the coarse saliency grid $\mathcal{G}$ , the feature hashgrid $\mathcal{H}$ , and the MLP weights $\mathcal{W}$ to minimize the augmented loss function in Problem (5) using gradient descent, where the dual variable $\gamma$ is treated as a constant. $\gamma$ is then updated using gradient ascent with $\gamma^{(t + 1)} = \left[\gamma^{(t)} + \rho_{\gamma}\left(\| S_{\mathrm{sig}}(\mathcal{G})\| _1 - C\right)\right]_+$ + +# 4. Evaluation + +In this section, we first investigate the effect of different hyper-parameters, including hashgrid sizes, saliency grid sizes, and sparsity constraints, on the performance (§4.2). We then conduct an ablation study to validate each of HollowNeRF's design components (§4.3). Finally, we evaluate the performance of HollowNeRF on various 3D scenes and compare it with the state-of-the-art works (§4.4). + +# 4.1. Experimental setup + +We implement HollowNeRF on top of torch-ngp [25, 26], a PyTorch CUDA extension implementation of instant-NGP [16]. To ensure a fair comparison, we evaluated both HollowNeRF and instant-NGP using the same torch-ngp implementation. Both methods use a 16-level hashgrid with a feature dimension of 2, where the coarsest resolution of the feature grid is 16, and the finest resolution is 1024. Both methods employ a MLP decoder that has only two hidden layers with a width of 64 neurons, while HollowNeRF augments the MLP with a zero-skipping gate as described in §3.2. + +We use the peak signal-to-noise ratio (PSNR) and the Learned Perceptual Image Patch Similarity (LPIPS) [33] with the default AlexNet backend as the performance metrics to assess the rendering quality. Each experiments was conducted on a single NVIDIA Tesla A100 GPU for both training and evaluation. The machine used was equipped with an AMD EPYC 7742 64-Core CPU, 1TB of memory, and 8 GPU cards in total. + +![](images/fec9c0c29f1d0be494664c7383dcc15824fbd1ac9f6f9e7d0d7b93b1b8fef20d.jpg) + +![](images/f1315eb5158ec21a3c15de2ad3ecb11f5c8b4f4362c69cb00ea0cf85fa043166.jpg) +(a) PSNR $\uparrow$ (higher is better). +(b) LPIPS $\downarrow$ (lower is better). +Figure 4: Performance vs. saliency grid size $T$ . + +Unless otherwise specified, we train the model for 300,000 steps using the Adam optimizer with an initial learning rate of $1 \times 10^{-2}$ . Further details regarding the appropriate selection of HollowNeRF hyper-parameters, such as $T$ and $C$ , are discussed in §4.2. + +# 4.2. Investigating the cost-accuracy tradeoff + +We first investigate the cost-accuracy tradeoff of HollowNeRF to determine the appropriate hyper-parameter config + +![](images/9dbaf5eb65d4f1502ae0024b024709fbd374d2e26d68348c924e71c2aa26a471.jpg) +(a) PSNR $\uparrow$ (higher is better). + +![](images/efa6804fc207a5453459d80664d9ac47fb1f60fe32017b609be1af4d890a4651.jpg) +(b) LPIPS $\downarrow$ (lower is better). + +uration for larger scale experiments in §4.4. We conducted experiments on the chair scene from the NeRF synthetic dataset, varying the hashgrid size from $2^{14}$ to $2^{19}$ and testing the saliency grid resolution of 64, 96, and 128 for each hashgrid size. Table 1 summarizes the cost of different configurations, including the number of parameters in the hashgrid, MLP, and saliency grid. We use a sparsity constraint of $C = 0.04$ in this experiment. + +The results in Figure 4 indicate that increasing the saliency grid resolution $T$ from 64 to 96 or 128 does not improve accuracy much but results in a higher parameter count, which harms the cost-accuracy tradeoff. On the other hand, lowering $T$ below 64 causes unstable convergence. Consequently, we use a grid resolution of $T = 64$ for the subsequent experiments. Figure 4 further shows that HollowNeRF outperforms the baseline over a wide range of model sizes, achieving higher PSNR and lower LPIPS values for comparable parameter counts. The best overall model, HollowNeRF with $T = 64$ , uses 6.48M parameters to attain a 34.85dB PSNR, as compared to Instant-NGP's 33.86dB using 11.49M. In other words, it achieves about 1dB higher PSNR than Instant-NGP, while utilizing only 56% of the parameters employed by Instant-NGP. + +![](images/ac5e9e6d256221fc1eec60364db84c60553091efd7f38c5c44a1d72d53f3f9b0.jpg) +(a) PSNR $\uparrow$ (higher is better). + +![](images/f841c4e3717ef361fabb2406887c4c495123d69e3d776b763d6b00e2328204b0.jpg) +Figure 5: Performance vs. sparsity constraint $C$ . +(b) LPIPS $\downarrow$ (lower is better). +Figure 6: Ablation study for the design components. + +We also repeated the training and evaluation for each hashgrid size five times for HollowNeRF with $T = 64$ and instant-NGP, and plotted the standard deviations across the five tests as error bars in Figure 4. The results demonstrate that our approach is more stable than Instant-NGP in terms of convergence, with lower performance variation. + +To investigate the impact of the choice of sparsity constraint $C$ on performance, we conducted additional experiments following the same setup as previous tests, while varying the $C$ values, and report the results in Figure 5. First, from Figure 5a, we observe that the $C = 0.04$ configuration yields the highest PSNR, in other words, HollowNeRF achieves the highest accuracy when $4\%$ of the spatial voxels are non-zero. $C = 0.02$ exhibits slightly diminished accuracy, potentially due to excessive pruning in the 3D space, which compromises crucial information for scene reconstruction. Conversely, $C = 0.08$ shows the lowest accuracy across all tested HollowNeRF configurations, suggesting that keeping $8\%$ non-zero voxels causes excessive hash collisions when packing them into the hashgrid, and the 3D scene to capture could be sparser. We have similar observation for LPIPS, as depicted in Figure 5b. + +![](images/b398dc24f851b638ba4004a1fc9aa6645146ea1857ac019ae983e8a3e005e2c1.jpg) +Figure 7: Qualitative comparison with instant-NGP over scenes from the NeRF synthetic dataset. + +![](images/c6a24caaa983c0d5743c48e5622edb063d25b6cfb296aa9a1ec8541d200ed7ca.jpg) + +![](images/e9202d24c5e0f9f79f255c311d39ff25a9fa89ae4c57c33d8d9caec8df9d9adf.jpg) +Figure 8: Comparison of PSNR(↑) and LPIPS (↓) performance between HollowNeRF and Instant-NGP on various scenes. + +![](images/c8d0aca4088252db84726f814d567a6c69f22ce6b9963f4b6ef9d9116c773c86.jpg) +Figure 9: Comparison of HollowNeRF and recent methods in terms of the trade-off between PSNR and model size ("HollowNeRF-H19" means using a hashgrid size of $2^{19}$ ). + +# 4.3. Ablation study + +In this ablation study, we analyze the impact of each component in our solution by comparing the convergence speed and resulting performance of the configurations with and without each design component. We use a saliency grid resolution of 64 and a hashgrid size of $2^{18}$ (total 6.48M parameters) for all HollowNeRF configurations, while the baseline is instant-NGP with a hashgrid size of $2^{19}$ (total 11.49M parameters). To investigate how each component affects the convergence speed, we evaluate the performance on the test sets every 300 epochs during the total 3000 epochs. + +The results presented in Figure 6 show that each design component presence gracefully increases the overall performance, without any catastrophic failure. Each of the HollowNeRF ablated configurations still outperforms instant-NGP, while employing less parameters. + +# 4.4. Comparisons on standardized datasets + +Finally, we evaluate the performance of HollowNeRF on 3D scenes from the NeRF synthetic dataset [15] and the more complicated Tanks and Temples dataset [11]. Based on the cost-accuracy tradeoff investigated in Sec. 4.2, we set HollowNeRF's hyper-parameters to a sparsity constraint of $C = 0.04$ and a saliency grid resolution of $T = 64$ . + +We first provide qualitative example views comparing HollowNeRF with Instant-NGP. The distinct advantages of HollowNeRF are evident through the visuals showcased in Figure 7. For example, from the lego scene in column 1, we observe that HollowNeRF with 6.48M parameters achieves notably superior rendering accuracy over Instant-NGP with $77\%$ more parameters (11.49M). The advantage is particularly pronounced within the bulldozer track region. Furthermore, HollowNeRF with 11.75M parameters exhibits + +the highest accuracy, capturing fine details such as the tiny indentation on the edge of the black rod located at the left border of the view. + +Then we show the PSNR and LPIPS performance of HollowNeRF and Instant-NGP on various 3D scenes (we sample 10 different scenes due to the page limit) in Figure 8. For each scene, we test three hashgrid sizes $(2^{17}, 2^{18}, 2^{19})$ for both methods. We can observe that HollowNeRF consistently outperforms Instant-NGP on all the 10 scenes tested, achieving better PSNR and LPIPS performance with fewer parameters when using the same hashgrid size. For seven out of ten tested scenes, the HollowNeRF with the smallest hashgrid size $(2^{17}$ , for a total parameter count including MLP of $3.34\mathrm{M}$ ) achieves higher PSNR than the best Instant-NGP model (hashgrid size of $2^{19}$ , corresponding to $11.49\mathrm{M}$ parameters including the MLP). + +Finally, we compare the average PSNR across the NeRF synthetic dataset [16] and the corresponding model size for HollowNeRF and several recent works. Figure 9 shows that HollowNeRF achieves the best cost-accuracy tradeoff among all tested methods. In particular, HollowNeRF with a hashgrid size of $2^{17}$ achieves an average PSNR of 32.53dB with only a mode size of 14.0MB. + +# 5. Discussion + +The compression gains of HollowNeRF require the 3D scene being sparse, i.e. mostly filled with either empty or occluded space, which holds for typical 3D scenes. HollowNeRF can still operate when this assumption does not hold, such as for scenes containing smoke, fire, clouds, but its performance will regress to the baseline without pruning. + +The current version of HollowNeRF, like its predecessor Instant-NGP, faces challenges in modeling objects with reflective surfaces. In future works, we plan to extend the concept of HollowNeRF to the frameworks [8, 34, 21, 28, 14] with better support for reflective surfaces. + +# 6. Conclusions + +In this paper, we present HollowNeRF, a novel hashgrid-based NeRF technique that achieves superior rendering quality than state-of-the-art solutions like instant-NGP while using only a fraction of the parameters. HollowNeRF mitigates hash collisions by a simple feature pruning mechanism. Unlike existing methods that rely on explicit surface geometries for feature pruning, HollowNeRF learns a 3D saliency grid to guide feature compression during training, and employs an ADMM pruner to enforce a sparse feature domain. Our experiments demonstrate that HollowNeRF achieves a better balance between cost and accuracy than state-of-the-art solutions, making a solid step towards lightweight and ubiquitous NeRF applications. + +# References + +[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, October 2021. 1 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2 +[3] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 2 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 9 +[5] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6124-6132, June 2022. 2 +[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, June 2022. 2 +[7] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 2 +[8] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 2, 9 +[9] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2 +[10] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021. 2, 9 +[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9 +[12] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2 +[13] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021 + +IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 1, 2 +[14] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, June 2022. 2, 9 +[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 9 +[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 4, 6, 9 +[17] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, June 2022. 2 +[18] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 40(4), 2021. 1, 2 +[19] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, June 2022. 2 +[20] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In International Conference on Computer Vision (ICCV), 2021. 1 +[21] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2, 9 +[22] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 9 +[23] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 9 +[24] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, June 2022. 2 + +[25] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp.6 +[26] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6, 9 +[27] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12922–12931, June 2022. 2 +[28] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, June 2022. 2, 9 +[29] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 2 +[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, June 2022. 2 +[31] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 9 +[32] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2 +[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[34] Chengxuan Zhu, Renjie Wan, and Boxin Shi. Neural transmitted radiance fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 9 \ No newline at end of file diff --git a/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/images.zip b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2df5d06677a7135f7bf58297f69cf227a3c823c1 --- /dev/null +++ b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:220d0b15d35f04b7ed3f95885784c03dcbe174fc6bb1a4e78ea89901d1f22244 +size 761091 diff --git a/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/layout.json b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a46d07fe39147a87e692511df8007afabc066564 --- /dev/null +++ b/2023/HollowNeRF_ Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation/layout.json @@ -0,0 +1,10382 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 103, + 442, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 103, + 442, + 139 + ], + "spans": [ + { + "bbox": [ + 153, + 103, + 442, + 139 + ], + "type": "text", + "content": "HollowNeRF: Pruning Hashgrid-Based NeRFs with Trainable Collision Mitigation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 161, + 449, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 161, + 449, + 175 + ], + "spans": [ + { + "bbox": [ + 145, + 161, + 449, + 175 + ], + "type": "text", + "content": "Xiufeng Xie, Riccardo Gherardi, Zhihong Pan, Stephen Huang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 176, + 395, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 176, + 395, + 190 + ], + "spans": [ + { + "bbox": [ + 197, + 176, + 395, + 190 + ], + "type": "text", + "content": "Oppo Mobile Telecommunications Corp." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 190, + 417, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 190, + 417, + 203 + ], + "spans": [ + { + "bbox": [ + 176, + 190, + 417, + 203 + ], + "type": "text", + "content": "2479 E Bayshore Rd, Palo Alto, CA, 94303, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 252, + 206, + 341, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 206, + 341, + 217 + ], + "spans": [ + { + "bbox": [ + 252, + 206, + 341, + 217 + ], + "type": "text", + "content": "xxie28@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 271, + 290, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 290, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 290, + 571 + ], + "type": "text", + "content": "Neural radiance fields (NeRF) have garnered significant attention, with recent works such as Instant-NGP accelerating NeRF training and evaluation through a combination of hashgrid-based positional encoding and neural networks. However, effectively leveraging the spatial sparsity of 3D scenes remains a challenge. To cull away unnecessary regions of the feature grid, existing solutions rely on prior knowledge of object shape or periodically estimate object shape during training by repeated model evaluations, which are costly and wasteful. To address this issue, we propose HollowNeRF, a novel compression solution for hashgrid-based NeRF which automatically sparsifies the feature grid during the training phase. Instead of directly compressing dense features, HollowNeRF trains a coarse 3D saliency mask that guides efficient feature pruning, and employs an alternating direction method of multipliers (ADMM) pruner to sparsify the 3D saliency mask during training. By exploiting the sparsity in the 3D scene to redistribute hash collisions, HollowNeRF improves rendering quality while using a fraction of the parameters of comparable state-of-the-art solutions, leading to a better cost-accuracy trade-off. Our method delivers comparable rendering quality to Instant-NGP, while utilizing just " + }, + { + "bbox": [ + 47, + 271, + 290, + 571 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 47, + 271, + 290, + 571 + ], + "type": "text", + "content": " of the parameters. In addition, our solution can achieve a PSNR accuracy gain of up to 1dB using only " + }, + { + "bbox": [ + 47, + 271, + 290, + 571 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 47, + 271, + 290, + 571 + ], + "type": "text", + "content": " of the parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF [1, 15]) have gained widespread recognition across academia and industry due to their remarkable capability to generate photorealistic novel views of 3D scenes from a collection of 2D images. Inspired by the volumetric representation, NeRF models the scene as a continuous 5D plenoptic function, enabling the creation of high-fidelity renderings with accurate lighting and shading effects. This technique has found versatile ap" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 545, + 399 + ], + "type": "image", + "image_path": "ff438e54c8660f1313d917272bc057f83e8b0dbac6b448ddd4515490aab21663.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 416, + 547, + 477 + ], + "lines": [ + { + "bbox": [ + 304, + 416, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 547, + 477 + ], + "type": "text", + "content": "Figure 1: Cross-section views of rendered 3D scenes. The left column shows the hollow interior of HollowNeRF's rendering, while the middle column shows the solid interior of Instant-NGP's rendering. The right column displays HollowNeRF's rendering without slicing." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 521 + ], + "type": "text", + "content": "plications in fields such as computer graphics, virtual and augmented reality, as well as robotics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "Training and evaluating NeRF models can be computationally expensive, and many recent works [20, 16, 4, 22, 18, 23] have focused on improving the efficiency of NeRF. Instant-NGP [16] is an established solution with state-of-the-art training speed. It employs a lightweight hashgrid for input encoding and a small multi-layer perceptron (MLP) to disambiguate hash collisions. However, the lightweight hashgrid unavoidably causes severe collisions at fine resolutions. These collisions are evenly scattered across the occupied voxels, resulting in a suboptimal accuracy. Another line of research focuses on accelerating NeRF rendering by only sampling near the surface of interest, but has to rely on prior knowledge of surface geometries (either from conventional algorithms such as shape-from-silhouette [13] or an MLP predicting the depth distribution along each camera ray [18]). However, a coarse surface estimation degrades" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3480" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "content": "NeRF rendering quality, while a precise surface estimation adds too much complexity, defeating the purpose of acceleration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 289, + 407 + ], + "type": "text", + "content": "We propose HollowNeRF, a novel NeRF compression method using trainable hash collision mitigation to improve rendering accuracy while consuming less parameters than existing NeRF methods. Built on a hash-based pipeline from Instant-NGP, HollowNeRF prioritizes the important features (of visible voxels) and prunes unnecessary features (of invisible voxels), leading to a redistribution of hash collision probability across the 3D volume. When two voxels direct to the same hash bucket, Instant-NGP shapes the shared feature in this bucket as a mixture of the desired features, which harms the accuracy. In contrast, HollowNeRF steers the shared feature to fit the more important voxel, and prunes the feature of the less important voxel toward 0, reducing interference to features sharing the same bucket. Specifically, when reading the feature of a certain voxel, HollowNeRF further scales the feature by a trainable saliency weight whose value captures the voxel's visibility. To reduce the cost, we divide the 3D space into coarse grid regions and assign a saliency weight to capture each region's visibility, forming a trainable 3D saliency grid. Unlike existing methods [13, 18] that require prior knowledge of the surface geometries, HollowNeRF learns to prioritize the important features by training the saliency grid, which converges to a \"hollow\" saliency distribution across the 3D volume. Figure 1 showcases this \"hollow\" rendering result." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "text", + "content": "The proposed design consists of three main components: a trainable 3D saliency grid to guide the compression of dense features (" + }, + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "inline_equation", + "content": "\\S 3.1" + }, + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "text", + "content": "); a soft zero-skipping gate that enhances the MLP in the NeRF model to ensure a feature compressed to 0 translates to a zero density in the 3D space (" + }, + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "inline_equation", + "content": "\\S 3.2" + }, + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "text", + "content": "); a pruner to further push unnecessary features to exact 0 instead of a small non-zero value by alternating direction method of multipliers or ADMM [3] (" + }, + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "inline_equation", + "content": "\\S 3.3" + }, + { + "bbox": [ + 46, + 407, + 289, + 538 + ], + "type": "text", + "content": "). Our experiments demonstrate that HollowNeRF achieves better accuracy (PSNR and LPIPS) than state-of-the-art methods while using significantly fewer parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 539, + 218, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 539, + 218, + 550 + ], + "spans": [ + { + "bbox": [ + 59, + 539, + 218, + 550 + ], + "type": "text", + "content": "The key contributions of this work are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 556, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 556, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 556, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 58, + 556, + 287, + 604 + ], + "type": "text", + "content": "- We propose a novel NeRF compression solution, HollowNeRF, that learns to prioritize features defining the visible surface and prune invisible internal features without prior knowledge of surface geometries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 611, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 611, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 58, + 611, + 287, + 658 + ], + "type": "text", + "content": "- We use an ADMM-based optimization framework to prune unnecessary features during NeRF training and enhance the MLP with a soft zero-skipping gate to ensure that pruned features correctly map to zero density." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 287, + 712 + ], + "type": "text", + "content": "- We evaluate the performance of HollowNeRF on popular NeRF datasets and our solution demonstrates a significantly superior balance between cost and accuracy than state-of-the-art solutions." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 474, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 474, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 474, + 85 + ], + "type": "text", + "content": "2. Background and related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 222 + ], + "type": "text", + "content": "Differentiable rendering [15, 7, 5, 22, 17] has emerged as a prominent alternative framework for novel view synthesis, alongside the conventional 3D rendering approaches. Following the seminal NeRF paper [15], the research community has matured the approach and enables it to handle arbitrarily large scenes [24, 29, 27, 2], complex reflections [8, 34, 21, 28, 14], and small training datasets with a limited number of views [10, 6, 19]. Recent works also substantially improve the efficiency of NeRF, reducing the run-time complexity at training [16, 22] and inference [16, 12, 18, 30, 32, 31, 9]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "text", + "content": "Comparatively fewer works have tackled the problem of maximizing visual fidelity for a given space complexity budget [16, 26], which is the focus of this paper. The original NeRF technique [15] encodes the entire 3D scene in the weights of an MLP that predicts opacity " + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "inline_equation", + "content": "(r,g,b)" + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "text", + "content": " given a 3D position " + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 304, + 223, + 546, + 414 + ], + "type": "text", + "content": ". While spatial complexity (438K parameters) is not the primary bottleneck, the substantial limitation of NeRF lies in the remarkably slow training and inference speeds. This is primarily attributed to the deep MLP with 8 hidden layers, each having a width of 256 neurons. It has since been shown [22] that differentiable rendering does not necessarily require neural networks. In [22], a scene is encoded as a neuron-free sparse voxel grid of opacities and RGB spherical harmonics. Without the MLP, training and inference run fast, at the cost of a significantly high spatial complexity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 546, + 605 + ], + "type": "text", + "content": "In the continuum between the two extremes discussed above, there are techniques that harness the strengths of both ends and fare better in the balance between training speed and spatial complexity. Instant-NGP [16] is a representative example, which encodes features using a multi-resolution hashgrid and then feed the features into a lightweight MLP with only 2 hidden layers to decode the color and density; both the features in the hashgrid and the MLP are trained in conjunction. Our work extends the hash-based pipeline from Instant-NGP by achieving higher quality (PSNR & LPIPS) with less number of parameters. We achieve this goal by reclaiming the resources spent on empty, invisible, or internal regions. Instead of using an auxiliary MLP to predict space occupancy like DONeRF [18], we learn it through a trainable lightweight volumetric saliency grid." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "Some existing literature has applied model compression techniques to NeRF: CC-NeRF [26] uses tensor decomposition to obtain a low-rank approximation of the learned network. Similarly, TensoRF [4] represents the volume as a 4D tensor and factorizes it into low-rank components. Unlike these compression works, HollowNeRF vets information based on its actual impact on the rendering accuracy, rather than raw entropy. The relative performance of these approaches are investigated in §4.4." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3481" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 72, + 509, + 227 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 509, + 227 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 509, + 227 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 509, + 227 + ], + "type": "image", + "image_path": "3ca2d7e5c5b0f5261f90d2957e8bf5d044998ade43110992e4a123fdd358d1c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 235, + 400, + 247 + ], + "lines": [ + { + "bbox": [ + 193, + 235, + 400, + 247 + ], + "spans": [ + { + "bbox": [ + 193, + 235, + 400, + 247 + ], + "type": "text", + "content": "Figure 2: Overview of the HollowNeRF workflow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 165, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 165, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 165, + 281 + ], + "type": "text", + "content": "3. HollowNeRF design" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": "This section presents the system design of HollowNeRF, as outlined in Figure 2: (i) Given an input coordinate " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} = (x,y,z)" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": ", we fetch the corresponding feature " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " from the multi-level hashgrid, following the Instant-NGP method [16]. (ii) We predict the saliency weight " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " of position " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " by processing the information from a trainable 3D saliency grid " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " through trilinear interpolation and a sigmoid function (§3.1), and use " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " to scale the feature " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v} = p\\boldsymbol{f}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": ". Figure 3a showcases a trained saliency grid " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": ", demonstrating that training " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " can mitigate hash collisions by suppressing the unnecessary features in empty or invisible regions. (iii) The weighted feature " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " is fed into an MLP, which decodes " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " to obtain the density " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " and color at position " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": ". To ensure that " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v} = \\mathbf{0}" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": " maps to " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "\\sigma = 0" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": ", we introduce a zero-skipping gate (§3.2). (iv) The density and color outputs of the MLP are used for volume rendering, and the resulting image is compared to the ground truth to obtain a loss function " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": ". During training, an ADMM pruner (§3.3) sparsifies the saliency value distribution across the 3D space by enforcing a sparsity constraint when optimizing the loss " + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 288, + 290, + 601 + ], + "type": "text", + "content": ". Figure 3a shows that a large portion of the unnecessary features are not entirely eliminated by training the saliency grid, with saliency weights diminished but not reaching 0. Therefore, we introduce the ADMM pruner to explicitly prune saliency weights of the empty or invisible regions to 0, as shown in Figure 3b." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 610, + 194, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 194, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 194, + 624 + ], + "type": "text", + "content": "3.1. Trainable 3D saliency grid" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "In typical 3D scenes without large transparent objects, most regions of the space do not contribute to the final rendering as they are either empty or invisible from any view angle. To leverage this sparsity, methods such as Instant-NGP use a coarse-grained binary mask to capture space occupancy, which is periodically updated during training by evaluating the NeRF model at each voxel and checking if the density is" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 266, + 354, + 363 + ], + "blocks": [ + { + "bbox": [ + 312, + 266, + 354, + 363 + ], + "lines": [ + { + "bbox": [ + 312, + 266, + 354, + 363 + ], + "spans": [ + { + "bbox": [ + 312, + 266, + 354, + 363 + ], + "type": "image", + "image_path": "54dd887c1bdb73629434dc99e8034bb76592f0a2f4b4d4af68209341634de09a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 371, + 419, + 383 + ], + "lines": [ + { + "bbox": [ + 317, + 371, + 419, + 383 + ], + "spans": [ + { + "bbox": [ + 317, + 371, + 419, + 383 + ], + "type": "text", + "content": "(a) Without ADMM pruner." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 361, + 267, + 422, + 364 + ], + "blocks": [ + { + "bbox": [ + 361, + 267, + 422, + 364 + ], + "lines": [ + { + "bbox": [ + 361, + 267, + 422, + 364 + ], + "spans": [ + { + "bbox": [ + 361, + 267, + 422, + 364 + ], + "type": "image", + "image_path": "252d8b4c9f4d1fba89113877adbbe843346d1f634954c62beb788dd92b424405.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 371, + 529, + 383 + ], + "lines": [ + { + "bbox": [ + 440, + 371, + 529, + 383 + ], + "spans": [ + { + "bbox": [ + 440, + 371, + 529, + 383 + ], + "type": "text", + "content": "(b) With ADMM pruner." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 424, + 268, + 471, + 363 + ], + "blocks": [ + { + "bbox": [ + 424, + 268, + 471, + 363 + ], + "lines": [ + { + "bbox": [ + 424, + 268, + 471, + 363 + ], + "spans": [ + { + "bbox": [ + 424, + 268, + 471, + 363 + ], + "type": "image", + "image_path": "ca227756aa46323dccd8640169d8ee11dc735ca070c41aaec0b287913d16a594.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 392, + 546, + 417 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 546, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 546, + 417 + ], + "type": "text", + "content": "Figure 3: Slices of a " + }, + { + "bbox": [ + 305, + 392, + 546, + 417 + ], + "type": "inline_equation", + "content": "64 \\times 64 \\times 64" + }, + { + "bbox": [ + 305, + 392, + 546, + 417 + ], + "type": "text", + "content": " saliency grid trained on the \"Chair\" scene from the NeRF synthetic dataset[15]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 477, + 268, + 541, + 364 + ], + "blocks": [ + { + "bbox": [ + 477, + 268, + 541, + 364 + ], + "lines": [ + { + "bbox": [ + 477, + 268, + 541, + 364 + ], + "spans": [ + { + "bbox": [ + 477, + 268, + 541, + 364 + ], + "type": "image", + "image_path": "31f42bb1348737bb4fdbda2a83573ccdc7c685d8eb4279f1099b658e0b4869d7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 546, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 509 + ], + "type": "text", + "content": "above a threshold. This occupancy mask guides sampling in ray marching by skipping unoccupied regions. However, it still keeps unnecessary features in internal occluded regions that have non-zero densities but make no contributions to the visible surfaces from any view angle, as shown in Figure 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 605 + ], + "type": "text", + "content": "With this insight, we propose to reclaim the capacity spent on internal regions by learning the sparsity pattern through training, effectively making objects hollow. The benefits of this approach are twofold: it reduces space complexity by requiring fewer features, and improves accuracy by reducing hash collisions. However, precisely locating the invisible regions of objects can be challenging without any prior knowledge of the object's shape." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": "To addresses this challenge, our approach utilizes the hashgrid-based positional encoding from Instant-NGP [16] and extends it by introducing a trainable saliency grid " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " with a resolution of " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "T\\times T\\times T" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": ". Specifically, HollowNeRF trains a 3D tensor data structure " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " alongside the NeRF model (feature hashgrid " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " and MLP weights " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": "), which gradually learns which regions to prioritize, such as a nontransparent object surface. We split the space coarsely as " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "T\\times T\\times T" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " uniform grids and the 3D tensor " + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": " stores train" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3482" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "able values representing the saliency of features near each grid voxel. Instead of letting multiple dense features in the same coarse grid share one value, we interpolate the values associated to the 8 voxels surrounding the input coordinate " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " to get a smoothed saliency distribution " + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": " over the 3D space. The trainable saliency grid represents in itself an opportunity to compress the NeRF model using differentiable model compression techniques, as discussed later in §3.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": "We initialize " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " as an all-ones tensor, where " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " is typically much smaller than the resolution of the dense feature grid (In §4.2, we further investigate how different " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " values affect the performance). Given a 3D coordinate " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{x} = (x,y,z)" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ", we fetch its corresponding feature vector " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " from the multi-level hashgrid [16]. We then locate the 8 voxels " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_1,\\pmb{x}_2,\\dots,\\pmb{x}_8)" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " surrounding " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " in the 3D saliency grid " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ", and fetch its corresponding saliency value " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ". Specifically, the coordinate " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i = (x_i,y_i,z_i)" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " are quantized to the values " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "(\\hat{x}_i,\\hat{y}_i,\\hat{z}_i)" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " between 0 and " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ", which are then used to index the saliency grid " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ". This gives us the saliency value " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "g_i = \\mathcal{G}(\\hat{x}_i,\\hat{y}_i,\\hat{z}_i)" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ". After obtaining " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " for each of the 8 saliency grid voxels surrounding " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ", we calculate the saliency weight " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " at coordinate " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " by trilinear interpolation followed by a sigmoid operator, so that " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "0 < p < 1" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": ". The final weighted feature vector " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " is obtained by multiplying the saliency weight " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " with the feature vector " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": " fetched from the hashgrid, i.e., " + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\pmb{v} = p\\pmb{f}" + }, + { + "bbox": [ + 46, + 167, + 288, + 371 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": "For mathematical analysis, we define the saliency distribution " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " as the collection of saliency weights " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " at every possible coordinate " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "p = \\mathcal{P}(\\boldsymbol{x})" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ". Our proposed method, HollowNeRF, learns the saliency distribution " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " by training the saliency grid " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " using gradient descent. It is worth noting that we inject " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " (and thus " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ") to the training pipeline by feeding the weighted feature " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " instead of the raw feature " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " to the MLP for decoding; the rationale for this choice is that since the sigmoid function in the above workflow projects the saliency weights to the range " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ", we can view the 3D saliency distribution as a probability distribution " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\boldsymbol{x})" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " indicating the likelihood " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " that a feature is non-zero at a given 3D position " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ". There are only two possibilities for the feature at a given position: either the position is visible with probability " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "p = \\mathcal{P}(\\boldsymbol{x})" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ", and we retrieve the feature vector " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " from the multi-level hashgrid as in [16]; otherwise the position is empty or occluded with probability " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "1 - p" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ", and the corresponding feature vector is an all-zero vector " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": ". Taking both cases into account, we can calculate the expectation " + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 46, + 371, + 287, + 610 + ], + "type": "text", + "content": " of the feature as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 620, + 240, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 620, + 240, + 633 + ], + "spans": [ + { + "bbox": [ + 95, + 620, + 240, + 633 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} = \\mathcal {P} (\\boldsymbol {x}) \\boldsymbol {f} + (1 - \\mathcal {P} (\\boldsymbol {x})) \\boldsymbol {0} = p \\boldsymbol {f}", + "image_path": "879d7adcc6ee71a73d01cd1a49625fbb24ea9d45d266f1c2ef1c3a7e5cfe1c72.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Thus the weighted feature " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " fed to MLP decoder has a physical meaning, namely the expectation of the feature at " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " being either " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " fetched from feature hashgrid or 0. Then training the saliency grid is equivalent to learning the probability distribution " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\pmb{x})" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": ". This probability distribution represents the likelihood that a given voxel " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " contains a non-zero fea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "ture. A feature with a lower saliency weight " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " is more likely to be 0 and will cause less interference to other features assigned to the same bucket in the feature hashgrid. When the saliency weight converges to 0, that feature is pruned with a probability of 1 and becomes a zero vector 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "type": "text", + "content": "Training the saliency distribution can also be viewed as hash collision mitigation. Instant-NGP employs hash encoding and distributes hash collisions evenly across voxels with non-zero density. By learning to shuffle hash collisions across the space, HollowNeRF guarantees less interference for more important features. When the feature of a voxel is scaled by a higher saliency weight, it tends to accumulate more gradients during backward propagation, and the information at this voxel will dominate the trained feature value inside the hash bucket shared with other voxels. On the other hand, a feature scaled by zero saliency weight has a zero gradient, hence causes no harm to colliding features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 282, + 437, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 437, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 437, + 296 + ], + "type": "text", + "content": "3.2. Soft zero-skipping gate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": "While HollowNeRF mitigates hash collisions by pruning unnecessary features to zero, the MLP used to decode these features into densities can break the sparsity in the 3D space: the MLP can be viewed as a function " + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\sigma = \\mathcal{M}(\\boldsymbol{v})" + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": " mapping the feature vector " + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": " to the density " + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": ", and typically " + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{0}) \\neq 0" + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": ". Even though the saliency grid and ADMM pruner ensure most voxels direct to " + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": " features in the hash-grid, a typical MLP translates such a sparse feature domain to a non-sparse 3D space with non-zero densities scattered everywhere, which is evidently inaccurate. To enforce a 0 output density when the input feature is " + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": ", we introduce a zero-skipping gate " + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "inline_equation", + "content": "g(\\boldsymbol{v})" + }, + { + "bbox": [ + 304, + 300, + 545, + 445 + ], + "type": "text", + "content": " to the MLP decoder:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 396, + 453, + 489, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 453, + 489, + 465 + ], + "spans": [ + { + "bbox": [ + 396, + 453, + 489, + 465 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {M}} (\\boldsymbol {v}) = \\hat {g} (\\boldsymbol {v}) \\cdot \\mathcal {M} (\\boldsymbol {v})", + "image_path": "80f03d17c0ebace977c7251bc90b3a39424bfdb84691e32c732b1819f0b6507e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 468, + 545, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 468, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 361, + 468, + 545, + 482 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad \\hat {g} (\\boldsymbol {v}) := \\tanh (\\alpha \\| \\boldsymbol {v} \\| _ {2}) \\tag {1}", + "image_path": "47faf062c35ed8584146f0760bae71832273c01f759ef67d02bc0be89c7a6489.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "text", + "content": "The gate " + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\hat{g}(\\pmb{v})" + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "text", + "content": " is a differentiable approximation of the \"hard\" gate in Eq. (2) to enforce the desired property " + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}(\\mathbf{0}) = 0" + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "text", + "content": ", and we use a constant " + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 489, + 545, + 536 + ], + "type": "text", + "content": " to control the steepness of the 0-to-1 transition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 543, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 543, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 364, + 543, + 545, + 574 + ], + "type": "interline_equation", + "content": "g (\\boldsymbol {v}) := \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\| \\boldsymbol {v} \\| _ {2} = 0 \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "f069ba211d057702e5d26f30ebb0b724b3f84724d25f9fffaed2d097ab45de8a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": "The non-differentiable \"hard\" gate in Eq. (2) may cause abrupt density value changes during training. Also, the gradient " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\frac{\\partial g}{\\partial p}" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "g(\\boldsymbol{v}) = g(p\\boldsymbol{f})" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " is 0 with " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "p \\neq 0" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": ", this 0 gradient prevents the ADMM pruner from suppressing a small saliency value " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " to 0, resulting in insufficient feature pruning and little improvement in hash collision. Our experiments in §4.3 confirm that adding the \"hard\" gate " + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "inline_equation", + "content": "g(\\boldsymbol{v})" + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": " provides little gain over adding no gate." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Thus, we propose a \"soft\" zero-skipping gate " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{g}(\\pmb{v})" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " in Eq. (1) that smoothly reduces the density output to zero for input features being pruned. As an approximation of " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "g(\\pmb{v})" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3483" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "the value of " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\hat{g}(\\boldsymbol{v})" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " should smoothly transition from 0 to 1 for input " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " with small magnitudes " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\|\\boldsymbol{v}\\|_2" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", while otherwise staying close to 1 to avoid perturbing necessary information. The tanh function is a suitable choice. In our implementation we adjust the " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " value with a schedule, to gradually \"harden\" the soft gate as the training progresses. Specifically, for epochs below 1000, we set " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "10^4" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " to ensure fast convergence. Afterward we increase " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "10^5" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " to minimize the perturbation to the fine-tuning process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 194, + 141, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 141, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 141, + 208 + ], + "type": "text", + "content": "3.3. ADMM pruner" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 216, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 276 + ], + "type": "text", + "content": "Training the 3D saliency grid improves the accuracy by redistributing hash collisions, which, However, cannot fully eliminate unnecessary features, as shown in Figure 3a. To address this, we introduce an ADMM pruning that enforces sparsity in the saliency grid " + }, + { + "bbox": [ + 46, + 216, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 216, + 287, + 276 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": "We first motivate the use of ADMM pruner rigorously. Assuming a " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "K \\times K \\times K" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " volume where each voxel " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " corresponds to a feature " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ", and the ground truth feature " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " at voxel " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1 = (x_1, y_1, z_1)" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " at a different voxel " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x}_2 = (x_2, y_2, z_2)" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " are typically different for a 3D scene. Instant-NGP encodes this " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "K \\times K \\times K" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " feature grid into a small hashgrid " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " whose size " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "<< (K \\times K \\times K)" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " collide with each other by directing to the same hash bucket, their values queried from the hashgrid become identical " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1^* = \\pmb{f}_2^* = \\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " is the trained feature stored in the shared hash bucket. Since the ground truth features " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1 \\neq \\pmb{f}_2" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " while the queried features " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1^* = \\pmb{f}_2^*" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ", the queried values differ from the ground truth, causing compression artifacts. To solve this problem, HollowNeRF allows the colliding features to be different by introducing the scalar saliency weight " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " to scale the queried features: " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1^* = p(\\pmb{x}_1)\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2^* = p(\\pmb{x}_2)\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ". However, during training, the share feature " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " may not find a value that satisfies both equations, as they can be written as " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1^* = \\frac{p(\\pmb{x}_1)}{p(\\pmb{x}_2)}\\pmb{f}_2^*" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " which is not always true especially when " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " are nonzero. Then the queried features " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1^*" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2^*" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " may still cause compression artifacts (although less than instant-NGP). The ADMM pruner addresses this challenge by pruning the ground truth " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " to 0 when " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x}_2" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " is empty/invisible. For example, if voxel " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " is invisible from any view angle " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "(\\pmb{f}_1 = 0)" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ", there exists a single " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " value to satisfy both " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1 = p(\\pmb{x}_1)\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2 = p(\\pmb{x}_2)\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ", that is " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_H = \\pmb{f}_2" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": ", and there is no compression artifacts. The optimizer can converge to this optimal " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{f}_H" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " value through gradient descent by learning that " + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_1)" + }, + { + "bbox": [ + 46, + 277, + 289, + 663 + ], + "type": "text", + "content": " is 0. In summary, making the feature grid sparse reduces the compression artifacts when encoding features into a hashgrid." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "In what follows, we describe two methods for achieving sparsity: a basic L1-regularization approach and our ADMM pruner. To sparsify the saliency grid " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", one simple approach is to add an L1 regularization term to the original" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 524, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 524, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 524, + 85 + ], + "type": "text", + "content": "MSE loss function " + }, + { + "bbox": [ + 306, + 72, + 524, + 85 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 306, + 72, + 524, + 85 + ], + "type": "text", + "content": " typically used in NeRF training:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 370, + 94, + 545, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 94, + 545, + 113 + ], + "spans": [ + { + "bbox": [ + 370, + 94, + 545, + 113 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {W}, \\mathcal {H}, \\mathcal {G}} \\left(L + \\lambda \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1}\\right) \\tag {3}", + "image_path": "d4d25e3bd39a6c8955526e3be207f478ccc3444eaff77727cda5c441bc5f5199.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": " is a constant weight to control the amount of sparsity and compression, and " + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": " denote respectively the embeddings in the multi-level hashgrid and the weights in the MLP. We use the sigmoid operator " + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{sig}}(\\cdot)" + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": " here because the forward pipeline (Figure 2) uses the saliency value after the sigmoid operator to weight the feature vector, and our ultimate goal is a sparse distribution of saliency value to prune unnecessary features. Note that " + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{sig}}(\\mathcal{G})" + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": " here denotes applying sigmoid operator on each element " + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": " of the 3D tensor " + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 122, + 547, + 266 + ], + "type": "text", + "content": ", and the result is a 3D tensor with all non-negative elements. Therefore, the absolute operator in the L1 regularization term can be removed:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 340, + 274, + 512, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 274, + 512, + 300 + ], + "spans": [ + { + "bbox": [ + 340, + 274, + 512, + 300 + ], + "type": "interline_equation", + "content": "\\| S _ {\\operatorname {s i g}} (\\mathcal {G}) \\| _ {1} = \\sum_ {g \\in \\mathcal {G}} | S _ {\\operatorname {s i g}} (g) | = \\sum_ {g \\in \\mathcal {G}} S _ {\\operatorname {s i g}} (g)", + "image_path": "b48a578c7f97332eada5552003a39852c93d73dc67eb0b40131da733aee3f325.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 309, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 369 + ], + "type": "text", + "content": "Removing the absolute operator from the regularization term makes it differentiable and compatible with the training pipeline, eliminating the need to use more complex optimization algorithms such as proximal gradient descent to handle non-differentiable regularization terms." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "text", + "content": "Although the usage of L1 regularization is already a marked improved over the baseline (see §4.3), we find it is often difficult to select a single " + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "text", + "content": " hyperparameter to control sparsity and compression. During training when the 3D saliency grid " + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "text", + "content": " has already become sparse enough, the regularization term " + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\lambda \\| S_{\\mathrm{sig}}(\\mathcal{G})\\|_1" + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "text", + "content": " may still provide gradients that further prune " + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 369, + 545, + 477 + ], + "type": "text", + "content": " is set too high or if the algorithm is let to run for too long, visible features may be pruned resulting in performance degradation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 477, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 526 + ], + "type": "text", + "content": "We use an ADMM pruner to overcome the limitations of simple L1 regularization. By introducing a configurable sparsity constraint " + }, + { + "bbox": [ + 304, + 477, + 545, + 526 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 477, + 545, + 526 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 477, + 545, + 526 + ], + "type": "inline_equation", + "content": "\\| S_{\\mathrm{sig}}(\\mathcal{G})\\|_1 < C" + }, + { + "bbox": [ + 304, + 477, + 545, + 526 + ], + "type": "text", + "content": ", feature pruning becomes the constrained optimization problem:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 548, + 499, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 548, + 499, + 582 + ], + "spans": [ + { + "bbox": [ + 353, + 548, + 499, + 582 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\mathcal {M}, \\mathcal {H}, \\mathcal {G}} L \\left(\\mathcal {M}, \\mathcal {H}, \\mathcal {G}\\right) \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1} < C \\end{array} \\\\ \\end{array}", + "image_path": "c4ce6f952bc5b23920f4219ee78d1243d6f1fa5b7f86c197e91192e50dd3d1ab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "text", + "content": "We transform this constrained optimization problem into an equivalent unconstrained minimax problem in Problem (5) by using the augmented Lagrangian method which employs a trainable dual variable " + }, + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "text", + "content": ", also known as a Lagrange multiplier. In essence, " + }, + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "text", + "content": " serves as a trainable alternative to the fixed hyperparameter " + }, + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 592, + 545, + 664 + ], + "type": "text", + "content": " in Problem (3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 673, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 673, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 317, + 673, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\mathcal {W}, \\mathcal {H}, \\mathcal {G}} \\max _ {\\gamma \\geq 0} L \\left(\\mathcal {W}, \\mathcal {H}, \\mathcal {G}\\right) + \\frac {\\rho_ {\\gamma}}{2} \\left[ \\| S _ {\\text {s i g}} (\\mathcal {G}) \\| _ {1} - C \\right] _ {+} ^ {2} \\\\ + \\gamma \\left(\\left\\| S _ {\\text {s i g}} (\\mathcal {G}) \\right\\| _ {1} - C\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "83ea234c5f39ce4cebfca284a97a21e45ade40cdd00cdd221d31b7f6ec157477.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3484" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 127, + 70, + 468, + 164 + ], + "blocks": [ + { + "bbox": [ + 127, + 70, + 468, + 164 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 468, + 164 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 468, + 164 + ], + "type": "table", + "html": "
MethodHashgrid size214215216217218219
Param #
Instant-NGP0.50M0.94M1.77M3.34M6.22M11.49M
HollowNeRFT=640.76M1.21M2.04M3.60M6.48M11.75M
T=961.39M1.83M2.66M4.22M7.10M12.37M
T=1282.60M3.04M3.87M5.43M8.32M13.58M
", + "image_path": "90012d9cda07e88f118596a16af90f890752d6776442d5996673628c8f720d8d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 172, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 547, + 222 + ], + "type": "text", + "content": "Table 1: The total number of model parameters, depending on the hashgrid size and saliency grid resolution " + }, + { + "bbox": [ + 46, + 172, + 547, + 222 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 172, + 547, + 222 + ], + "type": "text", + "content": ". This number counts the parameters in the hashgrid, MLP, and saliency grid. (both HollowNeRF and Instant-NGP employ a multi-resolution hashgrid with 16 levels. We use the maximum entries per level to denote the hashgrid size, following the notation in the Instant-NGP paper [16].)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "inline_equation", + "content": "\\rho_{\\gamma}" + }, + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "text", + "content": " denotes the learning rate of dual variable " + }, + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "inline_equation", + "content": "[\\cdot ]_{+}" + }, + { + "bbox": [ + 47, + 242, + 288, + 266 + ], + "type": "text", + "content": " clamps the input to non-negative values." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "content": "At each iteration " + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "content": ", the optimizer first updates the coarse saliency grid " + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "content": ", the feature hashgrid " + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "content": ", and the MLP weights " + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "content": " to minimize the augmented loss function in Problem (5) using gradient descent, where the dual variable " + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "content": " is treated as a constant. " + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "text", + "content": " is then updated using gradient ascent with " + }, + { + "bbox": [ + 46, + 266, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\gamma^{(t + 1)} = \\left[\\gamma^{(t)} + \\rho_{\\gamma}\\left(\\| S_{\\mathrm{sig}}(\\mathcal{G})\\| _1 - C\\right)\\right]_+" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 351, + 119, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 119, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 119, + 364 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 372, + 289, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 289, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 289, + 458 + ], + "type": "text", + "content": "In this section, we first investigate the effect of different hyper-parameters, including hashgrid sizes, saliency grid sizes, and sparsity constraints, on the performance (§4.2). We then conduct an ablation study to validate each of HollowNeRF's design components (§4.3). Finally, we evaluate the performance of HollowNeRF on various 3D scenes and compare it with the state-of-the-art works (§4.4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 466, + 163, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 163, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 163, + 479 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 485, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 616 + ], + "type": "text", + "content": "We implement HollowNeRF on top of torch-ngp [25, 26], a PyTorch CUDA extension implementation of instant-NGP [16]. To ensure a fair comparison, we evaluated both HollowNeRF and instant-NGP using the same torch-ngp implementation. Both methods use a 16-level hashgrid with a feature dimension of 2, where the coarsest resolution of the feature grid is 16, and the finest resolution is 1024. Both methods employ a MLP decoder that has only two hidden layers with a width of 64 neurons, while HollowNeRF augments the MLP with a zero-skipping gate as described in §3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "We use the peak signal-to-noise ratio (PSNR) and the Learned Perceptual Image Patch Similarity (LPIPS) [33] with the default AlexNet backend as the performance metrics to assess the rendering quality. Each experiments was conducted on a single NVIDIA Tesla A100 GPU for both training and evaluation. The machine used was equipped with an AMD EPYC 7742 64-Core CPU, 1TB of memory, and 8 GPU cards in total." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 243, + 537, + 381 + ], + "blocks": [ + { + "bbox": [ + 315, + 243, + 537, + 381 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 537, + 381 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 537, + 381 + ], + "type": "image", + "image_path": "fec9c0c29f1d0be494664c7383dcc15824fbd1ac9f6f9e7d0d7b93b1b8fef20d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 399, + 544, + 540 + ], + "blocks": [ + { + "bbox": [ + 372, + 387, + 480, + 397 + ], + "lines": [ + { + "bbox": [ + 372, + 387, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 372, + 387, + 480, + 397 + ], + "type": "text", + "content": "(a) PSNR " + }, + { + "bbox": [ + 372, + 387, + 480, + 397 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 372, + 387, + 480, + 397 + ], + "type": "text", + "content": " (higher is better)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 399, + 544, + 540 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 544, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 544, + 540 + ], + "type": "image", + "image_path": "f1315eb5158ec21a3c15de2ad3ecb11f5c8b4f4362c69cb00ea0cf85fa043166.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 544, + 479, + 555 + ], + "lines": [ + { + "bbox": [ + 372, + 544, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 372, + 544, + 479, + 555 + ], + "type": "text", + "content": "(b) LPIPS " + }, + { + "bbox": [ + 372, + 544, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 372, + 544, + 479, + 555 + ], + "type": "text", + "content": " (lower is better)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 566, + 522, + 578 + ], + "lines": [ + { + "bbox": [ + 329, + 566, + 522, + 578 + ], + "spans": [ + { + "bbox": [ + 329, + 566, + 522, + 578 + ], + "type": "text", + "content": "Figure 4: Performance vs. saliency grid size " + }, + { + "bbox": [ + 329, + 566, + 522, + 578 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 329, + 566, + 522, + 578 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "text", + "content": "Unless otherwise specified, we train the model for 300,000 steps using the Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-2}" + }, + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "text", + "content": ". Further details regarding the appropriate selection of HollowNeRF hyper-parameters, such as " + }, + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 601, + 547, + 661 + ], + "type": "text", + "content": ", are discussed in §4.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 670, + 515, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 670, + 515, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 670, + 515, + 684 + ], + "type": "text", + "content": "4.2. Investigating the cost-accuracy tradeoff" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "content": "We first investigate the cost-accuracy tradeoff of HollowNeRF to determine the appropriate hyper-parameter config" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3485" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 285, + 212 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 285, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 285, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 285, + 212 + ], + "type": "image", + "image_path": "9dbaf5eb65d4f1502ae0024b024709fbd374d2e26d68348c924e71c2aa26a471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 217, + 221, + 228 + ], + "lines": [ + { + "bbox": [ + 113, + 217, + 221, + 228 + ], + "spans": [ + { + "bbox": [ + 113, + 217, + 221, + 228 + ], + "type": "text", + "content": "(a) PSNR " + }, + { + "bbox": [ + 113, + 217, + 221, + 228 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 113, + 217, + 221, + 228 + ], + "type": "text", + "content": " (higher is better)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 232, + 282, + 370 + ], + "blocks": [ + { + "bbox": [ + 56, + 232, + 282, + 370 + ], + "lines": [ + { + "bbox": [ + 56, + 232, + 282, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 282, + 370 + ], + "type": "image", + "image_path": "efa6804fc207a5453459d80664d9ac47fb1f60fe32017b609be1af4d890a4651.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 375, + 220, + 386 + ], + "lines": [ + { + "bbox": [ + 113, + 375, + 220, + 386 + ], + "spans": [ + { + "bbox": [ + 113, + 375, + 220, + 386 + ], + "type": "text", + "content": "(b) LPIPS " + }, + { + "bbox": [ + 113, + 375, + 220, + 386 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 113, + 375, + 220, + 386 + ], + "type": "text", + "content": " (lower is better)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "text", + "content": "uration for larger scale experiments in §4.4. We conducted experiments on the chair scene from the NeRF synthetic dataset, varying the hashgrid size from " + }, + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "inline_equation", + "content": "2^{14}" + }, + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "inline_equation", + "content": "2^{19}" + }, + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "text", + "content": " and testing the saliency grid resolution of 64, 96, and 128 for each hashgrid size. Table 1 summarizes the cost of different configurations, including the number of parameters in the hashgrid, MLP, and saliency grid. We use a sparsity constraint of " + }, + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "inline_equation", + "content": "C = 0.04" + }, + { + "bbox": [ + 46, + 435, + 287, + 532 + ], + "type": "text", + "content": " in this experiment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": "The results in Figure 4 indicate that increasing the saliency grid resolution " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " from 64 to 96 or 128 does not improve accuracy much but results in a higher parameter count, which harms the cost-accuracy tradeoff. On the other hand, lowering " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " below 64 causes unstable convergence. Consequently, we use a grid resolution of " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "T = 64" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " for the subsequent experiments. Figure 4 further shows that HollowNeRF outperforms the baseline over a wide range of model sizes, achieving higher PSNR and lower LPIPS values for comparable parameter counts. The best overall model, HollowNeRF with " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "T = 64" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": ", uses 6.48M parameters to attain a 34.85dB PSNR, as compared to Instant-NGP's 33.86dB using 11.49M. In other words, it achieves about 1dB higher PSNR than Instant-NGP, while utilizing only 56% of the parameters employed by Instant-NGP." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 74, + 541, + 212 + ], + "blocks": [ + { + "bbox": [ + 315, + 74, + 541, + 212 + ], + "lines": [ + { + "bbox": [ + 315, + 74, + 541, + 212 + ], + "spans": [ + { + "bbox": [ + 315, + 74, + 541, + 212 + ], + "type": "image", + "image_path": "ac5e9e6d256221fc1eec60364db84c60553091efd7f38c5c44a1d72d53f3f9b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 217, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 372, + 217, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 372, + 217, + 481, + 228 + ], + "type": "text", + "content": "(a) PSNR " + }, + { + "bbox": [ + 372, + 217, + 481, + 228 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 372, + 217, + 481, + 228 + ], + "type": "text", + "content": " (higher is better)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 232, + 543, + 370 + ], + "blocks": [ + { + "bbox": [ + 69, + 396, + 265, + 408 + ], + "lines": [ + { + "bbox": [ + 69, + 396, + 265, + 408 + ], + "spans": [ + { + "bbox": [ + 69, + 396, + 265, + 408 + ], + "type": "text", + "content": "Figure 5: Performance vs. sparsity constraint " + }, + { + "bbox": [ + 69, + 396, + 265, + 408 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 69, + 396, + 265, + 408 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 232, + 543, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 232, + 543, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 543, + 370 + ], + "type": "image", + "image_path": "f841c4e3717ef361fabb2406887c4c495123d69e3d776b763d6b00e2328204b0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 375, + 479, + 386 + ], + "lines": [ + { + "bbox": [ + 372, + 375, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 372, + 375, + 479, + 386 + ], + "type": "text", + "content": "(b) LPIPS " + }, + { + "bbox": [ + 372, + 375, + 479, + 386 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 372, + 375, + 479, + 386 + ], + "type": "text", + "content": " (lower is better)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 396, + 531, + 408 + ], + "lines": [ + { + "bbox": [ + 319, + 396, + 531, + 408 + ], + "spans": [ + { + "bbox": [ + 319, + 396, + 531, + 408 + ], + "type": "text", + "content": "Figure 6: Ablation study for the design components." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 443, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 515 + ], + "type": "text", + "content": "We also repeated the training and evaluation for each hashgrid size five times for HollowNeRF with " + }, + { + "bbox": [ + 304, + 443, + 545, + 515 + ], + "type": "inline_equation", + "content": "T = 64" + }, + { + "bbox": [ + 304, + 443, + 545, + 515 + ], + "type": "text", + "content": " and instant-NGP, and plotted the standard deviations across the five tests as error bars in Figure 4. The results demonstrate that our approach is more stable than Instant-NGP in terms of convergence, with lower performance variation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": "To investigate the impact of the choice of sparsity constraint " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " on performance, we conducted additional experiments following the same setup as previous tests, while varying the " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " values, and report the results in Figure 5. First, from Figure 5a, we observe that the " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "C = 0.04" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " configuration yields the highest PSNR, in other words, HollowNeRF achieves the highest accuracy when " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " of the spatial voxels are non-zero. " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "C = 0.02" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " exhibits slightly diminished accuracy, potentially due to excessive pruning in the 3D space, which compromises crucial information for scene reconstruction. Conversely, " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "C = 0.08" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " shows the lowest accuracy across all tested HollowNeRF configurations, suggesting that keeping " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " non-zero voxels causes excessive hash collisions when packing them into the hashgrid, and the 3D scene to capture could be sparser. We have similar observation for LPIPS, as depicted in Figure 5b." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3486" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 302, + 398 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 302, + 398 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 302, + 398 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 302, + 398 + ], + "type": "image", + "image_path": "b398dc24f851b638ba4004a1fc9aa6645146ea1857ac019ae983e8a3e005e2c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 406, + 492, + 419 + ], + "lines": [ + { + "bbox": [ + 100, + 406, + 492, + 419 + ], + "spans": [ + { + "bbox": [ + 100, + 406, + 492, + 419 + ], + "type": "text", + "content": "Figure 7: Qualitative comparison with instant-NGP over scenes from the NeRF synthetic dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 544, + 398 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 544, + 398 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 544, + 398 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 544, + 398 + ], + "type": "image", + "image_path": "c6a24caaa983c0d5743c48e5622edb063d25b6cfb296aa9a1ec8541d200ed7ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 427, + 539, + 674 + ], + "blocks": [ + { + "bbox": [ + 58, + 427, + 539, + 674 + ], + "lines": [ + { + "bbox": [ + 58, + 427, + 539, + 674 + ], + "spans": [ + { + "bbox": [ + 58, + 427, + 539, + 674 + ], + "type": "image", + "image_path": "e9202d24c5e0f9f79f255c311d39ff25a9fa89ae4c57c33d8d9caec8df9d9adf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 681, + 541, + 693 + ], + "lines": [ + { + "bbox": [ + 51, + 681, + 541, + 693 + ], + "spans": [ + { + "bbox": [ + 51, + 681, + 541, + 693 + ], + "type": "text", + "content": "Figure 8: Comparison of PSNR(↑) and LPIPS (↓) performance between HollowNeRF and Instant-NGP on various scenes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3487" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 281, + 229 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 281, + 229 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 281, + 229 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 281, + 229 + ], + "type": "image", + "image_path": "c8d0aca4088252db84726f814d567a6c69f22ce6b9963f4b6ef9d9116c773c86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 288, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 277 + ], + "type": "text", + "content": "Figure 9: Comparison of HollowNeRF and recent methods in terms of the trade-off between PSNR and model size (\"HollowNeRF-H19\" means using a hashgrid size of " + }, + { + "bbox": [ + 46, + 240, + 288, + 277 + ], + "type": "inline_equation", + "content": "2^{19}" + }, + { + "bbox": [ + 46, + 240, + 288, + 277 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 298, + 139, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 139, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 139, + 310 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "text", + "content": "In this ablation study, we analyze the impact of each component in our solution by comparing the convergence speed and resulting performance of the configurations with and without each design component. We use a saliency grid resolution of 64 and a hashgrid size of " + }, + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "inline_equation", + "content": "2^{18}" + }, + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "text", + "content": " (total 6.48M parameters) for all HollowNeRF configurations, while the baseline is instant-NGP with a hashgrid size of " + }, + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "inline_equation", + "content": "2^{19}" + }, + { + "bbox": [ + 46, + 316, + 287, + 448 + ], + "type": "text", + "content": " (total 11.49M parameters). To investigate how each component affects the convergence speed, we evaluate the performance on the test sets every 300 epochs during the total 3000 epochs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 448, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 509 + ], + "type": "text", + "content": "The results presented in Figure 6 show that each design component presence gracefully increases the overall performance, without any catastrophic failure. Each of the HollowNeRF ablated configurations still outperforms instant-NGP, while employing less parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 515, + 251, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 251, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 251, + 528 + ], + "type": "text", + "content": "4.4. Comparisons on standardized datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "Finally, we evaluate the performance of HollowNeRF on 3D scenes from the NeRF synthetic dataset [15] and the more complicated Tanks and Temples dataset [11]. Based on the cost-accuracy tradeoff investigated in Sec. 4.2, we set HollowNeRF's hyper-parameters to a sparsity constraint of " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "C = 0.04" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": " and a saliency grid resolution of " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "T = 64" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 606, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 288, + 714 + ], + "type": "text", + "content": "We first provide qualitative example views comparing HollowNeRF with Instant-NGP. The distinct advantages of HollowNeRF are evident through the visuals showcased in Figure 7. For example, from the lego scene in column 1, we observe that HollowNeRF with 6.48M parameters achieves notably superior rendering accuracy over Instant-NGP with " + }, + { + "bbox": [ + 46, + 606, + 288, + 714 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 46, + 606, + 288, + 714 + ], + "type": "text", + "content": " more parameters (11.49M). The advantage is particularly pronounced within the bulldozer track region. Furthermore, HollowNeRF with 11.75M parameters exhibits" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "the highest accuracy, capturing fine details such as the tiny indentation on the edge of the black rod located at the left border of the view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "text", + "content": "Then we show the PSNR and LPIPS performance of HollowNeRF and Instant-NGP on various 3D scenes (we sample 10 different scenes due to the page limit) in Figure 8. For each scene, we test three hashgrid sizes " + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "inline_equation", + "content": "(2^{17}, 2^{18}, 2^{19})" + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "text", + "content": " for both methods. We can observe that HollowNeRF consistently outperforms Instant-NGP on all the 10 scenes tested, achieving better PSNR and LPIPS performance with fewer parameters when using the same hashgrid size. For seven out of ten tested scenes, the HollowNeRF with the smallest hashgrid size " + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "inline_equation", + "content": "(2^{17}" + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "text", + "content": ", for a total parameter count including MLP of " + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "inline_equation", + "content": "3.34\\mathrm{M}" + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "text", + "content": ") achieves higher PSNR than the best Instant-NGP model (hashgrid size of " + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "inline_equation", + "content": "2^{19}" + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "text", + "content": ", corresponding to " + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "inline_equation", + "content": "11.49\\mathrm{M}" + }, + { + "bbox": [ + 304, + 110, + 546, + 266 + ], + "type": "text", + "content": " parameters including the MLP)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 266, + 547, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 350 + ], + "type": "text", + "content": "Finally, we compare the average PSNR across the NeRF synthetic dataset [16] and the corresponding model size for HollowNeRF and several recent works. Figure 9 shows that HollowNeRF achieves the best cost-accuracy tradeoff among all tested methods. In particular, HollowNeRF with a hashgrid size of " + }, + { + "bbox": [ + 304, + 266, + 547, + 350 + ], + "type": "inline_equation", + "content": "2^{17}" + }, + { + "bbox": [ + 304, + 266, + 547, + 350 + ], + "type": "text", + "content": " achieves an average PSNR of 32.53dB with only a mode size of 14.0MB." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 366, + 375, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 375, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 375, + 379 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 460 + ], + "type": "text", + "content": "The compression gains of HollowNeRF require the 3D scene being sparse, i.e. mostly filled with either empty or occluded space, which holds for typical 3D scenes. HollowNeRF can still operate when this assumption does not hold, such as for scenes containing smoke, fire, clouds, but its performance will regress to the baseline without pruning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 521 + ], + "type": "text", + "content": "The current version of HollowNeRF, like its predecessor Instant-NGP, faces challenges in modeling objects with reflective surfaces. In future works, we plan to extend the concept of HollowNeRF to the frameworks [8, 34, 21, 28, 14] with better support for reflective surfaces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 536, + 384, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 384, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 384, + 549 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "content": "In this paper, we present HollowNeRF, a novel hashgrid-based NeRF technique that achieves superior rendering quality than state-of-the-art solutions like instant-NGP while using only a fraction of the parameters. HollowNeRF mitigates hash collisions by a simple feature pruning mechanism. Unlike existing methods that rely on explicit surface geometries for feature pruning, HollowNeRF learns a 3D saliency grid to guide feature compression during training, and employs an ADMM pruner to enforce a sparse feature domain. Our experiments demonstrate that HollowNeRF achieves a better balance between cost and accuracy than state-of-the-art solutions, making a solid step towards lightweight and ubiquitous NeRF applications." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3488" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, October 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 245 + ], + "type": "text", + "content": "[3] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, Jonathan Eckstein, et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends® in Machine learning, 3(1):1-122, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 279 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 324 + ], + "type": "text", + "content": "[5] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6124-6132, June 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 378 + ], + "type": "text", + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, June 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "text", + "content": "[7] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 490 + ], + "type": "text", + "content": "[8] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 2, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "type": "text", + "content": "[9] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "text", + "content": "[10] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, October 2021. 2, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 645 + ], + "type": "text", + "content": "[11] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[12] Andreas Kurz, Thomas Neff, Zhaoyang Lv, Michael Zollhöfer, and Markus Steinberger. Adanerf: Adaptive sampling for real-time rendering of neural radiance fields. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 174 + ], + "type": "text", + "content": "[14] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, June 2022. 2, 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 220 + ], + "type": "text", + "content": "[15] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "text", + "content": "[16] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 4, 6, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 269, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 334 + ], + "type": "text", + "content": "[17] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, June 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 337, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 403 + ], + "type": "text", + "content": "[18] Thomas Neff, Pascal Stadlbauer, Mathias Parger, Andreas Kurz, Joerg H. Mueller, Chakravarty R. Alla Chaitanya, Anton S. Kaplanyan, and Markus Steinberger. DONeRF: Towards Real-Time Rendering of Compact Neural Radiance Fields using Depth Oracle Networks. Computer Graphics Forum, 40(4), 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 405, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 471 + ], + "type": "text", + "content": "[19] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5480-5490, June 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 473, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 473, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 473, + 545, + 517 + ], + "type": "text", + "content": "[20] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 519, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 519, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 545, + 563 + ], + "type": "text", + "content": "[21] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 2, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "type": "text", + "content": "[22] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "type": "text", + "content": "[23] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "text", + "content": "[24] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, June 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3489" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 597 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[25] Jiaxiang Tang. Torch-ngp: a pytorch implementation of instant-ngp, 2022. https://github.com/ashawkey/torch-ngp.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "text", + "content": "[26] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable neRF via rank-residual decomposition. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6, 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "type": "text", + "content": "[27] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12922–12931, June 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 283 + ], + "type": "text", + "content": "[28] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, June 2022. 2, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 285, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 285, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 287, + 339 + ], + "type": "text", + "content": "[29] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 342, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 287, + 396 + ], + "type": "text", + "content": "[30] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, June 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 397, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 397, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 49, + 397, + 287, + 451 + ], + "type": "text", + "content": "[31] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 453, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 453, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 49, + 453, + 287, + 518 + ], + "type": "text", + "content": "[32] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 520, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 520, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 49, + 520, + 287, + 552 + ], + "type": "text", + "content": "[33] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 554, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 554, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 554, + 287, + 597 + ], + "type": "text", + "content": "[34] Chengxuan Zhu, Renjie Wan, and Boxin Shi. Neural transmitted radiance fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 9" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3490" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_content_list.json b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a7e0bd04850f86f57d0295f479fa5ecd3265000b --- /dev/null +++ b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_content_list.json @@ -0,0 +1,1653 @@ +[ + { + "type": "text", + "text": "HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World", + "text_level": 1, + "bbox": [ + 212, + 130, + 759, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Wang $^{1*}$ Taein Kwon $^{1,2*}$ Mahdi Rad $^{1}$ Bowen Pan $^{1\\dagger}$ Ishani Chakraborty $^{1}$ Sean Andrist $^{1}$ Dan Bohus $^{1}$ Ashley Feniello $^{1}$ Bugra Tekin $^{1\\dagger}$ Felipe Vieira Frujeri $^{1}$ Neel Joshi $^{1}$ Marc Pollefeys $^{1,2}$ Microsoft ${}^{2}$ ETH Zurich", + "bbox": [ + 101, + 202, + 913, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building an interactive AI assistant that can perceive, reason, and collaborate with humans in the real world has been a long-standing pursuit in the AI community. This work is part of a broader research effort to develop intelligent agents that can interactively guide humans through performing tasks in the physical world. As a first step in this direction, we introduce HoloAssist, a large-scale egocentric human interaction dataset, where two people collaboratively complete physical manipulation tasks. The task performer executes the task while wearing a mixed-reality headset that captures seven synchronized data streams. The task instructor watches the performer's egocentric video in real time and guides them verbally. By augmenting the data with action and conversational annotations and observing the rich behaviors of various participants, we present key insights into how human assistants correct mistakes, intervene in the task completion procedure, and ground their instructions to the environment. HoloAssist spans 166 hours of data captured by 350 unique instructor-performer pairs. Furthermore, we construct and present benchmarks on mistake detection, intervention type prediction, and hand forecasting, along with detailed analysis. We expect HoloAssist will provide an important resource for building AI assistants that can fluidly collaborate with humans in the real world. Data can be downloaded at https://holoassist.github.io/.", + "bbox": [ + 76, + 323, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed incredible progress in general-purpose AI agents that assist humans with various open-world tasks, especially in the digital world. AI systems powered by large language models (LLMs) like ChatGPT [27] can answer users' questions and assist them with various text-based tasks. However, these AI assistants do not have sufficient first-hand experience in the physical world and thus cannot perceive world states and actively intervene", + "bbox": [ + 75, + 756, + 470, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two-Person Interactive Assistive Task Completion", + "text_level": 1, + "bbox": [ + 557, + 296, + 833, + 308 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1dda76f6bd78d81d0df44b1d520ffd96c73633b26be16a4df7c6b6a0a1076a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 310, + 696, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7633283a5d89767f41f1e99ea3c91a2917798a2d8eab6c4d403ec5ff07f4303d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 310, + 890, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87fcba3aa50b0e7330ac13ab705e577d5cecca34184a6521f6678610e9f82686.jpg", + "image_caption": [ + "(1) RGB, (2) eye gaze, (3) hand pose", + "Figure 1: HoloAssist features a two-person interactive assistive task completion setting. The task performer wears an AR device and completes the tasks while the captured data is streamed over the network to a remote instructor watching it on the laptop. The instructor provides verbal guidance to the student. HoloAssist includes seven modalities captured live and human annotated text descriptions as the 8th modality." + ], + "image_footnote": [], + "bbox": [ + 501, + 508, + 606, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e501e0ad178f56304f6fd0c67d2171de4501fbb580efbe8924d44c112ed7af3.jpg", + "image_caption": [ + "(4) head pose, (5) depth" + ], + "image_footnote": [], + "bbox": [ + 607, + 508, + 712, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61e5bda88e7a330a35e155563e39782e0835a37940b97082f348a3448971c1a7.jpg", + "image_caption": [ + "(6) IMU, (7) audio" + ], + "image_footnote": [], + "bbox": [ + 714, + 508, + 831, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2feb99fad4a437d27f11d01deddfaa2377790abb610f10e00484df729312c4be.jpg", + "image_caption": [ + "(8) text" + ], + "image_footnote": [], + "bbox": [ + 834, + 511, + 888, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in the task completion procedure.", + "bbox": [ + 500, + 733, + 720, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building an AI assistant that can perceive, reason and interact in the physical world has attracted attention from researchers across different fields in computer vision [7, 23, 37, 38], human-computer interaction [6, 8, 15, 29], robotics [5, 32], and industrial practitioners. For example, AR Guides [1], which aims to guide users to complete complex tasks, has become popular with the development of augmented reality (AR) devices. However, existing systems often rely on pre-defined instructions or formulate the virtual assistant as a question answering [37, 38] or video under", + "bbox": [ + 496, + 750, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Co-first authors; †Work done at Microsoft", + "bbox": [ + 107, + 886, + 336, + 898 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "20270", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "standing problem [7, 23] without real-world interaction.", + "bbox": [ + 76, + 90, + 444, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In another line of work, researchers have developed simulation environments like Habitat [21, 36], VirtualHome [26], and AI2-Thor [17] to build AI agents that can interact with the physical world and collaboratively achieve new tasks [42]. Still, a large gap remains in transferring these agents to the real world, and the interaction between agents is largely simplified compared to real-world human interaction.", + "bbox": [ + 75, + 106, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on the challenges of developing intelligent agents that share perspectives with humans and interactively guide human users through performing tasks in the physical world. As a first step, we introduce HoloAssist, a large-scale egocentric human interaction dataset to explore and identify the open problems in this direction. As shown in Figure 1, the task performer wears an AR headset* to capture data while completing the tasks. An instructor watches the real-time egocentric video feed remotely and verbally guides the performer. We have developed and open-sourced a data capture tool [3] using a distributed server-client setup to enable data streaming and multimodal data capture.", + "bbox": [ + 75, + 212, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HoloAssist contains 166 hours of data captured by 222 diverse participants forming 350 unique instructor-performer pairs and carrying out 20 object-centric manipulation tasks. The objects range from common electronic devices to rare objects in factories and specialized labs. The tasks are generally challenging for first-time participants, requiring instructor assistance for successful completion. Seven raw sensor modalities are captured, including RGB, depth, head pose, 3D hand pose, eye gaze, audio, and IMU, to aid in the understanding of human intentions, estimating world states, predicting future actions, and so on. Finally, the dataset is augmented with third-person manual annotations consisting of a text summary, intervention types, mistake annotation, and action segments of the videos as illustrated in Figure 2.", + "bbox": [ + 75, + 393, + 470, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We have observed several characteristics demonstrated by human instructors from HoloAssist. First, instructors are often proactive with precisely timed interventions. Instead of waiting until mistakes happen, instructors provide follow-up instructions when the task performer appears confused. Second, the verbal guidance from the instructors tends to be concise and grounded in the task performer's environment. The instructions are often framed as spatial deictics to aid the task performer in spatial directions and distances in the 3D world. Moreover, instructors often have a good world model estimation and can detect whether mistakes disrupt task completion and then adjust the guidance.", + "bbox": [ + 75, + 604, + 470, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We take a step further and introduce new tasks and benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we conjecture are essential modules for an intelligent assistant. Additionally, we benchmark the dataset on action classification and anticipation tasks and provide empirical results to understand the role of", + "bbox": [ + 75, + 787, + 470, + 878 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e59c7abb08bdb1905e93e9f0daa7a4c7c694db4ee2cfec00f526720044a5cacf.jpg", + "image_caption": [ + "Figure 2: HoloAssist includes action and conversational annotations, in addition to text summaries of the videos, to indicate the mistakes and interventions in task completion. mistake or correct attributes are associated with each fine-grained action. A purpose label is associated with every utterance to indicate the type of verbal intervention." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 888, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "different modalities in various tasks. We hope our dataset, findings, and tooling can inspire and provide rich resources for future work on designing interactive AI assistants and situated AI assistance applications in the real world.", + "bbox": [ + 496, + 446, + 893, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 523, + 640, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work closely connects with several lines of work in computer vision, especially egocentric vision, embodied AI, and human-computer interaction.", + "bbox": [ + 496, + 544, + 893, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interactive AI assistants. Building interactive agents that can assist humans to carry out tasks in the world—real or virtual—has been a long-standing problem in different areas of AI and HCI [6, 11, 15, 22, 25, 29, 30]. As far back as 1997, Johnson and Rickel introduced \"Steve\" [15], an early pedagogical agent that aims to help students learn procedural tasks in VR. Recent efforts have focused on new modeling approaches and data collection techniques for training conversational task guidance assistants, such as model-in-the-loop wizard-of-oz [22] and human-human interaction to mimic robot actions in simulated environments [25]. In this work, we revisit this problem and provide a systematic study of real-world human interaction, and we also provide rich sensor information to push the frontiers of the research.", + "bbox": [ + 496, + 594, + 893, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Egocentric video datasets. Egocentric perspectives often convey rich information about the users' intentions. A shared perspective between the users and the human or AI assistants is useful for the assistants to provide more timely and grounded guidance. In computer vision, several egocentric video datasets [12, 14, 18, 20, 28, 31, 39] have emerged in", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "*We use HoloLens 2 [2] for data capture in this work.", + "bbox": [ + 99, + 887, + 383, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20271", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d315d1ce8ba45d9afd6d5dff34e8b2d7fb05c5687289a28292445d0e73f834d2.jpg", + "image_caption": [ + "Figure 3: HoloAssist was collected by participants diverse in ages, occupations, genders, and geography. This helps us to study a diverse set of users with different backgrounds." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 450, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the community. EPIC-KICHENS [12] is a widely adopted egocentric video dataset capturing kitchen activities. The recent Ego4D dataset [14] is the largest egocentric video dataset in the wild that provides a comprehensive database for egocentric perception in the 3D world. In contrast to earlier egocentric video datasets, HoloAssist features a multiperson interactive task completion setting, where human interaction during the procedure provides a rich source for designing AI assistants to be more proactive and grounded in the environment. Yet our work can benefit from the rich knowledge and representation learned from existing datasets like Ego4D and is complementary in nature.", + "bbox": [ + 75, + 347, + 472, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mistake detection. One of the key observations in human interaction is that human assistants tend to correct mistakes and proactively intervene in the task completion procedure. While there has been a large body of work for video-based anomaly detection [24, 40, 43, 44], mistake detection in procedural settings has been under-explored. The Assembly101 dataset [31] proposes a mistake detection task to predict if a coarse-grained action segment is a mistake or correction. By contrast, HoloAssist emphasizes fine-grained actions since instructors may intervene when they spot a student's mistake in an active intervention setting rather than wait until the whole step (i.e., coarse-grained action) is completed. In addition, we propose a new intervention prediction task and, in combination, enable a more comprehensive understanding of interactions in an assistive task completion setting.", + "bbox": [ + 75, + 532, + 470, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodality and interaction. Human interaction with the world is multimodal as we see, speak, and touch objects in the environment. In HoloAssist, we collect seven raw sensor modalities that might help understand humans' intentions, estimate the world states, predict future actions, etc. Previous datasets [12, 14, 18, 39] often provide a limited subset of modalities. Although not every sensor may currently be relevant for the downstream tasks, the seven synchronized sensor modalities provided in HoloAssist will", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c07a3d583e7ac31dc183a8e1ea9d4e067f35775db90e5275f3c2fdd86d28a633.jpg", + "table_caption": [ + "OBJECT SCALES OBJECT CATEGORIES" + ], + "table_footnote": [], + "table_body": "
SmallGoPro, Nintendo Switch, DSLR
MediumPortable printer, Computer, Nespresso machine
BigStandalone printer, big coffee machine, IKEA fur-niture (stool, utility cart, tray Table, nightstand)
RareNavVis laser scanner, ATV motor cycle, wheel belt, circuit breaker
", + "bbox": [ + 503, + 106, + 893, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: HoloAssist includes 16 objects with diverse scales. Apart from common objects used in daily life, HoloAssist includes rare equipment from mechanical labs. 20 tasks are object-centric manipulation tasks for each object and the 4 IKEA furniture has both assembly and disassembly tasks.", + "bbox": [ + 496, + 233, + 893, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "give practitioners more potential for designing multimodal agents and models even beyond the scope of this work.", + "bbox": [ + 496, + 321, + 890, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Embodied simulation platforms. There is an emerging interest in embodied agents that can perceive, reason, and act in the 3D world. Researchers [17, 21, 26, 34, 36, 42] build various simulation environments to learn such embodied agents. IGLU [42] aims to build interactive agents that learn to solve a task while being provided with grounded natural language instructions in a collaborative environment based on Minecraft, a popular video game. HoloAssist complements this line of work by providing more realistic human interaction and real-world sensor perception.", + "bbox": [ + 496, + 354, + 893, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. HoloAssist: Human Assistance Dataset", + "text_level": 1, + "bbox": [ + 500, + 515, + 849, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we introduce HoloAssist which features a two-person collaboration scenario and can be used to situate AI assistance in the physical world. We will start by describing the data collection and statistics in Section 3.1 and annotations in Section 3.2, before diving into the observations and benchmarks in the following sections.", + "bbox": [ + 496, + 537, + 893, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection and Statistics", + "text_level": 1, + "bbox": [ + 500, + 632, + 764, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tasks and objects. We consider multi-step goal-oriented tasks involving 16 objects ranging from familiar objects often used in daily life to rare objects sometimes used in labs and factories as summarized in Table 1. We consider small electronics like a GoPro, DSLR camera, and Nintendo, office appliances like a Nespresso machine and printer, IKEA furniture, and objects in labs such as a laser scanner, motor cycle, and circuit breaker. We have designed 20 tasks involving physical manipulation of these objects, e.g., changing batteries, changing belts, furniture assembly, machine setup, etc. There is one task per object except for the IKEA furniture, which has assembly and disassembly tasks. Detailed task instructions are in supplementary materials.", + "bbox": [ + 496, + 655, + 893, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Participants and collection procedure. We recruited 222 participants to form 350 unique pairs of instructors and performers for data collection. Figure 3 shows the demograph-", + "bbox": [ + 496, + 854, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20272", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d5ccbe40edbf69e07f487a1f0638393bb22c06e1b99f0dfe1b8de8a310128bc5.jpg", + "image_caption": [ + "Figure 4: The skill level of participants (0-10) for the tasks is self-reported by the participants. The skill levels roughly reflect the length of the sessions though they might be noisy." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 467, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ics of the participants. Before data collection, the participants review the IRB forms to acknowledge the privacy and ethics standards (more details in supplementary materials). The instructors are informed about the task in detail. The participants playing the role of performers are only given a rough description of the tasks and scenarios beforehand and interacted with the objects based on their understanding. The instructors provide verbal guidance as the performers set out to complete the tasks.", + "bbox": [ + 75, + 342, + 470, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Figure 4 (top), we show the distribution of the performers' familiarity with the tasks measured by a self-reported score (0-10) by the participants. We show the average length and the outliers of the recorded sessions in Figure 4 (bottom) to give a rough idea of how the participant's skill levels may lead to increased session variance. The participants' diverse skill levels and backgrounds provide rich information about the user behaviors and diverse interaction between the instructors and performers.", + "bbox": [ + 75, + 479, + 470, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data capture tool. We leveraged the Platform for Situated Intelligence framework [10] to develop and open-source a distributed application for data capture using HoloLens 2 [3]. A client process running on the device captured the sensor data while displaying a rectangular hologram frame around the user's visual field of view to guide their attention downward and keep the task actions in view of the sensors. Sensor data was streamed live over the network to a server application that ran on a PC and persisted the data to disk. This distributed setup allows for collecting longer uninterrupted sessions without reaching the device storage capacity limits.", + "bbox": [ + 75, + 622, + 470, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison with other datasets. While there is no direct comparison of datasets with the same setup with HoloAssist, we list out different aspects of our dataset and compare it with related datasets in Table 2. HoloAssist is among the largest egocentric video datasets and features a multi-person collaboration setting, which is a unique addition to the field. In addition, HoloAssist relates to work on multi-agent collab-", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2167e90cae7efab397313ab6c45ebb97871cd1263b186cc50add8293825b9f9a.jpg", + "image_caption": [ + "Figure 5: Data distribution of 166 hours captured by HoloAssist. (left) number of sessions per activity, and (right) total length of sessions in minutes." + ], + "image_footnote": [], + "bbox": [ + 506, + 93, + 890, + 214 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/baf6155d9b16c2a2a27ce778542a091da73d08cf84dababc4af4b1fb31da6ab0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSettingsCollaborative &InteractiveInstructional &Procedural# real video hours
Epic-Kitchen-100 [12]CookingXX100
Assembly101 [31]Toy assemblyX✓167
Ego4D [14]Daily-life task++3,670
VirtualHome [26]Household taskX✓$
ALFRED [33]Household taskX✓$
Habitat [36]Home assistanceX✓$
BEHAVIOR [34]Daily-life task✓✓$
IGLU [16]Collaborative building*✓✓$
TEACH [25]Household task✓✓$
HoloAssist (ours)Assistive task✓✓166
", + "bbox": [ + 504, + 287, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: Comparison to related datasets and simulation platforms. HoloAssist features a multi-person collaborative setting which is a unique addition to existing egocentric datasets in the real world. HoloAssist provides a set of instructional and procedure videos with multi-turn dialogues. Procedure tasks are defined as following a set of defined steps or procedures to achieve a specific goal, deviation from the procedure can be construed to be a mistake. HoloAssist spans 166 hours and 2,221 sessions. §: simulation, *: Minecraft-like, +: partially included.", + "bbox": [ + 498, + 487, + 893, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "orative simulation environments with a distinct characteristic of real-world sensor data and real-world human interaction.", + "bbox": [ + 498, + 667, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Annotations", + "text_level": 1, + "bbox": [ + 500, + 709, + 632, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better understand the actions and interactions in the dataset, we provide several sets of third-person manual annotations for text summaries, action segments, mistake attributes, and intervention attributes.", + "bbox": [ + 498, + 729, + 893, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language annotations. We asked the annotators to watch the video and write a paragraph to describe the activities in the videos. The description focuses on describing the hand actions in the procedure. The third-person post hoc summary provides insights into the key moments during the interactive task completion. These could be used to build a comprehensive set of instructions for task completion. We", + "bbox": [ + 498, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20273", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/375558581e7d68ca45d5197ce1d72ed02d74bf0b5c65005114d38c3d1044f44a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 348, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/819b3b8f63b91102c59be1acb7da84f6bc09cfc260b9b339f28b39029b4ca4af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 92, + 614, + 238 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c9bf663b27903c482cd3796d12cf365ae1042926e1f93e99a58b0887f7d210cc.jpg", + "image_caption": [ + "Figure 6: Data distribution of the coarse-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns." + ], + "image_footnote": [], + "bbox": [ + 622, + 114, + 888, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b72b865bf72ee6741949174835bc214a4abffbf66a45dd99f28e8baec40ffb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 316, + 348, + 439 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5eb05a5bb82c1824cca9087c0f5fccead80dd303fd6545cb855a9b40fbbb269b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 316, + 614, + 459 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dba5cfe3da510c226c789124af38c68913adf962104a4d000a36313a43a88442.jpg", + "image_caption": [ + "Figure 7: Data distribution of the fine-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns. We can see that most fine-grained actions last less than 2 seconds, and there is a long tail distribution in actions, verbs, and nouns." + ], + "image_footnote": [], + "bbox": [ + 620, + 316, + 888, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "also provide the transcriptions of the conversations in the video. With this set of annotations, we can understand the difference between third-person post hoc summaries and real-time conversations during the activities. More examples are in supplementary materials.", + "bbox": [ + 75, + 550, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse-grained action annotations. The coarse actions usually describe a high-level step in the task (e.g., change battery of a GoPro in the GoPro set up task) and can be divided into multiple fine-grained actions. To deal with the open-world setting, we ask the annotators to write a sequence (e.g., man changes the battery of GoPro) to describe the coarse actions and also identify the active verb-noun pair and optionally an adjective for the noun for benchmarking purposes (e.g., change battery). The dataset includes 414 coarse-grained actions with 90 nouns and 39 verbs. The distribution of the actions follows a long-tail distribution shown in Figure 6, where 185 actions are considered head classes while the rest are considered tail classes according to the action frequency for evaluation purposes.", + "bbox": [ + 75, + 635, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-grained action annotations. Fine-grained actions are the low-level atomic actions (e.g., press button, grab screw, etc.) for completing a step in the task, usually lasting for 1-2", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "seconds. The fine-grained actions are presented in a verb (adj.)-noun pair format. There are 1887 fine-grained actions with 165 nouns and 49 verbs. For a more comprehensive evaluation, we create a split of head actions with 1082 top actions and 805 tail actions. Distributions of fine-grained actions are shown in Figure 7.", + "bbox": [ + 496, + 550, + 892, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned earlier, noun and verb vocabularies are not pre-defined but gradually built through annotation. We ask the annotators to enter a new verb and noun if they cannot find it in the vocabulary. After the data is annotated, we ask the annotators to revisit and check the tail classes and see if they are repetitive to head classes. Due to the open-world nature, some verb and noun combinations might be interchangeable with others in the list. We show some examples in the supplementary materials.", + "bbox": [ + 496, + 648, + 892, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mistake annotation. Each fine-grained action is labeled as either correct or mistake, as indicated in Figure 2. Mistakes include the ones that are \"self-corrected by the task performers\", are \"verbally corrected by the instructors\", and \"are not corrected labeled\". Our human annotators annotate all three mistake types separately, but for benchmark evaluation, we will consolidate them into one mistake class.", + "bbox": [ + 496, + 795, + 893, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20274", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8287c72822c297f46dcda59a370f2f50ef45e8fd1d9dd794a1dcd44c15721390.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IMMEDIATE INTERVENTIONLAZY INTERVENTION
insert screwdrop allen wrench
approach buttondrop hex socket head
insert joy con controllerdrop screw
place trayscrew hex socket head
pull battery doorscrew screw
", + "bbox": [ + 101, + 89, + 450, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We defer the detailed study of differentiating whether and how the mistakes are corrected to future work. To ensure the annotation quality, we additionally ask the third-person annotators to explain why the action is a mistake and also assign a mapping to every mistake that is corrected by an instructor verbally to the conversation sentence whose type is \"instructor correcting mistakes\".", + "bbox": [ + 75, + 253, + 468, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intervention annotation. Since instructors assist the task performers verbally, we annotate the conversation between the instructors and performers to reflect the interventions in task completion. We annotate each conversation sentence with two attributes to indicate the conversation types and the conversation initiator. The conversation initiators can either be the \"task performer\" or the \"instructor\". And the sentence purpose types can be the instructor \"correcting mistakes\", \"answering questions\", \"following up with more instructions\", \"confirming previous actions\". \"describing the high-level task\", \"opening/closing remarks\", or the task performer starting the conversation to ask questions. We asked human annotators to watch the videos and use their best judgment to annotate the roles of different conversations in the videos.", + "bbox": [ + 75, + 364, + 468, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our benchmarks, we consider 3 intervention types: correcting mistakes, following up with more instructions, and confirming the previous action as they are more related to physical actions in the task procedure. Figure 2 shows examples of conversational interventions. More examples can be found in the supplementary materials.", + "bbox": [ + 75, + 594, + 468, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audit process. Annotations are done by professional annotators based on the following process. The annotators first take a pass on the video to add the fine-grained actions, coarse-grained actions, conversation, and text summary, along with the associated annotation elements for each event. After self-review, the annotated data is passed to an independent reviewer for auditing, and the mistakes are fixed directly or sent back to the original annotator for updates. The annotations finally go through a targeted review to check the open-ended text fields like narration, action sentences, and conversation transcriptions to ensure consistency. Before the annotations were delivered, we applied a list of constraints to systematically check the annotations to further dig out the wrong annotations.", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial deixis from the intervention transcripts", + "text_level": 1, + "bbox": [ + 509, + 90, + 808, + 104 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/610a33ebf1d71a8ba687d719d88b0f9897f64913e1cef3d0c6f00f3a1207224c.jpg", + "table_caption": [ + "Table 3: Top mistakes that are corrected immediately (left) or later and sometimes through self-correction (right)." + ], + "table_footnote": [], + "table_body": "
"You should press the button that's on the body of the camera just at the right of the lens."
"You should leave the bolt, like it was before."
"The button is on the other side of the Switch."
"The SD card comes from the right slot, on the right hand side and it opens by using a knob next to the screen on the right bottom side of the screen."
"Currently the tray is facing you, please rotate so the back is facing you."
"You should put it the other way around. It's upside down."
"Please start by removing the screw of the top shelf first."
"And now, ehm, it is, it should be the other one that should be on top of the other."
"To the right of that, there is a tiny little square."
", + "bbox": [ + 508, + 108, + 883, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4: Examples of deictic phrases that help in grounded guidance by specifying contextual spatial locations.", + "bbox": [ + 498, + 383, + 890, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Observations and Tasks from HoloAssist", + "text_level": 1, + "bbox": [ + 500, + 440, + 862, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we present observations from HoloAssist in Section 4.1 and identify a few open problems that are necessary components for an interactive AI assistant. In Section 4.2, we define new tasks and benchmarks with HoloAssist.", + "bbox": [ + 498, + 465, + 893, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Observations from HoloAssist", + "text_level": 1, + "bbox": [ + 500, + 537, + 766, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Correlation between mistakes and intervention. We find that the response time for the human assistant to intervene in the procedure and correct mistakes depends on the severity level of the mistakes. If mistakes are critical, the instructors proactively interrupt the student immediately (within less than 5 seconds duration) while other mistakes are either self-corrected by the task performer or corrected later by the instructor.", + "bbox": [ + 496, + 564, + 893, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3, we present the top actions that need immediate intervention and the top actions that are self-corrected by the task performers or corrected later by the instructors. We notice that actions related to linear tasks, where the task progression is stalled if steps are not followed in order, are often intervened immediately by the instructors. For example, \"insert joy con controller\", and \"place tray\" etc. In contrast, the lazily edited corrections are related to furniture assembly such as the actions \"drop allen wrench\", \"drop hex socket head\", etc. These are tasks where mistakes are unclear in every stage, and the user often intuitively adjusts their steps.", + "bbox": [ + 496, + 686, + 893, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Grounded guidance. We also notice that the instructions from human instructors are often grounded in the 3D environment. An important aspect of grounded guidance is the", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20275", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ability to communicate about the physical world by pointing to things in context. Spatial deixis refers to phrases that are used to locate things in space and to express direction and distance. The deictic analysis of the transcripts in HoloAssist reveals a wide set of words that indicate specific and relative locations and directions, especially during interventions to correct mistakes. A list of some prominent examples is shown in Table 4.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Benchmark Tasks", + "text_level": 1, + "bbox": [ + 76, + 239, + 250, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inspired by the observations above, we think it is important for an interactive AI assistant to have a good world state estimation model that can detect mistakes and predict whether to intervene in the task procedure. Besides, augmenting instructions with spatial guidance can be useful for AI agents. To this end, we introduce new mistake detection, intervention prediction and 3D hand pose forecasting tasks for interactive and grounded guidance. Additionally, we benchmark models on action recognition tasks following the convention in [12].", + "bbox": [ + 75, + 268, + 472, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mistake detection is defined following the convention [31] but applied to fine-grained actions in our benchmark. We take the features from the fine-grained action clips from the beginning of the coarse-grained action until the end of the current action clip, and the model predicts a label from {correct, mistake}. The task is challenging given that the class distribution is highly skewed, with around $6\\%$ mistakes among the fine-grained actions.", + "bbox": [ + 75, + 428, + 472, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intervention type prediction is to predict the intervention types given an input of a window of 1, 3, or 5 seconds before the intervention. This newly proposed benchmark is to test if the model can correctly figure out the correct intervention types during task completion. Currently, HoloAssist includes 3 intervention types, and we report the precision and recall of each intervention type.", + "bbox": [ + 75, + 559, + 470, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D hand pose forecasting is another new benchmark introduced by HoloAssist. Existing action forecasting work [12] mostly focuses on providing semantic labels of future actions and does not provide explicit 3D guidance on hand poses. Predicting 3D hand poses can be useful for various applications [4], and it can augment instructions and spatially guide users in different tasks. In this benchmark, we take 3 seconds inputs similar to other 3D body location forecasting literature [41] and forecast the continuous 3D hand poses for the next 0.5, 1.0, and 1.5 seconds. The evaluation metric is the average of mean per joint position error over time in centimeters compared to ground truth. To have a proper evaluation metric that can help 3D action guidance, we remove the mistakes from the action sequences and only forecast 3D hand pose for the correct labels.", + "bbox": [ + 75, + 674, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 632, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide the evaluation results of the proposed benchmarks. We will start with the standard action recognition benchmarks, and then we will present the results of the newly proposed benchmarks. We also provide ablations of different sensor modalities to understand the roles of different sensors in various tasks. We hope the baseline results can guide future research in this space.", + "bbox": [ + 496, + 116, + 893, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. We adopt TimeSformer [9], a state-of-the-art vision transformer (ViT) [13] based video model, as the backbone and change the head with a different number of classes for different benchmarks. We modify the original TimeSformers to perform multimodal learning by introducing additional tokens for different modalities and embedding layers to encode the additional sensor modalities. Specifically, we use $26 \\times 2$ tokens for both left and right hands (one token for one hand joint), one token for eye gaze, one token for head poses, and 196 tokens for depth. We can enable and/or disable different modalities during training and evaluation. Detailed configurations are available in supplementary materials. Note that we consider the resulting model a vanilla multimodal model and serve as a baseline for future studies.", + "bbox": [ + 496, + 227, + 893, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We randomly split the 2221 sessions into train, validation, and test sets following a ratio of $70\\%$ , $10\\%$ , and $20\\%$ on a per-task basis, which includes 1545 sessions for training, 213 sessions for validation, and 463 sessions for testing. We also synchronize all the modalities according to the video stream and keep the frame rate at 30 fps by sub-sampling other modalities in our experiments.", + "bbox": [ + 496, + 454, + 893, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, the model is trained for 15 epochs with an initial learning rate of 0.01 and a batch size of 64 using stochastic gradient descent. We divide the learning rate by 10 at the epoch 11 and 14. For each input segment, we randomly sample 8 frames/data points within the segment. We train our models with 4xA6000 GPU machines, and the fine-grained action recognition runs usually take about one day. We also train the model with random initialization and an ImageNet pre-trained ViT backbone.", + "bbox": [ + 496, + 561, + 893, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For our hand pose forecasting benchmark, which is a regression task, we adopt Seq2Seq model [35] following [19].", + "bbox": [ + 496, + 698, + 895, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Action recognition. We show the fine-grained action recognition results in Table 5 and the coarse-grained action recognition results in Table 6. We can see that initialized with a pre-trained ViT model on ImageNet; all models can achieve around $35\\%$ top-1 accuracy on fine-grained action recognition and $50\\%$ top-1 accuracy on coarse-grained action recognition. This is comparable to baselines of other fine-grained action benchmarks [31] ( $23\\%$ top-1 accuracy),", + "bbox": [ + 496, + 733, + 893, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We notice that the pre-trained model on ImageNet may reduce the influence of other modalities that are not pretrained. If trained from scratch, we can see from Table 5", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "footer", + "text": "20276", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb02190f86d99bac4ce16252bfc9d8110408c195083ff3c62bf7b298e805cc48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB34.83/68.6042.14/78.9666.81/90.0435.26/69.3442.56/79.5367.19/90.360.03/0.1710.86/36.3338.01/66.48
Hands20.86/43.9235.38/65.7637.10/63.4221.13/44.5035.72/66.3037.50/64.060.00/0.0110.11/25.477.30/15.54
R+H35.06/68.9542.45/79.4267.05/90.0135.49/69.7142.87/80.0367.43/90.320.03/0.1611.05/33.7138.39/66.85
R+H+E35.27/68.6942.92/79.1167.03/89.9635.72/69.4243.33/79.6767.45/90.290.03/0.1811.99/37.6435.96/65.17
R+H+E+I34.80/68.2642.24/78.8866.65/89.7635.23/69.0042.63/79.4667.03/90.080.03/0.1712.92/35.9638.20/65.17
ScratchRGB18.78/48.0928.45/65.4343.72/73.6919.03/48.7228.74/66.0044.09/74.210.00/0.016.55/22.6615.73/34.27
Hands23.94/47.5839.79/68.7639.34/65.7624.26/48.2140.18/69.3039.79/66.430.00/0.0010.67/28.655.62/16.29
RGB + Eyes20.86/50.2730.92/67.5945.28/75.2921.13/50.9331.22/68.2045.65/75.800.00/0.018.05/21.7217.60/37.08
", + "bbox": [ + 80, + 89, + 893, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6e1742f3edfc8b6590cfc46293b88ef240be4d62bb39d5b52c466f981c31c252.jpg", + "table_caption": [ + "Table 5: Fine-grained action recognition benchmark results. We report baseline results of (multimodal) TimeSformer [9] models on HoloAssist. Here R, H, I denote RGB, hands, and head pose (whose sensor source is similar as IMUs). When starting from a pre-trained ViT model on ImageNet, all the models have higher performance than training from scratch. However, with a pre-trained image model, the influence of other modalities is reduced. If the models are trained from scratch, we can see that adding hands can improve the prediction of verbs and lead to better results than the baseline RGB only model. Simply concatenating more sensors as inputs may not necessarily lead to improved performance." + ], + "table_footnote": [], + "table_body": "
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB50.91/86.8960.51/93.4573.35/95.9053.40/89.5362.78/95.3975.00/96.700.37/2.2420.00/58.8943.89/81.67
Hands22.20/54.1635.12/72.3437.43/68.3323.44/57.0836.75/75.2538.84/70.450.00/0.126.11/20.5612.22/30.56
R+H50.80/86.5459.71/93.3673.20/95.7853.27/89.1261.94/95.2075.09/96.570.37/2.2820.00/60.5639.44/81.67
R+H+E50.35/85.5158.85/93.2173.67/95.6352.65/88.0360.85/94.8975.28/96.320.53/2.2823.33/63.3345.00/83.33
R+H+E+I50.18/86.0759.00/93.2173.64/96.2552.49/88.7861.22/95.0775.41/96.910.50/2.1219.44/60.0042.22/84.44
", + "bbox": [ + 80, + 358, + 893, + 474 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e6fcef07bc401cf82a65c78022b47f933f6d3cc8c36e6605a7e769813eee32a4.jpg", + "table_caption": [ + "Table 6: Coarse-grained action recognition results. The overall trend is similar to fine-grained action recognition." + ], + "table_footnote": [], + "table_body": "
Mods.F-scoreCorrectMistake
PrecisionRecallPrecisionRecall
Random27.7160.8710.2215.0046.15
RGB35.1182.5651.8212.9626.92
Hands40.1992.6852.4112.531.25
R+H36.1885.5143.079.6811.54
R+H+E32.0888.5742.7611.4350.00
", + "bbox": [ + 99, + 516, + 447, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(bottom) that adding hands can improve the prediction of verbs and lead to better results than the RGB-only model.", + "bbox": [ + 75, + 741, + 470, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mistake detection. In Table $7^{\\dagger}$ , we present the results of mistake detection where the features are extracted from the pre-trained action recognition models shown in Table 5. Here we find the hand poses information benefits the task and outperforms the other modalities.", + "bbox": [ + 75, + 776, + 470, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intervention type prediction. For intervention prediction,", + "bbox": [ + 76, + 857, + 470, + 873 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89c68b7886cd30f48f92d09c1e4c60ac708934364c545b7716176b04547ebb9e.jpg", + "table_caption": [ + "Table 7: Mistake detection results. Similar to action recognition, adding hand poses to RGB $(\\mathbb{R} + \\mathbb{H})$ improves an RGB-only model from 35.11 to 36.18 points. What's different is that we find the hands only model outperforms other modalities in mistake detection, achieving 40.19 points." + ], + "table_footnote": [], + "table_body": "
Mods.OverallConfirmCorrectFollow-up
Prec.RecallPrec.RecallPrec.RecallPrec.Recall
RGB13.9333.330.000.000.000.0041.79100
R+H24.8933.6432.144.500.000.0042.5296.43
R+E25.5533.7333.9010.360.000.0042.7690.83
R+H+E48.3137.5939.1140.9361.119.9144.761.93
RGB(Pt)37.5437.7442.3127.5027.3336.6142.9749.11
", + "bbox": [ + 503, + 517, + 893, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: Intervention type prediction results. The classes in benchmark are highly skewed. Here E refers to eye gaze. We can see the predictions of the baseline models (i.e., RGB, R+H, R+E) favor predicting follow-up instructions while ignoring correcting mistakes (recall is 0). Training more modalities $(\\mathbb{R} + \\mathbb{H} + \\mathbb{E})$ alleviates this issue and even outperforms the RGB model (RGB (Pt)) in overall precision.", + "bbox": [ + 496, + 659, + 893, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we show the results in Table $8^{\\dagger}$ . We can see adding hands and eye gaze $(\\mathbb{R} + \\mathbb{H} + \\mathbb{E})$ can significantly boost the overall precision and recall to $48.31\\%$ and $37.59\\%$ , improving about 35 and 4 percentage points over RGB. This may be because eye gaze is a forecasting signal, as people often look at the regions before the action starts, which can assist the models to attend to important regions for better anticipation.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Evaluated on $10\\%$ of the entire data", + "bbox": [ + 101, + 886, + 297, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20277", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a74fddbfad7c04202018bde1239cb7b5bc2874a9e3084f517958d0212b8c949.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 87, + 893, + 162 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/596ddaea51e8b8524c682d58e0abdfe8d8a84c9e0e953952144dc7dac79a7f37.jpg", + "image_caption": [ + "Time", + "Figure 8: Qualitative visualization for 3D hand pose forecasting. We visualize the ground-truth hand joint positions (i.e., hand pose, visualized in Green) and the prediction of the hand pose for the next 1.5 seconds (visualized in Red). The input to the model is a 3-second long clip ahead of the prediction. The task is challenging as hands often move quickly. As we can see from the figure, the predicted hand pose is more off from the grounded truth in the longer future." + ], + "image_footnote": [], + "bbox": [ + 80, + 166, + 890, + 242 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0f6d3b71983cf546ecee133acd3b9805d2bfe00b638e8b0819d2878d6cc617ad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mods.Mean Error Distance (cm, ↓)
0.5 sec1.0 sec1.5 sec
Static-H9.3413.9116.70
Hands9.8010.6811.25
H+E9.8010.7011.25
H+E+I9.8010.6911.25
R+H+E9.7310.6511.22
R+H+E+I9.7210.6211.19
", + "bbox": [ + 101, + 353, + 450, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 9: 3D hand pose forecasting benchmark results. We report the mean per joint position error for 0.5 1.0, 1.5 seconds (lower the better). The static hand baseline (Static-H) refers to using the last frame of the input. The Seq2Seq [35] model trained using hands only H, and using only hand H achieves better results than Static-H.", + "bbox": [ + 75, + 491, + 470, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3D hand pose forecasting. Table 9 shows that given only the 3D poses of hands as input, the model can perform with the accuracy of 9.80, 10.68, and 11.25 centimeters, the average of mean per joint position error, for 0.5, 1, and 1.5 seconds, respectively. We should note that this task is challenging as hands can move quickly within a window of 0.5 seconds. Compared to static-H, which uses the last 3D hand pose, our baseline (H) already outperforms it. In Figure 8, we show the visualization of the hand pose forecasting.", + "bbox": [ + 75, + 593, + 470, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Importance of hand poses and eye gaze. As we can already see from the Tables 5, 7, 8, 9, the 3D hand poses and eye gaze can help the model prediction to recognize actions, detect mistakes, and understand users' intentions. These modalities can augment the commonly used RGB images for better performance. Simply concatenating more modalities (e.g., depth, head poses) as inputs may not necessarily lead to a more capable model, as those modalities may need specialized encoders or model architectures to process them simultaneously. We believe HoloAssist will enable and foster further research in multi-modal learning in this direction.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 354, + 769, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we identified and explored several important problems with building an interactive AI assistant in the physical world. We introduced a large-scale multimodal egocentric video dataset, HoloAssist, containing rich information about human interaction in an assistive task completion setting. The task performer wears a HoloLens 2 headset while completing various object-centric manipulation tasks. The real-time video feed from the headset is sent to a remote instructor who provides verbal guidance to the task performer. HoloAssist captures seven raw sensor modalities during the interaction, and among them, we found hand pose and eye gaze are useful information sources for an interactive AI agent. By augmenting the data with additional third-person manual annotations on action segments, mistakes, and intervention types, we constructed new benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we believe is a necessary component for an interactive and grounded AI assistant. As a first step in this direction, this work also leaves room for future work to improve upon (e.g., annotating object poses in the data, investigating object-centric models of affordance and manipulations in AI assistance, etc.). We believe HoloAssist, coupled with the associated benchmarks and tooling will benefit future research into building competent AI assistants for everyday tasks in the real world.", + "bbox": [ + 496, + 382, + 895, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 777, + 648, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank all the 222 data collectors who participated in the study and acknowledge the hard work of the annotation team led by Megan Yuan and Dan Luo from DataTang Technology Inc. We thank Nick Saw for the help in the data collection software and Yale Song and Vibhav Vineet from the Computer Vision Group at Microsoft Research for the early discussion. We also thank the feedback from colleagues at MSR.", + "bbox": [ + 496, + 803, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20278", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Microsoft dynamics 365 guides. https://dynamics.microsoft.com/en-us/mixed-reality/guides/.1", + "[2] Microsoft hololens. https://www.microsoft.com/en-us/hololens.2", + "[3] Platform for situated intelligence hololens capture tools. https://github.com/microsoft/psi/tree/master/Sources/MixedReality/HoloLensCapture.2,4", + "[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 7", + "[5] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 1", + "[6] Gregory Aist, John Dowding, Beth Ann Hockey, Manny Rayner, James Hieronymus, D Bohus, B Boven, Nate Blaylock, Ellen Campana, S Early, et al. Talking through procedures: An intelligent space station procedure assistant. In Proceedings of the tenth conference on European chapter of the Association for Computational Linguistics-Volume 2, pages 187-190, 2003. 1, 2", + "[7] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Ivan Laptev, Josef Sivic, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2", + "[8] Sean Andrist, Michael Gleicher, and Bilge Mutlu. Looking coordinated: Bidirectional gaze mechanisms for collaborative interaction with virtual characters. In Proceedings of the 2017 CHI conference on human factors in computing systems, pages 2571-2582, 2017. 1", + "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 7, 8", + "[10] Dan Bohus, Sean Andrist, Ashley Feniello, Nick Saw, Mihai Jalobeanu, Patrick Sweeney, Anne Loomis Thompson, and Eric Horvitz. Platform for situated intelligence, 2021. 4", + "[11] Dan Bohus and Alexander I Rudnicky. Larri: A language-based maintenance and repair assistant. Spoken multimodal human-computer dialogue in mobile environments, pages 203-218, 2005. 2", + "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonio Furnari, Jian Ma, Evangelos Kazakos, Davide Molti" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "santi, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 3, 4, 7", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7", + "[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4", + "[15] W Lewis Johnson and Jeff Rickel. Steve: An animated pedagogical agent for procedural training in virtual environments. ACM SIGART Bulletin, 8(1-4):16-21, 1997. 1, 2", + "[16] Julia Kiseleva, Ziming Li, Mohammad Aliannejadi, Shrestha Mohanty, Maartje ter Hoeve, Mikhail Burtsev, Alexey Skrynnik, Artem Zholus, Aleksandr Panov, Kavya Srinet, et al. Interactive grounded language understanding in a collaborative environment: Iglu 2021. In NeurIPS 2021 Competitions and Demonstrations Track, pages 146-161. PMLR, 2022. 4", + "[17] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2, 3", + "[18] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3", + "[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20279", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 7", + "[20] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2", + "[21] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3", + "[22] Ramesh Manuvinakurike, Sovan Biswas, Giuseppe Raffa, Richard Beckwith, Anthony Rhodes, Meng Shi, Gesem Gudino Mejia, Saurav Sahay, and Lama Nachman. Human in the loop approaches in multi-modal conversational task guidance system development. arXiv preprint arXiv:2211.01824, 2022. 2", + "[23] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl;dw? summarizing instructional videos with task relevance and cross-modal saliency. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 540–557, 2022. 1, 2", + "[24] Rashmiranjan Nayak, Umesh Chandra Pati, and Santos Kumar Das. A comprehensive review on deep learning-based methods for video anomaly detection. Image and Vision Computing, 106:104078, 2021. 3", + "[25] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2017-2025, 2022. 2, 4", + "[26] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8494-8502, 2018. 2, 3, 4", + "[27] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, June 2019. 1", + "[28] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 2", + "[29] Charles Rich and Candace L Sidner. Diamondhelp: A generic collaborative task guidance system. AI Magazine, 28(2):33-33, 2007. 1, 2", + "[30] Charles Rich, Candace L Sidner, and Neal Lesh. Collagen: Applying collaborative discourse theory to human-computer interaction. AI magazine, 22(4):15-15, 2001. 2" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] F. Sener, D. Chatterjee, D. Shelepop, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. CVPR 2022. 2, 3, 4, 7", + "[32] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022. 1", + "[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4", + "[34] Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martin-Martin, Fei Xia, Kent Elliott Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, Karen Liu, et al. Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In Conference on Robot Learning, pages 477-490. PMLR, 2022. 3, 4", + "[35] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27, 2014. 7, 9", + "[36] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 3, 4", + "[37] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning (CoRL), 2019. 1", + "[38] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, pages 485-501. Springer, 2022. 1", + "[39] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), Oct. 2022. 2, 3", + "[40] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 3", + "[41] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 7", + "[42] Artem Zholus, Alexey Skrynnik, Shrestha Mohanty, Zoya Volovikova, Julia Kiseleva, Artur Szlam, Marc-Alexandre Cote, and Aleksandr I Panov. Iglu gridworld: Simple and fast environment for embodied dialog agents. arXiv preprint arXiv:2206.00142, 2022. 2, 3", + "[43] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 14(10):2537-2550, 2019. 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20280", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[44] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner,", + "bbox": [ + 76, + 90, + 470, + 133 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "editors, Computer Vision - ECCV 2022, pages 395-412, 2022. 3", + "bbox": [ + 532, + 90, + 893, + 118 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20281", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_model.json b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a525d8dc40d10bfcef5a3a03c526ea2a3132057b --- /dev/null +++ b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_model.json @@ -0,0 +1,2204 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.213, + 0.131, + 0.761, + 0.175 + ], + "angle": 0, + "content": "HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.914, + 0.257 + ], + "angle": 0, + "content": "Xin Wang\\(^{1*}\\) Taein Kwon\\(^{1,2*}\\) Mahdi Rad\\(^{1}\\) Bowen Pan\\(^{1\\dagger}\\) Ishani Chakraborty\\(^{1}\\) Sean Andrist\\(^{1}\\) Dan Bohus\\(^{1}\\) Ashley Feniello\\(^{1}\\) Bugra Tekin\\(^{1\\dagger}\\) Felipe Vieira Frujeri\\(^{1}\\) Neel Joshi\\(^{1}\\) Marc Pollefeys\\(^{1,2}\\) Microsoft \\({}^{2}\\) ETH Zurich" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.324, + 0.474, + 0.702 + ], + "angle": 0, + "content": "Building an interactive AI assistant that can perceive, reason, and collaborate with humans in the real world has been a long-standing pursuit in the AI community. This work is part of a broader research effort to develop intelligent agents that can interactively guide humans through performing tasks in the physical world. As a first step in this direction, we introduce HoloAssist, a large-scale egocentric human interaction dataset, where two people collaboratively complete physical manipulation tasks. The task performer executes the task while wearing a mixed-reality headset that captures seven synchronized data streams. The task instructor watches the performer's egocentric video in real time and guides them verbally. By augmenting the data with action and conversational annotations and observing the rich behaviors of various participants, we present key insights into how human assistants correct mistakes, intervene in the task completion procedure, and ground their instructions to the environment. HoloAssist spans 166 hours of data captured by 350 unique instructor-performer pairs. Furthermore, we construct and present benchmarks on mistake detection, intervention type prediction, and hand forecasting, along with detailed analysis. We expect HoloAssist will provide an important resource for building AI assistants that can fluidly collaborate with humans in the real world. Data can be downloaded at https://holoassist.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.731, + 0.21, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.471, + 0.878 + ], + "angle": 0, + "content": "Recent years have witnessed incredible progress in general-purpose AI agents that assist humans with various open-world tasks, especially in the digital world. AI systems powered by large language models (LLMs) like ChatGPT [27] can answer users' questions and assist them with various text-based tasks. However, these AI assistants do not have sufficient first-hand experience in the physical world and thus cannot perceive world states and actively intervene" + }, + { + "type": "title", + "bbox": [ + 0.558, + 0.297, + 0.834, + 0.309 + ], + "angle": 0, + "content": "Two-Person Interactive Assistive Task Completion" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.311, + 0.697, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.311, + 0.891, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.509, + 0.607, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.577, + 0.61, + 0.599 + ], + "angle": 0, + "content": "(1) RGB, (2) eye gaze, (3) hand pose" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.509, + 0.714, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.577, + 0.698, + 0.599 + ], + "angle": 0, + "content": "(4) head pose, (5) depth" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.509, + 0.832, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.582, + 0.816, + 0.593 + ], + "angle": 0, + "content": "(6) IMU, (7) audio" + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.512, + 0.89, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.843, + 0.582, + 0.879, + 0.592 + ], + "angle": 0, + "content": "(8) text" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.719 + ], + "angle": 0, + "content": "Figure 1: HoloAssist features a two-person interactive assistive task completion setting. The task performer wears an AR device and completes the tasks while the captured data is streamed over the network to a remote instructor watching it on the laptop. The instructor provides verbal guidance to the student. HoloAssist includes seven modalities captured live and human annotated text descriptions as the 8th modality." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.734, + 0.722, + 0.749 + ], + "angle": 0, + "content": "in the task completion procedure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Building an AI assistant that can perceive, reason and interact in the physical world has attracted attention from researchers across different fields in computer vision [7, 23, 37, 38], human-computer interaction [6, 8, 15, 29], robotics [5, 32], and industrial practitioners. For example, AR Guides [1], which aims to guide users to complete complex tasks, has become popular with the development of augmented reality (AR) devices. However, existing systems often rely on pre-defined instructions or formulate the virtual assistant as a question answering [37, 38] or video under" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.338, + 0.9 + ], + "angle": 0, + "content": "*Co-first authors; †Work done at Microsoft" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "20270" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.446, + 0.106 + ], + "angle": 0, + "content": "standing problem [7, 23] without real-world interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.471, + 0.212 + ], + "angle": 0, + "content": "In another line of work, researchers have developed simulation environments like Habitat [21, 36], VirtualHome [26], and AI2-Thor [17] to build AI agents that can interact with the physical world and collaboratively achieve new tasks [42]. Still, a large gap remains in transferring these agents to the real world, and the interaction between agents is largely simplified compared to real-world human interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.394 + ], + "angle": 0, + "content": "In this work, we focus on the challenges of developing intelligent agents that share perspectives with humans and interactively guide human users through performing tasks in the physical world. As a first step, we introduce HoloAssist, a large-scale egocentric human interaction dataset to explore and identify the open problems in this direction. As shown in Figure 1, the task performer wears an AR headset* to capture data while completing the tasks. An instructor watches the real-time egocentric video feed remotely and verbally guides the performer. We have developed and open-sourced a data capture tool [3] using a distributed server-client setup to enable data streaming and multimodal data capture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.605 + ], + "angle": 0, + "content": "HoloAssist contains 166 hours of data captured by 222 diverse participants forming 350 unique instructor-performer pairs and carrying out 20 object-centric manipulation tasks. The objects range from common electronic devices to rare objects in factories and specialized labs. The tasks are generally challenging for first-time participants, requiring instructor assistance for successful completion. Seven raw sensor modalities are captured, including RGB, depth, head pose, 3D hand pose, eye gaze, audio, and IMU, to aid in the understanding of human intentions, estimating world states, predicting future actions, and so on. Finally, the dataset is augmented with third-person manual annotations consisting of a text summary, intervention types, mistake annotation, and action segments of the videos as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.471, + 0.787 + ], + "angle": 0, + "content": "We have observed several characteristics demonstrated by human instructors from HoloAssist. First, instructors are often proactive with precisely timed interventions. Instead of waiting until mistakes happen, instructors provide follow-up instructions when the task performer appears confused. Second, the verbal guidance from the instructors tends to be concise and grounded in the task performer's environment. The instructions are often framed as spatial deictics to aid the task performer in spatial directions and distances in the 3D world. Moreover, instructors often have a good world model estimation and can detect whether mistakes disrupt task completion and then adjust the guidance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.471, + 0.879 + ], + "angle": 0, + "content": "We take a step further and introduce new tasks and benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we conjecture are essential modules for an intelligent assistant. Additionally, we benchmark the dataset on action classification and anticipation tasks and provide empirical results to understand the role of" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.89, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.34, + 0.895, + 0.431 + ], + "angle": 0, + "content": "Figure 2: HoloAssist includes action and conversational annotations, in addition to text summaries of the videos, to indicate the mistakes and interventions in task completion. mistake or correct attributes are associated with each fine-grained action. A purpose label is associated with every utterance to indicate the type of verbal intervention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.894, + 0.508 + ], + "angle": 0, + "content": "different modalities in various tasks. We hope our dataset, findings, and tooling can inspire and provide rich resources for future work on designing interactive AI assistants and situated AI assistance applications in the real world." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.642, + 0.54 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.894, + 0.59 + ], + "angle": 0, + "content": "Our work closely connects with several lines of work in computer vision, especially egocentric vision, embodied AI, and human-computer interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.806 + ], + "angle": 0, + "content": "Interactive AI assistants. Building interactive agents that can assist humans to carry out tasks in the world—real or virtual—has been a long-standing problem in different areas of AI and HCI [6, 11, 15, 22, 25, 29, 30]. As far back as 1997, Johnson and Rickel introduced \"Steve\" [15], an early pedagogical agent that aims to help students learn procedural tasks in VR. Recent efforts have focused on new modeling approaches and data collection techniques for training conversational task guidance assistants, such as model-in-the-loop wizard-of-oz [22] and human-human interaction to mimic robot actions in simulated environments [25]. In this work, we revisit this problem and provide a systematic study of real-world human interaction, and we also provide rich sensor information to push the frontiers of the research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Egocentric video datasets. Egocentric perspectives often convey rich information about the users' intentions. A shared perspective between the users and the human or AI assistants is useful for the assistants to provide more timely and grounded guidance. In computer vision, several egocentric video datasets [12, 14, 18, 20, 28, 31, 39] have emerged in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.385, + 0.9 + ], + "angle": 0, + "content": "*We use HoloLens 2 [2] for data capture in this work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "20271" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.452, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.471, + 0.332 + ], + "angle": 0, + "content": "Figure 3: HoloAssist was collected by participants diverse in ages, occupations, genders, and geography. This helps us to study a diverse set of users with different backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.529 + ], + "angle": 0, + "content": "the community. EPIC-KICHENS [12] is a widely adopted egocentric video dataset capturing kitchen activities. The recent Ego4D dataset [14] is the largest egocentric video dataset in the wild that provides a comprehensive database for egocentric perception in the 3D world. In contrast to earlier egocentric video datasets, HoloAssist features a multiperson interactive task completion setting, where human interaction during the procedure provides a rich source for designing AI assistants to be more proactive and grounded in the environment. Yet our work can benefit from the rich knowledge and representation learned from existing datasets like Ego4D and is complementary in nature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.472, + 0.761 + ], + "angle": 0, + "content": "Mistake detection. One of the key observations in human interaction is that human assistants tend to correct mistakes and proactively intervene in the task completion procedure. While there has been a large body of work for video-based anomaly detection [24, 40, 43, 44], mistake detection in procedural settings has been under-explored. The Assembly101 dataset [31] proposes a mistake detection task to predict if a coarse-grained action segment is a mistake or correction. By contrast, HoloAssist emphasizes fine-grained actions since instructors may intervene when they spot a student's mistake in an active intervention setting rather than wait until the whole step (i.e., coarse-grained action) is completed. In addition, we propose a new intervention prediction task and, in combination, enable a more comprehensive understanding of interactions in an assistive task completion setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Multimodality and interaction. Human interaction with the world is multimodal as we see, speak, and touch objects in the environment. In HoloAssist, we collect seven raw sensor modalities that might help understand humans' intentions, estimate the world states, predict future actions, etc. Previous datasets [12, 14, 18, 39] often provide a limited subset of modalities. Although not every sensor may currently be relevant for the downstream tasks, the seven synchronized sensor modalities provided in HoloAssist will" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.094, + 0.744, + 0.106 + ], + "angle": 0, + "content": "OBJECT SCALES OBJECT CATEGORIES" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.107, + 0.895, + 0.224 + ], + "angle": 0, + "content": "
SmallGoPro, Nintendo Switch, DSLR
MediumPortable printer, Computer, Nespresso machine
BigStandalone printer, big coffee machine, IKEA fur-niture (stool, utility cart, tray Table, nightstand)
RareNavVis laser scanner, ATV motor cycle, wheel belt, circuit breaker
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.234, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Table 1: HoloAssist includes 16 objects with diverse scales. Apart from common objects used in daily life, HoloAssist includes rare equipment from mechanical labs. 20 tasks are object-centric manipulation tasks for each object and the 4 IKEA furniture has both assembly and disassembly tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.352 + ], + "angle": 0, + "content": "give practitioners more potential for designing multimodal agents and models even beyond the scope of this work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.895, + 0.507 + ], + "angle": 0, + "content": "Embodied simulation platforms. There is an emerging interest in embodied agents that can perceive, reason, and act in the 3D world. Researchers [17, 21, 26, 34, 36, 42] build various simulation environments to learn such embodied agents. IGLU [42] aims to build interactive agents that learn to solve a task while being provided with grounded natural language instructions in a collaborative environment based on Minecraft, a popular video game. HoloAssist complements this line of work by providing more realistic human interaction and real-world sensor perception." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.516, + 0.851, + 0.531 + ], + "angle": 0, + "content": "3. HoloAssist: Human Assistance Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.629 + ], + "angle": 0, + "content": "In this work, we introduce HoloAssist which features a two-person collaboration scenario and can be used to situate AI assistance in the physical world. We will start by describing the data collection and statistics in Section 3.1 and annotations in Section 3.2, before diving into the observations and benchmarks in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.766, + 0.647 + ], + "angle": 0, + "content": "3.1. Data Collection and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.852 + ], + "angle": 0, + "content": "Tasks and objects. We consider multi-step goal-oriented tasks involving 16 objects ranging from familiar objects often used in daily life to rare objects sometimes used in labs and factories as summarized in Table 1. We consider small electronics like a GoPro, DSLR camera, and Nintendo, office appliances like a Nespresso machine and printer, IKEA furniture, and objects in labs such as a laser scanner, motor cycle, and circuit breaker. We have designed 20 tasks involving physical manipulation of these objects, e.g., changing batteries, changing belts, furniture assembly, machine setup, etc. There is one task per object except for the IKEA furniture, which has assembly and disassembly tasks. Detailed task instructions are in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Participants and collection procedure. We recruited 222 participants to form 350 unique pairs of instructors and performers for data collection. Figure 3 shows the demograph-" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "20272" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.468, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.472, + 0.312 + ], + "angle": 0, + "content": "Figure 4: The skill level of participants (0-10) for the tasks is self-reported by the participants. The skill levels roughly reflect the length of the sessions though they might be noisy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.472, + 0.478 + ], + "angle": 0, + "content": "ics of the participants. Before data collection, the participants review the IRB forms to acknowledge the privacy and ethics standards (more details in supplementary materials). The instructors are informed about the task in detail. The participants playing the role of performers are only given a rough description of the tasks and scenarios beforehand and interacted with the objects based on their understanding. The instructors provide verbal guidance as the performers set out to complete the tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.472, + 0.619 + ], + "angle": 0, + "content": "In Figure 4 (top), we show the distribution of the performers' familiarity with the tasks measured by a self-reported score (0-10) by the participants. We show the average length and the outliers of the recorded sessions in Figure 4 (bottom) to give a rough idea of how the participant's skill levels may lead to increased session variance. The participants' diverse skill levels and backgrounds provide rich information about the user behaviors and diverse interaction between the instructors and performers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.472, + 0.79 + ], + "angle": 0, + "content": "Data capture tool. We leveraged the Platform for Situated Intelligence framework [10] to develop and open-source a distributed application for data capture using HoloLens 2 [3]. A client process running on the device captured the sensor data while displaying a rectangular hologram frame around the user's visual field of view to guide their attention downward and keep the task actions in view of the sensors. Sensor data was streamed live over the network to a server application that ran on a PC and persisted the data to disk. This distributed setup allows for collecting longer uninterrupted sessions without reaching the device storage capacity limits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Comparison with other datasets. While there is no direct comparison of datasets with the same setup with HoloAssist, we list out different aspects of our dataset and compare it with related datasets in Table 2. HoloAssist is among the largest egocentric video datasets and features a multi-person collaboration setting, which is a unique addition to the field. In addition, HoloAssist relates to work on multi-agent collab-" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.094, + 0.891, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.226, + 0.895, + 0.271 + ], + "angle": 0, + "content": "Figure 5: Data distribution of 166 hours captured by HoloAssist. (left) number of sessions per activity, and (right) total length of sessions in minutes." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.288, + 0.892, + 0.478 + ], + "angle": 0, + "content": "
DatasetSettingsCollaborative &InteractiveInstructional &Procedural# real video hours
Epic-Kitchen-100 [12]CookingXX100
Assembly101 [31]Toy assemblyX✓167
Ego4D [14]Daily-life task++3,670
VirtualHome [26]Household taskX✓$
ALFRED [33]Household taskX✓$
Habitat [36]Home assistanceX✓$
BEHAVIOR [34]Daily-life task✓✓$
IGLU [16]Collaborative building*✓✓$
TEACH [25]Household task✓✓$
HoloAssist (ours)Assistive task✓✓166
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.488, + 0.895, + 0.64 + ], + "angle": 0, + "content": "Table 2: Comparison to related datasets and simulation platforms. HoloAssist features a multi-person collaborative setting which is a unique addition to existing egocentric datasets in the real world. HoloAssist provides a set of instructional and procedure videos with multi-turn dialogues. Procedure tasks are defined as following a set of defined steps or procedures to achieve a specific goal, deviation from the procedure can be construed to be a mistake. HoloAssist spans 166 hours and 2,221 sessions. §: simulation, *: Minecraft-like, +: partially included." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.669, + 0.892, + 0.699 + ], + "angle": 0, + "content": "orative simulation environments with a distinct characteristic of real-world sensor data and real-world human interaction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.633, + 0.724 + ], + "angle": 0, + "content": "3.2. Annotations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.73, + 0.895, + 0.79 + ], + "angle": 0, + "content": "To better understand the actions and interactions in the dataset, we provide several sets of third-person manual annotations for text summaries, action segments, mistake attributes, and intervention attributes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Language annotations. We asked the annotators to watch the video and write a paragraph to describe the activities in the videos. The description focuses on describing the hand actions in the procedure. The third-person post hoc summary provides insights into the key moments during the interactive task completion. These could be used to build a comprehensive set of instructions for task completion. We" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20273" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.349, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.093, + 0.615, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.115, + 0.889, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.893, + 0.3 + ], + "angle": 0, + "content": "Figure 6: Data distribution of the coarse-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.317, + 0.349, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.318, + 0.616, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.317, + 0.889, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.479, + 0.893, + 0.524 + ], + "angle": 0, + "content": "Figure 7: Data distribution of the fine-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns. We can see that most fine-grained actions last less than 2 seconds, and there is a long tail distribution in actions, verbs, and nouns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.626 + ], + "angle": 0, + "content": "also provide the transcriptions of the conversations in the video. With this set of annotations, we can understand the difference between third-person post hoc summaries and real-time conversations during the activities. More examples are in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Coarse-grained action annotations. The coarse actions usually describe a high-level step in the task (e.g., change battery of a GoPro in the GoPro set up task) and can be divided into multiple fine-grained actions. To deal with the open-world setting, we ask the annotators to write a sequence (e.g., man changes the battery of GoPro) to describe the coarse actions and also identify the active verb-noun pair and optionally an adjective for the noun for benchmarking purposes (e.g., change battery). The dataset includes 414 coarse-grained actions with 90 nouns and 39 verbs. The distribution of the actions follows a long-tail distribution shown in Figure 6, where 185 actions are considered head classes while the rest are considered tail classes according to the action frequency for evaluation purposes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Fine-grained action annotations. Fine-grained actions are the low-level atomic actions (e.g., press button, grab screw, etc.) for completing a step in the task, usually lasting for 1-2" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.893, + 0.642 + ], + "angle": 0, + "content": "seconds. The fine-grained actions are presented in a verb (adj.)-noun pair format. There are 1887 fine-grained actions with 165 nouns and 49 verbs. For a more comprehensive evaluation, we create a split of head actions with 1082 top actions and 805 tail actions. Distributions of fine-grained actions are shown in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.893, + 0.785 + ], + "angle": 0, + "content": "As mentioned earlier, noun and verb vocabularies are not pre-defined but gradually built through annotation. We ask the annotators to enter a new verb and noun if they cannot find it in the vocabulary. After the data is annotated, we ask the annotators to revisit and check the tail classes and see if they are repetitive to head classes. Due to the open-world nature, some verb and noun combinations might be interchangeable with others in the list. We show some examples in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Mistake annotation. Each fine-grained action is labeled as either correct or mistake, as indicated in Figure 2. Mistakes include the ones that are \"self-corrected by the task performers\", are \"verbally corrected by the instructors\", and \"are not corrected labeled\". Our human annotators annotate all three mistake types separately, but for benchmark evaluation, we will consolidate them into one mistake class." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20274" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.09, + 0.451, + 0.187 + ], + "angle": 0, + "content": "
IMMEDIATE INTERVENTIONLAZY INTERVENTION
insert screwdrop allen wrench
approach buttondrop hex socket head
insert joy con controllerdrop screw
place trayscrew hex socket head
pull battery doorscrew screw
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.198, + 0.471, + 0.228 + ], + "angle": 0, + "content": "Table 3: Top mistakes that are corrected immediately (left) or later and sometimes through self-correction (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.36 + ], + "angle": 0, + "content": "We defer the detailed study of differentiating whether and how the mistakes are corrected to future work. To ensure the annotation quality, we additionally ask the third-person annotators to explain why the action is a mistake and also assign a mapping to every mistake that is corrected by an instructor verbally to the conversation sentence whose type is \"instructor correcting mistakes\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.591 + ], + "angle": 0, + "content": "Intervention annotation. Since instructors assist the task performers verbally, we annotate the conversation between the instructors and performers to reflect the interventions in task completion. We annotate each conversation sentence with two attributes to indicate the conversation types and the conversation initiator. The conversation initiators can either be the \"task performer\" or the \"instructor\". And the sentence purpose types can be the instructor \"correcting mistakes\", \"answering questions\", \"following up with more instructions\", \"confirming previous actions\". \"describing the high-level task\", \"opening/closing remarks\", or the task performer starting the conversation to ask questions. We asked human annotators to watch the videos and use their best judgment to annotate the roles of different conversations in the videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.684 + ], + "angle": 0, + "content": "In our benchmarks, we consider 3 intervention types: correcting mistakes, following up with more instructions, and confirming the previous action as they are more related to physical actions in the task procedure. Figure 2 shows examples of conversational interventions. More examples can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Audit process. Annotations are done by professional annotators based on the following process. The annotators first take a pass on the video to add the fine-grained actions, coarse-grained actions, conversation, and text summary, along with the associated annotation elements for each event. After self-review, the annotated data is passed to an independent reviewer for auditing, and the mistakes are fixed directly or sent back to the original annotator for updates. The annotations finally go through a targeted review to check the open-ended text fields like narration, action sentences, and conversation transcriptions to ensure consistency. Before the annotations were delivered, we applied a list of constraints to systematically check the annotations to further dig out the wrong annotations." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.091, + 0.809, + 0.105 + ], + "angle": 0, + "content": "Spatial deixis from the intervention transcripts" + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.109, + 0.884, + 0.373 + ], + "angle": 0, + "content": "
"You should press the button that's on the body of the camera just at the right of the lens."
"You should leave the bolt, like it was before."
"The button is on the other side of the Switch."
"The SD card comes from the right slot, on the right hand side and it opens by using a knob next to the screen on the right bottom side of the screen."
"Currently the tray is facing you, please rotate so the back is facing you."
"You should put it the other way around. It's upside down."
"Please start by removing the screw of the top shelf first."
"And now, ehm, it is, it should be the other one that should be on top of the other."
"To the right of that, there is a tiny little square."
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.384, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Table 4: Examples of deictic phrases that help in grounded guidance by specifying contextual spatial locations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.441, + 0.864, + 0.457 + ], + "angle": 0, + "content": "4. Observations and Tasks from HoloAssist" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.467, + 0.894, + 0.526 + ], + "angle": 0, + "content": "Here we present observations from HoloAssist in Section 4.1 and identify a few open problems that are necessary components for an interactive AI assistant. In Section 4.2, we define new tasks and benchmarks with HoloAssist." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.767, + 0.552 + ], + "angle": 0, + "content": "4.1. Observations from HoloAssist" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.894, + 0.684 + ], + "angle": 0, + "content": "Correlation between mistakes and intervention. We find that the response time for the human assistant to intervene in the procedure and correct mistakes depends on the severity level of the mistakes. If mistakes are critical, the instructors proactively interrupt the student immediately (within less than 5 seconds duration) while other mistakes are either self-corrected by the task performer or corrected later by the instructor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.895, + 0.852 + ], + "angle": 0, + "content": "In Table 3, we present the top actions that need immediate intervention and the top actions that are self-corrected by the task performers or corrected later by the instructors. We notice that actions related to linear tasks, where the task progression is stalled if steps are not followed in order, are often intervened immediately by the instructors. For example, \"insert joy con controller\", and \"place tray\" etc. In contrast, the lazily edited corrections are related to furniture assembly such as the actions \"drop allen wrench\", \"drop hex socket head\", etc. These are tasks where mistakes are unclear in every stage, and the user often intuitively adjusts their steps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Grounded guidance. We also notice that the instructions from human instructors are often grounded in the 3D environment. An important aspect of grounded guidance is the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20275" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "ability to communicate about the physical world by pointing to things in context. Spatial deixis refers to phrases that are used to locate things in space and to express direction and distance. The deictic analysis of the transcripts in HoloAssist reveals a wide set of words that indicate specific and relative locations and directions, especially during interventions to correct mistakes. A list of some prominent examples is shown in Table 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.24, + 0.251, + 0.254 + ], + "angle": 0, + "content": "4.2. Benchmark Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.473, + 0.42 + ], + "angle": 0, + "content": "Inspired by the observations above, we think it is important for an interactive AI assistant to have a good world state estimation model that can detect mistakes and predict whether to intervene in the task procedure. Besides, augmenting instructions with spatial guidance can be useful for AI agents. To this end, we introduce new mistake detection, intervention prediction and 3D hand pose forecasting tasks for interactive and grounded guidance. Additionally, we benchmark models on action recognition tasks following the convention in [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.473, + 0.55 + ], + "angle": 0, + "content": "Mistake detection is defined following the convention [31] but applied to fine-grained actions in our benchmark. We take the features from the fine-grained action clips from the beginning of the coarse-grained action until the end of the current action clip, and the model predicts a label from {correct, mistake}. The task is challenging given that the class distribution is highly skewed, with around \\(6\\%\\) mistakes among the fine-grained actions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.471, + 0.666 + ], + "angle": 0, + "content": "Intervention type prediction is to predict the intervention types given an input of a window of 1, 3, or 5 seconds before the intervention. This newly proposed benchmark is to test if the model can correctly figure out the correct intervention types during task completion. Currently, HoloAssist includes 3 intervention types, and we report the precision and recall of each intervention type." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.901 + ], + "angle": 0, + "content": "3D hand pose forecasting is another new benchmark introduced by HoloAssist. Existing action forecasting work [12] mostly focuses on providing semantic labels of future actions and does not provide explicit 3D guidance on hand poses. Predicting 3D hand poses can be useful for various applications [4], and it can augment instructions and spatially guide users in different tasks. In this benchmark, we take 3 seconds inputs similar to other 3D body location forecasting literature [41] and forecast the continuous 3D hand poses for the next 0.5, 1.0, and 1.5 seconds. The evaluation metric is the average of mean per joint position error over time in centimeters compared to ground truth. To have a proper evaluation metric that can help 3D action guidance, we remove the mistakes from the action sequences and only forecast 3D hand pose for the correct labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.633, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.117, + 0.895, + 0.223 + ], + "angle": 0, + "content": "In this section, we provide the evaluation results of the proposed benchmarks. We will start with the standard action recognition benchmarks, and then we will present the results of the newly proposed benchmarks. We also provide ablations of different sensor modalities to understand the roles of different sensors in various tasks. We hope the baseline results can guide future research in this space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.452 + ], + "angle": 0, + "content": "Implementation details. We adopt TimeSformer [9], a state-of-the-art vision transformer (ViT) [13] based video model, as the backbone and change the head with a different number of classes for different benchmarks. We modify the original TimeSformers to perform multimodal learning by introducing additional tokens for different modalities and embedding layers to encode the additional sensor modalities. Specifically, we use \\(26 \\times 2\\) tokens for both left and right hands (one token for one hand joint), one token for eye gaze, one token for head poses, and 196 tokens for depth. We can enable and/or disable different modalities during training and evaluation. Detailed configurations are available in supplementary materials. Note that we consider the resulting model a vanilla multimodal model and serve as a baseline for future studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.56 + ], + "angle": 0, + "content": "We randomly split the 2221 sessions into train, validation, and test sets following a ratio of \\(70\\%\\), \\(10\\%\\), and \\(20\\%\\) on a per-task basis, which includes 1545 sessions for training, 213 sessions for validation, and 463 sessions for testing. We also synchronize all the modalities according to the video stream and keep the frame rate at 30 fps by sub-sampling other modalities in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.895, + 0.697 + ], + "angle": 0, + "content": "During training, the model is trained for 15 epochs with an initial learning rate of 0.01 and a batch size of 64 using stochastic gradient descent. We divide the learning rate by 10 at the epoch 11 and 14. For each input segment, we randomly sample 8 frames/data points within the segment. We train our models with 4xA6000 GPU machines, and the fine-grained action recognition runs usually take about one day. We also train the model with random initialization and an ImageNet pre-trained ViT backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.897, + 0.73 + ], + "angle": 0, + "content": "For our hand pose forecasting benchmark, which is a regression task, we adopt Seq2Seq model [35] following [19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Action recognition. We show the fine-grained action recognition results in Table 5 and the coarse-grained action recognition results in Table 6. We can see that initialized with a pre-trained ViT model on ImageNet; all models can achieve around \\(35\\%\\) top-1 accuracy on fine-grained action recognition and \\(50\\%\\) top-1 accuracy on coarse-grained action recognition. This is comparable to baselines of other fine-grained action benchmarks [31] (\\(23\\%\\) top-1 accuracy)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We notice that the pre-trained model on ImageNet may reduce the influence of other modalities that are not pretrained. If trained from scratch, we can see from Table 5" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "20276" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.251 + ], + "angle": 0, + "content": "
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB34.83/68.6042.14/78.9666.81/90.0435.26/69.3442.56/79.5367.19/90.360.03/0.1710.86/36.3338.01/66.48
Hands20.86/43.9235.38/65.7637.10/63.4221.13/44.5035.72/66.3037.50/64.060.00/0.0110.11/25.477.30/15.54
R+H35.06/68.9542.45/79.4267.05/90.0135.49/69.7142.87/80.0367.43/90.320.03/0.1611.05/33.7138.39/66.85
R+H+E35.27/68.6942.92/79.1167.03/89.9635.72/69.4243.33/79.6767.45/90.290.03/0.1811.99/37.6435.96/65.17
R+H+E+I34.80/68.2642.24/78.8866.65/89.7635.23/69.0042.63/79.4667.03/90.080.03/0.1712.92/35.9638.20/65.17
ScratchRGB18.78/48.0928.45/65.4343.72/73.6919.03/48.7228.74/66.0044.09/74.210.00/0.016.55/22.6615.73/34.27
Hands23.94/47.5839.79/68.7639.34/65.7624.26/48.2140.18/69.3039.79/66.430.00/0.0010.67/28.655.62/16.29
RGB + Eyes20.86/50.2730.92/67.5945.28/75.2921.13/50.9331.22/68.2045.65/75.800.00/0.018.05/21.7217.60/37.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.261, + 0.895, + 0.354 + ], + "angle": 0, + "content": "Table 5: Fine-grained action recognition benchmark results. We report baseline results of (multimodal) TimeSformer [9] models on HoloAssist. Here R, H, I denote RGB, hands, and head pose (whose sensor source is similar as IMUs). When starting from a pre-trained ViT model on ImageNet, all the models have higher performance than training from scratch. However, with a pre-trained image model, the influence of other modalities is reduced. If the models are trained from scratch, we can see that adding hands can improve the prediction of verbs and lead to better results than the baseline RGB only model. Simply concatenating more sensors as inputs may not necessarily lead to improved performance." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.359, + 0.894, + 0.476 + ], + "angle": 0, + "content": "
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB50.91/86.8960.51/93.4573.35/95.9053.40/89.5362.78/95.3975.00/96.700.37/2.2420.00/58.8943.89/81.67
Hands22.20/54.1635.12/72.3437.43/68.3323.44/57.0836.75/75.2538.84/70.450.00/0.126.11/20.5612.22/30.56
R+H50.80/86.5459.71/93.3673.20/95.7853.27/89.1261.94/95.2075.09/96.570.37/2.2820.00/60.5639.44/81.67
R+H+E50.35/85.5158.85/93.2173.67/95.6352.65/88.0360.85/94.8975.28/96.320.53/2.2823.33/63.3345.00/83.33
R+H+E+I50.18/86.0759.00/93.2173.64/96.2552.49/88.7861.22/95.0775.41/96.910.50/2.1219.44/60.0042.22/84.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.486, + 0.865, + 0.502 + ], + "angle": 0, + "content": "Table 6: Coarse-grained action recognition results. The overall trend is similar to fine-grained action recognition." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.517, + 0.449, + 0.641 + ], + "angle": 0, + "content": "
Mods.F-scoreCorrectMistake
PrecisionRecallPrecisionRecall
Random27.7160.8710.2215.0046.15
RGB35.1182.5651.8212.9626.92
Hands40.1992.6852.4112.531.25
R+H36.1885.5143.079.6811.54
R+H+E32.0888.5742.7611.4350.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.651, + 0.472, + 0.728 + ], + "angle": 0, + "content": "Table 7: Mistake detection results. Similar to action recognition, adding hand poses to RGB \\((\\mathbb{R} + \\mathbb{H})\\) improves an RGB-only model from 35.11 to 36.18 points. What's different is that we find the hands only model outperforms other modalities in mistake detection, achieving 40.19 points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.471, + 0.773 + ], + "angle": 0, + "content": "(bottom) that adding hands can improve the prediction of verbs and lead to better results than the RGB-only model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Mistake detection. In Table \\(7^{\\dagger}\\), we present the results of mistake detection where the features are extracted from the pre-trained action recognition models shown in Table 5. Here we find the hand poses information benefits the task and outperforms the other modalities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.858, + 0.472, + 0.874 + ], + "angle": 0, + "content": "Intervention type prediction. For intervention prediction," + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.518, + 0.894, + 0.649 + ], + "angle": 0, + "content": "
Mods.OverallConfirmCorrectFollow-up
Prec.RecallPrec.RecallPrec.RecallPrec.Recall
RGB13.9333.330.000.000.000.0041.79100
R+H24.8933.6432.144.500.000.0042.5296.43
R+E25.5533.7333.9010.360.000.0042.7690.83
R+H+E48.3137.5939.1140.9361.119.9144.761.93
RGB(Pt)37.5437.7442.3127.5027.3336.6142.9749.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.782 + ], + "angle": 0, + "content": "Table 8: Intervention type prediction results. The classes in benchmark are highly skewed. Here E refers to eye gaze. We can see the predictions of the baseline models (i.e., RGB, R+H, R+E) favor predicting follow-up instructions while ignoring correcting mistakes (recall is 0). Training more modalities \\((\\mathbb{R} + \\mathbb{H} + \\mathbb{E})\\) alleviates this issue and even outperforms the RGB model (RGB (Pt)) in overall precision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "we show the results in Table \\(8^{\\dagger}\\). We can see adding hands and eye gaze \\((\\mathbb{R} + \\mathbb{H} + \\mathbb{E})\\) can significantly boost the overall precision and recall to \\(48.31\\%\\) and \\(37.59\\%\\), improving about 35 and 4 percentage points over RGB. This may be because eye gaze is a forecasting signal, as people often look at the regions before the action starts, which can assist the models to attend to important regions for better anticipation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.299, + 0.9 + ], + "angle": 0, + "content": "Evaluated on \\(10\\%\\) of the entire data" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "20277" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.895, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.167, + 0.892, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.249, + 0.495, + 0.261 + ], + "angle": 0, + "content": "Time" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.897, + 0.339 + ], + "angle": 0, + "content": "Figure 8: Qualitative visualization for 3D hand pose forecasting. We visualize the ground-truth hand joint positions (i.e., hand pose, visualized in Green) and the prediction of the hand pose for the next 1.5 seconds (visualized in Red). The input to the model is a 3-second long clip ahead of the prediction. The task is challenging as hands often move quickly. As we can see from the figure, the predicted hand pose is more off from the grounded truth in the longer future." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.354, + 0.451, + 0.48 + ], + "angle": 0, + "content": "
Mods.Mean Error Distance (cm, ↓)
0.5 sec1.0 sec1.5 sec
Static-H9.3413.9116.70
Hands9.8010.6811.25
H+E9.8010.7011.25
H+E+I9.8010.6911.25
R+H+E9.7310.6511.22
R+H+E+I9.7210.6211.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.492, + 0.472, + 0.584 + ], + "angle": 0, + "content": "Table 9: 3D hand pose forecasting benchmark results. We report the mean per joint position error for 0.5 1.0, 1.5 seconds (lower the better). The static hand baseline (Static-H) refers to using the last frame of the input. The Seq2Seq [35] model trained using hands only H, and using only hand H achieves better results than Static-H." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.472, + 0.732 + ], + "angle": 0, + "content": "3D hand pose forecasting. Table 9 shows that given only the 3D poses of hands as input, the model can perform with the accuracy of 9.80, 10.68, and 11.25 centimeters, the average of mean per joint position error, for 0.5, 1, and 1.5 seconds, respectively. We should note that this task is challenging as hands can move quickly within a window of 0.5 seconds. Compared to static-H, which uses the last 3D hand pose, our baseline (H) already outperforms it. In Figure 8, we show the visualization of the hand pose forecasting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Importance of hand poses and eye gaze. As we can already see from the Tables 5, 7, 8, 9, the 3D hand poses and eye gaze can help the model prediction to recognize actions, detect mistakes, and understand users' intentions. These modalities can augment the commonly used RGB images for better performance. Simply concatenating more modalities (e.g., depth, head poses) as inputs may not necessarily lead to a more capable model, as those modalities may need specialized encoders or model architectures to process them simultaneously. We believe HoloAssist will enable and foster further research in multi-modal learning in this direction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.355, + 0.771, + 0.372 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.896, + 0.761 + ], + "angle": 0, + "content": "In this work, we identified and explored several important problems with building an interactive AI assistant in the physical world. We introduced a large-scale multimodal egocentric video dataset, HoloAssist, containing rich information about human interaction in an assistive task completion setting. The task performer wears a HoloLens 2 headset while completing various object-centric manipulation tasks. The real-time video feed from the headset is sent to a remote instructor who provides verbal guidance to the task performer. HoloAssist captures seven raw sensor modalities during the interaction, and among them, we found hand pose and eye gaze are useful information sources for an interactive AI agent. By augmenting the data with additional third-person manual annotations on action segments, mistakes, and intervention types, we constructed new benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we believe is a necessary component for an interactive and grounded AI assistant. As a first step in this direction, this work also leaves room for future work to improve upon (e.g., annotating object poses in the data, investigating object-centric models of affordance and manipulations in AI assistance, etc.). We believe HoloAssist, coupled with the associated benchmarks and tooling will benefit future research into building competent AI assistants for everyday tasks in the real world." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.649, + 0.794 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We thank all the 222 data collectors who participated in the study and acknowledge the hard work of the annotation team led by Megan Yuan and Dan Luo from DataTang Technology Inc. We thank Nick Saw for the help in the data collection software and Yale Song and Vibhav Vineet from the Computer Vision Group at Microsoft Research for the early discussion. We also thank the feedback from colleagues at MSR." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "20278" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.156 + ], + "angle": 0, + "content": "[1] Microsoft dynamics 365 guides. https://dynamics.microsoft.com/en-us/mixed-reality/guides/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.185 + ], + "angle": 0, + "content": "[2] Microsoft hololens. https://www.microsoft.com/en-us/hololens.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[3] Platform for situated intelligence hololens capture tools. https://github.com/microsoft/psi/tree/master/Sources/MixedReality/HoloLensCapture.2,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.472, + 0.436 + ], + "angle": 0, + "content": "[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.472, + 0.507 + ], + "angle": 0, + "content": "[5] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.472, + 0.605 + ], + "angle": 0, + "content": "[6] Gregory Aist, John Dowding, Beth Ann Hockey, Manny Rayner, James Hieronymus, D Bohus, B Boven, Nate Blaylock, Ellen Campana, S Early, et al. Talking through procedures: An intelligent space station procedure assistant. In Proceedings of the tenth conference on European chapter of the Association for Computational Linguistics-Volume 2, pages 187-190, 2003. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.606, + 0.472, + 0.661 + ], + "angle": 0, + "content": "[7] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Ivan Laptev, Josef Sivic, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[8] Sean Andrist, Michael Gleicher, and Bilge Mutlu. Looking coordinated: Bidirectional gaze mechanisms for collaborative interaction with virtual characters. In Proceedings of the 2017 CHI conference on human factors in computing systems, pages 2571-2582, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.733, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[10] Dan Bohus, Sean Andrist, Ashley Feniello, Nick Saw, Mihai Jalobeanu, Patrick Sweeney, Anne Loomis Thompson, and Eric Horvitz. Platform for situated intelligence, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[11] Dan Bohus and Alexander I Rudnicky. Larri: A language-based maintenance and repair assistant. Spoken multimodal human-computer dialogue in mobile environments, pages 203-218, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonio Furnari, Jian Ma, Evangelos Kazakos, Davide Molti" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "santi, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.594, + 0.894, + 0.635 + ], + "angle": 0, + "content": "[15] W Lewis Johnson and Jeff Rickel. Steve: An animated pedagogical agent for procedural training in virtual environments. ACM SIGART Bulletin, 8(1-4):16-21, 1997. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[16] Julia Kiseleva, Ziming Li, Mohammad Aliannejadi, Shrestha Mohanty, Maartje ter Hoeve, Mikhail Burtsev, Alexey Skrynnik, Artem Zholus, Aleksandr Panov, Kavya Srinet, et al. Interactive grounded language understanding in a collaborative environment: Iglu 2021. In NeurIPS 2021 Competitions and Demonstrations Track, pages 146-161. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[17] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[18] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20279" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[20] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[21] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[22] Ramesh Manuvinakurike, Sovan Biswas, Giuseppe Raffa, Richard Beckwith, Anthony Rhodes, Meng Shi, Gesem Gudino Mejia, Saurav Sahay, and Lama Nachman. Human in the loop approaches in multi-modal conversational task guidance system development. arXiv preprint arXiv:2211.01824, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[23] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl;dw? summarizing instructional videos with task relevance and cross-modal saliency. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 540–557, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[24] Rashmiranjan Nayak, Umesh Chandra Pati, and Santos Kumar Das. A comprehensive review on deep learning-based methods for video anomaly detection. Image and Vision Computing, 106:104078, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[25] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2017-2025, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[26] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8494-8502, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[27] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, June 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[28] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[29] Charles Rich and Candace L Sidner. Diamondhelp: A generic collaborative task guidance system. AI Magazine, 28(2):33-33, 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[30] Charles Rich, Candace L Sidner, and Neal Lesh. Collagen: Applying collaborative discourse theory to human-computer interaction. AI magazine, 22(4):15-15, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[31] F. Sener, D. Chatterjee, D. Shelepop, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. CVPR 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[32] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[34] Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martin-Martin, Fei Xia, Kent Elliott Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, Karen Liu, et al. Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In Conference on Robot Learning, pages 477-490. PMLR, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[35] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27, 2014. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[36] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.893, + 0.525 + ], + "angle": 0, + "content": "[37] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning (CoRL), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[38] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, pages 485-501. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.894, + 0.663 + ], + "angle": 0, + "content": "[39] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), Oct. 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.894, + 0.706 + ], + "angle": 0, + "content": "[40] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.894, + 0.762 + ], + "angle": 0, + "content": "[41] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[42] Artem Zholus, Alexey Skrynnik, Shrestha Mohanty, Zoya Volovikova, Julia Kiseleva, Artur Szlam, Marc-Alexandre Cote, and Aleksandr I Panov. Iglu gridworld: Simple and fast environment for embodied dialog agents. arXiv preprint arXiv:2206.00142, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[43] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 14(10):2537-2550, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.926, + 0.52, + 0.937 + ], + "angle": 0, + "content": "20280" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.092, + 0.472, + 0.135 + ], + "angle": 0, + "content": "[44] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner," + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.894, + 0.119 + ], + "angle": 0, + "content": "editors, Computer Vision - ECCV 2022, pages 395-412, 2022. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "20281" + } + ] +] \ No newline at end of file diff --git a/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_origin.pdf b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b1d842265269bf776690c380153e2bfe699ae362 --- /dev/null +++ b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/7b800728-4ea2-4a01-b47d-4a862ae2ec61_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3f5c7ad2af0f13ccdbe4b40d156911c5d852141fa0ba194ef6cc6d85178ff2b +size 4573383 diff --git a/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/full.md b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75f5140638a095daa904fc9487208c67f3aa1ae1 --- /dev/null +++ b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/full.md @@ -0,0 +1,299 @@ +# HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World + +Xin Wang $^{1*}$ Taein Kwon $^{1,2*}$ Mahdi Rad $^{1}$ Bowen Pan $^{1\dagger}$ Ishani Chakraborty $^{1}$ Sean Andrist $^{1}$ Dan Bohus $^{1}$ Ashley Feniello $^{1}$ Bugra Tekin $^{1\dagger}$ Felipe Vieira Frujeri $^{1}$ Neel Joshi $^{1}$ Marc Pollefeys $^{1,2}$ Microsoft ${}^{2}$ ETH Zurich + +# Abstract + +Building an interactive AI assistant that can perceive, reason, and collaborate with humans in the real world has been a long-standing pursuit in the AI community. This work is part of a broader research effort to develop intelligent agents that can interactively guide humans through performing tasks in the physical world. As a first step in this direction, we introduce HoloAssist, a large-scale egocentric human interaction dataset, where two people collaboratively complete physical manipulation tasks. The task performer executes the task while wearing a mixed-reality headset that captures seven synchronized data streams. The task instructor watches the performer's egocentric video in real time and guides them verbally. By augmenting the data with action and conversational annotations and observing the rich behaviors of various participants, we present key insights into how human assistants correct mistakes, intervene in the task completion procedure, and ground their instructions to the environment. HoloAssist spans 166 hours of data captured by 350 unique instructor-performer pairs. Furthermore, we construct and present benchmarks on mistake detection, intervention type prediction, and hand forecasting, along with detailed analysis. We expect HoloAssist will provide an important resource for building AI assistants that can fluidly collaborate with humans in the real world. Data can be downloaded at https://holoassist.github.io/. + +# 1. Introduction + +Recent years have witnessed incredible progress in general-purpose AI agents that assist humans with various open-world tasks, especially in the digital world. AI systems powered by large language models (LLMs) like ChatGPT [27] can answer users' questions and assist them with various text-based tasks. However, these AI assistants do not have sufficient first-hand experience in the physical world and thus cannot perceive world states and actively intervene + +# Two-Person Interactive Assistive Task Completion + +![](images/a1dda76f6bd78d81d0df44b1d520ffd96c73633b26be16a4df7c6b6a0a1076a3.jpg) + +![](images/7633283a5d89767f41f1e99ea3c91a2917798a2d8eab6c4d403ec5ff07f4303d.jpg) + +![](images/87fcba3aa50b0e7330ac13ab705e577d5cecca34184a6521f6678610e9f82686.jpg) +(1) RGB, (2) eye gaze, (3) hand pose +Figure 1: HoloAssist features a two-person interactive assistive task completion setting. The task performer wears an AR device and completes the tasks while the captured data is streamed over the network to a remote instructor watching it on the laptop. The instructor provides verbal guidance to the student. HoloAssist includes seven modalities captured live and human annotated text descriptions as the 8th modality. + +![](images/5e501e0ad178f56304f6fd0c67d2171de4501fbb580efbe8924d44c112ed7af3.jpg) +(4) head pose, (5) depth + +![](images/61e5bda88e7a330a35e155563e39782e0835a37940b97082f348a3448971c1a7.jpg) +(6) IMU, (7) audio + +![](images/2feb99fad4a437d27f11d01deddfaa2377790abb610f10e00484df729312c4be.jpg) +(8) text + +in the task completion procedure. + +Building an AI assistant that can perceive, reason and interact in the physical world has attracted attention from researchers across different fields in computer vision [7, 23, 37, 38], human-computer interaction [6, 8, 15, 29], robotics [5, 32], and industrial practitioners. For example, AR Guides [1], which aims to guide users to complete complex tasks, has become popular with the development of augmented reality (AR) devices. However, existing systems often rely on pre-defined instructions or formulate the virtual assistant as a question answering [37, 38] or video under + +standing problem [7, 23] without real-world interaction. + +In another line of work, researchers have developed simulation environments like Habitat [21, 36], VirtualHome [26], and AI2-Thor [17] to build AI agents that can interact with the physical world and collaboratively achieve new tasks [42]. Still, a large gap remains in transferring these agents to the real world, and the interaction between agents is largely simplified compared to real-world human interaction. + +In this work, we focus on the challenges of developing intelligent agents that share perspectives with humans and interactively guide human users through performing tasks in the physical world. As a first step, we introduce HoloAssist, a large-scale egocentric human interaction dataset to explore and identify the open problems in this direction. As shown in Figure 1, the task performer wears an AR headset* to capture data while completing the tasks. An instructor watches the real-time egocentric video feed remotely and verbally guides the performer. We have developed and open-sourced a data capture tool [3] using a distributed server-client setup to enable data streaming and multimodal data capture. + +HoloAssist contains 166 hours of data captured by 222 diverse participants forming 350 unique instructor-performer pairs and carrying out 20 object-centric manipulation tasks. The objects range from common electronic devices to rare objects in factories and specialized labs. The tasks are generally challenging for first-time participants, requiring instructor assistance for successful completion. Seven raw sensor modalities are captured, including RGB, depth, head pose, 3D hand pose, eye gaze, audio, and IMU, to aid in the understanding of human intentions, estimating world states, predicting future actions, and so on. Finally, the dataset is augmented with third-person manual annotations consisting of a text summary, intervention types, mistake annotation, and action segments of the videos as illustrated in Figure 2. + +We have observed several characteristics demonstrated by human instructors from HoloAssist. First, instructors are often proactive with precisely timed interventions. Instead of waiting until mistakes happen, instructors provide follow-up instructions when the task performer appears confused. Second, the verbal guidance from the instructors tends to be concise and grounded in the task performer's environment. The instructions are often framed as spatial deictics to aid the task performer in spatial directions and distances in the 3D world. Moreover, instructors often have a good world model estimation and can detect whether mistakes disrupt task completion and then adjust the guidance. + +We take a step further and introduce new tasks and benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we conjecture are essential modules for an intelligent assistant. Additionally, we benchmark the dataset on action classification and anticipation tasks and provide empirical results to understand the role of + +![](images/e59c7abb08bdb1905e93e9f0daa7a4c7c694db4ee2cfec00f526720044a5cacf.jpg) +Figure 2: HoloAssist includes action and conversational annotations, in addition to text summaries of the videos, to indicate the mistakes and interventions in task completion. mistake or correct attributes are associated with each fine-grained action. A purpose label is associated with every utterance to indicate the type of verbal intervention. + +different modalities in various tasks. We hope our dataset, findings, and tooling can inspire and provide rich resources for future work on designing interactive AI assistants and situated AI assistance applications in the real world. + +# 2. Related Work + +Our work closely connects with several lines of work in computer vision, especially egocentric vision, embodied AI, and human-computer interaction. + +Interactive AI assistants. Building interactive agents that can assist humans to carry out tasks in the world—real or virtual—has been a long-standing problem in different areas of AI and HCI [6, 11, 15, 22, 25, 29, 30]. As far back as 1997, Johnson and Rickel introduced "Steve" [15], an early pedagogical agent that aims to help students learn procedural tasks in VR. Recent efforts have focused on new modeling approaches and data collection techniques for training conversational task guidance assistants, such as model-in-the-loop wizard-of-oz [22] and human-human interaction to mimic robot actions in simulated environments [25]. In this work, we revisit this problem and provide a systematic study of real-world human interaction, and we also provide rich sensor information to push the frontiers of the research. + +Egocentric video datasets. Egocentric perspectives often convey rich information about the users' intentions. A shared perspective between the users and the human or AI assistants is useful for the assistants to provide more timely and grounded guidance. In computer vision, several egocentric video datasets [12, 14, 18, 20, 28, 31, 39] have emerged in + +![](images/d315d1ce8ba45d9afd6d5dff34e8b2d7fb05c5687289a28292445d0e73f834d2.jpg) +Figure 3: HoloAssist was collected by participants diverse in ages, occupations, genders, and geography. This helps us to study a diverse set of users with different backgrounds. + +the community. EPIC-KICHENS [12] is a widely adopted egocentric video dataset capturing kitchen activities. The recent Ego4D dataset [14] is the largest egocentric video dataset in the wild that provides a comprehensive database for egocentric perception in the 3D world. In contrast to earlier egocentric video datasets, HoloAssist features a multiperson interactive task completion setting, where human interaction during the procedure provides a rich source for designing AI assistants to be more proactive and grounded in the environment. Yet our work can benefit from the rich knowledge and representation learned from existing datasets like Ego4D and is complementary in nature. + +Mistake detection. One of the key observations in human interaction is that human assistants tend to correct mistakes and proactively intervene in the task completion procedure. While there has been a large body of work for video-based anomaly detection [24, 40, 43, 44], mistake detection in procedural settings has been under-explored. The Assembly101 dataset [31] proposes a mistake detection task to predict if a coarse-grained action segment is a mistake or correction. By contrast, HoloAssist emphasizes fine-grained actions since instructors may intervene when they spot a student's mistake in an active intervention setting rather than wait until the whole step (i.e., coarse-grained action) is completed. In addition, we propose a new intervention prediction task and, in combination, enable a more comprehensive understanding of interactions in an assistive task completion setting. + +Multimodality and interaction. Human interaction with the world is multimodal as we see, speak, and touch objects in the environment. In HoloAssist, we collect seven raw sensor modalities that might help understand humans' intentions, estimate the world states, predict future actions, etc. Previous datasets [12, 14, 18, 39] often provide a limited subset of modalities. Although not every sensor may currently be relevant for the downstream tasks, the seven synchronized sensor modalities provided in HoloAssist will + +OBJECT SCALES OBJECT CATEGORIES + +
SmallGoPro, Nintendo Switch, DSLR
MediumPortable printer, Computer, Nespresso machine
BigStandalone printer, big coffee machine, IKEA fur-niture (stool, utility cart, tray Table, nightstand)
RareNavVis laser scanner, ATV motor cycle, wheel belt, circuit breaker
+ +Table 1: HoloAssist includes 16 objects with diverse scales. Apart from common objects used in daily life, HoloAssist includes rare equipment from mechanical labs. 20 tasks are object-centric manipulation tasks for each object and the 4 IKEA furniture has both assembly and disassembly tasks. + +give practitioners more potential for designing multimodal agents and models even beyond the scope of this work. + +Embodied simulation platforms. There is an emerging interest in embodied agents that can perceive, reason, and act in the 3D world. Researchers [17, 21, 26, 34, 36, 42] build various simulation environments to learn such embodied agents. IGLU [42] aims to build interactive agents that learn to solve a task while being provided with grounded natural language instructions in a collaborative environment based on Minecraft, a popular video game. HoloAssist complements this line of work by providing more realistic human interaction and real-world sensor perception. + +# 3. HoloAssist: Human Assistance Dataset + +In this work, we introduce HoloAssist which features a two-person collaboration scenario and can be used to situate AI assistance in the physical world. We will start by describing the data collection and statistics in Section 3.1 and annotations in Section 3.2, before diving into the observations and benchmarks in the following sections. + +# 3.1. Data Collection and Statistics + +Tasks and objects. We consider multi-step goal-oriented tasks involving 16 objects ranging from familiar objects often used in daily life to rare objects sometimes used in labs and factories as summarized in Table 1. We consider small electronics like a GoPro, DSLR camera, and Nintendo, office appliances like a Nespresso machine and printer, IKEA furniture, and objects in labs such as a laser scanner, motor cycle, and circuit breaker. We have designed 20 tasks involving physical manipulation of these objects, e.g., changing batteries, changing belts, furniture assembly, machine setup, etc. There is one task per object except for the IKEA furniture, which has assembly and disassembly tasks. Detailed task instructions are in supplementary materials. + +Participants and collection procedure. We recruited 222 participants to form 350 unique pairs of instructors and performers for data collection. Figure 3 shows the demograph- + +![](images/d5ccbe40edbf69e07f487a1f0638393bb22c06e1b99f0dfe1b8de8a310128bc5.jpg) +Figure 4: The skill level of participants (0-10) for the tasks is self-reported by the participants. The skill levels roughly reflect the length of the sessions though they might be noisy. + +ics of the participants. Before data collection, the participants review the IRB forms to acknowledge the privacy and ethics standards (more details in supplementary materials). The instructors are informed about the task in detail. The participants playing the role of performers are only given a rough description of the tasks and scenarios beforehand and interacted with the objects based on their understanding. The instructors provide verbal guidance as the performers set out to complete the tasks. + +In Figure 4 (top), we show the distribution of the performers' familiarity with the tasks measured by a self-reported score (0-10) by the participants. We show the average length and the outliers of the recorded sessions in Figure 4 (bottom) to give a rough idea of how the participant's skill levels may lead to increased session variance. The participants' diverse skill levels and backgrounds provide rich information about the user behaviors and diverse interaction between the instructors and performers. + +Data capture tool. We leveraged the Platform for Situated Intelligence framework [10] to develop and open-source a distributed application for data capture using HoloLens 2 [3]. A client process running on the device captured the sensor data while displaying a rectangular hologram frame around the user's visual field of view to guide their attention downward and keep the task actions in view of the sensors. Sensor data was streamed live over the network to a server application that ran on a PC and persisted the data to disk. This distributed setup allows for collecting longer uninterrupted sessions without reaching the device storage capacity limits. + +Comparison with other datasets. While there is no direct comparison of datasets with the same setup with HoloAssist, we list out different aspects of our dataset and compare it with related datasets in Table 2. HoloAssist is among the largest egocentric video datasets and features a multi-person collaboration setting, which is a unique addition to the field. In addition, HoloAssist relates to work on multi-agent collab- + +![](images/2167e90cae7efab397313ab6c45ebb97871cd1263b186cc50add8293825b9f9a.jpg) +Figure 5: Data distribution of 166 hours captured by HoloAssist. (left) number of sessions per activity, and (right) total length of sessions in minutes. + +
DatasetSettingsCollaborative &InteractiveInstructional &Procedural# real video hours
Epic-Kitchen-100 [12]CookingXX100
Assembly101 [31]Toy assemblyX✓167
Ego4D [14]Daily-life task++3,670
VirtualHome [26]Household taskX✓$
ALFRED [33]Household taskX✓$
Habitat [36]Home assistanceX✓$
BEHAVIOR [34]Daily-life task✓✓$
IGLU [16]Collaborative building*✓✓$
TEACH [25]Household task✓✓$
HoloAssist (ours)Assistive task✓✓166
+ +Table 2: Comparison to related datasets and simulation platforms. HoloAssist features a multi-person collaborative setting which is a unique addition to existing egocentric datasets in the real world. HoloAssist provides a set of instructional and procedure videos with multi-turn dialogues. Procedure tasks are defined as following a set of defined steps or procedures to achieve a specific goal, deviation from the procedure can be construed to be a mistake. HoloAssist spans 166 hours and 2,221 sessions. §: simulation, *: Minecraft-like, +: partially included. + +orative simulation environments with a distinct characteristic of real-world sensor data and real-world human interaction. + +# 3.2. Annotations + +To better understand the actions and interactions in the dataset, we provide several sets of third-person manual annotations for text summaries, action segments, mistake attributes, and intervention attributes. + +Language annotations. We asked the annotators to watch the video and write a paragraph to describe the activities in the videos. The description focuses on describing the hand actions in the procedure. The third-person post hoc summary provides insights into the key moments during the interactive task completion. These could be used to build a comprehensive set of instructions for task completion. We + +![](images/375558581e7d68ca45d5197ce1d72ed02d74bf0b5c65005114d38c3d1044f44a.jpg) + +![](images/819b3b8f63b91102c59be1acb7da84f6bc09cfc260b9b339f28b39029b4ca4af.jpg) + +![](images/c9bf663b27903c482cd3796d12cf365ae1042926e1f93e99a58b0887f7d210cc.jpg) +Figure 6: Data distribution of the coarse-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns. + +![](images/7b72b865bf72ee6741949174835bc214a4abffbf66a45dd99f28e8baec40ffb4.jpg) + +![](images/5eb05a5bb82c1824cca9087c0f5fccead80dd303fd6545cb855a9b40fbbb269b.jpg) + +![](images/dba5cfe3da510c226c789124af38c68913adf962104a4d000a36313a43a88442.jpg) +Figure 7: Data distribution of the fine-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns. We can see that most fine-grained actions last less than 2 seconds, and there is a long tail distribution in actions, verbs, and nouns. + +also provide the transcriptions of the conversations in the video. With this set of annotations, we can understand the difference between third-person post hoc summaries and real-time conversations during the activities. More examples are in supplementary materials. + +Coarse-grained action annotations. The coarse actions usually describe a high-level step in the task (e.g., change battery of a GoPro in the GoPro set up task) and can be divided into multiple fine-grained actions. To deal with the open-world setting, we ask the annotators to write a sequence (e.g., man changes the battery of GoPro) to describe the coarse actions and also identify the active verb-noun pair and optionally an adjective for the noun for benchmarking purposes (e.g., change battery). The dataset includes 414 coarse-grained actions with 90 nouns and 39 verbs. The distribution of the actions follows a long-tail distribution shown in Figure 6, where 185 actions are considered head classes while the rest are considered tail classes according to the action frequency for evaluation purposes. + +Fine-grained action annotations. Fine-grained actions are the low-level atomic actions (e.g., press button, grab screw, etc.) for completing a step in the task, usually lasting for 1-2 + +seconds. The fine-grained actions are presented in a verb (adj.)-noun pair format. There are 1887 fine-grained actions with 165 nouns and 49 verbs. For a more comprehensive evaluation, we create a split of head actions with 1082 top actions and 805 tail actions. Distributions of fine-grained actions are shown in Figure 7. + +As mentioned earlier, noun and verb vocabularies are not pre-defined but gradually built through annotation. We ask the annotators to enter a new verb and noun if they cannot find it in the vocabulary. After the data is annotated, we ask the annotators to revisit and check the tail classes and see if they are repetitive to head classes. Due to the open-world nature, some verb and noun combinations might be interchangeable with others in the list. We show some examples in the supplementary materials. + +Mistake annotation. Each fine-grained action is labeled as either correct or mistake, as indicated in Figure 2. Mistakes include the ones that are "self-corrected by the task performers", are "verbally corrected by the instructors", and "are not corrected labeled". Our human annotators annotate all three mistake types separately, but for benchmark evaluation, we will consolidate them into one mistake class. + +
IMMEDIATE INTERVENTIONLAZY INTERVENTION
insert screwdrop allen wrench
approach buttondrop hex socket head
insert joy con controllerdrop screw
place trayscrew hex socket head
pull battery doorscrew screw
+ +We defer the detailed study of differentiating whether and how the mistakes are corrected to future work. To ensure the annotation quality, we additionally ask the third-person annotators to explain why the action is a mistake and also assign a mapping to every mistake that is corrected by an instructor verbally to the conversation sentence whose type is "instructor correcting mistakes". + +Intervention annotation. Since instructors assist the task performers verbally, we annotate the conversation between the instructors and performers to reflect the interventions in task completion. We annotate each conversation sentence with two attributes to indicate the conversation types and the conversation initiator. The conversation initiators can either be the "task performer" or the "instructor". And the sentence purpose types can be the instructor "correcting mistakes", "answering questions", "following up with more instructions", "confirming previous actions". "describing the high-level task", "opening/closing remarks", or the task performer starting the conversation to ask questions. We asked human annotators to watch the videos and use their best judgment to annotate the roles of different conversations in the videos. + +In our benchmarks, we consider 3 intervention types: correcting mistakes, following up with more instructions, and confirming the previous action as they are more related to physical actions in the task procedure. Figure 2 shows examples of conversational interventions. More examples can be found in the supplementary materials. + +Audit process. Annotations are done by professional annotators based on the following process. The annotators first take a pass on the video to add the fine-grained actions, coarse-grained actions, conversation, and text summary, along with the associated annotation elements for each event. After self-review, the annotated data is passed to an independent reviewer for auditing, and the mistakes are fixed directly or sent back to the original annotator for updates. The annotations finally go through a targeted review to check the open-ended text fields like narration, action sentences, and conversation transcriptions to ensure consistency. Before the annotations were delivered, we applied a list of constraints to systematically check the annotations to further dig out the wrong annotations. + +# Spatial deixis from the intervention transcripts + +Table 3: Top mistakes that are corrected immediately (left) or later and sometimes through self-correction (right). + +
"You should press the button that's on the body of the camera just at the right of the lens."
"You should leave the bolt, like it was before."
"The button is on the other side of the Switch."
"The SD card comes from the right slot, on the right hand side and it opens by using a knob next to the screen on the right bottom side of the screen."
"Currently the tray is facing you, please rotate so the back is facing you."
"You should put it the other way around. It's upside down."
"Please start by removing the screw of the top shelf first."
"And now, ehm, it is, it should be the other one that should be on top of the other."
"To the right of that, there is a tiny little square."
+ +Table 4: Examples of deictic phrases that help in grounded guidance by specifying contextual spatial locations. + +# 4. Observations and Tasks from HoloAssist + +Here we present observations from HoloAssist in Section 4.1 and identify a few open problems that are necessary components for an interactive AI assistant. In Section 4.2, we define new tasks and benchmarks with HoloAssist. + +# 4.1. Observations from HoloAssist + +Correlation between mistakes and intervention. We find that the response time for the human assistant to intervene in the procedure and correct mistakes depends on the severity level of the mistakes. If mistakes are critical, the instructors proactively interrupt the student immediately (within less than 5 seconds duration) while other mistakes are either self-corrected by the task performer or corrected later by the instructor. + +In Table 3, we present the top actions that need immediate intervention and the top actions that are self-corrected by the task performers or corrected later by the instructors. We notice that actions related to linear tasks, where the task progression is stalled if steps are not followed in order, are often intervened immediately by the instructors. For example, "insert joy con controller", and "place tray" etc. In contrast, the lazily edited corrections are related to furniture assembly such as the actions "drop allen wrench", "drop hex socket head", etc. These are tasks where mistakes are unclear in every stage, and the user often intuitively adjusts their steps. + +Grounded guidance. We also notice that the instructions from human instructors are often grounded in the 3D environment. An important aspect of grounded guidance is the + +ability to communicate about the physical world by pointing to things in context. Spatial deixis refers to phrases that are used to locate things in space and to express direction and distance. The deictic analysis of the transcripts in HoloAssist reveals a wide set of words that indicate specific and relative locations and directions, especially during interventions to correct mistakes. A list of some prominent examples is shown in Table 4. + +# 4.2. Benchmark Tasks + +Inspired by the observations above, we think it is important for an interactive AI assistant to have a good world state estimation model that can detect mistakes and predict whether to intervene in the task procedure. Besides, augmenting instructions with spatial guidance can be useful for AI agents. To this end, we introduce new mistake detection, intervention prediction and 3D hand pose forecasting tasks for interactive and grounded guidance. Additionally, we benchmark models on action recognition tasks following the convention in [12]. + +Mistake detection is defined following the convention [31] but applied to fine-grained actions in our benchmark. We take the features from the fine-grained action clips from the beginning of the coarse-grained action until the end of the current action clip, and the model predicts a label from {correct, mistake}. The task is challenging given that the class distribution is highly skewed, with around $6\%$ mistakes among the fine-grained actions. + +Intervention type prediction is to predict the intervention types given an input of a window of 1, 3, or 5 seconds before the intervention. This newly proposed benchmark is to test if the model can correctly figure out the correct intervention types during task completion. Currently, HoloAssist includes 3 intervention types, and we report the precision and recall of each intervention type. + +3D hand pose forecasting is another new benchmark introduced by HoloAssist. Existing action forecasting work [12] mostly focuses on providing semantic labels of future actions and does not provide explicit 3D guidance on hand poses. Predicting 3D hand poses can be useful for various applications [4], and it can augment instructions and spatially guide users in different tasks. In this benchmark, we take 3 seconds inputs similar to other 3D body location forecasting literature [41] and forecast the continuous 3D hand poses for the next 0.5, 1.0, and 1.5 seconds. The evaluation metric is the average of mean per joint position error over time in centimeters compared to ground truth. To have a proper evaluation metric that can help 3D action guidance, we remove the mistakes from the action sequences and only forecast 3D hand pose for the correct labels. + +# 5. Experiments + +In this section, we provide the evaluation results of the proposed benchmarks. We will start with the standard action recognition benchmarks, and then we will present the results of the newly proposed benchmarks. We also provide ablations of different sensor modalities to understand the roles of different sensors in various tasks. We hope the baseline results can guide future research in this space. + +Implementation details. We adopt TimeSformer [9], a state-of-the-art vision transformer (ViT) [13] based video model, as the backbone and change the head with a different number of classes for different benchmarks. We modify the original TimeSformers to perform multimodal learning by introducing additional tokens for different modalities and embedding layers to encode the additional sensor modalities. Specifically, we use $26 \times 2$ tokens for both left and right hands (one token for one hand joint), one token for eye gaze, one token for head poses, and 196 tokens for depth. We can enable and/or disable different modalities during training and evaluation. Detailed configurations are available in supplementary materials. Note that we consider the resulting model a vanilla multimodal model and serve as a baseline for future studies. + +We randomly split the 2221 sessions into train, validation, and test sets following a ratio of $70\%$ , $10\%$ , and $20\%$ on a per-task basis, which includes 1545 sessions for training, 213 sessions for validation, and 463 sessions for testing. We also synchronize all the modalities according to the video stream and keep the frame rate at 30 fps by sub-sampling other modalities in our experiments. + +During training, the model is trained for 15 epochs with an initial learning rate of 0.01 and a batch size of 64 using stochastic gradient descent. We divide the learning rate by 10 at the epoch 11 and 14. For each input segment, we randomly sample 8 frames/data points within the segment. We train our models with 4xA6000 GPU machines, and the fine-grained action recognition runs usually take about one day. We also train the model with random initialization and an ImageNet pre-trained ViT backbone. + +For our hand pose forecasting benchmark, which is a regression task, we adopt Seq2Seq model [35] following [19]. + +Action recognition. We show the fine-grained action recognition results in Table 5 and the coarse-grained action recognition results in Table 6. We can see that initialized with a pre-trained ViT model on ImageNet; all models can achieve around $35\%$ top-1 accuracy on fine-grained action recognition and $50\%$ top-1 accuracy on coarse-grained action recognition. This is comparable to baselines of other fine-grained action benchmarks [31] ( $23\%$ top-1 accuracy), + +We notice that the pre-trained model on ImageNet may reduce the influence of other modalities that are not pretrained. If trained from scratch, we can see from Table 5 + +
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB34.83/68.6042.14/78.9666.81/90.0435.26/69.3442.56/79.5367.19/90.360.03/0.1710.86/36.3338.01/66.48
Hands20.86/43.9235.38/65.7637.10/63.4221.13/44.5035.72/66.3037.50/64.060.00/0.0110.11/25.477.30/15.54
R+H35.06/68.9542.45/79.4267.05/90.0135.49/69.7142.87/80.0367.43/90.320.03/0.1611.05/33.7138.39/66.85
R+H+E35.27/68.6942.92/79.1167.03/89.9635.72/69.4243.33/79.6767.45/90.290.03/0.1811.99/37.6435.96/65.17
R+H+E+I34.80/68.2642.24/78.8866.65/89.7635.23/69.0042.63/79.4667.03/90.080.03/0.1712.92/35.9638.20/65.17
ScratchRGB18.78/48.0928.45/65.4343.72/73.6919.03/48.7228.74/66.0044.09/74.210.00/0.016.55/22.6615.73/34.27
Hands23.94/47.5839.79/68.7639.34/65.7624.26/48.2140.18/69.3039.79/66.430.00/0.0010.67/28.655.62/16.29
RGB + Eyes20.86/50.2730.92/67.5945.28/75.2921.13/50.9331.22/68.2045.65/75.800.00/0.018.05/21.7217.60/37.08
+ +Table 5: Fine-grained action recognition benchmark results. We report baseline results of (multimodal) TimeSformer [9] models on HoloAssist. Here R, H, I denote RGB, hands, and head pose (whose sensor source is similar as IMUs). When starting from a pre-trained ViT model on ImageNet, all the models have higher performance than training from scratch. However, with a pre-trained image model, the influence of other modalities is reduced. If the models are trained from scratch, we can see that adding hands can improve the prediction of verbs and lead to better results than the baseline RGB only model. Simply concatenating more sensors as inputs may not necessarily lead to improved performance. + +
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB50.91/86.8960.51/93.4573.35/95.9053.40/89.5362.78/95.3975.00/96.700.37/2.2420.00/58.8943.89/81.67
Hands22.20/54.1635.12/72.3437.43/68.3323.44/57.0836.75/75.2538.84/70.450.00/0.126.11/20.5612.22/30.56
R+H50.80/86.5459.71/93.3673.20/95.7853.27/89.1261.94/95.2075.09/96.570.37/2.2820.00/60.5639.44/81.67
R+H+E50.35/85.5158.85/93.2173.67/95.6352.65/88.0360.85/94.8975.28/96.320.53/2.2823.33/63.3345.00/83.33
R+H+E+I50.18/86.0759.00/93.2173.64/96.2552.49/88.7861.22/95.0775.41/96.910.50/2.1219.44/60.0042.22/84.44
+ +Table 6: Coarse-grained action recognition results. The overall trend is similar to fine-grained action recognition. + +
Mods.F-scoreCorrectMistake
PrecisionRecallPrecisionRecall
Random27.7160.8710.2215.0046.15
RGB35.1182.5651.8212.9626.92
Hands40.1992.6852.4112.531.25
R+H36.1885.5143.079.6811.54
R+H+E32.0888.5742.7611.4350.00
+ +(bottom) that adding hands can improve the prediction of verbs and lead to better results than the RGB-only model. + +Mistake detection. In Table $7^{\dagger}$ , we present the results of mistake detection where the features are extracted from the pre-trained action recognition models shown in Table 5. Here we find the hand poses information benefits the task and outperforms the other modalities. + +Intervention type prediction. For intervention prediction, + +Table 7: Mistake detection results. Similar to action recognition, adding hand poses to RGB $(\mathbb{R} + \mathbb{H})$ improves an RGB-only model from 35.11 to 36.18 points. What's different is that we find the hands only model outperforms other modalities in mistake detection, achieving 40.19 points. + +
Mods.OverallConfirmCorrectFollow-up
Prec.RecallPrec.RecallPrec.RecallPrec.Recall
RGB13.9333.330.000.000.000.0041.79100
R+H24.8933.6432.144.500.000.0042.5296.43
R+E25.5533.7333.9010.360.000.0042.7690.83
R+H+E48.3137.5939.1140.9361.119.9144.761.93
RGB(Pt)37.5437.7442.3127.5027.3336.6142.9749.11
+ +Table 8: Intervention type prediction results. The classes in benchmark are highly skewed. Here E refers to eye gaze. We can see the predictions of the baseline models (i.e., RGB, R+H, R+E) favor predicting follow-up instructions while ignoring correcting mistakes (recall is 0). Training more modalities $(\mathbb{R} + \mathbb{H} + \mathbb{E})$ alleviates this issue and even outperforms the RGB model (RGB (Pt)) in overall precision. + +we show the results in Table $8^{\dagger}$ . We can see adding hands and eye gaze $(\mathbb{R} + \mathbb{H} + \mathbb{E})$ can significantly boost the overall precision and recall to $48.31\%$ and $37.59\%$ , improving about 35 and 4 percentage points over RGB. This may be because eye gaze is a forecasting signal, as people often look at the regions before the action starts, which can assist the models to attend to important regions for better anticipation. + +![](images/0a74fddbfad7c04202018bde1239cb7b5bc2874a9e3084f517958d0212b8c949.jpg) + +![](images/596ddaea51e8b8524c682d58e0abdfe8d8a84c9e0e953952144dc7dac79a7f37.jpg) +Time +Figure 8: Qualitative visualization for 3D hand pose forecasting. We visualize the ground-truth hand joint positions (i.e., hand pose, visualized in Green) and the prediction of the hand pose for the next 1.5 seconds (visualized in Red). The input to the model is a 3-second long clip ahead of the prediction. The task is challenging as hands often move quickly. As we can see from the figure, the predicted hand pose is more off from the grounded truth in the longer future. + +
Mods.Mean Error Distance (cm, ↓)
0.5 sec1.0 sec1.5 sec
Static-H9.3413.9116.70
Hands9.8010.6811.25
H+E9.8010.7011.25
H+E+I9.8010.6911.25
R+H+E9.7310.6511.22
R+H+E+I9.7210.6211.19
+ +Table 9: 3D hand pose forecasting benchmark results. We report the mean per joint position error for 0.5 1.0, 1.5 seconds (lower the better). The static hand baseline (Static-H) refers to using the last frame of the input. The Seq2Seq [35] model trained using hands only H, and using only hand H achieves better results than Static-H. + +3D hand pose forecasting. Table 9 shows that given only the 3D poses of hands as input, the model can perform with the accuracy of 9.80, 10.68, and 11.25 centimeters, the average of mean per joint position error, for 0.5, 1, and 1.5 seconds, respectively. We should note that this task is challenging as hands can move quickly within a window of 0.5 seconds. Compared to static-H, which uses the last 3D hand pose, our baseline (H) already outperforms it. In Figure 8, we show the visualization of the hand pose forecasting. + +Importance of hand poses and eye gaze. As we can already see from the Tables 5, 7, 8, 9, the 3D hand poses and eye gaze can help the model prediction to recognize actions, detect mistakes, and understand users' intentions. These modalities can augment the commonly used RGB images for better performance. Simply concatenating more modalities (e.g., depth, head poses) as inputs may not necessarily lead to a more capable model, as those modalities may need specialized encoders or model architectures to process them simultaneously. We believe HoloAssist will enable and foster further research in multi-modal learning in this direction. + +# 6. Conclusion and Future Work + +In this work, we identified and explored several important problems with building an interactive AI assistant in the physical world. We introduced a large-scale multimodal egocentric video dataset, HoloAssist, containing rich information about human interaction in an assistive task completion setting. The task performer wears a HoloLens 2 headset while completing various object-centric manipulation tasks. The real-time video feed from the headset is sent to a remote instructor who provides verbal guidance to the task performer. HoloAssist captures seven raw sensor modalities during the interaction, and among them, we found hand pose and eye gaze are useful information sources for an interactive AI agent. By augmenting the data with additional third-person manual annotations on action segments, mistakes, and intervention types, we constructed new benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we believe is a necessary component for an interactive and grounded AI assistant. As a first step in this direction, this work also leaves room for future work to improve upon (e.g., annotating object poses in the data, investigating object-centric models of affordance and manipulations in AI assistance, etc.). We believe HoloAssist, coupled with the associated benchmarks and tooling will benefit future research into building competent AI assistants for everyday tasks in the real world. + +# Acknowledgement + +We thank all the 222 data collectors who participated in the study and acknowledge the hard work of the annotation team led by Megan Yuan and Dan Luo from DataTang Technology Inc. We thank Nick Saw for the help in the data collection software and Yale Song and Vibhav Vineet from the Computer Vision Group at Microsoft Research for the early discussion. We also thank the feedback from colleagues at MSR. + +# References + +[1] Microsoft dynamics 365 guides. https://dynamics.microsoft.com/en-us/mixed-reality/guides/.1 +[2] Microsoft hololens. https://www.microsoft.com/en-us/hololens.2 +[3] Platform for situated intelligence hololens capture tools. https://github.com/microsoft/psi/tree/master/Sources/MixedReality/HoloLensCapture.2,4 +[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 7 +[5] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 1 +[6] Gregory Aist, John Dowding, Beth Ann Hockey, Manny Rayner, James Hieronymus, D Bohus, B Boven, Nate Blaylock, Ellen Campana, S Early, et al. Talking through procedures: An intelligent space station procedure assistant. In Proceedings of the tenth conference on European chapter of the Association for Computational Linguistics-Volume 2, pages 187-190, 2003. 1, 2 +[7] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Ivan Laptev, Josef Sivic, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2 +[8] Sean Andrist, Michael Gleicher, and Bilge Mutlu. Looking coordinated: Bidirectional gaze mechanisms for collaborative interaction with virtual characters. In Proceedings of the 2017 CHI conference on human factors in computing systems, pages 2571-2582, 2017. 1 +[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 7, 8 +[10] Dan Bohus, Sean Andrist, Ashley Feniello, Nick Saw, Mihai Jalobeanu, Patrick Sweeney, Anne Loomis Thompson, and Eric Horvitz. Platform for situated intelligence, 2021. 4 +[11] Dan Bohus and Alexander I Rudnicky. Larri: A language-based maintenance and repair assistant. Spoken multimodal human-computer dialogue in mobile environments, pages 203-218, 2005. 2 +[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonio Furnari, Jian Ma, Evangelos Kazakos, Davide Molti + +santi, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 3, 4, 7 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7 +[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4 +[15] W Lewis Johnson and Jeff Rickel. Steve: An animated pedagogical agent for procedural training in virtual environments. ACM SIGART Bulletin, 8(1-4):16-21, 1997. 1, 2 +[16] Julia Kiseleva, Ziming Li, Mohammad Aliannejadi, Shrestha Mohanty, Maartje ter Hoeve, Mikhail Burtsev, Alexey Skrynnik, Artem Zholus, Aleksandr Panov, Kavya Srinet, et al. Interactive grounded language understanding in a collaborative environment: Iglu 2021. In NeurIPS 2021 Competitions and Demonstrations Track, pages 146-161. PMLR, 2022. 4 +[17] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2, 3 +[18] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3 +[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 7 +[20] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2 +[21] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3 +[22] Ramesh Manuvinakurike, Sovan Biswas, Giuseppe Raffa, Richard Beckwith, Anthony Rhodes, Meng Shi, Gesem Gudino Mejia, Saurav Sahay, and Lama Nachman. Human in the loop approaches in multi-modal conversational task guidance system development. arXiv preprint arXiv:2211.01824, 2022. 2 +[23] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl;dw? summarizing instructional videos with task relevance and cross-modal saliency. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 540–557, 2022. 1, 2 +[24] Rashmiranjan Nayak, Umesh Chandra Pati, and Santos Kumar Das. A comprehensive review on deep learning-based methods for video anomaly detection. Image and Vision Computing, 106:104078, 2021. 3 +[25] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2017-2025, 2022. 2, 4 +[26] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8494-8502, 2018. 2, 3, 4 +[27] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, June 2019. 1 +[28] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 2 +[29] Charles Rich and Candace L Sidner. Diamondhelp: A generic collaborative task guidance system. AI Magazine, 28(2):33-33, 2007. 1, 2 +[30] Charles Rich, Candace L Sidner, and Neal Lesh. Collagen: Applying collaborative discourse theory to human-computer interaction. AI magazine, 22(4):15-15, 2001. 2 + +[31] F. Sener, D. Chatterjee, D. Shelepop, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. CVPR 2022. 2, 3, 4, 7 +[32] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022. 1 +[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4 +[34] Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martin-Martin, Fei Xia, Kent Elliott Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, Karen Liu, et al. Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In Conference on Robot Learning, pages 477-490. PMLR, 2022. 3, 4 +[35] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27, 2014. 7, 9 +[36] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 3, 4 +[37] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning (CoRL), 2019. 1 +[38] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, pages 485-501. Springer, 2022. 1 +[39] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), Oct. 2022. 2, 3 +[40] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 3 +[41] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 7 +[42] Artem Zholus, Alexey Skrynnik, Shrestha Mohanty, Zoya Volovikova, Julia Kiseleva, Artur Szlam, Marc-Alexandre Cote, and Aleksandr I Panov. Iglu gridworld: Simple and fast environment for embodied dialog agents. arXiv preprint arXiv:2206.00142, 2022. 2, 3 +[43] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 14(10):2537-2550, 2019. 3 + +[44] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, + +editors, Computer Vision - ECCV 2022, pages 395-412, 2022. 3 \ No newline at end of file diff --git a/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/images.zip b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c7f1386ab7a68a63ae17a6b24de033e30f3390b5 --- /dev/null +++ b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc1bf9fa084b8255364b99297a6f673be5f51192323ab651f9c2ff9945cbeb04 +size 879995 diff --git a/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/layout.json b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17984e7c6a9430c20b4826fda04c920f5024be30 --- /dev/null +++ b/2023/HoloAssist_ an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World/layout.json @@ -0,0 +1,7509 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 465, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 465, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 465, + 138 + ], + "type": "text", + "content": "HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": "Xin Wang" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Taein Kwon" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Mahdi Rad" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Bowen Pan" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Ishani Chakraborty" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Sean Andrist" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Dan Bohus" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Ashley Feniello" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Bugra Tekin" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Felipe Vieira Frujeri" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Neel Joshi" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Marc Pollefeys" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " Microsoft " + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 62, + 160, + 559, + 203 + ], + "type": "text", + "content": " ETH Zurich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 290, + 555 + ], + "type": "text", + "content": "Building an interactive AI assistant that can perceive, reason, and collaborate with humans in the real world has been a long-standing pursuit in the AI community. This work is part of a broader research effort to develop intelligent agents that can interactively guide humans through performing tasks in the physical world. As a first step in this direction, we introduce HoloAssist, a large-scale egocentric human interaction dataset, where two people collaboratively complete physical manipulation tasks. The task performer executes the task while wearing a mixed-reality headset that captures seven synchronized data streams. The task instructor watches the performer's egocentric video in real time and guides them verbally. By augmenting the data with action and conversational annotations and observing the rich behaviors of various participants, we present key insights into how human assistants correct mistakes, intervene in the task completion procedure, and ground their instructions to the environment. HoloAssist spans 166 hours of data captured by 350 unique instructor-performer pairs. Furthermore, we construct and present benchmarks on mistake detection, intervention type prediction, and hand forecasting, along with detailed analysis. We expect HoloAssist will provide an important resource for building AI assistants that can fluidly collaborate with humans in the real world. Data can be downloaded at https://holoassist.github.io/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 599, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 288, + 695 + ], + "type": "text", + "content": "Recent years have witnessed incredible progress in general-purpose AI agents that assist humans with various open-world tasks, especially in the digital world. AI systems powered by large language models (LLMs) like ChatGPT [27] can answer users' questions and assist them with various text-based tasks. However, these AI assistants do not have sufficient first-hand experience in the physical world and thus cannot perceive world states and actively intervene" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 341, + 235, + 510, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 235, + 510, + 244 + ], + "spans": [ + { + "bbox": [ + 341, + 235, + 510, + 244 + ], + "type": "text", + "content": "Two-Person Interactive Assistive Task Completion" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 246, + 426, + 389 + ], + "blocks": [ + { + "bbox": [ + 307, + 246, + 426, + 389 + ], + "lines": [ + { + "bbox": [ + 307, + 246, + 426, + 389 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 426, + 389 + ], + "type": "image", + "image_path": "a1dda76f6bd78d81d0df44b1d520ffd96c73633b26be16a4df7c6b6a0a1076a3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 246, + 545, + 389 + ], + "blocks": [ + { + "bbox": [ + 427, + 246, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 427, + 246, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 427, + 246, + 545, + 389 + ], + "type": "image", + "image_path": "7633283a5d89767f41f1e99ea3c91a2917798a2d8eab6c4d403ec5ff07f4303d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 403, + 371, + 456 + ], + "blocks": [ + { + "bbox": [ + 307, + 403, + 371, + 456 + ], + "lines": [ + { + "bbox": [ + 307, + 403, + 371, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 371, + 456 + ], + "type": "image", + "image_path": "87fcba3aa50b0e7330ac13ab705e577d5cecca34184a6521f6678610e9f82686.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 456, + 373, + 474 + ], + "lines": [ + { + "bbox": [ + 312, + 456, + 373, + 474 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 373, + 474 + ], + "type": "text", + "content": "(1) RGB, (2) eye gaze, (3) hand pose" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 485, + 547, + 569 + ], + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 569 + ], + "type": "text", + "content": "Figure 1: HoloAssist features a two-person interactive assistive task completion setting. The task performer wears an AR device and completes the tasks while the captured data is streamed over the network to a remote instructor watching it on the laptop. The instructor provides verbal guidance to the student. HoloAssist includes seven modalities captured live and human annotated text descriptions as the 8th modality." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 372, + 403, + 436, + 456 + ], + "blocks": [ + { + "bbox": [ + 372, + 403, + 436, + 456 + ], + "lines": [ + { + "bbox": [ + 372, + 403, + 436, + 456 + ], + "spans": [ + { + "bbox": [ + 372, + 403, + 436, + 456 + ], + "type": "image", + "image_path": "5e501e0ad178f56304f6fd0c67d2171de4501fbb580efbe8924d44c112ed7af3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 456, + 427, + 474 + ], + "lines": [ + { + "bbox": [ + 383, + 456, + 427, + 474 + ], + "spans": [ + { + "bbox": [ + 383, + 456, + 427, + 474 + ], + "type": "text", + "content": "(4) head pose, (5) depth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 437, + 403, + 509, + 456 + ], + "blocks": [ + { + "bbox": [ + 437, + 403, + 509, + 456 + ], + "lines": [ + { + "bbox": [ + 437, + 403, + 509, + 456 + ], + "spans": [ + { + "bbox": [ + 437, + 403, + 509, + 456 + ], + "type": "image", + "image_path": "61e5bda88e7a330a35e155563e39782e0835a37940b97082f348a3448971c1a7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 460, + 499, + 469 + ], + "lines": [ + { + "bbox": [ + 446, + 460, + 499, + 469 + ], + "spans": [ + { + "bbox": [ + 446, + 460, + 499, + 469 + ], + "type": "text", + "content": "(6) IMU, (7) audio" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 511, + 405, + 544, + 456 + ], + "blocks": [ + { + "bbox": [ + 511, + 405, + 544, + 456 + ], + "lines": [ + { + "bbox": [ + 511, + 405, + 544, + 456 + ], + "spans": [ + { + "bbox": [ + 511, + 405, + 544, + 456 + ], + "type": "image", + "image_path": "2feb99fad4a437d27f11d01deddfaa2377790abb610f10e00484df729312c4be.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 460, + 537, + 468 + ], + "lines": [ + { + "bbox": [ + 515, + 460, + 537, + 468 + ], + "spans": [ + { + "bbox": [ + 515, + 460, + 537, + 468 + ], + "type": "text", + "content": "(8) text" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 581, + 441, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 581, + 441, + 593 + ], + "spans": [ + { + "bbox": [ + 306, + 581, + 441, + 593 + ], + "type": "text", + "content": "in the task completion procedure." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": "Building an AI assistant that can perceive, reason and interact in the physical world has attracted attention from researchers across different fields in computer vision [7, 23, 37, 38], human-computer interaction [6, 8, 15, 29], robotics [5, 32], and industrial practitioners. For example, AR Guides [1], which aims to guide users to complete complex tasks, has become popular with the development of augmented reality (AR) devices. However, existing systems often rely on pre-defined instructions or formulate the virtual assistant as a question answering [37, 38] or video under" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 206, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 206, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 206, + 712 + ], + "type": "text", + "content": "*Co-first authors; †Work done at Microsoft" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20270" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 272, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 272, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 272, + 83 + ], + "type": "text", + "content": "standing problem [7, 23] without real-world interaction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 288, + 167 + ], + "type": "text", + "content": "In another line of work, researchers have developed simulation environments like Habitat [21, 36], VirtualHome [26], and AI2-Thor [17] to build AI agents that can interact with the physical world and collaboratively achieve new tasks [42]. Still, a large gap remains in transferring these agents to the real world, and the interaction between agents is largely simplified compared to real-world human interaction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 168, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 312 + ], + "type": "text", + "content": "In this work, we focus on the challenges of developing intelligent agents that share perspectives with humans and interactively guide human users through performing tasks in the physical world. As a first step, we introduce HoloAssist, a large-scale egocentric human interaction dataset to explore and identify the open problems in this direction. As shown in Figure 1, the task performer wears an AR headset* to capture data while completing the tasks. An instructor watches the real-time egocentric video feed remotely and verbally guides the performer. We have developed and open-sourced a data capture tool [3] using a distributed server-client setup to enable data streaming and multimodal data capture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 479 + ], + "type": "text", + "content": "HoloAssist contains 166 hours of data captured by 222 diverse participants forming 350 unique instructor-performer pairs and carrying out 20 object-centric manipulation tasks. The objects range from common electronic devices to rare objects in factories and specialized labs. The tasks are generally challenging for first-time participants, requiring instructor assistance for successful completion. Seven raw sensor modalities are captured, including RGB, depth, head pose, 3D hand pose, eye gaze, audio, and IMU, to aid in the understanding of human intentions, estimating world states, predicting future actions, and so on. Finally, the dataset is augmented with third-person manual annotations consisting of a text summary, intervention types, mistake annotation, and action segments of the videos as illustrated in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 288, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 623 + ], + "type": "text", + "content": "We have observed several characteristics demonstrated by human instructors from HoloAssist. First, instructors are often proactive with precisely timed interventions. Instead of waiting until mistakes happen, instructors provide follow-up instructions when the task performer appears confused. Second, the verbal guidance from the instructors tends to be concise and grounded in the task performer's environment. The instructions are often framed as spatial deictics to aid the task performer in spatial directions and distances in the 3D world. Moreover, instructors often have a good world model estimation and can detect whether mistakes disrupt task completion and then adjust the guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "text", + "content": "We take a step further and introduce new tasks and benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we conjecture are essential modules for an intelligent assistant. Additionally, we benchmark the dataset on action classification and anticipation tasks and provide empirical results to understand the role of" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 544, + 266 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 544, + 266 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 544, + 266 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 544, + 266 + ], + "type": "image", + "image_path": "e59c7abb08bdb1905e93e9f0daa7a4c7c694db4ee2cfec00f526720044a5cacf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 269, + 547, + 341 + ], + "lines": [ + { + "bbox": [ + 304, + 269, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 547, + 341 + ], + "type": "text", + "content": "Figure 2: HoloAssist includes action and conversational annotations, in addition to text summaries of the videos, to indicate the mistakes and interventions in task completion. mistake or correct attributes are associated with each fine-grained action. A purpose label is associated with every utterance to indicate the type of verbal intervention." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 354, + 547, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 402 + ], + "type": "text", + "content": "different modalities in various tasks. We hope our dataset, findings, and tooling can inspire and provide rich resources for future work on designing interactive AI assistants and situated AI assistance applications in the real world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 415, + 392, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 392, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 392, + 427 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 431, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 467 + ], + "type": "text", + "content": "Our work closely connects with several lines of work in computer vision, especially egocentric vision, embodied AI, and human-computer interaction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 471, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 638 + ], + "type": "text", + "content": "Interactive AI assistants. Building interactive agents that can assist humans to carry out tasks in the world—real or virtual—has been a long-standing problem in different areas of AI and HCI [6, 11, 15, 22, 25, 29, 30]. As far back as 1997, Johnson and Rickel introduced \"Steve\" [15], an early pedagogical agent that aims to help students learn procedural tasks in VR. Recent efforts have focused on new modeling approaches and data collection techniques for training conversational task guidance assistants, such as model-in-the-loop wizard-of-oz [22] and human-human interaction to mimic robot actions in simulated environments [25]. In this work, we revisit this problem and provide a systematic study of real-world human interaction, and we also provide rich sensor information to push the frontiers of the research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Egocentric video datasets. Egocentric perspectives often convey rich information about the users' intentions. A shared perspective between the users and the human or AI assistants is useful for the assistants to provide more timely and grounded guidance. In computer vision, several egocentric video datasets [12, 14, 18, 20, 28, 31, 39] have emerged in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 703, + 235, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 235, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 235, + 712 + ], + "type": "text", + "content": "*We use HoloLens 2 [2] for data capture in this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20271" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 276, + 220 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 276, + 220 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 276, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 276, + 220 + ], + "type": "image", + "image_path": "d315d1ce8ba45d9afd6d5dff34e8b2d7fb05c5687289a28292445d0e73f834d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 288, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 288, + 262 + ], + "type": "text", + "content": "Figure 3: HoloAssist was collected by participants diverse in ages, occupations, genders, and geography. This helps us to study a diverse set of users with different backgrounds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 418 + ], + "type": "text", + "content": "the community. EPIC-KICHENS [12] is a widely adopted egocentric video dataset capturing kitchen activities. The recent Ego4D dataset [14] is the largest egocentric video dataset in the wild that provides a comprehensive database for egocentric perception in the 3D world. In contrast to earlier egocentric video datasets, HoloAssist features a multiperson interactive task completion setting, where human interaction during the procedure provides a rich source for designing AI assistants to be more proactive and grounded in the environment. Yet our work can benefit from the rich knowledge and representation learned from existing datasets like Ego4D and is complementary in nature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 288, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 602 + ], + "type": "text", + "content": "Mistake detection. One of the key observations in human interaction is that human assistants tend to correct mistakes and proactively intervene in the task completion procedure. While there has been a large body of work for video-based anomaly detection [24, 40, 43, 44], mistake detection in procedural settings has been under-explored. The Assembly101 dataset [31] proposes a mistake detection task to predict if a coarse-grained action segment is a mistake or correction. By contrast, HoloAssist emphasizes fine-grained actions since instructors may intervene when they spot a student's mistake in an active intervention setting rather than wait until the whole step (i.e., coarse-grained action) is completed. In addition, we propose a new intervention prediction task and, in combination, enable a more comprehensive understanding of interactions in an assistive task completion setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Multimodality and interaction. Human interaction with the world is multimodal as we see, speak, and touch objects in the environment. In HoloAssist, we collect seven raw sensor modalities that might help understand humans' intentions, estimate the world states, predict future actions, etc. Previous datasets [12, 14, 18, 39] often provide a limited subset of modalities. Although not every sensor may currently be relevant for the downstream tasks, the seven synchronized sensor modalities provided in HoloAssist will" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 308, + 84, + 547, + 177 + ], + "blocks": [ + { + "bbox": [ + 309, + 74, + 455, + 83 + ], + "lines": [ + { + "bbox": [ + 309, + 74, + 455, + 83 + ], + "spans": [ + { + "bbox": [ + 309, + 74, + 455, + 83 + ], + "type": "text", + "content": "OBJECT SCALES OBJECT CATEGORIES" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 84, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 308, + 84, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 308, + 84, + 547, + 177 + ], + "type": "table", + "html": "
SmallGoPro, Nintendo Switch, DSLR
MediumPortable printer, Computer, Nespresso machine
BigStandalone printer, big coffee machine, IKEA fur-niture (stool, utility cart, tray Table, nightstand)
RareNavVis laser scanner, ATV motor cycle, wheel belt, circuit breaker
", + "image_path": "c07a3d583e7ac31dc183a8e1ea9d4e067f35775db90e5275f3c2fdd86d28a633.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 185, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 245 + ], + "type": "text", + "content": "Table 1: HoloAssist includes 16 objects with diverse scales. Apart from common objects used in daily life, HoloAssist includes rare equipment from mechanical labs. 20 tasks are object-centric manipulation tasks for each object and the 4 IKEA furniture has both assembly and disassembly tasks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 255, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 278 + ], + "type": "text", + "content": "give practitioners more potential for designing multimodal agents and models even beyond the scope of this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 281, + 547, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 401 + ], + "type": "text", + "content": "Embodied simulation platforms. There is an emerging interest in embodied agents that can perceive, reason, and act in the 3D world. Researchers [17, 21, 26, 34, 36, 42] build various simulation environments to learn such embodied agents. IGLU [42] aims to build interactive agents that learn to solve a task while being provided with grounded natural language instructions in a collaborative environment based on Minecraft, a popular video game. HoloAssist complements this line of work by providing more realistic human interaction and real-world sensor perception." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 408, + 520, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 408, + 520, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 520, + 420 + ], + "type": "text", + "content": "3. HoloAssist: Human Assistance Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 498 + ], + "type": "text", + "content": "In this work, we introduce HoloAssist which features a two-person collaboration scenario and can be used to situate AI assistance in the physical world. We will start by describing the data collection and statistics in Section 3.1 and annotations in Section 3.2, before diving into the observations and benchmarks in the following sections." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 501, + 468, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 468, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 468, + 512 + ], + "type": "text", + "content": "3.1. Data Collection and Statistics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 519, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 674 + ], + "type": "text", + "content": "Tasks and objects. We consider multi-step goal-oriented tasks involving 16 objects ranging from familiar objects often used in daily life to rare objects sometimes used in labs and factories as summarized in Table 1. We consider small electronics like a GoPro, DSLR camera, and Nintendo, office appliances like a Nespresso machine and printer, IKEA furniture, and objects in labs such as a laser scanner, motor cycle, and circuit breaker. We have designed 20 tasks involving physical manipulation of these objects, e.g., changing batteries, changing belts, furniture assembly, machine setup, etc. There is one task per object except for the IKEA furniture, which has assembly and disassembly tasks. Detailed task instructions are in supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "content": "Participants and collection procedure. We recruited 222 participants to form 350 unique pairs of instructors and performers for data collection. Figure 3 shows the demograph-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20272" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 286, + 205 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 286, + 205 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 286, + 205 + ], + "type": "image", + "image_path": "d5ccbe40edbf69e07f487a1f0638393bb22c06e1b99f0dfe1b8de8a310128bc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 288, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 288, + 247 + ], + "type": "text", + "content": "Figure 4: The skill level of participants (0-10) for the tasks is self-reported by the participants. The skill levels roughly reflect the length of the sessions though they might be noisy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 288, + 378 + ], + "type": "text", + "content": "ics of the participants. Before data collection, the participants review the IRB forms to acknowledge the privacy and ethics standards (more details in supplementary materials). The instructors are informed about the task in detail. The participants playing the role of performers are only given a rough description of the tasks and scenarios beforehand and interacted with the objects based on their understanding. The instructors provide verbal guidance as the performers set out to complete the tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 380, + 288, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 490 + ], + "type": "text", + "content": "In Figure 4 (top), we show the distribution of the performers' familiarity with the tasks measured by a self-reported score (0-10) by the participants. We show the average length and the outliers of the recorded sessions in Figure 4 (bottom) to give a rough idea of how the participant's skill levels may lead to increased session variance. The participants' diverse skill levels and backgrounds provide rich information about the user behaviors and diverse interaction between the instructors and performers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 493, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 288, + 625 + ], + "type": "text", + "content": "Data capture tool. We leveraged the Platform for Situated Intelligence framework [10] to develop and open-source a distributed application for data capture using HoloLens 2 [3]. A client process running on the device captured the sensor data while displaying a rectangular hologram frame around the user's visual field of view to guide their attention downward and keep the task actions in view of the sensors. Sensor data was streamed live over the network to a server application that ran on a PC and persisted the data to disk. This distributed setup allows for collecting longer uninterrupted sessions without reaching the device storage capacity limits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "Comparison with other datasets. While there is no direct comparison of datasets with the same setup with HoloAssist, we list out different aspects of our dataset and compare it with related datasets in Table 2. HoloAssist is among the largest egocentric video datasets and features a multi-person collaboration setting, which is a unique addition to the field. In addition, HoloAssist relates to work on multi-agent collab-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 74, + 545, + 170 + ], + "blocks": [ + { + "bbox": [ + 310, + 74, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 310, + 74, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 310, + 74, + 545, + 170 + ], + "type": "image", + "image_path": "2167e90cae7efab397313ab6c45ebb97871cd1263b186cc50add8293825b9f9a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 547, + 214 + ], + "type": "text", + "content": "Figure 5: Data distribution of 166 hours captured by HoloAssist. (left) number of sessions per activity, and (right) total length of sessions in minutes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 228, + 545, + 378 + ], + "blocks": [ + { + "bbox": [ + 309, + 228, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 309, + 228, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 545, + 378 + ], + "type": "table", + "html": "
DatasetSettingsCollaborative &InteractiveInstructional &Procedural# real video hours
Epic-Kitchen-100 [12]CookingXX100
Assembly101 [31]Toy assemblyX✓167
Ego4D [14]Daily-life task++3,670
VirtualHome [26]Household taskX✓$
ALFRED [33]Household taskX✓$
Habitat [36]Home assistanceX✓$
BEHAVIOR [34]Daily-life task✓✓$
IGLU [16]Collaborative building*✓✓$
TEACH [25]Household task✓✓$
HoloAssist (ours)Assistive task✓✓166
", + "image_path": "baf6155d9b16c2a2a27ce778542a091da73d08cf84dababc4af4b1fb31da6ab0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 386, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 547, + 506 + ], + "type": "text", + "content": "Table 2: Comparison to related datasets and simulation platforms. HoloAssist features a multi-person collaborative setting which is a unique addition to existing egocentric datasets in the real world. HoloAssist provides a set of instructional and procedure videos with multi-turn dialogues. Procedure tasks are defined as following a set of defined steps or procedures to achieve a specific goal, deviation from the procedure can be construed to be a mistake. HoloAssist spans 166 hours and 2,221 sessions. §: simulation, *: Minecraft-like, +: partially included." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 529, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 529, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 545, + 553 + ], + "type": "text", + "content": "orative simulation environments with a distinct characteristic of real-world sensor data and real-world human interaction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 562, + 387, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 387, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 387, + 573 + ], + "type": "text", + "content": "3.2. Annotations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 578, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 547, + 625 + ], + "type": "text", + "content": "To better understand the actions and interactions in the dataset, we provide several sets of third-person manual annotations for text summaries, action segments, mistake attributes, and intervention attributes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 630, + 545, + 713 + ], + "type": "text", + "content": "Language annotations. We asked the annotators to watch the video and write a paragraph to describe the activities in the videos. The description focuses on describing the hand actions in the procedure. The third-person post hoc summary provides insights into the key moments during the interactive task completion. These could be used to build a comprehensive set of instructions for task completion. We" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20273" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 213, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 213, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 213, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 213, + 173 + ], + "type": "image", + "image_path": "375558581e7d68ca45d5197ce1d72ed02d74bf0b5c65005114d38c3d1044f44a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 216, + 73, + 376, + 189 + ], + "blocks": [ + { + "bbox": [ + 216, + 73, + 376, + 189 + ], + "lines": [ + { + "bbox": [ + 216, + 73, + 376, + 189 + ], + "spans": [ + { + "bbox": [ + 216, + 73, + 376, + 189 + ], + "type": "image", + "image_path": "819b3b8f63b91102c59be1acb7da84f6bc09cfc260b9b339f28b39029b4ca4af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 91, + 544, + 200 + ], + "blocks": [ + { + "bbox": [ + 381, + 91, + 544, + 200 + ], + "lines": [ + { + "bbox": [ + 381, + 91, + 544, + 200 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 544, + 200 + ], + "type": "image", + "image_path": "c9bf663b27903c482cd3796d12cf365ae1042926e1f93e99a58b0887f7d210cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 546, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 546, + 237 + ], + "type": "text", + "content": "Figure 6: Data distribution of the coarse-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 251, + 213, + 348 + ], + "blocks": [ + { + "bbox": [ + 50, + 251, + 213, + 348 + ], + "lines": [ + { + "bbox": [ + 50, + 251, + 213, + 348 + ], + "spans": [ + { + "bbox": [ + 50, + 251, + 213, + 348 + ], + "type": "image", + "image_path": "7b72b865bf72ee6741949174835bc214a4abffbf66a45dd99f28e8baec40ffb4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 251, + 376, + 364 + ], + "blocks": [ + { + "bbox": [ + 215, + 251, + 376, + 364 + ], + "lines": [ + { + "bbox": [ + 215, + 251, + 376, + 364 + ], + "spans": [ + { + "bbox": [ + 215, + 251, + 376, + 364 + ], + "type": "image", + "image_path": "5eb05a5bb82c1824cca9087c0f5fccead80dd303fd6545cb855a9b40fbbb269b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 380, + 251, + 544, + 376 + ], + "blocks": [ + { + "bbox": [ + 380, + 251, + 544, + 376 + ], + "lines": [ + { + "bbox": [ + 380, + 251, + 544, + 376 + ], + "spans": [ + { + "bbox": [ + 380, + 251, + 544, + 376 + ], + "type": "image", + "image_path": "dba5cfe3da510c226c789124af38c68913adf962104a4d000a36313a43a88442.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 379, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 379, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 546, + 415 + ], + "type": "text", + "content": "Figure 7: Data distribution of the fine-grained actions. (left) duration of the actions in seconds, (middle) 30 most frequently occurring verbs, and (right) 30 most frequently occurring nouns. We can see that most fine-grained actions last less than 2 seconds, and there is a long tail distribution in actions, verbs, and nouns." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "type": "text", + "content": "also provide the transcriptions of the conversations in the video. With this set of annotations, we can understand the difference between third-person post hoc summaries and real-time conversations during the activities. More examples are in supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 503, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 670 + ], + "type": "text", + "content": "Coarse-grained action annotations. The coarse actions usually describe a high-level step in the task (e.g., change battery of a GoPro in the GoPro set up task) and can be divided into multiple fine-grained actions. To deal with the open-world setting, we ask the annotators to write a sequence (e.g., man changes the battery of GoPro) to describe the coarse actions and also identify the active verb-noun pair and optionally an adjective for the noun for benchmarking purposes (e.g., change battery). The dataset includes 414 coarse-grained actions with 90 nouns and 39 verbs. The distribution of the actions follows a long-tail distribution shown in Figure 6, where 185 actions are considered head classes while the rest are considered tail classes according to the action frequency for evaluation purposes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Fine-grained action annotations. Fine-grained actions are the low-level atomic actions (e.g., press button, grab screw, etc.) for completing a step in the task, usually lasting for 1-2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 436, + 546, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 546, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 546, + 508 + ], + "type": "text", + "content": "seconds. The fine-grained actions are presented in a verb (adj.)-noun pair format. There are 1887 fine-grained actions with 165 nouns and 49 verbs. For a more comprehensive evaluation, we create a split of head actions with 1082 top actions and 805 tail actions. Distributions of fine-grained actions are shown in Figure 7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 514, + 546, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 546, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 546, + 621 + ], + "type": "text", + "content": "As mentioned earlier, noun and verb vocabularies are not pre-defined but gradually built through annotation. We ask the annotators to enter a new verb and noun if they cannot find it in the vocabulary. After the data is annotated, we ask the annotators to revisit and check the tail classes and see if they are repetitive to head classes. Due to the open-world nature, some verb and noun combinations might be interchangeable with others in the list. We show some examples in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": "Mistake annotation. Each fine-grained action is labeled as either correct or mistake, as indicated in Figure 2. Mistakes include the ones that are \"self-corrected by the task performers\", are \"verbally corrected by the instructors\", and \"are not corrected labeled\". Our human annotators annotate all three mistake types separately, but for benchmark evaluation, we will consolidate them into one mistake class." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20274" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 71, + 276, + 148 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 276, + 148 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 276, + 148 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 276, + 148 + ], + "type": "table", + "html": "
IMMEDIATE INTERVENTIONLAZY INTERVENTION
insert screwdrop allen wrench
approach buttondrop hex socket head
insert joy con controllerdrop screw
place trayscrew hex socket head
pull battery doorscrew screw
", + "image_path": "8287c72822c297f46dcda59a370f2f50ef45e8fd1d9dd794a1dcd44c15721390.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 285 + ], + "type": "text", + "content": "We defer the detailed study of differentiating whether and how the mistakes are corrected to future work. To ensure the annotation quality, we additionally ask the third-person annotators to explain why the action is a mistake and also assign a mapping to every mistake that is corrected by an instructor verbally to the conversation sentence whose type is \"instructor correcting mistakes\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 468 + ], + "type": "text", + "content": "Intervention annotation. Since instructors assist the task performers verbally, we annotate the conversation between the instructors and performers to reflect the interventions in task completion. We annotate each conversation sentence with two attributes to indicate the conversation types and the conversation initiator. The conversation initiators can either be the \"task performer\" or the \"instructor\". And the sentence purpose types can be the instructor \"correcting mistakes\", \"answering questions\", \"following up with more instructions\", \"confirming previous actions\". \"describing the high-level task\", \"opening/closing remarks\", or the task performer starting the conversation to ask questions. We asked human annotators to watch the videos and use their best judgment to annotate the roles of different conversations in the videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 471, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 541 + ], + "type": "text", + "content": "In our benchmarks, we consider 3 intervention types: correcting mistakes, following up with more instructions, and confirming the previous action as they are more related to physical actions in the task procedure. Figure 2 shows examples of conversational interventions. More examples can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "Audit process. Annotations are done by professional annotators based on the following process. The annotators first take a pass on the video to add the fine-grained actions, coarse-grained actions, conversation, and text summary, along with the associated annotation elements for each event. After self-review, the annotated data is passed to an independent reviewer for auditing, and the mistakes are fixed directly or sent back to the original annotator for updates. The annotations finally go through a targeted review to check the open-ended text fields like narration, action sentences, and conversation transcriptions to ensure consistency. Before the annotations were delivered, we applied a list of constraints to systematically check the annotations to further dig out the wrong annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 72, + 495, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 495, + 83 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 495, + 83 + ], + "type": "text", + "content": "Spatial deixis from the intervention transcripts" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 311, + 86, + 541, + 295 + ], + "blocks": [ + { + "bbox": [ + 47, + 156, + 288, + 180 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 288, + 180 + ], + "type": "text", + "content": "Table 3: Top mistakes that are corrected immediately (left) or later and sometimes through self-correction (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 86, + 541, + 295 + ], + "lines": [ + { + "bbox": [ + 311, + 86, + 541, + 295 + ], + "spans": [ + { + "bbox": [ + 311, + 86, + 541, + 295 + ], + "type": "table", + "html": "
"You should press the button that's on the body of the camera just at the right of the lens."
"You should leave the bolt, like it was before."
"The button is on the other side of the Switch."
"The SD card comes from the right slot, on the right hand side and it opens by using a knob next to the screen on the right bottom side of the screen."
"Currently the tray is facing you, please rotate so the back is facing you."
"You should put it the other way around. It's upside down."
"Please start by removing the screw of the top shelf first."
"And now, ehm, it is, it should be the other one that should be on top of the other."
"To the right of that, there is a tiny little square."
", + "image_path": "610a33ebf1d71a8ba687d719d88b0f9897f64913e1cef3d0c6f00f3a1207224c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "type": "text", + "content": "Table 4: Examples of deictic phrases that help in grounded guidance by specifying contextual spatial locations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 349, + 528, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 528, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 528, + 361 + ], + "type": "text", + "content": "4. Observations and Tasks from HoloAssist" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 369, + 547, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 547, + 416 + ], + "type": "text", + "content": "Here we present observations from HoloAssist in Section 4.1 and identify a few open problems that are necessary components for an interactive AI assistant. In Section 4.2, we define new tasks and benchmarks with HoloAssist." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 426, + 469, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 469, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 469, + 437 + ], + "type": "text", + "content": "4.1. Observations from HoloAssist" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 447, + 547, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 541 + ], + "type": "text", + "content": "Correlation between mistakes and intervention. We find that the response time for the human assistant to intervene in the procedure and correct mistakes depends on the severity level of the mistakes. If mistakes are critical, the instructors proactively interrupt the student immediately (within less than 5 seconds duration) while other mistakes are either self-corrected by the task performer or corrected later by the instructor." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 674 + ], + "type": "text", + "content": "In Table 3, we present the top actions that need immediate intervention and the top actions that are self-corrected by the task performers or corrected later by the instructors. We notice that actions related to linear tasks, where the task progression is stalled if steps are not followed in order, are often intervened immediately by the instructors. For example, \"insert joy con controller\", and \"place tray\" etc. In contrast, the lazily edited corrections are related to furniture assembly such as the actions \"drop allen wrench\", \"drop hex socket head\", etc. These are tasks where mistakes are unclear in every stage, and the user often intuitively adjusts their steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Grounded guidance. We also notice that the instructions from human instructors are often grounded in the 3D environment. An important aspect of grounded guidance is the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20275" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "ability to communicate about the physical world by pointing to things in context. Spatial deixis refers to phrases that are used to locate things in space and to express direction and distance. The deictic analysis of the transcripts in HoloAssist reveals a wide set of words that indicate specific and relative locations and directions, especially during interventions to correct mistakes. A list of some prominent examples is shown in Table 4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 153, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 153, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 153, + 201 + ], + "type": "text", + "content": "4.2. Benchmark Tasks" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 213, + 289, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 332 + ], + "type": "text", + "content": "Inspired by the observations above, we think it is important for an interactive AI assistant to have a good world state estimation model that can detect mistakes and predict whether to intervene in the task procedure. Besides, augmenting instructions with spatial guidance can be useful for AI agents. To this end, we introduce new mistake detection, intervention prediction and 3D hand pose forecasting tasks for interactive and grounded guidance. Additionally, we benchmark models on action recognition tasks following the convention in [12]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 289, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 289, + 435 + ], + "type": "text", + "content": "Mistake detection is defined following the convention [31] but applied to fine-grained actions in our benchmark. We take the features from the fine-grained action clips from the beginning of the coarse-grained action until the end of the current action clip, and the model predicts a label from {correct, mistake}. The task is challenging given that the class distribution is highly skewed, with around " + }, + { + "bbox": [ + 46, + 339, + 289, + 435 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 339, + 289, + 435 + ], + "type": "text", + "content": " mistakes among the fine-grained actions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 443, + 288, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 288, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 288, + 527 + ], + "type": "text", + "content": "Intervention type prediction is to predict the intervention types given an input of a window of 1, 3, or 5 seconds before the intervention. This newly proposed benchmark is to test if the model can correctly figure out the correct intervention types during task completion. Currently, HoloAssist includes 3 intervention types, and we report the precision and recall of each intervention type." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": "3D hand pose forecasting is another new benchmark introduced by HoloAssist. Existing action forecasting work [12] mostly focuses on providing semantic labels of future actions and does not provide explicit 3D guidance on hand poses. Predicting 3D hand poses can be useful for various applications [4], and it can augment instructions and spatially guide users in different tasks. In this benchmark, we take 3 seconds inputs similar to other 3D body location forecasting literature [41] and forecast the continuous 3D hand poses for the next 0.5, 1.0, and 1.5 seconds. The evaluation metric is the average of mean per joint position error over time in centimeters compared to ground truth. To have a proper evaluation metric that can help 3D action guidance, we remove the mistakes from the action sequences and only forecast 3D hand pose for the correct labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 92, + 547, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 92, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 304, + 92, + 547, + 176 + ], + "type": "text", + "content": "In this section, we provide the evaluation results of the proposed benchmarks. We will start with the standard action recognition benchmarks, and then we will present the results of the newly proposed benchmarks. We also provide ablations of different sensor modalities to understand the roles of different sensors in various tasks. We hope the baseline results can guide future research in this space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 180, + 547, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 357 + ], + "type": "text", + "content": "Implementation details. We adopt TimeSformer [9], a state-of-the-art vision transformer (ViT) [13] based video model, as the backbone and change the head with a different number of classes for different benchmarks. We modify the original TimeSformers to perform multimodal learning by introducing additional tokens for different modalities and embedding layers to encode the additional sensor modalities. Specifically, we use " + }, + { + "bbox": [ + 304, + 180, + 547, + 357 + ], + "type": "inline_equation", + "content": "26 \\times 2" + }, + { + "bbox": [ + 304, + 180, + 547, + 357 + ], + "type": "text", + "content": " tokens for both left and right hands (one token for one hand joint), one token for eye gaze, one token for head poses, and 196 tokens for depth. We can enable and/or disable different modalities during training and evaluation. Detailed configurations are available in supplementary materials. Note that we consider the resulting model a vanilla multimodal model and serve as a baseline for future studies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "text", + "content": "We randomly split the 2221 sessions into train, validation, and test sets following a ratio of " + }, + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 360, + 547, + 443 + ], + "type": "text", + "content": " on a per-task basis, which includes 1545 sessions for training, 213 sessions for validation, and 463 sessions for testing. We also synchronize all the modalities according to the video stream and keep the frame rate at 30 fps by sub-sampling other modalities in our experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 445, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 552 + ], + "type": "text", + "content": "During training, the model is trained for 15 epochs with an initial learning rate of 0.01 and a batch size of 64 using stochastic gradient descent. We divide the learning rate by 10 at the epoch 11 and 14. For each input segment, we randomly sample 8 frames/data points within the segment. We train our models with 4xA6000 GPU machines, and the fine-grained action recognition runs usually take about one day. We also train the model with random initialization and an ImageNet pre-trained ViT backbone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 553, + 548, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 548, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 548, + 578 + ], + "type": "text", + "content": "For our hand pose forecasting benchmark, which is a regression task, we adopt Seq2Seq model [35] following [19]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "text", + "content": "Action recognition. We show the fine-grained action recognition results in Table 5 and the coarse-grained action recognition results in Table 6. We can see that initialized with a pre-trained ViT model on ImageNet; all models can achieve around " + }, + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "text", + "content": " top-1 accuracy on fine-grained action recognition and " + }, + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "text", + "content": " top-1 accuracy on coarse-grained action recognition. This is comparable to baselines of other fine-grained action benchmarks [31] (" + }, + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 304, + 581, + 547, + 676 + ], + "type": "text", + "content": " top-1 accuracy)," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "We notice that the pre-trained model on ImageNet may reduce the influence of other modalities that are not pretrained. If trained from scratch, we can see from Table 5" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20276" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 198 + ], + "type": "table", + "html": "
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB34.83/68.6042.14/78.9666.81/90.0435.26/69.3442.56/79.5367.19/90.360.03/0.1710.86/36.3338.01/66.48
Hands20.86/43.9235.38/65.7637.10/63.4221.13/44.5035.72/66.3037.50/64.060.00/0.0110.11/25.477.30/15.54
R+H35.06/68.9542.45/79.4267.05/90.0135.49/69.7142.87/80.0367.43/90.320.03/0.1611.05/33.7138.39/66.85
R+H+E35.27/68.6942.92/79.1167.03/89.9635.72/69.4243.33/79.6767.45/90.290.03/0.1811.99/37.6435.96/65.17
R+H+E+I34.80/68.2642.24/78.8866.65/89.7635.23/69.0042.63/79.4667.03/90.080.03/0.1712.92/35.9638.20/65.17
ScratchRGB18.78/48.0928.45/65.4343.72/73.6919.03/48.7228.74/66.0044.09/74.210.00/0.016.55/22.6615.73/34.27
Hands23.94/47.5839.79/68.7639.34/65.7624.26/48.2140.18/69.3039.79/66.430.00/0.0010.67/28.655.62/16.29
RGB + Eyes20.86/50.2730.92/67.5945.28/75.2921.13/50.9331.22/68.2045.65/75.800.00/0.018.05/21.7217.60/37.08
", + "image_path": "fb02190f86d99bac4ce16252bfc9d8110408c195083ff3c62bf7b298e805cc48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 284, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 46, + 206, + 547, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 280 + ], + "type": "text", + "content": "Table 5: Fine-grained action recognition benchmark results. We report baseline results of (multimodal) TimeSformer [9] models on HoloAssist. Here R, H, I denote RGB, hands, and head pose (whose sensor source is similar as IMUs). When starting from a pre-trained ViT model on ImageNet, all the models have higher performance than training from scratch. However, with a pre-trained image model, the influence of other modalities is reduced. If the models are trained from scratch, we can see that adding hands can improve the prediction of verbs and lead to better results than the baseline RGB only model. Simply concatenating more sensors as inputs may not necessarily lead to improved performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 284, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 49, + 284, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 547, + 376 + ], + "type": "table", + "html": "
Mods.All Classes AccuracyHead Classes AccuracyTail Classes Accuracy
Top1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 NounTop1 / 5 ActTop1 / 5 VerbTop1 / 5 Noun
PretrainedRGB50.91/86.8960.51/93.4573.35/95.9053.40/89.5362.78/95.3975.00/96.700.37/2.2420.00/58.8943.89/81.67
Hands22.20/54.1635.12/72.3437.43/68.3323.44/57.0836.75/75.2538.84/70.450.00/0.126.11/20.5612.22/30.56
R+H50.80/86.5459.71/93.3673.20/95.7853.27/89.1261.94/95.2075.09/96.570.37/2.2820.00/60.5639.44/81.67
R+H+E50.35/85.5158.85/93.2173.67/95.6352.65/88.0360.85/94.8975.28/96.320.53/2.2823.33/63.3345.00/83.33
R+H+E+I50.18/86.0759.00/93.2173.64/96.2552.49/88.7861.22/95.0775.41/96.910.50/2.1219.44/60.0042.22/84.44
", + "image_path": "6e1742f3edfc8b6590cfc46293b88ef240be4d62bb39d5b52c466f981c31c252.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 409, + 274, + 507 + ], + "blocks": [ + { + "bbox": [ + 63, + 384, + 529, + 397 + ], + "lines": [ + { + "bbox": [ + 63, + 384, + 529, + 397 + ], + "spans": [ + { + "bbox": [ + 63, + 384, + 529, + 397 + ], + "type": "text", + "content": "Table 6: Coarse-grained action recognition results. The overall trend is similar to fine-grained action recognition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 409, + 274, + 507 + ], + "lines": [ + { + "bbox": [ + 61, + 409, + 274, + 507 + ], + "spans": [ + { + "bbox": [ + 61, + 409, + 274, + 507 + ], + "type": "table", + "html": "
Mods.F-scoreCorrectMistake
PrecisionRecallPrecisionRecall
Random27.7160.8710.2215.0046.15
RGB35.1182.5651.8212.9626.92
Hands40.1992.6852.4112.531.25
R+H36.1885.5143.079.6811.54
R+H+E32.0888.5742.7611.4350.00
", + "image_path": "e6fcef07bc401cf82a65c78022b47f933f6d3cc8c36e6605a7e769813eee32a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 587, + 288, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 288, + 612 + ], + "type": "text", + "content": "(bottom) that adding hands can improve the prediction of verbs and lead to better results than the RGB-only model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "content": "Mistake detection. In Table " + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "inline_equation", + "content": "7^{\\dagger}" + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "content": ", we present the results of mistake detection where the features are extracted from the pre-trained action recognition models shown in Table 5. Here we find the hand poses information benefits the task and outperforms the other modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 679, + 288, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 679, + 288, + 692 + ], + "spans": [ + { + "bbox": [ + 47, + 679, + 288, + 692 + ], + "type": "text", + "content": "Intervention type prediction. For intervention prediction," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 410, + 547, + 514 + ], + "blocks": [ + { + "bbox": [ + 46, + 515, + 288, + 576 + ], + "lines": [ + { + "bbox": [ + 46, + 515, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 288, + 576 + ], + "type": "text", + "content": "Table 7: Mistake detection results. Similar to action recognition, adding hand poses to RGB " + }, + { + "bbox": [ + 46, + 515, + 288, + 576 + ], + "type": "inline_equation", + "content": "(\\mathbb{R} + \\mathbb{H})" + }, + { + "bbox": [ + 46, + 515, + 288, + 576 + ], + "type": "text", + "content": " improves an RGB-only model from 35.11 to 36.18 points. What's different is that we find the hands only model outperforms other modalities in mistake detection, achieving 40.19 points." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 410, + 547, + 514 + ], + "lines": [ + { + "bbox": [ + 308, + 410, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 547, + 514 + ], + "type": "table", + "html": "
Mods.OverallConfirmCorrectFollow-up
Prec.RecallPrec.RecallPrec.RecallPrec.Recall
RGB13.9333.330.000.000.000.0041.79100
R+H24.8933.6432.144.500.000.0042.5296.43
R+E25.5533.7333.9010.360.000.0042.7690.83
R+H+E48.3137.5939.1140.9361.119.9144.761.93
RGB(Pt)37.5437.7442.3127.5027.3336.6142.9749.11
", + "image_path": "89c68b7886cd30f48f92d09c1e4c60ac708934364c545b7716176b04547ebb9e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 547, + 619 + ], + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 619 + ], + "type": "text", + "content": "Table 8: Intervention type prediction results. The classes in benchmark are highly skewed. Here E refers to eye gaze. We can see the predictions of the baseline models (i.e., RGB, R+H, R+E) favor predicting follow-up instructions while ignoring correcting mistakes (recall is 0). Training more modalities " + }, + { + "bbox": [ + 304, + 522, + 547, + 619 + ], + "type": "inline_equation", + "content": "(\\mathbb{R} + \\mathbb{H} + \\mathbb{E})" + }, + { + "bbox": [ + 304, + 522, + 547, + 619 + ], + "type": "text", + "content": " alleviates this issue and even outperforms the RGB model (RGB (Pt)) in overall precision." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "we show the results in Table " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "8^{\\dagger}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ". We can see adding hands and eye gaze " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbb{R} + \\mathbb{H} + \\mathbb{E})" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " can significantly boost the overall precision and recall to " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "48.31\\%" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "37.59\\%" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", improving about 35 and 4 percentage points over RGB. This may be because eye gaze is a forecasting signal, as people often look at the regions before the action starts, which can assist the models to attend to important regions for better anticipation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 62, + 702, + 182, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 182, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 182, + 712 + ], + "type": "text", + "content": "Evaluated on " + }, + { + "bbox": [ + 62, + 702, + 182, + 712 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 62, + 702, + 182, + 712 + ], + "type": "text", + "content": " of the entire data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20277" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 547, + 129 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 547, + 129 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 547, + 129 + ], + "type": "image", + "image_path": "0a74fddbfad7c04202018bde1239cb7b5bc2874a9e3084f517958d0212b8c949.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 132, + 545, + 192 + ], + "blocks": [ + { + "bbox": [ + 49, + 132, + 545, + 192 + ], + "lines": [ + { + "bbox": [ + 49, + 132, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 49, + 132, + 545, + 192 + ], + "type": "image", + "image_path": "596ddaea51e8b8524c682d58e0abdfe8d8a84c9e0e953952144dc7dac79a7f37.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 197, + 302, + 206 + ], + "lines": [ + { + "bbox": [ + 280, + 197, + 302, + 206 + ], + "spans": [ + { + "bbox": [ + 280, + 197, + 302, + 206 + ], + "type": "text", + "content": "Time" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 219, + 548, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 548, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 548, + 268 + ], + "type": "text", + "content": "Figure 8: Qualitative visualization for 3D hand pose forecasting. We visualize the ground-truth hand joint positions (i.e., hand pose, visualized in Green) and the prediction of the hand pose for the next 1.5 seconds (visualized in Red). The input to the model is a 3-second long clip ahead of the prediction. The task is challenging as hands often move quickly. As we can see from the figure, the predicted hand pose is more off from the grounded truth in the longer future." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 62, + 280, + 276, + 380 + ], + "blocks": [ + { + "bbox": [ + 62, + 280, + 276, + 380 + ], + "lines": [ + { + "bbox": [ + 62, + 280, + 276, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 276, + 380 + ], + "type": "table", + "html": "
Mods.Mean Error Distance (cm, ↓)
0.5 sec1.0 sec1.5 sec
Static-H9.3413.9116.70
Hands9.8010.6811.25
H+E9.8010.7011.25
H+E+I9.8010.6911.25
R+H+E9.7310.6511.22
R+H+E+I9.7210.6211.19
", + "image_path": "0f6d3b71983cf546ecee133acd3b9805d2bfe00b638e8b0819d2878d6cc617ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 389, + 288, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 462 + ], + "type": "text", + "content": "Table 9: 3D hand pose forecasting benchmark results. We report the mean per joint position error for 0.5 1.0, 1.5 seconds (lower the better). The static hand baseline (Static-H) refers to using the last frame of the input. The Seq2Seq [35] model trained using hands only H, and using only hand H achieves better results than Static-H." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 470, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 288, + 579 + ], + "type": "text", + "content": "3D hand pose forecasting. Table 9 shows that given only the 3D poses of hands as input, the model can perform with the accuracy of 9.80, 10.68, and 11.25 centimeters, the average of mean per joint position error, for 0.5, 1, and 1.5 seconds, respectively. We should note that this task is challenging as hands can move quickly within a window of 0.5 seconds. Compared to static-H, which uses the last 3D hand pose, our baseline (H) already outperforms it. In Figure 8, we show the visualization of the hand pose forecasting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Importance of hand poses and eye gaze. As we can already see from the Tables 5, 7, 8, 9, the 3D hand poses and eye gaze can help the model prediction to recognize actions, detect mistakes, and understand users' intentions. These modalities can augment the commonly used RGB images for better performance. Simply concatenating more modalities (e.g., depth, head poses) as inputs may not necessarily lead to a more capable model, as those modalities may need specialized encoders or model architectures to process them simultaneously. We believe HoloAssist will enable and foster further research in multi-modal learning in this direction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 281, + 471, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 471, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 471, + 294 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 303, + 548, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 548, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 548, + 602 + ], + "type": "text", + "content": "In this work, we identified and explored several important problems with building an interactive AI assistant in the physical world. We introduced a large-scale multimodal egocentric video dataset, HoloAssist, containing rich information about human interaction in an assistive task completion setting. The task performer wears a HoloLens 2 headset while completing various object-centric manipulation tasks. The real-time video feed from the headset is sent to a remote instructor who provides verbal guidance to the task performer. HoloAssist captures seven raw sensor modalities during the interaction, and among them, we found hand pose and eye gaze are useful information sources for an interactive AI agent. By augmenting the data with additional third-person manual annotations on action segments, mistakes, and intervention types, we constructed new benchmarks on mistake detection, intervention type prediction, and 3D hand pose forecasting, which we believe is a necessary component for an interactive and grounded AI assistant. As a first step in this direction, this work also leaves room for future work to improve upon (e.g., annotating object poses in the data, investigating object-centric models of affordance and manipulations in AI assistance, etc.). We believe HoloAssist, coupled with the associated benchmarks and tooling will benefit future research into building competent AI assistants for everyday tasks in the real world." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 616, + 397, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 397, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 397, + 628 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 636, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 547, + 715 + ], + "type": "text", + "content": "We thank all the 222 data collectors who participated in the study and acknowledge the hard work of the annotation team led by Megan Yuan and Dan Luo from DataTang Technology Inc. We thank Nick Saw for the help in the data collection software and Yale Song and Vibhav Vineet from the Computer Vision Group at Microsoft Research for the early discussion. We also thank the feedback from colleagues at MSR." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "20278" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Microsoft dynamics 365 guides. https://dynamics.microsoft.com/en-us/mixed-reality/guides/.1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 146 + ], + "type": "text", + "content": "[2] Microsoft hololens. https://www.microsoft.com/en-us/hololens.2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "text", + "content": "[3] Platform for situated intelligence hololens capture tools. https://github.com/microsoft/psi/tree/master/Sources/MixedReality/HoloLensCapture.2,4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 193, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 345 + ], + "type": "text", + "content": "[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do as i can and not as i say: Grounding language in robotic affordances. In arXiv preprint arXiv:2204.01691, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "text", + "content": "[5] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 402, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 479 + ], + "type": "text", + "content": "[6] Gregory Aist, John Dowding, Beth Ann Hockey, Manny Rayner, James Hieronymus, D Bohus, B Boven, Nate Blaylock, Ellen Campana, S Early, et al. Talking through procedures: An intelligent space station procedure assistant. In Proceedings of the tenth conference on European chapter of the Association for Computational Linguistics-Volume 2, pages 187-190, 2003. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 479, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 288, + 523 + ], + "type": "text", + "content": "[7] Jean-Baptiste Alayrac, Piotr Bojanowski, Nishant Agrawal, Ivan Laptev, Josef Sivic, and Simon Lacoste-Julien. Unsupervised learning from narrated instruction videos. In Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 524, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 288, + 579 + ], + "type": "text", + "content": "[8] Sean Andrist, Michael Gleicher, and Bilge Mutlu. Looking coordinated: Bidirectional gaze mechanisms for collaborative interaction with virtual characters. In Proceedings of the 2017 CHI conference on human factors in computing systems, pages 2571-2582, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 580, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 288, + 612 + ], + "type": "text", + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, page 4, 2021. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "type": "text", + "content": "[10] Dan Bohus, Sean Andrist, Ashley Feniello, Nick Saw, Mihai Jalobeanu, Patrick Sweeney, Anne Loomis Thompson, and Eric Horvitz. Platform for situated intelligence, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[11] Dan Bohus and Alexander I Rudnicky. Larri: A language-based maintenance and repair assistant. Spoken multimodal human-computer dialogue in mobile environments, pages 203-218, 2005. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonio Furnari, Jian Ma, Evangelos Kazakos, Davide Molti" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 117 + ], + "type": "text", + "content": "santi, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 547, + 182 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 469 + ], + "type": "text", + "content": "[14] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 470, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 470, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 470, + 547, + 502 + ], + "type": "text", + "content": "[15] W Lewis Johnson and Jeff Rickel. Steve: An animated pedagogical agent for procedural training in virtual environments. ACM SIGART Bulletin, 8(1-4):16-21, 1997. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 503, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 547, + 568 + ], + "type": "text", + "content": "[16] Julia Kiseleva, Ziming Li, Mohammad Aliannejadi, Shrestha Mohanty, Maartje ter Hoeve, Mikhail Burtsev, Alexey Skrynnik, Artem Zholus, Aleksandr Panov, Kavya Srinet, et al. Interactive grounded language understanding in a collaborative environment: Iglu 2021. In NeurIPS 2021 Competitions and Demonstrations Track, pages 146-161. PMLR, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 569, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 547, + 624 + ], + "type": "text", + "content": "[17] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 625, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 680 + ], + "type": "text", + "content": "[18] Taein Kwon, Bugra Tekin, Jan Stühmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "text", + "content": "[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20279" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 162 + ], + "type": "text", + "content": "[20] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "text", + "content": "[21] Manolis Savva*, Abhishek Kadian*, Oleksandr Maksymets*, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 297 + ], + "type": "text", + "content": "[22] Ramesh Manuvinakurike, Sovan Biswas, Giuseppe Raffa, Richard Beckwith, Anthony Rhodes, Meng Shi, Gesem Gudino Mejia, Saurav Sahay, and Lama Nachman. Human in the loop approaches in multi-modal conversational task guidance system development. arXiv preprint arXiv:2211.01824, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 374 + ], + "type": "text", + "content": "[23] Medhini Narasimhan, Arsha Nagrani, Chen Sun, Michael Rubinstein, Trevor Darrell, Anna Rohrbach, and Cordelia Schmid. Tl;dw? summarizing instructional videos with task relevance and cross-modal saliency. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 540–557, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 421 + ], + "type": "text", + "content": "[24] Rashmiranjan Nayak, Umesh Chandra Pati, and Santos Kumar Das. A comprehensive review on deep learning-based methods for video anomaly detection. Image and Vision Computing, 106:104078, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "type": "text", + "content": "[25] Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. Teach: Task-driven embodied agents that chat. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2017-2025, 2022. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 488, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 544 + ], + "type": "text", + "content": "[26] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8494-8502, 2018. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 544, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 578 + ], + "type": "text", + "content": "[27] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, June 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 579, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 644 + ], + "type": "text", + "content": "[28] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1569-1578, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "text", + "content": "[29] Charles Rich and Candace L Sidner. Diamondhelp: A generic collaborative task guidance system. AI Magazine, 28(2):33-33, 2007. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[30] Charles Rich, Candace L Sidner, and Neal Lesh. Collagen: Applying collaborative discourse theory to human-computer interaction. AI magazine, 22(4):15-15, 2001. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[31] F. Sener, D. Chatterjee, D. Shelepop, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. CVPR 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 150 + ], + "type": "text", + "content": "[32] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. arXiv preprint arXiv:2209.05451, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 215 + ], + "type": "text", + "content": "[33] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 217, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 282 + ], + "type": "text", + "content": "[34] Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martin-Martin, Fei Xia, Kent Elliott Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, Karen Liu, et al. Behavior: Benchmark for everyday household activities in virtual, interactive, and ecological environments. In Conference on Robot Learning, pages 477-490. PMLR, 2022. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 283, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 547, + 316 + ], + "type": "text", + "content": "[35] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27, 2014. 7, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 316, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 381 + ], + "type": "text", + "content": "[36] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 383, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 546, + 415 + ], + "type": "text", + "content": "[37] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation. In Conference on Robot Learning (CoRL), 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 416, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 470 + ], + "type": "text", + "content": "[38] Benita Wong, Joya Chen, You Wu, Stan Weixian Lei, Dongxing Mao, Difei Gao, and Mike Zheng Shou. Assistq: Affordance-centric question-driven task completion for egocentric assistant. In European Conference on Computer Vision, pages 485-501. Springer, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 472, + 547, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 547, + 525 + ], + "type": "text", + "content": "[39] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In European conference on computer vision (ECCV), Oct. 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 526, + 547, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 547, + 559 + ], + "type": "text", + "content": "[40] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 559, + 547, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 547, + 603 + ], + "type": "text", + "content": "[41] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao Yu, Yebin Liu, Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. arXiv preprint arXiv:2204.09443, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 604, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 546, + 657 + ], + "type": "text", + "content": "[42] Artem Zholus, Alexey Skrynnik, Shrestha Mohanty, Zoya Volovikova, Julia Kiseleva, Artur Szlam, Marc-Alexandre Cote, and Aleksandr I Panov. Iglu gridworld: Simple and fast environment for embodied dialog agents. arXiv preprint arXiv:2206.00142, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 712 + ], + "type": "text", + "content": "[43] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 14(10):2537-2550, 2019. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "20280" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "text", + "content": "[44] Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 326, + 72, + 547, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 547, + 94 + ], + "type": "text", + "content": "editors, Computer Vision - ECCV 2022, pages 395-412, 2022. 3" + } + ] + } + ], + "index": 1 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20281" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_content_list.json b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c0d252163ad91a8eab00d59766a577fbf4dcc7d --- /dev/null +++ b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_content_list.json @@ -0,0 +1,1546 @@ +[ + { + "type": "text", + "text": "HOLOFUSION: Towards Photo-realistic 3D Generative Modeling", + "text_level": 1, + "bbox": [ + 156, + 130, + 813, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animesh Karnewar UCL", + "bbox": [ + 122, + 181, + 279, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a.karnewar@ucl.ac.uk", + "bbox": [ + 109, + 219, + 290, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Niloy J. Mitra \nUCL", + "bbox": [ + 348, + 181, + 462, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "n.mitra@ucl.ac.uk", + "bbox": [ + 330, + 219, + 482, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrea Vedaldi Meta AI", + "bbox": [ + 531, + 181, + 656, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vedaldi@meta.com", + "bbox": [ + 522, + 219, + 666, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Novotny Meta AI", + "bbox": [ + 720, + 181, + 841, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dnovotny@meta.com", + "bbox": [ + 705, + 219, + 857, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fb702bffad84786de6fa40d73a8b07283a2b36f371f8e95035cf4d4eec0ed98.jpg", + "image_caption": [ + "HoloDiffusion samples", + "Figure 1: We propose HoloFUSION to generate photo-realistic 3D radiance fields by extending the HoloDiffusion method with a jointly trained 2D 'super resolution' network. The independently super-resolved images are fused back into the 3D representation to improve the fidelity of the 3D model via distillation, while preserving the consistency across view changes." + ], + "image_footnote": [], + "bbox": [ + 80, + 263, + 403, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f65173352270c65d0f173fb8aa431b48f989e6d57923901dca62ba92cc94ffa1.jpg", + "image_caption": [ + "Distillation" + ], + "image_footnote": [], + "bbox": [ + 423, + 266, + 557, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6dac8a8c53b07a863c6d44e0d090043390ec06a00c6092664c7f048155911b79.jpg", + "image_caption": [ + "HoloFusion high quality samples" + ], + "image_footnote": [], + "bbox": [ + 573, + 263, + 900, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 594, + 312, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based image generators can now produce high-quality and diverse samples, but their success has yet to fully translate to 3D generation: existing diffusion methods can either generate low-resolution but 3D consistent outputs, or detailed 2D views of 3D objects but with potential structural defects and lacking view consistency or realism. We present HoloFUSION, a method that combines the best of these approaches to produce high-fidelity, plausible, and diverse 3D samples while learning from a collection of multi-view 2D images only. The method first generates coarse 3D samples using a variant of the recently proposed HoloDiffusion generator. Then, it independently renders and upsamples a large number of views of the coarse 3D model, super-resolves them to add detail, and distills those into a single, high-fidelity implicit 3D representation, which also ensures view-consistency of the final renders. The super-resolution network is trained as an integral part of HoloFUSION, end-to-end, and the final distillation uses a", + "bbox": [ + 75, + 628, + 473, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "new sampling scheme to capture the space of super-resolved signals. We compare our method against existing baselines, including DreamFusion, Get3D, EG3D, and HoloDiffusion, and achieve, to the best of our knowledge, the most realistic results on the challenging CO3Dv2 dataset.", + "bbox": [ + 498, + 595, + 893, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 688, + 632, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [32, 7, 31] are at the basis of state-of-the-art 2D image generators which can now produce very high-quality and diverse outputs. However, their success has yet to be translated to 3D and there is no generator that can produce 3D assets of a comparable quality.", + "bbox": [ + 496, + 715, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent attempts at extending diffusion to 3D generation have reported mixed success. Some authors have attempted to apply diffusion directly in 3D [18], or still in 2D but using a 3D-aware neural network [42, 1]. This requires solving two problems: first, finding a suitable 3D representation (e.g., triplane features [4], mesh [19], voxels [18]) that scales well with resolution and is amenable to diffusion;", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22976", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and, second, obtaining a large amount of 3D training data, for example using synthetic models [41, 27], or training the model using only 2D images [18], often via differentiable (volume) rendering [13, 26]. However, the quality of results so far is limited, especially when training on real images.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other authors have proposed to distill 3D objects from pre-trained 2D image generators. For instance Score Distillation Sampling (SDS) [29] can sample 3D objects from a high-quality off-the-shelf 2D diffusion model while requiring no (re)training. However, without any 3D guidance, distillation methods often produce implausible results; for example, they suffer from the 'Janus effect', where details of the front of the object are replicated on its back. They also create overly-smooth outputs that average out inconsistencies arising from the fact that the signal obtained from the 2D model is analogous to sampling independent views of the object (see Sec. 4.2 for examples). Furthermore, distillation methods do not support unconditional sampling, even if the underlying image generator does, as strong language guidance is required to stabilise the 3D reconstruction.", + "bbox": [ + 75, + 167, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose HLOFUSION, a method that combines the best of both approaches. We start from HoloDiffusion [18], a diffusion-based 3D generator. This model can be trained using only a multiview image dataset like [30] and produces outputs that are 3D consistent. However, the output resolution is limited by computation and memory. We augment the base model with a lightweight super-resolution network that upscales the initial renders. Crucially, the 2D super-resolution model is integrated and trained jointly with the 3D generator, end-to-end.", + "bbox": [ + 75, + 397, + 468, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The super-resolution network outputs detailed views of the 3D object, and the underlying 3D generator ensures that the coarse structure of these views is indeed consistent (e.g., avoiding the Janus effect and other structural artifacts). However, the 2D upscaling still progresses independently for different views, which means that fine grained details may still be inconsistent between views. We address this issue by distilling a single, coherent, high quality 3D model of the object from the output of the upsampler. For this, we propose a new distillation technique that efficiently combines several putative super-resolved views of the object into a single, coherent 3D reconstruction.", + "bbox": [ + 75, + 549, + 468, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With this, we are able to train a high-quality 3D generator model purely from real 2D data. This model is capable of generating consistent and detailed 3D objects, which in turn result in view-consistent renderings (see Fig. 1) at a quality not achievable by prior methods.", + "bbox": [ + 75, + 732, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate HoloFusion on real images (CO3Dv2 dataset [30]) and compare with a variety of competing alternatives (e.g., HoloDiffusion [18], Get3D [9], EG3D [4], DreamFusion [40]) demonstrating that view-consistent high-quality 3D generation is possible using our simple, effective, easy-to-implement hybrid approach.", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D generators that use adversarial learning. Generative Adversarial Learning (GAN) [10] learns a generator network so that its \"fake\" samples cannot be distinguished from real images by a second discriminator network. Approaches such as PlatonicGAN [13], HoloGAN [28], and PrGAN [8] introduced 3D structure into the generator network, achieving 3D shape generation with only image-level supervision. Our method is related to those as it renders images from a generated voxel grid, as well as to HoloGAN [28], which renders features and then converts them into an image by a lightweight 2D convolutional network. Other voxel-based 3D generators include VoxGRAF [34] and NeuralVolumes [21].", + "bbox": [ + 496, + 116, + 890, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, 3D generators have built on neural radiance fields [26]. GRAF [33] was the first to adopt the NeRF framework; analogous to PlatonicGAN, they generate the parameters of an MLP which renders realistic images of the object from a random viewpoint. This idea has been improved in StyleNeRF [11] and EG3D [4] by adding a 2D convolutional post-processing step after emission-absorption rendering, which is analogous to our super-resolution network. EG3D also introduced a novel 'Triplane' representation of the radiance field which, in a memory efficient manner, factorises the latter into a triplet of 2D feature planes. EG3D inspired several improvements such as GAUDI [2] and EpiGRAF [37].", + "bbox": [ + 496, + 313, + 890, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mesh-based 3D generators have been explored in [43]. Recently, GET3D [9] replaced the radiance field with a signed distance function to regularise the representation of geometry. The latter is converted into a mesh and rendered in a differentiable manner by using the marching tetrahedral representation [35].", + "bbox": [ + 496, + 510, + 890, + 601 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modeling 3D with diffusion. Diffusion methods [38] have recently become the go-to framework for generative modeling of any kind, including 3D generative modeling. The first applications of diffusion to 3D considered pointcloud generators trained on synthetic data [23, 45, 44].", + "bbox": [ + 496, + 604, + 890, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D distillation of 2D diffusion models. More recently, DreamFusion [29] ported the idea of distillation to diffusion models: they extract a neural radiance field so that its renders match the belief of a pre-trained 2D diffusion generator [32, 7, 31]. They introduce the Score Distillation Sampling (SDS) loss which makes distillation relatively efficient (but still in the order of several minutes for a single 3D sample). Their generation can be conditioned by an image or by a textual description, making the process rather flexible. Magic3D [19] further increases the quality of the output by distilling a mesh-based 3D representation instead of a radiance field.", + "bbox": [ + 496, + 685, + 890, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-conditioned 3D diffusion. The idea of distillation has been applied to few-view conditioned reconstruction in", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22977", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[42, 12, 25, 46, 6]. SparseFusion [46] employs a 3D-based new-view synthesis model [39] followed by a 2D diffusion upsampler. They complete the process by 3D distillation, ensuring that the generated views of the object are consistent. NeRFDiff [12] and 3DiM [42] bypass an explicit 3D model and directly generate new views of an object using a 2D image generator and, in the case of NeRFDiff, refine the results using distillation.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While SparseFusion and NeRFDiff need to be trained on a dataset of object-centric multi-view images with pose information, RealFusion [25] and NeRDi [6] can be used for zero-shot monocular 3D reconstruction, starting from a pretrained 2D diffusion model. Given a single image as input, they automatically generate a prompt for the diffusion model, using a form of prompt inversion, and then use distillation to extract a radiance field.", + "bbox": [ + 75, + 210, + 470, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unconditional generation. Most relevantly to us, unconditional generation, i.e., generation which does not require either text or image conditioning, was explored in [41, 36, 27, 18]. While [27, 36, 41] train generators given synthetic 3D ground truth, similar to us, HoloDiffusion [18] is supervised only with real object-centric images and camera poses. While HoloDiffusion was the first to demonstrate successful training on real image data, its renders contain considerably lower amount of detail than samples from a conventional 2D image generator that uses diffusion. We thus leverage a 2D diffusion upsampler, conditioned on the lower-fidelity HoloDiffusion renders, to distill higher resolution images and, eventually, 3D models.", + "bbox": [ + 75, + 335, + 472, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. HOOFUSION", + "text_level": 1, + "bbox": [ + 76, + 546, + 223, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present HOOFUSION, a method that can learn a high-quality diffusion-based 3D generator from a collection of multiview 2D images. HOOFUSION first obtains an unconditional low-resolution 3D sample using diffusion and then distills a high-resolution 3D radiance field representing a higher-quality version of the generated object. We first summarize the Denoising Diffusion Probabilistic Models (DDPMs) [16] that we utilize in Sec. 3.1. Then, we discuss the low-resolution 3D generator in Sec. 3.2 followed by super-resolution distillation in Sec. 3.3.", + "bbox": [ + 75, + 571, + 468, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: DDPMs", + "text_level": 1, + "bbox": [ + 76, + 731, + 290, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $x = x_0$ be a random vector whose probability distribution $p(x|y)$ we seek to model. The DDPM [16] defines a hierarchy of latent variables $x_{t}$ , $t = 0,\\dots ,T$ and an encoder $q$ comprising a sequence of Gaussian distributions", + "bbox": [ + 75, + 753, + 468, + 815 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} x _ {t - 1}, (1 - \\alpha_ {t}) \\mathbb {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 827, + 468, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t},\\ldots ,\\alpha_{T}$ is a predefined 'noising schedule'. Given knowledge of $x_0$ , a sample $x_{t}$ can be drawn in a closed-form directly from $q(x_{t}|x_{0}) = \\mathcal{N}(x_{t};\\sqrt{\\bar{\\alpha}_{t}} x_{0},(1 - \\bar{\\alpha}_{t})\\mathbb{I})$ , where", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}$ . Hence, we can express $x_{t}$ as:", + "bbox": [ + 500, + 89, + 795, + 108 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\hat {\\epsilon} _ {t} \\left(x _ {0}\\right) = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 114, + 890, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The noising schedule is chosen such that $\\bar{\\alpha}_{T} \\approx 0$ . In this manner, $q(x_{T}|x_{0}) \\approx \\mathcal{N}(x_{T};0,\\mathbb{I})$ is approximately normal, and so is $q(x_{T})$ . To generate a sample $x = x_0$ , we start by sampling $x_{T}$ from this normal distribution and then sample the intermediate latent variables backward. This is done by using a variational approximation of the probabilities $q(x_{t - 1}|x_t)$ given by the Gaussian factors:", + "bbox": [ + 496, + 140, + 890, + 247 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\sqrt {\\bar {\\alpha} _ {t - 1}} D _ {\\theta} \\left(x _ {t}, t\\right), (1 - \\bar {\\alpha} _ {t - 1}) \\mathbb {I}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 255, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{\\theta}$ is a neural network with parameters $\\theta$", + "bbox": [ + 500, + 280, + 823, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The network $D_{\\theta}$ is trained by maximizing the ELBO (Evidence Lower Bound), which reduces to the denoising objective [22]:", + "bbox": [ + 496, + 295, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, \\epsilon , x} \\left[ \\frac {\\bar {\\alpha} _ {t - 1}}{2 (1 - \\bar {\\alpha} _ {t - 1}) ^ {2}} \\| D _ {\\theta} \\left(\\hat {\\epsilon} _ {t} (x), t\\right) - x \\| ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 348, + 890, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t$ is sampled uniformly from $U[1,T]$ . Hence, $D_{\\theta}(x_t,t)$ approximates the clean sample $x_0$ given the noisy sample $x_{t}$ (obtained using (2)).", + "bbox": [ + 500, + 388, + 890, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. HoloDiffusion revisited", + "text_level": 1, + "bbox": [ + 500, + 443, + 714, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a large dataset of 3D models, the framework of Sec. 3.1 could be used to train a corresponding probability distribution. However, such a dataset is not available, and we must instead learn from 2D images of physical 3D objects. Given a dataset containing several views of a large number of objects, we could use image-based reconstruction (e.g., using neural rendering) to obtain corresponding 3D models first, and then use those to train a diffusion model. Instead, we adopt, and slightly upgrade, the HoloDiffusion method [18], which learns a 3D diffusion model directly from the 2D images.", + "bbox": [ + 496, + 465, + 890, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training data. HoloDiffusion learns from a collection $\\mathcal{D}$ of $N$ image sequences $s_i = (I_j^i, C_j^i)_{j=1}^{N_{\\mathrm{frame}}}$ , $i = 1, \\ldots, N$ , where frame $I_j^i \\in \\mathbb{R}^{3 \\times H \\times W}$ is an RGB image and $C_j^i \\in \\mathbb{R}^{4 \\times 4}$ is the corresponding camera projection matrix, collectively defining the motion of the camera.", + "bbox": [ + 496, + 635, + 890, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D representation and rendering. The shape and appearance of the object are represented by a voxel grid $V \\in \\mathbb{R}^{d \\times S \\times S \\times S}$ with resolution $S$ containing a $d$ -dimensional feature vector per voxel. Given a 3D point $\\mathbf{p} \\in \\mathbb{R}^3$ , its opacity $\\sigma(\\mathbf{p}) \\in \\mathbb{R}_+$ and color $c(\\mathbf{p}) \\in \\mathbb{R}_{[0,1]}^3$ are obtained from the voxel grid by an MLP $M_{\\eta}(V(\\mathbf{p}))$ that takes as input the $d$ -dimensional feature vector $V(\\mathbf{p})$ extracted from the grid via trilinear interpolation [20]. The usual emission-absorption model [26, 24] is then used to implement a differentiable rendering function $R_{\\eta}$ , mapping the voxel grid $V$ and the camera viewpoint $C$ into an image $\\hat{I} = R_{\\eta}(V,C)$ , where $\\eta$ are the parameters of the MLP.", + "bbox": [ + 496, + 717, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22978", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8ea85f537f37312e8a7c69fdbeec3bffe7e20f1e7221e40bb31e3cd5b96cc611.jpg", + "image_caption": [ + "Figure 2: Overview. HOOFUSION, which trains the 3D denoiser network $D_{\\theta}$ , is augmented with the 2D 'super-resolution' diffusion model $D_{\\beta}$ . Both models are trained end-to-end by supervising their outputs with 2D photometric error." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training scheme. HoloDiffusion leverages the DDPM framework (revised in the previous paragraphs) to recover the density $p(V)$ over voxel grids $x = V$ encoding plausible real-life objects. In order to train a DDPM on such 3D data, we would need access to ground-truth 3D models $V$ , which are not available. HoloDiffusion addresses this problem by making three changes to DDPM.", + "bbox": [ + 75, + 446, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, it replaces the data denoising loss with a photometric reconstruction loss. Given a pair $(I,C)\\in s$ from one of the training sequences $s$ , it replaces Eq. (4) with $\\mathbb{E}_{t,\\epsilon ,C}\\left[\\| I - R_{\\eta}(D_{\\theta}(\\hat{\\epsilon}_t(V),t),C)\\| ^2\\right]$ where the goal is not to reconstruct the 'clean' volume $V$ (which is unknown), but rather its image $I$ (which is known).", + "bbox": [ + 75, + 551, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, also because the 'clean' volume $V$ is not available, we cannot use Eq. (3) to generate the noisy volumes $V_{t}$ to denoise; the only exception is the last sample $V_{T}$ , which is pure noise. This suggests to adopt a 'double denoising' step. First, pure noise $V_{T}$ is fed into the denoiser to obtain an (approximate) version of $V_{0} = D_{\\theta}(V_{T}, T)$ of the clean volume $V_{0} = V$ . Then, noise is applied to obtain $V_{t} = \\hat{\\epsilon}_{t}(V_{0}) = \\sqrt{\\bar{\\alpha}_{T}} V_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{t}$ according to Eq. (3), and the latter is fed back into the denoiser as above.", + "bbox": [ + 75, + 643, + 470, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, there is the issue that unconditional generation of the clean volume $V_{0}$ from pure noise $V_{T}$ is difficult, especially in the first iterations of training. On the other hand, the problem of view-conditioned generation is considerably easier. Hence, the third idea is to learn a conditional generator, using a variable number of input views. Specifically, given a training sequence $s$ , the method extracts a random subset of frames $\\bar{s} \\subset s$ (which could be empty, which cor-", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "responds to unconditional generation). Then, a feature volume $\\bar{V} = \\Phi (\\bar{s})\\in \\mathbb{R}^{d\\times S\\times S\\times S}$ is obtained from the selected frames. This extracts 2D image features using a pre-trained and frozen 2D image encoder and then pools them in 3D via 'unprojection' [17, 14] into $\\bar{V}$ , where $\\bar{V} = 0$ if $\\bar{s}$ is empty. Finally, these pooled features are used to condition the denoising network $V_{0} = D_{\\theta}(V_{T},\\bar{V},T)$ , which, on average, leads to a simpler reconstruction problem.", + "bbox": [ + 496, + 446, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Putting it all together, the training loss becomes:", + "bbox": [ + 517, + 566, + 839, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta | I, C, \\bar {s}) = \\mathbb {E} _ {t, \\hat {\\epsilon}, V _ {T}} \\left[ \\| \\hat {I} - I \\| ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 590, + 890, + 616 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\hat {I} = R _ {\\eta} \\left(D _ {\\theta} \\left(V _ {t}, \\bar {V}, t\\right), C\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 619, + 890, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {t} = \\hat {\\epsilon} _ {t} (V _ {0}),\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 638, + 723, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {0} = D _ {\\theta} (V _ {T}, \\bar {V}, T),\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 657, + 774, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {V} = \\Phi (\\bar {s}).\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 676, + 715, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This loss is averaged over training sequences $s$ , subsequences $\\bar{s} \\subset s$ , and views $(I, C) \\in s$ therein. Note that this is slightly different than the original HoloDiffusion, where feature volume $\\tilde{V}$ and reconstructed volumes $V_{t}$ overlap as arguments of the denoiser; we found that keeping them separated in the formulation leads to more stable training and additionally allows for view-conditioned generation.", + "bbox": [ + 496, + 702, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. HOOFUSION", + "text_level": 1, + "bbox": [ + 500, + 816, + 633, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The method of Sec. 3.2 learns to generate 3D objects from 2D image supervision only, but the fidelity of the output is limited by the resolution at which the operations are carried out. Increasing resolution is difficult due to the GPU", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22979", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "memory impact of the voxel-based representation, so we seek a more efficient way to do so. The idea is to incorporate a 2D super-resolution network (Sec. 3.3.1), trained end-to-end, that improves the output from the base model. The super-resolved images are eventually fused back in an improved 3D model, which also has the benefit of further increasing view consistency (Sec. 3.3.2).", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Integrating super-resolution", + "text_level": 1, + "bbox": [ + 76, + 220, + 328, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we augment the method of Sec. 3.2 with a lightweight refinement post-processor network that takes the 2D image $\\hat{I}$ generated by the base model and outputs a higher quality version $\\hat{I}_{\\mathrm{super}}$ of the same. This can be thought of as a form of super-resolution; however, due to the particular statistics of the input ('low-res') images $\\hat{I}$ that HoloDiffusion generates, it is necessary to train this superresolution network in an end-to-end fashion with HoloDiffusion, integrating the two models.", + "bbox": [ + 75, + 244, + 468, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To make this integration seamless, we formulate super-resolution as another diffusion process that runs 'in parallel' with 3D reconstruction. Hence, the super-resolved image $\\hat{I}_{\\mathrm{super}} = D_{\\beta}(I_t,\\hat{I},t)$ is the output of a denoiser network (a lightweight U-Net), which takes as input the noised target image $I_{t} = \\hat{\\epsilon}_{t}(I)$ and is also conditioned on the 'low-res' output $\\hat{I} = R_{\\eta}(V,C)$ of HoloDiffusion from Eq. (6). This denoiser is trained with the DDPM loss:", + "bbox": [ + 75, + 381, + 468, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\beta | I) = \\mathbb {E} _ {t, \\hat {\\epsilon}} \\left[ \\| D _ {\\beta} \\left(\\hat {\\epsilon} _ {t} (I), \\hat {I}, t\\right) - I \\| ^ {2} \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 510, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details The overall model $(D_{\\beta}$ and $D_{\\theta})$ is trained end-to-end by optimising the loss $\\mathcal{L}(\\theta |I,C,\\bar{s}) + \\mathcal{L}(\\beta |I)$ obtained by summing Eqs. (5) and (7).", + "bbox": [ + 76, + 549, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As training data, we use a large dataset of images capturing object-centric scenes ([30]). In each training batch, we pick a random training scene $s$ and sample 15 different source images $\\bar{s}_{\\mathrm{src}} \\subset s$ which are unprojected to generate the feature volume conditioning $\\bar{V}$ . Then, $\\bar{V}$ is rendered into 4 random target views $\\bar{s}_{\\mathrm{tgt}} \\subset (s \\setminus \\bar{s}_{\\mathrm{src}})$ which allows to optimize the training image reconstruction loss $\\mathcal{L}(\\theta | I, C, \\bar{s}) + \\mathcal{L}(\\beta | I)$ . The latter uses the Adam optimizer with an initial learning rate of $5 \\cdot 10^{-5}$ decaying tenfold whenever the loss plateaus until convergence.", + "bbox": [ + 75, + 595, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Fusing super-resolved views in 3D", + "text_level": 1, + "bbox": [ + 76, + 768, + 369, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The method of Sec. 3.3.1 leaves us with high-resolution views $\\hat{I}_{\\mathrm{super}}$ of the generated 3D object. However, we would like to obtain a single, high-quality 3D model, not just individual views of it. In this section, we discuss how to take the super-resolved images and fuse them into such a model, while addressing the issues that these images are not perfectly view-consistent.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ecabef4deb273ec11a28889a588e4e8b4c8c8e637b4fbc031a3307b76289271.jpg", + "image_caption": [ + "Figure 3: Distillation. HOOFUSION distills a single high-resolution voxel grid $V_0^H$ by minimizing a top-k patch-remix loss $\\mathcal{L}_{\\mathrm{distil}}$ between the grid renders $R_{\\eta'}(V_0^H, C)$ and a bank $\\mathcal{I}_C$ of $K = 5$ high-res images output by the 2D diffusion upsampler $D_{\\beta}$ for each scene camera $C$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The basic idea is simple. We can generate a certain number (e.g., 100) high-resolution images of the object from different viewpoints $C$ and then use a technique, akin to neural rendering, to fuse them back into a single 3D model. However, there is a problem with this idea: The model of Sec. 3.3.1 generates high-quality views $I_{\\mathrm{super}}$ , but these are view-dependent samples from the distribution $p(\\hat{I}_{\\mathrm{super}}|\\hat{I})$ where $\\hat{I} = R_{\\eta}(V,C)$ is the 'low-res' output form HoloDiffusion. Because super-resolving details is intrinsically ambiguous, there is no reason why samples $I_{\\mathrm{super}}$ taken from different viewpoints $C$ would be consistent (Fig. 7). Fusing them into a single 3D model would then result in a blurry appearance yet again.", + "bbox": [ + 496, + 377, + 892, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As described in Fig. 3, we address this issue in a principled manner by considering several possible super-resolved images $\\mathcal{I}_C = \\{I_{\\mathrm{super}}\\sim p(\\hat{I}_{\\mathrm{super}}|\\hat{I})\\}$ sampled from each given viewpoint $C$ . Then, we optimize a high-resolution voxel grid $V_0^H$ by minimizing the photometric loss:", + "bbox": [ + 496, + 578, + 890, + 655 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l}} \\left(\\eta^ {\\prime}, V _ {0} ^ {H} \\mid \\mathcal {I} _ {C}\\right) = \\mathbb {E} _ {C} \\left[ \\min _ {I _ {\\text {s u p e r}} \\in \\mathcal {I} _ {C}} \\| I _ {\\text {s u p e r}} - R _ {\\eta^ {\\prime}} \\left(V _ {0} ^ {H}, C\\right) \\| ^ {2} \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 667, + 893, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{\\eta'}(V_0^H, C)$ is the render of a high-resolution voxel grid $V_0^H \\in \\mathbb{R}^{d \\times S' \\times S' \\times S'}$ , $S' > S$ using the learnable renderer $R_{\\eta'}$ with scene specific parameters $\\eta'$ . Minimizing with respect to $I_{\\mathrm{super}}$ means that the 3D model must be consistent with at least one of the possible super-resolved images, drawn from the distribution of super-resolved samples, for each view $C$ .", + "bbox": [ + 496, + 713, + 890, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch remix. In practice, this approach requires a very large number of super resolved images $\\mathcal{I}_C$ to be effective. We found that we can significantly improve the statistical efficiency by performing the minimization at the level of individual patches. Namely, we produce a stack of only", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22980", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6caeb0cb9e20918eeba8abf09ce78107c7727018f0f599c8f0526df050d2022e.jpg", + "image_caption": [ + "Figure 4: Generated 3D samples visualized from a moving camera. $\\pi$ -GAN and HoloDiffusion* fail to produce 3D view consistent samples, while DreamFusion suffers from the \"Janus\" problem (multiple heads)." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 460, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$K = |\\mathcal{I}_C| = 5$ super resolved images and perform the minimization in Eq. (8) at the level of small $16 \\times 16$ patches independently (effectively allowing super-resolved images to 'remix' as needed to fit the generated view $R_{\\eta'}(V_0^H, C)$ ).", + "bbox": [ + 75, + 453, + 468, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Distillation details. $\\mathcal{L}_{\\mathrm{distil}}$ is optimized independently for each generated scene with Adam $(1\\mathrm{r} = 2\\cdot 10^{-4})$ for $25\\mathrm{K}$ steps until convergence. While $\\eta^\\prime$ is initialized using the pretrained multi-sequence weights $\\eta$ , $V_0^H$ is initialized by tri-linearly upsampling the low-resolution volume $V_{0}$ output by HoloDiffusion. Cameras $C$ are sampled at uniform azimuths with elevation fixed at object's equator.", + "bbox": [ + 75, + 527, + 468, + 635 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19d0793b669aa1114c512b9643516e5b124e74034a0bb4f066efa75cf2b01315.jpg", + "image_caption": [ + "Figure 5: Fusing views. Our patch-remix (Sec. 3.3.2) compared to the SDS and MSE distillation. MSE has \"floaters\" and viewpoint inconsistencies, SDS over-smooths the texture. Ours is robust and produces superior quality." + ], + "image_footnote": [], + "bbox": [ + 84, + 685, + 464, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 632, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin with a description of the experiments conducted in Sec. 4.1, followed by an analysis and discussion of the results in Sec. 4.2.", + "bbox": [ + 498, + 116, + 890, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Details", + "text_level": 1, + "bbox": [ + 500, + 174, + 589, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We experiment on the challenging large-scale Co3Dv2 [30] dataset which is a popular choice for a real-world 3D reconstruction benchmark. More specifically, 4 categories are selected, Apple, Hydrant, TeddyBear, and Donut, with 500 3D-scenes per category for training. Each 3D scene contains $\\sim 200$ images of the object of interest along with poses of their corresponding cameras.", + "bbox": [ + 496, + 199, + 890, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We use two sets of baselines for comparison (Tab. 1): (i) general 3D generative modeling baselines and (ii) diffusion distillation based baselines. $\\pi$ -GAN [5], EG3D [4], GET3D [9], and HoloDiffusion [18] are considered as the 3D generative baselines. Along with HoloDiffusion, we also test the super-resolution integrated model (described in Sec. 3.3.1) HoloDiffusion*. For the distillation-based baselines, we consider the open-source implementation of DreamFusion [29] titled Stable-DreamFusion [40]. For the latter, scenes are generated by conditioning on prompts comprising names of Co3Dv2 categories extended with color and style modifier phrases leading to $\\sim 200$ prompts / 3D shapes per class. More details regarding the prompt creation are in the supplementary.", + "bbox": [ + 496, + 309, + 892, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We use FID [15] and KID [3] to compare the quality of our 2D renders, as these are commonly used to assess 2D and 3D generators.", + "bbox": [ + 498, + 525, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative and qualitative analysis", + "text_level": 1, + "bbox": [ + 500, + 584, + 816, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 evaluates quantitatively while Fig. 6 qualitatively. Furthermore, Fig. 4 compares rendering view-consistency.", + "bbox": [ + 498, + 609, + 888, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HoloFusion (Ours) yields better FID/KID scores than the general 3D generative baselines except for $\\pi$ -GAN on Apple and Donut classes. However, since $\\pi$ -GAN does not guarantee view consistency by design, it essentially acts as a 2D image GAN, and thus does better on the 2D FID/KID metrics, but it generates significantly view-inconsistent renders (see Fig. 4 and the supplementary).", + "bbox": [ + 496, + 641, + 890, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observed that the other 3D-GAN baselines, EG3D and GET3D, are prone to collapsing to a single adversarial sample leading to poor FID/KID scores. The latter is probably due to the 3D misalignment of the CO3Dv2 sequences across instances, which makes training harder.", + "bbox": [ + 496, + 748, + 890, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HoloFusion also outperforms the text-to-3D StableDreamFusion on both FID/KID. Stable-DreamFusion yields good shapes, but produces synthetic-looking and overly-smooth textures and thus performs poorly when compared to the real-world images of Co3Dv2. As evi", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22981", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/382b412cae39adb21545ba1a92e4e4c3d05a0f609a7303a75bfd6f8c55058c48.jpg", + "image_caption": [ + "HoloFusion HoloDiffusion* HoloDiffusion DreamFusion GET-3D EG3D $\\pi$ -GAN", + "Figure 6: 3D samples generated by our HoloFusion compared to $\\pi$ -GAN, EG3D, GET3D, HoloDiffusion, HoloDiffusion*, and the text-to-3D Stable-DreamFusion." + ], + "image_footnote": [], + "bbox": [ + 106, + 108, + 890, + 829 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22982", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/01622b1493a2fcdab1ea6af1fba0ba6b2dc77e7da0dc32898a839812f3e5b3e0.jpg", + "table_caption": [ + "Table 1: FID (↓) and KID (↓) on 4 classes of Co3Dv2 [30]. We compare with 3D generative modeling baselines (rows 1–5); with an SDS distillation-based Stable-DreamFusion (row 6); and with ablations of our HoloFusion (rows 7–8). The column “VP” denotes whether renders of a method are 3D view-consistent or not." + ], + "table_footnote": [], + "table_body": "
methodVPAppleHydrantTeddyBearDonutMean
FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
π-GAN [5]X49.30.04292.10.080125.80.11899.40.06991.70.077
EG3D [4]✓170.50.203229.50.253236.10.239222.30.237214.60.233
GET3D [9]✓179.10.190303.30.380244.50.280209.90.230234.20.270
HoloDiffusion [18]✓94.50.095100.50.079109.20.106115.40.085122.50.102
HoloDiffusion*X55.90.04562.60.045116.60.10199.60.07983.70.068
Stable-DreamFusion [40]✓139.00.104185.20.132183.40.125169.30.114169.20.119
HoloFusion (MSE)X72.70.06762.20.04587.20.076109.00.09982.80.072
HoloFusion (SDS)✓123.00.10577.10.058117.80.090142.80.087115.20.085
HoloFusion (Ours)✓69.20.06366.80.04787.60.075109.70.09883.30.071
", + "bbox": [ + 76, + 138, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17b4ed1e7f688b0fa3b2eb10816ca8efd1afcc85b08419cb2e334b250096c7b6.jpg", + "image_caption": [ + "Figure 7: Heatmaps illustrating the per-pixel color variance of $K = 10$ hypothesis produced by the upsampler $D_{\\beta}$ . Some samples contain artifacts around the object boundaries which correspond to the high-variance regions in the figure. Our top-K patch-remix increases robustness by allowing the loss to discard such artifacts during distillation." + ], + "image_footnote": [], + "bbox": [ + 86, + 376, + 460, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dent from the TeddyBear samples, the method also suffers from the \"Janus\" issue.", + "bbox": [ + 75, + 700, + 468, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to HoloDiffusion, we improve the FID/KID scores by a significant margin, mainly due to the more photo-realistic renders that include high-frequency details. Although the 2D Diffusion upsampler of HoloDiffusion* produces renders with the highest amount of details yielding scores similar to ours, they are not 3D view-consistent (as apparent from Fig. 4 and as explained in Sec. 3.3).", + "bbox": [ + 75, + 729, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations. In Tab. 1 and in Fig. 5 we ablate components of our HoloFusion to verify their contribution.", + "bbox": [ + 76, + 839, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The first variant, HoloFusion (SDS), replaces the Topk patch-remixed distillation loss with the score distillation", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sampling (SDS) gradient as proposed in [29]. As apparent from Fig. 5 and from the lower scores, SDS washes out all the high-frequency details in the textures.", + "bbox": [ + 498, + 376, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Secondly, HoloFusion (MSE) reduces the number of upsampling hypotheses $\\mathcal{I}$ to the minimum of $|\\mathcal{I}| = 1$ . Even though this slightly improves the 2D metrics, as can be seen from Fig. 5, the samples lack view-consistency and introduce \"floaters\". In Fig. 7 we further illustrate the variability of the upsampling hypotheses.", + "bbox": [ + 498, + 422, + 890, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 529, + 619, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a hybrid diffusion-based method that can generate high-quality 3D neural radiance fields of real-life object categories. Our method starts by producing coarse 3D models whose renders are independently superresolved, and finally consolidated using a robust distillation process. We evaluated our method on the Co3D v2 dataset and presented 3D-consistent, diverse, and high-quality results superior to all competing baselines.", + "bbox": [ + 496, + 555, + 890, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method suffers from limitations that can be addressed in future work. First, our method is slow to sample from as the sampling process takes about 30 mins for each generation, because it is still a distillation-based method. An interesting extension would be to train another network to directly distill a set of super-resolved images, without requiring explicit optimization during inference. Second, we do not produce an explicit surface representation (e.g., a mesh), which could be done by integrating a differentiable mesh render in the loop as done in some prior work.", + "bbox": [ + 496, + 676, + 890, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 844, + 687, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Animesh and Niloy were partially funded by the European Union's Horizon 2020 research and innovation pro", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22983", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gramme under the Marie Skłodowska-Curie grant agreement No. 956585. This research has been partly supported by MetaAI and the UCL AI Centre. Finally Animesh is grateful to The Rabin Ezra Scholarship Fund being a recipient of their esteemed fellowship for the year 2023.", + "bbox": [ + 76, + 90, + 472, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 181, + 174, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv.cs, abs/2211.09869, 2022. 1", + "[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. arXiv.cs, abs/2207.13751, 2022. 2", + "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6", + "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 6, 8", + "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 6, 8", + "[6] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv preprint arXiv:2212.03267, 2022. 3", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2", + "[8] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3D shape induction from 2D views of multiple objects. In arXiv, 2016. 2", + "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 6, 8", + "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. CoRR, abs/2110.08985, 2021. 2" + ], + "bbox": [ + 78, + 205, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerfguided distillation from 3d-aware diffusion. arXiv preprint arXiv:2302.10109, 2023. 3", + "[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 2", + "[14] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3", + "[17] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 4", + "[18] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6, 8", + "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3d content creation. arXiv.cs, abs/2211.10440, 2022. 1, 2", + "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3", + "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2", + "[22] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3", + "[23] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2", + "[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3", + "[25] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: $360^{\\circ}$ reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 3", + "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF:" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22984", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 2, 3", + "[27] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3d radiance field diffusion. In Proc. CVPR, 2022. 2, 3", + "[28] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. arXiv.cs, abs/1904.01326, 2019. 2", + "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 8", + "[30] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2, 5, 6, 8", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2", + "[33] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3D-aware image synthesis. arXiv.cs, abs/2007.02442, 2020. 2", + "[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. ARXIV, 2022. 2", + "[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Proc. NeurIPS, 2021. 2", + "[36] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 3", + "[37] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d gans. CoRR, abs/2206.10535, 2022. 2", + "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2", + "[39] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 156-174. Springer, 2022. 3", + "[40] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.2,6,8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2, 3", + "[42] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv.cs, abs/2210.04628, 2022. 1, 3", + "[43] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2", + "[44] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 2", + "[45] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2", + "[46] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. arXiv preprint arXiv:2212.00792, 2022. 3" + ], + "bbox": [ + 501, + 92, + 892, + 444 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22985", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_model.json b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb8a3a2e4916499753f815e409445f533bd49893 --- /dev/null +++ b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.815, + 0.154 + ], + "angle": 0, + "content": "HOLOFUSION: Towards Photo-realistic 3D Generative Modeling" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.182, + 0.28, + 0.215 + ], + "angle": 0, + "content": "Animesh Karnewar UCL" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.22, + 0.292, + 0.233 + ], + "angle": 0, + "content": "a.karnewar@ucl.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.182, + 0.464, + 0.215 + ], + "angle": 0, + "content": "Niloy J. Mitra \nUCL" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.22, + 0.483, + 0.233 + ], + "angle": 0, + "content": "n.mitra@ucl.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.182, + 0.658, + 0.215 + ], + "angle": 0, + "content": "Andrea Vedaldi Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.22, + 0.667, + 0.233 + ], + "angle": 0, + "content": "vedaldi@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.182, + 0.843, + 0.215 + ], + "angle": 0, + "content": "David Novotny Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.707, + 0.22, + 0.859, + 0.234 + ], + "angle": 0, + "content": "dnovotny@meta.com" + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.248, + 0.312, + 0.261 + ], + "angle": 0, + "content": "HoloDiffusion samples" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.264, + 0.405, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.248, + 0.527, + 0.259 + ], + "angle": 0, + "content": "Distillation" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.267, + 0.558, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.248, + 0.837, + 0.261 + ], + "angle": 0, + "content": "HoloFusion high quality samples" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.265, + 0.901, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.536, + 0.895, + 0.582 + ], + "angle": 0, + "content": "Figure 1: We propose HoloFUSION to generate photo-realistic 3D radiance fields by extending the HoloDiffusion method with a jointly trained 2D 'super resolution' network. The independently super-resolved images are fused back into the 3D representation to improve the fidelity of the 3D model via distillation, while preserving the consistency across view changes." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.595, + 0.313, + 0.61 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.474, + 0.902 + ], + "angle": 0, + "content": "Diffusion-based image generators can now produce high-quality and diverse samples, but their success has yet to fully translate to 3D generation: existing diffusion methods can either generate low-resolution but 3D consistent outputs, or detailed 2D views of 3D objects but with potential structural defects and lacking view consistency or realism. We present HoloFUSION, a method that combines the best of these approaches to produce high-fidelity, plausible, and diverse 3D samples while learning from a collection of multi-view 2D images only. The method first generates coarse 3D samples using a variant of the recently proposed HoloDiffusion generator. Then, it independently renders and upsamples a large number of views of the coarse 3D model, super-resolves them to add detail, and distills those into a single, high-fidelity implicit 3D representation, which also ensures view-consistency of the final renders. The super-resolution network is trained as an integral part of HoloFUSION, end-to-end, and the final distillation uses a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.596, + 0.895, + 0.673 + ], + "angle": 0, + "content": "new sampling scheme to capture the space of super-resolved signals. We compare our method against existing baselines, including DreamFusion, Get3D, EG3D, and HoloDiffusion, and achieve, to the best of our knowledge, the most realistic results on the challenging CO3Dv2 dataset." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.689, + 0.633, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.893, + 0.793 + ], + "angle": 0, + "content": "Diffusion models [32, 7, 31] are at the basis of state-of-the-art 2D image generators which can now produce very high-quality and diverse outputs. However, their success has yet to be translated to 3D and there is no generator that can produce 3D assets of a comparable quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Recent attempts at extending diffusion to 3D generation have reported mixed success. Some authors have attempted to apply diffusion directly in 3D [18], or still in 2D but using a 3D-aware neural network [42, 1]. This requires solving two problems: first, finding a suitable 3D representation (e.g., triplane features [4], mesh [19], voxels [18]) that scales well with resolution and is amenable to diffusion;" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "and, second, obtaining a large amount of 3D training data, for example using synthetic models [41, 27], or training the model using only 2D images [18], often via differentiable (volume) rendering [13, 26]. However, the quality of results so far is limited, especially when training on real images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.471, + 0.395 + ], + "angle": 0, + "content": "Other authors have proposed to distill 3D objects from pre-trained 2D image generators. For instance Score Distillation Sampling (SDS) [29] can sample 3D objects from a high-quality off-the-shelf 2D diffusion model while requiring no (re)training. However, without any 3D guidance, distillation methods often produce implausible results; for example, they suffer from the 'Janus effect', where details of the front of the object are replicated on its back. They also create overly-smooth outputs that average out inconsistencies arising from the fact that the signal obtained from the 2D model is analogous to sampling independent views of the object (see Sec. 4.2 for examples). Furthermore, distillation methods do not support unconditional sampling, even if the underlying image generator does, as strong language guidance is required to stabilise the 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.548 + ], + "angle": 0, + "content": "In this work, we propose HLOFUSION, a method that combines the best of both approaches. We start from HoloDiffusion [18], a diffusion-based 3D generator. This model can be trained using only a multiview image dataset like [30] and produces outputs that are 3D consistent. However, the output resolution is limited by computation and memory. We augment the base model with a lightweight super-resolution network that upscales the initial renders. Crucially, the 2D super-resolution model is integrated and trained jointly with the 3D generator, end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.732 + ], + "angle": 0, + "content": "The super-resolution network outputs detailed views of the 3D object, and the underlying 3D generator ensures that the coarse structure of these views is indeed consistent (e.g., avoiding the Janus effect and other structural artifacts). However, the 2D upscaling still progresses independently for different views, which means that fine grained details may still be inconsistent between views. We address this issue by distilling a single, coherent, high quality 3D model of the object from the output of the upsampler. For this, we propose a new distillation technique that efficiently combines several putative super-resolved views of the object into a single, coherent 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.809 + ], + "angle": 0, + "content": "With this, we are able to train a high-quality 3D generator model purely from real 2D data. This model is capable of generating consistent and detailed 3D objects, which in turn result in view-consistent renderings (see Fig. 1) at a quality not achievable by prior methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We evaluate HoloFusion on real images (CO3Dv2 dataset [30]) and compare with a variety of competing alternatives (e.g., HoloDiffusion [18], Get3D [9], EG3D [4], DreamFusion [40]) demonstrating that view-consistent high-quality 3D generation is possible using our simple, effective, easy-to-implement hybrid approach." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.117, + 0.892, + 0.313 + ], + "angle": 0, + "content": "3D generators that use adversarial learning. Generative Adversarial Learning (GAN) [10] learns a generator network so that its \"fake\" samples cannot be distinguished from real images by a second discriminator network. Approaches such as PlatonicGAN [13], HoloGAN [28], and PrGAN [8] introduced 3D structure into the generator network, achieving 3D shape generation with only image-level supervision. Our method is related to those as it renders images from a generated voxel grid, as well as to HoloGAN [28], which renders features and then converts them into an image by a lightweight 2D convolutional network. Other voxel-based 3D generators include VoxGRAF [34] and NeuralVolumes [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.51 + ], + "angle": 0, + "content": "More recently, 3D generators have built on neural radiance fields [26]. GRAF [33] was the first to adopt the NeRF framework; analogous to PlatonicGAN, they generate the parameters of an MLP which renders realistic images of the object from a random viewpoint. This idea has been improved in StyleNeRF [11] and EG3D [4] by adding a 2D convolutional post-processing step after emission-absorption rendering, which is analogous to our super-resolution network. EG3D also introduced a novel 'Triplane' representation of the radiance field which, in a memory efficient manner, factorises the latter into a triplet of 2D feature planes. EG3D inspired several improvements such as GAUDI [2] and EpiGRAF [37]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Mesh-based 3D generators have been explored in [43]. Recently, GET3D [9] replaced the radiance field with a signed distance function to regularise the representation of geometry. The latter is converted into a mesh and rendered in a differentiable manner by using the marching tetrahedral representation [35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Modeling 3D with diffusion. Diffusion methods [38] have recently become the go-to framework for generative modeling of any kind, including 3D generative modeling. The first applications of diffusion to 3D considered pointcloud generators trained on synthetic data [23, 45, 44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.866 + ], + "angle": 0, + "content": "3D distillation of 2D diffusion models. More recently, DreamFusion [29] ported the idea of distillation to diffusion models: they extract a neural radiance field so that its renders match the belief of a pre-trained 2D diffusion generator [32, 7, 31]. They introduce the Score Distillation Sampling (SDS) loss which makes distillation relatively efficient (but still in the order of several minutes for a single 3D sample). Their generation can be conditioned by an image or by a textual description, making the process rather flexible. Magic3D [19] further increases the quality of the output by distilling a mesh-based 3D representation instead of a radiance field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image-conditioned 3D diffusion. The idea of distillation has been applied to few-view conditioned reconstruction in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22977" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.212 + ], + "angle": 0, + "content": "[42, 12, 25, 46, 6]. SparseFusion [46] employs a 3D-based new-view synthesis model [39] followed by a 2D diffusion upsampler. They complete the process by 3D distillation, ensuring that the generated views of the object are consistent. NeRFDiff [12] and 3DiM [42] bypass an explicit 3D model and directly generate new views of an object using a 2D image generator and, in the case of NeRFDiff, refine the results using distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.332 + ], + "angle": 0, + "content": "While SparseFusion and NeRFDiff need to be trained on a dataset of object-centric multi-view images with pose information, RealFusion [25] and NeRDi [6] can be used for zero-shot monocular 3D reconstruction, starting from a pretrained 2D diffusion model. Given a single image as input, they automatically generate a prompt for the diffusion model, using a form of prompt inversion, and then use distillation to extract a radiance field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.534 + ], + "angle": 0, + "content": "Unconditional generation. Most relevantly to us, unconditional generation, i.e., generation which does not require either text or image conditioning, was explored in [41, 36, 27, 18]. While [27, 36, 41] train generators given synthetic 3D ground truth, similar to us, HoloDiffusion [18] is supervised only with real object-centric images and camera poses. While HoloDiffusion was the first to demonstrate successful training on real image data, its renders contain considerably lower amount of detail than samples from a conventional 2D image generator that uses diffusion. We thus leverage a 2D diffusion upsampler, conditioned on the lower-fidelity HoloDiffusion renders, to distill higher resolution images and, eventually, 3D models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.547, + 0.224, + 0.563 + ], + "angle": 0, + "content": "3. HOOFUSION" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.723 + ], + "angle": 0, + "content": "We present HOOFUSION, a method that can learn a high-quality diffusion-based 3D generator from a collection of multiview 2D images. HOOFUSION first obtains an unconditional low-resolution 3D sample using diffusion and then distills a high-resolution 3D radiance field representing a higher-quality version of the generated object. We first summarize the Denoising Diffusion Probabilistic Models (DDPMs) [16] that we utilize in Sec. 3.1. Then, we discuss the low-resolution 3D generator in Sec. 3.2 followed by super-resolution distillation in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.292, + 0.747 + ], + "angle": 0, + "content": "3.1. Preliminaries: DDPMs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.469, + 0.816 + ], + "angle": 0, + "content": "Let \\( x = x_0 \\) be a random vector whose probability distribution \\( p(x|y) \\) we seek to model. The DDPM [16] defines a hierarchy of latent variables \\( x_{t} \\), \\( t = 0,\\dots ,T \\) and an encoder \\( q \\) comprising a sequence of Gaussian distributions" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.828, + 0.47, + 0.845 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} x _ {t - 1}, (1 - \\alpha_ {t}) \\mathbb {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\alpha_{t},\\ldots ,\\alpha_{T}\\) is a predefined 'noising schedule'. Given knowledge of \\(x_0\\) , a sample \\(x_{t}\\) can be drawn in a closed-form directly from \\(q(x_{t}|x_{0}) = \\mathcal{N}(x_{t};\\sqrt{\\bar{\\alpha}_{t}} x_{0},(1 - \\bar{\\alpha}_{t})\\mathbb{I})\\) , where" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.09, + 0.797, + 0.109 + ], + "angle": 0, + "content": "\\(\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}\\) . Hence, we can express \\(x_{t}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.115, + 0.892, + 0.133 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\hat {\\epsilon} _ {t} \\left(x _ {0}\\right) = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.141, + 0.892, + 0.248 + ], + "angle": 0, + "content": "The noising schedule is chosen such that \\(\\bar{\\alpha}_{T} \\approx 0\\). In this manner, \\(q(x_{T}|x_{0}) \\approx \\mathcal{N}(x_{T};0,\\mathbb{I})\\) is approximately normal, and so is \\(q(x_{T})\\). To generate a sample \\(x = x_0\\), we start by sampling \\(x_{T}\\) from this normal distribution and then sample the intermediate latent variables backward. This is done by using a variational approximation of the probabilities \\(q(x_{t - 1}|x_t)\\) given by the Gaussian factors:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.256, + 0.892, + 0.273 + ], + "angle": 0, + "content": "\\[\np \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\sqrt {\\bar {\\alpha} _ {t - 1}} D _ {\\theta} \\left(x _ {t}, t\\right), (1 - \\bar {\\alpha} _ {t - 1}) \\mathbb {I}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.281, + 0.824, + 0.295 + ], + "angle": 0, + "content": "where \\(D_{\\theta}\\) is a neural network with parameters \\(\\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.341 + ], + "angle": 0, + "content": "The network \\(D_{\\theta}\\) is trained by maximizing the ELBO (Evidence Lower Bound), which reduces to the denoising objective [22]:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.349, + 0.892, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, \\epsilon , x} \\left[ \\frac {\\bar {\\alpha} _ {t - 1}}{2 (1 - \\bar {\\alpha} _ {t - 1}) ^ {2}} \\| D _ {\\theta} \\left(\\hat {\\epsilon} _ {t} (x), t\\right) - x \\| ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.389, + 0.892, + 0.436 + ], + "angle": 0, + "content": "where \\( t \\) is sampled uniformly from \\( U[1,T] \\). Hence, \\( D_{\\theta}(x_t,t) \\) approximates the clean sample \\( x_0 \\) given the noisy sample \\( x_{t} \\) (obtained using (2))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.715, + 0.458 + ], + "angle": 0, + "content": "3.2. HoloDiffusion revisited" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Given a large dataset of 3D models, the framework of Sec. 3.1 could be used to train a corresponding probability distribution. However, such a dataset is not available, and we must instead learn from 2D images of physical 3D objects. Given a dataset containing several views of a large number of objects, we could use image-based reconstruction (e.g., using neural rendering) to obtain corresponding 3D models first, and then use those to train a diffusion model. Instead, we adopt, and slightly upgrade, the HoloDiffusion method [18], which learns a 3D diffusion model directly from the 2D images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Training data. HoloDiffusion learns from a collection \\(\\mathcal{D}\\) of \\(N\\) image sequences \\(s_i = (I_j^i, C_j^i)_{j=1}^{N_{\\mathrm{frame}}}\\), \\(i = 1, \\ldots, N\\), where frame \\(I_j^i \\in \\mathbb{R}^{3 \\times H \\times W}\\) is an RGB image and \\(C_j^i \\in \\mathbb{R}^{4 \\times 4}\\) is the corresponding camera projection matrix, collectively defining the motion of the camera." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.902 + ], + "angle": 0, + "content": "3D representation and rendering. The shape and appearance of the object are represented by a voxel grid \\(V \\in \\mathbb{R}^{d \\times S \\times S \\times S}\\) with resolution \\(S\\) containing a \\(d\\)-dimensional feature vector per voxel. Given a 3D point \\(\\mathbf{p} \\in \\mathbb{R}^3\\), its opacity \\(\\sigma(\\mathbf{p}) \\in \\mathbb{R}_+\\) and color \\(c(\\mathbf{p}) \\in \\mathbb{R}_{[0,1]}^3\\) are obtained from the voxel grid by an MLP \\(M_{\\eta}(V(\\mathbf{p}))\\) that takes as input the \\(d\\)-dimensional feature vector \\(V(\\mathbf{p})\\) extracted from the grid via trilinear interpolation [20]. The usual emission-absorption model [26, 24] is then used to implement a differentiable rendering function \\(R_{\\eta}\\), mapping the voxel grid \\(V\\) and the camera viewpoint \\(C\\) into an image \\(\\hat{I} = R_{\\eta}(V,C)\\), where \\(\\eta\\) are the parameters of the MLP." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22978" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.895, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.89, + 0.421 + ], + "angle": 0, + "content": "Figure 2: Overview. HOOFUSION, which trains the 3D denoiser network \\( D_{\\theta} \\), is augmented with the 2D 'super-resolution' diffusion model \\( D_{\\beta} \\). Both models are trained end-to-end by supervising their outputs with 2D photometric error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Training scheme. HoloDiffusion leverages the DDPM framework (revised in the previous paragraphs) to recover the density \\( p(V) \\) over voxel grids \\( x = V \\) encoding plausible real-life objects. In order to train a DDPM on such 3D data, we would need access to ground-truth 3D models \\( V \\), which are not available. HoloDiffusion addresses this problem by making three changes to DDPM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.644 + ], + "angle": 0, + "content": "First, it replaces the data denoising loss with a photometric reconstruction loss. Given a pair \\((I,C)\\in s\\) from one of the training sequences \\(s\\), it replaces Eq. (4) with \\(\\mathbb{E}_{t,\\epsilon ,C}\\left[\\| I - R_{\\eta}(D_{\\theta}(\\hat{\\epsilon}_t(V),t),C)\\| ^2\\right]\\) where the goal is not to reconstruct the 'clean' volume \\(V\\) (which is unknown), but rather its image \\(I\\) (which is known)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Second, also because the 'clean' volume \\( V \\) is not available, we cannot use Eq. (3) to generate the noisy volumes \\( V_{t} \\) to denoise; the only exception is the last sample \\( V_{T} \\), which is pure noise. This suggests to adopt a 'double denoising' step. First, pure noise \\( V_{T} \\) is fed into the denoiser to obtain an (approximate) version of \\( V_{0} = D_{\\theta}(V_{T}, T) \\) of the clean volume \\( V_{0} = V \\). Then, noise is applied to obtain \\( V_{t} = \\hat{\\epsilon}_{t}(V_{0}) = \\sqrt{\\bar{\\alpha}_{T}} V_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{t} \\) according to Eq. (3), and the latter is fed back into the denoiser as above." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Finally, there is the issue that unconditional generation of the clean volume \\( V_{0} \\) from pure noise \\( V_{T} \\) is difficult, especially in the first iterations of training. On the other hand, the problem of view-conditioned generation is considerably easier. Hence, the third idea is to learn a conditional generator, using a variable number of input views. Specifically, given a training sequence \\( s \\), the method extracts a random subset of frames \\( \\bar{s} \\subset s \\) (which could be empty, which cor-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.568 + ], + "angle": 0, + "content": "responds to unconditional generation). Then, a feature volume \\(\\bar{V} = \\Phi (\\bar{s})\\in \\mathbb{R}^{d\\times S\\times S\\times S}\\) is obtained from the selected frames. This extracts 2D image features using a pre-trained and frozen 2D image encoder and then pools them in 3D via 'unprojection' [17, 14] into \\(\\bar{V}\\), where \\(\\bar{V} = 0\\) if \\(\\bar{s}\\) is empty. Finally, these pooled features are used to condition the denoising network \\(V_{0} = D_{\\theta}(V_{T},\\bar{V},T)\\), which, on average, leads to a simpler reconstruction problem." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.568, + 0.84, + 0.584 + ], + "angle": 0, + "content": "Putting it all together, the training loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.592, + 0.891, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta | I, C, \\bar {s}) = \\mathbb {E} _ {t, \\hat {\\epsilon}, V _ {T}} \\left[ \\| \\hat {I} - I \\| ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.62, + 0.891, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\hat {I} = R _ {\\eta} \\left(D _ {\\theta} \\left(V _ {t}, \\bar {V}, t\\right), C\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.64, + 0.725, + 0.656 + ], + "angle": 0, + "content": "\\[\nV _ {t} = \\hat {\\epsilon} _ {t} (V _ {0}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.658, + 0.776, + 0.675 + ], + "angle": 0, + "content": "\\[\nV _ {0} = D _ {\\theta} (V _ {T}, \\bar {V}, T),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.678, + 0.716, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\bar {V} = \\Phi (\\bar {s}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.809 + ], + "angle": 0, + "content": "This loss is averaged over training sequences \\( s \\), subsequences \\( \\bar{s} \\subset s \\), and views \\( (I, C) \\in s \\) therein. Note that this is slightly different than the original HoloDiffusion, where feature volume \\( \\tilde{V} \\) and reconstructed volumes \\( V_{t} \\) overlap as arguments of the denoiser; we found that keeping them separated in the formulation leads to more stable training and additionally allows for view-conditioned generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.634, + 0.832 + ], + "angle": 0, + "content": "3.3. HOOFUSION" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The method of Sec. 3.2 learns to generate 3D objects from 2D image supervision only, but the fidelity of the output is limited by the resolution at which the operations are carried out. Increasing resolution is difficult due to the GPU" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22979" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "memory impact of the voxel-based representation, so we seek a more efficient way to do so. The idea is to incorporate a 2D super-resolution network (Sec. 3.3.1), trained end-to-end, that improves the output from the base model. The super-resolved images are eventually fused back in an improved 3D model, which also has the benefit of further increasing view consistency (Sec. 3.3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.221, + 0.329, + 0.236 + ], + "angle": 0, + "content": "3.3.1 Integrating super-resolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.47, + 0.381 + ], + "angle": 0, + "content": "As shown in Fig. 2, we augment the method of Sec. 3.2 with a lightweight refinement post-processor network that takes the 2D image \\(\\hat{I}\\) generated by the base model and outputs a higher quality version \\(\\hat{I}_{\\mathrm{super}}\\) of the same. This can be thought of as a form of super-resolution; however, due to the particular statistics of the input ('low-res') images \\(\\hat{I}\\) that HoloDiffusion generates, it is necessary to train this superresolution network in an end-to-end fashion with HoloDiffusion, integrating the two models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.47, + 0.502 + ], + "angle": 0, + "content": "To make this integration seamless, we formulate super-resolution as another diffusion process that runs 'in parallel' with 3D reconstruction. Hence, the super-resolved image \\(\\hat{I}_{\\mathrm{super}} = D_{\\beta}(I_t,\\hat{I},t)\\) is the output of a denoiser network (a lightweight U-Net), which takes as input the noised target image \\(I_{t} = \\hat{\\epsilon}_{t}(I)\\) and is also conditioned on the 'low-res' output \\(\\hat{I} = R_{\\eta}(V,C)\\) of HoloDiffusion from Eq. (6). This denoiser is trained with the DDPM loss:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.511, + 0.47, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\beta | I) = \\mathbb {E} _ {t, \\hat {\\epsilon}} \\left[ \\| D _ {\\beta} \\left(\\hat {\\epsilon} _ {t} (I), \\hat {I}, t\\right) - I \\| ^ {2} \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.55, + 0.469, + 0.596 + ], + "angle": 0, + "content": "Training details The overall model \\((D_{\\beta}\\) and \\(D_{\\theta})\\) is trained end-to-end by optimising the loss \\(\\mathcal{L}(\\theta |I,C,\\bar{s}) + \\mathcal{L}(\\beta |I)\\) obtained by summing Eqs. (5) and (7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.749 + ], + "angle": 0, + "content": "As training data, we use a large dataset of images capturing object-centric scenes ([30]). In each training batch, we pick a random training scene \\(s\\) and sample 15 different source images \\(\\bar{s}_{\\mathrm{src}} \\subset s\\) which are unprojected to generate the feature volume conditioning \\(\\bar{V}\\). Then, \\(\\bar{V}\\) is rendered into 4 random target views \\(\\bar{s}_{\\mathrm{tgt}} \\subset (s \\setminus \\bar{s}_{\\mathrm{src}})\\) which allows to optimize the training image reconstruction loss \\(\\mathcal{L}(\\theta | I, C, \\bar{s}) + \\mathcal{L}(\\beta | I)\\). The latter uses the Adam optimizer with an initial learning rate of \\(5 \\cdot 10^{-5}\\) decaying tenfold whenever the loss plateaus until convergence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.37, + 0.785 + ], + "angle": 0, + "content": "3.3.2 Fusing super-resolved views in 3D" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The method of Sec. 3.3.1 leaves us with high-resolution views \\(\\hat{I}_{\\mathrm{super}}\\) of the generated 3D object. However, we would like to obtain a single, high-quality 3D model, not just individual views of it. In this section, we discuss how to take the super-resolved images and fuse them into such a model, while addressing the issues that these images are not perfectly view-consistent." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.891, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.273, + 0.893, + 0.35 + ], + "angle": 0, + "content": "Figure 3: Distillation. HOOFUSION distills a single high-resolution voxel grid \\(V_0^H\\) by minimizing a top-k patch-remix loss \\(\\mathcal{L}_{\\mathrm{distil}}\\) between the grid renders \\(R_{\\eta'}(V_0^H, C)\\) and a bank \\(\\mathcal{I}_C\\) of \\(K = 5\\) high-res images output by the 2D diffusion upsampler \\(D_{\\beta}\\) for each scene camera \\(C\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.893, + 0.578 + ], + "angle": 0, + "content": "The basic idea is simple. We can generate a certain number (e.g., 100) high-resolution images of the object from different viewpoints \\( C \\) and then use a technique, akin to neural rendering, to fuse them back into a single 3D model. However, there is a problem with this idea: The model of Sec. 3.3.1 generates high-quality views \\( I_{\\mathrm{super}} \\), but these are view-dependent samples from the distribution \\( p(\\hat{I}_{\\mathrm{super}}|\\hat{I}) \\) where \\( \\hat{I} = R_{\\eta}(V,C) \\) is the 'low-res' output form HoloDiffusion. Because super-resolving details is intrinsically ambiguous, there is no reason why samples \\( I_{\\mathrm{super}} \\) taken from different viewpoints \\( C \\) would be consistent (Fig. 7). Fusing them into a single 3D model would then result in a blurry appearance yet again." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.656 + ], + "angle": 0, + "content": "As described in Fig. 3, we address this issue in a principled manner by considering several possible super-resolved images \\(\\mathcal{I}_C = \\{I_{\\mathrm{super}}\\sim p(\\hat{I}_{\\mathrm{super}}|\\hat{I})\\}\\) sampled from each given viewpoint \\(C\\). Then, we optimize a high-resolution voxel grid \\(V_0^H\\) by minimizing the photometric loss:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.668, + 0.895, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l}} \\left(\\eta^ {\\prime}, V _ {0} ^ {H} \\mid \\mathcal {I} _ {C}\\right) = \\mathbb {E} _ {C} \\left[ \\min _ {I _ {\\text {s u p e r}} \\in \\mathcal {I} _ {C}} \\| I _ {\\text {s u p e r}} - R _ {\\eta^ {\\prime}} \\left(V _ {0} ^ {H}, C\\right) \\| ^ {2} \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.821 + ], + "angle": 0, + "content": "where \\(R_{\\eta'}(V_0^H, C)\\) is the render of a high-resolution voxel grid \\(V_0^H \\in \\mathbb{R}^{d \\times S' \\times S' \\times S'}\\), \\(S' > S\\) using the learnable renderer \\(R_{\\eta'}\\) with scene specific parameters \\(\\eta'\\). Minimizing with respect to \\(I_{\\mathrm{super}}\\) means that the 3D model must be consistent with at least one of the possible super-resolved images, drawn from the distribution of super-resolved samples, for each view \\(C\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Patch remix. In practice, this approach requires a very large number of super resolved images \\(\\mathcal{I}_C\\) to be effective. We found that we can significantly improve the statistical efficiency by performing the minimization at the level of individual patches. Namely, we produce a stack of only" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22980" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.462, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.47, + 0.404 + ], + "angle": 0, + "content": "Figure 4: Generated 3D samples visualized from a moving camera. \\(\\pi\\)-GAN and HoloDiffusion* fail to produce 3D view consistent samples, while DreamFusion suffers from the \"Janus\" problem (multiple heads)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.516 + ], + "angle": 0, + "content": "\\(K = |\\mathcal{I}_C| = 5\\) super resolved images and perform the minimization in Eq. (8) at the level of small \\(16 \\times 16\\) patches independently (effectively allowing super-resolved images to 'remix' as needed to fit the generated view \\(R_{\\eta'}(V_0^H, C)\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.636 + ], + "angle": 0, + "content": "Distillation details. \\(\\mathcal{L}_{\\mathrm{distil}}\\) is optimized independently for each generated scene with Adam \\((1\\mathrm{r} = 2\\cdot 10^{-4})\\) for \\(25\\mathrm{K}\\) steps until convergence. While \\(\\eta^\\prime\\) is initialized using the pretrained multi-sequence weights \\(\\eta\\), \\(V_0^H\\) is initialized by tri-linearly upsampling the low-resolution volume \\(V_{0}\\) output by HoloDiffusion. Cameras \\(C\\) are sampled at uniform azimuths with elevation fixed at object's equator." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.686, + 0.465, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.836, + 0.468, + 0.896 + ], + "angle": 0, + "content": "Figure 5: Fusing views. Our patch-remix (Sec. 3.3.2) compared to the SDS and MSE distillation. MSE has \"floaters\" and viewpoint inconsistencies, SDS over-smooths the texture. Ours is robust and produces superior quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.633, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.117, + 0.892, + 0.161 + ], + "angle": 0, + "content": "We begin with a description of the experiments conducted in Sec. 4.1, followed by an analysis and discussion of the results in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.175, + 0.59, + 0.19 + ], + "angle": 0, + "content": "4.1. Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.306 + ], + "angle": 0, + "content": "Dataset. We experiment on the challenging large-scale Co3Dv2 [30] dataset which is a popular choice for a real-world 3D reconstruction benchmark. More specifically, 4 categories are selected, Apple, Hydrant, TeddyBear, and Donut, with 500 3D-scenes per category for training. Each 3D scene contains \\(\\sim 200\\) images of the object of interest along with poses of their corresponding cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.893, + 0.522 + ], + "angle": 0, + "content": "Baselines. We use two sets of baselines for comparison (Tab. 1): (i) general 3D generative modeling baselines and (ii) diffusion distillation based baselines. \\(\\pi\\)-GAN [5], EG3D [4], GET3D [9], and HoloDiffusion [18] are considered as the 3D generative baselines. Along with HoloDiffusion, we also test the super-resolution integrated model (described in Sec. 3.3.1) HoloDiffusion*. For the distillation-based baselines, we consider the open-source implementation of DreamFusion [29] titled Stable-DreamFusion [40]. For the latter, scenes are generated by conditioning on prompts comprising names of Co3Dv2 categories extended with color and style modifier phrases leading to \\(\\sim 200\\) prompts / 3D shapes per class. More details regarding the prompt creation are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.572 + ], + "angle": 0, + "content": "Metrics. We use FID [15] and KID [3] to compare the quality of our 2D renders, as these are commonly used to assess 2D and 3D generators." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.585, + 0.818, + 0.602 + ], + "angle": 0, + "content": "4.2. Quantitative and qualitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.89, + 0.64 + ], + "angle": 0, + "content": "Tab. 1 evaluates quantitatively while Fig. 6 qualitatively. Furthermore, Fig. 4 compares rendering view-consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.747 + ], + "angle": 0, + "content": "HoloFusion (Ours) yields better FID/KID scores than the general 3D generative baselines except for \\(\\pi\\)-GAN on Apple and Donut classes. However, since \\(\\pi\\)-GAN does not guarantee view consistency by design, it essentially acts as a 2D image GAN, and thus does better on the 2D FID/KID metrics, but it generates significantly view-inconsistent renders (see Fig. 4 and the supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.824 + ], + "angle": 0, + "content": "We observed that the other 3D-GAN baselines, EG3D and GET3D, are prone to collapsing to a single adversarial sample leading to poor FID/KID scores. The latter is probably due to the 3D misalignment of the CO3Dv2 sequences across instances, which makes training harder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "HoloFusion also outperforms the text-to-3D StableDreamFusion on both FID/KID. Stable-DreamFusion yields good shapes, but produces synthetic-looking and overly-smooth textures and thus performs poorly when compared to the real-world images of Co3Dv2. As evi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22981" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.155, + 0.101, + 0.814 + ], + "angle": 270, + "content": "HoloFusion HoloDiffusion* HoloDiffusion DreamFusion GET-3D EG3D \\(\\pi\\) -GAN" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.109, + 0.892, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.834, + 0.893, + 0.865 + ], + "angle": 0, + "content": "Figure 6: 3D samples generated by our HoloFusion compared to \\(\\pi\\)-GAN, EG3D, GET3D, HoloDiffusion, HoloDiffusion*, and the text-to-3D Stable-DreamFusion." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22982" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.135 + ], + "angle": 0, + "content": "Table 1: FID (↓) and KID (↓) on 4 classes of Co3Dv2 [30]. We compare with 3D generative modeling baselines (rows 1–5); with an SDS distillation-based Stable-DreamFusion (row 6); and with ablations of our HoloFusion (rows 7–8). The column “VP” denotes whether renders of a method are 3D view-consistent or not." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.139, + 0.891, + 0.354 + ], + "angle": 0, + "content": "
methodVPAppleHydrantTeddyBearDonutMean
FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
π-GAN [5]X49.30.04292.10.080125.80.11899.40.06991.70.077
EG3D [4]✓170.50.203229.50.253236.10.239222.30.237214.60.233
GET3D [9]✓179.10.190303.30.380244.50.280209.90.230234.20.270
HoloDiffusion [18]✓94.50.095100.50.079109.20.106115.40.085122.50.102
HoloDiffusion*X55.90.04562.60.045116.60.10199.60.07983.70.068
Stable-DreamFusion [40]✓139.00.104185.20.132183.40.125169.30.114169.20.119
HoloFusion (MSE)X72.70.06762.20.04587.20.076109.00.09982.80.072
HoloFusion (SDS)✓123.00.10577.10.058117.80.090142.80.087115.20.085
HoloFusion (Ours)✓69.20.06366.80.04787.60.075109.70.09883.30.071
" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.377, + 0.462, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.472, + 0.675 + ], + "angle": 0, + "content": "Figure 7: Heatmaps illustrating the per-pixel color variance of \\( K = 10 \\) hypothesis produced by the upsampler \\( D_{\\beta} \\). Some samples contain artifacts around the object boundaries which correspond to the high-variance regions in the figure. Our top-K patch-remix increases robustness by allowing the loss to discard such artifacts during distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.729 + ], + "angle": 0, + "content": "dent from the TeddyBear samples, the method also suffers from the \"Janus\" issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Compared to HoloDiffusion, we improve the FID/KID scores by a significant margin, mainly due to the more photo-realistic renders that include high-frequency details. Although the 2D Diffusion upsampler of HoloDiffusion* produces renders with the highest amount of details yielding scores similar to ours, they are not 3D view-consistent (as apparent from Fig. 4 and as explained in Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.871 + ], + "angle": 0, + "content": "Ablations. In Tab. 1 and in Fig. 5 we ablate components of our HoloFusion to verify their contribution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The first variant, HoloFusion (SDS), replaces the Topk patch-remixed distillation loss with the score distillation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.422 + ], + "angle": 0, + "content": "sampling (SDS) gradient as proposed in [29]. As apparent from Fig. 5 and from the lower scores, SDS washes out all the high-frequency details in the textures." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.423, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Secondly, HoloFusion (MSE) reduces the number of upsampling hypotheses \\(\\mathcal{I}\\) to the minimum of \\(|\\mathcal{I}| = 1\\). Even though this slightly improves the 2D metrics, as can be seen from Fig. 5, the samples lack view-consistency and introduce \"floaters\". In Fig. 7 we further illustrate the variability of the upsampling hypotheses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.62, + 0.545 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.677 + ], + "angle": 0, + "content": "We have presented a hybrid diffusion-based method that can generate high-quality 3D neural radiance fields of real-life object categories. Our method starts by producing coarse 3D models whose renders are independently superresolved, and finally consolidated using a robust distillation process. We evaluated our method on the Co3D v2 dataset and presented 3D-consistent, diverse, and high-quality results superior to all competing baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Our method suffers from limitations that can be addressed in future work. First, our method is slow to sample from as the sampling process takes about 30 mins for each generation, because it is still a distillation-based method. An interesting extension would be to train another network to directly distill a set of super-resolved images, without requiring explicit optimization during inference. Second, we do not produce an explicit surface representation (e.g., a mesh), which could be done by integrating a differentiable mesh render in the loop as done in some prior work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.688, + 0.862 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Animesh and Niloy were partially funded by the European Union's Horizon 2020 research and innovation pro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22983" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "gramme under the Marie Skłodowska-Curie grant agreement No. 956585. This research has been partly supported by MetaAI and the UCL AI Centre. Finally Animesh is grateful to The Rabin Ezra Scholarship Fund being a recipient of their esteemed fellowship for the year 2023." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.182, + 0.175, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.207, + 0.47, + 0.263 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv.cs, abs/2211.09869, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.264, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. arXiv.cs, abs/2207.13751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.472, + 0.629 + ], + "angle": 0, + "content": "[6] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv preprint arXiv:2212.03267, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.632, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.675, + 0.472, + 0.715 + ], + "angle": 0, + "content": "[8] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3D shape induction from 2D views of multiple objects. In arXiv, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.718, + 0.472, + 0.787 + ], + "angle": 0, + "content": "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. CoRR, abs/2110.08985, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.207, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerfguided distillation from 3d-aware diffusion. arXiv preprint arXiv:2302.10109, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[14] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.894, + 0.37 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[17] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.894, + 0.523 + ], + "angle": 0, + "content": "[18] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3d content creation. arXiv.cs, abs/2211.10440, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.894, + 0.648 + ], + "angle": 0, + "content": "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.894, + 0.704 + ], + "angle": 0, + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.894, + 0.734 + ], + "angle": 0, + "content": "[22] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.894, + 0.789 + ], + "angle": 0, + "content": "[23] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.894, + 0.817 + ], + "angle": 0, + "content": "[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[25] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: \\(360^{\\circ}\\) reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22984" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[27] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3d radiance field diffusion. In Proc. CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[28] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. arXiv.cs, abs/1904.01326, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[30] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[33] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3D-aware image synthesis. arXiv.cs, abs/2007.02442, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. ARXIV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Proc. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[36] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[37] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d gans. CoRR, abs/2206.10535, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[39] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 156-174. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.2,6,8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[41] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[42] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv.cs, abs/2210.04628, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[43] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[44] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[45] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[46] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. arXiv preprint arXiv:2212.00792, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22985" + } + ] +] \ No newline at end of file diff --git a/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_origin.pdf b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab365c82744a0b1b5791e81af50d26ec6121ec56 --- /dev/null +++ b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/ded240ab-1a84-4b65-9fba-38d695e9e368_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:031421a88e3965a902f0a1b521c97ab323f88ba709fc737a5c4de49d9dfcc478 +size 5909398 diff --git a/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/full.md b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cae8ff8f9c075fb569de9befe7f89da5fd3d566b --- /dev/null +++ b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/full.md @@ -0,0 +1,312 @@ +# HOLOFUSION: Towards Photo-realistic 3D Generative Modeling + +Animesh Karnewar UCL + +a.karnewar@ucl.ac.uk + +Niloy J. Mitra +UCL + +n.mitra@ucl.ac.uk + +Andrea Vedaldi Meta AI + +vedaldi@meta.com + +David Novotny Meta AI + +dnovotny@meta.com + +![](images/8fb702bffad84786de6fa40d73a8b07283a2b36f371f8e95035cf4d4eec0ed98.jpg) +HoloDiffusion samples +Figure 1: We propose HoloFUSION to generate photo-realistic 3D radiance fields by extending the HoloDiffusion method with a jointly trained 2D 'super resolution' network. The independently super-resolved images are fused back into the 3D representation to improve the fidelity of the 3D model via distillation, while preserving the consistency across view changes. + +![](images/f65173352270c65d0f173fb8aa431b48f989e6d57923901dca62ba92cc94ffa1.jpg) +Distillation + +![](images/6dac8a8c53b07a863c6d44e0d090043390ec06a00c6092664c7f048155911b79.jpg) +HoloFusion high quality samples + +# Abstract + +Diffusion-based image generators can now produce high-quality and diverse samples, but their success has yet to fully translate to 3D generation: existing diffusion methods can either generate low-resolution but 3D consistent outputs, or detailed 2D views of 3D objects but with potential structural defects and lacking view consistency or realism. We present HoloFUSION, a method that combines the best of these approaches to produce high-fidelity, plausible, and diverse 3D samples while learning from a collection of multi-view 2D images only. The method first generates coarse 3D samples using a variant of the recently proposed HoloDiffusion generator. Then, it independently renders and upsamples a large number of views of the coarse 3D model, super-resolves them to add detail, and distills those into a single, high-fidelity implicit 3D representation, which also ensures view-consistency of the final renders. The super-resolution network is trained as an integral part of HoloFUSION, end-to-end, and the final distillation uses a + +new sampling scheme to capture the space of super-resolved signals. We compare our method against existing baselines, including DreamFusion, Get3D, EG3D, and HoloDiffusion, and achieve, to the best of our knowledge, the most realistic results on the challenging CO3Dv2 dataset. + +# 1. Introduction + +Diffusion models [32, 7, 31] are at the basis of state-of-the-art 2D image generators which can now produce very high-quality and diverse outputs. However, their success has yet to be translated to 3D and there is no generator that can produce 3D assets of a comparable quality. + +Recent attempts at extending diffusion to 3D generation have reported mixed success. Some authors have attempted to apply diffusion directly in 3D [18], or still in 2D but using a 3D-aware neural network [42, 1]. This requires solving two problems: first, finding a suitable 3D representation (e.g., triplane features [4], mesh [19], voxels [18]) that scales well with resolution and is amenable to diffusion; + +and, second, obtaining a large amount of 3D training data, for example using synthetic models [41, 27], or training the model using only 2D images [18], often via differentiable (volume) rendering [13, 26]. However, the quality of results so far is limited, especially when training on real images. + +Other authors have proposed to distill 3D objects from pre-trained 2D image generators. For instance Score Distillation Sampling (SDS) [29] can sample 3D objects from a high-quality off-the-shelf 2D diffusion model while requiring no (re)training. However, without any 3D guidance, distillation methods often produce implausible results; for example, they suffer from the 'Janus effect', where details of the front of the object are replicated on its back. They also create overly-smooth outputs that average out inconsistencies arising from the fact that the signal obtained from the 2D model is analogous to sampling independent views of the object (see Sec. 4.2 for examples). Furthermore, distillation methods do not support unconditional sampling, even if the underlying image generator does, as strong language guidance is required to stabilise the 3D reconstruction. + +In this work, we propose HLOFUSION, a method that combines the best of both approaches. We start from HoloDiffusion [18], a diffusion-based 3D generator. This model can be trained using only a multiview image dataset like [30] and produces outputs that are 3D consistent. However, the output resolution is limited by computation and memory. We augment the base model with a lightweight super-resolution network that upscales the initial renders. Crucially, the 2D super-resolution model is integrated and trained jointly with the 3D generator, end-to-end. + +The super-resolution network outputs detailed views of the 3D object, and the underlying 3D generator ensures that the coarse structure of these views is indeed consistent (e.g., avoiding the Janus effect and other structural artifacts). However, the 2D upscaling still progresses independently for different views, which means that fine grained details may still be inconsistent between views. We address this issue by distilling a single, coherent, high quality 3D model of the object from the output of the upsampler. For this, we propose a new distillation technique that efficiently combines several putative super-resolved views of the object into a single, coherent 3D reconstruction. + +With this, we are able to train a high-quality 3D generator model purely from real 2D data. This model is capable of generating consistent and detailed 3D objects, which in turn result in view-consistent renderings (see Fig. 1) at a quality not achievable by prior methods. + +We evaluate HoloFusion on real images (CO3Dv2 dataset [30]) and compare with a variety of competing alternatives (e.g., HoloDiffusion [18], Get3D [9], EG3D [4], DreamFusion [40]) demonstrating that view-consistent high-quality 3D generation is possible using our simple, effective, easy-to-implement hybrid approach. + +# 2. Related Work + +3D generators that use adversarial learning. Generative Adversarial Learning (GAN) [10] learns a generator network so that its "fake" samples cannot be distinguished from real images by a second discriminator network. Approaches such as PlatonicGAN [13], HoloGAN [28], and PrGAN [8] introduced 3D structure into the generator network, achieving 3D shape generation with only image-level supervision. Our method is related to those as it renders images from a generated voxel grid, as well as to HoloGAN [28], which renders features and then converts them into an image by a lightweight 2D convolutional network. Other voxel-based 3D generators include VoxGRAF [34] and NeuralVolumes [21]. + +More recently, 3D generators have built on neural radiance fields [26]. GRAF [33] was the first to adopt the NeRF framework; analogous to PlatonicGAN, they generate the parameters of an MLP which renders realistic images of the object from a random viewpoint. This idea has been improved in StyleNeRF [11] and EG3D [4] by adding a 2D convolutional post-processing step after emission-absorption rendering, which is analogous to our super-resolution network. EG3D also introduced a novel 'Triplane' representation of the radiance field which, in a memory efficient manner, factorises the latter into a triplet of 2D feature planes. EG3D inspired several improvements such as GAUDI [2] and EpiGRAF [37]. + +Mesh-based 3D generators have been explored in [43]. Recently, GET3D [9] replaced the radiance field with a signed distance function to regularise the representation of geometry. The latter is converted into a mesh and rendered in a differentiable manner by using the marching tetrahedral representation [35]. + +Modeling 3D with diffusion. Diffusion methods [38] have recently become the go-to framework for generative modeling of any kind, including 3D generative modeling. The first applications of diffusion to 3D considered pointcloud generators trained on synthetic data [23, 45, 44]. + +3D distillation of 2D diffusion models. More recently, DreamFusion [29] ported the idea of distillation to diffusion models: they extract a neural radiance field so that its renders match the belief of a pre-trained 2D diffusion generator [32, 7, 31]. They introduce the Score Distillation Sampling (SDS) loss which makes distillation relatively efficient (but still in the order of several minutes for a single 3D sample). Their generation can be conditioned by an image or by a textual description, making the process rather flexible. Magic3D [19] further increases the quality of the output by distilling a mesh-based 3D representation instead of a radiance field. + +Image-conditioned 3D diffusion. The idea of distillation has been applied to few-view conditioned reconstruction in + +[42, 12, 25, 46, 6]. SparseFusion [46] employs a 3D-based new-view synthesis model [39] followed by a 2D diffusion upsampler. They complete the process by 3D distillation, ensuring that the generated views of the object are consistent. NeRFDiff [12] and 3DiM [42] bypass an explicit 3D model and directly generate new views of an object using a 2D image generator and, in the case of NeRFDiff, refine the results using distillation. + +While SparseFusion and NeRFDiff need to be trained on a dataset of object-centric multi-view images with pose information, RealFusion [25] and NeRDi [6] can be used for zero-shot monocular 3D reconstruction, starting from a pretrained 2D diffusion model. Given a single image as input, they automatically generate a prompt for the diffusion model, using a form of prompt inversion, and then use distillation to extract a radiance field. + +Unconditional generation. Most relevantly to us, unconditional generation, i.e., generation which does not require either text or image conditioning, was explored in [41, 36, 27, 18]. While [27, 36, 41] train generators given synthetic 3D ground truth, similar to us, HoloDiffusion [18] is supervised only with real object-centric images and camera poses. While HoloDiffusion was the first to demonstrate successful training on real image data, its renders contain considerably lower amount of detail than samples from a conventional 2D image generator that uses diffusion. We thus leverage a 2D diffusion upsampler, conditioned on the lower-fidelity HoloDiffusion renders, to distill higher resolution images and, eventually, 3D models. + +# 3. HOOFUSION + +We present HOOFUSION, a method that can learn a high-quality diffusion-based 3D generator from a collection of multiview 2D images. HOOFUSION first obtains an unconditional low-resolution 3D sample using diffusion and then distills a high-resolution 3D radiance field representing a higher-quality version of the generated object. We first summarize the Denoising Diffusion Probabilistic Models (DDPMs) [16] that we utilize in Sec. 3.1. Then, we discuss the low-resolution 3D generator in Sec. 3.2 followed by super-resolution distillation in Sec. 3.3. + +# 3.1. Preliminaries: DDPMs + +Let $x = x_0$ be a random vector whose probability distribution $p(x|y)$ we seek to model. The DDPM [16] defines a hierarchy of latent variables $x_{t}$ , $t = 0,\dots ,T$ and an encoder $q$ comprising a sequence of Gaussian distributions + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {\alpha_ {t}} x _ {t - 1}, (1 - \alpha_ {t}) \mathbb {I}\right), \tag {1} +$$ + +where $\alpha_{t},\ldots ,\alpha_{T}$ is a predefined 'noising schedule'. Given knowledge of $x_0$ , a sample $x_{t}$ can be drawn in a closed-form directly from $q(x_{t}|x_{0}) = \mathcal{N}(x_{t};\sqrt{\bar{\alpha}_{t}} x_{0},(1 - \bar{\alpha}_{t})\mathbb{I})$ , where + +$\bar{\alpha}_{t} = \prod_{i = 1}^{t}\alpha_{i}$ . Hence, we can express $x_{t}$ as: + +$$ +x _ {t} = \hat {\epsilon} _ {t} \left(x _ {0}\right) = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} \tag {2} +$$ + +The noising schedule is chosen such that $\bar{\alpha}_{T} \approx 0$ . In this manner, $q(x_{T}|x_{0}) \approx \mathcal{N}(x_{T};0,\mathbb{I})$ is approximately normal, and so is $q(x_{T})$ . To generate a sample $x = x_0$ , we start by sampling $x_{T}$ from this normal distribution and then sample the intermediate latent variables backward. This is done by using a variational approximation of the probabilities $q(x_{t - 1}|x_t)$ given by the Gaussian factors: + +$$ +p \left(x _ {t - 1} \mid x _ {t}\right) = \mathcal {N} \left(x _ {t - 1}; \sqrt {\bar {\alpha} _ {t - 1}} D _ {\theta} \left(x _ {t}, t\right), (1 - \bar {\alpha} _ {t - 1}) \mathbb {I}\right) \tag {3} +$$ + +where $D_{\theta}$ is a neural network with parameters $\theta$ + +The network $D_{\theta}$ is trained by maximizing the ELBO (Evidence Lower Bound), which reduces to the denoising objective [22]: + +$$ +\mathcal {L} (\theta) = \mathbb {E} _ {t, \epsilon , x} \left[ \frac {\bar {\alpha} _ {t - 1}}{2 (1 - \bar {\alpha} _ {t - 1}) ^ {2}} \| D _ {\theta} \left(\hat {\epsilon} _ {t} (x), t\right) - x \| ^ {2} \right], \tag {4} +$$ + +where $t$ is sampled uniformly from $U[1,T]$ . Hence, $D_{\theta}(x_t,t)$ approximates the clean sample $x_0$ given the noisy sample $x_{t}$ (obtained using (2)). + +# 3.2. HoloDiffusion revisited + +Given a large dataset of 3D models, the framework of Sec. 3.1 could be used to train a corresponding probability distribution. However, such a dataset is not available, and we must instead learn from 2D images of physical 3D objects. Given a dataset containing several views of a large number of objects, we could use image-based reconstruction (e.g., using neural rendering) to obtain corresponding 3D models first, and then use those to train a diffusion model. Instead, we adopt, and slightly upgrade, the HoloDiffusion method [18], which learns a 3D diffusion model directly from the 2D images. + +Training data. HoloDiffusion learns from a collection $\mathcal{D}$ of $N$ image sequences $s_i = (I_j^i, C_j^i)_{j=1}^{N_{\mathrm{frame}}}$ , $i = 1, \ldots, N$ , where frame $I_j^i \in \mathbb{R}^{3 \times H \times W}$ is an RGB image and $C_j^i \in \mathbb{R}^{4 \times 4}$ is the corresponding camera projection matrix, collectively defining the motion of the camera. + +3D representation and rendering. The shape and appearance of the object are represented by a voxel grid $V \in \mathbb{R}^{d \times S \times S \times S}$ with resolution $S$ containing a $d$ -dimensional feature vector per voxel. Given a 3D point $\mathbf{p} \in \mathbb{R}^3$ , its opacity $\sigma(\mathbf{p}) \in \mathbb{R}_+$ and color $c(\mathbf{p}) \in \mathbb{R}_{[0,1]}^3$ are obtained from the voxel grid by an MLP $M_{\eta}(V(\mathbf{p}))$ that takes as input the $d$ -dimensional feature vector $V(\mathbf{p})$ extracted from the grid via trilinear interpolation [20]. The usual emission-absorption model [26, 24] is then used to implement a differentiable rendering function $R_{\eta}$ , mapping the voxel grid $V$ and the camera viewpoint $C$ into an image $\hat{I} = R_{\eta}(V,C)$ , where $\eta$ are the parameters of the MLP. + +![](images/8ea85f537f37312e8a7c69fdbeec3bffe7e20f1e7221e40bb31e3cd5b96cc611.jpg) +Figure 2: Overview. HOOFUSION, which trains the 3D denoiser network $D_{\theta}$ , is augmented with the 2D 'super-resolution' diffusion model $D_{\beta}$ . Both models are trained end-to-end by supervising their outputs with 2D photometric error. + +Training scheme. HoloDiffusion leverages the DDPM framework (revised in the previous paragraphs) to recover the density $p(V)$ over voxel grids $x = V$ encoding plausible real-life objects. In order to train a DDPM on such 3D data, we would need access to ground-truth 3D models $V$ , which are not available. HoloDiffusion addresses this problem by making three changes to DDPM. + +First, it replaces the data denoising loss with a photometric reconstruction loss. Given a pair $(I,C)\in s$ from one of the training sequences $s$ , it replaces Eq. (4) with $\mathbb{E}_{t,\epsilon ,C}\left[\| I - R_{\eta}(D_{\theta}(\hat{\epsilon}_t(V),t),C)\| ^2\right]$ where the goal is not to reconstruct the 'clean' volume $V$ (which is unknown), but rather its image $I$ (which is known). + +Second, also because the 'clean' volume $V$ is not available, we cannot use Eq. (3) to generate the noisy volumes $V_{t}$ to denoise; the only exception is the last sample $V_{T}$ , which is pure noise. This suggests to adopt a 'double denoising' step. First, pure noise $V_{T}$ is fed into the denoiser to obtain an (approximate) version of $V_{0} = D_{\theta}(V_{T}, T)$ of the clean volume $V_{0} = V$ . Then, noise is applied to obtain $V_{t} = \hat{\epsilon}_{t}(V_{0}) = \sqrt{\bar{\alpha}_{T}} V_{0} + \sqrt{1 - \bar{\alpha}_{t}} \epsilon_{t}$ according to Eq. (3), and the latter is fed back into the denoiser as above. + +Finally, there is the issue that unconditional generation of the clean volume $V_{0}$ from pure noise $V_{T}$ is difficult, especially in the first iterations of training. On the other hand, the problem of view-conditioned generation is considerably easier. Hence, the third idea is to learn a conditional generator, using a variable number of input views. Specifically, given a training sequence $s$ , the method extracts a random subset of frames $\bar{s} \subset s$ (which could be empty, which cor- + +responds to unconditional generation). Then, a feature volume $\bar{V} = \Phi (\bar{s})\in \mathbb{R}^{d\times S\times S\times S}$ is obtained from the selected frames. This extracts 2D image features using a pre-trained and frozen 2D image encoder and then pools them in 3D via 'unprojection' [17, 14] into $\bar{V}$ , where $\bar{V} = 0$ if $\bar{s}$ is empty. Finally, these pooled features are used to condition the denoising network $V_{0} = D_{\theta}(V_{T},\bar{V},T)$ , which, on average, leads to a simpler reconstruction problem. + +Putting it all together, the training loss becomes: + +$$ +\mathcal {L} (\theta | I, C, \bar {s}) = \mathbb {E} _ {t, \hat {\epsilon}, V _ {T}} \left[ \| \hat {I} - I \| ^ {2} \right], \tag {5} +$$ + +$$ +\text {w h e r e} \hat {I} = R _ {\eta} \left(D _ {\theta} \left(V _ {t}, \bar {V}, t\right), C\right), \tag {6} +$$ + +$$ +V _ {t} = \hat {\epsilon} _ {t} (V _ {0}), +$$ + +$$ +V _ {0} = D _ {\theta} (V _ {T}, \bar {V}, T), +$$ + +$$ +\bar {V} = \Phi (\bar {s}). +$$ + +This loss is averaged over training sequences $s$ , subsequences $\bar{s} \subset s$ , and views $(I, C) \in s$ therein. Note that this is slightly different than the original HoloDiffusion, where feature volume $\tilde{V}$ and reconstructed volumes $V_{t}$ overlap as arguments of the denoiser; we found that keeping them separated in the formulation leads to more stable training and additionally allows for view-conditioned generation. + +# 3.3. HOOFUSION + +The method of Sec. 3.2 learns to generate 3D objects from 2D image supervision only, but the fidelity of the output is limited by the resolution at which the operations are carried out. Increasing resolution is difficult due to the GPU + +memory impact of the voxel-based representation, so we seek a more efficient way to do so. The idea is to incorporate a 2D super-resolution network (Sec. 3.3.1), trained end-to-end, that improves the output from the base model. The super-resolved images are eventually fused back in an improved 3D model, which also has the benefit of further increasing view consistency (Sec. 3.3.2). + +# 3.3.1 Integrating super-resolution + +As shown in Fig. 2, we augment the method of Sec. 3.2 with a lightweight refinement post-processor network that takes the 2D image $\hat{I}$ generated by the base model and outputs a higher quality version $\hat{I}_{\mathrm{super}}$ of the same. This can be thought of as a form of super-resolution; however, due to the particular statistics of the input ('low-res') images $\hat{I}$ that HoloDiffusion generates, it is necessary to train this superresolution network in an end-to-end fashion with HoloDiffusion, integrating the two models. + +To make this integration seamless, we formulate super-resolution as another diffusion process that runs 'in parallel' with 3D reconstruction. Hence, the super-resolved image $\hat{I}_{\mathrm{super}} = D_{\beta}(I_t,\hat{I},t)$ is the output of a denoiser network (a lightweight U-Net), which takes as input the noised target image $I_{t} = \hat{\epsilon}_{t}(I)$ and is also conditioned on the 'low-res' output $\hat{I} = R_{\eta}(V,C)$ of HoloDiffusion from Eq. (6). This denoiser is trained with the DDPM loss: + +$$ +\mathcal {L} (\beta | I) = \mathbb {E} _ {t, \hat {\epsilon}} \left[ \| D _ {\beta} \left(\hat {\epsilon} _ {t} (I), \hat {I}, t\right) - I \| ^ {2} \right]. \tag {7} +$$ + +Training details The overall model $(D_{\beta}$ and $D_{\theta})$ is trained end-to-end by optimising the loss $\mathcal{L}(\theta |I,C,\bar{s}) + \mathcal{L}(\beta |I)$ obtained by summing Eqs. (5) and (7). + +As training data, we use a large dataset of images capturing object-centric scenes ([30]). In each training batch, we pick a random training scene $s$ and sample 15 different source images $\bar{s}_{\mathrm{src}} \subset s$ which are unprojected to generate the feature volume conditioning $\bar{V}$ . Then, $\bar{V}$ is rendered into 4 random target views $\bar{s}_{\mathrm{tgt}} \subset (s \setminus \bar{s}_{\mathrm{src}})$ which allows to optimize the training image reconstruction loss $\mathcal{L}(\theta | I, C, \bar{s}) + \mathcal{L}(\beta | I)$ . The latter uses the Adam optimizer with an initial learning rate of $5 \cdot 10^{-5}$ decaying tenfold whenever the loss plateaus until convergence. + +# 3.3.2 Fusing super-resolved views in 3D + +The method of Sec. 3.3.1 leaves us with high-resolution views $\hat{I}_{\mathrm{super}}$ of the generated 3D object. However, we would like to obtain a single, high-quality 3D model, not just individual views of it. In this section, we discuss how to take the super-resolved images and fuse them into such a model, while addressing the issues that these images are not perfectly view-consistent. + +![](images/0ecabef4deb273ec11a28889a588e4e8b4c8c8e637b4fbc031a3307b76289271.jpg) +Figure 3: Distillation. HOOFUSION distills a single high-resolution voxel grid $V_0^H$ by minimizing a top-k patch-remix loss $\mathcal{L}_{\mathrm{distil}}$ between the grid renders $R_{\eta'}(V_0^H, C)$ and a bank $\mathcal{I}_C$ of $K = 5$ high-res images output by the 2D diffusion upsampler $D_{\beta}$ for each scene camera $C$ . + +The basic idea is simple. We can generate a certain number (e.g., 100) high-resolution images of the object from different viewpoints $C$ and then use a technique, akin to neural rendering, to fuse them back into a single 3D model. However, there is a problem with this idea: The model of Sec. 3.3.1 generates high-quality views $I_{\mathrm{super}}$ , but these are view-dependent samples from the distribution $p(\hat{I}_{\mathrm{super}}|\hat{I})$ where $\hat{I} = R_{\eta}(V,C)$ is the 'low-res' output form HoloDiffusion. Because super-resolving details is intrinsically ambiguous, there is no reason why samples $I_{\mathrm{super}}$ taken from different viewpoints $C$ would be consistent (Fig. 7). Fusing them into a single 3D model would then result in a blurry appearance yet again. + +As described in Fig. 3, we address this issue in a principled manner by considering several possible super-resolved images $\mathcal{I}_C = \{I_{\mathrm{super}}\sim p(\hat{I}_{\mathrm{super}}|\hat{I})\}$ sampled from each given viewpoint $C$ . Then, we optimize a high-resolution voxel grid $V_0^H$ by minimizing the photometric loss: + +$$ +\mathcal {L} _ {\text {d i s t i l}} \left(\eta^ {\prime}, V _ {0} ^ {H} \mid \mathcal {I} _ {C}\right) = \mathbb {E} _ {C} \left[ \min _ {I _ {\text {s u p e r}} \in \mathcal {I} _ {C}} \| I _ {\text {s u p e r}} - R _ {\eta^ {\prime}} \left(V _ {0} ^ {H}, C\right) \| ^ {2} \right] \tag {8} +$$ + +where $R_{\eta'}(V_0^H, C)$ is the render of a high-resolution voxel grid $V_0^H \in \mathbb{R}^{d \times S' \times S' \times S'}$ , $S' > S$ using the learnable renderer $R_{\eta'}$ with scene specific parameters $\eta'$ . Minimizing with respect to $I_{\mathrm{super}}$ means that the 3D model must be consistent with at least one of the possible super-resolved images, drawn from the distribution of super-resolved samples, for each view $C$ . + +Patch remix. In practice, this approach requires a very large number of super resolved images $\mathcal{I}_C$ to be effective. We found that we can significantly improve the statistical efficiency by performing the minimization at the level of individual patches. Namely, we produce a stack of only + +![](images/6caeb0cb9e20918eeba8abf09ce78107c7727018f0f599c8f0526df050d2022e.jpg) +Figure 4: Generated 3D samples visualized from a moving camera. $\pi$ -GAN and HoloDiffusion* fail to produce 3D view consistent samples, while DreamFusion suffers from the "Janus" problem (multiple heads). + +$K = |\mathcal{I}_C| = 5$ super resolved images and perform the minimization in Eq. (8) at the level of small $16 \times 16$ patches independently (effectively allowing super-resolved images to 'remix' as needed to fit the generated view $R_{\eta'}(V_0^H, C)$ ). + +Distillation details. $\mathcal{L}_{\mathrm{distil}}$ is optimized independently for each generated scene with Adam $(1\mathrm{r} = 2\cdot 10^{-4})$ for $25\mathrm{K}$ steps until convergence. While $\eta^\prime$ is initialized using the pretrained multi-sequence weights $\eta$ , $V_0^H$ is initialized by tri-linearly upsampling the low-resolution volume $V_{0}$ output by HoloDiffusion. Cameras $C$ are sampled at uniform azimuths with elevation fixed at object's equator. + +![](images/19d0793b669aa1114c512b9643516e5b124e74034a0bb4f066efa75cf2b01315.jpg) +Figure 5: Fusing views. Our patch-remix (Sec. 3.3.2) compared to the SDS and MSE distillation. MSE has "floaters" and viewpoint inconsistencies, SDS over-smooths the texture. Ours is robust and produces superior quality. + +# 4. Experiments + +We begin with a description of the experiments conducted in Sec. 4.1, followed by an analysis and discussion of the results in Sec. 4.2. + +# 4.1. Details + +Dataset. We experiment on the challenging large-scale Co3Dv2 [30] dataset which is a popular choice for a real-world 3D reconstruction benchmark. More specifically, 4 categories are selected, Apple, Hydrant, TeddyBear, and Donut, with 500 3D-scenes per category for training. Each 3D scene contains $\sim 200$ images of the object of interest along with poses of their corresponding cameras. + +Baselines. We use two sets of baselines for comparison (Tab. 1): (i) general 3D generative modeling baselines and (ii) diffusion distillation based baselines. $\pi$ -GAN [5], EG3D [4], GET3D [9], and HoloDiffusion [18] are considered as the 3D generative baselines. Along with HoloDiffusion, we also test the super-resolution integrated model (described in Sec. 3.3.1) HoloDiffusion*. For the distillation-based baselines, we consider the open-source implementation of DreamFusion [29] titled Stable-DreamFusion [40]. For the latter, scenes are generated by conditioning on prompts comprising names of Co3Dv2 categories extended with color and style modifier phrases leading to $\sim 200$ prompts / 3D shapes per class. More details regarding the prompt creation are in the supplementary. + +Metrics. We use FID [15] and KID [3] to compare the quality of our 2D renders, as these are commonly used to assess 2D and 3D generators. + +# 4.2. Quantitative and qualitative analysis + +Tab. 1 evaluates quantitatively while Fig. 6 qualitatively. Furthermore, Fig. 4 compares rendering view-consistency. + +HoloFusion (Ours) yields better FID/KID scores than the general 3D generative baselines except for $\pi$ -GAN on Apple and Donut classes. However, since $\pi$ -GAN does not guarantee view consistency by design, it essentially acts as a 2D image GAN, and thus does better on the 2D FID/KID metrics, but it generates significantly view-inconsistent renders (see Fig. 4 and the supplementary). + +We observed that the other 3D-GAN baselines, EG3D and GET3D, are prone to collapsing to a single adversarial sample leading to poor FID/KID scores. The latter is probably due to the 3D misalignment of the CO3Dv2 sequences across instances, which makes training harder. + +HoloFusion also outperforms the text-to-3D StableDreamFusion on both FID/KID. Stable-DreamFusion yields good shapes, but produces synthetic-looking and overly-smooth textures and thus performs poorly when compared to the real-world images of Co3Dv2. As evi + +![](images/382b412cae39adb21545ba1a92e4e4c3d05a0f609a7303a75bfd6f8c55058c48.jpg) +HoloFusion HoloDiffusion* HoloDiffusion DreamFusion GET-3D EG3D $\pi$ -GAN +Figure 6: 3D samples generated by our HoloFusion compared to $\pi$ -GAN, EG3D, GET3D, HoloDiffusion, HoloDiffusion*, and the text-to-3D Stable-DreamFusion. + +Table 1: FID (↓) and KID (↓) on 4 classes of Co3Dv2 [30]. We compare with 3D generative modeling baselines (rows 1–5); with an SDS distillation-based Stable-DreamFusion (row 6); and with ablations of our HoloFusion (rows 7–8). The column “VP” denotes whether renders of a method are 3D view-consistent or not. + +
methodVPAppleHydrantTeddyBearDonutMean
FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
π-GAN [5]X49.30.04292.10.080125.80.11899.40.06991.70.077
EG3D [4]✓170.50.203229.50.253236.10.239222.30.237214.60.233
GET3D [9]✓179.10.190303.30.380244.50.280209.90.230234.20.270
HoloDiffusion [18]✓94.50.095100.50.079109.20.106115.40.085122.50.102
HoloDiffusion*X55.90.04562.60.045116.60.10199.60.07983.70.068
Stable-DreamFusion [40]✓139.00.104185.20.132183.40.125169.30.114169.20.119
HoloFusion (MSE)X72.70.06762.20.04587.20.076109.00.09982.80.072
HoloFusion (SDS)✓123.00.10577.10.058117.80.090142.80.087115.20.085
HoloFusion (Ours)✓69.20.06366.80.04787.60.075109.70.09883.30.071
+ +![](images/17b4ed1e7f688b0fa3b2eb10816ca8efd1afcc85b08419cb2e334b250096c7b6.jpg) +Figure 7: Heatmaps illustrating the per-pixel color variance of $K = 10$ hypothesis produced by the upsampler $D_{\beta}$ . Some samples contain artifacts around the object boundaries which correspond to the high-variance regions in the figure. Our top-K patch-remix increases robustness by allowing the loss to discard such artifacts during distillation. + +dent from the TeddyBear samples, the method also suffers from the "Janus" issue. + +Compared to HoloDiffusion, we improve the FID/KID scores by a significant margin, mainly due to the more photo-realistic renders that include high-frequency details. Although the 2D Diffusion upsampler of HoloDiffusion* produces renders with the highest amount of details yielding scores similar to ours, they are not 3D view-consistent (as apparent from Fig. 4 and as explained in Sec. 3.3). + +Ablations. In Tab. 1 and in Fig. 5 we ablate components of our HoloFusion to verify their contribution. + +The first variant, HoloFusion (SDS), replaces the Topk patch-remixed distillation loss with the score distillation + +sampling (SDS) gradient as proposed in [29]. As apparent from Fig. 5 and from the lower scores, SDS washes out all the high-frequency details in the textures. + +Secondly, HoloFusion (MSE) reduces the number of upsampling hypotheses $\mathcal{I}$ to the minimum of $|\mathcal{I}| = 1$ . Even though this slightly improves the 2D metrics, as can be seen from Fig. 5, the samples lack view-consistency and introduce "floaters". In Fig. 7 we further illustrate the variability of the upsampling hypotheses. + +# 5. Conclusion + +We have presented a hybrid diffusion-based method that can generate high-quality 3D neural radiance fields of real-life object categories. Our method starts by producing coarse 3D models whose renders are independently superresolved, and finally consolidated using a robust distillation process. We evaluated our method on the Co3D v2 dataset and presented 3D-consistent, diverse, and high-quality results superior to all competing baselines. + +Our method suffers from limitations that can be addressed in future work. First, our method is slow to sample from as the sampling process takes about 30 mins for each generation, because it is still a distillation-based method. An interesting extension would be to train another network to directly distill a set of super-resolved images, without requiring explicit optimization during inference. Second, we do not produce an explicit surface representation (e.g., a mesh), which could be done by integrating a differentiable mesh render in the loop as done in some prior work. + +# 6. Acknowledgements + +Animesh and Niloy were partially funded by the European Union's Horizon 2020 research and innovation pro + +gramme under the Marie Skłodowska-Curie grant agreement No. 956585. This research has been partly supported by MetaAI and the UCL AI Centre. Finally Animesh is grateful to The Rabin Ezra Scholarship Fund being a recipient of their esteemed fellowship for the year 2023. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv.cs, abs/2211.09869, 2022. 1 +[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. arXiv.cs, abs/2207.13751, 2022. 2 +[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6 +[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 6, 8 +[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 6, 8 +[6] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv preprint arXiv:2212.03267, 2022. 3 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2 +[8] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3D shape induction from 2D views of multiple objects. In arXiv, 2016. 2 +[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 6, 8 +[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. CoRR, abs/2110.08985, 2021. 2 + +[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerfguided distillation from 3d-aware diffusion. arXiv preprint arXiv:2302.10109, 2023. 3 +[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 2 +[14] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3 +[17] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 4 +[18] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6, 8 +[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3d content creation. arXiv.cs, abs/2211.10440, 2022. 1, 2 +[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3 +[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2 +[22] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3 +[23] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2 +[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3 +[25] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: $360^{\circ}$ reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 3 +[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: + +Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 2, 3 +[27] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3d radiance field diffusion. In Proc. CVPR, 2022. 2, 3 +[28] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. arXiv.cs, abs/1904.01326, 2019. 2 +[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 8 +[30] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2, 5, 6, 8 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2 +[33] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3D-aware image synthesis. arXiv.cs, abs/2007.02442, 2020. 2 +[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. ARXIV, 2022. 2 +[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Proc. NeurIPS, 2021. 2 +[36] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 3 +[37] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d gans. CoRR, abs/2206.10535, 2022. 2 +[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2 +[39] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 156-174. Springer, 2022. 3 +[40] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.2,6,8 + +[41] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2, 3 +[42] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv.cs, abs/2210.04628, 2022. 1, 3 +[43] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2 +[44] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 2 +[45] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2 +[46] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. arXiv preprint arXiv:2212.00792, 2022. 3 \ No newline at end of file diff --git a/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/images.zip b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..31ec9dce76c7c40e3e3fd2363463fabd7be9e866 --- /dev/null +++ b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4345e0d1e4a3a365702ac0369e82f6796ae5d3fbaf64d41dc59326c92d533007 +size 713707 diff --git a/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/layout.json b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..68a88244f6475b2a4e3c2c72f4e396946d15f046 --- /dev/null +++ b/2023/HoloFusion_ Towards Photo-realistic 3D Generative Modeling/layout.json @@ -0,0 +1,9166 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "text", + "content": "HOLOFUSION: Towards Photo-realistic 3D Generative Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 144, + 171, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 144, + 171, + 170 + ], + "spans": [ + { + "bbox": [ + 75, + 144, + 171, + 170 + ], + "type": "text", + "content": "Animesh Karnewar UCL" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 174, + 178, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 174, + 178, + 184 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 178, + 184 + ], + "type": "text", + "content": "a.karnewar@ucl.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 213, + 144, + 283, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 144, + 283, + 170 + ], + "spans": [ + { + "bbox": [ + 213, + 144, + 283, + 170 + ], + "type": "text", + "content": "Niloy J. Mitra \nUCL" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 174, + 295, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 174, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 202, + 174, + 295, + 184 + ], + "type": "text", + "content": "n.mitra@ucl.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 144, + 402, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 144, + 402, + 170 + ], + "spans": [ + { + "bbox": [ + 325, + 144, + 402, + 170 + ], + "type": "text", + "content": "Andrea Vedaldi Meta AI" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 320, + 174, + 408, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 174, + 408, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 174, + 408, + 184 + ], + "type": "text", + "content": "vedaldi@meta.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 441, + 144, + 515, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 144, + 515, + 170 + ], + "spans": [ + { + "bbox": [ + 441, + 144, + 515, + 170 + ], + "type": "text", + "content": "David Novotny Meta AI" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 432, + 174, + 525, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 174, + 525, + 185 + ], + "spans": [ + { + "bbox": [ + 432, + 174, + 525, + 185 + ], + "type": "text", + "content": "dnovotny@meta.com" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 49, + 209, + 247, + 423 + ], + "blocks": [ + { + "bbox": [ + 105, + 196, + 190, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 190, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 190, + 206 + ], + "type": "text", + "content": "HoloDiffusion samples" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 209, + 247, + 423 + ], + "lines": [ + { + "bbox": [ + 49, + 209, + 247, + 423 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 247, + 423 + ], + "type": "image", + "image_path": "8fb702bffad84786de6fa40d73a8b07283a2b36f371f8e95035cf4d4eec0ed98.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 424, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 424, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 547, + 460 + ], + "type": "text", + "content": "Figure 1: We propose HoloFUSION to generate photo-realistic 3D radiance fields by extending the HoloDiffusion method with a jointly trained 2D 'super resolution' network. The independently super-resolved images are fused back into the 3D representation to improve the fidelity of the 3D model via distillation, while preserving the consistency across view changes." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 259, + 211, + 341, + 422 + ], + "blocks": [ + { + "bbox": [ + 280, + 196, + 322, + 205 + ], + "lines": [ + { + "bbox": [ + 280, + 196, + 322, + 205 + ], + "spans": [ + { + "bbox": [ + 280, + 196, + 322, + 205 + ], + "type": "text", + "content": "Distillation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 259, + 211, + 341, + 422 + ], + "lines": [ + { + "bbox": [ + 259, + 211, + 341, + 422 + ], + "spans": [ + { + "bbox": [ + 259, + 211, + 341, + 422 + ], + "type": "image", + "image_path": "f65173352270c65d0f173fb8aa431b48f989e6d57923901dca62ba92cc94ffa1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 351, + 209, + 551, + 422 + ], + "blocks": [ + { + "bbox": [ + 391, + 196, + 512, + 206 + ], + "lines": [ + { + "bbox": [ + 391, + 196, + 512, + 206 + ], + "spans": [ + { + "bbox": [ + 391, + 196, + 512, + 206 + ], + "type": "text", + "content": "HoloFusion high quality samples" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 209, + 551, + 422 + ], + "lines": [ + { + "bbox": [ + 351, + 209, + 551, + 422 + ], + "spans": [ + { + "bbox": [ + 351, + 209, + 551, + 422 + ], + "type": "image", + "image_path": "6dac8a8c53b07a863c6d44e0d090043390ec06a00c6092664c7f048155911b79.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 471, + 191, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 471, + 191, + 483 + ], + "spans": [ + { + "bbox": [ + 143, + 471, + 191, + 483 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 498, + 290, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 290, + 714 + ], + "type": "text", + "content": "Diffusion-based image generators can now produce high-quality and diverse samples, but their success has yet to fully translate to 3D generation: existing diffusion methods can either generate low-resolution but 3D consistent outputs, or detailed 2D views of 3D objects but with potential structural defects and lacking view consistency or realism. We present HoloFUSION, a method that combines the best of these approaches to produce high-fidelity, plausible, and diverse 3D samples while learning from a collection of multi-view 2D images only. The method first generates coarse 3D samples using a variant of the recently proposed HoloDiffusion generator. Then, it independently renders and upsamples a large number of views of the coarse 3D model, super-resolves them to add detail, and distills those into a single, high-fidelity implicit 3D representation, which also ensures view-consistency of the final renders. The super-resolution network is trained as an integral part of HoloFUSION, end-to-end, and the final distillation uses a" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 472, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 547, + 533 + ], + "type": "text", + "content": "new sampling scheme to capture the space of super-resolved signals. We compare our method against existing baselines, including DreamFusion, Get3D, EG3D, and HoloDiffusion, and achieve, to the best of our knowledge, the most realistic results on the challenging CO3Dv2 dataset." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 545, + 387, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 387, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 387, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 567, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 546, + 628 + ], + "type": "text", + "content": "Diffusion models [32, 7, 31] are at the basis of state-of-the-art 2D image generators which can now produce very high-quality and diverse outputs. However, their success has yet to be translated to 3D and there is no generator that can produce 3D assets of a comparable quality." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Recent attempts at extending diffusion to 3D generation have reported mixed success. Some authors have attempted to apply diffusion directly in 3D [18], or still in 2D but using a 3D-aware neural network [42, 1]. This requires solving two problems: first, finding a suitable 3D representation (e.g., triplane features [4], mesh [19], voxels [18]) that scales well with resolution and is amenable to diffusion;" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22976" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "and, second, obtaining a large amount of 3D training data, for example using synthetic models [41, 27], or training the model using only 2D images [18], often via differentiable (volume) rendering [13, 26]. However, the quality of results so far is limited, especially when training on real images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 312 + ], + "type": "text", + "content": "Other authors have proposed to distill 3D objects from pre-trained 2D image generators. For instance Score Distillation Sampling (SDS) [29] can sample 3D objects from a high-quality off-the-shelf 2D diffusion model while requiring no (re)training. However, without any 3D guidance, distillation methods often produce implausible results; for example, they suffer from the 'Janus effect', where details of the front of the object are replicated on its back. They also create overly-smooth outputs that average out inconsistencies arising from the fact that the signal obtained from the 2D model is analogous to sampling independent views of the object (see Sec. 4.2 for examples). Furthermore, distillation methods do not support unconditional sampling, even if the underlying image generator does, as strong language guidance is required to stabilise the 3D reconstruction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 434 + ], + "type": "text", + "content": "In this work, we propose HLOFUSION, a method that combines the best of both approaches. We start from HoloDiffusion [18], a diffusion-based 3D generator. This model can be trained using only a multiview image dataset like [30] and produces outputs that are 3D consistent. However, the output resolution is limited by computation and memory. We augment the base model with a lightweight super-resolution network that upscales the initial renders. Crucially, the 2D super-resolution model is integrated and trained jointly with the 3D generator, end-to-end." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 579 + ], + "type": "text", + "content": "The super-resolution network outputs detailed views of the 3D object, and the underlying 3D generator ensures that the coarse structure of these views is indeed consistent (e.g., avoiding the Janus effect and other structural artifacts). However, the 2D upscaling still progresses independently for different views, which means that fine grained details may still be inconsistent between views. We address this issue by distilling a single, coherent, high quality 3D model of the object from the output of the upsampler. For this, we propose a new distillation technique that efficiently combines several putative super-resolved views of the object into a single, coherent 3D reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": "With this, we are able to train a high-quality 3D generator model purely from real 2D data. This model is capable of generating consistent and detailed 3D objects, which in turn result in view-consistent renderings (see Fig. 1) at a quality not achievable by prior methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "We evaluate HoloFusion on real images (CO3Dv2 dataset [30]) and compare with a variety of competing alternatives (e.g., HoloDiffusion [18], Get3D [9], EG3D [4], DreamFusion [40]) demonstrating that view-consistent high-quality 3D generation is possible using our simple, effective, easy-to-implement hybrid approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 92, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 92, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 92, + 545, + 247 + ], + "type": "text", + "content": "3D generators that use adversarial learning. Generative Adversarial Learning (GAN) [10] learns a generator network so that its \"fake\" samples cannot be distinguished from real images by a second discriminator network. Approaches such as PlatonicGAN [13], HoloGAN [28], and PrGAN [8] introduced 3D structure into the generator network, achieving 3D shape generation with only image-level supervision. Our method is related to those as it renders images from a generated voxel grid, as well as to HoloGAN [28], which renders features and then converts them into an image by a lightweight 2D convolutional network. Other voxel-based 3D generators include VoxGRAF [34] and NeuralVolumes [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 248, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 403 + ], + "type": "text", + "content": "More recently, 3D generators have built on neural radiance fields [26]. GRAF [33] was the first to adopt the NeRF framework; analogous to PlatonicGAN, they generate the parameters of an MLP which renders realistic images of the object from a random viewpoint. This idea has been improved in StyleNeRF [11] and EG3D [4] by adding a 2D convolutional post-processing step after emission-absorption rendering, which is analogous to our super-resolution network. EG3D also introduced a novel 'Triplane' representation of the radiance field which, in a memory efficient manner, factorises the latter into a triplet of 2D feature planes. EG3D inspired several improvements such as GAUDI [2] and EpiGRAF [37]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 404, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 476 + ], + "type": "text", + "content": "Mesh-based 3D generators have been explored in [43]. Recently, GET3D [9] replaced the radiance field with a signed distance function to regularise the representation of geometry. The latter is converted into a mesh and rendered in a differentiable manner by using the marching tetrahedral representation [35]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 479, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 540 + ], + "type": "text", + "content": "Modeling 3D with diffusion. Diffusion methods [38] have recently become the go-to framework for generative modeling of any kind, including 3D generative modeling. The first applications of diffusion to 3D considered pointcloud generators trained on synthetic data [23, 45, 44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 543, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 685 + ], + "type": "text", + "content": "3D distillation of 2D diffusion models. More recently, DreamFusion [29] ported the idea of distillation to diffusion models: they extract a neural radiance field so that its renders match the belief of a pre-trained 2D diffusion generator [32, 7, 31]. They introduce the Score Distillation Sampling (SDS) loss which makes distillation relatively efficient (but still in the order of several minutes for a single 3D sample). Their generation can be conditioned by an image or by a textual description, making the process rather flexible. Magic3D [19] further increases the quality of the output by distilling a mesh-based 3D representation instead of a radiance field." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Image-conditioned 3D diffusion. The idea of distillation has been applied to few-view conditioned reconstruction in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22977" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "[42, 12, 25, 46, 6]. SparseFusion [46] employs a 3D-based new-view synthesis model [39] followed by a 2D diffusion upsampler. They complete the process by 3D distillation, ensuring that the generated views of the object are consistent. NeRFDiff [12] and 3DiM [42] bypass an explicit 3D model and directly generate new views of an object using a 2D image generator and, in the case of NeRFDiff, refine the results using distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 288, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 262 + ], + "type": "text", + "content": "While SparseFusion and NeRFDiff need to be trained on a dataset of object-centric multi-view images with pose information, RealFusion [25] and NeRDi [6] can be used for zero-shot monocular 3D reconstruction, starting from a pretrained 2D diffusion model. Given a single image as input, they automatically generate a prompt for the diffusion model, using a form of prompt inversion, and then use distillation to extract a radiance field." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 422 + ], + "type": "text", + "content": "Unconditional generation. Most relevantly to us, unconditional generation, i.e., generation which does not require either text or image conditioning, was explored in [41, 36, 27, 18]. While [27, 36, 41] train generators given synthetic 3D ground truth, similar to us, HoloDiffusion [18] is supervised only with real object-centric images and camera poses. While HoloDiffusion was the first to demonstrate successful training on real image data, its renders contain considerably lower amount of detail than samples from a conventional 2D image generator that uses diffusion. We thus leverage a 2D diffusion upsampler, conditioned on the lower-fidelity HoloDiffusion renders, to distill higher resolution images and, eventually, 3D models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 433, + 137, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 137, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 137, + 445 + ], + "type": "text", + "content": "3. HOOFUSION" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 453, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 572 + ], + "type": "text", + "content": "We present HOOFUSION, a method that can learn a high-quality diffusion-based 3D generator from a collection of multiview 2D images. HOOFUSION first obtains an unconditional low-resolution 3D sample using diffusion and then distills a high-resolution 3D radiance field representing a higher-quality version of the generated object. We first summarize the Denoising Diffusion Probabilistic Models (DDPMs) [16] that we utilize in Sec. 3.1. Then, we discuss the low-resolution 3D generator in Sec. 3.2 followed by super-resolution distillation in Sec. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 579, + 178, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 178, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 178, + 591 + ], + "type": "text", + "content": "3.1. Preliminaries: DDPMs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "x = x_0" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": " be a random vector whose probability distribution " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "p(x|y)" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": " we seek to model. The DDPM [16] defines a hierarchy of latent variables " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "t = 0,\\dots ,T" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": " and an encoder " + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 597, + 287, + 646 + ], + "type": "text", + "content": " comprising a sequence of Gaussian distributions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 655, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 655, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 79, + 655, + 287, + 669 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} x _ {t - 1}, (1 - \\alpha_ {t}) \\mathbb {I}\\right), \\tag {1}", + "image_path": "ab480560d98c88991f9c2d4228c8cec81d3a37d5762fcd519329fff1d090ee2f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\ldots ,\\alpha_{T}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " is a predefined 'noising schedule'. Given knowledge of " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " , a sample " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " can be drawn in a closed-form directly from " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{0}) = \\mathcal{N}(x_{t};\\sqrt{\\bar{\\alpha}_{t}} x_{0},(1 - \\bar{\\alpha}_{t})\\mathbb{I})" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " , where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 71, + 487, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 487, + 86 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 487, + 86 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}" + }, + { + "bbox": [ + 306, + 71, + 487, + 86 + ], + "type": "text", + "content": " . Hence, we can express " + }, + { + "bbox": [ + 306, + 71, + 487, + 86 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 306, + 71, + 487, + 86 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 91, + 545, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 91, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 351, + 91, + 545, + 105 + ], + "type": "interline_equation", + "content": "x _ {t} = \\hat {\\epsilon} _ {t} \\left(x _ {0}\\right) = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {2}", + "image_path": "1195ffa3004f25f9e024147f7ca8422697ab4ac4b03e5514a0d15cbae5eb9e74.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "content": "The noising schedule is chosen such that " + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{T} \\approx 0" + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "content": ". In this manner, " + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "inline_equation", + "content": "q(x_{T}|x_{0}) \\approx \\mathcal{N}(x_{T};0,\\mathbb{I})" + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "content": " is approximately normal, and so is " + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "inline_equation", + "content": "q(x_{T})" + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "content": ". To generate a sample " + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "inline_equation", + "content": "x = x_0" + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "content": ", we start by sampling " + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "content": " from this normal distribution and then sample the intermediate latent variables backward. This is done by using a variational approximation of the probabilities " + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "inline_equation", + "content": "q(x_{t - 1}|x_t)" + }, + { + "bbox": [ + 304, + 111, + 545, + 196 + ], + "type": "text", + "content": " given by the Gaussian factors:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 202, + 545, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 202, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 310, + 202, + 545, + 216 + ], + "type": "interline_equation", + "content": "p \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\sqrt {\\bar {\\alpha} _ {t - 1}} D _ {\\theta} \\left(x _ {t}, t\\right), (1 - \\bar {\\alpha} _ {t - 1}) \\mathbb {I}\\right) \\tag {3}", + "image_path": "2194c43ecff810739b1ce357c92828b44d3d816650828951d0e7c2de04c31a6f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 222, + 504, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 504, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 222, + 504, + 233 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 306, + 222, + 504, + 233 + ], + "type": "text", + "content": " is a neural network with parameters " + }, + { + "bbox": [ + 306, + 222, + 504, + 233 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "type": "text", + "content": "The network " + }, + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 234, + 545, + 270 + ], + "type": "text", + "content": " is trained by maximizing the ELBO (Evidence Lower Bound), which reduces to the denoising objective [22]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 276, + 545, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 276, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 312, + 276, + 545, + 303 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, \\epsilon , x} \\left[ \\frac {\\bar {\\alpha} _ {t - 1}}{2 (1 - \\bar {\\alpha} _ {t - 1}) ^ {2}} \\| D _ {\\theta} \\left(\\hat {\\epsilon} _ {t} (x), t\\right) - x \\| ^ {2} \\right], \\tag {4}", + "image_path": "6dcd5a30933bab1ad2bc48870785e7eac86c7bb196cc0a2ef7ba98ffc4d36db0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "text", + "content": " is sampled uniformly from " + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "inline_equation", + "content": "U[1,T]" + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "text", + "content": ". Hence, " + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "inline_equation", + "content": "D_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "text", + "content": " approximates the clean sample " + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "text", + "content": " given the noisy sample " + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 306, + 308, + 545, + 345 + ], + "type": "text", + "content": " (obtained using (2))." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 351, + 437, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 437, + 362 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 437, + 362 + ], + "type": "text", + "content": "3.2. HoloDiffusion revisited" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 369, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 501 + ], + "type": "text", + "content": "Given a large dataset of 3D models, the framework of Sec. 3.1 could be used to train a corresponding probability distribution. However, such a dataset is not available, and we must instead learn from 2D images of physical 3D objects. Given a dataset containing several views of a large number of objects, we could use image-based reconstruction (e.g., using neural rendering) to obtain corresponding 3D models first, and then use those to train a diffusion model. Instead, we adopt, and slightly upgrade, the HoloDiffusion method [18], which learns a 3D diffusion model directly from the 2D images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "content": "Training data. HoloDiffusion learns from a collection " + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "content": " image sequences " + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "inline_equation", + "content": "s_i = (I_j^i, C_j^i)_{j=1}^{N_{\\mathrm{frame}}}" + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "inline_equation", + "content": "i = 1, \\ldots, N" + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "content": ", where frame " + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "inline_equation", + "content": "I_j^i \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "content": " is an RGB image and " + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "inline_equation", + "content": "C_j^i \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 304, + 503, + 545, + 566 + ], + "type": "text", + "content": " is the corresponding camera projection matrix, collectively defining the motion of the camera." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": "3D representation and rendering. The shape and appearance of the object are represented by a voxel grid " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{d \\times S \\times S \\times S}" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " with resolution " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " containing a " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": "-dimensional feature vector per voxel. Given a 3D point " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": ", its opacity " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathbf{p}) \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "c(\\mathbf{p}) \\in \\mathbb{R}_{[0,1]}^3" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " are obtained from the voxel grid by an MLP " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "M_{\\eta}(V(\\mathbf{p}))" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " that takes as input the " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": "-dimensional feature vector " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "V(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " extracted from the grid via trilinear interpolation [20]. The usual emission-absorption model [26, 24] is then used to implement a differentiable rendering function " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "R_{\\eta}" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": ", mapping the voxel grid " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " and the camera viewpoint " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " into an image " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{I} = R_{\\eta}(V,C)" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 568, + 545, + 714 + ], + "type": "text", + "content": " are the parameters of the MLP." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22978" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 305 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 305 + ], + "type": "image", + "image_path": "8ea85f537f37312e8a7c69fdbeec3bffe7e20f1e7221e40bb31e3cd5b96cc611.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 544, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 544, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 544, + 333 + ], + "type": "text", + "content": "Figure 2: Overview. HOOFUSION, which trains the 3D denoiser network " + }, + { + "bbox": [ + 46, + 308, + 544, + 333 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 308, + 544, + 333 + ], + "type": "text", + "content": ", is augmented with the 2D 'super-resolution' diffusion model " + }, + { + "bbox": [ + 46, + 308, + 544, + 333 + ], + "type": "inline_equation", + "content": "D_{\\beta}" + }, + { + "bbox": [ + 46, + 308, + 544, + 333 + ], + "type": "text", + "content": ". Both models are trained end-to-end by supervising their outputs with 2D photometric error." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "content": "Training scheme. HoloDiffusion leverages the DDPM framework (revised in the previous paragraphs) to recover the density " + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "inline_equation", + "content": "p(V)" + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "content": " over voxel grids " + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "inline_equation", + "content": "x = V" + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "content": " encoding plausible real-life objects. In order to train a DDPM on such 3D data, we would need access to ground-truth 3D models " + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "content": ", which are not available. HoloDiffusion addresses this problem by making three changes to DDPM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": "First, it replaces the data denoising loss with a photometric reconstruction loss. Given a pair " + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "(I,C)\\in s" + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": " from one of the training sequences " + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": ", it replaces Eq. (4) with " + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{t,\\epsilon ,C}\\left[\\| I - R_{\\eta}(D_{\\theta}(\\hat{\\epsilon}_t(V),t),C)\\| ^2\\right]" + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": " where the goal is not to reconstruct the 'clean' volume " + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": " (which is unknown), but rather its image " + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": " (which is known)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": "Second, also because the 'clean' volume " + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": " is not available, we cannot use Eq. (3) to generate the noisy volumes " + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": " to denoise; the only exception is the last sample " + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "inline_equation", + "content": "V_{T}" + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": ", which is pure noise. This suggests to adopt a 'double denoising' step. First, pure noise " + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "inline_equation", + "content": "V_{T}" + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": " is fed into the denoiser to obtain an (approximate) version of " + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "inline_equation", + "content": "V_{0} = D_{\\theta}(V_{T}, T)" + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": " of the clean volume " + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "inline_equation", + "content": "V_{0} = V" + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": ". Then, noise is applied to obtain " + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "inline_equation", + "content": "V_{t} = \\hat{\\epsilon}_{t}(V_{0}) = \\sqrt{\\bar{\\alpha}_{T}} V_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{t}" + }, + { + "bbox": [ + 46, + 510, + 288, + 616 + ], + "type": "text", + "content": " according to Eq. (3), and the latter is fed back into the denoiser as above." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Finally, there is the issue that unconditional generation of the clean volume " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " from pure noise " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "V_{T}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " is difficult, especially in the first iterations of training. On the other hand, the problem of view-conditioned generation is considerably easier. Hence, the third idea is to learn a conditional generator, using a variable number of input views. Specifically, given a training sequence " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", the method extracts a random subset of frames " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\bar{s} \\subset s" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " (which could be empty, which cor-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": "responds to unconditional generation). Then, a feature volume " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\bar{V} = \\Phi (\\bar{s})\\in \\mathbb{R}^{d\\times S\\times S\\times S}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": " is obtained from the selected frames. This extracts 2D image features using a pre-trained and frozen 2D image encoder and then pools them in 3D via 'unprojection' [17, 14] into " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\bar{V}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\bar{V} = 0" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": " is empty. Finally, these pooled features are used to condition the denoising network " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "V_{0} = D_{\\theta}(V_{T},\\bar{V},T)" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": ", which, on average, leads to a simpler reconstruction problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 449, + 514, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 514, + 462 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 514, + 462 + ], + "type": "text", + "content": "Putting it all together, the training loss becomes:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 468, + 545, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 468, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 350, + 468, + 545, + 488 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta | I, C, \\bar {s}) = \\mathbb {E} _ {t, \\hat {\\epsilon}, V _ {T}} \\left[ \\| \\hat {I} - I \\| ^ {2} \\right], \\tag {5}", + "image_path": "444f4dd5deda89ed48a64f208defc48cf89cb9117a4bcf67a1a03127dc22f164.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 364, + 491, + 545, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 491, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 364, + 491, + 545, + 504 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\hat {I} = R _ {\\eta} \\left(D _ {\\theta} \\left(V _ {t}, \\bar {V}, t\\right), C\\right), \\tag {6}", + "image_path": "c7a11e23780689904587ccca6a507513df203cb4e571c278bb1ad7d2ff01bddc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 390, + 506, + 443, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 506, + 443, + 519 + ], + "spans": [ + { + "bbox": [ + 390, + 506, + 443, + 519 + ], + "type": "interline_equation", + "content": "V _ {t} = \\hat {\\epsilon} _ {t} (V _ {0}),", + "image_path": "498b20727bdae1313f054783bf1157f61bf9cc11bf27019f3979c94f5ad26f2e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 390, + 521, + 474, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 521, + 474, + 534 + ], + "spans": [ + { + "bbox": [ + 390, + 521, + 474, + 534 + ], + "type": "interline_equation", + "content": "V _ {0} = D _ {\\theta} (V _ {T}, \\bar {V}, T),", + "image_path": "9829b452e5c4e003633c816b8971927dae02894f4cddb5a1fd67661c01476a87.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 392, + 536, + 438, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 536, + 438, + 549 + ], + "spans": [ + { + "bbox": [ + 392, + 536, + 438, + 549 + ], + "type": "interline_equation", + "content": "\\bar {V} = \\Phi (\\bar {s}).", + "image_path": "96b6c3e526d5be1e7963f7d39991559d330a22bf19734cee615a57d5a12bdff3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": "This loss is averaged over training sequences " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": ", subsequences " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\bar{s} \\subset s" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": ", and views " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "(I, C) \\in s" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": " therein. Note that this is slightly different than the original HoloDiffusion, where feature volume " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\tilde{V}" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": " and reconstructed volumes " + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 304, + 556, + 545, + 640 + ], + "type": "text", + "content": " overlap as arguments of the denoiser; we found that keeping them separated in the formulation leads to more stable training and additionally allows for view-conditioned generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 647, + 388, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 388, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 388, + 658 + ], + "type": "text", + "content": "3.3. HOOFUSION" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The method of Sec. 3.2 learns to generate 3D objects from 2D image supervision only, but the fidelity of the output is limited by the resolution at which the operations are carried out. Increasing resolution is difficult due to the GPU" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22979" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "memory impact of the voxel-based representation, so we seek a more efficient way to do so. The idea is to incorporate a 2D super-resolution network (Sec. 3.3.1), trained end-to-end, that improves the output from the base model. The super-resolved images are eventually fused back in an improved 3D model, which also has the benefit of further increasing view consistency (Sec. 3.3.2)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 201, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 201, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 201, + 186 + ], + "type": "text", + "content": "3.3.1 Integrating super-resolution" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "text", + "content": "As shown in Fig. 2, we augment the method of Sec. 3.2 with a lightweight refinement post-processor network that takes the 2D image " + }, + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "text", + "content": " generated by the base model and outputs a higher quality version " + }, + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{super}}" + }, + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "text", + "content": " of the same. This can be thought of as a form of super-resolution; however, due to the particular statistics of the input ('low-res') images " + }, + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 46, + 194, + 287, + 301 + ], + "type": "text", + "content": " that HoloDiffusion generates, it is necessary to train this superresolution network in an end-to-end fashion with HoloDiffusion, integrating the two models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "text", + "content": "To make this integration seamless, we formulate super-resolution as another diffusion process that runs 'in parallel' with 3D reconstruction. Hence, the super-resolved image " + }, + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{super}} = D_{\\beta}(I_t,\\hat{I},t)" + }, + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "text", + "content": " is the output of a denoiser network (a lightweight U-Net), which takes as input the noised target image " + }, + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "inline_equation", + "content": "I_{t} = \\hat{\\epsilon}_{t}(I)" + }, + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "text", + "content": " and is also conditioned on the 'low-res' output " + }, + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\hat{I} = R_{\\eta}(V,C)" + }, + { + "bbox": [ + 46, + 302, + 287, + 397 + ], + "type": "text", + "content": " of HoloDiffusion from Eq. (6). This denoiser is trained with the DDPM loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 404, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 404, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 83, + 404, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\beta | I) = \\mathbb {E} _ {t, \\hat {\\epsilon}} \\left[ \\| D _ {\\beta} \\left(\\hat {\\epsilon} _ {t} (I), \\hat {I}, t\\right) - I \\| ^ {2} \\right]. \\tag {7}", + "image_path": "7c9fda3a9276991dbfdf0b549b921608e6f1517973e38ff36dd326c6518e5236.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "text", + "content": "Training details The overall model " + }, + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "inline_equation", + "content": "(D_{\\beta}" + }, + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "inline_equation", + "content": "D_{\\theta})" + }, + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "text", + "content": " is trained end-to-end by optimising the loss " + }, + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta |I,C,\\bar{s}) + \\mathcal{L}(\\beta |I)" + }, + { + "bbox": [ + 47, + 435, + 287, + 472 + ], + "type": "text", + "content": " obtained by summing Eqs. (5) and (7)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": "As training data, we use a large dataset of images capturing object-centric scenes ([30]). In each training batch, we pick a random training scene " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": " and sample 15 different source images " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\bar{s}_{\\mathrm{src}} \\subset s" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": " which are unprojected to generate the feature volume conditioning " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\bar{V}" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\bar{V}" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": " is rendered into 4 random target views " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\bar{s}_{\\mathrm{tgt}} \\subset (s \\setminus \\bar{s}_{\\mathrm{src}})" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": " which allows to optimize the training image reconstruction loss " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta | I, C, \\bar{s}) + \\mathcal{L}(\\beta | I)" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": ". The latter uses the Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "5 \\cdot 10^{-5}" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": " decaying tenfold whenever the loss plateaus until convergence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 226, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 226, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 226, + 621 + ], + "type": "text", + "content": "3.3.2 Fusing super-resolved views in 3D" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "The method of Sec. 3.3.1 leaves us with high-resolution views " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{super}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " of the generated 3D object. However, we would like to obtain a single, high-quality 3D model, not just individual views of it. In this section, we discuss how to take the super-resolved images and fuse them into such a model, while addressing the issues that these images are not perfectly view-consistent." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 213 + ], + "type": "image", + "image_path": "0ecabef4deb273ec11a28889a588e4e8b4c8c8e637b4fbc031a3307b76289271.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": "Figure 3: Distillation. HOOFUSION distills a single high-resolution voxel grid " + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "inline_equation", + "content": "V_0^H" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": " by minimizing a top-k patch-remix loss " + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{distil}}" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": " between the grid renders " + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "inline_equation", + "content": "R_{\\eta'}(V_0^H, C)" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": " and a bank " + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_C" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": " high-res images output by the 2D diffusion upsampler " + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "inline_equation", + "content": "D_{\\beta}" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": " for each scene camera " + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 216, + 546, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "content": "The basic idea is simple. We can generate a certain number (e.g., 100) high-resolution images of the object from different viewpoints " + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "content": " and then use a technique, akin to neural rendering, to fuse them back into a single 3D model. However, there is a problem with this idea: The model of Sec. 3.3.1 generates high-quality views " + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{super}}" + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "content": ", but these are view-dependent samples from the distribution " + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "inline_equation", + "content": "p(\\hat{I}_{\\mathrm{super}}|\\hat{I})" + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "inline_equation", + "content": "\\hat{I} = R_{\\eta}(V,C)" + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "content": " is the 'low-res' output form HoloDiffusion. Because super-resolving details is intrinsically ambiguous, there is no reason why samples " + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{super}}" + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "content": " taken from different viewpoints " + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 299, + 546, + 457 + ], + "type": "text", + "content": " would be consistent (Fig. 7). Fusing them into a single 3D model would then result in a blurry appearance yet again." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "text", + "content": "As described in Fig. 3, we address this issue in a principled manner by considering several possible super-resolved images " + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_C = \\{I_{\\mathrm{super}}\\sim p(\\hat{I}_{\\mathrm{super}}|\\hat{I})\\}" + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "text", + "content": " sampled from each given viewpoint " + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "text", + "content": ". Then, we optimize a high-resolution voxel grid " + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "inline_equation", + "content": "V_0^H" + }, + { + "bbox": [ + 304, + 458, + 545, + 519 + ], + "type": "text", + "content": " by minimizing the photometric loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 529, + 547, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 547, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l}} \\left(\\eta^ {\\prime}, V _ {0} ^ {H} \\mid \\mathcal {I} _ {C}\\right) = \\mathbb {E} _ {C} \\left[ \\min _ {I _ {\\text {s u p e r}} \\in \\mathcal {I} _ {C}} \\| I _ {\\text {s u p e r}} - R _ {\\eta^ {\\prime}} \\left(V _ {0} ^ {H}, C\\right) \\| ^ {2} \\right] \\tag {8}", + "image_path": "8b5d2d3e4cc8cd517a19d2799df115cd2a0f55e5bc46e91f62bcc19f9122aa2c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "R_{\\eta'}(V_0^H, C)" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": " is the render of a high-resolution voxel grid " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "V_0^H \\in \\mathbb{R}^{d \\times S' \\times S' \\times S'}" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "S' > S" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": " using the learnable renderer " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "R_{\\eta'}" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": " with scene specific parameters " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\eta'" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": ". Minimizing with respect to " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{super}}" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": " means that the 3D model must be consistent with at least one of the possible super-resolved images, drawn from the distribution of super-resolved samples, for each view " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "Patch remix. In practice, this approach requires a very large number of super resolved images " + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_C" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": " to be effective. We found that we can significantly improve the statistical efficiency by performing the minimization at the level of individual patches. Namely, we produce a stack of only" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22980" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 282, + 268 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 282, + 268 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 282, + 268 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 282, + 268 + ], + "type": "image", + "image_path": "6caeb0cb9e20918eeba8abf09ce78107c7727018f0f599c8f0526df050d2022e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "type": "text", + "content": "Figure 4: Generated 3D samples visualized from a moving camera. " + }, + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "type": "text", + "content": "-GAN and HoloDiffusion* fail to produce 3D view consistent samples, while DreamFusion suffers from the \"Janus\" problem (multiple heads)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "type": "inline_equation", + "content": "K = |\\mathcal{I}_C| = 5" + }, + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "type": "text", + "content": " super resolved images and perform the minimization in Eq. (8) at the level of small " + }, + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "type": "text", + "content": " patches independently (effectively allowing super-resolved images to 'remix' as needed to fit the generated view " + }, + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "type": "inline_equation", + "content": "R_{\\eta'}(V_0^H, C)" + }, + { + "bbox": [ + 46, + 359, + 287, + 408 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": "Distillation details. " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{distil}}" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " is optimized independently for each generated scene with Adam " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "(1\\mathrm{r} = 2\\cdot 10^{-4})" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "25\\mathrm{K}" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " steps until convergence. While " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\eta^\\prime" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " is initialized using the pretrained multi-sequence weights " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "V_0^H" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " is initialized by tri-linearly upsampling the low-resolution volume " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " output by HoloDiffusion. Cameras " + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 418, + 287, + 503 + ], + "type": "text", + "content": " are sampled at uniform azimuths with elevation fixed at object's equator." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 543, + 284, + 657 + ], + "blocks": [ + { + "bbox": [ + 52, + 543, + 284, + 657 + ], + "lines": [ + { + "bbox": [ + 52, + 543, + 284, + 657 + ], + "spans": [ + { + "bbox": [ + 52, + 543, + 284, + 657 + ], + "type": "image", + "image_path": "19d0793b669aa1114c512b9643516e5b124e74034a0bb4f066efa75cf2b01315.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 662, + 286, + 709 + ], + "lines": [ + { + "bbox": [ + 46, + 662, + 286, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 286, + 709 + ], + "type": "text", + "content": "Figure 5: Fusing views. Our patch-remix (Sec. 3.3.2) compared to the SDS and MSE distillation. MSE has \"floaters\" and viewpoint inconsistencies, SDS over-smooths the texture. Ours is robust and produces superior quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 92, + 545, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 92, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 305, + 92, + 545, + 127 + ], + "type": "text", + "content": "We begin with a description of the experiments conducted in Sec. 4.1, followed by an analysis and discussion of the results in Sec. 4.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 138, + 361, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 138, + 361, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 361, + 150 + ], + "type": "text", + "content": "4.1. Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 158, + 545, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 242 + ], + "type": "text", + "content": "Dataset. We experiment on the challenging large-scale Co3Dv2 [30] dataset which is a popular choice for a real-world 3D reconstruction benchmark. More specifically, 4 categories are selected, Apple, Hydrant, TeddyBear, and Donut, with 500 3D-scenes per category for training. Each 3D scene contains " + }, + { + "bbox": [ + 304, + 158, + 545, + 242 + ], + "type": "inline_equation", + "content": "\\sim 200" + }, + { + "bbox": [ + 304, + 158, + 545, + 242 + ], + "type": "text", + "content": " images of the object of interest along with poses of their corresponding cameras." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 245, + 546, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 546, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 546, + 413 + ], + "type": "text", + "content": "Baselines. We use two sets of baselines for comparison (Tab. 1): (i) general 3D generative modeling baselines and (ii) diffusion distillation based baselines. " + }, + { + "bbox": [ + 304, + 245, + 546, + 413 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 245, + 546, + 413 + ], + "type": "text", + "content": "-GAN [5], EG3D [4], GET3D [9], and HoloDiffusion [18] are considered as the 3D generative baselines. Along with HoloDiffusion, we also test the super-resolution integrated model (described in Sec. 3.3.1) HoloDiffusion*. For the distillation-based baselines, we consider the open-source implementation of DreamFusion [29] titled Stable-DreamFusion [40]. For the latter, scenes are generated by conditioning on prompts comprising names of Co3Dv2 categories extended with color and style modifier phrases leading to " + }, + { + "bbox": [ + 304, + 245, + 546, + 413 + ], + "type": "inline_equation", + "content": "\\sim 200" + }, + { + "bbox": [ + 304, + 245, + 546, + 413 + ], + "type": "text", + "content": " prompts / 3D shapes per class. More details regarding the prompt creation are in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 453 + ], + "type": "text", + "content": "Metrics. We use FID [15] and KID [3] to compare the quality of our 2D renders, as these are commonly used to assess 2D and 3D generators." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 463, + 500, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 500, + 476 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 500, + 476 + ], + "type": "text", + "content": "4.2. Quantitative and qualitative analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 483, + 544, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 544, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 544, + 506 + ], + "type": "text", + "content": "Tab. 1 evaluates quantitatively while Fig. 6 qualitatively. Furthermore, Fig. 4 compares rendering view-consistency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "text", + "content": "HoloFusion (Ours) yields better FID/KID scores than the general 3D generative baselines except for " + }, + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "text", + "content": "-GAN on Apple and Donut classes. However, since " + }, + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 508, + 545, + 591 + ], + "type": "text", + "content": "-GAN does not guarantee view consistency by design, it essentially acts as a 2D image GAN, and thus does better on the 2D FID/KID metrics, but it generates significantly view-inconsistent renders (see Fig. 4 and the supplementary)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 593, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 652 + ], + "type": "text", + "content": "We observed that the other 3D-GAN baselines, EG3D and GET3D, are prone to collapsing to a single adversarial sample leading to poor FID/KID scores. The latter is probably due to the 3D misalignment of the CO3Dv2 sequences across instances, which makes training harder." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "HoloFusion also outperforms the text-to-3D StableDreamFusion on both FID/KID. Stable-DreamFusion yields good shapes, but produces synthetic-looking and overly-smooth textures and thus performs poorly when compared to the real-world images of Co3Dv2. As evi" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22981" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 86, + 545, + 657 + ], + "blocks": [ + { + "bbox": [ + 47, + 122, + 61, + 644 + ], + "lines": [ + { + "bbox": [ + 47, + 122, + 61, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 61, + 644 + ], + "type": "text", + "content": "HoloFusion HoloDiffusion* HoloDiffusion DreamFusion GET-3D EG3D " + }, + { + "bbox": [ + 47, + 122, + 61, + 644 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 122, + 61, + 644 + ], + "type": "text", + "content": " -GAN" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 86, + 545, + 657 + ], + "lines": [ + { + "bbox": [ + 65, + 86, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 65, + 86, + 545, + 657 + ], + "type": "image", + "image_path": "382b412cae39adb21545ba1a92e4e4c3d05a0f609a7303a75bfd6f8c55058c48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 660, + 546, + 685 + ], + "lines": [ + { + "bbox": [ + 46, + 660, + 546, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 660, + 546, + 685 + ], + "type": "text", + "content": "Figure 6: 3D samples generated by our HoloFusion compared to " + }, + { + "bbox": [ + 46, + 660, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 660, + 546, + 685 + ], + "type": "text", + "content": "-GAN, EG3D, GET3D, HoloDiffusion, HoloDiffusion*, and the text-to-3D Stable-DreamFusion." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22982" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 110, + 545, + 280 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 106 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 106 + ], + "type": "text", + "content": "Table 1: FID (↓) and KID (↓) on 4 classes of Co3Dv2 [30]. We compare with 3D generative modeling baselines (rows 1–5); with an SDS distillation-based Stable-DreamFusion (row 6); and with ablations of our HoloFusion (rows 7–8). The column “VP” denotes whether renders of a method are 3D view-consistent or not." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 110, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 47, + 110, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 545, + 280 + ], + "type": "table", + "html": "
methodVPAppleHydrantTeddyBearDonutMean
FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
π-GAN [5]X49.30.04292.10.080125.80.11899.40.06991.70.077
EG3D [4]✓170.50.203229.50.253236.10.239222.30.237214.60.233
GET3D [9]✓179.10.190303.30.380244.50.280209.90.230234.20.270
HoloDiffusion [18]✓94.50.095100.50.079109.20.106115.40.085122.50.102
HoloDiffusion*X55.90.04562.60.045116.60.10199.60.07983.70.068
Stable-DreamFusion [40]✓139.00.104185.20.132183.40.125169.30.114169.20.119
HoloFusion (MSE)X72.70.06762.20.04587.20.076109.00.09982.80.072
HoloFusion (SDS)✓123.00.10577.10.058117.80.090142.80.087115.20.085
HoloFusion (Ours)✓69.20.06366.80.04787.60.075109.70.09883.30.071
", + "image_path": "01622b1493a2fcdab1ea6af1fba0ba6b2dc77e7da0dc32898a839812f3e5b3e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 298, + 282, + 456 + ], + "blocks": [ + { + "bbox": [ + 53, + 298, + 282, + 456 + ], + "lines": [ + { + "bbox": [ + 53, + 298, + 282, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 282, + 456 + ], + "type": "image", + "image_path": "17b4ed1e7f688b0fa3b2eb10816ca8efd1afcc85b08419cb2e334b250096c7b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 460, + 288, + 534 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 534 + ], + "type": "text", + "content": "Figure 7: Heatmaps illustrating the per-pixel color variance of " + }, + { + "bbox": [ + 46, + 460, + 288, + 534 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 46, + 460, + 288, + 534 + ], + "type": "text", + "content": " hypothesis produced by the upsampler " + }, + { + "bbox": [ + 46, + 460, + 288, + 534 + ], + "type": "inline_equation", + "content": "D_{\\beta}" + }, + { + "bbox": [ + 46, + 460, + 288, + 534 + ], + "type": "text", + "content": ". Some samples contain artifacts around the object boundaries which correspond to the high-variance regions in the figure. Our top-K patch-remix increases robustness by allowing the loss to discard such artifacts during distillation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 555, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 577 + ], + "type": "text", + "content": "dent from the TeddyBear samples, the method also suffers from the \"Janus\" issue." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 578, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 664 + ], + "type": "text", + "content": "Compared to HoloDiffusion, we improve the FID/KID scores by a significant margin, mainly due to the more photo-realistic renders that include high-frequency details. Although the 2D Diffusion upsampler of HoloDiffusion* produces renders with the highest amount of details yielding scores similar to ours, they are not 3D view-consistent (as apparent from Fig. 4 and as explained in Sec. 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 689 + ], + "type": "text", + "content": "Ablations. In Tab. 1 and in Fig. 5 we ablate components of our HoloFusion to verify their contribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The first variant, HoloFusion (SDS), replaces the Topk patch-remixed distillation loss with the score distillation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 298, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 334 + ], + "type": "text", + "content": "sampling (SDS) gradient as proposed in [29]. As apparent from Fig. 5 and from the lower scores, SDS washes out all the high-frequency details in the textures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 335, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 407 + ], + "type": "text", + "content": "Secondly, HoloFusion (MSE) reduces the number of upsampling hypotheses " + }, + { + "bbox": [ + 305, + 335, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 335, + 545, + 407 + ], + "type": "text", + "content": " to the minimum of " + }, + { + "bbox": [ + 305, + 335, + 545, + 407 + ], + "type": "inline_equation", + "content": "|\\mathcal{I}| = 1" + }, + { + "bbox": [ + 305, + 335, + 545, + 407 + ], + "type": "text", + "content": ". Even though this slightly improves the 2D metrics, as can be seen from Fig. 5, the samples lack view-consistency and introduce \"floaters\". In Fig. 7 we further illustrate the variability of the upsampling hypotheses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 419, + 379, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 379, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 379, + 431 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 440, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 536 + ], + "type": "text", + "content": "We have presented a hybrid diffusion-based method that can generate high-quality 3D neural radiance fields of real-life object categories. Our method starts by producing coarse 3D models whose renders are independently superresolved, and finally consolidated using a robust distillation process. We evaluated our method on the Co3D v2 dataset and presented 3D-consistent, diverse, and high-quality results superior to all competing baselines." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 536, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 657 + ], + "type": "text", + "content": "Our method suffers from limitations that can be addressed in future work. First, our method is slow to sample from as the sampling process takes about 30 mins for each generation, because it is still a distillation-based method. An interesting extension would be to train another network to directly distill a set of super-resolved images, without requiring explicit optimization during inference. Second, we do not produce an explicit surface representation (e.g., a mesh), which could be done by integrating a differentiable mesh render in the loop as done in some prior work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 669, + 421, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 421, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 421, + 682 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "Animesh and Niloy were partially funded by the European Union's Horizon 2020 research and innovation pro" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22983" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": "gramme under the Marie Skłodowska-Curie grant agreement No. 956585. This research has been partly supported by MetaAI and the UCL AI Centre. Finally Animesh is grateful to The Rabin Ezra Scholarship Fund being a recipient of their esteemed fellowship for the year 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 144, + 107, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 144, + 107, + 156 + ], + "spans": [ + { + "bbox": [ + 48, + 144, + 107, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 163, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 287, + 208 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv.cs, abs/2211.09869, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 209, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 274 + ], + "type": "text", + "content": "[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. arXiv.cs, abs/2207.13751, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "type": "text", + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 310, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 288, + 385 + ], + "type": "text", + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 388, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 443 + ], + "type": "text", + "content": "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 288, + 498 + ], + "type": "text", + "content": "[6] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv preprint arXiv:2212.03267, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 500, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 288, + 533 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 534, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 288, + 566 + ], + "type": "text", + "content": "[8] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3D shape induction from 2D views of multiple objects. In arXiv, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 568, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 568, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 53, + 568, + 288, + 623 + ], + "type": "text", + "content": "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 668 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3d-aware generator for high-resolution image synthesis. CoRR, abs/2110.08985, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[12] Jiatao Gu, Alex Trevithick, Kai-En Lin, Josh Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerfguided distillation from 3d-aware diffusion. arXiv preprint arXiv:2302.10109, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[13] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9984-9993, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 237 + ], + "type": "text", + "content": "[14] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 239, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 547, + 293 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 294, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 547, + 326 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 547, + 360 + ], + "type": "text", + "content": "[17] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 361, + 547, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 547, + 414 + ], + "type": "text", + "content": "[18] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy Mitra. Holodiffusion: Training a 3D diffusion model using 2D images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 415, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 547, + 469 + ], + "type": "text", + "content": "[19] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-resolution text-to-3d content creation. arXiv.cs, abs/2211.10440, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 547, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 547, + 513 + ], + "type": "text", + "content": "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 514, + 547, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 547, + 557 + ], + "type": "text", + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 547, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 547, + 581 + ], + "type": "text", + "content": "[22] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 582, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 547, + 624 + ], + "type": "text", + "content": "[23] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 547, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 647 + ], + "type": "text", + "content": "[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "text", + "content": "[25] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: " + }, + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "text", + "content": " reconstruction of any object from a single image. arXiv e-prints, pages arXiv-2302, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF:" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22984" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[27] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. DiffRF: Rendering-guided 3d radiance field diffusion. In Proc. CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[28] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. arXiv.cs, abs/1904.01326, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 219 + ], + "type": "text", + "content": "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "text", + "content": "[30] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 308 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 375 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 409 + ], + "type": "text", + "content": "[33] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3D-aware image synthesis. arXiv.cs, abs/2007.02442, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. ARXIV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Proc. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 532 + ], + "type": "text", + "content": "[36] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "text", + "content": "[37] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. EpiGRAF: Rethinking training of 3d gans. CoRR, abs/2206.10535, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 621 + ], + "type": "text", + "content": "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "text", + "content": "[39] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 156-174. Springer, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[40] Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stabledreamfusion.2,6,8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 352 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[41] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[42] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv.cs, abs/2210.04628, 2022. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 228 + ], + "type": "text", + "content": "[43] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-10, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "text", + "content": "[44] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 318 + ], + "type": "text", + "content": "[45] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 352 + ], + "type": "text", + "content": "[46] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. arXiv preprint arXiv:2212.00792, 2022. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22985" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_content_list.json b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..42fa8b323e732f163cd9066bba5b887826a5c6da --- /dev/null +++ b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_content_list.json @@ -0,0 +1,1689 @@ +[ + { + "type": "text", + "text": "Homeomorphism Alignment for Unsupervised Domain Adaptation", + "text_level": 1, + "bbox": [ + 148, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lihua Zhou $^{1}$ , Mao Ye $^{1,*}$ , Xiatian Zhu $^{3}$ , Siying Xiao $^{1}$ , Xu-Qian Fan $^{2}$ , Ferrante Neri $^{3}$", + "bbox": [ + 155, + 179, + 812, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Electronic Science and Technology of China", + "bbox": [ + 251, + 198, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Jinan University", + "bbox": [ + 415, + 215, + 553, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Surrey", + "bbox": [ + 398, + 233, + 570, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lihua.zhou@std.uestc.edu.cn, cvlab.uestc@gmail.com, xiatian.zhu@surrey.ac.uk", + "bbox": [ + 166, + 251, + 800, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2018270101016@std.uestc.edu.cn, txqfan@jnu.edu.cn, f.neri@surrey.ac.uk", + "bbox": [ + 187, + 268, + 782, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing unsupervised domain adaptation (UDA) methods rely on aligning the features from the source and target domains explicitly or implicitly in a common space (i.e., the domain invariant space). Explicit distribution matching ignores the discriminability of learned features, while the implicit counterpart such as self-supervised learning suffers from pseudo-label noises. With distribution alignment, it is challenging to acquire a common space which maintains fully the discriminative structure of both domains. In this work, we propose a novel HomeomorphismM Alignment (HMA) approach characterized by aligning the source and target data in two separate spaces. Specifically, an invertible neural network based homeomorphism is constructed. Distribution matching is then used as a sewing up tool for connecting this homeomorphism mapping between the source and target feature spaces. Theoretically, we show that this mapping can preserve the data topological structure (e.g., the cluster/group structure). This property allows for more discriminative model adaptation by leveraging both the original and transformed features of source data in a supervised manner, and those of target domain in an unsupervised manner (e.g., prediction consistency). Extensive experiments demonstrate that our method can achieve the state-of-the-art results. Code is released at https://github.com/buerzlh/HMA.", + "bbox": [ + 75, + 353, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 760, + 209, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning methods rely mostly on a large quantity of manually labeled data [13, 37, 21, 55] which however could be prohibitively expensive or impossible to collect in many scenarios. One effective strategy is to exploit pre-existing labeled data (i.e., the source domain) for a target domain without need for manual labeling. Due to the domain shift", + "bbox": [ + 75, + 786, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenge [35], a model pretrained on a source domain often suffers from drastic performance degradation when directly applied on a target domain. This gives rise to the research attention of Unsupervised Domain Adaptation (UDA).", + "bbox": [ + 500, + 323, + 890, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing UDA methods can be roughly divided into two categories. One is based on distribution alignment [31, 18, 10, 32] which minimizes domain discrepancy by aligning the distributions between two domains. They usually align two different distributions to a single distribution. However, this strategy could distort the original structural information, potentially hurting the final model generalization [5, 11, 44]. The other category is based on self-supervised learning [9, 26, 43] which also learns a single common feature space using pseudo labels or other self-supervision information. Given inevitable noise with self-supervision, it is difficult to obtain a common feature space with discriminative structure well kept. As shown in Fig. 1(a), adapting a model in a common space cannot guarantee better classification performance.", + "bbox": [ + 496, + 383, + 892, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A natural solution for the above problem is to learn a latent distribution transformation function without destroying the original distribution of both domains. There are a few works in this line. For example, CyCADA [14] uses ordinary bijection implemented by two different networks to transform the images of the source domain to the target domain and vice versa. However, its learned two networks are not strictly inverse mappings, making the transformed images not necessarily semantically consistent through the transformation. That is, an image cannot be reconstructed after going through the two learned mappings. In addition, CyCADA learns two networks which is expensive.", + "bbox": [ + 496, + 612, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the above limitations, we leverage a stricter bijection, namely Homeomorphism mapping, a concept borrowed from the topology field [34]. If a bijection satisfies the definition of homeomorphism (i.e., one-to-one correspondence and continuous), theoretically we prove that the data topological structure can be well preserved in the projected space (i.e., the samples in the same cluster are still", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* corresponding author.", + "bbox": [ + 94, + 887, + 218, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18699", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5579a389849140fca7601d675483580f69e2c232a4a5508b7e2094af2cf39ed2.jpg", + "image_caption": [ + "(a) Previous methods" + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 450, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/832384117019492cb04034a06b27f978f03d913af2a8a802a5ba94ad32b2520e.jpg", + "image_caption": [ + "(b) Our method" + ], + "image_footnote": [], + "bbox": [ + 450, + 87, + 883, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d292abfd4c9c31d079be4d7735a0cc518257679fc183b1b605ef5ff392bf7884.jpg", + "image_caption": [ + "Source feature" + ], + "image_footnote": [], + "bbox": [ + 91, + 291, + 145, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3db759acd1eb22e4d614d687c1ea4b4b141c5eec5a18f8f07989a86473050d0f.jpg", + "image_caption": [ + "Transformed source feature", + "Target feature" + ], + "image_footnote": [], + "bbox": [ + 400, + 291, + 452, + 304 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5e8ec6dc831894c9fdefe638e59c2498b32a5983e8e19267b53b3cc500f5db76.jpg", + "image_caption": [ + "Transformed target feature", + "Figure 1. Illustration of previous UDA methods and our homeomorphism alignment. (a) Previous methods align the distributions between two domains in a single common feature space, leading to that per-domain data discriminative structure is not well preserved. To overcome this problem, (b) our method leverages a homeomorphism mapping as a bridge for alignment across the source and target feature spaces, since the homeomorphism mapping can preserve the original data topological structure." + ], + "image_footnote": [], + "bbox": [ + 594, + 308, + 647, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the same projected cluster). By this property, we further introduce a noise-free self-supervised learning task. With such a reversible mapping, semantic consistency is guaranteed. As shown in Fig.1(b), with our homeomorphism mapping, the adapted model works better along with both source and target feature spaces.", + "bbox": [ + 75, + 414, + 468, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above analysis, we propose a novel unsupervised domain adaptation method, called HomeomorphismM Alignment (HMA). Our method consists of three components. The first is to construct a homeomorphism mapping for connecting the source and target feature spaces. We show that an Invertible Neural Network (INN) [20] can be used to derive a pair of mutually invertible functions (i.e., homeomorphism mapping) through its forward and invertible processes. The second is sewing up across the two feature spaces with using distribution matching, as it is required that the cross-space transformed features are aligned with the original features. For discriminative learning, class semantic information is also considered during homeomorphic mapping. The third is to train the model in a self-supervised manner with the labels of source domain and the predictive consistency of unlabeled target domain as supervision. This is noise-free due to the preserved topological structure.", + "bbox": [ + 75, + 506, + 470, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (1) We theoretically prove that homeomorphism mapping can guarantee the topological structure of the mapped data. This is an important property yet ignored in the existing researches. We show that an INN can implement a homeomorphism. (2) We propose a novel UDA method with homeomorphism, the first attempt to consider the UDA problem from the viewpoint of topology by conducting domain alignment", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "across two feature spaces. This design differs from the previous alternative methods typically learning a single common feature space for domain alignment. (3) Extensive experiments demonstrate the superiority of our HMA over prior art alternatives, along with in-depth ablation studies.", + "bbox": [ + 496, + 414, + 892, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 508, + 637, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Domain Adaptation. UDA aims to improve the generalization ability of a model on an unlabeled target domain by leveraging the labeled source domain. Existing methods can be roughly divided into two categories. The first category adopts the idea of distribution alignment that minimizes the source error and the discrepancy between source and target domains concurrently. There exist five main strategies: statistic moment matching, adversarial learning, optimal transport, bi-classifier adversarial learning and adversarial generation.", + "bbox": [ + 496, + 536, + 892, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Statistic moment matching based methods minimize the statistic discrepancy to align the distributions between two domains (e.g., DAN [31], CORAL [42] and CAN [18]). Adversarial learning based methods are inspired by GAN [12], which plays a minimax game between feature extractor and discriminator to learn domain invariant features (e.g., DANN [10], CDAN [32] and ADDA [46]). Optimal transport based methods generally consist of two steps: Finding a coupling matrix for connecting each source sample and target sample, followed by minimizing the cost of these pair-wise connections (e.g., DeepJDOT [7], RWOT [50] and ETD [23]). Bi-classifier adversarial learning based methods play a minimax game with a single feature extractor and two distinct classifiers during adaptation.", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18700", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Commonly, they maximize the prediction discrepancy when training the classifiers and minimize the prediction discrepancy when training the feature extractor (e.g., MCD [39], SWD [22], BCDM [24] and CDAL [56]). Adversarial generation methods combine the domain discriminator and a generator, and generate fake data to align the distributions across domains at the pixel level (e.g., CoGAN [29], SimGAN [41] and CycleGAN [57]). Overall, the first four strategies do feature-level distribution alignment while the last one considers pixel-level alignment. And in this work, we focus on feature-level distribution alignment.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second category of UDA methods consider domain adaptation as a self-supervised learning problem. The key of this strategy is to obtain accurate pseudo-labels or self-supervision information. Similarly, such methods also aim to find a common space to implicitly align the source and target features so that the source and target domain features projected by the feature extractor have better discriminability. For example, SE [9] uses the mean teacher framework with a student and a teacher. For the update of the student, it uses the cross-entropy of source samples and the consistency constraints of target samples. While the teacher is updated by exponential moving average of the student. ATDOC [26] assigns a pseudo-label for each target sample by employing a memory mechanism or neighborhood aggregation. ssUDA [43] performs self-supervised tasks (e.g., rotation, flip and patch location predictions) to improve the model generalization. SHOT [27] adopts the information maximization to solve the source-data free UDA problem. BNM [6] proposes Batch Nuclear-norm Maximization on the output matrix to improve both discriminability and diversity. CDTRANS [51] designs a two-way center-aware labeling algorithm to produce pseudo labels for target samples.", + "bbox": [ + 75, + 261, + 472, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Invertible Neural Network (INN) is a flow-based model that transforms a probability distribution to another by a sequence of invertible and differentiable mappings. This model has been applied in a variety of problems. For example, HCFlow [28] utilizes the hierarchical conditional flow as a unified framework for image super-resolution and image rescaling. NCSR [19] proposes noise conditional flow model for super-resolution for increasing the visual quality and diversity of images. Derived from general volume preserving flows, iVPF [54] solve lossless compression by an exact bijective mapping without any numerical error. iFlow [53] also solves lossless compression by a modular scale transform with numerically invertible flow transformations. DIST [2] a diverse image style transfer framework by enforcing invertible cross-space mapping. Additionally, invertible networks play an important role in protecting privacy, such as invertible de-identification and image hiding [1, 17]. To the best of our knowledge, this is first attempt that leverages INN for unsupervised domain adap", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion.", + "bbox": [ + 500, + 92, + 535, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Analysis of Distribution Alignment", + "text_level": 1, + "bbox": [ + 500, + 119, + 816, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem statements. In UDA, we have a source domain $D_{s} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{n_{s}}$ with $n_s$ labeled samples and a target domain $D_{t} = \\{(x_{i}^{t})\\}_{i = 1}^{n_{t}}$ with $n_t$ unlabeled samples. The two domains share the same label space $\\{1,2,\\dots ,K\\}$ , but in different data distributions. The source model $\\Gamma_{s}$ , pretrained on labeled source data, is composed of a feature extractor $F$ and a classifier $C$ . The goal of UDA is to adapt the source model for the unlabeled target domain.", + "bbox": [ + 498, + 143, + 893, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ideally, given ground-truth labels, the trained model can work well in both the source and target domains. To verify this, we conduct experiments on the Office-31 dataset [38]. Using ground-truth labels of both source and target domains, supervised training is applied. It is found that $100\\%$ accuracy on both domains can be achieved. Next, we will investigate the adapted models trained by varying distribution alignment methods. See Appendix for more algorithm details.", + "bbox": [ + 498, + 266, + 893, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Can previous distribution alignment methods really achieve $100\\%$ accuracy given ground-truth target labels? As mentioned earlier, existing UDA methods usually adapt the source domain model to the target domain through two strategies: distribution alignment and self-supervised learning. Commonly, both tackle domain adaptation by projecting source and target samples into a common space (i.e., a domain invariant space). We conduct an experiment on the top-4 most challenging tasks from Office-31. We use the ground-truth target labels for domain alignment. The results are shown in Table 1. The first row shows the results of CAN [18] which aligns feature distributions based on the ground-truth target labels. The second row shows the results with an adversarial learning method CDAN [32]. The third row gives the performance of an optimal transport method DeepJDOT [7]. Note, the bi-classifier adversarial learning methods are excluded as they are designed to use the predictions of the classifiers but not pseudo-labels, and thus ground-truth labels can not be used. It is observed that these explicit distribution alignment methods cannot achieve $100\\%$ accuracy, meaning that the discriminative data structure is not well preserved during alignment. The fourth row gives the results of a self-supervised learning approach using the ground-truth target sample labels. It works well which means that the feature extractor can find the domain invariant space while keeping data discriminative structure. However, in practice the label noise cannot be eliminated fully.", + "bbox": [ + 498, + 402, + 893, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Can bijection achieve $100\\%$ accuracy? It is difficult to find a feature extractor to obtain a common space while keeping discriminative data structure. We then investigate whether source and target samples can be aligned in two feature spaces in a way of bijection for superior domain", + "bbox": [ + 498, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e2d59fb2731f166e3e992b10d3f8b9a8a94842eb035765a5e67818c389c1a664.jpg", + "table_caption": [ + "Table 1. Up-bound performance probing: Comparing different distribution alignment strategies on Office-31 using the groundtruth target sample labels. SMM: Statistic moment matching; AL: Adversarial learning; OP: Optimal transport; SL: Self-supervised learning; BA:Bijection alignment." + ], + "table_footnote": [], + "table_body": "
ComponentA→DA→WD→AW→A
SMM99.9±0.199.9±0.092.6±0.293.8±0.1
AL99.2±0.199.8±0.190.9±0.292.3±0.1
OP96.2±0.198.8±0.189.9±0.290.9±0.1
SL100.0±0.0100.0±0.0100.0±0.0100.0±0.0
BA97.8±0.298.9±0.190.7±0.193.2±0.2
Ours100.0±0.0100.0±0.0100.0±0.0100.0±0.0
", + "bbox": [ + 88, + 155, + 459, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "adaptation. The fifth row in Table 1 shows the performance of using two different networks to map the source features to the target feature space and vice versa. In this test, we use ground-truth target labels to train these two networks and align the conditional distributions between the transformed features and original features. This bijection made by the two networks is not a homeomorphism. Thus, the data topological structure can not be preserved in the process of bidirectional projections. It is shown that despite access to all ground-truth labels of target domain, this method still cannot achieve $100\\%$ accuracy.", + "bbox": [ + 75, + 273, + 467, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Homeomorphism alignment can achieve $100\\%$ accuracy. A homeomorphism (also known as a continuous transformation) is a one-to-one correspondence mapping between the points in two topological spaces. It is continuous in both directions (refer to [34] for more details). Let $M$ and $N$ be two topological spaces, and $g: M \\to N$ be a bijection. If both the function $g$ and its inverse function $g^{-1}: N \\to M$ are continuous, then $g$ is called a homeomorphism. That being said, a homeomorphism is a bijective correspondence $g: M \\to N$ such that $g(U)$ is open if and only if $U$ is open. By the definition above, $g$ is a homeomorphism if and only if $g^{-1}$ is a homeomorphism. Based on this definition, the following theorem can be derived.", + "bbox": [ + 75, + 441, + 467, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. The set boundary corresponds to the set boundary by homeomorphism. More precisely, let $(M,d_M)$ and $(N,d_N)$ be two metric spaces where $d_M$ and $d_N$ are the metrics on $M,N$ respectively. Suppose there is a homeomorphism $g:M\\to N$ , and $A$ is an open subset in $(M,d_{M})$ , we have that its image $B\\coloneqq g(A)$ is an open subset in $(N,d_{N})$ , and", + "bbox": [ + 76, + 637, + 467, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (\\partial A) = \\partial B = \\partial g (A).\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 753, + 351, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\partial$ means the boundary (see more details in Sec. 6).", + "bbox": [ + 76, + 779, + 455, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Theorem 1, data topological structure is preserved by homeomorphism mapping, i.e., the samples in the same cluster are still in the same projected cluster. In the literature, we find that there exists a network satisfying homeomorphism definition – Invertible Neural Network (INN) [20]. We validate that an INN satisfies the following theorem.", + "bbox": [ + 75, + 795, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2. Invertible Neural Network is a homeomorphism (see more details in Sec. 6).", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in the last row of Table 1, we use an INN to connect two domains at the feature level. With the ground-truth target sample labels, the homeomorphism alignment can achieve $100\\%$ accuracy thanks to the topological structure preserving property.", + "bbox": [ + 496, + 121, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method of Homeomorphism Alignment", + "text_level": 1, + "bbox": [ + 498, + 210, + 854, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview: Based on the analysis of distribution alignment, as shown in Fig. 2, we propose a Homeomorphism Alignment (HMA) method with three components. The first is homeomorphism mapping construction using an INN for connecting the source and target feature spaces. The second is sewing up across the two feature spaces with using distribution matching. The third is training the source model in the source and target feature spaces with a noise-free self-supervised task by using the property of homeomorphism.", + "bbox": [ + 496, + 237, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Homeomorphism implemented by invertible neural network (INN)", + "text_level": 1, + "bbox": [ + 498, + 383, + 890, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each iteration, we randomly sample a batch of source and target samples. We use ResNet [13] as the feature extractor $F$ to map a sample $x$ to the feature space: $f^{s/t} = F(x^{s/t})$ where $s/t$ represents the source and target domain respectively. Due to the distribution discrepancy between the source and target domains, we consider that the source and target features reside on two different spaces (manifolds) respectively.", + "bbox": [ + 496, + 422, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A homeomorphism $g$ consists of $m$ blocks of INN. $m$ is a hyperparameter discussed in Appendix. In each block, we use an affine network to implement the INN [8], as shown in Fig. 2. In the forward process, we transform from the source feature space $\\pmb{f}^s$ to the target feature space $\\pmb{f}^t$ . Specifically, for the $i$ -th block, we denote the input $\\mu_{1:2d}^i$ with $2d$ dimension. And we split evenly $\\mu_{1:2d}^i$ to two parts $[\\mu_{1:d}^i, \\mu_{d+1:2d}^i]$ , and further transform them with two respective linear neural networks $s(\\cdot), t(\\cdot)$ . The output of $i$ -th block $\\mu_{1:2d}^{i+1}$ is then obtained with residual as follows:", + "bbox": [ + 496, + 544, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {1: d} ^ {i + 1} = \\mu_ {1: d} ^ {i} + s (\\mu_ {d + 1: 2 d} ^ {i}), \\\\ \\mu_ {d + 1: 2 d} ^ {i + 1} = \\mu_ {d + 1: 2 d} ^ {i} + t (\\mu_ {1: d} ^ {i + 1}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 708, + 795, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output $\\mu_{1:2d}^{i+1}$ (a concatenation of $\\mu_{1:d}^{i+1}$ and $\\mu_{d+1:2d}^{i+1}$ ) will be set as the input of the next block.", + "bbox": [ + 496, + 758, + 888, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the inverse projection process, we transform from the target feature space $\\pmb{f}^t$ to the source feature space $\\pmb{f}^s$ . For the $i$ -th block, we map $\\mu_{1:2d}^{i+1}$ to $\\mu_{1:2d}^i$ in the opposite away around. According to Eq.(1), we can get the following equation:", + "bbox": [ + 496, + 792, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {d + 1: 2 d} ^ {i} = \\mu_ {d + 1: 2 d} ^ {i + 1} - t (\\mu_ {1: d} ^ {i + 1}), \\\\ \\mu_ {1: d} ^ {i} = \\mu_ {1: d} ^ {i + 1} - s \\left(\\mu_ {d + 1: 2 d} ^ {i}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 864, + 795, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18702", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/783f10a10654f4b9c9113e9af8b0c4eb8dc58b6723f236cffd8573760a2d71d9.jpg", + "image_caption": [ + "Figure 2. The framework of the proposed Homeomorphism Alignment (HMA). (a) We cascade $m$ invertible neural networks to implement a homeomorphism. (b) The transformed features are sewed up with the corresponding feature spaces by category. (c) The source model is iteratively trained in the two spaces concurrently." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We similarly split $\\mu_{1:2d}^{i+1}$ into two parts $[\\mu_{1:d}^{i+1}, \\mu_{d+1:2d}^{i+1}]$ and follow Eq.(2) to get the original input $\\mu_{1:2d}^{i}$ ((concatenation of $\\mu_{1:d}^{i}$ and $\\mu_{d+1:2d}^{i}$ ). Obviously Eqs. (1) and (2) are inverse functions of each other. We denote the forward process of the $m$ INNs as the function $g$ , while the inverse process as $g^{-1}$ . Hence, the function $g$ is a bijection. Since the functions $s(\\cdot)$ , $t(\\cdot)$ are implemented by two linear connected neural networks, they are continuous; This means both $g$ and $g^{-1}$ are continuous, too. According to the definition of homeomorphism, this INN is a homeomorphism.", + "bbox": [ + 75, + 496, + 472, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Sewing up", + "text_level": 1, + "bbox": [ + 76, + 670, + 192, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next we will sew up homeomorphism mapping $g$ to the source and target feature spaces such that the corresponding classes are aligned. Suppose the transformed feature $\\pmb{f}^{s2t} = g(\\pmb{f}^s)$ according to the source feature $\\pmb{f}^s$ and the transformed feature $\\pmb{f}^{t2s} = g^{-1}(\\pmb{f}^t)$ according to target feature $\\pmb{f}^t$ . According to the homeomorphism mapping $g$ , $\\pmb{f}^s = g^{-1}(\\pmb{f}^{s2t})$ and $\\pmb{f}^t = g(\\pmb{f}^{t2s})$ . To guarantee that the transformed features are located discriminatively (i.e., aligned with the corresponding classes), distribution matching is used for sewing up. The loss function is defined as follows,", + "bbox": [ + 75, + 696, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {g} \\operatorname {L o s s} _ {S e w} = D M \\left(\\boldsymbol {f} ^ {s 2 t}, \\boldsymbol {f} ^ {t}\\right) + D M \\left(\\boldsymbol {f} ^ {t 2 s}, \\boldsymbol {f} ^ {s}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 878, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $DM(\\cdot, \\cdot)$ refers to any existing distribution matching method (e.g., DAN [31], CAN [18], DANN [10], CDAN [32], DeepJDOT [7] and MCD [39]). Note, in case of using the adversarial learning method or bi-classifier adversarial learning method, an additional discriminator network or two auxiliary classifiers are needed.", + "bbox": [ + 496, + 498, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that only marginal distribution matching methods (e.g., DAN and DANN) cannot achieve satisfactory results, because they focus on overall distribution alignment instead of class-wise alignment. Although our homeomorphism mapping $g$ can ensure that the transformed features preserve the topological structure, without discriminative sewing up, it is still hard to achieve lossless transformation across domains. Thus class conditional distribution matching is a better choice as validated in ours experiment (Table 2-4).", + "bbox": [ + 496, + 594, + 892, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Model Training", + "text_level": 1, + "bbox": [ + 500, + 751, + 656, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A model often suffers performance degradation from the domain shift. To address this problem, we leverage the homeomorphic property (preserved topological structure) to perform a noise-free self-supervised training. Specifically, as proved by Theorem 1, $f^s$ has the same structural information as $f^{s2t}$ . Concretely, for a specific labeled source sample $x$ , the corresponding feature $f^s$ and $f^{s2t}$ share the same label. The following loss function is applied to the", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "supervised training of feature extractor $F$ and classifier $C$ ,", + "bbox": [ + 76, + 90, + 464, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL o s s _ {S} = \\mathcal {L} ^ {c e} \\left(C \\left(\\boldsymbol {f} ^ {s}\\right), \\boldsymbol {y} ^ {s}\\right) + \\mathcal {L} ^ {c e} \\left(C \\left(\\boldsymbol {f} ^ {s 2 t}\\right), \\boldsymbol {y} ^ {s}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 114, + 468, + 133 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{y}^s$ is the corresponding label of the source sample $x^{s}$ and $\\mathcal{L}^{ce}(\\cdot ,\\cdot)$ denotes the cross entropy function. In particular, the term $\\mathcal{L}^{ce}(C(\\pmb {f}^s),\\pmb {y}^s)$ focuses on the classification of the source domain, whilst $\\mathcal{L}^{ce}(C(\\pmb {f}^{s2t}),\\pmb {y}^s)$ is concerned with the classification of the target domain since $\\pmb{f}^{s2t}$ and $\\pmb{f}^{t}$ have been aligned.", + "bbox": [ + 76, + 143, + 468, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering that our homeomorphism preserves the structure across the mapping and no label information in the target domain, unsupervised consistency constraint is a natural strategy for optimization. Formally, for an unlabeled target sample $x^{t}$ , we formulate the consistency constraint on the predictions between $f^{t}$ and $f^{t2s}$ as:", + "bbox": [ + 75, + 234, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL o s s _ {T} = L _ {C} \\left(C \\left(\\boldsymbol {f} ^ {t}\\right), C \\left(\\boldsymbol {f} ^ {t 2 s}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 334, + 468, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{C}(\\cdot ,\\cdot)$ is a consistency constraint such as $L_{1}$ -Norm and $L_{2}$ -Norm. In practice, we found $L_{2}$ -Norm suffices. By combining (4) and (5), the overall loss is defined as follows,", + "bbox": [ + 76, + 363, + 468, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {F, C} \\operatorname {L o s s} _ {S} + \\operatorname {L o s s} _ {T}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 420, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Summary. At the training phase, in each iteration, we first train an INN based homeomorphism mapping, followed by model training in two spaces concurrently. At the inference phase, both the target features $\\pmb{f}^t$ and the transformed target features $\\pmb{f}^{t2s}$ can be used to make the prediction. Also, average based ensemble can be used to obtain the final prediction.", + "bbox": [ + 75, + 446, + 468, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remarks. Our model is trained in the source and target feature spaces concurrently. Compared with the existing alignment based UDA methods in a common space, this design naturally overcomes the intrinsic challenges of projecting the source and target domain samples into a single shared feature space using a feature extraction network while keeping their respective discriminative structures. When the distributions between the two feature spaces are not originally aligned typical in practice (e.g., due to domain-specific characteristics such as different background, viewing conditions, etc.), the homeomorphism provides a flexible non-invasive means for cross-domain relating via transforming their individual features from each other externally. Critically, this alignment in two spaces allows to fully keep the original per-domain characteristics including some discriminative information. Compared with the self-supervised learning methods suffering the noises of pseudo-labeling, our transformed source features in the target domain can directly use the ground-truth source labels, in addition to additionally exploiting the topological structure of the source domain. Further, our consistency constraint can exploit the unlabeled target training data (i.e., the original and transformed target features with shared topological structure)", + "bbox": [ + 75, + 551, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in an unsupervised manner, without the notorious pseudolabel noise issue.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 136, + 633, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 161, + 689, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: In our experiments, three standard datasets are used. Office-31 [38] is a popular benchmark. It contains a total of 4110 images of 31 office environment objects from 3 domains: Amazon (A), Webcam (W), Dslr(D). Office-Home [47] is a more challenging dataset which contains 15588 images within 65 classes from 4 domains: Artistic images (A), Clip-Art images (C), Product images (P) and RealWorld images (R). Visda-17 [36] is a widely used benchmark for domain adaptation with focus on a 12-class synthesis-to-real object classification task. The source domain contains 152,397 synthetic images and the target domain has 55,388 real object images.", + "bbox": [ + 496, + 185, + 890, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details: Our experiment is performed in Pytorch. Each task is run 5 times to enhance the robustness of the results. The same backbone network is selected as other compared methods for fair comparison. Specifically, Resnet-50 is selected as the backbone on Office-31 and Office-Home, and Resnet-101 is selected on Visda-17. It is worth noting that the output dimension of the classifier in the original backbone is replaced by the number of categories to fit each task. The SGD optimizer is chosen to update the network and the CosineAnnealingLR [33] is used to update the learning rate of the SGD optimizer.", + "bbox": [ + 496, + 367, + 892, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competitors: For extensive evaluation, we compare our model with five groups of state-of-the-art methods. The first is based on distribution alignment, such as statistic moment matching methods DAN [31], CAN [18], and TSA [25]. The second uses adversarial learning including DANN [10], CDAN [32], MDD+IA [16], DADA [45], CLS [30], ILA [40], MetaAlign [48], DWL [49], and DALN [3]. The third exploits optimal transport: DeepJ-DOT [7]. The fourth is based on bi-classifier adversarial learning, MCD [39]. The fifth adopts self-supervised learning, including ALDA [4], ATDOC [26], CaCo [15], and SUDA [52].", + "bbox": [ + 496, + 534, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons to State-of-the-Art", + "text_level": 1, + "bbox": [ + 500, + 724, + 784, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance comparison with the state-of-the-art methods on Office-31, Office-home and Visda-17 are shown in Table 2, Table 3 and Table 4 respectively. The methods HMA(DANN) and HMA(CDAN) mean the sewing up by distribution matching method based on adversarial learning, with DANN and CDAN focusing on marginal distribution alignment and conditional distribution alignment respectively. HMA(DeepHDOT) integrates the optimal transport method DeepJDOT, and HMA(MCD) the bi-classifier adversarial learning method MCD, as the sewing up tool.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18704", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6a11f5a07117492650a3fe890df793e059cdcf5dadb62da569a5130eb59058ac.jpg", + "table_caption": [ + "Table 2. Comparison with the state-of-the-art methods on Office-31 dataset. Metric: classification accuracy (%); Backbone: ResNet-50." + ], + "table_footnote": [], + "table_body": "
MethodVenueA→DA→WD→AD→WW→AW→Davg
ResNet-50 [13]CVPR1668.968.462.596.760.799.376.1
DAN [31]ICML1578.680.563.697.162.899.680.4
CAN [18]CVPR1995.094.578.099.177.099.890.6
TSA [25]CVPR2192.694.874.999.174.4100.089.3
DANN [10]JMLR1679.782.068.296.967.499.182.2
CDAN [32]NIPS1889.893.170.198.268.0100.086.5
DADA [45]AAAI2093.992.374.499.274.2100.089.0
MDD+IA [16]ICML2092.190.375.398.774.999.888.8
ILA [40]CVPR2193.495.772.199.375.4100.089.3
MetaAlign [48]CVPR2194.593.075.098.673.6100.089.2
DWL [49]CVPR2191.289.273.199.269.8100.087.1
DALN [3]CVPR2295.495.276.499.176.5100.090.4
DeepJDOT [7]ECCV1888.288.972.198.570.199.686.2
MCD [39]CVPR1892.288.669.598.569.7100.086.5
ALDA [4]AAAI2094.095.672.297.772.5100.088.7
ATDOC [26]CVPR2194.494.575.698.975.299.689.7
CaCo [15]CVPR2291.789.773.198.472.8100.087.6
SUDA [52]CVPR2291.290.872.298.771.4100.087.4
HMA(DANN)Ours83.9±0.183.5±0.270.5±0.198.2±0.170.1±0.2100.0±0.084.4
HMA(CDAN)Ours92.4±0.295.1±0.273.7±0.199.2±0.172.8±0.2100.0±0.088.9
HMA(DeepJDOT)Ours90.9±0.191.6±0.274.6±0.299.0±0.173.8±0.199.8±0.088.3
HMA(MCD)Ours93.5±0.191.3±0.173.1±0.199.2±0.173.5±0.1100.0±0.088.4
HMA(DAN)Ours85.1±0.284.5±0.267.9±0.398.9±0.266.7±0.3100.0±0.083.9
HMA(CAN)Ours95.8±0.395.1±0.179.3±0.399.3±0.177.6±0.2100.0±0.091.2
", + "bbox": [ + 153, + 101, + 815, + 452 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53001e3e3c0b121bf8977847df80e9139935360db7a7199d6f0305b7f91110be.jpg", + "table_caption": [ + "Table 3. Comparisons with the state-of-the-art methods on Office-Home dataset. Metric: classification accuracy (\\%); Backbone: ResNet-50." + ], + "table_footnote": [], + "table_body": "
MethodVenueA→CA→PA→RC→AC→PC→RP→AP→CP→RR→AR→CR→Pavg
ResNet-50 [13]CVPR1634.950.058.037.441.946.238.531.260.453.941.259.946.1
DAN [31]ICML1543.657.067.945.856.560.444.043.667.763.151.574.356.3
CAN [18]CVPR1958.778.182.167.475.778.167.254.282.573.460.983.571.8
TSA [25]CVPR2153.675.178.364.473.772.562.349.477.572.258.882.168.3
DANN [10]JMLR1645.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [32]NIPS1849.069.374.554.466.068.455.648.375.968.455.480.563.8
MDD+IA [16]ICML2056.277.979.264.473.174.464.254.279.971.258.183.169.5
MetaAlign [48]CVPR2159.376.080.265.774.775.165.756.581.674.161.185.271.3
DALN [3]CVPR2257.879.982.066.376.277.266.755.581.373.560.485.271.8
DeepJDOT [7]ECCV1850.768.674.459.965.868.155.246.373.866.054.978.363.5
MCD [39]CVPR1848.968.374.661.367.668.857.047.175.169.152.279.664.1
ALDAAAAI2053.770.176.460.272.671.556.851.977.170.256.382.166.6
ATDOC [26]CVPR2158.378.882.369.478.278.267.156.082.772.058.285.572.2
HMA(DANN)Ours48.265.175.457.065.068.355.645.273.566.654.378.462.7
HMA(CDAN)Ours58.778.181.667.475.878.166.854.282.573.459.783.571.7
HMA(DeepJDOT)Ours57.574.780.766.472.174.261.252.880.672.861.784.870.0
HMA(MCD)Ours55.574.581.367.674.175.563.653.982.275.358.782.470.4
HMA(DAN)Ours46.263.573.958.165.368.355.343.974.867.253.478.462.4
HMA(CAN)Ours60.679.182.968.977.579.369.155.983.574.662.384.473.2
", + "bbox": [ + 140, + 502, + 828, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While HMA(DAN) and HMA(CAN) apply the statistic moment matching methods as the sewing up tool, where DAN and CAN focus on marginal distribution alignment and conditional distribution alignment respectively.", + "bbox": [ + 76, + 821, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It can be observed that HMA(CAN) yields the best av", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "erage performance on both three datasets. This also confirms our previous analysis. Different sewing up methods will affect the final performance. In general, conditional distribution alignment is better than marginal distribution alignment on the two kinds of methods based on adversar", + "bbox": [ + 498, + 823, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18705", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6175440619d57fa8d54da680eca3cc5d696155bb5a5d39343d0f95b90e1c0382.jpg", + "table_caption": [ + "Table 4. Comparison with the state-of-the-art methods on Visda-17 dataset. Metric: per-class classification accuracy (%); Backbone: ResNet-101." + ], + "table_footnote": [], + "table_body": "
MethodVenueplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckavg
ResNet-101 [13]CVPR1655.153.361.959.180.617.979.731.281.026.573.58.552.4
DAN [31]ICML1584.842.175.453.077.962.686.650.759.752.982.526.062.9
CAN [18]CVPR1997.087.282.574.397.896.290.880.796.696.387.559.987.2
TSA [25]CVPR21------------78.6
DANN [10]JMLR1681.977.782.844.381.229.565.128.651.954.682.87.857.4
CDAN [32]NIPS1885.266.983.050.884.274.988.174.583.476.081.938.073.9
DWL [49]CVPR2190.780.286.167.692.481.586.878.090.657.185.628.777.1
CLS [30]ICCV2192.684.573.772.788.583.389.177.689.589.285.872.781.6
DALN [3]CVPR22------------80.6
DeepJDOT [7]ECCV1885.473.477.387.384.164.791.579.391.944.488.561.877.4
MCD [39]CVPR1887.060.983.764.088.979.684.776.988.640.383.025.871.9
ALDA [4]AAAI2093.874.182.469.490.687.289.067.693.476.187.722.277.8
ATDOC [26]CVPR2193.783.076.958.789.795.184.471.489.480.086.755.180.3
CaCo [15]CVPR2290.480.778.857.088.987.081.379.488.788.186.863.980.9
SUDA [52]CVPR2288.379.366.264.787.480.185.978.386.387.578.874.579.8
HMA(DANN)Ours86.979.183.550.586.747.386.155.164.659.884.636.268.4
HMA(CDAN)Ours88.371.285.166.486.379.388.887.683.979.383.446.278.8
HMA(DeepJDOT)Ours89.376.880.668.387.975.690.686.289.673.582.249.879.2
HMA(MCD)Ours89.072.285.570.390.486.687.683.892.255.589.142.178.7
HMA(DAN)Ours87.549.280.253.881.871.887.857.660.957.085.332.867.1
HMA(CAN)Ours97.688.484.376.098.497.191.381.497.096.788.860.788.1
", + "bbox": [ + 151, + 112, + 818, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8995bb3bb0dd3c4fc6965fb16221c59437656540da8b6d5292ae3aa5c18ddef9.jpg", + "table_caption": [ + "Table 5. Homeomorphism mapping vs ordinary bijection on Office-31. OB(CAN) means ordinary bijection sewed by the distribution matching method CAN." + ], + "table_footnote": [], + "table_body": "
ComponentA→DA→WD→AW→AParameters
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.220992000
OB(CAN)90.3±0.391.7±0.276.6±0.175.8±0.233603584
", + "bbox": [ + 78, + 507, + 488, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ial learning strategy and statistic moment matching strategy because conditional distribution alignment can stitch homeomorphism mapping with two spaces by category. Furthermore, for condition distribution matching method, statistic moment matching strategy is better than adversarial learning strategy. The reason is that statistic moment matching strategy stitches homeomorphism mapping with two feature spaces explicitly by category. Interestingly, the alignment method CAN published in 2019 still achieves SOTA results. It can be seen that alignment by category is very important for extracting domain invariant features. Our method HMA(CAN) further boosts the CAN performance, because we realize the difficulty of extracting a domain invariant space and we do alignment in the two spaces by homeomorphism mapping.", + "bbox": [ + 75, + 593, + 468, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Analysis and Discussion", + "text_level": 1, + "bbox": [ + 76, + 830, + 369, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Homeomorphism map is better than ordinary bijection. As mentioned in Section 3, there does not exist many methods which do alignments in two feature spaces based on bi-", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9d39cf39792a097cd69fbc601f85c2ceb684f6d86b6fcef7460fbafcfa18120b.jpg", + "table_caption": [ + "Table 6. Ablation study on Office-31." + ], + "table_footnote": [], + "table_body": "
ComponentA→DA→WD→AW→A
CAN95.0±0.394.5±0.378.0±0.377.0±0.3
INN(CAN)94.8±0.394.1±0.277.3±0.476.7±0.2
INN(CAN)+S2T95.6±0.294.9±0.378.9±0.277.4±0.2
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.2
DAN78.6±0.280.5±0.463.6±0.362.8±0.2
INN(DAN)78.3±0.379.9±0.262.9±0.262.6±0.1
INN(DAN)+S2T84.1±0.283.8±0.366.5±0.266.1±0.2
INN(DAN)+S2T+T2S85.1±0.284.5±0.267.9±0.366.7±0.3
", + "bbox": [ + 500, + 479, + 903, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "jection. In this experiment, we will apply the ordinary bijection method to top-4 challenging tasks on Office-31 dataset, the difference between homeomorphism implemented by INN and ordinary bijection is mainly reflected in topological structure maintenance. The results are shown in Table 5. It is obvious that homeomorphism mapping is superior to the ordinary bijection in both accuracy and model parameters. For the model size, the INN based homeomorphism mapping consumes around half of the parameters compared to using two neural networks.", + "bbox": [ + 496, + 656, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. To show the effectiveness of alignment in two spaces, we conduct an experiment on top-4 challenging tasks of Office-31 dataset. The results are shown in Table 6. The method CAN is considered as the baseline. INN(CAN) means just sewing up homeomorphism mapping to the two feature spaces and the source model is retrained based on", + "bbox": [ + 496, + 809, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18706", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the source labels. INN(CAN) $^+$ S2T means the transformed source features are used to learn the model in the target feature space compared with INN(CAN). HMA(CAN) uses all features in two spaces. From Table 6, we can find that simply using INN can achieve similar performance as the feature distribution alignment method CAN. As shown in the third row in Table 6, by transferring the source features to the target feature space, the performance is greatly improved. For this case, the data topological structure and label information from source domain can be correctly transformed to the target domain, which allows that the learned model works well in the target domain. The fourth row in Table 6 shows that the performance can be further improved if both transformed features are used. Because our homeomorphism mapping keeps the corresponding relationship by category, with the help of supervision information in the source feature spaces, the generalization performance of the model in the two domains is improved. To further verify our ideas, we test DAN as a sewing tool, giving the consistent conclusion.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Proof of Theorem", + "text_level": 1, + "bbox": [ + 76, + 409, + 250, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.1. Proof of Theorem 1", + "text_level": 1, + "bbox": [ + 76, + 433, + 261, + 448 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Theorem 1. Let $(M, d_M)$ and $(N, d_N)$ be two metric spaces with a homeomorphism", + "bbox": [ + 76, + 455, + 468, + 488 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ng: M \\to N,\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 501, + 318, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and $A$ is an open subset in $(M,d_M)$ , we have that its image $B\\coloneqq g(A)$ is an open subset in $(N,d_N)$ , and", + "bbox": [ + 76, + 529, + 468, + 561 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ng (\\partial A) = \\partial B = \\partial g (A),\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 573, + 354, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\partial$ means the boundary.", + "bbox": [ + 76, + 603, + 272, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Proof. It is sufficient to show that", + "bbox": [ + 76, + 618, + 303, + 632 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ng (\\partial A) \\subset \\partial B. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 646, + 468, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Assume this is true, then we can imply Eq. 7 to $g^{-1}$ , and obtain", + "bbox": [ + 76, + 676, + 468, + 704 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ng ^ {- 1} (\\partial B) \\subset \\partial A.\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 705, + 330, + 723 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hence we have $\\partial B\\subset g(\\partial A)$ . Combing this with Eq. 7, we have $g(\\partial A) = \\partial B$", + "bbox": [ + 76, + 731, + 468, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Now we want to show Eq. 7, that is, for any $x \\in \\partial A$ , we have $g(x) \\in \\partial B$ . Since $x \\in \\partial A$ , but $x \\notin A$ , then $g(x) \\notin B$ , and there is a sequence $\\{x_i\\} \\subset A$ such that $\\lim_{i \\to \\infty} x_i = x$ . By the continuity of the function $g$ , we have", + "bbox": [ + 76, + 762, + 468, + 823 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ng (x) = g \\left(\\lim _ {i \\rightarrow \\infty} x _ {i}\\right) = \\lim _ {i \\rightarrow \\infty} g \\left(x _ {i}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 835, + 383, + 858 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Noting that $g(x_{i}) \\in B$ , we get $g(x) \\in \\partial B$ . This completes the proof.", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.2. Proof of Theorem 2", + "text_level": 1, + "bbox": [ + 500, + 90, + 687, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Theorem 2. Invertible Neural Network is a homeomorphism.", + "bbox": [ + 498, + 114, + 890, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Proof. The definition of homeomorphism is that, a function $g: M \\to N$ between two topological spaces is a homeomorphism if it has the following properties: (1) $g$ is a bijection; (2) $g$ is continuous; (3) The inverse function $g^{-1}$ is continuous.", + "bbox": [ + 498, + 143, + 890, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For any invertible neural network, assuming that its forward process is $g$ , then its invertible process can be represented as $g^{-1}$ , so $g$ is a bijection. Because the function of each part of the invertible neural network is continuous, as $s(\\cdot)$ and $t(\\cdot)$ in our method, so both $g$ and $g^{-1}$ are continuous. To sum up, the invertible network is a homeomorphism.", + "bbox": [ + 498, + 220, + 890, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 325, + 617, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have proposed a new unsupervised domain adaptation method, termed as Homeomorphism Alignment in two spaces (HMA). By analyzing previous alignment based methods, we argue that it is difficult to find a common space or domain invariant space to adapt the pretrained source model. So the alignment is performed in two spaces. The extracted source and target features can be further transformed respectively by a homeomorphism mapping so that they can be aligned semantically. Our method consists of three steps, i.e., constructing an INN based homeomorphism mapping, sewing up by category and retraining iteratively training the model in two spaces. In this way, the source labels can be fully used even in the target feature space for improving the model generalization for the target domain. Extensive experimental results demonstrate the effectiveness of our method.", + "bbox": [ + 498, + 349, + 890, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 606, + 669, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (2018YFE0203900), National Natural Science Foundation of China (62276048).", + "bbox": [ + 500, + 632, + 890, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 691, + 596, + 707 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3334–3342, October 2021. 3", + "[2] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. *Diverse image style transfer via invertible cross-space mapping.* In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, October 2021. 3", + "[3] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific clas" + ], + "bbox": [ + 506, + 715, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18707", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sifier as a discriminator: Discriminator-free adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7181-7190, 2022. 6, 7, 8", + "[4] Minghao Chen, Shuai Zhao, Haifeng Liu, and Deng Cai. Adversarial-learned loss for domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3521–3528, 2020. 6, 7, 8", + "[5] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 1", + "[6] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 3", + "[7] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 2, 3, 5, 6, 7, 8", + "[8] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 4", + "[9] Geoff French, Michal Mackiewicz, and Mark Fisher. Self-ensembleing for visual domain adaptation. In International Conference on Learning Representations, 2018. 1, 3", + "[10] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 2, 5, 6, 7, 8", + "[11] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 1", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, 2014. 2", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 4, 7, 8", + "[14] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1", + "[15] Jiaxing Huang, Dayan Guan, Aoran Xiao, Shijian Lu, and Ling Shao. Category contrast for unsupervised domain adaptation in visual tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1203-1214, 2022. 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In International Conference on Machine Learning, pages 4816-4827. PMLR, 2020. 6, 7", + "[17] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4733-4742, October 2021. 3", + "[18] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4893-4902, 2019. 1, 2, 3, 5, 6, 7, 8", + "[19] Younggeun Kim and Donghee Son. Noise conditional flow model for learning the super-resolution space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 424-432, 2021. 3", + "[20] Diederik P Kingma and Prafulla Dhariwal. Glow: generative flow with invertible $1 \\times 1$ convolutions. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 10236-10245, 2018. 2, 4", + "[21] Sajja Tulasi Krishna and Hemantha Kumar Kalluri. Deep learning and transfer learning approaches for image classification. International Journal of Recent Technology and Engineering (IJRTE), 7(5S4):427-432, 2019. 1", + "[22] Chen-Yu Lee, Tanmay Batra, Mohammad Haris Baig, and Daniel Ulbricht. Sliced Wasserstein discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10285-10295, 2019. 3", + "[23] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13936–13944, 2020. 2", + "[24] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8455-8464, 2021. 3", + "[25] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11516-11525, 2021. 6, 7, 8", + "[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Domain adaptation with auxiliary target domain-oriented classifier. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2021. 1, 3, 6, 7, 8", + "[27] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 3", + "[28] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchi" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18708", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cal conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4076-4085, 2021. 3", + "[29] Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 469-477, 2016. 3", + "[30] Xiaofeng Liu, Zhenhua Guo, Site Li, Fangxu Xing, Jane You, C.-C. Jay Kuo, Georges El Fakhri, and Jonghye Woo. Adversarial unsupervised domain adaptation with conditional and label shift: Infer, align and iterate. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10367-10376, October 2021. 6, 8", + "[31] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 1, 2, 5, 6, 7, 8", + "[32] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in neural information processing systems, 2018. 1, 2, 3, 5, 6, 7, 8", + "[33] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[34] James R. Munkres. Topology (2nd edition). In Prentice Hall, 2000. 1, 4", + "[35] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, pages 1345-1359, 2009. 1", + "[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.visda06924, 2017. 6", + "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 1", + "[38] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. European conference on computer vision, pages 213-226. Springer, 2010. 3, 6", + "[39] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 3, 5, 6, 7, 8", + "[40] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5361-5371, 2021. 6, 7", + "[41] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2107-2116, 2017. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. In Domain Adaptation in Computer Vision Applications, pages 153-171. Springer, 2017. 2", + "[43] Yu Sun, Eric Tzeng, Trevor Darrell, and Alexei A Efros. Unsupervised domain adaptation through self-supervision. arXiv preprint arXiv:1909.11825, 2019. 1, 3", + "[44] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 1", + "[45] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 6, 7", + "[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 2", + "[47] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 6", + "[48] Guoqiang Wei, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Metaalign: Coordinating domain alignment and classification for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16643-16653, 2021. 6, 7", + "[49] Ni Xiao and Lei Zhang. Dynamic weighted learning for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15242-15251, 2021. 6, 7, 8", + "[50] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4394-4403, 2020. 2", + "[51] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3", + "[52] Jingyi Zhang, Jiaxing Huang, Zichen Tian, and Shijian Lu. Spectral unsupervised domain adaptation for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9829-9840, 2022. 6, 7, 8", + "[53] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 3", + "[54] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18709", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Bo Zhao, Jiashi Feng, Xiao Wu, and Shuicheng Yan. A survey on deep learning-based fine-grained object classification and semantic segmentation. International Journal of Automation and Computing, 14(2):119-135, 2017. 1", + "[56] Lihua Zhou, Mao Ye, Xiatian Zhu, Shuaifeng Li, and Yiguang Liu. Class discriminative adversarial learning for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4318-4326, 2022. 3", + "[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + ], + "bbox": [ + 78, + 90, + 470, + 287 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18710", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_model.json b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9c1960fbad32816e7df1a539f32749e384aadac --- /dev/null +++ b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_model.json @@ -0,0 +1,2358 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "Homeomorphism Alignment for Unsupervised Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.18, + 0.813, + 0.2 + ], + "angle": 0, + "content": "Lihua Zhou\\(^{1}\\), Mao Ye\\(^{1,*}\\), Xiatian Zhu\\(^{3}\\), Siying Xiao\\(^{1}\\), Xu-Qian Fan\\(^{2}\\), Ferrante Neri\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.199, + 0.721, + 0.217 + ], + "angle": 0, + "content": "1University of Electronic Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.217, + 0.555, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)Jinan University" + }, + { + "type": "text", + "bbox": [ + 0.4, + 0.234, + 0.571, + 0.252 + ], + "angle": 0, + "content": "3University of Surrey" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.252, + 0.802, + 0.269 + ], + "angle": 0, + "content": "lihua.zhou@std.uestc.edu.cn, cvlab.uestc@gmail.com, xiatian.zhu@surrey.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.27, + 0.783, + 0.288 + ], + "angle": 0, + "content": "2018270101016@std.uestc.edu.cn, txqfan@jnu.edu.cn, f.neri@surrey.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Existing unsupervised domain adaptation (UDA) methods rely on aligning the features from the source and target domains explicitly or implicitly in a common space (i.e., the domain invariant space). Explicit distribution matching ignores the discriminability of learned features, while the implicit counterpart such as self-supervised learning suffers from pseudo-label noises. With distribution alignment, it is challenging to acquire a common space which maintains fully the discriminative structure of both domains. In this work, we propose a novel HomeomorphismM Alignment (HMA) approach characterized by aligning the source and target data in two separate spaces. Specifically, an invertible neural network based homeomorphism is constructed. Distribution matching is then used as a sewing up tool for connecting this homeomorphism mapping between the source and target feature spaces. Theoretically, we show that this mapping can preserve the data topological structure (e.g., the cluster/group structure). This property allows for more discriminative model adaptation by leveraging both the original and transformed features of source data in a supervised manner, and those of target domain in an unsupervised manner (e.g., prediction consistency). Extensive experiments demonstrate that our method can achieve the state-of-the-art results. Code is released at https://github.com/buerzlh/HMA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.761, + 0.21, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Deep learning methods rely mostly on a large quantity of manually labeled data [13, 37, 21, 55] which however could be prohibitively expensive or impossible to collect in many scenarios. One effective strategy is to exploit pre-existing labeled data (i.e., the source domain) for a target domain without need for manual labeling. Due to the domain shift" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.892, + 0.383 + ], + "angle": 0, + "content": "challenge [35], a model pretrained on a source domain often suffers from drastic performance degradation when directly applied on a target domain. This gives rise to the research attention of Unsupervised Domain Adaptation (UDA)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.612 + ], + "angle": 0, + "content": "Existing UDA methods can be roughly divided into two categories. One is based on distribution alignment [31, 18, 10, 32] which minimizes domain discrepancy by aligning the distributions between two domains. They usually align two different distributions to a single distribution. However, this strategy could distort the original structural information, potentially hurting the final model generalization [5, 11, 44]. The other category is based on self-supervised learning [9, 26, 43] which also learns a single common feature space using pseudo labels or other self-supervision information. Given inevitable noise with self-supervision, it is difficult to obtain a common feature space with discriminative structure well kept. As shown in Fig. 1(a), adapting a model in a common space cannot guarantee better classification performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.794 + ], + "angle": 0, + "content": "A natural solution for the above problem is to learn a latent distribution transformation function without destroying the original distribution of both domains. There are a few works in this line. For example, CyCADA [14] uses ordinary bijection implemented by two different networks to transform the images of the source domain to the target domain and vice versa. However, its learned two networks are not strictly inverse mappings, making the transformed images not necessarily semantically consistent through the transformation. That is, an image cannot be reconstructed after going through the two learned mappings. In addition, CyCADA learns two networks which is expensive." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To address the above limitations, we leverage a stricter bijection, namely Homeomorphism mapping, a concept borrowed from the topology field [34]. If a bijection satisfies the definition of homeomorphism (i.e., one-to-one correspondence and continuous), theoretically we prove that the data topological structure can be well preserved in the projected space (i.e., the samples in the same cluster are still" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.22, + 0.901 + ], + "angle": 0, + "content": "* corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18699" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.452, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.266, + 0.301, + 0.28 + ], + "angle": 0, + "content": "(a) Previous methods" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.088, + 0.884, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.266, + 0.761, + 0.28 + ], + "angle": 0, + "content": "(b) Our method" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.292, + 0.147, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.292, + 0.271, + 0.305 + ], + "angle": 0, + "content": "Source feature" + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.31, + 0.364, + 0.324 + ], + "angle": 0, + "content": "Transformed source feature" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.292, + 0.453, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.292, + 0.578, + 0.306 + ], + "angle": 0, + "content": "Target feature" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.309, + 0.648, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.31, + 0.865, + 0.325 + ], + "angle": 0, + "content": "Transformed target feature" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.334, + 0.894, + 0.391 + ], + "angle": 0, + "content": "Figure 1. Illustration of previous UDA methods and our homeomorphism alignment. (a) Previous methods align the distributions between two domains in a single common feature space, leading to that per-domain data discriminative structure is not well preserved. To overcome this problem, (b) our method leverages a homeomorphism mapping as a bridge for alignment across the source and target feature spaces, since the homeomorphism mapping can preserve the original data topological structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.505 + ], + "angle": 0, + "content": "in the same projected cluster). By this property, we further introduce a noise-free self-supervised learning task. With such a reversible mapping, semantic consistency is guaranteed. As shown in Fig.1(b), with our homeomorphism mapping, the adapted model works better along with both source and target feature spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Motivated by the above analysis, we propose a novel unsupervised domain adaptation method, called HomeomorphismM Alignment (HMA). Our method consists of three components. The first is to construct a homeomorphism mapping for connecting the source and target feature spaces. We show that an Invertible Neural Network (INN) [20] can be used to derive a pair of mutually invertible functions (i.e., homeomorphism mapping) through its forward and invertible processes. The second is sewing up across the two feature spaces with using distribution matching, as it is required that the cross-space transformed features are aligned with the original features. For discriminative learning, class semantic information is also considered during homeomorphic mapping. The third is to train the model in a self-supervised manner with the labels of source domain and the predictive consistency of unlabeled target domain as supervision. This is noise-free due to the preserved topological structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (1) We theoretically prove that homeomorphism mapping can guarantee the topological structure of the mapped data. This is an important property yet ignored in the existing researches. We show that an INN can implement a homeomorphism. (2) We propose a novel UDA method with homeomorphism, the first attempt to consider the UDA problem from the viewpoint of topology by conducting domain alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.893, + 0.492 + ], + "angle": 0, + "content": "across two feature spaces. This design differs from the previous alternative methods typically learning a single common feature space for domain alignment. (3) Extensive experiments demonstrate the superiority of our HMA over prior art alternatives, along with in-depth ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.638, + 0.525 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.893, + 0.688 + ], + "angle": 0, + "content": "Unsupervised Domain Adaptation. UDA aims to improve the generalization ability of a model on an unlabeled target domain by leveraging the labeled source domain. Existing methods can be roughly divided into two categories. The first category adopts the idea of distribution alignment that minimizes the source error and the discrepancy between source and target domains concurrently. There exist five main strategies: statistic moment matching, adversarial learning, optimal transport, bi-classifier adversarial learning and adversarial generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Statistic moment matching based methods minimize the statistic discrepancy to align the distributions between two domains (e.g., DAN [31], CORAL [42] and CAN [18]). Adversarial learning based methods are inspired by GAN [12], which plays a minimax game between feature extractor and discriminator to learn domain invariant features (e.g., DANN [10], CDAN [32] and ADDA [46]). Optimal transport based methods generally consist of two steps: Finding a coupling matrix for connecting each source sample and target sample, followed by minimizing the cost of these pair-wise connections (e.g., DeepJDOT [7], RWOT [50] and ETD [23]). Bi-classifier adversarial learning based methods play a minimax game with a single feature extractor and two distinct classifiers during adaptation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18700" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Commonly, they maximize the prediction discrepancy when training the classifiers and minimize the prediction discrepancy when training the feature extractor (e.g., MCD [39], SWD [22], BCDM [24] and CDAL [56]). Adversarial generation methods combine the domain discriminator and a generator, and generate fake data to align the distributions across domains at the pixel level (e.g., CoGAN [29], SimGAN [41] and CycleGAN [57]). Overall, the first four strategies do feature-level distribution alignment while the last one considers pixel-level alignment. And in this work, we focus on feature-level distribution alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.61 + ], + "angle": 0, + "content": "The second category of UDA methods consider domain adaptation as a self-supervised learning problem. The key of this strategy is to obtain accurate pseudo-labels or self-supervision information. Similarly, such methods also aim to find a common space to implicitly align the source and target features so that the source and target domain features projected by the feature extractor have better discriminability. For example, SE [9] uses the mean teacher framework with a student and a teacher. For the update of the student, it uses the cross-entropy of source samples and the consistency constraints of target samples. While the teacher is updated by exponential moving average of the student. ATDOC [26] assigns a pseudo-label for each target sample by employing a memory mechanism or neighborhood aggregation. ssUDA [43] performs self-supervised tasks (e.g., rotation, flip and patch location predictions) to improve the model generalization. SHOT [27] adopts the information maximization to solve the source-data free UDA problem. BNM [6] proposes Batch Nuclear-norm Maximization on the output matrix to improve both discriminability and diversity. CDTRANS [51] designs a two-way center-aware labeling algorithm to produce pseudo labels for target samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Invertible Neural Network (INN) is a flow-based model that transforms a probability distribution to another by a sequence of invertible and differentiable mappings. This model has been applied in a variety of problems. For example, HCFlow [28] utilizes the hierarchical conditional flow as a unified framework for image super-resolution and image rescaling. NCSR [19] proposes noise conditional flow model for super-resolution for increasing the visual quality and diversity of images. Derived from general volume preserving flows, iVPF [54] solve lossless compression by an exact bijective mapping without any numerical error. iFlow [53] also solves lossless compression by a modular scale transform with numerically invertible flow transformations. DIST [2] a diverse image style transfer framework by enforcing invertible cross-space mapping. Additionally, invertible networks play an important role in protecting privacy, such as invertible de-identification and image hiding [1, 17]. To the best of our knowledge, this is first attempt that leverages INN for unsupervised domain adap" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.093, + 0.536, + 0.105 + ], + "angle": 0, + "content": "tion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.12, + 0.818, + 0.138 + ], + "angle": 0, + "content": "3. Analysis of Distribution Alignment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.145, + 0.895, + 0.267 + ], + "angle": 0, + "content": "Problem statements. In UDA, we have a source domain \\( D_{s} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{n_{s}} \\) with \\( n_s \\) labeled samples and a target domain \\( D_{t} = \\{(x_{i}^{t})\\}_{i = 1}^{n_{t}} \\) with \\( n_t \\) unlabeled samples. The two domains share the same label space \\( \\{1,2,\\dots ,K\\} \\), but in different data distributions. The source model \\( \\Gamma_{s} \\), pretrained on labeled source data, is composed of a feature extractor \\( F \\) and a classifier \\( C \\). The goal of UDA is to adapt the source model for the unlabeled target domain." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.267, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Ideally, given ground-truth labels, the trained model can work well in both the source and target domains. To verify this, we conduct experiments on the Office-31 dataset [38]. Using ground-truth labels of both source and target domains, supervised training is applied. It is found that \\(100\\%\\) accuracy on both domains can be achieved. Next, we will investigate the adapted models trained by varying distribution alignment methods. See Appendix for more algorithm details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Can previous distribution alignment methods really achieve \\(100\\%\\) accuracy given ground-truth target labels? As mentioned earlier, existing UDA methods usually adapt the source domain model to the target domain through two strategies: distribution alignment and self-supervised learning. Commonly, both tackle domain adaptation by projecting source and target samples into a common space (i.e., a domain invariant space). We conduct an experiment on the top-4 most challenging tasks from Office-31. We use the ground-truth target labels for domain alignment. The results are shown in Table 1. The first row shows the results of CAN [18] which aligns feature distributions based on the ground-truth target labels. The second row shows the results with an adversarial learning method CDAN [32]. The third row gives the performance of an optimal transport method DeepJDOT [7]. Note, the bi-classifier adversarial learning methods are excluded as they are designed to use the predictions of the classifiers but not pseudo-labels, and thus ground-truth labels can not be used. It is observed that these explicit distribution alignment methods cannot achieve \\(100\\%\\) accuracy, meaning that the discriminative data structure is not well preserved during alignment. The fourth row gives the results of a self-supervised learning approach using the ground-truth target sample labels. It works well which means that the feature extractor can find the domain invariant space while keeping data discriminative structure. However, in practice the label noise cannot be eliminated fully." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Can bijection achieve \\(100\\%\\) accuracy? It is difficult to find a feature extractor to obtain a common space while keeping discriminative data structure. We then investigate whether source and target samples can be aligned in two feature spaces in a way of bijection for superior domain" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18701" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.156 + ], + "angle": 0, + "content": "Table 1. Up-bound performance probing: Comparing different distribution alignment strategies on Office-31 using the groundtruth target sample labels. SMM: Statistic moment matching; AL: Adversarial learning; OP: Optimal transport; SL: Self-supervised learning; BA:Bijection alignment." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.156, + 0.46, + 0.257 + ], + "angle": 0, + "content": "
ComponentA→DA→WD→AW→A
SMM99.9±0.199.9±0.092.6±0.293.8±0.1
AL99.2±0.199.8±0.190.9±0.292.3±0.1
OP96.2±0.198.8±0.189.9±0.290.9±0.1
SL100.0±0.0100.0±0.0100.0±0.0100.0±0.0
BA97.8±0.298.9±0.190.7±0.193.2±0.2
Ours100.0±0.0100.0±0.0100.0±0.0100.0±0.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.468, + 0.44 + ], + "angle": 0, + "content": "adaptation. The fifth row in Table 1 shows the performance of using two different networks to map the source features to the target feature space and vice versa. In this test, we use ground-truth target labels to train these two networks and align the conditional distributions between the transformed features and original features. This bijection made by the two networks is not a homeomorphism. Thus, the data topological structure can not be preserved in the process of bidirectional projections. It is shown that despite access to all ground-truth labels of target domain, this method still cannot achieve \\(100\\%\\) accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.468, + 0.637 + ], + "angle": 0, + "content": "Homeomorphism alignment can achieve \\(100\\%\\) accuracy. A homeomorphism (also known as a continuous transformation) is a one-to-one correspondence mapping between the points in two topological spaces. It is continuous in both directions (refer to [34] for more details). Let \\(M\\) and \\(N\\) be two topological spaces, and \\(g: M \\to N\\) be a bijection. If both the function \\(g\\) and its inverse function \\(g^{-1}: N \\to M\\) are continuous, then \\(g\\) is called a homeomorphism. That being said, a homeomorphism is a bijective correspondence \\(g: M \\to N\\) such that \\(g(U)\\) is open if and only if \\(U\\) is open. By the definition above, \\(g\\) is a homeomorphism if and only if \\(g^{-1}\\) is a homeomorphism. Based on this definition, the following theorem can be derived." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.638, + 0.468, + 0.743 + ], + "angle": 0, + "content": "Theorem 1. The set boundary corresponds to the set boundary by homeomorphism. More precisely, let \\((M,d_M)\\) and \\((N,d_N)\\) be two metric spaces where \\(d_M\\) and \\(d_N\\) are the metrics on \\(M,N\\) respectively. Suppose there is a homeomorphism \\(g:M\\to N\\), and \\(A\\) is an open subset in \\((M,d_{M})\\), we have that its image \\(B\\coloneqq g(A)\\) is an open subset in \\((N,d_{N})\\), and" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.754, + 0.352, + 0.77 + ], + "angle": 0, + "content": "\\[\ng (\\partial A) = \\partial B = \\partial g (A).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.456, + 0.795 + ], + "angle": 0, + "content": "where \\(\\partial\\) means the boundary (see more details in Sec. 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.9 + ], + "angle": 0, + "content": "As shown in Theorem 1, data topological structure is preserved by homeomorphism mapping, i.e., the samples in the same cluster are still in the same projected cluster. In the literature, we find that there exists a network satisfying homeomorphism definition – Invertible Neural Network (INN) [20]. We validate that an INN satisfies the following theorem." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "Theorem 2. Invertible Neural Network is a homeomorphism (see more details in Sec. 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.198 + ], + "angle": 0, + "content": "As shown in the last row of Table 1, we use an INN to connect two domains at the feature level. With the ground-truth target sample labels, the homeomorphism alignment can achieve \\(100\\%\\) accuracy thanks to the topological structure preserving property." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.212, + 0.855, + 0.23 + ], + "angle": 0, + "content": "4. Method of Homeomorphism Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Overview: Based on the analysis of distribution alignment, as shown in Fig. 2, we propose a Homeomorphism Alignment (HMA) method with three components. The first is homeomorphism mapping construction using an INN for connecting the source and target feature spaces. The second is sewing up across the two feature spaces with using distribution matching. The third is training the source model in the source and target feature spaces with a noise-free self-supervised task by using the property of homeomorphism." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.385, + 0.891, + 0.415 + ], + "angle": 0, + "content": "4.1. Homeomorphism implemented by invertible neural network (INN)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.544 + ], + "angle": 0, + "content": "In each iteration, we randomly sample a batch of source and target samples. We use ResNet [13] as the feature extractor \\( F \\) to map a sample \\( x \\) to the feature space: \\( f^{s/t} = F(x^{s/t}) \\) where \\( s/t \\) represents the source and target domain respectively. Due to the distribution discrepancy between the source and target domains, we consider that the source and target features reside on two different spaces (manifolds) respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.698 + ], + "angle": 0, + "content": "A homeomorphism \\( g \\) consists of \\( m \\) blocks of INN. \\( m \\) is a hyperparameter discussed in Appendix. In each block, we use an affine network to implement the INN [8], as shown in Fig. 2. In the forward process, we transform from the source feature space \\( \\pmb{f}^s \\) to the target feature space \\( \\pmb{f}^t \\). Specifically, for the \\( i \\)-th block, we denote the input \\( \\mu_{1:2d}^i \\) with \\( 2d \\) dimension. And we split evenly \\( \\mu_{1:2d}^i \\) to two parts \\( [\\mu_{1:d}^i, \\mu_{d+1:2d}^i] \\), and further transform them with two respective linear neural networks \\( s(\\cdot), t(\\cdot) \\). The output of \\( i \\)-th block \\( \\mu_{1:2d}^{i+1} \\) is then obtained with residual as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.709, + 0.796, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {1: d} ^ {i + 1} = \\mu_ {1: d} ^ {i} + s (\\mu_ {d + 1: 2 d} ^ {i}), \\\\ \\mu_ {d + 1: 2 d} ^ {i + 1} = \\mu_ {d + 1: 2 d} ^ {i} + t (\\mu_ {1: d} ^ {i + 1}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.89, + 0.791 + ], + "angle": 0, + "content": "The output \\(\\mu_{1:2d}^{i+1}\\) (a concatenation of \\(\\mu_{1:d}^{i+1}\\) and \\(\\mu_{d+1:2d}^{i+1}\\)) will be set as the input of the next block." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.892, + 0.865 + ], + "angle": 0, + "content": "In the inverse projection process, we transform from the target feature space \\( \\pmb{f}^t \\) to the source feature space \\( \\pmb{f}^s \\). For the \\( i \\)-th block, we map \\( \\mu_{1:2d}^{i+1} \\) to \\( \\mu_{1:2d}^i \\) in the opposite away around. According to Eq.(1), we can get the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.866, + 0.796, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {d + 1: 2 d} ^ {i} = \\mu_ {d + 1: 2 d} ^ {i + 1} - t (\\mu_ {1: d} ^ {i + 1}), \\\\ \\mu_ {1: d} ^ {i} = \\mu_ {1: d} ^ {i + 1} - s \\left(\\mu_ {d + 1: 2 d} ^ {i}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18702" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.892, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.894, + 0.476 + ], + "angle": 0, + "content": "Figure 2. The framework of the proposed Homeomorphism Alignment (HMA). (a) We cascade \\( m \\) invertible neural networks to implement a homeomorphism. (b) The transformed features are sewed up with the corresponding feature spaces by category. (c) The source model is iteratively trained in the two spaces concurrently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.473, + 0.653 + ], + "angle": 0, + "content": "We similarly split \\(\\mu_{1:2d}^{i+1}\\) into two parts \\([\\mu_{1:d}^{i+1}, \\mu_{d+1:2d}^{i+1}]\\) and follow Eq.(2) to get the original input \\(\\mu_{1:2d}^{i}\\) ((concatenation of \\(\\mu_{1:d}^{i}\\) and \\(\\mu_{d+1:2d}^{i}\\)). Obviously Eqs. (1) and (2) are inverse functions of each other. We denote the forward process of the \\(m\\) INNs as the function \\(g\\), while the inverse process as \\(g^{-1}\\). Hence, the function \\(g\\) is a bijection. Since the functions \\(s(\\cdot)\\), \\(t(\\cdot)\\) are implemented by two linear connected neural networks, they are continuous; This means both \\(g\\) and \\(g^{-1}\\) are continuous, too. According to the definition of homeomorphism, this INN is a homeomorphism." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.194, + 0.687 + ], + "angle": 0, + "content": "4.2. Sewing up" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.863 + ], + "angle": 0, + "content": "Next we will sew up homeomorphism mapping \\( g \\) to the source and target feature spaces such that the corresponding classes are aligned. Suppose the transformed feature \\( \\pmb{f}^{s2t} = g(\\pmb{f}^s) \\) according to the source feature \\( \\pmb{f}^s \\) and the transformed feature \\( \\pmb{f}^{t2s} = g^{-1}(\\pmb{f}^t) \\) according to target feature \\( \\pmb{f}^t \\). According to the homeomorphism mapping \\( g \\), \\( \\pmb{f}^s = g^{-1}(\\pmb{f}^{s2t}) \\) and \\( \\pmb{f}^t = g(\\pmb{f}^{t2s}) \\). To guarantee that the transformed features are located discriminatively (i.e., aligned with the corresponding classes), distribution matching is used for sewing up. The loss function is defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.88, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\min _ {g} \\operatorname {L o s s} _ {S e w} = D M \\left(\\boldsymbol {f} ^ {s 2 t}, \\boldsymbol {f} ^ {t}\\right) + D M \\left(\\boldsymbol {f} ^ {t 2 s}, \\boldsymbol {f} ^ {s}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.59 + ], + "angle": 0, + "content": "where \\(DM(\\cdot, \\cdot)\\) refers to any existing distribution matching method (e.g., DAN [31], CAN [18], DANN [10], CDAN [32], DeepJDOT [7] and MCD [39]). Note, in case of using the adversarial learning method or bi-classifier adversarial learning method, an additional discriminator network or two auxiliary classifiers are needed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.893, + 0.73 + ], + "angle": 0, + "content": "Recall that only marginal distribution matching methods (e.g., DAN and DANN) cannot achieve satisfactory results, because they focus on overall distribution alignment instead of class-wise alignment. Although our homeomorphism mapping \\( g \\) can ensure that the transformed features preserve the topological structure, without discriminative sewing up, it is still hard to achieve lossless transformation across domains. Thus class conditional distribution matching is a better choice as validated in ours experiment (Table 2-4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.658, + 0.769 + ], + "angle": 0, + "content": "4.3. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A model often suffers performance degradation from the domain shift. To address this problem, we leverage the homeomorphic property (preserved topological structure) to perform a noise-free self-supervised training. Specifically, as proved by Theorem 1, \\( f^s \\) has the same structural information as \\( f^{s2t} \\). Concretely, for a specific labeled source sample \\( x \\), the corresponding feature \\( f^s \\) and \\( f^{s2t} \\) share the same label. The following loss function is applied to the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18703" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.465, + 0.108 + ], + "angle": 0, + "content": "supervised training of feature extractor \\(F\\) and classifier \\(C\\)," + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.116, + 0.47, + 0.135 + ], + "angle": 0, + "content": "\\[\nL o s s _ {S} = \\mathcal {L} ^ {c e} \\left(C \\left(\\boldsymbol {f} ^ {s}\\right), \\boldsymbol {y} ^ {s}\\right) + \\mathcal {L} ^ {c e} \\left(C \\left(\\boldsymbol {f} ^ {s 2 t}\\right), \\boldsymbol {y} ^ {s}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.144, + 0.469, + 0.234 + ], + "angle": 0, + "content": "where \\(\\pmb{y}^s\\) is the corresponding label of the source sample \\(x^{s}\\) and \\(\\mathcal{L}^{ce}(\\cdot ,\\cdot)\\) denotes the cross entropy function. In particular, the term \\(\\mathcal{L}^{ce}(C(\\pmb {f}^s),\\pmb {y}^s)\\) focuses on the classification of the source domain, whilst \\(\\mathcal{L}^{ce}(C(\\pmb {f}^{s2t}),\\pmb {y}^s)\\) is concerned with the classification of the target domain since \\(\\pmb{f}^{s2t}\\) and \\(\\pmb{f}^{t}\\) have been aligned." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.469, + 0.326 + ], + "angle": 0, + "content": "Considering that our homeomorphism preserves the structure across the mapping and no label information in the target domain, unsupervised consistency constraint is a natural strategy for optimization. Formally, for an unlabeled target sample \\( x^{t} \\), we formulate the consistency constraint on the predictions between \\( f^{t} \\) and \\( f^{t2s} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.335, + 0.469, + 0.354 + ], + "angle": 0, + "content": "\\[\nL o s s _ {T} = L _ {C} \\left(C \\left(\\boldsymbol {f} ^ {t}\\right), C \\left(\\boldsymbol {f} ^ {t 2 s}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.364, + 0.469, + 0.409 + ], + "angle": 0, + "content": "where \\(L_{C}(\\cdot ,\\cdot)\\) is a consistency constraint such as \\(L_{1}\\) -Norm and \\(L_{2}\\) -Norm. In practice, we found \\(L_{2}\\) -Norm suffices. By combining (4) and (5), the overall loss is defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.421, + 0.469, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\min _ {F, C} \\operatorname {L o s s} _ {S} + \\operatorname {L o s s} _ {T}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.469, + 0.551 + ], + "angle": 0, + "content": "Summary. At the training phase, in each iteration, we first train an INN based homeomorphism mapping, followed by model training in two spaces concurrently. At the inference phase, both the target features \\( \\pmb{f}^t \\) and the transformed target features \\( \\pmb{f}^{t2s} \\) can be used to make the prediction. Also, average based ensemble can be used to obtain the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Remarks. Our model is trained in the source and target feature spaces concurrently. Compared with the existing alignment based UDA methods in a common space, this design naturally overcomes the intrinsic challenges of projecting the source and target domain samples into a single shared feature space using a feature extraction network while keeping their respective discriminative structures. When the distributions between the two feature spaces are not originally aligned typical in practice (e.g., due to domain-specific characteristics such as different background, viewing conditions, etc.), the homeomorphism provides a flexible non-invasive means for cross-domain relating via transforming their individual features from each other externally. Critically, this alignment in two spaces allows to fully keep the original per-domain characteristics including some discriminative information. Compared with the self-supervised learning methods suffering the noises of pseudo-labeling, our transformed source features in the target domain can directly use the ground-truth source labels, in addition to additionally exploiting the topological structure of the source domain. Further, our consistency constraint can exploit the unlabeled target training data (i.e., the original and transformed target features with shared topological structure)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "in an unsupervised manner, without the notorious pseudolabel noise issue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.137, + 0.634, + 0.154 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.162, + 0.691, + 0.179 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.367 + ], + "angle": 0, + "content": "Datasets: In our experiments, three standard datasets are used. Office-31 [38] is a popular benchmark. It contains a total of 4110 images of 31 office environment objects from 3 domains: Amazon (A), Webcam (W), Dslr(D). Office-Home [47] is a more challenging dataset which contains 15588 images within 65 classes from 4 domains: Artistic images (A), Clip-Art images (C), Product images (P) and RealWorld images (R). Visda-17 [36] is a widely used benchmark for domain adaptation with focus on a 12-class synthesis-to-real object classification task. The source domain contains 152,397 synthetic images and the target domain has 55,388 real object images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Implementation details: Our experiment is performed in Pytorch. Each task is run 5 times to enhance the robustness of the results. The same backbone network is selected as other compared methods for fair comparison. Specifically, Resnet-50 is selected as the backbone on Office-31 and Office-Home, and Resnet-101 is selected on Visda-17. It is worth noting that the output dimension of the classifier in the original backbone is replaced by the number of categories to fit each task. The SGD optimizer is chosen to update the network and the CosineAnnealingLR [33] is used to update the learning rate of the SGD optimizer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Competitors: For extensive evaluation, we compare our model with five groups of state-of-the-art methods. The first is based on distribution alignment, such as statistic moment matching methods DAN [31], CAN [18], and TSA [25]. The second uses adversarial learning including DANN [10], CDAN [32], MDD+IA [16], DADA [45], CLS [30], ILA [40], MetaAlign [48], DWL [49], and DALN [3]. The third exploits optimal transport: DeepJ-DOT [7]. The fourth is based on bi-classifier adversarial learning, MCD [39]. The fifth adopts self-supervised learning, including ALDA [4], ATDOC [26], CaCo [15], and SUDA [52]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.785, + 0.743 + ], + "angle": 0, + "content": "5.2. Comparisons to State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The performance comparison with the state-of-the-art methods on Office-31, Office-home and Visda-17 are shown in Table 2, Table 3 and Table 4 respectively. The methods HMA(DANN) and HMA(CDAN) mean the sewing up by distribution matching method based on adversarial learning, with DANN and CDAN focusing on marginal distribution alignment and conditional distribution alignment respectively. HMA(DeepHDOT) integrates the optimal transport method DeepJDOT, and HMA(MCD) the bi-classifier adversarial learning method MCD, as the sewing up tool." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18704" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.088, + 0.877, + 0.101 + ], + "angle": 0, + "content": "Table 2. Comparison with the state-of-the-art methods on Office-31 dataset. Metric: classification accuracy (%); Backbone: ResNet-50." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.102, + 0.816, + 0.453 + ], + "angle": 0, + "content": "
MethodVenueA→DA→WD→AD→WW→AW→Davg
ResNet-50 [13]CVPR1668.968.462.596.760.799.376.1
DAN [31]ICML1578.680.563.697.162.899.680.4
CAN [18]CVPR1995.094.578.099.177.099.890.6
TSA [25]CVPR2192.694.874.999.174.4100.089.3
DANN [10]JMLR1679.782.068.296.967.499.182.2
CDAN [32]NIPS1889.893.170.198.268.0100.086.5
DADA [45]AAAI2093.992.374.499.274.2100.089.0
MDD+IA [16]ICML2092.190.375.398.774.999.888.8
ILA [40]CVPR2193.495.772.199.375.4100.089.3
MetaAlign [48]CVPR2194.593.075.098.673.6100.089.2
DWL [49]CVPR2191.289.273.199.269.8100.087.1
DALN [3]CVPR2295.495.276.499.176.5100.090.4
DeepJDOT [7]ECCV1888.288.972.198.570.199.686.2
MCD [39]CVPR1892.288.669.598.569.7100.086.5
ALDA [4]AAAI2094.095.672.297.772.5100.088.7
ATDOC [26]CVPR2194.494.575.698.975.299.689.7
CaCo [15]CVPR2291.789.773.198.472.8100.087.6
SUDA [52]CVPR2291.290.872.298.771.4100.087.4
HMA(DANN)Ours83.9±0.183.5±0.270.5±0.198.2±0.170.1±0.2100.0±0.084.4
HMA(CDAN)Ours92.4±0.295.1±0.273.7±0.199.2±0.172.8±0.2100.0±0.088.9
HMA(DeepJDOT)Ours90.9±0.191.6±0.274.6±0.299.0±0.173.8±0.199.8±0.088.3
HMA(MCD)Ours93.5±0.191.3±0.173.1±0.199.2±0.173.5±0.1100.0±0.088.4
HMA(DAN)Ours85.1±0.284.5±0.267.9±0.398.9±0.266.7±0.3100.0±0.083.9
HMA(CAN)Ours95.8±0.395.1±0.179.3±0.399.3±0.177.6±0.2100.0±0.091.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.477, + 0.891, + 0.503 + ], + "angle": 0, + "content": "Table 3. Comparisons with the state-of-the-art methods on Office-Home dataset. Metric: classification accuracy (\\%); Backbone: ResNet-50." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.503, + 0.83, + 0.785 + ], + "angle": 0, + "content": "
MethodVenueA→CA→PA→RC→AC→PC→RP→AP→CP→RR→AR→CR→Pavg
ResNet-50 [13]CVPR1634.950.058.037.441.946.238.531.260.453.941.259.946.1
DAN [31]ICML1543.657.067.945.856.560.444.043.667.763.151.574.356.3
CAN [18]CVPR1958.778.182.167.475.778.167.254.282.573.460.983.571.8
TSA [25]CVPR2153.675.178.364.473.772.562.349.477.572.258.882.168.3
DANN [10]JMLR1645.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [32]NIPS1849.069.374.554.466.068.455.648.375.968.455.480.563.8
MDD+IA [16]ICML2056.277.979.264.473.174.464.254.279.971.258.183.169.5
MetaAlign [48]CVPR2159.376.080.265.774.775.165.756.581.674.161.185.271.3
DALN [3]CVPR2257.879.982.066.376.277.266.755.581.373.560.485.271.8
DeepJDOT [7]ECCV1850.768.674.459.965.868.155.246.373.866.054.978.363.5
MCD [39]CVPR1848.968.374.661.367.668.857.047.175.169.152.279.664.1
ALDAAAAI2053.770.176.460.272.671.556.851.977.170.256.382.166.6
ATDOC [26]CVPR2158.378.882.369.478.278.267.156.082.772.058.285.572.2
HMA(DANN)Ours48.265.175.457.065.068.355.645.273.566.654.378.462.7
HMA(CDAN)Ours58.778.181.667.475.878.166.854.282.573.459.783.571.7
HMA(DeepJDOT)Ours57.574.780.766.472.174.261.252.880.672.861.784.870.0
HMA(MCD)Ours55.574.581.367.674.175.563.653.982.275.358.782.470.4
HMA(DAN)Ours46.263.573.958.165.368.355.343.974.867.253.478.462.4
HMA(CAN)Ours60.679.182.968.977.579.369.155.983.574.662.384.473.2
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.469, + 0.884 + ], + "angle": 0, + "content": "While HMA(DAN) and HMA(CAN) apply the statistic moment matching methods as the sewing up tool, where DAN and CAN focus on marginal distribution alignment and conditional distribution alignment respectively." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "It can be observed that HMA(CAN) yields the best av" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.824, + 0.892, + 0.899 + ], + "angle": 0, + "content": "erage performance on both three datasets. This also confirms our previous analysis. Different sewing up methods will affect the final performance. In general, conditional distribution alignment is better than marginal distribution alignment on the two kinds of methods based on adversar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18705" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.113 + ], + "angle": 0, + "content": "Table 4. Comparison with the state-of-the-art methods on Visda-17 dataset. Metric: per-class classification accuracy (%); Backbone: ResNet-101." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.113, + 0.819, + 0.424 + ], + "angle": 0, + "content": "
MethodVenueplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckavg
ResNet-101 [13]CVPR1655.153.361.959.180.617.979.731.281.026.573.58.552.4
DAN [31]ICML1584.842.175.453.077.962.686.650.759.752.982.526.062.9
CAN [18]CVPR1997.087.282.574.397.896.290.880.796.696.387.559.987.2
TSA [25]CVPR21------------78.6
DANN [10]JMLR1681.977.782.844.381.229.565.128.651.954.682.87.857.4
CDAN [32]NIPS1885.266.983.050.884.274.988.174.583.476.081.938.073.9
DWL [49]CVPR2190.780.286.167.692.481.586.878.090.657.185.628.777.1
CLS [30]ICCV2192.684.573.772.788.583.389.177.689.589.285.872.781.6
DALN [3]CVPR22------------80.6
DeepJDOT [7]ECCV1885.473.477.387.384.164.791.579.391.944.488.561.877.4
MCD [39]CVPR1887.060.983.764.088.979.684.776.988.640.383.025.871.9
ALDA [4]AAAI2093.874.182.469.490.687.289.067.693.476.187.722.277.8
ATDOC [26]CVPR2193.783.076.958.789.795.184.471.489.480.086.755.180.3
CaCo [15]CVPR2290.480.778.857.088.987.081.379.488.788.186.863.980.9
SUDA [52]CVPR2288.379.366.264.787.480.185.978.386.387.578.874.579.8
HMA(DANN)Ours86.979.183.550.586.747.386.155.164.659.884.636.268.4
HMA(CDAN)Ours88.371.285.166.486.379.388.887.683.979.383.446.278.8
HMA(DeepJDOT)Ours89.376.880.668.387.975.690.686.289.673.582.249.879.2
HMA(MCD)Ours89.072.285.570.390.486.687.683.892.255.589.142.178.7
HMA(DAN)Ours87.549.280.253.881.871.887.857.660.957.085.332.867.1
HMA(CAN)Ours97.688.484.376.098.497.191.381.497.096.788.860.788.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.457, + 0.47, + 0.499 + ], + "angle": 0, + "content": "Table 5. Homeomorphism mapping vs ordinary bijection on Office-31. OB(CAN) means ordinary bijection sewed by the distribution matching method CAN." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.508, + 0.49, + 0.555 + ], + "angle": 0, + "content": "
ComponentA→DA→WD→AW→AParameters
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.220992000
OB(CAN)90.3±0.391.7±0.276.6±0.175.8±0.233603584
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.822 + ], + "angle": 0, + "content": "ial learning strategy and statistic moment matching strategy because conditional distribution alignment can stitch homeomorphism mapping with two spaces by category. Furthermore, for condition distribution matching method, statistic moment matching strategy is better than adversarial learning strategy. The reason is that statistic moment matching strategy stitches homeomorphism mapping with two feature spaces explicitly by category. Interestingly, the alignment method CAN published in 2019 still achieves SOTA results. It can be seen that alignment by category is very important for extracting domain invariant features. Our method HMA(CAN) further boosts the CAN performance, because we realize the difficulty of extracting a domain invariant space and we do alignment in the two spaces by homeomorphism mapping." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.37, + 0.848 + ], + "angle": 0, + "content": "5.3. Ablation Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Homeomorphism map is better than ordinary bijection. As mentioned in Section 3, there does not exist many methods which do alignments in two feature spaces based on bi-" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.457, + 0.804, + 0.472 + ], + "angle": 0, + "content": "Table 6. Ablation study on Office-31." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.481, + 0.905, + 0.613 + ], + "angle": 0, + "content": "
ComponentA→DA→WD→AW→A
CAN95.0±0.394.5±0.378.0±0.377.0±0.3
INN(CAN)94.8±0.394.1±0.277.3±0.476.7±0.2
INN(CAN)+S2T95.6±0.294.9±0.378.9±0.277.4±0.2
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.2
DAN78.6±0.280.5±0.463.6±0.362.8±0.2
INN(DAN)78.3±0.379.9±0.262.9±0.262.6±0.1
INN(DAN)+S2T84.1±0.283.8±0.366.5±0.266.1±0.2
INN(DAN)+S2T+T2S85.1±0.284.5±0.267.9±0.366.7±0.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.807 + ], + "angle": 0, + "content": "jection. In this experiment, we will apply the ordinary bijection method to top-4 challenging tasks on Office-31 dataset, the difference between homeomorphism implemented by INN and ordinary bijection is mainly reflected in topological structure maintenance. The results are shown in Table 5. It is obvious that homeomorphism mapping is superior to the ordinary bijection in both accuracy and model parameters. For the model size, the INN based homeomorphism mapping consumes around half of the parameters compared to using two neural networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Ablation study. To show the effectiveness of alignment in two spaces, we conduct an experiment on top-4 challenging tasks of Office-31 dataset. The results are shown in Table 6. The method CAN is considered as the baseline. INN(CAN) means just sewing up homeomorphism mapping to the two feature spaces and the source model is retrained based on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18706" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.394 + ], + "angle": 0, + "content": "the source labels. INN(CAN) \\(^+\\) S2T means the transformed source features are used to learn the model in the target feature space compared with INN(CAN). HMA(CAN) uses all features in two spaces. From Table 6, we can find that simply using INN can achieve similar performance as the feature distribution alignment method CAN. As shown in the third row in Table 6, by transferring the source features to the target feature space, the performance is greatly improved. For this case, the data topological structure and label information from source domain can be correctly transformed to the target domain, which allows that the learned model works well in the target domain. The fourth row in Table 6 shows that the performance can be further improved if both transformed features are used. Because our homeomorphism mapping keeps the corresponding relationship by category, with the help of supervision information in the source feature spaces, the generalization performance of the model in the two domains is improved. To further verify our ideas, we test DAN as a sewing tool, giving the consistent conclusion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.41, + 0.251, + 0.424 + ], + "angle": 0, + "content": "6. Proof of Theorem" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.263, + 0.449 + ], + "angle": 0, + "content": "6.1. Proof of Theorem 1" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.457, + 0.469, + 0.489 + ], + "angle": 0, + "content": "Theorem 1. Let \\((M, d_M)\\) and \\((N, d_N)\\) be two metric spaces with a homeomorphism" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.502, + 0.319, + 0.518 + ], + "angle": 0, + "content": "\\[\ng: M \\to N,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.469, + 0.562 + ], + "angle": 0, + "content": "and \\(A\\) is an open subset in \\((M,d_M)\\) , we have that its image \\(B\\coloneqq g(A)\\) is an open subset in \\((N,d_N)\\) , and" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.574, + 0.355, + 0.591 + ], + "angle": 0, + "content": "\\[\ng (\\partial A) = \\partial B = \\partial g (A),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.604, + 0.274, + 0.618 + ], + "angle": 0, + "content": "where \\(\\partial\\) means the boundary." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.619, + 0.304, + 0.633 + ], + "angle": 0, + "content": "Proof. It is sufficient to show that" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.647, + 0.469, + 0.664 + ], + "angle": 0, + "content": "\\[\ng (\\partial A) \\subset \\partial B. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.677, + 0.469, + 0.705 + ], + "angle": 0, + "content": "Assume this is true, then we can imply Eq. 7 to \\(g^{-1}\\), and obtain" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.707, + 0.331, + 0.724 + ], + "angle": 0, + "content": "\\[\ng ^ {- 1} (\\partial B) \\subset \\partial A.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.732, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Hence we have \\(\\partial B\\subset g(\\partial A)\\) . Combing this with Eq. 7, we have \\(g(\\partial A) = \\partial B\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.763, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Now we want to show Eq. 7, that is, for any \\( x \\in \\partial A \\), we have \\( g(x) \\in \\partial B \\). Since \\( x \\in \\partial A \\), but \\( x \\notin A \\), then \\( g(x) \\notin B \\), and there is a sequence \\( \\{x_i\\} \\subset A \\) such that \\( \\lim_{i \\to \\infty} x_i = x \\). By the continuity of the function \\( g \\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.837, + 0.384, + 0.859 + ], + "angle": 0, + "content": "\\[\ng (x) = g \\left(\\lim _ {i \\rightarrow \\infty} x _ {i}\\right) = \\lim _ {i \\rightarrow \\infty} g \\left(x _ {i}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Noting that \\( g(x_{i}) \\in B \\), we get \\( g(x) \\in \\partial B \\). This completes the proof." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.688, + 0.106 + ], + "angle": 0, + "content": "6.2. Proof of Theorem 2" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.144 + ], + "angle": 0, + "content": "Theorem 2. Invertible Neural Network is a homeomorphism." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.145, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Proof. The definition of homeomorphism is that, a function \\( g: M \\to N \\) between two topological spaces is a homeomorphism if it has the following properties: (1) \\( g \\) is a bijection; (2) \\( g \\) is continuous; (3) The inverse function \\( g^{-1} \\) is continuous." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.312 + ], + "angle": 0, + "content": "For any invertible neural network, assuming that its forward process is \\( g \\), then its invertible process can be represented as \\( g^{-1} \\), so \\( g \\) is a bijection. Because the function of each part of the invertible neural network is continuous, as \\( s(\\cdot) \\) and \\( t(\\cdot) \\) in our method, so both \\( g \\) and \\( g^{-1} \\) are continuous. To sum up, the invertible network is a homeomorphism." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.326, + 0.619, + 0.341 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.351, + 0.892, + 0.593 + ], + "angle": 0, + "content": "In this paper, we have proposed a new unsupervised domain adaptation method, termed as Homeomorphism Alignment in two spaces (HMA). By analyzing previous alignment based methods, we argue that it is difficult to find a common space or domain invariant space to adapt the pretrained source model. So the alignment is performed in two spaces. The extracted source and target features can be further transformed respectively by a homeomorphism mapping so that they can be aligned semantically. Our method consists of three steps, i.e., constructing an INN based homeomorphism mapping, sewing up by category and retraining iteratively training the model in two spaces. In this way, the source labels can be fully used even in the target feature space for improving the model generalization for the target domain. Extensive experimental results demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.607, + 0.671, + 0.625 + ], + "angle": 0, + "content": "8. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.633, + 0.892, + 0.678 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (2018YFE0203900), National Natural Science Foundation of China (62276048)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.693, + 0.597, + 0.708 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.717, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[1] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3334–3342, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.789, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[2] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. *Diverse image style transfer via invertible cross-space mapping.* In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[3] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific clas" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.717, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18707" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "sifier as a discriminator: Discriminator-free adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7181-7190, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.151, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[4] Minghao Chen, Shuai Zhao, Haifeng Liu, and Deng Cai. Adversarial-learned loss for domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3521–3528, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.209, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[5] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.281, + 0.472, + 0.363 + ], + "angle": 0, + "content": "[6] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.366, + 0.472, + 0.449 + ], + "angle": 0, + "content": "[7] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.472, + 0.494 + ], + "angle": 0, + "content": "[8] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.497, + 0.472, + 0.538 + ], + "angle": 0, + "content": "[9] Geoff French, Michal Mackiewicz, and Mark Fisher. Self-ensembleing for visual domain adaptation. In International Conference on Learning Representations, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.47, + 0.61 + ], + "angle": 0, + "content": "[10] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[11] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[14] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[15] Jiaxing Huang, Dayan Guan, Aoran Xiao, Shijian Lu, and Ling Shao. Category contrast for unsupervised domain adaptation in visual tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1203-1214, 2022. 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In International Conference on Machine Learning, pages 4816-4827. PMLR, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[17] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4733-4742, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[18] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4893-4902, 2019. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[19] Younggeun Kim and Donghee Son. Noise conditional flow model for learning the super-resolution space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 424-432, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Prafulla Dhariwal. Glow: generative flow with invertible \\(1 \\times 1\\) convolutions. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 10236-10245, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[21] Sajja Tulasi Krishna and Hemantha Kumar Kalluri. Deep learning and transfer learning approaches for image classification. International Journal of Recent Technology and Engineering (IJRTE), 7(5S4):427-432, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[22] Chen-Yu Lee, Tanmay Batra, Mohammad Haris Baig, and Daniel Ulbricht. Sliced Wasserstein discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10285-10295, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[23] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13936–13944, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[24] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8455-8464, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[25] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11516-11525, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Domain adaptation with auxiliary target domain-oriented classifier. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2021. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchi" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18708" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "cal conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4076-4085, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[29] Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 469-477, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[30] Xiaofeng Liu, Zhenhua Guo, Site Li, Fangxu Xing, Jane You, C.-C. Jay Kuo, Georges El Fakhri, and Jonghye Woo. Adversarial unsupervised domain adaptation with conditional and label shift: Infer, align and iterate. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10367-10376, October 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[31] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[32] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in neural information processing systems, 2018. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[34] James R. Munkres. Topology (2nd edition). In Prentice Hall, 2000. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[35] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, pages 1345-1359, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.visda06924, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.638, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[38] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. European conference on computer vision, pages 213-226. Springer, 2010. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[39] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[40] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5361-5371, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2107-2116, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[42] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. In Domain Adaptation in Computer Vision Applications, pages 153-171. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[43] Yu Sun, Eric Tzeng, Trevor Darrell, and Alexei A Efros. Unsupervised domain adaptation through self-supervision. arXiv preprint arXiv:1909.11825, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[44] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.254, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[45] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.367 + ], + "angle": 0, + "content": "[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.37, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[47] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[48] Guoqiang Wei, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Metaalign: Coordinating domain alignment and classification for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16643-16653, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[49] Ni Xiao and Lei Zhang. Dynamic weighted learning for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15242-15251, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[50] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4394-4403, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[51] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[52] Jingyi Zhang, Jiaxing Huang, Zichen Tian, and Shijian Lu. Spectral unsupervised domain adaptation for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9829-9840, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[53] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[55] Bo Zhao, Jiashi Feng, Xiao Wu, and Shuicheng Yan. A survey on deep learning-based fine-grained object classification and semantic segmentation. International Journal of Automation and Computing, 14(2):119-135, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[56] Lihua Zhou, Mao Ye, Xiatian Zhu, Shuaifeng Li, and Yiguang Liu. Class discriminative adversarial learning for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4318-4326, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18710" + } + ] +] \ No newline at end of file diff --git a/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_origin.pdf b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7bc42065d391adbf528b2199f24fe4af6ede497e --- /dev/null +++ b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/3818e9e0-d874-4841-b0b1-98c499dc8740_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfc2a5af30c5df9217cfac5f8655b59118ebfe16090a0e215a519e27254dfe78 +size 272894 diff --git a/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/full.md b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c9850a2b08ff98bbdc72a9dd75112703f5b1c41 --- /dev/null +++ b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/full.md @@ -0,0 +1,344 @@ +# Homeomorphism Alignment for Unsupervised Domain Adaptation + +Lihua Zhou $^{1}$ , Mao Ye $^{1,*}$ , Xiatian Zhu $^{3}$ , Siying Xiao $^{1}$ , Xu-Qian Fan $^{2}$ , Ferrante Neri $^{3}$ + +1University of Electronic Science and Technology of China + +$^{2}$ Jinan University + +3University of Surrey + +lihua.zhou@std.uestc.edu.cn, cvlab.uestc@gmail.com, xiatian.zhu@surrey.ac.uk + +2018270101016@std.uestc.edu.cn, txqfan@jnu.edu.cn, f.neri@surrey.ac.uk + +# Abstract + +Existing unsupervised domain adaptation (UDA) methods rely on aligning the features from the source and target domains explicitly or implicitly in a common space (i.e., the domain invariant space). Explicit distribution matching ignores the discriminability of learned features, while the implicit counterpart such as self-supervised learning suffers from pseudo-label noises. With distribution alignment, it is challenging to acquire a common space which maintains fully the discriminative structure of both domains. In this work, we propose a novel HomeomorphismM Alignment (HMA) approach characterized by aligning the source and target data in two separate spaces. Specifically, an invertible neural network based homeomorphism is constructed. Distribution matching is then used as a sewing up tool for connecting this homeomorphism mapping between the source and target feature spaces. Theoretically, we show that this mapping can preserve the data topological structure (e.g., the cluster/group structure). This property allows for more discriminative model adaptation by leveraging both the original and transformed features of source data in a supervised manner, and those of target domain in an unsupervised manner (e.g., prediction consistency). Extensive experiments demonstrate that our method can achieve the state-of-the-art results. Code is released at https://github.com/buerzlh/HMA. + +# 1. Introduction + +Deep learning methods rely mostly on a large quantity of manually labeled data [13, 37, 21, 55] which however could be prohibitively expensive or impossible to collect in many scenarios. One effective strategy is to exploit pre-existing labeled data (i.e., the source domain) for a target domain without need for manual labeling. Due to the domain shift + +challenge [35], a model pretrained on a source domain often suffers from drastic performance degradation when directly applied on a target domain. This gives rise to the research attention of Unsupervised Domain Adaptation (UDA). + +Existing UDA methods can be roughly divided into two categories. One is based on distribution alignment [31, 18, 10, 32] which minimizes domain discrepancy by aligning the distributions between two domains. They usually align two different distributions to a single distribution. However, this strategy could distort the original structural information, potentially hurting the final model generalization [5, 11, 44]. The other category is based on self-supervised learning [9, 26, 43] which also learns a single common feature space using pseudo labels or other self-supervision information. Given inevitable noise with self-supervision, it is difficult to obtain a common feature space with discriminative structure well kept. As shown in Fig. 1(a), adapting a model in a common space cannot guarantee better classification performance. + +A natural solution for the above problem is to learn a latent distribution transformation function without destroying the original distribution of both domains. There are a few works in this line. For example, CyCADA [14] uses ordinary bijection implemented by two different networks to transform the images of the source domain to the target domain and vice versa. However, its learned two networks are not strictly inverse mappings, making the transformed images not necessarily semantically consistent through the transformation. That is, an image cannot be reconstructed after going through the two learned mappings. In addition, CyCADA learns two networks which is expensive. + +To address the above limitations, we leverage a stricter bijection, namely Homeomorphism mapping, a concept borrowed from the topology field [34]. If a bijection satisfies the definition of homeomorphism (i.e., one-to-one correspondence and continuous), theoretically we prove that the data topological structure can be well preserved in the projected space (i.e., the samples in the same cluster are still + +![](images/5579a389849140fca7601d675483580f69e2c232a4a5508b7e2094af2cf39ed2.jpg) +(a) Previous methods + +![](images/832384117019492cb04034a06b27f978f03d913af2a8a802a5ba94ad32b2520e.jpg) +(b) Our method + +![](images/d292abfd4c9c31d079be4d7735a0cc518257679fc183b1b605ef5ff392bf7884.jpg) +Source feature + +![](images/3db759acd1eb22e4d614d687c1ea4b4b141c5eec5a18f8f07989a86473050d0f.jpg) +Transformed source feature +Target feature + +![](images/5e8ec6dc831894c9fdefe638e59c2498b32a5983e8e19267b53b3cc500f5db76.jpg) +Transformed target feature +Figure 1. Illustration of previous UDA methods and our homeomorphism alignment. (a) Previous methods align the distributions between two domains in a single common feature space, leading to that per-domain data discriminative structure is not well preserved. To overcome this problem, (b) our method leverages a homeomorphism mapping as a bridge for alignment across the source and target feature spaces, since the homeomorphism mapping can preserve the original data topological structure. + +in the same projected cluster). By this property, we further introduce a noise-free self-supervised learning task. With such a reversible mapping, semantic consistency is guaranteed. As shown in Fig.1(b), with our homeomorphism mapping, the adapted model works better along with both source and target feature spaces. + +Motivated by the above analysis, we propose a novel unsupervised domain adaptation method, called HomeomorphismM Alignment (HMA). Our method consists of three components. The first is to construct a homeomorphism mapping for connecting the source and target feature spaces. We show that an Invertible Neural Network (INN) [20] can be used to derive a pair of mutually invertible functions (i.e., homeomorphism mapping) through its forward and invertible processes. The second is sewing up across the two feature spaces with using distribution matching, as it is required that the cross-space transformed features are aligned with the original features. For discriminative learning, class semantic information is also considered during homeomorphic mapping. The third is to train the model in a self-supervised manner with the labels of source domain and the predictive consistency of unlabeled target domain as supervision. This is noise-free due to the preserved topological structure. + +Our contributions are summarized as follows: (1) We theoretically prove that homeomorphism mapping can guarantee the topological structure of the mapped data. This is an important property yet ignored in the existing researches. We show that an INN can implement a homeomorphism. (2) We propose a novel UDA method with homeomorphism, the first attempt to consider the UDA problem from the viewpoint of topology by conducting domain alignment + +across two feature spaces. This design differs from the previous alternative methods typically learning a single common feature space for domain alignment. (3) Extensive experiments demonstrate the superiority of our HMA over prior art alternatives, along with in-depth ablation studies. + +# 2. Related work + +Unsupervised Domain Adaptation. UDA aims to improve the generalization ability of a model on an unlabeled target domain by leveraging the labeled source domain. Existing methods can be roughly divided into two categories. The first category adopts the idea of distribution alignment that minimizes the source error and the discrepancy between source and target domains concurrently. There exist five main strategies: statistic moment matching, adversarial learning, optimal transport, bi-classifier adversarial learning and adversarial generation. + +Statistic moment matching based methods minimize the statistic discrepancy to align the distributions between two domains (e.g., DAN [31], CORAL [42] and CAN [18]). Adversarial learning based methods are inspired by GAN [12], which plays a minimax game between feature extractor and discriminator to learn domain invariant features (e.g., DANN [10], CDAN [32] and ADDA [46]). Optimal transport based methods generally consist of two steps: Finding a coupling matrix for connecting each source sample and target sample, followed by minimizing the cost of these pair-wise connections (e.g., DeepJDOT [7], RWOT [50] and ETD [23]). Bi-classifier adversarial learning based methods play a minimax game with a single feature extractor and two distinct classifiers during adaptation. + +Commonly, they maximize the prediction discrepancy when training the classifiers and minimize the prediction discrepancy when training the feature extractor (e.g., MCD [39], SWD [22], BCDM [24] and CDAL [56]). Adversarial generation methods combine the domain discriminator and a generator, and generate fake data to align the distributions across domains at the pixel level (e.g., CoGAN [29], SimGAN [41] and CycleGAN [57]). Overall, the first four strategies do feature-level distribution alignment while the last one considers pixel-level alignment. And in this work, we focus on feature-level distribution alignment. + +The second category of UDA methods consider domain adaptation as a self-supervised learning problem. The key of this strategy is to obtain accurate pseudo-labels or self-supervision information. Similarly, such methods also aim to find a common space to implicitly align the source and target features so that the source and target domain features projected by the feature extractor have better discriminability. For example, SE [9] uses the mean teacher framework with a student and a teacher. For the update of the student, it uses the cross-entropy of source samples and the consistency constraints of target samples. While the teacher is updated by exponential moving average of the student. ATDOC [26] assigns a pseudo-label for each target sample by employing a memory mechanism or neighborhood aggregation. ssUDA [43] performs self-supervised tasks (e.g., rotation, flip and patch location predictions) to improve the model generalization. SHOT [27] adopts the information maximization to solve the source-data free UDA problem. BNM [6] proposes Batch Nuclear-norm Maximization on the output matrix to improve both discriminability and diversity. CDTRANS [51] designs a two-way center-aware labeling algorithm to produce pseudo labels for target samples. + +Invertible Neural Network (INN) is a flow-based model that transforms a probability distribution to another by a sequence of invertible and differentiable mappings. This model has been applied in a variety of problems. For example, HCFlow [28] utilizes the hierarchical conditional flow as a unified framework for image super-resolution and image rescaling. NCSR [19] proposes noise conditional flow model for super-resolution for increasing the visual quality and diversity of images. Derived from general volume preserving flows, iVPF [54] solve lossless compression by an exact bijective mapping without any numerical error. iFlow [53] also solves lossless compression by a modular scale transform with numerically invertible flow transformations. DIST [2] a diverse image style transfer framework by enforcing invertible cross-space mapping. Additionally, invertible networks play an important role in protecting privacy, such as invertible de-identification and image hiding [1, 17]. To the best of our knowledge, this is first attempt that leverages INN for unsupervised domain adap + +tion. + +# 3. Analysis of Distribution Alignment + +Problem statements. In UDA, we have a source domain $D_{s} = \{(x_{i}^{s},y_{i}^{s})\}_{i = 1}^{n_{s}}$ with $n_s$ labeled samples and a target domain $D_{t} = \{(x_{i}^{t})\}_{i = 1}^{n_{t}}$ with $n_t$ unlabeled samples. The two domains share the same label space $\{1,2,\dots ,K\}$ , but in different data distributions. The source model $\Gamma_{s}$ , pretrained on labeled source data, is composed of a feature extractor $F$ and a classifier $C$ . The goal of UDA is to adapt the source model for the unlabeled target domain. + +Ideally, given ground-truth labels, the trained model can work well in both the source and target domains. To verify this, we conduct experiments on the Office-31 dataset [38]. Using ground-truth labels of both source and target domains, supervised training is applied. It is found that $100\%$ accuracy on both domains can be achieved. Next, we will investigate the adapted models trained by varying distribution alignment methods. See Appendix for more algorithm details. + +Can previous distribution alignment methods really achieve $100\%$ accuracy given ground-truth target labels? As mentioned earlier, existing UDA methods usually adapt the source domain model to the target domain through two strategies: distribution alignment and self-supervised learning. Commonly, both tackle domain adaptation by projecting source and target samples into a common space (i.e., a domain invariant space). We conduct an experiment on the top-4 most challenging tasks from Office-31. We use the ground-truth target labels for domain alignment. The results are shown in Table 1. The first row shows the results of CAN [18] which aligns feature distributions based on the ground-truth target labels. The second row shows the results with an adversarial learning method CDAN [32]. The third row gives the performance of an optimal transport method DeepJDOT [7]. Note, the bi-classifier adversarial learning methods are excluded as they are designed to use the predictions of the classifiers but not pseudo-labels, and thus ground-truth labels can not be used. It is observed that these explicit distribution alignment methods cannot achieve $100\%$ accuracy, meaning that the discriminative data structure is not well preserved during alignment. The fourth row gives the results of a self-supervised learning approach using the ground-truth target sample labels. It works well which means that the feature extractor can find the domain invariant space while keeping data discriminative structure. However, in practice the label noise cannot be eliminated fully. + +Can bijection achieve $100\%$ accuracy? It is difficult to find a feature extractor to obtain a common space while keeping discriminative data structure. We then investigate whether source and target samples can be aligned in two feature spaces in a way of bijection for superior domain + +Table 1. Up-bound performance probing: Comparing different distribution alignment strategies on Office-31 using the groundtruth target sample labels. SMM: Statistic moment matching; AL: Adversarial learning; OP: Optimal transport; SL: Self-supervised learning; BA:Bijection alignment. + +
ComponentA→DA→WD→AW→A
SMM99.9±0.199.9±0.092.6±0.293.8±0.1
AL99.2±0.199.8±0.190.9±0.292.3±0.1
OP96.2±0.198.8±0.189.9±0.290.9±0.1
SL100.0±0.0100.0±0.0100.0±0.0100.0±0.0
BA97.8±0.298.9±0.190.7±0.193.2±0.2
Ours100.0±0.0100.0±0.0100.0±0.0100.0±0.0
+ +adaptation. The fifth row in Table 1 shows the performance of using two different networks to map the source features to the target feature space and vice versa. In this test, we use ground-truth target labels to train these two networks and align the conditional distributions between the transformed features and original features. This bijection made by the two networks is not a homeomorphism. Thus, the data topological structure can not be preserved in the process of bidirectional projections. It is shown that despite access to all ground-truth labels of target domain, this method still cannot achieve $100\%$ accuracy. + +Homeomorphism alignment can achieve $100\%$ accuracy. A homeomorphism (also known as a continuous transformation) is a one-to-one correspondence mapping between the points in two topological spaces. It is continuous in both directions (refer to [34] for more details). Let $M$ and $N$ be two topological spaces, and $g: M \to N$ be a bijection. If both the function $g$ and its inverse function $g^{-1}: N \to M$ are continuous, then $g$ is called a homeomorphism. That being said, a homeomorphism is a bijective correspondence $g: M \to N$ such that $g(U)$ is open if and only if $U$ is open. By the definition above, $g$ is a homeomorphism if and only if $g^{-1}$ is a homeomorphism. Based on this definition, the following theorem can be derived. + +Theorem 1. The set boundary corresponds to the set boundary by homeomorphism. More precisely, let $(M,d_M)$ and $(N,d_N)$ be two metric spaces where $d_M$ and $d_N$ are the metrics on $M,N$ respectively. Suppose there is a homeomorphism $g:M\to N$ , and $A$ is an open subset in $(M,d_{M})$ , we have that its image $B\coloneqq g(A)$ is an open subset in $(N,d_{N})$ , and + +$$ +g (\partial A) = \partial B = \partial g (A). +$$ + +where $\partial$ means the boundary (see more details in Sec. 6). + +As shown in Theorem 1, data topological structure is preserved by homeomorphism mapping, i.e., the samples in the same cluster are still in the same projected cluster. In the literature, we find that there exists a network satisfying homeomorphism definition – Invertible Neural Network (INN) [20]. We validate that an INN satisfies the following theorem. + +Theorem 2. Invertible Neural Network is a homeomorphism (see more details in Sec. 6). + +As shown in the last row of Table 1, we use an INN to connect two domains at the feature level. With the ground-truth target sample labels, the homeomorphism alignment can achieve $100\%$ accuracy thanks to the topological structure preserving property. + +# 4. Method of Homeomorphism Alignment + +Overview: Based on the analysis of distribution alignment, as shown in Fig. 2, we propose a Homeomorphism Alignment (HMA) method with three components. The first is homeomorphism mapping construction using an INN for connecting the source and target feature spaces. The second is sewing up across the two feature spaces with using distribution matching. The third is training the source model in the source and target feature spaces with a noise-free self-supervised task by using the property of homeomorphism. + +# 4.1. Homeomorphism implemented by invertible neural network (INN) + +In each iteration, we randomly sample a batch of source and target samples. We use ResNet [13] as the feature extractor $F$ to map a sample $x$ to the feature space: $f^{s/t} = F(x^{s/t})$ where $s/t$ represents the source and target domain respectively. Due to the distribution discrepancy between the source and target domains, we consider that the source and target features reside on two different spaces (manifolds) respectively. + +A homeomorphism $g$ consists of $m$ blocks of INN. $m$ is a hyperparameter discussed in Appendix. In each block, we use an affine network to implement the INN [8], as shown in Fig. 2. In the forward process, we transform from the source feature space $\pmb{f}^s$ to the target feature space $\pmb{f}^t$ . Specifically, for the $i$ -th block, we denote the input $\mu_{1:2d}^i$ with $2d$ dimension. And we split evenly $\mu_{1:2d}^i$ to two parts $[\mu_{1:d}^i, \mu_{d+1:2d}^i]$ , and further transform them with two respective linear neural networks $s(\cdot), t(\cdot)$ . The output of $i$ -th block $\mu_{1:2d}^{i+1}$ is then obtained with residual as follows: + +$$ +\begin{array}{l} \mu_ {1: d} ^ {i + 1} = \mu_ {1: d} ^ {i} + s (\mu_ {d + 1: 2 d} ^ {i}), \\ \mu_ {d + 1: 2 d} ^ {i + 1} = \mu_ {d + 1: 2 d} ^ {i} + t (\mu_ {1: d} ^ {i + 1}). \\ \end{array} +$$ + +The output $\mu_{1:2d}^{i+1}$ (a concatenation of $\mu_{1:d}^{i+1}$ and $\mu_{d+1:2d}^{i+1}$ ) will be set as the input of the next block. + +In the inverse projection process, we transform from the target feature space $\pmb{f}^t$ to the source feature space $\pmb{f}^s$ . For the $i$ -th block, we map $\mu_{1:2d}^{i+1}$ to $\mu_{1:2d}^i$ in the opposite away around. According to Eq.(1), we can get the following equation: + +$$ +\begin{array}{l} \mu_ {d + 1: 2 d} ^ {i} = \mu_ {d + 1: 2 d} ^ {i + 1} - t (\mu_ {1: d} ^ {i + 1}), \\ \mu_ {1: d} ^ {i} = \mu_ {1: d} ^ {i + 1} - s \left(\mu_ {d + 1: 2 d} ^ {i}\right). \\ \end{array} +$$ + +![](images/783f10a10654f4b9c9113e9af8b0c4eb8dc58b6723f236cffd8573760a2d71d9.jpg) +Figure 2. The framework of the proposed Homeomorphism Alignment (HMA). (a) We cascade $m$ invertible neural networks to implement a homeomorphism. (b) The transformed features are sewed up with the corresponding feature spaces by category. (c) The source model is iteratively trained in the two spaces concurrently. + +We similarly split $\mu_{1:2d}^{i+1}$ into two parts $[\mu_{1:d}^{i+1}, \mu_{d+1:2d}^{i+1}]$ and follow Eq.(2) to get the original input $\mu_{1:2d}^{i}$ ((concatenation of $\mu_{1:d}^{i}$ and $\mu_{d+1:2d}^{i}$ ). Obviously Eqs. (1) and (2) are inverse functions of each other. We denote the forward process of the $m$ INNs as the function $g$ , while the inverse process as $g^{-1}$ . Hence, the function $g$ is a bijection. Since the functions $s(\cdot)$ , $t(\cdot)$ are implemented by two linear connected neural networks, they are continuous; This means both $g$ and $g^{-1}$ are continuous, too. According to the definition of homeomorphism, this INN is a homeomorphism. + +# 4.2. Sewing up + +Next we will sew up homeomorphism mapping $g$ to the source and target feature spaces such that the corresponding classes are aligned. Suppose the transformed feature $\pmb{f}^{s2t} = g(\pmb{f}^s)$ according to the source feature $\pmb{f}^s$ and the transformed feature $\pmb{f}^{t2s} = g^{-1}(\pmb{f}^t)$ according to target feature $\pmb{f}^t$ . According to the homeomorphism mapping $g$ , $\pmb{f}^s = g^{-1}(\pmb{f}^{s2t})$ and $\pmb{f}^t = g(\pmb{f}^{t2s})$ . To guarantee that the transformed features are located discriminatively (i.e., aligned with the corresponding classes), distribution matching is used for sewing up. The loss function is defined as follows, + +$$ +\min _ {g} \operatorname {L o s s} _ {S e w} = D M \left(\boldsymbol {f} ^ {s 2 t}, \boldsymbol {f} ^ {t}\right) + D M \left(\boldsymbol {f} ^ {t 2 s}, \boldsymbol {f} ^ {s}\right), \tag {3} +$$ + +where $DM(\cdot, \cdot)$ refers to any existing distribution matching method (e.g., DAN [31], CAN [18], DANN [10], CDAN [32], DeepJDOT [7] and MCD [39]). Note, in case of using the adversarial learning method or bi-classifier adversarial learning method, an additional discriminator network or two auxiliary classifiers are needed. + +Recall that only marginal distribution matching methods (e.g., DAN and DANN) cannot achieve satisfactory results, because they focus on overall distribution alignment instead of class-wise alignment. Although our homeomorphism mapping $g$ can ensure that the transformed features preserve the topological structure, without discriminative sewing up, it is still hard to achieve lossless transformation across domains. Thus class conditional distribution matching is a better choice as validated in ours experiment (Table 2-4). + +# 4.3. Model Training + +A model often suffers performance degradation from the domain shift. To address this problem, we leverage the homeomorphic property (preserved topological structure) to perform a noise-free self-supervised training. Specifically, as proved by Theorem 1, $f^s$ has the same structural information as $f^{s2t}$ . Concretely, for a specific labeled source sample $x$ , the corresponding feature $f^s$ and $f^{s2t}$ share the same label. The following loss function is applied to the + +supervised training of feature extractor $F$ and classifier $C$ , + +$$ +L o s s _ {S} = \mathcal {L} ^ {c e} \left(C \left(\boldsymbol {f} ^ {s}\right), \boldsymbol {y} ^ {s}\right) + \mathcal {L} ^ {c e} \left(C \left(\boldsymbol {f} ^ {s 2 t}\right), \boldsymbol {y} ^ {s}\right), \tag {4} +$$ + +where $\pmb{y}^s$ is the corresponding label of the source sample $x^{s}$ and $\mathcal{L}^{ce}(\cdot ,\cdot)$ denotes the cross entropy function. In particular, the term $\mathcal{L}^{ce}(C(\pmb {f}^s),\pmb {y}^s)$ focuses on the classification of the source domain, whilst $\mathcal{L}^{ce}(C(\pmb {f}^{s2t}),\pmb {y}^s)$ is concerned with the classification of the target domain since $\pmb{f}^{s2t}$ and $\pmb{f}^{t}$ have been aligned. + +Considering that our homeomorphism preserves the structure across the mapping and no label information in the target domain, unsupervised consistency constraint is a natural strategy for optimization. Formally, for an unlabeled target sample $x^{t}$ , we formulate the consistency constraint on the predictions between $f^{t}$ and $f^{t2s}$ as: + +$$ +L o s s _ {T} = L _ {C} \left(C \left(\boldsymbol {f} ^ {t}\right), C \left(\boldsymbol {f} ^ {t 2 s}\right)\right), \tag {5} +$$ + +where $L_{C}(\cdot ,\cdot)$ is a consistency constraint such as $L_{1}$ -Norm and $L_{2}$ -Norm. In practice, we found $L_{2}$ -Norm suffices. By combining (4) and (5), the overall loss is defined as follows, + +$$ +\min _ {F, C} \operatorname {L o s s} _ {S} + \operatorname {L o s s} _ {T}. \tag {6} +$$ + +Summary. At the training phase, in each iteration, we first train an INN based homeomorphism mapping, followed by model training in two spaces concurrently. At the inference phase, both the target features $\pmb{f}^t$ and the transformed target features $\pmb{f}^{t2s}$ can be used to make the prediction. Also, average based ensemble can be used to obtain the final prediction. + +Remarks. Our model is trained in the source and target feature spaces concurrently. Compared with the existing alignment based UDA methods in a common space, this design naturally overcomes the intrinsic challenges of projecting the source and target domain samples into a single shared feature space using a feature extraction network while keeping their respective discriminative structures. When the distributions between the two feature spaces are not originally aligned typical in practice (e.g., due to domain-specific characteristics such as different background, viewing conditions, etc.), the homeomorphism provides a flexible non-invasive means for cross-domain relating via transforming their individual features from each other externally. Critically, this alignment in two spaces allows to fully keep the original per-domain characteristics including some discriminative information. Compared with the self-supervised learning methods suffering the noises of pseudo-labeling, our transformed source features in the target domain can directly use the ground-truth source labels, in addition to additionally exploiting the topological structure of the source domain. Further, our consistency constraint can exploit the unlabeled target training data (i.e., the original and transformed target features with shared topological structure) + +in an unsupervised manner, without the notorious pseudolabel noise issue. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets: In our experiments, three standard datasets are used. Office-31 [38] is a popular benchmark. It contains a total of 4110 images of 31 office environment objects from 3 domains: Amazon (A), Webcam (W), Dslr(D). Office-Home [47] is a more challenging dataset which contains 15588 images within 65 classes from 4 domains: Artistic images (A), Clip-Art images (C), Product images (P) and RealWorld images (R). Visda-17 [36] is a widely used benchmark for domain adaptation with focus on a 12-class synthesis-to-real object classification task. The source domain contains 152,397 synthetic images and the target domain has 55,388 real object images. + +Implementation details: Our experiment is performed in Pytorch. Each task is run 5 times to enhance the robustness of the results. The same backbone network is selected as other compared methods for fair comparison. Specifically, Resnet-50 is selected as the backbone on Office-31 and Office-Home, and Resnet-101 is selected on Visda-17. It is worth noting that the output dimension of the classifier in the original backbone is replaced by the number of categories to fit each task. The SGD optimizer is chosen to update the network and the CosineAnnealingLR [33] is used to update the learning rate of the SGD optimizer. + +Competitors: For extensive evaluation, we compare our model with five groups of state-of-the-art methods. The first is based on distribution alignment, such as statistic moment matching methods DAN [31], CAN [18], and TSA [25]. The second uses adversarial learning including DANN [10], CDAN [32], MDD+IA [16], DADA [45], CLS [30], ILA [40], MetaAlign [48], DWL [49], and DALN [3]. The third exploits optimal transport: DeepJ-DOT [7]. The fourth is based on bi-classifier adversarial learning, MCD [39]. The fifth adopts self-supervised learning, including ALDA [4], ATDOC [26], CaCo [15], and SUDA [52]. + +# 5.2. Comparisons to State-of-the-Art + +The performance comparison with the state-of-the-art methods on Office-31, Office-home and Visda-17 are shown in Table 2, Table 3 and Table 4 respectively. The methods HMA(DANN) and HMA(CDAN) mean the sewing up by distribution matching method based on adversarial learning, with DANN and CDAN focusing on marginal distribution alignment and conditional distribution alignment respectively. HMA(DeepHDOT) integrates the optimal transport method DeepJDOT, and HMA(MCD) the bi-classifier adversarial learning method MCD, as the sewing up tool. + +Table 2. Comparison with the state-of-the-art methods on Office-31 dataset. Metric: classification accuracy (%); Backbone: ResNet-50. + +
MethodVenueA→DA→WD→AD→WW→AW→Davg
ResNet-50 [13]CVPR1668.968.462.596.760.799.376.1
DAN [31]ICML1578.680.563.697.162.899.680.4
CAN [18]CVPR1995.094.578.099.177.099.890.6
TSA [25]CVPR2192.694.874.999.174.4100.089.3
DANN [10]JMLR1679.782.068.296.967.499.182.2
CDAN [32]NIPS1889.893.170.198.268.0100.086.5
DADA [45]AAAI2093.992.374.499.274.2100.089.0
MDD+IA [16]ICML2092.190.375.398.774.999.888.8
ILA [40]CVPR2193.495.772.199.375.4100.089.3
MetaAlign [48]CVPR2194.593.075.098.673.6100.089.2
DWL [49]CVPR2191.289.273.199.269.8100.087.1
DALN [3]CVPR2295.495.276.499.176.5100.090.4
DeepJDOT [7]ECCV1888.288.972.198.570.199.686.2
MCD [39]CVPR1892.288.669.598.569.7100.086.5
ALDA [4]AAAI2094.095.672.297.772.5100.088.7
ATDOC [26]CVPR2194.494.575.698.975.299.689.7
CaCo [15]CVPR2291.789.773.198.472.8100.087.6
SUDA [52]CVPR2291.290.872.298.771.4100.087.4
HMA(DANN)Ours83.9±0.183.5±0.270.5±0.198.2±0.170.1±0.2100.0±0.084.4
HMA(CDAN)Ours92.4±0.295.1±0.273.7±0.199.2±0.172.8±0.2100.0±0.088.9
HMA(DeepJDOT)Ours90.9±0.191.6±0.274.6±0.299.0±0.173.8±0.199.8±0.088.3
HMA(MCD)Ours93.5±0.191.3±0.173.1±0.199.2±0.173.5±0.1100.0±0.088.4
HMA(DAN)Ours85.1±0.284.5±0.267.9±0.398.9±0.266.7±0.3100.0±0.083.9
HMA(CAN)Ours95.8±0.395.1±0.179.3±0.399.3±0.177.6±0.2100.0±0.091.2
+ +Table 3. Comparisons with the state-of-the-art methods on Office-Home dataset. Metric: classification accuracy (\%); Backbone: ResNet-50. + +
MethodVenueA→CA→PA→RC→AC→PC→RP→AP→CP→RR→AR→CR→Pavg
ResNet-50 [13]CVPR1634.950.058.037.441.946.238.531.260.453.941.259.946.1
DAN [31]ICML1543.657.067.945.856.560.444.043.667.763.151.574.356.3
CAN [18]CVPR1958.778.182.167.475.778.167.254.282.573.460.983.571.8
TSA [25]CVPR2153.675.178.364.473.772.562.349.477.572.258.882.168.3
DANN [10]JMLR1645.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [32]NIPS1849.069.374.554.466.068.455.648.375.968.455.480.563.8
MDD+IA [16]ICML2056.277.979.264.473.174.464.254.279.971.258.183.169.5
MetaAlign [48]CVPR2159.376.080.265.774.775.165.756.581.674.161.185.271.3
DALN [3]CVPR2257.879.982.066.376.277.266.755.581.373.560.485.271.8
DeepJDOT [7]ECCV1850.768.674.459.965.868.155.246.373.866.054.978.363.5
MCD [39]CVPR1848.968.374.661.367.668.857.047.175.169.152.279.664.1
ALDAAAAI2053.770.176.460.272.671.556.851.977.170.256.382.166.6
ATDOC [26]CVPR2158.378.882.369.478.278.267.156.082.772.058.285.572.2
HMA(DANN)Ours48.265.175.457.065.068.355.645.273.566.654.378.462.7
HMA(CDAN)Ours58.778.181.667.475.878.166.854.282.573.459.783.571.7
HMA(DeepJDOT)Ours57.574.780.766.472.174.261.252.880.672.861.784.870.0
HMA(MCD)Ours55.574.581.367.674.175.563.653.982.275.358.782.470.4
HMA(DAN)Ours46.263.573.958.165.368.355.343.974.867.253.478.462.4
HMA(CAN)Ours60.679.182.968.977.579.369.155.983.574.662.384.473.2
+ +While HMA(DAN) and HMA(CAN) apply the statistic moment matching methods as the sewing up tool, where DAN and CAN focus on marginal distribution alignment and conditional distribution alignment respectively. + +It can be observed that HMA(CAN) yields the best av + +erage performance on both three datasets. This also confirms our previous analysis. Different sewing up methods will affect the final performance. In general, conditional distribution alignment is better than marginal distribution alignment on the two kinds of methods based on adversar + +Table 4. Comparison with the state-of-the-art methods on Visda-17 dataset. Metric: per-class classification accuracy (%); Backbone: ResNet-101. + +
MethodVenueplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckavg
ResNet-101 [13]CVPR1655.153.361.959.180.617.979.731.281.026.573.58.552.4
DAN [31]ICML1584.842.175.453.077.962.686.650.759.752.982.526.062.9
CAN [18]CVPR1997.087.282.574.397.896.290.880.796.696.387.559.987.2
TSA [25]CVPR21------------78.6
DANN [10]JMLR1681.977.782.844.381.229.565.128.651.954.682.87.857.4
CDAN [32]NIPS1885.266.983.050.884.274.988.174.583.476.081.938.073.9
DWL [49]CVPR2190.780.286.167.692.481.586.878.090.657.185.628.777.1
CLS [30]ICCV2192.684.573.772.788.583.389.177.689.589.285.872.781.6
DALN [3]CVPR22------------80.6
DeepJDOT [7]ECCV1885.473.477.387.384.164.791.579.391.944.488.561.877.4
MCD [39]CVPR1887.060.983.764.088.979.684.776.988.640.383.025.871.9
ALDA [4]AAAI2093.874.182.469.490.687.289.067.693.476.187.722.277.8
ATDOC [26]CVPR2193.783.076.958.789.795.184.471.489.480.086.755.180.3
CaCo [15]CVPR2290.480.778.857.088.987.081.379.488.788.186.863.980.9
SUDA [52]CVPR2288.379.366.264.787.480.185.978.386.387.578.874.579.8
HMA(DANN)Ours86.979.183.550.586.747.386.155.164.659.884.636.268.4
HMA(CDAN)Ours88.371.285.166.486.379.388.887.683.979.383.446.278.8
HMA(DeepJDOT)Ours89.376.880.668.387.975.690.686.289.673.582.249.879.2
HMA(MCD)Ours89.072.285.570.390.486.687.683.892.255.589.142.178.7
HMA(DAN)Ours87.549.280.253.881.871.887.857.660.957.085.332.867.1
HMA(CAN)Ours97.688.484.376.098.497.191.381.497.096.788.860.788.1
+ +Table 5. Homeomorphism mapping vs ordinary bijection on Office-31. OB(CAN) means ordinary bijection sewed by the distribution matching method CAN. + +
ComponentA→DA→WD→AW→AParameters
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.220992000
OB(CAN)90.3±0.391.7±0.276.6±0.175.8±0.233603584
+ +ial learning strategy and statistic moment matching strategy because conditional distribution alignment can stitch homeomorphism mapping with two spaces by category. Furthermore, for condition distribution matching method, statistic moment matching strategy is better than adversarial learning strategy. The reason is that statistic moment matching strategy stitches homeomorphism mapping with two feature spaces explicitly by category. Interestingly, the alignment method CAN published in 2019 still achieves SOTA results. It can be seen that alignment by category is very important for extracting domain invariant features. Our method HMA(CAN) further boosts the CAN performance, because we realize the difficulty of extracting a domain invariant space and we do alignment in the two spaces by homeomorphism mapping. + +# 5.3. Ablation Analysis and Discussion + +Homeomorphism map is better than ordinary bijection. As mentioned in Section 3, there does not exist many methods which do alignments in two feature spaces based on bi- + +Table 6. Ablation study on Office-31. + +
ComponentA→DA→WD→AW→A
CAN95.0±0.394.5±0.378.0±0.377.0±0.3
INN(CAN)94.8±0.394.1±0.277.3±0.476.7±0.2
INN(CAN)+S2T95.6±0.294.9±0.378.9±0.277.4±0.2
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.2
DAN78.6±0.280.5±0.463.6±0.362.8±0.2
INN(DAN)78.3±0.379.9±0.262.9±0.262.6±0.1
INN(DAN)+S2T84.1±0.283.8±0.366.5±0.266.1±0.2
INN(DAN)+S2T+T2S85.1±0.284.5±0.267.9±0.366.7±0.3
+ +jection. In this experiment, we will apply the ordinary bijection method to top-4 challenging tasks on Office-31 dataset, the difference between homeomorphism implemented by INN and ordinary bijection is mainly reflected in topological structure maintenance. The results are shown in Table 5. It is obvious that homeomorphism mapping is superior to the ordinary bijection in both accuracy and model parameters. For the model size, the INN based homeomorphism mapping consumes around half of the parameters compared to using two neural networks. + +Ablation study. To show the effectiveness of alignment in two spaces, we conduct an experiment on top-4 challenging tasks of Office-31 dataset. The results are shown in Table 6. The method CAN is considered as the baseline. INN(CAN) means just sewing up homeomorphism mapping to the two feature spaces and the source model is retrained based on + +the source labels. INN(CAN) $^+$ S2T means the transformed source features are used to learn the model in the target feature space compared with INN(CAN). HMA(CAN) uses all features in two spaces. From Table 6, we can find that simply using INN can achieve similar performance as the feature distribution alignment method CAN. As shown in the third row in Table 6, by transferring the source features to the target feature space, the performance is greatly improved. For this case, the data topological structure and label information from source domain can be correctly transformed to the target domain, which allows that the learned model works well in the target domain. The fourth row in Table 6 shows that the performance can be further improved if both transformed features are used. Because our homeomorphism mapping keeps the corresponding relationship by category, with the help of supervision information in the source feature spaces, the generalization performance of the model in the two domains is improved. To further verify our ideas, we test DAN as a sewing tool, giving the consistent conclusion. + +# 6. Proof of Theorem + +# 6.1. Proof of Theorem 1 + +Theorem 1. Let $(M, d_M)$ and $(N, d_N)$ be two metric spaces with a homeomorphism + +$$ +g: M \to N, +$$ + +and $A$ is an open subset in $(M,d_M)$ , we have that its image $B\coloneqq g(A)$ is an open subset in $(N,d_N)$ , and + +$$ +g (\partial A) = \partial B = \partial g (A), +$$ + +where $\partial$ means the boundary. + +Proof. It is sufficient to show that + +$$ +g (\partial A) \subset \partial B. \tag {7} +$$ + +Assume this is true, then we can imply Eq. 7 to $g^{-1}$ , and obtain + +$$ +g ^ {- 1} (\partial B) \subset \partial A. +$$ + +Hence we have $\partial B\subset g(\partial A)$ . Combing this with Eq. 7, we have $g(\partial A) = \partial B$ + +Now we want to show Eq. 7, that is, for any $x \in \partial A$ , we have $g(x) \in \partial B$ . Since $x \in \partial A$ , but $x \notin A$ , then $g(x) \notin B$ , and there is a sequence $\{x_i\} \subset A$ such that $\lim_{i \to \infty} x_i = x$ . By the continuity of the function $g$ , we have + +$$ +g (x) = g \left(\lim _ {i \rightarrow \infty} x _ {i}\right) = \lim _ {i \rightarrow \infty} g \left(x _ {i}\right). +$$ + +Noting that $g(x_{i}) \in B$ , we get $g(x) \in \partial B$ . This completes the proof. + +# 6.2. Proof of Theorem 2 + +Theorem 2. Invertible Neural Network is a homeomorphism. + +Proof. The definition of homeomorphism is that, a function $g: M \to N$ between two topological spaces is a homeomorphism if it has the following properties: (1) $g$ is a bijection; (2) $g$ is continuous; (3) The inverse function $g^{-1}$ is continuous. + +For any invertible neural network, assuming that its forward process is $g$ , then its invertible process can be represented as $g^{-1}$ , so $g$ is a bijection. Because the function of each part of the invertible neural network is continuous, as $s(\cdot)$ and $t(\cdot)$ in our method, so both $g$ and $g^{-1}$ are continuous. To sum up, the invertible network is a homeomorphism. + +# 7. Conclusion + +In this paper, we have proposed a new unsupervised domain adaptation method, termed as Homeomorphism Alignment in two spaces (HMA). By analyzing previous alignment based methods, we argue that it is difficult to find a common space or domain invariant space to adapt the pretrained source model. So the alignment is performed in two spaces. The extracted source and target features can be further transformed respectively by a homeomorphism mapping so that they can be aligned semantically. Our method consists of three steps, i.e., constructing an INN based homeomorphism mapping, sewing up by category and retraining iteratively training the model in two spaces. In this way, the source labels can be fully used even in the target feature space for improving the model generalization for the target domain. Extensive experimental results demonstrate the effectiveness of our method. + +# 8. Acknowledgment + +This work was supported in part by the National Key R&D Program of China (2018YFE0203900), National Natural Science Foundation of China (62276048). + +# References + +[1] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3334–3342, October 2021. 3 +[2] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. *Diverse image style transfer via invertible cross-space mapping.* In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, October 2021. 3 +[3] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific clas + +sifier as a discriminator: Discriminator-free adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7181-7190, 2022. 6, 7, 8 +[4] Minghao Chen, Shuai Zhao, Haifeng Liu, and Deng Cai. Adversarial-learned loss for domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3521–3528, 2020. 6, 7, 8 +[5] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 1 +[6] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 3 +[7] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 2, 3, 5, 6, 7, 8 +[8] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 4 +[9] Geoff French, Michal Mackiewicz, and Mark Fisher. Self-ensembleing for visual domain adaptation. In International Conference on Learning Representations, 2018. 1, 3 +[10] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 2, 5, 6, 7, 8 +[11] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 1 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, 2014. 2 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 4, 7, 8 +[14] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1 +[15] Jiaxing Huang, Dayan Guan, Aoran Xiao, Shijian Lu, and Ling Shao. Category contrast for unsupervised domain adaptation in visual tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1203-1214, 2022. 6, 7, 8 + +[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In International Conference on Machine Learning, pages 4816-4827. PMLR, 2020. 6, 7 +[17] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4733-4742, October 2021. 3 +[18] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4893-4902, 2019. 1, 2, 3, 5, 6, 7, 8 +[19] Younggeun Kim and Donghee Son. Noise conditional flow model for learning the super-resolution space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 424-432, 2021. 3 +[20] Diederik P Kingma and Prafulla Dhariwal. Glow: generative flow with invertible $1 \times 1$ convolutions. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 10236-10245, 2018. 2, 4 +[21] Sajja Tulasi Krishna and Hemantha Kumar Kalluri. Deep learning and transfer learning approaches for image classification. International Journal of Recent Technology and Engineering (IJRTE), 7(5S4):427-432, 2019. 1 +[22] Chen-Yu Lee, Tanmay Batra, Mohammad Haris Baig, and Daniel Ulbricht. Sliced Wasserstein discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10285-10295, 2019. 3 +[23] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13936–13944, 2020. 2 +[24] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8455-8464, 2021. 3 +[25] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11516-11525, 2021. 6, 7, 8 +[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Domain adaptation with auxiliary target domain-oriented classifier. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2021. 1, 3, 6, 7, 8 +[27] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 3 +[28] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchi + +cal conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4076-4085, 2021. 3 +[29] Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 469-477, 2016. 3 +[30] Xiaofeng Liu, Zhenhua Guo, Site Li, Fangxu Xing, Jane You, C.-C. Jay Kuo, Georges El Fakhri, and Jonghye Woo. Adversarial unsupervised domain adaptation with conditional and label shift: Infer, align and iterate. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10367-10376, October 2021. 6, 8 +[31] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 1, 2, 5, 6, 7, 8 +[32] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in neural information processing systems, 2018. 1, 2, 3, 5, 6, 7, 8 +[33] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[34] James R. Munkres. Topology (2nd edition). In Prentice Hall, 2000. 1, 4 +[35] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, pages 1345-1359, 2009. 1 +[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.visda06924, 2017. 6 +[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 1 +[38] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. European conference on computer vision, pages 213-226. Springer, 2010. 3, 6 +[39] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 3, 5, 6, 7, 8 +[40] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5361-5371, 2021. 6, 7 +[41] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2107-2116, 2017. 3 + +[42] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. In Domain Adaptation in Computer Vision Applications, pages 153-171. Springer, 2017. 2 +[43] Yu Sun, Eric Tzeng, Trevor Darrell, and Alexei A Efros. Unsupervised domain adaptation through self-supervision. arXiv preprint arXiv:1909.11825, 2019. 1, 3 +[44] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 1 +[45] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 6, 7 +[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 2 +[47] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 6 +[48] Guoqiang Wei, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Metaalign: Coordinating domain alignment and classification for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16643-16653, 2021. 6, 7 +[49] Ni Xiao and Lei Zhang. Dynamic weighted learning for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15242-15251, 2021. 6, 7, 8 +[50] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4394-4403, 2020. 2 +[51] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3 +[52] Jingyi Zhang, Jiaxing Huang, Zichen Tian, and Shijian Lu. Spectral unsupervised domain adaptation for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9829-9840, 2022. 6, 7, 8 +[53] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 3 +[54] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 3 + +[55] Bo Zhao, Jiashi Feng, Xiao Wu, and Shuicheng Yan. A survey on deep learning-based fine-grained object classification and semantic segmentation. International Journal of Automation and Computing, 14(2):119-135, 2017. 1 +[56] Lihua Zhou, Mao Ye, Xiatian Zhu, Shuaifeng Li, and Yiguang Liu. Class discriminative adversarial learning for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4318-4326, 2022. 3 +[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3 \ No newline at end of file diff --git a/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/images.zip b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e74461e73a6cdd9f67fc28ce0b8e2724d11bf24f --- /dev/null +++ b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:762aa1100fbf1f1b89a8a35dd97ecae9684957e57f416023459fbcb3b36c3dc8 +size 836647 diff --git a/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/layout.json b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..31321e8c698f93cb08417697e5ff683eaa17cb0a --- /dev/null +++ b/2023/Homeomorphism Alignment for Unsupervised Domain Adaptation/layout.json @@ -0,0 +1,10132 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 503, + 121 + ], + "type": "text", + "content": "Homeomorphism Alignment for Unsupervised Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "text", + "content": "Lihua Zhou" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "text", + "content": ", Mao Ye" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "text", + "content": ", Xiatian Zhu" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "text", + "content": ", Siying Xiao" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "text", + "content": ", Xu-Qian Fan" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "text", + "content": ", Ferrante Neri" + }, + { + "bbox": [ + 95, + 142, + 497, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 157, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 157, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 154, + 157, + 441, + 171 + ], + "type": "text", + "content": "1University of Electronic Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 254, + 171, + 339, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 171, + 339, + 185 + ], + "spans": [ + { + "bbox": [ + 254, + 171, + 339, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 254, + 171, + 339, + 185 + ], + "type": "text", + "content": "Jinan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 185, + 349, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 185, + 349, + 199 + ], + "spans": [ + { + "bbox": [ + 244, + 185, + 349, + 199 + ], + "type": "text", + "content": "3University of Surrey" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 199, + 490, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 199, + 490, + 213 + ], + "spans": [ + { + "bbox": [ + 102, + 199, + 490, + 213 + ], + "type": "text", + "content": "lihua.zhou@std.uestc.edu.cn, cvlab.uestc@gmail.com, xiatian.zhu@surrey.ac.uk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 213, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 213, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 479, + 228 + ], + "type": "text", + "content": "2018270101016@std.uestc.edu.cn, txqfan@jnu.edu.cn, f.neri@surrey.ac.uk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 280, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 579 + ], + "type": "text", + "content": "Existing unsupervised domain adaptation (UDA) methods rely on aligning the features from the source and target domains explicitly or implicitly in a common space (i.e., the domain invariant space). Explicit distribution matching ignores the discriminability of learned features, while the implicit counterpart such as self-supervised learning suffers from pseudo-label noises. With distribution alignment, it is challenging to acquire a common space which maintains fully the discriminative structure of both domains. In this work, we propose a novel HomeomorphismM Alignment (HMA) approach characterized by aligning the source and target data in two separate spaces. Specifically, an invertible neural network based homeomorphism is constructed. Distribution matching is then used as a sewing up tool for connecting this homeomorphism mapping between the source and target feature spaces. Theoretically, we show that this mapping can preserve the data topological structure (e.g., the cluster/group structure). This property allows for more discriminative model adaptation by leveraging both the original and transformed features of source data in a supervised manner, and those of target domain in an unsupervised manner (e.g., prediction consistency). Extensive experiments demonstrate that our method can achieve the state-of-the-art results. Code is released at https://github.com/buerzlh/HMA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "content": "Deep learning methods rely mostly on a large quantity of manually labeled data [13, 37, 21, 55] which however could be prohibitively expensive or impossible to collect in many scenarios. One effective strategy is to exploit pre-existing labeled data (i.e., the source domain) for a target domain without need for manual labeling. Due to the domain shift" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 256, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 545, + 303 + ], + "type": "text", + "content": "challenge [35], a model pretrained on a source domain often suffers from drastic performance degradation when directly applied on a target domain. This gives rise to the research attention of Unsupervised Domain Adaptation (UDA)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 484 + ], + "type": "text", + "content": "Existing UDA methods can be roughly divided into two categories. One is based on distribution alignment [31, 18, 10, 32] which minimizes domain discrepancy by aligning the distributions between two domains. They usually align two different distributions to a single distribution. However, this strategy could distort the original structural information, potentially hurting the final model generalization [5, 11, 44]. The other category is based on self-supervised learning [9, 26, 43] which also learns a single common feature space using pseudo labels or other self-supervision information. Given inevitable noise with self-supervision, it is difficult to obtain a common feature space with discriminative structure well kept. As shown in Fig. 1(a), adapting a model in a common space cannot guarantee better classification performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 628 + ], + "type": "text", + "content": "A natural solution for the above problem is to learn a latent distribution transformation function without destroying the original distribution of both domains. There are a few works in this line. For example, CyCADA [14] uses ordinary bijection implemented by two different networks to transform the images of the source domain to the target domain and vice versa. However, its learned two networks are not strictly inverse mappings, making the transformed images not necessarily semantically consistent through the transformation. That is, an image cannot be reconstructed after going through the two learned mappings. In addition, CyCADA learns two networks which is expensive." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "To address the above limitations, we leverage a stricter bijection, namely Homeomorphism mapping, a concept borrowed from the topology field [34]. If a bijection satisfies the definition of homeomorphism (i.e., one-to-one correspondence and continuous), theoretically we prove that the data topological structure can be well preserved in the projected space (i.e., the samples in the same cluster are still" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 134, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 134, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 134, + 713 + ], + "type": "text", + "content": "* corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18699" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 276, + 209 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 276, + 209 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 276, + 209 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 276, + 209 + ], + "type": "image", + "image_path": "5579a389849140fca7601d675483580f69e2c232a4a5508b7e2094af2cf39ed2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 210, + 184, + 221 + ], + "lines": [ + { + "bbox": [ + 89, + 210, + 184, + 221 + ], + "spans": [ + { + "bbox": [ + 89, + 210, + 184, + 221 + ], + "type": "text", + "content": "(a) Previous methods" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 69, + 541, + 209 + ], + "blocks": [ + { + "bbox": [ + 276, + 69, + 541, + 209 + ], + "lines": [ + { + "bbox": [ + 276, + 69, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 276, + 69, + 541, + 209 + ], + "type": "image", + "image_path": "832384117019492cb04034a06b27f978f03d913af2a8a802a5ba94ad32b2520e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 210, + 465, + 221 + ], + "lines": [ + { + "bbox": [ + 395, + 210, + 465, + 221 + ], + "spans": [ + { + "bbox": [ + 395, + 210, + 465, + 221 + ], + "type": "text", + "content": "(b) Our method" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 231, + 89, + 258 + ], + "blocks": [ + { + "bbox": [ + 56, + 231, + 89, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 231, + 89, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 89, + 258 + ], + "type": "image", + "image_path": "d292abfd4c9c31d079be4d7735a0cc518257679fc183b1b605ef5ff392bf7884.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 231, + 165, + 241 + ], + "lines": [ + { + "bbox": [ + 94, + 231, + 165, + 241 + ], + "spans": [ + { + "bbox": [ + 94, + 231, + 165, + 241 + ], + "type": "text", + "content": "Source feature" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 245, + 231, + 277, + 241 + ], + "blocks": [ + { + "bbox": [ + 97, + 245, + 222, + 256 + ], + "lines": [ + { + "bbox": [ + 97, + 245, + 222, + 256 + ], + "spans": [ + { + "bbox": [ + 97, + 245, + 222, + 256 + ], + "type": "text", + "content": "Transformed source feature" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 245, + 231, + 277, + 241 + ], + "lines": [ + { + "bbox": [ + 245, + 231, + 277, + 241 + ], + "spans": [ + { + "bbox": [ + 245, + 231, + 277, + 241 + ], + "type": "image", + "image_path": "3db759acd1eb22e4d614d687c1ea4b4b141c5eec5a18f8f07989a86473050d0f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 231, + 353, + 242 + ], + "lines": [ + { + "bbox": [ + 277, + 231, + 353, + 242 + ], + "spans": [ + { + "bbox": [ + 277, + 231, + 353, + 242 + ], + "type": "text", + "content": "Target feature" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 364, + 244, + 396, + 257 + ], + "blocks": [ + { + "bbox": [ + 364, + 244, + 396, + 257 + ], + "lines": [ + { + "bbox": [ + 364, + 244, + 396, + 257 + ], + "spans": [ + { + "bbox": [ + 364, + 244, + 396, + 257 + ], + "type": "image", + "image_path": "5e8ec6dc831894c9fdefe638e59c2498b32a5983e8e19267b53b3cc500f5db76.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 245, + 529, + 257 + ], + "lines": [ + { + "bbox": [ + 402, + 245, + 529, + 257 + ], + "spans": [ + { + "bbox": [ + 402, + 245, + 529, + 257 + ], + "type": "text", + "content": "Transformed target feature" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 264, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 45, + 264, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 45, + 264, + 547, + 309 + ], + "type": "text", + "content": "Figure 1. Illustration of previous UDA methods and our homeomorphism alignment. (a) Previous methods align the distributions between two domains in a single common feature space, leading to that per-domain data discriminative structure is not well preserved. To overcome this problem, (b) our method leverages a homeomorphism mapping as a bridge for alignment across the source and target feature spaces, since the homeomorphism mapping can preserve the original data topological structure." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 328, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 399 + ], + "type": "text", + "content": "in the same projected cluster). By this property, we further introduce a noise-free self-supervised learning task. With such a reversible mapping, semantic consistency is guaranteed. As shown in Fig.1(b), with our homeomorphism mapping, the adapted model works better along with both source and target feature spaces." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 401, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 616 + ], + "type": "text", + "content": "Motivated by the above analysis, we propose a novel unsupervised domain adaptation method, called HomeomorphismM Alignment (HMA). Our method consists of three components. The first is to construct a homeomorphism mapping for connecting the source and target feature spaces. We show that an Invertible Neural Network (INN) [20] can be used to derive a pair of mutually invertible functions (i.e., homeomorphism mapping) through its forward and invertible processes. The second is sewing up across the two feature spaces with using distribution matching, as it is required that the cross-space transformed features are aligned with the original features. For discriminative learning, class semantic information is also considered during homeomorphic mapping. The third is to train the model in a self-supervised manner with the labels of source domain and the predictive consistency of unlabeled target domain as supervision. This is noise-free due to the preserved topological structure." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (1) We theoretically prove that homeomorphism mapping can guarantee the topological structure of the mapped data. This is an important property yet ignored in the existing researches. We show that an INN can implement a homeomorphism. (2) We propose a novel UDA method with homeomorphism, the first attempt to consider the UDA problem from the viewpoint of topology by conducting domain alignment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 328, + 546, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 546, + 389 + ], + "type": "text", + "content": "across two feature spaces. This design differs from the previous alternative methods typically learning a single common feature space for domain alignment. (3) Extensive experiments demonstrate the superiority of our HMA over prior art alternatives, along with in-depth ablation studies." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 403, + 390, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 390, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 390, + 415 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 425, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 546, + 544 + ], + "type": "text", + "content": "Unsupervised Domain Adaptation. UDA aims to improve the generalization ability of a model on an unlabeled target domain by leveraging the labeled source domain. Existing methods can be roughly divided into two categories. The first category adopts the idea of distribution alignment that minimizes the source error and the discrepancy between source and target domains concurrently. There exist five main strategies: statistic moment matching, adversarial learning, optimal transport, bi-classifier adversarial learning and adversarial generation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "Statistic moment matching based methods minimize the statistic discrepancy to align the distributions between two domains (e.g., DAN [31], CORAL [42] and CAN [18]). Adversarial learning based methods are inspired by GAN [12], which plays a minimax game between feature extractor and discriminator to learn domain invariant features (e.g., DANN [10], CDAN [32] and ADDA [46]). Optimal transport based methods generally consist of two steps: Finding a coupling matrix for connecting each source sample and target sample, followed by minimizing the cost of these pair-wise connections (e.g., DeepJDOT [7], RWOT [50] and ETD [23]). Bi-classifier adversarial learning based methods play a minimax game with a single feature extractor and two distinct classifiers during adaptation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18700" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Commonly, they maximize the prediction discrepancy when training the classifiers and minimize the prediction discrepancy when training the feature extractor (e.g., MCD [39], SWD [22], BCDM [24] and CDAL [56]). Adversarial generation methods combine the domain discriminator and a generator, and generate fake data to align the distributions across domains at the pixel level (e.g., CoGAN [29], SimGAN [41] and CycleGAN [57]). Overall, the first four strategies do feature-level distribution alignment while the last one considers pixel-level alignment. And in this work, we focus on feature-level distribution alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 483 + ], + "type": "text", + "content": "The second category of UDA methods consider domain adaptation as a self-supervised learning problem. The key of this strategy is to obtain accurate pseudo-labels or self-supervision information. Similarly, such methods also aim to find a common space to implicitly align the source and target features so that the source and target domain features projected by the feature extractor have better discriminability. For example, SE [9] uses the mean teacher framework with a student and a teacher. For the update of the student, it uses the cross-entropy of source samples and the consistency constraints of target samples. While the teacher is updated by exponential moving average of the student. ATDOC [26] assigns a pseudo-label for each target sample by employing a memory mechanism or neighborhood aggregation. ssUDA [43] performs self-supervised tasks (e.g., rotation, flip and patch location predictions) to improve the model generalization. SHOT [27] adopts the information maximization to solve the source-data free UDA problem. BNM [6] proposes Batch Nuclear-norm Maximization on the output matrix to improve both discriminability and diversity. CDTRANS [51] designs a two-way center-aware labeling algorithm to produce pseudo labels for target samples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Invertible Neural Network (INN) is a flow-based model that transforms a probability distribution to another by a sequence of invertible and differentiable mappings. This model has been applied in a variety of problems. For example, HCFlow [28] utilizes the hierarchical conditional flow as a unified framework for image super-resolution and image rescaling. NCSR [19] proposes noise conditional flow model for super-resolution for increasing the visual quality and diversity of images. Derived from general volume preserving flows, iVPF [54] solve lossless compression by an exact bijective mapping without any numerical error. iFlow [53] also solves lossless compression by a modular scale transform with numerically invertible flow transformations. DIST [2] a diverse image style transfer framework by enforcing invertible cross-space mapping. Additionally, invertible networks play an important role in protecting privacy, such as invertible de-identification and image hiding [1, 17]. To the best of our knowledge, this is first attempt that leverages INN for unsupervised domain adap" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 73, + 328, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 328, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 328, + 83 + ], + "type": "text", + "content": "tion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 95, + 500, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 95, + 500, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 95, + 500, + 109 + ], + "type": "text", + "content": "3. Analysis of Distribution Alignment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": "Problem statements. In UDA, we have a source domain " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "D_{s} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{n_{s}}" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": " labeled samples and a target domain " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "D_{t} = \\{(x_{i}^{t})\\}_{i = 1}^{n_{t}}" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": " unlabeled samples. The two domains share the same label space " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots ,K\\}" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": ", but in different data distributions. The source model " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\Gamma_{s}" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": ", pretrained on labeled source data, is composed of a feature extractor " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 114, + 547, + 211 + ], + "type": "text", + "content": ". The goal of UDA is to adapt the source model for the unlabeled target domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 211, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 211, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 547, + 319 + ], + "type": "text", + "content": "Ideally, given ground-truth labels, the trained model can work well in both the source and target domains. To verify this, we conduct experiments on the Office-31 dataset [38]. Using ground-truth labels of both source and target domains, supervised training is applied. It is found that " + }, + { + "bbox": [ + 305, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 211, + 547, + 319 + ], + "type": "text", + "content": " accuracy on both domains can be achieved. Next, we will investigate the adapted models trained by varying distribution alignment methods. See Appendix for more algorithm details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 319, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 547, + 653 + ], + "type": "text", + "content": "Can previous distribution alignment methods really achieve " + }, + { + "bbox": [ + 305, + 319, + 547, + 653 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 319, + 547, + 653 + ], + "type": "text", + "content": " accuracy given ground-truth target labels? As mentioned earlier, existing UDA methods usually adapt the source domain model to the target domain through two strategies: distribution alignment and self-supervised learning. Commonly, both tackle domain adaptation by projecting source and target samples into a common space (i.e., a domain invariant space). We conduct an experiment on the top-4 most challenging tasks from Office-31. We use the ground-truth target labels for domain alignment. The results are shown in Table 1. The first row shows the results of CAN [18] which aligns feature distributions based on the ground-truth target labels. The second row shows the results with an adversarial learning method CDAN [32]. The third row gives the performance of an optimal transport method DeepJDOT [7]. Note, the bi-classifier adversarial learning methods are excluded as they are designed to use the predictions of the classifiers but not pseudo-labels, and thus ground-truth labels can not be used. It is observed that these explicit distribution alignment methods cannot achieve " + }, + { + "bbox": [ + 305, + 319, + 547, + 653 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 319, + 547, + 653 + ], + "type": "text", + "content": " accuracy, meaning that the discriminative data structure is not well preserved during alignment. The fourth row gives the results of a self-supervised learning approach using the ground-truth target sample labels. It works well which means that the feature extractor can find the domain invariant space while keeping data discriminative structure. However, in practice the label noise cannot be eliminated fully." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "type": "text", + "content": "Can bijection achieve " + }, + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "type": "text", + "content": " accuracy? It is difficult to find a feature extractor to obtain a common space while keeping discriminative data structure. We then investigate whether source and target samples can be aligned in two feature spaces in a way of bijection for superior domain" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18701" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 123, + 281, + 203 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 123 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 123 + ], + "type": "text", + "content": "Table 1. Up-bound performance probing: Comparing different distribution alignment strategies on Office-31 using the groundtruth target sample labels. SMM: Statistic moment matching; AL: Adversarial learning; OP: Optimal transport; SL: Self-supervised learning; BA:Bijection alignment." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 123, + 281, + 203 + ], + "lines": [ + { + "bbox": [ + 54, + 123, + 281, + 203 + ], + "spans": [ + { + "bbox": [ + 54, + 123, + 281, + 203 + ], + "type": "table", + "html": "
ComponentA→DA→WD→AW→A
SMM99.9±0.199.9±0.092.6±0.293.8±0.1
AL99.2±0.199.8±0.190.9±0.292.3±0.1
OP96.2±0.198.8±0.189.9±0.290.9±0.1
SL100.0±0.0100.0±0.0100.0±0.0100.0±0.0
BA97.8±0.298.9±0.190.7±0.193.2±0.2
Ours100.0±0.0100.0±0.0100.0±0.0100.0±0.0
", + "image_path": "e2d59fb2731f166e3e992b10d3f8b9a8a94842eb035765a5e67818c389c1a664.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 286, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 286, + 348 + ], + "type": "text", + "content": "adaptation. The fifth row in Table 1 shows the performance of using two different networks to map the source features to the target feature space and vice versa. In this test, we use ground-truth target labels to train these two networks and align the conditional distributions between the transformed features and original features. This bijection made by the two networks is not a homeomorphism. Thus, the data topological structure can not be preserved in the process of bidirectional projections. It is shown that despite access to all ground-truth labels of target domain, this method still cannot achieve " + }, + { + "bbox": [ + 46, + 217, + 286, + 348 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 217, + 286, + 348 + ], + "type": "text", + "content": " accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": "Homeomorphism alignment can achieve " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " accuracy. A homeomorphism (also known as a continuous transformation) is a one-to-one correspondence mapping between the points in two topological spaces. It is continuous in both directions (refer to [34] for more details). Let " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " be two topological spaces, and " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g: M \\to N" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " be a bijection. If both the function " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " and its inverse function " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g^{-1}: N \\to M" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " are continuous, then " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " is called a homeomorphism. That being said, a homeomorphism is a bijective correspondence " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g: M \\to N" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g(U)" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " is open if and only if " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " is open. By the definition above, " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " is a homeomorphism if and only if " + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 46, + 350, + 286, + 504 + ], + "type": "text", + "content": " is a homeomorphism. Based on this definition, the following theorem can be derived." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": "Theorem 1. The set boundary corresponds to the set boundary by homeomorphism. More precisely, let " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "(M,d_M)" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "(N,d_N)" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": " be two metric spaces where " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "d_M" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "d_N" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": " are the metrics on " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "M,N" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": " respectively. Suppose there is a homeomorphism " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "g:M\\to N" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": " is an open subset in " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "(M,d_{M})" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": ", we have that its image " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "B\\coloneqq g(A)" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": " is an open subset in " + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "inline_equation", + "content": "(N,d_{N})" + }, + { + "bbox": [ + 47, + 505, + 286, + 588 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 597, + 215, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 597, + 215, + 609 + ], + "spans": [ + { + "bbox": [ + 116, + 597, + 215, + 609 + ], + "type": "interline_equation", + "content": "g (\\partial A) = \\partial B = \\partial g (A).", + "image_path": "4544433cd50a5014b7cce7fce47cce8854972e7aeb081626d90da251e2d06b3d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 617, + 279, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 279, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 279, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 279, + 629 + ], + "type": "inline_equation", + "content": "\\partial" + }, + { + "bbox": [ + 47, + 617, + 279, + 629 + ], + "type": "text", + "content": " means the boundary (see more details in Sec. 6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "text", + "content": "As shown in Theorem 1, data topological structure is preserved by homeomorphism mapping, i.e., the samples in the same cluster are still in the same projected cluster. In the literature, we find that there exists a network satisfying homeomorphism definition – Invertible Neural Network (INN) [20]. We validate that an INN satisfies the following theorem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "Theorem 2. Invertible Neural Network is a homeomorphism (see more details in Sec. 6)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 96, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 156 + ], + "type": "text", + "content": "As shown in the last row of Table 1, we use an INN to connect two domains at the feature level. With the ground-truth target sample labels, the homeomorphism alignment can achieve " + }, + { + "bbox": [ + 304, + 96, + 545, + 156 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 96, + 545, + 156 + ], + "type": "text", + "content": " accuracy thanks to the topological structure preserving property." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 167, + 523, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 523, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 523, + 182 + ], + "type": "text", + "content": "4. Method of Homeomorphism Alignment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 188, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 545, + 297 + ], + "type": "text", + "content": "Overview: Based on the analysis of distribution alignment, as shown in Fig. 2, we propose a Homeomorphism Alignment (HMA) method with three components. The first is homeomorphism mapping construction using an INN for connecting the source and target feature spaces. The second is sewing up across the two feature spaces with using distribution matching. The third is training the source model in the source and target feature spaces with a noise-free self-supervised task by using the property of homeomorphism." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 328 + ], + "type": "text", + "content": "4.1. Homeomorphism implemented by invertible neural network (INN)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": "In each iteration, we randomly sample a batch of source and target samples. We use ResNet [13] as the feature extractor " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": " to map a sample " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": " to the feature space: " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "f^{s/t} = F(x^{s/t})" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "inline_equation", + "content": "s/t" + }, + { + "bbox": [ + 304, + 335, + 545, + 430 + ], + "type": "text", + "content": " represents the source and target domain respectively. Due to the distribution discrepancy between the source and target domains, we consider that the source and target features reside on two different spaces (manifolds) respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": "A homeomorphism " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " blocks of INN. " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " is a hyperparameter discussed in Appendix. In each block, we use an affine network to implement the INN [8], as shown in Fig. 2. In the forward process, we transform from the source feature space " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{f}^s" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " to the target feature space " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{f}^t" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": ". Specifically, for the " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": "-th block, we denote the input " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^i" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "2d" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " dimension. And we split evenly " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^i" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " to two parts " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "[\\mu_{1:d}^i, \\mu_{d+1:2d}^i]" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": ", and further transform them with two respective linear neural networks " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "s(\\cdot), t(\\cdot)" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": ". The output of " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": "-th block " + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^{i+1}" + }, + { + "bbox": [ + 304, + 431, + 545, + 552 + ], + "type": "text", + "content": " is then obtained with residual as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 561, + 487, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 561, + 487, + 591 + ], + "spans": [ + { + "bbox": [ + 365, + 561, + 487, + 591 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {1: d} ^ {i + 1} = \\mu_ {1: d} ^ {i} + s (\\mu_ {d + 1: 2 d} ^ {i}), \\\\ \\mu_ {d + 1: 2 d} ^ {i + 1} = \\mu_ {d + 1: 2 d} ^ {i} + t (\\mu_ {1: d} ^ {i + 1}). \\\\ \\end{array}", + "image_path": "91a8a948a34e7dccd6c8053e34f90b9dcf25010f16526043f0a91e828c01c9af.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "text", + "content": "The output " + }, + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^{i+1}" + }, + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "text", + "content": " (a concatenation of " + }, + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "inline_equation", + "content": "\\mu_{1:d}^{i+1}" + }, + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "inline_equation", + "content": "\\mu_{d+1:2d}^{i+1}" + }, + { + "bbox": [ + 304, + 601, + 544, + 626 + ], + "type": "text", + "content": ") will be set as the input of the next block." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "text", + "content": "In the inverse projection process, we transform from the target feature space " + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\pmb{f}^t" + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "text", + "content": " to the source feature space " + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\pmb{f}^s" + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "text", + "content": "-th block, we map " + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^{i+1}" + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^i" + }, + { + "bbox": [ + 304, + 628, + 545, + 685 + ], + "type": "text", + "content": " in the opposite away around. According to Eq.(1), we can get the following equation:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 685, + 487, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 685, + 487, + 715 + ], + "spans": [ + { + "bbox": [ + 365, + 685, + 487, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {d + 1: 2 d} ^ {i} = \\mu_ {d + 1: 2 d} ^ {i + 1} - t (\\mu_ {1: d} ^ {i + 1}), \\\\ \\mu_ {1: d} ^ {i} = \\mu_ {1: d} ^ {i + 1} - s \\left(\\mu_ {d + 1: 2 d} ^ {i}\\right). \\\\ \\end{array}", + "image_path": "f1620339a01247fe7a73a88ec81b59c2ebd3af448704c89491dd8a55da1bc687.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "18702" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 339 + ], + "type": "image", + "image_path": "783f10a10654f4b9c9113e9af8b0c4eb8dc58b6723f236cffd8573760a2d71d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "type": "text", + "content": "Figure 2. The framework of the proposed Homeomorphism Alignment (HMA). (a) We cascade " + }, + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 342, + 547, + 376 + ], + "type": "text", + "content": " invertible neural networks to implement a homeomorphism. (b) The transformed features are sewed up with the corresponding feature spaces by category. (c) The source model is iteratively trained in the two spaces concurrently." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": "We similarly split " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^{i+1}" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " into two parts " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "[\\mu_{1:d}^{i+1}, \\mu_{d+1:2d}^{i+1}]" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " and follow Eq.(2) to get the original input " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "\\mu_{1:2d}^{i}" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " ((concatenation of " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "\\mu_{1:d}^{i}" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "\\mu_{d+1:2d}^{i}" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": "). Obviously Eqs. (1) and (2) are inverse functions of each other. We denote the forward process of the " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " INNs as the function " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": ", while the inverse process as " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": ". Hence, the function " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " is a bijection. Since the functions " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "s(\\cdot)" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "t(\\cdot)" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " are implemented by two linear connected neural networks, they are continuous; This means both " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 46, + 393, + 289, + 517 + ], + "type": "text", + "content": " are continuous, too. According to the definition of homeomorphism, this INN is a homeomorphism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 531, + 118, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 118, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 118, + 544 + ], + "type": "text", + "content": "4.2. Sewing up" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": "Next we will sew up homeomorphism mapping " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": " to the source and target feature spaces such that the corresponding classes are aligned. Suppose the transformed feature " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{s2t} = g(\\pmb{f}^s)" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": " according to the source feature " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{f}^s" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": " and the transformed feature " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{t2s} = g^{-1}(\\pmb{f}^t)" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": " according to target feature " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{f}^t" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": ". According to the homeomorphism mapping " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{f}^s = g^{-1}(\\pmb{f}^{s2t})" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\pmb{f}^t = g(\\pmb{f}^{t2s})" + }, + { + "bbox": [ + 46, + 552, + 287, + 683 + ], + "type": "text", + "content": ". To guarantee that the transformed features are located discriminatively (i.e., aligned with the corresponding classes), distribution matching is used for sewing up. The loss function is defined as follows," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 696, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 696, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 58, + 696, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\min _ {g} \\operatorname {L o s s} _ {S e w} = D M \\left(\\boldsymbol {f} ^ {s 2 t}, \\boldsymbol {f} ^ {t}\\right) + D M \\left(\\boldsymbol {f} ^ {t 2 s}, \\boldsymbol {f} ^ {s}\\right), \\tag {3}", + "image_path": "9955c85870d1e45bb9e817d4a319fbddbecd744e77d444b3681e5572a84e4aa1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 395, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 395, + 545, + 467 + ], + "type": "inline_equation", + "content": "DM(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 395, + 545, + 467 + ], + "type": "text", + "content": " refers to any existing distribution matching method (e.g., DAN [31], CAN [18], DANN [10], CDAN [32], DeepJDOT [7] and MCD [39]). Note, in case of using the adversarial learning method or bi-classifier adversarial learning method, an additional discriminator network or two auxiliary classifiers are needed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 471, + 546, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 546, + 578 + ], + "type": "text", + "content": "Recall that only marginal distribution matching methods (e.g., DAN and DANN) cannot achieve satisfactory results, because they focus on overall distribution alignment instead of class-wise alignment. Although our homeomorphism mapping " + }, + { + "bbox": [ + 304, + 471, + 546, + 578 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 471, + 546, + 578 + ], + "type": "text", + "content": " can ensure that the transformed features preserve the topological structure, without discriminative sewing up, it is still hard to achieve lossless transformation across domains. Thus class conditional distribution matching is a better choice as validated in ours experiment (Table 2-4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 595, + 402, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 402, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 402, + 609 + ], + "type": "text", + "content": "4.3. Model Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "A model often suffers performance degradation from the domain shift. To address this problem, we leverage the homeomorphic property (preserved topological structure) to perform a noise-free self-supervised training. Specifically, as proved by Theorem 1, " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^s" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " has the same structural information as " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^{s2t}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". Concretely, for a specific labeled source sample " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", the corresponding feature " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^s" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^{s2t}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " share the same label. The following loss function is applied to the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18703" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "text", + "content": "supervised training of feature extractor " + }, + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 91, + 287, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 91, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 65, + 91, + 287, + 106 + ], + "type": "interline_equation", + "content": "L o s s _ {S} = \\mathcal {L} ^ {c e} \\left(C \\left(\\boldsymbol {f} ^ {s}\\right), \\boldsymbol {y} ^ {s}\\right) + \\mathcal {L} ^ {c e} \\left(C \\left(\\boldsymbol {f} ^ {s 2 t}\\right), \\boldsymbol {y} ^ {s}\\right), \\tag {4}", + "image_path": "05f966ec0a30a0c2a224222ee3e14b9bfb5229876c16ccacf422a334ac269e1e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\pmb{y}^s" + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": " is the corresponding label of the source sample " + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "inline_equation", + "content": "x^{s}" + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ce}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": " denotes the cross entropy function. In particular, the term " + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ce}(C(\\pmb {f}^s),\\pmb {y}^s)" + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": " focuses on the classification of the source domain, whilst " + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ce}(C(\\pmb {f}^{s2t}),\\pmb {y}^s)" + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": " is concerned with the classification of the target domain since " + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{s2t}" + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{t}" + }, + { + "bbox": [ + 47, + 114, + 287, + 185 + ], + "type": "text", + "content": " have been aligned." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": "Considering that our homeomorphism preserves the structure across the mapping and no label information in the target domain, unsupervised consistency constraint is a natural strategy for optimization. Formally, for an unlabeled target sample " + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": ", we formulate the consistency constraint on the predictions between " + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "inline_equation", + "content": "f^{t}" + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "inline_equation", + "content": "f^{t2s}" + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 265, + 287, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 265, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 101, + 265, + 287, + 280 + ], + "type": "interline_equation", + "content": "L o s s _ {T} = L _ {C} \\left(C \\left(\\boldsymbol {f} ^ {t}\\right), C \\left(\\boldsymbol {f} ^ {t 2 s}\\right)\\right), \\tag {5}", + "image_path": "126f4013680fc23ff04ca519ac6bad45f7fae0758b236adfaa52e7a8efdeca61.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "inline_equation", + "content": "L_{C}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "text", + "content": " is a consistency constraint such as " + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "text", + "content": " -Norm and " + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "text", + "content": " -Norm. In practice, we found " + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 288, + 287, + 323 + ], + "type": "text", + "content": " -Norm suffices. By combining (4) and (5), the overall loss is defined as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 333, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 333, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 121, + 333, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\min _ {F, C} \\operatorname {L o s s} _ {S} + \\operatorname {L o s s} _ {T}. \\tag {6}", + "image_path": "ad1ab1be8dbaaf5605a1c57e871707b378c6e1117d136e9c98a6b3362e5d5952.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": "Summary. At the training phase, in each iteration, we first train an INN based homeomorphism mapping, followed by model training in two spaces concurrently. At the inference phase, both the target features " + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\pmb{f}^t" + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": " and the transformed target features " + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{t2s}" + }, + { + "bbox": [ + 46, + 354, + 287, + 436 + ], + "type": "text", + "content": " can be used to make the prediction. Also, average based ensemble can be used to obtain the final prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "type": "text", + "content": "Remarks. Our model is trained in the source and target feature spaces concurrently. Compared with the existing alignment based UDA methods in a common space, this design naturally overcomes the intrinsic challenges of projecting the source and target domain samples into a single shared feature space using a feature extraction network while keeping their respective discriminative structures. When the distributions between the two feature spaces are not originally aligned typical in practice (e.g., due to domain-specific characteristics such as different background, viewing conditions, etc.), the homeomorphism provides a flexible non-invasive means for cross-domain relating via transforming their individual features from each other externally. Critically, this alignment in two spaces allows to fully keep the original per-domain characteristics including some discriminative information. Compared with the self-supervised learning methods suffering the noises of pseudo-labeling, our transformed source features in the target domain can directly use the ground-truth source labels, in addition to additionally exploiting the topological structure of the source domain. Further, our consistency constraint can exploit the unlabeled target training data (i.e., the original and transformed target features with shared topological structure)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "in an unsupervised manner, without the notorious pseudolabel noise issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 108, + 388, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 388, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 388, + 121 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 128, + 422, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 128, + 422, + 141 + ], + "spans": [ + { + "bbox": [ + 306, + 128, + 422, + 141 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 147, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 290 + ], + "type": "text", + "content": "Datasets: In our experiments, three standard datasets are used. Office-31 [38] is a popular benchmark. It contains a total of 4110 images of 31 office environment objects from 3 domains: Amazon (A), Webcam (W), Dslr(D). Office-Home [47] is a more challenging dataset which contains 15588 images within 65 classes from 4 domains: Artistic images (A), Clip-Art images (C), Product images (P) and RealWorld images (R). Visda-17 [36] is a widely used benchmark for domain adaptation with focus on a 12-class synthesis-to-real object classification task. The source domain contains 152,397 synthetic images and the target domain has 55,388 real object images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 291, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 546, + 422 + ], + "type": "text", + "content": "Implementation details: Our experiment is performed in Pytorch. Each task is run 5 times to enhance the robustness of the results. The same backbone network is selected as other compared methods for fair comparison. Specifically, Resnet-50 is selected as the backbone on Office-31 and Office-Home, and Resnet-101 is selected on Visda-17. It is worth noting that the output dimension of the classifier in the original backbone is replaced by the number of categories to fit each task. The SGD optimizer is chosen to update the network and the CosineAnnealingLR [33] is used to update the learning rate of the SGD optimizer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "text", + "content": "Competitors: For extensive evaluation, we compare our model with five groups of state-of-the-art methods. The first is based on distribution alignment, such as statistic moment matching methods DAN [31], CAN [18], and TSA [25]. The second uses adversarial learning including DANN [10], CDAN [32], MDD+IA [16], DADA [45], CLS [30], ILA [40], MetaAlign [48], DWL [49], and DALN [3]. The third exploits optimal transport: DeepJ-DOT [7]. The fourth is based on bi-classifier adversarial learning, MCD [39]. The fifth adopts self-supervised learning, including ALDA [4], ATDOC [26], CaCo [15], and SUDA [52]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 574, + 480, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 480, + 588 + ], + "type": "text", + "content": "5.2. Comparisons to State-of-the-Art" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "The performance comparison with the state-of-the-art methods on Office-31, Office-home and Visda-17 are shown in Table 2, Table 3 and Table 4 respectively. The methods HMA(DANN) and HMA(CDAN) mean the sewing up by distribution matching method based on adversarial learning, with DANN and CDAN focusing on marginal distribution alignment and conditional distribution alignment respectively. HMA(DeepHDOT) integrates the optimal transport method DeepJDOT, and HMA(MCD) the bi-classifier adversarial learning method MCD, as the sewing up tool." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18704" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 80, + 499, + 358 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 536, + 79 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 536, + 79 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 536, + 79 + ], + "type": "text", + "content": "Table 2. Comparison with the state-of-the-art methods on Office-31 dataset. Metric: classification accuracy (%); Backbone: ResNet-50." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 80, + 499, + 358 + ], + "lines": [ + { + "bbox": [ + 94, + 80, + 499, + 358 + ], + "spans": [ + { + "bbox": [ + 94, + 80, + 499, + 358 + ], + "type": "table", + "html": "
MethodVenueA→DA→WD→AD→WW→AW→Davg
ResNet-50 [13]CVPR1668.968.462.596.760.799.376.1
DAN [31]ICML1578.680.563.697.162.899.680.4
CAN [18]CVPR1995.094.578.099.177.099.890.6
TSA [25]CVPR2192.694.874.999.174.4100.089.3
DANN [10]JMLR1679.782.068.296.967.499.182.2
CDAN [32]NIPS1889.893.170.198.268.0100.086.5
DADA [45]AAAI2093.992.374.499.274.2100.089.0
MDD+IA [16]ICML2092.190.375.398.774.999.888.8
ILA [40]CVPR2193.495.772.199.375.4100.089.3
MetaAlign [48]CVPR2194.593.075.098.673.6100.089.2
DWL [49]CVPR2191.289.273.199.269.8100.087.1
DALN [3]CVPR2295.495.276.499.176.5100.090.4
DeepJDOT [7]ECCV1888.288.972.198.570.199.686.2
MCD [39]CVPR1892.288.669.598.569.7100.086.5
ALDA [4]AAAI2094.095.672.297.772.5100.088.7
ATDOC [26]CVPR2194.494.575.698.975.299.689.7
CaCo [15]CVPR2291.789.773.198.472.8100.087.6
SUDA [52]CVPR2291.290.872.298.771.4100.087.4
HMA(DANN)Ours83.9±0.183.5±0.270.5±0.198.2±0.170.1±0.2100.0±0.084.4
HMA(CDAN)Ours92.4±0.295.1±0.273.7±0.199.2±0.172.8±0.2100.0±0.088.9
HMA(DeepJDOT)Ours90.9±0.191.6±0.274.6±0.299.0±0.173.8±0.199.8±0.088.3
HMA(MCD)Ours93.5±0.191.3±0.173.1±0.199.2±0.173.5±0.1100.0±0.088.4
HMA(DAN)Ours85.1±0.284.5±0.267.9±0.398.9±0.266.7±0.3100.0±0.083.9
HMA(CAN)Ours95.8±0.395.1±0.179.3±0.399.3±0.177.6±0.2100.0±0.091.2
", + "image_path": "6a11f5a07117492650a3fe890df793e059cdcf5dadb62da569a5130eb59058ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 86, + 398, + 507, + 621 + ], + "blocks": [ + { + "bbox": [ + 47, + 377, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 47, + 377, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 545, + 398 + ], + "type": "text", + "content": "Table 3. Comparisons with the state-of-the-art methods on Office-Home dataset. Metric: classification accuracy (\\%); Backbone: ResNet-50." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 398, + 507, + 621 + ], + "lines": [ + { + "bbox": [ + 86, + 398, + 507, + 621 + ], + "spans": [ + { + "bbox": [ + 86, + 398, + 507, + 621 + ], + "type": "table", + "html": "
MethodVenueA→CA→PA→RC→AC→PC→RP→AP→CP→RR→AR→CR→Pavg
ResNet-50 [13]CVPR1634.950.058.037.441.946.238.531.260.453.941.259.946.1
DAN [31]ICML1543.657.067.945.856.560.444.043.667.763.151.574.356.3
CAN [18]CVPR1958.778.182.167.475.778.167.254.282.573.460.983.571.8
TSA [25]CVPR2153.675.178.364.473.772.562.349.477.572.258.882.168.3
DANN [10]JMLR1645.659.370.147.058.560.946.143.768.563.251.876.857.6
CDAN [32]NIPS1849.069.374.554.466.068.455.648.375.968.455.480.563.8
MDD+IA [16]ICML2056.277.979.264.473.174.464.254.279.971.258.183.169.5
MetaAlign [48]CVPR2159.376.080.265.774.775.165.756.581.674.161.185.271.3
DALN [3]CVPR2257.879.982.066.376.277.266.755.581.373.560.485.271.8
DeepJDOT [7]ECCV1850.768.674.459.965.868.155.246.373.866.054.978.363.5
MCD [39]CVPR1848.968.374.661.367.668.857.047.175.169.152.279.664.1
ALDAAAAI2053.770.176.460.272.671.556.851.977.170.256.382.166.6
ATDOC [26]CVPR2158.378.882.369.478.278.267.156.082.772.058.285.572.2
HMA(DANN)Ours48.265.175.457.065.068.355.645.273.566.654.378.462.7
HMA(CDAN)Ours58.778.181.667.475.878.166.854.282.573.459.783.571.7
HMA(DeepJDOT)Ours57.574.780.766.472.174.261.252.880.672.861.784.870.0
HMA(MCD)Ours55.574.581.367.674.175.563.653.982.275.358.782.470.4
HMA(DAN)Ours46.263.573.958.165.368.355.343.974.867.253.478.462.4
HMA(CAN)Ours60.679.182.968.977.579.369.155.983.574.662.384.473.2
", + "image_path": "53001e3e3c0b121bf8977847df80e9139935360db7a7199d6f0305b7f91110be.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 651, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 700 + ], + "type": "text", + "content": "While HMA(DAN) and HMA(CAN) apply the statistic moment matching methods as the sewing up tool, where DAN and CAN focus on marginal distribution alignment and conditional distribution alignment respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "It can be observed that HMA(CAN) yields the best av" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 652, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 652, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 652, + 545, + 712 + ], + "type": "text", + "content": "erage performance on both three datasets. This also confirms our previous analysis. Different sewing up methods will affect the final performance. In general, conditional distribution alignment is better than marginal distribution alignment on the two kinds of methods based on adversar" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18705" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 89, + 501, + 335 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 89 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 89 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 89 + ], + "type": "text", + "content": "Table 4. Comparison with the state-of-the-art methods on Visda-17 dataset. Metric: per-class classification accuracy (%); Backbone: ResNet-101." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 89, + 501, + 335 + ], + "lines": [ + { + "bbox": [ + 93, + 89, + 501, + 335 + ], + "spans": [ + { + "bbox": [ + 93, + 89, + 501, + 335 + ], + "type": "table", + "html": "
MethodVenueplanebcyclbuscarhorseknifemcyclpersonplantsktbrdtraintruckavg
ResNet-101 [13]CVPR1655.153.361.959.180.617.979.731.281.026.573.58.552.4
DAN [31]ICML1584.842.175.453.077.962.686.650.759.752.982.526.062.9
CAN [18]CVPR1997.087.282.574.397.896.290.880.796.696.387.559.987.2
TSA [25]CVPR21------------78.6
DANN [10]JMLR1681.977.782.844.381.229.565.128.651.954.682.87.857.4
CDAN [32]NIPS1885.266.983.050.884.274.988.174.583.476.081.938.073.9
DWL [49]CVPR2190.780.286.167.692.481.586.878.090.657.185.628.777.1
CLS [30]ICCV2192.684.573.772.788.583.389.177.689.589.285.872.781.6
DALN [3]CVPR22------------80.6
DeepJDOT [7]ECCV1885.473.477.387.384.164.791.579.391.944.488.561.877.4
MCD [39]CVPR1887.060.983.764.088.979.684.776.988.640.383.025.871.9
ALDA [4]AAAI2093.874.182.469.490.687.289.067.693.476.187.722.277.8
ATDOC [26]CVPR2193.783.076.958.789.795.184.471.489.480.086.755.180.3
CaCo [15]CVPR2290.480.778.857.088.987.081.379.488.788.186.863.980.9
SUDA [52]CVPR2288.379.366.264.787.480.185.978.386.387.578.874.579.8
HMA(DANN)Ours86.979.183.550.586.747.386.155.164.659.884.636.268.4
HMA(CDAN)Ours88.371.285.166.486.379.388.887.683.979.383.446.278.8
HMA(DeepJDOT)Ours89.376.880.668.387.975.690.686.289.673.582.249.879.2
HMA(MCD)Ours89.072.285.570.390.486.687.683.892.255.589.142.178.7
HMA(DAN)Ours87.549.280.253.881.871.887.857.660.957.085.332.867.1
HMA(CAN)Ours97.688.484.376.098.497.191.381.497.096.788.860.788.1
", + "image_path": "6175440619d57fa8d54da680eca3cc5d696155bb5a5d39343d0f95b90e1c0382.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 402, + 299, + 439 + ], + "blocks": [ + { + "bbox": [ + 47, + 361, + 287, + 395 + ], + "lines": [ + { + "bbox": [ + 47, + 361, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 287, + 395 + ], + "type": "text", + "content": "Table 5. Homeomorphism mapping vs ordinary bijection on Office-31. OB(CAN) means ordinary bijection sewed by the distribution matching method CAN." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 402, + 299, + 439 + ], + "lines": [ + { + "bbox": [ + 48, + 402, + 299, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 402, + 299, + 439 + ], + "type": "table", + "html": "
ComponentA→DA→WD→AW→AParameters
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.220992000
OB(CAN)90.3±0.391.7±0.276.6±0.175.8±0.233603584
", + "image_path": "8995bb3bb0dd3c4fc6965fb16221c59437656540da8b6d5292ae3aa5c18ddef9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 470, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 651 + ], + "type": "text", + "content": "ial learning strategy and statistic moment matching strategy because conditional distribution alignment can stitch homeomorphism mapping with two spaces by category. Furthermore, for condition distribution matching method, statistic moment matching strategy is better than adversarial learning strategy. The reason is that statistic moment matching strategy stitches homeomorphism mapping with two feature spaces explicitly by category. Interestingly, the alignment method CAN published in 2019 still achieves SOTA results. It can be seen that alignment by category is very important for extracting domain invariant features. Our method HMA(CAN) further boosts the CAN performance, because we realize the difficulty of extracting a domain invariant space and we do alignment in the two spaces by homeomorphism mapping." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 226, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 226, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 226, + 671 + ], + "type": "text", + "content": "5.3. Ablation Analysis and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "Homeomorphism map is better than ordinary bijection. As mentioned in Section 3, there does not exist many methods which do alignments in two feature spaces based on bi-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 380, + 553, + 485 + ], + "blocks": [ + { + "bbox": [ + 356, + 361, + 492, + 373 + ], + "lines": [ + { + "bbox": [ + 356, + 361, + 492, + 373 + ], + "spans": [ + { + "bbox": [ + 356, + 361, + 492, + 373 + ], + "type": "text", + "content": "Table 6. Ablation study on Office-31." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 380, + 553, + 485 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 553, + 485 + ], + "type": "table", + "html": "
ComponentA→DA→WD→AW→A
CAN95.0±0.394.5±0.378.0±0.377.0±0.3
INN(CAN)94.8±0.394.1±0.277.3±0.476.7±0.2
INN(CAN)+S2T95.6±0.294.9±0.378.9±0.277.4±0.2
HMA(CAN)95.8±0.395.1±0.179.3±0.377.6±0.2
DAN78.6±0.280.5±0.463.6±0.362.8±0.2
INN(DAN)78.3±0.379.9±0.262.9±0.262.6±0.1
INN(DAN)+S2T84.1±0.283.8±0.366.5±0.266.1±0.2
INN(DAN)+S2T+T2S85.1±0.284.5±0.267.9±0.366.7±0.3
", + "image_path": "9d39cf39792a097cd69fbc601f85c2ceb684f6d86b6fcef7460fbafcfa18120b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 520, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 639 + ], + "type": "text", + "content": "jection. In this experiment, we will apply the ordinary bijection method to top-4 challenging tasks on Office-31 dataset, the difference between homeomorphism implemented by INN and ordinary bijection is mainly reflected in topological structure maintenance. The results are shown in Table 5. It is obvious that homeomorphism mapping is superior to the ordinary bijection in both accuracy and model parameters. For the model size, the INN based homeomorphism mapping consumes around half of the parameters compared to using two neural networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": "Ablation study. To show the effectiveness of alignment in two spaces, we conduct an experiment on top-4 challenging tasks of Office-31 dataset. The results are shown in Table 6. The method CAN is considered as the baseline. INN(CAN) means just sewing up homeomorphism mapping to the two feature spaces and the source model is retrained based on" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18706" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "the source labels. INN(CAN) " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": " S2T means the transformed source features are used to learn the model in the target feature space compared with INN(CAN). HMA(CAN) uses all features in two spaces. From Table 6, we can find that simply using INN can achieve similar performance as the feature distribution alignment method CAN. As shown in the third row in Table 6, by transferring the source features to the target feature space, the performance is greatly improved. For this case, the data topological structure and label information from source domain can be correctly transformed to the target domain, which allows that the learned model works well in the target domain. The fourth row in Table 6 shows that the performance can be further improved if both transformed features are used. Because our homeomorphism mapping keeps the corresponding relationship by category, with the help of supervision information in the source feature spaces, the generalization performance of the model in the two domains is improved. To further verify our ideas, we test DAN as a sewing tool, giving the consistent conclusion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 324, + 153, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 153, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 153, + 335 + ], + "type": "text", + "content": "6. Proof of Theorem" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 343, + 160, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 160, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 160, + 355 + ], + "type": "text", + "content": "6.1. Proof of Theorem 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 361, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 287, + 387 + ], + "type": "text", + "content": "Theorem 1. Let " + }, + { + "bbox": [ + 47, + 361, + 287, + 387 + ], + "type": "inline_equation", + "content": "(M, d_M)" + }, + { + "bbox": [ + 47, + 361, + 287, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 361, + 287, + 387 + ], + "type": "inline_equation", + "content": "(N, d_N)" + }, + { + "bbox": [ + 47, + 361, + 287, + 387 + ], + "type": "text", + "content": " be two metric spaces with a homeomorphism" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 397, + 195, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 397, + 195, + 410 + ], + "spans": [ + { + "bbox": [ + 139, + 397, + 195, + 410 + ], + "type": "interline_equation", + "content": "g: M \\to N,", + "image_path": "8be1c3de5eefe0021d32fc2af87260cdfaf5fe7d3ea9460f504b37e5b50a45fb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "text", + "content": " is an open subset in " + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "inline_equation", + "content": "(M,d_M)" + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "text", + "content": " , we have that its image " + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "inline_equation", + "content": "B\\coloneqq g(A)" + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "text", + "content": " is an open subset in " + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "inline_equation", + "content": "(N,d_N)" + }, + { + "bbox": [ + 47, + 419, + 287, + 445 + ], + "type": "text", + "content": " , and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 454, + 217, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 454, + 217, + 468 + ], + "spans": [ + { + "bbox": [ + 116, + 454, + 217, + 468 + ], + "type": "interline_equation", + "content": "g (\\partial A) = \\partial B = \\partial g (A),", + "image_path": "d9f24cb2297ae392cf0d6fe13856117be9baaf894e4e4f237c1997bc95e80c18.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 478, + 167, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 167, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 167, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 478, + 167, + 489 + ], + "type": "inline_equation", + "content": "\\partial" + }, + { + "bbox": [ + 47, + 478, + 167, + 489 + ], + "type": "text", + "content": " means the boundary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 490, + 186, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 186, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 186, + 501 + ], + "type": "text", + "content": "Proof. It is sufficient to show that" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 512, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 512, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 136, + 512, + 287, + 525 + ], + "type": "interline_equation", + "content": "g (\\partial A) \\subset \\partial B. \\tag {7}", + "image_path": "c5890e4b8801d63406f14ef433be99f01d492401a3e4f2670dc37b1380595aff.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 536, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 558 + ], + "type": "text", + "content": "Assume this is true, then we can imply Eq. 7 to " + }, + { + "bbox": [ + 47, + 536, + 287, + 558 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 47, + 536, + 287, + 558 + ], + "type": "text", + "content": ", and obtain" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 559, + 202, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 202, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 202, + 573 + ], + "type": "interline_equation", + "content": "g ^ {- 1} (\\partial B) \\subset \\partial A.", + "image_path": "89b04931e6dd2ecc8f9089ca10d0ee861ea2b0bcddf07a230d41abcf823c80d2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 579, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 287, + 604 + ], + "type": "text", + "content": "Hence we have " + }, + { + "bbox": [ + 47, + 579, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\partial B\\subset g(\\partial A)" + }, + { + "bbox": [ + 47, + 579, + 287, + 604 + ], + "type": "text", + "content": " . Combing this with Eq. 7, we have " + }, + { + "bbox": [ + 47, + 579, + 287, + 604 + ], + "type": "inline_equation", + "content": "g(\\partial A) = \\partial B" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": "Now we want to show Eq. 7, that is, for any " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "x \\in \\partial A" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "g(x) \\in \\partial B" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "x \\in \\partial A" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": ", but " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "x \\notin A" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "g(x) \\notin B" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": ", and there is a sequence " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\{x_i\\} \\subset A" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\lim_{i \\to \\infty} x_i = x" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": ". By the continuity of the function " + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 604, + 287, + 652 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 98, + 662, + 235, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 662, + 235, + 680 + ], + "spans": [ + { + "bbox": [ + 98, + 662, + 235, + 680 + ], + "type": "interline_equation", + "content": "g (x) = g \\left(\\lim _ {i \\rightarrow \\infty} x _ {i}\\right) = \\lim _ {i \\rightarrow \\infty} g \\left(x _ {i}\\right).", + "image_path": "7bcb3344aa6df97af835296617bfb48e26e44810a36aae9fbc9be15df801e39b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Noting that " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "g(x_{i}) \\in B" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", we get " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "g(x) \\in \\partial B" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ". This completes the proof." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 72, + 421, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 421, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 421, + 83 + ], + "type": "text", + "content": "6.2. Proof of Theorem 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 91, + 545, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 114 + ], + "type": "text", + "content": "Theorem 2. Invertible Neural Network is a homeomorphism." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "text", + "content": "Proof. The definition of homeomorphism is that, a function " + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "inline_equation", + "content": "g: M \\to N" + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "text", + "content": " between two topological spaces is a homeomorphism if it has the following properties: (1) " + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "text", + "content": " is a bijection; (2) " + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "text", + "content": " is continuous; (3) The inverse function " + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 305, + 114, + 545, + 173 + ], + "type": "text", + "content": " is continuous." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": "For any invertible neural network, assuming that its forward process is " + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": ", then its invertible process can be represented as " + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": ", so " + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": " is a bijection. Because the function of each part of the invertible neural network is continuous, as " + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "s(\\cdot)" + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "t(\\cdot)" + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": " in our method, so both " + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 305, + 175, + 545, + 247 + ], + "type": "text", + "content": " are continuous. To sum up, the invertible network is a homeomorphism." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 258, + 378, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 378, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 378, + 270 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 277, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 469 + ], + "type": "text", + "content": "In this paper, we have proposed a new unsupervised domain adaptation method, termed as Homeomorphism Alignment in two spaces (HMA). By analyzing previous alignment based methods, we argue that it is difficult to find a common space or domain invariant space to adapt the pretrained source model. So the alignment is performed in two spaces. The extracted source and target features can be further transformed respectively by a homeomorphism mapping so that they can be aligned semantically. Our method consists of three steps, i.e., constructing an INN based homeomorphism mapping, sewing up by category and retraining iteratively training the model in two spaces. In this way, the source labels can be fully used even in the target feature space for improving the model generalization for the target domain. Extensive experimental results demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 480, + 410, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 410, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 410, + 495 + ], + "type": "text", + "content": "8. Acknowledgment" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 501, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 545, + 536 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (2018YFE0203900), National Natural Science Foundation of China (62276048)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 548, + 365, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 365, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 365, + 560 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 310, + 567, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 310, + 567, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 567, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 310, + 567, + 545, + 623 + ], + "type": "text", + "content": "[1] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3334–3342, October 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 311, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 545, + 689 + ], + "type": "text", + "content": "[2] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. *Diverse image style transfer via invertible cross-space mapping.* In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14880-14889, October 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "type": "text", + "content": "[3] Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, and Enhong Chen. Reusing the task-specific clas" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18707" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "sifier as a discriminator: Discriminator-free adversarial domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7181-7190, 2022. 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 119, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 119, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 119, + 288, + 163 + ], + "type": "text", + "content": "[4] Minghao Chen, Shuai Zhao, Haifeng Liu, and Deng Cai. Adversarial-learned loss for domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3521–3528, 2020. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 165, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 165, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 165, + 288, + 219 + ], + "type": "text", + "content": "[5] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International conference on machine learning, pages 1081-1090. PMLR, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 222, + 288, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 222, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 53, + 222, + 288, + 287 + ], + "type": "text", + "content": "[6] Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3941-3950, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 289, + 288, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 288, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 288, + 355 + ], + "type": "text", + "content": "[7] Bharath Bhushan Damodaran, Benjamin Kellenberger, Rémi Flamary, Devis Tuia, and Nicolas Courty. Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 447-463, 2018. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 358, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 391 + ], + "type": "text", + "content": "[8] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 393, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 288, + 426 + ], + "type": "text", + "content": "[9] Geoff French, Michal Mackiewicz, and Mark Fisher. Self-ensembleing for visual domain adaptation. In International Conference on Learning Representations, 2018. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "type": "text", + "content": "[10] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 518 + ], + "type": "text", + "content": "[11] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 564 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 4, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[14] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[15] Jiaxing Huang, Dayan Guan, Aoran Xiao, Shijian Lu, and Ling Shao. Category contrast for unsupervised domain adaptation in visual tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1203-1214, 2022. 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In International Conference on Machine Learning, pages 4816-4827. PMLR, 2020. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[17] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4733-4742, October 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[18] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4893-4902, 2019. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[19] Younggeun Kim and Donghee Son. Noise conditional flow model for learning the super-resolution space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 424-432, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Prafulla Dhariwal. Glow: generative flow with invertible " + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": " convolutions. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 10236-10245, 2018. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "text", + "content": "[21] Sajja Tulasi Krishna and Hemantha Kumar Kalluri. Deep learning and transfer learning approaches for image classification. International Journal of Recent Technology and Engineering (IJRTE), 7(5S4):427-432, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[22] Chen-Yu Lee, Tanmay Batra, Mohammad Haris Baig, and Daniel Ulbricht. Sliced Wasserstein discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10285-10295, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "text", + "content": "[23] Mengxue Li, Yi-Ming Zhai, You-Wei Luo, Peng-Fei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13936–13944, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "text", + "content": "[24] Shuang Li, Fangrui Lv, Binhui Xie, Chi Harold Liu, Jian Liang, and Chen Qin. Bi-classifier determinacy maximization for unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8455-8464, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "text", + "content": "[25] Shuang Li, Mixue Xie, Kaixiong Gong, Chi Harold Liu, Yulin Wang, and Wei Li. Transferable semantic augmentation for domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11516-11525, 2021. 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Domain adaptation with auxiliary target domain-oriented classifier. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2021. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[27] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchi" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18708" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "cal conditional flow: A unified framework for image superresolution and image rescaling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4076-4085, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "text", + "content": "[29] Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 469-477, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 227 + ], + "type": "text", + "content": "[30] Xiaofeng Liu, Zhenhua Guo, Site Li, Fangxu Xing, Jane You, C.-C. Jay Kuo, Georges El Fakhri, and Jonghye Woo. Adversarial unsupervised domain adaptation with conditional and label shift: Infer, align and iterate. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10367-10376, October 2021. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[31] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 315 + ], + "type": "text", + "content": "[32] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. In Advances in neural information processing systems, 2018. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 370 + ], + "type": "text", + "content": "[34] James R. Munkres. Topology (2nd edition). In Prentice Hall, 2000. 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 372, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 404 + ], + "type": "text", + "content": "[35] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, pages 1345-1359, 2009. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "text", + "content": "[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.visda06924, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "text", + "content": "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 505, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 505, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 505, + 287, + 547 + ], + "type": "text", + "content": "[38] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. European conference on computer vision, pages 213-226. Springer, 2010. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[39] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3723-3732, 2018. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "type": "text", + "content": "[40] Astuti Sharma, Tarun Kalluri, and Manmohan Chandraker. Instance level affinity-based transfer for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5361-5371, 2021. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[41] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2107-2116, 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[42] Baochen Sun, Jiashi Feng, and Kate Saenko. Correlation alignment for unsupervised domain adaptation. In Domain Adaptation in Computer Vision Applications, pages 153-171. Springer, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 119, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 152 + ], + "type": "text", + "content": "[43] Yu Sun, Eric Tzeng, Trevor Darrell, and Alexei A Efros. Unsupervised domain adaptation through self-supervision. arXiv preprint arXiv:1909.11825, 2019. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "text", + "content": "[44] Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8725-8735, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 201, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 545, + 243 + ], + "type": "text", + "content": "[45] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 246, + 545, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 290 + ], + "type": "text", + "content": "[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 293, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 347 + ], + "type": "text", + "content": "[47] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 404 + ], + "type": "text", + "content": "[48] Guoqiang Wei, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Metaalign: Coordinating domain alignment and classification for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16643-16653, 2021. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 406, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 450 + ], + "type": "text", + "content": "[49] Ni Xiao and Lei Zhang. Dynamic weighted learning for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15242-15251, 2021. 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 453, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 545, + 506 + ], + "type": "text", + "content": "[50] Renjun Xu, Pelen Liu, Liyan Wang, Chao Chen, and Jindong Wang. Reliable weighted optimal transport for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4394-4403, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "text", + "content": "[51] Tongkun Xu, Weihua Chen, Pichao Wang, Fan Wang, Hao Li, and Rong Jin. Cdtrans: Cross-domain transformer for unsupervised domain adaptation. arXiv preprint arXiv:2109.06165, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[52] Jingyi Zhang, Jiaxing Huang, Zichen Tian, and Shijian Lu. Spectral unsupervised domain adaptation for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9829-9840, 2022. 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 612, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 657 + ], + "type": "text", + "content": "[53] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[54] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18709" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[55] Bo Zhao, Jiashi Feng, Xiao Wu, and Shuicheng Yan. A survey on deep learning-based fine-grained object classification and semantic segmentation. International Journal of Automation and Computing, 14(2):119-135, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "text", + "content": "[56] Lihua Zhou, Mao Ye, Xiatian Zhu, Shuaifeng Li, and Yiguang Liu. Class discriminative adversarial learning for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4318-4326, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 228 + ], + "type": "text", + "content": "[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18710" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_content_list.json b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84e9df860f13134a3b3412c3564b9c012b455fc8 --- /dev/null +++ b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_content_list.json @@ -0,0 +1,1610 @@ +[ + { + "type": "text", + "text": "Homography Guided Temporal Fusion for Road Line and Marking Segmentation", + "text_level": 1, + "bbox": [ + 147, + 128, + 823, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shan Wang $^{1,2}$ Chuong Nguyen $^{1}$ Jiawei Liu $^{2}$ Kaihao Zhang $^{2}$ Wenhan Luo $^{3}$ Yanhao Zhang $^{2}$ Sundaram Muthu $^{1}$ Fahira Afzal Maken $^{1}$ Hongdong Li $^{2}$ $^{1}$ Data61, CSIRO $^{2}$ Australian National University $^{3}$ Sun Yat-sen University", + "bbox": [ + 151, + 202, + 815, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reliable segmentation of road lines and markings is critical to autonomous driving. Our work is motivated by the observations that road lines and markings are (1) frequently occluded in the presence of moving vehicles, shadow, and glare and (2) highly structured with low intra-class shape variance and overall high appearance consistency. To solve these issues, we propose a Homography Guided Fusion (HomoFusion) module to exploit temporally-adjacent video frames for complementary cues facilitating the correct classification of the partially occluded road lines or markings. To reduce computational complexity, a novel surface normal estimator is proposed to establish spatial correspondences between the sampled frames, allowing the HomoFusion module to perform a pixel-to-pixel attention mechanism in updating the representation of the occluded road lines or markings. Experiments on ApolloScape, a largescale lane mark segmentation dataset, and ApolloScape Night with artificial simulated night-time road conditions, demonstrate that our method outperforms other existing SOTA lane mark segmentation models with less than $9\\%$ of their parameters and computational complexity. We show that exploiting available camera intrinsic data and ground plane assumption for cross-frame correspondence can lead to a light-weight network with significantly improved performances in speed and accuracy. We also prove the versatility of our HomoFusion approach by applying it to the problem of water puddle segmentation and achieving SOTA performance1.", + "bbox": [ + 75, + 323, + 473, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lane mark segmentation aims to achieve pixel-wise classification of road lines and markings simultaneously. Known for its quintessential importance in autonomous driving, lane mark segmentation is also an effective tool for constructing accurate High Definition (HD) maps. Existing works solve individual sub-tasks, e.g. drivable area segmen", + "bbox": [ + 75, + 787, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebb967392b1de54130f4ddb8f633a65435ddcef2c0f073ca19f88caf61981358.jpg", + "image_caption": [ + "Figure 1: Illustration of the effect of the proposed HomoFusion module that explores the adjacent frames for cues, facilitating the correct classification of (1) a \"Straight Arrow\", which, with its bottom half occluded by a vehicle, is mistakenly classified as a \"Right Turn & Straight Arrow\" without the HomoFusion module, and (2) a partially occluded \"Dotted Line\", which is incorrectly classified as a \"Solid Line\" without the HomoFusion module. The fused frame in the $4^{th}$ row and the $2^{nd}$ column demonstrates the recovered road lines and markings after projecting the previous frames onto the current frame with the estimated homography matrices. The yellow box enlarges the area where mistake classifications are corrected. The red box indicates the spatially corresponding area across the frames. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 504, + 291, + 890, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tation [33, 9] and lane detection [19, 35], while few address the lane mark segmentation in its entirety [53, 16].", + "bbox": [ + 496, + 733, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the tremendous progresses in semantic/scene segmentation [52, 42, 26], little attention has been paid to the lane mark segmentation task [53, 16]. Yin et al. [53] leverages additional LiDAR information, merging the segmented visual information with the point clouds, to achieve lane mark segmentation. IntRA-KD [16] applies the knowledge distillation technique to improve the efficiency of the lane mark segmentation model.", + "bbox": [ + 496, + 763, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A major challenge in the lane mark segmentation task is", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at https://github.com/ShanWang-Shan/HomoFusion.", + "bbox": [ + 94, + 886, + 467, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1075", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "partial occlusion of the road lines and markings caused by the vehicle and surrounding environment leading to false classifications. For example, a partially occluded straight arrow can be easily mistaken as a right turn & straight arrow, as depicted in Fig. 1. However, existing lane mark segmentation methods have not yet made use of the following observations: (1) occlusion of the road lines and markings (due to nearby moving vehicles, shadow, and glare) can be reduced by cross-frame consistency; (2) road lines and markings are highly structured, maintaining high consistency in intra-class shapes as well as in overall appearance, alleviating the requirements on learning complicated features to distinguish different classes with high intra-class variance as in semantic/scene segmentation.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To find complementary information from adjacent frames, we use the ground plane assumption for the road region immediately in front of the camera. Our Homography Guided Fusion (HomoFusion) module achieves temporally consistent representation, recovers partially occluded road lines or markings, and leads to correct classification. The module's success depends on accurate cross-frame spatial correspondence, which we achieve using a homography transformation matrix estimated with available intrinsic and extrinsic parameters of the on-vehicle camera, and the road surface normal estimated by our novel optimization method called Road Surface Normal Estimator (RSNE).", + "bbox": [ + 75, + 301, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using the highly structured nature of road lines and markings could reduce the complexity of the detection problem and make it suitable to run on edge devices. To this end, we employ (1) a lightweight encoder to represent the visual information in feature spaces, and (2) a cross-frame pixel-to-pixel attention mechanism in our HomoFusion module, instead of a more computationally expensive global attention mechanism employed by other methods. These design decisions allow the proposed model to outperform the existing lane mark segmentation models with less than $9\\%$ of their parameters and Giga Floating Point Operations (GFLOPs) We summarize our contributions as follows:", + "bbox": [ + 75, + 484, + 472, + 680 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a HomoFusion module that uses ground plane assumption and the adjacent frames for temporally consistent representations for accurate classification of partially occluded road lines and markings.", + "- We design a novel estimator RSNE for road surface normal, which, combined with camera intrinsic and extrinsic parameters readily available on an autonomous vehicle, yields accurate homography matrix between frame pairs. RSNE simplifies the 8 degrees of freedom (DoF) of homography problem to a 2 DoF of normal vector problem.", + "- We present a lightweight lane mark segmentation model that achieves better performance than the state" + ], + "bbox": [ + 94, + 686, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of-the-art (SOTA) methods with significantly reduced model complexity and computation requirements.", + "bbox": [ + 531, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 135, + 640, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Lane Detection", + "text_level": 1, + "bbox": [ + 500, + 160, + 653, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional lane detection approaches rely on handcrafted features such as color [44, 51], edge [23] and texture [25], which are limited in complex scenarios [48, 7]. Recent advances in deep neural networks have led to significant performance improvements, with methods like message-passing networks [36] and attention-based modules [45]. [16, 37, 38, 54, 27] formulate lane detection as a row-wise classification task based on grid division of the input image. PolyLaneNet [46] is the first parametric prediction method, which outputs polynomials to represent each lane. BEV-Former [24] and PETRv2 [28] employ transformer mechanisms to streamline the conversion of perspective views into bird's-eye views, thereby eliminating variations in object size caused by perspective. Although the above methods achieve impressive performance, all of them focus on detecting the lane lines and ignore the basic road elements like arrow signal in Fig. 1. Furthermore, lane detection differs from segmentation tasks in that it focuses on identifying the boundaries of the drivable area, rather than the real shape of the lanes.", + "bbox": [ + 496, + 183, + 893, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Road Line and Marking Segmentation", + "text_level": 1, + "bbox": [ + 500, + 496, + 831, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The identification of road elements is crucial for ensuring safety in autonomous driving systems, but there are few works addressing this problem. Hou et al. [16] introduced a distillation approach that demonstrates competitive performance in lane mark segmentation. Yin et al. [53] used an LSTM-based network to segment images with DeepLabv3+ [8] and then merged them with point clouds to assist with lane segmentation. Unfortunately, these methods do not address the challenges of occlusion, shadows, and glare that frequently occur in real-world driving scenarios. Recently, Zhang et al. [56] introduced global memory information from previous frames to enhance local information for video lane segmentation. However, this method introduces attention to the entire image, making it memory-intensive and inefficient for autonomous driving scenarios.", + "bbox": [ + 496, + 518, + 893, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Homography Estimation", + "text_level": 1, + "bbox": [ + 500, + 756, + 725, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Homography estimation methods can be divided into non-deep and deep learning-based approaches. Non-deep methods estimate the homography using feature extraction, feature matching, and outlier rejection. SIFT [29], SURF [5, 4], ORB [39], LPPM [30], GMS [6], and BEBLID [41] are commonly used for feature extraction, and RANSAC [12], MAGAC [2], and LRLS [15] are applied for outlier rejection. Recently, many deep learning-based approaches", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1076", + "bbox": [ + 483, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/043b4b782b333726fa0cdd0c646fe96ea82b98dda31efae2f4d8de990b5b8b36.jpg", + "image_caption": [ + "Figure 2: Overview of our proposed model consisting of a pair of lightweight encoder and decoder, our proposed HomoFusion module, and our proposed Road Surface Normal Estimator (RSNE). A sequence of frames $\\mathbf{I}$ , including a target frame $\\mathbf{I}_{\\mathbf{t}}$ and $n - 1$ previous frames, are encoded into the feature representations $(\\mathbf{F}^l)$ . RNSE estimates the road surface normal vector, which, combined with the camera intrinsic and extrinsic parameters, yields a homography matrix between each frame pair, establishing cross-frame spatial correspondences. HomoFusion uses pixel-to-pixel attention mechanism to obtain temporally consistent representation for on-road pixels of the current frame with the spatial correspondence across frames as guidance. Finally, the decoder decodes and upsamples the temporally consistent feature representations to produce the lane mark segmentation prediction $(\\mathbf{S}_{\\mathbf{t}}^{\\mathrm{prd}})$ ." + ], + "image_footnote": [], + "bbox": [ + 158, + 85, + 803, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have been proposed, such as DeTone et al. [10]'s VGG-like architecture, Nowruzi et al. [11], Le et al. [20]'s and Man et al. [31]'s cascaded VGG-like networks, and Chang et al. and Zhao et al.'s incorporation of the Lucas-Kanade (LK) algorithm with deep networks. However, these methods do not utilize known camera intrinsic and extrinsic parameters across frames for homography matrix estimation, which can effectively reduce the search dimensions.", + "bbox": [ + 75, + 458, + 472, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 595, + 169, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 619, + 187, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, our proposed framework takes a sequence of frames, denoted as $\\mathbf{I} = \\{\\mathbf{I}_i|i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}$ , as input. This sequence consists of a current frame $\\mathbf{I}_t$ and $n - 1$ previous frames sampled at a fixed time interval $\\Delta t$ . The output of the framework is a lane mark segmentation map $\\mathbf{S}_t$ for the current frame. In addition, the framework includes a novel HomoFusion module, which uses the homography transformation guided by the proposed RSNE to fuse the feature map of the current frame with those of the previous frames. We denote the encoded feature representation for the $t^{th}$ frame as $\\mathbf{F}_t = \\{\\mathbf{F}_t^l\\in \\mathbb{R}^{H_l\\times W_l\\times C_l}\\}_{l = 1}^L$ where $l$ indicates the", + "bbox": [ + 75, + 643, + 468, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "level of the feature map. The shallowest and deepest levels are represented by 1 and $L$ , respectively. $H_{l}$ , $W_{l}$ , and $C_l$ represent the height, width, and channel number of feature maps in level $l$ . We denote all encoded features as $\\mathbf{F} = \\{\\mathbf{F}_i^l | i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}$ . The proposed HomoFusion module, detailed in Sec. 3.2, uses a cross-frame pixel-to-pixel attention mechanism to fuse the feature map $(\\mathbf{F}_f = \\{\\mathbf{F}_f^l\\}_{l = 1}^L)$ of the current frame with those of the previous frames. The homography transformation matrix between the target frame and each of the previous frames is calculated using the estimated normal vector of the road surface $\\mathbf{n}$ obtained by the proposed RSNE, as described in Sec. 3.3. Finally, the decoder produces the segmentation prediction $(\\mathbf{S}_t)$ for the target frame based on the fused feature map.", + "bbox": [ + 496, + 458, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Homography Guided Fusion (HomoFusion)", + "text_level": 1, + "bbox": [ + 498, + 695, + 870, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed HomoFusion module employs a pixel-to-pixel attention mechanism to fuse spatially corresponding pixels across frames. This mechanism is achieved by projecting the pixels of the current frame onto the previous frames through an accurate homography transformation. Since road lines and markings are strictly on-road, the search area can be limited to the road surface, which is mostly a plane, at least within the immediate front of the vehicle where markings are readable. Therefore, only the homography transformation is needed to accurately map road pixels between frames. According to the standard inverse homography [14, 57, 49], for each on-road pixel $p_t =$", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2In our experiments, both camera intrinsic and extrinsic parameters are provided by the dataset. In real-world scenarios, these parameters can be obtained through camera calibration and autonomous vehicle pose obtained from the pose estimation framework, which is already known in the in-car system and free of cost.", + "bbox": [ + 75, + 838, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1077", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2a75d7d795a37273ec68a561ec4cc0ba9232f02ea232f8f195b482d5ce704e33.jpg", + "image_caption": [ + "Figure 3: Illustration of sample points. (Right) Sample points in the current frame. (Left/Middle) Correspondence of a sample point in previous frames. The red point coordinate is calculated by using the correct normal, while the cyan point coordinate is calculated by using the initial (incorrect) normal." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(u,v)^{\\top}$ of the current frame, its spatially corresponding pixels of the reference frames $\\{p_i|i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}$ can be computed using Eq. (1).", + "bbox": [ + 76, + 281, + 470, + 329 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} \\propto \\mathbf {K} \\left(\\mathbf {R} _ {i} - \\frac {\\mathbf {t} _ {i} \\mathbf {n} ^ {\\top}}{d}\\right) \\mathbf {K} ^ {- 1} \\left(p _ {t} \\oplus 1\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 339, + 468, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\propto$ represents proportional, $\\mathbf{K}$ is the intrinsic matrix of the on-vehicle camera, $\\mathbf{R}_i$ and $\\mathbf{t}_i$ represent the relative rotation and translation between the current frame $\\mathbf{I}_t$ and the $i^{th}$ previous frame $\\mathbf{I}_i$ . $\\mathbf{n}$ is the normal of the observed road surface, which is estimated in Sec. 3.3. $d$ is the vertical distance between the on-vehicle camera and the road surface obtained from the camera calibration. $\\oplus 1$ converts the point coordinates to homogeneous coordinates. The on-road pixels of the current frame are sampled from a triangular area in front of the car, as illustrated in Fig. 3. Figure 3 also demonstrates that the shape and category of a left-turn can be recovered by exploring the spatially corresponding area in the previous frames under the guidance of an accurate homography transformation.", + "bbox": [ + 75, + 371, + 468, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the spatial correspondence across frames, the proposed HomoFusion module computes a temporally consistent representation for each sampled pixel of the current frame by employing temporal attention mechanism on the feature presentations of spatially corresponding pixels. The mechanism assigns the pixel representation of current frame $(\\mathbf{F}_t[p_t])$ as a query and all frames $(\\{\\mathbf{F}_i[p_i]\\})$ as keys and values. In contrast to the existing attention mechanism [50], we propose to (1) omit the spatial encoding in the presence of obtained spatial correspondence across frames, and (2) apply an L2 normalization on the query and the keys to obtain more robust (deep) feature similarities (Eq. 2) across various lighting environments.", + "bbox": [ + 75, + 583, + 470, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} _ {i} = \\frac {\\mathbf {F} _ {t} [ p _ {t} ]}{\\| \\mathbf {F} _ {t} [ p _ {t} ] \\| _ {2}} \\cdot \\frac {\\mathbf {F} _ {i} [ p _ {i} ]}{\\| \\mathbf {F} _ {i} [ p _ {i} ] \\| _ {2}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 791, + 468, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\| .\\| _2$ indicates an $L_{2}$ normalization. The similarity is normalized by softmax function as given in Eq. 3:", + "bbox": [ + 76, + 830, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {i} = \\frac {\\exp \\mathbf {a} _ {i}}{\\sum_ {i} \\exp \\mathbf {a} _ {i}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 873, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "before being used as weight to fuse the spatially corresponding pixels across the frames as shown in Eq. 4:", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {f} [ p _ {t} ] = \\mathbf {F} _ {t} [ p _ {t} ] + \\sum_ {i} \\mathbf {W} _ {i} \\mathbf {F} _ {i} [ p _ {i} ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 132, + 890, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Road Surface Normal Estimator (RSNE)", + "text_level": 1, + "bbox": [ + 500, + 172, + 849, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accurate estimation of road surface normal is crucial for establishing spatial correspondence across frames. Generally, on-vehicle cameras have a fixed angle with respect to the ground level. However, the actual road surface normal can vary due to the unevenness of the road surface, e.g. uphill or downhill roads, and sloped turns, etc. Assuming that the initial road surface normal is a vector being perpendicular to a horizontal plane and pointing upwards, we propose to obtain an accurate road surface normal with an optimization process that repeats for each new coming frame. The proposed RSNE iteratively updates the estimated road surface normal with the projection error of on-road pixels from the current frame onto the previous frames. Specifically, we sample $m$ pixels $\\{p_t^j\\}_{j=1}^m$ from the current frame that are likely to be on-road, as illustrated in Fig. 3, and compute the spatially corresponding pixels $\\{p_i^j\\}_{j=1}^m$ in the previous frames using the homography matrix computed with the estimated road surface normal as defined in Eq. 1. The estimated road surface normal is optimized in an iterative manner through the following procedure: (a) computing a residual between the low-level features of spatially corresponding pixels across the frames with Eq. 5", + "bbox": [ + 496, + 196, + 890, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {i} ^ {j} = \\mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ] - \\mathbf {F} _ {t} ^ {1} [ p _ {t} ^ {j} ] \\in \\mathbb {R} ^ {c}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 539, + 890, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the residual is set to 0 for the sample point whose projections in previous frames fall outside the frame area, and computing the overall error with Eq. 6", + "bbox": [ + 498, + 566, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} = \\sum_ {i, j} \\rho \\left(\\left\\| \\mathbf {r} _ {i} ^ {j} \\right\\| _ {2} ^ {2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 623, + 890, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\| .\\| _2^2$ shows squared norm and $\\rho (\\cdot)$ is a robust cost function proposed by Barron [3]. This nonlinear least-square cost is iteratively minimized towards correct normal estimation using the Levenberg-Marquardt algorithm [21, 32]; (b) computing a new road surface normal value by adding the update given by Eq. 12.", + "bbox": [ + 496, + 660, + 890, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The road surface normal vector is a unit vector with two degrees of freedom (DoF), which can be decomposed into a pitch angle $\\theta$ and a roll angle $\\phi^3$ , as shown in Eq. 7.", + "bbox": [ + 496, + 751, + 890, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} = \\left(- \\sin \\phi \\cos \\theta , - \\cos \\phi \\cos \\theta , \\sin \\theta\\right) ^ {\\top}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 806, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Jacobian of the residual function w.r.t. the pitch angle and the roll angle is defined in Eq. 8 by applying the chain", + "bbox": [ + 498, + 835, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3An explanation of the pitch angle $\\theta$ and roll angle $\\phi$ is provided in the appendix.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1078", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rule,", + "bbox": [ + 76, + 92, + 112, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} _ {i} ^ {j} = \\frac {\\partial \\mathbf {r} _ {i} ^ {j}}{\\partial \\delta} = \\frac {\\partial \\mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ]}{\\partial p _ {i} ^ {j}} \\frac {\\partial p _ {i} ^ {j}}{\\partial \\mathbf {n}} \\frac {\\partial \\mathbf {n}}{\\partial \\delta}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 102, + 468, + 141 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ represents the to-be-optimized target ( $\\theta$ or $\\phi$ ), and $\\frac{\\partial \\mathbf{F}_i^1[p_i^j]}{\\partial p_i^j}$ is gradient from 2D interpolation. $\\frac{\\partial p_i^j}{\\partial \\mathbf{n}}$ is the Jacobian of the homography transformation (Eq. (1)) w.r.t. $\\mathbf{n}$ , as defined in Eq. 9", + "bbox": [ + 76, + 145, + 468, + 214 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial p _ {i} ^ {j}}{\\partial \\mathbf {n}} = - \\frac {1}{d} \\mathbf {K t} _ {i} \\left(\\mathbf {K} ^ {- 1} \\left(p _ {t} ^ {j} \\oplus 1\\right)\\right) ^ {\\top}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 224, + 468, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\frac{\\partial\\mathbf{n}}{\\partial\\delta}$ , defined in Eq. 10 and Eq. 11, is the Jacobian of Eq. (7) w.r.t. $\\theta$ and $\\phi$ respectively", + "bbox": [ + 76, + 268, + 468, + 300 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathbf {n}}{\\partial \\theta} = \\left(- \\frac {\\mathbf {n} _ {1} \\mathbf {n} _ {3}}{\\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}}, - \\frac {\\mathbf {n} _ {2} \\mathbf {n} _ {3}}{\\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}}, \\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 309, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathbf {n}}{\\partial \\phi} = \\left(\\mathbf {n} _ {2}, - \\mathbf {n} _ {1}, 0\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 354, + 468, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{n}_k$ represents the $k$ -th element of $\\mathbf{n}$ . The detailed derivation is included in appendix (Sec. 1). We obtain Hessian matrices as $\\mathbb{H} = \\mathbf{J}^{\\top}\\rho^{\\prime}\\mathbf{J}$ , where $\\rho^\\prime$ is the derivative of function $\\rho (\\cdot)$ . We compute the update by damping the Hessian and solving the linear system, as shown in", + "bbox": [ + 76, + 391, + 468, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\triangle \\delta = - (\\mathbb {H} + \\lambda \\operatorname {d i a g} (\\mathbb {H})) ^ {- 1} \\mathbf {J} ^ {\\top} \\rho^ {\\prime} \\mathbf {r}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 477, + 468, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the damping factor [40], balancing between the Gauss-Newton $(\\lambda = 0)$ and gradient descent $(\\lambda = \\infty)$ . Our entire optimization process is differentiable.", + "bbox": [ + 76, + 506, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The algorithm is illustrated in Alg. 1. The normal vector $\\mathbf{n}$ uses low-level features $\\mathbf{F}^1$ . While the high-level features $\\mathbf{F}^2$ focus on segmentation feature expression, the low-level features $\\mathbf{F}^1$ are supervised to extract useful features for both segmentation and road surface normal estimation. As both tasks focus on on-road objects, they can benefit each other.", + "bbox": [ + 76, + 551, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 656, + 209, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 681, + 294, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a pre-trained EfficientNet [47] to extract image features at two different scales, $4 \\times$ and $16 \\times$ down-scaling, with 64 and 128 channels, respectively. To focus on the bottom region of the input image, where road marks are typically visible, we crop only the bottom $40\\%$ of the input image, following the same approach in the prior work [16]. The processed image size is set to $272 \\times 848$ . Our decoder consists of bi-linear up-sampling layers and convolution layers, which up-sample the high-level features and increase the resolution by a factor of 4 at each level. We trained our model using the AdamW optimizer [55] for 30,000 iterations on two NVIDIA RTX 3090 GPUs, with a learning rate of $4 \\times 10^{-3}$ .", + "bbox": [ + 76, + 704, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Normal Optimization" + ], + "code_body": "Input: \nThe 1-st Level Feature Maps: $\\mathbf{F}^1 = \\{\\mathbf{F}_i^1\\}$ \nSample Points Coordinates: $p_t = \\{p_t^j\\}_{j = 1}^m$ \nInitial Pitch and Roll Value: $(\\theta_0,\\phi_0) = (0.15,0)$ \nDamping Factor: $\\lambda = (\\lambda_{\\theta}\\lambda_{\\phi})$ \nHyper-parameter: \nk; $\\triangleright$ Maximum loop count, empirically set to 20 \n $\\alpha$ ; $\\triangleright$ Convergence threshold, empirically set to 0.0001 \nOutput: Optimized Pitch and Roll $(\\theta ,\\phi)$ . \n1: function OPT( $\\theta_0,\\phi_0,\\mathbf{F}^1,p_t,\\lambda$ \n2: Derive point features $\\mathbf{F}_t^1 [p_t]$ from $\\mathbf{F}_t^1$ . \n3: for $k\\gets 1$ to k do \n4: Calculate point coordinates $p_i$ in $\\mathbf{F}_i^1$ (Eq. 1); \n5: Derive point features $\\mathbf{F}_i^1[p_i]$ from $\\mathbf{F}_i^1$ . \n6: Calculate residual $\\mathbf{r}_i$ (Eq. 5); \n7: Calculate observe error E (Eq. 6); \n8: Calculate robust cost $\\rho (\\mathbf{E})$ and its derivation $\\rho^\\prime$ \n9: Construct J and its Hessian matrices $\\mathbb{H}$ (Eq. 8); \n10: Obtain $\\triangle \\delta$ by Cholesky decomposition (Eq. 12); \n11: Update Normal as $(\\theta ,\\phi)_k\\gets (\\theta ,\\phi)_{k - 1} + \\triangle \\delta$ \n12: if MAX(△δ)< α then \n13: Break; \n14: end if \n15: end for \n16: end function", + "bbox": [ + 500, + 108, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 500, + 549, + 601, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct our experiments on two datasets, ApolloScape [17] and ApolloScape Night. Other datasets were deemed unsuitable due to the absence of camera extrinsics, as observed in SDLane [18] and VLI-100 [56], inadequate frame overlap, or the lack of diverse segmentation labels, exemplified by Waymo Open Dataset [43]. Further elaboration on these reasons can be found in the appendix.", + "bbox": [ + 496, + 566, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ApolloScape. The ApolloScape dataset is a large-scale dataset that can be used for localization and segmentation tasks. It consists of 38 distinct classes and poses various challenges including occlusions and tiny road markings. To provide accurate camera poses, the vehicle is equipped with customer-grade GPS/IMU. However, for the lane mark segmentation task, only 41, 201 annotated images out of more than $110,000+$ are associated with camera pose information. We used these 41, 201 images for our experiments since our approach relies on camera poses. We divided them into 35, 173 training images and 6, 028 validation images. Because our approach uses sequential information, we select the validation set with completely different trajectories to ensure a fair comparison.", + "bbox": [ + 496, + 672, + 892, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ApolloScape Night. Since there is a lack of datasets for", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1079", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0b3f7423ba7b6bd15d299ce3d9e592bf807345ab40298c7505e5823920175334.jpg", + "table_caption": [ + "Table 1: Comparison with SOTA Methods on the ApolloScape [17] and Apolloscape Night Datasets and Running on an NVIDIA RTX 3090 GPU. \"18 mIoU\" Represents the Mean IoU of 18 Types of Lane Markers Selected by ApolloScape Official Metrics. \"36 mIoU\" Represents the Mean IoU of All Unignorable 36 Types." + ], + "table_footnote": [], + "table_body": "
MethodsFrame CountBackboneParams (M)↓GFLOPs↓FPS (f/s)↑ApolloScape [17]ApolloScape Night
18 mIoU↑36 mIoU↑18 mIoU↑36 mIoU↑
IntRA-KD[16]1ResNet-10165.65159.410.842.124.629.816.7
SegFormer[52]1MiT-B113.51048.843.852.332.138.323.1
CFFM [42]4MiT-B115.31192.622.753.232.739.123.6
MMA-Net [56]4ResNet-5057.9723.220.652.931.438.823.2
HomoFusion(ours)4EfficientNet-B61.2461.225.459.335.944.826.6
", + "bbox": [ + 76, + 138, + 893, + 232 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5bc9f79a72ab999a35f2820e2b4cc188e81285273cf6ba6e6044218e683cb632.jpg", + "image_caption": [ + "Figure 4: Sample images from the ApolloScape Night dataset. Top: original daytime images from the ApolloScape dataset. Bottom: synthesized night-time images." + ], + "image_footnote": [], + "bbox": [ + 81, + 234, + 467, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "night-time lane mark segmentations, we created an artificial dataset called ApolloScape Night from the ApolloScape dataset using a cross-domain generation network [1]. This allows us to evaluate our proposed model on a challenging dataset with poor lighting conditions, road reflection, and glare. Fig. 4 displays some sample images from the dataset.", + "bbox": [ + 75, + 500, + 468, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 599, + 261, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In accordance with the guidelines of the ApolloScape benchmark [17], we used mean intersection-over-union (mIoU) as the evaluation criterion. ApolloScape contains 38 different types of lane markers, including two ignorable labels (noise and ignored). Of these labels, 18 categories are used in the official evaluation metrics. To provide a comprehensive evaluation of our approach, we report the evaluated mIoU results for both the selected 18 categories, denoted as \"18 mIoU\", and all 36 categories, denoted as \"36 mIoU\". By reporting both sets of results, we provide a more complete picture of the performance of our method on the ApolloScape and ApolloScape Night datasets.", + "bbox": [ + 75, + 625, + 470, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 816, + 441, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance of our proposed method, we compared it with SOTA algorithms, including (a) IntRA-KD [16], (b) SegFormer [52], (c) CFFM [42], and (d) MMA-Net [56], on the ApolloScape [17] and ApolloScape", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Night datasets. To ensure a fair comparison, we retrained each model on the same subset of the training set with the same input resolution.", + "bbox": [ + 496, + 238, + 890, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results of the comparison are presented in Table 1. Our approach outperforms the existing SOTA methods on both the ApolloScape [17] and ApolloScape Night datasets while having less than $9\\%$ of their parameters or computational overhead.", + "bbox": [ + 496, + 285, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we compared our proposed method with CFFM [42] which also uses adjacent frames to enhance the representation of the current frame. Our method is more efficient in terms of model complexity, with a complexity of $\\mathcal{O}(HWC)$ for feature extraction and $\\mathcal{O}(HWC^2)$ for cross-frame feature fusion. In contrast, the complexity of CFFM is $\\mathcal{O}(H^2 W^2 C) + \\mathcal{O}(HWC^2)$ for feature extraction and $\\mathcal{O}(HWEC) + \\mathcal{O}(HWC^2)$ for cross-frame feature fusion, where $E$ is calculated by their receptive field and pooling kernel size. This explains why our method has a lower computational requirement for the cross-frame pixel-to-pixel attention mechanism compared to the global attention mechanism used by CFFM.", + "bbox": [ + 496, + 362, + 892, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative results of both the existing SOTA models and our proposed model are shown in Fig. 5. The results indicate that our approach achieves superior segmentation performance on road lines and markings, even under adverse conditions such as occlusion, road reflection, and poor lighting. Additional visualization examples for various categories can be found in the appendix.", + "bbox": [ + 496, + 559, + 890, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Study on Hyper-parameters", + "text_level": 1, + "bbox": [ + 500, + 678, + 751, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Number of Frames $(n)$ . We conducted a study on the impact of the number of frames while setting the sampling frame gap $\\triangle t = 2$ . The results presented in Fig. 6 (a)-(b) demonstrate that the performance of our method improves on both ApolloScape and ApolloScape Night datasets with an increase in the number of explored adjacent frames. However, the performance improvement becomes negligible when the number of frames exceeds 4. Considering the monotonically increasing relationship between the model complexity and the number of frames, we set $n = 4$ to strike a balance between model performance and complexity.", + "bbox": [ + 496, + 703, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sampling Frame Gap $(\\triangle t)$ . In the study on the effect of the sampling frame gap, we set the number of frames", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1080", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/046f66b36e313b4e0fc1979894796229282eace5ffa43923391516698af4f7c3.jpg", + "image_caption": [ + "Figure 5: Qualitative comparison with SOTA methods. The top two examples are from the ApolloScape [17] dataset, and the bottom two examples are from the ApolloScape Night dataset. Yellow boxes highlight the area of interest for better visualization. Red boxes indicate false-positive segmentation predictions. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 890, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/103c965265be01852973f356baec3bb1419bbb9021e1fe98c4410ea4afe1a6ef.jpg", + "image_caption": [ + "(a) Number of frames (AS))" + ], + "image_footnote": [], + "bbox": [ + 102, + 469, + 269, + 547 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a63926b1780a60b9eb07e97b60b996f7c5dc010d54b4f73ce472af1b67e06b53.jpg", + "image_caption": [ + "(b) Number of frames (ASN)" + ], + "image_footnote": [], + "bbox": [ + 279, + 469, + 444, + 547 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06eb612b9cb19730fb07a488a6076dff637c62626bf4ff98cb2ef760d9140b31.jpg", + "image_caption": [ + "(c) Sampling frame gap (AS)" + ], + "image_footnote": [], + "bbox": [ + 101, + 580, + 267, + 659 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46770a1b70147801b354797c55666d9aaf5cdf1b069122278de9ab7012510501.jpg", + "image_caption": [ + "(d) Sampling frame gap (ASN)", + "Figure 6: The impact of the number of frames $n$ on ApolloScape (AS) and ApolloScape Night (ASN) datasets is shown in (a) and (b), while (c) and (d) show the effect of the sampling frame gap $\\triangle t$ ." + ], + "image_footnote": [], + "bbox": [ + 281, + 582, + 444, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$n = 4$ . Fig. 6 (c)-(d) indicates that the optimal performance is achieved at $\\triangle t = 2$ . This can be attributed to the balance between the sufficient overlap of road area across the frames and a wide enough temporal gap that allows the displacement of the source of occlusion, e.g. moving vehicle.", + "bbox": [ + 75, + 758, + 468, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Discussion", + "text_level": 1, + "bbox": [ + 76, + 845, + 194, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All ablation studies are conducted on the ApolloScape dataset with the same training strategies described in", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4836c4b5744e0c461d1404acda0a4283e95cef7eaa4c0a020764334fbfa92012.jpg", + "table_caption": [ + "Table 2: Ablation Study of our Proposed HomoFusion and RSNE on the ApolloScape Dataset" + ], + "table_footnote": [], + "table_body": "
18 mIoU↑36 mIoU↑
w/o HomoFusion49.431.0
HomoFusion w/o HomoGuide51.831.3
HomoFusion w/o RSNE57.435.3
HomoFusion w/ RGB RSNE54.933.8
Noisy Extrinsic57.535.1
HomoFusion Full59.335.9
", + "bbox": [ + 535, + 492, + 856, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sec. 4.1. We set the number of frames to $n = 4$ and a sampling frame gap of $\\triangle t = 2$ .", + "bbox": [ + 498, + 592, + 890, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Homofusion. The core component of our approach is the Homofusion module, which allows the use of complementary information from sequential frames. Tab. 2 demonstrates that our model with the Homofusion module (4 frames) achieves better performance than the variant without Homofusion module (1 frame), illustrating the effectiveness of the temporally consistent feature representation obtained with Homofusion module in addressing the partial occlusion issues. We also study the effect directly associating the pixels across frames based on their coordinates by using identity matrix as the homography matrix, denoted as \"w/o HomoGuide\". It shows that such an incorrect cross-frame spatial correspondence, although leveraging additional frames, provides limited help in improving the model performance.", + "bbox": [ + 496, + 625, + 892, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of RSNE. We study the impact of the proposed RSNE in the following model variants each comprising the HomoFusion modules: (a) without RSNE (\"HomoFusion", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1081", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "w/o RSNE\"), where the homography matrix is estimated with the initial road surface normal $(\\mathbf{n}_0)$ , (b) with RSNE using RGB information (\"HomoFusion w/ RGB RSNE\") and (c) with RSNE with features (\"HomoFusion Full\"). Results in Tab. 2 show that \"HomoFusion Full\" achieves better performance than \"HomoFusion w/o RSNE\", demonstrating the effectiveness of the iterative optimization in estimating a more accurate homography matrix that subsequently provides more accurate guidance on cross-frame spatial correspondences. It can be also observed that \"HomoFusion Full\" significantly outperforms \"HomoFusion w/ RGB RSNE\" indicating that deep features are more robust than RGB information in estimating an accurate road surface normal. This can be attributed to the fact the feature representations are less prone to adverse noises caused by environmental factors, e.g. brightness. Fig. 7 shows that the fused frame with road surface normal estimated with features has better alignment in road lines and markings than that with RGB information.", + "bbox": [ + 75, + 90, + 472, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e6f905dac6f480fd4bfb46d1b768cf0e903e46531c93fda4c6cd168dc297fb1.jpg", + "image_caption": [ + "(a) Normal estimation with RGB" + ], + "image_footnote": [], + "bbox": [ + 83, + 388, + 277, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/400830723900dd0292a3d335e18fd6abd7cc11ae9670381bc54e9b6aa19b9042.jpg", + "image_caption": [ + "(b) Normal estimation with feature", + "Figure 7: Fused frame sequence with the homography transformation between the frame pairs. The road surface normal estimated with feature representation enables more accurate homography transformation than with RGB information, resulting in better alignment of road lines and markings across frames." + ], + "image_footnote": [], + "bbox": [ + 285, + 388, + 480, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness against Extrinsic Noise. We investigate the robustness of our method to extrinsic noise by introducing random translation errors of less than 1 meter and rotation errors of up to $30^{\\circ}$ . The results are shown in \"Noisy Extrinsic\" of Tab. 2, which demonstrate that our method is capable of mitigating extrinsic noise and remains robust.", + "bbox": [ + 75, + 574, + 468, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Backbone. In order to ensure a fair comparison with CFFM [42] and MMA-Net [56], we implement our method with 'MiT-B1' and 'ResNet-50' backbone and report its performance in Tab. 3. Our method still outperforms CFFM [42] and MMA-Net [56] significantly. Additionally, we test our method with 'EfficientNet-B4/5' backbones, and the results show that our method has relatively stable performance even when using smaller backbones.", + "bbox": [ + 75, + 666, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application to Another Task. Given our method's ability to accurately align road surface objects across adjacent frames, we also try it with the task of detecting water hazards. Specifically, our approach aims to perform binary segmentation to identify water puddles in the road. The previ", + "bbox": [ + 75, + 787, + 468, + 863 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/13bc54305710c96d2d87225d83d6b3e2efa0c118e4865492b78ad7208585dcea.jpg", + "table_caption": [ + "Table 3: Comparison of our Proposed HomoFusion with Different Backbones on ApolloScape Dataset" + ], + "table_footnote": [], + "table_body": "
BackbonesParams(M)↓GFLOPs↓18 mIoU↑36 mIoU↑
MiT-B113.6489.857.235.3
ResNet-509.12189.755.334.6
EfficientNet-B40.860.857.035.3
EfficientNet-B51.0261.257.935.2
EfficientNet-B61.2461.259.335.9
", + "bbox": [ + 500, + 126, + 903, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ous work on this task T3D-FCN [22] also exploits temporal information, as water puddles can appear differently at different angles and distances as a vehicle moves. We use the Puddle-1000 dataset [13], which includes On-road and Offroad datasets. The frames obtained from an onboard camera on bumpy roads are not stable and contain small extrinsic noise. The intrinsic information of the camera is provided by the dataset, while the global extrinsic information of the camera is obtained using ORB-SLAM [34]. Our experiments are conducted using the current and seven previously consecutive frames with a $240 \\times 320$ image size, the same settings as T3D-FCN [22]. The results, presented in Tab. 4, demonstrate the high effectiveness of our method. More visual details of this task can be found in the Appendix.", + "bbox": [ + 496, + 210, + 893, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/78ebd398965f0fae4e66e75373e0b6ab8add981479f1916cdc40c8b4fda09788.jpg", + "table_caption": [ + "Table 4: Comparison of water puddle segmentations on Puddle-1000 dataset" + ], + "table_footnote": [], + "table_body": "
DatasetMethodsF1-meas↑Prec↑Rec↑
On-roadFCN-8s-FL-RAU [13]0.700.680.72
T3D-FCN [22]0.680.790.62
HomoFusion (ours)0.800.810.78
Off-roadFCN-8s-FL-RAU [13]0.810.870.77
T3D-FCNU [22]0.730.870.63
HomoFusion (ours)0.870.870.87
", + "bbox": [ + 513, + 477, + 877, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 573, + 617, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed lightweight lane mark segmentation model presented in this paper offers superior performance with reduced model complexity compared to SOTA approaches. The integration of the HomoFusion module and Road Surface Normal Estimator provides an accurate classification of partially occluded, shadowed, and/or glare-affected road lines and markings by leveraging adjacent frames and pixel-to-pixel attention mechanisms. Moreover, the novel approach of using the road surface normal to guide spatial correspondences across frames has significant implications for a wide range of applications in autonomous driving, including road surface inspection, water, and other hazard detections. However, the current model relies on available camera extrinsics and is somewhat impacted by extrinsic noise. Future work will focus on optimizing both the normal vector and camera extrinsic parameters to improve the performance and make the method more robust to challenging environmental conditions. Overall, the proposed method has the potential to significantly advance the field of autonomous driving and related applications.", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4We use a heavy backbone in the competing methods as their publicly available code uses this backbone", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1082", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 90, + 266, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The research is funded in part by an ARC Discovery Grant (grant ID: DP220100800) to HL.", + "bbox": [ + 76, + 114, + 468, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 157, + 174, + 174 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019.", + "[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10197-10205, 2019.", + "[3] Jonathan T. Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008.", + "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European conference on computer vision, pages 404-417. Springer, 2006.", + "[6] JiaWang Bian, Wen-Yan Lin, Yasuyuki Matsushita, Sai-Kit Yeung, Tan-Dat Nguyen, and Ming-Ming Cheng. Gms: Grid-based motion statistics for fast, ultra-robust feature correspondence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4181–4190, 2017.", + "[7] Amol Borkar, Monson Hayes, and Mark T Smith. A novel lane detection system with efficient ground truth generation. IEEE Transactions on Intelligent Transportation Systems, 13(1):365-374, 2011.", + "[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018.", + "[9] Zhe Chen, Jing Zhang, and Dacheng Tao. Progressive lidar adaptation for road detection. IEEE/CAA Journal of Automatica Sinica, 6(3):693-702, 2019.", + "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016.", + "[11] Farzan Erlik Nowruzi, Robert Laganiere, and Nathalie Japkowicz. Homography estimation from image pairs with hierarchical convolutional networks. In Proceedings of the IEEE international conference on computer vision workshops, pages 913-920, 2017.", + "[12] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981.", + "[13] Xiaofeng Han, Chuong Nguyen, Shaodi You, and Jianfeng Lu. Single image water hazard detection using fcn with re" + ], + "bbox": [ + 76, + 183, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "flection attention units. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-120, 2018.", + "[14] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004.", + "[15] Paul W Holland and Roy E Welsch. Robust regression using iteratively reweighted least-squares. Communications in Statistics-theory and Methods, 6(9):813-827, 1977.", + "[16] Yuenan Hou, Zheng Ma, Chunxiao Liu, Tak-Wai Hui, and Chen Change Loy. Inter-region affinity distillation for road marking segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12486-12495, 2020.", + "[17] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018.", + "[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In CVPR, 2022.", + "[19] Yeongmin Ko, Younkwan Lee, Shoaib Azam, Farzeen Munir, Moongu Jeon, and Witold Pedrycz. Key points estimation and point instance segmentation approach for lane detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):8949-8958, 2022.", + "[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020.", + "[21] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of applied mathematics, 2(2):164-168, 1944.", + "[22] Juntao Li, Chuong Nguyen, and Shaodi You. Temporal 3d fully connected network for water-hazard detection. In 2019 Digital Image Computing: Techniques and Applications (DICTA), pages 1-5, 2019.", + "[23] Yadi Li, Liguo Chen, Haibo Huang, Xiangpeng Li, Wenkui Xu, Liang Zheng, and Jiaqi Huang. Nighttime lane markings recognition based on canny detection and hough transform. In IEEE International Conference on Real-time Computing and Robotics, pages 411-415. IEEE, 2016.", + "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022.", + "[25] Zuo-Quan Li, Hui-Min Ma, and Zheng-Yu Liu. Road lane detection with gabor filters. In International Conference on Information System and Artificial Intelligence, pages 436-440. IEEE, 2016.", + "[26] Jiawei Liu, Jing Zhang, Yicong Hong, and Nick Barnes. Learning structure-aware semantic segmentation with image-level supervision. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021.", + "[27] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1083", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021.", + "[28] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022.", + "[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004.", + "[30] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019.", + "[31] Yunze Man, Xinshuo Weng, Xi Li, and Kris Kitani. Ground-net: Monocular ground plane normal estimation with geometric consistency. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2170-2178, 2019.", + "[32] Donald W Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the society for Industrial and Applied Mathematics, 11(2):431-441, 1963.", + "[33] Annika Meyer, N. Ole Salscheider, Piotr F. Orzechowski, and Christoph Stiller. Deep semantic lane segmentation for mapless driving. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 869-875, 2018.", + "[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015.", + "[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 286-291, 2018.", + "[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. volume 32, 2018.", + "[37] Jonah Philion. Fastdraw: Addressing the long tail of lane detection by adapting a sequential prediction network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11582-11591, 2019.", + "[38] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In European Conference on Computer Vision, pages 276-291. Springer, 2020.", + "[39] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011.", + "[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning robust camera localization from pixels to pose. In CVPR, 2021.", + "[41] Iago Suárez, Ghesn Sfeir, José M Buenasada, and Luis Baumela. Beblid: Boosted efficient binary local image descriptor. Pattern recognition letters, 133:366-372, 2020." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020.", + "[44] Tsung-Ying Sun, Shang-Jeng Tsai, and Vincent Chan. Hsi color model based lane-marking detection. In IEEE Intelligent Transportation Systems Conference, pages 1168-1172. IEEE, 2006.", + "[45] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021.", + "[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021.", + "[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019.", + "[48] Luo-Wei Tsai, Jun-Wei Hsieh, Chi-Hung Chuang, and Kuo-Chin Fan. Lane detection using directional random walks. In IEEE Intelligent Vehicles Symposium, pages 303–306. IEEE, 2008.", + "[49] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020.", + "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[51] Jun Wang, Tao Mei, Bin Kong, and Hu Wei. An approach of lane detection based on inverse perspective mapping. In International IEEE Conference on Intelligent Transportation Systems, pages 35-38. IEEE, 2014.", + "[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021.", + "[53] Ruochen Yin, Yong Cheng, Huapeng Wu, Yuntao Song, Biao Yu, and Runxin Niu. Fusionlane: Multi-sensor fusion for lane marking semantic segmentation using deep neural networks. IEEE Transactions on Intelligent Transportation Systems, 2020.", + "[54] Seungwoo Yoo, Hee Seok Lee, Heesoo Myeong, Sungrack Yun, Youngwoo Park, Janghoon Cho, and Duck Hoon" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1084", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kim. End-to-end lane marker detection via row-wise classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1006-1007, 2020.", + "[55] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. arXiv preprint arXiv:1904.00962, 2019.", + "[56] Yujun Zhang, Lei Zhu, Wei Feng, Huazhu Fu, Mingqian Wang, Qingxia Li, Cheng Li, and Song Wang. Vil-100: A new dataset and a baseline model for video instance lane detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15681–15690, 2021.", + "[57] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018." + ], + "bbox": [ + 78, + 90, + 468, + 343 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1085", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_model.json b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..884f780c450e67ac039ff9fbfc7366e6cf89c5e1 --- /dev/null +++ b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.13, + 0.825, + 0.176 + ], + "angle": 0, + "content": "Homography Guided Temporal Fusion for Road Line and Marking Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.203, + 0.816, + 0.257 + ], + "angle": 0, + "content": "Shan Wang\\(^{1,2}\\) Chuong Nguyen\\(^{1}\\) Jiawei Liu\\(^{2}\\) Kaihao Zhang\\(^{2}\\) Wenhan Luo\\(^{3}\\) Yanhao Zhang\\(^{2}\\) Sundaram Muthu\\(^{1}\\) Fahira Afzal Maken\\(^{1}\\) Hongdong Li\\(^{2}\\) \n\\(^{1}\\)Data61, CSIRO \\(^{2}\\)Australian National University \\(^{3}\\)Sun Yat-sen University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.747 + ], + "angle": 0, + "content": "Reliable segmentation of road lines and markings is critical to autonomous driving. Our work is motivated by the observations that road lines and markings are (1) frequently occluded in the presence of moving vehicles, shadow, and glare and (2) highly structured with low intra-class shape variance and overall high appearance consistency. To solve these issues, we propose a Homography Guided Fusion (HomoFusion) module to exploit temporally-adjacent video frames for complementary cues facilitating the correct classification of the partially occluded road lines or markings. To reduce computational complexity, a novel surface normal estimator is proposed to establish spatial correspondences between the sampled frames, allowing the HomoFusion module to perform a pixel-to-pixel attention mechanism in updating the representation of the occluded road lines or markings. Experiments on ApolloScape, a largescale lane mark segmentation dataset, and ApolloScape Night with artificial simulated night-time road conditions, demonstrate that our method outperforms other existing SOTA lane mark segmentation models with less than \\(9\\%\\) of their parameters and computational complexity. We show that exploiting available camera intrinsic data and ground plane assumption for cross-frame correspondence can lead to a light-weight network with significantly improved performances in speed and accuracy. We also prove the versatility of our HomoFusion approach by applying it to the problem of water puddle segmentation and achieving SOTA performance1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.21, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Lane mark segmentation aims to achieve pixel-wise classification of road lines and markings simultaneously. Known for its quintessential importance in autonomous driving, lane mark segmentation is also an effective tool for constructing accurate High Definition (HD) maps. Existing works solve individual sub-tasks, e.g. drivable area segmen" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.292, + 0.891, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.718 + ], + "angle": 0, + "content": "Figure 1: Illustration of the effect of the proposed HomoFusion module that explores the adjacent frames for cues, facilitating the correct classification of (1) a \"Straight Arrow\", which, with its bottom half occluded by a vehicle, is mistakenly classified as a \"Right Turn & Straight Arrow\" without the HomoFusion module, and (2) a partially occluded \"Dotted Line\", which is incorrectly classified as a \"Solid Line\" without the HomoFusion module. The fused frame in the \\(4^{th}\\) row and the \\(2^{nd}\\) column demonstrates the recovered road lines and markings after projecting the previous frames onto the current frame with the estimated homography matrices. The yellow box enlarges the area where mistake classifications are corrected. The red box indicates the spatially corresponding area across the frames. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.765 + ], + "angle": 0, + "content": "tation [33, 9] and lane detection [19, 35], while few address the lane mark segmentation in its entirety [53, 16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Despite the tremendous progresses in semantic/scene segmentation [52, 42, 26], little attention has been paid to the lane mark segmentation task [53, 16]. Yin et al. [53] leverages additional LiDAR information, merging the segmented visual information with the point clouds, to achieve lane mark segmentation. IntRA-KD [16] applies the knowledge distillation technique to improve the efficiency of the lane mark segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "A major challenge in the lane mark segmentation task is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at https://github.com/ShanWang-Shan/HomoFusion." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "partial occlusion of the road lines and markings caused by the vehicle and surrounding environment leading to false classifications. For example, a partially occluded straight arrow can be easily mistaken as a right turn & straight arrow, as depicted in Fig. 1. However, existing lane mark segmentation methods have not yet made use of the following observations: (1) occlusion of the road lines and markings (due to nearby moving vehicles, shadow, and glare) can be reduced by cross-frame consistency; (2) road lines and markings are highly structured, maintaining high consistency in intra-class shapes as well as in overall appearance, alleviating the requirements on learning complicated features to distinguish different classes with high intra-class variance as in semantic/scene segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.485 + ], + "angle": 0, + "content": "To find complementary information from adjacent frames, we use the ground plane assumption for the road region immediately in front of the camera. Our Homography Guided Fusion (HomoFusion) module achieves temporally consistent representation, recovers partially occluded road lines or markings, and leads to correct classification. The module's success depends on accurate cross-frame spatial correspondence, which we achieve using a homography transformation matrix estimated with available intrinsic and extrinsic parameters of the on-vehicle camera, and the road surface normal estimated by our novel optimization method called Road Surface Normal Estimator (RSNE)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.473, + 0.681 + ], + "angle": 0, + "content": "Using the highly structured nature of road lines and markings could reduce the complexity of the detection problem and make it suitable to run on edge devices. To this end, we employ (1) a lightweight encoder to represent the visual information in feature spaces, and (2) a cross-frame pixel-to-pixel attention mechanism in our HomoFusion module, instead of a more computationally expensive global attention mechanism employed by other methods. These design decisions allow the proposed model to outperform the existing lane mark segmentation models with less than \\(9\\%\\) of their parameters and Giga Floating Point Operations (GFLOPs) We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.688, + 0.47, + 0.75 + ], + "angle": 0, + "content": "- We propose a HomoFusion module that uses ground plane assumption and the adjacent frames for temporally consistent representations for accurate classification of partially occluded road lines and markings." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.757, + 0.47, + 0.864 + ], + "angle": 0, + "content": "- We design a novel estimator RSNE for road surface normal, which, combined with camera intrinsic and extrinsic parameters readily available on an autonomous vehicle, yields accurate homography matrix between frame pairs. RSNE simplifies the 8 degrees of freedom (DoF) of homography problem to a 2 DoF of normal vector problem." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We present a lightweight lane mark segmentation model that achieves better performance than the state" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.688, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "of-the-art (SOTA) methods with significantly reduced model complexity and computation requirements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.136, + 0.642, + 0.152 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.161, + 0.655, + 0.176 + ], + "angle": 0, + "content": "2.1. Lane Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.895, + 0.487 + ], + "angle": 0, + "content": "Traditional lane detection approaches rely on handcrafted features such as color [44, 51], edge [23] and texture [25], which are limited in complex scenarios [48, 7]. Recent advances in deep neural networks have led to significant performance improvements, with methods like message-passing networks [36] and attention-based modules [45]. [16, 37, 38, 54, 27] formulate lane detection as a row-wise classification task based on grid division of the input image. PolyLaneNet [46] is the first parametric prediction method, which outputs polynomials to represent each lane. BEV-Former [24] and PETRv2 [28] employ transformer mechanisms to streamline the conversion of perspective views into bird's-eye views, thereby eliminating variations in object size caused by perspective. Although the above methods achieve impressive performance, all of them focus on detecting the lane lines and ignore the basic road elements like arrow signal in Fig. 1. Furthermore, lane detection differs from segmentation tasks in that it focuses on identifying the boundaries of the drivable area, rather than the real shape of the lanes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.833, + 0.513 + ], + "angle": 0, + "content": "2.2. Road Line and Marking Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.747 + ], + "angle": 0, + "content": "The identification of road elements is crucial for ensuring safety in autonomous driving systems, but there are few works addressing this problem. Hou et al. [16] introduced a distillation approach that demonstrates competitive performance in lane mark segmentation. Yin et al. [53] used an LSTM-based network to segment images with DeepLabv3+ [8] and then merged them with point clouds to assist with lane segmentation. Unfortunately, these methods do not address the challenges of occlusion, shadows, and glare that frequently occur in real-world driving scenarios. Recently, Zhang et al. [56] introduced global memory information from previous frames to enhance local information for video lane segmentation. However, this method introduces attention to the entire image, making it memory-intensive and inefficient for autonomous driving scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.727, + 0.773 + ], + "angle": 0, + "content": "2.3. Homography Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Homography estimation methods can be divided into non-deep and deep learning-based approaches. Non-deep methods estimate the homography using feature extraction, feature matching, and outlier rejection. SIFT [29], SURF [5, 4], ORB [39], LPPM [30], GMS [6], and BEBLID [41] are commonly used for feature extraction, and RANSAC [12], MAGAC [2], and LRLS [15] are applied for outlier rejection. Recently, many deep learning-based approaches" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "1076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.086, + 0.805, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.451 + ], + "angle": 0, + "content": "Figure 2: Overview of our proposed model consisting of a pair of lightweight encoder and decoder, our proposed HomoFusion module, and our proposed Road Surface Normal Estimator (RSNE). A sequence of frames \\(\\mathbf{I}\\), including a target frame \\(\\mathbf{I}_{\\mathbf{t}}\\) and \\(n - 1\\) previous frames, are encoded into the feature representations \\((\\mathbf{F}^l)\\). RNSE estimates the road surface normal vector, which, combined with the camera intrinsic and extrinsic parameters, yields a homography matrix between each frame pair, establishing cross-frame spatial correspondences. HomoFusion uses pixel-to-pixel attention mechanism to obtain temporally consistent representation for on-road pixels of the current frame with the spatial correspondence across frames as guidance. Finally, the decoder decodes and upsamples the temporally consistent feature representations to produce the lane mark segmentation prediction \\((\\mathbf{S}_{\\mathbf{t}}^{\\mathrm{prd}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.473, + 0.581 + ], + "angle": 0, + "content": "have been proposed, such as DeTone et al. [10]'s VGG-like architecture, Nowruzi et al. [11], Le et al. [20]'s and Man et al. [31]'s cascaded VGG-like networks, and Chang et al. and Zhao et al.'s incorporation of the Lucas-Kanade (LK) algorithm with deep networks. However, these methods do not utilize known camera intrinsic and extrinsic parameters across frames for homography matrix estimation, which can effectively reduce the search dimensions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.596, + 0.17, + 0.611 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.188, + 0.637 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.829 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, our proposed framework takes a sequence of frames, denoted as \\(\\mathbf{I} = \\{\\mathbf{I}_i|i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}\\), as input. This sequence consists of a current frame \\(\\mathbf{I}_t\\) and \\(n - 1\\) previous frames sampled at a fixed time interval \\(\\Delta t\\). The output of the framework is a lane mark segmentation map \\(\\mathbf{S}_t\\) for the current frame. In addition, the framework includes a novel HomoFusion module, which uses the homography transformation guided by the proposed RSNE to fuse the feature map of the current frame with those of the previous frames. We denote the encoded feature representation for the \\(t^{th}\\) frame as \\(\\mathbf{F}_t = \\{\\mathbf{F}_t^l\\in \\mathbb{R}^{H_l\\times W_l\\times C_l}\\}_{l = 1}^L\\) where \\(l\\) indicates the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.901 + ], + "angle": 0, + "content": "2In our experiments, both camera intrinsic and extrinsic parameters are provided by the dataset. In real-world scenarios, these parameters can be obtained through camera calibration and autonomous vehicle pose obtained from the pose estimation framework, which is already known in the in-car system and free of cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.687 + ], + "angle": 0, + "content": "level of the feature map. The shallowest and deepest levels are represented by 1 and \\(L\\), respectively. \\(H_{l}\\), \\(W_{l}\\), and \\(C_l\\) represent the height, width, and channel number of feature maps in level \\(l\\). We denote all encoded features as \\(\\mathbf{F} = \\{\\mathbf{F}_i^l | i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}\\). The proposed HomoFusion module, detailed in Sec. 3.2, uses a cross-frame pixel-to-pixel attention mechanism to fuse the feature map \\((\\mathbf{F}_f = \\{\\mathbf{F}_f^l\\}_{l = 1}^L)\\) of the current frame with those of the previous frames. The homography transformation matrix between the target frame and each of the previous frames is calculated using the estimated normal vector of the road surface \\(\\mathbf{n}\\) obtained by the proposed RSNE, as described in Sec. 3.3. Finally, the decoder produces the segmentation prediction \\((\\mathbf{S}_t)\\) for the target frame based on the fused feature map." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.696, + 0.871, + 0.713 + ], + "angle": 0, + "content": "3.2. Homography Guided Fusion (HomoFusion)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our proposed HomoFusion module employs a pixel-to-pixel attention mechanism to fuse spatially corresponding pixels across frames. This mechanism is achieved by projecting the pixels of the current frame onto the previous frames through an accurate homography transformation. Since road lines and markings are strictly on-road, the search area can be limited to the road surface, which is mostly a plane, at least within the immediate front of the vehicle where markings are readable. Therefore, only the homography transformation is needed to accurately map road pixels between frames. According to the standard inverse homography [14, 57, 49], for each on-road pixel \\( p_t = \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1077" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.181, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 3: Illustration of sample points. (Right) Sample points in the current frame. (Left/Middle) Correspondence of a sample point in previous frames. The red point coordinate is calculated by using the correct normal, while the cyan point coordinate is calculated by using the initial (incorrect) normal." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.471, + 0.33 + ], + "angle": 0, + "content": "\\((u,v)^{\\top}\\) of the current frame, its spatially corresponding pixels of the reference frames \\(\\{p_i|i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}\\) can be computed using Eq. (1)." + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.34, + 0.47, + 0.372 + ], + "angle": 0, + "content": "\\[\np _ {i} \\propto \\mathbf {K} \\left(\\mathbf {R} _ {i} - \\frac {\\mathbf {t} _ {i} \\mathbf {n} ^ {\\top}}{d}\\right) \\mathbf {K} ^ {- 1} \\left(p _ {t} \\oplus 1\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.583 + ], + "angle": 0, + "content": "where \\(\\propto\\) represents proportional, \\(\\mathbf{K}\\) is the intrinsic matrix of the on-vehicle camera, \\(\\mathbf{R}_i\\) and \\(\\mathbf{t}_i\\) represent the relative rotation and translation between the current frame \\(\\mathbf{I}_t\\) and the \\(i^{th}\\) previous frame \\(\\mathbf{I}_i\\). \\(\\mathbf{n}\\) is the normal of the observed road surface, which is estimated in Sec. 3.3. \\(d\\) is the vertical distance between the on-vehicle camera and the road surface obtained from the camera calibration. \\(\\oplus 1\\) converts the point coordinates to homogeneous coordinates. The on-road pixels of the current frame are sampled from a triangular area in front of the car, as illustrated in Fig. 3. Figure 3 also demonstrates that the shape and category of a left-turn can be recovered by exploring the spatially corresponding area in the previous frames under the guidance of an accurate homography transformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.78 + ], + "angle": 0, + "content": "With the spatial correspondence across frames, the proposed HomoFusion module computes a temporally consistent representation for each sampled pixel of the current frame by employing temporal attention mechanism on the feature presentations of spatially corresponding pixels. The mechanism assigns the pixel representation of current frame \\((\\mathbf{F}_t[p_t])\\) as a query and all frames \\((\\{\\mathbf{F}_i[p_i]\\})\\) as keys and values. In contrast to the existing attention mechanism [50], we propose to (1) omit the spatial encoding in the presence of obtained spatial correspondence across frames, and (2) apply an L2 normalization on the query and the keys to obtain more robust (deep) feature similarities (Eq. 2) across various lighting environments." + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.792, + 0.47, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} _ {i} = \\frac {\\mathbf {F} _ {t} [ p _ {t} ]}{\\| \\mathbf {F} _ {t} [ p _ {t} ] \\| _ {2}} \\cdot \\frac {\\mathbf {F} _ {i} [ p _ {i} ]}{\\| \\mathbf {F} _ {i} [ p _ {i} ] \\| _ {2}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.832, + 0.47, + 0.865 + ], + "angle": 0, + "content": "where \\(\\| .\\| _2\\) indicates an \\(L_{2}\\) normalization. The similarity is normalized by softmax function as given in Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.874, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {i} = \\frac {\\exp \\mathbf {a} _ {i}}{\\sum_ {i} \\exp \\mathbf {a} _ {i}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "before being used as weight to fuse the spatially corresponding pixels across the frames as shown in Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.133, + 0.892, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {f} [ p _ {t} ] = \\mathbf {F} _ {t} [ p _ {t} ] + \\sum_ {i} \\mathbf {W} _ {i} \\mathbf {F} _ {i} [ p _ {i} ]. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.174, + 0.85, + 0.19 + ], + "angle": 0, + "content": "3.3. Road Surface Normal Estimator (RSNE)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Accurate estimation of road surface normal is crucial for establishing spatial correspondence across frames. Generally, on-vehicle cameras have a fixed angle with respect to the ground level. However, the actual road surface normal can vary due to the unevenness of the road surface, e.g. uphill or downhill roads, and sloped turns, etc. Assuming that the initial road surface normal is a vector being perpendicular to a horizontal plane and pointing upwards, we propose to obtain an accurate road surface normal with an optimization process that repeats for each new coming frame. The proposed RSNE iteratively updates the estimated road surface normal with the projection error of on-road pixels from the current frame onto the previous frames. Specifically, we sample \\( m \\) pixels \\( \\{p_t^j\\}_{j=1}^m \\) from the current frame that are likely to be on-road, as illustrated in Fig. 3, and compute the spatially corresponding pixels \\( \\{p_i^j\\}_{j=1}^m \\) in the previous frames using the homography matrix computed with the estimated road surface normal as defined in Eq. 1. The estimated road surface normal is optimized in an iterative manner through the following procedure: (a) computing a residual between the low-level features of spatially corresponding pixels across the frames with Eq. 5" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.54, + 0.892, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {i} ^ {j} = \\mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ] - \\mathbf {F} _ {t} ^ {1} [ p _ {t} ^ {j} ] \\in \\mathbb {R} ^ {c}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.892, + 0.614 + ], + "angle": 0, + "content": "where the residual is set to 0 for the sample point whose projections in previous frames fall outside the frame area, and computing the overall error with Eq. 6" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.624, + 0.892, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} = \\sum_ {i, j} \\rho \\left(\\left\\| \\mathbf {r} _ {i} ^ {j} \\right\\| _ {2} ^ {2}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.752 + ], + "angle": 0, + "content": "where \\(\\| .\\| _2^2\\) shows squared norm and \\(\\rho (\\cdot)\\) is a robust cost function proposed by Barron [3]. This nonlinear least-square cost is iteratively minimized towards correct normal estimation using the Levenberg-Marquardt algorithm [21, 32]; (b) computing a new road surface normal value by adding the update given by Eq. 12." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.799 + ], + "angle": 0, + "content": "The road surface normal vector is a unit vector with two degrees of freedom (DoF), which can be decomposed into a pitch angle \\(\\theta\\) and a roll angle \\(\\phi^3\\), as shown in Eq. 7." + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.807, + 0.892, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} = \\left(- \\sin \\phi \\cos \\theta , - \\cos \\phi \\cos \\theta , \\sin \\theta\\right) ^ {\\top}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.836, + 0.892, + 0.868 + ], + "angle": 0, + "content": "The Jacobian of the residual function w.r.t. the pitch angle and the roll angle is defined in Eq. 8 by applying the chain" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3An explanation of the pitch angle \\( \\theta \\) and roll angle \\( \\phi \\) is provided in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.113, + 0.105 + ], + "angle": 0, + "content": "rule," + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.103, + 0.47, + 0.142 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} _ {i} ^ {j} = \\frac {\\partial \\mathbf {r} _ {i} ^ {j}}{\\partial \\delta} = \\frac {\\partial \\mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ]}{\\partial p _ {i} ^ {j}} \\frac {\\partial p _ {i} ^ {j}}{\\partial \\mathbf {n}} \\frac {\\partial \\mathbf {n}}{\\partial \\delta}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.146, + 0.47, + 0.215 + ], + "angle": 0, + "content": "where \\(\\delta\\) represents the to-be-optimized target (\\(\\theta\\) or \\(\\phi\\)), and \\(\\frac{\\partial \\mathbf{F}_i^1[p_i^j]}{\\partial p_i^j}\\) is gradient from 2D interpolation. \\(\\frac{\\partial p_i^j}{\\partial \\mathbf{n}}\\) is the Jacobian of the homography transformation (Eq. (1)) w.r.t. \\(\\mathbf{n}\\), as defined in Eq. 9" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.225, + 0.469, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial p _ {i} ^ {j}}{\\partial \\mathbf {n}} = - \\frac {1}{d} \\mathbf {K t} _ {i} \\left(\\mathbf {K} ^ {- 1} \\left(p _ {t} ^ {j} \\oplus 1\\right)\\right) ^ {\\top}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.47, + 0.301 + ], + "angle": 0, + "content": "and \\(\\frac{\\partial\\mathbf{n}}{\\partial\\delta}\\), defined in Eq. 10 and Eq. 11, is the Jacobian of Eq. (7) w.r.t. \\(\\theta\\) and \\(\\phi\\) respectively" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.31, + 0.47, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathbf {n}}{\\partial \\theta} = \\left(- \\frac {\\mathbf {n} _ {1} \\mathbf {n} _ {3}}{\\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}}, - \\frac {\\mathbf {n} _ {2} \\mathbf {n} _ {3}}{\\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}}, \\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.356, + 0.469, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathbf {n}}{\\partial \\phi} = \\left(\\mathbf {n} _ {2}, - \\mathbf {n} _ {1}, 0\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.392, + 0.47, + 0.468 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}_k\\) represents the \\(k\\)-th element of \\(\\mathbf{n}\\). The detailed derivation is included in appendix (Sec. 1). We obtain Hessian matrices as \\(\\mathbb{H} = \\mathbf{J}^{\\top}\\rho^{\\prime}\\mathbf{J}\\), where \\(\\rho^\\prime\\) is the derivative of function \\(\\rho (\\cdot)\\). We compute the update by damping the Hessian and solving the linear system, as shown in" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.478, + 0.469, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\triangle \\delta = - (\\mathbb {H} + \\lambda \\operatorname {d i a g} (\\mathbb {H})) ^ {- 1} \\mathbf {J} ^ {\\top} \\rho^ {\\prime} \\mathbf {r}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.507, + 0.47, + 0.553 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the damping factor [40], balancing between the Gauss-Newton \\((\\lambda = 0)\\) and gradient descent \\((\\lambda = \\infty)\\). Our entire optimization process is differentiable." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.47, + 0.644 + ], + "angle": 0, + "content": "The algorithm is illustrated in Alg. 1. The normal vector \\(\\mathbf{n}\\) uses low-level features \\(\\mathbf{F}^1\\). While the high-level features \\(\\mathbf{F}^2\\) focus on segmentation feature expression, the low-level features \\(\\mathbf{F}^1\\) are supervised to extract useful features for both segmentation and road surface normal estimation. As both tasks focus on on-road objects, they can benefit each other." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.657, + 0.21, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.295, + 0.697 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We use a pre-trained EfficientNet [47] to extract image features at two different scales, \\( 4 \\times \\) and \\( 16 \\times \\) down-scaling, with 64 and 128 channels, respectively. To focus on the bottom region of the input image, where road marks are typically visible, we crop only the bottom \\( 40\\% \\) of the input image, following the same approach in the prior work [16]. The processed image size is set to \\( 272 \\times 848 \\). Our decoder consists of bi-linear up-sampling layers and convolution layers, which up-sample the high-level features and increase the resolution by a factor of 4 at each level. We trained our model using the AdamW optimizer [55] for 30,000 iterations on two NVIDIA RTX 3090 GPUs, with a learning rate of \\( 4 \\times 10^{-3} \\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.734, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Normal Optimization" + }, + { + "type": "algorithm", + "bbox": [ + 0.5, + 0.109, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Input: \nThe 1-st Level Feature Maps: \\(\\mathbf{F}^1 = \\{\\mathbf{F}_i^1\\}\\) \nSample Points Coordinates: \\(p_t = \\{p_t^j\\}_{j = 1}^m\\) \nInitial Pitch and Roll Value: \\((\\theta_0,\\phi_0) = (0.15,0)\\) \nDamping Factor: \\(\\lambda = (\\lambda_{\\theta}\\lambda_{\\phi})\\) \nHyper-parameter: \nk; \\(\\triangleright\\) Maximum loop count, empirically set to 20 \n\\(\\alpha\\) ; \\(\\triangleright\\) Convergence threshold, empirically set to 0.0001 \nOutput: Optimized Pitch and Roll \\((\\theta ,\\phi)\\) . \n1: function OPT( \\(\\theta_0,\\phi_0,\\mathbf{F}^1,p_t,\\lambda\\) \n2: Derive point features \\(\\mathbf{F}_t^1 [p_t]\\) from \\(\\mathbf{F}_t^1\\) . \n3: for \\(k\\gets 1\\) to k do \n4: Calculate point coordinates \\(p_i\\) in \\(\\mathbf{F}_i^1\\) (Eq. 1); \n5: Derive point features \\(\\mathbf{F}_i^1[p_i]\\) from \\(\\mathbf{F}_i^1\\) . \n6: Calculate residual \\(\\mathbf{r}_i\\) (Eq. 5); \n7: Calculate observe error E (Eq. 6); \n8: Calculate robust cost \\(\\rho (\\mathbf{E})\\) and its derivation \\(\\rho^\\prime\\) \n9: Construct J and its Hessian matrices \\(\\mathbb{H}\\) (Eq. 8); \n10: Obtain \\(\\triangle \\delta\\) by Cholesky decomposition (Eq. 12); \n11: Update Normal as \\((\\theta ,\\phi)_k\\gets (\\theta ,\\phi)_{k - 1} + \\triangle \\delta\\) \n12: if MAX(△δ)< α then \n13: Break; \n14: end if \n15: end for \n16: end function" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.602, + 0.563 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.673 + ], + "angle": 0, + "content": "We conduct our experiments on two datasets, ApolloScape [17] and ApolloScape Night. Other datasets were deemed unsuitable due to the absence of camera extrinsics, as observed in SDLane [18] and VLI-100 [56], inadequate frame overlap, or the lack of diverse segmentation labels, exemplified by Waymo Open Dataset [43]. Further elaboration on these reasons can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.893, + 0.884 + ], + "angle": 0, + "content": "ApolloScape. The ApolloScape dataset is a large-scale dataset that can be used for localization and segmentation tasks. It consists of 38 distinct classes and poses various challenges including occlusions and tiny road markings. To provide accurate camera poses, the vehicle is equipped with customer-grade GPS/IMU. However, for the lane mark segmentation task, only 41, 201 annotated images out of more than \\(110,000+\\) are associated with camera pose information. We used these 41, 201 images for our experiments since our approach relies on camera poses. We divided them into 35, 173 training images and 6, 028 validation images. Because our approach uses sequential information, we select the validation set with completely different trajectories to ensure a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "ApolloScape Night. Since there is a lack of datasets for" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1079" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.137 + ], + "angle": 0, + "content": "Table 1: Comparison with SOTA Methods on the ApolloScape [17] and Apolloscape Night Datasets and Running on an NVIDIA RTX 3090 GPU. \"18 mIoU\" Represents the Mean IoU of 18 Types of Lane Markers Selected by ApolloScape Official Metrics. \"36 mIoU\" Represents the Mean IoU of All Unignorable 36 Types." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.139, + 0.895, + 0.233 + ], + "angle": 0, + "content": "
MethodsFrame CountBackboneParams (M)↓GFLOPs↓FPS (f/s)↑ApolloScape [17]ApolloScape Night
18 mIoU↑36 mIoU↑18 mIoU↑36 mIoU↑
IntRA-KD[16]1ResNet-10165.65159.410.842.124.629.816.7
SegFormer[52]1MiT-B113.51048.843.852.332.138.323.1
CFFM [42]4MiT-B115.31192.622.753.232.739.123.6
MMA-Net [56]4ResNet-5057.9723.220.652.931.438.823.2
HomoFusion(ours)4EfficientNet-B61.2461.225.459.335.944.826.6
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.236, + 0.468, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.446, + 0.472, + 0.494 + ], + "angle": 0, + "content": "Figure 4: Sample images from the ApolloScape Night dataset. Top: original daytime images from the ApolloScape dataset. Bottom: synthesized night-time images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.592 + ], + "angle": 0, + "content": "night-time lane mark segmentations, we created an artificial dataset called ApolloScape Night from the ApolloScape dataset using a cross-domain generation network [1]. This allows us to evaluate our proposed model on a challenging dataset with poor lighting conditions, road reflection, and glare. Fig. 4 displays some sample images from the dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.263, + 0.617 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.471, + 0.807 + ], + "angle": 0, + "content": "In accordance with the guidelines of the ApolloScape benchmark [17], we used mean intersection-over-union (mIoU) as the evaluation criterion. ApolloScape contains 38 different types of lane markers, including two ignorable labels (noise and ignored). Of these labels, 18 categories are used in the official evaluation metrics. To provide a comprehensive evaluation of our approach, we report the evaluated mIoU results for both the selected 18 categories, denoted as \"18 mIoU\", and all 36 categories, denoted as \"36 mIoU\". By reporting both sets of results, we provide a more complete picture of the performance of our method on the ApolloScape and ApolloScape Night datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.442, + 0.833 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To evaluate the performance of our proposed method, we compared it with SOTA algorithms, including (a) IntRA-KD [16], (b) SegFormer [52], (c) CFFM [42], and (d) MMA-Net [56], on the ApolloScape [17] and ApolloScape" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.239, + 0.892, + 0.284 + ], + "angle": 0, + "content": "Night datasets. To ensure a fair comparison, we retrained each model on the same subset of the training set with the same input resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.892, + 0.36 + ], + "angle": 0, + "content": "The results of the comparison are presented in Table 1. Our approach outperforms the existing SOTA methods on both the ApolloScape [17] and ApolloScape Night datasets while having less than \\(9\\%\\) of their parameters or computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.893, + 0.559 + ], + "angle": 0, + "content": "In addition, we compared our proposed method with CFFM [42] which also uses adjacent frames to enhance the representation of the current frame. Our method is more efficient in terms of model complexity, with a complexity of \\(\\mathcal{O}(HWC)\\) for feature extraction and \\(\\mathcal{O}(HWC^2)\\) for cross-frame feature fusion. In contrast, the complexity of CFFM is \\(\\mathcal{O}(H^2 W^2 C) + \\mathcal{O}(HWC^2)\\) for feature extraction and \\(\\mathcal{O}(HWEC) + \\mathcal{O}(HWC^2)\\) for cross-frame feature fusion, where \\(E\\) is calculated by their receptive field and pooling kernel size. This explains why our method has a lower computational requirement for the cross-frame pixel-to-pixel attention mechanism compared to the global attention mechanism used by CFFM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.667 + ], + "angle": 0, + "content": "The qualitative results of both the existing SOTA models and our proposed model are shown in Fig. 5. The results indicate that our approach achieves superior segmentation performance on road lines and markings, even under adverse conditions such as occlusion, road reflection, and poor lighting. Additional visualization examples for various categories can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.753, + 0.696 + ], + "angle": 0, + "content": "4.5. Study on Hyper-parameters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Number of Frames \\((n)\\). We conducted a study on the impact of the number of frames while setting the sampling frame gap \\(\\triangle t = 2\\). The results presented in Fig. 6 (a)-(b) demonstrate that the performance of our method improves on both ApolloScape and ApolloScape Night datasets with an increase in the number of explored adjacent frames. However, the performance improvement becomes negligible when the number of frames exceeds 4. Considering the monotonically increasing relationship between the model complexity and the number of frames, we set \\(n = 4\\) to strike a balance between model performance and complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Sampling Frame Gap \\((\\triangle t)\\). In the study on the effect of the sampling frame gap, we set the number of frames" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1080" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.891, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.403, + 0.895, + 0.449 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparison with SOTA methods. The top two examples are from the ApolloScape [17] dataset, and the bottom two examples are from the ApolloScape Night dataset. Yellow boxes highlight the area of interest for better visualization. Red boxes indicate false-positive segmentation predictions. Best viewed in color." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.47, + 0.271, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.554, + 0.262, + 0.567 + ], + "angle": 0, + "content": "(a) Number of frames (AS))" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.47, + 0.445, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.554, + 0.441, + 0.567 + ], + "angle": 0, + "content": "(b) Number of frames (ASN)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.582, + 0.268, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.666, + 0.263, + 0.68 + ], + "angle": 0, + "content": "(c) Sampling frame gap (AS)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.583, + 0.446, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.666, + 0.445, + 0.68 + ], + "angle": 0, + "content": "(d) Sampling frame gap (ASN)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.691, + 0.47, + 0.752 + ], + "angle": 0, + "content": "Figure 6: The impact of the number of frames \\( n \\) on ApolloScape (AS) and ApolloScape Night (ASN) datasets is shown in (a) and (b), while (c) and (d) show the effect of the sampling frame gap \\( \\triangle t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.835 + ], + "angle": 0, + "content": "\\(n = 4\\). Fig. 6 (c)-(d) indicates that the optimal performance is achieved at \\(\\triangle t = 2\\). This can be attributed to the balance between the sufficient overlap of road area across the frames and a wide enough temporal gap that allows the displacement of the source of occlusion, e.g. moving vehicle." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.195, + 0.86 + ], + "angle": 0, + "content": "4.6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "All ablation studies are conducted on the ApolloScape dataset with the same training strategies described in" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.456, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Table 2: Ablation Study of our Proposed HomoFusion and RSNE on the ApolloScape Dataset" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.493, + 0.857, + 0.582 + ], + "angle": 0, + "content": "
18 mIoU↑36 mIoU↑
w/o HomoFusion49.431.0
HomoFusion w/o HomoGuide51.831.3
HomoFusion w/o RSNE57.435.3
HomoFusion w/ RGB RSNE54.933.8
Noisy Extrinsic57.535.1
HomoFusion Full59.335.9
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.593, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Sec. 4.1. We set the number of frames to \\( n = 4 \\) and a sampling frame gap of \\( \\triangle t = 2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.893, + 0.852 + ], + "angle": 0, + "content": "Impact of Homofusion. The core component of our approach is the Homofusion module, which allows the use of complementary information from sequential frames. Tab. 2 demonstrates that our model with the Homofusion module (4 frames) achieves better performance than the variant without Homofusion module (1 frame), illustrating the effectiveness of the temporally consistent feature representation obtained with Homofusion module in addressing the partial occlusion issues. We also study the effect directly associating the pixels across frames based on their coordinates by using identity matrix as the homography matrix, denoted as \"w/o HomoGuide\". It shows that such an incorrect cross-frame spatial correspondence, although leveraging additional frames, provides limited help in improving the model performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Impact of RSNE. We study the impact of the proposed RSNE in the following model variants each comprising the HomoFusion modules: (a) without RSNE (\"HomoFusion" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1081" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.378 + ], + "angle": 0, + "content": "w/o RSNE\"), where the homography matrix is estimated with the initial road surface normal \\((\\mathbf{n}_0)\\), (b) with RSNE using RGB information (\"HomoFusion w/ RGB RSNE\") and (c) with RSNE with features (\"HomoFusion Full\"). Results in Tab. 2 show that \"HomoFusion Full\" achieves better performance than \"HomoFusion w/o RSNE\", demonstrating the effectiveness of the iterative optimization in estimating a more accurate homography matrix that subsequently provides more accurate guidance on cross-frame spatial correspondences. It can be also observed that \"HomoFusion Full\" significantly outperforms \"HomoFusion w/ RGB RSNE\" indicating that deep features are more robust than RGB information in estimating an accurate road surface normal. This can be attributed to the fact the feature representations are less prone to adverse noises caused by environmental factors, e.g. brightness. Fig. 7 shows that the fused frame with road surface normal estimated with features has better alignment in road lines and markings than that with RGB information." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.389, + 0.279, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.467, + 0.27, + 0.478 + ], + "angle": 0, + "content": "(a) Normal estimation with RGB" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.389, + 0.482, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.467, + 0.477, + 0.478 + ], + "angle": 0, + "content": "(b) Normal estimation with feature" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.562 + ], + "angle": 0, + "content": "Figure 7: Fused frame sequence with the homography transformation between the frame pairs. The road surface normal estimated with feature representation enables more accurate homography transformation than with RGB information, resulting in better alignment of road lines and markings across frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.469, + 0.666 + ], + "angle": 0, + "content": "Robustness against Extrinsic Noise. We investigate the robustness of our method to extrinsic noise by introducing random translation errors of less than 1 meter and rotation errors of up to \\(30^{\\circ}\\). The results are shown in \"Noisy Extrinsic\" of Tab. 2, which demonstrate that our method is capable of mitigating extrinsic noise and remains robust." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Impact of Backbone. In order to ensure a fair comparison with CFFM [42] and MMA-Net [56], we implement our method with 'MiT-B1' and 'ResNet-50' backbone and report its performance in Tab. 3. Our method still outperforms CFFM [42] and MMA-Net [56] significantly. Additionally, we test our method with 'EfficientNet-B4/5' backbones, and the results show that our method has relatively stable performance even when using smaller backbones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.469, + 0.864 + ], + "angle": 0, + "content": "Application to Another Task. Given our method's ability to accurately align road surface objects across adjacent frames, we also try it with the task of detecting water hazards. Specifically, our approach aims to perform binary segmentation to identify water puddles in the road. The previ" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Table 3: Comparison of our Proposed HomoFusion with Different Backbones on ApolloScape Dataset" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.127, + 0.905, + 0.204 + ], + "angle": 0, + "content": "
BackbonesParams(M)↓GFLOPs↓18 mIoU↑36 mIoU↑
MiT-B113.6489.857.235.3
ResNet-509.12189.755.334.6
EfficientNet-B40.860.857.035.3
EfficientNet-B51.0261.257.935.2
EfficientNet-B61.2461.259.335.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.895, + 0.424 + ], + "angle": 0, + "content": "ous work on this task T3D-FCN [22] also exploits temporal information, as water puddles can appear differently at different angles and distances as a vehicle moves. We use the Puddle-1000 dataset [13], which includes On-road and Offroad datasets. The frames obtained from an onboard camera on bumpy roads are not stable and contain small extrinsic noise. The intrinsic information of the camera is provided by the dataset, while the global extrinsic information of the camera is obtained using ORB-SLAM [34]. Our experiments are conducted using the current and seven previously consecutive frames with a \\(240 \\times 320\\) image size, the same settings as T3D-FCN [22]. The results, presented in Tab. 4, demonstrate the high effectiveness of our method. More visual details of this task can be found in the Appendix." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.437, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Table 4: Comparison of water puddle segmentations on Puddle-1000 dataset" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.478, + 0.878, + 0.568 + ], + "angle": 0, + "content": "
DatasetMethodsF1-meas↑Prec↑Rec↑
On-roadFCN-8s-FL-RAU [13]0.700.680.72
T3D-FCN [22]0.680.790.62
HomoFusion (ours)0.800.810.78
Off-roadFCN-8s-FL-RAU [13]0.810.870.77
T3D-FCNU [22]0.730.870.63
HomoFusion (ours)0.870.870.87
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.619, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposed lightweight lane mark segmentation model presented in this paper offers superior performance with reduced model complexity compared to SOTA approaches. The integration of the HomoFusion module and Road Surface Normal Estimator provides an accurate classification of partially occluded, shadowed, and/or glare-affected road lines and markings by leveraging adjacent frames and pixel-to-pixel attention mechanisms. Moreover, the novel approach of using the road surface normal to guide spatial correspondences across frames has significant implications for a wide range of applications in autonomous driving, including road surface inspection, water, and other hazard detections. However, the current model relies on available camera extrinsics and is somewhat impacted by extrinsic noise. Future work will focus on optimizing both the normal vector and camera extrinsic parameters to improve the performance and make the method more robust to challenging environmental conditions. Overall, the proposed method has the potential to significantly advance the field of autonomous driving and related applications." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "4We use a heavy backbone in the competing methods as their publicly available code uses this backbone" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1082" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.267, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.115, + 0.47, + 0.146 + ], + "angle": 0, + "content": "The research is funded in part by an ARC Discovery Grant (grant ID: DP220100800) to HL." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.159, + 0.175, + 0.175 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.184, + 0.47, + 0.267 + ], + "angle": 0, + "content": "[1] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.268, + 0.47, + 0.323 + ], + "angle": 0, + "content": "[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10197-10205, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.324, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.367, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European conference on computer vision, pages 404-417. Springer, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[6] JiaWang Bian, Wen-Yan Lin, Yasuyuki Matsushita, Sai-Kit Yeung, Tan-Dat Nguyen, and Ming-Ming Cheng. Gms: Grid-based motion statistics for fast, ultra-robust feature correspondence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4181–4190, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.535, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[7] Amol Borkar, Monson Hayes, and Mark T Smith. A novel lane detection system with efficient ground truth generation. IEEE Transactions on Intelligent Transportation Systems, 13(1):365-374, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[9] Zhe Chen, Jing Zhang, and Dacheng Tao. Progressive lidar adaptation for road detection. IEEE/CAA Journal of Automatica Sinica, 6(3):693-702, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.704, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.746, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[11] Farzan Erlik Nowruzi, Robert Laganiere, and Nathalie Japkowicz. Homography estimation from image pairs with hierarchical convolutional networks. In Proceedings of the IEEE international conference on computer vision workshops, pages 913-920, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.817, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[12] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[13] Xiaofeng Han, Chuong Nguyen, Shaodi You, and Jianfeng Lu. Single image water hazard detection using fcn with re" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.184, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "flection attention units. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-120, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[14] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Paul W Holland and Roy E Welsch. Robust regression using iteratively reweighted least-squares. Communications in Statistics-theory and Methods, 6(9):813-827, 1977." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[16] Yuenan Hou, Zheng Ma, Chunxiao Liu, Tak-Wai Hui, and Chen Change Loy. Inter-region affinity distillation for road marking segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12486-12495, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[17] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Yeongmin Ko, Younkwan Lee, Shoaib Azam, Farzeen Munir, Moongu Jeon, and Witold Pedrycz. Key points estimation and point instance segmentation approach for lane detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):8949-8958, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of applied mathematics, 2(2):164-168, 1944." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[22] Juntao Li, Chuong Nguyen, and Shaodi You. Temporal 3d fully connected network for water-hazard detection. In 2019 Digital Image Computing: Techniques and Applications (DICTA), pages 1-5, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[23] Yadi Li, Liguo Chen, Haibo Huang, Xiangpeng Li, Wenkui Xu, Liang Zheng, and Jiaqi Huang. Nighttime lane markings recognition based on canny detection and hough transform. In IEEE International Conference on Real-time Computing and Robotics, pages 411-415. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Zuo-Quan Li, Hui-Min Ma, and Zheng-Yu Liu. Road lane detection with gabor filters. In International Conference on Information System and Artificial Intelligence, pages 436-440. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Jiawei Liu, Jing Zhang, Yicong Hong, and Nick Barnes. Learning structure-aware semantic segmentation with image-level supervision. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1083" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[28] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.47, + 0.235 + ], + "angle": 0, + "content": "[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.472, + 0.28 + ], + "angle": 0, + "content": "[30] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[31] Yunze Man, Xinshuo Weng, Xi Li, and Kris Kitani. Ground-net: Monocular ground plane normal estimation with geometric consistency. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2170-2178, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[32] Donald W Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the society for Industrial and Applied Mathematics, 11(2):431-441, 1963." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.384, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[33] Annika Meyer, N. Ole Salscheider, Piotr F. Orzechowski, and Christoph Stiller. Deep semantic lane segmentation for mapless driving. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 869-875, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.47, + 0.581 + ], + "angle": 0, + "content": "[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 286-291, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.47, + 0.626 + ], + "angle": 0, + "content": "[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. volume 32, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.47, + 0.684 + ], + "angle": 0, + "content": "[37] Jonah Philion. Fastdraw: Addressing the long tail of lane detection by adapting a sequential prediction network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11582-11591, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[38] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In European Conference on Computer Vision, pages 276-291. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[39] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning robust camera localization from pixels to pose. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[41] Iago Suárez, Ghesn Sfeir, José M Buenasada, and Luis Baumela. Beblid: Boosted efficient binary local image descriptor. Pattern recognition letters, 133:366-372, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[42] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.893, + 0.3 + ], + "angle": 0, + "content": "[44] Tsung-Ying Sun, Shang-Jeng Tsai, and Vincent Chan. Hsi color model based lane-marking detection. In IEEE Intelligent Transportation Systems Conference, pages 1168-1172. IEEE, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.893, + 0.383 + ], + "angle": 0, + "content": "[45] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.564 + ], + "angle": 0, + "content": "[48] Luo-Wei Tsai, Jun-Wei Hsieh, Chi-Hung Chuang, and Kuo-Chin Fan. Lane detection using directional random walks. In IEEE Intelligent Vehicles Symposium, pages 303–306. IEEE, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.893, + 0.621 + ], + "angle": 0, + "content": "[49] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.893, + 0.677 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.734 + ], + "angle": 0, + "content": "[51] Jun Wang, Tao Mei, Bin Kong, and Hu Wei. An approach of lane detection based on inverse perspective mapping. In International IEEE Conference on Intelligent Transportation Systems, pages 35-38. IEEE, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.893, + 0.802 + ], + "angle": 0, + "content": "[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[53] Ruochen Yin, Yong Cheng, Huapeng Wu, Yuntao Song, Biao Yu, and Runxin Niu. Fusionlane: Multi-sensor fusion for lane marking semantic segmentation using deep neural networks. IEEE Transactions on Intelligent Transportation Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[54] Seungwoo Yoo, Hee Seok Lee, Heesoo Myeong, Sungrack Yun, Youngwoo Park, Janghoon Cho, and Duck Hoon" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1084" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.147 + ], + "angle": 0, + "content": "Kim. End-to-end lane marker detection via row-wise classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1006-1007, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[55] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. arXiv preprint arXiv:1904.00962, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[56] Yujun Zhang, Lei Zhu, Wei Feng, Huazhu Fu, Mingqian Wang, Qingxia Li, Cheng Li, and Song Wang. Vil-100: A new dataset and a baseline model for video instance lane detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15681–15690, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.29, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[57] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1085" + } + ] +] \ No newline at end of file diff --git a/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_origin.pdf b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c1828865d690e16a0b6511c9faef3d86a82c753f --- /dev/null +++ b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/44fc0793-39bb-4f98-abb1-f6daedfcefc2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81250f11e35d425c1affc482337ed0a2fd22103ff30795bc1712ddc621d4b1b2 +size 7112992 diff --git a/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/full.md b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35bd9e598d73fe00282289e677c198b2d2afec28 --- /dev/null +++ b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/full.md @@ -0,0 +1,352 @@ +# Homography Guided Temporal Fusion for Road Line and Marking Segmentation + +Shan Wang $^{1,2}$ Chuong Nguyen $^{1}$ Jiawei Liu $^{2}$ Kaihao Zhang $^{2}$ Wenhan Luo $^{3}$ Yanhao Zhang $^{2}$ Sundaram Muthu $^{1}$ Fahira Afzal Maken $^{1}$ Hongdong Li $^{2}$ $^{1}$ Data61, CSIRO $^{2}$ Australian National University $^{3}$ Sun Yat-sen University + +# Abstract + +Reliable segmentation of road lines and markings is critical to autonomous driving. Our work is motivated by the observations that road lines and markings are (1) frequently occluded in the presence of moving vehicles, shadow, and glare and (2) highly structured with low intra-class shape variance and overall high appearance consistency. To solve these issues, we propose a Homography Guided Fusion (HomoFusion) module to exploit temporally-adjacent video frames for complementary cues facilitating the correct classification of the partially occluded road lines or markings. To reduce computational complexity, a novel surface normal estimator is proposed to establish spatial correspondences between the sampled frames, allowing the HomoFusion module to perform a pixel-to-pixel attention mechanism in updating the representation of the occluded road lines or markings. Experiments on ApolloScape, a largescale lane mark segmentation dataset, and ApolloScape Night with artificial simulated night-time road conditions, demonstrate that our method outperforms other existing SOTA lane mark segmentation models with less than $9\%$ of their parameters and computational complexity. We show that exploiting available camera intrinsic data and ground plane assumption for cross-frame correspondence can lead to a light-weight network with significantly improved performances in speed and accuracy. We also prove the versatility of our HomoFusion approach by applying it to the problem of water puddle segmentation and achieving SOTA performance1. + +# 1. Introduction + +Lane mark segmentation aims to achieve pixel-wise classification of road lines and markings simultaneously. Known for its quintessential importance in autonomous driving, lane mark segmentation is also an effective tool for constructing accurate High Definition (HD) maps. Existing works solve individual sub-tasks, e.g. drivable area segmen + +![](images/ebb967392b1de54130f4ddb8f633a65435ddcef2c0f073ca19f88caf61981358.jpg) +Figure 1: Illustration of the effect of the proposed HomoFusion module that explores the adjacent frames for cues, facilitating the correct classification of (1) a "Straight Arrow", which, with its bottom half occluded by a vehicle, is mistakenly classified as a "Right Turn & Straight Arrow" without the HomoFusion module, and (2) a partially occluded "Dotted Line", which is incorrectly classified as a "Solid Line" without the HomoFusion module. The fused frame in the $4^{th}$ row and the $2^{nd}$ column demonstrates the recovered road lines and markings after projecting the previous frames onto the current frame with the estimated homography matrices. The yellow box enlarges the area where mistake classifications are corrected. The red box indicates the spatially corresponding area across the frames. Best viewed in color. + +tation [33, 9] and lane detection [19, 35], while few address the lane mark segmentation in its entirety [53, 16]. + +Despite the tremendous progresses in semantic/scene segmentation [52, 42, 26], little attention has been paid to the lane mark segmentation task [53, 16]. Yin et al. [53] leverages additional LiDAR information, merging the segmented visual information with the point clouds, to achieve lane mark segmentation. IntRA-KD [16] applies the knowledge distillation technique to improve the efficiency of the lane mark segmentation model. + +A major challenge in the lane mark segmentation task is + +partial occlusion of the road lines and markings caused by the vehicle and surrounding environment leading to false classifications. For example, a partially occluded straight arrow can be easily mistaken as a right turn & straight arrow, as depicted in Fig. 1. However, existing lane mark segmentation methods have not yet made use of the following observations: (1) occlusion of the road lines and markings (due to nearby moving vehicles, shadow, and glare) can be reduced by cross-frame consistency; (2) road lines and markings are highly structured, maintaining high consistency in intra-class shapes as well as in overall appearance, alleviating the requirements on learning complicated features to distinguish different classes with high intra-class variance as in semantic/scene segmentation. + +To find complementary information from adjacent frames, we use the ground plane assumption for the road region immediately in front of the camera. Our Homography Guided Fusion (HomoFusion) module achieves temporally consistent representation, recovers partially occluded road lines or markings, and leads to correct classification. The module's success depends on accurate cross-frame spatial correspondence, which we achieve using a homography transformation matrix estimated with available intrinsic and extrinsic parameters of the on-vehicle camera, and the road surface normal estimated by our novel optimization method called Road Surface Normal Estimator (RSNE). + +Using the highly structured nature of road lines and markings could reduce the complexity of the detection problem and make it suitable to run on edge devices. To this end, we employ (1) a lightweight encoder to represent the visual information in feature spaces, and (2) a cross-frame pixel-to-pixel attention mechanism in our HomoFusion module, instead of a more computationally expensive global attention mechanism employed by other methods. These design decisions allow the proposed model to outperform the existing lane mark segmentation models with less than $9\%$ of their parameters and Giga Floating Point Operations (GFLOPs) We summarize our contributions as follows: + +- We propose a HomoFusion module that uses ground plane assumption and the adjacent frames for temporally consistent representations for accurate classification of partially occluded road lines and markings. +- We design a novel estimator RSNE for road surface normal, which, combined with camera intrinsic and extrinsic parameters readily available on an autonomous vehicle, yields accurate homography matrix between frame pairs. RSNE simplifies the 8 degrees of freedom (DoF) of homography problem to a 2 DoF of normal vector problem. +- We present a lightweight lane mark segmentation model that achieves better performance than the state + +of-the-art (SOTA) methods with significantly reduced model complexity and computation requirements. + +# 2. Related Work + +# 2.1. Lane Detection + +Traditional lane detection approaches rely on handcrafted features such as color [44, 51], edge [23] and texture [25], which are limited in complex scenarios [48, 7]. Recent advances in deep neural networks have led to significant performance improvements, with methods like message-passing networks [36] and attention-based modules [45]. [16, 37, 38, 54, 27] formulate lane detection as a row-wise classification task based on grid division of the input image. PolyLaneNet [46] is the first parametric prediction method, which outputs polynomials to represent each lane. BEV-Former [24] and PETRv2 [28] employ transformer mechanisms to streamline the conversion of perspective views into bird's-eye views, thereby eliminating variations in object size caused by perspective. Although the above methods achieve impressive performance, all of them focus on detecting the lane lines and ignore the basic road elements like arrow signal in Fig. 1. Furthermore, lane detection differs from segmentation tasks in that it focuses on identifying the boundaries of the drivable area, rather than the real shape of the lanes. + +# 2.2. Road Line and Marking Segmentation + +The identification of road elements is crucial for ensuring safety in autonomous driving systems, but there are few works addressing this problem. Hou et al. [16] introduced a distillation approach that demonstrates competitive performance in lane mark segmentation. Yin et al. [53] used an LSTM-based network to segment images with DeepLabv3+ [8] and then merged them with point clouds to assist with lane segmentation. Unfortunately, these methods do not address the challenges of occlusion, shadows, and glare that frequently occur in real-world driving scenarios. Recently, Zhang et al. [56] introduced global memory information from previous frames to enhance local information for video lane segmentation. However, this method introduces attention to the entire image, making it memory-intensive and inefficient for autonomous driving scenarios. + +# 2.3. Homography Estimation + +Homography estimation methods can be divided into non-deep and deep learning-based approaches. Non-deep methods estimate the homography using feature extraction, feature matching, and outlier rejection. SIFT [29], SURF [5, 4], ORB [39], LPPM [30], GMS [6], and BEBLID [41] are commonly used for feature extraction, and RANSAC [12], MAGAC [2], and LRLS [15] are applied for outlier rejection. Recently, many deep learning-based approaches + +![](images/043b4b782b333726fa0cdd0c646fe96ea82b98dda31efae2f4d8de990b5b8b36.jpg) +Figure 2: Overview of our proposed model consisting of a pair of lightweight encoder and decoder, our proposed HomoFusion module, and our proposed Road Surface Normal Estimator (RSNE). A sequence of frames $\mathbf{I}$ , including a target frame $\mathbf{I}_{\mathbf{t}}$ and $n - 1$ previous frames, are encoded into the feature representations $(\mathbf{F}^l)$ . RNSE estimates the road surface normal vector, which, combined with the camera intrinsic and extrinsic parameters, yields a homography matrix between each frame pair, establishing cross-frame spatial correspondences. HomoFusion uses pixel-to-pixel attention mechanism to obtain temporally consistent representation for on-road pixels of the current frame with the spatial correspondence across frames as guidance. Finally, the decoder decodes and upsamples the temporally consistent feature representations to produce the lane mark segmentation prediction $(\mathbf{S}_{\mathbf{t}}^{\mathrm{prd}})$ . + +have been proposed, such as DeTone et al. [10]'s VGG-like architecture, Nowruzi et al. [11], Le et al. [20]'s and Man et al. [31]'s cascaded VGG-like networks, and Chang et al. and Zhao et al.'s incorporation of the Lucas-Kanade (LK) algorithm with deep networks. However, these methods do not utilize known camera intrinsic and extrinsic parameters across frames for homography matrix estimation, which can effectively reduce the search dimensions. + +# 3. Method + +# 3.1. Overview + +As illustrated in Fig. 2, our proposed framework takes a sequence of frames, denoted as $\mathbf{I} = \{\mathbf{I}_i|i = \{t - (s - 1)\Delta t\}_{s = 1}^n\}$ , as input. This sequence consists of a current frame $\mathbf{I}_t$ and $n - 1$ previous frames sampled at a fixed time interval $\Delta t$ . The output of the framework is a lane mark segmentation map $\mathbf{S}_t$ for the current frame. In addition, the framework includes a novel HomoFusion module, which uses the homography transformation guided by the proposed RSNE to fuse the feature map of the current frame with those of the previous frames. We denote the encoded feature representation for the $t^{th}$ frame as $\mathbf{F}_t = \{\mathbf{F}_t^l\in \mathbb{R}^{H_l\times W_l\times C_l}\}_{l = 1}^L$ where $l$ indicates the + +level of the feature map. The shallowest and deepest levels are represented by 1 and $L$ , respectively. $H_{l}$ , $W_{l}$ , and $C_l$ represent the height, width, and channel number of feature maps in level $l$ . We denote all encoded features as $\mathbf{F} = \{\mathbf{F}_i^l | i = \{t - (s - 1)\Delta t\}_{s = 1}^n\}$ . The proposed HomoFusion module, detailed in Sec. 3.2, uses a cross-frame pixel-to-pixel attention mechanism to fuse the feature map $(\mathbf{F}_f = \{\mathbf{F}_f^l\}_{l = 1}^L)$ of the current frame with those of the previous frames. The homography transformation matrix between the target frame and each of the previous frames is calculated using the estimated normal vector of the road surface $\mathbf{n}$ obtained by the proposed RSNE, as described in Sec. 3.3. Finally, the decoder produces the segmentation prediction $(\mathbf{S}_t)$ for the target frame based on the fused feature map. + +# 3.2. Homography Guided Fusion (HomoFusion) + +Our proposed HomoFusion module employs a pixel-to-pixel attention mechanism to fuse spatially corresponding pixels across frames. This mechanism is achieved by projecting the pixels of the current frame onto the previous frames through an accurate homography transformation. Since road lines and markings are strictly on-road, the search area can be limited to the road surface, which is mostly a plane, at least within the immediate front of the vehicle where markings are readable. Therefore, only the homography transformation is needed to accurately map road pixels between frames. According to the standard inverse homography [14, 57, 49], for each on-road pixel $p_t =$ + +![](images/2a75d7d795a37273ec68a561ec4cc0ba9232f02ea232f8f195b482d5ce704e33.jpg) +Figure 3: Illustration of sample points. (Right) Sample points in the current frame. (Left/Middle) Correspondence of a sample point in previous frames. The red point coordinate is calculated by using the correct normal, while the cyan point coordinate is calculated by using the initial (incorrect) normal. + +$(u,v)^{\top}$ of the current frame, its spatially corresponding pixels of the reference frames $\{p_i|i = \{t - (s - 1)\Delta t\}_{s = 1}^n\}$ can be computed using Eq. (1). + +$$ +p _ {i} \propto \mathbf {K} \left(\mathbf {R} _ {i} - \frac {\mathbf {t} _ {i} \mathbf {n} ^ {\top}}{d}\right) \mathbf {K} ^ {- 1} \left(p _ {t} \oplus 1\right), \tag {1} +$$ + +where $\propto$ represents proportional, $\mathbf{K}$ is the intrinsic matrix of the on-vehicle camera, $\mathbf{R}_i$ and $\mathbf{t}_i$ represent the relative rotation and translation between the current frame $\mathbf{I}_t$ and the $i^{th}$ previous frame $\mathbf{I}_i$ . $\mathbf{n}$ is the normal of the observed road surface, which is estimated in Sec. 3.3. $d$ is the vertical distance between the on-vehicle camera and the road surface obtained from the camera calibration. $\oplus 1$ converts the point coordinates to homogeneous coordinates. The on-road pixels of the current frame are sampled from a triangular area in front of the car, as illustrated in Fig. 3. Figure 3 also demonstrates that the shape and category of a left-turn can be recovered by exploring the spatially corresponding area in the previous frames under the guidance of an accurate homography transformation. + +With the spatial correspondence across frames, the proposed HomoFusion module computes a temporally consistent representation for each sampled pixel of the current frame by employing temporal attention mechanism on the feature presentations of spatially corresponding pixels. The mechanism assigns the pixel representation of current frame $(\mathbf{F}_t[p_t])$ as a query and all frames $(\{\mathbf{F}_i[p_i]\})$ as keys and values. In contrast to the existing attention mechanism [50], we propose to (1) omit the spatial encoding in the presence of obtained spatial correspondence across frames, and (2) apply an L2 normalization on the query and the keys to obtain more robust (deep) feature similarities (Eq. 2) across various lighting environments. + +$$ +\mathbf {a} _ {i} = \frac {\mathbf {F} _ {t} [ p _ {t} ]}{\| \mathbf {F} _ {t} [ p _ {t} ] \| _ {2}} \cdot \frac {\mathbf {F} _ {i} [ p _ {i} ]}{\| \mathbf {F} _ {i} [ p _ {i} ] \| _ {2}}, \tag {2} +$$ + +where $\| .\| _2$ indicates an $L_{2}$ normalization. The similarity is normalized by softmax function as given in Eq. 3: + +$$ +\mathbf {W} _ {i} = \frac {\exp \mathbf {a} _ {i}}{\sum_ {i} \exp \mathbf {a} _ {i}}, \tag {3} +$$ + +before being used as weight to fuse the spatially corresponding pixels across the frames as shown in Eq. 4: + +$$ +\mathbf {F} _ {f} [ p _ {t} ] = \mathbf {F} _ {t} [ p _ {t} ] + \sum_ {i} \mathbf {W} _ {i} \mathbf {F} _ {i} [ p _ {i} ]. \tag {4} +$$ + +# 3.3. Road Surface Normal Estimator (RSNE) + +Accurate estimation of road surface normal is crucial for establishing spatial correspondence across frames. Generally, on-vehicle cameras have a fixed angle with respect to the ground level. However, the actual road surface normal can vary due to the unevenness of the road surface, e.g. uphill or downhill roads, and sloped turns, etc. Assuming that the initial road surface normal is a vector being perpendicular to a horizontal plane and pointing upwards, we propose to obtain an accurate road surface normal with an optimization process that repeats for each new coming frame. The proposed RSNE iteratively updates the estimated road surface normal with the projection error of on-road pixels from the current frame onto the previous frames. Specifically, we sample $m$ pixels $\{p_t^j\}_{j=1}^m$ from the current frame that are likely to be on-road, as illustrated in Fig. 3, and compute the spatially corresponding pixels $\{p_i^j\}_{j=1}^m$ in the previous frames using the homography matrix computed with the estimated road surface normal as defined in Eq. 1. The estimated road surface normal is optimized in an iterative manner through the following procedure: (a) computing a residual between the low-level features of spatially corresponding pixels across the frames with Eq. 5 + +$$ +\mathbf {r} _ {i} ^ {j} = \mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ] - \mathbf {F} _ {t} ^ {1} [ p _ {t} ^ {j} ] \in \mathbb {R} ^ {c}, \tag {5} +$$ + +where the residual is set to 0 for the sample point whose projections in previous frames fall outside the frame area, and computing the overall error with Eq. 6 + +$$ +\mathbf {E} = \sum_ {i, j} \rho \left(\left\| \mathbf {r} _ {i} ^ {j} \right\| _ {2} ^ {2}\right), \tag {6} +$$ + +where $\| .\| _2^2$ shows squared norm and $\rho (\cdot)$ is a robust cost function proposed by Barron [3]. This nonlinear least-square cost is iteratively minimized towards correct normal estimation using the Levenberg-Marquardt algorithm [21, 32]; (b) computing a new road surface normal value by adding the update given by Eq. 12. + +The road surface normal vector is a unit vector with two degrees of freedom (DoF), which can be decomposed into a pitch angle $\theta$ and a roll angle $\phi^3$ , as shown in Eq. 7. + +$$ +\mathbf {n} = \left(- \sin \phi \cos \theta , - \cos \phi \cos \theta , \sin \theta\right) ^ {\top}. \tag {7} +$$ + +The Jacobian of the residual function w.r.t. the pitch angle and the roll angle is defined in Eq. 8 by applying the chain + +rule, + +$$ +\mathbf {J} _ {i} ^ {j} = \frac {\partial \mathbf {r} _ {i} ^ {j}}{\partial \delta} = \frac {\partial \mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ]}{\partial p _ {i} ^ {j}} \frac {\partial p _ {i} ^ {j}}{\partial \mathbf {n}} \frac {\partial \mathbf {n}}{\partial \delta}, \tag {8} +$$ + +where $\delta$ represents the to-be-optimized target ( $\theta$ or $\phi$ ), and $\frac{\partial \mathbf{F}_i^1[p_i^j]}{\partial p_i^j}$ is gradient from 2D interpolation. $\frac{\partial p_i^j}{\partial \mathbf{n}}$ is the Jacobian of the homography transformation (Eq. (1)) w.r.t. $\mathbf{n}$ , as defined in Eq. 9 + +$$ +\frac {\partial p _ {i} ^ {j}}{\partial \mathbf {n}} = - \frac {1}{d} \mathbf {K t} _ {i} \left(\mathbf {K} ^ {- 1} \left(p _ {t} ^ {j} \oplus 1\right)\right) ^ {\top}, \tag {9} +$$ + +and $\frac{\partial\mathbf{n}}{\partial\delta}$ , defined in Eq. 10 and Eq. 11, is the Jacobian of Eq. (7) w.r.t. $\theta$ and $\phi$ respectively + +$$ +\frac {\partial \mathbf {n}}{\partial \theta} = \left(- \frac {\mathbf {n} _ {1} \mathbf {n} _ {3}}{\sqrt {1 - \mathbf {n} _ {3} ^ {2}}}, - \frac {\mathbf {n} _ {2} \mathbf {n} _ {3}}{\sqrt {1 - \mathbf {n} _ {3} ^ {2}}}, \sqrt {1 - \mathbf {n} _ {3} ^ {2}}\right), \tag {10} +$$ + +$$ +\frac {\partial \mathbf {n}}{\partial \phi} = \left(\mathbf {n} _ {2}, - \mathbf {n} _ {1}, 0\right), \tag {11} +$$ + +where $\mathbf{n}_k$ represents the $k$ -th element of $\mathbf{n}$ . The detailed derivation is included in appendix (Sec. 1). We obtain Hessian matrices as $\mathbb{H} = \mathbf{J}^{\top}\rho^{\prime}\mathbf{J}$ , where $\rho^\prime$ is the derivative of function $\rho (\cdot)$ . We compute the update by damping the Hessian and solving the linear system, as shown in + +$$ +\triangle \delta = - (\mathbb {H} + \lambda \operatorname {d i a g} (\mathbb {H})) ^ {- 1} \mathbf {J} ^ {\top} \rho^ {\prime} \mathbf {r}, \tag {12} +$$ + +where $\lambda$ is the damping factor [40], balancing between the Gauss-Newton $(\lambda = 0)$ and gradient descent $(\lambda = \infty)$ . Our entire optimization process is differentiable. + +The algorithm is illustrated in Alg. 1. The normal vector $\mathbf{n}$ uses low-level features $\mathbf{F}^1$ . While the high-level features $\mathbf{F}^2$ focus on segmentation feature expression, the low-level features $\mathbf{F}^1$ are supervised to extract useful features for both segmentation and road surface normal estimation. As both tasks focus on on-road objects, they can benefit each other. + +# 4. Experiments + +# 4.1. Implementation Details + +We use a pre-trained EfficientNet [47] to extract image features at two different scales, $4 \times$ and $16 \times$ down-scaling, with 64 and 128 channels, respectively. To focus on the bottom region of the input image, where road marks are typically visible, we crop only the bottom $40\%$ of the input image, following the same approach in the prior work [16]. The processed image size is set to $272 \times 848$ . Our decoder consists of bi-linear up-sampling layers and convolution layers, which up-sample the high-level features and increase the resolution by a factor of 4 at each level. We trained our model using the AdamW optimizer [55] for 30,000 iterations on two NVIDIA RTX 3090 GPUs, with a learning rate of $4 \times 10^{-3}$ . + +Algorithm 1 Normal Optimization +Input: +The 1-st Level Feature Maps: $\mathbf{F}^1 = \{\mathbf{F}_i^1\}$ +Sample Points Coordinates: $p_t = \{p_t^j\}_{j = 1}^m$ +Initial Pitch and Roll Value: $(\theta_0,\phi_0) = (0.15,0)$ +Damping Factor: $\lambda = (\lambda_{\theta}\lambda_{\phi})$ +Hyper-parameter: +k; $\triangleright$ Maximum loop count, empirically set to 20 + $\alpha$ ; $\triangleright$ Convergence threshold, empirically set to 0.0001 +Output: Optimized Pitch and Roll $(\theta ,\phi)$ . +1: function OPT( $\theta_0,\phi_0,\mathbf{F}^1,p_t,\lambda$ +2: Derive point features $\mathbf{F}_t^1 [p_t]$ from $\mathbf{F}_t^1$ . +3: for $k\gets 1$ to k do +4: Calculate point coordinates $p_i$ in $\mathbf{F}_i^1$ (Eq. 1); +5: Derive point features $\mathbf{F}_i^1[p_i]$ from $\mathbf{F}_i^1$ . +6: Calculate residual $\mathbf{r}_i$ (Eq. 5); +7: Calculate observe error E (Eq. 6); +8: Calculate robust cost $\rho (\mathbf{E})$ and its derivation $\rho^\prime$ +9: Construct J and its Hessian matrices $\mathbb{H}$ (Eq. 8); +10: Obtain $\triangle \delta$ by Cholesky decomposition (Eq. 12); +11: Update Normal as $(\theta ,\phi)_k\gets (\theta ,\phi)_{k - 1} + \triangle \delta$ +12: if MAX(△δ)< α then +13: Break; +14: end if +15: end for +16: end function + +# 4.2. Datasets + +We conduct our experiments on two datasets, ApolloScape [17] and ApolloScape Night. Other datasets were deemed unsuitable due to the absence of camera extrinsics, as observed in SDLane [18] and VLI-100 [56], inadequate frame overlap, or the lack of diverse segmentation labels, exemplified by Waymo Open Dataset [43]. Further elaboration on these reasons can be found in the appendix. + +ApolloScape. The ApolloScape dataset is a large-scale dataset that can be used for localization and segmentation tasks. It consists of 38 distinct classes and poses various challenges including occlusions and tiny road markings. To provide accurate camera poses, the vehicle is equipped with customer-grade GPS/IMU. However, for the lane mark segmentation task, only 41, 201 annotated images out of more than $110,000+$ are associated with camera pose information. We used these 41, 201 images for our experiments since our approach relies on camera poses. We divided them into 35, 173 training images and 6, 028 validation images. Because our approach uses sequential information, we select the validation set with completely different trajectories to ensure a fair comparison. + +ApolloScape Night. Since there is a lack of datasets for + +Table 1: Comparison with SOTA Methods on the ApolloScape [17] and Apolloscape Night Datasets and Running on an NVIDIA RTX 3090 GPU. "18 mIoU" Represents the Mean IoU of 18 Types of Lane Markers Selected by ApolloScape Official Metrics. "36 mIoU" Represents the Mean IoU of All Unignorable 36 Types. + +
MethodsFrame CountBackboneParams (M)↓GFLOPs↓FPS (f/s)↑ApolloScape [17]ApolloScape Night
18 mIoU↑36 mIoU↑18 mIoU↑36 mIoU↑
IntRA-KD[16]1ResNet-10165.65159.410.842.124.629.816.7
SegFormer[52]1MiT-B113.51048.843.852.332.138.323.1
CFFM [42]4MiT-B115.31192.622.753.232.739.123.6
MMA-Net [56]4ResNet-5057.9723.220.652.931.438.823.2
HomoFusion(ours)4EfficientNet-B61.2461.225.459.335.944.826.6
+ +![](images/5bc9f79a72ab999a35f2820e2b4cc188e81285273cf6ba6e6044218e683cb632.jpg) +Figure 4: Sample images from the ApolloScape Night dataset. Top: original daytime images from the ApolloScape dataset. Bottom: synthesized night-time images. + +night-time lane mark segmentations, we created an artificial dataset called ApolloScape Night from the ApolloScape dataset using a cross-domain generation network [1]. This allows us to evaluate our proposed model on a challenging dataset with poor lighting conditions, road reflection, and glare. Fig. 4 displays some sample images from the dataset. + +# 4.3. Evaluation Metrics + +In accordance with the guidelines of the ApolloScape benchmark [17], we used mean intersection-over-union (mIoU) as the evaluation criterion. ApolloScape contains 38 different types of lane markers, including two ignorable labels (noise and ignored). Of these labels, 18 categories are used in the official evaluation metrics. To provide a comprehensive evaluation of our approach, we report the evaluated mIoU results for both the selected 18 categories, denoted as "18 mIoU", and all 36 categories, denoted as "36 mIoU". By reporting both sets of results, we provide a more complete picture of the performance of our method on the ApolloScape and ApolloScape Night datasets. + +# 4.4. Comparison with State-of-the-art Methods + +To evaluate the performance of our proposed method, we compared it with SOTA algorithms, including (a) IntRA-KD [16], (b) SegFormer [52], (c) CFFM [42], and (d) MMA-Net [56], on the ApolloScape [17] and ApolloScape + +Night datasets. To ensure a fair comparison, we retrained each model on the same subset of the training set with the same input resolution. + +The results of the comparison are presented in Table 1. Our approach outperforms the existing SOTA methods on both the ApolloScape [17] and ApolloScape Night datasets while having less than $9\%$ of their parameters or computational overhead. + +In addition, we compared our proposed method with CFFM [42] which also uses adjacent frames to enhance the representation of the current frame. Our method is more efficient in terms of model complexity, with a complexity of $\mathcal{O}(HWC)$ for feature extraction and $\mathcal{O}(HWC^2)$ for cross-frame feature fusion. In contrast, the complexity of CFFM is $\mathcal{O}(H^2 W^2 C) + \mathcal{O}(HWC^2)$ for feature extraction and $\mathcal{O}(HWEC) + \mathcal{O}(HWC^2)$ for cross-frame feature fusion, where $E$ is calculated by their receptive field and pooling kernel size. This explains why our method has a lower computational requirement for the cross-frame pixel-to-pixel attention mechanism compared to the global attention mechanism used by CFFM. + +The qualitative results of both the existing SOTA models and our proposed model are shown in Fig. 5. The results indicate that our approach achieves superior segmentation performance on road lines and markings, even under adverse conditions such as occlusion, road reflection, and poor lighting. Additional visualization examples for various categories can be found in the appendix. + +# 4.5. Study on Hyper-parameters + +Number of Frames $(n)$ . We conducted a study on the impact of the number of frames while setting the sampling frame gap $\triangle t = 2$ . The results presented in Fig. 6 (a)-(b) demonstrate that the performance of our method improves on both ApolloScape and ApolloScape Night datasets with an increase in the number of explored adjacent frames. However, the performance improvement becomes negligible when the number of frames exceeds 4. Considering the monotonically increasing relationship between the model complexity and the number of frames, we set $n = 4$ to strike a balance between model performance and complexity. + +Sampling Frame Gap $(\triangle t)$ . In the study on the effect of the sampling frame gap, we set the number of frames + +![](images/046f66b36e313b4e0fc1979894796229282eace5ffa43923391516698af4f7c3.jpg) +Figure 5: Qualitative comparison with SOTA methods. The top two examples are from the ApolloScape [17] dataset, and the bottom two examples are from the ApolloScape Night dataset. Yellow boxes highlight the area of interest for better visualization. Red boxes indicate false-positive segmentation predictions. Best viewed in color. + +![](images/103c965265be01852973f356baec3bb1419bbb9021e1fe98c4410ea4afe1a6ef.jpg) +(a) Number of frames (AS)) + +![](images/a63926b1780a60b9eb07e97b60b996f7c5dc010d54b4f73ce472af1b67e06b53.jpg) +(b) Number of frames (ASN) + +![](images/06eb612b9cb19730fb07a488a6076dff637c62626bf4ff98cb2ef760d9140b31.jpg) +(c) Sampling frame gap (AS) + +![](images/46770a1b70147801b354797c55666d9aaf5cdf1b069122278de9ab7012510501.jpg) +(d) Sampling frame gap (ASN) +Figure 6: The impact of the number of frames $n$ on ApolloScape (AS) and ApolloScape Night (ASN) datasets is shown in (a) and (b), while (c) and (d) show the effect of the sampling frame gap $\triangle t$ . + +$n = 4$ . Fig. 6 (c)-(d) indicates that the optimal performance is achieved at $\triangle t = 2$ . This can be attributed to the balance between the sufficient overlap of road area across the frames and a wide enough temporal gap that allows the displacement of the source of occlusion, e.g. moving vehicle. + +# 4.6. Discussion + +All ablation studies are conducted on the ApolloScape dataset with the same training strategies described in + +Table 2: Ablation Study of our Proposed HomoFusion and RSNE on the ApolloScape Dataset + +
18 mIoU↑36 mIoU↑
w/o HomoFusion49.431.0
HomoFusion w/o HomoGuide51.831.3
HomoFusion w/o RSNE57.435.3
HomoFusion w/ RGB RSNE54.933.8
Noisy Extrinsic57.535.1
HomoFusion Full59.335.9
+ +Sec. 4.1. We set the number of frames to $n = 4$ and a sampling frame gap of $\triangle t = 2$ . + +Impact of Homofusion. The core component of our approach is the Homofusion module, which allows the use of complementary information from sequential frames. Tab. 2 demonstrates that our model with the Homofusion module (4 frames) achieves better performance than the variant without Homofusion module (1 frame), illustrating the effectiveness of the temporally consistent feature representation obtained with Homofusion module in addressing the partial occlusion issues. We also study the effect directly associating the pixels across frames based on their coordinates by using identity matrix as the homography matrix, denoted as "w/o HomoGuide". It shows that such an incorrect cross-frame spatial correspondence, although leveraging additional frames, provides limited help in improving the model performance. + +Impact of RSNE. We study the impact of the proposed RSNE in the following model variants each comprising the HomoFusion modules: (a) without RSNE ("HomoFusion + +w/o RSNE"), where the homography matrix is estimated with the initial road surface normal $(\mathbf{n}_0)$ , (b) with RSNE using RGB information ("HomoFusion w/ RGB RSNE") and (c) with RSNE with features ("HomoFusion Full"). Results in Tab. 2 show that "HomoFusion Full" achieves better performance than "HomoFusion w/o RSNE", demonstrating the effectiveness of the iterative optimization in estimating a more accurate homography matrix that subsequently provides more accurate guidance on cross-frame spatial correspondences. It can be also observed that "HomoFusion Full" significantly outperforms "HomoFusion w/ RGB RSNE" indicating that deep features are more robust than RGB information in estimating an accurate road surface normal. This can be attributed to the fact the feature representations are less prone to adverse noises caused by environmental factors, e.g. brightness. Fig. 7 shows that the fused frame with road surface normal estimated with features has better alignment in road lines and markings than that with RGB information. + +![](images/0e6f905dac6f480fd4bfb46d1b768cf0e903e46531c93fda4c6cd168dc297fb1.jpg) +(a) Normal estimation with RGB + +![](images/400830723900dd0292a3d335e18fd6abd7cc11ae9670381bc54e9b6aa19b9042.jpg) +(b) Normal estimation with feature +Figure 7: Fused frame sequence with the homography transformation between the frame pairs. The road surface normal estimated with feature representation enables more accurate homography transformation than with RGB information, resulting in better alignment of road lines and markings across frames. + +Robustness against Extrinsic Noise. We investigate the robustness of our method to extrinsic noise by introducing random translation errors of less than 1 meter and rotation errors of up to $30^{\circ}$ . The results are shown in "Noisy Extrinsic" of Tab. 2, which demonstrate that our method is capable of mitigating extrinsic noise and remains robust. + +Impact of Backbone. In order to ensure a fair comparison with CFFM [42] and MMA-Net [56], we implement our method with 'MiT-B1' and 'ResNet-50' backbone and report its performance in Tab. 3. Our method still outperforms CFFM [42] and MMA-Net [56] significantly. Additionally, we test our method with 'EfficientNet-B4/5' backbones, and the results show that our method has relatively stable performance even when using smaller backbones. + +Application to Another Task. Given our method's ability to accurately align road surface objects across adjacent frames, we also try it with the task of detecting water hazards. Specifically, our approach aims to perform binary segmentation to identify water puddles in the road. The previ + +Table 3: Comparison of our Proposed HomoFusion with Different Backbones on ApolloScape Dataset + +
BackbonesParams(M)↓GFLOPs↓18 mIoU↑36 mIoU↑
MiT-B113.6489.857.235.3
ResNet-509.12189.755.334.6
EfficientNet-B40.860.857.035.3
EfficientNet-B51.0261.257.935.2
EfficientNet-B61.2461.259.335.9
+ +ous work on this task T3D-FCN [22] also exploits temporal information, as water puddles can appear differently at different angles and distances as a vehicle moves. We use the Puddle-1000 dataset [13], which includes On-road and Offroad datasets. The frames obtained from an onboard camera on bumpy roads are not stable and contain small extrinsic noise. The intrinsic information of the camera is provided by the dataset, while the global extrinsic information of the camera is obtained using ORB-SLAM [34]. Our experiments are conducted using the current and seven previously consecutive frames with a $240 \times 320$ image size, the same settings as T3D-FCN [22]. The results, presented in Tab. 4, demonstrate the high effectiveness of our method. More visual details of this task can be found in the Appendix. + +Table 4: Comparison of water puddle segmentations on Puddle-1000 dataset + +
DatasetMethodsF1-meas↑Prec↑Rec↑
On-roadFCN-8s-FL-RAU [13]0.700.680.72
T3D-FCN [22]0.680.790.62
HomoFusion (ours)0.800.810.78
Off-roadFCN-8s-FL-RAU [13]0.810.870.77
T3D-FCNU [22]0.730.870.63
HomoFusion (ours)0.870.870.87
+ +# 5. Conclusion + +The proposed lightweight lane mark segmentation model presented in this paper offers superior performance with reduced model complexity compared to SOTA approaches. The integration of the HomoFusion module and Road Surface Normal Estimator provides an accurate classification of partially occluded, shadowed, and/or glare-affected road lines and markings by leveraging adjacent frames and pixel-to-pixel attention mechanisms. Moreover, the novel approach of using the road surface normal to guide spatial correspondences across frames has significant implications for a wide range of applications in autonomous driving, including road surface inspection, water, and other hazard detections. However, the current model relies on available camera extrinsics and is somewhat impacted by extrinsic noise. Future work will focus on optimizing both the normal vector and camera extrinsic parameters to improve the performance and make the method more robust to challenging environmental conditions. Overall, the proposed method has the potential to significantly advance the field of autonomous driving and related applications. + +# 6. Acknowledgements + +The research is funded in part by an ARC Discovery Grant (grant ID: DP220100800) to HL. + +# References + +[1] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. +[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10197-10205, 2019. +[3] Jonathan T. Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008. +[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European conference on computer vision, pages 404-417. Springer, 2006. +[6] JiaWang Bian, Wen-Yan Lin, Yasuyuki Matsushita, Sai-Kit Yeung, Tan-Dat Nguyen, and Ming-Ming Cheng. Gms: Grid-based motion statistics for fast, ultra-robust feature correspondence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4181–4190, 2017. +[7] Amol Borkar, Monson Hayes, and Mark T Smith. A novel lane detection system with efficient ground truth generation. IEEE Transactions on Intelligent Transportation Systems, 13(1):365-374, 2011. +[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. +[9] Zhe Chen, Jing Zhang, and Dacheng Tao. Progressive lidar adaptation for road detection. IEEE/CAA Journal of Automatica Sinica, 6(3):693-702, 2019. +[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. +[11] Farzan Erlik Nowruzi, Robert Laganiere, and Nathalie Japkowicz. Homography estimation from image pairs with hierarchical convolutional networks. In Proceedings of the IEEE international conference on computer vision workshops, pages 913-920, 2017. +[12] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. +[13] Xiaofeng Han, Chuong Nguyen, Shaodi You, and Jianfeng Lu. Single image water hazard detection using fcn with re + +flection attention units. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-120, 2018. +[14] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. +[15] Paul W Holland and Roy E Welsch. Robust regression using iteratively reweighted least-squares. Communications in Statistics-theory and Methods, 6(9):813-827, 1977. +[16] Yuenan Hou, Zheng Ma, Chunxiao Liu, Tak-Wai Hui, and Chen Change Loy. Inter-region affinity distillation for road marking segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12486-12495, 2020. +[17] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018. +[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In CVPR, 2022. +[19] Yeongmin Ko, Younkwan Lee, Shoaib Azam, Farzeen Munir, Moongu Jeon, and Witold Pedrycz. Key points estimation and point instance segmentation approach for lane detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):8949-8958, 2022. +[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020. +[21] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of applied mathematics, 2(2):164-168, 1944. +[22] Juntao Li, Chuong Nguyen, and Shaodi You. Temporal 3d fully connected network for water-hazard detection. In 2019 Digital Image Computing: Techniques and Applications (DICTA), pages 1-5, 2019. +[23] Yadi Li, Liguo Chen, Haibo Huang, Xiangpeng Li, Wenkui Xu, Liang Zheng, and Jiaqi Huang. Nighttime lane markings recognition based on canny detection and hough transform. In IEEE International Conference on Real-time Computing and Robotics, pages 411-415. IEEE, 2016. +[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. +[25] Zuo-Quan Li, Hui-Min Ma, and Zheng-Yu Liu. Road lane detection with gabor filters. In International Conference on Information System and Artificial Intelligence, pages 436-440. IEEE, 2016. +[26] Jiawei Liu, Jing Zhang, Yicong Hong, and Nick Barnes. Learning structure-aware semantic segmentation with image-level supervision. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. +[27] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on + +conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021. +[28] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. +[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004. +[30] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019. +[31] Yunze Man, Xinshuo Weng, Xi Li, and Kris Kitani. Ground-net: Monocular ground plane normal estimation with geometric consistency. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2170-2178, 2019. +[32] Donald W Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the society for Industrial and Applied Mathematics, 11(2):431-441, 1963. +[33] Annika Meyer, N. Ole Salscheider, Piotr F. Orzechowski, and Christoph Stiller. Deep semantic lane segmentation for mapless driving. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 869-875, 2018. +[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. +[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 286-291, 2018. +[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. volume 32, 2018. +[37] Jonah Philion. Fastdraw: Addressing the long tail of lane detection by adapting a sequential prediction network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11582-11591, 2019. +[38] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In European Conference on Computer Vision, pages 276-291. Springer, 2020. +[39] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011. +[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning robust camera localization from pixels to pose. In CVPR, 2021. +[41] Iago Suárez, Ghesn Sfeir, José M Buenasada, and Luis Baumela. Beblid: Boosted efficient binary local image descriptor. Pattern recognition letters, 133:366-372, 2020. + +[42] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. +[44] Tsung-Ying Sun, Shang-Jeng Tsai, and Vincent Chan. Hsi color model based lane-marking detection. In IEEE Intelligent Transportation Systems Conference, pages 1168-1172. IEEE, 2006. +[45] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021. +[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021. +[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. +[48] Luo-Wei Tsai, Jun-Wei Hsieh, Chi-Hung Chuang, and Kuo-Chin Fan. Lane detection using directional random walks. In IEEE Intelligent Vehicles Symposium, pages 303–306. IEEE, 2008. +[49] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[51] Jun Wang, Tao Mei, Bin Kong, and Hu Wei. An approach of lane detection based on inverse perspective mapping. In International IEEE Conference on Intelligent Transportation Systems, pages 35-38. IEEE, 2014. +[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. +[53] Ruochen Yin, Yong Cheng, Huapeng Wu, Yuntao Song, Biao Yu, and Runxin Niu. Fusionlane: Multi-sensor fusion for lane marking semantic segmentation using deep neural networks. IEEE Transactions on Intelligent Transportation Systems, 2020. +[54] Seungwoo Yoo, Hee Seok Lee, Heesoo Myeong, Sungrack Yun, Youngwoo Park, Janghoon Cho, and Duck Hoon + +Kim. End-to-end lane marker detection via row-wise classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1006-1007, 2020. +[55] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. arXiv preprint arXiv:1904.00962, 2019. +[56] Yujun Zhang, Lei Zhu, Wei Feng, Huazhu Fu, Mingqian Wang, Qingxia Li, Cheng Li, and Song Wang. Vil-100: A new dataset and a baseline model for video instance lane detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15681–15690, 2021. +[57] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. \ No newline at end of file diff --git a/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/images.zip b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80c1d9e6832aaf845dcc416b40ec5bab3f70dbf2 --- /dev/null +++ b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:baa7458a49ac946957cd6a5a2d1694db8e15724c69c38764e14f35bf33bcc44f +size 517650 diff --git a/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/layout.json b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..96c095ddb7366118112ecdc10bbade653a7312b1 --- /dev/null +++ b/2023/Homography Guided Temporal Fusion for Road Line and Marking Segmentation/layout.json @@ -0,0 +1,9815 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 102, + 504, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 102, + 504, + 139 + ], + "spans": [ + { + "bbox": [ + 90, + 102, + 504, + 139 + ], + "type": "text", + "content": "Homography Guided Temporal Fusion for Road Line and Marking Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "spans": [ + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": "Shan Wang" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Chuong Nguyen" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Jiawei Liu" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Kaihao Zhang" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Wenhan Luo" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Yanhao Zhang" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Sundaram Muthu" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Fahira Afzal Maken" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": " Hongdong Li" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": "Data61, CSIRO " + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": "Australian National University " + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 93, + 160, + 499, + 203 + ], + "type": "text", + "content": "Sun Yat-sen University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 290, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 591 + ], + "type": "text", + "content": "Reliable segmentation of road lines and markings is critical to autonomous driving. Our work is motivated by the observations that road lines and markings are (1) frequently occluded in the presence of moving vehicles, shadow, and glare and (2) highly structured with low intra-class shape variance and overall high appearance consistency. To solve these issues, we propose a Homography Guided Fusion (HomoFusion) module to exploit temporally-adjacent video frames for complementary cues facilitating the correct classification of the partially occluded road lines or markings. To reduce computational complexity, a novel surface normal estimator is proposed to establish spatial correspondences between the sampled frames, allowing the HomoFusion module to perform a pixel-to-pixel attention mechanism in updating the representation of the occluded road lines or markings. Experiments on ApolloScape, a largescale lane mark segmentation dataset, and ApolloScape Night with artificial simulated night-time road conditions, demonstrate that our method outperforms other existing SOTA lane mark segmentation models with less than " + }, + { + "bbox": [ + 46, + 256, + 290, + 591 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 46, + 256, + 290, + 591 + ], + "type": "text", + "content": " of their parameters and computational complexity. We show that exploiting available camera intrinsic data and ground plane assumption for cross-frame correspondence can lead to a light-weight network with significantly improved performances in speed and accuracy. We also prove the versatility of our HomoFusion approach by applying it to the problem of water puddle segmentation and achieving SOTA performance1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "type": "text", + "content": "Lane mark segmentation aims to achieve pixel-wise classification of road lines and markings simultaneously. Known for its quintessential importance in autonomous driving, lane mark segmentation is also an effective tool for constructing accurate High Definition (HD) maps. Existing works solve individual sub-tasks, e.g. drivable area segmen" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 231, + 545, + 380 + ], + "blocks": [ + { + "bbox": [ + 309, + 231, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 309, + 231, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 309, + 231, + 545, + 380 + ], + "type": "image", + "image_path": "ebb967392b1de54130f4ddb8f633a65435ddcef2c0f073ca19f88caf61981358.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 388, + 547, + 568 + ], + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 568 + ], + "type": "text", + "content": "Figure 1: Illustration of the effect of the proposed HomoFusion module that explores the adjacent frames for cues, facilitating the correct classification of (1) a \"Straight Arrow\", which, with its bottom half occluded by a vehicle, is mistakenly classified as a \"Right Turn & Straight Arrow\" without the HomoFusion module, and (2) a partially occluded \"Dotted Line\", which is incorrectly classified as a \"Solid Line\" without the HomoFusion module. The fused frame in the " + }, + { + "bbox": [ + 304, + 388, + 547, + 568 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 304, + 388, + 547, + 568 + ], + "type": "text", + "content": " row and the " + }, + { + "bbox": [ + 304, + 388, + 547, + 568 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 304, + 388, + 547, + 568 + ], + "type": "text", + "content": " column demonstrates the recovered road lines and markings after projecting the previous frames onto the current frame with the estimated homography matrices. The yellow box enlarges the area where mistake classifications are corrected. The red box indicates the spatially corresponding area across the frames. Best viewed in color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 581, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 605 + ], + "type": "text", + "content": "tation [33, 9] and lane detection [19, 35], while few address the lane mark segmentation in its entirety [53, 16]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": "Despite the tremendous progresses in semantic/scene segmentation [52, 42, 26], little attention has been paid to the lane mark segmentation task [53, 16]. Yin et al. [53] leverages additional LiDAR information, merging the segmented visual information with the point clouds, to achieve lane mark segmentation. IntRA-KD [16] applies the knowledge distillation technique to improve the efficiency of the lane mark segmentation model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "A major challenge in the lane mark segmentation task is" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 286, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 286, + 713 + ], + "type": "text", + "content": "Code is available at https://github.com/ShanWang-Shan/HomoFusion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1075" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "partial occlusion of the road lines and markings caused by the vehicle and surrounding environment leading to false classifications. For example, a partially occluded straight arrow can be easily mistaken as a right turn & straight arrow, as depicted in Fig. 1. However, existing lane mark segmentation methods have not yet made use of the following observations: (1) occlusion of the road lines and markings (due to nearby moving vehicles, shadow, and glare) can be reduced by cross-frame consistency; (2) road lines and markings are highly structured, maintaining high consistency in intra-class shapes as well as in overall appearance, alleviating the requirements on learning complicated features to distinguish different classes with high intra-class variance as in semantic/scene segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 384 + ], + "type": "text", + "content": "To find complementary information from adjacent frames, we use the ground plane assumption for the road region immediately in front of the camera. Our Homography Guided Fusion (HomoFusion) module achieves temporally consistent representation, recovers partially occluded road lines or markings, and leads to correct classification. The module's success depends on accurate cross-frame spatial correspondence, which we achieve using a homography transformation matrix estimated with available intrinsic and extrinsic parameters of the on-vehicle camera, and the road surface normal estimated by our novel optimization method called Road Surface Normal Estimator (RSNE)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 539 + ], + "type": "text", + "content": "Using the highly structured nature of road lines and markings could reduce the complexity of the detection problem and make it suitable to run on edge devices. To this end, we employ (1) a lightweight encoder to represent the visual information in feature spaces, and (2) a cross-frame pixel-to-pixel attention mechanism in our HomoFusion module, instead of a more computationally expensive global attention mechanism employed by other methods. These design decisions allow the proposed model to outperform the existing lane mark segmentation models with less than " + }, + { + "bbox": [ + 46, + 384, + 289, + 539 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 46, + 384, + 289, + 539 + ], + "type": "text", + "content": " of their parameters and Giga Floating Point Operations (GFLOPs) We summarize our contributions as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 544, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 544, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 544, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 58, + 544, + 287, + 594 + ], + "type": "text", + "content": "- We propose a HomoFusion module that uses ground plane assumption and the adjacent frames for temporally consistent representations for accurate classification of partially occluded road lines and markings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 599, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 599, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 599, + 287, + 684 + ], + "type": "text", + "content": "- We design a novel estimator RSNE for road surface normal, which, combined with camera intrinsic and extrinsic parameters readily available on an autonomous vehicle, yields accurate homography matrix between frame pairs. RSNE simplifies the 8 degrees of freedom (DoF) of homography problem to a 2 DoF of normal vector problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We present a lightweight lane mark segmentation model that achieves better performance than the state" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 96 + ], + "type": "text", + "content": "of-the-art (SOTA) methods with significantly reduced model complexity and computation requirements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 107, + 392, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 107, + 392, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 107, + 392, + 120 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 127, + 400, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 127, + 400, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 127, + 400, + 139 + ], + "type": "text", + "content": "2.1. Lane Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 145, + 547, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 547, + 385 + ], + "type": "text", + "content": "Traditional lane detection approaches rely on handcrafted features such as color [44, 51], edge [23] and texture [25], which are limited in complex scenarios [48, 7]. Recent advances in deep neural networks have led to significant performance improvements, with methods like message-passing networks [36] and attention-based modules [45]. [16, 37, 38, 54, 27] formulate lane detection as a row-wise classification task based on grid division of the input image. PolyLaneNet [46] is the first parametric prediction method, which outputs polynomials to represent each lane. BEV-Former [24] and PETRv2 [28] employ transformer mechanisms to streamline the conversion of perspective views into bird's-eye views, thereby eliminating variations in object size caused by perspective. Although the above methods achieve impressive performance, all of them focus on detecting the lane lines and ignore the basic road elements like arrow signal in Fig. 1. Furthermore, lane detection differs from segmentation tasks in that it focuses on identifying the boundaries of the drivable area, rather than the real shape of the lanes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 393, + 509, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 509, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 509, + 406 + ], + "type": "text", + "content": "2.2. Road Line and Marking Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 591 + ], + "type": "text", + "content": "The identification of road elements is crucial for ensuring safety in autonomous driving systems, but there are few works addressing this problem. Hou et al. [16] introduced a distillation approach that demonstrates competitive performance in lane mark segmentation. Yin et al. [53] used an LSTM-based network to segment images with DeepLabv3+ [8] and then merged them with point clouds to assist with lane segmentation. Unfortunately, these methods do not address the challenges of occlusion, shadows, and glare that frequently occur in real-world driving scenarios. Recently, Zhang et al. [56] introduced global memory information from previous frames to enhance local information for video lane segmentation. However, this method introduces attention to the entire image, making it memory-intensive and inefficient for autonomous driving scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 599, + 444, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 444, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 444, + 612 + ], + "type": "text", + "content": "2.3. Homography Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Homography estimation methods can be divided into non-deep and deep learning-based approaches. Non-deep methods estimate the homography using feature extraction, feature matching, and outlier rejection. SIFT [29], SURF [5, 4], ORB [39], LPPM [30], GMS [6], and BEBLID [41] are commonly used for feature extraction, and RANSAC [12], MAGAC [2], and LRLS [15] are applied for outlier rejection. Recently, many deep learning-based approaches" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 317, + 758 + ], + "type": "text", + "content": "1076" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 68, + 492, + 252 + ], + "blocks": [ + { + "bbox": [ + 97, + 68, + 492, + 252 + ], + "lines": [ + { + "bbox": [ + 97, + 68, + 492, + 252 + ], + "spans": [ + { + "bbox": [ + 97, + 68, + 492, + 252 + ], + "type": "image", + "image_path": "043b4b782b333726fa0cdd0c646fe96ea82b98dda31efae2f4d8de990b5b8b36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "text", + "content": "Figure 2: Overview of our proposed model consisting of a pair of lightweight encoder and decoder, our proposed HomoFusion module, and our proposed Road Surface Normal Estimator (RSNE). A sequence of frames " + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "text", + "content": ", including a target frame " + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{t}}" + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "text", + "content": " previous frames, are encoded into the feature representations " + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}^l)" + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "text", + "content": ". RNSE estimates the road surface normal vector, which, combined with the camera intrinsic and extrinsic parameters, yields a homography matrix between each frame pair, establishing cross-frame spatial correspondences. HomoFusion uses pixel-to-pixel attention mechanism to obtain temporally consistent representation for on-road pixels of the current frame with the spatial correspondence across frames as guidance. Finally, the decoder decodes and upsamples the temporally consistent feature representations to produce the lane mark segmentation prediction " + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}_{\\mathbf{t}}^{\\mathrm{prd}})" + }, + { + "bbox": [ + 46, + 261, + 547, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 363, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 289, + 460 + ], + "type": "text", + "content": "have been proposed, such as DeTone et al. [10]'s VGG-like architecture, Nowruzi et al. [11], Le et al. [20]'s and Man et al. [31]'s cascaded VGG-like networks, and Chang et al. and Zhao et al.'s incorporation of the Lucas-Kanade (LK) algorithm with deep networks. However, these methods do not utilize known camera intrinsic and extrinsic parameters across frames for homography matrix estimation, which can effectively reduce the search dimensions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 472, + 104, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 104, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 104, + 483 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 115, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 115, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 115, + 504 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": "As illustrated in Fig. 2, our proposed framework takes a sequence of frames, denoted as " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{\\mathbf{I}_i|i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": ", as input. This sequence consists of a current frame " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": " previous frames sampled at a fixed time interval " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": ". The output of the framework is a lane mark segmentation map " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": " for the current frame. In addition, the framework includes a novel HomoFusion module, which uses the homography transformation guided by the proposed RSNE to fuse the feature map of the current frame with those of the previous frames. We denote the encoded feature representation for the " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": " frame as " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t = \\{\\mathbf{F}_t^l\\in \\mathbb{R}^{H_l\\times W_l\\times C_l}\\}_{l = 1}^L" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 510, + 287, + 656 + ], + "type": "text", + "content": " indicates the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": "level of the feature map. The shallowest and deepest levels are represented by 1 and " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": " represent the height, width, and channel number of feature maps in level " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": ". We denote all encoded features as " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{\\mathbf{F}_i^l | i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": ". The proposed HomoFusion module, detailed in Sec. 3.2, uses a cross-frame pixel-to-pixel attention mechanism to fuse the feature map " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_f = \\{\\mathbf{F}_f^l\\}_{l = 1}^L)" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": " of the current frame with those of the previous frames. The homography transformation matrix between the target frame and each of the previous frames is calculated using the estimated normal vector of the road surface " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": " obtained by the proposed RSNE, as described in Sec. 3.3. Finally, the decoder produces the segmentation prediction " + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}_t)" + }, + { + "bbox": [ + 304, + 363, + 545, + 544 + ], + "type": "text", + "content": " for the target frame based on the fused feature map." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 551, + 533, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 551, + 533, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 551, + 533, + 564 + ], + "type": "text", + "content": "3.2. Homography Guided Fusion (HomoFusion)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Our proposed HomoFusion module employs a pixel-to-pixel attention mechanism to fuse spatially corresponding pixels across frames. This mechanism is achieved by projecting the pixels of the current frame onto the previous frames through an accurate homography transformation. Since road lines and markings are strictly on-road, the search area can be limited to the road surface, which is mostly a plane, at least within the immediate front of the vehicle where markings are readable. Therefore, only the homography transformation is needed to accurately map road pixels between frames. According to the standard inverse homography [14, 57, 49], for each on-road pixel " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_t =" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "text", + "content": "2In our experiments, both camera intrinsic and extrinsic parameters are provided by the dataset. In real-world scenarios, these parameters can be obtained through camera calibration and autonomous vehicle pose obtained from the pose estimation framework, which is already known in the in-car system and free of cost." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1077" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 135 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 135 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 135 + ], + "type": "image", + "image_path": "2a75d7d795a37273ec68a561ec4cc0ba9232f02ea232f8f195b482d5ce704e33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 143, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 143, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 287, + 216 + ], + "type": "text", + "content": "Figure 3: Illustration of sample points. (Right) Sample points in the current frame. (Left/Middle) Correspondence of a sample point in previous frames. The red point coordinate is calculated by using the correct normal, while the cyan point coordinate is calculated by using the initial (incorrect) normal." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 223, + 288, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 288, + 261 + ], + "type": "inline_equation", + "content": "(u,v)^{\\top}" + }, + { + "bbox": [ + 47, + 223, + 288, + 261 + ], + "type": "text", + "content": " of the current frame, its spatially corresponding pixels of the reference frames " + }, + { + "bbox": [ + 47, + 223, + 288, + 261 + ], + "type": "inline_equation", + "content": "\\{p_i|i = \\{t - (s - 1)\\Delta t\\}_{s = 1}^n\\}" + }, + { + "bbox": [ + 47, + 223, + 288, + 261 + ], + "type": "text", + "content": " can be computed using Eq. (1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 269, + 287, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 269, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 93, + 269, + 287, + 294 + ], + "type": "interline_equation", + "content": "p _ {i} \\propto \\mathbf {K} \\left(\\mathbf {R} _ {i} - \\frac {\\mathbf {t} _ {i} \\mathbf {n} ^ {\\top}}{d}\\right) \\mathbf {K} ^ {- 1} \\left(p _ {t} \\oplus 1\\right), \\tag {1}", + "image_path": "10beac51847fb86b40db5a790086ea4ddbefd6939cf96715717d5e1e27062631.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\propto" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " represents proportional, " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " is the intrinsic matrix of the on-vehicle camera, " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " represent the relative rotation and translation between the current frame " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " previous frame " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " is the normal of the observed road surface, which is estimated in Sec. 3.3. " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " is the vertical distance between the on-vehicle camera and the road surface obtained from the camera calibration. " + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\oplus 1" + }, + { + "bbox": [ + 46, + 294, + 287, + 461 + ], + "type": "text", + "content": " converts the point coordinates to homogeneous coordinates. The on-road pixels of the current frame are sampled from a triangular area in front of the car, as illustrated in Fig. 3. Figure 3 also demonstrates that the shape and category of a left-turn can be recovered by exploring the spatially corresponding area in the previous frames under the guidance of an accurate homography transformation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 617 + ], + "type": "text", + "content": "With the spatial correspondence across frames, the proposed HomoFusion module computes a temporally consistent representation for each sampled pixel of the current frame by employing temporal attention mechanism on the feature presentations of spatially corresponding pixels. The mechanism assigns the pixel representation of current frame " + }, + { + "bbox": [ + 46, + 462, + 288, + 617 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_t[p_t])" + }, + { + "bbox": [ + 46, + 462, + 288, + 617 + ], + "type": "text", + "content": " as a query and all frames " + }, + { + "bbox": [ + 46, + 462, + 288, + 617 + ], + "type": "inline_equation", + "content": "(\\{\\mathbf{F}_i[p_i]\\})" + }, + { + "bbox": [ + 46, + 462, + 288, + 617 + ], + "type": "text", + "content": " as keys and values. In contrast to the existing attention mechanism [50], we propose to (1) omit the spatial encoding in the presence of obtained spatial correspondence across frames, and (2) apply an L2 normalization on the query and the keys to obtain more robust (deep) feature similarities (Eq. 2) across various lighting environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 627, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 627, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 107, + 627, + 287, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {a} _ {i} = \\frac {\\mathbf {F} _ {t} [ p _ {t} ]}{\\| \\mathbf {F} _ {t} [ p _ {t} ] \\| _ {2}} \\cdot \\frac {\\mathbf {F} _ {i} [ p _ {i} ]}{\\| \\mathbf {F} _ {i} [ p _ {i} ] \\| _ {2}}, \\tag {2}", + "image_path": "188b469cb08524b5138dc4e19f4836c2231276e014198be7c6caf70aba774f91.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 658, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\| .\\| _2" + }, + { + "bbox": [ + 47, + 658, + 287, + 685 + ], + "type": "text", + "content": " indicates an " + }, + { + "bbox": [ + 47, + 658, + 287, + 685 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 658, + 287, + 685 + ], + "type": "text", + "content": " normalization. The similarity is normalized by softmax function as given in Eq. 3:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 692, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 692, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 127, + 692, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {i} = \\frac {\\exp \\mathbf {a} _ {i}}{\\sum_ {i} \\exp \\mathbf {a} _ {i}}, \\tag {3}", + "image_path": "fb8df54a90f776920e336ba1ce240771c4f77f120a3e0589b798e40e8b045988.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "before being used as weight to fuse the spatially corresponding pixels across the frames as shown in Eq. 4:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 105, + 545, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 105, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 356, + 105, + 545, + 129 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {f} [ p _ {t} ] = \\mathbf {F} _ {t} [ p _ {t} ] + \\sum_ {i} \\mathbf {W} _ {i} \\mathbf {F} _ {i} [ p _ {i} ]. \\tag {4}", + "image_path": "7f7eb490421168f335ff069e875a65b8d6624162b86a62ad07dcb9b2896718d4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 137, + 520, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 137, + 520, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 137, + 520, + 150 + ], + "type": "text", + "content": "3.3. Road Surface Normal Estimator (RSNE)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "text", + "content": "Accurate estimation of road surface normal is crucial for establishing spatial correspondence across frames. Generally, on-vehicle cameras have a fixed angle with respect to the ground level. However, the actual road surface normal can vary due to the unevenness of the road surface, e.g. uphill or downhill roads, and sloped turns, etc. Assuming that the initial road surface normal is a vector being perpendicular to a horizontal plane and pointing upwards, we propose to obtain an accurate road surface normal with an optimization process that repeats for each new coming frame. The proposed RSNE iteratively updates the estimated road surface normal with the projection error of on-road pixels from the current frame onto the previous frames. Specifically, we sample " + }, + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "text", + "content": " pixels " + }, + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\{p_t^j\\}_{j=1}^m" + }, + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "text", + "content": " from the current frame that are likely to be on-road, as illustrated in Fig. 3, and compute the spatially corresponding pixels " + }, + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\{p_i^j\\}_{j=1}^m" + }, + { + "bbox": [ + 304, + 156, + 545, + 419 + ], + "type": "text", + "content": " in the previous frames using the homography matrix computed with the estimated road surface normal as defined in Eq. 1. The estimated road surface normal is optimized in an iterative manner through the following procedure: (a) computing a residual between the low-level features of spatially corresponding pixels across the frames with Eq. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 427, + 545, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 427, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 367, + 427, + 545, + 442 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {i} ^ {j} = \\mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ] - \\mathbf {F} _ {t} ^ {1} [ p _ {t} ^ {j} ] \\in \\mathbb {R} ^ {c}, \\tag {5}", + "image_path": "1bc8948425610e9a3cdc1dd1734a19cab26efaa6002fa72dd9c2f6681450fe7a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 449, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 486 + ], + "type": "text", + "content": "where the residual is set to 0 for the sample point whose projections in previous frames fall outside the frame area, and computing the overall error with Eq. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 386, + 494, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 494, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 386, + 494, + 545, + 520 + ], + "type": "interline_equation", + "content": "\\mathbf {E} = \\sum_ {i, j} \\rho \\left(\\left\\| \\mathbf {r} _ {i} ^ {j} \\right\\| _ {2} ^ {2}\\right), \\tag {6}", + "image_path": "4f77ec7039aa5189637300887966618c7cc0623e122c4258761a2bb8dea869d6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 523, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 523, + 545, + 595 + ], + "type": "inline_equation", + "content": "\\| .\\| _2^2" + }, + { + "bbox": [ + 304, + 523, + 545, + 595 + ], + "type": "text", + "content": " shows squared norm and " + }, + { + "bbox": [ + 304, + 523, + 545, + 595 + ], + "type": "inline_equation", + "content": "\\rho (\\cdot)" + }, + { + "bbox": [ + 304, + 523, + 545, + 595 + ], + "type": "text", + "content": " is a robust cost function proposed by Barron [3]. This nonlinear least-square cost is iteratively minimized towards correct normal estimation using the Levenberg-Marquardt algorithm [21, 32]; (b) computing a new road surface normal value by adding the update given by Eq. 12." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 595, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 632 + ], + "type": "text", + "content": "The road surface normal vector is a unit vector with two degrees of freedom (DoF), which can be decomposed into a pitch angle " + }, + { + "bbox": [ + 304, + 595, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 595, + 545, + 632 + ], + "type": "text", + "content": " and a roll angle " + }, + { + "bbox": [ + 304, + 595, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\phi^3" + }, + { + "bbox": [ + 304, + 595, + 545, + 632 + ], + "type": "text", + "content": ", as shown in Eq. 7." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 335, + 639, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 639, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 335, + 639, + 545, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {n} = \\left(- \\sin \\phi \\cos \\theta , - \\cos \\phi \\cos \\theta , \\sin \\theta\\right) ^ {\\top}. \\tag {7}", + "image_path": "355bc8a9a2774035ac31ce1b5457e57d8e783f4d23f06547f8461ef11fe158db.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "text", + "content": "The Jacobian of the residual function w.r.t. the pitch angle and the roll angle is defined in Eq. 8 by applying the chain" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "3An explanation of the pitch angle " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " and roll angle " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " is provided in the appendix." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1078" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 69, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 69, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 69, + 83 + ], + "type": "text", + "content": "rule," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 103, + 81, + 287, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 81, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 103, + 81, + 287, + 112 + ], + "type": "interline_equation", + "content": "\\mathbf {J} _ {i} ^ {j} = \\frac {\\partial \\mathbf {r} _ {i} ^ {j}}{\\partial \\delta} = \\frac {\\partial \\mathbf {F} _ {i} ^ {1} [ p _ {i} ^ {j} ]}{\\partial p _ {i} ^ {j}} \\frac {\\partial p _ {i} ^ {j}}{\\partial \\mathbf {n}} \\frac {\\partial \\mathbf {n}}{\\partial \\delta}, \\tag {8}", + "image_path": "7eb29d3e9e583e99130ea0b990091924e6d8367feda3c4a2c65549efdc6c36ca.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "content": " represents the to-be-optimized target (" + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\mathbf{F}_i^1[p_i^j]}{\\partial p_i^j}" + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "content": " is gradient from 2D interpolation. " + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "inline_equation", + "content": "\\frac{\\partial p_i^j}{\\partial \\mathbf{n}}" + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "content": " is the Jacobian of the homography transformation (Eq. (1)) w.r.t. " + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 47, + 115, + 287, + 170 + ], + "type": "text", + "content": ", as defined in Eq. 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 178, + 287, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 178, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 99, + 178, + 287, + 204 + ], + "type": "interline_equation", + "content": "\\frac {\\partial p _ {i} ^ {j}}{\\partial \\mathbf {n}} = - \\frac {1}{d} \\mathbf {K t} _ {i} \\left(\\mathbf {K} ^ {- 1} \\left(p _ {t} ^ {j} \\oplus 1\\right)\\right) ^ {\\top}, \\tag {9}", + "image_path": "f74822d9704b9dbb5875e063f1455f06339e1459398c766eec3d3e034f0725c1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathbf{n}}{\\partial\\delta}" + }, + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "text", + "content": ", defined in Eq. 10 and Eq. 11, is the Jacobian of Eq. (7) w.r.t. " + }, + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 213, + 287, + 238 + ], + "type": "text", + "content": " respectively" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 245, + 287, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 245, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 245, + 287, + 274 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathbf {n}}{\\partial \\theta} = \\left(- \\frac {\\mathbf {n} _ {1} \\mathbf {n} _ {3}}{\\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}}, - \\frac {\\mathbf {n} _ {2} \\mathbf {n} _ {3}}{\\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}}, \\sqrt {1 - \\mathbf {n} _ {3} ^ {2}}\\right), \\tag {10}", + "image_path": "b7119dd76ebf67f8025334aea17a83247e921b62b52e70a8f6ff87c2a85b0b5f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 281, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 281, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 121, + 281, + 287, + 306 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathbf {n}}{\\partial \\phi} = \\left(\\mathbf {n} _ {2}, - \\mathbf {n} _ {1}, 0\\right), \\tag {11}", + "image_path": "145a9b0499ab86f784544b16e7878552d03ce7430c22223a0d4322ecb6372ead.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_k" + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "content": ". The detailed derivation is included in appendix (Sec. 1). We obtain Hessian matrices as " + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\mathbb{H} = \\mathbf{J}^{\\top}\\rho^{\\prime}\\mathbf{J}" + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\rho^\\prime" + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "content": " is the derivative of function " + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "inline_equation", + "content": "\\rho (\\cdot)" + }, + { + "bbox": [ + 47, + 310, + 287, + 370 + ], + "type": "text", + "content": ". We compute the update by damping the Hessian and solving the linear system, as shown in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 378, + 287, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 378, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 94, + 378, + 287, + 393 + ], + "type": "interline_equation", + "content": "\\triangle \\delta = - (\\mathbb {H} + \\lambda \\operatorname {d i a g} (\\mathbb {H})) ^ {- 1} \\mathbf {J} ^ {\\top} \\rho^ {\\prime} \\mathbf {r}, \\tag {12}", + "image_path": "c2f9ad4eeb30da841d61ff5280f0571dd25323f9294cd27b673d665eb2ee0735.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": " is the damping factor [40], balancing between the Gauss-Newton " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "(\\lambda = 0)" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": " and gradient descent " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "(\\lambda = \\infty)" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": ". Our entire optimization process is differentiable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "text", + "content": "The algorithm is illustrated in Alg. 1. The normal vector " + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "text", + "content": " uses low-level features " + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^1" + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "text", + "content": ". While the high-level features " + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^2" + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "text", + "content": " focus on segmentation feature expression, the low-level features " + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^1" + }, + { + "bbox": [ + 47, + 437, + 287, + 510 + ], + "type": "text", + "content": " are supervised to extract useful features for both segmentation and road surface normal estimation. As both tasks focus on on-road objects, they can benefit each other." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 128, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 540, + 180, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 180, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 180, + 552 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "text", + "content": "We use a pre-trained EfficientNet [47] to extract image features at two different scales, " + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "text", + "content": " down-scaling, with 64 and 128 channels, respectively. To focus on the bottom region of the input image, where road marks are typically visible, we crop only the bottom " + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "text", + "content": " of the input image, following the same approach in the prior work [16]. The processed image size is set to " + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "272 \\times 848" + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "text", + "content": ". Our decoder consists of bi-linear up-sampling layers and convolution layers, which up-sample the high-level features and increase the resolution by a factor of 4 at each level. We trained our model using the AdamW optimizer [55] for 30,000 iterations on two NVIDIA RTX 3090 GPUs, with a learning rate of " + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-3}" + }, + { + "bbox": [ + 47, + 558, + 288, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "type": "code", + "bbox": [ + 306, + 86, + 545, + 413 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 449, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 449, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 449, + 84 + ], + "type": "text", + "content": "Algorithm 1 Normal Optimization" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": "Input: \nThe 1-st Level Feature Maps: " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^1 = \\{\\mathbf{F}_i^1\\}" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " \nSample Points Coordinates: " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "p_t = \\{p_t^j\\}_{j = 1}^m" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " \nInitial Pitch and Roll Value: " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "(\\theta_0,\\phi_0) = (0.15,0)" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " \nDamping Factor: " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\lambda = (\\lambda_{\\theta}\\lambda_{\\phi})" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " \nHyper-parameter: \nk; " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " Maximum loop count, empirically set to 20 \n" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " ; " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " Convergence threshold, empirically set to 0.0001 \nOutput: Optimized Pitch and Roll " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " . \n1: function OPT( " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\theta_0,\\phi_0,\\mathbf{F}^1,p_t,\\lambda" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " \n2: Derive point features " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^1 [p_t]" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^1" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " . \n3: for " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "k\\gets 1" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " to k do \n4: Calculate point coordinates " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^1" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " (Eq. 1); \n5: Derive point features " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^1[p_i]" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^1" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " . \n6: Calculate residual " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " (Eq. 5); \n7: Calculate observe error E (Eq. 6); \n8: Calculate robust cost " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\rho (\\mathbf{E})" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " and its derivation " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\rho^\\prime" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " \n9: Construct J and its Hessian matrices " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{H}" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " (Eq. 8); \n10: Obtain " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\triangle \\delta" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " by Cholesky decomposition (Eq. 12); \n11: Update Normal as " + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)_k\\gets (\\theta ,\\phi)_{k - 1} + \\triangle \\delta" + }, + { + "bbox": [ + 306, + 86, + 545, + 413 + ], + "type": "text", + "content": " \n12: if MAX(△δ)< α then \n13: Break; \n14: end if \n15: end for \n16: end function" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_body" + } + ], + "index": 15, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 435, + 368, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 368, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 368, + 445 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": "We conduct our experiments on two datasets, ApolloScape [17] and ApolloScape Night. Other datasets were deemed unsuitable due to the absence of camera extrinsics, as observed in SDLane [18] and VLI-100 [56], inadequate frame overlap, or the lack of diverse segmentation labels, exemplified by Waymo Open Dataset [43]. Further elaboration on these reasons can be found in the appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "text", + "content": "ApolloScape. The ApolloScape dataset is a large-scale dataset that can be used for localization and segmentation tasks. It consists of 38 distinct classes and poses various challenges including occlusions and tiny road markings. To provide accurate camera poses, the vehicle is equipped with customer-grade GPS/IMU. However, for the lane mark segmentation task, only 41, 201 annotated images out of more than " + }, + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "inline_equation", + "content": "110,000+" + }, + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "text", + "content": " are associated with camera pose information. We used these 41, 201 images for our experiments since our approach relies on camera poses. We divided them into 35, 173 training images and 6, 028 validation images. Because our approach uses sequential information, we select the validation set with completely different trajectories to ensure a fair comparison." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "ApolloScape Night. Since there is a lack of datasets for" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1079" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 110, + 547, + 184 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "type": "text", + "content": "Table 1: Comparison with SOTA Methods on the ApolloScape [17] and Apolloscape Night Datasets and Running on an NVIDIA RTX 3090 GPU. \"18 mIoU\" Represents the Mean IoU of 18 Types of Lane Markers Selected by ApolloScape Official Metrics. \"36 mIoU\" Represents the Mean IoU of All Unignorable 36 Types." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 110, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 47, + 110, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 547, + 184 + ], + "type": "table", + "html": "
MethodsFrame CountBackboneParams (M)↓GFLOPs↓FPS (f/s)↑ApolloScape [17]ApolloScape Night
18 mIoU↑36 mIoU↑18 mIoU↑36 mIoU↑
IntRA-KD[16]1ResNet-10165.65159.410.842.124.629.816.7
SegFormer[52]1MiT-B113.51048.843.852.332.138.323.1
CFFM [42]4MiT-B115.31192.622.753.232.739.123.6
MMA-Net [56]4ResNet-5057.9723.220.652.931.438.823.2
HomoFusion(ours)4EfficientNet-B61.2461.225.459.335.944.826.6
", + "image_path": "0b3f7423ba7b6bd15d299ce3d9e592bf807345ab40298c7505e5823920175334.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 186, + 286, + 345 + ], + "blocks": [ + { + "bbox": [ + 50, + 186, + 286, + 345 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 286, + 345 + ], + "type": "image", + "image_path": "5bc9f79a72ab999a35f2820e2b4cc188e81285273cf6ba6e6044218e683cb632.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 353, + 288, + 391 + ], + "lines": [ + { + "bbox": [ + 46, + 353, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 288, + 391 + ], + "type": "text", + "content": "Figure 4: Sample images from the ApolloScape Night dataset. Top: original daytime images from the ApolloScape dataset. Bottom: synthesized night-time images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 468 + ], + "type": "text", + "content": "night-time lane mark segmentations, we created an artificial dataset called ApolloScape Night from the ApolloScape dataset using a cross-domain generation network [1]. This allows us to evaluate our proposed model on a challenging dataset with poor lighting conditions, road reflection, and glare. Fig. 4 displays some sample images from the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 475, + 160, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 160, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 160, + 488 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 288, + 639 + ], + "type": "text", + "content": "In accordance with the guidelines of the ApolloScape benchmark [17], we used mean intersection-over-union (mIoU) as the evaluation criterion. ApolloScape contains 38 different types of lane markers, including two ignorable labels (noise and ignored). Of these labels, 18 categories are used in the official evaluation metrics. To provide a comprehensive evaluation of our approach, we report the evaluated mIoU results for both the selected 18 categories, denoted as \"18 mIoU\", and all 36 categories, denoted as \"36 mIoU\". By reporting both sets of results, we provide a more complete picture of the performance of our method on the ApolloScape and ApolloScape Night datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 270, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 270, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 270, + 659 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "To evaluate the performance of our proposed method, we compared it with SOTA algorithms, including (a) IntRA-KD [16], (b) SegFormer [52], (c) CFFM [42], and (d) MMA-Net [56], on the ApolloScape [17] and ApolloScape" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 189, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 545, + 224 + ], + "type": "text", + "content": "Night datasets. To ensure a fair comparison, we retrained each model on the same subset of the training set with the same input resolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "type": "text", + "content": "The results of the comparison are presented in Table 1. Our approach outperforms the existing SOTA methods on both the ApolloScape [17] and ApolloScape Night datasets while having less than " + }, + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 304, + 226, + 545, + 285 + ], + "type": "text", + "content": " of their parameters or computational overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "text", + "content": "In addition, we compared our proposed method with CFFM [42] which also uses adjacent frames to enhance the representation of the current frame. Our method is more efficient in terms of model complexity, with a complexity of " + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(HWC)" + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "text", + "content": " for feature extraction and " + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(HWC^2)" + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "text", + "content": " for cross-frame feature fusion. In contrast, the complexity of CFFM is " + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(H^2 W^2 C) + \\mathcal{O}(HWC^2)" + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "text", + "content": " for feature extraction and " + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(HWEC) + \\mathcal{O}(HWC^2)" + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "text", + "content": " for cross-frame feature fusion, where " + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 287, + 546, + 442 + ], + "type": "text", + "content": " is calculated by their receptive field and pooling kernel size. This explains why our method has a lower computational requirement for the cross-frame pixel-to-pixel attention mechanism compared to the global attention mechanism used by CFFM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 443, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 528 + ], + "type": "text", + "content": "The qualitative results of both the existing SOTA models and our proposed model are shown in Fig. 5. The results indicate that our approach achieves superior segmentation performance on road lines and markings, even under adverse conditions such as occlusion, road reflection, and poor lighting. Additional visualization examples for various categories can be found in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 537, + 460, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 460, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 460, + 551 + ], + "type": "text", + "content": "4.5. Study on Hyper-parameters" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": "Number of Frames " + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": ". We conducted a study on the impact of the number of frames while setting the sampling frame gap " + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\triangle t = 2" + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": ". The results presented in Fig. 6 (a)-(b) demonstrate that the performance of our method improves on both ApolloScape and ApolloScape Night datasets with an increase in the number of explored adjacent frames. However, the performance improvement becomes negligible when the number of frames exceeds 4. Considering the monotonically increasing relationship between the model complexity and the number of frames, we set " + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": " to strike a balance between model performance and complexity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Sampling Frame Gap " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\triangle t)" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": ". In the study on the effect of the sampling frame gap, we set the number of frames" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1080" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 545, + 312 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 312 + ], + "type": "image", + "image_path": "046f66b36e313b4e0fc1979894796229282eace5ffa43923391516698af4f7c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 547, + 355 + ], + "type": "text", + "content": "Figure 5: Qualitative comparison with SOTA methods. The top two examples are from the ApolloScape [17] dataset, and the bottom two examples are from the ApolloScape Night dataset. Yellow boxes highlight the area of interest for better visualization. Red boxes indicate false-positive segmentation predictions. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 372, + 165, + 434 + ], + "blocks": [ + { + "bbox": [ + 63, + 372, + 165, + 434 + ], + "lines": [ + { + "bbox": [ + 63, + 372, + 165, + 434 + ], + "spans": [ + { + "bbox": [ + 63, + 372, + 165, + 434 + ], + "type": "image", + "image_path": "103c965265be01852973f356baec3bb1419bbb9021e1fe98c4410ea4afe1a6ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 438, + 160, + 449 + ], + "lines": [ + { + "bbox": [ + 67, + 438, + 160, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 438, + 160, + 449 + ], + "type": "text", + "content": "(a) Number of frames (AS))" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 372, + 272, + 434 + ], + "blocks": [ + { + "bbox": [ + 171, + 372, + 272, + 434 + ], + "lines": [ + { + "bbox": [ + 171, + 372, + 272, + 434 + ], + "spans": [ + { + "bbox": [ + 171, + 372, + 272, + 434 + ], + "type": "image", + "image_path": "a63926b1780a60b9eb07e97b60b996f7c5dc010d54b4f73ce472af1b67e06b53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 438, + 269, + 449 + ], + "lines": [ + { + "bbox": [ + 174, + 438, + 269, + 449 + ], + "spans": [ + { + "bbox": [ + 174, + 438, + 269, + 449 + ], + "type": "text", + "content": "(b) Number of frames (ASN)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 460, + 164, + 522 + ], + "blocks": [ + { + "bbox": [ + 62, + 460, + 164, + 522 + ], + "lines": [ + { + "bbox": [ + 62, + 460, + 164, + 522 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 164, + 522 + ], + "type": "image", + "image_path": "06eb612b9cb19730fb07a488a6076dff637c62626bf4ff98cb2ef760d9140b31.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 527, + 160, + 538 + ], + "lines": [ + { + "bbox": [ + 65, + 527, + 160, + 538 + ], + "spans": [ + { + "bbox": [ + 65, + 527, + 160, + 538 + ], + "type": "text", + "content": "(c) Sampling frame gap (AS)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 172, + 461, + 272, + 521 + ], + "blocks": [ + { + "bbox": [ + 172, + 461, + 272, + 521 + ], + "lines": [ + { + "bbox": [ + 172, + 461, + 272, + 521 + ], + "spans": [ + { + "bbox": [ + 172, + 461, + 272, + 521 + ], + "type": "image", + "image_path": "46770a1b70147801b354797c55666d9aaf5cdf1b069122278de9ab7012510501.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 527, + 272, + 538 + ], + "lines": [ + { + "bbox": [ + 171, + 527, + 272, + 538 + ], + "spans": [ + { + "bbox": [ + 171, + 527, + 272, + 538 + ], + "type": "text", + "content": "(d) Sampling frame gap (ASN)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "type": "text", + "content": "Figure 6: The impact of the number of frames " + }, + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "type": "text", + "content": " on ApolloScape (AS) and ApolloScape Night (ASN) datasets is shown in (a) and (b), while (c) and (d) show the effect of the sampling frame gap " + }, + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\triangle t" + }, + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "text", + "content": ". Fig. 6 (c)-(d) indicates that the optimal performance is achieved at " + }, + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\triangle t = 2" + }, + { + "bbox": [ + 46, + 601, + 287, + 661 + ], + "type": "text", + "content": ". This can be attributed to the balance between the sufficient overlap of road area across the frames and a wide enough temporal gap that allows the displacement of the source of occlusion, e.g. moving vehicle." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "text", + "content": "4.6. Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "All ablation studies are conducted on the ApolloScape dataset with the same training strategies described in" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 328, + 390, + 524, + 460 + ], + "blocks": [ + { + "bbox": [ + 306, + 361, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 384 + ], + "type": "text", + "content": "Table 2: Ablation Study of our Proposed HomoFusion and RSNE on the ApolloScape Dataset" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 390, + 524, + 460 + ], + "lines": [ + { + "bbox": [ + 328, + 390, + 524, + 460 + ], + "spans": [ + { + "bbox": [ + 328, + 390, + 524, + 460 + ], + "type": "table", + "html": "
18 mIoU↑36 mIoU↑
w/o HomoFusion49.431.0
HomoFusion w/o HomoGuide51.831.3
HomoFusion w/o RSNE57.435.3
HomoFusion w/ RGB RSNE54.933.8
Noisy Extrinsic57.535.1
HomoFusion Full59.335.9
", + "image_path": "4836c4b5744e0c461d1404acda0a4283e95cef7eaa4c0a020764334fbfa92012.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "text", + "content": "Sec. 4.1. We set the number of frames to " + }, + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "text", + "content": " and a sampling frame gap of " + }, + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\triangle t = 2" + }, + { + "bbox": [ + 305, + 469, + 545, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "text", + "content": "Impact of Homofusion. The core component of our approach is the Homofusion module, which allows the use of complementary information from sequential frames. Tab. 2 demonstrates that our model with the Homofusion module (4 frames) achieves better performance than the variant without Homofusion module (1 frame), illustrating the effectiveness of the temporally consistent feature representation obtained with Homofusion module in addressing the partial occlusion issues. We also study the effect directly associating the pixels across frames based on their coordinates by using identity matrix as the homography matrix, denoted as \"w/o HomoGuide\". It shows that such an incorrect cross-frame spatial correspondence, although leveraging additional frames, provides limited help in improving the model performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Impact of RSNE. We study the impact of the proposed RSNE in the following model variants each comprising the HomoFusion modules: (a) without RSNE (\"HomoFusion" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1081" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": "w/o RSNE\"), where the homography matrix is estimated with the initial road surface normal " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "(\\mathbf{n}_0)" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", (b) with RSNE using RGB information (\"HomoFusion w/ RGB RSNE\") and (c) with RSNE with features (\"HomoFusion Full\"). Results in Tab. 2 show that \"HomoFusion Full\" achieves better performance than \"HomoFusion w/o RSNE\", demonstrating the effectiveness of the iterative optimization in estimating a more accurate homography matrix that subsequently provides more accurate guidance on cross-frame spatial correspondences. It can be also observed that \"HomoFusion Full\" significantly outperforms \"HomoFusion w/ RGB RSNE\" indicating that deep features are more robust than RGB information in estimating an accurate road surface normal. This can be attributed to the fact the feature representations are less prone to adverse noises caused by environmental factors, e.g. brightness. Fig. 7 shows that the fused frame with road surface normal estimated with features has better alignment in road lines and markings than that with RGB information." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 308, + 170, + 366 + ], + "blocks": [ + { + "bbox": [ + 51, + 308, + 170, + 366 + ], + "lines": [ + { + "bbox": [ + 51, + 308, + 170, + 366 + ], + "spans": [ + { + "bbox": [ + 51, + 308, + 170, + 366 + ], + "type": "image", + "image_path": "0e6f905dac6f480fd4bfb46d1b768cf0e903e46531c93fda4c6cd168dc297fb1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 369, + 165, + 378 + ], + "lines": [ + { + "bbox": [ + 58, + 369, + 165, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 369, + 165, + 378 + ], + "type": "text", + "content": "(a) Normal estimation with RGB" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 175, + 308, + 294, + 366 + ], + "blocks": [ + { + "bbox": [ + 175, + 308, + 294, + 366 + ], + "lines": [ + { + "bbox": [ + 175, + 308, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 175, + 308, + 294, + 366 + ], + "type": "image", + "image_path": "400830723900dd0292a3d335e18fd6abd7cc11ae9670381bc54e9b6aa19b9042.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 369, + 291, + 378 + ], + "lines": [ + { + "bbox": [ + 179, + 369, + 291, + 378 + ], + "spans": [ + { + "bbox": [ + 179, + 369, + 291, + 378 + ], + "type": "text", + "content": "(b) Normal estimation with feature" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 389, + 287, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 445 + ], + "type": "text", + "content": "Figure 7: Fused frame sequence with the homography transformation between the frame pairs. The road surface normal estimated with feature representation enables more accurate homography transformation than with RGB information, resulting in better alignment of road lines and markings across frames." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 455, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 527 + ], + "type": "text", + "content": "Robustness against Extrinsic Noise. We investigate the robustness of our method to extrinsic noise by introducing random translation errors of less than 1 meter and rotation errors of up to " + }, + { + "bbox": [ + 46, + 455, + 287, + 527 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 46, + 455, + 287, + 527 + ], + "type": "text", + "content": ". The results are shown in \"Noisy Extrinsic\" of Tab. 2, which demonstrate that our method is capable of mitigating extrinsic noise and remains robust." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "type": "text", + "content": "Impact of Backbone. In order to ensure a fair comparison with CFFM [42] and MMA-Net [56], we implement our method with 'MiT-B1' and 'ResNet-50' backbone and report its performance in Tab. 3. Our method still outperforms CFFM [42] and MMA-Net [56] significantly. Additionally, we test our method with 'EfficientNet-B4/5' backbones, and the results show that our method has relatively stable performance even when using smaller backbones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 624, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 684 + ], + "type": "text", + "content": "Application to Another Task. Given our method's ability to accurately align road surface objects across adjacent frames, we also try it with the task of detecting water hazards. Specifically, our approach aims to perform binary segmentation to identify water puddles in the road. The previ" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 100, + 553, + 161 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 95 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 95 + ], + "type": "text", + "content": "Table 3: Comparison of our Proposed HomoFusion with Different Backbones on ApolloScape Dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 100, + 553, + 161 + ], + "lines": [ + { + "bbox": [ + 306, + 100, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 306, + 100, + 553, + 161 + ], + "type": "table", + "html": "
BackbonesParams(M)↓GFLOPs↓18 mIoU↑36 mIoU↑
MiT-B113.6489.857.235.3
ResNet-509.12189.755.334.6
EfficientNet-B40.860.857.035.3
EfficientNet-B51.0261.257.935.2
EfficientNet-B61.2461.259.335.9
", + "image_path": "13bc54305710c96d2d87225d83d6b3e2efa0c118e4865492b78ad7208585dcea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 167, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 547, + 335 + ], + "type": "text", + "content": "ous work on this task T3D-FCN [22] also exploits temporal information, as water puddles can appear differently at different angles and distances as a vehicle moves. We use the Puddle-1000 dataset [13], which includes On-road and Offroad datasets. The frames obtained from an onboard camera on bumpy roads are not stable and contain small extrinsic noise. The intrinsic information of the camera is provided by the dataset, while the global extrinsic information of the camera is obtained using ORB-SLAM [34]. Our experiments are conducted using the current and seven previously consecutive frames with a " + }, + { + "bbox": [ + 304, + 167, + 547, + 335 + ], + "type": "inline_equation", + "content": "240 \\times 320" + }, + { + "bbox": [ + 304, + 167, + 547, + 335 + ], + "type": "text", + "content": " image size, the same settings as T3D-FCN [22]. The results, presented in Tab. 4, demonstrate the high effectiveness of our method. More visual details of this task can be found in the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 314, + 378, + 537, + 449 + ], + "blocks": [ + { + "bbox": [ + 305, + 346, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 369 + ], + "type": "text", + "content": "Table 4: Comparison of water puddle segmentations on Puddle-1000 dataset" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 378, + 537, + 449 + ], + "lines": [ + { + "bbox": [ + 314, + 378, + 537, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 378, + 537, + 449 + ], + "type": "table", + "html": "
DatasetMethodsF1-meas↑Prec↑Rec↑
On-roadFCN-8s-FL-RAU [13]0.700.680.72
T3D-FCN [22]0.680.790.62
HomoFusion (ours)0.800.810.78
Off-roadFCN-8s-FL-RAU [13]0.810.870.77
T3D-FCNU [22]0.730.870.63
HomoFusion (ours)0.870.870.87
", + "image_path": "78ebd398965f0fae4e66e75373e0b6ab8add981479f1916cdc40c8b4fda09788.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 378, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "The proposed lightweight lane mark segmentation model presented in this paper offers superior performance with reduced model complexity compared to SOTA approaches. The integration of the HomoFusion module and Road Surface Normal Estimator provides an accurate classification of partially occluded, shadowed, and/or glare-affected road lines and markings by leveraging adjacent frames and pixel-to-pixel attention mechanisms. Moreover, the novel approach of using the road surface normal to guide spatial correspondences across frames has significant implications for a wide range of applications in autonomous driving, including road surface inspection, water, and other hazard detections. However, the current model relies on available camera extrinsics and is somewhat impacted by extrinsic noise. Future work will focus on optimizing both the normal vector and camera extrinsic parameters to improve the performance and make the method more robust to challenging environmental conditions. Overall, the proposed method has the potential to significantly advance the field of autonomous driving and related applications." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "4We use a heavy backbone in the competing methods as their publicly available code uses this backbone" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1082" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 115 + ], + "type": "text", + "content": "The research is funded in part by an ARC Discovery Grant (grant ID: DP220100800) to HL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 125, + 107, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 125, + 107, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 125, + 107, + 138 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 145, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 145, + 287, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 145, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 53, + 145, + 287, + 211 + ], + "type": "text", + "content": "[1] Vinicius F Arruda, Thiago M Paixão, Rodrigo F Berriel, Alberto F De Souza, Claudine Badue, Nicu Sebe, and Thiago Oliveira-Santos. Cross-domain car detection using unsupervised image-to-image translation: From day to night. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 212, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 212, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 212, + 287, + 255 + ], + "type": "text", + "content": "[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10197-10205, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 256, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 287, + 290 + ], + "type": "text", + "content": "[3] Jonathan T. Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 290, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 290, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 290, + 287, + 323 + ], + "type": "text", + "content": "[4] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 323, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 287, + 357 + ], + "type": "text", + "content": "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European conference on computer vision, pages 404-417. Springer, 2006." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 422 + ], + "type": "text", + "content": "[6] JiaWang Bian, Wen-Yan Lin, Yasuyuki Matsushita, Sai-Kit Yeung, Tan-Dat Nguyen, and Ming-Ming Cheng. Gms: Grid-based motion statistics for fast, ultra-robust feature correspondence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4181–4190, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 423, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 287, + 467 + ], + "type": "text", + "content": "[7] Amol Borkar, Monson Hayes, and Mark T Smith. A novel lane detection system with efficient ground truth generation. IEEE Transactions on Intelligent Transportation Systems, 13(1):365-374, 2011." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 468, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 523 + ], + "type": "text", + "content": "[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 524, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 287, + 556 + ], + "type": "text", + "content": "[9] Zhe Chen, Jing Zhang, and Dacheng Tao. Progressive lidar adaptation for road detection. IEEE/CAA Journal of Automatica Sinica, 6(3):693-702, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 557, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 287, + 590 + ], + "type": "text", + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 590, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 646 + ], + "type": "text", + "content": "[11] Farzan Erlik Nowruzi, Robert Laganiere, and Nathalie Japkowicz. Homography estimation from image pairs with hierarchical convolutional networks. In Proceedings of the IEEE international conference on computer vision workshops, pages 913-920, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 690 + ], + "type": "text", + "content": "[12] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 287, + 715 + ], + "type": "text", + "content": "[13] Xiaofeng Han, Chuong Nguyen, Shaodi You, and Jianfeng Lu. Single image water hazard detection using fcn with re" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "flection attention units. In Proceedings of the European Conference on Computer Vision (ECCV), pages 105-120, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "text", + "content": "[14] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[15] Paul W Holland and Roy E Welsch. Robust regression using iteratively reweighted least-squares. Communications in Statistics-theory and Methods, 6(9):813-827, 1977." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 162, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 216 + ], + "type": "text", + "content": "[16] Yuenan Hou, Zheng Ma, Chunxiao Liu, Tak-Wai Hui, and Chen Change Loy. Inter-region affinity distillation for road marking segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12486-12495, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 217, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 272 + ], + "type": "text", + "content": "[17] Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954-960, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "text", + "content": "[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In CVPR, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 360 + ], + "type": "text", + "content": "[19] Yeongmin Ko, Younkwan Lee, Shoaib Azam, Farzeen Munir, Moongu Jeon, and Witold Pedrycz. Key points estimation and point instance segmentation approach for lane detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):8949-8958, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 405 + ], + "type": "text", + "content": "[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "text", + "content": "[21] Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of applied mathematics, 2(2):164-168, 1944." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "type": "text", + "content": "[22] Juntao Li, Chuong Nguyen, and Shaodi You. Temporal 3d fully connected network for water-hazard detection. In 2019 Digital Image Computing: Techniques and Applications (DICTA), pages 1-5, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 536 + ], + "type": "text", + "content": "[23] Yadi Li, Liguo Chen, Haibo Huang, Xiangpeng Li, Wenkui Xu, Liang Zheng, and Jiaqi Huang. Nighttime lane markings recognition based on canny detection and hough transform. In IEEE International Conference on Real-time Computing and Robotics, pages 411-415. IEEE, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 537, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 592 + ], + "type": "text", + "content": "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "text", + "content": "[25] Zuo-Quan Li, Hui-Min Ma, and Zheng-Yu Liu. Road lane detection with gabor filters. In International Conference on Information System and Artificial Intelligence, pages 436-440. IEEE, 2016." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[26] Jiawei Liu, Jing Zhang, Yicong Hong, and Nick Barnes. Learning structure-aware semantic segmentation with image-level supervision. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[27] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1083" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[28] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "text", + "content": "[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 288, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 288, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 288, + 221 + ], + "type": "text", + "content": "[30] Jiayi Ma, Ji Zhao, Junjun Jiang, Huabing Zhou, and Xiaojie Guo. Locality preserving matching. International Journal of Computer Vision, 127(5):512-531, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "text", + "content": "[31] Yunze Man, Xinshuo Weng, Xi Li, and Kris Kitani. Ground-net: Monocular ground plane normal estimation with geometric consistency. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2170-2178, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 269, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 302 + ], + "type": "text", + "content": "[32] Donald W Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the society for Industrial and Applied Mathematics, 11(2):431-441, 1963." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 357 + ], + "type": "text", + "content": "[33] Annika Meyer, N. Ole Salscheider, Piotr F. Orzechowski, and Christoph Stiller. Deep semantic lane segmentation for mapless driving. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 869-875, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 406, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 460 + ], + "type": "text", + "content": "[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In 2018 IEEE Intelligent Vehicles Symposium (IV), pages 286-291, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "text", + "content": "[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. volume 32, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "type": "text", + "content": "[37] Jonah Philion. Fastdraw: Addressing the long tail of lane detection by adapting a sequential prediction network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11582-11591, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "type": "text", + "content": "[38] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In European Conference on Computer Vision, pages 276-291. Springer, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[39] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International conference on computer vision, pages 2564-2571. Ieee, 2011." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[40] Paul-Edouard Sarlin, Ajaykumar Unagar, Måns Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning robust camera localization from pixels to pose. In CVPR, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[41] Iago Suárez, Ghesn Sfeir, José M Buenasada, and Luis Baumela. Beblid: Boosted efficient binary local image descriptor. Pattern recognition letters, 133:366-372, 2020." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[42] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 194 + ], + "type": "text", + "content": "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 195, + 546, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 546, + 237 + ], + "type": "text", + "content": "[44] Tsung-Ying Sun, Shang-Jeng Tsai, and Vincent Chan. Hsi color model based lane-marking detection. In IEEE Intelligent Transportation Systems Conference, pages 1168-1172. IEEE, 2006." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 239, + 546, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 546, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 546, + 303 + ], + "type": "text", + "content": "[45] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 360 + ], + "type": "text", + "content": "[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "type": "text", + "content": "[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 546, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 446 + ], + "type": "text", + "content": "[48] Luo-Wei Tsai, Jun-Wei Hsieh, Chi-Hung Chuang, and Kuo-Chin Fan. Lane detection using directional random walks. In IEEE Intelligent Vehicles Symposium, pages 303–306. IEEE, 2008." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "type": "text", + "content": "[49] Richard Tucker and Noah Snively. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 537, + 546, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 581 + ], + "type": "text", + "content": "[51] Jun Wang, Tao Mei, Bin Kong, and Hu Wei. An approach of lane detection based on inverse perspective mapping. In International IEEE Conference on Intelligent Transportation Systems, pages 35-38. IEEE, 2014." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 582, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 546, + 635 + ], + "type": "text", + "content": "[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 546, + 690 + ], + "type": "text", + "content": "[53] Ruochen Yin, Yong Cheng, Huapeng Wu, Yuntao Song, Biao Yu, and Runxin Niu. Fusionlane: Multi-sensor fusion for lane marking semantic segmentation using deep neural networks. IEEE Transactions on Intelligent Transportation Systems, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "type": "text", + "content": "[54] Seungwoo Yoo, Hee Seok Lee, Heesoo Myeong, Sungrack Yun, Youngwoo Park, Janghoon Cho, and Duck Hoon" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1084" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 272 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Kim. End-to-end lane marker detection via row-wise classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1006-1007, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[55] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. arXiv preprint arXiv:1904.00962, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 287, + 228 + ], + "type": "text", + "content": "[56] Yujun Zhang, Lei Zhu, Wei Feng, Huazhu Fu, Mingqian Wang, Qingxia Li, Cheng Li, and Song Wang. Vil-100: A new dataset and a baseline model for video instance lane detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15681–15690, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 229, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 229, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 229, + 287, + 272 + ], + "type": "text", + "content": "[57] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1085" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_content_list.json b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4684cfabbbacf3a616dbebf46fea3fabcd383c00 --- /dev/null +++ b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_content_list.json @@ -0,0 +1,1641 @@ +[ + { + "type": "text", + "text": "HopFIR: Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Zhai $^{1*}$ Qiang Nie $^{2*}$ Bo Ouyang $^{1\\dagger}$ Xiang Li $^{3}$ Shanlin Yang $^{1}$ $^{1}$ Hefei University of Technology $^{2}$ Youtu Lab, Tencent $^{3}$ Tsinghua University", + "bbox": [ + 169, + 202, + 800, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 287, + 313, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2D-to-3D human pose lifting is fundamental for 3D human pose estimation (HPE), for which graph convolutional networks (GCNs) have proven inherently suitable for modeling the human skeletal topology. However, the current GCN-based 3D HPE methods update the node features by aggregating their neighbors' information without considering the interaction of joints in different joint synergies. Although some studies have proposed importing limb information to learn the movement patterns, the latent synergies among joints, such as maintaining balance are seldom investigated. We propose the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture to tackle the 3D HPE problem. HopFIR mainly consists of a novel hop-wise GraphFormer (HGF) module and an intragroup joint refinement (IJR) module. The HGF module groups the joints by $k$ -hop neighbors and applies a hopwise transformer-like attention mechanism to these groups to discover latent joint synergies. The IJR module leverages the prior limb information for peripheral joint refinement. Extensive experimental results show that HopFIR outperforms the SOTA methods by a large margin, with a mean per-joint position error (MPJPE) on the Human3.6M dataset of $32.67\\mathrm{mm}$ . We also demonstrate that the state-of-the-art GCN-based methods can benefit from the proposed hop-wise attention mechanism with a significant improvement in performance: SemGCN [42] and MGCN [49] are improved by $8.9\\%$ and $4.5\\%$ , respectively.", + "bbox": [ + 75, + 321, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 760, + 209, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D human pose estimation aims to accurately regress the 3D locations of human joints in the camera coordinate system from a single image. It plays an important role in many applications, such as action recognition and human-computer interaction. Compared with the monocu", + "bbox": [ + 75, + 786, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5bc2f34248cc90ef9114cd7610308005b82d2dfd435c4e252cc1d04f0ffa1cd2.jpg", + "image_caption": [ + "Figure 1. Comparison of performance and model size between the proposed HopFIR and SOTA methods, namely Modulated GCN (MGCN) [49], SemGCN [42], Weight Unsharing [18], SR-Net [39], GraphSH [36], and GraFormer [43]. The methods are evaluated on the Human3.6M dataset [10] with ground truth 2D joints as input. The arrow shows the performance improvement obtained by inserting the HA layer into the other networks. Tiny, small, and normal denote the feature dimension of the HopFIR are 32, 64, and 128, respectively." + ], + "image_footnote": [], + "bbox": [ + 516, + 282, + 864, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lar systems, multi-view capture systems are expensive and inconvenient to set up and operate, which prevents them from being widely used in practice. To tackle the monocular 3D HPE task, some approaches [2, 15, 31, 32, 38] estimates 3D joint coordinates or heat maps directly from an image via a convolutional neural network (CNN) [13, 14]. However, direct regression from the image space suffers from the problem of a large parameter searching space, which always leads to a sub-optimal solution. Recently, Martinez et al. [20] constructed a simple fully connected network using only 2D keypoints as input and achieved promising 3D HPE performance, showing that the 3D human pose can be efficiently and accurately estimated from 2D joint po", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution †Corresponding Author", + "bbox": [ + 107, + 875, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14985", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sitions. Inspired by them and considering the thoroughly investigated 2D HPE, many works decompose the problem into two subtasks, i.e., the 2D HPE and 2D-to-3D pose lifting [18, 40, 49, 36]. 2D-to-3D pose lifting, therefore, has emerged as a fundamental task in this area that our work devotes to.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key consideration in 2D-to-3D pose lifting is that the human skeleton topology is inherently sparse and graph-structured. Fully connected neural networks are less effective in modeling graph-structured data due to their simple connections among all nodes and the probability of overfitting. To leverage the information of the human skeletal topology, some works [36, 40, 49] have proposed to model the human body with GCNs and have achieved SOTA results. For example, Ci et al. [4] introduced a locally connected network to enhance the representation capability of the GCN, and Liu et al. [18] explored the weight sharing and feature transformation that occurs before or after feature aggregation in the GCN. One limitation of these GCN-based 3D HPE methods, however, is that they update the node features by aggregating their neighbors' information without considering the different contributions of these nodes to different joint synergies.", + "bbox": [ + 75, + 184, + 470, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of considering all the joints of a skeleton as a whole, Xue et al. [37] demonstrated that the human skeleton exhibits obvious part-wise inconsistency in its motion patterns, as also reported in SRNet [39]. However, these works are limited to considering prior structural information of limb groupings and ignore investigating the latent groups underlying joint synergies. For example, the relative positions of the 1-hop neighbors of joint 0 are almost constant in \"Discussion\" subject, which can be a latent group, as shown in Fig. 2. Moreover, these works consider the joints in a limb as a whole to calculate the relationship with other limbs, which resulted in lower accuracy for peripheral joints, such as wrists and feet.", + "bbox": [ + 75, + 444, + 468, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the abovementioned problems in monocular 3D HPE, we propose a novel architecture: the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR). The first key component of HopFIR is a novel hop-wise GraphFormer (HGF) module that considers $k$ -hop neighbors. In the HGF module, the information of every hop of every joint is aggregated into the hidden space, such that $N \\times k$ groups of features are obtained for a skeleton model with $N$ joints. Meanwhile, a hop-wise transformer-like attention mechanism is designed to extract the correlation among feature groups, which computes similarity by the dot product of the node feature and the group feature. The proposed HGF module enables the network to discover latent joint interactions considering human joint synergy. Because the HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interactions of peripheral joints associated with", + "bbox": [ + 75, + 643, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a limb, we introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information. Specifically, a residual block is built from two HGF modules followed by one IJR module. The proposed HopFIR architecture achieves optimal regression accuracy with a stack of three blocks.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our work makes the following contributions:", + "bbox": [ + 498, + 181, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we design the first Hopwise GraphFormer module to explore potential joint correlations underlying human joint synergy. We also prove that other GCN-based methods can benefit from the proposed HGF module efficiently, as shown in Fig. 1.", + "- We design an Intragroup Joint Refinement module, which attends to intragroup joints to refine joint features through the associated limb, especially the wrists and feet. The IJR module enables HGF modules to discover the latent synergies among joints.", + "- We propose the novel Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture for 3D HPE, which is built entirely from HGF and IJR modules. Specifically, two HGF modules and one IJR module are coupled into a block.", + "- Extensive experiments demonstrate the effectiveness and generalizability of the proposed modules and HopFIR architecture by providing new state-of-the-art results on two challenging datasets, i.e., Human3.6M [10] and MPI-INF-3DHP [21]." + ], + "bbox": [ + 496, + 212, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 527, + 640, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Human Pose Estimation. Early works [26, 29] use handcrafted features, perspective relationships, and geometric constraints to estimate the 3D human pose. Recent pose estimation approaches can be generally divided into two categories. The first category of networks regresses 3D human joints directly from the image [24, 46]. Pavlakos et al. [24] adopted a CNN to predict the voxel-wise likelihoods for each joint, and Zhou et al. [46] directly embedded a kinematic object model into the networks to learn the general multi-articulate object pose. Approaches in the second category decouple the 3D HPE task into 2D pose estimation from an image and 3D pose estimation from the detected 2D joints (2D-to-3D). For example, Martinez et al. [20] proposed a simple yet effective baseline with fully-connect networks and proved that 3D human poses can be regressed simply and effectively from 2D keypoints. Our paper follows this pipeline and focuses on the 2D-to-3D pose lifting. For promoting 3D human pose regression accuracy, it is crucial to group joints with consideration of their interactions rather than treating all joints of a skeleton as a whole. Xue et al. [37] divided the human skeleton graph into five groups according to the limbs to explore part-wise motion inconsistency, and Zeng et al. [39] split the human joints into", + "bbox": [ + 496, + 553, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14986", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "local regions and recombined the global information from the rest of the joints. Our proposed HopFIR differs from these approaches by grouping joints by the $k$ -hop neighbors of each joint and prior limb information, which enables the network to discover latent connections between groups in different human joint synergies.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Graph Convolutional Networks. GCNs [5, 7, 12, 27] generalize the capability of CNNs by performing convolution operations on graph-structured data. GCNs can be divided into two categories: the spectral-based approaches [5] and the spatial-based approaches [12]. Our approach falls into the second category, which applies message-passing operations on the graph nodes and their neighbors.", + "bbox": [ + 75, + 184, + 467, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the graph-structure topology of the human skeleton, many works [4, 18] have introduced GCN to tackle the 3D HPE task. Zhao et al. [42] proposed a SemGCN to learn the semantic relationships between human joints, and Zou et al. [49] proposed a weight modulation and an affinity modulation based on the SemGCN. These methods aggregate the first-order neighborhood messages to update the feature matrix by assigning different weights to different nodes. Some works [47, 48] have extended the first-order neighbors to high-order neighbors in the spatial domain directly. Zeng et al. [40] designed a hierarchical fusion block by dividing the fusion procedure into two stages, where all the high-order neighbors of a node are aggregated into a feature in the first stage and fuse it with the node feature and the first-order neighbor in the second stage. Zhao et al. [43] introduced Chebyshev graph convolution to fuse information among the $k$ -hop neighbors of a joint directly. These works updated features by aggregating each node's own $k$ -hop neighborhood information in a GCN layer. However, HopFIR considers the $k$ -hop groups of all nodes to reconstruct the $k$ -hop feature of a node through the proposed attention mechanism, which can enhance the representation capability of GCNs.", + "bbox": [ + 75, + 292, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Graph Attention. Graph attention networks [34] is a pioneer work that pay attention to the data in a graph structure by assigning an attention weight to each node. The introduction of the transformer [33] for machine translation tasks has proven the capacity of attention for sequential input. ViT [6] introduces the transformer into computer vision and achieves excellent performance. Inspired by them, some researchers have adopted the transformer for 3D HPE. Zhao et al. [43] applied self-attention to capture global information by calculating the similarity of all nodes. PoseFormer [44] directly applies Transformer Encoder (TE) by viewing each joint and each frame as tokens in the spatial and temporal domains, respectively. MixSTE [41] is based on [44], where each joint feature is represented by the temporal TE to model the joint motion. P-stmo [28] replaces the spatial TE with MLP and applies Stride Transformer Encoder to map $N$ frames to one frame. MHFormer [17] generates", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cb670ec8dfb645aeddec880737bf4316cadc122c63e7044bff95e1e5ac76c551.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 524, + 89, + 627, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b3d9bfec6aa4c9b4191bc1f4e7059caf8f7eece3beb678b2e70b8085e238d37.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 642, + 90, + 764, + 196 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c8971dbdb6be3b194413e8c566fedf1c02c2dc63ffd51ebaa72088d8bd92efe.jpg", + "image_caption": [ + "(c)", + "Figure 2. An illustration of the human skeleton graph and the groups in HGF and IJR modules. The $k$ -hop neighbors of a joint are set as a group in HGF modules where (a) indicates the 1-hop groups of different joints and (b) indicates several latent groups because of physical limitations. (c) indicates the joints grouped by prior limb information." + ], + "image_footnote": [], + "bbox": [ + 764, + 90, + 867, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "multi-hypothesis at different depths of stacked spatial TEs by viewing all frames of each joint feature as a token, then communicates these hypotheses with cross-attention and self-attention. Different from the existing graph attention for 3D HPE, we propose the intergroup multi-head attention mechanism among the $k$ -hop groups of all nodes, which assigns the attention weights by computing the similarity between the node feature and $k$ -hop group feature. Moreover, we introduce the intragroup multi-head self-attention in limb groups to refine the joint features and promote the HGF module to discover the latent synergies among joints.", + "bbox": [ + 496, + 338, + 890, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed HopFIR", + "text_level": 1, + "bbox": [ + 500, + 517, + 714, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper proposes a novel architecture to regress the 3D human pose from $N$ given 2D keypoints $X \\in \\mathbb{R}^{N \\times 2}$ . The proposed framework mainly consists of the HGF and IJR modules. In this section, we first review the vanilla Graph Convolutional Network and Transformer in Sec. 3.1. We then introduce the HGF and IJR modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we present the network architecture in Sec. 3.4.", + "bbox": [ + 496, + 542, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vanilla GCN and Transformer", + "text_level": 1, + "bbox": [ + 500, + 672, + 769, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GCN. A graph is defined as $\\mathcal{G} = (\\mathcal{V},\\mathcal{E})$ , where $\\mathcal{V}$ is a set of $N$ nodes and $\\mathcal{E}$ is the adjacency matrix representing the edges between the nodes. Given a collection of input features $H^{l}\\in \\mathbb{R}^{N\\times D}$ , a generic GCN layer that aggregates neighborhood information can be formulated as follows:", + "bbox": [ + 496, + 696, + 890, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {(l + 1)} = \\sigma (\\tilde {A} H ^ {l} W) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 780, + 890, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W\\in \\mathbb{R}^{D\\times D^{\\prime}}$ is the learnable weight matrix that transforms the feature dimension from $D$ to $D^{\\prime}$ , and $\\sigma (.)$ is the activation function, such as ReLU [22]. $H^{(l + 1)}$ is the updated feature matrix, $\\tilde{A}\\in \\mathbb{R}^{N\\times N}$ is the symmetrically normalized affinity matrix [12] with added selfconnections, and $A\\in \\{0,1\\}^{N\\times N}$ is the adjacency ma", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14987", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b73bf7f1727b044f6557a45cfa24667f94059486ae1ba38047561c596cfe3203.jpg", + "image_caption": [ + "Figure 3. The HopFIR architecture, with details of the HGF and IJR modules. The designed block is a residual block [9] built by two HGF modules followed by one IJR module. The proposed architecture with three blocks achieves optimal performance. Arrows of different colors represent different hops and groups in HGF and IJR, respectively." + ], + "image_footnote": [], + "bbox": [ + 169, + 90, + 803, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trix. The $(i,j)th$ entry $a_{ij} = 1$ representing node $j$ is the neighbor of node $i$ . Otherwise, they are not connected and $a_{ij} = 0$ . Therefore, the none-neighbor nodes have a weak influence on each other in the vanilla GCN, which hinders the modeling of underlying joint synergy in 3D HPE.", + "bbox": [ + 75, + 335, + 468, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer. Transformer architecture relies entirely on self-attention to compute representations of its input and output. The self-attention function maps the inputs to the queries $Q$ , keys $K$ , and values $V$ by weight matrices $W_{Q}$ , $W_{K}$ , and $W_{V}$ , respectively, and the matrix of outputs is calculated as:", + "bbox": [ + 75, + 412, + 470, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(Q K ^ {T} / \\sqrt {d}\\right) V \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 515, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ is the feature dimension of $Q$ , and $\\frac{1}{\\sqrt{d}}$ is a scaling factor to prevent extremely small gradients. The multihead self-attention (MHSA), which performs self-attention in parallel, projects the queries, keys, and values $P$ times with different linear projections to the respective subspaces, as follows:", + "bbox": [ + 75, + 546, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(D _ {1}, \\dots , D _ {P}\\right) W _ {o} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 652, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_O$ is the projection matrix of outputs, $D_p = \\text{Attention}(QW_Q^p, KW_K^p, VW_V^p)$ , and $p \\in [1,..,P]$ .", + "bbox": [ + 76, + 683, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hop-wise GraphFormer", + "text_level": 1, + "bbox": [ + 76, + 724, + 299, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous GCN studies for 3D HPE aggregate multi-hop neighborhood information [40, 48] or assign an attention weight to each first-order neighbor [42, 49]. To effectively capture the node's neighborhood message and increase the representational capacity of GCNs, we introduce the hopwise GraphFormer (HGF) module, which treats each hop as a group (Fig.2 a,b) and computes the attention weights for each hop (more intuitive descriptions for $k$ -hop can be found in the supplementary material). By considering the relationship within $k$ hops, we can obtain $N \\times k$ groups for a", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "skeleton graph with $N$ joints, which provides enough combinations of joints to discover the latent correlations among joints in different human joint synergy.", + "bbox": [ + 496, + 335, + 890, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first define the $k$ -hop matrix $A^k$ as", + "bbox": [ + 517, + 382, + 774, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {i j} ^ {k} = \\left\\{ \\begin{array}{l l} 1, & d \\left(v _ {i}, v _ {j}\\right) = k \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 404, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d(v_{i}, v_{j})$ denotes the distance of the shortest path between $v_{i}$ and $v_{j}$ on the skeleton graph. The $k$ -hop neighborhood information is aggregated with a weighted sum of the target node's $k$ -hop neighbors, named HopGCN:", + "bbox": [ + 496, + 452, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {k} = \\sum_ {j} a _ {i j} ^ {k} h _ {j} W ^ {k} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 520, + 890, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_i^k \\in \\mathbb{R}^D$ is a hidden representation of the $k$ -hop neighborhood information. Eq. 5 is similar to Eq. 1 but with the extended definition of $A$ to $k$ hops. The weight matrix $W$ is assigned to the respective hops.", + "bbox": [ + 496, + 561, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before aggregating the hidden representation $s_i^k$ to the target node, we propose a transformer-like attention mechanism computing the similarity between the node feature $h_i$ and the $k$ -hop hidden representation, as shown in Eq. 6:", + "bbox": [ + 496, + 623, + 890, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} ^ {k} = \\sum_ {j} \\operatorname {s o f t m a x} _ {j} \\left(\\frac {h _ {i} s _ {j} ^ {k T}}{\\sqrt {d}}\\right) s _ {j} ^ {k}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 690, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the structure of the human skeleton and the joint synergy, the $k$ -hop neighborhood of joint $i$ is related to that of other joints. In Fig. 2, for example, the synergy of hop $S_5^1$ is related to hop $S_6^1$ and $S_3^1$ . By packing $z_i^k$ , $h_i$ , and $s_i^k$ into matrices $Z^k$ , $H$ , and $S^k$ , we can calculate $Z^k$ in parallel:", + "bbox": [ + 496, + 741, + 890, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ ^ {k} = \\operatorname {s o f t m a x} \\left(\\frac {H S ^ {k ^ {T}}}{\\sqrt {d}}\\right) S ^ {k} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 824, + 890, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$Z^k$ can also be reconstructed using the self-attention mechanism purely on hop features or swap the positions of $H$ and", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14988", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$S^k$ . A performance comparison of these cases is discussed in the experimental results.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further reduce the dimension of hidden representation by using a fully connected(FC) layer based on the order of the neighborhood, considering different amounts of information related to the target node, as formulated in Eq. 8.", + "bbox": [ + 76, + 121, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} ^ {k} = F ^ {k} \\left(z _ {i} ^ {k}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 194, + 468, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_i^k$ is the refined representation of the $k$ -hop neighborhood information with respect to node $i$ , and $F^k$ is the mapping function. The refined representation is then concatenated to the updated node feature.", + "bbox": [ + 76, + 217, + 468, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All the refined representations of hop-wise neighborhoods and the target node feature $h_i$ are encoded to a $D$ -dimensional vector:", + "bbox": [ + 76, + 277, + 468, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {\\prime} = F \\left(h _ {i}, r _ {i} ^ {1}, r _ {i} ^ {2}, r _ {i} ^ {3}, \\dots , r _ {i} ^ {k}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 330, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_i'$ is the final updated feature of node $i$ in this layer, and $F$ is the aggregation function. Packing together the updated features of all nodes, Eq. 9 can be rewritten as:", + "bbox": [ + 76, + 357, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {\\prime} = F \\left(H, R ^ {1}, R ^ {2}, R ^ {3}, \\dots , R ^ {k}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 410, + 468, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thereby obtain as the core layer of the HGF module the hop-wise attention (HA) layer, which extracts latent correlations between feature groups and aggregates $k$ -hop neighborhood information. With the increment of $k$ , we get more $N$ groups on which to explore the underlying joint synergies. The value of $k$ can be adjusted based on the tasks and pipelines. According to the experimental results, three hops achieve optimal performance in the HopFIR architecture.", + "bbox": [ + 76, + 436, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fuse the current global information and original 2D information, we concatenate all the joint features in a batch and feed them into a FC layer to extract the current global information. We then concatenate the extracted global information, 2D information, and the output of the HA layer on the feature dimension of the joint feature to fuse all the information in another FC layer.", + "bbox": [ + 76, + 558, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Intragroup Joint Refinement", + "text_level": 1, + "bbox": [ + 76, + 671, + 333, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed HGF module splits the skeleton graph into different groups based on the $k$ -hop neighborhood of each joint, which attends to the key groups in joint synergies. However, HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interaction of the peripheral joints associated with limbs, such as the wrists and feet. We introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information, as shown in Fig. 2(c). The HGF features of the joints in each limb group are refined in the IJR module by using multi-head self-attention [33], as in Eq. 11.", + "bbox": [ + 75, + 694, + 468, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {g} = M H S A \\left(M H S A (H) _ {g}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ is the feature matrix from the HGF module, $MHSA(H)_g$ is the feature matrix of the group $g$ updated by global multi-head self-attention, and $H_{g}$ is the final feature matrix of the group $g$ updated by the IJR module. More details of the IJR module are provided in Fig. 3.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. The HopFIR Architecture", + "text_level": 1, + "bbox": [ + 500, + 174, + 736, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The HopFIR architecture consists of the proposed HGF and IJR modules, as illustrated in Fig. 3. The residual block, which contains two HGF modules and one IJR module, is designed as the basic block in HopFIR. Moreover, we define a linear embedding layer to map the input to the latent space and a HG module to transform the output into 3D space. The HG module is a variant of HGF designed for the output layer, more details of which are provided in the supplementary material. HopFIR accepts 2D keypoints as input, which can be obtained via an off-the-shelf 2D detector. The graph is obtained by adding a normalized globally learnable $k$ -hop graph to the skeleton graph and then symmetrically normalizing it, as in [49]. We use the L1-norm loss and L2-norm loss to compute the error between ground truth and prediction with the weighted sum as follows:", + "bbox": [ + 498, + 196, + 890, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\alpha \\sum_ {n = 1} ^ {N} \\| Y _ {n} - \\hat {Y _ {n}} \\| _ {2} + \\beta \\sum_ {n = 1} ^ {N} \\| Y _ {n} - \\hat {Y _ {n}} \\| _ {1} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 431, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the joint number, $\\hat{Y}_n$ is the predicted 3D position of joint $n$ , $Y_n$ is the ground truth, $\\alpha = 1$ , and $\\beta = 0.1$ .", + "bbox": [ + 498, + 483, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 526, + 632, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the experimental setup and implementation details of the HopFIR networks. We then present our experimental results and comparisons with state-of-the-art methods. Finally, we conduct several ablation studies of the proposed architecture.", + "bbox": [ + 498, + 551, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 500, + 635, + 797, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Human3.6M [10] is currently the largest publicly available dataset for 3D human pose estimation, with 3.6 million video frames. It captures accurate 3D human joint positions from four camera viewpoints and records 11 subjects performing 15 assigned actions. Following previous works [49, 36, 42], we train our model on five subjects (S1, S5, S6, S7, S8) and test it on two subjects (S9, S11). In contrast to Human3.6M, MPI-INF-3DHP [21] includes complex outdoor scenes, which are commonly used to evaluate the generalizability of proposed methods. Accordingly, we use the test set of MPI-INF-3DHP to verify the generalizability of our model.", + "bbox": [ + 498, + 657, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. For Human3.6M [10], we evaluate our model on two standard evaluation protocols: the mean per-joint position error (MPJPE) and the mean per-joint position error after Procrustes alignment (P-MPJPE).", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14989", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4cc59a59509aca93b72db87c83a1f642cb4ecb8fa472d5bd328bc366cbf4f952.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Martinez et al. [20]ICCV201751.856.258.159.069.578.455.258.174.094.662.359.165.149.552.462.9
Zhao et al. [42](†)CVPR201947.360.751.460.561.149.947.368.186.255.067.861.042.160.645.357.6
Ci et al. [4](†)ICCV201946.852.344.750.452.968.949.646.460.278.951.250.054.840.443.352.7
Pavlo et al. [25]CVPR201947.150.649.051.853.661.449.447.459.367.452.449.555.339.542.751.8
Cai et al. [2](†)ICCV201946.548.847.650.952.961.348.345.859.264.451.248.453.539.241.250.6
Liu et al. [18](†)ECCV202046.352.247.350.755.567.149.246.060.471.151.550.154.540.343.752.4
zeng et al. [39]ECCV202044.548.247.147.851.256.850.145.659.966.452.145.354.239.140.349.9
Zou et al. [49](†)ICCV202145.449.245.749.450.458.247.946.057.563.049.746.652.238.940.849.4
Xu et al. [36](†)CVPR202145.249.947.550.954.966.148.546.359.771.551.448.653.939.944.151.9
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202245.250.848.050.054.965.048.247.160.270.051.648.754.139.743.151.8
Ours(Δ)(†)43.947.645.548.950.158.046.244.555.762.949.045.851.838.039.948.5
", + "bbox": [ + 81, + 89, + 890, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5d34869a30dcc425f42bc8842ac9fe600896a4a7dee51ab101b62a3d12f3c9dd.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on Human3.6M with detected 2D poses as input under Protocol #1, in millimeters. The best results are highlighted in bold and the second-best results are underlined. $(\\dagger)$ indicates GCN-based methods and $(\\Delta)$ indicates Transformer-based methods." + ], + "table_footnote": [], + "table_body": "
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Zhou et al. [45](+)ICCV201934.442.436.642.138.239.834.740.245.660.839.042.642.029.831.739.9
Ci et al. [4](+)(*)†ICCV201936.338.829.737.834.642.539.832.536.239.534.438.438.231.334.236.3
Martinez et al. [20]ICCV201737.744.440.342.148.254.944.442.154.658.045.146.447.636.440.445.5
Zhao et al. [42](†)CVPR201937.849.437.640.945.141.440.148.350.142.253.544.340.547.339.043.8
Cai et al. [2](†)ICCV201933.439.033.837.038.147.339.537.343.246.237.738.038.630.432.138.1
Liu et al. [18](†)ECCV202036.840.333.036.337.545.039.734.940.347.737.438.538.629.632.037.8
zeng et al. [39]ECCV202035.936.729.334.536.042.837.731.740.144.335.837.236.233.734.036.4
Zou et al. [49](†)ICCV2021---------------37.4
Xu et al. [36](†)CVPR202135.838.131.035.335.843.237.331.738.445.535.436.736.827.930.735.8
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202232.038.030.434.434.743.335.231.438.046.234.235.736.127.430.635.2
Ours(Δ)(†)31.334.028.032.033.142.134.128.133.639.831.732.933.826.728.932.7
", + "bbox": [ + 80, + 287, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison on Human3.6M with ground truth 2D keypoints as input under Protocol #1, in millimeters. (+) uses additional data from MPII [1]. (*) uses pose scales in both training and testing. The best results are highlighted in bold.", + "bbox": [ + 75, + 436, + 892, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These are referred to Protocol #1 and Protocol #2, respectively. The MPJPE and P-MPJPE are given in millimeters. For MPI-INF-3DHP [21], we follow previous works [4, 42, 49] by reporting the percentage of correct keypoints (PCK) with a threshold of $150\\mathrm{mm}$ and the area under the curve (AUC) for a range of PCK thresholds.", + "bbox": [ + 75, + 489, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 590, + 294, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous work [25], we obtain the detected 2D poses using the cascaded pyramid network(CPN) [3]. We do not use data augmentation during training and testing with the 2D ground truth input to verify the efficacy of our model. We adopt Adam [11] optimizer and all experiments are conducted on a single NVIDIA RTX 3090 GPU. 3D pose regression from 2D detections is more challenging than that from 2D ground truth because the former needs to deal with some extra uncertainty in the 2D space. To manage this uncertainty, we set different configurations for them. In experiments with 2D ground truth as the input, we train the HopFIR networks with an initial learning rate of 0.001, a decay factor of 0.90 per 4 epochs, a batch size of 64, channels of 128, and PReLU activation [8]. When using detected 2D poses, we train the HopFIR networks with the initial learning rate of 0.006, a decay factor of 0.95 per 4 epochs (but 0.2 for the first 4 epochs), a batch size of 256, 256 channels, and LeaklyReLU activation [19]. To avoid overfitting, we apply Dropout[30] with a dropout rate of 0.5.", + "bbox": [ + 73, + 613, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art", + "text_level": 1, + "bbox": [ + 500, + 489, + 792, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the performance of HopFIR with some SOTA methods on Human3.6M under Protocol #1 and Protocol #2, with the results shown in Table 1. Our method reaches an MPJPE of $48.50\\mathrm{mm}$ and outperforms the best of the existing approaches [49] on all 15 actions. Given the uncertainty of 2D detections, we also investigate the capability of HopFIR networks using ground truth 2D key points as input. As shown in Table 2, HopFIR obtains surprisingly better performance when given precise 2D joint information and produces SOTA results, which verifies its effectiveness.", + "bbox": [ + 496, + 513, + 890, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 678, + 653, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a comprehensive ablation study on Human3.6M to validate the individual effectiveness of each component of the proposed HopFIR architecture under controlled settings. We follow previous works [18, 36, 49] conducting the ablation experiments using GT as inputs to avoid the influence of the 2D pose detector.", + "bbox": [ + 496, + 702, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of Different Modules. We separately verify the effectiveness of the HGF module and IJR module and conduct experiments on first-order neighbors, removing all modules and the multi-hop mechanism. Note that GCN and HopGCN refer to applying only vanilla GCN (Eq. 1) and HopGCN (Eq. 5) to Fig. 3, respectively. The results in Table 3 show that each module improves the performance", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14990", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ca4356cddc9a1cdaa0350cef2d4e30755386f4013353dda55d59dccd1baa982f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodChannelsParamsMPJPEP-MPJPE
GCN1280.36M40.6331.65
HopGCN1280.59M39.1531.40
HopGCN & IJR1282.15M36.6229.23
HopGCN & HGF1281.05M35.1928.81
HopGCN & IJR640.57M36.6929.68
HopGCN & HGF800.48M36.0129.58
HopGCN & HopFIR1282.15M32.6726.20
HopGCN & HopFIR640.54M33.5227.37
HopGCN & HopFIR320.14M35.1928.71
", + "bbox": [ + 101, + 88, + 446, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "over the GCN-only approach, and coupling the two layers to form a HopFIR block achieves further performance improvement. The HopFIR networks reduce the MPJPE to $32.67\\mathrm{mm}$ , which represents a $7.2\\%$ improvement over GraFormer [43]. By reducing feature channels in HopGCN & IJR and HopGCN & HGF, we decrease parameters to 0.57M and 0.48M. However, the models still achieved errors of $36.69\\mathrm{mm}$ and $36.01\\mathrm{mm}$ , compared to the HopGCN with $39.15\\mathrm{mm}$ error and 0.59M parameters, which should be attributed to the structure design other than the model size. Moreover, we reduce the HopFIR network parameters by changing the channels to 64 and 32, respectively, which are also superior to SOTA.", + "bbox": [ + 75, + 261, + 468, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the HA Layer. In sec. 3.2, we introduce $k$ -hop groups to discover latent joint interactions in human joint synergies. The attention matrices in Fig. 5 show the latent joint synergies captured by HopFIR, in which each weight of row $i$ indicates a discovered latent group for the corresponding joint $i$ . To verify the effectiveness of HA layer, we explore the correlation between all joints with a transformer encoder instead of HA layer, which explores the correlation between individual nodes but ignores the synergy between groups of joints in the human body and obtained $36.21 \\mathrm{~mm}$ error. Moreover, we remove the human body prior by using random graph instead of skeleton graph without changing the HopFIR architecture and reached $34.68 \\mathrm{~mm}$ error, suggesting that latent group correlations can be explored by $k$ -hop groups, but group correlations underlying joint synergies can be better explored based on the human body prior.", + "bbox": [ + 75, + 459, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 4, we show the experimental results of three different ways to design the HA layer. HSS is the method selected in this paper, where H, S, and S are $Q$ , $K$ , and $V$ , and H and S represent the node feature and the $k$ -hop group feature, respectively. As we do not follow [33] in applying a linear transformation of $Q$ , $K$ , and $V$ , we also show the result of such a linear transformation, which is denoted as HA+W. The experiment results show that HSS similarity achieves better performance in the HopFIR architecture, but one can choose the type of similarity according to the network property.", + "bbox": [ + 75, + 717, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further insert the HA layer into SOTA pose esti", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a41b765206e7960839280adf1a49d261651442772cd06a5bc263ea3585cf2d9.jpg", + "table_caption": [ + "Table 3. Ablation experiments on the proposed modules." + ], + "table_footnote": [], + "table_body": "
AttentionHAHA+W
HSSSSSSHHHSSSSSSHH
Params2.15M2.50M
MPJPE32.6734.1834.2833.2933.5333.90
P-MPJPE26.2027.7027.7127.1627.4027.41
", + "bbox": [ + 524, + 88, + 867, + 152 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4af30901527f078a841ac5eec718959f8634fc7b33b2c88bb8178a21c2a66af5.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of HA layers with different similarity computing approaches." + ], + "table_footnote": [], + "table_body": "
MethodChannelsParamsMPJPEP-MPJPE
SemGCN [42]1280.27M42.1433.53
SemGCN + HA(HSS)1280.49M38.4130.56
SemGCN + HA(SSS)1280.49M41.3033.07
SemGCN + HA(SHH)1280.49M38.8131.05
SemGCN [42] w/ Non-local [35]1280.43M40.7831.46
SemGCN w/ Non-local +HA(HSS)1280.66M38.0330.50
SemGCN w/ Non-local +HA(SSS)1280.66M37.7530.17
SemGCN w/ Non-local +HA(SHH)1280.66M37.9429.71
Modulated GCN [49]1280.29M38.2530.06
Modulated GCN +HA(HSS)+W1280.96M36.5429.09
Modulated GCN +HA(SSS)+W1280.96M36.1429.02
Modulated GCN +HA(SHH)+W1280.96M37.3830.02
", + "bbox": [ + 504, + 204, + 887, + 347 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ff54f58c35a59cb09a4884c39b39ff8c73e1ce9379aceef76a9f95fbd4206d85.jpg", + "table_caption": [ + "Table 5. Comparison of the improved performance of proposed HA layer added on different methods. We test on two GCN-based methods: SemGCN [42] and MGCN [49]." + ], + "table_footnote": [], + "table_body": "
Num-kChannelsParamsMPJPEP-MPJPE
11281.88M35.8828.76
21282.03M34.9627.74
31282.15M32.6726.20
41282.27M35.5828.35
Num-BlockChannelsParamsMPJPEP-MPJPE
11280.80M37.3330.71
21281.47M34.2127.63
31282.15M32.6726.20
41282.82M33.8427.63
", + "bbox": [ + 545, + 411, + 846, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study for number of $k$ -hop and designed blocks. The units of MPJPE and P-MPJPE are millimeters (mm).", + "bbox": [ + 500, + 542, + 888, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mation methods, namely SemGCN [42] and Modulated GCN [49], to investigate its generalizability. No changes are made to their source code, with the HA layer inserted before the information aggregation stage of these methods. The experimental results in Table 5 show that the HA layer improves these previous SOTA networks to a large degree; especially, the MPJPE of SemGCN [42], is reduced from $42.14\\mathrm{mm}$ to $38.41\\mathrm{mm}$ , representing an $8.9\\%$ improvement. Moreover, HA layer with linear transformation (W) makes learning more stable, so We test HA and $\\mathrm{HA + W}$ in SemGCN and MGCN, respectively, to show both of them are effective. For a fair comparison of parameters, we tested MLP and MHSA instead of HA, both of them with 1.03M parameters, and achieved $39.12\\mathrm{mm}$ and $39.22\\mathrm{mm}$ errors, respectively. More details can be found in the supplementary material. Experiments on the above methods of aggregating first-order neighbor information demonstrate the effectiveness of HA and also indicate that latent joint grouping can recognize the human joint synergies.", + "bbox": [ + 496, + 597, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Error on Peripheral Joints. We report the regres", + "bbox": [ + 517, + 885, + 888, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14991", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c5a5aa59debb94a5f73b5d1bef79d5835e3620abd2cc57b6041fcf28716a763.jpg", + "image_caption": [ + "Figure 4. Comparison of MPJPE for peripheral joints on the test set of the Human3.6M. R and L denote right and left, respectively." + ], + "image_footnote": [], + "bbox": [ + 120, + 90, + 429, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7c05d19b47f8b72565c420b403d9986caa7f9b957d5f7711721ea52196f428ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BlockChannelsParamsMPJPEP-MPJPE
(H)(I)1281.78M34.4327.79
(I)(H)1281.78M34.1327.56
(H)(I)(H)1282.15M35.0827.98
(I)(H)(H)1282.15M34.2027.60
(H)(H)(I)1282.15M32.6726.20
", + "bbox": [ + 122, + 308, + 424, + 386 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8acee63fa452851815403c3cc7ae484f6c04213c3e542b5102ae4027ac0ac547.jpg", + "table_caption": [ + "Table 7. Ablation study for arrangements of the designed block. (I) and (H) denote IJR module and HGF module, respectively." + ], + "table_footnote": [], + "table_body": "
MPJPEP-MPJPEMPJVE
PoseFormer [44] (T=81)44.336.53.1
MixSTE [41] (T=243)40.932.62.3
MHFormer [17] (T=351)43.0--
P-STMO [28] (T=243)42.134.4-
Ours (T=243)41.132.52.1
", + "bbox": [ + 112, + 436, + 434, + 515 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f8d33cad84c71ee807a4c58b335ea4a092e0f1ca87f676372ec4e3d43981cd8.jpg", + "table_caption": [ + "Table 8. Quantitative comparison on Human3.6M with detected 2D pose (CPN) in video. $T$ denotes the number of input frames." + ], + "table_footnote": [], + "table_body": "
MethodsPCKAUC
GSno GSOutdoorAll
Martinez et al. [20]49.842.531.242.517.0
Ci et al. [4]74.870.877.374.036.7
zeng et al. [39]--80.377.643.8
Li et al. [16]70.168.266.666.9-
Zhao et al. [43]80.177.974.179.043.8
Liu et al. [18] (weight unsharing)77.680.580.179.347.6
Xu et al. [36]81.581.775.280.145.8
Nie et al. [23]---83.545.9
Zou et al. [49]86.486.085.786.153.7
Ours89.185.985.987.257.0
", + "bbox": [ + 88, + 566, + 457, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Quantitative comparisons on the MPI-INF-3DHP test set. GS denotes green screen.", + "bbox": [ + 76, + 709, + 467, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sion accuracy for the peripheral joints (wrist and feet) in Fig. 4 in comparison with some previously proposed methods [49, 42, 39]. The HopFIR network with IJR modules outperforms SOTA methods on the right foot (RFoot) by 6.7 mm, left foot (LFoot) by 7.7 mm, left wrist (LWrist) by 9.6 mm, and right wrist (RWrist) by 7.7 mm. The experimental results verify that the intragroup joint attention within each limb group strengthens the capabilities of the HopFIR.", + "bbox": [ + 75, + 763, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Numbers of $k$ -Hops. As HopFIR is designed", + "bbox": [ + 96, + 883, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30a7532c27391266c82b4ff8f610415ae589b0ed161638b9c984660f9f9d0763.jpg", + "image_caption": [ + "Figure 5. Attention weight of the $j$ -th $k$ -hop for the $i$ -th joint, deeper color indicates higher correlation. $i$ -th row and $j$ -th col represent $i$ -th joint and $k$ -hop of $j$ -th joint, respectively." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 887, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a6921bff2b26e547624cd71f09eaec86ad8f3791dba539140d71055830ec885.jpg", + "image_caption": [ + "Figure 6. Qualitative visual results for HopFIR and MGCN [49] on the Human3.6M. The black lines are the ground truth (GT) and the red lines are the predictions of HopFIR and MGCN. Wrong predictions are circled. The bottom row shows our failure case." + ], + "image_footnote": [], + "bbox": [ + 506, + 258, + 888, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to extract the correlation between feature groups, we set different $k$ values to discover various latent connections underlying the human joint synergies, with the results shown at the top of Table 6. The MPJPE gradually decreases as the number of hops increases, and reaches the best performance at 3 hops. Therefore, the optimal number of hops for 3D HPE is 3, which entails that we obtain $16 \\times 3$ groups from the skeleton graph. Each of the 16 groups corresponds to a potential correlation among coupled nodes at different distances, and three hops is sufficient to recognize the joint synergies.", + "bbox": [ + 496, + 598, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Arrangement of the Designed Block. To investigate the optimal structure of the designed block, experiments are performed with various block numbers and various combinations of HGF and IJR modules. As shown at the bottom of Table 6, the error gradually decreases as the number of blocks increases, until the best performance is achieved at 3 blocks. As shown in Table 7, the (H)(H)(I) arrangement achieves the optimal results by reducing the error to 32.67 mm. HGF treats each hop as a group and applies a hop", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14992", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "wise attention mechanism to these groups to discover latent joint synergy. IJR utilizes the limb prior for peripheral joint refinement. Thus, (H)(H)(I) first integrates the complete joint information and then refines it by IJR. While (I)(H)(H) reverses this procedure resulting in the insufficient utilization of joint information. That the results are superior to GraFormer [43] in all cases except for the single-block model indicates that the HopFIR has significant human pose representation capabilities.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Extend to Temporal Domain. Without a specific design to integrate temporal information, we extend to the temporal domain by adding two TEs after each block and replacing the HG module with a linear layer, and achieved competitive results as shown in table 8.", + "bbox": [ + 75, + 227, + 470, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cross-Dataset Results on MPI-INF-3DHP. Table 9 further compares HopFIR with previous methods on cross-dataset scenarios to validate its generalizability. For these experiments, we train our model on the Human3.6M dataset and test it on the test set of the MPI-INF-3DHP dataset. The results show that our approach obtains better results than other methods, which verifies the generalizability of our approach to unseen scenarios.", + "bbox": [ + 75, + 303, + 470, + 425 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 6, we show the visual results on Human3.6M in the world space. The bottom of the figure shows some failure cases of HopFIR, which predict some wrong joint positions. The figure shows that HopFIR is able to predict 3D joint positions more accurately, even for poses that cause difficulties for MGCN.", + "bbox": [ + 75, + 425, + 470, + 515 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 531, + 204, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) as a novel architecture for 3D human pose estimation. The proposed architecture mainly comprises the HGF and IJR modules. The HGF module improves on the GCN-based pose estimation networks by grouping the joints by $k$ -hop neighborhood and capturing the potential joint correlations in the different joint synergies. Because the peripheral joints strongly interact with intra-limb joints, the proposed IJR module applies intragroup attention to refine the peripheral joint features through the associated limb. The proposed method achieves new state-of-the-art results while maintaining a modest model size.", + "bbox": [ + 75, + 556, + 472, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 768, + 236, + 786 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China (Grant No. 2021YFC0122602), in part by the Joint Funds Program of the National Natural Science Foundation of China (Grant No. U21A20517), in part by the Basic Science Centre Program of National Natural Science Foundation of China (Grant No. 72188101).", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 90, + 596, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 6", + "[2] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2272-2281, 2019. 1, 6", + "[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 6", + "[4] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019. 2, 3, 6, 8", + "[5] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[7] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., 2:729-734 vol. 2, 2005. 3", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1, 2, 5", + "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[12] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3" + ], + "bbox": [ + 501, + 116, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14993", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017. 1", + "[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1", + "[15] Kyoungoh Lee, Inwooong Lee, and Sanghoon Lee. Propagating lstm: 3d pose estimation based on joint interdependency. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 1", + "[16] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 8", + "[17] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 3, 8", + "[18] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In European Conference on Computer Vision, pages 318-334. Springer, 2020. 1, 2, 3, 6, 8", + "[19] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, volume 30, page 3. Atlanta, Georgia, USA, 2013. 6", + "[20] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1, 2, 6, 8", + "[21] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 2, 5, 6", + "[22] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010. 3", + "[23] Qiang Nie, Ziwei Liu, and Yunhui Liu. Lifting 2d human pose to 3d with domain adapted 3d body concept. International Journal of Computer Vision, 131(5):1250-1268, 2023. 8", + "[24] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2", + "[25] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7753-7762, 2019. 6", + "[26] Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Reconstructing 3d human pose from 2d image landmarks. In" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European conference on computer vision, pages 573-586. Springer, 2012. 2", + "[27] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008. 3", + "[28] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 461-478. Springer, 2022. 3, 8", + "[29] Cristian Sminchisescu. 3d human motion analysis in monocular video: techniques and challenges. In Human motion, pages 185-211. Springer, 2008. 2", + "[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958, 2014. 6", + "[31] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 1", + "[32] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 1", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 7", + "[34] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 3", + "[35] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 7", + "[36] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16105-16114, 2021. 1, 2, 5, 6, 8", + "[37] Youze Xue, Jiansheng Chen, Xiangming Gu, Huimin Ma, and Hongbing Ma. Boosting monocular 3d human pose estimation with part aware attention. IEEE Transactions on Image Processing, 31:4278-4291, 2022. 2", + "[38] Wei Yang, Wanli Ouyang, Xiaolong Wang, Jimmy Ren, Hongsheng Li, and Xiaogang Wang. 3d human pose estimation in the wild by adversarial learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5255-5264, 2018. 1", + "[39] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Snet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In European Conference on Computer Vision, pages 507-523. Springer, 2020. 1, 2, 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14994", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11436-11445, 2021. 2, 3, 4", + "[41] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13232-13242, 2022. 3, 8", + "[42] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3425-3435, 2019. 1, 3, 4, 5, 6, 7, 8", + "[43] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 1, 3, 6, 7, 8, 9", + "[44] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 3, 8", + "[45] Kun Zhou, Xiaoguang Han, Nianjuan Jiang, Kui Jia, and Jiangbo Lu. Hemlets pose: Learning part-centric heatmap triplets for accurate 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2344-2353, 2019. 6", + "[46] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In Proceedings of the IEEE International Conference on Computer Vision, pages 398-407, 2017. 2", + "[47] Yiran Zhu, Xing Xu, Fumin Shen, Yanli Ji, Lianli Gao, and Heng Tao Shen. PosegTac: Graph transformer encoder-decoder with atrous convolution for 3d human pose estimation. In IJCAI, pages 1359-1365, 2021. 3", + "[48] Zhiming Zou, Kenkun Liu, Le Wang 0003, and Wei Tang. High-order graph convolutional networks for 3d human pose estimation. In BMVC, 2020. 3, 4", + "[49] Zhiming Zou and Wei Tang. Modulated graph convolutional network for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11477-11487, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 739 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14995", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_model.json b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f3f10cc3b0b8745af3a5cf4e106f60217f38f5d --- /dev/null +++ b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.175 + ], + "angle": 0, + "content": "HopFIR: Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.203, + 0.802, + 0.243 + ], + "angle": 0, + "content": "Kai Zhai\\(^{1*}\\) Qiang Nie\\(^{2*}\\) Bo Ouyang\\(^{1\\dagger}\\) Xiang Li\\(^{3}\\) Shanlin Yang\\(^{1}\\) \n\\(^{1}\\)Hefei University of Technology \\(^{2}\\)Youtu Lab, Tencent \\(^{3}\\)Tsinghua University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.289, + 0.314, + 0.306 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.474, + 0.732 + ], + "angle": 0, + "content": "2D-to-3D human pose lifting is fundamental for 3D human pose estimation (HPE), for which graph convolutional networks (GCNs) have proven inherently suitable for modeling the human skeletal topology. However, the current GCN-based 3D HPE methods update the node features by aggregating their neighbors' information without considering the interaction of joints in different joint synergies. Although some studies have proposed importing limb information to learn the movement patterns, the latent synergies among joints, such as maintaining balance are seldom investigated. We propose the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture to tackle the 3D HPE problem. HopFIR mainly consists of a novel hop-wise GraphFormer (HGF) module and an intragroup joint refinement (IJR) module. The HGF module groups the joints by \\( k \\)-hop neighbors and applies a hopwise transformer-like attention mechanism to these groups to discover latent joint synergies. The IJR module leverages the prior limb information for peripheral joint refinement. Extensive experimental results show that HopFIR outperforms the SOTA methods by a large margin, with a mean per-joint position error (MPJPE) on the Human3.6M dataset of \\( 32.67\\mathrm{mm} \\). We also demonstrate that the state-of-the-art GCN-based methods can benefit from the proposed hop-wise attention mechanism with a significant improvement in performance: SemGCN [42] and MGCN [49] are improved by \\( 8.9\\% \\) and \\( 4.5\\% \\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.761, + 0.21, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Monocular 3D human pose estimation aims to accurately regress the 3D locations of human joints in the camera coordinate system from a single image. It plays an important role in many applications, such as action recognition and human-computer interaction. Compared with the monocu" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.283, + 0.865, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.542, + 0.895, + 0.669 + ], + "angle": 0, + "content": "Figure 1. Comparison of performance and model size between the proposed HopFIR and SOTA methods, namely Modulated GCN (MGCN) [49], SemGCN [42], Weight Unsharing [18], SR-Net [39], GraphSH [36], and GraFormer [43]. The methods are evaluated on the Human3.6M dataset [10] with ground truth 2D joints as input. The arrow shows the performance improvement obtained by inserting the HA layer into the other networks. Tiny, small, and normal denote the feature dimension of the HopFIR are 32, 64, and 128, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "lar systems, multi-view capture systems are expensive and inconvenient to set up and operate, which prevents them from being widely used in practice. To tackle the monocular 3D HPE task, some approaches [2, 15, 31, 32, 38] estimates 3D joint coordinates or heat maps directly from an image via a convolutional neural network (CNN) [13, 14]. However, direct regression from the image space suffers from the problem of a large parameter searching space, which always leads to a sub-optimal solution. Recently, Martinez et al. [20] constructed a simple fully connected network using only 2D keypoints as input and achieved promising 3D HPE performance, showing that the 3D human pose can be efficiently and accurately estimated from 2D joint po" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.876, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution †Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14985" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "sitions. Inspired by them and considering the thoroughly investigated 2D HPE, many works decompose the problem into two subtasks, i.e., the 2D HPE and 2D-to-3D pose lifting [18, 40, 49, 36]. 2D-to-3D pose lifting, therefore, has emerged as a fundamental task in this area that our work devotes to." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.471, + 0.442 + ], + "angle": 0, + "content": "A key consideration in 2D-to-3D pose lifting is that the human skeleton topology is inherently sparse and graph-structured. Fully connected neural networks are less effective in modeling graph-structured data due to their simple connections among all nodes and the probability of overfitting. To leverage the information of the human skeletal topology, some works [36, 40, 49] have proposed to model the human body with GCNs and have achieved SOTA results. For example, Ci et al. [4] introduced a locally connected network to enhance the representation capability of the GCN, and Liu et al. [18] explored the weight sharing and feature transformation that occurs before or after feature aggregation in the GCN. One limitation of these GCN-based 3D HPE methods, however, is that they update the node features by aggregating their neighbors' information without considering the different contributions of these nodes to different joint synergies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Instead of considering all the joints of a skeleton as a whole, Xue et al. [37] demonstrated that the human skeleton exhibits obvious part-wise inconsistency in its motion patterns, as also reported in SRNet [39]. However, these works are limited to considering prior structural information of limb groupings and ignore investigating the latent groups underlying joint synergies. For example, the relative positions of the 1-hop neighbors of joint 0 are almost constant in \"Discussion\" subject, which can be a latent group, as shown in Fig. 2. Moreover, these works consider the joints in a limb as a whole to calculate the relationship with other limbs, which resulted in lower accuracy for peripheral joints, such as wrists and feet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To address the abovementioned problems in monocular 3D HPE, we propose a novel architecture: the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR). The first key component of HopFIR is a novel hop-wise GraphFormer (HGF) module that considers \\( k \\)-hop neighbors. In the HGF module, the information of every hop of every joint is aggregated into the hidden space, such that \\( N \\times k \\) groups of features are obtained for a skeleton model with \\( N \\) joints. Meanwhile, a hop-wise transformer-like attention mechanism is designed to extract the correlation among feature groups, which computes similarity by the dot product of the node feature and the group feature. The proposed HGF module enables the network to discover latent joint interactions considering human joint synergy. Because the HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interactions of peripheral joints associated with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "a limb, we introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information. Specifically, a residual block is built from two HGF modules followed by one IJR module. The proposed HopFIR architecture achieves optimal regression accuracy with a stack of three blocks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.211 + ], + "angle": 0, + "content": "To summarize, our work makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.288 + ], + "angle": 0, + "content": "- To the best of our knowledge, we design the first Hopwise GraphFormer module to explore potential joint correlations underlying human joint synergy. We also prove that other GCN-based methods can benefit from the proposed HGF module efficiently, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.363 + ], + "angle": 0, + "content": "- We design an Intragroup Joint Refinement module, which attends to intragroup joints to refine joint features through the associated limb, especially the wrists and feet. The IJR module enables HGF modules to discover the latent synergies among joints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.438 + ], + "angle": 0, + "content": "- We propose the novel Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture for 3D HPE, which is built entirely from HGF and IJR modules. Specifically, two HGF modules and one IJR module are coupled into a block." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.514 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness and generalizability of the proposed modules and HopFIR architecture by providing new state-of-the-art results on two challenging datasets, i.e., Human3.6M [10] and MPI-INF-3DHP [21]." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.642, + 0.544 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.903 + ], + "angle": 0, + "content": "3D Human Pose Estimation. Early works [26, 29] use handcrafted features, perspective relationships, and geometric constraints to estimate the 3D human pose. Recent pose estimation approaches can be generally divided into two categories. The first category of networks regresses 3D human joints directly from the image [24, 46]. Pavlakos et al. [24] adopted a CNN to predict the voxel-wise likelihoods for each joint, and Zhou et al. [46] directly embedded a kinematic object model into the networks to learn the general multi-articulate object pose. Approaches in the second category decouple the 3D HPE task into 2D pose estimation from an image and 3D pose estimation from the detected 2D joints (2D-to-3D). For example, Martinez et al. [20] proposed a simple yet effective baseline with fully-connect networks and proved that 3D human poses can be regressed simply and effectively from 2D keypoints. Our paper follows this pipeline and focuses on the 2D-to-3D pose lifting. For promoting 3D human pose regression accuracy, it is crucial to group joints with consideration of their interactions rather than treating all joints of a skeleton as a whole. Xue et al. [37] divided the human skeleton graph into five groups according to the limbs to explore part-wise motion inconsistency, and Zeng et al. [39] split the human joints into" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14986" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.183 + ], + "angle": 0, + "content": "local regions and recombined the global information from the rest of the joints. Our proposed HopFIR differs from these approaches by grouping joints by the \\(k\\)-hop neighbors of each joint and prior limb information, which enables the network to discover latent connections between groups in different human joint synergies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.468, + 0.292 + ], + "angle": 0, + "content": "Graph Convolutional Networks. GCNs [5, 7, 12, 27] generalize the capability of CNNs by performing convolution operations on graph-structured data. GCNs can be divided into two categories: the spectral-based approaches [5] and the spatial-based approaches [12]. Our approach falls into the second category, which applies message-passing operations on the graph nodes and their neighbors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Due to the graph-structure topology of the human skeleton, many works [4, 18] have introduced GCN to tackle the 3D HPE task. Zhao et al. [42] proposed a SemGCN to learn the semantic relationships between human joints, and Zou et al. [49] proposed a weight modulation and an affinity modulation based on the SemGCN. These methods aggregate the first-order neighborhood messages to update the feature matrix by assigning different weights to different nodes. Some works [47, 48] have extended the first-order neighbors to high-order neighbors in the spatial domain directly. Zeng et al. [40] designed a hierarchical fusion block by dividing the fusion procedure into two stages, where all the high-order neighbors of a node are aggregated into a feature in the first stage and fuse it with the node feature and the first-order neighbor in the second stage. Zhao et al. [43] introduced Chebyshev graph convolution to fuse information among the \\(k\\)-hop neighbors of a joint directly. These works updated features by aggregating each node's own \\(k\\)-hop neighborhood information in a GCN layer. However, HopFIR considers the \\(k\\)-hop groups of all nodes to reconstruct the \\(k\\)-hop feature of a node through the proposed attention mechanism, which can enhance the representation capability of GCNs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Graph Attention. Graph attention networks [34] is a pioneer work that pay attention to the data in a graph structure by assigning an attention weight to each node. The introduction of the transformer [33] for machine translation tasks has proven the capacity of attention for sequential input. ViT [6] introduces the transformer into computer vision and achieves excellent performance. Inspired by them, some researchers have adopted the transformer for 3D HPE. Zhao et al. [43] applied self-attention to capture global information by calculating the similarity of all nodes. PoseFormer [44] directly applies Transformer Encoder (TE) by viewing each joint and each frame as tokens in the spatial and temporal domains, respectively. MixSTE [41] is based on [44], where each joint feature is represented by the temporal TE to model the joint motion. P-stmo [28] replaces the spatial TE with MLP and applies Stride Transformer Encoder to map \\(N\\) frames to one frame. MHFormer [17] generates" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.09, + 0.628, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.204, + 0.585, + 0.213 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.092, + 0.765, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.203, + 0.709, + 0.213 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.092, + 0.869, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.203, + 0.823, + 0.213 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Figure 2. An illustration of the human skeleton graph and the groups in HGF and IJR modules. The \\(k\\)-hop neighbors of a joint are set as a group in HGF modules where (a) indicates the 1-hop groups of different joints and (b) indicates several latent groups because of physical limitations. (c) indicates the joints grouped by prior limb information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.506 + ], + "angle": 0, + "content": "multi-hypothesis at different depths of stacked spatial TEs by viewing all frames of each joint feature as a token, then communicates these hypotheses with cross-attention and self-attention. Different from the existing graph attention for 3D HPE, we propose the intergroup multi-head attention mechanism among the \\(k\\)-hop groups of all nodes, which assigns the attention weights by computing the similarity between the node feature and \\(k\\)-hop group feature. Moreover, we introduce the intragroup multi-head self-attention in limb groups to refine the joint features and promote the HGF module to discover the latent synergies among joints." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.715, + 0.536 + ], + "angle": 0, + "content": "3. The Proposed HopFIR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.664 + ], + "angle": 0, + "content": "This paper proposes a novel architecture to regress the 3D human pose from \\(N\\) given 2D keypoints \\(X \\in \\mathbb{R}^{N \\times 2}\\). The proposed framework mainly consists of the HGF and IJR modules. In this section, we first review the vanilla Graph Convolutional Network and Transformer in Sec. 3.1. We then introduce the HGF and IJR modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we present the network architecture in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.771, + 0.688 + ], + "angle": 0, + "content": "3.1. Vanilla GCN and Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.772 + ], + "angle": 0, + "content": "GCN. A graph is defined as \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E})\\), where \\(\\mathcal{V}\\) is a set of \\(N\\) nodes and \\(\\mathcal{E}\\) is the adjacency matrix representing the edges between the nodes. Given a collection of input features \\(H^{l}\\in \\mathbb{R}^{N\\times D}\\), a generic GCN layer that aggregates neighborhood information can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.781, + 0.891, + 0.8 + ], + "angle": 0, + "content": "\\[\nH ^ {(l + 1)} = \\sigma (\\tilde {A} H ^ {l} W) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(W\\in \\mathbb{R}^{D\\times D^{\\prime}}\\) is the learnable weight matrix that transforms the feature dimension from \\(D\\) to \\(D^{\\prime}\\) , and \\(\\sigma (.)\\) is the activation function, such as ReLU [22]. \\(H^{(l + 1)}\\) is the updated feature matrix, \\(\\tilde{A}\\in \\mathbb{R}^{N\\times N}\\) is the symmetrically normalized affinity matrix [12] with added selfconnections, and \\(A\\in \\{0,1\\}^{N\\times N}\\) is the adjacency ma" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14987" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.17, + 0.091, + 0.805, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Figure 3. The HopFIR architecture, with details of the HGF and IJR modules. The designed block is a residual block [9] built by two HGF modules followed by one IJR module. The proposed architecture with three blocks achieves optimal performance. Arrows of different colors represent different hops and groups in HGF and IJR, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.412 + ], + "angle": 0, + "content": "trix. The \\((i,j)th\\) entry \\(a_{ij} = 1\\) representing node \\(j\\) is the neighbor of node \\(i\\). Otherwise, they are not connected and \\(a_{ij} = 0\\). Therefore, the none-neighbor nodes have a weak influence on each other in the vanilla GCN, which hinders the modeling of underlying joint synergy in 3D HPE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.471, + 0.504 + ], + "angle": 0, + "content": "Transformer. Transformer architecture relies entirely on self-attention to compute representations of its input and output. The self-attention function maps the inputs to the queries \\( Q \\), keys \\( K \\), and values \\( V \\) by weight matrices \\( W_{Q} \\), \\( W_{K} \\), and \\( W_{V} \\), respectively, and the matrix of outputs is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.516, + 0.47, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(Q K ^ {T} / \\sqrt {d}\\right) V \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.64 + ], + "angle": 0, + "content": "where \\(d\\) is the feature dimension of \\(Q\\), and \\(\\frac{1}{\\sqrt{d}}\\) is a scaling factor to prevent extremely small gradients. The multihead self-attention (MHSA), which performs self-attention in parallel, projects the queries, keys, and values \\(P\\) times with different linear projections to the respective subspaces, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.654, + 0.47, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(D _ {1}, \\dots , D _ {P}\\right) W _ {o} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.684, + 0.47, + 0.716 + ], + "angle": 0, + "content": "where \\( W_O \\) is the projection matrix of outputs, \\( D_p = \\text{Attention}(QW_Q^p, KW_K^p, VW_V^p) \\), and \\( p \\in [1,..,P] \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.3, + 0.743 + ], + "angle": 0, + "content": "3.2. Hop-wise GraphFormer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Previous GCN studies for 3D HPE aggregate multi-hop neighborhood information [40, 48] or assign an attention weight to each first-order neighbor [42, 49]. To effectively capture the node's neighborhood message and increase the representational capacity of GCNs, we introduce the hopwise GraphFormer (HGF) module, which treats each hop as a group (Fig.2 a,b) and computes the attention weights for each hop (more intuitive descriptions for \\( k \\)-hop can be found in the supplementary material). By considering the relationship within \\( k \\) hops, we can obtain \\( N \\times k \\) groups for a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.382 + ], + "angle": 0, + "content": "skeleton graph with \\(N\\) joints, which provides enough combinations of joints to discover the latent correlations among joints in different human joint synergy." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.383, + 0.776, + 0.398 + ], + "angle": 0, + "content": "We first define the \\(k\\)-hop matrix \\(A^k\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "\\[\na _ {i j} ^ {k} = \\left\\{ \\begin{array}{l l} 1, & d \\left(v _ {i}, v _ {j}\\right) = k \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.514 + ], + "angle": 0, + "content": "where \\( d(v_{i}, v_{j}) \\) denotes the distance of the shortest path between \\( v_{i} \\) and \\( v_{j} \\) on the skeleton graph. The \\( k \\)-hop neighborhood information is aggregated with a weighted sum of the target node's \\( k \\)-hop neighbors, named HopGCN:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.521, + 0.892, + 0.555 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {k} = \\sum_ {j} a _ {i j} ^ {k} h _ {j} W ^ {k} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.623 + ], + "angle": 0, + "content": "where \\( s_i^k \\in \\mathbb{R}^D \\) is a hidden representation of the \\( k \\)-hop neighborhood information. Eq. 5 is similar to Eq. 1 but with the extended definition of \\( A \\) to \\( k \\) hops. The weight matrix \\( W \\) is assigned to the respective hops." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Before aggregating the hidden representation \\( s_i^k \\) to the target node, we propose a transformer-like attention mechanism computing the similarity between the node feature \\( h_i \\) and the \\( k \\)-hop hidden representation, as shown in Eq. 6:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.691, + 0.892, + 0.735 + ], + "angle": 0, + "content": "\\[\nz _ {i} ^ {k} = \\sum_ {j} \\operatorname {s o f t m a x} _ {j} \\left(\\frac {h _ {i} s _ {j} ^ {k T}}{\\sqrt {d}}\\right) s _ {j} ^ {k}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.819 + ], + "angle": 0, + "content": "Due to the structure of the human skeleton and the joint synergy, the \\(k\\)-hop neighborhood of joint \\(i\\) is related to that of other joints. In Fig. 2, for example, the synergy of hop \\(S_5^1\\) is related to hop \\(S_6^1\\) and \\(S_3^1\\). By packing \\(z_i^k\\), \\(h_i\\), and \\(s_i^k\\) into matrices \\(Z^k\\), \\(H\\), and \\(S^k\\), we can calculate \\(Z^k\\) in parallel:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.825, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\nZ ^ {k} = \\operatorname {s o f t m a x} \\left(\\frac {H S ^ {k ^ {T}}}{\\sqrt {d}}\\right) S ^ {k} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\(Z^k\\) can also be reconstructed using the self-attention mechanism purely on hop features or swap the positions of \\(H\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14988" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.121 + ], + "angle": 0, + "content": "\\(S^k\\). A performance comparison of these cases is discussed in the experimental results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.122, + 0.469, + 0.182 + ], + "angle": 0, + "content": "We further reduce the dimension of hidden representation by using a fully connected(FC) layer based on the order of the neighborhood, considering different amounts of information related to the target node, as formulated in Eq. 8." + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.195, + 0.469, + 0.214 + ], + "angle": 0, + "content": "\\[\nr _ {i} ^ {k} = F ^ {k} \\left(z _ {i} ^ {k}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.218, + 0.469, + 0.278 + ], + "angle": 0, + "content": "where \\( r_i^k \\) is the refined representation of the \\( k \\)-hop neighborhood information with respect to node \\( i \\), and \\( F^k \\) is the mapping function. The refined representation is then concatenated to the updated node feature." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.279, + 0.469, + 0.323 + ], + "angle": 0, + "content": "All the refined representations of hop-wise neighborhoods and the target node feature \\( h_i \\) are encoded to a \\( D \\)-dimensional vector:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.331, + 0.469, + 0.35 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {\\prime} = F \\left(h _ {i}, r _ {i} ^ {1}, r _ {i} ^ {2}, r _ {i} ^ {3}, \\dots , r _ {i} ^ {k}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.358, + 0.469, + 0.403 + ], + "angle": 0, + "content": "where \\( h_i' \\) is the final updated feature of node \\( i \\) in this layer, and \\( F \\) is the aggregation function. Packing together the updated features of all nodes, Eq. 9 can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.411, + 0.469, + 0.429 + ], + "angle": 0, + "content": "\\[\nH ^ {\\prime} = F \\left(H, R ^ {1}, R ^ {2}, R ^ {3}, \\dots , R ^ {k}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.437, + 0.469, + 0.558 + ], + "angle": 0, + "content": "We thereby obtain as the core layer of the HGF module the hop-wise attention (HA) layer, which extracts latent correlations between feature groups and aggregates \\( k \\)-hop neighborhood information. With the increment of \\( k \\), we get more \\( N \\) groups on which to explore the underlying joint synergies. The value of \\( k \\) can be adjusted based on the tasks and pipelines. According to the experimental results, three hops achieve optimal performance in the HopFIR architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.559, + 0.469, + 0.663 + ], + "angle": 0, + "content": "To fuse the current global information and original 2D information, we concatenate all the joint features in a batch and feed them into a FC layer to extract the current global information. We then concatenate the extracted global information, 2D information, and the output of the HA layer on the feature dimension of the joint feature to fuse all the information in another FC layer." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.672, + 0.334, + 0.687 + ], + "angle": 0, + "content": "3.3. Intragroup Joint Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.469, + 0.876 + ], + "angle": 0, + "content": "The proposed HGF module splits the skeleton graph into different groups based on the \\(k\\)-hop neighborhood of each joint, which attends to the key groups in joint synergies. However, HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interaction of the peripheral joints associated with limbs, such as the wrists and feet. We introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information, as shown in Fig. 2(c). The HGF features of the joints in each limb group are refined in the IJR module by using multi-head self-attention [33], as in Eq. 11." + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nH _ {g} = M H S A \\left(M H S A (H) _ {g}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "where \\(H\\) is the feature matrix from the HGF module, \\(MHSA(H)_g\\) is the feature matrix of the group \\(g\\) updated by global multi-head self-attention, and \\(H_{g}\\) is the final feature matrix of the group \\(g\\) updated by the IJR module. More details of the IJR module are provided in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.175, + 0.738, + 0.191 + ], + "angle": 0, + "content": "3.4. The HopFIR Architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.198, + 0.892, + 0.426 + ], + "angle": 0, + "content": "The HopFIR architecture consists of the proposed HGF and IJR modules, as illustrated in Fig. 3. The residual block, which contains two HGF modules and one IJR module, is designed as the basic block in HopFIR. Moreover, we define a linear embedding layer to map the input to the latent space and a HG module to transform the output into 3D space. The HG module is a variant of HGF designed for the output layer, more details of which are provided in the supplementary material. HopFIR accepts 2D keypoints as input, which can be obtained via an off-the-shelf 2D detector. The graph is obtained by adding a normalized globally learnable \\( k \\)-hop graph to the skeleton graph and then symmetrically normalizing it, as in [49]. We use the L1-norm loss and L2-norm loss to compute the error between ground truth and prediction with the weighted sum as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.433, + 0.892, + 0.475 + ], + "angle": 0, + "content": "\\[\nL = \\alpha \\sum_ {n = 1} ^ {N} \\| Y _ {n} - \\hat {Y _ {n}} \\| _ {2} + \\beta \\sum_ {n = 1} ^ {N} \\| Y _ {n} - \\hat {Y _ {n}} \\| _ {1} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.484, + 0.891, + 0.515 + ], + "angle": 0, + "content": "where \\(N\\) is the joint number, \\(\\hat{Y}_n\\) is the predicted 3D position of joint \\(n\\), \\(Y_n\\) is the ground truth, \\(\\alpha = 1\\), and \\(\\beta = 0.1\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.633, + 0.544 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.628 + ], + "angle": 0, + "content": "In this section, we first introduce the experimental setup and implementation details of the HopFIR networks. We then present our experimental results and comparisons with state-of-the-art methods. Finally, we conduct several ablation studies of the proposed architecture." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.799, + 0.651 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Datasets. Human3.6M [10] is currently the largest publicly available dataset for 3D human pose estimation, with 3.6 million video frames. It captures accurate 3D human joint positions from four camera viewpoints and records 11 subjects performing 15 assigned actions. Following previous works [49, 36, 42], we train our model on five subjects (S1, S5, S6, S7, S8) and test it on two subjects (S9, S11). In contrast to Human3.6M, MPI-INF-3DHP [21] includes complex outdoor scenes, which are commonly used to evaluate the generalizability of proposed methods. Accordingly, we use the test set of MPI-INF-3DHP to verify the generalizability of our model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Protocols. For Human3.6M [10], we evaluate our model on two standard evaluation protocols: the mean per-joint position error (MPJPE) and the mean per-joint position error after Procrustes alignment (P-MPJPE)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14989" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.226 + ], + "angle": 0, + "content": "
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Martinez et al. [20]ICCV201751.856.258.159.069.578.455.258.174.094.662.359.165.149.552.462.9
Zhao et al. [42](†)CVPR201947.360.751.460.561.149.947.368.186.255.067.861.042.160.645.357.6
Ci et al. [4](†)ICCV201946.852.344.750.452.968.949.646.460.278.951.250.054.840.443.352.7
Pavlo et al. [25]CVPR201947.150.649.051.853.661.449.447.459.367.452.449.555.339.542.751.8
Cai et al. [2](†)ICCV201946.548.847.650.952.961.348.345.859.264.451.248.453.539.241.250.6
Liu et al. [18](†)ECCV202046.352.247.350.755.567.149.246.060.471.151.550.154.540.343.752.4
zeng et al. [39]ECCV202044.548.247.147.851.256.850.145.659.966.452.145.354.239.140.349.9
Zou et al. [49](†)ICCV202145.449.245.749.450.458.247.946.057.563.049.746.652.238.940.849.4
Xu et al. [36](†)CVPR202145.249.947.550.954.966.148.546.359.771.551.448.653.939.944.151.9
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202245.250.848.050.054.965.048.247.160.270.051.648.754.139.743.151.8
Ours(Δ)(†)43.947.645.548.950.158.046.244.555.762.949.045.851.838.039.948.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.237, + 0.893, + 0.278 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on Human3.6M with detected 2D poses as input under Protocol #1, in millimeters. The best results are highlighted in bold and the second-best results are underlined. \\((\\dagger)\\) indicates GCN-based methods and \\((\\Delta)\\) indicates Transformer-based methods." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.289, + 0.891, + 0.428 + ], + "angle": 0, + "content": "
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Zhou et al. [45](+)ICCV201934.442.436.642.138.239.834.740.245.660.839.042.642.029.831.739.9
Ci et al. [4](+)(*)†ICCV201936.338.829.737.834.642.539.832.536.239.534.438.438.231.334.236.3
Martinez et al. [20]ICCV201737.744.440.342.148.254.944.442.154.658.045.146.447.636.440.445.5
Zhao et al. [42](†)CVPR201937.849.437.640.945.141.440.148.350.142.253.544.340.547.339.043.8
Cai et al. [2](†)ICCV201933.439.033.837.038.147.339.537.343.246.237.738.038.630.432.138.1
Liu et al. [18](†)ECCV202036.840.333.036.337.545.039.734.940.347.737.438.538.629.632.037.8
zeng et al. [39]ECCV202035.936.729.334.536.042.837.731.740.144.335.837.236.233.734.036.4
Zou et al. [49](†)ICCV2021---------------37.4
Xu et al. [36](†)CVPR202135.838.131.035.335.843.237.331.738.445.535.436.736.827.930.735.8
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202232.038.030.434.434.743.335.231.438.046.234.235.736.127.430.635.2
Ours(Δ)(†)31.334.028.032.033.142.134.128.133.639.831.732.933.826.728.932.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.437, + 0.893, + 0.466 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on Human3.6M with ground truth 2D keypoints as input under Protocol #1, in millimeters. (+) uses additional data from MPII [1]. (*) uses pose scales in both training and testing. The best results are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.581 + ], + "angle": 0, + "content": "These are referred to Protocol #1 and Protocol #2, respectively. The MPJPE and P-MPJPE are given in millimeters. For MPI-INF-3DHP [21], we follow previous works [4, 42, 49] by reporting the percentage of correct keypoints (PCK) with a threshold of \\(150\\mathrm{mm}\\) and the area under the curve (AUC) for a range of PCK thresholds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.295, + 0.607 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.614, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Following previous work [25], we obtain the detected 2D poses using the cascaded pyramid network(CPN) [3]. We do not use data augmentation during training and testing with the 2D ground truth input to verify the efficacy of our model. We adopt Adam [11] optimizer and all experiments are conducted on a single NVIDIA RTX 3090 GPU. 3D pose regression from 2D detections is more challenging than that from 2D ground truth because the former needs to deal with some extra uncertainty in the 2D space. To manage this uncertainty, we set different configurations for them. In experiments with 2D ground truth as the input, we train the HopFIR networks with an initial learning rate of 0.001, a decay factor of 0.90 per 4 epochs, a batch size of 64, channels of 128, and PReLU activation [8]. When using detected 2D poses, we train the HopFIR networks with the initial learning rate of 0.006, a decay factor of 0.95 per 4 epochs (but 0.2 for the first 4 epochs), a batch size of 256, 256 channels, and LeaklyReLU activation [19]. To avoid overfitting, we apply Dropout[30] with a dropout rate of 0.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.49, + 0.794, + 0.505 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.665 + ], + "angle": 0, + "content": "We compare the performance of HopFIR with some SOTA methods on Human3.6M under Protocol #1 and Protocol #2, with the results shown in Table 1. Our method reaches an MPJPE of \\(48.50\\mathrm{mm}\\) and outperforms the best of the existing approaches [49] on all 15 actions. Given the uncertainty of 2D detections, we also investigate the capability of HopFIR networks using ground truth 2D key points as input. As shown in Table 2, HopFIR obtains surprisingly better performance when given precise 2D joint information and produces SOTA results, which verifies its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.655, + 0.696 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.794 + ], + "angle": 0, + "content": "We conduct a comprehensive ablation study on Human3.6M to validate the individual effectiveness of each component of the proposed HopFIR architecture under controlled settings. We follow previous works [18, 36, 49] conducting the ablation experiments using GT as inputs to avoid the influence of the 2D pose detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Different Modules. We separately verify the effectiveness of the HGF module and IJR module and conduct experiments on first-order neighbors, removing all modules and the multi-hop mechanism. Note that GCN and HopGCN refer to applying only vanilla GCN (Eq. 1) and HopGCN (Eq. 5) to Fig. 3, respectively. The results in Table 3 show that each module improves the performance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14990" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.447, + 0.21 + ], + "angle": 0, + "content": "
MethodChannelsParamsMPJPEP-MPJPE
GCN1280.36M40.6331.65
HopGCN1280.59M39.1531.40
HopGCN & IJR1282.15M36.6229.23
HopGCN & HGF1281.05M35.1928.81
HopGCN & IJR640.57M36.6929.68
HopGCN & HGF800.48M36.0129.58
HopGCN & HopFIR1282.15M32.6726.20
HopGCN & HopFIR640.54M33.5227.37
HopGCN & HopFIR320.14M35.1928.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.22, + 0.438, + 0.234 + ], + "angle": 0, + "content": "Table 3. Ablation experiments on the proposed modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.457 + ], + "angle": 0, + "content": "over the GCN-only approach, and coupling the two layers to form a HopFIR block achieves further performance improvement. The HopFIR networks reduce the MPJPE to \\(32.67\\mathrm{mm}\\), which represents a \\(7.2\\%\\) improvement over GraFormer [43]. By reducing feature channels in HopGCN & IJR and HopGCN & HGF, we decrease parameters to 0.57M and 0.48M. However, the models still achieved errors of \\(36.69\\mathrm{mm}\\) and \\(36.01\\mathrm{mm}\\), compared to the HopGCN with \\(39.15\\mathrm{mm}\\) error and 0.59M parameters, which should be attributed to the structure design other than the model size. Moreover, we reduce the HopFIR network parameters by changing the channels to 64 and 32, respectively, which are also superior to SOTA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.469, + 0.717 + ], + "angle": 0, + "content": "Effectiveness of the HA Layer. In sec. 3.2, we introduce \\( k \\)-hop groups to discover latent joint interactions in human joint synergies. The attention matrices in Fig. 5 show the latent joint synergies captured by HopFIR, in which each weight of row \\( i \\) indicates a discovered latent group for the corresponding joint \\( i \\). To verify the effectiveness of HA layer, we explore the correlation between all joints with a transformer encoder instead of HA layer, which explores the correlation between individual nodes but ignores the synergy between groups of joints in the human body and obtained \\( 36.21 \\mathrm{~mm} \\) error. Moreover, we remove the human body prior by using random graph instead of skeleton graph without changing the HopFIR architecture and reached \\( 34.68 \\mathrm{~mm} \\) error, suggesting that latent group correlations can be explored by \\( k \\)-hop groups, but group correlations underlying joint synergies can be better explored based on the human body prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.884 + ], + "angle": 0, + "content": "In Table 4, we show the experimental results of three different ways to design the HA layer. HSS is the method selected in this paper, where H, S, and S are \\( Q \\), \\( K \\), and \\( V \\), and H and S represent the node feature and the \\( k \\)-hop group feature, respectively. As we do not follow [33] in applying a linear transformation of \\( Q \\), \\( K \\), and \\( V \\), we also show the result of such a linear transformation, which is denoted as HA+W. The experiment results show that HSS similarity achieves better performance in the HopFIR architecture, but one can choose the type of similarity according to the network property." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We further insert the HA layer into SOTA pose esti" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.089, + 0.869, + 0.153 + ], + "angle": 0, + "content": "
AttentionHAHA+W
HSSSSSSHHHSSSSSSHH
Params2.15M2.50M
MPJPE32.6734.1834.2833.2933.5333.90
P-MPJPE26.2027.7027.7127.1627.4027.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.164, + 0.892, + 0.192 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of HA layers with different similarity computing approaches." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.205, + 0.888, + 0.348 + ], + "angle": 0, + "content": "
MethodChannelsParamsMPJPEP-MPJPE
SemGCN [42]1280.27M42.1433.53
SemGCN + HA(HSS)1280.49M38.4130.56
SemGCN + HA(SSS)1280.49M41.3033.07
SemGCN + HA(SHH)1280.49M38.8131.05
SemGCN [42] w/ Non-local [35]1280.43M40.7831.46
SemGCN w/ Non-local +HA(HSS)1280.66M38.0330.50
SemGCN w/ Non-local +HA(SSS)1280.66M37.7530.17
SemGCN w/ Non-local +HA(SHH)1280.66M37.9429.71
Modulated GCN [49]1280.29M38.2530.06
Modulated GCN +HA(HSS)+W1280.96M36.5429.09
Modulated GCN +HA(SSS)+W1280.96M36.1429.02
Modulated GCN +HA(SHH)+W1280.96M37.3830.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.358, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Table 5. Comparison of the improved performance of proposed HA layer added on different methods. We test on two GCN-based methods: SemGCN [42] and MGCN [49]." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.412, + 0.847, + 0.535 + ], + "angle": 0, + "content": "
Num-kChannelsParamsMPJPEP-MPJPE
11281.88M35.8828.76
21282.03M34.9627.74
31282.15M32.6726.20
41282.27M35.5828.35
Num-BlockChannelsParamsMPJPEP-MPJPE
11280.80M37.3330.71
21281.47M34.2127.63
31282.15M32.6726.20
41282.82M33.8427.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.544, + 0.89, + 0.573 + ], + "angle": 0, + "content": "Table 6. Ablation study for number of \\( k \\)-hop and designed blocks. The units of MPJPE and P-MPJPE are millimeters (mm)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.884 + ], + "angle": 0, + "content": "mation methods, namely SemGCN [42] and Modulated GCN [49], to investigate its generalizability. No changes are made to their source code, with the HA layer inserted before the information aggregation stage of these methods. The experimental results in Table 5 show that the HA layer improves these previous SOTA networks to a large degree; especially, the MPJPE of SemGCN [42], is reduced from \\(42.14\\mathrm{mm}\\) to \\(38.41\\mathrm{mm}\\), representing an \\(8.9\\%\\) improvement. Moreover, HA layer with linear transformation (W) makes learning more stable, so We test HA and \\(\\mathrm{HA + W}\\) in SemGCN and MGCN, respectively, to show both of them are effective. For a fair comparison of parameters, we tested MLP and MHSA instead of HA, both of them with 1.03M parameters, and achieved \\(39.12\\mathrm{mm}\\) and \\(39.22\\mathrm{mm}\\) errors, respectively. More details can be found in the supplementary material. Experiments on the above methods of aggregating first-order neighbor information demonstrate the effectiveness of HA and also indicate that latent joint grouping can recognize the human joint synergies." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.89, + 0.902 + ], + "angle": 0, + "content": "Error on Peripheral Joints. We report the regres" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14991" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.091, + 0.43, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.268, + 0.471, + 0.297 + ], + "angle": 0, + "content": "Figure 4. Comparison of MPJPE for peripheral joints on the test set of the Human3.6M. R and L denote right and left, respectively." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.309, + 0.426, + 0.387 + ], + "angle": 0, + "content": "
BlockChannelsParamsMPJPEP-MPJPE
(H)(I)1281.78M34.4327.79
(I)(H)1281.78M34.1327.56
(H)(I)(H)1282.15M35.0827.98
(I)(H)(H)1282.15M34.2027.60
(H)(H)(I)1282.15M32.6726.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.396, + 0.468, + 0.425 + ], + "angle": 0, + "content": "Table 7. Ablation study for arrangements of the designed block. (I) and (H) denote IJR module and HGF module, respectively." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.437, + 0.435, + 0.516 + ], + "angle": 0, + "content": "
MPJPEP-MPJPEMPJVE
PoseFormer [44] (T=81)44.336.53.1
MixSTE [41] (T=243)40.932.62.3
MHFormer [17] (T=351)43.0--
P-STMO [28] (T=243)42.134.4-
Ours (T=243)41.132.52.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.525, + 0.47, + 0.555 + ], + "angle": 0, + "content": "Table 8. Quantitative comparison on Human3.6M with detected 2D pose (CPN) in video. \\( T \\) denotes the number of input frames." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.567, + 0.458, + 0.7 + ], + "angle": 0, + "content": "
MethodsPCKAUC
GSno GSOutdoorAll
Martinez et al. [20]49.842.531.242.517.0
Ci et al. [4]74.870.877.374.036.7
zeng et al. [39]--80.377.643.8
Li et al. [16]70.168.266.666.9-
Zhao et al. [43]80.177.974.179.043.8
Liu et al. [18] (weight unsharing)77.680.580.179.347.6
Xu et al. [36]81.581.775.280.145.8
Nie et al. [23]---83.545.9
Zou et al. [49]86.486.085.786.153.7
Ours89.185.985.987.257.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.71, + 0.468, + 0.739 + ], + "angle": 0, + "content": "Table 9. Quantitative comparisons on the MPI-INF-3DHP test set. GS denotes green screen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.468, + 0.884 + ], + "angle": 0, + "content": "sion accuracy for the peripheral joints (wrist and feet) in Fig. 4 in comparison with some previously proposed methods [49, 42, 39]. The HopFIR network with IJR modules outperforms SOTA methods on the right foot (RFoot) by 6.7 mm, left foot (LFoot) by 7.7 mm, left wrist (LWrist) by 9.6 mm, and right wrist (RWrist) by 7.7 mm. The experimental results verify that the intragroup joint attention within each limb group strengthens the capabilities of the HopFIR." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.885, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Different Numbers of \\( k \\)-Hops. As HopFIR is designed" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.888, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.204, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Figure 5. Attention weight of the \\(j\\)-th \\(k\\)-hop for the \\(i\\)-th joint, deeper color indicates higher correlation. \\(i\\)-th row and \\(j\\)-th col represent \\(i\\)-th joint and \\(k\\)-hop of \\(j\\)-th joint, respectively." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.259, + 0.889, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Figure 6. Qualitative visual results for HopFIR and MGCN [49] on the Human3.6M. The black lines are the ground truth (GT) and the red lines are the predictions of HopFIR and MGCN. Wrong predictions are circled. The bottom row shows our failure case." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.764 + ], + "angle": 0, + "content": "to extract the correlation between feature groups, we set different \\( k \\) values to discover various latent connections underlying the human joint synergies, with the results shown at the top of Table 6. The MPJPE gradually decreases as the number of hops increases, and reaches the best performance at 3 hops. Therefore, the optimal number of hops for 3D HPE is 3, which entails that we obtain \\( 16 \\times 3 \\) groups from the skeleton graph. Each of the 16 groups corresponds to a potential correlation among coupled nodes at different distances, and three hops is sufficient to recognize the joint synergies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Arrangement of the Designed Block. To investigate the optimal structure of the designed block, experiments are performed with various block numbers and various combinations of HGF and IJR modules. As shown at the bottom of Table 6, the error gradually decreases as the number of blocks increases, until the best performance is achieved at 3 blocks. As shown in Table 7, the (H)(H)(I) arrangement achieves the optimal results by reducing the error to 32.67 mm. HGF treats each hop as a group and applies a hop" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14992" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "wise attention mechanism to these groups to discover latent joint synergy. IJR utilizes the limb prior for peripheral joint refinement. Thus, (H)(H)(I) first integrates the complete joint information and then refines it by IJR. While (I)(H)(H) reverses this procedure resulting in the insufficient utilization of joint information. That the results are superior to GraFormer [43] in all cases except for the single-block model indicates that the HopFIR has significant human pose representation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.472, + 0.303 + ], + "angle": 0, + "content": "Extend to Temporal Domain. Without a specific design to integrate temporal information, we extend to the temporal domain by adding two TEs after each block and replacing the HG module with a linear layer, and achieved competitive results as shown in table 8." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.472, + 0.426 + ], + "angle": 0, + "content": "Cross-Dataset Results on MPI-INF-3DHP. Table 9 further compares HopFIR with previous methods on cross-dataset scenarios to validate its generalizability. For these experiments, we train our model on the Human3.6M dataset and test it on the test set of the MPI-INF-3DHP dataset. The results show that our approach obtains better results than other methods, which verifies the generalizability of our approach to unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.472, + 0.516 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 6, we show the visual results on Human3.6M in the world space. The bottom of the figure shows some failure cases of HopFIR, which predict some wrong joint positions. The figure shows that HopFIR is able to predict 3D joint positions more accurately, even for poses that cause difficulties for MGCN." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.532, + 0.205, + 0.547 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.473, + 0.753 + ], + "angle": 0, + "content": "We present the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) as a novel architecture for 3D human pose estimation. The proposed architecture mainly comprises the HGF and IJR modules. The HGF module improves on the GCN-based pose estimation networks by grouping the joints by \\( k \\)-hop neighborhood and capturing the potential joint correlations in the different joint synergies. Because the peripheral joints strongly interact with intra-limb joints, the proposed IJR module applies intragroup attention to refine the peripheral joint features through the associated limb. The proposed method achieves new state-of-the-art results while maintaining a modest model size." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.238, + 0.787 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China (Grant No. 2021YFC0122602), in part by the Joint Funds Program of the National Natural Science Foundation of China (Grant No. U21A20517), in part by the Basic Science Centre Program of National Natural Science Foundation of China (Grant No. 72188101)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.091, + 0.597, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.117, + 0.895, + 0.186 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.189, + 0.895, + 0.274 + ], + "angle": 0, + "content": "[2] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2272-2281, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.276, + 0.895, + 0.346 + ], + "angle": 0, + "content": "[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.349, + 0.894, + 0.405 + ], + "angle": 0, + "content": "[4] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.407, + 0.895, + 0.464 + ], + "angle": 0, + "content": "[5] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.466, + 0.895, + 0.549 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.552, + 0.895, + 0.607 + ], + "angle": 0, + "content": "[7] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., 2:729-734 vol. 2, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.61, + 0.895, + 0.68 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.683, + 0.895, + 0.739 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.742, + 0.895, + 0.811 + ], + "angle": 0, + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.814, + 0.895, + 0.855 + ], + "angle": 0, + "content": "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.858, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[12] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.117, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14993" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[15] Kyoungoh Lee, Inwooong Lee, and Sanghoon Lee. Propagating lstm: 3d pose estimation based on joint interdependency. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[16] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[17] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[18] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In European Conference on Computer Vision, pages 318-334. Springer, 2020. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[19] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, volume 30, page 3. Atlanta, Georgia, USA, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[20] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[21] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[22] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[23] Qiang Nie, Ziwei Liu, and Yunhui Liu. Lifting 2d human pose to 3d with domain adapted 3d body concept. International Journal of Computer Vision, 131(5):1250-1268, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[24] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[25] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7753-7762, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[26] Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Reconstructing 3d human pose from 2d image landmarks. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "European conference on computer vision, pages 573-586. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[27] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[28] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 461-478. Springer, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[29] Cristian Sminchisescu. 3d human motion analysis in monocular video: techniques and challenges. In Human motion, pages 185-211. Springer, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[31] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[32] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[34] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[35] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[36] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16105-16114, 2021. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[37] Youze Xue, Jiansheng Chen, Xiangming Gu, Huimin Ma, and Hongbing Ma. Boosting monocular 3d human pose estimation with part aware attention. IEEE Transactions on Image Processing, 31:4278-4291, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[38] Wei Yang, Wanli Ouyang, Xiaolong Wang, Jimmy Ren, Hongsheng Li, and Xiaogang Wang. 3d human pose estimation in the wild by adversarial learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5255-5264, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[39] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Snet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In European Conference on Computer Vision, pages 507-523. Springer, 2020. 1, 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14994" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[40] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11436-11445, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[41] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13232-13242, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[42] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3425-3435, 2019. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[43] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 1, 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.374, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[44] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.445, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[45] Kun Zhou, Xiaoguang Han, Nianjuan Jiang, Kui Jia, and Jiangbo Lu. Hemlets pose: Learning part-centric heatmap triplets for accurate 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2344-2353, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.515, + 0.47, + 0.583 + ], + "angle": 0, + "content": "[46] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In Proceedings of the IEEE International Conference on Computer Vision, pages 398-407, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[47] Yiran Zhu, Xing Xu, Fumin Shen, Yanli Ji, Lianli Gao, and Heng Tao Shen. PosegTac: Graph transformer encoder-decoder with atrous convolution for 3d human pose estimation. In IJCAI, pages 1359-1365, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.47, + 0.684 + ], + "angle": 0, + "content": "[48] Zhiming Zou, Kenkun Liu, Le Wang 0003, and Wei Tang. High-order graph convolutional networks for 3d human pose estimation. In BMVC, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[49] Zhiming Zou and Wei Tang. Modulated graph convolutional network for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11477-11487, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14995" + } + ] +] \ No newline at end of file diff --git a/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_origin.pdf b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..13d3ed6ab6f8c817a9b137d118809dbc039bbdd6 --- /dev/null +++ b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/f663f584-1a4b-4181-b11b-1abe3d533ef4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5709d896b550128f084265512122ad07327f568e92aa56292cf6cb40fd9d17e +size 1321388 diff --git a/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/full.md b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..679cd3050376e79b6b4961d732d08e97dbd89e8d --- /dev/null +++ b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/full.md @@ -0,0 +1,333 @@ +# HopFIR: Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation + +Kai Zhai $^{1*}$ Qiang Nie $^{2*}$ Bo Ouyang $^{1\dagger}$ Xiang Li $^{3}$ Shanlin Yang $^{1}$ $^{1}$ Hefei University of Technology $^{2}$ Youtu Lab, Tencent $^{3}$ Tsinghua University + +# Abstract + +2D-to-3D human pose lifting is fundamental for 3D human pose estimation (HPE), for which graph convolutional networks (GCNs) have proven inherently suitable for modeling the human skeletal topology. However, the current GCN-based 3D HPE methods update the node features by aggregating their neighbors' information without considering the interaction of joints in different joint synergies. Although some studies have proposed importing limb information to learn the movement patterns, the latent synergies among joints, such as maintaining balance are seldom investigated. We propose the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture to tackle the 3D HPE problem. HopFIR mainly consists of a novel hop-wise GraphFormer (HGF) module and an intragroup joint refinement (IJR) module. The HGF module groups the joints by $k$ -hop neighbors and applies a hopwise transformer-like attention mechanism to these groups to discover latent joint synergies. The IJR module leverages the prior limb information for peripheral joint refinement. Extensive experimental results show that HopFIR outperforms the SOTA methods by a large margin, with a mean per-joint position error (MPJPE) on the Human3.6M dataset of $32.67\mathrm{mm}$ . We also demonstrate that the state-of-the-art GCN-based methods can benefit from the proposed hop-wise attention mechanism with a significant improvement in performance: SemGCN [42] and MGCN [49] are improved by $8.9\%$ and $4.5\%$ , respectively. + +# 1. Introduction + +Monocular 3D human pose estimation aims to accurately regress the 3D locations of human joints in the camera coordinate system from a single image. It plays an important role in many applications, such as action recognition and human-computer interaction. Compared with the monocu + +![](images/5bc2f34248cc90ef9114cd7610308005b82d2dfd435c4e252cc1d04f0ffa1cd2.jpg) +Figure 1. Comparison of performance and model size between the proposed HopFIR and SOTA methods, namely Modulated GCN (MGCN) [49], SemGCN [42], Weight Unsharing [18], SR-Net [39], GraphSH [36], and GraFormer [43]. The methods are evaluated on the Human3.6M dataset [10] with ground truth 2D joints as input. The arrow shows the performance improvement obtained by inserting the HA layer into the other networks. Tiny, small, and normal denote the feature dimension of the HopFIR are 32, 64, and 128, respectively. + +lar systems, multi-view capture systems are expensive and inconvenient to set up and operate, which prevents them from being widely used in practice. To tackle the monocular 3D HPE task, some approaches [2, 15, 31, 32, 38] estimates 3D joint coordinates or heat maps directly from an image via a convolutional neural network (CNN) [13, 14]. However, direct regression from the image space suffers from the problem of a large parameter searching space, which always leads to a sub-optimal solution. Recently, Martinez et al. [20] constructed a simple fully connected network using only 2D keypoints as input and achieved promising 3D HPE performance, showing that the 3D human pose can be efficiently and accurately estimated from 2D joint po + +sitions. Inspired by them and considering the thoroughly investigated 2D HPE, many works decompose the problem into two subtasks, i.e., the 2D HPE and 2D-to-3D pose lifting [18, 40, 49, 36]. 2D-to-3D pose lifting, therefore, has emerged as a fundamental task in this area that our work devotes to. + +A key consideration in 2D-to-3D pose lifting is that the human skeleton topology is inherently sparse and graph-structured. Fully connected neural networks are less effective in modeling graph-structured data due to their simple connections among all nodes and the probability of overfitting. To leverage the information of the human skeletal topology, some works [36, 40, 49] have proposed to model the human body with GCNs and have achieved SOTA results. For example, Ci et al. [4] introduced a locally connected network to enhance the representation capability of the GCN, and Liu et al. [18] explored the weight sharing and feature transformation that occurs before or after feature aggregation in the GCN. One limitation of these GCN-based 3D HPE methods, however, is that they update the node features by aggregating their neighbors' information without considering the different contributions of these nodes to different joint synergies. + +Instead of considering all the joints of a skeleton as a whole, Xue et al. [37] demonstrated that the human skeleton exhibits obvious part-wise inconsistency in its motion patterns, as also reported in SRNet [39]. However, these works are limited to considering prior structural information of limb groupings and ignore investigating the latent groups underlying joint synergies. For example, the relative positions of the 1-hop neighbors of joint 0 are almost constant in "Discussion" subject, which can be a latent group, as shown in Fig. 2. Moreover, these works consider the joints in a limb as a whole to calculate the relationship with other limbs, which resulted in lower accuracy for peripheral joints, such as wrists and feet. + +To address the abovementioned problems in monocular 3D HPE, we propose a novel architecture: the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR). The first key component of HopFIR is a novel hop-wise GraphFormer (HGF) module that considers $k$ -hop neighbors. In the HGF module, the information of every hop of every joint is aggregated into the hidden space, such that $N \times k$ groups of features are obtained for a skeleton model with $N$ joints. Meanwhile, a hop-wise transformer-like attention mechanism is designed to extract the correlation among feature groups, which computes similarity by the dot product of the node feature and the group feature. The proposed HGF module enables the network to discover latent joint interactions considering human joint synergy. Because the HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interactions of peripheral joints associated with + +a limb, we introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information. Specifically, a residual block is built from two HGF modules followed by one IJR module. The proposed HopFIR architecture achieves optimal regression accuracy with a stack of three blocks. + +To summarize, our work makes the following contributions: + +- To the best of our knowledge, we design the first Hopwise GraphFormer module to explore potential joint correlations underlying human joint synergy. We also prove that other GCN-based methods can benefit from the proposed HGF module efficiently, as shown in Fig. 1. +- We design an Intragroup Joint Refinement module, which attends to intragroup joints to refine joint features through the associated limb, especially the wrists and feet. The IJR module enables HGF modules to discover the latent synergies among joints. +- We propose the novel Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture for 3D HPE, which is built entirely from HGF and IJR modules. Specifically, two HGF modules and one IJR module are coupled into a block. +- Extensive experiments demonstrate the effectiveness and generalizability of the proposed modules and HopFIR architecture by providing new state-of-the-art results on two challenging datasets, i.e., Human3.6M [10] and MPI-INF-3DHP [21]. + +# 2. Related Work + +3D Human Pose Estimation. Early works [26, 29] use handcrafted features, perspective relationships, and geometric constraints to estimate the 3D human pose. Recent pose estimation approaches can be generally divided into two categories. The first category of networks regresses 3D human joints directly from the image [24, 46]. Pavlakos et al. [24] adopted a CNN to predict the voxel-wise likelihoods for each joint, and Zhou et al. [46] directly embedded a kinematic object model into the networks to learn the general multi-articulate object pose. Approaches in the second category decouple the 3D HPE task into 2D pose estimation from an image and 3D pose estimation from the detected 2D joints (2D-to-3D). For example, Martinez et al. [20] proposed a simple yet effective baseline with fully-connect networks and proved that 3D human poses can be regressed simply and effectively from 2D keypoints. Our paper follows this pipeline and focuses on the 2D-to-3D pose lifting. For promoting 3D human pose regression accuracy, it is crucial to group joints with consideration of their interactions rather than treating all joints of a skeleton as a whole. Xue et al. [37] divided the human skeleton graph into five groups according to the limbs to explore part-wise motion inconsistency, and Zeng et al. [39] split the human joints into + +local regions and recombined the global information from the rest of the joints. Our proposed HopFIR differs from these approaches by grouping joints by the $k$ -hop neighbors of each joint and prior limb information, which enables the network to discover latent connections between groups in different human joint synergies. + +Graph Convolutional Networks. GCNs [5, 7, 12, 27] generalize the capability of CNNs by performing convolution operations on graph-structured data. GCNs can be divided into two categories: the spectral-based approaches [5] and the spatial-based approaches [12]. Our approach falls into the second category, which applies message-passing operations on the graph nodes and their neighbors. + +Due to the graph-structure topology of the human skeleton, many works [4, 18] have introduced GCN to tackle the 3D HPE task. Zhao et al. [42] proposed a SemGCN to learn the semantic relationships between human joints, and Zou et al. [49] proposed a weight modulation and an affinity modulation based on the SemGCN. These methods aggregate the first-order neighborhood messages to update the feature matrix by assigning different weights to different nodes. Some works [47, 48] have extended the first-order neighbors to high-order neighbors in the spatial domain directly. Zeng et al. [40] designed a hierarchical fusion block by dividing the fusion procedure into two stages, where all the high-order neighbors of a node are aggregated into a feature in the first stage and fuse it with the node feature and the first-order neighbor in the second stage. Zhao et al. [43] introduced Chebyshev graph convolution to fuse information among the $k$ -hop neighbors of a joint directly. These works updated features by aggregating each node's own $k$ -hop neighborhood information in a GCN layer. However, HopFIR considers the $k$ -hop groups of all nodes to reconstruct the $k$ -hop feature of a node through the proposed attention mechanism, which can enhance the representation capability of GCNs. + +Graph Attention. Graph attention networks [34] is a pioneer work that pay attention to the data in a graph structure by assigning an attention weight to each node. The introduction of the transformer [33] for machine translation tasks has proven the capacity of attention for sequential input. ViT [6] introduces the transformer into computer vision and achieves excellent performance. Inspired by them, some researchers have adopted the transformer for 3D HPE. Zhao et al. [43] applied self-attention to capture global information by calculating the similarity of all nodes. PoseFormer [44] directly applies Transformer Encoder (TE) by viewing each joint and each frame as tokens in the spatial and temporal domains, respectively. MixSTE [41] is based on [44], where each joint feature is represented by the temporal TE to model the joint motion. P-stmo [28] replaces the spatial TE with MLP and applies Stride Transformer Encoder to map $N$ frames to one frame. MHFormer [17] generates + +![](images/cb670ec8dfb645aeddec880737bf4316cadc122c63e7044bff95e1e5ac76c551.jpg) +(a) + +![](images/4b3d9bfec6aa4c9b4191bc1f4e7059caf8f7eece3beb678b2e70b8085e238d37.jpg) +(b) + +![](images/2c8971dbdb6be3b194413e8c566fedf1c02c2dc63ffd51ebaa72088d8bd92efe.jpg) +(c) +Figure 2. An illustration of the human skeleton graph and the groups in HGF and IJR modules. The $k$ -hop neighbors of a joint are set as a group in HGF modules where (a) indicates the 1-hop groups of different joints and (b) indicates several latent groups because of physical limitations. (c) indicates the joints grouped by prior limb information. + +multi-hypothesis at different depths of stacked spatial TEs by viewing all frames of each joint feature as a token, then communicates these hypotheses with cross-attention and self-attention. Different from the existing graph attention for 3D HPE, we propose the intergroup multi-head attention mechanism among the $k$ -hop groups of all nodes, which assigns the attention weights by computing the similarity between the node feature and $k$ -hop group feature. Moreover, we introduce the intragroup multi-head self-attention in limb groups to refine the joint features and promote the HGF module to discover the latent synergies among joints. + +# 3. The Proposed HopFIR + +This paper proposes a novel architecture to regress the 3D human pose from $N$ given 2D keypoints $X \in \mathbb{R}^{N \times 2}$ . The proposed framework mainly consists of the HGF and IJR modules. In this section, we first review the vanilla Graph Convolutional Network and Transformer in Sec. 3.1. We then introduce the HGF and IJR modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we present the network architecture in Sec. 3.4. + +# 3.1. Vanilla GCN and Transformer + +GCN. A graph is defined as $\mathcal{G} = (\mathcal{V},\mathcal{E})$ , where $\mathcal{V}$ is a set of $N$ nodes and $\mathcal{E}$ is the adjacency matrix representing the edges between the nodes. Given a collection of input features $H^{l}\in \mathbb{R}^{N\times D}$ , a generic GCN layer that aggregates neighborhood information can be formulated as follows: + +$$ +H ^ {(l + 1)} = \sigma (\tilde {A} H ^ {l} W) \tag {1} +$$ + +where $W\in \mathbb{R}^{D\times D^{\prime}}$ is the learnable weight matrix that transforms the feature dimension from $D$ to $D^{\prime}$ , and $\sigma (.)$ is the activation function, such as ReLU [22]. $H^{(l + 1)}$ is the updated feature matrix, $\tilde{A}\in \mathbb{R}^{N\times N}$ is the symmetrically normalized affinity matrix [12] with added selfconnections, and $A\in \{0,1\}^{N\times N}$ is the adjacency ma + +![](images/b73bf7f1727b044f6557a45cfa24667f94059486ae1ba38047561c596cfe3203.jpg) +Figure 3. The HopFIR architecture, with details of the HGF and IJR modules. The designed block is a residual block [9] built by two HGF modules followed by one IJR module. The proposed architecture with three blocks achieves optimal performance. Arrows of different colors represent different hops and groups in HGF and IJR, respectively. + +trix. The $(i,j)th$ entry $a_{ij} = 1$ representing node $j$ is the neighbor of node $i$ . Otherwise, they are not connected and $a_{ij} = 0$ . Therefore, the none-neighbor nodes have a weak influence on each other in the vanilla GCN, which hinders the modeling of underlying joint synergy in 3D HPE. + +Transformer. Transformer architecture relies entirely on self-attention to compute representations of its input and output. The self-attention function maps the inputs to the queries $Q$ , keys $K$ , and values $V$ by weight matrices $W_{Q}$ , $W_{K}$ , and $W_{V}$ , respectively, and the matrix of outputs is calculated as: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(Q K ^ {T} / \sqrt {d}\right) V \tag {2} +$$ + +where $d$ is the feature dimension of $Q$ , and $\frac{1}{\sqrt{d}}$ is a scaling factor to prevent extremely small gradients. The multihead self-attention (MHSA), which performs self-attention in parallel, projects the queries, keys, and values $P$ times with different linear projections to the respective subspaces, as follows: + +$$ +\operatorname {M u l t i H e a d} (Q, K, V) = \operatorname {C o n c a t} \left(D _ {1}, \dots , D _ {P}\right) W _ {o} \tag {3} +$$ + +where $W_O$ is the projection matrix of outputs, $D_p = \text{Attention}(QW_Q^p, KW_K^p, VW_V^p)$ , and $p \in [1,..,P]$ . + +# 3.2. Hop-wise GraphFormer + +Previous GCN studies for 3D HPE aggregate multi-hop neighborhood information [40, 48] or assign an attention weight to each first-order neighbor [42, 49]. To effectively capture the node's neighborhood message and increase the representational capacity of GCNs, we introduce the hopwise GraphFormer (HGF) module, which treats each hop as a group (Fig.2 a,b) and computes the attention weights for each hop (more intuitive descriptions for $k$ -hop can be found in the supplementary material). By considering the relationship within $k$ hops, we can obtain $N \times k$ groups for a + +skeleton graph with $N$ joints, which provides enough combinations of joints to discover the latent correlations among joints in different human joint synergy. + +We first define the $k$ -hop matrix $A^k$ as + +$$ +a _ {i j} ^ {k} = \left\{ \begin{array}{l l} 1, & d \left(v _ {i}, v _ {j}\right) = k \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +where $d(v_{i}, v_{j})$ denotes the distance of the shortest path between $v_{i}$ and $v_{j}$ on the skeleton graph. The $k$ -hop neighborhood information is aggregated with a weighted sum of the target node's $k$ -hop neighbors, named HopGCN: + +$$ +s _ {i} ^ {k} = \sum_ {j} a _ {i j} ^ {k} h _ {j} W ^ {k} \tag {5} +$$ + +where $s_i^k \in \mathbb{R}^D$ is a hidden representation of the $k$ -hop neighborhood information. Eq. 5 is similar to Eq. 1 but with the extended definition of $A$ to $k$ hops. The weight matrix $W$ is assigned to the respective hops. + +Before aggregating the hidden representation $s_i^k$ to the target node, we propose a transformer-like attention mechanism computing the similarity between the node feature $h_i$ and the $k$ -hop hidden representation, as shown in Eq. 6: + +$$ +z _ {i} ^ {k} = \sum_ {j} \operatorname {s o f t m a x} _ {j} \left(\frac {h _ {i} s _ {j} ^ {k T}}{\sqrt {d}}\right) s _ {j} ^ {k}. \tag {6} +$$ + +Due to the structure of the human skeleton and the joint synergy, the $k$ -hop neighborhood of joint $i$ is related to that of other joints. In Fig. 2, for example, the synergy of hop $S_5^1$ is related to hop $S_6^1$ and $S_3^1$ . By packing $z_i^k$ , $h_i$ , and $s_i^k$ into matrices $Z^k$ , $H$ , and $S^k$ , we can calculate $Z^k$ in parallel: + +$$ +Z ^ {k} = \operatorname {s o f t m a x} \left(\frac {H S ^ {k ^ {T}}}{\sqrt {d}}\right) S ^ {k} \tag {7} +$$ + +$Z^k$ can also be reconstructed using the self-attention mechanism purely on hop features or swap the positions of $H$ and + +$S^k$ . A performance comparison of these cases is discussed in the experimental results. + +We further reduce the dimension of hidden representation by using a fully connected(FC) layer based on the order of the neighborhood, considering different amounts of information related to the target node, as formulated in Eq. 8. + +$$ +r _ {i} ^ {k} = F ^ {k} \left(z _ {i} ^ {k}\right) \tag {8} +$$ + +where $r_i^k$ is the refined representation of the $k$ -hop neighborhood information with respect to node $i$ , and $F^k$ is the mapping function. The refined representation is then concatenated to the updated node feature. + +All the refined representations of hop-wise neighborhoods and the target node feature $h_i$ are encoded to a $D$ -dimensional vector: + +$$ +h _ {i} ^ {\prime} = F \left(h _ {i}, r _ {i} ^ {1}, r _ {i} ^ {2}, r _ {i} ^ {3}, \dots , r _ {i} ^ {k}\right) \tag {9} +$$ + +where $h_i'$ is the final updated feature of node $i$ in this layer, and $F$ is the aggregation function. Packing together the updated features of all nodes, Eq. 9 can be rewritten as: + +$$ +H ^ {\prime} = F \left(H, R ^ {1}, R ^ {2}, R ^ {3}, \dots , R ^ {k}\right). \tag {10} +$$ + +We thereby obtain as the core layer of the HGF module the hop-wise attention (HA) layer, which extracts latent correlations between feature groups and aggregates $k$ -hop neighborhood information. With the increment of $k$ , we get more $N$ groups on which to explore the underlying joint synergies. The value of $k$ can be adjusted based on the tasks and pipelines. According to the experimental results, three hops achieve optimal performance in the HopFIR architecture. + +To fuse the current global information and original 2D information, we concatenate all the joint features in a batch and feed them into a FC layer to extract the current global information. We then concatenate the extracted global information, 2D information, and the output of the HA layer on the feature dimension of the joint feature to fuse all the information in another FC layer. + +# 3.3. Intragroup Joint Refinement + +The proposed HGF module splits the skeleton graph into different groups based on the $k$ -hop neighborhood of each joint, which attends to the key groups in joint synergies. However, HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interaction of the peripheral joints associated with limbs, such as the wrists and feet. We introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information, as shown in Fig. 2(c). The HGF features of the joints in each limb group are refined in the IJR module by using multi-head self-attention [33], as in Eq. 11. + +$$ +H _ {g} = M H S A \left(M H S A (H) _ {g}\right) \tag {11} +$$ + +where $H$ is the feature matrix from the HGF module, $MHSA(H)_g$ is the feature matrix of the group $g$ updated by global multi-head self-attention, and $H_{g}$ is the final feature matrix of the group $g$ updated by the IJR module. More details of the IJR module are provided in Fig. 3. + +# 3.4. The HopFIR Architecture + +The HopFIR architecture consists of the proposed HGF and IJR modules, as illustrated in Fig. 3. The residual block, which contains two HGF modules and one IJR module, is designed as the basic block in HopFIR. Moreover, we define a linear embedding layer to map the input to the latent space and a HG module to transform the output into 3D space. The HG module is a variant of HGF designed for the output layer, more details of which are provided in the supplementary material. HopFIR accepts 2D keypoints as input, which can be obtained via an off-the-shelf 2D detector. The graph is obtained by adding a normalized globally learnable $k$ -hop graph to the skeleton graph and then symmetrically normalizing it, as in [49]. We use the L1-norm loss and L2-norm loss to compute the error between ground truth and prediction with the weighted sum as follows: + +$$ +L = \alpha \sum_ {n = 1} ^ {N} \| Y _ {n} - \hat {Y _ {n}} \| _ {2} + \beta \sum_ {n = 1} ^ {N} \| Y _ {n} - \hat {Y _ {n}} \| _ {1} \tag {12} +$$ + +where $N$ is the joint number, $\hat{Y}_n$ is the predicted 3D position of joint $n$ , $Y_n$ is the ground truth, $\alpha = 1$ , and $\beta = 0.1$ . + +# 4. Experiments + +In this section, we first introduce the experimental setup and implementation details of the HopFIR networks. We then present our experimental results and comparisons with state-of-the-art methods. Finally, we conduct several ablation studies of the proposed architecture. + +# 4.1. Datasets and Evaluation Protocols + +Datasets. Human3.6M [10] is currently the largest publicly available dataset for 3D human pose estimation, with 3.6 million video frames. It captures accurate 3D human joint positions from four camera viewpoints and records 11 subjects performing 15 assigned actions. Following previous works [49, 36, 42], we train our model on five subjects (S1, S5, S6, S7, S8) and test it on two subjects (S9, S11). In contrast to Human3.6M, MPI-INF-3DHP [21] includes complex outdoor scenes, which are commonly used to evaluate the generalizability of proposed methods. Accordingly, we use the test set of MPI-INF-3DHP to verify the generalizability of our model. + +Evaluation Protocols. For Human3.6M [10], we evaluate our model on two standard evaluation protocols: the mean per-joint position error (MPJPE) and the mean per-joint position error after Procrustes alignment (P-MPJPE). + +
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Martinez et al. [20]ICCV201751.856.258.159.069.578.455.258.174.094.662.359.165.149.552.462.9
Zhao et al. [42](†)CVPR201947.360.751.460.561.149.947.368.186.255.067.861.042.160.645.357.6
Ci et al. [4](†)ICCV201946.852.344.750.452.968.949.646.460.278.951.250.054.840.443.352.7
Pavlo et al. [25]CVPR201947.150.649.051.853.661.449.447.459.367.452.449.555.339.542.751.8
Cai et al. [2](†)ICCV201946.548.847.650.952.961.348.345.859.264.451.248.453.539.241.250.6
Liu et al. [18](†)ECCV202046.352.247.350.755.567.149.246.060.471.151.550.154.540.343.752.4
zeng et al. [39]ECCV202044.548.247.147.851.256.850.145.659.966.452.145.354.239.140.349.9
Zou et al. [49](†)ICCV202145.449.245.749.450.458.247.946.057.563.049.746.652.238.940.849.4
Xu et al. [36](†)CVPR202145.249.947.550.954.966.148.546.359.771.551.448.653.939.944.151.9
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202245.250.848.050.054.965.048.247.160.270.051.648.754.139.743.151.8
Ours(Δ)(†)43.947.645.548.950.158.046.244.555.762.949.045.851.838.039.948.5
+ +Table 1. Quantitative comparison on Human3.6M with detected 2D poses as input under Protocol #1, in millimeters. The best results are highlighted in bold and the second-best results are underlined. $(\dagger)$ indicates GCN-based methods and $(\Delta)$ indicates Transformer-based methods. + +
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Zhou et al. [45](+)ICCV201934.442.436.642.138.239.834.740.245.660.839.042.642.029.831.739.9
Ci et al. [4](+)(*)†ICCV201936.338.829.737.834.642.539.832.536.239.534.438.438.231.334.236.3
Martinez et al. [20]ICCV201737.744.440.342.148.254.944.442.154.658.045.146.447.636.440.445.5
Zhao et al. [42](†)CVPR201937.849.437.640.945.141.440.148.350.142.253.544.340.547.339.043.8
Cai et al. [2](†)ICCV201933.439.033.837.038.147.339.537.343.246.237.738.038.630.432.138.1
Liu et al. [18](†)ECCV202036.840.333.036.337.545.039.734.940.347.737.438.538.629.632.037.8
zeng et al. [39]ECCV202035.936.729.334.536.042.837.731.740.144.335.837.236.233.734.036.4
Zou et al. [49](†)ICCV2021---------------37.4
Xu et al. [36](†)CVPR202135.838.131.035.335.843.237.331.738.445.535.436.736.827.930.735.8
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202232.038.030.434.434.743.335.231.438.046.234.235.736.127.430.635.2
Ours(Δ)(†)31.334.028.032.033.142.134.128.133.639.831.732.933.826.728.932.7
+ +Table 2. Quantitative comparison on Human3.6M with ground truth 2D keypoints as input under Protocol #1, in millimeters. (+) uses additional data from MPII [1]. (*) uses pose scales in both training and testing. The best results are highlighted in bold. + +These are referred to Protocol #1 and Protocol #2, respectively. The MPJPE and P-MPJPE are given in millimeters. For MPI-INF-3DHP [21], we follow previous works [4, 42, 49] by reporting the percentage of correct keypoints (PCK) with a threshold of $150\mathrm{mm}$ and the area under the curve (AUC) for a range of PCK thresholds. + +# 4.2. Implementation Details + +Following previous work [25], we obtain the detected 2D poses using the cascaded pyramid network(CPN) [3]. We do not use data augmentation during training and testing with the 2D ground truth input to verify the efficacy of our model. We adopt Adam [11] optimizer and all experiments are conducted on a single NVIDIA RTX 3090 GPU. 3D pose regression from 2D detections is more challenging than that from 2D ground truth because the former needs to deal with some extra uncertainty in the 2D space. To manage this uncertainty, we set different configurations for them. In experiments with 2D ground truth as the input, we train the HopFIR networks with an initial learning rate of 0.001, a decay factor of 0.90 per 4 epochs, a batch size of 64, channels of 128, and PReLU activation [8]. When using detected 2D poses, we train the HopFIR networks with the initial learning rate of 0.006, a decay factor of 0.95 per 4 epochs (but 0.2 for the first 4 epochs), a batch size of 256, 256 channels, and LeaklyReLU activation [19]. To avoid overfitting, we apply Dropout[30] with a dropout rate of 0.5. + +# 4.3. Comparison with State-of-the-art + +We compare the performance of HopFIR with some SOTA methods on Human3.6M under Protocol #1 and Protocol #2, with the results shown in Table 1. Our method reaches an MPJPE of $48.50\mathrm{mm}$ and outperforms the best of the existing approaches [49] on all 15 actions. Given the uncertainty of 2D detections, we also investigate the capability of HopFIR networks using ground truth 2D key points as input. As shown in Table 2, HopFIR obtains surprisingly better performance when given precise 2D joint information and produces SOTA results, which verifies its effectiveness. + +# 4.4. Ablation Study + +We conduct a comprehensive ablation study on Human3.6M to validate the individual effectiveness of each component of the proposed HopFIR architecture under controlled settings. We follow previous works [18, 36, 49] conducting the ablation experiments using GT as inputs to avoid the influence of the 2D pose detector. + +Effectiveness of Different Modules. We separately verify the effectiveness of the HGF module and IJR module and conduct experiments on first-order neighbors, removing all modules and the multi-hop mechanism. Note that GCN and HopGCN refer to applying only vanilla GCN (Eq. 1) and HopGCN (Eq. 5) to Fig. 3, respectively. The results in Table 3 show that each module improves the performance + +
MethodChannelsParamsMPJPEP-MPJPE
GCN1280.36M40.6331.65
HopGCN1280.59M39.1531.40
HopGCN & IJR1282.15M36.6229.23
HopGCN & HGF1281.05M35.1928.81
HopGCN & IJR640.57M36.6929.68
HopGCN & HGF800.48M36.0129.58
HopGCN & HopFIR1282.15M32.6726.20
HopGCN & HopFIR640.54M33.5227.37
HopGCN & HopFIR320.14M35.1928.71
+ +over the GCN-only approach, and coupling the two layers to form a HopFIR block achieves further performance improvement. The HopFIR networks reduce the MPJPE to $32.67\mathrm{mm}$ , which represents a $7.2\%$ improvement over GraFormer [43]. By reducing feature channels in HopGCN & IJR and HopGCN & HGF, we decrease parameters to 0.57M and 0.48M. However, the models still achieved errors of $36.69\mathrm{mm}$ and $36.01\mathrm{mm}$ , compared to the HopGCN with $39.15\mathrm{mm}$ error and 0.59M parameters, which should be attributed to the structure design other than the model size. Moreover, we reduce the HopFIR network parameters by changing the channels to 64 and 32, respectively, which are also superior to SOTA. + +Effectiveness of the HA Layer. In sec. 3.2, we introduce $k$ -hop groups to discover latent joint interactions in human joint synergies. The attention matrices in Fig. 5 show the latent joint synergies captured by HopFIR, in which each weight of row $i$ indicates a discovered latent group for the corresponding joint $i$ . To verify the effectiveness of HA layer, we explore the correlation between all joints with a transformer encoder instead of HA layer, which explores the correlation between individual nodes but ignores the synergy between groups of joints in the human body and obtained $36.21 \mathrm{~mm}$ error. Moreover, we remove the human body prior by using random graph instead of skeleton graph without changing the HopFIR architecture and reached $34.68 \mathrm{~mm}$ error, suggesting that latent group correlations can be explored by $k$ -hop groups, but group correlations underlying joint synergies can be better explored based on the human body prior. + +In Table 4, we show the experimental results of three different ways to design the HA layer. HSS is the method selected in this paper, where H, S, and S are $Q$ , $K$ , and $V$ , and H and S represent the node feature and the $k$ -hop group feature, respectively. As we do not follow [33] in applying a linear transformation of $Q$ , $K$ , and $V$ , we also show the result of such a linear transformation, which is denoted as HA+W. The experiment results show that HSS similarity achieves better performance in the HopFIR architecture, but one can choose the type of similarity according to the network property. + +We further insert the HA layer into SOTA pose esti + +Table 3. Ablation experiments on the proposed modules. + +
AttentionHAHA+W
HSSSSSSHHHSSSSSSHH
Params2.15M2.50M
MPJPE32.6734.1834.2833.2933.5333.90
P-MPJPE26.2027.7027.7127.1627.4027.41
+ +Table 4. Quantitative comparison of HA layers with different similarity computing approaches. + +
MethodChannelsParamsMPJPEP-MPJPE
SemGCN [42]1280.27M42.1433.53
SemGCN + HA(HSS)1280.49M38.4130.56
SemGCN + HA(SSS)1280.49M41.3033.07
SemGCN + HA(SHH)1280.49M38.8131.05
SemGCN [42] w/ Non-local [35]1280.43M40.7831.46
SemGCN w/ Non-local +HA(HSS)1280.66M38.0330.50
SemGCN w/ Non-local +HA(SSS)1280.66M37.7530.17
SemGCN w/ Non-local +HA(SHH)1280.66M37.9429.71
Modulated GCN [49]1280.29M38.2530.06
Modulated GCN +HA(HSS)+W1280.96M36.5429.09
Modulated GCN +HA(SSS)+W1280.96M36.1429.02
Modulated GCN +HA(SHH)+W1280.96M37.3830.02
+ +Table 5. Comparison of the improved performance of proposed HA layer added on different methods. We test on two GCN-based methods: SemGCN [42] and MGCN [49]. + +
Num-kChannelsParamsMPJPEP-MPJPE
11281.88M35.8828.76
21282.03M34.9627.74
31282.15M32.6726.20
41282.27M35.5828.35
Num-BlockChannelsParamsMPJPEP-MPJPE
11280.80M37.3330.71
21281.47M34.2127.63
31282.15M32.6726.20
41282.82M33.8427.63
+ +Table 6. Ablation study for number of $k$ -hop and designed blocks. The units of MPJPE and P-MPJPE are millimeters (mm). + +mation methods, namely SemGCN [42] and Modulated GCN [49], to investigate its generalizability. No changes are made to their source code, with the HA layer inserted before the information aggregation stage of these methods. The experimental results in Table 5 show that the HA layer improves these previous SOTA networks to a large degree; especially, the MPJPE of SemGCN [42], is reduced from $42.14\mathrm{mm}$ to $38.41\mathrm{mm}$ , representing an $8.9\%$ improvement. Moreover, HA layer with linear transformation (W) makes learning more stable, so We test HA and $\mathrm{HA + W}$ in SemGCN and MGCN, respectively, to show both of them are effective. For a fair comparison of parameters, we tested MLP and MHSA instead of HA, both of them with 1.03M parameters, and achieved $39.12\mathrm{mm}$ and $39.22\mathrm{mm}$ errors, respectively. More details can be found in the supplementary material. Experiments on the above methods of aggregating first-order neighbor information demonstrate the effectiveness of HA and also indicate that latent joint grouping can recognize the human joint synergies. + +Error on Peripheral Joints. We report the regres + +![](images/3c5a5aa59debb94a5f73b5d1bef79d5835e3620abd2cc57b6041fcf28716a763.jpg) +Figure 4. Comparison of MPJPE for peripheral joints on the test set of the Human3.6M. R and L denote right and left, respectively. + +
BlockChannelsParamsMPJPEP-MPJPE
(H)(I)1281.78M34.4327.79
(I)(H)1281.78M34.1327.56
(H)(I)(H)1282.15M35.0827.98
(I)(H)(H)1282.15M34.2027.60
(H)(H)(I)1282.15M32.6726.20
+ +Table 7. Ablation study for arrangements of the designed block. (I) and (H) denote IJR module and HGF module, respectively. + +
MPJPEP-MPJPEMPJVE
PoseFormer [44] (T=81)44.336.53.1
MixSTE [41] (T=243)40.932.62.3
MHFormer [17] (T=351)43.0--
P-STMO [28] (T=243)42.134.4-
Ours (T=243)41.132.52.1
+ +Table 8. Quantitative comparison on Human3.6M with detected 2D pose (CPN) in video. $T$ denotes the number of input frames. + +
MethodsPCKAUC
GSno GSOutdoorAll
Martinez et al. [20]49.842.531.242.517.0
Ci et al. [4]74.870.877.374.036.7
zeng et al. [39]--80.377.643.8
Li et al. [16]70.168.266.666.9-
Zhao et al. [43]80.177.974.179.043.8
Liu et al. [18] (weight unsharing)77.680.580.179.347.6
Xu et al. [36]81.581.775.280.145.8
Nie et al. [23]---83.545.9
Zou et al. [49]86.486.085.786.153.7
Ours89.185.985.987.257.0
+ +Table 9. Quantitative comparisons on the MPI-INF-3DHP test set. GS denotes green screen. + +sion accuracy for the peripheral joints (wrist and feet) in Fig. 4 in comparison with some previously proposed methods [49, 42, 39]. The HopFIR network with IJR modules outperforms SOTA methods on the right foot (RFoot) by 6.7 mm, left foot (LFoot) by 7.7 mm, left wrist (LWrist) by 9.6 mm, and right wrist (RWrist) by 7.7 mm. The experimental results verify that the intragroup joint attention within each limb group strengthens the capabilities of the HopFIR. + +Different Numbers of $k$ -Hops. As HopFIR is designed + +![](images/30a7532c27391266c82b4ff8f610415ae589b0ed161638b9c984660f9f9d0763.jpg) +Figure 5. Attention weight of the $j$ -th $k$ -hop for the $i$ -th joint, deeper color indicates higher correlation. $i$ -th row and $j$ -th col represent $i$ -th joint and $k$ -hop of $j$ -th joint, respectively. + +![](images/6a6921bff2b26e547624cd71f09eaec86ad8f3791dba539140d71055830ec885.jpg) +Figure 6. Qualitative visual results for HopFIR and MGCN [49] on the Human3.6M. The black lines are the ground truth (GT) and the red lines are the predictions of HopFIR and MGCN. Wrong predictions are circled. The bottom row shows our failure case. + +to extract the correlation between feature groups, we set different $k$ values to discover various latent connections underlying the human joint synergies, with the results shown at the top of Table 6. The MPJPE gradually decreases as the number of hops increases, and reaches the best performance at 3 hops. Therefore, the optimal number of hops for 3D HPE is 3, which entails that we obtain $16 \times 3$ groups from the skeleton graph. Each of the 16 groups corresponds to a potential correlation among coupled nodes at different distances, and three hops is sufficient to recognize the joint synergies. + +Arrangement of the Designed Block. To investigate the optimal structure of the designed block, experiments are performed with various block numbers and various combinations of HGF and IJR modules. As shown at the bottom of Table 6, the error gradually decreases as the number of blocks increases, until the best performance is achieved at 3 blocks. As shown in Table 7, the (H)(H)(I) arrangement achieves the optimal results by reducing the error to 32.67 mm. HGF treats each hop as a group and applies a hop + +wise attention mechanism to these groups to discover latent joint synergy. IJR utilizes the limb prior for peripheral joint refinement. Thus, (H)(H)(I) first integrates the complete joint information and then refines it by IJR. While (I)(H)(H) reverses this procedure resulting in the insufficient utilization of joint information. That the results are superior to GraFormer [43] in all cases except for the single-block model indicates that the HopFIR has significant human pose representation capabilities. + +Extend to Temporal Domain. Without a specific design to integrate temporal information, we extend to the temporal domain by adding two TEs after each block and replacing the HG module with a linear layer, and achieved competitive results as shown in table 8. + +Cross-Dataset Results on MPI-INF-3DHP. Table 9 further compares HopFIR with previous methods on cross-dataset scenarios to validate its generalizability. For these experiments, we train our model on the Human3.6M dataset and test it on the test set of the MPI-INF-3DHP dataset. The results show that our approach obtains better results than other methods, which verifies the generalizability of our approach to unseen scenarios. + +Qualitative Results. In Fig. 6, we show the visual results on Human3.6M in the world space. The bottom of the figure shows some failure cases of HopFIR, which predict some wrong joint positions. The figure shows that HopFIR is able to predict 3D joint positions more accurately, even for poses that cause difficulties for MGCN. + +# 5. Conclusions + +We present the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) as a novel architecture for 3D human pose estimation. The proposed architecture mainly comprises the HGF and IJR modules. The HGF module improves on the GCN-based pose estimation networks by grouping the joints by $k$ -hop neighborhood and capturing the potential joint correlations in the different joint synergies. Because the peripheral joints strongly interact with intra-limb joints, the proposed IJR module applies intragroup attention to refine the peripheral joint features through the associated limb. The proposed method achieves new state-of-the-art results while maintaining a modest model size. + +# Acknowledgement + +This work was supported in part by the National Key Research and Development Program of China (Grant No. 2021YFC0122602), in part by the Joint Funds Program of the National Natural Science Foundation of China (Grant No. U21A20517), in part by the Basic Science Centre Program of National Natural Science Foundation of China (Grant No. 72188101). + +# References + +[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 6 +[2] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2272-2281, 2019. 1, 6 +[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 6 +[4] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019. 2, 3, 6, 8 +[5] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[7] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., 2:729-734 vol. 2, 2005. 3 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1, 2, 5 +[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[12] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3 + +[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017. 1 +[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1 +[15] Kyoungoh Lee, Inwooong Lee, and Sanghoon Lee. Propagating lstm: 3d pose estimation based on joint interdependency. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 1 +[16] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 8 +[17] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 3, 8 +[18] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In European Conference on Computer Vision, pages 318-334. Springer, 2020. 1, 2, 3, 6, 8 +[19] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, volume 30, page 3. Atlanta, Georgia, USA, 2013. 6 +[20] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1, 2, 6, 8 +[21] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 2, 5, 6 +[22] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010. 3 +[23] Qiang Nie, Ziwei Liu, and Yunhui Liu. Lifting 2d human pose to 3d with domain adapted 3d body concept. International Journal of Computer Vision, 131(5):1250-1268, 2023. 8 +[24] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2 +[25] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7753-7762, 2019. 6 +[26] Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Reconstructing 3d human pose from 2d image landmarks. In + +European conference on computer vision, pages 573-586. Springer, 2012. 2 +[27] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008. 3 +[28] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 461-478. Springer, 2022. 3, 8 +[29] Cristian Sminchisescu. 3d human motion analysis in monocular video: techniques and challenges. In Human motion, pages 185-211. Springer, 2008. 2 +[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958, 2014. 6 +[31] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 1 +[32] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 1 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 7 +[34] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 3 +[35] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 7 +[36] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16105-16114, 2021. 1, 2, 5, 6, 8 +[37] Youze Xue, Jiansheng Chen, Xiangming Gu, Huimin Ma, and Hongbing Ma. Boosting monocular 3d human pose estimation with part aware attention. IEEE Transactions on Image Processing, 31:4278-4291, 2022. 2 +[38] Wei Yang, Wanli Ouyang, Xiaolong Wang, Jimmy Ren, Hongsheng Li, and Xiaogang Wang. 3d human pose estimation in the wild by adversarial learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5255-5264, 2018. 1 +[39] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Snet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In European Conference on Computer Vision, pages 507-523. Springer, 2020. 1, 2, 6, 8 + +[40] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11436-11445, 2021. 2, 3, 4 +[41] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13232-13242, 2022. 3, 8 +[42] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3425-3435, 2019. 1, 3, 4, 5, 6, 7, 8 +[43] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 1, 3, 6, 7, 8, 9 +[44] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 3, 8 +[45] Kun Zhou, Xiaoguang Han, Nianjuan Jiang, Kui Jia, and Jiangbo Lu. Hemlets pose: Learning part-centric heatmap triplets for accurate 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2344-2353, 2019. 6 +[46] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In Proceedings of the IEEE International Conference on Computer Vision, pages 398-407, 2017. 2 +[47] Yiran Zhu, Xing Xu, Fumin Shen, Yanli Ji, Lianli Gao, and Heng Tao Shen. PosegTac: Graph transformer encoder-decoder with atrous convolution for 3d human pose estimation. In IJCAI, pages 1359-1365, 2021. 3 +[48] Zhiming Zou, Kenkun Liu, Le Wang 0003, and Wei Tang. High-order graph convolutional networks for 3d human pose estimation. In BMVC, 2020. 3, 4 +[49] Zhiming Zou and Wei Tang. Modulated graph convolutional network for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11477-11487, 2021. 1, 2, 3, 4, 5, 6, 7, 8 \ No newline at end of file diff --git a/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/images.zip b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc98e45fb21d9bed141fd59333c2ea56f7bcb01b --- /dev/null +++ b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6655cbc8a86c8cd340ff42847f88b57d2b85ad67425406479046ad95c43057ed +size 640610 diff --git a/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/layout.json b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4bdd927673b5f117e750144d57783f72c4690dee --- /dev/null +++ b/2023/HopFIR_ Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation/layout.json @@ -0,0 +1,10086 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 138 + ], + "type": "text", + "content": "HopFIR: Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": "Kai Zhai" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": " Qiang Nie" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": " Bo Ouyang" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": " Xiang Li" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": " Shanlin Yang" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": "Hefei University of Technology " + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": "Youtu Lab, Tencent " + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 160, + 490, + 192 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 228, + 192, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 228, + 192, + 242 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 192, + 242 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "text", + "content": "2D-to-3D human pose lifting is fundamental for 3D human pose estimation (HPE), for which graph convolutional networks (GCNs) have proven inherently suitable for modeling the human skeletal topology. However, the current GCN-based 3D HPE methods update the node features by aggregating their neighbors' information without considering the interaction of joints in different joint synergies. Although some studies have proposed importing limb information to learn the movement patterns, the latent synergies among joints, such as maintaining balance are seldom investigated. We propose the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture to tackle the 3D HPE problem. HopFIR mainly consists of a novel hop-wise GraphFormer (HGF) module and an intragroup joint refinement (IJR) module. The HGF module groups the joints by " + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "text", + "content": "-hop neighbors and applies a hopwise transformer-like attention mechanism to these groups to discover latent joint synergies. The IJR module leverages the prior limb information for peripheral joint refinement. Extensive experimental results show that HopFIR outperforms the SOTA methods by a large margin, with a mean per-joint position error (MPJPE) on the Human3.6M dataset of " + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "inline_equation", + "content": "32.67\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "text", + "content": ". We also demonstrate that the state-of-the-art GCN-based methods can benefit from the proposed hop-wise attention mechanism with a significant improvement in performance: SemGCN [42] and MGCN [49] are improved by " + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 255, + 290, + 579 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 602, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 128, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 623, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 288, + 684 + ], + "type": "text", + "content": "Monocular 3D human pose estimation aims to accurately regress the 3D locations of human joints in the camera coordinate system from a single image. It plays an important role in many applications, such as action recognition and human-computer interaction. Compared with the monocu" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 224, + 529, + 414 + ], + "blocks": [ + { + "bbox": [ + 316, + 224, + 529, + 414 + ], + "lines": [ + { + "bbox": [ + 316, + 224, + 529, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 529, + 414 + ], + "type": "image", + "image_path": "5bc2f34248cc90ef9114cd7610308005b82d2dfd435c4e252cc1d04f0ffa1cd2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 429, + 547, + 529 + ], + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 529 + ], + "type": "text", + "content": "Figure 1. Comparison of performance and model size between the proposed HopFIR and SOTA methods, namely Modulated GCN (MGCN) [49], SemGCN [42], Weight Unsharing [18], SR-Net [39], GraphSH [36], and GraFormer [43]. The methods are evaluated on the Human3.6M dataset [10] with ground truth 2D joints as input. The arrow shows the performance improvement obtained by inserting the HA layer into the other networks. Tiny, small, and normal denote the feature dimension of the HopFIR are 32, 64, and 128, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "lar systems, multi-view capture systems are expensive and inconvenient to set up and operate, which prevents them from being widely used in practice. To tackle the monocular 3D HPE task, some approaches [2, 15, 31, 32, 38] estimates 3D joint coordinates or heat maps directly from an image via a convolutional neural network (CNN) [13, 14]. However, direct regression from the image space suffers from the problem of a large parameter searching space, which always leads to a sub-optimal solution. Recently, Martinez et al. [20] constructed a simple fully connected network using only 2D keypoints as input and achieved promising 3D HPE performance, showing that the 3D human pose can be efficiently and accurately estimated from 2D joint po" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 144, + 713 + ], + "type": "text", + "content": "*Equal Contribution †Corresponding Author" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14985" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "sitions. Inspired by them and considering the thoroughly investigated 2D HPE, many works decompose the problem into two subtasks, i.e., the 2D HPE and 2D-to-3D pose lifting [18, 40, 49, 36]. 2D-to-3D pose lifting, therefore, has emerged as a fundamental task in this area that our work devotes to." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 288, + 350 + ], + "type": "text", + "content": "A key consideration in 2D-to-3D pose lifting is that the human skeleton topology is inherently sparse and graph-structured. Fully connected neural networks are less effective in modeling graph-structured data due to their simple connections among all nodes and the probability of overfitting. To leverage the information of the human skeletal topology, some works [36, 40, 49] have proposed to model the human body with GCNs and have achieved SOTA results. For example, Ci et al. [4] introduced a locally connected network to enhance the representation capability of the GCN, and Liu et al. [18] explored the weight sharing and feature transformation that occurs before or after feature aggregation in the GCN. One limitation of these GCN-based 3D HPE methods, however, is that they update the node features by aggregating their neighbors' information without considering the different contributions of these nodes to different joint synergies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 508 + ], + "type": "text", + "content": "Instead of considering all the joints of a skeleton as a whole, Xue et al. [37] demonstrated that the human skeleton exhibits obvious part-wise inconsistency in its motion patterns, as also reported in SRNet [39]. However, these works are limited to considering prior structural information of limb groupings and ignore investigating the latent groups underlying joint synergies. For example, the relative positions of the 1-hop neighbors of joint 0 are almost constant in \"Discussion\" subject, which can be a latent group, as shown in Fig. 2. Moreover, these works consider the joints in a limb as a whole to calculate the relationship with other limbs, which resulted in lower accuracy for peripheral joints, such as wrists and feet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "To address the abovementioned problems in monocular 3D HPE, we propose a novel architecture: the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR). The first key component of HopFIR is a novel hop-wise GraphFormer (HGF) module that considers " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "-hop neighbors. In the HGF module, the information of every hop of every joint is aggregated into the hidden space, such that " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "N \\times k" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": " groups of features are obtained for a skeleton model with " + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": " joints. Meanwhile, a hop-wise transformer-like attention mechanism is designed to extract the correlation among feature groups, which computes similarity by the dot product of the node feature and the group feature. The proposed HGF module enables the network to discover latent joint interactions considering human joint synergy. Because the HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interactions of peripheral joints associated with" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "a limb, we introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information. Specifically, a residual block is built from two HGF modules followed by one IJR module. The proposed HopFIR architecture achieves optimal regression accuracy with a stack of three blocks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "type": "text", + "content": "To summarize, our work makes the following contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 168, + 545, + 407 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 168, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 228 + ], + "type": "text", + "content": "- To the best of our knowledge, we design the first Hopwise GraphFormer module to explore potential joint correlations underlying human joint synergy. We also prove that other GCN-based methods can benefit from the proposed HGF module efficiently, as shown in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "type": "text", + "content": "- We design an Intragroup Joint Refinement module, which attends to intragroup joints to refine joint features through the associated limb, especially the wrists and feet. The IJR module enables HGF modules to discover the latent synergies among joints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 288, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 346 + ], + "type": "text", + "content": "- We propose the novel Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture for 3D HPE, which is built entirely from HGF and IJR modules. Specifically, two HGF modules and one IJR module are coupled into a block." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 348, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 407 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness and generalizability of the proposed modules and HopFIR architecture by providing new state-of-the-art results on two challenging datasets, i.e., Human3.6M [10] and MPI-INF-3DHP [21]." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 715 + ], + "type": "text", + "content": "3D Human Pose Estimation. Early works [26, 29] use handcrafted features, perspective relationships, and geometric constraints to estimate the 3D human pose. Recent pose estimation approaches can be generally divided into two categories. The first category of networks regresses 3D human joints directly from the image [24, 46]. Pavlakos et al. [24] adopted a CNN to predict the voxel-wise likelihoods for each joint, and Zhou et al. [46] directly embedded a kinematic object model into the networks to learn the general multi-articulate object pose. Approaches in the second category decouple the 3D HPE task into 2D pose estimation from an image and 3D pose estimation from the detected 2D joints (2D-to-3D). For example, Martinez et al. [20] proposed a simple yet effective baseline with fully-connect networks and proved that 3D human poses can be regressed simply and effectively from 2D keypoints. Our paper follows this pipeline and focuses on the 2D-to-3D pose lifting. For promoting 3D human pose regression accuracy, it is crucial to group joints with consideration of their interactions rather than treating all joints of a skeleton as a whole. Xue et al. [37] divided the human skeleton graph into five groups according to the limbs to explore part-wise motion inconsistency, and Zeng et al. [39] split the human joints into" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14986" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "local regions and recombined the global information from the rest of the joints. Our proposed HopFIR differs from these approaches by grouping joints by the " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "-hop neighbors of each joint and prior limb information, which enables the network to discover latent connections between groups in different human joint synergies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 286, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 286, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 286, + 231 + ], + "type": "text", + "content": "Graph Convolutional Networks. GCNs [5, 7, 12, 27] generalize the capability of CNNs by performing convolution operations on graph-structured data. GCNs can be divided into two categories: the spectral-based approaches [5] and the spatial-based approaches [12]. Our approach falls into the second category, which applies message-passing operations on the graph nodes and their neighbors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "text", + "content": "Due to the graph-structure topology of the human skeleton, many works [4, 18] have introduced GCN to tackle the 3D HPE task. Zhao et al. [42] proposed a SemGCN to learn the semantic relationships between human joints, and Zou et al. [49] proposed a weight modulation and an affinity modulation based on the SemGCN. These methods aggregate the first-order neighborhood messages to update the feature matrix by assigning different weights to different nodes. Some works [47, 48] have extended the first-order neighbors to high-order neighbors in the spatial domain directly. Zeng et al. [40] designed a hierarchical fusion block by dividing the fusion procedure into two stages, where all the high-order neighbors of a node are aggregated into a feature in the first stage and fuse it with the node feature and the first-order neighbor in the second stage. Zhao et al. [43] introduced Chebyshev graph convolution to fuse information among the " + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "text", + "content": "-hop neighbors of a joint directly. These works updated features by aggregating each node's own " + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "text", + "content": "-hop neighborhood information in a GCN layer. However, HopFIR considers the " + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "text", + "content": "-hop groups of all nodes to reconstruct the " + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 232, + 287, + 508 + ], + "type": "text", + "content": "-hop feature of a node through the proposed attention mechanism, which can enhance the representation capability of GCNs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Graph Attention. Graph attention networks [34] is a pioneer work that pay attention to the data in a graph structure by assigning an attention weight to each node. The introduction of the transformer [33] for machine translation tasks has proven the capacity of attention for sequential input. ViT [6] introduces the transformer into computer vision and achieves excellent performance. Inspired by them, some researchers have adopted the transformer for 3D HPE. Zhao et al. [43] applied self-attention to capture global information by calculating the similarity of all nodes. PoseFormer [44] directly applies Transformer Encoder (TE) by viewing each joint and each frame as tokens in the spatial and temporal domains, respectively. MixSTE [41] is based on [44], where each joint feature is represented by the temporal TE to model the joint motion. P-stmo [28] replaces the spatial TE with MLP and applies Stride Transformer Encoder to map " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " frames to one frame. MHFormer [17] generates" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 321, + 71, + 384, + 155 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 384, + 155 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 384, + 155 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 384, + 155 + ], + "type": "image", + "image_path": "cb670ec8dfb645aeddec880737bf4316cadc122c63e7044bff95e1e5ac76c551.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 161, + 358, + 168 + ], + "lines": [ + { + "bbox": [ + 348, + 161, + 358, + 168 + ], + "spans": [ + { + "bbox": [ + 348, + 161, + 358, + 168 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 393, + 72, + 468, + 156 + ], + "blocks": [ + { + "bbox": [ + 393, + 72, + 468, + 156 + ], + "lines": [ + { + "bbox": [ + 393, + 72, + 468, + 156 + ], + "spans": [ + { + "bbox": [ + 393, + 72, + 468, + 156 + ], + "type": "image", + "image_path": "4b3d9bfec6aa4c9b4191bc1f4e7059caf8f7eece3beb678b2e70b8085e238d37.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 160, + 433, + 168 + ], + "lines": [ + { + "bbox": [ + 424, + 160, + 433, + 168 + ], + "spans": [ + { + "bbox": [ + 424, + 160, + 433, + 168 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 468, + 72, + 531, + 156 + ], + "blocks": [ + { + "bbox": [ + 468, + 72, + 531, + 156 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 531, + 156 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 531, + 156 + ], + "type": "image", + "image_path": "2c8971dbdb6be3b194413e8c566fedf1c02c2dc63ffd51ebaa72088d8bd92efe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 160, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 495, + 160, + 503, + 168 + ], + "spans": [ + { + "bbox": [ + 495, + 160, + 503, + 168 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 182, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 248 + ], + "type": "text", + "content": "Figure 2. An illustration of the human skeleton graph and the groups in HGF and IJR modules. The " + }, + { + "bbox": [ + 304, + 182, + 545, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 182, + 545, + 248 + ], + "type": "text", + "content": "-hop neighbors of a joint are set as a group in HGF modules where (a) indicates the 1-hop groups of different joints and (b) indicates several latent groups because of physical limitations. (c) indicates the joints grouped by prior limb information." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 400 + ], + "type": "text", + "content": "multi-hypothesis at different depths of stacked spatial TEs by viewing all frames of each joint feature as a token, then communicates these hypotheses with cross-attention and self-attention. Different from the existing graph attention for 3D HPE, we propose the intergroup multi-head attention mechanism among the " + }, + { + "bbox": [ + 304, + 268, + 545, + 400 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 268, + 545, + 400 + ], + "type": "text", + "content": "-hop groups of all nodes, which assigns the attention weights by computing the similarity between the node feature and " + }, + { + "bbox": [ + 304, + 268, + 545, + 400 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 268, + 545, + 400 + ], + "type": "text", + "content": "-hop group feature. Moreover, we introduce the intragroup multi-head self-attention in limb groups to refine the joint features and promote the HGF module to discover the latent synergies among joints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 410, + 437, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 437, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 437, + 424 + ], + "type": "text", + "content": "3. The Proposed HopFIR" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": "This paper proposes a novel architecture to regress the 3D human pose from " + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": " given 2D keypoints " + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times 2}" + }, + { + "bbox": [ + 304, + 430, + 545, + 525 + ], + "type": "text", + "content": ". The proposed framework mainly consists of the HGF and IJR modules. In this section, we first review the vanilla Graph Convolutional Network and Transformer in Sec. 3.1. We then introduce the HGF and IJR modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we present the network architecture in Sec. 3.4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 533, + 471, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 471, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 471, + 544 + ], + "type": "text", + "content": "3.1. Vanilla GCN and Transformer" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": "GCN. A graph is defined as " + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": " is a set of " + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": " nodes and " + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": " is the adjacency matrix representing the edges between the nodes. Given a collection of input features " + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "inline_equation", + "content": "H^{l}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": ", a generic GCN layer that aggregates neighborhood information can be formulated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 618, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 618, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 381, + 618, + 545, + 633 + ], + "type": "interline_equation", + "content": "H ^ {(l + 1)} = \\sigma (\\tilde {A} H ^ {l} W) \\tag {1}", + "image_path": "3a0f9b0ef276868f3fdff380623848a83e9c43895a389e8d655dd8cb5a2ea8fd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "W\\in \\mathbb{R}^{D\\times D^{\\prime}}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the learnable weight matrix that transforms the feature dimension from " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "D^{\\prime}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\sigma (.)" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the activation function, such as ReLU [22]. " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "H^{(l + 1)}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the updated feature matrix, " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{A}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the symmetrically normalized affinity matrix [12] with added selfconnections, and " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "A\\in \\{0,1\\}^{N\\times N}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the adjacency ma" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14987" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 72, + 492, + 201 + ], + "blocks": [ + { + "bbox": [ + 104, + 72, + 492, + 201 + ], + "lines": [ + { + "bbox": [ + 104, + 72, + 492, + 201 + ], + "spans": [ + { + "bbox": [ + 104, + 72, + 492, + 201 + ], + "type": "image", + "image_path": "b73bf7f1727b044f6557a45cfa24667f94059486ae1ba38047561c596cfe3203.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 248 + ], + "type": "text", + "content": "Figure 3. The HopFIR architecture, with details of the HGF and IJR modules. The designed block is a residual block [9] built by two HGF modules followed by one IJR module. The proposed architecture with three blocks achieves optimal performance. Arrows of different colors represent different hops and groups in HGF and IJR, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "text", + "content": "trix. The " + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "inline_equation", + "content": "(i,j)th" + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "text", + "content": " entry " + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "inline_equation", + "content": "a_{ij} = 1" + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "text", + "content": " representing node " + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "text", + "content": " is the neighbor of node " + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "text", + "content": ". Otherwise, they are not connected and " + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "inline_equation", + "content": "a_{ij} = 0" + }, + { + "bbox": [ + 46, + 266, + 287, + 326 + ], + "type": "text", + "content": ". Therefore, the none-neighbor nodes have a weak influence on each other in the vanilla GCN, which hinders the modeling of underlying joint synergy in 3D HPE." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "content": "Transformer. Transformer architecture relies entirely on self-attention to compute representations of its input and output. The self-attention function maps the inputs to the queries " + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "content": ", and values " + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "content": " by weight matrices " + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "inline_equation", + "content": "W_{Q}" + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "inline_equation", + "content": "W_{K}" + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 46, + 327, + 288, + 399 + ], + "type": "text", + "content": ", respectively, and the matrix of outputs is calculated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 408, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(Q K ^ {T} / \\sqrt {d}\\right) V \\tag {2}", + "image_path": "6f7303f92af5167d0e63191f8ee6b2908ec7b4c6281fb3ec97d1d2cb3c9b1b17.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "text", + "content": " is the feature dimension of " + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\sqrt{d}}" + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "text", + "content": " is a scaling factor to prevent extremely small gradients. The multihead self-attention (MHSA), which performs self-attention in parallel, projects the queries, keys, and values " + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 433, + 287, + 506 + ], + "type": "text", + "content": " times with different linear projections to the respective subspaces, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 517, + 287, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 517, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 58, + 517, + 287, + 531 + ], + "type": "interline_equation", + "content": "\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(D _ {1}, \\dots , D _ {P}\\right) W _ {o} \\tag {3}", + "image_path": "093dd45d913cb38912b9f3448b180ae005ec7120105748ec0fbb562dfc7e4976.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "inline_equation", + "content": "W_O" + }, + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "text", + "content": " is the projection matrix of outputs, " + }, + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "inline_equation", + "content": "D_p = \\text{Attention}(QW_Q^p, KW_K^p, VW_V^p)" + }, + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "inline_equation", + "content": "p \\in [1,..,P]" + }, + { + "bbox": [ + 47, + 541, + 287, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 574, + 183, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 183, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 183, + 588 + ], + "type": "text", + "content": "3.2. Hop-wise GraphFormer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Previous GCN studies for 3D HPE aggregate multi-hop neighborhood information [40, 48] or assign an attention weight to each first-order neighbor [42, 49]. To effectively capture the node's neighborhood message and increase the representational capacity of GCNs, we introduce the hopwise GraphFormer (HGF) module, which treats each hop as a group (Fig.2 a,b) and computes the attention weights for each hop (more intuitive descriptions for " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "-hop can be found in the supplementary material). By considering the relationship within " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " hops, we can obtain " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "N \\times k" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " groups for a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 266, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 302 + ], + "type": "text", + "content": "skeleton graph with " + }, + { + "bbox": [ + 304, + 266, + 545, + 302 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 266, + 545, + 302 + ], + "type": "text", + "content": " joints, which provides enough combinations of joints to discover the latent correlations among joints in different human joint synergy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 303, + 474, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 303, + 474, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 303, + 474, + 315 + ], + "type": "text", + "content": "We first define the " + }, + { + "bbox": [ + 317, + 303, + 474, + 315 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 317, + 303, + 474, + 315 + ], + "type": "text", + "content": "-hop matrix " + }, + { + "bbox": [ + 317, + 303, + 474, + 315 + ], + "type": "inline_equation", + "content": "A^k" + }, + { + "bbox": [ + 317, + 303, + 474, + 315 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 320, + 545, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 370, + 320, + 545, + 353 + ], + "type": "interline_equation", + "content": "a _ {i j} ^ {k} = \\left\\{ \\begin{array}{l l} 1, & d \\left(v _ {i}, v _ {j}\\right) = k \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "9ffd63d1870cc9359d2a71c1f5697e507373a6fb10e599e14ad6ed4236e68c44.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "inline_equation", + "content": "d(v_{i}, v_{j})" + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "content": " denotes the distance of the shortest path between " + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "content": " on the skeleton graph. The " + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "content": "-hop neighborhood information is aggregated with a weighted sum of the target node's " + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 358, + 545, + 407 + ], + "type": "text", + "content": "-hop neighbors, named HopGCN:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 385, + 412, + 545, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 412, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 385, + 412, + 545, + 439 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {k} = \\sum_ {j} a _ {i j} ^ {k} h _ {j} W ^ {k} \\tag {5}", + "image_path": "5bae5b8caa00e357227dbeefccd2de51428c8381b04ab7784405deb1d6668c40.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "inline_equation", + "content": "s_i^k \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": " is a hidden representation of the " + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": "-hop neighborhood information. Eq. 5 is similar to Eq. 1 but with the extended definition of " + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": " hops. The weight matrix " + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": " is assigned to the respective hops." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "text", + "content": "Before aggregating the hidden representation " + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "inline_equation", + "content": "s_i^k" + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "text", + "content": " to the target node, we propose a transformer-like attention mechanism computing the similarity between the node feature " + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 494, + 545, + 542 + ], + "type": "text", + "content": "-hop hidden representation, as shown in Eq. 6:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 547, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 547, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 359, + 547, + 545, + 582 + ], + "type": "interline_equation", + "content": "z _ {i} ^ {k} = \\sum_ {j} \\operatorname {s o f t m a x} _ {j} \\left(\\frac {h _ {i} s _ {j} ^ {k T}}{\\sqrt {d}}\\right) s _ {j} ^ {k}. \\tag {6}", + "image_path": "f4541ad9c549286117d1d46d17bb1ded673ffecce47f0139c20e22c9239bb398.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": "Due to the structure of the human skeleton and the joint synergy, the " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": "-hop neighborhood of joint " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": " is related to that of other joints. In Fig. 2, for example, the synergy of hop " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "S_5^1" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": " is related to hop " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "S_6^1" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "S_3^1" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": ". By packing " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "z_i^k" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "s_i^k" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": " into matrices " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "Z^k" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "S^k" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": ", we can calculate " + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "inline_equation", + "content": "Z^k" + }, + { + "bbox": [ + 304, + 587, + 545, + 648 + ], + "type": "text", + "content": " in parallel:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 653, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 653, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 367, + 653, + 545, + 682 + ], + "type": "interline_equation", + "content": "Z ^ {k} = \\operatorname {s o f t m a x} \\left(\\frac {H S ^ {k ^ {T}}}{\\sqrt {d}}\\right) S ^ {k} \\tag {7}", + "image_path": "89e0b066722d052cf2ae2f58645350c81d89e92bf0945ad0252b1be0099eddc7.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "Z^k" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " can also be reconstructed using the self-attention mechanism purely on hop features or swap the positions of " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14988" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "S^k" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": ". A performance comparison of these cases is discussed in the experimental results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 287, + 144 + ], + "type": "text", + "content": "We further reduce the dimension of hidden representation by using a fully connected(FC) layer based on the order of the neighborhood, considering different amounts of information related to the target node, as formulated in Eq. 8." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 154, + 287, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 154, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 154, + 287, + 169 + ], + "type": "interline_equation", + "content": "r _ {i} ^ {k} = F ^ {k} \\left(z _ {i} ^ {k}\\right) \\tag {8}", + "image_path": "3091d9df60eabc070c24d6189495b707de20ab77d96516e3ecba32584e793c14.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "inline_equation", + "content": "r_i^k" + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "text", + "content": " is the refined representation of the " + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "text", + "content": "-hop neighborhood information with respect to node " + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "inline_equation", + "content": "F^k" + }, + { + "bbox": [ + 47, + 172, + 287, + 220 + ], + "type": "text", + "content": " is the mapping function. The refined representation is then concatenated to the updated node feature." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 220, + 287, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 255 + ], + "type": "text", + "content": "All the refined representations of hop-wise neighborhoods and the target node feature " + }, + { + "bbox": [ + 47, + 220, + 287, + 255 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 47, + 220, + 287, + 255 + ], + "type": "text", + "content": " are encoded to a " + }, + { + "bbox": [ + 47, + 220, + 287, + 255 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 220, + 287, + 255 + ], + "type": "text", + "content": "-dimensional vector:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 262, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 262, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 107, + 262, + 287, + 277 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {\\prime} = F \\left(h _ {i}, r _ {i} ^ {1}, r _ {i} ^ {2}, r _ {i} ^ {3}, \\dots , r _ {i} ^ {k}\\right) \\tag {9}", + "image_path": "0a3ab186e6c03ac98766326b308d8038b547aaf02f497b01fdf8e774a093205d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "inline_equation", + "content": "h_i'" + }, + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "text", + "content": " is the final updated feature of node " + }, + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "text", + "content": " in this layer, and " + }, + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 283, + 287, + 319 + ], + "type": "text", + "content": " is the aggregation function. Packing together the updated features of all nodes, Eq. 9 can be rewritten as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 325, + 287, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 325, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 100, + 325, + 287, + 339 + ], + "type": "interline_equation", + "content": "H ^ {\\prime} = F \\left(H, R ^ {1}, R ^ {2}, R ^ {3}, \\dots , R ^ {k}\\right). \\tag {10}", + "image_path": "7e35cecba480ca75b63d1ae1f390fe0121c25c9159ed4c0e55ea8a78640cb112.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "text", + "content": "We thereby obtain as the core layer of the HGF module the hop-wise attention (HA) layer, which extracts latent correlations between feature groups and aggregates " + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "text", + "content": "-hop neighborhood information. With the increment of " + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "text", + "content": ", we get more " + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "text", + "content": " groups on which to explore the underlying joint synergies. The value of " + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 346, + 287, + 441 + ], + "type": "text", + "content": " can be adjusted based on the tasks and pipelines. According to the experimental results, three hops achieve optimal performance in the HopFIR architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 442, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 287, + 525 + ], + "type": "text", + "content": "To fuse the current global information and original 2D information, we concatenate all the joint features in a batch and feed them into a FC layer to extract the current global information. We then concatenate the extracted global information, 2D information, and the output of the HA layer on the feature dimension of the joint feature to fuse all the information in another FC layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 532, + 204, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 204, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 204, + 544 + ], + "type": "text", + "content": "3.3. Intragroup Joint Refinement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "type": "text", + "content": "The proposed HGF module splits the skeleton graph into different groups based on the " + }, + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 550, + 287, + 693 + ], + "type": "text", + "content": "-hop neighborhood of each joint, which attends to the key groups in joint synergies. However, HGF leverages little prior information about the human body and ignores the interaction among joints in a limb, especially the interaction of the peripheral joints associated with limbs, such as the wrists and feet. We introduce an intragroup joint refinement (IJR) module to strengthen the intragroup correlation of joints grouped by limb prior information, as shown in Fig. 2(c). The HGF features of the joints in each limb group are refined in the IJR module by using multi-head self-attention [33], as in Eq. 11." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "H _ {g} = M H S A \\left(M H S A (H) _ {g}\\right) \\tag {11}", + "image_path": "276e5d01eee8f71cd2294c9921cec6db0e2db0e152f4593bda90bf4d2f16b754.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the feature matrix from the HGF module, " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "MHSA(H)_g" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the feature matrix of the group " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " updated by global multi-head self-attention, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "H_{g}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the final feature matrix of the group " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " updated by the IJR module. More details of the IJR module are provided in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 138, + 451, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 138, + 451, + 151 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 451, + 151 + ], + "type": "text", + "content": "3.4. The HopFIR Architecture" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 156, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 156, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 545, + 337 + ], + "type": "text", + "content": "The HopFIR architecture consists of the proposed HGF and IJR modules, as illustrated in Fig. 3. The residual block, which contains two HGF modules and one IJR module, is designed as the basic block in HopFIR. Moreover, we define a linear embedding layer to map the input to the latent space and a HG module to transform the output into 3D space. The HG module is a variant of HGF designed for the output layer, more details of which are provided in the supplementary material. HopFIR accepts 2D keypoints as input, which can be obtained via an off-the-shelf 2D detector. The graph is obtained by adding a normalized globally learnable " + }, + { + "bbox": [ + 305, + 156, + 545, + 337 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 156, + 545, + 337 + ], + "type": "text", + "content": "-hop graph to the skeleton graph and then symmetrically normalizing it, as in [49]. We use the L1-norm loss and L2-norm loss to compute the error between ground truth and prediction with the weighted sum as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 342, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 342, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 326, + 342, + 545, + 376 + ], + "type": "interline_equation", + "content": "L = \\alpha \\sum_ {n = 1} ^ {N} \\| Y _ {n} - \\hat {Y _ {n}} \\| _ {2} + \\beta \\sum_ {n = 1} ^ {N} \\| Y _ {n} - \\hat {Y _ {n}} \\| _ {1} \\tag {12}", + "image_path": "2d461ea3a5e5faddeb1a96d5eb2c056df402f9a550b461562db1ab596b993a58.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "content": " is the joint number, " + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\hat{Y}_n" + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "content": " is the predicted 3D position of joint " + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "Y_n" + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "content": " is the ground truth, " + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 305, + 383, + 545, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 417, + 387, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 387, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 387, + 430 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "text", + "content": "In this section, we first introduce the experimental setup and implementation details of the HopFIR networks. We then present our experimental results and comparisons with state-of-the-art methods. Finally, we conduct several ablation studies of the proposed architecture." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 503, + 488, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 488, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 488, + 515 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 521, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 665 + ], + "type": "text", + "content": "Datasets. Human3.6M [10] is currently the largest publicly available dataset for 3D human pose estimation, with 3.6 million video frames. It captures accurate 3D human joint positions from four camera viewpoints and records 11 subjects performing 15 assigned actions. Following previous works [49, 36, 42], we train our model on five subjects (S1, S5, S6, S7, S8) and test it on two subjects (S9, S11). In contrast to Human3.6M, MPI-INF-3DHP [21] includes complex outdoor scenes, which are commonly used to evaluate the generalizability of proposed methods. Accordingly, we use the test set of MPI-INF-3DHP to verify the generalizability of our model." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Protocols. For Human3.6M [10], we evaluate our model on two standard evaluation protocols: the mean per-joint position error (MPJPE) and the mean per-joint position error after Procrustes alignment (P-MPJPE)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14989" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 178 + ], + "type": "table", + "html": "
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Martinez et al. [20]ICCV201751.856.258.159.069.578.455.258.174.094.662.359.165.149.552.462.9
Zhao et al. [42](†)CVPR201947.360.751.460.561.149.947.368.186.255.067.861.042.160.645.357.6
Ci et al. [4](†)ICCV201946.852.344.750.452.968.949.646.460.278.951.250.054.840.443.352.7
Pavlo et al. [25]CVPR201947.150.649.051.853.661.449.447.459.367.452.449.555.339.542.751.8
Cai et al. [2](†)ICCV201946.548.847.650.952.961.348.345.859.264.451.248.453.539.241.250.6
Liu et al. [18](†)ECCV202046.352.247.350.755.567.149.246.060.471.151.550.154.540.343.752.4
zeng et al. [39]ECCV202044.548.247.147.851.256.850.145.659.966.452.145.354.239.140.349.9
Zou et al. [49](†)ICCV202145.449.245.749.450.458.247.946.057.563.049.746.652.238.940.849.4
Xu et al. [36](†)CVPR202145.249.947.550.954.966.148.546.359.771.551.448.653.939.944.151.9
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202245.250.848.050.054.965.048.247.160.270.051.648.754.139.743.151.8
Ours(Δ)(†)43.947.645.548.950.158.046.244.555.762.949.045.851.838.039.948.5
", + "image_path": "4cc59a59509aca93b72db87c83a1f642cb4ecb8fa472d5bd328bc366cbf4f952.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 228, + 545, + 338 + ], + "blocks": [ + { + "bbox": [ + 46, + 187, + 546, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 546, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 546, + 220 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on Human3.6M with detected 2D poses as input under Protocol #1, in millimeters. The best results are highlighted in bold and the second-best results are underlined. " + }, + { + "bbox": [ + 46, + 187, + 546, + 220 + ], + "type": "inline_equation", + "content": "(\\dagger)" + }, + { + "bbox": [ + 46, + 187, + 546, + 220 + ], + "type": "text", + "content": " indicates GCN-based methods and " + }, + { + "bbox": [ + 46, + 187, + 546, + 220 + ], + "type": "inline_equation", + "content": "(\\Delta)" + }, + { + "bbox": [ + 46, + 187, + 546, + 220 + ], + "type": "text", + "content": " indicates Transformer-based methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 228, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 49, + 228, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 228, + 545, + 338 + ], + "type": "table", + "html": "
Protocol #1Dir.Disc.EatGreetPhonePhotoPosePur.SitSitD.SmokeWaitWalkD.WalkWalkT.Avg.
Zhou et al. [45](+)ICCV201934.442.436.642.138.239.834.740.245.660.839.042.642.029.831.739.9
Ci et al. [4](+)(*)†ICCV201936.338.829.737.834.642.539.832.536.239.534.438.438.231.334.236.3
Martinez et al. [20]ICCV201737.744.440.342.148.254.944.442.154.658.045.146.447.636.440.445.5
Zhao et al. [42](†)CVPR201937.849.437.640.945.141.440.148.350.142.253.544.340.547.339.043.8
Cai et al. [2](†)ICCV201933.439.033.837.038.147.339.537.343.246.237.738.038.630.432.138.1
Liu et al. [18](†)ECCV202036.840.333.036.337.545.039.734.940.347.737.438.538.629.632.037.8
zeng et al. [39]ECCV202035.936.729.334.536.042.837.731.740.144.335.837.236.233.734.036.4
Zou et al. [49](†)ICCV2021---------------37.4
Xu et al. [36](†)CVPR202135.838.131.035.335.843.237.331.738.445.535.436.736.827.930.735.8
Zhao et al. [43] (Δ)(†) CVPR2022CVPR202232.038.030.434.434.743.335.231.438.046.234.235.736.127.430.635.2
Ours(Δ)(†)31.334.028.032.033.142.134.128.133.639.831.732.933.826.728.932.7
", + "image_path": "5d34869a30dcc425f42bc8842ac9fe600896a4a7dee51ab101b62a3d12f3c9dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 546, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 546, + 369 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on Human3.6M with ground truth 2D keypoints as input under Protocol #1, in millimeters. (+) uses additional data from MPII [1]. (*) uses pose scales in both training and testing. The best results are highlighted in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "text", + "content": "These are referred to Protocol #1 and Protocol #2, respectively. The MPJPE and P-MPJPE are given in millimeters. For MPI-INF-3DHP [21], we follow previous works [4, 42, 49] by reporting the percentage of correct keypoints (PCK) with a threshold of " + }, + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "inline_equation", + "content": "150\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 388, + 287, + 460 + ], + "type": "text", + "content": " and the area under the curve (AUC) for a range of PCK thresholds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 468, + 180, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 180, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 180, + 480 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 486, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 486, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 486, + 288, + 715 + ], + "type": "text", + "content": "Following previous work [25], we obtain the detected 2D poses using the cascaded pyramid network(CPN) [3]. We do not use data augmentation during training and testing with the 2D ground truth input to verify the efficacy of our model. We adopt Adam [11] optimizer and all experiments are conducted on a single NVIDIA RTX 3090 GPU. 3D pose regression from 2D detections is more challenging than that from 2D ground truth because the former needs to deal with some extra uncertainty in the 2D space. To manage this uncertainty, we set different configurations for them. In experiments with 2D ground truth as the input, we train the HopFIR networks with an initial learning rate of 0.001, a decay factor of 0.90 per 4 epochs, a batch size of 64, channels of 128, and PReLU activation [8]. When using detected 2D poses, we train the HopFIR networks with the initial learning rate of 0.006, a decay factor of 0.95 per 4 epochs (but 0.2 for the first 4 epochs), a batch size of 256, 256 channels, and LeaklyReLU activation [19]. To avoid overfitting, we apply Dropout[30] with a dropout rate of 0.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 388, + 485, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 485, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 485, + 399 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 407, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 526 + ], + "type": "text", + "content": "We compare the performance of HopFIR with some SOTA methods on Human3.6M under Protocol #1 and Protocol #2, with the results shown in Table 1. Our method reaches an MPJPE of " + }, + { + "bbox": [ + 304, + 407, + 545, + 526 + ], + "type": "inline_equation", + "content": "48.50\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 407, + 545, + 526 + ], + "type": "text", + "content": " and outperforms the best of the existing approaches [49] on all 15 actions. Given the uncertainty of 2D detections, we also investigate the capability of HopFIR networks using ground truth 2D key points as input. As shown in Table 2, HopFIR obtains surprisingly better performance when given precise 2D joint information and produces SOTA results, which verifies its effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "content": "We conduct a comprehensive ablation study on Human3.6M to validate the individual effectiveness of each component of the proposed HopFIR architecture under controlled settings. We follow previous works [18, 36, 49] conducting the ablation experiments using GT as inputs to avoid the influence of the 2D pose detector." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of Different Modules. We separately verify the effectiveness of the HGF module and IJR module and conduct experiments on first-order neighbors, removing all modules and the multi-hop mechanism. Note that GCN and HopGCN refer to applying only vanilla GCN (Eq. 1) and HopGCN (Eq. 5) to Fig. 3, respectively. The results in Table 3 show that each module improves the performance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14990" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 273, + 166 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 273, + 166 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 273, + 166 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 273, + 166 + ], + "type": "table", + "html": "
MethodChannelsParamsMPJPEP-MPJPE
GCN1280.36M40.6331.65
HopGCN1280.59M39.1531.40
HopGCN & IJR1282.15M36.6229.23
HopGCN & HGF1281.05M35.1928.81
HopGCN & IJR640.57M36.6929.68
HopGCN & HGF800.48M36.0129.58
HopGCN & HopFIR1282.15M32.6726.20
HopGCN & HopFIR640.54M33.5227.37
HopGCN & HopFIR320.14M35.1928.71
", + "image_path": "ca4356cddc9a1cdaa0350cef2d4e30755386f4013353dda55d59dccd1baa982f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "text", + "content": "over the GCN-only approach, and coupling the two layers to form a HopFIR block achieves further performance improvement. The HopFIR networks reduce the MPJPE to " + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "inline_equation", + "content": "32.67\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "text", + "content": ", which represents a " + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "text", + "content": " improvement over GraFormer [43]. By reducing feature channels in HopGCN & IJR and HopGCN & HGF, we decrease parameters to 0.57M and 0.48M. However, the models still achieved errors of " + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "inline_equation", + "content": "36.69\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "inline_equation", + "content": "36.01\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "text", + "content": ", compared to the HopGCN with " + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "inline_equation", + "content": "39.15\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 207, + 287, + 361 + ], + "type": "text", + "content": " error and 0.59M parameters, which should be attributed to the structure design other than the model size. Moreover, we reduce the HopFIR network parameters by changing the channels to 64 and 32, respectively, which are also superior to SOTA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "content": "Effectiveness of the HA Layer. In sec. 3.2, we introduce " + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "content": "-hop groups to discover latent joint interactions in human joint synergies. The attention matrices in Fig. 5 show the latent joint synergies captured by HopFIR, in which each weight of row " + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "content": " indicates a discovered latent group for the corresponding joint " + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "content": ". To verify the effectiveness of HA layer, we explore the correlation between all joints with a transformer encoder instead of HA layer, which explores the correlation between individual nodes but ignores the synergy between groups of joints in the human body and obtained " + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "inline_equation", + "content": "36.21 \\mathrm{~mm}" + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "content": " error. Moreover, we remove the human body prior by using random graph instead of skeleton graph without changing the HopFIR architecture and reached " + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "inline_equation", + "content": "34.68 \\mathrm{~mm}" + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "content": " error, suggesting that latent group correlations can be explored by " + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 364, + 287, + 567 + ], + "type": "text", + "content": "-hop groups, but group correlations underlying joint synergies can be better explored based on the human body prior." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": "In Table 4, we show the experimental results of three different ways to design the HA layer. HSS is the method selected in this paper, where H, S, and S are " + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": ", and H and S represent the node feature and the " + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": "-hop group feature, respectively. As we do not follow [33] in applying a linear transformation of " + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": ", we also show the result of such a linear transformation, which is denoted as HA+W. The experiment results show that HSS similarity achieves better performance in the HopFIR architecture, but one can choose the type of similarity according to the network property." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "We further insert the HA layer into SOTA pose esti" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 531, + 121 + ], + "blocks": [ + { + "bbox": [ + 63, + 174, + 268, + 185 + ], + "lines": [ + { + "bbox": [ + 63, + 174, + 268, + 185 + ], + "spans": [ + { + "bbox": [ + 63, + 174, + 268, + 185 + ], + "type": "text", + "content": "Table 3. Ablation experiments on the proposed modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 70, + 531, + 121 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 531, + 121 + ], + "type": "table", + "html": "
AttentionHAHA+W
HSSSSSSHHHSSSSSSHH
Params2.15M2.50M
MPJPE32.6734.1834.2833.2933.5333.90
P-MPJPE26.2027.7027.7127.1627.4027.41
", + "image_path": "9a41b765206e7960839280adf1a49d261651442772cd06a5bc263ea3585cf2d9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 162, + 543, + 275 + ], + "blocks": [ + { + "bbox": [ + 306, + 129, + 545, + 152 + ], + "lines": [ + { + "bbox": [ + 306, + 129, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 545, + 152 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of HA layers with different similarity computing approaches." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 162, + 543, + 275 + ], + "lines": [ + { + "bbox": [ + 309, + 162, + 543, + 275 + ], + "spans": [ + { + "bbox": [ + 309, + 162, + 543, + 275 + ], + "type": "table", + "html": "
MethodChannelsParamsMPJPEP-MPJPE
SemGCN [42]1280.27M42.1433.53
SemGCN + HA(HSS)1280.49M38.4130.56
SemGCN + HA(SSS)1280.49M41.3033.07
SemGCN + HA(SHH)1280.49M38.8131.05
SemGCN [42] w/ Non-local [35]1280.43M40.7831.46
SemGCN w/ Non-local +HA(HSS)1280.66M38.0330.50
SemGCN w/ Non-local +HA(SSS)1280.66M37.7530.17
SemGCN w/ Non-local +HA(SHH)1280.66M37.9429.71
Modulated GCN [49]1280.29M38.2530.06
Modulated GCN +HA(HSS)+W1280.96M36.5429.09
Modulated GCN +HA(SSS)+W1280.96M36.1429.02
Modulated GCN +HA(SHH)+W1280.96M37.3830.02
", + "image_path": "4af30901527f078a841ac5eec718959f8634fc7b33b2c88bb8178a21c2a66af5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 334, + 326, + 518, + 423 + ], + "blocks": [ + { + "bbox": [ + 306, + 283, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 306, + 283, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 545, + 316 + ], + "type": "text", + "content": "Table 5. Comparison of the improved performance of proposed HA layer added on different methods. We test on two GCN-based methods: SemGCN [42] and MGCN [49]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 326, + 518, + 423 + ], + "lines": [ + { + "bbox": [ + 334, + 326, + 518, + 423 + ], + "spans": [ + { + "bbox": [ + 334, + 326, + 518, + 423 + ], + "type": "table", + "html": "
Num-kChannelsParamsMPJPEP-MPJPE
11281.88M35.8828.76
21282.03M34.9627.74
31282.15M32.6726.20
41282.27M35.5828.35
Num-BlockChannelsParamsMPJPEP-MPJPE
11280.80M37.3330.71
21281.47M34.2127.63
31282.15M32.6726.20
41282.82M33.8427.63
", + "image_path": "ff54f58c35a59cb09a4884c39b39ff8c73e1ce9379aceef76a9f95fbd4206d85.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 430, + 544, + 453 + ], + "lines": [ + { + "bbox": [ + 306, + 430, + 544, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 544, + 453 + ], + "type": "text", + "content": "Table 6. Ablation study for number of " + }, + { + "bbox": [ + 306, + 430, + 544, + 453 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 430, + 544, + 453 + ], + "type": "text", + "content": "-hop and designed blocks. The units of MPJPE and P-MPJPE are millimeters (mm)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": "mation methods, namely SemGCN [42] and Modulated GCN [49], to investigate its generalizability. No changes are made to their source code, with the HA layer inserted before the information aggregation stage of these methods. The experimental results in Table 5 show that the HA layer improves these previous SOTA networks to a large degree; especially, the MPJPE of SemGCN [42], is reduced from " + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "inline_equation", + "content": "42.14\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "inline_equation", + "content": "38.41\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": ", representing an " + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": " improvement. Moreover, HA layer with linear transformation (W) makes learning more stable, so We test HA and " + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{HA + W}" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": " in SemGCN and MGCN, respectively, to show both of them are effective. For a fair comparison of parameters, we tested MLP and MHSA instead of HA, both of them with 1.03M parameters, and achieved " + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "inline_equation", + "content": "39.12\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "inline_equation", + "content": "39.22\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 473, + 545, + 700 + ], + "type": "text", + "content": " errors, respectively. More details can be found in the supplementary material. Experiments on the above methods of aggregating first-order neighbor information demonstrate the effectiveness of HA and also indicate that latent joint grouping can recognize the human joint synergies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 544, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 544, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 544, + 714 + ], + "type": "text", + "content": "Error on Peripheral Joints. We report the regres" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14991" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 72, + 263, + 199 + ], + "blocks": [ + { + "bbox": [ + 74, + 72, + 263, + 199 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 263, + 199 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 263, + 199 + ], + "type": "image", + "image_path": "3c5a5aa59debb94a5f73b5d1bef79d5835e3620abd2cc57b6041fcf28716a763.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 212, + 288, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 212, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 288, + 235 + ], + "type": "text", + "content": "Figure 4. Comparison of MPJPE for peripheral joints on the test set of the Human3.6M. R and L denote right and left, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 244, + 260, + 306 + ], + "blocks": [ + { + "bbox": [ + 75, + 244, + 260, + 306 + ], + "lines": [ + { + "bbox": [ + 75, + 244, + 260, + 306 + ], + "spans": [ + { + "bbox": [ + 75, + 244, + 260, + 306 + ], + "type": "table", + "html": "
BlockChannelsParamsMPJPEP-MPJPE
(H)(I)1281.78M34.4327.79
(I)(H)1281.78M34.1327.56
(H)(I)(H)1282.15M35.0827.98
(I)(H)(H)1282.15M34.2027.60
(H)(H)(I)1282.15M32.6726.20
", + "image_path": "7c05d19b47f8b72565c420b403d9986caa7f9b957d5f7711721ea52196f428ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 346, + 266, + 408 + ], + "blocks": [ + { + "bbox": [ + 47, + 313, + 286, + 336 + ], + "lines": [ + { + "bbox": [ + 47, + 313, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 286, + 336 + ], + "type": "text", + "content": "Table 7. Ablation study for arrangements of the designed block. (I) and (H) denote IJR module and HGF module, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 346, + 266, + 408 + ], + "lines": [ + { + "bbox": [ + 69, + 346, + 266, + 408 + ], + "spans": [ + { + "bbox": [ + 69, + 346, + 266, + 408 + ], + "type": "table", + "html": "
MPJPEP-MPJPEMPJVE
PoseFormer [44] (T=81)44.336.53.1
MixSTE [41] (T=243)40.932.62.3
MHFormer [17] (T=351)43.0--
P-STMO [28] (T=243)42.134.4-
Ours (T=243)41.132.52.1
", + "image_path": "8acee63fa452851815403c3cc7ae484f6c04213c3e542b5102ae4027ac0ac547.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 54, + 449, + 280, + 554 + ], + "blocks": [ + { + "bbox": [ + 47, + 415, + 287, + 439 + ], + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 439 + ], + "type": "text", + "content": "Table 8. Quantitative comparison on Human3.6M with detected 2D pose (CPN) in video. " + }, + { + "bbox": [ + 47, + 415, + 287, + 439 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 415, + 287, + 439 + ], + "type": "text", + "content": " denotes the number of input frames." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 449, + 280, + 554 + ], + "lines": [ + { + "bbox": [ + 54, + 449, + 280, + 554 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 280, + 554 + ], + "type": "table", + "html": "
MethodsPCKAUC
GSno GSOutdoorAll
Martinez et al. [20]49.842.531.242.517.0
Ci et al. [4]74.870.877.374.036.7
zeng et al. [39]--80.377.643.8
Li et al. [16]70.168.266.666.9-
Zhao et al. [43]80.177.974.179.043.8
Liu et al. [18] (weight unsharing)77.680.580.179.347.6
Xu et al. [36]81.581.775.280.145.8
Nie et al. [23]---83.545.9
Zou et al. [49]86.486.085.786.153.7
Ours89.185.985.987.257.0
", + "image_path": "9f8d33cad84c71ee807a4c58b335ea4a092e0f1ca87f676372ec4e3d43981cd8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 562, + 286, + 585 + ], + "lines": [ + { + "bbox": [ + 47, + 562, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 286, + 585 + ], + "type": "text", + "content": "Table 9. Quantitative comparisons on the MPI-INF-3DHP test set. GS denotes green screen." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "type": "text", + "content": "sion accuracy for the peripheral joints (wrist and feet) in Fig. 4 in comparison with some previously proposed methods [49, 42, 39]. The HopFIR network with IJR modules outperforms SOTA methods on the right foot (RFoot) by 6.7 mm, left foot (LFoot) by 7.7 mm, left wrist (LWrist) by 9.6 mm, and right wrist (RWrist) by 7.7 mm. The experimental results verify that the intragroup joint attention within each limb group strengthens the capabilities of the HopFIR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "content": "Different Numbers of " + }, + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "content": "-Hops. As HopFIR is designed" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 543, + 149 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 543, + 149 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 543, + 149 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 543, + 149 + ], + "type": "image", + "image_path": "30a7532c27391266c82b4ff8f610415ae589b0ed161638b9c984660f9f9d0763.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "Figure 5. Attention weight of the " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-hop for the " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-th joint, deeper color indicates higher correlation. " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-th col represent " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-th joint and " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-hop of " + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 161, + 545, + 196 + ], + "type": "text", + "content": "-th joint, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 205, + 544, + 398 + ], + "blocks": [ + { + "bbox": [ + 310, + 205, + 544, + 398 + ], + "lines": [ + { + "bbox": [ + 310, + 205, + 544, + 398 + ], + "spans": [ + { + "bbox": [ + 310, + 205, + 544, + 398 + ], + "type": "image", + "image_path": "6a6921bff2b26e547624cd71f09eaec86ad8f3791dba539140d71055830ec885.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 408, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 453 + ], + "type": "text", + "content": "Figure 6. Qualitative visual results for HopFIR and MGCN [49] on the Human3.6M. The black lines are the ground truth (GT) and the red lines are the predictions of HopFIR and MGCN. Wrong predictions are circled. The bottom row shows our failure case." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": "to extract the correlation between feature groups, we set different " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " values to discover various latent connections underlying the human joint synergies, with the results shown at the top of Table 6. The MPJPE gradually decreases as the number of hops increases, and reaches the best performance at 3 hops. Therefore, the optimal number of hops for 3D HPE is 3, which entails that we obtain " + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "inline_equation", + "content": "16 \\times 3" + }, + { + "bbox": [ + 304, + 474, + 545, + 605 + ], + "type": "text", + "content": " groups from the skeleton graph. Each of the 16 groups corresponds to a potential correlation among coupled nodes at different distances, and three hops is sufficient to recognize the joint synergies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "Arrangement of the Designed Block. To investigate the optimal structure of the designed block, experiments are performed with various block numbers and various combinations of HGF and IJR modules. As shown at the bottom of Table 6, the error gradually decreases as the number of blocks increases, until the best performance is achieved at 3 blocks. As shown in Table 7, the (H)(H)(I) arrangement achieves the optimal results by reducing the error to 32.67 mm. HGF treats each hop as a group and applies a hop" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14992" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "wise attention mechanism to these groups to discover latent joint synergy. IJR utilizes the limb prior for peripheral joint refinement. Thus, (H)(H)(I) first integrates the complete joint information and then refines it by IJR. While (I)(H)(H) reverses this procedure resulting in the insufficient utilization of joint information. That the results are superior to GraFormer [43] in all cases except for the single-block model indicates that the HopFIR has significant human pose representation capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 288, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 288, + 239 + ], + "type": "text", + "content": "Extend to Temporal Domain. Without a specific design to integrate temporal information, we extend to the temporal domain by adding two TEs after each block and replacing the HG module with a linear layer, and achieved competitive results as shown in table 8." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 337 + ], + "type": "text", + "content": "Cross-Dataset Results on MPI-INF-3DHP. Table 9 further compares HopFIR with previous methods on cross-dataset scenarios to validate its generalizability. For these experiments, we train our model on the Human3.6M dataset and test it on the test set of the MPI-INF-3DHP dataset. The results show that our approach obtains better results than other methods, which verifies the generalizability of our approach to unseen scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 337, + 288, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 408 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 6, we show the visual results on Human3.6M in the world space. The bottom of the figure shows some failure cases of HopFIR, which predict some wrong joint positions. The figure shows that HopFIR is able to predict 3D joint positions more accurately, even for poses that cause difficulties for MGCN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 421, + 125, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 125, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 125, + 433 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 289, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 289, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 289, + 596 + ], + "type": "text", + "content": "We present the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) as a novel architecture for 3D human pose estimation. The proposed architecture mainly comprises the HGF and IJR modules. The HGF module improves on the GCN-based pose estimation networks by grouping the joints by " + }, + { + "bbox": [ + 46, + 441, + 289, + 596 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 441, + 289, + 596 + ], + "type": "text", + "content": "-hop neighborhood and capturing the potential joint correlations in the different joint synergies. Because the peripheral joints strongly interact with intra-limb joints, the proposed IJR module applies intragroup attention to refine the peripheral joint features through the associated limb. The proposed method achieves new state-of-the-art results while maintaining a modest model size." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 609, + 145, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 145, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 145, + 623 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China (Grant No. 2021YFC0122602), in part by the Joint Funds Program of the National Natural Science Foundation of China (Grant No. U21A20517), in part by the Basic Science Centre Program of National Natural Science Foundation of China (Grant No. 72188101)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 92, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 311, + 92, + 547, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 92, + 547, + 147 + ], + "spans": [ + { + "bbox": [ + 311, + 92, + 547, + 147 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 149, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 149, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 311, + 149, + 547, + 217 + ], + "type": "text", + "content": "[2] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2272-2281, 2019. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 218, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 218, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 311, + 218, + 547, + 274 + ], + "type": "text", + "content": "[3] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cascaded pyramid network for multi-person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7103-7112, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 276, + 547, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 276, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 311, + 276, + 547, + 320 + ], + "type": "text", + "content": "[4] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019. 2, 3, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 322, + 547, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 322, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 311, + 322, + 547, + 367 + ], + "type": "text", + "content": "[5] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 369, + 547, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 369, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 311, + 369, + 547, + 434 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 437, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 437, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 311, + 437, + 547, + 480 + ], + "type": "text", + "content": "[7] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., 2:729-734 vol. 2, 2005. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 483, + 547, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 483, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 311, + 483, + 547, + 538 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 540, + 547, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 540, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 311, + 540, + 547, + 585 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 587, + 547, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 547, + 642 + ], + "type": "text", + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 644, + 547, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 644, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 644, + 547, + 677 + ], + "type": "text", + "content": "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 679, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 547, + 713 + ], + "type": "text", + "content": "[12] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14993" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[15] Kyoungoh Lee, Inwooong Lee, and Sanghoon Lee. Propagating lstm: 3d pose estimation based on joint interdependency. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[16] Chen Li and Gim Hee Lee. Generating multiple hypotheses for 3d human pose estimation with mixture density network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9887-9895, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "text", + "content": "[17] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13147-13156, 2022. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "text", + "content": "[18] Kenkun Liu, Rongqi Ding, Zhiming Zou, Le Wang, and Wei Tang. A comprehensive study of weight sharing in graph networks for 3d human pose estimation. In European Conference on Computer Vision, pages 318-334. Springer, 2020. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[19] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, volume 30, page 3. Atlanta, Georgia, USA, 2013. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[20] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1, 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "type": "text", + "content": "[21] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 533 + ], + "type": "text", + "content": "[22] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 577 + ], + "type": "text", + "content": "[23] Qiang Nie, Ziwei Liu, and Yunhui Liu. Lifting 2d human pose to 3d with domain adapted 3d body concept. International Journal of Computer Vision, 131(5):1250-1268, 2023. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "text", + "content": "[24] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7025-7034, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[25] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7753-7762, 2019. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[26] Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Reconstructing 3d human pose from 2d image landmarks. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "European conference on computer vision, pages 573-586. Springer, 2012. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[27] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 207 + ], + "type": "text", + "content": "[28] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 461-478. Springer, 2022. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[29] Cristian Sminchisescu. 3d human motion analysis in monocular video: techniques and challenges. In Human motion, pages 185-211. Springer, 2008. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929-1958, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "text", + "content": "[31] Xiao Sun, Jiaxiang Shang, Shuang Liang, and Yichen Wei. Compositional human pose regression. In Proceedings of the IEEE International Conference on Computer Vision, pages 2602-2611, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "text", + "content": "[32] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Proceedings of the European conference on computer vision (ECCV), pages 529-545, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[34] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "text", + "content": "[35] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 513, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 555 + ], + "type": "text", + "content": "[36] Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16105-16114, 2021. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[37] Youze Xue, Jiansheng Chen, Xiangming Gu, Huimin Ma, and Hongbing Ma. Boosting monocular 3d human pose estimation with part aware attention. IEEE Transactions on Image Processing, 31:4278-4291, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 657 + ], + "type": "text", + "content": "[38] Wei Yang, Wanli Ouyang, Xiaolong Wang, Jimmy Ren, Hongsheng Li, and Xiaogang Wang. 3d human pose estimation in the wild by adversarial learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5255-5264, 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[39] Ailing Zeng, Xiao Sun, Fuyang Huang, Minhao Liu, Qiang Xu, and Stephen Lin. Snet: Improving generalization in 3d human pose estimation with a split-and-recombine approach. In European Conference on Computer Vision, pages 507-523. Springer, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14994" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 586 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[40] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11436-11445, 2021. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[41] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13232-13242, 2022. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[42] Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dimitris N Metaxas. Semantic graph convolutional networks for 3d human pose regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3425-3435, 2019. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 295 + ], + "type": "text", + "content": "[43] Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20438-20447, 2022. 1, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 296, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 296, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 296, + 287, + 351 + ], + "type": "text", + "content": "[44] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11656-11665, 2021. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 287, + 407 + ], + "type": "text", + "content": "[45] Kun Zhou, Xiaoguang Han, Nianjuan Jiang, Kui Jia, and Jiangbo Lu. Hemlets pose: Learning part-centric heatmap triplets for accurate 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2344-2353, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 407, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 287, + 461 + ], + "type": "text", + "content": "[46] Xingyi Zhou, Qixing Huang, Xiao Sun, Xiangyang Xue, and Yichen Wei. Towards 3d human pose estimation in the wild: a weakly-supervised approach. In Proceedings of the IEEE International Conference on Computer Vision, pages 398-407, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 463, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 508 + ], + "type": "text", + "content": "[47] Yiran Zhu, Xing Xu, Fumin Shen, Yanli Ji, Lianli Gao, and Heng Tao Shen. PosegTac: Graph transformer encoder-decoder with atrous convolution for 3d human pose estimation. In IJCAI, pages 1359-1365, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 509, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 541 + ], + "type": "text", + "content": "[48] Zhiming Zou, Kenkun Liu, Le Wang 0003, and Wei Tang. High-order graph convolutional networks for 3d human pose estimation. In BMVC, 2020. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 542, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 287, + 586 + ], + "type": "text", + "content": "[49] Zhiming Zou and Wei Tang. Modulated graph convolutional network for 3d human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11477-11487, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14995" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_content_list.json b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..68d070fdcf7844d85bd450a5dad5dfdbdb35459b --- /dev/null +++ b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_content_list.json @@ -0,0 +1,1392 @@ +[ + { + "type": "text", + "text": "Householder Projector for Unsupervised Latent Semantics Discovery", + "text_level": 1, + "bbox": [ + 137, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yue Song $^{1}$ , Jichao Zhang $^{1}$ , Nicu Sebe $^{1}$ , and Wei Wang $^{2}$", + "bbox": [ + 264, + 191, + 705, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Information Engineering and Computer Science, University of Trento, Italy $^{2}$ Beijing Jiaotong University, China yue.song@unitn.it", + "bbox": [ + 124, + 222, + 844, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Adversarial Networks (GANs), especially the recent style-based generators (StyleGANs), have versatile semantics in the structured latent space. Latent semantics discovery methods emerge to move around the latent code such that only one factor varies during the traversal. Recently, an unsupervised method proposed a promising direction to directly use the eigenvectors of the projection matrix that maps latent codes to features as the interpretable directions. However, one overlooked fact is that the projection matrix is non-orthogonal and the number of eigenvectors is too large. The non-orthogonality would entangle semantic attributes in the top few eigenvectors, and the large dimensionality might result in meaningless variations among the directions even if the matrix is orthogonal. To avoid these issues, we propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix. The orthogonality guarantees that the eigenvectors correspond to disentangled interpretable semantics, while the low-rank property encourages that each identified direction has meaningful variations. We integrate our projector into pre-trained StyleGAN2/StyleGAN3 and evaluate the models on several benchmarks. Within only $1\\%$ of the original training steps for fine-tuning, our projector helps StyleGANs to discover more disentangled and precise semantic attributes without sacrificing image fidelity. Code is publicly available via https://github.com/KingJamesSong/HouseholderGAN.", + "bbox": [ + 75, + 353, + 473, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Adversarial Networks (GANs) [15] and the recent style-based generative models (StyleGANs) [27, 28,", + "bbox": [ + 76, + 832, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df0ec7839c5f7f07dcc9cdaef12468d57362fd1e62ab465516f6ea383b487253.jpg", + "image_caption": [ + "Figure 1: Motivation of our proposed Householder Projector. Here \"Projector\" denotes the projection matrix that maps latent codes to features, i.e., the modulation weight of StyleGANs. (Top Left) The singular value imbalance of the non-orthogonal projector would entangle multiple semantics in the top interpretable directions. (Top Right) Due to the large dimensionality of the projector, directly enforcing vanilla orthogonality would spread the data variations among all the eigenvectors, leading to imperceptible and meaningless traversal. (Bottom) Our Householder Projector equips the projection matrix with low-rank orthogonal properties, which simultaneously disentangles semantics into multiple equally-important eigenvectors and guarantees that each direction could correspond to semanticallymeaningful variations." + ], + "image_footnote": [], + "bbox": [ + 506, + 323, + 890, + 536 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[26] in particular have become the leading paradigm of generative modeling in the vision domain. The latent spaces of StyleGANs are known to embed rich and hierarchical semantics [14, 21]; moving the latent code in certain directions could trigger meaningful variations in the output images. Therefore, latent semantics discovery methods", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The first two authors contribute equally; Wei Wang is the corresponding author.", + "bbox": [ + 75, + 875, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7712", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "emerge to identify such interpretable directions that each variation factor is disentangled and the generation process can be precisely controlled [14, 47, 42, 61, 48, 1, 58, 37, 8].", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among the recent explorations of unsupervised interpretable semantics discovery methods [56, 48, 57, 41], SeFa [48] pointed out a promising direction to discover semantically meaningful concepts by computing the eigenvector of the projector. Here we refer to the projection matrix that maps latent codes to features as the projector. The key observation is that using the eigenvectors of the projector for latent perturbation would maximize the data variations. Such identified eigenvectors/directions would correspond to meaningful semantic concepts. However, as shown in Fig. 1 top left, this is likely to cause semantics entangled in the top few eigenvectors. This phenomenon stems from the fact that the variation caused by an eigenvector is actually determined by the associated eigenvalue. Due to the imbalanced eigenvalue distribution, the discovered directions are not equally-important, and the top few ones would simultaneously manipulate multiple attributes. This eigenvalue discrepancy can be mitigated by enforcing orthogonality constraint to the projector. Nonetheless, since standard orthogonal matrices have as many equally-important eigenvectors as the dimensionality, there would not be enough semantics to mine in practice when the method scales to large models such as StyleGANs whose projector dimension is 1024 or 512. Consequently, as an accompanying limitation, the data variations would be split among all the eigenvectors and none of them could produce meaningful output changes (see Fig. 1 top right).", + "bbox": [ + 75, + 141, + 472, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To resolve the above issues, we propose Householder Projector, a low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix that maps latent codes to features. The projector is first decomposed to its SVD form $(\\mathbf{USV}^T)$ . Next, the orthogonal singular vectors $\\mathbf{U}$ and $\\mathbf{V}$ are represented by a series of Householder reflectors, respectively. Thanks to the normalization of Householder reflections, the orthogonality is also preserved during backpropagation. For the singular value $\\mathbf{S}$ , we explicitly set it as a low-rank identity matrix (i.e., $\\mathbf{S} = \\mathrm{diag}(1,\\dots ,1,0,\\dots ,0)$ ) whose rank defines exactly the number of semantic concepts. As shown in Fig. 1 bottom, the low-rank property guarantees that the identified directions would cause meaningful variations, while the orthogonality encourages that each semantic attribute is disentangled from the others. Moreover, a proper initialization scheme is proposed to leverage the statistics of pre-trained weights, and an acceleration technique is applied to speed up the computation. We also propose a metric dedicated to measuring the smoothness of latent space to interpretable directions based perturbations. Our Householder Projector is integrated into pre-trained StyleGANs [28, 26] at multiple different layers to mine the di", + "bbox": [ + 75, + 553, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "verse and hierarchical semantics. Since our projector inevitably changes the pre-trained parameters, the modified models incorporated with our projector are fine-tuned for limited steps to maintain the original image fidelity. Both quantitative and qualitative results on several widely used benchmarks [29, 63, 10, 25, 13] show that within marginal fine-tuning steps (1% of the training steps), our Householder Projector improves the latent semantics discovery of StyleGANs to have more precise attribute control while not impairing the quality of generated images.", + "bbox": [ + 496, + 90, + 893, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key results and main contributions are as follows:", + "bbox": [ + 517, + 242, + 877, + 257 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projector of generative models for latent semantics discovery.", + "- Our Householder Projector can be easily integrated into pre-trained GAN models. With the acceleration technique, it can be efficiently fine-tuned for very limited additional steps, which paves the way for applying latent semantics discovery and orthogonality techniques to large-scale generative models such as StyleGANs.", + "- Extensive experiments on two popular backbones (StyleGAN2 [28] and StyleGAN3 [26]) and six benchmarks (FFHQ [29], LSUN Church and Cat [63], MetFaces [25], AFHQ [10], and SHHQ [13]) demonstrate that within marginal extra fine-tuning steps (1% of the original training steps), our method could both achieve precise attribute control and preserve the original image quality." + ], + "bbox": [ + 500, + 267, + 890, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 558, + 640, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Adversarial Networks. In the past few years, GAN-based generative models [15] have achieved remarkable progress in high-fidelity image synthesis [43, 24, 5, 27, 28, 26, 25, 39, 6, 16, 50, 46]. The generation process usually takes the following procedure: a randomly-sampled latent code is fed into the generator through a projection step, and then the generator outputs realistic-looking images. Recently, the style-based generators [27, 28, 26] that gradually absorb layer-wise latent style codes are becoming the de facto GAN backbones. StyleGAN2 [28] improves the original StyleGAN [27] by redesigning generator normalization and training techniques, and StyleGAN3 [26] further explores some equivariance properties. We use the popular StyleGAN2 and StyleGAN3 as our backbones in this paper. Latent Semantics Discovery. Recently, lots of methods explore disentangling the latent space to achieve image editing by moving the latent code in the identified interpretable directions [7, 3, 21, 47, 17, 56, 67, 48, 18, 41, 54, 57, 55, 30, 55, 31, 60, 22, 44, 9, 53, 51]. Supervised methods rely on human annotations (i.e., segmentation masks, attribute categories, 3D priors, and text descriptions) to de", + "bbox": [ + 496, + 583, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7713", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fine the semantic labels [14, 47, 61, 21, 42, 33, 8, 12, 49, 40, 34, 62, 59, 23]. Here we mainly highlight some relevant unsupervised approaches that are free of annotations. Voynov et al. [56] proposed to jointly learn a set of directions and a classifier such that the interpretable directions can be recognized by the classifier. More recently, Peebles et al. [41] and Wei et al. [57] proposed to add orthogonal Hessian and Jacobian regularization to encourage disentangled representations, respectively. Song et al. [51] proposed to use wave equations to model the spatiotemporal dynamic non-linear latent traversals in generative models. SeFa [48] showed that the eigenvectors of the projector after the latent code could maximize the data variations and proposed to directly use them as the interpretable directions. However, due to the imbalanced eigenvalues, the image attributes would be entangled in the top few eigenvectors. Our proposed Householder Projector solved this issue by parameterizing the projector to a low-rank but orthogonal matrix. Notice that the used orthogonality technique in [56] is different from ours. In [56], the authors use matrix exponential $\\exp (\\mathbf{A} - \\mathbf{A}^T)$ to generate skew orthogonal matrices where the skew-symmetry could limit the representation power. Also, the technique cannot parameterize given matrices and cannot explicitly control the rank. Our Householder representation is more general and flexible, allowing for controllable rank and parameterization of given matrices. Compared with the soft orthogonality regularization used in [41, 57, 18], the hard orthogonality of our method helps the model to learn more uncorrelated attributes within less fine-tuning steps.", + "bbox": [ + 76, + 90, + 472, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to global editing approaches discussed above, some methods can perform local image editing in a post hoc way: they first define or learn a segment of regions of interests, and then manipulate the masked intermediate features for local editing [2, 11, 65, 66, 38]. Empowered by the precise control of attributes, our method can also achieve competitive performance in local image editing (see Sec. 4.2).", + "bbox": [ + 76, + 545, + 472, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Householder Projector", + "text_level": 1, + "bbox": [ + 76, + 667, + 294, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section starts with the preliminary introduction of the closed-form latent semantics discovery. Next, we analyze its inherent limitation on entangled semantics and then illustrate our proposed Householder Projector in detail.", + "bbox": [ + 76, + 695, + 470, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Closed-form Latent Discovery", + "text_level": 1, + "bbox": [ + 76, + 770, + 449, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous latent semantics discovery approaches [14, 56, 17, 47, 48] consider the GAN manipulation as $\\text{edit}(G(\\mathbf{z})) = G(\\mathbf{z} + \\alpha \\mathbf{n})$ where $G(\\cdot)$ represents the generator, $\\mathbf{z} \\in \\mathbb{R}^d$ denotes the latent code of dimension $d$ , $\\mathbf{n} \\in \\mathbb{R}^d$ is the identified semantically meaningful direction, and $\\alpha$ represents the perturbation strength. If one views the GAN as a multi-step projection function, the first projection can", + "bbox": [ + 76, + 794, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be expressed as:", + "bbox": [ + 498, + 90, + 612, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG _ {1} (\\mathbf {z} + \\alpha \\mathbf {n}) = \\mathbf {A} \\mathbf {z} + \\mathbf {b} + \\alpha \\mathbf {A} \\mathbf {n}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 109, + 890, + 125 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ and $\\mathbf{b}$ denote the weight and bias of the projection step (e.g., convolution or linear transform), respectively. As can be seen from eq. (1), the resultant manipulation depends on the term $\\alpha \\mathbf{A}\\mathbf{n}$ . Intuitively, an interpretable direction $\\mathbf{n}$ should cause large variations of $G_{1}(\\mathbf{z} + \\alpha \\mathbf{n})$ , which is equivalent to maximizing the impact of $\\alpha \\mathbf{A}\\mathbf{n}$ . Motivated by this observation, SeFa [48] proposed to consider the formulation as the following constrained optimization problem:", + "bbox": [ + 496, + 128, + 893, + 250 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} ^ {\\star} = \\arg \\max | | \\mathbf {A n} | | _ {2} ^ {2} s. t. \\mathbf {n} ^ {T} \\mathbf {n} = 1, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 251, + 890, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the constraint $\\mathbf{n}^T\\mathbf{n} = 1$ is set to satisfy vector orthogonality, and $||\\cdot ||$ denotes the $l_{2}$ norm. Introducing a Lagrange multiplier $\\lambda$ leads to $2\\mathbf{A}^{T}\\mathbf{A}\\mathbf{n} - 2\\lambda \\mathbf{n} = 0$ . The closed-form solutions all correspond to the eigenvectors of $\\mathbf{A}^T\\mathbf{A}$ . This presents a promising approach to identify the semantics by exploiting the projector $\\mathbf{A}$ that projects latent codes. However, one fact overlooked by [48] is that the eigenvectors would cause different extents of variations due to the discrepancy of associated eigenvalues. Supposing that $\\mathbf{n}$ is an eigenvector of $\\mathbf{A}^T\\mathbf{A}$ , then we would have $||\\mathbf{A}\\mathbf{n}||_2^2 = \\sigma^2$ where $\\sigma$ is the corresponding singular value of $\\mathbf{A}$ . For non-orthogonal matrices, the singular values are exponentially decreasing, i.e., $\\sigma_1\\geq \\sigma_2\\geq \\ldots \\geq \\sigma_d$ . This would cause most variations captured by the first few interpretable directions. The imbalance is thus likely to make semantic attributes entangled in the top eigenvectors (see Fig. 1 top left).", + "bbox": [ + 496, + 271, + 893, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Householder Low-rank Orthogonal Projector", + "text_level": 1, + "bbox": [ + 500, + 521, + 885, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The eigenvalue discrepancy can be eliminated by enforcing strict orthogonality. Orthogonal matrices have the property of identical eigenvalues, which assigns equal importance to each eigenvector. The low-rank constraint could further limit the number of semantics to mine. Therefore, we propose to use Householder representation, a flexible and general framework to parameterize matrices, to endow the projection matrix with low-rank orthogonality.", + "bbox": [ + 496, + 545, + 890, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Householder Parameterization. Householder representations can parameterize any matrices by using a series of Householder reflectors to represent the orthogonal singular vectors of its Singular Value Decomposition (SVD) form. In the field of deep learning, it has been used to parameterize the transition matrix and to stabilize gradients of recurrent neural networks [36, 64, 35]. The key to the orthogonality representation relies on the following theorem:", + "bbox": [ + 496, + 666, + 893, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 (Householder representation [20, 32]). Given any square orthogonal matrix $\\mathbf{M} \\in \\mathbb{R}^{m \\times m}$ , it can be represented by the product of Householder matrices $\\mathbf{M} = \\mathbf{H}_1\\mathbf{H}_2 \\ldots \\mathbf{H}_m$ where each Householder matrix is parameterized by a vector as $\\mathbf{H}_i = \\mathbf{I} - 2\\frac{\\mathbf{h}_i\\mathbf{h}_i^T}{||\\mathbf{h}_i||_2^2}$ .", + "bbox": [ + 496, + 794, + 893, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{USV}^T$ denote the SVD of the projector $\\mathbf{A}$ where $\\mathbf{S}$", + "bbox": [ + 517, + 883, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7714", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d8e11d6c7fb6a29907ef099e538b9d71d280386a276eb3af226bab384fe5801.jpg", + "image_caption": [ + "Figure 2: Gallery of some semantic attributes discovered by our Householder Projector across all used datasets (FFHQ [29] in the top row, LSUN Church [63] and LSUN Cat [63] in the $2_{nd}$ row, SHHQ [13] in the $3_{rd}$ row, and MetFaces [25] and AFHQ [10] in the bottom row). These semantic attributes are sorted from low-level layers (left) to high-level layers (right)." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 883, + 825 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7715", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "denotes the diagonal singular value, and $\\mathbf{U}$ and $\\mathbf{V}$ are left and right orthogonal singular vectors. We use the accumulation of Householder reflectors (i.e., $\\prod_{i=0}^{w}\\left(\\mathbf{I}-2\\frac{\\mathbf{u}_i\\mathbf{u}_i^T}{||\\mathbf{u}_i||_2^2}\\right)$ and $\\prod_{i=0}^{h}\\left(\\mathbf{I}-2\\frac{\\mathbf{v}_i\\mathbf{v}_i^T}{||\\mathbf{v}_i||_2^2}\\right)$ where $w,h$ denotes the width and height of $\\mathbf{A}$ ) to parameterize $\\mathbf{U}$ and $\\mathbf{V}$ , respectively. Notice that only $\\mathbf{u}_i$ and $\\mathbf{v}_i$ are the actual learnable parameters. For $\\mathbf{S}$ , we explicitly set it to a diagonal matrix and keep the weight fixed. The projector is thus represented by our Householder parameterizations.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Low-rank Constraint. To achieve the property of disentangled attributes, one straightforward approach is to parameterize $\\mathbf{A}$ as an orthogonal matrix, i.e., to set $\\mathbf{S}$ to an identity matrix $\\mathbf{I}$ where $\\mathbf{I}_{i,j} = 1$ for $i = j$ and $\\mathbf{I}_{i,j} = 0$ otherwise. This would lead to equally-important semantic attributes whose number is exactly the projector dimension. However, for large generative models such as StyleGANs, the projector dimension is typically very large (e.g., 512 or 1024). It is not likely to have enough attributes to edit in practice. Setting $\\mathbf{S}$ to a full-rank diagonal matrix would spread data variations among all the eigenvectors, resulting in trivial and imperceptible traversal (see Fig. 1 top right). To avoid this issue, we propose to define $\\mathbf{S}$ as a low-rank identity matrix:", + "bbox": [ + 75, + 243, + 472, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\operatorname {d i a g} (\\underbrace {1 , \\dots , 1} _ {N}, 0, \\dots , 0), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 457, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ defines the rank and also the number of semantic attributes to mine. By restricting the rank of the projector, we explicitly limit the number of interpretable directions. As shown in Fig. 1 bottom, this would benefit the latent traversal for more meaningful output variations.", + "bbox": [ + 75, + 500, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Orthogonality Preservation. One advantage of our Householder representation is that the orthogonality can be kept during backpropagation. Since we have the vector normalization $\\frac{\\mathbf{h}_i\\mathbf{h}_i^T}{||\\mathbf{h}_i||_2^2}$ , the impact of any gradient descent step $\\mathbf{h}_i - \\eta \\nabla \\mathbf{h}_i$ on the orthogonality would be cancelled, i.e., the updated vector remains orthogonal after normalization.", + "bbox": [ + 75, + 575, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization. When our method is applied to pre-trained models, the well-trained network weights could be leveraged to initialize our Householder Projection. To this end, we propose to use the nearest-orthogonal mapping [52] to project the weight matrix into its orthogonal form that has the nearest distance in the Frobenius norm (i.e., $\\min_{\\mathbf{R}}||\\mathbf{R} - \\mathbf{A}||_{\\mathrm{F}}$ where $\\mathbf{R}^T\\mathbf{R} = \\mathbf{I}$ ). The closed-form solution is given by $\\mathbf{U}\\mathbf{V}^T$ where $\\mathbf{USV}^T$ is the SVD of the original weight matrix $\\mathbf{A}$ . Next, we decompose $\\mathbf{U}$ and $\\mathbf{V}$ into their Householder reflectors and use them to initialize our projector. Such an initialization scheme leverages the statistics of the original weight matrix, which might give our projector a good starting point and improve the performance (see the ablation study of Sec. D.3 in the supplementary).", + "bbox": [ + 75, + 672, + 470, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acceleration. The accumulated product of elementary", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d6b66fa10c03b7a4919c43e42619846c3a5b6da75936eff8b883a4ee6d10a60.jpg", + "image_caption": [ + "Figure 3: Illustration on how our Householder Projector represents the modulation weight $\\mathbf{A}$ of StyleGANs. Here \"Demod\", \"EMA\", and \"FN\" denote Demodulation, Exponential Moving Average, and Filtered Non-linearities, respectively. The projector is parameterized by its SVD form where $\\mathbf{U}$ and $\\mathbf{V}$ are represented by the accumulation of Householder reflectors, and $\\mathbf{S}$ is set to a low-rank identity matrix. Our projector is applied at multiple different layers of StyleGANs to explore the diverse and hierarchical semantics. The actual learnable parameters are $\\mathbf{u}_i$ and $\\mathbf{v}_i$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2b6616c60b43c442c09eb5107a69155c47113f61e028f220c90e233381ddf32.jpg", + "image_caption": [ + "Figure 4: Comparison with ReSeFa [66]. The blue lines indicate the specific regions changed by our method, and the red box indicates the region of interest that is needed as input to ReSeFa [66]." + ], + "image_footnote": [], + "bbox": [ + 506, + 347, + 890, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Householder matrices can be accelerated via the theorem:", + "bbox": [ + 500, + 536, + 880, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2 (Compact WY representation [4]). For any accumulation of $m$ Householder matrices $\\mathbf{H}_1 \\ldots \\mathbf{H}_m$ , there exists $\\mathbf{W}, \\mathbf{Y} \\in \\mathbb{R}^{d \\times m}$ such that $\\mathbf{I} - 2\\mathbf{W}\\mathbf{Y}^T = \\mathbf{H}_1 \\ldots \\mathbf{H}_m$ where computing $\\mathbf{W}$ and $\\mathbf{Y}$ takes $O(dm^2)$ time and $m$ sequential Householder multiplications.", + "bbox": [ + 496, + 559, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This theorem indicates the possibility to repeatedly split the accumulation $\\mathbf{H}_1\\mathbf{H}_2\\dots \\mathbf{H}_m$ into multiple subsequences until irreducible. Then these sub-sequences can be computed in parallel and gradually merged. As revealed in the ablation of Sec. D.3 in the supplementary, this technique could fully exploit the parallel computational power of GPUs and greatly speed up the aggregation, particularly in our case where the projector dimension is large.", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation in StyleGANs. Fig. 3 depicts how our Householder Projector modifies the original StyleGAN architectures. The projector used in the weight modulation module is represented by our proposed projector. The weight matrix is thus endowed with low-rank orthogonal properties. We insert the proposed projector at every layer of StyleGAN2 and every four layers of StyleGAN3. Since StyleGAN3 has 15 intermediate layers, the adjacent layers have very similar and even repeated semantics. Therefore,", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7716", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6676f9bc3bc7787058978d422ea5c986847fdedcf28593f16111be81721b9aa0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IdentityPoseAgeGenderGlassesSmile
Identity0.650.240.030.040.010.03
Pose0.110.570.040.040.110.13
Age0.020.050.670.030.190.03
Gender0.030.320.020.520.100.00
Glasses0.010.080.000.020.880.01
Smile0.020.010.010.000.010.95
", + "bbox": [ + 86, + 88, + 279, + 142 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/da3ccaac59a2ff50a64a16edfab91ebeeab7ff6ff76a634bf1001f20b433f586.jpg", + "table_caption": [ + "(a) Our method" + ], + "table_footnote": [], + "table_body": "
IdentityPoseAgeGenderGlassesSmile
Identity0.560.060.050.020.050.26
Pose0.440.480.050.010.010.00
Age0.430.270.220.050.030.01
Gender0.330.180.040.400.040.02
Glasses0.270.140.040.100.230.22
Smile0.200.080.090.000.030.61
", + "bbox": [ + 285, + 88, + 478, + 142 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/662d05b222066e9ff88c805525a8f048ce5ecfec2550419e3a06f1910b115feb.jpg", + "table_caption": [ + "(b) SeFa" + ], + "table_footnote": [], + "table_body": "
IdentityPoseAgeGenderGlassesSmile
Identity0.510.270.040.050.030.10
Pose0.390.350.050.020.070.11
Age0.260.140.320.100.050.12
Gender0.200.060.340.290.040.07
Glasses0.180.060.120.070.420.14
Smile0.130.050.070.020.030.70
", + "bbox": [ + 483, + 88, + 676, + 142 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/007412e07a25b9cdddfeeef02bf8b8c8cee4e027182b4bb0a3c79d261f9e4a69.jpg", + "table_caption": [ + "(c) OrJaR" + ], + "table_footnote": [], + "table_body": "
IdentityPoseAgeGenderGlassesSmile
Identity0.540.250.040.060.020.08
Pose0.420.300.080.040.050.12
Age0.210.150.280.110.080.17
Gender0.270.070.230.250.040.13
Glasses0.150.120.160.040.370.16
Smile0.110.090.120.010.010.66
", + "bbox": [ + 681, + 88, + 875, + 142 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(d) HP", + "bbox": [ + 756, + 145, + 803, + 157 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: The $l_{1}$ normalized attribute correlations based on $2K$ same samples generated by GAN inversion.", + "bbox": [ + 135, + 161, + 831, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9a110296133d286bd9fb2ac780c9061af04bfedba34a618129827073ae09371.jpg", + "image_caption": [ + "Pose Unambiguity (1st Layer 8th Direction)", + "Figure 5: Interpretable directions identified by our method are semantically consistent among different samples." + ], + "image_footnote": [], + "bbox": [ + 81, + 213, + 467, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ba7e1065ca5a3b630e7008a875742abbc5973ca09da1ea3d7a8e84935d8b07b.jpg", + "image_caption": [ + "Figure 6: Qualitative comparisons of the same samples." + ], + "image_footnote": [], + "bbox": [ + 81, + 397, + 467, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we choose to integrate our projector every few layers to obtain interpretable directions of different semantics.", + "bbox": [ + 76, + 678, + 468, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 722, + 209, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first introduce the experimental setup, followed by the quantitative and qualitative evaluation. We defer the full ablation studies to Sec. D of supplementary.", + "bbox": [ + 75, + 747, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 801, + 158, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We evaluate our Householder Projector on StyleGAN2 [28] and StyleGAN3 [26], i.e., the challenging state-of-the-art GAN backbones in the field of computer vision.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Baselines. For StyleGAN2, we conduct experiments on FFHQ [29], LSUN Church [63], and LSUN", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cat [63]. The experiments of StyleGAN3 are performed on SHHQ [13], MetFaces [25] and AFHQv2 [10]. We mainly compare our method with representative unsupervised latent semantics discovery approaches, including SeFa [48], Orthogonal Jacobian Regularization (OrJaR) [57], and Hessian Penalty (HP) [41]. SeFa [48] can be directly applied to pre-trained models, while OrJaR and HP need extra finetuning or training from scratch due to the regularization.", + "bbox": [ + 496, + 202, + 890, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We conduct quantitative evaluation using (1) Fréchet Inception Distance (FID) [19], (2) Perceptual Path Length (PPL) [27], (3) Perceptual Interpretable Path Length (PIPL), and (4) Face Attribute Correlation. FID aims to measure the image quality and diversity by computing the distance between the real and fake distributions, and PPL is designed to assess the perceptual smoothness of the latent space where the smoothness can reflect the disentanglement ability. Our proposed PIPL is a natural modification of PPL by adapting the latent manipulation from random interpolation-based perturbations to vector-based perturbations using interpretable directions. Compared with PPL, our PIPL can better measure the latent space smoothness when the latent code is perturbed along with specific vectors, which suits more vector-based semantic discovery methods like SeFa [48] and ours. Furthermore, for StyleGAN2 trained on FFHQ, we rely on pre-trained face attribute estimators to compute the correlation coefficient between the traversal steps and the face attributes. Besides the above four metrics, we also assess the disentanglement through visual observation. We defer the details of used datasets and metrics to Sec. C of the supplementary.", + "bbox": [ + 496, + 325, + 892, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the widely used Pytorch implementation of StyleGAN $^{1}$ and convert the official TensorFlow pre-trained models into PyTorch for FFHQ [29], LSUN Church [63], and LSUN Cat [63]. For StyleGAN3, we use the official code and pre-trained models of AFHQv2 [10] and MetFaces [25]. As for SHHQ [13], we also use the official pre-trained model $^{2}$ . Following the original optimization strategy, we finetune all the parameters of the pre-trained generator and discriminator within $1\\%$ of the total training steps (kimgs for StyleGAN3). For instance, the fine-tuning process takes $5K$ steps for StyleGAN2 with FFHQ and 250 kimgs for StyleGAN3 with AFHQ. The fine-tuning time is thus very limited due to the small number", + "bbox": [ + 496, + 660, + 892, + 857 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/rosinality/stylegan2-pytorch", + "bbox": [ + 517, + 875, + 774, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/stylegan-human/StyleGAN-Human", + "bbox": [ + 517, + 887, + 810, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7717", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c2452bad3806cbe8ec8196ecc12255fc2c002b8091e389610dfbff5b04f25f0.jpg", + "image_caption": [ + "Figure 7: Exemplary qualitative comparison of two different semantics on FFHQ [29] with StyleGAN2 [28]. Our Householder Projector can precisely control the image attributes without changing the face identity. The direction index denotes the index of eigenvectors." + ], + "image_footnote": [], + "bbox": [ + 147, + 88, + 836, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of training steps. To give concrete examples, fine-tuning models on FFHQ and AFHQ takes 1.5 and 2.5 hours, respectively. The rank $N$ of S is set to 10 for all experiments based on our empirical observation of the number of semantics of StyleGANs. Our editing strength is set the same as SeFa. We use 4 RTX A6000 GPUs for the training. For the comparison fairness, the baseline methods OrJaR [57] and HP [41] are also fine-tuned with the same steps.", + "bbox": [ + 75, + 622, + 472, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Evaluation and Discussion", + "text_level": 1, + "bbox": [ + 76, + 755, + 405, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diverse and Precise Attributes. Fig. 2 exhibits some semantic attributes discovered by our Householder Projector on all the datasets. Our method mines a diverse set of disentangled semantics, enabling a wide range of attribute manipulation. Take as an example the first row of attributes discovered on FFHQ [29]. The left columns present diverse high-level semantic concepts such as \"Pose\" and \"Shape\", while the right columns show low-level attributes such as", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f7aa6923f1a9e414e0b3ca5d99185d402a679afedd7f79f45238563461ed7a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
FFHQ [29]1024 × 1024SeFa [48]4.481579.760.227
OrJaR [57]4.51987.800.204
HP [41]4.66993.170.207
Ours4.40966.230.141
LSUN Church [63]256 × 256SeFa [48]4.61530.680.069
OrJaR [57]3.77474.770.065
HP [41]3.78486.930.058
Ours3.72457.520.030
LSUN Cat [63]256 × 256SeFa [48]8.37722.240.141
OrJaR [57]8.39562.980.134
HP [41]8.31573.710.136
Ours8.46526.260.057
", + "bbox": [ + 522, + 612, + 870, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Evaluation results on StyleGAN2.", + "bbox": [ + 552, + 787, + 836, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"Color\" and \"Lighting\". This hierarchy also meets the same trend of original StyleGANs. The manipulation of the diverse and highly-disentangled semantics would give users more precise control on the image generation process.", + "bbox": [ + 496, + 821, + 890, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Unambiguity. Importantly, our interpretable di", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7718", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e216ea49735de13264ad94fcd7daf3a3caa73e953b78fc5adc39b3698042a1fa.jpg", + "image_caption": [ + "Figure 8: Exemplary visual comparison of three different semantics on SHHQ [13] with StyleGAN3 [26]. Our method is able to mine more disentangled interpretable directions and have more precise control on the attributes. The direction index denotes the index of eigenvectors." + ], + "image_footnote": [], + "bbox": [ + 147, + 89, + 383, + 489 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/523d20aae279e8338ad53c7e694d8e8c631b504ab4bad992f05868296779643b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 88, + 604, + 491 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28d1586cd879d45ff75e34f3f35bb0b675c89b902698d6f280a13f8457caba7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 88, + 826, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rections are unambiguous to different samples. As shown in Fig. 5, the image variations manipulated by our discovered directions all correspond to the same semantic attribute, i.e., the head pose. The other non-target attributes are untouched, such as the background and face identity.", + "bbox": [ + 75, + 559, + 468, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison against Other Methods. Fig. 7 and Fig. 8 compare the latent traversal of some directions against other baselines on FFHQ [29] and SHHQ [13], respectively. All the methods can discover similar attributes in the same layer, such as the head length in Fig. 7 left. However, the baselines suffer from entangled semantics and the other attributes vary during the traversal, such as hairstyle for SeFA [48] and OrJaR [57], and expression for HP [41]. By contrast, our method is able to discover more precise semantics and preserve other non-target attributes.", + "bbox": [ + 75, + 638, + 472, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of Same Samples. To better compare the qualitative disentanglement performance, we further use a GAN inversion technique (PTI [45]) to create nearly the same images from FFHQ for different methods. Fig. 6 displays some qualitative comparisons. For the same samples, our method still has more precise attribute control.", + "bbox": [ + 75, + 791, + 468, + 882 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Local Editing Applications. With precise attribute control,", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our method is even able to edit local regions by simply performing latent traversal. Fig. 4 displays such a use case. Our method achieves very competitive performance against local editing methods such as the recent ReSeFa [66]. In addition, our method is free of extra bounding box as input.", + "bbox": [ + 496, + 559, + 890, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Please refer to Sec. E of the supplementary for more visualizations about comparisons on other datasets, and semantic diversity/unambiguity/hierarchy.", + "bbox": [ + 496, + 636, + 890, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 694, + 720, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "StyleGAN2 Results. Table 2 presents the quantitative evaluation results on FFHQ [29], LSUN Church [63] and LSUN Cat [63] datasets. Our proposed Householder Projector outperforms other baselines in terms of both PPL and PIPL. This demonstrates that our method has a smoother and more structured latent space, which corresponds to more disentangled representations. In particular, our approach surpasses SeFa [48] by a large margin, which indicates the benefit of enforcing low-rank orthogonality to the projection matrix. Compared with the soft orthogonality regularization used in OrJaR [57] and HP [41], the hard orthogonality of our projector also has an advantage in latent smoothness", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7719", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "due to the strict orthogonality preservation and the additional low-rank constraint. Moreover, our FID score is also very competitive, implying that our method could simultaneously improve the disentanglement performance while keeping the quality of generated images unharmed. For the attribute correlation, we also use GAN inversion to create a dataset of $2K$ identical images for each method. Table 1 compares the correlation results on FFHQ. Our method outperforms other unsupervised baselines and preserves the attribute well in particular.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4b7d145fd52a86c2f40a5131b9341d61a081f7cc260813a1e1d691d8fa302a56.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
MetFaces [25]1024 × 1024SeFa [48]15.335626.312.991
OrJaR [57]17.555754.443.700
HP [41]17.325323.273.465
Ours16.894192.520.099
AFHQv2 [10]512 × 512SeFa [48]4.402193.740.470
OrJaR [57]5.452103.470.440
HP [41]5.332198.460.463
Ours4.982052.380.070
SHHQ [13]512 × 256SeFa [48]2.541621.070.370
OrJaR [57]4.781614.560.245
HP [41]5.381648.740.216
Ours4.171549.010.119
", + "bbox": [ + 99, + 258, + 446, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Evaluation results on StyleGAN3.", + "bbox": [ + 130, + 446, + 413, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "StyleGAN3 Results. Table 3 compares the performance of our method against other baselines on MetFaces [25], AFHQv2 [10], and SHHQ [13] datasets. The results are very coherent with those on StyleGAN2: our Householder Projector improves the latent space smoothness without harming the image fidelity. Our method as well as the two baselines that involve fine-tuning have slightly worse FID than the original StyleGAN3. This might stem from the fact that due to the limited computational resources, our used batch size (16) is actually smaller than the original setting of StyleGAN3 (32). As revealed in the ablation study of Sec. D.2 of the supplementary, the hyper-parameter batch size has a substantial impact on FID. We expect that increasing the batch size would further boost the image quality of our method and lead to a more competitive FID score.", + "bbox": [ + 75, + 476, + 472, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 722, + 194, + 738 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper proposes a general and flexible low-rank orthogonal matrix representation coined as Householder Projector for unsupervised latent semantics discovery of generative models. The proposed method endows the projection matrix with low-rank orthogonality. This superiority helps pre-trained GANs to achieve precise and diverse semantics control within limited fine-tuning steps. Extensive experiments of StyleGANs on various benchmarks demonstrate that our method could simultaneously identify the disentangled attributes while maintaining image fidelity.", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 90, + 658, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9).", + "bbox": [ + 498, + 114, + 892, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 172, + 596, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM TOG, 2021. 2", + "[2] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 2020. 3", + "[3] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B Tenenbaum, William T Freeman, and Antonio Torralba. Gan dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2", + "[4] Christian Bischof and Charles Van Loan. The wy representation for products of householder matrices. SIAM Journal on Scientific and Statistical Computing, 8(1):s2-s13, 1987. 5", + "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2", + "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 2", + "[7] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. NeurIPS, 2016. 2", + "[8]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. ECCV, 2022. 2, 3", + "[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. *ICLR*, 2022. 2", + "[10] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 2, 4, 6, 9", + "[11] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In CVPR, 2020. 3", + "[12] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In CVPR, 2020. 3", + "[13] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. ECCV, 2022. 2, 4, 6, 8, 9", + "[14] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In ICCV, 2019. 1, 2, 3" + ], + "bbox": [ + 501, + 196, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7720", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 1, 2", + "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. *ICLR*, 2022. 2", + "[17] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. NeurIPS, 2020. 2, 3", + "[18] Zhenliang He, Meina Kan, and Shiguang Shan. Eigengan: Layer-wise eigen-learning for gans. In ICCV, 2021. 2, 3", + "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6", + "[20] Alston S Householder. Unitary triangularization of a nonsymmetric matrix. Journal of the ACM (JACM), 5(4):339-342, 1958. 3", + "[21] Ali Jahanian, Lucy Chai, and Phillip Isola. On the\"steerability\" of generative adversarial networks. In ICLR, 2020. 1, 2, 3", + "[22] Adarsh Kappiyath, Silpa Vadakkeeveetil Sreelatha, and S Sumitra. Self-supervised enhancement of latent discovery in gans. In AAAI, 2022. 2", + "[23] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Singh, and R Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In ECCV, 2022. 3", + "[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. *ICLR*, 2018. 2", + "[25] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2, 4, 6, 9", + "[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1, 2, 6, 8", + "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6", + "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 1, 2, 6, 7", + "[29] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In CVPR, 2014. 2, 4, 6, 7, 8", + "[30] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2", + "[31] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In CVPR, 2021. 2", + "[32] Richard B Lehoucq. The computation of elementary unitary matrices. ACM Transactions on Mathematical Software (TOMS), 22(4):393-400, 1996. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Minjun Li, Yanghua Jin, and Huachun Zhu. Surrogate gradient field for latent space manipulation. In CVPR, 2021. 3", + "[34] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. NeurIPS, 2021. 3", + "[35] Alexander Mathiasen, Frederik Hvilshøj, Jakob Rødsgaard Jørgensen, Anshul Nasery, and Davide Mottin. What if neural networks had svds? NeurIPS, 2020. 3", + "[36] Zakaria Mhammedi, Andrew Hellicar, Ashfaqur Rahman, and James Bailey. Efficient orthogonal parametrisation of recurrent neural networks using householder reflections. In ICML, 2017. 3", + "[37] James Oldfield, Markos Georgopoulos, Yannis Panagakis, Mihalis A Nicolaou, and Ioannis Patras. Tensor component analysis for interpreting the latent space of gans. BMVC, 2021. 2", + "[38] James Oldfield, Christos Tzelepis, Yannis Panagakis, Michaelis A Nicolaou, and Ioannis Patras. Panda: Unsupervised learning of parts and appearances in the feature maps of gans. *ICLR*, 2023. 3", + "[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2", + "[40] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3", + "[41] William Peebles, John Peebles, Jun-Yan Zhu, Alexei Efros, and Antonio Torralba. The hessian penalty: A weak prior for unsupervised disentanglement. In ECCV. Springer, 2020. 2, 3, 6, 7, 8, 9", + "[42] Antoine Plumerault, Hervé Le Borgne, and Céline Hudelot. Controlling generative models with continuous factors of variations. *ICLR*, 2020. 2, 3", + "[43] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. *ICLR*, 2016. 2", + "[44] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Learning disentangled representation by exploiting pretrained generative models: A contrastive learning view. In ICLR, 2022. 2", + "[45] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM TOG, 2022. 8", + "[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 2", + "[47] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2, 3", + "[48] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2, 3, 6, 7, 8, 9", + "[49] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7721", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2", + "[51] Yue Song, Andy Keller, Nicu Sebe, and Max Welling. Latent traversals in generative models as potential flows. In ICML. PMLR, 2023. 2, 3", + "[52] Yue Song, Nicu Sebe, and Wei Wang. Improving covariance conditioning of the svd meta-layer by orthogonality. In ECCV, 2022. 5", + "[53] Yue Song, Nicu Sebe, and Wei Wang. Orthogonal svd covariance conditioning and latent disentanglement. IEEE TPAMI, 2022. 2", + "[54] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. \"Gan\" steerability\" without optimization. *ICLR*, 2021. 2", + "[55] Christos Tzelepis, Georgios Tzimiropoulos, and Ioannis Pañas. Warpedganspace: Finding non-linear rbf paths in gan latent space. In ICCV, 2021. 2", + "[56] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, 2020. 2, 3", + "[57] Yuxiang Wei, Yupeng Shi, Xiao Liu, Zhilong Ji, Yuan Gao, Zhongqin Wu, and Wangmeng Zuo. Orthogonal jacobian regularization for unsupervised disentanglement in image generation. In ICCV, 2021. 2, 3, 6, 7, 8, 9", + "[58] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In CVPR, 2021. 2", + "[59] Jianjin Xu and Changxi Zheng. Linear semantics in generative adversarial networks. In CVPR, 2021. 3", + "[60] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, 2022. 2", + "[61] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 2, 3", + "[62] Huiting Yang, Liangyu Chai, Qiang Wen, Shuang Zhao, Zixun Sun, and Shengfeng He. Discovering interpretable latent space directions of gans beyond binary attributes. In CVPR, 2021. 3", + "[63] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 2, 4, 6, 7, 8", + "[64] Jiong Zhang, Qi Lei, and Inderjit Dhillon. Stabilizing gradients for deep neural networks via efficient svd parameterization. In ICML, 2018. 3", + "[65] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in gans. NeurIPS, 2021. 3", + "[66] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in gans. ICML, 2022. 3, 5, 8", + "[67] Xinqi Zhu, Chang Xu, and Dacheng Tao. Learning disentangled representations with latent variation predictability. In ECCV. Springer, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 896 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7722", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_model.json b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b67033a6b764c87ed837e28b37d83d4a4a347dcb --- /dev/null +++ b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Householder Projector for Unsupervised Latent Semantics Discovery" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.193, + 0.707, + 0.213 + ], + "angle": 0, + "content": "Yue Song\\(^{1}\\), Jichao Zhang\\(^{1}\\), Nicu Sebe\\(^{1}\\), and Wei Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.223, + 0.846, + 0.28 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Information Engineering and Computer Science, University of Trento, Italy \\(^{2}\\)Beijing Jiaotong University, China yue.song@unitn.it" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.337 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.778 + ], + "angle": 0, + "content": "Generative Adversarial Networks (GANs), especially the recent style-based generators (StyleGANs), have versatile semantics in the structured latent space. Latent semantics discovery methods emerge to move around the latent code such that only one factor varies during the traversal. Recently, an unsupervised method proposed a promising direction to directly use the eigenvectors of the projection matrix that maps latent codes to features as the interpretable directions. However, one overlooked fact is that the projection matrix is non-orthogonal and the number of eigenvectors is too large. The non-orthogonality would entangle semantic attributes in the top few eigenvectors, and the large dimensionality might result in meaningless variations among the directions even if the matrix is orthogonal. To avoid these issues, we propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix. The orthogonality guarantees that the eigenvectors correspond to disentangled interpretable semantics, while the low-rank property encourages that each identified direction has meaningful variations. We integrate our projector into pre-trained StyleGAN2/StyleGAN3 and evaluate the models on several benchmarks. Within only \\(1\\%\\) of the original training steps for fine-tuning, our projector helps StyleGANs to discover more disentangled and precise semantic attributes without sacrificing image fidelity. Code is publicly available via https://github.com/KingJamesSong/HouseholderGAN." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.21, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.833, + 0.47, + 0.865 + ], + "angle": 0, + "content": "Generative Adversarial Networks (GANs) [15] and the recent style-based generative models (StyleGANs) [27, 28," + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.324, + 0.891, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.778 + ], + "angle": 0, + "content": "Figure 1: Motivation of our proposed Householder Projector. Here \"Projector\" denotes the projection matrix that maps latent codes to features, i.e., the modulation weight of StyleGANs. (Top Left) The singular value imbalance of the non-orthogonal projector would entangle multiple semantics in the top interpretable directions. (Top Right) Due to the large dimensionality of the projector, directly enforcing vanilla orthogonality would spread the data variations among all the eigenvectors, leading to imperceptible and meaningless traversal. (Bottom) Our Householder Projector equips the projection matrix with low-rank orthogonal properties, which simultaneously disentangles semantics into multiple equally-important eigenvectors and guarantees that each direction could correspond to semanticallymeaningful variations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "[26] in particular have become the leading paradigm of generative modeling in the vision domain. The latent spaces of StyleGANs are known to embed rich and hierarchical semantics [14, 21]; moving the latent code in certain directions could trigger meaningful variations in the output images. Therefore, latent semantics discovery methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The first two authors contribute equally; Wei Wang is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7712" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "emerge to identify such interpretable directions that each variation factor is disentangled and the generation process can be precisely controlled [14, 47, 42, 61, 48, 1, 58, 37, 8]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.142, + 0.473, + 0.55 + ], + "angle": 0, + "content": "Among the recent explorations of unsupervised interpretable semantics discovery methods [56, 48, 57, 41], SeFa [48] pointed out a promising direction to discover semantically meaningful concepts by computing the eigenvector of the projector. Here we refer to the projection matrix that maps latent codes to features as the projector. The key observation is that using the eigenvectors of the projector for latent perturbation would maximize the data variations. Such identified eigenvectors/directions would correspond to meaningful semantic concepts. However, as shown in Fig. 1 top left, this is likely to cause semantics entangled in the top few eigenvectors. This phenomenon stems from the fact that the variation caused by an eigenvector is actually determined by the associated eigenvalue. Due to the imbalanced eigenvalue distribution, the discovered directions are not equally-important, and the top few ones would simultaneously manipulate multiple attributes. This eigenvalue discrepancy can be mitigated by enforcing orthogonality constraint to the projector. Nonetheless, since standard orthogonal matrices have as many equally-important eigenvectors as the dimensionality, there would not be enough semantics to mine in practice when the method scales to large models such as StyleGANs whose projector dimension is 1024 or 512. Consequently, as an accompanying limitation, the data variations would be split among all the eigenvectors and none of them could produce meaningful output changes (see Fig. 1 top right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To resolve the above issues, we propose Householder Projector, a low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix that maps latent codes to features. The projector is first decomposed to its SVD form \\((\\mathbf{USV}^T)\\). Next, the orthogonal singular vectors \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\) are represented by a series of Householder reflectors, respectively. Thanks to the normalization of Householder reflections, the orthogonality is also preserved during backpropagation. For the singular value \\(\\mathbf{S}\\), we explicitly set it as a low-rank identity matrix (i.e., \\(\\mathbf{S} = \\mathrm{diag}(1,\\dots ,1,0,\\dots ,0)\\)) whose rank defines exactly the number of semantic concepts. As shown in Fig. 1 bottom, the low-rank property guarantees that the identified directions would cause meaningful variations, while the orthogonality encourages that each semantic attribute is disentangled from the others. Moreover, a proper initialization scheme is proposed to leverage the statistics of pre-trained weights, and an acceleration technique is applied to speed up the computation. We also propose a metric dedicated to measuring the smoothness of latent space to interpretable directions based perturbations. Our Householder Projector is integrated into pre-trained StyleGANs [28, 26] at multiple different layers to mine the di" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.243 + ], + "angle": 0, + "content": "verse and hierarchical semantics. Since our projector inevitably changes the pre-trained parameters, the modified models incorporated with our projector are fine-tuned for limited steps to maintain the original image fidelity. Both quantitative and qualitative results on several widely used benchmarks [29, 63, 10, 25, 13] show that within marginal fine-tuning steps (1% of the training steps), our Householder Projector improves the latent semantics discovery of StyleGANs to have more precise attribute control while not impairing the quality of generated images." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.243, + 0.878, + 0.258 + ], + "angle": 0, + "content": "The key results and main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.268, + 0.892, + 0.329 + ], + "angle": 0, + "content": "- We propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projector of generative models for latent semantics discovery." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.339, + 0.892, + 0.43 + ], + "angle": 0, + "content": "- Our Householder Projector can be easily integrated into pre-trained GAN models. With the acceleration technique, it can be efficiently fine-tuned for very limited additional steps, which paves the way for applying latent semantics discovery and orthogonality techniques to large-scale generative models such as StyleGANs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.439, + 0.892, + 0.547 + ], + "angle": 0, + "content": "- Extensive experiments on two popular backbones (StyleGAN2 [28] and StyleGAN3 [26]) and six benchmarks (FFHQ [29], LSUN Church and Cat [63], MetFaces [25], AFHQ [10], and SHHQ [13]) demonstrate that within marginal extra fine-tuning steps (1% of the original training steps), our method could both achieve precise attribute control and preserve the original image quality." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.268, + 0.892, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.559, + 0.642, + 0.574 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Generative Adversarial Networks. In the past few years, GAN-based generative models [15] have achieved remarkable progress in high-fidelity image synthesis [43, 24, 5, 27, 28, 26, 25, 39, 6, 16, 50, 46]. The generation process usually takes the following procedure: a randomly-sampled latent code is fed into the generator through a projection step, and then the generator outputs realistic-looking images. Recently, the style-based generators [27, 28, 26] that gradually absorb layer-wise latent style codes are becoming the de facto GAN backbones. StyleGAN2 [28] improves the original StyleGAN [27] by redesigning generator normalization and training techniques, and StyleGAN3 [26] further explores some equivariance properties. We use the popular StyleGAN2 and StyleGAN3 as our backbones in this paper. Latent Semantics Discovery. Recently, lots of methods explore disentangling the latent space to achieve image editing by moving the latent code in the identified interpretable directions [7, 3, 21, 47, 17, 56, 67, 48, 18, 41, 54, 57, 55, 30, 55, 31, 60, 22, 44, 9, 53, 51]. Supervised methods rely on human annotations (i.e., segmentation masks, attribute categories, 3D priors, and text descriptions) to de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7713" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.544 + ], + "angle": 0, + "content": "fine the semantic labels [14, 47, 61, 21, 42, 33, 8, 12, 49, 40, 34, 62, 59, 23]. Here we mainly highlight some relevant unsupervised approaches that are free of annotations. Voynov et al. [56] proposed to jointly learn a set of directions and a classifier such that the interpretable directions can be recognized by the classifier. More recently, Peebles et al. [41] and Wei et al. [57] proposed to add orthogonal Hessian and Jacobian regularization to encourage disentangled representations, respectively. Song et al. [51] proposed to use wave equations to model the spatiotemporal dynamic non-linear latent traversals in generative models. SeFa [48] showed that the eigenvectors of the projector after the latent code could maximize the data variations and proposed to directly use them as the interpretable directions. However, due to the imbalanced eigenvalues, the image attributes would be entangled in the top few eigenvectors. Our proposed Householder Projector solved this issue by parameterizing the projector to a low-rank but orthogonal matrix. Notice that the used orthogonality technique in [56] is different from ours. In [56], the authors use matrix exponential \\(\\exp (\\mathbf{A} - \\mathbf{A}^T)\\) to generate skew orthogonal matrices where the skew-symmetry could limit the representation power. Also, the technique cannot parameterize given matrices and cannot explicitly control the rank. Our Householder representation is more general and flexible, allowing for controllable rank and parameterization of given matrices. Compared with the soft orthogonality regularization used in [41, 57, 18], the hard orthogonality of our method helps the model to learn more uncorrelated attributes within less fine-tuning steps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.473, + 0.654 + ], + "angle": 0, + "content": "In contrast to global editing approaches discussed above, some methods can perform local image editing in a post hoc way: they first define or learn a segment of regions of interests, and then manipulate the masked intermediate features for local editing [2, 11, 65, 66, 38]. Empowered by the precise control of attributes, our method can also achieve competitive performance in local image editing (see Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.295, + 0.687 + ], + "angle": 0, + "content": "3. Householder Projector" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.696, + 0.472, + 0.758 + ], + "angle": 0, + "content": "This section starts with the preliminary introduction of the closed-form latent semantics discovery. Next, we analyze its inherent limitation on entangled semantics and then illustrate our proposed Householder Projector in detail." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.45, + 0.788 + ], + "angle": 0, + "content": "3.1. Preliminary: Closed-form Latent Discovery" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Previous latent semantics discovery approaches [14, 56, 17, 47, 48] consider the GAN manipulation as \\( \\text{edit}(G(\\mathbf{z})) = G(\\mathbf{z} + \\alpha \\mathbf{n}) \\) where \\( G(\\cdot) \\) represents the generator, \\( \\mathbf{z} \\in \\mathbb{R}^d \\) denotes the latent code of dimension \\( d \\), \\( \\mathbf{n} \\in \\mathbb{R}^d \\) is the identified semantically meaningful direction, and \\( \\alpha \\) represents the perturbation strength. If one views the GAN as a multi-step projection function, the first projection can" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.613, + 0.107 + ], + "angle": 0, + "content": "be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.11, + 0.892, + 0.126 + ], + "angle": 0, + "content": "\\[\nG _ {1} (\\mathbf {z} + \\alpha \\mathbf {n}) = \\mathbf {A} \\mathbf {z} + \\mathbf {b} + \\alpha \\mathbf {A} \\mathbf {n}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.129, + 0.894, + 0.25 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\) and \\(\\mathbf{b}\\) denote the weight and bias of the projection step (e.g., convolution or linear transform), respectively. As can be seen from eq. (1), the resultant manipulation depends on the term \\(\\alpha \\mathbf{A}\\mathbf{n}\\). Intuitively, an interpretable direction \\(\\mathbf{n}\\) should cause large variations of \\(G_{1}(\\mathbf{z} + \\alpha \\mathbf{n})\\), which is equivalent to maximizing the impact of \\(\\alpha \\mathbf{A}\\mathbf{n}\\). Motivated by this observation, SeFa [48] proposed to consider the formulation as the following constrained optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.252, + 0.892, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} ^ {\\star} = \\arg \\max | | \\mathbf {A n} | | _ {2} ^ {2} s. t. \\mathbf {n} ^ {T} \\mathbf {n} = 1, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.895, + 0.516 + ], + "angle": 0, + "content": "where the constraint \\(\\mathbf{n}^T\\mathbf{n} = 1\\) is set to satisfy vector orthogonality, and \\(||\\cdot ||\\) denotes the \\(l_{2}\\) norm. Introducing a Lagrange multiplier \\(\\lambda\\) leads to \\(2\\mathbf{A}^{T}\\mathbf{A}\\mathbf{n} - 2\\lambda \\mathbf{n} = 0\\). The closed-form solutions all correspond to the eigenvectors of \\(\\mathbf{A}^T\\mathbf{A}\\). This presents a promising approach to identify the semantics by exploiting the projector \\(\\mathbf{A}\\) that projects latent codes. However, one fact overlooked by [48] is that the eigenvectors would cause different extents of variations due to the discrepancy of associated eigenvalues. Supposing that \\(\\mathbf{n}\\) is an eigenvector of \\(\\mathbf{A}^T\\mathbf{A}\\), then we would have \\(||\\mathbf{A}\\mathbf{n}||_2^2 = \\sigma^2\\) where \\(\\sigma\\) is the corresponding singular value of \\(\\mathbf{A}\\). For non-orthogonal matrices, the singular values are exponentially decreasing, i.e., \\(\\sigma_1\\geq \\sigma_2\\geq \\ldots \\geq \\sigma_d\\). This would cause most variations captured by the first few interpretable directions. The imbalance is thus likely to make semantic attributes entangled in the top eigenvectors (see Fig. 1 top left)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.887, + 0.54 + ], + "angle": 0, + "content": "3.2. Householder Low-rank Orthogonal Projector" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.667 + ], + "angle": 0, + "content": "The eigenvalue discrepancy can be eliminated by enforcing strict orthogonality. Orthogonal matrices have the property of identical eigenvalues, which assigns equal importance to each eigenvector. The low-rank constraint could further limit the number of semantics to mine. Therefore, we propose to use Householder representation, a flexible and general framework to parameterize matrices, to endow the projection matrix with low-rank orthogonality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.894, + 0.789 + ], + "angle": 0, + "content": "Householder Parameterization. Householder representations can parameterize any matrices by using a series of Householder reflectors to represent the orthogonal singular vectors of its Singular Value Decomposition (SVD) form. In the field of deep learning, it has been used to parameterize the transition matrix and to stabilize gradients of recurrent neural networks [36, 64, 35]. The key to the orthogonality representation relies on the following theorem:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.88 + ], + "angle": 0, + "content": "Theorem 1 (Householder representation [20, 32]). Given any square orthogonal matrix \\(\\mathbf{M} \\in \\mathbb{R}^{m \\times m}\\), it can be represented by the product of Householder matrices \\(\\mathbf{M} = \\mathbf{H}_1\\mathbf{H}_2 \\ldots \\mathbf{H}_m\\) where each Householder matrix is parameterized by a vector as \\(\\mathbf{H}_i = \\mathbf{I} - 2\\frac{\\mathbf{h}_i\\mathbf{h}_i^T}{||\\mathbf{h}_i||_2^2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Let \\(\\mathbf{USV}^T\\) denote the SVD of the projector \\(\\mathbf{A}\\) where \\(\\mathbf{S}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7714" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.885, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.837, + 0.893, + 0.884 + ], + "angle": 0, + "content": "Figure 2: Gallery of some semantic attributes discovered by our Householder Projector across all used datasets (FFHQ [29] in the top row, LSUN Church [63] and LSUN Cat [63] in the \\(2_{nd}\\) row, SHHQ [13] in the \\(3_{rd}\\) row, and MetFaces [25] and AFHQ [10] in the bottom row). These semantic attributes are sorted from low-level layers (left) to high-level layers (right)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7715" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "denotes the diagonal singular value, and \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\) are left and right orthogonal singular vectors. We use the accumulation of Householder reflectors (i.e., \\(\\prod_{i=0}^{w}\\left(\\mathbf{I}-2\\frac{\\mathbf{u}_i\\mathbf{u}_i^T}{||\\mathbf{u}_i||_2^2}\\right)\\) and \\(\\prod_{i=0}^{h}\\left(\\mathbf{I}-2\\frac{\\mathbf{v}_i\\mathbf{v}_i^T}{||\\mathbf{v}_i||_2^2}\\right)\\) where \\(w,h\\) denotes the width and height of \\(\\mathbf{A}\\)) to parameterize \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\), respectively. Notice that only \\(\\mathbf{u}_i\\) and \\(\\mathbf{v}_i\\) are the actual learnable parameters. For \\(\\mathbf{S}\\), we explicitly set it to a diagonal matrix and keep the weight fixed. The projector is thus represented by our Householder parameterizations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Low-rank Constraint. To achieve the property of disentangled attributes, one straightforward approach is to parameterize \\(\\mathbf{A}\\) as an orthogonal matrix, i.e., to set \\(\\mathbf{S}\\) to an identity matrix \\(\\mathbf{I}\\) where \\(\\mathbf{I}_{i,j} = 1\\) for \\(i = j\\) and \\(\\mathbf{I}_{i,j} = 0\\) otherwise. This would lead to equally-important semantic attributes whose number is exactly the projector dimension. However, for large generative models such as StyleGANs, the projector dimension is typically very large (e.g., 512 or 1024). It is not likely to have enough attributes to edit in practice. Setting \\(\\mathbf{S}\\) to a full-rank diagonal matrix would spread data variations among all the eigenvectors, resulting in trivial and imperceptible traversal (see Fig. 1 top right). To avoid this issue, we propose to define \\(\\mathbf{S}\\) as a low-rank identity matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.458, + 0.47, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\operatorname {d i a g} (\\underbrace {1 , \\dots , 1} _ {N}, 0, \\dots , 0), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.576 + ], + "angle": 0, + "content": "where \\( N \\) defines the rank and also the number of semantic attributes to mine. By restricting the rank of the projector, we explicitly limit the number of interpretable directions. As shown in Fig. 1 bottom, this would benefit the latent traversal for more meaningful output variations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Orthogonality Preservation. One advantage of our Householder representation is that the orthogonality can be kept during backpropagation. Since we have the vector normalization \\(\\frac{\\mathbf{h}_i\\mathbf{h}_i^T}{||\\mathbf{h}_i||_2^2}\\), the impact of any gradient descent step \\(\\mathbf{h}_i - \\eta \\nabla \\mathbf{h}_i\\) on the orthogonality would be cancelled, i.e., the updated vector remains orthogonal after normalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.886 + ], + "angle": 0, + "content": "Initialization. When our method is applied to pre-trained models, the well-trained network weights could be leveraged to initialize our Householder Projection. To this end, we propose to use the nearest-orthogonal mapping [52] to project the weight matrix into its orthogonal form that has the nearest distance in the Frobenius norm (i.e., \\(\\min_{\\mathbf{R}}||\\mathbf{R} - \\mathbf{A}||_{\\mathrm{F}}\\) where \\(\\mathbf{R}^T\\mathbf{R} = \\mathbf{I}\\)). The closed-form solution is given by \\(\\mathbf{U}\\mathbf{V}^T\\) where \\(\\mathbf{USV}^T\\) is the SVD of the original weight matrix \\(\\mathbf{A}\\). Next, we decompose \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\) into their Householder reflectors and use them to initialize our projector. Such an initialization scheme leverages the statistics of the original weight matrix, which might give our projector a good starting point and improve the performance (see the ablation study of Sec. D.3 in the supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Acceleration. The accumulated product of elementary" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.891, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.189, + 0.895, + 0.342 + ], + "angle": 0, + "content": "Figure 3: Illustration on how our Householder Projector represents the modulation weight \\(\\mathbf{A}\\) of StyleGANs. Here \"Demod\", \"EMA\", and \"FN\" denote Demodulation, Exponential Moving Average, and Filtered Non-linearities, respectively. The projector is parameterized by its SVD form where \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\) are represented by the accumulation of Householder reflectors, and \\(\\mathbf{S}\\) is set to a low-rank identity matrix. Our projector is applied at multiple different layers of StyleGANs to explore the diverse and hierarchical semantics. The actual learnable parameters are \\(\\mathbf{u}_i\\) and \\(\\mathbf{v}_i\\)." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.348, + 0.891, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.458, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 4: Comparison with ReSeFa [66]. The blue lines indicate the specific regions changed by our method, and the red box indicates the region of interest that is needed as input to ReSeFa [66]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.537, + 0.882, + 0.552 + ], + "angle": 0, + "content": "Householder matrices can be accelerated via the theorem:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Theorem 2 (Compact WY representation [4]). For any accumulation of \\( m \\) Householder matrices \\( \\mathbf{H}_1 \\ldots \\mathbf{H}_m \\), there exists \\( \\mathbf{W}, \\mathbf{Y} \\in \\mathbb{R}^{d \\times m} \\) such that \\( \\mathbf{I} - 2\\mathbf{W}\\mathbf{Y}^T = \\mathbf{H}_1 \\ldots \\mathbf{H}_m \\) where computing \\( \\mathbf{W} \\) and \\( \\mathbf{Y} \\) takes \\( O(dm^2) \\) time and \\( m \\) sequential Householder multiplications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.765 + ], + "angle": 0, + "content": "This theorem indicates the possibility to repeatedly split the accumulation \\(\\mathbf{H}_1\\mathbf{H}_2\\dots \\mathbf{H}_m\\) into multiple subsequences until irreducible. Then these sub-sequences can be computed in parallel and gradually merged. As revealed in the ablation of Sec. D.3 in the supplementary, this technique could fully exploit the parallel computational power of GPUs and greatly speed up the aggregation, particularly in our case where the projector dimension is large." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Implementation in StyleGANs. Fig. 3 depicts how our Householder Projector modifies the original StyleGAN architectures. The projector used in the weight modulation module is represented by our proposed projector. The weight matrix is thus endowed with low-rank orthogonal properties. We insert the proposed projector at every layer of StyleGAN2 and every four layers of StyleGAN3. Since StyleGAN3 has 15 intermediate layers, the adjacent layers have very similar and even repeated semantics. Therefore," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7716" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.28, + 0.143 + ], + "angle": 0, + "content": "
IdentityPoseAgeGenderGlassesSmile
Identity0.650.240.030.040.010.03
Pose0.110.570.040.040.110.13
Age0.020.050.670.030.190.03
Gender0.030.320.020.520.100.00
Glasses0.010.080.000.020.880.01
Smile0.020.010.010.000.010.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.146, + 0.237, + 0.158 + ], + "angle": 0, + "content": "(a) Our method" + }, + { + "type": "table", + "bbox": [ + 0.287, + 0.089, + 0.48, + 0.143 + ], + "angle": 0, + "content": "
IdentityPoseAgeGenderGlassesSmile
Identity0.560.060.050.020.050.26
Pose0.440.480.050.010.010.00
Age0.430.270.220.050.030.01
Gender0.330.180.040.400.040.02
Glasses0.270.140.040.100.230.22
Smile0.200.080.090.000.030.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.356, + 0.146, + 0.413, + 0.159 + ], + "angle": 0, + "content": "(b) SeFa" + }, + { + "type": "table", + "bbox": [ + 0.485, + 0.089, + 0.677, + 0.143 + ], + "angle": 0, + "content": "
IdentityPoseAgeGenderGlassesSmile
Identity0.510.270.040.050.030.10
Pose0.390.350.050.020.070.11
Age0.260.140.320.100.050.12
Gender0.200.060.340.290.040.07
Glasses0.180.060.120.070.420.14
Smile0.130.050.070.020.030.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.146, + 0.615, + 0.159 + ], + "angle": 0, + "content": "(c) OrJaR" + }, + { + "type": "table", + "bbox": [ + 0.683, + 0.089, + 0.877, + 0.143 + ], + "angle": 0, + "content": "
IdentityPoseAgeGenderGlassesSmile
Identity0.540.250.040.060.020.08
Pose0.420.300.080.040.050.12
Age0.210.150.280.110.080.17
Gender0.270.070.230.250.040.13
Glasses0.150.120.160.040.370.16
Smile0.110.090.120.010.010.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.758, + 0.146, + 0.804, + 0.159 + ], + "angle": 0, + "content": "(d) HP" + }, + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.162, + 0.833, + 0.177 + ], + "angle": 0, + "content": "Table 1: The \\( l_{1} \\) normalized attribute correlations based on \\( 2K \\) same samples generated by GAN inversion." + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.203, + 0.402, + 0.214 + ], + "angle": 0, + "content": "Pose Unambiguity (1st Layer 8th Direction)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.214, + 0.468, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.352, + 0.47, + 0.384 + ], + "angle": 0, + "content": "Figure 5: Interpretable directions identified by our method are semantically consistent among different samples." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.398, + 0.468, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.638, + 0.456, + 0.653 + ], + "angle": 0, + "content": "Figure 6: Qualitative comparisons of the same samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.679, + 0.47, + 0.71 + ], + "angle": 0, + "content": "we choose to integrate our projector every few layers to obtain interpretable directions of different semantics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.21, + 0.74 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.794 + ], + "angle": 0, + "content": "In this section, we first introduce the experimental setup, followed by the quantitative and qualitative evaluation. We defer the full ablation studies to Sec. D of supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.159, + 0.818 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Models. We evaluate our Householder Projector on StyleGAN2 [28] and StyleGAN3 [26], i.e., the challenging state-of-the-art GAN backbones in the field of computer vision." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets and Baselines. For StyleGAN2, we conduct experiments on FFHQ [29], LSUN Church [63], and LSUN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.203, + 0.892, + 0.324 + ], + "angle": 0, + "content": "Cat [63]. The experiments of StyleGAN3 are performed on SHHQ [13], MetFaces [25] and AFHQv2 [10]. We mainly compare our method with representative unsupervised latent semantics discovery approaches, including SeFa [48], Orthogonal Jacobian Regularization (OrJaR) [57], and Hessian Penalty (HP) [41]. SeFa [48] can be directly applied to pre-trained models, while OrJaR and HP need extra finetuning or training from scratch due to the regularization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.893, + 0.659 + ], + "angle": 0, + "content": "Metrics. We conduct quantitative evaluation using (1) Fréchet Inception Distance (FID) [19], (2) Perceptual Path Length (PPL) [27], (3) Perceptual Interpretable Path Length (PIPL), and (4) Face Attribute Correlation. FID aims to measure the image quality and diversity by computing the distance between the real and fake distributions, and PPL is designed to assess the perceptual smoothness of the latent space where the smoothness can reflect the disentanglement ability. Our proposed PIPL is a natural modification of PPL by adapting the latent manipulation from random interpolation-based perturbations to vector-based perturbations using interpretable directions. Compared with PPL, our PIPL can better measure the latent space smoothness when the latent code is perturbed along with specific vectors, which suits more vector-based semantic discovery methods like SeFa [48] and ours. Furthermore, for StyleGAN2 trained on FFHQ, we rely on pre-trained face attribute estimators to compute the correlation coefficient between the traversal steps and the face attributes. Besides the above four metrics, we also assess the disentanglement through visual observation. We defer the details of used datasets and metrics to Sec. C of the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.893, + 0.858 + ], + "angle": 0, + "content": "Implementation Details. We adopt the widely used Pytorch implementation of StyleGAN\\(^{1}\\) and convert the official TensorFlow pre-trained models into PyTorch for FFHQ [29], LSUN Church [63], and LSUN Cat [63]. For StyleGAN3, we use the official code and pre-trained models of AFHQv2 [10] and MetFaces [25]. As for SHHQ [13], we also use the official pre-trained model\\(^{2}\\). Following the original optimization strategy, we finetune all the parameters of the pre-trained generator and discriminator within \\(1\\%\\) of the total training steps (kimgs for StyleGAN3). For instance, the fine-tuning process takes \\(5K\\) steps for StyleGAN2 with FFHQ and 250 kimgs for StyleGAN3 with AFHQ. The fine-tuning time is thus very limited due to the small number" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.875, + 0.775, + 0.888 + ], + "angle": 0, + "content": "1https://github.com/rosinality/stylegan2-pytorch" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.812, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/stylegan-human/StyleGAN-Human" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.875, + 0.812, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7717" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.089, + 0.837, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.565, + 0.895, + 0.611 + ], + "angle": 0, + "content": "Figure 7: Exemplary qualitative comparison of two different semantics on FFHQ [29] with StyleGAN2 [28]. Our Householder Projector can precisely control the image attributes without changing the face identity. The direction index denotes the index of eigenvectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.473, + 0.745 + ], + "angle": 0, + "content": "of training steps. To give concrete examples, fine-tuning models on FFHQ and AFHQ takes 1.5 and 2.5 hours, respectively. The rank \\( N \\) of S is set to 10 for all experiments based on our empirical observation of the number of semantics of StyleGANs. Our editing strength is set the same as SeFa. We use 4 RTX A6000 GPUs for the training. For the comparison fairness, the baseline methods OrJaR [57] and HP [41] are also fine-tuned with the same steps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.406, + 0.772 + ], + "angle": 0, + "content": "4.2. Qualitative Evaluation and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Diverse and Precise Attributes. Fig. 2 exhibits some semantic attributes discovered by our Householder Projector on all the datasets. Our method mines a diverse set of disentangled semantics, enabling a wide range of attribute manipulation. Take as an example the first row of attributes discovered on FFHQ [29]. The left columns present diverse high-level semantic concepts such as \"Pose\" and \"Shape\", while the right columns show low-level attributes such as" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.613, + 0.872, + 0.778 + ], + "angle": 0, + "content": "
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
FFHQ [29]1024 × 1024SeFa [48]4.481579.760.227
OrJaR [57]4.51987.800.204
HP [41]4.66993.170.207
Ours4.40966.230.141
LSUN Church [63]256 × 256SeFa [48]4.61530.680.069
OrJaR [57]3.77474.770.065
HP [41]3.78486.930.058
Ours3.72457.520.030
LSUN Cat [63]256 × 256SeFa [48]8.37722.240.141
OrJaR [57]8.39562.980.134
HP [41]8.31573.710.136
Ours8.46526.260.057
" + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.789, + 0.838, + 0.804 + ], + "angle": 0, + "content": "Table 2: Evaluation results on StyleGAN2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.823, + 0.892, + 0.883 + ], + "angle": 0, + "content": "\"Color\" and \"Lighting\". This hierarchy also meets the same trend of original StyleGANs. The manipulation of the diverse and highly-disentangled semantics would give users more precise control on the image generation process." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Semantic Unambiguity. Importantly, our interpretable di" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7718" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.09, + 0.385, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.089, + 0.605, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.089, + 0.828, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.501, + 0.896, + 0.548 + ], + "angle": 0, + "content": "Figure 8: Exemplary visual comparison of three different semantics on SHHQ [13] with StyleGAN3 [26]. Our method is able to mine more disentangled interpretable directions and have more precise control on the attributes. The direction index denotes the index of eigenvectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.636 + ], + "angle": 0, + "content": "rections are unambiguous to different samples. As shown in Fig. 5, the image variations manipulated by our discovered directions all correspond to the same semantic attribute, i.e., the head pose. The other non-target attributes are untouched, such as the background and face identity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.473, + 0.789 + ], + "angle": 0, + "content": "Comparison against Other Methods. Fig. 7 and Fig. 8 compare the latent traversal of some directions against other baselines on FFHQ [29] and SHHQ [13], respectively. All the methods can discover similar attributes in the same layer, such as the head length in Fig. 7 left. However, the baselines suffer from entangled semantics and the other attributes vary during the traversal, such as hairstyle for SeFA [48] and OrJaR [57], and expression for HP [41]. By contrast, our method is able to discover more precise semantics and preserve other non-target attributes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.47, + 0.883 + ], + "angle": 0, + "content": "Comparison of Same Samples. To better compare the qualitative disentanglement performance, we further use a GAN inversion technique (PTI [45]) to create nearly the same images from FFHQ for different methods. Fig. 6 displays some qualitative comparisons. For the same samples, our method still has more precise attribute control." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Local Editing Applications. With precise attribute control," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.636 + ], + "angle": 0, + "content": "our method is even able to edit local regions by simply performing latent traversal. Fig. 4 displays such a use case. Our method achieves very competitive performance against local editing methods such as the recent ReSeFa [66]. In addition, our method is free of extra bounding box as input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Please refer to Sec. E of the supplementary for more visualizations about comparisons on other datasets, and semantic diversity/unambiguity/hierarchy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.722, + 0.712 + ], + "angle": 0, + "content": "4.3. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "StyleGAN2 Results. Table 2 presents the quantitative evaluation results on FFHQ [29], LSUN Church [63] and LSUN Cat [63] datasets. Our proposed Householder Projector outperforms other baselines in terms of both PPL and PIPL. This demonstrates that our method has a smoother and more structured latent space, which corresponds to more disentangled representations. In particular, our approach surpasses SeFa [48] by a large margin, which indicates the benefit of enforcing low-rank orthogonality to the projection matrix. Compared with the soft orthogonality regularization used in OrJaR [57] and HP [41], the hard orthogonality of our projector also has an advantage in latent smoothness" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7719" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "due to the strict orthogonality preservation and the additional low-rank constraint. Moreover, our FID score is also very competitive, implying that our method could simultaneously improve the disentanglement performance while keeping the quality of generated images unharmed. For the attribute correlation, we also use GAN inversion to create a dataset of \\(2K\\) identical images for each method. Table 1 compares the correlation results on FFHQ. Our method outperforms other unsupervised baselines and preserves the attribute well in particular." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.259, + 0.447, + 0.438 + ], + "angle": 0, + "content": "
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
MetFaces [25]1024 × 1024SeFa [48]15.335626.312.991
OrJaR [57]17.555754.443.700
HP [41]17.325323.273.465
Ours16.894192.520.099
AFHQv2 [10]512 × 512SeFa [48]4.402193.740.470
OrJaR [57]5.452103.470.440
HP [41]5.332198.460.463
Ours4.982052.380.070
SHHQ [13]512 × 256SeFa [48]2.541621.070.370
OrJaR [57]4.781614.560.245
HP [41]5.381648.740.216
Ours4.171549.010.119
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.448, + 0.415, + 0.463 + ], + "angle": 0, + "content": "Table 3: Evaluation results on StyleGAN3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.473, + 0.704 + ], + "angle": 0, + "content": "StyleGAN3 Results. Table 3 compares the performance of our method against other baselines on MetFaces [25], AFHQv2 [10], and SHHQ [13] datasets. The results are very coherent with those on StyleGAN2: our Householder Projector improves the latent space smoothness without harming the image fidelity. Our method as well as the two baselines that involve fine-tuning have slightly worse FID than the original StyleGAN3. This might stem from the fact that due to the limited computational resources, our used batch size (16) is actually smaller than the original setting of StyleGAN3 (32). As revealed in the ablation study of Sec. D.2 of the supplementary, the hyper-parameter batch size has a substantial impact on FID. We expect that increasing the batch size would further boost the image quality of our method and lead to a more competitive FID score." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.196, + 0.739 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "This paper proposes a general and flexible low-rank orthogonal matrix representation coined as Householder Projector for unsupervised latent semantics discovery of generative models. The proposed method endows the projection matrix with low-rank orthogonality. This superiority helps pre-trained GANs to achieve precise and diverse semantics control within limited fine-tuning steps. Extensive experiments of StyleGANs on various benchmarks demonstrate that our method could simultaneously identify the disentangled attributes while maintaining image fidelity." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.66, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.893, + 0.162 + ], + "angle": 0, + "content": "This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.173, + 0.597, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.198, + 0.893, + 0.253 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.254, + 0.894, + 0.31 + ], + "angle": 0, + "content": "[2] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.311, + 0.894, + 0.366 + ], + "angle": 0, + "content": "[3] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B Tenenbaum, William T Freeman, and Antonio Torralba. Gan dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.367, + 0.894, + 0.409 + ], + "angle": 0, + "content": "[4] Christian Bischof and Charles Van Loan. The wy representation for products of householder matrices. SIAM Journal on Scientific and Statistical Computing, 8(1):s2-s13, 1987. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.41, + 0.893, + 0.45 + ], + "angle": 0, + "content": "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.451, + 0.893, + 0.521 + ], + "angle": 0, + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.522, + 0.893, + 0.577 + ], + "angle": 0, + "content": "[7] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.578, + 0.893, + 0.62 + ], + "angle": 0, + "content": "[8]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.621, + 0.893, + 0.675 + ], + "angle": 0, + "content": "[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. *ICLR*, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.676, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[10] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 2, 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[11] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.76, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[12] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[13] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. ECCV, 2022. 2, 4, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[14] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In ICCV, 2019. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.198, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7720" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. *ICLR*, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[17] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[18] Zhenliang He, Meina Kan, and Shiguang Shan. Eigengan: Layer-wise eigen-learning for gans. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.468, + 0.359 + ], + "angle": 0, + "content": "[20] Alston S Householder. Unitary triangularization of a nonsymmetric matrix. Journal of the ACM (JACM), 5(4):339-342, 1958. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[21] Ali Jahanian, Lucy Chai, and Phillip Isola. On the\"steerability\" of generative adversarial networks. In ICLR, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[22] Adarsh Kappiyath, Silpa Vadakkeeveetil Sreelatha, and S Sumitra. Self-supervised enhancement of latent discovery in gans. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.468, + 0.503 + ], + "angle": 0, + "content": "[23] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Singh, and R Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. *ICLR*, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.468, + 0.599 + ], + "angle": 0, + "content": "[25] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2, 4, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.603, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.468, + 0.686 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.468, + 0.73 + ], + "angle": 0, + "content": "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.732, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[29] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In CVPR, 2014. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[30] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[31] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[32] Richard B Lehoucq. The computation of elementary unitary matrices. ACM Transactions on Mathematical Software (TOMS), 22(4):393-400, 1996. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[33] Minjun Li, Yanghua Jin, and Huachun Zhu. Surrogate gradient field for latent space manipulation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[34] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[35] Alexander Mathiasen, Frederik Hvilshøj, Jakob Rødsgaard Jørgensen, Anshul Nasery, and Davide Mottin. What if neural networks had svds? NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[36] Zakaria Mhammedi, Andrew Hellicar, Ashfaqur Rahman, and James Bailey. Efficient orthogonal parametrisation of recurrent neural networks using householder reflections. In ICML, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[37] James Oldfield, Markos Georgopoulos, Yannis Panagakis, Mihalis A Nicolaou, and Ioannis Patras. Tensor component analysis for interpreting the latent space of gans. BMVC, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[38] James Oldfield, Christos Tzelepis, Yannis Panagakis, Michaelis A Nicolaou, and Ioannis Patras. Panda: Unsupervised learning of parts and appearances in the feature maps of gans. *ICLR*, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[40] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[41] William Peebles, John Peebles, Jun-Yan Zhu, Alexei Efros, and Antonio Torralba. The hessian penalty: A weak prior for unsupervised disentanglement. In ECCV. Springer, 2020. 2, 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[42] Antoine Plumerault, Hervé Le Borgne, and Céline Hudelot. Controlling generative models with continuous factors of variations. *ICLR*, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[43] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. *ICLR*, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[44] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Learning disentangled representation by exploiting pretrained generative models: A contrastive learning view. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[45] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM TOG, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[47] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[48] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2, 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.898 + ], + "angle": 0, + "content": "[49] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7721" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.148, + 0.468, + 0.188 + ], + "angle": 0, + "content": "[51] Yue Song, Andy Keller, Nicu Sebe, and Max Welling. Latent traversals in generative models as potential flows. In ICML. PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.19, + 0.468, + 0.231 + ], + "angle": 0, + "content": "[52] Yue Song, Nicu Sebe, and Wei Wang. Improving covariance conditioning of the svd meta-layer by orthogonality. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.232, + 0.468, + 0.271 + ], + "angle": 0, + "content": "[53] Yue Song, Nicu Sebe, and Wei Wang. Orthogonal svd covariance conditioning and latent disentanglement. IEEE TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.273, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[54] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. \"Gan\" steerability\" without optimization. *ICLR*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.301, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[55] Christos Tzelepis, Georgios Tzimiropoulos, and Ioannis Pañas. Warpedganspace: Finding non-linear rbf paths in gan latent space. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.343, + 0.468, + 0.383 + ], + "angle": 0, + "content": "[56] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.384, + 0.468, + 0.438 + ], + "angle": 0, + "content": "[57] Yuxiang Wei, Yupeng Shi, Xiao Liu, Zhilong Ji, Yuan Gao, Zhongqin Wu, and Wangmeng Zuo. Orthogonal jacobian regularization for unsupervised disentanglement in image generation. In ICCV, 2021. 2, 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.44, + 0.468, + 0.48 + ], + "angle": 0, + "content": "[58] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.481, + 0.468, + 0.508 + ], + "angle": 0, + "content": "[59] Jianjin Xu and Changxi Zheng. Linear semantics in generative adversarial networks. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.509, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[60] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.565, + 0.468, + 0.605 + ], + "angle": 0, + "content": "[61] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.606, + 0.468, + 0.66 + ], + "angle": 0, + "content": "[62] Huiting Yang, Liangyu Chai, Qiang Wen, Shuang Zhao, Zixun Sun, and Shengfeng He. Discovering interpretable latent space directions of gans beyond binary attributes. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.468, + 0.729 + ], + "angle": 0, + "content": "[63] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.73, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[64] Jiong Zhang, Qi Lei, and Inderjit Dhillon. Stabilizing gradients for deep neural networks via efficient svd parameterization. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.772, + 0.468, + 0.813 + ], + "angle": 0, + "content": "[65] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in gans. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.814, + 0.468, + 0.854 + ], + "angle": 0, + "content": "[66] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in gans. ICML, 2022. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.855, + 0.468, + 0.897 + ], + "angle": 0, + "content": "[67] Xinqi Zhu, Chang Xu, and Dacheng Tao. Learning disentangled representations with latent variation predictability. In ECCV. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7722" + } + ] +] \ No newline at end of file diff --git a/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_origin.pdf b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2ae35300883175cc341cc9fe73414354207e25a5 --- /dev/null +++ b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/58ee27b2-47d7-4507-9418-162f31fa3470_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d1f13a4f925d2a76f7d43451e5db5b5e726e09842f9c3f45136a0f5cc00c06e +size 9587689 diff --git a/2023/Householder Projector for Unsupervised Latent Semantics Discovery/full.md b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ee2b121915a1c87221ae1406f9264ec668d90a8 --- /dev/null +++ b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/full.md @@ -0,0 +1,277 @@ +# Householder Projector for Unsupervised Latent Semantics Discovery + +Yue Song $^{1}$ , Jichao Zhang $^{1}$ , Nicu Sebe $^{1}$ , and Wei Wang $^{2}$ + +$^{1}$ Department of Information Engineering and Computer Science, University of Trento, Italy $^{2}$ Beijing Jiaotong University, China yue.song@unitn.it + +# Abstract + +Generative Adversarial Networks (GANs), especially the recent style-based generators (StyleGANs), have versatile semantics in the structured latent space. Latent semantics discovery methods emerge to move around the latent code such that only one factor varies during the traversal. Recently, an unsupervised method proposed a promising direction to directly use the eigenvectors of the projection matrix that maps latent codes to features as the interpretable directions. However, one overlooked fact is that the projection matrix is non-orthogonal and the number of eigenvectors is too large. The non-orthogonality would entangle semantic attributes in the top few eigenvectors, and the large dimensionality might result in meaningless variations among the directions even if the matrix is orthogonal. To avoid these issues, we propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix. The orthogonality guarantees that the eigenvectors correspond to disentangled interpretable semantics, while the low-rank property encourages that each identified direction has meaningful variations. We integrate our projector into pre-trained StyleGAN2/StyleGAN3 and evaluate the models on several benchmarks. Within only $1\%$ of the original training steps for fine-tuning, our projector helps StyleGANs to discover more disentangled and precise semantic attributes without sacrificing image fidelity. Code is publicly available via https://github.com/KingJamesSong/HouseholderGAN. + +# 1. Introduction + +Generative Adversarial Networks (GANs) [15] and the recent style-based generative models (StyleGANs) [27, 28, + +![](images/df0ec7839c5f7f07dcc9cdaef12468d57362fd1e62ab465516f6ea383b487253.jpg) +Figure 1: Motivation of our proposed Householder Projector. Here "Projector" denotes the projection matrix that maps latent codes to features, i.e., the modulation weight of StyleGANs. (Top Left) The singular value imbalance of the non-orthogonal projector would entangle multiple semantics in the top interpretable directions. (Top Right) Due to the large dimensionality of the projector, directly enforcing vanilla orthogonality would spread the data variations among all the eigenvectors, leading to imperceptible and meaningless traversal. (Bottom) Our Householder Projector equips the projection matrix with low-rank orthogonal properties, which simultaneously disentangles semantics into multiple equally-important eigenvectors and guarantees that each direction could correspond to semanticallymeaningful variations. + +[26] in particular have become the leading paradigm of generative modeling in the vision domain. The latent spaces of StyleGANs are known to embed rich and hierarchical semantics [14, 21]; moving the latent code in certain directions could trigger meaningful variations in the output images. Therefore, latent semantics discovery methods + +emerge to identify such interpretable directions that each variation factor is disentangled and the generation process can be precisely controlled [14, 47, 42, 61, 48, 1, 58, 37, 8]. + +Among the recent explorations of unsupervised interpretable semantics discovery methods [56, 48, 57, 41], SeFa [48] pointed out a promising direction to discover semantically meaningful concepts by computing the eigenvector of the projector. Here we refer to the projection matrix that maps latent codes to features as the projector. The key observation is that using the eigenvectors of the projector for latent perturbation would maximize the data variations. Such identified eigenvectors/directions would correspond to meaningful semantic concepts. However, as shown in Fig. 1 top left, this is likely to cause semantics entangled in the top few eigenvectors. This phenomenon stems from the fact that the variation caused by an eigenvector is actually determined by the associated eigenvalue. Due to the imbalanced eigenvalue distribution, the discovered directions are not equally-important, and the top few ones would simultaneously manipulate multiple attributes. This eigenvalue discrepancy can be mitigated by enforcing orthogonality constraint to the projector. Nonetheless, since standard orthogonal matrices have as many equally-important eigenvectors as the dimensionality, there would not be enough semantics to mine in practice when the method scales to large models such as StyleGANs whose projector dimension is 1024 or 512. Consequently, as an accompanying limitation, the data variations would be split among all the eigenvectors and none of them could produce meaningful output changes (see Fig. 1 top right). + +To resolve the above issues, we propose Householder Projector, a low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix that maps latent codes to features. The projector is first decomposed to its SVD form $(\mathbf{USV}^T)$ . Next, the orthogonal singular vectors $\mathbf{U}$ and $\mathbf{V}$ are represented by a series of Householder reflectors, respectively. Thanks to the normalization of Householder reflections, the orthogonality is also preserved during backpropagation. For the singular value $\mathbf{S}$ , we explicitly set it as a low-rank identity matrix (i.e., $\mathbf{S} = \mathrm{diag}(1,\dots ,1,0,\dots ,0)$ ) whose rank defines exactly the number of semantic concepts. As shown in Fig. 1 bottom, the low-rank property guarantees that the identified directions would cause meaningful variations, while the orthogonality encourages that each semantic attribute is disentangled from the others. Moreover, a proper initialization scheme is proposed to leverage the statistics of pre-trained weights, and an acceleration technique is applied to speed up the computation. We also propose a metric dedicated to measuring the smoothness of latent space to interpretable directions based perturbations. Our Householder Projector is integrated into pre-trained StyleGANs [28, 26] at multiple different layers to mine the di + +verse and hierarchical semantics. Since our projector inevitably changes the pre-trained parameters, the modified models incorporated with our projector are fine-tuned for limited steps to maintain the original image fidelity. Both quantitative and qualitative results on several widely used benchmarks [29, 63, 10, 25, 13] show that within marginal fine-tuning steps (1% of the training steps), our Householder Projector improves the latent semantics discovery of StyleGANs to have more precise attribute control while not impairing the quality of generated images. + +The key results and main contributions are as follows: + +- We propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projector of generative models for latent semantics discovery. +- Our Householder Projector can be easily integrated into pre-trained GAN models. With the acceleration technique, it can be efficiently fine-tuned for very limited additional steps, which paves the way for applying latent semantics discovery and orthogonality techniques to large-scale generative models such as StyleGANs. +- Extensive experiments on two popular backbones (StyleGAN2 [28] and StyleGAN3 [26]) and six benchmarks (FFHQ [29], LSUN Church and Cat [63], MetFaces [25], AFHQ [10], and SHHQ [13]) demonstrate that within marginal extra fine-tuning steps (1% of the original training steps), our method could both achieve precise attribute control and preserve the original image quality. + +# 2. Related Work + +Generative Adversarial Networks. In the past few years, GAN-based generative models [15] have achieved remarkable progress in high-fidelity image synthesis [43, 24, 5, 27, 28, 26, 25, 39, 6, 16, 50, 46]. The generation process usually takes the following procedure: a randomly-sampled latent code is fed into the generator through a projection step, and then the generator outputs realistic-looking images. Recently, the style-based generators [27, 28, 26] that gradually absorb layer-wise latent style codes are becoming the de facto GAN backbones. StyleGAN2 [28] improves the original StyleGAN [27] by redesigning generator normalization and training techniques, and StyleGAN3 [26] further explores some equivariance properties. We use the popular StyleGAN2 and StyleGAN3 as our backbones in this paper. Latent Semantics Discovery. Recently, lots of methods explore disentangling the latent space to achieve image editing by moving the latent code in the identified interpretable directions [7, 3, 21, 47, 17, 56, 67, 48, 18, 41, 54, 57, 55, 30, 55, 31, 60, 22, 44, 9, 53, 51]. Supervised methods rely on human annotations (i.e., segmentation masks, attribute categories, 3D priors, and text descriptions) to de + +fine the semantic labels [14, 47, 61, 21, 42, 33, 8, 12, 49, 40, 34, 62, 59, 23]. Here we mainly highlight some relevant unsupervised approaches that are free of annotations. Voynov et al. [56] proposed to jointly learn a set of directions and a classifier such that the interpretable directions can be recognized by the classifier. More recently, Peebles et al. [41] and Wei et al. [57] proposed to add orthogonal Hessian and Jacobian regularization to encourage disentangled representations, respectively. Song et al. [51] proposed to use wave equations to model the spatiotemporal dynamic non-linear latent traversals in generative models. SeFa [48] showed that the eigenvectors of the projector after the latent code could maximize the data variations and proposed to directly use them as the interpretable directions. However, due to the imbalanced eigenvalues, the image attributes would be entangled in the top few eigenvectors. Our proposed Householder Projector solved this issue by parameterizing the projector to a low-rank but orthogonal matrix. Notice that the used orthogonality technique in [56] is different from ours. In [56], the authors use matrix exponential $\exp (\mathbf{A} - \mathbf{A}^T)$ to generate skew orthogonal matrices where the skew-symmetry could limit the representation power. Also, the technique cannot parameterize given matrices and cannot explicitly control the rank. Our Householder representation is more general and flexible, allowing for controllable rank and parameterization of given matrices. Compared with the soft orthogonality regularization used in [41, 57, 18], the hard orthogonality of our method helps the model to learn more uncorrelated attributes within less fine-tuning steps. + +In contrast to global editing approaches discussed above, some methods can perform local image editing in a post hoc way: they first define or learn a segment of regions of interests, and then manipulate the masked intermediate features for local editing [2, 11, 65, 66, 38]. Empowered by the precise control of attributes, our method can also achieve competitive performance in local image editing (see Sec. 4.2). + +# 3. Householder Projector + +This section starts with the preliminary introduction of the closed-form latent semantics discovery. Next, we analyze its inherent limitation on entangled semantics and then illustrate our proposed Householder Projector in detail. + +# 3.1. Preliminary: Closed-form Latent Discovery + +Previous latent semantics discovery approaches [14, 56, 17, 47, 48] consider the GAN manipulation as $\text{edit}(G(\mathbf{z})) = G(\mathbf{z} + \alpha \mathbf{n})$ where $G(\cdot)$ represents the generator, $\mathbf{z} \in \mathbb{R}^d$ denotes the latent code of dimension $d$ , $\mathbf{n} \in \mathbb{R}^d$ is the identified semantically meaningful direction, and $\alpha$ represents the perturbation strength. If one views the GAN as a multi-step projection function, the first projection can + +be expressed as: + +$$ +G _ {1} (\mathbf {z} + \alpha \mathbf {n}) = \mathbf {A} \mathbf {z} + \mathbf {b} + \alpha \mathbf {A} \mathbf {n}, \tag {1} +$$ + +where $\mathbf{A}$ and $\mathbf{b}$ denote the weight and bias of the projection step (e.g., convolution or linear transform), respectively. As can be seen from eq. (1), the resultant manipulation depends on the term $\alpha \mathbf{A}\mathbf{n}$ . Intuitively, an interpretable direction $\mathbf{n}$ should cause large variations of $G_{1}(\mathbf{z} + \alpha \mathbf{n})$ , which is equivalent to maximizing the impact of $\alpha \mathbf{A}\mathbf{n}$ . Motivated by this observation, SeFa [48] proposed to consider the formulation as the following constrained optimization problem: + +$$ +\mathbf {n} ^ {\star} = \arg \max | | \mathbf {A n} | | _ {2} ^ {2} s. t. \mathbf {n} ^ {T} \mathbf {n} = 1, \tag {2} +$$ + +where the constraint $\mathbf{n}^T\mathbf{n} = 1$ is set to satisfy vector orthogonality, and $||\cdot ||$ denotes the $l_{2}$ norm. Introducing a Lagrange multiplier $\lambda$ leads to $2\mathbf{A}^{T}\mathbf{A}\mathbf{n} - 2\lambda \mathbf{n} = 0$ . The closed-form solutions all correspond to the eigenvectors of $\mathbf{A}^T\mathbf{A}$ . This presents a promising approach to identify the semantics by exploiting the projector $\mathbf{A}$ that projects latent codes. However, one fact overlooked by [48] is that the eigenvectors would cause different extents of variations due to the discrepancy of associated eigenvalues. Supposing that $\mathbf{n}$ is an eigenvector of $\mathbf{A}^T\mathbf{A}$ , then we would have $||\mathbf{A}\mathbf{n}||_2^2 = \sigma^2$ where $\sigma$ is the corresponding singular value of $\mathbf{A}$ . For non-orthogonal matrices, the singular values are exponentially decreasing, i.e., $\sigma_1\geq \sigma_2\geq \ldots \geq \sigma_d$ . This would cause most variations captured by the first few interpretable directions. The imbalance is thus likely to make semantic attributes entangled in the top eigenvectors (see Fig. 1 top left). + +# 3.2. Householder Low-rank Orthogonal Projector + +The eigenvalue discrepancy can be eliminated by enforcing strict orthogonality. Orthogonal matrices have the property of identical eigenvalues, which assigns equal importance to each eigenvector. The low-rank constraint could further limit the number of semantics to mine. Therefore, we propose to use Householder representation, a flexible and general framework to parameterize matrices, to endow the projection matrix with low-rank orthogonality. + +Householder Parameterization. Householder representations can parameterize any matrices by using a series of Householder reflectors to represent the orthogonal singular vectors of its Singular Value Decomposition (SVD) form. In the field of deep learning, it has been used to parameterize the transition matrix and to stabilize gradients of recurrent neural networks [36, 64, 35]. The key to the orthogonality representation relies on the following theorem: + +Theorem 1 (Householder representation [20, 32]). Given any square orthogonal matrix $\mathbf{M} \in \mathbb{R}^{m \times m}$ , it can be represented by the product of Householder matrices $\mathbf{M} = \mathbf{H}_1\mathbf{H}_2 \ldots \mathbf{H}_m$ where each Householder matrix is parameterized by a vector as $\mathbf{H}_i = \mathbf{I} - 2\frac{\mathbf{h}_i\mathbf{h}_i^T}{||\mathbf{h}_i||_2^2}$ . + +Let $\mathbf{USV}^T$ denote the SVD of the projector $\mathbf{A}$ where $\mathbf{S}$ + +![](images/0d8e11d6c7fb6a29907ef099e538b9d71d280386a276eb3af226bab384fe5801.jpg) +Figure 2: Gallery of some semantic attributes discovered by our Householder Projector across all used datasets (FFHQ [29] in the top row, LSUN Church [63] and LSUN Cat [63] in the $2_{nd}$ row, SHHQ [13] in the $3_{rd}$ row, and MetFaces [25] and AFHQ [10] in the bottom row). These semantic attributes are sorted from low-level layers (left) to high-level layers (right). + +denotes the diagonal singular value, and $\mathbf{U}$ and $\mathbf{V}$ are left and right orthogonal singular vectors. We use the accumulation of Householder reflectors (i.e., $\prod_{i=0}^{w}\left(\mathbf{I}-2\frac{\mathbf{u}_i\mathbf{u}_i^T}{||\mathbf{u}_i||_2^2}\right)$ and $\prod_{i=0}^{h}\left(\mathbf{I}-2\frac{\mathbf{v}_i\mathbf{v}_i^T}{||\mathbf{v}_i||_2^2}\right)$ where $w,h$ denotes the width and height of $\mathbf{A}$ ) to parameterize $\mathbf{U}$ and $\mathbf{V}$ , respectively. Notice that only $\mathbf{u}_i$ and $\mathbf{v}_i$ are the actual learnable parameters. For $\mathbf{S}$ , we explicitly set it to a diagonal matrix and keep the weight fixed. The projector is thus represented by our Householder parameterizations. + +Low-rank Constraint. To achieve the property of disentangled attributes, one straightforward approach is to parameterize $\mathbf{A}$ as an orthogonal matrix, i.e., to set $\mathbf{S}$ to an identity matrix $\mathbf{I}$ where $\mathbf{I}_{i,j} = 1$ for $i = j$ and $\mathbf{I}_{i,j} = 0$ otherwise. This would lead to equally-important semantic attributes whose number is exactly the projector dimension. However, for large generative models such as StyleGANs, the projector dimension is typically very large (e.g., 512 or 1024). It is not likely to have enough attributes to edit in practice. Setting $\mathbf{S}$ to a full-rank diagonal matrix would spread data variations among all the eigenvectors, resulting in trivial and imperceptible traversal (see Fig. 1 top right). To avoid this issue, we propose to define $\mathbf{S}$ as a low-rank identity matrix: + +$$ +\mathbf {S} = \operatorname {d i a g} (\underbrace {1 , \dots , 1} _ {N}, 0, \dots , 0), \tag {3} +$$ + +where $N$ defines the rank and also the number of semantic attributes to mine. By restricting the rank of the projector, we explicitly limit the number of interpretable directions. As shown in Fig. 1 bottom, this would benefit the latent traversal for more meaningful output variations. + +Orthogonality Preservation. One advantage of our Householder representation is that the orthogonality can be kept during backpropagation. Since we have the vector normalization $\frac{\mathbf{h}_i\mathbf{h}_i^T}{||\mathbf{h}_i||_2^2}$ , the impact of any gradient descent step $\mathbf{h}_i - \eta \nabla \mathbf{h}_i$ on the orthogonality would be cancelled, i.e., the updated vector remains orthogonal after normalization. + +Initialization. When our method is applied to pre-trained models, the well-trained network weights could be leveraged to initialize our Householder Projection. To this end, we propose to use the nearest-orthogonal mapping [52] to project the weight matrix into its orthogonal form that has the nearest distance in the Frobenius norm (i.e., $\min_{\mathbf{R}}||\mathbf{R} - \mathbf{A}||_{\mathrm{F}}$ where $\mathbf{R}^T\mathbf{R} = \mathbf{I}$ ). The closed-form solution is given by $\mathbf{U}\mathbf{V}^T$ where $\mathbf{USV}^T$ is the SVD of the original weight matrix $\mathbf{A}$ . Next, we decompose $\mathbf{U}$ and $\mathbf{V}$ into their Householder reflectors and use them to initialize our projector. Such an initialization scheme leverages the statistics of the original weight matrix, which might give our projector a good starting point and improve the performance (see the ablation study of Sec. D.3 in the supplementary). + +Acceleration. The accumulated product of elementary + +![](images/2d6b66fa10c03b7a4919c43e42619846c3a5b6da75936eff8b883a4ee6d10a60.jpg) +Figure 3: Illustration on how our Householder Projector represents the modulation weight $\mathbf{A}$ of StyleGANs. Here "Demod", "EMA", and "FN" denote Demodulation, Exponential Moving Average, and Filtered Non-linearities, respectively. The projector is parameterized by its SVD form where $\mathbf{U}$ and $\mathbf{V}$ are represented by the accumulation of Householder reflectors, and $\mathbf{S}$ is set to a low-rank identity matrix. Our projector is applied at multiple different layers of StyleGANs to explore the diverse and hierarchical semantics. The actual learnable parameters are $\mathbf{u}_i$ and $\mathbf{v}_i$ . + +![](images/a2b6616c60b43c442c09eb5107a69155c47113f61e028f220c90e233381ddf32.jpg) +Figure 4: Comparison with ReSeFa [66]. The blue lines indicate the specific regions changed by our method, and the red box indicates the region of interest that is needed as input to ReSeFa [66]. + +Householder matrices can be accelerated via the theorem: + +Theorem 2 (Compact WY representation [4]). For any accumulation of $m$ Householder matrices $\mathbf{H}_1 \ldots \mathbf{H}_m$ , there exists $\mathbf{W}, \mathbf{Y} \in \mathbb{R}^{d \times m}$ such that $\mathbf{I} - 2\mathbf{W}\mathbf{Y}^T = \mathbf{H}_1 \ldots \mathbf{H}_m$ where computing $\mathbf{W}$ and $\mathbf{Y}$ takes $O(dm^2)$ time and $m$ sequential Householder multiplications. + +This theorem indicates the possibility to repeatedly split the accumulation $\mathbf{H}_1\mathbf{H}_2\dots \mathbf{H}_m$ into multiple subsequences until irreducible. Then these sub-sequences can be computed in parallel and gradually merged. As revealed in the ablation of Sec. D.3 in the supplementary, this technique could fully exploit the parallel computational power of GPUs and greatly speed up the aggregation, particularly in our case where the projector dimension is large. + +Implementation in StyleGANs. Fig. 3 depicts how our Householder Projector modifies the original StyleGAN architectures. The projector used in the weight modulation module is represented by our proposed projector. The weight matrix is thus endowed with low-rank orthogonal properties. We insert the proposed projector at every layer of StyleGAN2 and every four layers of StyleGAN3. Since StyleGAN3 has 15 intermediate layers, the adjacent layers have very similar and even repeated semantics. Therefore, + +
IdentityPoseAgeGenderGlassesSmile
Identity0.650.240.030.040.010.03
Pose0.110.570.040.040.110.13
Age0.020.050.670.030.190.03
Gender0.030.320.020.520.100.00
Glasses0.010.080.000.020.880.01
Smile0.020.010.010.000.010.95
+ +(a) Our method + +
IdentityPoseAgeGenderGlassesSmile
Identity0.560.060.050.020.050.26
Pose0.440.480.050.010.010.00
Age0.430.270.220.050.030.01
Gender0.330.180.040.400.040.02
Glasses0.270.140.040.100.230.22
Smile0.200.080.090.000.030.61
+ +(b) SeFa + +
IdentityPoseAgeGenderGlassesSmile
Identity0.510.270.040.050.030.10
Pose0.390.350.050.020.070.11
Age0.260.140.320.100.050.12
Gender0.200.060.340.290.040.07
Glasses0.180.060.120.070.420.14
Smile0.130.050.070.020.030.70
+ +(c) OrJaR + +
IdentityPoseAgeGenderGlassesSmile
Identity0.540.250.040.060.020.08
Pose0.420.300.080.040.050.12
Age0.210.150.280.110.080.17
Gender0.270.070.230.250.040.13
Glasses0.150.120.160.040.370.16
Smile0.110.090.120.010.010.66
+ +(d) HP + +Table 1: The $l_{1}$ normalized attribute correlations based on $2K$ same samples generated by GAN inversion. + +![](images/b9a110296133d286bd9fb2ac780c9061af04bfedba34a618129827073ae09371.jpg) +Pose Unambiguity (1st Layer 8th Direction) +Figure 5: Interpretable directions identified by our method are semantically consistent among different samples. + +![](images/5ba7e1065ca5a3b630e7008a875742abbc5973ca09da1ea3d7a8e84935d8b07b.jpg) +Figure 6: Qualitative comparisons of the same samples. + +we choose to integrate our projector every few layers to obtain interpretable directions of different semantics. + +# 4. Experiments + +In this section, we first introduce the experimental setup, followed by the quantitative and qualitative evaluation. We defer the full ablation studies to Sec. D of supplementary. + +# 4.1. Setup + +Models. We evaluate our Householder Projector on StyleGAN2 [28] and StyleGAN3 [26], i.e., the challenging state-of-the-art GAN backbones in the field of computer vision. + +Datasets and Baselines. For StyleGAN2, we conduct experiments on FFHQ [29], LSUN Church [63], and LSUN + +Cat [63]. The experiments of StyleGAN3 are performed on SHHQ [13], MetFaces [25] and AFHQv2 [10]. We mainly compare our method with representative unsupervised latent semantics discovery approaches, including SeFa [48], Orthogonal Jacobian Regularization (OrJaR) [57], and Hessian Penalty (HP) [41]. SeFa [48] can be directly applied to pre-trained models, while OrJaR and HP need extra finetuning or training from scratch due to the regularization. + +Metrics. We conduct quantitative evaluation using (1) Fréchet Inception Distance (FID) [19], (2) Perceptual Path Length (PPL) [27], (3) Perceptual Interpretable Path Length (PIPL), and (4) Face Attribute Correlation. FID aims to measure the image quality and diversity by computing the distance between the real and fake distributions, and PPL is designed to assess the perceptual smoothness of the latent space where the smoothness can reflect the disentanglement ability. Our proposed PIPL is a natural modification of PPL by adapting the latent manipulation from random interpolation-based perturbations to vector-based perturbations using interpretable directions. Compared with PPL, our PIPL can better measure the latent space smoothness when the latent code is perturbed along with specific vectors, which suits more vector-based semantic discovery methods like SeFa [48] and ours. Furthermore, for StyleGAN2 trained on FFHQ, we rely on pre-trained face attribute estimators to compute the correlation coefficient between the traversal steps and the face attributes. Besides the above four metrics, we also assess the disentanglement through visual observation. We defer the details of used datasets and metrics to Sec. C of the supplementary. + +Implementation Details. We adopt the widely used Pytorch implementation of StyleGAN $^{1}$ and convert the official TensorFlow pre-trained models into PyTorch for FFHQ [29], LSUN Church [63], and LSUN Cat [63]. For StyleGAN3, we use the official code and pre-trained models of AFHQv2 [10] and MetFaces [25]. As for SHHQ [13], we also use the official pre-trained model $^{2}$ . Following the original optimization strategy, we finetune all the parameters of the pre-trained generator and discriminator within $1\%$ of the total training steps (kimgs for StyleGAN3). For instance, the fine-tuning process takes $5K$ steps for StyleGAN2 with FFHQ and 250 kimgs for StyleGAN3 with AFHQ. The fine-tuning time is thus very limited due to the small number + +![](images/5c2452bad3806cbe8ec8196ecc12255fc2c002b8091e389610dfbff5b04f25f0.jpg) +Figure 7: Exemplary qualitative comparison of two different semantics on FFHQ [29] with StyleGAN2 [28]. Our Householder Projector can precisely control the image attributes without changing the face identity. The direction index denotes the index of eigenvectors. + +of training steps. To give concrete examples, fine-tuning models on FFHQ and AFHQ takes 1.5 and 2.5 hours, respectively. The rank $N$ of S is set to 10 for all experiments based on our empirical observation of the number of semantics of StyleGANs. Our editing strength is set the same as SeFa. We use 4 RTX A6000 GPUs for the training. For the comparison fairness, the baseline methods OrJaR [57] and HP [41] are also fine-tuned with the same steps. + +# 4.2. Qualitative Evaluation and Discussion + +Diverse and Precise Attributes. Fig. 2 exhibits some semantic attributes discovered by our Householder Projector on all the datasets. Our method mines a diverse set of disentangled semantics, enabling a wide range of attribute manipulation. Take as an example the first row of attributes discovered on FFHQ [29]. The left columns present diverse high-level semantic concepts such as "Pose" and "Shape", while the right columns show low-level attributes such as + +
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
FFHQ [29]1024 × 1024SeFa [48]4.481579.760.227
OrJaR [57]4.51987.800.204
HP [41]4.66993.170.207
Ours4.40966.230.141
LSUN Church [63]256 × 256SeFa [48]4.61530.680.069
OrJaR [57]3.77474.770.065
HP [41]3.78486.930.058
Ours3.72457.520.030
LSUN Cat [63]256 × 256SeFa [48]8.37722.240.141
OrJaR [57]8.39562.980.134
HP [41]8.31573.710.136
Ours8.46526.260.057
+ +Table 2: Evaluation results on StyleGAN2. + +"Color" and "Lighting". This hierarchy also meets the same trend of original StyleGANs. The manipulation of the diverse and highly-disentangled semantics would give users more precise control on the image generation process. + +Semantic Unambiguity. Importantly, our interpretable di + +![](images/e216ea49735de13264ad94fcd7daf3a3caa73e953b78fc5adc39b3698042a1fa.jpg) +Figure 8: Exemplary visual comparison of three different semantics on SHHQ [13] with StyleGAN3 [26]. Our method is able to mine more disentangled interpretable directions and have more precise control on the attributes. The direction index denotes the index of eigenvectors. + +![](images/523d20aae279e8338ad53c7e694d8e8c631b504ab4bad992f05868296779643b.jpg) + +![](images/28d1586cd879d45ff75e34f3f35bb0b675c89b902698d6f280a13f8457caba7d.jpg) + +rections are unambiguous to different samples. As shown in Fig. 5, the image variations manipulated by our discovered directions all correspond to the same semantic attribute, i.e., the head pose. The other non-target attributes are untouched, such as the background and face identity. + +Comparison against Other Methods. Fig. 7 and Fig. 8 compare the latent traversal of some directions against other baselines on FFHQ [29] and SHHQ [13], respectively. All the methods can discover similar attributes in the same layer, such as the head length in Fig. 7 left. However, the baselines suffer from entangled semantics and the other attributes vary during the traversal, such as hairstyle for SeFA [48] and OrJaR [57], and expression for HP [41]. By contrast, our method is able to discover more precise semantics and preserve other non-target attributes. + +Comparison of Same Samples. To better compare the qualitative disentanglement performance, we further use a GAN inversion technique (PTI [45]) to create nearly the same images from FFHQ for different methods. Fig. 6 displays some qualitative comparisons. For the same samples, our method still has more precise attribute control. + +Local Editing Applications. With precise attribute control, + +our method is even able to edit local regions by simply performing latent traversal. Fig. 4 displays such a use case. Our method achieves very competitive performance against local editing methods such as the recent ReSeFa [66]. In addition, our method is free of extra bounding box as input. + +Please refer to Sec. E of the supplementary for more visualizations about comparisons on other datasets, and semantic diversity/unambiguity/hierarchy. + +# 4.3. Quantitative Evaluation + +StyleGAN2 Results. Table 2 presents the quantitative evaluation results on FFHQ [29], LSUN Church [63] and LSUN Cat [63] datasets. Our proposed Householder Projector outperforms other baselines in terms of both PPL and PIPL. This demonstrates that our method has a smoother and more structured latent space, which corresponds to more disentangled representations. In particular, our approach surpasses SeFa [48] by a large margin, which indicates the benefit of enforcing low-rank orthogonality to the projection matrix. Compared with the soft orthogonality regularization used in OrJaR [57] and HP [41], the hard orthogonality of our projector also has an advantage in latent smoothness + +due to the strict orthogonality preservation and the additional low-rank constraint. Moreover, our FID score is also very competitive, implying that our method could simultaneously improve the disentanglement performance while keeping the quality of generated images unharmed. For the attribute correlation, we also use GAN inversion to create a dataset of $2K$ identical images for each method. Table 1 compares the correlation results on FFHQ. Our method outperforms other unsupervised baselines and preserves the attribute well in particular. + +
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
MetFaces [25]1024 × 1024SeFa [48]15.335626.312.991
OrJaR [57]17.555754.443.700
HP [41]17.325323.273.465
Ours16.894192.520.099
AFHQv2 [10]512 × 512SeFa [48]4.402193.740.470
OrJaR [57]5.452103.470.440
HP [41]5.332198.460.463
Ours4.982052.380.070
SHHQ [13]512 × 256SeFa [48]2.541621.070.370
OrJaR [57]4.781614.560.245
HP [41]5.381648.740.216
Ours4.171549.010.119
+ +Table 3: Evaluation results on StyleGAN3. + +StyleGAN3 Results. Table 3 compares the performance of our method against other baselines on MetFaces [25], AFHQv2 [10], and SHHQ [13] datasets. The results are very coherent with those on StyleGAN2: our Householder Projector improves the latent space smoothness without harming the image fidelity. Our method as well as the two baselines that involve fine-tuning have slightly worse FID than the original StyleGAN3. This might stem from the fact that due to the limited computational resources, our used batch size (16) is actually smaller than the original setting of StyleGAN3 (32). As revealed in the ablation study of Sec. D.2 of the supplementary, the hyper-parameter batch size has a substantial impact on FID. We expect that increasing the batch size would further boost the image quality of our method and lead to a more competitive FID score. + +# 5. Conclusion + +This paper proposes a general and flexible low-rank orthogonal matrix representation coined as Householder Projector for unsupervised latent semantics discovery of generative models. The proposed method endows the projection matrix with low-rank orthogonality. This superiority helps pre-trained GANs to achieve precise and diverse semantics control within limited fine-tuning steps. Extensive experiments of StyleGANs on various benchmarks demonstrate that our method could simultaneously identify the disentangled attributes while maintaining image fidelity. + +# Acknowledgement + +This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9). + +# References + +[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM TOG, 2021. 2 +[2] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 2020. 3 +[3] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B Tenenbaum, William T Freeman, and Antonio Torralba. Gan dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2 +[4] Christian Bischof and Charles Van Loan. The wy representation for products of householder matrices. SIAM Journal on Scientific and Statistical Computing, 8(1):s2-s13, 1987. 5 +[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2 +[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 2 +[7] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. NeurIPS, 2016. 2 +[8]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. ECCV, 2022. 2, 3 +[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. *ICLR*, 2022. 2 +[10] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 2, 4, 6, 9 +[11] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In CVPR, 2020. 3 +[12] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In CVPR, 2020. 3 +[13] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. ECCV, 2022. 2, 4, 6, 8, 9 +[14] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In ICCV, 2019. 1, 2, 3 + +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 1, 2 +[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. *ICLR*, 2022. 2 +[17] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. NeurIPS, 2020. 2, 3 +[18] Zhenliang He, Meina Kan, and Shiguang Shan. Eigengan: Layer-wise eigen-learning for gans. In ICCV, 2021. 2, 3 +[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6 +[20] Alston S Householder. Unitary triangularization of a nonsymmetric matrix. Journal of the ACM (JACM), 5(4):339-342, 1958. 3 +[21] Ali Jahanian, Lucy Chai, and Phillip Isola. On the"steerability" of generative adversarial networks. In ICLR, 2020. 1, 2, 3 +[22] Adarsh Kappiyath, Silpa Vadakkeeveetil Sreelatha, and S Sumitra. Self-supervised enhancement of latent discovery in gans. In AAAI, 2022. 2 +[23] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Singh, and R Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In ECCV, 2022. 3 +[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. *ICLR*, 2018. 2 +[25] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2, 4, 6, 9 +[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1, 2, 6, 8 +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6 +[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 1, 2, 6, 7 +[29] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In CVPR, 2014. 2, 4, 6, 7, 8 +[30] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2 +[31] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In CVPR, 2021. 2 +[32] Richard B Lehoucq. The computation of elementary unitary matrices. ACM Transactions on Mathematical Software (TOMS), 22(4):393-400, 1996. 3 + +[33] Minjun Li, Yanghua Jin, and Huachun Zhu. Surrogate gradient field for latent space manipulation. In CVPR, 2021. 3 +[34] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. NeurIPS, 2021. 3 +[35] Alexander Mathiasen, Frederik Hvilshøj, Jakob Rødsgaard Jørgensen, Anshul Nasery, and Davide Mottin. What if neural networks had svds? NeurIPS, 2020. 3 +[36] Zakaria Mhammedi, Andrew Hellicar, Ashfaqur Rahman, and James Bailey. Efficient orthogonal parametrisation of recurrent neural networks using householder reflections. In ICML, 2017. 3 +[37] James Oldfield, Markos Georgopoulos, Yannis Panagakis, Mihalis A Nicolaou, and Ioannis Patras. Tensor component analysis for interpreting the latent space of gans. BMVC, 2021. 2 +[38] James Oldfield, Christos Tzelepis, Yannis Panagakis, Michaelis A Nicolaou, and Ioannis Patras. Panda: Unsupervised learning of parts and appearances in the feature maps of gans. *ICLR*, 2023. 3 +[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2 +[40] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3 +[41] William Peebles, John Peebles, Jun-Yan Zhu, Alexei Efros, and Antonio Torralba. The hessian penalty: A weak prior for unsupervised disentanglement. In ECCV. Springer, 2020. 2, 3, 6, 7, 8, 9 +[42] Antoine Plumerault, Hervé Le Borgne, and Céline Hudelot. Controlling generative models with continuous factors of variations. *ICLR*, 2020. 2, 3 +[43] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. *ICLR*, 2016. 2 +[44] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Learning disentangled representation by exploiting pretrained generative models: A contrastive learning view. In ICLR, 2022. 2 +[45] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM TOG, 2022. 8 +[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 2 +[47] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2, 3 +[48] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2, 3, 6, 7, 8, 9 +[49] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 3 + +[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2 +[51] Yue Song, Andy Keller, Nicu Sebe, and Max Welling. Latent traversals in generative models as potential flows. In ICML. PMLR, 2023. 2, 3 +[52] Yue Song, Nicu Sebe, and Wei Wang. Improving covariance conditioning of the svd meta-layer by orthogonality. In ECCV, 2022. 5 +[53] Yue Song, Nicu Sebe, and Wei Wang. Orthogonal svd covariance conditioning and latent disentanglement. IEEE TPAMI, 2022. 2 +[54] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. "Gan" steerability" without optimization. *ICLR*, 2021. 2 +[55] Christos Tzelepis, Georgios Tzimiropoulos, and Ioannis Pañas. Warpedganspace: Finding non-linear rbf paths in gan latent space. In ICCV, 2021. 2 +[56] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, 2020. 2, 3 +[57] Yuxiang Wei, Yupeng Shi, Xiao Liu, Zhilong Ji, Yuan Gao, Zhongqin Wu, and Wangmeng Zuo. Orthogonal jacobian regularization for unsupervised disentanglement in image generation. In ICCV, 2021. 2, 3, 6, 7, 8, 9 +[58] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In CVPR, 2021. 2 +[59] Jianjin Xu and Changxi Zheng. Linear semantics in generative adversarial networks. In CVPR, 2021. 3 +[60] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, 2022. 2 +[61] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 2, 3 +[62] Huiting Yang, Liangyu Chai, Qiang Wen, Shuang Zhao, Zixun Sun, and Shengfeng He. Discovering interpretable latent space directions of gans beyond binary attributes. In CVPR, 2021. 3 +[63] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 2, 4, 6, 7, 8 +[64] Jiong Zhang, Qi Lei, and Inderjit Dhillon. Stabilizing gradients for deep neural networks via efficient svd parameterization. In ICML, 2018. 3 +[65] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in gans. NeurIPS, 2021. 3 +[66] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in gans. ICML, 2022. 3, 5, 8 +[67] Xinqi Zhu, Chang Xu, and Dacheng Tao. Learning disentangled representations with latent variation predictability. In ECCV. Springer, 2020. 2 \ No newline at end of file diff --git a/2023/Householder Projector for Unsupervised Latent Semantics Discovery/images.zip b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..413ebcfc1b1adffc4ee7deba5041f90e4a73295f --- /dev/null +++ b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c725852878aa47bc6240ffae85c6d721859416a403d3ed9dcec6f7d688b7ed75 +size 1116870 diff --git a/2023/Householder Projector for Unsupervised Latent Semantics Discovery/layout.json b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ef11f06a8ff5d85a087c483b03fe7a073ea138b2 --- /dev/null +++ b/2023/Householder Projector for Unsupervised Latent Semantics Discovery/layout.json @@ -0,0 +1,8575 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 510, + 121 + ], + "type": "text", + "content": "Householder Projector for Unsupervised Latent Semantics Discovery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "spans": [ + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "text", + "content": "Yue Song" + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "text", + "content": ", Jichao Zhang" + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "text", + "content": ", Nicu Sebe" + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "text", + "content": ", and Wei Wang" + }, + { + "bbox": [ + 162, + 152, + 432, + 168 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 176, + 517, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 176, + 517, + 221 + ], + "spans": [ + { + "bbox": [ + 76, + 176, + 517, + 221 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 176, + 517, + 221 + ], + "type": "text", + "content": "Department of Information Engineering and Computer Science, University of Trento, Italy " + }, + { + "bbox": [ + 76, + 176, + 517, + 221 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 176, + 517, + 221 + ], + "type": "text", + "content": "Beijing Jiaotong University, China yue.song@unitn.it" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 255, + 192, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 266 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 266 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 280, + 290, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 616 + ], + "type": "text", + "content": "Generative Adversarial Networks (GANs), especially the recent style-based generators (StyleGANs), have versatile semantics in the structured latent space. Latent semantics discovery methods emerge to move around the latent code such that only one factor varies during the traversal. Recently, an unsupervised method proposed a promising direction to directly use the eigenvectors of the projection matrix that maps latent codes to features as the interpretable directions. However, one overlooked fact is that the projection matrix is non-orthogonal and the number of eigenvectors is too large. The non-orthogonality would entangle semantic attributes in the top few eigenvectors, and the large dimensionality might result in meaningless variations among the directions even if the matrix is orthogonal. To avoid these issues, we propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix. The orthogonality guarantees that the eigenvectors correspond to disentangled interpretable semantics, while the low-rank property encourages that each identified direction has meaningful variations. We integrate our projector into pre-trained StyleGAN2/StyleGAN3 and evaluate the models on several benchmarks. Within only " + }, + { + "bbox": [ + 46, + 280, + 290, + 616 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 280, + 290, + 616 + ], + "type": "text", + "content": " of the original training steps for fine-tuning, our projector helps StyleGANs to discover more disentangled and precise semantic attributes without sacrificing image fidelity. Code is publicly available via https://github.com/KingJamesSong/HouseholderGAN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 287, + 685 + ], + "type": "text", + "content": "Generative Adversarial Networks (GANs) [15] and the recent style-based generative models (StyleGANs) [27, 28," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 256, + 545, + 425 + ], + "blocks": [ + { + "bbox": [ + 310, + 256, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 310, + 256, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 310, + 256, + 545, + 425 + ], + "type": "image", + "image_path": "df0ec7839c5f7f07dcc9cdaef12468d57362fd1e62ab465516f6ea383b487253.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 616 + ], + "type": "text", + "content": "Figure 1: Motivation of our proposed Householder Projector. Here \"Projector\" denotes the projection matrix that maps latent codes to features, i.e., the modulation weight of StyleGANs. (Top Left) The singular value imbalance of the non-orthogonal projector would entangle multiple semantics in the top interpretable directions. (Top Right) Due to the large dimensionality of the projector, directly enforcing vanilla orthogonality would spread the data variations among all the eigenvectors, leading to imperceptible and meaningless traversal. (Bottom) Our Householder Projector equips the projection matrix with low-rank orthogonal properties, which simultaneously disentangles semantics into multiple equally-important eigenvectors and guarantees that each direction could correspond to semanticallymeaningful variations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "[26] in particular have become the leading paradigm of generative modeling in the vision domain. The latent spaces of StyleGANs are known to embed rich and hierarchical semantics [14, 21]; moving the latent code in certain directions could trigger meaningful variations in the output images. Therefore, latent semantics discovery methods" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 714 + ], + "type": "text", + "content": "The first two authors contribute equally; Wei Wang is the corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7712" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "emerge to identify such interpretable directions that each variation factor is disentangled and the generation process can be precisely controlled [14, 47, 42, 61, 48, 1, 58, 37, 8]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 112, + 289, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 112, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 112, + 289, + 435 + ], + "type": "text", + "content": "Among the recent explorations of unsupervised interpretable semantics discovery methods [56, 48, 57, 41], SeFa [48] pointed out a promising direction to discover semantically meaningful concepts by computing the eigenvector of the projector. Here we refer to the projection matrix that maps latent codes to features as the projector. The key observation is that using the eigenvectors of the projector for latent perturbation would maximize the data variations. Such identified eigenvectors/directions would correspond to meaningful semantic concepts. However, as shown in Fig. 1 top left, this is likely to cause semantics entangled in the top few eigenvectors. This phenomenon stems from the fact that the variation caused by an eigenvector is actually determined by the associated eigenvalue. Due to the imbalanced eigenvalue distribution, the discovered directions are not equally-important, and the top few ones would simultaneously manipulate multiple attributes. This eigenvalue discrepancy can be mitigated by enforcing orthogonality constraint to the projector. Nonetheless, since standard orthogonal matrices have as many equally-important eigenvectors as the dimensionality, there would not be enough semantics to mine in practice when the method scales to large models such as StyleGANs whose projector dimension is 1024 or 512. Consequently, as an accompanying limitation, the data variations would be split among all the eigenvectors and none of them could produce meaningful output changes (see Fig. 1 top right)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "content": "To resolve the above issues, we propose Householder Projector, a low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projection matrix that maps latent codes to features. The projector is first decomposed to its SVD form " + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{USV}^T)" + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "content": ". Next, the orthogonal singular vectors " + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "content": " are represented by a series of Householder reflectors, respectively. Thanks to the normalization of Householder reflections, the orthogonality is also preserved during backpropagation. For the singular value " + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "content": ", we explicitly set it as a low-rank identity matrix (i.e., " + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathrm{diag}(1,\\dots ,1,0,\\dots ,0)" + }, + { + "bbox": [ + 46, + 438, + 289, + 715 + ], + "type": "text", + "content": ") whose rank defines exactly the number of semantic concepts. As shown in Fig. 1 bottom, the low-rank property guarantees that the identified directions would cause meaningful variations, while the orthogonality encourages that each semantic attribute is disentangled from the others. Moreover, a proper initialization scheme is proposed to leverage the statistics of pre-trained weights, and an acceleration technique is applied to speed up the computation. We also propose a metric dedicated to measuring the smoothness of latent space to interpretable directions based perturbations. Our Householder Projector is integrated into pre-trained StyleGANs [28, 26] at multiple different layers to mine the di" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "content": "verse and hierarchical semantics. Since our projector inevitably changes the pre-trained parameters, the modified models incorporated with our projector are fine-tuned for limited steps to maintain the original image fidelity. Both quantitative and qualitative results on several widely used benchmarks [29, 63, 10, 25, 13] show that within marginal fine-tuning steps (1% of the training steps), our Householder Projector improves the latent semantics discovery of StyleGANs to have more precise attribute control while not impairing the quality of generated images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 192, + 537, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 192, + 537, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 192, + 537, + 204 + ], + "type": "text", + "content": "The key results and main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 212, + 545, + 433 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 212, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 545, + 260 + ], + "type": "text", + "content": "- We propose Householder Projector, a flexible and general low-rank orthogonal matrix representation based on Householder transformations, to parameterize the projector of generative models for latent semantics discovery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 268, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 545, + 340 + ], + "type": "text", + "content": "- Our Householder Projector can be easily integrated into pre-trained GAN models. With the acceleration technique, it can be efficiently fine-tuned for very limited additional steps, which paves the way for applying latent semantics discovery and orthogonality techniques to large-scale generative models such as StyleGANs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 347, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 545, + 433 + ], + "type": "text", + "content": "- Extensive experiments on two popular backbones (StyleGAN2 [28] and StyleGAN3 [26]) and six benchmarks (FFHQ [29], LSUN Church and Cat [63], MetFaces [25], AFHQ [10], and SHHQ [13]) demonstrate that within marginal extra fine-tuning steps (1% of the original training steps), our method could both achieve precise attribute control and preserve the original image quality." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 442, + 392, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 392, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 392, + 454 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 462, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 714 + ], + "type": "text", + "content": "Generative Adversarial Networks. In the past few years, GAN-based generative models [15] have achieved remarkable progress in high-fidelity image synthesis [43, 24, 5, 27, 28, 26, 25, 39, 6, 16, 50, 46]. The generation process usually takes the following procedure: a randomly-sampled latent code is fed into the generator through a projection step, and then the generator outputs realistic-looking images. Recently, the style-based generators [27, 28, 26] that gradually absorb layer-wise latent style codes are becoming the de facto GAN backbones. StyleGAN2 [28] improves the original StyleGAN [27] by redesigning generator normalization and training techniques, and StyleGAN3 [26] further explores some equivariance properties. We use the popular StyleGAN2 and StyleGAN3 as our backbones in this paper. Latent Semantics Discovery. Recently, lots of methods explore disentangling the latent space to achieve image editing by moving the latent code in the identified interpretable directions [7, 3, 21, 47, 17, 56, 67, 48, 18, 41, 54, 57, 55, 30, 55, 31, 60, 22, 44, 9, 53, 51]. Supervised methods rely on human annotations (i.e., segmentation masks, attribute categories, 3D priors, and text descriptions) to de" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7713" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "text", + "content": "fine the semantic labels [14, 47, 61, 21, 42, 33, 8, 12, 49, 40, 34, 62, 59, 23]. Here we mainly highlight some relevant unsupervised approaches that are free of annotations. Voynov et al. [56] proposed to jointly learn a set of directions and a classifier such that the interpretable directions can be recognized by the classifier. More recently, Peebles et al. [41] and Wei et al. [57] proposed to add orthogonal Hessian and Jacobian regularization to encourage disentangled representations, respectively. Song et al. [51] proposed to use wave equations to model the spatiotemporal dynamic non-linear latent traversals in generative models. SeFa [48] showed that the eigenvectors of the projector after the latent code could maximize the data variations and proposed to directly use them as the interpretable directions. However, due to the imbalanced eigenvalues, the image attributes would be entangled in the top few eigenvectors. Our proposed Householder Projector solved this issue by parameterizing the projector to a low-rank but orthogonal matrix. Notice that the used orthogonality technique in [56] is different from ours. In [56], the authors use matrix exponential " + }, + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "inline_equation", + "content": "\\exp (\\mathbf{A} - \\mathbf{A}^T)" + }, + { + "bbox": [ + 47, + 72, + 289, + 430 + ], + "type": "text", + "content": " to generate skew orthogonal matrices where the skew-symmetry could limit the representation power. Also, the technique cannot parameterize given matrices and cannot explicitly control the rank. Our Householder representation is more general and flexible, allowing for controllable rank and parameterization of given matrices. Compared with the soft orthogonality regularization used in [41, 57, 18], the hard orthogonality of our method helps the model to learn more uncorrelated attributes within less fine-tuning steps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 432, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 289, + 517 + ], + "type": "text", + "content": "In contrast to global editing approaches discussed above, some methods can perform local image editing in a post hoc way: they first define or learn a segment of regions of interests, and then manipulate the masked intermediate features for local editing [2, 11, 65, 66, 38]. Empowered by the precise control of attributes, our method can also achieve competitive performance in local image editing (see Sec. 4.2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 529, + 180, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 180, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 180, + 544 + ], + "type": "text", + "content": "3. Householder Projector" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 551, + 288, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 288, + 600 + ], + "type": "text", + "content": "This section starts with the preliminary introduction of the closed-form latent semantics discovery. Next, we analyze its inherent limitation on entangled semantics and then illustrate our proposed Householder Projector in detail." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 610, + 275, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 275, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 275, + 624 + ], + "type": "text", + "content": "3.1. Preliminary: Closed-form Latent Discovery" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": "Previous latent semantics discovery approaches [14, 56, 17, 47, 48] consider the GAN manipulation as " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\text{edit}(G(\\mathbf{z})) = G(\\mathbf{z} + \\alpha \\mathbf{n})" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " represents the generator, " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " denotes the latent code of dimension " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " is the identified semantically meaningful direction, and " + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 629, + 289, + 715 + ], + "type": "text", + "content": " represents the perturbation strength. If one views the GAN as a multi-step projection function, the first projection can" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 375, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 375, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 375, + 84 + ], + "type": "text", + "content": "be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 358, + 87, + 545, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 87, + 545, + 99 + ], + "spans": [ + { + "bbox": [ + 358, + 87, + 545, + 99 + ], + "type": "interline_equation", + "content": "G _ {1} (\\mathbf {z} + \\alpha \\mathbf {n}) = \\mathbf {A} \\mathbf {z} + \\mathbf {b} + \\alpha \\mathbf {A} \\mathbf {n}, \\tag {1}", + "image_path": "00cf71c8d041b4cb10b2d21096afc84a1c7d4c5ee19da64123d58a1ee0e68084.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "content": " denote the weight and bias of the projection step (e.g., convolution or linear transform), respectively. As can be seen from eq. (1), the resultant manipulation depends on the term " + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "inline_equation", + "content": "\\alpha \\mathbf{A}\\mathbf{n}" + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "content": ". Intuitively, an interpretable direction " + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "content": " should cause large variations of " + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "inline_equation", + "content": "G_{1}(\\mathbf{z} + \\alpha \\mathbf{n})" + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "content": ", which is equivalent to maximizing the impact of " + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "inline_equation", + "content": "\\alpha \\mathbf{A}\\mathbf{n}" + }, + { + "bbox": [ + 304, + 102, + 547, + 198 + ], + "type": "text", + "content": ". Motivated by this observation, SeFa [48] proposed to consider the formulation as the following constrained optimization problem:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 199, + 545, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 199, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 348, + 199, + 545, + 214 + ], + "type": "interline_equation", + "content": "\\mathbf {n} ^ {\\star} = \\arg \\max | | \\mathbf {A n} | | _ {2} ^ {2} s. t. \\mathbf {n} ^ {T} \\mathbf {n} = 1, \\tag {2}", + "image_path": "686c2526ed64868a0cca0425f765c004a87966bf07b65983c454347dcba53644.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": "where the constraint " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^T\\mathbf{n} = 1" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " is set to satisfy vector orthogonality, and " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "||\\cdot ||" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " norm. Introducing a Lagrange multiplier " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " leads to " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "2\\mathbf{A}^{T}\\mathbf{A}\\mathbf{n} - 2\\lambda \\mathbf{n} = 0" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": ". The closed-form solutions all correspond to the eigenvectors of " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^T\\mathbf{A}" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": ". This presents a promising approach to identify the semantics by exploiting the projector " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " that projects latent codes. However, one fact overlooked by [48] is that the eigenvectors would cause different extents of variations due to the discrepancy of associated eigenvalues. Supposing that " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " is an eigenvector of " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^T\\mathbf{A}" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": ", then we would have " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "||\\mathbf{A}\\mathbf{n}||_2^2 = \\sigma^2" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": " is the corresponding singular value of " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": ". For non-orthogonal matrices, the singular values are exponentially decreasing, i.e., " + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\sigma_1\\geq \\sigma_2\\geq \\ldots \\geq \\sigma_d" + }, + { + "bbox": [ + 304, + 215, + 547, + 408 + ], + "type": "text", + "content": ". This would cause most variations captured by the first few interpretable directions. The imbalance is thus likely to make semantic attributes entangled in the top eigenvectors (see Fig. 1 top left)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 413, + 542, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 542, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 542, + 427 + ], + "type": "text", + "content": "3.2. Householder Low-rank Orthogonal Projector" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 432, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 528 + ], + "type": "text", + "content": "The eigenvalue discrepancy can be eliminated by enforcing strict orthogonality. Orthogonal matrices have the property of identical eigenvalues, which assigns equal importance to each eigenvector. The low-rank constraint could further limit the number of semantics to mine. Therefore, we propose to use Householder representation, a flexible and general framework to parameterize matrices, to endow the projection matrix with low-rank orthogonality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 528, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 547, + 624 + ], + "type": "text", + "content": "Householder Parameterization. Householder representations can parameterize any matrices by using a series of Householder reflectors to represent the orthogonal singular vectors of its Singular Value Decomposition (SVD) form. In the field of deep learning, it has been used to parameterize the transition matrix and to stabilize gradients of recurrent neural networks [36, 64, 35]. The key to the orthogonality representation relies on the following theorem:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "text", + "content": "Theorem 1 (Householder representation [20, 32]). Given any square orthogonal matrix " + }, + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{m \\times m}" + }, + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "text", + "content": ", it can be represented by the product of Householder matrices " + }, + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\mathbf{H}_1\\mathbf{H}_2 \\ldots \\mathbf{H}_m" + }, + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "text", + "content": " where each Householder matrix is parameterized by a vector as " + }, + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_i = \\mathbf{I} - 2\\frac{\\mathbf{h}_i\\mathbf{h}_i^T}{||\\mathbf{h}_i||_2^2}" + }, + { + "bbox": [ + 304, + 629, + 547, + 696 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{USV}^T" + }, + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "text", + "content": " denote the SVD of the projector " + }, + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 317, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7714" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 541, + 654 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 541, + 654 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 541, + 654 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 541, + 654 + ], + "type": "image", + "image_path": "0d8e11d6c7fb6a29907ef099e538b9d71d280386a276eb3af226bab384fe5801.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 662, + 546, + 700 + ], + "lines": [ + { + "bbox": [ + 46, + 662, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 546, + 700 + ], + "type": "text", + "content": "Figure 2: Gallery of some semantic attributes discovered by our Householder Projector across all used datasets (FFHQ [29] in the top row, LSUN Church [63] and LSUN Cat [63] in the " + }, + { + "bbox": [ + 46, + 662, + 546, + 700 + ], + "type": "inline_equation", + "content": "2_{nd}" + }, + { + "bbox": [ + 46, + 662, + 546, + 700 + ], + "type": "text", + "content": " row, SHHQ [13] in the " + }, + { + "bbox": [ + 46, + 662, + 546, + 700 + ], + "type": "inline_equation", + "content": "3_{rd}" + }, + { + "bbox": [ + 46, + 662, + 546, + 700 + ], + "type": "text", + "content": " row, and MetFaces [25] and AFHQ [10] in the bottom row). These semantic attributes are sorted from low-level layers (left) to high-level layers (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7715" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "denotes the diagonal singular value, and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " are left and right orthogonal singular vectors. We use the accumulation of Householder reflectors (i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\prod_{i=0}^{w}\\left(\\mathbf{I}-2\\frac{\\mathbf{u}_i\\mathbf{u}_i^T}{||\\mathbf{u}_i||_2^2}\\right)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\prod_{i=0}^{h}\\left(\\mathbf{I}-2\\frac{\\mathbf{v}_i\\mathbf{v}_i^T}{||\\mathbf{v}_i||_2^2}\\right)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "w,h" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " denotes the width and height of " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ") to parameterize " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", respectively. Notice that only " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " are the actual learnable parameters. For " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", we explicitly set it to a diagonal matrix and keep the weight fixed. The projector is thus represented by our Householder parameterizations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": "Low-rank Constraint. To achieve the property of disentangled attributes, one straightforward approach is to parameterize " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " as an orthogonal matrix, i.e., to set " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " to an identity matrix " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i,j} = 1" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i,j} = 0" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " otherwise. This would lead to equally-important semantic attributes whose number is exactly the projector dimension. However, for large generative models such as StyleGANs, the projector dimension is typically very large (e.g., 512 or 1024). It is not likely to have enough attributes to edit in practice. Setting " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " to a full-rank diagonal matrix would spread data variations among all the eigenvectors, resulting in trivial and imperceptible traversal (see Fig. 1 top right). To avoid this issue, we propose to define " + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 193, + 289, + 350 + ], + "type": "text", + "content": " as a low-rank identity matrix:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 362, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 362, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 287, + 389 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\operatorname {d i a g} (\\underbrace {1 , \\dots , 1} _ {N}, 0, \\dots , 0), \\tag {3}", + "image_path": "711e8562f7897196c8e0a42a2b0ad9d74bc4a54d4e19dfee3194f6e90ceb36ff.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": " defines the rank and also the number of semantic attributes to mine. By restricting the rank of the projector, we explicitly limit the number of interpretable directions. As shown in Fig. 1 bottom, this would benefit the latent traversal for more meaningful output variations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 533 + ], + "type": "text", + "content": "Orthogonality Preservation. One advantage of our Householder representation is that the orthogonality can be kept during backpropagation. Since we have the vector normalization " + }, + { + "bbox": [ + 46, + 456, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\frac{\\mathbf{h}_i\\mathbf{h}_i^T}{||\\mathbf{h}_i||_2^2}" + }, + { + "bbox": [ + 46, + 456, + 287, + 533 + ], + "type": "text", + "content": ", the impact of any gradient descent step " + }, + { + "bbox": [ + 46, + 456, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i - \\eta \\nabla \\mathbf{h}_i" + }, + { + "bbox": [ + 46, + 456, + 287, + 533 + ], + "type": "text", + "content": " on the orthogonality would be cancelled, i.e., the updated vector remains orthogonal after normalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": "Initialization. When our method is applied to pre-trained models, the well-trained network weights could be leveraged to initialize our Householder Projection. To this end, we propose to use the nearest-orthogonal mapping [52] to project the weight matrix into its orthogonal form that has the nearest distance in the Frobenius norm (i.e., " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\min_{\\mathbf{R}}||\\mathbf{R} - \\mathbf{A}||_{\\mathrm{F}}" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^T\\mathbf{R} = \\mathbf{I}" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": "). The closed-form solution is given by " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{U}\\mathbf{V}^T" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{USV}^T" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " is the SVD of the original weight matrix " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": ". Next, we decompose " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " into their Householder reflectors and use them to initialize our projector. Such an initialization scheme leverages the statistics of the original weight matrix, which might give our projector a good starting point and improve the performance (see the ablation study of Sec. D.3 in the supplementary)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Acceleration. The accumulated product of elementary" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 140 + ], + "type": "image", + "image_path": "2d6b66fa10c03b7a4919c43e42619846c3a5b6da75936eff8b883a4ee6d10a60.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "text", + "content": "Figure 3: Illustration on how our Householder Projector represents the modulation weight " + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "text", + "content": " of StyleGANs. Here \"Demod\", \"EMA\", and \"FN\" denote Demodulation, Exponential Moving Average, and Filtered Non-linearities, respectively. The projector is parameterized by its SVD form where " + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "text", + "content": " are represented by the accumulation of Householder reflectors, and " + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "text", + "content": " is set to a low-rank identity matrix. Our projector is applied at multiple different layers of StyleGANs to explore the diverse and hierarchical semantics. The actual learnable parameters are " + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 304, + 149, + 547, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 275, + 545, + 356 + ], + "blocks": [ + { + "bbox": [ + 310, + 275, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 310, + 275, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 310, + 275, + 545, + 356 + ], + "type": "image", + "image_path": "a2b6616c60b43c442c09eb5107a69155c47113f61e028f220c90e233381ddf32.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 362, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 304, + 362, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 546, + 411 + ], + "type": "text", + "content": "Figure 4: Comparison with ReSeFa [66]. The blue lines indicate the specific regions changed by our method, and the red box indicates the region of interest that is needed as input to ReSeFa [66]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 425, + 539, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 539, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 539, + 437 + ], + "type": "text", + "content": "Householder matrices can be accelerated via the theorem:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": "Theorem 2 (Compact WY representation [4]). For any accumulation of " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": " Householder matrices " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_1 \\ldots \\mathbf{H}_m" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": ", there exists " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{W}, \\mathbf{Y} \\in \\mathbb{R}^{d \\times m}" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{I} - 2\\mathbf{W}\\mathbf{Y}^T = \\mathbf{H}_1 \\ldots \\mathbf{H}_m" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": " where computing " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "O(dm^2)" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": " time and " + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 443, + 545, + 504 + ], + "type": "text", + "content": " sequential Householder multiplications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "This theorem indicates the possibility to repeatedly split the accumulation " + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_1\\mathbf{H}_2\\dots \\mathbf{H}_m" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": " into multiple subsequences until irreducible. Then these sub-sequences can be computed in parallel and gradually merged. As revealed in the ablation of Sec. D.3 in the supplementary, this technique could fully exploit the parallel computational power of GPUs and greatly speed up the aggregation, particularly in our case where the projector dimension is large." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Implementation in StyleGANs. Fig. 3 depicts how our Householder Projector modifies the original StyleGAN architectures. The projector used in the weight modulation module is represented by our proposed projector. The weight matrix is thus endowed with low-rank orthogonal properties. We insert the proposed projector at every layer of StyleGAN2 and every four layers of StyleGAN3. Since StyleGAN3 has 15 intermediate layers, the adjacent layers have very similar and even repeated semantics. Therefore," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7716" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 171, + 113 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 171, + 113 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 171, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 171, + 113 + ], + "type": "table", + "html": "
IdentityPoseAgeGenderGlassesSmile
Identity0.650.240.030.040.010.03
Pose0.110.570.040.040.110.13
Age0.020.050.670.030.190.03
Gender0.030.320.020.520.100.00
Glasses0.010.080.000.020.880.01
Smile0.020.010.010.000.010.95
", + "image_path": "6676f9bc3bc7787058978d422ea5c986847fdedcf28593f16111be81721b9aa0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 175, + 70, + 293, + 113 + ], + "blocks": [ + { + "bbox": [ + 82, + 115, + 145, + 125 + ], + "lines": [ + { + "bbox": [ + 82, + 115, + 145, + 125 + ], + "spans": [ + { + "bbox": [ + 82, + 115, + 145, + 125 + ], + "type": "text", + "content": "(a) Our method" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 70, + 293, + 113 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 293, + 113 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 293, + 113 + ], + "type": "table", + "html": "
IdentityPoseAgeGenderGlassesSmile
Identity0.560.060.050.020.050.26
Pose0.440.480.050.010.010.00
Age0.430.270.220.050.030.01
Gender0.330.180.040.400.040.02
Glasses0.270.140.040.100.230.22
Smile0.200.080.090.000.030.61
", + "image_path": "da3ccaac59a2ff50a64a16edfab91ebeeab7ff6ff76a634bf1001f20b433f586.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 296, + 70, + 414, + 113 + ], + "blocks": [ + { + "bbox": [ + 217, + 115, + 252, + 125 + ], + "lines": [ + { + "bbox": [ + 217, + 115, + 252, + 125 + ], + "spans": [ + { + "bbox": [ + 217, + 115, + 252, + 125 + ], + "type": "text", + "content": "(b) SeFa" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 296, + 70, + 414, + 113 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 414, + 113 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 414, + 113 + ], + "type": "table", + "html": "
IdentityPoseAgeGenderGlassesSmile
Identity0.510.270.040.050.030.10
Pose0.390.350.050.020.070.11
Age0.260.140.320.100.050.12
Gender0.200.060.340.290.040.07
Glasses0.180.060.120.070.420.14
Smile0.130.050.070.020.030.70
", + "image_path": "662d05b222066e9ff88c805525a8f048ce5ecfec2550419e3a06f1910b115feb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 417, + 70, + 536, + 113 + ], + "blocks": [ + { + "bbox": [ + 336, + 115, + 376, + 125 + ], + "lines": [ + { + "bbox": [ + 336, + 115, + 376, + 125 + ], + "spans": [ + { + "bbox": [ + 336, + 115, + 376, + 125 + ], + "type": "text", + "content": "(c) OrJaR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 417, + 70, + 536, + 113 + ], + "lines": [ + { + "bbox": [ + 417, + 70, + 536, + 113 + ], + "spans": [ + { + "bbox": [ + 417, + 70, + 536, + 113 + ], + "type": "table", + "html": "
IdentityPoseAgeGenderGlassesSmile
Identity0.540.250.040.060.020.08
Pose0.420.300.080.040.050.12
Age0.210.150.280.110.080.17
Gender0.270.070.230.250.040.13
Glasses0.150.120.160.040.370.16
Smile0.110.090.120.010.010.66
", + "image_path": "007412e07a25b9cdddfeeef02bf8b8c8cee4e027182b4bb0a3c79d261f9e4a69.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 463, + 115, + 492, + 125 + ], + "lines": [ + { + "bbox": [ + 463, + 115, + 492, + 125 + ], + "spans": [ + { + "bbox": [ + 463, + 115, + 492, + 125 + ], + "type": "text", + "content": "(d) HP" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 83, + 128, + 509, + 140 + ], + "lines": [ + { + "bbox": [ + 83, + 128, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 128, + 509, + 140 + ], + "type": "text", + "content": "Table 1: The " + }, + { + "bbox": [ + 83, + 128, + 509, + 140 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 83, + 128, + 509, + 140 + ], + "type": "text", + "content": " normalized attribute correlations based on " + }, + { + "bbox": [ + 83, + 128, + 509, + 140 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 83, + 128, + 509, + 140 + ], + "type": "text", + "content": " same samples generated by GAN inversion." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 169, + 286, + 270 + ], + "blocks": [ + { + "bbox": [ + 88, + 160, + 246, + 169 + ], + "lines": [ + { + "bbox": [ + 88, + 160, + 246, + 169 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 246, + 169 + ], + "type": "text", + "content": "Pose Unambiguity (1st Layer 8th Direction)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 169, + 286, + 270 + ], + "lines": [ + { + "bbox": [ + 50, + 169, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 50, + 169, + 286, + 270 + ], + "type": "image", + "image_path": "b9a110296133d286bd9fb2ac780c9061af04bfedba34a618129827073ae09371.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 278, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 304 + ], + "type": "text", + "content": "Figure 5: Interpretable directions identified by our method are semantically consistent among different samples." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 315, + 286, + 497 + ], + "blocks": [ + { + "bbox": [ + 50, + 315, + 286, + 497 + ], + "lines": [ + { + "bbox": [ + 50, + 315, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 50, + 315, + 286, + 497 + ], + "type": "image", + "image_path": "5ba7e1065ca5a3b630e7008a875742abbc5973ca09da1ea3d7a8e84935d8b07b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 505, + 279, + 517 + ], + "lines": [ + { + "bbox": [ + 54, + 505, + 279, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 279, + 517 + ], + "type": "text", + "content": "Figure 6: Qualitative comparisons of the same samples." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 537, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 287, + 562 + ], + "type": "text", + "content": "we choose to integrate our projector every few layers to obtain interpretable directions of different semantics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 572, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 128, + 586 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "content": "In this section, we first introduce the experimental setup, followed by the quantitative and qualitative evaluation. We defer the full ablation studies to Sec. D of supplementary." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 635, + 97, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 97, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 97, + 647 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "Models. We evaluate our Householder Projector on StyleGAN2 [28] and StyleGAN3 [26], i.e., the challenging state-of-the-art GAN backbones in the field of computer vision." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Datasets and Baselines. For StyleGAN2, we conduct experiments on FFHQ [29], LSUN Church [63], and LSUN" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 160, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 160, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 545, + 256 + ], + "type": "text", + "content": "Cat [63]. The experiments of StyleGAN3 are performed on SHHQ [13], MetFaces [25] and AFHQv2 [10]. We mainly compare our method with representative unsupervised latent semantics discovery approaches, including SeFa [48], Orthogonal Jacobian Regularization (OrJaR) [57], and Hessian Penalty (HP) [41]. SeFa [48] can be directly applied to pre-trained models, while OrJaR and HP need extra finetuning or training from scratch due to the regularization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 258, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 546, + 521 + ], + "type": "text", + "content": "Metrics. We conduct quantitative evaluation using (1) Fréchet Inception Distance (FID) [19], (2) Perceptual Path Length (PPL) [27], (3) Perceptual Interpretable Path Length (PIPL), and (4) Face Attribute Correlation. FID aims to measure the image quality and diversity by computing the distance between the real and fake distributions, and PPL is designed to assess the perceptual smoothness of the latent space where the smoothness can reflect the disentanglement ability. Our proposed PIPL is a natural modification of PPL by adapting the latent manipulation from random interpolation-based perturbations to vector-based perturbations using interpretable directions. Compared with PPL, our PIPL can better measure the latent space smoothness when the latent code is perturbed along with specific vectors, which suits more vector-based semantic discovery methods like SeFa [48] and ours. Furthermore, for StyleGAN2 trained on FFHQ, we rely on pre-trained face attribute estimators to compute the correlation coefficient between the traversal steps and the face attributes. Besides the above four metrics, we also assess the disentanglement through visual observation. We defer the details of used datasets and metrics to Sec. C of the supplementary." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "text", + "content": "Implementation Details. We adopt the widely used Pytorch implementation of StyleGAN" + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "text", + "content": " and convert the official TensorFlow pre-trained models into PyTorch for FFHQ [29], LSUN Church [63], and LSUN Cat [63]. For StyleGAN3, we use the official code and pre-trained models of AFHQv2 [10] and MetFaces [25]. As for SHHQ [13], we also use the official pre-trained model" + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "text", + "content": ". Following the original optimization strategy, we finetune all the parameters of the pre-trained generator and discriminator within " + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "text", + "content": " of the total training steps (kimgs for StyleGAN3). For instance, the fine-tuning process takes " + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "inline_equation", + "content": "5K" + }, + { + "bbox": [ + 304, + 523, + 546, + 679 + ], + "type": "text", + "content": " steps for StyleGAN2 with FFHQ and 250 kimgs for StyleGAN3 with AFHQ. The fine-tuning time is thus very limited due to the small number" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 693, + 474, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 474, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 474, + 703 + ], + "type": "text", + "content": "1https://github.com/rosinality/stylegan2-pytorch" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 703, + 496, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 496, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 496, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 317, + 703, + 496, + 713 + ], + "type": "text", + "content": "https://github.com/stylegan-human/StyleGAN-Human" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7717" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 70, + 512, + 441 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 512, + 441 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 512, + 441 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 512, + 441 + ], + "type": "image", + "image_path": "5c2452bad3806cbe8ec8196ecc12255fc2c002b8091e389610dfbff5b04f25f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 447, + 547, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 447, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 547, + 483 + ], + "type": "text", + "content": "Figure 7: Exemplary qualitative comparison of two different semantics on FFHQ [29] with StyleGAN2 [28]. Our Householder Projector can precisely control the image attributes without changing the face identity. The direction index denotes the index of eigenvectors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 493, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 289, + 590 + ], + "type": "text", + "content": "of training steps. To give concrete examples, fine-tuning models on FFHQ and AFHQ takes 1.5 and 2.5 hours, respectively. The rank " + }, + { + "bbox": [ + 46, + 493, + 289, + 590 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 493, + 289, + 590 + ], + "type": "text", + "content": " of S is set to 10 for all experiments based on our empirical observation of the number of semantics of StyleGANs. Our editing strength is set the same as SeFa. We use 4 RTX A6000 GPUs for the training. For the comparison fairness, the baseline methods OrJaR [57] and HP [41] are also fine-tuned with the same steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 598, + 248, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 248, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 248, + 611 + ], + "type": "text", + "content": "4.2. Qualitative Evaluation and Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "Diverse and Precise Attributes. Fig. 2 exhibits some semantic attributes discovered by our Householder Projector on all the datasets. Our method mines a diverse set of disentangled semantics, enabling a wide range of attribute manipulation. Take as an example the first row of attributes discovered on FFHQ [29]. The left columns present diverse high-level semantic concepts such as \"Pose\" and \"Shape\", while the right columns show low-level attributes such as" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 320, + 485, + 533, + 616 + ], + "blocks": [ + { + "bbox": [ + 320, + 485, + 533, + 616 + ], + "lines": [ + { + "bbox": [ + 320, + 485, + 533, + 616 + ], + "spans": [ + { + "bbox": [ + 320, + 485, + 533, + 616 + ], + "type": "table", + "html": "
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
FFHQ [29]1024 × 1024SeFa [48]4.481579.760.227
OrJaR [57]4.51987.800.204
HP [41]4.66993.170.207
Ours4.40966.230.141
LSUN Church [63]256 × 256SeFa [48]4.61530.680.069
OrJaR [57]3.77474.770.065
HP [41]3.78486.930.058
Ours3.72457.520.030
LSUN Cat [63]256 × 256SeFa [48]8.37722.240.141
OrJaR [57]8.39562.980.134
HP [41]8.31573.710.136
Ours8.46526.260.057
", + "image_path": "0f7aa6923f1a9e414e0b3ca5d99185d402a679afedd7f79f45238563461ed7a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 338, + 624, + 512, + 636 + ], + "lines": [ + { + "bbox": [ + 338, + 624, + 512, + 636 + ], + "spans": [ + { + "bbox": [ + 338, + 624, + 512, + 636 + ], + "type": "text", + "content": "Table 2: Evaluation results on StyleGAN2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 651, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 651, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 545, + 699 + ], + "type": "text", + "content": "\"Color\" and \"Lighting\". This hierarchy also meets the same trend of original StyleGANs. The manipulation of the diverse and highly-disentangled semantics would give users more precise control on the image generation process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Semantic Unambiguity. Importantly, our interpretable di" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7718" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 71, + 235, + 388 + ], + "blocks": [ + { + "bbox": [ + 90, + 71, + 235, + 388 + ], + "lines": [ + { + "bbox": [ + 90, + 71, + 235, + 388 + ], + "spans": [ + { + "bbox": [ + 90, + 71, + 235, + 388 + ], + "type": "image", + "image_path": "e216ea49735de13264ad94fcd7daf3a3caa73e953b78fc5adc39b3698042a1fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 548, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 548, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 548, + 434 + ], + "type": "text", + "content": "Figure 8: Exemplary visual comparison of three different semantics on SHHQ [13] with StyleGAN3 [26]. Our method is able to mine more disentangled interpretable directions and have more precise control on the attributes. The direction index denotes the index of eigenvectors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 70, + 370, + 389 + ], + "blocks": [ + { + "bbox": [ + 238, + 70, + 370, + 389 + ], + "lines": [ + { + "bbox": [ + 238, + 70, + 370, + 389 + ], + "spans": [ + { + "bbox": [ + 238, + 70, + 370, + 389 + ], + "type": "image", + "image_path": "523d20aae279e8338ad53c7e694d8e8c631b504ab4bad992f05868296779643b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 373, + 70, + 506, + 389 + ], + "blocks": [ + { + "bbox": [ + 373, + 70, + 506, + 389 + ], + "lines": [ + { + "bbox": [ + 373, + 70, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 373, + 70, + 506, + 389 + ], + "type": "image", + "image_path": "28d1586cd879d45ff75e34f3f35bb0b675c89b902698d6f280a13f8457caba7d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 443, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 503 + ], + "type": "text", + "content": "rections are unambiguous to different samples. As shown in Fig. 5, the image variations manipulated by our discovered directions all correspond to the same semantic attribute, i.e., the head pose. The other non-target attributes are untouched, such as the background and face identity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 624 + ], + "type": "text", + "content": "Comparison against Other Methods. Fig. 7 and Fig. 8 compare the latent traversal of some directions against other baselines on FFHQ [29] and SHHQ [13], respectively. All the methods can discover similar attributes in the same layer, such as the head length in Fig. 7 left. However, the baselines suffer from entangled semantics and the other attributes vary during the traversal, such as hairstyle for SeFA [48] and OrJaR [57], and expression for HP [41]. By contrast, our method is able to discover more precise semantics and preserve other non-target attributes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 627, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 699 + ], + "type": "text", + "content": "Comparison of Same Samples. To better compare the qualitative disentanglement performance, we further use a GAN inversion technique (PTI [45]) to create nearly the same images from FFHQ for different methods. Fig. 6 displays some qualitative comparisons. For the same samples, our method still has more precise attribute control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Local Editing Applications. With precise attribute control," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 443, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 503 + ], + "type": "text", + "content": "our method is even able to edit local regions by simply performing latent traversal. Fig. 4 displays such a use case. Our method achieves very competitive performance against local editing methods such as the recent ReSeFa [66]. In addition, our method is free of extra bounding box as input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 504, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 540 + ], + "type": "text", + "content": "Please refer to Sec. E of the supplementary for more visualizations about comparisons on other datasets, and semantic diversity/unambiguity/hierarchy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 550, + 441, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 441, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 441, + 563 + ], + "type": "text", + "content": "4.3. Quantitative Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "StyleGAN2 Results. Table 2 presents the quantitative evaluation results on FFHQ [29], LSUN Church [63] and LSUN Cat [63] datasets. Our proposed Householder Projector outperforms other baselines in terms of both PPL and PIPL. This demonstrates that our method has a smoother and more structured latent space, which corresponds to more disentangled representations. In particular, our approach surpasses SeFa [48] by a large margin, which indicates the benefit of enforcing low-rank orthogonality to the projection matrix. Compared with the soft orthogonality regularization used in OrJaR [57] and HP [41], the hard orthogonality of our projector also has an advantage in latent smoothness" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7719" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "due to the strict orthogonality preservation and the additional low-rank constraint. Moreover, our FID score is also very competitive, implying that our method could simultaneously improve the disentanglement performance while keeping the quality of generated images unharmed. For the attribute correlation, we also use GAN inversion to create a dataset of " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " identical images for each method. Table 1 compares the correlation results on FFHQ. Our method outperforms other unsupervised baselines and preserves the attribute well in particular." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 205, + 273, + 346 + ], + "blocks": [ + { + "bbox": [ + 61, + 205, + 273, + 346 + ], + "lines": [ + { + "bbox": [ + 61, + 205, + 273, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 273, + 346 + ], + "type": "table", + "html": "
DatasetsMethodsFID (↓)PPL (↓)PIPL (↓)
MetFaces [25]1024 × 1024SeFa [48]15.335626.312.991
OrJaR [57]17.555754.443.700
HP [41]17.325323.273.465
Ours16.894192.520.099
AFHQv2 [10]512 × 512SeFa [48]4.402193.740.470
OrJaR [57]5.452103.470.440
HP [41]5.332198.460.463
Ours4.982052.380.070
SHHQ [13]512 × 256SeFa [48]2.541621.070.370
OrJaR [57]4.781614.560.245
HP [41]5.381648.740.216
Ours4.171549.010.119
", + "image_path": "4b7d145fd52a86c2f40a5131b9341d61a081f7cc260813a1e1d691d8fa302a56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 354, + 253, + 366 + ], + "lines": [ + { + "bbox": [ + 80, + 354, + 253, + 366 + ], + "spans": [ + { + "bbox": [ + 80, + 354, + 253, + 366 + ], + "type": "text", + "content": "Table 3: Evaluation results on StyleGAN3." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 377, + 289, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 289, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 289, + 557 + ], + "type": "text", + "content": "StyleGAN3 Results. Table 3 compares the performance of our method against other baselines on MetFaces [25], AFHQv2 [10], and SHHQ [13] datasets. The results are very coherent with those on StyleGAN2: our Householder Projector improves the latent space smoothness without harming the image fidelity. Our method as well as the two baselines that involve fine-tuning have slightly worse FID than the original StyleGAN3. This might stem from the fact that due to the limited computational resources, our used batch size (16) is actually smaller than the original setting of StyleGAN3 (32). As revealed in the ablation study of Sec. D.2 of the supplementary, the hyper-parameter batch size has a substantial impact on FID. We expect that increasing the batch size would further boost the image quality of our method and lead to a more competitive FID score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 572, + 119, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 119, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 119, + 585 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "This paper proposes a general and flexible low-rank orthogonal matrix representation coined as Householder Projector for unsupervised latent semantics discovery of generative models. The proposed method endows the projection matrix with low-rank orthogonality. This superiority helps pre-trained GANs to achieve precise and diverse semantics control within limited fine-tuning steps. Extensive experiments of StyleGANs on various benchmarks demonstrate that our method could simultaneously identify the disentangled attributes while maintaining image fidelity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 403, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 403, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 403, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 91, + 546, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 546, + 128 + ], + "type": "text", + "content": "This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 137, + 365, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 365, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 365, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 156, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 311, + 156, + 546, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 156, + 546, + 200 + ], + "spans": [ + { + "bbox": [ + 311, + 156, + 546, + 200 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM TOG, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 201, + 547, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 201, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 311, + 201, + 547, + 245 + ], + "type": "text", + "content": "[2] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 246, + 547, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 246, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 311, + 246, + 547, + 289 + ], + "type": "text", + "content": "[3] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B Tenenbaum, William T Freeman, and Antonio Torralba. Gan dissection: Visualizing and understanding generative adversarial networks. In ICLR, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 290, + 547, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 290, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 311, + 290, + 547, + 323 + ], + "type": "text", + "content": "[4] Christian Bischof and Charles Van Loan. The wy representation for products of householder matrices. SIAM Journal on Scientific and Statistical Computing, 8(1):s2-s13, 1987. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 324, + 546, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 324, + 546, + 356 + ], + "spans": [ + { + "bbox": [ + 311, + 324, + 546, + 356 + ], + "type": "text", + "content": "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 357, + 546, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 357, + 546, + 412 + ], + "spans": [ + { + "bbox": [ + 311, + 357, + 546, + 412 + ], + "type": "text", + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 413, + 546, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 413, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 311, + 413, + 546, + 456 + ], + "type": "text", + "content": "[7] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. NeurIPS, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 457, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 457, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 311, + 457, + 546, + 491 + ], + "type": "text", + "content": "[8]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 491, + 546, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 491, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 311, + 491, + 546, + 534 + ], + "type": "text", + "content": "[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. *ICLR*, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "text", + "content": "[10] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In CVPR, 2020. 2, 4, 6, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[11] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In CVPR, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 635 + ], + "type": "text", + "content": "[12] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In CVPR, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[13] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. ECCV, 2022. 2, 4, 6, 8, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "text", + "content": "[14] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In ICCV, 2019. 1, 2, 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7720" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. *ICLR*, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "text", + "content": "[17] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 207 + ], + "type": "text", + "content": "[18] Zhenliang He, Meina Kan, and Shiguang Shan. Eigengan: Layer-wise eigen-learning for gans. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 209, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 251 + ], + "type": "text", + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 253, + 286, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 286, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 286, + 284 + ], + "type": "text", + "content": "[20] Alston S Householder. Unitary triangularization of a nonsymmetric matrix. Journal of the ACM (JACM), 5(4):339-342, 1958. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 286, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 318 + ], + "type": "text", + "content": "[21] Ali Jahanian, Lucy Chai, and Phillip Isola. On the\"steerability\" of generative adversarial networks. In ICLR, 2020. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 320, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 286, + 353 + ], + "type": "text", + "content": "[22] Adarsh Kappiyath, Silpa Vadakkeeveetil Sreelatha, and S Sumitra. Self-supervised enhancement of latent discovery in gans. In AAAI, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 354, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 398 + ], + "type": "text", + "content": "[23] Tejan Karmali, Rishubh Parihar, Susmit Agrawal, Harsh Rangwani, Varun Jampani, Maneesh Singh, and R Venkatesh Babu. Hierarchical semantic regularization of latent spaces in stylegans. In ECCV, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 286, + 430 + ], + "type": "text", + "content": "[24] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. *ICLR*, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 433, + 286, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 286, + 474 + ], + "type": "text", + "content": "[25] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. NeurIPS, 2020. 2, 4, 6, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 477, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 477, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 477, + 286, + 510 + ], + "type": "text", + "content": "[26] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 512, + 286, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 286, + 543 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 545, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 286, + 578 + ], + "type": "text", + "content": "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 579, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 579, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 49, + 579, + 286, + 611 + ], + "type": "text", + "content": "[29] Vahid Kazemi and Josephine Sullivan. One millisecond face alignment with an ensemble of regression trees. In CVPR, 2014. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 613, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 286, + 645 + ], + "type": "text", + "content": "[30] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "type": "text", + "content": "[31] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "text", + "content": "[32] Richard B Lehoucq. The computation of elementary unitary matrices. ACM Transactions on Mathematical Software (TOMS), 22(4):393-400, 1996. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 711 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[33] Minjun Li, Yanghua Jin, and Huachun Zhu. Surrogate gradient field for latent space manipulation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[34] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[35] Alexander Mathiasen, Frederik Hvilshøj, Jakob Rødsgaard Jørgensen, Anshul Nasery, and Davide Mottin. What if neural networks had svds? NeurIPS, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[36] Zakaria Mhammedi, Andrew Hellicar, Ashfaqur Rahman, and James Bailey. Efficient orthogonal parametrisation of recurrent neural networks using householder reflections. In ICML, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "type": "text", + "content": "[37] James Oldfield, Markos Georgopoulos, Yannis Panagakis, Mihalis A Nicolaou, and Ioannis Patras. Tensor component analysis for interpreting the latent space of gans. BMVC, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "type": "text", + "content": "[38] James Oldfield, Christos Tzelepis, Yannis Panagakis, Michaelis A Nicolaou, and Ioannis Patras. Panda: Unsupervised learning of parts and appearances in the feature maps of gans. *ICLR*, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 352 + ], + "type": "text", + "content": "[39] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "text", + "content": "[40] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[41] William Peebles, John Peebles, Jun-Yan Zhu, Alexei Efros, and Antonio Torralba. The hessian penalty: A weak prior for unsupervised disentanglement. In ECCV. Springer, 2020. 2, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "text", + "content": "[42] Antoine Plumerault, Hervé Le Borgne, and Céline Hudelot. Controlling generative models with continuous factors of variations. *ICLR*, 2020. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 467, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 498 + ], + "type": "text", + "content": "[43] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. *ICLR*, 2016. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 501, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 542 + ], + "type": "text", + "content": "[44] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Learning disentangled representation by exploiting pretrained generative models: A contrastive learning view. In ICLR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 545, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 577 + ], + "type": "text", + "content": "[45] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM TOG, 2022. 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 579, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 611 + ], + "type": "text", + "content": "[46] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[47] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 647, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 667 + ], + "type": "text", + "content": "[48] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 670, + 545, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 711 + ], + "type": "text", + "content": "[49] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "7721" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 710 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 117, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 117, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 117, + 286, + 148 + ], + "type": "text", + "content": "[51] Yue Song, Andy Keller, Nicu Sebe, and Max Welling. Latent traversals in generative models as potential flows. In ICML. PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 150, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 150, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 49, + 150, + 286, + 182 + ], + "type": "text", + "content": "[52] Yue Song, Nicu Sebe, and Wei Wang. Improving covariance conditioning of the svd meta-layer by orthogonality. In ECCV, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 183, + 286, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 183, + 286, + 214 + ], + "spans": [ + { + "bbox": [ + 49, + 183, + 286, + 214 + ], + "type": "text", + "content": "[53] Yue Song, Nicu Sebe, and Wei Wang. Orthogonal svd covariance conditioning and latent disentanglement. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 216, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 216, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 216, + 286, + 237 + ], + "type": "text", + "content": "[54] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. \"Gan\" steerability\" without optimization. *ICLR*, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 238, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 238, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 238, + 286, + 270 + ], + "type": "text", + "content": "[55] Christos Tzelepis, Georgios Tzimiropoulos, and Ioannis Pañas. Warpedganspace: Finding non-linear rbf paths in gan latent space. In ICCV, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 271, + 286, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 271, + 286, + 303 + ], + "spans": [ + { + "bbox": [ + 49, + 271, + 286, + 303 + ], + "type": "text", + "content": "[56] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 304, + 286, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 304, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 286, + 346 + ], + "type": "text", + "content": "[57] Yuxiang Wei, Yupeng Shi, Xiao Liu, Zhilong Ji, Yuan Gao, Zhongqin Wu, and Wangmeng Zuo. Orthogonal jacobian regularization for unsupervised disentanglement in image generation. In ICCV, 2021. 2, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 348, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 348, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 348, + 286, + 380 + ], + "type": "text", + "content": "[58] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 380, + 286, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 380, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 49, + 380, + 286, + 402 + ], + "type": "text", + "content": "[59] Jianjin Xu and Changxi Zheng. Linear semantics in generative adversarial networks. In CVPR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 403, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 403, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 49, + 403, + 286, + 446 + ], + "type": "text", + "content": "[60] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. Transeditor: Transformer-based dual-space gan for highly controllable facial editing. In CVPR, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 447, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 447, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 49, + 447, + 286, + 479 + ], + "type": "text", + "content": "[61] Ceyuan Yang, Yujun Shen, and Bolei Zhou. Semantic hierarchy emerges in deep generative representations for scene synthesis. IJCV, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 479, + 286, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 479, + 286, + 522 + ], + "spans": [ + { + "bbox": [ + 49, + 479, + 286, + 522 + ], + "type": "text", + "content": "[62] Huiting Yang, Liangyu Chai, Qiang Wen, Shuang Zhao, Zixun Sun, and Shengfeng He. Discovering interpretable latent space directions of gans beyond binary attributes. In CVPR, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 523, + 286, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 286, + 577 + ], + "type": "text", + "content": "[63] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 578, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 286, + 610 + ], + "type": "text", + "content": "[64] Jiong Zhang, Qi Lei, and Inderjit Dhillon. Stabilizing gradients for deep neural networks via efficient svd parameterization. In ICML, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 611, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 611, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 49, + 611, + 286, + 643 + ], + "type": "text", + "content": "[65] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces in gans. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 644, + 286, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 644, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 49, + 644, + 286, + 676 + ], + "type": "text", + "content": "[66] Jiapeng Zhu, Yujun Shen, Yinghao Xu, Deli Zhao, and Qifeng Chen. Region-based semantic factorization in gans. ICML, 2022. 3, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 677, + 286, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 677, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 49, + 677, + 286, + 710 + ], + "type": "text", + "content": "[67] Xinqi Zhu, Chang Xu, and Dacheng Tao. Learning disentangled representations with latent variation predictability. In ECCV. Springer, 2020. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7722" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_content_list.json b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f0a68522461c70bf569d16b00e8a79217b61718a --- /dev/null +++ b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_content_list.json @@ -0,0 +1,1705 @@ +[ + { + "type": "text", + "text": "How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability", + "text_level": 1, + "bbox": [ + 96, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijian Wang $^{1}$ Yadan Luo $^{1}$ Liang Zheng $^{2}$ Zi Huang $^{1}$ Mahsa Baktashmotlagh $^{1}$ $^{1}$ The University of Queensland $^{2}$ Australian National University", + "bbox": [ + 158, + 202, + 805, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zijian.wang, y.luo, helen.huang, m.baktashmotlagh}@uq.edu.au, liang.zheng@anu.edu.au", + "bbox": [ + 109, + 241, + 859, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper focuses on model transferability estimation, i.e., assessing the performance of pre-trained models on a downstream task without performing fine-tuning. Motivated by neural collapse (NC) [25] that reveals particular feature geometry at the terminal stage of training, we consider model transferability as how far the target activations obtained by pre-trained models are from their hypothetical state in the terminal phase of the model fine-tuned on the target domain. We propose a metric that measures this proximity based on three phenomena of NC: within-class variability collapse, simplex encoded label interpolation geometry structure is formed, and the nearest center classifier becomes optimal on training data. Through experiments on 11 datasets, we confirm none of the three NC proxies are dispensable, which allows us to obtain very competitive transferability estimation accuracy with approximately $10 \\times$ wall-clock time speed up compared to state-of-the-art approaches.", + "bbox": [ + 75, + 324, + 473, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The past decade has witnessed a surge in the availability of off-the-shelf pre-trained models in public repositories. They are adapted and re-tuned to facilitate new tasks, which has become a standard practice [12]. Nevertheless, this approach raises an important practical question of how to effectively select the best pre-trained model from a large model pool to be carried onto a downstream task.", + "bbox": [ + 75, + 657, + 468, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While precise ranking of these pre-trained models on a new task can be obtained by a greedy approach of fineturning each model and comparing the test accuracy, it is often prohibitive as it requires a vast amount of computing resources. Therefore, an early line of research has been developed referred to as transferability estimation [22, 41, 24, 29], where the aim is to design efficient methods for ranking the performances of pre-trained models on a downstream task without fine-tuning. Existing", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a74933bc6341a105e9721e5c4ab6f3ebdadd1f7f32acdeb840c1f175888a8886.jpg", + "image_caption": [ + "Figure 1: Illustration of the transferability estimation problem. Given a pool of pre-trained models and labeled target data, the objective of transferability estimation is to predict which models can achieve higher performance on the target data after fine-tuning. This task is much more time efficient than fine-tuning the models to obtain ground truth rankings." + ], + "image_footnote": [], + "bbox": [ + 522, + 291, + 870, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "approaches in transferability estimation typically consider the current status of the pre-trained models (w.r.t the target task) and can be generally categorized into two streams of probabilistic-based [22, 36, 41, 1] and feature distribution-based techniques [24, 29]. Approaches from the former category model the expected conditional probability of target labels given the extracted target features [41, 20] or predictions [36, 22]. The latter category models transferability as the class-wise separation of features, in which Pandey et al. [24] employed the Bhattacharyya coefficient to approximate the overlap between class distributions.", + "bbox": [ + 496, + 597, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From a different viewpoint for transferability estimation, this paper measures the difference between the current status and the hypothetical terminal status of pre-trained models after fine-tuning. The latter status refers to neural collapse (NC) [25, 33] on the target domain, which commonly occurs when the training loss approaches zero. Specifically, there are three distinct characteristics of NC: 1) the within-class feature variability drops towards zero; 2) class means gradually converge to a simplex encoded label interpolation", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5549", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(SELI) geometry (def. in Sec.3); and 3) the behavior of the final classifier layer is similar to the nearest centroid classifier. Meanwhile, it is empirically demonstrated in [25] that stopping training at the terminal phase achieves higher test performance compared to the models that stop at an earlier stage (i.e., zero prediction error).", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these interesting findings, we formalize a transferability metric based on the observations of NC discussed above. The metric measures how close the geometry of the target features is to their hypothetical state in the terminal stage of the fine-tuned model. Specifically, we analyze the spectral property of the class-wise feature matrix to understand to what extent the current status is from zero within-class variation. Secondly, we assess the gap between the features extracted by the pre-trained model and SELI geometry with the rank of the feature matrix. Thirdly, we calculate the difference between the hypothetical classifier layer and the nearest neighbor classifier w.r.t. the empirical class-conditional distribution. Collectively, these three different measurements allow us to understand how far the pre-trained models are from NC on the target domain. On a series of standard benchmarks, we demonstrate that the metric exhibits a strong correlation with the transferability of supervised and self-supervised pre-trained models and thus is a very useful transferability estimator. Compared with the state-of-the-art, our method gives stronger correlations while being computationally efficient. We summarize our main contributions below.", + "bbox": [ + 75, + 181, + 468, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the neural collapse transferability index (NCTI) for the transferability estimation. NCTI approximately measures the model w.r.t the three prominent characteristics of NC, which can be efficiently computed.", + "- Experimental results on 11 datasets demonstrate strong correlations between NCTI and the transferability of models to the target dataset, which outperforms existing transferability estimation methods.", + "- We give interesting insights when the NC components are computed on a subset of the pre-training dataset. Our findings suggest a potential for an unsupervised transferability estimation metric, solely based on measuring within-class variability collapse of the source features." + ], + "bbox": [ + 76, + 520, + 467, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 739, + 217, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural collapse. Initially observed empirically by Papyan et. al [25], the phenomenon of neural collapse has garnered significant research attention within a short time frame. Neural collapse shows that at the terminal phase of deep network training (i.e., driving the training loss towards zero), the last layer activations have the following four properties: 1) the variation of activation drop to negligible; 2) the class mean of activation converge to simplex Equiangular Tight Frame (ETF); 3) the weight of linear classifier", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and class means converge to a dual-vector space; and 4) the final linear classifier behaves similar to a nearest class center classifier. Recent follow-up works focus on using the unconstrained feature model with cross-entropy [44, 9] or square loss [43, 35] training to theoretically analyze the NC phenomenon. [33, 8] extend the NC phenomenon from a class-balanced scenario to an imbalanced one. Particularly, Thrangpoulidis et. al [33] demonstrated a more generalized geometry than simplex ETF, named simplex encoded labels interpolation (SELI). SELI shows the invariant property on balanced datasets, as well as imbalanced datasets. The authors also proved that SELI recovers the ETF geometry for the balanced dataset.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transferability estimation. Early works [36, 1, 22] aim to assess the transferability of pre-trained models by calculating the expected conditional distribution of the target label space given the target prediction obtained from the pre-trained classifier. Specifically, [36] leverages negative conditional entropy to evaluate the amount of information shared by pre-training label space and the target task. [22] computes the empirical conditional distribution from the joint distribution of the target task and the pre-training task. Since the early methods heavily rely on the pre-trained predictor to obtain dummy source label distribution of target samples, they cannot directly evaluate the transferability of unsupervised or self-supervised models, where no classifier is available to predict dummy source label distributions.", + "bbox": [ + 496, + 287, + 890, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent transferability estimation methods broaden the applicability by nullifying the dependency on the classifier of the pre-training task. In order to achieve this, NLEEP [20] substitutes the output layer with a Gaussian mixture model. LogME [41] formalizes the transferability estimation as the maximum label marginalized likelihood and adopts a directed graphical model to solve it. Huang et. al [15] proposed to approximate transferability with the mutual information between the pre-trained model extracted features and the corresponding label. On par with the metrics the probabilistic point of view, [29, 24] take the class separability into consideration for the transferability estimation metric. Particularly, Shao et. al [29] incorporates the self-challenge noise augmentation to encourage the classifier to discriminate the hard samples. However, none of the existing works model the transferability as the gap between the current target feature geometry and the hypothetical target feature geometry at the terminal stage of the fine-tuning.", + "bbox": [ + 496, + 500, + 890, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 784, + 633, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transferability estimation is defined as ranking a set of $M$ pre-trained models $\\{\\phi^m (\\cdot)\\}_{m = 1}^M$ given a labeled target dataset $D = \\{\\mathbf{X},\\mathbf{Y}\\} = \\{(x_i,y_i)\\}_{i = 1}^N$ with $N$ being the number of samples. An evaluation metric (accuracy, mAP, etc.) is associated with the dataset, which measures the ground-truth performance $T^{m}$ of the $m$ -th pre", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5550", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained model $\\phi^m (\\cdot)$ after fine-tuning. The feature extracted by the $m$ -th pre-trained model $\\phi^m (\\cdot)$ is denoted as $\\mathbf{H}^m$ where $\\mathbf{H}^m = \\phi^m (\\mathbf{X})\\in \\mathbb{R}^{N\\times d}$ and $d$ is the feature dimension. With the objective of truly reflecting the ground truth model ranking, transferability estimation methods calculate a score $S^m$ for each pre-trained model $\\phi^m (\\cdot)$ . Ideally, the calculated scores $\\{S^m\\}_{m = 1}^M$ are supposed to be highly correlated with the ground truth performance $\\{T^m\\}_{m = 1}^M$ , and able to determine the best pre-trained model to be fine-tuned on the target data.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The central idea of our proposed method is to employ neural collapse phenomena to assess the proximity of the current state of a pre-trained model $\\phi^m (\\cdot)$ to its terminal stage of fine-tuning on the target. Particularly, the proximity is measured from three dimensions, which are: 1) distance of within-class variability of features from zero; 2) simplicity for the pre-trained model to produce simplex encoded labels interpolation (SELI) geometry structure [33]; and 3) the applicability of the nearest centroid classifier. The final metric jointly considers these three terms, with each term contributing equally to the overall transferability score.", + "bbox": [ + 75, + 242, + 472, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The subsequent sections elaborate on each of these three perspectives in detail. Here, we start with two definitions that will be used in the sections: the simplex encoded label (SEL) matrix, and the corresponding simplex encoded label interpolation (SELI) geometry structure.", + "bbox": [ + 76, + 409, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3.1 (Simplex Encoded Label (SEL) Matrix [33])", + "text_level": 1, + "bbox": [ + 76, + 500, + 483, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The SEL matrix $\\hat{\\mathbf{Z}}\\in \\mathbb{R}^{C\\times N}$ is defined as,", + "bbox": [ + 76, + 515, + 352, + 531 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} [ c, i ] = \\left\\{ \\begin{array}{l l} 1 - 1 / C & , c = y _ {i} \\\\ - 1 / C & , c \\neq y _ {i} \\end{array} . \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 542, + 468, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $c \\in [C]$ and $i \\in [N]$ denote the $c$ -th class and $i$ -th observation, respectively. $C$ is the number of classes. Note that $\\hat{\\mathbf{Z}}^{\\top} \\mathbf{1} = 0$ .", + "bbox": [ + 76, + 595, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the definition of SEL matrix $\\hat{\\mathbf{Z}}$ , we can observe that the matrix is almost a full rank matrix ( $rank(\\hat{\\mathbf{Z}}) = \\min(C, N) - 1$ ).", + "bbox": [ + 76, + 656, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3.2 (SEL Interpolation (SELI) Geometry [33])", + "text_level": 1, + "bbox": [ + 76, + 717, + 473, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The embeddings $\\mathbf{H} \\in \\mathbb{R}^{d \\times N}$ and classifier weight matrices $\\mathbf{W} \\in \\mathbb{R}^{d \\times C}$ follow the simplex-encoded-labels interpolation geometry when for some scaling $\\alpha > 0$ :", + "bbox": [ + 76, + 732, + 472, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} ^ {\\top} \\mathbf {W} = \\alpha \\mathbf {V} \\boldsymbol {\\Lambda} \\mathbf {V} ^ {\\top}, \\mathbf {H} ^ {\\top} \\mathbf {H} = \\alpha \\mathbf {U} \\boldsymbol {\\Lambda} \\mathbf {U} ^ {\\top}, a n d \\mathbf {W} ^ {\\top} \\mathbf {H} = \\alpha \\hat {\\mathbf {Z}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 805, + 478, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{Z}} = \\mathbf{V}\\boldsymbol{\\Lambda}\\mathbf{U}^{\\top}$ is the SEL matrix. $\\mathbf{U}$ and $\\mathbf{V}$ denote the left and right singular vector matrix of $\\hat{\\mathbf{Z}}$ . $\\boldsymbol{\\Lambda}$ represents the diagonal singular value matrix.", + "bbox": [ + 76, + 837, + 468, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Within-class Variability Collapse", + "text_level": 1, + "bbox": [ + 500, + 90, + 790, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As described in the neural collapse phenomenon, at the terminal stage of training, the variability of within-class features becomes negligible and the features collapse towards the corresponding class mean. Motivated by this, we develop a score function $S_{vc}$ to track the within-class variability of features, which indicates how close a pretrained model is to the neural collapse state. Specifically, the within-class covariance $\\mathbf{\\Sigma}_c$ of $c$ -th class is denoted as:", + "bbox": [ + 496, + 114, + 890, + 234 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {c} = \\frac {1}{n _ {c}} \\left(\\mathbf {H} _ {\\mathbf {c}} - \\mu_ {c}\\right) ^ {\\top} \\left(\\mathbf {H} _ {\\mathbf {c}} - \\mu_ {c}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 250, + 890, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{\\Sigma}_{c} \\in \\mathbb{R}^{d \\times d}$ , and $\\mathbf{H}_{\\mathbf{c}} \\in \\mathbb{R}^{N_{c} \\times d}$ consists $N_{c}$ features that belong to $c$ -th class. $\\mu_{c}$ denotes the mean of $c$ -th class. By applying singular value decomposition of $(\\mathbf{H}_{\\mathbf{c}} - \\mu_{c}) = \\mathbf{U} \\pmb{\\Lambda} \\mathbf{V}^{T}$ , the covariance matrix can be written as:", + "bbox": [ + 498, + 287, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {c} = \\mathbf {V} \\boldsymbol {\\Lambda} \\mathbf {U} ^ {\\top} \\mathbf {U} \\boldsymbol {\\Lambda} \\mathbf {V} ^ {\\top} = \\mathbf {V} \\boldsymbol {\\Lambda} ^ {2} \\mathbf {V}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 362, + 890, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathbf{U}$ and $\\mathbf{V}$ represent the left and right singular vector matrices. $\\Lambda$ denotes the diagonal singular value matrix.", + "bbox": [ + 496, + 393, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From Eq. 4, we observe that larger singular values mean higher within-class variability. Since SVD is computationally prohibitive on large matrices, it motivates us to use the nuclear norm, which calculates the sum of singular values of a matrix with lower time complexity. Meanwhile, as feature space is usually with high dimensionality, there might exist noises that are easily mitigated by fine-tuning but negatively affects variability calculation. Therefore, we compute the feature variability by substituting $\\mathbf{H}_c^m$ into class-wise logits $\\mathbf{Z}_c^m$ . We define the within-class variability score of $S_{vc}$ using the nuclear norm as follows:", + "bbox": [ + 496, + 425, + 890, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {v c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = - \\sum_ {c = 1} ^ {C} \\left\\| \\mathbf {Z} _ {c} ^ {m} \\right\\| _ {*}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 599, + 890, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{Z}_c^m$ denotes the logits of $c$ -th class extracted by the $m$ -th model. We defer the formulation of $z_{i,c}$ to section 3.3. Since $S_{vc}^{m}$ directly reflects the within-class variability, when a pre-trained model $\\phi^m (\\cdot)$ is close to the terminal stage of the fine-tuning, a larger $S_{vc}^{m}$ should be achieved.", + "bbox": [ + 498, + 652, + 893, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Simplex Encoded Labels Interpolation (SELI)", + "text_level": 1, + "bbox": [ + 500, + 739, + 887, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thrampoulidis et al. [33] proposed SELI as a generalized geometry structure version of the simplex equiangular tight frame (ETF) that is defined in the neural collapse phenomenon. While the simplex ETF geometry observed in neural collapse applies only to balanced datasets, SELI can be established in both balanced and imbalanced datasets at the neural collapse state. As such, we propose a score function $S_{seli}^{m}$ to assess the SELI geometry structure of the target features extracted from the $m$ -th pre-training model. In the", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5551", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41fdd5cc049dd0f628a6eb83df8f3d250b69fb077de4312f2c48f9977bbe1b73.jpg", + "image_caption": [ + "Figure 2: TSNE visualization of feature distributions produced by four pre-trained models with CIFAR-10 as the downstream task. Different colors mean different classes in CIFAR-10. From left to right, target recognition accuracy (or transferability) decreases. At the same time, the proposed three scores, i.e., $S_{vc}$ , $S_{seli}$ , and $S_{ncc}$ also decrease." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 269, + 251 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/202c64155dd48ef7ee69021c461344ad4c3c59d7fba56520ecfa8e7831fa2f38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 88, + 459, + 251 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b8f812051387f72a518d1aacafa7a49ce9e510b1a874894affea420750677e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 88, + 666, + 251 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a15a5fa761e78720e5abf6e43bf17c5f9f796af9306efde1e1a9c7d947c87c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 88, + 859, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rest of this section, we first review the role of SELI geometry in cross-entropy minimization, followed by discussing our SELI-based metric.", + "bbox": [ + 75, + 321, + 468, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under the unconstrained feature (UF) model [44], Thrampoulidis et al. shows that minimizing the cross-entropy loss converges to a KKT point of the SVM classifier, where the global minimizer of this optimization problem is the SEL matrix.", + "bbox": [ + 75, + 367, + 468, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 (Structure of the UF-SVM minimizers) [33] Let $(\\hat{\\mathbf{W}},\\hat{\\mathbf{H}})$ be any solution and $p^*$ be the optimal cost of the following unconstrained feature SVM (UF-SVM) classifier:", + "bbox": [ + 73, + 450, + 470, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(\\hat {\\mathbf {W}}, \\hat {\\mathbf {H}}) \\in \\underset {\\mathbf {W}, \\mathbf {H}} {\\arg \\min} | | \\mathbf {W} | | _ {F} ^ {2} / 2 + | | \\mathbf {H} | | _ {F} ^ {2} / 2,\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 500, + 416, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\left(w _ {y _ {i}} - w _ {c}\\right) ^ {T} h _ {i} \\geq 1, i \\in [ n ], c \\neq y _ {i}.\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 530, + 398, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $\\hat{\\mathbf{Z}} = \\hat{\\mathbf{W}}^{\\top}\\hat{\\mathbf{H}}$ holds for optimal logits, the optimal cost can be obtained as $p^{*} = ||\\hat{\\mathbf{Z}} ||_{*} = ||\\hat{\\mathbf{H}} ||_{F}^{2} = ||\\hat{\\mathbf{W}} ||_{F}^{2}$ .", + "bbox": [ + 75, + 556, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A straightforward way to assess the SELI geometry structure of the target features is to compute the difference between logits $\\mathbf{Z}^m$ extracted from $m$ -th pre-trained model and the optimal logits $\\hat{\\mathbf{Z}}$ . However, since the source and target data might not share similar classes, we cannot directly obtain the logits $\\mathbf{Z}^m$ without the time-consuming fine-tuning process on the target dataset. Therefore, we extract features $\\mathbf{H}^m$ of the model and measure its discrepancy to form the SELI structure.", + "bbox": [ + 75, + 594, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we use the proven fact that, at the terminal stage of the fine-tuning where the model achieves the optimal cost $p^*$ , we have $||\\hat{\\mathbf{Z}}||_* = ||\\hat{\\mathbf{H}}||_F^2$ [33]. Therefore, we approximate the complexity of achieving the optimal logits $\\hat{\\mathbf{Z}}$ via features $\\mathbf{H}^m$ extracted from the pre-trained model $\\phi_m(\\mathbf{X})$ as:", + "bbox": [ + 75, + 729, + 468, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {s e l i} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = \\left\\| \\mathbf {H} ^ {m} \\right\\| _ {*}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 820, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the nuclear norm is the upper bound of $||\\cdot ||_F^2$ and is the convex envelope for rank(), a higher $S_{seli}^{m}$ indicates a higher rank of the feature matrix $\\mathbf{H}_m$ , which makes $\\mathbf{Z}$ closer to a full rank matrix.", + "bbox": [ + 75, + 839, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Simplicity to Nearest Centroid Classifier", + "text_level": 1, + "bbox": [ + 498, + 321, + 846, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the neural collapse phenomenon shown in [25], at the terminal stage of training, the linear classifier behaves similarly to the nearest centroid classifier. Therefore, a score function $S_{ncc}^{m}$ is proposed to measure the applicability of the nearest centroid classifier on the $m$ -th pre-trained model. We start with defining the nearest centroid classifier as [28]:", + "bbox": [ + 496, + 345, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\underset {c \\in [ C ]} {\\arg \\max } \\cos \\left(\\mu_ {c}, x\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 450, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{y}$ is the assigned class label to a sample $x$ , and $\\cos$ indicates the cosine similarity.", + "bbox": [ + 496, + 483, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the neural collapse state, the nearest centroid classifier can achieve the optimal classification performance on the training set. However, in transferability estimation, where there is no actual fine-tuning on the target dataset, directly comparing the performance of applying the nearest centroid classifier to the extracted feature with that of the optimal solution can lead to inaccurate estimation results. To mitigate this issue, we consider two relaxations to the nearest centroid classifier: Instead of using cosine similarity which only involves the mean of the classes, we adopt Mahalanobis distance as the nearest centroid classifier that further involves the correlation between the classes. Moreover, we leverage the expected conditional distribution of labels given input to replace the hard label. Under a mild assumption, the class distribution follows a multivariate Gaussian distribution, thus, we can obtain the following posterior by applying Bayes' Rules:", + "bbox": [ + 496, + 513, + 892, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log P (y = c | h) = \\frac {1}{2} \\left(h _ {i} - \\mu_ {c}\\right) ^ {\\top} \\boldsymbol {\\Sigma} \\left(h _ {j} - \\mu_ {c}\\right) + \\log P (y = c). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 782, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\log P(y = c)$ denotes the prior probability of class $c$ , which can be computed as $N_{c} / N$ . $N_{c}$ represents the number of samples belonging to class $c$ .", + "bbox": [ + 496, + 824, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with the discrete output of argmax function in the nearest centroid classifier formulation, we adopt the", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5552", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1283e8036759425ee21c30d93df920ceda565b6e6236a6ca1c8275a56be52653.jpg", + "image_caption": [ + "Figure 3: Strong correlation between the proposed NCTI score and model recognition performance (\\%) on the target dataset after fine-tuning. Each marker represents a different pre-trained model. The target domains are (from left to right): Cifar100 [17], DTD [6], Food101 [2], Pets [26], SUN397 [39], and VOC2007 [7]. We show Spearman's coefficient $\\rho$ and weighted Kendall's $\\tau$ here. More correlation results are presented in supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 228, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d0a56fa941d4a334ac75e2df4034ba70c796c3b2d61521cacd8d905689f242b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 88, + 362, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a89d69ea03a4aeae14f7043b30c79ac2fb6d522802c069721492d2d041d3518f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 88, + 491, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/93ec1d46fdd5000ce2501591110d3bdd0c40874fdbe47eaa04fd669c0c096e45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 88, + 622, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/097e90b7c951a6ec888a373fa2f678b7fb9378960e4f3ee1881139a69d02a56a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 754, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63065a634a78e719a231439407626b2f3ba1c0993e9502d92bbc7593d2749f9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 88, + 885, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "softmax function as it preserves the relative order information of the prediction. Therefore, even if the corresponding class of a sample is not assigned with the highest probability, the probability value still reflects the distance of the sample to that class. The $c$ -th class normalized posterior probability of $i$ -th sample can be calculated as follows:", + "bbox": [ + 75, + 276, + 468, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i, c} ^ {m} = \\frac {\\exp \\left(\\log P \\left(y = c \\mid h _ {i} ^ {m}\\right)\\right)}{\\sum_ {k = 1} ^ {C} \\exp \\left(\\log P \\left(y = k \\mid h _ {i} ^ {m}\\right)\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 373, + 468, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As such, we have the score $S_{ncc}^{m}$ to describe the discrepancy between the current state of the $m$ -th pre-trained model $\\phi_{m}(X)$ and the optimal nearest centroid classifier as follows:", + "bbox": [ + 75, + 415, + 468, + 473 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n c c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} z _ {i} ^ {m} y _ {i}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 473, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{i}$ denotes the corresponding one-hot ground truth label of $x_{i}$ and $z_{i}^{m}$ is the $C$ dimensional logits of $x_{i}$ , predicted by $m$ -th model. A greater $S_{ncc}^{m}$ indicates a smaller discrepancy towards an optimal nearest centroid classifier, which translates to greater transferability on the target dataset.", + "bbox": [ + 75, + 516, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Score Function", + "text_level": 1, + "bbox": [ + 76, + 599, + 290, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we model the transferability of a pre-trained model as how similar the current state of the network is to its terminal phase of fine-tuning. Specifically, we expect the terminal phase of fine-tuning to hold the so-called three properties of within-class variability, SELI geometry, and the nearest centroid classifier. Correspondingly, we design three score functions, which are $S_{vc}(\\cdot)$ , $S_{seli}(\\cdot)$ , and $S_{ncc}(\\cdot)$ to measure the proximity of the state of the pre-trained network applied on the target data versus its optimal state at the terminal phase. However, since the values of score functions have different scales, summing up the scores may hurt the transferability estimation performance (e.g. $S_{seli}(\\cdot)$ ) can numerically dominate over $S_{ncc}(\\cdot)$ ). Instead of having manually defined hyper-parameters to balance the contribution of each score, we propose to normalize each score into a unit range (0 to 1) as follows:", + "bbox": [ + 75, + 622, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {s e l i} ^ {m} \\leftarrow \\frac {S _ {s e l i} ^ {m} - \\operatorname* {m i n} \\left(S _ {s e l i}\\right)}{\\operatorname* {m a x} \\left(S _ {s e l i}\\right) - \\operatorname* {m i n} \\left(S _ {s e l i}\\right)}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 869, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we can have the normalized score $S_{vc}(H^m)$ and $S_{ncc}(H^m)$ for the $m$ -th pretrained model. By summing the normalized scores, we have the final transferability estimation metric:", + "bbox": [ + 498, + 276, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t o t a l} ^ {m} = S _ {v c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) + S _ {s e l i} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) + S _ {n c c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 345, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $m$ -th pre-trained model with a higher overall score $S_{total}^{m}$ indicates a better transferability in the model pool for the target dataset $D$ .", + "bbox": [ + 498, + 371, + 890, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 430, + 624, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 454, + 785, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We consider a wide range of classification benchmark datasets, including five fine-grained classifications (FGVC Aircraft [21], Standford Cars [16], Food101 [2], Oxford-IIIT Pets [26], Oxford-102 Flowers [23]), four coarse-grained classifications (Caltech101 [18], CIFAR-10 [17], CIFAR-100 [17], VOC2007 [7]), one scene classification (SUN397 [39]), and one texture classification (DTD [6]). In total, we adopt 11 benchmark datasets that are broadly used in transfer learning. A detailed description of each dataset can be found in the supplementary material. All models have undergone pre-training on the ImageNet dataset, and the fine-tuned performance of the model on target datasets is obtained from [29]. Without affecting the fine-tuning performance, we note that only the training and validation splits of the dataset are used for transferability estimation, and we hold the test split out of the estimation process.", + "bbox": [ + 496, + 477, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correlation measurement. Following the previous work [41, 24, 29], we adopt weighted Kendall's $\\tau$ [37] to measure the correlation between the estimated model ranking and the ground-truth performance ranking. Each pairwise comparison between items in two rankings is assigned a weight based on the distance between their ranks, whereas top-ranked items will be assigned with higher weight.", + "bbox": [ + 496, + 734, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. In order to replicate the fine-tuning process, our proposed method aims to find a projection matrix $W$ that maximizes the discriminative power of the extracted features $\\phi^m (X)$ from the $m$ -th pre-trained", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5553", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/29a7a3e72824c653ade8ab69fe2b41460949a20dbd348324ca4493398130b31e.jpg", + "table_caption": [ + "Table 1: Method comparison of their correlation strength with target accuracy for supervised models. Weighted Kendall's $\\tau$ on 11 target test sets and their average are shown. For each column, the best, and second-best results are in bold, and underlined, respectively. Our method achieves the best overall weighted Kendall's $\\tau$ ." + ], + "table_footnote": [], + "table_body": "
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NCE [36]-0.1610.4650.6850.7090.7230.302-0.4820.6270.7720.7600.5710.452
LEEP [22]-0.2770.6050.3670.8240.6770.486-0.2910.4340.3890.6580.4130.390
LogME [41]0.4390.4630.6050.8520.7250.7000.1470.3850.4110.5110.6950.539
NLEEP [20]-0.5310.6140.4890.8250.7310.8200.0540.5290.9550.8480.6990.548
SFDA [29]0.6140.6960.5180.9490.8660.5750.5140.8150.5220.5580.6710.663
NCTI0.4960.6790.6470.8430.8790.7040.5410.7730.9240.7560.7410.726
", + "bbox": [ + 112, + 146, + 859, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fcb74c476c13916e1472bb9e9e934ad04dbf15c25784f3a64d9b4fedc8479830.jpg", + "image_caption": [ + "(a) Supervised models", + "Figure 4: Method comparison w.r.t running time (seconds) vs. average Weighted Kendall's $\\tau$ on 11 datasets on (a) supervised and (b) self-supervised models. Our method obtains the highest ranking correlation within a small time budget." + ], + "image_footnote": [], + "bbox": [ + 81, + 270, + 272, + 411 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d2ef7393490b3ee5cd0506f046ff77b4a3f4a0179bdea8e6c8615c96e7d19c6.jpg", + "image_caption": [ + "(b) Self-supervised models" + ], + "image_footnote": [], + "bbox": [ + 277, + 270, + 467, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model. To achieve this, we employ linear discriminant analysis (LDA) to maximize the separation between the class means of the projected data. However, LDA can be prone to overfitting when applied to high-dimensional feature spaces or when the covariance matrix is singular [40, 42]. To mitigate the overfitting issue on high-dimensional feature space, we follow the previous transferability estimation methods [20, 24] which incorporate principal component analysis (PCA) into the framework to project the original features onto a lower-dimensional subspace. We further address the singularity issue by shrinking the covariance matrix toward a full-rank identity matrix as follows:", + "bbox": [ + 75, + 545, + 472, + 728 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {w} ^ {\\prime} = (1 - \\alpha) \\boldsymbol {\\Sigma} _ {w} + \\alpha \\mathbf {I}, \\text {w h e r e} \\alpha = \\exp \\left(\\sum_ {i} ^ {d ^ {*}} \\lambda_ {i}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 750, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\lambda_{i}$ denotes the $i$ -th largest eigenvalue of the PCA projected feature matrix $PCA(\\phi_m(X))$ . $d^{*}$ indicates the number of eigenvalues to be considered in the calculation. We set the dimension of the PCA projected subspace as 64 and set $d^{*}$ as 32 throughout all the experiments. $\\alpha$ controls the trade-off between the covariance and the identity matrices.", + "bbox": [ + 75, + 811, + 470, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on Supervised Models", + "text_level": 1, + "bbox": [ + 498, + 271, + 794, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview. To initiate our evaluation of transferability estimation metrics for ranking supervised models, we assess the performance of these metrics on 11 widely adopted models, which are ResNet-34 [13], ResNet-50, ResNet-101, ResNet-151, DenseNet-121 [14], DenseNet-169, DenseNet-201, MNet-A1 [32], MobileNet-v2 [27], GoogleNet [30], and Inception-v3 [31].", + "bbox": [ + 496, + 294, + 890, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance comparison. We compare our proposed NCTI with existing transferability estimation metrics, including NCE [36], LEEP [22], LogME [41], NLEEP [20], and SFDA [29]. The experimental results are shown in Tab. 1. From the table, we can observe that empirical conditional probability-based methods (i.e., NCE, LEEP, and LogME) generally achieve lower ranking correlations than our method and SFDA, indicating the importance of modeling the feature space in transferability estimation. NCTI achieves the best or the second best average weighted Kendall's ranking correlation on nine datasets and obtains a $6.9\\%$ averaged performance gain compared to the second-best method SFDA. We argue that while the neural collapse explicitly reflects the maximum class separation, a feature space with a high class separation score can not necessarily leads to the neural collapse phenomenon. Therefore, our method can achieve superior performance than class separation-based methods.", + "bbox": [ + 496, + 400, + 892, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation on Self-supervised Models", + "text_level": 1, + "bbox": [ + 498, + 681, + 823, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview. We further evaluate the effectiveness of NCTI in assessing the transferability of models pre-trained by self-supervised learning (SSL). We construct a pool with 10 SSL pre-trained models, including MoCo-v1 [11], MoCo-v2 [5], PCL-v2 [19], SELA-V2, Deepcluster-v2 [4], BYOL [10], Infomin [34], SWAV [3], and Insdis [38].", + "bbox": [ + 496, + 704, + 890, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance comparison. Since NCE and LEEP require the classifier on the pre-training task, it is not directly applicable to self-supervised model ranking tasks. Therefore, to evaluate the ranking performance of the self-supervised models, we compare our method with NLEEP, LogME, and SFDA in Tab. 2. Similarly, we observe that the average ranking correlation of SFDA and the proposed", + "bbox": [ + 496, + 796, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5554", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/190573cd9d692a6d5917227d1c0b7658085e6a2ed7b8483c216718d08f667d92.jpg", + "table_caption": [ + "Table 2: Method comparison of their correlation strength with target accuracy for self-supervised models. All the settings and notations are the same as Table 1. Our method achieves the best overall weighted Kendall's $\\tau$ ." + ], + "table_footnote": [], + "table_body": "
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NLEEP [20]-0.2860.6620.5950.1080.3740.7790.5980.7160.8640.880-0.0910.473
LogME [41]0.0210.0750.6270.4170.1460.7430.7630.6860.7380.2600.1810.423
SFDA [29]0.1670.6740.6830.8460.7890.8820.8970.8370.5640.8310.6210.708
NCTI0.0360.8110.7960.7580.8110.7960.7620.9450.8050.7740.6060.719
", + "bbox": [ + 112, + 130, + 859, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbb75c0f615981a7c90c3e7589c2fde51296ea98406e2a37924f3bbfd8fe9493.jpg", + "image_caption": [ + "(a) Supervised models", + "Figure 5: Effectiveness of each individual component in NCTI. We use the three terms individually or remove them one at a time from the full system. Results on (a) supervised and (b) self-supervised models are shown." + ], + "image_footnote": [], + "bbox": [ + 83, + 241, + 272, + 387 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49eeae66bb7b0b011d6710737bb95996babcd910b729a5f0fcc30a09356b4e4d.jpg", + "image_caption": [ + "(b) Self-supervised models" + ], + "image_footnote": [], + "bbox": [ + 279, + 241, + 464, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method can achieve superior results than that of the conditional probability-based method. NCTI surpasses the second-best method by $1.5\\%$ on average and scores the highest overall weighted Kendall's $\\tau$ , which verifies the effectiveness of the proposed method.", + "bbox": [ + 75, + 500, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 76, + 588, + 243, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tranferability estimation with fewer target data. We test a practical but challenging transferability estimation scenario, where only limited target data are available for the estimation. We randomly sample 2 to 500 images per class from the full SUN397 and examine how limited data affects the performance of different methods. As we can see from Fig. 6(d), although fewer data generally trigger performance degradation for transferability estimation methods, our method is able to achieve a larger performance gain under the low shot scheme than using the full dataset. Particularly, we highlight that our method can achieve SOTA performance on SUN397 with only 10 samples per class, which indicates a strong potential for practical applications.", + "bbox": [ + 75, + 612, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational efficiency comparison. We compare the time efficiency versus ranking correlation of different transferability estimation metrics in Fig. 4. The $x$ -axis is the total time cost for estimating transferability over 11 datasets, while the $y$ -axis indicates the average ranking correlation on 11 datasets. We adopt a logarithmic scale for", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the $x$ -axis. The figure shows that within a similar time budget, NCTI can achieve significantly higher ranking correlations compared with empirical conditional probability-based methods, such as NCE, LEEP, and LogME. SFDA can achieve a comparable ranking correlation with NCTI on self-supervised model ranking tasks. However, it consumes more than $10 \\times$ more time to estimate the transferability, showing that it is less efficient than our proposed metric. Meanwhile, we can observe from the figure that our method scores the highest ranking correlation to the running time ratio under both settings, emphasizing the effectiveness and efficiency of our method.", + "bbox": [ + 496, + 242, + 890, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study. The NCTI metric proposed in this work is composed of three terms, each term is designed to describe a specific aspect of the gap between the current feature geometry and the geometry of neural collapse. In order to illustrate the impact of each individual term in NCTI, we conduct ablation studies to show how each individual term contributes to the overall metric. The ablation studies are conducted on both supervised and self-supervision scenarios, and the results are shown in Fig. 5. Among these three individual terms of NCTI, although $S_{vc}$ can attain a positive averaged ranking correlation, the contribution towards the full metric is less than the other two terms. $S_{seli}$ delivers the highest average ranking correlation on supervised model estimation, while $S_{ncc}$ attains the best-averaged performance on self-supervised model estimation. We infer this is due to the difference in the supervision signal. As the label information is not available for self-supervised training, the discriminative power of the pre-trained feature is more critical for the downstream task than that in supervised models. We also validate the effectiveness of two-term combinations. As we can see from the figure, the combination between $S_{seli}$ and $S_{vc}$ always achieves higher correlations than its individual component, confirming the terms are complementary. Although NCTI obtains a slightly lower correlation than w/o $S_{ncc}$ on the supervised models, it achieves an absolute $4.2\\%$ higher averaged correlation of both tasks, which verifies the importance of $S_{ncc}$ .", + "bbox": [ + 496, + 428, + 892, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameter analysis. NCTI simplifies the hyperparameter tuning process by assigning equal weights to each term. In order to validate this weight assignment, we investigate the impact of weight changes on the ranking cor", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5555", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80bb27f4f86a341e03df4a7e9768fc1ff224819faaf26ca56de8026ee2ebc9a3.jpg", + "image_caption": [ + "(a) Weight of $S_{vc}$" + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 267, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbe66a0ed849e023f5ca5feb85f9d702c310dd6e9b917c454ffe19ac6d018d34.jpg", + "image_caption": [ + "(b) Weight of $S_{ncc}$" + ], + "image_footnote": [], + "bbox": [ + 303, + 88, + 472, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8d0266025764c280a17f9ad46d81acd3eb7ffcf762637cac72c02878357b8ae.jpg", + "image_caption": [ + "(c) Weight of $S_{seli}$" + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 671, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4356a40cebfc48c525fa089e0e1e614bb3a89dd3f045698fc19ec339f5e77cfb.jpg", + "image_caption": [ + "(d) SUN397" + ], + "image_footnote": [], + "bbox": [ + 707, + 89, + 870, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5873dfbea9ec3ec17a8d902ed440ebb4aa36169161ef8adc2ecc9bde897fcb83.jpg", + "image_caption": [ + "Figure 6: (a)-(c) Hyperparameter analysis of three components in NCTI. We report correlation strength $\\tau$ computed on 11 downstream datasets against weight values. Impact on both supervised and self-supervised pre-trained models are drawn. We study each weight by fixing the rest as 1. In this paper, by default, we set the weight to 1 for all components. (d) Influence of target data size for estimation on the correlation strength. We vary the data size from 2 to 500 samples per class and observe that our method can achieve a larger performance margin compared with existing methods under the low-shot scenario.", + "(a) Supervised models", + "Figure 7: Effectiveness of the three components computed on a subset of the pre-training dataset. Average correlation strength $(\\tau)$ on 11 target datasets reported. We assess (a) supervised models, and (b) self-supervised models." + ], + "image_footnote": [], + "bbox": [ + 81, + 335, + 272, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/544fe4283bc99cea3cff9a2b469723aca4e32906376c9785d9bc4ed6a65b7de1.jpg", + "image_caption": [ + "(b) Self-supervised models" + ], + "image_footnote": [], + "bbox": [ + 279, + 337, + 467, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "relation. Specifically, we vary the weight of the term of interest within the range of $\\{0.1, 0.4, 0.8, 1.2, 1.6, 2\\}$ while keeping the weight of the remaining terms fixed at 1. The experimental results of the hyperparameter analysis are reported in Fig.6 (a)-(c). Our findings indicate that NCTI achieves high averaged correlations ( $\\tau \\geq 0.7$ ) with the most weight configurations. While we observe different monotonicity between the correlations of supervised and self-supervised tasks, the results demonstrate that harmonic points (i.e., high mean and low variance) are consistently achieved when all weights are close to 1, thereby confirming the appropriateness of the equal weight setting.", + "bbox": [ + 75, + 593, + 468, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of NCTI computed using source features. The extensive experiments conducted so far have demonstrated the effectiveness of each individual term in the proposed NCTI metric for transferability estimation. All the previous experiments have been conducted by using the target features extracted from the pre-trained models. It remains important to investigate what the properties of neural collapse on the source (i.e., pre-training) features tell us", + "bbox": [ + 75, + 780, + 467, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "about the transferability of a model to a downstream task. To this end, we measure each individual term in NCTI on a subset of the source dataset and report the ranking correlation of the supervised model estimation task in Fig.7(a) and (b). On average, the transferability of a supervised model to a downstream task is consistently positively related to $-S_{vc}$ and $S_{ncc}$ on the source dataset on both supervised and self-supervised, which scores 0.725 and 0.598 weighted Kendall $\\tau$ , respectively. While the figure shows $S_{seli}$ can achieve a positive correlation on supervised tasks, no significant correlation on $S_{seli}$ is observed in the supervised transfer estimation task. Note that, $-S_{vc}$ achieves the highest ranking correlation on average, which shows that a pre-trained model with higher within-class variability can better transfer to a new task. This indicates that a fully unsupervised metric can be designed for transferability estimation which is solely based on measuring the class-wise compactness of the source features.", + "bbox": [ + 496, + 338, + 893, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 621, + 617, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How to efficiently choose the optimal pre-trained model from a vast collection of models to use in a downstream task? To rank models based on their transferability, we propose a metric named Neural Collapse informed Transferability Index (NCTI). Our method models model transferability as the gap between the current feature geometry and the geometry at the terminal stage (i.e., neural collapse) after hypothetical fine-tuning on the downstream task. Specifically, we model the gap from three perspectives, including the formation of SELI geometry structure, the within-class variability, and the applicability of the nearest center classifier. We show that our method is light to compute and that the ranking of model transferability has a very strong correlation with the ground truth ranking and compares favorably against the state-of-the-art methods.", + "bbox": [ + 496, + 641, + 890, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was partially supported by ARC DP 230101196", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5556", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In 2019 IEEE international conference on image processing (ICIP), pages 2309-2313. IEEE, 2019. 1, 2", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 5", + "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6", + "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6", + "[5] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5", + "[7] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 5", + "[8] Cong Fang, Hangfeng He, Qi Long, and Weijie J Su. Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training. Proceedings of the National Academy of Sciences, 118(43):e2103091118, 2021. 2", + "[9] Florian Graf, Christoph Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised contrastive learning. In International Conference on Machine Learning, pages 3821-3830. PMLR, 2021. 2", + "[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6", + "[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6", + "[12] Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imageNet pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4918-4927, 2019. 1", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6", + "[15] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 2", + "[16] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5", + "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[18] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, Apr 2022. 5", + "[19] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966, 2020. 6", + "[20] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2663-2673, 2021. 1, 2, 6, 7", + "[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[22] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 1, 2, 6", + "[23] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5", + "[24] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9172-9182, 2022. 1, 2, 5, 6", + "[25] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 1, 2, 4", + "[26] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5", + "[27] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6", + "[28] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval, volume 39. Cambridge University Press Cambridge, 2008. 4" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5557", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 286-302. Springer, 2022. 1, 2, 5, 6, 7", + "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6", + "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6", + "[32] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 6", + "[33] Christos Thrampoulidis, Ganesh R Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. arXiv preprint arXiv:2208.05512, 2022. 1, 2, 3, 4", + "[34] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 6", + "[35] Tom Tirer and Joan Bruna. Extended unconstrained features model for exploring deep neural collapse. In International Conference on Machine Learning, pages 21478-21505. PMLR, 2022. 2", + "[36] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1395-1405, 2019. 1, 2, 6", + "[37] Sebastiano Vigna. A weighted correlation index for rankings with ties. In Proceedings of the 24th international conference on World Wide Web, pages 1166-1176, 2015. 5", + "[38] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 6", + "[39] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5", + "[40] Jian Yang and Jing-yu Yang. Why can lda be performed in pca transformed space? Pattern recognition, 36(2):563-566, 2003. 6", + "[41] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Ma" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chine Learning, pages 12133-12143. PMLR, 2021. 1, 2, 5, 6, 7", + "[42] Hua Yu and Jie Yang. A direct lda algorithm for high-dimensional data—with application to face recognition. Pattern recognition, 34(10):2067-2070, 2001. 6", + "[43] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse undermse loss: Global optimality with unconstrained features. In International Conference on Machine Learning, pages 27179-27202. PMLR, 2022. 2", + "[44] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. Advances in Neural Information Processing Systems, 34:29820-29834, 2021. 2, 4" + ], + "bbox": [ + 501, + 92, + 890, + 301 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5558", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_model.json b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b8f04e83da9bafacb6cece205ec043d456d030f8 --- /dev/null +++ b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.203, + 0.806, + 0.24 + ], + "angle": 0, + "content": "Zijian Wang\\(^{1}\\) Yadan Luo\\(^{1}\\) Liang Zheng\\(^{2}\\) Zi Huang\\(^{1}\\) Mahsa Baktashmotlagh\\(^{1}\\) \n\\(^{1}\\)The University of Queensland \\(^{2}\\)Australian National University" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.242, + 0.86, + 0.258 + ], + "angle": 0, + "content": "{zijian.wang, y.luo, helen.huang, m.baktashmotlagh}@uq.edu.au, liang.zheng@anu.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.599 + ], + "angle": 0, + "content": "This paper focuses on model transferability estimation, i.e., assessing the performance of pre-trained models on a downstream task without performing fine-tuning. Motivated by neural collapse (NC) [25] that reveals particular feature geometry at the terminal stage of training, we consider model transferability as how far the target activations obtained by pre-trained models are from their hypothetical state in the terminal phase of the model fine-tuned on the target domain. We propose a metric that measures this proximity based on three phenomena of NC: within-class variability collapse, simplex encoded label interpolation geometry structure is formed, and the nearest center classifier becomes optimal on training data. Through experiments on 11 datasets, we confirm none of the three NC proxies are dispensable, which allows us to obtain very competitive transferability estimation accuracy with approximately \\(10 \\times\\) wall-clock time speed up compared to state-of-the-art approaches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.21, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.764 + ], + "angle": 0, + "content": "The past decade has witnessed a surge in the availability of off-the-shelf pre-trained models in public repositories. They are adapted and re-tuned to facilitate new tasks, which has become a standard practice [12]. Nevertheless, this approach raises an important practical question of how to effectively select the best pre-trained model from a large model pool to be carried onto a downstream task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "While precise ranking of these pre-trained models on a new task can be obtained by a greedy approach of fineturning each model and comparing the test accuracy, it is often prohibitive as it requires a vast amount of computing resources. Therefore, an early line of research has been developed referred to as transferability estimation [22, 41, 24, 29], where the aim is to design efficient methods for ranking the performances of pre-trained models on a downstream task without fine-tuning. Existing" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.292, + 0.872, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.489, + 0.895, + 0.583 + ], + "angle": 0, + "content": "Figure 1: Illustration of the transferability estimation problem. Given a pool of pre-trained models and labeled target data, the objective of transferability estimation is to predict which models can achieve higher performance on the target data after fine-tuning. This task is much more time efficient than fine-tuning the models to obtain ground truth rankings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.764 + ], + "angle": 0, + "content": "approaches in transferability estimation typically consider the current status of the pre-trained models (w.r.t the target task) and can be generally categorized into two streams of probabilistic-based [22, 36, 41, 1] and feature distribution-based techniques [24, 29]. Approaches from the former category model the expected conditional probability of target labels given the extracted target features [41, 20] or predictions [36, 22]. The latter category models transferability as the class-wise separation of features, in which Pandey et al. [24] employed the Bhattacharyya coefficient to approximate the overlap between class distributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "From a different viewpoint for transferability estimation, this paper measures the difference between the current status and the hypothetical terminal status of pre-trained models after fine-tuning. The latter status refers to neural collapse (NC) [25, 33] on the target domain, which commonly occurs when the training loss approaches zero. Specifically, there are three distinct characteristics of NC: 1) the within-class feature variability drops towards zero; 2) class means gradually converge to a simplex encoded label interpolation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5549" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "(SELI) geometry (def. in Sec.3); and 3) the behavior of the final classifier layer is similar to the nearest centroid classifier. Meanwhile, it is empirically demonstrated in [25] that stopping training at the terminal phase achieves higher test performance compared to the models that stop at an earlier stage (i.e., zero prediction error)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.47, + 0.513 + ], + "angle": 0, + "content": "Motivated by these interesting findings, we formalize a transferability metric based on the observations of NC discussed above. The metric measures how close the geometry of the target features is to their hypothetical state in the terminal stage of the fine-tuned model. Specifically, we analyze the spectral property of the class-wise feature matrix to understand to what extent the current status is from zero within-class variation. Secondly, we assess the gap between the features extracted by the pre-trained model and SELI geometry with the rank of the feature matrix. Thirdly, we calculate the difference between the hypothetical classifier layer and the nearest neighbor classifier w.r.t. the empirical class-conditional distribution. Collectively, these three different measurements allow us to understand how far the pre-trained models are from NC on the target domain. On a series of standard benchmarks, we demonstrate that the metric exhibits a strong correlation with the transferability of supervised and self-supervised pre-trained models and thus is a very useful transferability estimator. Compared with the state-of-the-art, our method gives stronger correlations while being computationally efficient. We summarize our main contributions below." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.521, + 0.468, + 0.58 + ], + "angle": 0, + "content": "- We propose the neural collapse transferability index (NCTI) for the transferability estimation. NCTI approximately measures the model w.r.t the three prominent characteristics of NC, which can be efficiently computed." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.588, + 0.468, + 0.647 + ], + "angle": 0, + "content": "- Experimental results on 11 datasets demonstrate strong correlations between NCTI and the transferability of models to the target dataset, which outperforms existing transferability estimation methods." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.654, + 0.468, + 0.729 + ], + "angle": 0, + "content": "- We give interesting insights when the NC components are computed on a subset of the pre-training dataset. Our findings suggest a potential for an unsupervised transferability estimation metric, solely based on measuring within-class variability collapse of the source features." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.521, + 0.468, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.218, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Neural collapse. Initially observed empirically by Papyan et. al [25], the phenomenon of neural collapse has garnered significant research attention within a short time frame. Neural collapse shows that at the terminal phase of deep network training (i.e., driving the training loss towards zero), the last layer activations have the following four properties: 1) the variation of activation drop to negligible; 2) the class mean of activation converge to simplex Equiangular Tight Frame (ETF); 3) the weight of linear classifier" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "and class means converge to a dual-vector space; and 4) the final linear classifier behaves similar to a nearest class center classifier. Recent follow-up works focus on using the unconstrained feature model with cross-entropy [44, 9] or square loss [43, 35] training to theoretically analyze the NC phenomenon. [33, 8] extend the NC phenomenon from a class-balanced scenario to an imbalanced one. Particularly, Thrangpoulidis et. al [33] demonstrated a more generalized geometry than simplex ETF, named simplex encoded labels interpolation (SELI). SELI shows the invariant property on balanced datasets, as well as imbalanced datasets. The authors also proved that SELI recovers the ETF geometry for the balanced dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Transferability estimation. Early works [36, 1, 22] aim to assess the transferability of pre-trained models by calculating the expected conditional distribution of the target label space given the target prediction obtained from the pre-trained classifier. Specifically, [36] leverages negative conditional entropy to evaluate the amount of information shared by pre-training label space and the target task. [22] computes the empirical conditional distribution from the joint distribution of the target task and the pre-training task. Since the early methods heavily rely on the pre-trained predictor to obtain dummy source label distribution of target samples, they cannot directly evaluate the transferability of unsupervised or self-supervised models, where no classifier is available to predict dummy source label distributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Recent transferability estimation methods broaden the applicability by nullifying the dependency on the classifier of the pre-training task. In order to achieve this, NLEEP [20] substitutes the output layer with a Gaussian mixture model. LogME [41] formalizes the transferability estimation as the maximum label marginalized likelihood and adopts a directed graphical model to solve it. Huang et. al [15] proposed to approximate transferability with the mutual information between the pre-trained model extracted features and the corresponding label. On par with the metrics the probabilistic point of view, [29, 24] take the class separability into consideration for the transferability estimation metric. Particularly, Shao et. al [29] incorporates the self-challenge noise augmentation to encourage the classifier to discriminate the hard samples. However, none of the existing works model the transferability as the gap between the current target feature geometry and the hypothetical target feature geometry at the terminal stage of the fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.634, + 0.803 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Transferability estimation is defined as ranking a set of \\(M\\) pre-trained models \\(\\{\\phi^m (\\cdot)\\}_{m = 1}^M\\) given a labeled target dataset \\(D = \\{\\mathbf{X},\\mathbf{Y}\\} = \\{(x_i,y_i)\\}_{i = 1}^N\\) with \\(N\\) being the number of samples. An evaluation metric (accuracy, mAP, etc.) is associated with the dataset, which measures the ground-truth performance \\(T^{m}\\) of the \\(m\\)-th pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5550" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "trained model \\(\\phi^m (\\cdot)\\) after fine-tuning. The feature extracted by the \\(m\\)-th pre-trained model \\(\\phi^m (\\cdot)\\) is denoted as \\(\\mathbf{H}^m\\) where \\(\\mathbf{H}^m = \\phi^m (\\mathbf{X})\\in \\mathbb{R}^{N\\times d}\\) and \\(d\\) is the feature dimension. With the objective of truly reflecting the ground truth model ranking, transferability estimation methods calculate a score \\(S^m\\) for each pre-trained model \\(\\phi^m (\\cdot)\\). Ideally, the calculated scores \\(\\{S^m\\}_{m = 1}^M\\) are supposed to be highly correlated with the ground truth performance \\(\\{T^m\\}_{m = 1}^M\\), and able to determine the best pre-trained model to be fine-tuned on the target data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.409 + ], + "angle": 0, + "content": "The central idea of our proposed method is to employ neural collapse phenomena to assess the proximity of the current state of a pre-trained model \\(\\phi^m (\\cdot)\\) to its terminal stage of fine-tuning on the target. Particularly, the proximity is measured from three dimensions, which are: 1) distance of within-class variability of features from zero; 2) simplicity for the pre-trained model to produce simplex encoded labels interpolation (SELI) geometry structure [33]; and 3) the applicability of the nearest centroid classifier. The final metric jointly considers these three terms, with each term contributing equally to the overall transferability score." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.473, + 0.486 + ], + "angle": 0, + "content": "The subsequent sections elaborate on each of these three perspectives in detail. Here, we start with two definitions that will be used in the sections: the simplex encoded label (SEL) matrix, and the corresponding simplex encoded label interpolation (SELI) geometry structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.501, + 0.485, + 0.516 + ], + "angle": 0, + "content": "Definition 3.1 (Simplex Encoded Label (SEL) Matrix [33])" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.516, + 0.354, + 0.532 + ], + "angle": 0, + "content": "The SEL matrix \\(\\hat{\\mathbf{Z}}\\in \\mathbb{R}^{C\\times N}\\) is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.544, + 0.47, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} [ c, i ] = \\left\\{ \\begin{array}{l l} 1 - 1 / C & , c = y _ {i} \\\\ - 1 / C & , c \\neq y _ {i} \\end{array} . \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.596, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Here, \\(c \\in [C]\\) and \\(i \\in [N]\\) denote the \\(c\\)-th class and \\(i\\)-th observation, respectively. \\(C\\) is the number of classes. Note that \\(\\hat{\\mathbf{Z}}^{\\top} \\mathbf{1} = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.657, + 0.47, + 0.703 + ], + "angle": 0, + "content": "From the definition of SEL matrix \\(\\hat{\\mathbf{Z}}\\), we can observe that the matrix is almost a full rank matrix (\\(rank(\\hat{\\mathbf{Z}}) = \\min(C, N) - 1\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.718, + 0.475, + 0.733 + ], + "angle": 0, + "content": "Definition 3.2 (SEL Interpolation (SELI) Geometry [33])" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.473, + 0.78 + ], + "angle": 0, + "content": "The embeddings \\(\\mathbf{H} \\in \\mathbb{R}^{d \\times N}\\) and classifier weight matrices \\(\\mathbf{W} \\in \\mathbb{R}^{d \\times C}\\) follow the simplex-encoded-labels interpolation geometry when for some scaling \\(\\alpha > 0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.806, + 0.48, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} ^ {\\top} \\mathbf {W} = \\alpha \\mathbf {V} \\boldsymbol {\\Lambda} \\mathbf {V} ^ {\\top}, \\mathbf {H} ^ {\\top} \\mathbf {H} = \\alpha \\mathbf {U} \\boldsymbol {\\Lambda} \\mathbf {U} ^ {\\top}, a n d \\mathbf {W} ^ {\\top} \\mathbf {H} = \\alpha \\hat {\\mathbf {Z}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.838, + 0.47, + 0.884 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{Z}} = \\mathbf{V}\\boldsymbol{\\Lambda}\\mathbf{U}^{\\top}\\) is the SEL matrix. \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\) denote the left and right singular vector matrix of \\(\\hat{\\mathbf{Z}}\\). \\(\\boldsymbol{\\Lambda}\\) represents the diagonal singular value matrix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.791, + 0.108 + ], + "angle": 0, + "content": "3.1. Within-class Variability Collapse" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.236 + ], + "angle": 0, + "content": "As described in the neural collapse phenomenon, at the terminal stage of training, the variability of within-class features becomes negligible and the features collapse towards the corresponding class mean. Motivated by this, we develop a score function \\( S_{vc} \\) to track the within-class variability of features, which indicates how close a pretrained model is to the neural collapse state. Specifically, the within-class covariance \\( \\mathbf{\\Sigma}_c \\) of \\( c \\)-th class is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.25, + 0.892, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {c} = \\frac {1}{n _ {c}} \\left(\\mathbf {H} _ {\\mathbf {c}} - \\mu_ {c}\\right) ^ {\\top} \\left(\\mathbf {H} _ {\\mathbf {c}} - \\mu_ {c}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.351 + ], + "angle": 0, + "content": "where \\(\\pmb{\\Sigma}_{c} \\in \\mathbb{R}^{d \\times d}\\), and \\(\\mathbf{H}_{\\mathbf{c}} \\in \\mathbb{R}^{N_{c} \\times d}\\) consists \\(N_{c}\\) features that belong to \\(c\\)-th class. \\(\\mu_{c}\\) denotes the mean of \\(c\\)-th class. By applying singular value decomposition of \\((\\mathbf{H}_{\\mathbf{c}} - \\mu_{c}) = \\mathbf{U} \\pmb{\\Lambda} \\mathbf{V}^{T}\\), the covariance matrix can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.363, + 0.892, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {c} = \\mathbf {V} \\boldsymbol {\\Lambda} \\mathbf {U} ^ {\\top} \\mathbf {U} \\boldsymbol {\\Lambda} \\mathbf {V} ^ {\\top} = \\mathbf {V} \\boldsymbol {\\Lambda} ^ {2} \\mathbf {V}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\) represent the left and right singular vector matrices. \\(\\Lambda\\) denotes the diagonal singular value matrix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.593 + ], + "angle": 0, + "content": "From Eq. 4, we observe that larger singular values mean higher within-class variability. Since SVD is computationally prohibitive on large matrices, it motivates us to use the nuclear norm, which calculates the sum of singular values of a matrix with lower time complexity. Meanwhile, as feature space is usually with high dimensionality, there might exist noises that are easily mitigated by fine-tuning but negatively affects variability calculation. Therefore, we compute the feature variability by substituting \\(\\mathbf{H}_c^m\\) into class-wise logits \\(\\mathbf{Z}_c^m\\). We define the within-class variability score of \\(S_{vc}\\) using the nuclear norm as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.6, + 0.892, + 0.642 + ], + "angle": 0, + "content": "\\[\nS _ {v c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = - \\sum_ {c = 1} ^ {C} \\left\\| \\mathbf {Z} _ {c} ^ {m} \\right\\| _ {*}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.653, + 0.894, + 0.73 + ], + "angle": 0, + "content": "where \\(\\mathbf{Z}_c^m\\) denotes the logits of \\(c\\)-th class extracted by the \\(m\\)-th model. We defer the formulation of \\(z_{i,c}\\) to section 3.3. Since \\(S_{vc}^{m}\\) directly reflects the within-class variability, when a pre-trained model \\(\\phi^m (\\cdot)\\) is close to the terminal stage of the fine-tuning, a larger \\(S_{vc}^{m}\\) should be achieved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.888, + 0.757 + ], + "angle": 0, + "content": "3.2. Simplex Encoded Labels Interpolation (SELI)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Thrampoulidis et al. [33] proposed SELI as a generalized geometry structure version of the simplex equiangular tight frame (ETF) that is defined in the neural collapse phenomenon. While the simplex ETF geometry observed in neural collapse applies only to balanced datasets, SELI can be established in both balanced and imbalanced datasets at the neural collapse state. As such, we propose a score function \\( S_{seli}^{m} \\) to assess the SELI geometry structure of the target features extracted from the \\( m \\)-th pre-training model. In the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5551" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.271, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.089, + 0.46, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.089, + 0.667, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.089, + 0.86, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Figure 2: TSNE visualization of feature distributions produced by four pre-trained models with CIFAR-10 as the downstream task. Different colors mean different classes in CIFAR-10. From left to right, target recognition accuracy (or transferability) decreases. At the same time, the proposed three scores, i.e., \\( S_{vc} \\), \\( S_{seli} \\), and \\( S_{ncc} \\) also decrease." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.367 + ], + "angle": 0, + "content": "rest of this section, we first review the role of SELI geometry in cross-entropy minimization, followed by discussing our SELI-based metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.444 + ], + "angle": 0, + "content": "Under the unconstrained feature (UF) model [44], Thrampoulidis et al. shows that minimizing the cross-entropy loss converges to a KKT point of the SVM classifier, where the global minimizer of this optimization problem is the SEL matrix." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.451, + 0.471, + 0.497 + ], + "angle": 0, + "content": "Theorem 1 (Structure of the UF-SVM minimizers) [33] Let \\((\\hat{\\mathbf{W}},\\hat{\\mathbf{H}})\\) be any solution and \\(p^*\\) be the optimal cost of the following unconstrained feature SVM (UF-SVM) classifier:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.5, + 0.418, + 0.528 + ], + "angle": 0, + "content": "\\[\n(\\hat {\\mathbf {W}}, \\hat {\\mathbf {H}}) \\in \\underset {\\mathbf {W}, \\mathbf {H}} {\\arg \\min} | | \\mathbf {W} | | _ {F} ^ {2} / 2 + | | \\mathbf {H} | | _ {F} ^ {2} / 2,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.531, + 0.399, + 0.551 + ], + "angle": 0, + "content": "\\[\ns. t. \\left(w _ {y _ {i}} - w _ {c}\\right) ^ {T} h _ {i} \\geq 1, i \\in [ n ], c \\neq y _ {i}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Since \\(\\hat{\\mathbf{Z}} = \\hat{\\mathbf{W}}^{\\top}\\hat{\\mathbf{H}}\\) holds for optimal logits, the optimal cost can be obtained as \\(p^{*} = ||\\hat{\\mathbf{Z}} ||_{*} = ||\\hat{\\mathbf{H}} ||_{F}^{2} = ||\\hat{\\mathbf{W}} ||_{F}^{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.73 + ], + "angle": 0, + "content": "A straightforward way to assess the SELI geometry structure of the target features is to compute the difference between logits \\(\\mathbf{Z}^m\\) extracted from \\(m\\)-th pre-trained model and the optimal logits \\(\\hat{\\mathbf{Z}}\\). However, since the source and target data might not share similar classes, we cannot directly obtain the logits \\(\\mathbf{Z}^m\\) without the time-consuming fine-tuning process on the target dataset. Therefore, we extract features \\(\\mathbf{H}^m\\) of the model and measure its discrepancy to form the SELI structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Specifically, we use the proven fact that, at the terminal stage of the fine-tuning where the model achieves the optimal cost \\( p^* \\), we have \\( ||\\hat{\\mathbf{Z}}||_* = ||\\hat{\\mathbf{H}}||_F^2 \\) [33]. Therefore, we approximate the complexity of achieving the optimal logits \\( \\hat{\\mathbf{Z}} \\) via features \\( \\mathbf{H}^m \\) extracted from the pre-trained model \\( \\phi_m(\\mathbf{X}) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.821, + 0.469, + 0.838 + ], + "angle": 0, + "content": "\\[\nS _ {s e l i} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = \\left\\| \\mathbf {H} ^ {m} \\right\\| _ {*}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Since the nuclear norm is the upper bound of \\(||\\cdot ||_F^2\\) and is the convex envelope for rank(), a higher \\(S_{seli}^{m}\\) indicates a higher rank of the feature matrix \\(\\mathbf{H}_m\\), which makes \\(\\mathbf{Z}\\) closer to a full rank matrix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.322, + 0.847, + 0.339 + ], + "angle": 0, + "content": "3.3. Simplicity to Nearest Centroid Classifier" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Based on the neural collapse phenomenon shown in [25], at the terminal stage of training, the linear classifier behaves similarly to the nearest centroid classifier. Therefore, a score function \\( S_{ncc}^{m} \\) is proposed to measure the applicability of the nearest centroid classifier on the \\( m \\)-th pre-trained model. We start with defining the nearest centroid classifier as [28]:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.451, + 0.892, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\underset {c \\in [ C ]} {\\arg \\max } \\cos \\left(\\mu_ {c}, x\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.891, + 0.514 + ], + "angle": 0, + "content": "where \\(\\hat{y}\\) is the assigned class label to a sample \\(x\\), and \\(\\cos\\) indicates the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.893, + 0.772 + ], + "angle": 0, + "content": "In the neural collapse state, the nearest centroid classifier can achieve the optimal classification performance on the training set. However, in transferability estimation, where there is no actual fine-tuning on the target dataset, directly comparing the performance of applying the nearest centroid classifier to the extracted feature with that of the optimal solution can lead to inaccurate estimation results. To mitigate this issue, we consider two relaxations to the nearest centroid classifier: Instead of using cosine similarity which only involves the mean of the classes, we adopt Mahalanobis distance as the nearest centroid classifier that further involves the correlation between the classes. Moreover, we leverage the expected conditional distribution of labels given input to replace the hard label. Under a mild assumption, the class distribution follows a multivariate Gaussian distribution, thus, we can obtain the following posterior by applying Bayes' Rules:" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\log P (y = c | h) = \\frac {1}{2} \\left(h _ {i} - \\mu_ {c}\\right) ^ {\\top} \\boldsymbol {\\Sigma} \\left(h _ {j} - \\mu_ {c}\\right) + \\log P (y = c). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Here, \\(\\log P(y = c)\\) denotes the prior probability of class \\(c\\), which can be computed as \\(N_{c} / N\\). \\(N_{c}\\) represents the number of samples belonging to class \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Compared with the discrete output of argmax function in the nearest centroid classifier formulation, we adopt the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5552" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.23, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.089, + 0.363, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.089, + 0.493, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.089, + 0.623, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.755, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.089, + 0.887, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.202, + 0.894, + 0.263 + ], + "angle": 0, + "content": "Figure 3: Strong correlation between the proposed NCTI score and model recognition performance (\\%) on the target dataset after fine-tuning. Each marker represents a different pre-trained model. The target domains are (from left to right): Cifar100 [17], DTD [6], Food101 [2], Pets [26], SUN397 [39], and VOC2007 [7]. We show Spearman's coefficient \\(\\rho\\) and weighted Kendall's \\(\\tau\\) here. More correlation results are presented in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.368 + ], + "angle": 0, + "content": "softmax function as it preserves the relative order information of the prediction. Therefore, even if the corresponding class of a sample is not assigned with the highest probability, the probability value still reflects the distance of the sample to that class. The \\( c \\)-th class normalized posterior probability of \\( i \\)-th sample can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.374, + 0.47, + 0.411 + ], + "angle": 0, + "content": "\\[\nz _ {i, c} ^ {m} = \\frac {\\exp \\left(\\log P \\left(y = c \\mid h _ {i} ^ {m}\\right)\\right)}{\\sum_ {k = 1} ^ {C} \\exp \\left(\\log P \\left(y = k \\mid h _ {i} ^ {m}\\right)\\right)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.474 + ], + "angle": 0, + "content": "As such, we have the score \\( S_{ncc}^{m} \\) to describe the discrepancy between the current state of the \\( m \\)-th pre-trained model \\( \\phi_{m}(X) \\) and the optimal nearest centroid classifier as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.474, + 0.469, + 0.514 + ], + "angle": 0, + "content": "\\[\nS _ {n c c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} z _ {i} ^ {m} y _ {i}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.593 + ], + "angle": 0, + "content": "where \\( y_{i} \\) denotes the corresponding one-hot ground truth label of \\( x_{i} \\) and \\( z_{i}^{m} \\) is the \\( C \\) dimensional logits of \\( x_{i} \\), predicted by \\( m \\)-th model. A greater \\( S_{ncc}^{m} \\) indicates a smaller discrepancy towards an optimal nearest centroid classifier, which translates to greater transferability on the target dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.6, + 0.292, + 0.615 + ], + "angle": 0, + "content": "3.4. Overall Score Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.865 + ], + "angle": 0, + "content": "In this work, we model the transferability of a pre-trained model as how similar the current state of the network is to its terminal phase of fine-tuning. Specifically, we expect the terminal phase of fine-tuning to hold the so-called three properties of within-class variability, SELI geometry, and the nearest centroid classifier. Correspondingly, we design three score functions, which are \\( S_{vc}(\\cdot) \\), \\( S_{seli}(\\cdot) \\), and \\( S_{ncc}(\\cdot) \\) to measure the proximity of the state of the pre-trained network applied on the target data versus its optimal state at the terminal phase. However, since the values of score functions have different scales, summing up the scores may hurt the transferability estimation performance (e.g. \\( S_{seli}(\\cdot) \\)) can numerically dominate over \\( S_{ncc}(\\cdot) \\)). Instead of having manually defined hyper-parameters to balance the contribution of each score, we propose to normalize each score into a unit range (0 to 1) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.871, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\nS _ {s e l i} ^ {m} \\leftarrow \\frac {S _ {s e l i} ^ {m} - \\operatorname* {m i n} \\left(S _ {s e l i}\\right)}{\\operatorname* {m a x} \\left(S _ {s e l i}\\right) - \\operatorname* {m i n} \\left(S _ {s e l i}\\right)}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Similarly, we can have the normalized score \\( S_{vc}(H^m) \\) and \\( S_{ncc}(H^m) \\) for the \\( m \\)-th pretrained model. By summing the normalized scores, we have the final transferability estimation metric:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.346, + 0.892, + 0.365 + ], + "angle": 0, + "content": "\\[\nS _ {t o t a l} ^ {m} = S _ {v c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) + S _ {s e l i} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) + S _ {n c c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.372, + 0.892, + 0.418 + ], + "angle": 0, + "content": "where the \\(m\\)-th pre-trained model with a higher overall score \\(S_{total}^{m}\\) indicates a better transferability in the model pool for the target dataset \\(D\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.625, + 0.447 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.787, + 0.47 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Datasets. We consider a wide range of classification benchmark datasets, including five fine-grained classifications (FGVC Aircraft [21], Standford Cars [16], Food101 [2], Oxford-IIIT Pets [26], Oxford-102 Flowers [23]), four coarse-grained classifications (Caltech101 [18], CIFAR-10 [17], CIFAR-100 [17], VOC2007 [7]), one scene classification (SUN397 [39]), and one texture classification (DTD [6]). In total, we adopt 11 benchmark datasets that are broadly used in transfer learning. A detailed description of each dataset can be found in the supplementary material. All models have undergone pre-training on the ImageNet dataset, and the fine-tuned performance of the model on target datasets is obtained from [29]. Without affecting the fine-tuning performance, we note that only the training and validation splits of the dataset are used for transferability estimation, and we hold the test split out of the estimation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Correlation measurement. Following the previous work [41, 24, 29], we adopt weighted Kendall's \\(\\tau\\) [37] to measure the correlation between the estimated model ranking and the ground-truth performance ranking. Each pairwise comparison between items in two rankings is assigned a weight based on the distance between their ranks, whereas top-ranked items will be assigned with higher weight." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. In order to replicate the fine-tuning process, our proposed method aims to find a projection matrix \\(W\\) that maximizes the discriminative power of the extracted features \\(\\phi^m (X)\\) from the \\(m\\)-th pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5553" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 1: Method comparison of their correlation strength with target accuracy for supervised models. Weighted Kendall's \\(\\tau\\) on 11 target test sets and their average are shown. For each column, the best, and second-best results are in bold, and underlined, respectively. Our method achieves the best overall weighted Kendall's \\(\\tau\\)." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.147, + 0.86, + 0.248 + ], + "angle": 0, + "content": "
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NCE [36]-0.1610.4650.6850.7090.7230.302-0.4820.6270.7720.7600.5710.452
LEEP [22]-0.2770.6050.3670.8240.6770.486-0.2910.4340.3890.6580.4130.390
LogME [41]0.4390.4630.6050.8520.7250.7000.1470.3850.4110.5110.6950.539
NLEEP [20]-0.5310.6140.4890.8250.7310.8200.0540.5290.9550.8480.6990.548
SFDA [29]0.6140.6960.5180.9490.8660.5750.5140.8150.5220.5580.6710.663
NCTI0.4960.6790.6470.8430.8790.7040.5410.7730.9240.7560.7410.726
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.271, + 0.273, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.419, + 0.243, + 0.432 + ], + "angle": 0, + "content": "(a) Supervised models" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.271, + 0.468, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.418, + 0.453, + 0.432 + ], + "angle": 0, + "content": "(b) Self-supervised models" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Figure 4: Method comparison w.r.t running time (seconds) vs. average Weighted Kendall's \\(\\tau\\) on 11 datasets on (a) supervised and (b) self-supervised models. Our method obtains the highest ranking correlation within a small time budget." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.473, + 0.729 + ], + "angle": 0, + "content": "model. To achieve this, we employ linear discriminant analysis (LDA) to maximize the separation between the class means of the projected data. However, LDA can be prone to overfitting when applied to high-dimensional feature spaces or when the covariance matrix is singular [40, 42]. To mitigate the overfitting issue on high-dimensional feature space, we follow the previous transferability estimation methods [20, 24] which incorporate principal component analysis (PCA) into the framework to project the original features onto a lower-dimensional subspace. We further address the singularity issue by shrinking the covariance matrix toward a full-rank identity matrix as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.75, + 0.47, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {w} ^ {\\prime} = (1 - \\alpha) \\boldsymbol {\\Sigma} _ {w} + \\alpha \\mathbf {I}, \\text {w h e r e} \\alpha = \\exp \\left(\\sum_ {i} ^ {d ^ {*}} \\lambda_ {i}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.471, + 0.906 + ], + "angle": 0, + "content": "Here, \\(\\lambda_{i}\\) denotes the \\(i\\)-th largest eigenvalue of the PCA projected feature matrix \\(PCA(\\phi_m(X))\\). \\(d^{*}\\) indicates the number of eigenvalues to be considered in the calculation. We set the dimension of the PCA projected subspace as 64 and set \\(d^{*}\\) as 32 throughout all the experiments. \\(\\alpha\\) controls the trade-off between the covariance and the identity matrices." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.272, + 0.795, + 0.289 + ], + "angle": 0, + "content": "4.2. Evaluation on Supervised Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.401 + ], + "angle": 0, + "content": "Overview. To initiate our evaluation of transferability estimation metrics for ranking supervised models, we assess the performance of these metrics on 11 widely adopted models, which are ResNet-34 [13], ResNet-50, ResNet-101, ResNet-151, DenseNet-121 [14], DenseNet-169, DenseNet-201, MNet-A1 [32], MobileNet-v2 [27], GoogleNet [30], and Inception-v3 [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.893, + 0.673 + ], + "angle": 0, + "content": "Performance comparison. We compare our proposed NCTI with existing transferability estimation metrics, including NCE [36], LEEP [22], LogME [41], NLEEP [20], and SFDA [29]. The experimental results are shown in Tab. 1. From the table, we can observe that empirical conditional probability-based methods (i.e., NCE, LEEP, and LogME) generally achieve lower ranking correlations than our method and SFDA, indicating the importance of modeling the feature space in transferability estimation. NCTI achieves the best or the second best average weighted Kendall's ranking correlation on nine datasets and obtains a \\(6.9\\%\\) averaged performance gain compared to the second-best method SFDA. We argue that while the neural collapse explicitly reflects the maximum class separation, a feature space with a high class separation score can not necessarily leads to the neural collapse phenomenon. Therefore, our method can achieve superior performance than class separation-based methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.682, + 0.825, + 0.699 + ], + "angle": 0, + "content": "4.3. Evaluation on Self-supervised Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Overview. We further evaluate the effectiveness of NCTI in assessing the transferability of models pre-trained by self-supervised learning (SSL). We construct a pool with 10 SSL pre-trained models, including MoCo-v1 [11], MoCo-v2 [5], PCL-v2 [19], SELA-V2, Deepcluster-v2 [4], BYOL [10], Infomin [34], SWAV [3], and Insdis [38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Performance comparison. Since NCE and LEEP require the classifier on the pre-training task, it is not directly applicable to self-supervised model ranking tasks. Therefore, to evaluate the ranking performance of the self-supervised models, we compare our method with NLEEP, LogME, and SFDA in Tab. 2. Similarly, we observe that the average ranking correlation of SFDA and the proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5554" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.121 + ], + "angle": 0, + "content": "Table 2: Method comparison of their correlation strength with target accuracy for self-supervised models. All the settings and notations are the same as Table 1. Our method achieves the best overall weighted Kendall's \\( \\tau \\) ." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.131, + 0.86, + 0.206 + ], + "angle": 0, + "content": "
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NLEEP [20]-0.2860.6620.5950.1080.3740.7790.5980.7160.8640.880-0.0910.473
LogME [41]0.0210.0750.6270.4170.1460.7430.7630.6860.7380.2600.1810.423
SFDA [29]0.1670.6740.6830.8460.7890.8820.8970.8370.5640.8310.6210.708
NCTI0.0360.8110.7960.7580.8110.7960.7620.9450.8050.7740.6060.719
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.242, + 0.273, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.396, + 0.243, + 0.41 + ], + "angle": 0, + "content": "(a) Supervised models" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.242, + 0.465, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.396, + 0.453, + 0.41 + ], + "angle": 0, + "content": "(b) Self-supervised models" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.471, + 0.482 + ], + "angle": 0, + "content": "Figure 5: Effectiveness of each individual component in NCTI. We use the three terms individually or remove them one at a time from the full system. Results on (a) supervised and (b) self-supervised models are shown." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.577 + ], + "angle": 0, + "content": "method can achieve superior results than that of the conditional probability-based method. NCTI surpasses the second-best method by \\(1.5\\%\\) on average and scores the highest overall weighted Kendall's \\(\\tau\\), which verifies the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.245, + 0.605 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Tranferability estimation with fewer target data. We test a practical but challenging transferability estimation scenario, where only limited target data are available for the estimation. We randomly sample 2 to 500 images per class from the full SUN397 and examine how limited data affects the performance of different methods. As we can see from Fig. 6(d), although fewer data generally trigger performance degradation for transferability estimation methods, our method is able to achieve a larger performance gain under the low shot scheme than using the full dataset. Particularly, we highlight that our method can achieve SOTA performance on SUN397 with only 10 samples per class, which indicates a strong potential for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Computational efficiency comparison. We compare the time efficiency versus ranking correlation of different transferability estimation metrics in Fig. 4. The \\( x \\)-axis is the total time cost for estimating transferability over 11 datasets, while the \\( y \\)-axis indicates the average ranking correlation on 11 datasets. We adopt a logarithmic scale for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.426 + ], + "angle": 0, + "content": "the \\(x\\)-axis. The figure shows that within a similar time budget, NCTI can achieve significantly higher ranking correlations compared with empirical conditional probability-based methods, such as NCE, LEEP, and LogME. SFDA can achieve a comparable ranking correlation with NCTI on self-supervised model ranking tasks. However, it consumes more than \\(10 \\times\\) more time to estimate the transferability, showing that it is less efficient than our proposed metric. Meanwhile, we can observe from the figure that our method scores the highest ranking correlation to the running time ratio under both settings, emphasizing the effectiveness and efficiency of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.893, + 0.837 + ], + "angle": 0, + "content": "Ablation study. The NCTI metric proposed in this work is composed of three terms, each term is designed to describe a specific aspect of the gap between the current feature geometry and the geometry of neural collapse. In order to illustrate the impact of each individual term in NCTI, we conduct ablation studies to show how each individual term contributes to the overall metric. The ablation studies are conducted on both supervised and self-supervision scenarios, and the results are shown in Fig. 5. Among these three individual terms of NCTI, although \\( S_{vc} \\) can attain a positive averaged ranking correlation, the contribution towards the full metric is less than the other two terms. \\( S_{seli} \\) delivers the highest average ranking correlation on supervised model estimation, while \\( S_{ncc} \\) attains the best-averaged performance on self-supervised model estimation. We infer this is due to the difference in the supervision signal. As the label information is not available for self-supervised training, the discriminative power of the pre-trained feature is more critical for the downstream task than that in supervised models. We also validate the effectiveness of two-term combinations. As we can see from the figure, the combination between \\( S_{seli} \\) and \\( S_{vc} \\) always achieves higher correlations than its individual component, confirming the terms are complementary. Although NCTI obtains a slightly lower correlation than w/o \\( S_{ncc} \\) on the supervised models, it achieves an absolute \\( 4.2\\% \\) higher averaged correlation of both tasks, which verifies the importance of \\( S_{ncc} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Hyperparameter analysis. NCTI simplifies the hyperparameter tuning process by assigning equal weights to each term. In order to validate this weight assignment, we investigate the impact of weight changes on the ranking cor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.268, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.222, + 0.235, + 0.236 + ], + "angle": 0, + "content": "(a) Weight of \\( S_{vc} \\)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.089, + 0.473, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.222, + 0.441, + 0.237 + ], + "angle": 0, + "content": "(b) Weight of \\(S_{ncc}\\)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.673, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.222, + 0.643, + 0.237 + ], + "angle": 0, + "content": "(c) Weight of \\(S_{seli}\\)" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.09, + 0.872, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.222, + 0.827, + 0.236 + ], + "angle": 0, + "content": "(d) SUN397" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.895, + 0.326 + ], + "angle": 0, + "content": "Figure 6: (a)-(c) Hyperparameter analysis of three components in NCTI. We report correlation strength \\(\\tau\\) computed on 11 downstream datasets against weight values. Impact on both supervised and self-supervised pre-trained models are drawn. We study each weight by fixing the rest as 1. In this paper, by default, we set the weight to 1 for all components. (d) Influence of target data size for estimation on the correlation strength. We vary the data size from 2 to 500 samples per class and observe that our method can achieve a larger performance margin compared with existing methods under the low-shot scenario." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.337, + 0.273, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.487, + 0.243, + 0.501 + ], + "angle": 0, + "content": "(a) Supervised models" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.338, + 0.468, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.487, + 0.453, + 0.501 + ], + "angle": 0, + "content": "(b) Self-supervised models" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.575 + ], + "angle": 0, + "content": "Figure 7: Effectiveness of the three components computed on a subset of the pre-training dataset. Average correlation strength \\((\\tau)\\) on 11 target datasets reported. We assess (a) supervised models, and (b) self-supervised models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.776 + ], + "angle": 0, + "content": "relation. Specifically, we vary the weight of the term of interest within the range of \\(\\{0.1, 0.4, 0.8, 1.2, 1.6, 2\\}\\) while keeping the weight of the remaining terms fixed at 1. The experimental results of the hyperparameter analysis are reported in Fig.6 (a)-(c). Our findings indicate that NCTI achieves high averaged correlations (\\(\\tau \\geq 0.7\\)) with the most weight configurations. While we observe different monotonicity between the correlations of supervised and self-supervised tasks, the results demonstrate that harmonic points (i.e., high mean and low variance) are consistently achieved when all weights are close to 1, thereby confirming the appropriateness of the equal weight setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Analysis of NCTI computed using source features. The extensive experiments conducted so far have demonstrated the effectiveness of each individual term in the proposed NCTI metric for transferability estimation. All the previous experiments have been conducted by using the target features extracted from the pre-trained models. It remains important to investigate what the properties of neural collapse on the source (i.e., pre-training) features tell us" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.61 + ], + "angle": 0, + "content": "about the transferability of a model to a downstream task. To this end, we measure each individual term in NCTI on a subset of the source dataset and report the ranking correlation of the supervised model estimation task in Fig.7(a) and (b). On average, the transferability of a supervised model to a downstream task is consistently positively related to \\(-S_{vc}\\) and \\(S_{ncc}\\) on the source dataset on both supervised and self-supervised, which scores 0.725 and 0.598 weighted Kendall \\(\\tau\\), respectively. While the figure shows \\(S_{seli}\\) can achieve a positive correlation on supervised tasks, no significant correlation on \\(S_{seli}\\) is observed in the supervised transfer estimation task. Note that, \\(-S_{vc}\\) achieves the highest ranking correlation on average, which shows that a pre-trained model with higher within-class variability can better transfer to a new task. This indicates that a fully unsupervised metric can be designed for transferability estimation which is solely based on measuring the class-wise compactness of the source features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.622, + 0.619, + 0.637 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.868 + ], + "angle": 0, + "content": "How to efficiently choose the optimal pre-trained model from a vast collection of models to use in a downstream task? To rank models based on their transferability, we propose a metric named Neural Collapse informed Transferability Index (NCTI). Our method models model transferability as the gap between the current feature geometry and the geometry at the terminal stage (i.e., neural collapse) after hypothetical fine-tuning on the downstream task. Specifically, we model the gap from three perspectives, including the formation of SELI geometry structure, the within-class variability, and the applicability of the nearest center classifier. We show that our method is light to compute and that the ranking of model transferability has a very strong correlation with the ground truth ranking and compares favorably against the state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This work was partially supported by ARC DP 230101196" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5556" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In 2019 IEEE international conference on image processing (ICIP), pages 2309-2313. IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[5] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.47, + 0.48 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[7] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[8] Cong Fang, Hangfeng He, Qi Long, and Weijie J Su. Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training. Proceedings of the National Academy of Sciences, 118(43):e2103091118, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[9] Florian Graf, Christoph Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised contrastive learning. In International Conference on Machine Learning, pages 3821-3830. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[12] Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imageNet pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4918-4927, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[15] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[16] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[18] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, Apr 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[19] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[20] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2663-2673, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[22] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[23] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[24] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9172-9182, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[25] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[26] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[27] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval, volume 39. Cambridge University Press Cambridge, 2008. 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5557" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[29] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 286-302. Springer, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.19, + 0.472, + 0.274 + ], + "angle": 0, + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.276, + 0.472, + 0.344 + ], + "angle": 0, + "content": "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.346, + 0.472, + 0.416 + ], + "angle": 0, + "content": "[32] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.472, + 0.46 + ], + "angle": 0, + "content": "[33] Christos Thrampoulidis, Ganesh R Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. arXiv preprint arXiv:2208.05512, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[34] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[35] Tom Tirer and Joan Bruna. Extended unconstrained features model for exploring deep neural collapse. In International Conference on Machine Learning, pages 21478-21505. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[36] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1395-1405, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[37] Sebastiano Vigna. A weighted correlation index for rankings with ties. In Proceedings of the 24th international conference on World Wide Web, pages 1166-1176, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.741 + ], + "angle": 0, + "content": "[38] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[39] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[40] Jian Yang and Jing-yu Yang. Why can lda be performed in pca transformed space? Pattern recognition, 36(2):563-566, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[41] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Ma" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "chine Learning, pages 12133-12143. PMLR, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[42] Hua Yu and Jie Yang. A direct lda algorithm for high-dimensional data—with application to face recognition. Pattern recognition, 34(10):2067-2070, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[43] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse undermse loss: Global optimality with unconstrained features. In International Conference on Machine Learning, pages 27179-27202. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[44] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. Advances in Neural Information Processing Systems, 34:29820-29834, 2021. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5558" + } + ] +] \ No newline at end of file diff --git a/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_origin.pdf b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1b7f4921982bf168f5e327e25273136601107f50 --- /dev/null +++ b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/336dd7d0-3ba4-4bc4-a56f-4e1851a25562_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ad7198adb7050ac1ab80a27b679cd1b216f5ca73c9c395f5a62529c32532625 +size 690815 diff --git a/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/full.md b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..084db8064bf23bf05a14fd7c3edd89eff0177f9c --- /dev/null +++ b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/full.md @@ -0,0 +1,348 @@ +# How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability + +Zijian Wang $^{1}$ Yadan Luo $^{1}$ Liang Zheng $^{2}$ Zi Huang $^{1}$ Mahsa Baktashmotlagh $^{1}$ $^{1}$ The University of Queensland $^{2}$ Australian National University + +{zijian.wang, y.luo, helen.huang, m.baktashmotlagh}@uq.edu.au, liang.zheng@anu.edu.au + +# Abstract + +This paper focuses on model transferability estimation, i.e., assessing the performance of pre-trained models on a downstream task without performing fine-tuning. Motivated by neural collapse (NC) [25] that reveals particular feature geometry at the terminal stage of training, we consider model transferability as how far the target activations obtained by pre-trained models are from their hypothetical state in the terminal phase of the model fine-tuned on the target domain. We propose a metric that measures this proximity based on three phenomena of NC: within-class variability collapse, simplex encoded label interpolation geometry structure is formed, and the nearest center classifier becomes optimal on training data. Through experiments on 11 datasets, we confirm none of the three NC proxies are dispensable, which allows us to obtain very competitive transferability estimation accuracy with approximately $10 \times$ wall-clock time speed up compared to state-of-the-art approaches. + +# 1. Introduction + +The past decade has witnessed a surge in the availability of off-the-shelf pre-trained models in public repositories. They are adapted and re-tuned to facilitate new tasks, which has become a standard practice [12]. Nevertheless, this approach raises an important practical question of how to effectively select the best pre-trained model from a large model pool to be carried onto a downstream task. + +While precise ranking of these pre-trained models on a new task can be obtained by a greedy approach of fineturning each model and comparing the test accuracy, it is often prohibitive as it requires a vast amount of computing resources. Therefore, an early line of research has been developed referred to as transferability estimation [22, 41, 24, 29], where the aim is to design efficient methods for ranking the performances of pre-trained models on a downstream task without fine-tuning. Existing + +![](images/a74933bc6341a105e9721e5c4ab6f3ebdadd1f7f32acdeb840c1f175888a8886.jpg) +Figure 1: Illustration of the transferability estimation problem. Given a pool of pre-trained models and labeled target data, the objective of transferability estimation is to predict which models can achieve higher performance on the target data after fine-tuning. This task is much more time efficient than fine-tuning the models to obtain ground truth rankings. + +approaches in transferability estimation typically consider the current status of the pre-trained models (w.r.t the target task) and can be generally categorized into two streams of probabilistic-based [22, 36, 41, 1] and feature distribution-based techniques [24, 29]. Approaches from the former category model the expected conditional probability of target labels given the extracted target features [41, 20] or predictions [36, 22]. The latter category models transferability as the class-wise separation of features, in which Pandey et al. [24] employed the Bhattacharyya coefficient to approximate the overlap between class distributions. + +From a different viewpoint for transferability estimation, this paper measures the difference between the current status and the hypothetical terminal status of pre-trained models after fine-tuning. The latter status refers to neural collapse (NC) [25, 33] on the target domain, which commonly occurs when the training loss approaches zero. Specifically, there are three distinct characteristics of NC: 1) the within-class feature variability drops towards zero; 2) class means gradually converge to a simplex encoded label interpolation + +(SELI) geometry (def. in Sec.3); and 3) the behavior of the final classifier layer is similar to the nearest centroid classifier. Meanwhile, it is empirically demonstrated in [25] that stopping training at the terminal phase achieves higher test performance compared to the models that stop at an earlier stage (i.e., zero prediction error). + +Motivated by these interesting findings, we formalize a transferability metric based on the observations of NC discussed above. The metric measures how close the geometry of the target features is to their hypothetical state in the terminal stage of the fine-tuned model. Specifically, we analyze the spectral property of the class-wise feature matrix to understand to what extent the current status is from zero within-class variation. Secondly, we assess the gap between the features extracted by the pre-trained model and SELI geometry with the rank of the feature matrix. Thirdly, we calculate the difference between the hypothetical classifier layer and the nearest neighbor classifier w.r.t. the empirical class-conditional distribution. Collectively, these three different measurements allow us to understand how far the pre-trained models are from NC on the target domain. On a series of standard benchmarks, we demonstrate that the metric exhibits a strong correlation with the transferability of supervised and self-supervised pre-trained models and thus is a very useful transferability estimator. Compared with the state-of-the-art, our method gives stronger correlations while being computationally efficient. We summarize our main contributions below. + +- We propose the neural collapse transferability index (NCTI) for the transferability estimation. NCTI approximately measures the model w.r.t the three prominent characteristics of NC, which can be efficiently computed. +- Experimental results on 11 datasets demonstrate strong correlations between NCTI and the transferability of models to the target dataset, which outperforms existing transferability estimation methods. +- We give interesting insights when the NC components are computed on a subset of the pre-training dataset. Our findings suggest a potential for an unsupervised transferability estimation metric, solely based on measuring within-class variability collapse of the source features. + +# 2. Related Work + +Neural collapse. Initially observed empirically by Papyan et. al [25], the phenomenon of neural collapse has garnered significant research attention within a short time frame. Neural collapse shows that at the terminal phase of deep network training (i.e., driving the training loss towards zero), the last layer activations have the following four properties: 1) the variation of activation drop to negligible; 2) the class mean of activation converge to simplex Equiangular Tight Frame (ETF); 3) the weight of linear classifier + +and class means converge to a dual-vector space; and 4) the final linear classifier behaves similar to a nearest class center classifier. Recent follow-up works focus on using the unconstrained feature model with cross-entropy [44, 9] or square loss [43, 35] training to theoretically analyze the NC phenomenon. [33, 8] extend the NC phenomenon from a class-balanced scenario to an imbalanced one. Particularly, Thrangpoulidis et. al [33] demonstrated a more generalized geometry than simplex ETF, named simplex encoded labels interpolation (SELI). SELI shows the invariant property on balanced datasets, as well as imbalanced datasets. The authors also proved that SELI recovers the ETF geometry for the balanced dataset. + +Transferability estimation. Early works [36, 1, 22] aim to assess the transferability of pre-trained models by calculating the expected conditional distribution of the target label space given the target prediction obtained from the pre-trained classifier. Specifically, [36] leverages negative conditional entropy to evaluate the amount of information shared by pre-training label space and the target task. [22] computes the empirical conditional distribution from the joint distribution of the target task and the pre-training task. Since the early methods heavily rely on the pre-trained predictor to obtain dummy source label distribution of target samples, they cannot directly evaluate the transferability of unsupervised or self-supervised models, where no classifier is available to predict dummy source label distributions. + +Recent transferability estimation methods broaden the applicability by nullifying the dependency on the classifier of the pre-training task. In order to achieve this, NLEEP [20] substitutes the output layer with a Gaussian mixture model. LogME [41] formalizes the transferability estimation as the maximum label marginalized likelihood and adopts a directed graphical model to solve it. Huang et. al [15] proposed to approximate transferability with the mutual information between the pre-trained model extracted features and the corresponding label. On par with the metrics the probabilistic point of view, [29, 24] take the class separability into consideration for the transferability estimation metric. Particularly, Shao et. al [29] incorporates the self-challenge noise augmentation to encourage the classifier to discriminate the hard samples. However, none of the existing works model the transferability as the gap between the current target feature geometry and the hypothetical target feature geometry at the terminal stage of the fine-tuning. + +# 3. Methodology + +Transferability estimation is defined as ranking a set of $M$ pre-trained models $\{\phi^m (\cdot)\}_{m = 1}^M$ given a labeled target dataset $D = \{\mathbf{X},\mathbf{Y}\} = \{(x_i,y_i)\}_{i = 1}^N$ with $N$ being the number of samples. An evaluation metric (accuracy, mAP, etc.) is associated with the dataset, which measures the ground-truth performance $T^{m}$ of the $m$ -th pre + +trained model $\phi^m (\cdot)$ after fine-tuning. The feature extracted by the $m$ -th pre-trained model $\phi^m (\cdot)$ is denoted as $\mathbf{H}^m$ where $\mathbf{H}^m = \phi^m (\mathbf{X})\in \mathbb{R}^{N\times d}$ and $d$ is the feature dimension. With the objective of truly reflecting the ground truth model ranking, transferability estimation methods calculate a score $S^m$ for each pre-trained model $\phi^m (\cdot)$ . Ideally, the calculated scores $\{S^m\}_{m = 1}^M$ are supposed to be highly correlated with the ground truth performance $\{T^m\}_{m = 1}^M$ , and able to determine the best pre-trained model to be fine-tuned on the target data. + +The central idea of our proposed method is to employ neural collapse phenomena to assess the proximity of the current state of a pre-trained model $\phi^m (\cdot)$ to its terminal stage of fine-tuning on the target. Particularly, the proximity is measured from three dimensions, which are: 1) distance of within-class variability of features from zero; 2) simplicity for the pre-trained model to produce simplex encoded labels interpolation (SELI) geometry structure [33]; and 3) the applicability of the nearest centroid classifier. The final metric jointly considers these three terms, with each term contributing equally to the overall transferability score. + +The subsequent sections elaborate on each of these three perspectives in detail. Here, we start with two definitions that will be used in the sections: the simplex encoded label (SEL) matrix, and the corresponding simplex encoded label interpolation (SELI) geometry structure. + +# Definition 3.1 (Simplex Encoded Label (SEL) Matrix [33]) + +The SEL matrix $\hat{\mathbf{Z}}\in \mathbb{R}^{C\times N}$ is defined as, + +$$ +\hat {\mathbf {Z}} [ c, i ] = \left\{ \begin{array}{l l} 1 - 1 / C & , c = y _ {i} \\ - 1 / C & , c \neq y _ {i} \end{array} . \right. \tag {1} +$$ + +Here, $c \in [C]$ and $i \in [N]$ denote the $c$ -th class and $i$ -th observation, respectively. $C$ is the number of classes. Note that $\hat{\mathbf{Z}}^{\top} \mathbf{1} = 0$ . + +From the definition of SEL matrix $\hat{\mathbf{Z}}$ , we can observe that the matrix is almost a full rank matrix ( $rank(\hat{\mathbf{Z}}) = \min(C, N) - 1$ ). + +# Definition 3.2 (SEL Interpolation (SELI) Geometry [33]) + +The embeddings $\mathbf{H} \in \mathbb{R}^{d \times N}$ and classifier weight matrices $\mathbf{W} \in \mathbb{R}^{d \times C}$ follow the simplex-encoded-labels interpolation geometry when for some scaling $\alpha > 0$ : + +$$ +\mathbf {W} ^ {\top} \mathbf {W} = \alpha \mathbf {V} \boldsymbol {\Lambda} \mathbf {V} ^ {\top}, \mathbf {H} ^ {\top} \mathbf {H} = \alpha \mathbf {U} \boldsymbol {\Lambda} \mathbf {U} ^ {\top}, a n d \mathbf {W} ^ {\top} \mathbf {H} = \alpha \hat {\mathbf {Z}}, \tag {2} +$$ + +where $\hat{\mathbf{Z}} = \mathbf{V}\boldsymbol{\Lambda}\mathbf{U}^{\top}$ is the SEL matrix. $\mathbf{U}$ and $\mathbf{V}$ denote the left and right singular vector matrix of $\hat{\mathbf{Z}}$ . $\boldsymbol{\Lambda}$ represents the diagonal singular value matrix. + +# 3.1. Within-class Variability Collapse + +As described in the neural collapse phenomenon, at the terminal stage of training, the variability of within-class features becomes negligible and the features collapse towards the corresponding class mean. Motivated by this, we develop a score function $S_{vc}$ to track the within-class variability of features, which indicates how close a pretrained model is to the neural collapse state. Specifically, the within-class covariance $\mathbf{\Sigma}_c$ of $c$ -th class is denoted as: + +$$ +\boldsymbol {\Sigma} _ {c} = \frac {1}{n _ {c}} \left(\mathbf {H} _ {\mathbf {c}} - \mu_ {c}\right) ^ {\top} \left(\mathbf {H} _ {\mathbf {c}} - \mu_ {c}\right), \tag {3} +$$ + +where $\pmb{\Sigma}_{c} \in \mathbb{R}^{d \times d}$ , and $\mathbf{H}_{\mathbf{c}} \in \mathbb{R}^{N_{c} \times d}$ consists $N_{c}$ features that belong to $c$ -th class. $\mu_{c}$ denotes the mean of $c$ -th class. By applying singular value decomposition of $(\mathbf{H}_{\mathbf{c}} - \mu_{c}) = \mathbf{U} \pmb{\Lambda} \mathbf{V}^{T}$ , the covariance matrix can be written as: + +$$ +\boldsymbol {\Sigma} _ {c} = \mathbf {V} \boldsymbol {\Lambda} \mathbf {U} ^ {\top} \mathbf {U} \boldsymbol {\Lambda} \mathbf {V} ^ {\top} = \mathbf {V} \boldsymbol {\Lambda} ^ {2} \mathbf {V}. \tag {4} +$$ + +Here, $\mathbf{U}$ and $\mathbf{V}$ represent the left and right singular vector matrices. $\Lambda$ denotes the diagonal singular value matrix. + +From Eq. 4, we observe that larger singular values mean higher within-class variability. Since SVD is computationally prohibitive on large matrices, it motivates us to use the nuclear norm, which calculates the sum of singular values of a matrix with lower time complexity. Meanwhile, as feature space is usually with high dimensionality, there might exist noises that are easily mitigated by fine-tuning but negatively affects variability calculation. Therefore, we compute the feature variability by substituting $\mathbf{H}_c^m$ into class-wise logits $\mathbf{Z}_c^m$ . We define the within-class variability score of $S_{vc}$ using the nuclear norm as follows: + +$$ +S _ {v c} ^ {m} \left(\mathbf {H} ^ {m}\right) = - \sum_ {c = 1} ^ {C} \left\| \mathbf {Z} _ {c} ^ {m} \right\| _ {*}, \tag {5} +$$ + +where $\mathbf{Z}_c^m$ denotes the logits of $c$ -th class extracted by the $m$ -th model. We defer the formulation of $z_{i,c}$ to section 3.3. Since $S_{vc}^{m}$ directly reflects the within-class variability, when a pre-trained model $\phi^m (\cdot)$ is close to the terminal stage of the fine-tuning, a larger $S_{vc}^{m}$ should be achieved. + +# 3.2. Simplex Encoded Labels Interpolation (SELI) + +Thrampoulidis et al. [33] proposed SELI as a generalized geometry structure version of the simplex equiangular tight frame (ETF) that is defined in the neural collapse phenomenon. While the simplex ETF geometry observed in neural collapse applies only to balanced datasets, SELI can be established in both balanced and imbalanced datasets at the neural collapse state. As such, we propose a score function $S_{seli}^{m}$ to assess the SELI geometry structure of the target features extracted from the $m$ -th pre-training model. In the + +![](images/41fdd5cc049dd0f628a6eb83df8f3d250b69fb077de4312f2c48f9977bbe1b73.jpg) +Figure 2: TSNE visualization of feature distributions produced by four pre-trained models with CIFAR-10 as the downstream task. Different colors mean different classes in CIFAR-10. From left to right, target recognition accuracy (or transferability) decreases. At the same time, the proposed three scores, i.e., $S_{vc}$ , $S_{seli}$ , and $S_{ncc}$ also decrease. + +![](images/202c64155dd48ef7ee69021c461344ad4c3c59d7fba56520ecfa8e7831fa2f38.jpg) + +![](images/9b8f812051387f72a518d1aacafa7a49ce9e510b1a874894affea420750677e8.jpg) + +![](images/3a15a5fa761e78720e5abf6e43bf17c5f9f796af9306efde1e1a9c7d947c87c0.jpg) + +rest of this section, we first review the role of SELI geometry in cross-entropy minimization, followed by discussing our SELI-based metric. + +Under the unconstrained feature (UF) model [44], Thrampoulidis et al. shows that minimizing the cross-entropy loss converges to a KKT point of the SVM classifier, where the global minimizer of this optimization problem is the SEL matrix. + +Theorem 1 (Structure of the UF-SVM minimizers) [33] Let $(\hat{\mathbf{W}},\hat{\mathbf{H}})$ be any solution and $p^*$ be the optimal cost of the following unconstrained feature SVM (UF-SVM) classifier: + +$$ +(\hat {\mathbf {W}}, \hat {\mathbf {H}}) \in \underset {\mathbf {W}, \mathbf {H}} {\arg \min} | | \mathbf {W} | | _ {F} ^ {2} / 2 + | | \mathbf {H} | | _ {F} ^ {2} / 2, +$$ + +$$ +s. t. \left(w _ {y _ {i}} - w _ {c}\right) ^ {T} h _ {i} \geq 1, i \in [ n ], c \neq y _ {i}. +$$ + +Since $\hat{\mathbf{Z}} = \hat{\mathbf{W}}^{\top}\hat{\mathbf{H}}$ holds for optimal logits, the optimal cost can be obtained as $p^{*} = ||\hat{\mathbf{Z}} ||_{*} = ||\hat{\mathbf{H}} ||_{F}^{2} = ||\hat{\mathbf{W}} ||_{F}^{2}$ . + +A straightforward way to assess the SELI geometry structure of the target features is to compute the difference between logits $\mathbf{Z}^m$ extracted from $m$ -th pre-trained model and the optimal logits $\hat{\mathbf{Z}}$ . However, since the source and target data might not share similar classes, we cannot directly obtain the logits $\mathbf{Z}^m$ without the time-consuming fine-tuning process on the target dataset. Therefore, we extract features $\mathbf{H}^m$ of the model and measure its discrepancy to form the SELI structure. + +Specifically, we use the proven fact that, at the terminal stage of the fine-tuning where the model achieves the optimal cost $p^*$ , we have $||\hat{\mathbf{Z}}||_* = ||\hat{\mathbf{H}}||_F^2$ [33]. Therefore, we approximate the complexity of achieving the optimal logits $\hat{\mathbf{Z}}$ via features $\mathbf{H}^m$ extracted from the pre-trained model $\phi_m(\mathbf{X})$ as: + +$$ +S _ {s e l i} ^ {m} \left(\mathbf {H} ^ {m}\right) = \left\| \mathbf {H} ^ {m} \right\| _ {*}. \tag {6} +$$ + +Since the nuclear norm is the upper bound of $||\cdot ||_F^2$ and is the convex envelope for rank(), a higher $S_{seli}^{m}$ indicates a higher rank of the feature matrix $\mathbf{H}_m$ , which makes $\mathbf{Z}$ closer to a full rank matrix. + +# 3.3. Simplicity to Nearest Centroid Classifier + +Based on the neural collapse phenomenon shown in [25], at the terminal stage of training, the linear classifier behaves similarly to the nearest centroid classifier. Therefore, a score function $S_{ncc}^{m}$ is proposed to measure the applicability of the nearest centroid classifier on the $m$ -th pre-trained model. We start with defining the nearest centroid classifier as [28]: + +$$ +\hat {y} = \underset {c \in [ C ]} {\arg \max } \cos \left(\mu_ {c}, x\right), \tag {7} +$$ + +where $\hat{y}$ is the assigned class label to a sample $x$ , and $\cos$ indicates the cosine similarity. + +In the neural collapse state, the nearest centroid classifier can achieve the optimal classification performance on the training set. However, in transferability estimation, where there is no actual fine-tuning on the target dataset, directly comparing the performance of applying the nearest centroid classifier to the extracted feature with that of the optimal solution can lead to inaccurate estimation results. To mitigate this issue, we consider two relaxations to the nearest centroid classifier: Instead of using cosine similarity which only involves the mean of the classes, we adopt Mahalanobis distance as the nearest centroid classifier that further involves the correlation between the classes. Moreover, we leverage the expected conditional distribution of labels given input to replace the hard label. Under a mild assumption, the class distribution follows a multivariate Gaussian distribution, thus, we can obtain the following posterior by applying Bayes' Rules: + +$$ +\log P (y = c | h) = \frac {1}{2} \left(h _ {i} - \mu_ {c}\right) ^ {\top} \boldsymbol {\Sigma} \left(h _ {j} - \mu_ {c}\right) + \log P (y = c). \tag {8} +$$ + +Here, $\log P(y = c)$ denotes the prior probability of class $c$ , which can be computed as $N_{c} / N$ . $N_{c}$ represents the number of samples belonging to class $c$ . + +Compared with the discrete output of argmax function in the nearest centroid classifier formulation, we adopt the + +![](images/1283e8036759425ee21c30d93df920ceda565b6e6236a6ca1c8275a56be52653.jpg) +Figure 3: Strong correlation between the proposed NCTI score and model recognition performance (\%) on the target dataset after fine-tuning. Each marker represents a different pre-trained model. The target domains are (from left to right): Cifar100 [17], DTD [6], Food101 [2], Pets [26], SUN397 [39], and VOC2007 [7]. We show Spearman's coefficient $\rho$ and weighted Kendall's $\tau$ here. More correlation results are presented in supplementary materials. + +![](images/d0a56fa941d4a334ac75e2df4034ba70c796c3b2d61521cacd8d905689f242b5.jpg) + +![](images/a89d69ea03a4aeae14f7043b30c79ac2fb6d522802c069721492d2d041d3518f.jpg) + +![](images/93ec1d46fdd5000ce2501591110d3bdd0c40874fdbe47eaa04fd669c0c096e45.jpg) + +![](images/097e90b7c951a6ec888a373fa2f678b7fb9378960e4f3ee1881139a69d02a56a.jpg) + +![](images/63065a634a78e719a231439407626b2f3ba1c0993e9502d92bbc7593d2749f9d.jpg) + +softmax function as it preserves the relative order information of the prediction. Therefore, even if the corresponding class of a sample is not assigned with the highest probability, the probability value still reflects the distance of the sample to that class. The $c$ -th class normalized posterior probability of $i$ -th sample can be calculated as follows: + +$$ +z _ {i, c} ^ {m} = \frac {\exp \left(\log P \left(y = c \mid h _ {i} ^ {m}\right)\right)}{\sum_ {k = 1} ^ {C} \exp \left(\log P \left(y = k \mid h _ {i} ^ {m}\right)\right)}, \tag {9} +$$ + +As such, we have the score $S_{ncc}^{m}$ to describe the discrepancy between the current state of the $m$ -th pre-trained model $\phi_{m}(X)$ and the optimal nearest centroid classifier as follows: + +$$ +S _ {n c c} ^ {m} \left(\mathbf {H} ^ {m}\right) = \frac {1}{N} \sum_ {i = 1} ^ {N} z _ {i} ^ {m} y _ {i}, \tag {10} +$$ + +where $y_{i}$ denotes the corresponding one-hot ground truth label of $x_{i}$ and $z_{i}^{m}$ is the $C$ dimensional logits of $x_{i}$ , predicted by $m$ -th model. A greater $S_{ncc}^{m}$ indicates a smaller discrepancy towards an optimal nearest centroid classifier, which translates to greater transferability on the target dataset. + +# 3.4. Overall Score Function + +In this work, we model the transferability of a pre-trained model as how similar the current state of the network is to its terminal phase of fine-tuning. Specifically, we expect the terminal phase of fine-tuning to hold the so-called three properties of within-class variability, SELI geometry, and the nearest centroid classifier. Correspondingly, we design three score functions, which are $S_{vc}(\cdot)$ , $S_{seli}(\cdot)$ , and $S_{ncc}(\cdot)$ to measure the proximity of the state of the pre-trained network applied on the target data versus its optimal state at the terminal phase. However, since the values of score functions have different scales, summing up the scores may hurt the transferability estimation performance (e.g. $S_{seli}(\cdot)$ ) can numerically dominate over $S_{ncc}(\cdot)$ ). Instead of having manually defined hyper-parameters to balance the contribution of each score, we propose to normalize each score into a unit range (0 to 1) as follows: + +$$ +S _ {s e l i} ^ {m} \leftarrow \frac {S _ {s e l i} ^ {m} - \operatorname* {m i n} \left(S _ {s e l i}\right)}{\operatorname* {m a x} \left(S _ {s e l i}\right) - \operatorname* {m i n} \left(S _ {s e l i}\right)}. \tag {11} +$$ + +Similarly, we can have the normalized score $S_{vc}(H^m)$ and $S_{ncc}(H^m)$ for the $m$ -th pretrained model. By summing the normalized scores, we have the final transferability estimation metric: + +$$ +S _ {t o t a l} ^ {m} = S _ {v c} ^ {m} \left(\mathbf {H} ^ {m}\right) + S _ {s e l i} ^ {m} \left(\mathbf {H} ^ {m}\right) + S _ {n c c} ^ {m} \left(\mathbf {H} ^ {m}\right), \tag {12} +$$ + +where the $m$ -th pre-trained model with a higher overall score $S_{total}^{m}$ indicates a better transferability in the model pool for the target dataset $D$ . + +# 4. Experiment + +# 4.1. Datasets and Evaluation Metrics + +Datasets. We consider a wide range of classification benchmark datasets, including five fine-grained classifications (FGVC Aircraft [21], Standford Cars [16], Food101 [2], Oxford-IIIT Pets [26], Oxford-102 Flowers [23]), four coarse-grained classifications (Caltech101 [18], CIFAR-10 [17], CIFAR-100 [17], VOC2007 [7]), one scene classification (SUN397 [39]), and one texture classification (DTD [6]). In total, we adopt 11 benchmark datasets that are broadly used in transfer learning. A detailed description of each dataset can be found in the supplementary material. All models have undergone pre-training on the ImageNet dataset, and the fine-tuned performance of the model on target datasets is obtained from [29]. Without affecting the fine-tuning performance, we note that only the training and validation splits of the dataset are used for transferability estimation, and we hold the test split out of the estimation process. + +Correlation measurement. Following the previous work [41, 24, 29], we adopt weighted Kendall's $\tau$ [37] to measure the correlation between the estimated model ranking and the ground-truth performance ranking. Each pairwise comparison between items in two rankings is assigned a weight based on the distance between their ranks, whereas top-ranked items will be assigned with higher weight. + +Implementation details. In order to replicate the fine-tuning process, our proposed method aims to find a projection matrix $W$ that maximizes the discriminative power of the extracted features $\phi^m (X)$ from the $m$ -th pre-trained + +Table 1: Method comparison of their correlation strength with target accuracy for supervised models. Weighted Kendall's $\tau$ on 11 target test sets and their average are shown. For each column, the best, and second-best results are in bold, and underlined, respectively. Our method achieves the best overall weighted Kendall's $\tau$ . + +
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NCE [36]-0.1610.4650.6850.7090.7230.302-0.4820.6270.7720.7600.5710.452
LEEP [22]-0.2770.6050.3670.8240.6770.486-0.2910.4340.3890.6580.4130.390
LogME [41]0.4390.4630.6050.8520.7250.7000.1470.3850.4110.5110.6950.539
NLEEP [20]-0.5310.6140.4890.8250.7310.8200.0540.5290.9550.8480.6990.548
SFDA [29]0.6140.6960.5180.9490.8660.5750.5140.8150.5220.5580.6710.663
NCTI0.4960.6790.6470.8430.8790.7040.5410.7730.9240.7560.7410.726
+ +![](images/fcb74c476c13916e1472bb9e9e934ad04dbf15c25784f3a64d9b4fedc8479830.jpg) +(a) Supervised models +Figure 4: Method comparison w.r.t running time (seconds) vs. average Weighted Kendall's $\tau$ on 11 datasets on (a) supervised and (b) self-supervised models. Our method obtains the highest ranking correlation within a small time budget. + +![](images/0d2ef7393490b3ee5cd0506f046ff77b4a3f4a0179bdea8e6c8615c96e7d19c6.jpg) +(b) Self-supervised models + +model. To achieve this, we employ linear discriminant analysis (LDA) to maximize the separation between the class means of the projected data. However, LDA can be prone to overfitting when applied to high-dimensional feature spaces or when the covariance matrix is singular [40, 42]. To mitigate the overfitting issue on high-dimensional feature space, we follow the previous transferability estimation methods [20, 24] which incorporate principal component analysis (PCA) into the framework to project the original features onto a lower-dimensional subspace. We further address the singularity issue by shrinking the covariance matrix toward a full-rank identity matrix as follows: + +$$ +\boldsymbol {\Sigma} _ {w} ^ {\prime} = (1 - \alpha) \boldsymbol {\Sigma} _ {w} + \alpha \mathbf {I}, \text {w h e r e} \alpha = \exp \left(\sum_ {i} ^ {d ^ {*}} \lambda_ {i}\right), \tag {13} +$$ + +Here, $\lambda_{i}$ denotes the $i$ -th largest eigenvalue of the PCA projected feature matrix $PCA(\phi_m(X))$ . $d^{*}$ indicates the number of eigenvalues to be considered in the calculation. We set the dimension of the PCA projected subspace as 64 and set $d^{*}$ as 32 throughout all the experiments. $\alpha$ controls the trade-off between the covariance and the identity matrices. + +# 4.2. Evaluation on Supervised Models + +Overview. To initiate our evaluation of transferability estimation metrics for ranking supervised models, we assess the performance of these metrics on 11 widely adopted models, which are ResNet-34 [13], ResNet-50, ResNet-101, ResNet-151, DenseNet-121 [14], DenseNet-169, DenseNet-201, MNet-A1 [32], MobileNet-v2 [27], GoogleNet [30], and Inception-v3 [31]. + +Performance comparison. We compare our proposed NCTI with existing transferability estimation metrics, including NCE [36], LEEP [22], LogME [41], NLEEP [20], and SFDA [29]. The experimental results are shown in Tab. 1. From the table, we can observe that empirical conditional probability-based methods (i.e., NCE, LEEP, and LogME) generally achieve lower ranking correlations than our method and SFDA, indicating the importance of modeling the feature space in transferability estimation. NCTI achieves the best or the second best average weighted Kendall's ranking correlation on nine datasets and obtains a $6.9\%$ averaged performance gain compared to the second-best method SFDA. We argue that while the neural collapse explicitly reflects the maximum class separation, a feature space with a high class separation score can not necessarily leads to the neural collapse phenomenon. Therefore, our method can achieve superior performance than class separation-based methods. + +# 4.3. Evaluation on Self-supervised Models + +Overview. We further evaluate the effectiveness of NCTI in assessing the transferability of models pre-trained by self-supervised learning (SSL). We construct a pool with 10 SSL pre-trained models, including MoCo-v1 [11], MoCo-v2 [5], PCL-v2 [19], SELA-V2, Deepcluster-v2 [4], BYOL [10], Infomin [34], SWAV [3], and Insdis [38]. + +Performance comparison. Since NCE and LEEP require the classifier on the pre-training task, it is not directly applicable to self-supervised model ranking tasks. Therefore, to evaluate the ranking performance of the self-supervised models, we compare our method with NLEEP, LogME, and SFDA in Tab. 2. Similarly, we observe that the average ranking correlation of SFDA and the proposed + +Table 2: Method comparison of their correlation strength with target accuracy for self-supervised models. All the settings and notations are the same as Table 1. Our method achieves the best overall weighted Kendall's $\tau$ . + +
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NLEEP [20]-0.2860.6620.5950.1080.3740.7790.5980.7160.8640.880-0.0910.473
LogME [41]0.0210.0750.6270.4170.1460.7430.7630.6860.7380.2600.1810.423
SFDA [29]0.1670.6740.6830.8460.7890.8820.8970.8370.5640.8310.6210.708
NCTI0.0360.8110.7960.7580.8110.7960.7620.9450.8050.7740.6060.719
+ +![](images/dbb75c0f615981a7c90c3e7589c2fde51296ea98406e2a37924f3bbfd8fe9493.jpg) +(a) Supervised models +Figure 5: Effectiveness of each individual component in NCTI. We use the three terms individually or remove them one at a time from the full system. Results on (a) supervised and (b) self-supervised models are shown. + +![](images/49eeae66bb7b0b011d6710737bb95996babcd910b729a5f0fcc30a09356b4e4d.jpg) +(b) Self-supervised models + +method can achieve superior results than that of the conditional probability-based method. NCTI surpasses the second-best method by $1.5\%$ on average and scores the highest overall weighted Kendall's $\tau$ , which verifies the effectiveness of the proposed method. + +# 4.4. Further Analysis + +Tranferability estimation with fewer target data. We test a practical but challenging transferability estimation scenario, where only limited target data are available for the estimation. We randomly sample 2 to 500 images per class from the full SUN397 and examine how limited data affects the performance of different methods. As we can see from Fig. 6(d), although fewer data generally trigger performance degradation for transferability estimation methods, our method is able to achieve a larger performance gain under the low shot scheme than using the full dataset. Particularly, we highlight that our method can achieve SOTA performance on SUN397 with only 10 samples per class, which indicates a strong potential for practical applications. + +Computational efficiency comparison. We compare the time efficiency versus ranking correlation of different transferability estimation metrics in Fig. 4. The $x$ -axis is the total time cost for estimating transferability over 11 datasets, while the $y$ -axis indicates the average ranking correlation on 11 datasets. We adopt a logarithmic scale for + +the $x$ -axis. The figure shows that within a similar time budget, NCTI can achieve significantly higher ranking correlations compared with empirical conditional probability-based methods, such as NCE, LEEP, and LogME. SFDA can achieve a comparable ranking correlation with NCTI on self-supervised model ranking tasks. However, it consumes more than $10 \times$ more time to estimate the transferability, showing that it is less efficient than our proposed metric. Meanwhile, we can observe from the figure that our method scores the highest ranking correlation to the running time ratio under both settings, emphasizing the effectiveness and efficiency of our method. + +Ablation study. The NCTI metric proposed in this work is composed of three terms, each term is designed to describe a specific aspect of the gap between the current feature geometry and the geometry of neural collapse. In order to illustrate the impact of each individual term in NCTI, we conduct ablation studies to show how each individual term contributes to the overall metric. The ablation studies are conducted on both supervised and self-supervision scenarios, and the results are shown in Fig. 5. Among these three individual terms of NCTI, although $S_{vc}$ can attain a positive averaged ranking correlation, the contribution towards the full metric is less than the other two terms. $S_{seli}$ delivers the highest average ranking correlation on supervised model estimation, while $S_{ncc}$ attains the best-averaged performance on self-supervised model estimation. We infer this is due to the difference in the supervision signal. As the label information is not available for self-supervised training, the discriminative power of the pre-trained feature is more critical for the downstream task than that in supervised models. We also validate the effectiveness of two-term combinations. As we can see from the figure, the combination between $S_{seli}$ and $S_{vc}$ always achieves higher correlations than its individual component, confirming the terms are complementary. Although NCTI obtains a slightly lower correlation than w/o $S_{ncc}$ on the supervised models, it achieves an absolute $4.2\%$ higher averaged correlation of both tasks, which verifies the importance of $S_{ncc}$ . + +Hyperparameter analysis. NCTI simplifies the hyperparameter tuning process by assigning equal weights to each term. In order to validate this weight assignment, we investigate the impact of weight changes on the ranking cor + +![](images/80bb27f4f86a341e03df4a7e9768fc1ff224819faaf26ca56de8026ee2ebc9a3.jpg) +(a) Weight of $S_{vc}$ + +![](images/dbe66a0ed849e023f5ca5feb85f9d702c310dd6e9b917c454ffe19ac6d018d34.jpg) +(b) Weight of $S_{ncc}$ + +![](images/a8d0266025764c280a17f9ad46d81acd3eb7ffcf762637cac72c02878357b8ae.jpg) +(c) Weight of $S_{seli}$ + +![](images/4356a40cebfc48c525fa089e0e1e614bb3a89dd3f045698fc19ec339f5e77cfb.jpg) +(d) SUN397 + +![](images/5873dfbea9ec3ec17a8d902ed440ebb4aa36169161ef8adc2ecc9bde897fcb83.jpg) +Figure 6: (a)-(c) Hyperparameter analysis of three components in NCTI. We report correlation strength $\tau$ computed on 11 downstream datasets against weight values. Impact on both supervised and self-supervised pre-trained models are drawn. We study each weight by fixing the rest as 1. In this paper, by default, we set the weight to 1 for all components. (d) Influence of target data size for estimation on the correlation strength. We vary the data size from 2 to 500 samples per class and observe that our method can achieve a larger performance margin compared with existing methods under the low-shot scenario. +(a) Supervised models +Figure 7: Effectiveness of the three components computed on a subset of the pre-training dataset. Average correlation strength $(\tau)$ on 11 target datasets reported. We assess (a) supervised models, and (b) self-supervised models. + +![](images/544fe4283bc99cea3cff9a2b469723aca4e32906376c9785d9bc4ed6a65b7de1.jpg) +(b) Self-supervised models + +relation. Specifically, we vary the weight of the term of interest within the range of $\{0.1, 0.4, 0.8, 1.2, 1.6, 2\}$ while keeping the weight of the remaining terms fixed at 1. The experimental results of the hyperparameter analysis are reported in Fig.6 (a)-(c). Our findings indicate that NCTI achieves high averaged correlations ( $\tau \geq 0.7$ ) with the most weight configurations. While we observe different monotonicity between the correlations of supervised and self-supervised tasks, the results demonstrate that harmonic points (i.e., high mean and low variance) are consistently achieved when all weights are close to 1, thereby confirming the appropriateness of the equal weight setting. + +Analysis of NCTI computed using source features. The extensive experiments conducted so far have demonstrated the effectiveness of each individual term in the proposed NCTI metric for transferability estimation. All the previous experiments have been conducted by using the target features extracted from the pre-trained models. It remains important to investigate what the properties of neural collapse on the source (i.e., pre-training) features tell us + +about the transferability of a model to a downstream task. To this end, we measure each individual term in NCTI on a subset of the source dataset and report the ranking correlation of the supervised model estimation task in Fig.7(a) and (b). On average, the transferability of a supervised model to a downstream task is consistently positively related to $-S_{vc}$ and $S_{ncc}$ on the source dataset on both supervised and self-supervised, which scores 0.725 and 0.598 weighted Kendall $\tau$ , respectively. While the figure shows $S_{seli}$ can achieve a positive correlation on supervised tasks, no significant correlation on $S_{seli}$ is observed in the supervised transfer estimation task. Note that, $-S_{vc}$ achieves the highest ranking correlation on average, which shows that a pre-trained model with higher within-class variability can better transfer to a new task. This indicates that a fully unsupervised metric can be designed for transferability estimation which is solely based on measuring the class-wise compactness of the source features. + +# 5. Conclusion + +How to efficiently choose the optimal pre-trained model from a vast collection of models to use in a downstream task? To rank models based on their transferability, we propose a metric named Neural Collapse informed Transferability Index (NCTI). Our method models model transferability as the gap between the current feature geometry and the geometry at the terminal stage (i.e., neural collapse) after hypothetical fine-tuning on the downstream task. Specifically, we model the gap from three perspectives, including the formation of SELI geometry structure, the within-class variability, and the applicability of the nearest center classifier. We show that our method is light to compute and that the ranking of model transferability has a very strong correlation with the ground truth ranking and compares favorably against the state-of-the-art methods. + +Acknowledgement This work was partially supported by ARC DP 230101196 + +# References + +[1] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In 2019 IEEE international conference on image processing (ICIP), pages 2309-2313. IEEE, 2019. 1, 2 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 5 +[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6 +[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6 +[5] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5 +[7] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 5 +[8] Cong Fang, Hangfeng He, Qi Long, and Weijie J Su. Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training. Proceedings of the National Academy of Sciences, 118(43):e2103091118, 2021. 2 +[9] Florian Graf, Christoph Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised contrastive learning. In International Conference on Machine Learning, pages 3821-3830. PMLR, 2021. 2 +[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6 +[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6 +[12] Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imageNet pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4918-4927, 2019. 1 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6 +[15] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 2 +[16] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5 +[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[18] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, Apr 2022. 5 +[19] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966, 2020. 6 +[20] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2663-2673, 2021. 1, 2, 6, 7 +[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[22] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 1, 2, 6 +[23] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5 +[24] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9172-9182, 2022. 1, 2, 5, 6 +[25] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 1, 2, 4 +[26] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5 +[27] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6 +[28] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval, volume 39. Cambridge University Press Cambridge, 2008. 4 + +[29] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 286-302. Springer, 2022. 1, 2, 5, 6, 7 +[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6 +[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6 +[32] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 6 +[33] Christos Thrampoulidis, Ganesh R Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. arXiv preprint arXiv:2208.05512, 2022. 1, 2, 3, 4 +[34] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 6 +[35] Tom Tirer and Joan Bruna. Extended unconstrained features model for exploring deep neural collapse. In International Conference on Machine Learning, pages 21478-21505. PMLR, 2022. 2 +[36] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1395-1405, 2019. 1, 2, 6 +[37] Sebastiano Vigna. A weighted correlation index for rankings with ties. In Proceedings of the 24th international conference on World Wide Web, pages 1166-1176, 2015. 5 +[38] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 6 +[39] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5 +[40] Jian Yang and Jing-yu Yang. Why can lda be performed in pca transformed space? Pattern recognition, 36(2):563-566, 2003. 6 +[41] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Ma + +chine Learning, pages 12133-12143. PMLR, 2021. 1, 2, 5, 6, 7 +[42] Hua Yu and Jie Yang. A direct lda algorithm for high-dimensional data—with application to face recognition. Pattern recognition, 34(10):2067-2070, 2001. 6 +[43] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse undermse loss: Global optimality with unconstrained features. In International Conference on Machine Learning, pages 27179-27202. PMLR, 2022. 2 +[44] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. Advances in Neural Information Processing Systems, 34:29820-29834, 2021. 2, 4 \ No newline at end of file diff --git a/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/images.zip b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dcbc85ac750de5bcb4b9b677514e1888f11c3bc6 --- /dev/null +++ b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2cfb3b982f61a47a9be35a2727880dc3d26501ed3dbbe9d6b70c3ab9d931492 +size 478482 diff --git a/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/layout.json b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17c4bacaf24b5c126ba80bd7dbb1855a96b60784 --- /dev/null +++ b/2023/How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability/layout.json @@ -0,0 +1,10405 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "text", + "content": "How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "content": "Zijian Wang" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "content": " Yadan Luo" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "content": " Liang Zheng" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "content": " Zi Huang" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "content": " Mahsa Baktashmotlagh" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "content": "The University of Queensland " + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 493, + 190 + ], + "type": "text", + "content": "Australian National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 191, + 526, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 191, + 526, + 204 + ], + "spans": [ + { + "bbox": [ + 67, + 191, + 526, + 204 + ], + "type": "text", + "content": "{zijian.wang, y.luo, helen.huang, m.baktashmotlagh}@uq.edu.au, liang.zheng@anu.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 257, + 290, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 474 + ], + "type": "text", + "content": "This paper focuses on model transferability estimation, i.e., assessing the performance of pre-trained models on a downstream task without performing fine-tuning. Motivated by neural collapse (NC) [25] that reveals particular feature geometry at the terminal stage of training, we consider model transferability as how far the target activations obtained by pre-trained models are from their hypothetical state in the terminal phase of the model fine-tuned on the target domain. We propose a metric that measures this proximity based on three phenomena of NC: within-class variability collapse, simplex encoded label interpolation geometry structure is formed, and the nearest center classifier becomes optimal on training data. Through experiments on 11 datasets, we confirm none of the three NC proxies are dispensable, which allows us to obtain very competitive transferability estimation accuracy with approximately " + }, + { + "bbox": [ + 46, + 257, + 290, + 474 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 257, + 290, + 474 + ], + "type": "text", + "content": " wall-clock time speed up compared to state-of-the-art approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": "The past decade has witnessed a surge in the availability of off-the-shelf pre-trained models in public repositories. They are adapted and re-tuned to facilitate new tasks, which has become a standard practice [12]. Nevertheless, this approach raises an important practical question of how to effectively select the best pre-trained model from a large model pool to be carried onto a downstream task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "While precise ranking of these pre-trained models on a new task can be obtained by a greedy approach of fineturning each model and comparing the test accuracy, it is often prohibitive as it requires a vast amount of computing resources. Therefore, an early line of research has been developed referred to as transferability estimation [22, 41, 24, 29], where the aim is to design efficient methods for ranking the performances of pre-trained models on a downstream task without fine-tuning. Existing" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 231, + 533, + 379 + ], + "blocks": [ + { + "bbox": [ + 320, + 231, + 533, + 379 + ], + "lines": [ + { + "bbox": [ + 320, + 231, + 533, + 379 + ], + "spans": [ + { + "bbox": [ + 320, + 231, + 533, + 379 + ], + "type": "image", + "image_path": "a74933bc6341a105e9721e5c4ab6f3ebdadd1f7f32acdeb840c1f175888a8886.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 387, + 547, + 461 + ], + "lines": [ + { + "bbox": [ + 304, + 387, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 547, + 461 + ], + "type": "text", + "content": "Figure 1: Illustration of the transferability estimation problem. Given a pool of pre-trained models and labeled target data, the objective of transferability estimation is to predict which models can achieve higher performance on the target data after fine-tuning. This task is much more time efficient than fine-tuning the models to obtain ground truth rankings." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 605 + ], + "type": "text", + "content": "approaches in transferability estimation typically consider the current status of the pre-trained models (w.r.t the target task) and can be generally categorized into two streams of probabilistic-based [22, 36, 41, 1] and feature distribution-based techniques [24, 29]. Approaches from the former category model the expected conditional probability of target labels given the extracted target features [41, 20] or predictions [36, 22]. The latter category models transferability as the class-wise separation of features, in which Pandey et al. [24] employed the Bhattacharyya coefficient to approximate the overlap between class distributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "From a different viewpoint for transferability estimation, this paper measures the difference between the current status and the hypothetical terminal status of pre-trained models after fine-tuning. The latter status refers to neural collapse (NC) [25, 33] on the target domain, which commonly occurs when the training loss approaches zero. Specifically, there are three distinct characteristics of NC: 1) the within-class feature variability drops towards zero; 2) class means gradually converge to a simplex encoded label interpolation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5549" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "(SELI) geometry (def. in Sec.3); and 3) the behavior of the final classifier layer is similar to the nearest centroid classifier. Meanwhile, it is empirically demonstrated in [25] that stopping training at the terminal phase achieves higher test performance compared to the models that stop at an earlier stage (i.e., zero prediction error)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 287, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 406 + ], + "type": "text", + "content": "Motivated by these interesting findings, we formalize a transferability metric based on the observations of NC discussed above. The metric measures how close the geometry of the target features is to their hypothetical state in the terminal stage of the fine-tuned model. Specifically, we analyze the spectral property of the class-wise feature matrix to understand to what extent the current status is from zero within-class variation. Secondly, we assess the gap between the features extracted by the pre-trained model and SELI geometry with the rank of the feature matrix. Thirdly, we calculate the difference between the hypothetical classifier layer and the nearest neighbor classifier w.r.t. the empirical class-conditional distribution. Collectively, these three different measurements allow us to understand how far the pre-trained models are from NC on the target domain. On a series of standard benchmarks, we demonstrate that the metric exhibits a strong correlation with the transferability of supervised and self-supervised pre-trained models and thus is a very useful transferability estimator. Compared with the state-of-the-art, our method gives stronger correlations while being computationally efficient. We summarize our main contributions below." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 412, + 286, + 577 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 412, + 286, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 286, + 459 + ], + "type": "text", + "content": "- We propose the neural collapse transferability index (NCTI) for the transferability estimation. NCTI approximately measures the model w.r.t the three prominent characteristics of NC, which can be efficiently computed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 465, + 286, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 286, + 512 + ], + "type": "text", + "content": "- Experimental results on 11 datasets demonstrate strong correlations between NCTI and the transferability of models to the target dataset, which outperforms existing transferability estimation methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 517, + 286, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 286, + 577 + ], + "type": "text", + "content": "- We give interesting insights when the NC components are computed on a subset of the pre-training dataset. Our findings suggest a potential for an unsupervised transferability estimation metric, solely based on measuring within-class variability collapse of the source features." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 586, + 133, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 133, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 133, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Neural collapse. Initially observed empirically by Papyan et. al [25], the phenomenon of neural collapse has garnered significant research attention within a short time frame. Neural collapse shows that at the terminal phase of deep network training (i.e., driving the training loss towards zero), the last layer activations have the following four properties: 1) the variation of activation drop to negligible; 2) the class mean of activation converge to simplex Equiangular Tight Frame (ETF); 3) the weight of linear classifier" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "and class means converge to a dual-vector space; and 4) the final linear classifier behaves similar to a nearest class center classifier. Recent follow-up works focus on using the unconstrained feature model with cross-entropy [44, 9] or square loss [43, 35] training to theoretically analyze the NC phenomenon. [33, 8] extend the NC phenomenon from a class-balanced scenario to an imbalanced one. Particularly, Thrangpoulidis et. al [33] demonstrated a more generalized geometry than simplex ETF, named simplex encoded labels interpolation (SELI). SELI shows the invariant property on balanced datasets, as well as imbalanced datasets. The authors also proved that SELI recovers the ETF geometry for the balanced dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 395 + ], + "type": "text", + "content": "Transferability estimation. Early works [36, 1, 22] aim to assess the transferability of pre-trained models by calculating the expected conditional distribution of the target label space given the target prediction obtained from the pre-trained classifier. Specifically, [36] leverages negative conditional entropy to evaluate the amount of information shared by pre-training label space and the target task. [22] computes the empirical conditional distribution from the joint distribution of the target task and the pre-training task. Since the early methods heavily rely on the pre-trained predictor to obtain dummy source label distribution of target samples, they cannot directly evaluate the transferability of unsupervised or self-supervised models, where no classifier is available to predict dummy source label distributions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 611 + ], + "type": "text", + "content": "Recent transferability estimation methods broaden the applicability by nullifying the dependency on the classifier of the pre-training task. In order to achieve this, NLEEP [20] substitutes the output layer with a Gaussian mixture model. LogME [41] formalizes the transferability estimation as the maximum label marginalized likelihood and adopts a directed graphical model to solve it. Huang et. al [15] proposed to approximate transferability with the mutual information between the pre-trained model extracted features and the corresponding label. On par with the metrics the probabilistic point of view, [29, 24] take the class separability into consideration for the transferability estimation metric. Particularly, Shao et. al [29] incorporates the self-challenge noise augmentation to encourage the classifier to discriminate the hard samples. However, none of the existing works model the transferability as the gap between the current target feature geometry and the hypothetical target feature geometry at the terminal stage of the fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Transferability estimation is defined as ranking a set of " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " pre-trained models " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\phi^m (\\cdot)\\}_{m = 1}^M" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " given a labeled target dataset " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "D = \\{\\mathbf{X},\\mathbf{Y}\\} = \\{(x_i,y_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " being the number of samples. An evaluation metric (accuracy, mAP, etc.) is associated with the dataset, which measures the ground-truth performance " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "T^{m}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "-th pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5550" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "trained model " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\phi^m (\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " after fine-tuning. The feature extracted by the " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "-th pre-trained model " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\phi^m (\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^m" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^m = \\phi^m (\\mathbf{X})\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " is the feature dimension. With the objective of truly reflecting the ground truth model ranking, transferability estimation methods calculate a score " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "S^m" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " for each pre-trained model " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\phi^m (\\cdot)" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". Ideally, the calculated scores " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\{S^m\\}_{m = 1}^M" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " are supposed to be highly correlated with the ground truth performance " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\{T^m\\}_{m = 1}^M" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", and able to determine the best pre-trained model to be fine-tuned on the target data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 323 + ], + "type": "text", + "content": "The central idea of our proposed method is to employ neural collapse phenomena to assess the proximity of the current state of a pre-trained model " + }, + { + "bbox": [ + 46, + 192, + 289, + 323 + ], + "type": "inline_equation", + "content": "\\phi^m (\\cdot)" + }, + { + "bbox": [ + 46, + 192, + 289, + 323 + ], + "type": "text", + "content": " to its terminal stage of fine-tuning on the target. Particularly, the proximity is measured from three dimensions, which are: 1) distance of within-class variability of features from zero; 2) simplicity for the pre-trained model to produce simplex encoded labels interpolation (SELI) geometry structure [33]; and 3) the applicability of the nearest centroid classifier. The final metric jointly considers these three terms, with each term contributing equally to the overall transferability score." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 324, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 289, + 384 + ], + "type": "text", + "content": "The subsequent sections elaborate on each of these three perspectives in detail. Here, we start with two definitions that will be used in the sections: the simplex encoded label (SEL) matrix, and the corresponding simplex encoded label interpolation (SELI) geometry structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 396, + 296, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 296, + 408 + ], + "type": "text", + "content": "Definition 3.1 (Simplex Encoded Label (SEL) Matrix [33])" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 408, + 216, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 216, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 216, + 421 + ], + "type": "text", + "content": "The SEL matrix " + }, + { + "bbox": [ + 47, + 408, + 216, + 421 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}\\in \\mathbb{R}^{C\\times N}" + }, + { + "bbox": [ + 47, + 408, + 216, + 421 + ], + "type": "text", + "content": " is defined as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 430, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 430, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 101, + 430, + 287, + 462 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} [ c, i ] = \\left\\{ \\begin{array}{l l} 1 - 1 / C & , c = y _ {i} \\\\ - 1 / C & , c \\neq y _ {i} \\end{array} . \\right. \\tag {1}", + "image_path": "93bd66d9a082d053d8a8b43b547370c9fa79d0a29135491ec48daa7d64ad282b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "inline_equation", + "content": "c \\in [C]" + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "inline_equation", + "content": "i \\in [N]" + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": "-th class and " + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": "-th observation, respectively. " + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": " is the number of classes. Note that " + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^{\\top} \\mathbf{1} = 0" + }, + { + "bbox": [ + 47, + 472, + 287, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 520, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 287, + 556 + ], + "type": "text", + "content": "From the definition of SEL matrix " + }, + { + "bbox": [ + 47, + 520, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 47, + 520, + 287, + 556 + ], + "type": "text", + "content": ", we can observe that the matrix is almost a full rank matrix (" + }, + { + "bbox": [ + 47, + 520, + 287, + 556 + ], + "type": "inline_equation", + "content": "rank(\\hat{\\mathbf{Z}}) = \\min(C, N) - 1" + }, + { + "bbox": [ + 47, + 520, + 287, + 556 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 568, + 290, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 290, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 290, + 580 + ], + "type": "text", + "content": "Definition 3.2 (SEL Interpolation (SELI) Geometry [33])" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "text", + "content": "The embeddings " + }, + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{H} \\in \\mathbb{R}^{d \\times N}" + }, + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "text", + "content": " and classifier weight matrices " + }, + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d \\times C}" + }, + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "text", + "content": " follow the simplex-encoded-labels interpolation geometry when for some scaling " + }, + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "inline_equation", + "content": "\\alpha > 0" + }, + { + "bbox": [ + 47, + 580, + 289, + 617 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 638, + 293, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 293, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 293, + 663 + ], + "type": "interline_equation", + "content": "\\mathbf {W} ^ {\\top} \\mathbf {W} = \\alpha \\mathbf {V} \\boldsymbol {\\Lambda} \\mathbf {V} ^ {\\top}, \\mathbf {H} ^ {\\top} \\mathbf {H} = \\alpha \\mathbf {U} \\boldsymbol {\\Lambda} \\mathbf {U} ^ {\\top}, a n d \\mathbf {W} ^ {\\top} \\mathbf {H} = \\alpha \\hat {\\mathbf {Z}}, \\tag {2}", + "image_path": "7359e24cfb262da2b49fd72dce4ebe43cc6fccb8f6509a2c35f67fe13c6168b1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}} = \\mathbf{V}\\boldsymbol{\\Lambda}\\mathbf{U}^{\\top}" + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "text", + "content": " is the SEL matrix. " + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "text", + "content": " denote the left and right singular vector matrix of " + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Lambda}" + }, + { + "bbox": [ + 47, + 663, + 287, + 700 + ], + "type": "text", + "content": " represents the diagonal singular value matrix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "type": "text", + "content": "3.1. Within-class Variability Collapse" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "content": "As described in the neural collapse phenomenon, at the terminal stage of training, the variability of within-class features becomes negligible and the features collapse towards the corresponding class mean. Motivated by this, we develop a score function " + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "inline_equation", + "content": "S_{vc}" + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "content": " to track the within-class variability of features, which indicates how close a pretrained model is to the neural collapse state. Specifically, the within-class covariance " + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{\\Sigma}_c" + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "content": "-th class is denoted as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 198, + 545, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 198, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 356, + 198, + 545, + 223 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {c} = \\frac {1}{n _ {c}} \\left(\\mathbf {H} _ {\\mathbf {c}} - \\mu_ {c}\\right) ^ {\\top} \\left(\\mathbf {H} _ {\\mathbf {c}} - \\mu_ {c}\\right), \\tag {3}", + "image_path": "6341fe540e25f3ed2399563321367daeb4825e45360fc94aba3733cd9a7ba9bd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{c} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathbf{c}} \\in \\mathbb{R}^{N_{c} \\times d}" + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": " consists " + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": " features that belong to " + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": "-th class. " + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mu_{c}" + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": " denotes the mean of " + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": "-th class. By applying singular value decomposition of " + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "inline_equation", + "content": "(\\mathbf{H}_{\\mathbf{c}} - \\mu_{c}) = \\mathbf{U} \\pmb{\\Lambda} \\mathbf{V}^{T}" + }, + { + "bbox": [ + 305, + 228, + 545, + 277 + ], + "type": "text", + "content": ", the covariance matrix can be written as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 287, + 545, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 287, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 356, + 287, + 545, + 301 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {c} = \\mathbf {V} \\boldsymbol {\\Lambda} \\mathbf {U} ^ {\\top} \\mathbf {U} \\boldsymbol {\\Lambda} \\mathbf {V} ^ {\\top} = \\mathbf {V} \\boldsymbol {\\Lambda} ^ {2} \\mathbf {V}. \\tag {4}", + "image_path": "77dace7a10b1f7d2d0e73b9a261b2d17f08e399a0a01632374d8e52a73fd4097.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "text", + "content": " represent the left and right singular vector matrices. " + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "text", + "content": " denotes the diagonal singular value matrix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "text", + "content": "From Eq. 4, we observe that larger singular values mean higher within-class variability. Since SVD is computationally prohibitive on large matrices, it motivates us to use the nuclear norm, which calculates the sum of singular values of a matrix with lower time complexity. Meanwhile, as feature space is usually with high dimensionality, there might exist noises that are easily mitigated by fine-tuning but negatively affects variability calculation. Therefore, we compute the feature variability by substituting " + }, + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c^m" + }, + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "text", + "content": " into class-wise logits " + }, + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_c^m" + }, + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "text", + "content": ". We define the within-class variability score of " + }, + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "inline_equation", + "content": "S_{vc}" + }, + { + "bbox": [ + 304, + 337, + 545, + 469 + ], + "type": "text", + "content": " using the nuclear norm as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 368, + 475, + 545, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 475, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 368, + 475, + 545, + 508 + ], + "type": "interline_equation", + "content": "S _ {v c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = - \\sum_ {c = 1} ^ {C} \\left\\| \\mathbf {Z} _ {c} ^ {m} \\right\\| _ {*}, \\tag {5}", + "image_path": "1ae16e3f44f3b10a0fa751094db952bbd4da9dd9a69ed8fd505d36c5ad924c72.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_c^m" + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": " denotes the logits of " + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": "-th class extracted by the " + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": "-th model. We defer the formulation of " + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "inline_equation", + "content": "z_{i,c}" + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": " to section 3.3. Since " + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "inline_equation", + "content": "S_{vc}^{m}" + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": " directly reflects the within-class variability, when a pre-trained model " + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "inline_equation", + "content": "\\phi^m (\\cdot)" + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": " is close to the terminal stage of the fine-tuning, a larger " + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "inline_equation", + "content": "S_{vc}^{m}" + }, + { + "bbox": [ + 305, + 517, + 547, + 578 + ], + "type": "text", + "content": " should be achieved." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 586, + 543, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 543, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 543, + 599 + ], + "type": "text", + "content": "3.2. Simplex Encoded Labels Interpolation (SELI)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Thrampoulidis et al. [33] proposed SELI as a generalized geometry structure version of the simplex equiangular tight frame (ETF) that is defined in the neural collapse phenomenon. While the simplex ETF geometry observed in neural collapse applies only to balanced datasets, SELI can be established in both balanced and imbalanced datasets at the neural collapse state. As such, we propose a score function " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{seli}^{m}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to assess the SELI geometry structure of the target features extracted from the " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "-th pre-training model. In the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5551" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 165, + 199 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 165, + 199 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 165, + 199 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 165, + 199 + ], + "type": "image", + "image_path": "41fdd5cc049dd0f628a6eb83df8f3d250b69fb077de4312f2c48f9977bbe1b73.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "type": "text", + "content": "Figure 2: TSNE visualization of feature distributions produced by four pre-trained models with CIFAR-10 as the downstream task. Different colors mean different classes in CIFAR-10. From left to right, target recognition accuracy (or transferability) decreases. At the same time, the proposed three scores, i.e., " + }, + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "type": "inline_equation", + "content": "S_{vc}" + }, + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "type": "inline_equation", + "content": "S_{seli}" + }, + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "type": "inline_equation", + "content": "S_{ncc}" + }, + { + "bbox": [ + 46, + 208, + 547, + 246 + ], + "type": "text", + "content": " also decrease." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 70, + 281, + 199 + ], + "blocks": [ + { + "bbox": [ + 187, + 70, + 281, + 199 + ], + "lines": [ + { + "bbox": [ + 187, + 70, + 281, + 199 + ], + "spans": [ + { + "bbox": [ + 187, + 70, + 281, + 199 + ], + "type": "image", + "image_path": "202c64155dd48ef7ee69021c461344ad4c3c59d7fba56520ecfa8e7831fa2f38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 408, + 199 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 408, + 199 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 408, + 199 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 408, + 199 + ], + "type": "image", + "image_path": "9b8f812051387f72a518d1aacafa7a49ce9e510b1a874894affea420750677e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 426, + 70, + 526, + 199 + ], + "blocks": [ + { + "bbox": [ + 426, + 70, + 526, + 199 + ], + "lines": [ + { + "bbox": [ + 426, + 70, + 526, + 199 + ], + "spans": [ + { + "bbox": [ + 426, + 70, + 526, + 199 + ], + "type": "image", + "image_path": "3a15a5fa761e78720e5abf6e43bf17c5f9f796af9306efde1e1a9c7d947c87c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 290 + ], + "type": "text", + "content": "rest of this section, we first review the role of SELI geometry in cross-entropy minimization, followed by discussing our SELI-based metric." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "content": "Under the unconstrained feature (UF) model [44], Thrampoulidis et al. shows that minimizing the cross-entropy loss converges to a KKT point of the SVM classifier, where the global minimizer of this optimization problem is the SEL matrix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 357, + 288, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 357, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 45, + 357, + 288, + 393 + ], + "type": "text", + "content": "Theorem 1 (Structure of the UF-SVM minimizers) [33] Let " + }, + { + "bbox": [ + 45, + 357, + 288, + 393 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{W}},\\hat{\\mathbf{H}})" + }, + { + "bbox": [ + 45, + 357, + 288, + 393 + ], + "type": "text", + "content": " be any solution and " + }, + { + "bbox": [ + 45, + 357, + 288, + 393 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 45, + 357, + 288, + 393 + ], + "type": "text", + "content": " be the optimal cost of the following unconstrained feature SVM (UF-SVM) classifier:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 396, + 255, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 396, + 255, + 418 + ], + "spans": [ + { + "bbox": [ + 80, + 396, + 255, + 418 + ], + "type": "interline_equation", + "content": "(\\hat {\\mathbf {W}}, \\hat {\\mathbf {H}}) \\in \\underset {\\mathbf {W}, \\mathbf {H}} {\\arg \\min} | | \\mathbf {W} | | _ {F} ^ {2} / 2 + | | \\mathbf {H} | | _ {F} ^ {2} / 2,", + "image_path": "4b0412db3d78f7cfadc019d6ed3372633b54c0dc194a99b1143b710dd87b1f59.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 80, + 420, + 244, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 420, + 244, + 436 + ], + "spans": [ + { + "bbox": [ + 80, + 420, + 244, + 436 + ], + "type": "interline_equation", + "content": "s. t. \\left(w _ {y _ {i}} - w _ {c}\\right) ^ {T} h _ {i} \\geq 1, i \\in [ n ], c \\neq y _ {i}.", + "image_path": "8dfb7d213b5155f5fd23c2d9421d9f971016e819836876d69524777efaf382b8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 441, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 467 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 46, + 441, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}} = \\hat{\\mathbf{W}}^{\\top}\\hat{\\mathbf{H}}" + }, + { + "bbox": [ + 46, + 441, + 287, + 467 + ], + "type": "text", + "content": " holds for optimal logits, the optimal cost can be obtained as " + }, + { + "bbox": [ + 46, + 441, + 287, + 467 + ], + "type": "inline_equation", + "content": "p^{*} = ||\\hat{\\mathbf{Z}} ||_{*} = ||\\hat{\\mathbf{H}} ||_{F}^{2} = ||\\hat{\\mathbf{W}} ||_{F}^{2}" + }, + { + "bbox": [ + 46, + 441, + 287, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": "A straightforward way to assess the SELI geometry structure of the target features is to compute the difference between logits " + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^m" + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": "-th pre-trained model and the optimal logits " + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": ". However, since the source and target data might not share similar classes, we cannot directly obtain the logits " + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^m" + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": " without the time-consuming fine-tuning process on the target dataset. Therefore, we extract features " + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^m" + }, + { + "bbox": [ + 46, + 471, + 287, + 578 + ], + "type": "text", + "content": " of the model and measure its discrepancy to form the SELI structure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": "Specifically, we use the proven fact that, at the terminal stage of the fine-tuning where the model achieves the optimal cost " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "||\\hat{\\mathbf{Z}}||_* = ||\\hat{\\mathbf{H}}||_F^2" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " [33]. Therefore, we approximate the complexity of achieving the optimal logits " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " via features " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^m" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " extracted from the pre-trained model " + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\phi_m(\\mathbf{X})" + }, + { + "bbox": [ + 46, + 578, + 287, + 650 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 119, + 650, + 287, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 650, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 119, + 650, + 287, + 663 + ], + "type": "interline_equation", + "content": "S _ {s e l i} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = \\left\\| \\mathbf {H} ^ {m} \\right\\| _ {*}. \\tag {6}", + "image_path": "33bc89ae5de4b2e100648edbe627a35aae5da263eab3dd0cc41316fa3f53be2e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": "Since the nuclear norm is the upper bound of " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "||\\cdot ||_F^2" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": " and is the convex envelope for rank(), a higher " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "S_{seli}^{m}" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": " indicates a higher rank of the feature matrix " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_m" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": ", which makes " + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": " closer to a full rank matrix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 255, + 518, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 518, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 518, + 268 + ], + "type": "text", + "content": "3.3. Simplicity to Nearest Centroid Classifier" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": "Based on the neural collapse phenomenon shown in [25], at the terminal stage of training, the linear classifier behaves similarly to the nearest centroid classifier. Therefore, a score function " + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "inline_equation", + "content": "S_{ncc}^{m}" + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": " is proposed to measure the applicability of the nearest centroid classifier on the " + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": "-th pre-trained model. We start with defining the nearest centroid classifier as [28]:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 357, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 357, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 375, + 357, + 545, + 376 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\underset {c \\in [ C ]} {\\arg \\max } \\cos \\left(\\mu_ {c}, x\\right), \\tag {7}", + "image_path": "a0e1aef5618866e880add12d227453771355c10057679fb85622840e248f098d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "text", + "content": " is the assigned class label to a sample " + }, + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 304, + 383, + 545, + 407 + ], + "type": "text", + "content": " indicates the cosine similarity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 407, + 546, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 546, + 611 + ], + "type": "text", + "content": "In the neural collapse state, the nearest centroid classifier can achieve the optimal classification performance on the training set. However, in transferability estimation, where there is no actual fine-tuning on the target dataset, directly comparing the performance of applying the nearest centroid classifier to the extracted feature with that of the optimal solution can lead to inaccurate estimation results. To mitigate this issue, we consider two relaxations to the nearest centroid classifier: Instead of using cosine similarity which only involves the mean of the classes, we adopt Mahalanobis distance as the nearest centroid classifier that further involves the correlation between the classes. Moreover, we leverage the expected conditional distribution of labels given input to replace the hard label. Under a mild assumption, the class distribution follows a multivariate Gaussian distribution, thus, we can obtain the following posterior by applying Bayes' Rules:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 620, + 545, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 545, + 653 + ], + "type": "interline_equation", + "content": "\\log P (y = c | h) = \\frac {1}{2} \\left(h _ {i} - \\mu_ {c}\\right) ^ {\\top} \\boldsymbol {\\Sigma} \\left(h _ {j} - \\mu_ {c}\\right) + \\log P (y = c). \\tag {8}", + "image_path": "4acac6c2c76745814ae961daaafd9c57c374705c37edbda6010d0954cc819409.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\log P(y = c)" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " denotes the prior probability of class " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": ", which can be computed as " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "N_{c} / N" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " represents the number of samples belonging to class " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Compared with the discrete output of argmax function in the nearest centroid classifier formulation, we adopt the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5552" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 140, + 148 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 140, + 148 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 140, + 148 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 140, + 148 + ], + "type": "image", + "image_path": "1283e8036759425ee21c30d93df920ceda565b6e6236a6ca1c8275a56be52653.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 159, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 547, + 208 + ], + "type": "text", + "content": "Figure 3: Strong correlation between the proposed NCTI score and model recognition performance (\\%) on the target dataset after fine-tuning. Each marker represents a different pre-trained model. The target domains are (from left to right): Cifar100 [17], DTD [6], Food101 [2], Pets [26], SUN397 [39], and VOC2007 [7]. We show Spearman's coefficient " + }, + { + "bbox": [ + 46, + 159, + 547, + 208 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 159, + 547, + 208 + ], + "type": "text", + "content": " and weighted Kendall's " + }, + { + "bbox": [ + 46, + 159, + 547, + 208 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 159, + 547, + 208 + ], + "type": "text", + "content": " here. More correlation results are presented in supplementary materials." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 70, + 222, + 149 + ], + "blocks": [ + { + "bbox": [ + 141, + 70, + 222, + 149 + ], + "lines": [ + { + "bbox": [ + 141, + 70, + 222, + 149 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 222, + 149 + ], + "type": "image", + "image_path": "d0a56fa941d4a334ac75e2df4034ba70c796c3b2d61521cacd8d905689f242b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 222, + 70, + 301, + 149 + ], + "blocks": [ + { + "bbox": [ + 222, + 70, + 301, + 149 + ], + "lines": [ + { + "bbox": [ + 222, + 70, + 301, + 149 + ], + "spans": [ + { + "bbox": [ + 222, + 70, + 301, + 149 + ], + "type": "image", + "image_path": "a89d69ea03a4aeae14f7043b30c79ac2fb6d522802c069721492d2d041d3518f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 381, + 149 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 381, + 149 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 381, + 149 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 381, + 149 + ], + "type": "image", + "image_path": "93ec1d46fdd5000ce2501591110d3bdd0c40874fdbe47eaa04fd669c0c096e45.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 462, + 149 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 462, + 149 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 462, + 149 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 462, + 149 + ], + "type": "image", + "image_path": "097e90b7c951a6ec888a373fa2f678b7fb9378960e4f3ee1881139a69d02a56a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 462, + 70, + 542, + 149 + ], + "blocks": [ + { + "bbox": [ + 462, + 70, + 542, + 149 + ], + "lines": [ + { + "bbox": [ + 462, + 70, + 542, + 149 + ], + "spans": [ + { + "bbox": [ + 462, + 70, + 542, + 149 + ], + "type": "image", + "image_path": "63065a634a78e719a231439407626b2f3ba1c0993e9502d92bbc7593d2749f9d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 219, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 291 + ], + "type": "text", + "content": "softmax function as it preserves the relative order information of the prediction. Therefore, even if the corresponding class of a sample is not assigned with the highest probability, the probability value still reflects the distance of the sample to that class. The " + }, + { + "bbox": [ + 46, + 219, + 287, + 291 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 219, + 287, + 291 + ], + "type": "text", + "content": "-th class normalized posterior probability of " + }, + { + "bbox": [ + 46, + 219, + 287, + 291 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 219, + 287, + 291 + ], + "type": "text", + "content": "-th sample can be calculated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 296, + 287, + 325 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 296, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 89, + 296, + 287, + 325 + ], + "type": "interline_equation", + "content": "z _ {i, c} ^ {m} = \\frac {\\exp \\left(\\log P \\left(y = c \\mid h _ {i} ^ {m}\\right)\\right)}{\\sum_ {k = 1} ^ {C} \\exp \\left(\\log P \\left(y = k \\mid h _ {i} ^ {m}\\right)\\right)}, \\tag {9}", + "image_path": "69b777a36840f124c9a2557e7b3b9045e9d1e94093765bbbd3627e4246e03f8b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "text", + "content": "As such, we have the score " + }, + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "inline_equation", + "content": "S_{ncc}^{m}" + }, + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "text", + "content": " to describe the discrepancy between the current state of the " + }, + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "text", + "content": "-th pre-trained model " + }, + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "inline_equation", + "content": "\\phi_{m}(X)" + }, + { + "bbox": [ + 46, + 329, + 287, + 375 + ], + "type": "text", + "content": " and the optimal nearest centroid classifier as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 375, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 375, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 110, + 375, + 287, + 407 + ], + "type": "interline_equation", + "content": "S _ {n c c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} z _ {i} ^ {m} y _ {i}, \\tag {10}", + "image_path": "e6b59cf8bd1da3e514b34293d9168651cd1529f8c14982119e8838f2ccb40405.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": " denotes the corresponding one-hot ground truth label of " + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "inline_equation", + "content": "z_{i}^{m}" + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": " dimensional logits of " + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": ", predicted by " + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": "-th model. A greater " + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "inline_equation", + "content": "S_{ncc}^{m}" + }, + { + "bbox": [ + 46, + 409, + 287, + 469 + ], + "type": "text", + "content": " indicates a smaller discrepancy towards an optimal nearest centroid classifier, which translates to greater transferability on the target dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 475, + 178, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 178, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 178, + 487 + ], + "type": "text", + "content": "3.4. Overall Score Function" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "text", + "content": "In this work, we model the transferability of a pre-trained model as how similar the current state of the network is to its terminal phase of fine-tuning. Specifically, we expect the terminal phase of fine-tuning to hold the so-called three properties of within-class variability, SELI geometry, and the nearest centroid classifier. Correspondingly, we design three score functions, which are " + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "inline_equation", + "content": "S_{vc}(\\cdot)" + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "inline_equation", + "content": "S_{seli}(\\cdot)" + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "inline_equation", + "content": "S_{ncc}(\\cdot)" + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "text", + "content": " to measure the proximity of the state of the pre-trained network applied on the target data versus its optimal state at the terminal phase. However, since the values of score functions have different scales, summing up the scores may hurt the transferability estimation performance (e.g. " + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "inline_equation", + "content": "S_{seli}(\\cdot)" + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "text", + "content": ") can numerically dominate over " + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "inline_equation", + "content": "S_{ncc}(\\cdot)" + }, + { + "bbox": [ + 46, + 493, + 287, + 685 + ], + "type": "text", + "content": "). Instead of having manually defined hyper-parameters to balance the contribution of each score, we propose to normalize each score into a unit range (0 to 1) as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 94, + 689, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 689, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 94, + 689, + 287, + 717 + ], + "type": "interline_equation", + "content": "S _ {s e l i} ^ {m} \\leftarrow \\frac {S _ {s e l i} ^ {m} - \\operatorname* {m i n} \\left(S _ {s e l i}\\right)}{\\operatorname* {m a x} \\left(S _ {s e l i}\\right) - \\operatorname* {m i n} \\left(S _ {s e l i}\\right)}. \\tag {11}", + "image_path": "d9d721fd09bf738fb993c441f2c445ead494732386d106beca5a04e54dacdbcc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "text", + "content": "Similarly, we can have the normalized score " + }, + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "inline_equation", + "content": "S_{vc}(H^m)" + }, + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "inline_equation", + "content": "S_{ncc}(H^m)" + }, + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 219, + 545, + 266 + ], + "type": "text", + "content": "-th pretrained model. By summing the normalized scores, we have the final transferability estimation metric:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 274, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 274, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 321, + 274, + 545, + 289 + ], + "type": "interline_equation", + "content": "S _ {t o t a l} ^ {m} = S _ {v c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) + S _ {s e l i} ^ {m} \\left(\\mathbf {H} ^ {m}\\right) + S _ {n c c} ^ {m} \\left(\\mathbf {H} ^ {m}\\right), \\tag {12}", + "image_path": "d089781e0c03d853b8cc08b3fab481aec0a66620bb116fb5da7de06ac6c71089.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "text", + "content": "-th pre-trained model with a higher overall score " + }, + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "inline_equation", + "content": "S_{total}^{m}" + }, + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "text", + "content": " indicates a better transferability in the model pool for the target dataset " + }, + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 294, + 545, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 341, + 382, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 382, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 382, + 354 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 360, + 481, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 481, + 372 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 378, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 581 + ], + "type": "text", + "content": "Datasets. We consider a wide range of classification benchmark datasets, including five fine-grained classifications (FGVC Aircraft [21], Standford Cars [16], Food101 [2], Oxford-IIIT Pets [26], Oxford-102 Flowers [23]), four coarse-grained classifications (Caltech101 [18], CIFAR-10 [17], CIFAR-100 [17], VOC2007 [7]), one scene classification (SUN397 [39]), and one texture classification (DTD [6]). In total, we adopt 11 benchmark datasets that are broadly used in transfer learning. A detailed description of each dataset can be found in the supplementary material. All models have undergone pre-training on the ImageNet dataset, and the fine-tuned performance of the model on target datasets is obtained from [29]. Without affecting the fine-tuning performance, we note that only the training and validation splits of the dataset are used for transferability estimation, and we hold the test split out of the estimation process." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": "Correlation measurement. Following the previous work [41, 24, 29], we adopt weighted Kendall's " + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": " [37] to measure the correlation between the estimated model ranking and the ground-truth performance ranking. Each pairwise comparison between items in two rankings is assigned a weight based on the distance between their ranks, whereas top-ranked items will be assigned with higher weight." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. In order to replicate the fine-tuning process, our proposed method aims to find a projection matrix " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " that maximizes the discriminative power of the extracted features " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi^m (X)" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "-th pre-trained" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5553" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 116, + 526, + 196 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 1: Method comparison of their correlation strength with target accuracy for supervised models. Weighted Kendall's " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": " on 11 target test sets and their average are shown. For each column, the best, and second-best results are in bold, and underlined, respectively. Our method achieves the best overall weighted Kendall's " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 116, + 526, + 196 + ], + "lines": [ + { + "bbox": [ + 69, + 116, + 526, + 196 + ], + "spans": [ + { + "bbox": [ + 69, + 116, + 526, + 196 + ], + "type": "table", + "html": "
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NCE [36]-0.1610.4650.6850.7090.7230.302-0.4820.6270.7720.7600.5710.452
LEEP [22]-0.2770.6050.3670.8240.6770.486-0.2910.4340.3890.6580.4130.390
LogME [41]0.4390.4630.6050.8520.7250.7000.1470.3850.4110.5110.6950.539
NLEEP [20]-0.5310.6140.4890.8250.7310.8200.0540.5290.9550.8480.6990.548
SFDA [29]0.6140.6960.5180.9490.8660.5750.5140.8150.5220.5580.6710.663
NCTI0.4960.6790.6470.8430.8790.7040.5410.7730.9240.7560.7410.726
", + "image_path": "29a7a3e72824c653ade8ab69fe2b41460949a20dbd348324ca4493398130b31e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 214, + 167, + 326 + ], + "blocks": [ + { + "bbox": [ + 50, + 214, + 167, + 326 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 167, + 326 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 167, + 326 + ], + "type": "image", + "image_path": "fcb74c476c13916e1472bb9e9e934ad04dbf15c25784f3a64d9b4fedc8479830.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 331, + 148, + 342 + ], + "lines": [ + { + "bbox": [ + 66, + 331, + 148, + 342 + ], + "spans": [ + { + "bbox": [ + 66, + 331, + 148, + 342 + ], + "type": "text", + "content": "(a) Supervised models" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 352, + 289, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 413 + ], + "type": "text", + "content": "Figure 4: Method comparison w.r.t running time (seconds) vs. average Weighted Kendall's " + }, + { + "bbox": [ + 46, + 352, + 289, + 413 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 352, + 289, + 413 + ], + "type": "text", + "content": " on 11 datasets on (a) supervised and (b) self-supervised models. Our method obtains the highest ranking correlation within a small time budget." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 214, + 286, + 325 + ], + "blocks": [ + { + "bbox": [ + 170, + 214, + 286, + 325 + ], + "lines": [ + { + "bbox": [ + 170, + 214, + 286, + 325 + ], + "spans": [ + { + "bbox": [ + 170, + 214, + 286, + 325 + ], + "type": "image", + "image_path": "0d2ef7393490b3ee5cd0506f046ff77b4a3f4a0179bdea8e6c8615c96e7d19c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 331, + 277, + 342 + ], + "lines": [ + { + "bbox": [ + 178, + 331, + 277, + 342 + ], + "spans": [ + { + "bbox": [ + 178, + 331, + 277, + 342 + ], + "type": "text", + "content": "(b) Self-supervised models" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 432, + 289, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 289, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 289, + 577 + ], + "type": "text", + "content": "model. To achieve this, we employ linear discriminant analysis (LDA) to maximize the separation between the class means of the projected data. However, LDA can be prone to overfitting when applied to high-dimensional feature spaces or when the covariance matrix is singular [40, 42]. To mitigate the overfitting issue on high-dimensional feature space, we follow the previous transferability estimation methods [20, 24] which incorporate principal component analysis (PCA) into the framework to project the original features onto a lower-dimensional subspace. We further address the singularity issue by shrinking the covariance matrix toward a full-rank identity matrix as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 594, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 594, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 58, + 594, + 287, + 625 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {w} ^ {\\prime} = (1 - \\alpha) \\boldsymbol {\\Sigma} _ {w} + \\alpha \\mathbf {I}, \\text {w h e r e} \\alpha = \\exp \\left(\\sum_ {i} ^ {d ^ {*}} \\lambda_ {i}\\right), \\tag {13}", + "image_path": "9655fa49d0a1558cab0883bcb1498c3b40736a6637a6df468795e6f8f2bf4b71.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": "-th largest eigenvalue of the PCA projected feature matrix " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "PCA(\\phi_m(X))" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "d^{*}" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": " indicates the number of eigenvalues to be considered in the calculation. We set the dimension of the PCA projected subspace as 64 and set " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "d^{*}" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": " as 32 throughout all the experiments. " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": " controls the trade-off between the covariance and the identity matrices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 215, + 486, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 215, + 486, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 215, + 486, + 228 + ], + "type": "text", + "content": "4.2. Evaluation on Supervised Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 317 + ], + "type": "text", + "content": "Overview. To initiate our evaluation of transferability estimation metrics for ranking supervised models, we assess the performance of these metrics on 11 widely adopted models, which are ResNet-34 [13], ResNet-50, ResNet-101, ResNet-151, DenseNet-121 [14], DenseNet-169, DenseNet-201, MNet-A1 [32], MobileNet-v2 [27], GoogleNet [30], and Inception-v3 [31]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 317, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 546, + 533 + ], + "type": "text", + "content": "Performance comparison. We compare our proposed NCTI with existing transferability estimation metrics, including NCE [36], LEEP [22], LogME [41], NLEEP [20], and SFDA [29]. The experimental results are shown in Tab. 1. From the table, we can observe that empirical conditional probability-based methods (i.e., NCE, LEEP, and LogME) generally achieve lower ranking correlations than our method and SFDA, indicating the importance of modeling the feature space in transferability estimation. NCTI achieves the best or the second best average weighted Kendall's ranking correlation on nine datasets and obtains a " + }, + { + "bbox": [ + 304, + 317, + 546, + 533 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 304, + 317, + 546, + 533 + ], + "type": "text", + "content": " averaged performance gain compared to the second-best method SFDA. We argue that while the neural collapse explicitly reflects the maximum class separation, a feature space with a high class separation score can not necessarily leads to the neural collapse phenomenon. Therefore, our method can achieve superior performance than class separation-based methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 540, + 504, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 504, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 504, + 553 + ], + "type": "text", + "content": "4.3. Evaluation on Self-supervised Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": "Overview. We further evaluate the effectiveness of NCTI in assessing the transferability of models pre-trained by self-supervised learning (SSL). We construct a pool with 10 SSL pre-trained models, including MoCo-v1 [11], MoCo-v2 [5], PCL-v2 [19], SELA-V2, Deepcluster-v2 [4], BYOL [10], Infomin [34], SWAV [3], and Insdis [38]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "content": "Performance comparison. Since NCE and LEEP require the classifier on the pre-training task, it is not directly applicable to self-supervised model ranking tasks. Therefore, to evaluate the ranking performance of the self-supervised models, we compare our method with NLEEP, LogME, and SFDA in Tab. 2. Similarly, we observe that the average ranking correlation of SFDA and the proposed" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5554" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 103, + 526, + 163 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 2: Method comparison of their correlation strength with target accuracy for self-supervised models. All the settings and notations are the same as Table 1. Our method achieves the best overall weighted Kendall's " + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 103, + 526, + 163 + ], + "lines": [ + { + "bbox": [ + 69, + 103, + 526, + 163 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 526, + 163 + ], + "type": "table", + "html": "
AircraftCaltechCarsCifar10Cifar100DTDFlowersFoodPetsSUNVOCAverage
NLEEP [20]-0.2860.6620.5950.1080.3740.7790.5980.7160.8640.880-0.0910.473
LogME [41]0.0210.0750.6270.4170.1460.7430.7630.6860.7380.2600.1810.423
SFDA [29]0.1670.6740.6830.8460.7890.8820.8970.8370.5640.8310.6210.708
NCTI0.0360.8110.7960.7580.8110.7960.7620.9450.8050.7740.6060.719
", + "image_path": "190573cd9d692a6d5917227d1c0b7658085e6a2ed7b8483c216718d08f667d92.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 191, + 167, + 307 + ], + "blocks": [ + { + "bbox": [ + 51, + 191, + 167, + 307 + ], + "lines": [ + { + "bbox": [ + 51, + 191, + 167, + 307 + ], + "spans": [ + { + "bbox": [ + 51, + 191, + 167, + 307 + ], + "type": "image", + "image_path": "dbb75c0f615981a7c90c3e7589c2fde51296ea98406e2a37924f3bbfd8fe9493.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 313, + 148, + 324 + ], + "lines": [ + { + "bbox": [ + 66, + 313, + 148, + 324 + ], + "spans": [ + { + "bbox": [ + 66, + 313, + 148, + 324 + ], + "type": "text", + "content": "(a) Supervised models" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 334, + 288, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 381 + ], + "type": "text", + "content": "Figure 5: Effectiveness of each individual component in NCTI. We use the three terms individually or remove them one at a time from the full system. Results on (a) supervised and (b) self-supervised models are shown." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 191, + 284, + 306 + ], + "blocks": [ + { + "bbox": [ + 171, + 191, + 284, + 306 + ], + "lines": [ + { + "bbox": [ + 171, + 191, + 284, + 306 + ], + "spans": [ + { + "bbox": [ + 171, + 191, + 284, + 306 + ], + "type": "image", + "image_path": "49eeae66bb7b0b011d6710737bb95996babcd910b729a5f0fcc30a09356b4e4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 313, + 277, + 324 + ], + "lines": [ + { + "bbox": [ + 178, + 313, + 277, + 324 + ], + "spans": [ + { + "bbox": [ + 178, + 313, + 277, + 324 + ], + "type": "text", + "content": "(b) Self-supervised models" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": "method can achieve superior results than that of the conditional probability-based method. NCTI surpasses the second-best method by " + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": " on average and scores the highest overall weighted Kendall's " + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": ", which verifies the effectiveness of the proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 466, + 149, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 149, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 149, + 479 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 485, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 640 + ], + "type": "text", + "content": "Tranferability estimation with fewer target data. We test a practical but challenging transferability estimation scenario, where only limited target data are available for the estimation. We randomly sample 2 to 500 images per class from the full SUN397 and examine how limited data affects the performance of different methods. As we can see from Fig. 6(d), although fewer data generally trigger performance degradation for transferability estimation methods, our method is able to achieve a larger performance gain under the low shot scheme than using the full dataset. Particularly, we highlight that our method can achieve SOTA performance on SUN397 with only 10 samples per class, which indicates a strong potential for practical applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Computational efficiency comparison. We compare the time efficiency versus ranking correlation of different transferability estimation metrics in Fig. 4. The " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "-axis is the total time cost for estimating transferability over 11 datasets, while the " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "-axis indicates the average ranking correlation on 11 datasets. We adopt a logarithmic scale for" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "text", + "content": "-axis. The figure shows that within a similar time budget, NCTI can achieve significantly higher ranking correlations compared with empirical conditional probability-based methods, such as NCE, LEEP, and LogME. SFDA can achieve a comparable ranking correlation with NCTI on self-supervised model ranking tasks. However, it consumes more than " + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 304, + 192, + 545, + 337 + ], + "type": "text", + "content": " more time to estimate the transferability, showing that it is less efficient than our proposed metric. Meanwhile, we can observe from the figure that our method scores the highest ranking correlation to the running time ratio under both settings, emphasizing the effectiveness and efficiency of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": "Ablation study. The NCTI metric proposed in this work is composed of three terms, each term is designed to describe a specific aspect of the gap between the current feature geometry and the geometry of neural collapse. In order to illustrate the impact of each individual term in NCTI, we conduct ablation studies to show how each individual term contributes to the overall metric. The ablation studies are conducted on both supervised and self-supervision scenarios, and the results are shown in Fig. 5. Among these three individual terms of NCTI, although " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "S_{vc}" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": " can attain a positive averaged ranking correlation, the contribution towards the full metric is less than the other two terms. " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "S_{seli}" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": " delivers the highest average ranking correlation on supervised model estimation, while " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "S_{ncc}" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": " attains the best-averaged performance on self-supervised model estimation. We infer this is due to the difference in the supervision signal. As the label information is not available for self-supervised training, the discriminative power of the pre-trained feature is more critical for the downstream task than that in supervised models. We also validate the effectiveness of two-term combinations. As we can see from the figure, the combination between " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "S_{seli}" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "S_{vc}" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": " always achieves higher correlations than its individual component, confirming the terms are complementary. Although NCTI obtains a slightly lower correlation than w/o " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "S_{ncc}" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": " on the supervised models, it achieves an absolute " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": " higher averaged correlation of both tasks, which verifies the importance of " + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "inline_equation", + "content": "S_{ncc}" + }, + { + "bbox": [ + 304, + 339, + 546, + 662 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Hyperparameter analysis. NCTI simplifies the hyperparameter tuning process by assigning equal weights to each term. In order to validate this weight assignment, we investigate the impact of weight changes on the ranking cor" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5555" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 164, + 171 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 164, + 171 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 164, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 164, + 171 + ], + "type": "image", + "image_path": "80bb27f4f86a341e03df4a7e9768fc1ff224819faaf26ca56de8026ee2ebc9a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 175, + 143, + 186 + ], + "lines": [ + { + "bbox": [ + 78, + 175, + 143, + 186 + ], + "spans": [ + { + "bbox": [ + 78, + 175, + 143, + 186 + ], + "type": "text", + "content": "(a) Weight of " + }, + { + "bbox": [ + 78, + 175, + 143, + 186 + ], + "type": "inline_equation", + "content": "S_{vc}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 70, + 289, + 171 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 289, + 171 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 289, + 171 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 289, + 171 + ], + "type": "image", + "image_path": "dbe66a0ed849e023f5ca5feb85f9d702c310dd6e9b917c454ffe19ac6d018d34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 175, + 269, + 187 + ], + "lines": [ + { + "bbox": [ + 200, + 175, + 269, + 187 + ], + "spans": [ + { + "bbox": [ + 200, + 175, + 269, + 187 + ], + "type": "text", + "content": "(b) Weight of " + }, + { + "bbox": [ + 200, + 175, + 269, + 187 + ], + "type": "inline_equation", + "content": "S_{ncc}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 411, + 171 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 411, + 171 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 411, + 171 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 411, + 171 + ], + "type": "image", + "image_path": "a8d0266025764c280a17f9ad46d81acd3eb7ffcf762637cac72c02878357b8ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 175, + 393, + 187 + ], + "lines": [ + { + "bbox": [ + 324, + 175, + 393, + 187 + ], + "spans": [ + { + "bbox": [ + 324, + 175, + 393, + 187 + ], + "type": "text", + "content": "(c) Weight of " + }, + { + "bbox": [ + 324, + 175, + 393, + 187 + ], + "type": "inline_equation", + "content": "S_{seli}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 71, + 533, + 171 + ], + "blocks": [ + { + "bbox": [ + 433, + 71, + 533, + 171 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 533, + 171 + ], + "type": "image", + "image_path": "4356a40cebfc48c525fa089e0e1e614bb3a89dd3f045698fc19ec339f5e77cfb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 175, + 506, + 186 + ], + "lines": [ + { + "bbox": [ + 460, + 175, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 460, + 175, + 506, + 186 + ], + "type": "text", + "content": "(d) SUN397" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 266, + 167, + 380 + ], + "blocks": [ + { + "bbox": [ + 46, + 197, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 547, + 258 + ], + "type": "text", + "content": "Figure 6: (a)-(c) Hyperparameter analysis of three components in NCTI. We report correlation strength " + }, + { + "bbox": [ + 46, + 197, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 197, + 547, + 258 + ], + "type": "text", + "content": " computed on 11 downstream datasets against weight values. Impact on both supervised and self-supervised pre-trained models are drawn. We study each weight by fixing the rest as 1. In this paper, by default, we set the weight to 1 for all components. (d) Influence of target data size for estimation on the correlation strength. We vary the data size from 2 to 500 samples per class and observe that our method can achieve a larger performance margin compared with existing methods under the low-shot scenario." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 266, + 167, + 380 + ], + "lines": [ + { + "bbox": [ + 50, + 266, + 167, + 380 + ], + "spans": [ + { + "bbox": [ + 50, + 266, + 167, + 380 + ], + "type": "image", + "image_path": "5873dfbea9ec3ec17a8d902ed440ebb4aa36169161ef8adc2ecc9bde897fcb83.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 385, + 148, + 396 + ], + "lines": [ + { + "bbox": [ + 66, + 385, + 148, + 396 + ], + "spans": [ + { + "bbox": [ + 66, + 385, + 148, + 396 + ], + "type": "text", + "content": "(a) Supervised models" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 406, + 287, + 455 + ], + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 455 + ], + "type": "text", + "content": "Figure 7: Effectiveness of the three components computed on a subset of the pre-training dataset. Average correlation strength " + }, + { + "bbox": [ + 46, + 406, + 287, + 455 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 46, + 406, + 287, + 455 + ], + "type": "text", + "content": " on 11 target datasets reported. We assess (a) supervised models, and (b) self-supervised models." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 171, + 267, + 286, + 380 + ], + "blocks": [ + { + "bbox": [ + 171, + 267, + 286, + 380 + ], + "lines": [ + { + "bbox": [ + 171, + 267, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 171, + 267, + 286, + 380 + ], + "type": "image", + "image_path": "544fe4283bc99cea3cff9a2b469723aca4e32906376c9785d9bc4ed6a65b7de1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 385, + 277, + 396 + ], + "lines": [ + { + "bbox": [ + 178, + 385, + 277, + 396 + ], + "spans": [ + { + "bbox": [ + 178, + 385, + 277, + 396 + ], + "type": "text", + "content": "(b) Self-supervised models" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 470, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 614 + ], + "type": "text", + "content": "relation. Specifically, we vary the weight of the term of interest within the range of " + }, + { + "bbox": [ + 46, + 470, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.4, 0.8, 1.2, 1.6, 2\\}" + }, + { + "bbox": [ + 46, + 470, + 287, + 614 + ], + "type": "text", + "content": " while keeping the weight of the remaining terms fixed at 1. The experimental results of the hyperparameter analysis are reported in Fig.6 (a)-(c). Our findings indicate that NCTI achieves high averaged correlations (" + }, + { + "bbox": [ + 46, + 470, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\tau \\geq 0.7" + }, + { + "bbox": [ + 46, + 470, + 287, + 614 + ], + "type": "text", + "content": ") with the most weight configurations. While we observe different monotonicity between the correlations of supervised and self-supervised tasks, the results demonstrate that harmonic points (i.e., high mean and low variance) are consistently achieved when all weights are close to 1, thereby confirming the appropriateness of the equal weight setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 618, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 714 + ], + "type": "text", + "content": "Analysis of NCTI computed using source features. The extensive experiments conducted so far have demonstrated the effectiveness of each individual term in the proposed NCTI metric for transferability estimation. All the previous experiments have been conducted by using the target features extracted from the pre-trained models. It remains important to investigate what the properties of neural collapse on the source (i.e., pre-training) features tell us" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "content": "about the transferability of a model to a downstream task. To this end, we measure each individual term in NCTI on a subset of the source dataset and report the ranking correlation of the supervised model estimation task in Fig.7(a) and (b). On average, the transferability of a supervised model to a downstream task is consistently positively related to " + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "inline_equation", + "content": "-S_{vc}" + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "inline_equation", + "content": "S_{ncc}" + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "content": " on the source dataset on both supervised and self-supervised, which scores 0.725 and 0.598 weighted Kendall " + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "content": ", respectively. While the figure shows " + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "inline_equation", + "content": "S_{seli}" + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "content": " can achieve a positive correlation on supervised tasks, no significant correlation on " + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "inline_equation", + "content": "S_{seli}" + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "content": " is observed in the supervised transfer estimation task. Note that, " + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "inline_equation", + "content": "-S_{vc}" + }, + { + "bbox": [ + 304, + 268, + 547, + 483 + ], + "type": "text", + "content": " achieves the highest ranking correlation on average, which shows that a pre-trained model with higher within-class variability can better transfer to a new task. This indicates that a fully unsupervised metric can be designed for transferability estimation which is solely based on measuring the class-wise compactness of the source features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 492, + 378, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 492, + 378, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 378, + 504 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 508, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 687 + ], + "type": "text", + "content": "How to efficiently choose the optimal pre-trained model from a vast collection of models to use in a downstream task? To rank models based on their transferability, we propose a metric named Neural Collapse informed Transferability Index (NCTI). Our method models model transferability as the gap between the current feature geometry and the geometry at the terminal stage (i.e., neural collapse) after hypothetical fine-tuning on the downstream task. Specifically, we model the gap from three perspectives, including the formation of SELI geometry structure, the within-class variability, and the applicability of the nearest center classifier. We show that our method is light to compute and that the ranking of model transferability has a very strong correlation with the ground truth ranking and compares favorably against the state-of-the-art methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement This work was partially supported by ARC DP 230101196" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5556" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In 2019 IEEE international conference on image processing (ICIP), pages 2309-2313. IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "text", + "content": "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[5] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 380 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "text", + "content": "[7] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "text", + "content": "[8] Cong Fang, Hangfeng He, Qi Long, and Weijie J Su. Exploring deep neural networks via layer-peeled model: Minority collapse in imbalanced training. Proceedings of the National Academy of Sciences, 118(43):e2103091118, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 523 + ], + "type": "text", + "content": "[9] Florian Graf, Christoph Hofer, Marc Niethammer, and Roland Kwitt. Dissecting supervised contrastive learning. In International Conference on Machine Learning, pages 3821-3830. PMLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 590 + ], + "type": "text", + "content": "[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[12] Kaiming He, Ross Girshick, and Piotr Dollar. Rethinking imageNet pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4918-4927, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[15] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 183, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 545, + 205 + ], + "type": "text", + "content": "[16] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 228 + ], + "type": "text", + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 250 + ], + "type": "text", + "content": "[18] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, Apr 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "type": "text", + "content": "[19] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv preprint arXiv:2005.04966, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "text", + "content": "[20] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2663-2673, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "type": "text", + "content": "[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "text", + "content": "[22] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "text", + "content": "[23] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[24] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9172-9182, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 580 + ], + "type": "text", + "content": "[25] Vardan Papyan, XY Han, and David L Donoho. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652-24663, 2020. 1, 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[26] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "text", + "content": "[27] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval, volume 39. Cambridge University Press Cambridge, 2008. 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5557" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "text", + "content": "[29] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 286-302. Springer, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 150, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 217 + ], + "type": "text", + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 218, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 218, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 288, + 272 + ], + "type": "text", + "content": "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 274, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 288, + 329 + ], + "type": "text", + "content": "[32] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 364 + ], + "type": "text", + "content": "[33] Christos Thrampoulidis, Ganesh R Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. arXiv preprint arXiv:2208.05512, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "text", + "content": "[34] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 453 + ], + "type": "text", + "content": "[35] Tom Tirer and Joan Bruna. Extended unconstrained features model for exploring deep neural collapse. In International Conference on Machine Learning, pages 21478-21505. PMLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "text", + "content": "[36] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1395-1405, 2019. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 533 + ], + "type": "text", + "content": "[37] Sebastiano Vigna. A weighted correlation index for rankings with ties. In Proceedings of the 24th international conference on World Wide Web, pages 1166-1176, 2015. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 586 + ], + "type": "text", + "content": "[38] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "text", + "content": "[39] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "text", + "content": "[40] Jian Yang and Jing-yu Yang. Why can lda be performed in pca transformed space? Pattern recognition, 36(2):563-566, 2003. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[41] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Ma" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 239 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "chine Learning, pages 12133-12143. PMLR, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[42] Hua Yu and Jie Yang. A direct lda algorithm for high-dimensional data—with application to face recognition. Pattern recognition, 34(10):2067-2070, 2001. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 130, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 184 + ], + "type": "text", + "content": "[43] Jinxin Zhou, Xiao Li, Tianyu Ding, Chong You, Qing Qu, and Zhihui Zhu. On the optimization landscape of neural collapse undermse loss: Global optimality with unconstrained features. In International Conference on Machine Learning, pages 27179-27202. PMLR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[44] Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A geometric analysis of neural collapse with unconstrained features. Advances in Neural Information Processing Systems, 34:29820-29834, 2021. 2, 4" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5558" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_content_list.json b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..22bbfd724194cbf9f2aad29f0ecd561d531af9d4 --- /dev/null +++ b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_content_list.json @@ -0,0 +1,1552 @@ +[ + { + "type": "text", + "text": "How Much Temporal Long-Term Context is Needed for Action Segmentation?", + "text_level": 1, + "bbox": [ + 91, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emad Bahrami", + "bbox": [ + 230, + 179, + 362, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gianpiero Francesca", + "bbox": [ + 401, + 180, + 575, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juergen Gall1,3", + "bbox": [ + 614, + 180, + 736, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Bonn, Germany", + "bbox": [ + 256, + 199, + 462, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Toyota Motor Europe, Belgium", + "bbox": [ + 495, + 199, + 712, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Lamarr Institute for Machine Learning and Artificial Intelligence, Germany", + "bbox": [ + 230, + 217, + 736, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modeling long-term context in videos is crucial for many fine-grained tasks including temporal action segmentation. An interesting question that is still open is how much long-term temporal context is needed for optimal performance. While transformers can model the long-term context of a video, this becomes computationally prohibitive for long videos. Recent works on temporal action segmentation thus combine temporal convolutional networks with self-attention that are computed only for a local temporal window. While these approaches show good results, their performance is limited by their inability to capture the full context of a video. In this work, we try to answer how much long-term temporal context is required for temporal action segmentation by introducing a transformer-based model that leverages sparse attention to capture the full context of a video. We compare our model with the current state of the art on three datasets for temporal action segmentation, namely 50Salads, Breakfast, and Assembly101. Our experiments show that modeling the full context of a video is necessary to obtain the best performance for temporal action segmentation.", + "bbox": [ + 73, + 300, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal action segmentation can be used in many real-world applications such as monitoring production lines or studying animal behavior. In these settings, the videos can be very long and it is required to recognize the start and end of all actions that occur in a video as illustrated in Fig. 1.", + "bbox": [ + 75, + 672, + 470, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, combinations of temporal convolutional networks [1, 27] with self- and cross-attention from transformers [61, 4] have shown impressive results for temporal action segmentation. These works are in line with other hybrid models [15, 9, 61, 58] that combine the attention modules with convolutions to compensate for the lack of strong inductive bias of pure transformers. However, the emergence of datasets like Assembly101 [41], where subjects perform assembly tasks, poses a new challenge in the area of temporal action segmentation due to the existence of long videos", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/89c6531ce91098468580331fe78d56c67f734046b049397c4eb0f2690eb7c0e4.jpg", + "image_caption": [ + "Figure 1: Datasets like Assembly101 contain long videos of assembly tasks and an action label needs to be predicted for each frame. The first row shows some frames of a video. The second row shows the ground-truth labels for all frames of the video where different colors correspond to different action labels. Rows 3-6 show the predictions of the proposed model for different amounts of long-term context where $100\\%$ means the temporal context of the full video." + ], + "image_footnote": [], + "bbox": [ + 501, + 268, + 893, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that can last up to 25 minutes. Since modeling the long-term context of such a video is very expensive, Yi et al. [61] proposed to compute the attention for a local temporal window. In order to understand the impact of the window size on the temporal action segmentation accuracy, we analyze the impact of the window size on two datasets with long videos in Section 4.1. Fig. 1 shows some qualitative results of this study on Assembly101. Indeed, the results show that modeling the long-term context of an entire video is very important for temporal action segmentation.", + "bbox": [ + 496, + 547, + 893, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on this finding, we revisit how temporal attention is modeled in transformer architectures for temporal action segmentation. [61, 4] use a hierarchy of temporal windows, which makes the training on long video sequences as they occur in Assembly101 [41] very expensive. Inspired by works that decompose attention over the spatial and temporal domain for short video clips [6, 62], we propose to iterate between computing windowed local attention and sparse long-term context attention such that both short and long-term context are modeled. This approach is particularly suitable for temporal action segmentation since the local attentions focus on the similarity or dissimilarity of features within an action segment or between neighboring action", + "bbox": [ + 496, + 703, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10351", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segments whereas the long-term context attention focuses on the relations between actions within the entire video. The source code is available at https://github.com/LTContext/LTContext.", + "bbox": [ + 76, + 90, + 470, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 165, + 218, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The traditional sliding window with non-maxima suppression has been among the early approaches for action segmentation [19, 40]. [22, 20, 46] adopted hidden Markov Models (HMMs) for temporal modeling. [39] incorporates length and temporal context and uses dynamic programming for inference. Temporal convolutional networks (TCN) with temporal pooling are used by [23] to classify each video frame. Later, [1] introduced a multi-stage TCN capable of maintaining high temporal resolution, which is necessary for fine-grained recognition. [18] reduces the over-segmentation error by adding an action boundary regression network to refine the frame-wise results. [17] proposed another refinement method based on a graph convolutional network.", + "bbox": [ + 76, + 191, + 470, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers [52] were originally used for natural language processing and only rely on attention for sequence modeling. Recently, transformers have been widely adopted in vision [11, 57, 31, 54], speech recognition [15, 37], and action recognition [6, 2]. The original transformers suffered from two issues. First, the cost of the self-attention operation is quadratic with respect to the sequence length. This has been addressed by many methods that improve the memory efficiency of transformers [5, 34, 10]. We refer to [47] for a comprehensive survey of efficient transformers. For instance, restricting the attention to a fixed window size [5] is one approach. Multi-axis self-attention [51] combines block-attention with grid-attention, which is based on a spatial grid overlaid on the entire 2D space. Adapting sparse attention for capturing global information has shown to be an effective solution in vision tasks such as high-resolution image generation [63, 50], object detection [51, 60], and instance segmentation [51, 60]. MViT [12] uses a pooling operation to reduce the space-time resolution before attention and MViTv2 [28] improves MViT by adding the relative position along (x,y,t) and residual connections. The second issue of transformers is the poor generalization due to a relatively weak inductive bias [9, 49] compared to convolutional neural networks (CNN). Hybrid models that combine self-attention and convolution layers have thus been proposed for vision tasks [58, 9, 51, 57] and speech recognition [15, 37].", + "bbox": [ + 76, + 388, + 470, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In action recognition, works such as [6, 62] adopt the idea of using attention for video understanding. In Times-Former [6], the self-attention is first applied along the temporal dimension for each single patch (time attention), i.e., over all frames of a short video clip of a few seconds. In a second step, the attention is applied over the patches of each frame (space attention). VIDTR [62] also decomposes self-attention into spatial and temporal attention, but additionally", + "bbox": [ + 76, + 780, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "down-samples the temporal dimension. In our work, we do not apply attention spatially and temporally, but we address the question of how temporal attention can be computed for very long sequences that last 25 minutes as it is required for temporal action segmentation.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For temporal action segmentation, [61] proposed an architecture, called ASFormer, which is based on a multi-stage TCN [1] and equips the temporal convolutions with a local window attention [5]. This is done in a hierarchy where the size of the local window grows with each layer. Recently, [4] proposed a decoder on top of the encoder that generates action segments in an autoregressive manner. It uses two heads where the head on top of the encoder predicts framewise probabilities and the head of the decoder predicts the sequence of actions. Finally, an alignment decoder fuses the output of the two heads and aligns the predicted sequence of actions to the frames. [3] introduced a hybrid Temporal Convolution Transformer (TCTr) where they adapt an action boundary detector to adaptively estimate attention from local neighboring frames. [59] proposed to use additional constraints during training, but the approach assumes that action sequences can be modelled by a directed acyclic graph, which does not allow that actions occur more than once in a video. Recently, multi-modal approaches that combine language models with vision transformers have been proposed. For instance, [26] uses prompt engineering to extract features from pre-trained vision-language models such as ActionCLIP [55].", + "bbox": [ + 496, + 167, + 895, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Long-Term Context for Action Segmentation", + "text_level": 1, + "bbox": [ + 498, + 527, + 890, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, transformers combined with temporal convolutional neural networks [61, 4, 3] have shown a very good performance for temporal action segmentation. For this task, the frame-wise labels $c_{1}, \\ldots, c_{T}$ , where $c_{t} \\in \\mathcal{C}$ and $C = |\\mathcal{C}|$ denotes the number of action classes, need to be predicted for a given a video $X = (x_{1}, \\ldots, x_{T})$ with $T$ frames, where $x_{t}$ represents a feature map of size $D$ at frame $t$ . Since $T$ can be very large, [61, 4, 3] limit the self-attention to a local temporal window.", + "bbox": [ + 496, + 553, + 893, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to understand how much temporal long-term context is needed for temporal action segmentation, we limited the temporal input window and evaluated the quality of the temporal action segmentation in Section 4.1. The results in Fig. 4 show that temporal long-term context has a strong impact on the performance. Based on our analysis, we thus revisit the windowed attention of previous works for temporal action segmentation and propose to model the temporal long-term context of a video using sparse attentions [5, 38, 36]. Additionally, we equip our method with windowed attention to capture the locality between neighboring frames. In this way, we obtain a flexible design that is capable of providing long-term and local temporal context. While we describe first the Long-Term Context (LTCon", + "bbox": [ + 496, + 689, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10352", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/995082b14912e24ab0b5a9300955eea2b4b6ece7b0df240175baeb26000878e7.jpg", + "image_caption": [ + "Figure 2: Illustration of windowed and long-term context attentions with a window of size 2. For the windowed attention, the sequence is partitioned into small windows and the attentions are computed for each window. For long-term context attention, the sequence is reordered such that the attentions are computed over the whole, but sparsely sampled sequence. After the attention, the output is reordered again to preserve the original order. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 86, + 92, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text) block in Section 3.1, the entire network is described in Section 3.2.", + "bbox": [ + 75, + 400, + 470, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal Context Attention", + "text_level": 1, + "bbox": [ + 76, + 440, + 331, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The self-attention blocks of transformers are advantageous over convolutions in the aggregation of global information. However, applying attention to long sequences such as untrimmed videos is impractical due to the quadratic complexity of the self-attention blocks. To address this issue, we adopt an attention mechanism where we leverage sparse and windowed attention to model long-term and local temporal context.", + "bbox": [ + 75, + 464, + 470, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attention function transforms the input into a query, key, and value and computes the output as a weighted sum of the values. For example, given a sequence of features, $X\\in \\mathbb{R}^{T\\times D}$ , the attention can be written as follows:", + "bbox": [ + 75, + 584, + 470, + 645 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA t t e n t i o n (Q, K, V) = \\mathrm {s o f t m a x} _ {r o w} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right) V \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 654, + 472, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q, K, V \\in \\mathbb{R}^{T \\times D}$ are linearly transformed from $X$ . Since $T$ is very large for long sequences, we need to modify $Q, K, V$ to enable the modeling of long-term and local temporal context as illustrated in Fig. 2.", + "bbox": [ + 75, + 699, + 470, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Windowed Attention For the temporal Windowed Attention (WA), we partition the sequence into nonoverlapping windows of size $W$ . Fig. 2 illustrates the case for $W = 2$ , but we use $W = 64$ in practice. The impact of $W$ is evaluated in Section 4. Instead of computing the attention over the entire sequence of length $T$ , we compute the attention $\\frac{T}{W}$ times, where each query $Q \\in \\mathbb{R}^{W \\times D}$ corresponds to each window. For the keys and values, we use", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an overlap where we use the next window in addition, i.e., for each query $Q$ we have $K,V\\in \\mathbb{R}^{2W\\times D}$ . We perform masking when $K$ and $V$ exceed the input sequence. We evaluate the impact of the overlap in Section 4.", + "bbox": [ + 496, + 400, + 893, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Long-Term Context Attention For the temporal Long-Term Context (LTC) attention, the input is also partitioned into non-overlapping windows of size $G$ . However, instead of computing the attention over each window, the attention is computed over all windows where from each window only one element is taken. In the illustration of Fig. 2 with $G = 2$ , we compute the attention over the first feature of all windows and the attention over the second feature of all windows. In general, we compute the attentions for $G$ queries $Q \\in \\mathbb{R}^{\\frac{T}{G} \\times D}$ where the keys and values are the same, i.e., $K, V \\in \\mathbb{R}^{\\frac{T}{G} \\times D}$ . The parameter $G$ provides the flexibility to adjust the sparseness based on the available memory budget, e.g., $G = 1$ corresponds to the case where the attention is applied over the full sequence. In practice, we use $G = 64$ and we evaluate the impact of $G$ in Section 4.", + "bbox": [ + 496, + 484, + 895, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LTContext block The top of Fig. 3 illustrates the entire LTContext block. As in previous works [1, 61, 4], we use a 1D dilated temporal convolution with kernel size 3, where the dilation factor increases by factor 2 for each layer. The dilated temporal convolution is followed by a Gaussian Error Linear Unit (GELU). In the LTContext block, we first use the windowed and then the long-term context attention, which are shown in Fig. 2. We evaluate the impact of the order in Section 4. Finally, we use a linear layer with a residual connection to output the features for each frame, $F \\in \\mathbb{R}^{T \\times D}$ .", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10353", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/687ec45d91111480e73d139494f70f7219ea3b1c910f95828a75487ac011f750.jpg", + "image_caption": [ + "Figure 3: The network architecture of LTContext with LTContext blocks (top)." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 890, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. LTContext Architecture", + "text_level": 1, + "bbox": [ + 76, + 333, + 299, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The entire LTContext network is depicted in Fig. 3. For a fair comparison, we will use the features that are provided for the corresponding datasets as input. In all cases, the dimensionality of the features is 2048. As in previous works, we use a linear layer to reduce the dimensionality of the features to 64. The output of each LTContext block is the feature map $F \\in \\mathbb{R}^{T \\times D}$ . We repeat each LTContext block $N$ times where the dilation factor of the temporal convolution increases in each layer. In practice, we use $N = 9$ and we evaluate the impact of $N$ in Section 4. After the first $N$ layers of LTContext blocks, we use an additional linear layer to reduce the dimensionality $D$ further to 32. The dimensionality reduction reduces the number of parameters from 1.42 million to 0.72 million without reducing the accuracy. We also use an additional linear layer followed by a softmax layer to generate the frame-wise class probabilities $P \\in \\mathbb{R}^{T \\times C}$ .", + "bbox": [ + 75, + 364, + 473, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We continue with three additional stages where each stage consists of $N$ layers of LTContext blocks. Note that we reset the dilation factor to 1 for the temporal convolution at the beginning of each stage and we compute the frame-wise class probabilities $P \\in \\mathbb{R}^{T \\times C}$ after each stage, which contributes to the multi-stage loss. We use the cross-entropy loss combined with the mean squared error smoothing loss as introduced by [1] and used in [27, 61] for a fair comparison. Inspired by [61], we use the cross-attention for the LTContext blocks in stages 2 to 4. Instead of using the features $F$ for the queries and keys for windowed and long-term context attention, the predictions $P$ are used. We thus have $Q \\in \\mathbb{R}^{W \\times C}$ , $K \\in \\mathbb{R}^{2W \\times C}$ , and $V \\in \\mathbb{R}^{2W \\times D}$ for the windowed attention and $Q \\in \\mathbb{R}_{G}^{T \\times C}$ , $K \\in \\mathbb{R}_{G}^{T \\times C}$ , and $V \\in \\mathbb{R}_{G}^{\\frac{T}{G} \\times D}$ for the long-term context attention. While the cross-attention is not shown in Fig. 3, it only means that $P$ is an additional input for the windowed and long-term context attention in stages 2-4. We evaluate the impact of the number", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of stages in Section 4.", + "bbox": [ + 498, + 335, + 650, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 362, + 633, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We evaluate the performance of our proposed model on three challenging action segmentation datasets: 50Salads [45], Breakfast [21], and Assembly101 [41].", + "bbox": [ + 498, + 387, + 893, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "50Salads [45] contains 50 videos annotated with 17 action classes. On average, each video is 6.4 minutes long and has 18 action segments. Following previous works [1, 61, 4], we use five-fold cross-validation and report the average.", + "bbox": [ + 496, + 431, + 893, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Breakfast contains 1,712 videos of breakfast preparation activities with an average length of 2.3 minutes. There are 48 action classes and each video has on average 6.6 action segments. For evaluation, we report the average of the 4 splits for cross-validation as in [1].", + "bbox": [ + 496, + 492, + 893, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assembly101 [41] is the largest dataset among the three datasets with 4,321 videos and 202 coarse action classes composed of 11 verbs and 61 objects. Assembly101 is a procedural activity dataset containing videos of people assembling and disassembling 101 toy vehicles. On average, each video includes 24 action segments and is 7.1 minutes long. Compared to Breakfast, Assembly101 has 2.5 times more videos, 6.7 times more hours of video footage, 9.3 times more action segments, and 4.2 more action classes. For our evaluation, we follow the setting for temporal action segmentation [41] and report the results on the validation set.", + "bbox": [ + 496, + 568, + 895, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a fair comparison, we use the features that are provided for the datasets and that have been used in previous works. We use the I3D [7] features for the 50Salads and Breakfast datasets and TSM [29] features for the Assembly101 dataset [41]. Both features are 2048 dimensional vectors. Following [1], we also used the temporally downsampled features for 50Salads to compensate for the different frame-rates of the datasets.", + "bbox": [ + 496, + 750, + 895, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation Metrics. We report the frame-wise accuracy (Acc), segmental Edit distance, and segmental F1 score at", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10354", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the overlapping thresholds of $10\\%$ , $25\\%$ , and $50\\%$ denoted by $F1@{10,25,50}$ . The intersection over the union (IoU) ratio is used as the overlapping threshold. The edit distance measures only the order of the actions but not the duration. The frame-wise accuracy measures the accuracy per frame. It is dominated by actions that have a long duration and it is not very sensitive to over-segmentation errors that occur when a few frames within a ground-truth action segment are wrongly classified. The F1 score is the most reliable measure.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. How much temporal long-term context is needed?", + "text_level": 1, + "bbox": [ + 76, + 252, + 470, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first present the results of our analysis on the impact of using the full sequence as input compared to using a temporal window. The goal of this experiment is to shed light on how much temporal long-term context is needed for the task of temporal action segmentation. For the analysis, we use only 50Salads and Assembly101 since the videos in Breakfast are too short. For the experiments, we train our approach (LTContext) and ASFormer [61] either on the full video sequences or we divide the videos into shorter sequences where the full context is lacking.", + "bbox": [ + 75, + 291, + 470, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 shows the result of this experiment on the 50Salads and Assembly101 dataset. We report the window size as percentage of the average length of a video in the corresponding dataset and $100\\%$ means that the full sequence has been used. The results clearly show that the full context of an entire sequence is advantageous over allowing the model to see only a window of the input sequence even if the window is large $(50\\%)$ . We can also observe that our approach benefits more from the full sequence than ASFormer.", + "bbox": [ + 75, + 443, + 470, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We furthermore evaluated whether the impact of the window size is stronger for longer videos. To this end, we sorted all test videos into four quarters by their length. Fig. 5 (left) shows that the difference between the window size $50\\%$ and the full video $(100\\%)$ is larger for long videos. This shows that long-term context is in particular for long videos important. We also evaluated whether choosing a window for each video instead of choosing a window based on the average video length (fixed) performs better. For the video-specific window size, the window size is set to the percentage of each video. Fig. 5 (right) shows that a video-specific window size performs much worse than a fixed window. Varying the amount of context for each video is thus not beneficial. Fig. 1 shows qualitative results of our approach for different amounts of temporal context with a fixed window for a video from the Assembly101 dataset.", + "bbox": [ + 75, + 579, + 470, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State of the Art", + "text_level": 1, + "bbox": [ + 76, + 830, + 369, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the performance comparison of our method with state-of-the-art methods on the Breakfast and 50Salads datasets in Table 1 and on the Assembly101 dataset [41] in", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0765e105ea9b7ae9c71a297cc6b7c8095f7e31fd26afcbb752ddeb183343379a.jpg", + "image_caption": [ + "Figure 4: Impact of different sizes of the input window on the 50Salads dataset (left) and the Assembly101 dataset (right). The window size is given in percentage of the average length of a video in the corresponding dataset. $100\\%$ denotes entire video." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 697, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f87ca3a0883076c899867d20007ef72ceada82d0d70be014eb62520dbfe1972.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 90, + 888, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e3e0ed7221c707ec6f2d85ffa1433634f4307174c50be9c8d55495d3e5fe275.jpg", + "image_caption": [ + "Figure 5: Results for different video lengths (rights). Comparison of a fixed with a video-specific input window size (left). Both plots are for 50Salads." + ], + "image_footnote": [], + "bbox": [ + 504, + 339, + 696, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e6986f0237908ffdfa23405ce3cd8d13ce60fb069613063cc94d672650b5a5c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 340, + 887, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2.", + "text_level": 1, + "bbox": [ + 500, + 507, + 555, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On Breakfast and 50Salads, our method outperforms all comparable methods in terms of F1 score at all thresholds, which is the most important measure. Our method also achieves a better Edit score than all methods except for UVAST [4] and TCTr [3]. UVAST uses ASFormer [61] as encoder and an additional alignment decoder. The higher Edit score of UVAST is expected since the approach has an additional head that predicts the sequence of actions and thus maximizes the Edit score by an additional loss. This, however, comes at the cost of a much lower frame accuracy. TCTr [3] achieves a higher frame-wise accuracy on Breakfast and a higher Edit score on 50 Salads, but it performs worse for the other metrics. The approach uses a boundary detection module and compresses the temporal features, which is a complementary approach. It needs to be noted that our approach outperforms TCTr for all metrics on 50Salads if we use 10 instead of 9 layers as shown in Table 9. However, 9 layers perform better than 10 layers for the other datasets. We therefore report in Tables 1 and 2 only the results with 9 layers.", + "bbox": [ + 496, + 522, + 893, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also compare to approaches that use different input features, perform additional test augmentation, or add additional constraints. On Breakfast, only DTL [59] reports better results. DTL [59] uses additional logic-based constraints for training ASFormer and assumes that action se", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10355", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1852c614f5c537f382d6d880d5197d11b9ccecbd1362737a97f4545a92cec6ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNoteBreakfast50Salads
F1@{10,25,50}EditAccF1@{10,25,50}EditAcc
IDT+LM [39]DF1-----44.438.927.845.848.7
ST-CNN [24]DF2-----55.949.637.145.959.4
ED-TCN [23]DF2-----68.063.952.659.864.7
TDRN [25]DF2-----72.968.557.266.068.1
SSA-GAN [13]DF3----43.374.971.767.069.873.3
Bridge-Prompt [26]DF4-----89.287.881.383.888.1
C2F-TCN [42]TA72.268.757.669.676.084.381.872.676.484.9
UVAST [4]+ViterbiP75.970.057.276.566.089.187.681.783.987.4
UVAST [4]+FIFA [43]P76.971.558.077.169.788.987.078.583.984.5
Liu [32] + ASRF [18]P77.572.359.576.773.787.986.680.582.786.6
DTL [59]C78.874.562.977.775.887.185.778.580.586.9
MS-TCN [1]-52.648.137.961.766.376.374.064.567.980.7
MS-TCN++ [27]-64.158.645.965.667.680.778.570.174.383.7
DTGRM [53]-68.761.946.668.968.379.175.966.172.080.0
MuCon [44]-73.266.148.476.362.8-----
Gao et al. [14]-74.969.055.273.370.780.378.069.873.482.2
BCN [56]-68.765.555.066.270.482.381.374.074.384.4
SSTDA [8]-75.069.155.273.770.283.081.573.875.882.2
C2F-TCN [42]-70.166.656.268.273.576.673.062.569.280.1
ASRF [18]-74.368.956.172.467.684.983.577.379.384.5
UVAST [4]-76.770.056.677.268.286.281.270.483.979.5
DPRN [35]-75.670.557.675.171.787.886.379.482.087.2
LGTNN [48]-76.271.557.575.272.587.586.279.882.086.1
ASFormer [61]-76.070.657.475.073.585.183.476.079.685.6
TCTr [3]-76.671.158.576.177.587.586.180.283.486.6
LTContext (Ours)-77.672.660.177.074.289.487.782.083.287.7
", + "bbox": [ + 81, + 88, + 465, + 359 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bf5ec1333a1673e1cb5bbb4936471c013330cb4238dccfe69ce90b70980054bf.jpg", + "table_caption": [ + "Table 1: Results on the Breakfast and 50Salads datasets. The best and second best results for the methods in the bottom half are shown in bold and underlined since only methods without additional notes are directly comparable. P: additional post-processing; C: additional constraints; DF: different features (DF1: Improved Dense Trajectories (IDT), DF2: Spatio-temporal VGG-style CNN, DF3: Generative Adversarial Network (GAN), DF4: ActionCLIP); TA: test augmentation. In the top half, we highlight results that are better in italic bold formatting." + ], + "table_footnote": [], + "table_body": "
MethodAssembly101
F1@{10,25,50}EditAccParams (M)Inference (sec)
MS-TCN++ [27]31.627.820.630.737.11.080.1
UVAST [4]*32.128.320.831.537.41.223.2
C2F-TCN [42]33.329.021.332.439.26.890.1
ASFormer [61]*33.429.221.430.538.81.132.1
LTContext (Ours)33.930.022.630.441.20.720.6
", + "bbox": [ + 78, + 532, + 465, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Comparison with state-of-the-art methods on the Assembly101 dataset. The best and second best results are shown in bold and underlined. *We trained UVAST [4] and ASFormer [61] using the code of the authors.", + "bbox": [ + 75, + 619, + 468, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quences can be modeled by a directed acyclic graph. This is very efficient on the Breakfast dataset, which has relatively few actions per video compared to the other datasets. However, our approach outperforms DTL for all metrics on 50Salads where the structure of the action sequences is more complex. Furthermore, DTL cannot be applied to datasets like Assembly101 where actions occur more than once in a video. On 50Salads, only Bridge-Prompt [26] performs slightly better for some metrics. Bridge-Prompt proposes an approach for feature learning using a vision-language model such as ActionCLIP [55] in combination with ASFormer. The approach is thus complementary and not comparable to our approach. Nevertheless, our approach achieves higher", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$F1@10$ and $F1@50$ scores on 50Salads. If we use 10 instead of 9 layers as shown in Table 9, we even outperform Bridge-Prompt for all measures except for frame-wise accuracy. In summary, our approach outperforms all methods in terms of F1 score, which is the most precise measure of the segmentation quality, at all thresholds on 50Salads.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assembly101 is the largest dataset both in terms of the number of videos and their length, and it is the most challenging dataset. Since ASFormer [61] and UVAST [4] have not been evaluated on this dataset, we trained ASFormer and UVAST on Assembly101 using the publicly available source code and report the results in Table 2 as well. We outperform all methods on Assembly101 in F1 scores. C2F-TCN [42] achieves the best edit score. Since we trained ASFormer [61] and UVAST [4] by ourselves, we can compare the training time on Assembly101. While our approach requires 1 day and 18 hours, ASFormer [61] and UVAST [4] needed 4 weeks and 2 weeks, respectively.", + "bbox": [ + 496, + 181, + 893, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 372, + 710, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figs. 6-8, we present some qualitative results for the Assembly101, Breakfast, and 50Salads dataset. The first row of each figure shows the middle frame of each ground-truth action segment. In the second, third, and fourth rows, the ground truth segmentation, the predictions of ASFormer [61], and the prediction of our model (LTContext) are shown, respectively. In Fig. 6, our approach shows much fewer errors than ASFormer. Although our approach recognizes all action classes that occur in the video, there are several errors where some instances are missed like the last action segment of the bottom row which corresponds to the action 'detach base'. This indicates how challenging the Assembly101 dataset is. In Fig. 7, the predictions of our model are very close to the ground-truth. ASFormer overestimates the duration of the green segment, which corresponds to the action 'spoon powder', and hallucinates purple segments at the beginning and end of the video. In Fig. 8, both methods estimate the segments well, but ASFormer predicts wrongly two orange segments, which correspond to the action 'mix dressing', and the olive segment is too short, which corresponds to the action 'add salt'.", + "bbox": [ + 496, + 395, + 893, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 722, + 663, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We finally evaluate the impact of each component of our architecture. For the ablation studies, we report the results averaged over the 5 splits of the 50Salads dataset.", + "bbox": [ + 496, + 744, + 893, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of attention types In Table 3, we show the impact of using the combination of windowed and long-term context (LTContext) attention in the LTContext block illustrated in Fig. 3. We first compare it to two variants where we use only windowed or only long-term context attention. In order to keep the number of parameters the same, we still use", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10356", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19a37f9fa2afed00f320001205ad8a91eb6b1dbd4db0d8d1396c9684d254007a.jpg", + "image_caption": [ + "Figure 6: Qualitative results on Assembly101. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext. It can be best viewed by using the zoom function of a PDF viewer." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 893, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a5e768fa14b18d64125e296077d1579173596435439661b9103cb8d1525804b.jpg", + "image_caption": [ + "Figure 7: Qualitative results on Breakfast. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext." + ], + "image_footnote": [], + "bbox": [ + 94, + 372, + 890, + 506 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e1d8beb1401b6c3319e81f658c554d410d646b47db3cb1bd13dd1c7bbfdde24.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model ArchitectureF1@{10, 25, 50}EditAcc
LTContext89.487.782.083.287.7
- Windowed Attention87.785.177.482.185.5
- LTContext Attention84.182.674.178.285.3
WA (S1) + LTContext (S2-S4)87.885.079.481.485.2
", + "bbox": [ + 78, + 571, + 467, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Impact of using LTContext and Windowed Attention (WA) on 50Salads. In the case of Windowed Attention, we use two windowed attention blocks instead of a combination of windowed and LTContext attention. In the case of LTContext Attention, we use two LTContext attention blocks. The last row corresponds to using only windowed attention in stage 1 and only LTContext attention in stages 2 to 4.", + "bbox": [ + 75, + 660, + 470, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in these cases two attention blocks within LTContext. The results show that combining both types of attention leads to better results. In particular, the F1@50 score is substantially higher.", + "bbox": [ + 75, + 806, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we use four LTContext blocks. We also evaluate what happens if we vary the attention not within", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an LTContext block but between the four LTContext blocks. For this, we use only windowed attention for the first LTContext block and only long-term context attention for the other three LTContext blocks. The last row in Table 3 shows that this does not perform better than using only windowed attention and it is worse than combining windowed and long-term context attention within an LTContext block.", + "bbox": [ + 496, + 574, + 893, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of different values of $W$ and $G$ The parameter $W$ controls the size of the local window for the local attention and the parameter $G$ controls the sparseness of the global attention. If not otherwise specified, we use $W = G = 64$ in our experiments. The results in Table 4 show that the F1 score drops when the size of the local window $W$ becomes smaller. Note that larger values of $W$ increase the memory and computational cost. When we decrease $G$ , the F1 score also drops but not as drastic as for $W$ .", + "bbox": [ + 496, + 699, + 893, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of overlaps for windowed attention As described in Section 3.1, we use an overlap for the keys and values for the windowed attention. In Table 5, we also report the result", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10357", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3b4b14841ba7291dc5af17a582d79e4b482f90702d664c41c88792dee8eec74.jpg", + "image_caption": [ + "Figure 8: Qualitative results on 50Salads. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext." + ], + "image_footnote": [], + "bbox": [ + 130, + 89, + 851, + 214 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/643744abcb0856330f91cc2406152cd460f8bb17315b753b9e79357279791066.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
WGF1@{10, 25, 50}EditAcc
86483.882.174.577.087.0
166485.482.977.177.187.0
326487.685.879.881.586.6
646489.487.782.083.287.7
643288.887.281.383.287.7
641688.687.080.382.686.2
64888.687.179.982.986.6
", + "bbox": [ + 116, + 277, + 426, + 412 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f354e03d20f7005a2716be7506936b8bd472247fee008c5f934a9f0d963c52ba.jpg", + "table_caption": [ + "Table 4: Impact of the parameters $W$ and $G$ on 50 Salads." + ], + "table_footnote": [], + "table_body": "
F1@{10, 25, 50}EditAcc
Non-Overlapping87.985.579.181.485.1
Overlap (2 Windows)89.487.782.083.287.7
Overlap (3 Windows)89.288.080.883.286.6
", + "bbox": [ + 107, + 450, + 437, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "when the keys and values do not overlap, i.e., they are the same as the queries. It is interesting to note that our attention can be interpreted as a combination of a reshaping operation with axial attention [16] in the special case without overlap. The results, however, show that an overlap improves the results. Using a larger overlap where the keys and values consist of three consecutive windows does not improve the results further.", + "bbox": [ + 75, + 579, + 472, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the order of the attention The LTContext block shown in Fig. 3 uses first windowed attention and then long-term context attention. Table 6 shows the results when we change the order of windowed and long-term context attention within the LTContext block. If the order is changed, the performance decreases. Since the LTContext blocks are repeated, the drop in performance is moderate.", + "bbox": [ + 75, + 724, + 470, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the cross-attention As described in Section 3.2, we use cross-attention in stages 2 to 4. Table 7 shows that the performance drastically decreases without cross-attention.", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a59f647768768f3fc5a49fbaf8db2674cba8af25d575441f8518360af31fa2a0.jpg", + "table_caption": [ + "Table 5: Impact of using an overlap for the keys and values for the local attention on 50Salads." + ], + "table_footnote": [], + "table_body": "
F1@{10, 25, 50}EditAcc
WA-LTContext89.487.782.083.287.7
LTContext-WA88.386.480.481.686.1
", + "bbox": [ + 519, + 279, + 870, + 339 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0312fb8748ad8d72b04bd31ec90605b2b76bda1afb4713016822c1d9c2faaf3a.jpg", + "table_caption": [ + "Table 6: Impact of the attention order within the LTContext block on 50Salads. Our model computes first windowed attention (WA) and then long-term context attention (row 1)." + ], + "table_footnote": [], + "table_body": "
F1@{10, 25, 50}EditAcc
Without Cross-Attention82.278.570.473.680.2
LTContext89.487.782.083.287.7
", + "bbox": [ + 519, + 405, + 870, + 457 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65139d554878af774ffac09fd24e1094d7875dc447e86c007b2aab0392c97db9.jpg", + "table_caption": [ + "Table 7: Impact of the cross-attention on 50Salads." + ], + "table_footnote": [], + "table_body": "
F1@{10, 25, 50}EditAcc
Conv with Dilation89.487.782.083.287.7
Conv without Dilation80.178.067.772.679.3
Without Conv81.879.267.172.778.9
", + "bbox": [ + 519, + 493, + 870, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: Impact of using LTContext blocks with 1D convolution but without dilation (row 2) and using LTContext blocks without 1D convolution (row 3).", + "bbox": [ + 498, + 569, + 893, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of using convolutions The LTContext block shown in Fig. 3 starts with a dilated 1D convolution with kernel size 3. In Table 8, we evaluate the impact of the dilated convolution by comparing to a LTContext block that uses a convolutional kernel of the same size but without dilation factor, and a LTContext block without 1D convolution. The results show that dilated convolutions have a very high impact on the performance.", + "bbox": [ + 496, + 637, + 895, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the number of layers As shown in Fig. 3, we repeat the LTContext blocks at each stage $N$ times. We used $N = 9$ layers in all experiments. Table 9 shows the impact of varying the number of layers. On 50Salads, all measures improve by increasing $N$ . It needs to be noted that we use $N = 9$ for the results reported in Table 1 although we can get even better results with 10 layers on 50 Salads. For the Breakfast and Assembly101 dataset, the best performance is", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10358", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e1000f31c3996c65a0eca89021b26cf3d5522e8c78718a1dd2a9fd860890506.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
number of layers (N)50Salads
F1@{10, 25, 50}EditAcc
887.285.177.680.885.7
989.487.782.083.287.7
1089.588.182.484.187.7
Breakfast
875.770.557.774.573.1
977.672.660.177.074.2
1077.372.459.776.473.5
Assembly101
831.928.421.327.841.0
933.930.022.630.441.2
1032.629.321.928.741.5
", + "bbox": [ + 98, + 88, + 446, + 292 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/41d2456dc440e9776237cbdd1dfce763ebda033e70312d11f1905ec4274fd697.jpg", + "table_caption": [ + "Table 9: Impact of the number of layers on 50Salads, Breakfast, and Assembly101." + ], + "table_footnote": [], + "table_body": "
number of headsF1@{10, 25, 50}EditAcc
189.487.782.083.287.7
288.887.080.582.987.0
489.287.180.883.186.9
888.286.480.582.186.1
", + "bbox": [ + 98, + 344, + 446, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "achieved with 9 layers.", + "bbox": [ + 76, + 478, + 233, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Impact of the number of attention heads In our implementation, we do not use multiple attention heads. Nevertheless, we evaluated the impact of using multiple heads in Table 10 since most transformers use multiple heads. The results, however, show that there is no benefit in using multiple heads. The results are consistent with the observations in [33, 30]. For example, [33] shows that many attention heads can often be reduced to a single head without impacting the performance. They also argue that some tasks are more reliant on multiple heads than others. Temporal action segmentation seems to be a task where one head is sufficient.", + "bbox": [ + 75, + 511, + 470, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Impact of the number of stages As shown in Fig. 3, we use four stages of LTContext blocks, each of them with 9 layers. We evaluate the impact of the number of stages in Table 11. As can be seen, using multiple stages helps to reduce the over-segmentation error and improves the F1 score and Edit score significantly compared to using only one stage. Increasing the number of stages up to 4 improves all metrics, but using 5 stages decreases the performance and the network starts to overfit.", + "bbox": [ + 75, + 696, + 468, + 832 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 861 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we addressed the question of how much temporal long-term context is needed for action segmentation.", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c18a2b1d2fb5f13bec40c1e9d247571b5e3bc2cb120c6c0205cc06cb6b7f0817.jpg", + "table_caption": [ + "Table 10: Impact of the number of attention heads." + ], + "table_footnote": [], + "table_body": "
number of stagesF1@{10, 25, 50}EditAcc
156.354.149.945.684.7
286.284.578.280.586.7
387.685.778.381.285.5
489.487.782.083.287.7
589.086.880.083.085.7
", + "bbox": [ + 531, + 88, + 859, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 11: Impact of the number of stages on 50Salads.", + "bbox": [ + 514, + 193, + 874, + 209 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our analysis indicates that allowing networks to operate on the full input sequence is more beneficial compared to the case where the model has only access to a subset of the input. Based on our analysis, we presented LTContext, an approach for temporal action segmentation, where we leverage sparse attention to capture the long-term context of a video and windowed attention to model the local information in the neighboring frames. Our approach achieves state-of-the-art segmental F1 scores on the 50Salads and Assembly101 datasets, which contain long videos.", + "bbox": [ + 496, + 229, + 893, + 380 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 392, + 660, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), the project iBehave (receiving funding from the programme \"Netzwerke 2021\", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors.", + "bbox": [ + 496, + 417, + 893, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 566, + 596, + 580 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yazan Abu Farha and Juergen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. ViViT: A video vision transformer. In IEEE International Conference on Computer Vision (ICCV), 2021.", + "[3] Nicolas Aziere and Sinisa Todorovic. Multistage temporal convolution transformer for action segmentation. Image and Vision Computing, 2022.", + "[4] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In European Conference on Computer Vision (ECCV), 2022.", + "[5] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document. arXiv preprint arXiv:2004.05150, 2020.", + "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021." + ], + "bbox": [ + 506, + 590, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10359", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[7] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.", + "[8] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[9] Zihang Dai, Hanxiao Liu, Quoc Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems (NeurIPS), 2021.", + "[10] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics (ACL), 2019.", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021.", + "[12] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021.", + "[13] Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Fine-grained action segmentation using the semi-supervised action gan. Pattern Recognition, 2020.", + "[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[15] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. Conformer: Convolution-augmented transformer for speech recognition. In INTERSPEECH, 2020.", + "[16] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv, 2019.", + "[17] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[18] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021.", + "[19] Svebor Karaman, Lorenzo Seidenari, and Alberto Del Bimbo. Fast saliency based pooling of fisher encoded dense trajectories. In European Conference on Computer Vision (ECCV) Workshops, 2014.", + "[20] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014.", + "[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 780-787, 2014.", + "[22] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2016.", + "[23] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.", + "[24] Colin Lea, Austin Reiter, René Vidal, and Gregory D Hager. Segmental spatiotemporal cnns for fine-grained action segmentation. In European Conference on Computer Vision (ECCV), 2016.", + "[25] Peng Lei and Sinisa Todorovic. Temporal deformable residual networks for action segmentation in videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.", + "[26] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[27] Shi-Jie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. MS-TCN++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020.", + "[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE International Conference on Computer Vision (ICCV), 2019.", + "[30] Liyuan Liu, Jialu Liu, and Jiawei Han. Multi-head or singlehead? an empirical comparison for transformer training. arXiv, 2021.", + "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021.", + "[32] Zhichao Liu, Leshan Wang, Desen Zhou, Jian Wang, Songyang Zhang, Yang Bai, Errui Ding, and Rui Fan. Temporal segment transformer for action segmentation. arXiv preprint arXiv:2302.13074, 2023.", + "[33] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems (NeurIPS), 2019.", + "[34] Piotr Nawrot, Szymon Tworkowski, Michal Tyrolski, Lukasz Kaiser, Yuhuai Wu, Christian Szegedy, and Henryk Michalewski. Hierarchical transformers are more efficient language models. arXiv preprint arXiv:2110.13711, 2021." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10360", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Junyong Park, Daekyum Kim, Sejoon Huh, and Sungho Jo. Maximization and restoration: Action segmentation through dilation passing and temporal reconstruction. Pattern Recognition, 2022.", + "[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning (ICML), 2018.", + "[37] Yifan Peng, Siddharth Dalmia, Ian Lane, and Shinji Watanabe. Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding. In International Conference on Machine Learning (ICML), 2022.", + "[38] Jiezhong Qiu, Hao Ma, Omer Levy, Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. Findings of Empirical Methods in Natural Language Processing (EMNLP), 2020.", + "[39] Alexander Richard and Juergen Gall. Temporal action detection using a statistical language model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.", + "[40] Marcus Rohrbach, Sikandar Amin, Mykhaylo Andriluka, and Bernt Schiele. A database for fine grained activity detection of cooking activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012.", + "[41] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[42] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. arXiv preprint, 2021.", + "[43] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fifa: Fast inference approximation for action segmentation. In DAGM German Conference on Pattern Recognition (GCR), 2021.", + "[44] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021.", + "[45] Sebastian Stein and Stephen J. McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In ACM International Joint Conference on Pervasive and Ubiquitous Computing, page 729-738, 2013.", + "[46] Kevin Tang, Li Fei-Fei, and Daphne Koller. Learning latent temporal structure for complex event detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012.", + "[47] Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. ACM Comput. Surv., 2022.", + "[48] Xiaoyan Tian, Ye Jin, and Xianglong Tang. Local-global transformer neural network for temporal action segmentation. Multimedia Systems, 2022.", + "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021.", + "[50] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[51] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European Conference on Computer Vision (ECCV), 2022.", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017.", + "[53] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. AAAI Conference on Artificial Intelligence, 2020.", + "[54] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021.", + "[56] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In European Conference on Computer Vision (ECCV), 2020.", + "[57] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In IEEE International Conference on Computer Vision (ICCV), 2021.", + "[58] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. In Advances in Neural Information Processing Systems (NeurIPS), 2021.", + "[59] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2022.", + "[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021.", + "[61] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In British Machine Vision Conference (BMVC), 2021.", + "[62] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021.", + "[63] Long Zhao, Zizhao Zhang, Ting Chen, and Dimitris Metaxas abd Han Zhang. Improved transformer for high-resolution GANs. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + ], + "bbox": [ + 503, + 92, + 893, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10361", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_model.json b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_model.json new file mode 100644 index 0000000000000000000000000000000000000000..658a1ac375992bd05b3ade1fd0654208180461c4 --- /dev/null +++ b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.88, + 0.153 + ], + "angle": 0, + "content": "How Much Temporal Long-Term Context is Needed for Action Segmentation?" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.18, + 0.363, + 0.198 + ], + "angle": 0, + "content": "Emad Bahrami" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.181, + 0.576, + 0.199 + ], + "angle": 0, + "content": "Gianpiero Francesca" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.181, + 0.737, + 0.199 + ], + "angle": 0, + "content": "Juergen Gall1,3" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.2, + 0.464, + 0.217 + ], + "angle": 0, + "content": "1University of Bonn, Germany" + }, + { + "type": "text", + "bbox": [ + 0.496, + 0.2, + 0.714, + 0.217 + ], + "angle": 0, + "content": "\\(^{2}\\)Toyota Motor Europe, Belgium" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.218, + 0.737, + 0.235 + ], + "angle": 0, + "content": "\\(^{3}\\)Lamarr Institute for Machine Learning and Artificial Intelligence, Germany" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.301, + 0.474, + 0.619 + ], + "angle": 0, + "content": "Modeling long-term context in videos is crucial for many fine-grained tasks including temporal action segmentation. An interesting question that is still open is how much long-term temporal context is needed for optimal performance. While transformers can model the long-term context of a video, this becomes computationally prohibitive for long videos. Recent works on temporal action segmentation thus combine temporal convolutional networks with self-attention that are computed only for a local temporal window. While these approaches show good results, their performance is limited by their inability to capture the full context of a video. In this work, we try to answer how much long-term temporal context is required for temporal action segmentation by introducing a transformer-based model that leverages sparse attention to capture the full context of a video. We compare our model with the current state of the art on three datasets for temporal action segmentation, namely 50Salads, Breakfast, and Assembly101. Our experiments show that modeling the full context of a video is necessary to obtain the best performance for temporal action segmentation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.648, + 0.21, + 0.664 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Temporal action segmentation can be used in many real-world applications such as monitoring production lines or studying animal behavior. In these settings, the videos can be very long and it is required to recognize the start and end of all actions that occur in a video as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Recently, combinations of temporal convolutional networks [1, 27] with self- and cross-attention from transformers [61, 4] have shown impressive results for temporal action segmentation. These works are in line with other hybrid models [15, 9, 61, 58] that combine the attention modules with convolutions to compensate for the lack of strong inductive bias of pure transformers. However, the emergence of datasets like Assembly101 [41], where subjects perform assembly tasks, poses a new challenge in the area of temporal action segmentation due to the existence of long videos" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.269, + 0.895, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Figure 1: Datasets like Assembly101 contain long videos of assembly tasks and an action label needs to be predicted for each frame. The first row shows some frames of a video. The second row shows the ground-truth labels for all frames of the video where different colors correspond to different action labels. Rows 3-6 show the predictions of the proposed model for different amounts of long-term context where \\(100\\%\\) means the temporal context of the full video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.701 + ], + "angle": 0, + "content": "that can last up to 25 minutes. Since modeling the long-term context of such a video is very expensive, Yi et al. [61] proposed to compute the attention for a local temporal window. In order to understand the impact of the window size on the temporal action segmentation accuracy, we analyze the impact of the window size on two datasets with long videos in Section 4.1. Fig. 1 shows some qualitative results of this study on Assembly101. Indeed, the results show that modeling the long-term context of an entire video is very important for temporal action segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Based on this finding, we revisit how temporal attention is modeled in transformer architectures for temporal action segmentation. [61, 4] use a hierarchy of temporal windows, which makes the training on long video sequences as they occur in Assembly101 [41] very expensive. Inspired by works that decompose attention over the spatial and temporal domain for short video clips [6, 62], we propose to iterate between computing windowed local attention and sparse long-term context attention such that both short and long-term context are modeled. This approach is particularly suitable for temporal action segmentation since the local attentions focus on the similarity or dissimilarity of features within an action segment or between neighboring action" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10351" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.472, + 0.152 + ], + "angle": 0, + "content": "segments whereas the long-term context attention focuses on the relations between actions within the entire video. The source code is available at https://github.com/LTContext/LTContext." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.166, + 0.22, + 0.182 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.192, + 0.472, + 0.388 + ], + "angle": 0, + "content": "The traditional sliding window with non-maxima suppression has been among the early approaches for action segmentation [19, 40]. [22, 20, 46] adopted hidden Markov Models (HMMs) for temporal modeling. [39] incorporates length and temporal context and uses dynamic programming for inference. Temporal convolutional networks (TCN) with temporal pooling are used by [23] to classify each video frame. Later, [1] introduced a multi-stage TCN capable of maintaining high temporal resolution, which is necessary for fine-grained recognition. [18] reduces the over-segmentation error by adding an action boundary regression network to refine the frame-wise results. [17] proposed another refinement method based on a graph convolutional network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.389, + 0.472, + 0.78 + ], + "angle": 0, + "content": "Transformers [52] were originally used for natural language processing and only rely on attention for sequence modeling. Recently, transformers have been widely adopted in vision [11, 57, 31, 54], speech recognition [15, 37], and action recognition [6, 2]. The original transformers suffered from two issues. First, the cost of the self-attention operation is quadratic with respect to the sequence length. This has been addressed by many methods that improve the memory efficiency of transformers [5, 34, 10]. We refer to [47] for a comprehensive survey of efficient transformers. For instance, restricting the attention to a fixed window size [5] is one approach. Multi-axis self-attention [51] combines block-attention with grid-attention, which is based on a spatial grid overlaid on the entire 2D space. Adapting sparse attention for capturing global information has shown to be an effective solution in vision tasks such as high-resolution image generation [63, 50], object detection [51, 60], and instance segmentation [51, 60]. MViT [12] uses a pooling operation to reduce the space-time resolution before attention and MViTv2 [28] improves MViT by adding the relative position along (x,y,t) and residual connections. The second issue of transformers is the poor generalization due to a relatively weak inductive bias [9, 49] compared to convolutional neural networks (CNN). Hybrid models that combine self-attention and convolution layers have thus been proposed for vision tasks [58, 9, 51, 57] and speech recognition [15, 37]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In action recognition, works such as [6, 62] adopt the idea of using attention for video understanding. In Times-Former [6], the self-attention is first applied along the temporal dimension for each single patch (time attention), i.e., over all frames of a short video clip of a few seconds. In a second step, the attention is applied over the patches of each frame (space attention). VIDTR [62] also decomposes self-attention into spatial and temporal attention, but additionally" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "down-samples the temporal dimension. In our work, we do not apply attention spatially and temporally, but we address the question of how temporal attention can be computed for very long sequences that last 25 minutes as it is required for temporal action segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.897, + 0.514 + ], + "angle": 0, + "content": "For temporal action segmentation, [61] proposed an architecture, called ASFormer, which is based on a multi-stage TCN [1] and equips the temporal convolutions with a local window attention [5]. This is done in a hierarchy where the size of the local window grows with each layer. Recently, [4] proposed a decoder on top of the encoder that generates action segments in an autoregressive manner. It uses two heads where the head on top of the encoder predicts framewise probabilities and the head of the decoder predicts the sequence of actions. Finally, an alignment decoder fuses the output of the two heads and aligns the predicted sequence of actions to the frames. [3] introduced a hybrid Temporal Convolution Transformer (TCTr) where they adapt an action boundary detector to adaptively estimate attention from local neighboring frames. [59] proposed to use additional constraints during training, but the approach assumes that action sequences can be modelled by a directed acyclic graph, which does not allow that actions occur more than once in a video. Recently, multi-modal approaches that combine language models with vision transformers have been proposed. For instance, [26] uses prompt engineering to extract features from pre-trained vision-language models such as ActionCLIP [55]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.546 + ], + "angle": 0, + "content": "3. Long-Term Context for Action Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.689 + ], + "angle": 0, + "content": "Recently, transformers combined with temporal convolutional neural networks [61, 4, 3] have shown a very good performance for temporal action segmentation. For this task, the frame-wise labels \\( c_{1}, \\ldots, c_{T} \\), where \\( c_{t} \\in \\mathcal{C} \\) and \\( C = |\\mathcal{C}| \\) denotes the number of action classes, need to be predicted for a given a video \\( X = (x_{1}, \\ldots, x_{T}) \\) with \\( T \\) frames, where \\( x_{t} \\) represents a feature map of size \\( D \\) at frame \\( t \\). Since \\( T \\) can be very large, [61, 4, 3] limit the self-attention to a local temporal window." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.897, + 0.901 + ], + "angle": 0, + "content": "In order to understand how much temporal long-term context is needed for temporal action segmentation, we limited the temporal input window and evaluated the quality of the temporal action segmentation in Section 4.1. The results in Fig. 4 show that temporal long-term context has a strong impact on the performance. Based on our analysis, we thus revisit the windowed attention of previous works for temporal action segmentation and propose to model the temporal long-term context of a video using sparse attentions [5, 38, 36]. Additionally, we equip our method with windowed attention to capture the locality between neighboring frames. In this way, we obtain a flexible design that is capable of providing long-term and local temporal context. While we describe first the Long-Term Context (LTCon" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10352" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.093, + 0.891, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.897, + 0.377 + ], + "angle": 0, + "content": "Figure 2: Illustration of windowed and long-term context attentions with a window of size 2. For the windowed attention, the sequence is partitioned into small windows and the attentions are computed for each window. For long-term context attention, the sequence is reordered such that the attentions are computed over the whole, but sparsely sampled sequence. After the attention, the output is reordered again to preserve the original order. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.471, + 0.431 + ], + "angle": 0, + "content": "text) block in Section 3.1, the entire network is described in Section 3.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.441, + 0.332, + 0.457 + ], + "angle": 0, + "content": "3.1. Temporal Context Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.471, + 0.584 + ], + "angle": 0, + "content": "The self-attention blocks of transformers are advantageous over convolutions in the aggregation of global information. However, applying attention to long sequences such as untrimmed videos is impractical due to the quadratic complexity of the self-attention blocks. To address this issue, we adopt an attention mechanism where we leverage sparse and windowed attention to model long-term and local temporal context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.472, + 0.646 + ], + "angle": 0, + "content": "The attention function transforms the input into a query, key, and value and computes the output as a weighted sum of the values. For example, given a sequence of features, \\(X\\in \\mathbb{R}^{T\\times D}\\), the attention can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.655, + 0.473, + 0.693 + ], + "angle": 0, + "content": "\\[\nA t t e n t i o n (Q, K, V) = \\mathrm {s o f t m a x} _ {r o w} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right) V \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.472, + 0.763 + ], + "angle": 0, + "content": "where \\(Q, K, V \\in \\mathbb{R}^{T \\times D}\\) are linearly transformed from \\(X\\). Since \\(T\\) is very large for long sequences, we need to modify \\(Q, K, V\\) to enable the modeling of long-term and local temporal context as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Temporal Windowed Attention For the temporal Windowed Attention (WA), we partition the sequence into nonoverlapping windows of size \\( W \\). Fig. 2 illustrates the case for \\( W = 2 \\), but we use \\( W = 64 \\) in practice. The impact of \\( W \\) is evaluated in Section 4. Instead of computing the attention over the entire sequence of length \\( T \\), we compute the attention \\( \\frac{T}{W} \\) times, where each query \\( Q \\in \\mathbb{R}^{W \\times D} \\) corresponds to each window. For the keys and values, we use" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.895, + 0.464 + ], + "angle": 0, + "content": "an overlap where we use the next window in addition, i.e., for each query \\(Q\\) we have \\(K,V\\in \\mathbb{R}^{2W\\times D}\\). We perform masking when \\(K\\) and \\(V\\) exceed the input sequence. We evaluate the impact of the overlap in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.897, + 0.713 + ], + "angle": 0, + "content": "Temporal Long-Term Context Attention For the temporal Long-Term Context (LTC) attention, the input is also partitioned into non-overlapping windows of size \\( G \\). However, instead of computing the attention over each window, the attention is computed over all windows where from each window only one element is taken. In the illustration of Fig. 2 with \\( G = 2 \\), we compute the attention over the first feature of all windows and the attention over the second feature of all windows. In general, we compute the attentions for \\( G \\) queries \\( Q \\in \\mathbb{R}^{\\frac{T}{G} \\times D} \\) where the keys and values are the same, i.e., \\( K, V \\in \\mathbb{R}^{\\frac{T}{G} \\times D} \\). The parameter \\( G \\) provides the flexibility to adjust the sparseness based on the available memory budget, e.g., \\( G = 1 \\) corresponds to the case where the attention is applied over the full sequence. In practice, we use \\( G = 64 \\) and we evaluate the impact of \\( G \\) in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "LTContext block The top of Fig. 3 illustrates the entire LTContext block. As in previous works [1, 61, 4], we use a 1D dilated temporal convolution with kernel size 3, where the dilation factor increases by factor 2 for each layer. The dilated temporal convolution is followed by a Gaussian Error Linear Unit (GELU). In the LTContext block, we first use the windowed and then the long-term context attention, which are shown in Fig. 2. We evaluate the impact of the order in Section 4. Finally, we use a linear layer with a residual connection to output the features for each frame, \\( F \\in \\mathbb{R}^{T \\times D} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10353" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.891, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.293, + 0.745, + 0.309 + ], + "angle": 0, + "content": "Figure 3: The network architecture of LTContext with LTContext blocks (top)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.334, + 0.3, + 0.35 + ], + "angle": 0, + "content": "3.2. LTContext Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.474, + 0.622 + ], + "angle": 0, + "content": "The entire LTContext network is depicted in Fig. 3. For a fair comparison, we will use the features that are provided for the corresponding datasets as input. In all cases, the dimensionality of the features is 2048. As in previous works, we use a linear layer to reduce the dimensionality of the features to 64. The output of each LTContext block is the feature map \\( F \\in \\mathbb{R}^{T \\times D} \\). We repeat each LTContext block \\( N \\) times where the dilation factor of the temporal convolution increases in each layer. In practice, we use \\( N = 9 \\) and we evaluate the impact of \\( N \\) in Section 4. After the first \\( N \\) layers of LTContext blocks, we use an additional linear layer to reduce the dimensionality \\( D \\) further to 32. The dimensionality reduction reduces the number of parameters from 1.42 million to 0.72 million without reducing the accuracy. We also use an additional linear layer followed by a softmax layer to generate the frame-wise class probabilities \\( P \\in \\mathbb{R}^{T \\times C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We continue with three additional stages where each stage consists of \\(N\\) layers of LTContext blocks. Note that we reset the dilation factor to 1 for the temporal convolution at the beginning of each stage and we compute the frame-wise class probabilities \\(P \\in \\mathbb{R}^{T \\times C}\\) after each stage, which contributes to the multi-stage loss. We use the cross-entropy loss combined with the mean squared error smoothing loss as introduced by [1] and used in [27, 61] for a fair comparison. Inspired by [61], we use the cross-attention for the LTContext blocks in stages 2 to 4. Instead of using the features \\(F\\) for the queries and keys for windowed and long-term context attention, the predictions \\(P\\) are used. We thus have \\(Q \\in \\mathbb{R}^{W \\times C}\\), \\(K \\in \\mathbb{R}^{2W \\times C}\\), and \\(V \\in \\mathbb{R}^{2W \\times D}\\) for the windowed attention and \\(Q \\in \\mathbb{R}_{G}^{T \\times C}\\), \\(K \\in \\mathbb{R}_{G}^{T \\times C}\\), and \\(V \\in \\mathbb{R}_{G}^{\\frac{T}{G} \\times D}\\) for the long-term context attention. While the cross-attention is not shown in Fig. 3, it only means that \\(P\\) is an additional input for the windowed and long-term context attention in stages 2-4. We evaluate the impact of the number" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.336, + 0.651, + 0.351 + ], + "angle": 0, + "content": "of stages in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.363, + 0.634, + 0.38 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.895, + 0.433 + ], + "angle": 0, + "content": "Datasets. We evaluate the performance of our proposed model on three challenging action segmentation datasets: 50Salads [45], Breakfast [21], and Assembly101 [41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.493 + ], + "angle": 0, + "content": "50Salads [45] contains 50 videos annotated with 17 action classes. On average, each video is 6.4 minutes long and has 18 action segments. Following previous works [1, 61, 4], we use five-fold cross-validation and report the average." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.569 + ], + "angle": 0, + "content": "Breakfast contains 1,712 videos of breakfast preparation activities with an average length of 2.3 minutes. There are 48 action classes and each video has on average 6.6 action segments. For evaluation, we report the average of the 4 splits for cross-validation as in [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.896, + 0.749 + ], + "angle": 0, + "content": "Assembly101 [41] is the largest dataset among the three datasets with 4,321 videos and 202 coarse action classes composed of 11 verbs and 61 objects. Assembly101 is a procedural activity dataset containing videos of people assembling and disassembling 101 toy vehicles. On average, each video includes 24 action segments and is 7.1 minutes long. Compared to Breakfast, Assembly101 has 2.5 times more videos, 6.7 times more hours of video footage, 9.3 times more action segments, and 4.2 more action classes. For our evaluation, we follow the setting for temporal action segmentation [41] and report the results on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.87 + ], + "angle": 0, + "content": "For a fair comparison, we use the features that are provided for the datasets and that have been used in previous works. We use the I3D [7] features for the 50Salads and Breakfast datasets and TSM [29] features for the Assembly101 dataset [41]. Both features are 2048 dimensional vectors. Following [1], we also used the temporally downsampled features for 50Salads to compensate for the different frame-rates of the datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We report the frame-wise accuracy (Acc), segmental Edit distance, and segmental F1 score at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10354" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "the overlapping thresholds of \\(10\\%\\), \\(25\\%\\), and \\(50\\%\\) denoted by \\(F1@{10,25,50}\\). The intersection over the union (IoU) ratio is used as the overlapping threshold. The edit distance measures only the order of the actions but not the duration. The frame-wise accuracy measures the accuracy per frame. It is dominated by actions that have a long duration and it is not very sensitive to over-segmentation errors that occur when a few frames within a ground-truth action segment are wrongly classified. The F1 score is the most reliable measure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.253, + 0.471, + 0.283 + ], + "angle": 0, + "content": "4.1. How much temporal long-term context is needed?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.471, + 0.442 + ], + "angle": 0, + "content": "We first present the results of our analysis on the impact of using the full sequence as input compared to using a temporal window. The goal of this experiment is to shed light on how much temporal long-term context is needed for the task of temporal action segmentation. For the analysis, we use only 50Salads and Assembly101 since the videos in Breakfast are too short. For the experiments, we train our approach (LTContext) and ASFormer [61] either on the full video sequences or we divide the videos into shorter sequences where the full context is lacking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.471, + 0.579 + ], + "angle": 0, + "content": "Fig. 4 shows the result of this experiment on the 50Salads and Assembly101 dataset. We report the window size as percentage of the average length of a video in the corresponding dataset and \\(100\\%\\) means that the full sequence has been used. The results clearly show that the full context of an entire sequence is advantageous over allowing the model to see only a window of the input sequence even if the window is large \\((50\\%)\\). We can also observe that our approach benefits more from the full sequence than ASFormer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.471, + 0.822 + ], + "angle": 0, + "content": "We furthermore evaluated whether the impact of the window size is stronger for longer videos. To this end, we sorted all test videos into four quarters by their length. Fig. 5 (left) shows that the difference between the window size \\(50\\%\\) and the full video \\((100\\%)\\) is larger for long videos. This shows that long-term context is in particular for long videos important. We also evaluated whether choosing a window for each video instead of choosing a window based on the average video length (fixed) performs better. For the video-specific window size, the window size is set to the percentage of each video. Fig. 5 (right) shows that a video-specific window size performs much worse than a fixed window. Varying the amount of context for each video is thus not beneficial. Fig. 1 shows qualitative results of our approach for different amounts of temporal context with a fixed window for a video from the Assembly101 dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.37, + 0.848 + ], + "angle": 0, + "content": "4.2. Comparison with State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We present the performance comparison of our method with state-of-the-art methods on the Breakfast and 50Salads datasets in Table 1 and on the Assembly101 dataset [41] in" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.699, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.091, + 0.889, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.247, + 0.895, + 0.321 + ], + "angle": 0, + "content": "Figure 4: Impact of different sizes of the input window on the 50Salads dataset (left) and the Assembly101 dataset (right). The window size is given in percentage of the average length of a video in the corresponding dataset. \\(100\\%\\) denotes entire video." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.34, + 0.697, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.341, + 0.888, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.449, + 0.895, + 0.494 + ], + "angle": 0, + "content": "Figure 5: Results for different video lengths (rights). Comparison of a fixed with a video-specific input window size (left). Both plots are for 50Salads." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.508, + 0.557, + 0.521 + ], + "angle": 0, + "content": "Table 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.895, + 0.825 + ], + "angle": 0, + "content": "On Breakfast and 50Salads, our method outperforms all comparable methods in terms of F1 score at all thresholds, which is the most important measure. Our method also achieves a better Edit score than all methods except for UVAST [4] and TCTr [3]. UVAST uses ASFormer [61] as encoder and an additional alignment decoder. The higher Edit score of UVAST is expected since the approach has an additional head that predicts the sequence of actions and thus maximizes the Edit score by an additional loss. This, however, comes at the cost of a much lower frame accuracy. TCTr [3] achieves a higher frame-wise accuracy on Breakfast and a higher Edit score on 50 Salads, but it performs worse for the other metrics. The approach uses a boundary detection module and compresses the temporal features, which is a complementary approach. It needs to be noted that our approach outperforms TCTr for all metrics on 50Salads if we use 10 instead of 9 layers as shown in Table 9. However, 9 layers perform better than 10 layers for the other datasets. We therefore report in Tables 1 and 2 only the results with 9 layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We also compare to approaches that use different input features, perform additional test augmentation, or add additional constraints. On Breakfast, only DTL [59] reports better results. DTL [59] uses additional logic-based constraints for training ASFormer and assumes that action se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10355" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.467, + 0.36 + ], + "angle": 0, + "content": "
MethodNoteBreakfast50Salads
F1@{10,25,50}EditAccF1@{10,25,50}EditAcc
IDT+LM [39]DF1-----44.438.927.845.848.7
ST-CNN [24]DF2-----55.949.637.145.959.4
ED-TCN [23]DF2-----68.063.952.659.864.7
TDRN [25]DF2-----72.968.557.266.068.1
SSA-GAN [13]DF3----43.374.971.767.069.873.3
Bridge-Prompt [26]DF4-----89.287.881.383.888.1
C2F-TCN [42]TA72.268.757.669.676.084.381.872.676.484.9
UVAST [4]+ViterbiP75.970.057.276.566.089.187.681.783.987.4
UVAST [4]+FIFA [43]P76.971.558.077.169.788.987.078.583.984.5
Liu [32] + ASRF [18]P77.572.359.576.773.787.986.680.582.786.6
DTL [59]C78.874.562.977.775.887.185.778.580.586.9
MS-TCN [1]-52.648.137.961.766.376.374.064.567.980.7
MS-TCN++ [27]-64.158.645.965.667.680.778.570.174.383.7
DTGRM [53]-68.761.946.668.968.379.175.966.172.080.0
MuCon [44]-73.266.148.476.362.8-----
Gao et al. [14]-74.969.055.273.370.780.378.069.873.482.2
BCN [56]-68.765.555.066.270.482.381.374.074.384.4
SSTDA [8]-75.069.155.273.770.283.081.573.875.882.2
C2F-TCN [42]-70.166.656.268.273.576.673.062.569.280.1
ASRF [18]-74.368.956.172.467.684.983.577.379.384.5
UVAST [4]-76.770.056.677.268.286.281.270.483.979.5
DPRN [35]-75.670.557.675.171.787.886.379.482.087.2
LGTNN [48]-76.271.557.575.272.587.586.279.882.086.1
ASFormer [61]-76.070.657.475.073.585.183.476.079.685.6
TCTr [3]-76.671.158.576.177.587.586.180.283.486.6
LTContext (Ours)-77.672.660.177.074.289.487.782.083.287.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.371, + 0.472, + 0.522 + ], + "angle": 0, + "content": "Table 1: Results on the Breakfast and 50Salads datasets. The best and second best results for the methods in the bottom half are shown in bold and underlined since only methods without additional notes are directly comparable. P: additional post-processing; C: additional constraints; DF: different features (DF1: Improved Dense Trajectories (IDT), DF2: Spatio-temporal VGG-style CNN, DF3: Generative Adversarial Network (GAN), DF4: ActionCLIP); TA: test augmentation. In the top half, we highlight results that are better in italic bold formatting." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.533, + 0.466, + 0.61 + ], + "angle": 0, + "content": "
MethodAssembly101
F1@{10,25,50}EditAccParams (M)Inference (sec)
MS-TCN++ [27]31.627.820.630.737.11.080.1
UVAST [4]*32.128.320.831.537.41.223.2
C2F-TCN [42]33.329.021.332.439.26.890.1
ASFormer [61]*33.429.221.430.538.81.132.1
LTContext (Ours)33.930.022.630.441.20.720.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Table 2: Comparison with state-of-the-art methods on the Assembly101 dataset. The best and second best results are shown in bold and underlined. *We trained UVAST [4] and ASFormer [61] using the code of the authors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.472, + 0.903 + ], + "angle": 0, + "content": "quences can be modeled by a directed acyclic graph. This is very efficient on the Breakfast dataset, which has relatively few actions per video compared to the other datasets. However, our approach outperforms DTL for all metrics on 50Salads where the structure of the action sequences is more complex. Furthermore, DTL cannot be applied to datasets like Assembly101 where actions occur more than once in a video. On 50Salads, only Bridge-Prompt [26] performs slightly better for some metrics. Bridge-Prompt proposes an approach for feature learning using a vision-language model such as ActionCLIP [55] in combination with ASFormer. The approach is thus complementary and not comparable to our approach. Nevertheless, our approach achieves higher" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.182 + ], + "angle": 0, + "content": "\\(F1@10\\) and \\(F1@50\\) scores on 50Salads. If we use 10 instead of 9 layers as shown in Table 9, we even outperform Bridge-Prompt for all measures except for frame-wise accuracy. In summary, our approach outperforms all methods in terms of F1 score, which is the most precise measure of the segmentation quality, at all thresholds on 50Salads." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Assembly101 is the largest dataset both in terms of the number of videos and their length, and it is the most challenging dataset. Since ASFormer [61] and UVAST [4] have not been evaluated on this dataset, we trained ASFormer and UVAST on Assembly101 using the publicly available source code and report the results in Table 2 as well. We outperform all methods on Assembly101 in F1 scores. C2F-TCN [42] achieves the best edit score. Since we trained ASFormer [61] and UVAST [4] by ourselves, we can compare the training time on Assembly101. While our approach requires 1 day and 18 hours, ASFormer [61] and UVAST [4] needed 4 weeks and 2 weeks, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.373, + 0.712, + 0.389 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.895, + 0.713 + ], + "angle": 0, + "content": "In Figs. 6-8, we present some qualitative results for the Assembly101, Breakfast, and 50Salads dataset. The first row of each figure shows the middle frame of each ground-truth action segment. In the second, third, and fourth rows, the ground truth segmentation, the predictions of ASFormer [61], and the prediction of our model (LTContext) are shown, respectively. In Fig. 6, our approach shows much fewer errors than ASFormer. Although our approach recognizes all action classes that occur in the video, there are several errors where some instances are missed like the last action segment of the bottom row which corresponds to the action 'detach base'. This indicates how challenging the Assembly101 dataset is. In Fig. 7, the predictions of our model are very close to the ground-truth. ASFormer overestimates the duration of the green segment, which corresponds to the action 'spoon powder', and hallucinates purple segments at the beginning and end of the video. In Fig. 8, both methods estimate the segments well, but ASFormer predicts wrongly two orange segments, which correspond to the action 'mix dressing', and the olive segment is too short, which corresponds to the action 'add salt'." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.665, + 0.738 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.894, + 0.793 + ], + "angle": 0, + "content": "We finally evaluate the impact of each component of our architecture. For the ablation studies, we report the results averaged over the 5 splits of the 50Salads dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Impact of attention types In Table 3, we show the impact of using the combination of windowed and long-term context (LTContext) attention in the LTContext block illustrated in Fig. 3. We first compare it to two variants where we use only windowed or only long-term context attention. In order to keep the number of parameters the same, we still use" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10356" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.895, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 6: Qualitative results on Assembly101. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext. It can be best viewed by using the zoom function of a PDF viewer." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.373, + 0.891, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.518, + 0.895, + 0.551 + ], + "angle": 0, + "content": "Figure 7: Qualitative results on Breakfast. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.573, + 0.468, + 0.651 + ], + "angle": 0, + "content": "
Model ArchitectureF1@{10, 25, 50}EditAcc
LTContext89.487.782.083.287.7
- Windowed Attention87.785.177.482.185.5
- LTContext Attention84.182.674.178.285.3
WA (S1) + LTContext (S2-S4)87.885.079.481.485.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.661, + 0.472, + 0.781 + ], + "angle": 0, + "content": "Table 3: Impact of using LTContext and Windowed Attention (WA) on 50Salads. In the case of Windowed Attention, we use two windowed attention blocks instead of a combination of windowed and LTContext attention. In the case of LTContext Attention, we use two LTContext attention blocks. The last row corresponds to using only windowed attention in stage 1 and only LTContext attention in stages 2 to 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.47, + 0.869 + ], + "angle": 0, + "content": "in these cases two attention blocks within LTContext. The results show that combining both types of attention leads to better results. In particular, the F1@50 score is substantially higher." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 3, we use four LTContext blocks. We also evaluate what happens if we vary the attention not within" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.895, + 0.682 + ], + "angle": 0, + "content": "an LTContext block but between the four LTContext blocks. For this, we use only windowed attention for the first LTContext block and only long-term context attention for the other three LTContext blocks. The last row in Table 3 shows that this does not perform better than using only windowed attention and it is worse than combining windowed and long-term context attention within an LTContext block." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Impact of different values of \\( W \\) and \\( G \\) The parameter \\( W \\) controls the size of the local window for the local attention and the parameter \\( G \\) controls the sparseness of the global attention. If not otherwise specified, we use \\( W = G = 64 \\) in our experiments. The results in Table 4 show that the F1 score drops when the size of the local window \\( W \\) becomes smaller. Note that larger values of \\( W \\) increase the memory and computational cost. When we decrease \\( G \\), the F1 score also drops but not as drastic as for \\( W \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Impact of overlaps for windowed attention As described in Section 3.1, we use an overlap for the keys and values for the windowed attention. In Table 5, we also report the result" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.09, + 0.852, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Figure 8: Qualitative results on 50Salads. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.279, + 0.428, + 0.414 + ], + "angle": 0, + "content": "
WGF1@{10, 25, 50}EditAcc
86483.882.174.577.087.0
166485.482.977.177.187.0
326487.685.879.881.586.6
646489.487.782.083.287.7
643288.887.281.383.287.7
641688.687.080.382.686.2
64888.687.179.982.986.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.424, + 0.465, + 0.44 + ], + "angle": 0, + "content": "Table 4: Impact of the parameters \\( W \\) and \\( G \\) on 50 Salads." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.451, + 0.438, + 0.515 + ], + "angle": 0, + "content": "
F1@{10, 25, 50}EditAcc
Non-Overlapping87.985.579.181.485.1
Overlap (2 Windows)89.487.782.083.287.7
Overlap (3 Windows)89.288.080.883.286.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Table 5: Impact of using an overlap for the keys and values for the local attention on 50Salads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.473, + 0.701 + ], + "angle": 0, + "content": "when the keys and values do not overlap, i.e., they are the same as the queries. It is interesting to note that our attention can be interpreted as a combination of a reshaping operation with axial attention [16] in the special case without overlap. The results, however, show that an overlap improves the results. Using a larger overlap where the keys and values consist of three consecutive windows does not improve the results further." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.471, + 0.834 + ], + "angle": 0, + "content": "Impact of the order of the attention The LTContext block shown in Fig. 3 uses first windowed attention and then long-term context attention. Table 6 shows the results when we change the order of windowed and long-term context attention within the LTContext block. If the order is changed, the performance decreases. Since the LTContext blocks are repeated, the drop in performance is moderate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Impact of the cross-attention As described in Section 3.2, we use cross-attention in stages 2 to 4. Table 7 shows that the performance drastically decreases without cross-attention." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.28, + 0.871, + 0.34 + ], + "angle": 0, + "content": "
F1@{10, 25, 50}EditAcc
WA-LTContext89.487.782.083.287.7
LTContext-WA88.386.480.481.686.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.349, + 0.895, + 0.396 + ], + "angle": 0, + "content": "Table 6: Impact of the attention order within the LTContext block on 50Salads. Our model computes first windowed attention (WA) and then long-term context attention (row 1)." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.406, + 0.871, + 0.458 + ], + "angle": 0, + "content": "
F1@{10, 25, 50}EditAcc
Without Cross-Attention82.278.570.473.680.2
LTContext89.487.782.083.287.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.468, + 0.864, + 0.484 + ], + "angle": 0, + "content": "Table 7: Impact of the cross-attention on 50Salads." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.494, + 0.871, + 0.56 + ], + "angle": 0, + "content": "
F1@{10, 25, 50}EditAcc
Conv with Dilation89.487.782.083.287.7
Conv without Dilation80.178.067.772.679.3
Without Conv81.879.267.172.778.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.57, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Table 8: Impact of using LTContext blocks with 1D convolution but without dilation (row 2) and using LTContext blocks without 1D convolution (row 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.897, + 0.76 + ], + "angle": 0, + "content": "Impact of using convolutions The LTContext block shown in Fig. 3 starts with a dilated 1D convolution with kernel size 3. In Table 8, we evaluate the impact of the dilated convolution by comparing to a LTContext block that uses a convolutional kernel of the same size but without dilation factor, and a LTContext block without 1D convolution. The results show that dilated convolutions have a very high impact on the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Impact of the number of layers As shown in Fig. 3, we repeat the LTContext blocks at each stage \\( N \\) times. We used \\( N = 9 \\) layers in all experiments. Table 9 shows the impact of varying the number of layers. On 50Salads, all measures improve by increasing \\( N \\). It needs to be noted that we use \\( N = 9 \\) for the results reported in Table 1 although we can get even better results with 10 layers on 50 Salads. For the Breakfast and Assembly101 dataset, the best performance is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10358" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.447, + 0.294 + ], + "angle": 0, + "content": "
number of layers (N)50Salads
F1@{10, 25, 50}EditAcc
887.285.177.680.885.7
989.487.782.083.287.7
1089.588.182.484.187.7
Breakfast
875.770.557.774.573.1
977.672.660.177.074.2
1077.372.459.776.473.5
Assembly101
831.928.421.327.841.0
933.930.022.630.441.2
1032.629.321.928.741.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.306, + 0.472, + 0.339 + ], + "angle": 0, + "content": "Table 9: Impact of the number of layers on 50Salads, Breakfast, and Assembly101." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.345, + 0.447, + 0.432 + ], + "angle": 0, + "content": "
number of headsF1@{10, 25, 50}EditAcc
189.487.782.083.287.7
288.887.080.582.987.0
489.287.180.883.186.9
888.286.480.582.186.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.442, + 0.441, + 0.459 + ], + "angle": 0, + "content": "Table 10: Impact of the number of attention heads." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.479, + 0.234, + 0.495 + ], + "angle": 0, + "content": "achieved with 9 layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.472, + 0.68 + ], + "angle": 0, + "content": "Impact of the number of attention heads In our implementation, we do not use multiple attention heads. Nevertheless, we evaluated the impact of using multiple heads in Table 10 since most transformers use multiple heads. The results, however, show that there is no benefit in using multiple heads. The results are consistent with the observations in [33, 30]. For example, [33] shows that many attention heads can often be reduced to a single head without impacting the performance. They also argue that some tasks are more reliant on multiple heads than others. Temporal action segmentation seems to be a task where one head is sufficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Impact of the number of stages As shown in Fig. 3, we use four stages of LTContext blocks, each of them with 9 layers. We evaluate the impact of the number of stages in Table 11. As can be seen, using multiple stages helps to reduce the over-segmentation error and improves the F1 score and Edit score significantly compared to using only one stage. Increasing the number of stages up to 4 improves all metrics, but using 5 stages decreases the performance and the network starts to overfit." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.196, + 0.862 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this work, we addressed the question of how much temporal long-term context is needed for action segmentation." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.089, + 0.86, + 0.183 + ], + "angle": 0, + "content": "
number of stagesF1@{10, 25, 50}EditAcc
156.354.149.945.684.7
286.284.578.280.586.7
387.685.778.381.285.5
489.487.782.083.287.7
589.086.880.083.085.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.194, + 0.875, + 0.21 + ], + "angle": 0, + "content": "Table 11: Impact of the number of stages on 50Salads." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Our analysis indicates that allowing networks to operate on the full input sequence is more beneficial compared to the case where the model has only access to a subset of the input. Based on our analysis, we presented LTContext, an approach for temporal action segmentation, where we leverage sparse attention to capture the long-term context of a video and windowed attention to model the local information in the neighboring frames. Our approach achieves state-of-the-art segmental F1 scores on the 50Salads and Assembly101 datasets, which contain long videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.661, + 0.41 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), the project iBehave (receiving funding from the programme \"Netzwerke 2021\", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.567, + 0.597, + 0.582 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.592, + 0.895, + 0.647 + ], + "angle": 0, + "content": "[1] Yazan Abu Farha and Juergen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.649, + 0.895, + 0.703 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. ViViT: A video vision transformer. In IEEE International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.706, + 0.895, + 0.746 + ], + "angle": 0, + "content": "[3] Nicolas Aziere and Sinisa Todorovic. Multistage temporal convolution transformer for action segmentation. Image and Vision Computing, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.748, + 0.895, + 0.815 + ], + "angle": 0, + "content": "[4] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.818, + 0.895, + 0.845 + ], + "angle": 0, + "content": "[5] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document. arXiv preprint arXiv:2004.05150, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.847, + 0.895, + 0.899 + ], + "angle": 0, + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.592, + 0.895, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10359" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[7] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.15, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[8] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.207, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[9] Zihang Dai, Hanxiao Liu, Quoc Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems (NeurIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[10] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics (ACL), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.43 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[12] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.472, + 0.531 + ], + "angle": 0, + "content": "[13] Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Fine-grained action segmentation using the semi-supervised action gan. Pattern Recognition, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[15] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. Conformer: Convolution-augmented transformer for speech recognition. In INTERSPEECH, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[17] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[18] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[19] Svebor Karaman, Lorenzo Seidenari, and Alberto Del Bimbo. Fast saliency based pooling of fisher encoded dense trajectories. In European Conference on Computer Vision (ECCV) Workshops, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[20] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 780-787, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[22] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[23] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[24] Colin Lea, Austin Reiter, René Vidal, and Gregory D Hager. Segmental spatiotemporal cnns for fine-grained action segmentation. In European Conference on Computer Vision (ECCV), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.345, + 0.894, + 0.397 + ], + "angle": 0, + "content": "[25] Peng Lei and Sinisa Todorovic. Temporal deformable residual networks for action segmentation in videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.467 + ], + "angle": 0, + "content": "[26] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[27] Shi-Jie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. MS-TCN++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.894, + 0.636 + ], + "angle": 0, + "content": "[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.894, + 0.677 + ], + "angle": 0, + "content": "[30] Liyuan Liu, Jialu Liu, and Jiawei Han. Multi-head or singlehead? an empirical comparison for transformer training. arXiv, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.747 + ], + "angle": 0, + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[32] Zhichao Liu, Leshan Wang, Desen Zhou, Jian Wang, Songyang Zhang, Yang Bai, Errui Ding, and Rui Fan. Temporal segment transformer for action segmentation. arXiv preprint arXiv:2302.13074, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[33] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems (NeurIPS), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[34] Piotr Nawrot, Szymon Tworkowski, Michal Tyrolski, Lukasz Kaiser, Yuhuai Wu, Christian Szegedy, and Henryk Michalewski. Hierarchical transformers are more efficient language models. arXiv preprint arXiv:2110.13711, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10360" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[35] Junyong Park, Daekyum Kim, Sejoon Huh, and Sungho Jo. Maximization and restoration: Action segmentation through dilation passing and temporal reconstruction. Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning (ICML), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.472, + 0.277 + ], + "angle": 0, + "content": "[37] Yifan Peng, Siddharth Dalmia, Ian Lane, and Shinji Watanabe. Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding. In International Conference on Machine Learning (ICML), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.472, + 0.336 + ], + "angle": 0, + "content": "[38] Jiezhong Qiu, Hao Ma, Omer Levy, Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. Findings of Empirical Methods in Natural Language Processing (EMNLP), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.472, + 0.379 + ], + "angle": 0, + "content": "[39] Alexander Richard and Juergen Gall. Temporal action detection using a statistical language model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.472, + 0.436 + ], + "angle": 0, + "content": "[40] Marcus Rohrbach, Sikandar Amin, Mykhaylo Andriluka, and Bernt Schiele. A database for fine grained activity detection of cooking activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.472, + 0.508 + ], + "angle": 0, + "content": "[41] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[42] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. arXiv preprint, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.472, + 0.61 + ], + "angle": 0, + "content": "[43] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fifa: Fast inference approximation for action segmentation. In DAGM German Conference on Pattern Recognition (GCR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.472, + 0.668 + ], + "angle": 0, + "content": "[44] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.472, + 0.738 + ], + "angle": 0, + "content": "[45] Sebastian Stein and Stephen J. McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In ACM International Joint Conference on Pervasive and Ubiquitous Computing, page 729-738, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.472, + 0.796 + ], + "angle": 0, + "content": "[46] Kevin Tang, Li Fei-Fei, and Daphne Koller. Learning latent temporal structure for complex event detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.8, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[47] Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. ACM Comput. Surv., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[48] Xiaoyan Tian, Ye Jin, and Xianglong Tang. Local-global transformer neural network for temporal action segmentation. Multimedia Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[50] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.894, + 0.246 + ], + "angle": 0, + "content": "[51] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[53] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. AAAI Conference on Artificial Intelligence, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.894, + 0.402 + ], + "angle": 0, + "content": "[54] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.894, + 0.443 + ], + "angle": 0, + "content": "[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.894, + 0.5 + ], + "angle": 0, + "content": "[56] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.894, + 0.557 + ], + "angle": 0, + "content": "[57] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In IEEE International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.558, + 0.894, + 0.613 + ], + "angle": 0, + "content": "[58] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.894, + 0.683 + ], + "angle": 0, + "content": "[59] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.685, + 0.894, + 0.739 + ], + "angle": 0, + "content": "[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.741, + 0.894, + 0.781 + ], + "angle": 0, + "content": "[61] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In British Machine Vision Conference (BMVC), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.784, + 0.894, + 0.838 + ], + "angle": 0, + "content": "[62] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.84, + 0.894, + 0.894 + ], + "angle": 0, + "content": "[63] Long Zhao, Zizhao Zhang, Ting Chen, and Dimitris Metaxas abd Han Zhang. Improved transformer for high-resolution GANs. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10361" + } + ] +] \ No newline at end of file diff --git a/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_origin.pdf b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4b0abb7c739243919f150fa4149e962ba77b7b9f --- /dev/null +++ b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/94b447e5-f23e-4bb3-8d5c-681df55d2119_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c086f88cf0d4bb7b806da5f49d39e2e5db638efc7bff6cf897e275f53a7ff81 +size 5891738 diff --git a/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/full.md b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e987cbd3083c72c07685f0f9e465a225d7ef6268 --- /dev/null +++ b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/full.md @@ -0,0 +1,307 @@ +# How Much Temporal Long-Term Context is Needed for Action Segmentation? + +Emad Bahrami + +Gianpiero Francesca + +Juergen Gall1,3 + +1University of Bonn, Germany + +$^{2}$ Toyota Motor Europe, Belgium + +$^{3}$ Lamarr Institute for Machine Learning and Artificial Intelligence, Germany + +# Abstract + +Modeling long-term context in videos is crucial for many fine-grained tasks including temporal action segmentation. An interesting question that is still open is how much long-term temporal context is needed for optimal performance. While transformers can model the long-term context of a video, this becomes computationally prohibitive for long videos. Recent works on temporal action segmentation thus combine temporal convolutional networks with self-attention that are computed only for a local temporal window. While these approaches show good results, their performance is limited by their inability to capture the full context of a video. In this work, we try to answer how much long-term temporal context is required for temporal action segmentation by introducing a transformer-based model that leverages sparse attention to capture the full context of a video. We compare our model with the current state of the art on three datasets for temporal action segmentation, namely 50Salads, Breakfast, and Assembly101. Our experiments show that modeling the full context of a video is necessary to obtain the best performance for temporal action segmentation. + +# 1. Introduction + +Temporal action segmentation can be used in many real-world applications such as monitoring production lines or studying animal behavior. In these settings, the videos can be very long and it is required to recognize the start and end of all actions that occur in a video as illustrated in Fig. 1. + +Recently, combinations of temporal convolutional networks [1, 27] with self- and cross-attention from transformers [61, 4] have shown impressive results for temporal action segmentation. These works are in line with other hybrid models [15, 9, 61, 58] that combine the attention modules with convolutions to compensate for the lack of strong inductive bias of pure transformers. However, the emergence of datasets like Assembly101 [41], where subjects perform assembly tasks, poses a new challenge in the area of temporal action segmentation due to the existence of long videos + +![](images/89c6531ce91098468580331fe78d56c67f734046b049397c4eb0f2690eb7c0e4.jpg) +Figure 1: Datasets like Assembly101 contain long videos of assembly tasks and an action label needs to be predicted for each frame. The first row shows some frames of a video. The second row shows the ground-truth labels for all frames of the video where different colors correspond to different action labels. Rows 3-6 show the predictions of the proposed model for different amounts of long-term context where $100\%$ means the temporal context of the full video. + +that can last up to 25 minutes. Since modeling the long-term context of such a video is very expensive, Yi et al. [61] proposed to compute the attention for a local temporal window. In order to understand the impact of the window size on the temporal action segmentation accuracy, we analyze the impact of the window size on two datasets with long videos in Section 4.1. Fig. 1 shows some qualitative results of this study on Assembly101. Indeed, the results show that modeling the long-term context of an entire video is very important for temporal action segmentation. + +Based on this finding, we revisit how temporal attention is modeled in transformer architectures for temporal action segmentation. [61, 4] use a hierarchy of temporal windows, which makes the training on long video sequences as they occur in Assembly101 [41] very expensive. Inspired by works that decompose attention over the spatial and temporal domain for short video clips [6, 62], we propose to iterate between computing windowed local attention and sparse long-term context attention such that both short and long-term context are modeled. This approach is particularly suitable for temporal action segmentation since the local attentions focus on the similarity or dissimilarity of features within an action segment or between neighboring action + +segments whereas the long-term context attention focuses on the relations between actions within the entire video. The source code is available at https://github.com/LTContext/LTContext. + +# 2. Related Work + +The traditional sliding window with non-maxima suppression has been among the early approaches for action segmentation [19, 40]. [22, 20, 46] adopted hidden Markov Models (HMMs) for temporal modeling. [39] incorporates length and temporal context and uses dynamic programming for inference. Temporal convolutional networks (TCN) with temporal pooling are used by [23] to classify each video frame. Later, [1] introduced a multi-stage TCN capable of maintaining high temporal resolution, which is necessary for fine-grained recognition. [18] reduces the over-segmentation error by adding an action boundary regression network to refine the frame-wise results. [17] proposed another refinement method based on a graph convolutional network. + +Transformers [52] were originally used for natural language processing and only rely on attention for sequence modeling. Recently, transformers have been widely adopted in vision [11, 57, 31, 54], speech recognition [15, 37], and action recognition [6, 2]. The original transformers suffered from two issues. First, the cost of the self-attention operation is quadratic with respect to the sequence length. This has been addressed by many methods that improve the memory efficiency of transformers [5, 34, 10]. We refer to [47] for a comprehensive survey of efficient transformers. For instance, restricting the attention to a fixed window size [5] is one approach. Multi-axis self-attention [51] combines block-attention with grid-attention, which is based on a spatial grid overlaid on the entire 2D space. Adapting sparse attention for capturing global information has shown to be an effective solution in vision tasks such as high-resolution image generation [63, 50], object detection [51, 60], and instance segmentation [51, 60]. MViT [12] uses a pooling operation to reduce the space-time resolution before attention and MViTv2 [28] improves MViT by adding the relative position along (x,y,t) and residual connections. The second issue of transformers is the poor generalization due to a relatively weak inductive bias [9, 49] compared to convolutional neural networks (CNN). Hybrid models that combine self-attention and convolution layers have thus been proposed for vision tasks [58, 9, 51, 57] and speech recognition [15, 37]. + +In action recognition, works such as [6, 62] adopt the idea of using attention for video understanding. In Times-Former [6], the self-attention is first applied along the temporal dimension for each single patch (time attention), i.e., over all frames of a short video clip of a few seconds. In a second step, the attention is applied over the patches of each frame (space attention). VIDTR [62] also decomposes self-attention into spatial and temporal attention, but additionally + +down-samples the temporal dimension. In our work, we do not apply attention spatially and temporally, but we address the question of how temporal attention can be computed for very long sequences that last 25 minutes as it is required for temporal action segmentation. + +For temporal action segmentation, [61] proposed an architecture, called ASFormer, which is based on a multi-stage TCN [1] and equips the temporal convolutions with a local window attention [5]. This is done in a hierarchy where the size of the local window grows with each layer. Recently, [4] proposed a decoder on top of the encoder that generates action segments in an autoregressive manner. It uses two heads where the head on top of the encoder predicts framewise probabilities and the head of the decoder predicts the sequence of actions. Finally, an alignment decoder fuses the output of the two heads and aligns the predicted sequence of actions to the frames. [3] introduced a hybrid Temporal Convolution Transformer (TCTr) where they adapt an action boundary detector to adaptively estimate attention from local neighboring frames. [59] proposed to use additional constraints during training, but the approach assumes that action sequences can be modelled by a directed acyclic graph, which does not allow that actions occur more than once in a video. Recently, multi-modal approaches that combine language models with vision transformers have been proposed. For instance, [26] uses prompt engineering to extract features from pre-trained vision-language models such as ActionCLIP [55]. + +# 3. Long-Term Context for Action Segmentation + +Recently, transformers combined with temporal convolutional neural networks [61, 4, 3] have shown a very good performance for temporal action segmentation. For this task, the frame-wise labels $c_{1}, \ldots, c_{T}$ , where $c_{t} \in \mathcal{C}$ and $C = |\mathcal{C}|$ denotes the number of action classes, need to be predicted for a given a video $X = (x_{1}, \ldots, x_{T})$ with $T$ frames, where $x_{t}$ represents a feature map of size $D$ at frame $t$ . Since $T$ can be very large, [61, 4, 3] limit the self-attention to a local temporal window. + +In order to understand how much temporal long-term context is needed for temporal action segmentation, we limited the temporal input window and evaluated the quality of the temporal action segmentation in Section 4.1. The results in Fig. 4 show that temporal long-term context has a strong impact on the performance. Based on our analysis, we thus revisit the windowed attention of previous works for temporal action segmentation and propose to model the temporal long-term context of a video using sparse attentions [5, 38, 36]. Additionally, we equip our method with windowed attention to capture the locality between neighboring frames. In this way, we obtain a flexible design that is capable of providing long-term and local temporal context. While we describe first the Long-Term Context (LTCon + +![](images/995082b14912e24ab0b5a9300955eea2b4b6ece7b0df240175baeb26000878e7.jpg) +Figure 2: Illustration of windowed and long-term context attentions with a window of size 2. For the windowed attention, the sequence is partitioned into small windows and the attentions are computed for each window. For long-term context attention, the sequence is reordered such that the attentions are computed over the whole, but sparsely sampled sequence. After the attention, the output is reordered again to preserve the original order. Best viewed in color. + +text) block in Section 3.1, the entire network is described in Section 3.2. + +# 3.1. Temporal Context Attention + +The self-attention blocks of transformers are advantageous over convolutions in the aggregation of global information. However, applying attention to long sequences such as untrimmed videos is impractical due to the quadratic complexity of the self-attention blocks. To address this issue, we adopt an attention mechanism where we leverage sparse and windowed attention to model long-term and local temporal context. + +The attention function transforms the input into a query, key, and value and computes the output as a weighted sum of the values. For example, given a sequence of features, $X\in \mathbb{R}^{T\times D}$ , the attention can be written as follows: + +$$ +A t t e n t i o n (Q, K, V) = \mathrm {s o f t m a x} _ {r o w} \left(\frac {Q K ^ {T}}{\sqrt {D}}\right) V \quad (1) +$$ + +where $Q, K, V \in \mathbb{R}^{T \times D}$ are linearly transformed from $X$ . Since $T$ is very large for long sequences, we need to modify $Q, K, V$ to enable the modeling of long-term and local temporal context as illustrated in Fig. 2. + +Temporal Windowed Attention For the temporal Windowed Attention (WA), we partition the sequence into nonoverlapping windows of size $W$ . Fig. 2 illustrates the case for $W = 2$ , but we use $W = 64$ in practice. The impact of $W$ is evaluated in Section 4. Instead of computing the attention over the entire sequence of length $T$ , we compute the attention $\frac{T}{W}$ times, where each query $Q \in \mathbb{R}^{W \times D}$ corresponds to each window. For the keys and values, we use + +an overlap where we use the next window in addition, i.e., for each query $Q$ we have $K,V\in \mathbb{R}^{2W\times D}$ . We perform masking when $K$ and $V$ exceed the input sequence. We evaluate the impact of the overlap in Section 4. + +Temporal Long-Term Context Attention For the temporal Long-Term Context (LTC) attention, the input is also partitioned into non-overlapping windows of size $G$ . However, instead of computing the attention over each window, the attention is computed over all windows where from each window only one element is taken. In the illustration of Fig. 2 with $G = 2$ , we compute the attention over the first feature of all windows and the attention over the second feature of all windows. In general, we compute the attentions for $G$ queries $Q \in \mathbb{R}^{\frac{T}{G} \times D}$ where the keys and values are the same, i.e., $K, V \in \mathbb{R}^{\frac{T}{G} \times D}$ . The parameter $G$ provides the flexibility to adjust the sparseness based on the available memory budget, e.g., $G = 1$ corresponds to the case where the attention is applied over the full sequence. In practice, we use $G = 64$ and we evaluate the impact of $G$ in Section 4. + +LTContext block The top of Fig. 3 illustrates the entire LTContext block. As in previous works [1, 61, 4], we use a 1D dilated temporal convolution with kernel size 3, where the dilation factor increases by factor 2 for each layer. The dilated temporal convolution is followed by a Gaussian Error Linear Unit (GELU). In the LTContext block, we first use the windowed and then the long-term context attention, which are shown in Fig. 2. We evaluate the impact of the order in Section 4. Finally, we use a linear layer with a residual connection to output the features for each frame, $F \in \mathbb{R}^{T \times D}$ . + +![](images/687ec45d91111480e73d139494f70f7219ea3b1c910f95828a75487ac011f750.jpg) +Figure 3: The network architecture of LTContext with LTContext blocks (top). + +# 3.2. LTContext Architecture + +The entire LTContext network is depicted in Fig. 3. For a fair comparison, we will use the features that are provided for the corresponding datasets as input. In all cases, the dimensionality of the features is 2048. As in previous works, we use a linear layer to reduce the dimensionality of the features to 64. The output of each LTContext block is the feature map $F \in \mathbb{R}^{T \times D}$ . We repeat each LTContext block $N$ times where the dilation factor of the temporal convolution increases in each layer. In practice, we use $N = 9$ and we evaluate the impact of $N$ in Section 4. After the first $N$ layers of LTContext blocks, we use an additional linear layer to reduce the dimensionality $D$ further to 32. The dimensionality reduction reduces the number of parameters from 1.42 million to 0.72 million without reducing the accuracy. We also use an additional linear layer followed by a softmax layer to generate the frame-wise class probabilities $P \in \mathbb{R}^{T \times C}$ . + +We continue with three additional stages where each stage consists of $N$ layers of LTContext blocks. Note that we reset the dilation factor to 1 for the temporal convolution at the beginning of each stage and we compute the frame-wise class probabilities $P \in \mathbb{R}^{T \times C}$ after each stage, which contributes to the multi-stage loss. We use the cross-entropy loss combined with the mean squared error smoothing loss as introduced by [1] and used in [27, 61] for a fair comparison. Inspired by [61], we use the cross-attention for the LTContext blocks in stages 2 to 4. Instead of using the features $F$ for the queries and keys for windowed and long-term context attention, the predictions $P$ are used. We thus have $Q \in \mathbb{R}^{W \times C}$ , $K \in \mathbb{R}^{2W \times C}$ , and $V \in \mathbb{R}^{2W \times D}$ for the windowed attention and $Q \in \mathbb{R}_{G}^{T \times C}$ , $K \in \mathbb{R}_{G}^{T \times C}$ , and $V \in \mathbb{R}_{G}^{\frac{T}{G} \times D}$ for the long-term context attention. While the cross-attention is not shown in Fig. 3, it only means that $P$ is an additional input for the windowed and long-term context attention in stages 2-4. We evaluate the impact of the number + +of stages in Section 4. + +# 4. Experiments + +Datasets. We evaluate the performance of our proposed model on three challenging action segmentation datasets: 50Salads [45], Breakfast [21], and Assembly101 [41]. + +50Salads [45] contains 50 videos annotated with 17 action classes. On average, each video is 6.4 minutes long and has 18 action segments. Following previous works [1, 61, 4], we use five-fold cross-validation and report the average. + +Breakfast contains 1,712 videos of breakfast preparation activities with an average length of 2.3 minutes. There are 48 action classes and each video has on average 6.6 action segments. For evaluation, we report the average of the 4 splits for cross-validation as in [1]. + +Assembly101 [41] is the largest dataset among the three datasets with 4,321 videos and 202 coarse action classes composed of 11 verbs and 61 objects. Assembly101 is a procedural activity dataset containing videos of people assembling and disassembling 101 toy vehicles. On average, each video includes 24 action segments and is 7.1 minutes long. Compared to Breakfast, Assembly101 has 2.5 times more videos, 6.7 times more hours of video footage, 9.3 times more action segments, and 4.2 more action classes. For our evaluation, we follow the setting for temporal action segmentation [41] and report the results on the validation set. + +For a fair comparison, we use the features that are provided for the datasets and that have been used in previous works. We use the I3D [7] features for the 50Salads and Breakfast datasets and TSM [29] features for the Assembly101 dataset [41]. Both features are 2048 dimensional vectors. Following [1], we also used the temporally downsampled features for 50Salads to compensate for the different frame-rates of the datasets. + +Evaluation Metrics. We report the frame-wise accuracy (Acc), segmental Edit distance, and segmental F1 score at + +the overlapping thresholds of $10\%$ , $25\%$ , and $50\%$ denoted by $F1@{10,25,50}$ . The intersection over the union (IoU) ratio is used as the overlapping threshold. The edit distance measures only the order of the actions but not the duration. The frame-wise accuracy measures the accuracy per frame. It is dominated by actions that have a long duration and it is not very sensitive to over-segmentation errors that occur when a few frames within a ground-truth action segment are wrongly classified. The F1 score is the most reliable measure. + +# 4.1. How much temporal long-term context is needed? + +We first present the results of our analysis on the impact of using the full sequence as input compared to using a temporal window. The goal of this experiment is to shed light on how much temporal long-term context is needed for the task of temporal action segmentation. For the analysis, we use only 50Salads and Assembly101 since the videos in Breakfast are too short. For the experiments, we train our approach (LTContext) and ASFormer [61] either on the full video sequences or we divide the videos into shorter sequences where the full context is lacking. + +Fig. 4 shows the result of this experiment on the 50Salads and Assembly101 dataset. We report the window size as percentage of the average length of a video in the corresponding dataset and $100\%$ means that the full sequence has been used. The results clearly show that the full context of an entire sequence is advantageous over allowing the model to see only a window of the input sequence even if the window is large $(50\%)$ . We can also observe that our approach benefits more from the full sequence than ASFormer. + +We furthermore evaluated whether the impact of the window size is stronger for longer videos. To this end, we sorted all test videos into four quarters by their length. Fig. 5 (left) shows that the difference between the window size $50\%$ and the full video $(100\%)$ is larger for long videos. This shows that long-term context is in particular for long videos important. We also evaluated whether choosing a window for each video instead of choosing a window based on the average video length (fixed) performs better. For the video-specific window size, the window size is set to the percentage of each video. Fig. 5 (right) shows that a video-specific window size performs much worse than a fixed window. Varying the amount of context for each video is thus not beneficial. Fig. 1 shows qualitative results of our approach for different amounts of temporal context with a fixed window for a video from the Assembly101 dataset. + +# 4.2. Comparison with State of the Art + +We present the performance comparison of our method with state-of-the-art methods on the Breakfast and 50Salads datasets in Table 1 and on the Assembly101 dataset [41] in + +![](images/0765e105ea9b7ae9c71a297cc6b7c8095f7e31fd26afcbb752ddeb183343379a.jpg) +Figure 4: Impact of different sizes of the input window on the 50Salads dataset (left) and the Assembly101 dataset (right). The window size is given in percentage of the average length of a video in the corresponding dataset. $100\%$ denotes entire video. + +![](images/7f87ca3a0883076c899867d20007ef72ceada82d0d70be014eb62520dbfe1972.jpg) + +![](images/1e3e0ed7221c707ec6f2d85ffa1433634f4307174c50be9c8d55495d3e5fe275.jpg) +Figure 5: Results for different video lengths (rights). Comparison of a fixed with a video-specific input window size (left). Both plots are for 50Salads. + +![](images/e6986f0237908ffdfa23405ce3cd8d13ce60fb069613063cc94d672650b5a5c9.jpg) + +# Table 2. + +On Breakfast and 50Salads, our method outperforms all comparable methods in terms of F1 score at all thresholds, which is the most important measure. Our method also achieves a better Edit score than all methods except for UVAST [4] and TCTr [3]. UVAST uses ASFormer [61] as encoder and an additional alignment decoder. The higher Edit score of UVAST is expected since the approach has an additional head that predicts the sequence of actions and thus maximizes the Edit score by an additional loss. This, however, comes at the cost of a much lower frame accuracy. TCTr [3] achieves a higher frame-wise accuracy on Breakfast and a higher Edit score on 50 Salads, but it performs worse for the other metrics. The approach uses a boundary detection module and compresses the temporal features, which is a complementary approach. It needs to be noted that our approach outperforms TCTr for all metrics on 50Salads if we use 10 instead of 9 layers as shown in Table 9. However, 9 layers perform better than 10 layers for the other datasets. We therefore report in Tables 1 and 2 only the results with 9 layers. + +We also compare to approaches that use different input features, perform additional test augmentation, or add additional constraints. On Breakfast, only DTL [59] reports better results. DTL [59] uses additional logic-based constraints for training ASFormer and assumes that action se + +
MethodNoteBreakfast50Salads
F1@{10,25,50}EditAccF1@{10,25,50}EditAcc
IDT+LM [39]DF1-----44.438.927.845.848.7
ST-CNN [24]DF2-----55.949.637.145.959.4
ED-TCN [23]DF2-----68.063.952.659.864.7
TDRN [25]DF2-----72.968.557.266.068.1
SSA-GAN [13]DF3----43.374.971.767.069.873.3
Bridge-Prompt [26]DF4-----89.287.881.383.888.1
C2F-TCN [42]TA72.268.757.669.676.084.381.872.676.484.9
UVAST [4]+ViterbiP75.970.057.276.566.089.187.681.783.987.4
UVAST [4]+FIFA [43]P76.971.558.077.169.788.987.078.583.984.5
Liu [32] + ASRF [18]P77.572.359.576.773.787.986.680.582.786.6
DTL [59]C78.874.562.977.775.887.185.778.580.586.9
MS-TCN [1]-52.648.137.961.766.376.374.064.567.980.7
MS-TCN++ [27]-64.158.645.965.667.680.778.570.174.383.7
DTGRM [53]-68.761.946.668.968.379.175.966.172.080.0
MuCon [44]-73.266.148.476.362.8-----
Gao et al. [14]-74.969.055.273.370.780.378.069.873.482.2
BCN [56]-68.765.555.066.270.482.381.374.074.384.4
SSTDA [8]-75.069.155.273.770.283.081.573.875.882.2
C2F-TCN [42]-70.166.656.268.273.576.673.062.569.280.1
ASRF [18]-74.368.956.172.467.684.983.577.379.384.5
UVAST [4]-76.770.056.677.268.286.281.270.483.979.5
DPRN [35]-75.670.557.675.171.787.886.379.482.087.2
LGTNN [48]-76.271.557.575.272.587.586.279.882.086.1
ASFormer [61]-76.070.657.475.073.585.183.476.079.685.6
TCTr [3]-76.671.158.576.177.587.586.180.283.486.6
LTContext (Ours)-77.672.660.177.074.289.487.782.083.287.7
+ +Table 1: Results on the Breakfast and 50Salads datasets. The best and second best results for the methods in the bottom half are shown in bold and underlined since only methods without additional notes are directly comparable. P: additional post-processing; C: additional constraints; DF: different features (DF1: Improved Dense Trajectories (IDT), DF2: Spatio-temporal VGG-style CNN, DF3: Generative Adversarial Network (GAN), DF4: ActionCLIP); TA: test augmentation. In the top half, we highlight results that are better in italic bold formatting. + +
MethodAssembly101
F1@{10,25,50}EditAccParams (M)Inference (sec)
MS-TCN++ [27]31.627.820.630.737.11.080.1
UVAST [4]*32.128.320.831.537.41.223.2
C2F-TCN [42]33.329.021.332.439.26.890.1
ASFormer [61]*33.429.221.430.538.81.132.1
LTContext (Ours)33.930.022.630.441.20.720.6
+ +Table 2: Comparison with state-of-the-art methods on the Assembly101 dataset. The best and second best results are shown in bold and underlined. *We trained UVAST [4] and ASFormer [61] using the code of the authors. + +quences can be modeled by a directed acyclic graph. This is very efficient on the Breakfast dataset, which has relatively few actions per video compared to the other datasets. However, our approach outperforms DTL for all metrics on 50Salads where the structure of the action sequences is more complex. Furthermore, DTL cannot be applied to datasets like Assembly101 where actions occur more than once in a video. On 50Salads, only Bridge-Prompt [26] performs slightly better for some metrics. Bridge-Prompt proposes an approach for feature learning using a vision-language model such as ActionCLIP [55] in combination with ASFormer. The approach is thus complementary and not comparable to our approach. Nevertheless, our approach achieves higher + +$F1@10$ and $F1@50$ scores on 50Salads. If we use 10 instead of 9 layers as shown in Table 9, we even outperform Bridge-Prompt for all measures except for frame-wise accuracy. In summary, our approach outperforms all methods in terms of F1 score, which is the most precise measure of the segmentation quality, at all thresholds on 50Salads. + +Assembly101 is the largest dataset both in terms of the number of videos and their length, and it is the most challenging dataset. Since ASFormer [61] and UVAST [4] have not been evaluated on this dataset, we trained ASFormer and UVAST on Assembly101 using the publicly available source code and report the results in Table 2 as well. We outperform all methods on Assembly101 in F1 scores. C2F-TCN [42] achieves the best edit score. Since we trained ASFormer [61] and UVAST [4] by ourselves, we can compare the training time on Assembly101. While our approach requires 1 day and 18 hours, ASFormer [61] and UVAST [4] needed 4 weeks and 2 weeks, respectively. + +# 4.3. Qualitative Evaluation + +In Figs. 6-8, we present some qualitative results for the Assembly101, Breakfast, and 50Salads dataset. The first row of each figure shows the middle frame of each ground-truth action segment. In the second, third, and fourth rows, the ground truth segmentation, the predictions of ASFormer [61], and the prediction of our model (LTContext) are shown, respectively. In Fig. 6, our approach shows much fewer errors than ASFormer. Although our approach recognizes all action classes that occur in the video, there are several errors where some instances are missed like the last action segment of the bottom row which corresponds to the action 'detach base'. This indicates how challenging the Assembly101 dataset is. In Fig. 7, the predictions of our model are very close to the ground-truth. ASFormer overestimates the duration of the green segment, which corresponds to the action 'spoon powder', and hallucinates purple segments at the beginning and end of the video. In Fig. 8, both methods estimate the segments well, but ASFormer predicts wrongly two orange segments, which correspond to the action 'mix dressing', and the olive segment is too short, which corresponds to the action 'add salt'. + +# 4.4. Ablation Studies + +We finally evaluate the impact of each component of our architecture. For the ablation studies, we report the results averaged over the 5 splits of the 50Salads dataset. + +Impact of attention types In Table 3, we show the impact of using the combination of windowed and long-term context (LTContext) attention in the LTContext block illustrated in Fig. 3. We first compare it to two variants where we use only windowed or only long-term context attention. In order to keep the number of parameters the same, we still use + +![](images/19a37f9fa2afed00f320001205ad8a91eb6b1dbd4db0d8d1396c9684d254007a.jpg) +Figure 6: Qualitative results on Assembly101. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext. It can be best viewed by using the zoom function of a PDF viewer. + +![](images/4a5e768fa14b18d64125e296077d1579173596435439661b9103cb8d1525804b.jpg) +Figure 7: Qualitative results on Breakfast. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext. + +
Model ArchitectureF1@{10, 25, 50}EditAcc
LTContext89.487.782.083.287.7
- Windowed Attention87.785.177.482.185.5
- LTContext Attention84.182.674.178.285.3
WA (S1) + LTContext (S2-S4)87.885.079.481.485.2
+ +Table 3: Impact of using LTContext and Windowed Attention (WA) on 50Salads. In the case of Windowed Attention, we use two windowed attention blocks instead of a combination of windowed and LTContext attention. In the case of LTContext Attention, we use two LTContext attention blocks. The last row corresponds to using only windowed attention in stage 1 and only LTContext attention in stages 2 to 4. + +in these cases two attention blocks within LTContext. The results show that combining both types of attention leads to better results. In particular, the F1@50 score is substantially higher. + +As shown in Fig. 3, we use four LTContext blocks. We also evaluate what happens if we vary the attention not within + +an LTContext block but between the four LTContext blocks. For this, we use only windowed attention for the first LTContext block and only long-term context attention for the other three LTContext blocks. The last row in Table 3 shows that this does not perform better than using only windowed attention and it is worse than combining windowed and long-term context attention within an LTContext block. + +Impact of different values of $W$ and $G$ The parameter $W$ controls the size of the local window for the local attention and the parameter $G$ controls the sparseness of the global attention. If not otherwise specified, we use $W = G = 64$ in our experiments. The results in Table 4 show that the F1 score drops when the size of the local window $W$ becomes smaller. Note that larger values of $W$ increase the memory and computational cost. When we decrease $G$ , the F1 score also drops but not as drastic as for $W$ . + +Impact of overlaps for windowed attention As described in Section 3.1, we use an overlap for the keys and values for the windowed attention. In Table 5, we also report the result + +![](images/d3b4b14841ba7291dc5af17a582d79e4b482f90702d664c41c88792dee8eec74.jpg) +Figure 8: Qualitative results on 50Salads. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext. + +
WGF1@{10, 25, 50}EditAcc
86483.882.174.577.087.0
166485.482.977.177.187.0
326487.685.879.881.586.6
646489.487.782.083.287.7
643288.887.281.383.287.7
641688.687.080.382.686.2
64888.687.179.982.986.6
+ +Table 4: Impact of the parameters $W$ and $G$ on 50 Salads. + +
F1@{10, 25, 50}EditAcc
Non-Overlapping87.985.579.181.485.1
Overlap (2 Windows)89.487.782.083.287.7
Overlap (3 Windows)89.288.080.883.286.6
+ +when the keys and values do not overlap, i.e., they are the same as the queries. It is interesting to note that our attention can be interpreted as a combination of a reshaping operation with axial attention [16] in the special case without overlap. The results, however, show that an overlap improves the results. Using a larger overlap where the keys and values consist of three consecutive windows does not improve the results further. + +Impact of the order of the attention The LTContext block shown in Fig. 3 uses first windowed attention and then long-term context attention. Table 6 shows the results when we change the order of windowed and long-term context attention within the LTContext block. If the order is changed, the performance decreases. Since the LTContext blocks are repeated, the drop in performance is moderate. + +Impact of the cross-attention As described in Section 3.2, we use cross-attention in stages 2 to 4. Table 7 shows that the performance drastically decreases without cross-attention. + +Table 5: Impact of using an overlap for the keys and values for the local attention on 50Salads. + +
F1@{10, 25, 50}EditAcc
WA-LTContext89.487.782.083.287.7
LTContext-WA88.386.480.481.686.1
+ +Table 6: Impact of the attention order within the LTContext block on 50Salads. Our model computes first windowed attention (WA) and then long-term context attention (row 1). + +
F1@{10, 25, 50}EditAcc
Without Cross-Attention82.278.570.473.680.2
LTContext89.487.782.083.287.7
+ +Table 7: Impact of the cross-attention on 50Salads. + +
F1@{10, 25, 50}EditAcc
Conv with Dilation89.487.782.083.287.7
Conv without Dilation80.178.067.772.679.3
Without Conv81.879.267.172.778.9
+ +Table 8: Impact of using LTContext blocks with 1D convolution but without dilation (row 2) and using LTContext blocks without 1D convolution (row 3). + +Impact of using convolutions The LTContext block shown in Fig. 3 starts with a dilated 1D convolution with kernel size 3. In Table 8, we evaluate the impact of the dilated convolution by comparing to a LTContext block that uses a convolutional kernel of the same size but without dilation factor, and a LTContext block without 1D convolution. The results show that dilated convolutions have a very high impact on the performance. + +Impact of the number of layers As shown in Fig. 3, we repeat the LTContext blocks at each stage $N$ times. We used $N = 9$ layers in all experiments. Table 9 shows the impact of varying the number of layers. On 50Salads, all measures improve by increasing $N$ . It needs to be noted that we use $N = 9$ for the results reported in Table 1 although we can get even better results with 10 layers on 50 Salads. For the Breakfast and Assembly101 dataset, the best performance is + +
number of layers (N)50Salads
F1@{10, 25, 50}EditAcc
887.285.177.680.885.7
989.487.782.083.287.7
1089.588.182.484.187.7
Breakfast
875.770.557.774.573.1
977.672.660.177.074.2
1077.372.459.776.473.5
Assembly101
831.928.421.327.841.0
933.930.022.630.441.2
1032.629.321.928.741.5
+ +Table 9: Impact of the number of layers on 50Salads, Breakfast, and Assembly101. + +
number of headsF1@{10, 25, 50}EditAcc
189.487.782.083.287.7
288.887.080.582.987.0
489.287.180.883.186.9
888.286.480.582.186.1
+ +achieved with 9 layers. + +Impact of the number of attention heads In our implementation, we do not use multiple attention heads. Nevertheless, we evaluated the impact of using multiple heads in Table 10 since most transformers use multiple heads. The results, however, show that there is no benefit in using multiple heads. The results are consistent with the observations in [33, 30]. For example, [33] shows that many attention heads can often be reduced to a single head without impacting the performance. They also argue that some tasks are more reliant on multiple heads than others. Temporal action segmentation seems to be a task where one head is sufficient. + +Impact of the number of stages As shown in Fig. 3, we use four stages of LTContext blocks, each of them with 9 layers. We evaluate the impact of the number of stages in Table 11. As can be seen, using multiple stages helps to reduce the over-segmentation error and improves the F1 score and Edit score significantly compared to using only one stage. Increasing the number of stages up to 4 improves all metrics, but using 5 stages decreases the performance and the network starts to overfit. + +# 5. Conclusion + +In this work, we addressed the question of how much temporal long-term context is needed for action segmentation. + +Table 10: Impact of the number of attention heads. + +
number of stagesF1@{10, 25, 50}EditAcc
156.354.149.945.684.7
286.284.578.280.586.7
387.685.778.381.285.5
489.487.782.083.287.7
589.086.880.083.085.7
+ +Table 11: Impact of the number of stages on 50Salads. + +Our analysis indicates that allowing networks to operate on the full input sequence is more beneficial compared to the case where the model has only access to a subset of the input. Based on our analysis, we presented LTContext, an approach for temporal action segmentation, where we leverage sparse attention to capture the long-term context of a video and windowed attention to model the local information in the neighboring frames. Our approach achieves state-of-the-art segmental F1 scores on the 50Salads and Assembly101 datasets, which contain long videos. + +# Acknowledgement + +Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), the project iBehave (receiving funding from the programme "Netzwerke 2021", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors. + +# References + +[1] Yazan Abu Farha and Juergen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. ViViT: A video vision transformer. In IEEE International Conference on Computer Vision (ICCV), 2021. +[3] Nicolas Aziere and Sinisa Todorovic. Multistage temporal convolution transformer for action segmentation. Image and Vision Computing, 2022. +[4] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In European Conference on Computer Vision (ECCV), 2022. +[5] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document. arXiv preprint arXiv:2004.05150, 2020. +[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021. + +[7] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. +[8] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[9] Zihang Dai, Hanxiao Liu, Quoc Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems (NeurIPS), 2021. +[10] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics (ACL), 2019. +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. +[12] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. +[13] Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Fine-grained action segmentation using the semi-supervised action gan. Pattern Recognition, 2020. +[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[15] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. Conformer: Convolution-augmented transformer for speech recognition. In INTERSPEECH, 2020. +[16] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv, 2019. +[17] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[18] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. +[19] Svebor Karaman, Lorenzo Seidenari, and Alberto Del Bimbo. Fast saliency based pooling of fisher encoded dense trajectories. In European Conference on Computer Vision (ECCV) Workshops, 2014. +[20] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal + +directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. +[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 780-787, 2014. +[22] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2016. +[23] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. +[24] Colin Lea, Austin Reiter, René Vidal, and Gregory D Hager. Segmental spatiotemporal cnns for fine-grained action segmentation. In European Conference on Computer Vision (ECCV), 2016. +[25] Peng Lei and Sinisa Todorovic. Temporal deformable residual networks for action segmentation in videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. +[26] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[27] Shi-Jie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. MS-TCN++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. +[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE International Conference on Computer Vision (ICCV), 2019. +[30] Liyuan Liu, Jialu Liu, and Jiawei Han. Multi-head or singlehead? an empirical comparison for transformer training. arXiv, 2021. +[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. +[32] Zhichao Liu, Leshan Wang, Desen Zhou, Jian Wang, Songyang Zhang, Yang Bai, Errui Ding, and Rui Fan. Temporal segment transformer for action segmentation. arXiv preprint arXiv:2302.13074, 2023. +[33] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems (NeurIPS), 2019. +[34] Piotr Nawrot, Szymon Tworkowski, Michal Tyrolski, Lukasz Kaiser, Yuhuai Wu, Christian Szegedy, and Henryk Michalewski. Hierarchical transformers are more efficient language models. arXiv preprint arXiv:2110.13711, 2021. + +[35] Junyong Park, Daekyum Kim, Sejoon Huh, and Sungho Jo. Maximization and restoration: Action segmentation through dilation passing and temporal reconstruction. Pattern Recognition, 2022. +[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning (ICML), 2018. +[37] Yifan Peng, Siddharth Dalmia, Ian Lane, and Shinji Watanabe. Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding. In International Conference on Machine Learning (ICML), 2022. +[38] Jiezhong Qiu, Hao Ma, Omer Levy, Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. Findings of Empirical Methods in Natural Language Processing (EMNLP), 2020. +[39] Alexander Richard and Juergen Gall. Temporal action detection using a statistical language model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. +[40] Marcus Rohrbach, Sikandar Amin, Mykhaylo Andriluka, and Bernt Schiele. A database for fine grained activity detection of cooking activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. +[41] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[42] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. arXiv preprint, 2021. +[43] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fifa: Fast inference approximation for action segmentation. In DAGM German Conference on Pattern Recognition (GCR), 2021. +[44] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. +[45] Sebastian Stein and Stephen J. McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In ACM International Joint Conference on Pervasive and Ubiquitous Computing, page 729-738, 2013. +[46] Kevin Tang, Li Fei-Fei, and Daphne Koller. Learning latent temporal structure for complex event detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. +[47] Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. ACM Comput. Surv., 2022. +[48] Xiaoyan Tian, Ye Jin, and Xianglong Tang. Local-global transformer neural network for temporal action segmentation. Multimedia Systems, 2022. +[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training + +data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. +[50] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[51] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European Conference on Computer Vision (ECCV), 2022. +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. +[53] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. AAAI Conference on Artificial Intelligence, 2020. +[54] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. +[56] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In European Conference on Computer Vision (ECCV), 2020. +[57] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In IEEE International Conference on Computer Vision (ICCV), 2021. +[58] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. In Advances in Neural Information Processing Systems (NeurIPS), 2021. +[59] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. +[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021. +[61] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In British Machine Vision Conference (BMVC), 2021. +[62] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. +[63] Long Zhao, Zizhao Zhang, Ting Chen, and Dimitris Metaxas abd Han Zhang. Improved transformer for high-resolution GANs. In Advances in Neural Information Processing Systems (NeurIPS), 2021. \ No newline at end of file diff --git a/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/images.zip b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48c676be4b1325932ce8aad6d384fe0534f2a5d0 --- /dev/null +++ b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6c2f9a1b9b748632263ef8c13f716c9f3b5c2d8eb02a077c5e5f4ac51e3737e +size 721096 diff --git a/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/layout.json b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d1c94596746361ec2d1fe5b75f719e116d7c9b00 --- /dev/null +++ b/2023/How Much Temporal Long-Term Context is Needed for Action Segmentation_/layout.json @@ -0,0 +1,8671 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 538, + 121 + ], + "type": "text", + "content": "How Much Temporal Long-Term Context is Needed for Action Segmentation?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 222, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 222, + 156 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 222, + 156 + ], + "type": "text", + "content": "Emad Bahrami" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 143, + 352, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 143, + 352, + 157 + ], + "spans": [ + { + "bbox": [ + 246, + 143, + 352, + 157 + ], + "type": "text", + "content": "Gianpiero Francesca" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 376, + 143, + 451, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 143, + 451, + 157 + ], + "spans": [ + { + "bbox": [ + 376, + 143, + 451, + 157 + ], + "type": "text", + "content": "Juergen Gall1,3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 157, + 158, + 283, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 158, + 283, + 171 + ], + "spans": [ + { + "bbox": [ + 157, + 158, + 283, + 171 + ], + "type": "text", + "content": "1University of Bonn, Germany" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 158, + 436, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 158, + 436, + 171 + ], + "spans": [ + { + "bbox": [ + 303, + 158, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 303, + 158, + 436, + 171 + ], + "type": "text", + "content": "Toyota Motor Europe, Belgium" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 172, + 451, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 172, + 451, + 186 + ], + "spans": [ + { + "bbox": [ + 141, + 172, + 451, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 172, + 451, + 186 + ], + "type": "text", + "content": "Lamarr Institute for Machine Learning and Artificial Intelligence, Germany" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 238, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 290, + 490 + ], + "type": "text", + "content": "Modeling long-term context in videos is crucial for many fine-grained tasks including temporal action segmentation. An interesting question that is still open is how much long-term temporal context is needed for optimal performance. While transformers can model the long-term context of a video, this becomes computationally prohibitive for long videos. Recent works on temporal action segmentation thus combine temporal convolutional networks with self-attention that are computed only for a local temporal window. While these approaches show good results, their performance is limited by their inability to capture the full context of a video. In this work, we try to answer how much long-term temporal context is required for temporal action segmentation by introducing a transformer-based model that leverages sparse attention to capture the full context of a video. We compare our model with the current state of the art on three datasets for temporal action segmentation, namely 50Salads, Breakfast, and Assembly101. Our experiments show that modeling the full context of a video is necessary to obtain the best performance for temporal action segmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 533, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 593 + ], + "type": "text", + "content": "Temporal action segmentation can be used in many real-world applications such as monitoring production lines or studying animal behavior. In these settings, the videos can be very long and it is required to recognize the start and end of all actions that occur in a video as illustrated in Fig. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "Recently, combinations of temporal convolutional networks [1, 27] with self- and cross-attention from transformers [61, 4] have shown impressive results for temporal action segmentation. These works are in line with other hybrid models [15, 9, 61, 58] that combine the attention modules with convolutions to compensate for the lack of strong inductive bias of pure transformers. However, the emergence of datasets like Assembly101 [41], where subjects perform assembly tasks, poses a new challenge in the area of temporal action segmentation due to the existence of long videos" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 213, + 547, + 303 + ], + "blocks": [ + { + "bbox": [ + 307, + 213, + 547, + 303 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 547, + 303 + ], + "type": "image", + "image_path": "89c6531ce91098468580331fe78d56c67f734046b049397c4eb0f2690eb7c0e4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 310, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 407 + ], + "type": "text", + "content": "Figure 1: Datasets like Assembly101 contain long videos of assembly tasks and an action label needs to be predicted for each frame. The first row shows some frames of a video. The second row shows the ground-truth labels for all frames of the video where different colors correspond to different action labels. Rows 3-6 show the predictions of the proposed model for different amounts of long-term context where " + }, + { + "bbox": [ + 304, + 310, + 547, + 407 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 310, + 547, + 407 + ], + "type": "text", + "content": " means the temporal context of the full video." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 434, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 555 + ], + "type": "text", + "content": "that can last up to 25 minutes. Since modeling the long-term context of such a video is very expensive, Yi et al. [61] proposed to compute the attention for a local temporal window. In order to understand the impact of the window size on the temporal action segmentation accuracy, we analyze the impact of the window size on two datasets with long videos in Section 4.1. Fig. 1 shows some qualitative results of this study on Assembly101. Indeed, the results show that modeling the long-term context of an entire video is very important for temporal action segmentation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "content": "Based on this finding, we revisit how temporal attention is modeled in transformer architectures for temporal action segmentation. [61, 4] use a hierarchy of temporal windows, which makes the training on long video sequences as they occur in Assembly101 [41] very expensive. Inspired by works that decompose attention over the spatial and temporal domain for short video clips [6, 62], we propose to iterate between computing windowed local attention and sparse long-term context attention such that both short and long-term context are modeled. This approach is particularly suitable for temporal action segmentation since the local attentions focus on the similarity or dissimilarity of features within an action segment or between neighboring action" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10351" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "content": "segments whereas the long-term context attention focuses on the relations between actions within the entire video. The source code is available at https://github.com/LTContext/LTContext." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 131, + 134, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 131, + 134, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 131, + 134, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 152, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 288, + 307 + ], + "type": "text", + "content": "The traditional sliding window with non-maxima suppression has been among the early approaches for action segmentation [19, 40]. [22, 20, 46] adopted hidden Markov Models (HMMs) for temporal modeling. [39] incorporates length and temporal context and uses dynamic programming for inference. Temporal convolutional networks (TCN) with temporal pooling are used by [23] to classify each video frame. Later, [1] introduced a multi-stage TCN capable of maintaining high temporal resolution, which is necessary for fine-grained recognition. [18] reduces the over-segmentation error by adding an action boundary regression network to refine the frame-wise results. [17] proposed another refinement method based on a graph convolutional network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 308, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 288, + 617 + ], + "type": "text", + "content": "Transformers [52] were originally used for natural language processing and only rely on attention for sequence modeling. Recently, transformers have been widely adopted in vision [11, 57, 31, 54], speech recognition [15, 37], and action recognition [6, 2]. The original transformers suffered from two issues. First, the cost of the self-attention operation is quadratic with respect to the sequence length. This has been addressed by many methods that improve the memory efficiency of transformers [5, 34, 10]. We refer to [47] for a comprehensive survey of efficient transformers. For instance, restricting the attention to a fixed window size [5] is one approach. Multi-axis self-attention [51] combines block-attention with grid-attention, which is based on a spatial grid overlaid on the entire 2D space. Adapting sparse attention for capturing global information has shown to be an effective solution in vision tasks such as high-resolution image generation [63, 50], object detection [51, 60], and instance segmentation [51, 60]. MViT [12] uses a pooling operation to reduce the space-time resolution before attention and MViTv2 [28] improves MViT by adding the relative position along (x,y,t) and residual connections. The second issue of transformers is the poor generalization due to a relatively weak inductive bias [9, 49] compared to convolutional neural networks (CNN). Hybrid models that combine self-attention and convolution layers have thus been proposed for vision tasks [58, 9, 51, 57] and speech recognition [15, 37]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 288, + 715 + ], + "type": "text", + "content": "In action recognition, works such as [6, 62] adopt the idea of using attention for video understanding. In Times-Former [6], the self-attention is first applied along the temporal dimension for each single patch (time attention), i.e., over all frames of a short video clip of a few seconds. In a second step, the attention is applied over the patches of each frame (space attention). VIDTR [62] also decomposes self-attention into spatial and temporal attention, but additionally" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "down-samples the temporal dimension. In our work, we do not apply attention spatially and temporally, but we address the question of how temporal attention can be computed for very long sequences that last 25 minutes as it is required for temporal action segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 548, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 548, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 548, + 407 + ], + "type": "text", + "content": "For temporal action segmentation, [61] proposed an architecture, called ASFormer, which is based on a multi-stage TCN [1] and equips the temporal convolutions with a local window attention [5]. This is done in a hierarchy where the size of the local window grows with each layer. Recently, [4] proposed a decoder on top of the encoder that generates action segments in an autoregressive manner. It uses two heads where the head on top of the encoder predicts framewise probabilities and the head of the decoder predicts the sequence of actions. Finally, an alignment decoder fuses the output of the two heads and aligns the predicted sequence of actions to the frames. [3] introduced a hybrid Temporal Convolution Transformer (TCTr) where they adapt an action boundary detector to adaptively estimate attention from local neighboring frames. [59] proposed to use additional constraints during training, but the approach assumes that action sequences can be modelled by a directed acyclic graph, which does not allow that actions occur more than once in a video. Recently, multi-modal approaches that combine language models with vision transformers have been proposed. For instance, [26] uses prompt engineering to extract features from pre-trained vision-language models such as ActionCLIP [55]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 418, + 545, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 432 + ], + "type": "text", + "content": "3. Long-Term Context for Action Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": "Recently, transformers combined with temporal convolutional neural networks [61, 4, 3] have shown a very good performance for temporal action segmentation. For this task, the frame-wise labels " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "c_{1}, \\ldots, c_{T}" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "c_{t} \\in \\mathcal{C}" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "C = |\\mathcal{C}|" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": " denotes the number of action classes, need to be predicted for a given a video " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "X = (x_{1}, \\ldots, x_{T})" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": " frames, where " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": " represents a feature map of size " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 438, + 547, + 545 + ], + "type": "text", + "content": " can be very large, [61, 4, 3] limit the self-attention to a local temporal window." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "content": "In order to understand how much temporal long-term context is needed for temporal action segmentation, we limited the temporal input window and evaluated the quality of the temporal action segmentation in Section 4.1. The results in Fig. 4 show that temporal long-term context has a strong impact on the performance. Based on our analysis, we thus revisit the windowed attention of previous works for temporal action segmentation and propose to model the temporal long-term context of a video using sparse attentions [5, 38, 36]. Additionally, we equip our method with windowed attention to capture the locality between neighboring frames. In this way, we obtain a flexible design that is capable of providing long-term and local temporal context. While we describe first the Long-Term Context (LTCon" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10352" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 545, + 240 + ], + "type": "image", + "image_path": "995082b14912e24ab0b5a9300955eea2b4b6ece7b0df240175baeb26000878e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 548, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 548, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 548, + 298 + ], + "type": "text", + "content": "Figure 2: Illustration of windowed and long-term context attentions with a window of size 2. For the windowed attention, the sequence is partitioned into small windows and the attentions are computed for each window. For long-term context attention, the sequence is reordered such that the attentions are computed over the whole, but sparsely sampled sequence. After the attention, the output is reordered again to preserve the original order. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 317, + 288, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 288, + 341 + ], + "type": "text", + "content": "text) block in Section 3.1, the entire network is described in Section 3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 349, + 203, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 203, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 203, + 361 + ], + "type": "text", + "content": "3.1. Temporal Context Attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 368, + 288, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 288, + 462 + ], + "type": "text", + "content": "The self-attention blocks of transformers are advantageous over convolutions in the aggregation of global information. However, applying attention to long sequences such as untrimmed videos is impractical due to the quadratic complexity of the self-attention blocks. To address this issue, we adopt an attention mechanism where we leverage sparse and windowed attention to model long-term and local temporal context." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 463, + 288, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 288, + 511 + ], + "type": "text", + "content": "The attention function transforms the input into a query, key, and value and computes the output as a weighted sum of the values. For example, given a sequence of features, " + }, + { + "bbox": [ + 46, + 463, + 288, + 511 + ], + "type": "inline_equation", + "content": "X\\in \\mathbb{R}^{T\\times D}" + }, + { + "bbox": [ + 46, + 463, + 288, + 511 + ], + "type": "text", + "content": ", the attention can be written as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 518, + 289, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 518, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 61, + 518, + 289, + 548 + ], + "type": "interline_equation", + "content": "A t t e n t i o n (Q, K, V) = \\mathrm {s o f t m a x} _ {r o w} \\left(\\frac {Q K ^ {T}}{\\sqrt {D}}\\right) V \\quad (1)", + "image_path": "62ddccaeb364b8b4d67ca2a878e0fd52fb95e7a8afb5ba8f4916d701e15609b0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "inline_equation", + "content": "Q, K, V \\in \\mathbb{R}^{T \\times D}" + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "content": " are linearly transformed from " + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "content": " is very large for long sequences, we need to modify " + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 46, + 554, + 288, + 604 + ], + "type": "text", + "content": " to enable the modeling of long-term and local temporal context as illustrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Temporal Windowed Attention For the temporal Windowed Attention (WA), we partition the sequence into nonoverlapping windows of size " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ". Fig. 2 illustrates the case for " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "W = 2" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ", but we use " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "W = 64" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " in practice. The impact of " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " is evaluated in Section 4. Instead of computing the attention over the entire sequence of length " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ", we compute the attention " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{T}{W}" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " times, where each query " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{W \\times D}" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " corresponds to each window. For the keys and values, we use" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "text", + "content": "an overlap where we use the next window in addition, i.e., for each query " + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "text", + "content": " we have " + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "inline_equation", + "content": "K,V\\in \\mathbb{R}^{2W\\times D}" + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "text", + "content": ". We perform masking when " + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 317, + 547, + 367 + ], + "type": "text", + "content": " exceed the input sequence. We evaluate the impact of the overlap in Section 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": "Temporal Long-Term Context Attention For the temporal Long-Term Context (LTC) attention, the input is also partitioned into non-overlapping windows of size " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": ". However, instead of computing the attention over each window, the attention is computed over all windows where from each window only one element is taken. In the illustration of Fig. 2 with " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "G = 2" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": ", we compute the attention over the first feature of all windows and the attention over the second feature of all windows. In general, we compute the attentions for " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": " queries " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{\\frac{T}{G} \\times D}" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": " where the keys and values are the same, i.e., " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "K, V \\in \\mathbb{R}^{\\frac{T}{G} \\times D}" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": " provides the flexibility to adjust the sparseness based on the available memory budget, e.g., " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "G = 1" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": " corresponds to the case where the attention is applied over the full sequence. In practice, we use " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "G = 64" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": " and we evaluate the impact of " + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 384, + 548, + 564 + ], + "type": "text", + "content": " in Section 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "LTContext block The top of Fig. 3 illustrates the entire LTContext block. As in previous works [1, 61, 4], we use a 1D dilated temporal convolution with kernel size 3, where the dilation factor increases by factor 2 for each layer. The dilated temporal convolution is followed by a Gaussian Error Linear Unit (GELU). In the LTContext block, we first use the windowed and then the long-term context attention, which are shown in Fig. 2. We evaluate the impact of the order in Section 4. Finally, we use a linear layer with a residual connection to output the features for each frame, " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{T \\times D}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10353" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 545, + 223 + ], + "type": "image", + "image_path": "687ec45d91111480e73d139494f70f7219ea3b1c910f95828a75487ac011f750.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 232, + 455, + 244 + ], + "lines": [ + { + "bbox": [ + 137, + 232, + 455, + 244 + ], + "spans": [ + { + "bbox": [ + 137, + 232, + 455, + 244 + ], + "type": "text", + "content": "Figure 3: The network architecture of LTContext with LTContext blocks (top)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 264, + 183, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 183, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 183, + 277 + ], + "type": "text", + "content": "3.2. LTContext Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": "The entire LTContext network is depicted in Fig. 3. For a fair comparison, we will use the features that are provided for the corresponding datasets as input. In all cases, the dimensionality of the features is 2048. As in previous works, we use a linear layer to reduce the dimensionality of the features to 64. The output of each LTContext block is the feature map " + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{T \\times D}" + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": ". We repeat each LTContext block " + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": " times where the dilation factor of the temporal convolution increases in each layer. In practice, we use " + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "inline_equation", + "content": "N = 9" + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": " and we evaluate the impact of " + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": " in Section 4. After the first " + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": " layers of LTContext blocks, we use an additional linear layer to reduce the dimensionality " + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": " further to 32. The dimensionality reduction reduces the number of parameters from 1.42 million to 0.72 million without reducing the accuracy. We also use an additional linear layer followed by a softmax layer to generate the frame-wise class probabilities " + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 46, + 289, + 290, + 492 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "We continue with three additional stages where each stage consists of " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " layers of LTContext blocks. Note that we reset the dilation factor to 1 for the temporal convolution at the beginning of each stage and we compute the frame-wise class probabilities " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " after each stage, which contributes to the multi-stage loss. We use the cross-entropy loss combined with the mean squared error smoothing loss as introduced by [1] and used in [27, 61] for a fair comparison. Inspired by [61], we use the cross-attention for the LTContext blocks in stages 2 to 4. Instead of using the features " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " for the queries and keys for windowed and long-term context attention, the predictions " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " are used. We thus have " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{W \\times C}" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{2W \\times C}" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{2W \\times D}" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " for the windowed attention and " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}_{G}^{T \\times C}" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}_{G}^{T \\times C}" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}_{G}^{\\frac{T}{G} \\times D}" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " for the long-term context attention. While the cross-attention is not shown in Fig. 3, it only means that " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": " is an additional input for the windowed and long-term context attention in stages 2-4. We evaluate the impact of the number" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 266, + 398, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 398, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 398, + 277 + ], + "type": "text", + "content": "of stages in Section 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 287, + 388, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 388, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 388, + 300 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 307, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 547, + 342 + ], + "type": "text", + "content": "Datasets. We evaluate the performance of our proposed model on three challenging action segmentation datasets: 50Salads [45], Breakfast [21], and Assembly101 [41]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 390 + ], + "type": "text", + "content": "50Salads [45] contains 50 videos annotated with 17 action classes. On average, each video is 6.4 minutes long and has 18 action segments. Following previous works [1, 61, 4], we use five-fold cross-validation and report the average." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 390, + 547, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 450 + ], + "type": "text", + "content": "Breakfast contains 1,712 videos of breakfast preparation activities with an average length of 2.3 minutes. There are 48 action classes and each video has on average 6.6 action segments. For evaluation, we report the average of the 4 splits for cross-validation as in [1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 450, + 548, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 548, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 548, + 593 + ], + "type": "text", + "content": "Assembly101 [41] is the largest dataset among the three datasets with 4,321 videos and 202 coarse action classes composed of 11 verbs and 61 objects. Assembly101 is a procedural activity dataset containing videos of people assembling and disassembling 101 toy vehicles. On average, each video includes 24 action segments and is 7.1 minutes long. Compared to Breakfast, Assembly101 has 2.5 times more videos, 6.7 times more hours of video footage, 9.3 times more action segments, and 4.2 more action classes. For our evaluation, we follow the setting for temporal action segmentation [41] and report the results on the validation set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 548, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 689 + ], + "type": "text", + "content": "For a fair comparison, we use the features that are provided for the datasets and that have been used in previous works. We use the I3D [7] features for the 50Salads and Breakfast datasets and TSM [29] features for the Assembly101 dataset [41]. Both features are 2048 dimensional vectors. Following [1], we also used the temporally downsampled features for 50Salads to compensate for the different frame-rates of the datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We report the frame-wise accuracy (Acc), segmental Edit distance, and segmental F1 score at" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10354" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "the overlapping thresholds of " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " denoted by " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "F1@{10,25,50}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". The intersection over the union (IoU) ratio is used as the overlapping threshold. The edit distance measures only the order of the actions but not the duration. The frame-wise accuracy measures the accuracy per frame. It is dominated by actions that have a long duration and it is not very sensitive to over-segmentation errors that occur when a few frames within a ground-truth action segment are wrongly classified. The F1 score is the most reliable measure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 288, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 288, + 224 + ], + "type": "text", + "content": "4.1. How much temporal long-term context is needed?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 350 + ], + "type": "text", + "content": "We first present the results of our analysis on the impact of using the full sequence as input compared to using a temporal window. The goal of this experiment is to shed light on how much temporal long-term context is needed for the task of temporal action segmentation. For the analysis, we use only 50Salads and Assembly101 since the videos in Breakfast are too short. For the experiments, we train our approach (LTContext) and ASFormer [61] either on the full video sequences or we divide the videos into shorter sequences where the full context is lacking." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 288, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 288, + 458 + ], + "type": "text", + "content": "Fig. 4 shows the result of this experiment on the 50Salads and Assembly101 dataset. We report the window size as percentage of the average length of a video in the corresponding dataset and " + }, + { + "bbox": [ + 46, + 351, + 288, + 458 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 351, + 288, + 458 + ], + "type": "text", + "content": " means that the full sequence has been used. The results clearly show that the full context of an entire sequence is advantageous over allowing the model to see only a window of the input sequence even if the window is large " + }, + { + "bbox": [ + 46, + 351, + 288, + 458 + ], + "type": "inline_equation", + "content": "(50\\%)" + }, + { + "bbox": [ + 46, + 351, + 288, + 458 + ], + "type": "text", + "content": ". We can also observe that our approach benefits more from the full sequence than ASFormer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 459, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 288, + 651 + ], + "type": "text", + "content": "We furthermore evaluated whether the impact of the window size is stronger for longer videos. To this end, we sorted all test videos into four quarters by their length. Fig. 5 (left) shows that the difference between the window size " + }, + { + "bbox": [ + 46, + 459, + 288, + 651 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 459, + 288, + 651 + ], + "type": "text", + "content": " and the full video " + }, + { + "bbox": [ + 46, + 459, + 288, + 651 + ], + "type": "inline_equation", + "content": "(100\\%)" + }, + { + "bbox": [ + 46, + 459, + 288, + 651 + ], + "type": "text", + "content": " is larger for long videos. This shows that long-term context is in particular for long videos important. We also evaluated whether choosing a window for each video instead of choosing a window based on the average video length (fixed) performs better. For the video-specific window size, the window size is set to the percentage of each video. Fig. 5 (right) shows that a video-specific window size performs much worse than a fixed window. Varying the amount of context for each video is thus not beneficial. Fig. 1 shows qualitative results of our approach for different amounts of temporal context with a fixed window for a video from the Assembly101 dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 226, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 226, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 226, + 671 + ], + "type": "text", + "content": "4.2. Comparison with State of the Art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "We present the performance comparison of our method with state-of-the-art methods on the Breakfast and 50Salads datasets in Table 1 and on the Assembly101 dataset [41] in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 427, + 185 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 427, + 185 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 427, + 185 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 427, + 185 + ], + "type": "image", + "image_path": "0765e105ea9b7ae9c71a297cc6b7c8095f7e31fd26afcbb752ddeb183343379a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 195, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 304, + 195, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 547, + 254 + ], + "type": "text", + "content": "Figure 4: Impact of different sizes of the input window on the 50Salads dataset (left) and the Assembly101 dataset (right). The window size is given in percentage of the average length of a video in the corresponding dataset. " + }, + { + "bbox": [ + 304, + 195, + 547, + 254 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 195, + 547, + 254 + ], + "type": "text", + "content": " denotes entire video." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 429, + 72, + 544, + 185 + ], + "blocks": [ + { + "bbox": [ + 429, + 72, + 544, + 185 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 544, + 185 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 544, + 185 + ], + "type": "image", + "image_path": "7f87ca3a0883076c899867d20007ef72ceada82d0d70be014eb62520dbfe1972.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 269, + 426, + 354 + ], + "blocks": [ + { + "bbox": [ + 309, + 269, + 426, + 354 + ], + "lines": [ + { + "bbox": [ + 309, + 269, + 426, + 354 + ], + "spans": [ + { + "bbox": [ + 309, + 269, + 426, + 354 + ], + "type": "image", + "image_path": "1e3e0ed7221c707ec6f2d85ffa1433634f4307174c50be9c8d55495d3e5fe275.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 355, + 547, + 391 + ], + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 391 + ], + "type": "text", + "content": "Figure 5: Results for different video lengths (rights). Comparison of a fixed with a video-specific input window size (left). Both plots are for 50Salads." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 270, + 543, + 354 + ], + "blocks": [ + { + "bbox": [ + 430, + 270, + 543, + 354 + ], + "lines": [ + { + "bbox": [ + 430, + 270, + 543, + 354 + ], + "spans": [ + { + "bbox": [ + 430, + 270, + 543, + 354 + ], + "type": "image", + "image_path": "e6986f0237908ffdfa23405ce3cd8d13ce60fb069613063cc94d672650b5a5c9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 402, + 340, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 340, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 340, + 412 + ], + "type": "text", + "content": "Table 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 414, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 653 + ], + "type": "text", + "content": "On Breakfast and 50Salads, our method outperforms all comparable methods in terms of F1 score at all thresholds, which is the most important measure. Our method also achieves a better Edit score than all methods except for UVAST [4] and TCTr [3]. UVAST uses ASFormer [61] as encoder and an additional alignment decoder. The higher Edit score of UVAST is expected since the approach has an additional head that predicts the sequence of actions and thus maximizes the Edit score by an additional loss. This, however, comes at the cost of a much lower frame accuracy. TCTr [3] achieves a higher frame-wise accuracy on Breakfast and a higher Edit score on 50 Salads, but it performs worse for the other metrics. The approach uses a boundary detection module and compresses the temporal features, which is a complementary approach. It needs to be noted that our approach outperforms TCTr for all metrics on 50Salads if we use 10 instead of 9 layers as shown in Table 9. However, 9 layers perform better than 10 layers for the other datasets. We therefore report in Tables 1 and 2 only the results with 9 layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "We also compare to approaches that use different input features, perform additional test augmentation, or add additional constraints. On Breakfast, only DTL [59] reports better results. DTL [59] uses additional logic-based constraints for training ASFormer and assumes that action se" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10355" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 285, + 285 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 285, + 285 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 285, + 285 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 285, + 285 + ], + "type": "table", + "html": "
MethodNoteBreakfast50Salads
F1@{10,25,50}EditAccF1@{10,25,50}EditAcc
IDT+LM [39]DF1-----44.438.927.845.848.7
ST-CNN [24]DF2-----55.949.637.145.959.4
ED-TCN [23]DF2-----68.063.952.659.864.7
TDRN [25]DF2-----72.968.557.266.068.1
SSA-GAN [13]DF3----43.374.971.767.069.873.3
Bridge-Prompt [26]DF4-----89.287.881.383.888.1
C2F-TCN [42]TA72.268.757.669.676.084.381.872.676.484.9
UVAST [4]+ViterbiP75.970.057.276.566.089.187.681.783.987.4
UVAST [4]+FIFA [43]P76.971.558.077.169.788.987.078.583.984.5
Liu [32] + ASRF [18]P77.572.359.576.773.787.986.680.582.786.6
DTL [59]C78.874.562.977.775.887.185.778.580.586.9
MS-TCN [1]-52.648.137.961.766.376.374.064.567.980.7
MS-TCN++ [27]-64.158.645.965.667.680.778.570.174.383.7
DTGRM [53]-68.761.946.668.968.379.175.966.172.080.0
MuCon [44]-73.266.148.476.362.8-----
Gao et al. [14]-74.969.055.273.370.780.378.069.873.482.2
BCN [56]-68.765.555.066.270.482.381.374.074.384.4
SSTDA [8]-75.069.155.273.770.283.081.573.875.882.2
C2F-TCN [42]-70.166.656.268.273.576.673.062.569.280.1
ASRF [18]-74.368.956.172.467.684.983.577.379.384.5
UVAST [4]-76.770.056.677.268.286.281.270.483.979.5
DPRN [35]-75.670.557.675.171.787.886.379.482.087.2
LGTNN [48]-76.271.557.575.272.587.586.279.882.086.1
ASFormer [61]-76.070.657.475.073.585.183.476.079.685.6
TCTr [3]-76.671.158.576.177.587.586.180.283.486.6
LTContext (Ours)-77.672.660.177.074.289.487.782.083.287.7
", + "image_path": "1852c614f5c537f382d6d880d5197d11b9ccecbd1362737a97f4545a92cec6ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 422, + 285, + 483 + ], + "blocks": [ + { + "bbox": [ + 46, + 293, + 288, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 413 + ], + "type": "text", + "content": "Table 1: Results on the Breakfast and 50Salads datasets. The best and second best results for the methods in the bottom half are shown in bold and underlined since only methods without additional notes are directly comparable. P: additional post-processing; C: additional constraints; DF: different features (DF1: Improved Dense Trajectories (IDT), DF2: Spatio-temporal VGG-style CNN, DF3: Generative Adversarial Network (GAN), DF4: ActionCLIP); TA: test augmentation. In the top half, we highlight results that are better in italic bold formatting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 422, + 285, + 483 + ], + "lines": [ + { + "bbox": [ + 48, + 422, + 285, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 285, + 483 + ], + "type": "table", + "html": "
MethodAssembly101
F1@{10,25,50}EditAccParams (M)Inference (sec)
MS-TCN++ [27]31.627.820.630.737.11.080.1
UVAST [4]*32.128.320.831.537.41.223.2
C2F-TCN [42]33.329.021.332.439.26.890.1
ASFormer [61]*33.429.221.430.538.81.132.1
LTContext (Ours)33.930.022.630.441.20.720.6
", + "image_path": "bf5ec1333a1673e1cb5bbb4936471c013330cb4238dccfe69ce90b70980054bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "type": "text", + "content": "Table 2: Comparison with state-of-the-art methods on the Assembly101 dataset. The best and second best results are shown in bold and underlined. *We trained UVAST [4] and ASFormer [61] using the code of the authors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "quences can be modeled by a directed acyclic graph. This is very efficient on the Breakfast dataset, which has relatively few actions per video compared to the other datasets. However, our approach outperforms DTL for all metrics on 50Salads where the structure of the action sequences is more complex. Furthermore, DTL cannot be applied to datasets like Assembly101 where actions occur more than once in a video. On 50Salads, only Bridge-Prompt [26] performs slightly better for some metrics. Bridge-Prompt proposes an approach for feature learning using a vision-language model such as ActionCLIP [55] in combination with ASFormer. The approach is thus complementary and not comparable to our approach. Nevertheless, our approach achieves higher" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "F1@10" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "F1@50" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " scores on 50Salads. If we use 10 instead of 9 layers as shown in Table 9, we even outperform Bridge-Prompt for all measures except for frame-wise accuracy. In summary, our approach outperforms all methods in terms of F1 score, which is the most precise measure of the segmentation quality, at all thresholds on 50Salads." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 144, + 547, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 288 + ], + "type": "text", + "content": "Assembly101 is the largest dataset both in terms of the number of videos and their length, and it is the most challenging dataset. Since ASFormer [61] and UVAST [4] have not been evaluated on this dataset, we trained ASFormer and UVAST on Assembly101 using the publicly available source code and report the results in Table 2 as well. We outperform all methods on Assembly101 in F1 scores. C2F-TCN [42] achieves the best edit score. Since we trained ASFormer [61] and UVAST [4] by ourselves, we can compare the training time on Assembly101. While our approach requires 1 day and 18 hours, ASFormer [61] and UVAST [4] needed 4 weeks and 2 weeks, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 295, + 435, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 295, + 435, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 295, + 435, + 308 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 313, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 547, + 564 + ], + "type": "text", + "content": "In Figs. 6-8, we present some qualitative results for the Assembly101, Breakfast, and 50Salads dataset. The first row of each figure shows the middle frame of each ground-truth action segment. In the second, third, and fourth rows, the ground truth segmentation, the predictions of ASFormer [61], and the prediction of our model (LTContext) are shown, respectively. In Fig. 6, our approach shows much fewer errors than ASFormer. Although our approach recognizes all action classes that occur in the video, there are several errors where some instances are missed like the last action segment of the bottom row which corresponds to the action 'detach base'. This indicates how challenging the Assembly101 dataset is. In Fig. 7, the predictions of our model are very close to the ground-truth. ASFormer overestimates the duration of the green segment, which corresponds to the action 'spoon powder', and hallucinates purple segments at the beginning and end of the video. In Fig. 8, both methods estimate the segments well, but ASFormer predicts wrongly two orange segments, which correspond to the action 'mix dressing', and the olive segment is too short, which corresponds to the action 'add salt'." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 572, + 406, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 406, + 584 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 406, + 584 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 590, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 547, + 628 + ], + "type": "text", + "content": "We finally evaluate the impact of each component of our architecture. For the ablation studies, we report the results averaged over the 5 splits of the 50Salads dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Impact of attention types In Table 3, we show the impact of using the combination of windowed and long-term context (LTContext) attention in the LTContext block illustrated in Fig. 3. We first compare it to two variants where we use only windowed or only long-term context attention. In order to keep the number of parameters the same, we still use" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10356" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 547, + 250 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 547, + 250 + ], + "type": "image", + "image_path": "19a37f9fa2afed00f320001205ad8a91eb6b1dbd4db0d8d1396c9684d254007a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 259, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 547, + 285 + ], + "type": "text", + "content": "Figure 6: Qualitative results on Assembly101. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext. It can be best viewed by using the zoom function of a PDF viewer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 295, + 545, + 401 + ], + "blocks": [ + { + "bbox": [ + 58, + 295, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 58, + 295, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 58, + 295, + 545, + 401 + ], + "type": "image", + "image_path": "4a5e768fa14b18d64125e296077d1579173596435439661b9103cb8d1525804b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 410, + 547, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 410, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 547, + 436 + ], + "type": "text", + "content": "Figure 7: Qualitative results on Breakfast. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 453, + 286, + 515 + ], + "blocks": [ + { + "bbox": [ + 48, + 453, + 286, + 515 + ], + "lines": [ + { + "bbox": [ + 48, + 453, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 286, + 515 + ], + "type": "table", + "html": "
Model ArchitectureF1@{10, 25, 50}EditAcc
LTContext89.487.782.083.287.7
- Windowed Attention87.785.177.482.185.5
- LTContext Attention84.182.674.178.285.3
WA (S1) + LTContext (S2-S4)87.885.079.481.485.2
", + "image_path": "1e1d8beb1401b6c3319e81f658c554d410d646b47db3cb1bd13dd1c7bbfdde24.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 523, + 288, + 618 + ], + "lines": [ + { + "bbox": [ + 46, + 523, + 288, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 288, + 618 + ], + "type": "text", + "content": "Table 3: Impact of using LTContext and Windowed Attention (WA) on 50Salads. In the case of Windowed Attention, we use two windowed attention blocks instead of a combination of windowed and LTContext attention. In the case of LTContext Attention, we use two LTContext attention blocks. The last row corresponds to using only windowed attention in stage 1 and only LTContext attention in stages 2 to 4." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 639, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 287, + 688 + ], + "type": "text", + "content": "in these cases two attention blocks within LTContext. The results show that combining both types of attention leads to better results. In particular, the F1@50 score is substantially higher." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "As shown in Fig. 3, we use four LTContext blocks. We also evaluate what happens if we vary the attention not within" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 455, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 540 + ], + "type": "text", + "content": "an LTContext block but between the four LTContext blocks. For this, we use only windowed attention for the first LTContext block and only long-term context attention for the other three LTContext blocks. The last row in Table 3 shows that this does not perform better than using only windowed attention and it is worse than combining windowed and long-term context attention within an LTContext block." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": "Impact of different values of " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": " The parameter " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": " controls the size of the local window for the local attention and the parameter " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": " controls the sparseness of the global attention. If not otherwise specified, we use " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "W = G = 64" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": " in our experiments. The results in Table 4 show that the F1 score drops when the size of the local window " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": " becomes smaller. Note that larger values of " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": " increase the memory and computational cost. When we decrease " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": ", the F1 score also drops but not as drastic as for " + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 554, + 547, + 662 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Impact of overlaps for windowed attention As described in Section 3.1, we use an overlap for the keys and values for the windowed attention. In Table 5, we also report the result" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10357" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 521, + 170 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 521, + 170 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 521, + 170 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 521, + 170 + ], + "type": "image", + "image_path": "d3b4b14841ba7291dc5af17a582d79e4b482f90702d664c41c88792dee8eec74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 205 + ], + "type": "text", + "content": "Figure 8: Qualitative results on 50Salads. The three rows show the ground-truth labels, the predictions by ASFormer, and the predictions by the proposed approach LTContext." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 220, + 261, + 327 + ], + "blocks": [ + { + "bbox": [ + 71, + 220, + 261, + 327 + ], + "lines": [ + { + "bbox": [ + 71, + 220, + 261, + 327 + ], + "spans": [ + { + "bbox": [ + 71, + 220, + 261, + 327 + ], + "type": "table", + "html": "
WGF1@{10, 25, 50}EditAcc
86483.882.174.577.087.0
166485.482.977.177.187.0
326487.685.879.881.586.6
646489.487.782.083.287.7
643288.887.281.383.287.7
641688.687.080.382.686.2
64888.687.179.982.986.6
", + "image_path": "643744abcb0856330f91cc2406152cd460f8bb17315b753b9e79357279791066.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 66, + 357, + 268, + 407 + ], + "blocks": [ + { + "bbox": [ + 50, + 335, + 284, + 348 + ], + "lines": [ + { + "bbox": [ + 50, + 335, + 284, + 348 + ], + "spans": [ + { + "bbox": [ + 50, + 335, + 284, + 348 + ], + "type": "text", + "content": "Table 4: Impact of the parameters " + }, + { + "bbox": [ + 50, + 335, + 284, + 348 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 50, + 335, + 284, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 335, + 284, + 348 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 50, + 335, + 284, + 348 + ], + "type": "text", + "content": " on 50 Salads." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 357, + 268, + 407 + ], + "lines": [ + { + "bbox": [ + 66, + 357, + 268, + 407 + ], + "spans": [ + { + "bbox": [ + 66, + 357, + 268, + 407 + ], + "type": "table", + "html": "
F1@{10, 25, 50}EditAcc
Non-Overlapping87.985.579.181.485.1
Overlap (2 Windows)89.487.782.083.287.7
Overlap (3 Windows)89.288.080.883.286.6
", + "image_path": "f354e03d20f7005a2716be7506936b8bd472247fee008c5f934a9f0d963c52ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 289, + 555 + ], + "type": "text", + "content": "when the keys and values do not overlap, i.e., they are the same as the queries. It is interesting to note that our attention can be interpreted as a combination of a reshaping operation with axial attention [16] in the special case without overlap. The results, however, show that an overlap improves the results. Using a larger overlap where the keys and values consist of three consecutive windows does not improve the results further." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 574, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 660 + ], + "type": "text", + "content": "Impact of the order of the attention The LTContext block shown in Fig. 3 uses first windowed attention and then long-term context attention. Table 6 shows the results when we change the order of windowed and long-term context attention within the LTContext block. If the order is changed, the performance decreases. Since the LTContext blocks are repeated, the drop in performance is moderate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Impact of the cross-attention As described in Section 3.2, we use cross-attention in stages 2 to 4. Table 7 shows that the performance drastically decreases without cross-attention." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 221, + 533, + 269 + ], + "blocks": [ + { + "bbox": [ + 46, + 415, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 440 + ], + "type": "text", + "content": "Table 5: Impact of using an overlap for the keys and values for the local attention on 50Salads." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 221, + 533, + 269 + ], + "lines": [ + { + "bbox": [ + 318, + 221, + 533, + 269 + ], + "spans": [ + { + "bbox": [ + 318, + 221, + 533, + 269 + ], + "type": "table", + "html": "
F1@{10, 25, 50}EditAcc
WA-LTContext89.487.782.083.287.7
LTContext-WA88.386.480.481.686.1
", + "image_path": "a59f647768768f3fc5a49fbaf8db2674cba8af25d575441f8518360af31fa2a0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 321, + 533, + 362 + ], + "blocks": [ + { + "bbox": [ + 305, + 276, + 547, + 313 + ], + "lines": [ + { + "bbox": [ + 305, + 276, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 276, + 547, + 313 + ], + "type": "text", + "content": "Table 6: Impact of the attention order within the LTContext block on 50Salads. Our model computes first windowed attention (WA) and then long-term context attention (row 1)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 321, + 533, + 362 + ], + "lines": [ + { + "bbox": [ + 318, + 321, + 533, + 362 + ], + "spans": [ + { + "bbox": [ + 318, + 321, + 533, + 362 + ], + "type": "table", + "html": "
F1@{10, 25, 50}EditAcc
Without Cross-Attention82.278.570.473.680.2
LTContext89.487.782.083.287.7
", + "image_path": "0312fb8748ad8d72b04bd31ec90605b2b76bda1afb4713016822c1d9c2faaf3a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 318, + 391, + 533, + 443 + ], + "blocks": [ + { + "bbox": [ + 322, + 370, + 528, + 383 + ], + "lines": [ + { + "bbox": [ + 322, + 370, + 528, + 383 + ], + "spans": [ + { + "bbox": [ + 322, + 370, + 528, + 383 + ], + "type": "text", + "content": "Table 7: Impact of the cross-attention on 50Salads." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 391, + 533, + 443 + ], + "lines": [ + { + "bbox": [ + 318, + 391, + 533, + 443 + ], + "spans": [ + { + "bbox": [ + 318, + 391, + 533, + 443 + ], + "type": "table", + "html": "
F1@{10, 25, 50}EditAcc
Conv with Dilation89.487.782.083.287.7
Conv without Dilation80.178.067.772.679.3
Without Conv81.879.267.172.778.9
", + "image_path": "65139d554878af774ffac09fd24e1094d7875dc447e86c007b2aab0392c97db9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 451, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 305, + 451, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 547, + 487 + ], + "type": "text", + "content": "Table 8: Impact of using LTContext blocks with 1D convolution but without dilation (row 2) and using LTContext blocks without 1D convolution (row 3)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 505, + 548, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 548, + 601 + ], + "type": "text", + "content": "Impact of using convolutions The LTContext block shown in Fig. 3 starts with a dilated 1D convolution with kernel size 3. In Table 8, we evaluate the impact of the dilated convolution by comparing to a LTContext block that uses a convolutional kernel of the same size but without dilation factor, and a LTContext block without 1D convolution. The results show that dilated convolutions have a very high impact on the performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Impact of the number of layers As shown in Fig. 3, we repeat the LTContext blocks at each stage " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " times. We used " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "N = 9" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " layers in all experiments. Table 9 shows the impact of varying the number of layers. On 50Salads, all measures improve by increasing " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ". It needs to be noted that we use " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "N = 9" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " for the results reported in Table 1 although we can get even better results with 10 layers on 50 Salads. For the Breakfast and Assembly101 dataset, the best performance is" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10358" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 273, + 232 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 273, + 232 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 273, + 232 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 273, + 232 + ], + "type": "table", + "html": "
number of layers (N)50Salads
F1@{10, 25, 50}EditAcc
887.285.177.680.885.7
989.487.782.083.287.7
1089.588.182.484.187.7
Breakfast
875.770.557.774.573.1
977.672.660.177.074.2
1077.372.459.776.473.5
Assembly101
831.928.421.327.841.0
933.930.022.630.441.2
1032.629.321.928.741.5
", + "image_path": "4e1000f31c3996c65a0eca89021b26cf3d5522e8c78718a1dd2a9fd860890506.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 273, + 273, + 342 + ], + "blocks": [ + { + "bbox": [ + 47, + 242, + 288, + 268 + ], + "lines": [ + { + "bbox": [ + 47, + 242, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 288, + 268 + ], + "type": "text", + "content": "Table 9: Impact of the number of layers on 50Salads, Breakfast, and Assembly101." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 273, + 273, + 342 + ], + "lines": [ + { + "bbox": [ + 60, + 273, + 273, + 342 + ], + "spans": [ + { + "bbox": [ + 60, + 273, + 273, + 342 + ], + "type": "table", + "html": "
number of headsF1@{10, 25, 50}EditAcc
189.487.782.083.287.7
288.887.080.582.987.0
489.287.180.883.186.9
888.286.480.582.186.1
", + "image_path": "41d2456dc440e9776237cbdd1dfce763ebda033e70312d11f1905ec4274fd697.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 379, + 143, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 143, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 143, + 392 + ], + "type": "text", + "content": "achieved with 9 layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 405, + 288, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 288, + 538 + ], + "type": "text", + "content": "Impact of the number of attention heads In our implementation, we do not use multiple attention heads. Nevertheless, we evaluated the impact of using multiple heads in Table 10 since most transformers use multiple heads. The results, however, show that there is no benefit in using multiple heads. The results are consistent with the observations in [33, 30]. For example, [33] shows that many attention heads can often be reduced to a single head without impacting the performance. They also argue that some tasks are more reliant on multiple heads than others. Temporal action segmentation seems to be a task where one head is sufficient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 552, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 659 + ], + "type": "text", + "content": "Impact of the number of stages As shown in Fig. 3, we use four stages of LTContext blocks, each of them with 9 layers. We evaluate the impact of the number of stages in Table 11. As can be seen, using multiple stages helps to reduce the over-segmentation error and improves the F1 score and Edit score significantly compared to using only one stage. Increasing the number of stages up to 4 improves all metrics, but using 5 stages decreases the performance and the network starts to overfit." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 669, + 119, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 682 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In this work, we addressed the question of how much temporal long-term context is needed for action segmentation." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 325, + 70, + 526, + 144 + ], + "blocks": [ + { + "bbox": [ + 64, + 350, + 269, + 363 + ], + "lines": [ + { + "bbox": [ + 64, + 350, + 269, + 363 + ], + "spans": [ + { + "bbox": [ + 64, + 350, + 269, + 363 + ], + "type": "text", + "content": "Table 10: Impact of the number of attention heads." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 70, + 526, + 144 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 526, + 144 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 526, + 144 + ], + "type": "table", + "html": "
number of stagesF1@{10, 25, 50}EditAcc
156.354.149.945.684.7
286.284.578.280.586.7
387.685.778.381.285.5
489.487.782.083.287.7
589.086.880.083.085.7
", + "image_path": "c18a2b1d2fb5f13bec40c1e9d247571b5e3bc2cb120c6c0205cc06cb6b7f0817.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 153, + 535, + 166 + ], + "lines": [ + { + "bbox": [ + 315, + 153, + 535, + 166 + ], + "spans": [ + { + "bbox": [ + 315, + 153, + 535, + 166 + ], + "type": "text", + "content": "Table 11: Impact of the number of stages on 50Salads." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 182, + 547, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 301 + ], + "type": "text", + "content": "Our analysis indicates that allowing networks to operate on the full input sequence is more beneficial compared to the case where the model has only access to a subset of the input. Based on our analysis, we presented LTContext, an approach for temporal action segmentation, where we leverage sparse attention to capture the long-term context of a video and windowed attention to model the local information in the neighboring frames. Our approach achieves state-of-the-art segmental F1 scores on the 50Salads and Assembly101 datasets, which contain long videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 311, + 404, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 404, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 404, + 324 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 331, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 437 + ], + "type": "text", + "content": "Juergen Gall has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) GA 1927/4-2 (FOR 2535 Anticipating Human Behavior), the project iBehave (receiving funding from the programme \"Netzwerke 2021\", an initiative of the Ministry of Culture and Science of the State of Northrhine Westphalia), and the ERC Consolidator Grant FORHUE (101044724). The sole responsibility for the content of this publication lies with the authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 449, + 365, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 365, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 365, + 460 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 468, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 310, + 468, + 547, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 468, + 547, + 512 + ], + "spans": [ + { + "bbox": [ + 310, + 468, + 547, + 512 + ], + "type": "text", + "content": "[1] Yazan Abu Farha and Juergen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 514, + 547, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 514, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 311, + 514, + 547, + 556 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. ViViT: A video vision transformer. In IEEE International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 559, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 559, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 311, + 559, + 547, + 590 + ], + "type": "text", + "content": "[3] Nicolas Aziere and Sinisa Todorovic. Multistage temporal convolution transformer for action segmentation. Image and Vision Computing, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 592, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 592, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 311, + 592, + 547, + 645 + ], + "type": "text", + "content": "[4] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 647, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 647, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 311, + 647, + 547, + 669 + ], + "type": "text", + "content": "[5] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document. arXiv preprint arXiv:2004.05150, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 670, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 670, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 311, + 670, + 547, + 712 + ], + "type": "text", + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10359" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[7] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 162 + ], + "type": "text", + "content": "[8] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 205 + ], + "type": "text", + "content": "[9] Zihang Dai, Hanxiao Liu, Quoc Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems (NeurIPS), 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "text", + "content": "[10] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics (ACL), 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 340 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 385 + ], + "type": "text", + "content": "[12] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 387, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 288, + 420 + ], + "type": "text", + "content": "[13] Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Fine-grained action segmentation using the semi-supervised action gan. Pattern Recognition, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 464 + ], + "type": "text", + "content": "[14] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2local: Efficient structure search for video action segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 520 + ], + "type": "text", + "content": "[15] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. Conformer: Convolution-augmented transformer for speech recognition. In INTERSPEECH, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "type": "text", + "content": "[16] Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. Axial attention in multidimensional transformers. arXiv, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "text", + "content": "[17] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "type": "text", + "content": "[18] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[19] Svebor Karaman, Lorenzo Seidenari, and Alberto Del Bimbo. Fast saliency based pooling of fisher encoded dense trajectories. In European Conference on Computer Vision (ECCV) Workshops, 2014." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[20] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "text", + "content": "[21] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 780-787, 2014." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 182 + ], + "type": "text", + "content": "[22] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 547, + 227 + ], + "type": "text", + "content": "[23] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 271 + ], + "type": "text", + "content": "[24] Colin Lea, Austin Reiter, René Vidal, and Gregory D Hager. Segmental spatiotemporal cnns for fine-grained action segmentation. In European Conference on Computer Vision (ECCV), 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 273, + 547, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 547, + 314 + ], + "type": "text", + "content": "[25] Peng Lei and Sinisa Todorovic. Temporal deformable residual networks for action segmentation in videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 547, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 369 + ], + "type": "text", + "content": "[26] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 371, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 547, + 415 + ], + "type": "text", + "content": "[27] Shi-Jie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. MS-TCN++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "type": "text", + "content": "[28] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 547, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 547, + 503 + ], + "type": "text", + "content": "[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In IEEE International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 505, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 547, + 536 + ], + "type": "text", + "content": "[30] Liyuan Liu, Jialu Liu, and Jiawei Han. Multi-head or singlehead? an empirical comparison for transformer training. arXiv, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 547, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 591 + ], + "type": "text", + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "type": "text", + "content": "[32] Zhichao Liu, Leshan Wang, Desen Zhou, Jian Wang, Songyang Zhang, Yang Bai, Errui Ding, and Rui Fan. Temporal segment transformer for action segmentation. arXiv preprint arXiv:2302.13074, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 636, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 547, + 668 + ], + "type": "text", + "content": "[33] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems (NeurIPS), 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[34] Piotr Nawrot, Szymon Tworkowski, Michal Tyrolski, Lukasz Kaiser, Yuhuai Wu, Christian Szegedy, and Henryk Michalewski. Hierarchical transformers are more efficient language models. arXiv preprint arXiv:2110.13711, 2021." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10360" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[35] Junyong Park, Daekyum Kim, Sejoon Huh, and Sungho Jo. Maximization and restoration: Action segmentation through dilation passing and temporal reconstruction. Pattern Recognition, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 163 + ], + "type": "text", + "content": "[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning (ICML), 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 219 + ], + "type": "text", + "content": "[37] Yifan Peng, Siddharth Dalmia, Ian Lane, and Shinji Watanabe. Branchformer: Parallel MLP-attention architectures to capture local and global context for speech recognition and understanding. In International Conference on Machine Learning (ICML), 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 288, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 288, + 266 + ], + "type": "text", + "content": "[38] Jiezhong Qiu, Hao Ma, Omer Levy, Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. Findings of Empirical Methods in Natural Language Processing (EMNLP), 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 288, + 300 + ], + "type": "text", + "content": "[39] Alexander Richard and Juergen Gall. Temporal action detection using a statistical language model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 302, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 288, + 345 + ], + "type": "text", + "content": "[40] Marcus Rohrbach, Sikandar Amin, Mykhaylo Andriluka, and Bernt Schiele. A database for fine grained activity detection of cooking activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 348, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 288, + 402 + ], + "type": "text", + "content": "[41] F. Sener, D. Chatterjee, D. Shelepov, K. He, D. Singhania, R. Wang, and A. Yao. Assembly101: A large-scale multiview video dataset for understanding procedural activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "type": "text", + "content": "[42] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. arXiv preprint, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 439, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 288, + 483 + ], + "type": "text", + "content": "[43] Yaser Souri, Yazan Abu Farha, Fabien Despinoy, Gianpiero Francesca, and Juergen Gall. Fifa: Fast inference approximation for action segmentation. In DAGM German Conference on Pattern Recognition (GCR), 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 485, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 288, + 529 + ], + "type": "text", + "content": "[44] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast weakly supervised action segmentation using mutual consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 531, + 288, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 288, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 288, + 584 + ], + "type": "text", + "content": "[45] Sebastian Stein and Stephen J. McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In ACM International Joint Conference on Pervasive and Ubiquitous Computing, page 729-738, 2013." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 587, + 288, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 288, + 630 + ], + "type": "text", + "content": "[46] Kevin Tang, Li Fei-Fei, and Daphne Koller. Learning latent temporal structure for complex event detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 633, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 633, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 633, + 288, + 654 + ], + "type": "text", + "content": "[47] Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. ACM Comput. Surv., 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "text", + "content": "[48] Xiaoyan Tian, Ye Jin, and Xianglong Tang. Local-global transformer neural network for temporal action segmentation. Multimedia Systems, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 714 + ], + "type": "text", + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 708 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "type": "text", + "content": "[50] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 194 + ], + "type": "text", + "content": "[51] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 197, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 547, + 239 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 241, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 547, + 274 + ], + "type": "text", + "content": "[53] Dong Wang, Di Hu, Xingjian Li, and Dejing Dou. Temporal relational modeling with self-supervision for action segmentation. AAAI Conference on Artificial Intelligence, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 275, + 547, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 547, + 318 + ], + "type": "text", + "content": "[54] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. MaX-DeepLab: End-to-end panoptic segmentation with mask transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 319, + 547, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 547, + 350 + ], + "type": "text", + "content": "[55] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 353, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 547, + 396 + ], + "type": "text", + "content": "[56] Zhenzhi Wang, Ziteng Gao, Limin Wang, Zhifeng Li, and Gangshan Wu. Boundary-aware cascade networks for temporal action segmentation. In European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 397, + 547, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 547, + 441 + ], + "type": "text", + "content": "[57] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In IEEE International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 441, + 547, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 547, + 485 + ], + "type": "text", + "content": "[58] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 487, + 547, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 547, + 540 + ], + "type": "text", + "content": "[59] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 542, + 547, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 542, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 308, + 542, + 547, + 585 + ], + "type": "text", + "content": "[60] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 586, + 547, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 586, + 547, + 618 + ], + "spans": [ + { + "bbox": [ + 308, + 586, + 547, + 618 + ], + "type": "text", + "content": "[61] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In British Machine Vision Conference (BMVC), 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 620, + 547, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 620, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 308, + 620, + 547, + 663 + ], + "type": "text", + "content": "[62] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 665, + 547, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 665, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 665, + 547, + 708 + ], + "type": "text", + "content": "[63] Long Zhao, Zizhao Zhang, Ting Chen, and Dimitris Metaxas abd Han Zhang. Improved transformer for high-resolution GANs. In Advances in Neural Information Processing Systems (NeurIPS), 2021." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10361" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_content_list.json b/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c027762060a37c764b7786905f69df8ff2eaddd --- /dev/null +++ b/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_content_list.json @@ -0,0 +1,1354 @@ +[ + { + "type": "text", + "text": "How to Boost Face Recognition with StyleGAN?", + "text_level": 1, + "bbox": [ + 241, + 130, + 727, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Artem Sevastopolsky1 Yury Malkov2,* Nikita Durasov3 Luisa Verdoliva1,4 Matthias Nießner1", + "bbox": [ + 225, + 179, + 743, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Technical University of Munich, Germany 2 Twitter, US 3 École polytechnique fédérale de Lausanne, Switzerland 4 University Federico II of Naples, Italy", + "bbox": [ + 236, + 232, + 733, + 290 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e00cfd1644ea9e081094344af373d3dc09ed5284de51c502c7b5f0b50fb7315.jpg", + "image_caption": [ + "Figure 1: Our method is aimed at boosting the performance of face recognition. This is achieved by gathering a random image collection without face recognition labels (unlabeled set) and then fitting a mapping from an image to the StyleGAN latent space onto that collection. To learn this mapping, we use the pSp encoder architecture. For the downstream face recognition task, the same encoder is then fine-tuned on a (potentially much smaller) face recognition dataset with identity labels (labeled set)." + ], + "image_footnote": [], + "bbox": [ + 196, + 305, + 769, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 540, + 313, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art face recognition systems require vast amounts of labeled training data. Given the priority of privacy in face recognition applications, the data is limited to celebrity web crawls, which have issues such as limited numbers of identities. On the other hand, self-supervised revolution in the industry motivates research on the adaptation of related techniques to facial recognition. One of the most popular practical tricks is to augment the dataset by the samples drawn from generative models while preserving the identity. We show that a simple approach based on fine-tuning $pSp$ encoder for StyleGAN allows to improve upon the state-of-the-art facial recognition and performs better compared to training on synthetic face identities. We also collect large-scale unlabeled datasets with controllable ethnic constitution - AfricanFaceSet-5M (5 million images of different people) and AsianFaceSet-3M (3 million images of different people) - and we show that pretraining on each of them improves recognition of the respective ethnicities (as well as others), while combining all unlabeled datasets results in the biggest performance increase. Our self-supervised strategy is the most useful with limited amounts of labeled training data,", + "bbox": [ + 73, + 569, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which can be beneficial for more tailored face recognition tasks and when facing privacy concerns. Evaluation is based on a standard RFW dataset and a new large-scale RB-WebFace benchmark. The code and data are made publicly available at https://github.com/seval00/stylegan-for-facerec.", + "bbox": [ + 498, + 542, + 893, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 648, + 632, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern face recognition methods rely on deep convolutional networks trained on large-scale datasets [54, 7, 23, 60]. These methods are now being integrated into a vast number of real-world applications, ranging from face unlock for smartphones and photo organizers to law enforcement systems and border control. A typical open face recognition dataset consists of web-crawled images of celebrities, leading to limited size and lack of balance in subgroups, such as ethnicity, age, etc. Training a state-of-the-art solution, however, requires enormous amounts of labeled data, scraping which may lead to privacy and legal issues. We suggest and study an alternative solution to using celebrity photos - pretraining the face recognition backbone on a generative task. Specifically, we first train StyleGAN2-ADA [29] on collected", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20924", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce4aea9a1ba9b95c3edeace262404e5f6d803ae6dd38f975fa79f4dd5b12353a.jpg", + "image_caption": [ + "Figure 2: Our method is trained in three consecutive steps. First, we fit StyleGAN2-ADA to the face image distribution of the unlabeled prior dataset $\\mathcal{D}^{prior}$ . Second, the pSp encoder is trained (also on $\\mathcal{D}^{prior}$ ) to map images to the latent codes in the learned latent space. Finally, the encoder, which is pretrained to extract meaningful features from an image, is fine-tuned for the downstream face recognition task with the ArcFace loss (similar losses can be used instead) on $\\mathcal{D}^{facerec}$ . The two first steps comprise the self-supervised pretraining stage; i.e., no identity labels are required for them." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 336, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e6ae3af18f67dfb67f11c10417def155595b4175cdd61330fabc429e7a7bfe8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 90, + 666, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a1ca5781353c1f3c74e4af754ea3f7ea328cecc1ab98deaf1d7c7f2c243a3e99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 90, + 888, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unlabeled data (which we later refer to as an unlabeled prior dataset) to fit the face image distribution. Subsequently, we train an encoder (following pixel2style2pixel (pSp) architecture [41]) that maps input images to vectors in the learned StyleGAN2-ADA latent space. Importantly, during the pretraining steps, no identity labels are used, so we can use diverse datasets crawled from the Internet without compromising privacy. Finally, we transfer the learned pSp encoder convolutional weights into the face recognition network and train it in a standard face recognition setup.", + "bbox": [ + 75, + 352, + 470, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that, in contrast to training face recognition tasks on StyleGAN generated data (also demonstrated in [38, 39] and studied e.g. in [36]), our encoder pretraining step significantly boosts the final performance. The idea of augmenting face recognition datasets with synthetic data is widespread and constitutes many approaches, however, unclear and heuristic definition of the target label limits the amount of useful signal that can be transferred into the face recognition model this way. Our approach goes hand-in-hand with the current development of self-supervised learning [15, 35, 12, 16] and makes up one of the first approaches of its application to face recognition [26, 33]. This allows us to demonstrate vast improvements on limited labeled training data compared to the setup without self-supervised pretraining (for instance, $10\\%$ verification accuracy increase for only $1\\%$ of the labeled data used).", + "bbox": [ + 75, + 506, + 468, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The simplicity of the data collection procedure also allows us to control the distribution of the unlabeled data and thus influence the decrease of the error rates for specific demographic groups. Despite the fact that the current state-of-the-art algorithms often demonstrate very low average error rates [13, 47, 51, 48], it is considered unethical to integrate face recognition solutions that exhibit significant ethnic, age, or gender bias. Such bias is present both in open-source face recognition methods [20, 53, 49, 19, 28, 44, 46] and in comprehensively evaluated commercial face recog", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nition systems [22], resulting in significantly different error rates measured for the groups of interest. The topic has attracted significant attention in other areas of computer vision operating in face domain [8, 14, 21, 40] and in the media.", + "bbox": [ + 496, + 352, + 890, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We constructively demonstrate that collecting large amounts of in-the-wild face images of a given group of interest is feasible (and can be done semi-automatically), while collecting datasets with identity labels is problematic. The labels require linking photos of the same person taken in different conditions, which means the person must be tracked. This typically constrains public datasets to celebrities, gathered using search engines [23, 54, 7, 60], while social networks and companies that provide services with photos use input from users. Second, the collected in-the-wild data, treated as a set of faces without identity labels (but labeled with the group attribution), can be efficiently used for self-supervised pretraining for face recognition networks, subsequently fine-tuned on the standard face recognition datasets (see Fig. 1).", + "bbox": [ + 496, + 429, + 892, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize our main contribution, we present a novel self-supervised method for improving the performance of face recognition based on StyleGAN pretraining. This allows to leverage large-scale amounts of available unlabeled data for face recognition. While the improvement is the most significant on limited data, pretraining is also helpful for large-scale labeled datasets.", + "bbox": [ + 496, + 655, + 890, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 768, + 635, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face recognition datasets. Several large-scale datasets of faces with identity labels have been released publicly, such as CASIA-WebFace [54], VGGFace2 [7], MS-Celeb-1M [23] and the very recent million-scale WebFace-42M dataset [60]. However, these datasets have been collected \"in the wild\", hence, they inevitably suffer from biases in terms of age, gender,", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20925", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/b0367a326945200d30a3b55d0b197774fe884de55d2536f1498aa5dfb90a06fd.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Overview of the publicly available training sets in facial domain. Typically, the quality of a face recognition dataset is described by several factors: the number of people in the dataset, the number of images per person, and the diversity of capture conditions. While our AfricanFaceSet-5M and AsianFaceSet-3M datasets are not directly suitable for face recognition due to the absence of identity labels, they comprise more distinct people than existing large-scale datasets and contain a more diverse distribution of faces than only celebrities." + ], + "table_body": "
Dataset name# people# images# pic./personethnic diversityacquisition
MS-Celeb-1M100K8.2M82unevenmostly American and British actors
VGGFace29.1K3.3M362unevenGoogle Image search
CASIA-WebFace10K494K49unevencelebrities from IMDb
CASIA-Face-Africa1.1K38K34all Africancontrolled, indoor and outdoor
MegaFace (unavailable)672K4.7M7unevenYahoo Flickr website search queries
DiveFace24K72K3balanced across 3 ethnicitiessubset of MegaFace
BUPT-BalancedFace28K1.3M46balanced across 4 ethnicitiescelebrities from MS-Celeb-1M
BUPT-GlobalFace38K2M52matches global populationcelebrities from MS-Celeb-1M
BUPT-TransferFace>10K600K6075% Caucasian. vs. otherscelebrities from MS-Celeb-1M
AfricanFaceSet-5M5M5MunlabeledAfrican majorityrandom faces from YouTube news videos
AsianFaceSet-3M3M3MunlabeledAsian majorityrandom faces from YouTube news videos
", + "bbox": [ + 76, + 89, + 885, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "or race. To better understand and stimulate research in fairness about face recognition, several datasets have been proposed. Examples of such labeled datasets include BUPT-Globalface [49] (2M images with ethnic distribution matching the world population) and BUPT-Balancedface [49] (1.3M images with the perfect ethnic split). Racial Faces in-the-Wild (RFW) [50] is a verification database that has been constructed from MS-Celeb-1M. Currently, it serves a standard fairness benchmark. Another verification dataset BFW has been introduced in [42], which, similarly to RFW, contains eight subgroups balanced across gender and ethnicity. The main feature of all these datasets – predefined ethnic split – allows one to disambiguate dataset bias and model bias for more precise methods evaluation. Still, the currently available data possesses a number of limitations. For instance, the number of distinct people is typically limited and is often many orders of magnitude smaller than the dataset size. Additionally, both identification and verification datasets require many images per person, which usually restricts the construction of open-source datasets to the search of celebrity pictures by text queries. Our two photo collections, released to the public, – AfricanFaceSet-5M and AsianFaceSet-3M, – fill a different gap in the space of available datasets. On the one hand, they neither have identity labels nor feature many images per person. On the other hand, these collections are large and focused on groups and conditions that are underrepresented in general face recognition datasets. Our evaluation dataset RB-WebFace, assembled from large-scale WebFace-42M, is a new verification dataset containing a significantly larger number of pairs and does not use external models to select the negatives (which introduces certain selection bias) compared to RFW, which is currently used as the main benchmark in the bias mitigation-focused branch of works.", + "bbox": [ + 75, + 367, + 472, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data augmentation. Generating synthetic data is a possi-", + "bbox": [ + 76, + 885, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ble solution to improving the performance of face recognition and in some cases reducing the negative effect of the dataset bias. An idea of this kind is pursued in [37], a nonlinear 3DMM texture model is proposed to produce sharp renderings of faces from novel poses. This technique improves generalization with respect to the head pose and illumination. A number of approaches synthesize data via face generative models. For instance, in SynFace [39] random face images are constructed by a GAN with identity control, and labels are constructed by a procedure similar to MixUp augmentation [55]. Similarly, in [18] a style transfer GAN is used to simultaneously transfer multiple facial demographic attributes and generate diverse images for each attribute class. Authors of Virface [31] suggested a method for incorporating additional negative (impostor) pairs from the unlabeled data showing a boost in metrics. Zhang et al. [56] study the applicability of the data generated by StyleGAN by inspecting the distributions of downstream models but do not study its effect on face recognition metrics.", + "bbox": [ + 496, + 367, + 893, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach is inspired by the latter idea of generating synthetic samples but takes a step forward by utilizing the generative model itself (StyleGAN with an encoder in our case). As shown in [41, 45, 2, 3], the StyleGAN encoders have high potential for both generating new realistic faces from a latent code and solving the inversion problem. This highlights the expressiveness of these models and richness of their internal representations.", + "bbox": [ + 496, + 656, + 893, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised pretraining. Despite the general dominance of supervised learning in practical ML and CV, approaches based on self-supervised learning have been evolving in various forms, such as self-organizing or siamese networks [4, 5, 9, 24]. Currently, self-supervised learning is the dominant approach in NLP with a wide spectrum of possible approaches [35, 12, 11] and is being actively adapted in computer vision [16].", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20926", + "bbox": [ + 478, + 945, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/260341261875363e7bea12e74e27db49afbc07d0094fd2608d21960a0aab139e.jpg", + "image_caption": [ + "Figure 3: Our data collection process starts by manually specifying a set of YouTube channels with a specific topic, e.g. a set of news channels of a desired town or part of the world. All videos are downloaded in the highest available quality, one frame per $P$ seconds is extracted from each video, and all the faces found in the extracted frames are cropped, aligned by landmarks and resized to the target resolution. This way, we obtain millions of random faces with a desired demographic distribution." + ], + "image_footnote": [], + "bbox": [ + 84, + 103, + 883, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analogously, we are witnessing the first attempts to apply self-supervised learning now being integrated into the face recognition frameworks. In 3D-BERL [26], the performance over multiple benchmarks is improved by a separate 3D reconstruction network branch. The work [33] studies the effect of self-supervised learning for domain transfer in face recognition. Incremental learning [52] can be bridged with self-supervised approaches to adapt to large number of target classes [59]. The procedure for collecting large-scale unlabeled datasets allows us to adapt self-supervised training in a more conventional fashion while proving its efficacy for our application.", + "bbox": [ + 75, + 305, + 472, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 494, + 169, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our pipeline is comprised of several stages. First, we train StyleGAN2-ADA to fit the face distribution on the unlabeled dataset. Second, the pSp encoder is trained (also on the unlabeled dataset), that will define a feature extractor well-suited for the group of interest. Finally, the encoder is fine-tuned for the downstream face recognition task. The entire procedure is outlined in Subsec. 3.2 and visually described in Fig. 2. Before describing the method itself, we outline the schematics of our prior dataset collection procedures in Subsec. 3.1.", + "bbox": [ + 75, + 521, + 470, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Prior dataset collection", + "text_level": 1, + "bbox": [ + 76, + 681, + 290, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the nature of the task, the prior dataset must contain the samples from the group of interest. At the same time, samples from the prior dataset do not require identity labels of any kind, unlike samples of the face recognition datasets. In practice, this removes the restriction of linking photos of the same person (something that might be considered a violation of privacy), and thus enlarges the search space and simplifies data collection. Still, fulfilling several requirements for the prior dataset remains a challenge, such as: collecting faces only from a specific group (e.g. an ethnic or gender group), obtaining a large number of them (preferably, an order of magnitude more than in the face recognition dataset to perform subsequent transfer", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "learning), and using only data legally allowed for collection. We found semi-automatic YouTube channel scraping to be an efficient solution that satisfies these requirements. In particular, we propose to select a set of publicly available YouTube channels dedicated to the desired group of interest. The channel names are the only entry point and the only manually performed step for the scraping procedure (see Fig. 3). Typically, the requirement for the channels is having them systematically featuring different people; a possible example would be a set of news channels released in a country of choice. All videos are downloaded from every channel, one frame per $P$ seconds of each video is extracted, and all faces are detected, cropped and aligned by landmarks via MTCNN [57] library. Our data has been scraped from a predefined set of channels such as news channels and others.", + "bbox": [ + 496, + 305, + 893, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The unlabeled prior dataset will later be referred to as $\\mathcal{D}^{prior}$ . Table 1 describes the relative difference of the collected data to the datasets typically used for face recognition. Note however, that despite the latter being directly inapplicable for training due to the lack of the labels, they potentially contain more people than the others.", + "bbox": [ + 496, + 547, + 893, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Architecture and the training procedure", + "text_level": 1, + "bbox": [ + 498, + 650, + 844, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The central part of our pipeline is a single face recognition convnet $f_{\\theta, \\psi}(I)$ (a backbone), that takes a single RGB image $I$ as an input, and outputs a 512-dimensional vector $e \\in \\mathbb{R}^{512}$ (face embedding). Typically, the backbone is trained on a training dataset with angular margin based losses, such as SphereFace [34], ArcFace [13], and others, in standard works on face recognition. In our pipeline, we follow the same procedure, but only after a pretraining stage is performed. In order to do so, we first train the StyleGAN2-ADA [29] generator $g_{\\phi}(w(z))$ that transforms a latent vector $z \\in \\mathcal{Z} \\subseteq \\mathbb{R}^{512}$ into unfolded latent space $w(z) \\in \\mathcal{W} + \\subseteq \\mathbb{R}^{L \\times 512}$ , and then into a realistic face image $\\hat{I}$ . The generator $g_{\\phi}(w(z))$ is trained together with a discriminator that allows the generator to learn the distribution of faces in $D^{prior}$ . The training procedure for StyleGAN2-", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20927", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ADA follows the one in the corresponding paper. Secondly, we introduce a network $f_{\\theta, \\omega}'(x)$ , which is trained as an encoder for the StyleGAN generator, i.e. a network that solves an inverse problem: given an RGB image $I$ , predict an unfolded latent code $w \\in \\mathcal{W}+$ , such that the corresponding reconstruction $\\hat{I} = g_{\\phi}(w)$ is as close to the input $I$ as possible. A set of different possible approaches has appeared recently for training an encoder for StyleGAN2, ranging from applying a ConvNet with style-predicting layers [41] to employing hyper-networks [2]. In our method, the encoder architecture follows the pSp method [41], that proposes predicting a latent code by a ConvNet divided into a convolution part (with parameters $\\theta$ ), following a ResNet architecture, and a set of fully-convolutional style predictors - map2style blocks (with parameters $\\omega$ ). The training procedure for the encoder follows the one outlined in [41] with a few variations. Namely, we only keep the fidelity losses (standard L2 and neural-based LPIPS [58]) and disable the identity loss proposed by the authors. During the encoder training, the StyleGAN generator remains frozen in order to fix the latent space. A pair of networks $f_{\\theta, \\omega}'(x)$ and $g_{\\phi}(w(z))$ can be seen as an asymmetric autoencoder (since the number of encoder parameters highly surpasses the number of generator parameters) with the decoder pretrained for a generative task and the encoder subsequently trained for a discriminative (regressive) task. The visual quality and the identity preservation of the reconstruction $\\hat{I}$ indirectly defines the expressiveness of the latent code $l$ predicted by the encoder.", + "bbox": [ + 76, + 90, + 472, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, after the encoder $f_{\\theta,\\omega}'(I)$ is trained, we transfer its convolutional parameters $\\theta$ into the face recognition backbone $f_{\\theta,\\psi}(I)$ , which also comprises a set of new parameters $\\psi$ , corresponding to a fully-connected layer in the end of the network. We repeat the standard training procedure for face recognition, e.g. described in ArcFace [13] and others. The backbone is trained on a face recognition dataset $D^{facerec} = \\{\\mathcal{D}_1, \\dots, \\mathcal{D}_N\\}$ , where each group $\\mathcal{D}_i$ corresponds to a set of images of the same person $\\# i$ . The slight differences in hyperparameters that we found beneficial when used with pretraining and technical details are described in the Supplementary [1].", + "bbox": [ + 75, + 530, + 472, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 729, + 163, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data and evaluation protocol", + "text_level": 1, + "bbox": [ + 76, + 755, + 336, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our system requires two datasets - a labeled face recognition dataset $\\mathcal{D}^{facerec}$ and an unlabeled prior dataset $\\mathcal{D}^{prior}$ . As for the $\\mathcal{D}^{facerec}$ , we employ BUPT-BalancedFace [49] due to its ethnic and gender balance guarantees. Namely, it consists of 1.3 million images belonging to 28K different people, divided into 4 ethnic groups of equal size - 7K African, East Asian, Indian, and Caucasian people each. Other state-of-the-art methods can", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/64a4f9442e2d2ef36f6aa40e39d71dcaebf7423c91310ab57e73719d48d06052.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RFW, accuracy %,↑
Cauc.AfricanAsianIndianAvg ↑Std ↓
ArcFace R-50* [13]96.1893.9893.7294.6794.641.11
CosFace R-50* [47]95.1293.9392.9892.9393.740.89
DebFace* [19]95.9593.6794.3394.7894.680.83
ACNN* [28]96.1294.0093.6794.5594.580.94
PFE* [44]96.3895.1794.2794.6095.110.93
RL-RBN* [49]96.2795.0094.8294.6895.190.63
GAC R-50* [20]96.2794.4094.3294.7794.940.79
Baseline (ArcFace)96.0094.0093.0894.4894.391.06
+ Dprior (African)96.3594.3793.6294.8894.811.00
+ Dprior (Asian)96.3894.6794.0395.0395.030.86
+ Dprior (Afr+Asian)96.5295.0093.9094.9395.090.94
increase+0.52+1.00+0.82+0.45
", + "bbox": [ + 504, + 71, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Comparison of the verification accuracy of the methods on RFW validation set. Asterisk “*” indicates that the numbers are directly taken from the respective table in [20]. In case of the ArcFace R-50 baseline, it also corresponds to a slightly different reported training procedure. For other methods, the epoch with the best RFW African accuracy score is taken. For GAC [20], we consider the experiment with Estimated ethnic labels in order to compare in the same setting when ground truth labels are not given as an input to the model. The pretraining scheme is efficient when applied to the common choice of the FR method (ArcFace) and the best quality increase is for the ethnicities seen during pretraining (African and Asian).", + "bbox": [ + 496, + 244, + 893, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "only take advantage of a face recognition dataset, and compare with them on BUPT-BalancedFace.", + "bbox": [ + 496, + 459, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prior dataset $\\mathcal{D}^{prior}$ for our work has been collected from YouTube via the procedure outlined in Subsec. 3.1. The dataset consists of two parts - African and East Asian ethnic groups - corresponding to two of the four benchmarks, these groups demonstrate higher error rates for face recognition systems compared to others in the branch of works [20, 49, 19]. Analogously, these two groups are known to be challenging according to NIST evaluation on demographics [22] and benchmarks of other tasks [6]. As an input for the scraping routine, we have used a set of 30-40 news channels corresponding to the respective part of the world. One frame per $P = 5$ seconds was taken, and only frames from the start through the 20th minute of each video have been considered. Since the target resolution for our network training is $112 \\times 112$ , we only approve face bounding boxes selected by MTCNN if the face occupies at least $100 \\mathrm{px}$ on each side. Additionally, we increase the decision thresholds for MTCNN to reduce the number of false positives during the detection process. The exact set of the news channels is enlisted in the Supplementary [1]. In total, 5 million images for the African group (AfricanFaceSet-5M) and 3 million images for the East Asian group (AsianFaceSet-3M) have been collected.", + "bbox": [ + 496, + 491, + 892, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One face recognition validation protocol is based on a balanced verification set Racial Faces In-the-Wild (RFW) [50] (following [53, 20, 19, 49]). It consists of 3K positive pairs of samples (pairs of images of the same per", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20928", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/804a7d1c1a168d040f257b14d55d16e8dbb9ad7224969f03acbb3f95b7d6514a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
ArcFace89.8686.7394.3193.82
ArcFace + Dprior(African)92.6290.2096.3195.85
ArcFace + Dprior(Asian)92.9190.6196.1595.83
ArcFace + Dprior(Afr+Asian)93.5491.5196.5196.17
increase+3.68+4.78+2.20+2.35
SphereFace92.0889.5195.5295.02
SphereFace + Dprior(Afr+Asian)92.8390.5296.2295.65
increase+0.75+1.01+0.70+0.63
GAC92.4889.1396.3394.66
GAC + Dprior(Afr+Asian)93.6090.8096.7695.46
increase+1.12+1.67+0.43+0.80
", + "bbox": [ + 81, + 88, + 467, + 263 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6afa4758851964cc581f0f262759cec2476b7f9d43f80f8b8cf4221b836b1dd.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Comparison of the methods on the newly assembled RB-WebFace validation set. Publicly available authors' implementation of GAC [20] has been used to retrain the method and evaluate the quality. Here, we showcase the TPR given two pre-selected FPR thresholds to highlight the error rate difference separately for each ethnic group. For GAC, the setting GAC (Estimated) (see [20]) is used. This table indicates how the proposed pretraining scheme provides an additional boost for different pipelines." + ], + "table_body": "
TPR @ FPR=10-4↑
ArcFace81.4876.8087.4786.89
ArcFace + D prior (African)85.5681.9991.4290.67
ArcFace + D prior (Asian)85.7982.4191.1290.36
ArcFace + D prior (Af+Asian)86.8483.7491.7491.19
increase+5.36+6.94+4.27+4.30
SphereFace83.6380.0489.1089.07
SphereFace + D prior (Af+Asian)85.1481.4790.6590.16
increase+1.51+1.43+1.55+1.09
GAC85.6480.2491.7288.12
GAC + D prior (Af+Asian)86.9182.4792.3989.48
increase+1.27+2.23+0.67+1.36
", + "bbox": [ + 81, + 266, + 465, + 422 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/393219b238b78ed44dd0e5b47d944a16bfab1348d9dfee851a7426654498f459.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4: Ablation study of the quality dependence on the $\\left| {\\mathcal{D}}_{\\text{prior }}\\right|$ prior dataset size on the newly assembled RB-WebFace validation set. ${\\mathcal{D}}_{\\text{prior }} = \\left\\{ \\begin{array}{ll} \\text{AfricanFaceSet - } 5\\mathrm{M} & \\cup \\\\ \\text{AsianFaceSet - } 3\\mathrm{M} & \\end{array}\\right\\}$ was used in this set of experiments." + ], + "table_body": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ 1% Dprior92.5990.4496.0695.72
+ 10% Dprior92.8490.7296.1295.64
+ 100% Dprior93.5491.5196.5196.17
", + "bbox": [ + 120, + 577, + 426, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "son) and 3K negative pairs (pairs of images of different but similarly looking people, selected by a face recognition algorithm which can introduce selection bias) for each ethnic group, summing up to 24K pairs. Evaluation is carried out in an LFW-like protocol [27] that involves evaluating the backbone $f_{\\theta,\\psi}(I)$ on images of each pair and thresholding the resulting cosine distances between embeddings. Each set of 3K+3K pairs is constructed from 3K people comprising a subset of MS-Celeb-1M [23].", + "bbox": [ + 75, + 750, + 467, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to that, we propose a new test set $RB$", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WebFace constructed in a similar fashion from the recently proposed million-scale identification dataset WebFace-42M (a cleaned version of WebFace-260M [60]). RB-WebFace is constructed by evaluating a pretrained ethnic group classifier on WebFace-42M to separate it into four ethnic groups (the group assignment is later refined by a consensus procedure). For each group, the largest possible number of distinct people is taken to construct positive and negative pairs. The details are provided in Supplementary [1], as well as the example pictures and comparison to typically used test datasets by the number of people and pairs.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 268, + 619, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with the state-of-the-art on RFW. The Table 2 contains the comparison to baseline (ArcFace R-50) and a few state-of-the-art methods (GAC [20], DebFace [19], and others). RFW verification accuracy is reported for each of the four ethnic groups. We demonstrate an increase in accuracy for all races compared to the baseline, which is on par with the state-of-the-art and outperforming on Caucasian and Indian ethnic groups. The largest increase w.r.t. the baseline is observed for the African group. It is important to note that our pretraining scheme can also be used to initialize other methods and enhance their results in a similar fashion.", + "bbox": [ + 496, + 292, + 890, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on newly assembled RB-WebFace test set. Since the number of positive and negative pairs is orders of magnitude different, we report the ROC curves values (TPR vs. FPR), better suited for the class-imbalanced evaluations, instead of reporting accuracy as for RFW. By sweeping the threshold (in [0.1, 0.75] range), we obtain the (TPR, FPR) pairs, and TPR @ FPR = $\\{10^{-3}, 10^{-4}\\}$ is reported in Table 3. Our method (denoted as Baseline + Dprior) is compared to the Baseline (ArcFace R-50). Additionally, we report the enhancement that our procedures provide to other methods [34, 20]. The main conceptual result that we demonstrate is the significant increase of the TPR vs. given FPR for all races and decrease of ethnic bias on RB-WebFace. A separate plot in the Supplementary [1] comprises the TPR measurements for a wide range of possible FPR values. We also provide the results for ResNet $\\{34,100\\}$ backbones in Table 6.", + "bbox": [ + 496, + 476, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limited labeled data. The study in Fig. 4 describes the dependency of our method's accuracy on the number of samples in $\\mathcal{D}^{facerec}$ . We demonstrate the difference between our method and the ArcFace R-50 baseline for each of the benchmarks, both on RFW and RB-WebFace. As expected, pretraining helps the most when the network is fine-tuned on more limited amounts of labeled data. This especially highlights the benefits of using self-supervised learning in these scenarios.", + "bbox": [ + 496, + 733, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limited prior dataset. An ablation in Table 4 describes the quality dependence on the size of the prior dataset. As", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20929", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/679df20abf3b04148a9bedefbaaf1a419a40a9bf10025e853a11d56d222e9057.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 88, + 209, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab3cdb9b485f32bdecbd869a5b91605316e6b06e1cb47adfc045fd308a795566.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 339, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/559836407d0f87f2ba7c894306097d7813976c9b06e5e89244f1cce1ff9dc4a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 88, + 480, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d702ba6bd0d22c523a6c8ecdcaf49c91f560cc772924dbc94eeec6c784dae7da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 88, + 612, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ec474258f7e2d6354bedb891ed6eaecacff9f7f758a4c964be8bbd46c62bfd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 88, + 754, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcda3b971923562fdc3c02e20e2652797cf347073ed4b04c07f4b41bc792c4a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 88, + 888, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79224daf0f3b09991b8b03ceb38d5d6491edfa32fbaaeb76e7829d657f008430.jpg", + "image_caption": [ + "Figure 4: Ablation study of the dependence of the test quality, evaluated for RFW and RB-WebFace datasets, on the number of labeled samples from $|\\mathcal{D}^{facerec}|$ used for fine-tuning. In this experiment, $\\mathcal{D}^{prior} = \\{\\text{AfricanFaceSet-5M} \\cup \\text{AsianFaceSet-3M}\\}$ , while $\\mathcal{D}^{facerec} = \\text{BUPT-BalancedFace}$ . For $\\mathcal{D}^{facerec}$ , fraction of the data (100%, 10%, 1%) defines the number of uniformly sampled people left in the dataset. In the plot legend, Baseline is ArcFace R-50." + ], + "image_footnote": [], + "bbox": [ + 76, + 205, + 205, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12a3da1131f8093263c2a04ab381ff411ba4638ba43fc73b567f50da2950d7c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 205, + 341, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b48df4ab325a9637731ad7e84cc7c45627f540538914bb2585b4f0b55195fcca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 205, + 480, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee33af30be71be7c5596f67250025d79e4570f886631eb597c53227d06925c11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 205, + 612, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa01872f2da9baec2dd3da6bd37eda04b364e57cf803ef9de2ad867bd8dc0c12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 205, + 754, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/931351f4d3e0e2e80d60d79d581c3ff8f61c982f5353596a6f0fed24c3317d17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 205, + 888, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shown, gradually enlarging it results in a monotonous performance increase for all ethnic groups. This indicates that the use of large amounts of diverse unlabeled data significantly strengthens the method in the end.", + "bbox": [ + 75, + 406, + 468, + 467 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e3b1b8ffbe287ed315370f1c3af212102752d7947db2a2f1882ebdf05d42fe2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ AE on Dprior92.9390.8395.7395.68
+ VAE on Dprior92.4590.3995.6495.40
+ ours on Dprior93.5491.5196.5196.17
", + "bbox": [ + 109, + 474, + 436, + 570 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/442fe9c8bbc7ead18a507a98e3bdf7baa7094900351f5406b9f812d9a600e568.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TPR @ FPR=10-4↑
Baseline81.4876.8087.4786.89
+ AE on Dprior85.8583.0990.4689.86
+ VAE on Dprior85.1982.0890.3889.25
+ ours on Dprior86.8483.7491.7491.19
", + "bbox": [ + 109, + 578, + 436, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can StyleGAN be replaced with a simpler encoder-decoder architecture? In Table 5, we compare our pretraining procedure to more conventional ones where the ResNet-50 encoder is pretrained in a symmetric autoencoder (AE) or in a symmetric variational autoencoder (VAE) setting. While these models can be pretrained in a single stage, our proposed StyleGAN+encoder approach re", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8fa7838bace5caec5484773e22ae68182f32ded4118c5517b3ca8f83441cad2.jpg", + "table_caption": [ + "Table 5: Comparison of different pretraining strategies. Along with ours based on consequent training of StyleGAN and a ResNet encoder, one could employ vanilla (AE) and variational (VAE) autoencoders for pretraining (with the same ResNet-50 architecture as an encoder). Despite simpler simultaneous training of an encoder and a decoder in a single stage, the StyleGAN-based procedure is a more powerful prior. Baseline refers to the ArcFace R-50." + ], + "table_footnote": [], + "table_body": "
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
R-3484.5781.4791.1589.9173.6069.2383.1180.40
+ D prior90.9288.3495.2794.3983.0878.6989.2587.98
R-5089.8686.7394.3193.8281.4876.8087.4786.89
+ D prior93.5491.5196.5196.1786.8483.7491.7491.19
R-10093.7791.5596.3296.6287.2884.2891.4691.66
+ D prior94.8293.0096.7096.9288.7686.0092.2492.70
", + "bbox": [ + 504, + 404, + 888, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6: Comparison of different ResNet backbones on RB-WebFace. The baseline is ArcFace and $D^{prior} =$ AfricanFaceSet-5M U AsianFaceSet-3M.", + "bbox": [ + 498, + 515, + 893, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "quires two separate pretraining stages but demonstrates better performance, which motivates the use of state-of-the-art generative models for face recognition.", + "bbox": [ + 496, + 577, + 890, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Why is AE/VAE less suitable than StyleGAN when finetuned for face recognition? We see two possible reasons why our procedure yields better results. First, StyleGAN-based architectures are tailored to be the state-of-the-art of face generation, which leads to the possible assumption that encoder + StyleGAN pipeline is capable of saving more useful information about the face features than e.g. VAE. Second, we observe that the quality of StyleGAN generations significantly improves when it is trained on larger amount of data, while for VAE it is not the case (6% over 32% improvement when 100x more data given – see Fig. 5). This scalability issue is also highlighted by the fact that AE/VAE, trained on the 100% of $\\mathcal{D}^{prior}$ , perform for the face recognition task on par with StyleGAN, trained on only 1% of $\\mathcal{D}^{prior}$ (see the Tables 5 and 4).", + "bbox": [ + 496, + 625, + 892, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is it helpful to draw samples from StyleGAN instead of training an encoder for it? In this experiment, we used our trained StyleGAN to generate synthetic faces and add", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20930", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/263ae190c8fbae7922cee2c6c12b891a16d7e70e39fcc6566de50fe50d8bce0c.jpg", + "image_caption": [ + "Figure 5: FID score, calculated over 100K random pictures from $\\mathcal{D}^{prior} = \\{\\text{AfricanFaceSet-5M} \\cup \\text{AsianFaceSet-3M}\\}$ vs. 100K random samples from VAE or StyleGAN. The score (the less the better) is given w.r.t. the percentage of $\\mathcal{D}^{prior}$ used to train these models. We observe that the image generation quality, that FID represents, drops significantly when StyleGAN is provided with more data, while VAE does not scale to larger data sets that well." + ], + "image_footnote": [], + "bbox": [ + 145, + 78, + 379, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "them to the training set. We first infer StyleGAN latent codes for all African and Asian samples in our training set (BUPT-BalancedFace). Since face recognition training always requires labeled images, we generate samples using latent codes closer to the inverted training images. In particular, we take random pairs $(I_1, I_2)$ of images of the same ethnic group (either African or Asian) and generate an interpolated sample by calculating convex combinations of their StyleGAN latent codes: $I_{\\mathrm{comb}} = \\lambda \\cdot f_{\\theta, \\omega}'(I_1) + (1 - \\lambda) \\cdot f_{\\theta, \\omega}'(I_2)$ , $\\lambda \\sim U[0, 1]$ . The ground-truth label (always required for face recognition training) is constructed as a two-hot vector $(0, \\dots, 0, 1 - \\lambda, 0, \\dots, 0, \\lambda, 0, \\dots, 0)$ (as influenced by MixUp [55]). In total, we augment the training set with an equal amount of synthetic interpolations (1.3 M), obtained from randomly drawn African-African and Asian-Asian pairs. The results in Table. 7 show that the added interpolations, despite the created imbalance and increased variability in the data, do not yield a comparable performance increase. In addition, we show that increasing the number of interpolations doesn't help, while pretraining scheme benefits from more pretraining data (see Table 4).", + "bbox": [ + 75, + 433, + 468, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide additional ablations and comparisons in the Supplementary [1].", + "bbox": [ + 76, + 752, + 468, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 799, + 204, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As an increasing number of modern computer vision methods become production-ready, new challenges regarding their final use are posed. In this work, we present a solution for improving the quality of facial recognition by pretraining the face image encoder with unlabeled data using", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/345384b233806c0dde4b6e94ce95474622764c239272bf210fd5705ca907f90c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
ArcFace R-50 + Dprior89.8686.7394.3193.8281.4876.8087.4786.89
93.5491.5196.5196.1786.8483.7491.7491.19
+ 1.3M interps91.8789.0796.0895.5584.6579.9791.1390.30
+ 6M interps89.4886.8594.9294.2280.9577.0088.7587.94
", + "bbox": [ + 504, + 88, + 888, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7: In this experiment, instead of pretraining our model as an encoder for StyleGAN, we use the StyleGAN and the trained encoder to draw samples from the learned face distribution and add them to our dataset. The results indicate that interpolation-based procedure, despite being more straightforward than the proposed pretraining-based scheme, is actually less efficient.", + "bbox": [ + 498, + 171, + 890, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "StyleGAN and encoder. Furthermore, we additionally release two training datasets for unsupervised pretraining and one large-scale protocol for bias estimation. We show that we are able to tune the performance on different ethnicities by altering the composition of the unlabeled prior datasets. We hope that the released datasets, together with the protocol, will drive forward the research on mitigating the face recognition biases using self-supervised approaches.", + "bbox": [ + 496, + 297, + 890, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are several directions for possible future work that we foresee. First, integrating the two steps of pretraining and fine-tuning into one would help to avoid forgetting of the pretrained weights. Second, publicly available datasets often don't possess enough variability (both in terms of capture conditions and demographics) [25] and typically feature only low-resolution images which might be a blocking factor for the fairness research [49, 30]. Accordingly, the results on the newly assembled RB-WebFace benchmark can be analyzed more thoroughly, which can bring new insights about the factors positively affecting bias mitigation in real-world scenarios. Finally, various architectures typically used for self-supervised learning (such as transformers or highly scalable generative models [43]), might allow for the construction of more flexible methods due to heterogeneous inputs and outputs.", + "bbox": [ + 496, + 424, + 892, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Legal concerns. The collected unlabeled data was collected anonymously in accordance with Standard YouTube License and CC YouTube License and does not contain personally identifiable information (PII). The data is released only in the form of links to YouTube videos and the corresponding timestamps, fully following good practice of similar data collection [10, 17, 32] and with the data subjects protection as per Art. 14 5(b) GDPR law.", + "bbox": [ + 496, + 669, + 890, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We gratefully acknowledge the support of this research by a TUM-IAS Hans Fischer Senior Fellowship, the ERC Starting Grant Scan2CAD (804724) and the Horizon Europe vera.ai project (101070093). We also thank Yawar Siddiqui for the helpful technical advice on StyleGAN part, Angela Dai for the video voiceover, and Dmitriy Karfagenskiy for the proofreading and corrections.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20931", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] How to boost face recognition with stylegan?. Supplemental material, 2022.", + "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711-6720, 2021.", + "[3] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[4] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992.", + "[5] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993.", + "[6] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018.", + "[7] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGface2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face and Gesture Recognition, pages 3414-3424, 2018.", + "[8] Yunliang Chen and Jungseock Joo. Understanding and mitigating annotation bias in facial expression recognition. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14980-14991, 2021.", + "[9] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 539-546, 2005.", + "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In Interspeech, 2018.", + "[11] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020.", + "[12] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In Proc. of the 58th Annual Meeting of the Association for Computational Linguistics, 2020.", + "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019.", + "[14] Emily Denton, Ben Hutchinson, Margaret Mitchell, Timnit Gebru, and Andrew Zaldivar. Image counterfactual sensitivity analysis for detecting unintended bias. arXiv preprint arXiv:1906.06439, 2019." + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, 2019.", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[17] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018.", + "[18] Markos Georgopoulos, James Oldfield, Mihalis A Nicolaou, Yannis Panagakis, and Maja Pantic. Mitigating demographic bias in facial datasets with style-based multi-attribute transfer. International Journal of Computer Vision, 129:2288-2307, 2021.", + "[19] Sixue Gong, Xiaoming Liu, and Anil K Jain. Debface: Debiasing face recognition. arXiv preprint arXiv:1911.08080, 2019.", + "[20] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating Face Recognition Bias via Group Adaptive Classifier. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3414-3424, 2021.", + "[21] Google. Real Tone on Pixel 6. Be seen as you truly are., 2022.", + "[22] Patrick Grother, Mei Ngan, and Kayee Hanaoka. Face recognition vendor test (FVRT): Part 3, demographic effects. 2019.", + "[23] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-Celeb-1M: A dataset and benchmark for large-scale face recognition. In European Conference on Computer Vision (ECCV), 2016.", + "[24] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, pages 1735-1742, 2006.", + "[25] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in AI: the Casual Conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021.", + "[26] Mingjie He, Jie Zhang, Shiguang Shan, and Xilin Chen. Enhancing face recognition with self-supervised 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4062-4071, 2022.", + "[27] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008.", + "[28] Di Kang, Debarun Dhar, and Antoni Chan. Incorporating side information by adaptive convolution. Advances in Neural Information Processing Systems, 30, 2017." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20932", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020.", + "[30] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Image resolution susceptibility of face recognition models. arXiv preprint arXiv:2107.03769, 2021.", + "[31] Wenyu Li, Tianchu Guo, Pengyu Li, Binghui Chen, Biao Wang, Wangmeng Zuo, and Lei Zhang. Virface: Enhancing face recognition via unlabeled shallow data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14729-14738, 2021.", + "[32] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3207-3216, 2020.", + "[33] Chun-Hsien Lin and Bing-Fei Wu. Domain adapting ability of self-supervised learning for face recognition. In IEEE International Conference on Image Processing (ICIP), pages 479–483, 2021.", + "[34] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 212-220, 2017.", + "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv:1907.11692, 2019.", + "[36] Vongani Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying Bias in GANs through the Lens of Race. In European Conference on Computer Vision (ECCV), pages 344-360, 2022.", + "[37] Richard T Marriott, Sami Romdhani, and Liming Chen. A 3d gan for improved large-posed facial recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13445–13455, 2021.", + "[38] Tuan-Duy H Nguyen, Huu-Nghia H Nguyen, and Hieu Dao. Recognizing families through images with pretrained encoder. In IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 887-891, 2020.", + "[39] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10880-10890, 2021.", + "[40] Christian Rathgeb, Pawel Drozdowski, Naser Damer, Dinusha C Frings, and Christoph Busch. Demographic fairness in biometric systems: What do the experts say? IEEE Technology and Society Magazine, 41(4):71-82, 2022.", + "[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021.", + "[42] Joseph P Robinson, Gennady Livitz, Yann Henon, Can Qin, Yun Fu, and Samson Timoner. Face Recognition: Too Bias," + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "or Not Too Bias? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020.", + "[43] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH Conference Proceedings, pages 1-10, 2022.", + "[44] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6902-6911, 2019.", + "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021.", + "[46] Fu-En Wang, Chien-Yi Wang, Min Sun, and Shang-Hong Lai. Mixfairface: Towards ultimate fairness via mixfair adapter in face recognition. In AAAI Conference on Artificial Intelligence, volume 37, pages 14531-14538, 2023.", + "[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5265-5274, 2018.", + "[48] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In ACM International Conference on Multimedia, pages 3779-3782, 2021.", + "[49] Mei Wang and Weihong Deng. Mitigating Bias in Face Recognition using Skewness-Aware Reinforcement Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9322-9331, 2020.", + "[50] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial bias by information maximization adaptation network. In International Conference on Computer Vision (ICCV), pages 692-702, 2019.", + "[51] Qingzhong Wang, Pengfei Zhang, Haoyi Xiong, and Jian Zhao. Face evolves: A high-performance face recognition library. arXiv preprint arXiv:2107.08621, 2021.", + "[52] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019.", + "[53] Xingkun Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, Jilin Li, Feiyue Huang, Yong Li, and Zhen Cui. Consistent Instance False Positive Improves Fairness in Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 578-586, 2021.", + "[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv:1411.7923, 2014.", + "[55] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018.", + "[56] Haoyu Zhang, Marcel Grimmer, Raghavendra Ramachandra, Kiran Raja, and Christoph Busch. On the applicability of synthetic data for face recognition. In IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20933", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10):1499-1503, 2016.", + "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018.", + "[59] Song Zhang, Gehui Shen, Jinsong Huang, and Zhi-Hong Deng. Self-supervised learning aided class-incremental life-long learning. arXiv preprint arXiv:2006.05882, 2020.", + "[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + ], + "bbox": [ + 78, + 90, + 470, + 344 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20934", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_model.json b/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..97495a20029b86d0434a702c2bfc58f95913727c --- /dev/null +++ b/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_model.json @@ -0,0 +1,1982 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.242, + 0.131, + 0.728, + 0.154 + ], + "angle": 0, + "content": "How to Boost Face Recognition with StyleGAN?" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.18, + 0.744, + 0.217 + ], + "angle": 0, + "content": "Artem Sevastopolsky1 Yury Malkov2,* Nikita Durasov3 Luisa Verdoliva1,4 Matthias Nießner1" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.233, + 0.734, + 0.291 + ], + "angle": 0, + "content": "1 Technical University of Munich, Germany 2 Twitter, US 3 École polytechnique fédérale de Lausanne, Switzerland 4 University Federico II of Naples, Italy" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.306, + 0.77, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.452, + 0.895, + 0.528 + ], + "angle": 0, + "content": "Figure 1: Our method is aimed at boosting the performance of face recognition. This is achieved by gathering a random image collection without face recognition labels (unlabeled set) and then fitting a mapping from an image to the StyleGAN latent space onto that collection. To learn this mapping, we use the pSp encoder architecture. For the downstream face recognition task, the same encoder is then fine-tuned on a (potentially much smaller) face recognition dataset with identity labels (labeled set)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.541, + 0.314, + 0.557 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.57, + 0.474, + 0.903 + ], + "angle": 0, + "content": "State-of-the-art face recognition systems require vast amounts of labeled training data. Given the priority of privacy in face recognition applications, the data is limited to celebrity web crawls, which have issues such as limited numbers of identities. On the other hand, self-supervised revolution in the industry motivates research on the adaptation of related techniques to facial recognition. One of the most popular practical tricks is to augment the dataset by the samples drawn from generative models while preserving the identity. We show that a simple approach based on fine-tuning \\( pSp \\) encoder for StyleGAN allows to improve upon the state-of-the-art facial recognition and performs better compared to training on synthetic face identities. We also collect large-scale unlabeled datasets with controllable ethnic constitution - AfricanFaceSet-5M (5 million images of different people) and AsianFaceSet-3M (3 million images of different people) - and we show that pretraining on each of them improves recognition of the respective ethnicities (as well as others), while combining all unlabeled datasets results in the biggest performance increase. Our self-supervised strategy is the most useful with limited amounts of labeled training data," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.543, + 0.895, + 0.633 + ], + "angle": 0, + "content": "which can be beneficial for more tailored face recognition tasks and when facing privacy concerns. Evaluation is based on a standard RFW dataset and a new large-scale RB-WebFace benchmark. The code and data are made publicly available at https://github.com/seval00/stylegan-for-facerec." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.649, + 0.633, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Modern face recognition methods rely on deep convolutional networks trained on large-scale datasets [54, 7, 23, 60]. These methods are now being integrated into a vast number of real-world applications, ranging from face unlock for smartphones and photo organizers to law enforcement systems and border control. A typical open face recognition dataset consists of web-crawled images of celebrities, leading to limited size and lack of balance in subgroups, such as ethnicity, age, etc. Training a state-of-the-art solution, however, requires enormous amounts of labeled data, scraping which may lead to privacy and legal issues. We suggest and study an alternative solution to using celebrity photos - pretraining the face recognition backbone on a generative task. Specifically, we first train StyleGAN2-ADA [29] on collected" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.337, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.091, + 0.668, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.092, + 0.89, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Figure 2: Our method is trained in three consecutive steps. First, we fit StyleGAN2-ADA to the face image distribution of the unlabeled prior dataset \\( \\mathcal{D}^{prior} \\). Second, the pSp encoder is trained (also on \\( \\mathcal{D}^{prior} \\)) to map images to the latent codes in the learned latent space. Finally, the encoder, which is pretrained to extract meaningful features from an image, is fine-tuned for the downstream face recognition task with the ArcFace loss (similar losses can be used instead) on \\( \\mathcal{D}^{facerec} \\). The two first steps comprise the self-supervised pretraining stage; i.e., no identity labels are required for them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.471, + 0.504 + ], + "angle": 0, + "content": "unlabeled data (which we later refer to as an unlabeled prior dataset) to fit the face image distribution. Subsequently, we train an encoder (following pixel2style2pixel (pSp) architecture [41]) that maps input images to vectors in the learned StyleGAN2-ADA latent space. Importantly, during the pretraining steps, no identity labels are used, so we can use diverse datasets crawled from the Internet without compromising privacy. Finally, we transfer the learned pSp encoder convolutional weights into the face recognition network and train it in a standard face recognition setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.748 + ], + "angle": 0, + "content": "We show that, in contrast to training face recognition tasks on StyleGAN generated data (also demonstrated in [38, 39] and studied e.g. in [36]), our encoder pretraining step significantly boosts the final performance. The idea of augmenting face recognition datasets with synthetic data is widespread and constitutes many approaches, however, unclear and heuristic definition of the target label limits the amount of useful signal that can be transferred into the face recognition model this way. Our approach goes hand-in-hand with the current development of self-supervised learning [15, 35, 12, 16] and makes up one of the first approaches of its application to face recognition [26, 33]. This allows us to demonstrate vast improvements on limited labeled training data compared to the setup without self-supervised pretraining (for instance, \\(10\\%\\) verification accuracy increase for only \\(1\\%\\) of the labeled data used)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The simplicity of the data collection procedure also allows us to control the distribution of the unlabeled data and thus influence the decrease of the error rates for specific demographic groups. Despite the fact that the current state-of-the-art algorithms often demonstrate very low average error rates [13, 47, 51, 48], it is considered unethical to integrate face recognition solutions that exhibit significant ethnic, age, or gender bias. Such bias is present both in open-source face recognition methods [20, 53, 49, 19, 28, 44, 46] and in comprehensively evaluated commercial face recog" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.428 + ], + "angle": 0, + "content": "nition systems [22], resulting in significantly different error rates measured for the groups of interest. The topic has attracted significant attention in other areas of computer vision operating in face domain [8, 14, 21, 40] and in the media." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.893, + 0.655 + ], + "angle": 0, + "content": "We constructively demonstrate that collecting large amounts of in-the-wild face images of a given group of interest is feasible (and can be done semi-automatically), while collecting datasets with identity labels is problematic. The labels require linking photos of the same person taken in different conditions, which means the person must be tracked. This typically constrains public datasets to celebrities, gathered using search engines [23, 54, 7, 60], while social networks and companies that provide services with photos use input from users. Second, the collected in-the-wild data, treated as a set of faces without identity labels (but labeled with the group attribution), can be efficiently used for self-supervised pretraining for face recognition networks, subsequently fine-tuned on the standard face recognition datasets (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.762 + ], + "angle": 0, + "content": "To summarize our main contribution, we present a novel self-supervised method for improving the performance of face recognition based on StyleGAN pretraining. This allows to leverage large-scale amounts of available unlabeled data for face recognition. While the improvement is the most significant on limited data, pretraining is also helpful for large-scale labeled datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.637, + 0.785 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Face recognition datasets. Several large-scale datasets of faces with identity labels have been released publicly, such as CASIA-WebFace [54], VGGFace2 [7], MS-Celeb-1M [23] and the very recent million-scale WebFace-42M dataset [60]. However, these datasets have been collected \"in the wild\", hence, they inevitably suffer from biases in terms of age, gender," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20925" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.09, + 0.887, + 0.281 + ], + "angle": 0, + "content": "
Dataset name# people# images# pic./personethnic diversityacquisition
MS-Celeb-1M100K8.2M82unevenmostly American and British actors
VGGFace29.1K3.3M362unevenGoogle Image search
CASIA-WebFace10K494K49unevencelebrities from IMDb
CASIA-Face-Africa1.1K38K34all Africancontrolled, indoor and outdoor
MegaFace (unavailable)672K4.7M7unevenYahoo Flickr website search queries
DiveFace24K72K3balanced across 3 ethnicitiessubset of MegaFace
BUPT-BalancedFace28K1.3M46balanced across 4 ethnicitiescelebrities from MS-Celeb-1M
BUPT-GlobalFace38K2M52matches global populationcelebrities from MS-Celeb-1M
BUPT-TransferFace>10K600K6075% Caucasian. vs. otherscelebrities from MS-Celeb-1M
AfricanFaceSet-5M5M5MunlabeledAfrican majorityrandom faces from YouTube news videos
AsianFaceSet-3M3M3MunlabeledAsian majorityrandom faces from YouTube news videos
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.283, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Table 1: Overview of the publicly available training sets in facial domain. Typically, the quality of a face recognition dataset is described by several factors: the number of people in the dataset, the number of images per person, and the diversity of capture conditions. While our AfricanFaceSet-5M and AsianFaceSet-3M datasets are not directly suitable for face recognition due to the absence of identity labels, they comprise more distinct people than existing large-scale datasets and contain a more diverse distribution of faces than only celebrities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.882 + ], + "angle": 0, + "content": "or race. To better understand and stimulate research in fairness about face recognition, several datasets have been proposed. Examples of such labeled datasets include BUPT-Globalface [49] (2M images with ethnic distribution matching the world population) and BUPT-Balancedface [49] (1.3M images with the perfect ethnic split). Racial Faces in-the-Wild (RFW) [50] is a verification database that has been constructed from MS-Celeb-1M. Currently, it serves a standard fairness benchmark. Another verification dataset BFW has been introduced in [42], which, similarly to RFW, contains eight subgroups balanced across gender and ethnicity. The main feature of all these datasets – predefined ethnic split – allows one to disambiguate dataset bias and model bias for more precise methods evaluation. Still, the currently available data possesses a number of limitations. For instance, the number of distinct people is typically limited and is often many orders of magnitude smaller than the dataset size. Additionally, both identification and verification datasets require many images per person, which usually restricts the construction of open-source datasets to the search of celebrity pictures by text queries. Our two photo collections, released to the public, – AfricanFaceSet-5M and AsianFaceSet-3M, – fill a different gap in the space of available datasets. On the one hand, they neither have identity labels nor feature many images per person. On the other hand, these collections are large and focused on groups and conditions that are underrepresented in general face recognition datasets. Our evaluation dataset RB-WebFace, assembled from large-scale WebFace-42M, is a new verification dataset containing a significantly larger number of pairs and does not use external models to select the negatives (which introduces certain selection bias) compared to RFW, which is currently used as the main benchmark in the bias mitigation-focused branch of works." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Data augmentation. Generating synthetic data is a possi-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.895, + 0.655 + ], + "angle": 0, + "content": "ble solution to improving the performance of face recognition and in some cases reducing the negative effect of the dataset bias. An idea of this kind is pursued in [37], a nonlinear 3DMM texture model is proposed to produce sharp renderings of faces from novel poses. This technique improves generalization with respect to the head pose and illumination. A number of approaches synthesize data via face generative models. For instance, in SynFace [39] random face images are constructed by a GAN with identity control, and labels are constructed by a procedure similar to MixUp augmentation [55]. Similarly, in [18] a style transfer GAN is used to simultaneously transfer multiple facial demographic attributes and generate diverse images for each attribute class. Authors of Virface [31] suggested a method for incorporating additional negative (impostor) pairs from the unlabeled data showing a boost in metrics. Zhang et al. [56] study the applicability of the data generated by StyleGAN by inspecting the distributions of downstream models but do not study its effect on face recognition metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.779 + ], + "angle": 0, + "content": "Our approach is inspired by the latter idea of generating synthetic samples but takes a step forward by utilizing the generative model itself (StyleGAN with an encoder in our case). As shown in [41, 45, 2, 3], the StyleGAN encoders have high potential for both generating new realistic faces from a latent code and solving the inversion problem. This highlights the expressiveness of these models and richness of their internal representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Self-supervised pretraining. Despite the general dominance of supervised learning in practical ML and CV, approaches based on self-supervised learning have been evolving in various forms, such as self-organizing or siamese networks [4, 5, 9, 24]. Currently, self-supervised learning is the dominant approach in NLP with a wide spectrum of possible approaches [35, 12, 11] and is being actively adapted in computer vision [16]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.104, + 0.885, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 3: Our data collection process starts by manually specifying a set of YouTube channels with a specific topic, e.g. a set of news channels of a desired town or part of the world. All videos are downloaded in the highest available quality, one frame per \\(P\\) seconds is extracted from each video, and all the faces found in the extracted frames are cropped, aligned by landmarks and resized to the target resolution. This way, we obtain millions of random faces with a desired demographic distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.488 + ], + "angle": 0, + "content": "Analogously, we are witnessing the first attempts to apply self-supervised learning now being integrated into the face recognition frameworks. In 3D-BERL [26], the performance over multiple benchmarks is improved by a separate 3D reconstruction network branch. The work [33] studies the effect of self-supervised learning for domain transfer in face recognition. Incremental learning [52] can be bridged with self-supervised approaches to adapt to large number of target classes [59]. The procedure for collecting large-scale unlabeled datasets allows us to adapt self-supervised training in a more conventional fashion while proving its efficacy for our application." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.17, + 0.512 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.471, + 0.673 + ], + "angle": 0, + "content": "Our pipeline is comprised of several stages. First, we train StyleGAN2-ADA to fit the face distribution on the unlabeled dataset. Second, the pSp encoder is trained (also on the unlabeled dataset), that will define a feature extractor well-suited for the group of interest. Finally, the encoder is fine-tuned for the downstream face recognition task. The entire procedure is outlined in Subsec. 3.2 and visually described in Fig. 2. Before describing the method itself, we outline the schematics of our prior dataset collection procedures in Subsec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.292, + 0.697 + ], + "angle": 0, + "content": "3.1. Prior dataset collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Due to the nature of the task, the prior dataset must contain the samples from the group of interest. At the same time, samples from the prior dataset do not require identity labels of any kind, unlike samples of the face recognition datasets. In practice, this removes the restriction of linking photos of the same person (something that might be considered a violation of privacy), and thus enlarges the search space and simplifies data collection. Still, fulfilling several requirements for the prior dataset remains a challenge, such as: collecting faces only from a specific group (e.g. an ethnic or gender group), obtaining a large number of them (preferably, an order of magnitude more than in the face recognition dataset to perform subsequent transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.547 + ], + "angle": 0, + "content": "learning), and using only data legally allowed for collection. We found semi-automatic YouTube channel scraping to be an efficient solution that satisfies these requirements. In particular, we propose to select a set of publicly available YouTube channels dedicated to the desired group of interest. The channel names are the only entry point and the only manually performed step for the scraping procedure (see Fig. 3). Typically, the requirement for the channels is having them systematically featuring different people; a possible example would be a set of news channels released in a country of choice. All videos are downloaded from every channel, one frame per \\( P \\) seconds of each video is extracted, and all faces are detected, cropped and aligned by landmarks via MTCNN [57] library. Our data has been scraped from a predefined set of channels such as news channels and others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.895, + 0.641 + ], + "angle": 0, + "content": "The unlabeled prior dataset will later be referred to as \\( \\mathcal{D}^{prior} \\). Table 1 describes the relative difference of the collected data to the datasets typically used for face recognition. Note however, that despite the latter being directly inapplicable for training due to the lack of the labels, they potentially contain more people than the others." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.845, + 0.667 + ], + "angle": 0, + "content": "3.2. Architecture and the training procedure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The central part of our pipeline is a single face recognition convnet \\( f_{\\theta, \\psi}(I) \\) (a backbone), that takes a single RGB image \\( I \\) as an input, and outputs a 512-dimensional vector \\( e \\in \\mathbb{R}^{512} \\) (face embedding). Typically, the backbone is trained on a training dataset with angular margin based losses, such as SphereFace [34], ArcFace [13], and others, in standard works on face recognition. In our pipeline, we follow the same procedure, but only after a pretraining stage is performed. In order to do so, we first train the StyleGAN2-ADA [29] generator \\( g_{\\phi}(w(z)) \\) that transforms a latent vector \\( z \\in \\mathcal{Z} \\subseteq \\mathbb{R}^{512} \\) into unfolded latent space \\( w(z) \\in \\mathcal{W} + \\subseteq \\mathbb{R}^{L \\times 512} \\), and then into a realistic face image \\( \\hat{I} \\). The generator \\( g_{\\phi}(w(z)) \\) is trained together with a discriminator that allows the generator to learn the distribution of faces in \\( D^{prior} \\). The training procedure for StyleGAN2-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20927" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.53 + ], + "angle": 0, + "content": "ADA follows the one in the corresponding paper. Secondly, we introduce a network \\( f_{\\theta, \\omega}'(x) \\), which is trained as an encoder for the StyleGAN generator, i.e. a network that solves an inverse problem: given an RGB image \\( I \\), predict an unfolded latent code \\( w \\in \\mathcal{W}+ \\), such that the corresponding reconstruction \\( \\hat{I} = g_{\\phi}(w) \\) is as close to the input \\( I \\) as possible. A set of different possible approaches has appeared recently for training an encoder for StyleGAN2, ranging from applying a ConvNet with style-predicting layers [41] to employing hyper-networks [2]. In our method, the encoder architecture follows the pSp method [41], that proposes predicting a latent code by a ConvNet divided into a convolution part (with parameters \\( \\theta \\)), following a ResNet architecture, and a set of fully-convolutional style predictors - map2style blocks (with parameters \\( \\omega \\)). The training procedure for the encoder follows the one outlined in [41] with a few variations. Namely, we only keep the fidelity losses (standard L2 and neural-based LPIPS [58]) and disable the identity loss proposed by the authors. During the encoder training, the StyleGAN generator remains frozen in order to fix the latent space. A pair of networks \\( f_{\\theta, \\omega}'(x) \\) and \\( g_{\\phi}(w(z)) \\) can be seen as an asymmetric autoencoder (since the number of encoder parameters highly surpasses the number of generator parameters) with the decoder pretrained for a generative task and the encoder subsequently trained for a discriminative (regressive) task. The visual quality and the identity preservation of the reconstruction \\( \\hat{I} \\) indirectly defines the expressiveness of the latent code \\( l \\) predicted by the encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.473, + 0.713 + ], + "angle": 0, + "content": "Finally, after the encoder \\( f_{\\theta,\\omega}'(I) \\) is trained, we transfer its convolutional parameters \\( \\theta \\) into the face recognition backbone \\( f_{\\theta,\\psi}(I) \\), which also comprises a set of new parameters \\( \\psi \\), corresponding to a fully-connected layer in the end of the network. We repeat the standard training procedure for face recognition, e.g. described in ArcFace [13] and others. The backbone is trained on a face recognition dataset \\( D^{facerec} = \\{\\mathcal{D}_1, \\dots, \\mathcal{D}_N\\} \\), where each group \\( \\mathcal{D}_i \\) corresponds to a set of images of the same person \\( \\# i \\). The slight differences in hyperparameters that we found beneficial when used with pretraining and technical details are described in the Supplementary [1]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.73, + 0.165, + 0.745 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.338, + 0.772 + ], + "angle": 0, + "content": "4.1. Data and evaluation protocol" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our system requires two datasets - a labeled face recognition dataset \\( \\mathcal{D}^{facerec} \\) and an unlabeled prior dataset \\( \\mathcal{D}^{prior} \\). As for the \\( \\mathcal{D}^{facerec} \\), we employ BUPT-BalancedFace [49] due to its ethnic and gender balance guarantees. Namely, it consists of 1.3 million images belonging to 28K different people, divided into 4 ethnic groups of equal size - 7K African, East Asian, Indian, and Caucasian people each. Other state-of-the-art methods can" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.072, + 0.891, + 0.234 + ], + "angle": 0, + "content": "
RFW, accuracy %,↑
Cauc.AfricanAsianIndianAvg ↑Std ↓
ArcFace R-50* [13]96.1893.9893.7294.6794.641.11
CosFace R-50* [47]95.1293.9392.9892.9393.740.89
DebFace* [19]95.9593.6794.3394.7894.680.83
ACNN* [28]96.1294.0093.6794.5594.580.94
PFE* [44]96.3895.1794.2794.6095.110.93
RL-RBN* [49]96.2795.0094.8294.6895.190.63
GAC R-50* [20]96.2794.4094.3294.7794.940.79
Baseline (ArcFace)96.0094.0093.0894.4894.391.06
+ Dprior (African)96.3594.3793.6294.8894.811.00
+ Dprior (Asian)96.3894.6794.0395.0395.030.86
+ Dprior (Afr+Asian)96.5295.0093.9094.9395.090.94
increase+0.52+1.00+0.82+0.45
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Table 2: Comparison of the verification accuracy of the methods on RFW validation set. Asterisk “*” indicates that the numbers are directly taken from the respective table in [20]. In case of the ArcFace R-50 baseline, it also corresponds to a slightly different reported training procedure. For other methods, the epoch with the best RFW African accuracy score is taken. For GAC [20], we consider the experiment with Estimated ethnic labels in order to compare in the same setting when ground truth labels are not given as an input to the model. The pretraining scheme is efficient when applied to the common choice of the FR method (ArcFace) and the best quality increase is for the ethnicities seen during pretraining (African and Asian)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.489 + ], + "angle": 0, + "content": "only take advantage of a face recognition dataset, and compare with them on BUPT-BalancedFace." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.893, + 0.839 + ], + "angle": 0, + "content": "The prior dataset \\( \\mathcal{D}^{prior} \\) for our work has been collected from YouTube via the procedure outlined in Subsec. 3.1. The dataset consists of two parts - African and East Asian ethnic groups - corresponding to two of the four benchmarks, these groups demonstrate higher error rates for face recognition systems compared to others in the branch of works [20, 49, 19]. Analogously, these two groups are known to be challenging according to NIST evaluation on demographics [22] and benchmarks of other tasks [6]. As an input for the scraping routine, we have used a set of 30-40 news channels corresponding to the respective part of the world. One frame per \\( P = 5 \\) seconds was taken, and only frames from the start through the 20th minute of each video have been considered. Since the target resolution for our network training is \\( 112 \\times 112 \\), we only approve face bounding boxes selected by MTCNN if the face occupies at least \\( 100 \\mathrm{px} \\) on each side. Additionally, we increase the decision thresholds for MTCNN to reduce the number of false positives during the detection process. The exact set of the news channels is enlisted in the Supplementary [1]. In total, 5 million images for the African group (AfricanFaceSet-5M) and 3 million images for the East Asian group (AsianFaceSet-3M) have been collected." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "One face recognition validation protocol is based on a balanced verification set Racial Faces In-the-Wild (RFW) [50] (following [53, 20, 19, 49]). It consists of 3K positive pairs of samples (pairs of images of the same per" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20928" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.264 + ], + "angle": 0, + "content": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
ArcFace89.8686.7394.3193.82
ArcFace + Dprior(African)92.6290.2096.3195.85
ArcFace + Dprior(Asian)92.9190.6196.1595.83
ArcFace + Dprior(Afr+Asian)93.5491.5196.5196.17
increase+3.68+4.78+2.20+2.35
SphereFace92.0889.5195.5295.02
SphereFace + Dprior(Afr+Asian)92.8390.5296.2295.65
increase+0.75+1.01+0.70+0.63
GAC92.4889.1396.3394.66
GAC + Dprior(Afr+Asian)93.6090.8096.7695.46
increase+1.12+1.67+0.43+0.80
" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.267, + 0.466, + 0.424 + ], + "angle": 0, + "content": "
TPR @ FPR=10-4↑
ArcFace81.4876.8087.4786.89
ArcFace + D prior (African)85.5681.9991.4290.67
ArcFace + D prior (Asian)85.7982.4191.1290.36
ArcFace + D prior (Af+Asian)86.8483.7491.7491.19
increase+5.36+6.94+4.27+4.30
SphereFace83.6380.0489.1089.07
SphereFace + D prior (Af+Asian)85.1481.4790.6590.16
increase+1.51+1.43+1.55+1.09
GAC85.6480.2491.7288.12
GAC + D prior (Af+Asian)86.9182.4792.3989.48
increase+1.27+2.23+0.67+1.36
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Table 3: Comparison of the methods on the newly assembled RB-WebFace validation set. Publicly available authors' implementation of GAC [20] has been used to retrain the method and evaluate the quality. Here, we showcase the TPR given two pre-selected FPR thresholds to highlight the error rate difference separately for each ethnic group. For GAC, the setting GAC (Estimated) (see [20]) is used. This table indicates how the proposed pretraining scheme provides an additional boost for different pipelines." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.578, + 0.428, + 0.669 + ], + "angle": 0, + "content": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ 1% Dprior92.5990.4496.0695.72
+ 10% Dprior92.8490.7296.1295.64
+ 100% Dprior93.5491.5196.5196.17
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.68, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Table 4: Ablation study of the quality dependence on the \\( \\left| {\\mathcal{D}}_{\\text{prior }}\\right| \\) prior dataset size on the newly assembled RB-WebFace validation set. \\( {\\mathcal{D}}_{\\text{prior }} = \\left\\{ \\begin{array}{ll} \\text{AfricanFaceSet - } 5\\mathrm{M} & \\cup \\\\ \\text{AsianFaceSet - } 3\\mathrm{M} & \\end{array}\\right\\} \\) was used in this set of experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.884 + ], + "angle": 0, + "content": "son) and 3K negative pairs (pairs of images of different but similarly looking people, selected by a face recognition algorithm which can introduce selection bias) for each ethnic group, summing up to 24K pairs. Evaluation is carried out in an LFW-like protocol [27] that involves evaluating the backbone \\( f_{\\theta,\\psi}(I) \\) on images of each pair and thresholding the resulting cosine distances between embeddings. Each set of 3K+3K pairs is constructed from 3K people comprising a subset of MS-Celeb-1M [23]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In addition to that, we propose a new test set \\( RB \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "WebFace constructed in a similar fashion from the recently proposed million-scale identification dataset WebFace-42M (a cleaned version of WebFace-260M [60]). RB-WebFace is constructed by evaluating a pretrained ethnic group classifier on WebFace-42M to separate it into four ethnic groups (the group assignment is later refined by a consensus procedure). For each group, the largest possible number of distinct people is taken to construct positive and negative pairs. The details are provided in Supplementary [1], as well as the example pictures and comparison to typically used test datasets by the number of people and pairs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.27, + 0.62, + 0.284 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.475 + ], + "angle": 0, + "content": "Comparison with the state-of-the-art on RFW. The Table 2 contains the comparison to baseline (ArcFace R-50) and a few state-of-the-art methods (GAC [20], DebFace [19], and others). RFW verification accuracy is reported for each of the four ethnic groups. We demonstrate an increase in accuracy for all races compared to the baseline, which is on par with the state-of-the-art and outperforming on Caucasian and Indian ethnic groups. The largest increase w.r.t. the baseline is observed for the African group. It is important to note that our pretraining scheme can also be used to initialize other methods and enhance their results in a similar fashion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Comparison on newly assembled RB-WebFace test set. Since the number of positive and negative pairs is orders of magnitude different, we report the ROC curves values (TPR vs. FPR), better suited for the class-imbalanced evaluations, instead of reporting accuracy as for RFW. By sweeping the threshold (in [0.1, 0.75] range), we obtain the (TPR, FPR) pairs, and TPR @ FPR = \\(\\{10^{-3}, 10^{-4}\\}\\) is reported in Table 3. Our method (denoted as Baseline + Dprior) is compared to the Baseline (ArcFace R-50). Additionally, we report the enhancement that our procedures provide to other methods [34, 20]. The main conceptual result that we demonstrate is the significant increase of the TPR vs. given FPR for all races and decrease of ethnic bias on RB-WebFace. A separate plot in the Supplementary [1] comprises the TPR measurements for a wide range of possible FPR values. We also provide the results for ResNet\\(\\{34,100\\}\\) backbones in Table 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Limited labeled data. The study in Fig. 4 describes the dependency of our method's accuracy on the number of samples in \\( \\mathcal{D}^{facerec} \\). We demonstrate the difference between our method and the ArcFace R-50 baseline for each of the benchmarks, both on RFW and RB-WebFace. As expected, pretraining helps the most when the network is fine-tuned on more limited amounts of labeled data. This especially highlights the benefits of using self-supervised learning in these scenarios." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limited prior dataset. An ablation in Table 4 describes the quality dependence on the size of the prior dataset. As" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.21, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.34, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.481, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.089, + 0.614, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.089, + 0.756, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.089, + 0.889, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.206, + 0.207, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.206, + 0.342, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.206, + 0.481, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.206, + 0.614, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.206, + 0.756, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.206, + 0.889, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.894, + 0.391 + ], + "angle": 0, + "content": "Figure 4: Ablation study of the dependence of the test quality, evaluated for RFW and RB-WebFace datasets, on the number of labeled samples from \\( |\\mathcal{D}^{facerec}| \\) used for fine-tuning. In this experiment, \\( \\mathcal{D}^{prior} = \\{\\text{AfricanFaceSet-5M} \\cup \\text{AsianFaceSet-3M}\\} \\), while \\( \\mathcal{D}^{facerec} = \\text{BUPT-BalancedFace} \\). For \\( \\mathcal{D}^{facerec} \\), fraction of the data (100%, 10%, 1%) defines the number of uniformly sampled people left in the dataset. In the plot legend, Baseline is ArcFace R-50." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.468 + ], + "angle": 0, + "content": "shown, gradually enlarging it results in a monotonous performance increase for all ethnic groups. This indicates that the use of large amounts of diverse unlabeled data significantly strengthens the method in the end." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.475, + 0.437, + 0.571 + ], + "angle": 0, + "content": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ AE on Dprior92.9390.8395.7395.68
+ VAE on Dprior92.4590.3995.6495.40
+ ours on Dprior93.5491.5196.5196.17
" + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.579, + 0.437, + 0.648 + ], + "angle": 0, + "content": "
TPR @ FPR=10-4↑
Baseline81.4876.8087.4786.89
+ AE on Dprior85.8583.0990.4689.86
+ VAE on Dprior85.1982.0890.3889.25
+ ours on Dprior86.8483.7491.7491.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Table 5: Comparison of different pretraining strategies. Along with ours based on consequent training of StyleGAN and a ResNet encoder, one could employ vanilla (AE) and variational (VAE) autoencoders for pretraining (with the same ResNet-50 architecture as an encoder). Despite simpler simultaneous training of an encoder and a decoder in a single stage, the StyleGAN-based procedure is a more powerful prior. Baseline refers to the ArcFace R-50." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Can StyleGAN be replaced with a simpler encoder-decoder architecture? In Table 5, we compare our pretraining procedure to more conventional ones where the ResNet-50 encoder is pretrained in a symmetric autoencoder (AE) or in a symmetric variational autoencoder (VAE) setting. While these models can be pretrained in a single stage, our proposed StyleGAN+encoder approach re" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.405, + 0.89, + 0.505 + ], + "angle": 0, + "content": "
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
R-3484.5781.4791.1589.9173.6069.2383.1180.40
+ D prior90.9288.3495.2794.3983.0878.6989.2587.98
R-5089.8686.7394.3193.8281.4876.8087.4786.89
+ D prior93.5491.5196.5196.1786.8483.7491.7491.19
R-10093.7791.5596.3296.6287.2884.2891.4691.66
+ D prior94.8293.0096.7096.9288.7686.0092.2492.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.516, + 0.894, + 0.561 + ], + "angle": 0, + "content": "Table 6: Comparison of different ResNet backbones on RB-WebFace. The baseline is ArcFace and \\( D^{prior} = \\) AfricanFaceSet-5M U AsianFaceSet-3M." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.623 + ], + "angle": 0, + "content": "quires two separate pretraining stages but demonstrates better performance, which motivates the use of state-of-the-art generative models for face recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.893, + 0.852 + ], + "angle": 0, + "content": "Why is AE/VAE less suitable than StyleGAN when finetuned for face recognition? We see two possible reasons why our procedure yields better results. First, StyleGAN-based architectures are tailored to be the state-of-the-art of face generation, which leads to the possible assumption that encoder + StyleGAN pipeline is capable of saving more useful information about the face features than e.g. VAE. Second, we observe that the quality of StyleGAN generations significantly improves when it is trained on larger amount of data, while for VAE it is not the case (6% over 32% improvement when 100x more data given – see Fig. 5). This scalability issue is also highlighted by the fact that AE/VAE, trained on the 100% of \\( \\mathcal{D}^{prior} \\), perform for the face recognition task on par with StyleGAN, trained on only 1% of \\( \\mathcal{D}^{prior} \\) (see the Tables 5 and 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Is it helpful to draw samples from StyleGAN instead of training an encoder for it? In this experiment, we used our trained StyleGAN to generate synthetic faces and add" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20930" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.146, + 0.079, + 0.38, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.473, + 0.425 + ], + "angle": 0, + "content": "Figure 5: FID score, calculated over 100K random pictures from \\( \\mathcal{D}^{prior} = \\{\\text{AfricanFaceSet-5M} \\cup \\text{AsianFaceSet-3M}\\} \\) vs. 100K random samples from VAE or StyleGAN. The score (the less the better) is given w.r.t. the percentage of \\( \\mathcal{D}^{prior} \\) used to train these models. We observe that the image generation quality, that FID represents, drops significantly when StyleGAN is provided with more data, while VAE does not scale to larger data sets that well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.752 + ], + "angle": 0, + "content": "them to the training set. We first infer StyleGAN latent codes for all African and Asian samples in our training set (BUPT-BalancedFace). Since face recognition training always requires labeled images, we generate samples using latent codes closer to the inverted training images. In particular, we take random pairs \\((I_1, I_2)\\) of images of the same ethnic group (either African or Asian) and generate an interpolated sample by calculating convex combinations of their StyleGAN latent codes: \\(I_{\\mathrm{comb}} = \\lambda \\cdot f_{\\theta, \\omega}'(I_1) + (1 - \\lambda) \\cdot f_{\\theta, \\omega}'(I_2)\\), \\(\\lambda \\sim U[0, 1]\\). The ground-truth label (always required for face recognition training) is constructed as a two-hot vector \\((0, \\dots, 0, 1 - \\lambda, 0, \\dots, 0, \\lambda, 0, \\dots, 0)\\) (as influenced by MixUp [55]). In total, we augment the training set with an equal amount of synthetic interpolations (1.3 M), obtained from randomly drawn African-African and Asian-Asian pairs. The results in Table. 7 show that the added interpolations, despite the created imbalance and increased variability in the data, do not yield a comparable performance increase. In addition, we show that increasing the number of interpolations doesn't help, while pretraining scheme benefits from more pretraining data (see Table 4)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.753, + 0.47, + 0.784 + ], + "angle": 0, + "content": "We provide additional ablations and comparisons in the Supplementary [1]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.205, + 0.815 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As an increasing number of modern computer vision methods become production-ready, new challenges regarding their final use are posed. In this work, we present a solution for improving the quality of facial recognition by pretraining the face image encoder with unlabeled data using" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.162 + ], + "angle": 0, + "content": "
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
ArcFace R-50 + Dprior89.8686.7394.3193.8281.4876.8087.4786.89
93.5491.5196.5196.1786.8483.7491.7491.19
+ 1.3M interps91.8789.0796.0895.5584.6579.9791.1390.30
+ 6M interps89.4886.8594.9294.2280.9577.0088.7587.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.172, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Table 7: In this experiment, instead of pretraining our model as an encoder for StyleGAN, we use the StyleGAN and the trained encoder to draw samples from the learned face distribution and add them to our dataset. The results indicate that interpolation-based procedure, despite being more straightforward than the proposed pretraining-based scheme, is actually less efficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.421 + ], + "angle": 0, + "content": "StyleGAN and encoder. Furthermore, we additionally release two training datasets for unsupervised pretraining and one large-scale protocol for bias estimation. We show that we are able to tune the performance on different ethnicities by altering the composition of the unlabeled prior datasets. We hope that the released datasets, together with the protocol, will drive forward the research on mitigating the face recognition biases using self-supervised approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.893, + 0.667 + ], + "angle": 0, + "content": "There are several directions for possible future work that we foresee. First, integrating the two steps of pretraining and fine-tuning into one would help to avoid forgetting of the pretrained weights. Second, publicly available datasets often don't possess enough variability (both in terms of capture conditions and demographics) [25] and typically feature only low-resolution images which might be a blocking factor for the fairness research [49, 30]. Accordingly, the results on the newly assembled RB-WebFace benchmark can be analyzed more thoroughly, which can bring new insights about the factors positively affecting bias mitigation in real-world scenarios. Finally, various architectures typically used for self-supervised learning (such as transformers or highly scalable generative models [43]), might allow for the construction of more flexible methods due to heterogeneous inputs and outputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Legal concerns. The collected unlabeled data was collected anonymously in accordance with Standard YouTube License and CC YouTube License and does not contain personally identifiable information (PII). The data is released only in the form of links to YouTube videos and the corresponding timestamps, fully following good practice of similar data collection [10, 17, 32] and with the data subjects protection as per Art. 14 5(b) GDPR law." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. We gratefully acknowledge the support of this research by a TUM-IAS Hans Fischer Senior Fellowship, the ERC Starting Grant Scan2CAD (804724) and the Horizon Europe vera.ai project (101070093). We also thank Yawar Siddiqui for the helpful technical advice on StyleGAN part, Angela Dai for the video voiceover, and Dmitriy Karfagenskiy for the proofreading and corrections." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20931" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] How to boost face recognition with stylegan?. Supplemental material, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.199 + ], + "angle": 0, + "content": "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711-6720, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.468, + 0.268 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[4] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.468, + 0.368 + ], + "angle": 0, + "content": "[5] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.424 + ], + "angle": 0, + "content": "[6] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[7] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGface2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face and Gesture Recognition, pages 3414-3424, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[8] Yunliang Chen and Jungseock Joo. Understanding and mitigating annotation bias in facial expression recognition. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14980-14991, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[9] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 539-546, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.649 + ], + "angle": 0, + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In Interspeech, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[11] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.789 + ], + "angle": 0, + "content": "[12] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In Proc. of the 58th Annual Meeting of the Association for Computational Linguistics, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.845 + ], + "angle": 0, + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Emily Denton, Ben Hutchinson, Margaret Mitchell, Timnit Gebru, and Andrew Zaldivar. Image counterfactual sensitivity analysis for detecting unintended bias. arXiv preprint arXiv:1906.06439, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[17] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[18] Markos Georgopoulos, James Oldfield, Mihalis A Nicolaou, Yannis Panagakis, and Maja Pantic. Mitigating demographic bias in facial datasets with style-based multi-attribute transfer. International Journal of Computer Vision, 129:2288-2307, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[19] Sixue Gong, Xiaoming Liu, and Anil K Jain. Debface: Debiasing face recognition. arXiv preprint arXiv:1911.08080, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[20] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating Face Recognition Bias via Group Adaptive Classifier. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3414-3424, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] Google. Real Tone on Pixel 6. Be seen as you truly are., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[22] Patrick Grother, Mei Ngan, and Kayee Hanaoka. Face recognition vendor test (FVRT): Part 3, demographic effects. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[23] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-Celeb-1M: A dataset and benchmark for large-scale face recognition. In European Conference on Computer Vision (ECCV), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[24] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, pages 1735-1742, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[25] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in AI: the Casual Conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.79 + ], + "angle": 0, + "content": "[26] Mingjie He, Jie Zhang, Shiguang Shan, and Xilin Chen. Enhancing face recognition with self-supervised 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4062-4071, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[27] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Di Kang, Debarun Dhar, and Antoni Chan. Incorporating side information by adaptive convolution. Advances in Neural Information Processing Systems, 30, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[29] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[30] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Image resolution susceptibility of face recognition models. arXiv preprint arXiv:2107.03769, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[31] Wenyu Li, Tianchu Guo, Pengyu Li, Binghui Chen, Biao Wang, Wangmeng Zuo, and Lei Zhang. Virface: Enhancing face recognition via unlabeled shallow data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14729-14738, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[32] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3207-3216, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[33] Chun-Hsien Lin and Bing-Fei Wu. Domain adapting ability of self-supervised learning for face recognition. In IEEE International Conference on Image Processing (ICIP), pages 479–483, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[34] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 212-220, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv:1907.11692, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.471, + 0.574 + ], + "angle": 0, + "content": "[36] Vongani Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying Bias in GANs through the Lens of Race. In European Conference on Computer Vision (ECCV), pages 344-360, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.471, + 0.631 + ], + "angle": 0, + "content": "[37] Richard T Marriott, Sami Romdhani, and Liming Chen. A 3d gan for improved large-posed facial recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13445–13455, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[38] Tuan-Duy H Nguyen, Huu-Nghia H Nguyen, and Hieu Dao. Recognizing families through images with pretrained encoder. In IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 887-891, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[39] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10880-10890, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[40] Christian Rathgeb, Pawel Drozdowski, Naser Damer, Dinusha C Frings, and Christoph Busch. Demographic fairness in biometric systems: What do the experts say? IEEE Technology and Society Magazine, 41(4):71-82, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[42] Joseph P Robinson, Gennady Livitz, Yann Henon, Can Qin, Yun Fu, and Samson Timoner. Face Recognition: Too Bias," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "or Not Too Bias? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[43] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH Conference Proceedings, pages 1-10, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[44] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6902-6911, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[46] Fu-En Wang, Chien-Yi Wang, Min Sun, and Shang-Hong Lai. Mixfairface: Towards ultimate fairness via mixfair adapter in face recognition. In AAAI Conference on Artificial Intelligence, volume 37, pages 14531-14538, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5265-5274, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[48] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In ACM International Conference on Multimedia, pages 3779-3782, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[49] Mei Wang and Weihong Deng. Mitigating Bias in Face Recognition using Skewness-Aware Reinforcement Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9322-9331, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[50] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial bias by information maximization adaptation network. In International Conference on Computer Vision (ICCV), pages 692-702, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[51] Qingzhong Wang, Pengfei Zhang, Haoyi Xiong, and Jian Zhao. Face evolves: A high-performance face recognition library. arXiv preprint arXiv:2107.08621, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[52] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[53] Xingkun Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, Jilin Li, Feiyue Huang, Yong Li, and Zhen Cui. Consistent Instance False Positive Improves Fairness in Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 578-586, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv:1411.7923, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[55] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[56] Haoyu Zhang, Marcel Grimmer, Raghavendra Ramachandra, Kiran Raja, and Christoph Busch. On the applicability of synthetic data for face recognition. In IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20933" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10):1499-1503, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[59] Song Zhang, Gehui Shen, Jinsong Huang, and Zhi-Hong Deng. Self-supervised learning aided class-incremental life-long learning. arXiv preprint arXiv:2006.05882, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20934" + } + ] +] \ No newline at end of file diff --git a/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_origin.pdf b/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..081ef15384beffc5d09d210ddd68fe7298dd9912 --- /dev/null +++ b/2023/How to Boost Face Recognition with StyleGAN_/aa9ef81d-3574-478e-bb1c-7bc6363afaee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72a79af115586cd13c45adc5ba5b260bae49367133cd6010eb135400211bf9aa +size 1077545 diff --git a/2023/How to Boost Face Recognition with StyleGAN_/full.md b/2023/How to Boost Face Recognition with StyleGAN_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e6f63908cafe398362e89b0845abd3a230e34c6 --- /dev/null +++ b/2023/How to Boost Face Recognition with StyleGAN_/full.md @@ -0,0 +1,261 @@ +# How to Boost Face Recognition with StyleGAN? + +Artem Sevastopolsky1 Yury Malkov2,* Nikita Durasov3 Luisa Verdoliva1,4 Matthias Nießner1 + +1 Technical University of Munich, Germany 2 Twitter, US 3 École polytechnique fédérale de Lausanne, Switzerland 4 University Federico II of Naples, Italy + +![](images/7e00cfd1644ea9e081094344af373d3dc09ed5284de51c502c7b5f0b50fb7315.jpg) +Figure 1: Our method is aimed at boosting the performance of face recognition. This is achieved by gathering a random image collection without face recognition labels (unlabeled set) and then fitting a mapping from an image to the StyleGAN latent space onto that collection. To learn this mapping, we use the pSp encoder architecture. For the downstream face recognition task, the same encoder is then fine-tuned on a (potentially much smaller) face recognition dataset with identity labels (labeled set). + +# Abstract + +State-of-the-art face recognition systems require vast amounts of labeled training data. Given the priority of privacy in face recognition applications, the data is limited to celebrity web crawls, which have issues such as limited numbers of identities. On the other hand, self-supervised revolution in the industry motivates research on the adaptation of related techniques to facial recognition. One of the most popular practical tricks is to augment the dataset by the samples drawn from generative models while preserving the identity. We show that a simple approach based on fine-tuning $pSp$ encoder for StyleGAN allows to improve upon the state-of-the-art facial recognition and performs better compared to training on synthetic face identities. We also collect large-scale unlabeled datasets with controllable ethnic constitution - AfricanFaceSet-5M (5 million images of different people) and AsianFaceSet-3M (3 million images of different people) - and we show that pretraining on each of them improves recognition of the respective ethnicities (as well as others), while combining all unlabeled datasets results in the biggest performance increase. Our self-supervised strategy is the most useful with limited amounts of labeled training data, + +which can be beneficial for more tailored face recognition tasks and when facing privacy concerns. Evaluation is based on a standard RFW dataset and a new large-scale RB-WebFace benchmark. The code and data are made publicly available at https://github.com/seval00/stylegan-for-facerec. + +# 1. Introduction + +Modern face recognition methods rely on deep convolutional networks trained on large-scale datasets [54, 7, 23, 60]. These methods are now being integrated into a vast number of real-world applications, ranging from face unlock for smartphones and photo organizers to law enforcement systems and border control. A typical open face recognition dataset consists of web-crawled images of celebrities, leading to limited size and lack of balance in subgroups, such as ethnicity, age, etc. Training a state-of-the-art solution, however, requires enormous amounts of labeled data, scraping which may lead to privacy and legal issues. We suggest and study an alternative solution to using celebrity photos - pretraining the face recognition backbone on a generative task. Specifically, we first train StyleGAN2-ADA [29] on collected + +![](images/ce4aea9a1ba9b95c3edeace262404e5f6d803ae6dd38f975fa79f4dd5b12353a.jpg) +Figure 2: Our method is trained in three consecutive steps. First, we fit StyleGAN2-ADA to the face image distribution of the unlabeled prior dataset $\mathcal{D}^{prior}$ . Second, the pSp encoder is trained (also on $\mathcal{D}^{prior}$ ) to map images to the latent codes in the learned latent space. Finally, the encoder, which is pretrained to extract meaningful features from an image, is fine-tuned for the downstream face recognition task with the ArcFace loss (similar losses can be used instead) on $\mathcal{D}^{facerec}$ . The two first steps comprise the self-supervised pretraining stage; i.e., no identity labels are required for them. + +![](images/4e6ae3af18f67dfb67f11c10417def155595b4175cdd61330fabc429e7a7bfe8.jpg) + +![](images/a1ca5781353c1f3c74e4af754ea3f7ea328cecc1ab98deaf1d7c7f2c243a3e99.jpg) + +unlabeled data (which we later refer to as an unlabeled prior dataset) to fit the face image distribution. Subsequently, we train an encoder (following pixel2style2pixel (pSp) architecture [41]) that maps input images to vectors in the learned StyleGAN2-ADA latent space. Importantly, during the pretraining steps, no identity labels are used, so we can use diverse datasets crawled from the Internet without compromising privacy. Finally, we transfer the learned pSp encoder convolutional weights into the face recognition network and train it in a standard face recognition setup. + +We show that, in contrast to training face recognition tasks on StyleGAN generated data (also demonstrated in [38, 39] and studied e.g. in [36]), our encoder pretraining step significantly boosts the final performance. The idea of augmenting face recognition datasets with synthetic data is widespread and constitutes many approaches, however, unclear and heuristic definition of the target label limits the amount of useful signal that can be transferred into the face recognition model this way. Our approach goes hand-in-hand with the current development of self-supervised learning [15, 35, 12, 16] and makes up one of the first approaches of its application to face recognition [26, 33]. This allows us to demonstrate vast improvements on limited labeled training data compared to the setup without self-supervised pretraining (for instance, $10\%$ verification accuracy increase for only $1\%$ of the labeled data used). + +The simplicity of the data collection procedure also allows us to control the distribution of the unlabeled data and thus influence the decrease of the error rates for specific demographic groups. Despite the fact that the current state-of-the-art algorithms often demonstrate very low average error rates [13, 47, 51, 48], it is considered unethical to integrate face recognition solutions that exhibit significant ethnic, age, or gender bias. Such bias is present both in open-source face recognition methods [20, 53, 49, 19, 28, 44, 46] and in comprehensively evaluated commercial face recog + +nition systems [22], resulting in significantly different error rates measured for the groups of interest. The topic has attracted significant attention in other areas of computer vision operating in face domain [8, 14, 21, 40] and in the media. + +We constructively demonstrate that collecting large amounts of in-the-wild face images of a given group of interest is feasible (and can be done semi-automatically), while collecting datasets with identity labels is problematic. The labels require linking photos of the same person taken in different conditions, which means the person must be tracked. This typically constrains public datasets to celebrities, gathered using search engines [23, 54, 7, 60], while social networks and companies that provide services with photos use input from users. Second, the collected in-the-wild data, treated as a set of faces without identity labels (but labeled with the group attribution), can be efficiently used for self-supervised pretraining for face recognition networks, subsequently fine-tuned on the standard face recognition datasets (see Fig. 1). + +To summarize our main contribution, we present a novel self-supervised method for improving the performance of face recognition based on StyleGAN pretraining. This allows to leverage large-scale amounts of available unlabeled data for face recognition. While the improvement is the most significant on limited data, pretraining is also helpful for large-scale labeled datasets. + +# 2. Related work + +Face recognition datasets. Several large-scale datasets of faces with identity labels have been released publicly, such as CASIA-WebFace [54], VGGFace2 [7], MS-Celeb-1M [23] and the very recent million-scale WebFace-42M dataset [60]. However, these datasets have been collected "in the wild", hence, they inevitably suffer from biases in terms of age, gender, + +
Dataset name# people# images# pic./personethnic diversityacquisition
MS-Celeb-1M100K8.2M82unevenmostly American and British actors
VGGFace29.1K3.3M362unevenGoogle Image search
CASIA-WebFace10K494K49unevencelebrities from IMDb
CASIA-Face-Africa1.1K38K34all Africancontrolled, indoor and outdoor
MegaFace (unavailable)672K4.7M7unevenYahoo Flickr website search queries
DiveFace24K72K3balanced across 3 ethnicitiessubset of MegaFace
BUPT-BalancedFace28K1.3M46balanced across 4 ethnicitiescelebrities from MS-Celeb-1M
BUPT-GlobalFace38K2M52matches global populationcelebrities from MS-Celeb-1M
BUPT-TransferFace>10K600K6075% Caucasian. vs. otherscelebrities from MS-Celeb-1M
AfricanFaceSet-5M5M5MunlabeledAfrican majorityrandom faces from YouTube news videos
AsianFaceSet-3M3M3MunlabeledAsian majorityrandom faces from YouTube news videos
+ +Table 1: Overview of the publicly available training sets in facial domain. Typically, the quality of a face recognition dataset is described by several factors: the number of people in the dataset, the number of images per person, and the diversity of capture conditions. While our AfricanFaceSet-5M and AsianFaceSet-3M datasets are not directly suitable for face recognition due to the absence of identity labels, they comprise more distinct people than existing large-scale datasets and contain a more diverse distribution of faces than only celebrities. + +or race. To better understand and stimulate research in fairness about face recognition, several datasets have been proposed. Examples of such labeled datasets include BUPT-Globalface [49] (2M images with ethnic distribution matching the world population) and BUPT-Balancedface [49] (1.3M images with the perfect ethnic split). Racial Faces in-the-Wild (RFW) [50] is a verification database that has been constructed from MS-Celeb-1M. Currently, it serves a standard fairness benchmark. Another verification dataset BFW has been introduced in [42], which, similarly to RFW, contains eight subgroups balanced across gender and ethnicity. The main feature of all these datasets – predefined ethnic split – allows one to disambiguate dataset bias and model bias for more precise methods evaluation. Still, the currently available data possesses a number of limitations. For instance, the number of distinct people is typically limited and is often many orders of magnitude smaller than the dataset size. Additionally, both identification and verification datasets require many images per person, which usually restricts the construction of open-source datasets to the search of celebrity pictures by text queries. Our two photo collections, released to the public, – AfricanFaceSet-5M and AsianFaceSet-3M, – fill a different gap in the space of available datasets. On the one hand, they neither have identity labels nor feature many images per person. On the other hand, these collections are large and focused on groups and conditions that are underrepresented in general face recognition datasets. Our evaluation dataset RB-WebFace, assembled from large-scale WebFace-42M, is a new verification dataset containing a significantly larger number of pairs and does not use external models to select the negatives (which introduces certain selection bias) compared to RFW, which is currently used as the main benchmark in the bias mitigation-focused branch of works. + +Data augmentation. Generating synthetic data is a possi- + +ble solution to improving the performance of face recognition and in some cases reducing the negative effect of the dataset bias. An idea of this kind is pursued in [37], a nonlinear 3DMM texture model is proposed to produce sharp renderings of faces from novel poses. This technique improves generalization with respect to the head pose and illumination. A number of approaches synthesize data via face generative models. For instance, in SynFace [39] random face images are constructed by a GAN with identity control, and labels are constructed by a procedure similar to MixUp augmentation [55]. Similarly, in [18] a style transfer GAN is used to simultaneously transfer multiple facial demographic attributes and generate diverse images for each attribute class. Authors of Virface [31] suggested a method for incorporating additional negative (impostor) pairs from the unlabeled data showing a boost in metrics. Zhang et al. [56] study the applicability of the data generated by StyleGAN by inspecting the distributions of downstream models but do not study its effect on face recognition metrics. + +Our approach is inspired by the latter idea of generating synthetic samples but takes a step forward by utilizing the generative model itself (StyleGAN with an encoder in our case). As shown in [41, 45, 2, 3], the StyleGAN encoders have high potential for both generating new realistic faces from a latent code and solving the inversion problem. This highlights the expressiveness of these models and richness of their internal representations. + +Self-supervised pretraining. Despite the general dominance of supervised learning in practical ML and CV, approaches based on self-supervised learning have been evolving in various forms, such as self-organizing or siamese networks [4, 5, 9, 24]. Currently, self-supervised learning is the dominant approach in NLP with a wide spectrum of possible approaches [35, 12, 11] and is being actively adapted in computer vision [16]. + +![](images/260341261875363e7bea12e74e27db49afbc07d0094fd2608d21960a0aab139e.jpg) +Figure 3: Our data collection process starts by manually specifying a set of YouTube channels with a specific topic, e.g. a set of news channels of a desired town or part of the world. All videos are downloaded in the highest available quality, one frame per $P$ seconds is extracted from each video, and all the faces found in the extracted frames are cropped, aligned by landmarks and resized to the target resolution. This way, we obtain millions of random faces with a desired demographic distribution. + +Analogously, we are witnessing the first attempts to apply self-supervised learning now being integrated into the face recognition frameworks. In 3D-BERL [26], the performance over multiple benchmarks is improved by a separate 3D reconstruction network branch. The work [33] studies the effect of self-supervised learning for domain transfer in face recognition. Incremental learning [52] can be bridged with self-supervised approaches to adapt to large number of target classes [59]. The procedure for collecting large-scale unlabeled datasets allows us to adapt self-supervised training in a more conventional fashion while proving its efficacy for our application. + +# 3. Method + +Our pipeline is comprised of several stages. First, we train StyleGAN2-ADA to fit the face distribution on the unlabeled dataset. Second, the pSp encoder is trained (also on the unlabeled dataset), that will define a feature extractor well-suited for the group of interest. Finally, the encoder is fine-tuned for the downstream face recognition task. The entire procedure is outlined in Subsec. 3.2 and visually described in Fig. 2. Before describing the method itself, we outline the schematics of our prior dataset collection procedures in Subsec. 3.1. + +# 3.1. Prior dataset collection + +Due to the nature of the task, the prior dataset must contain the samples from the group of interest. At the same time, samples from the prior dataset do not require identity labels of any kind, unlike samples of the face recognition datasets. In practice, this removes the restriction of linking photos of the same person (something that might be considered a violation of privacy), and thus enlarges the search space and simplifies data collection. Still, fulfilling several requirements for the prior dataset remains a challenge, such as: collecting faces only from a specific group (e.g. an ethnic or gender group), obtaining a large number of them (preferably, an order of magnitude more than in the face recognition dataset to perform subsequent transfer + +learning), and using only data legally allowed for collection. We found semi-automatic YouTube channel scraping to be an efficient solution that satisfies these requirements. In particular, we propose to select a set of publicly available YouTube channels dedicated to the desired group of interest. The channel names are the only entry point and the only manually performed step for the scraping procedure (see Fig. 3). Typically, the requirement for the channels is having them systematically featuring different people; a possible example would be a set of news channels released in a country of choice. All videos are downloaded from every channel, one frame per $P$ seconds of each video is extracted, and all faces are detected, cropped and aligned by landmarks via MTCNN [57] library. Our data has been scraped from a predefined set of channels such as news channels and others. + +The unlabeled prior dataset will later be referred to as $\mathcal{D}^{prior}$ . Table 1 describes the relative difference of the collected data to the datasets typically used for face recognition. Note however, that despite the latter being directly inapplicable for training due to the lack of the labels, they potentially contain more people than the others. + +# 3.2. Architecture and the training procedure + +The central part of our pipeline is a single face recognition convnet $f_{\theta, \psi}(I)$ (a backbone), that takes a single RGB image $I$ as an input, and outputs a 512-dimensional vector $e \in \mathbb{R}^{512}$ (face embedding). Typically, the backbone is trained on a training dataset with angular margin based losses, such as SphereFace [34], ArcFace [13], and others, in standard works on face recognition. In our pipeline, we follow the same procedure, but only after a pretraining stage is performed. In order to do so, we first train the StyleGAN2-ADA [29] generator $g_{\phi}(w(z))$ that transforms a latent vector $z \in \mathcal{Z} \subseteq \mathbb{R}^{512}$ into unfolded latent space $w(z) \in \mathcal{W} + \subseteq \mathbb{R}^{L \times 512}$ , and then into a realistic face image $\hat{I}$ . The generator $g_{\phi}(w(z))$ is trained together with a discriminator that allows the generator to learn the distribution of faces in $D^{prior}$ . The training procedure for StyleGAN2- + +ADA follows the one in the corresponding paper. Secondly, we introduce a network $f_{\theta, \omega}'(x)$ , which is trained as an encoder for the StyleGAN generator, i.e. a network that solves an inverse problem: given an RGB image $I$ , predict an unfolded latent code $w \in \mathcal{W}+$ , such that the corresponding reconstruction $\hat{I} = g_{\phi}(w)$ is as close to the input $I$ as possible. A set of different possible approaches has appeared recently for training an encoder for StyleGAN2, ranging from applying a ConvNet with style-predicting layers [41] to employing hyper-networks [2]. In our method, the encoder architecture follows the pSp method [41], that proposes predicting a latent code by a ConvNet divided into a convolution part (with parameters $\theta$ ), following a ResNet architecture, and a set of fully-convolutional style predictors - map2style blocks (with parameters $\omega$ ). The training procedure for the encoder follows the one outlined in [41] with a few variations. Namely, we only keep the fidelity losses (standard L2 and neural-based LPIPS [58]) and disable the identity loss proposed by the authors. During the encoder training, the StyleGAN generator remains frozen in order to fix the latent space. A pair of networks $f_{\theta, \omega}'(x)$ and $g_{\phi}(w(z))$ can be seen as an asymmetric autoencoder (since the number of encoder parameters highly surpasses the number of generator parameters) with the decoder pretrained for a generative task and the encoder subsequently trained for a discriminative (regressive) task. The visual quality and the identity preservation of the reconstruction $\hat{I}$ indirectly defines the expressiveness of the latent code $l$ predicted by the encoder. + +Finally, after the encoder $f_{\theta,\omega}'(I)$ is trained, we transfer its convolutional parameters $\theta$ into the face recognition backbone $f_{\theta,\psi}(I)$ , which also comprises a set of new parameters $\psi$ , corresponding to a fully-connected layer in the end of the network. We repeat the standard training procedure for face recognition, e.g. described in ArcFace [13] and others. The backbone is trained on a face recognition dataset $D^{facerec} = \{\mathcal{D}_1, \dots, \mathcal{D}_N\}$ , where each group $\mathcal{D}_i$ corresponds to a set of images of the same person $\# i$ . The slight differences in hyperparameters that we found beneficial when used with pretraining and technical details are described in the Supplementary [1]. + +# 4. Results + +# 4.1. Data and evaluation protocol + +Our system requires two datasets - a labeled face recognition dataset $\mathcal{D}^{facerec}$ and an unlabeled prior dataset $\mathcal{D}^{prior}$ . As for the $\mathcal{D}^{facerec}$ , we employ BUPT-BalancedFace [49] due to its ethnic and gender balance guarantees. Namely, it consists of 1.3 million images belonging to 28K different people, divided into 4 ethnic groups of equal size - 7K African, East Asian, Indian, and Caucasian people each. Other state-of-the-art methods can + +
RFW, accuracy %,↑
Cauc.AfricanAsianIndianAvg ↑Std ↓
ArcFace R-50* [13]96.1893.9893.7294.6794.641.11
CosFace R-50* [47]95.1293.9392.9892.9393.740.89
DebFace* [19]95.9593.6794.3394.7894.680.83
ACNN* [28]96.1294.0093.6794.5594.580.94
PFE* [44]96.3895.1794.2794.6095.110.93
RL-RBN* [49]96.2795.0094.8294.6895.190.63
GAC R-50* [20]96.2794.4094.3294.7794.940.79
Baseline (ArcFace)96.0094.0093.0894.4894.391.06
+ Dprior (African)96.3594.3793.6294.8894.811.00
+ Dprior (Asian)96.3894.6794.0395.0395.030.86
+ Dprior (Afr+Asian)96.5295.0093.9094.9395.090.94
increase+0.52+1.00+0.82+0.45
+ +Table 2: Comparison of the verification accuracy of the methods on RFW validation set. Asterisk “*” indicates that the numbers are directly taken from the respective table in [20]. In case of the ArcFace R-50 baseline, it also corresponds to a slightly different reported training procedure. For other methods, the epoch with the best RFW African accuracy score is taken. For GAC [20], we consider the experiment with Estimated ethnic labels in order to compare in the same setting when ground truth labels are not given as an input to the model. The pretraining scheme is efficient when applied to the common choice of the FR method (ArcFace) and the best quality increase is for the ethnicities seen during pretraining (African and Asian). + +only take advantage of a face recognition dataset, and compare with them on BUPT-BalancedFace. + +The prior dataset $\mathcal{D}^{prior}$ for our work has been collected from YouTube via the procedure outlined in Subsec. 3.1. The dataset consists of two parts - African and East Asian ethnic groups - corresponding to two of the four benchmarks, these groups demonstrate higher error rates for face recognition systems compared to others in the branch of works [20, 49, 19]. Analogously, these two groups are known to be challenging according to NIST evaluation on demographics [22] and benchmarks of other tasks [6]. As an input for the scraping routine, we have used a set of 30-40 news channels corresponding to the respective part of the world. One frame per $P = 5$ seconds was taken, and only frames from the start through the 20th minute of each video have been considered. Since the target resolution for our network training is $112 \times 112$ , we only approve face bounding boxes selected by MTCNN if the face occupies at least $100 \mathrm{px}$ on each side. Additionally, we increase the decision thresholds for MTCNN to reduce the number of false positives during the detection process. The exact set of the news channels is enlisted in the Supplementary [1]. In total, 5 million images for the African group (AfricanFaceSet-5M) and 3 million images for the East Asian group (AsianFaceSet-3M) have been collected. + +One face recognition validation protocol is based on a balanced verification set Racial Faces In-the-Wild (RFW) [50] (following [53, 20, 19, 49]). It consists of 3K positive pairs of samples (pairs of images of the same per + +
RB-WebFace +TPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
ArcFace89.8686.7394.3193.82
ArcFace + Dprior(African)92.6290.2096.3195.85
ArcFace + Dprior(Asian)92.9190.6196.1595.83
ArcFace + Dprior(Afr+Asian)93.5491.5196.5196.17
increase+3.68+4.78+2.20+2.35
SphereFace92.0889.5195.5295.02
SphereFace + Dprior(Afr+Asian)92.8390.5296.2295.65
increase+0.75+1.01+0.70+0.63
GAC92.4889.1396.3394.66
GAC + Dprior(Afr+Asian)93.6090.8096.7695.46
increase+1.12+1.67+0.43+0.80
+ +
TPR @ FPR=10-4↑
ArcFace81.4876.8087.4786.89
ArcFace + D prior (African)85.5681.9991.4290.67
ArcFace + D prior (Asian)85.7982.4191.1290.36
ArcFace + D prior (Af+Asian)86.8483.7491.7491.19
increase+5.36+6.94+4.27+4.30
SphereFace83.6380.0489.1089.07
SphereFace + D prior (Af+Asian)85.1481.4790.6590.16
increase+1.51+1.43+1.55+1.09
GAC85.6480.2491.7288.12
GAC + D prior (Af+Asian)86.9182.4792.3989.48
increase+1.27+2.23+0.67+1.36
+ +Table 3: Comparison of the methods on the newly assembled RB-WebFace validation set. Publicly available authors' implementation of GAC [20] has been used to retrain the method and evaluate the quality. Here, we showcase the TPR given two pre-selected FPR thresholds to highlight the error rate difference separately for each ethnic group. For GAC, the setting GAC (Estimated) (see [20]) is used. This table indicates how the proposed pretraining scheme provides an additional boost for different pipelines. + +
RB-WebFace +TPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ 1% Dprior92.5990.4496.0695.72
+ 10% Dprior92.8490.7296.1295.64
+ 100% Dprior93.5491.5196.5196.17
+ +Table 4: Ablation study of the quality dependence on the $\left| {\mathcal{D}}_{\text{prior }}\right|$ prior dataset size on the newly assembled RB-WebFace validation set. ${\mathcal{D}}_{\text{prior }} = \left\{ \begin{array}{ll} \text{AfricanFaceSet - } 5\mathrm{M} & \cup \\ \text{AsianFaceSet - } 3\mathrm{M} & \end{array}\right\}$ was used in this set of experiments. + +son) and 3K negative pairs (pairs of images of different but similarly looking people, selected by a face recognition algorithm which can introduce selection bias) for each ethnic group, summing up to 24K pairs. Evaluation is carried out in an LFW-like protocol [27] that involves evaluating the backbone $f_{\theta,\psi}(I)$ on images of each pair and thresholding the resulting cosine distances between embeddings. Each set of 3K+3K pairs is constructed from 3K people comprising a subset of MS-Celeb-1M [23]. + +In addition to that, we propose a new test set $RB$ + +WebFace constructed in a similar fashion from the recently proposed million-scale identification dataset WebFace-42M (a cleaned version of WebFace-260M [60]). RB-WebFace is constructed by evaluating a pretrained ethnic group classifier on WebFace-42M to separate it into four ethnic groups (the group assignment is later refined by a consensus procedure). For each group, the largest possible number of distinct people is taken to construct positive and negative pairs. The details are provided in Supplementary [1], as well as the example pictures and comparison to typically used test datasets by the number of people and pairs. + +# 4.2. Evaluation + +Comparison with the state-of-the-art on RFW. The Table 2 contains the comparison to baseline (ArcFace R-50) and a few state-of-the-art methods (GAC [20], DebFace [19], and others). RFW verification accuracy is reported for each of the four ethnic groups. We demonstrate an increase in accuracy for all races compared to the baseline, which is on par with the state-of-the-art and outperforming on Caucasian and Indian ethnic groups. The largest increase w.r.t. the baseline is observed for the African group. It is important to note that our pretraining scheme can also be used to initialize other methods and enhance their results in a similar fashion. + +Comparison on newly assembled RB-WebFace test set. Since the number of positive and negative pairs is orders of magnitude different, we report the ROC curves values (TPR vs. FPR), better suited for the class-imbalanced evaluations, instead of reporting accuracy as for RFW. By sweeping the threshold (in [0.1, 0.75] range), we obtain the (TPR, FPR) pairs, and TPR @ FPR = $\{10^{-3}, 10^{-4}\}$ is reported in Table 3. Our method (denoted as Baseline + Dprior) is compared to the Baseline (ArcFace R-50). Additionally, we report the enhancement that our procedures provide to other methods [34, 20]. The main conceptual result that we demonstrate is the significant increase of the TPR vs. given FPR for all races and decrease of ethnic bias on RB-WebFace. A separate plot in the Supplementary [1] comprises the TPR measurements for a wide range of possible FPR values. We also provide the results for ResNet $\{34,100\}$ backbones in Table 6. + +Limited labeled data. The study in Fig. 4 describes the dependency of our method's accuracy on the number of samples in $\mathcal{D}^{facerec}$ . We demonstrate the difference between our method and the ArcFace R-50 baseline for each of the benchmarks, both on RFW and RB-WebFace. As expected, pretraining helps the most when the network is fine-tuned on more limited amounts of labeled data. This especially highlights the benefits of using self-supervised learning in these scenarios. + +Limited prior dataset. An ablation in Table 4 describes the quality dependence on the size of the prior dataset. As + +![](images/679df20abf3b04148a9bedefbaaf1a419a40a9bf10025e853a11d56d222e9057.jpg) + +![](images/ab3cdb9b485f32bdecbd869a5b91605316e6b06e1cb47adfc045fd308a795566.jpg) + +![](images/559836407d0f87f2ba7c894306097d7813976c9b06e5e89244f1cce1ff9dc4a9.jpg) + +![](images/d702ba6bd0d22c523a6c8ecdcaf49c91f560cc772924dbc94eeec6c784dae7da.jpg) + +![](images/8ec474258f7e2d6354bedb891ed6eaecacff9f7f758a4c964be8bbd46c62bfd0.jpg) + +![](images/dcda3b971923562fdc3c02e20e2652797cf347073ed4b04c07f4b41bc792c4a6.jpg) + +![](images/79224daf0f3b09991b8b03ceb38d5d6491edfa32fbaaeb76e7829d657f008430.jpg) +Figure 4: Ablation study of the dependence of the test quality, evaluated for RFW and RB-WebFace datasets, on the number of labeled samples from $|\mathcal{D}^{facerec}|$ used for fine-tuning. In this experiment, $\mathcal{D}^{prior} = \{\text{AfricanFaceSet-5M} \cup \text{AsianFaceSet-3M}\}$ , while $\mathcal{D}^{facerec} = \text{BUPT-BalancedFace}$ . For $\mathcal{D}^{facerec}$ , fraction of the data (100%, 10%, 1%) defines the number of uniformly sampled people left in the dataset. In the plot legend, Baseline is ArcFace R-50. + +![](images/12a3da1131f8093263c2a04ab381ff411ba4638ba43fc73b567f50da2950d7c3.jpg) + +![](images/b48df4ab325a9637731ad7e84cc7c45627f540538914bb2585b4f0b55195fcca.jpg) + +![](images/ee33af30be71be7c5596f67250025d79e4570f886631eb597c53227d06925c11.jpg) + +![](images/fa01872f2da9baec2dd3da6bd37eda04b364e57cf803ef9de2ad867bd8dc0c12.jpg) + +![](images/931351f4d3e0e2e80d60d79d581c3ff8f61c982f5353596a6f0fed24c3317d17.jpg) + +shown, gradually enlarging it results in a monotonous performance increase for all ethnic groups. This indicates that the use of large amounts of diverse unlabeled data significantly strengthens the method in the end. + +
RB-WebFace +TPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ AE on Dprior92.9390.8395.7395.68
+ VAE on Dprior92.4590.3995.6495.40
+ ours on Dprior93.5491.5196.5196.17
+ +
TPR @ FPR=10-4↑
Baseline81.4876.8087.4786.89
+ AE on Dprior85.8583.0990.4689.86
+ VAE on Dprior85.1982.0890.3889.25
+ ours on Dprior86.8483.7491.7491.19
+ +Can StyleGAN be replaced with a simpler encoder-decoder architecture? In Table 5, we compare our pretraining procedure to more conventional ones where the ResNet-50 encoder is pretrained in a symmetric autoencoder (AE) or in a symmetric variational autoencoder (VAE) setting. While these models can be pretrained in a single stage, our proposed StyleGAN+encoder approach re + +Table 5: Comparison of different pretraining strategies. Along with ours based on consequent training of StyleGAN and a ResNet encoder, one could employ vanilla (AE) and variational (VAE) autoencoders for pretraining (with the same ResNet-50 architecture as an encoder). Despite simpler simultaneous training of an encoder and a decoder in a single stage, the StyleGAN-based procedure is a more powerful prior. Baseline refers to the ArcFace R-50. + +
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
R-3484.5781.4791.1589.9173.6069.2383.1180.40
+ D prior90.9288.3495.2794.3983.0878.6989.2587.98
R-5089.8686.7394.3193.8281.4876.8087.4786.89
+ D prior93.5491.5196.5196.1786.8483.7491.7491.19
R-10093.7791.5596.3296.6287.2884.2891.4691.66
+ D prior94.8293.0096.7096.9288.7686.0092.2492.70
+ +Table 6: Comparison of different ResNet backbones on RB-WebFace. The baseline is ArcFace and $D^{prior} =$ AfricanFaceSet-5M U AsianFaceSet-3M. + +quires two separate pretraining stages but demonstrates better performance, which motivates the use of state-of-the-art generative models for face recognition. + +Why is AE/VAE less suitable than StyleGAN when finetuned for face recognition? We see two possible reasons why our procedure yields better results. First, StyleGAN-based architectures are tailored to be the state-of-the-art of face generation, which leads to the possible assumption that encoder + StyleGAN pipeline is capable of saving more useful information about the face features than e.g. VAE. Second, we observe that the quality of StyleGAN generations significantly improves when it is trained on larger amount of data, while for VAE it is not the case (6% over 32% improvement when 100x more data given – see Fig. 5). This scalability issue is also highlighted by the fact that AE/VAE, trained on the 100% of $\mathcal{D}^{prior}$ , perform for the face recognition task on par with StyleGAN, trained on only 1% of $\mathcal{D}^{prior}$ (see the Tables 5 and 4). + +Is it helpful to draw samples from StyleGAN instead of training an encoder for it? In this experiment, we used our trained StyleGAN to generate synthetic faces and add + +![](images/263ae190c8fbae7922cee2c6c12b891a16d7e70e39fcc6566de50fe50d8bce0c.jpg) +Figure 5: FID score, calculated over 100K random pictures from $\mathcal{D}^{prior} = \{\text{AfricanFaceSet-5M} \cup \text{AsianFaceSet-3M}\}$ vs. 100K random samples from VAE or StyleGAN. The score (the less the better) is given w.r.t. the percentage of $\mathcal{D}^{prior}$ used to train these models. We observe that the image generation quality, that FID represents, drops significantly when StyleGAN is provided with more data, while VAE does not scale to larger data sets that well. + +them to the training set. We first infer StyleGAN latent codes for all African and Asian samples in our training set (BUPT-BalancedFace). Since face recognition training always requires labeled images, we generate samples using latent codes closer to the inverted training images. In particular, we take random pairs $(I_1, I_2)$ of images of the same ethnic group (either African or Asian) and generate an interpolated sample by calculating convex combinations of their StyleGAN latent codes: $I_{\mathrm{comb}} = \lambda \cdot f_{\theta, \omega}'(I_1) + (1 - \lambda) \cdot f_{\theta, \omega}'(I_2)$ , $\lambda \sim U[0, 1]$ . The ground-truth label (always required for face recognition training) is constructed as a two-hot vector $(0, \dots, 0, 1 - \lambda, 0, \dots, 0, \lambda, 0, \dots, 0)$ (as influenced by MixUp [55]). In total, we augment the training set with an equal amount of synthetic interpolations (1.3 M), obtained from randomly drawn African-African and Asian-Asian pairs. The results in Table. 7 show that the added interpolations, despite the created imbalance and increased variability in the data, do not yield a comparable performance increase. In addition, we show that increasing the number of interpolations doesn't help, while pretraining scheme benefits from more pretraining data (see Table 4). + +We provide additional ablations and comparisons in the Supplementary [1]. + +# 5. Conclusions + +As an increasing number of modern computer vision methods become production-ready, new challenges regarding their final use are posed. In this work, we present a solution for improving the quality of facial recognition by pretraining the face image encoder with unlabeled data using + +
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
ArcFace R-50 + Dprior89.8686.7394.3193.8281.4876.8087.4786.89
93.5491.5196.5196.1786.8483.7491.7491.19
+ 1.3M interps91.8789.0796.0895.5584.6579.9791.1390.30
+ 6M interps89.4886.8594.9294.2280.9577.0088.7587.94
+ +Table 7: In this experiment, instead of pretraining our model as an encoder for StyleGAN, we use the StyleGAN and the trained encoder to draw samples from the learned face distribution and add them to our dataset. The results indicate that interpolation-based procedure, despite being more straightforward than the proposed pretraining-based scheme, is actually less efficient. + +StyleGAN and encoder. Furthermore, we additionally release two training datasets for unsupervised pretraining and one large-scale protocol for bias estimation. We show that we are able to tune the performance on different ethnicities by altering the composition of the unlabeled prior datasets. We hope that the released datasets, together with the protocol, will drive forward the research on mitigating the face recognition biases using self-supervised approaches. + +There are several directions for possible future work that we foresee. First, integrating the two steps of pretraining and fine-tuning into one would help to avoid forgetting of the pretrained weights. Second, publicly available datasets often don't possess enough variability (both in terms of capture conditions and demographics) [25] and typically feature only low-resolution images which might be a blocking factor for the fairness research [49, 30]. Accordingly, the results on the newly assembled RB-WebFace benchmark can be analyzed more thoroughly, which can bring new insights about the factors positively affecting bias mitigation in real-world scenarios. Finally, various architectures typically used for self-supervised learning (such as transformers or highly scalable generative models [43]), might allow for the construction of more flexible methods due to heterogeneous inputs and outputs. + +Legal concerns. The collected unlabeled data was collected anonymously in accordance with Standard YouTube License and CC YouTube License and does not contain personally identifiable information (PII). The data is released only in the form of links to YouTube videos and the corresponding timestamps, fully following good practice of similar data collection [10, 17, 32] and with the data subjects protection as per Art. 14 5(b) GDPR law. + +Acknowledgments. We gratefully acknowledge the support of this research by a TUM-IAS Hans Fischer Senior Fellowship, the ERC Starting Grant Scan2CAD (804724) and the Horizon Europe vera.ai project (101070093). We also thank Yawar Siddiqui for the helpful technical advice on StyleGAN part, Angela Dai for the video voiceover, and Dmitriy Karfagenskiy for the proofreading and corrections. + +# References + +[1] How to boost face recognition with stylegan?. Supplemental material, 2022. +[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711-6720, 2021. +[3] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[4] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992. +[5] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a" siamese" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993. +[6] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018. +[7] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGface2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face and Gesture Recognition, pages 3414-3424, 2018. +[8] Yunliang Chen and Jungseock Joo. Understanding and mitigating annotation bias in facial expression recognition. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14980-14991, 2021. +[9] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 539-546, 2005. +[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In Interspeech, 2018. +[11] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020. +[12] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In Proc. of the 58th Annual Meeting of the Association for Computational Linguistics, 2020. +[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. +[14] Emily Denton, Ben Hutchinson, Margaret Mitchell, Timnit Gebru, and Andrew Zaldivar. Image counterfactual sensitivity analysis for detecting unintended bias. arXiv preprint arXiv:1906.06439, 2019. + +[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, 2019. +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[17] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. +[18] Markos Georgopoulos, James Oldfield, Mihalis A Nicolaou, Yannis Panagakis, and Maja Pantic. Mitigating demographic bias in facial datasets with style-based multi-attribute transfer. International Journal of Computer Vision, 129:2288-2307, 2021. +[19] Sixue Gong, Xiaoming Liu, and Anil K Jain. Debface: Debiasing face recognition. arXiv preprint arXiv:1911.08080, 2019. +[20] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating Face Recognition Bias via Group Adaptive Classifier. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3414-3424, 2021. +[21] Google. Real Tone on Pixel 6. Be seen as you truly are., 2022. +[22] Patrick Grother, Mei Ngan, and Kayee Hanaoka. Face recognition vendor test (FVRT): Part 3, demographic effects. 2019. +[23] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-Celeb-1M: A dataset and benchmark for large-scale face recognition. In European Conference on Computer Vision (ECCV), 2016. +[24] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, pages 1735-1742, 2006. +[25] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in AI: the Casual Conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021. +[26] Mingjie He, Jie Zhang, Shiguang Shan, and Xilin Chen. Enhancing face recognition with self-supervised 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4062-4071, 2022. +[27] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. +[28] Di Kang, Debarun Dhar, and Antoni Chan. Incorporating side information by adaptive convolution. Advances in Neural Information Processing Systems, 30, 2017. + +[29] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. +[30] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Image resolution susceptibility of face recognition models. arXiv preprint arXiv:2107.03769, 2021. +[31] Wenyu Li, Tianchu Guo, Pengyu Li, Binghui Chen, Biao Wang, Wangmeng Zuo, and Lei Zhang. Virface: Enhancing face recognition via unlabeled shallow data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14729-14738, 2021. +[32] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3207-3216, 2020. +[33] Chun-Hsien Lin and Bing-Fei Wu. Domain adapting ability of self-supervised learning for face recognition. In IEEE International Conference on Image Processing (ICIP), pages 479–483, 2021. +[34] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 212-220, 2017. +[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv:1907.11692, 2019. +[36] Vongani Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying Bias in GANs through the Lens of Race. In European Conference on Computer Vision (ECCV), pages 344-360, 2022. +[37] Richard T Marriott, Sami Romdhani, and Liming Chen. A 3d gan for improved large-posed facial recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13445–13455, 2021. +[38] Tuan-Duy H Nguyen, Huu-Nghia H Nguyen, and Hieu Dao. Recognizing families through images with pretrained encoder. In IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 887-891, 2020. +[39] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10880-10890, 2021. +[40] Christian Rathgeb, Pawel Drozdowski, Naser Damer, Dinusha C Frings, and Christoph Busch. Demographic fairness in biometric systems: What do the experts say? IEEE Technology and Society Magazine, 41(4):71-82, 2022. +[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021. +[42] Joseph P Robinson, Gennady Livitz, Yann Henon, Can Qin, Yun Fu, and Samson Timoner. Face Recognition: Too Bias, + +or Not Too Bias? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020. +[43] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH Conference Proceedings, pages 1-10, 2022. +[44] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6902-6911, 2019. +[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. +[46] Fu-En Wang, Chien-Yi Wang, Min Sun, and Shang-Hong Lai. Mixfairface: Towards ultimate fairness via mixfair adapter in face recognition. In AAAI Conference on Artificial Intelligence, volume 37, pages 14531-14538, 2023. +[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5265-5274, 2018. +[48] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In ACM International Conference on Multimedia, pages 3779-3782, 2021. +[49] Mei Wang and Weihong Deng. Mitigating Bias in Face Recognition using Skewness-Aware Reinforcement Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9322-9331, 2020. +[50] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial bias by information maximization adaptation network. In International Conference on Computer Vision (ICCV), pages 692-702, 2019. +[51] Qingzhong Wang, Pengfei Zhang, Haoyi Xiong, and Jian Zhao. Face evolves: A high-performance face recognition library. arXiv preprint arXiv:2107.08621, 2021. +[52] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019. +[53] Xingkun Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, Jilin Li, Feiyue Huang, Yong Li, and Zhen Cui. Consistent Instance False Positive Improves Fairness in Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 578-586, 2021. +[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv:1411.7923, 2014. +[55] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. +[56] Haoyu Zhang, Marcel Grimmer, Raghavendra Ramachandra, Kiran Raja, and Christoph Busch. On the applicability of synthetic data for face recognition. In IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6, 2021. + +[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10):1499-1503, 2016. +[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. +[59] Song Zhang, Gehui Shen, Jinsong Huang, and Zhi-Hong Deng. Self-supervised learning aided class-incremental life-long learning. arXiv preprint arXiv:2006.05882, 2020. +[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. \ No newline at end of file diff --git a/2023/How to Boost Face Recognition with StyleGAN_/images.zip b/2023/How to Boost Face Recognition with StyleGAN_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1bbbc987065ff7d07a18824621536c0b06caf9da --- /dev/null +++ b/2023/How to Boost Face Recognition with StyleGAN_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16c410d80c28901f4f231d9f27584dcffdc9813f9369edf1f61a09dca5e61aab +size 628257 diff --git a/2023/How to Boost Face Recognition with StyleGAN_/layout.json b/2023/How to Boost Face Recognition with StyleGAN_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b7dad2de46021c33e2ffc4a26b9f5a6dd350c277 --- /dev/null +++ b/2023/How to Boost Face Recognition with StyleGAN_/layout.json @@ -0,0 +1,7489 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 445, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 445, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 445, + 121 + ], + "type": "text", + "content": "How to Boost Face Recognition with StyleGAN?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 455, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 455, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 455, + 171 + ], + "type": "text", + "content": "Artem Sevastopolsky1 Yury Malkov2,* Nikita Durasov3 Luisa Verdoliva1,4 Matthias Nießner1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 184, + 449, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 184, + 449, + 230 + ], + "spans": [ + { + "bbox": [ + 145, + 184, + 449, + 230 + ], + "type": "text", + "content": "1 Technical University of Munich, Germany 2 Twitter, US 3 École polytechnique fédérale de Lausanne, Switzerland 4 University Federico II of Naples, Italy" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 120, + 242, + 471, + 352 + ], + "blocks": [ + { + "bbox": [ + 120, + 242, + 471, + 352 + ], + "lines": [ + { + "bbox": [ + 120, + 242, + 471, + 352 + ], + "spans": [ + { + "bbox": [ + 120, + 242, + 471, + 352 + ], + "type": "image", + "image_path": "7e00cfd1644ea9e081094344af373d3dc09ed5284de51c502c7b5f0b50fb7315.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 357, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 45, + 357, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 45, + 357, + 547, + 418 + ], + "type": "text", + "content": "Figure 1: Our method is aimed at boosting the performance of face recognition. This is achieved by gathering a random image collection without face recognition labels (unlabeled set) and then fitting a mapping from an image to the StyleGAN latent space onto that collection. To learn this mapping, we use the pSp encoder architecture. For the downstream face recognition task, the same encoder is then fine-tuned on a (potentially much smaller) face recognition dataset with identity labels (labeled set)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 428, + 192, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 428, + 192, + 441 + ], + "spans": [ + { + "bbox": [ + 143, + 428, + 192, + 441 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 451, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 451, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 451, + 290, + 715 + ], + "type": "text", + "content": "State-of-the-art face recognition systems require vast amounts of labeled training data. Given the priority of privacy in face recognition applications, the data is limited to celebrity web crawls, which have issues such as limited numbers of identities. On the other hand, self-supervised revolution in the industry motivates research on the adaptation of related techniques to facial recognition. One of the most popular practical tricks is to augment the dataset by the samples drawn from generative models while preserving the identity. We show that a simple approach based on fine-tuning " + }, + { + "bbox": [ + 45, + 451, + 290, + 715 + ], + "type": "inline_equation", + "content": "pSp" + }, + { + "bbox": [ + 45, + 451, + 290, + 715 + ], + "type": "text", + "content": " encoder for StyleGAN allows to improve upon the state-of-the-art facial recognition and performs better compared to training on synthetic face identities. We also collect large-scale unlabeled datasets with controllable ethnic constitution - AfricanFaceSet-5M (5 million images of different people) and AsianFaceSet-3M (3 million images of different people) - and we show that pretraining on each of them improves recognition of the respective ethnicities (as well as others), while combining all unlabeled datasets results in the biggest performance increase. Our self-supervised strategy is the most useful with limited amounts of labeled training data," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 430, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 547, + 501 + ], + "type": "text", + "content": "which can be beneficial for more tailored face recognition tasks and when facing privacy concerns. Evaluation is based on a standard RFW dataset and a new large-scale RB-WebFace benchmark. The code and data are made publicly available at https://github.com/seval00/stylegan-for-facerec." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 514, + 387, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 387, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 387, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Modern face recognition methods rely on deep convolutional networks trained on large-scale datasets [54, 7, 23, 60]. These methods are now being integrated into a vast number of real-world applications, ranging from face unlock for smartphones and photo organizers to law enforcement systems and border control. A typical open face recognition dataset consists of web-crawled images of celebrities, leading to limited size and lack of balance in subgroups, such as ethnicity, age, etc. Training a state-of-the-art solution, however, requires enormous amounts of labeled data, scraping which may lead to privacy and legal issues. We suggest and study an alternative solution to using celebrity photos - pretraining the face recognition backbone on a generative task. Specifically, we first train StyleGAN2-ADA [29] on collected" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20924" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 206, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 206, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 206, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 206, + 194 + ], + "type": "image", + "image_path": "ce4aea9a1ba9b95c3edeace262404e5f6d803ae6dd38f975fa79f4dd5b12353a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "type": "text", + "content": "Figure 2: Our method is trained in three consecutive steps. First, we fit StyleGAN2-ADA to the face image distribution of the unlabeled prior dataset " + }, + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "type": "text", + "content": ". Second, the pSp encoder is trained (also on " + }, + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "type": "text", + "content": ") to map images to the latent codes in the learned latent space. Finally, the encoder, which is pretrained to extract meaningful features from an image, is fine-tuned for the downstream face recognition task with the ArcFace loss (similar losses can be used instead) on " + }, + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{facerec}" + }, + { + "bbox": [ + 46, + 206, + 547, + 268 + ], + "type": "text", + "content": ". The two first steps comprise the self-supervised pretraining stage; i.e., no identity labels are required for them." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 72, + 408, + 194 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 408, + 194 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 408, + 194 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 408, + 194 + ], + "type": "image", + "image_path": "4e6ae3af18f67dfb67f11c10417def155595b4175cdd61330fabc429e7a7bfe8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 413, + 72, + 544, + 194 + ], + "blocks": [ + { + "bbox": [ + 413, + 72, + 544, + 194 + ], + "lines": [ + { + "bbox": [ + 413, + 72, + 544, + 194 + ], + "spans": [ + { + "bbox": [ + 413, + 72, + 544, + 194 + ], + "type": "image", + "image_path": "a1ca5781353c1f3c74e4af754ea3f7ea328cecc1ab98deaf1d7c7f2c243a3e99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 288, + 399 + ], + "type": "text", + "content": "unlabeled data (which we later refer to as an unlabeled prior dataset) to fit the face image distribution. Subsequently, we train an encoder (following pixel2style2pixel (pSp) architecture [41]) that maps input images to vectors in the learned StyleGAN2-ADA latent space. Importantly, during the pretraining steps, no identity labels are used, so we can use diverse datasets crawled from the Internet without compromising privacy. Finally, we transfer the learned pSp encoder convolutional weights into the face recognition network and train it in a standard face recognition setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 592 + ], + "type": "text", + "content": "We show that, in contrast to training face recognition tasks on StyleGAN generated data (also demonstrated in [38, 39] and studied e.g. in [36]), our encoder pretraining step significantly boosts the final performance. The idea of augmenting face recognition datasets with synthetic data is widespread and constitutes many approaches, however, unclear and heuristic definition of the target label limits the amount of useful signal that can be transferred into the face recognition model this way. Our approach goes hand-in-hand with the current development of self-supervised learning [15, 35, 12, 16] and makes up one of the first approaches of its application to face recognition [26, 33]. This allows us to demonstrate vast improvements on limited labeled training data compared to the setup without self-supervised pretraining (for instance, " + }, + { + "bbox": [ + 46, + 401, + 287, + 592 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 401, + 287, + 592 + ], + "type": "text", + "content": " verification accuracy increase for only " + }, + { + "bbox": [ + 46, + 401, + 287, + 592 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 401, + 287, + 592 + ], + "type": "text", + "content": " of the labeled data used)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "The simplicity of the data collection procedure also allows us to control the distribution of the unlabeled data and thus influence the decrease of the error rates for specific demographic groups. Despite the fact that the current state-of-the-art algorithms often demonstrate very low average error rates [13, 47, 51, 48], it is considered unethical to integrate face recognition solutions that exhibit significant ethnic, age, or gender bias. Such bias is present both in open-source face recognition methods [20, 53, 49, 19, 28, 44, 46] and in comprehensively evaluated commercial face recog" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 279, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 338 + ], + "type": "text", + "content": "nition systems [22], resulting in significantly different error rates measured for the groups of interest. The topic has attracted significant attention in other areas of computer vision operating in face domain [8, 14, 21, 40] and in the media." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 340, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 546, + 518 + ], + "type": "text", + "content": "We constructively demonstrate that collecting large amounts of in-the-wild face images of a given group of interest is feasible (and can be done semi-automatically), while collecting datasets with identity labels is problematic. The labels require linking photos of the same person taken in different conditions, which means the person must be tracked. This typically constrains public datasets to celebrities, gathered using search engines [23, 54, 7, 60], while social networks and companies that provide services with photos use input from users. Second, the collected in-the-wild data, treated as a set of faces without identity labels (but labeled with the group attribution), can be efficiently used for self-supervised pretraining for face recognition networks, subsequently fine-tuned on the standard face recognition datasets (see Fig. 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "type": "text", + "content": "To summarize our main contribution, we present a novel self-supervised method for improving the performance of face recognition based on StyleGAN pretraining. This allows to leverage large-scale amounts of available unlabeled data for face recognition. While the improvement is the most significant on limited data, pretraining is also helpful for large-scale labeled datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 609, + 389, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 389, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 389, + 621 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Face recognition datasets. Several large-scale datasets of faces with identity labels have been released publicly, such as CASIA-WebFace [54], VGGFace2 [7], MS-Celeb-1M [23] and the very recent million-scale WebFace-42M dataset [60]. However, these datasets have been collected \"in the wild\", hence, they inevitably suffer from biases in terms of age, gender," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20925" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 71, + 542, + 222 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 542, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 542, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 542, + 222 + ], + "type": "table", + "html": "
Dataset name# people# images# pic./personethnic diversityacquisition
MS-Celeb-1M100K8.2M82unevenmostly American and British actors
VGGFace29.1K3.3M362unevenGoogle Image search
CASIA-WebFace10K494K49unevencelebrities from IMDb
CASIA-Face-Africa1.1K38K34all Africancontrolled, indoor and outdoor
MegaFace (unavailable)672K4.7M7unevenYahoo Flickr website search queries
DiveFace24K72K3balanced across 3 ethnicitiessubset of MegaFace
BUPT-BalancedFace28K1.3M46balanced across 4 ethnicitiescelebrities from MS-Celeb-1M
BUPT-GlobalFace38K2M52matches global populationcelebrities from MS-Celeb-1M
BUPT-TransferFace>10K600K6075% Caucasian. vs. otherscelebrities from MS-Celeb-1M
AfricanFaceSet-5M5M5MunlabeledAfrican majorityrandom faces from YouTube news videos
AsianFaceSet-3M3M3MunlabeledAsian majorityrandom faces from YouTube news videos
", + "image_path": "b0367a326945200d30a3b55d0b197774fe884de55d2536f1498aa5dfb90a06fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 224, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 547, + 285 + ], + "type": "text", + "content": "Table 1: Overview of the publicly available training sets in facial domain. Typically, the quality of a face recognition dataset is described by several factors: the number of people in the dataset, the number of images per person, and the diversity of capture conditions. While our AfricanFaceSet-5M and AsianFaceSet-3M datasets are not directly suitable for face recognition due to the absence of identity labels, they comprise more distinct people than existing large-scale datasets and contain a more diverse distribution of faces than only celebrities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 698 + ], + "type": "text", + "content": "or race. To better understand and stimulate research in fairness about face recognition, several datasets have been proposed. Examples of such labeled datasets include BUPT-Globalface [49] (2M images with ethnic distribution matching the world population) and BUPT-Balancedface [49] (1.3M images with the perfect ethnic split). Racial Faces in-the-Wild (RFW) [50] is a verification database that has been constructed from MS-Celeb-1M. Currently, it serves a standard fairness benchmark. Another verification dataset BFW has been introduced in [42], which, similarly to RFW, contains eight subgroups balanced across gender and ethnicity. The main feature of all these datasets – predefined ethnic split – allows one to disambiguate dataset bias and model bias for more precise methods evaluation. Still, the currently available data possesses a number of limitations. For instance, the number of distinct people is typically limited and is often many orders of magnitude smaller than the dataset size. Additionally, both identification and verification datasets require many images per person, which usually restricts the construction of open-source datasets to the search of celebrity pictures by text queries. Our two photo collections, released to the public, – AfricanFaceSet-5M and AsianFaceSet-3M, – fill a different gap in the space of available datasets. On the one hand, they neither have identity labels nor feature many images per person. On the other hand, these collections are large and focused on groups and conditions that are underrepresented in general face recognition datasets. Our evaluation dataset RB-WebFace, assembled from large-scale WebFace-42M, is a new verification dataset containing a significantly larger number of pairs and does not use external models to select the negatives (which introduces certain selection bias) compared to RFW, which is currently used as the main benchmark in the bias mitigation-focused branch of works." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 715 + ], + "type": "text", + "content": "Data augmentation. Generating synthetic data is a possi-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 291, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 547, + 518 + ], + "type": "text", + "content": "ble solution to improving the performance of face recognition and in some cases reducing the negative effect of the dataset bias. An idea of this kind is pursued in [37], a nonlinear 3DMM texture model is proposed to produce sharp renderings of faces from novel poses. This technique improves generalization with respect to the head pose and illumination. A number of approaches synthesize data via face generative models. For instance, in SynFace [39] random face images are constructed by a GAN with identity control, and labels are constructed by a procedure similar to MixUp augmentation [55]. Similarly, in [18] a style transfer GAN is used to simultaneously transfer multiple facial demographic attributes and generate diverse images for each attribute class. Authors of Virface [31] suggested a method for incorporating additional negative (impostor) pairs from the unlabeled data showing a boost in metrics. Zhang et al. [56] study the applicability of the data generated by StyleGAN by inspecting the distributions of downstream models but do not study its effect on face recognition metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 520, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 616 + ], + "type": "text", + "content": "Our approach is inspired by the latter idea of generating synthetic samples but takes a step forward by utilizing the generative model itself (StyleGAN with an encoder in our case). As shown in [41, 45, 2, 3], the StyleGAN encoders have high potential for both generating new realistic faces from a latent code and solving the inversion problem. This highlights the expressiveness of these models and richness of their internal representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Self-supervised pretraining. Despite the general dominance of supervised learning in practical ML and CV, approaches based on self-supervised learning have been evolving in various forms, such as self-organizing or siamese networks [4, 5, 9, 24]. Currently, self-supervised learning is the dominant approach in NLP with a wide spectrum of possible approaches [35, 12, 11] and is being actively adapted in computer vision [16]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "text", + "content": "20926" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 82, + 541, + 177 + ], + "blocks": [ + { + "bbox": [ + 52, + 82, + 541, + 177 + ], + "lines": [ + { + "bbox": [ + 52, + 82, + 541, + 177 + ], + "spans": [ + { + "bbox": [ + 52, + 82, + 541, + 177 + ], + "type": "image", + "image_path": "260341261875363e7bea12e74e27db49afbc07d0094fd2608d21960a0aab139e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 547, + 234 + ], + "type": "text", + "content": "Figure 3: Our data collection process starts by manually specifying a set of YouTube channels with a specific topic, e.g. a set of news channels of a desired town or part of the world. All videos are downloaded in the highest available quality, one frame per " + }, + { + "bbox": [ + 46, + 183, + 547, + 234 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 183, + 547, + 234 + ], + "type": "text", + "content": " seconds is extracted from each video, and all the faces found in the extracted frames are cropped, aligned by landmarks and resized to the target resolution. This way, we obtain millions of random faces with a desired demographic distribution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 386 + ], + "type": "text", + "content": "Analogously, we are witnessing the first attempts to apply self-supervised learning now being integrated into the face recognition frameworks. In 3D-BERL [26], the performance over multiple benchmarks is improved by a separate 3D reconstruction network branch. The work [33] studies the effect of self-supervised learning for domain transfer in face recognition. Incremental learning [52] can be bridged with self-supervised approaches to adapt to large number of target classes [59]. The procedure for collecting large-scale unlabeled datasets allows us to adapt self-supervised training in a more conventional fashion while proving its efficacy for our application." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 104, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 104, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 104, + 405 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 413, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 288, + 533 + ], + "type": "text", + "content": "Our pipeline is comprised of several stages. First, we train StyleGAN2-ADA to fit the face distribution on the unlabeled dataset. Second, the pSp encoder is trained (also on the unlabeled dataset), that will define a feature extractor well-suited for the group of interest. Finally, the encoder is fine-tuned for the downstream face recognition task. The entire procedure is outlined in Subsec. 3.2 and visually described in Fig. 2. Before describing the method itself, we outline the schematics of our prior dataset collection procedures in Subsec. 3.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 178, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 178, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 178, + 552 + ], + "type": "text", + "content": "3.1. Prior dataset collection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Due to the nature of the task, the prior dataset must contain the samples from the group of interest. At the same time, samples from the prior dataset do not require identity labels of any kind, unlike samples of the face recognition datasets. In practice, this removes the restriction of linking photos of the same person (something that might be considered a violation of privacy), and thus enlarges the search space and simplifies data collection. Still, fulfilling several requirements for the prior dataset remains a challenge, such as: collecting faces only from a specific group (e.g. an ethnic or gender group), obtaining a large number of them (preferably, an order of magnitude more than in the face recognition dataset to perform subsequent transfer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 242, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 433 + ], + "type": "text", + "content": "learning), and using only data legally allowed for collection. We found semi-automatic YouTube channel scraping to be an efficient solution that satisfies these requirements. In particular, we propose to select a set of publicly available YouTube channels dedicated to the desired group of interest. The channel names are the only entry point and the only manually performed step for the scraping procedure (see Fig. 3). Typically, the requirement for the channels is having them systematically featuring different people; a possible example would be a set of news channels released in a country of choice. All videos are downloaded from every channel, one frame per " + }, + { + "bbox": [ + 304, + 242, + 547, + 433 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 242, + 547, + 433 + ], + "type": "text", + "content": " seconds of each video is extracted, and all faces are detected, cropped and aligned by landmarks via MTCNN [57] library. Our data has been scraped from a predefined set of channels such as news channels and others." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "content": "The unlabeled prior dataset will later be referred to as " + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "content": ". Table 1 describes the relative difference of the collected data to the datasets typically used for face recognition. Note however, that despite the latter being directly inapplicable for training due to the lack of the labels, they potentially contain more people than the others." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 515, + 517, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 517, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 517, + 528 + ], + "type": "text", + "content": "3.2. Architecture and the training procedure" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "The central part of our pipeline is a single face recognition convnet " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta, \\psi}(I)" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " (a backbone), that takes a single RGB image " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " as an input, and outputs a 512-dimensional vector " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "e \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " (face embedding). Typically, the backbone is trained on a training dataset with angular margin based losses, such as SphereFace [34], ArcFace [13], and others, in standard works on face recognition. In our pipeline, we follow the same procedure, but only after a pretraining stage is performed. In order to do so, we first train the StyleGAN2-ADA [29] generator " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "g_{\\phi}(w(z))" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " that transforms a latent vector " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z} \\subseteq \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " into unfolded latent space " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "w(z) \\in \\mathcal{W} + \\subseteq \\mathbb{R}^{L \\times 512}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ", and then into a realistic face image " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". The generator " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "g_{\\phi}(w(z))" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " is trained together with a discriminator that allows the generator to learn the distribution of faces in " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "D^{prior}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". The training procedure for StyleGAN2-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20927" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": "ADA follows the one in the corresponding paper. Secondly, we introduce a network " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "f_{\\theta, \\omega}'(x)" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": ", which is trained as an encoder for the StyleGAN generator, i.e. a network that solves an inverse problem: given an RGB image " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": ", predict an unfolded latent code " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "w \\in \\mathcal{W}+" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": ", such that the corresponding reconstruction " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\hat{I} = g_{\\phi}(w)" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": " is as close to the input " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": " as possible. A set of different possible approaches has appeared recently for training an encoder for StyleGAN2, ranging from applying a ConvNet with style-predicting layers [41] to employing hyper-networks [2]. In our method, the encoder architecture follows the pSp method [41], that proposes predicting a latent code by a ConvNet divided into a convolution part (with parameters " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": "), following a ResNet architecture, and a set of fully-convolutional style predictors - map2style blocks (with parameters " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": "). The training procedure for the encoder follows the one outlined in [41] with a few variations. Namely, we only keep the fidelity losses (standard L2 and neural-based LPIPS [58]) and disable the identity loss proposed by the authors. During the encoder training, the StyleGAN generator remains frozen in order to fix the latent space. A pair of networks " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "f_{\\theta, \\omega}'(x)" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "g_{\\phi}(w(z))" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": " can be seen as an asymmetric autoencoder (since the number of encoder parameters highly surpasses the number of generator parameters) with the decoder pretrained for a generative task and the encoder subsequently trained for a discriminative (regressive) task. The visual quality and the identity preservation of the reconstruction " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": " indirectly defines the expressiveness of the latent code " + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 72, + 289, + 419 + ], + "type": "text", + "content": " predicted by the encoder." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": "Finally, after the encoder " + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta,\\omega}'(I)" + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": " is trained, we transfer its convolutional parameters " + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": " into the face recognition backbone " + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta,\\psi}(I)" + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": ", which also comprises a set of new parameters " + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": ", corresponding to a fully-connected layer in the end of the network. We repeat the standard training procedure for face recognition, e.g. described in ArcFace [13] and others. The backbone is trained on a face recognition dataset " + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "inline_equation", + "content": "D^{facerec} = \\{\\mathcal{D}_1, \\dots, \\mathcal{D}_N\\}" + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": ", where each group " + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": " corresponds to a set of images of the same person " + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "inline_equation", + "content": "\\# i" + }, + { + "bbox": [ + 46, + 420, + 289, + 564 + ], + "type": "text", + "content": ". The slight differences in hyperparameters that we found beneficial when used with pretraining and technical details are described in the Supplementary [1]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 578, + 100, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 100, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 100, + 590 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 598, + 206, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 206, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 206, + 611 + ], + "type": "text", + "content": "4.1. Data and evaluation protocol" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "Our system requires two datasets - a labeled face recognition dataset " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{facerec}" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": " and an unlabeled prior dataset " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": ". As for the " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{facerec}" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": ", we employ BUPT-BalancedFace [49] due to its ethnic and gender balance guarantees. Namely, it consists of 1.3 million images belonging to 28K different people, divided into 4 ethnic groups of equal size - 7K African, East Asian, Indian, and Caucasian people each. Other state-of-the-art methods can" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 57, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 309, + 57, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 309, + 57, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 309, + 57, + 545, + 185 + ], + "type": "table", + "html": "
RFW, accuracy %,↑
Cauc.AfricanAsianIndianAvg ↑Std ↓
ArcFace R-50* [13]96.1893.9893.7294.6794.641.11
CosFace R-50* [47]95.1293.9392.9892.9393.740.89
DebFace* [19]95.9593.6794.3394.7894.680.83
ACNN* [28]96.1294.0093.6794.5594.580.94
PFE* [44]96.3895.1794.2794.6095.110.93
RL-RBN* [49]96.2795.0094.8294.6895.190.63
GAC R-50* [20]96.2794.4094.3294.7794.940.79
Baseline (ArcFace)96.0094.0093.0894.4894.391.06
+ Dprior (African)96.3594.3793.6294.8894.811.00
+ Dprior (Asian)96.3894.6794.0395.0395.030.86
+ Dprior (Afr+Asian)96.5295.0093.9094.9395.090.94
increase+0.52+1.00+0.82+0.45
", + "image_path": "64a4f9442e2d2ef36f6aa40e39d71dcaebf7423c91310ab57e73719d48d06052.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 194, + 547, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 350 + ], + "type": "text", + "content": "Table 2: Comparison of the verification accuracy of the methods on RFW validation set. Asterisk “*” indicates that the numbers are directly taken from the respective table in [20]. In case of the ArcFace R-50 baseline, it also corresponds to a slightly different reported training procedure. For other methods, the epoch with the best RFW African accuracy score is taken. For GAC [20], we consider the experiment with Estimated ethnic labels in order to compare in the same setting when ground truth labels are not given as an input to the model. The pretraining scheme is efficient when applied to the common choice of the FR method (ArcFace) and the best quality increase is for the ethnicities seen during pretraining (African and Asian)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 364, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 387 + ], + "type": "text", + "content": "only take advantage of a face recognition dataset, and compare with them on BUPT-BalancedFace." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "text", + "content": "The prior dataset " + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "text", + "content": " for our work has been collected from YouTube via the procedure outlined in Subsec. 3.1. The dataset consists of two parts - African and East Asian ethnic groups - corresponding to two of the four benchmarks, these groups demonstrate higher error rates for face recognition systems compared to others in the branch of works [20, 49, 19]. Analogously, these two groups are known to be challenging according to NIST evaluation on demographics [22] and benchmarks of other tasks [6]. As an input for the scraping routine, we have used a set of 30-40 news channels corresponding to the respective part of the world. One frame per " + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "inline_equation", + "content": "P = 5" + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "text", + "content": " seconds was taken, and only frames from the start through the 20th minute of each video have been considered. Since the target resolution for our network training is " + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "text", + "content": ", we only approve face bounding boxes selected by MTCNN if the face occupies at least " + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "inline_equation", + "content": "100 \\mathrm{px}" + }, + { + "bbox": [ + 304, + 389, + 546, + 664 + ], + "type": "text", + "content": " on each side. Additionally, we increase the decision thresholds for MTCNN to reduce the number of false positives during the detection process. The exact set of the news channels is enlisted in the Supplementary [1]. In total, 5 million images for the African group (AfricanFaceSet-5M) and 3 million images for the East Asian group (AsianFaceSet-3M) have been collected." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "One face recognition validation protocol is based on a balanced verification set Racial Faces In-the-Wild (RFW) [50] (following [53, 20, 19, 49]). It consists of 3K positive pairs of samples (pairs of images of the same per" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20928" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 209 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 209 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 209 + ], + "type": "table", + "html": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
ArcFace89.8686.7394.3193.82
ArcFace + Dprior(African)92.6290.2096.3195.85
ArcFace + Dprior(Asian)92.9190.6196.1595.83
ArcFace + Dprior(Afr+Asian)93.5491.5196.5196.17
increase+3.68+4.78+2.20+2.35
SphereFace92.0889.5195.5295.02
SphereFace + Dprior(Afr+Asian)92.8390.5296.2295.65
increase+0.75+1.01+0.70+0.63
GAC92.4889.1396.3394.66
GAC + Dprior(Afr+Asian)93.6090.8096.7695.46
increase+1.12+1.67+0.43+0.80
", + "image_path": "804a7d1c1a168d040f257b14d55d16e8dbb9ad7224969f03acbb3f95b7d6514a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 211, + 285, + 335 + ], + "blocks": [ + { + "bbox": [ + 50, + 211, + 285, + 335 + ], + "lines": [ + { + "bbox": [ + 50, + 211, + 285, + 335 + ], + "spans": [ + { + "bbox": [ + 50, + 211, + 285, + 335 + ], + "type": "table", + "html": "
TPR @ FPR=10-4↑
ArcFace81.4876.8087.4786.89
ArcFace + D prior (African)85.5681.9991.4290.67
ArcFace + D prior (Asian)85.7982.4191.1290.36
ArcFace + D prior (Af+Asian)86.8483.7491.7491.19
increase+5.36+6.94+4.27+4.30
SphereFace83.6380.0489.1089.07
SphereFace + D prior (Af+Asian)85.1481.4790.6590.16
increase+1.51+1.43+1.55+1.09
GAC85.6480.2491.7288.12
GAC + D prior (Af+Asian)86.9182.4792.3989.48
increase+1.27+2.23+0.67+1.36
", + "image_path": "e6afa4758851964cc581f0f262759cec2476b7f9d43f80f8b8cf4221b836b1dd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 346, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 453 + ], + "type": "text", + "content": "Table 3: Comparison of the methods on the newly assembled RB-WebFace validation set. Publicly available authors' implementation of GAC [20] has been used to retrain the method and evaluate the quality. Here, we showcase the TPR given two pre-selected FPR thresholds to highlight the error rate difference separately for each ethnic group. For GAC, the setting GAC (Estimated) (see [20]) is used. This table indicates how the proposed pretraining scheme provides an additional boost for different pipelines." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 74, + 457, + 261, + 529 + ], + "blocks": [ + { + "bbox": [ + 74, + 457, + 261, + 529 + ], + "lines": [ + { + "bbox": [ + 74, + 457, + 261, + 529 + ], + "spans": [ + { + "bbox": [ + 74, + 457, + 261, + 529 + ], + "type": "table", + "html": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ 1% Dprior92.5990.4496.0695.72
+ 10% Dprior92.8490.7296.1295.64
+ 100% Dprior93.5491.5196.5196.17
", + "image_path": "393219b238b78ed44dd0e5b47d944a16bfab1348d9dfee851a7426654498f459.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 538, + 287, + 586 + ], + "lines": [ + { + "bbox": [ + 47, + 538, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 287, + 586 + ], + "type": "text", + "content": "Table 4: Ablation study of the quality dependence on the " + }, + { + "bbox": [ + 47, + 538, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\left| {\\mathcal{D}}_{\\text{prior }}\\right|" + }, + { + "bbox": [ + 47, + 538, + 287, + 586 + ], + "type": "text", + "content": " prior dataset size on the newly assembled RB-WebFace validation set. " + }, + { + "bbox": [ + 47, + 538, + 287, + 586 + ], + "type": "inline_equation", + "content": "{\\mathcal{D}}_{\\text{prior }} = \\left\\{ \\begin{array}{ll} \\text{AfricanFaceSet - } 5\\mathrm{M} & \\cup \\\\ \\text{AsianFaceSet - } 3\\mathrm{M} & \\end{array}\\right\\}" + }, + { + "bbox": [ + 47, + 538, + 287, + 586 + ], + "type": "text", + "content": " was used in this set of experiments." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 700 + ], + "type": "text", + "content": "son) and 3K negative pairs (pairs of images of different but similarly looking people, selected by a face recognition algorithm which can introduce selection bias) for each ethnic group, summing up to 24K pairs. Evaluation is carried out in an LFW-like protocol [27] that involves evaluating the backbone " + }, + { + "bbox": [ + 46, + 594, + 286, + 700 + ], + "type": "inline_equation", + "content": "f_{\\theta,\\psi}(I)" + }, + { + "bbox": [ + 46, + 594, + 286, + 700 + ], + "type": "text", + "content": " on images of each pair and thresholding the resulting cosine distances between embeddings. Each set of 3K+3K pairs is constructed from 3K people comprising a subset of MS-Celeb-1M [23]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "In addition to that, we propose a new test set " + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "inline_equation", + "content": "RB" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "WebFace constructed in a similar fashion from the recently proposed million-scale identification dataset WebFace-42M (a cleaned version of WebFace-260M [60]). RB-WebFace is constructed by evaluating a pretrained ethnic group classifier on WebFace-42M to separate it into four ethnic groups (the group assignment is later refined by a consensus procedure). For each group, the largest possible number of distinct people is taken to construct positive and negative pairs. The details are provided in Supplementary [1], as well as the example pictures and comparison to typically used test datasets by the number of people and pairs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 213, + 379, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 379, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 379, + 224 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 232, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 376 + ], + "type": "text", + "content": "Comparison with the state-of-the-art on RFW. The Table 2 contains the comparison to baseline (ArcFace R-50) and a few state-of-the-art methods (GAC [20], DebFace [19], and others). RFW verification accuracy is reported for each of the four ethnic groups. We demonstrate an increase in accuracy for all races compared to the baseline, which is on par with the state-of-the-art and outperforming on Caucasian and Indian ethnic groups. The largest increase w.r.t. the baseline is observed for the African group. It is important to note that our pretraining scheme can also be used to initialize other methods and enhance their results in a similar fashion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": "Comparison on newly assembled RB-WebFace test set. Since the number of positive and negative pairs is orders of magnitude different, we report the ROC curves values (TPR vs. FPR), better suited for the class-imbalanced evaluations, instead of reporting accuracy as for RFW. By sweeping the threshold (in [0.1, 0.75] range), we obtain the (TPR, FPR) pairs, and TPR @ FPR = " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\{10^{-3}, 10^{-4}\\}" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": " is reported in Table 3. Our method (denoted as Baseline + Dprior) is compared to the Baseline (ArcFace R-50). Additionally, we report the enhancement that our procedures provide to other methods [34, 20]. The main conceptual result that we demonstrate is the significant increase of the TPR vs. given FPR for all races and decrease of ethnic bias on RB-WebFace. A separate plot in the Supplementary [1] comprises the TPR measurements for a wide range of possible FPR values. We also provide the results for ResNet" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\{34,100\\}" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": " backbones in Table 6." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 581, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 687 + ], + "type": "text", + "content": "Limited labeled data. The study in Fig. 4 describes the dependency of our method's accuracy on the number of samples in " + }, + { + "bbox": [ + 304, + 581, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{facerec}" + }, + { + "bbox": [ + 304, + 581, + 545, + 687 + ], + "type": "text", + "content": ". We demonstrate the difference between our method and the ArcFace R-50 baseline for each of the benchmarks, both on RFW and RB-WebFace. As expected, pretraining helps the most when the network is fine-tuned on more limited amounts of labeled data. This especially highlights the benefits of using self-supervised learning in these scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Limited prior dataset. An ablation in Table 4 describes the quality dependence on the size of the prior dataset. As" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20929" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 128, + 160 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 128, + 160 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 128, + 160 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 128, + 160 + ], + "type": "image", + "image_path": "679df20abf3b04148a9bedefbaaf1a419a40a9bf10025e853a11d56d222e9057.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 208, + 160 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 208, + 160 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 208, + 160 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 208, + 160 + ], + "type": "image", + "image_path": "ab3cdb9b485f32bdecbd869a5b91605316e6b06e1cb47adfc045fd308a795566.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 294, + 160 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 294, + 160 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 294, + 160 + ], + "type": "image", + "image_path": "559836407d0f87f2ba7c894306097d7813976c9b06e5e89244f1cce1ff9dc4a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 295, + 70, + 375, + 160 + ], + "blocks": [ + { + "bbox": [ + 295, + 70, + 375, + 160 + ], + "lines": [ + { + "bbox": [ + 295, + 70, + 375, + 160 + ], + "spans": [ + { + "bbox": [ + 295, + 70, + 375, + 160 + ], + "type": "image", + "image_path": "d702ba6bd0d22c523a6c8ecdcaf49c91f560cc772924dbc94eeec6c784dae7da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 383, + 70, + 462, + 160 + ], + "blocks": [ + { + "bbox": [ + 383, + 70, + 462, + 160 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 462, + 160 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 462, + 160 + ], + "type": "image", + "image_path": "8ec474258f7e2d6354bedb891ed6eaecacff9f7f758a4c964be8bbd46c62bfd0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 544, + 160 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 544, + 160 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 544, + 160 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 544, + 160 + ], + "type": "image", + "image_path": "dcda3b971923562fdc3c02e20e2652797cf347073ed4b04c07f4b41bc792c4a6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 47, + 163, + 126, + 254 + ], + "blocks": [ + { + "bbox": [ + 47, + 163, + 126, + 254 + ], + "lines": [ + { + "bbox": [ + 47, + 163, + 126, + 254 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 126, + 254 + ], + "type": "image", + "image_path": "79224daf0f3b09991b8b03ceb38d5d6491edfa32fbaaeb76e7829d657f008430.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "text", + "content": "Figure 4: Ablation study of the dependence of the test quality, evaluated for RFW and RB-WebFace datasets, on the number of labeled samples from " + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}^{facerec}|" + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "text", + "content": " used for fine-tuning. In this experiment, " + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior} = \\{\\text{AfricanFaceSet-5M} \\cup \\text{AsianFaceSet-3M}\\}" + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{facerec} = \\text{BUPT-BalancedFace}" + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{facerec}" + }, + { + "bbox": [ + 46, + 261, + 547, + 309 + ], + "type": "text", + "content": ", fraction of the data (100%, 10%, 1%) defines the number of uniformly sampled people left in the dataset. In the plot legend, Baseline is ArcFace R-50." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 129, + 163, + 209, + 254 + ], + "blocks": [ + { + "bbox": [ + 129, + 163, + 209, + 254 + ], + "lines": [ + { + "bbox": [ + 129, + 163, + 209, + 254 + ], + "spans": [ + { + "bbox": [ + 129, + 163, + 209, + 254 + ], + "type": "image", + "image_path": "12a3da1131f8093263c2a04ab381ff411ba4638ba43fc73b567f50da2950d7c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 215, + 163, + 294, + 254 + ], + "blocks": [ + { + "bbox": [ + 215, + 163, + 294, + 254 + ], + "lines": [ + { + "bbox": [ + 215, + 163, + 294, + 254 + ], + "spans": [ + { + "bbox": [ + 215, + 163, + 294, + 254 + ], + "type": "image", + "image_path": "b48df4ab325a9637731ad7e84cc7c45627f540538914bb2585b4f0b55195fcca.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 295, + 163, + 375, + 254 + ], + "blocks": [ + { + "bbox": [ + 295, + 163, + 375, + 254 + ], + "lines": [ + { + "bbox": [ + 295, + 163, + 375, + 254 + ], + "spans": [ + { + "bbox": [ + 295, + 163, + 375, + 254 + ], + "type": "image", + "image_path": "ee33af30be71be7c5596f67250025d79e4570f886631eb597c53227d06925c11.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 383, + 163, + 462, + 254 + ], + "blocks": [ + { + "bbox": [ + 383, + 163, + 462, + 254 + ], + "lines": [ + { + "bbox": [ + 383, + 163, + 462, + 254 + ], + "spans": [ + { + "bbox": [ + 383, + 163, + 462, + 254 + ], + "type": "image", + "image_path": "fa01872f2da9baec2dd3da6bd37eda04b364e57cf803ef9de2ad867bd8dc0c12.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 465, + 163, + 544, + 254 + ], + "blocks": [ + { + "bbox": [ + 465, + 163, + 544, + 254 + ], + "lines": [ + { + "bbox": [ + 465, + 163, + 544, + 254 + ], + "spans": [ + { + "bbox": [ + 465, + 163, + 544, + 254 + ], + "type": "image", + "image_path": "931351f4d3e0e2e80d60d79d581c3ff8f61c982f5353596a6f0fed24c3317d17.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 322, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 370 + ], + "type": "text", + "content": "shown, gradually enlarging it results in a monotonous performance increase for all ethnic groups. This indicates that the use of large amounts of diverse unlabeled data significantly strengthens the method in the end." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 67, + 376, + 267, + 452 + ], + "blocks": [ + { + "bbox": [ + 67, + 376, + 267, + 452 + ], + "lines": [ + { + "bbox": [ + 67, + 376, + 267, + 452 + ], + "spans": [ + { + "bbox": [ + 67, + 376, + 267, + 452 + ], + "type": "table", + "html": "
RB-WebFace\nTPR @ FPR=10-3↑
Cauc.AfricanAsianIndian
Baseline89.8686.7394.3193.82
+ AE on Dprior92.9390.8395.7395.68
+ VAE on Dprior92.4590.3995.6495.40
+ ours on Dprior93.5491.5196.5196.17
", + "image_path": "4e3b1b8ffbe287ed315370f1c3af212102752d7947db2a2f1882ebdf05d42fe2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 67, + 458, + 267, + 513 + ], + "blocks": [ + { + "bbox": [ + 67, + 458, + 267, + 513 + ], + "lines": [ + { + "bbox": [ + 67, + 458, + 267, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 458, + 267, + 513 + ], + "type": "table", + "html": "
TPR @ FPR=10-4↑
Baseline81.4876.8087.4786.89
+ AE on Dprior85.8583.0990.4689.86
+ VAE on Dprior85.1982.0890.3889.25
+ ours on Dprior86.8483.7491.7491.19
", + "image_path": "442fe9c8bbc7ead18a507a98e3bdf7baa7094900351f5406b9f812d9a600e568.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Can StyleGAN be replaced with a simpler encoder-decoder architecture? In Table 5, we compare our pretraining procedure to more conventional ones where the ResNet-50 encoder is pretrained in a symmetric autoencoder (AE) or in a symmetric variational autoencoder (VAE) setting. While these models can be pretrained in a single stage, our proposed StyleGAN+encoder approach re" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 309, + 320, + 544, + 399 + ], + "blocks": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 616 + ], + "type": "text", + "content": "Table 5: Comparison of different pretraining strategies. Along with ours based on consequent training of StyleGAN and a ResNet encoder, one could employ vanilla (AE) and variational (VAE) autoencoders for pretraining (with the same ResNet-50 architecture as an encoder). Despite simpler simultaneous training of an encoder and a decoder in a single stage, the StyleGAN-based procedure is a more powerful prior. Baseline refers to the ArcFace R-50." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 320, + 544, + 399 + ], + "lines": [ + { + "bbox": [ + 309, + 320, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 309, + 320, + 544, + 399 + ], + "type": "table", + "html": "
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
R-3484.5781.4791.1589.9173.6069.2383.1180.40
+ D prior90.9288.3495.2794.3983.0878.6989.2587.98
R-5089.8686.7394.3193.8281.4876.8087.4786.89
+ D prior93.5491.5196.5196.1786.8483.7491.7491.19
R-10093.7791.5596.3296.6287.2884.2891.4691.66
+ D prior94.8293.0096.7096.9288.7686.0092.2492.70
", + "image_path": "f8fa7838bace5caec5484773e22ae68182f32ded4118c5517b3ca8f83441cad2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 408, + 547, + 444 + ], + "lines": [ + { + "bbox": [ + 305, + 408, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 547, + 444 + ], + "type": "text", + "content": "Table 6: Comparison of different ResNet backbones on RB-WebFace. The baseline is ArcFace and " + }, + { + "bbox": [ + 305, + 408, + 547, + 444 + ], + "type": "inline_equation", + "content": "D^{prior} =" + }, + { + "bbox": [ + 305, + 408, + 547, + 444 + ], + "type": "text", + "content": " AfricanFaceSet-5M U AsianFaceSet-3M." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 457, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 493 + ], + "type": "text", + "content": "quires two separate pretraining stages but demonstrates better performance, which motivates the use of state-of-the-art generative models for face recognition." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "text", + "content": "Why is AE/VAE less suitable than StyleGAN when finetuned for face recognition? We see two possible reasons why our procedure yields better results. First, StyleGAN-based architectures are tailored to be the state-of-the-art of face generation, which leads to the possible assumption that encoder + StyleGAN pipeline is capable of saving more useful information about the face features than e.g. VAE. Second, we observe that the quality of StyleGAN generations significantly improves when it is trained on larger amount of data, while for VAE it is not the case (6% over 32% improvement when 100x more data given – see Fig. 5). This scalability issue is also highlighted by the fact that AE/VAE, trained on the 100% of " + }, + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "text", + "content": ", perform for the face recognition task on par with StyleGAN, trained on only 1% of " + }, + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 304, + 495, + 546, + 674 + ], + "type": "text", + "content": " (see the Tables 5 and 4)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Is it helpful to draw samples from StyleGAN instead of training an encoder for it? In this experiment, we used our trained StyleGAN to generate synthetic faces and add" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20930" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 62, + 232, + 235 + ], + "blocks": [ + { + "bbox": [ + 89, + 62, + 232, + 235 + ], + "lines": [ + { + "bbox": [ + 89, + 62, + 232, + 235 + ], + "spans": [ + { + "bbox": [ + 89, + 62, + 232, + 235 + ], + "type": "image", + "image_path": "263ae190c8fbae7922cee2c6c12b891a16d7e70e39fcc6566de50fe50d8bce0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 289, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 336 + ], + "type": "text", + "content": "Figure 5: FID score, calculated over 100K random pictures from " + }, + { + "bbox": [ + 46, + 239, + 289, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior} = \\{\\text{AfricanFaceSet-5M} \\cup \\text{AsianFaceSet-3M}\\}" + }, + { + "bbox": [ + 46, + 239, + 289, + 336 + ], + "type": "text", + "content": " vs. 100K random samples from VAE or StyleGAN. The score (the less the better) is given w.r.t. the percentage of " + }, + { + "bbox": [ + 46, + 239, + 289, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{prior}" + }, + { + "bbox": [ + 46, + 239, + 289, + 336 + ], + "type": "text", + "content": " used to train these models. We observe that the image generation quality, that FID represents, drops significantly when StyleGAN is provided with more data, while VAE does not scale to larger data sets that well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "text", + "content": "them to the training set. We first infer StyleGAN latent codes for all African and Asian samples in our training set (BUPT-BalancedFace). Since face recognition training always requires labeled images, we generate samples using latent codes closer to the inverted training images. In particular, we take random pairs " + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "inline_equation", + "content": "(I_1, I_2)" + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "text", + "content": " of images of the same ethnic group (either African or Asian) and generate an interpolated sample by calculating convex combinations of their StyleGAN latent codes: " + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{comb}} = \\lambda \\cdot f_{\\theta, \\omega}'(I_1) + (1 - \\lambda) \\cdot f_{\\theta, \\omega}'(I_2)" + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\lambda \\sim U[0, 1]" + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "text", + "content": ". The ground-truth label (always required for face recognition training) is constructed as a two-hot vector " + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "inline_equation", + "content": "(0, \\dots, 0, 1 - \\lambda, 0, \\dots, 0, \\lambda, 0, \\dots, 0)" + }, + { + "bbox": [ + 46, + 343, + 287, + 595 + ], + "type": "text", + "content": " (as influenced by MixUp [55]). In total, we augment the training set with an equal amount of synthetic interpolations (1.3 M), obtained from randomly drawn African-African and Asian-Asian pairs. The results in Table. 7 show that the added interpolations, despite the created imbalance and increased variability in the data, do not yield a comparable performance increase. In addition, we show that increasing the number of interpolations doesn't help, while pretraining scheme benefits from more pretraining data (see Table 4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 596, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 287, + 620 + ], + "type": "text", + "content": "We provide additional ablations and comparisons in the Supplementary [1]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 633, + 125, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 125, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 125, + 645 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "As an increasing number of modern computer vision methods become production-ready, new challenges regarding their final use are posed. In this work, we present a solution for improving the quality of facial recognition by pretraining the face image encoder with unlabeled data using" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 128 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 128 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 128 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 128 + ], + "type": "table", + "html": "
RB-WebFace
TPR @ FPR=10-3↑TPR @ FPR=10-4↑
Cauc.Afr.AsianIndianCauc.Afr.AsianIndian
ArcFace R-50 + Dprior89.8686.7394.3193.8281.4876.8087.4786.89
93.5491.5196.5196.1786.8483.7491.7491.19
+ 1.3M interps91.8789.0796.0895.5584.6579.9791.1390.30
+ 6M interps89.4886.8594.9294.2280.9577.0088.7587.94
", + "image_path": "345384b233806c0dde4b6e94ce95474622764c239272bf210fd5705ca907f90c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 136, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 136, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 136, + 545, + 220 + ], + "type": "text", + "content": "Table 7: In this experiment, instead of pretraining our model as an encoder for StyleGAN, we use the StyleGAN and the trained encoder to draw samples from the learned face distribution and add them to our dataset. The results indicate that interpolation-based procedure, despite being more straightforward than the proposed pretraining-based scheme, is actually less efficient." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 236, + 545, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 333 + ], + "type": "text", + "content": "StyleGAN and encoder. Furthermore, we additionally release two training datasets for unsupervised pretraining and one large-scale protocol for bias estimation. We show that we are able to tune the performance on different ethnicities by altering the composition of the unlabeled prior datasets. We hope that the released datasets, together with the protocol, will drive forward the research on mitigating the face recognition biases using self-supervised approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 336, + 546, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 546, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 546, + 528 + ], + "type": "text", + "content": "There are several directions for possible future work that we foresee. First, integrating the two steps of pretraining and fine-tuning into one would help to avoid forgetting of the pretrained weights. Second, publicly available datasets often don't possess enough variability (both in terms of capture conditions and demographics) [25] and typically feature only low-resolution images which might be a blocking factor for the fairness research [49, 30]. Accordingly, the results on the newly assembled RB-WebFace benchmark can be analyzed more thoroughly, which can bring new insights about the factors positively affecting bias mitigation in real-world scenarios. Finally, various architectures typically used for self-supervised learning (such as transformers or highly scalable generative models [43]), might allow for the construction of more flexible methods due to heterogeneous inputs and outputs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 530, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 626 + ], + "type": "text", + "content": "Legal concerns. The collected unlabeled data was collected anonymously in accordance with Standard YouTube License and CC YouTube License and does not contain personally identifiable information (PII). The data is released only in the form of links to YouTube videos and the corresponding timestamps, fully following good practice of similar data collection [10, 17, 32] and with the data subjects protection as per Art. 14 5(b) GDPR law." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments. We gratefully acknowledge the support of this research by a TUM-IAS Hans Fischer Senior Fellowship, the ERC Starting Grant Scan2CAD (804724) and the Horizon Europe vera.ai project (101070093). We also thank Yawar Siddiqui for the helpful technical advice on StyleGAN part, Angela Dai for the video voiceover, and Dmitriy Karfagenskiy for the proofreading and corrections." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20931" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] How to boost face recognition with stylegan?. Supplemental material, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 157 + ], + "type": "text", + "content": "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6711-6720, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 286, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 286, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 286, + 212 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 286, + 246 + ], + "type": "text", + "content": "[4] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161-163, 1992." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 286, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 291 + ], + "type": "text", + "content": "[5] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. Advances in Neural Information Processing Systems, 6, 1993." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 292, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 335 + ], + "type": "text", + "content": "[6] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, pages 77-91, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "type": "text", + "content": "[7] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. VGGface2: A dataset for recognising faces across pose and age. In IEEE International Conference on Automatic Face and Gesture Recognition, pages 3414-3424, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 435 + ], + "type": "text", + "content": "[8] Yunliang Chen and Jungseock Joo. Understanding and mitigating annotation bias in facial expression recognition. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 14980-14991, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 286, + 491 + ], + "type": "text", + "content": "[9] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 539-546, 2005." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 514 + ], + "type": "text", + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In Interspeech, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "text", + "content": "[11] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 624 + ], + "type": "text", + "content": "[12] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In Proc. of the 58th Annual Meeting of the Association for Computational Linguistics, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 286, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 669 + ], + "type": "text", + "content": "[13] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] Emily Denton, Ben Hutchinson, Margaret Mitchell, Timnit Gebru, and Andrew Zaldivar. Image counterfactual sensitivity analysis for detecting unintended bias. arXiv preprint arXiv:1906.06439, 2019." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 182 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 183, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 239 + ], + "type": "text", + "content": "[17] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[18] Markos Georgopoulos, James Oldfield, Mihalis A Nicolaou, Yannis Panagakis, and Maja Pantic. Mitigating demographic bias in facial datasets with style-based multi-attribute transfer. International Journal of Computer Vision, 129:2288-2307, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "type": "text", + "content": "[19] Sixue Gong, Xiaoming Liu, and Anil K Jain. Debface: Debiasing face recognition. arXiv preprint arXiv:1911.08080, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "text", + "content": "[20] Sixue Gong, Xiaoming Liu, and Anil K Jain. Mitigating Face Recognition Bias via Group Adaptive Classifier. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3414-3424, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 392 + ], + "type": "text", + "content": "[21] Google. Real Tone on Pixel 6. Be seen as you truly are., 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 393, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 425 + ], + "type": "text", + "content": "[22] Patrick Grother, Mei Ngan, and Kayee Hanaoka. Face recognition vendor test (FVRT): Part 3, demographic effects. 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 426, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 471 + ], + "type": "text", + "content": "[23] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-Celeb-1M: A dataset and benchmark for large-scale face recognition. In European Conference on Computer Vision (ECCV), 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 525 + ], + "type": "text", + "content": "[24] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, pages 1735-1742, 2006." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 526, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 581 + ], + "type": "text", + "content": "[25] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in AI: the Casual Conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 582, + 545, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 625 + ], + "type": "text", + "content": "[26] Mingjie He, Jie Zhang, Shiguang Shan, and Xilin Chen. Enhancing face recognition with self-supervised 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4062-4071, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 680 + ], + "type": "text", + "content": "[27] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Di Kang, Debarun Dhar, and Antoni Chan. Incorporating side information by adaptive convolution. Advances in Neural Information Processing Systems, 30, 2017." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20932" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[29] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "text", + "content": "[30] Martin Knoche, Stefan Hörmann, and Gerhard Rigoll. Image resolution susceptibility of face recognition models. arXiv preprint arXiv:2107.03769, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[31] Wenyu Li, Tianchu Guo, Pengyu Li, Binghui Chen, Biao Wang, Wangmeng Zuo, and Lei Zhang. Virface: Enhancing face recognition via unlabeled shallow data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14729-14738, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 252 + ], + "type": "text", + "content": "[32] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3207-3216, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "text", + "content": "[33] Chun-Hsien Lin and Bing-Fei Wu. Domain adapting ability of self-supervised learning for face recognition. In IEEE International Conference on Image Processing (ICIP), pages 479–483, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[34] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 212-220, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 398 + ], + "type": "text", + "content": "[35] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv:1907.11692, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 454 + ], + "type": "text", + "content": "[36] Vongani Maluleke, Neerja Thakkar, Tim Brooks, Ethan Weber, Trevor Darrell, Alexei A Efros, Angjoo Kanazawa, and Devin Guillery. Studying Bias in GANs through the Lens of Race. In European Conference on Computer Vision (ECCV), pages 344-360, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "text", + "content": "[37] Richard T Marriott, Sami Romdhani, and Liming Chen. A 3d gan for improved large-posed facial recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13445–13455, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[38] Tuan-Duy H Nguyen, Huu-Nghia H Nguyen, and Hieu Dao. Recognizing families through images with pretrained encoder. In IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 887-891, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "type": "text", + "content": "[39] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10880-10890, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "text", + "content": "[40] Christian Rathgeb, Pawel Drozdowski, Naser Damer, Dinusha C Frings, and Christoph Busch. Demographic fairness in biometric systems: What do the experts say? IEEE Technology and Society Magazine, 41(4):71-82, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2287–2296, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[42] Joseph P Robinson, Gennady Livitz, Yann Henon, Can Qin, Yun Fu, and Samson Timoner. Face Recognition: Too Bias," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "or Not Too Bias? In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[43] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH Conference Proceedings, pages 1-10, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[44] Yichun Shi and Anil K Jain. Probabilistic face embeddings. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6902-6911, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "text", + "content": "[46] Fu-En Wang, Chien-Yi Wang, Min Sun, and Shang-Hong Lai. Mixfairface: Towards ultimate fairness via mixfair adapter in face recognition. In AAAI Conference on Artificial Intelligence, volume 37, pages 14531-14538, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[47] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5265-5274, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "text", + "content": "[48] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In ACM International Conference on Multimedia, pages 3779-3782, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "text", + "content": "[49] Mei Wang and Weihong Deng. Mitigating Bias in Face Recognition using Skewness-Aware Reinforcement Learning. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9322-9331, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 454 + ], + "type": "text", + "content": "[50] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, and Yaohai Huang. Racial faces in the wild: Reducing racial bias by information maximization adaptation network. In International Conference on Computer Vision (ICCV), pages 692-702, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[51] Qingzhong Wang, Pengfei Zhang, Haoyi Xiong, and Jian Zhao. Face evolves: A high-performance face recognition library. arXiv preprint arXiv:2107.08621, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 534 + ], + "type": "text", + "content": "[52] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 535, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 590 + ], + "type": "text", + "content": "[53] Xingkun Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, Jilin Li, Feiyue Huang, Yong Li, and Zhen Cui. Consistent Instance False Positive Improves Fairness in Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 578-586, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 612 + ], + "type": "text", + "content": "[54] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv:1411.7923, 2014." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "text", + "content": "[55] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[56] Haoyu Zhang, Marcel Grimmer, Raghavendra Ramachandra, Kiran Raja, and Christoph Busch. On the applicability of synthetic data for face recognition. In IEEE International Workshop on Biometrics and Forensics (IWBF), pages 1-6, 2021." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20933" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 273 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10):1499-1503, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "text", + "content": "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "type": "text", + "content": "[59] Song Zhang, Gehui Shen, Jinsong Huang, and Zhi-Hong Deng. Self-supervised learning aided class-incremental life-long learning. arXiv preprint arXiv:2006.05882, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 273 + ], + "type": "text", + "content": "[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20934" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_content_list.json b/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5f05cf556379e445191bb63b4e5456026606e6f3 --- /dev/null +++ b/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_content_list.json @@ -0,0 +1,1735 @@ +[ + { + "type": "text", + "text": "How to choose your best allies for a transferable attack?", + "text_level": 1, + "bbox": [ + 200, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thibault Maho* \nUniv. Rennes, Inria, CNRS IRISA, Rennes, France \nthibault.maho@inria.fr", + "bbox": [ + 96, + 180, + 316, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seyed-Mohsen Moosavi-Dezfooli \nImperial College London, UK \nseyed.moosavi@imperial.ac.uk", + "bbox": [ + 349, + 180, + 614, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Teddy Furon† \nUniv. Rennes, Inria, CNRS IRISA, Rennes, France \nteddy.furon@inria.fr", + "bbox": [ + 655, + 180, + 872, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The transferability of adversarial examples is a key issue in the security of deep neural networks. The possibility of an adversarial example crafted for a source model fooling another targeted model makes the threat of adversarial attacks more realistic. Measuring transferability is a crucial problem, but the Attack Success Rate alone does not provide a sound evaluation. This paper proposes a new methodology for evaluating transferability by putting distortion in a central position. This new tool shows that transferable attacks may perform far worse than a black box attack if the attacker randomly picks the source model. To address this issue, we propose a new selection mechanism, called $F \\vdash T$ , which aims at choosing the best source model with only a few preliminary queries to the target. Our experimental results show that $F \\vdash T$ is highly effective at selecting the best source model for multiple scenarios such as single-model attacks, ensemble-model attacks and multiple attacks.", + "bbox": [ + 75, + 318, + 473, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 607, + 209, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transferability is one of the most intriguing properties of adversarial examples. A white box attack crafting adversarial examples for an open-source model is likely to fool other models too [1, 14, 20, 25, 34]. This makes the threat of adversarial examples more realistic. In practice, the model targeted is usually unknown but accessible as a black box. This prevents directly applying any white box gradient-based attack [10, 16, 19, 35]. Black box attacks do exist but they require some thousands of queries to find an adversarial example of low distortion [4, 11, 17, 24]. Transferable attacks require no or few queries to fine-tune an adversarial example thanks to the help of a publicly available model similar enough to the target.", + "bbox": [ + 75, + 633, + 468, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transferability is usually measured by the Attack Success Rate (ASR), i.e., the probability that the adversarial", + "bbox": [ + 76, + 830, + 468, + 862 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c65d3a539418e84b1bd55dfff8da234f0f3adbfe25969e02c5ea9121e13af76.jpg", + "image_caption": [ + "Figure 1: Evaluation of transferability by comparing the Attack Success Rate vs. distortion trade-off of a white box, transferable, and black box attacks against model CoatLitesmall (See Sect. 4.1 for details). The blue area is the range of trade-off operated by a transferable attack with random source models. A transferable attack may be worse than a black box attack without a good source selection (like FiT)." + ], + "image_footnote": [], + "bbox": [ + 522, + 287, + 867, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "example crafted for the source model also deludes the target model. We argue that this measure leads to an unfair evaluation of transferability. In the context of adversarial examples, it is not just a matter of discovering data that is not well classified, but rather identifying the perturbation that can fool a classifier with minimal distortion. This principle should also apply to transferable attacks.", + "bbox": [ + 496, + 638, + 893, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For illustration purposes, let us consider two models, one is robust in the sense that the necessary amount of adversarial perturbation is large, whereas the other model is weak. If the attacker uses the robust model as the source to attack the weak target network, the ASR of the transferable attack will certainly be big. It does not mean that this is the right choice. The ASR is high because the robust source model needs large perturbation to be deluded, which will fool any weaker model. The ASR alone does not reflect the overshooting in distortion. The converse, using the weak to at", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Thanks to Rennes Métropole for its funding for international mobility.", + "bbox": [ + 93, + 875, + 467, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Thanks to ANR and AID french agencies for funding Chaire SAIDA.", + "bbox": [ + 96, + 887, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4542", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tack the robust, would yield a low ASR. To summarize, the ASR alone fails to capture how relevant the direction of the perturbation given by the source is for attacking the target.", + "bbox": [ + 75, + 90, + 467, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first contribution of this paper is to put distortion back into the picture. Section 3 evaluates transferability by comparing the distortion of a transferable attack to the ones of two reference attacks: On one hand, the strongest attack, i.e., the white box attack directly applied on the target model; on the other hand, the weakest attack, i.e., the black box attack.", + "bbox": [ + 75, + 137, + 467, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second contribution shows the great variability of the performance of transferable attacks. Figure 1 summarizes this observation by plotting the ASR as a function of the distortion (the experimental protocol is explained in Sect. 4.1). Naturally, the black box attack needs much more distortion than the white box attack. For instance, the white box attack yields an ASR of $50\\%$ with a distortion of 0.19, whereas the black box attack needs a distortion of 9.7. The surprise is that if the attacker resorts to a transferable attack and picks a source model at random, there is almost a $50\\%$ chance that the attack performs even worse than the black box attack. Section 4 outlines a triad of factors: input, source model, and attack.", + "bbox": [ + 75, + 243, + 467, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This observation challenges the prevalent notion that adversarial examples transfer easily between models, and highlights the need to carefully choose the source model to attack a target. Under the assumption that the attacker has indeed several candidate models, our third contribution, named FiT in Sect. 5, provides an affordable measure for model selection, allowing the attacker to choose a good source model with only a few queries to the target.", + "bbox": [ + 75, + 441, + 467, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 577, + 212, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. White box, black box, and transferable attacks", + "text_level": 1, + "bbox": [ + 76, + 603, + 467, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The threat analysis related to adversarial examples usually considers two scenarios: the white box assumption where the attacker knows the internals of the target model, and the black box assumption where he does not but has limited access to the target.", + "bbox": [ + 75, + 627, + 467, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "White box attacks are now performing very well in terms of Attack Success Rate and distortion. We distinguish two kinds of attacks: 1) attacks constrained by a distortion budget like PGD [16] whose performance is measured by the ASR, and 2) attacks yielding almost surely an adversarial example like DeepFool [19] and CW [3] whose performance is measured by the distortion. The trend is to make them speed efficient as well with fewer computations of the gradient of the neural network, like FMN [23] and BP [35].", + "bbox": [ + 75, + 703, + 467, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our paper also considers so-called decision-based black box attacks like GeoDA [24], SurFree [17], and RayS [4], where the attacker only sees the class predicted by the model, but not the confidence score, neither the pro", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bit nor the logit. A decision-based black box attack typically needs thousands of queries to find an adversarial example with a relatively low distortion, yet, much higher than with a white box attack. This is the price the attacker has to pay when he does not know the target model.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The transferable attacks, which pertain to the black box scenario, disrupt this last statement. The assumption is that attacker knows another model, called source, trained for the same classification problem as the target. Yet, leading a white box attack on the source model with the hope that this adversarial example will also fool the target yields poor results because the adversarial perturbation is too specific to the source. Modern transferable attacks increase the ASR by avoiding the overfitting of examples on the source. Input transformations have been purported by DI [32], TI [6] and Admix [27]. Papers [8, 13, 28] stabilized the gradient while [7, 29, 36, 39] focus on the importance of intermediate features. The combination of several methods has also been investigated [9, 40]. For instance, TAIG [9] uses input transformation, integrated gradients, and attention reduction.", + "bbox": [ + 496, + 167, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ensemble-model attacks make the assumption that the attacker leverages not a single source but several source models. Paper [28] averages the logits of several sources and leads a white box attack on this aggregation of models, whereas [33] reduces model variance. The main obstacle to these methods is the computational complexity. To counteract this drawback, some methods design a specific ensemble of sources like ghost networks [12] or pruned network [26]. This amounts to creating several sources without the need to compute the gradient of each model.", + "bbox": [ + 496, + 395, + 890, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All this literature uses the ASR as the figure of merit for a given distortion while our paper draws the full operating characteristic of ASR vs. distortion. Moreover, the performance of the transferability is compared neither to the white box nor the decision-based black box attack.", + "bbox": [ + 496, + 547, + 890, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Fingerprinting", + "text_level": 1, + "bbox": [ + 500, + 633, + 650, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Models are valuable assets because of their training costs, be it the expertise, the annotated data, or the computational power. Fingerprinting methods have been proposed to protect this intellectual property. They provide a similarity score between the two models by looking at similar decisions in submitted queries. They often use adversarial examples specifically designed to be unique for a model. The models are considered similar if they share the same decision for these inputs. IPGuard [2] creates adversarial examples close to the decision boundary to frame it. Universal adversarial perturbations are used to characterize the boundary in [21]. The paper [15] built adversarial examples with a transferable property called conferrable adversarial examples. AFA [37] crafts adversarial examples model-specific using dropout. FBI [18] is the only method using unmodified images. It estimates the mutual information to measure", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4543", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the statistical dependence of two models' predictions.", + "bbox": [ + 76, + 90, + 434, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our paper investigates the idea that a model similar to the target may be a good source for transferable attacks.", + "bbox": [ + 76, + 106, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 152, + 212, + 169 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper introduces a new way to gauge transferability. The main motivation is that transferability should indicate whether a source $s$ is useful for attacking a target $t$ , independently from the inherent robustness of $t$ .", + "bbox": [ + 76, + 178, + 468, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notations", + "text_level": 1, + "bbox": [ + 76, + 250, + 187, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $f:[0,255]^N \\to \\mathcal{R}^K$ denote a classifier computing the predicted probabilities $f(x)$ of $K$ classes for input $x$ . For a given input $x$ of class $y$ , the adversarial example $x_a$ is the result of an attack (be it white box, black box, or by transferability) such that", + "bbox": [ + 76, + 272, + 468, + 349 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\max _ {1 \\leq k \\leq K} f _ {k} (x _ {a}) \\neq \\arg \\max _ {1 \\leq k \\leq K} f _ {k} (x) = y. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 362, + 468, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the case of transferable attacks, the attacker uses a source model $f_{s}$ to build a transferable adversarial example against the target model $f_{t}$ . This paper considers a collection of $m$ source models denoted by $\\mathcal{F}_s = \\{f_s^1,\\dots,f_s^m\\}$ , and a set of $n$ inputs $\\mathcal{X}$ .", + "bbox": [ + 76, + 398, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Measurement", + "text_level": 1, + "bbox": [ + 76, + 484, + 218, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distortion. We measure the distortion of an adversarial perturbation using the Euclidean norm which is a common choice as we deal with natural images:", + "bbox": [ + 76, + 508, + 468, + 554 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s t} \\left(x _ {a}, x\\right) := \\| x _ {a} - x \\| _ {2} / \\sqrt {N}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 565, + 468, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $N = 3\\times H\\times L$ pixels. This distortion can be seen as the average amplitude of the perturbation per pixel.", + "bbox": [ + 76, + 595, + 468, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Operating characteristic. Given an attack (be it white box, black box, or by transferability) and a target model $f_{t}$ , we call operating characteristic the following function:", + "bbox": [ + 76, + 627, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP (D) := \\mathbb {P} (\\operatorname {d i s t} \\left(x _ {a}, x\\right) < D). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 685, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is thus the Attack Success Rate as a function of the distortion. Contrary to the literature which measures the ASR for a few distortion levels, we consider the full range of $D$ values.", + "bbox": [ + 76, + 714, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Transferability", + "text_level": 1, + "bbox": [ + 76, + 785, + 228, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a white box and a black box attacks, The proposed methodology first computes the operating characteristic $P_{t}^{\\mathrm{wb}}(D)$ of a state-of-the-art white box attack directly applied to the target model, and the operating characteristic $P_{t}^{\\mathrm{bb}}(D)$ of a state-of-the-art black box attack. It then measures where the operating characteristic of the transferable", + "bbox": [ + 76, + 809, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attack from the source $s$ to the target $t$ lies in between the two characteristics as follows:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {s, t} := \\frac {\\int_ {0} ^ {\\infty} P _ {s , t} (u) - P _ {t} ^ {\\mathrm {b b}} (u) d u}{\\int_ {0} ^ {\\infty} P _ {t} ^ {\\mathrm {w b}} (u) - P _ {t} ^ {\\mathrm {b b}} (u) d u}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 131, + 890, + 169 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed score is calculated as the ratio of the areas between the different operational characteristics, which are defined as Cumulative Distribution Functions (CDFs). The numerator is therefore close to the 1-Wasserstein distance [5] between the ASR function of the distortions of the transferable attack and that of the black box attack, where the absolute value is removed to obtain a signed score. If the transferable attack performs as well as the white box attack (resp. as bad as the black box attack), then $T_{s,t} = 1$ (resp. $T_{s,t} = 0$ ). Figure 1 shows that the numerator of (4) can indeed be negative as transferability can be even worse than the black box if the source $s$ is not well chosen. Therefore, zero is not a lower bound for $T_{s,t}$ . In the denominator, the 1-Wasserstein distance is obtained exactly between the white box and the black box because the white box consistently generates adversarial examples with lower distortion.", + "bbox": [ + 496, + 180, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Practical implementation", + "text_level": 1, + "bbox": [ + 500, + 434, + 730, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attacks. We only consider state-of-the-art attacks which almost surely deliver an adversarial example. As said in Sect. 2.1, this is usually the case for black box attacks, but we are limited to using white box attacks that are not subject to a distortion budget. As for transferability, the attack uses the publicly available model $s$ and input $x$ to craft an adversarial direction $u_{x,s}$ of Euclidean norm $\\sqrt{N}$ . The use of $\\sqrt{N}$ for the norm is for the sake of simplicity in notation. We assume that there is an oracle giving the minimal distortion along this direction to fool the target $t$ . In other words, $x_{a} = x + d u_{x,s}$ , with", + "bbox": [ + 496, + 458, + 890, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd = \\min \\left\\{\\delta : \\arg \\max _ {k} f _ {t, k} (x + \\delta u _ {x, s}) \\neq y \\right\\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 638, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that $\\mathrm{dist}(x_a,x) = d$ . This definition favours transferability as its best. In practice, such an oracle does not exist but the attacker finds a good estimate of (5) thanks to a line search within a few queries to the target.", + "bbox": [ + 496, + 674, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transferability. We run a given attack over a collection $\\mathcal{X}$ of $n$ inputs correctly classified by the target model. We compute the distortions $d(j) = \\mathrm{dist}(x_{a,j}, x_j)$ with $x_j \\in \\mathcal{X}$ and sort them so that $d(1) \\leq d(2) \\leq \\ldots \\leq d(n)$ . We set $d(0) = 0$ and $d(n + 1) = \\infty$ to properly define the empirical operating characteristic by the following step function:", + "bbox": [ + 496, + 734, + 890, + 827 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} (D) := j / n \\quad \\forall D \\in [ d (j), d (j + 1)). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 839, + 890, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is then easier to estimate the integrals appearing in (4), which are indeed areas under two curves, by a Lebesgue", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4544", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sum rather than a Riemann sum. This gives:", + "bbox": [ + 76, + 90, + 372, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {T} _ {s, t} := \\frac {\\sum_ {j = 1} ^ {n} d _ {s , t} (j) - d _ {t} ^ {\\mathrm {b b}} (j)}{\\sum_ {j = 1} ^ {n} d _ {t} ^ {\\mathrm {w b}} (j) - d _ {t} ^ {\\mathrm {b b}} (j)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 113, + 468, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the distortions $(d_t^{\\mathrm{bb}}(j))_j$ (resp. $(d_t^{\\mathrm{wb}}(j))_j$ ) resulting from the black box (resp. white box) attack against model $t$ are also sorted in increasing order.", + "bbox": [ + 76, + 160, + 468, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Triad of transferability: data, model, attack", + "text_level": 1, + "bbox": [ + 76, + 218, + 467, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section is an experimental investigation of the factors affecting transferability.", + "bbox": [ + 76, + 244, + 468, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 282, + 264, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The study assesses transferable attacks on a total of 48 models, with 47 of them sourced from the Timm library [30] and one very robust, namely ResNet50AdvTrain, obtained from the GitHub repository 1. Appendix A lists all the models. The experiments utilize 100 images from the validation set of ILSVRC'12, all of which are correctly classified by all models under consideration.", + "bbox": [ + 75, + 305, + 468, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The study considers three transferable attacks - DI [32], TAIG [9], and DWP [26]- each using a different approach to improve transferability (see Sect. 2.1). These attacks are selected as the best in their categories in [38]. They all share an $\\epsilon$ parameter to control the maximum perturbation added per pixel. The effect of the parameter $\\epsilon$ indeed happens to be negligible in our protocol, as demonstrated in Appendix B. We choose $\\epsilon = 8$ .", + "bbox": [ + 75, + 411, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To measure transferability (7), we need state-of-the-art black box and white box attacks. Certain methods may exhibit a preference towards one model over the other, necessitating the use of multiple attacks. Our study employs four white box attacks (BP [35], DeepFool [19], I-FGSM [10], and PGD [16]) and three black box attacks (SurFree [17], RayS [4], and GeoDA [24]). All black box attacks are run with 2,000 queries, which has been determined to be sufficient for achieving convergence. For distortion-constrained white box attack, the $\\epsilon$ parameter is set to 4. We record the smallest perturbation distortion over the black box (white box) for each image and draw the operating characteristic (6) as appearing in green (resp. red) dashed line in Fig. 1.", + "bbox": [ + 75, + 532, + 468, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for the transferable attack, for a given target model, the attacker has access to a subset of all the other models whose architecture differs from the target. This amounts to an average of 45 models out of 48. For instance, in Fig. 1, CoatLitesmall being the target, we exclude all other CoatLite models from being a source. The light blue area delimits the operating characteristics of transferable attack DI [32] using as the source one of the 45 remaining models.", + "bbox": [ + 75, + 743, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Model dependence", + "text_level": 1, + "bbox": [ + 500, + 90, + 679, + 106 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/79c1ed355cd42f1ac9c7073c690859e106db534d9f137ba089de4233d6566483.jpg", + "image_caption": [ + "Figure 2: Transferability score $\\hat{T}_{s,t}$ matrix of 48 sources and 48 targets listed in App. A with attack DI [32]." + ], + "image_footnote": [], + "bbox": [ + 501, + 127, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large transferability variation. Figure 2 shows the matrix $\\hat{T}_{s,t}$ , where $s$ and $t$ are any of the $48^2$ pairs, for attack D I [32]. Not all models possess the same transferability capabilities. At first look, the figure contains more blue than red cells which means that transferability takes negative value more often.", + "bbox": [ + 496, + 465, + 890, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some rare models, like PiTsmall-dist, exhibit good transferability towards any target. On the contrary, DPN68b is always a bad source. On the other hand, ResNet50AdvTrain is a very difficult target (note how", + "bbox": [ + 498, + 556, + 890, + 617 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5300a085f9c1a7cac86c7950c1f852ab2c7d4e0cb7bd6d63f57a2a43d5fa5ce6.jpg", + "image_caption": [ + "Figure 3: $\\hat{T}_{s,t}$ function of the number of available sources and attack DI [32]. The best or worst model selected per image (solid line) or on average (dashed line)." + ], + "image_footnote": [], + "bbox": [ + 501, + 659, + 854, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://github.com/MadryLab/robustness", + "bbox": [ + 94, + 886, + 401, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4545", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/22747ae36c47dbdbf308284fcae58d936c22d3ad31956ac7f846a275bdc86dda.jpg", + "image_caption": [ + "Figure 4: Distribution of best image-model selection function of the number of sources. Adversarial examples obtained with DI [32]." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 444, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ever that its accuracy is low), followed by PiTlight and ConViTtiny, whereas the family of ResNet are fairly easy target.", + "bbox": [ + 75, + 357, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior works have shown that models with architectures similar to the target transfer better [22, 31]. The red squares appearing in the matrix confirm this (models are ordered according to architecture family). For instance, MobileNetV2 and RexNet architectures exhibit transferability close to 0.7. However, this is not an absolute truth. For example, EfficientNet-B0 has better transferability against MixNetlarge than MixNetmedium, even though they have similar architectures. ConViT transfers exceptionally well to Twins although their architectures are very different. Similar observations hold for the other attacks but with lower transferability scores (see Appendix C).", + "bbox": [ + 75, + 412, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From now on, we exclude models whose architecture is similar to the target.", + "bbox": [ + 76, + 604, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Impact on the attack. Selecting the right source among several available models is critical for achieving high transferability. Figure 3 displays the transferability score (7) for the best and the worst choices (shown in dotted lines) as the number of source candidates increases. In this evaluation, a single model is selected as the source for building all adversarial examples. On average, the transferability is lower than zero, regardless of the attack method. This means that transferable attacks perform worst than black box attacks on average. If the attacker knows how to select the best source model, $\\hat{T}_{s,t}$ quickly converges to a maximum which is positive but below 0.5. This means that transferability at its best performs closer to a black box attack rather than a white box attack. These remarks highly mitigate the threat of transferability and put emphasis on the crucial selection of the best source model. As far as we know, these facts are not reported in the adversarial examples literature.", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c36484bfba84df3025f192834fb37ff19e02a31036c7cac698e3f45655215c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 90, + 692, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5876eed1cf796cf9d0972e5f086a19cad7abdd427e716d816a800fa985dcf0f6.jpg", + "image_caption": [ + "Figure 5: 2D Histogram of the minimum distortion for transferable and white box attacks." + ], + "image_footnote": [], + "bbox": [ + 516, + 220, + 692, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e11cac767e0f0bae8379a1e7726d996ea3ee6678f6ad18e77b83ad17ab42d642.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 90, + 870, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/01e31915b18a074af56f26cba78f58fdc470bcdd15c6dc7ac4887ab8f20b5a49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 220, + 870, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Image dependence", + "text_level": 1, + "bbox": [ + 500, + 417, + 679, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The difficulty of transferring adversarial examples from a source to a target varies significantly depending on the input $x$ . This is illustrated in Fig. 4, which shows the distribution of the best-performing model for each image based on the available sources. Even with a large number of sources available, there is typically one source that performs significantly better than the others, but only over $60\\%$ of the images on average. However, this superiority decreases rapidly as the number of available sources increases. It means that a better source exists for $40\\%$ of the images on average.", + "bbox": [ + 496, + 441, + 890, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supposing that the attacker knows the best source of each input, Figure 3 shows in solid line that the transferability converges to a value close to 0.5. The performance of the transferable attack lies in between the ones of the white box and black box attacks. This highlights the importance of selecting an appropriate source model for a given target and input.", + "bbox": [ + 496, + 608, + 890, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Attack dependence", + "text_level": 1, + "bbox": [ + 500, + 726, + 681, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transferable attack. Figure 5 compares transferable attacks with a 2D histogram of the distortion pair for two attacks. This is computed over all inputs in $\\mathcal{X}$ and all pairs of source and target models. The results show that DWP [26] exhibits poor transferability compared to DI [32] and TAIG [9], which produce adversarial perturbations with similar distortion. Additionally, regardless of the attack complexity or method, the challenging images remain consistent. If one attack requires a high distortion for a given input / source / target, other attacks are likely to encounter", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4546", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "similar difficulties.", + "bbox": [ + 76, + 90, + 204, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Traditional white box attack. We now compare the methods designed for transferability with a naive approach. A white box attack (ie. not specific to transferability) is executed on the source, and then the direction found serves as $u_{x,s}$ to place the adversarial examples on the target boundary with (5). The best $\\hat{T}_{s,t}$ score is 0.27 for BP [35] compared to 0.52 for DI [32]. This confirms the superiority of the recent methods designed for transferability. However, even though DI [32] is on average better, Figure 5 shows that when the necessary distortion is small, traditional white box attacks like DeepFool [19] and BP [35] indeed beats DI [32]. On the other hand, DI [32] performs much better for inputs requiring more distortion.", + "bbox": [ + 76, + 106, + 470, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. How to choose the best source", + "text_level": 1, + "bbox": [ + 76, + 316, + 352, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Section 4 outlines that the choice of the source is of utmost importance thanks to the transferability measure defined in Sect. 3. This section investigates whether the attacker can guess which model is the best source.", + "bbox": [ + 76, + 340, + 468, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now propose a procedure which combines the dependences with respect to the source and target models (Sect. 4.2), and the input (Sect. 4.3). Model dependence is measured by evaluating the similarity between the source and target models, which is denoted as ModSim. On the other hand, the image dependence is measured by the quality of the adversarial example, denoted as TransQ. Both metrics are combined into the following score:", + "bbox": [ + 76, + 402, + 468, + 523 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F i T} (s, t, x) := \\operatorname {M o d S i m} (s, t) \\times \\operatorname {T r a n s Q} (s, x). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 532, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These indicators should be easy to compute. We especially pay attention to the number of queries to the target. This score opens the door to a new strategy for the attacker which first selects the best source among the available models", + "bbox": [ + 76, + 556, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {\\star} (t, x) = \\underset {\\sigma \\in \\mathcal {F} _ {s}} {\\arg \\max } \\mathrm {F i T} (\\sigma , t, x), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 627, + 468, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and then crafts the adversarial direction $u_{x,s^*(t,x)}$ .", + "bbox": [ + 76, + 657, + 406, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Criterion ModSim(s,t)", + "text_level": 1, + "bbox": [ + 76, + 681, + 282, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Section 4.2 highlights the correlation between the transferability and the similarity between the source and target models. Gauging model similarity has been previously studied in the context of fingerprinting as a defense to protect intellectual property (see Sect. 2.2). This paper uses fingerprinting methods as an attack that leaks information about the target.", + "bbox": [ + 76, + 704, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider the fingerprinting method [18] because it works in a decision-based setup since the target is a black box in our application. Querying two models $s$ and $t$ with few natural images, it computes a distance $\\operatorname{Dist}(s, t) \\in [0, 1]$ by comparing their outputs. Since we look for a similarity, we set $\\operatorname{ModSim}(s, t) = 1 - \\operatorname{Dist}(s, t)$ . However,", + "bbox": [ + 76, + 810, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "this provides a symmetrical similarity, ie. $\\operatorname{ModSim}(s,t) = \\operatorname{ModSim}(t,s)$ , while transferability is not (see Fig. 2). This shows that this criterion alone is not sufficient.", + "bbox": [ + 498, + 90, + 892, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Criterion TransQ(s, x)", + "text_level": 1, + "bbox": [ + 500, + 145, + 707, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This criterion evaluates the general transferability of a given adversarial example crafted by a source. Our idea is to leverage the assumption that the attacker has a set of models $\\mathcal{F}_s$ . Consequently, we can evaluate the transferability thanks to the other models of this set, without querying the target.", + "bbox": [ + 498, + 167, + 893, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a given input $x$ , the source $s$ provides the adversarial direction $u_{x,s}$ and we compute the distortion $d_{s,\\sigma}$ necessary to delude classifier $\\sigma \\in \\mathcal{F}_s$ with (5). We then aggregate these distortions into a single score with two flavours:", + "bbox": [ + 498, + 258, + 892, + 319 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T r a n s Q} ^ {(1)} (s, x) := \\left(\\frac {1}{| \\mathcal {F} _ {s} |} \\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {s, \\sigma}\\right) ^ {- 1}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 328, + 892, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A good source for input $x$ gives birth to lower distortions, so that $\\mathrm{TransQ}^{(1)}(s,x)$ is large.", + "bbox": [ + 498, + 382, + 890, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T r a n s} Q ^ {(2)} (s, x) := \\frac {\\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {s , \\sigma} - d _ {\\sigma} ^ {\\mathrm {b b}}}{\\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {\\sigma} ^ {\\mathrm {w b}} - d _ {\\sigma} ^ {\\mathrm {b b}}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 422, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This measure is similar to (7) except that it is computed over the set of models instead of a set of inputs.", + "bbox": [ + 498, + 464, + 892, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 500, + 503, + 591, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental setup is the same as in Sect. 4.1. We select the fingerprinting method FBI [18] with 200 benign natural images to compute the criterion ModSim(s, t). It implies that the attacker first makes 200 queries to the target in a preliminary step before forging any adversarial example. Appendix D.1 shows that more images improve the fingerprinting accuracy hence the model selection, but the results converge after a number of 200 images.", + "bbox": [ + 498, + 526, + 890, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section is structured as follows: we use FiT to select the best model for single-model attacks, then we employ it to identify the best subset of models for ensemble-model attacks and different combinations of attacks.", + "bbox": [ + 498, + 648, + 890, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3.1 Single-model attacks", + "text_level": 1, + "bbox": [ + 500, + 724, + 697, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Criterion ModSim(s, t). Table 1 indicates that architectural similarity is a reliable measure of transferability between two models. It can drive the selection of a good source giving birth to a transferable attack outperforming the black box attack since $\\hat{T}_{s,t}$ is larger than 0 (except for DWP [26]). Yet, the results remain low compared to the best results obtained. As discussed in Sec 4.2, similarity may not suffice because it implies the selection of one unique for a given target. Better results are achieved when adapting the source to the input.", + "bbox": [ + 498, + 750, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4547", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb42459f66547bf0379a5583e7bfdcbe03deb4b35007d68d09448120cc918a8f.jpg", + "table_caption": [ + "Table 1: Transferability ${\\widehat{T}}_{s,t}$ for DI [32],TAIG [9] and DWP [26] for single and ensemble-model attacks." + ], + "table_footnote": [], + "table_body": "
CategorySelection MethodDI [32]TAIG [9]DWP [26]
Single-model attackBest0.52±0.120.46±0.120.34±0.13
Random-0.16±0.26-0.12±0.21-0.72±0.32
ModSimFBI [18]0.18±0.170.12±0.18-0.39±0.19
TransQASR-0.21±0.39-0.24±0.30-1.14±0.95
TransQ(1)(10)0.38±0.150.24±0.160.10±0.20
TransQ(2)(11)0.37±0.180.23±0.180.08±0.22
FiTTransQ(1)(10)0.40±0.130.27±0.170.12±0.20
TransQ(2)(11)0.39±0.160.25±0.180.10±0.22
Ensemble-model attack with three sourcesBest0.72±0.080.62±0.080.46±0.1
Random0.43±0.120.40±0.110.02±0.17
ModSimFBI [18]0.59±0.100.49±0.100.05±0.11
TransQASR0.53±0.210.45±0.270.06±0.27
TransQ(1)(10)0.62±0.110.54±0.130.33±0.13
TransQ(2)(11)0.61±0.110.54±0.120.33±0.17
FiTTransQ(1)(10)0.64±0.110.55±0.140.35±0.15
TransQ(2)(11)0.64±0.100.57±0.140.36±0.15
", + "bbox": [ + 114, + 128, + 854, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c63842bc1ec68e5899d631f05e4a9fd010da1807ee7e51175712e54e5aa38577.jpg", + "image_caption": [ + "Figure 6: $\\hat{T}_{s,t}$ as a function of the number of available sources for several selection methods. Dotted lines refer to the selection of a unique model for all images and solid lines refer to a model selected per image. Attack is DI [32]." + ], + "image_footnote": [], + "bbox": [ + 94, + 438, + 444, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Criterion TransQ(s,x). Improving transferability without querying the target model in a preliminary step is possible thanks to TransQ(s,x). Adversarial examples that exhibit good transferability on multiple models are more likely to also deceive the unknown targeted model. Figure 6 shows that a significant improvement in transferability is achieved even with only a few available models. Table 1 confirms this observation for the two other attack methods. This strategy is indeed better than the selection based on model similarity.", + "bbox": [ + 75, + 718, + 467, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Score FiT (s,t,x). Combining both criteria together as in (8) leads to a slight improvement in transferability com", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pared to $\\mathrm{TransQ}(s,x)$ alone. Fig. 6 confirms this holds over a wide range of numbers of available sources. For single-model attacks, $\\mathrm{TransQ}^{(1)}$ gives slightly better results than $\\mathrm{TransQ}^{(2)}$ .", + "bbox": [ + 498, + 436, + 890, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual results Figure 2 visually demonstrates the impact of different model selection methods on the quality of adversarial examples. Even when FiT is not accurate, the resulting adversarial examples are still close to the best ones obtained from the source models, and the perturbation remains imperceptible. However, random selection generates noisy perturbations, and the worst-case scenario destroys the image entirely.", + "bbox": [ + 496, + 497, + 890, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.2 Ensemble-model attacks", + "text_level": 1, + "bbox": [ + 500, + 635, + 723, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of model selection in ensemble-model attack. Ensemble attacks remain an under-studied area due to the significant computational resources required for evaluating attacks. Consequently, these attacks have only been evaluated with a limited number of sources. Appendix D.2 shows that increasing the number of sources is not necessarily beneficial. The FiT score provides a scalable solution: we select a small subset of sources based on their FiT scores (top-3) from the bigger set of available sources. In a way, it is better to put quality above quantity. For example, when running an ensemble-model attack against $\\times \\mathrm{CiT}_{\\mathrm{nano}}$ selecting only the three best models using the FiT measure can lead to a significant improvement in transferability compared to using a larger set of models. In our experiments, an ensemble-model of 20 random models achieved a $\\hat{T}_{s,t}$ of 0.47, while an ensemble-model of only three models se", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4548", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3ff88b575b363ce4a6f5f6680b7ed97f6b04653660aba440a1e0614b97598eb4.jpg", + "table_caption": [ + "Table 2: Visual impact of the source selection with DI [32] attacking ConViTbase (first row) and DPN92 (second row)." + ], + "table_footnote": [], + "table_body": "
OriginalBestFiT(TransQ(1))RandomWorst
label source distortionmouseoil filter PiTsmall-dist 5.62purse ConViTbase 8.13purse MobileNetV2110d 35.9jigsaw puzzle DenseNet121 87.8
label source distortiontramelec. locomotive ResNetV250x1-dist 6.94elec. locomotive PiTsmall-dist 7.50racing car DLA60 33.3jigsaw puzzle MixNetmedium 128.7
", + "bbox": [ + 165, + 114, + 803, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lected with $\\mathbb{F}\\mathbb{i}\\mathbb{T}$ was able to achieve a $\\hat{T}_{s,t}$ of 0.56 against the same target.", + "bbox": [ + 75, + 425, + 470, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance of ensemble-model attacks. Table. 1 shows that ensemble-model transferability surpasses that of single-model attacks. While the average results over a random selection of 3 sources increase, they remain closer to the black box results than the white box ones (transferability lower than 0.5). Choosing the top-3 sources returned by our scores for leading the ensemble-model attack yields better performance, comparable to the best possible results obtainable with ensemble-model attacks. Notably, for the DI [32] and TAIG [9] method, it approaches the performance of white box results (transferability greater than 0.5).", + "bbox": [ + 75, + 470, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7e484f0799afeaf937ecc9202a6ae9307259c7b4ba91c43d701c7b0a7bf291a5.jpg", + "table_caption": [ + "Table 3: Transferability $\\hat{T}_{s,t}$ for white box and transferable attacks (single-model). The FiT score uses $\\mathrm{TransQ}^{(1)}$ ." + ], + "table_footnote": [], + "table_body": "
AttackBestFiT
W.B.BP [35]0.27 ±0.18-0.06±0.31
DeepFool [19]0.03 ±0.20-0.34±0.34
PGD [16]0.29 ±0.20-0.01±0.36
I-FGSM [10]0.29 ±0.20-0.03±0.34
Trans.DI [32]0.52 ±0.120.40±0.13
TAIG [9]0.46 ±0.120.27±0.17
DWP [26]0.34 ±0.130.12±0.20
All Attacks0.65 ±0.090.48±0.14
", + "bbox": [ + 78, + 744, + 465, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.3 White box vs. transferable attacks", + "text_level": 1, + "bbox": [ + 500, + 426, + 794, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our last result is that the selection of a single source with the FiT score does not make traditional white box attacks transferable. Table 3 shows that BP [35], DeepFool [19], PGD [16], and I-FGSM [10] yield negative transferability values. More precisely, BP [35], I-FGSM [10], and PGD [16] perform better with FiT than the transferable attacks without any selection mechanism (ie. on average with a random source). However, section 4.4 highlights that the traditional white box attacks may be competitive for some inputs demanding low adversarial perturbation distortion. It is valuable to add them to the options of the attacker and let the FiT score decide the preferable option (source and attack method). This provides our best transferability score for a single model attack, close to 0.5.", + "bbox": [ + 496, + 450, + 890, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 619, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability is a crucial feature of adversarial examples as it allows a single perturbation to deceive multiple models. However, solely relying on Attack Success Rate (ASR) to measure transferability overlooks the degree of distortion needed to fool a model. This paper introduces a novel approach to assess transferability by comparing it to the distortion of two reference attacks: white box and black box attacks. We show that transferable attacks can perform worse than black box attacks without an appropriate selection of the source model, highlighting the need to choose the best source model to target a specific model.", + "bbox": [ + 496, + 703, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed solution, named F i T, allows the attacker to choose one of the best source models with minimal", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4549", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "queries to the target. Our experiments demonstrate that the proposed solution performs well in multiple attack scenarios.", + "bbox": [ + 76, + 90, + 468, + 133 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This study has highlighted the differences in transferability between images for the same source model and their particularity for this specific network. Further research could focus on addressing this issue and investigating its underlying causes, with the hope of designing an even better selection mechanism able to spot the best source model.", + "bbox": [ + 76, + 136, + 468, + 226 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 239, + 171, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mauro Barni, Kassem Kallas, Ehsan Nowroozi, and Benedetta Tondi. On the transferability of adversarial examples against CNN-based image forensics. In ICASSP. IEEE, 2019.", + "[2] Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Ipguard: Protecting intellectual property of deep neural networks via fingerprinting the classification boundary. In ACM Asia Conference on Computer and Communications Security, 2021.", + "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In Symposium on Security and Privacy. IEEE, 2017.", + "[4] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In ACM SIGKDD, 2020.", + "[5] Marco De Angelis and Ander Gray. Why the 1-wasserstein distance is the area between the two marginal cdfs, 2021.", + "[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019.", + "[7] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019.", + "[8] Tianjin Huang, Vlado Menkovski, Yulong Pei, Yuhao Wang, and Mykola Pechenizkiy. Direction-aggregated attack for transferable adversarial examples. JETC, 2022.", + "[9] Yi Huang and Adams Wai-Kin Kong. Transferable adversarial attack based on integrated gradients. In ICLR, 2022.", + "[10] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018.", + "[11] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In CVPR, 2020.", + "[12] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020.", + "[13] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020.", + "[14] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017." + ], + "bbox": [ + 76, + 263, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Nils Lukas, Yuxuan Zhang, and Florian Kerschbaum. Deep neural network fingerprinting by conferrable adversarial examples. In ICLR, 2021.", + "[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2017.", + "[17] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Surfree: a fast surrogate-free black-box attack. In CVPR, 2021.", + "[18] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Fbi: Fingerprinting models with benign inputs. arXiv preprint arXiv:2208.03169, 2022.", + "[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016.", + "[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In ICLR, 2022.", + "[21] Zirui Peng, Shaofeng Li, Guoxing Chen, Cheng Zhang, Haojin Zhu, and Minhui Xue. Fingerprinting deep neural networks globally via universal adversarial perturbations, 2022.", + "[22] Deyan Petrov and Timothy M Hospedales. Measuring the transferability of adversarial examples. arXiv preprint arXiv:1907.06291, 2019.", + "[23] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. In NeurIPS, 2021.", + "[24] Ali Rahmati, Seyed-Mohsen Moosavi-Dezfooli, Pascal Frossard, and Huaiyu Dai. Geoda: a geometric framework for black-box adversarial attacks. In CVPR, 2020.", + "[25] Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017.", + "[26] Hung-Jui Wang, Yu-Yu Wu, and Shang-Tse Chen. Enhancing targeted attack transferability via diversified weight pruning. arXiv preprint arXiv:2208.08677, 2022.", + "[27] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021.", + "[28] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting adversarial transferability through enhanced momentum. In BMVC, 2021.", + "[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021.", + "[30] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.", + "[31] Lei Wu and Zhanxing Zhu. Towards understanding and improving the transferability of adversarial examples in deep neural networks. In ACML, pages 837-850. PMLR, 2020.", + "[32] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019.", + "[33] Yifeng Xiong, Jiadong Lin, Min Zhang, John E Hopcroft, and Kun He. Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In CVPR, 2022." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4550", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Liping Yuan, Xiaqing Zheng, Yi Zhou, Cho-Jui Hsieh, and Kai-Wei Chang. On the transferability of adversarial attacks against neural text classifier. In EMNLP, 2021.", + "[35] Hanwei Zhang, Yannis Avrithis, Teddy Furon, and Laurent Amsaleg. Walking on the edge: Fast, low-distortion adversarial examples. IEEE Trans. on IFS, 2020.", + "[36] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022.", + "[37] Jingjing Zhao, Qingyue Hu, Gaoyang Liu, Xiaoqiang Ma, Fei Chen, and Mohammad Hassan. AFA: Adversarial Fingerprinting Authentication for deep neural networks. Computer Communications, 2020.", + "[38] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, and Michael Backes. Towards good practices in evaluating transfer adversarial attacks. arXiv preprint arXiv:2211.09565, 2022.", + "[39] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018.", + "[40] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020." + ], + "bbox": [ + 78, + 90, + 467, + 445 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4551", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_model.json b/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f927b8555bb7d3bb1d734266083045fe14a3c09 --- /dev/null +++ b/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.768, + 0.153 + ], + "angle": 0, + "content": "How to choose your best allies for a transferable attack?" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.181, + 0.317, + 0.252 + ], + "angle": 0, + "content": "Thibault Maho* \nUniv. Rennes, Inria, CNRS IRISA, Rennes, France \nthibault.maho@inria.fr" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.181, + 0.616, + 0.234 + ], + "angle": 0, + "content": "Seyed-Mohsen Moosavi-Dezfooli \nImperial College London, UK \nseyed.moosavi@imperial.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.656, + 0.181, + 0.874, + 0.252 + ], + "angle": 0, + "content": "Teddy Furon† \nUniv. Rennes, Inria, CNRS IRISA, Rennes, France \nteddy.furon@inria.fr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.577 + ], + "angle": 0, + "content": "The transferability of adversarial examples is a key issue in the security of deep neural networks. The possibility of an adversarial example crafted for a source model fooling another targeted model makes the threat of adversarial attacks more realistic. Measuring transferability is a crucial problem, but the Attack Success Rate alone does not provide a sound evaluation. This paper proposes a new methodology for evaluating transferability by putting distortion in a central position. This new tool shows that transferable attacks may perform far worse than a black box attack if the attacker randomly picks the source model. To address this issue, we propose a new selection mechanism, called \\( F \\vdash T \\), which aims at choosing the best source model with only a few preliminary queries to the target. Our experimental results show that \\( F \\vdash T \\) is highly effective at selecting the best source model for multiple scenarios such as single-model attacks, ensemble-model attacks and multiple attacks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.608, + 0.21, + 0.624 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Transferability is one of the most intriguing properties of adversarial examples. A white box attack crafting adversarial examples for an open-source model is likely to fool other models too [1, 14, 20, 25, 34]. This makes the threat of adversarial examples more realistic. In practice, the model targeted is usually unknown but accessible as a black box. This prevents directly applying any white box gradient-based attack [10, 16, 19, 35]. Black box attacks do exist but they require some thousands of queries to find an adversarial example of low distortion [4, 11, 17, 24]. Transferable attacks require no or few queries to fine-tune an adversarial example thanks to the help of a publicly available model similar enough to the target." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.832, + 0.47, + 0.863 + ], + "angle": 0, + "content": "Transferability is usually measured by the Attack Success Rate (ASR), i.e., the probability that the adversarial" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.289, + 0.868, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.606 + ], + "angle": 0, + "content": "Figure 1: Evaluation of transferability by comparing the Attack Success Rate vs. distortion trade-off of a white box, transferable, and black box attacks against model CoatLitesmall (See Sect. 4.1 for details). The blue area is the range of trade-off operated by a transferable attack with random source models. A transferable attack may be worse than a black box attack without a good source selection (like FiT)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.894, + 0.747 + ], + "angle": 0, + "content": "example crafted for the source model also deludes the target model. We argue that this measure leads to an unfair evaluation of transferability. In the context of adversarial examples, it is not just a matter of discovering data that is not well classified, but rather identifying the perturbation that can fool a classifier with minimal distortion. This principle should also apply to transferable attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "For illustration purposes, let us consider two models, one is robust in the sense that the necessary amount of adversarial perturbation is large, whereas the other model is weak. If the attacker uses the robust model as the source to attack the weak target network, the ASR of the transferable attack will certainly be big. It does not mean that this is the right choice. The ASR is high because the robust source model needs large perturbation to be deluded, which will fool any weaker model. The ASR alone does not reflect the overshooting in distortion. The converse, using the weak to at" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.468, + 0.888 + ], + "angle": 0, + "content": "*Thanks to Rennes Métropole for its funding for international mobility." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.468, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\)Thanks to ANR and AID french agencies for funding Chaire SAIDA." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4542" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "tack the robust, would yield a low ASR. To summarize, the ASR alone fails to capture how relevant the direction of the perturbation given by the source is for attacking the target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.242 + ], + "angle": 0, + "content": "The first contribution of this paper is to put distortion back into the picture. Section 3 evaluates transferability by comparing the distortion of a transferable attack to the ones of two reference attacks: On one hand, the strongest attack, i.e., the white box attack directly applied on the target model; on the other hand, the weakest attack, i.e., the black box attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.468, + 0.44 + ], + "angle": 0, + "content": "The second contribution shows the great variability of the performance of transferable attacks. Figure 1 summarizes this observation by plotting the ASR as a function of the distortion (the experimental protocol is explained in Sect. 4.1). Naturally, the black box attack needs much more distortion than the white box attack. For instance, the white box attack yields an ASR of \\(50\\%\\) with a distortion of 0.19, whereas the black box attack needs a distortion of 9.7. The surprise is that if the attacker resorts to a transferable attack and picks a source model at random, there is almost a \\(50\\%\\) chance that the attack performs even worse than the black box attack. Section 4 outlines a triad of factors: input, source model, and attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.468, + 0.562 + ], + "angle": 0, + "content": "This observation challenges the prevalent notion that adversarial examples transfer easily between models, and highlights the need to carefully choose the source model to attack a target. Under the assumption that the attacker has indeed several candidate models, our third contribution, named FiT in Sect. 5, provides an affordable measure for model selection, allowing the attacker to choose a good source model with only a few queries to the target." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.578, + 0.214, + 0.594 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.468, + 0.619 + ], + "angle": 0, + "content": "2.1. White box, black box, and transferable attacks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.468, + 0.703 + ], + "angle": 0, + "content": "The threat analysis related to adversarial examples usually considers two scenarios: the white box assumption where the attacker knows the internals of the target model, and the black box assumption where he does not but has limited access to the target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.84 + ], + "angle": 0, + "content": "White box attacks are now performing very well in terms of Attack Success Rate and distortion. We distinguish two kinds of attacks: 1) attacks constrained by a distortion budget like PGD [16] whose performance is measured by the ASR, and 2) attacks yielding almost surely an adversarial example like DeepFool [19] and CW [3] whose performance is measured by the distortion. The trend is to make them speed efficient as well with fewer computations of the gradient of the neural network, like FMN [23] and BP [35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our paper also considers so-called decision-based black box attacks like GeoDA [24], SurFree [17], and RayS [4], where the attacker only sees the class predicted by the model, but not the confidence score, neither the pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "bit nor the logit. A decision-based black box attack typically needs thousands of queries to find an adversarial example with a relatively low distortion, yet, much higher than with a white box attack. This is the price the attacker has to pay when he does not know the target model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.394 + ], + "angle": 0, + "content": "The transferable attacks, which pertain to the black box scenario, disrupt this last statement. The assumption is that attacker knows another model, called source, trained for the same classification problem as the target. Yet, leading a white box attack on the source model with the hope that this adversarial example will also fool the target yields poor results because the adversarial perturbation is too specific to the source. Modern transferable attacks increase the ASR by avoiding the overfitting of examples on the source. Input transformations have been purported by DI [32], TI [6] and Admix [27]. Papers [8, 13, 28] stabilized the gradient while [7, 29, 36, 39] focus on the importance of intermediate features. The combination of several methods has also been investigated [9, 40]. For instance, TAIG [9] uses input transformation, integrated gradients, and attention reduction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.892, + 0.546 + ], + "angle": 0, + "content": "Ensemble-model attacks make the assumption that the attacker leverages not a single source but several source models. Paper [28] averages the logits of several sources and leads a white box attack on this aggregation of models, whereas [33] reduces model variance. The main obstacle to these methods is the computational complexity. To counteract this drawback, some methods design a specific ensemble of sources like ghost networks [12] or pruned network [26]. This amounts to creating several sources without the need to compute the gradient of each model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.622 + ], + "angle": 0, + "content": "All this literature uses the ASR as the figure of merit for a given distortion while our paper draws the full operating characteristic of ASR vs. distortion. Moreover, the performance of the transferability is compared neither to the white box nor the decision-based black box attack." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.651, + 0.652 + ], + "angle": 0, + "content": "2.2. Fingerprinting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Models are valuable assets because of their training costs, be it the expertise, the annotated data, or the computational power. Fingerprinting methods have been proposed to protect this intellectual property. They provide a similarity score between the two models by looking at similar decisions in submitted queries. They often use adversarial examples specifically designed to be unique for a model. The models are considered similar if they share the same decision for these inputs. IPGuard [2] creates adversarial examples close to the decision boundary to frame it. Universal adversarial perturbations are used to characterize the boundary in [21]. The paper [15] built adversarial examples with a transferable property called conferrable adversarial examples. AFA [37] crafts adversarial examples model-specific using dropout. FBI [18] is the only method using unmodified images. It estimates the mutual information to measure" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4543" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.435, + 0.106 + ], + "angle": 0, + "content": "the statistical dependence of two models' predictions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.107, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Our paper investigates the idea that a model similar to the target may be a good source for transferable attacks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.153, + 0.214, + 0.17 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.179, + 0.469, + 0.24 + ], + "angle": 0, + "content": "This paper introduces a new way to gauge transferability. The main motivation is that transferability should indicate whether a source \\( s \\) is useful for attacking a target \\( t \\), independently from the inherent robustness of \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.25, + 0.188, + 0.265 + ], + "angle": 0, + "content": "3.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.469, + 0.35 + ], + "angle": 0, + "content": "Let \\( f:[0,255]^N \\to \\mathcal{R}^K \\) denote a classifier computing the predicted probabilities \\( f(x) \\) of \\( K \\) classes for input \\( x \\). For a given input \\( x \\) of class \\( y \\), the adversarial example \\( x_a \\) is the result of an attack (be it white box, black box, or by transferability) such that" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.363, + 0.47, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\arg \\max _ {1 \\leq k \\leq K} f _ {k} (x _ {a}) \\neq \\arg \\max _ {1 \\leq k \\leq K} f _ {k} (x) = y. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.469, + 0.475 + ], + "angle": 0, + "content": "In the case of transferable attacks, the attacker uses a source model \\( f_{s} \\) to build a transferable adversarial example against the target model \\( f_{t} \\). This paper considers a collection of \\( m \\) source models denoted by \\( \\mathcal{F}_s = \\{f_s^1,\\dots,f_s^m\\} \\), and a set of \\( n \\) inputs \\( \\mathcal{X} \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.485, + 0.22, + 0.5 + ], + "angle": 0, + "content": "3.2. Measurement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.469, + 0.555 + ], + "angle": 0, + "content": "Distortion. We measure the distortion of an adversarial perturbation using the Euclidean norm which is a common choice as we deal with natural images:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.566, + 0.469, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s t} \\left(x _ {a}, x\\right) := \\| x _ {a} - x \\| _ {2} / \\sqrt {N}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.597, + 0.469, + 0.627 + ], + "angle": 0, + "content": "with \\(N = 3\\times H\\times L\\) pixels. This distortion can be seen as the average amplitude of the perturbation per pixel." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.469, + 0.673 + ], + "angle": 0, + "content": "Operating characteristic. Given an attack (be it white box, black box, or by transferability) and a target model \\( f_{t} \\), we call operating characteristic the following function:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.686, + 0.469, + 0.703 + ], + "angle": 0, + "content": "\\[\nP (D) := \\mathbb {P} (\\operatorname {d i s t} \\left(x _ {a}, x\\right) < D). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.715, + 0.469, + 0.775 + ], + "angle": 0, + "content": "It is thus the Attack Success Rate as a function of the distortion. Contrary to the literature which measures the ASR for a few distortion levels, we consider the full range of \\( D \\) values." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.786, + 0.229, + 0.803 + ], + "angle": 0, + "content": "3.3. Transferability" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given a white box and a black box attacks, The proposed methodology first computes the operating characteristic \\( P_{t}^{\\mathrm{wb}}(D) \\) of a state-of-the-art white box attack directly applied to the target model, and the operating characteristic \\( P_{t}^{\\mathrm{bb}}(D) \\) of a state-of-the-art black box attack. It then measures where the operating characteristic of the transferable" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "attack from the source \\(s\\) to the target \\(t\\) lies in between the two characteristics as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.132, + 0.892, + 0.17 + ], + "angle": 0, + "content": "\\[\nT _ {s, t} := \\frac {\\int_ {0} ^ {\\infty} P _ {s , t} (u) - P _ {t} ^ {\\mathrm {b b}} (u) d u}{\\int_ {0} ^ {\\infty} P _ {t} ^ {\\mathrm {w b}} (u) - P _ {t} ^ {\\mathrm {b b}} (u) d u}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.181, + 0.892, + 0.424 + ], + "angle": 0, + "content": "The proposed score is calculated as the ratio of the areas between the different operational characteristics, which are defined as Cumulative Distribution Functions (CDFs). The numerator is therefore close to the 1-Wasserstein distance [5] between the ASR function of the distortions of the transferable attack and that of the black box attack, where the absolute value is removed to obtain a signed score. If the transferable attack performs as well as the white box attack (resp. as bad as the black box attack), then \\( T_{s,t} = 1 \\) (resp. \\( T_{s,t} = 0 \\)). Figure 1 shows that the numerator of (4) can indeed be negative as transferability can be even worse than the black box if the source \\( s \\) is not well chosen. Therefore, zero is not a lower bound for \\( T_{s,t} \\). In the denominator, the 1-Wasserstein distance is obtained exactly between the white box and the black box because the white box consistently generates adversarial examples with lower distortion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.731, + 0.451 + ], + "angle": 0, + "content": "3.4. Practical implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Attacks. We only consider state-of-the-art attacks which almost surely deliver an adversarial example. As said in Sect. 2.1, this is usually the case for black box attacks, but we are limited to using white box attacks that are not subject to a distortion budget. As for transferability, the attack uses the publicly available model \\( s \\) and input \\( x \\) to craft an adversarial direction \\( u_{x,s} \\) of Euclidean norm \\( \\sqrt{N} \\). The use of \\( \\sqrt{N} \\) for the norm is for the sake of simplicity in notation. We assume that there is an oracle giving the minimal distortion along this direction to fool the target \\( t \\). In other words, \\( x_{a} = x + d u_{x,s} \\), with" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.64, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\nd = \\min \\left\\{\\delta : \\arg \\max _ {k} f _ {t, k} (x + \\delta u _ {x, s}) \\neq y \\right\\}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.891, + 0.736 + ], + "angle": 0, + "content": "Note that \\(\\mathrm{dist}(x_a,x) = d\\). This definition favours transferability as its best. In practice, such an oracle does not exist but the attacker finds a good estimate of (5) thanks to a line search within a few queries to the target." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Transferability. We run a given attack over a collection \\(\\mathcal{X}\\) of \\(n\\) inputs correctly classified by the target model. We compute the distortions \\(d(j) = \\mathrm{dist}(x_{a,j}, x_j)\\) with \\(x_j \\in \\mathcal{X}\\) and sort them so that \\(d(1) \\leq d(2) \\leq \\ldots \\leq d(n)\\). We set \\(d(0) = 0\\) and \\(d(n + 1) = \\infty\\) to properly define the empirical operating characteristic by the following step function:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.84, + 0.892, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\hat {P} (D) := j / n \\quad \\forall D \\in [ d (j), d (j + 1)). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "It is then easier to estimate the integrals appearing in (4), which are indeed areas under two curves, by a Lebesgue" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4544" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.373, + 0.108 + ], + "angle": 0, + "content": "sum rather than a Riemann sum. This gives:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.114, + 0.47, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\hat {T} _ {s, t} := \\frac {\\sum_ {j = 1} ^ {n} d _ {s , t} (j) - d _ {t} ^ {\\mathrm {b b}} (j)}{\\sum_ {j = 1} ^ {n} d _ {t} ^ {\\mathrm {w b}} (j) - d _ {t} ^ {\\mathrm {b b}} (j)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.161, + 0.47, + 0.207 + ], + "angle": 0, + "content": "where the distortions \\((d_t^{\\mathrm{bb}}(j))_j\\) (resp. \\((d_t^{\\mathrm{wb}}(j))_j\\)) resulting from the black box (resp. white box) attack against model \\(t\\) are also sorted in increasing order." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.219, + 0.468, + 0.236 + ], + "angle": 0, + "content": "4. Triad of transferability: data, model, attack" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.245, + 0.469, + 0.275 + ], + "angle": 0, + "content": "This section is an experimental investigation of the factors affecting transferability." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.283, + 0.266, + 0.299 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.469, + 0.41 + ], + "angle": 0, + "content": "The study assesses transferable attacks on a total of 48 models, with 47 of them sourced from the Timm library [30] and one very robust, namely ResNet50AdvTrain, obtained from the GitHub repository 1. Appendix A lists all the models. The experiments utilize 100 images from the validation set of ILSVRC'12, all of which are correctly classified by all models under consideration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.469, + 0.531 + ], + "angle": 0, + "content": "The study considers three transferable attacks - DI [32], TAIG [9], and DWP [26]- each using a different approach to improve transferability (see Sect. 2.1). These attacks are selected as the best in their categories in [38]. They all share an \\(\\epsilon\\) parameter to control the maximum perturbation added per pixel. The effect of the parameter \\(\\epsilon\\) indeed happens to be negligible in our protocol, as demonstrated in Appendix B. We choose \\(\\epsilon = 8\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.469, + 0.744 + ], + "angle": 0, + "content": "To measure transferability (7), we need state-of-the-art black box and white box attacks. Certain methods may exhibit a preference towards one model over the other, necessitating the use of multiple attacks. Our study employs four white box attacks (BP [35], DeepFool [19], I-FGSM [10], and PGD [16]) and three black box attacks (SurFree [17], RayS [4], and GeoDA [24]). All black box attacks are run with 2,000 queries, which has been determined to be sufficient for achieving convergence. For distortion-constrained white box attack, the \\(\\epsilon\\) parameter is set to 4. We record the smallest perturbation distortion over the black box (white box) for each image and draw the operating characteristic (6) as appearing in green (resp. red) dashed line in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.469, + 0.879 + ], + "angle": 0, + "content": "As for the transferable attack, for a given target model, the attacker has access to a subset of all the other models whose architecture differs from the target. This amounts to an average of 45 models out of 48. For instance, in Fig. 1, CoatLitesmall being the target, we exclude all other CoatLite models from being a source. The light blue area delimits the operating characteristics of transferable attack DI [32] using as the source one of the 45 remaining models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.681, + 0.107 + ], + "angle": 0, + "content": "4.2. Model dependence" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.128, + 0.891, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.419, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 2: Transferability score \\(\\hat{T}_{s,t}\\) matrix of 48 sources and 48 targets listed in App. A with attack DI [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Large transferability variation. Figure 2 shows the matrix \\(\\hat{T}_{s,t}\\), where \\(s\\) and \\(t\\) are any of the \\(48^2\\) pairs, for attack D I [32]. Not all models possess the same transferability capabilities. At first look, the figure contains more blue than red cells which means that transferability takes negative value more often." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Some rare models, like PiTsmall-dist, exhibit good transferability towards any target. On the contrary, DPN68b is always a bad source. On the other hand, ResNet50AdvTrain is a very difficult target (note how" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.66, + 0.856, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.851, + 0.892, + 0.897 + ], + "angle": 0, + "content": "Figure 3: \\(\\hat{T}_{s,t}\\) function of the number of available sources and attack DI [32]. The best or worst model selected per image (solid line) or on average (dashed line)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.403, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/MadryLab/robustness" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4545" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.445, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Figure 4: Distribution of best image-model selection function of the number of sources. Adversarial examples obtained with DI [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.404 + ], + "angle": 0, + "content": "ever that its accuracy is low), followed by PiTlight and ConViTtiny, whereas the family of ResNet are fairly easy target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.595 + ], + "angle": 0, + "content": "Prior works have shown that models with architectures similar to the target transfer better [22, 31]. The red squares appearing in the matrix confirm this (models are ordered according to architecture family). For instance, MobileNetV2 and RexNet architectures exhibit transferability close to 0.7. However, this is not an absolute truth. For example, EfficientNet-B0 has better transferability against MixNetlarge than MixNetmedium, even though they have similar architectures. ConViT transfers exceptionally well to Twins although their architectures are very different. Similar observations hold for the other attacks but with lower transferability scores (see Appendix C)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.605, + 0.469, + 0.636 + ], + "angle": 0, + "content": "From now on, we exclude models whose architecture is similar to the target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Impact on the attack. Selecting the right source among several available models is critical for achieving high transferability. Figure 3 displays the transferability score (7) for the best and the worst choices (shown in dotted lines) as the number of source candidates increases. In this evaluation, a single model is selected as the source for building all adversarial examples. On average, the transferability is lower than zero, regardless of the attack method. This means that transferable attacks perform worst than black box attacks on average. If the attacker knows how to select the best source model, \\(\\hat{T}_{s,t}\\) quickly converges to a maximum which is positive but below 0.5. This means that transferability at its best performs closer to a black box attack rather than a white box attack. These remarks highly mitigate the threat of transferability and put emphasis on the crucial selection of the best source model. As far as we know, these facts are not reported in the adversarial examples literature." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.693, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.221, + 0.693, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.091, + 0.871, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.221, + 0.871, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.359, + 0.892, + 0.39 + ], + "angle": 0, + "content": "Figure 5: 2D Histogram of the minimum distortion for transferable and white box attacks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.418, + 0.68, + 0.435 + ], + "angle": 0, + "content": "4.3. Image dependence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.608 + ], + "angle": 0, + "content": "The difficulty of transferring adversarial examples from a source to a target varies significantly depending on the input \\( x \\). This is illustrated in Fig. 4, which shows the distribution of the best-performing model for each image based on the available sources. Even with a large number of sources available, there is typically one source that performs significantly better than the others, but only over \\( 60\\% \\) of the images on average. However, this superiority decreases rapidly as the number of available sources increases. It means that a better source exists for \\( 40\\% \\) of the images on average." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.714 + ], + "angle": 0, + "content": "Supposing that the attacker knows the best source of each input, Figure 3 shows in solid line that the transferability converges to a value close to 0.5. The performance of the transferable attack lies in between the ones of the white box and black box attacks. This highlights the importance of selecting an appropriate source model for a given target and input." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.683, + 0.743 + ], + "angle": 0, + "content": "4.4. Attack dependence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Transferable attack. Figure 5 compares transferable attacks with a 2D histogram of the distortion pair for two attacks. This is computed over all inputs in \\(\\mathcal{X}\\) and all pairs of source and target models. The results show that DWP [26] exhibits poor transferability compared to DI [32] and TAIG [9], which produce adversarial perturbations with similar distortion. Additionally, regardless of the attack complexity or method, the challenging images remain consistent. If one attack requires a high distortion for a given input / source / target, other attacks are likely to encounter" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4546" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.205, + 0.105 + ], + "angle": 0, + "content": "similar difficulties." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.107, + 0.471, + 0.306 + ], + "angle": 0, + "content": "Traditional white box attack. We now compare the methods designed for transferability with a naive approach. A white box attack (ie. not specific to transferability) is executed on the source, and then the direction found serves as \\( u_{x,s} \\) to place the adversarial examples on the target boundary with (5). The best \\( \\hat{T}_{s,t} \\) score is 0.27 for BP [35] compared to 0.52 for DI [32]. This confirms the superiority of the recent methods designed for transferability. However, even though DI [32] is on average better, Figure 5 shows that when the necessary distortion is small, traditional white box attacks like DeepFool [19] and BP [35] indeed beats DI [32]. On the other hand, DI [32] performs much better for inputs requiring more distortion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.318, + 0.353, + 0.334 + ], + "angle": 0, + "content": "5. How to choose the best source" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.47, + 0.403 + ], + "angle": 0, + "content": "Section 4 outlines that the choice of the source is of utmost importance thanks to the transferability measure defined in Sect. 3. This section investigates whether the attacker can guess which model is the best source." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.403, + 0.47, + 0.525 + ], + "angle": 0, + "content": "We now propose a procedure which combines the dependences with respect to the source and target models (Sect. 4.2), and the input (Sect. 4.3). Model dependence is measured by evaluating the similarity between the source and target models, which is denoted as ModSim. On the other hand, the image dependence is measured by the quality of the adversarial example, denoted as TransQ. Both metrics are combined into the following score:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.533, + 0.47, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F i T} (s, t, x) := \\operatorname {M o d S i m} (s, t) \\times \\operatorname {T r a n s Q} (s, x). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.47, + 0.62 + ], + "angle": 0, + "content": "These indicators should be easy to compute. We especially pay attention to the number of queries to the target. This score opens the door to a new strategy for the attacker which first selects the best source among the available models" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.628, + 0.47, + 0.651 + ], + "angle": 0, + "content": "\\[\ns ^ {\\star} (t, x) = \\underset {\\sigma \\in \\mathcal {F} _ {s}} {\\arg \\max } \\mathrm {F i T} (\\sigma , t, x), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.408, + 0.675 + ], + "angle": 0, + "content": "and then crafts the adversarial direction \\( u_{x,s^*(t,x)} \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.283, + 0.699 + ], + "angle": 0, + "content": "5.1. Criterion ModSim(s,t)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Section 4.2 highlights the correlation between the transferability and the similarity between the source and target models. Gauging model similarity has been previously studied in the context of fingerprinting as a defense to protect intellectual property (see Sect. 2.2). This paper uses fingerprinting methods as an attack that leaks information about the target." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We consider the fingerprinting method [18] because it works in a decision-based setup since the target is a black box in our application. Querying two models \\( s \\) and \\( t \\) with few natural images, it computes a distance \\( \\operatorname{Dist}(s, t) \\in [0, 1] \\) by comparing their outputs. Since we look for a similarity, we set \\( \\operatorname{ModSim}(s, t) = 1 - \\operatorname{Dist}(s, t) \\). However," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.893, + 0.137 + ], + "angle": 0, + "content": "this provides a symmetrical similarity, ie. \\(\\operatorname{ModSim}(s,t) = \\operatorname{ModSim}(t,s)\\), while transferability is not (see Fig. 2). This shows that this criterion alone is not sufficient." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.146, + 0.709, + 0.163 + ], + "angle": 0, + "content": "5.2. Criterion TransQ(s, x)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.169, + 0.894, + 0.259 + ], + "angle": 0, + "content": "This criterion evaluates the general transferability of a given adversarial example crafted by a source. Our idea is to leverage the assumption that the attacker has a set of models \\(\\mathcal{F}_s\\). Consequently, we can evaluate the transferability thanks to the other models of this set, without querying the target." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.26, + 0.893, + 0.32 + ], + "angle": 0, + "content": "For a given input \\( x \\), the source \\( s \\) provides the adversarial direction \\( u_{x,s} \\) and we compute the distortion \\( d_{s,\\sigma} \\) necessary to delude classifier \\( \\sigma \\in \\mathcal{F}_s \\) with (5). We then aggregate these distortions into a single score with two flavours:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.329, + 0.893, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T r a n s Q} ^ {(1)} (s, x) := \\left(\\frac {1}{| \\mathcal {F} _ {s} |} \\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {s, \\sigma}\\right) ^ {- 1}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.414 + ], + "angle": 0, + "content": "A good source for input \\( x \\) gives birth to lower distortions, so that \\( \\mathrm{TransQ}^{(1)}(s,x) \\) is large." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.424, + 0.892, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T r a n s} Q ^ {(2)} (s, x) := \\frac {\\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {s , \\sigma} - d _ {\\sigma} ^ {\\mathrm {b b}}}{\\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {\\sigma} ^ {\\mathrm {w b}} - d _ {\\sigma} ^ {\\mathrm {b b}}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.465, + 0.893, + 0.496 + ], + "angle": 0, + "content": "This measure is similar to (7) except that it is computed over the set of models instead of a set of inputs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.504, + 0.593, + 0.518 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.892, + 0.648 + ], + "angle": 0, + "content": "The experimental setup is the same as in Sect. 4.1. We select the fingerprinting method FBI [18] with 200 benign natural images to compute the criterion ModSim(s, t). It implies that the attacker first makes 200 queries to the target in a preliminary step before forging any adversarial example. Appendix D.1 shows that more images improve the fingerprinting accuracy hence the model selection, but the results converge after a number of 200 images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.649, + 0.892, + 0.708 + ], + "angle": 0, + "content": "This section is structured as follows: we use FiT to select the best model for single-model attacks, then we employ it to identify the best subset of models for ensemble-model attacks and different combinations of attacks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.699, + 0.741 + ], + "angle": 0, + "content": "5.3.1 Single-model attacks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Criterion ModSim(s, t). Table 1 indicates that architectural similarity is a reliable measure of transferability between two models. It can drive the selection of a good source giving birth to a transferable attack outperforming the black box attack since \\(\\hat{T}_{s,t}\\) is larger than 0 (except for DWP [26]). Yet, the results remain low compared to the best results obtained. As discussed in Sec 4.2, similarity may not suffice because it implies the selection of one unique for a given target. Better results are achieved when adapting the source to the input." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4547" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.09, + 0.828, + 0.107 + ], + "angle": 0, + "content": "Table 1: Transferability \\( {\\widehat{T}}_{s,t} \\) for DI [32],TAIG [9] and DWP [26] for single and ensemble-model attacks." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.13, + 0.855, + 0.4 + ], + "angle": 0, + "content": "
CategorySelection MethodDI [32]TAIG [9]DWP [26]
Single-model attackBest0.52±0.120.46±0.120.34±0.13
Random-0.16±0.26-0.12±0.21-0.72±0.32
ModSimFBI [18]0.18±0.170.12±0.18-0.39±0.19
TransQASR-0.21±0.39-0.24±0.30-1.14±0.95
TransQ(1)(10)0.38±0.150.24±0.160.10±0.20
TransQ(2)(11)0.37±0.180.23±0.180.08±0.22
FiTTransQ(1)(10)0.40±0.130.27±0.170.12±0.20
TransQ(2)(11)0.39±0.160.25±0.180.10±0.22
Ensemble-model attack with three sourcesBest0.72±0.080.62±0.080.46±0.1
Random0.43±0.120.40±0.110.02±0.17
ModSimFBI [18]0.59±0.100.49±0.100.05±0.11
TransQASR0.53±0.210.45±0.270.06±0.27
TransQ(1)(10)0.62±0.110.54±0.130.33±0.13
TransQ(2)(11)0.61±0.110.54±0.120.33±0.17
FiTTransQ(1)(10)0.64±0.110.55±0.140.35±0.15
TransQ(2)(11)0.64±0.100.57±0.140.36±0.15
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.439, + 0.445, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.692 + ], + "angle": 0, + "content": "Figure 6: \\(\\hat{T}_{s,t}\\) as a function of the number of available sources for several selection methods. Dotted lines refer to the selection of a unique model for all images and solid lines refer to a model selected per image. Attack is DI [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Criterion TransQ(s,x). Improving transferability without querying the target model in a preliminary step is possible thanks to TransQ(s,x). Adversarial examples that exhibit good transferability on multiple models are more likely to also deceive the unknown targeted model. Figure 6 shows that a significant improvement in transferability is achieved even with only a few available models. Table 1 confirms this observation for the two other attack methods. This strategy is indeed better than the selection based on model similarity." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Score FiT (s,t,x). Combining both criteria together as in (8) leads to a slight improvement in transferability com" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.892, + 0.497 + ], + "angle": 0, + "content": "pared to \\(\\mathrm{TransQ}(s,x)\\) alone. Fig. 6 confirms this holds over a wide range of numbers of available sources. For single-model attacks, \\(\\mathrm{TransQ}^{(1)}\\) gives slightly better results than \\(\\mathrm{TransQ}^{(2)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Visual results Figure 2 visually demonstrates the impact of different model selection methods on the quality of adversarial examples. Even when FiT is not accurate, the resulting adversarial examples are still close to the best ones obtained from the source models, and the perturbation remains imperceptible. However, random selection generates noisy perturbations, and the worst-case scenario destroys the image entirely." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.725, + 0.65 + ], + "angle": 0, + "content": "5.3.2 Ensemble-model attacks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Importance of model selection in ensemble-model attack. Ensemble attacks remain an under-studied area due to the significant computational resources required for evaluating attacks. Consequently, these attacks have only been evaluated with a limited number of sources. Appendix D.2 shows that increasing the number of sources is not necessarily beneficial. The FiT score provides a scalable solution: we select a small subset of sources based on their FiT scores (top-3) from the bigger set of available sources. In a way, it is better to put quality above quantity. For example, when running an ensemble-model attack against \\(\\times \\mathrm{CiT}_{\\mathrm{nano}}\\) selecting only the three best models using the FiT measure can lead to a significant improvement in transferability compared to using a larger set of models. In our experiments, an ensemble-model of 20 random models achieved a \\(\\hat{T}_{s,t}\\) of 0.47, while an ensemble-model of only three models se" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4548" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.089, + 0.877, + 0.105 + ], + "angle": 0, + "content": "Table 2: Visual impact of the source selection with DI [32] attacking ConViTbase (first row) and DPN92 (second row)." + }, + { + "type": "table", + "bbox": [ + 0.166, + 0.116, + 0.805, + 0.403 + ], + "angle": 0, + "content": "
OriginalBestFiT(TransQ(1))RandomWorst
label source distortionmouseoil filter PiTsmall-dist 5.62purse ConViTbase 8.13purse MobileNetV2110d 35.9jigsaw puzzle DenseNet121 87.8
label source distortiontramelec. locomotive ResNetV250x1-dist 6.94elec. locomotive PiTsmall-dist 7.50racing car DLA60 33.3jigsaw puzzle MixNetmedium 128.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.471, + 0.458 + ], + "angle": 0, + "content": "lected with \\(\\mathbb{F}\\mathbb{i}\\mathbb{T}\\) was able to achieve a \\(\\hat{T}_{s,t}\\) of 0.56 against the same target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Performance of ensemble-model attacks. Table. 1 shows that ensemble-model transferability surpasses that of single-model attacks. While the average results over a random selection of 3 sources increase, they remain closer to the black box results than the white box ones (transferability lower than 0.5). Choosing the top-3 sources returned by our scores for leading the ensemble-model attack yields better performance, comparable to the best possible results obtainable with ensemble-model attacks. Notably, for the DI [32] and TAIG [9] method, it approaches the performance of white box results (transferability greater than 0.5)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.722 + ], + "angle": 0, + "content": "Table 3: Transferability \\(\\hat{T}_{s,t}\\) for white box and transferable attacks (single-model). The FiT score uses \\(\\mathrm{TransQ}^{(1)}\\)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.745, + 0.466, + 0.886 + ], + "angle": 0, + "content": "
AttackBestFiT
W.B.BP [35]0.27 ±0.18-0.06±0.31
DeepFool [19]0.03 ±0.20-0.34±0.34
PGD [16]0.29 ±0.20-0.01±0.36
I-FGSM [10]0.29 ±0.20-0.03±0.34
Trans.DI [32]0.52 ±0.120.40±0.13
TAIG [9]0.46 ±0.120.27±0.17
DWP [26]0.34 ±0.130.12±0.20
All Attacks0.65 ±0.090.48±0.14
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.795, + 0.442 + ], + "angle": 0, + "content": "5.3.3 White box vs. transferable attacks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.663 + ], + "angle": 0, + "content": "Our last result is that the selection of a single source with the FiT score does not make traditional white box attacks transferable. Table 3 shows that BP [35], DeepFool [19], PGD [16], and I-FGSM [10] yield negative transferability values. More precisely, BP [35], I-FGSM [10], and PGD [16] perform better with FiT than the transferable attacks without any selection mechanism (ie. on average with a random source). However, section 4.4 highlights that the traditional white box attacks may be competitive for some inputs demanding low adversarial perturbation distortion. It is valuable to add them to the options of the attacker and let the FiT score decide the preferable option (source and attack method). This provides our best transferability score for a single model attack, close to 0.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.62, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Transferability is a crucial feature of adversarial examples as it allows a single perturbation to deceive multiple models. However, solely relying on Attack Success Rate (ASR) to measure transferability overlooks the degree of distortion needed to fool a model. This paper introduces a novel approach to assess transferability by comparing it to the distortion of two reference attacks: white box and black box attacks. We show that transferable attacks can perform worse than black box attacks without an appropriate selection of the source model, highlighting the need to choose the best source model to target a specific model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "The proposed solution, named F i T, allows the attacker to choose one of the best source models with minimal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4549" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "queries to the target. Our experiments demonstrate that the proposed solution performs well in multiple attack scenarios." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.137, + 0.47, + 0.227 + ], + "angle": 0, + "content": "This study has highlighted the differences in transferability between images for the same source model and their particularity for this specific network. Further research could focus on addressing this issue and investigating its underlying causes, with the hope of designing an even better selection mechanism able to spot the best source model." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.24, + 0.173, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.265, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[1] Mauro Barni, Kassem Kallas, Ehsan Nowroozi, and Benedetta Tondi. On the transferability of adversarial examples against CNN-based image forensics. In ICASSP. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.322, + 0.472, + 0.39 + ], + "angle": 0, + "content": "[2] Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Ipguard: Protecting intellectual property of deep neural networks via fingerprinting the classification boundary. In ACM Asia Conference on Computer and Communications Security, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.47, + 0.434 + ], + "angle": 0, + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In Symposium on Security and Privacy. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.435, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[4] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In ACM SIGKDD, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.477, + 0.469, + 0.505 + ], + "angle": 0, + "content": "[5] Marco De Angelis and Ander Gray. Why the 1-wasserstein distance is the area between the two marginal cdfs, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.506, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.548, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[7] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.605, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[8] Tianjin Huang, Vlado Menkovski, Yulong Pei, Yuhao Wang, and Mykola Pechenizkiy. Direction-aggregated attack for transferable adversarial examples. JETC, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.647, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[9] Yi Huang and Adams Wai-Kin Kong. Transferable adversarial attack based on integrated gradients. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.676, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[10] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.718, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[11] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[12] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[13] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.265, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Nils Lukas, Yuxuan Zhang, and Florian Kerschbaum. Deep neural network fingerprinting by conferrable adversarial examples. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.891, + 0.204 + ], + "angle": 0, + "content": "[17] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Surfree: a fast surrogate-free black-box attack. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[18] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Fbi: Fingerprinting models with benign inputs. arXiv preprint arXiv:2208.03169, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[21] Zirui Peng, Shaofeng Li, Guoxing Chen, Cheng Zhang, Haojin Zhu, and Minhui Xue. Fingerprinting deep neural networks globally via universal adversarial perturbations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[22] Deyan Petrov and Timothy M Hospedales. Measuring the transferability of adversarial examples. arXiv preprint arXiv:1907.06291, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[23] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[24] Ali Rahmati, Seyed-Mohsen Moosavi-Dezfooli, Pascal Frossard, and Huaiyu Dai. Geoda: a geometric framework for black-box adversarial attacks. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[25] Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[26] Hung-Jui Wang, Yu-Yu Wu, and Shang-Tse Chen. Enhancing targeted attack transferability via diversified weight pruning. arXiv preprint arXiv:2208.08677, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[27] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[28] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting adversarial transferability through enhanced momentum. In BMVC, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[30] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[31] Lei Wu and Zhanxing Zhu. Towards understanding and improving the transferability of adversarial examples in deep neural networks. In ACML, pages 837-850. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[32] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[33] Yifeng Xiong, Jiadong Lin, Min Zhang, John E Hopcroft, and Kun He. Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In CVPR, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4550" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[34] Liping Yuan, Xiaqing Zheng, Yi Zhou, Cho-Jui Hsieh, and Kai-Wei Chang. On the transferability of adversarial attacks against neural text classifier. In EMNLP, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[35] Hanwei Zhang, Yannis Avrithis, Teddy Furon, and Laurent Amsaleg. Walking on the edge: Fast, low-distortion adversarial examples. IEEE Trans. on IFS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[36] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[37] Jingjing Zhao, Qingyue Hu, Gaoyang Liu, Xiaoqiang Ma, Fei Chen, and Mohammad Hassan. AFA: Adversarial Fingerprinting Authentication for deep neural networks. Computer Communications, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[38] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, and Michael Backes. Towards good practices in evaluating transfer adversarial attacks. arXiv preprint arXiv:2211.09565, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[39] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[40] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4551" + } + ] +] \ No newline at end of file diff --git a/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_origin.pdf b/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b9a54a853978b7f95bf2418de7b6a5b11f7cdae --- /dev/null +++ b/2023/How to Choose your Best Allies for a Transferable Attack_/b91b40ef-6207-4d32-b5eb-5667ebb674ec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faad1085f0195308df41e3be8232dd7ca73c420eb5f763e9dbc900c44ee82963 +size 2946655 diff --git a/2023/How to Choose your Best Allies for a Transferable Attack_/full.md b/2023/How to Choose your Best Allies for a Transferable Attack_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..12aeaab57150e470fa92f67e72d1c62ee641638a --- /dev/null +++ b/2023/How to Choose your Best Allies for a Transferable Attack_/full.md @@ -0,0 +1,338 @@ +# How to choose your best allies for a transferable attack? + +Thibault Maho* +Univ. Rennes, Inria, CNRS IRISA, Rennes, France +thibault.maho@inria.fr + +Seyed-Mohsen Moosavi-Dezfooli +Imperial College London, UK +seyed.moosavi@imperial.ac.uk + +Teddy Furon† +Univ. Rennes, Inria, CNRS IRISA, Rennes, France +teddy.furon@inria.fr + +# Abstract + +The transferability of adversarial examples is a key issue in the security of deep neural networks. The possibility of an adversarial example crafted for a source model fooling another targeted model makes the threat of adversarial attacks more realistic. Measuring transferability is a crucial problem, but the Attack Success Rate alone does not provide a sound evaluation. This paper proposes a new methodology for evaluating transferability by putting distortion in a central position. This new tool shows that transferable attacks may perform far worse than a black box attack if the attacker randomly picks the source model. To address this issue, we propose a new selection mechanism, called $F \vdash T$ , which aims at choosing the best source model with only a few preliminary queries to the target. Our experimental results show that $F \vdash T$ is highly effective at selecting the best source model for multiple scenarios such as single-model attacks, ensemble-model attacks and multiple attacks. + +# 1. Introduction + +Transferability is one of the most intriguing properties of adversarial examples. A white box attack crafting adversarial examples for an open-source model is likely to fool other models too [1, 14, 20, 25, 34]. This makes the threat of adversarial examples more realistic. In practice, the model targeted is usually unknown but accessible as a black box. This prevents directly applying any white box gradient-based attack [10, 16, 19, 35]. Black box attacks do exist but they require some thousands of queries to find an adversarial example of low distortion [4, 11, 17, 24]. Transferable attacks require no or few queries to fine-tune an adversarial example thanks to the help of a publicly available model similar enough to the target. + +Transferability is usually measured by the Attack Success Rate (ASR), i.e., the probability that the adversarial + +![](images/8c65d3a539418e84b1bd55dfff8da234f0f3adbfe25969e02c5ea9121e13af76.jpg) +Figure 1: Evaluation of transferability by comparing the Attack Success Rate vs. distortion trade-off of a white box, transferable, and black box attacks against model CoatLitesmall (See Sect. 4.1 for details). The blue area is the range of trade-off operated by a transferable attack with random source models. A transferable attack may be worse than a black box attack without a good source selection (like FiT). + +example crafted for the source model also deludes the target model. We argue that this measure leads to an unfair evaluation of transferability. In the context of adversarial examples, it is not just a matter of discovering data that is not well classified, but rather identifying the perturbation that can fool a classifier with minimal distortion. This principle should also apply to transferable attacks. + +For illustration purposes, let us consider two models, one is robust in the sense that the necessary amount of adversarial perturbation is large, whereas the other model is weak. If the attacker uses the robust model as the source to attack the weak target network, the ASR of the transferable attack will certainly be big. It does not mean that this is the right choice. The ASR is high because the robust source model needs large perturbation to be deluded, which will fool any weaker model. The ASR alone does not reflect the overshooting in distortion. The converse, using the weak to at + +tack the robust, would yield a low ASR. To summarize, the ASR alone fails to capture how relevant the direction of the perturbation given by the source is for attacking the target. + +The first contribution of this paper is to put distortion back into the picture. Section 3 evaluates transferability by comparing the distortion of a transferable attack to the ones of two reference attacks: On one hand, the strongest attack, i.e., the white box attack directly applied on the target model; on the other hand, the weakest attack, i.e., the black box attack. + +The second contribution shows the great variability of the performance of transferable attacks. Figure 1 summarizes this observation by plotting the ASR as a function of the distortion (the experimental protocol is explained in Sect. 4.1). Naturally, the black box attack needs much more distortion than the white box attack. For instance, the white box attack yields an ASR of $50\%$ with a distortion of 0.19, whereas the black box attack needs a distortion of 9.7. The surprise is that if the attacker resorts to a transferable attack and picks a source model at random, there is almost a $50\%$ chance that the attack performs even worse than the black box attack. Section 4 outlines a triad of factors: input, source model, and attack. + +This observation challenges the prevalent notion that adversarial examples transfer easily between models, and highlights the need to carefully choose the source model to attack a target. Under the assumption that the attacker has indeed several candidate models, our third contribution, named FiT in Sect. 5, provides an affordable measure for model selection, allowing the attacker to choose a good source model with only a few queries to the target. + +# 2. Related work + +# 2.1. White box, black box, and transferable attacks + +The threat analysis related to adversarial examples usually considers two scenarios: the white box assumption where the attacker knows the internals of the target model, and the black box assumption where he does not but has limited access to the target. + +White box attacks are now performing very well in terms of Attack Success Rate and distortion. We distinguish two kinds of attacks: 1) attacks constrained by a distortion budget like PGD [16] whose performance is measured by the ASR, and 2) attacks yielding almost surely an adversarial example like DeepFool [19] and CW [3] whose performance is measured by the distortion. The trend is to make them speed efficient as well with fewer computations of the gradient of the neural network, like FMN [23] and BP [35]. + +Our paper also considers so-called decision-based black box attacks like GeoDA [24], SurFree [17], and RayS [4], where the attacker only sees the class predicted by the model, but not the confidence score, neither the pro + +bit nor the logit. A decision-based black box attack typically needs thousands of queries to find an adversarial example with a relatively low distortion, yet, much higher than with a white box attack. This is the price the attacker has to pay when he does not know the target model. + +The transferable attacks, which pertain to the black box scenario, disrupt this last statement. The assumption is that attacker knows another model, called source, trained for the same classification problem as the target. Yet, leading a white box attack on the source model with the hope that this adversarial example will also fool the target yields poor results because the adversarial perturbation is too specific to the source. Modern transferable attacks increase the ASR by avoiding the overfitting of examples on the source. Input transformations have been purported by DI [32], TI [6] and Admix [27]. Papers [8, 13, 28] stabilized the gradient while [7, 29, 36, 39] focus on the importance of intermediate features. The combination of several methods has also been investigated [9, 40]. For instance, TAIG [9] uses input transformation, integrated gradients, and attention reduction. + +Ensemble-model attacks make the assumption that the attacker leverages not a single source but several source models. Paper [28] averages the logits of several sources and leads a white box attack on this aggregation of models, whereas [33] reduces model variance. The main obstacle to these methods is the computational complexity. To counteract this drawback, some methods design a specific ensemble of sources like ghost networks [12] or pruned network [26]. This amounts to creating several sources without the need to compute the gradient of each model. + +All this literature uses the ASR as the figure of merit for a given distortion while our paper draws the full operating characteristic of ASR vs. distortion. Moreover, the performance of the transferability is compared neither to the white box nor the decision-based black box attack. + +# 2.2. Fingerprinting + +Models are valuable assets because of their training costs, be it the expertise, the annotated data, or the computational power. Fingerprinting methods have been proposed to protect this intellectual property. They provide a similarity score between the two models by looking at similar decisions in submitted queries. They often use adversarial examples specifically designed to be unique for a model. The models are considered similar if they share the same decision for these inputs. IPGuard [2] creates adversarial examples close to the decision boundary to frame it. Universal adversarial perturbations are used to characterize the boundary in [21]. The paper [15] built adversarial examples with a transferable property called conferrable adversarial examples. AFA [37] crafts adversarial examples model-specific using dropout. FBI [18] is the only method using unmodified images. It estimates the mutual information to measure + +the statistical dependence of two models' predictions. + +Our paper investigates the idea that a model similar to the target may be a good source for transferable attacks. + +# 3. Methodology + +This paper introduces a new way to gauge transferability. The main motivation is that transferability should indicate whether a source $s$ is useful for attacking a target $t$ , independently from the inherent robustness of $t$ . + +# 3.1. Notations + +Let $f:[0,255]^N \to \mathcal{R}^K$ denote a classifier computing the predicted probabilities $f(x)$ of $K$ classes for input $x$ . For a given input $x$ of class $y$ , the adversarial example $x_a$ is the result of an attack (be it white box, black box, or by transferability) such that + +$$ +\arg \max _ {1 \leq k \leq K} f _ {k} (x _ {a}) \neq \arg \max _ {1 \leq k \leq K} f _ {k} (x) = y. \tag {1} +$$ + +In the case of transferable attacks, the attacker uses a source model $f_{s}$ to build a transferable adversarial example against the target model $f_{t}$ . This paper considers a collection of $m$ source models denoted by $\mathcal{F}_s = \{f_s^1,\dots,f_s^m\}$ , and a set of $n$ inputs $\mathcal{X}$ . + +# 3.2. Measurement + +Distortion. We measure the distortion of an adversarial perturbation using the Euclidean norm which is a common choice as we deal with natural images: + +$$ +\operatorname {d i s t} \left(x _ {a}, x\right) := \| x _ {a} - x \| _ {2} / \sqrt {N}, \tag {2} +$$ + +with $N = 3\times H\times L$ pixels. This distortion can be seen as the average amplitude of the perturbation per pixel. + +Operating characteristic. Given an attack (be it white box, black box, or by transferability) and a target model $f_{t}$ , we call operating characteristic the following function: + +$$ +P (D) := \mathbb {P} (\operatorname {d i s t} \left(x _ {a}, x\right) < D). \tag {3} +$$ + +It is thus the Attack Success Rate as a function of the distortion. Contrary to the literature which measures the ASR for a few distortion levels, we consider the full range of $D$ values. + +# 3.3. Transferability + +Given a white box and a black box attacks, The proposed methodology first computes the operating characteristic $P_{t}^{\mathrm{wb}}(D)$ of a state-of-the-art white box attack directly applied to the target model, and the operating characteristic $P_{t}^{\mathrm{bb}}(D)$ of a state-of-the-art black box attack. It then measures where the operating characteristic of the transferable + +attack from the source $s$ to the target $t$ lies in between the two characteristics as follows: + +$$ +T _ {s, t} := \frac {\int_ {0} ^ {\infty} P _ {s , t} (u) - P _ {t} ^ {\mathrm {b b}} (u) d u}{\int_ {0} ^ {\infty} P _ {t} ^ {\mathrm {w b}} (u) - P _ {t} ^ {\mathrm {b b}} (u) d u}. \tag {4} +$$ + +The proposed score is calculated as the ratio of the areas between the different operational characteristics, which are defined as Cumulative Distribution Functions (CDFs). The numerator is therefore close to the 1-Wasserstein distance [5] between the ASR function of the distortions of the transferable attack and that of the black box attack, where the absolute value is removed to obtain a signed score. If the transferable attack performs as well as the white box attack (resp. as bad as the black box attack), then $T_{s,t} = 1$ (resp. $T_{s,t} = 0$ ). Figure 1 shows that the numerator of (4) can indeed be negative as transferability can be even worse than the black box if the source $s$ is not well chosen. Therefore, zero is not a lower bound for $T_{s,t}$ . In the denominator, the 1-Wasserstein distance is obtained exactly between the white box and the black box because the white box consistently generates adversarial examples with lower distortion. + +# 3.4. Practical implementation + +Attacks. We only consider state-of-the-art attacks which almost surely deliver an adversarial example. As said in Sect. 2.1, this is usually the case for black box attacks, but we are limited to using white box attacks that are not subject to a distortion budget. As for transferability, the attack uses the publicly available model $s$ and input $x$ to craft an adversarial direction $u_{x,s}$ of Euclidean norm $\sqrt{N}$ . The use of $\sqrt{N}$ for the norm is for the sake of simplicity in notation. We assume that there is an oracle giving the minimal distortion along this direction to fool the target $t$ . In other words, $x_{a} = x + d u_{x,s}$ , with + +$$ +d = \min \left\{\delta : \arg \max _ {k} f _ {t, k} (x + \delta u _ {x, s}) \neq y \right\}. \tag {5} +$$ + +Note that $\mathrm{dist}(x_a,x) = d$ . This definition favours transferability as its best. In practice, such an oracle does not exist but the attacker finds a good estimate of (5) thanks to a line search within a few queries to the target. + +Transferability. We run a given attack over a collection $\mathcal{X}$ of $n$ inputs correctly classified by the target model. We compute the distortions $d(j) = \mathrm{dist}(x_{a,j}, x_j)$ with $x_j \in \mathcal{X}$ and sort them so that $d(1) \leq d(2) \leq \ldots \leq d(n)$ . We set $d(0) = 0$ and $d(n + 1) = \infty$ to properly define the empirical operating characteristic by the following step function: + +$$ +\hat {P} (D) := j / n \quad \forall D \in [ d (j), d (j + 1)). \tag {6} +$$ + +It is then easier to estimate the integrals appearing in (4), which are indeed areas under two curves, by a Lebesgue + +sum rather than a Riemann sum. This gives: + +$$ +\hat {T} _ {s, t} := \frac {\sum_ {j = 1} ^ {n} d _ {s , t} (j) - d _ {t} ^ {\mathrm {b b}} (j)}{\sum_ {j = 1} ^ {n} d _ {t} ^ {\mathrm {w b}} (j) - d _ {t} ^ {\mathrm {b b}} (j)}, \tag {7} +$$ + +where the distortions $(d_t^{\mathrm{bb}}(j))_j$ (resp. $(d_t^{\mathrm{wb}}(j))_j$ ) resulting from the black box (resp. white box) attack against model $t$ are also sorted in increasing order. + +# 4. Triad of transferability: data, model, attack + +This section is an experimental investigation of the factors affecting transferability. + +# 4.1. Experimental setup + +The study assesses transferable attacks on a total of 48 models, with 47 of them sourced from the Timm library [30] and one very robust, namely ResNet50AdvTrain, obtained from the GitHub repository 1. Appendix A lists all the models. The experiments utilize 100 images from the validation set of ILSVRC'12, all of which are correctly classified by all models under consideration. + +The study considers three transferable attacks - DI [32], TAIG [9], and DWP [26]- each using a different approach to improve transferability (see Sect. 2.1). These attacks are selected as the best in their categories in [38]. They all share an $\epsilon$ parameter to control the maximum perturbation added per pixel. The effect of the parameter $\epsilon$ indeed happens to be negligible in our protocol, as demonstrated in Appendix B. We choose $\epsilon = 8$ . + +To measure transferability (7), we need state-of-the-art black box and white box attacks. Certain methods may exhibit a preference towards one model over the other, necessitating the use of multiple attacks. Our study employs four white box attacks (BP [35], DeepFool [19], I-FGSM [10], and PGD [16]) and three black box attacks (SurFree [17], RayS [4], and GeoDA [24]). All black box attacks are run with 2,000 queries, which has been determined to be sufficient for achieving convergence. For distortion-constrained white box attack, the $\epsilon$ parameter is set to 4. We record the smallest perturbation distortion over the black box (white box) for each image and draw the operating characteristic (6) as appearing in green (resp. red) dashed line in Fig. 1. + +As for the transferable attack, for a given target model, the attacker has access to a subset of all the other models whose architecture differs from the target. This amounts to an average of 45 models out of 48. For instance, in Fig. 1, CoatLitesmall being the target, we exclude all other CoatLite models from being a source. The light blue area delimits the operating characteristics of transferable attack DI [32] using as the source one of the 45 remaining models. + +# 4.2. Model dependence + +![](images/79c1ed355cd42f1ac9c7073c690859e106db534d9f137ba089de4233d6566483.jpg) +Figure 2: Transferability score $\hat{T}_{s,t}$ matrix of 48 sources and 48 targets listed in App. A with attack DI [32]. + +Large transferability variation. Figure 2 shows the matrix $\hat{T}_{s,t}$ , where $s$ and $t$ are any of the $48^2$ pairs, for attack D I [32]. Not all models possess the same transferability capabilities. At first look, the figure contains more blue than red cells which means that transferability takes negative value more often. + +Some rare models, like PiTsmall-dist, exhibit good transferability towards any target. On the contrary, DPN68b is always a bad source. On the other hand, ResNet50AdvTrain is a very difficult target (note how + +![](images/5300a085f9c1a7cac86c7950c1f852ab2c7d4e0cb7bd6d63f57a2a43d5fa5ce6.jpg) +Figure 3: $\hat{T}_{s,t}$ function of the number of available sources and attack DI [32]. The best or worst model selected per image (solid line) or on average (dashed line). + +![](images/22747ae36c47dbdbf308284fcae58d936c22d3ad31956ac7f846a275bdc86dda.jpg) +Figure 4: Distribution of best image-model selection function of the number of sources. Adversarial examples obtained with DI [32]. + +ever that its accuracy is low), followed by PiTlight and ConViTtiny, whereas the family of ResNet are fairly easy target. + +Prior works have shown that models with architectures similar to the target transfer better [22, 31]. The red squares appearing in the matrix confirm this (models are ordered according to architecture family). For instance, MobileNetV2 and RexNet architectures exhibit transferability close to 0.7. However, this is not an absolute truth. For example, EfficientNet-B0 has better transferability against MixNetlarge than MixNetmedium, even though they have similar architectures. ConViT transfers exceptionally well to Twins although their architectures are very different. Similar observations hold for the other attacks but with lower transferability scores (see Appendix C). + +From now on, we exclude models whose architecture is similar to the target. + +Impact on the attack. Selecting the right source among several available models is critical for achieving high transferability. Figure 3 displays the transferability score (7) for the best and the worst choices (shown in dotted lines) as the number of source candidates increases. In this evaluation, a single model is selected as the source for building all adversarial examples. On average, the transferability is lower than zero, regardless of the attack method. This means that transferable attacks perform worst than black box attacks on average. If the attacker knows how to select the best source model, $\hat{T}_{s,t}$ quickly converges to a maximum which is positive but below 0.5. This means that transferability at its best performs closer to a black box attack rather than a white box attack. These remarks highly mitigate the threat of transferability and put emphasis on the crucial selection of the best source model. As far as we know, these facts are not reported in the adversarial examples literature. + +![](images/8c36484bfba84df3025f192834fb37ff19e02a31036c7cac698e3f45655215c5.jpg) + +![](images/5876eed1cf796cf9d0972e5f086a19cad7abdd427e716d816a800fa985dcf0f6.jpg) +Figure 5: 2D Histogram of the minimum distortion for transferable and white box attacks. + +![](images/e11cac767e0f0bae8379a1e7726d996ea3ee6678f6ad18e77b83ad17ab42d642.jpg) + +![](images/01e31915b18a074af56f26cba78f58fdc470bcdd15c6dc7ac4887ab8f20b5a49.jpg) + +# 4.3. Image dependence + +The difficulty of transferring adversarial examples from a source to a target varies significantly depending on the input $x$ . This is illustrated in Fig. 4, which shows the distribution of the best-performing model for each image based on the available sources. Even with a large number of sources available, there is typically one source that performs significantly better than the others, but only over $60\%$ of the images on average. However, this superiority decreases rapidly as the number of available sources increases. It means that a better source exists for $40\%$ of the images on average. + +Supposing that the attacker knows the best source of each input, Figure 3 shows in solid line that the transferability converges to a value close to 0.5. The performance of the transferable attack lies in between the ones of the white box and black box attacks. This highlights the importance of selecting an appropriate source model for a given target and input. + +# 4.4. Attack dependence + +Transferable attack. Figure 5 compares transferable attacks with a 2D histogram of the distortion pair for two attacks. This is computed over all inputs in $\mathcal{X}$ and all pairs of source and target models. The results show that DWP [26] exhibits poor transferability compared to DI [32] and TAIG [9], which produce adversarial perturbations with similar distortion. Additionally, regardless of the attack complexity or method, the challenging images remain consistent. If one attack requires a high distortion for a given input / source / target, other attacks are likely to encounter + +similar difficulties. + +Traditional white box attack. We now compare the methods designed for transferability with a naive approach. A white box attack (ie. not specific to transferability) is executed on the source, and then the direction found serves as $u_{x,s}$ to place the adversarial examples on the target boundary with (5). The best $\hat{T}_{s,t}$ score is 0.27 for BP [35] compared to 0.52 for DI [32]. This confirms the superiority of the recent methods designed for transferability. However, even though DI [32] is on average better, Figure 5 shows that when the necessary distortion is small, traditional white box attacks like DeepFool [19] and BP [35] indeed beats DI [32]. On the other hand, DI [32] performs much better for inputs requiring more distortion. + +# 5. How to choose the best source + +Section 4 outlines that the choice of the source is of utmost importance thanks to the transferability measure defined in Sect. 3. This section investigates whether the attacker can guess which model is the best source. + +We now propose a procedure which combines the dependences with respect to the source and target models (Sect. 4.2), and the input (Sect. 4.3). Model dependence is measured by evaluating the similarity between the source and target models, which is denoted as ModSim. On the other hand, the image dependence is measured by the quality of the adversarial example, denoted as TransQ. Both metrics are combined into the following score: + +$$ +\operatorname {F i T} (s, t, x) := \operatorname {M o d S i m} (s, t) \times \operatorname {T r a n s Q} (s, x). \tag {8} +$$ + +These indicators should be easy to compute. We especially pay attention to the number of queries to the target. This score opens the door to a new strategy for the attacker which first selects the best source among the available models + +$$ +s ^ {\star} (t, x) = \underset {\sigma \in \mathcal {F} _ {s}} {\arg \max } \mathrm {F i T} (\sigma , t, x), \tag {9} +$$ + +and then crafts the adversarial direction $u_{x,s^*(t,x)}$ . + +# 5.1. Criterion ModSim(s,t) + +Section 4.2 highlights the correlation between the transferability and the similarity between the source and target models. Gauging model similarity has been previously studied in the context of fingerprinting as a defense to protect intellectual property (see Sect. 2.2). This paper uses fingerprinting methods as an attack that leaks information about the target. + +We consider the fingerprinting method [18] because it works in a decision-based setup since the target is a black box in our application. Querying two models $s$ and $t$ with few natural images, it computes a distance $\operatorname{Dist}(s, t) \in [0, 1]$ by comparing their outputs. Since we look for a similarity, we set $\operatorname{ModSim}(s, t) = 1 - \operatorname{Dist}(s, t)$ . However, + +this provides a symmetrical similarity, ie. $\operatorname{ModSim}(s,t) = \operatorname{ModSim}(t,s)$ , while transferability is not (see Fig. 2). This shows that this criterion alone is not sufficient. + +# 5.2. Criterion TransQ(s, x) + +This criterion evaluates the general transferability of a given adversarial example crafted by a source. Our idea is to leverage the assumption that the attacker has a set of models $\mathcal{F}_s$ . Consequently, we can evaluate the transferability thanks to the other models of this set, without querying the target. + +For a given input $x$ , the source $s$ provides the adversarial direction $u_{x,s}$ and we compute the distortion $d_{s,\sigma}$ necessary to delude classifier $\sigma \in \mathcal{F}_s$ with (5). We then aggregate these distortions into a single score with two flavours: + +$$ +\operatorname {T r a n s Q} ^ {(1)} (s, x) := \left(\frac {1}{| \mathcal {F} _ {s} |} \sum_ {\sigma \in \mathcal {F} _ {s}} d _ {s, \sigma}\right) ^ {- 1}. \tag {10} +$$ + +A good source for input $x$ gives birth to lower distortions, so that $\mathrm{TransQ}^{(1)}(s,x)$ is large. + +$$ +\operatorname {T r a n s} Q ^ {(2)} (s, x) := \frac {\sum_ {\sigma \in \mathcal {F} _ {s}} d _ {s , \sigma} - d _ {\sigma} ^ {\mathrm {b b}}}{\sum_ {\sigma \in \mathcal {F} _ {s}} d _ {\sigma} ^ {\mathrm {w b}} - d _ {\sigma} ^ {\mathrm {b b}}}. \tag {11} +$$ + +This measure is similar to (7) except that it is computed over the set of models instead of a set of inputs. + +# 5.3. Results + +The experimental setup is the same as in Sect. 4.1. We select the fingerprinting method FBI [18] with 200 benign natural images to compute the criterion ModSim(s, t). It implies that the attacker first makes 200 queries to the target in a preliminary step before forging any adversarial example. Appendix D.1 shows that more images improve the fingerprinting accuracy hence the model selection, but the results converge after a number of 200 images. + +This section is structured as follows: we use FiT to select the best model for single-model attacks, then we employ it to identify the best subset of models for ensemble-model attacks and different combinations of attacks. + +# 5.3.1 Single-model attacks + +Criterion ModSim(s, t). Table 1 indicates that architectural similarity is a reliable measure of transferability between two models. It can drive the selection of a good source giving birth to a transferable attack outperforming the black box attack since $\hat{T}_{s,t}$ is larger than 0 (except for DWP [26]). Yet, the results remain low compared to the best results obtained. As discussed in Sec 4.2, similarity may not suffice because it implies the selection of one unique for a given target. Better results are achieved when adapting the source to the input. + +Table 1: Transferability ${\widehat{T}}_{s,t}$ for DI [32],TAIG [9] and DWP [26] for single and ensemble-model attacks. + +
CategorySelection MethodDI [32]TAIG [9]DWP [26]
Single-model attackBest0.52±0.120.46±0.120.34±0.13
Random-0.16±0.26-0.12±0.21-0.72±0.32
ModSimFBI [18]0.18±0.170.12±0.18-0.39±0.19
TransQASR-0.21±0.39-0.24±0.30-1.14±0.95
TransQ(1)(10)0.38±0.150.24±0.160.10±0.20
TransQ(2)(11)0.37±0.180.23±0.180.08±0.22
FiTTransQ(1)(10)0.40±0.130.27±0.170.12±0.20
TransQ(2)(11)0.39±0.160.25±0.180.10±0.22
Ensemble-model attack with three sourcesBest0.72±0.080.62±0.080.46±0.1
Random0.43±0.120.40±0.110.02±0.17
ModSimFBI [18]0.59±0.100.49±0.100.05±0.11
TransQASR0.53±0.210.45±0.270.06±0.27
TransQ(1)(10)0.62±0.110.54±0.130.33±0.13
TransQ(2)(11)0.61±0.110.54±0.120.33±0.17
FiTTransQ(1)(10)0.64±0.110.55±0.140.35±0.15
TransQ(2)(11)0.64±0.100.57±0.140.36±0.15
+ +![](images/c63842bc1ec68e5899d631f05e4a9fd010da1807ee7e51175712e54e5aa38577.jpg) +Figure 6: $\hat{T}_{s,t}$ as a function of the number of available sources for several selection methods. Dotted lines refer to the selection of a unique model for all images and solid lines refer to a model selected per image. Attack is DI [32]. + +Criterion TransQ(s,x). Improving transferability without querying the target model in a preliminary step is possible thanks to TransQ(s,x). Adversarial examples that exhibit good transferability on multiple models are more likely to also deceive the unknown targeted model. Figure 6 shows that a significant improvement in transferability is achieved even with only a few available models. Table 1 confirms this observation for the two other attack methods. This strategy is indeed better than the selection based on model similarity. + +Score FiT (s,t,x). Combining both criteria together as in (8) leads to a slight improvement in transferability com + +pared to $\mathrm{TransQ}(s,x)$ alone. Fig. 6 confirms this holds over a wide range of numbers of available sources. For single-model attacks, $\mathrm{TransQ}^{(1)}$ gives slightly better results than $\mathrm{TransQ}^{(2)}$ . + +Visual results Figure 2 visually demonstrates the impact of different model selection methods on the quality of adversarial examples. Even when FiT is not accurate, the resulting adversarial examples are still close to the best ones obtained from the source models, and the perturbation remains imperceptible. However, random selection generates noisy perturbations, and the worst-case scenario destroys the image entirely. + +# 5.3.2 Ensemble-model attacks + +Importance of model selection in ensemble-model attack. Ensemble attacks remain an under-studied area due to the significant computational resources required for evaluating attacks. Consequently, these attacks have only been evaluated with a limited number of sources. Appendix D.2 shows that increasing the number of sources is not necessarily beneficial. The FiT score provides a scalable solution: we select a small subset of sources based on their FiT scores (top-3) from the bigger set of available sources. In a way, it is better to put quality above quantity. For example, when running an ensemble-model attack against $\times \mathrm{CiT}_{\mathrm{nano}}$ selecting only the three best models using the FiT measure can lead to a significant improvement in transferability compared to using a larger set of models. In our experiments, an ensemble-model of 20 random models achieved a $\hat{T}_{s,t}$ of 0.47, while an ensemble-model of only three models se + +Table 2: Visual impact of the source selection with DI [32] attacking ConViTbase (first row) and DPN92 (second row). + +
OriginalBestFiT(TransQ(1))RandomWorst
label source distortionmouseoil filter PiTsmall-dist 5.62purse ConViTbase 8.13purse MobileNetV2110d 35.9jigsaw puzzle DenseNet121 87.8
label source distortiontramelec. locomotive ResNetV250x1-dist 6.94elec. locomotive PiTsmall-dist 7.50racing car DLA60 33.3jigsaw puzzle MixNetmedium 128.7
+ +lected with $\mathbb{F}\mathbb{i}\mathbb{T}$ was able to achieve a $\hat{T}_{s,t}$ of 0.56 against the same target. + +Performance of ensemble-model attacks. Table. 1 shows that ensemble-model transferability surpasses that of single-model attacks. While the average results over a random selection of 3 sources increase, they remain closer to the black box results than the white box ones (transferability lower than 0.5). Choosing the top-3 sources returned by our scores for leading the ensemble-model attack yields better performance, comparable to the best possible results obtainable with ensemble-model attacks. Notably, for the DI [32] and TAIG [9] method, it approaches the performance of white box results (transferability greater than 0.5). + +Table 3: Transferability $\hat{T}_{s,t}$ for white box and transferable attacks (single-model). The FiT score uses $\mathrm{TransQ}^{(1)}$ . + +
AttackBestFiT
W.B.BP [35]0.27 ±0.18-0.06±0.31
DeepFool [19]0.03 ±0.20-0.34±0.34
PGD [16]0.29 ±0.20-0.01±0.36
I-FGSM [10]0.29 ±0.20-0.03±0.34
Trans.DI [32]0.52 ±0.120.40±0.13
TAIG [9]0.46 ±0.120.27±0.17
DWP [26]0.34 ±0.130.12±0.20
All Attacks0.65 ±0.090.48±0.14
+ +# 5.3.3 White box vs. transferable attacks + +Our last result is that the selection of a single source with the FiT score does not make traditional white box attacks transferable. Table 3 shows that BP [35], DeepFool [19], PGD [16], and I-FGSM [10] yield negative transferability values. More precisely, BP [35], I-FGSM [10], and PGD [16] perform better with FiT than the transferable attacks without any selection mechanism (ie. on average with a random source). However, section 4.4 highlights that the traditional white box attacks may be competitive for some inputs demanding low adversarial perturbation distortion. It is valuable to add them to the options of the attacker and let the FiT score decide the preferable option (source and attack method). This provides our best transferability score for a single model attack, close to 0.5. + +# 6. Conclusion + +Transferability is a crucial feature of adversarial examples as it allows a single perturbation to deceive multiple models. However, solely relying on Attack Success Rate (ASR) to measure transferability overlooks the degree of distortion needed to fool a model. This paper introduces a novel approach to assess transferability by comparing it to the distortion of two reference attacks: white box and black box attacks. We show that transferable attacks can perform worse than black box attacks without an appropriate selection of the source model, highlighting the need to choose the best source model to target a specific model. + +The proposed solution, named F i T, allows the attacker to choose one of the best source models with minimal + +queries to the target. Our experiments demonstrate that the proposed solution performs well in multiple attack scenarios. + +This study has highlighted the differences in transferability between images for the same source model and their particularity for this specific network. Further research could focus on addressing this issue and investigating its underlying causes, with the hope of designing an even better selection mechanism able to spot the best source model. + +# References + +[1] Mauro Barni, Kassem Kallas, Ehsan Nowroozi, and Benedetta Tondi. On the transferability of adversarial examples against CNN-based image forensics. In ICASSP. IEEE, 2019. +[2] Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Ipguard: Protecting intellectual property of deep neural networks via fingerprinting the classification boundary. In ACM Asia Conference on Computer and Communications Security, 2021. +[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In Symposium on Security and Privacy. IEEE, 2017. +[4] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In ACM SIGKDD, 2020. +[5] Marco De Angelis and Ander Gray. Why the 1-wasserstein distance is the area between the two marginal cdfs, 2021. +[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. +[7] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. +[8] Tianjin Huang, Vlado Menkovski, Yulong Pei, Yuhao Wang, and Mykola Pechenizkiy. Direction-aggregated attack for transferable adversarial examples. JETC, 2022. +[9] Yi Huang and Adams Wai-Kin Kong. Transferable adversarial attack based on integrated gradients. In ICLR, 2022. +[10] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018. +[11] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In CVPR, 2020. +[12] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. +[13] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. +[14] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. + +[15] Nils Lukas, Yuxuan Zhang, and Florian Kerschbaum. Deep neural network fingerprinting by conferrable adversarial examples. In ICLR, 2021. +[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2017. +[17] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Surfree: a fast surrogate-free black-box attack. In CVPR, 2021. +[18] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Fbi: Fingerprinting models with benign inputs. arXiv preprint arXiv:2208.03169, 2022. +[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016. +[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In ICLR, 2022. +[21] Zirui Peng, Shaofeng Li, Guoxing Chen, Cheng Zhang, Haojin Zhu, and Minhui Xue. Fingerprinting deep neural networks globally via universal adversarial perturbations, 2022. +[22] Deyan Petrov and Timothy M Hospedales. Measuring the transferability of adversarial examples. arXiv preprint arXiv:1907.06291, 2019. +[23] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. In NeurIPS, 2021. +[24] Ali Rahmati, Seyed-Mohsen Moosavi-Dezfooli, Pascal Frossard, and Huaiyu Dai. Geoda: a geometric framework for black-box adversarial attacks. In CVPR, 2020. +[25] Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017. +[26] Hung-Jui Wang, Yu-Yu Wu, and Shang-Tse Chen. Enhancing targeted attack transferability via diversified weight pruning. arXiv preprint arXiv:2208.08677, 2022. +[27] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. +[28] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting adversarial transferability through enhanced momentum. In BMVC, 2021. +[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021. +[30] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019. +[31] Lei Wu and Zhanxing Zhu. Towards understanding and improving the transferability of adversarial examples in deep neural networks. In ACML, pages 837-850. PMLR, 2020. +[32] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. +[33] Yifeng Xiong, Jiadong Lin, Min Zhang, John E Hopcroft, and Kun He. Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In CVPR, 2022. + +[34] Liping Yuan, Xiaqing Zheng, Yi Zhou, Cho-Jui Hsieh, and Kai-Wei Chang. On the transferability of adversarial attacks against neural text classifier. In EMNLP, 2021. +[35] Hanwei Zhang, Yannis Avrithis, Teddy Furon, and Laurent Amsaleg. Walking on the edge: Fast, low-distortion adversarial examples. IEEE Trans. on IFS, 2020. +[36] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022. +[37] Jingjing Zhao, Qingyue Hu, Gaoyang Liu, Xiaoqiang Ma, Fei Chen, and Mohammad Hassan. AFA: Adversarial Fingerprinting Authentication for deep neural networks. Computer Communications, 2020. +[38] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, and Michael Backes. Towards good practices in evaluating transfer adversarial attacks. arXiv preprint arXiv:2211.09565, 2022. +[39] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018. +[40] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020. \ No newline at end of file diff --git a/2023/How to Choose your Best Allies for a Transferable Attack_/images.zip b/2023/How to Choose your Best Allies for a Transferable Attack_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7da95a82e0703447d1ffa13296fe9d21f19fd87 --- /dev/null +++ b/2023/How to Choose your Best Allies for a Transferable Attack_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:182316713426f47454d2474b09763a6a55489c296d492a65161d206074eed528 +size 511661 diff --git a/2023/How to Choose your Best Allies for a Transferable Attack_/layout.json b/2023/How to Choose your Best Allies for a Transferable Attack_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..be7ec070dc870795fbb7c9c3d695d65fee6ceae4 --- /dev/null +++ b/2023/How to Choose your Best Allies for a Transferable Attack_/layout.json @@ -0,0 +1,8502 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 470, + 121 + ], + "type": "text", + "content": "How to choose your best allies for a transferable attack?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 143, + 194, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 143, + 194, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 143, + 194, + 199 + ], + "type": "text", + "content": "Thibault Maho* \nUniv. Rennes, Inria, CNRS IRISA, Rennes, France \nthibault.maho@inria.fr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 143, + 376, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 143, + 376, + 185 + ], + "spans": [ + { + "bbox": [ + 214, + 143, + 376, + 185 + ], + "type": "text", + "content": "Seyed-Mohsen Moosavi-Dezfooli \nImperial College London, UK \nseyed.moosavi@imperial.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 401, + 143, + 534, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 143, + 534, + 199 + ], + "spans": [ + { + "bbox": [ + 401, + 143, + 534, + 199 + ], + "type": "text", + "content": "Teddy Furon† \nUniv. Rennes, Inria, CNRS IRISA, Rennes, France \nteddy.furon@inria.fr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "content": "The transferability of adversarial examples is a key issue in the security of deep neural networks. The possibility of an adversarial example crafted for a source model fooling another targeted model makes the threat of adversarial attacks more realistic. Measuring transferability is a crucial problem, but the Attack Success Rate alone does not provide a sound evaluation. This paper proposes a new methodology for evaluating transferability by putting distortion in a central position. This new tool shows that transferable attacks may perform far worse than a black box attack if the attacker randomly picks the source model. To address this issue, we propose a new selection mechanism, called " + }, + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "inline_equation", + "content": "F \\vdash T" + }, + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "content": ", which aims at choosing the best source model with only a few preliminary queries to the target. Our experimental results show that " + }, + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "inline_equation", + "content": "F \\vdash T" + }, + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "content": " is highly effective at selecting the best source model for multiple scenarios such as single-model attacks, ensemble-model attacks and multiple attacks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 481, + 128, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 128, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 128, + 494 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 502, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 658 + ], + "type": "text", + "content": "Transferability is one of the most intriguing properties of adversarial examples. A white box attack crafting adversarial examples for an open-source model is likely to fool other models too [1, 14, 20, 25, 34]. This makes the threat of adversarial examples more realistic. In practice, the model targeted is usually unknown but accessible as a black box. This prevents directly applying any white box gradient-based attack [10, 16, 19, 35]. Black box attacks do exist but they require some thousands of queries to find an adversarial example of low distortion [4, 11, 17, 24]. Transferable attacks require no or few queries to fine-tune an adversarial example thanks to the help of a publicly available model similar enough to the target." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "content": "Transferability is usually measured by the Attack Success Rate (ASR), i.e., the probability that the adversarial" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 228, + 531, + 373 + ], + "blocks": [ + { + "bbox": [ + 320, + 228, + 531, + 373 + ], + "lines": [ + { + "bbox": [ + 320, + 228, + 531, + 373 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 531, + 373 + ], + "type": "image", + "image_path": "8c65d3a539418e84b1bd55dfff8da234f0f3adbfe25969e02c5ea9121e13af76.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 384, + 547, + 479 + ], + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 479 + ], + "type": "text", + "content": "Figure 1: Evaluation of transferability by comparing the Attack Success Rate vs. distortion trade-off of a white box, transferable, and black box attacks against model CoatLitesmall (See Sect. 4.1 for details). The blue area is the range of trade-off operated by a transferable attack with random source models. A transferable attack may be worse than a black box attack without a good source selection (like FiT)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 506, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 591 + ], + "type": "text", + "content": "example crafted for the source model also deludes the target model. We argue that this measure leads to an unfair evaluation of transferability. In the context of adversarial examples, it is not just a matter of discovering data that is not well classified, but rather identifying the perturbation that can fool a classifier with minimal distortion. This principle should also apply to transferable attacks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "For illustration purposes, let us consider two models, one is robust in the sense that the necessary amount of adversarial perturbation is large, whereas the other model is weak. If the attacker uses the robust model as the source to attack the weak target network, the ASR of the transferable attack will certainly be big. It does not mean that this is the right choice. The ASR is high because the robust source model needs large perturbation to be deluded, which will fool any weaker model. The ASR alone does not reflect the overshooting in distortion. The converse, using the weak to at" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 286, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 286, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 286, + 703 + ], + "type": "text", + "content": "*Thanks to Rennes Métropole for its funding for international mobility." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 703, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 59, + 703, + 286, + 712 + ], + "type": "text", + "content": "Thanks to ANR and AID french agencies for funding Chaire SAIDA." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4542" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": "tack the robust, would yield a low ASR. To summarize, the ASR alone fails to capture how relevant the direction of the perturbation given by the source is for attacking the target." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 191 + ], + "type": "text", + "content": "The first contribution of this paper is to put distortion back into the picture. Section 3 evaluates transferability by comparing the distortion of a transferable attack to the ones of two reference attacks: On one hand, the strongest attack, i.e., the white box attack directly applied on the target model; on the other hand, the weakest attack, i.e., the black box attack." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 286, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 286, + 348 + ], + "type": "text", + "content": "The second contribution shows the great variability of the performance of transferable attacks. Figure 1 summarizes this observation by plotting the ASR as a function of the distortion (the experimental protocol is explained in Sect. 4.1). Naturally, the black box attack needs much more distortion than the white box attack. For instance, the white box attack yields an ASR of " + }, + { + "bbox": [ + 46, + 193, + 286, + 348 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 193, + 286, + 348 + ], + "type": "text", + "content": " with a distortion of 0.19, whereas the black box attack needs a distortion of 9.7. The surprise is that if the attacker resorts to a transferable attack and picks a source model at random, there is almost a " + }, + { + "bbox": [ + 46, + 193, + 286, + 348 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 193, + 286, + 348 + ], + "type": "text", + "content": " chance that the attack performs even worse than the black box attack. Section 4 outlines a triad of factors: input, source model, and attack." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 445 + ], + "type": "text", + "content": "This observation challenges the prevalent notion that adversarial examples transfer easily between models, and highlights the need to carefully choose the source model to attack a target. Under the assumption that the attacker has indeed several candidate models, our third contribution, named FiT in Sect. 5, provides an affordable measure for model selection, allowing the attacker to choose a good source model with only a few queries to the target." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 457, + 130, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 130, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 130, + 470 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 478, + 286, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 286, + 490 + ], + "type": "text", + "content": "2.1. White box, black box, and transferable attacks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 497, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 286, + 556 + ], + "type": "text", + "content": "The threat analysis related to adversarial examples usually considers two scenarios: the white box assumption where the attacker knows the internals of the target model, and the black box assumption where he does not but has limited access to the target." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 665 + ], + "type": "text", + "content": "White box attacks are now performing very well in terms of Attack Success Rate and distortion. We distinguish two kinds of attacks: 1) attacks constrained by a distortion budget like PGD [16] whose performance is measured by the ASR, and 2) attacks yielding almost surely an adversarial example like DeepFool [19] and CW [3] whose performance is measured by the distortion. The trend is to make them speed efficient as well with fewer computations of the gradient of the neural network, like FMN [23] and BP [35]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Our paper also considers so-called decision-based black box attacks like GeoDA [24], SurFree [17], and RayS [4], where the attacker only sees the class predicted by the model, but not the confidence score, neither the pro" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "bit nor the logit. A decision-based black box attack typically needs thousands of queries to find an adversarial example with a relatively low distortion, yet, much higher than with a white box attack. This is the price the attacker has to pay when he does not know the target model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 133, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 312 + ], + "type": "text", + "content": "The transferable attacks, which pertain to the black box scenario, disrupt this last statement. The assumption is that attacker knows another model, called source, trained for the same classification problem as the target. Yet, leading a white box attack on the source model with the hope that this adversarial example will also fool the target yields poor results because the adversarial perturbation is too specific to the source. Modern transferable attacks increase the ASR by avoiding the overfitting of examples on the source. Input transformations have been purported by DI [32], TI [6] and Admix [27]. Papers [8, 13, 28] stabilized the gradient while [7, 29, 36, 39] focus on the importance of intermediate features. The combination of several methods has also been investigated [9, 40]. For instance, TAIG [9] uses input transformation, integrated gradients, and attention reduction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 313, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 545, + 432 + ], + "type": "text", + "content": "Ensemble-model attacks make the assumption that the attacker leverages not a single source but several source models. Paper [28] averages the logits of several sources and leads a white box attack on this aggregation of models, whereas [33] reduces model variance. The main obstacle to these methods is the computational complexity. To counteract this drawback, some methods design a specific ensemble of sources like ghost networks [12] or pruned network [26]. This amounts to creating several sources without the need to compute the gradient of each model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 434, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 492 + ], + "type": "text", + "content": "All this literature uses the ASR as the figure of merit for a given distortion while our paper draws the full operating characteristic of ASR vs. distortion. Moreover, the performance of the transferability is compared neither to the white box nor the decision-based black box attack." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 502, + 398, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 398, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 398, + 516 + ], + "type": "text", + "content": "2.2. Fingerprinting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Models are valuable assets because of their training costs, be it the expertise, the annotated data, or the computational power. Fingerprinting methods have been proposed to protect this intellectual property. They provide a similarity score between the two models by looking at similar decisions in submitted queries. They often use adversarial examples specifically designed to be unique for a model. The models are considered similar if they share the same decision for these inputs. IPGuard [2] creates adversarial examples close to the decision boundary to frame it. Universal adversarial perturbations are used to characterize the boundary in [21]. The paper [15] built adversarial examples with a transferable property called conferrable adversarial examples. AFA [37] crafts adversarial examples model-specific using dropout. FBI [18] is the only method using unmodified images. It estimates the mutual information to measure" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4543" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 266, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 266, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 266, + 83 + ], + "type": "text", + "content": "the statistical dependence of two models' predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 84, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 84, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 84, + 287, + 109 + ], + "type": "text", + "content": "Our paper investigates the idea that a model similar to the target may be a good source for transferable attacks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 121, + 130, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 130, + 134 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 130, + 134 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "text", + "content": "This paper introduces a new way to gauge transferability. The main motivation is that transferability should indicate whether a source " + }, + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "text", + "content": " is useful for attacking a target " + }, + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "text", + "content": ", independently from the inherent robustness of " + }, + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 141, + 287, + 190 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 198, + 115, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 115, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 115, + 209 + ], + "type": "text", + "content": "3.1. Notations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "f:[0,255]^N \\to \\mathcal{R}^K" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": " denote a classifier computing the predicted probabilities " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": " classes for input " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": ". For a given input " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": ", the adversarial example " + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "inline_equation", + "content": "x_a" + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": " is the result of an attack (be it white box, black box, or by transferability) such that" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 287, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 287, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 77, + 287, + 287, + 306 + ], + "type": "interline_equation", + "content": "\\arg \\max _ {1 \\leq k \\leq K} f _ {k} (x _ {a}) \\neq \\arg \\max _ {1 \\leq k \\leq K} f _ {k} (x) = y. \\tag {1}", + "image_path": "25e6a6e9c2654f21e816c61f5ad5ecf0b4fc14b35ff8991c6fc4072a81c11eee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": "In the case of transferable attacks, the attacker uses a source model " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": " to build a transferable adversarial example against the target model " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": ". This paper considers a collection of " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": " source models denoted by " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s = \\{f_s^1,\\dots,f_s^m\\}" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": ", and a set of " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": " inputs " + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 47, + 316, + 287, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 384, + 134, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 134, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 134, + 396 + ], + "type": "text", + "content": "3.2. Measurement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 403, + 287, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 439 + ], + "type": "text", + "content": "Distortion. We measure the distortion of an adversarial perturbation using the Euclidean norm which is a common choice as we deal with natural images:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 101, + 448, + 287, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 448, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 101, + 448, + 287, + 463 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s t} \\left(x _ {a}, x\\right) := \\| x _ {a} - x \\| _ {2} / \\sqrt {N}, \\tag {2}", + "image_path": "bfd0c41b067bd8dcbba0403dc1d8ddca48ed7782da1f6254a31a66569be9d878.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 472, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 496 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 472, + 287, + 496 + ], + "type": "inline_equation", + "content": "N = 3\\times H\\times L" + }, + { + "bbox": [ + 47, + 472, + 287, + 496 + ], + "type": "text", + "content": " pixels. This distortion can be seen as the average amplitude of the perturbation per pixel." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "content": "Operating characteristic. Given an attack (be it white box, black box, or by transferability) and a target model " + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "content": ", we call operating characteristic the following function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 543, + 287, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 543, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 543, + 287, + 556 + ], + "type": "interline_equation", + "content": "P (D) := \\mathbb {P} (\\operatorname {d i s t} \\left(x _ {a}, x\\right) < D). \\tag {3}", + "image_path": "a5b6bea6691427aef32f75a7653506943c2cd9bfac768f65321f5adf44e9808e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 566, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 287, + 613 + ], + "type": "text", + "content": "It is thus the Attack Success Rate as a function of the distortion. Contrary to the literature which measures the ASR for a few distortion levels, we consider the full range of " + }, + { + "bbox": [ + 47, + 566, + 287, + 613 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 566, + 287, + 613 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 622, + 140, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 140, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 140, + 635 + ], + "type": "text", + "content": "3.3. Transferability" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": "Given a white box and a black box attacks, The proposed methodology first computes the operating characteristic " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "P_{t}^{\\mathrm{wb}}(D)" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " of a state-of-the-art white box attack directly applied to the target model, and the operating characteristic " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "P_{t}^{\\mathrm{bb}}(D)" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " of a state-of-the-art black box attack. It then measures where the operating characteristic of the transferable" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "attack from the source " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " to the target " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " lies in between the two characteristics as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 355, + 104, + 545, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 104, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 355, + 104, + 545, + 134 + ], + "type": "interline_equation", + "content": "T _ {s, t} := \\frac {\\int_ {0} ^ {\\infty} P _ {s , t} (u) - P _ {t} ^ {\\mathrm {b b}} (u) d u}{\\int_ {0} ^ {\\infty} P _ {t} ^ {\\mathrm {w b}} (u) - P _ {t} ^ {\\mathrm {b b}} (u) d u}. \\tag {4}", + "image_path": "885f0f710bbc3ee99f25d283d16a9339010b6e678830f3f7bcb65c16f4054edb.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "text", + "content": "The proposed score is calculated as the ratio of the areas between the different operational characteristics, which are defined as Cumulative Distribution Functions (CDFs). The numerator is therefore close to the 1-Wasserstein distance [5] between the ASR function of the distortions of the transferable attack and that of the black box attack, where the absolute value is removed to obtain a signed score. If the transferable attack performs as well as the white box attack (resp. as bad as the black box attack), then " + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "inline_equation", + "content": "T_{s,t} = 1" + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "inline_equation", + "content": "T_{s,t} = 0" + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "text", + "content": "). Figure 1 shows that the numerator of (4) can indeed be negative as transferability can be even worse than the black box if the source " + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "text", + "content": " is not well chosen. Therefore, zero is not a lower bound for " + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "inline_equation", + "content": "T_{s,t}" + }, + { + "bbox": [ + 304, + 143, + 545, + 335 + ], + "type": "text", + "content": ". In the denominator, the 1-Wasserstein distance is obtained exactly between the white box and the black box because the white box consistently generates adversarial examples with lower distortion." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 344, + 447, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 447, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 447, + 357 + ], + "type": "text", + "content": "3.4. Practical implementation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": "Attacks. We only consider state-of-the-art attacks which almost surely deliver an adversarial example. As said in Sect. 2.1, this is usually the case for black box attacks, but we are limited to using white box attacks that are not subject to a distortion budget. As for transferability, the attack uses the publicly available model " + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": " and input " + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": " to craft an adversarial direction " + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "inline_equation", + "content": "u_{x,s}" + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": " of Euclidean norm " + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\sqrt{N}" + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": ". The use of " + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\sqrt{N}" + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": " for the norm is for the sake of simplicity in notation. We assume that there is an oracle giving the minimal distortion along this direction to fool the target " + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": ". In other words, " + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "inline_equation", + "content": "x_{a} = x + d u_{x,s}" + }, + { + "bbox": [ + 304, + 363, + 545, + 496 + ], + "type": "text", + "content": ", with" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 333, + 506, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 506, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 333, + 506, + 545, + 525 + ], + "type": "interline_equation", + "content": "d = \\min \\left\\{\\delta : \\arg \\max _ {k} f _ {t, k} (x + \\delta u _ {x, s}) \\neq y \\right\\}. \\tag {5}", + "image_path": "e272f437073aae78ce8c1516ce98b9098975c237c4d3c7b0a6b50fcfe1a757d2.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 534, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 582 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 304, + 534, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(x_a,x) = d" + }, + { + "bbox": [ + 304, + 534, + 545, + 582 + ], + "type": "text", + "content": ". This definition favours transferability as its best. In practice, such an oracle does not exist but the attacker finds a good estimate of (5) thanks to a line search within a few queries to the target." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": "Transferability. We run a given attack over a collection " + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": " inputs correctly classified by the target model. We compute the distortions " + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "inline_equation", + "content": "d(j) = \\mathrm{dist}(x_{a,j}, x_j)" + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "inline_equation", + "content": "x_j \\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": " and sort them so that " + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "inline_equation", + "content": "d(1) \\leq d(2) \\leq \\ldots \\leq d(n)" + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "inline_equation", + "content": "d(0) = 0" + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "inline_equation", + "content": "d(n + 1) = \\infty" + }, + { + "bbox": [ + 304, + 582, + 545, + 655 + ], + "type": "text", + "content": " to properly define the empirical operating characteristic by the following step function:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 345, + 665, + 545, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 665, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 345, + 665, + 545, + 679 + ], + "type": "interline_equation", + "content": "\\hat {P} (D) := j / n \\quad \\forall D \\in [ d (j), d (j + 1)). \\tag {6}", + "image_path": "80896d9f586d58e06757cf928283922820223df0dbf6bf4d7a15b2e06716a075.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "It is then easier to estimate the integrals appearing in (4), which are indeed areas under two curves, by a Lebesgue" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4544" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 228, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 228, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 228, + 85 + ], + "type": "text", + "content": "sum rather than a Riemann sum. This gives:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 90, + 287, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 90, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 90, + 287, + 121 + ], + "type": "interline_equation", + "content": "\\hat {T} _ {s, t} := \\frac {\\sum_ {j = 1} ^ {n} d _ {s , t} (j) - d _ {t} ^ {\\mathrm {b b}} (j)}{\\sum_ {j = 1} ^ {n} d _ {t} ^ {\\mathrm {w b}} (j) - d _ {t} ^ {\\mathrm {b b}} (j)}, \\tag {7}", + "image_path": "1c0e74f6d70717d81d71f9316691bbe00317e9e62d64f91302d09b30c9eec7aa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "text", + "content": "where the distortions " + }, + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "inline_equation", + "content": "(d_t^{\\mathrm{bb}}(j))_j" + }, + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "inline_equation", + "content": "(d_t^{\\mathrm{wb}}(j))_j" + }, + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "text", + "content": ") resulting from the black box (resp. white box) attack against model " + }, + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 127, + 287, + 163 + ], + "type": "text", + "content": " are also sorted in increasing order." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 173, + 286, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 286, + 186 + ], + "type": "text", + "content": "4. Triad of transferability: data, model, attack" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 194, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 217 + ], + "type": "text", + "content": "This section is an experimental investigation of the factors affecting transferability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 224, + 162, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 162, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 162, + 236 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 242, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 324 + ], + "type": "text", + "content": "The study assesses transferable attacks on a total of 48 models, with 47 of them sourced from the Timm library [30] and one very robust, namely ResNet50AdvTrain, obtained from the GitHub repository 1. Appendix A lists all the models. The experiments utilize 100 images from the validation set of ILSVRC'12, all of which are correctly classified by all models under consideration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "text", + "content": "The study considers three transferable attacks - DI [32], TAIG [9], and DWP [26]- each using a different approach to improve transferability (see Sect. 2.1). These attacks are selected as the best in their categories in [38]. They all share an " + }, + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "text", + "content": " parameter to control the maximum perturbation added per pixel. The effect of the parameter " + }, + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "text", + "content": " indeed happens to be negligible in our protocol, as demonstrated in Appendix B. We choose " + }, + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\epsilon = 8" + }, + { + "bbox": [ + 46, + 326, + 287, + 420 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "type": "text", + "content": "To measure transferability (7), we need state-of-the-art black box and white box attacks. Certain methods may exhibit a preference towards one model over the other, necessitating the use of multiple attacks. Our study employs four white box attacks (BP [35], DeepFool [19], I-FGSM [10], and PGD [16]) and three black box attacks (SurFree [17], RayS [4], and GeoDA [24]). All black box attacks are run with 2,000 queries, which has been determined to be sufficient for achieving convergence. For distortion-constrained white box attack, the " + }, + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "type": "text", + "content": " parameter is set to 4. We record the smallest perturbation distortion over the black box (white box) for each image and draw the operating characteristic (6) as appearing in green (resp. red) dashed line in Fig. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "type": "text", + "content": "As for the transferable attack, for a given target model, the attacker has access to a subset of all the other models whose architecture differs from the target. This amounts to an average of 45 models out of 48. For instance, in Fig. 1, CoatLitesmall being the target, we exclude all other CoatLite models from being a source. The light blue area delimits the operating characteristics of transferable attack DI [32] using as the source one of the 45 remaining models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 416, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 416, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 416, + 84 + ], + "type": "text", + "content": "4.2. Model dependence" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 101, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 307, + 101, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 307, + 101, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 101, + 545, + 324 + ], + "type": "image", + "image_path": "79c1ed355cd42f1ac9c7073c690859e106db534d9f137ba089de4233d6566483.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 331, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 357 + ], + "type": "text", + "content": "Figure 2: Transferability score " + }, + { + "bbox": [ + 305, + 331, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 305, + 331, + 545, + 357 + ], + "type": "text", + "content": " matrix of 48 sources and 48 targets listed in App. A with attack DI [32]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "text", + "content": "Large transferability variation. Figure 2 shows the matrix " + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "text", + "content": " are any of the " + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "inline_equation", + "content": "48^2" + }, + { + "bbox": [ + 304, + 369, + 545, + 440 + ], + "type": "text", + "content": " pairs, for attack D I [32]. Not all models possess the same transferability capabilities. At first look, the figure contains more blue than red cells which means that transferability takes negative value more often." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 441, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 489 + ], + "type": "text", + "content": "Some rare models, like PiTsmall-dist, exhibit good transferability towards any target. On the contrary, DPN68b is always a bad source. On the other hand, ResNet50AdvTrain is a very difficult target (note how" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 522, + 523, + 666 + ], + "blocks": [ + { + "bbox": [ + 307, + 522, + 523, + 666 + ], + "lines": [ + { + "bbox": [ + 307, + 522, + 523, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 523, + 666 + ], + "type": "image", + "image_path": "5300a085f9c1a7cac86c7950c1f852ab2c7d4e0cb7bd6d63f57a2a43d5fa5ce6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 673, + 545, + 710 + ], + "lines": [ + { + "bbox": [ + 305, + 673, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 545, + 710 + ], + "type": "text", + "content": "Figure 3: " + }, + { + "bbox": [ + 305, + 673, + 545, + 710 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 305, + 673, + 545, + 710 + ], + "type": "text", + "content": " function of the number of available sources and attack DI [32]. The best or worst model selected per image (solid line) or on average (dashed line)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 246, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 246, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 246, + 712 + ], + "type": "text", + "content": "1https://github.com/MadryLab/robustness" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4545" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 272, + 200 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 272, + 200 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 272, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 272, + 200 + ], + "type": "image", + "image_path": "22747ae36c47dbdbf308284fcae58d936c22d3ad31956ac7f846a275bdc86dda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 245 + ], + "type": "text", + "content": "Figure 4: Distribution of best image-model selection function of the number of sources. Adversarial examples obtained with DI [32]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 319 + ], + "type": "text", + "content": "ever that its accuracy is low), followed by PiTlight and ConViTtiny, whereas the family of ResNet are fairly easy target." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 471 + ], + "type": "text", + "content": "Prior works have shown that models with architectures similar to the target transfer better [22, 31]. The red squares appearing in the matrix confirm this (models are ordered according to architecture family). For instance, MobileNetV2 and RexNet architectures exhibit transferability close to 0.7. However, this is not an absolute truth. For example, EfficientNet-B0 has better transferability against MixNetlarge than MixNetmedium, even though they have similar architectures. ConViT transfers exceptionally well to Twins although their architectures are very different. Similar observations hold for the other attacks but with lower transferability scores (see Appendix C)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 503 + ], + "type": "text", + "content": "From now on, we exclude models whose architecture is similar to the target." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "Impact on the attack. Selecting the right source among several available models is critical for achieving high transferability. Figure 3 displays the transferability score (7) for the best and the worst choices (shown in dotted lines) as the number of source candidates increases. In this evaluation, a single model is selected as the source for building all adversarial examples. On average, the transferability is lower than zero, regardless of the attack method. This means that transferable attacks perform worst than black box attacks on average. If the attacker knows how to select the best source model, " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " quickly converges to a maximum which is positive but below 0.5. This means that transferability at its best performs closer to a black box attack rather than a white box attack. These remarks highly mitigate the threat of transferability and put emphasis on the crucial selection of the best source model. As far as we know, these facts are not reported in the adversarial examples literature." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 424, + 175 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 424, + 175 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 424, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 424, + 175 + ], + "type": "image", + "image_path": "8c36484bfba84df3025f192834fb37ff19e02a31036c7cac698e3f45655215c5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 175, + 424, + 277 + ], + "blocks": [ + { + "bbox": [ + 316, + 175, + 424, + 277 + ], + "lines": [ + { + "bbox": [ + 316, + 175, + 424, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 424, + 277 + ], + "type": "image", + "image_path": "5876eed1cf796cf9d0972e5f086a19cad7abdd427e716d816a800fa985dcf0f6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 284, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 306, + 284, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 545, + 308 + ], + "type": "text", + "content": "Figure 5: 2D Histogram of the minimum distortion for transferable and white box attacks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 533, + 175 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 533, + 175 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 533, + 175 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 533, + 175 + ], + "type": "image", + "image_path": "e11cac767e0f0bae8379a1e7726d996ea3ee6678f6ad18e77b83ad17ab42d642.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 175, + 533, + 277 + ], + "blocks": [ + { + "bbox": [ + 425, + 175, + 533, + 277 + ], + "lines": [ + { + "bbox": [ + 425, + 175, + 533, + 277 + ], + "spans": [ + { + "bbox": [ + 425, + 175, + 533, + 277 + ], + "type": "image", + "image_path": "01e31915b18a074af56f26cba78f58fdc470bcdd15c6dc7ac4887ab8f20b5a49.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 331, + 416, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 416, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 416, + 344 + ], + "type": "text", + "content": "4.3. Image dependence" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "text", + "content": "The difficulty of transferring adversarial examples from a source to a target varies significantly depending on the input " + }, + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "text", + "content": ". This is illustrated in Fig. 4, which shows the distribution of the best-performing model for each image based on the available sources. Even with a large number of sources available, there is typically one source that performs significantly better than the others, but only over " + }, + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "text", + "content": " of the images on average. However, this superiority decreases rapidly as the number of available sources increases. It means that a better source exists for " + }, + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 350, + 545, + 481 + ], + "type": "text", + "content": " of the images on average." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 482, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 565 + ], + "type": "text", + "content": "Supposing that the attacker knows the best source of each input, Figure 3 shows in solid line that the transferability converges to a value close to 0.5. The performance of the transferable attack lies in between the ones of the white box and black box attacks. This highlights the importance of selecting an appropriate source model for a given target and input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 575, + 417, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 417, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 417, + 588 + ], + "type": "text", + "content": "4.4. Attack dependence" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Transferable attack. Figure 5 compares transferable attacks with a 2D histogram of the distortion pair for two attacks. This is computed over all inputs in " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and all pairs of source and target models. The results show that DWP [26] exhibits poor transferability compared to DI [32] and TAIG [9], which produce adversarial perturbations with similar distortion. Additionally, regardless of the attack complexity or method, the challenging images remain consistent. If one attack requires a high distortion for a given input / source / target, other attacks are likely to encounter" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4546" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "type": "text", + "content": "similar difficulties." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 84, + 288, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 84, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 84, + 288, + 242 + ], + "type": "text", + "content": "Traditional white box attack. We now compare the methods designed for transferability with a naive approach. A white box attack (ie. not specific to transferability) is executed on the source, and then the direction found serves as " + }, + { + "bbox": [ + 47, + 84, + 288, + 242 + ], + "type": "inline_equation", + "content": "u_{x,s}" + }, + { + "bbox": [ + 47, + 84, + 288, + 242 + ], + "type": "text", + "content": " to place the adversarial examples on the target boundary with (5). The best " + }, + { + "bbox": [ + 47, + 84, + 288, + 242 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 47, + 84, + 288, + 242 + ], + "type": "text", + "content": " score is 0.27 for BP [35] compared to 0.52 for DI [32]. This confirms the superiority of the recent methods designed for transferability. However, even though DI [32] is on average better, Figure 5 shows that when the necessary distortion is small, traditional white box attacks like DeepFool [19] and BP [35] indeed beats DI [32]. On the other hand, DI [32] performs much better for inputs requiring more distortion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 251, + 216, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 216, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 216, + 264 + ], + "type": "text", + "content": "5. How to choose the best source" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 319 + ], + "type": "text", + "content": "Section 4 outlines that the choice of the source is of utmost importance thanks to the transferability measure defined in Sect. 3. This section investigates whether the attacker can guess which model is the best source." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 319, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 415 + ], + "type": "text", + "content": "We now propose a procedure which combines the dependences with respect to the source and target models (Sect. 4.2), and the input (Sect. 4.3). Model dependence is measured by evaluating the similarity between the source and target models, which is denoted as ModSim. On the other hand, the image dependence is measured by the quality of the adversarial example, denoted as TransQ. Both metrics are combined into the following score:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 422, + 287, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 287, + 435 + ], + "type": "interline_equation", + "content": "\\operatorname {F i T} (s, t, x) := \\operatorname {M o d S i m} (s, t) \\times \\operatorname {T r a n s Q} (s, x). \\tag {8}", + "image_path": "1723ee7f93b6d37edd14c0217bfe1b11373b3208512145796d6f51afaa9553d0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 491 + ], + "type": "text", + "content": "These indicators should be easy to compute. We especially pay attention to the number of queries to the target. This score opens the door to a new strategy for the attacker which first selects the best source among the available models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 497, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 497, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 98, + 497, + 287, + 515 + ], + "type": "interline_equation", + "content": "s ^ {\\star} (t, x) = \\underset {\\sigma \\in \\mathcal {F} _ {s}} {\\arg \\max } \\mathrm {F i T} (\\sigma , t, x), \\tag {9}", + "image_path": "2d533fec41bafe93ab8f398943babb23524ed221bd2598adf921bc2ad02669ee.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 521, + 249, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 249, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 249, + 534 + ], + "type": "text", + "content": "and then crafts the adversarial direction " + }, + { + "bbox": [ + 47, + 521, + 249, + 534 + ], + "type": "inline_equation", + "content": "u_{x,s^*(t,x)}" + }, + { + "bbox": [ + 47, + 521, + 249, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 540, + 173, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 173, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 173, + 553 + ], + "type": "text", + "content": "5.1. Criterion ModSim(s,t)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 558, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 641 + ], + "type": "text", + "content": "Section 4.2 highlights the correlation between the transferability and the similarity between the source and target models. Gauging model similarity has been previously studied in the context of fingerprinting as a defense to protect intellectual property (see Sect. 2.2). This paper uses fingerprinting methods as an attack that leaks information about the target." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": "We consider the fingerprinting method [18] because it works in a decision-based setup since the target is a black box in our application. Querying two models " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": " with few natural images, it computes a distance " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{Dist}(s, t) \\in [0, 1]" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": " by comparing their outputs. Since we look for a similarity, we set " + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{ModSim}(s, t) = 1 - \\operatorname{Dist}(s, t)" + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "text", + "content": ". However," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 546, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 546, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 546, + 108 + ], + "type": "text", + "content": "this provides a symmetrical similarity, ie. " + }, + { + "bbox": [ + 305, + 72, + 546, + 108 + ], + "type": "inline_equation", + "content": "\\operatorname{ModSim}(s,t) = \\operatorname{ModSim}(t,s)" + }, + { + "bbox": [ + 305, + 72, + 546, + 108 + ], + "type": "text", + "content": ", while transferability is not (see Fig. 2). This shows that this criterion alone is not sufficient." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 115, + 433, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 115, + 433, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 433, + 129 + ], + "type": "text", + "content": "5.2. Criterion TransQ(s, x)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": "This criterion evaluates the general transferability of a given adversarial example crafted by a source. Our idea is to leverage the assumption that the attacker has a set of models " + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 305, + 133, + 547, + 205 + ], + "type": "text", + "content": ". Consequently, we can evaluate the transferability thanks to the other models of this set, without querying the target." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "text", + "content": "For a given input " + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "text", + "content": ", the source " + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "text", + "content": " provides the adversarial direction " + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "inline_equation", + "content": "u_{x,s}" + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "text", + "content": " and we compute the distortion " + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "inline_equation", + "content": "d_{s,\\sigma}" + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "text", + "content": " necessary to delude classifier " + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathcal{F}_s" + }, + { + "bbox": [ + 305, + 205, + 546, + 253 + ], + "type": "text", + "content": " with (5). We then aggregate these distortions into a single score with two flavours:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 260, + 546, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 260, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 328, + 260, + 546, + 297 + ], + "type": "interline_equation", + "content": "\\operatorname {T r a n s Q} ^ {(1)} (s, x) := \\left(\\frac {1}{| \\mathcal {F} _ {s} |} \\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {s, \\sigma}\\right) ^ {- 1}. \\tag {10}", + "image_path": "48c93bb5cabcd215d62b4d33dce86e0814268fc6b00aeb2df8fd27a775d99b91.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 303, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 327 + ], + "type": "text", + "content": "A good source for input " + }, + { + "bbox": [ + 305, + 303, + 545, + 327 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 303, + 545, + 327 + ], + "type": "text", + "content": " gives birth to lower distortions, so that " + }, + { + "bbox": [ + 305, + 303, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathrm{TransQ}^{(1)}(s,x)" + }, + { + "bbox": [ + 305, + 303, + 545, + 327 + ], + "type": "text", + "content": " is large." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 341, + 335, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 335, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 341, + 335, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\operatorname {T r a n s} Q ^ {(2)} (s, x) := \\frac {\\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {s , \\sigma} - d _ {\\sigma} ^ {\\mathrm {b b}}}{\\sum_ {\\sigma \\in \\mathcal {F} _ {s}} d _ {\\sigma} ^ {\\mathrm {w b}} - d _ {\\sigma} ^ {\\mathrm {b b}}}. \\tag {11}", + "image_path": "612d355e8098877554cf08b308ab7b675401136f3f95bcc9d71c003f9ba92e02.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 368, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 368, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 546, + 392 + ], + "type": "text", + "content": "This measure is similar to (7) except that it is computed over the set of models instead of a set of inputs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 399, + 362, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 362, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 362, + 410 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 417, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 545, + 513 + ], + "type": "text", + "content": "The experimental setup is the same as in Sect. 4.1. We select the fingerprinting method FBI [18] with 200 benign natural images to compute the criterion ModSim(s, t). It implies that the attacker first makes 200 queries to the target in a preliminary step before forging any adversarial example. Appendix D.1 shows that more images improve the fingerprinting accuracy hence the model selection, but the results converge after a number of 200 images." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 514, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 545, + 560 + ], + "type": "text", + "content": "This section is structured as follows: we use FiT to select the best model for single-model attacks, then we employ it to identify the best subset of models for ensemble-model attacks and different combinations of attacks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 574, + 427, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 427, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 427, + 586 + ], + "type": "text", + "content": "5.3.1 Single-model attacks" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 547, + 714 + ], + "type": "text", + "content": "Criterion ModSim(s, t). Table 1 indicates that architectural similarity is a reliable measure of transferability between two models. It can drive the selection of a good source giving birth to a transferable attack outperforming the black box attack since " + }, + { + "bbox": [ + 305, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 305, + 594, + 547, + 714 + ], + "type": "text", + "content": " is larger than 0 (except for DWP [26]). Yet, the results remain low compared to the best results obtained. As discussed in Sec 4.2, similarity may not suffice because it implies the selection of one unique for a given target. Better results are achieved when adapting the source to the input." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4547" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 102, + 523, + 316 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 506, + 84 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 506, + 84 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 506, + 84 + ], + "type": "text", + "content": "Table 1: Transferability " + }, + { + "bbox": [ + 85, + 71, + 506, + 84 + ], + "type": "inline_equation", + "content": "{\\widehat{T}}_{s,t}" + }, + { + "bbox": [ + 85, + 71, + 506, + 84 + ], + "type": "text", + "content": " for DI [32],TAIG [9] and DWP [26] for single and ensemble-model attacks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 102, + 523, + 316 + ], + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 316 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 316 + ], + "type": "table", + "html": "
CategorySelection MethodDI [32]TAIG [9]DWP [26]
Single-model attackBest0.52±0.120.46±0.120.34±0.13
Random-0.16±0.26-0.12±0.21-0.72±0.32
ModSimFBI [18]0.18±0.170.12±0.18-0.39±0.19
TransQASR-0.21±0.39-0.24±0.30-1.14±0.95
TransQ(1)(10)0.38±0.150.24±0.160.10±0.20
TransQ(2)(11)0.37±0.180.23±0.180.08±0.22
FiTTransQ(1)(10)0.40±0.130.27±0.170.12±0.20
TransQ(2)(11)0.39±0.160.25±0.180.10±0.22
Ensemble-model attack with three sourcesBest0.72±0.080.62±0.080.46±0.1
Random0.43±0.120.40±0.110.02±0.17
ModSimFBI [18]0.59±0.100.49±0.100.05±0.11
TransQASR0.53±0.210.45±0.270.06±0.27
TransQ(1)(10)0.62±0.110.54±0.130.33±0.13
TransQ(2)(11)0.61±0.110.54±0.120.33±0.17
FiTTransQ(1)(10)0.64±0.110.55±0.140.35±0.15
TransQ(2)(11)0.64±0.100.57±0.140.36±0.15
", + "image_path": "cb42459f66547bf0379a5583e7bfdcbe03deb4b35007d68d09448120cc918a8f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 347, + 272, + 489 + ], + "blocks": [ + { + "bbox": [ + 58, + 347, + 272, + 489 + ], + "lines": [ + { + "bbox": [ + 58, + 347, + 272, + 489 + ], + "spans": [ + { + "bbox": [ + 58, + 347, + 272, + 489 + ], + "type": "image", + "image_path": "c63842bc1ec68e5899d631f05e4a9fd010da1807ee7e51175712e54e5aa38577.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 499, + 287, + 548 + ], + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 548 + ], + "type": "text", + "content": "Figure 6: " + }, + { + "bbox": [ + 46, + 499, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 46, + 499, + 287, + 548 + ], + "type": "text", + "content": " as a function of the number of available sources for several selection methods. Dotted lines refer to the selection of a unique model for all images and solid lines refer to a model selected per image. Attack is DI [32]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 569, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 286, + 688 + ], + "type": "text", + "content": "Criterion TransQ(s,x). Improving transferability without querying the target model in a preliminary step is possible thanks to TransQ(s,x). Adversarial examples that exhibit good transferability on multiple models are more likely to also deceive the unknown targeted model. Figure 6 shows that a significant improvement in transferability is achieved even with only a few available models. Table 1 confirms this observation for the two other attack methods. This strategy is indeed better than the selection based on model similarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Score FiT (s,t,x). Combining both criteria together as in (8) leads to a slight improvement in transferability com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "text", + "content": "pared to " + }, + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "inline_equation", + "content": "\\mathrm{TransQ}(s,x)" + }, + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "text", + "content": " alone. Fig. 6 confirms this holds over a wide range of numbers of available sources. For single-model attacks, " + }, + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "inline_equation", + "content": "\\mathrm{TransQ}^{(1)}" + }, + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "text", + "content": " gives slightly better results than " + }, + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "inline_equation", + "content": "\\mathrm{TransQ}^{(2)}" + }, + { + "bbox": [ + 305, + 346, + 545, + 393 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "content": "Visual results Figure 2 visually demonstrates the impact of different model selection methods on the quality of adversarial examples. Even when FiT is not accurate, the resulting adversarial examples are still close to the best ones obtained from the source models, and the perturbation remains imperceptible. However, random selection generates noisy perturbations, and the worst-case scenario destroys the image entirely." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 503, + 443, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 443, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 443, + 514 + ], + "type": "text", + "content": "5.3.2 Ensemble-model attacks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Importance of model selection in ensemble-model attack. Ensemble attacks remain an under-studied area due to the significant computational resources required for evaluating attacks. Consequently, these attacks have only been evaluated with a limited number of sources. Appendix D.2 shows that increasing the number of sources is not necessarily beneficial. The FiT score provides a scalable solution: we select a small subset of sources based on their FiT scores (top-3) from the bigger set of available sources. In a way, it is better to put quality above quantity. For example, when running an ensemble-model attack against " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\times \\mathrm{CiT}_{\\mathrm{nano}}" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " selecting only the three best models using the FiT measure can lead to a significant improvement in transferability compared to using a larger set of models. In our experiments, an ensemble-model of 20 random models achieved a " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " of 0.47, while an ensemble-model of only three models se" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4548" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 101, + 91, + 492, + 319 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 536, + 83 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 536, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 536, + 83 + ], + "type": "text", + "content": "Table 2: Visual impact of the source selection with DI [32] attacking ConViTbase (first row) and DPN92 (second row)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 101, + 91, + 492, + 319 + ], + "lines": [ + { + "bbox": [ + 101, + 91, + 492, + 319 + ], + "spans": [ + { + "bbox": [ + 101, + 91, + 492, + 319 + ], + "type": "table", + "html": "
OriginalBestFiT(TransQ(1))RandomWorst
label source distortionmouseoil filter PiTsmall-dist 5.62purse ConViTbase 8.13purse MobileNetV2110d 35.9jigsaw puzzle DenseNet121 87.8
label source distortiontramelec. locomotive ResNetV250x1-dist 6.94elec. locomotive PiTsmall-dist 7.50racing car DLA60 33.3jigsaw puzzle MixNetmedium 128.7
", + "image_path": "3ff88b575b363ce4a6f5f6680b7ed97f6b04653660aba440a1e0614b97598eb4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 288, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 362 + ], + "type": "text", + "content": "lected with " + }, + { + "bbox": [ + 46, + 337, + 288, + 362 + ], + "type": "inline_equation", + "content": "\\mathbb{F}\\mathbb{i}\\mathbb{T}" + }, + { + "bbox": [ + 46, + 337, + 288, + 362 + ], + "type": "text", + "content": " was able to achieve a " + }, + { + "bbox": [ + 46, + 337, + 288, + 362 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 46, + 337, + 288, + 362 + ], + "type": "text", + "content": " of 0.56 against the same target." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 506 + ], + "type": "text", + "content": "Performance of ensemble-model attacks. Table. 1 shows that ensemble-model transferability surpasses that of single-model attacks. While the average results over a random selection of 3 sources increase, they remain closer to the black box results than the white box ones (transferability lower than 0.5). Choosing the top-3 sources returned by our scores for leading the ensemble-model attack yields better performance, comparable to the best possible results obtainable with ensemble-model attacks. Notably, for the DI [32] and TAIG [9] method, it approaches the performance of white box results (transferability greater than 0.5)." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 590, + 285, + 701 + ], + "blocks": [ + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "type": "text", + "content": "Table 3: Transferability " + }, + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\hat{T}_{s,t}" + }, + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "type": "text", + "content": " for white box and transferable attacks (single-model). The FiT score uses " + }, + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{TransQ}^{(1)}" + }, + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 590, + 285, + 701 + ], + "lines": [ + { + "bbox": [ + 48, + 590, + 285, + 701 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 285, + 701 + ], + "type": "table", + "html": "
AttackBestFiT
W.B.BP [35]0.27 ±0.18-0.06±0.31
DeepFool [19]0.03 ±0.20-0.34±0.34
PGD [16]0.29 ±0.20-0.01±0.36
I-FGSM [10]0.29 ±0.20-0.03±0.34
Trans.DI [32]0.52 ±0.120.40±0.13
TAIG [9]0.46 ±0.120.27±0.17
DWP [26]0.34 ±0.130.12±0.20
All Attacks0.65 ±0.090.48±0.14
", + "image_path": "7e484f0799afeaf937ecc9202a6ae9307259c7b4ba91c43d701c7b0a7bf291a5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 338, + 486, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 486, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 486, + 350 + ], + "type": "text", + "content": "5.3.3 White box vs. transferable attacks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 357, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 525 + ], + "type": "text", + "content": "Our last result is that the selection of a single source with the FiT score does not make traditional white box attacks transferable. Table 3 shows that BP [35], DeepFool [19], PGD [16], and I-FGSM [10] yield negative transferability values. More precisely, BP [35], I-FGSM [10], and PGD [16] perform better with FiT than the transferable attacks without any selection mechanism (ie. on average with a random source). However, section 4.4 highlights that the traditional white box attacks may be competitive for some inputs demanding low adversarial perturbation distortion. It is valuable to add them to the options of the attacker and let the FiT score decide the preferable option (source and attack method). This provides our best transferability score for a single model attack, close to 0.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": "Transferability is a crucial feature of adversarial examples as it allows a single perturbation to deceive multiple models. However, solely relying on Attack Success Rate (ASR) to measure transferability overlooks the degree of distortion needed to fool a model. This paper introduces a novel approach to assess transferability by comparing it to the distortion of two reference attacks: white box and black box attacks. We show that transferable attacks can perform worse than black box attacks without an appropriate selection of the source model, highlighting the need to choose the best source model to target a specific model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "The proposed solution, named F i T, allows the attacker to choose one of the best source models with minimal" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4549" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "content": "queries to the target. Our experiments demonstrate that the proposed solution performs well in multiple attack scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 179 + ], + "type": "text", + "content": "This study has highlighted the differences in transferability between images for the same source model and their particularity for this specific network. Further research could focus on addressing this issue and investigating its underlying causes, with the hope of designing an even better selection mechanism able to spot the best source model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 190, + 105, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 190, + 105, + 202 + ], + "spans": [ + { + "bbox": [ + 48, + 190, + 105, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 209, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 209, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 287, + 253 + ], + "type": "text", + "content": "[1] Mauro Barni, Kassem Kallas, Ehsan Nowroozi, and Benedetta Tondi. On the transferability of adversarial examples against CNN-based image forensics. In ICASSP. IEEE, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 255, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 288, + 308 + ], + "type": "text", + "content": "[2] Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Ipguard: Protecting intellectual property of deep neural networks via fingerprinting the classification boundary. In ACM Asia Conference on Computer and Communications Security, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 310, + 287, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 287, + 343 + ], + "type": "text", + "content": "[3] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In Symposium on Security and Privacy. IEEE, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 344, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 344, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 344, + 287, + 376 + ], + "type": "text", + "content": "[4] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In ACM SIGKDD, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 377, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 377, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 53, + 377, + 287, + 399 + ], + "type": "text", + "content": "[5] Marco De Angelis and Ander Gray. Why the 1-wasserstein distance is the area between the two marginal cdfs, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 400, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 400, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 53, + 400, + 287, + 433 + ], + "type": "text", + "content": "[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 287, + 477 + ], + "type": "text", + "content": "[7] Qian Huang, Isay Katsman, Horace He, Zeqi Gu, Serge Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 479, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 287, + 511 + ], + "type": "text", + "content": "[8] Tianjin Huang, Vlado Menkovski, Yulong Pei, Yuhao Wang, and Mykola Pechenizkiy. Direction-aggregated attack for transferable adversarial examples. JETC, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 512, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 534 + ], + "type": "text", + "content": "[9] Yi Huang and Adams Wai-Kin Kong. Transferable adversarial attack based on integrated gradients. In ICLR, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 535, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 287, + 567 + ], + "type": "text", + "content": "[10] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security. Chapman and Hall/CRC, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "text", + "content": "[11] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In CVPR, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[12] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan Yuille. Learning transferable adversarial examples via ghost networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[13] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[14] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Nils Lukas, Yuxuan Zhang, and Florian Kerschbaum. Deep neural network fingerprinting by conferrable adversarial examples. In ICLR, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "text", + "content": "[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 140, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 161 + ], + "type": "text", + "content": "[17] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Surfree: a fast surrogate-free black-box attack. In CVPR, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[18] Thibault Maho, Teddy Furon, and Erwan Le Merrer. Fbi: Fingerprinting models with benign inputs. arXiv preprint arXiv:2208.03169, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "text", + "content": "[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[20] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Fahad Khan, and Fatih Porikli. On improving adversarial transferability of vision transformers. In ICLR, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "text", + "content": "[21] Zirui Peng, Shaofeng Li, Guoxing Chen, Cheng Zhang, Haojin Zhu, and Minhui Xue. Fingerprinting deep neural networks globally via universal adversarial perturbations, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "text", + "content": "[22] Deyan Petrov and Timothy M Hospedales. Measuring the transferability of adversarial examples. arXiv preprint arXiv:1907.06291, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 360 + ], + "type": "text", + "content": "[23] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints. In NeurIPS, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "text", + "content": "[24] Ali Rahmati, Seyed-Mohsen Moosavi-Dezfooli, Pascal Frossard, and Huaiyu Dai. Geoda: a geometric framework for black-box adversarial attacks. In CVPR, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "text", + "content": "[25] Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[26] Hung-Jui Wang, Yu-Yu Wu, and Shang-Tse Chen. Enhancing targeted attack transferability via diversified weight pruning. arXiv preprint arXiv:2208.08677, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 491 + ], + "type": "text", + "content": "[27] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "text", + "content": "[28] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting adversarial transferability through enhanced momentum. In BMVC, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 525, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 558 + ], + "type": "text", + "content": "[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In ICCV, 2021." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "text", + "content": "[30] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "text", + "content": "[31] Lei Wu and Zhanxing Zhu. Towards understanding and improving the transferability of adversarial examples in deep neural networks. In ACML, pages 837-850. PMLR, 2020." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[32] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[33] Yifeng Xiong, Jiadong Lin, Min Zhang, John E Hopcroft, and Kun He. Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In CVPR, 2022." + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4550" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 353 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[34] Liping Yuan, Xiaqing Zheng, Yi Zhou, Cho-Jui Hsieh, and Kai-Wei Chang. On the transferability of adversarial attacks against neural text classifier. In EMNLP, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[35] Hanwei Zhang, Yannis Avrithis, Teddy Furon, and Laurent Amsaleg. Walking on the edge: Fast, low-distortion adversarial examples. IEEE Trans. on IFS, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "text", + "content": "[36] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In CVPR, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "text", + "content": "[37] Jingjing Zhao, Qingyue Hu, Gaoyang Liu, Xiaoqiang Ma, Fei Chen, and Mohammad Hassan. AFA: Adversarial Fingerprinting Authentication for deep neural networks. Computer Communications, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "text", + "content": "[38] Zhengyu Zhao, Hanwei Zhang, Renjue Li, Ronan Sicre, Laurent Amsaleg, and Michael Backes. Towards good practices in evaluating transfer adversarial attacks. arXiv preprint arXiv:2211.09565, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "type": "text", + "content": "[39] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 353 + ], + "type": "text", + "content": "[40] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the transferability of adversarial examples with resized-diverse-inputs, diversity-ensemble and region fitting. In ECCV, 2020." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4551" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_content_list.json b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..289d9b0618811cd6e9e1398040c8efa96b7a091f --- /dev/null +++ b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_content_list.json @@ -0,0 +1,1374 @@ +[ + { + "type": "text", + "text": "Human Part-wise 3D Motion Context Learning for Sign Language Recognition", + "text_level": 1, + "bbox": [ + 246, + 130, + 725, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taeryung Lee", + "bbox": [ + 251, + 202, + 372, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yeonguk Oh²", + "bbox": [ + 413, + 203, + 524, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kyoung Mu Lee $^{1,2}$", + "bbox": [ + 563, + 203, + 714, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1IPAI, 2Dept. of ECE & ASRI, Seoul National University, Seoul, Korea", + "bbox": [ + 202, + 220, + 767, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{trlee94, namepllet, kyoungmu}@snu.ac.kr", + "bbox": [ + 307, + 241, + 661, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose P3D, the human part-wise motion context learning framework for sign language recognition. Our main contributions lie in two dimensions: learning the part-wise motion context and employing the pose ensemble to utilize 2D and 3D pose jointly. First, our empirical observation implies that part-wise context encoding benefits the performance of sign language recognition. While previous methods of sign language recognition learned motion context from the sequence of the entire pose, we argue that such methods cannot exploit part-specific motion context. In order to utilize part-wise motion context, we propose the alternating combination of a part-wise encoding Transformer (PET) and a whole-body encoding Transformer (WET). PET encodes the motion contexts from a part sequence, while WET merges them into a unified context. By learning part-wise motion context, our P3D achieves superior performance on WLASL compared to previous state-of-the-art methods. Second, our framework is the first to ensemble 2D and 3D poses for sign language recognition. Since the 3D pose holds rich motion context and depth information to distinguish the words, our P3D outperformed the previous state-of-the-art methods employing a pose ensemble.", + "bbox": [ + 75, + 324, + 473, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 705, + 209, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding and translating sign language is essential due to the huge potential benefit to society. Sign language recognition (SLR) can be considered as a specific kind of action recognition (AR) since both tasks aim to classify human motion. However, SLR has a characteristic that is distinguished from AR; Actions mainly focus on the motion of the human body, while the SLR system should be aware of more detailed information such as hand gestures and facial expressions. Thus, a specified system for SLR will provide further gain over using the AR models on SLR.", + "bbox": [ + 75, + 731, + 470, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous approaches on SLR problem are mainly cat-", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36f551132757c4f08a38b8a6b46bc1bb40fda5afe7af3c332a471f50c0e43b91.jpg", + "image_caption": [ + "Figure 1. Overall pipeline of proposed system. We first extract expressive 3D human pose from RGB video using off-the-shelf pose estimation methods [5, 36]. The expressive 3D human pose consists of 2D pose, 3D pose and facial expression. Then our proposed P3D predicts the word from the expressive 3D human pose." + ], + "image_footnote": [], + "bbox": [ + 506, + 300, + 888, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "egorized into RGB-based [1, 4, 31] and pose-based methods [25, 29, 45, 47, 50]. Pose-based methods are known to have the advantage over RGB-based approaches in terms of the robustness on domain gap [14, 37]. For that, we aimed to enhance the performance of SLR, while preserving the benefits inherent in the pose-based approach.", + "bbox": [ + 496, + 714, + 893, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although recent pose-based SLR methods have shown noticeable progress, there are two main limitations of existing methods. First, they utilize only whole-body representation (e.g., skeleton) to understand the motion context of human [20, 30, 47]. Since such approaches overlook the detailed motion context within a specific part", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20740", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c20a43ea8f81905307522bb4d63b6a22d2fb6a0f24bbbc8871d5cfa02d403e79.jpg", + "image_caption": [ + "RGB Video", + "2D Pose", + "3D Pose", + "Figure 2. Effectiveness of 3D pose for sign language recognition. The 2D poses of the words \"Avoid\" and \"Behind\" are almost identical. Due to the absence of depth information, solely relying on 2D pose makes it hard to distinguish the words. On the other hand, as shown in the red boxes, additionally employing the 3D pose with rich depth information can effectively resolve the ambiguity in sign language recognition." + ], + "image_footnote": [], + "bbox": [ + 196, + 97, + 509, + 426 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d51f5a1814776b316bc991fb7e10f444b1817324401eb1934b0a65484091baab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 95, + 887, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of human [18, 32], it leads to a decrease in SLR accuracy. Second, existing pose-based methods only employ 2D pose [3, 20, 30, 47], which induces ambiguities in distinguishing the different words. As shown in Figure 2, the 2D pose of different words can be almost identical due to the absence of depth information.", + "bbox": [ + 75, + 526, + 468, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the above limitations, we propose P3D, the human part-wise motion context learning framework for accurate sign language recognition. The proposed P3D employs the part-wise encoding Transformer (PET) and whole-body encoding Transformer (WET) as main components. Given the pose sequence, we first divide it into four parts: body, right hand, left hand, and face. In contrast to previous methods where entire human poses are jointly processed, we utilize part-specific PET to encode the motion context of each part. As each PET takes a pose sequence of a single part, it can learn more detailed motion context, which is pivotal for sign language recognition. Then, the encoded features from each part are merged by WET. While PET only considers the intra-part motion context, WET covers the entire part. For the effective fusion of the inter-part and intra-part motion context, we alternately combined PET and WET inspired by spatio-temporal Transformers [2,51]. As a result, our proposed P3D has out", + "bbox": [ + 75, + 628, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performed the previous state-of-the-art [3, 50] on WLASL. Our ablation study on layer design supports part-wise motion context learning benefits SLR performance. Especially, we observed that solely forming the model with WET (i.e., without part-wise motion context) degrades the performance, thus directly induces that our PET contributed to improving the accuracy.", + "bbox": [ + 496, + 526, + 890, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We alleviate the second limitation of previous methods through the ensemble of 2D and 3D poses. In terms of 3D pose, we utilize both 3D positional pose and 3D rotational pose (e.g., SMPL-X [41] pose parameters). The 3D positional pose complements depth information that the 2D pose does not contain, and the 3D rotational pose provides rich motion context on the kinematic chain. Our experiment shows that our P3D outperforms the previous methods by a large margin with using identical inputs; our model shows the superior result from only 2D pose, and also from ensemble of 2D and 3D pose. Moreover, our framework is the first pose-based method that recorded a comparable score or even outperformed the RGB-based methods. We claim our main contributions as follows:", + "bbox": [ + 496, + 633, + 892, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We proposed P3D, the human part-wise motion context learning framework for sign language recognition. Our experiment supports that part-wise learning highly", + "bbox": [ + 517, + 854, + 892, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20741", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "benefits recognition performance. As a result, our proposed P3D outperformed the previous SOTA on the WLASL benchmark.", + "bbox": [ + 107, + 90, + 467, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We employ the pose ensemble by joint-wise pose concatenation in order to exploit both 2D and 3D poses. We observed significant performance gain via pose ensemble, and P3D outperformed the previous methods by a large margin by jointly using 2D and 3D poses.", + "bbox": [ + 96, + 148, + 467, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 76, + 241, + 220, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Sign language recognition", + "text_level": 1, + "bbox": [ + 76, + 266, + 310, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Word-level sign language recognition task aims to classify short-term videos by the corresponding word label. Previous works on sign language recognition can be mainly categorized into two groups by the input modality: videobased and pose-based approaches.", + "bbox": [ + 76, + 290, + 467, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-based sign language recognition. With the significant performance of the CNN-based action recognition, the majority of the prior works have designed their models based on CNN backbone [9, 27, 53]. Primary works [11, 12, 26, 44] first used convolutional neural networks to extract the frame-wise features. Then, such methods encoded the temporal semantics along the time dimension using recurrent neural networks. On the other hand, similar to action recognition, 3D CNNs are widely used in sign language recognition. C3D [21, 23, 46] is the first 3D CNN model for action recognition, and it is widely used as a baseline model for quantitative comparison. I3D [7] is proposed based on the C3D [46] architecture and applied on sign language recognition in [22]. Li et al. [30] have compared various architectures, including 3D CNN and 2D CNN combined with RNN. Hosain et al. [19] proposed Fusion-3 model based on I3D Inception-v1 [7] with pose-guided pooling.", + "bbox": [ + 76, + 366, + 467, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose-based sign language recognition. Compared to the RGB video, the human pose sequence holds a compact and semantic-aware representation of human motion. ST-GCN [50] is the first pose-based approach to action recognition, utilizing the spatio-temporal graph convolutional network to encode the motion on kinetic chain. Following works also have used the spatio-temporal architecture based on graph convolutional network [30, 47] and Transformers [49] [3, 20].", + "bbox": [ + 76, + 638, + 467, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D human pose and shape estimation", + "text_level": 1, + "bbox": [ + 76, + 786, + 397, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D human pose and shape estimation aims to localize 3D coordinates of vertices of the human mesh. Considering the various applications such as VR/AR and human action recognition, research on 3D human pose and shape estimation has been widely conducted. Most works tried to reconstruct 3D human pose and shape for specific part of human,", + "bbox": [ + 76, + 810, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "e.g. body [10, 34, 38], hand [8, 28], or face [43]. Several works [36, 40] proposed methods that simultaneously estimate the 3D pose and shape for all human parts, i.e., whole-body. By reconstructing 3D poses and shapes for all human parts, we can understand human intention and feelings more precisely. In this work, we leverage Hand4Whole [36] to acquire a 3D pose of all human parts from RGB videos. To the best of our knowledge, we first utilize 3D pose for sign language recognition.", + "bbox": [ + 500, + 90, + 890, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Transformers for human motion.", + "text_level": 1, + "bbox": [ + 500, + 238, + 785, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After the remarkable success of Transformers in natural language processing (NLP) [49], vision researchers have employed Transformers for various vision tasks [6, 13, 34]. Transformers also have been widely adopted for tasks that require human motion understanding, such as 2D-to-3D pose sequence lifting [33, 51, 52] and skeleton-based action recognition [16, 42]. In the field of 2D-to-3D pose sequence lifting, PoseFormer [52] proposed a pure Transformer-based approach for 2D-to-3D pose sequence lifting. MixSTE [51] utilized separated joint-wise temporal Transformer and spatial Transformer alternately. MH-Former [33] employed Transformers to generate and refine multiple hypotheses for 3D pose sequences from 2D pose sequences. In the field of skeleton-based action recognition, ST-TR [42] proposed Transformer-based two-stream networks where each stream learns spatial and temporal correlations in a skeleton sequence. FG-STFormer [16] presented a variation of vanilla Transformer to capture local and global information effectively. Our P3D also leverages Transformer-based modules to learn the human motion. However, most previous works only focused on the motion of the human body, while proposed P3D learns the motion of all human parts, including hands and facial expressions.", + "bbox": [ + 498, + 262, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 638, + 589, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 665, + 607, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the input data format and architecture of our P3D model. Figure 3 illustrates the overview of our model. We provide a sequence of human poses to our model as an input P3D leverages alternating structure of Transformer layers [49], inspired by the spatiotemporal architectures from the previous works [2,51]. The first layer, the part-wise encoding Transformer (PET), is designed to learn intra-part motion context. To capture the specific motion context from each part, we assign the dedicated Transformers. Meanwhile, the second whole-body encoding Transformer (WET) encodes the inter-part motion context, taking account of every human part used in the model. As the last step, we take mean along the time axis on encoded representations, then sequentially pass through", + "bbox": [ + 498, + 689, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20742", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/45217de8a3a149ffce1ddca5e57d5d2d2ab4054588dc19503294084dbe976709.jpg", + "image_caption": [ + "Figure 3. Overall architecture of P3D. Our P3D aims to predict the word label corresponding to a given pose sequence. We first provide joint-wise linearly embedded feature vectors as an input, along with linearly encoded facial expression features. P3D consists of $2N$ alternating layers of PET, which are assigned for each part, and WET, which covers entire parts. After the encoder, the prediction head takes the mean along the time dimension, then sequentially passes through batch normalization, linear layer, and softmax layer to produce the word prediction." + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 893, + 359 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/08a61810a0b220b36d9a35453cd2adb752e3ed6af2523e9d86f3943f2745b207.jpg", + "image_caption": [ + "Figure 4. Overall architectures of PET and WET. We adapt the Transformer encoder layer [49] with vanilla self-attention for both PET (above) and WET (below). For the PET, the subscript $p \\in \\{B, RH, LH, \\psi\\}$ denotes the assigned human part." + ], + "image_footnote": [], + "bbox": [ + 81, + 478, + 467, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "batch normalization, linear layer, and softmax layer to get the final output.", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Input format", + "text_level": 1, + "bbox": [ + 500, + 468, + 635, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human motion representation. The overview of the input data structure is illustrated on the left of Figure 3. As a pose-based sign language model, our P3D takes a sequence $(X_{1},\\dots,X_{T})\\in \\mathbb{R}^{T\\times (J\\times F + E)}$ of $T$ human poses as an input. Pose representation $X_{t} = (\\mathbf{x}_{1},\\dots,\\mathbf{x}_{J},e)\\in \\mathbb{R}^{J\\times F + E}$ for each time $t\\in \\{1,2,\\dots,T\\}$ consists of joint-wise pose representation $\\mathbf{x}_j\\in \\mathbb{R}^F$ for each joint $j\\in \\{1,2,\\dots,J\\}$ , and facial expression feature $e\\in \\mathbb{R}^E$ [41]. From our experiment on the pose fusion method, we have observed that joint-wise concatenation of the 2D positional pose (2D pos.), 3D positional pose (3D pos.), and 3D rotational pose (3D rot.) performed the best on WLASL2000. Thus, we form the joint-wise pose representation $\\mathbf{x}_j = (\\mathbf{x}_{j,2\\mathrm{D}\\mathrm{pos}},\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{pos}},\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{rot}})\\in \\mathbb{R}^F$ where $\\mathbf{x}_{j,2\\mathrm{D}\\mathrm{pos}}\\in \\mathbb{R}^2$ , $\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{pos}}\\in \\mathbb{R}^3$ , $\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{rot}}\\in \\mathbb{R}^6$ represents the 2D positional pose, 3D positional pose and 3D rotational pose [54] of joint $j$ , respectively. Please refer to our experiment on Section 4.3 for the details of pose fusion methods. Following the literature [20], we use $T = 32$ for both training and testing. Joint set and human part. We use $J = 40$ joints and 4 human parts to represent the human motion of the sign language. Ten joints above the pelvis are selected from the SMPL [35] joint set, defined as a body part. Fifteen joints, three per finger, represent each hands. For the face, we use the 10-dimensional feature vector from SMPL-X [41] to represent the facial expression. Please refer to Section 4.2 for the details of human pose preprocessing.", + "bbox": [ + 496, + 492, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20743", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5f00008c5beaeb8e9ca0566a224d0c86d61cd3484bb03538b5c4c09fe1bc2997.jpg", + "image_caption": [ + "Figure 5. Visualization. The 3D expressive pose of the signer (b) is estimated from the given sign language video (a) with off-the-shelf estimator Hand4Whole [36]. The corresponding word label (c) is denoted in the right." + ], + "image_footnote": [], + "bbox": [ + 98, + 97, + 880, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Architecture", + "text_level": 1, + "bbox": [ + 76, + 493, + 212, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linear layer. Our model first linearly encodes each jointwise pose representation $\\mathbf{x}_j \\in \\mathbb{R}^F$ into D dimensional vector. We also assign a dedicated linear layer to encode facial expressions into $\\alpha D$ dimensional vector. Empirically, we decided to use $D = 8$ and $\\alpha = 10$ . As a result, pose $X_t = (\\mathbf{x}_1, \\dots, \\mathbf{x}_J, e) \\in \\mathbb{R}^{J \\times F + E}$ at time $t$ is encoded into $X_t^{(0)} \\in \\mathbb{R}^{(J \\times D + \\alpha D) = ((10 + 15 + 15) \\times 8 + 80)}$ . Note that 10, 15, and 15 represent the number of joints in the body, left hand and right hand, and thus can be separated into each human part via indexing. Finally we form the input $X^{(0)} = (X_1^{(0)}, X_2^{(0)}, \\dots, X_T^{(0)}) \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)}$ with positional encoding [49].", + "bbox": [ + 75, + 516, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alternating Transformers. We adapted the idea of alternating two types of layers [2, 51], which have different purposes and expertise. Our P3D consists of $2N$ alternating layers: $N$ Part-wise Encoding Transformer (PET) and $N$ Whole-body Encoding Transformer (WET). Starting from linear encoded input $X^{(0)}$ , we denote the output of $i$ 'th PET as $Y^{(i)}$ and the output of $i$ 'th WET as $X^{(i)}$ . As the dimension remains identical through Transformers, our features $X^{(i)}$ and $Y^{(i)}$ remain in a same dimension as $X^{(0)}$ . Empirically, we use $N = 3$ layers for WLASL.", + "bbox": [ + 75, + 703, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Part-wise encoding Transformer (PET). The architecture of PET is illustrated in the upper part of Figure 4. From previous layer, PET takes 2D tensor $X^{(i)} \\in$", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbb{R}^T \\times ((10 + 15 + 15) \\times 8 + 80)$ as an input. The first (horizontal) and second (vertical) dimensions represent the time and feature axis, respectively. We first separate features for each human part $X_B^{(i)} \\in \\mathbb{R}^{T \\times 80}$ , $X_{LH}^{(i)} \\in \\mathbb{R}^{T \\times 120}$ , $X_{RH}^{(i)} \\in \\mathbb{R}^{T \\times 120}$ , $X_{\\psi}^{(i)} \\in \\mathbb{R}^{T \\times 80}$ via slicing on second (feature) dimension by indices (80, 200, 320, 400), respectively. Then we pass each feature through dedicated Transformer encoder layers $\\mathrm{PET}_B$ , $\\mathrm{PET}_{RH}$ , $\\mathrm{PET}_{LH}$ , and $\\mathrm{PET}_{\\psi}$ assigned for each human part, respectively. Finally, we concatenate the outputs along the second (feature) dimension into a single 2D tensor $Y^{(i)} \\in \\mathbb{R}^T \\times ((10 + 15 + 15) \\times 8 + 80)$ .", + "bbox": [ + 496, + 491, + 893, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whole-body encoding Transformer (WET). The architecture of WET is illustrated in the lower part of Figure 4. From the previous layer PET, WET takes 2D tensor $Y^{(i)} \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)}$ as an input. Since WET aims to learn inter-part motion context, we assign a single Transformer to encode the whole motion sequence. Output of WET $X^{(i + 1)} \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)}$ is passed to next layer without modification.", + "bbox": [ + 496, + 669, + 893, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prediction head. From Transformer-encoded feature $X^{(N)}\\in \\mathbb{R}^{T\\times 400}$ , we first calculate the mean feature vector $\\mathbf{x}_{\\mathrm{feat}}\\in \\mathbb{R}^{400}$ by taking the mean along the first (time) dimension. Then, we pass $\\mathbf{x}_{\\mathrm{feat}}$ sequentially through 1D batch normalization, a single FC-layer, and a softmax layer. As a result, our model produces the final prediction $\\mathbf{x}_{\\mathrm{class}}\\in \\mathbb{R}^C$ , where $C$ is the number of target classes (glosses) in the", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20744", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset.", + "bbox": [ + 76, + 92, + 130, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and inference. We train P3D with a cross-entropy loss between predicted $\\mathbf{x}_{\\mathrm{class}}$ and one-hot encoded GT class vector. We randomly sample continuous $T$ frames to form the training inputs. During the inference phase, we sample four chunks of $T$ continuous frames from the human pose sequence in the test set following the previous work [30]. Then, we use P3D to generate the class prediction $\\mathbf{x}_{\\mathrm{class}}$ for each $T$ continuous frame, and average them out to produce the final prediction for the entire sequence.", + "bbox": [ + 76, + 106, + 468, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 257, + 200, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 282, + 290, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used PyTorch [39] Transformer implementation for our PET and WET, with 256 inner dimensions, 0.3 dropout rate, GELU activation layer, and 4 multi-head attentions. We trained our model for 500 epochs using Adam [24] optimizer with a mini-batch size of 512, a learning rate of $5 \\times 10^{-4}$ , the weight decay rate of $5 \\times 10^{-3}$ .", + "bbox": [ + 76, + 306, + 468, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets, evaluation metrics and preprocessing", + "text_level": 1, + "bbox": [ + 76, + 407, + 467, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WLASL dataset. WLASL [30] is a large-scale word-level sign language dataset that contains 21,083 videos with 119 signers. The videos in WLASL are collected from the Internet, and each video depicts a signer performing a single sign. The WLASL dataset consists of 2,000 words, and it also provides subsets of the dataset, WLASL100, and WLASL300. Nevertheless, our main target is to push through the limit of recognition performance on WLASL2000.", + "bbox": [ + 76, + 431, + 467, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics For the evaluation metrics, we employ Top-1 and Top-5 accuracy. Following the previous work [20], we calculate both metrics under two different settings, i.e., per-instance and per-class. In per-instance setting, we measure the percentage of accurately classified video over the entire test instances. Per-class setting is also utilized since each class (gloss) has a different number of instances. In a per-class setting, we calculate the accuracy of each class and average them over all classes.", + "bbox": [ + 76, + 568, + 467, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose preprocessing Regarding the 2D pose, we utilize the dedicated 2D pose estimator for better accuracy, since direct estimation of 2D pose is more accurate compared to projection from estimated 3D pose. The 2D pose is aligned based on the bounding box of the signer by applying the affine transform. After then, we rescaled the 2D pose sequence from the bounding box into the [0, 1] range. The 3D poses, including the positional pose, the rotational pose and the facial expression, we utilize the off-the-shelf 3D expressive human pose estimator Hand4Whole [36] as illustrated in Figure 5. For the 3D positional pose, we aligned the translation by placing the root joint of the first frame at the origin without normalization. In addition, the 3D rotational", + "bbox": [ + 76, + 704, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pose and the facial expression features are used without preprocessing.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 133, + 648, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct an ablation study on two main contributions of our framework: part-wise motion context learning and pose ensemble. For each experiment, the contribution that is not being tested remains identical to our proposed P3D.", + "bbox": [ + 500, + 157, + 890, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of part-wise motion context learning. In Table 1, we compare the results of various pose encoding methods in single iteration of P3D. The MLP is the simple alternative to the proposed PET or WET, which is constructed of stacked fully-connected layers, batch normalization, and dropout. Experimental results demonstrate that employing our PET and WET gives a large performance gain over MLP alternatives.", + "bbox": [ + 500, + 219, + 890, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also report that a combination of PET and WET is essential in sign language recognition. When the number of classes is small (WLASL 100, WLASL 300), utilizing one of the WET (the fourth row) or PET (the third row) provides a similar result to utilizing both PET and WET (the fifth row). However, when the number of classes is large (WLASL 2000), utilizing both PET and WET performs the best. As the number of classes increases, learning highly distinguishable features from motion context becomes crucial. For that, the PET and WET are dedicated to capturing intra-part context and merging them into a unified inter-part context, respectively. Such design benefits the task-specific representation power on sign language recognition, that each part (i.e. body, hands, and facial expressions) holds its own important motion context. As a result, our PET-WET encoder outperforms PET-PET and WET-WET by $3.75\\%$ and $12.08\\%$ , respectively.", + "bbox": [ + 500, + 340, + 890, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of each human part. In Table 2, we conduct an ablation study on employing different human part inputs. As our task is sign language recognition, hands are the most essential part of accurate recognition. However, the additional gain can be obtained by leveraging body pose and facial expression on top of hands. It demonstrates that body pose and facial expression, as well as hands, also provide helpful information for sign language recognition.", + "bbox": [ + 500, + 598, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of each pose representation. Table 3 shows that jointly utilizing 2D positional pose, 3D positional pose, and 3D rotational pose achieves the best performance. In terms of single representation, the 3D rotational pose is the most effective pose representation for sign language recognition. Furthermore, additionally using the 3D rotational pose to other pose representations (the first row & the fifth row, the second row & the sixth row, and the fourth row & last row) consistently improves the accuracy. As a reason for the effectiveness of 3D rotational pose for sign language recognition, we argue that 3D rotational pose is less sensitive to signer variation than 2D positional pose and 3D positional", + "bbox": [ + 500, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20745", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/407965c4d30116382a1155174bb29da0d7ad7dedae9f09e029549ab52655ca96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pose encoding methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
MLP-WET71.8985.9472.3387.2159.4482.2059.7082.4334.9770.5732.7869.41
MLP-PET72.2888.7673.7689.5557.8982.5058.5682.9634.7270.7132.4969.66
PET-PET77.9194.7778.3095.2567.3388.6868.0688.0632.3963.6130.3462.20
WET-WET75.5091.5776.4392.6866.8788.2466.9288.5640.7274.7038.2673.54
PET-WET (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
", + "bbox": [ + 96, + 89, + 870, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/00aeb808166628aa9f8243a800addc05a6e232c6017dc1a610c6854a639c5c65.jpg", + "table_caption": [ + "Table 1. Ablation study on pose encoding methods. We ablate the performance of pose encoding methods in single iteration of P3D. Our part-wise context learning (PET-WET) scored the best performance. The observation that ours (PET-WET) has outperformed WET-WET (Transformer without part) directly supports our main contribution of part-wise context learning." + ], + "table_footnote": [], + "table_body": "
Input human partsWLASL 100WLASL 300WLASL 2000
BodyHandsExpr.Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
✓44.9874.7044.7575.8537.6270.4337.4070.1717.8040.7615.7037.57
✓✓46.9972.2945.5772.4338.7070.4338.7370.4520.9847.0518.7343.94
✓71.8991.1673.1092.0862.0787.1562.6987.9339.0073.4036.9772.16
✓✓75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
", + "bbox": [ + 93, + 263, + 874, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on input human parts for P3D. We ablate the performance of provided input human parts. Check marks in the first column implies the given input part. We abbreviate the facial expression to Expr.", + "bbox": [ + 75, + 385, + 892, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cee37f4d5c5b01a574e2559128833b559f9c77fe73c8a90f319818629e00b37f.jpg", + "image_caption": [ + "Figure 6. t-SNE [48] visualization of features from different pose representations. Each plot corresponds to a single word (\"Corn\", \"Language\", \"Many\", and \"Thanksgiving\", respectively), and each point is feature $X^{(N)}$ from each video. All test videos of the corresponding word in WLASL 2000 are employed for the visualization." + ], + "image_footnote": [], + "bbox": [ + 81, + 446, + 464, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pose. Figure 6 illustrates that features of single words from different signers are closely embedded when input pose representation is a 3D rotational pose. On the other hand, in the case of the 2D and 3D positional poses, the embedded features of the single words show large discrepancies when the", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "signers are different. Such advantage of 3D rotational pose leads to improved performance over other pose representations since the sign languages in the WLASL dataset are signer-independent [30].", + "bbox": [ + 496, + 438, + 890, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pose ensemble methods. In Table 4, we compare the performances of different pose ensemble methods. Since we employ three types of pose representations (2D positional pose, 3D positional pose, and 3D rotational pose), how to ensemble information in each pose representation is an important issue. We design three strategies for ensemble methods, and the description for each ensemble method is as follows.", + "bbox": [ + 496, + 500, + 892, + 619 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Early ensemble (Ours): All pose representations of each joint are concatenated before being fed to the PET-WET blocks. Thus, PET-WET blocks the process motion context of all poses representations simultaneously.", + "- Middle ensemble: Separate PET-WET blocks are utilized to extract feature $X^{(n)}$ for each pose representation: $X_{2\\mathrm{D}\\mathrm{pos}}^{(n)}$ , $X_{3\\mathrm{D}\\mathrm{pos}}^{(n)}$ , and $X_{3\\mathrm{D}\\mathrm{rot}}^{(n)}$ . The extracted features of all pose representations are averaged before being fed to the prediction head.", + "- Late ensemble: Same as the middle ensemble, but each feature of each pose representation is fed to its own prediction head. The output of all prediction heads is averaged to regress a single word." + ], + "bbox": [ + 517, + 628, + 890, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results show that Early ensemble performs best for the WLASL 2000. In Table 5, we compare", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20746", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/784aa2c91181ba8717ff92773b03211a6cad7e6b0b68a610b876886212be2e18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input pose representationsWLASL 100WLASL 300WLASL 2000
2D pos.3D pos.3D rot.Per-instancePer-classPer-instancePer-classPer-instancePer-class
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
✓60.2483.1360.4582.6352.1779.7252.5979.7935.9967.2633.4765.15
✓41.0966.6740.9867.0239.9769.7640.0170.0131.6962.8228.4259.83
✓72.8788.3773.6088.9361.6882.3461.9382.7838.6073.4235.5771.86
✓✓55.0281.1255.7382.0150.1580.1949.7481.0936.9169.0934.0067.06
✓✓75.5091.1676.8891.5865.9488.8566.2688.7543.4877.6441.3976.38
✓✓74.4288.3775.1589.2662.7283.5363.4384.0141.1476.7938.4775.34
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
", + "bbox": [ + 93, + 89, + 874, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/01878a6e6eed9d73dc1d232c03a2ed687d34b35db097d7ddfa174cf99b780a75.jpg", + "table_caption": [ + "Table 3. Ablation study on input pose representations. We ablate the performance of provided input human pose representations. Check marks in the first column implies the given input pose representation. Employing every pose representation scored the best result." + ], + "table_footnote": [], + "table_body": "
Ensemble methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
Late77.9093.1778.3095.2567.3488.2368.0688.6832.7164.0430.3463.67
Middle79.9293.5781.6694.3867.8089.9467.8690.1342.6371.7140.1776.43
Early (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
", + "bbox": [ + 96, + 277, + 870, + 363 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7365ddd32041143b3c3b793a994affea9e3f11f16607fe93ea032641898d6103.jpg", + "table_caption": [ + "Table 4. Ablation study on pose representation ensemble methods. We report the quantitative result on pose representation ensemble methods. Early ensemble by joint-wise concatenation outperforms the middle (between encoder & head) and late (after head) ensemble." + ], + "table_footnote": [], + "table_body": "
Ensemble methodsFLOPs (G)# Params (M)
Late182.0610.71
Middle182.0210.65
Early (Ours)83.664.94
", + "bbox": [ + 81, + 422, + 465, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Computational costs of pose ensemble methods. We used the FLOPs counter for PyTorch [39] from fvcore package to measure the computational costs of the models.", + "bbox": [ + 76, + 523, + 468, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the computational costs of ensemble methods. As separate PET-WET blocks are utilized in Middle ensemble and Late ensemble, Early ensemble is most efficient in terms of the number of model parameters and FLOPs. Considering the accuracy and efficiency, we adopt Early ensemble as our ensemble strategy.", + "bbox": [ + 75, + 590, + 468, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 76, + 694, + 436, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 shows that P3D outperforms previous pose-based methods. We first demonstrate the capability of our proposed part-wise motion context learning model via quantitative comparison on the WLASL dataset using only 2D pose. Then we show that 3D pose fusion is more effective in our P3D than in other models, leading to our model outperforming others on the WLASL dataset using 2D and 3D pose.", + "bbox": [ + 75, + 718, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using only 2D pose. The second block of Table 6 shows that our P3D outperforms existing methods on WLASL 300 and WLASL 2000 split with 2D pose input. In this experiment, we only use 2D pose for the fair comparison of", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models to learn highly distinguishable features from motion context. Prior to our work, state-of-the-art methods were SPOTER [3] in WLASL 100, and ST-GCN [50] in WLASL 300 and WLASL 2000. From the benefits of its simple and intuitive Transformer architecture, SPOTER achieved the best performance on WLASL 100. However, due to the limitation on representation power, SPOTER did not perform well on more complex problems such as WLASL 300 and WLASL 2000. Meanwhile, graph convolution via spatiotemporal adjacency used in ST-GCN is observed to be effective on a large number of classes (WLASL 300 and WLASL 2000). Yet, employing the spatio-temporal adjacency graph leads to two cons: it cannot model the long-term dependency, and cannot learn the part-specific context. In contrast, the proposed part-aware Transformer architecture of our P3D is capable of handling the long-term dependency and part-specific context at the same time. As a result, our P3D shows better performance compared to ST-GCN on every split of WLASL, proving the better representation power of proposed part-wise motion context learning.", + "bbox": [ + 496, + 424, + 890, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Jointly using 2D and 3D pose. The third block of Table 6 shows that our P3D scores highly superior results compared to previous methods with using both 2D and 3D poses as an input. Again, in order to conduct fair comparison, identical input features are used and the same pose ensemble method is used for every tested models. We performed a comparative analysis in two aspects as follows.", + "bbox": [ + 496, + 729, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, our P3D resulted in a much higher performance gain from jointly using 2D and 3D poses. While our P3D gained $6.39\\%$ top-1 per-instance recognition accuracy on WLASL2000 by pose ensemble, gains of previous state", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20747", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/859c141ede4c5662590a1b9aa14ba5aeb1e37ab0702cdfcf5c7916e7ec862af4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
RGB-based
TK-3D ConvNet [17]77.55---68.75-------
Fusion-3 [19]75.67---68.3-------
I3D [7]65.8984.1167.0184.5856.1479.9456.2478.3832.4857.31--
Pose-based (2D)
Pose-TGCN [30]55.4378.68--38.4267.51--23.6551.75--
GCN-BERT [15]60.1583.98--42.1871.71------
SPOTER [3]63.18---43.18---28.8661.7828.2160.26
ST-GCN [50]50.7879.0751.6279.4744.4673.0545.2973.1634.4066.4732.5365.45
P3D (Ours)60.2485.9460.4885.6650.9377.7150.8178.5535.9668.3133.2266.20
Pose-based (2D + 3D)
SPOTER [3]73.9091.5774.5191.5554.0282.2055.2882.6430.9164.2529.8862.7
ST-GCN [50]56.6379.5156.9379.7145.8276.0145.9676.2237.3772.1335.6170.66
P3D (Ours)75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
Pose-based (2D + 3D + facial expression)
P3D (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
", + "bbox": [ + 84, + 88, + 880, + 368 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 6. Comparisons with state-of-the-art methods. We report the recognition accuracy of previous methods for comparison with ours. The results in each block are trained and evaluated with using exact same input denoted in the first column. Results in the first block show the recognition accuracy of existing RGB-based approaches. The second block represents the pose-based approach with 2D pose inputs. Our result in the second block is also generated solely with 2D poses. Scores in the third block are obtained from employing pose ensemble on previous methods and ours, using both 2D and 3D poses. Note that previous methods did not utilize 3D pose in their works, thus we applied pose ensemble based on released codes. Finally, the fourth block shows the best performance of our model using 2D, 3D poses and facial expression. Blank scores denoted with a hyphen are not reported in previous works.", + "bbox": [ + 75, + 377, + 890, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "of-the-art methods SPOTER [3] and ST-GCN [50] are confined to $2.05\\%$ and $2.97\\%$ , respectively. Such observation suggests that our part-wise motion context learning is more capable of acquiring contextual information from various types of pose representations, which distinguishes it from the basic Transformer architecture of SPOTER [3] or the spatio-temporal graph convolution of ST-GCN [50].", + "bbox": [ + 75, + 500, + 468, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Second, P3D achieved superior performance compared to existing methods, particularly by a large margin on a large number of classes such as WLASL300 and WLASL2000. As demonstrated in the third block of Table 6, our P3D outperformed previous pose-based methods SPOTER [3] and ST-GCN [50] on every subset of WLASL.", + "bbox": [ + 75, + 609, + 468, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Using 3D expressive pose. The fourth block of Table 6 shows the performance with full input features: 2D pose, 3D positional and rotational pose, and facial expression. With additional usage of facial expression over the third block of Table 6, we observed the performance gain of $0.79\\%$ , $0.93\\%$ , and $2.12\\%$ , respectively for each subset of WLASL. Moreover, when compared to the RGB-based models presented in the first block of Table 6, our P3D scored comparable results in WLASL100 and WLASL300, and outperformed in WLASL2000. Our framework is the first to outperform (WASL 2000) or achieve comparable performance (WASL 100, WASL 300) with RGB-based methods.", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 497, + 617, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we proposed the sign language recognition framework P3D by leveraging part-wise 3D motion context learning. We utilize the alternating PET and WET layers, specializing in learning intra-part and inter-part motion contexts, respectively. Learning part-wise motion context resulted in superior performance on the WLASL benchmark compared to previous methods. Moreover, we jointly utilized 2D and 3D poses via joint-wise pose concatenation to fully exploit the performance of pose-based sign language recognition. We observed that joint-wise concatenation has advantages on both sides of performance and efficiency. As a result, our P3D outperformed the previous methods by a large margin on WLASL, jointly using 2D and 3D poses.", + "bbox": [ + 496, + 532, + 890, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 773, + 666, + 791 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], the NRF grant [No. 2021M3A9E4080782] funded by the Korea government (MSIT), and the SNU-Samsung SDS Research Center.", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20748", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In ECCV, 2020. 1", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 3, 5", + "[3] Matyáš Boháček and Marek Hrúz. Sign pose-based transformer for word-level sign language recognition. In WACV, 2022. 2, 3, 8, 9", + "[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, 2020. 1", + "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 1", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 3, 9", + "[8] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 3", + "[9] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV, 2020. 3", + "[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 3", + "[11] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In CVPR, 2017. 3", + "[12] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia (ToM), 21(7):1880-1891, 2019. 3", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[14] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1", + "[15] Haodong Duan, Yue Zhao, Kai Chen, Dian Shao, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 9", + "[16] Zhimin Gao, Peitao Wang, Pei Lv, Xiaoheng Jiang, Qidong Liu, Pichao Wang, Mingliang Xu, and Wanqing Li. Focal" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and global spatial-temporal transformer for skeleton-based action recognition. In ACCV, 2022. 3", + "[17] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In CVPR, 2017. 9", + "[18] Xiao Guo and Jongmoo Choi. Human motion prediction via learning local structure representations and temporal dependencies. In AAAI, 2019. 2", + "[19] Al Amin Hosain, Panneer Selvam Santhalingam, Parth Pathak, Huzefa Rangwala, and Jana Kosecka. Hand pose guided 3d pooling for word-level sign language recognition. In WACV, 2021. 3, 9", + "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: pre-training of hand-model-aware representation for sign language recognition. In ICCV, 2021. 1, 2, 3, 4, 6", + "[21] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding, 2014. 3", + "[22] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. In BMVC, 2018. 3", + "[23] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2014. 3", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[25] Oscar Koller. Quantitative survey of the state of the art in sign language recognition, 2020. 1", + "[26] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(9):2306-2320, 2019. 3", + "[27] Oscar Koller, Sepehr Zargaran, Hermann Ney, and Richard Bowden. Deep sign: Enabling robust statistical continuous sign language recognition via hybrid cnn-hmms. International Journal of Computer Vision (IJCV), 126:1311-1325, 2018. 3", + "[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 3", + "[29] Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Cooccurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation. arXiv preprint arXiv:1804.06055, 2018. 1", + "[30] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, 2020. 1, 2, 3, 6, 7, 9" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20749", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In CVPR, 2020. 1", + "[32] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-parted graph scattering networks for 3d human motion prediction. In ECCV, 2022. 2", + "[33] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In CVPR, 2022. 3", + "[34] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 3", + "[35] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, Oct. 2015. 4", + "[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR Workshops, 2022. 1, 3, 5, 6", + "[37] Gyeongsik Moon, Heeseung Kwon, Kyoung Mu Lee, and Minsu Cho. Integralaction: Pose-driven feature integration for robust human action recognition in videos. In CVPRW, 2021. 1", + "[38] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 3", + "[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. *NeurIPS Workshop on Autodiff*, 2017. 6, 8", + "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3", + "[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019, 2, 4", + "[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In ICPR Workshops, 2021. 3", + "[43] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 3", + "[44] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3d multi-channel sign language production via progressive transformers and mixture density networks. International Journal of Computer Vision (IJCV), 129(7):2113-2135, 2021. 3", + "[45] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In AAAI, 2017. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Du Tran, Lubomir Bourdev, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. ICCV, 2015. 3", + "[47] Anirudh Tunga, Sai Vidyaranya Nuthalapati, and Juan Wachs. Pose-based sign language recognition using gcn and bert. In WACV, 2021. 1, 2, 3", + "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 2008. 7", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3, 4, 5", + "[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3, 8, 9", + "[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2, 3, 5", + "[52] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 3", + "[53] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 3", + "[54] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4" + ], + "bbox": [ + 501, + 92, + 890, + 530 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20750", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_model.json b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28894409a28598a1613a79333bf415692402e7f3 --- /dev/null +++ b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_model.json @@ -0,0 +1,2026 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.247, + 0.131, + 0.726, + 0.177 + ], + "angle": 0, + "content": "Human Part-wise 3D Motion Context Learning for Sign Language Recognition" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.203, + 0.374, + 0.222 + ], + "angle": 0, + "content": "Taeryung Lee" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.204, + 0.525, + 0.222 + ], + "angle": 0, + "content": "Yeonguk Oh²" + }, + { + "type": "text", + "bbox": [ + 0.565, + 0.204, + 0.715, + 0.222 + ], + "angle": 0, + "content": "Kyoung Mu Lee\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.222, + 0.769, + 0.24 + ], + "angle": 0, + "content": "1IPAI, 2Dept. of ECE & ASRI, Seoul National University, Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.242, + 0.662, + 0.256 + ], + "angle": 0, + "content": "{trlee94, namepllet, kyoungmu}@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.673 + ], + "angle": 0, + "content": "In this paper, we propose P3D, the human part-wise motion context learning framework for sign language recognition. Our main contributions lie in two dimensions: learning the part-wise motion context and employing the pose ensemble to utilize 2D and 3D pose jointly. First, our empirical observation implies that part-wise context encoding benefits the performance of sign language recognition. While previous methods of sign language recognition learned motion context from the sequence of the entire pose, we argue that such methods cannot exploit part-specific motion context. In order to utilize part-wise motion context, we propose the alternating combination of a part-wise encoding Transformer (PET) and a whole-body encoding Transformer (WET). PET encodes the motion contexts from a part sequence, while WET merges them into a unified context. By learning part-wise motion context, our P3D achieves superior performance on WLASL compared to previous state-of-the-art methods. Second, our framework is the first to ensemble 2D and 3D poses for sign language recognition. Since the 3D pose holds rich motion context and depth information to distinguish the words, our P3D outperformed the previous state-of-the-art methods employing a pose ensemble." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.706, + 0.21, + 0.722 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Understanding and translating sign language is essential due to the huge potential benefit to society. Sign language recognition (SLR) can be considered as a specific kind of action recognition (AR) since both tasks aim to classify human motion. However, SLR has a characteristic that is distinguished from AR; Actions mainly focus on the motion of the human body, while the SLR system should be aware of more detailed information such as hand gestures and facial expressions. Thus, a specified system for SLR will provide further gain over using the AR models on SLR." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Previous approaches on SLR problem are mainly cat-" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.301, + 0.889, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.684 + ], + "angle": 0, + "content": "Figure 1. Overall pipeline of proposed system. We first extract expressive 3D human pose from RGB video using off-the-shelf pose estimation methods [5, 36]. The expressive 3D human pose consists of 2D pose, 3D pose and facial expression. Then our proposed P3D predicts the word from the expressive 3D human pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.895, + 0.807 + ], + "angle": 0, + "content": "egorized into RGB-based [1, 4, 31] and pose-based methods [25, 29, 45, 47, 50]. Pose-based methods are known to have the advantage over RGB-based approaches in terms of the robustness on domain gap [14, 37]. For that, we aimed to enhance the performance of SLR, while preserving the benefits inherent in the pose-based approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Although recent pose-based SLR methods have shown noticeable progress, there are two main limitations of existing methods. First, they utilize only whole-body representation (e.g., skeleton) to understand the motion context of human [20, 30, 47]. Since such approaches overlook the detailed motion context within a specific part" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20740" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.168, + 0.173, + 0.184 + ], + "angle": 0, + "content": "RGB Video" + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.262, + 0.152, + 0.277 + ], + "angle": 0, + "content": "2D Pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.355, + 0.152, + 0.369 + ], + "angle": 0, + "content": "3D Pose" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.098, + 0.511, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.097, + 0.888, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.447, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Figure 2. Effectiveness of 3D pose for sign language recognition. The 2D poses of the words \"Avoid\" and \"Behind\" are almost identical. Due to the absence of depth information, solely relying on 2D pose makes it hard to distinguish the words. On the other hand, as shown in the red boxes, additionally employing the 3D pose with rich depth information can effectively resolve the ambiguity in sign language recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.47, + 0.618 + ], + "angle": 0, + "content": "of human [18, 32], it leads to a decrease in SLR accuracy. Second, existing pose-based methods only employ 2D pose [3, 20, 30, 47], which induces ambiguities in distinguishing the different words. As shown in Figure 2, the 2D pose of different words can be almost identical due to the absence of depth information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To overcome the above limitations, we propose P3D, the human part-wise motion context learning framework for accurate sign language recognition. The proposed P3D employs the part-wise encoding Transformer (PET) and whole-body encoding Transformer (WET) as main components. Given the pose sequence, we first divide it into four parts: body, right hand, left hand, and face. In contrast to previous methods where entire human poses are jointly processed, we utilize part-specific PET to encode the motion context of each part. As each PET takes a pose sequence of a single part, it can learn more detailed motion context, which is pivotal for sign language recognition. Then, the encoded features from each part are merged by WET. While PET only considers the intra-part motion context, WET covers the entire part. For the effective fusion of the inter-part and intra-part motion context, we alternately combined PET and WET inspired by spatio-temporal Transformers [2,51]. As a result, our proposed P3D has out" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.633 + ], + "angle": 0, + "content": "performed the previous state-of-the-art [3, 50] on WLASL. Our ablation study on layer design supports part-wise motion context learning benefits SLR performance. Especially, we observed that solely forming the model with WET (i.e., without part-wise motion context) degrades the performance, thus directly induces that our PET contributed to improving the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.893, + 0.844 + ], + "angle": 0, + "content": "We alleviate the second limitation of previous methods through the ensemble of 2D and 3D poses. In terms of 3D pose, we utilize both 3D positional pose and 3D rotational pose (e.g., SMPL-X [41] pose parameters). The 3D positional pose complements depth information that the 2D pose does not contain, and the 3D rotational pose provides rich motion context on the kinematic chain. Our experiment shows that our P3D outperforms the previous methods by a large margin with using identical inputs; our model shows the superior result from only 2D pose, and also from ensemble of 2D and 3D pose. Moreover, our framework is the first pose-based method that recorded a comparable score or even outperformed the RGB-based methods. We claim our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "- We proposed P3D, the human part-wise motion context learning framework for sign language recognition. Our experiment supports that part-wise learning highly" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20741" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.136 + ], + "angle": 0, + "content": "benefits recognition performance. As a result, our proposed P3D outperformed the previous SOTA on the WLASL benchmark." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.15, + 0.468, + 0.225 + ], + "angle": 0, + "content": "- We employ the pose ensemble by joint-wise pose concatenation in order to exploit both 2D and 3D poses. We observed significant performance gain via pose ensemble, and P3D outperformed the previous methods by a large margin by jointly using 2D and 3D poses." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.242, + 0.221, + 0.256 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.267, + 0.311, + 0.283 + ], + "angle": 0, + "content": "2.1. Sign language recognition" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.291, + 0.468, + 0.366 + ], + "angle": 0, + "content": "Word-level sign language recognition task aims to classify short-term videos by the corresponding word label. Previous works on sign language recognition can be mainly categorized into two groups by the input modality: videobased and pose-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.367, + 0.468, + 0.638 + ], + "angle": 0, + "content": "Video-based sign language recognition. With the significant performance of the CNN-based action recognition, the majority of the prior works have designed their models based on CNN backbone [9, 27, 53]. Primary works [11, 12, 26, 44] first used convolutional neural networks to extract the frame-wise features. Then, such methods encoded the temporal semantics along the time dimension using recurrent neural networks. On the other hand, similar to action recognition, 3D CNNs are widely used in sign language recognition. C3D [21, 23, 46] is the first 3D CNN model for action recognition, and it is widely used as a baseline model for quantitative comparison. I3D [7] is proposed based on the C3D [46] architecture and applied on sign language recognition in [22]. Li et al. [30] have compared various architectures, including 3D CNN and 2D CNN combined with RNN. Hosain et al. [19] proposed Fusion-3 model based on I3D Inception-v1 [7] with pose-guided pooling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.468, + 0.775 + ], + "angle": 0, + "content": "Pose-based sign language recognition. Compared to the RGB video, the human pose sequence holds a compact and semantic-aware representation of human motion. ST-GCN [50] is the first pose-based approach to action recognition, utilizing the spatio-temporal graph convolutional network to encode the motion on kinetic chain. Following works also have used the spatio-temporal architecture based on graph convolutional network [30, 47] and Transformers [49] [3, 20]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.398, + 0.803 + ], + "angle": 0, + "content": "2.2. 3D human pose and shape estimation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "3D human pose and shape estimation aims to localize 3D coordinates of vertices of the human mesh. Considering the various applications such as VR/AR and human action recognition, research on 3D human pose and shape estimation has been widely conducted. Most works tried to reconstruct 3D human pose and shape for specific part of human," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "e.g. body [10, 34, 38], hand [8, 28], or face [43]. Several works [36, 40] proposed methods that simultaneously estimate the 3D pose and shape for all human parts, i.e., whole-body. By reconstructing 3D poses and shapes for all human parts, we can understand human intention and feelings more precisely. In this work, we leverage Hand4Whole [36] to acquire a 3D pose of all human parts from RGB videos. To the best of our knowledge, we first utilize 3D pose for sign language recognition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.239, + 0.787, + 0.254 + ], + "angle": 0, + "content": "2.3. Transformers for human motion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.626 + ], + "angle": 0, + "content": "After the remarkable success of Transformers in natural language processing (NLP) [49], vision researchers have employed Transformers for various vision tasks [6, 13, 34]. Transformers also have been widely adopted for tasks that require human motion understanding, such as 2D-to-3D pose sequence lifting [33, 51, 52] and skeleton-based action recognition [16, 42]. In the field of 2D-to-3D pose sequence lifting, PoseFormer [52] proposed a pure Transformer-based approach for 2D-to-3D pose sequence lifting. MixSTE [51] utilized separated joint-wise temporal Transformer and spatial Transformer alternately. MH-Former [33] employed Transformers to generate and refine multiple hypotheses for 3D pose sequences from 2D pose sequences. In the field of skeleton-based action recognition, ST-TR [42] proposed Transformer-based two-stream networks where each stream learns spatial and temporal correlations in a skeleton sequence. FG-STFormer [16] presented a variation of vanilla Transformer to capture local and global information effectively. Our P3D also leverages Transformer-based modules to learn the human motion. However, most previous works only focused on the motion of the human body, while proposed P3D learns the motion of all human parts, including hands and facial expressions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.59, + 0.655 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.609, + 0.68 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we describe the input data format and architecture of our P3D model. Figure 3 illustrates the overview of our model. We provide a sequence of human poses to our model as an input P3D leverages alternating structure of Transformer layers [49], inspired by the spatiotemporal architectures from the previous works [2,51]. The first layer, the part-wise encoding Transformer (PET), is designed to learn intra-part motion context. To capture the specific motion context from each part, we assign the dedicated Transformers. Meanwhile, the second whole-body encoding Transformer (WET) encodes the inter-part motion context, taking account of every human part used in the model. As the last step, we take mean along the time axis on encoded representations, then sequentially pass through" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20742" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.096, + 0.895, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.375, + 0.895, + 0.446 + ], + "angle": 0, + "content": "Figure 3. Overall architecture of P3D. Our P3D aims to predict the word label corresponding to a given pose sequence. We first provide joint-wise linearly embedded feature vectors as an input, along with linearly encoded facial expression features. P3D consists of \\(2N\\) alternating layers of PET, which are assigned for each part, and WET, which covers entire parts. After the encoder, the prediction head takes the mean along the time dimension, then sequentially passes through batch normalization, linear layer, and softmax layer to produce the word prediction." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.479, + 0.468, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.749, + 0.471, + 0.807 + ], + "angle": 0, + "content": "Figure 4. Overall architectures of PET and WET. We adapt the Transformer encoder layer [49] with vanilla self-attention for both PET (above) and WET (below). For the PET, the subscript \\( p \\in \\{B, RH, LH, \\psi\\} \\) denotes the assigned human part." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "batch normalization, linear layer, and softmax layer to get the final output." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.637, + 0.485 + ], + "angle": 0, + "content": "3.2. Input format" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Human motion representation. The overview of the input data structure is illustrated on the left of Figure 3. As a pose-based sign language model, our P3D takes a sequence \\((X_{1},\\dots,X_{T})\\in \\mathbb{R}^{T\\times (J\\times F + E)}\\) of \\(T\\) human poses as an input. Pose representation \\(X_{t} = (\\mathbf{x}_{1},\\dots,\\mathbf{x}_{J},e)\\in \\mathbb{R}^{J\\times F + E}\\) for each time \\(t\\in \\{1,2,\\dots,T\\}\\) consists of joint-wise pose representation \\(\\mathbf{x}_j\\in \\mathbb{R}^F\\) for each joint \\(j\\in \\{1,2,\\dots,J\\}\\), and facial expression feature \\(e\\in \\mathbb{R}^E\\) [41]. From our experiment on the pose fusion method, we have observed that joint-wise concatenation of the 2D positional pose (2D pos.), 3D positional pose (3D pos.), and 3D rotational pose (3D rot.) performed the best on WLASL2000. Thus, we form the joint-wise pose representation \\(\\mathbf{x}_j = (\\mathbf{x}_{j,2\\mathrm{D}\\mathrm{pos}},\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{pos}},\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{rot}})\\in \\mathbb{R}^F\\) where \\(\\mathbf{x}_{j,2\\mathrm{D}\\mathrm{pos}}\\in \\mathbb{R}^2\\), \\(\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{pos}}\\in \\mathbb{R}^3\\), \\(\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{rot}}\\in \\mathbb{R}^6\\) represents the 2D positional pose, 3D positional pose and 3D rotational pose [54] of joint \\(j\\), respectively. Please refer to our experiment on Section 4.3 for the details of pose fusion methods. Following the literature [20], we use \\(T = 32\\) for both training and testing. Joint set and human part. We use \\(J = 40\\) joints and 4 human parts to represent the human motion of the sign language. Ten joints above the pelvis are selected from the SMPL [35] joint set, defined as a body part. Fifteen joints, three per finger, represent each hands. For the face, we use the 10-dimensional feature vector from SMPL-X [41] to represent the facial expression. Please refer to Section 4.2 for the details of human pose preprocessing." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.098, + 0.881, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.441, + 0.895, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Visualization. The 3D expressive pose of the signer (b) is estimated from the given sign language video (a) with off-the-shelf estimator Hand4Whole [36]. The corresponding word label (c) is denoted in the right." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.214, + 0.509 + ], + "angle": 0, + "content": "3.3. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Linear layer. Our model first linearly encodes each jointwise pose representation \\(\\mathbf{x}_j \\in \\mathbb{R}^F\\) into D dimensional vector. We also assign a dedicated linear layer to encode facial expressions into \\(\\alpha D\\) dimensional vector. Empirically, we decided to use \\(D = 8\\) and \\(\\alpha = 10\\). As a result, pose \\(X_t = (\\mathbf{x}_1, \\dots, \\mathbf{x}_J, e) \\in \\mathbb{R}^{J \\times F + E}\\) at time \\(t\\) is encoded into \\(X_t^{(0)} \\in \\mathbb{R}^{(J \\times D + \\alpha D) = ((10 + 15 + 15) \\times 8 + 80)}\\). Note that 10, 15, and 15 represent the number of joints in the body, left hand and right hand, and thus can be separated into each human part via indexing. Finally we form the input \\(X^{(0)} = (X_1^{(0)}, X_2^{(0)}, \\dots, X_T^{(0)}) \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)}\\) with positional encoding [49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Alternating Transformers. We adapted the idea of alternating two types of layers [2, 51], which have different purposes and expertise. Our P3D consists of \\(2N\\) alternating layers: \\(N\\) Part-wise Encoding Transformer (PET) and \\(N\\) Whole-body Encoding Transformer (WET). Starting from linear encoded input \\(X^{(0)}\\), we denote the output of \\(i\\)'th PET as \\(Y^{(i)}\\) and the output of \\(i\\)'th WET as \\(X^{(i)}\\). As the dimension remains identical through Transformers, our features \\(X^{(i)}\\) and \\(Y^{(i)}\\) remain in a same dimension as \\(X^{(0)}\\). Empirically, we use \\(N = 3\\) layers for WLASL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Part-wise encoding Transformer (PET). The architecture of PET is illustrated in the upper part of Figure 4. From previous layer, PET takes 2D tensor \\( X^{(i)} \\in \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.895, + 0.666 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^T \\times ((10 + 15 + 15) \\times 8 + 80)\\) as an input. The first (horizontal) and second (vertical) dimensions represent the time and feature axis, respectively. We first separate features for each human part \\(X_B^{(i)} \\in \\mathbb{R}^{T \\times 80}\\), \\(X_{LH}^{(i)} \\in \\mathbb{R}^{T \\times 120}\\), \\(X_{RH}^{(i)} \\in \\mathbb{R}^{T \\times 120}\\), \\(X_{\\psi}^{(i)} \\in \\mathbb{R}^{T \\times 80}\\) via slicing on second (feature) dimension by indices (80, 200, 320, 400), respectively. Then we pass each feature through dedicated Transformer encoder layers \\(\\mathrm{PET}_B\\), \\(\\mathrm{PET}_{RH}\\), \\(\\mathrm{PET}_{LH}\\), and \\(\\mathrm{PET}_{\\psi}\\) assigned for each human part, respectively. Finally, we concatenate the outputs along the second (feature) dimension into a single 2D tensor \\(Y^{(i)} \\in \\mathbb{R}^T \\times ((10 + 15 + 15) \\times 8 + 80)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.895, + 0.79 + ], + "angle": 0, + "content": "Whole-body encoding Transformer (WET). The architecture of WET is illustrated in the lower part of Figure 4. From the previous layer PET, WET takes 2D tensor \\( Y^{(i)} \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)} \\) as an input. Since WET aims to learn inter-part motion context, we assign a single Transformer to encode the whole motion sequence. Output of WET \\( X^{(i + 1)} \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)} \\) is passed to next layer without modification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Prediction head. From Transformer-encoded feature \\(X^{(N)}\\in \\mathbb{R}^{T\\times 400}\\), we first calculate the mean feature vector \\(\\mathbf{x}_{\\mathrm{feat}}\\in \\mathbb{R}^{400}\\) by taking the mean along the first (time) dimension. Then, we pass \\(\\mathbf{x}_{\\mathrm{feat}}\\) sequentially through 1D batch normalization, a single FC-layer, and a softmax layer. As a result, our model produces the final prediction \\(\\mathbf{x}_{\\mathrm{class}}\\in \\mathbb{R}^C\\), where \\(C\\) is the number of target classes (glosses) in the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20744" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.131, + 0.105 + ], + "angle": 0, + "content": "dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.107, + 0.47, + 0.244 + ], + "angle": 0, + "content": "Training and inference. We train P3D with a cross-entropy loss between predicted \\(\\mathbf{x}_{\\mathrm{class}}\\) and one-hot encoded GT class vector. We randomly sample continuous \\(T\\) frames to form the training inputs. During the inference phase, we sample four chunks of \\(T\\) continuous frames from the human pose sequence in the test set following the previous work [30]. Then, we use P3D to generate the class prediction \\(\\mathbf{x}_{\\mathrm{class}}\\) for each \\(T\\) continuous frame, and average them out to produce the final prediction for the entire sequence." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.258, + 0.201, + 0.275 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.283, + 0.292, + 0.3 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.307, + 0.47, + 0.398 + ], + "angle": 0, + "content": "We used PyTorch [39] Transformer implementation for our PET and WET, with 256 inner dimensions, 0.3 dropout rate, GELU activation layer, and 4 multi-head attentions. We trained our model for 500 epochs using Adam [24] optimizer with a mini-batch size of 512, a learning rate of \\(5 \\times 10^{-4}\\), the weight decay rate of \\(5 \\times 10^{-3}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.408, + 0.468, + 0.425 + ], + "angle": 0, + "content": "4.2. Datasets, evaluation metrics and preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.432, + 0.468, + 0.567 + ], + "angle": 0, + "content": "WLASL dataset. WLASL [30] is a large-scale word-level sign language dataset that contains 21,083 videos with 119 signers. The videos in WLASL are collected from the Internet, and each video depicts a signer performing a single sign. The WLASL dataset consists of 2,000 words, and it also provides subsets of the dataset, WLASL100, and WLASL300. Nevertheless, our main target is to push through the limit of recognition performance on WLASL2000." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.468, + 0.704 + ], + "angle": 0, + "content": "Evaluation metrics For the evaluation metrics, we employ Top-1 and Top-5 accuracy. Following the previous work [20], we calculate both metrics under two different settings, i.e., per-instance and per-class. In per-instance setting, we measure the percentage of accurately classified video over the entire test instances. Per-class setting is also utilized since each class (gloss) has a different number of instances. In a per-class setting, we calculate the accuracy of each class and average them over all classes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Pose preprocessing Regarding the 2D pose, we utilize the dedicated 2D pose estimator for better accuracy, since direct estimation of 2D pose is more accurate compared to projection from estimated 3D pose. The 2D pose is aligned based on the bounding box of the signer by applying the affine transform. After then, we rescaled the 2D pose sequence from the bounding box into the [0, 1] range. The 3D poses, including the positional pose, the rotational pose and the facial expression, we utilize the off-the-shelf 3D expressive human pose estimator Hand4Whole [36] as illustrated in Figure 5. For the 3D positional pose, we aligned the translation by placing the root joint of the first frame at the origin without normalization. In addition, the 3D rotational" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "pose and the facial expression features are used without preprocessing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.134, + 0.65, + 0.15 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.158, + 0.892, + 0.219 + ], + "angle": 0, + "content": "We conduct an ablation study on two main contributions of our framework: part-wise motion context learning and pose ensemble. For each experiment, the contribution that is not being tested remains identical to our proposed P3D." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.22, + 0.892, + 0.339 + ], + "angle": 0, + "content": "Effect of part-wise motion context learning. In Table 1, we compare the results of various pose encoding methods in single iteration of P3D. The MLP is the simple alternative to the proposed PET or WET, which is constructed of stacked fully-connected layers, batch normalization, and dropout. Experimental results demonstrate that employing our PET and WET gives a large performance gain over MLP alternatives." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.341, + 0.892, + 0.598 + ], + "angle": 0, + "content": "We also report that a combination of PET and WET is essential in sign language recognition. When the number of classes is small (WLASL 100, WLASL 300), utilizing one of the WET (the fourth row) or PET (the third row) provides a similar result to utilizing both PET and WET (the fifth row). However, when the number of classes is large (WLASL 2000), utilizing both PET and WET performs the best. As the number of classes increases, learning highly distinguishable features from motion context becomes crucial. For that, the PET and WET are dedicated to capturing intra-part context and merging them into a unified inter-part context, respectively. Such design benefits the task-specific representation power on sign language recognition, that each part (i.e. body, hands, and facial expressions) holds its own important motion context. As a result, our PET-WET encoder outperforms PET-PET and WET-WET by \\(3.75\\%\\) and \\(12.08\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.599, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Effect of each human part. In Table 2, we conduct an ablation study on employing different human part inputs. As our task is sign language recognition, hands are the most essential part of accurate recognition. However, the additional gain can be obtained by leveraging body pose and facial expression on top of hands. It demonstrates that body pose and facial expression, as well as hands, also provide helpful information for sign language recognition." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of each pose representation. Table 3 shows that jointly utilizing 2D positional pose, 3D positional pose, and 3D rotational pose achieves the best performance. In terms of single representation, the 3D rotational pose is the most effective pose representation for sign language recognition. Furthermore, additionally using the 3D rotational pose to other pose representations (the first row & the fifth row, the second row & the sixth row, and the fourth row & last row) consistently improves the accuracy. As a reason for the effectiveness of 3D rotational pose for sign language recognition, we argue that 3D rotational pose is less sensitive to signer variation than 2D positional pose and 3D positional" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20745" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.872, + 0.201 + ], + "angle": 0, + "content": "
Pose encoding methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
MLP-WET71.8985.9472.3387.2159.4482.2059.7082.4334.9770.5732.7869.41
MLP-PET72.2888.7673.7689.5557.8982.5058.5682.9634.7270.7132.4969.66
PET-PET77.9194.7778.3095.2567.3388.6868.0688.0632.3963.6130.3462.20
WET-WET75.5091.5776.4392.6866.8788.2466.9288.5640.7274.7038.2673.54
PET-WET (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.21, + 0.894, + 0.255 + ], + "angle": 0, + "content": "Table 1. Ablation study on pose encoding methods. We ablate the performance of pose encoding methods in single iteration of P3D. Our part-wise context learning (PET-WET) scored the best performance. The observation that ours (PET-WET) has outperformed WET-WET (Transformer without part) directly supports our main contribution of part-wise context learning." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.265, + 0.875, + 0.378 + ], + "angle": 0, + "content": "
Input human partsWLASL 100WLASL 300WLASL 2000
BodyHandsExpr.Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
✓44.9874.7044.7575.8537.6270.4337.4070.1717.8040.7615.7037.57
✓✓46.9972.2945.5772.4338.7070.4338.7370.4520.9847.0518.7343.94
✓71.8991.1673.1092.0862.0787.1562.6987.9339.0073.4036.9772.16
✓✓75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.386, + 0.893, + 0.416 + ], + "angle": 0, + "content": "Table 2. Ablation study on input human parts for P3D. We ablate the performance of provided input human parts. Check marks in the first column implies the given input part. We abbreviate the facial expression to Expr." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.447, + 0.465, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.708, + 0.471, + 0.79 + ], + "angle": 0, + "content": "Figure 6. t-SNE [48] visualization of features from different pose representations. Each plot corresponds to a single word (\"Corn\", \"Language\", \"Many\", and \"Thanksgiving\", respectively), and each point is feature \\( X^{(N)} \\) from each video. All test videos of the corresponding word in WLASL 2000 are employed for the visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "pose. Figure 6 illustrates that features of single words from different signers are closely embedded when input pose representation is a 3D rotational pose. On the other hand, in the case of the 2D and 3D positional poses, the embedded features of the single words show large discrepancies when the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.499 + ], + "angle": 0, + "content": "signers are different. Such advantage of 3D rotational pose leads to improved performance over other pose representations since the sign languages in the WLASL dataset are signer-independent [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.893, + 0.62 + ], + "angle": 0, + "content": "Pose ensemble methods. In Table 4, we compare the performances of different pose ensemble methods. Since we employ three types of pose representations (2D positional pose, 3D positional pose, and 3D rotational pose), how to ensemble information in each pose representation is an important issue. We design three strategies for ensemble methods, and the description for each ensemble method is as follows." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.629, + 0.892, + 0.706 + ], + "angle": 0, + "content": "- Early ensemble (Ours): All pose representations of each joint are concatenated before being fed to the PET-WET blocks. Thus, PET-WET blocks the process motion context of all poses representations simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.714, + 0.892, + 0.792 + ], + "angle": 0, + "content": "- Middle ensemble: Separate PET-WET blocks are utilized to extract feature \\( X^{(n)} \\) for each pose representation: \\( X_{2\\mathrm{D}\\mathrm{pos}}^{(n)} \\), \\( X_{3\\mathrm{D}\\mathrm{pos}}^{(n)} \\), and \\( X_{3\\mathrm{D}\\mathrm{rot}}^{(n)} \\). The extracted features of all pose representations are averaged before being fed to the prediction head." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.801, + 0.892, + 0.862 + ], + "angle": 0, + "content": "- Late ensemble: Same as the middle ensemble, but each feature of each pose representation is fed to its own prediction head. The output of all prediction heads is averaged to regress a single word." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.629, + 0.892, + 0.862 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The experimental results show that Early ensemble performs best for the WLASL 2000. In Table 5, we compare" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20746" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.875, + 0.228 + ], + "angle": 0, + "content": "
Input pose representationsWLASL 100WLASL 300WLASL 2000
2D pos.3D pos.3D rot.Per-instancePer-classPer-instancePer-classPer-instancePer-class
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
✓60.2483.1360.4582.6352.1779.7252.5979.7935.9967.2633.4765.15
✓41.0966.6740.9867.0239.9769.7640.0170.0131.6962.8228.4259.83
✓72.8788.3773.6088.9361.6882.3461.9382.7838.6073.4235.5771.86
✓✓55.0281.1255.7382.0150.1580.1949.7481.0936.9169.0934.0067.06
✓✓75.5091.1676.8891.5865.9488.8566.2688.7543.4877.6441.3976.38
✓✓74.4288.3775.1589.2662.7283.5363.4384.0141.1476.7938.4775.34
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.237, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Table 3. Ablation study on input pose representations. We ablate the performance of provided input human pose representations. Check marks in the first column implies the given input pose representation. Employing every pose representation scored the best result." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.278, + 0.872, + 0.364 + ], + "angle": 0, + "content": "
Ensemble methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
Late77.9093.1778.3095.2567.3488.2368.0688.6832.7164.0430.3463.67
Middle79.9293.5781.6694.3867.8089.9467.8690.1342.6371.7140.1776.43
Early (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.372, + 0.892, + 0.402 + ], + "angle": 0, + "content": "Table 4. Ablation study on pose representation ensemble methods. We report the quantitative result on pose representation ensemble methods. Early ensemble by joint-wise concatenation outperforms the middle (between encoder & head) and late (after head) ensemble." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.424, + 0.467, + 0.513 + ], + "angle": 0, + "content": "
Ensemble methodsFLOPs (G)# Params (M)
Late182.0610.71
Middle182.0210.65
Early (Ours)83.664.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.524, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Table 5. Computational costs of pose ensemble methods. We used the FLOPs counter for PyTorch [39] from fvcore package to measure the computational costs of the models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.684 + ], + "angle": 0, + "content": "the computational costs of ensemble methods. As separate PET-WET blocks are utilized in Middle ensemble and Late ensemble, Early ensemble is most efficient in terms of the number of model parameters and FLOPs. Considering the accuracy and efficiency, we adopt Early ensemble as our ensemble strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.437, + 0.71 + ], + "angle": 0, + "content": "4.4. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Table 6 shows that P3D outperforms previous pose-based methods. We first demonstrate the capability of our proposed part-wise motion context learning model via quantitative comparison on the WLASL dataset using only 2D pose. Then we show that 3D pose fusion is more effective in our P3D than in other models, leading to our model outperforming others on the WLASL dataset using 2D and 3D pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Using only 2D pose. The second block of Table 6 shows that our P3D outperforms existing methods on WLASL 300 and WLASL 2000 split with 2D pose input. In this experiment, we only use 2D pose for the fair comparison of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.728 + ], + "angle": 0, + "content": "models to learn highly distinguishable features from motion context. Prior to our work, state-of-the-art methods were SPOTER [3] in WLASL 100, and ST-GCN [50] in WLASL 300 and WLASL 2000. From the benefits of its simple and intuitive Transformer architecture, SPOTER achieved the best performance on WLASL 100. However, due to the limitation on representation power, SPOTER did not perform well on more complex problems such as WLASL 300 and WLASL 2000. Meanwhile, graph convolution via spatiotemporal adjacency used in ST-GCN is observed to be effective on a large number of classes (WLASL 300 and WLASL 2000). Yet, employing the spatio-temporal adjacency graph leads to two cons: it cannot model the long-term dependency, and cannot learn the part-specific context. In contrast, the proposed part-aware Transformer architecture of our P3D is capable of handling the long-term dependency and part-specific context at the same time. As a result, our P3D shows better performance compared to ST-GCN on every split of WLASL, proving the better representation power of proposed part-wise motion context learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Jointly using 2D and 3D pose. The third block of Table 6 shows that our P3D scores highly superior results compared to previous methods with using both 2D and 3D poses as an input. Again, in order to conduct fair comparison, identical input features are used and the same pose ensemble method is used for every tested models. We performed a comparative analysis in two aspects as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "First, our P3D resulted in a much higher performance gain from jointly using 2D and 3D poses. While our P3D gained \\(6.39\\%\\) top-1 per-instance recognition accuracy on WLASL2000 by pose ensemble, gains of previous state" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20747" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.882, + 0.369 + ], + "angle": 0, + "content": "
MethodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
RGB-based
TK-3D ConvNet [17]77.55---68.75-------
Fusion-3 [19]75.67---68.3-------
I3D [7]65.8984.1167.0184.5856.1479.9456.2478.3832.4857.31--
Pose-based (2D)
Pose-TGCN [30]55.4378.68--38.4267.51--23.6551.75--
GCN-BERT [15]60.1583.98--42.1871.71------
SPOTER [3]63.18---43.18---28.8661.7828.2160.26
ST-GCN [50]50.7879.0751.6279.4744.4673.0545.2973.1634.4066.4732.5365.45
P3D (Ours)60.2485.9460.4885.6650.9377.7150.8178.5535.9668.3133.2266.20
Pose-based (2D + 3D)
SPOTER [3]73.9091.5774.5191.5554.0282.2055.2882.6430.9164.2529.8862.7
ST-GCN [50]56.6379.5156.9379.7145.8276.0145.9676.2237.3772.1335.6170.66
P3D (Ours)75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
Pose-based (2D + 3D + facial expression)
P3D (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.378, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Table 6. Comparisons with state-of-the-art methods. We report the recognition accuracy of previous methods for comparison with ours. The results in each block are trained and evaluated with using exact same input denoted in the first column. Results in the first block show the recognition accuracy of existing RGB-based approaches. The second block represents the pose-based approach with 2D pose inputs. Our result in the second block is also generated solely with 2D poses. Scores in the third block are obtained from employing pose ensemble on previous methods and ours, using both 2D and 3D poses. Note that previous methods did not utilize 3D pose in their works, thus we applied pose ensemble based on released codes. Finally, the fourth block shows the best performance of our model using 2D, 3D poses and facial expression. Blank scores denoted with a hyphen are not reported in previous works." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.606 + ], + "angle": 0, + "content": "of-the-art methods SPOTER [3] and ST-GCN [50] are confined to \\(2.05\\%\\) and \\(2.97\\%\\), respectively. Such observation suggests that our part-wise motion context learning is more capable of acquiring contextual information from various types of pose representations, which distinguishes it from the basic Transformer architecture of SPOTER [3] or the spatio-temporal graph convolution of ST-GCN [50]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Second, P3D achieved superior performance compared to existing methods, particularly by a large margin on a large number of classes such as WLASL300 and WLASL2000. As demonstrated in the third block of Table 6, our P3D outperformed previous pose-based methods SPOTER [3] and ST-GCN [50] on every subset of WLASL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Using 3D expressive pose. The fourth block of Table 6 shows the performance with full input features: 2D pose, 3D positional and rotational pose, and facial expression. With additional usage of facial expression over the third block of Table 6, we observed the performance gain of \\(0.79\\%\\), \\(0.93\\%\\), and \\(2.12\\%\\), respectively for each subset of WLASL. Moreover, when compared to the RGB-based models presented in the first block of Table 6, our P3D scored comparable results in WLASL100 and WLASL300, and outperformed in WLASL2000. Our framework is the first to outperform (WASL 2000) or achieve comparable performance (WASL 100, WASL 300) with RGB-based methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.619, + 0.515 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.732 + ], + "angle": 0, + "content": "In this work, we proposed the sign language recognition framework P3D by leveraging part-wise 3D motion context learning. We utilize the alternating PET and WET layers, specializing in learning intra-part and inter-part motion contexts, respectively. Learning part-wise motion context resulted in superior performance on the WLASL benchmark compared to previous methods. Moreover, we jointly utilized 2D and 3D poses via joint-wise pose concatenation to fully exploit the performance of pose-based sign language recognition. We observed that joint-wise concatenation has advantages on both sides of performance and efficiency. As a result, our P3D outperformed the previous methods by a large margin on WLASL, jointly using 2D and 3D poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.775, + 0.667, + 0.792 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], the NRF grant [No. 2021M3A9E4080782] funded by the Korea government (MSIT), and the SNU-Samsung SDS Research Center." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20748" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] Matyáš Boháček and Marek Hrúz. Sign pose-based transformer for word-level sign language recognition. In WACV, 2022. 2, 3, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[8] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[9] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.603, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[11] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.645, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[12] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia (ToM), 21(7):1880-1891, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.702, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[14] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[15] Haodong Duan, Yue Zhao, Kai Chen, Dian Shao, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[16] Zhimin Gao, Peitao Wang, Pei Lv, Xiaoheng Jiang, Qidong Liu, Pichao Wang, Mingliang Xu, and Wanqing Li. Focal" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and global spatial-temporal transformer for skeleton-based action recognition. In ACCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[17] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In CVPR, 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Xiao Guo and Jongmoo Choi. Human motion prediction via learning local structure representations and temporal dependencies. In AAAI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[19] Al Amin Hosain, Panneer Selvam Santhalingam, Parth Pathak, Huzefa Rangwala, and Jana Kosecka. Hand pose guided 3d pooling for word-level sign language recognition. In WACV, 2021. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: pre-training of hand-model-aware representation for sign language recognition. In ICCV, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[21] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[22] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. In BMVC, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[23] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.576 + ], + "angle": 0, + "content": "[25] Oscar Koller. Quantitative survey of the state of the art in sign language recognition, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.578, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[26] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(9):2306-2320, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[27] Oscar Koller, Sepehr Zargaran, Hermann Ney, and Richard Bowden. Deep sign: Enabling robust statistical continuous sign language recognition via hybrid cnn-hmms. International Journal of Computer Vision (IJCV), 126:1311-1325, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[29] Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Cooccurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation. arXiv preprint arXiv:1804.06055, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, 2020. 1, 2, 3, 6, 7, 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20749" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[31] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[32] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-parted graph scattering networks for 3d human motion prediction. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.222 + ], + "angle": 0, + "content": "[33] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.469, + 0.264 + ], + "angle": 0, + "content": "[34] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.469, + 0.323 + ], + "angle": 0, + "content": "[35] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, Oct. 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.326, + 0.469, + 0.368 + ], + "angle": 0, + "content": "[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR Workshops, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.37, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[37] Gyeongsik Moon, Heeseung Kwon, Kyoung Mu Lee, and Minsu Cho. Integralaction: Pose-driven feature integration for robust human action recognition in videos. In CVPRW, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[38] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. *NeurIPS Workshop on Autodiff*, 2017. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.543, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In ICPR Workshops, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[43] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[44] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3d multi-channel sign language production via progressive transformers and mixture density networks. International Journal of Computer Vision (IJCV), 129(7):2113-2135, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[45] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In AAAI, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[46] Du Tran, Lubomir Bourdev, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[47] Anirudh Tunga, Sai Vidyaranya Nuthalapati, and Juan Wachs. Pose-based sign language recognition using gcn and bert. In WACV, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[52] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[53] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[54] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20750" + } + ] +] \ No newline at end of file diff --git a/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_origin.pdf b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e8c9ca419eefb83fdf80253133b651012cb93af --- /dev/null +++ b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/7a02bc72-2ece-4aac-a870-d0710b03a665_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcd20e2e70c3e86602dbeedbc1e4febb8d60b831fbcd2c282794f59e7c7fbb7f +size 2279143 diff --git a/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/full.md b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c71f42afa05c0382dcb4d58b60ecd60702f24d97 --- /dev/null +++ b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/full.md @@ -0,0 +1,265 @@ +# Human Part-wise 3D Motion Context Learning for Sign Language Recognition + +Taeryung Lee + +Yeonguk Oh² + +Kyoung Mu Lee $^{1,2}$ + +1IPAI, 2Dept. of ECE & ASRI, Seoul National University, Seoul, Korea + +{trlee94, namepllet, kyoungmu}@snu.ac.kr + +# Abstract + +In this paper, we propose P3D, the human part-wise motion context learning framework for sign language recognition. Our main contributions lie in two dimensions: learning the part-wise motion context and employing the pose ensemble to utilize 2D and 3D pose jointly. First, our empirical observation implies that part-wise context encoding benefits the performance of sign language recognition. While previous methods of sign language recognition learned motion context from the sequence of the entire pose, we argue that such methods cannot exploit part-specific motion context. In order to utilize part-wise motion context, we propose the alternating combination of a part-wise encoding Transformer (PET) and a whole-body encoding Transformer (WET). PET encodes the motion contexts from a part sequence, while WET merges them into a unified context. By learning part-wise motion context, our P3D achieves superior performance on WLASL compared to previous state-of-the-art methods. Second, our framework is the first to ensemble 2D and 3D poses for sign language recognition. Since the 3D pose holds rich motion context and depth information to distinguish the words, our P3D outperformed the previous state-of-the-art methods employing a pose ensemble. + +# 1. Introduction + +Understanding and translating sign language is essential due to the huge potential benefit to society. Sign language recognition (SLR) can be considered as a specific kind of action recognition (AR) since both tasks aim to classify human motion. However, SLR has a characteristic that is distinguished from AR; Actions mainly focus on the motion of the human body, while the SLR system should be aware of more detailed information such as hand gestures and facial expressions. Thus, a specified system for SLR will provide further gain over using the AR models on SLR. + +Previous approaches on SLR problem are mainly cat- + +![](images/36f551132757c4f08a38b8a6b46bc1bb40fda5afe7af3c332a471f50c0e43b91.jpg) +Figure 1. Overall pipeline of proposed system. We first extract expressive 3D human pose from RGB video using off-the-shelf pose estimation methods [5, 36]. The expressive 3D human pose consists of 2D pose, 3D pose and facial expression. Then our proposed P3D predicts the word from the expressive 3D human pose. + +egorized into RGB-based [1, 4, 31] and pose-based methods [25, 29, 45, 47, 50]. Pose-based methods are known to have the advantage over RGB-based approaches in terms of the robustness on domain gap [14, 37]. For that, we aimed to enhance the performance of SLR, while preserving the benefits inherent in the pose-based approach. + +Although recent pose-based SLR methods have shown noticeable progress, there are two main limitations of existing methods. First, they utilize only whole-body representation (e.g., skeleton) to understand the motion context of human [20, 30, 47]. Since such approaches overlook the detailed motion context within a specific part + +![](images/c20a43ea8f81905307522bb4d63b6a22d2fb6a0f24bbbc8871d5cfa02d403e79.jpg) +RGB Video +2D Pose +3D Pose +Figure 2. Effectiveness of 3D pose for sign language recognition. The 2D poses of the words "Avoid" and "Behind" are almost identical. Due to the absence of depth information, solely relying on 2D pose makes it hard to distinguish the words. On the other hand, as shown in the red boxes, additionally employing the 3D pose with rich depth information can effectively resolve the ambiguity in sign language recognition. + +![](images/d51f5a1814776b316bc991fb7e10f444b1817324401eb1934b0a65484091baab.jpg) + +of human [18, 32], it leads to a decrease in SLR accuracy. Second, existing pose-based methods only employ 2D pose [3, 20, 30, 47], which induces ambiguities in distinguishing the different words. As shown in Figure 2, the 2D pose of different words can be almost identical due to the absence of depth information. + +To overcome the above limitations, we propose P3D, the human part-wise motion context learning framework for accurate sign language recognition. The proposed P3D employs the part-wise encoding Transformer (PET) and whole-body encoding Transformer (WET) as main components. Given the pose sequence, we first divide it into four parts: body, right hand, left hand, and face. In contrast to previous methods where entire human poses are jointly processed, we utilize part-specific PET to encode the motion context of each part. As each PET takes a pose sequence of a single part, it can learn more detailed motion context, which is pivotal for sign language recognition. Then, the encoded features from each part are merged by WET. While PET only considers the intra-part motion context, WET covers the entire part. For the effective fusion of the inter-part and intra-part motion context, we alternately combined PET and WET inspired by spatio-temporal Transformers [2,51]. As a result, our proposed P3D has out + +performed the previous state-of-the-art [3, 50] on WLASL. Our ablation study on layer design supports part-wise motion context learning benefits SLR performance. Especially, we observed that solely forming the model with WET (i.e., without part-wise motion context) degrades the performance, thus directly induces that our PET contributed to improving the accuracy. + +We alleviate the second limitation of previous methods through the ensemble of 2D and 3D poses. In terms of 3D pose, we utilize both 3D positional pose and 3D rotational pose (e.g., SMPL-X [41] pose parameters). The 3D positional pose complements depth information that the 2D pose does not contain, and the 3D rotational pose provides rich motion context on the kinematic chain. Our experiment shows that our P3D outperforms the previous methods by a large margin with using identical inputs; our model shows the superior result from only 2D pose, and also from ensemble of 2D and 3D pose. Moreover, our framework is the first pose-based method that recorded a comparable score or even outperformed the RGB-based methods. We claim our main contributions as follows: + +- We proposed P3D, the human part-wise motion context learning framework for sign language recognition. Our experiment supports that part-wise learning highly + +benefits recognition performance. As a result, our proposed P3D outperformed the previous SOTA on the WLASL benchmark. + +- We employ the pose ensemble by joint-wise pose concatenation in order to exploit both 2D and 3D poses. We observed significant performance gain via pose ensemble, and P3D outperformed the previous methods by a large margin by jointly using 2D and 3D poses. + +# 2. Related works + +# 2.1. Sign language recognition + +Word-level sign language recognition task aims to classify short-term videos by the corresponding word label. Previous works on sign language recognition can be mainly categorized into two groups by the input modality: videobased and pose-based approaches. + +Video-based sign language recognition. With the significant performance of the CNN-based action recognition, the majority of the prior works have designed their models based on CNN backbone [9, 27, 53]. Primary works [11, 12, 26, 44] first used convolutional neural networks to extract the frame-wise features. Then, such methods encoded the temporal semantics along the time dimension using recurrent neural networks. On the other hand, similar to action recognition, 3D CNNs are widely used in sign language recognition. C3D [21, 23, 46] is the first 3D CNN model for action recognition, and it is widely used as a baseline model for quantitative comparison. I3D [7] is proposed based on the C3D [46] architecture and applied on sign language recognition in [22]. Li et al. [30] have compared various architectures, including 3D CNN and 2D CNN combined with RNN. Hosain et al. [19] proposed Fusion-3 model based on I3D Inception-v1 [7] with pose-guided pooling. + +Pose-based sign language recognition. Compared to the RGB video, the human pose sequence holds a compact and semantic-aware representation of human motion. ST-GCN [50] is the first pose-based approach to action recognition, utilizing the spatio-temporal graph convolutional network to encode the motion on kinetic chain. Following works also have used the spatio-temporal architecture based on graph convolutional network [30, 47] and Transformers [49] [3, 20]. + +# 2.2. 3D human pose and shape estimation + +3D human pose and shape estimation aims to localize 3D coordinates of vertices of the human mesh. Considering the various applications such as VR/AR and human action recognition, research on 3D human pose and shape estimation has been widely conducted. Most works tried to reconstruct 3D human pose and shape for specific part of human, + +e.g. body [10, 34, 38], hand [8, 28], or face [43]. Several works [36, 40] proposed methods that simultaneously estimate the 3D pose and shape for all human parts, i.e., whole-body. By reconstructing 3D poses and shapes for all human parts, we can understand human intention and feelings more precisely. In this work, we leverage Hand4Whole [36] to acquire a 3D pose of all human parts from RGB videos. To the best of our knowledge, we first utilize 3D pose for sign language recognition. + +# 2.3. Transformers for human motion. + +After the remarkable success of Transformers in natural language processing (NLP) [49], vision researchers have employed Transformers for various vision tasks [6, 13, 34]. Transformers also have been widely adopted for tasks that require human motion understanding, such as 2D-to-3D pose sequence lifting [33, 51, 52] and skeleton-based action recognition [16, 42]. In the field of 2D-to-3D pose sequence lifting, PoseFormer [52] proposed a pure Transformer-based approach for 2D-to-3D pose sequence lifting. MixSTE [51] utilized separated joint-wise temporal Transformer and spatial Transformer alternately. MH-Former [33] employed Transformers to generate and refine multiple hypotheses for 3D pose sequences from 2D pose sequences. In the field of skeleton-based action recognition, ST-TR [42] proposed Transformer-based two-stream networks where each stream learns spatial and temporal correlations in a skeleton sequence. FG-STFormer [16] presented a variation of vanilla Transformer to capture local and global information effectively. Our P3D also leverages Transformer-based modules to learn the human motion. However, most previous works only focused on the motion of the human body, while proposed P3D learns the motion of all human parts, including hands and facial expressions. + +# 3. Method + +# 3.1. Overview + +In this section, we describe the input data format and architecture of our P3D model. Figure 3 illustrates the overview of our model. We provide a sequence of human poses to our model as an input P3D leverages alternating structure of Transformer layers [49], inspired by the spatiotemporal architectures from the previous works [2,51]. The first layer, the part-wise encoding Transformer (PET), is designed to learn intra-part motion context. To capture the specific motion context from each part, we assign the dedicated Transformers. Meanwhile, the second whole-body encoding Transformer (WET) encodes the inter-part motion context, taking account of every human part used in the model. As the last step, we take mean along the time axis on encoded representations, then sequentially pass through + +![](images/45217de8a3a149ffce1ddca5e57d5d2d2ab4054588dc19503294084dbe976709.jpg) +Figure 3. Overall architecture of P3D. Our P3D aims to predict the word label corresponding to a given pose sequence. We first provide joint-wise linearly embedded feature vectors as an input, along with linearly encoded facial expression features. P3D consists of $2N$ alternating layers of PET, which are assigned for each part, and WET, which covers entire parts. After the encoder, the prediction head takes the mean along the time dimension, then sequentially passes through batch normalization, linear layer, and softmax layer to produce the word prediction. + +![](images/08a61810a0b220b36d9a35453cd2adb752e3ed6af2523e9d86f3943f2745b207.jpg) +Figure 4. Overall architectures of PET and WET. We adapt the Transformer encoder layer [49] with vanilla self-attention for both PET (above) and WET (below). For the PET, the subscript $p \in \{B, RH, LH, \psi\}$ denotes the assigned human part. + +batch normalization, linear layer, and softmax layer to get the final output. + +# 3.2. Input format + +Human motion representation. The overview of the input data structure is illustrated on the left of Figure 3. As a pose-based sign language model, our P3D takes a sequence $(X_{1},\dots,X_{T})\in \mathbb{R}^{T\times (J\times F + E)}$ of $T$ human poses as an input. Pose representation $X_{t} = (\mathbf{x}_{1},\dots,\mathbf{x}_{J},e)\in \mathbb{R}^{J\times F + E}$ for each time $t\in \{1,2,\dots,T\}$ consists of joint-wise pose representation $\mathbf{x}_j\in \mathbb{R}^F$ for each joint $j\in \{1,2,\dots,J\}$ , and facial expression feature $e\in \mathbb{R}^E$ [41]. From our experiment on the pose fusion method, we have observed that joint-wise concatenation of the 2D positional pose (2D pos.), 3D positional pose (3D pos.), and 3D rotational pose (3D rot.) performed the best on WLASL2000. Thus, we form the joint-wise pose representation $\mathbf{x}_j = (\mathbf{x}_{j,2\mathrm{D}\mathrm{pos}},\mathbf{x}_{j,3\mathrm{D}\mathrm{pos}},\mathbf{x}_{j,3\mathrm{D}\mathrm{rot}})\in \mathbb{R}^F$ where $\mathbf{x}_{j,2\mathrm{D}\mathrm{pos}}\in \mathbb{R}^2$ , $\mathbf{x}_{j,3\mathrm{D}\mathrm{pos}}\in \mathbb{R}^3$ , $\mathbf{x}_{j,3\mathrm{D}\mathrm{rot}}\in \mathbb{R}^6$ represents the 2D positional pose, 3D positional pose and 3D rotational pose [54] of joint $j$ , respectively. Please refer to our experiment on Section 4.3 for the details of pose fusion methods. Following the literature [20], we use $T = 32$ for both training and testing. Joint set and human part. We use $J = 40$ joints and 4 human parts to represent the human motion of the sign language. Ten joints above the pelvis are selected from the SMPL [35] joint set, defined as a body part. Fifteen joints, three per finger, represent each hands. For the face, we use the 10-dimensional feature vector from SMPL-X [41] to represent the facial expression. Please refer to Section 4.2 for the details of human pose preprocessing. + +![](images/5f00008c5beaeb8e9ca0566a224d0c86d61cd3484bb03538b5c4c09fe1bc2997.jpg) +Figure 5. Visualization. The 3D expressive pose of the signer (b) is estimated from the given sign language video (a) with off-the-shelf estimator Hand4Whole [36]. The corresponding word label (c) is denoted in the right. + +# 3.3. Architecture + +Linear layer. Our model first linearly encodes each jointwise pose representation $\mathbf{x}_j \in \mathbb{R}^F$ into D dimensional vector. We also assign a dedicated linear layer to encode facial expressions into $\alpha D$ dimensional vector. Empirically, we decided to use $D = 8$ and $\alpha = 10$ . As a result, pose $X_t = (\mathbf{x}_1, \dots, \mathbf{x}_J, e) \in \mathbb{R}^{J \times F + E}$ at time $t$ is encoded into $X_t^{(0)} \in \mathbb{R}^{(J \times D + \alpha D) = ((10 + 15 + 15) \times 8 + 80)}$ . Note that 10, 15, and 15 represent the number of joints in the body, left hand and right hand, and thus can be separated into each human part via indexing. Finally we form the input $X^{(0)} = (X_1^{(0)}, X_2^{(0)}, \dots, X_T^{(0)}) \in \mathbb{R}^{T \times ((10 + 15 + 15) \times 8 + 80)}$ with positional encoding [49]. + +Alternating Transformers. We adapted the idea of alternating two types of layers [2, 51], which have different purposes and expertise. Our P3D consists of $2N$ alternating layers: $N$ Part-wise Encoding Transformer (PET) and $N$ Whole-body Encoding Transformer (WET). Starting from linear encoded input $X^{(0)}$ , we denote the output of $i$ 'th PET as $Y^{(i)}$ and the output of $i$ 'th WET as $X^{(i)}$ . As the dimension remains identical through Transformers, our features $X^{(i)}$ and $Y^{(i)}$ remain in a same dimension as $X^{(0)}$ . Empirically, we use $N = 3$ layers for WLASL. + +Part-wise encoding Transformer (PET). The architecture of PET is illustrated in the upper part of Figure 4. From previous layer, PET takes 2D tensor $X^{(i)} \in$ + +$\mathbb{R}^T \times ((10 + 15 + 15) \times 8 + 80)$ as an input. The first (horizontal) and second (vertical) dimensions represent the time and feature axis, respectively. We first separate features for each human part $X_B^{(i)} \in \mathbb{R}^{T \times 80}$ , $X_{LH}^{(i)} \in \mathbb{R}^{T \times 120}$ , $X_{RH}^{(i)} \in \mathbb{R}^{T \times 120}$ , $X_{\psi}^{(i)} \in \mathbb{R}^{T \times 80}$ via slicing on second (feature) dimension by indices (80, 200, 320, 400), respectively. Then we pass each feature through dedicated Transformer encoder layers $\mathrm{PET}_B$ , $\mathrm{PET}_{RH}$ , $\mathrm{PET}_{LH}$ , and $\mathrm{PET}_{\psi}$ assigned for each human part, respectively. Finally, we concatenate the outputs along the second (feature) dimension into a single 2D tensor $Y^{(i)} \in \mathbb{R}^T \times ((10 + 15 + 15) \times 8 + 80)$ . + +Whole-body encoding Transformer (WET). The architecture of WET is illustrated in the lower part of Figure 4. From the previous layer PET, WET takes 2D tensor $Y^{(i)} \in \mathbb{R}^{T \times ((10 + 15 + 15) \times 8 + 80)}$ as an input. Since WET aims to learn inter-part motion context, we assign a single Transformer to encode the whole motion sequence. Output of WET $X^{(i + 1)} \in \mathbb{R}^{T \times ((10 + 15 + 15) \times 8 + 80)}$ is passed to next layer without modification. + +Prediction head. From Transformer-encoded feature $X^{(N)}\in \mathbb{R}^{T\times 400}$ , we first calculate the mean feature vector $\mathbf{x}_{\mathrm{feat}}\in \mathbb{R}^{400}$ by taking the mean along the first (time) dimension. Then, we pass $\mathbf{x}_{\mathrm{feat}}$ sequentially through 1D batch normalization, a single FC-layer, and a softmax layer. As a result, our model produces the final prediction $\mathbf{x}_{\mathrm{class}}\in \mathbb{R}^C$ , where $C$ is the number of target classes (glosses) in the + +dataset. + +Training and inference. We train P3D with a cross-entropy loss between predicted $\mathbf{x}_{\mathrm{class}}$ and one-hot encoded GT class vector. We randomly sample continuous $T$ frames to form the training inputs. During the inference phase, we sample four chunks of $T$ continuous frames from the human pose sequence in the test set following the previous work [30]. Then, we use P3D to generate the class prediction $\mathbf{x}_{\mathrm{class}}$ for each $T$ continuous frame, and average them out to produce the final prediction for the entire sequence. + +# 4. Experiment + +# 4.1. Implementation details + +We used PyTorch [39] Transformer implementation for our PET and WET, with 256 inner dimensions, 0.3 dropout rate, GELU activation layer, and 4 multi-head attentions. We trained our model for 500 epochs using Adam [24] optimizer with a mini-batch size of 512, a learning rate of $5 \times 10^{-4}$ , the weight decay rate of $5 \times 10^{-3}$ . + +# 4.2. Datasets, evaluation metrics and preprocessing + +WLASL dataset. WLASL [30] is a large-scale word-level sign language dataset that contains 21,083 videos with 119 signers. The videos in WLASL are collected from the Internet, and each video depicts a signer performing a single sign. The WLASL dataset consists of 2,000 words, and it also provides subsets of the dataset, WLASL100, and WLASL300. Nevertheless, our main target is to push through the limit of recognition performance on WLASL2000. + +Evaluation metrics For the evaluation metrics, we employ Top-1 and Top-5 accuracy. Following the previous work [20], we calculate both metrics under two different settings, i.e., per-instance and per-class. In per-instance setting, we measure the percentage of accurately classified video over the entire test instances. Per-class setting is also utilized since each class (gloss) has a different number of instances. In a per-class setting, we calculate the accuracy of each class and average them over all classes. + +Pose preprocessing Regarding the 2D pose, we utilize the dedicated 2D pose estimator for better accuracy, since direct estimation of 2D pose is more accurate compared to projection from estimated 3D pose. The 2D pose is aligned based on the bounding box of the signer by applying the affine transform. After then, we rescaled the 2D pose sequence from the bounding box into the [0, 1] range. The 3D poses, including the positional pose, the rotational pose and the facial expression, we utilize the off-the-shelf 3D expressive human pose estimator Hand4Whole [36] as illustrated in Figure 5. For the 3D positional pose, we aligned the translation by placing the root joint of the first frame at the origin without normalization. In addition, the 3D rotational + +pose and the facial expression features are used without preprocessing. + +# 4.3. Ablation study + +We conduct an ablation study on two main contributions of our framework: part-wise motion context learning and pose ensemble. For each experiment, the contribution that is not being tested remains identical to our proposed P3D. + +Effect of part-wise motion context learning. In Table 1, we compare the results of various pose encoding methods in single iteration of P3D. The MLP is the simple alternative to the proposed PET or WET, which is constructed of stacked fully-connected layers, batch normalization, and dropout. Experimental results demonstrate that employing our PET and WET gives a large performance gain over MLP alternatives. + +We also report that a combination of PET and WET is essential in sign language recognition. When the number of classes is small (WLASL 100, WLASL 300), utilizing one of the WET (the fourth row) or PET (the third row) provides a similar result to utilizing both PET and WET (the fifth row). However, when the number of classes is large (WLASL 2000), utilizing both PET and WET performs the best. As the number of classes increases, learning highly distinguishable features from motion context becomes crucial. For that, the PET and WET are dedicated to capturing intra-part context and merging them into a unified inter-part context, respectively. Such design benefits the task-specific representation power on sign language recognition, that each part (i.e. body, hands, and facial expressions) holds its own important motion context. As a result, our PET-WET encoder outperforms PET-PET and WET-WET by $3.75\%$ and $12.08\%$ , respectively. + +Effect of each human part. In Table 2, we conduct an ablation study on employing different human part inputs. As our task is sign language recognition, hands are the most essential part of accurate recognition. However, the additional gain can be obtained by leveraging body pose and facial expression on top of hands. It demonstrates that body pose and facial expression, as well as hands, also provide helpful information for sign language recognition. + +Effect of each pose representation. Table 3 shows that jointly utilizing 2D positional pose, 3D positional pose, and 3D rotational pose achieves the best performance. In terms of single representation, the 3D rotational pose is the most effective pose representation for sign language recognition. Furthermore, additionally using the 3D rotational pose to other pose representations (the first row & the fifth row, the second row & the sixth row, and the fourth row & last row) consistently improves the accuracy. As a reason for the effectiveness of 3D rotational pose for sign language recognition, we argue that 3D rotational pose is less sensitive to signer variation than 2D positional pose and 3D positional + +
Pose encoding methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
MLP-WET71.8985.9472.3387.2159.4482.2059.7082.4334.9770.5732.7869.41
MLP-PET72.2888.7673.7689.5557.8982.5058.5682.9634.7270.7132.4969.66
PET-PET77.9194.7778.3095.2567.3388.6868.0688.0632.3963.6130.3462.20
WET-WET75.5091.5776.4392.6866.8788.2466.9288.5640.7274.7038.2673.54
PET-WET (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
+ +Table 1. Ablation study on pose encoding methods. We ablate the performance of pose encoding methods in single iteration of P3D. Our part-wise context learning (PET-WET) scored the best performance. The observation that ours (PET-WET) has outperformed WET-WET (Transformer without part) directly supports our main contribution of part-wise context learning. + +
Input human partsWLASL 100WLASL 300WLASL 2000
BodyHandsExpr.Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
✓44.9874.7044.7575.8537.6270.4337.4070.1717.8040.7615.7037.57
✓✓46.9972.2945.5772.4338.7070.4338.7370.4520.9847.0518.7343.94
✓71.8991.1673.1092.0862.0787.1562.6987.9339.0073.4036.9772.16
✓✓75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
+ +Table 2. Ablation study on input human parts for P3D. We ablate the performance of provided input human parts. Check marks in the first column implies the given input part. We abbreviate the facial expression to Expr. + +![](images/cee37f4d5c5b01a574e2559128833b559f9c77fe73c8a90f319818629e00b37f.jpg) +Figure 6. t-SNE [48] visualization of features from different pose representations. Each plot corresponds to a single word ("Corn", "Language", "Many", and "Thanksgiving", respectively), and each point is feature $X^{(N)}$ from each video. All test videos of the corresponding word in WLASL 2000 are employed for the visualization. + +pose. Figure 6 illustrates that features of single words from different signers are closely embedded when input pose representation is a 3D rotational pose. On the other hand, in the case of the 2D and 3D positional poses, the embedded features of the single words show large discrepancies when the + +signers are different. Such advantage of 3D rotational pose leads to improved performance over other pose representations since the sign languages in the WLASL dataset are signer-independent [30]. + +Pose ensemble methods. In Table 4, we compare the performances of different pose ensemble methods. Since we employ three types of pose representations (2D positional pose, 3D positional pose, and 3D rotational pose), how to ensemble information in each pose representation is an important issue. We design three strategies for ensemble methods, and the description for each ensemble method is as follows. + +- Early ensemble (Ours): All pose representations of each joint are concatenated before being fed to the PET-WET blocks. Thus, PET-WET blocks the process motion context of all poses representations simultaneously. +- Middle ensemble: Separate PET-WET blocks are utilized to extract feature $X^{(n)}$ for each pose representation: $X_{2\mathrm{D}\mathrm{pos}}^{(n)}$ , $X_{3\mathrm{D}\mathrm{pos}}^{(n)}$ , and $X_{3\mathrm{D}\mathrm{rot}}^{(n)}$ . The extracted features of all pose representations are averaged before being fed to the prediction head. +- Late ensemble: Same as the middle ensemble, but each feature of each pose representation is fed to its own prediction head. The output of all prediction heads is averaged to regress a single word. + +The experimental results show that Early ensemble performs best for the WLASL 2000. In Table 5, we compare + +
Input pose representationsWLASL 100WLASL 300WLASL 2000
2D pos.3D pos.3D rot.Per-instancePer-classPer-instancePer-classPer-instancePer-class
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
✓60.2483.1360.4582.6352.1779.7252.5979.7935.9967.2633.4765.15
✓41.0966.6740.9867.0239.9769.7640.0170.0131.6962.8228.4259.83
✓72.8788.3773.6088.9361.6882.3461.9382.7838.6073.4235.5771.86
✓✓55.0281.1255.7382.0150.1580.1949.7481.0936.9169.0934.0067.06
✓✓75.5091.1676.8891.5865.9488.8566.2688.7543.4877.6441.3976.38
✓✓74.4288.3775.1589.2662.7283.5363.4384.0141.1476.7938.4775.34
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
+ +Table 3. Ablation study on input pose representations. We ablate the performance of provided input human pose representations. Check marks in the first column implies the given input pose representation. Employing every pose representation scored the best result. + +
Ensemble methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
Late77.9093.1778.3095.2567.3488.2368.0688.6832.7164.0430.3463.67
Middle79.9293.5781.6694.3867.8089.9467.8690.1342.6371.7140.1776.43
Early (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
+ +Table 4. Ablation study on pose representation ensemble methods. We report the quantitative result on pose representation ensemble methods. Early ensemble by joint-wise concatenation outperforms the middle (between encoder & head) and late (after head) ensemble. + +
Ensemble methodsFLOPs (G)# Params (M)
Late182.0610.71
Middle182.0210.65
Early (Ours)83.664.94
+ +Table 5. Computational costs of pose ensemble methods. We used the FLOPs counter for PyTorch [39] from fvcore package to measure the computational costs of the models. + +the computational costs of ensemble methods. As separate PET-WET blocks are utilized in Middle ensemble and Late ensemble, Early ensemble is most efficient in terms of the number of model parameters and FLOPs. Considering the accuracy and efficiency, we adopt Early ensemble as our ensemble strategy. + +# 4.4. Comparison with state-of-the-art methods + +Table 6 shows that P3D outperforms previous pose-based methods. We first demonstrate the capability of our proposed part-wise motion context learning model via quantitative comparison on the WLASL dataset using only 2D pose. Then we show that 3D pose fusion is more effective in our P3D than in other models, leading to our model outperforming others on the WLASL dataset using 2D and 3D pose. + +Using only 2D pose. The second block of Table 6 shows that our P3D outperforms existing methods on WLASL 300 and WLASL 2000 split with 2D pose input. In this experiment, we only use 2D pose for the fair comparison of + +models to learn highly distinguishable features from motion context. Prior to our work, state-of-the-art methods were SPOTER [3] in WLASL 100, and ST-GCN [50] in WLASL 300 and WLASL 2000. From the benefits of its simple and intuitive Transformer architecture, SPOTER achieved the best performance on WLASL 100. However, due to the limitation on representation power, SPOTER did not perform well on more complex problems such as WLASL 300 and WLASL 2000. Meanwhile, graph convolution via spatiotemporal adjacency used in ST-GCN is observed to be effective on a large number of classes (WLASL 300 and WLASL 2000). Yet, employing the spatio-temporal adjacency graph leads to two cons: it cannot model the long-term dependency, and cannot learn the part-specific context. In contrast, the proposed part-aware Transformer architecture of our P3D is capable of handling the long-term dependency and part-specific context at the same time. As a result, our P3D shows better performance compared to ST-GCN on every split of WLASL, proving the better representation power of proposed part-wise motion context learning. + +Jointly using 2D and 3D pose. The third block of Table 6 shows that our P3D scores highly superior results compared to previous methods with using both 2D and 3D poses as an input. Again, in order to conduct fair comparison, identical input features are used and the same pose ensemble method is used for every tested models. We performed a comparative analysis in two aspects as follows. + +First, our P3D resulted in a much higher performance gain from jointly using 2D and 3D poses. While our P3D gained $6.39\%$ top-1 per-instance recognition accuracy on WLASL2000 by pose ensemble, gains of previous state + +
MethodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
RGB-based
TK-3D ConvNet [17]77.55---68.75-------
Fusion-3 [19]75.67---68.3-------
I3D [7]65.8984.1167.0184.5856.1479.9456.2478.3832.4857.31--
Pose-based (2D)
Pose-TGCN [30]55.4378.68--38.4267.51--23.6551.75--
GCN-BERT [15]60.1583.98--42.1871.71------
SPOTER [3]63.18---43.18---28.8661.7828.2160.26
ST-GCN [50]50.7879.0751.6279.4744.4673.0545.2973.1634.4066.4732.5365.45
P3D (Ours)60.2485.9460.4885.6650.9377.7150.8178.5535.9668.3133.2266.20
Pose-based (2D + 3D)
SPOTER [3]73.9091.5774.5191.5554.0282.2055.2882.6430.9164.2529.8862.7
ST-GCN [50]56.6379.5156.9379.7145.8276.0145.9676.2237.3772.1335.6170.66
P3D (Ours)75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
Pose-based (2D + 3D + facial expression)
P3D (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
+ +Table 6. Comparisons with state-of-the-art methods. We report the recognition accuracy of previous methods for comparison with ours. The results in each block are trained and evaluated with using exact same input denoted in the first column. Results in the first block show the recognition accuracy of existing RGB-based approaches. The second block represents the pose-based approach with 2D pose inputs. Our result in the second block is also generated solely with 2D poses. Scores in the third block are obtained from employing pose ensemble on previous methods and ours, using both 2D and 3D poses. Note that previous methods did not utilize 3D pose in their works, thus we applied pose ensemble based on released codes. Finally, the fourth block shows the best performance of our model using 2D, 3D poses and facial expression. Blank scores denoted with a hyphen are not reported in previous works. + +of-the-art methods SPOTER [3] and ST-GCN [50] are confined to $2.05\%$ and $2.97\%$ , respectively. Such observation suggests that our part-wise motion context learning is more capable of acquiring contextual information from various types of pose representations, which distinguishes it from the basic Transformer architecture of SPOTER [3] or the spatio-temporal graph convolution of ST-GCN [50]. + +Second, P3D achieved superior performance compared to existing methods, particularly by a large margin on a large number of classes such as WLASL300 and WLASL2000. As demonstrated in the third block of Table 6, our P3D outperformed previous pose-based methods SPOTER [3] and ST-GCN [50] on every subset of WLASL. + +Using 3D expressive pose. The fourth block of Table 6 shows the performance with full input features: 2D pose, 3D positional and rotational pose, and facial expression. With additional usage of facial expression over the third block of Table 6, we observed the performance gain of $0.79\%$ , $0.93\%$ , and $2.12\%$ , respectively for each subset of WLASL. Moreover, when compared to the RGB-based models presented in the first block of Table 6, our P3D scored comparable results in WLASL100 and WLASL300, and outperformed in WLASL2000. Our framework is the first to outperform (WASL 2000) or achieve comparable performance (WASL 100, WASL 300) with RGB-based methods. + +# 5. Conclusion + +In this work, we proposed the sign language recognition framework P3D by leveraging part-wise 3D motion context learning. We utilize the alternating PET and WET layers, specializing in learning intra-part and inter-part motion contexts, respectively. Learning part-wise motion context resulted in superior performance on the WLASL benchmark compared to previous methods. Moreover, we jointly utilized 2D and 3D poses via joint-wise pose concatenation to fully exploit the performance of pose-based sign language recognition. We observed that joint-wise concatenation has advantages on both sides of performance and efficiency. As a result, our P3D outperformed the previous methods by a large margin on WLASL, jointly using 2D and 3D poses. + +# Acknowledgements + +This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], the NRF grant [No. 2021M3A9E4080782] funded by the Korea government (MSIT), and the SNU-Samsung SDS Research Center. + +# References + +[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In ECCV, 2020. 1 +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 3, 5 +[3] Matyáš Boháček and Marek Hrúz. Sign pose-based transformer for word-level sign language recognition. In WACV, 2022. 2, 3, 8, 9 +[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, 2020. 1 +[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 1 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 3, 9 +[8] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 3 +[9] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV, 2020. 3 +[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 3 +[11] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In CVPR, 2017. 3 +[12] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia (ToM), 21(7):1880-1891, 2019. 3 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[14] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1 +[15] Haodong Duan, Yue Zhao, Kai Chen, Dian Shao, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 9 +[16] Zhimin Gao, Peitao Wang, Pei Lv, Xiaoheng Jiang, Qidong Liu, Pichao Wang, Mingliang Xu, and Wanqing Li. Focal + +and global spatial-temporal transformer for skeleton-based action recognition. In ACCV, 2022. 3 +[17] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In CVPR, 2017. 9 +[18] Xiao Guo and Jongmoo Choi. Human motion prediction via learning local structure representations and temporal dependencies. In AAAI, 2019. 2 +[19] Al Amin Hosain, Panneer Selvam Santhalingam, Parth Pathak, Huzefa Rangwala, and Jana Kosecka. Hand pose guided 3d pooling for word-level sign language recognition. In WACV, 2021. 3, 9 +[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: pre-training of hand-model-aware representation for sign language recognition. In ICCV, 2021. 1, 2, 3, 4, 6 +[21] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding, 2014. 3 +[22] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. In BMVC, 2018. 3 +[23] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2014. 3 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[25] Oscar Koller. Quantitative survey of the state of the art in sign language recognition, 2020. 1 +[26] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(9):2306-2320, 2019. 3 +[27] Oscar Koller, Sepehr Zargaran, Hermann Ney, and Richard Bowden. Deep sign: Enabling robust statistical continuous sign language recognition via hybrid cnn-hmms. International Journal of Computer Vision (IJCV), 126:1311-1325, 2018. 3 +[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 3 +[29] Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Cooccurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation. arXiv preprint arXiv:1804.06055, 2018. 1 +[30] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, 2020. 1, 2, 3, 6, 7, 9 + +[31] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In CVPR, 2020. 1 +[32] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-parted graph scattering networks for 3d human motion prediction. In ECCV, 2022. 2 +[33] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In CVPR, 2022. 3 +[34] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 3 +[35] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, Oct. 2015. 4 +[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR Workshops, 2022. 1, 3, 5, 6 +[37] Gyeongsik Moon, Heeseung Kwon, Kyoung Mu Lee, and Minsu Cho. Integralaction: Pose-driven feature integration for robust human action recognition in videos. In CVPRW, 2021. 1 +[38] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 3 +[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. *NeurIPS Workshop on Autodiff*, 2017. 6, 8 +[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3 +[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019, 2, 4 +[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In ICPR Workshops, 2021. 3 +[43] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 3 +[44] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3d multi-channel sign language production via progressive transformers and mixture density networks. International Journal of Computer Vision (IJCV), 129(7):2113-2135, 2021. 3 +[45] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In AAAI, 2017. 1 + +[46] Du Tran, Lubomir Bourdev, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. ICCV, 2015. 3 +[47] Anirudh Tunga, Sai Vidyaranya Nuthalapati, and Juan Wachs. Pose-based sign language recognition using gcn and bert. In WACV, 2021. 1, 2, 3 +[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 2008. 7 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3, 4, 5 +[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3, 8, 9 +[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2, 3, 5 +[52] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 3 +[53] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 3 +[54] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4 \ No newline at end of file diff --git a/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/images.zip b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..35aef7b43229a80feae709ee8075d3bd6746ab80 --- /dev/null +++ b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f96f73a69787d1b061676d15bed1093484004547317879fdab971393c5cfd333 +size 649685 diff --git a/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/layout.json b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..51457317c0e8c83a1adc327237b1980289e09531 --- /dev/null +++ b/2023/Human Part-wise 3D Motion Context Learning for Sign Language Recognition/layout.json @@ -0,0 +1,7670 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 103, + 444, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 103, + 444, + 140 + ], + "spans": [ + { + "bbox": [ + 151, + 103, + 444, + 140 + ], + "type": "text", + "content": "Human Part-wise 3D Motion Context Learning for Sign Language Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 160, + 228, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 160, + 228, + 175 + ], + "spans": [ + { + "bbox": [ + 154, + 160, + 228, + 175 + ], + "type": "text", + "content": "Taeryung Lee" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 161, + 321, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 161, + 321, + 175 + ], + "spans": [ + { + "bbox": [ + 253, + 161, + 321, + 175 + ], + "type": "text", + "content": "Yeonguk Oh²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 345, + 161, + 437, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 161, + 437, + 175 + ], + "spans": [ + { + "bbox": [ + 345, + 161, + 437, + 175 + ], + "type": "text", + "content": "Kyoung Mu Lee" + }, + { + "bbox": [ + 345, + 161, + 437, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 175, + 470, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 175, + 470, + 190 + ], + "spans": [ + { + "bbox": [ + 124, + 175, + 470, + 190 + ], + "type": "text", + "content": "1IPAI, 2Dept. of ECE & ASRI, Seoul National University, Seoul, Korea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 191, + 405, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 191, + 405, + 202 + ], + "spans": [ + { + "bbox": [ + 188, + 191, + 405, + 202 + ], + "type": "text", + "content": "{trlee94, namepllet, kyoungmu}@snu.ac.kr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 257, + 290, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 533 + ], + "type": "text", + "content": "In this paper, we propose P3D, the human part-wise motion context learning framework for sign language recognition. Our main contributions lie in two dimensions: learning the part-wise motion context and employing the pose ensemble to utilize 2D and 3D pose jointly. First, our empirical observation implies that part-wise context encoding benefits the performance of sign language recognition. While previous methods of sign language recognition learned motion context from the sequence of the entire pose, we argue that such methods cannot exploit part-specific motion context. In order to utilize part-wise motion context, we propose the alternating combination of a part-wise encoding Transformer (PET) and a whole-body encoding Transformer (WET). PET encodes the motion contexts from a part sequence, while WET merges them into a unified context. By learning part-wise motion context, our P3D achieves superior performance on WLASL compared to previous state-of-the-art methods. Second, our framework is the first to ensemble 2D and 3D poses for sign language recognition. Since the 3D pose holds rich motion context and depth information to distinguish the words, our P3D outperformed the previous state-of-the-art methods employing a pose ensemble." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 559, + 128, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 128, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 128, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 579, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 288, + 700 + ], + "type": "text", + "content": "Understanding and translating sign language is essential due to the huge potential benefit to society. Sign language recognition (SLR) can be considered as a specific kind of action recognition (AR) since both tasks aim to classify human motion. However, SLR has a characteristic that is distinguished from AR; Actions mainly focus on the motion of the human body, while the SLR system should be aware of more detailed information such as hand gestures and facial expressions. Thus, a specified system for SLR will provide further gain over using the AR models on SLR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Previous approaches on SLR problem are mainly cat-" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 238, + 544, + 473 + ], + "blocks": [ + { + "bbox": [ + 310, + 238, + 544, + 473 + ], + "lines": [ + { + "bbox": [ + 310, + 238, + 544, + 473 + ], + "spans": [ + { + "bbox": [ + 310, + 238, + 544, + 473 + ], + "type": "image", + "image_path": "36f551132757c4f08a38b8a6b46bc1bb40fda5afe7af3c332a471f50c0e43b91.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 485, + 547, + 541 + ], + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 541 + ], + "type": "text", + "content": "Figure 1. Overall pipeline of proposed system. We first extract expressive 3D human pose from RGB video using off-the-shelf pose estimation methods [5, 36]. The expressive 3D human pose consists of 2D pose, 3D pose and facial expression. Then our proposed P3D predicts the word from the expressive 3D human pose." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "text", + "content": "egorized into RGB-based [1, 4, 31] and pose-based methods [25, 29, 45, 47, 50]. Pose-based methods are known to have the advantage over RGB-based approaches in terms of the robustness on domain gap [14, 37]. For that, we aimed to enhance the performance of SLR, while preserving the benefits inherent in the pose-based approach." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Although recent pose-based SLR methods have shown noticeable progress, there are two main limitations of existing methods. First, they utilize only whole-body representation (e.g., skeleton) to understand the motion context of human [20, 30, 47]. Since such approaches overlook the detailed motion context within a specific part" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20740" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 77, + 312, + 338 + ], + "blocks": [ + { + "bbox": [ + 52, + 133, + 105, + 145 + ], + "lines": [ + { + "bbox": [ + 52, + 133, + 105, + 145 + ], + "spans": [ + { + "bbox": [ + 52, + 133, + 105, + 145 + ], + "type": "text", + "content": "RGB Video" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 207, + 93, + 219 + ], + "lines": [ + { + "bbox": [ + 53, + 207, + 93, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 93, + 219 + ], + "type": "text", + "content": "2D Pose" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 281, + 93, + 292 + ], + "lines": [ + { + "bbox": [ + 54, + 281, + 93, + 292 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 93, + 292 + ], + "type": "text", + "content": "3D Pose" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 120, + 77, + 312, + 338 + ], + "lines": [ + { + "bbox": [ + 120, + 77, + 312, + 338 + ], + "spans": [ + { + "bbox": [ + 120, + 77, + 312, + 338 + ], + "type": "image", + "image_path": "c20a43ea8f81905307522bb4d63b6a22d2fb6a0f24bbbc8871d5cfa02d403e79.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 546, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 546, + 398 + ], + "type": "text", + "content": "Figure 2. Effectiveness of 3D pose for sign language recognition. The 2D poses of the words \"Avoid\" and \"Behind\" are almost identical. Due to the absence of depth information, solely relying on 2D pose makes it hard to distinguish the words. On the other hand, as shown in the red boxes, additionally employing the 3D pose with rich depth information can effectively resolve the ambiguity in sign language recognition." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 345, + 76, + 543, + 338 + ], + "blocks": [ + { + "bbox": [ + 345, + 76, + 543, + 338 + ], + "lines": [ + { + "bbox": [ + 345, + 76, + 543, + 338 + ], + "spans": [ + { + "bbox": [ + 345, + 76, + 543, + 338 + ], + "type": "image", + "image_path": "d51f5a1814776b316bc991fb7e10f444b1817324401eb1934b0a65484091baab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 417, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 489 + ], + "type": "text", + "content": "of human [18, 32], it leads to a decrease in SLR accuracy. Second, existing pose-based methods only employ 2D pose [3, 20, 30, 47], which induces ambiguities in distinguishing the different words. As shown in Figure 2, the 2D pose of different words can be almost identical due to the absence of depth information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": "To overcome the above limitations, we propose P3D, the human part-wise motion context learning framework for accurate sign language recognition. The proposed P3D employs the part-wise encoding Transformer (PET) and whole-body encoding Transformer (WET) as main components. Given the pose sequence, we first divide it into four parts: body, right hand, left hand, and face. In contrast to previous methods where entire human poses are jointly processed, we utilize part-specific PET to encode the motion context of each part. As each PET takes a pose sequence of a single part, it can learn more detailed motion context, which is pivotal for sign language recognition. Then, the encoded features from each part are merged by WET. While PET only considers the intra-part motion context, WET covers the entire part. For the effective fusion of the inter-part and intra-part motion context, we alternately combined PET and WET inspired by spatio-temporal Transformers [2,51]. As a result, our proposed P3D has out" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "content": "performed the previous state-of-the-art [3, 50] on WLASL. Our ablation study on layer design supports part-wise motion context learning benefits SLR performance. Especially, we observed that solely forming the model with WET (i.e., without part-wise motion context) degrades the performance, thus directly induces that our PET contributed to improving the accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 502, + 546, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 546, + 668 + ], + "type": "text", + "content": "We alleviate the second limitation of previous methods through the ensemble of 2D and 3D poses. In terms of 3D pose, we utilize both 3D positional pose and 3D rotational pose (e.g., SMPL-X [41] pose parameters). The 3D positional pose complements depth information that the 2D pose does not contain, and the 3D rotational pose provides rich motion context on the kinematic chain. Our experiment shows that our P3D outperforms the previous methods by a large margin with using identical inputs; our model shows the superior result from only 2D pose, and also from ensemble of 2D and 3D pose. Moreover, our framework is the first pose-based method that recorded a comparable score or even outperformed the RGB-based methods. We claim our main contributions as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 317, + 677, + 546, + 715 + ], + "type": "text", + "content": "- We proposed P3D, the human part-wise motion context learning framework for sign language recognition. Our experiment supports that part-wise learning highly" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20741" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 107 + ], + "type": "text", + "content": "benefits recognition performance. As a result, our proposed P3D outperformed the previous SOTA on the WLASL benchmark." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 118, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 118, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 118, + 286, + 178 + ], + "type": "text", + "content": "- We employ the pose ensemble by joint-wise pose concatenation in order to exploit both 2D and 3D poses. We observed significant performance gain via pose ensemble, and P3D outperformed the previous methods by a large margin by jointly using 2D and 3D poses." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 191, + 135, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 135, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 135, + 202 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 211, + 190, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 190, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 190, + 224 + ], + "type": "text", + "content": "2.1. Sign language recognition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 230, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 230, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 230, + 286, + 289 + ], + "type": "text", + "content": "Word-level sign language recognition task aims to classify short-term videos by the corresponding word label. Previous works on sign language recognition can be mainly categorized into two groups by the input modality: videobased and pose-based approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 290, + 286, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 290, + 286, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 286, + 505 + ], + "type": "text", + "content": "Video-based sign language recognition. With the significant performance of the CNN-based action recognition, the majority of the prior works have designed their models based on CNN backbone [9, 27, 53]. Primary works [11, 12, 26, 44] first used convolutional neural networks to extract the frame-wise features. Then, such methods encoded the temporal semantics along the time dimension using recurrent neural networks. On the other hand, similar to action recognition, 3D CNNs are widely used in sign language recognition. C3D [21, 23, 46] is the first 3D CNN model for action recognition, and it is widely used as a baseline model for quantitative comparison. I3D [7] is proposed based on the C3D [46] architecture and applied on sign language recognition in [22]. Li et al. [30] have compared various architectures, including 3D CNN and 2D CNN combined with RNN. Hosain et al. [19] proposed Fusion-3 model based on I3D Inception-v1 [7] with pose-guided pooling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 506, + 286, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 286, + 613 + ], + "type": "text", + "content": "Pose-based sign language recognition. Compared to the RGB video, the human pose sequence holds a compact and semantic-aware representation of human motion. ST-GCN [50] is the first pose-based approach to action recognition, utilizing the spatio-temporal graph convolutional network to encode the motion on kinetic chain. Following works also have used the spatio-temporal architecture based on graph convolutional network [30, 47] and Transformers [49] [3, 20]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 623, + 243, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 243, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 243, + 635 + ], + "type": "text", + "content": "2.2. 3D human pose and shape estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "type": "text", + "content": "3D human pose and shape estimation aims to localize 3D coordinates of vertices of the human mesh. Considering the various applications such as VR/AR and human action recognition, research on 3D human pose and shape estimation has been widely conducted. Most works tried to reconstruct 3D human pose and shape for specific part of human," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": "e.g. body [10, 34, 38], hand [8, 28], or face [43]. Several works [36, 40] proposed methods that simultaneously estimate the 3D pose and shape for all human parts, i.e., whole-body. By reconstructing 3D poses and shapes for all human parts, we can understand human intention and feelings more precisely. In this work, we leverage Hand4Whole [36] to acquire a 3D pose of all human parts from RGB videos. To the best of our knowledge, we first utilize 3D pose for sign language recognition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 189, + 481, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 481, + 201 + ], + "type": "text", + "content": "2.3. Transformers for human motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 208, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 495 + ], + "type": "text", + "content": "After the remarkable success of Transformers in natural language processing (NLP) [49], vision researchers have employed Transformers for various vision tasks [6, 13, 34]. Transformers also have been widely adopted for tasks that require human motion understanding, such as 2D-to-3D pose sequence lifting [33, 51, 52] and skeleton-based action recognition [16, 42]. In the field of 2D-to-3D pose sequence lifting, PoseFormer [52] proposed a pure Transformer-based approach for 2D-to-3D pose sequence lifting. MixSTE [51] utilized separated joint-wise temporal Transformer and spatial Transformer alternately. MH-Former [33] employed Transformers to generate and refine multiple hypotheses for 3D pose sequences from 2D pose sequences. In the field of skeleton-based action recognition, ST-TR [42] proposed Transformer-based two-stream networks where each stream learns spatial and temporal correlations in a skeleton sequence. FG-STFormer [16] presented a variation of vanilla Transformer to capture local and global information effectively. Our P3D also leverages Transformer-based modules to learn the human motion. However, most previous works only focused on the motion of the human body, while proposed P3D learns the motion of all human parts, including hands and facial expressions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 506, + 361, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 361, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 361, + 518 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 527, + 372, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 372, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 372, + 538 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 545, + 713 + ], + "type": "text", + "content": "In this section, we describe the input data format and architecture of our P3D model. Figure 3 illustrates the overview of our model. We provide a sequence of human poses to our model as an input P3D leverages alternating structure of Transformer layers [49], inspired by the spatiotemporal architectures from the previous works [2,51]. The first layer, the part-wise encoding Transformer (PET), is designed to learn intra-part motion context. To capture the specific motion context from each part, we assign the dedicated Transformers. Meanwhile, the second whole-body encoding Transformer (WET) encodes the inter-part motion context, taking account of every human part used in the model. As the last step, we take mean along the time axis on encoded representations, then sequentially pass through" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20742" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 547, + 285 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 547, + 285 + ], + "type": "image", + "image_path": "45217de8a3a149ffce1ddca5e57d5d2d2ab4054588dc19503294084dbe976709.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 297, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 45, + 297, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 45, + 297, + 547, + 353 + ], + "type": "text", + "content": "Figure 3. Overall architecture of P3D. Our P3D aims to predict the word label corresponding to a given pose sequence. We first provide joint-wise linearly embedded feature vectors as an input, along with linearly encoded facial expression features. P3D consists of " + }, + { + "bbox": [ + 45, + 297, + 547, + 353 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 45, + 297, + 547, + 353 + ], + "type": "text", + "content": " alternating layers of PET, which are assigned for each part, and WET, which covers entire parts. After the encoder, the prediction head takes the mean along the time dimension, then sequentially passes through batch normalization, linear layer, and softmax layer to produce the word prediction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 379, + 286, + 582 + ], + "blocks": [ + { + "bbox": [ + 50, + 379, + 286, + 582 + ], + "lines": [ + { + "bbox": [ + 50, + 379, + 286, + 582 + ], + "spans": [ + { + "bbox": [ + 50, + 379, + 286, + 582 + ], + "type": "image", + "image_path": "08a61810a0b220b36d9a35453cd2adb752e3ed6af2523e9d86f3943f2745b207.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 593, + 288, + 639 + ], + "lines": [ + { + "bbox": [ + 46, + 593, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 288, + 639 + ], + "type": "text", + "content": "Figure 4. Overall architectures of PET and WET. We adapt the Transformer encoder layer [49] with vanilla self-attention for both PET (above) and WET (below). For the PET, the subscript " + }, + { + "bbox": [ + 46, + 593, + 288, + 639 + ], + "type": "inline_equation", + "content": "p \\in \\{B, RH, LH, \\psi\\}" + }, + { + "bbox": [ + 46, + 593, + 288, + 639 + ], + "type": "text", + "content": " denotes the assigned human part." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "batch normalization, linear layer, and softmax layer to get the final output." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 371, + 389, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 389, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 389, + 384 + ], + "type": "text", + "content": "3.2. Input format" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": "Human motion representation. The overview of the input data structure is illustrated on the left of Figure 3. As a pose-based sign language model, our P3D takes a sequence " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "(X_{1},\\dots,X_{T})\\in \\mathbb{R}^{T\\times (J\\times F + E)}" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " human poses as an input. Pose representation " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "X_{t} = (\\mathbf{x}_{1},\\dots,\\mathbf{x}_{J},e)\\in \\mathbb{R}^{J\\times F + E}" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " for each time " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "t\\in \\{1,2,\\dots,T\\}" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " consists of joint-wise pose representation " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j\\in \\mathbb{R}^F" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " for each joint " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "j\\in \\{1,2,\\dots,J\\}" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": ", and facial expression feature " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "e\\in \\mathbb{R}^E" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " [41]. From our experiment on the pose fusion method, we have observed that joint-wise concatenation of the 2D positional pose (2D pos.), 3D positional pose (3D pos.), and 3D rotational pose (3D rot.) performed the best on WLASL2000. Thus, we form the joint-wise pose representation " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j = (\\mathbf{x}_{j,2\\mathrm{D}\\mathrm{pos}},\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{pos}},\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{rot}})\\in \\mathbb{R}^F" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{j,2\\mathrm{D}\\mathrm{pos}}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{pos}}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{j,3\\mathrm{D}\\mathrm{rot}}\\in \\mathbb{R}^6" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " represents the 2D positional pose, 3D positional pose and 3D rotational pose [54] of joint " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": ", respectively. Please refer to our experiment on Section 4.3 for the details of pose fusion methods. Following the literature [20], we use " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "T = 32" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " for both training and testing. Joint set and human part. We use " + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "inline_equation", + "content": "J = 40" + }, + { + "bbox": [ + 304, + 390, + 547, + 715 + ], + "type": "text", + "content": " joints and 4 human parts to represent the human motion of the sign language. Ten joints above the pelvis are selected from the SMPL [35] joint set, defined as a body part. Fifteen joints, three per finger, represent each hands. For the face, we use the 10-dimensional feature vector from SMPL-X [41] to represent the facial expression. Please refer to Section 4.2 for the details of human pose preprocessing." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20743" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 77, + 539, + 335 + ], + "blocks": [ + { + "bbox": [ + 60, + 77, + 539, + 335 + ], + "lines": [ + { + "bbox": [ + 60, + 77, + 539, + 335 + ], + "spans": [ + { + "bbox": [ + 60, + 77, + 539, + 335 + ], + "type": "image", + "image_path": "5f00008c5beaeb8e9ca0566a224d0c86d61cd3484bb03538b5c4c09fe1bc2997.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 349, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 349, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 547, + 373 + ], + "type": "text", + "content": "Figure 5. Visualization. The 3D expressive pose of the signer (b) is estimated from the given sign language video (a) with off-the-shelf estimator Hand4Whole [36]. The corresponding word label (c) is denoted in the right." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 391, + 130, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 130, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 130, + 403 + ], + "type": "text", + "content": "3.3. Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": "Linear layer. Our model first linearly encodes each jointwise pose representation " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j \\in \\mathbb{R}^F" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": " into D dimensional vector. We also assign a dedicated linear layer to encode facial expressions into " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\alpha D" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": " dimensional vector. Empirically, we decided to use " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "D = 8" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": ". As a result, pose " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "X_t = (\\mathbf{x}_1, \\dots, \\mathbf{x}_J, e) \\in \\mathbb{R}^{J \\times F + E}" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": " is encoded into " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "X_t^{(0)} \\in \\mathbb{R}^{(J \\times D + \\alpha D) = ((10 + 15 + 15) \\times 8 + 80)}" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": ". Note that 10, 15, and 15 represent the number of joints in the body, left hand and right hand, and thus can be separated into each human part via indexing. Finally we form the input " + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "inline_equation", + "content": "X^{(0)} = (X_1^{(0)}, X_2^{(0)}, \\dots, X_T^{(0)}) \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)}" + }, + { + "bbox": [ + 46, + 409, + 287, + 556 + ], + "type": "text", + "content": " with positional encoding [49]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": "Alternating Transformers. We adapted the idea of alternating two types of layers [2, 51], which have different purposes and expertise. Our P3D consists of " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " alternating layers: " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " Part-wise Encoding Transformer (PET) and " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " Whole-body Encoding Transformer (WET). Starting from linear encoded input " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "X^{(0)}" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": ", we denote the output of " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": "'th PET as " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "Y^{(i)}" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " and the output of " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": "'th WET as " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "X^{(i)}" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": ". As the dimension remains identical through Transformers, our features " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "X^{(i)}" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "Y^{(i)}" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " remain in a same dimension as " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "X^{(0)}" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": ". Empirically, we use " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " layers for WLASL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Part-wise encoding Transformer (PET). The architecture of PET is illustrated in the upper part of Figure 4. From previous layer, PET takes 2D tensor " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "X^{(i)} \\in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^T \\times ((10 + 15 + 15) \\times 8 + 80)" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": " as an input. The first (horizontal) and second (vertical) dimensions represent the time and feature axis, respectively. We first separate features for each human part " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "X_B^{(i)} \\in \\mathbb{R}^{T \\times 80}" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "X_{LH}^{(i)} \\in \\mathbb{R}^{T \\times 120}" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "X_{RH}^{(i)} \\in \\mathbb{R}^{T \\times 120}" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "X_{\\psi}^{(i)} \\in \\mathbb{R}^{T \\times 80}" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": " via slicing on second (feature) dimension by indices (80, 200, 320, 400), respectively. Then we pass each feature through dedicated Transformer encoder layers " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{PET}_B" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{PET}_{RH}" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{PET}_{LH}" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{PET}_{\\psi}" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": " assigned for each human part, respectively. Finally, we concatenate the outputs along the second (feature) dimension into a single 2D tensor " + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "inline_equation", + "content": "Y^{(i)} \\in \\mathbb{R}^T \\times ((10 + 15 + 15) \\times 8 + 80)" + }, + { + "bbox": [ + 304, + 389, + 547, + 527 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 530, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 547, + 625 + ], + "type": "text", + "content": "Whole-body encoding Transformer (WET). The architecture of WET is illustrated in the lower part of Figure 4. From the previous layer PET, WET takes 2D tensor " + }, + { + "bbox": [ + 304, + 530, + 547, + 625 + ], + "type": "inline_equation", + "content": "Y^{(i)} \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)}" + }, + { + "bbox": [ + 304, + 530, + 547, + 625 + ], + "type": "text", + "content": " as an input. Since WET aims to learn inter-part motion context, we assign a single Transformer to encode the whole motion sequence. Output of WET " + }, + { + "bbox": [ + 304, + 530, + 547, + 625 + ], + "type": "inline_equation", + "content": "X^{(i + 1)} \\in \\mathbb{R}^{T \\times ((10 + 15 + 15) \\times 8 + 80)}" + }, + { + "bbox": [ + 304, + 530, + 547, + 625 + ], + "type": "text", + "content": " is passed to next layer without modification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Prediction head. From Transformer-encoded feature " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "X^{(N)}\\in \\mathbb{R}^{T\\times 400}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": ", we first calculate the mean feature vector " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{feat}}\\in \\mathbb{R}^{400}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " by taking the mean along the first (time) dimension. Then, we pass " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " sequentially through 1D batch normalization, a single FC-layer, and a softmax layer. As a result, our model produces the final prediction " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{class}}\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " is the number of target classes (glosses) in the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20744" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 80, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 80, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 80, + 83 + ], + "type": "text", + "content": "dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "text", + "content": "Training and inference. We train P3D with a cross-entropy loss between predicted " + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{class}}" + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "text", + "content": " and one-hot encoded GT class vector. We randomly sample continuous " + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "text", + "content": " frames to form the training inputs. During the inference phase, we sample four chunks of " + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "text", + "content": " continuous frames from the human pose sequence in the test set following the previous work [30]. Then, we use P3D to generate the class prediction " + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{class}}" + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 84, + 287, + 193 + ], + "type": "text", + "content": " continuous frame, and average them out to produce the final prediction for the entire sequence." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 204, + 123, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 123, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 123, + 217 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 224, + 178, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 178, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 178, + 237 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 243, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 287, + 315 + ], + "type": "text", + "content": "We used PyTorch [39] Transformer implementation for our PET and WET, with 256 inner dimensions, 0.3 dropout rate, GELU activation layer, and 4 multi-head attentions. We trained our model for 500 epochs using Adam [24] optimizer with a mini-batch size of 512, a learning rate of " + }, + { + "bbox": [ + 47, + 243, + 287, + 315 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 47, + 243, + 287, + 315 + ], + "type": "text", + "content": ", the weight decay rate of " + }, + { + "bbox": [ + 47, + 243, + 287, + 315 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-3}" + }, + { + "bbox": [ + 47, + 243, + 287, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 323, + 286, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 286, + 336 + ], + "type": "text", + "content": "4.2. Datasets, evaluation metrics and preprocessing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 342, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 286, + 449 + ], + "type": "text", + "content": "WLASL dataset. WLASL [30] is a large-scale word-level sign language dataset that contains 21,083 videos with 119 signers. The videos in WLASL are collected from the Internet, and each video depicts a signer performing a single sign. The WLASL dataset consists of 2,000 words, and it also provides subsets of the dataset, WLASL100, and WLASL300. Nevertheless, our main target is to push through the limit of recognition performance on WLASL2000." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 450, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 286, + 557 + ], + "type": "text", + "content": "Evaluation metrics For the evaluation metrics, we employ Top-1 and Top-5 accuracy. Following the previous work [20], we calculate both metrics under two different settings, i.e., per-instance and per-class. In per-instance setting, we measure the percentage of accurately classified video over the entire test instances. Per-class setting is also utilized since each class (gloss) has a different number of instances. In a per-class setting, we calculate the accuracy of each class and average them over all classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 558, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 286, + 713 + ], + "type": "text", + "content": "Pose preprocessing Regarding the 2D pose, we utilize the dedicated 2D pose estimator for better accuracy, since direct estimation of 2D pose is more accurate compared to projection from estimated 3D pose. The 2D pose is aligned based on the bounding box of the signer by applying the affine transform. After then, we rescaled the 2D pose sequence from the bounding box into the [0, 1] range. The 3D poses, including the positional pose, the rotational pose and the facial expression, we utilize the off-the-shelf 3D expressive human pose estimator Hand4Whole [36] as illustrated in Figure 5. For the 3D positional pose, we aligned the translation by placing the root joint of the first frame at the origin without normalization. In addition, the 3D rotational" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "pose and the facial expression features are used without preprocessing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 106, + 397, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 397, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 397, + 118 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 125, + 545, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 125, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 306, + 125, + 545, + 173 + ], + "type": "text", + "content": "We conduct an ablation study on two main contributions of our framework: part-wise motion context learning and pose ensemble. For each experiment, the contribution that is not being tested remains identical to our proposed P3D." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 174, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 174, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 306, + 174, + 545, + 268 + ], + "type": "text", + "content": "Effect of part-wise motion context learning. In Table 1, we compare the results of various pose encoding methods in single iteration of P3D. The MLP is the simple alternative to the proposed PET or WET, which is constructed of stacked fully-connected layers, batch normalization, and dropout. Experimental results demonstrate that employing our PET and WET gives a large performance gain over MLP alternatives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 270, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 473 + ], + "type": "text", + "content": "We also report that a combination of PET and WET is essential in sign language recognition. When the number of classes is small (WLASL 100, WLASL 300), utilizing one of the WET (the fourth row) or PET (the third row) provides a similar result to utilizing both PET and WET (the fifth row). However, when the number of classes is large (WLASL 2000), utilizing both PET and WET performs the best. As the number of classes increases, learning highly distinguishable features from motion context becomes crucial. For that, the PET and WET are dedicated to capturing intra-part context and merging them into a unified inter-part context, respectively. Such design benefits the task-specific representation power on sign language recognition, that each part (i.e. body, hands, and facial expressions) holds its own important motion context. As a result, our PET-WET encoder outperforms PET-PET and WET-WET by " + }, + { + "bbox": [ + 306, + 270, + 545, + 473 + ], + "type": "inline_equation", + "content": "3.75\\%" + }, + { + "bbox": [ + 306, + 270, + 545, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 270, + 545, + 473 + ], + "type": "inline_equation", + "content": "12.08\\%" + }, + { + "bbox": [ + 306, + 270, + 545, + 473 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 474, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 545, + 569 + ], + "type": "text", + "content": "Effect of each human part. In Table 2, we conduct an ablation study on employing different human part inputs. As our task is sign language recognition, hands are the most essential part of accurate recognition. However, the additional gain can be obtained by leveraging body pose and facial expression on top of hands. It demonstrates that body pose and facial expression, as well as hands, also provide helpful information for sign language recognition." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 545, + 713 + ], + "type": "text", + "content": "Effect of each pose representation. Table 3 shows that jointly utilizing 2D positional pose, 3D positional pose, and 3D rotational pose achieves the best performance. In terms of single representation, the 3D rotational pose is the most effective pose representation for sign language recognition. Furthermore, additionally using the 3D rotational pose to other pose representations (the first row & the fifth row, the second row & the sixth row, and the fourth row & last row) consistently improves the accuracy. As a reason for the effectiveness of 3D rotational pose for sign language recognition, we argue that 3D rotational pose is less sensitive to signer variation than 2D positional pose and 3D positional" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20745" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 533, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 533, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 533, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 533, + 159 + ], + "type": "table", + "html": "
Pose encoding methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
MLP-WET71.8985.9472.3387.2159.4482.2059.7082.4334.9770.5732.7869.41
MLP-PET72.2888.7673.7689.5557.8982.5058.5682.9634.7270.7132.4969.66
PET-PET77.9194.7778.3095.2567.3388.6868.0688.0632.3963.6130.3462.20
WET-WET75.5091.5776.4392.6866.8788.2466.9288.5640.7274.7038.2673.54
PET-WET (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
", + "image_path": "407965c4d30116382a1155174bb29da0d7ad7dedae9f09e029549ab52655ca96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 209, + 535, + 299 + ], + "blocks": [ + { + "bbox": [ + 46, + 166, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 201 + ], + "type": "text", + "content": "Table 1. Ablation study on pose encoding methods. We ablate the performance of pose encoding methods in single iteration of P3D. Our part-wise context learning (PET-WET) scored the best performance. The observation that ours (PET-WET) has outperformed WET-WET (Transformer without part) directly supports our main contribution of part-wise context learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 209, + 535, + 299 + ], + "lines": [ + { + "bbox": [ + 57, + 209, + 535, + 299 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 535, + 299 + ], + "type": "table", + "html": "
Input human partsWLASL 100WLASL 300WLASL 2000
BodyHandsExpr.Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
✓44.9874.7044.7575.8537.6270.4337.4070.1717.8040.7615.7037.57
✓✓46.9972.2945.5772.4338.7070.4338.7370.4520.9847.0518.7343.94
✓71.8991.1673.1092.0862.0787.1562.6987.9339.0073.4036.9772.16
✓✓75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.9742.1878.52
", + "image_path": "00aeb808166628aa9f8243a800addc05a6e232c6017dc1a610c6854a639c5c65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 546, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 305, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 546, + 329 + ], + "type": "text", + "content": "Table 2. Ablation study on input human parts for P3D. We ablate the performance of provided input human parts. Check marks in the first column implies the given input part. We abbreviate the facial expression to Expr." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 354, + 284, + 548 + ], + "blocks": [ + { + "bbox": [ + 50, + 354, + 284, + 548 + ], + "lines": [ + { + "bbox": [ + 50, + 354, + 284, + 548 + ], + "spans": [ + { + "bbox": [ + 50, + 354, + 284, + 548 + ], + "type": "image", + "image_path": "cee37f4d5c5b01a574e2559128833b559f9c77fe73c8a90f319818629e00b37f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 560, + 288, + 625 + ], + "lines": [ + { + "bbox": [ + 46, + 560, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 288, + 625 + ], + "type": "text", + "content": "Figure 6. t-SNE [48] visualization of features from different pose representations. Each plot corresponds to a single word (\"Corn\", \"Language\", \"Many\", and \"Thanksgiving\", respectively), and each point is feature " + }, + { + "bbox": [ + 46, + 560, + 288, + 625 + ], + "type": "inline_equation", + "content": "X^{(N)}" + }, + { + "bbox": [ + 46, + 560, + 288, + 625 + ], + "type": "text", + "content": " from each video. All test videos of the corresponding word in WLASL 2000 are employed for the visualization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "pose. Figure 6 illustrates that features of single words from different signers are closely embedded when input pose representation is a 3D rotational pose. On the other hand, in the case of the 2D and 3D positional poses, the embedded features of the single words show large discrepancies when the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 347, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 395 + ], + "type": "text", + "content": "signers are different. Such advantage of 3D rotational pose leads to improved performance over other pose representations since the sign languages in the WLASL dataset are signer-independent [30]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 396, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 491 + ], + "type": "text", + "content": "Pose ensemble methods. In Table 4, we compare the performances of different pose ensemble methods. Since we employ three types of pose representations (2D positional pose, 3D positional pose, and 3D rotational pose), how to ensemble information in each pose representation is an important issue. We design three strategies for ensemble methods, and the description for each ensemble method is as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 498, + 545, + 682 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 498, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 545, + 559 + ], + "type": "text", + "content": "- Early ensemble (Ours): All pose representations of each joint are concatenated before being fed to the PET-WET blocks. Thus, PET-WET blocks the process motion context of all poses representations simultaneously." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "text", + "content": "- Middle ensemble: Separate PET-WET blocks are utilized to extract feature " + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "inline_equation", + "content": "X^{(n)}" + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "text", + "content": " for each pose representation: " + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "inline_equation", + "content": "X_{2\\mathrm{D}\\mathrm{pos}}^{(n)}" + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "inline_equation", + "content": "X_{3\\mathrm{D}\\mathrm{pos}}^{(n)}" + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "inline_equation", + "content": "X_{3\\mathrm{D}\\mathrm{rot}}^{(n)}" + }, + { + "bbox": [ + 317, + 565, + 545, + 627 + ], + "type": "text", + "content": ". The extracted features of all pose representations are averaged before being fed to the prediction head." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 634, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 634, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 317, + 634, + 545, + 682 + ], + "type": "text", + "content": "- Late ensemble: Same as the middle ensemble, but each feature of each pose representation is fed to its own prediction head. The output of all prediction heads is averaged to regress a single word." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "The experimental results show that Early ensemble performs best for the WLASL 2000. In Table 5, we compare" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20746" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 535, + 180 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 535, + 180 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 535, + 180 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 535, + 180 + ], + "type": "table", + "html": "
Input pose representationsWLASL 100WLASL 300WLASL 2000
2D pos.3D pos.3D rot.Per-instancePer-classPer-instancePer-classPer-instancePer-class
Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5Top-1Top-5
✓60.2483.1360.4582.6352.1779.7252.5979.7935.9967.2633.4765.15
✓41.0966.6740.9867.0239.9769.7640.0170.0131.6962.8228.4259.83
✓72.8788.3773.6088.9361.6882.3461.9382.7838.6073.4235.5771.86
✓✓55.0281.1255.7382.0150.1580.1949.7481.0936.9169.0934.0067.06
✓✓75.5091.1676.8891.5865.9488.8566.2688.7543.4877.6441.3976.38
✓✓74.4288.3775.1589.2662.7283.5363.4384.0141.1476.7938.4775.34
✓✓✓76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
", + "image_path": "784aa2c91181ba8717ff92773b03211a6cad7e6b0b68a610b876886212be2e18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 220, + 533, + 288 + ], + "blocks": [ + { + "bbox": [ + 47, + 187, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 47, + 187, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 187, + 545, + 210 + ], + "type": "text", + "content": "Table 3. Ablation study on input pose representations. We ablate the performance of provided input human pose representations. Check marks in the first column implies the given input pose representation. Employing every pose representation scored the best result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 220, + 533, + 288 + ], + "lines": [ + { + "bbox": [ + 59, + 220, + 533, + 288 + ], + "spans": [ + { + "bbox": [ + 59, + 220, + 533, + 288 + ], + "type": "table", + "html": "
Ensemble methodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
Late77.9093.1778.3095.2567.3488.2368.0688.6832.7164.0430.3463.67
Middle79.9293.5781.6694.3867.8089.9467.8690.1342.6371.7140.1776.43
Early (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
", + "image_path": "01878a6e6eed9d73dc1d232c03a2ed687d34b35db097d7ddfa174cf99b780a75.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 335, + 285, + 406 + ], + "blocks": [ + { + "bbox": [ + 47, + 294, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 47, + 294, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 545, + 318 + ], + "type": "text", + "content": "Table 4. Ablation study on pose representation ensemble methods. We report the quantitative result on pose representation ensemble methods. Early ensemble by joint-wise concatenation outperforms the middle (between encoder & head) and late (after head) ensemble." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 335, + 285, + 406 + ], + "lines": [ + { + "bbox": [ + 50, + 335, + 285, + 406 + ], + "spans": [ + { + "bbox": [ + 50, + 335, + 285, + 406 + ], + "type": "table", + "html": "
Ensemble methodsFLOPs (G)# Params (M)
Late182.0610.71
Middle182.0210.65
Early (Ours)83.664.94
", + "image_path": "7365ddd32041143b3c3b793a994affea9e3f11f16607fe93ea032641898d6103.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 415, + 287, + 448 + ], + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 448 + ], + "type": "text", + "content": "Table 5. Computational costs of pose ensemble methods. We used the FLOPs counter for PyTorch [39] from fvcore package to measure the computational costs of the models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "type": "text", + "content": "the computational costs of ensemble methods. As separate PET-WET blocks are utilized in Middle ensemble and Late ensemble, Early ensemble is most efficient in terms of the number of model parameters and FLOPs. Considering the accuracy and efficiency, we adopt Early ensemble as our ensemble strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 550, + 267, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 267, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 267, + 562 + ], + "type": "text", + "content": "4.4. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 569, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 664 + ], + "type": "text", + "content": "Table 6 shows that P3D outperforms previous pose-based methods. We first demonstrate the capability of our proposed part-wise motion context learning model via quantitative comparison on the WLASL dataset using only 2D pose. Then we show that 3D pose fusion is more effective in our P3D than in other models, leading to our model outperforming others on the WLASL dataset using 2D and 3D pose." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Using only 2D pose. The second block of Table 6 shows that our P3D outperforms existing methods on WLASL 300 and WLASL 2000 split with 2D pose input. In this experiment, we only use 2D pose for the fair comparison of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 336, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 576 + ], + "type": "text", + "content": "models to learn highly distinguishable features from motion context. Prior to our work, state-of-the-art methods were SPOTER [3] in WLASL 100, and ST-GCN [50] in WLASL 300 and WLASL 2000. From the benefits of its simple and intuitive Transformer architecture, SPOTER achieved the best performance on WLASL 100. However, due to the limitation on representation power, SPOTER did not perform well on more complex problems such as WLASL 300 and WLASL 2000. Meanwhile, graph convolution via spatiotemporal adjacency used in ST-GCN is observed to be effective on a large number of classes (WLASL 300 and WLASL 2000). Yet, employing the spatio-temporal adjacency graph leads to two cons: it cannot model the long-term dependency, and cannot learn the part-specific context. In contrast, the proposed part-aware Transformer architecture of our P3D is capable of handling the long-term dependency and part-specific context at the same time. As a result, our P3D shows better performance compared to ST-GCN on every split of WLASL, proving the better representation power of proposed part-wise motion context learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "type": "text", + "content": "Jointly using 2D and 3D pose. The third block of Table 6 shows that our P3D scores highly superior results compared to previous methods with using both 2D and 3D poses as an input. Again, in order to conduct fair comparison, identical input features are used and the same pose ensemble method is used for every tested models. We performed a comparative analysis in two aspects as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "First, our P3D resulted in a much higher performance gain from jointly using 2D and 3D poses. While our P3D gained " + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "inline_equation", + "content": "6.39\\%" + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": " top-1 per-instance recognition accuracy on WLASL2000 by pose ensemble, gains of previous state" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20747" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 539, + 292 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 539, + 292 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 539, + 292 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 539, + 292 + ], + "type": "table", + "html": "
MethodsWLASL 100WLASL 300WLASL 2000
Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5Per-instance Top-1Top-5Per-class Top-1Top-5
RGB-based
TK-3D ConvNet [17]77.55---68.75-------
Fusion-3 [19]75.67---68.3-------
I3D [7]65.8984.1167.0184.5856.1479.9456.2478.3832.4857.31--
Pose-based (2D)
Pose-TGCN [30]55.4378.68--38.4267.51--23.6551.75--
GCN-BERT [15]60.1583.98--42.1871.71------
SPOTER [3]63.18---43.18---28.8661.7828.2160.26
ST-GCN [50]50.7879.0751.6279.4744.4673.0545.2973.1634.4066.4732.5365.45
P3D (Ours)60.2485.9460.4885.6650.9377.7150.8178.5535.9668.3133.2266.20
Pose-based (2D + 3D)
SPOTER [3]73.9091.5774.5191.5554.0282.2055.2882.6430.9164.2529.8862.7
ST-GCN [50]56.6379.5156.9379.7145.8276.0145.9676.2237.3772.1335.6170.66
P3D (Ours)75.9091.5678.1091.9766.2589.1665.9289.5242.3577.1540.7476.67
Pose-based (2D + 3D + facial expression)
P3D (Ours)76.7191.9778.2792.9767.1889.0167.6289.2444.4779.6942.1878.52
", + "image_path": "859c141ede4c5662590a1b9aa14ba5aeb1e37ab0702cdfcf5c7916e7ec862af4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 299, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 545, + 376 + ], + "type": "text", + "content": "Table 6. Comparisons with state-of-the-art methods. We report the recognition accuracy of previous methods for comparison with ours. The results in each block are trained and evaluated with using exact same input denoted in the first column. Results in the first block show the recognition accuracy of existing RGB-based approaches. The second block represents the pose-based approach with 2D pose inputs. Our result in the second block is also generated solely with 2D poses. Scores in the third block are obtained from employing pose ensemble on previous methods and ours, using both 2D and 3D poses. Note that previous methods did not utilize 3D pose in their works, thus we applied pose ensemble based on released codes. Finally, the fourth block shows the best performance of our model using 2D, 3D poses and facial expression. Blank scores denoted with a hyphen are not reported in previous works." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": "of-the-art methods SPOTER [3] and ST-GCN [50] are confined to " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "2.05\\%" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "inline_equation", + "content": "2.97\\%" + }, + { + "bbox": [ + 46, + 396, + 287, + 479 + ], + "type": "text", + "content": ", respectively. Such observation suggests that our part-wise motion context learning is more capable of acquiring contextual information from various types of pose representations, which distinguishes it from the basic Transformer architecture of SPOTER [3] or the spatio-temporal graph convolution of ST-GCN [50]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 555 + ], + "type": "text", + "content": "Second, P3D achieved superior performance compared to existing methods, particularly by a large margin on a large number of classes such as WLASL300 and WLASL2000. As demonstrated in the third block of Table 6, our P3D outperformed previous pose-based methods SPOTER [3] and ST-GCN [50] on every subset of WLASL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Using 3D expressive pose. The fourth block of Table 6 shows the performance with full input features: 2D pose, 3D positional and rotational pose, and facial expression. With additional usage of facial expression over the third block of Table 6, we observed the performance gain of " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.79\\%" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.93\\%" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "2.12\\%" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", respectively for each subset of WLASL. Moreover, when compared to the RGB-based models presented in the first block of Table 6, our P3D scored comparable results in WLASL100 and WLASL300, and outperformed in WLASL2000. Our framework is the first to outperform (WASL 2000) or achieve comparable performance (WASL 100, WASL 300) with RGB-based methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 394, + 378, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 378, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 378, + 407 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 422, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 579 + ], + "type": "text", + "content": "In this work, we proposed the sign language recognition framework P3D by leveraging part-wise 3D motion context learning. We utilize the alternating PET and WET layers, specializing in learning intra-part and inter-part motion contexts, respectively. Learning part-wise motion context resulted in superior performance on the WLASL benchmark compared to previous methods. Moreover, we jointly utilized 2D and 3D poses via joint-wise pose concatenation to fully exploit the performance of pose-based sign language recognition. We observed that joint-wise concatenation has advantages on both sides of performance and efficiency. As a result, our P3D outperformed the previous methods by a large margin on WLASL, jointly using 2D and 3D poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 613, + 408, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 613, + 408, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 613, + 408, + 627 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], the NRF grant [No. 2021M3A9E4080782] funded by the Korea government (MSIT), and the SNU-Samsung SDS Research Center." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20748" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In ECCV, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 168 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 203 + ], + "type": "text", + "content": "[3] Matyáš Boháček and Marek Hrúz. Sign pose-based transformer for word-level sign language recognition. In WACV, 2022. 2, 3, 8, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 247 + ], + "type": "text", + "content": "[4] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In CVPR, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "text", + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 317 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 350 + ], + "type": "text", + "content": "[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. 3, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "type": "text", + "content": "[8] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "text", + "content": "[9] Ka Leong Cheng, Zhaoyang Yang, Qifeng Chen, and Yu-Wing Tai. Fully convolutional networks for continuous sign language recognition. In ECCV, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 432, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 474 + ], + "type": "text", + "content": "[10] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "text", + "content": "[11] Runpeng Cui, Hu Liu, and Changshui Zhang. Recurrent convolutional neural networks for continuous sign language recognition by staged optimization. In CVPR, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "text", + "content": "[12] Runpeng Cui, Hu Liu, and Changshui Zhang. A deep neural framework for continuous sign language recognition by iterative training. IEEE Transactions on Multimedia (ToM), 21(7):1880-1891, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 555, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 621 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "text", + "content": "[14] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[15] Haodong Duan, Yue Zhao, Kai Chen, Dian Shao, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In CVPR, 2022. 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[16] Zhimin Gao, Peitao Wang, Pei Lv, Xiaoheng Jiang, Qidong Liu, Pichao Wang, Mingliang Xu, and Wanqing Li. Focal" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "and global spatial-temporal transformer for skeleton-based action recognition. In ACCV, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[17] Liuhao Ge, Hui Liang, Junsong Yuan, and Daniel Thalmann. 3d convolutional neural networks for efficient and robust hand pose estimation from single depth images. In CVPR, 2017. 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[18] Xiao Guo and Jongmoo Choi. Human motion prediction via learning local structure representations and temporal dependencies. In AAAI, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[19] Al Amin Hosain, Panneer Selvam Santhalingam, Parth Pathak, Huzefa Rangwala, and Jana Kosecka. Hand pose guided 3d pooling for word-level sign language recognition. In WACV, 2021. 3, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: pre-training of hand-model-aware representation for sign language recognition. In ICCV, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[21] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding, 2014. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "type": "text", + "content": "[22] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. In BMVC, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 344, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 398 + ], + "type": "text", + "content": "[23] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2014. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 434, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 456 + ], + "type": "text", + "content": "[25] Oscar Koller. Quantitative survey of the state of the art in sign language recognition, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 457, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 457, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 457, + 545, + 521 + ], + "type": "text", + "content": "[26] Oscar Koller, Necati Cihan Camgoz, Hermann Ney, and Richard Bowden. Weakly supervised learning with multistream cnn-lstm-hmms to discover sequential parallelism in sign language videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(9):2306-2320, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[27] Oscar Koller, Sepehr Zargaran, Hermann Ney, and Richard Bowden. Deep sign: Enabling robust statistical continuous sign language recognition via hybrid cnn-hmms. International Journal of Computer Vision (IJCV), 126:1311-1325, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[28] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[29] Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Cooccurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation. arXiv preprint arXiv:1804.06055, 2018. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[30] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, 2020. 1, 2, 3, 6, 7, 9" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20749" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[31] Dongxu Li, Xin Yu, Chenchen Xu, Lars Petersson, and Hongdong Li. Transferring cross-domain knowledge for video sign language recognition. In CVPR, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[32] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-parted graph scattering networks for 3d human motion prediction. In ECCV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 175 + ], + "type": "text", + "content": "[33] Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 209 + ], + "type": "text", + "content": "[34] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 287, + 255 + ], + "type": "text", + "content": "[35] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, Oct. 2015. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 258, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 258, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 287, + 291 + ], + "type": "text", + "content": "[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR Workshops, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 293, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 335 + ], + "type": "text", + "content": "[37] Gyeongsik Moon, Heeseung Kwon, Kyoung Mu Lee, and Minsu Cho. Integralaction: Pose-driven feature integration for robust human action recognition in videos. In CVPRW, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "text", + "content": "[38] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 373, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 426 + ], + "type": "text", + "content": "[39] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. *NeurIPS Workshop on Autodiff*, 2017. 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 430, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 473 + ], + "type": "text", + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 540 + ], + "type": "text", + "content": "[41] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "text", + "content": "[42] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In ICPR Workshops, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[43] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In CVPR, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[44] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3d multi-channel sign language production via progressive transformers and mixture density networks. International Journal of Computer Vision (IJCV), 129(7):2113-2135, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[45] Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. In AAAI, 2017. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 420 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[46] Du Tran, Lubomir Bourdev, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. ICCV, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[47] Anirudh Tunga, Sai Vidyaranya Nuthalapati, and Juan Wachs. Pose-based sign language recognition using gcn and bert. In WACV, 2021. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "text", + "content": "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 2008. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 1, 2, 3, 8, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "text", + "content": "[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In CVPR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "type": "text", + "content": "[52] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In ICCV, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 385 + ], + "type": "text", + "content": "[53] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[54] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20750" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file